diff --git a/.gitattributes b/.gitattributes index 5ddbc4bde21a254eb519dea7a06b42345c8c981c..1fcffcb2691c3a2e91f4932d0f79d8e1f643f418 100644 --- a/.gitattributes +++ b/.gitattributes @@ -1717,3 +1717,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2024/Pseudo-Embedding[[:space:]]for[[:space:]]Generalized[[:space:]]Few-Shot[[:space:]]Point[[:space:]]Cloud[[:space:]]Segmentation/46d5c6d5-8872-42b2-84ce-48ea440f2084_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/Pseudo-Labelling[[:space:]]Should[[:space:]]Be[[:space:]]Aware[[:space:]]of[[:space:]]Disguising[[:space:]]Channel[[:space:]]Activations/6a83183f-b3b5-4621-912f-6229a2e9823f_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/Pseudo-RIS_[[:space:]]Distinctive[[:space:]]Pseudo-supervision[[:space:]]Generation[[:space:]]for[[:space:]]Referring[[:space:]]Image[[:space:]]Segmentation/412ae3f8-6c16-4b34-b46d-6f75350352b8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Pseudo-keypoint[[:space:]]RKHS[[:space:]]Learning[[:space:]]for[[:space:]]Self-supervised[[:space:]]6DoF[[:space:]]Pose[[:space:]]Estimation/e3acf3fe-7127-4839-bf51-8bc132ae63e7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Put[[:space:]]Myself[[:space:]]in[[:space:]]Your[[:space:]]Shoes_[[:space:]]Lifting[[:space:]]the[[:space:]]Egocentric[[:space:]]Perspective[[:space:]]from[[:space:]]Exocentric[[:space:]]Videos/cf2ae3fa-fba8-4d91-a4a1-6c60e6a8095b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Pyramid[[:space:]]Diffusion[[:space:]]for[[:space:]]Fine[[:space:]]3D[[:space:]]Large[[:space:]]Scene[[:space:]]Generation/0615ce43-c6d2-40fd-ba98-d83c475c6139_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Q&A[[:space:]]Prompts_[[:space:]]Discovering[[:space:]]Rich[[:space:]]Visual[[:space:]]Clues[[:space:]]through[[:space:]]Mining[[:space:]]Question-Answer[[:space:]]Prompts[[:space:]]for[[:space:]]VQA[[:space:]]requiring[[:space:]]Diverse[[:space:]]World[[:space:]]Knowledge/d6b849aa-8c9f-421c-b457-5802b3bd4aca_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/QUAR-VLA_[[:space:]]Vision-Language-Action[[:space:]]Model[[:space:]]for[[:space:]]Quadruped[[:space:]]Robots/7817397f-b04e-4f74-8ceb-92291da9f1da_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Quality[[:space:]]Assured_[[:space:]]Rethinking[[:space:]]Annotation[[:space:]]Strategies[[:space:]]in[[:space:]]Imaging[[:space:]]AI/1dd3eeac-8808-45a9-8928-d1c207e0686a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Quanta[[:space:]]Video[[:space:]]Restoration/c31fed1f-5357-4c1e-adf0-39121e1936f9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Quantization-Friendly[[:space:]]Winograd[[:space:]]Transformations[[:space:]]for[[:space:]]Convolutional[[:space:]]Neural[[:space:]]Networks/90e7434b-cd12-4e62-9976-08dc6035e0cd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Quantized[[:space:]]Prompt[[:space:]]for[[:space:]]Efficient[[:space:]]Generalization[[:space:]]of[[:space:]]Vision-Language[[:space:]]Models/d406d327-feb0-4aee-91eb-de3562f63166_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/QueryCDR_[[:space:]]Query-based[[:space:]]Controllable[[:space:]]Distortion[[:space:]]Rectification[[:space:]]Network[[:space:]]for[[:space:]]Fisheye[[:space:]]Images/8182781d-9f7c-4e2a-86d5-0c2e6b9982a2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/R.A.C.E._[[:space:]]Robust[[:space:]]Adversarial[[:space:]]Concept[[:space:]]Erasure[[:space:]]for[[:space:]]Secure[[:space:]]Text-to-Image[[:space:]]Diffusion[[:space:]]Model/547862c1-28d0-440c-98c1-6204b0fe03dd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/R3D-AD_[[:space:]]Reconstruction[[:space:]]via[[:space:]]Diffusion[[:space:]]for[[:space:]]3D[[:space:]]Anomaly[[:space:]]Detection/a722039e-753c-4cd1-9dab-7069c8f25f9e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/R3DS_[[:space:]]Reality-linked[[:space:]]3D[[:space:]]Scenes[[:space:]]for[[:space:]]Panoramic[[:space:]]Scene[[:space:]]Understanding/7708992f-16b8-4792-b8c3-10a97f869f8d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/RANRAC_[[:space:]]Robust[[:space:]]Neural[[:space:]]Scene[[:space:]]Representations[[:space:]]via[[:space:]]Random[[:space:]]Ray[[:space:]]Consensus/de9512b1-9e5d-48ed-a48f-d5995f4754ec_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/RAP_[[:space:]]Retrieval-Augmented[[:space:]]Planner[[:space:]]for[[:space:]]Adaptive[[:space:]]Procedure[[:space:]]Planning[[:space:]]in[[:space:]]Instructional[[:space:]]Videos/8baf20b2-0efa-428f-a4d5-15cb7958ae91_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/RAPiD-Seg_[[:space:]]Range-Aware[[:space:]]Pointwise[[:space:]]Distance[[:space:]]Distribution[[:space:]]Networks[[:space:]]for[[:space:]]3D[[:space:]]LiDAR[[:space:]]Segmentation/b19a59c8-eca1-4aae-977c-6a71eef1622b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/RAVE_[[:space:]]Residual[[:space:]]Vector[[:space:]]Embedding[[:space:]]for[[:space:]]CLIP-Guided[[:space:]]Backlit[[:space:]]Image[[:space:]]Enhancement/7f0c9e4e-7b51-4d42-aac8-de63b12e0ec8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/RAW-Adapter_[[:space:]]Adapting[[:space:]]Pretrained[[:space:]]Visual[[:space:]]Model[[:space:]]to[[:space:]]Camera[[:space:]]RAW[[:space:]]Images/a7bce0f0-b64e-4428-b3dc-7b7e2a217569_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/RCS-Prompt_[[:space:]]Learning[[:space:]]Prompt[[:space:]]to[[:space:]]Rearrange[[:space:]]Class[[:space:]]Space[[:space:]]for[[:space:]]Prompt-based[[:space:]]Continual[[:space:]]Learning/6bf9dca4-8b37-4308-ad56-3bfced51f84f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/REDIR_[[:space:]]Refocus-free[[:space:]]Event-based[[:space:]]De-occlusion[[:space:]]Image[[:space:]]Reconstruction/59c82a8e-f2d6-49fa-a29f-6faeb9b1bbb2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/REFRAME_[[:space:]]Reflective[[:space:]]Surface[[:space:]]Real-Time[[:space:]]Rendering[[:space:]]for[[:space:]]Mobile[[:space:]]Devices/f5175c1c-adca-4484-8e6b-cf9e011d3e14_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/REVISION_[[:space:]]Rendering[[:space:]]Tools[[:space:]]Enable[[:space:]]Spatial[[:space:]]Fidelity[[:space:]]in[[:space:]]Vision-Language[[:space:]]Models/f77a4b99-c39a-4693-bb9e-aff60ba50459_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/RGBD[[:space:]]GS-ICP[[:space:]]SLAM/794618d9-8655-4def-b2e9-3b1608333e79_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/RGNet_[[:space:]]A[[:space:]]Unified[[:space:]]Clip[[:space:]]Retrieval[[:space:]]and[[:space:]]Grounding[[:space:]]Network[[:space:]]for[[:space:]]Long[[:space:]]Videos/f091b053-bc8d-4374-9687-69b5bf80ade6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/RICA^2_[[:space:]]Rubric-Informed,[[:space:]]Calibrated[[:space:]]Assessment[[:space:]]of[[:space:]]Actions/805cf405-290f-425f-a265-d9aa43a99eb8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/RING-NeRF[[:space:]]_[[:space:]]Rethinking[[:space:]]Inductive[[:space:]]Biases[[:space:]]for[[:space:]]Versatile[[:space:]]and[[:space:]]Efficient[[:space:]]Neural[[:space:]]Fields/5e093f34-06af-49a4-a5fe-2e79d5819847_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/RISurConv_[[:space:]]Rotation[[:space:]]Invariant[[:space:]]Surface[[:space:]]Attention-Augmented[[:space:]]Convolutions[[:space:]]for[[:space:]]3D[[:space:]]Point[[:space:]]Cloud[[:space:]]Classification[[:space:]]and[[:space:]]Segmentation/85a1ebc6-6b7d-4e16-8cd8-c420684dc80d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/RPBG_[[:space:]]Towards[[:space:]]Robust[[:space:]]Neural[[:space:]]Point-based[[:space:]]Graphics[[:space:]]in[[:space:]]the[[:space:]]Wild/ced6023f-74f5-424c-b6a1-715ff57d9512_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/RS-NeRF_[[:space:]]Neural[[:space:]]Radiance[[:space:]]Fields[[:space:]]from[[:space:]]Rolling[[:space:]]Shutter[[:space:]]Images/cafbceee-b1fc-44b7-b4f6-4691cb686eec_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/RSL-BA_[[:space:]]Rolling[[:space:]]Shutter[[:space:]]Line[[:space:]]Bundle[[:space:]]Adjustment/25dc3962-792b-46d2-8db3-b73a1efe60d9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/RT-Pose_[[:space:]]A[[:space:]]4D[[:space:]]Radar-Tensor[[:space:]]based[[:space:]]3D[[:space:]]Human[[:space:]]Pose[[:space:]]Estimation[[:space:]]and[[:space:]]Localization[[:space:]]Benchmark/e4fd786f-06ad-4f57-8db1-c3bab37f210b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/R^2-Bench_[[:space:]]Benchmarking[[:space:]]the[[:space:]]Robustness[[:space:]]of[[:space:]]Referring[[:space:]]Perception[[:space:]]Models[[:space:]]under[[:space:]]Perturbations/3cf440ac-b4a0-433a-943f-d4d176a64132_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/R^2-Tuning_[[:space:]]Efficient[[:space:]]Image-to-Video[[:space:]]Transfer[[:space:]]Learning[[:space:]]for[[:space:]]Video[[:space:]]Temporal[[:space:]]Grounding/148e844e-b528-47e5-9db4-f50b0b9b4b9c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/RaFE_[[:space:]]Generative[[:space:]]Radiance[[:space:]]Fields[[:space:]]Restoration/8e187213-9e85-483c-8762-09f958b4b9d0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/RadEdit_[[:space:]]stress-testing[[:space:]]biomedical[[:space:]]vision[[:space:]]models[[:space:]]via[[:space:]]diffusion[[:space:]]image[[:space:]]editing/730452c5-1675-40d8-8c46-f903c2fa2703_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Radiative[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]for[[:space:]]Efficient[[:space:]]X-ray[[:space:]]Novel[[:space:]]View[[:space:]]Synthesis/f615a129-2ee2-41a8-9f1c-82a796edae0e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Raindrop[[:space:]]Clarity_[[:space:]]A[[:space:]]Dual-Focused[[:space:]]Dataset[[:space:]]for[[:space:]]Day[[:space:]]and[[:space:]]Night[[:space:]]Raindrop[[:space:]]Removal/64be914f-41c4-4c4e-9cfc-b492879688db_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Raising[[:space:]]the[[:space:]]Ceiling_[[:space:]]Conflict-Free[[:space:]]Local[[:space:]]Feature[[:space:]]Matching[[:space:]]with[[:space:]]Dynamic[[:space:]]View[[:space:]]Switching/f888146a-2080-4739-ac33-923c8980cbb0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Random[[:space:]]Walk[[:space:]]on[[:space:]]Pixel[[:space:]]Manifolds[[:space:]]for[[:space:]]Anomaly[[:space:]]Segmentation[[:space:]]of[[:space:]]Complex[[:space:]]Driving[[:space:]]Scenes/40e3d166-d87c-43d0-aeb7-8851ed372fb1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/RangeLDM_[[:space:]]Fast[[:space:]]Realistic[[:space:]]LiDAR[[:space:]]Point[[:space:]]Cloud[[:space:]]Generation/ec4004eb-e76a-4f41-9891-f055a3eeaf7c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Rasterized[[:space:]]Edge[[:space:]]Gradients_[[:space:]]Handling[[:space:]]Discontinuities[[:space:]]Differentially/ac83eb0e-ca7d-4471-8db5-05179e149303_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Rate-Distortion-Cognition[[:space:]]Controllable[[:space:]]Versatile[[:space:]]Neural[[:space:]]Image[[:space:]]Compression/d4e52d64-79ee-43c9-8953-86dbf5d724e6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Rawformer_[[:space:]]Unpaired[[:space:]]Raw-to-Raw[[:space:]]Translation[[:space:]]for[[:space:]]Learnable[[:space:]]Camera[[:space:]]ISPs/c3c8340d-834b-4e08-bdff-c9515d0d2660_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Ray[[:space:]]Denoising_[[:space:]]Depth-aware[[:space:]]Hard[[:space:]]Negative[[:space:]]Sampling[[:space:]]for[[:space:]]Multi-view[[:space:]]3D[[:space:]]Object[[:space:]]Detection/ebbe937b-35fd-41db-b04a-6c0c4e0d90c8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Ray-Distance[[:space:]]Volume[[:space:]]Rendering[[:space:]]for[[:space:]]Neural[[:space:]]Scene[[:space:]]Reconstruction/e689642e-0cb9-456e-9b14-7ba1bd1312a3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ReALFRED_[[:space:]]An[[:space:]]Embodied[[:space:]]Instruction[[:space:]]Following[[:space:]]Benchmark[[:space:]]in[[:space:]]Photo-Realistic[[:space:]]Environments/0e4c20a8-1ef4-4401-b5cc-41d9bc314f7b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ReCON_[[:space:]]Training-Free[[:space:]]Acceleration[[:space:]]for[[:space:]]Text-to-Image[[:space:]]Synthesis[[:space:]]with[[:space:]]Retrieval[[:space:]]of[[:space:]]Concept[[:space:]]Prompt[[:space:]]Trajectories/3287df3f-9065-4d00-b93c-c529c48f7c92_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ReGround_[[:space:]]Improving[[:space:]]Textual[[:space:]]and[[:space:]]Spatial[[:space:]]Grounding[[:space:]]at[[:space:]]No[[:space:]]Cost/e369c9ee-bed9-4dbc-92b4-a4d9da7bccb2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ReLoo_[[:space:]]Reconstructing[[:space:]]Humans[[:space:]]Dressed[[:space:]]in[[:space:]]Loose[[:space:]]Garments[[:space:]]from[[:space:]]Monocular[[:space:]]Video[[:space:]]in[[:space:]]the[[:space:]]Wild/e08d3ae3-38a4-42db-b159-2c5a46fac242_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ReMamber_[[:space:]]Referring[[:space:]]Image[[:space:]]Segmentation[[:space:]]with[[:space:]]Mamba[[:space:]]Twister/04366454-5dd3-4b5b-8a0b-a5b01b0efbcd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ReMatching_[[:space:]]Low-Resolution[[:space:]]Representations[[:space:]]for[[:space:]]Scalable[[:space:]]Shape[[:space:]]Correspondence/8dd15a7c-48bf-42e4-a752-0b0dbad83c91_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ReMoS_[[:space:]]3D[[:space:]]Motion-Conditioned[[:space:]]Reaction[[:space:]]Synthesis[[:space:]]for[[:space:]]Two-Person[[:space:]]Interactions/a734f12f-489f-47c4-b3e9-744b3d03eee5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ReNoise_[[:space:]]Real[[:space:]]Image[[:space:]]Inversion[[:space:]]Through[[:space:]]Iterative[[:space:]]Noising/4a999b7d-d6fa-416b-8345-a221ff72abc9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/RePOSE_[[:space:]]3D[[:space:]]Human[[:space:]]Pose[[:space:]]Estimation[[:space:]]via[[:space:]]Spatio-Temporal[[:space:]]Depth[[:space:]]Relational[[:space:]]Consistency/7257623d-d6ee-4c47-b288-fda8cfd3af7e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ReSyncer_[[:space:]]Rewiring[[:space:]]Style-based[[:space:]]Generator[[:space:]]for[[:space:]]Unified[[:space:]]Audio-Visually[[:space:]]Synced[[:space:]]Facial[[:space:]]Performer/6fe8e31e-41da-4533-8af1-187e2be9d843_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Real[[:space:]]Appearance[[:space:]]Modeling[[:space:]]for[[:space:]]More[[:space:]]General[[:space:]]Deepfake[[:space:]]Detection/2bfe37cc-d80a-473f-a5d8-70fca9af4b74_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Real-data-driven[[:space:]]2000[[:space:]]FPS[[:space:]]Color[[:space:]]Video[[:space:]]from[[:space:]]Mosaicked[[:space:]]Chromatic[[:space:]]Spikes/6731bc50-147c-4c01-a7c6-498656963ba3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Real-time[[:space:]]3D-aware[[:space:]]Portrait[[:space:]]Editing[[:space:]]from[[:space:]]a[[:space:]]Single[[:space:]]Image/c7f6568e-666e-490d-b059-9d93ade8c6f3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Real-time[[:space:]]Holistic[[:space:]]Robot[[:space:]]Pose[[:space:]]Estimation[[:space:]]with[[:space:]]Unknown[[:space:]]States/0b349d89-54a2-48da-a5a9-0c888f46a6ff_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/RealGen_[[:space:]]Retrieval[[:space:]]Augmented[[:space:]]Generation[[:space:]]for[[:space:]]Controllable[[:space:]]Traffic[[:space:]]Scenarios/4b2a9c03-0a68-49b3-9375-225cd793a8b3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/RealViformer_[[:space:]]Investigating[[:space:]]Attention[[:space:]]for[[:space:]]Real-World[[:space:]]Video[[:space:]]Super-Resolution/39444a96-c968-4d03-9b6c-b416cbbaab18_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Realistic[[:space:]]Human[[:space:]]Motion[[:space:]]Generation[[:space:]]with[[:space:]]Cross-Diffusion[[:space:]]Models/5b1805dd-89b3-4df9-9e0a-1ed375e3f44c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Reason2Drive_[[:space:]]Towards[[:space:]]Interpretable[[:space:]]and[[:space:]]Chain-based[[:space:]]Reasoning[[:space:]]for[[:space:]]Autonomous[[:space:]]Driving/66f968e7-ebed-40e1-a3ff-b107a9813c0e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Rebalancing[[:space:]]Using[[:space:]]Estimated[[:space:]]Class[[:space:]]Distribution[[:space:]]for[[:space:]]Imbalanced[[:space:]]Semi-Supervised[[:space:]]Learning[[:space:]]under[[:space:]]Class[[:space:]]Distribution[[:space:]]Mismatch/c1f56d15-2542-45a0-80c7-7ebab6520f43_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2024/Pseudo-keypoint RKHS Learning for Self-supervised 6DoF Pose Estimation/e3acf3fe-7127-4839-bf51-8bc132ae63e7_content_list.json b/2024/Pseudo-keypoint RKHS Learning for Self-supervised 6DoF Pose Estimation/e3acf3fe-7127-4839-bf51-8bc132ae63e7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1bdc2a24b61dcfaec36a5a1f5cbcb798abe9baaf --- /dev/null +++ b/2024/Pseudo-keypoint RKHS Learning for Self-supervised 6DoF Pose Estimation/e3acf3fe-7127-4839-bf51-8bc132ae63e7_content_list.json @@ -0,0 +1,1905 @@ +[ + { + "type": "text", + "text": "Pseudo-keypoint RKHS Learning for Self-supervised 6DoF Pose Estimation", + "text_level": 1, + "bbox": [ + 176, + 48, + 815, + 104 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yangzheng Wu and Michael Greenspan", + "bbox": [ + 276, + 139, + 721, + 159 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "RCVLab, Dept. of Electrical and Computer Engineering, Ingenuity Labs, Queen's University, Kingston, Ontario, Canada {y.wu, michael.greenspan}@queensu.ca", + "bbox": [ + 142, + 173, + 854, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We address the simulation-to-real domain gap in six degree-of-freedom pose estimation (6DoF PE), and propose a novel self-supervised keypoint voting-based 6DoF PE framework, effectively narrowing this gap using a learnable kernel in RKHS. We formulate this domain gap as a distance in high-dimensional feature space, distinct from previous iterative matching methods. We propose an adapter network, which is pre-trained on purely synthetic data with synthetic ground truth poses, and which evolves the network parameters from this source synthetic domain to the target real domain. Importantly, the real data training only uses pseudo-poses estimated by pseudo-keypoints, and thereby requires no real ground truth data annotations. Our proposed method is called RKHSPose, and achieves state-of-the-art performance among self-supervised methods on three commonly used 6DoF PE datasets including LINEMOD $(+4.2\\%)$ , Occlusion LINEMOD $(+2\\%)$ , and YCB-Video $(+3\\%)$ . It also compares favorably to fully supervised methods on all six applicable BOP core datasets, achieving within $-11.3\\%$ to $+0.2\\%$ of the top fully supervised results.", + "bbox": [ + 147, + 269, + 861, + 572 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: pose estimation $\\cdot$ self-supervision $\\cdot$ domain adaptation $\\cdot$ key-point estimation", + "bbox": [ + 147, + 590, + 849, + 626 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 82, + 656, + 307, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "RGB-D Six Degree-of-Freedom Pose Estimation (6DoF PE) is a problem being actively explored in computer vision research. Given an RGB image and its associated depth map, the task is to detect scene objects and estimate their poses comprising 3DoF rotational angles and 3DoF translational offsets in the camera reference frame. This task enables many applications such as augmented reality [34,36,40,84], robotic bin picking [19,35,43], autonomous driving [51,85] and image-guided surgeries [21,25].", + "bbox": [ + 79, + 694, + 917, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As with other machine learning (ML) tasks, fully supervised 6DoF PE requires large annotated datasets. This requirement is particularly challenging for 6DoF PE, as the annotations comprise not only the identity of the objects in the scene, but also their 6DoF pose, which makes the data relatively expensive to annotate compared to related tasks such as classification, detection and segmentation. This is due to the fact that humans are not able to qualitatively or", + "bbox": [ + 79, + 832, + 917, + 947 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6d7dd66c9b3040ba2b96e07a7a37270b2ed6f71f87197eeee34a608ef6ebd606.jpg", + "image_caption": [ + "Fig. 1: RKHSPose adapts the network pretrained on synthetic data to real test scenes (left), by comparing network feature spaces with real image inputs (solid arrows), against those with syn/real image (right) inputs (dashed arrows). $M_r$ regresses radial quantities, $M_A$ is the Adapter network, and RKHS maps features into a higher dimensional space." + ], + "image_footnote": [], + "bbox": [ + 96, + 50, + 903, + 195 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "intuitively estimate 6DoF pose, which therefore requires additional instrumentation of the scene at data collection, and/or more sophisticated user annotation tools [34,84]. Consequently, synthetic 6DoF PE datasets [36,43] have been introduced, either as an additional complement to real datasets, or as standalone purely synthetic datasets. Annotated synthetic datasets are of course trivially inexpensive to collect, simply because precise synthetic pose annotation in a simulated environment is fully automatic. A known challenge in using synthetic data, however, is that there typically exists a domain gap between the real and synthetic data, which makes results less accurate when inferring in real data using models trained on purely synthetic datasets. Expectations of the potential benefit of synthetic datasets has led to the exploration of a rich set of Domain Adaptation (DA) methods, which specifically aim to reduce the domain gap [6,62,88] using inexpensive synthetic data for a wide variety of tasks, recently including 6DoF PE [28,45,47].", + "bbox": [ + 72, + 373, + 922, + 647 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early methods [42, 83] ignored the simulation-to-real (sim2real) domain gap and nevertheless improved performance by training on both synthetic and real annotated data, effectively augmenting the real images with the synthetic. However, these methods still required real labels to be sufficiently accurate and robust for practical applications, partially due to the domain gap. As shown in Fig. 1, the rendered synthetic objects (right image) have a slightly different appearance than the real objects (left image). The details of the CAD models, both geometry and texture, are not precise, as can be seen for the can object (which lacks a mouth and shadows), the benchwise (which is missing a handle), and the holepuncher (which has coarse geometric resolution). Several recent methods [9, 44, 73, 75, 80] have started to address the sim2real gap for 6DoF PE by first training on labeled synthetic data and then fine-tuning on unlabeled real data. Commonly known as self-supervised, these methods reduce the domain gap by adding extra supervision using features extracted from real images without requiring real ground truth (GT) labels. The majority of these methods are view-", + "bbox": [ + 72, + 656, + 922, + 949 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 77, + 17, + 94, + 29 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y.Wu and M. Greenspan", + "bbox": [ + 159, + 14, + 406, + 32 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "point/template matching-based, and the self-supervision commonly iteratively matches 2D masks or 3D point clouds [44,73,80].", + "bbox": [ + 77, + 55, + 917, + 95 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While the above-mentioned self-supervision works have shown promise, there exist a wealth of DA techniques that can be brought to bear to improve performance further for this task. One such technique is Reproducing Kernel Hilbert Space (RKHS), which is a kernel method that has been shown to be effective for DA [1,5,41]. RKHS was initially used to create decision boundaries for non-separable data [12,60], and has been shown to be effective at reducing the domain gap for various tasks and applications [7,65,89]. The reproducing kernel guarantees that the domain gap can be statistically measured, allowing network parameters trained on synthetic data to be effectively adapted to the real data, using specifically tailored metrics.", + "bbox": [ + 77, + 95, + 917, + 289 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the sim2real domain gap in 6DoF PE, we propose RKHSPose, which is a keypoint-based method [82, 83] trained on a mixture of a large collection of labeled synthetic data, and a small handful of unlabeled real data. RKHSPose estimates the intermediate radial voting quantity, which has been shown to be effective for estimating keypoints [83], by first training a modified FCN-Resnet-18 on purely synthetic data, with automatically labeled synthetic GT poses. The radial quantity is a map of the distance from each image pixel to each keypoint. Next, real images are passed through the synthetically trained network, resulting in a set of pseudo-keypoints. The real images and their corresponding pseudo-keypoints are used to render a set of 'synthetic-overreal' (syn/real) images, by first estimating the pseudo-pose from the pseudo-keypoints, and then overlaying the synthetic object, rendered with the pseudopose, onto the real image. The network training then continues on the syn/real images, invoking an RKHS network module with a trainable linear product kernel, which minimizes the Maximum Mean Discrepancy (MMD) loss. At the front end, a proposed keypoint radial voting network learns to cast votes to estimate keypoints from the backend-generated radial maps. The final pose is then determined using ePnP [46] based on the estimated and corresponding object keypoints.", + "bbox": [ + 77, + 289, + 917, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main contributions of this work are:", + "bbox": [ + 116, + 658, + 544, + 676 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A novel learnable RKHS-based Adapter backend network architecture to minimize the sim2real domain gap in 6DoF PE;", + "- A novel CNN-based frontend network for keypoint radial voting;", + "- A self-supervised keypoint-based 6DoF PE method, RKHSPose, which is shown to have state-of-the-art (SOTA) performance, based on our experiments and several ablation studies." + ], + "bbox": [ + 89, + 684, + 917, + 798 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 77, + 824, + 331, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 6DoF PE", + "text_level": 1, + "bbox": [ + 79, + 861, + 259, + 879 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ML-based 6DoF PE methods [57,61,84] all train a network to regress quantities, such as keypoints and camera viewpoints, as have been used in classical algorithms [34]. ML-based methods, which initially became popular for the general", + "bbox": [ + 79, + 889, + 917, + 947 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "RKHPose", + "bbox": [ + 726, + 14, + 837, + 30 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 903, + 16, + 917, + 30 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/086f41ad63e4e6f3f883610df8965df6ce7c48a1f886130a8b4ec37c03fa14c0.jpg", + "table_caption": [ + "Table 1: Existing self-supervised 6DoF PE methods. Some methods use CAD models, labeled synthetic (syn) data (images+poses), and real data without GT labels, while others use only labeled synthetic data. Many methods require a ROI such as a bounding box or a semantic mask, manually labeled or estimated by an existing framework. SO-Pose [18] used very few real labels to improve performance." + ], + "table_footnote": [], + "table_body": "
methodmodeCAD modelsyn datareal imagesreal posesROI
SO-Pose [18]RGB✓✓✓✓✓
TexPose [9]RGB✓✓✓X✓
SMOC-Net [75]RGB✓✓✓X✓
FS6D [33]RGBD✓✓XX✓
AAE [73]RGB✓✓XX✓
MHP [54]RGB✓✓XX✓
Sock et al. [68]RGB✓✓XX✓
DSC [86]RGBD✓✓XX✓
Sundermeyer [71]RGB✓✓XX✓
LatentFusion [59]RGBX✓XX✓
OSOP [66]RGBD✓✓XXX
Kleeberger et al. [44]D✓✓XXX
Su et al. [69]RGB✓✓✓XX
Self6D [80]RGBD✓✓✓XX
Self6D++ [79]RGB✓✓✓XX
Deng et al. [15]RGBD✓✓✓XX
RKHSPose (Ours)RGBD✓✓✓XX
", + "bbox": [ + 205, + 157, + 789, + 505 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "object detection task, have started to dominate the 6DoF PE literature due to their accuracy and efficiency.", + "bbox": [ + 79, + 567, + 917, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "There are two main categories of ML-based fully supervised methods: feature matching-based [29, 30, 57, 70, 84], and keypoint-based methods [31, 32, 61, 87]. Feature matching-based methods make use of the structures from general object detection networks most directly. The network encodes and matches features and estimates pose by either regressing elements of the pose (e.g. the transformation matrix [75], rotational angles and translational offsets [84] or 3D vertices [70]) directly, or by regressing some intermediate feature-matching representations, such as viewpoints or segments.", + "bbox": [ + 79, + 609, + 917, + 765 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast, keypoint-based methods encode features to estimate keypoints which are predefined within the reference frame of an object's CAD model. These methods then use (modified) classical algorithms such as PnP [22, 46], Horn's method [38], and ICP [3] to estimate the final pose from corresponding image and model keypoints. Unlike feature-matching methods, keypoint-based methods are typically more accurate due to redundancies encountered through voting schemes [32, 61, 83] and by generating confidence hypotheses of keypoints [31, 87].", + "bbox": [ + 79, + 769, + 917, + 905 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recently, self-supervised 6DoF PE methods have been explored in order to reduce the reliance on labeled real data, which is expensive to acquire. As sum-", + "bbox": [ + 79, + 908, + 917, + 947 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 79, + 17, + 94, + 29 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y.Wu and M. Greenspan", + "bbox": [ + 159, + 14, + 404, + 30 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "marized in Table 1, these methods commonly use real images without GT labels. Some methods use pure synthetic data and CAD models only, with the exception of LatentFusion [59] which trained the model using only synthetic data. The majority of these methods [9, 15, 33, 48, 66, 68, 69, 71, 75, 79, 80] are inspired by fully supervised feature-matching methods, except DPODv2 [67] and DSC [86], in which keypoint correspondences are rendered and matched.", + "bbox": [ + 77, + 53, + 917, + 172 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A few methods [15,66] fine-tuned the pose trivially by iteratively matching the template/viewpoint, whereas others [33,44,59,86] augmented the training data by adding noise [44,86], rendering textures [33,67] and creating a latent space [59]. Some methods [54,71,73] also implemented DA techniques such as codebook encoding [71], Principle Component Analysis (PCA) [71] and symmetric Bingham distributions [54]. Most methods [66,68,69,79,80,86] used rendering techniques to render and match a template. There are a few methods that combined 3D reconstruction techniques, such as Neural radiance fields (Nerf) [55] and Structure from Motion (SfM) [76]. TexPose [9] matched CAD models to segments generated by Nerf, and SMOC-Net [75] used SfM to create the 3D segment and matched with the CAD model.", + "bbox": [ + 77, + 178, + 917, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.2 Kernel Methods and Deep Learning", + "text_level": 1, + "bbox": [ + 79, + 456, + 581, + 478 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While deep learning is the most common ML technique within the computer vision literature, kernel methods [12,60,64] have also been actively explored. Kernel methods are typically in RKHS space [74] with reproducing properties that facilitate solving non-linear problems by mapping input data into high dimensional spaces that can be linearly separated [11,23]. Well-known early kernel methods that have been applied to computer vision are Support Vector Machines [12] and PCA [60]. A recent method [64] linked energy distance and MMD in RKHS, and showed the effectiveness of kernels in statistical hypothesis testing.", + "bbox": [ + 79, + 525, + 917, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "More recent studies compare kernel methods with deep learning networks [1, 4,5,41]. RKHS is found to perform better on classification tasks than one single block of a CNN comprising convolution, pooling and downsampling (linear) layers [41]. RKHS can also help with CNN generalization by meta-regularization on image registration problems [1]. Similarly, norms (magnitude of trainable weights) defined in RKHS help with CNN regularization [4,5]. Further, discriminant information in label distributions in unsupervised DA is addressed and RKHS-based Conditional Kernel Bures metric is proposed [50]. Lastly, the connection between Neural Tangent Kernel and MMD is established, and an efficient MMD for two-sample tests is developed [10].", + "bbox": [ + 79, + 687, + 917, + 882 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by the previous work, RKHSPose applies concepts of kernel learning to keypoint-based 6DoF PE, to provide an effective means to self-supervise a synthetically trained network on unlabeled real data.", + "bbox": [ + 79, + 889, + 917, + 947 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "RKHPose", + "bbox": [ + 726, + 14, + 837, + 30 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 903, + 16, + 917, + 30 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fe1acf770a9f08cf67b785bce1bf8b9cfa652b18d5242f45ceffc0883be82151.jpg", + "image_caption": [ + "Fig. 2: RKHSPose architecture. RKHPose is first trained on synthetic labeled data (solid arrows), and then finetuned on alternating syn/real and (unlabeled) real images (dashed arrows). $M_A$ is measured by MMD in RKHS by densely mapping the intermediate features of $M_r$ into high dimensional spaces with conv blocks. The distance is treated as $\\mathcal{L}_{M_A}$ and back-propagated through $M_A$ and $M_r$ ." + ], + "image_footnote": [], + "bbox": [ + 106, + 52, + 707, + 349 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b39f96337e2343f7784c0e00651d956680d5ccf78ea772c2927bf9a92dd30c72.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 58, + 888, + 139 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dd85496e0dc56714813d5722c57ab3aebe87c652c635f9995a8cb599294accce.jpg", + "image_caption": [ + "$K_{syn}$", + "$K_{syn / real}$" + ], + "image_footnote": [], + "bbox": [ + 728, + 152, + 891, + 235 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f704538c2f497cab3eea1ec82dc383a7e159ba121a1acb97190844cf6e4a056e.jpg", + "image_caption": [ + "$K_{real}$" + ], + "image_footnote": [], + "bbox": [ + 728, + 247, + 888, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 75, + 504, + 244, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Network Overview", + "text_level": 1, + "bbox": [ + 75, + 546, + 372, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Fig. 2, RKHSPose is made up of two networks, a main network $M_{rv}$ for keypoint regression and classification, and an Adapter network $M_A$ . The input of $M_{rv}$ with shape $W \\times H \\times 4$ , is the concatenation of an RGB image $I$ and its corresponding depth map $D$ . This input can be synthetically generated with an arbitrary background ( $I_{syn}$ and $D_{syn}$ ), a synthetic mask overlayed on a real background ( $I_{syn/real}$ and $D_{syn/real}$ ), or a real ( $I_{real}$ and $D_{real}$ ) image. The outputs are $n$ projected 2D keypoints $K$ , along with corresponding classification labels $C$ and confidence scores $S$ . $K$ are organized into instance sets based on $C$ and geometric constraints.", + "bbox": [ + 72, + 578, + 920, + 752 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$M_{rv}$ comprises two sub-networks, regression network $M_r$ and voting network $M_v$ . Inspired by recent voting techniques [32, 61, 82, 83, 90], $M_r$ estimates an intermediate voting quantity, which is a radial distance map $V_r$ [83], by using a modified Fully Connected ResNet 18 (FCN-ResNet-18). The radial voting map $V_r$ , with shape $W \\times H$ , stores the Euclidean distance from each object point to each keypoint in the 3D camera world reference frame. The voting network $M_v$ (described in Sec. 3.3) then takes $V_r$ as input, accumulates votes, and detects peaks to estimate $K$ , $C$ and $S$ .", + "bbox": [ + 72, + 754, + 917, + 908 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The Adapter network $M_A$ consists of a series of CNNs which encode pairs of feature maps from $M_{rv}$ , and are trained on both synthetic overlayed and pure real", + "bbox": [ + 75, + 910, + 920, + 949 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 77, + 17, + 94, + 29 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y.Wu and M. Greenspan", + "bbox": [ + 159, + 14, + 406, + 32 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "data. $M_A$ encodes feature map pairs $(f_{syn / real}, f_{real})$ into corresponding high-dimensional feature maps $(h_{syn / real}, h_{real})$ . The input data, $(I_{syn / real}, I_{real})$ and $(D_{syn / real}, D_{real})$ , are also treated as $(f_{syn / real}, f_{real})$ during the learning of $M_A$ . Each of these networks creates a high-dimensional latent space, essentially the Reproducing Kernel Hilbert Space (RKHS) [2] and contributes to the learning of both $M_{rv}$ and $M_A$ by calculating the MMD [27]. While RKHS has been applied effectively to other DA tasks, to our knowledge, this is the first time that it has been applied to 6DoF PE, and the adapter network architecture is novel.", + "bbox": [ + 72, + 55, + 920, + 209 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The loss function $\\mathcal{L}$ is made up of five elements: Radial regression loss $\\mathcal{L}_r$ for $V_r$ ; Keypoint projection loss $\\mathcal{L}_k$ for $K$ ; Classification loss $\\mathcal{L}_c$ for $C$ ; Confidence loss $\\mathcal{L}_s$ for $S$ , and finally; Adapter loss $\\mathcal{L}_A$ for the comparison of intermediate feature maps. The regression losses $\\mathcal{L}_r$ , $\\mathcal{L}_k$ , and $\\mathcal{L}_s$ all use the smooth L1 metric, whereas classification loss $\\mathcal{L}_c$ uses the cross-entropy metric $H(\\cdot)$ . RKHSPose losses can then be denoted as:", + "bbox": [ + 72, + 211, + 920, + 326 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r} = \\operatorname {s m o o t h} _ {\\mathcal {L} _ {1}} \\left(V _ {r}, \\widehat {V} _ {r}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 353, + 336, + 917, + 357 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {k} = \\operatorname {s m o o t h} _ {\\mathcal {L} _ {1}} (K, \\widehat {K}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 351, + 362, + 917, + 383 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} = H (p (C), \\widehat {p} (C)) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 353, + 388, + 917, + 406 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} = \\operatorname {s m o o t h} _ {\\mathcal {L} _ {1}} (S, \\widehat {S}) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 353, + 411, + 917, + 432 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {A} = M M D \\left(\\widehat {f} _ {\\text {s y n / r e a l}}, \\widehat {f} _ {\\text {r e a l}}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 348, + 437, + 917, + 461 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {r} \\mathcal {L} _ {r} + \\lambda_ {k} \\mathcal {L} _ {k} + \\lambda_ {c} \\mathcal {L} _ {c} + \\lambda_ {s} \\mathcal {L} _ {s} + \\lambda_ {A} \\mathcal {L} _ {A} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 278, + 466, + 917, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\lambda_r, \\lambda_k, \\lambda_c, \\lambda_s$ , and $\\lambda_A$ are weights for adjustment during training, and all non-hatted quantities are GT values. At inference, $M_{rv}$ takes $I_{real}$ and $D_{real}$ as input, and outputs $K, C$ and $S$ . The keypoints $K$ are ranked and grouped by $S$ and $C$ , and are then forwarded into the ePnP [46] algorithm which estimates 6DoF pose values. ICP can then be optionally applied using the depth data to refine the estimated pose.", + "bbox": [ + 72, + 492, + 917, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Convolutional RKHS Adapter", + "text_level": 1, + "bbox": [ + 75, + 632, + 510, + 653 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reproducing Kernel Hilbert Space $\\mathcal{H}$ is a commonly used vector space for Domain Adaptation [58, 77]. Hilbert Space is a complete metric space (in which every Cauchy sequence of points has a limit within the metric) represented by the inner product of vectors. For a non-empty set of data $\\mathcal{X}$ , a function $\\mathcal{K}_{\\mathcal{X}}: \\mathcal{X} \\times \\mathcal{X} \\to \\mathcal{R}$ is a reproducing kernel if:", + "bbox": [ + 72, + 658, + 917, + 756 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} k (\\cdot , x) \\in \\mathcal {H} \\quad \\forall \\quad x \\in \\mathcal {X} \\\\ \\langle f (\\cdot), k (\\cdot , x) \\rangle = f (x) \\quad \\forall \\quad x \\in \\mathcal {X}, f \\in \\mathcal {H} \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 273, + 765, + 917, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\langle a,b\\rangle$ denotes the inner product of two vectors $a$ and $b$ , $k(\\cdot ,x) = \\mathcal{K}_{\\mathcal{X}}$ for each $x\\in \\mathcal{X}$ , and $f$ is a function in $\\mathcal{H}$ . The second equation in Eq. 7 is an expression of the reproducing property of $\\mathcal{H}$ .", + "bbox": [ + 72, + 822, + 917, + 879 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In order to utilize $\\mathcal{H}$ for DA, we expand the kernel definition to two sets of data $\\mathcal{X}$ and $\\mathcal{Y}$ . The reproducing kernel can then be defined as:", + "bbox": [ + 72, + 881, + 920, + 920 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {K} (\\mathcal {X}, \\mathcal {Y}) = \\left\\langle \\mathcal {K} _ {\\mathcal {X}}, \\mathcal {K} _ {\\mathcal {Y}} \\right\\rangle_ {H} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 930, + 917, + 949 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "RKHPose", + "bbox": [ + 726, + 14, + 837, + 30 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 903, + 16, + 917, + 30 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, $\\mathcal{K}_{\\mathcal{X}}$ and $\\mathcal{K}_{\\mathcal{Y}}$ are themselves inner product kernels that map $\\mathcal{X}$ and $\\mathcal{Y}$ respectively into their own Hilbert spaces, and $\\mathcal{K}(\\mathcal{X},\\mathcal{Y})$ is the joint Hilbert space kernel of $\\mathcal{X}$ and $\\mathcal{Y}$ . Note that $\\mathcal{K}(\\mathcal{X},\\mathcal{Y})$ is not the kernel commonly defined for CNNs. Rather, it is the similarity function defined in kernel methods, such as is used by Support Vector Machine (SVM) techniques to calculate similarity measurements. Some recent methods [8, 52, 53, 56] named it a Convolutional Kernel Network (CKN) to distinguish it from CNNs.", + "bbox": [ + 72, + 55, + 920, + 190 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Various kernels of CKN, such as the Gaussian Kernel [53], the RBF Kernel [52] and the Inner Product Kernel [56], have been shown to be comparable to shallow CNNs for various tasks, especially for Domain Adaption. The most intuitive inner product kernel is mathematically similar to a fully connected layer, where trainable weights are multiplied by the input feature map. To allow the application of RKHS methods into our CNN based Adapter network $M_A$ , trainable weights are added to $\\mathcal{K}(\\mathcal{X},\\mathcal{Y})$ [49]. The trainable Kernel $\\mathcal{K}_w(\\mathcal{X},\\mathcal{Y})$ can then be denoted as:", + "bbox": [ + 72, + 191, + 920, + 347 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {K} _ {w} (\\mathcal {X}, \\mathcal {Y}) = \\langle \\langle \\mathcal {X}, W _ {\\mathcal {X}} \\rangle , \\langle \\mathcal {Y}, W _ {\\mathcal {Y}} \\rangle \\rangle_ {H} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 317, + 360, + 917, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $W_{\\mathcal{X}}$ and $W_{\\mathcal{Y}}$ are trainable weights. By adding $W_{\\mathcal{X}}$ and $W_{\\mathcal{Y}}$ , $K_{w}(\\mathcal{X},\\mathcal{Y})$ still satisfies the RKHS constraints.", + "bbox": [ + 72, + 390, + 917, + 427 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The sim2real domain gap of feature maps $f$ being trained in $M_{rv}$ (with few real images and no real GT labels) are hard to measure using trivial distance metrics. In contrast, RKHS can be a more accurate and robust space for comparison, since it is known to be capable of handling high-dimensional data with a low number of samples [24, 56]. To compare $(f_{syn / real}, f_{real})$ in RKHS, a series of CNN layers encodes $(f_{syn / real}, f_{real})$ into higher-dimensional features $(h_{syn / real}, h_{real})$ , followed by the trainable $\\mathcal{K}_w(\\mathcal{X}, \\mathcal{Y})$ . Once mapped into RKHS, $h_{syn / real} = \\{sr_i\\}_{i=1}^m$ and $h_{real} = \\{r_i\\}_{i=1}^m$ can then be measured by Maximum Mean Discrepancy (MMD), a common DA measurement [27, 49], which is the square distance between the kernel embedding [26]:", + "bbox": [ + 72, + 429, + 920, + 624 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} M M D \\left(h _ {s y n / r e a l}, h _ {r e a l}\\right) = \\frac {1}{m} \\left[ \\left(\\sum_ {i = 1} ^ {m} \\sum_ {j = 1} ^ {m} k _ {w} \\left(s r _ {i}, s r _ {j}\\right) - \\sum_ {i = 1} ^ {m} k _ {w} \\left(s r _ {i}, s r _ {i}\\right)\\right) \\right. \\\\ \\left. - \\left(\\sum_ {i = 1} ^ {m} \\sum_ {j = 1} ^ {m} k _ {w} \\left(s r _ {i}, r _ {j}\\right) - \\sum_ {i = 1} ^ {m} k _ {w} \\left(s r _ {i}, r _ {i}\\right)\\right) + \\left(\\sum_ {i = 1} ^ {m} \\sum_ {j = 1} ^ {m} k _ {w} \\left(r _ {i}, r _ {j}\\right) - \\sum_ {i = 1} ^ {m} k _ {w} \\left(r _ {i}, r _ {i}\\right)\\right) \\right] ^ {\\frac {1}{2}} \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 82, + 635, + 929, + 780 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $k_{w}()$ is the feature element of $\\mathcal{K}_w($", + "bbox": [ + 75, + 793, + 530, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In summary, the Adapter $M_A$ shown in Fig. 2 measures MMD for each $(h_{syn / real}, h_{real})$ in RKHS, by increasing the $f_{syn / real}$ and $f_{real}$ dimension using a CNN and thereby constructing a learnable kernel $\\mathcal{K}_w$ . The outputs are the feature maps $h_{syn / real}$ and $h_{real}$ , which are supervised by loss $\\mathcal{L}_A$ during training of the real data epochs. Based on the experiments in Sec. 5.3, our trainable inner product kernel is shown to be more accurate for our task than other known kernels [52, 53, 56] that we tested, that are often used in such kernel methods.", + "bbox": [ + 72, + 811, + 920, + 949 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 77, + 16, + 94, + 29 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y.Wu and M. Greenspan", + "bbox": [ + 157, + 14, + 406, + 32 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Keypoint Radial Voting Network", + "text_level": 1, + "bbox": [ + 77, + 53, + 547, + 74 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The network $M_v$ votes for keypoints using a CNN architecture, taking the radial voting quantity $V_r$ resulting from $M_r$ as input. VoteNet [63] previously used a CNN approach to vote for object centers, whereas other keypoint-based techniques have implemented GPU-based parallel RANSAC [32, 61] methods for offset and vector quantities. The radial quantity is known to be more accurate than the vector or offset quantities, and has been previously implemented with a CPU-based parallel accumulator space method [82, 83]. Given its superior accuracy, $M_v$ implements radial voting using a CNN to improve efficiency. Given a 2D radial map $\\widehat{V}_r$ estimated by $M_r$ and supervised by GT radial maps $V_r$ , the task is to accumulate votes, find the peak, and estimate the keypoint location. The $V_r$ foreground pixels (which lie on the target object) store the Euclidean distance from these pixels to each of the keypoints, with background (non-object) pixels set to value -1.", + "bbox": [ + 77, + 87, + 917, + 338 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The estimated radial map $\\widehat{V}_r$ is indeed an inverse heat map of the candidate keypoints' locations, distributed in a radial pattern centered at the keypoints. To forward $\\widehat{V}_r$ into a CNN voting module, it is inversely normalized so that it becomes a heat map. Let $v_r^{max}$ and $v_r^{min}$ be the maximum and minimum global radial distances for all objects in a dataset, which can be calculated by iterating through all GT radial maps, or alternately generated from the object CAD models. An inverse radial map $\\widehat{V}_r^{-1}$ can then be denoted as $\\widehat{V}_r^{-1} = (v_r^{max} - \\widehat{V}_r) / (v_r^{max} - v_r^{min})$ . Voting network $M_v$ takes $\\widehat{V}_r^{-1}$ as the input and generates the accumulated vote map by a series of convolution, ReLu, and batch normalization layers. The complete network architecture is provided in the Supplementary material Sec. S.2. The background pixels are filtered out by a ReLu layer, and only foreground pixels contribute to voting. The accumulated vote map is then max-pooled for peak extraction, and reshaped using a fully connected layer into a $n\\times 4$ output. The output represents $n$ keypoints and comprises $n\\times 2$ projected 2D keypoints $K$ , $n$ classification labels $C$ , and $n$ confidence scores $S$ . The labels $C$ indicate which object the corresponding keypoints belongs to, and $S$ ranks the confidence level of keypoints before being forwarded into ePnP [46] for pose estimation. $M_v$ is supervised by $\\mathcal{L}_k$ , $\\mathcal{L}_c$ , and $\\mathcal{L}_s$ (Eqs. 2-4) and is trained end-to-end along with $M_r$ .", + "bbox": [ + 77, + 339, + 917, + 710 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 77, + 739, + 309, + 762 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 77, + 780, + 542, + 798 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "RKHSPose uses BOP Procedural Blender [17] (PBR) synthetic images [16, 36, 37, 72] for the synthetic training phase. The images are generated by dropping synthetic objects (CAD models) onto a plane in a simulated environment using PyBullet [13], and then rendering them with synthetic textures. All objects in the synthetic images are thus automatically labeled with precise GT poses. We evaluated RKHSPose for the six BOP [37, 72] core datasets (LMO [34], YCB [84], TLESS [35], TUDL [36], ITODO [20], and HB [40]), all except IC-BIN [19],", + "bbox": [ + 77, + 811, + 917, + 949 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "RKHPose", + "bbox": [ + 726, + 14, + 837, + 30 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 900, + 16, + 917, + 30 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "which does not include any real training or validation images and is therefore not applicable. ITODO and HB have no real images in the training set, and so for training we instead used the real images in their validation sets, which were disjoint from their test sets.", + "bbox": [ + 77, + 53, + 917, + 131 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Our main results are evaluated with the ADD(S) [34] metric for the LM and LMO dataset, and the ADD(S) AUC [84] metric for the YCB dataset. These are the standard metrics commonly used to compare self-supervised 6DoF PE methods. ADD(S) is based on the mean distance (minimum distance for symmetry) between the object surfaces for GT and estimated poses, whereas ADD(S) AUC plots a curve formed by ADD(S) for various object diameter thresholds. We use the BOP average recall ( $AR$ ) metrics for our ablation studies. The $AR$ metric, based on the original ADD(S) [34], evaluates three aspects, including Visible Surface Discrepancy ( $AR_{VSD}$ ), Maximum Symmetry-Aware Surface/Projection Distance ( $AR_{MSSD}$ and $AR_{MSPD}$ ) [36].", + "bbox": [ + 77, + 134, + 917, + 330 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 79, + 362, + 430, + 382 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "RKHSPose is trained on a server with an Intel Xeon 5218 CPU and two RTX6000 GPUs with a batch size of 32. The Adam optimizer is used for the training of $M_{rv}$ , on both synthetic and real data, and $M_A$ is optimized by SGD. Both of the optimizers have an initial learning rate of $lr = 1e - 3$ and weight decay 1e-4 for 80 and 20 epochs respectively.", + "bbox": [ + 79, + 400, + 917, + 497 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The input of the network is normalized before training, as follows. The RGB images $I$ are normalized and standardized using ImageNet [14] specifications. The depth maps are each individually normalized by their local minima and maxima, to lie within a range of 0 to 1. The radial distances in radial map $V_{r}$ and 2D projected keypoints are both normalized by the width and height of $I$ .", + "bbox": [ + 77, + 499, + 917, + 596 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A single set of four keypoints is chosen by KeyGNet [81] for the set of all objects in each dataset. One extra background class is added to $C$ in order to filter out the redundant background points in $K$ . $M_{rv}$ is first trained for 120 epochs on synthetic data, during which $M_A$ remains frozen. Following this, training proceeds for an additional 80 epochs which alternate between real and synthetic data. When training on real data, both $M_A$ and $M_{rv}$ weights are learned, whereas $M_A$ is frozen for the alternating synthetic data training.", + "bbox": [ + 77, + 596, + 917, + 733 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "During real data epochs, $M_{rv}$ initially estimates pseudo-keypoints for each real image. These pseudo-keypoints are then forwarded into ePnP for pseudopose estimation. Each estimated pseudo-pose is then augmented into a set of poses $P_{aug}$ by applying arbitrary rotational and translational perturbations with respective ranges of $[-\\frac{\\pi}{18}, \\frac{\\pi}{18}]$ radians and $[-0.1, 0.1]$ along three axes within the normalized model frame, which is defined using the largest object in the dataset. The set of syn/real images are rendered by overlaying onto the real image the CAD model of each object using each augmented pose value in $P_{aug}$ . The cardinality of $P_{aug}$ is set to be one less than the batch size, and the adapter $M_A$ is trained on a mini-batch of the hybrid images resulting from $P_{aug}$ , plus the image resulting from the original estimated pseudo-pose.", + "bbox": [ + 79, + 734, + 917, + 947 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y.Wu and M. Greenspan", + "bbox": [ + 157, + 14, + 406, + 32 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/fa28ed9b27e25c67b517edceaa2e90a1ad1899ac0d3d9154aeff3176ac2af7cf.jpg", + "table_caption": [ + "Table 2: Comparison with other methods. Accuracy of RKHSPose for LM and LMO is evaluated with ADD(S), and for YCB is evaluated with ADD(S) AUC. All 'Supervision: Syn + Self' methods use real images without real labels." + ], + "table_footnote": [], + "table_body": "
MethodReal data image labelDataset/Metric
LM ADD(S)LMOYCB ADD(S)ADD-S AUC
Supervision: Syn (lower bound)
AAEXX31.4---
MHPXX38.8---
GDR (TexPose version)XX77.452.9--
Self6D++XX77.452.977.889.4
Self6D++ with DrefXX88.062.579.290.1
OursXX78.254.376.590.2
Ours+ICPXX87.955.778.391.3
Supervision: Syn + Self
Sock et al.✓X60.622.8--
DSC✓X58.624.8--
Self6D✓X58.932.1--
SMOC-Net✓X91.363.3--
Self6D++✓X88.564.780.091.4
TexPose✓X91.766.7--
Ours✓X95.868.682.892.4
Ours+ICP✓X95.968.783.092.6
Supervision: Syn + Real GT (upper bound)
SO-Pose✓✓96.062.383.990.9
Self6D++✓✓91.074.482.690.7
Ours✓✓96.770.885.492.2
Ours+ICP✓✓96.871.385.692.4
", + "bbox": [ + 181, + 105, + 818, + 583 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Initially $M_A$ is frozen, and for the first 80 epochs, the loss is set to emphasize the classification and the visibility score regression, i.e. $\\lambda_c = \\lambda_s = 0.6$ and $\\lambda_r = \\lambda_k = 0.4$ . Following this, up to epoch 200, the scales of losses are then exchanged to fine-tune the localization of the keypoints, i.e. $\\lambda_c = \\lambda_s = 0.4$ and $\\lambda_r = \\lambda_k = 0.6$ . After epoch 120, $M_A$ is unfrozen each alternating epoch, and $\\lambda_D$ is set to 1 during the remaining $M_A$ training epochs. This training strategy, shown in Fig. 2, minimizes the sim2real gap without any real image GT labels, and using very few (320) real images.", + "bbox": [ + 72, + 635, + 920, + 793 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Results", + "text_level": 1, + "bbox": [ + 75, + 819, + 237, + 837 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The results are summarized in Tables 2 and 3. To our knowledge, RKHSPose outperforms all existing self-supervised 6DoF PE methods. In Table 2, the upper bound fully supervised, lower bound synthetically supervised, and middle self-supervised methods (including ours) are compared. On LM and LMO, our ADD(S) is $+4.2\\%$ and $+2\\%$ better than the second best method TexPose [9]. We", + "bbox": [ + 72, + 850, + 922, + 949 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "RKHPose", + "bbox": [ + 726, + 14, + 840, + 30 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 893, + 16, + 915, + 30 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/190cbb24832e94571ceba9401c6ee62fe09632621598dbcfc72cca519347751d.jpg", + "table_caption": [ + "Table 3: Comparison with fully supervised methods. RKHSPose results on TLESS $(-1.8)$ , TUDL $(-0.4)$ , ITODD $(-4.6)$ and HB $(+0.1)$ compares to SOTA methods with full supervision of real GT labels. Methods annotated with * use the detection results from other detection methods." + ], + "table_footnote": [], + "table_body": "
Methodreal labelLMOTLESSDataset
TUDLITODDHBYCB
SurfEmb* [30]✓-76.082.885.465.986.679.9
RCVPose3D [82]✓-72.970.896.673.386.384.3
RADet [87]+PFA* [39]✓-79.785.096.067.686.988.8
ZebraPose [70]✓-78.086.295.665.492.189.9
Ours✗95.768.285.596.268.692.283.6
Ours+ICP✗95.868.485.696.268.792.383.8
", + "bbox": [ + 130, + 141, + 866, + 297 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/12f1a860821259363b9ce9e8fa2e3d636edb840ef5654196f8ff86ed9ae4784b.jpg", + "image_caption": [ + "Fig. 3: Qualitative overlay results on selected images. Red dots and blue dots are projected surface points from GT poses and estimated poses, respectively." + ], + "image_footnote": [], + "bbox": [ + 77, + 330, + 217, + 426 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c8e452a90ee70b8162806bcab95a05890b7ae4d8358c542d23f3e244ab86431b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 330, + 355, + 424 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/243dda54444d5ca87002df56e5e61697d016711f6e09f3e84d9a5c175c79b206.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 360, + 331, + 496, + 424 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/222bb3aa03cc0d98fdda13e3782932a1008bbceab48df717e16cfb32bb20e597.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 331, + 636, + 424 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6c80e3d267335e01caa2245aa89098bd693bb1be49d82798c7582dae71fc694e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 641, + 330, + 777, + 424 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/9256f29dad8d7bd713e1dd1603d26852183ecbdc77afac167ec1918370b65e59.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 330, + 917, + 424 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "compared our performance to Self6D++ [79], which was the only other method that evaluated using YCB, and saw a $3\\%$ improvement after ICP on ADD(S) AUC. Last but not least, in Table 3 our performance evaluated on the other four BOP core datasets is comparable to several upper bound methods from the BOP leaderboard. Some test scenes with RKHSPose results are shown in Fig. 3.", + "bbox": [ + 72, + 530, + 917, + 627 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "RKHSPose runs at 34 fps on an Intel i7 2.5GHz CPU and an RTX 3090 GPU with 24G VRAM. It takes on average $8.7\\mathrm{ms}$ for loading data, $4.5\\mathrm{ms}$ for forward inference through $M_{rv}$ ( $\\times 10$ faster compared to the analytical radial voting in RCVPose [83]), and $16.2\\mathrm{ms}$ for ePnP.", + "bbox": [ + 72, + 627, + 917, + 705 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5 Ablation Studies", + "text_level": 1, + "bbox": [ + 75, + 736, + 370, + 757 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.1 Dense Vs. Sparse Adapter", + "text_level": 1, + "bbox": [ + 75, + 778, + 464, + 800 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The Adapter $M_A$ densely matches the intermediate feature maps, whereas the majority of other methods [15,75,80] only compare the final output. To show the benefits of dense comparison, we conduct an experiment with different variations of $M_A$ . A sparse matching $M_A^s$ network is trained on synthetic and real data comparing only a single feature map, which is the intermediate radial map. $M_A^s$ has the exact same overall learning capacity (number of parameters) as the dense matching $M_A$ described in Sec. 3.3. The results in Table 4 show that $M_A$", + "bbox": [ + 72, + 811, + 920, + 949 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y.Wu and M. Greenspan", + "bbox": [ + 157, + 14, + 406, + 32 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/d57edd734d3cfa5a4ddbe47c39c2f2bde889c92dcee02a1fab6bbb07c8cf2ec7.jpg", + "table_caption": [ + "Table 4: ${AR}$ of different adapters on LM and five BOP core datasets." + ], + "table_footnote": [], + "table_body": "
AdapterARVSDARMSSDARMSPDAR
MA78.177.877.877.9
MA84.984.184.384.4
", + "bbox": [ + 77, + 105, + 476, + 170 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/4feef9e91556a63b42dd09d1b8e61331cf7fc64fac546cc6319ca7b65e6394b9.jpg", + "table_caption": [ + "Table 5: ${AR}$ of different training strategies on LM and five BOP core datasets." + ], + "table_footnote": [], + "table_body": "
Training TypeARVSDARMSSDARMSPDAR
Mixed84.984.184.384.4
Sequential82.381.781.781.9
", + "bbox": [ + 520, + 105, + 920, + 162 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/c2b16c73cd552ab1dc078e1c46882079172f0ff6f42b96029ff4648b7b326565.jpg", + "table_caption": [ + "Table 6: [R Tab A] $AR$ of different kernel- Table 7: $AR$ of different metrics on LM nets on LM and five BOP core datasets. and five BOP core datasets." + ], + "table_footnote": [], + "table_body": "
KernelwARVSDARMSSDARMSPDAR
LinearX71.670.870.671.0
✓84.984.184.384.4
RBFX73.472.973.273.2
✓82.581.381.581.6
", + "bbox": [ + 77, + 239, + 486, + 336 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/08167831f2a1dae82367e2665a0220538f06fbf6e04a300f420d95688fc0a174.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MetricARVSDARMSSDARMSPDAR
MMD84.984.184.384.4
KL Div78.077.878.077.9
Wass80.980.680.980.8
", + "bbox": [ + 486, + 239, + 886, + 321 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "surpassed $M_A^s$ on all six datasets tested. Specifically, on ITODD, $M_A$ is $12.1\\%$ more accurate than $M_A^s$ . This experiment shows the effectiveness of our densely matched $M_A$ .", + "bbox": [ + 72, + 373, + 920, + 434 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.2 Syn/Real Synchronized Training", + "text_level": 1, + "bbox": [ + 75, + 463, + 542, + 484 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "When training RKHSPose, real epochs are alternated with synthetic epochs. In contrast, some other methods [9, 75, 79, 80] separate the synthetic/real training. We conducted an experiment to compare these two different training strategies, the results of which are shown in Table 5. The alternating training performs slightly better ( $+2.5\\%$ on average) than the sequential training, possibly due to the early access to real scenes thereby avoiding local minima.", + "bbox": [ + 72, + 499, + 917, + 617 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Adapter Kernels and Metrics", + "text_level": 1, + "bbox": [ + 75, + 648, + 501, + 668 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We use a linear (dot product) kernel and MMD in RKHS for domain gap measurements. There are various other kernels and similarity measurements that can be implemented in RKHS as described in Sec. 3.2. First, we add trainable weights to the radial basis function (RBF) kernel in a similar manner as $K_{w}$ defined in Eq. 9. The trainable RBF kernel on two sets of data $X$ and $Y$ is denoted as:", + "bbox": [ + 72, + 682, + 917, + 798 + ], + "page_idx": 12 + }, + { + "type": "equation", + "text": "\n$$\nK _ {r b f} (X, Y) = \\exp (- w \\| X - Y \\| ^ {2}) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 317, + 798, + 917, + 822 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "where $w$ are the trainable weights, which replaces the original adjustable parameter in the classical RBF kernel. We also experiment with the classical RKHS kernel functions without trainable weights, including the inner product kernel and the original RBF kernel [78], for comparison. Further, we experiment on other commonly used distance measures, including Kullback-Leibler Divergence (KL Div, i.e. relative-entropy) and Wasserstein (Wass) Distance.", + "bbox": [ + 72, + 830, + 917, + 949 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "RKHPose", + "bbox": [ + 726, + 14, + 837, + 30 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 893, + 16, + 917, + 30 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/85734d7328823939c2c8a4cace4cb00d1ae19d2781561399992531721a03abc4.jpg", + "image_caption": [ + "Fig. 4: Impact of # of real images with/without GT labels used during training. All datasets are evaluated by the BOP AR metric. We conduct experiments from 0 to 640 real images on all datasets, except ITODD which contained only 357 real images." + ], + "image_footnote": [], + "bbox": [ + 200, + 50, + 794, + 261 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In Table 6, the RBF kernel performs similar to the linear product kernel with a slight performance dip. In Table 7, MMD minimizes the domain gap better than the Wass Distance, followed by the KL Div metric, leading to a better overall performance on $AR$ . Based on these results, we used the linear product kernel with trainable weights and selected MMD as the main loss metric.", + "bbox": [ + 72, + 378, + 922, + 479 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5.4 Number of Real Images and Real Labels", + "text_level": 1, + "bbox": [ + 75, + 502, + 641, + 523 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The objective of RKHSPose is to reduce real data usage and train without any real GT labels. To show the effectiveness of the approach, we conducted an experiment by training on different numbers of real images, the results of which are shown in Fig. 4. We used up to 640 real images in all cases, except for that of ITODD which contains only 357 real images. The $AR$ of all datasets saturates at 160 images except YCB. The further improvement of YCB beyond 160 images is also only $+0.1\\%$ and saturates after 320 real images. We nevertheless use 320 real unlabeled images for our main results. This experiment showed that adding more than 320 real labeled images did not significantly improve performance.", + "bbox": [ + 72, + 531, + 924, + 708 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 75, + 734, + 290, + 757 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "To sum up, we propose a novel self-supervised keypoint radial voting-based 6DoF PE method using RGB-D data called RKHSPose. RKHSPose fine-tunes poses pre-trained on synthetic data by densely matching features with a learnable kernel in RKHS, using real data albeit without any real GT poses. By applying this DA technique in feature space, RKHSPose achieved SOTA performance on the six applicable BOP core datasets, surpassing the performance of all other self-supervised methods. Notably, the RKHSPose performance closely approaches that of several fully-supervised methods, which indicates the strength of the approach at reducing the sim2real domain gap for this problem.", + "bbox": [ + 72, + 773, + 924, + 949 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y.Wu and M. Greenspan", + "bbox": [ + 157, + 14, + 406, + 32 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements: Thanks to Bluewrist Inc. and NSERC for their support of this work.", + "bbox": [ + 75, + 55, + 920, + 95 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 77, + 125, + 237, + 144 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Al Safadi, E., Song, X.: Learning-based image registration with meta-regularization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10928-10937 (2021)", + "2. Aronszajn, N.: Theory of reproducing kernels. Transactions of the American mathematical society 68(3), 337-404 (1950)", + "3. Besl, P.J., McKay, N.D.: Method for registration of 3-d shapes. In: Sensor fusion IV: control paradigms and data structures. vol. 1611, pp. 586-606. International Society for Optics and Photonics (1992)", + "4. Bietti, A., Mairal, J.: Group invariance, stability to deformations, and complexity of deep convolutional representations. The Journal of Machine Learning Research 20(1), 876-924 (2019)", + "5. Bietti, A., Mialon, G., Chen, D., Mairal, J.: A kernel perspective for regularizing deep neural networks. In: International Conference on Machine Learning. pp. 664-674. PMLR (2019)", + "6. Bozorgtabar, B., Mahapatra, D., Thiran, J.P.: Exprada: Adversarial domain adaptation for facial expression analysis. Pattern Recognition 100, 107111 (2020)", + "7. Chen, C., Fu, Z., Chen, Z., Jin, S., Cheng, Z., Jin, X., Hua, X.S.: Homm: Higher-order moment matching for unsupervised domain adaptation. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 3422-3429 (2020)", + "8. Chen, D., Jacob, L., Mairal, J.: Convolutional kernel networks for graph-structured data. In: International Conference on Machine Learning. pp. 1576–1586. PMLR (2020)", + "9. Chen, H., Manhardt, F., Navab, N., Busam, B.: Texpose: Neural texture learning for self-supervised 6d object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4841-4852 (2023)", + "10. Cheng, X., Xie, Y.: Neural tangent kernel maximum mean discrepancy. Advances in Neural Information Processing Systems 34, 6658-6670 (2021)", + "11. Corcoran, P.: An end-to-end graph convolutional kernel support vector machine. Applied Network Science 5(1), 1-15 (2020)", + "12. Cortes, C., Vapnik, V.: Support-vector networks. Machine learning 20, 273-297 (1995)", + "13. Coumans, E., Bai, Y.: Pybullet, a python module for physics simulation for games, robotics and machine learning. http://pybullet.org (2016-2021)", + "14. Deng, J., Dong, W., Socher, R., Li, L., Kai Li, Li Fei-Fei: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. pp. 248-255 (2009). https://doi.org/10.1109/CVPR.2009.5206848", + "15. Deng, X., Xiang, Y., Mousavian, A., Eppner, C., Bretl, T., Fox, D.: Self-supervised 6d object pose estimation for robot manipulation. In: 2020 IEEE International Conference on Robotics and Automation (ICRA). pp. 3665-3671. IEEE (2020)", + "16. Denninger, M., Sundermeyer, M., Winkelbauer, D., Olefir, D., Hodan, T., Zidan, Y., Elbadrawy, M., Knauer, M., Katam, H., Lodhi, A.: Blenderproc: Reducing the reality gap with photorealistic rendering. In: International Conference on Robotics: Science and Systems, RSS 2020 (2020)" + ], + "bbox": [ + 79, + 164, + 920, + 947 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "RKHPose", + "bbox": [ + 726, + 16, + 837, + 30 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 893, + 16, + 917, + 30 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Denninger, M., Winkelbauer, D., Sundermeyer, M., Boerdijk, W., Knauer, M., Strobl, K.H., Humt, M., Triebel, R.: Blenderproc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software 8(82), 4901 (2023). https://doi.org/10.21105/joss.04901, https://doi.org/10.21105/joss.04901", + "18. Di, Y., Manhardt, F., Wang, G., Ji, X., Navab, N., Tombari, F.: So-posed: Exploiting self-occlusion for direct 6d pose estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 12396-12405 (October 2021)", + "19. Doumanoglou, A., Kouskouridas, R., Malassiotis, S., Kim, T.K.: Recovering 6d object pose and predicting next-best-view in the crowd. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3583-3592 (2016)", + "20. Drost, B., Ulrich, M., Bergmann, P., Hartinger, P., Steger, C.: Introducing mvtec itodd-a dataset for 3d object recognition in industry. In: Proceedings of the IEEE international conference on computer vision workshops. pp. 2200-2208 (2017)", + "21. Gadwe, A., Ren, H.: Real-time 6 dof pose estimation of endoscopic instruments using printable markers. IEEE Sensors Journal 19(6), 2338-2346 (2018)", + "22. Gao, X., Hou, X., Tang, J., Cheng, H.: Complete solution classification for the perspective-three-point problem. IEEE Trans. Pattern Anal. Mach. Intell. 25, 930-943 (2003)", + "23. Ghojogh, B., Ghodsi, A., Karray, F., Crowley, M.: Reproducing kernel hilbert space, mercer's theorem, eigenfunctions, nystr\\'' om method, and use of kernels in machine learning: Tutorial and survey. arXiv preprint arXiv:2106.08443 (2021)", + "24. Ghorbani, B., Mei, S., Misiakiewicz, T., Montanari, A.: When do neural networks outperform kernel methods? Advances in Neural Information Processing Systems 33, 14820-14830 (2020)", + "25. Greene, N., Luo, W., Kazanzides, P.: dvpose: Automated data collection and dataset for 6d pose estimation of robotic surgical instruments. In: 2023 International Symposium on Medical Robotics (ISMR). pp. 1-7. IEEE (2023)", + "26. Gretton, A., Borgwardt, K., Rasch, M., Scholkopf, B., Smola, A.: A kernel method for the two-sample-problem. Advances in neural information processing systems 19 (2006)", + "27. Gretton, A., Borgwardt, K.M., Rasch, M.J., Scholkopf, B., Smola, A.: A kernel two-sample test. The Journal of Machine Learning Research 13(1), 723-773 (2012)", + "28. Guo, S., Hu, Y., Alvarez, J.M., Salzmann, M.: Knowledge distillation for 6d pose estimation by aligning distributions of local predictions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18633-18642 (2023)", + "29. Hai, Y., Song, R., Li, J., Salzmann, M., Hu, Y.: Rigidity-aware detection for 6d object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8927-8936 (2023)", + "30. Haugaard, R.L., Buch, A.G.: Surfemb: Dense and continuous correspondence distributions for object pose estimation with learnt surface embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6749-6758 (2022)", + "31. He, Y., Huang, H., Fan, H., Chen, Q., Sun, J.: Ffb6d: A full flow bidirectional fusion network for 6d pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3003-3013 (2021)", + "32. He, Y., Sun, W., Huang, H., Liu, J., Fan, H., Sun, J.: Pvn3d: A deep pointwise 3d keypoints voting network for 6 dof pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + ], + "bbox": [ + 79, + 56, + 917, + 947 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y.Wu and M. Greenspan", + "bbox": [ + 157, + 14, + 406, + 32 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. He, Y., Wang, Y., Fan, H., Sun, J., Chen, Q.: Fs6d: Few-shot 6d pose estimation of novel objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6814-6824 (2022)", + "34. Hinterstoisser, S., Lepetit, V., Ilic, S., Holzer, S., Bradski, G., Konolige, K., Navab, N.: Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In: Asian conference on computer vision. pp. 548-562. Springer (2012)", + "35. Hodañ, T., Haluza, P., Obdržálek, Š., Matas, J., Lourakis, M., Zabulis, X.: T-LESS: An RGB-D dataset for 6D pose estimation of texture-less objects. IEEE Winter Conference on Applications of Computer Vision (WACV) (2017)", + "36. Hodan, T., Michel, F., Brachmann, E., Kehl, W., GlentBuch, A., Kraft, D., Drost, B., Vidal, J., Ihrke, S., Zabulis, X., et al.: Bop: Benchmark for 6d object pose estimation. In: Proceedings of the European conference on computer vision (ECCV). pp. 19-34 (2018)", + "37. Hodač, T., Sundermeyer, M., Drost, B., Labbé, Y., Brachmann, E., Michel, F., Rother, C., Matas, J.: Bop challenge 2020 on 6d object localization. In: European Conference on Computer Vision. pp. 577-594. Springer (2020)", + "38. Horn, B.K., Hilden, H.M., Negahdaripour, S.: Closed-form solution of absolute orientation using orthonormal matrices. JOSA A 5(7), 1127-1135 (1988)", + "39. Hu, Y., Fua, P., Salzmann, M.: Perspective flow aggregation for data-limited 6d object pose estimation. In: European Conference on Computer Vision. pp. 89-106. Springer (2022)", + "40. Kaskman, R., Zakharov, S., Shugurov, I., Ilic, S.: Homebreweddb: RGB-d dataset for 6d pose estimation of 3d objects. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. pp. 0-0 (2019)", + "41. Khosravi, M., Smith, R.S.: The existence and uniqueness of solutions for kernel-based system identification. Automatica 148, 110728 (2023)", + "42. Kleeberger, K., Huber, M.F.: Single shot 6d object pose estimation. In: 2020 IEEE International Conference on Robotics and Automation (ICRA). pp. 6239-6245. IEEE (2020)", + "43. Kleeberger, K., Landgraf, C., Huber, M.F.: Large-scale 6d object pose estimation dataset for industrial bin-picking. In: 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 2573-2578. IEEE (2019)", + "44. Kleeberger, K., Völk, M., Bormann, R., Huber, M.F.: Investigations on output parameterizations of neural networks for single shot 6d object pose estimation. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 13916-13922. IEEE (2021)", + "45. Lee, T., Lee, B.U., Shin, I., Choe, J., Shin, U., Kweon, I.S., Yoon, K.J.: Udacope: unsupervised domain adaptation for category-level object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14891-14900 (2022)", + "46. Lepetit, V., Moreno-Noguer, F., Fua, P.: Ep n p: An accurate o (n) solution to the p n p problem. International journal of computer vision 81, 155-166 (2009)", + "47. Li, Z., Hu, Y., Salzmann, M., Ji, X.: Sd-pose: Semantic decomposition for cross-domain 6d object pose estimation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 2020–2028 (2021)", + "48. Lin, J., Wei, Z., Ding, C., Jia, K.: Category-level 6d object pose and size estimation using self-supervised deep prior deformation networks. In: European Conference on Computer Vision. pp. 19-34. Springer (2022)" + ], + "bbox": [ + 75, + 56, + 917, + 947 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "RKHPose", + "bbox": [ + 726, + 14, + 840, + 30 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 893, + 16, + 917, + 30 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "49. Liu, F., Xu, W., Lu, J., Zhang, G., Gretton, A., Sutherland, D.J.: Learning deep kernels for non-parametric two-sample tests. In: International conference on machine learning. pp. 6316-6326. PMLR (2020)", + "50. Luo, Y.W., Ren, C.X.: Conditional bures metric for domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13989-13998 (2021)", + "51. Ma, X., Wang, Z., Li, H., Zhang, P., Ouyang, W., Fan, X.: Accurate monocular 3d object detection via color-embedded 3d reconstruction for autonomous driving. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6851-6860 (2019)", + "52. Mairal, J.: End-to-end kernel learning with supervised convolutional kernel networks. Advances in neural information processing systems 29 (2016)", + "53. Mairal, J., Koniusz, P., Harchaoui, Z., Schmid, C.: Convolutional kernel networks. Advances in neural information processing systems 27 (2014)", + "54. Manhardt, F., Arroyo, D.M., Rupprecht, C., Busam, B., Birdal, T., Navab, N., Tombari, F.: Explaining the ambiguity of object detection and 6d pose from visual data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019)", + "55. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)", + "56. Misiakiewicz, T., Mei, S.: Learning with convolution and pooling operations in kernel methods. Advances in Neural Information Processing Systems 35, 29014-29025 (2022)", + "57. Oberweger, M., Rad, M., Lepetit, V.: Making deep heatmaps robust to partial occlusions for 3d object pose estimation. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 119-134 (2018)", + "58. Pan, S.J., Tsang, I.W., Kwok, J.T., Yang, Q.: Domain adaptation via transfer component analysis. IEEE transactions on neural networks 22(2), 199-210 (2010)", + "59. Park, K., Mousavian, A., Xiang, Y., Fox, D.: Latentfusion: End-to-end differentiable reconstruction and rendering for unseen object pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10710-10719 (2020)", + "60. Pearson, K.: Liii. on lines and planes of closest fit to systems of points in space. The London, Edinburgh, and Dublin philosophical magazine and journal of science 2(11), 559-572 (1901)", + "61. Peng, S., Liu, Y., Huang, Q., Zhou, X., Bao, H.: Pynet: Pixel-wise voting network for 6dof pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4561-4570 (2019)", + "62. Pinheiro, P.O., Rostamzadeh, N., Ahn, S.: Domain-adaptive single-view 3d reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7638-7647 (2019)", + "63. Qi, C.R., Litany, O., He, K., Guibas, L.J.: Deep hough voting for 3d object detection in point clouds. In: proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9277-9286 (2019)", + "64. Sejdinovic, D., Sriperumbudur, B., Gretton, A., Fukumizu, K.: Equivalence of distance-based and rkhs-based statistics in hypothesis testing. The annals of statistics pp. 2263-2291 (2013)", + "65. Shan, P., Bi, Y., Li, Z., Wang, Q., He, Z., Zhao, Y., Peng, S.: Unsupervised model adaptation for multivariate calibration by domain adaptation-regularization based" + ], + "bbox": [ + 77, + 56, + 917, + 947 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y.Wu and M. Greenspan", + "bbox": [ + 157, + 14, + 406, + 32 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "kernel partial least square. Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 292, 122418 (2023)", + "66. Shugurov, I., Li, F., Busam, B., Ilic, S.: Osop: A multi-stage one shot object pose estimation framework. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6835-6844 (2022)", + "67. Shugurov, I., Zakharov, S., Ilic, S.: Dpodv2: Dense correspondence-based 6 dof pose estimation. IEEE transactions on pattern analysis and machine intelligence 44(11), 7417-7435 (2021)", + "68. Sock, J., Garcia-Hernando, G., Armagan, A., Kim, T.K.: Introducing pose consistency and warp-alignment for self-supervised 6d object pose estimation in color images. In: 2020 International Conference on 3D Vision (3DV). pp. 291-300. IEEE (2020)", + "69. Su, H., Qi, C.R., Li, Y., Guibas, L.J.: Render for cnn: Viewpoint estimation in images using cnns trained with rendered 3d model views. In: Proceedings of the IEEE international conference on computer vision. pp. 2686-2694 (2015)", + "70. Su, Y., Saleh, M., Fetzer, T., Rambach, J., Navab, N., Busam, B., Stricker, D., Tombari, F.: Zebrapose: Coarse to fine surface encoding for 6dof object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6738-6748 (2022)", + "71. Sundermeyer, M., Durner, M., Huang, E.Y., Marton, Z.C., Vaskevicius, N., Arras, K.O., Triebel, R.: Multi-path learning for object pose estimation across domains. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13916-13925 (2020)", + "72. Sundermeyer, M., Hodač, T., Labbe, Y., Wang, G., Brachmann, E., Drost, B., Rother, C., Matas, J.: Bop challenge 2022 on detection, segmentation and pose estimation of specific rigid objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2784-2793 (2023)", + "73. Sundermeyer, M., Marton, Z.C., Durner, M., Brucker, M., Triebel, R.: Implicit 3d orientation learning for 6d object detection from rgb images. In: Proceedings of the european conference on computer vision (ECCV). pp. 699-715 (2018)", + "74. Szafraniec, F.H.: The reproducing kernel hilbert space and its multiplication operators. Complex Analysis and Related Topics pp. 253-263 (2000)", + "75. Tan, T., Dong, Q.: Smoc-net: Leveraging camera pose for self-supervised monocular object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21307-21316 (2023)", + "76. Ullman, S.: The interpretation of structure from motion. Proceedings of the Royal Society of London. Series B. Biological Sciences 203(1153), 405-426 (1979)", + "77. Venkateswara, H., Eusebio, J., Chakraborty, S., Panchanathan, S.: Deep hashing network for unsupervised domain adaptation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5018-5027 (2017)", + "78. Vert, J.P., Tsuda, K., Scholkopf, B.: A primer on kernel methods. Kernel methods in computational biology 47, 35-70 (2004)", + "79. Wang, G., Manhardt, F., Liu, X., Ji, X., Tombari, F.: Occlusion-aware self-supervised monocular 6d object pose estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021)", + "80. Wang, G., Manhardt, F., Shao, J., Ji, X., Navab, N., Tombari, F.: Self6d: Self-supervised monocular 6d object pose estimation. In: European Conference on Computer Vision. pp. 108-125. Springer (2020)", + "81. Wu, Y., Greenspan, M.: Learning better keypoints for multi-object 6dof pose estimation. arXiv preprint arXiv:2308.07827 (2023)" + ], + "bbox": [ + 75, + 56, + 917, + 947 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "RKHPose", + "bbox": [ + 726, + 14, + 840, + 30 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 893, + 16, + 917, + 30 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "82. Wu, Y., Javaheri, A., Zand, M., Greenspan, M.: Keypoint cascade voting for point cloud based 6dof pose estimation. In: 2022 International Conference on 3D Vision (3DV). pp. 176-186. IEEE (2022)", + "83. Wu, Y., Zand, M., Etemad, A., Greenspan, M.: Vote from the center: 6 dof pose estimation in rgb-d images by radial keypoint voting. In: European Conference on Computer Vision. pp. 335-352. Springer (2022)", + "84. Xiang, Y., Schmidt, T., Narayanan, V., Fox, D.: PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes (2018)", + "85. Xiao, F., Liu, H., Lee, Y.J.: Identity from here, pose from there: Self-supervised disentanglement and generation of objects using unlabeled videos. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7013-7022 (2019)", + "86. Xiao, Y., Qiu, X., Langlois, P.A., Aubry, M., Marlet, R.: Pose from shape: Deep pose estimation for arbitrary 3d objects. arXiv preprint arXiv:1906.05105 (2019)", + "87. Yang, H., Pavone, M.: Object pose estimation with statistical guarantees: Conformal keypoint detection and geometric uncertainty propagation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8947-8958 (2023)", + "88. Zhang, Y., David, P., Gong, B.: Curriculum domain adaptation for semantic segmentation of urban scenes. In: Proceedings of the IEEE international conference on computer vision. pp. 2020-2030 (2017)", + "89. Zhang, Z., Wang, M., Huang, Y., Nehorai, A.: Aligning infinite-dimensional covariance matrices in reproducing kernel hilbert spaces for domain adaptation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3437-3445 (2018)", + "90. Zhou, J., Chen, K., Xu, L., Dou, Q., Qin, J.: Deep fusion transformer network with weighted vector-wise keypoints voting for robust 6d object pose estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13967-13977 (2023)" + ], + "bbox": [ + 77, + 56, + 917, + 573 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 77, + 16, + 104, + 30 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "Y.Wu and M. Greenspan", + "bbox": [ + 157, + 14, + 406, + 32 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/2024/Pseudo-keypoint RKHS Learning for Self-supervised 6DoF Pose Estimation/e3acf3fe-7127-4839-bf51-8bc132ae63e7_model.json b/2024/Pseudo-keypoint RKHS Learning for Self-supervised 6DoF Pose Estimation/e3acf3fe-7127-4839-bf51-8bc132ae63e7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..42ba7bb75e1311317698a4dd91f08993d4110337 --- /dev/null +++ b/2024/Pseudo-keypoint RKHS Learning for Self-supervised 6DoF Pose Estimation/e3acf3fe-7127-4839-bf51-8bc132ae63e7_model.json @@ -0,0 +1,2913 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.179, + 0.049, + 0.817, + 0.105 + ], + "angle": 0, + "content": "Pseudo-keypoint RKHS Learning for Self-supervised 6DoF Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.141, + 0.723, + 0.16 + ], + "angle": 0, + "content": "Yangzheng Wu and Michael Greenspan" + }, + { + "type": "text", + "bbox": [ + 0.143, + 0.175, + 0.855, + 0.23 + ], + "angle": 0, + "content": "RCVLab, Dept. of Electrical and Computer Engineering, Ingenuity Labs, Queen's University, Kingston, Ontario, Canada {y.wu, michael.greenspan}@queensu.ca" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.271, + 0.862, + 0.573 + ], + "angle": 0, + "content": "Abstract. We address the simulation-to-real domain gap in six degree-of-freedom pose estimation (6DoF PE), and propose a novel self-supervised keypoint voting-based 6DoF PE framework, effectively narrowing this gap using a learnable kernel in RKHS. We formulate this domain gap as a distance in high-dimensional feature space, distinct from previous iterative matching methods. We propose an adapter network, which is pre-trained on purely synthetic data with synthetic ground truth poses, and which evolves the network parameters from this source synthetic domain to the target real domain. Importantly, the real data training only uses pseudo-poses estimated by pseudo-keypoints, and thereby requires no real ground truth data annotations. Our proposed method is called RKHSPose, and achieves state-of-the-art performance among self-supervised methods on three commonly used 6DoF PE datasets including LINEMOD \\((+4.2\\%)\\), Occlusion LINEMOD \\((+2\\%)\\), and YCB-Video \\((+3\\%)\\). It also compares favorably to fully supervised methods on all six applicable BOP core datasets, achieving within \\(-11.3\\%\\) to \\(+0.2\\%\\) of the top fully supervised results." + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.591, + 0.851, + 0.627 + ], + "angle": 0, + "content": "Keywords: pose estimation \\(\\cdot\\) self-supervision \\(\\cdot\\) domain adaptation \\(\\cdot\\) key-point estimation" + }, + { + "type": "title", + "bbox": [ + 0.083, + 0.658, + 0.309, + 0.678 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.695, + 0.919, + 0.832 + ], + "angle": 0, + "content": "RGB-D Six Degree-of-Freedom Pose Estimation (6DoF PE) is a problem being actively explored in computer vision research. Given an RGB image and its associated depth map, the task is to detect scene objects and estimate their poses comprising 3DoF rotational angles and 3DoF translational offsets in the camera reference frame. This task enables many applications such as augmented reality [34,36,40,84], robotic bin picking [19,35,43], autonomous driving [51,85] and image-guided surgeries [21,25]." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.833, + 0.919, + 0.948 + ], + "angle": 0, + "content": "As with other machine learning (ML) tasks, fully supervised 6DoF PE requires large annotated datasets. This requirement is particularly challenging for 6DoF PE, as the annotations comprise not only the identity of the objects in the scene, but also their 6DoF pose, which makes the data relatively expensive to annotate compared to related tasks such as classification, detection and segmentation. This is due to the fact that humans are not able to qualitatively or" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.079, + 0.018, + 0.095, + 0.03 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.16, + 0.016, + 0.409, + 0.034 + ], + "angle": 0, + "content": "Y.Wu and M. Greenspan" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.051, + 0.904, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.207, + 0.924, + 0.3 + ], + "angle": 0, + "content": "Fig. 1: RKHSPose adapts the network pretrained on synthetic data to real test scenes (left), by comparing network feature spaces with real image inputs (solid arrows), against those with syn/real image (right) inputs (dashed arrows). \\( M_r \\) regresses radial quantities, \\( M_A \\) is the Adapter network, and RKHS maps features into a higher dimensional space." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.374, + 0.923, + 0.648 + ], + "angle": 0, + "content": "intuitively estimate 6DoF pose, which therefore requires additional instrumentation of the scene at data collection, and/or more sophisticated user annotation tools [34,84]. Consequently, synthetic 6DoF PE datasets [36,43] have been introduced, either as an additional complement to real datasets, or as standalone purely synthetic datasets. Annotated synthetic datasets are of course trivially inexpensive to collect, simply because precise synthetic pose annotation in a simulated environment is fully automatic. A known challenge in using synthetic data, however, is that there typically exists a domain gap between the real and synthetic data, which makes results less accurate when inferring in real data using models trained on purely synthetic datasets. Expectations of the potential benefit of synthetic datasets has led to the exploration of a rich set of Domain Adaptation (DA) methods, which specifically aim to reduce the domain gap [6,62,88] using inexpensive synthetic data for a wide variety of tasks, recently including 6DoF PE [28,45,47]." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.657, + 0.924, + 0.951 + ], + "angle": 0, + "content": "Early methods [42, 83] ignored the simulation-to-real (sim2real) domain gap and nevertheless improved performance by training on both synthetic and real annotated data, effectively augmenting the real images with the synthetic. However, these methods still required real labels to be sufficiently accurate and robust for practical applications, partially due to the domain gap. As shown in Fig. 1, the rendered synthetic objects (right image) have a slightly different appearance than the real objects (left image). The details of the CAD models, both geometry and texture, are not precise, as can be seen for the can object (which lacks a mouth and shadows), the benchwise (which is missing a handle), and the holepuncher (which has coarse geometric resolution). Several recent methods [9, 44, 73, 75, 80] have started to address the sim2real gap for 6DoF PE by first training on labeled synthetic data and then fine-tuning on unlabeled real data. Commonly known as self-supervised, these methods reduce the domain gap by adding extra supervision using features extracted from real images without requiring real ground truth (GT) labels. The majority of these methods are view-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.728, + 0.016, + 0.84, + 0.032 + ], + "angle": 0, + "content": "RKHPose" + }, + { + "type": "page_number", + "bbox": [ + 0.904, + 0.017, + 0.92, + 0.031 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.056, + 0.918, + 0.096 + ], + "angle": 0, + "content": "point/template matching-based, and the self-supervision commonly iteratively matches 2D masks or 3D point clouds [44,73,80]." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.097, + 0.92, + 0.29 + ], + "angle": 0, + "content": "While the above-mentioned self-supervision works have shown promise, there exist a wealth of DA techniques that can be brought to bear to improve performance further for this task. One such technique is Reproducing Kernel Hilbert Space (RKHS), which is a kernel method that has been shown to be effective for DA [1,5,41]. RKHS was initially used to create decision boundaries for non-separable data [12,60], and has been shown to be effective at reducing the domain gap for various tasks and applications [7,65,89]. The reproducing kernel guarantees that the domain gap can be statistically measured, allowing network parameters trained on synthetic data to be effectively adapted to the real data, using specifically tailored metrics." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.291, + 0.92, + 0.658 + ], + "angle": 0, + "content": "To address the sim2real domain gap in 6DoF PE, we propose RKHSPose, which is a keypoint-based method [82, 83] trained on a mixture of a large collection of labeled synthetic data, and a small handful of unlabeled real data. RKHSPose estimates the intermediate radial voting quantity, which has been shown to be effective for estimating keypoints [83], by first training a modified FCN-Resnet-18 on purely synthetic data, with automatically labeled synthetic GT poses. The radial quantity is a map of the distance from each image pixel to each keypoint. Next, real images are passed through the synthetically trained network, resulting in a set of pseudo-keypoints. The real images and their corresponding pseudo-keypoints are used to render a set of 'synthetic-overreal' (syn/real) images, by first estimating the pseudo-pose from the pseudo-keypoints, and then overlaying the synthetic object, rendered with the pseudopose, onto the real image. The network training then continues on the syn/real images, invoking an RKHS network module with a trainable linear product kernel, which minimizes the Maximum Mean Discrepancy (MMD) loss. At the front end, a proposed keypoint radial voting network learns to cast votes to estimate keypoints from the backend-generated radial maps. The final pose is then determined using ePnP [46] based on the estimated and corresponding object keypoints." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.66, + 0.547, + 0.678 + ], + "angle": 0, + "content": "The main contributions of this work are:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.918, + 0.724 + ], + "angle": 0, + "content": "- A novel learnable RKHS-based Adapter backend network architecture to minimize the sim2real domain gap in 6DoF PE;" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.724, + 0.805, + 0.743 + ], + "angle": 0, + "content": "- A novel CNN-based frontend network for keypoint radial voting;" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.743, + 0.918, + 0.799 + ], + "angle": 0, + "content": "- A self-supervised keypoint-based 6DoF PE method, RKHSPose, which is shown to have state-of-the-art (SOTA) performance, based on our experiments and several ablation studies." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.686, + 0.918, + 0.799 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.826, + 0.332, + 0.847 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.863, + 0.26, + 0.881 + ], + "angle": 0, + "content": "2.1 6DoF PE" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.891, + 0.918, + 0.949 + ], + "angle": 0, + "content": "ML-based 6DoF PE methods [57,61,84] all train a network to regress quantities, such as keypoints and camera viewpoints, as have been used in classical algorithms [34]. ML-based methods, which initially became popular for the general" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.018, + 0.095, + 0.03 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.16, + 0.016, + 0.405, + 0.032 + ], + "angle": 0, + "content": "Y.Wu and M. Greenspan" + }, + { + "type": "table_caption", + "bbox": [ + 0.079, + 0.053, + 0.92, + 0.143 + ], + "angle": 0, + "content": "Table 1: Existing self-supervised 6DoF PE methods. Some methods use CAD models, labeled synthetic (syn) data (images+poses), and real data without GT labels, while others use only labeled synthetic data. Many methods require a ROI such as a bounding box or a semantic mask, manually labeled or estimated by an existing framework. SO-Pose [18] used very few real labels to improve performance." + }, + { + "type": "table", + "bbox": [ + 0.208, + 0.159, + 0.79, + 0.507 + ], + "angle": 0, + "content": "
methodmodeCAD modelsyn datareal imagesreal posesROI
SO-Pose [18]RGB✓✓✓✓✓
TexPose [9]RGB✓✓✓X✓
SMOC-Net [75]RGB✓✓✓X✓
FS6D [33]RGBD✓✓XX✓
AAE [73]RGB✓✓XX✓
MHP [54]RGB✓✓XX✓
Sock et al. [68]RGB✓✓XX✓
DSC [86]RGBD✓✓XX✓
Sundermeyer [71]RGB✓✓XX✓
LatentFusion [59]RGBX✓XX✓
OSOP [66]RGBD✓✓XXX
Kleeberger et al. [44]D✓✓XXX
Su et al. [69]RGB✓✓✓XX
Self6D [80]RGBD✓✓✓XX
Self6D++ [79]RGB✓✓✓XX
Deng et al. [15]RGBD✓✓✓XX
RKHSPose (Ours)RGBD✓✓✓XX
" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.568, + 0.918, + 0.607 + ], + "angle": 0, + "content": "object detection task, have started to dominate the 6DoF PE literature due to their accuracy and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.611, + 0.92, + 0.766 + ], + "angle": 0, + "content": "There are two main categories of ML-based fully supervised methods: feature matching-based [29, 30, 57, 70, 84], and keypoint-based methods [31, 32, 61, 87]. Feature matching-based methods make use of the structures from general object detection networks most directly. The network encodes and matches features and estimates pose by either regressing elements of the pose (e.g. the transformation matrix [75], rotational angles and translational offsets [84] or 3D vertices [70]) directly, or by regressing some intermediate feature-matching representations, such as viewpoints or segments." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.77, + 0.92, + 0.907 + ], + "angle": 0, + "content": "In contrast, keypoint-based methods encode features to estimate keypoints which are predefined within the reference frame of an object's CAD model. These methods then use (modified) classical algorithms such as PnP [22, 46], Horn's method [38], and ICP [3] to estimate the final pose from corresponding image and model keypoints. Unlike feature-matching methods, keypoint-based methods are typically more accurate due to redundancies encountered through voting schemes [32, 61, 83] and by generating confidence hypotheses of keypoints [31, 87]." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.91, + 0.918, + 0.948 + ], + "angle": 0, + "content": "Recently, self-supervised 6DoF PE methods have been explored in order to reduce the reliance on labeled real data, which is expensive to acquire. As sum-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.728, + 0.016, + 0.84, + 0.032 + ], + "angle": 0, + "content": "RKHPose" + }, + { + "type": "page_number", + "bbox": [ + 0.904, + 0.017, + 0.919, + 0.031 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.055, + 0.919, + 0.173 + ], + "angle": 0, + "content": "marized in Table 1, these methods commonly use real images without GT labels. Some methods use pure synthetic data and CAD models only, with the exception of LatentFusion [59] which trained the model using only synthetic data. The majority of these methods [9, 15, 33, 48, 66, 68, 69, 71, 75, 79, 80] are inspired by fully supervised feature-matching methods, except DPODv2 [67] and DSC [86], in which keypoint correspondences are rendered and matched." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.18, + 0.92, + 0.394 + ], + "angle": 0, + "content": "A few methods [15,66] fine-tuned the pose trivially by iteratively matching the template/viewpoint, whereas others [33,44,59,86] augmented the training data by adding noise [44,86], rendering textures [33,67] and creating a latent space [59]. Some methods [54,71,73] also implemented DA techniques such as codebook encoding [71], Principle Component Analysis (PCA) [71] and symmetric Bingham distributions [54]. Most methods [66,68,69,79,80,86] used rendering techniques to render and match a template. There are a few methods that combined 3D reconstruction techniques, such as Neural radiance fields (Nerf) [55] and Structure from Motion (SfM) [76]. TexPose [9] matched CAD models to segments generated by Nerf, and SMOC-Net [75] used SfM to create the 3D segment and matched with the CAD model." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.458, + 0.583, + 0.479 + ], + "angle": 0, + "content": "2.2 Kernel Methods and Deep Learning" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.526, + 0.92, + 0.682 + ], + "angle": 0, + "content": "While deep learning is the most common ML technique within the computer vision literature, kernel methods [12,60,64] have also been actively explored. Kernel methods are typically in RKHS space [74] with reproducing properties that facilitate solving non-linear problems by mapping input data into high dimensional spaces that can be linearly separated [11,23]. Well-known early kernel methods that have been applied to computer vision are Support Vector Machines [12] and PCA [60]. A recent method [64] linked energy distance and MMD in RKHS, and showed the effectiveness of kernels in statistical hypothesis testing." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.689, + 0.92, + 0.884 + ], + "angle": 0, + "content": "More recent studies compare kernel methods with deep learning networks [1, 4,5,41]. RKHS is found to perform better on classification tasks than one single block of a CNN comprising convolution, pooling and downsampling (linear) layers [41]. RKHS can also help with CNN generalization by meta-regularization on image registration problems [1]. Similarly, norms (magnitude of trainable weights) defined in RKHS help with CNN regularization [4,5]. Further, discriminant information in label distributions in unsupervised DA is addressed and RKHS-based Conditional Kernel Bures metric is proposed [50]. Lastly, the connection between Neural Tangent Kernel and MMD is established, and an efficient MMD for two-sample tests is developed [10]." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.89, + 0.919, + 0.948 + ], + "angle": 0, + "content": "Inspired by the previous work, RKHSPose applies concepts of kernel learning to keypoint-based 6DoF PE, to provide an effective means to self-supervise a synthetically trained network on unlabeled real data." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.079, + 0.018, + 0.095, + 0.03 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.16, + 0.016, + 0.408, + 0.034 + ], + "angle": 0, + "content": "Y.Wu and M. Greenspan" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.053, + 0.709, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.06, + 0.891, + 0.141 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.795, + 0.142, + 0.826, + 0.153 + ], + "angle": 0, + "content": "\\(K_{syn}\\)" + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.153, + 0.892, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.785, + 0.236, + 0.839, + 0.248 + ], + "angle": 0, + "content": "\\(K_{syn / real}\\)" + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.248, + 0.891, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.796, + 0.335, + 0.828, + 0.347 + ], + "angle": 0, + "content": "\\(K_{real}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.363, + 0.921, + 0.455 + ], + "angle": 0, + "content": "Fig. 2: RKHSPose architecture. RKHPose is first trained on synthetic labeled data (solid arrows), and then finetuned on alternating syn/real and (unlabeled) real images (dashed arrows). \\( M_A \\) is measured by MMD in RKHS by densely mapping the intermediate features of \\( M_r \\) into high dimensional spaces with conv blocks. The distance is treated as \\( \\mathcal{L}_{M_A} \\) and back-propagated through \\( M_A \\) and \\( M_r \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.505, + 0.245, + 0.525 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.547, + 0.373, + 0.566 + ], + "angle": 0, + "content": "3.1 Network Overview" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.579, + 0.922, + 0.753 + ], + "angle": 0, + "content": "As shown in Fig. 2, RKHSPose is made up of two networks, a main network \\( M_{rv} \\) for keypoint regression and classification, and an Adapter network \\( M_A \\). The input of \\( M_{rv} \\) with shape \\( W \\times H \\times 4 \\), is the concatenation of an RGB image \\( I \\) and its corresponding depth map \\( D \\). This input can be synthetically generated with an arbitrary background (\\( I_{syn} \\) and \\( D_{syn} \\)), a synthetic mask overlayed on a real background (\\( I_{syn/real} \\) and \\( D_{syn/real} \\)), or a real (\\( I_{real} \\) and \\( D_{real} \\)) image. The outputs are \\( n \\) projected 2D keypoints \\( K \\), along with corresponding classification labels \\( C \\) and confidence scores \\( S \\). \\( K \\) are organized into instance sets based on \\( C \\) and geometric constraints." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.755, + 0.92, + 0.91 + ], + "angle": 0, + "content": "\\(M_{rv}\\) comprises two sub-networks, regression network \\(M_r\\) and voting network \\(M_v\\). Inspired by recent voting techniques [32, 61, 82, 83, 90], \\(M_r\\) estimates an intermediate voting quantity, which is a radial distance map \\(V_r\\) [83], by using a modified Fully Connected ResNet 18 (FCN-ResNet-18). The radial voting map \\(V_r\\), with shape \\(W \\times H\\), stores the Euclidean distance from each object point to each keypoint in the 3D camera world reference frame. The voting network \\(M_v\\) (described in Sec. 3.3) then takes \\(V_r\\) as input, accumulates votes, and detects peaks to estimate \\(K\\), \\(C\\) and \\(S\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.911, + 0.922, + 0.951 + ], + "angle": 0, + "content": "The Adapter network \\( M_A \\) consists of a series of CNNs which encode pairs of feature maps from \\( M_{rv} \\), and are trained on both synthetic overlayed and pure real" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.728, + 0.016, + 0.84, + 0.032 + ], + "angle": 0, + "content": "RKHPose" + }, + { + "type": "page_number", + "bbox": [ + 0.904, + 0.017, + 0.92, + 0.031 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.056, + 0.921, + 0.211 + ], + "angle": 0, + "content": "data. \\( M_A \\) encodes feature map pairs \\( (f_{syn / real}, f_{real}) \\) into corresponding high-dimensional feature maps \\( (h_{syn / real}, h_{real}) \\). The input data, \\( (I_{syn / real}, I_{real}) \\) and \\( (D_{syn / real}, D_{real}) \\), are also treated as \\( (f_{syn / real}, f_{real}) \\) during the learning of \\( M_A \\). Each of these networks creates a high-dimensional latent space, essentially the Reproducing Kernel Hilbert Space (RKHS) [2] and contributes to the learning of both \\( M_{rv} \\) and \\( M_A \\) by calculating the MMD [27]. While RKHS has been applied effectively to other DA tasks, to our knowledge, this is the first time that it has been applied to 6DoF PE, and the adapter network architecture is novel." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.212, + 0.921, + 0.328 + ], + "angle": 0, + "content": "The loss function \\(\\mathcal{L}\\) is made up of five elements: Radial regression loss \\(\\mathcal{L}_r\\) for \\(V_r\\); Keypoint projection loss \\(\\mathcal{L}_k\\) for \\(K\\); Classification loss \\(\\mathcal{L}_c\\) for \\(C\\); Confidence loss \\(\\mathcal{L}_s\\) for \\(S\\), and finally; Adapter loss \\(\\mathcal{L}_A\\) for the comparison of intermediate feature maps. The regression losses \\(\\mathcal{L}_r\\), \\(\\mathcal{L}_k\\), and \\(\\mathcal{L}_s\\) all use the smooth L1 metric, whereas classification loss \\(\\mathcal{L}_c\\) uses the cross-entropy metric \\(H(\\cdot)\\). RKHSPose losses can then be denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.337, + 0.918, + 0.358 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r} = \\operatorname {s m o o t h} _ {\\mathcal {L} _ {1}} \\left(V _ {r}, \\widehat {V} _ {r}\\right) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.352, + 0.363, + 0.918, + 0.384 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {k} = \\operatorname {s m o o t h} _ {\\mathcal {L} _ {1}} (K, \\widehat {K}) \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.39, + 0.918, + 0.408 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} = H (p (C), \\widehat {p} (C)) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.413, + 0.918, + 0.434 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} = \\operatorname {s m o o t h} _ {\\mathcal {L} _ {1}} (S, \\widehat {S}) \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.349, + 0.439, + 0.918, + 0.462 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {A} = M M D \\left(\\widehat {f} _ {\\text {s y n / r e a l}}, \\widehat {f} _ {\\text {r e a l}}\\right) \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.28, + 0.468, + 0.918, + 0.485 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {r} \\mathcal {L} _ {r} + \\lambda_ {k} \\mathcal {L} _ {k} + \\lambda_ {c} \\mathcal {L} _ {c} + \\lambda_ {s} \\mathcal {L} _ {s} + \\lambda_ {A} \\mathcal {L} _ {A} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.494, + 0.92, + 0.611 + ], + "angle": 0, + "content": "where \\(\\lambda_r, \\lambda_k, \\lambda_c, \\lambda_s\\), and \\(\\lambda_A\\) are weights for adjustment during training, and all non-hatted quantities are GT values. At inference, \\(M_{rv}\\) takes \\(I_{real}\\) and \\(D_{real}\\) as input, and outputs \\(K, C\\) and \\(S\\). The keypoints \\(K\\) are ranked and grouped by \\(S\\) and \\(C\\), and are then forwarded into the ePnP [46] algorithm which estimates 6DoF pose values. ICP can then be optionally applied using the depth data to refine the estimated pose." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.634, + 0.512, + 0.654 + ], + "angle": 0, + "content": "3.2 Convolutional RKHS Adapter" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.66, + 0.92, + 0.757 + ], + "angle": 0, + "content": "Reproducing Kernel Hilbert Space \\(\\mathcal{H}\\) is a commonly used vector space for Domain Adaptation [58, 77]. Hilbert Space is a complete metric space (in which every Cauchy sequence of points has a limit within the metric) represented by the inner product of vectors. For a non-empty set of data \\(\\mathcal{X}\\), a function \\(\\mathcal{K}_{\\mathcal{X}}: \\mathcal{X} \\times \\mathcal{X} \\to \\mathcal{R}\\) is a reproducing kernel if:" + }, + { + "type": "equation", + "bbox": [ + 0.276, + 0.767, + 0.918, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} k (\\cdot , x) \\in \\mathcal {H} \\quad \\forall \\quad x \\in \\mathcal {X} \\\\ \\langle f (\\cdot), k (\\cdot , x) \\rangle = f (x) \\quad \\forall \\quad x \\in \\mathcal {X}, f \\in \\mathcal {H} \\end{array} \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.823, + 0.919, + 0.881 + ], + "angle": 0, + "content": "where \\(\\langle a,b\\rangle\\) denotes the inner product of two vectors \\(a\\) and \\(b\\), \\(k(\\cdot ,x) = \\mathcal{K}_{\\mathcal{X}}\\) for each \\(x\\in \\mathcal{X}\\), and \\(f\\) is a function in \\(\\mathcal{H}\\). The second equation in Eq. 7 is an expression of the reproducing property of \\(\\mathcal{H}\\)." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.882, + 0.921, + 0.921 + ], + "angle": 0, + "content": "In order to utilize \\(\\mathcal{H}\\) for DA, we expand the kernel definition to two sets of data \\(\\mathcal{X}\\) and \\(\\mathcal{Y}\\). The reproducing kernel can then be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.931, + 0.918, + 0.95 + ], + "angle": 0, + "content": "\\[\n\\mathcal {K} (\\mathcal {X}, \\mathcal {Y}) = \\left\\langle \\mathcal {K} _ {\\mathcal {X}}, \\mathcal {K} _ {\\mathcal {Y}} \\right\\rangle_ {H} \\tag {8}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.079, + 0.017, + 0.096, + 0.03 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.407, + 0.033 + ], + "angle": 0, + "content": "Y.Wu and M. Greenspan" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.056, + 0.921, + 0.191 + ], + "angle": 0, + "content": "Here, \\(\\mathcal{K}_{\\mathcal{X}}\\) and \\(\\mathcal{K}_{\\mathcal{Y}}\\) are themselves inner product kernels that map \\(\\mathcal{X}\\) and \\(\\mathcal{Y}\\) respectively into their own Hilbert spaces, and \\(\\mathcal{K}(\\mathcal{X},\\mathcal{Y})\\) is the joint Hilbert space kernel of \\(\\mathcal{X}\\) and \\(\\mathcal{Y}\\). Note that \\(\\mathcal{K}(\\mathcal{X},\\mathcal{Y})\\) is not the kernel commonly defined for CNNs. Rather, it is the similarity function defined in kernel methods, such as is used by Support Vector Machine (SVM) techniques to calculate similarity measurements. Some recent methods [8, 52, 53, 56] named it a Convolutional Kernel Network (CKN) to distinguish it from CNNs." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.192, + 0.921, + 0.348 + ], + "angle": 0, + "content": "Various kernels of CKN, such as the Gaussian Kernel [53], the RBF Kernel [52] and the Inner Product Kernel [56], have been shown to be comparable to shallow CNNs for various tasks, especially for Domain Adaption. The most intuitive inner product kernel is mathematically similar to a fully connected layer, where trainable weights are multiplied by the input feature map. To allow the application of RKHS methods into our CNN based Adapter network \\( M_A \\), trainable weights are added to \\( \\mathcal{K}(\\mathcal{X},\\mathcal{Y}) \\) [49]. The trainable Kernel \\( \\mathcal{K}_w(\\mathcal{X},\\mathcal{Y}) \\) can then be denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.318, + 0.362, + 0.918, + 0.381 + ], + "angle": 0, + "content": "\\[\n\\mathcal {K} _ {w} (\\mathcal {X}, \\mathcal {Y}) = \\langle \\langle \\mathcal {X}, W _ {\\mathcal {X}} \\rangle , \\langle \\mathcal {Y}, W _ {\\mathcal {Y}} \\rangle \\rangle_ {H} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.391, + 0.919, + 0.429 + ], + "angle": 0, + "content": "where \\( W_{\\mathcal{X}} \\) and \\( W_{\\mathcal{Y}} \\) are trainable weights. By adding \\( W_{\\mathcal{X}} \\) and \\( W_{\\mathcal{Y}} \\), \\( K_{w}(\\mathcal{X},\\mathcal{Y}) \\) still satisfies the RKHS constraints." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.43, + 0.921, + 0.626 + ], + "angle": 0, + "content": "The sim2real domain gap of feature maps \\( f \\) being trained in \\( M_{rv} \\) (with few real images and no real GT labels) are hard to measure using trivial distance metrics. In contrast, RKHS can be a more accurate and robust space for comparison, since it is known to be capable of handling high-dimensional data with a low number of samples [24, 56]. To compare \\( (f_{syn / real}, f_{real}) \\) in RKHS, a series of CNN layers encodes \\( (f_{syn / real}, f_{real}) \\) into higher-dimensional features \\( (h_{syn / real}, h_{real}) \\), followed by the trainable \\( \\mathcal{K}_w(\\mathcal{X}, \\mathcal{Y}) \\). Once mapped into RKHS, \\( h_{syn / real} = \\{sr_i\\}_{i=1}^m \\) and \\( h_{real} = \\{r_i\\}_{i=1}^m \\) can then be measured by Maximum Mean Discrepancy (MMD), a common DA measurement [27, 49], which is the square distance between the kernel embedding [26]:" + }, + { + "type": "equation", + "bbox": [ + 0.083, + 0.637, + 0.93, + 0.782 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} M M D \\left(h _ {s y n / r e a l}, h _ {r e a l}\\right) = \\frac {1}{m} \\left[ \\left(\\sum_ {i = 1} ^ {m} \\sum_ {j = 1} ^ {m} k _ {w} \\left(s r _ {i}, s r _ {j}\\right) - \\sum_ {i = 1} ^ {m} k _ {w} \\left(s r _ {i}, s r _ {i}\\right)\\right) \\right. \\\\ \\left. - \\left(\\sum_ {i = 1} ^ {m} \\sum_ {j = 1} ^ {m} k _ {w} \\left(s r _ {i}, r _ {j}\\right) - \\sum_ {i = 1} ^ {m} k _ {w} \\left(s r _ {i}, r _ {i}\\right)\\right) + \\left(\\sum_ {i = 1} ^ {m} \\sum_ {j = 1} ^ {m} k _ {w} \\left(r _ {i}, r _ {j}\\right) - \\sum_ {i = 1} ^ {m} k _ {w} \\left(r _ {i}, r _ {i}\\right)\\right) \\right] ^ {\\frac {1}{2}} \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.794, + 0.531, + 0.813 + ], + "angle": 0, + "content": "where \\(k_{w}()\\) is the feature element of \\(\\mathcal{K}_w(\\)" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.813, + 0.921, + 0.95 + ], + "angle": 0, + "content": "In summary, the Adapter \\( M_A \\) shown in Fig. 2 measures MMD for each \\( (h_{syn / real}, h_{real}) \\) in RKHS, by increasing the \\( f_{syn / real} \\) and \\( f_{real} \\) dimension using a CNN and thereby constructing a learnable kernel \\( \\mathcal{K}_w \\). The outputs are the feature maps \\( h_{syn / real} \\) and \\( h_{real} \\), which are supervised by loss \\( \\mathcal{L}_A \\) during training of the real data epochs. Based on the experiments in Sec. 5.3, our trainable inner product kernel is shown to be more accurate for our task than other known kernels [52, 53, 56] that we tested, that are often used in such kernel methods." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.728, + 0.016, + 0.84, + 0.032 + ], + "angle": 0, + "content": "RKHPose" + }, + { + "type": "page_number", + "bbox": [ + 0.903, + 0.017, + 0.92, + 0.031 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.055, + 0.548, + 0.075 + ], + "angle": 0, + "content": "3.3 Keypoint Radial Voting Network" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.088, + 0.92, + 0.339 + ], + "angle": 0, + "content": "The network \\( M_v \\) votes for keypoints using a CNN architecture, taking the radial voting quantity \\( V_r \\) resulting from \\( M_r \\) as input. VoteNet [63] previously used a CNN approach to vote for object centers, whereas other keypoint-based techniques have implemented GPU-based parallel RANSAC [32, 61] methods for offset and vector quantities. The radial quantity is known to be more accurate than the vector or offset quantities, and has been previously implemented with a CPU-based parallel accumulator space method [82, 83]. Given its superior accuracy, \\( M_v \\) implements radial voting using a CNN to improve efficiency. Given a 2D radial map \\( \\widehat{V}_r \\) estimated by \\( M_r \\) and supervised by GT radial maps \\( V_r \\), the task is to accumulate votes, find the peak, and estimate the keypoint location. The \\( V_r \\) foreground pixels (which lie on the target object) store the Euclidean distance from these pixels to each of the keypoints, with background (non-object) pixels set to value -1." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.341, + 0.92, + 0.711 + ], + "angle": 0, + "content": "The estimated radial map \\(\\widehat{V}_r\\) is indeed an inverse heat map of the candidate keypoints' locations, distributed in a radial pattern centered at the keypoints. To forward \\(\\widehat{V}_r\\) into a CNN voting module, it is inversely normalized so that it becomes a heat map. Let \\(v_r^{max}\\) and \\(v_r^{min}\\) be the maximum and minimum global radial distances for all objects in a dataset, which can be calculated by iterating through all GT radial maps, or alternately generated from the object CAD models. An inverse radial map \\(\\widehat{V}_r^{-1}\\) can then be denoted as \\(\\widehat{V}_r^{-1} = (v_r^{max} - \\widehat{V}_r) / (v_r^{max} - v_r^{min})\\). Voting network \\(M_v\\) takes \\(\\widehat{V}_r^{-1}\\) as the input and generates the accumulated vote map by a series of convolution, ReLu, and batch normalization layers. The complete network architecture is provided in the Supplementary material Sec. S.2. The background pixels are filtered out by a ReLu layer, and only foreground pixels contribute to voting. The accumulated vote map is then max-pooled for peak extraction, and reshaped using a fully connected layer into a \\(n\\times 4\\) output. The output represents \\(n\\) keypoints and comprises \\(n\\times 2\\) projected 2D keypoints \\(K\\), \\(n\\) classification labels \\(C\\), and \\(n\\) confidence scores \\(S\\). The labels \\(C\\) indicate which object the corresponding keypoints belongs to, and \\(S\\) ranks the confidence level of keypoints before being forwarded into ePnP [46] for pose estimation. \\(M_v\\) is supervised by \\(\\mathcal{L}_k\\), \\(\\mathcal{L}_c\\), and \\(\\mathcal{L}_s\\) (Eqs. 2-4) and is trained end-to-end along with \\(M_r\\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.741, + 0.311, + 0.763 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.781, + 0.543, + 0.799 + ], + "angle": 0, + "content": "4.1 Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.812, + 0.92, + 0.95 + ], + "angle": 0, + "content": "RKHSPose uses BOP Procedural Blender [17] (PBR) synthetic images [16, 36, 37, 72] for the synthetic training phase. The images are generated by dropping synthetic objects (CAD models) onto a plane in a simulated environment using PyBullet [13], and then rendering them with synthetic textures. All objects in the synthetic images are thus automatically labeled with precise GT poses. We evaluated RKHSPose for the six BOP [37, 72] core datasets (LMO [34], YCB [84], TLESS [35], TUDL [36], ITODO [20], and HB [40]), all except IC-BIN [19]," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.106, + 0.032 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.409, + 0.034 + ], + "angle": 0, + "content": "Y.Wu and M. Greenspan" + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.055, + 0.92, + 0.133 + ], + "angle": 0, + "content": "which does not include any real training or validation images and is therefore not applicable. ITODO and HB have no real images in the training set, and so for training we instead used the real images in their validation sets, which were disjoint from their test sets." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.135, + 0.92, + 0.331 + ], + "angle": 0, + "content": "Our main results are evaluated with the ADD(S) [34] metric for the LM and LMO dataset, and the ADD(S) AUC [84] metric for the YCB dataset. These are the standard metrics commonly used to compare self-supervised 6DoF PE methods. ADD(S) is based on the mean distance (minimum distance for symmetry) between the object surfaces for GT and estimated poses, whereas ADD(S) AUC plots a curve formed by ADD(S) for various object diameter thresholds. We use the BOP average recall (\\( AR \\)) metrics for our ablation studies. The \\( AR \\) metric, based on the original ADD(S) [34], evaluates three aspects, including Visible Surface Discrepancy (\\( AR_{VSD} \\)), Maximum Symmetry-Aware Surface/Projection Distance (\\( AR_{MSSD} \\) and \\( AR_{MSPD} \\)) [36]." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.364, + 0.432, + 0.383 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.4, + 0.92, + 0.499 + ], + "angle": 0, + "content": "RKHSPose is trained on a server with an Intel Xeon 5218 CPU and two RTX6000 GPUs with a batch size of 32. The Adam optimizer is used for the training of \\( M_{rv} \\), on both synthetic and real data, and \\( M_A \\) is optimized by SGD. Both of the optimizers have an initial learning rate of \\( lr = 1e - 3 \\) and weight decay 1e-4 for 80 and 20 epochs respectively." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.5, + 0.92, + 0.597 + ], + "angle": 0, + "content": "The input of the network is normalized before training, as follows. The RGB images \\(I\\) are normalized and standardized using ImageNet [14] specifications. The depth maps are each individually normalized by their local minima and maxima, to lie within a range of 0 to 1. The radial distances in radial map \\(V_{r}\\) and 2D projected keypoints are both normalized by the width and height of \\(I\\)." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.598, + 0.92, + 0.734 + ], + "angle": 0, + "content": "A single set of four keypoints is chosen by KeyGNet [81] for the set of all objects in each dataset. One extra background class is added to \\( C \\) in order to filter out the redundant background points in \\( K \\). \\( M_{rv} \\) is first trained for 120 epochs on synthetic data, during which \\( M_A \\) remains frozen. Following this, training proceeds for an additional 80 epochs which alternate between real and synthetic data. When training on real data, both \\( M_A \\) and \\( M_{rv} \\) weights are learned, whereas \\( M_A \\) is frozen for the alternating synthetic data training." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.735, + 0.92, + 0.948 + ], + "angle": 0, + "content": "During real data epochs, \\( M_{rv} \\) initially estimates pseudo-keypoints for each real image. These pseudo-keypoints are then forwarded into ePnP for pseudopose estimation. Each estimated pseudo-pose is then augmented into a set of poses \\( P_{aug} \\) by applying arbitrary rotational and translational perturbations with respective ranges of \\( [-\\frac{\\pi}{18}, \\frac{\\pi}{18}] \\) radians and \\( [-0.1, 0.1] \\) along three axes within the normalized model frame, which is defined using the largest object in the dataset. The set of syn/real images are rendered by overlaying onto the real image the CAD model of each object using each augmented pose value in \\( P_{aug} \\). The cardinality of \\( P_{aug} \\) is set to be one less than the batch size, and the adapter \\( M_A \\) is trained on a mini-batch of the hybrid images resulting from \\( P_{aug} \\), plus the image resulting from the original estimated pseudo-pose." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.728, + 0.016, + 0.842, + 0.032 + ], + "angle": 0, + "content": "RKHPose" + }, + { + "type": "page_number", + "bbox": [ + 0.894, + 0.017, + 0.917, + 0.031 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.075, + 0.035, + 0.921, + 0.09 + ], + "angle": 0, + "content": "Table 2: Comparison with other methods. Accuracy of RKHSPose for LM and LMO is evaluated with ADD(S), and for YCB is evaluated with ADD(S) AUC. All 'Supervision: Syn + Self' methods use real images without real labels." + }, + { + "type": "table", + "bbox": [ + 0.182, + 0.106, + 0.819, + 0.585 + ], + "angle": 0, + "content": "
MethodReal data image labelDataset/Metric
LM ADD(S)LMOYCB ADD(S)ADD-S AUC
Supervision: Syn (lower bound)
AAEXX31.4---
MHPXX38.8---
GDR (TexPose version)XX77.452.9--
Self6D++XX77.452.977.889.4
Self6D++ with DrefXX88.062.579.290.1
OursXX78.254.376.590.2
Ours+ICPXX87.955.778.391.3
Supervision: Syn + Self
Sock et al.✓X60.622.8--
DSC✓X58.624.8--
Self6D✓X58.932.1--
SMOC-Net✓X91.363.3--
Self6D++✓X88.564.780.091.4
TexPose✓X91.766.7--
Ours✓X95.868.682.892.4
Ours+ICP✓X95.968.783.092.6
Supervision: Syn + Real GT (upper bound)
SO-Pose✓✓96.062.383.990.9
Self6D++✓✓91.074.482.690.7
Ours✓✓96.770.885.492.2
Ours+ICP✓✓96.871.385.692.4
" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.637, + 0.922, + 0.794 + ], + "angle": 0, + "content": "Initially \\( M_A \\) is frozen, and for the first 80 epochs, the loss is set to emphasize the classification and the visibility score regression, i.e. \\( \\lambda_c = \\lambda_s = 0.6 \\) and \\( \\lambda_r = \\lambda_k = 0.4 \\). Following this, up to epoch 200, the scales of losses are then exchanged to fine-tune the localization of the keypoints, i.e. \\( \\lambda_c = \\lambda_s = 0.4 \\) and \\( \\lambda_r = \\lambda_k = 0.6 \\). After epoch 120, \\( M_A \\) is unfrozen each alternating epoch, and \\( \\lambda_D \\) is set to 1 during the remaining \\( M_A \\) training epochs. This training strategy, shown in Fig. 2, minimizes the sim2real gap without any real image GT labels, and using very few (320) real images." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.821, + 0.238, + 0.839 + ], + "angle": 0, + "content": "4.3 Results" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.852, + 0.923, + 0.951 + ], + "angle": 0, + "content": "The results are summarized in Tables 2 and 3. To our knowledge, RKHSPose outperforms all existing self-supervised 6DoF PE methods. In Table 2, the upper bound fully supervised, lower bound synthetically supervised, and middle self-supervised methods (including ours) are compared. On LM and LMO, our ADD(S) is \\(+4.2\\%\\) and \\(+2\\%\\) better than the second best method TexPose [9]. We" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.105, + 0.031 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.409, + 0.034 + ], + "angle": 0, + "content": "Y.Wu and M. Greenspan" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.054, + 0.921, + 0.127 + ], + "angle": 0, + "content": "Table 3: Comparison with fully supervised methods. RKHSPose results on TLESS \\((-1.8)\\), TUDL \\((-0.4)\\), ITODD \\((-4.6)\\) and HB \\((+0.1)\\) compares to SOTA methods with full supervision of real GT labels. Methods annotated with * use the detection results from other detection methods." + }, + { + "type": "table", + "bbox": [ + 0.133, + 0.142, + 0.868, + 0.299 + ], + "angle": 0, + "content": "
Methodreal labelLMOTLESSDataset
TUDLITODDHBYCB
SurfEmb* [30]✓-76.082.885.465.986.679.9
RCVPose3D [82]✓-72.970.896.673.386.384.3
RADet [87]+PFA* [39]✓-79.785.096.067.686.988.8
ZebraPose [70]✓-78.086.295.665.492.189.9
Ours✗95.768.285.596.268.692.283.6
Ours+ICP✗95.868.485.696.268.792.383.8
" + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.331, + 0.218, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.331, + 0.357, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.332, + 0.498, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.332, + 0.639, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.331, + 0.779, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.782, + 0.331, + 0.92, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.44, + 0.921, + 0.477 + ], + "angle": 0, + "content": "Fig. 3: Qualitative overlay results on selected images. Red dots and blue dots are projected surface points from GT poses and estimated poses, respectively." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.531, + 0.92, + 0.628 + ], + "angle": 0, + "content": "compared our performance to Self6D++ [79], which was the only other method that evaluated using YCB, and saw a \\(3\\%\\) improvement after ICP on ADD(S) AUC. Last but not least, in Table 3 our performance evaluated on the other four BOP core datasets is comparable to several upper bound methods from the BOP leaderboard. Some test scenes with RKHSPose results are shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.629, + 0.92, + 0.707 + ], + "angle": 0, + "content": "RKHSPose runs at 34 fps on an Intel i7 2.5GHz CPU and an RTX 3090 GPU with 24G VRAM. It takes on average \\(8.7\\mathrm{ms}\\) for loading data, \\(4.5\\mathrm{ms}\\) for forward inference through \\(M_{rv}\\) (\\(\\times 10\\) faster compared to the analytical radial voting in RCVPose [83]), and \\(16.2\\mathrm{ms}\\) for ePnP." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.738, + 0.371, + 0.758 + ], + "angle": 0, + "content": "5 Ablation Studies" + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.78, + 0.465, + 0.8 + ], + "angle": 0, + "content": "5.1 Dense Vs. Sparse Adapter" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.813, + 0.922, + 0.95 + ], + "angle": 0, + "content": "The Adapter \\( M_A \\) densely matches the intermediate feature maps, whereas the majority of other methods [15,75,80] only compare the final output. To show the benefits of dense comparison, we conduct an experiment with different variations of \\( M_A \\). A sparse matching \\( M_A^s \\) network is trained on synthetic and real data comparing only a single feature map, which is the intermediate radial map. \\( M_A^s \\) has the exact same overall learning capacity (number of parameters) as the dense matching \\( M_A \\) described in Sec. 3.3. The results in Table 4 show that \\( M_A \\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.728, + 0.016, + 0.84, + 0.032 + ], + "angle": 0, + "content": "RKHPose" + }, + { + "type": "page_number", + "bbox": [ + 0.894, + 0.017, + 0.92, + 0.031 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.054, + 0.479, + 0.089 + ], + "angle": 0, + "content": "Table 4: \\( {AR} \\) of different adapters on LM and five BOP core datasets." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.106, + 0.479, + 0.171 + ], + "angle": 0, + "content": "
AdapterARVSDARMSSDARMSPDAR
MA78.177.877.877.9
MA84.984.184.384.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.521, + 0.054, + 0.92, + 0.09 + ], + "angle": 0, + "content": "Table 5: \\( {AR} \\) of different training strategies on LM and five BOP core datasets." + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.107, + 0.922, + 0.163 + ], + "angle": 0, + "content": "
Training TypeARVSDARMSSDARMSPDAR
Mixed84.984.184.384.4
Sequential82.381.781.781.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.188, + 0.887, + 0.223 + ], + "angle": 0, + "content": "Table 6: [R Tab A] \\( AR \\) of different kernel- Table 7: \\( AR \\) of different metrics on LM nets on LM and five BOP core datasets. and five BOP core datasets." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.24, + 0.487, + 0.337 + ], + "angle": 0, + "content": "
KernelwARVSDARMSSDARMSPDAR
LinearX71.670.870.671.0
✓84.984.184.384.4
RBFX73.472.973.273.2
✓82.581.381.581.6
" + }, + { + "type": "table", + "bbox": [ + 0.487, + 0.24, + 0.887, + 0.322 + ], + "angle": 0, + "content": "
MetricARVSDARMSSDARMSPDAR
MMD84.984.184.384.4
KL Div78.077.878.077.9
Wass80.980.680.980.8
" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.375, + 0.921, + 0.435 + ], + "angle": 0, + "content": "surpassed \\( M_A^s \\) on all six datasets tested. Specifically, on ITODD, \\( M_A \\) is \\( 12.1\\% \\) more accurate than \\( M_A^s \\). This experiment shows the effectiveness of our densely matched \\( M_A \\)." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.465, + 0.543, + 0.486 + ], + "angle": 0, + "content": "5.2 Syn/Real Synchronized Training" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.5, + 0.92, + 0.618 + ], + "angle": 0, + "content": "When training RKHSPose, real epochs are alternated with synthetic epochs. In contrast, some other methods [9, 75, 79, 80] separate the synthetic/real training. We conducted an experiment to compare these two different training strategies, the results of which are shown in Table 5. The alternating training performs slightly better (\\(+2.5\\%\\) on average) than the sequential training, possibly due to the early access to real scenes thereby avoiding local minima." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.649, + 0.502, + 0.669 + ], + "angle": 0, + "content": "5.3 Adapter Kernels and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.683, + 0.92, + 0.799 + ], + "angle": 0, + "content": "We use a linear (dot product) kernel and MMD in RKHS for domain gap measurements. There are various other kernels and similarity measurements that can be implemented in RKHS as described in Sec. 3.2. First, we add trainable weights to the radial basis function (RBF) kernel in a similar manner as \\( K_{w} \\) defined in Eq. 9. The trainable RBF kernel on two sets of data \\( X \\) and \\( Y \\) is denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.318, + 0.799, + 0.919, + 0.824 + ], + "angle": 0, + "content": "\\[\nK _ {r b f} (X, Y) = \\exp (- w \\| X - Y \\| ^ {2}) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.832, + 0.92, + 0.95 + ], + "angle": 0, + "content": "where \\( w \\) are the trainable weights, which replaces the original adjustable parameter in the classical RBF kernel. We also experiment with the classical RKHS kernel functions without trainable weights, including the inner product kernel and the original RBF kernel [78], for comparison. Further, we experiment on other commonly used distance measures, including Kullback-Leibler Divergence (KL Div, i.e. relative-entropy) and Wasserstein (Wass) Distance." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.106, + 0.032 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.409, + 0.034 + ], + "angle": 0, + "content": "Y.Wu and M. Greenspan" + }, + { + "type": "image", + "bbox": [ + 0.202, + 0.052, + 0.796, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.273, + 0.923, + 0.33 + ], + "angle": 0, + "content": "Fig. 4: Impact of # of real images with/without GT labels used during training. All datasets are evaluated by the BOP AR metric. We conduct experiments from 0 to 640 real images on all datasets, except ITODD which contained only 357 real images." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.38, + 0.923, + 0.481 + ], + "angle": 0, + "content": "In Table 6, the RBF kernel performs similar to the linear product kernel with a slight performance dip. In Table 7, MMD minimizes the domain gap better than the Wass Distance, followed by the KL Div metric, leading to a better overall performance on \\( AR \\). Based on these results, we used the linear product kernel with trainable weights and selected MMD as the main loss metric." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.503, + 0.642, + 0.525 + ], + "angle": 0, + "content": "5.4 Number of Real Images and Real Labels" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.533, + 0.925, + 0.71 + ], + "angle": 0, + "content": "The objective of RKHSPose is to reduce real data usage and train without any real GT labels. To show the effectiveness of the approach, we conducted an experiment by training on different numbers of real images, the results of which are shown in Fig. 4. We used up to 640 real images in all cases, except for that of ITODD which contains only 357 real images. The \\( AR \\) of all datasets saturates at 160 images except YCB. The further improvement of YCB beyond 160 images is also only \\( +0.1\\% \\) and saturates after 320 real images. We nevertheless use 320 real unlabeled images for our main results. This experiment showed that adding more than 320 real labeled images did not significantly improve performance." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.736, + 0.291, + 0.758 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.774, + 0.925, + 0.951 + ], + "angle": 0, + "content": "To sum up, we propose a novel self-supervised keypoint radial voting-based 6DoF PE method using RGB-D data called RKHSPose. RKHSPose fine-tunes poses pre-trained on synthetic data by densely matching features with a learnable kernel in RKHS, using real data albeit without any real GT poses. By applying this DA technique in feature space, RKHSPose achieved SOTA performance on the six applicable BOP core datasets, surpassing the performance of all other self-supervised methods. Notably, the RKHSPose performance closely approaches that of several fully-supervised methods, which indicates the strength of the approach at reducing the sim2real domain gap for this problem." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.728, + 0.017, + 0.84, + 0.032 + ], + "angle": 0, + "content": "RKHPose" + }, + { + "type": "page_number", + "bbox": [ + 0.894, + 0.017, + 0.919, + 0.031 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.056, + 0.921, + 0.096 + ], + "angle": 0, + "content": "Acknowledgements: Thanks to Bluewrist Inc. and NSERC for their support of this work." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.126, + 0.238, + 0.145 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.165, + 0.92, + 0.219 + ], + "angle": 0, + "content": "1. Al Safadi, E., Song, X.: Learning-based image registration with meta-regularization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10928-10937 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.221, + 0.918, + 0.254 + ], + "angle": 0, + "content": "2. Aronszajn, N.: Theory of reproducing kernels. Transactions of the American mathematical society 68(3), 337-404 (1950)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.256, + 0.921, + 0.308 + ], + "angle": 0, + "content": "3. Besl, P.J., McKay, N.D.: Method for registration of 3-d shapes. In: Sensor fusion IV: control paradigms and data structures. vol. 1611, pp. 586-606. International Society for Optics and Photonics (1992)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.31, + 0.92, + 0.362 + ], + "angle": 0, + "content": "4. Bietti, A., Mairal, J.: Group invariance, stability to deformations, and complexity of deep convolutional representations. The Journal of Machine Learning Research 20(1), 876-924 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.363, + 0.921, + 0.415 + ], + "angle": 0, + "content": "5. Bietti, A., Mialon, G., Chen, D., Mairal, J.: A kernel perspective for regularizing deep neural networks. In: International Conference on Machine Learning. pp. 664-674. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.417, + 0.918, + 0.451 + ], + "angle": 0, + "content": "6. Bozorgtabar, B., Mahapatra, D., Thiran, J.P.: Exprada: Adversarial domain adaptation for facial expression analysis. Pattern Recognition 100, 107111 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.452, + 0.921, + 0.504 + ], + "angle": 0, + "content": "7. Chen, C., Fu, Z., Chen, Z., Jin, S., Cheng, Z., Jin, X., Hua, X.S.: Homm: Higher-order moment matching for unsupervised domain adaptation. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 3422-3429 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.505, + 0.92, + 0.558 + ], + "angle": 0, + "content": "8. Chen, D., Jacob, L., Mairal, J.: Convolutional kernel networks for graph-structured data. In: International Conference on Machine Learning. pp. 1576–1586. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.559, + 0.92, + 0.611 + ], + "angle": 0, + "content": "9. Chen, H., Manhardt, F., Navab, N., Busam, B.: Texpose: Neural texture learning for self-supervised 6d object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4841-4852 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.612, + 0.918, + 0.646 + ], + "angle": 0, + "content": "10. Cheng, X., Xie, Y.: Neural tangent kernel maximum mean discrepancy. Advances in Neural Information Processing Systems 34, 6658-6670 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.647, + 0.917, + 0.682 + ], + "angle": 0, + "content": "11. Corcoran, P.: An end-to-end graph convolutional kernel support vector machine. Applied Network Science 5(1), 1-15 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.683, + 0.918, + 0.717 + ], + "angle": 0, + "content": "12. Cortes, C., Vapnik, V.: Support-vector networks. Machine learning 20, 273-297 (1995)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.718, + 0.918, + 0.754 + ], + "angle": 0, + "content": "13. Coumans, E., Bai, Y.: Pybullet, a python module for physics simulation for games, robotics and machine learning. http://pybullet.org (2016-2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.755, + 0.918, + 0.823 + ], + "angle": 0, + "content": "14. Deng, J., Dong, W., Socher, R., Li, L., Kai Li, Li Fei-Fei: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. pp. 248-255 (2009). https://doi.org/10.1109/CVPR.2009.5206848" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.825, + 0.918, + 0.878 + ], + "angle": 0, + "content": "15. Deng, X., Xiang, Y., Mousavian, A., Eppner, C., Bretl, T., Fox, D.: Self-supervised 6d object pose estimation for robot manipulation. In: 2020 IEEE International Conference on Robotics and Automation (ICRA). pp. 3665-3671. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.879, + 0.918, + 0.949 + ], + "angle": 0, + "content": "16. Denninger, M., Sundermeyer, M., Winkelbauer, D., Olefir, D., Hodan, T., Zidan, Y., Elbadrawy, M., Knauer, M., Katam, H., Lodhi, A.: Blenderproc: Reducing the reality gap with photorealistic rendering. In: International Conference on Robotics: Science and Systems, RSS 2020 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.081, + 0.165, + 0.921, + 0.949 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.106, + 0.031 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.409, + 0.034 + ], + "angle": 0, + "content": "Y.Wu and M. Greenspan" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.057, + 0.918, + 0.129 + ], + "angle": 0, + "content": "17. Denninger, M., Winkelbauer, D., Sundermeyer, M., Boerdijk, W., Knauer, M., Strobl, K.H., Humt, M., Triebel, R.: Blenderproc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software 8(82), 4901 (2023). https://doi.org/10.21105/joss.04901, https://doi.org/10.21105/joss.04901" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.129, + 0.92, + 0.199 + ], + "angle": 0, + "content": "18. Di, Y., Manhardt, F., Wang, G., Ji, X., Navab, N., Tombari, F.: So-posed: Exploiting self-occlusion for direct 6d pose estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 12396-12405 (October 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.199, + 0.918, + 0.252 + ], + "angle": 0, + "content": "19. Doumanoglou, A., Kouskouridas, R., Malassiotis, S., Kim, T.K.: Recovering 6d object pose and predicting next-best-view in the crowd. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3583-3592 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.252, + 0.918, + 0.303 + ], + "angle": 0, + "content": "20. Drost, B., Ulrich, M., Bergmann, P., Hartinger, P., Steger, C.: Introducing mvtec itodd-a dataset for 3d object recognition in industry. In: Proceedings of the IEEE international conference on computer vision workshops. pp. 2200-2208 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.303, + 0.917, + 0.338 + ], + "angle": 0, + "content": "21. Gadwe, A., Ren, H.: Real-time 6 dof pose estimation of endoscopic instruments using printable markers. IEEE Sensors Journal 19(6), 2338-2346 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.338, + 0.918, + 0.39 + ], + "angle": 0, + "content": "22. Gao, X., Hou, X., Tang, J., Cheng, H.: Complete solution classification for the perspective-three-point problem. IEEE Trans. Pattern Anal. Mach. Intell. 25, 930-943 (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.918, + 0.442 + ], + "angle": 0, + "content": "23. Ghojogh, B., Ghodsi, A., Karray, F., Crowley, M.: Reproducing kernel hilbert space, mercer's theorem, eigenfunctions, nystr\\'' om method, and use of kernels in machine learning: Tutorial and survey. arXiv preprint arXiv:2106.08443 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.918, + 0.495 + ], + "angle": 0, + "content": "24. Ghorbani, B., Mei, S., Misiakiewicz, T., Montanari, A.: When do neural networks outperform kernel methods? Advances in Neural Information Processing Systems 33, 14820-14830 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.495, + 0.918, + 0.547 + ], + "angle": 0, + "content": "25. Greene, N., Luo, W., Kazanzides, P.: dvpose: Automated data collection and dataset for 6d pose estimation of robotic surgical instruments. In: 2023 International Symposium on Medical Robotics (ISMR). pp. 1-7. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.547, + 0.918, + 0.599 + ], + "angle": 0, + "content": "26. Gretton, A., Borgwardt, K., Rasch, M., Scholkopf, B., Smola, A.: A kernel method for the two-sample-problem. Advances in neural information processing systems 19 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.599, + 0.918, + 0.634 + ], + "angle": 0, + "content": "27. Gretton, A., Borgwardt, K.M., Rasch, M.J., Scholkopf, B., Smola, A.: A kernel two-sample test. The Journal of Machine Learning Research 13(1), 723-773 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.634, + 0.918, + 0.704 + ], + "angle": 0, + "content": "28. Guo, S., Hu, Y., Alvarez, J.M., Salzmann, M.: Knowledge distillation for 6d pose estimation by aligning distributions of local predictions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18633-18642 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.918, + 0.757 + ], + "angle": 0, + "content": "29. Hai, Y., Song, R., Li, J., Salzmann, M., Hu, Y.: Rigidity-aware detection for 6d object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8927-8936 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.757, + 0.918, + 0.826 + ], + "angle": 0, + "content": "30. Haugaard, R.L., Buch, A.G.: Surfemb: Dense and continuous correspondence distributions for object pose estimation with learnt surface embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6749-6758 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.826, + 0.918, + 0.879 + ], + "angle": 0, + "content": "31. He, Y., Huang, H., Fan, H., Chen, Q., Sun, J.: Ffb6d: A full flow bidirectional fusion network for 6d pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3003-3013 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.879, + 0.918, + 0.949 + ], + "angle": 0, + "content": "32. He, Y., Sun, W., Huang, H., Liu, J., Fan, H., Sun, J.: Pvn3d: A deep pointwise 3d keypoints voting network for 6 dof pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.057, + 0.92, + 0.949 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.728, + 0.016, + 0.842, + 0.032 + ], + "angle": 0, + "content": "RKHPose" + }, + { + "type": "page_number", + "bbox": [ + 0.894, + 0.017, + 0.92, + 0.031 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.057, + 0.92, + 0.111 + ], + "angle": 0, + "content": "33. He, Y., Wang, Y., Fan, H., Sun, J., Chen, Q.: Fs6d: Few-shot 6d pose estimation of novel objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6814-6824 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.113, + 0.92, + 0.183 + ], + "angle": 0, + "content": "34. Hinterstoisser, S., Lepetit, V., Ilic, S., Holzer, S., Bradski, G., Konolige, K., Navab, N.: Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In: Asian conference on computer vision. pp. 548-562. Springer (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.185, + 0.918, + 0.239 + ], + "angle": 0, + "content": "35. Hodañ, T., Haluza, P., Obdržálek, Š., Matas, J., Lourakis, M., Zabulis, X.: T-LESS: An RGB-D dataset for 6D pose estimation of texture-less objects. IEEE Winter Conference on Applications of Computer Vision (WACV) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.241, + 0.918, + 0.311 + ], + "angle": 0, + "content": "36. Hodan, T., Michel, F., Brachmann, E., Kehl, W., GlentBuch, A., Kraft, D., Drost, B., Vidal, J., Ihrke, S., Zabulis, X., et al.: Bop: Benchmark for 6d object pose estimation. In: Proceedings of the European conference on computer vision (ECCV). pp. 19-34 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.313, + 0.918, + 0.366 + ], + "angle": 0, + "content": "37. Hodač, T., Sundermeyer, M., Drost, B., Labbé, Y., Brachmann, E., Michel, F., Rother, C., Matas, J.: Bop challenge 2020 on 6d object localization. In: European Conference on Computer Vision. pp. 577-594. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.368, + 0.918, + 0.403 + ], + "angle": 0, + "content": "38. Horn, B.K., Hilden, H.M., Negahdaripour, S.: Closed-form solution of absolute orientation using orthonormal matrices. JOSA A 5(7), 1127-1135 (1988)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.405, + 0.918, + 0.457 + ], + "angle": 0, + "content": "39. Hu, Y., Fua, P., Salzmann, M.: Perspective flow aggregation for data-limited 6d object pose estimation. In: European Conference on Computer Vision. pp. 89-106. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.459, + 0.918, + 0.512 + ], + "angle": 0, + "content": "40. Kaskman, R., Zakharov, S., Shugurov, I., Ilic, S.: Homebreweddb: RGB-d dataset for 6d pose estimation of 3d objects. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. pp. 0-0 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.514, + 0.918, + 0.548 + ], + "angle": 0, + "content": "41. Khosravi, M., Smith, R.S.: The existence and uniqueness of solutions for kernel-based system identification. Automatica 148, 110728 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.551, + 0.918, + 0.603 + ], + "angle": 0, + "content": "42. Kleeberger, K., Huber, M.F.: Single shot 6d object pose estimation. In: 2020 IEEE International Conference on Robotics and Automation (ICRA). pp. 6239-6245. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.605, + 0.918, + 0.658 + ], + "angle": 0, + "content": "43. Kleeberger, K., Landgraf, C., Huber, M.F.: Large-scale 6d object pose estimation dataset for industrial bin-picking. In: 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 2573-2578. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.66, + 0.918, + 0.731 + ], + "angle": 0, + "content": "44. Kleeberger, K., Völk, M., Bormann, R., Huber, M.F.: Investigations on output parameterizations of neural networks for single shot 6d object pose estimation. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 13916-13922. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.732, + 0.918, + 0.803 + ], + "angle": 0, + "content": "45. Lee, T., Lee, B.U., Shin, I., Choe, J., Shin, U., Kweon, I.S., Yoon, K.J.: Udacope: unsupervised domain adaptation for category-level object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14891-14900 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.805, + 0.918, + 0.84 + ], + "angle": 0, + "content": "46. Lepetit, V., Moreno-Noguer, F., Fua, P.: Ep n p: An accurate o (n) solution to the p n p problem. International journal of computer vision 81, 155-166 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.842, + 0.918, + 0.894 + ], + "angle": 0, + "content": "47. Li, Z., Hu, Y., Salzmann, M., Ji, X.: Sd-pose: Semantic decomposition for cross-domain 6d object pose estimation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 2020–2028 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.896, + 0.918, + 0.949 + ], + "angle": 0, + "content": "48. Lin, J., Wei, Z., Ding, C., Jia, K.: Category-level 6d object pose and size estimation using self-supervised deep prior deformation networks. In: European Conference on Computer Vision. pp. 19-34. Springer (2022)" + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.057, + 0.92, + 0.949 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.106, + 0.032 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.409, + 0.034 + ], + "angle": 0, + "content": "Y.Wu and M. Greenspan" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.057, + 0.918, + 0.111 + ], + "angle": 0, + "content": "49. Liu, F., Xu, W., Lu, J., Zhang, G., Gretton, A., Sutherland, D.J.: Learning deep kernels for non-parametric two-sample tests. In: International conference on machine learning. pp. 6316-6326. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.112, + 0.918, + 0.164 + ], + "angle": 0, + "content": "50. Luo, Y.W., Ren, C.X.: Conditional bures metric for domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13989-13998 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.918, + 0.236 + ], + "angle": 0, + "content": "51. Ma, X., Wang, Z., Li, H., Zhang, P., Ouyang, W., Fan, X.: Accurate monocular 3d object detection via color-embedded 3d reconstruction for autonomous driving. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6851-6860 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.237, + 0.918, + 0.273 + ], + "angle": 0, + "content": "52. Mairal, J.: End-to-end kernel learning with supervised convolutional kernel networks. Advances in neural information processing systems 29 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.273, + 0.917, + 0.308 + ], + "angle": 0, + "content": "53. Mairal, J., Koniusz, P., Harchaoui, Z., Schmid, C.: Convolutional kernel networks. Advances in neural information processing systems 27 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.308, + 0.918, + 0.379 + ], + "angle": 0, + "content": "54. Manhardt, F., Arroyo, D.M., Rupprecht, C., Busam, B., Birdal, T., Navab, N., Tombari, F.: Explaining the ambiguity of object detection and 6d pose from visual data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.379, + 0.918, + 0.432 + ], + "angle": 0, + "content": "55. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.918, + 0.486 + ], + "angle": 0, + "content": "56. Misiakiewicz, T., Mei, S.: Learning with convolution and pooling operations in kernel methods. Advances in Neural Information Processing Systems 35, 29014-29025 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.486, + 0.918, + 0.539 + ], + "angle": 0, + "content": "57. Oberweger, M., Rad, M., Lepetit, V.: Making deep heatmaps robust to partial occlusions for 3d object pose estimation. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 119-134 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.918, + 0.575 + ], + "angle": 0, + "content": "58. Pan, S.J., Tsang, I.W., Kwok, J.T., Yang, Q.: Domain adaptation via transfer component analysis. IEEE transactions on neural networks 22(2), 199-210 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.918, + 0.646 + ], + "angle": 0, + "content": "59. Park, K., Mousavian, A., Xiang, Y., Fox, D.: Latentfusion: End-to-end differentiable reconstruction and rendering for unseen object pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10710-10719 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.918, + 0.699 + ], + "angle": 0, + "content": "60. Pearson, K.: Liii. on lines and planes of closest fit to systems of points in space. The London, Edinburgh, and Dublin philosophical magazine and journal of science 2(11), 559-572 (1901)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.7, + 0.918, + 0.753 + ], + "angle": 0, + "content": "61. Peng, S., Liu, Y., Huang, Q., Zhou, X., Bao, H.: Pynet: Pixel-wise voting network for 6dof pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4561-4570 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.754, + 0.918, + 0.807 + ], + "angle": 0, + "content": "62. Pinheiro, P.O., Rostamzadeh, N., Ahn, S.: Domain-adaptive single-view 3d reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7638-7647 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.807, + 0.918, + 0.86 + ], + "angle": 0, + "content": "63. Qi, C.R., Litany, O., He, K., Guibas, L.J.: Deep hough voting for 3d object detection in point clouds. In: proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9277-9286 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.861, + 0.918, + 0.913 + ], + "angle": 0, + "content": "64. Sejdinovic, D., Sriperumbudur, B., Gretton, A., Fukumizu, K.: Equivalence of distance-based and rkhs-based statistics in hypothesis testing. The annals of statistics pp. 2263-2291 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.914, + 0.918, + 0.949 + ], + "angle": 0, + "content": "65. Shan, P., Bi, Y., Li, Z., Wang, Q., He, Z., Zhao, Y., Peng, S.: Unsupervised model adaptation for multivariate calibration by domain adaptation-regularization based" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.057, + 0.918, + 0.949 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.728, + 0.016, + 0.842, + 0.032 + ], + "angle": 0, + "content": "RKHPose" + }, + { + "type": "page_number", + "bbox": [ + 0.894, + 0.017, + 0.92, + 0.031 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.119, + 0.057, + 0.918, + 0.093 + ], + "angle": 0, + "content": "kernel partial least square. Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 292, 122418 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.094, + 0.92, + 0.147 + ], + "angle": 0, + "content": "66. Shugurov, I., Li, F., Busam, B., Ilic, S.: Osop: A multi-stage one shot object pose estimation framework. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6835-6844 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.147, + 0.92, + 0.2 + ], + "angle": 0, + "content": "67. Shugurov, I., Zakharov, S., Ilic, S.: Dpodv2: Dense correspondence-based 6 dof pose estimation. IEEE transactions on pattern analysis and machine intelligence 44(11), 7417-7435 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.201, + 0.92, + 0.272 + ], + "angle": 0, + "content": "68. Sock, J., Garcia-Hernando, G., Armagan, A., Kim, T.K.: Introducing pose consistency and warp-alignment for self-supervised 6d object pose estimation in color images. In: 2020 International Conference on 3D Vision (3DV). pp. 291-300. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.272, + 0.918, + 0.325 + ], + "angle": 0, + "content": "69. Su, H., Qi, C.R., Li, Y., Guibas, L.J.: Render for cnn: Viewpoint estimation in images using cnns trained with rendered 3d model views. In: Proceedings of the IEEE international conference on computer vision. pp. 2686-2694 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.326, + 0.918, + 0.396 + ], + "angle": 0, + "content": "70. Su, Y., Saleh, M., Fetzer, T., Rambach, J., Navab, N., Busam, B., Stricker, D., Tombari, F.: Zebrapose: Coarse to fine surface encoding for 6dof object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6738-6748 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.397, + 0.918, + 0.468 + ], + "angle": 0, + "content": "71. Sundermeyer, M., Durner, M., Huang, E.Y., Marton, Z.C., Vaskevicius, N., Arras, K.O., Triebel, R.: Multi-path learning for object pose estimation across domains. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13916-13925 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.469, + 0.918, + 0.539 + ], + "angle": 0, + "content": "72. Sundermeyer, M., Hodač, T., Labbe, Y., Wang, G., Brachmann, E., Drost, B., Rother, C., Matas, J.: Bop challenge 2022 on detection, segmentation and pose estimation of specific rigid objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2784-2793 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.539, + 0.918, + 0.592 + ], + "angle": 0, + "content": "73. Sundermeyer, M., Marton, Z.C., Durner, M., Brucker, M., Triebel, R.: Implicit 3d orientation learning for 6d object detection from rgb images. In: Proceedings of the european conference on computer vision (ECCV). pp. 699-715 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.593, + 0.918, + 0.628 + ], + "angle": 0, + "content": "74. Szafraniec, F.H.: The reproducing kernel hilbert space and its multiplication operators. Complex Analysis and Related Topics pp. 253-263 (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.629, + 0.918, + 0.682 + ], + "angle": 0, + "content": "75. Tan, T., Dong, Q.: Smoc-net: Leveraging camera pose for self-supervised monocular object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21307-21316 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.682, + 0.918, + 0.718 + ], + "angle": 0, + "content": "76. Ullman, S.: The interpretation of structure from motion. Proceedings of the Royal Society of London. Series B. Biological Sciences 203(1153), 405-426 (1979)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.718, + 0.918, + 0.77 + ], + "angle": 0, + "content": "77. Venkateswara, H., Eusebio, J., Chakraborty, S., Panchanathan, S.: Deep hashing network for unsupervised domain adaptation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5018-5027 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.771, + 0.918, + 0.806 + ], + "angle": 0, + "content": "78. Vert, J.P., Tsuda, K., Scholkopf, B.: A primer on kernel methods. Kernel methods in computational biology 47, 35-70 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.806, + 0.918, + 0.86 + ], + "angle": 0, + "content": "79. Wang, G., Manhardt, F., Liu, X., Ji, X., Tombari, F.: Occlusion-aware self-supervised monocular 6d object pose estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.86, + 0.918, + 0.914 + ], + "angle": 0, + "content": "80. Wang, G., Manhardt, F., Shao, J., Ji, X., Navab, N., Tombari, F.: Self6d: Self-supervised monocular 6d object pose estimation. In: European Conference on Computer Vision. pp. 108-125. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.914, + 0.918, + 0.949 + ], + "angle": 0, + "content": "81. Wu, Y., Greenspan, M.: Learning better keypoints for multi-object 6dof pose estimation. arXiv preprint arXiv:2308.07827 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.057, + 0.92, + 0.949 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.079, + 0.017, + 0.106, + 0.031 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.409, + 0.034 + ], + "angle": 0, + "content": "Y.Wu and M. Greenspan" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.057, + 0.918, + 0.111 + ], + "angle": 0, + "content": "82. Wu, Y., Javaheri, A., Zand, M., Greenspan, M.: Keypoint cascade voting for point cloud based 6dof pose estimation. In: 2022 International Conference on 3D Vision (3DV). pp. 176-186. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.112, + 0.918, + 0.164 + ], + "angle": 0, + "content": "83. Wu, Y., Zand, M., Etemad, A., Greenspan, M.: Vote from the center: 6 dof pose estimation in rgb-d images by radial keypoint voting. In: European Conference on Computer Vision. pp. 335-352. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.918, + 0.2 + ], + "angle": 0, + "content": "84. Xiang, Y., Schmidt, T., Narayanan, V., Fox, D.: PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.201, + 0.918, + 0.271 + ], + "angle": 0, + "content": "85. Xiao, F., Liu, H., Lee, Y.J.: Identity from here, pose from there: Self-supervised disentanglement and generation of objects using unlabeled videos. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7013-7022 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.272, + 0.918, + 0.308 + ], + "angle": 0, + "content": "86. Xiao, Y., Qiu, X., Langlois, P.A., Aubry, M., Marlet, R.: Pose from shape: Deep pose estimation for arbitrary 3d objects. arXiv preprint arXiv:1906.05105 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.308, + 0.918, + 0.378 + ], + "angle": 0, + "content": "87. Yang, H., Pavone, M.: Object pose estimation with statistical guarantees: Conformal keypoint detection and geometric uncertainty propagation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8947-8958 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.379, + 0.918, + 0.432 + ], + "angle": 0, + "content": "88. Zhang, Y., David, P., Gong, B.: Curriculum domain adaptation for semantic segmentation of urban scenes. In: Proceedings of the IEEE international conference on computer vision. pp. 2020-2030 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.918, + 0.503 + ], + "angle": 0, + "content": "89. Zhang, Z., Wang, M., Huang, Y., Nehorai, A.: Aligning infinite-dimensional covariance matrices in reproducing kernel hilbert spaces for domain adaptation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3437-3445 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.918, + 0.575 + ], + "angle": 0, + "content": "90. Zhou, J., Chen, K., Xu, L., Dou, Q., Qin, J.: Deep fusion transformer network with weighted vector-wise keypoints voting for robust 6d object pose estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13967-13977 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.057, + 0.918, + 0.575 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Pseudo-keypoint RKHS Learning for Self-supervised 6DoF Pose Estimation/e3acf3fe-7127-4839-bf51-8bc132ae63e7_origin.pdf b/2024/Pseudo-keypoint RKHS Learning for Self-supervised 6DoF Pose Estimation/e3acf3fe-7127-4839-bf51-8bc132ae63e7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c280995630b4a1d16e8f53b7a71232fe6e3d3422 --- /dev/null +++ b/2024/Pseudo-keypoint RKHS Learning for Self-supervised 6DoF Pose Estimation/e3acf3fe-7127-4839-bf51-8bc132ae63e7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:87c1d6248e55e75d1086fd172b1a7a2690e4be5b38820af300cb66733b63779b +size 20385204 diff --git a/2024/Pseudo-keypoint RKHS Learning for Self-supervised 6DoF Pose Estimation/full.md b/2024/Pseudo-keypoint RKHS Learning for Self-supervised 6DoF Pose Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..564c99fa5dd49f78dd09125379afde73e9d257b0 --- /dev/null +++ b/2024/Pseudo-keypoint RKHS Learning for Self-supervised 6DoF Pose Estimation/full.md @@ -0,0 +1,359 @@ +# Pseudo-keypoint RKHS Learning for Self-supervised 6DoF Pose Estimation + +Yangzheng Wu and Michael Greenspan + +RCVLab, Dept. of Electrical and Computer Engineering, Ingenuity Labs, Queen's University, Kingston, Ontario, Canada {y.wu, michael.greenspan}@queensu.ca + +Abstract. We address the simulation-to-real domain gap in six degree-of-freedom pose estimation (6DoF PE), and propose a novel self-supervised keypoint voting-based 6DoF PE framework, effectively narrowing this gap using a learnable kernel in RKHS. We formulate this domain gap as a distance in high-dimensional feature space, distinct from previous iterative matching methods. We propose an adapter network, which is pre-trained on purely synthetic data with synthetic ground truth poses, and which evolves the network parameters from this source synthetic domain to the target real domain. Importantly, the real data training only uses pseudo-poses estimated by pseudo-keypoints, and thereby requires no real ground truth data annotations. Our proposed method is called RKHSPose, and achieves state-of-the-art performance among self-supervised methods on three commonly used 6DoF PE datasets including LINEMOD $(+4.2\%)$ , Occlusion LINEMOD $(+2\%)$ , and YCB-Video $(+3\%)$ . It also compares favorably to fully supervised methods on all six applicable BOP core datasets, achieving within $-11.3\%$ to $+0.2\%$ of the top fully supervised results. + +Keywords: pose estimation $\cdot$ self-supervision $\cdot$ domain adaptation $\cdot$ key-point estimation + +# 1 Introduction + +RGB-D Six Degree-of-Freedom Pose Estimation (6DoF PE) is a problem being actively explored in computer vision research. Given an RGB image and its associated depth map, the task is to detect scene objects and estimate their poses comprising 3DoF rotational angles and 3DoF translational offsets in the camera reference frame. This task enables many applications such as augmented reality [34,36,40,84], robotic bin picking [19,35,43], autonomous driving [51,85] and image-guided surgeries [21,25]. + +As with other machine learning (ML) tasks, fully supervised 6DoF PE requires large annotated datasets. This requirement is particularly challenging for 6DoF PE, as the annotations comprise not only the identity of the objects in the scene, but also their 6DoF pose, which makes the data relatively expensive to annotate compared to related tasks such as classification, detection and segmentation. This is due to the fact that humans are not able to qualitatively or + +![](images/6d7dd66c9b3040ba2b96e07a7a37270b2ed6f71f87197eeee34a608ef6ebd606.jpg) +Fig. 1: RKHSPose adapts the network pretrained on synthetic data to real test scenes (left), by comparing network feature spaces with real image inputs (solid arrows), against those with syn/real image (right) inputs (dashed arrows). $M_r$ regresses radial quantities, $M_A$ is the Adapter network, and RKHS maps features into a higher dimensional space. + +intuitively estimate 6DoF pose, which therefore requires additional instrumentation of the scene at data collection, and/or more sophisticated user annotation tools [34,84]. Consequently, synthetic 6DoF PE datasets [36,43] have been introduced, either as an additional complement to real datasets, or as standalone purely synthetic datasets. Annotated synthetic datasets are of course trivially inexpensive to collect, simply because precise synthetic pose annotation in a simulated environment is fully automatic. A known challenge in using synthetic data, however, is that there typically exists a domain gap between the real and synthetic data, which makes results less accurate when inferring in real data using models trained on purely synthetic datasets. Expectations of the potential benefit of synthetic datasets has led to the exploration of a rich set of Domain Adaptation (DA) methods, which specifically aim to reduce the domain gap [6,62,88] using inexpensive synthetic data for a wide variety of tasks, recently including 6DoF PE [28,45,47]. + +Early methods [42, 83] ignored the simulation-to-real (sim2real) domain gap and nevertheless improved performance by training on both synthetic and real annotated data, effectively augmenting the real images with the synthetic. However, these methods still required real labels to be sufficiently accurate and robust for practical applications, partially due to the domain gap. As shown in Fig. 1, the rendered synthetic objects (right image) have a slightly different appearance than the real objects (left image). The details of the CAD models, both geometry and texture, are not precise, as can be seen for the can object (which lacks a mouth and shadows), the benchwise (which is missing a handle), and the holepuncher (which has coarse geometric resolution). Several recent methods [9, 44, 73, 75, 80] have started to address the sim2real gap for 6DoF PE by first training on labeled synthetic data and then fine-tuning on unlabeled real data. Commonly known as self-supervised, these methods reduce the domain gap by adding extra supervision using features extracted from real images without requiring real ground truth (GT) labels. The majority of these methods are view- + +point/template matching-based, and the self-supervision commonly iteratively matches 2D masks or 3D point clouds [44,73,80]. + +While the above-mentioned self-supervision works have shown promise, there exist a wealth of DA techniques that can be brought to bear to improve performance further for this task. One such technique is Reproducing Kernel Hilbert Space (RKHS), which is a kernel method that has been shown to be effective for DA [1,5,41]. RKHS was initially used to create decision boundaries for non-separable data [12,60], and has been shown to be effective at reducing the domain gap for various tasks and applications [7,65,89]. The reproducing kernel guarantees that the domain gap can be statistically measured, allowing network parameters trained on synthetic data to be effectively adapted to the real data, using specifically tailored metrics. + +To address the sim2real domain gap in 6DoF PE, we propose RKHSPose, which is a keypoint-based method [82, 83] trained on a mixture of a large collection of labeled synthetic data, and a small handful of unlabeled real data. RKHSPose estimates the intermediate radial voting quantity, which has been shown to be effective for estimating keypoints [83], by first training a modified FCN-Resnet-18 on purely synthetic data, with automatically labeled synthetic GT poses. The radial quantity is a map of the distance from each image pixel to each keypoint. Next, real images are passed through the synthetically trained network, resulting in a set of pseudo-keypoints. The real images and their corresponding pseudo-keypoints are used to render a set of 'synthetic-overreal' (syn/real) images, by first estimating the pseudo-pose from the pseudo-keypoints, and then overlaying the synthetic object, rendered with the pseudopose, onto the real image. The network training then continues on the syn/real images, invoking an RKHS network module with a trainable linear product kernel, which minimizes the Maximum Mean Discrepancy (MMD) loss. At the front end, a proposed keypoint radial voting network learns to cast votes to estimate keypoints from the backend-generated radial maps. The final pose is then determined using ePnP [46] based on the estimated and corresponding object keypoints. + +The main contributions of this work are: + +- A novel learnable RKHS-based Adapter backend network architecture to minimize the sim2real domain gap in 6DoF PE; +- A novel CNN-based frontend network for keypoint radial voting; +- A self-supervised keypoint-based 6DoF PE method, RKHSPose, which is shown to have state-of-the-art (SOTA) performance, based on our experiments and several ablation studies. + +# 2 Related Work + +# 2.1 6DoF PE + +ML-based 6DoF PE methods [57,61,84] all train a network to regress quantities, such as keypoints and camera viewpoints, as have been used in classical algorithms [34]. ML-based methods, which initially became popular for the general + +Table 1: Existing self-supervised 6DoF PE methods. Some methods use CAD models, labeled synthetic (syn) data (images+poses), and real data without GT labels, while others use only labeled synthetic data. Many methods require a ROI such as a bounding box or a semantic mask, manually labeled or estimated by an existing framework. SO-Pose [18] used very few real labels to improve performance. + +
methodmodeCAD modelsyn datareal imagesreal posesROI
SO-Pose [18]RGB✓✓✓✓✓
TexPose [9]RGB✓✓✓X✓
SMOC-Net [75]RGB✓✓✓X✓
FS6D [33]RGBD✓✓XX✓
AAE [73]RGB✓✓XX✓
MHP [54]RGB✓✓XX✓
Sock et al. [68]RGB✓✓XX✓
DSC [86]RGBD✓✓XX✓
Sundermeyer [71]RGB✓✓XX✓
LatentFusion [59]RGBX✓XX✓
OSOP [66]RGBD✓✓XXX
Kleeberger et al. [44]D✓✓XXX
Su et al. [69]RGB✓✓✓XX
Self6D [80]RGBD✓✓✓XX
Self6D++ [79]RGB✓✓✓XX
Deng et al. [15]RGBD✓✓✓XX
RKHSPose (Ours)RGBD✓✓✓XX
+ +object detection task, have started to dominate the 6DoF PE literature due to their accuracy and efficiency. + +There are two main categories of ML-based fully supervised methods: feature matching-based [29, 30, 57, 70, 84], and keypoint-based methods [31, 32, 61, 87]. Feature matching-based methods make use of the structures from general object detection networks most directly. The network encodes and matches features and estimates pose by either regressing elements of the pose (e.g. the transformation matrix [75], rotational angles and translational offsets [84] or 3D vertices [70]) directly, or by regressing some intermediate feature-matching representations, such as viewpoints or segments. + +In contrast, keypoint-based methods encode features to estimate keypoints which are predefined within the reference frame of an object's CAD model. These methods then use (modified) classical algorithms such as PnP [22, 46], Horn's method [38], and ICP [3] to estimate the final pose from corresponding image and model keypoints. Unlike feature-matching methods, keypoint-based methods are typically more accurate due to redundancies encountered through voting schemes [32, 61, 83] and by generating confidence hypotheses of keypoints [31, 87]. + +Recently, self-supervised 6DoF PE methods have been explored in order to reduce the reliance on labeled real data, which is expensive to acquire. As sum- + +marized in Table 1, these methods commonly use real images without GT labels. Some methods use pure synthetic data and CAD models only, with the exception of LatentFusion [59] which trained the model using only synthetic data. The majority of these methods [9, 15, 33, 48, 66, 68, 69, 71, 75, 79, 80] are inspired by fully supervised feature-matching methods, except DPODv2 [67] and DSC [86], in which keypoint correspondences are rendered and matched. + +A few methods [15,66] fine-tuned the pose trivially by iteratively matching the template/viewpoint, whereas others [33,44,59,86] augmented the training data by adding noise [44,86], rendering textures [33,67] and creating a latent space [59]. Some methods [54,71,73] also implemented DA techniques such as codebook encoding [71], Principle Component Analysis (PCA) [71] and symmetric Bingham distributions [54]. Most methods [66,68,69,79,80,86] used rendering techniques to render and match a template. There are a few methods that combined 3D reconstruction techniques, such as Neural radiance fields (Nerf) [55] and Structure from Motion (SfM) [76]. TexPose [9] matched CAD models to segments generated by Nerf, and SMOC-Net [75] used SfM to create the 3D segment and matched with the CAD model. + +# 2.2 Kernel Methods and Deep Learning + +While deep learning is the most common ML technique within the computer vision literature, kernel methods [12,60,64] have also been actively explored. Kernel methods are typically in RKHS space [74] with reproducing properties that facilitate solving non-linear problems by mapping input data into high dimensional spaces that can be linearly separated [11,23]. Well-known early kernel methods that have been applied to computer vision are Support Vector Machines [12] and PCA [60]. A recent method [64] linked energy distance and MMD in RKHS, and showed the effectiveness of kernels in statistical hypothesis testing. + +More recent studies compare kernel methods with deep learning networks [1, 4,5,41]. RKHS is found to perform better on classification tasks than one single block of a CNN comprising convolution, pooling and downsampling (linear) layers [41]. RKHS can also help with CNN generalization by meta-regularization on image registration problems [1]. Similarly, norms (magnitude of trainable weights) defined in RKHS help with CNN regularization [4,5]. Further, discriminant information in label distributions in unsupervised DA is addressed and RKHS-based Conditional Kernel Bures metric is proposed [50]. Lastly, the connection between Neural Tangent Kernel and MMD is established, and an efficient MMD for two-sample tests is developed [10]. + +Inspired by the previous work, RKHSPose applies concepts of kernel learning to keypoint-based 6DoF PE, to provide an effective means to self-supervise a synthetically trained network on unlabeled real data. + +![](images/fe1acf770a9f08cf67b785bce1bf8b9cfa652b18d5242f45ceffc0883be82151.jpg) +Fig. 2: RKHSPose architecture. RKHPose is first trained on synthetic labeled data (solid arrows), and then finetuned on alternating syn/real and (unlabeled) real images (dashed arrows). $M_A$ is measured by MMD in RKHS by densely mapping the intermediate features of $M_r$ into high dimensional spaces with conv blocks. The distance is treated as $\mathcal{L}_{M_A}$ and back-propagated through $M_A$ and $M_r$ . + +![](images/b39f96337e2343f7784c0e00651d956680d5ccf78ea772c2927bf9a92dd30c72.jpg) + +![](images/dd85496e0dc56714813d5722c57ab3aebe87c652c635f9995a8cb599294accce.jpg) +$K_{syn}$ +$K_{syn / real}$ + +![](images/f704538c2f497cab3eea1ec82dc383a7e159ba121a1acb97190844cf6e4a056e.jpg) +$K_{real}$ + +# 3 Method + +# 3.1 Network Overview + +As shown in Fig. 2, RKHSPose is made up of two networks, a main network $M_{rv}$ for keypoint regression and classification, and an Adapter network $M_A$ . The input of $M_{rv}$ with shape $W \times H \times 4$ , is the concatenation of an RGB image $I$ and its corresponding depth map $D$ . This input can be synthetically generated with an arbitrary background ( $I_{syn}$ and $D_{syn}$ ), a synthetic mask overlayed on a real background ( $I_{syn/real}$ and $D_{syn/real}$ ), or a real ( $I_{real}$ and $D_{real}$ ) image. The outputs are $n$ projected 2D keypoints $K$ , along with corresponding classification labels $C$ and confidence scores $S$ . $K$ are organized into instance sets based on $C$ and geometric constraints. + +$M_{rv}$ comprises two sub-networks, regression network $M_r$ and voting network $M_v$ . Inspired by recent voting techniques [32, 61, 82, 83, 90], $M_r$ estimates an intermediate voting quantity, which is a radial distance map $V_r$ [83], by using a modified Fully Connected ResNet 18 (FCN-ResNet-18). The radial voting map $V_r$ , with shape $W \times H$ , stores the Euclidean distance from each object point to each keypoint in the 3D camera world reference frame. The voting network $M_v$ (described in Sec. 3.3) then takes $V_r$ as input, accumulates votes, and detects peaks to estimate $K$ , $C$ and $S$ . + +The Adapter network $M_A$ consists of a series of CNNs which encode pairs of feature maps from $M_{rv}$ , and are trained on both synthetic overlayed and pure real + +data. $M_A$ encodes feature map pairs $(f_{syn / real}, f_{real})$ into corresponding high-dimensional feature maps $(h_{syn / real}, h_{real})$ . The input data, $(I_{syn / real}, I_{real})$ and $(D_{syn / real}, D_{real})$ , are also treated as $(f_{syn / real}, f_{real})$ during the learning of $M_A$ . Each of these networks creates a high-dimensional latent space, essentially the Reproducing Kernel Hilbert Space (RKHS) [2] and contributes to the learning of both $M_{rv}$ and $M_A$ by calculating the MMD [27]. While RKHS has been applied effectively to other DA tasks, to our knowledge, this is the first time that it has been applied to 6DoF PE, and the adapter network architecture is novel. + +The loss function $\mathcal{L}$ is made up of five elements: Radial regression loss $\mathcal{L}_r$ for $V_r$ ; Keypoint projection loss $\mathcal{L}_k$ for $K$ ; Classification loss $\mathcal{L}_c$ for $C$ ; Confidence loss $\mathcal{L}_s$ for $S$ , and finally; Adapter loss $\mathcal{L}_A$ for the comparison of intermediate feature maps. The regression losses $\mathcal{L}_r$ , $\mathcal{L}_k$ , and $\mathcal{L}_s$ all use the smooth L1 metric, whereas classification loss $\mathcal{L}_c$ uses the cross-entropy metric $H(\cdot)$ . RKHSPose losses can then be denoted as: + +$$ +\mathcal {L} _ {r} = \operatorname {s m o o t h} _ {\mathcal {L} _ {1}} \left(V _ {r}, \widehat {V} _ {r}\right) \tag {1} +$$ + +$$ +\mathcal {L} _ {k} = \operatorname {s m o o t h} _ {\mathcal {L} _ {1}} (K, \widehat {K}) \tag {2} +$$ + +$$ +\mathcal {L} _ {c} = H (p (C), \widehat {p} (C)) \tag {3} +$$ + +$$ +\mathcal {L} _ {s} = \operatorname {s m o o t h} _ {\mathcal {L} _ {1}} (S, \widehat {S}) \tag {4} +$$ + +$$ +\mathcal {L} _ {A} = M M D \left(\widehat {f} _ {\text {s y n / r e a l}}, \widehat {f} _ {\text {r e a l}}\right) \tag {5} +$$ + +$$ +\mathcal {L} = \lambda_ {r} \mathcal {L} _ {r} + \lambda_ {k} \mathcal {L} _ {k} + \lambda_ {c} \mathcal {L} _ {c} + \lambda_ {s} \mathcal {L} _ {s} + \lambda_ {A} \mathcal {L} _ {A} \tag {6} +$$ + +where $\lambda_r, \lambda_k, \lambda_c, \lambda_s$ , and $\lambda_A$ are weights for adjustment during training, and all non-hatted quantities are GT values. At inference, $M_{rv}$ takes $I_{real}$ and $D_{real}$ as input, and outputs $K, C$ and $S$ . The keypoints $K$ are ranked and grouped by $S$ and $C$ , and are then forwarded into the ePnP [46] algorithm which estimates 6DoF pose values. ICP can then be optionally applied using the depth data to refine the estimated pose. + +# 3.2 Convolutional RKHS Adapter + +Reproducing Kernel Hilbert Space $\mathcal{H}$ is a commonly used vector space for Domain Adaptation [58, 77]. Hilbert Space is a complete metric space (in which every Cauchy sequence of points has a limit within the metric) represented by the inner product of vectors. For a non-empty set of data $\mathcal{X}$ , a function $\mathcal{K}_{\mathcal{X}}: \mathcal{X} \times \mathcal{X} \to \mathcal{R}$ is a reproducing kernel if: + +$$ +\left\{ \begin{array}{l} k (\cdot , x) \in \mathcal {H} \quad \forall \quad x \in \mathcal {X} \\ \langle f (\cdot), k (\cdot , x) \rangle = f (x) \quad \forall \quad x \in \mathcal {X}, f \in \mathcal {H} \end{array} \right. \tag {7} +$$ + +where $\langle a,b\rangle$ denotes the inner product of two vectors $a$ and $b$ , $k(\cdot ,x) = \mathcal{K}_{\mathcal{X}}$ for each $x\in \mathcal{X}$ , and $f$ is a function in $\mathcal{H}$ . The second equation in Eq. 7 is an expression of the reproducing property of $\mathcal{H}$ . + +In order to utilize $\mathcal{H}$ for DA, we expand the kernel definition to two sets of data $\mathcal{X}$ and $\mathcal{Y}$ . The reproducing kernel can then be defined as: + +$$ +\mathcal {K} (\mathcal {X}, \mathcal {Y}) = \left\langle \mathcal {K} _ {\mathcal {X}}, \mathcal {K} _ {\mathcal {Y}} \right\rangle_ {H} \tag {8} +$$ + +Here, $\mathcal{K}_{\mathcal{X}}$ and $\mathcal{K}_{\mathcal{Y}}$ are themselves inner product kernels that map $\mathcal{X}$ and $\mathcal{Y}$ respectively into their own Hilbert spaces, and $\mathcal{K}(\mathcal{X},\mathcal{Y})$ is the joint Hilbert space kernel of $\mathcal{X}$ and $\mathcal{Y}$ . Note that $\mathcal{K}(\mathcal{X},\mathcal{Y})$ is not the kernel commonly defined for CNNs. Rather, it is the similarity function defined in kernel methods, such as is used by Support Vector Machine (SVM) techniques to calculate similarity measurements. Some recent methods [8, 52, 53, 56] named it a Convolutional Kernel Network (CKN) to distinguish it from CNNs. + +Various kernels of CKN, such as the Gaussian Kernel [53], the RBF Kernel [52] and the Inner Product Kernel [56], have been shown to be comparable to shallow CNNs for various tasks, especially for Domain Adaption. The most intuitive inner product kernel is mathematically similar to a fully connected layer, where trainable weights are multiplied by the input feature map. To allow the application of RKHS methods into our CNN based Adapter network $M_A$ , trainable weights are added to $\mathcal{K}(\mathcal{X},\mathcal{Y})$ [49]. The trainable Kernel $\mathcal{K}_w(\mathcal{X},\mathcal{Y})$ can then be denoted as: + +$$ +\mathcal {K} _ {w} (\mathcal {X}, \mathcal {Y}) = \langle \langle \mathcal {X}, W _ {\mathcal {X}} \rangle , \langle \mathcal {Y}, W _ {\mathcal {Y}} \rangle \rangle_ {H} \tag {9} +$$ + +where $W_{\mathcal{X}}$ and $W_{\mathcal{Y}}$ are trainable weights. By adding $W_{\mathcal{X}}$ and $W_{\mathcal{Y}}$ , $K_{w}(\mathcal{X},\mathcal{Y})$ still satisfies the RKHS constraints. + +The sim2real domain gap of feature maps $f$ being trained in $M_{rv}$ (with few real images and no real GT labels) are hard to measure using trivial distance metrics. In contrast, RKHS can be a more accurate and robust space for comparison, since it is known to be capable of handling high-dimensional data with a low number of samples [24, 56]. To compare $(f_{syn / real}, f_{real})$ in RKHS, a series of CNN layers encodes $(f_{syn / real}, f_{real})$ into higher-dimensional features $(h_{syn / real}, h_{real})$ , followed by the trainable $\mathcal{K}_w(\mathcal{X}, \mathcal{Y})$ . Once mapped into RKHS, $h_{syn / real} = \{sr_i\}_{i=1}^m$ and $h_{real} = \{r_i\}_{i=1}^m$ can then be measured by Maximum Mean Discrepancy (MMD), a common DA measurement [27, 49], which is the square distance between the kernel embedding [26]: + +$$ +\begin{array}{l} M M D \left(h _ {s y n / r e a l}, h _ {r e a l}\right) = \frac {1}{m} \left[ \left(\sum_ {i = 1} ^ {m} \sum_ {j = 1} ^ {m} k _ {w} \left(s r _ {i}, s r _ {j}\right) - \sum_ {i = 1} ^ {m} k _ {w} \left(s r _ {i}, s r _ {i}\right)\right) \right. \\ \left. - \left(\sum_ {i = 1} ^ {m} \sum_ {j = 1} ^ {m} k _ {w} \left(s r _ {i}, r _ {j}\right) - \sum_ {i = 1} ^ {m} k _ {w} \left(s r _ {i}, r _ {i}\right)\right) + \left(\sum_ {i = 1} ^ {m} \sum_ {j = 1} ^ {m} k _ {w} \left(r _ {i}, r _ {j}\right) - \sum_ {i = 1} ^ {m} k _ {w} \left(r _ {i}, r _ {i}\right)\right) \right] ^ {\frac {1}{2}} \tag {10} \\ \end{array} +$$ + +where $k_{w}()$ is the feature element of $\mathcal{K}_w($ + +In summary, the Adapter $M_A$ shown in Fig. 2 measures MMD for each $(h_{syn / real}, h_{real})$ in RKHS, by increasing the $f_{syn / real}$ and $f_{real}$ dimension using a CNN and thereby constructing a learnable kernel $\mathcal{K}_w$ . The outputs are the feature maps $h_{syn / real}$ and $h_{real}$ , which are supervised by loss $\mathcal{L}_A$ during training of the real data epochs. Based on the experiments in Sec. 5.3, our trainable inner product kernel is shown to be more accurate for our task than other known kernels [52, 53, 56] that we tested, that are often used in such kernel methods. + +# 3.3 Keypoint Radial Voting Network + +The network $M_v$ votes for keypoints using a CNN architecture, taking the radial voting quantity $V_r$ resulting from $M_r$ as input. VoteNet [63] previously used a CNN approach to vote for object centers, whereas other keypoint-based techniques have implemented GPU-based parallel RANSAC [32, 61] methods for offset and vector quantities. The radial quantity is known to be more accurate than the vector or offset quantities, and has been previously implemented with a CPU-based parallel accumulator space method [82, 83]. Given its superior accuracy, $M_v$ implements radial voting using a CNN to improve efficiency. Given a 2D radial map $\widehat{V}_r$ estimated by $M_r$ and supervised by GT radial maps $V_r$ , the task is to accumulate votes, find the peak, and estimate the keypoint location. The $V_r$ foreground pixels (which lie on the target object) store the Euclidean distance from these pixels to each of the keypoints, with background (non-object) pixels set to value -1. + +The estimated radial map $\widehat{V}_r$ is indeed an inverse heat map of the candidate keypoints' locations, distributed in a radial pattern centered at the keypoints. To forward $\widehat{V}_r$ into a CNN voting module, it is inversely normalized so that it becomes a heat map. Let $v_r^{max}$ and $v_r^{min}$ be the maximum and minimum global radial distances for all objects in a dataset, which can be calculated by iterating through all GT radial maps, or alternately generated from the object CAD models. An inverse radial map $\widehat{V}_r^{-1}$ can then be denoted as $\widehat{V}_r^{-1} = (v_r^{max} - \widehat{V}_r) / (v_r^{max} - v_r^{min})$ . Voting network $M_v$ takes $\widehat{V}_r^{-1}$ as the input and generates the accumulated vote map by a series of convolution, ReLu, and batch normalization layers. The complete network architecture is provided in the Supplementary material Sec. S.2. The background pixels are filtered out by a ReLu layer, and only foreground pixels contribute to voting. The accumulated vote map is then max-pooled for peak extraction, and reshaped using a fully connected layer into a $n\times 4$ output. The output represents $n$ keypoints and comprises $n\times 2$ projected 2D keypoints $K$ , $n$ classification labels $C$ , and $n$ confidence scores $S$ . The labels $C$ indicate which object the corresponding keypoints belongs to, and $S$ ranks the confidence level of keypoints before being forwarded into ePnP [46] for pose estimation. $M_v$ is supervised by $\mathcal{L}_k$ , $\mathcal{L}_c$ , and $\mathcal{L}_s$ (Eqs. 2-4) and is trained end-to-end along with $M_r$ . + +# 4 Experiments + +# 4.1 Datasets and Evaluation Metrics + +RKHSPose uses BOP Procedural Blender [17] (PBR) synthetic images [16, 36, 37, 72] for the synthetic training phase. The images are generated by dropping synthetic objects (CAD models) onto a plane in a simulated environment using PyBullet [13], and then rendering them with synthetic textures. All objects in the synthetic images are thus automatically labeled with precise GT poses. We evaluated RKHSPose for the six BOP [37, 72] core datasets (LMO [34], YCB [84], TLESS [35], TUDL [36], ITODO [20], and HB [40]), all except IC-BIN [19], + +which does not include any real training or validation images and is therefore not applicable. ITODO and HB have no real images in the training set, and so for training we instead used the real images in their validation sets, which were disjoint from their test sets. + +Our main results are evaluated with the ADD(S) [34] metric for the LM and LMO dataset, and the ADD(S) AUC [84] metric for the YCB dataset. These are the standard metrics commonly used to compare self-supervised 6DoF PE methods. ADD(S) is based on the mean distance (minimum distance for symmetry) between the object surfaces for GT and estimated poses, whereas ADD(S) AUC plots a curve formed by ADD(S) for various object diameter thresholds. We use the BOP average recall ( $AR$ ) metrics for our ablation studies. The $AR$ metric, based on the original ADD(S) [34], evaluates three aspects, including Visible Surface Discrepancy ( $AR_{VSD}$ ), Maximum Symmetry-Aware Surface/Projection Distance ( $AR_{MSSD}$ and $AR_{MSPD}$ ) [36]. + +# 4.2 Implementation Details + +RKHSPose is trained on a server with an Intel Xeon 5218 CPU and two RTX6000 GPUs with a batch size of 32. The Adam optimizer is used for the training of $M_{rv}$ , on both synthetic and real data, and $M_A$ is optimized by SGD. Both of the optimizers have an initial learning rate of $lr = 1e - 3$ and weight decay 1e-4 for 80 and 20 epochs respectively. + +The input of the network is normalized before training, as follows. The RGB images $I$ are normalized and standardized using ImageNet [14] specifications. The depth maps are each individually normalized by their local minima and maxima, to lie within a range of 0 to 1. The radial distances in radial map $V_{r}$ and 2D projected keypoints are both normalized by the width and height of $I$ . + +A single set of four keypoints is chosen by KeyGNet [81] for the set of all objects in each dataset. One extra background class is added to $C$ in order to filter out the redundant background points in $K$ . $M_{rv}$ is first trained for 120 epochs on synthetic data, during which $M_A$ remains frozen. Following this, training proceeds for an additional 80 epochs which alternate between real and synthetic data. When training on real data, both $M_A$ and $M_{rv}$ weights are learned, whereas $M_A$ is frozen for the alternating synthetic data training. + +During real data epochs, $M_{rv}$ initially estimates pseudo-keypoints for each real image. These pseudo-keypoints are then forwarded into ePnP for pseudopose estimation. Each estimated pseudo-pose is then augmented into a set of poses $P_{aug}$ by applying arbitrary rotational and translational perturbations with respective ranges of $[-\frac{\pi}{18}, \frac{\pi}{18}]$ radians and $[-0.1, 0.1]$ along three axes within the normalized model frame, which is defined using the largest object in the dataset. The set of syn/real images are rendered by overlaying onto the real image the CAD model of each object using each augmented pose value in $P_{aug}$ . The cardinality of $P_{aug}$ is set to be one less than the batch size, and the adapter $M_A$ is trained on a mini-batch of the hybrid images resulting from $P_{aug}$ , plus the image resulting from the original estimated pseudo-pose. + +Table 2: Comparison with other methods. Accuracy of RKHSPose for LM and LMO is evaluated with ADD(S), and for YCB is evaluated with ADD(S) AUC. All 'Supervision: Syn + Self' methods use real images without real labels. + +
MethodReal data image labelDataset/Metric
LM ADD(S)LMOYCB ADD(S)ADD-S AUC
Supervision: Syn (lower bound)
AAEXX31.4---
MHPXX38.8---
GDR (TexPose version)XX77.452.9--
Self6D++XX77.452.977.889.4
Self6D++ with DrefXX88.062.579.290.1
OursXX78.254.376.590.2
Ours+ICPXX87.955.778.391.3
Supervision: Syn + Self
Sock et al.✓X60.622.8--
DSC✓X58.624.8--
Self6D✓X58.932.1--
SMOC-Net✓X91.363.3--
Self6D++✓X88.564.780.091.4
TexPose✓X91.766.7--
Ours✓X95.868.682.892.4
Ours+ICP✓X95.968.783.092.6
Supervision: Syn + Real GT (upper bound)
SO-Pose✓✓96.062.383.990.9
Self6D++✓✓91.074.482.690.7
Ours✓✓96.770.885.492.2
Ours+ICP✓✓96.871.385.692.4
+ +Initially $M_A$ is frozen, and for the first 80 epochs, the loss is set to emphasize the classification and the visibility score regression, i.e. $\lambda_c = \lambda_s = 0.6$ and $\lambda_r = \lambda_k = 0.4$ . Following this, up to epoch 200, the scales of losses are then exchanged to fine-tune the localization of the keypoints, i.e. $\lambda_c = \lambda_s = 0.4$ and $\lambda_r = \lambda_k = 0.6$ . After epoch 120, $M_A$ is unfrozen each alternating epoch, and $\lambda_D$ is set to 1 during the remaining $M_A$ training epochs. This training strategy, shown in Fig. 2, minimizes the sim2real gap without any real image GT labels, and using very few (320) real images. + +# 4.3 Results + +The results are summarized in Tables 2 and 3. To our knowledge, RKHSPose outperforms all existing self-supervised 6DoF PE methods. In Table 2, the upper bound fully supervised, lower bound synthetically supervised, and middle self-supervised methods (including ours) are compared. On LM and LMO, our ADD(S) is $+4.2\%$ and $+2\%$ better than the second best method TexPose [9]. We + +Table 3: Comparison with fully supervised methods. RKHSPose results on TLESS $(-1.8)$ , TUDL $(-0.4)$ , ITODD $(-4.6)$ and HB $(+0.1)$ compares to SOTA methods with full supervision of real GT labels. Methods annotated with * use the detection results from other detection methods. + +
Methodreal labelLMOTLESSDataset
TUDLITODDHBYCB
SurfEmb* [30]✓-76.082.885.465.986.679.9
RCVPose3D [82]✓-72.970.896.673.386.384.3
RADet [87]+PFA* [39]✓-79.785.096.067.686.988.8
ZebraPose [70]✓-78.086.295.665.492.189.9
Ours✗95.768.285.596.268.692.283.6
Ours+ICP✗95.868.485.696.268.792.383.8
+ +![](images/12f1a860821259363b9ce9e8fa2e3d636edb840ef5654196f8ff86ed9ae4784b.jpg) +Fig. 3: Qualitative overlay results on selected images. Red dots and blue dots are projected surface points from GT poses and estimated poses, respectively. + +![](images/c8e452a90ee70b8162806bcab95a05890b7ae4d8358c542d23f3e244ab86431b.jpg) + +![](images/243dda54444d5ca87002df56e5e61697d016711f6e09f3e84d9a5c175c79b206.jpg) + +![](images/222bb3aa03cc0d98fdda13e3782932a1008bbceab48df717e16cfb32bb20e597.jpg) + +![](images/6c80e3d267335e01caa2245aa89098bd693bb1be49d82798c7582dae71fc694e.jpg) + +![](images/9256f29dad8d7bd713e1dd1603d26852183ecbdc77afac167ec1918370b65e59.jpg) + +compared our performance to Self6D++ [79], which was the only other method that evaluated using YCB, and saw a $3\%$ improvement after ICP on ADD(S) AUC. Last but not least, in Table 3 our performance evaluated on the other four BOP core datasets is comparable to several upper bound methods from the BOP leaderboard. Some test scenes with RKHSPose results are shown in Fig. 3. + +RKHSPose runs at 34 fps on an Intel i7 2.5GHz CPU and an RTX 3090 GPU with 24G VRAM. It takes on average $8.7\mathrm{ms}$ for loading data, $4.5\mathrm{ms}$ for forward inference through $M_{rv}$ ( $\times 10$ faster compared to the analytical radial voting in RCVPose [83]), and $16.2\mathrm{ms}$ for ePnP. + +# 5 Ablation Studies + +# 5.1 Dense Vs. Sparse Adapter + +The Adapter $M_A$ densely matches the intermediate feature maps, whereas the majority of other methods [15,75,80] only compare the final output. To show the benefits of dense comparison, we conduct an experiment with different variations of $M_A$ . A sparse matching $M_A^s$ network is trained on synthetic and real data comparing only a single feature map, which is the intermediate radial map. $M_A^s$ has the exact same overall learning capacity (number of parameters) as the dense matching $M_A$ described in Sec. 3.3. The results in Table 4 show that $M_A$ + +Table 4: ${AR}$ of different adapters on LM and five BOP core datasets. + +
AdapterARVSDARMSSDARMSPDAR
MA78.177.877.877.9
MA84.984.184.384.4
+ +Table 5: ${AR}$ of different training strategies on LM and five BOP core datasets. + +
Training TypeARVSDARMSSDARMSPDAR
Mixed84.984.184.384.4
Sequential82.381.781.781.9
+ +Table 6: [R Tab A] $AR$ of different kernel- Table 7: $AR$ of different metrics on LM nets on LM and five BOP core datasets. and five BOP core datasets. + +
KernelwARVSDARMSSDARMSPDAR
LinearX71.670.870.671.0
✓84.984.184.384.4
RBFX73.472.973.273.2
✓82.581.381.581.6
+ +
MetricARVSDARMSSDARMSPDAR
MMD84.984.184.384.4
KL Div78.077.878.077.9
Wass80.980.680.980.8
+ +surpassed $M_A^s$ on all six datasets tested. Specifically, on ITODD, $M_A$ is $12.1\%$ more accurate than $M_A^s$ . This experiment shows the effectiveness of our densely matched $M_A$ . + +# 5.2 Syn/Real Synchronized Training + +When training RKHSPose, real epochs are alternated with synthetic epochs. In contrast, some other methods [9, 75, 79, 80] separate the synthetic/real training. We conducted an experiment to compare these two different training strategies, the results of which are shown in Table 5. The alternating training performs slightly better ( $+2.5\%$ on average) than the sequential training, possibly due to the early access to real scenes thereby avoiding local minima. + +# 5.3 Adapter Kernels and Metrics + +We use a linear (dot product) kernel and MMD in RKHS for domain gap measurements. There are various other kernels and similarity measurements that can be implemented in RKHS as described in Sec. 3.2. First, we add trainable weights to the radial basis function (RBF) kernel in a similar manner as $K_{w}$ defined in Eq. 9. The trainable RBF kernel on two sets of data $X$ and $Y$ is denoted as: + +$$ +K _ {r b f} (X, Y) = \exp (- w \| X - Y \| ^ {2}) \tag {11} +$$ + +where $w$ are the trainable weights, which replaces the original adjustable parameter in the classical RBF kernel. We also experiment with the classical RKHS kernel functions without trainable weights, including the inner product kernel and the original RBF kernel [78], for comparison. Further, we experiment on other commonly used distance measures, including Kullback-Leibler Divergence (KL Div, i.e. relative-entropy) and Wasserstein (Wass) Distance. + +![](images/85734d7328823939c2c8a4cace4cb00d1ae19d2781561399992531721a03abc4.jpg) +Fig. 4: Impact of # of real images with/without GT labels used during training. All datasets are evaluated by the BOP AR metric. We conduct experiments from 0 to 640 real images on all datasets, except ITODD which contained only 357 real images. + +In Table 6, the RBF kernel performs similar to the linear product kernel with a slight performance dip. In Table 7, MMD minimizes the domain gap better than the Wass Distance, followed by the KL Div metric, leading to a better overall performance on $AR$ . Based on these results, we used the linear product kernel with trainable weights and selected MMD as the main loss metric. + +# 5.4 Number of Real Images and Real Labels + +The objective of RKHSPose is to reduce real data usage and train without any real GT labels. To show the effectiveness of the approach, we conducted an experiment by training on different numbers of real images, the results of which are shown in Fig. 4. We used up to 640 real images in all cases, except for that of ITODD which contains only 357 real images. The $AR$ of all datasets saturates at 160 images except YCB. The further improvement of YCB beyond 160 images is also only $+0.1\%$ and saturates after 320 real images. We nevertheless use 320 real unlabeled images for our main results. This experiment showed that adding more than 320 real labeled images did not significantly improve performance. + +# 6 Conclusion + +To sum up, we propose a novel self-supervised keypoint radial voting-based 6DoF PE method using RGB-D data called RKHSPose. RKHSPose fine-tunes poses pre-trained on synthetic data by densely matching features with a learnable kernel in RKHS, using real data albeit without any real GT poses. By applying this DA technique in feature space, RKHSPose achieved SOTA performance on the six applicable BOP core datasets, surpassing the performance of all other self-supervised methods. Notably, the RKHSPose performance closely approaches that of several fully-supervised methods, which indicates the strength of the approach at reducing the sim2real domain gap for this problem. + +Acknowledgements: Thanks to Bluewrist Inc. and NSERC for their support of this work. + +# References + +1. Al Safadi, E., Song, X.: Learning-based image registration with meta-regularization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10928-10937 (2021) +2. Aronszajn, N.: Theory of reproducing kernels. Transactions of the American mathematical society 68(3), 337-404 (1950) +3. Besl, P.J., McKay, N.D.: Method for registration of 3-d shapes. In: Sensor fusion IV: control paradigms and data structures. vol. 1611, pp. 586-606. International Society for Optics and Photonics (1992) +4. Bietti, A., Mairal, J.: Group invariance, stability to deformations, and complexity of deep convolutional representations. The Journal of Machine Learning Research 20(1), 876-924 (2019) +5. Bietti, A., Mialon, G., Chen, D., Mairal, J.: A kernel perspective for regularizing deep neural networks. In: International Conference on Machine Learning. pp. 664-674. PMLR (2019) +6. Bozorgtabar, B., Mahapatra, D., Thiran, J.P.: Exprada: Adversarial domain adaptation for facial expression analysis. Pattern Recognition 100, 107111 (2020) +7. Chen, C., Fu, Z., Chen, Z., Jin, S., Cheng, Z., Jin, X., Hua, X.S.: Homm: Higher-order moment matching for unsupervised domain adaptation. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 3422-3429 (2020) +8. Chen, D., Jacob, L., Mairal, J.: Convolutional kernel networks for graph-structured data. In: International Conference on Machine Learning. pp. 1576–1586. PMLR (2020) +9. Chen, H., Manhardt, F., Navab, N., Busam, B.: Texpose: Neural texture learning for self-supervised 6d object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4841-4852 (2023) +10. Cheng, X., Xie, Y.: Neural tangent kernel maximum mean discrepancy. Advances in Neural Information Processing Systems 34, 6658-6670 (2021) +11. Corcoran, P.: An end-to-end graph convolutional kernel support vector machine. Applied Network Science 5(1), 1-15 (2020) +12. Cortes, C., Vapnik, V.: Support-vector networks. Machine learning 20, 273-297 (1995) +13. Coumans, E., Bai, Y.: Pybullet, a python module for physics simulation for games, robotics and machine learning. http://pybullet.org (2016-2021) +14. Deng, J., Dong, W., Socher, R., Li, L., Kai Li, Li Fei-Fei: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. pp. 248-255 (2009). https://doi.org/10.1109/CVPR.2009.5206848 +15. Deng, X., Xiang, Y., Mousavian, A., Eppner, C., Bretl, T., Fox, D.: Self-supervised 6d object pose estimation for robot manipulation. In: 2020 IEEE International Conference on Robotics and Automation (ICRA). pp. 3665-3671. IEEE (2020) +16. Denninger, M., Sundermeyer, M., Winkelbauer, D., Olefir, D., Hodan, T., Zidan, Y., Elbadrawy, M., Knauer, M., Katam, H., Lodhi, A.: Blenderproc: Reducing the reality gap with photorealistic rendering. In: International Conference on Robotics: Science and Systems, RSS 2020 (2020) + +17. Denninger, M., Winkelbauer, D., Sundermeyer, M., Boerdijk, W., Knauer, M., Strobl, K.H., Humt, M., Triebel, R.: Blenderproc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software 8(82), 4901 (2023). https://doi.org/10.21105/joss.04901, https://doi.org/10.21105/joss.04901 +18. Di, Y., Manhardt, F., Wang, G., Ji, X., Navab, N., Tombari, F.: So-posed: Exploiting self-occlusion for direct 6d pose estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 12396-12405 (October 2021) +19. Doumanoglou, A., Kouskouridas, R., Malassiotis, S., Kim, T.K.: Recovering 6d object pose and predicting next-best-view in the crowd. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3583-3592 (2016) +20. Drost, B., Ulrich, M., Bergmann, P., Hartinger, P., Steger, C.: Introducing mvtec itodd-a dataset for 3d object recognition in industry. In: Proceedings of the IEEE international conference on computer vision workshops. pp. 2200-2208 (2017) +21. Gadwe, A., Ren, H.: Real-time 6 dof pose estimation of endoscopic instruments using printable markers. IEEE Sensors Journal 19(6), 2338-2346 (2018) +22. Gao, X., Hou, X., Tang, J., Cheng, H.: Complete solution classification for the perspective-three-point problem. IEEE Trans. Pattern Anal. Mach. Intell. 25, 930-943 (2003) +23. Ghojogh, B., Ghodsi, A., Karray, F., Crowley, M.: Reproducing kernel hilbert space, mercer's theorem, eigenfunctions, nystr\'' om method, and use of kernels in machine learning: Tutorial and survey. arXiv preprint arXiv:2106.08443 (2021) +24. Ghorbani, B., Mei, S., Misiakiewicz, T., Montanari, A.: When do neural networks outperform kernel methods? Advances in Neural Information Processing Systems 33, 14820-14830 (2020) +25. Greene, N., Luo, W., Kazanzides, P.: dvpose: Automated data collection and dataset for 6d pose estimation of robotic surgical instruments. In: 2023 International Symposium on Medical Robotics (ISMR). pp. 1-7. IEEE (2023) +26. Gretton, A., Borgwardt, K., Rasch, M., Scholkopf, B., Smola, A.: A kernel method for the two-sample-problem. Advances in neural information processing systems 19 (2006) +27. Gretton, A., Borgwardt, K.M., Rasch, M.J., Scholkopf, B., Smola, A.: A kernel two-sample test. The Journal of Machine Learning Research 13(1), 723-773 (2012) +28. Guo, S., Hu, Y., Alvarez, J.M., Salzmann, M.: Knowledge distillation for 6d pose estimation by aligning distributions of local predictions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18633-18642 (2023) +29. Hai, Y., Song, R., Li, J., Salzmann, M., Hu, Y.: Rigidity-aware detection for 6d object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8927-8936 (2023) +30. Haugaard, R.L., Buch, A.G.: Surfemb: Dense and continuous correspondence distributions for object pose estimation with learnt surface embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6749-6758 (2022) +31. He, Y., Huang, H., Fan, H., Chen, Q., Sun, J.: Ffb6d: A full flow bidirectional fusion network for 6d pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3003-3013 (2021) +32. He, Y., Sun, W., Huang, H., Liu, J., Fan, H., Sun, J.: Pvn3d: A deep pointwise 3d keypoints voting network for 6 dof pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020) + +33. He, Y., Wang, Y., Fan, H., Sun, J., Chen, Q.: Fs6d: Few-shot 6d pose estimation of novel objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6814-6824 (2022) +34. Hinterstoisser, S., Lepetit, V., Ilic, S., Holzer, S., Bradski, G., Konolige, K., Navab, N.: Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In: Asian conference on computer vision. pp. 548-562. Springer (2012) +35. Hodañ, T., Haluza, P., Obdržálek, Š., Matas, J., Lourakis, M., Zabulis, X.: T-LESS: An RGB-D dataset for 6D pose estimation of texture-less objects. IEEE Winter Conference on Applications of Computer Vision (WACV) (2017) +36. Hodan, T., Michel, F., Brachmann, E., Kehl, W., GlentBuch, A., Kraft, D., Drost, B., Vidal, J., Ihrke, S., Zabulis, X., et al.: Bop: Benchmark for 6d object pose estimation. In: Proceedings of the European conference on computer vision (ECCV). pp. 19-34 (2018) +37. Hodač, T., Sundermeyer, M., Drost, B., Labbé, Y., Brachmann, E., Michel, F., Rother, C., Matas, J.: Bop challenge 2020 on 6d object localization. In: European Conference on Computer Vision. pp. 577-594. Springer (2020) +38. Horn, B.K., Hilden, H.M., Negahdaripour, S.: Closed-form solution of absolute orientation using orthonormal matrices. JOSA A 5(7), 1127-1135 (1988) +39. Hu, Y., Fua, P., Salzmann, M.: Perspective flow aggregation for data-limited 6d object pose estimation. In: European Conference on Computer Vision. pp. 89-106. Springer (2022) +40. Kaskman, R., Zakharov, S., Shugurov, I., Ilic, S.: Homebreweddb: RGB-d dataset for 6d pose estimation of 3d objects. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. pp. 0-0 (2019) +41. Khosravi, M., Smith, R.S.: The existence and uniqueness of solutions for kernel-based system identification. Automatica 148, 110728 (2023) +42. Kleeberger, K., Huber, M.F.: Single shot 6d object pose estimation. In: 2020 IEEE International Conference on Robotics and Automation (ICRA). pp. 6239-6245. IEEE (2020) +43. Kleeberger, K., Landgraf, C., Huber, M.F.: Large-scale 6d object pose estimation dataset for industrial bin-picking. In: 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 2573-2578. IEEE (2019) +44. Kleeberger, K., Völk, M., Bormann, R., Huber, M.F.: Investigations on output parameterizations of neural networks for single shot 6d object pose estimation. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 13916-13922. IEEE (2021) +45. Lee, T., Lee, B.U., Shin, I., Choe, J., Shin, U., Kweon, I.S., Yoon, K.J.: Udacope: unsupervised domain adaptation for category-level object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14891-14900 (2022) +46. Lepetit, V., Moreno-Noguer, F., Fua, P.: Ep n p: An accurate o (n) solution to the p n p problem. International journal of computer vision 81, 155-166 (2009) +47. Li, Z., Hu, Y., Salzmann, M., Ji, X.: Sd-pose: Semantic decomposition for cross-domain 6d object pose estimation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 2020–2028 (2021) +48. Lin, J., Wei, Z., Ding, C., Jia, K.: Category-level 6d object pose and size estimation using self-supervised deep prior deformation networks. In: European Conference on Computer Vision. pp. 19-34. Springer (2022) + +49. Liu, F., Xu, W., Lu, J., Zhang, G., Gretton, A., Sutherland, D.J.: Learning deep kernels for non-parametric two-sample tests. In: International conference on machine learning. pp. 6316-6326. PMLR (2020) +50. Luo, Y.W., Ren, C.X.: Conditional bures metric for domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13989-13998 (2021) +51. Ma, X., Wang, Z., Li, H., Zhang, P., Ouyang, W., Fan, X.: Accurate monocular 3d object detection via color-embedded 3d reconstruction for autonomous driving. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6851-6860 (2019) +52. Mairal, J.: End-to-end kernel learning with supervised convolutional kernel networks. Advances in neural information processing systems 29 (2016) +53. Mairal, J., Koniusz, P., Harchaoui, Z., Schmid, C.: Convolutional kernel networks. Advances in neural information processing systems 27 (2014) +54. Manhardt, F., Arroyo, D.M., Rupprecht, C., Busam, B., Birdal, T., Navab, N., Tombari, F.: Explaining the ambiguity of object detection and 6d pose from visual data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019) +55. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) +56. Misiakiewicz, T., Mei, S.: Learning with convolution and pooling operations in kernel methods. Advances in Neural Information Processing Systems 35, 29014-29025 (2022) +57. Oberweger, M., Rad, M., Lepetit, V.: Making deep heatmaps robust to partial occlusions for 3d object pose estimation. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 119-134 (2018) +58. Pan, S.J., Tsang, I.W., Kwok, J.T., Yang, Q.: Domain adaptation via transfer component analysis. IEEE transactions on neural networks 22(2), 199-210 (2010) +59. Park, K., Mousavian, A., Xiang, Y., Fox, D.: Latentfusion: End-to-end differentiable reconstruction and rendering for unseen object pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10710-10719 (2020) +60. Pearson, K.: Liii. on lines and planes of closest fit to systems of points in space. The London, Edinburgh, and Dublin philosophical magazine and journal of science 2(11), 559-572 (1901) +61. Peng, S., Liu, Y., Huang, Q., Zhou, X., Bao, H.: Pynet: Pixel-wise voting network for 6dof pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4561-4570 (2019) +62. Pinheiro, P.O., Rostamzadeh, N., Ahn, S.: Domain-adaptive single-view 3d reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7638-7647 (2019) +63. Qi, C.R., Litany, O., He, K., Guibas, L.J.: Deep hough voting for 3d object detection in point clouds. In: proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9277-9286 (2019) +64. Sejdinovic, D., Sriperumbudur, B., Gretton, A., Fukumizu, K.: Equivalence of distance-based and rkhs-based statistics in hypothesis testing. The annals of statistics pp. 2263-2291 (2013) +65. Shan, P., Bi, Y., Li, Z., Wang, Q., He, Z., Zhao, Y., Peng, S.: Unsupervised model adaptation for multivariate calibration by domain adaptation-regularization based + +kernel partial least square. Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 292, 122418 (2023) +66. Shugurov, I., Li, F., Busam, B., Ilic, S.: Osop: A multi-stage one shot object pose estimation framework. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6835-6844 (2022) +67. Shugurov, I., Zakharov, S., Ilic, S.: Dpodv2: Dense correspondence-based 6 dof pose estimation. IEEE transactions on pattern analysis and machine intelligence 44(11), 7417-7435 (2021) +68. Sock, J., Garcia-Hernando, G., Armagan, A., Kim, T.K.: Introducing pose consistency and warp-alignment for self-supervised 6d object pose estimation in color images. In: 2020 International Conference on 3D Vision (3DV). pp. 291-300. IEEE (2020) +69. Su, H., Qi, C.R., Li, Y., Guibas, L.J.: Render for cnn: Viewpoint estimation in images using cnns trained with rendered 3d model views. In: Proceedings of the IEEE international conference on computer vision. pp. 2686-2694 (2015) +70. Su, Y., Saleh, M., Fetzer, T., Rambach, J., Navab, N., Busam, B., Stricker, D., Tombari, F.: Zebrapose: Coarse to fine surface encoding for 6dof object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6738-6748 (2022) +71. Sundermeyer, M., Durner, M., Huang, E.Y., Marton, Z.C., Vaskevicius, N., Arras, K.O., Triebel, R.: Multi-path learning for object pose estimation across domains. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13916-13925 (2020) +72. Sundermeyer, M., Hodač, T., Labbe, Y., Wang, G., Brachmann, E., Drost, B., Rother, C., Matas, J.: Bop challenge 2022 on detection, segmentation and pose estimation of specific rigid objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2784-2793 (2023) +73. Sundermeyer, M., Marton, Z.C., Durner, M., Brucker, M., Triebel, R.: Implicit 3d orientation learning for 6d object detection from rgb images. In: Proceedings of the european conference on computer vision (ECCV). pp. 699-715 (2018) +74. Szafraniec, F.H.: The reproducing kernel hilbert space and its multiplication operators. Complex Analysis and Related Topics pp. 253-263 (2000) +75. Tan, T., Dong, Q.: Smoc-net: Leveraging camera pose for self-supervised monocular object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21307-21316 (2023) +76. Ullman, S.: The interpretation of structure from motion. Proceedings of the Royal Society of London. Series B. Biological Sciences 203(1153), 405-426 (1979) +77. Venkateswara, H., Eusebio, J., Chakraborty, S., Panchanathan, S.: Deep hashing network for unsupervised domain adaptation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5018-5027 (2017) +78. Vert, J.P., Tsuda, K., Scholkopf, B.: A primer on kernel methods. Kernel methods in computational biology 47, 35-70 (2004) +79. Wang, G., Manhardt, F., Liu, X., Ji, X., Tombari, F.: Occlusion-aware self-supervised monocular 6d object pose estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021) +80. Wang, G., Manhardt, F., Shao, J., Ji, X., Navab, N., Tombari, F.: Self6d: Self-supervised monocular 6d object pose estimation. In: European Conference on Computer Vision. pp. 108-125. Springer (2020) +81. Wu, Y., Greenspan, M.: Learning better keypoints for multi-object 6dof pose estimation. arXiv preprint arXiv:2308.07827 (2023) + +82. Wu, Y., Javaheri, A., Zand, M., Greenspan, M.: Keypoint cascade voting for point cloud based 6dof pose estimation. In: 2022 International Conference on 3D Vision (3DV). pp. 176-186. IEEE (2022) +83. Wu, Y., Zand, M., Etemad, A., Greenspan, M.: Vote from the center: 6 dof pose estimation in rgb-d images by radial keypoint voting. In: European Conference on Computer Vision. pp. 335-352. Springer (2022) +84. Xiang, Y., Schmidt, T., Narayanan, V., Fox, D.: PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes (2018) +85. Xiao, F., Liu, H., Lee, Y.J.: Identity from here, pose from there: Self-supervised disentanglement and generation of objects using unlabeled videos. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7013-7022 (2019) +86. Xiao, Y., Qiu, X., Langlois, P.A., Aubry, M., Marlet, R.: Pose from shape: Deep pose estimation for arbitrary 3d objects. arXiv preprint arXiv:1906.05105 (2019) +87. Yang, H., Pavone, M.: Object pose estimation with statistical guarantees: Conformal keypoint detection and geometric uncertainty propagation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8947-8958 (2023) +88. Zhang, Y., David, P., Gong, B.: Curriculum domain adaptation for semantic segmentation of urban scenes. In: Proceedings of the IEEE international conference on computer vision. pp. 2020-2030 (2017) +89. Zhang, Z., Wang, M., Huang, Y., Nehorai, A.: Aligning infinite-dimensional covariance matrices in reproducing kernel hilbert spaces for domain adaptation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3437-3445 (2018) +90. Zhou, J., Chen, K., Xu, L., Dou, Q., Qin, J.: Deep fusion transformer network with weighted vector-wise keypoints voting for robust 6d object pose estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13967-13977 (2023) \ No newline at end of file diff --git a/2024/Pseudo-keypoint RKHS Learning for Self-supervised 6DoF Pose Estimation/images.zip b/2024/Pseudo-keypoint RKHS Learning for Self-supervised 6DoF Pose Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..57de406dafff12dfe8281df14a41376c7927cef8 --- /dev/null +++ b/2024/Pseudo-keypoint RKHS Learning for Self-supervised 6DoF Pose Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:01737b0c35003085b1abf86f541ee2b19ecd267ca096f062b4972604453d2f8d +size 518102 diff --git a/2024/Pseudo-keypoint RKHS Learning for Self-supervised 6DoF Pose Estimation/layout.json b/2024/Pseudo-keypoint RKHS Learning for Self-supervised 6DoF Pose Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..80c320aae98faefc20ebc097fa4d3a4205d219e4 --- /dev/null +++ b/2024/Pseudo-keypoint RKHS Learning for Self-supervised 6DoF Pose Estimation/layout.json @@ -0,0 +1,13501 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 30, + 337, + 64 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 30, + 337, + 64 + ], + "spans": [ + { + "bbox": [ + 73, + 30, + 337, + 64 + ], + "type": "text", + "content": "Pseudo-keypoint RKHS Learning for Self-supervised 6DoF Pose Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 114, + 86, + 298, + 98 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 86, + 298, + 98 + ], + "spans": [ + { + "bbox": [ + 114, + 86, + 298, + 98 + ], + "type": "text", + "content": "Yangzheng Wu and Michael Greenspan" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 107, + 353, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 107, + 353, + 141 + ], + "spans": [ + { + "bbox": [ + 59, + 107, + 353, + 141 + ], + "type": "text", + "content": "RCVLab, Dept. of Electrical and Computer Engineering, Ingenuity Labs, Queen's University, Kingston, Ontario, Canada {y.wu, michael.greenspan}@queensu.ca" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 166, + 356, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 166, + 356, + 352 + ], + "spans": [ + { + "bbox": [ + 61, + 166, + 356, + 352 + ], + "type": "text", + "content": "Abstract. We address the simulation-to-real domain gap in six degree-of-freedom pose estimation (6DoF PE), and propose a novel self-supervised keypoint voting-based 6DoF PE framework, effectively narrowing this gap using a learnable kernel in RKHS. We formulate this domain gap as a distance in high-dimensional feature space, distinct from previous iterative matching methods. We propose an adapter network, which is pre-trained on purely synthetic data with synthetic ground truth poses, and which evolves the network parameters from this source synthetic domain to the target real domain. Importantly, the real data training only uses pseudo-poses estimated by pseudo-keypoints, and thereby requires no real ground truth data annotations. Our proposed method is called RKHSPose, and achieves state-of-the-art performance among self-supervised methods on three commonly used 6DoF PE datasets including LINEMOD " + }, + { + "bbox": [ + 61, + 166, + 356, + 352 + ], + "type": "inline_equation", + "content": "(+4.2\\%)" + }, + { + "bbox": [ + 61, + 166, + 356, + 352 + ], + "type": "text", + "content": ", Occlusion LINEMOD " + }, + { + "bbox": [ + 61, + 166, + 356, + 352 + ], + "type": "inline_equation", + "content": "(+2\\%)" + }, + { + "bbox": [ + 61, + 166, + 356, + 352 + ], + "type": "text", + "content": ", and YCB-Video " + }, + { + "bbox": [ + 61, + 166, + 356, + 352 + ], + "type": "inline_equation", + "content": "(+3\\%)" + }, + { + "bbox": [ + 61, + 166, + 356, + 352 + ], + "type": "text", + "content": ". It also compares favorably to fully supervised methods on all six applicable BOP core datasets, achieving within " + }, + { + "bbox": [ + 61, + 166, + 356, + 352 + ], + "type": "inline_equation", + "content": "-11.3\\%" + }, + { + "bbox": [ + 61, + 166, + 356, + 352 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 61, + 166, + 356, + 352 + ], + "type": "inline_equation", + "content": "+0.2\\%" + }, + { + "bbox": [ + 61, + 166, + 356, + 352 + ], + "type": "text", + "content": " of the top fully supervised results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 363, + 351, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 363, + 351, + 385 + ], + "spans": [ + { + "bbox": [ + 61, + 363, + 351, + 385 + ], + "type": "text", + "content": "Keywords: pose estimation " + }, + { + "bbox": [ + 61, + 363, + 351, + 385 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 61, + 363, + 351, + 385 + ], + "type": "text", + "content": " self-supervision " + }, + { + "bbox": [ + 61, + 363, + 351, + 385 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 61, + 363, + 351, + 385 + ], + "type": "text", + "content": " domain adaptation " + }, + { + "bbox": [ + 61, + 363, + 351, + 385 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 61, + 363, + 351, + 385 + ], + "type": "text", + "content": " key-point estimation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 34, + 404, + 127, + 416 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 34, + 404, + 127, + 416 + ], + "spans": [ + { + "bbox": [ + 34, + 404, + 127, + 416 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 427, + 379, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 427, + 379, + 511 + ], + "spans": [ + { + "bbox": [ + 33, + 427, + 379, + 511 + ], + "type": "text", + "content": "RGB-D Six Degree-of-Freedom Pose Estimation (6DoF PE) is a problem being actively explored in computer vision research. Given an RGB image and its associated depth map, the task is to detect scene objects and estimate their poses comprising 3DoF rotational angles and 3DoF translational offsets in the camera reference frame. This task enables many applications such as augmented reality [34,36,40,84], robotic bin picking [19,35,43], autonomous driving [51,85] and image-guided surgeries [21,25]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 512, + 379, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 512, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 512, + 379, + 583 + ], + "type": "text", + "content": "As with other machine learning (ML) tasks, fully supervised 6DoF PE requires large annotated datasets. This requirement is particularly challenging for 6DoF PE, as the annotations comprise not only the identity of the objects in the scene, but also their 6DoF pose, which makes the data relatively expensive to annotate compared to related tasks such as classification, detection and segmentation. This is due to the fact that humans are not able to qualitatively or" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 413, + 615 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 40, + 31, + 373, + 120 + ], + "blocks": [ + { + "bbox": [ + 40, + 31, + 373, + 120 + ], + "lines": [ + { + "bbox": [ + 40, + 31, + 373, + 120 + ], + "spans": [ + { + "bbox": [ + 40, + 31, + 373, + 120 + ], + "type": "image", + "image_path": "6d7dd66c9b3040ba2b96e07a7a37270b2ed6f71f87197eeee34a608ef6ebd606.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 30, + 127, + 381, + 184 + ], + "lines": [ + { + "bbox": [ + 30, + 127, + 381, + 184 + ], + "spans": [ + { + "bbox": [ + 30, + 127, + 381, + 184 + ], + "type": "text", + "content": "Fig. 1: RKHSPose adapts the network pretrained on synthetic data to real test scenes (left), by comparing network feature spaces with real image inputs (solid arrows), against those with syn/real image (right) inputs (dashed arrows). " + }, + { + "bbox": [ + 30, + 127, + 381, + 184 + ], + "type": "inline_equation", + "content": "M_r" + }, + { + "bbox": [ + 30, + 127, + 381, + 184 + ], + "type": "text", + "content": " regresses radial quantities, " + }, + { + "bbox": [ + 30, + 127, + 381, + 184 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 30, + 127, + 381, + 184 + ], + "type": "text", + "content": " is the Adapter network, and RKHS maps features into a higher dimensional space." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 30, + 230, + 381, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 230, + 381, + 398 + ], + "spans": [ + { + "bbox": [ + 30, + 230, + 381, + 398 + ], + "type": "text", + "content": "intuitively estimate 6DoF pose, which therefore requires additional instrumentation of the scene at data collection, and/or more sophisticated user annotation tools [34,84]. Consequently, synthetic 6DoF PE datasets [36,43] have been introduced, either as an additional complement to real datasets, or as standalone purely synthetic datasets. Annotated synthetic datasets are of course trivially inexpensive to collect, simply because precise synthetic pose annotation in a simulated environment is fully automatic. A known challenge in using synthetic data, however, is that there typically exists a domain gap between the real and synthetic data, which makes results less accurate when inferring in real data using models trained on purely synthetic datasets. Expectations of the potential benefit of synthetic datasets has led to the exploration of a rich set of Domain Adaptation (DA) methods, which specifically aim to reduce the domain gap [6,62,88] using inexpensive synthetic data for a wide variety of tasks, recently including 6DoF PE [28,45,47]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 30, + 404, + 381, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 404, + 381, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 404, + 381, + 584 + ], + "type": "text", + "content": "Early methods [42, 83] ignored the simulation-to-real (sim2real) domain gap and nevertheless improved performance by training on both synthetic and real annotated data, effectively augmenting the real images with the synthetic. However, these methods still required real labels to be sufficiently accurate and robust for practical applications, partially due to the domain gap. As shown in Fig. 1, the rendered synthetic objects (right image) have a slightly different appearance than the real objects (left image). The details of the CAD models, both geometry and texture, are not precise, as can be seen for the can object (which lacks a mouth and shadows), the benchwise (which is missing a handle), and the holepuncher (which has coarse geometric resolution). Several recent methods [9, 44, 73, 75, 80] have started to address the sim2real gap for 6DoF PE by first training on labeled synthetic data and then fine-tuning on unlabeled real data. Commonly known as self-supervised, these methods reduce the domain gap by adding extra supervision using features extracted from real images without requiring real ground truth (GT) labels. The majority of these methods are view-" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 32, + 11, + 39, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 11, + 39, + 18 + ], + "spans": [ + { + "bbox": [ + 32, + 11, + 39, + 18 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 9, + 168, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 9, + 168, + 20 + ], + "spans": [ + { + "bbox": [ + 66, + 9, + 168, + 20 + ], + "type": "text", + "content": "Y.Wu and M. Greenspan" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 34, + 379, + 59 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 34, + 379, + 59 + ], + "spans": [ + { + "bbox": [ + 32, + 34, + 379, + 59 + ], + "type": "text", + "content": "point/template matching-based, and the self-supervision commonly iteratively matches 2D masks or 3D point clouds [44,73,80]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 32, + 59, + 379, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 59, + 379, + 178 + ], + "spans": [ + { + "bbox": [ + 32, + 59, + 379, + 178 + ], + "type": "text", + "content": "While the above-mentioned self-supervision works have shown promise, there exist a wealth of DA techniques that can be brought to bear to improve performance further for this task. One such technique is Reproducing Kernel Hilbert Space (RKHS), which is a kernel method that has been shown to be effective for DA [1,5,41]. RKHS was initially used to create decision boundaries for non-separable data [12,60], and has been shown to be effective at reducing the domain gap for various tasks and applications [7,65,89]. The reproducing kernel guarantees that the domain gap can be statistically measured, allowing network parameters trained on synthetic data to be effectively adapted to the real data, using specifically tailored metrics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 32, + 178, + 379, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 178, + 379, + 404 + ], + "spans": [ + { + "bbox": [ + 32, + 178, + 379, + 404 + ], + "type": "text", + "content": "To address the sim2real domain gap in 6DoF PE, we propose RKHSPose, which is a keypoint-based method [82, 83] trained on a mixture of a large collection of labeled synthetic data, and a small handful of unlabeled real data. RKHSPose estimates the intermediate radial voting quantity, which has been shown to be effective for estimating keypoints [83], by first training a modified FCN-Resnet-18 on purely synthetic data, with automatically labeled synthetic GT poses. The radial quantity is a map of the distance from each image pixel to each keypoint. Next, real images are passed through the synthetically trained network, resulting in a set of pseudo-keypoints. The real images and their corresponding pseudo-keypoints are used to render a set of 'synthetic-overreal' (syn/real) images, by first estimating the pseudo-pose from the pseudo-keypoints, and then overlaying the synthetic object, rendered with the pseudopose, onto the real image. The network training then continues on the syn/real images, invoking an RKHS network module with a trainable linear product kernel, which minimizes the Maximum Mean Discrepancy (MMD) loss. At the front end, a proposed keypoint radial voting network learns to cast votes to estimate keypoints from the backend-generated radial maps. The final pose is then determined using ePnP [46] based on the estimated and corresponding object keypoints." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 405, + 225, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 405, + 225, + 416 + ], + "spans": [ + { + "bbox": [ + 48, + 405, + 225, + 416 + ], + "type": "text", + "content": "The main contributions of this work are:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 37, + 421, + 379, + 491 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 37, + 421, + 379, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 421, + 379, + 445 + ], + "spans": [ + { + "bbox": [ + 37, + 421, + 379, + 445 + ], + "type": "text", + "content": "- A novel learnable RKHS-based Adapter backend network architecture to minimize the sim2real domain gap in 6DoF PE;" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 37, + 445, + 332, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 445, + 332, + 456 + ], + "spans": [ + { + "bbox": [ + 37, + 445, + 332, + 456 + ], + "type": "text", + "content": "- A novel CNN-based frontend network for keypoint radial voting;" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 37, + 456, + 379, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 456, + 379, + 491 + ], + "spans": [ + { + "bbox": [ + 37, + 456, + 379, + 491 + ], + "type": "text", + "content": "- A self-supervised keypoint-based 6DoF PE method, RKHSPose, which is shown to have state-of-the-art (SOTA) performance, based on our experiments and several ablation studies." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 32, + 507, + 137, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 507, + 137, + 520 + ], + "spans": [ + { + "bbox": [ + 32, + 507, + 137, + 520 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 33, + 530, + 107, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 530, + 107, + 541 + ], + "spans": [ + { + "bbox": [ + 33, + 530, + 107, + 541 + ], + "type": "text", + "content": "2.1 6DoF PE" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 33, + 547, + 379, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 547, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 547, + 379, + 583 + ], + "type": "text", + "content": "ML-based 6DoF PE methods [57,61,84] all train a network to regress quantities, such as keypoints and camera viewpoints, as have been used in classical algorithms [34]. ML-based methods, which initially became popular for the general" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 300, + 9, + 346, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 9, + 346, + 19 + ], + "spans": [ + { + "bbox": [ + 300, + 9, + 346, + 19 + ], + "type": "text", + "content": "RKHPose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 85, + 97, + 326, + 311 + ], + "blocks": [ + { + "bbox": [ + 32, + 32, + 379, + 87 + ], + "lines": [ + { + "bbox": [ + 32, + 32, + 379, + 87 + ], + "spans": [ + { + "bbox": [ + 32, + 32, + 379, + 87 + ], + "type": "text", + "content": "Table 1: Existing self-supervised 6DoF PE methods. Some methods use CAD models, labeled synthetic (syn) data (images+poses), and real data without GT labels, while others use only labeled synthetic data. Many methods require a ROI such as a bounding box or a semantic mask, manually labeled or estimated by an existing framework. SO-Pose [18] used very few real labels to improve performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 97, + 326, + 311 + ], + "lines": [ + { + "bbox": [ + 85, + 97, + 326, + 311 + ], + "spans": [ + { + "bbox": [ + 85, + 97, + 326, + 311 + ], + "type": "table", + "html": "
methodmodeCAD modelsyn datareal imagesreal posesROI
SO-Pose [18]RGB✓✓✓✓✓
TexPose [9]RGB✓✓✓X✓
SMOC-Net [75]RGB✓✓✓X✓
FS6D [33]RGBD✓✓XX✓
AAE [73]RGB✓✓XX✓
MHP [54]RGB✓✓XX✓
Sock et al. [68]RGB✓✓XX✓
DSC [86]RGBD✓✓XX✓
Sundermeyer [71]RGB✓✓XX✓
LatentFusion [59]RGBX✓XX✓
OSOP [66]RGBD✓✓XXX
Kleeberger et al. [44]D✓✓XXX
Su et al. [69]RGB✓✓✓XX
Self6D [80]RGBD✓✓✓XX
Self6D++ [79]RGB✓✓✓XX
Deng et al. [15]RGBD✓✓✓XX
RKHSPose (Ours)RGBD✓✓✓XX
", + "image_path": "086f41ad63e4e6f3f883610df8965df6ce7c48a1f886130a8b4ec37c03fa14c0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 349, + 379, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 349, + 379, + 373 + ], + "spans": [ + { + "bbox": [ + 33, + 349, + 379, + 373 + ], + "type": "text", + "content": "object detection task, have started to dominate the 6DoF PE literature due to their accuracy and efficiency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 375, + 379, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 375, + 379, + 471 + ], + "spans": [ + { + "bbox": [ + 33, + 375, + 379, + 471 + ], + "type": "text", + "content": "There are two main categories of ML-based fully supervised methods: feature matching-based [29, 30, 57, 70, 84], and keypoint-based methods [31, 32, 61, 87]. Feature matching-based methods make use of the structures from general object detection networks most directly. The network encodes and matches features and estimates pose by either regressing elements of the pose (e.g. the transformation matrix [75], rotational angles and translational offsets [84] or 3D vertices [70]) directly, or by regressing some intermediate feature-matching representations, such as viewpoints or segments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 473, + 379, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 473, + 379, + 557 + ], + "spans": [ + { + "bbox": [ + 33, + 473, + 379, + 557 + ], + "type": "text", + "content": "In contrast, keypoint-based methods encode features to estimate keypoints which are predefined within the reference frame of an object's CAD model. These methods then use (modified) classical algorithms such as PnP [22, 46], Horn's method [38], and ICP [3] to estimate the final pose from corresponding image and model keypoints. Unlike feature-matching methods, keypoint-based methods are typically more accurate due to redundancies encountered through voting schemes [32, 61, 83] and by generating confidence hypotheses of keypoints [31, 87]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 559, + 379, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 559, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 559, + 379, + 583 + ], + "type": "text", + "content": "Recently, self-supervised 6DoF PE methods have been explored in order to reduce the reliance on labeled real data, which is expensive to acquire. As sum-" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 11, + 39, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 11, + 39, + 18 + ], + "spans": [ + { + "bbox": [ + 33, + 11, + 39, + 18 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 9, + 167, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 9, + 167, + 19 + ], + "spans": [ + { + "bbox": [ + 66, + 9, + 167, + 19 + ], + "type": "text", + "content": "Y.Wu and M. Greenspan" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 33, + 379, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 33, + 379, + 106 + ], + "spans": [ + { + "bbox": [ + 32, + 33, + 379, + 106 + ], + "type": "text", + "content": "marized in Table 1, these methods commonly use real images without GT labels. Some methods use pure synthetic data and CAD models only, with the exception of LatentFusion [59] which trained the model using only synthetic data. The majority of these methods [9, 15, 33, 48, 66, 68, 69, 71, 75, 79, 80] are inspired by fully supervised feature-matching methods, except DPODv2 [67] and DSC [86], in which keypoint correspondences are rendered and matched." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 32, + 110, + 379, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 110, + 379, + 242 + ], + "spans": [ + { + "bbox": [ + 32, + 110, + 379, + 242 + ], + "type": "text", + "content": "A few methods [15,66] fine-tuned the pose trivially by iteratively matching the template/viewpoint, whereas others [33,44,59,86] augmented the training data by adding noise [44,86], rendering textures [33,67] and creating a latent space [59]. Some methods [54,71,73] also implemented DA techniques such as codebook encoding [71], Principle Component Analysis (PCA) [71] and symmetric Bingham distributions [54]. Most methods [66,68,69,79,80,86] used rendering techniques to render and match a template. There are a few methods that combined 3D reconstruction techniques, such as Neural radiance fields (Nerf) [55] and Structure from Motion (SfM) [76]. TexPose [9] matched CAD models to segments generated by Nerf, and SMOC-Net [75] used SfM to create the 3D segment and matched with the CAD model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 281, + 240, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 281, + 240, + 294 + ], + "spans": [ + { + "bbox": [ + 33, + 281, + 240, + 294 + ], + "type": "text", + "content": "2.2 Kernel Methods and Deep Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 323, + 379, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 323, + 379, + 419 + ], + "spans": [ + { + "bbox": [ + 33, + 323, + 379, + 419 + ], + "type": "text", + "content": "While deep learning is the most common ML technique within the computer vision literature, kernel methods [12,60,64] have also been actively explored. Kernel methods are typically in RKHS space [74] with reproducing properties that facilitate solving non-linear problems by mapping input data into high dimensional spaces that can be linearly separated [11,23]. Well-known early kernel methods that have been applied to computer vision are Support Vector Machines [12] and PCA [60]. A recent method [64] linked energy distance and MMD in RKHS, and showed the effectiveness of kernels in statistical hypothesis testing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 423, + 379, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 423, + 379, + 543 + ], + "spans": [ + { + "bbox": [ + 33, + 423, + 379, + 543 + ], + "type": "text", + "content": "More recent studies compare kernel methods with deep learning networks [1, 4,5,41]. RKHS is found to perform better on classification tasks than one single block of a CNN comprising convolution, pooling and downsampling (linear) layers [41]. RKHS can also help with CNN generalization by meta-regularization on image registration problems [1]. Similarly, norms (magnitude of trainable weights) defined in RKHS help with CNN regularization [4,5]. Further, discriminant information in label distributions in unsupervised DA is addressed and RKHS-based Conditional Kernel Bures metric is proposed [50]. Lastly, the connection between Neural Tangent Kernel and MMD is established, and an efficient MMD for two-sample tests is developed [10]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 547, + 379, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 547, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 547, + 379, + 583 + ], + "type": "text", + "content": "Inspired by the previous work, RKHSPose applies concepts of kernel learning to keypoint-based 6DoF PE, to provide an effective means to self-supervise a synthetically trained network on unlabeled real data." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 300, + 9, + 346, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 9, + 346, + 19 + ], + "spans": [ + { + "bbox": [ + 300, + 9, + 346, + 19 + ], + "type": "text", + "content": "RKHPose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 44, + 32, + 292, + 215 + ], + "blocks": [ + { + "bbox": [ + 44, + 32, + 292, + 215 + ], + "lines": [ + { + "bbox": [ + 44, + 32, + 292, + 215 + ], + "spans": [ + { + "bbox": [ + 44, + 32, + 292, + 215 + ], + "type": "image", + "image_path": "fe1acf770a9f08cf67b785bce1bf8b9cfa652b18d5242f45ceffc0883be82151.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 31, + 223, + 380, + 279 + ], + "lines": [ + { + "bbox": [ + 31, + 223, + 380, + 279 + ], + "spans": [ + { + "bbox": [ + 31, + 223, + 380, + 279 + ], + "type": "text", + "content": "Fig. 2: RKHSPose architecture. RKHPose is first trained on synthetic labeled data (solid arrows), and then finetuned on alternating syn/real and (unlabeled) real images (dashed arrows). " + }, + { + "bbox": [ + 31, + 223, + 380, + 279 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 31, + 223, + 380, + 279 + ], + "type": "text", + "content": " is measured by MMD in RKHS by densely mapping the intermediate features of " + }, + { + "bbox": [ + 31, + 223, + 380, + 279 + ], + "type": "inline_equation", + "content": "M_r" + }, + { + "bbox": [ + 31, + 223, + 380, + 279 + ], + "type": "text", + "content": " into high dimensional spaces with conv blocks. The distance is treated as " + }, + { + "bbox": [ + 31, + 223, + 380, + 279 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{M_A}" + }, + { + "bbox": [ + 31, + 223, + 380, + 279 + ], + "type": "text", + "content": " and back-propagated through " + }, + { + "bbox": [ + 31, + 223, + 380, + 279 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 31, + 223, + 380, + 279 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 31, + 223, + 380, + 279 + ], + "type": "inline_equation", + "content": "M_r" + }, + { + "bbox": [ + 31, + 223, + 380, + 279 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 301, + 36, + 367, + 86 + ], + "blocks": [ + { + "bbox": [ + 301, + 36, + 367, + 86 + ], + "lines": [ + { + "bbox": [ + 301, + 36, + 367, + 86 + ], + "spans": [ + { + "bbox": [ + 301, + 36, + 367, + 86 + ], + "type": "image", + "image_path": "b39f96337e2343f7784c0e00651d956680d5ccf78ea772c2927bf9a92dd30c72.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 301, + 94, + 368, + 145 + ], + "blocks": [ + { + "bbox": [ + 328, + 87, + 341, + 94 + ], + "lines": [ + { + "bbox": [ + 328, + 87, + 341, + 94 + ], + "spans": [ + { + "bbox": [ + 328, + 87, + 341, + 94 + ], + "type": "inline_equation", + "content": "K_{syn}" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 301, + 94, + 368, + 145 + ], + "lines": [ + { + "bbox": [ + 301, + 94, + 368, + 145 + ], + "spans": [ + { + "bbox": [ + 301, + 94, + 368, + 145 + ], + "type": "image", + "image_path": "dd85496e0dc56714813d5722c57ab3aebe87c652c635f9995a8cb599294accce.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 145, + 346, + 152 + ], + "lines": [ + { + "bbox": [ + 324, + 145, + 346, + 152 + ], + "spans": [ + { + "bbox": [ + 324, + 145, + 346, + 152 + ], + "type": "inline_equation", + "content": "K_{syn / real}" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 301, + 152, + 367, + 205 + ], + "blocks": [ + { + "bbox": [ + 301, + 152, + 367, + 205 + ], + "lines": [ + { + "bbox": [ + 301, + 152, + 367, + 205 + ], + "spans": [ + { + "bbox": [ + 301, + 152, + 367, + 205 + ], + "type": "image", + "image_path": "f704538c2f497cab3eea1ec82dc383a7e159ba121a1acb97190844cf6e4a056e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 206, + 341, + 213 + ], + "lines": [ + { + "bbox": [ + 328, + 206, + 341, + 213 + ], + "spans": [ + { + "bbox": [ + 328, + 206, + 341, + 213 + ], + "type": "inline_equation", + "content": "K_{real}" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 31, + 310, + 101, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 310, + 101, + 322 + ], + "spans": [ + { + "bbox": [ + 31, + 310, + 101, + 322 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 31, + 336, + 154, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 336, + 154, + 348 + ], + "spans": [ + { + "bbox": [ + 31, + 336, + 154, + 348 + ], + "type": "text", + "content": "3.1 Network Overview" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "spans": [ + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "text", + "content": "As shown in Fig. 2, RKHSPose is made up of two networks, a main network " + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "inline_equation", + "content": "M_{rv}" + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "text", + "content": " for keypoint regression and classification, and an Adapter network " + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "text", + "content": ". The input of " + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "inline_equation", + "content": "M_{rv}" + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "text", + "content": " with shape " + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "inline_equation", + "content": "W \\times H \\times 4" + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "text", + "content": ", is the concatenation of an RGB image " + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "text", + "content": " and its corresponding depth map " + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "text", + "content": ". This input can be synthetically generated with an arbitrary background (" + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "inline_equation", + "content": "I_{syn}" + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "inline_equation", + "content": "D_{syn}" + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "text", + "content": "), a synthetic mask overlayed on a real background (" + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "inline_equation", + "content": "I_{syn/real}" + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "inline_equation", + "content": "D_{syn/real}" + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "text", + "content": "), or a real (" + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "inline_equation", + "content": "I_{real}" + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "inline_equation", + "content": "D_{real}" + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "text", + "content": ") image. The outputs are " + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "text", + "content": " projected 2D keypoints " + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "text", + "content": ", along with corresponding classification labels " + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "text", + "content": " and confidence scores " + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "text", + "content": " are organized into instance sets based on " + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 30, + 356, + 380, + 463 + ], + "type": "text", + "content": " and geometric constraints." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 30, + 464, + 379, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 464, + 379, + 559 + ], + "spans": [ + { + "bbox": [ + 30, + 464, + 379, + 559 + ], + "type": "inline_equation", + "content": "M_{rv}" + }, + { + "bbox": [ + 30, + 464, + 379, + 559 + ], + "type": "text", + "content": " comprises two sub-networks, regression network " + }, + { + "bbox": [ + 30, + 464, + 379, + 559 + ], + "type": "inline_equation", + "content": "M_r" + }, + { + "bbox": [ + 30, + 464, + 379, + 559 + ], + "type": "text", + "content": " and voting network " + }, + { + "bbox": [ + 30, + 464, + 379, + 559 + ], + "type": "inline_equation", + "content": "M_v" + }, + { + "bbox": [ + 30, + 464, + 379, + 559 + ], + "type": "text", + "content": ". Inspired by recent voting techniques [32, 61, 82, 83, 90], " + }, + { + "bbox": [ + 30, + 464, + 379, + 559 + ], + "type": "inline_equation", + "content": "M_r" + }, + { + "bbox": [ + 30, + 464, + 379, + 559 + ], + "type": "text", + "content": " estimates an intermediate voting quantity, which is a radial distance map " + }, + { + "bbox": [ + 30, + 464, + 379, + 559 + ], + "type": "inline_equation", + "content": "V_r" + }, + { + "bbox": [ + 30, + 464, + 379, + 559 + ], + "type": "text", + "content": " [83], by using a modified Fully Connected ResNet 18 (FCN-ResNet-18). The radial voting map " + }, + { + "bbox": [ + 30, + 464, + 379, + 559 + ], + "type": "inline_equation", + "content": "V_r" + }, + { + "bbox": [ + 30, + 464, + 379, + 559 + ], + "type": "text", + "content": ", with shape " + }, + { + "bbox": [ + 30, + 464, + 379, + 559 + ], + "type": "inline_equation", + "content": "W \\times H" + }, + { + "bbox": [ + 30, + 464, + 379, + 559 + ], + "type": "text", + "content": ", stores the Euclidean distance from each object point to each keypoint in the 3D camera world reference frame. The voting network " + }, + { + "bbox": [ + 30, + 464, + 379, + 559 + ], + "type": "inline_equation", + "content": "M_v" + }, + { + "bbox": [ + 30, + 464, + 379, + 559 + ], + "type": "text", + "content": " (described in Sec. 3.3) then takes " + }, + { + "bbox": [ + 30, + 464, + 379, + 559 + ], + "type": "inline_equation", + "content": "V_r" + }, + { + "bbox": [ + 30, + 464, + 379, + 559 + ], + "type": "text", + "content": " as input, accumulates votes, and detects peaks to estimate " + }, + { + "bbox": [ + 30, + 464, + 379, + 559 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 30, + 464, + 379, + 559 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 30, + 464, + 379, + 559 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 30, + 464, + 379, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 464, + 379, + 559 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 30, + 464, + 379, + 559 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 31, + 560, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 560, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 31, + 560, + 380, + 584 + ], + "type": "text", + "content": "The Adapter network " + }, + { + "bbox": [ + 31, + 560, + 380, + 584 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 31, + 560, + 380, + 584 + ], + "type": "text", + "content": " consists of a series of CNNs which encode pairs of feature maps from " + }, + { + "bbox": [ + 31, + 560, + 380, + 584 + ], + "type": "inline_equation", + "content": "M_{rv}" + }, + { + "bbox": [ + 31, + 560, + 380, + 584 + ], + "type": "text", + "content": ", and are trained on both synthetic overlayed and pure real" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 32, + 11, + 39, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 11, + 39, + 18 + ], + "spans": [ + { + "bbox": [ + 32, + 11, + 39, + 18 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 9, + 168, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 9, + 168, + 20 + ], + "spans": [ + { + "bbox": [ + 66, + 9, + 168, + 20 + ], + "type": "text", + "content": "Y.Wu and M. Greenspan" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 30, + 34, + 380, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 34, + 380, + 129 + ], + "spans": [ + { + "bbox": [ + 30, + 34, + 380, + 129 + ], + "type": "text", + "content": "data. " + }, + { + "bbox": [ + 30, + 34, + 380, + 129 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 30, + 34, + 380, + 129 + ], + "type": "text", + "content": " encodes feature map pairs " + }, + { + "bbox": [ + 30, + 34, + 380, + 129 + ], + "type": "inline_equation", + "content": "(f_{syn / real}, f_{real})" + }, + { + "bbox": [ + 30, + 34, + 380, + 129 + ], + "type": "text", + "content": " into corresponding high-dimensional feature maps " + }, + { + "bbox": [ + 30, + 34, + 380, + 129 + ], + "type": "inline_equation", + "content": "(h_{syn / real}, h_{real})" + }, + { + "bbox": [ + 30, + 34, + 380, + 129 + ], + "type": "text", + "content": ". The input data, " + }, + { + "bbox": [ + 30, + 34, + 380, + 129 + ], + "type": "inline_equation", + "content": "(I_{syn / real}, I_{real})" + }, + { + "bbox": [ + 30, + 34, + 380, + 129 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 34, + 380, + 129 + ], + "type": "inline_equation", + "content": "(D_{syn / real}, D_{real})" + }, + { + "bbox": [ + 30, + 34, + 380, + 129 + ], + "type": "text", + "content": ", are also treated as " + }, + { + "bbox": [ + 30, + 34, + 380, + 129 + ], + "type": "inline_equation", + "content": "(f_{syn / real}, f_{real})" + }, + { + "bbox": [ + 30, + 34, + 380, + 129 + ], + "type": "text", + "content": " during the learning of " + }, + { + "bbox": [ + 30, + 34, + 380, + 129 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 30, + 34, + 380, + 129 + ], + "type": "text", + "content": ". Each of these networks creates a high-dimensional latent space, essentially the Reproducing Kernel Hilbert Space (RKHS) [2] and contributes to the learning of both " + }, + { + "bbox": [ + 30, + 34, + 380, + 129 + ], + "type": "inline_equation", + "content": "M_{rv}" + }, + { + "bbox": [ + 30, + 34, + 380, + 129 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 34, + 380, + 129 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 30, + 34, + 380, + 129 + ], + "type": "text", + "content": " by calculating the MMD [27]. While RKHS has been applied effectively to other DA tasks, to our knowledge, this is the first time that it has been applied to 6DoF PE, and the adapter network architecture is novel." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "spans": [ + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "text", + "content": "The loss function " + }, + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "text", + "content": " is made up of five elements: Radial regression loss " + }, + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "inline_equation", + "content": "V_r" + }, + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "text", + "content": "; Keypoint projection loss " + }, + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_k" + }, + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "text", + "content": "; Classification loss " + }, + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "text", + "content": "; Confidence loss " + }, + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "text", + "content": ", and finally; Adapter loss " + }, + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_A" + }, + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "text", + "content": " for the comparison of intermediate feature maps. The regression losses " + }, + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_k" + }, + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "text", + "content": " all use the smooth L1 metric, whereas classification loss " + }, + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "text", + "content": " uses the cross-entropy metric " + }, + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "inline_equation", + "content": "H(\\cdot)" + }, + { + "bbox": [ + 30, + 130, + 380, + 201 + ], + "type": "text", + "content": ". RKHSPose losses can then be denoted as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 207, + 379, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 207, + 379, + 220 + ], + "spans": [ + { + "bbox": [ + 146, + 207, + 379, + 220 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r} = \\operatorname {s m o o t h} _ {\\mathcal {L} _ {1}} \\left(V _ {r}, \\widehat {V} _ {r}\\right) \\tag {1}", + "image_path": "36eb78f25598203f3fbf502bc805ded5587f8089fa21763e8b1ba43d3649c3f9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 145, + 223, + 379, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 223, + 379, + 236 + ], + "spans": [ + { + "bbox": [ + 145, + 223, + 379, + 236 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {k} = \\operatorname {s m o o t h} _ {\\mathcal {L} _ {1}} (K, \\widehat {K}) \\tag {2}", + "image_path": "1b1f84c9cba12323dce7abc62eef82a71988e5803bc5684d62af1c33c47119d9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 146, + 239, + 379, + 250 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 239, + 379, + 250 + ], + "spans": [ + { + "bbox": [ + 146, + 239, + 379, + 250 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} = H (p (C), \\widehat {p} (C)) \\tag {3}", + "image_path": "fbb5ef9aa16f2026cbfaada233af6e13b1a6e3ab2b0d79f655b881b21ba1dcca.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 146, + 253, + 379, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 253, + 379, + 266 + ], + "spans": [ + { + "bbox": [ + 146, + 253, + 379, + 266 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} = \\operatorname {s m o o t h} _ {\\mathcal {L} _ {1}} (S, \\widehat {S}) \\tag {4}", + "image_path": "d3596666761b4e286f7b7a1922608e5d350f575ccfedf42b0cc65f24155fe5d1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 144, + 269, + 379, + 284 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 269, + 379, + 284 + ], + "spans": [ + { + "bbox": [ + 144, + 269, + 379, + 284 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {A} = M M D \\left(\\widehat {f} _ {\\text {s y n / r e a l}}, \\widehat {f} _ {\\text {r e a l}}\\right) \\tag {5}", + "image_path": "91cd09b60b7e3bb491dc6cb7f8e6e4d8f8633ba687c421d49ddf33e4d15a8986.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 287, + 379, + 298 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 287, + 379, + 298 + ], + "spans": [ + { + "bbox": [ + 115, + 287, + 379, + 298 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {r} \\mathcal {L} _ {r} + \\lambda_ {k} \\mathcal {L} _ {k} + \\lambda_ {c} \\mathcal {L} _ {c} + \\lambda_ {s} \\mathcal {L} _ {s} + \\lambda_ {A} \\mathcal {L} _ {A} \\tag {6}", + "image_path": "a50060711ebfa590884e4a419d3064dc069295417b0cd7be4ed2dbf62a2b7ec6.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 30, + 303, + 379, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 303, + 379, + 375 + ], + "spans": [ + { + "bbox": [ + 30, + 303, + 379, + 375 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 30, + 303, + 379, + 375 + ], + "type": "inline_equation", + "content": "\\lambda_r, \\lambda_k, \\lambda_c, \\lambda_s" + }, + { + "bbox": [ + 30, + 303, + 379, + 375 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 30, + 303, + 379, + 375 + ], + "type": "inline_equation", + "content": "\\lambda_A" + }, + { + "bbox": [ + 30, + 303, + 379, + 375 + ], + "type": "text", + "content": " are weights for adjustment during training, and all non-hatted quantities are GT values. At inference, " + }, + { + "bbox": [ + 30, + 303, + 379, + 375 + ], + "type": "inline_equation", + "content": "M_{rv}" + }, + { + "bbox": [ + 30, + 303, + 379, + 375 + ], + "type": "text", + "content": " takes " + }, + { + "bbox": [ + 30, + 303, + 379, + 375 + ], + "type": "inline_equation", + "content": "I_{real}" + }, + { + "bbox": [ + 30, + 303, + 379, + 375 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 303, + 379, + 375 + ], + "type": "inline_equation", + "content": "D_{real}" + }, + { + "bbox": [ + 30, + 303, + 379, + 375 + ], + "type": "text", + "content": " as input, and outputs " + }, + { + "bbox": [ + 30, + 303, + 379, + 375 + ], + "type": "inline_equation", + "content": "K, C" + }, + { + "bbox": [ + 30, + 303, + 379, + 375 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 303, + 379, + 375 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 30, + 303, + 379, + 375 + ], + "type": "text", + "content": ". The keypoints " + }, + { + "bbox": [ + 30, + 303, + 379, + 375 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 30, + 303, + 379, + 375 + ], + "type": "text", + "content": " are ranked and grouped by " + }, + { + "bbox": [ + 30, + 303, + 379, + 375 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 30, + 303, + 379, + 375 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 303, + 379, + 375 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 30, + 303, + 379, + 375 + ], + "type": "text", + "content": ", and are then forwarded into the ePnP [46] algorithm which estimates 6DoF pose values. ICP can then be optionally applied using the depth data to refine the estimated pose." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 31, + 389, + 211, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 389, + 211, + 402 + ], + "spans": [ + { + "bbox": [ + 31, + 389, + 211, + 402 + ], + "type": "text", + "content": "3.2 Convolutional RKHS Adapter" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 30, + 405, + 379, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 405, + 379, + 465 + ], + "spans": [ + { + "bbox": [ + 30, + 405, + 379, + 465 + ], + "type": "text", + "content": "Reproducing Kernel Hilbert Space " + }, + { + "bbox": [ + 30, + 405, + 379, + 465 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 30, + 405, + 379, + 465 + ], + "type": "text", + "content": " is a commonly used vector space for Domain Adaptation [58, 77]. Hilbert Space is a complete metric space (in which every Cauchy sequence of points has a limit within the metric) represented by the inner product of vectors. For a non-empty set of data " + }, + { + "bbox": [ + 30, + 405, + 379, + 465 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 30, + 405, + 379, + 465 + ], + "type": "text", + "content": ", a function " + }, + { + "bbox": [ + 30, + 405, + 379, + 465 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_{\\mathcal{X}}: \\mathcal{X} \\times \\mathcal{X} \\to \\mathcal{R}" + }, + { + "bbox": [ + 30, + 405, + 379, + 465 + ], + "type": "text", + "content": " is a reproducing kernel if:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 113, + 471, + 379, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 471, + 379, + 502 + ], + "spans": [ + { + "bbox": [ + 113, + 471, + 379, + 502 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} k (\\cdot , x) \\in \\mathcal {H} \\quad \\forall \\quad x \\in \\mathcal {X} \\\\ \\langle f (\\cdot), k (\\cdot , x) \\rangle = f (x) \\quad \\forall \\quad x \\in \\mathcal {X}, f \\in \\mathcal {H} \\end{array} \\right. \\tag {7}", + "image_path": "daeb1130d472f64fab441b68c451361d8d22c63a619425563af1166c197669db.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 30, + 506, + 379, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 506, + 379, + 541 + ], + "spans": [ + { + "bbox": [ + 30, + 506, + 379, + 541 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 30, + 506, + 379, + 541 + ], + "type": "inline_equation", + "content": "\\langle a,b\\rangle" + }, + { + "bbox": [ + 30, + 506, + 379, + 541 + ], + "type": "text", + "content": " denotes the inner product of two vectors " + }, + { + "bbox": [ + 30, + 506, + 379, + 541 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 30, + 506, + 379, + 541 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 506, + 379, + 541 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 30, + 506, + 379, + 541 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 30, + 506, + 379, + 541 + ], + "type": "inline_equation", + "content": "k(\\cdot ,x) = \\mathcal{K}_{\\mathcal{X}}" + }, + { + "bbox": [ + 30, + 506, + 379, + 541 + ], + "type": "text", + "content": " for each " + }, + { + "bbox": [ + 30, + 506, + 379, + 541 + ], + "type": "inline_equation", + "content": "x\\in \\mathcal{X}" + }, + { + "bbox": [ + 30, + 506, + 379, + 541 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 30, + 506, + 379, + 541 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 30, + 506, + 379, + 541 + ], + "type": "text", + "content": " is a function in " + }, + { + "bbox": [ + 30, + 506, + 379, + 541 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 30, + 506, + 379, + 541 + ], + "type": "text", + "content": ". The second equation in Eq. 7 is an expression of the reproducing property of " + }, + { + "bbox": [ + 30, + 506, + 379, + 541 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 30, + 506, + 379, + 541 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 30, + 542, + 380, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 542, + 380, + 566 + ], + "spans": [ + { + "bbox": [ + 30, + 542, + 380, + 566 + ], + "type": "text", + "content": "In order to utilize " + }, + { + "bbox": [ + 30, + 542, + 380, + 566 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 30, + 542, + 380, + 566 + ], + "type": "text", + "content": " for DA, we expand the kernel definition to two sets of data " + }, + { + "bbox": [ + 30, + 542, + 380, + 566 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 30, + 542, + 380, + 566 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 542, + 380, + 566 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 30, + 542, + 380, + 566 + ], + "type": "text", + "content": ". The reproducing kernel can then be defined as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 157, + 572, + 379, + 584 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 572, + 379, + 584 + ], + "spans": [ + { + "bbox": [ + 157, + 572, + 379, + 584 + ], + "type": "interline_equation", + "content": "\\mathcal {K} (\\mathcal {X}, \\mathcal {Y}) = \\left\\langle \\mathcal {K} _ {\\mathcal {X}}, \\mathcal {K} _ {\\mathcal {Y}} \\right\\rangle_ {H} \\tag {8}", + "image_path": "9ce0e37a1887eab226fb78589c14de0213f86e845c87a969a586157d585cbc1c.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 300, + 9, + 346, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 9, + 346, + 19 + ], + "spans": [ + { + "bbox": [ + 300, + 9, + 346, + 19 + ], + "type": "text", + "content": "RKHPose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 30, + 34, + 380, + 117 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 34, + 380, + 117 + ], + "spans": [ + { + "bbox": [ + 30, + 34, + 380, + 117 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 30, + 34, + 380, + 117 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_{\\mathcal{X}}" + }, + { + "bbox": [ + 30, + 34, + 380, + 117 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 34, + 380, + 117 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_{\\mathcal{Y}}" + }, + { + "bbox": [ + 30, + 34, + 380, + 117 + ], + "type": "text", + "content": " are themselves inner product kernels that map " + }, + { + "bbox": [ + 30, + 34, + 380, + 117 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 30, + 34, + 380, + 117 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 34, + 380, + 117 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 30, + 34, + 380, + 117 + ], + "type": "text", + "content": " respectively into their own Hilbert spaces, and " + }, + { + "bbox": [ + 30, + 34, + 380, + 117 + ], + "type": "inline_equation", + "content": "\\mathcal{K}(\\mathcal{X},\\mathcal{Y})" + }, + { + "bbox": [ + 30, + 34, + 380, + 117 + ], + "type": "text", + "content": " is the joint Hilbert space kernel of " + }, + { + "bbox": [ + 30, + 34, + 380, + 117 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 30, + 34, + 380, + 117 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 34, + 380, + 117 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 30, + 34, + 380, + 117 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 30, + 34, + 380, + 117 + ], + "type": "inline_equation", + "content": "\\mathcal{K}(\\mathcal{X},\\mathcal{Y})" + }, + { + "bbox": [ + 30, + 34, + 380, + 117 + ], + "type": "text", + "content": " is not the kernel commonly defined for CNNs. Rather, it is the similarity function defined in kernel methods, such as is used by Support Vector Machine (SVM) techniques to calculate similarity measurements. Some recent methods [8, 52, 53, 56] named it a Convolutional Kernel Network (CKN) to distinguish it from CNNs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 30, + 118, + 380, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 118, + 380, + 214 + ], + "spans": [ + { + "bbox": [ + 30, + 118, + 380, + 214 + ], + "type": "text", + "content": "Various kernels of CKN, such as the Gaussian Kernel [53], the RBF Kernel [52] and the Inner Product Kernel [56], have been shown to be comparable to shallow CNNs for various tasks, especially for Domain Adaption. The most intuitive inner product kernel is mathematically similar to a fully connected layer, where trainable weights are multiplied by the input feature map. To allow the application of RKHS methods into our CNN based Adapter network " + }, + { + "bbox": [ + 30, + 118, + 380, + 214 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 30, + 118, + 380, + 214 + ], + "type": "text", + "content": ", trainable weights are added to " + }, + { + "bbox": [ + 30, + 118, + 380, + 214 + ], + "type": "inline_equation", + "content": "\\mathcal{K}(\\mathcal{X},\\mathcal{Y})" + }, + { + "bbox": [ + 30, + 118, + 380, + 214 + ], + "type": "text", + "content": " [49]. The trainable Kernel " + }, + { + "bbox": [ + 30, + 118, + 380, + 214 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_w(\\mathcal{X},\\mathcal{Y})" + }, + { + "bbox": [ + 30, + 118, + 380, + 214 + ], + "type": "text", + "content": " can then be denoted as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 222, + 379, + 234 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 222, + 379, + 234 + ], + "spans": [ + { + "bbox": [ + 131, + 222, + 379, + 234 + ], + "type": "interline_equation", + "content": "\\mathcal {K} _ {w} (\\mathcal {X}, \\mathcal {Y}) = \\langle \\langle \\mathcal {X}, W _ {\\mathcal {X}} \\rangle , \\langle \\mathcal {Y}, W _ {\\mathcal {Y}} \\rangle \\rangle_ {H} \\tag {9}", + "image_path": "66a69f52376899cbc1ac6f72abcd8b351a80acff5ad10eb2f30bdf30adadc90b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 30, + 240, + 379, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 240, + 379, + 263 + ], + "spans": [ + { + "bbox": [ + 30, + 240, + 379, + 263 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 30, + 240, + 379, + 263 + ], + "type": "inline_equation", + "content": "W_{\\mathcal{X}}" + }, + { + "bbox": [ + 30, + 240, + 379, + 263 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 240, + 379, + 263 + ], + "type": "inline_equation", + "content": "W_{\\mathcal{Y}}" + }, + { + "bbox": [ + 30, + 240, + 379, + 263 + ], + "type": "text", + "content": " are trainable weights. By adding " + }, + { + "bbox": [ + 30, + 240, + 379, + 263 + ], + "type": "inline_equation", + "content": "W_{\\mathcal{X}}" + }, + { + "bbox": [ + 30, + 240, + 379, + 263 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 240, + 379, + 263 + ], + "type": "inline_equation", + "content": "W_{\\mathcal{Y}}" + }, + { + "bbox": [ + 30, + 240, + 379, + 263 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 30, + 240, + 379, + 263 + ], + "type": "inline_equation", + "content": "K_{w}(\\mathcal{X},\\mathcal{Y})" + }, + { + "bbox": [ + 30, + 240, + 379, + 263 + ], + "type": "text", + "content": " still satisfies the RKHS constraints." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 264, + 380, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 264, + 380, + 384 + ], + "spans": [ + { + "bbox": [ + 30, + 264, + 380, + 384 + ], + "type": "text", + "content": "The sim2real domain gap of feature maps " + }, + { + "bbox": [ + 30, + 264, + 380, + 384 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 30, + 264, + 380, + 384 + ], + "type": "text", + "content": " being trained in " + }, + { + "bbox": [ + 30, + 264, + 380, + 384 + ], + "type": "inline_equation", + "content": "M_{rv}" + }, + { + "bbox": [ + 30, + 264, + 380, + 384 + ], + "type": "text", + "content": " (with few real images and no real GT labels) are hard to measure using trivial distance metrics. In contrast, RKHS can be a more accurate and robust space for comparison, since it is known to be capable of handling high-dimensional data with a low number of samples [24, 56]. To compare " + }, + { + "bbox": [ + 30, + 264, + 380, + 384 + ], + "type": "inline_equation", + "content": "(f_{syn / real}, f_{real})" + }, + { + "bbox": [ + 30, + 264, + 380, + 384 + ], + "type": "text", + "content": " in RKHS, a series of CNN layers encodes " + }, + { + "bbox": [ + 30, + 264, + 380, + 384 + ], + "type": "inline_equation", + "content": "(f_{syn / real}, f_{real})" + }, + { + "bbox": [ + 30, + 264, + 380, + 384 + ], + "type": "text", + "content": " into higher-dimensional features " + }, + { + "bbox": [ + 30, + 264, + 380, + 384 + ], + "type": "inline_equation", + "content": "(h_{syn / real}, h_{real})" + }, + { + "bbox": [ + 30, + 264, + 380, + 384 + ], + "type": "text", + "content": ", followed by the trainable " + }, + { + "bbox": [ + 30, + 264, + 380, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_w(\\mathcal{X}, \\mathcal{Y})" + }, + { + "bbox": [ + 30, + 264, + 380, + 384 + ], + "type": "text", + "content": ". Once mapped into RKHS, " + }, + { + "bbox": [ + 30, + 264, + 380, + 384 + ], + "type": "inline_equation", + "content": "h_{syn / real} = \\{sr_i\\}_{i=1}^m" + }, + { + "bbox": [ + 30, + 264, + 380, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 264, + 380, + 384 + ], + "type": "inline_equation", + "content": "h_{real} = \\{r_i\\}_{i=1}^m" + }, + { + "bbox": [ + 30, + 264, + 380, + 384 + ], + "type": "text", + "content": " can then be measured by Maximum Mean Discrepancy (MMD), a common DA measurement [27, 49], which is the square distance between the kernel embedding [26]:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 34, + 391, + 384, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 34, + 391, + 384, + 480 + ], + "spans": [ + { + "bbox": [ + 34, + 391, + 384, + 480 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} M M D \\left(h _ {s y n / r e a l}, h _ {r e a l}\\right) = \\frac {1}{m} \\left[ \\left(\\sum_ {i = 1} ^ {m} \\sum_ {j = 1} ^ {m} k _ {w} \\left(s r _ {i}, s r _ {j}\\right) - \\sum_ {i = 1} ^ {m} k _ {w} \\left(s r _ {i}, s r _ {i}\\right)\\right) \\right. \\\\ \\left. - \\left(\\sum_ {i = 1} ^ {m} \\sum_ {j = 1} ^ {m} k _ {w} \\left(s r _ {i}, r _ {j}\\right) - \\sum_ {i = 1} ^ {m} k _ {w} \\left(s r _ {i}, r _ {i}\\right)\\right) + \\left(\\sum_ {i = 1} ^ {m} \\sum_ {j = 1} ^ {m} k _ {w} \\left(r _ {i}, r _ {j}\\right) - \\sum_ {i = 1} ^ {m} k _ {w} \\left(r _ {i}, r _ {i}\\right)\\right) \\right] ^ {\\frac {1}{2}} \\tag {10} \\\\ \\end{array}", + "image_path": "c275c7898cd29d07128aa6bc9f57244c9aae8a339a4e9292bfd6f80e5fcaded9.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 31, + 488, + 219, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 488, + 219, + 499 + ], + "spans": [ + { + "bbox": [ + 31, + 488, + 219, + 499 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 31, + 488, + 219, + 499 + ], + "type": "inline_equation", + "content": "k_{w}()" + }, + { + "bbox": [ + 31, + 488, + 219, + 499 + ], + "type": "text", + "content": " is the feature element of " + }, + { + "bbox": [ + 31, + 488, + 219, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_w(" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "type": "text", + "content": "In summary, the Adapter " + }, + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "type": "text", + "content": " shown in Fig. 2 measures MMD for each " + }, + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "type": "inline_equation", + "content": "(h_{syn / real}, h_{real})" + }, + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "type": "text", + "content": " in RKHS, by increasing the " + }, + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "type": "inline_equation", + "content": "f_{syn / real}" + }, + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "type": "inline_equation", + "content": "f_{real}" + }, + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "type": "text", + "content": " dimension using a CNN and thereby constructing a learnable kernel " + }, + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_w" + }, + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "type": "text", + "content": ". The outputs are the feature maps " + }, + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "type": "inline_equation", + "content": "h_{syn / real}" + }, + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "type": "inline_equation", + "content": "h_{real}" + }, + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "type": "text", + "content": ", which are supervised by loss " + }, + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_A" + }, + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "type": "text", + "content": " during training of the real data epochs. Based on the experiments in Sec. 5.3, our trainable inner product kernel is shown to be more accurate for our task than other known kernels [52, 53, 56] that we tested, that are often used in such kernel methods." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 32, + 10, + 39, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 10, + 39, + 18 + ], + "spans": [ + { + "bbox": [ + 32, + 10, + 39, + 18 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 168, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 168, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 168, + 20 + ], + "type": "text", + "content": "Y.Wu and M. Greenspan" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 33, + 226, + 46 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 33, + 226, + 46 + ], + "spans": [ + { + "bbox": [ + 32, + 33, + 226, + 46 + ], + "type": "text", + "content": "3.3 Keypoint Radial Voting Network" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 32, + 54, + 379, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 54, + 379, + 208 + ], + "spans": [ + { + "bbox": [ + 32, + 54, + 379, + 208 + ], + "type": "text", + "content": "The network " + }, + { + "bbox": [ + 32, + 54, + 379, + 208 + ], + "type": "inline_equation", + "content": "M_v" + }, + { + "bbox": [ + 32, + 54, + 379, + 208 + ], + "type": "text", + "content": " votes for keypoints using a CNN architecture, taking the radial voting quantity " + }, + { + "bbox": [ + 32, + 54, + 379, + 208 + ], + "type": "inline_equation", + "content": "V_r" + }, + { + "bbox": [ + 32, + 54, + 379, + 208 + ], + "type": "text", + "content": " resulting from " + }, + { + "bbox": [ + 32, + 54, + 379, + 208 + ], + "type": "inline_equation", + "content": "M_r" + }, + { + "bbox": [ + 32, + 54, + 379, + 208 + ], + "type": "text", + "content": " as input. VoteNet [63] previously used a CNN approach to vote for object centers, whereas other keypoint-based techniques have implemented GPU-based parallel RANSAC [32, 61] methods for offset and vector quantities. The radial quantity is known to be more accurate than the vector or offset quantities, and has been previously implemented with a CPU-based parallel accumulator space method [82, 83]. Given its superior accuracy, " + }, + { + "bbox": [ + 32, + 54, + 379, + 208 + ], + "type": "inline_equation", + "content": "M_v" + }, + { + "bbox": [ + 32, + 54, + 379, + 208 + ], + "type": "text", + "content": " implements radial voting using a CNN to improve efficiency. Given a 2D radial map " + }, + { + "bbox": [ + 32, + 54, + 379, + 208 + ], + "type": "inline_equation", + "content": "\\widehat{V}_r" + }, + { + "bbox": [ + 32, + 54, + 379, + 208 + ], + "type": "text", + "content": " estimated by " + }, + { + "bbox": [ + 32, + 54, + 379, + 208 + ], + "type": "inline_equation", + "content": "M_r" + }, + { + "bbox": [ + 32, + 54, + 379, + 208 + ], + "type": "text", + "content": " and supervised by GT radial maps " + }, + { + "bbox": [ + 32, + 54, + 379, + 208 + ], + "type": "inline_equation", + "content": "V_r" + }, + { + "bbox": [ + 32, + 54, + 379, + 208 + ], + "type": "text", + "content": ", the task is to accumulate votes, find the peak, and estimate the keypoint location. The " + }, + { + "bbox": [ + 32, + 54, + 379, + 208 + ], + "type": "inline_equation", + "content": "V_r" + }, + { + "bbox": [ + 32, + 54, + 379, + 208 + ], + "type": "text", + "content": " foreground pixels (which lie on the target object) store the Euclidean distance from these pixels to each of the keypoints, with background (non-object) pixels set to value -1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "spans": [ + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "text", + "content": "The estimated radial map " + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "inline_equation", + "content": "\\widehat{V}_r" + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "text", + "content": " is indeed an inverse heat map of the candidate keypoints' locations, distributed in a radial pattern centered at the keypoints. To forward " + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "inline_equation", + "content": "\\widehat{V}_r" + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "text", + "content": " into a CNN voting module, it is inversely normalized so that it becomes a heat map. Let " + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "inline_equation", + "content": "v_r^{max}" + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "inline_equation", + "content": "v_r^{min}" + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "text", + "content": " be the maximum and minimum global radial distances for all objects in a dataset, which can be calculated by iterating through all GT radial maps, or alternately generated from the object CAD models. An inverse radial map " + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "inline_equation", + "content": "\\widehat{V}_r^{-1}" + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "text", + "content": " can then be denoted as " + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "inline_equation", + "content": "\\widehat{V}_r^{-1} = (v_r^{max} - \\widehat{V}_r) / (v_r^{max} - v_r^{min})" + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "text", + "content": ". Voting network " + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "inline_equation", + "content": "M_v" + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "text", + "content": " takes " + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "inline_equation", + "content": "\\widehat{V}_r^{-1}" + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "text", + "content": " as the input and generates the accumulated vote map by a series of convolution, ReLu, and batch normalization layers. The complete network architecture is provided in the Supplementary material Sec. S.2. The background pixels are filtered out by a ReLu layer, and only foreground pixels contribute to voting. The accumulated vote map is then max-pooled for peak extraction, and reshaped using a fully connected layer into a " + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "inline_equation", + "content": "n\\times 4" + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "text", + "content": " output. The output represents " + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "text", + "content": " keypoints and comprises " + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "inline_equation", + "content": "n\\times 2" + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "text", + "content": " projected 2D keypoints " + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "text", + "content": " classification labels " + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "text", + "content": " confidence scores " + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "text", + "content": ". The labels " + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "text", + "content": " indicate which object the corresponding keypoints belongs to, and " + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "text", + "content": " ranks the confidence level of keypoints before being forwarded into ePnP [46] for pose estimation. " + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "inline_equation", + "content": "M_v" + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "text", + "content": " is supervised by " + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_k" + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "text", + "content": " (Eqs. 2-4) and is trained end-to-end along with " + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "inline_equation", + "content": "M_r" + }, + { + "bbox": [ + 32, + 209, + 379, + 437 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 32, + 455, + 128, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 455, + 128, + 469 + ], + "spans": [ + { + "bbox": [ + 32, + 455, + 128, + 469 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 32, + 480, + 224, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 480, + 224, + 491 + ], + "spans": [ + { + "bbox": [ + 32, + 480, + 224, + 491 + ], + "type": "text", + "content": "4.1 Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 32, + 499, + 379, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 499, + 379, + 584 + ], + "spans": [ + { + "bbox": [ + 32, + 499, + 379, + 584 + ], + "type": "text", + "content": "RKHSPose uses BOP Procedural Blender [17] (PBR) synthetic images [16, 36, 37, 72] for the synthetic training phase. The images are generated by dropping synthetic objects (CAD models) onto a plane in a simulated environment using PyBullet [13], and then rendering them with synthetic textures. All objects in the synthetic images are thus automatically labeled with precise GT poses. We evaluated RKHSPose for the six BOP [37, 72] core datasets (LMO [34], YCB [84], TLESS [35], TUDL [36], ITODO [20], and HB [40]), all except IC-BIN [19]," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 300, + 9, + 346, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 9, + 346, + 19 + ], + "spans": [ + { + "bbox": [ + 300, + 9, + 346, + 19 + ], + "type": "text", + "content": "RKHPose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 33, + 379, + 81 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 33, + 379, + 81 + ], + "spans": [ + { + "bbox": [ + 32, + 33, + 379, + 81 + ], + "type": "text", + "content": "which does not include any real training or validation images and is therefore not applicable. ITODO and HB have no real images in the training set, and so for training we instead used the real images in their validation sets, which were disjoint from their test sets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 32, + 83, + 379, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 83, + 379, + 203 + ], + "spans": [ + { + "bbox": [ + 32, + 83, + 379, + 203 + ], + "type": "text", + "content": "Our main results are evaluated with the ADD(S) [34] metric for the LM and LMO dataset, and the ADD(S) AUC [84] metric for the YCB dataset. These are the standard metrics commonly used to compare self-supervised 6DoF PE methods. ADD(S) is based on the mean distance (minimum distance for symmetry) between the object surfaces for GT and estimated poses, whereas ADD(S) AUC plots a curve formed by ADD(S) for various object diameter thresholds. We use the BOP average recall (" + }, + { + "bbox": [ + 32, + 83, + 379, + 203 + ], + "type": "inline_equation", + "content": "AR" + }, + { + "bbox": [ + 32, + 83, + 379, + 203 + ], + "type": "text", + "content": ") metrics for our ablation studies. The " + }, + { + "bbox": [ + 32, + 83, + 379, + 203 + ], + "type": "inline_equation", + "content": "AR" + }, + { + "bbox": [ + 32, + 83, + 379, + 203 + ], + "type": "text", + "content": " metric, based on the original ADD(S) [34], evaluates three aspects, including Visible Surface Discrepancy (" + }, + { + "bbox": [ + 32, + 83, + 379, + 203 + ], + "type": "inline_equation", + "content": "AR_{VSD}" + }, + { + "bbox": [ + 32, + 83, + 379, + 203 + ], + "type": "text", + "content": "), Maximum Symmetry-Aware Surface/Projection Distance (" + }, + { + "bbox": [ + 32, + 83, + 379, + 203 + ], + "type": "inline_equation", + "content": "AR_{MSSD}" + }, + { + "bbox": [ + 32, + 83, + 379, + 203 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 32, + 83, + 379, + 203 + ], + "type": "inline_equation", + "content": "AR_{MSPD}" + }, + { + "bbox": [ + 32, + 83, + 379, + 203 + ], + "type": "text", + "content": ") [36]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 223, + 178, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 223, + 178, + 235 + ], + "spans": [ + { + "bbox": [ + 33, + 223, + 178, + 235 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 246, + 379, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 246, + 379, + 306 + ], + "spans": [ + { + "bbox": [ + 33, + 246, + 379, + 306 + ], + "type": "text", + "content": "RKHSPose is trained on a server with an Intel Xeon 5218 CPU and two RTX6000 GPUs with a batch size of 32. The Adam optimizer is used for the training of " + }, + { + "bbox": [ + 33, + 246, + 379, + 306 + ], + "type": "inline_equation", + "content": "M_{rv}" + }, + { + "bbox": [ + 33, + 246, + 379, + 306 + ], + "type": "text", + "content": ", on both synthetic and real data, and " + }, + { + "bbox": [ + 33, + 246, + 379, + 306 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 33, + 246, + 379, + 306 + ], + "type": "text", + "content": " is optimized by SGD. Both of the optimizers have an initial learning rate of " + }, + { + "bbox": [ + 33, + 246, + 379, + 306 + ], + "type": "inline_equation", + "content": "lr = 1e - 3" + }, + { + "bbox": [ + 33, + 246, + 379, + 306 + ], + "type": "text", + "content": " and weight decay 1e-4 for 80 and 20 epochs respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 32, + 307, + 379, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 307, + 379, + 367 + ], + "spans": [ + { + "bbox": [ + 32, + 307, + 379, + 367 + ], + "type": "text", + "content": "The input of the network is normalized before training, as follows. The RGB images " + }, + { + "bbox": [ + 32, + 307, + 379, + 367 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 32, + 307, + 379, + 367 + ], + "type": "text", + "content": " are normalized and standardized using ImageNet [14] specifications. The depth maps are each individually normalized by their local minima and maxima, to lie within a range of 0 to 1. The radial distances in radial map " + }, + { + "bbox": [ + 32, + 307, + 379, + 367 + ], + "type": "inline_equation", + "content": "V_{r}" + }, + { + "bbox": [ + 32, + 307, + 379, + 367 + ], + "type": "text", + "content": " and 2D projected keypoints are both normalized by the width and height of " + }, + { + "bbox": [ + 32, + 307, + 379, + 367 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 32, + 307, + 379, + 367 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 32, + 367, + 379, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 367, + 379, + 451 + ], + "spans": [ + { + "bbox": [ + 32, + 367, + 379, + 451 + ], + "type": "text", + "content": "A single set of four keypoints is chosen by KeyGNet [81] for the set of all objects in each dataset. One extra background class is added to " + }, + { + "bbox": [ + 32, + 367, + 379, + 451 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 32, + 367, + 379, + 451 + ], + "type": "text", + "content": " in order to filter out the redundant background points in " + }, + { + "bbox": [ + 32, + 367, + 379, + 451 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 32, + 367, + 379, + 451 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 32, + 367, + 379, + 451 + ], + "type": "inline_equation", + "content": "M_{rv}" + }, + { + "bbox": [ + 32, + 367, + 379, + 451 + ], + "type": "text", + "content": " is first trained for 120 epochs on synthetic data, during which " + }, + { + "bbox": [ + 32, + 367, + 379, + 451 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 32, + 367, + 379, + 451 + ], + "type": "text", + "content": " remains frozen. Following this, training proceeds for an additional 80 epochs which alternate between real and synthetic data. When training on real data, both " + }, + { + "bbox": [ + 32, + 367, + 379, + 451 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 32, + 367, + 379, + 451 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 32, + 367, + 379, + 451 + ], + "type": "inline_equation", + "content": "M_{rv}" + }, + { + "bbox": [ + 32, + 367, + 379, + 451 + ], + "type": "text", + "content": " weights are learned, whereas " + }, + { + "bbox": [ + 32, + 367, + 379, + 451 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 32, + 367, + 379, + 451 + ], + "type": "text", + "content": " is frozen for the alternating synthetic data training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 33, + 452, + 379, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 452, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 452, + 379, + 583 + ], + "type": "text", + "content": "During real data epochs, " + }, + { + "bbox": [ + 33, + 452, + 379, + 583 + ], + "type": "inline_equation", + "content": "M_{rv}" + }, + { + "bbox": [ + 33, + 452, + 379, + 583 + ], + "type": "text", + "content": " initially estimates pseudo-keypoints for each real image. These pseudo-keypoints are then forwarded into ePnP for pseudopose estimation. Each estimated pseudo-pose is then augmented into a set of poses " + }, + { + "bbox": [ + 33, + 452, + 379, + 583 + ], + "type": "inline_equation", + "content": "P_{aug}" + }, + { + "bbox": [ + 33, + 452, + 379, + 583 + ], + "type": "text", + "content": " by applying arbitrary rotational and translational perturbations with respective ranges of " + }, + { + "bbox": [ + 33, + 452, + 379, + 583 + ], + "type": "inline_equation", + "content": "[-\\frac{\\pi}{18}, \\frac{\\pi}{18}]" + }, + { + "bbox": [ + 33, + 452, + 379, + 583 + ], + "type": "text", + "content": " radians and " + }, + { + "bbox": [ + 33, + 452, + 379, + 583 + ], + "type": "inline_equation", + "content": "[-0.1, 0.1]" + }, + { + "bbox": [ + 33, + 452, + 379, + 583 + ], + "type": "text", + "content": " along three axes within the normalized model frame, which is defined using the largest object in the dataset. The set of syn/real images are rendered by overlaying onto the real image the CAD model of each object using each augmented pose value in " + }, + { + "bbox": [ + 33, + 452, + 379, + 583 + ], + "type": "inline_equation", + "content": "P_{aug}" + }, + { + "bbox": [ + 33, + 452, + 379, + 583 + ], + "type": "text", + "content": ". The cardinality of " + }, + { + "bbox": [ + 33, + 452, + 379, + 583 + ], + "type": "inline_equation", + "content": "P_{aug}" + }, + { + "bbox": [ + 33, + 452, + 379, + 583 + ], + "type": "text", + "content": " is set to be one less than the batch size, and the adapter " + }, + { + "bbox": [ + 33, + 452, + 379, + 583 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 33, + 452, + 379, + 583 + ], + "type": "text", + "content": " is trained on a mini-batch of the hybrid images resulting from " + }, + { + "bbox": [ + 33, + 452, + 379, + 583 + ], + "type": "inline_equation", + "content": "P_{aug}" + }, + { + "bbox": [ + 33, + 452, + 379, + 583 + ], + "type": "text", + "content": ", plus the image resulting from the original estimated pseudo-pose." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 168, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 168, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 168, + 20 + ], + "type": "text", + "content": "Y.Wu and M. Greenspan" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 65, + 338, + 359 + ], + "blocks": [ + { + "bbox": [ + 30, + 21, + 380, + 55 + ], + "lines": [ + { + "bbox": [ + 30, + 21, + 380, + 55 + ], + "spans": [ + { + "bbox": [ + 30, + 21, + 380, + 55 + ], + "type": "text", + "content": "Table 2: Comparison with other methods. Accuracy of RKHSPose for LM and LMO is evaluated with ADD(S), and for YCB is evaluated with ADD(S) AUC. All 'Supervision: Syn + Self' methods use real images without real labels." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 65, + 338, + 359 + ], + "lines": [ + { + "bbox": [ + 75, + 65, + 338, + 359 + ], + "spans": [ + { + "bbox": [ + 75, + 65, + 338, + 359 + ], + "type": "table", + "html": "
MethodReal data image labelDataset/Metric
LM ADD(S)LMOYCB ADD(S)ADD-S AUC
Supervision: Syn (lower bound)
AAEXX31.4---
MHPXX38.8---
GDR (TexPose version)XX77.452.9--
Self6D++XX77.452.977.889.4
Self6D++ with DrefXX88.062.579.290.1
OursXX78.254.376.590.2
Ours+ICPXX87.955.778.391.3
Supervision: Syn + Self
Sock et al.✓X60.622.8--
DSC✓X58.624.8--
Self6D✓X58.932.1--
SMOC-Net✓X91.363.3--
Self6D++✓X88.564.780.091.4
TexPose✓X91.766.7--
Ours✓X95.868.682.892.4
Ours+ICP✓X95.968.783.092.6
Supervision: Syn + Real GT (upper bound)
SO-Pose✓✓96.062.383.990.9
Self6D++✓✓91.074.482.690.7
Ours✓✓96.770.885.492.2
Ours+ICP✓✓96.871.385.692.4
", + "image_path": "fa28ed9b27e25c67b517edceaa2e90a1ad1899ac0d3d9154aeff3176ac2af7cf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 30, + 391, + 380, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 391, + 380, + 488 + ], + "spans": [ + { + "bbox": [ + 30, + 391, + 380, + 488 + ], + "type": "text", + "content": "Initially " + }, + { + "bbox": [ + 30, + 391, + 380, + 488 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 30, + 391, + 380, + 488 + ], + "type": "text", + "content": " is frozen, and for the first 80 epochs, the loss is set to emphasize the classification and the visibility score regression, i.e. " + }, + { + "bbox": [ + 30, + 391, + 380, + 488 + ], + "type": "inline_equation", + "content": "\\lambda_c = \\lambda_s = 0.6" + }, + { + "bbox": [ + 30, + 391, + 380, + 488 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 391, + 380, + 488 + ], + "type": "inline_equation", + "content": "\\lambda_r = \\lambda_k = 0.4" + }, + { + "bbox": [ + 30, + 391, + 380, + 488 + ], + "type": "text", + "content": ". Following this, up to epoch 200, the scales of losses are then exchanged to fine-tune the localization of the keypoints, i.e. " + }, + { + "bbox": [ + 30, + 391, + 380, + 488 + ], + "type": "inline_equation", + "content": "\\lambda_c = \\lambda_s = 0.4" + }, + { + "bbox": [ + 30, + 391, + 380, + 488 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 391, + 380, + 488 + ], + "type": "inline_equation", + "content": "\\lambda_r = \\lambda_k = 0.6" + }, + { + "bbox": [ + 30, + 391, + 380, + 488 + ], + "type": "text", + "content": ". After epoch 120, " + }, + { + "bbox": [ + 30, + 391, + 380, + 488 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 30, + 391, + 380, + 488 + ], + "type": "text", + "content": " is unfrozen each alternating epoch, and " + }, + { + "bbox": [ + 30, + 391, + 380, + 488 + ], + "type": "inline_equation", + "content": "\\lambda_D" + }, + { + "bbox": [ + 30, + 391, + 380, + 488 + ], + "type": "text", + "content": " is set to 1 during the remaining " + }, + { + "bbox": [ + 30, + 391, + 380, + 488 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 30, + 391, + 380, + 488 + ], + "type": "text", + "content": " training epochs. This training strategy, shown in Fig. 2, minimizes the sim2real gap without any real image GT labels, and using very few (320) real images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 31, + 504, + 98, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 504, + 98, + 515 + ], + "spans": [ + { + "bbox": [ + 31, + 504, + 98, + 515 + ], + "type": "text", + "content": "4.3 Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 523, + 381, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 523, + 381, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 523, + 381, + 584 + ], + "type": "text", + "content": "The results are summarized in Tables 2 and 3. To our knowledge, RKHSPose outperforms all existing self-supervised 6DoF PE methods. In Table 2, the upper bound fully supervised, lower bound synthetically supervised, and middle self-supervised methods (including ours) are compared. On LM and LMO, our ADD(S) is " + }, + { + "bbox": [ + 30, + 523, + 381, + 584 + ], + "type": "inline_equation", + "content": "+4.2\\%" + }, + { + "bbox": [ + 30, + 523, + 381, + 584 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 523, + 381, + 584 + ], + "type": "inline_equation", + "content": "+2\\%" + }, + { + "bbox": [ + 30, + 523, + 381, + 584 + ], + "type": "text", + "content": " better than the second best method TexPose [9]. We" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 300, + 9, + 347, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 9, + 347, + 19 + ], + "spans": [ + { + "bbox": [ + 300, + 9, + 347, + 19 + ], + "type": "text", + "content": "RKHPose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 369, + 10, + 378, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 10, + 378, + 19 + ], + "spans": [ + { + "bbox": [ + 369, + 10, + 378, + 19 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 54, + 87, + 358, + 183 + ], + "blocks": [ + { + "bbox": [ + 31, + 33, + 380, + 78 + ], + "lines": [ + { + "bbox": [ + 31, + 33, + 380, + 78 + ], + "spans": [ + { + "bbox": [ + 31, + 33, + 380, + 78 + ], + "type": "text", + "content": "Table 3: Comparison with fully supervised methods. RKHSPose results on TLESS " + }, + { + "bbox": [ + 31, + 33, + 380, + 78 + ], + "type": "inline_equation", + "content": "(-1.8)" + }, + { + "bbox": [ + 31, + 33, + 380, + 78 + ], + "type": "text", + "content": ", TUDL " + }, + { + "bbox": [ + 31, + 33, + 380, + 78 + ], + "type": "inline_equation", + "content": "(-0.4)" + }, + { + "bbox": [ + 31, + 33, + 380, + 78 + ], + "type": "text", + "content": ", ITODD " + }, + { + "bbox": [ + 31, + 33, + 380, + 78 + ], + "type": "inline_equation", + "content": "(-4.6)" + }, + { + "bbox": [ + 31, + 33, + 380, + 78 + ], + "type": "text", + "content": " and HB " + }, + { + "bbox": [ + 31, + 33, + 380, + 78 + ], + "type": "inline_equation", + "content": "(+0.1)" + }, + { + "bbox": [ + 31, + 33, + 380, + 78 + ], + "type": "text", + "content": " compares to SOTA methods with full supervision of real GT labels. Methods annotated with * use the detection results from other detection methods." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 54, + 87, + 358, + 183 + ], + "lines": [ + { + "bbox": [ + 54, + 87, + 358, + 183 + ], + "spans": [ + { + "bbox": [ + 54, + 87, + 358, + 183 + ], + "type": "table", + "html": "
Methodreal labelLMOTLESSDataset
TUDLITODDHBYCB
SurfEmb* [30]✓-76.082.885.465.986.679.9
RCVPose3D [82]✓-72.970.896.673.386.384.3
RADet [87]+PFA* [39]✓-79.785.096.067.686.988.8
ZebraPose [70]✓-78.086.295.665.492.189.9
Ours✗95.768.285.596.268.692.283.6
Ours+ICP✗95.868.485.696.268.792.383.8
", + "image_path": "190cbb24832e94571ceba9401c6ee62fe09632621598dbcfc72cca519347751d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 32, + 203, + 90, + 262 + ], + "blocks": [ + { + "bbox": [ + 32, + 203, + 90, + 262 + ], + "lines": [ + { + "bbox": [ + 32, + 203, + 90, + 262 + ], + "spans": [ + { + "bbox": [ + 32, + 203, + 90, + 262 + ], + "type": "image", + "image_path": "12f1a860821259363b9ce9e8fa2e3d636edb840ef5654196f8ff86ed9ae4784b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 30, + 270, + 380, + 293 + ], + "lines": [ + { + "bbox": [ + 30, + 270, + 380, + 293 + ], + "spans": [ + { + "bbox": [ + 30, + 270, + 380, + 293 + ], + "type": "text", + "content": "Fig. 3: Qualitative overlay results on selected images. Red dots and blue dots are projected surface points from GT poses and estimated poses, respectively." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 90, + 203, + 147, + 261 + ], + "blocks": [ + { + "bbox": [ + 90, + 203, + 147, + 261 + ], + "lines": [ + { + "bbox": [ + 90, + 203, + 147, + 261 + ], + "spans": [ + { + "bbox": [ + 90, + 203, + 147, + 261 + ], + "type": "image", + "image_path": "c8e452a90ee70b8162806bcab95a05890b7ae4d8358c542d23f3e244ab86431b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 149, + 204, + 205, + 261 + ], + "blocks": [ + { + "bbox": [ + 149, + 204, + 205, + 261 + ], + "lines": [ + { + "bbox": [ + 149, + 204, + 205, + 261 + ], + "spans": [ + { + "bbox": [ + 149, + 204, + 205, + 261 + ], + "type": "image", + "image_path": "243dda54444d5ca87002df56e5e61697d016711f6e09f3e84d9a5c175c79b206.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 207, + 204, + 263, + 261 + ], + "blocks": [ + { + "bbox": [ + 207, + 204, + 263, + 261 + ], + "lines": [ + { + "bbox": [ + 207, + 204, + 263, + 261 + ], + "spans": [ + { + "bbox": [ + 207, + 204, + 263, + 261 + ], + "type": "image", + "image_path": "222bb3aa03cc0d98fdda13e3782932a1008bbceab48df717e16cfb32bb20e597.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 265, + 203, + 321, + 261 + ], + "blocks": [ + { + "bbox": [ + 265, + 203, + 321, + 261 + ], + "lines": [ + { + "bbox": [ + 265, + 203, + 321, + 261 + ], + "spans": [ + { + "bbox": [ + 265, + 203, + 321, + 261 + ], + "type": "image", + "image_path": "6c80e3d267335e01caa2245aa89098bd693bb1be49d82798c7582dae71fc694e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 322, + 203, + 379, + 261 + ], + "blocks": [ + { + "bbox": [ + 322, + 203, + 379, + 261 + ], + "lines": [ + { + "bbox": [ + 322, + 203, + 379, + 261 + ], + "spans": [ + { + "bbox": [ + 322, + 203, + 379, + 261 + ], + "type": "image", + "image_path": "9256f29dad8d7bd713e1dd1603d26852183ecbdc77afac167ec1918370b65e59.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 30, + 326, + 379, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 326, + 379, + 386 + ], + "spans": [ + { + "bbox": [ + 30, + 326, + 379, + 386 + ], + "type": "text", + "content": "compared our performance to Self6D++ [79], which was the only other method that evaluated using YCB, and saw a " + }, + { + "bbox": [ + 30, + 326, + 379, + 386 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 30, + 326, + 379, + 386 + ], + "type": "text", + "content": " improvement after ICP on ADD(S) AUC. Last but not least, in Table 3 our performance evaluated on the other four BOP core datasets is comparable to several upper bound methods from the BOP leaderboard. Some test scenes with RKHSPose results are shown in Fig. 3." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 30, + 386, + 379, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 386, + 379, + 434 + ], + "spans": [ + { + "bbox": [ + 30, + 386, + 379, + 434 + ], + "type": "text", + "content": "RKHSPose runs at 34 fps on an Intel i7 2.5GHz CPU and an RTX 3090 GPU with 24G VRAM. It takes on average " + }, + { + "bbox": [ + 30, + 386, + 379, + 434 + ], + "type": "inline_equation", + "content": "8.7\\mathrm{ms}" + }, + { + "bbox": [ + 30, + 386, + 379, + 434 + ], + "type": "text", + "content": " for loading data, " + }, + { + "bbox": [ + 30, + 386, + 379, + 434 + ], + "type": "inline_equation", + "content": "4.5\\mathrm{ms}" + }, + { + "bbox": [ + 30, + 386, + 379, + 434 + ], + "type": "text", + "content": " for forward inference through " + }, + { + "bbox": [ + 30, + 386, + 379, + 434 + ], + "type": "inline_equation", + "content": "M_{rv}" + }, + { + "bbox": [ + 30, + 386, + 379, + 434 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 30, + 386, + 379, + 434 + ], + "type": "inline_equation", + "content": "\\times 10" + }, + { + "bbox": [ + 30, + 386, + 379, + 434 + ], + "type": "text", + "content": " faster compared to the analytical radial voting in RCVPose [83]), and " + }, + { + "bbox": [ + 30, + 386, + 379, + 434 + ], + "type": "inline_equation", + "content": "16.2\\mathrm{ms}" + }, + { + "bbox": [ + 30, + 386, + 379, + 434 + ], + "type": "text", + "content": " for ePnP." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 31, + 453, + 153, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 453, + 153, + 466 + ], + "spans": [ + { + "bbox": [ + 31, + 453, + 153, + 466 + ], + "type": "text", + "content": "5 Ablation Studies" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 31, + 479, + 192, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 479, + 192, + 492 + ], + "spans": [ + { + "bbox": [ + 31, + 479, + 192, + 492 + ], + "type": "text", + "content": "5.1 Dense Vs. Sparse Adapter" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "type": "text", + "content": "The Adapter " + }, + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "type": "text", + "content": " densely matches the intermediate feature maps, whereas the majority of other methods [15,75,80] only compare the final output. To show the benefits of dense comparison, we conduct an experiment with different variations of " + }, + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "type": "text", + "content": ". A sparse matching " + }, + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "type": "inline_equation", + "content": "M_A^s" + }, + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "type": "text", + "content": " network is trained on synthetic and real data comparing only a single feature map, which is the intermediate radial map. " + }, + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "type": "inline_equation", + "content": "M_A^s" + }, + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "type": "text", + "content": " has the exact same overall learning capacity (number of parameters) as the dense matching " + }, + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "type": "text", + "content": " described in Sec. 3.3. The results in Table 4 show that " + }, + { + "bbox": [ + 30, + 499, + 380, + 584 + ], + "type": "inline_equation", + "content": "M_A" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 168, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 168, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 168, + 20 + ], + "type": "text", + "content": "Y.Wu and M. Greenspan" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 32, + 65, + 197, + 105 + ], + "blocks": [ + { + "bbox": [ + 31, + 33, + 197, + 54 + ], + "lines": [ + { + "bbox": [ + 31, + 33, + 197, + 54 + ], + "spans": [ + { + "bbox": [ + 31, + 33, + 197, + 54 + ], + "type": "text", + "content": "Table 4: " + }, + { + "bbox": [ + 31, + 33, + 197, + 54 + ], + "type": "inline_equation", + "content": "{AR}" + }, + { + "bbox": [ + 31, + 33, + 197, + 54 + ], + "type": "text", + "content": " of different adapters on LM and five BOP core datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 32, + 65, + 197, + 105 + ], + "lines": [ + { + "bbox": [ + 32, + 65, + 197, + 105 + ], + "spans": [ + { + "bbox": [ + 32, + 65, + 197, + 105 + ], + "type": "table", + "html": "
AdapterARVSDARMSSDARMSPDAR
MA78.177.877.877.9
MA84.984.184.384.4
", + "image_path": "d57edd734d3cfa5a4ddbe47c39c2f2bde889c92dcee02a1fab6bbb07c8cf2ec7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 215, + 65, + 380, + 100 + ], + "blocks": [ + { + "bbox": [ + 215, + 33, + 379, + 55 + ], + "lines": [ + { + "bbox": [ + 215, + 33, + 379, + 55 + ], + "spans": [ + { + "bbox": [ + 215, + 33, + 379, + 55 + ], + "type": "text", + "content": "Table 5: " + }, + { + "bbox": [ + 215, + 33, + 379, + 55 + ], + "type": "inline_equation", + "content": "{AR}" + }, + { + "bbox": [ + 215, + 33, + 379, + 55 + ], + "type": "text", + "content": " of different training strategies on LM and five BOP core datasets." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 215, + 65, + 380, + 100 + ], + "lines": [ + { + "bbox": [ + 215, + 65, + 380, + 100 + ], + "spans": [ + { + "bbox": [ + 215, + 65, + 380, + 100 + ], + "type": "table", + "html": "
Training TypeARVSDARMSSDARMSPDAR
Mixed84.984.184.384.4
Sequential82.381.781.781.9
", + "image_path": "4feef9e91556a63b42dd09d1b8e61331cf7fc64fac546cc6319ca7b65e6394b9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 32, + 147, + 201, + 207 + ], + "blocks": [ + { + "bbox": [ + 31, + 115, + 366, + 137 + ], + "lines": [ + { + "bbox": [ + 31, + 115, + 366, + 137 + ], + "spans": [ + { + "bbox": [ + 31, + 115, + 366, + 137 + ], + "type": "text", + "content": "Table 6: [R Tab A] " + }, + { + "bbox": [ + 31, + 115, + 366, + 137 + ], + "type": "inline_equation", + "content": "AR" + }, + { + "bbox": [ + 31, + 115, + 366, + 137 + ], + "type": "text", + "content": " of different kernel- Table 7: " + }, + { + "bbox": [ + 31, + 115, + 366, + 137 + ], + "type": "inline_equation", + "content": "AR" + }, + { + "bbox": [ + 31, + 115, + 366, + 137 + ], + "type": "text", + "content": " of different metrics on LM nets on LM and five BOP core datasets. and five BOP core datasets." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 32, + 147, + 201, + 207 + ], + "lines": [ + { + "bbox": [ + 32, + 147, + 201, + 207 + ], + "spans": [ + { + "bbox": [ + 32, + 147, + 201, + 207 + ], + "type": "table", + "html": "
KernelwARVSDARMSSDARMSPDAR
LinearX71.670.870.671.0
✓84.984.184.384.4
RBFX73.472.973.273.2
✓82.581.381.581.6
", + "image_path": "c2b16c73cd552ab1dc078e1c46882079172f0ff6f42b96029ff4648b7b326565.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 201, + 147, + 366, + 198 + ], + "blocks": [ + { + "bbox": [ + 201, + 147, + 366, + 198 + ], + "lines": [ + { + "bbox": [ + 201, + 147, + 366, + 198 + ], + "spans": [ + { + "bbox": [ + 201, + 147, + 366, + 198 + ], + "type": "table", + "html": "
MetricARVSDARMSSDARMSPDAR
MMD84.984.184.384.4
KL Div78.077.878.077.9
Wass80.980.680.980.8
", + "image_path": "08167831f2a1dae82367e2665a0220538f06fbf6e04a300f420d95688fc0a174.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 30, + 230, + 380, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 230, + 380, + 267 + ], + "spans": [ + { + "bbox": [ + 30, + 230, + 380, + 267 + ], + "type": "text", + "content": "surpassed " + }, + { + "bbox": [ + 30, + 230, + 380, + 267 + ], + "type": "inline_equation", + "content": "M_A^s" + }, + { + "bbox": [ + 30, + 230, + 380, + 267 + ], + "type": "text", + "content": " on all six datasets tested. Specifically, on ITODD, " + }, + { + "bbox": [ + 30, + 230, + 380, + 267 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 30, + 230, + 380, + 267 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 30, + 230, + 380, + 267 + ], + "type": "inline_equation", + "content": "12.1\\%" + }, + { + "bbox": [ + 30, + 230, + 380, + 267 + ], + "type": "text", + "content": " more accurate than " + }, + { + "bbox": [ + 30, + 230, + 380, + 267 + ], + "type": "inline_equation", + "content": "M_A^s" + }, + { + "bbox": [ + 30, + 230, + 380, + 267 + ], + "type": "text", + "content": ". This experiment shows the effectiveness of our densely matched " + }, + { + "bbox": [ + 30, + 230, + 380, + 267 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 30, + 230, + 380, + 267 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 31, + 285, + 224, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 285, + 224, + 298 + ], + "spans": [ + { + "bbox": [ + 31, + 285, + 224, + 298 + ], + "type": "text", + "content": "5.2 Syn/Real Synchronized Training" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 30, + 307, + 379, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 307, + 379, + 380 + ], + "spans": [ + { + "bbox": [ + 30, + 307, + 379, + 380 + ], + "type": "text", + "content": "When training RKHSPose, real epochs are alternated with synthetic epochs. In contrast, some other methods [9, 75, 79, 80] separate the synthetic/real training. We conducted an experiment to compare these two different training strategies, the results of which are shown in Table 5. The alternating training performs slightly better (" + }, + { + "bbox": [ + 30, + 307, + 379, + 380 + ], + "type": "inline_equation", + "content": "+2.5\\%" + }, + { + "bbox": [ + 30, + 307, + 379, + 380 + ], + "type": "text", + "content": " on average) than the sequential training, possibly due to the early access to real scenes thereby avoiding local minima." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 31, + 399, + 207, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 399, + 207, + 411 + ], + "spans": [ + { + "bbox": [ + 31, + 399, + 207, + 411 + ], + "type": "text", + "content": "5.3 Adapter Kernels and Metrics" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 30, + 420, + 379, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 420, + 379, + 491 + ], + "spans": [ + { + "bbox": [ + 30, + 420, + 379, + 491 + ], + "type": "text", + "content": "We use a linear (dot product) kernel and MMD in RKHS for domain gap measurements. There are various other kernels and similarity measurements that can be implemented in RKHS as described in Sec. 3.2. First, we add trainable weights to the radial basis function (RBF) kernel in a similar manner as " + }, + { + "bbox": [ + 30, + 420, + 379, + 491 + ], + "type": "inline_equation", + "content": "K_{w}" + }, + { + "bbox": [ + 30, + 420, + 379, + 491 + ], + "type": "text", + "content": " defined in Eq. 9. The trainable RBF kernel on two sets of data " + }, + { + "bbox": [ + 30, + 420, + 379, + 491 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 30, + 420, + 379, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 420, + 379, + 491 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 30, + 420, + 379, + 491 + ], + "type": "text", + "content": " is denoted as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 491, + 379, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 491, + 379, + 506 + ], + "spans": [ + { + "bbox": [ + 131, + 491, + 379, + 506 + ], + "type": "interline_equation", + "content": "K _ {r b f} (X, Y) = \\exp (- w \\| X - Y \\| ^ {2}) \\tag {11}", + "image_path": "cb0936685adbbb7ca793bbf553f852436da1faea3de29ab78eaf19cfbd3a7e2b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 30, + 511, + 379, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 511, + 379, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 511, + 379, + 584 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 30, + 511, + 379, + 584 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 30, + 511, + 379, + 584 + ], + "type": "text", + "content": " are the trainable weights, which replaces the original adjustable parameter in the classical RBF kernel. We also experiment with the classical RKHS kernel functions without trainable weights, including the inner product kernel and the original RBF kernel [78], for comparison. Further, we experiment on other commonly used distance measures, including Kullback-Leibler Divergence (KL Div, i.e. relative-entropy) and Wasserstein (Wass) Distance." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 300, + 9, + 346, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 9, + 346, + 19 + ], + "spans": [ + { + "bbox": [ + 300, + 9, + 346, + 19 + ], + "type": "text", + "content": "RKHPose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 31, + 328, + 161 + ], + "blocks": [ + { + "bbox": [ + 83, + 31, + 328, + 161 + ], + "lines": [ + { + "bbox": [ + 83, + 31, + 328, + 161 + ], + "spans": [ + { + "bbox": [ + 83, + 31, + 328, + 161 + ], + "type": "image", + "image_path": "85734d7328823939c2c8a4cace4cb00d1ae19d2781561399992531721a03abc4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 30, + 167, + 381, + 202 + ], + "lines": [ + { + "bbox": [ + 30, + 167, + 381, + 202 + ], + "spans": [ + { + "bbox": [ + 30, + 167, + 381, + 202 + ], + "type": "text", + "content": "Fig. 4: Impact of # of real images with/without GT labels used during training. All datasets are evaluated by the BOP AR metric. We conduct experiments from 0 to 640 real images on all datasets, except ITODD which contained only 357 real images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 30, + 233, + 381, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 233, + 381, + 295 + ], + "spans": [ + { + "bbox": [ + 30, + 233, + 381, + 295 + ], + "type": "text", + "content": "In Table 6, the RBF kernel performs similar to the linear product kernel with a slight performance dip. In Table 7, MMD minimizes the domain gap better than the Wass Distance, followed by the KL Div metric, leading to a better overall performance on " + }, + { + "bbox": [ + 30, + 233, + 381, + 295 + ], + "type": "inline_equation", + "content": "AR" + }, + { + "bbox": [ + 30, + 233, + 381, + 295 + ], + "type": "text", + "content": ". Based on these results, we used the linear product kernel with trainable weights and selected MMD as the main loss metric." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 31, + 309, + 265, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 309, + 265, + 322 + ], + "spans": [ + { + "bbox": [ + 31, + 309, + 265, + 322 + ], + "type": "text", + "content": "5.4 Number of Real Images and Real Labels" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 327, + 382, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 327, + 382, + 436 + ], + "spans": [ + { + "bbox": [ + 30, + 327, + 382, + 436 + ], + "type": "text", + "content": "The objective of RKHSPose is to reduce real data usage and train without any real GT labels. To show the effectiveness of the approach, we conducted an experiment by training on different numbers of real images, the results of which are shown in Fig. 4. We used up to 640 real images in all cases, except for that of ITODD which contains only 357 real images. The " + }, + { + "bbox": [ + 30, + 327, + 382, + 436 + ], + "type": "inline_equation", + "content": "AR" + }, + { + "bbox": [ + 30, + 327, + 382, + 436 + ], + "type": "text", + "content": " of all datasets saturates at 160 images except YCB. The further improvement of YCB beyond 160 images is also only " + }, + { + "bbox": [ + 30, + 327, + 382, + 436 + ], + "type": "inline_equation", + "content": "+0.1\\%" + }, + { + "bbox": [ + 30, + 327, + 382, + 436 + ], + "type": "text", + "content": " and saturates after 320 real images. We nevertheless use 320 real unlabeled images for our main results. This experiment showed that adding more than 320 real labeled images did not significantly improve performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 31, + 452, + 120, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 452, + 120, + 466 + ], + "spans": [ + { + "bbox": [ + 31, + 452, + 120, + 466 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 30, + 476, + 382, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 476, + 382, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 476, + 382, + 584 + ], + "type": "text", + "content": "To sum up, we propose a novel self-supervised keypoint radial voting-based 6DoF PE method using RGB-D data called RKHSPose. RKHSPose fine-tunes poses pre-trained on synthetic data by densely matching features with a learnable kernel in RKHS, using real data albeit without any real GT poses. By applying this DA technique in feature space, RKHSPose achieved SOTA performance on the six applicable BOP core datasets, surpassing the performance of all other self-supervised methods. Notably, the RKHSPose performance closely approaches that of several fully-supervised methods, which indicates the strength of the approach at reducing the sim2real domain gap for this problem." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 168, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 168, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 168, + 20 + ], + "type": "text", + "content": "Y.Wu and M. Greenspan" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 31, + 34, + 380, + 59 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 34, + 380, + 59 + ], + "spans": [ + { + "bbox": [ + 31, + 34, + 380, + 59 + ], + "type": "text", + "content": "Acknowledgements: Thanks to Bluewrist Inc. and NSERC for their support of this work." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 32, + 77, + 98, + 89 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 77, + 98, + 89 + ], + "spans": [ + { + "bbox": [ + 32, + 77, + 98, + 89 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 101, + 380, + 583 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 36, + 101, + 379, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 101, + 379, + 134 + ], + "spans": [ + { + "bbox": [ + 36, + 101, + 379, + 134 + ], + "type": "text", + "content": "1. Al Safadi, E., Song, X.: Learning-based image registration with meta-regularization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10928-10937 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 36, + 135, + 379, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 135, + 379, + 156 + ], + "spans": [ + { + "bbox": [ + 36, + 135, + 379, + 156 + ], + "type": "text", + "content": "2. Aronszajn, N.: Theory of reproducing kernels. Transactions of the American mathematical society 68(3), 337-404 (1950)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 37, + 157, + 380, + 189 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 157, + 380, + 189 + ], + "spans": [ + { + "bbox": [ + 37, + 157, + 380, + 189 + ], + "type": "text", + "content": "3. Besl, P.J., McKay, N.D.: Method for registration of 3-d shapes. In: Sensor fusion IV: control paradigms and data structures. vol. 1611, pp. 586-606. International Society for Optics and Photonics (1992)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 37, + 190, + 379, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 190, + 379, + 222 + ], + "spans": [ + { + "bbox": [ + 37, + 190, + 379, + 222 + ], + "type": "text", + "content": "4. Bietti, A., Mairal, J.: Group invariance, stability to deformations, and complexity of deep convolutional representations. The Journal of Machine Learning Research 20(1), 876-924 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 37, + 223, + 380, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 223, + 380, + 255 + ], + "spans": [ + { + "bbox": [ + 37, + 223, + 380, + 255 + ], + "type": "text", + "content": "5. Bietti, A., Mialon, G., Chen, D., Mairal, J.: A kernel perspective for regularizing deep neural networks. In: International Conference on Machine Learning. pp. 664-674. PMLR (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 37, + 256, + 379, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 256, + 379, + 277 + ], + "spans": [ + { + "bbox": [ + 37, + 256, + 379, + 277 + ], + "type": "text", + "content": "6. Bozorgtabar, B., Mahapatra, D., Thiran, J.P.: Exprada: Adversarial domain adaptation for facial expression analysis. Pattern Recognition 100, 107111 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 37, + 277, + 380, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 277, + 380, + 309 + ], + "spans": [ + { + "bbox": [ + 37, + 277, + 380, + 309 + ], + "type": "text", + "content": "7. Chen, C., Fu, Z., Chen, Z., Jin, S., Cheng, Z., Jin, X., Hua, X.S.: Homm: Higher-order moment matching for unsupervised domain adaptation. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 3422-3429 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 37, + 310, + 379, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 310, + 379, + 343 + ], + "spans": [ + { + "bbox": [ + 37, + 310, + 379, + 343 + ], + "type": "text", + "content": "8. Chen, D., Jacob, L., Mairal, J.: Convolutional kernel networks for graph-structured data. In: International Conference on Machine Learning. pp. 1576–1586. PMLR (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 37, + 343, + 379, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 343, + 379, + 375 + ], + "spans": [ + { + "bbox": [ + 37, + 343, + 379, + 375 + ], + "type": "text", + "content": "9. Chen, H., Manhardt, F., Navab, N., Busam, B.: Texpose: Neural texture learning for self-supervised 6d object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4841-4852 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 33, + 376, + 379, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 376, + 379, + 397 + ], + "spans": [ + { + "bbox": [ + 33, + 376, + 379, + 397 + ], + "type": "text", + "content": "10. Cheng, X., Xie, Y.: Neural tangent kernel maximum mean discrepancy. Advances in Neural Information Processing Systems 34, 6658-6670 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 33, + 397, + 378, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 397, + 378, + 419 + ], + "spans": [ + { + "bbox": [ + 33, + 397, + 378, + 419 + ], + "type": "text", + "content": "11. Corcoran, P.: An end-to-end graph convolutional kernel support vector machine. Applied Network Science 5(1), 1-15 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 33, + 420, + 379, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 420, + 379, + 440 + ], + "spans": [ + { + "bbox": [ + 33, + 420, + 379, + 440 + ], + "type": "text", + "content": "12. Cortes, C., Vapnik, V.: Support-vector networks. Machine learning 20, 273-297 (1995)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 33, + 441, + 379, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 441, + 379, + 463 + ], + "spans": [ + { + "bbox": [ + 33, + 441, + 379, + 463 + ], + "type": "text", + "content": "13. Coumans, E., Bai, Y.: Pybullet, a python module for physics simulation for games, robotics and machine learning. http://pybullet.org (2016-2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 33, + 464, + 379, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 464, + 379, + 506 + ], + "spans": [ + { + "bbox": [ + 33, + 464, + 379, + 506 + ], + "type": "text", + "content": "14. Deng, J., Dong, W., Socher, R., Li, L., Kai Li, Li Fei-Fei: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. pp. 248-255 (2009). https://doi.org/10.1109/CVPR.2009.5206848" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 33, + 507, + 379, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 507, + 379, + 539 + ], + "spans": [ + { + "bbox": [ + 33, + 507, + 379, + 539 + ], + "type": "text", + "content": "15. Deng, X., Xiang, Y., Mousavian, A., Eppner, C., Bretl, T., Fox, D.: Self-supervised 6d object pose estimation for robot manipulation. In: 2020 IEEE International Conference on Robotics and Automation (ICRA). pp. 3665-3671. IEEE (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 33, + 540, + 379, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 540, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 540, + 379, + 583 + ], + "type": "text", + "content": "16. Denninger, M., Sundermeyer, M., Winkelbauer, D., Olefir, D., Hodan, T., Zidan, Y., Elbadrawy, M., Knauer, M., Katam, H., Lodhi, A.: Blenderproc: Reducing the reality gap with photorealistic rendering. In: International Conference on Robotics: Science and Systems, RSS 2020 (2020)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 300, + 10, + 346, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 10, + 346, + 19 + ], + "spans": [ + { + "bbox": [ + 300, + 10, + 346, + 19 + ], + "type": "text", + "content": "RKHPose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 33, + 35, + 379, + 583 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 33, + 35, + 379, + 79 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 35, + 379, + 79 + ], + "spans": [ + { + "bbox": [ + 33, + 35, + 379, + 79 + ], + "type": "text", + "content": "17. Denninger, M., Winkelbauer, D., Sundermeyer, M., Boerdijk, W., Knauer, M., Strobl, K.H., Humt, M., Triebel, R.: Blenderproc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software 8(82), 4901 (2023). https://doi.org/10.21105/joss.04901, https://doi.org/10.21105/joss.04901" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 33, + 79, + 379, + 122 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 79, + 379, + 122 + ], + "spans": [ + { + "bbox": [ + 33, + 79, + 379, + 122 + ], + "type": "text", + "content": "18. Di, Y., Manhardt, F., Wang, G., Ji, X., Navab, N., Tombari, F.: So-posed: Exploiting self-occlusion for direct 6d pose estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 12396-12405 (October 2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 122, + 379, + 154 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 122, + 379, + 154 + ], + "spans": [ + { + "bbox": [ + 33, + 122, + 379, + 154 + ], + "type": "text", + "content": "19. Doumanoglou, A., Kouskouridas, R., Malassiotis, S., Kim, T.K.: Recovering 6d object pose and predicting next-best-view in the crowd. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3583-3592 (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 154, + 379, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 154, + 379, + 186 + ], + "spans": [ + { + "bbox": [ + 33, + 154, + 379, + 186 + ], + "type": "text", + "content": "20. Drost, B., Ulrich, M., Bergmann, P., Hartinger, P., Steger, C.: Introducing mvtec itodd-a dataset for 3d object recognition in industry. In: Proceedings of the IEEE international conference on computer vision workshops. pp. 2200-2208 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 186, + 378, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 186, + 378, + 207 + ], + "spans": [ + { + "bbox": [ + 33, + 186, + 378, + 207 + ], + "type": "text", + "content": "21. Gadwe, A., Ren, H.: Real-time 6 dof pose estimation of endoscopic instruments using printable markers. IEEE Sensors Journal 19(6), 2338-2346 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 207, + 379, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 207, + 379, + 239 + ], + "spans": [ + { + "bbox": [ + 33, + 207, + 379, + 239 + ], + "type": "text", + "content": "22. Gao, X., Hou, X., Tang, J., Cheng, H.: Complete solution classification for the perspective-three-point problem. IEEE Trans. Pattern Anal. Mach. Intell. 25, 930-943 (2003)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 33, + 239, + 379, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 239, + 379, + 271 + ], + "spans": [ + { + "bbox": [ + 33, + 239, + 379, + 271 + ], + "type": "text", + "content": "23. Ghojogh, B., Ghodsi, A., Karray, F., Crowley, M.: Reproducing kernel hilbert space, mercer's theorem, eigenfunctions, nystr\\'' om method, and use of kernels in machine learning: Tutorial and survey. arXiv preprint arXiv:2106.08443 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 33, + 271, + 379, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 271, + 379, + 304 + ], + "spans": [ + { + "bbox": [ + 33, + 271, + 379, + 304 + ], + "type": "text", + "content": "24. Ghorbani, B., Mei, S., Misiakiewicz, T., Montanari, A.: When do neural networks outperform kernel methods? Advances in Neural Information Processing Systems 33, 14820-14830 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 33, + 304, + 379, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 304, + 379, + 336 + ], + "spans": [ + { + "bbox": [ + 33, + 304, + 379, + 336 + ], + "type": "text", + "content": "25. Greene, N., Luo, W., Kazanzides, P.: dvpose: Automated data collection and dataset for 6d pose estimation of robotic surgical instruments. In: 2023 International Symposium on Medical Robotics (ISMR). pp. 1-7. IEEE (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 33, + 336, + 379, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 336, + 379, + 368 + ], + "spans": [ + { + "bbox": [ + 33, + 336, + 379, + 368 + ], + "type": "text", + "content": "26. Gretton, A., Borgwardt, K., Rasch, M., Scholkopf, B., Smola, A.: A kernel method for the two-sample-problem. Advances in neural information processing systems 19 (2006)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 33, + 368, + 379, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 368, + 379, + 389 + ], + "spans": [ + { + "bbox": [ + 33, + 368, + 379, + 389 + ], + "type": "text", + "content": "27. Gretton, A., Borgwardt, K.M., Rasch, M.J., Scholkopf, B., Smola, A.: A kernel two-sample test. The Journal of Machine Learning Research 13(1), 723-773 (2012)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 33, + 389, + 379, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 389, + 379, + 432 + ], + "spans": [ + { + "bbox": [ + 33, + 389, + 379, + 432 + ], + "type": "text", + "content": "28. Guo, S., Hu, Y., Alvarez, J.M., Salzmann, M.: Knowledge distillation for 6d pose estimation by aligning distributions of local predictions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18633-18642 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 33, + 432, + 379, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 432, + 379, + 465 + ], + "spans": [ + { + "bbox": [ + 33, + 432, + 379, + 465 + ], + "type": "text", + "content": "29. Hai, Y., Song, R., Li, J., Salzmann, M., Hu, Y.: Rigidity-aware detection for 6d object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8927-8936 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 33, + 465, + 379, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 465, + 379, + 507 + ], + "spans": [ + { + "bbox": [ + 33, + 465, + 379, + 507 + ], + "type": "text", + "content": "30. Haugaard, R.L., Buch, A.G.: Surfemb: Dense and continuous correspondence distributions for object pose estimation with learnt surface embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6749-6758 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 33, + 507, + 379, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 507, + 379, + 540 + ], + "spans": [ + { + "bbox": [ + 33, + 507, + 379, + 540 + ], + "type": "text", + "content": "31. He, Y., Huang, H., Fan, H., Chen, Q., Sun, J.: Ffb6d: A full flow bidirectional fusion network for 6d pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3003-3013 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 33, + 540, + 379, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 540, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 540, + 379, + 583 + ], + "type": "text", + "content": "32. He, Y., Sun, W., Huang, H., Liu, J., Fan, H., Sun, J.: Pvn3d: A deep pointwise 3d keypoints voting network for 6 dof pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 168, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 168, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 168, + 20 + ], + "type": "text", + "content": "Y.Wu and M. Greenspan" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 31, + 35, + 379, + 583 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 31, + 35, + 379, + 68 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 35, + 379, + 68 + ], + "spans": [ + { + "bbox": [ + 31, + 35, + 379, + 68 + ], + "type": "text", + "content": "33. He, Y., Wang, Y., Fan, H., Sun, J., Chen, Q.: Fs6d: Few-shot 6d pose estimation of novel objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6814-6824 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 31, + 69, + 379, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 69, + 379, + 112 + ], + "spans": [ + { + "bbox": [ + 31, + 69, + 379, + 112 + ], + "type": "text", + "content": "34. Hinterstoisser, S., Lepetit, V., Ilic, S., Holzer, S., Bradski, G., Konolige, K., Navab, N.: Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In: Asian conference on computer vision. pp. 548-562. Springer (2012)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 32, + 113, + 379, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 113, + 379, + 146 + ], + "spans": [ + { + "bbox": [ + 32, + 113, + 379, + 146 + ], + "type": "text", + "content": "35. Hodañ, T., Haluza, P., Obdržálek, Š., Matas, J., Lourakis, M., Zabulis, X.: T-LESS: An RGB-D dataset for 6D pose estimation of texture-less objects. IEEE Winter Conference on Applications of Computer Vision (WACV) (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 32, + 148, + 379, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 148, + 379, + 191 + ], + "spans": [ + { + "bbox": [ + 32, + 148, + 379, + 191 + ], + "type": "text", + "content": "36. Hodan, T., Michel, F., Brachmann, E., Kehl, W., GlentBuch, A., Kraft, D., Drost, B., Vidal, J., Ihrke, S., Zabulis, X., et al.: Bop: Benchmark for 6d object pose estimation. In: Proceedings of the European conference on computer vision (ECCV). pp. 19-34 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 32, + 192, + 379, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 192, + 379, + 225 + ], + "spans": [ + { + "bbox": [ + 32, + 192, + 379, + 225 + ], + "type": "text", + "content": "37. Hodač, T., Sundermeyer, M., Drost, B., Labbé, Y., Brachmann, E., Michel, F., Rother, C., Matas, J.: Bop challenge 2020 on 6d object localization. In: European Conference on Computer Vision. pp. 577-594. Springer (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 32, + 226, + 379, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 226, + 379, + 247 + ], + "spans": [ + { + "bbox": [ + 32, + 226, + 379, + 247 + ], + "type": "text", + "content": "38. Horn, B.K., Hilden, H.M., Negahdaripour, S.: Closed-form solution of absolute orientation using orthonormal matrices. JOSA A 5(7), 1127-1135 (1988)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 32, + 249, + 379, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 249, + 379, + 281 + ], + "spans": [ + { + "bbox": [ + 32, + 249, + 379, + 281 + ], + "type": "text", + "content": "39. Hu, Y., Fua, P., Salzmann, M.: Perspective flow aggregation for data-limited 6d object pose estimation. In: European Conference on Computer Vision. pp. 89-106. Springer (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 32, + 282, + 379, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 282, + 379, + 314 + ], + "spans": [ + { + "bbox": [ + 32, + 282, + 379, + 314 + ], + "type": "text", + "content": "40. Kaskman, R., Zakharov, S., Shugurov, I., Ilic, S.: Homebreweddb: RGB-d dataset for 6d pose estimation of 3d objects. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. pp. 0-0 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 32, + 316, + 379, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 316, + 379, + 337 + ], + "spans": [ + { + "bbox": [ + 32, + 316, + 379, + 337 + ], + "type": "text", + "content": "41. Khosravi, M., Smith, R.S.: The existence and uniqueness of solutions for kernel-based system identification. Automatica 148, 110728 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 32, + 338, + 379, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 338, + 379, + 370 + ], + "spans": [ + { + "bbox": [ + 32, + 338, + 379, + 370 + ], + "type": "text", + "content": "42. Kleeberger, K., Huber, M.F.: Single shot 6d object pose estimation. In: 2020 IEEE International Conference on Robotics and Automation (ICRA). pp. 6239-6245. IEEE (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 32, + 372, + 379, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 372, + 379, + 404 + ], + "spans": [ + { + "bbox": [ + 32, + 372, + 379, + 404 + ], + "type": "text", + "content": "43. Kleeberger, K., Landgraf, C., Huber, M.F.: Large-scale 6d object pose estimation dataset for industrial bin-picking. In: 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 2573-2578. IEEE (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 32, + 405, + 379, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 405, + 379, + 449 + ], + "spans": [ + { + "bbox": [ + 32, + 405, + 379, + 449 + ], + "type": "text", + "content": "44. Kleeberger, K., Völk, M., Bormann, R., Huber, M.F.: Investigations on output parameterizations of neural networks for single shot 6d object pose estimation. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 13916-13922. IEEE (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 32, + 450, + 379, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 450, + 379, + 493 + ], + "spans": [ + { + "bbox": [ + 32, + 450, + 379, + 493 + ], + "type": "text", + "content": "45. Lee, T., Lee, B.U., Shin, I., Choe, J., Shin, U., Kweon, I.S., Yoon, K.J.: Udacope: unsupervised domain adaptation for category-level object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14891-14900 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 32, + 495, + 379, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 495, + 379, + 516 + ], + "spans": [ + { + "bbox": [ + 32, + 495, + 379, + 516 + ], + "type": "text", + "content": "46. Lepetit, V., Moreno-Noguer, F., Fua, P.: Ep n p: An accurate o (n) solution to the p n p problem. International journal of computer vision 81, 155-166 (2009)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 32, + 517, + 379, + 549 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 517, + 379, + 549 + ], + "spans": [ + { + "bbox": [ + 32, + 517, + 379, + 549 + ], + "type": "text", + "content": "47. Li, Z., Hu, Y., Salzmann, M., Ji, X.: Sd-pose: Semantic decomposition for cross-domain 6d object pose estimation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 2020–2028 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 32, + 551, + 379, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 551, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 32, + 551, + 379, + 583 + ], + "type": "text", + "content": "48. Lin, J., Wei, Z., Ding, C., Jia, K.: Category-level 6d object pose and size estimation using self-supervised deep prior deformation networks. In: European Conference on Computer Vision. pp. 19-34. Springer (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 300, + 9, + 347, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 9, + 347, + 19 + ], + "spans": [ + { + "bbox": [ + 300, + 9, + 347, + 19 + ], + "type": "text", + "content": "RKHPose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 35, + 379, + 583 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 32, + 35, + 379, + 68 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 35, + 379, + 68 + ], + "spans": [ + { + "bbox": [ + 32, + 35, + 379, + 68 + ], + "type": "text", + "content": "49. Liu, F., Xu, W., Lu, J., Zhang, G., Gretton, A., Sutherland, D.J.: Learning deep kernels for non-parametric two-sample tests. In: International conference on machine learning. pp. 6316-6326. PMLR (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 32, + 68, + 379, + 100 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 68, + 379, + 100 + ], + "spans": [ + { + "bbox": [ + 32, + 68, + 379, + 100 + ], + "type": "text", + "content": "50. Luo, Y.W., Ren, C.X.: Conditional bures metric for domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13989-13998 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 101, + 379, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 101, + 379, + 145 + ], + "spans": [ + { + "bbox": [ + 33, + 101, + 379, + 145 + ], + "type": "text", + "content": "51. Ma, X., Wang, Z., Li, H., Zhang, P., Ouyang, W., Fan, X.: Accurate monocular 3d object detection via color-embedded 3d reconstruction for autonomous driving. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6851-6860 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 145, + 379, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 145, + 379, + 167 + ], + "spans": [ + { + "bbox": [ + 33, + 145, + 379, + 167 + ], + "type": "text", + "content": "52. Mairal, J.: End-to-end kernel learning with supervised convolutional kernel networks. Advances in neural information processing systems 29 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 167, + 378, + 189 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 167, + 378, + 189 + ], + "spans": [ + { + "bbox": [ + 33, + 167, + 378, + 189 + ], + "type": "text", + "content": "53. Mairal, J., Koniusz, P., Harchaoui, Z., Schmid, C.: Convolutional kernel networks. Advances in neural information processing systems 27 (2014)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 189, + 379, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 189, + 379, + 233 + ], + "spans": [ + { + "bbox": [ + 33, + 189, + 379, + 233 + ], + "type": "text", + "content": "54. Manhardt, F., Arroyo, D.M., Rupprecht, C., Busam, B., Birdal, T., Navab, N., Tombari, F.: Explaining the ambiguity of object detection and 6d pose from visual data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 33, + 233, + 379, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 233, + 379, + 265 + ], + "spans": [ + { + "bbox": [ + 33, + 233, + 379, + 265 + ], + "type": "text", + "content": "55. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 33, + 266, + 379, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 266, + 379, + 298 + ], + "spans": [ + { + "bbox": [ + 33, + 266, + 379, + 298 + ], + "type": "text", + "content": "56. Misiakiewicz, T., Mei, S.: Learning with convolution and pooling operations in kernel methods. Advances in Neural Information Processing Systems 35, 29014-29025 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 33, + 298, + 379, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 298, + 379, + 331 + ], + "spans": [ + { + "bbox": [ + 33, + 298, + 379, + 331 + ], + "type": "text", + "content": "57. Oberweger, M., Rad, M., Lepetit, V.: Making deep heatmaps robust to partial occlusions for 3d object pose estimation. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 119-134 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 33, + 331, + 379, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 331, + 379, + 353 + ], + "spans": [ + { + "bbox": [ + 33, + 331, + 379, + 353 + ], + "type": "text", + "content": "58. Pan, S.J., Tsang, I.W., Kwok, J.T., Yang, Q.: Domain adaptation via transfer component analysis. IEEE transactions on neural networks 22(2), 199-210 (2010)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 33, + 354, + 379, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 354, + 379, + 397 + ], + "spans": [ + { + "bbox": [ + 33, + 354, + 379, + 397 + ], + "type": "text", + "content": "59. Park, K., Mousavian, A., Xiang, Y., Fox, D.: Latentfusion: End-to-end differentiable reconstruction and rendering for unseen object pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10710-10719 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 33, + 397, + 379, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 397, + 379, + 429 + ], + "spans": [ + { + "bbox": [ + 33, + 397, + 379, + 429 + ], + "type": "text", + "content": "60. Pearson, K.: Liii. on lines and planes of closest fit to systems of points in space. The London, Edinburgh, and Dublin philosophical magazine and journal of science 2(11), 559-572 (1901)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 33, + 430, + 379, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 430, + 379, + 463 + ], + "spans": [ + { + "bbox": [ + 33, + 430, + 379, + 463 + ], + "type": "text", + "content": "61. Peng, S., Liu, Y., Huang, Q., Zhou, X., Bao, H.: Pynet: Pixel-wise voting network for 6dof pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4561-4570 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 33, + 463, + 379, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 463, + 379, + 496 + ], + "spans": [ + { + "bbox": [ + 33, + 463, + 379, + 496 + ], + "type": "text", + "content": "62. Pinheiro, P.O., Rostamzadeh, N., Ahn, S.: Domain-adaptive single-view 3d reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7638-7647 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 33, + 496, + 379, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 496, + 379, + 528 + ], + "spans": [ + { + "bbox": [ + 33, + 496, + 379, + 528 + ], + "type": "text", + "content": "63. Qi, C.R., Litany, O., He, K., Guibas, L.J.: Deep hough voting for 3d object detection in point clouds. In: proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9277-9286 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 33, + 529, + 379, + 561 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 529, + 379, + 561 + ], + "spans": [ + { + "bbox": [ + 33, + 529, + 379, + 561 + ], + "type": "text", + "content": "64. Sejdinovic, D., Sriperumbudur, B., Gretton, A., Fukumizu, K.: Equivalence of distance-based and rkhs-based statistics in hypothesis testing. The annals of statistics pp. 2263-2291 (2013)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 33, + 562, + 379, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 562, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 562, + 379, + 583 + ], + "type": "text", + "content": "65. Shan, P., Bi, Y., Li, Z., Wang, Q., He, Z., Zhao, Y., Peng, S.: Unsupervised model adaptation for multivariate calibration by domain adaptation-regularization based" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 168, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 168, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 168, + 20 + ], + "type": "text", + "content": "Y.Wu and M. Greenspan" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 31, + 35, + 379, + 583 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 49, + 35, + 379, + 57 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 35, + 379, + 57 + ], + "spans": [ + { + "bbox": [ + 49, + 35, + 379, + 57 + ], + "type": "text", + "content": "kernel partial least square. Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 292, 122418 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 31, + 57, + 379, + 90 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 57, + 379, + 90 + ], + "spans": [ + { + "bbox": [ + 31, + 57, + 379, + 90 + ], + "type": "text", + "content": "66. Shugurov, I., Li, F., Busam, B., Ilic, S.: Osop: A multi-stage one shot object pose estimation framework. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6835-6844 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 31, + 90, + 379, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 90, + 379, + 123 + ], + "spans": [ + { + "bbox": [ + 31, + 90, + 379, + 123 + ], + "type": "text", + "content": "67. Shugurov, I., Zakharov, S., Ilic, S.: Dpodv2: Dense correspondence-based 6 dof pose estimation. IEEE transactions on pattern analysis and machine intelligence 44(11), 7417-7435 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 32, + 123, + 379, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 123, + 379, + 167 + ], + "spans": [ + { + "bbox": [ + 32, + 123, + 379, + 167 + ], + "type": "text", + "content": "68. Sock, J., Garcia-Hernando, G., Armagan, A., Kim, T.K.: Introducing pose consistency and warp-alignment for self-supervised 6d object pose estimation in color images. In: 2020 International Conference on 3D Vision (3DV). pp. 291-300. IEEE (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 32, + 167, + 379, + 199 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 167, + 379, + 199 + ], + "spans": [ + { + "bbox": [ + 32, + 167, + 379, + 199 + ], + "type": "text", + "content": "69. Su, H., Qi, C.R., Li, Y., Guibas, L.J.: Render for cnn: Viewpoint estimation in images using cnns trained with rendered 3d model views. In: Proceedings of the IEEE international conference on computer vision. pp. 2686-2694 (2015)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 32, + 200, + 379, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 200, + 379, + 243 + ], + "spans": [ + { + "bbox": [ + 32, + 200, + 379, + 243 + ], + "type": "text", + "content": "70. Su, Y., Saleh, M., Fetzer, T., Rambach, J., Navab, N., Busam, B., Stricker, D., Tombari, F.: Zebrapose: Coarse to fine surface encoding for 6dof object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6738-6748 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 32, + 244, + 379, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 244, + 379, + 287 + ], + "spans": [ + { + "bbox": [ + 32, + 244, + 379, + 287 + ], + "type": "text", + "content": "71. Sundermeyer, M., Durner, M., Huang, E.Y., Marton, Z.C., Vaskevicius, N., Arras, K.O., Triebel, R.: Multi-path learning for object pose estimation across domains. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13916-13925 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 32, + 288, + 379, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 288, + 379, + 331 + ], + "spans": [ + { + "bbox": [ + 32, + 288, + 379, + 331 + ], + "type": "text", + "content": "72. Sundermeyer, M., Hodač, T., Labbe, Y., Wang, G., Brachmann, E., Drost, B., Rother, C., Matas, J.: Bop challenge 2022 on detection, segmentation and pose estimation of specific rigid objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2784-2793 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 32, + 331, + 379, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 331, + 379, + 364 + ], + "spans": [ + { + "bbox": [ + 32, + 331, + 379, + 364 + ], + "type": "text", + "content": "73. Sundermeyer, M., Marton, Z.C., Durner, M., Brucker, M., Triebel, R.: Implicit 3d orientation learning for 6d object detection from rgb images. In: Proceedings of the european conference on computer vision (ECCV). pp. 699-715 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 32, + 364, + 379, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 364, + 379, + 386 + ], + "spans": [ + { + "bbox": [ + 32, + 364, + 379, + 386 + ], + "type": "text", + "content": "74. Szafraniec, F.H.: The reproducing kernel hilbert space and its multiplication operators. Complex Analysis and Related Topics pp. 253-263 (2000)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 32, + 386, + 379, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 386, + 379, + 419 + ], + "spans": [ + { + "bbox": [ + 32, + 386, + 379, + 419 + ], + "type": "text", + "content": "75. Tan, T., Dong, Q.: Smoc-net: Leveraging camera pose for self-supervised monocular object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21307-21316 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 32, + 419, + 379, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 419, + 379, + 441 + ], + "spans": [ + { + "bbox": [ + 32, + 419, + 379, + 441 + ], + "type": "text", + "content": "76. Ullman, S.: The interpretation of structure from motion. Proceedings of the Royal Society of London. Series B. Biological Sciences 203(1153), 405-426 (1979)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 32, + 441, + 379, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 441, + 379, + 473 + ], + "spans": [ + { + "bbox": [ + 32, + 441, + 379, + 473 + ], + "type": "text", + "content": "77. Venkateswara, H., Eusebio, J., Chakraborty, S., Panchanathan, S.: Deep hashing network for unsupervised domain adaptation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5018-5027 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 32, + 474, + 379, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 474, + 379, + 495 + ], + "spans": [ + { + "bbox": [ + 32, + 474, + 379, + 495 + ], + "type": "text", + "content": "78. Vert, J.P., Tsuda, K., Scholkopf, B.: A primer on kernel methods. Kernel methods in computational biology 47, 35-70 (2004)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 32, + 495, + 379, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 495, + 379, + 528 + ], + "spans": [ + { + "bbox": [ + 32, + 495, + 379, + 528 + ], + "type": "text", + "content": "79. Wang, G., Manhardt, F., Liu, X., Ji, X., Tombari, F.: Occlusion-aware self-supervised monocular 6d object pose estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 32, + 528, + 379, + 562 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 528, + 379, + 562 + ], + "spans": [ + { + "bbox": [ + 32, + 528, + 379, + 562 + ], + "type": "text", + "content": "80. Wang, G., Manhardt, F., Shao, J., Ji, X., Navab, N., Tombari, F.: Self6d: Self-supervised monocular 6d object pose estimation. In: European Conference on Computer Vision. pp. 108-125. Springer (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 32, + 562, + 379, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 562, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 32, + 562, + 379, + 583 + ], + "type": "text", + "content": "81. Wu, Y., Greenspan, M.: Learning better keypoints for multi-object 6dof pose estimation. arXiv preprint arXiv:2308.07827 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 300, + 9, + 347, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 9, + 347, + 19 + ], + "spans": [ + { + "bbox": [ + 300, + 9, + 347, + 19 + ], + "type": "text", + "content": "RKHPose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 35, + 379, + 353 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 32, + 35, + 379, + 68 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 35, + 379, + 68 + ], + "spans": [ + { + "bbox": [ + 32, + 35, + 379, + 68 + ], + "type": "text", + "content": "82. Wu, Y., Javaheri, A., Zand, M., Greenspan, M.: Keypoint cascade voting for point cloud based 6dof pose estimation. In: 2022 International Conference on 3D Vision (3DV). pp. 176-186. IEEE (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 32, + 68, + 379, + 100 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 68, + 379, + 100 + ], + "spans": [ + { + "bbox": [ + 32, + 68, + 379, + 100 + ], + "type": "text", + "content": "83. Wu, Y., Zand, M., Etemad, A., Greenspan, M.: Vote from the center: 6 dof pose estimation in rgb-d images by radial keypoint voting. In: European Conference on Computer Vision. pp. 335-352. Springer (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 101, + 379, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 101, + 379, + 123 + ], + "spans": [ + { + "bbox": [ + 33, + 101, + 379, + 123 + ], + "type": "text", + "content": "84. Xiang, Y., Schmidt, T., Narayanan, V., Fox, D.: PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 123, + 379, + 166 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 123, + 379, + 166 + ], + "spans": [ + { + "bbox": [ + 33, + 123, + 379, + 166 + ], + "type": "text", + "content": "85. Xiao, F., Liu, H., Lee, Y.J.: Identity from here, pose from there: Self-supervised disentanglement and generation of objects using unlabeled videos. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7013-7022 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 167, + 379, + 189 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 167, + 379, + 189 + ], + "spans": [ + { + "bbox": [ + 33, + 167, + 379, + 189 + ], + "type": "text", + "content": "86. Xiao, Y., Qiu, X., Langlois, P.A., Aubry, M., Marlet, R.: Pose from shape: Deep pose estimation for arbitrary 3d objects. arXiv preprint arXiv:1906.05105 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 189, + 379, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 189, + 379, + 232 + ], + "spans": [ + { + "bbox": [ + 33, + 189, + 379, + 232 + ], + "type": "text", + "content": "87. Yang, H., Pavone, M.: Object pose estimation with statistical guarantees: Conformal keypoint detection and geometric uncertainty propagation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8947-8958 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 33, + 233, + 379, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 233, + 379, + 265 + ], + "spans": [ + { + "bbox": [ + 33, + 233, + 379, + 265 + ], + "type": "text", + "content": "88. Zhang, Y., David, P., Gong, B.: Curriculum domain adaptation for semantic segmentation of urban scenes. In: Proceedings of the IEEE international conference on computer vision. pp. 2020-2030 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 33, + 266, + 379, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 266, + 379, + 309 + ], + "spans": [ + { + "bbox": [ + 33, + 266, + 379, + 309 + ], + "type": "text", + "content": "89. Zhang, Z., Wang, M., Huang, Y., Nehorai, A.: Aligning infinite-dimensional covariance matrices in reproducing kernel hilbert spaces for domain adaptation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3437-3445 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 33, + 309, + 379, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 309, + 379, + 353 + ], + "spans": [ + { + "bbox": [ + 33, + 309, + 379, + 353 + ], + "type": "text", + "content": "90. Zhou, J., Chen, K., Xu, L., Dou, Q., Qin, J.: Deep fusion transformer network with weighted vector-wise keypoints voting for robust 6d object pose estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13967-13977 (2023)" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 32, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 32, + 10, + 43, + 19 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 168, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 168, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 168, + 20 + ], + "type": "text", + "content": "Y.Wu and M. Greenspan" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 19 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Put Myself in Your Shoes_ Lifting the Egocentric Perspective from Exocentric Videos/cf2ae3fa-fba8-4d91-a4a1-6c60e6a8095b_content_list.json b/2024/Put Myself in Your Shoes_ Lifting the Egocentric Perspective from Exocentric Videos/cf2ae3fa-fba8-4d91-a4a1-6c60e6a8095b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4a2bb795008cffd010e1f0e4f76ef07c4076c3a8 --- /dev/null +++ b/2024/Put Myself in Your Shoes_ Lifting the Egocentric Perspective from Exocentric Videos/cf2ae3fa-fba8-4d91-a4a1-6c60e6a8095b_content_list.json @@ -0,0 +1,1559 @@ +[ + { + "type": "text", + "text": "Put Myself in Your Shoes: Lifting the Egocentric Perspective from Exocentric Videos", + "text_level": 1, + "bbox": [ + 220, + 140, + 782, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mi Luo $^{1}$ , Zihui Xue $^{1,2}$ , Alex Dimakis $^{1}$ , and Kristen Grauman $^{1,2}$", + "bbox": [ + 269, + 210, + 732, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 The University of Texas at Austin \n2 FAIR, Meta AI", + "bbox": [ + 377, + 238, + 624, + 267 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a6dcc9c18039fafa2af0d5b6d13f26b60c4b5268fa152bec010bcc160d59e948.jpg", + "image_caption": [ + "Fig. 1: We consider cross-view translation from exocentric to egocentric video. Given an exocentric video sequence with hand-object interactions (left), the goal is to generate the corresponding egocentric perspective (right)." + ], + "image_footnote": [], + "bbox": [ + 230, + 289, + 782, + 377 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We investigate exocentric-to-egocentric cross-view translation, which aims to generate a first-person (egocentric) view of an actor based on a video recording that captures the actor from a third-person (exocentric) perspective. To this end, we propose a generative framework called Exo2Ego that decouples the translation process into two stages: high-level structure transformation, which explicitly encourages cross-view correspondence between exocentric and egocentric views, and a diffusion-based pixel-level hallucination, which incorporates a hand layout prior to enhance the fidelity of the generated egocentric view. To pave the way for future advancements in this field, we curate a comprehensive exo-to-ego cross-view translation benchmark focused on hand-object manipulations. It consists of a diverse collection of synchronized ego-exo video pairs from four public datasets: H2O, Aria Pilot, Assembly101, and Ego-Exo4D. The experimental results validate that Exo2Ego delivers photorealistic video results with clear hand manipulation details and outperforms several baselines in terms of both synthesis quality and generalization to new actions.", + "bbox": [ + 259, + 462, + 743, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Egocentric Hand-Object Interaction $\\cdot$ View Synthesis", + "bbox": [ + 261, + 710, + 705, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 751, + 375, + 767 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Given a third-person video capturing a person opening a milk carton, what would the visual world look like from his perspective? See Figure 1. Due to mirror neurons in the human brain [3], we can easily envision the appearance", + "bbox": [ + 212, + 782, + 787, + 829 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and spatial relationships of the person's hands and the milk carton from a first-person perspective. However, existing computer vision models struggle to do the same thing, owing to the stark distinctions between the two viewpoints.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We make a step towards addressing this underexplored problem: exocentric-to-egocentric cross-view translation. The goal is to synthesize the corresponding ego view of an actor from an exo video recording3, with minimal assumptions on the viewpoint relationships (e.g., camera parameters or accurate geometric scene structure). Specifically, we focus on synthesizing ego hand-object interactions in table-top activities, such as assembling toys, pouring milk, or picking up a box.", + "bbox": [ + 212, + 191, + 785, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This task can benefit many applications, ranging from robotics to virtual and augmented reality (VR/AR). For example, an AR assistant could transform the view shown in a third-person how-to video to demonstrate to a user how things should look from their own perspectives, e.g., revealing finger placements and strumming techniques for a guitar video. Similarly, in robot learning, the robot could better match its actions to those of a human instructor in the room by projecting his/her hand-object interactions to its own perspective. Indeed, recent work demonstrates that the human ego view is valuable for robot learning [4,6, 33,34,36], as robots also perceive the world through the egocentric lens.", + "bbox": [ + 212, + 282, + 785, + 417 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, exo-to-ego view translation is highly challenging. It requires understanding the spatial relationships of visible hands and objects and inferring their pixel-level appearance in the novel ego view. Further, the task is not strictly geometric; it is inherently underdetermined. Some parts of an object may not be visible in the exo view—such as the inner pages of a book when only the cover is observed in the exo view—requiring the model to extrapolate the occluded parts. As a result, the recent popular geometry-aware novel view synthesis approaches [23, 35, 37, 70] are ill-equipped to solve this problem. The key reason is that they are regressive rather than generative, which limits their ability to deal with sparse input views (single exo view in our case) and major occlusions.", + "bbox": [ + 212, + 417, + 785, + 568 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In light of this, we propose a probabilistic exo-to-ego generative framework (termed Exo2Ego) that learns the conditional distribution of the target ego video given the exo video. Exo2Ego differs from generative-based methods for general novel view synthesis [9, 29, 43, 45, 58, 64, 66] as it relaxes the requirement for camera parameters as input. This expands the potential application scenarios since camera parameters are rarely accessible in real-world scenarios, e.g., taking photos or videos with a mobile phone. Additionally, accurately inferring camera parameters remains challenging without precise correspondences or a rigid lab setting for calibration [63]. Moreover, Exo2Ego stands apart from recent cross-view image-to-image translation methods [42, 57], as it does not require the ground truth semantic map (object segmentation) of the target domain at the inference stage—typically an infeasible assumption.", + "bbox": [ + 212, + 569, + 785, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our key insights are to explicitly encourage cross-view correspondence by predicting the ego layout and to introduce a hand layout prior to boost the fidelity of generated hands in the ego view. Specifically, our Exo2Ego framework", + "bbox": [ + 212, + 750, + 785, + 796 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "M. Luo et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "3We use \"ego\" and \"exo\" as shorthand for egocentric (first-person) and exocentric (third-person).", + "bbox": [ + 214, + 810, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "decouples the exo-to-ego view translation problem into two stages: (1) High-level Structure Transformation, which infers the rough location and interaction manner of the hands and objects in the ego view, using a transformer-based model to transform the exo hand-object interaction layout into its ego counterpart; and (2) Diffusion-based Pixel Hallucination, which trains a conditional diffusion model to refine the details on top of the ego hand-object interaction layout. Overall, Exo2Ego is a generative framework that offers a simple but effective baseline approach for the exo-to-ego view translation problem, accounting for the centrality of hand-object manipulations in this domain.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To evaluate our approach and boost research in this field, we construct an exoto-ego cross-view synthesis benchmark, comprising synchronized ego-exo video pairs curated from four datasets: H2O [27], Aria Pilot [32], Assembly101 [47], and Ego-Exo4D [16]. These datasets offer a breadth of table-top activities, as well as some non-table-top activities where the camera wearer moves around the scene. Empirical results underscore the efficacy of our Exo2Ego framework, which produces realistic video outputs with distinct hand manipulation details. Exo2Ego surpasses single-view translation baselines based on GANs [61,62] and latent diffusion models [38], a recent cross-view synthesis approach [30], as well as a NeRF-based method [70], demonstrating superior generation quality and a marked improvement in generalization capability to new actions.", + "bbox": [ + 212, + 282, + 787, + 449 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 214, + 467, + 382, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Relating Egocentric and Exocentric Views", + "text_level": 1, + "bbox": [ + 214, + 494, + 609, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There have been several attempts to jointly understand ego and exo views [13, 15, 51]. Early efforts explore how to localize the person wearing a camera in a third-person view, given their egocentric video [1,2,14,65,67]. To bridge the ego-exo gap, other work learns view-invariant features from concurrent (paired) [3, 48,50,71,72] or unpaired [68] views, or boosts the latent ego signals in exo video during pretraining [28], summarization [17], and 3D pose estimation [60]. Additionally, fusing ego and exo views can improve action recognition [56] and robotic manipulation tasks [24]. Despite its significance, cross-view translation has received little attention. P-GAN [30] and STA-GAN [31] have explored cross-view image and video synthesis, respectively, yet they only examine basic activities such as walking, jogging, and running, limiting their applicability to more diverse scenarios. To address these limitations, we introduce a new benchmark for exoto-ego view synthesis covering some diverse activities, from assembling toys to manipulating everyday objects. Importantly, we relax STA-GAN's requirement for the ego semantic map during test time [31], enhancing the practical applicability. A concurrent work [10] examines the introduction of geometry priors to the diffusion model.", + "bbox": [ + 212, + 513, + 787, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Novel View Synthesis", + "text_level": 1, + "bbox": [ + 214, + 787, + 441, + 804 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Exo-to-ego view translation also relates to novel view synthesis, i.e., given a set of images of a scene, infer how the scene looks from novel viewpoints. State-of-the", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Lifting the Egocentric Perspective from Exocentric Videos", + "bbox": [ + 341, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "art geometry-aware approaches like Scene Representation Networks (SRN) [53] and Neural Radiance Fields (NeRF) [23,35,37,70] learn 3D representations from 2D images and camera parameters and perform differentiable neural rendering from one or a few input images at inference. Although they excel at interpolating near-input views, they are typically not well-suited for the large view changes required for exo-to-ego view synthesis. Trained purely with regression objectives, they have limited capacity to deal with sparse inputs, ambiguity caused by occluded parts of the scene, and long-range extrapolations caused by drastic camera transformations [9,64]—all challenges that are particularly prominent in exo-to-ego view translation. Alternatively, generative-based novel view synthesis methods [9,29,43,45,58,64,66] use generative priors to synthesize random plausible outputs from the conditional distribution, making them adept at handling ambiguity and extrapolating to occluded parts of the scene. Our Exo2Ego framework belongs to the generative-based vein, but unlike all the above, it does not depend on camera poses as critical conditional inputs. This improves scalability and widens the scope of potential applications, as camera parameters are seldom accessible in real-world scenarios and are challenging to infer from RGB images.", + "bbox": [ + 212, + 146, + 787, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Cross-view Image-to-Image Translation", + "text_level": 1, + "bbox": [ + 215, + 425, + 584, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Exo-to-ego view synthesis is inherently an image-to-image translation problem [5, 10, 22, 61, 62]—but with drastic viewpoint change. Prior work centers around aerial-to-ground translation [40–42, 57] or generating an egocentric view without human-object interactions, i.e. a person walking indoors [30, 31, 57]. Except for P-GAN [30], all of these approaches require the ground truth semantic map of the target domain at training or inference time to ensure high-quality generation—an unrealistic assumption that means a good portion of the translation task (the boundaries of the objects in the unobserved ego view or the input exo view) is already known, Compared with the semantic-map-free approach P-GAN, our Exo2Ego exhibits significantly better qualitative and quantitative results.", + "bbox": [ + 212, + 450, + 787, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.4 Diffusion Models", + "text_level": 1, + "bbox": [ + 215, + 625, + 401, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "DMs have shown competitive performance in image synthesis [11, 18, 19, 55], but their use for cross-view translation remains unexplored. We develop an exo-to-ego translation framework incorporating the diffusion model rather than GANs, considering potential advantages such as addressing mode collapse, promoting sample diversity, and maintaining stable training dynamics for improved image generation [44].", + "bbox": [ + 212, + 648, + 787, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 762, + 380, + 780 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, we formalize our problem (Sec. 3.1). Then we present the two stages of our model (Sec. 3.2 and 3.3), followed by an overview of training and inference (Sec. 3.4).", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "M. Luo et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/57b400d9f41d668deba24255b68beceb7a6c8c8388d993c8662050fffc0ce089.jpg", + "image_caption": [ + "(a) High-level Structure Transformation" + ], + "image_footnote": [], + "bbox": [ + 236, + 160, + 754, + 271 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/810ff5de80a80597b291fd03348103d831ab2fc360e06e47a952ed5109a671e8.jpg", + "image_caption": [ + "(b) Diffusion-based Pixel Hallucination", + "Fig. 2: Our Exo2Ego framework comprises two modules: (a) High-level Structure Transformation, which predicts the ego layout, capturing hand position and interactions using a transformer-based encoder-decoder architecture. (b) Diffusion-based Pixel Hallucination, which enhances pixel-level details on top of the ego layout using a conditional diffusion model." + ], + "image_footnote": [], + "bbox": [ + 243, + 290, + 763, + 416 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Problem Formulation", + "text_level": 1, + "bbox": [ + 214, + 508, + 437, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Whereas conventional geometry-aware approaches [10,35,70] generate novel views by specifying camera poses and performing volume rendering, Exo2Ego offers a different perspective. We propose tackling the exo-to-ego translation problem in a purely probabilistic framework. Intuitively, this is essential to address the inherent ambiguity in predicting the ego view from the exo view, e.g., due to entirely unseen portions of objects or the human body.", + "bbox": [ + 212, + 532, + 787, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We define $\\mathbf{X}_T = \\{x_1, \\ldots, x_i, \\ldots, x_T\\}$ as a series of $T$ video frames captured from an exo viewpoint, characterized by static backgrounds with dynamic actors and other objects present, where $i$ denotes the time index. This exo perspective reveals the actors' motions and (potentially) full-body poses within the scene. Let $\\mathbf{Y}_T = \\{y_1, \\ldots, y_i, \\ldots, y_T\\}$ denote the corresponding sequence of ego frames captured from a first-person perspective. This perspective emulates the view of a camera mounted on the head or body of the actor, with a focus on their actions and interactions.", + "bbox": [ + 212, + 625, + 787, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The goal of exo-to-ego view translation is to simulate the view of the ego camera-wearer in the scene recorded by an exo camera. Formally, we seek a translation model that can map $\\mathbf{X}_T$ into a series of output ego frames, $\\hat{\\mathbf{Y}}_T = \\{\\hat{y}_1,\\dots ,\\hat{y}_i,\\dots ,\\hat{y}_T\\}$ . As shown in Eqn. (1), the conditional distribution of $\\hat{\\mathbf{Y}}^T$ given $\\mathbf{X}^T$ should be indistinguishable from the conditional distribution of $\\mathbf{Y}^T$", + "bbox": [ + 214, + 744, + 787, + 821 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Lifting the Egocentric Perspective from Exocentric Videos", + "bbox": [ + 343, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "given $\\mathbf{X}^T$", + "bbox": [ + 215, + 145, + 290, + 161 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\mathbf {\\hat {Y}} _ {T} \\mid \\mathbf {X} _ {T}\\right) = p \\left(\\mathbf {Y} _ {T} \\mid \\mathbf {X} _ {T}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 161, + 784, + 178 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We prioritize translating daily tabletop activities from the exo to ego view, which is a prevalent setting in egocentric learning and frequently entails extensive interactions between hands and objects. This problem is highly challenging as the translation model must produce photorealistic hand-object interaction sequences in the ego view while also performing geometric and semantic reasoning, i.e., correctly predicting the spatial location and pixel-level appearance of visual concepts. Furthermore, it requires linking the actor's head/body motion captured by the exo camera with the viewpoint change in the ego video. Upon examining the application of the widely used pixel-to-pixel generation method [62] for our task, we found it struggles with intricate details of the hands, likely because treats all pixels equally and lacks geometric correspondence between views. At the same time, we know that (at least in manipulation-rich scenarios) the hands of the actor are a common ground between the ego and exo views, despite their many other differences.", + "bbox": [ + 212, + 184, + 787, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Motivated by these points, we propose the Exo2Ego framework, which disentangles the understanding of cross-view correspondences and pixel-level synthesis. It consists of two key modules, as illustrated in Figure 2: (1) High-level Structure Transformation, which addresses the task of inferring the location and interaction manner of hands and objects in the ego view. To accomplish this, we train a transformer-based encoder-decoder model to translate the exo frame into an ego hand-object interaction layout (detailed below). (2) Diffusion-based Pixel Hallucination, which learns to synthesize realistic and high-quality pixel-level details by training a conditional diffusion model operating on top of the ego hand layout.", + "bbox": [ + 212, + 395, + 787, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 High-level Structure Transformation", + "text_level": 1, + "bbox": [ + 215, + 569, + 563, + 584 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given an exo frame, the purpose of the high-level structure transformation is to train a layout translator which predicts the ego layout showing the location and rough contour of the hands. Specifically, to capture fine-grained hand-object interaction details, we propose to generate the hand layout, instantiated as 2D hand poses. The generated layout serves as a crucial reference for further pixel-level hallucination. To achieve this, we draw inspiration from a recent study [45] that highlights transformer-based architectures [12, 59] succeed in understanding cross-view correspondence, due to their reduced locality-bias. Considering transformers have the potential to represent the geometric information implicitly, we employ a pure transformer-based encoder-decoder architecture, which enables us to effectively incorporate and process the comprehensive context of the exo scene, and thus generate a precise ego hand layout.", + "bbox": [ + 212, + 594, + 787, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Exo Contextual Feature Extraction As shown in Figure 2, an exo frame $x_{i} \\in \\mathbb{R}^{H \\times W \\times 3}$ concatenated with the corresponding exo layout $x_{i}^{l} \\in \\mathbb{R}^{H \\times W \\times 3}$ (2D hand pose layout recorded as image size) is first split into a sequence of patches", + "bbox": [ + 214, + 790, + 787, + 835 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "M. Luo et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and then processed by an input embedding layer, such as patch embedding for ViTs [12] to get a tokenized input sequence $e_i \\in \\mathbb{R}^{M \\times D}$ . To capture the positional information which is critical for spatial relationship reasoning, we add a learnable position embedding to the sequence of patches. Then the sequence of tokens is passed into a transformer encoder consisting of $N$ repeated blocks. Each block contains a multi-head attention token mixer to first communicate information among tokens, which is represented as:", + "bbox": [ + 214, + 146, + 782, + 252 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nz _ {i} ^ {j - 1} = \\operatorname {A t t e n t i o n} \\left(\\operatorname {N o r m} \\left(e _ {i} ^ {j - 1}\\right)\\right) + e _ {i} ^ {j - 1}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 262, + 785, + 280 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $j \\in \\{1, \\dots, N\\}$ and $\\mathrm{Norm}(\\cdot)$ is the normalization method. Then, the mixed token $z_i^{j-1}$ is passed into a two-layer MLP expressed as follows:", + "bbox": [ + 214, + 290, + 782, + 320 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\ne _ {i} ^ {j} = \\operatorname {M L P} \\left(\\operatorname {N o r m} \\left(z _ {i} ^ {j - 1}\\right)\\right) + z _ {i} ^ {j - 1}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 332, + 784, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After iterating through all $N$ blocks, the input sequence $e_i$ undergoes a mapping process that transforms it into the final contextual tokens $c_{i}\\in \\mathbb{R}^{M\\times D}$ . We talk about how to decode the contextual tokens into the desired ego layout in the following part.", + "bbox": [ + 214, + 359, + 782, + 420 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Layout Decoder Consider a hand pose image with shape $\\{H\\times W\\times 3\\}$ as the target ego layout. The contextual tokens $c_{i}$ should be decoded into 2D hand joint coordinates and then rendered to the RGB image. First, the contextual tokens $c_{i}$ are combined with a sequence of learnable query embeddings $q_{i}\\in \\mathbb{R}^{E\\times D}$ and passed into a decoder architecture, consisting of $K$ transformer blocks, as described earlier. Following the transformer blocks, a regression head is employed to estimate the 2D coordinates of the joints whose range is restricted to [0,1]. Then we apply the classic bipartite matching loss [8, 39] to map the regressed hand joints to the ground truth joints.", + "bbox": [ + 214, + 434, + 784, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Diffusion-based Pixel Hallucination", + "text_level": 1, + "bbox": [ + 215, + 592, + 552, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Having inferred the ego hand pose layout $y_{i}^{l}$ , next the aim of pixel-level hallucination is to synthesize photorealistic ego frames by taking into account an exo frame $x_{i}$ and the target ego layout. We use the diffusion formulation proposed in the denoising diffusion probabilistic model (DDPM) [18] and train the diffusion model in the latent space. As shown in Figure 2, we first adopt a pre-trained variational autoencoder (VAE) model [25] as used in [44] to encode the original ego frames $y_{i}$ and conditional information exo frame $x_{i}$ and ego layout $y_{i}^{l}$ into the latent space. Then, we train a diffusion transformer [38] to learn the latent data distribution by denoising a latent vector sampled from a Gaussian distribution gradually. Specifically, given an initial noise map $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ and a conditioning vector $\\mathbf{d}$ , the diffusion model generates the corresponding ego latent $\\hat{\\mathbf{z}}_{\\theta}$ . A squared error loss is used to denoise a variably-noised latent code $\\mathbf{z}^{n} = \\alpha^{n}\\mathbf{z} + \\sigma^{n}\\epsilon$ as follows:", + "bbox": [ + 214, + 616, + 782, + 811 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} _ {\\mathbf {d}, \\mathbf {z}, \\boldsymbol {\\epsilon}, n} \\left[ \\| \\mathbf {z} - \\hat {\\mathbf {z}} _ {\\theta} \\left(\\alpha^ {n} \\mathbf {z} + \\sigma^ {n} \\boldsymbol {\\epsilon}, \\mathbf {d}\\right) \\| _ {2} ^ {2} \\right] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 824, + 784, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Lifting the Egocentric Perspective from Exocentric Videos", + "bbox": [ + 341, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{z}$ is the ground-truth latent vector, $\\mathbf{d}$ is the conditioning vector, and $\\alpha^n,\\sigma^n$ are functions of the diffusion process time step $n$ , which control the noise schedule and sample quality. We simply concatenate the noisy latent vector $\\mathbf{z}^n$ and the conditional embedding $\\mathbf{d}$ as the input of the denoising transformer. In contrast to earlier cross-view translation methods based on GANs [30, 61, 62], our conditional diffusion model sequentially updates the target ego outputs, excels in capturing intricate ego-exo dependencies and faithfully reproducing the mapping of the conditional distributions in Eqn. (1). Notably, our model exhibits enhanced stability throughout the training and consistently produces higher-quality samples, as verified by our experiments.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Training and Inference", + "text_level": 1, + "bbox": [ + 214, + 319, + 452, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "At the training stage, the two modules described in Section 3.2 and 3.3 are trained separately. During inference, we first generate the corresponding ego hand layout prediction $\\hat{y}_i^l$ for each input exo frame $x_i$ with the layout translator described in Section 3.2. Subsequently, $\\hat{y}_i^l$ is concatenated with $x_i$ in a channel-wise manner and then used as the conditional input of the pixel-level diffusion generator described in Section 3.3.", + "bbox": [ + 212, + 343, + 787, + 434 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experimental Evaluation", + "text_level": 1, + "bbox": [ + 214, + 455, + 493, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce an exo-to-ego synthesis benchmark (Sec. 4.1) and then validate our approach against state-of-the-art synthesis methods (Sec. 4.2).", + "bbox": [ + 212, + 488, + 785, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 A Benchmark for Exo-to-Ego View Translation", + "text_level": 1, + "bbox": [ + 214, + 540, + 650, + 555 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To facilitate new work on the exo-to-ego translation task, we contribute a new cross-view synthesis benchmark sourced from four public time-synchronized multi-view datasets: H2O [27], Aria Pilot [32], Assembly101 [47], and Ego-Exo4D [16]. H2O [27] has indoor videos featuring actors manipulating 8 different objects using both hands. Assembly101 [47] shows people (dis)assembling 101 take-apart toy vehicles. We take 6 activity sequences involving different individuals manipulating a toy roller. Aria Pilot [32] provides 16 multi-view recordings of an actor wearing Project Aria glasses [54] manipulating YCB [7] objects on a desktop. For Ego-Exo4D, we include the four (largely non-table-top) scenarios involving heavy hand-object interactions from the EgoHandPose Benchmark - bike repair, cooking, music, and health—where the camera wearer also moves about the scene. For all datasets, we select the exo camera viewpoints that offer the clearest capture of the action. All frames are cropped and resized into $256 \\times 256$ . We divide each video into 30-frame clips. Overall, this dataset showcases an array of human activities with various objects, from toy assembly to the manipulation of everyday objects. More details are provided in Appendix.", + "bbox": [ + 212, + 564, + 790, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We employ the following split settings to evaluate four kinds of generalization: (1) new actions, where we put the first $80\\%$ of a video's clips into the training", + "bbox": [ + 214, + 810, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "M. Luo et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3d6c4652f9367f296636c1487338979ce1c07a6f60c9d025c3966767eb60bc7f.jpg", + "image_caption": [ + "Fig. 3: Qualitative examples when generalizing to new actions on all datasets." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 802, + 537 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "set and the remaining $20\\%$ in the test set, hence splitting up action steps over time (e.g., for a video depicting picking up a coffee box and then taking coffee capsules out, the former will be assigned to the training set and the latter to the test set); (2) new objects, where we train with videos involving any of 7 objects and test with videos containing a novel 8th object; (3) new subjects, where we train with all clips from one subject and test on clips from another (subject $1 \\rightarrow$ subject 2); (4) new scenes (backgrounds), where we train and test with disjoint scenes (scene h1, h2, k1, k2 $\\rightarrow$ scene o1, o2). We analyze new action generalization on all 3 datasets, and the rest on H2O only, as it is the only dataset with the variations and meta-data to allow such split settings. For action generalization, H2O, Aria Pilot, Assembly101, and Ego-Exo4D have 704, 343, 682, and 543 training clips, and 199, 95, 205, and 211 test clips, respectively. For other generalization settings (new objects, subjects, and scenes), there are 691, 704, and 591 training clips, and 108, 194, and 312 test clips.", + "bbox": [ + 212, + 573, + 787, + 785 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Evaluating Synthesis Quality", + "text_level": 1, + "bbox": [ + 215, + 803, + 501, + 819 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "With this well-constructed benchmark in hand, we proceed to validate Exo2Ego.", + "bbox": [ + 215, + 824, + 784, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Lifting the Egocentric Perspective from Exocentric Videos", + "bbox": [ + 343, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/78205a34281d856ce75a95de6806ed950b33aef27d4e08a632019db8634fed1b.jpg", + "table_caption": [ + "Table 1: Generalizing to new actions on three datasets. *pixelNeRF requires privileged camera information relative to the other baseline models." + ], + "table_footnote": [], + "table_body": "
H2O
MethodSSIM↑PSNR↑FID↓P_Squeeze↓P_Alex↓P_Vgg↓Feasi.↑
pix2pixHD [62]0.42830.370132.030.1630.2200.3370.8952
P-GAN [30]0.27229.079192.260.2420.3080.4490.8353
vid2vid [61]0.40229.88285.760.1660.2260.3420.9116
pixelNeRF* [70]0.21927.871470.130.7110.8070.7310.0086
DiT [38]0.43630.55345.230.1290.1780.2970.9604
Exo2Ego (Ours)0.44330.58037.030.1040.1460.2620.9764
Aria Pilot
MethodSSIM↑PSNR↑FID↓P_Squeeze↓P_Alex↓P_Vgg↓Feasi.↑
pix2pixHD [62]0.35029.553159.910.3040.3620.4800.0790
P-GAN [30]0.34329.769122.400.2870.3450.4750.1014
vid2vid [61]0.35929.85852.380.2660.3280.4430.2646
DiT [38]0.36829.94532.710.2400.2980.4170.4784
Exo2Ego (Ours)0.38030.04725.990.2250.2790.4020.5977
Assembly101
MethodSSIM↑PSNR↑FID↓P_Squeeze↓P_Alex↓P_Vgg↓Feasi.↑
pix2pixHD [62]0.40529.909112.710.2160.2950.4010.1184
P-GAN [30]0.37229.303114.680.2160.3000.4360.0800
vid2vid [61]0.36829.48178.470.2240.3120.4240.1358
DiT [38]0.39829.99626.330.1790.2550.3700.3092
Exo2Ego (Ours)0.41330.15325.230.1680.2380.3550.4019
", + "bbox": [ + 261, + 184, + 736, + 507 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baselines We consider several baselines: (1) Pix2PixHD [62], a single-view image translation method that processes the videos frame by frame; (2) P-GAN [30], a recent exo-to-ego view translation method that proposes a parallel generative network to facilitate cross-view image translation; (3) DiT [38], a pixel-to-pixel generative model from the latent diffusion model family. (4) Vid2Vid [61], a single-view video translation method that models the temporal dynamics in the video; (5) pixelNeRF [70], a NeRF-like model known for its remarkable generalization capabilities and its accommodating stance toward sparser input views. Note that pixelNeRF assumes camera parameters are known, while other methods do not. We conduct experiments with pixelNeRF only on H2O, as the other two datasets lack the camera information needed by pixelNeRF.", + "bbox": [ + 212, + 512, + 787, + 679 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation Details For Exo2Ego's high-level structure transformation transformer, the number of blocks $N$ is set to 6. We use DiT-XL/2 [38] as the denoising architecture for Exo2Ego. The diffusion model is trained for 40,000 steps and pixelNeRF is trained for 70,000 steps for all datasets. For H2O and Assembly101, pix2pixHD, and P-GAN are trained for 100 epochs, and vid2vid is trained for 40 epochs. For Aria Pilot, we use 400 and 100 epochs respectively.", + "bbox": [ + 212, + 681, + 787, + 772 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Metrics We measure the quality of the synthesized videos as follows. First, we evaluate the feasibility of generated hands, denoted as Feasi. Inspired by [69], we adopt an off-the-shelf egocentric hand object detector [49] to measure the physical feasibility of the synthesized hands. This metric is calcu", + "bbox": [ + 212, + 773, + 787, + 835 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "M. Luo et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/30ef4e009d01ea1429b8a18ca20c08788b2cb3baa1dc401b63e979d95ef1daef.jpg", + "image_caption": [ + "Fig. 4: Qualitative examples when generalizing to new objects, subjects, and scenes (backgrounds) on the H2O dataset." + ], + "image_footnote": [], + "bbox": [ + 217, + 146, + 803, + 315 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "lated as the confidence score of hand detection results produced by the detector, averaged over all synthesized frames.", + "bbox": [ + 215, + 362, + 782, + 391 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We also use structural similarity (SSIM) and peak signal-to-noise ratio (PSNR) [20] to check the pixel-level similarity between the predicted ego frame and the ground truth (GT) ego frame. These two evaluation metrics should ideally remain stable under minor transformations (such as small translations or affine transformations). We further adopt perceptual metrics, including FID [46] and Learned Perceptual Image Patch Similarity (LPIPS) [74] which evaluates the feature-level similarity of the predicted ego frame and the true ego frame. LPIPS are denoted as $P_{\\text{Squeeze}}$ , $P_{\\text{Alex}}$ , and $P_{\\text{Vgg}}$ , which uses SqueezeNet [21], AlexNet [26], and VGG [52] as the feature extractor.", + "bbox": [ + 212, + 392, + 807, + 529 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Generalization to New Actions Figure 3 showcases qualitative results on H2O, Aria Pilot, and Assembly101, demonstrating the superiority of our Exo2Ego framework. Compared to all baselines, our approach produces realistic hands with correct poses, especially noticeable in the highlighted yellow circle regions when dealing with new actions during test time. Table 1 presents the feasibility of synthesized hands and the various measures of pixel-level similarity and perceptual similarity between the predicted ego frame and the ground truth ego frame. Our Exo2Ego greatly increases the feasibility of synthesized hands, with a confidence gain of up to $11.93\\%$ on Aria Pilot, on the basis of the second-highest score of $47.84\\%$ achieved by DiT. This improvement indicates enhanced realism in the generated hands, resulting in greater confidence in hand detection results. Additionally, our Exo2Ego can produce ego frames with higher pixel-level similarity to the ground truth frames and better perceptual metrics (much lower FID scores). This highlights the benefits of the 'decoupling' idea of our Exo2Ego framework in generating photorealistic ego frames.", + "bbox": [ + 212, + 551, + 787, + 779 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Looking at Table 1, there is a trend among our Exo2Ego and the baselines — doing better on H2O dataset. We conjecture that the superiority comes from H2O's higher hand quality, including clear pose visibility and larger spatial coverage within the field of view, in both exo and ego view.", + "bbox": [ + 212, + 779, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Lifting the Egocentric Perspective from Exocentric Videos", + "bbox": [ + 343, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/b4aaee35d2c9653dfd23f18ed6bd038216c9166c0ab54274255ef83c0b9770fc.jpg", + "table_caption": [ + "Table 2: Generalizing to different objects, subjects, and scenes. *pixelNeRF requires privileged camera information relative to the other baseline models." + ], + "table_footnote": [], + "table_body": "
H2O-unseen object
MethodSSIM↑PSNR↑FID↓P_Squeeze↓P_Alex↓P_Vgg↓Feasi.↑
pix2pixHD [62]0.33329.504253.730.2690.3600.4670.7527
P-GAN [30]0.27028.915264.760.2770.3500.4920.7720
vid2vid [61]0.34929.585145.390.2420.3240.4280.8612
pixelNeRF* [70]0.26227.912473.730.7500.8410.7360.0000
DiT [38]0.33829.740105.180.2140.2880.4100.9689
Exo2Ego (Ours)0.34429.786101.450.2080.2810.4030.9786
H2O-unseen subject
MethodSSIM↑PSNR↑FID↓P_Squeeze↓P_Alex↓P_Vgg↓Feasi.↑
pix2pixHD [62]0.06928.028477.470.6610.7540.7650.0140
P-GAN [30]0.08828.063451.370.6230.7080.7250.2002
vid2vid [61]0.13628.094308.520.5640.6900.6860.3838
pixelNeRF* [70]0.17027.874459.490.7690.8360.7680.0000
DiT [38]0.17228.185183.060.5240.6300.6820.8405
Exo2Ego (Ours)0.17828.266150.870.5030.6290.6650.9666
H2O-unseen scene
MethodSSIM↑PSNR↑FID↓P_Squeeze↓P_Alex↓P_Vgg↓Feasi.↑
pix2pixHD [62]0.02427.918558.870.7290.7890.7880.0013
P-GAN [30]0.08527.928325.340.6080.6990.7240.6041
vid2vid [61]0.03327.942398.780.6360.7310.7340.2095
pixelNeRF* [70]0.10327.810506.100.7730.9080.8000.0000
DiT [38]0.14027.874308.980.5840.7290.6780.8317
Exo2Ego (Ours)0.15727.943249.040.5310.7090.6580.9635
", + "bbox": [ + 259, + 184, + 736, + 536 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We further explored how Exo2Ego fares in the settings with heavy hand-object manipulation and frequent head motion, but minimal movement across the broader scene on Ego-Exo4D [16]. As shown in Figure 6, we compare our Exo2Ego with a strong image-to-image translation model ControlNet [73], which is built on a stable diffusion trained on large scales of data. It is observed that our Exo2Ego synthesizes feasible hands with correct poses, unlike ControlNet. The exo image's hand regions in Ego-Exo4D are small, so we crop it to use as additional input to offer a better view of the hands for both methods.", + "bbox": [ + 212, + 542, + 787, + 664 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Generalization to New Objects, Subjects, and Backgrounds Illustrated in Table 2, our Exo2Ego outperforms the baselines in terms of quantitative metrics for generalization to new objects/subjects/backgrounds. Figure 4 shows that our Exo2Ego can produce realistic hand-object interactions when encountering exo views with new distributions, while other baselines fail miserably. This is mainly because our diffusion-based generative mechanism gradually denoises the target ego view, capturing the inherent intricate mapping from exo to ego views better. Another observation is that pixelNeRF produces ego views characterized by noticeable blurring and a lack of fine details in hands and objects. This observation supports our earlier discussions on the limitations of 3D-aware approaches.", + "bbox": [ + 212, + 678, + 787, + 830 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "M. Luo et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/42f834b8cb09ccf73596a17383d0340687adc4c67584b1d5cc01ad0c4ebb423d.jpg", + "image_caption": [ + "Fig. 5: Exo2Ego framework generates ego videos with reasonable viewpoint changes." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 779, + 306 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/35de2f6f4c3b2e12a6f8f8d6a551b86955c7675b22d9e12c9a7e1dba93cb56ea.jpg", + "image_caption": [ + "Fig. 6: Qualitative examples when generalizing to new actions on Ego-Exo4D dataset." + ], + "image_footnote": [], + "bbox": [ + 220, + 332, + 782, + 452 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Modeling Egocentric Viewpoint Change Due to the probabilistic nature of our Exo2Ego framework, it excels in capturing the relationship between the viewpoint changes in the ego view and the head motions observed in the exo view, holding the potential to deliver more immersive virtual experiences. Intrinsically, our Exo2Ego framework leverages the fact that ego-camera pose is conditioned on the reference exo view, as the ego-camera wearer is usually visible in the exo view. This is highlighted by the yellow gaze direction and dashed box in Figure 5. As the actor moves their head when interacting with objects, the slope of the desk edge in the ego view also changes, indicating a corresponding change in ego viewpoint. This implies the ego-camera motion can be implicitly inferred via conditional generative modeling, even if it's not enforced explicitly. The ability to model ego viewpoint changes is crucial, especially in physical activities like playing ball games which involve heavy head and body motions.", + "bbox": [ + 212, + 488, + 787, + 686 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Limitations and Future Work", + "text_level": 1, + "bbox": [ + 215, + 708, + 535, + 724 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Temporal Coherence Our Exo2Ego operates on each frame individually — the generation of the current ego frame is solely conditioned on the current exo input frame. There is no explicit consideration of previously generated ego frames or past exo frames. Temporal coherence is orthogonal to our current contribution and can be used to improve any of the baseline models without such design.", + "bbox": [ + 212, + 739, + 787, + 816 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Lifting the Egocentric Perspective from Exocentric Videos", + "bbox": [ + 343, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2bee228fb7920b58dd07171f1d09bba823ab610d4d56ef0c6e7d90279e64b4be.jpg", + "image_caption": [ + "Fig. 7: Failure cases: Exo2Ego generates hands with reasonably accurate positions/poses but incomplete/wrong objects in the ego view." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 776, + 200 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Generalization Ability As an early-stage effort for exo-to-ego cross-view translation, our work makes non-trivial progress in producing reasonable hand manipulation details for novel actions. However, despite establishing stronger generalizations than the existing models, our model does not generalize perfectly to in-the-wild objects, subjects, and backgrounds, due to the limited scale of available training data. Considering the intricate nature of the generalization to new environments, we regard it as a challenging future research direction. Nonetheless, our focus on generalizing to new actions within the same environment is meaningful for several applications. Some examples: Real-time monitoring during physical rehabilitation (exo-to-ego view translation can help therapists guide patients during rehabilitation sessions), assistive cooking aid for visually impaired individuals (ego views are provided to the visually impaired individuals to cook independently), or real-time fitness training monitoring (coaches use synthesized ego views to help the trainee correct the fitness movements) — in each case, the environment is stable between training and testing, and yet the model must generalize to new actions for cross-view translation.", + "bbox": [ + 212, + 242, + 787, + 484 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "3D prior Our qualitative results expose limitations in existing baselines and our Exo2Ego framework in generating 3D-consistent views for objects during test time, attributed to the absence of geometric priors for common objects. See Figure 7. Future work involves integrating robust object geometric priors. While leveraging predicted camera parameters for encoding 3D geometry is a promising approach, as shown Figure 4, the popular 3D-aware method pixelNeRF [70] faces challenges in the dynamic exo-ego setting characterized by hand and object interactions, due to limitations in handling sparse inputs, occluded scenes, and long-range camera extrapolations. Notably, our Exo2Ego outperforms pixelNeRF, highlighting the potential for future work to explore explicit geometric reasoning within our framework.", + "bbox": [ + 212, + 498, + 787, + 664 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 686, + 359, + 704 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Overall, our work contributes to the growing body of research in cross-view translation and lays the groundwork for the important case of exo-to-ego. Our generative framework Exo2Ego integrates high-level structure transformation and pixel-level hallucination to yield very encouraging experimental results. We also provide a curated benchmark task for supporting continued work on this problem. The core technology holds immense potential for applications in robot learning and AR skill coaching, where an ego actor needs to replicate the actions of a demonstrator observed from the exo perspective.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "M. Luo et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgment: UT Austin is supported in part by NSF Grants AF 1901292, CNS 2148141, Tripods CCF 1934932, IFML CCF 2019844 and research gifts by Western Digital, Amazon, WNCG IAP, UT Austin Machine Learning Lab (MLL), Cisco, and the Stanly P. Finch Centennial Professorship in Engineering. K.G. is paid as a research scientist at Meta.", + "bbox": [ + 215, + 146, + 787, + 223 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 241, + 321, + 257 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ardeshir, S., Borji, A.: Ego2top: Matching viewers in egocentric and top-view videos. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14. pp. 253-268. Springer (2016) 3", + "2. Ardeshir, S., Borji, A.: Egocentric meets top-view. IEEE transactions on pattern analysis and machine intelligence 41(6), 1353-1366 (2018) 3", + "3. Ardeshir, S., Borji, A.: An exocentric look at egocentric actions and vice versa. Computer Vision and Image Understanding 171, 61-68 (2018) 1, 3", + "4. Bahl, S., Gupta, A., Pathak, D.: Human-to-robot imitation in the wild. arXiv preprint arXiv:2207.09450 (2022) 2", + "5. Bansal, A., Ma, S., Ramanan, D., Sheikh, Y.: Recycle-gan: Unsupervised video retargeting. In: Proceedings of the European conference on computer vision (ECCV). pp. 119-135 (2018) 4", + "6. Bharadhwaj, H., Gupta, A., Tulsiani, S., Kumar, V.: Zero-shot robot manipulation from passive human videos. arXiv preprint arXiv:2302.02011 (2023) 2", + "7. Calli, B., Singh, A., Walsman, A., Srinivasa, S., Abbeel, P., Dollar, A.M.: The ycb object and model set: Towards common benchmarks for manipulation research. In: 2015 international conference on advanced robotics (ICAR). pp. 510-517. IEEE (2015) 8", + "8. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 213-229. Springer (2020) 7", + "9. Chan, E.R., Nagano, K., Chan, M.A., Bergman, A.W., Park, J.J., Levy, A., Aittala, M., De Mello, S., Karras, T., Wetzstein, G.: Generative novel view synthesis with 3d-aware diffusion models. arXiv preprint arXiv:2304.02602 (2023) 2, 4", + "0. Cheng, F., Luo, M., Wang, H., Dimakis, A., Torresani, L., Bertasius, G., Grauman, K.: 4diff: 3d-aware diffusion model for third-to-first viewpoint translation. In: ECCV (2024) 3, 4, 5", + "1. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems 34, 8780-8794 (2021) 4", + "12. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 6, 7", + "13. Elfeki, M., Regmi, K., Ardeshir, S., Borji, A.: From third person to first person: Dataset and baselines for synthesis and retrieval. arXiv preprint arXiv:1812.00104 (2018) 3", + "14. Fan, C., Lee, J., Xu, M., Kumar Singh, K., Jae Lee, Y., Crandall, D.J., Ryoo, M.S.: Identifying first-person camera wearers in third-person videos. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5125-5133 (2017) 3" + ], + "bbox": [ + 218, + 268, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Lifting the Egocentric Perspective from Exocentric Videos", + "bbox": [ + 341, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Grauman, K., Westbury, A., Torresani, L., Kitani, K., Malik, J., Afouras, T., Ashutosh, K., Baiyya, V., Bansal, S., Boote, B., et al.: Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. arXiv preprint arXiv:2311.18259 (2023) 3", + "16. Grauman, K., Westbury, A., Torresani, L., Kitani, K., Malik, J., Afouras, T., Ashutosh, K., Baiyya, V., Bansal, S., Boote, B., et al.: Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19383-19400 (2024) 3, 8, 12", + "17. Ho, H.I., Chiu, W.C., Wang, Y.C.F.: Summarizing first-person videos from third persons' points of view. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 70-85 (2018) 3", + "18. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33, 6840-6851 (2020) 4, 7", + "19. Ho, J., Sahara, C., Chan, W., Fleet, D.J., Norouzi, M., Salimans, T.: Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research 23(1), 2249-2281 (2022) 4", + "20. Hore, A., Ziou, D.: Image quality metrics: Psnr vs. sSIM. In: 2010 20th international conference on pattern recognition. pp. 2366-2369. IEEE (2010) 11", + "21. Iandola, F.N., Han, S., Moskewicz, M.W., Ashraf, K., Dally, W.J., Keutzer, K.: SqueezeNET: Alexnet-level accuracy with 50x fewer parameters and $< 0.5$ mb model size. arXiv preprint arXiv:1602.07360 (2016) 11", + "22. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1125-1134 (2017) 4", + "23. Jang, W., Agapito, L.: Codenerf: Disentangled neural radiance fields for object categories. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12949-12958 (2021) 2, 4", + "24. Jangir, R., Hansen, N., Ghosal, S., Jain, M., Wang, X.: Look closer: Bridging egocentric and third-person views with transformers for robotic manipulation. IEEE Robotics and Automation Letters 7(2), 3046-3053 (2022) 3", + "25. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013) 7", + "26. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. Communications of the ACM 60(6), 84-90 (2017) 11", + "27. Kwon, T., Tekin, B., Stühmer, J., Bogo, F., Pollefeys, M.: H2o: Two hands manipulating objects for first person interaction recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10138-10148 (October 2021) 3, 8", + "28. Li, Y., Nagarajan, T., Xiong, B., Grauman, K.: Ego-exo: Transferring visual representations from third-person to first-person videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6943-6953 (2021) 3", + "29. Liu, A., Tucker, R., Jampani, V., Makadia, A., Snavely, N., Kanazawa, A.: Infinite nature: Perpetual view generation of natural scenes from a single image. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14458-14467 (2021) 2, 4", + "30. Liu, G., Tang, H., Latapie, H., Yan, Y.: Exocentric to egocentric image generation via parallel generative adversarial network. In: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1843-1847. IEEE (2020) 3, 4, 8, 10, 12" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "M. Luo et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Liu, G., Tang, H., Latapie, H.M., Corso, J.J., Yan, Y.: Cross-view exocentric to egocentric video synthesis. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 974-982 (2021) 3, 4", + "32. Lv, Z., Miller, E., Meissner, J., Pesqueira, L., Sweeney, C., Dong, J., Ma, L., Patel, P., Moulon, P., Somasundaram, K., Parkhi, O., Zou, Y., Raina, N., Saarinen, S., Mansour, Y.M., Huang, P.K., Wang, Z., Troynikov, A., Artal, R.M., DeTone, D., Barnes, D., Argall, E., Lobanovskiy, A., Kim, D.J., Bouttefroy, P., Straub, J., Engel, J.J., Gupta, P., Yan, M., Nardi, R.D., Newcombe, R.: Aria pilot dataset. https://about.facebook.com/realitylabs/projectaria/datasets (2022) 3, 8", + "33. Majumdar, A., Yadav, K., Arnaud, S., Ma, Y.J., Chen, C., Silwal, S., Jain, A., Berges, V.P., Abbeel, P., Malik, J., et al.: Where are we in the search for an artificial visual cortex for embodied intelligence? arXiv preprint arXiv:2303.18240 (2023) 2", + "34. Mandikal, P., Grauman, K.: Dexvip: Learning dexterous grasping with human hand pose priors from video. In: Conference on Robot Learning (2021) 2", + "35. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) 2, 4, 5", + "36. Nair, S., Rajeswaran, A., Kumar, V., Finn, C., Gupta, A.: R3m: A universal visual representation for robot manipulation. arXiv preprint arXiv:2203.12601 (2022) 2", + "37. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5480-5490 (2022) 2, 4", + "38. Peebles, W., Xie, S.: Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748 (2022) 3, 7, 10, 12", + "39. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified, real-time object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 779-788 (2016) 7", + "40. Regmi, K., Borji, A.: Cross-view image synthesis using conditional gans. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018) 4", + "41. Regmi, K., Borji, A.: Cross-view image synthesis using geometry-guided conditional gans. Computer Vision and Image Understanding (2019). https://doi.org/https://doi.org/10.1016/j.cviu.2019.07.008, http://www.sciencedirect.com/science/article/pii/S1077314219301043 4", + "42. Ren, B., Tang, H., Sebe, N.: Cascaded cross mlp-mixer gans for cross-view image translation. arXiv preprint arXiv:2110.10183 (2021) 2, 4", + "43. Ren, X., Wang, X.: Look outside the room: Synthesizing a consistent long-term 3d scene video from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3563-3573 (2022) 2, 4", + "44. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022) 4, 7", + "45. Rombach, R., Esser, P., Ommer, B.: Geometry-free view synthesis: Transformers and no 3d priors. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14356-14366 (2021) 2, 4, 6", + "46. Seitzer, M.: pytorch-fid: FID Score for PyTorch. https://github.com/mseitzer/pytorch-fid (August 2020), version 0.3.0 11" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Lifting the Egocentric Perspective from Exocentric Videos", + "bbox": [ + 343, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "47. Sener, F., Chatterjee, D., Shelepov, D., He, K., Singhania, D., Wang, R., Yao, A.: Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21096-21106 (2022) 3, 8", + "48. Sermanet, P., Lynch, C., Chebotar, Y., Hsu, J., Jang, E., Schaal, S., Levine, S.: Time-contrastive networks: Self-supervised learning from video. Proceedings of International Conference in Robotics and Automation (ICRA) (2018), http://arxiv.org/abs/1704.06888 3", + "49. Shan, D., Geng, J., Shu, M., Fouhey, D.F.: Understanding human hands in contact at internet scale. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020) 10", + "50. Sigurdsson, G.A., Gupta, A., Schmid, C., Farhadi, A., Alahari, K.: Actor and observer: Joint modeling of first and third-person videos. In: proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7396-7404 (2018) 3", + "51. Sigurdsson, G.A., Gupta, A., Schmid, C., Farhadi, A., Alahari, K.: Charades-ego: A large-scale dataset of paired third and first person videos. arXiv preprint arXiv:1804.09626 (2018) 3", + "52. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014) 11", + "53. Sitzmann, V., Zollhöfer, M., Wetzstein, G.: Scene representation networks: Continuous 3d-structure-aware neural scene representations. Advances in Neural Information Processing Systems 32 (2019) 4", + "54. Somasundaram, K., Dong, J., Tang, H., Straub, J., Yan, M., Goesele, M., Engel, J.J., De Nardi, R., Newcombe, R.: Project aria: A new tool for egocentric multimodal ai research. arXiv preprint arXiv:2308.13561 (2023) 8", + "55. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=PxTIG12RRHS 4", + "56. Soran, B., Farhadi, A., Shapiro, L.: Action recognition in the presence of one egocentric and multiple static cameras. In: Computer Vision-ACCV 2014: 12th Asian Conference on Computer Vision, Singapore, Singapore, November 1-5, 2014, Revised Selected Papers, Part V 12. pp. 178-193. Springer (2015) 3", + "57. Tang, H., Xu, D., Sebe, N., Wang, Y., Corso, J.J., Yan, Y.: Multi-channel attention selection gan with cascaded semantic guidance for cross-view image translation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2417-2426 (2019) 2, 4", + "58. Tseng, H.Y., Li, Q., Kim, C., Alsisan, S., Huang, J.B., Kopf, J.: Consistent view synthesis with pose-guided diffusion models. arXiv preprint arXiv:2303.17598 (2023) 2, 4", + "59. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) 6", + "60. Wang, J., Liu, L., Xu, W., Sarkar, K., Luvizon, D., Theobalt, C.: Estimating egocentric 3d human pose in the wild with external weak supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13157-13166 (2022) 3", + "61. Wang, T.C., Liu, M.Y., Zhu, J.Y., Liu, G., Tao, A., Kautz, J., Catanzaro, B.: Video-to-video synthesis. arXiv preprint arXiv:1808.06601 (2018) 3, 4, 8, 10, 12" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "M. Luo et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "62. Wang, T.C., Liu, M.Y., Zhu, J.Y., Tao, A., Kautz, J., Catanzaro, B.: High-resolution image synthesis and semantic manipulation with conditional gans. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8798-8807 (2018) 3, 4, 6, 8, 10, 12", + "63. Wang, Z., Wu, S., Xie, W., Chen, M., Prisacariu, V.A.: NeRF--: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064 (2021) 2", + "64. Watson, D., Chan, W., Martin-Brualla, R., Ho, J., Tagliasacchi, A., Norouzi, M.: Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628 (2022) 2, 4", + "65. Wen, Y., Singh, K.K., Anderson, M., Jan, W.P., Lee, Y.J.: Seeing the unseen: Predicting the first-person camera wearer's location and pose in third-person scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. pp. 3446-3455 (October 2021) 3", + "66. Wiles, O., Gkioxari, G., Szeliski, R., Johnson, J.: Synsin: End-to-end view synthesis from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7467-7477 (2020) 2, 4", + "67. Xu, M., Fan, C., Wang, Y., Ryoo, M.S., Crandall, D.J.: Joint person segmentation and identification in synchronized first-and third-person videos. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 637-652 (2018) 3", + "68. Xue, Z., Grauman, K.: Learning fine-grained view-invariant representations from unpaired ego-exo videos via temporal alignment. In: NeurIPS (2023) 3", + "69. Ye, Y., Li, X., Gupta, A., Mello, S.D., Birchfield, S., Song, J., Tulsiani, S., Liu, S.: Affordance diffusion: Synthesizing hand-object interactions. In: CVPR (2023) 10", + "70. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4578-4587 (June 2021) 2, 3, 4, 5, 10, 12, 14", + "71. Yu, H., Cai, M., Liu, Y., Lu, F.: What i see is what you see: Joint attention learning for first and third person video co-analysis. In: Proceedings of the 27th ACM International Conference on Multimedia. pp. 1358-1366 (2019) 3", + "72. Yu, H., Cai, M., Liu, Y., Lu, F.: First-and third-person video co-analysis by learning spatial-temporal joint attention. IEEE Transactions on Pattern Analysis and Machine Intelligence (2020) 3", + "73. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models (2023) 12", + "74. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018) 11" + ], + "bbox": [ + 212, + 146, + 787, + 672 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Lifting the Egocentric Perspective from Exocentric Videos", + "bbox": [ + 343, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Put Myself in Your Shoes_ Lifting the Egocentric Perspective from Exocentric Videos/cf2ae3fa-fba8-4d91-a4a1-6c60e6a8095b_model.json b/2024/Put Myself in Your Shoes_ Lifting the Egocentric Perspective from Exocentric Videos/cf2ae3fa-fba8-4d91-a4a1-6c60e6a8095b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f49e99c862e5045d06eacb9f5fa72ab9e30cebf2 --- /dev/null +++ b/2024/Put Myself in Your Shoes_ Lifting the Egocentric Perspective from Exocentric Videos/cf2ae3fa-fba8-4d91-a4a1-6c60e6a8095b_model.json @@ -0,0 +1,2383 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.222, + 0.141, + 0.783, + 0.187 + ], + "angle": 0, + "content": "Put Myself in Your Shoes: Lifting the Egocentric Perspective from Exocentric Videos" + }, + { + "type": "text", + "bbox": [ + 0.27, + 0.212, + 0.733, + 0.228 + ], + "angle": 0, + "content": "Mi Luo\\(^{1}\\), Zihui Xue\\(^{1,2}\\), Alex Dimakis\\(^{1}\\), and Kristen Grauman\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.379, + 0.239, + 0.625, + 0.268 + ], + "angle": 0, + "content": "1 The University of Texas at Austin \n2 FAIR, Meta AI" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.29, + 0.784, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.388, + 0.789, + 0.431 + ], + "angle": 0, + "content": "Fig. 1: We consider cross-view translation from exocentric to egocentric video. Given an exocentric video sequence with hand-object interactions (left), the goal is to generate the corresponding egocentric perspective (right)." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.463, + 0.744, + 0.699 + ], + "angle": 0, + "content": "Abstract. We investigate exocentric-to-egocentric cross-view translation, which aims to generate a first-person (egocentric) view of an actor based on a video recording that captures the actor from a third-person (exocentric) perspective. To this end, we propose a generative framework called Exo2Ego that decouples the translation process into two stages: high-level structure transformation, which explicitly encourages cross-view correspondence between exocentric and egocentric views, and a diffusion-based pixel-level hallucination, which incorporates a hand layout prior to enhance the fidelity of the generated egocentric view. To pave the way for future advancements in this field, we curate a comprehensive exo-to-ego cross-view translation benchmark focused on hand-object manipulations. It consists of a diverse collection of synchronized ego-exo video pairs from four public datasets: H2O, Aria Pilot, Assembly101, and Ego-Exo4D. The experimental results validate that Exo2Ego delivers photorealistic video results with clear hand manipulation details and outperforms several baselines in terms of both synthesis quality and generalization to new actions." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.712, + 0.707, + 0.726 + ], + "angle": 0, + "content": "Keywords: Egocentric Hand-Object Interaction \\(\\cdot\\) View Synthesis" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.752, + 0.377, + 0.768 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.784, + 0.789, + 0.83 + ], + "angle": 0, + "content": "Given a third-person video capturing a person opening a milk carton, what would the visual world look like from his perspective? See Figure 1. Due to mirror neurons in the human brain [3], we can easily envision the appearance" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "M. Luo et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "and spatial relationships of the person's hands and the milk carton from a first-person perspective. However, existing computer vision models struggle to do the same thing, owing to the stark distinctions between the two viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.787, + 0.284 + ], + "angle": 0, + "content": "We make a step towards addressing this underexplored problem: exocentric-to-egocentric cross-view translation. The goal is to synthesize the corresponding ego view of an actor from an exo video recording3, with minimal assumptions on the viewpoint relationships (e.g., camera parameters or accurate geometric scene structure). Specifically, we focus on synthesizing ego hand-object interactions in table-top activities, such as assembling toys, pouring milk, or picking up a box." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.787, + 0.419 + ], + "angle": 0, + "content": "This task can benefit many applications, ranging from robotics to virtual and augmented reality (VR/AR). For example, an AR assistant could transform the view shown in a third-person how-to video to demonstrate to a user how things should look from their own perspectives, e.g., revealing finger placements and strumming techniques for a guitar video. Similarly, in robot learning, the robot could better match its actions to those of a human instructor in the room by projecting his/her hand-object interactions to its own perspective. Indeed, recent work demonstrates that the human ego view is valuable for robot learning [4,6, 33,34,36], as robots also perceive the world through the egocentric lens." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.787, + 0.569 + ], + "angle": 0, + "content": "However, exo-to-ego view translation is highly challenging. It requires understanding the spatial relationships of visible hands and objects and inferring their pixel-level appearance in the novel ego view. Further, the task is not strictly geometric; it is inherently underdetermined. Some parts of an object may not be visible in the exo view—such as the inner pages of a book when only the cover is observed in the exo view—requiring the model to extrapolate the occluded parts. As a result, the recent popular geometry-aware novel view synthesis approaches [23, 35, 37, 70] are ill-equipped to solve this problem. The key reason is that they are regressive rather than generative, which limits their ability to deal with sparse input views (single exo view in our case) and major occlusions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.787, + 0.751 + ], + "angle": 0, + "content": "In light of this, we propose a probabilistic exo-to-ego generative framework (termed Exo2Ego) that learns the conditional distribution of the target ego video given the exo video. Exo2Ego differs from generative-based methods for general novel view synthesis [9, 29, 43, 45, 58, 64, 66] as it relaxes the requirement for camera parameters as input. This expands the potential application scenarios since camera parameters are rarely accessible in real-world scenarios, e.g., taking photos or videos with a mobile phone. Additionally, accurately inferring camera parameters remains challenging without precise correspondences or a rigid lab setting for calibration [63]. Moreover, Exo2Ego stands apart from recent cross-view image-to-image translation methods [42, 57], as it does not require the ground truth semantic map (object segmentation) of the target domain at the inference stage—typically an infeasible assumption." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.751, + 0.787, + 0.797 + ], + "angle": 0, + "content": "Our key insights are to explicitly encourage cross-view correspondence by predicting the ego layout and to introduce a hand layout prior to boost the fidelity of generated hands in the ego view. Specifically, our Exo2Ego framework" + }, + { + "type": "page_footnote", + "bbox": [ + 0.215, + 0.811, + 0.787, + 0.841 + ], + "angle": 0, + "content": "3We use \"ego\" and \"exo\" as shorthand for egocentric (first-person) and exocentric (third-person)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.343, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Lifting the Egocentric Perspective from Exocentric Videos" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "decouples the exo-to-ego view translation problem into two stages: (1) High-level Structure Transformation, which infers the rough location and interaction manner of the hands and objects in the ego view, using a transformer-based model to transform the exo hand-object interaction layout into its ego counterpart; and (2) Diffusion-based Pixel Hallucination, which trains a conditional diffusion model to refine the details on top of the ego hand-object interaction layout. Overall, Exo2Ego is a generative framework that offers a simple but effective baseline approach for the exo-to-ego view translation problem, accounting for the centrality of hand-object manipulations in this domain." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.45 + ], + "angle": 0, + "content": "To evaluate our approach and boost research in this field, we construct an exoto-ego cross-view synthesis benchmark, comprising synchronized ego-exo video pairs curated from four datasets: H2O [27], Aria Pilot [32], Assembly101 [47], and Ego-Exo4D [16]. These datasets offer a breadth of table-top activities, as well as some non-table-top activities where the camera wearer moves around the scene. Empirical results underscore the efficacy of our Exo2Ego framework, which produces realistic video outputs with distinct hand manipulation details. Exo2Ego surpasses single-view translation baselines based on GANs [61,62] and latent diffusion models [38], a recent cross-view synthesis approach [30], as well as a NeRF-based method [70], demonstrating superior generation quality and a marked improvement in generalization capability to new actions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.468, + 0.383, + 0.484 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.495, + 0.611, + 0.51 + ], + "angle": 0, + "content": "2.1 Relating Egocentric and Exocentric Views" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.788, + 0.773 + ], + "angle": 0, + "content": "There have been several attempts to jointly understand ego and exo views [13, 15, 51]. Early efforts explore how to localize the person wearing a camera in a third-person view, given their egocentric video [1,2,14,65,67]. To bridge the ego-exo gap, other work learns view-invariant features from concurrent (paired) [3, 48,50,71,72] or unpaired [68] views, or boosts the latent ego signals in exo video during pretraining [28], summarization [17], and 3D pose estimation [60]. Additionally, fusing ego and exo views can improve action recognition [56] and robotic manipulation tasks [24]. Despite its significance, cross-view translation has received little attention. P-GAN [30] and STA-GAN [31] have explored cross-view image and video synthesis, respectively, yet they only examine basic activities such as walking, jogging, and running, limiting their applicability to more diverse scenarios. To address these limitations, we introduce a new benchmark for exoto-ego view synthesis covering some diverse activities, from assembling toys to manipulating everyday objects. Importantly, we relax STA-GAN's requirement for the ego semantic map during test time [31], enhancing the practical applicability. A concurrent work [10] examines the introduction of geometry priors to the diffusion model." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.789, + 0.442, + 0.805 + ], + "angle": 0, + "content": "2.2 Novel View Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Exo-to-ego view translation also relates to novel view synthesis, i.e., given a set of images of a scene, infer how the scene looks from novel viewpoints. State-of-the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.361, + 0.127 + ], + "angle": 0, + "content": "M. Luo et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.404 + ], + "angle": 0, + "content": "art geometry-aware approaches like Scene Representation Networks (SRN) [53] and Neural Radiance Fields (NeRF) [23,35,37,70] learn 3D representations from 2D images and camera parameters and perform differentiable neural rendering from one or a few input images at inference. Although they excel at interpolating near-input views, they are typically not well-suited for the large view changes required for exo-to-ego view synthesis. Trained purely with regression objectives, they have limited capacity to deal with sparse inputs, ambiguity caused by occluded parts of the scene, and long-range extrapolations caused by drastic camera transformations [9,64]—all challenges that are particularly prominent in exo-to-ego view translation. Alternatively, generative-based novel view synthesis methods [9,29,43,45,58,64,66] use generative priors to synthesize random plausible outputs from the conditional distribution, making them adept at handling ambiguity and extrapolating to occluded parts of the scene. Our Exo2Ego framework belongs to the generative-based vein, but unlike all the above, it does not depend on camera poses as critical conditional inputs. This improves scalability and widens the scope of potential applications, as camera parameters are seldom accessible in real-world scenarios and are challenging to infer from RGB images." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.426, + 0.586, + 0.442 + ], + "angle": 0, + "content": "2.3 Cross-view Image-to-Image Translation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.451, + 0.788, + 0.603 + ], + "angle": 0, + "content": "Exo-to-ego view synthesis is inherently an image-to-image translation problem [5, 10, 22, 61, 62]—but with drastic viewpoint change. Prior work centers around aerial-to-ground translation [40–42, 57] or generating an egocentric view without human-object interactions, i.e. a person walking indoors [30, 31, 57]. Except for P-GAN [30], all of these approaches require the ground truth semantic map of the target domain at training or inference time to ensure high-quality generation—an unrealistic assumption that means a good portion of the translation task (the boundaries of the objects in the unobserved ego view or the input exo view) is already known, Compared with the semantic-map-free approach P-GAN, our Exo2Ego exhibits significantly better qualitative and quantitative results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.625, + 0.403, + 0.639 + ], + "angle": 0, + "content": "2.4 Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.65, + 0.788, + 0.742 + ], + "angle": 0, + "content": "DMs have shown competitive performance in image synthesis [11, 18, 19, 55], but their use for cross-view translation remains unexplored. We develop an exo-to-ego translation framework incorporating the diffusion model rather than GANs, considering potential advantages such as addressing mode collapse, promoting sample diversity, and maintaining stable training dynamics for improved image generation [44]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.763, + 0.381, + 0.781 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.841 + ], + "angle": 0, + "content": "First, we formalize our problem (Sec. 3.1). Then we present the two stages of our model (Sec. 3.2 and 3.3), followed by an overview of training and inference (Sec. 3.4)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.344, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Lifting the Egocentric Perspective from Exocentric Videos" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image_caption", + "bbox": [ + 0.237, + 0.148, + 0.41, + 0.159 + ], + "angle": 0, + "content": "(a) High-level Structure Transformation" + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.161, + 0.756, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.243, + 0.279, + 0.413, + 0.29 + ], + "angle": 0, + "content": "(b) Diffusion-based Pixel Hallucination" + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.291, + 0.764, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.431, + 0.788, + 0.502 + ], + "angle": 0, + "content": "Fig. 2: Our Exo2Ego framework comprises two modules: (a) High-level Structure Transformation, which predicts the ego layout, capturing hand position and interactions using a transformer-based encoder-decoder architecture. (b) Diffusion-based Pixel Hallucination, which enhances pixel-level details on top of the ego layout using a conditional diffusion model." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.509, + 0.439, + 0.523 + ], + "angle": 0, + "content": "3.1 Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.788, + 0.625 + ], + "angle": 0, + "content": "Whereas conventional geometry-aware approaches [10,35,70] generate novel views by specifying camera poses and performing volume rendering, Exo2Ego offers a different perspective. We propose tackling the exo-to-ego translation problem in a purely probabilistic framework. Intuitively, this is essential to address the inherent ambiguity in predicting the ego view from the exo view, e.g., due to entirely unseen portions of objects or the human body." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.625, + 0.788, + 0.744 + ], + "angle": 0, + "content": "We define \\(\\mathbf{X}_T = \\{x_1, \\ldots, x_i, \\ldots, x_T\\}\\) as a series of \\(T\\) video frames captured from an exo viewpoint, characterized by static backgrounds with dynamic actors and other objects present, where \\(i\\) denotes the time index. This exo perspective reveals the actors' motions and (potentially) full-body poses within the scene. Let \\(\\mathbf{Y}_T = \\{y_1, \\ldots, y_i, \\ldots, y_T\\}\\) denote the corresponding sequence of ego frames captured from a first-person perspective. This perspective emulates the view of a camera mounted on the head or body of the actor, with a focus on their actions and interactions." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.746, + 0.788, + 0.823 + ], + "angle": 0, + "content": "The goal of exo-to-ego view translation is to simulate the view of the ego camera-wearer in the scene recorded by an exo camera. Formally, we seek a translation model that can map \\(\\mathbf{X}_T\\) into a series of output ego frames, \\(\\hat{\\mathbf{Y}}_T = \\{\\hat{y}_1,\\dots ,\\hat{y}_i,\\dots ,\\hat{y}_T\\}\\). As shown in Eqn. (1), the conditional distribution of \\(\\hat{\\mathbf{Y}}^T\\) given \\(\\mathbf{X}^T\\) should be indistinguishable from the conditional distribution of \\(\\mathbf{Y}^T\\)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "M. Luo et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.146, + 0.292, + 0.162 + ], + "angle": 0, + "content": "given \\(\\mathbf{X}^T\\)" + }, + { + "type": "equation", + "bbox": [ + 0.414, + 0.162, + 0.785, + 0.179 + ], + "angle": 0, + "content": "\\[\np \\left(\\mathbf {\\hat {Y}} _ {T} \\mid \\mathbf {X} _ {T}\\right) = p \\left(\\mathbf {Y} _ {T} \\mid \\mathbf {X} _ {T}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.185, + 0.788, + 0.395 + ], + "angle": 0, + "content": "We prioritize translating daily tabletop activities from the exo to ego view, which is a prevalent setting in egocentric learning and frequently entails extensive interactions between hands and objects. This problem is highly challenging as the translation model must produce photorealistic hand-object interaction sequences in the ego view while also performing geometric and semantic reasoning, i.e., correctly predicting the spatial location and pixel-level appearance of visual concepts. Furthermore, it requires linking the actor's head/body motion captured by the exo camera with the viewpoint change in the ego video. Upon examining the application of the widely used pixel-to-pixel generation method [62] for our task, we found it struggles with intricate details of the hands, likely because treats all pixels equally and lacks geometric correspondence between views. At the same time, we know that (at least in manipulation-rich scenarios) the hands of the actor are a common ground between the ego and exo views, despite their many other differences." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.396, + 0.788, + 0.547 + ], + "angle": 0, + "content": "Motivated by these points, we propose the Exo2Ego framework, which disentangles the understanding of cross-view correspondences and pixel-level synthesis. It consists of two key modules, as illustrated in Figure 2: (1) High-level Structure Transformation, which addresses the task of inferring the location and interaction manner of hands and objects in the ego view. To accomplish this, we train a transformer-based encoder-decoder model to translate the exo frame into an ego hand-object interaction layout (detailed below). (2) Diffusion-based Pixel Hallucination, which learns to synthesize realistic and high-quality pixel-level details by training a conditional diffusion model operating on top of the ego hand layout." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.57, + 0.564, + 0.585 + ], + "angle": 0, + "content": "3.2 High-level Structure Transformation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.788, + 0.777 + ], + "angle": 0, + "content": "Given an exo frame, the purpose of the high-level structure transformation is to train a layout translator which predicts the ego layout showing the location and rough contour of the hands. Specifically, to capture fine-grained hand-object interaction details, we propose to generate the hand layout, instantiated as 2D hand poses. The generated layout serves as a crucial reference for further pixel-level hallucination. To achieve this, we draw inspiration from a recent study [45] that highlights transformer-based architectures [12, 59] succeed in understanding cross-view correspondence, due to their reduced locality-bias. Considering transformers have the potential to represent the geometric information implicitly, we employ a pure transformer-based encoder-decoder architecture, which enables us to effectively incorporate and process the comprehensive context of the exo scene, and thus generate a precise ego hand layout." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.791, + 0.788, + 0.837 + ], + "angle": 0, + "content": "Exo Contextual Feature Extraction As shown in Figure 2, an exo frame \\( x_{i} \\in \\mathbb{R}^{H \\times W \\times 3} \\) concatenated with the corresponding exo layout \\( x_{i}^{l} \\in \\mathbb{R}^{H \\times W \\times 3} \\) (2D hand pose layout recorded as image size) is first split into a sequence of patches" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.343, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Lifting the Egocentric Perspective from Exocentric Videos" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.253 + ], + "angle": 0, + "content": "and then processed by an input embedding layer, such as patch embedding for ViTs [12] to get a tokenized input sequence \\( e_i \\in \\mathbb{R}^{M \\times D} \\). To capture the positional information which is critical for spatial relationship reasoning, we add a learnable position embedding to the sequence of patches. Then the sequence of tokens is passed into a transformer encoder consisting of \\( N \\) repeated blocks. Each block contains a multi-head attention token mixer to first communicate information among tokens, which is represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.365, + 0.263, + 0.786, + 0.281 + ], + "angle": 0, + "content": "\\[\nz _ {i} ^ {j - 1} = \\operatorname {A t t e n t i o n} \\left(\\operatorname {N o r m} \\left(e _ {i} ^ {j - 1}\\right)\\right) + e _ {i} ^ {j - 1}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.291, + 0.784, + 0.321 + ], + "angle": 0, + "content": "where \\( j \\in \\{1, \\dots, N\\} \\) and \\( \\mathrm{Norm}(\\cdot) \\) is the normalization method. Then, the mixed token \\( z_i^{j-1} \\) is passed into a two-layer MLP expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.333, + 0.785, + 0.351 + ], + "angle": 0, + "content": "\\[\ne _ {i} ^ {j} = \\operatorname {M L P} \\left(\\operatorname {N o r m} \\left(z _ {i} ^ {j - 1}\\right)\\right) + z _ {i} ^ {j - 1}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.361, + 0.784, + 0.421 + ], + "angle": 0, + "content": "After iterating through all \\(N\\) blocks, the input sequence \\(e_i\\) undergoes a mapping process that transforms it into the final contextual tokens \\(c_{i}\\in \\mathbb{R}^{M\\times D}\\). We talk about how to decode the contextual tokens into the desired ego layout in the following part." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.435, + 0.785, + 0.572 + ], + "angle": 0, + "content": "Layout Decoder Consider a hand pose image with shape \\(\\{H\\times W\\times 3\\}\\) as the target ego layout. The contextual tokens \\(c_{i}\\) should be decoded into 2D hand joint coordinates and then rendered to the RGB image. First, the contextual tokens \\(c_{i}\\) are combined with a sequence of learnable query embeddings \\(q_{i}\\in \\mathbb{R}^{E\\times D}\\) and passed into a decoder architecture, consisting of \\(K\\) transformer blocks, as described earlier. Following the transformer blocks, a regression head is employed to estimate the 2D coordinates of the joints whose range is restricted to [0,1]. Then we apply the classic bipartite matching loss [8, 39] to map the regressed hand joints to the ground truth joints." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.593, + 0.553, + 0.607 + ], + "angle": 0, + "content": "3.3 Diffusion-based Pixel Hallucination" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.617, + 0.784, + 0.813 + ], + "angle": 0, + "content": "Having inferred the ego hand pose layout \\( y_{i}^{l} \\), next the aim of pixel-level hallucination is to synthesize photorealistic ego frames by taking into account an exo frame \\( x_{i} \\) and the target ego layout. We use the diffusion formulation proposed in the denoising diffusion probabilistic model (DDPM) [18] and train the diffusion model in the latent space. As shown in Figure 2, we first adopt a pre-trained variational autoencoder (VAE) model [25] as used in [44] to encode the original ego frames \\( y_{i} \\) and conditional information exo frame \\( x_{i} \\) and ego layout \\( y_{i}^{l} \\) into the latent space. Then, we train a diffusion transformer [38] to learn the latent data distribution by denoising a latent vector sampled from a Gaussian distribution gradually. Specifically, given an initial noise map \\( \\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I}) \\) and a conditioning vector \\( \\mathbf{d} \\), the diffusion model generates the corresponding ego latent \\( \\hat{\\mathbf{z}}_{\\theta} \\). A squared error loss is used to denoise a variably-noised latent code \\( \\mathbf{z}^{n} = \\alpha^{n}\\mathbf{z} + \\sigma^{n}\\epsilon \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.825, + 0.785, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} _ {\\mathbf {d}, \\mathbf {z}, \\boldsymbol {\\epsilon}, n} \\left[ \\| \\mathbf {z} - \\hat {\\mathbf {z}} _ {\\theta} \\left(\\alpha^ {n} \\mathbf {z} + \\sigma^ {n} \\boldsymbol {\\epsilon}, \\mathbf {d}\\right) \\| _ {2} ^ {2} \\right] \\tag {4}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "M. Luo et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.299 + ], + "angle": 0, + "content": "where \\(\\mathbf{z}\\) is the ground-truth latent vector, \\(\\mathbf{d}\\) is the conditioning vector, and \\(\\alpha^n,\\sigma^n\\) are functions of the diffusion process time step \\(n\\), which control the noise schedule and sample quality. We simply concatenate the noisy latent vector \\(\\mathbf{z}^n\\) and the conditional embedding \\(\\mathbf{d}\\) as the input of the denoising transformer. In contrast to earlier cross-view translation methods based on GANs [30, 61, 62], our conditional diffusion model sequentially updates the target ego outputs, excels in capturing intricate ego-exo dependencies and faithfully reproducing the mapping of the conditional distributions in Eqn. (1). Notably, our model exhibits enhanced stability throughout the training and consistently produces higher-quality samples, as verified by our experiments." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.32, + 0.453, + 0.336 + ], + "angle": 0, + "content": "3.4 Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.435 + ], + "angle": 0, + "content": "At the training stage, the two modules described in Section 3.2 and 3.3 are trained separately. During inference, we first generate the corresponding ego hand layout prediction \\(\\hat{y}_i^l\\) for each input exo frame \\(x_i\\) with the layout translator described in Section 3.2. Subsequently, \\(\\hat{y}_i^l\\) is concatenated with \\(x_i\\) in a channel-wise manner and then used as the conditional input of the pixel-level diffusion generator described in Section 3.3." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.457, + 0.494, + 0.475 + ], + "angle": 0, + "content": "4 Experimental Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.787, + 0.52 + ], + "angle": 0, + "content": "We introduce an exo-to-ego synthesis benchmark (Sec. 4.1) and then validate our approach against state-of-the-art synthesis methods (Sec. 4.2)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.541, + 0.651, + 0.556 + ], + "angle": 0, + "content": "4.1 A Benchmark for Exo-to-Ego View Translation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.791, + 0.81 + ], + "angle": 0, + "content": "To facilitate new work on the exo-to-ego translation task, we contribute a new cross-view synthesis benchmark sourced from four public time-synchronized multi-view datasets: H2O [27], Aria Pilot [32], Assembly101 [47], and Ego-Exo4D [16]. H2O [27] has indoor videos featuring actors manipulating 8 different objects using both hands. Assembly101 [47] shows people (dis)assembling 101 take-apart toy vehicles. We take 6 activity sequences involving different individuals manipulating a toy roller. Aria Pilot [32] provides 16 multi-view recordings of an actor wearing Project Aria glasses [54] manipulating YCB [7] objects on a desktop. For Ego-Exo4D, we include the four (largely non-table-top) scenarios involving heavy hand-object interactions from the EgoHandPose Benchmark - bike repair, cooking, music, and health—where the camera wearer also moves about the scene. For all datasets, we select the exo camera viewpoints that offer the clearest capture of the action. All frames are cropped and resized into \\(256 \\times 256\\). We divide each video into 30-frame clips. Overall, this dataset showcases an array of human activities with various objects, from toy assembly to the manipulation of everyday objects. More details are provided in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.811, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We employ the following split settings to evaluate four kinds of generalization: (1) new actions, where we put the first \\(80\\%\\) of a video's clips into the training" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.344, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Lifting the Egocentric Perspective from Exocentric Videos" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.803, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.553, + 0.761, + 0.568 + ], + "angle": 0, + "content": "Fig. 3: Qualitative examples when generalizing to new actions on all datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.574, + 0.788, + 0.786 + ], + "angle": 0, + "content": "set and the remaining \\(20\\%\\) in the test set, hence splitting up action steps over time (e.g., for a video depicting picking up a coffee box and then taking coffee capsules out, the former will be assigned to the training set and the latter to the test set); (2) new objects, where we train with videos involving any of 7 objects and test with videos containing a novel 8th object; (3) new subjects, where we train with all clips from one subject and test on clips from another (subject \\(1 \\rightarrow\\) subject 2); (4) new scenes (backgrounds), where we train and test with disjoint scenes (scene h1, h2, k1, k2 \\(\\rightarrow\\) scene o1, o2). We analyze new action generalization on all 3 datasets, and the rest on H2O only, as it is the only dataset with the variations and meta-data to allow such split settings. For action generalization, H2O, Aria Pilot, Assembly101, and Ego-Exo4D have 704, 343, 682, and 543 training clips, and 199, 95, 205, and 211 test clips, respectively. For other generalization settings (new objects, subjects, and scenes), there are 691, 704, and 591 training clips, and 108, 194, and 312 test clips." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.804, + 0.502, + 0.82 + ], + "angle": 0, + "content": "4.2 Evaluating Synthesis Quality" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.825, + 0.785, + 0.841 + ], + "angle": 0, + "content": "With this well-constructed benchmark in hand, we proceed to validate Exo2Ego." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "M. Luo et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.173 + ], + "angle": 0, + "content": "Table 1: Generalizing to new actions on three datasets. *pixelNeRF requires privileged camera information relative to the other baseline models." + }, + { + "type": "table", + "bbox": [ + 0.263, + 0.185, + 0.737, + 0.508 + ], + "angle": 0, + "content": "
H2O
MethodSSIM↑PSNR↑FID↓P_Squeeze↓P_Alex↓P_Vgg↓Feasi.↑
pix2pixHD [62]0.42830.370132.030.1630.2200.3370.8952
P-GAN [30]0.27229.079192.260.2420.3080.4490.8353
vid2vid [61]0.40229.88285.760.1660.2260.3420.9116
pixelNeRF* [70]0.21927.871470.130.7110.8070.7310.0086
DiT [38]0.43630.55345.230.1290.1780.2970.9604
Exo2Ego (Ours)0.44330.58037.030.1040.1460.2620.9764
Aria Pilot
MethodSSIM↑PSNR↑FID↓P_Squeeze↓P_Alex↓P_Vgg↓Feasi.↑
pix2pixHD [62]0.35029.553159.910.3040.3620.4800.0790
P-GAN [30]0.34329.769122.400.2870.3450.4750.1014
vid2vid [61]0.35929.85852.380.2660.3280.4430.2646
DiT [38]0.36829.94532.710.2400.2980.4170.4784
Exo2Ego (Ours)0.38030.04725.990.2250.2790.4020.5977
Assembly101
MethodSSIM↑PSNR↑FID↓P_Squeeze↓P_Alex↓P_Vgg↓Feasi.↑
pix2pixHD [62]0.40529.909112.710.2160.2950.4010.1184
P-GAN [30]0.37229.303114.680.2160.3000.4360.0800
vid2vid [61]0.36829.48178.470.2240.3120.4240.1358
DiT [38]0.39829.99626.330.1790.2550.3700.3092
Exo2Ego (Ours)0.41330.15325.230.1680.2380.3550.4019
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.513, + 0.789, + 0.68 + ], + "angle": 0, + "content": "Baselines We consider several baselines: (1) Pix2PixHD [62], a single-view image translation method that processes the videos frame by frame; (2) P-GAN [30], a recent exo-to-ego view translation method that proposes a parallel generative network to facilitate cross-view image translation; (3) DiT [38], a pixel-to-pixel generative model from the latent diffusion model family. (4) Vid2Vid [61], a single-view video translation method that models the temporal dynamics in the video; (5) pixelNeRF [70], a NeRF-like model known for its remarkable generalization capabilities and its accommodating stance toward sparser input views. Note that pixelNeRF assumes camera parameters are known, while other methods do not. We conduct experiments with pixelNeRF only on H2O, as the other two datasets lack the camera information needed by pixelNeRF." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.682, + 0.789, + 0.773 + ], + "angle": 0, + "content": "Implementation Details For Exo2Ego's high-level structure transformation transformer, the number of blocks \\(N\\) is set to 6. We use DiT-XL/2 [38] as the denoising architecture for Exo2Ego. The diffusion model is trained for 40,000 steps and pixelNeRF is trained for 70,000 steps for all datasets. For H2O and Assembly101, pix2pixHD, and P-GAN are trained for 100 epochs, and vid2vid is trained for 40 epochs. For Aria Pilot, we use 400 and 100 epochs respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.775, + 0.789, + 0.837 + ], + "angle": 0, + "content": "Evaluation Metrics We measure the quality of the synthesized videos as follows. First, we evaluate the feasibility of generated hands, denoted as Feasi. Inspired by [69], we adopt an off-the-shelf egocentric hand object detector [49] to measure the physical feasibility of the synthesized hands. This metric is calcu" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.344, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Lifting the Egocentric Perspective from Exocentric Videos" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.147, + 0.805, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.327, + 0.785, + 0.355 + ], + "angle": 0, + "content": "Fig. 4: Qualitative examples when generalizing to new objects, subjects, and scenes (backgrounds) on the H2O dataset." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.363, + 0.784, + 0.392 + ], + "angle": 0, + "content": "lated as the confidence score of hand detection results produced by the detector, averaged over all synthesized frames." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.393, + 0.808, + 0.53 + ], + "angle": 0, + "content": "We also use structural similarity (SSIM) and peak signal-to-noise ratio (PSNR) [20] to check the pixel-level similarity between the predicted ego frame and the ground truth (GT) ego frame. These two evaluation metrics should ideally remain stable under minor transformations (such as small translations or affine transformations). We further adopt perceptual metrics, including FID [46] and Learned Perceptual Image Patch Similarity (LPIPS) [74] which evaluates the feature-level similarity of the predicted ego frame and the true ego frame. LPIPS are denoted as \\( P_{\\text{Squeeze}} \\), \\( P_{\\text{Alex}} \\), and \\( P_{\\text{Vgg}} \\), which uses SqueezeNet [21], AlexNet [26], and VGG [52] as the feature extractor." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Generalization to New Actions Figure 3 showcases qualitative results on H2O, Aria Pilot, and Assembly101, demonstrating the superiority of our Exo2Ego framework. Compared to all baselines, our approach produces realistic hands with correct poses, especially noticeable in the highlighted yellow circle regions when dealing with new actions during test time. Table 1 presents the feasibility of synthesized hands and the various measures of pixel-level similarity and perceptual similarity between the predicted ego frame and the ground truth ego frame. Our Exo2Ego greatly increases the feasibility of synthesized hands, with a confidence gain of up to \\(11.93\\%\\) on Aria Pilot, on the basis of the second-highest score of \\(47.84\\%\\) achieved by DiT. This improvement indicates enhanced realism in the generated hands, resulting in greater confidence in hand detection results. Additionally, our Exo2Ego can produce ego frames with higher pixel-level similarity to the ground truth frames and better perceptual metrics (much lower FID scores). This highlights the benefits of the 'decoupling' idea of our Exo2Ego framework in generating photorealistic ego frames." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Looking at Table 1, there is a trend among our Exo2Ego and the baselines — doing better on H2O dataset. We conjecture that the superiority comes from H2O's higher hand quality, including clear pose visibility and larger spatial coverage within the field of view, in both exo and ego view." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "M. Luo et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 2: Generalizing to different objects, subjects, and scenes. *pixelNeRF requires privileged camera information relative to the other baseline models." + }, + { + "type": "table", + "bbox": [ + 0.26, + 0.185, + 0.737, + 0.537 + ], + "angle": 0, + "content": "
H2O-unseen object
MethodSSIM↑PSNR↑FID↓P_Squeeze↓P_Alex↓P_Vgg↓Feasi.↑
pix2pixHD [62]0.33329.504253.730.2690.3600.4670.7527
P-GAN [30]0.27028.915264.760.2770.3500.4920.7720
vid2vid [61]0.34929.585145.390.2420.3240.4280.8612
pixelNeRF* [70]0.26227.912473.730.7500.8410.7360.0000
DiT [38]0.33829.740105.180.2140.2880.4100.9689
Exo2Ego (Ours)0.34429.786101.450.2080.2810.4030.9786
H2O-unseen subject
MethodSSIM↑PSNR↑FID↓P_Squeeze↓P_Alex↓P_Vgg↓Feasi.↑
pix2pixHD [62]0.06928.028477.470.6610.7540.7650.0140
P-GAN [30]0.08828.063451.370.6230.7080.7250.2002
vid2vid [61]0.13628.094308.520.5640.6900.6860.3838
pixelNeRF* [70]0.17027.874459.490.7690.8360.7680.0000
DiT [38]0.17228.185183.060.5240.6300.6820.8405
Exo2Ego (Ours)0.17828.266150.870.5030.6290.6650.9666
H2O-unseen scene
MethodSSIM↑PSNR↑FID↓P_Squeeze↓P_Alex↓P_Vgg↓Feasi.↑
pix2pixHD [62]0.02427.918558.870.7290.7890.7880.0013
P-GAN [30]0.08527.928325.340.6080.6990.7240.6041
vid2vid [61]0.03327.942398.780.6360.7310.7340.2095
pixelNeRF* [70]0.10327.810506.100.7730.9080.8000.0000
DiT [38]0.14027.874308.980.5840.7290.6780.8317
Exo2Ego (Ours)0.15727.943249.040.5310.7090.6580.9635
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.543, + 0.788, + 0.665 + ], + "angle": 0, + "content": "We further explored how Exo2Ego fares in the settings with heavy hand-object manipulation and frequent head motion, but minimal movement across the broader scene on Ego-Exo4D [16]. As shown in Figure 6, we compare our Exo2Ego with a strong image-to-image translation model ControlNet [73], which is built on a stable diffusion trained on large scales of data. It is observed that our Exo2Ego synthesizes feasible hands with correct poses, unlike ControlNet. The exo image's hand regions in Ego-Exo4D are small, so we crop it to use as additional input to offer a better view of the hands for both methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.679, + 0.788, + 0.832 + ], + "angle": 0, + "content": "Generalization to New Objects, Subjects, and Backgrounds Illustrated in Table 2, our Exo2Ego outperforms the baselines in terms of quantitative metrics for generalization to new objects/subjects/backgrounds. Figure 4 shows that our Exo2Ego can produce realistic hand-object interactions when encountering exo views with new distributions, while other baselines fail miserably. This is mainly because our diffusion-based generative mechanism gradually denoises the target ego view, capturing the inherent intricate mapping from exo to ego views better. Another observation is that pixelNeRF produces ego views characterized by noticeable blurring and a lack of fine details in hands and objects. This observation supports our earlier discussions on the limitations of 3D-aware approaches." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.344, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Lifting the Egocentric Perspective from Exocentric Videos" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.144, + 0.78, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.22, + 0.313, + 0.781, + 0.327 + ], + "angle": 0, + "content": "Fig. 5: Exo2Ego framework generates ego videos with reasonable viewpoint changes." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.333, + 0.783, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.467, + 0.785, + 0.482 + ], + "angle": 0, + "content": "Fig. 6: Qualitative examples when generalizing to new actions on Ego-Exo4D dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.789, + 0.687 + ], + "angle": 0, + "content": "Modeling Egocentric Viewpoint Change Due to the probabilistic nature of our Exo2Ego framework, it excels in capturing the relationship between the viewpoint changes in the ego view and the head motions observed in the exo view, holding the potential to deliver more immersive virtual experiences. Intrinsically, our Exo2Ego framework leverages the fact that ego-camera pose is conditioned on the reference exo view, as the ego-camera wearer is usually visible in the exo view. This is highlighted by the yellow gaze direction and dashed box in Figure 5. As the actor moves their head when interacting with objects, the slope of the desk edge in the ego view also changes, indicating a corresponding change in ego viewpoint. This implies the ego-camera motion can be implicitly inferred via conditional generative modeling, even if it's not enforced explicitly. The ability to model ego viewpoint changes is crucial, especially in physical activities like playing ball games which involve heavy head and body motions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.709, + 0.536, + 0.725 + ], + "angle": 0, + "content": "5 Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.741, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Temporal Coherence Our Exo2Ego operates on each frame individually — the generation of the current ego frame is solely conditioned on the current exo input frame. There is no explicit consideration of previously generated ego frames or past exo frames. Temporal coherence is orthogonal to our current contribution and can be used to improve any of the baseline models without such design." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "M. Luo et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.777, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.207, + 0.785, + 0.234 + ], + "angle": 0, + "content": "Fig. 7: Failure cases: Exo2Ego generates hands with reasonably accurate positions/poses but incomplete/wrong objects in the ego view." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.243, + 0.788, + 0.485 + ], + "angle": 0, + "content": "Generalization Ability As an early-stage effort for exo-to-ego cross-view translation, our work makes non-trivial progress in producing reasonable hand manipulation details for novel actions. However, despite establishing stronger generalizations than the existing models, our model does not generalize perfectly to in-the-wild objects, subjects, and backgrounds, due to the limited scale of available training data. Considering the intricate nature of the generalization to new environments, we regard it as a challenging future research direction. Nonetheless, our focus on generalizing to new actions within the same environment is meaningful for several applications. Some examples: Real-time monitoring during physical rehabilitation (exo-to-ego view translation can help therapists guide patients during rehabilitation sessions), assistive cooking aid for visually impaired individuals (ego views are provided to the visually impaired individuals to cook independently), or real-time fitness training monitoring (coaches use synthesized ego views to help the trainee correct the fitness movements) — in each case, the environment is stable between training and testing, and yet the model must generalize to new actions for cross-view translation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.499, + 0.788, + 0.665 + ], + "angle": 0, + "content": "3D prior Our qualitative results expose limitations in existing baselines and our Exo2Ego framework in generating 3D-consistent views for objects during test time, attributed to the absence of geometric priors for common objects. See Figure 7. Future work involves integrating robust object geometric priors. While leveraging predicted camera parameters for encoding 3D geometry is a promising approach, as shown Figure 4, the popular 3D-aware method pixelNeRF [70] faces challenges in the dynamic exo-ego setting characterized by hand and object interactions, due to limitations in handling sparse inputs, occluded scenes, and long-range camera extrapolations. Notably, our Exo2Ego outperforms pixelNeRF, highlighting the potential for future work to explore explicit geometric reasoning within our framework." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.688, + 0.36, + 0.705 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Overall, our work contributes to the growing body of research in cross-view translation and lays the groundwork for the important case of exo-to-ego. Our generative framework Exo2Ego integrates high-level structure transformation and pixel-level hallucination to yield very encouraging experimental results. We also provide a curated benchmark task for supporting continued work on this problem. The core technology holds immense potential for applications in robot learning and AR skill coaching, where an ego actor needs to replicate the actions of a demonstrator observed from the exo perspective." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.343, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Lifting the Egocentric Perspective from Exocentric Videos" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "Acknowledgment: UT Austin is supported in part by NSF Grants AF 1901292, CNS 2148141, Tripods CCF 1934932, IFML CCF 2019844 and research gifts by Western Digital, Amazon, WNCG IAP, UT Austin Machine Learning Lab (MLL), Cisco, and the Stanly P. Finch Centennial Professorship in Engineering. K.G. is paid as a research scientist at Meta." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.242, + 0.323, + 0.258 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.269, + 0.786, + 0.325 + ], + "angle": 0, + "content": "1. Ardeshir, S., Borji, A.: Ego2top: Matching viewers in egocentric and top-view videos. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14. pp. 253-268. Springer (2016) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.326, + 0.786, + 0.352 + ], + "angle": 0, + "content": "2. Ardeshir, S., Borji, A.: Egocentric meets top-view. IEEE transactions on pattern analysis and machine intelligence 41(6), 1353-1366 (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.353, + 0.786, + 0.379 + ], + "angle": 0, + "content": "3. Ardeshir, S., Borji, A.: An exocentric look at egocentric actions and vice versa. Computer Vision and Image Understanding 171, 61-68 (2018) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.379, + 0.786, + 0.405 + ], + "angle": 0, + "content": "4. Bahl, S., Gupta, A., Pathak, D.: Human-to-robot imitation in the wild. arXiv preprint arXiv:2207.09450 (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.786, + 0.446 + ], + "angle": 0, + "content": "5. Bansal, A., Ma, S., Ramanan, D., Sheikh, Y.: Recycle-gan: Unsupervised video retargeting. In: Proceedings of the European conference on computer vision (ECCV). pp. 119-135 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.446, + 0.786, + 0.472 + ], + "angle": 0, + "content": "6. Bharadhwaj, H., Gupta, A., Tulsiani, S., Kumar, V.: Zero-shot robot manipulation from passive human videos. arXiv preprint arXiv:2302.02011 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.473, + 0.786, + 0.527 + ], + "angle": 0, + "content": "7. Calli, B., Singh, A., Walsman, A., Srinivasa, S., Abbeel, P., Dollar, A.M.: The ycb object and model set: Towards common benchmarks for manipulation research. In: 2015 international conference on advanced robotics (ICAR). pp. 510-517. IEEE (2015) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.528, + 0.786, + 0.582 + ], + "angle": 0, + "content": "8. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 213-229. Springer (2020) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.583, + 0.786, + 0.623 + ], + "angle": 0, + "content": "9. Chan, E.R., Nagano, K., Chan, M.A., Bergman, A.W., Park, J.J., Levy, A., Aittala, M., De Mello, S., Karras, T., Wetzstein, G.: Generative novel view synthesis with 3d-aware diffusion models. arXiv preprint arXiv:2304.02602 (2023) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.623, + 0.786, + 0.663 + ], + "angle": 0, + "content": "0. Cheng, F., Luo, M., Wang, H., Dimakis, A., Torresani, L., Bertasius, G., Grauman, K.: 4diff: 3d-aware diffusion model for third-to-first viewpoint translation. In: ECCV (2024) 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.663, + 0.786, + 0.69 + ], + "angle": 0, + "content": "1. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems 34, 8780-8794 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.691, + 0.786, + 0.745 + ], + "angle": 0, + "content": "12. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.745, + 0.786, + 0.786 + ], + "angle": 0, + "content": "13. Elfeki, M., Regmi, K., Ardeshir, S., Borji, A.: From third person to first person: Dataset and baselines for synthesis and retrieval. arXiv preprint arXiv:1812.00104 (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.786, + 0.786, + 0.84 + ], + "angle": 0, + "content": "14. Fan, C., Lee, J., Xu, M., Kumar Singh, K., Jae Lee, Y., Crandall, D.J., Ryoo, M.S.: Identifying first-person camera wearers in third-person videos. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5125-5133 (2017) 3" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.269, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "M. Luo et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "15. Grauman, K., Westbury, A., Torresani, L., Kitani, K., Malik, J., Afouras, T., Ashutosh, K., Baiyya, V., Bansal, S., Boote, B., et al.: Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. arXiv preprint arXiv:2311.18259 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.787, + 0.271 + ], + "angle": 0, + "content": "16. Grauman, K., Westbury, A., Torresani, L., Kitani, K., Malik, J., Afouras, T., Ashutosh, K., Baiyya, V., Bansal, S., Boote, B., et al.: Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19383-19400 (2024) 3, 8, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.787, + 0.312 + ], + "angle": 0, + "content": "17. Ho, H.I., Chiu, W.C., Wang, Y.C.F.: Summarizing first-person videos from third persons' points of view. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 70-85 (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.312, + 0.787, + 0.339 + ], + "angle": 0, + "content": "18. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33, 6840-6851 (2020) 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.339, + 0.787, + 0.379 + ], + "angle": 0, + "content": "19. Ho, J., Sahara, C., Chan, W., Fleet, D.J., Norouzi, M., Salimans, T.: Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research 23(1), 2249-2281 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.787, + 0.406 + ], + "angle": 0, + "content": "20. Hore, A., Ziou, D.: Image quality metrics: Psnr vs. sSIM. In: 2010 20th international conference on pattern recognition. pp. 2366-2369. IEEE (2010) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.406, + 0.787, + 0.447 + ], + "angle": 0, + "content": "21. Iandola, F.N., Han, S., Moskewicz, M.W., Ashraf, K., Dally, W.J., Keutzer, K.: SqueezeNET: Alexnet-level accuracy with 50x fewer parameters and \\(< 0.5\\) mb model size. arXiv preprint arXiv:1602.07360 (2016) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.787, + 0.488 + ], + "angle": 0, + "content": "22. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1125-1134 (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.528 + ], + "angle": 0, + "content": "23. Jang, W., Agapito, L.: Codenerf: Disentangled neural radiance fields for object categories. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12949-12958 (2021) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.787, + 0.569 + ], + "angle": 0, + "content": "24. Jangir, R., Hansen, N., Ghosal, S., Jain, M., Wang, X.: Look closer: Bridging egocentric and third-person views with transformers for robotic manipulation. IEEE Robotics and Automation Letters 7(2), 3046-3053 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.787, + 0.595 + ], + "angle": 0, + "content": "25. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.595, + 0.787, + 0.622 + ], + "angle": 0, + "content": "26. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. Communications of the ACM 60(6), 84-90 (2017) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.622, + 0.787, + 0.676 + ], + "angle": 0, + "content": "27. Kwon, T., Tekin, B., Stühmer, J., Bogo, F., Pollefeys, M.: H2o: Two hands manipulating objects for first person interaction recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10138-10148 (October 2021) 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.676, + 0.787, + 0.731 + ], + "angle": 0, + "content": "28. Li, Y., Nagarajan, T., Xiong, B., Grauman, K.: Ego-exo: Transferring visual representations from third-person to first-person videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6943-6953 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.731, + 0.787, + 0.785 + ], + "angle": 0, + "content": "29. Liu, A., Tucker, R., Jampani, V., Makadia, A., Snavely, N., Kanazawa, A.: Infinite nature: Perpetual view generation of natural scenes from a single image. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14458-14467 (2021) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "30. Liu, G., Tang, H., Latapie, H., Yan, Y.: Exocentric to egocentric image generation via parallel generative adversarial network. In: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1843-1847. IEEE (2020) 3, 4, 8, 10, 12" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.344, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Lifting the Egocentric Perspective from Exocentric Videos" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "31. Liu, G., Tang, H., Latapie, H.M., Corso, J.J., Yan, Y.: Cross-view exocentric to egocentric video synthesis. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 974-982 (2021) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.272 + ], + "angle": 0, + "content": "32. Lv, Z., Miller, E., Meissner, J., Pesqueira, L., Sweeney, C., Dong, J., Ma, L., Patel, P., Moulon, P., Somasundaram, K., Parkhi, O., Zou, Y., Raina, N., Saarinen, S., Mansour, Y.M., Huang, P.K., Wang, Z., Troynikov, A., Artal, R.M., DeTone, D., Barnes, D., Argall, E., Lobanovskiy, A., Kim, D.J., Bouttefroy, P., Straub, J., Engel, J.J., Gupta, P., Yan, M., Nardi, R.D., Newcombe, R.: Aria pilot dataset. https://about.facebook.com/realitylabs/projectaria/datasets (2022) 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.327 + ], + "angle": 0, + "content": "33. Majumdar, A., Yadav, K., Arnaud, S., Ma, Y.J., Chen, C., Silwal, S., Jain, A., Berges, V.P., Abbeel, P., Malik, J., et al.: Where are we in the search for an artificial visual cortex for embodied intelligence? arXiv preprint arXiv:2303.18240 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.355 + ], + "angle": 0, + "content": "34. Mandikal, P., Grauman, K.: Dexvip: Learning dexterous grasping with human hand pose priors from video. In: Conference on Robot Learning (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.396 + ], + "angle": 0, + "content": "35. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.397, + 0.788, + 0.424 + ], + "angle": 0, + "content": "36. Nair, S., Rajeswaran, A., Kumar, V., Finn, C., Gupta, A.: R3m: A universal visual representation for robot manipulation. arXiv preprint arXiv:2203.12601 (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.424, + 0.788, + 0.48 + ], + "angle": 0, + "content": "37. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5480-5490 (2022) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.48, + 0.788, + 0.507 + ], + "angle": 0, + "content": "38. Peebles, W., Xie, S.: Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748 (2022) 3, 7, 10, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.55 + ], + "angle": 0, + "content": "39. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified, real-time object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 779-788 (2016) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.55, + 0.788, + 0.59 + ], + "angle": 0, + "content": "40. Regmi, K., Borji, A.: Cross-view image synthesis using conditional gans. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.591, + 0.788, + 0.646 + ], + "angle": 0, + "content": "41. Regmi, K., Borji, A.: Cross-view image synthesis using geometry-guided conditional gans. Computer Vision and Image Understanding (2019). https://doi.org/https://doi.org/10.1016/j.cviu.2019.07.008, http://www.sciencedirect.com/science/article/pii/S1077314219301043 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.646, + 0.788, + 0.674 + ], + "angle": 0, + "content": "42. Ren, B., Tang, H., Sebe, N.: Cascaded cross mlp-mixer gans for cross-view image translation. arXiv preprint arXiv:2110.10183 (2021) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.715 + ], + "angle": 0, + "content": "43. Ren, X., Wang, X.: Look outside the room: Synthesizing a consistent long-term 3d scene video from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3563-3573 (2022) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.715, + 0.788, + 0.77 + ], + "angle": 0, + "content": "44. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022) 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.771, + 0.788, + 0.812 + ], + "angle": 0, + "content": "45. Rombach, R., Esser, P., Ommer, B.: Geometry-free view synthesis: Transformers and no 3d priors. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14356-14366 (2021) 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.812, + 0.788, + 0.841 + ], + "angle": 0, + "content": "46. Seitzer, M.: pytorch-fid: FID Score for PyTorch. https://github.com/mseitzer/pytorch-fid (August 2020), version 0.3.0 11" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "M. Luo et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "47. Sener, F., Chatterjee, D., Shelepov, D., He, K., Singhania, D., Wang, R., Yao, A.: Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21096-21106 (2022) 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.205, + 0.787, + 0.259 + ], + "angle": 0, + "content": "48. Sermanet, P., Lynch, C., Chebotar, Y., Hsu, J., Jang, E., Schaal, S., Levine, S.: Time-contrastive networks: Self-supervised learning from video. Proceedings of International Conference in Robotics and Automation (ICRA) (2018), http://arxiv.org/abs/1704.06888 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.787, + 0.3 + ], + "angle": 0, + "content": "49. Shan, D., Geng, J., Shu, M., Fouhey, D.F.: Understanding human hands in contact at internet scale. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.787, + 0.355 + ], + "angle": 0, + "content": "50. Sigurdsson, G.A., Gupta, A., Schmid, C., Farhadi, A., Alahari, K.: Actor and observer: Joint modeling of first and third-person videos. In: proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7396-7404 (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.787, + 0.397 + ], + "angle": 0, + "content": "51. Sigurdsson, G.A., Gupta, A., Schmid, C., Farhadi, A., Alahari, K.: Charades-ego: A large-scale dataset of paired third and first person videos. arXiv preprint arXiv:1804.09626 (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.424 + ], + "angle": 0, + "content": "52. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.467 + ], + "angle": 0, + "content": "53. Sitzmann, V., Zollhöfer, M., Wetzstein, G.: Scene representation networks: Continuous 3d-structure-aware neural scene representations. Advances in Neural Information Processing Systems 32 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.787, + 0.508 + ], + "angle": 0, + "content": "54. Somasundaram, K., Dong, J., Tang, H., Straub, J., Yan, M., Goesele, M., Engel, J.J., De Nardi, R., Newcombe, R.: Project aria: A new tool for egocentric multimodal ai research. arXiv preprint arXiv:2308.13561 (2023) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.787, + 0.562 + ], + "angle": 0, + "content": "55. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=PxTIG12RRHS 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.787, + 0.619 + ], + "angle": 0, + "content": "56. Soran, B., Farhadi, A., Shapiro, L.: Action recognition in the presence of one egocentric and multiple static cameras. In: Computer Vision-ACCV 2014: 12th Asian Conference on Computer Vision, Singapore, Singapore, November 1-5, 2014, Revised Selected Papers, Part V 12. pp. 178-193. Springer (2015) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.787, + 0.674 + ], + "angle": 0, + "content": "57. Tang, H., Xu, D., Sebe, N., Wang, Y., Corso, J.J., Yan, Y.: Multi-channel attention selection gan with cascaded semantic guidance for cross-view image translation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2417-2426 (2019) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.715 + ], + "angle": 0, + "content": "58. Tseng, H.Y., Li, Q., Kim, C., Alsisan, S., Huang, J.B., Kopf, J.: Consistent view synthesis with pose-guided diffusion models. arXiv preprint arXiv:2303.17598 (2023) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.757 + ], + "angle": 0, + "content": "59. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.812 + ], + "angle": 0, + "content": "60. Wang, J., Liu, L., Xu, W., Sarkar, K., Luvizon, D., Theobalt, C.: Estimating egocentric 3d human pose in the wild with external weak supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13157-13166 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.841 + ], + "angle": 0, + "content": "61. Wang, T.C., Liu, M.Y., Zhu, J.Y., Liu, G., Tao, A., Kautz, J., Catanzaro, B.: Video-to-video synthesis. arXiv preprint arXiv:1808.06601 (2018) 3, 4, 8, 10, 12" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.344, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Lifting the Egocentric Perspective from Exocentric Videos" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.204 + ], + "angle": 0, + "content": "62. Wang, T.C., Liu, M.Y., Zhu, J.Y., Tao, A., Kautz, J., Catanzaro, B.: High-resolution image synthesis and semantic manipulation with conditional gans. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8798-8807 (2018) 3, 4, 6, 8, 10, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.205, + 0.787, + 0.244 + ], + "angle": 0, + "content": "63. Wang, Z., Wu, S., Xie, W., Chen, M., Prisacariu, V.A.: NeRF--: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064 (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.245, + 0.786, + 0.285 + ], + "angle": 0, + "content": "64. Watson, D., Chan, W., Martin-Brualla, R., Ho, J., Tagliasacchi, A., Norouzi, M.: Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628 (2022) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.287, + 0.787, + 0.342 + ], + "angle": 0, + "content": "65. Wen, Y., Singh, K.K., Anderson, M., Jan, W.P., Lee, Y.J.: Seeing the unseen: Predicting the first-person camera wearer's location and pose in third-person scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. pp. 3446-3455 (October 2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.786, + 0.383 + ], + "angle": 0, + "content": "66. Wiles, O., Gkioxari, G., Szeliski, R., Johnson, J.: Synsin: End-to-end view synthesis from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7467-7477 (2020) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.425 + ], + "angle": 0, + "content": "67. Xu, M., Fan, C., Wang, Y., Ryoo, M.S., Crandall, D.J.: Joint person segmentation and identification in synchronized first-and third-person videos. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 637-652 (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.426, + 0.786, + 0.452 + ], + "angle": 0, + "content": "68. Xue, Z., Grauman, K.: Learning fine-grained view-invariant representations from unpaired ego-exo videos via temporal alignment. In: NeurIPS (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.786, + 0.479 + ], + "angle": 0, + "content": "69. Ye, Y., Li, X., Gupta, A., Mello, S.D., Birchfield, S., Song, J., Tulsiani, S., Liu, S.: Affordance diffusion: Synthesizing hand-object interactions. In: CVPR (2023) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.481, + 0.787, + 0.535 + ], + "angle": 0, + "content": "70. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4578-4587 (June 2021) 2, 3, 4, 5, 10, 12, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.536, + 0.787, + 0.577 + ], + "angle": 0, + "content": "71. Yu, H., Cai, M., Liu, Y., Lu, F.: What i see is what you see: Joint attention learning for first and third person video co-analysis. In: Proceedings of the 27th ACM International Conference on Multimedia. pp. 1358-1366 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.578, + 0.787, + 0.618 + ], + "angle": 0, + "content": "72. Yu, H., Cai, M., Liu, Y., Lu, F.: First-and third-person video co-analysis by learning spatial-temporal joint attention. IEEE Transactions on Pattern Analysis and Machine Intelligence (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.619, + 0.787, + 0.646 + ], + "angle": 0, + "content": "73. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models (2023) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.674 + ], + "angle": 0, + "content": "74. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018) 11" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.674 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Put Myself in Your Shoes_ Lifting the Egocentric Perspective from Exocentric Videos/cf2ae3fa-fba8-4d91-a4a1-6c60e6a8095b_origin.pdf b/2024/Put Myself in Your Shoes_ Lifting the Egocentric Perspective from Exocentric Videos/cf2ae3fa-fba8-4d91-a4a1-6c60e6a8095b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f361d61acad88e989c35ae107a2c75e440121d30 --- /dev/null +++ b/2024/Put Myself in Your Shoes_ Lifting the Egocentric Perspective from Exocentric Videos/cf2ae3fa-fba8-4d91-a4a1-6c60e6a8095b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fb4422aed140b1438cee68cba13f76165f7081501bd8db0d0515c799f4a35a56 +size 4437575 diff --git a/2024/Put Myself in Your Shoes_ Lifting the Egocentric Perspective from Exocentric Videos/full.md b/2024/Put Myself in Your Shoes_ Lifting the Egocentric Perspective from Exocentric Videos/full.md new file mode 100644 index 0000000000000000000000000000000000000000..291b5c98e69255bc29d3bf4d19eebe344f262998 --- /dev/null +++ b/2024/Put Myself in Your Shoes_ Lifting the Egocentric Perspective from Exocentric Videos/full.md @@ -0,0 +1,272 @@ +# Put Myself in Your Shoes: Lifting the Egocentric Perspective from Exocentric Videos + +Mi Luo $^{1}$ , Zihui Xue $^{1,2}$ , Alex Dimakis $^{1}$ , and Kristen Grauman $^{1,2}$ + +1 The University of Texas at Austin +2 FAIR, Meta AI + +![](images/a6dcc9c18039fafa2af0d5b6d13f26b60c4b5268fa152bec010bcc160d59e948.jpg) +Fig. 1: We consider cross-view translation from exocentric to egocentric video. Given an exocentric video sequence with hand-object interactions (left), the goal is to generate the corresponding egocentric perspective (right). + +Abstract. We investigate exocentric-to-egocentric cross-view translation, which aims to generate a first-person (egocentric) view of an actor based on a video recording that captures the actor from a third-person (exocentric) perspective. To this end, we propose a generative framework called Exo2Ego that decouples the translation process into two stages: high-level structure transformation, which explicitly encourages cross-view correspondence between exocentric and egocentric views, and a diffusion-based pixel-level hallucination, which incorporates a hand layout prior to enhance the fidelity of the generated egocentric view. To pave the way for future advancements in this field, we curate a comprehensive exo-to-ego cross-view translation benchmark focused on hand-object manipulations. It consists of a diverse collection of synchronized ego-exo video pairs from four public datasets: H2O, Aria Pilot, Assembly101, and Ego-Exo4D. The experimental results validate that Exo2Ego delivers photorealistic video results with clear hand manipulation details and outperforms several baselines in terms of both synthesis quality and generalization to new actions. + +Keywords: Egocentric Hand-Object Interaction $\cdot$ View Synthesis + +# 1 Introduction + +Given a third-person video capturing a person opening a milk carton, what would the visual world look like from his perspective? See Figure 1. Due to mirror neurons in the human brain [3], we can easily envision the appearance + +and spatial relationships of the person's hands and the milk carton from a first-person perspective. However, existing computer vision models struggle to do the same thing, owing to the stark distinctions between the two viewpoints. + +We make a step towards addressing this underexplored problem: exocentric-to-egocentric cross-view translation. The goal is to synthesize the corresponding ego view of an actor from an exo video recording3, with minimal assumptions on the viewpoint relationships (e.g., camera parameters or accurate geometric scene structure). Specifically, we focus on synthesizing ego hand-object interactions in table-top activities, such as assembling toys, pouring milk, or picking up a box. + +This task can benefit many applications, ranging from robotics to virtual and augmented reality (VR/AR). For example, an AR assistant could transform the view shown in a third-person how-to video to demonstrate to a user how things should look from their own perspectives, e.g., revealing finger placements and strumming techniques for a guitar video. Similarly, in robot learning, the robot could better match its actions to those of a human instructor in the room by projecting his/her hand-object interactions to its own perspective. Indeed, recent work demonstrates that the human ego view is valuable for robot learning [4,6, 33,34,36], as robots also perceive the world through the egocentric lens. + +However, exo-to-ego view translation is highly challenging. It requires understanding the spatial relationships of visible hands and objects and inferring their pixel-level appearance in the novel ego view. Further, the task is not strictly geometric; it is inherently underdetermined. Some parts of an object may not be visible in the exo view—such as the inner pages of a book when only the cover is observed in the exo view—requiring the model to extrapolate the occluded parts. As a result, the recent popular geometry-aware novel view synthesis approaches [23, 35, 37, 70] are ill-equipped to solve this problem. The key reason is that they are regressive rather than generative, which limits their ability to deal with sparse input views (single exo view in our case) and major occlusions. + +In light of this, we propose a probabilistic exo-to-ego generative framework (termed Exo2Ego) that learns the conditional distribution of the target ego video given the exo video. Exo2Ego differs from generative-based methods for general novel view synthesis [9, 29, 43, 45, 58, 64, 66] as it relaxes the requirement for camera parameters as input. This expands the potential application scenarios since camera parameters are rarely accessible in real-world scenarios, e.g., taking photos or videos with a mobile phone. Additionally, accurately inferring camera parameters remains challenging without precise correspondences or a rigid lab setting for calibration [63]. Moreover, Exo2Ego stands apart from recent cross-view image-to-image translation methods [42, 57], as it does not require the ground truth semantic map (object segmentation) of the target domain at the inference stage—typically an infeasible assumption. + +Our key insights are to explicitly encourage cross-view correspondence by predicting the ego layout and to introduce a hand layout prior to boost the fidelity of generated hands in the ego view. Specifically, our Exo2Ego framework + +decouples the exo-to-ego view translation problem into two stages: (1) High-level Structure Transformation, which infers the rough location and interaction manner of the hands and objects in the ego view, using a transformer-based model to transform the exo hand-object interaction layout into its ego counterpart; and (2) Diffusion-based Pixel Hallucination, which trains a conditional diffusion model to refine the details on top of the ego hand-object interaction layout. Overall, Exo2Ego is a generative framework that offers a simple but effective baseline approach for the exo-to-ego view translation problem, accounting for the centrality of hand-object manipulations in this domain. + +To evaluate our approach and boost research in this field, we construct an exoto-ego cross-view synthesis benchmark, comprising synchronized ego-exo video pairs curated from four datasets: H2O [27], Aria Pilot [32], Assembly101 [47], and Ego-Exo4D [16]. These datasets offer a breadth of table-top activities, as well as some non-table-top activities where the camera wearer moves around the scene. Empirical results underscore the efficacy of our Exo2Ego framework, which produces realistic video outputs with distinct hand manipulation details. Exo2Ego surpasses single-view translation baselines based on GANs [61,62] and latent diffusion models [38], a recent cross-view synthesis approach [30], as well as a NeRF-based method [70], demonstrating superior generation quality and a marked improvement in generalization capability to new actions. + +# 2 Related work + +# 2.1 Relating Egocentric and Exocentric Views + +There have been several attempts to jointly understand ego and exo views [13, 15, 51]. Early efforts explore how to localize the person wearing a camera in a third-person view, given their egocentric video [1,2,14,65,67]. To bridge the ego-exo gap, other work learns view-invariant features from concurrent (paired) [3, 48,50,71,72] or unpaired [68] views, or boosts the latent ego signals in exo video during pretraining [28], summarization [17], and 3D pose estimation [60]. Additionally, fusing ego and exo views can improve action recognition [56] and robotic manipulation tasks [24]. Despite its significance, cross-view translation has received little attention. P-GAN [30] and STA-GAN [31] have explored cross-view image and video synthesis, respectively, yet they only examine basic activities such as walking, jogging, and running, limiting their applicability to more diverse scenarios. To address these limitations, we introduce a new benchmark for exoto-ego view synthesis covering some diverse activities, from assembling toys to manipulating everyday objects. Importantly, we relax STA-GAN's requirement for the ego semantic map during test time [31], enhancing the practical applicability. A concurrent work [10] examines the introduction of geometry priors to the diffusion model. + +# 2.2 Novel View Synthesis + +Exo-to-ego view translation also relates to novel view synthesis, i.e., given a set of images of a scene, infer how the scene looks from novel viewpoints. State-of-the + +art geometry-aware approaches like Scene Representation Networks (SRN) [53] and Neural Radiance Fields (NeRF) [23,35,37,70] learn 3D representations from 2D images and camera parameters and perform differentiable neural rendering from one or a few input images at inference. Although they excel at interpolating near-input views, they are typically not well-suited for the large view changes required for exo-to-ego view synthesis. Trained purely with regression objectives, they have limited capacity to deal with sparse inputs, ambiguity caused by occluded parts of the scene, and long-range extrapolations caused by drastic camera transformations [9,64]—all challenges that are particularly prominent in exo-to-ego view translation. Alternatively, generative-based novel view synthesis methods [9,29,43,45,58,64,66] use generative priors to synthesize random plausible outputs from the conditional distribution, making them adept at handling ambiguity and extrapolating to occluded parts of the scene. Our Exo2Ego framework belongs to the generative-based vein, but unlike all the above, it does not depend on camera poses as critical conditional inputs. This improves scalability and widens the scope of potential applications, as camera parameters are seldom accessible in real-world scenarios and are challenging to infer from RGB images. + +# 2.3 Cross-view Image-to-Image Translation + +Exo-to-ego view synthesis is inherently an image-to-image translation problem [5, 10, 22, 61, 62]—but with drastic viewpoint change. Prior work centers around aerial-to-ground translation [40–42, 57] or generating an egocentric view without human-object interactions, i.e. a person walking indoors [30, 31, 57]. Except for P-GAN [30], all of these approaches require the ground truth semantic map of the target domain at training or inference time to ensure high-quality generation—an unrealistic assumption that means a good portion of the translation task (the boundaries of the objects in the unobserved ego view or the input exo view) is already known, Compared with the semantic-map-free approach P-GAN, our Exo2Ego exhibits significantly better qualitative and quantitative results. + +# 2.4 Diffusion Models + +DMs have shown competitive performance in image synthesis [11, 18, 19, 55], but their use for cross-view translation remains unexplored. We develop an exo-to-ego translation framework incorporating the diffusion model rather than GANs, considering potential advantages such as addressing mode collapse, promoting sample diversity, and maintaining stable training dynamics for improved image generation [44]. + +# 3 Methodology + +First, we formalize our problem (Sec. 3.1). Then we present the two stages of our model (Sec. 3.2 and 3.3), followed by an overview of training and inference (Sec. 3.4). + +![](images/57b400d9f41d668deba24255b68beceb7a6c8c8388d993c8662050fffc0ce089.jpg) +(a) High-level Structure Transformation + +![](images/810ff5de80a80597b291fd03348103d831ab2fc360e06e47a952ed5109a671e8.jpg) +(b) Diffusion-based Pixel Hallucination +Fig. 2: Our Exo2Ego framework comprises two modules: (a) High-level Structure Transformation, which predicts the ego layout, capturing hand position and interactions using a transformer-based encoder-decoder architecture. (b) Diffusion-based Pixel Hallucination, which enhances pixel-level details on top of the ego layout using a conditional diffusion model. + +# 3.1 Problem Formulation + +Whereas conventional geometry-aware approaches [10,35,70] generate novel views by specifying camera poses and performing volume rendering, Exo2Ego offers a different perspective. We propose tackling the exo-to-ego translation problem in a purely probabilistic framework. Intuitively, this is essential to address the inherent ambiguity in predicting the ego view from the exo view, e.g., due to entirely unseen portions of objects or the human body. + +We define $\mathbf{X}_T = \{x_1, \ldots, x_i, \ldots, x_T\}$ as a series of $T$ video frames captured from an exo viewpoint, characterized by static backgrounds with dynamic actors and other objects present, where $i$ denotes the time index. This exo perspective reveals the actors' motions and (potentially) full-body poses within the scene. Let $\mathbf{Y}_T = \{y_1, \ldots, y_i, \ldots, y_T\}$ denote the corresponding sequence of ego frames captured from a first-person perspective. This perspective emulates the view of a camera mounted on the head or body of the actor, with a focus on their actions and interactions. + +The goal of exo-to-ego view translation is to simulate the view of the ego camera-wearer in the scene recorded by an exo camera. Formally, we seek a translation model that can map $\mathbf{X}_T$ into a series of output ego frames, $\hat{\mathbf{Y}}_T = \{\hat{y}_1,\dots ,\hat{y}_i,\dots ,\hat{y}_T\}$ . As shown in Eqn. (1), the conditional distribution of $\hat{\mathbf{Y}}^T$ given $\mathbf{X}^T$ should be indistinguishable from the conditional distribution of $\mathbf{Y}^T$ + +given $\mathbf{X}^T$ + +$$ +p \left(\mathbf {\hat {Y}} _ {T} \mid \mathbf {X} _ {T}\right) = p \left(\mathbf {Y} _ {T} \mid \mathbf {X} _ {T}\right). \tag {1} +$$ + +We prioritize translating daily tabletop activities from the exo to ego view, which is a prevalent setting in egocentric learning and frequently entails extensive interactions between hands and objects. This problem is highly challenging as the translation model must produce photorealistic hand-object interaction sequences in the ego view while also performing geometric and semantic reasoning, i.e., correctly predicting the spatial location and pixel-level appearance of visual concepts. Furthermore, it requires linking the actor's head/body motion captured by the exo camera with the viewpoint change in the ego video. Upon examining the application of the widely used pixel-to-pixel generation method [62] for our task, we found it struggles with intricate details of the hands, likely because treats all pixels equally and lacks geometric correspondence between views. At the same time, we know that (at least in manipulation-rich scenarios) the hands of the actor are a common ground between the ego and exo views, despite their many other differences. + +Motivated by these points, we propose the Exo2Ego framework, which disentangles the understanding of cross-view correspondences and pixel-level synthesis. It consists of two key modules, as illustrated in Figure 2: (1) High-level Structure Transformation, which addresses the task of inferring the location and interaction manner of hands and objects in the ego view. To accomplish this, we train a transformer-based encoder-decoder model to translate the exo frame into an ego hand-object interaction layout (detailed below). (2) Diffusion-based Pixel Hallucination, which learns to synthesize realistic and high-quality pixel-level details by training a conditional diffusion model operating on top of the ego hand layout. + +# 3.2 High-level Structure Transformation + +Given an exo frame, the purpose of the high-level structure transformation is to train a layout translator which predicts the ego layout showing the location and rough contour of the hands. Specifically, to capture fine-grained hand-object interaction details, we propose to generate the hand layout, instantiated as 2D hand poses. The generated layout serves as a crucial reference for further pixel-level hallucination. To achieve this, we draw inspiration from a recent study [45] that highlights transformer-based architectures [12, 59] succeed in understanding cross-view correspondence, due to their reduced locality-bias. Considering transformers have the potential to represent the geometric information implicitly, we employ a pure transformer-based encoder-decoder architecture, which enables us to effectively incorporate and process the comprehensive context of the exo scene, and thus generate a precise ego hand layout. + +Exo Contextual Feature Extraction As shown in Figure 2, an exo frame $x_{i} \in \mathbb{R}^{H \times W \times 3}$ concatenated with the corresponding exo layout $x_{i}^{l} \in \mathbb{R}^{H \times W \times 3}$ (2D hand pose layout recorded as image size) is first split into a sequence of patches + +and then processed by an input embedding layer, such as patch embedding for ViTs [12] to get a tokenized input sequence $e_i \in \mathbb{R}^{M \times D}$ . To capture the positional information which is critical for spatial relationship reasoning, we add a learnable position embedding to the sequence of patches. Then the sequence of tokens is passed into a transformer encoder consisting of $N$ repeated blocks. Each block contains a multi-head attention token mixer to first communicate information among tokens, which is represented as: + +$$ +z _ {i} ^ {j - 1} = \operatorname {A t t e n t i o n} \left(\operatorname {N o r m} \left(e _ {i} ^ {j - 1}\right)\right) + e _ {i} ^ {j - 1}, \tag {2} +$$ + +where $j \in \{1, \dots, N\}$ and $\mathrm{Norm}(\cdot)$ is the normalization method. Then, the mixed token $z_i^{j-1}$ is passed into a two-layer MLP expressed as follows: + +$$ +e _ {i} ^ {j} = \operatorname {M L P} \left(\operatorname {N o r m} \left(z _ {i} ^ {j - 1}\right)\right) + z _ {i} ^ {j - 1}. \tag {3} +$$ + +After iterating through all $N$ blocks, the input sequence $e_i$ undergoes a mapping process that transforms it into the final contextual tokens $c_{i}\in \mathbb{R}^{M\times D}$ . We talk about how to decode the contextual tokens into the desired ego layout in the following part. + +Layout Decoder Consider a hand pose image with shape $\{H\times W\times 3\}$ as the target ego layout. The contextual tokens $c_{i}$ should be decoded into 2D hand joint coordinates and then rendered to the RGB image. First, the contextual tokens $c_{i}$ are combined with a sequence of learnable query embeddings $q_{i}\in \mathbb{R}^{E\times D}$ and passed into a decoder architecture, consisting of $K$ transformer blocks, as described earlier. Following the transformer blocks, a regression head is employed to estimate the 2D coordinates of the joints whose range is restricted to [0,1]. Then we apply the classic bipartite matching loss [8, 39] to map the regressed hand joints to the ground truth joints. + +# 3.3 Diffusion-based Pixel Hallucination + +Having inferred the ego hand pose layout $y_{i}^{l}$ , next the aim of pixel-level hallucination is to synthesize photorealistic ego frames by taking into account an exo frame $x_{i}$ and the target ego layout. We use the diffusion formulation proposed in the denoising diffusion probabilistic model (DDPM) [18] and train the diffusion model in the latent space. As shown in Figure 2, we first adopt a pre-trained variational autoencoder (VAE) model [25] as used in [44] to encode the original ego frames $y_{i}$ and conditional information exo frame $x_{i}$ and ego layout $y_{i}^{l}$ into the latent space. Then, we train a diffusion transformer [38] to learn the latent data distribution by denoising a latent vector sampled from a Gaussian distribution gradually. Specifically, given an initial noise map $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ and a conditioning vector $\mathbf{d}$ , the diffusion model generates the corresponding ego latent $\hat{\mathbf{z}}_{\theta}$ . A squared error loss is used to denoise a variably-noised latent code $\mathbf{z}^{n} = \alpha^{n}\mathbf{z} + \sigma^{n}\epsilon$ as follows: + +$$ +\mathcal {L} = \mathbb {E} _ {\mathbf {d}, \mathbf {z}, \boldsymbol {\epsilon}, n} \left[ \| \mathbf {z} - \hat {\mathbf {z}} _ {\theta} \left(\alpha^ {n} \mathbf {z} + \sigma^ {n} \boldsymbol {\epsilon}, \mathbf {d}\right) \| _ {2} ^ {2} \right] \tag {4} +$$ + +where $\mathbf{z}$ is the ground-truth latent vector, $\mathbf{d}$ is the conditioning vector, and $\alpha^n,\sigma^n$ are functions of the diffusion process time step $n$ , which control the noise schedule and sample quality. We simply concatenate the noisy latent vector $\mathbf{z}^n$ and the conditional embedding $\mathbf{d}$ as the input of the denoising transformer. In contrast to earlier cross-view translation methods based on GANs [30, 61, 62], our conditional diffusion model sequentially updates the target ego outputs, excels in capturing intricate ego-exo dependencies and faithfully reproducing the mapping of the conditional distributions in Eqn. (1). Notably, our model exhibits enhanced stability throughout the training and consistently produces higher-quality samples, as verified by our experiments. + +# 3.4 Training and Inference + +At the training stage, the two modules described in Section 3.2 and 3.3 are trained separately. During inference, we first generate the corresponding ego hand layout prediction $\hat{y}_i^l$ for each input exo frame $x_i$ with the layout translator described in Section 3.2. Subsequently, $\hat{y}_i^l$ is concatenated with $x_i$ in a channel-wise manner and then used as the conditional input of the pixel-level diffusion generator described in Section 3.3. + +# 4 Experimental Evaluation + +We introduce an exo-to-ego synthesis benchmark (Sec. 4.1) and then validate our approach against state-of-the-art synthesis methods (Sec. 4.2). + +# 4.1 A Benchmark for Exo-to-Ego View Translation + +To facilitate new work on the exo-to-ego translation task, we contribute a new cross-view synthesis benchmark sourced from four public time-synchronized multi-view datasets: H2O [27], Aria Pilot [32], Assembly101 [47], and Ego-Exo4D [16]. H2O [27] has indoor videos featuring actors manipulating 8 different objects using both hands. Assembly101 [47] shows people (dis)assembling 101 take-apart toy vehicles. We take 6 activity sequences involving different individuals manipulating a toy roller. Aria Pilot [32] provides 16 multi-view recordings of an actor wearing Project Aria glasses [54] manipulating YCB [7] objects on a desktop. For Ego-Exo4D, we include the four (largely non-table-top) scenarios involving heavy hand-object interactions from the EgoHandPose Benchmark - bike repair, cooking, music, and health—where the camera wearer also moves about the scene. For all datasets, we select the exo camera viewpoints that offer the clearest capture of the action. All frames are cropped and resized into $256 \times 256$ . We divide each video into 30-frame clips. Overall, this dataset showcases an array of human activities with various objects, from toy assembly to the manipulation of everyday objects. More details are provided in Appendix. + +We employ the following split settings to evaluate four kinds of generalization: (1) new actions, where we put the first $80\%$ of a video's clips into the training + +![](images/3d6c4652f9367f296636c1487338979ce1c07a6f60c9d025c3966767eb60bc7f.jpg) +Fig. 3: Qualitative examples when generalizing to new actions on all datasets. + +set and the remaining $20\%$ in the test set, hence splitting up action steps over time (e.g., for a video depicting picking up a coffee box and then taking coffee capsules out, the former will be assigned to the training set and the latter to the test set); (2) new objects, where we train with videos involving any of 7 objects and test with videos containing a novel 8th object; (3) new subjects, where we train with all clips from one subject and test on clips from another (subject $1 \rightarrow$ subject 2); (4) new scenes (backgrounds), where we train and test with disjoint scenes (scene h1, h2, k1, k2 $\rightarrow$ scene o1, o2). We analyze new action generalization on all 3 datasets, and the rest on H2O only, as it is the only dataset with the variations and meta-data to allow such split settings. For action generalization, H2O, Aria Pilot, Assembly101, and Ego-Exo4D have 704, 343, 682, and 543 training clips, and 199, 95, 205, and 211 test clips, respectively. For other generalization settings (new objects, subjects, and scenes), there are 691, 704, and 591 training clips, and 108, 194, and 312 test clips. + +# 4.2 Evaluating Synthesis Quality + +With this well-constructed benchmark in hand, we proceed to validate Exo2Ego. + +Table 1: Generalizing to new actions on three datasets. *pixelNeRF requires privileged camera information relative to the other baseline models. + +
H2O
MethodSSIM↑PSNR↑FID↓P_Squeeze↓P_Alex↓P_Vgg↓Feasi.↑
pix2pixHD [62]0.42830.370132.030.1630.2200.3370.8952
P-GAN [30]0.27229.079192.260.2420.3080.4490.8353
vid2vid [61]0.40229.88285.760.1660.2260.3420.9116
pixelNeRF* [70]0.21927.871470.130.7110.8070.7310.0086
DiT [38]0.43630.55345.230.1290.1780.2970.9604
Exo2Ego (Ours)0.44330.58037.030.1040.1460.2620.9764
Aria Pilot
MethodSSIM↑PSNR↑FID↓P_Squeeze↓P_Alex↓P_Vgg↓Feasi.↑
pix2pixHD [62]0.35029.553159.910.3040.3620.4800.0790
P-GAN [30]0.34329.769122.400.2870.3450.4750.1014
vid2vid [61]0.35929.85852.380.2660.3280.4430.2646
DiT [38]0.36829.94532.710.2400.2980.4170.4784
Exo2Ego (Ours)0.38030.04725.990.2250.2790.4020.5977
Assembly101
MethodSSIM↑PSNR↑FID↓P_Squeeze↓P_Alex↓P_Vgg↓Feasi.↑
pix2pixHD [62]0.40529.909112.710.2160.2950.4010.1184
P-GAN [30]0.37229.303114.680.2160.3000.4360.0800
vid2vid [61]0.36829.48178.470.2240.3120.4240.1358
DiT [38]0.39829.99626.330.1790.2550.3700.3092
Exo2Ego (Ours)0.41330.15325.230.1680.2380.3550.4019
+ +Baselines We consider several baselines: (1) Pix2PixHD [62], a single-view image translation method that processes the videos frame by frame; (2) P-GAN [30], a recent exo-to-ego view translation method that proposes a parallel generative network to facilitate cross-view image translation; (3) DiT [38], a pixel-to-pixel generative model from the latent diffusion model family. (4) Vid2Vid [61], a single-view video translation method that models the temporal dynamics in the video; (5) pixelNeRF [70], a NeRF-like model known for its remarkable generalization capabilities and its accommodating stance toward sparser input views. Note that pixelNeRF assumes camera parameters are known, while other methods do not. We conduct experiments with pixelNeRF only on H2O, as the other two datasets lack the camera information needed by pixelNeRF. + +Implementation Details For Exo2Ego's high-level structure transformation transformer, the number of blocks $N$ is set to 6. We use DiT-XL/2 [38] as the denoising architecture for Exo2Ego. The diffusion model is trained for 40,000 steps and pixelNeRF is trained for 70,000 steps for all datasets. For H2O and Assembly101, pix2pixHD, and P-GAN are trained for 100 epochs, and vid2vid is trained for 40 epochs. For Aria Pilot, we use 400 and 100 epochs respectively. + +Evaluation Metrics We measure the quality of the synthesized videos as follows. First, we evaluate the feasibility of generated hands, denoted as Feasi. Inspired by [69], we adopt an off-the-shelf egocentric hand object detector [49] to measure the physical feasibility of the synthesized hands. This metric is calcu + +![](images/30ef4e009d01ea1429b8a18ca20c08788b2cb3baa1dc401b63e979d95ef1daef.jpg) +Fig. 4: Qualitative examples when generalizing to new objects, subjects, and scenes (backgrounds) on the H2O dataset. + +lated as the confidence score of hand detection results produced by the detector, averaged over all synthesized frames. + +We also use structural similarity (SSIM) and peak signal-to-noise ratio (PSNR) [20] to check the pixel-level similarity between the predicted ego frame and the ground truth (GT) ego frame. These two evaluation metrics should ideally remain stable under minor transformations (such as small translations or affine transformations). We further adopt perceptual metrics, including FID [46] and Learned Perceptual Image Patch Similarity (LPIPS) [74] which evaluates the feature-level similarity of the predicted ego frame and the true ego frame. LPIPS are denoted as $P_{\text{Squeeze}}$ , $P_{\text{Alex}}$ , and $P_{\text{Vgg}}$ , which uses SqueezeNet [21], AlexNet [26], and VGG [52] as the feature extractor. + +Generalization to New Actions Figure 3 showcases qualitative results on H2O, Aria Pilot, and Assembly101, demonstrating the superiority of our Exo2Ego framework. Compared to all baselines, our approach produces realistic hands with correct poses, especially noticeable in the highlighted yellow circle regions when dealing with new actions during test time. Table 1 presents the feasibility of synthesized hands and the various measures of pixel-level similarity and perceptual similarity between the predicted ego frame and the ground truth ego frame. Our Exo2Ego greatly increases the feasibility of synthesized hands, with a confidence gain of up to $11.93\%$ on Aria Pilot, on the basis of the second-highest score of $47.84\%$ achieved by DiT. This improvement indicates enhanced realism in the generated hands, resulting in greater confidence in hand detection results. Additionally, our Exo2Ego can produce ego frames with higher pixel-level similarity to the ground truth frames and better perceptual metrics (much lower FID scores). This highlights the benefits of the 'decoupling' idea of our Exo2Ego framework in generating photorealistic ego frames. + +Looking at Table 1, there is a trend among our Exo2Ego and the baselines — doing better on H2O dataset. We conjecture that the superiority comes from H2O's higher hand quality, including clear pose visibility and larger spatial coverage within the field of view, in both exo and ego view. + +Table 2: Generalizing to different objects, subjects, and scenes. *pixelNeRF requires privileged camera information relative to the other baseline models. + +
H2O-unseen object
MethodSSIM↑PSNR↑FID↓P_Squeeze↓P_Alex↓P_Vgg↓Feasi.↑
pix2pixHD [62]0.33329.504253.730.2690.3600.4670.7527
P-GAN [30]0.27028.915264.760.2770.3500.4920.7720
vid2vid [61]0.34929.585145.390.2420.3240.4280.8612
pixelNeRF* [70]0.26227.912473.730.7500.8410.7360.0000
DiT [38]0.33829.740105.180.2140.2880.4100.9689
Exo2Ego (Ours)0.34429.786101.450.2080.2810.4030.9786
H2O-unseen subject
MethodSSIM↑PSNR↑FID↓P_Squeeze↓P_Alex↓P_Vgg↓Feasi.↑
pix2pixHD [62]0.06928.028477.470.6610.7540.7650.0140
P-GAN [30]0.08828.063451.370.6230.7080.7250.2002
vid2vid [61]0.13628.094308.520.5640.6900.6860.3838
pixelNeRF* [70]0.17027.874459.490.7690.8360.7680.0000
DiT [38]0.17228.185183.060.5240.6300.6820.8405
Exo2Ego (Ours)0.17828.266150.870.5030.6290.6650.9666
H2O-unseen scene
MethodSSIM↑PSNR↑FID↓P_Squeeze↓P_Alex↓P_Vgg↓Feasi.↑
pix2pixHD [62]0.02427.918558.870.7290.7890.7880.0013
P-GAN [30]0.08527.928325.340.6080.6990.7240.6041
vid2vid [61]0.03327.942398.780.6360.7310.7340.2095
pixelNeRF* [70]0.10327.810506.100.7730.9080.8000.0000
DiT [38]0.14027.874308.980.5840.7290.6780.8317
Exo2Ego (Ours)0.15727.943249.040.5310.7090.6580.9635
+ +We further explored how Exo2Ego fares in the settings with heavy hand-object manipulation and frequent head motion, but minimal movement across the broader scene on Ego-Exo4D [16]. As shown in Figure 6, we compare our Exo2Ego with a strong image-to-image translation model ControlNet [73], which is built on a stable diffusion trained on large scales of data. It is observed that our Exo2Ego synthesizes feasible hands with correct poses, unlike ControlNet. The exo image's hand regions in Ego-Exo4D are small, so we crop it to use as additional input to offer a better view of the hands for both methods. + +Generalization to New Objects, Subjects, and Backgrounds Illustrated in Table 2, our Exo2Ego outperforms the baselines in terms of quantitative metrics for generalization to new objects/subjects/backgrounds. Figure 4 shows that our Exo2Ego can produce realistic hand-object interactions when encountering exo views with new distributions, while other baselines fail miserably. This is mainly because our diffusion-based generative mechanism gradually denoises the target ego view, capturing the inherent intricate mapping from exo to ego views better. Another observation is that pixelNeRF produces ego views characterized by noticeable blurring and a lack of fine details in hands and objects. This observation supports our earlier discussions on the limitations of 3D-aware approaches. + +![](images/42f834b8cb09ccf73596a17383d0340687adc4c67584b1d5cc01ad0c4ebb423d.jpg) +Fig. 5: Exo2Ego framework generates ego videos with reasonable viewpoint changes. + +![](images/35de2f6f4c3b2e12a6f8f8d6a551b86955c7675b22d9e12c9a7e1dba93cb56ea.jpg) +Fig. 6: Qualitative examples when generalizing to new actions on Ego-Exo4D dataset. + +Modeling Egocentric Viewpoint Change Due to the probabilistic nature of our Exo2Ego framework, it excels in capturing the relationship between the viewpoint changes in the ego view and the head motions observed in the exo view, holding the potential to deliver more immersive virtual experiences. Intrinsically, our Exo2Ego framework leverages the fact that ego-camera pose is conditioned on the reference exo view, as the ego-camera wearer is usually visible in the exo view. This is highlighted by the yellow gaze direction and dashed box in Figure 5. As the actor moves their head when interacting with objects, the slope of the desk edge in the ego view also changes, indicating a corresponding change in ego viewpoint. This implies the ego-camera motion can be implicitly inferred via conditional generative modeling, even if it's not enforced explicitly. The ability to model ego viewpoint changes is crucial, especially in physical activities like playing ball games which involve heavy head and body motions. + +# 5 Limitations and Future Work + +Temporal Coherence Our Exo2Ego operates on each frame individually — the generation of the current ego frame is solely conditioned on the current exo input frame. There is no explicit consideration of previously generated ego frames or past exo frames. Temporal coherence is orthogonal to our current contribution and can be used to improve any of the baseline models without such design. + +![](images/2bee228fb7920b58dd07171f1d09bba823ab610d4d56ef0c6e7d90279e64b4be.jpg) +Fig. 7: Failure cases: Exo2Ego generates hands with reasonably accurate positions/poses but incomplete/wrong objects in the ego view. + +Generalization Ability As an early-stage effort for exo-to-ego cross-view translation, our work makes non-trivial progress in producing reasonable hand manipulation details for novel actions. However, despite establishing stronger generalizations than the existing models, our model does not generalize perfectly to in-the-wild objects, subjects, and backgrounds, due to the limited scale of available training data. Considering the intricate nature of the generalization to new environments, we regard it as a challenging future research direction. Nonetheless, our focus on generalizing to new actions within the same environment is meaningful for several applications. Some examples: Real-time monitoring during physical rehabilitation (exo-to-ego view translation can help therapists guide patients during rehabilitation sessions), assistive cooking aid for visually impaired individuals (ego views are provided to the visually impaired individuals to cook independently), or real-time fitness training monitoring (coaches use synthesized ego views to help the trainee correct the fitness movements) — in each case, the environment is stable between training and testing, and yet the model must generalize to new actions for cross-view translation. + +3D prior Our qualitative results expose limitations in existing baselines and our Exo2Ego framework in generating 3D-consistent views for objects during test time, attributed to the absence of geometric priors for common objects. See Figure 7. Future work involves integrating robust object geometric priors. While leveraging predicted camera parameters for encoding 3D geometry is a promising approach, as shown Figure 4, the popular 3D-aware method pixelNeRF [70] faces challenges in the dynamic exo-ego setting characterized by hand and object interactions, due to limitations in handling sparse inputs, occluded scenes, and long-range camera extrapolations. Notably, our Exo2Ego outperforms pixelNeRF, highlighting the potential for future work to explore explicit geometric reasoning within our framework. + +# 6 Conclusion + +Overall, our work contributes to the growing body of research in cross-view translation and lays the groundwork for the important case of exo-to-ego. Our generative framework Exo2Ego integrates high-level structure transformation and pixel-level hallucination to yield very encouraging experimental results. We also provide a curated benchmark task for supporting continued work on this problem. The core technology holds immense potential for applications in robot learning and AR skill coaching, where an ego actor needs to replicate the actions of a demonstrator observed from the exo perspective. + +Acknowledgment: UT Austin is supported in part by NSF Grants AF 1901292, CNS 2148141, Tripods CCF 1934932, IFML CCF 2019844 and research gifts by Western Digital, Amazon, WNCG IAP, UT Austin Machine Learning Lab (MLL), Cisco, and the Stanly P. Finch Centennial Professorship in Engineering. K.G. is paid as a research scientist at Meta. + +# References + +1. Ardeshir, S., Borji, A.: Ego2top: Matching viewers in egocentric and top-view videos. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14. pp. 253-268. Springer (2016) 3 +2. Ardeshir, S., Borji, A.: Egocentric meets top-view. IEEE transactions on pattern analysis and machine intelligence 41(6), 1353-1366 (2018) 3 +3. Ardeshir, S., Borji, A.: An exocentric look at egocentric actions and vice versa. Computer Vision and Image Understanding 171, 61-68 (2018) 1, 3 +4. Bahl, S., Gupta, A., Pathak, D.: Human-to-robot imitation in the wild. arXiv preprint arXiv:2207.09450 (2022) 2 +5. Bansal, A., Ma, S., Ramanan, D., Sheikh, Y.: Recycle-gan: Unsupervised video retargeting. In: Proceedings of the European conference on computer vision (ECCV). pp. 119-135 (2018) 4 +6. Bharadhwaj, H., Gupta, A., Tulsiani, S., Kumar, V.: Zero-shot robot manipulation from passive human videos. arXiv preprint arXiv:2302.02011 (2023) 2 +7. Calli, B., Singh, A., Walsman, A., Srinivasa, S., Abbeel, P., Dollar, A.M.: The ycb object and model set: Towards common benchmarks for manipulation research. In: 2015 international conference on advanced robotics (ICAR). pp. 510-517. IEEE (2015) 8 +8. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 213-229. Springer (2020) 7 +9. Chan, E.R., Nagano, K., Chan, M.A., Bergman, A.W., Park, J.J., Levy, A., Aittala, M., De Mello, S., Karras, T., Wetzstein, G.: Generative novel view synthesis with 3d-aware diffusion models. arXiv preprint arXiv:2304.02602 (2023) 2, 4 +0. Cheng, F., Luo, M., Wang, H., Dimakis, A., Torresani, L., Bertasius, G., Grauman, K.: 4diff: 3d-aware diffusion model for third-to-first viewpoint translation. In: ECCV (2024) 3, 4, 5 +1. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems 34, 8780-8794 (2021) 4 +12. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 6, 7 +13. Elfeki, M., Regmi, K., Ardeshir, S., Borji, A.: From third person to first person: Dataset and baselines for synthesis and retrieval. arXiv preprint arXiv:1812.00104 (2018) 3 +14. Fan, C., Lee, J., Xu, M., Kumar Singh, K., Jae Lee, Y., Crandall, D.J., Ryoo, M.S.: Identifying first-person camera wearers in third-person videos. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5125-5133 (2017) 3 + +15. Grauman, K., Westbury, A., Torresani, L., Kitani, K., Malik, J., Afouras, T., Ashutosh, K., Baiyya, V., Bansal, S., Boote, B., et al.: Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. arXiv preprint arXiv:2311.18259 (2023) 3 +16. Grauman, K., Westbury, A., Torresani, L., Kitani, K., Malik, J., Afouras, T., Ashutosh, K., Baiyya, V., Bansal, S., Boote, B., et al.: Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19383-19400 (2024) 3, 8, 12 +17. Ho, H.I., Chiu, W.C., Wang, Y.C.F.: Summarizing first-person videos from third persons' points of view. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 70-85 (2018) 3 +18. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33, 6840-6851 (2020) 4, 7 +19. Ho, J., Sahara, C., Chan, W., Fleet, D.J., Norouzi, M., Salimans, T.: Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research 23(1), 2249-2281 (2022) 4 +20. Hore, A., Ziou, D.: Image quality metrics: Psnr vs. sSIM. In: 2010 20th international conference on pattern recognition. pp. 2366-2369. IEEE (2010) 11 +21. Iandola, F.N., Han, S., Moskewicz, M.W., Ashraf, K., Dally, W.J., Keutzer, K.: SqueezeNET: Alexnet-level accuracy with 50x fewer parameters and $< 0.5$ mb model size. arXiv preprint arXiv:1602.07360 (2016) 11 +22. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1125-1134 (2017) 4 +23. Jang, W., Agapito, L.: Codenerf: Disentangled neural radiance fields for object categories. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12949-12958 (2021) 2, 4 +24. Jangir, R., Hansen, N., Ghosal, S., Jain, M., Wang, X.: Look closer: Bridging egocentric and third-person views with transformers for robotic manipulation. IEEE Robotics and Automation Letters 7(2), 3046-3053 (2022) 3 +25. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013) 7 +26. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. Communications of the ACM 60(6), 84-90 (2017) 11 +27. Kwon, T., Tekin, B., Stühmer, J., Bogo, F., Pollefeys, M.: H2o: Two hands manipulating objects for first person interaction recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10138-10148 (October 2021) 3, 8 +28. Li, Y., Nagarajan, T., Xiong, B., Grauman, K.: Ego-exo: Transferring visual representations from third-person to first-person videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6943-6953 (2021) 3 +29. Liu, A., Tucker, R., Jampani, V., Makadia, A., Snavely, N., Kanazawa, A.: Infinite nature: Perpetual view generation of natural scenes from a single image. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14458-14467 (2021) 2, 4 +30. Liu, G., Tang, H., Latapie, H., Yan, Y.: Exocentric to egocentric image generation via parallel generative adversarial network. In: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1843-1847. IEEE (2020) 3, 4, 8, 10, 12 + +31. Liu, G., Tang, H., Latapie, H.M., Corso, J.J., Yan, Y.: Cross-view exocentric to egocentric video synthesis. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 974-982 (2021) 3, 4 +32. Lv, Z., Miller, E., Meissner, J., Pesqueira, L., Sweeney, C., Dong, J., Ma, L., Patel, P., Moulon, P., Somasundaram, K., Parkhi, O., Zou, Y., Raina, N., Saarinen, S., Mansour, Y.M., Huang, P.K., Wang, Z., Troynikov, A., Artal, R.M., DeTone, D., Barnes, D., Argall, E., Lobanovskiy, A., Kim, D.J., Bouttefroy, P., Straub, J., Engel, J.J., Gupta, P., Yan, M., Nardi, R.D., Newcombe, R.: Aria pilot dataset. https://about.facebook.com/realitylabs/projectaria/datasets (2022) 3, 8 +33. Majumdar, A., Yadav, K., Arnaud, S., Ma, Y.J., Chen, C., Silwal, S., Jain, A., Berges, V.P., Abbeel, P., Malik, J., et al.: Where are we in the search for an artificial visual cortex for embodied intelligence? arXiv preprint arXiv:2303.18240 (2023) 2 +34. Mandikal, P., Grauman, K.: Dexvip: Learning dexterous grasping with human hand pose priors from video. In: Conference on Robot Learning (2021) 2 +35. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) 2, 4, 5 +36. Nair, S., Rajeswaran, A., Kumar, V., Finn, C., Gupta, A.: R3m: A universal visual representation for robot manipulation. arXiv preprint arXiv:2203.12601 (2022) 2 +37. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5480-5490 (2022) 2, 4 +38. Peebles, W., Xie, S.: Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748 (2022) 3, 7, 10, 12 +39. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified, real-time object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 779-788 (2016) 7 +40. Regmi, K., Borji, A.: Cross-view image synthesis using conditional gans. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018) 4 +41. Regmi, K., Borji, A.: Cross-view image synthesis using geometry-guided conditional gans. Computer Vision and Image Understanding (2019). https://doi.org/https://doi.org/10.1016/j.cviu.2019.07.008, http://www.sciencedirect.com/science/article/pii/S1077314219301043 4 +42. Ren, B., Tang, H., Sebe, N.: Cascaded cross mlp-mixer gans for cross-view image translation. arXiv preprint arXiv:2110.10183 (2021) 2, 4 +43. Ren, X., Wang, X.: Look outside the room: Synthesizing a consistent long-term 3d scene video from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3563-3573 (2022) 2, 4 +44. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022) 4, 7 +45. Rombach, R., Esser, P., Ommer, B.: Geometry-free view synthesis: Transformers and no 3d priors. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14356-14366 (2021) 2, 4, 6 +46. Seitzer, M.: pytorch-fid: FID Score for PyTorch. https://github.com/mseitzer/pytorch-fid (August 2020), version 0.3.0 11 + +47. Sener, F., Chatterjee, D., Shelepov, D., He, K., Singhania, D., Wang, R., Yao, A.: Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21096-21106 (2022) 3, 8 +48. Sermanet, P., Lynch, C., Chebotar, Y., Hsu, J., Jang, E., Schaal, S., Levine, S.: Time-contrastive networks: Self-supervised learning from video. Proceedings of International Conference in Robotics and Automation (ICRA) (2018), http://arxiv.org/abs/1704.06888 3 +49. Shan, D., Geng, J., Shu, M., Fouhey, D.F.: Understanding human hands in contact at internet scale. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020) 10 +50. Sigurdsson, G.A., Gupta, A., Schmid, C., Farhadi, A., Alahari, K.: Actor and observer: Joint modeling of first and third-person videos. In: proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7396-7404 (2018) 3 +51. Sigurdsson, G.A., Gupta, A., Schmid, C., Farhadi, A., Alahari, K.: Charades-ego: A large-scale dataset of paired third and first person videos. arXiv preprint arXiv:1804.09626 (2018) 3 +52. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014) 11 +53. Sitzmann, V., Zollhöfer, M., Wetzstein, G.: Scene representation networks: Continuous 3d-structure-aware neural scene representations. Advances in Neural Information Processing Systems 32 (2019) 4 +54. Somasundaram, K., Dong, J., Tang, H., Straub, J., Yan, M., Goesele, M., Engel, J.J., De Nardi, R., Newcombe, R.: Project aria: A new tool for egocentric multimodal ai research. arXiv preprint arXiv:2308.13561 (2023) 8 +55. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=PxTIG12RRHS 4 +56. Soran, B., Farhadi, A., Shapiro, L.: Action recognition in the presence of one egocentric and multiple static cameras. In: Computer Vision-ACCV 2014: 12th Asian Conference on Computer Vision, Singapore, Singapore, November 1-5, 2014, Revised Selected Papers, Part V 12. pp. 178-193. Springer (2015) 3 +57. Tang, H., Xu, D., Sebe, N., Wang, Y., Corso, J.J., Yan, Y.: Multi-channel attention selection gan with cascaded semantic guidance for cross-view image translation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2417-2426 (2019) 2, 4 +58. Tseng, H.Y., Li, Q., Kim, C., Alsisan, S., Huang, J.B., Kopf, J.: Consistent view synthesis with pose-guided diffusion models. arXiv preprint arXiv:2303.17598 (2023) 2, 4 +59. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) 6 +60. Wang, J., Liu, L., Xu, W., Sarkar, K., Luvizon, D., Theobalt, C.: Estimating egocentric 3d human pose in the wild with external weak supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13157-13166 (2022) 3 +61. Wang, T.C., Liu, M.Y., Zhu, J.Y., Liu, G., Tao, A., Kautz, J., Catanzaro, B.: Video-to-video synthesis. arXiv preprint arXiv:1808.06601 (2018) 3, 4, 8, 10, 12 + +62. Wang, T.C., Liu, M.Y., Zhu, J.Y., Tao, A., Kautz, J., Catanzaro, B.: High-resolution image synthesis and semantic manipulation with conditional gans. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8798-8807 (2018) 3, 4, 6, 8, 10, 12 +63. Wang, Z., Wu, S., Xie, W., Chen, M., Prisacariu, V.A.: NeRF--: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064 (2021) 2 +64. Watson, D., Chan, W., Martin-Brualla, R., Ho, J., Tagliasacchi, A., Norouzi, M.: Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628 (2022) 2, 4 +65. Wen, Y., Singh, K.K., Anderson, M., Jan, W.P., Lee, Y.J.: Seeing the unseen: Predicting the first-person camera wearer's location and pose in third-person scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. pp. 3446-3455 (October 2021) 3 +66. Wiles, O., Gkioxari, G., Szeliski, R., Johnson, J.: Synsin: End-to-end view synthesis from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7467-7477 (2020) 2, 4 +67. Xu, M., Fan, C., Wang, Y., Ryoo, M.S., Crandall, D.J.: Joint person segmentation and identification in synchronized first-and third-person videos. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 637-652 (2018) 3 +68. Xue, Z., Grauman, K.: Learning fine-grained view-invariant representations from unpaired ego-exo videos via temporal alignment. In: NeurIPS (2023) 3 +69. Ye, Y., Li, X., Gupta, A., Mello, S.D., Birchfield, S., Song, J., Tulsiani, S., Liu, S.: Affordance diffusion: Synthesizing hand-object interactions. In: CVPR (2023) 10 +70. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4578-4587 (June 2021) 2, 3, 4, 5, 10, 12, 14 +71. Yu, H., Cai, M., Liu, Y., Lu, F.: What i see is what you see: Joint attention learning for first and third person video co-analysis. In: Proceedings of the 27th ACM International Conference on Multimedia. pp. 1358-1366 (2019) 3 +72. Yu, H., Cai, M., Liu, Y., Lu, F.: First-and third-person video co-analysis by learning spatial-temporal joint attention. IEEE Transactions on Pattern Analysis and Machine Intelligence (2020) 3 +73. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models (2023) 12 +74. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018) 11 \ No newline at end of file diff --git a/2024/Put Myself in Your Shoes_ Lifting the Egocentric Perspective from Exocentric Videos/images.zip b/2024/Put Myself in Your Shoes_ Lifting the Egocentric Perspective from Exocentric Videos/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e6e1224f76de10c443a0e88c96ef912d01c6f01c --- /dev/null +++ b/2024/Put Myself in Your Shoes_ Lifting the Egocentric Perspective from Exocentric Videos/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1fb73e7602adf0cc3f42ed104e2411b7e19df70754921bd85afe8484f99626c3 +size 837334 diff --git a/2024/Put Myself in Your Shoes_ Lifting the Egocentric Perspective from Exocentric Videos/layout.json b/2024/Put Myself in Your Shoes_ Lifting the Egocentric Perspective from Exocentric Videos/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..888cf3f0e35b72b4e06ab29788ae41bfe4b5b7f5 --- /dev/null +++ b/2024/Put Myself in Your Shoes_ Lifting the Egocentric Perspective from Exocentric Videos/layout.json @@ -0,0 +1,8464 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 135, + 111, + 479, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 111, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 135, + 111, + 479, + 148 + ], + "type": "text", + "content": "Put Myself in Your Shoes: Lifting the Egocentric Perspective from Exocentric Videos" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 167, + 448, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 167, + 448, + 180 + ], + "spans": [ + { + "bbox": [ + 165, + 167, + 448, + 180 + ], + "type": "text", + "content": "Mi Luo" + }, + { + "bbox": [ + 165, + 167, + 448, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 165, + 167, + 448, + 180 + ], + "type": "text", + "content": ", Zihui Xue" + }, + { + "bbox": [ + 165, + 167, + 448, + 180 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 165, + 167, + 448, + 180 + ], + "type": "text", + "content": ", Alex Dimakis" + }, + { + "bbox": [ + 165, + 167, + 448, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 165, + 167, + 448, + 180 + ], + "type": "text", + "content": ", and Kristen Grauman" + }, + { + "bbox": [ + 165, + 167, + 448, + 180 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 231, + 189, + 382, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 189, + 382, + 212 + ], + "spans": [ + { + "bbox": [ + 231, + 189, + 382, + 212 + ], + "type": "text", + "content": "1 The University of Texas at Austin \n2 FAIR, Meta AI" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 141, + 229, + 479, + 299 + ], + "blocks": [ + { + "bbox": [ + 141, + 229, + 479, + 299 + ], + "lines": [ + { + "bbox": [ + 141, + 229, + 479, + 299 + ], + "spans": [ + { + "bbox": [ + 141, + 229, + 479, + 299 + ], + "type": "image", + "image_path": "a6dcc9c18039fafa2af0d5b6d13f26b60c4b5268fa152bec010bcc160d59e948.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 307, + 482, + 341 + ], + "lines": [ + { + "bbox": [ + 131, + 307, + 482, + 341 + ], + "spans": [ + { + "bbox": [ + 131, + 307, + 482, + 341 + ], + "type": "text", + "content": "Fig. 1: We consider cross-view translation from exocentric to egocentric video. Given an exocentric video sequence with hand-object interactions (left), the goal is to generate the corresponding egocentric perspective (right)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 366, + 455, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 366, + 455, + 553 + ], + "spans": [ + { + "bbox": [ + 159, + 366, + 455, + 553 + ], + "type": "text", + "content": "Abstract. We investigate exocentric-to-egocentric cross-view translation, which aims to generate a first-person (egocentric) view of an actor based on a video recording that captures the actor from a third-person (exocentric) perspective. To this end, we propose a generative framework called Exo2Ego that decouples the translation process into two stages: high-level structure transformation, which explicitly encourages cross-view correspondence between exocentric and egocentric views, and a diffusion-based pixel-level hallucination, which incorporates a hand layout prior to enhance the fidelity of the generated egocentric view. To pave the way for future advancements in this field, we curate a comprehensive exo-to-ego cross-view translation benchmark focused on hand-object manipulations. It consists of a diverse collection of synchronized ego-exo video pairs from four public datasets: H2O, Aria Pilot, Assembly101, and Ego-Exo4D. The experimental results validate that Exo2Ego delivers photorealistic video results with clear hand manipulation details and outperforms several baselines in terms of both synthesis quality and generalization to new actions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 563, + 432, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 563, + 432, + 574 + ], + "spans": [ + { + "bbox": [ + 160, + 563, + 432, + 574 + ], + "type": "text", + "content": "Keywords: Egocentric Hand-Object Interaction " + }, + { + "bbox": [ + 160, + 563, + 432, + 574 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 563, + 432, + 574 + ], + "type": "text", + "content": " View Synthesis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 595, + 230, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 595, + 230, + 608 + ], + "spans": [ + { + "bbox": [ + 132, + 595, + 230, + 608 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 620, + 482, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 620, + 482, + 657 + ], + "spans": [ + { + "bbox": [ + 130, + 620, + 482, + 657 + ], + "type": "text", + "content": "Given a third-person video capturing a person opening a milk carton, what would the visual world look like from his perspective? See Figure 1. Due to mirror neurons in the human brain [3], we can easily envision the appearance" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "and spatial relationships of the person's hands and the milk carton from a first-person perspective. However, existing computer vision models struggle to do the same thing, owing to the stark distinctions between the two viewpoints." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 481, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 481, + 224 + ], + "type": "text", + "content": "We make a step towards addressing this underexplored problem: exocentric-to-egocentric cross-view translation. The goal is to synthesize the corresponding ego view of an actor from an exo video recording3, with minimal assumptions on the viewpoint relationships (e.g., camera parameters or accurate geometric scene structure). Specifically, we focus on synthesizing ego hand-object interactions in table-top activities, such as assembling toys, pouring milk, or picking up a box." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 224, + 481, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 481, + 331 + ], + "type": "text", + "content": "This task can benefit many applications, ranging from robotics to virtual and augmented reality (VR/AR). For example, an AR assistant could transform the view shown in a third-person how-to video to demonstrate to a user how things should look from their own perspectives, e.g., revealing finger placements and strumming techniques for a guitar video. Similarly, in robot learning, the robot could better match its actions to those of a human instructor in the room by projecting his/her hand-object interactions to its own perspective. Indeed, recent work demonstrates that the human ego view is valuable for robot learning [4,6, 33,34,36], as robots also perceive the world through the egocentric lens." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 331, + 481, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 481, + 450 + ], + "type": "text", + "content": "However, exo-to-ego view translation is highly challenging. It requires understanding the spatial relationships of visible hands and objects and inferring their pixel-level appearance in the novel ego view. Further, the task is not strictly geometric; it is inherently underdetermined. Some parts of an object may not be visible in the exo view—such as the inner pages of a book when only the cover is observed in the exo view—requiring the model to extrapolate the occluded parts. As a result, the recent popular geometry-aware novel view synthesis approaches [23, 35, 37, 70] are ill-equipped to solve this problem. The key reason is that they are regressive rather than generative, which limits their ability to deal with sparse input views (single exo view in our case) and major occlusions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 451, + 481, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 481, + 594 + ], + "type": "text", + "content": "In light of this, we propose a probabilistic exo-to-ego generative framework (termed Exo2Ego) that learns the conditional distribution of the target ego video given the exo video. Exo2Ego differs from generative-based methods for general novel view synthesis [9, 29, 43, 45, 58, 64, 66] as it relaxes the requirement for camera parameters as input. This expands the potential application scenarios since camera parameters are rarely accessible in real-world scenarios, e.g., taking photos or videos with a mobile phone. Additionally, accurately inferring camera parameters remains challenging without precise correspondences or a rigid lab setting for calibration [63]. Moreover, Exo2Ego stands apart from recent cross-view image-to-image translation methods [42, 57], as it does not require the ground truth semantic map (object segmentation) of the target domain at the inference stage—typically an infeasible assumption." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 481, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 631 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 631 + ], + "type": "text", + "content": "Our key insights are to explicitly encourage cross-view correspondence by predicting the ego layout and to introduce a hand layout prior to boost the fidelity of generated hands in the ego view. Specifically, our Exo2Ego framework" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "M. Luo et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 131, + 642, + 481, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 642, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 642, + 481, + 666 + ], + "type": "text", + "content": "3We use \"ego\" and \"exo\" as shorthand for egocentric (first-person) and exocentric (third-person)." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "decouples the exo-to-ego view translation problem into two stages: (1) High-level Structure Transformation, which infers the rough location and interaction manner of the hands and objects in the ego view, using a transformer-based model to transform the exo hand-object interaction layout into its ego counterpart; and (2) Diffusion-based Pixel Hallucination, which trains a conditional diffusion model to refine the details on top of the ego hand-object interaction layout. Overall, Exo2Ego is a generative framework that offers a simple but effective baseline approach for the exo-to-ego view translation problem, accounting for the centrality of hand-object manipulations in this domain." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 482, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 356 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 356 + ], + "type": "text", + "content": "To evaluate our approach and boost research in this field, we construct an exoto-ego cross-view synthesis benchmark, comprising synchronized ego-exo video pairs curated from four datasets: H2O [27], Aria Pilot [32], Assembly101 [47], and Ego-Exo4D [16]. These datasets offer a breadth of table-top activities, as well as some non-table-top activities where the camera wearer moves around the scene. Empirical results underscore the efficacy of our Exo2Ego framework, which produces realistic video outputs with distinct hand manipulation details. Exo2Ego surpasses single-view translation baselines based on GANs [61,62] and latent diffusion models [38], a recent cross-view synthesis approach [30], as well as a NeRF-based method [70], demonstrating superior generation quality and a marked improvement in generalization capability to new actions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 370, + 234, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 370, + 234, + 383 + ], + "spans": [ + { + "bbox": [ + 131, + 370, + 234, + 383 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 392, + 373, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 392, + 373, + 403 + ], + "spans": [ + { + "bbox": [ + 131, + 392, + 373, + 403 + ], + "type": "text", + "content": "2.1 Relating Egocentric and Exocentric Views" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 407, + 482, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 612 + ], + "type": "text", + "content": "There have been several attempts to jointly understand ego and exo views [13, 15, 51]. Early efforts explore how to localize the person wearing a camera in a third-person view, given their egocentric video [1,2,14,65,67]. To bridge the ego-exo gap, other work learns view-invariant features from concurrent (paired) [3, 48,50,71,72] or unpaired [68] views, or boosts the latent ego signals in exo video during pretraining [28], summarization [17], and 3D pose estimation [60]. Additionally, fusing ego and exo views can improve action recognition [56] and robotic manipulation tasks [24]. Despite its significance, cross-view translation has received little attention. P-GAN [30] and STA-GAN [31] have explored cross-view image and video synthesis, respectively, yet they only examine basic activities such as walking, jogging, and running, limiting their applicability to more diverse scenarios. To address these limitations, we introduce a new benchmark for exoto-ego view synthesis covering some diverse activities, from assembling toys to manipulating everyday objects. Importantly, we relax STA-GAN's requirement for the ego semantic map during test time [31], enhancing the practical applicability. A concurrent work [10] examines the introduction of geometry priors to the diffusion model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 624, + 270, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 270, + 637 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 270, + 637 + ], + "type": "text", + "content": "2.2 Novel View Synthesis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Exo-to-ego view translation also relates to novel view synthesis, i.e., given a set of images of a scene, infer how the scene looks from novel viewpoints. State-of-the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "text", + "content": "Lifting the Egocentric Perspective from Exocentric Videos" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": "art geometry-aware approaches like Scene Representation Networks (SRN) [53] and Neural Radiance Fields (NeRF) [23,35,37,70] learn 3D representations from 2D images and camera parameters and perform differentiable neural rendering from one or a few input images at inference. Although they excel at interpolating near-input views, they are typically not well-suited for the large view changes required for exo-to-ego view synthesis. Trained purely with regression objectives, they have limited capacity to deal with sparse inputs, ambiguity caused by occluded parts of the scene, and long-range extrapolations caused by drastic camera transformations [9,64]—all challenges that are particularly prominent in exo-to-ego view translation. Alternatively, generative-based novel view synthesis methods [9,29,43,45,58,64,66] use generative priors to synthesize random plausible outputs from the conditional distribution, making them adept at handling ambiguity and extrapolating to occluded parts of the scene. Our Exo2Ego framework belongs to the generative-based vein, but unlike all the above, it does not depend on camera poses as critical conditional inputs. This improves scalability and widens the scope of potential applications, as camera parameters are seldom accessible in real-world scenarios and are challenging to infer from RGB images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 337, + 358, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 337, + 358, + 350 + ], + "spans": [ + { + "bbox": [ + 132, + 337, + 358, + 350 + ], + "type": "text", + "content": "2.3 Cross-view Image-to-Image Translation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 357, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 477 + ], + "type": "text", + "content": "Exo-to-ego view synthesis is inherently an image-to-image translation problem [5, 10, 22, 61, 62]—but with drastic viewpoint change. Prior work centers around aerial-to-ground translation [40–42, 57] or generating an egocentric view without human-object interactions, i.e. a person walking indoors [30, 31, 57]. Except for P-GAN [30], all of these approaches require the ground truth semantic map of the target domain at training or inference time to ensure high-quality generation—an unrealistic assumption that means a good portion of the translation task (the boundaries of the objects in the unobserved ego view or the input exo view) is already known, Compared with the semantic-map-free approach P-GAN, our Exo2Ego exhibits significantly better qualitative and quantitative results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 495, + 246, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 495, + 246, + 506 + ], + "spans": [ + { + "bbox": [ + 132, + 495, + 246, + 506 + ], + "type": "text", + "content": "2.4 Diffusion Models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 514, + 482, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 587 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 587 + ], + "type": "text", + "content": "DMs have shown competitive performance in image synthesis [11, 18, 19, 55], but their use for cross-view translation remains unexplored. We develop an exo-to-ego translation framework incorporating the diffusion model rather than GANs, considering potential advantages such as addressing mode collapse, promoting sample diversity, and maintaining stable training dynamics for improved image generation [44]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 604, + 233, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 604, + 233, + 618 + ], + "spans": [ + { + "bbox": [ + 132, + 604, + 233, + 618 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "First, we formalize our problem (Sec. 3.1). Then we present the two stages of our model (Sec. 3.2 and 3.3), followed by an overview of training and inference (Sec. 3.4)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "M. Luo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 127, + 462, + 215 + ], + "blocks": [ + { + "bbox": [ + 145, + 117, + 250, + 125 + ], + "lines": [ + { + "bbox": [ + 145, + 117, + 250, + 125 + ], + "spans": [ + { + "bbox": [ + 145, + 117, + 250, + 125 + ], + "type": "text", + "content": "(a) High-level Structure Transformation" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 145, + 127, + 462, + 215 + ], + "lines": [ + { + "bbox": [ + 145, + 127, + 462, + 215 + ], + "spans": [ + { + "bbox": [ + 145, + 127, + 462, + 215 + ], + "type": "image", + "image_path": "57b400d9f41d668deba24255b68beceb7a6c8c8388d993c8662050fffc0ce089.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 149, + 230, + 467, + 330 + ], + "blocks": [ + { + "bbox": [ + 148, + 220, + 252, + 229 + ], + "lines": [ + { + "bbox": [ + 148, + 220, + 252, + 229 + ], + "spans": [ + { + "bbox": [ + 148, + 220, + 252, + 229 + ], + "type": "text", + "content": "(b) Diffusion-based Pixel Hallucination" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 149, + 230, + 467, + 330 + ], + "lines": [ + { + "bbox": [ + 149, + 230, + 467, + 330 + ], + "spans": [ + { + "bbox": [ + 149, + 230, + 467, + 330 + ], + "type": "image", + "image_path": "810ff5de80a80597b291fd03348103d831ab2fc360e06e47a952ed5109a671e8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 341, + 482, + 397 + ], + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 397 + ], + "type": "text", + "content": "Fig. 2: Our Exo2Ego framework comprises two modules: (a) High-level Structure Transformation, which predicts the ego layout, capturing hand position and interactions using a transformer-based encoder-decoder architecture. (b) Diffusion-based Pixel Hallucination, which enhances pixel-level details on top of the ego layout using a conditional diffusion model." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 403, + 268, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 403, + 268, + 414 + ], + "spans": [ + { + "bbox": [ + 131, + 403, + 268, + 414 + ], + "type": "text", + "content": "3.1 Problem Formulation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 422, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 495 + ], + "type": "text", + "content": "Whereas conventional geometry-aware approaches [10,35,70] generate novel views by specifying camera poses and performing volume rendering, Exo2Ego offers a different perspective. We propose tackling the exo-to-ego translation problem in a purely probabilistic framework. Intuitively, this is essential to address the inherent ambiguity in predicting the ego view from the exo view, e.g., due to entirely unseen portions of objects or the human body." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 495, + 482, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 589 + ], + "type": "text", + "content": "We define " + }, + { + "bbox": [ + 130, + 495, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_T = \\{x_1, \\ldots, x_i, \\ldots, x_T\\}" + }, + { + "bbox": [ + 130, + 495, + 482, + 589 + ], + "type": "text", + "content": " as a series of " + }, + { + "bbox": [ + 130, + 495, + 482, + 589 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 495, + 482, + 589 + ], + "type": "text", + "content": " video frames captured from an exo viewpoint, characterized by static backgrounds with dynamic actors and other objects present, where " + }, + { + "bbox": [ + 130, + 495, + 482, + 589 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 495, + 482, + 589 + ], + "type": "text", + "content": " denotes the time index. This exo perspective reveals the actors' motions and (potentially) full-body poses within the scene. Let " + }, + { + "bbox": [ + 130, + 495, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_T = \\{y_1, \\ldots, y_i, \\ldots, y_T\\}" + }, + { + "bbox": [ + 130, + 495, + 482, + 589 + ], + "type": "text", + "content": " denote the corresponding sequence of ego frames captured from a first-person perspective. This perspective emulates the view of a camera mounted on the head or body of the actor, with a focus on their actions and interactions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 590, + 482, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 590, + 482, + 651 + ], + "spans": [ + { + "bbox": [ + 131, + 590, + 482, + 651 + ], + "type": "text", + "content": "The goal of exo-to-ego view translation is to simulate the view of the ego camera-wearer in the scene recorded by an exo camera. Formally, we seek a translation model that can map " + }, + { + "bbox": [ + 131, + 590, + 482, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_T" + }, + { + "bbox": [ + 131, + 590, + 482, + 651 + ], + "type": "text", + "content": " into a series of output ego frames, " + }, + { + "bbox": [ + 131, + 590, + 482, + 651 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}}_T = \\{\\hat{y}_1,\\dots ,\\hat{y}_i,\\dots ,\\hat{y}_T\\}" + }, + { + "bbox": [ + 131, + 590, + 482, + 651 + ], + "type": "text", + "content": ". As shown in Eqn. (1), the conditional distribution of " + }, + { + "bbox": [ + 131, + 590, + 482, + 651 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}}^T" + }, + { + "bbox": [ + 131, + 590, + 482, + 651 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 131, + 590, + 482, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^T" + }, + { + "bbox": [ + 131, + 590, + 482, + 651 + ], + "type": "text", + "content": " should be indistinguishable from the conditional distribution of " + }, + { + "bbox": [ + 131, + 590, + 482, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}^T" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 210, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 210, + 91, + 448, + 102 + ], + "type": "text", + "content": "Lifting the Egocentric Perspective from Exocentric Videos" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 178, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 178, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 178, + 128 + ], + "type": "text", + "content": "given " + }, + { + "bbox": [ + 132, + 115, + 178, + 128 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^T" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 253, + 128, + 480, + 141 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 128, + 480, + 141 + ], + "spans": [ + { + "bbox": [ + 253, + 128, + 480, + 141 + ], + "type": "interline_equation", + "content": "p \\left(\\mathbf {\\hat {Y}} _ {T} \\mid \\mathbf {X} _ {T}\\right) = p \\left(\\mathbf {Y} _ {T} \\mid \\mathbf {X} _ {T}\\right). \\tag {1}", + "image_path": "c97d4b0d51fdbc89c93e15687b43e6de4b106d596d0536553e71b8848afbf7b2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 146, + 482, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 146, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 146, + 482, + 312 + ], + "type": "text", + "content": "We prioritize translating daily tabletop activities from the exo to ego view, which is a prevalent setting in egocentric learning and frequently entails extensive interactions between hands and objects. This problem is highly challenging as the translation model must produce photorealistic hand-object interaction sequences in the ego view while also performing geometric and semantic reasoning, i.e., correctly predicting the spatial location and pixel-level appearance of visual concepts. Furthermore, it requires linking the actor's head/body motion captured by the exo camera with the viewpoint change in the ego video. Upon examining the application of the widely used pixel-to-pixel generation method [62] for our task, we found it struggles with intricate details of the hands, likely because treats all pixels equally and lacks geometric correspondence between views. At the same time, we know that (at least in manipulation-rich scenarios) the hands of the actor are a common ground between the ego and exo views, despite their many other differences." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 313, + 482, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 313, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 313, + 482, + 433 + ], + "type": "text", + "content": "Motivated by these points, we propose the Exo2Ego framework, which disentangles the understanding of cross-view correspondences and pixel-level synthesis. It consists of two key modules, as illustrated in Figure 2: (1) High-level Structure Transformation, which addresses the task of inferring the location and interaction manner of hands and objects in the ego view. To accomplish this, we train a transformer-based encoder-decoder model to translate the exo frame into an ego hand-object interaction layout (detailed below). (2) Diffusion-based Pixel Hallucination, which learns to synthesize realistic and high-quality pixel-level details by training a conditional diffusion model operating on top of the ego hand layout." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 451, + 345, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 451, + 345, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 451, + 345, + 463 + ], + "type": "text", + "content": "3.2 High-level Structure Transformation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 471, + 482, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 615 + ], + "type": "text", + "content": "Given an exo frame, the purpose of the high-level structure transformation is to train a layout translator which predicts the ego layout showing the location and rough contour of the hands. Specifically, to capture fine-grained hand-object interaction details, we propose to generate the hand layout, instantiated as 2D hand poses. The generated layout serves as a crucial reference for further pixel-level hallucination. To achieve this, we draw inspiration from a recent study [45] that highlights transformer-based architectures [12, 59] succeed in understanding cross-view correspondence, due to their reduced locality-bias. Considering transformers have the potential to represent the geometric information implicitly, we employ a pure transformer-based encoder-decoder architecture, which enables us to effectively incorporate and process the comprehensive context of the exo scene, and thus generate a precise ego hand layout." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 626, + 482, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 626, + 482, + 662 + ], + "spans": [ + { + "bbox": [ + 131, + 626, + 482, + 662 + ], + "type": "text", + "content": "Exo Contextual Feature Extraction As shown in Figure 2, an exo frame " + }, + { + "bbox": [ + 131, + 626, + 482, + 662 + ], + "type": "inline_equation", + "content": "x_{i} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 131, + 626, + 482, + 662 + ], + "type": "text", + "content": " concatenated with the corresponding exo layout " + }, + { + "bbox": [ + 131, + 626, + 482, + 662 + ], + "type": "inline_equation", + "content": "x_{i}^{l} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 131, + 626, + 482, + 662 + ], + "type": "text", + "content": " (2D hand pose layout recorded as image size) is first split into a sequence of patches" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "M. Luo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 200 + ], + "type": "text", + "content": "and then processed by an input embedding layer, such as patch embedding for ViTs [12] to get a tokenized input sequence " + }, + { + "bbox": [ + 131, + 116, + 479, + 200 + ], + "type": "inline_equation", + "content": "e_i \\in \\mathbb{R}^{M \\times D}" + }, + { + "bbox": [ + 131, + 116, + 479, + 200 + ], + "type": "text", + "content": ". To capture the positional information which is critical for spatial relationship reasoning, we add a learnable position embedding to the sequence of patches. Then the sequence of tokens is passed into a transformer encoder consisting of " + }, + { + "bbox": [ + 131, + 116, + 479, + 200 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 116, + 479, + 200 + ], + "type": "text", + "content": " repeated blocks. Each block contains a multi-head attention token mixer to first communicate information among tokens, which is represented as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 223, + 208, + 481, + 222 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 208, + 481, + 222 + ], + "spans": [ + { + "bbox": [ + 223, + 208, + 481, + 222 + ], + "type": "interline_equation", + "content": "z _ {i} ^ {j - 1} = \\operatorname {A t t e n t i o n} \\left(\\operatorname {N o r m} \\left(e _ {i} ^ {j - 1}\\right)\\right) + e _ {i} ^ {j - 1}, \\tag {2}", + "image_path": "d00d0b10d0da51aa2a76349d312039dc8b0771f0e33f3cae017aa2cac9de9ad6.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 230, + 479, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 230, + 479, + 254 + ], + "spans": [ + { + "bbox": [ + 131, + 230, + 479, + 254 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 230, + 479, + 254 + ], + "type": "inline_equation", + "content": "j \\in \\{1, \\dots, N\\}" + }, + { + "bbox": [ + 131, + 230, + 479, + 254 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 230, + 479, + 254 + ], + "type": "inline_equation", + "content": "\\mathrm{Norm}(\\cdot)" + }, + { + "bbox": [ + 131, + 230, + 479, + 254 + ], + "type": "text", + "content": " is the normalization method. Then, the mixed token " + }, + { + "bbox": [ + 131, + 230, + 479, + 254 + ], + "type": "inline_equation", + "content": "z_i^{j-1}" + }, + { + "bbox": [ + 131, + 230, + 479, + 254 + ], + "type": "text", + "content": " is passed into a two-layer MLP expressed as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 238, + 263, + 480, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 263, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 238, + 263, + 480, + 277 + ], + "type": "interline_equation", + "content": "e _ {i} ^ {j} = \\operatorname {M L P} \\left(\\operatorname {N o r m} \\left(z _ {i} ^ {j - 1}\\right)\\right) + z _ {i} ^ {j - 1}. \\tag {3}", + "image_path": "958202cc3de6b37a0b185ee8e4e3d699aefb05910b95cb9de5d05fbaef8cd399.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 285, + 479, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 285, + 479, + 333 + ], + "spans": [ + { + "bbox": [ + 131, + 285, + 479, + 333 + ], + "type": "text", + "content": "After iterating through all " + }, + { + "bbox": [ + 131, + 285, + 479, + 333 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 285, + 479, + 333 + ], + "type": "text", + "content": " blocks, the input sequence " + }, + { + "bbox": [ + 131, + 285, + 479, + 333 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 131, + 285, + 479, + 333 + ], + "type": "text", + "content": " undergoes a mapping process that transforms it into the final contextual tokens " + }, + { + "bbox": [ + 131, + 285, + 479, + 333 + ], + "type": "inline_equation", + "content": "c_{i}\\in \\mathbb{R}^{M\\times D}" + }, + { + "bbox": [ + 131, + 285, + 479, + 333 + ], + "type": "text", + "content": ". We talk about how to decode the contextual tokens into the desired ego layout in the following part." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 344, + 480, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 344, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 131, + 344, + 480, + 453 + ], + "type": "text", + "content": "Layout Decoder Consider a hand pose image with shape " + }, + { + "bbox": [ + 131, + 344, + 480, + 453 + ], + "type": "inline_equation", + "content": "\\{H\\times W\\times 3\\}" + }, + { + "bbox": [ + 131, + 344, + 480, + 453 + ], + "type": "text", + "content": " as the target ego layout. The contextual tokens " + }, + { + "bbox": [ + 131, + 344, + 480, + 453 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 131, + 344, + 480, + 453 + ], + "type": "text", + "content": " should be decoded into 2D hand joint coordinates and then rendered to the RGB image. First, the contextual tokens " + }, + { + "bbox": [ + 131, + 344, + 480, + 453 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 131, + 344, + 480, + 453 + ], + "type": "text", + "content": " are combined with a sequence of learnable query embeddings " + }, + { + "bbox": [ + 131, + 344, + 480, + 453 + ], + "type": "inline_equation", + "content": "q_{i}\\in \\mathbb{R}^{E\\times D}" + }, + { + "bbox": [ + 131, + 344, + 480, + 453 + ], + "type": "text", + "content": " and passed into a decoder architecture, consisting of " + }, + { + "bbox": [ + 131, + 344, + 480, + 453 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 344, + 480, + 453 + ], + "type": "text", + "content": " transformer blocks, as described earlier. Following the transformer blocks, a regression head is employed to estimate the 2D coordinates of the joints whose range is restricted to [0,1]. Then we apply the classic bipartite matching loss [8, 39] to map the regressed hand joints to the ground truth joints." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 469, + 338, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 469, + 338, + 480 + ], + "spans": [ + { + "bbox": [ + 132, + 469, + 338, + 480 + ], + "type": "text", + "content": "3.3 Diffusion-based Pixel Hallucination" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 488, + 479, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 488, + 479, + 643 + ], + "spans": [ + { + "bbox": [ + 131, + 488, + 479, + 643 + ], + "type": "text", + "content": "Having inferred the ego hand pose layout " + }, + { + "bbox": [ + 131, + 488, + 479, + 643 + ], + "type": "inline_equation", + "content": "y_{i}^{l}" + }, + { + "bbox": [ + 131, + 488, + 479, + 643 + ], + "type": "text", + "content": ", next the aim of pixel-level hallucination is to synthesize photorealistic ego frames by taking into account an exo frame " + }, + { + "bbox": [ + 131, + 488, + 479, + 643 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 131, + 488, + 479, + 643 + ], + "type": "text", + "content": " and the target ego layout. We use the diffusion formulation proposed in the denoising diffusion probabilistic model (DDPM) [18] and train the diffusion model in the latent space. As shown in Figure 2, we first adopt a pre-trained variational autoencoder (VAE) model [25] as used in [44] to encode the original ego frames " + }, + { + "bbox": [ + 131, + 488, + 479, + 643 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 131, + 488, + 479, + 643 + ], + "type": "text", + "content": " and conditional information exo frame " + }, + { + "bbox": [ + 131, + 488, + 479, + 643 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 131, + 488, + 479, + 643 + ], + "type": "text", + "content": " and ego layout " + }, + { + "bbox": [ + 131, + 488, + 479, + 643 + ], + "type": "inline_equation", + "content": "y_{i}^{l}" + }, + { + "bbox": [ + 131, + 488, + 479, + 643 + ], + "type": "text", + "content": " into the latent space. Then, we train a diffusion transformer [38] to learn the latent data distribution by denoising a latent vector sampled from a Gaussian distribution gradually. Specifically, given an initial noise map " + }, + { + "bbox": [ + 131, + 488, + 479, + 643 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 131, + 488, + 479, + 643 + ], + "type": "text", + "content": " and a conditioning vector " + }, + { + "bbox": [ + 131, + 488, + 479, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 131, + 488, + 479, + 643 + ], + "type": "text", + "content": ", the diffusion model generates the corresponding ego latent " + }, + { + "bbox": [ + 131, + 488, + 479, + 643 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_{\\theta}" + }, + { + "bbox": [ + 131, + 488, + 479, + 643 + ], + "type": "text", + "content": ". A squared error loss is used to denoise a variably-noised latent code " + }, + { + "bbox": [ + 131, + 488, + 479, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{n} = \\alpha^{n}\\mathbf{z} + \\sigma^{n}\\epsilon" + }, + { + "bbox": [ + 131, + 488, + 479, + 643 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 225, + 653, + 480, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 653, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 225, + 653, + 480, + 666 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} _ {\\mathbf {d}, \\mathbf {z}, \\boldsymbol {\\epsilon}, n} \\left[ \\| \\mathbf {z} - \\hat {\\mathbf {z}} _ {\\theta} \\left(\\alpha^ {n} \\mathbf {z} + \\sigma^ {n} \\boldsymbol {\\epsilon}, \\mathbf {d}\\right) \\| _ {2} ^ {2} \\right] \\tag {4}", + "image_path": "652e3ffab7c75d9247beccb2afa3419792843bd0dec21ccbfe00af7ba4fbd3b7.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 209, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 209, + 91, + 448, + 102 + ], + "type": "text", + "content": "Lifting the Egocentric Perspective from Exocentric Videos" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " is the ground-truth latent vector, " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " is the conditioning vector, and " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\alpha^n,\\sigma^n" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " are functions of the diffusion process time step " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": ", which control the noise schedule and sample quality. We simply concatenate the noisy latent vector " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^n" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " and the conditional embedding " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " as the input of the denoising transformer. In contrast to earlier cross-view translation methods based on GANs [30, 61, 62], our conditional diffusion model sequentially updates the target ego outputs, excels in capturing intricate ego-exo dependencies and faithfully reproducing the mapping of the conditional distributions in Eqn. (1). Notably, our model exhibits enhanced stability throughout the training and consistently produces higher-quality samples, as verified by our experiments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 253, + 277, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 253, + 277, + 266 + ], + "spans": [ + { + "bbox": [ + 131, + 253, + 277, + 266 + ], + "type": "text", + "content": "3.4 Training and Inference" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 272, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 344 + ], + "type": "text", + "content": "At the training stage, the two modules described in Section 3.2 and 3.3 are trained separately. During inference, we first generate the corresponding ego hand layout prediction " + }, + { + "bbox": [ + 130, + 272, + 482, + 344 + ], + "type": "inline_equation", + "content": "\\hat{y}_i^l" + }, + { + "bbox": [ + 130, + 272, + 482, + 344 + ], + "type": "text", + "content": " for each input exo frame " + }, + { + "bbox": [ + 130, + 272, + 482, + 344 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 130, + 272, + 482, + 344 + ], + "type": "text", + "content": " with the layout translator described in Section 3.2. Subsequently, " + }, + { + "bbox": [ + 130, + 272, + 482, + 344 + ], + "type": "inline_equation", + "content": "\\hat{y}_i^l" + }, + { + "bbox": [ + 130, + 272, + 482, + 344 + ], + "type": "text", + "content": " is concatenated with " + }, + { + "bbox": [ + 130, + 272, + 482, + 344 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 130, + 272, + 482, + 344 + ], + "type": "text", + "content": " in a channel-wise manner and then used as the conditional input of the pixel-level diffusion generator described in Section 3.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 361, + 302, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 361, + 302, + 376 + ], + "spans": [ + { + "bbox": [ + 131, + 361, + 302, + 376 + ], + "type": "text", + "content": "4 Experimental Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 387, + 481, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 481, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 481, + 411 + ], + "type": "text", + "content": "We introduce an exo-to-ego synthesis benchmark (Sec. 4.1) and then validate our approach against state-of-the-art synthesis methods (Sec. 4.2)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 428, + 398, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 428, + 398, + 440 + ], + "spans": [ + { + "bbox": [ + 131, + 428, + 398, + 440 + ], + "type": "text", + "content": "4.1 A Benchmark for Exo-to-Ego View Translation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 447, + 484, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 484, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 484, + 641 + ], + "type": "text", + "content": "To facilitate new work on the exo-to-ego translation task, we contribute a new cross-view synthesis benchmark sourced from four public time-synchronized multi-view datasets: H2O [27], Aria Pilot [32], Assembly101 [47], and Ego-Exo4D [16]. H2O [27] has indoor videos featuring actors manipulating 8 different objects using both hands. Assembly101 [47] shows people (dis)assembling 101 take-apart toy vehicles. We take 6 activity sequences involving different individuals manipulating a toy roller. Aria Pilot [32] provides 16 multi-view recordings of an actor wearing Project Aria glasses [54] manipulating YCB [7] objects on a desktop. For Ego-Exo4D, we include the four (largely non-table-top) scenarios involving heavy hand-object interactions from the EgoHandPose Benchmark - bike repair, cooking, music, and health—where the camera wearer also moves about the scene. For all datasets, we select the exo camera viewpoints that offer the clearest capture of the action. All frames are cropped and resized into " + }, + { + "bbox": [ + 130, + 447, + 484, + 641 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 447, + 484, + 641 + ], + "type": "text", + "content": ". We divide each video into 30-frame clips. Overall, this dataset showcases an array of human activities with various objects, from toy assembly to the manipulation of everyday objects. More details are provided in Appendix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 642, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 642, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 642, + 481, + 666 + ], + "type": "text", + "content": "We employ the following split settings to evaluate four kinds of generalization: (1) new actions, where we put the first " + }, + { + "bbox": [ + 131, + 642, + 481, + 666 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 131, + 642, + 481, + 666 + ], + "type": "text", + "content": " of a video's clips into the training" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "M. Luo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 491, + 426 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 491, + 426 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 491, + 426 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 491, + 426 + ], + "type": "image", + "image_path": "3d6c4652f9367f296636c1487338979ce1c07a6f60c9d025c3966767eb60bc7f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 437, + 465, + 449 + ], + "lines": [ + { + "bbox": [ + 146, + 437, + 465, + 449 + ], + "spans": [ + { + "bbox": [ + 146, + 437, + 465, + 449 + ], + "type": "text", + "content": "Fig. 3: Qualitative examples when generalizing to new actions on all datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 454, + 482, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 454, + 482, + 622 + ], + "spans": [ + { + "bbox": [ + 130, + 454, + 482, + 622 + ], + "type": "text", + "content": "set and the remaining " + }, + { + "bbox": [ + 130, + 454, + 482, + 622 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 454, + 482, + 622 + ], + "type": "text", + "content": " in the test set, hence splitting up action steps over time (e.g., for a video depicting picking up a coffee box and then taking coffee capsules out, the former will be assigned to the training set and the latter to the test set); (2) new objects, where we train with videos involving any of 7 objects and test with videos containing a novel 8th object; (3) new subjects, where we train with all clips from one subject and test on clips from another (subject " + }, + { + "bbox": [ + 130, + 454, + 482, + 622 + ], + "type": "inline_equation", + "content": "1 \\rightarrow" + }, + { + "bbox": [ + 130, + 454, + 482, + 622 + ], + "type": "text", + "content": " subject 2); (4) new scenes (backgrounds), where we train and test with disjoint scenes (scene h1, h2, k1, k2 " + }, + { + "bbox": [ + 130, + 454, + 482, + 622 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 454, + 482, + 622 + ], + "type": "text", + "content": " scene o1, o2). We analyze new action generalization on all 3 datasets, and the rest on H2O only, as it is the only dataset with the variations and meta-data to allow such split settings. For action generalization, H2O, Aria Pilot, Assembly101, and Ego-Exo4D have 704, 343, 682, and 543 training clips, and 199, 95, 205, and 211 test clips, respectively. For other generalization settings (new objects, subjects, and scenes), there are 691, 704, and 591 training clips, and 108, 194, and 312 test clips." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 636, + 307, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 636, + 307, + 649 + ], + "spans": [ + { + "bbox": [ + 132, + 636, + 307, + 649 + ], + "type": "text", + "content": "4.2 Evaluating Synthesis Quality" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 653, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 653, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 653, + 480, + 666 + ], + "type": "text", + "content": "With this well-constructed benchmark in hand, we proceed to validate Exo2Ego." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 210, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 210, + 91, + 448, + 102 + ], + "type": "text", + "content": "Lifting the Egocentric Perspective from Exocentric Videos" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 160, + 146, + 451, + 402 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 1: Generalizing to new actions on three datasets. *pixelNeRF requires privileged camera information relative to the other baseline models." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 160, + 146, + 451, + 402 + ], + "lines": [ + { + "bbox": [ + 160, + 146, + 451, + 402 + ], + "spans": [ + { + "bbox": [ + 160, + 146, + 451, + 402 + ], + "type": "table", + "html": "
H2O
MethodSSIM↑PSNR↑FID↓P_Squeeze↓P_Alex↓P_Vgg↓Feasi.↑
pix2pixHD [62]0.42830.370132.030.1630.2200.3370.8952
P-GAN [30]0.27229.079192.260.2420.3080.4490.8353
vid2vid [61]0.40229.88285.760.1660.2260.3420.9116
pixelNeRF* [70]0.21927.871470.130.7110.8070.7310.0086
DiT [38]0.43630.55345.230.1290.1780.2970.9604
Exo2Ego (Ours)0.44330.58037.030.1040.1460.2620.9764
Aria Pilot
MethodSSIM↑PSNR↑FID↓P_Squeeze↓P_Alex↓P_Vgg↓Feasi.↑
pix2pixHD [62]0.35029.553159.910.3040.3620.4800.0790
P-GAN [30]0.34329.769122.400.2870.3450.4750.1014
vid2vid [61]0.35929.85852.380.2660.3280.4430.2646
DiT [38]0.36829.94532.710.2400.2980.4170.4784
Exo2Ego (Ours)0.38030.04725.990.2250.2790.4020.5977
Assembly101
MethodSSIM↑PSNR↑FID↓P_Squeeze↓P_Alex↓P_Vgg↓Feasi.↑
pix2pixHD [62]0.40529.909112.710.2160.2950.4010.1184
P-GAN [30]0.37229.303114.680.2160.3000.4360.0800
vid2vid [61]0.36829.48178.470.2240.3120.4240.1358
DiT [38]0.39829.99626.330.1790.2550.3700.3092
Exo2Ego (Ours)0.41330.15325.230.1680.2380.3550.4019
", + "image_path": "78205a34281d856ce75a95de6806ed950b33aef27d4e08a632019db8634fed1b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 406, + 482, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 406, + 482, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 406, + 482, + 538 + ], + "type": "text", + "content": "Baselines We consider several baselines: (1) Pix2PixHD [62], a single-view image translation method that processes the videos frame by frame; (2) P-GAN [30], a recent exo-to-ego view translation method that proposes a parallel generative network to facilitate cross-view image translation; (3) DiT [38], a pixel-to-pixel generative model from the latent diffusion model family. (4) Vid2Vid [61], a single-view video translation method that models the temporal dynamics in the video; (5) pixelNeRF [70], a NeRF-like model known for its remarkable generalization capabilities and its accommodating stance toward sparser input views. Note that pixelNeRF assumes camera parameters are known, while other methods do not. We conduct experiments with pixelNeRF only on H2O, as the other two datasets lack the camera information needed by pixelNeRF." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 540, + 482, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 540, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 540, + 482, + 612 + ], + "type": "text", + "content": "Implementation Details For Exo2Ego's high-level structure transformation transformer, the number of blocks " + }, + { + "bbox": [ + 130, + 540, + 482, + 612 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 540, + 482, + 612 + ], + "type": "text", + "content": " is set to 6. We use DiT-XL/2 [38] as the denoising architecture for Exo2Ego. The diffusion model is trained for 40,000 steps and pixelNeRF is trained for 70,000 steps for all datasets. For H2O and Assembly101, pix2pixHD, and P-GAN are trained for 100 epochs, and vid2vid is trained for 40 epochs. For Aria Pilot, we use 400 and 100 epochs respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 613, + 482, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 613, + 482, + 662 + ], + "spans": [ + { + "bbox": [ + 130, + 613, + 482, + 662 + ], + "type": "text", + "content": "Evaluation Metrics We measure the quality of the synthesized videos as follows. First, we evaluate the feasibility of generated hands, denoted as Feasi. Inspired by [69], we adopt an off-the-shelf egocentric hand object detector [49] to measure the physical feasibility of the synthesized hands. This metric is calcu" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "M. Luo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 116, + 492, + 250 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 492, + 250 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 492, + 250 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 492, + 250 + ], + "type": "image", + "image_path": "30ef4e009d01ea1429b8a18ca20c08788b2cb3baa1dc401b63e979d95ef1daef.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 258, + 480, + 281 + ], + "lines": [ + { + "bbox": [ + 132, + 258, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 258, + 480, + 281 + ], + "type": "text", + "content": "Fig. 4: Qualitative examples when generalizing to new objects, subjects, and scenes (backgrounds) on the H2O dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 287, + 479, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 287, + 479, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 287, + 479, + 310 + ], + "type": "text", + "content": "lated as the confidence score of hand detection results produced by the detector, averaged over all synthesized frames." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 311, + 494, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 494, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 494, + 419 + ], + "type": "text", + "content": "We also use structural similarity (SSIM) and peak signal-to-noise ratio (PSNR) [20] to check the pixel-level similarity between the predicted ego frame and the ground truth (GT) ego frame. These two evaluation metrics should ideally remain stable under minor transformations (such as small translations or affine transformations). We further adopt perceptual metrics, including FID [46] and Learned Perceptual Image Patch Similarity (LPIPS) [74] which evaluates the feature-level similarity of the predicted ego frame and the true ego frame. LPIPS are denoted as " + }, + { + "bbox": [ + 130, + 311, + 494, + 419 + ], + "type": "inline_equation", + "content": "P_{\\text{Squeeze}}" + }, + { + "bbox": [ + 130, + 311, + 494, + 419 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 311, + 494, + 419 + ], + "type": "inline_equation", + "content": "P_{\\text{Alex}}" + }, + { + "bbox": [ + 130, + 311, + 494, + 419 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 311, + 494, + 419 + ], + "type": "inline_equation", + "content": "P_{\\text{Vgg}}" + }, + { + "bbox": [ + 130, + 311, + 494, + 419 + ], + "type": "text", + "content": ", which uses SqueezeNet [21], AlexNet [26], and VGG [52] as the feature extractor." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "text", + "content": "Generalization to New Actions Figure 3 showcases qualitative results on H2O, Aria Pilot, and Assembly101, demonstrating the superiority of our Exo2Ego framework. Compared to all baselines, our approach produces realistic hands with correct poses, especially noticeable in the highlighted yellow circle regions when dealing with new actions during test time. Table 1 presents the feasibility of synthesized hands and the various measures of pixel-level similarity and perceptual similarity between the predicted ego frame and the ground truth ego frame. Our Exo2Ego greatly increases the feasibility of synthesized hands, with a confidence gain of up to " + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "inline_equation", + "content": "11.93\\%" + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "text", + "content": " on Aria Pilot, on the basis of the second-highest score of " + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "inline_equation", + "content": "47.84\\%" + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "text", + "content": " achieved by DiT. This improvement indicates enhanced realism in the generated hands, resulting in greater confidence in hand detection results. Additionally, our Exo2Ego can produce ego frames with higher pixel-level similarity to the ground truth frames and better perceptual metrics (much lower FID scores). This highlights the benefits of the 'decoupling' idea of our Exo2Ego framework in generating photorealistic ego frames." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": "Looking at Table 1, there is a trend among our Exo2Ego and the baselines — doing better on H2O dataset. We conjecture that the superiority comes from H2O's higher hand quality, including clear pose visibility and larger spatial coverage within the field of view, in both exo and ego view." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 210, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 210, + 91, + 448, + 102 + ], + "type": "text", + "content": "Lifting the Egocentric Perspective from Exocentric Videos" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 159, + 146, + 451, + 425 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 2: Generalizing to different objects, subjects, and scenes. *pixelNeRF requires privileged camera information relative to the other baseline models." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 159, + 146, + 451, + 425 + ], + "lines": [ + { + "bbox": [ + 159, + 146, + 451, + 425 + ], + "spans": [ + { + "bbox": [ + 159, + 146, + 451, + 425 + ], + "type": "table", + "html": "
H2O-unseen object
MethodSSIM↑PSNR↑FID↓P_Squeeze↓P_Alex↓P_Vgg↓Feasi.↑
pix2pixHD [62]0.33329.504253.730.2690.3600.4670.7527
P-GAN [30]0.27028.915264.760.2770.3500.4920.7720
vid2vid [61]0.34929.585145.390.2420.3240.4280.8612
pixelNeRF* [70]0.26227.912473.730.7500.8410.7360.0000
DiT [38]0.33829.740105.180.2140.2880.4100.9689
Exo2Ego (Ours)0.34429.786101.450.2080.2810.4030.9786
H2O-unseen subject
MethodSSIM↑PSNR↑FID↓P_Squeeze↓P_Alex↓P_Vgg↓Feasi.↑
pix2pixHD [62]0.06928.028477.470.6610.7540.7650.0140
P-GAN [30]0.08828.063451.370.6230.7080.7250.2002
vid2vid [61]0.13628.094308.520.5640.6900.6860.3838
pixelNeRF* [70]0.17027.874459.490.7690.8360.7680.0000
DiT [38]0.17228.185183.060.5240.6300.6820.8405
Exo2Ego (Ours)0.17828.266150.870.5030.6290.6650.9666
H2O-unseen scene
MethodSSIM↑PSNR↑FID↓P_Squeeze↓P_Alex↓P_Vgg↓Feasi.↑
pix2pixHD [62]0.02427.918558.870.7290.7890.7880.0013
P-GAN [30]0.08527.928325.340.6080.6990.7240.6041
vid2vid [61]0.03327.942398.780.6360.7310.7340.2095
pixelNeRF* [70]0.10327.810506.100.7730.9080.8000.0000
DiT [38]0.14027.874308.980.5840.7290.6780.8317
Exo2Ego (Ours)0.15727.943249.040.5310.7090.6580.9635
", + "image_path": "b4aaee35d2c9653dfd23f18ed6bd038216c9166c0ab54274255ef83c0b9770fc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "text", + "content": "We further explored how Exo2Ego fares in the settings with heavy hand-object manipulation and frequent head motion, but minimal movement across the broader scene on Ego-Exo4D [16]. As shown in Figure 6, we compare our Exo2Ego with a strong image-to-image translation model ControlNet [73], which is built on a stable diffusion trained on large scales of data. It is observed that our Exo2Ego synthesizes feasible hands with correct poses, unlike ControlNet. The exo image's hand regions in Ego-Exo4D are small, so we crop it to use as additional input to offer a better view of the hands for both methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 537, + 482, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 537, + 482, + 658 + ], + "spans": [ + { + "bbox": [ + 130, + 537, + 482, + 658 + ], + "type": "text", + "content": "Generalization to New Objects, Subjects, and Backgrounds Illustrated in Table 2, our Exo2Ego outperforms the baselines in terms of quantitative metrics for generalization to new objects/subjects/backgrounds. Figure 4 shows that our Exo2Ego can produce realistic hand-object interactions when encountering exo views with new distributions, while other baselines fail miserably. This is mainly because our diffusion-based generative mechanism gradually denoises the target ego view, capturing the inherent intricate mapping from exo to ego views better. Another observation is that pixelNeRF produces ego views characterized by noticeable blurring and a lack of fine details in hands and objects. This observation supports our earlier discussions on the limitations of 3D-aware approaches." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "M. Luo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 477, + 243 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 477, + 243 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 477, + 243 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 477, + 243 + ], + "type": "image", + "image_path": "42f834b8cb09ccf73596a17383d0340687adc4c67584b1d5cc01ad0c4ebb423d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 134, + 247, + 477, + 258 + ], + "lines": [ + { + "bbox": [ + 134, + 247, + 477, + 258 + ], + "spans": [ + { + "bbox": [ + 134, + 247, + 477, + 258 + ], + "type": "text", + "content": "Fig. 5: Exo2Ego framework generates ego videos with reasonable viewpoint changes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 135, + 263, + 479, + 358 + ], + "blocks": [ + { + "bbox": [ + 135, + 263, + 479, + 358 + ], + "lines": [ + { + "bbox": [ + 135, + 263, + 479, + 358 + ], + "spans": [ + { + "bbox": [ + 135, + 263, + 479, + 358 + ], + "type": "image", + "image_path": "35de2f6f4c3b2e12a6f8f8d6a551b86955c7675b22d9e12c9a7e1dba93cb56ea.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 369, + 480, + 381 + ], + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 381 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 381 + ], + "type": "text", + "content": "Fig. 6: Qualitative examples when generalizing to new actions on Ego-Exo4D dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 387, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 544 + ], + "type": "text", + "content": "Modeling Egocentric Viewpoint Change Due to the probabilistic nature of our Exo2Ego framework, it excels in capturing the relationship between the viewpoint changes in the ego view and the head motions observed in the exo view, holding the potential to deliver more immersive virtual experiences. Intrinsically, our Exo2Ego framework leverages the fact that ego-camera pose is conditioned on the reference exo view, as the ego-camera wearer is usually visible in the exo view. This is highlighted by the yellow gaze direction and dashed box in Figure 5. As the actor moves their head when interacting with objects, the slope of the desk edge in the ego view also changes, indicating a corresponding change in ego viewpoint. This implies the ego-camera motion can be implicitly inferred via conditional generative modeling, even if it's not enforced explicitly. The ability to model ego viewpoint changes is crucial, especially in physical activities like playing ball games which involve heavy head and body motions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 561, + 328, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 561, + 328, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 561, + 328, + 574 + ], + "type": "text", + "content": "5 Limitations and Future Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": "Temporal Coherence Our Exo2Ego operates on each frame individually — the generation of the current ego frame is solely conditioned on the current exo input frame. There is no explicit consideration of previously generated ego frames or past exo frames. Temporal coherence is orthogonal to our current contribution and can be used to improve any of the baseline models without such design." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 210, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 210, + 91, + 448, + 102 + ], + "type": "text", + "content": "Lifting the Egocentric Perspective from Exocentric Videos" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 475, + 159 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 475, + 159 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 475, + 159 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 475, + 159 + ], + "type": "image", + "image_path": "2bee228fb7920b58dd07171f1d09bba823ab610d4d56ef0c6e7d90279e64b4be.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 163, + 480, + 185 + ], + "lines": [ + { + "bbox": [ + 132, + 163, + 480, + 185 + ], + "spans": [ + { + "bbox": [ + 132, + 163, + 480, + 185 + ], + "type": "text", + "content": "Fig. 7: Failure cases: Exo2Ego generates hands with reasonably accurate positions/poses but incomplete/wrong objects in the ego view." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 192, + 482, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 192, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 192, + 482, + 384 + ], + "type": "text", + "content": "Generalization Ability As an early-stage effort for exo-to-ego cross-view translation, our work makes non-trivial progress in producing reasonable hand manipulation details for novel actions. However, despite establishing stronger generalizations than the existing models, our model does not generalize perfectly to in-the-wild objects, subjects, and backgrounds, due to the limited scale of available training data. Considering the intricate nature of the generalization to new environments, we regard it as a challenging future research direction. Nonetheless, our focus on generalizing to new actions within the same environment is meaningful for several applications. Some examples: Real-time monitoring during physical rehabilitation (exo-to-ego view translation can help therapists guide patients during rehabilitation sessions), assistive cooking aid for visually impaired individuals (ego views are provided to the visually impaired individuals to cook independently), or real-time fitness training monitoring (coaches use synthesized ego views to help the trainee correct the fitness movements) — in each case, the environment is stable between training and testing, and yet the model must generalize to new actions for cross-view translation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 395, + 482, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 395, + 482, + 526 + ], + "spans": [ + { + "bbox": [ + 130, + 395, + 482, + 526 + ], + "type": "text", + "content": "3D prior Our qualitative results expose limitations in existing baselines and our Exo2Ego framework in generating 3D-consistent views for objects during test time, attributed to the absence of geometric priors for common objects. See Figure 7. Future work involves integrating robust object geometric priors. While leveraging predicted camera parameters for encoding 3D geometry is a promising approach, as shown Figure 4, the popular 3D-aware method pixelNeRF [70] faces challenges in the dynamic exo-ego setting characterized by hand and object interactions, due to limitations in handling sparse inputs, occluded scenes, and long-range camera extrapolations. Notably, our Exo2Ego outperforms pixelNeRF, highlighting the potential for future work to explore explicit geometric reasoning within our framework." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 544, + 220, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 220, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 220, + 558 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "Overall, our work contributes to the growing body of research in cross-view translation and lays the groundwork for the important case of exo-to-ego. Our generative framework Exo2Ego integrates high-level structure transformation and pixel-level hallucination to yield very encouraging experimental results. We also provide a curated benchmark task for supporting continued work on this problem. The core technology holds immense potential for applications in robot learning and AR skill coaching, where an ego actor needs to replicate the actions of a demonstrator observed from the exo perspective." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "M. Luo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 177 + ], + "type": "text", + "content": "Acknowledgment: UT Austin is supported in part by NSF Grants AF 1901292, CNS 2148141, Tripods CCF 1934932, IFML CCF 2019844 and research gifts by Western Digital, Amazon, WNCG IAP, UT Austin Machine Learning Lab (MLL), Cisco, and the Stanly P. Finch Centennial Professorship in Engineering. K.G. is paid as a research scientist at Meta." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 191, + 197, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 191, + 197, + 204 + ], + "spans": [ + { + "bbox": [ + 133, + 191, + 197, + 204 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 213, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 213, + 481, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 213, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 138, + 213, + 481, + 257 + ], + "type": "text", + "content": "1. Ardeshir, S., Borji, A.: Ego2top: Matching viewers in egocentric and top-view videos. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14. pp. 253-268. Springer (2016) 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 258, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 258, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 138, + 258, + 481, + 278 + ], + "type": "text", + "content": "2. Ardeshir, S., Borji, A.: Egocentric meets top-view. IEEE transactions on pattern analysis and machine intelligence 41(6), 1353-1366 (2018) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 279, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 279, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 279, + 481, + 300 + ], + "type": "text", + "content": "3. Ardeshir, S., Borji, A.: An exocentric look at egocentric actions and vice versa. Computer Vision and Image Understanding 171, 61-68 (2018) 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 300, + 481, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 300, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 481, + 320 + ], + "type": "text", + "content": "4. Bahl, S., Gupta, A., Pathak, D.: Human-to-robot imitation in the wild. arXiv preprint arXiv:2207.09450 (2022) 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 321, + 481, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 481, + 353 + ], + "type": "text", + "content": "5. Bansal, A., Ma, S., Ramanan, D., Sheikh, Y.: Recycle-gan: Unsupervised video retargeting. In: Proceedings of the European conference on computer vision (ECCV). pp. 119-135 (2018) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 353, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 353, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 353, + 481, + 373 + ], + "type": "text", + "content": "6. Bharadhwaj, H., Gupta, A., Tulsiani, S., Kumar, V.: Zero-shot robot manipulation from passive human videos. arXiv preprint arXiv:2302.02011 (2023) 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 374, + 481, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 374, + 481, + 417 + ], + "spans": [ + { + "bbox": [ + 138, + 374, + 481, + 417 + ], + "type": "text", + "content": "7. Calli, B., Singh, A., Walsman, A., Srinivasa, S., Abbeel, P., Dollar, A.M.: The ycb object and model set: Towards common benchmarks for manipulation research. In: 2015 international conference on advanced robotics (ICAR). pp. 510-517. IEEE (2015) 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 418, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 481, + 460 + ], + "type": "text", + "content": "8. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 213-229. Springer (2020) 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 461, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 461, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 138, + 461, + 481, + 493 + ], + "type": "text", + "content": "9. Chan, E.R., Nagano, K., Chan, M.A., Bergman, A.W., Park, J.J., Levy, A., Aittala, M., De Mello, S., Karras, T., Wetzstein, G.: Generative novel view synthesis with 3d-aware diffusion models. arXiv preprint arXiv:2304.02602 (2023) 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 493, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 493, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 493, + 481, + 525 + ], + "type": "text", + "content": "0. Cheng, F., Luo, M., Wang, H., Dimakis, A., Torresani, L., Bertasius, G., Grauman, K.: 4diff: 3d-aware diffusion model for third-to-first viewpoint translation. In: ECCV (2024) 3, 4, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 525, + 481, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 481, + 546 + ], + "type": "text", + "content": "1. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems 34, 8780-8794 (2021) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 547, + 481, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 547, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 134, + 547, + 481, + 590 + ], + "type": "text", + "content": "12. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 590, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 590, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 134, + 590, + 481, + 622 + ], + "type": "text", + "content": "13. Elfeki, M., Regmi, K., Ardeshir, S., Borji, A.: From third person to first person: Dataset and baselines for synthesis and retrieval. arXiv preprint arXiv:1812.00104 (2018) 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 622, + 481, + 665 + ], + "type": "text", + "content": "14. Fan, C., Lee, J., Xu, M., Kumar Singh, K., Jae Lee, Y., Crandall, D.J., Ryoo, M.S.: Identifying first-person camera wearers in third-person videos. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5125-5133 (2017) 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "text", + "content": "Lifting the Egocentric Perspective from Exocentric Videos" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "15. Grauman, K., Westbury, A., Torresani, L., Kitani, K., Malik, J., Afouras, T., Ashutosh, K., Baiyya, V., Bansal, S., Boote, B., et al.: Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. arXiv preprint arXiv:2311.18259 (2023) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 481, + 214 + ], + "type": "text", + "content": "16. Grauman, K., Westbury, A., Torresani, L., Kitani, K., Malik, J., Afouras, T., Ashutosh, K., Baiyya, V., Bansal, S., Boote, B., et al.: Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19383-19400 (2024) 3, 8, 12" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 214, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 481, + 247 + ], + "type": "text", + "content": "17. Ho, H.I., Chiu, W.C., Wang, Y.C.F.: Summarizing first-person videos from third persons' points of view. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 70-85 (2018) 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 247, + 481, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 247, + 481, + 268 + ], + "spans": [ + { + "bbox": [ + 133, + 247, + 481, + 268 + ], + "type": "text", + "content": "18. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33, 6840-6851 (2020) 4, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 268, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 268, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 133, + 268, + 481, + 300 + ], + "type": "text", + "content": "19. Ho, J., Sahara, C., Chan, W., Fleet, D.J., Norouzi, M., Salimans, T.: Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research 23(1), 2249-2281 (2022) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 300, + 481, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 321 + ], + "type": "text", + "content": "20. Hore, A., Ziou, D.: Image quality metrics: Psnr vs. sSIM. In: 2010 20th international conference on pattern recognition. pp. 2366-2369. IEEE (2010) 11" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 321, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 481, + 354 + ], + "type": "text", + "content": "21. Iandola, F.N., Han, S., Moskewicz, M.W., Ashraf, K., Dally, W.J., Keutzer, K.: SqueezeNET: Alexnet-level accuracy with 50x fewer parameters and " + }, + { + "bbox": [ + 132, + 321, + 481, + 354 + ], + "type": "inline_equation", + "content": "< 0.5" + }, + { + "bbox": [ + 132, + 321, + 481, + 354 + ], + "type": "text", + "content": " mb model size. arXiv preprint arXiv:1602.07360 (2016) 11" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "type": "text", + "content": "22. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1125-1134 (2017) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "text", + "content": "23. Jang, W., Agapito, L.: Codenerf: Disentangled neural radiance fields for object categories. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12949-12958 (2021) 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "type": "text", + "content": "24. Jangir, R., Hansen, N., Ghosal, S., Jain, M., Wang, X.: Look closer: Bridging egocentric and third-person views with transformers for robotic manipulation. IEEE Robotics and Automation Letters 7(2), 3046-3053 (2022) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 450, + 481, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 481, + 471 + ], + "type": "text", + "content": "25. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013) 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 471, + 481, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 481, + 492 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 481, + 492 + ], + "type": "text", + "content": "26. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. Communications of the ACM 60(6), 84-90 (2017) 11" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 492, + 481, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 492, + 481, + 535 + ], + "spans": [ + { + "bbox": [ + 132, + 492, + 481, + 535 + ], + "type": "text", + "content": "27. Kwon, T., Tekin, B., Stühmer, J., Bogo, F., Pollefeys, M.: H2o: Two hands manipulating objects for first person interaction recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10138-10148 (October 2021) 3, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 535, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 535, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 535, + 481, + 578 + ], + "type": "text", + "content": "28. Li, Y., Nagarajan, T., Xiong, B., Grauman, K.: Ego-exo: Transferring visual representations from third-person to first-person videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6943-6953 (2021) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 578, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 621 + ], + "type": "text", + "content": "29. Liu, A., Tucker, R., Jampani, V., Makadia, A., Snavely, N., Kanazawa, A.: Infinite nature: Perpetual view generation of natural scenes from a single image. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14458-14467 (2021) 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "30. Liu, G., Tang, H., Latapie, H., Yan, Y.: Exocentric to egocentric image generation via parallel generative adversarial network. In: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1843-1847. IEEE (2020) 3, 4, 8, 10, 12" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "M. Luo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "31. Liu, G., Tang, H., Latapie, H.M., Corso, J.J., Yan, Y.: Cross-view exocentric to egocentric video synthesis. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 974-982 (2021) 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 215 + ], + "type": "text", + "content": "32. Lv, Z., Miller, E., Meissner, J., Pesqueira, L., Sweeney, C., Dong, J., Ma, L., Patel, P., Moulon, P., Somasundaram, K., Parkhi, O., Zou, Y., Raina, N., Saarinen, S., Mansour, Y.M., Huang, P.K., Wang, Z., Troynikov, A., Artal, R.M., DeTone, D., Barnes, D., Argall, E., Lobanovskiy, A., Kim, D.J., Bouttefroy, P., Straub, J., Engel, J.J., Gupta, P., Yan, M., Nardi, R.D., Newcombe, R.: Aria pilot dataset. https://about.facebook.com/realitylabs/projectaria/datasets (2022) 3, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 215, + 482, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 258 + ], + "type": "text", + "content": "33. Majumdar, A., Yadav, K., Arnaud, S., Ma, Y.J., Chen, C., Silwal, S., Jain, A., Berges, V.P., Abbeel, P., Malik, J., et al.: Where are we in the search for an artificial visual cortex for embodied intelligence? arXiv preprint arXiv:2303.18240 (2023) 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 259, + 482, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 281 + ], + "type": "text", + "content": "34. Mandikal, P., Grauman, K.: Dexvip: Learning dexterous grasping with human hand pose priors from video. In: Conference on Robot Learning (2021) 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 281, + 482, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 313 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 313 + ], + "type": "text", + "content": "35. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) 2, 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 314, + 482, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 335 + ], + "type": "text", + "content": "36. Nair, S., Rajeswaran, A., Kumar, V., Finn, C., Gupta, A.: R3m: A universal visual representation for robot manipulation. arXiv preprint arXiv:2203.12601 (2022) 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 335, + 482, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 380 + ], + "type": "text", + "content": "37. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5480-5490 (2022) 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 380, + 482, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 401 + ], + "type": "text", + "content": "38. Peebles, W., Xie, S.: Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748 (2022) 3, 7, 10, 12" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 402, + 482, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 435 + ], + "type": "text", + "content": "39. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified, real-time object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 779-788 (2016) 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 435, + 482, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 467 + ], + "type": "text", + "content": "40. Regmi, K., Borji, A.: Cross-view image synthesis using conditional gans. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 468, + 482, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 511 + ], + "type": "text", + "content": "41. Regmi, K., Borji, A.: Cross-view image synthesis using geometry-guided conditional gans. Computer Vision and Image Understanding (2019). https://doi.org/https://doi.org/10.1016/j.cviu.2019.07.008, http://www.sciencedirect.com/science/article/pii/S1077314219301043 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 511, + 482, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 511, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 511, + 482, + 533 + ], + "type": "text", + "content": "42. Ren, B., Tang, H., Sebe, N.: Cascaded cross mlp-mixer gans for cross-view image translation. arXiv preprint arXiv:2110.10183 (2021) 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 533, + 482, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 566 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 566 + ], + "type": "text", + "content": "43. Ren, X., Wang, X.: Look outside the room: Synthesizing a consistent long-term 3d scene video from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3563-3573 (2022) 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 566, + 482, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 566, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 566, + 482, + 609 + ], + "type": "text", + "content": "44. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022) 4, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 610, + 482, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 610, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 610, + 482, + 643 + ], + "type": "text", + "content": "45. Rombach, R., Esser, P., Ommer, B.: Geometry-free view synthesis: Transformers and no 3d priors. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14356-14366 (2021) 2, 4, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 643, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 643, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 643, + 482, + 666 + ], + "type": "text", + "content": "46. Seitzer, M.: pytorch-fid: FID Score for PyTorch. https://github.com/mseitzer/pytorch-fid (August 2020), version 0.3.0 11" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 210, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 210, + 91, + 447, + 102 + ], + "type": "text", + "content": "Lifting the Egocentric Perspective from Exocentric Videos" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "47. Sener, F., Chatterjee, D., Shelepov, D., He, K., Singhania, D., Wang, R., Yao, A.: Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21096-21106 (2022) 3, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 162, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 162, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 481, + 205 + ], + "type": "text", + "content": "48. Sermanet, P., Lynch, C., Chebotar, Y., Hsu, J., Jang, E., Schaal, S., Levine, S.: Time-contrastive networks: Self-supervised learning from video. Proceedings of International Conference in Robotics and Automation (ICRA) (2018), http://arxiv.org/abs/1704.06888 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "49. Shan, D., Geng, J., Shu, M., Fouhey, D.F.: Understanding human hands in contact at internet scale. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020) 10" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 238, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 281 + ], + "type": "text", + "content": "50. Sigurdsson, G.A., Gupta, A., Schmid, C., Farhadi, A., Alahari, K.: Actor and observer: Joint modeling of first and third-person videos. In: proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7396-7404 (2018) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "type": "text", + "content": "51. Sigurdsson, G.A., Gupta, A., Schmid, C., Farhadi, A., Alahari, K.: Charades-ego: A large-scale dataset of paired third and first person videos. arXiv preprint arXiv:1804.09626 (2018) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "type": "text", + "content": "52. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014) 11" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "53. Sitzmann, V., Zollhöfer, M., Wetzstein, G.: Scene representation networks: Continuous 3d-structure-aware neural scene representations. Advances in Neural Information Processing Systems 32 (2019) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 370, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 481, + 402 + ], + "type": "text", + "content": "54. Somasundaram, K., Dong, J., Tang, H., Straub, J., Yan, M., Goesele, M., Engel, J.J., De Nardi, R., Newcombe, R.: Project aria: A new tool for egocentric multimodal ai research. arXiv preprint arXiv:2308.13561 (2023) 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 403, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 445 + ], + "type": "text", + "content": "55. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=PxTIG12RRHS 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 446, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 490 + ], + "type": "text", + "content": "56. Soran, B., Farhadi, A., Shapiro, L.: Action recognition in the presence of one egocentric and multiple static cameras. In: Computer Vision-ACCV 2014: 12th Asian Conference on Computer Vision, Singapore, Singapore, November 1-5, 2014, Revised Selected Papers, Part V 12. pp. 178-193. Springer (2015) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 491, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 533 + ], + "type": "text", + "content": "57. Tang, H., Xu, D., Sebe, N., Wang, Y., Corso, J.J., Yan, Y.: Multi-channel attention selection gan with cascaded semantic guidance for cross-view image translation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2417-2426 (2019) 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "text", + "content": "58. Tseng, H.Y., Li, Q., Kim, C., Alsisan, S., Huang, J.B., Kopf, J.: Consistent view synthesis with pose-guided diffusion models. arXiv preprint arXiv:2303.17598 (2023) 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "text", + "content": "59. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "type": "text", + "content": "60. Wang, J., Liu, L., Xu, W., Sarkar, K., Luvizon, D., Theobalt, C.: Estimating egocentric 3d human pose in the wild with external weak supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13157-13166 (2022) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "61. Wang, T.C., Liu, M.Y., Zhu, J.Y., Liu, G., Tao, A., Kautz, J., Catanzaro, B.: Video-to-video synthesis. arXiv preprint arXiv:1808.06601 (2018) 3, 4, 8, 10, 12" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "M. Luo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 533 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "type": "text", + "content": "62. Wang, T.C., Liu, M.Y., Zhu, J.Y., Tao, A., Kautz, J., Catanzaro, B.: High-resolution image synthesis and semantic manipulation with conditional gans. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8798-8807 (2018) 3, 4, 6, 8, 10, 12" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 162, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 162, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 130, + 162, + 481, + 193 + ], + "type": "text", + "content": "63. Wang, Z., Wu, S., Xie, W., Chen, M., Prisacariu, V.A.: NeRF--: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064 (2021) 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 225 + ], + "type": "text", + "content": "64. Watson, D., Chan, W., Martin-Brualla, R., Ho, J., Tagliasacchi, A., Norouzi, M.: Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628 (2022) 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "type": "text", + "content": "65. Wen, Y., Singh, K.K., Anderson, M., Jan, W.P., Lee, Y.J.: Seeing the unseen: Predicting the first-person camera wearer's location and pose in third-person scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. pp. 3446-3455 (October 2021) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "text", + "content": "66. Wiles, O., Gkioxari, G., Szeliski, R., Johnson, J.: Synsin: End-to-end view synthesis from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7467-7477 (2020) 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "text", + "content": "67. Xu, M., Fan, C., Wang, Y., Ryoo, M.S., Crandall, D.J.: Joint person segmentation and identification in synchronized first-and third-person videos. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 637-652 (2018) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 337, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 337, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 337, + 481, + 357 + ], + "type": "text", + "content": "68. Xue, Z., Grauman, K.: Learning fine-grained view-invariant representations from unpaired ego-exo videos via temporal alignment. In: NeurIPS (2023) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "type": "text", + "content": "69. Ye, Y., Li, X., Gupta, A., Mello, S.D., Birchfield, S., Song, J., Tulsiani, S., Liu, S.: Affordance diffusion: Synthesizing hand-object interactions. In: CVPR (2023) 10" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 380, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 380, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 133, + 380, + 481, + 423 + ], + "type": "text", + "content": "70. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4578-4587 (June 2021) 2, 3, 4, 5, 10, 12, 14" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 424, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 424, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 133, + 424, + 481, + 456 + ], + "type": "text", + "content": "71. Yu, H., Cai, M., Liu, Y., Lu, F.: What i see is what you see: Joint attention learning for first and third person video co-analysis. In: Proceedings of the 27th ACM International Conference on Multimedia. pp. 1358-1366 (2019) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 457, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 457, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 133, + 457, + 481, + 489 + ], + "type": "text", + "content": "72. Yu, H., Cai, M., Liu, Y., Lu, F.: First-and third-person video co-analysis by learning spatial-temporal joint attention. IEEE Transactions on Pattern Analysis and Machine Intelligence (2020) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 490, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 490, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 133, + 490, + 481, + 511 + ], + "type": "text", + "content": "73. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models (2023) 12" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "text", + "content": "74. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018) 11" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 210, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 210, + 91, + 447, + 102 + ], + "type": "text", + "content": "Lifting the Egocentric Perspective from Exocentric Videos" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Pyramid Diffusion for Fine 3D Large Scene Generation/0615ce43-c6d2-40fd-ba98-d83c475c6139_content_list.json b/2024/Pyramid Diffusion for Fine 3D Large Scene Generation/0615ce43-c6d2-40fd-ba98-d83c475c6139_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0173a94cb312b91d4e631d1f41beaa68c4f35755 --- /dev/null +++ b/2024/Pyramid Diffusion for Fine 3D Large Scene Generation/0615ce43-c6d2-40fd-ba98-d83c475c6139_content_list.json @@ -0,0 +1,1555 @@ +[ + { + "type": "text", + "text": "Pyramid Diffusion for Fine 3D Large Scene Generation", + "text_level": 1, + "bbox": [ + 254, + 140, + 750, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuheng Liu $^{1,2*}$ , Xinke Li $^{3*}$ , Xueting Li $^{4}$ , Lu Qi $^{5\\dagger}$ , Chongshou Li $^{1}$ , Ming-Hsuan Yang $^{5,6}$", + "bbox": [ + 290, + 210, + 709, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Southwest Jiaotong University $^{2}$ University of Leeds $^{3}$ City University of Hong Kong $^{4}$ NVIDIA $^{5}$ The University of California, Merced $^{6}$ Yonsei University", + "bbox": [ + 215, + 253, + 792, + 284 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5af58c26ba66a3ae081c745e97e1401a334775a1449f6987b379e03b46fce0a7.jpg", + "image_caption": [ + "Fig. 1: We present Pyramid Discrete Diffusion Model, a method that progresses from generating coarse- to fine-grained scenes, mirroring the top-down sequence of the pyramid structure shown. The model is extended for cross-dataset and infinite scene generation, with detailed scene intricacies illustrated on the flanking sides of the image. $\\mathcal{D}_s$ and $\\mathcal{D}_t$ refer to a source dataset and a target dataset, respectively." + ], + "image_footnote": [], + "bbox": [ + 235, + 325, + 769, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Diffusion models have shown remarkable results in generating 2D images and small-scale 3D objects. However, their application to the synthesis of large-scale 3D scenes has been rarely explored. This is mainly due to the inherent complexity and bulky size of 3D scenery data, particularly outdoor scenes, and the limited availability of comprehensive real-world datasets, which makes training a stable scene diffusion model challenging. In this work, we explore how to effectively generate large-scale 3D scenes using the coarse-to-fine paradigm. We introduce a framework, the Pyramid Discrete Diffusion model (PDD), which employs scale-varied diffusion models to progressively generate high-quality outdoor scenes. Experimental results of PDD demonstrate", + "bbox": [ + 259, + 648, + 740, + 801 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution.", + "bbox": [ + 232, + 811, + 377, + 824 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 232, + 824, + 395, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "our successful exploration in generating 3D scenes both unconditionally and conditionally. We further showcase the data compatibility of the PDD model, due to its multi-scale architecture: a PDD model trained on one dataset can be easily fine-tuned with another dataset. Code is available at https://github.com/yuhengliu02/pyramid-discrete-diffusion.", + "bbox": [ + 259, + 147, + 743, + 218 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Keywords: 3D Scene Generation $\\cdot$ Diffusion Models $\\cdot$ Transfer Learning", + "bbox": [ + 261, + 229, + 741, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 267, + 375, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D scene generation is the task of creating digital representations that mimic the three-dimensional complexities of our real-world environment, allowing for a more nuanced understanding of the tangible surroundings. This technique plays an essential role in fundamental computer vision tasks such as autonomous driving [21,42,44], virtual reality [30,32,39], and robotic manipulation [8,16,46]. However, high-quality large 3D scenes are extremely challenging to synthesize due to their inherently bulky size, and the lack of large-scale 3D scene datasets [43].", + "bbox": [ + 212, + 296, + 785, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Meanwhile, recent advances in diffusion models have shown impressive results in generating 2D images [35-37] or small-scale 3D objects [24, 33]. Yet, it is not a trivial task to employ diffusion models in 3D scene generation. On one hand, state-of-the-art diffusion models leave substantial memory footprints and demand considerable training time, posing a particular challenge when generating 3D scenes with large scales and intricate details. On the other hand, diffusion models require a large amount of training data [29, 48], while capturing large-scale 3D scenes is itself a challenging and ongoing research topic [4, 23]. As a result, only a few attempts have been made to apply diffusion models directly to 3D outdoor scenes [18], which resulted in unstable generation and thus suboptimal performance.", + "bbox": [ + 212, + 402, + 785, + 568 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To resolve these challenges, existing works focus on conditional generation and resort to additional signals such as Scene Graphs [41] or 2D maps [28] for guidance. Nonetheless, such conditional guidance is not always accessible, thereby restricting the generalizability of these approaches. Inspired by the coarse-to-fine philosophy widely used in image super-resolution [14, 31, 38], we introduce the Pyramid Discrete Diffusion model (PDD), a framework that progressively generates large 3D scenes without relying on additional guidance.", + "bbox": [ + 212, + 568, + 785, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We begin by generating small-scale 3D scenes and progressively increase the scale. At each scale level, we learn a separate diffusion model. This model takes the generated scene from the previous scale as a condition (except for the first diffusion model which takes noise as input) and synthesizes a 3D scene of a larger scale. Intuitively, this multi-scale generation process breaks down a challenging unconditional generation task (i.e., high-quality 3D scene generation) into several more manageable conditional generation tasks. This separation allows each diffusion model to specialize in generating either coarse structure (smaller scale) or intricate details (larger scale). Moreover, at the highest scale, we employ a technique known as scene subdivision, which involves dividing a large scene into multiple smaller segments that are synthesized using a shared diffusion model.", + "bbox": [ + 212, + 672, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This approach mitigates the issues of oversized models caused by the bulky size of the 3D scenes. Additionally, a noteworthy outcome of our multi-scale design is its capacity to facilitate cross-data transfer applications, which substantially reduces training resources. Lastly, we further propose a natural extension of our PDD framework with scene subdivision for infinite 3D scene generation, thereby demonstrating the scalability of the proposed method.", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main contributions of this work are as follows:", + "bbox": [ + 240, + 238, + 607, + 251 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We practically implement a coarse-to-fine strategy for 3D outdoor scene generation via designing a novel pyramid diffusion model.", + "- We conduct extensive experiments on our pyramid diffusion, demonstrating its generation of higher quality 3D scenes with comparable computational resources of existing approaches. In addition, we introduce new metrics to evaluate the quality of 3D scene generation from various perspectives.", + "- Our proposed method showcases broader applications, enabling the generation of scenes from synthetic datasets to real-world data. Furthermore, our approach can be extended to facilitate the creation of infinite scenes." + ], + "bbox": [ + 225, + 260, + 784, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 417, + 387, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Models for 2D Images. Recent advancements in the generative model have seen the diffusion models [14, 31, 38] rise to prominence, especially in applications in 2D image creation [9, 35, 36]. In order to generate high-fidelity images via diffusion models, a multi-stage diffusion process is proposed and employed as per [13, 15, 37]. This process starts with the generation of a coarse-resolution image using an initial diffusion model. Subsequently, a second diffusion model takes this initial output as input, refining it into a finer-resolution image. These cascaded diffusions can be iteratively applied to achieve the desired image resolution. We note that the generation of fine-grained 3D data presents more challenges than 2D due to the addition of an extra dimension. Consequently, our work is motivated by the aforementioned multistage 2D approaches to explore their applicability in 3D contexts. Furthermore, we aim to leverage the advantages of this structure to address the scarcity of datasets in 3D scenes.", + "bbox": [ + 212, + 446, + 787, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Models for 3D Generation. As a sparse and memory-efficient representation, 3D point clouds has been widely used in various computer vision applications such as digital human [27, 40, 49], autonomous driving [20], and 3D scene reconstruction [17]. Point clouds generation aims to synthesize a 3D point clouds from a random noise [5,6], or scanned lidar points [18]. Though the memory efficiency of point clouds is a valuable property, it poses high challenges in the task of point cloud generation. Existing works largely focus on using Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), or Vector Quantized Variational Autoencoders (VQ-VAEs) as the backbone for this task [1,5,6]. However, these models have limited capacity for high-fidelity generation and are notoriously known for unstable training. As an alternative to the generative models discussed above, diffusion models have revolutionized the computer vision community with their impressive performance in 2D image", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Pyramid Diffusion for Fine 3D Large Scene Generation", + "bbox": [ + 364, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/693ab21859dc9bb04e9636a414360c8969b6fae8a9325b8b64c77f92e2d4b569.jpg", + "image_caption": [ + "Fig. 2: Framework of the proposed Pyramid Discrete Diffusion model. In our structure, there are three different scales. Scenes generated by a previous scale can serve as a condition for the current scale after processing through our scale adaptive function. Furthermore, for the final scale processing, the scene from the previous scale is subdivided into four sub-scenes. The final scene is reconstructed into a large scene using our Scene Subdivision module." + ], + "image_footnote": [], + "bbox": [ + 263, + 148, + 769, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "generation [35-37]. Yet, applying diffusion models for 3D point cloud generation has not been thoroughly explored hitherto. Point-Voxel Diffusion [50] proposes to generate a raw point cloud through the diffusion process while LION [47] and DPM [26] use the latent representation of a point cloud during the denoising process. However, all these methods focus on object-level point clouds and cannot be naively extended to scene-level point clouds. Most relevant to our work is [18], where a diffusion model is trained on a scene-level point cloud dataset for the synthesis task. However, due to the capacity limitation of diffusion models, generating a scene-level point cloud with a single diffusion model leads to unsatisfying results, such as undesired wholes or the lack of fine-grained objects. In this work, we propose a pyramid discrete diffusion model that reduces the difficulty at each pyramid level, thus producing scene point clouds with more realistic and fine-grained details.", + "bbox": [ + 212, + 484, + 787, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D Large-scale Scene Generation. Generating large-scale 3D scenes is an important but highly challenging task. A generative model on 3D scenes potentially provides infinite training data for tasks such as scene segmentation, autonomous driving, etc. Existing works [4,22,23,45] simplify this task by first generating 2D scenes and then \"lifting\" them to 3D. Though such design is efficient for city scenes populated with regular geometries (e.g., buildings), it does not generalize easily to scenes with more fine-grained objects (e.g., pedestrians, cars, trees, etc.) In this paper, we directly generate 3D outdoor scenes using diffusion models, which include abundant small objects with semantics.", + "bbox": [ + 212, + 703, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Approach", + "text_level": 1, + "bbox": [ + 215, + 143, + 346, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proposed Pyramid Discrete Diffusion (PDD) model comprises multi-scale models capable of step-by-step generation of high-quality 3D scenes from smaller scales. The PDD first extends the standard discrete diffusion for 3D data (Section 3.2) and then proposes a scene subdivision method to further reduce memory requirements (Section 3.3). Finally, we demonstrate two practical applications of PDD in specific scenarios (Section 3.4).", + "bbox": [ + 212, + 185, + 787, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Discrete Diffusion", + "text_level": 1, + "bbox": [ + 214, + 306, + 410, + 320 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We focus on learning a data distribution based on 3D semantic scenes. Specifically, the semantic scene is represented in a one-hot format, i.e., $\\mathbf{X} \\in \\{0,1\\}^{h \\times w \\times d \\times c}$ , where $h$ , $w$ , and $d$ indicate the dimensions of the scene, respectively, and $c$ denotes the size of the one-hot label.", + "bbox": [ + 212, + 339, + 805, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discrete diffusion [2] has been proposed to generate discrete data including semantic scenes. It involves applying the Markov transition matrix on discrete states for noise diffusion. In the forward process, an original scene $\\mathbf{X}_0$ is gradually corrupted into a $t$ -step noised map $\\mathbf{X}_t$ with $t = 1,\\dots ,T$ . Each forward step can be defined by a Markov uniform transition matrix $\\mathbf{Q}_t$ as $\\mathbf{X}_t = \\mathbf{X}_{t - 1}\\mathbf{Q}_t$ . Based on the Markov property, we can derive the $t$ -step scene $\\mathbf{X}_t$ straight from $\\mathbf{X}_0$ with a cumulative transition matrix $\\bar{\\mathbf{Q}}_t = \\mathbf{Q}_1\\mathbf{Q}_2\\dots \\mathbf{Q}_t$ :", + "bbox": [ + 212, + 401, + 787, + 508 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {X} _ {t} \\mid \\mathbf {X} _ {0}\\right) = \\operatorname {C a t} \\left(\\mathbf {X} _ {t}; \\mathbf {P} = \\mathbf {X} _ {0} \\bar {\\mathbf {Q}} _ {t}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 369, + 521, + 785, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\operatorname{Cat}(\\mathbf{X};\\mathbf{P})$ is a multivariate categorical distribution over the one-hot semantic labels $\\mathbf{X}$ with probabilities given by $\\mathbf{P}$ . Finally, the semantic scene $\\mathbf{X}_T$ at the last step $T$ is supposed to be in the form of a uniform discrete noise. In the reverse process, a learnable model parametrized by $\\theta$ is used to predict denoised semantic labels by $\\tilde{p}_{\\theta}\\left(\\tilde{\\mathbf{X}}_0\\mid \\mathbf{X}_t\\right)$ . The reparametrization trick is applied subsequently to get the reverse process $p_{\\theta}\\left(\\mathbf{X}_{t - 1}\\mid \\mathbf{X}_t\\right)$ :", + "bbox": [ + 214, + 554, + 787, + 654 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\mathbf {X} _ {t - 1} \\mid \\mathbf {X} _ {t}\\right) = \\mathbb {E} _ {\\tilde {p} _ {\\theta} \\left(\\tilde {\\mathbf {X}} _ {0} \\mid \\mathbf {X} _ {t}\\right)} q \\left(\\mathbf {X} _ {t - 1} \\mid \\mathbf {X} _ {t}, \\tilde {\\mathbf {X}} _ {0}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 667, + 785, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A loss consisting of the two KL divergences is proposed to learn better reconstruction ability for the model, given by", + "bbox": [ + 214, + 705, + 785, + 737 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\theta} = d _ {\\mathrm {K L}} \\left(q \\left(\\mathbf {X} _ {t - 1} \\mid \\mathbf {X} _ {t}, \\mathbf {X} _ {0}\\right) \\| p _ {\\theta} \\left(\\mathbf {X} _ {t - 1} \\mid \\mathbf {X} _ {t}\\right)\\right) \\tag {3} \\\\ + \\lambda d _ {\\mathrm {K L}} \\left(q \\left(\\mathbf {X} _ {0}\\right) \\| \\tilde {p} _ {\\theta} \\left(\\tilde {\\mathbf {X}} _ {0} \\mid \\mathbf {X} _ {t}\\right)\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 343, + 751, + 784, + 796 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is an auxiliary loss weight and $d_{\\mathrm{KL}}$ stands for KL divergence. In the following, we focus on extending the discrete diffusion into the proposed PDD.", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Pyramid Diffusion for Fine 3D Large Scene Generation", + "bbox": [ + 362, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Pyramid Discrete Diffusion", + "text_level": 1, + "bbox": [ + 215, + 146, + 486, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We propose PDD that operates various diffusion processes across multiple scales (or resolutions), as depicted in Figure 2. Given a 3D scene data $\\mathbf{Z} \\in \\{0,1\\}^{h \\times w \\times d \\times c}$ , we define a 3D pyramid including different scales of $\\mathbf{Z}$ , i.e., $\\{\\mathbf{Z}^{(1)}, \\dots, \\mathbf{Z}^{(l)}, \\dots, \\mathbf{Z}^{(L)}\\}$ , where a larger $l$ indicates a larger scene scale. Formally, let $h_l \\times w_l \\times d_l \\times c$ denote the dimension of $\\mathbf{Z}^{(l)}$ , $h_{l+1} \\geq h_l$ , $w_{l+1} \\geq w_l$ and $d_{l+1} \\geq d_l$ are kept for $l = 1, \\dots, L-1$ . We note that such a pyramid can be obtained by applying different down-sample operators, such as pooling functions, on $\\mathbf{Z}$ . For each scale in the pyramid, we construct a conditional discrete diffusion model parameterized by $\\theta_l$ . The $l$ -th model for $l \\neq 1$ is given by:", + "bbox": [ + 212, + 172, + 812, + 309 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {p} _ {\\theta_ {l}} \\left(\\tilde {\\mathbf {X}} _ {0} ^ {(l)} \\mid \\mathbf {X} _ {t} ^ {(l)}, \\mathbf {Z} ^ {(l - 1)}\\right) = \\tilde {p} _ {\\theta_ {l}} \\left(\\tilde {\\mathbf {X}} _ {0} ^ {(l)} \\mid \\operatorname {C o n c a t} \\left(\\mathbf {X} _ {t} ^ {(l)}, \\phi^ {(l)} \\left(\\mathbf {Z} ^ {(l - 1)}\\right)\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 264, + 319, + 785, + 345 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{X}_t^{(l)}$ and $\\mathbf{X}_0^{(l)}$ are with the same size of $\\mathbf{Z}^{(l)}$ , and $\\phi^{(l)}$ is a Scale Adaptive Function (SAF) for upsampling $\\mathbf{Z}^{(l-1)}$ into the size of $\\mathbf{Z}^{(l)}$ . As a case in point, SAF can be a tri-linear interpolation function depending on the data. Additionally, we maintain the first model $\\tilde{p}_{\\theta_1}$ as the original non-conditional model.", + "bbox": [ + 212, + 357, + 784, + 419 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During the training process, PDD learns $L$ denoising models separately at varied scales of scene pyramids in the given dataset. Given that $\\mathbf{Z}^{(l - 1)}$ is essentially a lossy-compressed version of $\\mathbf{Z}^{(l)}$ , the model training can be viewed as learning to restore the details of a coarse scene. In the inference process, denoising model $p_{\\theta_1}$ is performed initially according to Equation (2) and the rest of PDD models are executed in sequence from $l = 2$ to $L$ via the sampling,", + "bbox": [ + 212, + 420, + 785, + 511 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {t - 1} ^ {(l)} \\sim p _ {\\theta_ {l}} (\\mathbf {X} _ {t - 1} ^ {(l)} \\mid \\mathbf {X} _ {t} ^ {(l)}, \\mathbf {X} _ {0} ^ {(l - 1)}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 522, + 784, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{X}_0^{(l - 1)}$ is the denoised result of $\\tilde{p}_{\\theta_{l - 1}}$ .", + "bbox": [ + 212, + 554, + 532, + 573 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Except for the high-quality generation, the proposed PDD bears two merits: 1) Diffusion models in PDD can be trained in parallel due to their independence, which allows for a flexible computation reallocation during training. 2) Due to its multi-stage generation process, PDD is fitting for restoring scenes of arbitrary coarse-grained scale by starting from the intermediate processes, thereby extending the method's versatility.", + "bbox": [ + 212, + 573, + 785, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Scene Subdivision", + "text_level": 1, + "bbox": [ + 215, + 686, + 410, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To overcome the memory constraint for generating large 3D scenes, we propose the scene subdivision method. We divide a 3D scene $\\mathbf{Z}^{(l)}$ along $z$ -axis into $I$ overlapped sub-components as $\\{\\mathbf{Z}_i^{(l)}\\}_{i=1}^I$ . For the instance of four subscenes case, let $\\mathbf{Z}_i^{(l)} \\in \\{0,1\\}^{(1 + \\delta_l)h_l \\setminus 2 \\times (1 + \\delta_l)w_l \\setminus 2 \\times d_l \\times c}$ denote one subscene and $\\delta_l$ denote the overlap ratio, the shared $l$ -th diffusion model in PDD is trained to reconstruct $\\mathbf{Z}_i^{(l)}$ for $i = 1, \\dots, 4$ . Subsequently, sub-scenes are merged into a holistic one by a fusion algorithm, i.e., voting on the overlapped parts to ensure the continuity of the 3D scene.", + "bbox": [ + 212, + 712, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the training process, to ensure context-awareness of the entire scene during the generation of a sub-scene, we train the model by adding the overlapped regions with other sub-scenes as the condition. In the inference process, the entire scene is generated in an autoregressive manner. Apart from the first sub-scene generated without context, all other sub-scenes utilize the already generated overlapped region as a condition, i.e.,", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {t - 1, i} ^ {(l)} \\sim p _ {\\theta} \\left(\\mathbf {X} _ {t - 1, i} ^ {(l)} \\mid \\mathbf {X} _ {t, i} ^ {(l)}, \\mathbf {X} _ {0, i} ^ {(l + 1)}, \\sum_ {j \\neq i} \\Delta_ {i j} \\odot \\mathbf {X} _ {0, j} ^ {(l + 1)}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 303, + 247, + 785, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $j$ is the index of generated sub-scenes before $i$ -th scene, and $\\Delta_{ij}$ is a binary mask between $\\mathbf{X}_{0,i}^{(l+1)}$ and $\\mathbf{X}_{0,j}^{(l+1)}$ representing the overlapped region on $\\mathbf{X}_{0,j}^{(l+1)}$ with 1 and the separate region with 0. Scene Subdivision module can reduce the model parameters, as diffusion model could be shared by four sub-scenes. In practice, we only implement the scene subdivision method on the largest scale which demands the largest memory.", + "bbox": [ + 212, + 306, + 787, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Applications", + "text_level": 1, + "bbox": [ + 214, + 425, + 366, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Beyond its primary function as a generative model, we introduce two novel applications for PDD. First, cross-dataset transfer aims at adapting a model trained on a source dataset to a target dataset [51]. Due to the flexibility of input scale, PDD can achieve this by retraining or fine-tuning the smaller-scale models in the new dataset while keeping the larger-scale models. The strategy leveraging PDD improves the efficiency of transferring 3D scene generation models between distinct datasets. Second, infinite scene generation is of great interest in fields such as autonomous driving [10] and urban modeling [19] which require a huge scale of 3D scenes. PDD can extend its scene subdivision technique. By using the edge of a previously generated scene as a condition as in Equation (6), it can iteratively create larger scenes, potentially without size limitations.", + "bbox": [ + 212, + 452, + 787, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experimental Results", + "text_level": 1, + "bbox": [ + 214, + 643, + 460, + 660 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Evaluation Protocols", + "text_level": 1, + "bbox": [ + 214, + 676, + 434, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since the metrics used in 2D generation such as FID [12] are not directly applicable in the 3D, we introduce and implement three metrics to assess the quality of the generated 3D scenes. We note that more implementation details can be found in the supplementary material.", + "bbox": [ + 212, + 703, + 785, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Semantic Segmentation results on the generated scenes are used to evaluate the effectiveness of models in creating semantically coherent scenes. Specifically, two architectures, the voxel-based SparseUNet [11] and point-based PointNet++ [34], are implemented to perform the segmentation tasks. We report the mean Intersection over Union (mIoU) and Mean Accuracy (MAs) for evaluation.", + "bbox": [ + 212, + 765, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Pyramid Diffusion for Fine 3D Large Scene Generation", + "bbox": [ + 362, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8d71a08cf6599b15b29865540ce8c0696b32b5aac939bb43fe3711927a106fea.jpg", + "table_caption": [ + "Table 1: Comparison of various diffusion models on 3D semantic scene generation of CarlaSC. DiscreteDiff [2], LatentDiff [18], and P-DiscreteDiff refer to the original discrete diffusion, latent discrete diffusion, and our approach, respectively. Conditioned models work based on the context of unlabeled point clouds or the coarse version of the ground truth scene. A higher Segmentation Metric value is better, indicating semantic consistency. A lower Feature-based Metric value is preferable, representing closer proximity to the original dataset. The brackets with $V$ represent voxel-based network and $P$ represent point-based network." + ], + "table_footnote": [], + "table_body": "
MethodModelConditionSegmentation MetricFeature-based Metric
mIoU (V)MA (V)mIoU (P)MA (P)F3D (↓)MMD (↓)
Ground Truth--52.1972.4032.9047.680.2460.108
UnconditionedDiscreteDiff [2]-40.0563.6525.5438.711.3610.599
LatentDiff [18]-38.0162.3926.6945.870.3310.211
P-DiscreteDiff (Ours)-68.0285.6633.8952.120.3150.200
ConditionedDiscreteDiff [2]Point cloud38.5559.9728.4144.060.3570.261
DiscreteDiff [2]Coarse scene (s1)52.5277.2327.9343.130.3590.284
P-DiscreteDiff (Ours)Coarse scene (s1)55.7578.7029.7846.610.3420.274
", + "bbox": [ + 222, + 267, + 781, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "F3D is a 3D adaption of the 2D Fréchet Inception Distance (FID) [12], which is based on a pre-trained autoencoder with an 3D CNN architecture. We calculate and report the Fréchet distance (by $10^{-3}$ ratio) between the generated scenes and real scenes in the feature domain.", + "bbox": [ + 212, + 405, + 782, + 463 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Maximum Mean Discrepancy (MMD) is a statistical measure to quantify the disparity between the distributions of generated and real scenes. Similar to our F3D approach, we extract features via the same pre-trained autoencoder and present the MMD between 3D scenes.", + "bbox": [ + 212, + 465, + 784, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Experiment Settings", + "text_level": 1, + "bbox": [ + 214, + 546, + 431, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets. We use CarlaSC [43] and SemanticKITTI [3] for experiments. Specifically, we conduct our main experiments as well as ablation studies on the synthesis dataset CarlaSC due to its large data volume and diverse semantic objects. Our primary model is trained on the training set of CarlaSC with 10 categories and 32,400 scans. SemanticKITTI, which is a real-world collected dataset, is used for our cross-dataset transfer experiment. Both datasets are adjusted to ensure consistency in semantic categories, with further details in the supplementary material.", + "bbox": [ + 212, + 566, + 787, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Model Architecture. The primary proposed PDD is performed on three scales of a 3D scene pyramid, i.e., $s_1$ , $s_2$ and $s_4$ in Table 2. We implement 3D-UNets [7] for three diffusion models in PDD based on the scales. Notably, the model applied on $s_4$ scale is with the input/output size of $s_3'$ due to the use of scene subdivision, while such a size of other models follows the working scale size. In the ablation study, we also introduce the scale $s_3$ in the experiment. Additionally, we implement two baseline methods merely on scale $s_4$ which are the original discrete diffusion [2] and the latent diffusion model with VQ-VAE decoder [18]. Training Setting. We train each PDD model using the same training setting except for the batch size. Specifically, we set the learning rate of $10^{-3}$ for the", + "bbox": [ + 212, + 688, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d358629d294c2cd5e888458f112762bb982265cb04b46ba63a0d54739bf4ffbb.jpg", + "table_caption": [ + "Table 2: Different scales in the 3D scene pyramid." + ], + "table_footnote": [], + "table_body": "
Scale Rep.3D Scene Size
s132 × 32 × 4
s264 × 64 × 8
s3128 × 128 × 8
s3'136 × 136 × 16
s4256 × 256 × 16
", + "bbox": [ + 233, + 186, + 457, + 287 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/949ccdf9d92e95db62fe13634daf2cb6e6b01fd6387ffe1cf0540fa60dd0e219.jpg", + "table_caption": [ + "Table 3: Average SSIM between each generated scene and the closest scene in the training set. We generate $1k$ scenes and calculate the average." + ], + "table_footnote": [], + "table_body": "
DatasetMethodSSIM
CarlaSCUncon-Gen0.72
Val Set0.65
Sem-KITTI (fine-tuned)Uncon-Gen0.74
Val Set0.67
", + "bbox": [ + 535, + 213, + 761, + 289 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/b0c2d6dbedfaf808d9adf43f1f810d82d04c842fa8cce266824042b56a500d50.jpg", + "table_caption": [ + "Table 4: Comparison of different diffusion pyramids on 3D semantic scene generation." + ], + "table_footnote": [], + "table_body": "
PyramidConditionedmIoU(V)mIoU(P)F3D(↓)MMD(↓)
s4×40.025.51.360.60
s1 → s4×67.032.10.320.24
s1 → s2 → s4×68.033.90.320.20
s1 → s2 → s3 → s4×68.033.40.320.23
s1 → s4✓52.527.90.360.28
s1 → s2 → s4✓55.829.80.340.27
s1 → s2 → s3 → s4✓55.929.60.340.28
", + "bbox": [ + 253, + 327, + 759, + 449 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "AdamW optimizer [25], and the time step $T = 100$ for the diffusion process, and 800 for the max epoch. The batch sizes are set to 128, 32, and 16 for the models working on $s_1$ , $s_2$ and $s_4$ scales. However, for the baseline method based on the $s_4$ scale, we use the batch size of 8 due to memory constraints. We note that all diffusion models are trained on four NVIDIA A100 GPUs. In addition, we apply the trilinear interpolation for the scene fusion algorithm and set the overlap ratio in scene subdivision, $\\delta_l$ to 0.0625.", + "bbox": [ + 212, + 476, + 787, + 582 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Main Results", + "text_level": 1, + "bbox": [ + 215, + 604, + 372, + 618 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Generation Quality. We compare our approach with two baselines, the original Discrete Diffusion [2] and the Latent Diffusion [18]. The result reported in Table 1 demonstrates the notable performance of our method across all metrics in both unconditional and conditional settings in comparable computational resources with existing method. Our proposed method demonstrates a notable advantage in segmentation tasks, especially when it reaches around $70\\%$ mIoU for SparseUNet, which reflects its ability to generate scenes with accurate semantic coherence. We also provide visualizations of different model results in Figure 3, where the proposed method demonstrates better performance in detail generation and scene diversity for random 3D scene generations.", + "bbox": [ + 212, + 628, + 785, + 779 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Additionally, we conduct the comparison on conditioned 3D scene generation. We leverage the flexibility of input scale for our method and perform the generation by models in $s_2$ and $s_4$ scales conditioned on a coarse ground truth scene in $s_1$ scale. We benchmark our method against the discrete diffusion conditioned on", + "bbox": [ + 214, + 779, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Pyramid Diffusion for Fine 3D Large Scene Generation", + "bbox": [ + 364, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/2ffbc795e494692ef7ddfed08133aed7acbcedabc4b3b4c6782e176a5570ee7e.jpg", + "image_caption": [ + "Fig. 3: Visualization of unconditional generation results on CarlaSC. We compare with two baseline models - DiscreteDiff [2] and LatentDiff [18] and show synthesis from our models with different scales. Our method produces more diverse scenes compared to the baseline models. Furthermore, with more levels, our model can synthesize scenes with more intricate details." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 781, + 436 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "unlabeled point clouds and the same coarse scenes. Results in Table 1 and Figure 5 present the impressive results of our conditional generation comparison. It is also observed that the point cloud-based model can achieve decent performance on F3D and MMD, which could be caused by 3D point conditions providing more structural information about the scene than the coarse scene. Despite the informative condition of the point cloud, our method can still outperform it across most metrics.", + "bbox": [ + 212, + 563, + 787, + 667 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "None-overfitting Verification. The MMD and F3D metrics (defined in Supp A) numerically illustrate the statistical feature distance between generated scenes and the training set. Our method achieves the lowest MMD and F3D among all baseline methods as shown in Table 1. However, we argue that this does not indicate overfitting to the dataset for the following reasons. First, our MMD and F3D are larger than those of the ground truth. Furthermore, we leverage structural similarity (SSIM) to search and show that a generated scene is different from its nearest neighbour in the dataset. Specifically, we generate $1k$ scenes and identify their closest matches in the training set using the SSIM metric. The average SSIM of these $1k$ scenes are calculated and presented in Table 3. Additionally, we apply the same methodology to the Validation Set to establish an", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/30a115586a27a16e20586cab35c9c1d25857903e10df245797c525847bb69ac4.jpg", + "image_caption": [ + "Fig. 4: Data retrieval visualization. We generate $1k$ scenes using PDD on CarlaSC dataset, and retrieve the most similar scene in training set for each scene using SSIM (SSIM = 1 means identical), and plot SSIM distribution. Scenes at various percentiles are displayed (red box: generated scenes; grey box: scenes in training set), those with the highest $10\\%$ similarity (i.e., $10^{th}$ percentile of SSIM) are very similar to the training set, but still not completely identical." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 782, + 300 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "oracle baseline. Table 3 shows that our generated scenes are comparable to the baseline, verifying that our method does not overfit the training set. To further support this, we use distribution plots, as shown in Figure 4, to validate the similarity between our generated scenes and the training set. We also display three pairs at different percentiles in this figure, showing that scenes with lower SSIM scores differ more from their nearest matches in the training set. This visual evidence reinforces that PDD effectively captures the distribution of the training set instead of merely memorizing it.", + "bbox": [ + 212, + 425, + 787, + 546 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 570, + 401, + 584 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Pyramid Diffusion. Our experiments explore the impact of varying refinement scales on the quality of generated scenes. According to Table 4, both conditional and unconditional scene generation quality show incremental improvements with additional scales. Balancing training overhead and generation quality, a three-scale model with the scale of $s_4$ progression offers an optimal compromise between performance and computational cost. We find that as the number of scales increases, there is indeed a rise in performance, particularly notable upon the addition of the second scale. However, the progression from a three-scale pyramid to a four-scale pyramid turns out to be insignificant. Given the greater training overhead for a four-scale pyramid compared to a three-scale one, we choose the latter as our main structure.", + "bbox": [ + 212, + 597, + 787, + 762 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Scene Subdivision. We explore the optimal mask ratio for scene subdivision and report on Figure 6, which shows an inverse correlation between the mask ratio and the effectiveness of F3D and MMD metrics; higher mask ratios may result in diminished outcomes due to increased overfitting, leading to reduced randomness in the generated results. The lowest mask ratio test, 0.0625, achieves", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Pyramid Diffusion for Fine 3D Large Scene Generation", + "bbox": [ + 364, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f53f936d2f764098f8fd2f4fd7f1a400309d691717a0b8406959e452efb13d20.jpg", + "image_caption": [ + "Fig. 5: Visualization of conditional generation results on CarlaSC. $PC$ stands for point cloud condition." + ], + "image_footnote": [], + "bbox": [ + 218, + 142, + 781, + 378 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/4a50b5423862a924bcb80b7dab4e0943e917c004752b808de0ee149fe234f444.jpg", + "table_caption": [ + "Table 5: Generation results on CarlaSC in different scales on the diffusion pyramid without any conditions. All output scales are lifted to $s_4$ using the upsampling method." + ], + "table_footnote": [], + "table_body": "
Model No.ScalemIoU(↑)MA(↑)F3D(↓)MMD(↓)
1s118.042.70.290.16
2s243.766.80.290.18
3s468.085.70.320.23
", + "bbox": [ + 272, + 489, + 730, + 565 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "the best results across all metrics with detailed retention. Thus, we set a mask ratio of 0.0625 as the standard for our scene subdivision module. Further analysis shows that higher overlap ratios in scene subdivision result in quality deterioration, mainly due to increased discontinuities when merging sub-scenes using scene fusion algorithm.", + "bbox": [ + 212, + 606, + 787, + 681 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Performance on Different Scales. To facilitate a comprehensive understanding of the progressive improvement achieved by our coarse-to-fine method, we evaluate the quality of scenes generated at different scales and present the findings in Table 5. Even at the smaller scale $s_1$ , we observe high F3D and MMD scores, indicating its capacity in synthesizing scenes with both reasonable and diverse layouts. As we advance to larger scales (i.e., $s_2$ and $s_3$ ), the mIoU and MA scores consistently increase, demonstrating that our model focuses on learning intricate details in later stages. Meanwhile, F3D and MMD metrics show stability without significant decline, indicating a balanced enhancement in scene complexity and fidelity.", + "bbox": [ + 212, + 689, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ac7c9dfae1a2fafb3f1efba87450d49c37bad5888a7860936e42a98afda9eda4.jpg", + "image_caption": [ + "3D and MMD Scores by Mask Ratio" + ], + "image_footnote": [], + "bbox": [ + 220, + 160, + 500, + 284 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/4a15fe1a66cd5979ff88a19b843fc32f6df1fa5f94aa9a0229a80a69e9440908.jpg", + "image_caption": [ + "(b) mIoU and OA Scores by Mask Ratio" + ], + "image_footnote": [], + "bbox": [ + 514, + 161, + 779, + 284 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/21ddd5bca6b2f442adbb717ad015fa2738375dc9eb941e6df3647863f1193c8e.jpg", + "image_caption": [ + "Fig. 6: Effects of mask ratio on unconditional generation results.", + "Fig. 7: Infinite Scene Generation. Thanks to the pyramid representation, PDD can be readily applied for unbounded scene generation. This involves the initial efficient synthesis of a large-scale coarse 3D scene, followed by subsequent refinement at higher levels." + ], + "image_footnote": [], + "bbox": [ + 225, + 325, + 779, + 460 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Applications", + "text_level": 1, + "bbox": [ + 215, + 560, + 366, + 574 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Cross-dataset. Figure 8 and Figure 9 showcase our model's performance on the transferred dataset from CarlaSC to SemanticKITTI for both unconditional and conditional scene generation. The PDD shows enhanced scene quality after finetuning with SemanticKITTI data, as indicated by improved results in Table 6. This fine-tuning effectively adapts the model to the dataset's complex object distributions and scene dynamics. Notably, despite the higher training effort of the Discrete Diffusion (DD) approach, our method outperforms DD even without fine-tuning, using only coarse scenes from SemanticKITTI. This demonstrates the strong cross-data transfer capability of our approach.", + "bbox": [ + 212, + 583, + 787, + 718 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Infinite Scene Generation. Figure 7 visualizes the process of generating large-scale infinite scenes using our PDD model. As discussed in Section 3.4, we first use the small-scale model to swiftly generate a coarse infinite 3D scene (bottom level in Figure 7). We then leverage larger-scale models to progressively add intricate details (middle and top level in Figure 7), enhancing realism. This approach enables our model to produce high-quality, continuous cityscapes without additional inputs, overcoming the limitations of conventional datasets with finite scenes and supporting downstream tasks like 3D scene segmentation.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Pyramid Diffusion for Fine 3D Large Scene Generation", + "bbox": [ + 364, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9926aa00e7c5b83b87c912a7f776e89cf588dfadbeb85dc1dac0d09d9597482a.jpg", + "image_caption": [ + "Fig. 8: SemanticKITTI unconditional generation. $FT$ stands for finetuning pretrained model from CarlaSC." + ], + "image_footnote": [], + "bbox": [ + 225, + 146, + 480, + 342 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/22d576bfcc5a8c80fcebf39cfdc082b22581abcade1d53d2fd59527249ee997b.jpg", + "image_caption": [ + "Fig.9: SemanticKITTI conditional generation. $FT$ stands for finetuning from CarlaSC models." + ], + "image_footnote": [], + "bbox": [ + 517, + 146, + 777, + 344 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/2410f20815ce02b9c45fd25ac6c2b81e421c65f7569d37b62cdf4f1d12e7277b.jpg", + "table_caption": [ + "Table 6: Generation results on SemanticKITTI. Finetuned Scales set to None indicates training from scratch and others stand for finetuning corresponding pre-trained CarlaSC model." + ], + "table_footnote": [], + "table_body": "
MethodFinetuned ScalesConditionedmIoU(V)mIoU(P)F3D(↓)MMD(↓)
DD [2]s4×29.116.00.460.31
PDDNone✓33.422.80.270.32
PDDs2, s4✓43.922.80.280.16
PDDs1×31.323.20.220.13
PDDs1, s2, s4×44.724.20.210.11
", + "bbox": [ + 228, + 468, + 782, + 566 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 593, + 359, + 609 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we introduce the Pyramid Discrete Diffusion model (PDD) to address the significant challenges associated with 3D large scene generation, particularly in the context of limited scale and available datasets. The PDD demonstrates a progressive approach to generating high-quality 3D outdoor scenes, transitioning seamlessly from coarse to fine details. Compared to the other methods, the PDD can generate high-quality scenes within limited resource constraints and does not require additional data sources. Our extensive experimental results show that PDD consistently performs favourably in both unconditional and conditional generation tasks, establishing itself as a reliable and robust solution for the creation of realistic and intricate scenes. Furthermore, the proposed PDD method has great potential in efficiently adapting models trained on synthetic data to real-world datasets and suggests a promising solution to the current challenge of limited real-world data.", + "bbox": [ + 212, + 625, + 787, + 821 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Supported by the Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/ Interior Business Center (DOI/IBC) contract number 140D0423C0074. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of IARPA, DOI/IBC, or the U.S. Government. Y. Liu and C. Li are supported in part by the National Science Foundation of China (Grant No: 62202395).", + "bbox": [ + 212, + 174, + 787, + 310 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 332, + 321, + 348 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Anvekar, T., Tabib, R.A., Hegde, D., Mudengudi, U.: Vg-vae: A venatus geometry point-cloud variational auto-encoder. In: CVPR (2022)", + "2. Austin, J., Johnson, D.D., Ho, J., Tarlow, D., Van Den Berg, R.: Structured denoising diffusion models in discrete state-spaces. In: NeurIPS (2021)", + "3. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: ICCV (2019)", + "4. Chen, Z., Wang, G., Liu, Z.: Scenedreamer: Unbounded 3d scene generation from 2d image collections. In: arXiv preprint arXiv: 2302.01330 (2023)", + "5. Cheng, A.C., Li, X., Liu, S., Sun, M., Yang, M.H.: Learning 3d dense correspondence via canonical point autoencoder. In: ECCV (2022)", + "6. Cheng, A.C., Li, X., Sun, M., Yang, M.H., Liu, S.: Learning 3d dense correspondence via canonical point autoencoder. In: NeurIPS (2021)", + "7. Çiçek, Ö., Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3d unet: learning dense volumetric segmentation from sparse annotation. In: MICCAI (2016)", + "8. Cong, Y., Chen, R., Ma, B., Liu, H., Hou, D., Yang, C.: A comprehensive study of 3-d vision-based robot manipulation. IEEE Transactions on Cybernetics (2021)", + "9. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: NeurIPS (2021)", + "0. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: CVPR (2012)", + "1. Graham, B., Engelcke, M., Van Der Maaten, L.: 3d semantic segmentation with submanifold sparse convolutional networks. In: CVPR (2018)", + "2. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017)", + "3. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)", + "4. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)", + "5. Ho, J., Saharia, C., Chan, W., Fleet, D.J., Norouzi, M., Salimans, T.: Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research (2022)" + ], + "bbox": [ + 225, + 361, + 784, + 838 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Pyramid Diffusion for Fine 3D Large Scene Generation", + "bbox": [ + 364, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Huang, W., Wang, C., Zhang, R., Li, Y., Wu, J., Fei-Fei, L.: Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973 (2023)", + "17. Lan, Z., Yew, Z.J., Lee, G.H.: Robust point cloud based reconstruction of large-scale outdoor scenes. In: CVPR (2019)", + "18. Lee, J., Im, W., Lee, S., Yoon, S.E.: Diffusion probabilistic models for scene-scale 3d categorical data. arXiv preprint arXiv:2301.00527 (2023)", + "19. Li, X., Li, C., Tong, Z., Lim, A., Yuan, J., Wu, Y., Tang, J., Huang, R.: Campus3d: A photogrammetry point cloud benchmark for hierarchical understanding of outdoor scene. In: ACM MM (2020)", + "20. Li, Y., Ma, L., Zhong, Z., Liu, F., Chapman, M.A., Cao, D., Li, J.: Deep learning for lidar point clouds in autonomous driving: A review. NeurIPS (2020)", + "21. Li, Y., Yu, A.W., Meng, T., Caine, B., Ngiam, J., Peng, D., Shen, J., Lu, Y., Zhou, D., Le, Q.V., et al.: Deepfusion: Lidar-camera deep fusion for multi-modal 3d object detection. In: CVPR (2022)", + "22. Li, Z., Wang, Q., Snavely, N., Kanazawa, A.: Infinitenature-zero: Learning perpetual view generation of natural scenes from single images. In: ECCV (2022)", + "23. Lin, C.H., Lee, H.Y., Menapace, W., Chai, M., Siarohin, A., Yang, M.H., Tulyakov, S.: Infinicity: Infinite-scale city synthesis. arXiv preprint arXiv:2301.09637 (2023)", + "24. Liu, M., Shi, R., Chen, L., Zhang, Z., Xu, C., Wei, X., Chen, H., Zeng, C., Gu, J., Su, H.: One-2-3-45++: Fast single image to 3d objects with consistent multi-view generation and 3d diffusion. arXiv preprint arXiv:2311.07885 (2023)", + "25. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)", + "26. Luo, S., Hu, W.: Diffusion probabilistic models for 3d point cloud generation. In: CVPR (2021)", + "27. Ma, Q., Yang, J., Tang, S., Black, M.J.: The power of points for modeling humans in clothing. In: ICCV (2021)", + "28. Mascaro, R., Teixeira, L., Chli, M.: Diffuser: Multi-view 2d-to-3d label diffusion for semantic scene segmentation. In: ICRA (2021)", + "29. Moon, T., Choi, M., Lee, G., Ha, J.W., Lee, J.: Fine-tuning diffusion models with limited data. In: NeurIPS 2022 Workshop on Score-Based Methods (2022)", + "30. Moro, S., Komuro, T.: Generation of virtual reality environment based on 3d scanned indoor physical space. In: ISVC (2021)", + "31. Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: ICML (2021)", + "32. Ögün, M.N., Kurul, R., Yasar, M.F., Turkoglu, S.A., Avci, Ş., Yildiz, N.: Effect of leap motion-based 3d immersive virtual reality usage on upper extremity function in ischemic stroke patients. Arquivos de neuro-psiquiatria (2019)", + "33. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv (2022)", + "34. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: NeurIPS (2017)", + "35. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)", + "36. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)", + "37. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "38. Sohl-Dickstein, J., Weiss, E.A., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. arXiv preprint arXiv:1503.03585 (2015)", + "39. Sra, M., Garrido-Jurado, S., Schmandt, C., Maes, P.: Procedurally generated virtual reality from 3d reconstructed physical space. In: Proceedings of the 22nd ACM Conference on Virtual Reality Software and Technology (2016)", + "40. Su, S.Y., Bagautdinov, T., Rhodin, H.: Npc: Neural point characters from video. In: ICCV (2023)", + "41. Tang, J., Nie, Y., Markhasin, L., Dai, A., Thies, J., Nießner, M.: Diffuscene: Scene graph denoising diffusion probabilistic model for generative indoor scene synthesis. arXiv preprint arXiv:2303.14207 (2023)", + "42. Tang, Y., He, H., Wang, Y., Mao, Z., Wang, H.: Multi-modality 3d object detection in autonomous driving: A review. Neurocomputing (2023)", + "43. Wilson, J., Song, J., Fu, Y., Zhang, A., Capodieci, A., Jayakumar, P., Barton, K., Ghaffari, M.: Motionsc: Data set and network for real-time semantic mapping in dynamic environments. IEEE Robotics and Automation Letters 7(3) (2022)", + "44. Wu, P., Jia, X., Chen, L., Yan, J., Li, H., Qiao, Y.: Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. In: NeurIPS (2022)", + "45. Xie, H., Chen, Z., Hong, F., Liu, Z.: Citydreamer: Compositional generative model of unbounded 3d cities. arXiv preprint arXiv:2309.00610 (2023)", + "46. Xu, Z., He, Z., Wu, J., Song, S.: Learning 3d dynamic scene representations for robot manipulation. arXiv preprint arXiv:2011.01968 (2020)", + "47. Zeng, X., Vahdat, A., Williams, F., Gojcic, Z., Litany, O., Fidler, S., Kreis, K.: Lion: Latent point diffusion models for 3d shape generation. In: NeurIPS (2022)", + "48. Zheng, H., Nie, W., Vahdat, A., Anandkumar, A.: Fast training of diffusion models with masked transformers. arXiv preprint arXiv:2306.09305 (2023)", + "49. Zheng, Y., Yifan, W., Wetzstein, G., Black, M.J., Hilliges, O.: Pointavatar: Deformable point-based head avatars from videos. In: CVPR (2023)", + "50. Zhou, L., Du, Y., Wu, J.: 3d shape generation and completion through point-voxel diffusion. In: ICCV (2021)", + "51. Zhuang, F., Qi, Z., Duan, K., Xi, D., Zhu, Y., Zhu, H., Xiong, H., He, Q.: A comprehensive survey on transfer learning. arXiv preprint arXiv:1911.02685 (2020)" + ], + "bbox": [ + 212, + 146, + 787, + 604 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Pyramid Diffusion for Fine 3D Large Scene Generation", + "bbox": [ + 364, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Pyramid Diffusion for Fine 3D Large Scene Generation/0615ce43-c6d2-40fd-ba98-d83c475c6139_model.json b/2024/Pyramid Diffusion for Fine 3D Large Scene Generation/0615ce43-c6d2-40fd-ba98-d83c475c6139_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3053a1ee299377e34c3f8a40109cf808ce2eb038 --- /dev/null +++ b/2024/Pyramid Diffusion for Fine 3D Large Scene Generation/0615ce43-c6d2-40fd-ba98-d83c475c6139_model.json @@ -0,0 +1,2225 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.255, + 0.141, + 0.75, + 0.185 + ], + "angle": 0, + "content": "Pyramid Diffusion for Fine 3D Large Scene Generation" + }, + { + "type": "text", + "bbox": [ + 0.292, + 0.212, + 0.71, + 0.245 + ], + "angle": 0, + "content": "Yuheng Liu\\(^{1,2*}\\), Xinke Li\\(^{3*}\\), Xueting Li\\(^{4}\\), Lu Qi\\(^{5\\dagger}\\), Chongshou Li\\(^{1}\\), Ming-Hsuan Yang\\(^{5,6}\\)" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.255, + 0.793, + 0.285 + ], + "angle": 0, + "content": "\\(^{1}\\) Southwest Jiaotong University \\(^{2}\\) University of Leeds \\(^{3}\\) City University of Hong Kong \\(^{4}\\) NVIDIA \\(^{5}\\) The University of California, Merced \\(^{6}\\) Yonsei University" + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.327, + 0.771, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.521, + 0.788, + 0.592 + ], + "angle": 0, + "content": "Fig. 1: We present Pyramid Discrete Diffusion Model, a method that progresses from generating coarse- to fine-grained scenes, mirroring the top-down sequence of the pyramid structure shown. The model is extended for cross-dataset and infinite scene generation, with detailed scene intricacies illustrated on the flanking sides of the image. \\(\\mathcal{D}_s\\) and \\(\\mathcal{D}_t\\) refer to a source dataset and a target dataset, respectively." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.649, + 0.741, + 0.802 + ], + "angle": 0, + "content": "Abstract. Diffusion models have shown remarkable results in generating 2D images and small-scale 3D objects. However, their application to the synthesis of large-scale 3D scenes has been rarely explored. This is mainly due to the inherent complexity and bulky size of 3D scenery data, particularly outdoor scenes, and the limited availability of comprehensive real-world datasets, which makes training a stable scene diffusion model challenging. In this work, we explore how to effectively generate large-scale 3D scenes using the coarse-to-fine paradigm. We introduce a framework, the Pyramid Discrete Diffusion model (PDD), which employs scale-varied diffusion models to progressively generate high-quality outdoor scenes. Experimental results of PDD demonstrate" + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.812, + 0.378, + 0.825 + ], + "angle": 0, + "content": "* Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.825, + 0.396, + 0.84 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.233, + 0.812, + 0.396, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.148, + 0.744, + 0.219 + ], + "angle": 0, + "content": "our successful exploration in generating 3D scenes both unconditionally and conditionally. We further showcase the data compatibility of the PDD model, due to its multi-scale architecture: a PDD model trained on one dataset can be easily fine-tuned with another dataset. Code is available at https://github.com/yuhengliu02/pyramid-discrete-diffusion." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.23, + 0.743, + 0.246 + ], + "angle": 0, + "content": "Keywords: 3D Scene Generation \\(\\cdot\\) Diffusion Models \\(\\cdot\\) Transfer Learning" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.268, + 0.377, + 0.284 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.297, + 0.787, + 0.403 + ], + "angle": 0, + "content": "3D scene generation is the task of creating digital representations that mimic the three-dimensional complexities of our real-world environment, allowing for a more nuanced understanding of the tangible surroundings. This technique plays an essential role in fundamental computer vision tasks such as autonomous driving [21,42,44], virtual reality [30,32,39], and robotic manipulation [8,16,46]. However, high-quality large 3D scenes are extremely challenging to synthesize due to their inherently bulky size, and the lack of large-scale 3D scene datasets [43]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.787, + 0.569 + ], + "angle": 0, + "content": "Meanwhile, recent advances in diffusion models have shown impressive results in generating 2D images [35-37] or small-scale 3D objects [24, 33]. Yet, it is not a trivial task to employ diffusion models in 3D scene generation. On one hand, state-of-the-art diffusion models leave substantial memory footprints and demand considerable training time, posing a particular challenge when generating 3D scenes with large scales and intricate details. On the other hand, diffusion models require a large amount of training data [29, 48], while capturing large-scale 3D scenes is itself a challenging and ongoing research topic [4, 23]. As a result, only a few attempts have been made to apply diffusion models directly to 3D outdoor scenes [18], which resulted in unstable generation and thus suboptimal performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.787, + 0.674 + ], + "angle": 0, + "content": "To resolve these challenges, existing works focus on conditional generation and resort to additional signals such as Scene Graphs [41] or 2D maps [28] for guidance. Nonetheless, such conditional guidance is not always accessible, thereby restricting the generalizability of these approaches. Inspired by the coarse-to-fine philosophy widely used in image super-resolution [14, 31, 38], we introduce the Pyramid Discrete Diffusion model (PDD), a framework that progressively generates large 3D scenes without relying on additional guidance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.841 + ], + "angle": 0, + "content": "We begin by generating small-scale 3D scenes and progressively increase the scale. At each scale level, we learn a separate diffusion model. This model takes the generated scene from the previous scale as a condition (except for the first diffusion model which takes noise as input) and synthesizes a 3D scene of a larger scale. Intuitively, this multi-scale generation process breaks down a challenging unconditional generation task (i.e., high-quality 3D scene generation) into several more manageable conditional generation tasks. This separation allows each diffusion model to specialize in generating either coarse structure (smaller scale) or intricate details (larger scale). Moreover, at the highest scale, we employ a technique known as scene subdivision, which involves dividing a large scene into multiple smaller segments that are synthesized using a shared diffusion model." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.365, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Pyramid Diffusion for Fine 3D Large Scene Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "This approach mitigates the issues of oversized models caused by the bulky size of the 3D scenes. Additionally, a noteworthy outcome of our multi-scale design is its capacity to facilitate cross-data transfer applications, which substantially reduces training resources. Lastly, we further propose a natural extension of our PDD framework with scene subdivision for infinite 3D scene generation, thereby demonstrating the scalability of the proposed method." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.239, + 0.608, + 0.252 + ], + "angle": 0, + "content": "The main contributions of this work are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.261, + 0.784, + 0.29 + ], + "angle": 0, + "content": "- We practically implement a coarse-to-fine strategy for 3D outdoor scene generation via designing a novel pyramid diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.291, + 0.785, + 0.35 + ], + "angle": 0, + "content": "- We conduct extensive experiments on our pyramid diffusion, demonstrating its generation of higher quality 3D scenes with comparable computational resources of existing approaches. In addition, we introduce new metrics to evaluate the quality of 3D scene generation from various perspectives." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.351, + 0.784, + 0.395 + ], + "angle": 0, + "content": "- Our proposed method showcases broader applications, enabling the generation of scenes from synthetic datasets to real-world data. Furthermore, our approach can be extended to facilitate the creation of infinite scenes." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.261, + 0.785, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.418, + 0.388, + 0.433 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.644 + ], + "angle": 0, + "content": "Diffusion Models for 2D Images. Recent advancements in the generative model have seen the diffusion models [14, 31, 38] rise to prominence, especially in applications in 2D image creation [9, 35, 36]. In order to generate high-fidelity images via diffusion models, a multi-stage diffusion process is proposed and employed as per [13, 15, 37]. This process starts with the generation of a coarse-resolution image using an initial diffusion model. Subsequently, a second diffusion model takes this initial output as input, refining it into a finer-resolution image. These cascaded diffusions can be iteratively applied to achieve the desired image resolution. We note that the generation of fine-grained 3D data presents more challenges than 2D due to the addition of an extra dimension. Consequently, our work is motivated by the aforementioned multistage 2D approaches to explore their applicability in 3D contexts. Furthermore, we aim to leverage the advantages of this structure to address the scarcity of datasets in 3D scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Diffusion Models for 3D Generation. As a sparse and memory-efficient representation, 3D point clouds has been widely used in various computer vision applications such as digital human [27, 40, 49], autonomous driving [20], and 3D scene reconstruction [17]. Point clouds generation aims to synthesize a 3D point clouds from a random noise [5,6], or scanned lidar points [18]. Though the memory efficiency of point clouds is a valuable property, it poses high challenges in the task of point cloud generation. Existing works largely focus on using Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), or Vector Quantized Variational Autoencoders (VQ-VAEs) as the backbone for this task [1,5,6]. However, these models have limited capacity for high-fidelity generation and are notoriously known for unstable training. As an alternative to the generative models discussed above, diffusion models have revolutionized the computer vision community with their impressive performance in 2D image" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.149, + 0.77, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.414 + ], + "angle": 0, + "content": "Fig. 2: Framework of the proposed Pyramid Discrete Diffusion model. In our structure, there are three different scales. Scenes generated by a previous scale can serve as a condition for the current scale after processing through our scale adaptive function. Furthermore, for the final scale processing, the scene from the previous scale is subdivided into four sub-scenes. The final scene is reconstructed into a large scene using our Scene Subdivision module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.789, + 0.683 + ], + "angle": 0, + "content": "generation [35-37]. Yet, applying diffusion models for 3D point cloud generation has not been thoroughly explored hitherto. Point-Voxel Diffusion [50] proposes to generate a raw point cloud through the diffusion process while LION [47] and DPM [26] use the latent representation of a point cloud during the denoising process. However, all these methods focus on object-level point clouds and cannot be naively extended to scene-level point clouds. Most relevant to our work is [18], where a diffusion model is trained on a scene-level point cloud dataset for the synthesis task. However, due to the capacity limitation of diffusion models, generating a scene-level point cloud with a single diffusion model leads to unsatisfying results, such as undesired wholes or the lack of fine-grained objects. In this work, we propose a pyramid discrete diffusion model that reduces the difficulty at each pyramid level, thus producing scene point clouds with more realistic and fine-grained details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.84 + ], + "angle": 0, + "content": "3D Large-scale Scene Generation. Generating large-scale 3D scenes is an important but highly challenging task. A generative model on 3D scenes potentially provides infinite training data for tasks such as scene segmentation, autonomous driving, etc. Existing works [4,22,23,45] simplify this task by first generating 2D scenes and then \"lifting\" them to 3D. Though such design is efficient for city scenes populated with regular geometries (e.g., buildings), it does not generalize easily to scenes with more fine-grained objects (e.g., pedestrians, cars, trees, etc.) In this paper, we directly generate 3D outdoor scenes using diffusion models, which include abundant small objects with semantics." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.364, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Pyramid Diffusion for Fine 3D Large Scene Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.347, + 0.163 + ], + "angle": 0, + "content": "3 Approach" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.186, + 0.788, + 0.277 + ], + "angle": 0, + "content": "The proposed Pyramid Discrete Diffusion (PDD) model comprises multi-scale models capable of step-by-step generation of high-quality 3D scenes from smaller scales. The PDD first extends the standard discrete diffusion for 3D data (Section 3.2) and then proposes a scene subdivision method to further reduce memory requirements (Section 3.3). Finally, we demonstrate two practical applications of PDD in specific scenarios (Section 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.307, + 0.411, + 0.321 + ], + "angle": 0, + "content": "3.1 Discrete Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.806, + 0.4 + ], + "angle": 0, + "content": "We focus on learning a data distribution based on 3D semantic scenes. Specifically, the semantic scene is represented in a one-hot format, i.e., \\(\\mathbf{X} \\in \\{0,1\\}^{h \\times w \\times d \\times c}\\), where \\(h\\), \\(w\\), and \\(d\\) indicate the dimensions of the scene, respectively, and \\(c\\) denotes the size of the one-hot label." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.789, + 0.509 + ], + "angle": 0, + "content": "Discrete diffusion [2] has been proposed to generate discrete data including semantic scenes. It involves applying the Markov transition matrix on discrete states for noise diffusion. In the forward process, an original scene \\(\\mathbf{X}_0\\) is gradually corrupted into a \\(t\\)-step noised map \\(\\mathbf{X}_t\\) with \\(t = 1,\\dots ,T\\). Each forward step can be defined by a Markov uniform transition matrix \\(\\mathbf{Q}_t\\) as \\(\\mathbf{X}_t = \\mathbf{X}_{t - 1}\\mathbf{Q}_t\\). Based on the Markov property, we can derive the \\(t\\)-step scene \\(\\mathbf{X}_t\\) straight from \\(\\mathbf{X}_0\\) with a cumulative transition matrix \\(\\bar{\\mathbf{Q}}_t = \\mathbf{Q}_1\\mathbf{Q}_2\\dots \\mathbf{Q}_t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.37, + 0.522, + 0.787, + 0.543 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {X} _ {t} \\mid \\mathbf {X} _ {0}\\right) = \\operatorname {C a t} \\left(\\mathbf {X} _ {t}; \\mathbf {P} = \\mathbf {X} _ {0} \\bar {\\mathbf {Q}} _ {t}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.555, + 0.788, + 0.655 + ], + "angle": 0, + "content": "where \\(\\operatorname{Cat}(\\mathbf{X};\\mathbf{P})\\) is a multivariate categorical distribution over the one-hot semantic labels \\(\\mathbf{X}\\) with probabilities given by \\(\\mathbf{P}\\). Finally, the semantic scene \\(\\mathbf{X}_T\\) at the last step \\(T\\) is supposed to be in the form of a uniform discrete noise. In the reverse process, a learnable model parametrized by \\(\\theta\\) is used to predict denoised semantic labels by \\(\\tilde{p}_{\\theta}\\left(\\tilde{\\mathbf{X}}_0\\mid \\mathbf{X}_t\\right)\\). The reparametrization trick is applied subsequently to get the reverse process \\(p_{\\theta}\\left(\\mathbf{X}_{t - 1}\\mid \\mathbf{X}_t\\right)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.329, + 0.668, + 0.787, + 0.695 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\mathbf {X} _ {t - 1} \\mid \\mathbf {X} _ {t}\\right) = \\mathbb {E} _ {\\tilde {p} _ {\\theta} \\left(\\tilde {\\mathbf {X}} _ {0} \\mid \\mathbf {X} _ {t}\\right)} q \\left(\\mathbf {X} _ {t - 1} \\mid \\mathbf {X} _ {t}, \\tilde {\\mathbf {X}} _ {0}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.706, + 0.787, + 0.738 + ], + "angle": 0, + "content": "A loss consisting of the two KL divergences is proposed to learn better reconstruction ability for the model, given by" + }, + { + "type": "equation", + "bbox": [ + 0.344, + 0.752, + 0.785, + 0.797 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\theta} = d _ {\\mathrm {K L}} \\left(q \\left(\\mathbf {X} _ {t - 1} \\mid \\mathbf {X} _ {t}, \\mathbf {X} _ {0}\\right) \\| p _ {\\theta} \\left(\\mathbf {X} _ {t - 1} \\mid \\mathbf {X} _ {t}\\right)\\right) \\tag {3} \\\\ + \\lambda d _ {\\mathrm {K L}} \\left(q \\left(\\mathbf {X} _ {0}\\right) \\| \\tilde {p} _ {\\theta} \\left(\\tilde {\\mathbf {X}} _ {0} \\mid \\mathbf {X} _ {t}\\right)\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is an auxiliary loss weight and \\(d_{\\mathrm{KL}}\\) stands for KL divergence. In the following, we focus on extending the discrete diffusion into the proposed PDD." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.488, + 0.163 + ], + "angle": 0, + "content": "3.2 Pyramid Discrete Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.813, + 0.31 + ], + "angle": 0, + "content": "We propose PDD that operates various diffusion processes across multiple scales (or resolutions), as depicted in Figure 2. Given a 3D scene data \\(\\mathbf{Z} \\in \\{0,1\\}^{h \\times w \\times d \\times c}\\), we define a 3D pyramid including different scales of \\(\\mathbf{Z}\\), i.e., \\(\\{\\mathbf{Z}^{(1)}, \\dots, \\mathbf{Z}^{(l)}, \\dots, \\mathbf{Z}^{(L)}\\}\\), where a larger \\(l\\) indicates a larger scene scale. Formally, let \\(h_l \\times w_l \\times d_l \\times c\\) denote the dimension of \\(\\mathbf{Z}^{(l)}\\), \\(h_{l+1} \\geq h_l\\), \\(w_{l+1} \\geq w_l\\) and \\(d_{l+1} \\geq d_l\\) are kept for \\(l = 1, \\dots, L-1\\). We note that such a pyramid can be obtained by applying different down-sample operators, such as pooling functions, on \\(\\mathbf{Z}\\). For each scale in the pyramid, we construct a conditional discrete diffusion model parameterized by \\(\\theta_l\\). The \\(l\\)-th model for \\(l \\neq 1\\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.266, + 0.32, + 0.786, + 0.346 + ], + "angle": 0, + "content": "\\[\n\\tilde {p} _ {\\theta_ {l}} \\left(\\tilde {\\mathbf {X}} _ {0} ^ {(l)} \\mid \\mathbf {X} _ {t} ^ {(l)}, \\mathbf {Z} ^ {(l - 1)}\\right) = \\tilde {p} _ {\\theta_ {l}} \\left(\\tilde {\\mathbf {X}} _ {0} ^ {(l)} \\mid \\operatorname {C o n c a t} \\left(\\mathbf {X} _ {t} ^ {(l)}, \\phi^ {(l)} \\left(\\mathbf {Z} ^ {(l - 1)}\\right)\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.785, + 0.42 + ], + "angle": 0, + "content": "where \\(\\mathbf{X}_t^{(l)}\\) and \\(\\mathbf{X}_0^{(l)}\\) are with the same size of \\(\\mathbf{Z}^{(l)}\\), and \\(\\phi^{(l)}\\) is a Scale Adaptive Function (SAF) for upsampling \\(\\mathbf{Z}^{(l-1)}\\) into the size of \\(\\mathbf{Z}^{(l)}\\). As a case in point, SAF can be a tri-linear interpolation function depending on the data. Additionally, we maintain the first model \\(\\tilde{p}_{\\theta_1}\\) as the original non-conditional model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.787, + 0.512 + ], + "angle": 0, + "content": "During the training process, PDD learns \\(L\\) denoising models separately at varied scales of scene pyramids in the given dataset. Given that \\(\\mathbf{Z}^{(l - 1)}\\) is essentially a lossy-compressed version of \\(\\mathbf{Z}^{(l)}\\), the model training can be viewed as learning to restore the details of a coarse scene. In the inference process, denoising model \\(p_{\\theta_1}\\) is performed initially according to Equation (2) and the rest of PDD models are executed in sequence from \\(l = 2\\) to \\(L\\) via the sampling," + }, + { + "type": "equation", + "bbox": [ + 0.387, + 0.523, + 0.785, + 0.543 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {t - 1} ^ {(l)} \\sim p _ {\\theta_ {l}} (\\mathbf {X} _ {t - 1} ^ {(l)} \\mid \\mathbf {X} _ {t} ^ {(l)}, \\mathbf {X} _ {0} ^ {(l - 1)}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.555, + 0.534, + 0.574 + ], + "angle": 0, + "content": "where \\(\\mathbf{X}_0^{(l - 1)}\\) is the denoised result of \\(\\tilde{p}_{\\theta_{l - 1}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.574, + 0.787, + 0.664 + ], + "angle": 0, + "content": "Except for the high-quality generation, the proposed PDD bears two merits: 1) Diffusion models in PDD can be trained in parallel due to their independence, which allows for a flexible computation reallocation during training. 2) Due to its multi-stage generation process, PDD is fitting for restoring scenes of arbitrary coarse-grained scale by starting from the intermediate processes, thereby extending the method's versatility." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.687, + 0.411, + 0.701 + ], + "angle": 0, + "content": "3.3 Scene Subdivision" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.713, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To overcome the memory constraint for generating large 3D scenes, we propose the scene subdivision method. We divide a 3D scene \\(\\mathbf{Z}^{(l)}\\) along \\(z\\)-axis into \\(I\\) overlapped sub-components as \\(\\{\\mathbf{Z}_i^{(l)}\\}_{i=1}^I\\). For the instance of four subscenes case, let \\(\\mathbf{Z}_i^{(l)} \\in \\{0,1\\}^{(1 + \\delta_l)h_l \\setminus 2 \\times (1 + \\delta_l)w_l \\setminus 2 \\times d_l \\times c}\\) denote one subscene and \\(\\delta_l\\) denote the overlap ratio, the shared \\(l\\)-th diffusion model in PDD is trained to reconstruct \\(\\mathbf{Z}_i^{(l)}\\) for \\(i = 1, \\dots, 4\\). Subsequently, sub-scenes are merged into a holistic one by a fusion algorithm, i.e., voting on the overlapped parts to ensure the continuity of the 3D scene." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.364, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Pyramid Diffusion for Fine 3D Large Scene Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "In the training process, to ensure context-awareness of the entire scene during the generation of a sub-scene, we train the model by adding the overlapped regions with other sub-scenes as the condition. In the inference process, the entire scene is generated in an autoregressive manner. Apart from the first sub-scene generated without context, all other sub-scenes utilize the already generated overlapped region as a condition, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.305, + 0.248, + 0.786, + 0.299 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {t - 1, i} ^ {(l)} \\sim p _ {\\theta} \\left(\\mathbf {X} _ {t - 1, i} ^ {(l)} \\mid \\mathbf {X} _ {t, i} ^ {(l)}, \\mathbf {X} _ {0, i} ^ {(l + 1)}, \\sum_ {j \\neq i} \\Delta_ {i j} \\odot \\mathbf {X} _ {0, j} ^ {(l + 1)}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.308, + 0.788, + 0.404 + ], + "angle": 0, + "content": "where \\( j \\) is the index of generated sub-scenes before \\( i \\)-th scene, and \\( \\Delta_{ij} \\) is a binary mask between \\( \\mathbf{X}_{0,i}^{(l+1)} \\) and \\( \\mathbf{X}_{0,j}^{(l+1)} \\) representing the overlapped region on \\( \\mathbf{X}_{0,j}^{(l+1)} \\) with 1 and the separate region with 0. Scene Subdivision module can reduce the model parameters, as diffusion model could be shared by four sub-scenes. In practice, we only implement the scene subdivision method on the largest scale which demands the largest memory." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.426, + 0.367, + 0.442 + ], + "angle": 0, + "content": "3.4 Applications" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.453, + 0.788, + 0.62 + ], + "angle": 0, + "content": "Beyond its primary function as a generative model, we introduce two novel applications for PDD. First, cross-dataset transfer aims at adapting a model trained on a source dataset to a target dataset [51]. Due to the flexibility of input scale, PDD can achieve this by retraining or fine-tuning the smaller-scale models in the new dataset while keeping the larger-scale models. The strategy leveraging PDD improves the efficiency of transferring 3D scene generation models between distinct datasets. Second, infinite scene generation is of great interest in fields such as autonomous driving [10] and urban modeling [19] which require a huge scale of 3D scenes. PDD can extend its scene subdivision technique. By using the edge of a previously generated scene as a condition as in Equation (6), it can iteratively create larger scenes, potentially without size limitations." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.644, + 0.462, + 0.661 + ], + "angle": 0, + "content": "4 Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.677, + 0.436, + 0.692 + ], + "angle": 0, + "content": "4.1 Evaluation Protocols" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.787, + 0.764 + ], + "angle": 0, + "content": "Since the metrics used in 2D generation such as FID [12] are not directly applicable in the 3D, we introduce and implement three metrics to assess the quality of the generated 3D scenes. We note that more implementation details can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Semantic Segmentation results on the generated scenes are used to evaluate the effectiveness of models in creating semantically coherent scenes. Specifically, two architectures, the voxel-based SparseUNet [11] and point-based PointNet++ [34], are implemented to perform the segmentation tasks. We report the mean Intersection over Union (mIoU) and Mean Accuracy (MAs) for evaluation." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.256 + ], + "angle": 0, + "content": "Table 1: Comparison of various diffusion models on 3D semantic scene generation of CarlaSC. DiscreteDiff [2], LatentDiff [18], and P-DiscreteDiff refer to the original discrete diffusion, latent discrete diffusion, and our approach, respectively. Conditioned models work based on the context of unlabeled point clouds or the coarse version of the ground truth scene. A higher Segmentation Metric value is better, indicating semantic consistency. A lower Feature-based Metric value is preferable, representing closer proximity to the original dataset. The brackets with \\(V\\) represent voxel-based network and \\(P\\) represent point-based network." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.268, + 0.782, + 0.381 + ], + "angle": 0, + "content": "
MethodModelConditionSegmentation MetricFeature-based Metric
mIoU (V)MA (V)mIoU (P)MA (P)F3D (↓)MMD (↓)
Ground Truth--52.1972.4032.9047.680.2460.108
UnconditionedDiscreteDiff [2]-40.0563.6525.5438.711.3610.599
LatentDiff [18]-38.0162.3926.6945.870.3310.211
P-DiscreteDiff (Ours)-68.0285.6633.8952.120.3150.200
ConditionedDiscreteDiff [2]Point cloud38.5559.9728.4144.060.3570.261
DiscreteDiff [2]Coarse scene (s1)52.5277.2327.9343.130.3590.284
P-DiscreteDiff (Ours)Coarse scene (s1)55.7578.7029.7846.610.3420.274
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.784, + 0.464 + ], + "angle": 0, + "content": "F3D is a 3D adaption of the 2D Fréchet Inception Distance (FID) [12], which is based on a pre-trained autoencoder with an 3D CNN architecture. We calculate and report the Fréchet distance (by \\(10^{-3}\\) ratio) between the generated scenes and real scenes in the feature domain." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.466, + 0.785, + 0.526 + ], + "angle": 0, + "content": "Maximum Mean Discrepancy (MMD) is a statistical measure to quantify the disparity between the distributions of generated and real scenes. Similar to our F3D approach, we extract features via the same pre-trained autoencoder and present the MMD between 3D scenes." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.547, + 0.433, + 0.563 + ], + "angle": 0, + "content": "4.2 Experiment Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.688 + ], + "angle": 0, + "content": "Datasets. We use CarlaSC [43] and SemanticKITTI [3] for experiments. Specifically, we conduct our main experiments as well as ablation studies on the synthesis dataset CarlaSC due to its large data volume and diverse semantic objects. Our primary model is trained on the training set of CarlaSC with 10 categories and 32,400 scans. SemanticKITTI, which is a real-world collected dataset, is used for our cross-dataset transfer experiment. Both datasets are adjusted to ensure consistency in semantic categories, with further details in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Model Architecture. The primary proposed PDD is performed on three scales of a 3D scene pyramid, i.e., \\( s_1 \\), \\( s_2 \\) and \\( s_4 \\) in Table 2. We implement 3D-UNets [7] for three diffusion models in PDD based on the scales. Notably, the model applied on \\( s_4 \\) scale is with the input/output size of \\( s_3' \\) due to the use of scene subdivision, while such a size of other models follows the working scale size. In the ablation study, we also introduce the scale \\( s_3 \\) in the experiment. Additionally, we implement two baseline methods merely on scale \\( s_4 \\) which are the original discrete diffusion [2] and the latent diffusion model with VQ-VAE decoder [18]. Training Setting. We train each PDD model using the same training setting except for the batch size. Specifically, we set the learning rate of \\( 10^{-3} \\) for the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.365, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Pyramid Diffusion for Fine 3D Large Scene Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.157, + 0.474, + 0.185 + ], + "angle": 0, + "content": "Table 2: Different scales in the 3D scene pyramid." + }, + { + "type": "table", + "bbox": [ + 0.235, + 0.187, + 0.458, + 0.289 + ], + "angle": 0, + "content": "
Scale Rep.3D Scene Size
s132 × 32 × 4
s264 × 64 × 8
s3128 × 128 × 8
s3'136 × 136 × 16
s4256 × 256 × 16
" + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.157, + 0.787, + 0.213 + ], + "angle": 0, + "content": "Table 3: Average SSIM between each generated scene and the closest scene in the training set. We generate \\(1k\\) scenes and calculate the average." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.214, + 0.763, + 0.29 + ], + "angle": 0, + "content": "
DatasetMethodSSIM
CarlaSCUncon-Gen0.72
Val Set0.65
Sem-KITTI (fine-tuned)Uncon-Gen0.74
Val Set0.67
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.302, + 0.784, + 0.317 + ], + "angle": 0, + "content": "Table 4: Comparison of different diffusion pyramids on 3D semantic scene generation." + }, + { + "type": "table", + "bbox": [ + 0.254, + 0.328, + 0.761, + 0.45 + ], + "angle": 0, + "content": "
PyramidConditionedmIoU(V)mIoU(P)F3D(↓)MMD(↓)
s4×40.025.51.360.60
s1 → s4×67.032.10.320.24
s1 → s2 → s4×68.033.90.320.20
s1 → s2 → s3 → s4×68.033.40.320.23
s1 → s4✓52.527.90.360.28
s1 → s2 → s4✓55.829.80.340.27
s1 → s2 → s3 → s4✓55.929.60.340.28
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.788, + 0.583 + ], + "angle": 0, + "content": "AdamW optimizer [25], and the time step \\( T = 100 \\) for the diffusion process, and 800 for the max epoch. The batch sizes are set to 128, 32, and 16 for the models working on \\( s_1 \\), \\( s_2 \\) and \\( s_4 \\) scales. However, for the baseline method based on the \\( s_4 \\) scale, we use the batch size of 8 due to memory constraints. We note that all diffusion models are trained on four NVIDIA A100 GPUs. In addition, we apply the trilinear interpolation for the scene fusion algorithm and set the overlap ratio in scene subdivision, \\( \\delta_l \\) to 0.0625." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.605, + 0.373, + 0.619 + ], + "angle": 0, + "content": "4.3 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.787, + 0.78 + ], + "angle": 0, + "content": "Generation Quality. We compare our approach with two baselines, the original Discrete Diffusion [2] and the Latent Diffusion [18]. The result reported in Table 1 demonstrates the notable performance of our method across all metrics in both unconditional and conditional settings in comparable computational resources with existing method. Our proposed method demonstrates a notable advantage in segmentation tasks, especially when it reaches around \\(70\\%\\) mIoU for SparseUNet, which reflects its ability to generate scenes with accurate semantic coherence. We also provide visualizations of different model results in Figure 3, where the proposed method demonstrates better performance in detail generation and scene diversity for random 3D scene generations." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Additionally, we conduct the comparison on conditioned 3D scene generation. We leverage the flexibility of input scale for our method and perform the generation by models in \\( s_2 \\) and \\( s_4 \\) scales conditioned on a coarse ground truth scene in \\( s_1 \\) scale. We benchmark our method against the discrete diffusion conditioned on" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.782, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.453, + 0.788, + 0.523 + ], + "angle": 0, + "content": "Fig. 3: Visualization of unconditional generation results on CarlaSC. We compare with two baseline models - DiscreteDiff [2] and LatentDiff [18] and show synthesis from our models with different scales. Our method produces more diverse scenes compared to the baseline models. Furthermore, with more levels, our model can synthesize scenes with more intricate details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.669 + ], + "angle": 0, + "content": "unlabeled point clouds and the same coarse scenes. Results in Table 1 and Figure 5 present the impressive results of our conditional generation comparison. It is also observed that the point cloud-based model can achieve decent performance on F3D and MMD, which could be caused by 3D point conditions providing more structural information about the scene than the coarse scene. Despite the informative condition of the point cloud, our method can still outperform it across most metrics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "None-overfitting Verification. The MMD and F3D metrics (defined in Supp A) numerically illustrate the statistical feature distance between generated scenes and the training set. Our method achieves the lowest MMD and F3D among all baseline methods as shown in Table 1. However, we argue that this does not indicate overfitting to the dataset for the following reasons. First, our MMD and F3D are larger than those of the ground truth. Furthermore, we leverage structural similarity (SSIM) to search and show that a generated scene is different from its nearest neighbour in the dataset. Specifically, we generate \\(1k\\) scenes and identify their closest matches in the training set using the SSIM metric. The average SSIM of these \\(1k\\) scenes are calculated and presented in Table 3. Additionally, we apply the same methodology to the Validation Set to establish an" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.365, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Pyramid Diffusion for Fine 3D Large Scene Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.144, + 0.784, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.311, + 0.788, + 0.396 + ], + "angle": 0, + "content": "Fig. 4: Data retrieval visualization. We generate \\(1k\\) scenes using PDD on CarlaSC dataset, and retrieve the most similar scene in training set for each scene using SSIM (SSIM = 1 means identical), and plot SSIM distribution. Scenes at various percentiles are displayed (red box: generated scenes; grey box: scenes in training set), those with the highest \\(10\\%\\) similarity (i.e., \\(10^{th}\\) percentile of SSIM) are very similar to the training set, but still not completely identical." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.426, + 0.788, + 0.547 + ], + "angle": 0, + "content": "oracle baseline. Table 3 shows that our generated scenes are comparable to the baseline, verifying that our method does not overfit the training set. To further support this, we use distribution plots, as shown in Figure 4, to validate the similarity between our generated scenes and the training set. We also display three pairs at different percentiles in this figure, showing that scenes with lower SSIM scores differ more from their nearest matches in the training set. This visual evidence reinforces that PDD effectively captures the distribution of the training set instead of merely memorizing it." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.571, + 0.402, + 0.585 + ], + "angle": 0, + "content": "4.4 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.788, + 0.763 + ], + "angle": 0, + "content": "Pyramid Diffusion. Our experiments explore the impact of varying refinement scales on the quality of generated scenes. According to Table 4, both conditional and unconditional scene generation quality show incremental improvements with additional scales. Balancing training overhead and generation quality, a three-scale model with the scale of \\( s_4 \\) progression offers an optimal compromise between performance and computational cost. We find that as the number of scales increases, there is indeed a rise in performance, particularly notable upon the addition of the second scale. However, the progression from a three-scale pyramid to a four-scale pyramid turns out to be insignificant. Given the greater training overhead for a four-scale pyramid compared to a three-scale one, we choose the latter as our main structure." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Scene Subdivision. We explore the optimal mask ratio for scene subdivision and report on Figure 6, which shows an inverse correlation between the mask ratio and the effectiveness of F3D and MMD metrics; higher mask ratios may result in diminished outcomes due to increased overfitting, leading to reduced randomness in the generated results. The lowest mask ratio test, 0.0625, achieves" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.143, + 0.782, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.394, + 0.788, + 0.421 + ], + "angle": 0, + "content": "Fig. 5: Visualization of conditional generation results on CarlaSC. \\(PC\\) stands for point cloud condition." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.449, + 0.788, + 0.478 + ], + "angle": 0, + "content": "Table 5: Generation results on CarlaSC in different scales on the diffusion pyramid without any conditions. All output scales are lifted to \\( s_4 \\) using the upsampling method." + }, + { + "type": "table", + "bbox": [ + 0.274, + 0.49, + 0.731, + 0.566 + ], + "angle": 0, + "content": "
Model No.ScalemIoU(↑)MA(↑)F3D(↓)MMD(↓)
1s118.042.70.290.16
2s243.766.80.290.18
3s468.085.70.320.23
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.607, + 0.788, + 0.682 + ], + "angle": 0, + "content": "the best results across all metrics with detailed retention. Thus, we set a mask ratio of 0.0625 as the standard for our scene subdivision module. Further analysis shows that higher overlap ratios in scene subdivision result in quality deterioration, mainly due to increased discontinuities when merging sub-scenes using scene fusion algorithm." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Performance on Different Scales. To facilitate a comprehensive understanding of the progressive improvement achieved by our coarse-to-fine method, we evaluate the quality of scenes generated at different scales and present the findings in Table 5. Even at the smaller scale \\( s_1 \\), we observe high F3D and MMD scores, indicating its capacity in synthesizing scenes with both reasonable and diverse layouts. As we advance to larger scales (i.e., \\( s_2 \\) and \\( s_3 \\)), the mIoU and MA scores consistently increase, demonstrating that our model focuses on learning intricate details in later stages. Meanwhile, F3D and MMD metrics show stability without significant decline, indicating a balanced enhancement in scene complexity and fidelity." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.365, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Pyramid Diffusion for Fine 3D Large Scene Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.147, + 0.476, + 0.159 + ], + "angle": 0, + "content": "3D and MMD Scores by Mask Ratio" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.161, + 0.501, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.555, + 0.147, + 0.756, + 0.159 + ], + "angle": 0, + "content": "(b) mIoU and OA Scores by Mask Ratio" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.162, + 0.78, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.284, + 0.299, + 0.719, + 0.314 + ], + "angle": 0, + "content": "Fig. 6: Effects of mask ratio on unconditional generation results." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.327, + 0.78, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.476, + 0.788, + 0.531 + ], + "angle": 0, + "content": "Fig. 7: Infinite Scene Generation. Thanks to the pyramid representation, PDD can be readily applied for unbounded scene generation. This involves the initial efficient synthesis of a large-scale coarse 3D scene, followed by subsequent refinement at higher levels." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.561, + 0.367, + 0.575 + ], + "angle": 0, + "content": "4.5 Applications" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.719 + ], + "angle": 0, + "content": "Cross-dataset. Figure 8 and Figure 9 showcase our model's performance on the transferred dataset from CarlaSC to SemanticKITTI for both unconditional and conditional scene generation. The PDD shows enhanced scene quality after finetuning with SemanticKITTI data, as indicated by improved results in Table 6. This fine-tuning effectively adapts the model to the dataset's complex object distributions and scene dynamics. Notably, despite the higher training effort of the Discrete Diffusion (DD) approach, our method outperforms DD even without fine-tuning, using only coarse scenes from SemanticKITTI. This demonstrates the strong cross-data transfer capability of our approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Infinite Scene Generation. Figure 7 visualizes the process of generating large-scale infinite scenes using our PDD model. As discussed in Section 3.4, we first use the small-scale model to swiftly generate a coarse infinite 3D scene (bottom level in Figure 7). We then leverage larger-scale models to progressively add intricate details (middle and top level in Figure 7), enhancing realism. This approach enables our model to produce high-quality, continuous cityscapes without additional inputs, overcoming the limitations of conventional datasets with finite scenes and supporting downstream tasks like 3D scene segmentation." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.147, + 0.482, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.358, + 0.493, + 0.399 + ], + "angle": 0, + "content": "Fig. 8: SemanticKITTI unconditional generation. \\(FT\\) stands for finetuning pretrained model from CarlaSC." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.147, + 0.778, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.358, + 0.788, + 0.399 + ], + "angle": 0, + "content": "Fig.9: SemanticKITTI conditional generation. \\(FT\\) stands for finetuning from CarlaSC models." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.414, + 0.789, + 0.456 + ], + "angle": 0, + "content": "Table 6: Generation results on SemanticKITTI. Finetuned Scales set to None indicates training from scratch and others stand for finetuning corresponding pre-trained CarlaSC model." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.469, + 0.784, + 0.567 + ], + "angle": 0, + "content": "
MethodFinetuned ScalesConditionedmIoU(V)mIoU(P)F3D(↓)MMD(↓)
DD [2]s4×29.116.00.460.31
PDDNone✓33.422.80.270.32
PDDs2, s4✓43.922.80.280.16
PDDs1×31.323.20.220.13
PDDs1, s2, s4×44.724.20.210.11
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.594, + 0.36, + 0.61 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.625, + 0.788, + 0.822 + ], + "angle": 0, + "content": "In this work, we introduce the Pyramid Discrete Diffusion model (PDD) to address the significant challenges associated with 3D large scene generation, particularly in the context of limited scale and available datasets. The PDD demonstrates a progressive approach to generating high-quality 3D outdoor scenes, transitioning seamlessly from coarse to fine details. Compared to the other methods, the PDD can generate high-quality scenes within limited resource constraints and does not require additional data sources. Our extensive experimental results show that PDD consistently performs favourably in both unconditional and conditional generation tasks, establishing itself as a reliable and robust solution for the creation of realistic and intricate scenes. Furthermore, the proposed PDD method has great potential in efficiently adapting models trained on synthetic data to real-world datasets and suggests a promising solution to the current challenge of limited real-world data." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.365, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Pyramid Diffusion for Fine 3D Large Scene Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.311 + ], + "angle": 0, + "content": "Supported by the Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/ Interior Business Center (DOI/IBC) contract number 140D0423C0074. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of IARPA, DOI/IBC, or the U.S. Government. Y. Liu and C. Li are supported in part by the National Science Foundation of China (Grant No: 62202395)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.333, + 0.323, + 0.349 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.362, + 0.785, + 0.389 + ], + "angle": 0, + "content": "1. Anvekar, T., Tabib, R.A., Hegde, D., Mudengudi, U.: Vg-vae: A venatus geometry point-cloud variational auto-encoder. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.39, + 0.785, + 0.416 + ], + "angle": 0, + "content": "2. Austin, J., Johnson, D.D., Ho, J., Tarlow, D., Van Den Berg, R.: Structured denoising diffusion models in discrete state-spaces. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.418, + 0.785, + 0.457 + ], + "angle": 0, + "content": "3. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.459, + 0.785, + 0.484 + ], + "angle": 0, + "content": "4. Chen, Z., Wang, G., Liu, Z.: Scenedreamer: Unbounded 3d scene generation from 2d image collections. In: arXiv preprint arXiv: 2302.01330 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.486, + 0.785, + 0.512 + ], + "angle": 0, + "content": "5. Cheng, A.C., Li, X., Liu, S., Sun, M., Yang, M.H.: Learning 3d dense correspondence via canonical point autoencoder. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.513, + 0.785, + 0.539 + ], + "angle": 0, + "content": "6. Cheng, A.C., Li, X., Sun, M., Yang, M.H., Liu, S.: Learning 3d dense correspondence via canonical point autoencoder. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.54, + 0.785, + 0.58 + ], + "angle": 0, + "content": "7. Çiçek, Ö., Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3d unet: learning dense volumetric segmentation from sparse annotation. In: MICCAI (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.582, + 0.785, + 0.607 + ], + "angle": 0, + "content": "8. Cong, Y., Chen, R., Ma, B., Liu, H., Hou, D., Yang, C.: A comprehensive study of 3-d vision-based robot manipulation. IEEE Transactions on Cybernetics (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.609, + 0.785, + 0.635 + ], + "angle": 0, + "content": "9. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.636, + 0.785, + 0.662 + ], + "angle": 0, + "content": "0. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: CVPR (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.663, + 0.785, + 0.689 + ], + "angle": 0, + "content": "1. Graham, B., Engelcke, M., Van Der Maaten, L.: 3d semantic segmentation with submanifold sparse convolutional networks. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.69, + 0.785, + 0.731 + ], + "angle": 0, + "content": "2. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.732, + 0.785, + 0.771 + ], + "angle": 0, + "content": "3. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "4. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.839 + ], + "angle": 0, + "content": "5. Ho, J., Saharia, C., Chan, W., Fleet, D.J., Norouzi, M., Salimans, T.: Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.362, + 0.785, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "16. Huang, W., Wang, C., Zhang, R., Li, Y., Wu, J., Fei-Fei, L.: Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.786, + 0.217 + ], + "angle": 0, + "content": "17. Lan, Z., Yew, Z.J., Lee, G.H.: Robust point cloud based reconstruction of large-scale outdoor scenes. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.217, + 0.787, + 0.243 + ], + "angle": 0, + "content": "18. Lee, J., Im, W., Lee, S., Yoon, S.E.: Diffusion probabilistic models for scene-scale 3d categorical data. arXiv preprint arXiv:2301.00527 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.787, + 0.284 + ], + "angle": 0, + "content": "19. Li, X., Li, C., Tong, Z., Lim, A., Yuan, J., Wu, Y., Tang, J., Huang, R.: Campus3d: A photogrammetry point cloud benchmark for hierarchical understanding of outdoor scene. In: ACM MM (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.787, + 0.311 + ], + "angle": 0, + "content": "20. Li, Y., Ma, L., Zhong, Z., Liu, F., Chapman, M.A., Cao, D., Li, J.: Deep learning for lidar point clouds in autonomous driving: A review. NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.787, + 0.352 + ], + "angle": 0, + "content": "21. Li, Y., Yu, A.W., Meng, T., Caine, B., Ngiam, J., Peng, D., Shen, J., Lu, Y., Zhou, D., Le, Q.V., et al.: Deepfusion: Lidar-camera deep fusion for multi-modal 3d object detection. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.787, + 0.38 + ], + "angle": 0, + "content": "22. Li, Z., Wang, Q., Snavely, N., Kanazawa, A.: Infinitenature-zero: Learning perpetual view generation of natural scenes from single images. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.38, + 0.787, + 0.406 + ], + "angle": 0, + "content": "23. Lin, C.H., Lee, H.Y., Menapace, W., Chai, M., Siarohin, A., Yang, M.H., Tulyakov, S.: Infinicity: Infinite-scale city synthesis. arXiv preprint arXiv:2301.09637 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.406, + 0.787, + 0.447 + ], + "angle": 0, + "content": "24. Liu, M., Shi, R., Chen, L., Zhang, Z., Xu, C., Wei, X., Chen, H., Zeng, C., Gu, J., Su, H.: One-2-3-45++: Fast single image to 3d objects with consistent multi-view generation and 3d diffusion. arXiv preprint arXiv:2311.07885 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.787, + 0.474 + ], + "angle": 0, + "content": "25. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.787, + 0.501 + ], + "angle": 0, + "content": "26. Luo, S., Hu, W.: Diffusion probabilistic models for 3d point cloud generation. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.787, + 0.528 + ], + "angle": 0, + "content": "27. Ma, Q., Yang, J., Tang, S., Black, M.J.: The power of points for modeling humans in clothing. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.787, + 0.555 + ], + "angle": 0, + "content": "28. Mascaro, R., Teixeira, L., Chli, M.: Diffuser: Multi-view 2d-to-3d label diffusion for semantic scene segmentation. In: ICRA (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.787, + 0.582 + ], + "angle": 0, + "content": "29. Moon, T., Choi, M., Lee, G., Ha, J.W., Lee, J.: Fine-tuning diffusion models with limited data. In: NeurIPS 2022 Workshop on Score-Based Methods (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.787, + 0.609 + ], + "angle": 0, + "content": "30. Moro, S., Komuro, T.: Generation of virtual reality environment based on 3d scanned indoor physical space. In: ISVC (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.787, + 0.636 + ], + "angle": 0, + "content": "31. Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.636, + 0.787, + 0.677 + ], + "angle": 0, + "content": "32. Ögün, M.N., Kurul, R., Yasar, M.F., Turkoglu, S.A., Avci, Ş., Yildiz, N.: Effect of leap motion-based 3d immersive virtual reality usage on upper extremity function in ischemic stroke patients. Arquivos de neuro-psiquiatria (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.677, + 0.787, + 0.704 + ], + "angle": 0, + "content": "33. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.787, + 0.731 + ], + "angle": 0, + "content": "34. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: NeurIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.731, + 0.787, + 0.772 + ], + "angle": 0, + "content": "35. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.799 + ], + "angle": 0, + "content": "36. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "37. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.365, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Pyramid Diffusion for Fine 3D Large Scene Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "38. Sohl-Dickstein, J., Weiss, E.A., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. arXiv preprint arXiv:1503.03585 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "39. Sra, M., Garrido-Jurado, S., Schmandt, C., Maes, P.: Procedurally generated virtual reality from 3d reconstructed physical space. In: Proceedings of the 22nd ACM Conference on Virtual Reality Software and Technology (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.786, + 0.259 + ], + "angle": 0, + "content": "40. Su, S.Y., Bagautdinov, T., Rhodin, H.: Npc: Neural point characters from video. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.787, + 0.3 + ], + "angle": 0, + "content": "41. Tang, J., Nie, Y., Markhasin, L., Dai, A., Thies, J., Nießner, M.: Diffuscene: Scene graph denoising diffusion probabilistic model for generative indoor scene synthesis. arXiv preprint arXiv:2303.14207 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.787, + 0.328 + ], + "angle": 0, + "content": "42. Tang, Y., He, H., Wang, Y., Mao, Z., Wang, H.: Multi-modality 3d object detection in autonomous driving: A review. Neurocomputing (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.329, + 0.787, + 0.37 + ], + "angle": 0, + "content": "43. Wilson, J., Song, J., Fu, Y., Zhang, A., Capodieci, A., Jayakumar, P., Barton, K., Ghaffari, M.: Motionsc: Data set and network for real-time semantic mapping in dynamic environments. IEEE Robotics and Automation Letters 7(3) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.411 + ], + "angle": 0, + "content": "44. Wu, P., Jia, X., Chen, L., Yan, J., Li, H., Qiao, Y.: Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.439 + ], + "angle": 0, + "content": "45. Xie, H., Chen, Z., Hong, F., Liu, Z.: Citydreamer: Compositional generative model of unbounded 3d cities. arXiv preprint arXiv:2309.00610 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.44, + 0.787, + 0.466 + ], + "angle": 0, + "content": "46. Xu, Z., He, Z., Wu, J., Song, S.: Learning 3d dynamic scene representations for robot manipulation. arXiv preprint arXiv:2011.01968 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.787, + 0.495 + ], + "angle": 0, + "content": "47. Zeng, X., Vahdat, A., Williams, F., Gojcic, Z., Litany, O., Fidler, S., Kreis, K.: Lion: Latent point diffusion models for 3d shape generation. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.496, + 0.787, + 0.522 + ], + "angle": 0, + "content": "48. Zheng, H., Nie, W., Vahdat, A., Anandkumar, A.: Fast training of diffusion models with masked transformers. arXiv preprint arXiv:2306.09305 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.787, + 0.55 + ], + "angle": 0, + "content": "49. Zheng, Y., Yifan, W., Wetzstein, G., Black, M.J., Hilliges, O.: Pointavatar: Deformable point-based head avatars from videos. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.551, + 0.787, + 0.577 + ], + "angle": 0, + "content": "50. Zhou, L., Du, Y., Wu, J.: 3d shape generation and completion through point-voxel diffusion. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.606 + ], + "angle": 0, + "content": "51. Zhuang, F., Qi, Z., Duan, K., Xi, D., Zhu, Y., Zhu, H., Xiong, H., He, Q.: A comprehensive survey on transfer learning. arXiv preprint arXiv:1911.02685 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.606 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Pyramid Diffusion for Fine 3D Large Scene Generation/0615ce43-c6d2-40fd-ba98-d83c475c6139_origin.pdf b/2024/Pyramid Diffusion for Fine 3D Large Scene Generation/0615ce43-c6d2-40fd-ba98-d83c475c6139_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9f335ccdbe6d9bc54d1cd1c170f4aef63d1aaa63 --- /dev/null +++ b/2024/Pyramid Diffusion for Fine 3D Large Scene Generation/0615ce43-c6d2-40fd-ba98-d83c475c6139_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:034d230c0fbda4564c154a59c34b2fd2466850c256486c1d0c4315f460cc6417 +size 9929270 diff --git a/2024/Pyramid Diffusion for Fine 3D Large Scene Generation/full.md b/2024/Pyramid Diffusion for Fine 3D Large Scene Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..adeff260ad32484a8ce41a5d23b58aaae699c9a6 --- /dev/null +++ b/2024/Pyramid Diffusion for Fine 3D Large Scene Generation/full.md @@ -0,0 +1,270 @@ +# Pyramid Diffusion for Fine 3D Large Scene Generation + +Yuheng Liu $^{1,2*}$ , Xinke Li $^{3*}$ , Xueting Li $^{4}$ , Lu Qi $^{5\dagger}$ , Chongshou Li $^{1}$ , Ming-Hsuan Yang $^{5,6}$ + +$^{1}$ Southwest Jiaotong University $^{2}$ University of Leeds $^{3}$ City University of Hong Kong $^{4}$ NVIDIA $^{5}$ The University of California, Merced $^{6}$ Yonsei University + +![](images/5af58c26ba66a3ae081c745e97e1401a334775a1449f6987b379e03b46fce0a7.jpg) +Fig. 1: We present Pyramid Discrete Diffusion Model, a method that progresses from generating coarse- to fine-grained scenes, mirroring the top-down sequence of the pyramid structure shown. The model is extended for cross-dataset and infinite scene generation, with detailed scene intricacies illustrated on the flanking sides of the image. $\mathcal{D}_s$ and $\mathcal{D}_t$ refer to a source dataset and a target dataset, respectively. + +Abstract. Diffusion models have shown remarkable results in generating 2D images and small-scale 3D objects. However, their application to the synthesis of large-scale 3D scenes has been rarely explored. This is mainly due to the inherent complexity and bulky size of 3D scenery data, particularly outdoor scenes, and the limited availability of comprehensive real-world datasets, which makes training a stable scene diffusion model challenging. In this work, we explore how to effectively generate large-scale 3D scenes using the coarse-to-fine paradigm. We introduce a framework, the Pyramid Discrete Diffusion model (PDD), which employs scale-varied diffusion models to progressively generate high-quality outdoor scenes. Experimental results of PDD demonstrate + +our successful exploration in generating 3D scenes both unconditionally and conditionally. We further showcase the data compatibility of the PDD model, due to its multi-scale architecture: a PDD model trained on one dataset can be easily fine-tuned with another dataset. Code is available at https://github.com/yuhengliu02/pyramid-discrete-diffusion. + +Keywords: 3D Scene Generation $\cdot$ Diffusion Models $\cdot$ Transfer Learning + +# 1 Introduction + +3D scene generation is the task of creating digital representations that mimic the three-dimensional complexities of our real-world environment, allowing for a more nuanced understanding of the tangible surroundings. This technique plays an essential role in fundamental computer vision tasks such as autonomous driving [21,42,44], virtual reality [30,32,39], and robotic manipulation [8,16,46]. However, high-quality large 3D scenes are extremely challenging to synthesize due to their inherently bulky size, and the lack of large-scale 3D scene datasets [43]. + +Meanwhile, recent advances in diffusion models have shown impressive results in generating 2D images [35-37] or small-scale 3D objects [24, 33]. Yet, it is not a trivial task to employ diffusion models in 3D scene generation. On one hand, state-of-the-art diffusion models leave substantial memory footprints and demand considerable training time, posing a particular challenge when generating 3D scenes with large scales and intricate details. On the other hand, diffusion models require a large amount of training data [29, 48], while capturing large-scale 3D scenes is itself a challenging and ongoing research topic [4, 23]. As a result, only a few attempts have been made to apply diffusion models directly to 3D outdoor scenes [18], which resulted in unstable generation and thus suboptimal performance. + +To resolve these challenges, existing works focus on conditional generation and resort to additional signals such as Scene Graphs [41] or 2D maps [28] for guidance. Nonetheless, such conditional guidance is not always accessible, thereby restricting the generalizability of these approaches. Inspired by the coarse-to-fine philosophy widely used in image super-resolution [14, 31, 38], we introduce the Pyramid Discrete Diffusion model (PDD), a framework that progressively generates large 3D scenes without relying on additional guidance. + +We begin by generating small-scale 3D scenes and progressively increase the scale. At each scale level, we learn a separate diffusion model. This model takes the generated scene from the previous scale as a condition (except for the first diffusion model which takes noise as input) and synthesizes a 3D scene of a larger scale. Intuitively, this multi-scale generation process breaks down a challenging unconditional generation task (i.e., high-quality 3D scene generation) into several more manageable conditional generation tasks. This separation allows each diffusion model to specialize in generating either coarse structure (smaller scale) or intricate details (larger scale). Moreover, at the highest scale, we employ a technique known as scene subdivision, which involves dividing a large scene into multiple smaller segments that are synthesized using a shared diffusion model. + +This approach mitigates the issues of oversized models caused by the bulky size of the 3D scenes. Additionally, a noteworthy outcome of our multi-scale design is its capacity to facilitate cross-data transfer applications, which substantially reduces training resources. Lastly, we further propose a natural extension of our PDD framework with scene subdivision for infinite 3D scene generation, thereby demonstrating the scalability of the proposed method. + +The main contributions of this work are as follows: + +- We practically implement a coarse-to-fine strategy for 3D outdoor scene generation via designing a novel pyramid diffusion model. +- We conduct extensive experiments on our pyramid diffusion, demonstrating its generation of higher quality 3D scenes with comparable computational resources of existing approaches. In addition, we introduce new metrics to evaluate the quality of 3D scene generation from various perspectives. +- Our proposed method showcases broader applications, enabling the generation of scenes from synthetic datasets to real-world data. Furthermore, our approach can be extended to facilitate the creation of infinite scenes. + +# 2 Related Work + +Diffusion Models for 2D Images. Recent advancements in the generative model have seen the diffusion models [14, 31, 38] rise to prominence, especially in applications in 2D image creation [9, 35, 36]. In order to generate high-fidelity images via diffusion models, a multi-stage diffusion process is proposed and employed as per [13, 15, 37]. This process starts with the generation of a coarse-resolution image using an initial diffusion model. Subsequently, a second diffusion model takes this initial output as input, refining it into a finer-resolution image. These cascaded diffusions can be iteratively applied to achieve the desired image resolution. We note that the generation of fine-grained 3D data presents more challenges than 2D due to the addition of an extra dimension. Consequently, our work is motivated by the aforementioned multistage 2D approaches to explore their applicability in 3D contexts. Furthermore, we aim to leverage the advantages of this structure to address the scarcity of datasets in 3D scenes. + +Diffusion Models for 3D Generation. As a sparse and memory-efficient representation, 3D point clouds has been widely used in various computer vision applications such as digital human [27, 40, 49], autonomous driving [20], and 3D scene reconstruction [17]. Point clouds generation aims to synthesize a 3D point clouds from a random noise [5,6], or scanned lidar points [18]. Though the memory efficiency of point clouds is a valuable property, it poses high challenges in the task of point cloud generation. Existing works largely focus on using Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), or Vector Quantized Variational Autoencoders (VQ-VAEs) as the backbone for this task [1,5,6]. However, these models have limited capacity for high-fidelity generation and are notoriously known for unstable training. As an alternative to the generative models discussed above, diffusion models have revolutionized the computer vision community with their impressive performance in 2D image + +![](images/693ab21859dc9bb04e9636a414360c8969b6fae8a9325b8b64c77f92e2d4b569.jpg) +Fig. 2: Framework of the proposed Pyramid Discrete Diffusion model. In our structure, there are three different scales. Scenes generated by a previous scale can serve as a condition for the current scale after processing through our scale adaptive function. Furthermore, for the final scale processing, the scene from the previous scale is subdivided into four sub-scenes. The final scene is reconstructed into a large scene using our Scene Subdivision module. + +generation [35-37]. Yet, applying diffusion models for 3D point cloud generation has not been thoroughly explored hitherto. Point-Voxel Diffusion [50] proposes to generate a raw point cloud through the diffusion process while LION [47] and DPM [26] use the latent representation of a point cloud during the denoising process. However, all these methods focus on object-level point clouds and cannot be naively extended to scene-level point clouds. Most relevant to our work is [18], where a diffusion model is trained on a scene-level point cloud dataset for the synthesis task. However, due to the capacity limitation of diffusion models, generating a scene-level point cloud with a single diffusion model leads to unsatisfying results, such as undesired wholes or the lack of fine-grained objects. In this work, we propose a pyramid discrete diffusion model that reduces the difficulty at each pyramid level, thus producing scene point clouds with more realistic and fine-grained details. + +3D Large-scale Scene Generation. Generating large-scale 3D scenes is an important but highly challenging task. A generative model on 3D scenes potentially provides infinite training data for tasks such as scene segmentation, autonomous driving, etc. Existing works [4,22,23,45] simplify this task by first generating 2D scenes and then "lifting" them to 3D. Though such design is efficient for city scenes populated with regular geometries (e.g., buildings), it does not generalize easily to scenes with more fine-grained objects (e.g., pedestrians, cars, trees, etc.) In this paper, we directly generate 3D outdoor scenes using diffusion models, which include abundant small objects with semantics. + +# 3 Approach + +The proposed Pyramid Discrete Diffusion (PDD) model comprises multi-scale models capable of step-by-step generation of high-quality 3D scenes from smaller scales. The PDD first extends the standard discrete diffusion for 3D data (Section 3.2) and then proposes a scene subdivision method to further reduce memory requirements (Section 3.3). Finally, we demonstrate two practical applications of PDD in specific scenarios (Section 3.4). + +# 3.1 Discrete Diffusion + +We focus on learning a data distribution based on 3D semantic scenes. Specifically, the semantic scene is represented in a one-hot format, i.e., $\mathbf{X} \in \{0,1\}^{h \times w \times d \times c}$ , where $h$ , $w$ , and $d$ indicate the dimensions of the scene, respectively, and $c$ denotes the size of the one-hot label. + +Discrete diffusion [2] has been proposed to generate discrete data including semantic scenes. It involves applying the Markov transition matrix on discrete states for noise diffusion. In the forward process, an original scene $\mathbf{X}_0$ is gradually corrupted into a $t$ -step noised map $\mathbf{X}_t$ with $t = 1,\dots ,T$ . Each forward step can be defined by a Markov uniform transition matrix $\mathbf{Q}_t$ as $\mathbf{X}_t = \mathbf{X}_{t - 1}\mathbf{Q}_t$ . Based on the Markov property, we can derive the $t$ -step scene $\mathbf{X}_t$ straight from $\mathbf{X}_0$ with a cumulative transition matrix $\bar{\mathbf{Q}}_t = \mathbf{Q}_1\mathbf{Q}_2\dots \mathbf{Q}_t$ : + +$$ +q \left(\mathbf {X} _ {t} \mid \mathbf {X} _ {0}\right) = \operatorname {C a t} \left(\mathbf {X} _ {t}; \mathbf {P} = \mathbf {X} _ {0} \bar {\mathbf {Q}} _ {t}\right), \tag {1} +$$ + +where $\operatorname{Cat}(\mathbf{X};\mathbf{P})$ is a multivariate categorical distribution over the one-hot semantic labels $\mathbf{X}$ with probabilities given by $\mathbf{P}$ . Finally, the semantic scene $\mathbf{X}_T$ at the last step $T$ is supposed to be in the form of a uniform discrete noise. In the reverse process, a learnable model parametrized by $\theta$ is used to predict denoised semantic labels by $\tilde{p}_{\theta}\left(\tilde{\mathbf{X}}_0\mid \mathbf{X}_t\right)$ . The reparametrization trick is applied subsequently to get the reverse process $p_{\theta}\left(\mathbf{X}_{t - 1}\mid \mathbf{X}_t\right)$ : + +$$ +p _ {\theta} \left(\mathbf {X} _ {t - 1} \mid \mathbf {X} _ {t}\right) = \mathbb {E} _ {\tilde {p} _ {\theta} \left(\tilde {\mathbf {X}} _ {0} \mid \mathbf {X} _ {t}\right)} q \left(\mathbf {X} _ {t - 1} \mid \mathbf {X} _ {t}, \tilde {\mathbf {X}} _ {0}\right). \tag {2} +$$ + +A loss consisting of the two KL divergences is proposed to learn better reconstruction ability for the model, given by + +$$ +\begin{array}{l} \mathcal {L} _ {\theta} = d _ {\mathrm {K L}} \left(q \left(\mathbf {X} _ {t - 1} \mid \mathbf {X} _ {t}, \mathbf {X} _ {0}\right) \| p _ {\theta} \left(\mathbf {X} _ {t - 1} \mid \mathbf {X} _ {t}\right)\right) \tag {3} \\ + \lambda d _ {\mathrm {K L}} \left(q \left(\mathbf {X} _ {0}\right) \| \tilde {p} _ {\theta} \left(\tilde {\mathbf {X}} _ {0} \mid \mathbf {X} _ {t}\right)\right), \\ \end{array} +$$ + +where $\lambda$ is an auxiliary loss weight and $d_{\mathrm{KL}}$ stands for KL divergence. In the following, we focus on extending the discrete diffusion into the proposed PDD. + +# 3.2 Pyramid Discrete Diffusion + +We propose PDD that operates various diffusion processes across multiple scales (or resolutions), as depicted in Figure 2. Given a 3D scene data $\mathbf{Z} \in \{0,1\}^{h \times w \times d \times c}$ , we define a 3D pyramid including different scales of $\mathbf{Z}$ , i.e., $\{\mathbf{Z}^{(1)}, \dots, \mathbf{Z}^{(l)}, \dots, \mathbf{Z}^{(L)}\}$ , where a larger $l$ indicates a larger scene scale. Formally, let $h_l \times w_l \times d_l \times c$ denote the dimension of $\mathbf{Z}^{(l)}$ , $h_{l+1} \geq h_l$ , $w_{l+1} \geq w_l$ and $d_{l+1} \geq d_l$ are kept for $l = 1, \dots, L-1$ . We note that such a pyramid can be obtained by applying different down-sample operators, such as pooling functions, on $\mathbf{Z}$ . For each scale in the pyramid, we construct a conditional discrete diffusion model parameterized by $\theta_l$ . The $l$ -th model for $l \neq 1$ is given by: + +$$ +\tilde {p} _ {\theta_ {l}} \left(\tilde {\mathbf {X}} _ {0} ^ {(l)} \mid \mathbf {X} _ {t} ^ {(l)}, \mathbf {Z} ^ {(l - 1)}\right) = \tilde {p} _ {\theta_ {l}} \left(\tilde {\mathbf {X}} _ {0} ^ {(l)} \mid \operatorname {C o n c a t} \left(\mathbf {X} _ {t} ^ {(l)}, \phi^ {(l)} \left(\mathbf {Z} ^ {(l - 1)}\right)\right)\right), \tag {4} +$$ + +where $\mathbf{X}_t^{(l)}$ and $\mathbf{X}_0^{(l)}$ are with the same size of $\mathbf{Z}^{(l)}$ , and $\phi^{(l)}$ is a Scale Adaptive Function (SAF) for upsampling $\mathbf{Z}^{(l-1)}$ into the size of $\mathbf{Z}^{(l)}$ . As a case in point, SAF can be a tri-linear interpolation function depending on the data. Additionally, we maintain the first model $\tilde{p}_{\theta_1}$ as the original non-conditional model. + +During the training process, PDD learns $L$ denoising models separately at varied scales of scene pyramids in the given dataset. Given that $\mathbf{Z}^{(l - 1)}$ is essentially a lossy-compressed version of $\mathbf{Z}^{(l)}$ , the model training can be viewed as learning to restore the details of a coarse scene. In the inference process, denoising model $p_{\theta_1}$ is performed initially according to Equation (2) and the rest of PDD models are executed in sequence from $l = 2$ to $L$ via the sampling, + +$$ +\mathbf {X} _ {t - 1} ^ {(l)} \sim p _ {\theta_ {l}} (\mathbf {X} _ {t - 1} ^ {(l)} \mid \mathbf {X} _ {t} ^ {(l)}, \mathbf {X} _ {0} ^ {(l - 1)}), \tag {5} +$$ + +where $\mathbf{X}_0^{(l - 1)}$ is the denoised result of $\tilde{p}_{\theta_{l - 1}}$ . + +Except for the high-quality generation, the proposed PDD bears two merits: 1) Diffusion models in PDD can be trained in parallel due to their independence, which allows for a flexible computation reallocation during training. 2) Due to its multi-stage generation process, PDD is fitting for restoring scenes of arbitrary coarse-grained scale by starting from the intermediate processes, thereby extending the method's versatility. + +# 3.3 Scene Subdivision + +To overcome the memory constraint for generating large 3D scenes, we propose the scene subdivision method. We divide a 3D scene $\mathbf{Z}^{(l)}$ along $z$ -axis into $I$ overlapped sub-components as $\{\mathbf{Z}_i^{(l)}\}_{i=1}^I$ . For the instance of four subscenes case, let $\mathbf{Z}_i^{(l)} \in \{0,1\}^{(1 + \delta_l)h_l \setminus 2 \times (1 + \delta_l)w_l \setminus 2 \times d_l \times c}$ denote one subscene and $\delta_l$ denote the overlap ratio, the shared $l$ -th diffusion model in PDD is trained to reconstruct $\mathbf{Z}_i^{(l)}$ for $i = 1, \dots, 4$ . Subsequently, sub-scenes are merged into a holistic one by a fusion algorithm, i.e., voting on the overlapped parts to ensure the continuity of the 3D scene. + +In the training process, to ensure context-awareness of the entire scene during the generation of a sub-scene, we train the model by adding the overlapped regions with other sub-scenes as the condition. In the inference process, the entire scene is generated in an autoregressive manner. Apart from the first sub-scene generated without context, all other sub-scenes utilize the already generated overlapped region as a condition, i.e., + +$$ +\mathbf {X} _ {t - 1, i} ^ {(l)} \sim p _ {\theta} \left(\mathbf {X} _ {t - 1, i} ^ {(l)} \mid \mathbf {X} _ {t, i} ^ {(l)}, \mathbf {X} _ {0, i} ^ {(l + 1)}, \sum_ {j \neq i} \Delta_ {i j} \odot \mathbf {X} _ {0, j} ^ {(l + 1)}\right), \tag {6} +$$ + +where $j$ is the index of generated sub-scenes before $i$ -th scene, and $\Delta_{ij}$ is a binary mask between $\mathbf{X}_{0,i}^{(l+1)}$ and $\mathbf{X}_{0,j}^{(l+1)}$ representing the overlapped region on $\mathbf{X}_{0,j}^{(l+1)}$ with 1 and the separate region with 0. Scene Subdivision module can reduce the model parameters, as diffusion model could be shared by four sub-scenes. In practice, we only implement the scene subdivision method on the largest scale which demands the largest memory. + +# 3.4 Applications + +Beyond its primary function as a generative model, we introduce two novel applications for PDD. First, cross-dataset transfer aims at adapting a model trained on a source dataset to a target dataset [51]. Due to the flexibility of input scale, PDD can achieve this by retraining or fine-tuning the smaller-scale models in the new dataset while keeping the larger-scale models. The strategy leveraging PDD improves the efficiency of transferring 3D scene generation models between distinct datasets. Second, infinite scene generation is of great interest in fields such as autonomous driving [10] and urban modeling [19] which require a huge scale of 3D scenes. PDD can extend its scene subdivision technique. By using the edge of a previously generated scene as a condition as in Equation (6), it can iteratively create larger scenes, potentially without size limitations. + +# 4 Experimental Results + +# 4.1 Evaluation Protocols + +Since the metrics used in 2D generation such as FID [12] are not directly applicable in the 3D, we introduce and implement three metrics to assess the quality of the generated 3D scenes. We note that more implementation details can be found in the supplementary material. + +Semantic Segmentation results on the generated scenes are used to evaluate the effectiveness of models in creating semantically coherent scenes. Specifically, two architectures, the voxel-based SparseUNet [11] and point-based PointNet++ [34], are implemented to perform the segmentation tasks. We report the mean Intersection over Union (mIoU) and Mean Accuracy (MAs) for evaluation. + +Table 1: Comparison of various diffusion models on 3D semantic scene generation of CarlaSC. DiscreteDiff [2], LatentDiff [18], and P-DiscreteDiff refer to the original discrete diffusion, latent discrete diffusion, and our approach, respectively. Conditioned models work based on the context of unlabeled point clouds or the coarse version of the ground truth scene. A higher Segmentation Metric value is better, indicating semantic consistency. A lower Feature-based Metric value is preferable, representing closer proximity to the original dataset. The brackets with $V$ represent voxel-based network and $P$ represent point-based network. + +
MethodModelConditionSegmentation MetricFeature-based Metric
mIoU (V)MA (V)mIoU (P)MA (P)F3D (↓)MMD (↓)
Ground Truth--52.1972.4032.9047.680.2460.108
UnconditionedDiscreteDiff [2]-40.0563.6525.5438.711.3610.599
LatentDiff [18]-38.0162.3926.6945.870.3310.211
P-DiscreteDiff (Ours)-68.0285.6633.8952.120.3150.200
ConditionedDiscreteDiff [2]Point cloud38.5559.9728.4144.060.3570.261
DiscreteDiff [2]Coarse scene (s1)52.5277.2327.9343.130.3590.284
P-DiscreteDiff (Ours)Coarse scene (s1)55.7578.7029.7846.610.3420.274
+ +F3D is a 3D adaption of the 2D Fréchet Inception Distance (FID) [12], which is based on a pre-trained autoencoder with an 3D CNN architecture. We calculate and report the Fréchet distance (by $10^{-3}$ ratio) between the generated scenes and real scenes in the feature domain. + +Maximum Mean Discrepancy (MMD) is a statistical measure to quantify the disparity between the distributions of generated and real scenes. Similar to our F3D approach, we extract features via the same pre-trained autoencoder and present the MMD between 3D scenes. + +# 4.2 Experiment Settings + +Datasets. We use CarlaSC [43] and SemanticKITTI [3] for experiments. Specifically, we conduct our main experiments as well as ablation studies on the synthesis dataset CarlaSC due to its large data volume and diverse semantic objects. Our primary model is trained on the training set of CarlaSC with 10 categories and 32,400 scans. SemanticKITTI, which is a real-world collected dataset, is used for our cross-dataset transfer experiment. Both datasets are adjusted to ensure consistency in semantic categories, with further details in the supplementary material. + +Model Architecture. The primary proposed PDD is performed on three scales of a 3D scene pyramid, i.e., $s_1$ , $s_2$ and $s_4$ in Table 2. We implement 3D-UNets [7] for three diffusion models in PDD based on the scales. Notably, the model applied on $s_4$ scale is with the input/output size of $s_3'$ due to the use of scene subdivision, while such a size of other models follows the working scale size. In the ablation study, we also introduce the scale $s_3$ in the experiment. Additionally, we implement two baseline methods merely on scale $s_4$ which are the original discrete diffusion [2] and the latent diffusion model with VQ-VAE decoder [18]. Training Setting. We train each PDD model using the same training setting except for the batch size. Specifically, we set the learning rate of $10^{-3}$ for the + +Table 2: Different scales in the 3D scene pyramid. + +
Scale Rep.3D Scene Size
s132 × 32 × 4
s264 × 64 × 8
s3128 × 128 × 8
s3'136 × 136 × 16
s4256 × 256 × 16
+ +Table 3: Average SSIM between each generated scene and the closest scene in the training set. We generate $1k$ scenes and calculate the average. + +
DatasetMethodSSIM
CarlaSCUncon-Gen0.72
Val Set0.65
Sem-KITTI (fine-tuned)Uncon-Gen0.74
Val Set0.67
+ +Table 4: Comparison of different diffusion pyramids on 3D semantic scene generation. + +
PyramidConditionedmIoU(V)mIoU(P)F3D(↓)MMD(↓)
s4×40.025.51.360.60
s1 → s4×67.032.10.320.24
s1 → s2 → s4×68.033.90.320.20
s1 → s2 → s3 → s4×68.033.40.320.23
s1 → s4✓52.527.90.360.28
s1 → s2 → s4✓55.829.80.340.27
s1 → s2 → s3 → s4✓55.929.60.340.28
+ +AdamW optimizer [25], and the time step $T = 100$ for the diffusion process, and 800 for the max epoch. The batch sizes are set to 128, 32, and 16 for the models working on $s_1$ , $s_2$ and $s_4$ scales. However, for the baseline method based on the $s_4$ scale, we use the batch size of 8 due to memory constraints. We note that all diffusion models are trained on four NVIDIA A100 GPUs. In addition, we apply the trilinear interpolation for the scene fusion algorithm and set the overlap ratio in scene subdivision, $\delta_l$ to 0.0625. + +# 4.3 Main Results + +Generation Quality. We compare our approach with two baselines, the original Discrete Diffusion [2] and the Latent Diffusion [18]. The result reported in Table 1 demonstrates the notable performance of our method across all metrics in both unconditional and conditional settings in comparable computational resources with existing method. Our proposed method demonstrates a notable advantage in segmentation tasks, especially when it reaches around $70\%$ mIoU for SparseUNet, which reflects its ability to generate scenes with accurate semantic coherence. We also provide visualizations of different model results in Figure 3, where the proposed method demonstrates better performance in detail generation and scene diversity for random 3D scene generations. + +Additionally, we conduct the comparison on conditioned 3D scene generation. We leverage the flexibility of input scale for our method and perform the generation by models in $s_2$ and $s_4$ scales conditioned on a coarse ground truth scene in $s_1$ scale. We benchmark our method against the discrete diffusion conditioned on + +![](images/2ffbc795e494692ef7ddfed08133aed7acbcedabc4b3b4c6782e176a5570ee7e.jpg) +Fig. 3: Visualization of unconditional generation results on CarlaSC. We compare with two baseline models - DiscreteDiff [2] and LatentDiff [18] and show synthesis from our models with different scales. Our method produces more diverse scenes compared to the baseline models. Furthermore, with more levels, our model can synthesize scenes with more intricate details. + +unlabeled point clouds and the same coarse scenes. Results in Table 1 and Figure 5 present the impressive results of our conditional generation comparison. It is also observed that the point cloud-based model can achieve decent performance on F3D and MMD, which could be caused by 3D point conditions providing more structural information about the scene than the coarse scene. Despite the informative condition of the point cloud, our method can still outperform it across most metrics. + +None-overfitting Verification. The MMD and F3D metrics (defined in Supp A) numerically illustrate the statistical feature distance between generated scenes and the training set. Our method achieves the lowest MMD and F3D among all baseline methods as shown in Table 1. However, we argue that this does not indicate overfitting to the dataset for the following reasons. First, our MMD and F3D are larger than those of the ground truth. Furthermore, we leverage structural similarity (SSIM) to search and show that a generated scene is different from its nearest neighbour in the dataset. Specifically, we generate $1k$ scenes and identify their closest matches in the training set using the SSIM metric. The average SSIM of these $1k$ scenes are calculated and presented in Table 3. Additionally, we apply the same methodology to the Validation Set to establish an + +![](images/30a115586a27a16e20586cab35c9c1d25857903e10df245797c525847bb69ac4.jpg) +Fig. 4: Data retrieval visualization. We generate $1k$ scenes using PDD on CarlaSC dataset, and retrieve the most similar scene in training set for each scene using SSIM (SSIM = 1 means identical), and plot SSIM distribution. Scenes at various percentiles are displayed (red box: generated scenes; grey box: scenes in training set), those with the highest $10\%$ similarity (i.e., $10^{th}$ percentile of SSIM) are very similar to the training set, but still not completely identical. + +oracle baseline. Table 3 shows that our generated scenes are comparable to the baseline, verifying that our method does not overfit the training set. To further support this, we use distribution plots, as shown in Figure 4, to validate the similarity between our generated scenes and the training set. We also display three pairs at different percentiles in this figure, showing that scenes with lower SSIM scores differ more from their nearest matches in the training set. This visual evidence reinforces that PDD effectively captures the distribution of the training set instead of merely memorizing it. + +# 4.4 Ablation Studies + +Pyramid Diffusion. Our experiments explore the impact of varying refinement scales on the quality of generated scenes. According to Table 4, both conditional and unconditional scene generation quality show incremental improvements with additional scales. Balancing training overhead and generation quality, a three-scale model with the scale of $s_4$ progression offers an optimal compromise between performance and computational cost. We find that as the number of scales increases, there is indeed a rise in performance, particularly notable upon the addition of the second scale. However, the progression from a three-scale pyramid to a four-scale pyramid turns out to be insignificant. Given the greater training overhead for a four-scale pyramid compared to a three-scale one, we choose the latter as our main structure. + +Scene Subdivision. We explore the optimal mask ratio for scene subdivision and report on Figure 6, which shows an inverse correlation between the mask ratio and the effectiveness of F3D and MMD metrics; higher mask ratios may result in diminished outcomes due to increased overfitting, leading to reduced randomness in the generated results. The lowest mask ratio test, 0.0625, achieves + +![](images/f53f936d2f764098f8fd2f4fd7f1a400309d691717a0b8406959e452efb13d20.jpg) +Fig. 5: Visualization of conditional generation results on CarlaSC. $PC$ stands for point cloud condition. + +Table 5: Generation results on CarlaSC in different scales on the diffusion pyramid without any conditions. All output scales are lifted to $s_4$ using the upsampling method. + +
Model No.ScalemIoU(↑)MA(↑)F3D(↓)MMD(↓)
1s118.042.70.290.16
2s243.766.80.290.18
3s468.085.70.320.23
+ +the best results across all metrics with detailed retention. Thus, we set a mask ratio of 0.0625 as the standard for our scene subdivision module. Further analysis shows that higher overlap ratios in scene subdivision result in quality deterioration, mainly due to increased discontinuities when merging sub-scenes using scene fusion algorithm. + +Performance on Different Scales. To facilitate a comprehensive understanding of the progressive improvement achieved by our coarse-to-fine method, we evaluate the quality of scenes generated at different scales and present the findings in Table 5. Even at the smaller scale $s_1$ , we observe high F3D and MMD scores, indicating its capacity in synthesizing scenes with both reasonable and diverse layouts. As we advance to larger scales (i.e., $s_2$ and $s_3$ ), the mIoU and MA scores consistently increase, demonstrating that our model focuses on learning intricate details in later stages. Meanwhile, F3D and MMD metrics show stability without significant decline, indicating a balanced enhancement in scene complexity and fidelity. + +![](images/ac7c9dfae1a2fafb3f1efba87450d49c37bad5888a7860936e42a98afda9eda4.jpg) +3D and MMD Scores by Mask Ratio + +![](images/4a15fe1a66cd5979ff88a19b843fc32f6df1fa5f94aa9a0229a80a69e9440908.jpg) +(b) mIoU and OA Scores by Mask Ratio + +![](images/21ddd5bca6b2f442adbb717ad015fa2738375dc9eb941e6df3647863f1193c8e.jpg) +Fig. 6: Effects of mask ratio on unconditional generation results. +Fig. 7: Infinite Scene Generation. Thanks to the pyramid representation, PDD can be readily applied for unbounded scene generation. This involves the initial efficient synthesis of a large-scale coarse 3D scene, followed by subsequent refinement at higher levels. + +# 4.5 Applications + +Cross-dataset. Figure 8 and Figure 9 showcase our model's performance on the transferred dataset from CarlaSC to SemanticKITTI for both unconditional and conditional scene generation. The PDD shows enhanced scene quality after finetuning with SemanticKITTI data, as indicated by improved results in Table 6. This fine-tuning effectively adapts the model to the dataset's complex object distributions and scene dynamics. Notably, despite the higher training effort of the Discrete Diffusion (DD) approach, our method outperforms DD even without fine-tuning, using only coarse scenes from SemanticKITTI. This demonstrates the strong cross-data transfer capability of our approach. + +Infinite Scene Generation. Figure 7 visualizes the process of generating large-scale infinite scenes using our PDD model. As discussed in Section 3.4, we first use the small-scale model to swiftly generate a coarse infinite 3D scene (bottom level in Figure 7). We then leverage larger-scale models to progressively add intricate details (middle and top level in Figure 7), enhancing realism. This approach enables our model to produce high-quality, continuous cityscapes without additional inputs, overcoming the limitations of conventional datasets with finite scenes and supporting downstream tasks like 3D scene segmentation. + +![](images/9926aa00e7c5b83b87c912a7f776e89cf588dfadbeb85dc1dac0d09d9597482a.jpg) +Fig. 8: SemanticKITTI unconditional generation. $FT$ stands for finetuning pretrained model from CarlaSC. + +![](images/22d576bfcc5a8c80fcebf39cfdc082b22581abcade1d53d2fd59527249ee997b.jpg) +Fig.9: SemanticKITTI conditional generation. $FT$ stands for finetuning from CarlaSC models. + +Table 6: Generation results on SemanticKITTI. Finetuned Scales set to None indicates training from scratch and others stand for finetuning corresponding pre-trained CarlaSC model. + +
MethodFinetuned ScalesConditionedmIoU(V)mIoU(P)F3D(↓)MMD(↓)
DD [2]s4×29.116.00.460.31
PDDNone✓33.422.80.270.32
PDDs2, s4✓43.922.80.280.16
PDDs1×31.323.20.220.13
PDDs1, s2, s4×44.724.20.210.11
+ +# 5 Conclusion + +In this work, we introduce the Pyramid Discrete Diffusion model (PDD) to address the significant challenges associated with 3D large scene generation, particularly in the context of limited scale and available datasets. The PDD demonstrates a progressive approach to generating high-quality 3D outdoor scenes, transitioning seamlessly from coarse to fine details. Compared to the other methods, the PDD can generate high-quality scenes within limited resource constraints and does not require additional data sources. Our extensive experimental results show that PDD consistently performs favourably in both unconditional and conditional generation tasks, establishing itself as a reliable and robust solution for the creation of realistic and intricate scenes. Furthermore, the proposed PDD method has great potential in efficiently adapting models trained on synthetic data to real-world datasets and suggests a promising solution to the current challenge of limited real-world data. + +# Acknowledgement + +Supported by the Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/ Interior Business Center (DOI/IBC) contract number 140D0423C0074. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of IARPA, DOI/IBC, or the U.S. Government. Y. Liu and C. Li are supported in part by the National Science Foundation of China (Grant No: 62202395). + +# References + +1. Anvekar, T., Tabib, R.A., Hegde, D., Mudengudi, U.: Vg-vae: A venatus geometry point-cloud variational auto-encoder. In: CVPR (2022) +2. Austin, J., Johnson, D.D., Ho, J., Tarlow, D., Van Den Berg, R.: Structured denoising diffusion models in discrete state-spaces. In: NeurIPS (2021) +3. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: ICCV (2019) +4. Chen, Z., Wang, G., Liu, Z.: Scenedreamer: Unbounded 3d scene generation from 2d image collections. In: arXiv preprint arXiv: 2302.01330 (2023) +5. Cheng, A.C., Li, X., Liu, S., Sun, M., Yang, M.H.: Learning 3d dense correspondence via canonical point autoencoder. In: ECCV (2022) +6. Cheng, A.C., Li, X., Sun, M., Yang, M.H., Liu, S.: Learning 3d dense correspondence via canonical point autoencoder. In: NeurIPS (2021) +7. Çiçek, Ö., Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3d unet: learning dense volumetric segmentation from sparse annotation. In: MICCAI (2016) +8. Cong, Y., Chen, R., Ma, B., Liu, H., Hou, D., Yang, C.: A comprehensive study of 3-d vision-based robot manipulation. IEEE Transactions on Cybernetics (2021) +9. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: NeurIPS (2021) +0. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: CVPR (2012) +1. Graham, B., Engelcke, M., Van Der Maaten, L.: 3d semantic segmentation with submanifold sparse convolutional networks. In: CVPR (2018) +2. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017) +3. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022) +4. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020) +5. Ho, J., Saharia, C., Chan, W., Fleet, D.J., Norouzi, M., Salimans, T.: Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research (2022) + +16. Huang, W., Wang, C., Zhang, R., Li, Y., Wu, J., Fei-Fei, L.: Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973 (2023) +17. Lan, Z., Yew, Z.J., Lee, G.H.: Robust point cloud based reconstruction of large-scale outdoor scenes. In: CVPR (2019) +18. Lee, J., Im, W., Lee, S., Yoon, S.E.: Diffusion probabilistic models for scene-scale 3d categorical data. arXiv preprint arXiv:2301.00527 (2023) +19. Li, X., Li, C., Tong, Z., Lim, A., Yuan, J., Wu, Y., Tang, J., Huang, R.: Campus3d: A photogrammetry point cloud benchmark for hierarchical understanding of outdoor scene. In: ACM MM (2020) +20. Li, Y., Ma, L., Zhong, Z., Liu, F., Chapman, M.A., Cao, D., Li, J.: Deep learning for lidar point clouds in autonomous driving: A review. NeurIPS (2020) +21. Li, Y., Yu, A.W., Meng, T., Caine, B., Ngiam, J., Peng, D., Shen, J., Lu, Y., Zhou, D., Le, Q.V., et al.: Deepfusion: Lidar-camera deep fusion for multi-modal 3d object detection. In: CVPR (2022) +22. Li, Z., Wang, Q., Snavely, N., Kanazawa, A.: Infinitenature-zero: Learning perpetual view generation of natural scenes from single images. In: ECCV (2022) +23. Lin, C.H., Lee, H.Y., Menapace, W., Chai, M., Siarohin, A., Yang, M.H., Tulyakov, S.: Infinicity: Infinite-scale city synthesis. arXiv preprint arXiv:2301.09637 (2023) +24. Liu, M., Shi, R., Chen, L., Zhang, Z., Xu, C., Wei, X., Chen, H., Zeng, C., Gu, J., Su, H.: One-2-3-45++: Fast single image to 3d objects with consistent multi-view generation and 3d diffusion. arXiv preprint arXiv:2311.07885 (2023) +25. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017) +26. Luo, S., Hu, W.: Diffusion probabilistic models for 3d point cloud generation. In: CVPR (2021) +27. Ma, Q., Yang, J., Tang, S., Black, M.J.: The power of points for modeling humans in clothing. In: ICCV (2021) +28. Mascaro, R., Teixeira, L., Chli, M.: Diffuser: Multi-view 2d-to-3d label diffusion for semantic scene segmentation. In: ICRA (2021) +29. Moon, T., Choi, M., Lee, G., Ha, J.W., Lee, J.: Fine-tuning diffusion models with limited data. In: NeurIPS 2022 Workshop on Score-Based Methods (2022) +30. Moro, S., Komuro, T.: Generation of virtual reality environment based on 3d scanned indoor physical space. In: ISVC (2021) +31. Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: ICML (2021) +32. Ögün, M.N., Kurul, R., Yasar, M.F., Turkoglu, S.A., Avci, Ş., Yildiz, N.: Effect of leap motion-based 3d immersive virtual reality usage on upper extremity function in ischemic stroke patients. Arquivos de neuro-psiquiatria (2019) +33. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv (2022) +34. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: NeurIPS (2017) +35. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022) +36. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) +37. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS (2022) + +38. Sohl-Dickstein, J., Weiss, E.A., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. arXiv preprint arXiv:1503.03585 (2015) +39. Sra, M., Garrido-Jurado, S., Schmandt, C., Maes, P.: Procedurally generated virtual reality from 3d reconstructed physical space. In: Proceedings of the 22nd ACM Conference on Virtual Reality Software and Technology (2016) +40. Su, S.Y., Bagautdinov, T., Rhodin, H.: Npc: Neural point characters from video. In: ICCV (2023) +41. Tang, J., Nie, Y., Markhasin, L., Dai, A., Thies, J., Nießner, M.: Diffuscene: Scene graph denoising diffusion probabilistic model for generative indoor scene synthesis. arXiv preprint arXiv:2303.14207 (2023) +42. Tang, Y., He, H., Wang, Y., Mao, Z., Wang, H.: Multi-modality 3d object detection in autonomous driving: A review. Neurocomputing (2023) +43. Wilson, J., Song, J., Fu, Y., Zhang, A., Capodieci, A., Jayakumar, P., Barton, K., Ghaffari, M.: Motionsc: Data set and network for real-time semantic mapping in dynamic environments. IEEE Robotics and Automation Letters 7(3) (2022) +44. Wu, P., Jia, X., Chen, L., Yan, J., Li, H., Qiao, Y.: Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. In: NeurIPS (2022) +45. Xie, H., Chen, Z., Hong, F., Liu, Z.: Citydreamer: Compositional generative model of unbounded 3d cities. arXiv preprint arXiv:2309.00610 (2023) +46. Xu, Z., He, Z., Wu, J., Song, S.: Learning 3d dynamic scene representations for robot manipulation. arXiv preprint arXiv:2011.01968 (2020) +47. Zeng, X., Vahdat, A., Williams, F., Gojcic, Z., Litany, O., Fidler, S., Kreis, K.: Lion: Latent point diffusion models for 3d shape generation. In: NeurIPS (2022) +48. Zheng, H., Nie, W., Vahdat, A., Anandkumar, A.: Fast training of diffusion models with masked transformers. arXiv preprint arXiv:2306.09305 (2023) +49. Zheng, Y., Yifan, W., Wetzstein, G., Black, M.J., Hilliges, O.: Pointavatar: Deformable point-based head avatars from videos. In: CVPR (2023) +50. Zhou, L., Du, Y., Wu, J.: 3d shape generation and completion through point-voxel diffusion. In: ICCV (2021) +51. Zhuang, F., Qi, Z., Duan, K., Xi, D., Zhu, Y., Zhu, H., Xiong, H., He, Q.: A comprehensive survey on transfer learning. arXiv preprint arXiv:1911.02685 (2020) \ No newline at end of file diff --git a/2024/Pyramid Diffusion for Fine 3D Large Scene Generation/images.zip b/2024/Pyramid Diffusion for Fine 3D Large Scene Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..abe9f97f03628ef3b8150c0f54263466562f2d35 --- /dev/null +++ b/2024/Pyramid Diffusion for Fine 3D Large Scene Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0d5f2af9778db0b8c36f006581eea2cd1d2564984567e21ab95c6b51c8b78e84 +size 676317 diff --git a/2024/Pyramid Diffusion for Fine 3D Large Scene Generation/layout.json b/2024/Pyramid Diffusion for Fine 3D Large Scene Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7c25811cb9410e83fc45183a5736c31772c7aaf2 --- /dev/null +++ b/2024/Pyramid Diffusion for Fine 3D Large Scene Generation/layout.json @@ -0,0 +1,9235 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 156, + 111, + 459, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 111, + 459, + 146 + ], + "spans": [ + { + "bbox": [ + 156, + 111, + 459, + 146 + ], + "type": "text", + "content": "Pyramid Diffusion for Fine 3D Large Scene Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 178, + 167, + 434, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 167, + 434, + 194 + ], + "spans": [ + { + "bbox": [ + 178, + 167, + 434, + 194 + ], + "type": "text", + "content": "Yuheng Liu" + }, + { + "bbox": [ + 178, + 167, + 434, + 194 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 178, + 167, + 434, + 194 + ], + "type": "text", + "content": ", Xinke Li" + }, + { + "bbox": [ + 178, + 167, + 434, + 194 + ], + "type": "inline_equation", + "content": "^{3*}" + }, + { + "bbox": [ + 178, + 167, + 434, + 194 + ], + "type": "text", + "content": ", Xueting Li" + }, + { + "bbox": [ + 178, + 167, + 434, + 194 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 178, + 167, + 434, + 194 + ], + "type": "text", + "content": ", Lu Qi" + }, + { + "bbox": [ + 178, + 167, + 434, + 194 + ], + "type": "inline_equation", + "content": "^{5\\dagger}" + }, + { + "bbox": [ + 178, + 167, + 434, + 194 + ], + "type": "text", + "content": ", Chongshou Li" + }, + { + "bbox": [ + 178, + 167, + 434, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 178, + 167, + 434, + 194 + ], + "type": "text", + "content": ", Ming-Hsuan Yang" + }, + { + "bbox": [ + 178, + 167, + 434, + 194 + ], + "type": "inline_equation", + "content": "^{5,6}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 201, + 485, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 201, + 485, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 201, + 485, + 225 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 201, + 485, + 225 + ], + "type": "text", + "content": " Southwest Jiaotong University " + }, + { + "bbox": [ + 132, + 201, + 485, + 225 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 201, + 485, + 225 + ], + "type": "text", + "content": " University of Leeds " + }, + { + "bbox": [ + 132, + 201, + 485, + 225 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 132, + 201, + 485, + 225 + ], + "type": "text", + "content": " City University of Hong Kong " + }, + { + "bbox": [ + 132, + 201, + 485, + 225 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 132, + 201, + 485, + 225 + ], + "type": "text", + "content": " NVIDIA " + }, + { + "bbox": [ + 132, + 201, + 485, + 225 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 132, + 201, + 485, + 225 + ], + "type": "text", + "content": " The University of California, Merced " + }, + { + "bbox": [ + 132, + 201, + 485, + 225 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 132, + 201, + 485, + 225 + ], + "type": "text", + "content": " Yonsei University" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 144, + 258, + 471, + 399 + ], + "blocks": [ + { + "bbox": [ + 144, + 258, + 471, + 399 + ], + "lines": [ + { + "bbox": [ + 144, + 258, + 471, + 399 + ], + "spans": [ + { + "bbox": [ + 144, + 258, + 471, + 399 + ], + "type": "image", + "image_path": "5af58c26ba66a3ae081c745e97e1401a334775a1449f6987b379e03b46fce0a7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 412, + 482, + 468 + ], + "lines": [ + { + "bbox": [ + 130, + 412, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 482, + 468 + ], + "type": "text", + "content": "Fig. 1: We present Pyramid Discrete Diffusion Model, a method that progresses from generating coarse- to fine-grained scenes, mirroring the top-down sequence of the pyramid structure shown. The model is extended for cross-dataset and infinite scene generation, with detailed scene intricacies illustrated on the flanking sides of the image. " + }, + { + "bbox": [ + 130, + 412, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s" + }, + { + "bbox": [ + 130, + 412, + 482, + 468 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 412, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t" + }, + { + "bbox": [ + 130, + 412, + 482, + 468 + ], + "type": "text", + "content": " refer to a source dataset and a target dataset, respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 514, + 453, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 514, + 453, + 635 + ], + "spans": [ + { + "bbox": [ + 159, + 514, + 453, + 635 + ], + "type": "text", + "content": "Abstract. Diffusion models have shown remarkable results in generating 2D images and small-scale 3D objects. However, their application to the synthesis of large-scale 3D scenes has been rarely explored. This is mainly due to the inherent complexity and bulky size of 3D scenery data, particularly outdoor scenes, and the limited availability of comprehensive real-world datasets, which makes training a stable scene diffusion model challenging. In this work, we explore how to effectively generate large-scale 3D scenes using the coarse-to-fine paradigm. We introduce a framework, the Pyramid Discrete Diffusion model (PDD), which employs scale-varied diffusion models to progressively generate high-quality outdoor scenes. Experimental results of PDD demonstrate" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 142, + 643, + 231, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 643, + 231, + 653 + ], + "spans": [ + { + "bbox": [ + 142, + 643, + 231, + 653 + ], + "type": "text", + "content": "* Equal contribution." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 142, + 653, + 242, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 653, + 242, + 665 + ], + "spans": [ + { + "bbox": [ + 142, + 653, + 242, + 665 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 159, + 117, + 455, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 117, + 455, + 173 + ], + "spans": [ + { + "bbox": [ + 159, + 117, + 455, + 173 + ], + "type": "text", + "content": "our successful exploration in generating 3D scenes both unconditionally and conditionally. We further showcase the data compatibility of the PDD model, due to its multi-scale architecture: a PDD model trained on one dataset can be easily fine-tuned with another dataset. Code is available at https://github.com/yuhengliu02/pyramid-discrete-diffusion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 182, + 454, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 182, + 454, + 194 + ], + "spans": [ + { + "bbox": [ + 160, + 182, + 454, + 194 + ], + "type": "text", + "content": "Keywords: 3D Scene Generation " + }, + { + "bbox": [ + 160, + 182, + 454, + 194 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 182, + 454, + 194 + ], + "type": "text", + "content": " Diffusion Models " + }, + { + "bbox": [ + 160, + 182, + 454, + 194 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 182, + 454, + 194 + ], + "type": "text", + "content": " Transfer Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 212, + 230, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 212, + 230, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 212, + 230, + 224 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 235, + 481, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 235, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 235, + 481, + 319 + ], + "type": "text", + "content": "3D scene generation is the task of creating digital representations that mimic the three-dimensional complexities of our real-world environment, allowing for a more nuanced understanding of the tangible surroundings. This technique plays an essential role in fundamental computer vision tasks such as autonomous driving [21,42,44], virtual reality [30,32,39], and robotic manipulation [8,16,46]. However, high-quality large 3D scenes are extremely challenging to synthesize due to their inherently bulky size, and the lack of large-scale 3D scene datasets [43]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 319, + 481, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 481, + 450 + ], + "type": "text", + "content": "Meanwhile, recent advances in diffusion models have shown impressive results in generating 2D images [35-37] or small-scale 3D objects [24, 33]. Yet, it is not a trivial task to employ diffusion models in 3D scene generation. On one hand, state-of-the-art diffusion models leave substantial memory footprints and demand considerable training time, posing a particular challenge when generating 3D scenes with large scales and intricate details. On the other hand, diffusion models require a large amount of training data [29, 48], while capturing large-scale 3D scenes is itself a challenging and ongoing research topic [4, 23]. As a result, only a few attempts have been made to apply diffusion models directly to 3D outdoor scenes [18], which resulted in unstable generation and thus suboptimal performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 450, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 481, + 533 + ], + "type": "text", + "content": "To resolve these challenges, existing works focus on conditional generation and resort to additional signals such as Scene Graphs [41] or 2D maps [28] for guidance. Nonetheless, such conditional guidance is not always accessible, thereby restricting the generalizability of these approaches. Inspired by the coarse-to-fine philosophy widely used in image super-resolution [14, 31, 38], we introduce the Pyramid Discrete Diffusion model (PDD), a framework that progressively generates large 3D scenes without relying on additional guidance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": "We begin by generating small-scale 3D scenes and progressively increase the scale. At each scale level, we learn a separate diffusion model. This model takes the generated scene from the previous scale as a condition (except for the first diffusion model which takes noise as input) and synthesizes a 3D scene of a larger scale. Intuitively, this multi-scale generation process breaks down a challenging unconditional generation task (i.e., high-quality 3D scene generation) into several more manageable conditional generation tasks. This separation allows each diffusion model to specialize in generating either coarse structure (smaller scale) or intricate details (larger scale). Moreover, at the highest scale, we employ a technique known as scene subdivision, which involves dividing a large scene into multiple smaller segments that are synthesized using a shared diffusion model." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "This approach mitigates the issues of oversized models caused by the bulky size of the 3D scenes. Additionally, a noteworthy outcome of our multi-scale design is its capacity to facilitate cross-data transfer applications, which substantially reduces training resources. Lastly, we further propose a natural extension of our PDD framework with scene subdivision for infinite 3D scene generation, thereby demonstrating the scalability of the proposed method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 147, + 189, + 372, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 189, + 372, + 199 + ], + "spans": [ + { + "bbox": [ + 147, + 189, + 372, + 199 + ], + "type": "text", + "content": "The main contributions of this work are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 206, + 480, + 312 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 206, + 479, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 479, + 229 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 479, + 229 + ], + "type": "text", + "content": "- We practically implement a coarse-to-fine strategy for 3D outdoor scene generation via designing a novel pyramid diffusion model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 230, + 480, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 230, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 230, + 480, + 277 + ], + "type": "text", + "content": "- We conduct extensive experiments on our pyramid diffusion, demonstrating its generation of higher quality 3D scenes with comparable computational resources of existing approaches. In addition, we introduce new metrics to evaluate the quality of 3D scene generation from various perspectives." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 277, + 479, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 479, + 312 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 479, + 312 + ], + "type": "text", + "content": "- Our proposed method showcases broader applications, enabling the generation of scenes from synthetic datasets to real-world data. Furthermore, our approach can be extended to facilitate the creation of infinite scenes." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 331, + 237, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 237, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 237, + 342 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 354, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 510 + ], + "type": "text", + "content": "Diffusion Models for 2D Images. Recent advancements in the generative model have seen the diffusion models [14, 31, 38] rise to prominence, especially in applications in 2D image creation [9, 35, 36]. In order to generate high-fidelity images via diffusion models, a multi-stage diffusion process is proposed and employed as per [13, 15, 37]. This process starts with the generation of a coarse-resolution image using an initial diffusion model. Subsequently, a second diffusion model takes this initial output as input, refining it into a finer-resolution image. These cascaded diffusions can be iteratively applied to achieve the desired image resolution. We note that the generation of fine-grained 3D data presents more challenges than 2D due to the addition of an extra dimension. Consequently, our work is motivated by the aforementioned multistage 2D approaches to explore their applicability in 3D contexts. Furthermore, we aim to leverage the advantages of this structure to address the scarcity of datasets in 3D scenes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Diffusion Models for 3D Generation. As a sparse and memory-efficient representation, 3D point clouds has been widely used in various computer vision applications such as digital human [27, 40, 49], autonomous driving [20], and 3D scene reconstruction [17]. Point clouds generation aims to synthesize a 3D point clouds from a random noise [5,6], or scanned lidar points [18]. Though the memory efficiency of point clouds is a valuable property, it poses high challenges in the task of point cloud generation. Existing works largely focus on using Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), or Vector Quantized Variational Autoencoders (VQ-VAEs) as the backbone for this task [1,5,6]. However, these models have limited capacity for high-fidelity generation and are notoriously known for unstable training. As an alternative to the generative models discussed above, diffusion models have revolutionized the computer vision community with their impressive performance in 2D image" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "type": "text", + "content": "Pyramid Diffusion for Fine 3D Large Scene Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 161, + 118, + 471, + 251 + ], + "blocks": [ + { + "bbox": [ + 161, + 118, + 471, + 251 + ], + "lines": [ + { + "bbox": [ + 161, + 118, + 471, + 251 + ], + "spans": [ + { + "bbox": [ + 161, + 118, + 471, + 251 + ], + "type": "image", + "image_path": "693ab21859dc9bb04e9636a414360c8969b6fae8a9325b8b64c77f92e2d4b569.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 261, + 482, + 327 + ], + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 327 + ], + "type": "text", + "content": "Fig. 2: Framework of the proposed Pyramid Discrete Diffusion model. In our structure, there are three different scales. Scenes generated by a previous scale can serve as a condition for the current scale after processing through our scale adaptive function. Furthermore, for the final scale processing, the scene from the previous scale is subdivided into four sub-scenes. The final scene is reconstructed into a large scene using our Scene Subdivision module." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 384, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 540 + ], + "type": "text", + "content": "generation [35-37]. Yet, applying diffusion models for 3D point cloud generation has not been thoroughly explored hitherto. Point-Voxel Diffusion [50] proposes to generate a raw point cloud through the diffusion process while LION [47] and DPM [26] use the latent representation of a point cloud during the denoising process. However, all these methods focus on object-level point clouds and cannot be naively extended to scene-level point clouds. Most relevant to our work is [18], where a diffusion model is trained on a scene-level point cloud dataset for the synthesis task. However, due to the capacity limitation of diffusion models, generating a scene-level point cloud with a single diffusion model leads to unsatisfying results, such as undesired wholes or the lack of fine-grained objects. In this work, we propose a pyramid discrete diffusion model that reduces the difficulty at each pyramid level, thus producing scene point clouds with more realistic and fine-grained details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "content": "3D Large-scale Scene Generation. Generating large-scale 3D scenes is an important but highly challenging task. A generative model on 3D scenes potentially provides infinite training data for tasks such as scene segmentation, autonomous driving, etc. Existing works [4,22,23,45] simplify this task by first generating 2D scenes and then \"lifting\" them to 3D. Though such design is efficient for city scenes populated with regular geometries (e.g., buildings), it does not generalize easily to scenes with more fine-grained objects (e.g., pedestrians, cars, trees, etc.) In this paper, we directly generate 3D outdoor scenes using diffusion models, which include abundant small objects with semantics." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 212, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 212, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 212, + 129 + ], + "type": "text", + "content": "3 Approach" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 147, + 482, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 147, + 482, + 219 + ], + "spans": [ + { + "bbox": [ + 130, + 147, + 482, + 219 + ], + "type": "text", + "content": "The proposed Pyramid Discrete Diffusion (PDD) model comprises multi-scale models capable of step-by-step generation of high-quality 3D scenes from smaller scales. The PDD first extends the standard discrete diffusion for 3D data (Section 3.2) and then proposes a scene subdivision method to further reduce memory requirements (Section 3.3). Finally, we demonstrate two practical applications of PDD in specific scenarios (Section 3.4)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 243, + 251, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 243, + 251, + 254 + ], + "spans": [ + { + "bbox": [ + 131, + 243, + 251, + 254 + ], + "type": "text", + "content": "3.1 Discrete Diffusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 269, + 493, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 493, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 493, + 316 + ], + "type": "text", + "content": "We focus on learning a data distribution based on 3D semantic scenes. Specifically, the semantic scene is represented in a one-hot format, i.e., " + }, + { + "bbox": [ + 130, + 269, + 493, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\{0,1\\}^{h \\times w \\times d \\times c}" + }, + { + "bbox": [ + 130, + 269, + 493, + 316 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 269, + 493, + 316 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 269, + 493, + 316 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 269, + 493, + 316 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 269, + 493, + 316 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 269, + 493, + 316 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 269, + 493, + 316 + ], + "type": "text", + "content": " indicate the dimensions of the scene, respectively, and " + }, + { + "bbox": [ + 130, + 269, + 493, + 316 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 269, + 493, + 316 + ], + "type": "text", + "content": " denotes the size of the one-hot label." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 318, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 482, + 403 + ], + "type": "text", + "content": "Discrete diffusion [2] has been proposed to generate discrete data including semantic scenes. It involves applying the Markov transition matrix on discrete states for noise diffusion. In the forward process, an original scene " + }, + { + "bbox": [ + 130, + 318, + 482, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_0" + }, + { + "bbox": [ + 130, + 318, + 482, + 403 + ], + "type": "text", + "content": " is gradually corrupted into a " + }, + { + "bbox": [ + 130, + 318, + 482, + 403 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 318, + 482, + 403 + ], + "type": "text", + "content": "-step noised map " + }, + { + "bbox": [ + 130, + 318, + 482, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t" + }, + { + "bbox": [ + 130, + 318, + 482, + 403 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 318, + 482, + 403 + ], + "type": "inline_equation", + "content": "t = 1,\\dots ,T" + }, + { + "bbox": [ + 130, + 318, + 482, + 403 + ], + "type": "text", + "content": ". Each forward step can be defined by a Markov uniform transition matrix " + }, + { + "bbox": [ + 130, + 318, + 482, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t" + }, + { + "bbox": [ + 130, + 318, + 482, + 403 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 318, + 482, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t = \\mathbf{X}_{t - 1}\\mathbf{Q}_t" + }, + { + "bbox": [ + 130, + 318, + 482, + 403 + ], + "type": "text", + "content": ". Based on the Markov property, we can derive the " + }, + { + "bbox": [ + 130, + 318, + 482, + 403 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 318, + 482, + 403 + ], + "type": "text", + "content": "-step scene " + }, + { + "bbox": [ + 130, + 318, + 482, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t" + }, + { + "bbox": [ + 130, + 318, + 482, + 403 + ], + "type": "text", + "content": " straight from " + }, + { + "bbox": [ + 130, + 318, + 482, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_0" + }, + { + "bbox": [ + 130, + 318, + 482, + 403 + ], + "type": "text", + "content": " with a cumulative transition matrix " + }, + { + "bbox": [ + 130, + 318, + 482, + 403 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{Q}}_t = \\mathbf{Q}_1\\mathbf{Q}_2\\dots \\mathbf{Q}_t" + }, + { + "bbox": [ + 130, + 318, + 482, + 403 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 226, + 413, + 481, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 413, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 226, + 413, + 481, + 430 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {X} _ {t} \\mid \\mathbf {X} _ {0}\\right) = \\operatorname {C a t} \\left(\\mathbf {X} _ {t}; \\mathbf {P} = \\mathbf {X} _ {0} \\bar {\\mathbf {Q}} _ {t}\\right), \\tag {1}", + "image_path": "c0b464e74dcac1f9ca531cd7b987ccf9725218a47fbcc9673cc6fa5e0535c69a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 439, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 439, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 131, + 439, + 482, + 518 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 439, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\operatorname{Cat}(\\mathbf{X};\\mathbf{P})" + }, + { + "bbox": [ + 131, + 439, + 482, + 518 + ], + "type": "text", + "content": " is a multivariate categorical distribution over the one-hot semantic labels " + }, + { + "bbox": [ + 131, + 439, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 131, + 439, + 482, + 518 + ], + "type": "text", + "content": " with probabilities given by " + }, + { + "bbox": [ + 131, + 439, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 131, + 439, + 482, + 518 + ], + "type": "text", + "content": ". Finally, the semantic scene " + }, + { + "bbox": [ + 131, + 439, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_T" + }, + { + "bbox": [ + 131, + 439, + 482, + 518 + ], + "type": "text", + "content": " at the last step " + }, + { + "bbox": [ + 131, + 439, + 482, + 518 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 439, + 482, + 518 + ], + "type": "text", + "content": " is supposed to be in the form of a uniform discrete noise. In the reverse process, a learnable model parametrized by " + }, + { + "bbox": [ + 131, + 439, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 131, + 439, + 482, + 518 + ], + "type": "text", + "content": " is used to predict denoised semantic labels by " + }, + { + "bbox": [ + 131, + 439, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\tilde{p}_{\\theta}\\left(\\tilde{\\mathbf{X}}_0\\mid \\mathbf{X}_t\\right)" + }, + { + "bbox": [ + 131, + 439, + 482, + 518 + ], + "type": "text", + "content": ". The reparametrization trick is applied subsequently to get the reverse process " + }, + { + "bbox": [ + 131, + 439, + 482, + 518 + ], + "type": "inline_equation", + "content": "p_{\\theta}\\left(\\mathbf{X}_{t - 1}\\mid \\mathbf{X}_t\\right)" + }, + { + "bbox": [ + 131, + 439, + 482, + 518 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 201, + 529, + 481, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 529, + 481, + 550 + ], + "spans": [ + { + "bbox": [ + 201, + 529, + 481, + 550 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\mathbf {X} _ {t - 1} \\mid \\mathbf {X} _ {t}\\right) = \\mathbb {E} _ {\\tilde {p} _ {\\theta} \\left(\\tilde {\\mathbf {X}} _ {0} \\mid \\mathbf {X} _ {t}\\right)} q \\left(\\mathbf {X} _ {t - 1} \\mid \\mathbf {X} _ {t}, \\tilde {\\mathbf {X}} _ {0}\\right). \\tag {2}", + "image_path": "a61aff49910aa9e62ca62848bf67323a87b689cb7d8c90a81628e28ffe38b174.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 559, + 481, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 559, + 481, + 584 + ], + "spans": [ + { + "bbox": [ + 131, + 559, + 481, + 584 + ], + "type": "text", + "content": "A loss consisting of the two KL divergences is proposed to learn better reconstruction ability for the model, given by" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 210, + 595, + 480, + 631 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 595, + 480, + 631 + ], + "spans": [ + { + "bbox": [ + 210, + 595, + 480, + 631 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\theta} = d _ {\\mathrm {K L}} \\left(q \\left(\\mathbf {X} _ {t - 1} \\mid \\mathbf {X} _ {t}, \\mathbf {X} _ {0}\\right) \\| p _ {\\theta} \\left(\\mathbf {X} _ {t - 1} \\mid \\mathbf {X} _ {t}\\right)\\right) \\tag {3} \\\\ + \\lambda d _ {\\mathrm {K L}} \\left(q \\left(\\mathbf {X} _ {0}\\right) \\| \\tilde {p} _ {\\theta} \\left(\\tilde {\\mathbf {X}} _ {0} \\mid \\mathbf {X} _ {t}\\right)\\right), \\\\ \\end{array}", + "image_path": "e88c3fa16da01309729276ef7f7ecd1f5676b7a58f23a4d2007f5d7375bc2742.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": " is an auxiliary loss weight and " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{KL}}" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": " stands for KL divergence. In the following, we focus on extending the discrete diffusion into the proposed PDD." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 222, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 222, + 91, + 447, + 102 + ], + "type": "text", + "content": "Pyramid Diffusion for Fine 3D Large Scene Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 298, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 298, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 298, + 129 + ], + "type": "text", + "content": "3.2 Pyramid Discrete Diffusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 497, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 497, + 245 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 497, + 245 + ], + "type": "text", + "content": "We propose PDD that operates various diffusion processes across multiple scales (or resolutions), as depicted in Figure 2. Given a 3D scene data " + }, + { + "bbox": [ + 130, + 137, + 497, + 245 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} \\in \\{0,1\\}^{h \\times w \\times d \\times c}" + }, + { + "bbox": [ + 130, + 137, + 497, + 245 + ], + "type": "text", + "content": ", we define a 3D pyramid including different scales of " + }, + { + "bbox": [ + 130, + 137, + 497, + 245 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 130, + 137, + 497, + 245 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 137, + 497, + 245 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Z}^{(1)}, \\dots, \\mathbf{Z}^{(l)}, \\dots, \\mathbf{Z}^{(L)}\\}" + }, + { + "bbox": [ + 130, + 137, + 497, + 245 + ], + "type": "text", + "content": ", where a larger " + }, + { + "bbox": [ + 130, + 137, + 497, + 245 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 137, + 497, + 245 + ], + "type": "text", + "content": " indicates a larger scene scale. Formally, let " + }, + { + "bbox": [ + 130, + 137, + 497, + 245 + ], + "type": "inline_equation", + "content": "h_l \\times w_l \\times d_l \\times c" + }, + { + "bbox": [ + 130, + 137, + 497, + 245 + ], + "type": "text", + "content": " denote the dimension of " + }, + { + "bbox": [ + 130, + 137, + 497, + 245 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^{(l)}" + }, + { + "bbox": [ + 130, + 137, + 497, + 245 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 137, + 497, + 245 + ], + "type": "inline_equation", + "content": "h_{l+1} \\geq h_l" + }, + { + "bbox": [ + 130, + 137, + 497, + 245 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 137, + 497, + 245 + ], + "type": "inline_equation", + "content": "w_{l+1} \\geq w_l" + }, + { + "bbox": [ + 130, + 137, + 497, + 245 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 137, + 497, + 245 + ], + "type": "inline_equation", + "content": "d_{l+1} \\geq d_l" + }, + { + "bbox": [ + 130, + 137, + 497, + 245 + ], + "type": "text", + "content": " are kept for " + }, + { + "bbox": [ + 130, + 137, + 497, + 245 + ], + "type": "inline_equation", + "content": "l = 1, \\dots, L-1" + }, + { + "bbox": [ + 130, + 137, + 497, + 245 + ], + "type": "text", + "content": ". We note that such a pyramid can be obtained by applying different down-sample operators, such as pooling functions, on " + }, + { + "bbox": [ + 130, + 137, + 497, + 245 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 130, + 137, + 497, + 245 + ], + "type": "text", + "content": ". For each scale in the pyramid, we construct a conditional discrete diffusion model parameterized by " + }, + { + "bbox": [ + 130, + 137, + 497, + 245 + ], + "type": "inline_equation", + "content": "\\theta_l" + }, + { + "bbox": [ + 130, + 137, + 497, + 245 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 130, + 137, + 497, + 245 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 137, + 497, + 245 + ], + "type": "text", + "content": "-th model for " + }, + { + "bbox": [ + 130, + 137, + 497, + 245 + ], + "type": "inline_equation", + "content": "l \\neq 1" + }, + { + "bbox": [ + 130, + 137, + 497, + 245 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 162, + 253, + 481, + 274 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 253, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 162, + 253, + 481, + 274 + ], + "type": "interline_equation", + "content": "\\tilde {p} _ {\\theta_ {l}} \\left(\\tilde {\\mathbf {X}} _ {0} ^ {(l)} \\mid \\mathbf {X} _ {t} ^ {(l)}, \\mathbf {Z} ^ {(l - 1)}\\right) = \\tilde {p} _ {\\theta_ {l}} \\left(\\tilde {\\mathbf {X}} _ {0} ^ {(l)} \\mid \\operatorname {C o n c a t} \\left(\\mathbf {X} _ {t} ^ {(l)}, \\phi^ {(l)} \\left(\\mathbf {Z} ^ {(l - 1)}\\right)\\right)\\right), \\tag {4}", + "image_path": "f33f4c49e438c5756721b33f675120a1c23206035580614c88271d21c0359128.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 283, + 480, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 480, + 332 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 283, + 480, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t^{(l)}" + }, + { + "bbox": [ + 130, + 283, + 480, + 332 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 283, + 480, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_0^{(l)}" + }, + { + "bbox": [ + 130, + 283, + 480, + 332 + ], + "type": "text", + "content": " are with the same size of " + }, + { + "bbox": [ + 130, + 283, + 480, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^{(l)}" + }, + { + "bbox": [ + 130, + 283, + 480, + 332 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 283, + 480, + 332 + ], + "type": "inline_equation", + "content": "\\phi^{(l)}" + }, + { + "bbox": [ + 130, + 283, + 480, + 332 + ], + "type": "text", + "content": " is a Scale Adaptive Function (SAF) for upsampling " + }, + { + "bbox": [ + 130, + 283, + 480, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^{(l-1)}" + }, + { + "bbox": [ + 130, + 283, + 480, + 332 + ], + "type": "text", + "content": " into the size of " + }, + { + "bbox": [ + 130, + 283, + 480, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^{(l)}" + }, + { + "bbox": [ + 130, + 283, + 480, + 332 + ], + "type": "text", + "content": ". As a case in point, SAF can be a tri-linear interpolation function depending on the data. Additionally, we maintain the first model " + }, + { + "bbox": [ + 130, + 283, + 480, + 332 + ], + "type": "inline_equation", + "content": "\\tilde{p}_{\\theta_1}" + }, + { + "bbox": [ + 130, + 283, + 480, + 332 + ], + "type": "text", + "content": " as the original non-conditional model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 333, + 481, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 481, + 405 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 481, + 405 + ], + "type": "text", + "content": "During the training process, PDD learns " + }, + { + "bbox": [ + 130, + 333, + 481, + 405 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 333, + 481, + 405 + ], + "type": "text", + "content": " denoising models separately at varied scales of scene pyramids in the given dataset. Given that " + }, + { + "bbox": [ + 130, + 333, + 481, + 405 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^{(l - 1)}" + }, + { + "bbox": [ + 130, + 333, + 481, + 405 + ], + "type": "text", + "content": " is essentially a lossy-compressed version of " + }, + { + "bbox": [ + 130, + 333, + 481, + 405 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^{(l)}" + }, + { + "bbox": [ + 130, + 333, + 481, + 405 + ], + "type": "text", + "content": ", the model training can be viewed as learning to restore the details of a coarse scene. In the inference process, denoising model " + }, + { + "bbox": [ + 130, + 333, + 481, + 405 + ], + "type": "inline_equation", + "content": "p_{\\theta_1}" + }, + { + "bbox": [ + 130, + 333, + 481, + 405 + ], + "type": "text", + "content": " is performed initially according to Equation (2) and the rest of PDD models are executed in sequence from " + }, + { + "bbox": [ + 130, + 333, + 481, + 405 + ], + "type": "inline_equation", + "content": "l = 2" + }, + { + "bbox": [ + 130, + 333, + 481, + 405 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 333, + 481, + 405 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 333, + 481, + 405 + ], + "type": "text", + "content": " via the sampling," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 236, + 414, + 480, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 414, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 236, + 414, + 480, + 430 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {t - 1} ^ {(l)} \\sim p _ {\\theta_ {l}} (\\mathbf {X} _ {t - 1} ^ {(l)} \\mid \\mathbf {X} _ {t} ^ {(l)}, \\mathbf {X} _ {0} ^ {(l - 1)}), \\tag {5}", + "image_path": "4fcda69585da422916bca68b87292e0bf04320ca0c7a68d6efa64a8c18742ec9.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 439, + 326, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 439, + 326, + 454 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 326, + 454 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 439, + 326, + 454 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_0^{(l - 1)}" + }, + { + "bbox": [ + 130, + 439, + 326, + 454 + ], + "type": "text", + "content": " is the denoised result of " + }, + { + "bbox": [ + 130, + 439, + 326, + 454 + ], + "type": "inline_equation", + "content": "\\tilde{p}_{\\theta_{l - 1}}" + }, + { + "bbox": [ + 130, + 439, + 326, + 454 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 454, + 481, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 454, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 454, + 481, + 525 + ], + "type": "text", + "content": "Except for the high-quality generation, the proposed PDD bears two merits: 1) Diffusion models in PDD can be trained in parallel due to their independence, which allows for a flexible computation reallocation during training. 2) Due to its multi-stage generation process, PDD is fitting for restoring scenes of arbitrary coarse-grained scale by starting from the intermediate processes, thereby extending the method's versatility." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 544, + 251, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 251, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 251, + 555 + ], + "type": "text", + "content": "3.3 Scene Subdivision" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 564, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 564, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 564, + 482, + 665 + ], + "type": "text", + "content": "To overcome the memory constraint for generating large 3D scenes, we propose the scene subdivision method. We divide a 3D scene " + }, + { + "bbox": [ + 130, + 564, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^{(l)}" + }, + { + "bbox": [ + 130, + 564, + 482, + 665 + ], + "type": "text", + "content": " along " + }, + { + "bbox": [ + 130, + 564, + 482, + 665 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 564, + 482, + 665 + ], + "type": "text", + "content": "-axis into " + }, + { + "bbox": [ + 130, + 564, + 482, + 665 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 564, + 482, + 665 + ], + "type": "text", + "content": " overlapped sub-components as " + }, + { + "bbox": [ + 130, + 564, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Z}_i^{(l)}\\}_{i=1}^I" + }, + { + "bbox": [ + 130, + 564, + 482, + 665 + ], + "type": "text", + "content": ". For the instance of four subscenes case, let " + }, + { + "bbox": [ + 130, + 564, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_i^{(l)} \\in \\{0,1\\}^{(1 + \\delta_l)h_l \\setminus 2 \\times (1 + \\delta_l)w_l \\setminus 2 \\times d_l \\times c}" + }, + { + "bbox": [ + 130, + 564, + 482, + 665 + ], + "type": "text", + "content": " denote one subscene and " + }, + { + "bbox": [ + 130, + 564, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\delta_l" + }, + { + "bbox": [ + 130, + 564, + 482, + 665 + ], + "type": "text", + "content": " denote the overlap ratio, the shared " + }, + { + "bbox": [ + 130, + 564, + 482, + 665 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 564, + 482, + 665 + ], + "type": "text", + "content": "-th diffusion model in PDD is trained to reconstruct " + }, + { + "bbox": [ + 130, + 564, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_i^{(l)}" + }, + { + "bbox": [ + 130, + 564, + 482, + 665 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 564, + 482, + 665 + ], + "type": "inline_equation", + "content": "i = 1, \\dots, 4" + }, + { + "bbox": [ + 130, + 564, + 482, + 665 + ], + "type": "text", + "content": ". Subsequently, sub-scenes are merged into a holistic one by a fusion algorithm, i.e., voting on the overlapped parts to ensure the continuity of the 3D scene." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "In the training process, to ensure context-awareness of the entire scene during the generation of a sub-scene, we train the model by adding the overlapped regions with other sub-scenes as the condition. In the inference process, the entire scene is generated in an autoregressive manner. Apart from the first sub-scene generated without context, all other sub-scenes utilize the already generated overlapped region as a condition, i.e.," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 186, + 196, + 481, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 196, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 186, + 196, + 481, + 236 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {t - 1, i} ^ {(l)} \\sim p _ {\\theta} \\left(\\mathbf {X} _ {t - 1, i} ^ {(l)} \\mid \\mathbf {X} _ {t, i} ^ {(l)}, \\mathbf {X} _ {0, i} ^ {(l + 1)}, \\sum_ {j \\neq i} \\Delta_ {i j} \\odot \\mathbf {X} _ {0, j} ^ {(l + 1)}\\right), \\tag {6}", + "image_path": "759f5a41511701388011eba0a376bcc8ae41f0ca668b48303e21ba27297ef7a2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 243, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 319 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 243, + 482, + 319 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 243, + 482, + 319 + ], + "type": "text", + "content": " is the index of generated sub-scenes before " + }, + { + "bbox": [ + 130, + 243, + 482, + 319 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 243, + 482, + 319 + ], + "type": "text", + "content": "-th scene, and " + }, + { + "bbox": [ + 130, + 243, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\Delta_{ij}" + }, + { + "bbox": [ + 130, + 243, + 482, + 319 + ], + "type": "text", + "content": " is a binary mask between " + }, + { + "bbox": [ + 130, + 243, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{0,i}^{(l+1)}" + }, + { + "bbox": [ + 130, + 243, + 482, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 243, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{0,j}^{(l+1)}" + }, + { + "bbox": [ + 130, + 243, + 482, + 319 + ], + "type": "text", + "content": " representing the overlapped region on " + }, + { + "bbox": [ + 130, + 243, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{0,j}^{(l+1)}" + }, + { + "bbox": [ + 130, + 243, + 482, + 319 + ], + "type": "text", + "content": " with 1 and the separate region with 0. Scene Subdivision module can reduce the model parameters, as diffusion model could be shared by four sub-scenes. In practice, we only implement the scene subdivision method on the largest scale which demands the largest memory." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 337, + 224, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 337, + 224, + 350 + ], + "spans": [ + { + "bbox": [ + 131, + 337, + 224, + 350 + ], + "type": "text", + "content": "3.4 Applications" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 358, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 491 + ], + "type": "text", + "content": "Beyond its primary function as a generative model, we introduce two novel applications for PDD. First, cross-dataset transfer aims at adapting a model trained on a source dataset to a target dataset [51]. Due to the flexibility of input scale, PDD can achieve this by retraining or fine-tuning the smaller-scale models in the new dataset while keeping the larger-scale models. The strategy leveraging PDD improves the efficiency of transferring 3D scene generation models between distinct datasets. Second, infinite scene generation is of great interest in fields such as autonomous driving [10] and urban modeling [19] which require a huge scale of 3D scenes. PDD can extend its scene subdivision technique. By using the edge of a previously generated scene as a condition as in Equation (6), it can iteratively create larger scenes, potentially without size limitations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 510, + 282, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 510, + 282, + 523 + ], + "spans": [ + { + "bbox": [ + 131, + 510, + 282, + 523 + ], + "type": "text", + "content": "4 Experimental Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 536, + 266, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 536, + 266, + 548 + ], + "spans": [ + { + "bbox": [ + 131, + 536, + 266, + 548 + ], + "type": "text", + "content": "4.1 Evaluation Protocols" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 557, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 605 + ], + "type": "text", + "content": "Since the metrics used in 2D generation such as FID [12] are not directly applicable in the 3D, we introduce and implement three metrics to assess the quality of the generated 3D scenes. We note that more implementation details can be found in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": "Semantic Segmentation results on the generated scenes are used to evaluate the effectiveness of models in creating semantically coherent scenes. Specifically, two architectures, the voxel-based SparseUNet [11] and point-based PointNet++ [34], are implemented to perform the segmentation tasks. We report the mean Intersection over Union (mIoU) and Mean Accuracy (MAs) for evaluation." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 222, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 222, + 91, + 447, + 102 + ], + "type": "text", + "content": "Pyramid Diffusion for Fine 3D Large Scene Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 212, + 478, + 301 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 202 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 202 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 202 + ], + "type": "text", + "content": "Table 1: Comparison of various diffusion models on 3D semantic scene generation of CarlaSC. DiscreteDiff [2], LatentDiff [18], and P-DiscreteDiff refer to the original discrete diffusion, latent discrete diffusion, and our approach, respectively. Conditioned models work based on the context of unlabeled point clouds or the coarse version of the ground truth scene. A higher Segmentation Metric value is better, indicating semantic consistency. A lower Feature-based Metric value is preferable, representing closer proximity to the original dataset. The brackets with " + }, + { + "bbox": [ + 130, + 114, + 482, + 202 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 114, + 482, + 202 + ], + "type": "text", + "content": " represent voxel-based network and " + }, + { + "bbox": [ + 130, + 114, + 482, + 202 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 114, + 482, + 202 + ], + "type": "text", + "content": " represent point-based network." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 212, + 478, + 301 + ], + "lines": [ + { + "bbox": [ + 136, + 212, + 478, + 301 + ], + "spans": [ + { + "bbox": [ + 136, + 212, + 478, + 301 + ], + "type": "table", + "html": "
MethodModelConditionSegmentation MetricFeature-based Metric
mIoU (V)MA (V)mIoU (P)MA (P)F3D (↓)MMD (↓)
Ground Truth--52.1972.4032.9047.680.2460.108
UnconditionedDiscreteDiff [2]-40.0563.6525.5438.711.3610.599
LatentDiff [18]-38.0162.3926.6945.870.3310.211
P-DiscreteDiff (Ours)-68.0285.6633.8952.120.3150.200
ConditionedDiscreteDiff [2]Point cloud38.5559.9728.4144.060.3570.261
DiscreteDiff [2]Coarse scene (s1)52.5277.2327.9343.130.3590.284
P-DiscreteDiff (Ours)Coarse scene (s1)55.7578.7029.7846.610.3420.274
", + "image_path": "8d71a08cf6599b15b29865540ce8c0696b32b5aac939bb43fe3711927a106fea.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 321, + 479, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 479, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 479, + 367 + ], + "type": "text", + "content": "F3D is a 3D adaption of the 2D Fréchet Inception Distance (FID) [12], which is based on a pre-trained autoencoder with an 3D CNN architecture. We calculate and report the Fréchet distance (by " + }, + { + "bbox": [ + 130, + 321, + 479, + 367 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 130, + 321, + 479, + 367 + ], + "type": "text", + "content": " ratio) between the generated scenes and real scenes in the feature domain." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 369, + 480, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 480, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 480, + 416 + ], + "type": "text", + "content": "Maximum Mean Discrepancy (MMD) is a statistical measure to quantify the disparity between the distributions of generated and real scenes. Similar to our F3D approach, we extract features via the same pre-trained autoencoder and present the MMD between 3D scenes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 433, + 264, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 433, + 264, + 445 + ], + "spans": [ + { + "bbox": [ + 131, + 433, + 264, + 445 + ], + "type": "text", + "content": "4.2 Experiment Settings" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "type": "text", + "content": "Datasets. We use CarlaSC [43] and SemanticKITTI [3] for experiments. Specifically, we conduct our main experiments as well as ablation studies on the synthesis dataset CarlaSC due to its large data volume and diverse semantic objects. Our primary model is trained on the training set of CarlaSC with 10 categories and 32,400 scans. SemanticKITTI, which is a real-world collected dataset, is used for our cross-dataset transfer experiment. Both datasets are adjusted to ensure consistency in semantic categories, with further details in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": "Model Architecture. The primary proposed PDD is performed on three scales of a 3D scene pyramid, i.e., " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "s_1" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "s_2" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "s_4" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": " in Table 2. We implement 3D-UNets [7] for three diffusion models in PDD based on the scales. Notably, the model applied on " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "s_4" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": " scale is with the input/output size of " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "s_3'" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": " due to the use of scene subdivision, while such a size of other models follows the working scale size. In the ablation study, we also introduce the scale " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "s_3" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": " in the experiment. Additionally, we implement two baseline methods merely on scale " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "s_4" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": " which are the original discrete diffusion [2] and the latent diffusion model with VQ-VAE decoder [18]. Training Setting. We train each PDD model using the same training setting except for the batch size. Specifically, we set the learning rate of " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": " for the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 148, + 280, + 228 + ], + "blocks": [ + { + "bbox": [ + 132, + 124, + 290, + 146 + ], + "lines": [ + { + "bbox": [ + 132, + 124, + 290, + 146 + ], + "spans": [ + { + "bbox": [ + 132, + 124, + 290, + 146 + ], + "type": "text", + "content": "Table 2: Different scales in the 3D scene pyramid." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 148, + 280, + 228 + ], + "lines": [ + { + "bbox": [ + 143, + 148, + 280, + 228 + ], + "spans": [ + { + "bbox": [ + 143, + 148, + 280, + 228 + ], + "type": "table", + "html": "
Scale Rep.3D Scene Size
s132 × 32 × 4
s264 × 64 × 8
s3128 × 128 × 8
s3'136 × 136 × 16
s4256 × 256 × 16
", + "image_path": "d358629d294c2cd5e888458f112762bb982265cb04b46ba63a0d54739bf4ffbb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 328, + 169, + 466, + 229 + ], + "blocks": [ + { + "bbox": [ + 311, + 124, + 481, + 168 + ], + "lines": [ + { + "bbox": [ + 311, + 124, + 481, + 168 + ], + "spans": [ + { + "bbox": [ + 311, + 124, + 481, + 168 + ], + "type": "text", + "content": "Table 3: Average SSIM between each generated scene and the closest scene in the training set. We generate " + }, + { + "bbox": [ + 311, + 124, + 481, + 168 + ], + "type": "inline_equation", + "content": "1k" + }, + { + "bbox": [ + 311, + 124, + 481, + 168 + ], + "type": "text", + "content": " scenes and calculate the average." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 169, + 466, + 229 + ], + "lines": [ + { + "bbox": [ + 328, + 169, + 466, + 229 + ], + "spans": [ + { + "bbox": [ + 328, + 169, + 466, + 229 + ], + "type": "table", + "html": "
DatasetMethodSSIM
CarlaSCUncon-Gen0.72
Val Set0.65
Sem-KITTI (fine-tuned)Uncon-Gen0.74
Val Set0.67
", + "image_path": "949ccdf9d92e95db62fe13634daf2cb6e6b01fd6387ffe1cf0540fa60dd0e219.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 155, + 259, + 465, + 356 + ], + "blocks": [ + { + "bbox": [ + 132, + 239, + 479, + 251 + ], + "lines": [ + { + "bbox": [ + 132, + 239, + 479, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 239, + 479, + 251 + ], + "type": "text", + "content": "Table 4: Comparison of different diffusion pyramids on 3D semantic scene generation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 155, + 259, + 465, + 356 + ], + "lines": [ + { + "bbox": [ + 155, + 259, + 465, + 356 + ], + "spans": [ + { + "bbox": [ + 155, + 259, + 465, + 356 + ], + "type": "table", + "html": "
PyramidConditionedmIoU(V)mIoU(P)F3D(↓)MMD(↓)
s4×40.025.51.360.60
s1 → s4×67.032.10.320.24
s1 → s2 → s4×68.033.90.320.20
s1 → s2 → s3 → s4×68.033.40.320.23
s1 → s4✓52.527.90.360.28
s1 → s2 → s4✓55.829.80.340.27
s1 → s2 → s3 → s4✓55.929.60.340.28
", + "image_path": "b0c2d6dbedfaf808d9adf43f1f810d82d04c842fa8cce266824042b56a500d50.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "type": "text", + "content": "AdamW optimizer [25], and the time step " + }, + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "type": "inline_equation", + "content": "T = 100" + }, + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "type": "text", + "content": " for the diffusion process, and 800 for the max epoch. The batch sizes are set to 128, 32, and 16 for the models working on " + }, + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "type": "inline_equation", + "content": "s_1" + }, + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "type": "inline_equation", + "content": "s_2" + }, + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "type": "inline_equation", + "content": "s_4" + }, + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "type": "text", + "content": " scales. However, for the baseline method based on the " + }, + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "type": "inline_equation", + "content": "s_4" + }, + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "type": "text", + "content": " scale, we use the batch size of 8 due to memory constraints. We note that all diffusion models are trained on four NVIDIA A100 GPUs. In addition, we apply the trilinear interpolation for the scene fusion algorithm and set the overlap ratio in scene subdivision, " + }, + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\delta_l" + }, + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "type": "text", + "content": " to 0.0625." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 479, + 228, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 228, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 228, + 490 + ], + "type": "text", + "content": "4.3 Main Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 498, + 481, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 617 + ], + "type": "text", + "content": "Generation Quality. We compare our approach with two baselines, the original Discrete Diffusion [2] and the Latent Diffusion [18]. The result reported in Table 1 demonstrates the notable performance of our method across all metrics in both unconditional and conditional settings in comparable computational resources with existing method. Our proposed method demonstrates a notable advantage in segmentation tasks, especially when it reaches around " + }, + { + "bbox": [ + 130, + 498, + 481, + 617 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 130, + 498, + 481, + 617 + ], + "type": "text", + "content": " mIoU for SparseUNet, which reflects its ability to generate scenes with accurate semantic coherence. We also provide visualizations of different model results in Figure 3, where the proposed method demonstrates better performance in detail generation and scene diversity for random 3D scene generations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "content": "Additionally, we conduct the comparison on conditioned 3D scene generation. We leverage the flexibility of input scale for our method and perform the generation by models in " + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "s_2" + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "s_4" + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "content": " scales conditioned on a coarse ground truth scene in " + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "s_1" + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "content": " scale. We benchmark our method against the discrete diffusion conditioned on" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "type": "text", + "content": "Pyramid Diffusion for Fine 3D Large Scene Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 478, + 346 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 478, + 346 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 478, + 346 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 478, + 346 + ], + "type": "image", + "image_path": "2ffbc795e494692ef7ddfed08133aed7acbcedabc4b3b4c6782e176a5570ee7e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 358, + 482, + 414 + ], + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 414 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 414 + ], + "type": "text", + "content": "Fig. 3: Visualization of unconditional generation results on CarlaSC. We compare with two baseline models - DiscreteDiff [2] and LatentDiff [18] and show synthesis from our models with different scales. Our method produces more diverse scenes compared to the baseline models. Furthermore, with more levels, our model can synthesize scenes with more intricate details." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 446, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 529 + ], + "type": "text", + "content": "unlabeled point clouds and the same coarse scenes. Results in Table 1 and Figure 5 present the impressive results of our conditional generation comparison. It is also observed that the point cloud-based model can achieve decent performance on F3D and MMD, which could be caused by 3D point conditions providing more structural information about the scene than the coarse scene. Despite the informative condition of the point cloud, our method can still outperform it across most metrics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "None-overfitting Verification. The MMD and F3D metrics (defined in Supp A) numerically illustrate the statistical feature distance between generated scenes and the training set. Our method achieves the lowest MMD and F3D among all baseline methods as shown in Table 1. However, we argue that this does not indicate overfitting to the dataset for the following reasons. First, our MMD and F3D are larger than those of the ground truth. Furthermore, we leverage structural similarity (SSIM) to search and show that a generated scene is different from its nearest neighbour in the dataset. Specifically, we generate " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "1k" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " scenes and identify their closest matches in the training set using the SSIM metric. The average SSIM of these " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "1k" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " scenes are calculated and presented in Table 3. Additionally, we apply the same methodology to the Validation Set to establish an" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 479, + 238 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 479, + 238 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 479, + 238 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 479, + 238 + ], + "type": "image", + "image_path": "30a115586a27a16e20586cab35c9c1d25857903e10df245797c525847bb69ac4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 246, + 482, + 313 + ], + "lines": [ + { + "bbox": [ + 130, + 246, + 482, + 313 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 482, + 313 + ], + "type": "text", + "content": "Fig. 4: Data retrieval visualization. We generate " + }, + { + "bbox": [ + 130, + 246, + 482, + 313 + ], + "type": "inline_equation", + "content": "1k" + }, + { + "bbox": [ + 130, + 246, + 482, + 313 + ], + "type": "text", + "content": " scenes using PDD on CarlaSC dataset, and retrieve the most similar scene in training set for each scene using SSIM (SSIM = 1 means identical), and plot SSIM distribution. Scenes at various percentiles are displayed (red box: generated scenes; grey box: scenes in training set), those with the highest " + }, + { + "bbox": [ + 130, + 246, + 482, + 313 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 246, + 482, + 313 + ], + "type": "text", + "content": " similarity (i.e., " + }, + { + "bbox": [ + 130, + 246, + 482, + 313 + ], + "type": "inline_equation", + "content": "10^{th}" + }, + { + "bbox": [ + 130, + 246, + 482, + 313 + ], + "type": "text", + "content": " percentile of SSIM) are very similar to the training set, but still not completely identical." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 337, + 482, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 337, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 337, + 482, + 433 + ], + "type": "text", + "content": "oracle baseline. Table 3 shows that our generated scenes are comparable to the baseline, verifying that our method does not overfit the training set. To further support this, we use distribution plots, as shown in Figure 4, to validate the similarity between our generated scenes and the training set. We also display three pairs at different percentiles in this figure, showing that scenes with lower SSIM scores differ more from their nearest matches in the training set. This visual evidence reinforces that PDD effectively captures the distribution of the training set instead of merely memorizing it." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 452, + 246, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 452, + 246, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 452, + 246, + 463 + ], + "type": "text", + "content": "4.4 Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 473, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 604 + ], + "type": "text", + "content": "Pyramid Diffusion. Our experiments explore the impact of varying refinement scales on the quality of generated scenes. According to Table 4, both conditional and unconditional scene generation quality show incremental improvements with additional scales. Balancing training overhead and generation quality, a three-scale model with the scale of " + }, + { + "bbox": [ + 130, + 473, + 482, + 604 + ], + "type": "inline_equation", + "content": "s_4" + }, + { + "bbox": [ + 130, + 473, + 482, + 604 + ], + "type": "text", + "content": " progression offers an optimal compromise between performance and computational cost. We find that as the number of scales increases, there is indeed a rise in performance, particularly notable upon the addition of the second scale. However, the progression from a three-scale pyramid to a four-scale pyramid turns out to be insignificant. Given the greater training overhead for a four-scale pyramid compared to a three-scale one, we choose the latter as our main structure." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Scene Subdivision. We explore the optimal mask ratio for scene subdivision and report on Figure 6, which shows an inverse correlation between the mask ratio and the effectiveness of F3D and MMD metrics; higher mask ratios may result in diminished outcomes due to increased overfitting, leading to reduced randomness in the generated results. The lowest mask ratio test, 0.0625, achieves" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "type": "text", + "content": "Pyramid Diffusion for Fine 3D Large Scene Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 113, + 478, + 300 + ], + "blocks": [ + { + "bbox": [ + 134, + 113, + 478, + 300 + ], + "lines": [ + { + "bbox": [ + 134, + 113, + 478, + 300 + ], + "spans": [ + { + "bbox": [ + 134, + 113, + 478, + 300 + ], + "type": "image", + "image_path": "f53f936d2f764098f8fd2f4fd7f1a400309d691717a0b8406959e452efb13d20.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 312, + 482, + 333 + ], + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 333 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 333 + ], + "type": "text", + "content": "Fig. 5: Visualization of conditional generation results on CarlaSC. " + }, + { + "bbox": [ + 130, + 312, + 482, + 333 + ], + "type": "inline_equation", + "content": "PC" + }, + { + "bbox": [ + 130, + 312, + 482, + 333 + ], + "type": "text", + "content": " stands for point cloud condition." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 167, + 388, + 447, + 448 + ], + "blocks": [ + { + "bbox": [ + 130, + 355, + 482, + 378 + ], + "lines": [ + { + "bbox": [ + 130, + 355, + 482, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 355, + 482, + 378 + ], + "type": "text", + "content": "Table 5: Generation results on CarlaSC in different scales on the diffusion pyramid without any conditions. All output scales are lifted to " + }, + { + "bbox": [ + 130, + 355, + 482, + 378 + ], + "type": "inline_equation", + "content": "s_4" + }, + { + "bbox": [ + 130, + 355, + 482, + 378 + ], + "type": "text", + "content": " using the upsampling method." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 388, + 447, + 448 + ], + "lines": [ + { + "bbox": [ + 167, + 388, + 447, + 448 + ], + "spans": [ + { + "bbox": [ + 167, + 388, + 447, + 448 + ], + "type": "table", + "html": "
Model No.ScalemIoU(↑)MA(↑)F3D(↓)MMD(↓)
1s118.042.70.290.16
2s243.766.80.290.18
3s468.085.70.320.23
", + "image_path": "4a50b5423862a924bcb80b7dab4e0943e917c004752b808de0ee149fe234f444.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 480, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 480, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 480, + 482, + 540 + ], + "type": "text", + "content": "the best results across all metrics with detailed retention. Thus, we set a mask ratio of 0.0625 as the standard for our scene subdivision module. Further analysis shows that higher overlap ratios in scene subdivision result in quality deterioration, mainly due to increased discontinuities when merging sub-scenes using scene fusion algorithm." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 546, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 665 + ], + "type": "text", + "content": "Performance on Different Scales. To facilitate a comprehensive understanding of the progressive improvement achieved by our coarse-to-fine method, we evaluate the quality of scenes generated at different scales and present the findings in Table 5. Even at the smaller scale " + }, + { + "bbox": [ + 130, + 546, + 482, + 665 + ], + "type": "inline_equation", + "content": "s_1" + }, + { + "bbox": [ + 130, + 546, + 482, + 665 + ], + "type": "text", + "content": ", we observe high F3D and MMD scores, indicating its capacity in synthesizing scenes with both reasonable and diverse layouts. As we advance to larger scales (i.e., " + }, + { + "bbox": [ + 130, + 546, + 482, + 665 + ], + "type": "inline_equation", + "content": "s_2" + }, + { + "bbox": [ + 130, + 546, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 546, + 482, + 665 + ], + "type": "inline_equation", + "content": "s_3" + }, + { + "bbox": [ + 130, + 546, + 482, + 665 + ], + "type": "text", + "content": "), the mIoU and MA scores consistently increase, demonstrating that our model focuses on learning intricate details in later stages. Meanwhile, F3D and MMD metrics show stability without significant decline, indicating a balanced enhancement in scene complexity and fidelity." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 127, + 306, + 225 + ], + "blocks": [ + { + "bbox": [ + 179, + 116, + 291, + 125 + ], + "lines": [ + { + "bbox": [ + 179, + 116, + 291, + 125 + ], + "spans": [ + { + "bbox": [ + 179, + 116, + 291, + 125 + ], + "type": "text", + "content": "3D and MMD Scores by Mask Ratio" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 135, + 127, + 306, + 225 + ], + "lines": [ + { + "bbox": [ + 135, + 127, + 306, + 225 + ], + "spans": [ + { + "bbox": [ + 135, + 127, + 306, + 225 + ], + "type": "image", + "image_path": "ac7c9dfae1a2fafb3f1efba87450d49c37bad5888a7860936e42a98afda9eda4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 315, + 128, + 477, + 225 + ], + "blocks": [ + { + "bbox": [ + 339, + 116, + 462, + 125 + ], + "lines": [ + { + "bbox": [ + 339, + 116, + 462, + 125 + ], + "spans": [ + { + "bbox": [ + 339, + 116, + 462, + 125 + ], + "type": "text", + "content": "(b) mIoU and OA Scores by Mask Ratio" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 128, + 477, + 225 + ], + "lines": [ + { + "bbox": [ + 315, + 128, + 477, + 225 + ], + "spans": [ + { + "bbox": [ + 315, + 128, + 477, + 225 + ], + "type": "image", + "image_path": "4a15fe1a66cd5979ff88a19b843fc32f6df1fa5f94aa9a0229a80a69e9440908.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 138, + 258, + 477, + 365 + ], + "blocks": [ + { + "bbox": [ + 173, + 236, + 440, + 248 + ], + "lines": [ + { + "bbox": [ + 173, + 236, + 440, + 248 + ], + "spans": [ + { + "bbox": [ + 173, + 236, + 440, + 248 + ], + "type": "text", + "content": "Fig. 6: Effects of mask ratio on unconditional generation results." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 138, + 258, + 477, + 365 + ], + "lines": [ + { + "bbox": [ + 138, + 258, + 477, + 365 + ], + "spans": [ + { + "bbox": [ + 138, + 258, + 477, + 365 + ], + "type": "image", + "image_path": "21ddd5bca6b2f442adbb717ad015fa2738375dc9eb941e6df3647863f1193c8e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 376, + 482, + 420 + ], + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 420 + ], + "type": "text", + "content": "Fig. 7: Infinite Scene Generation. Thanks to the pyramid representation, PDD can be readily applied for unbounded scene generation. This involves the initial efficient synthesis of a large-scale coarse 3D scene, followed by subsequent refinement at higher levels." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 444, + 224, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 444, + 224, + 455 + ], + "spans": [ + { + "bbox": [ + 132, + 444, + 224, + 455 + ], + "type": "text", + "content": "4.5 Applications" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 462, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 569 + ], + "type": "text", + "content": "Cross-dataset. Figure 8 and Figure 9 showcase our model's performance on the transferred dataset from CarlaSC to SemanticKITTI for both unconditional and conditional scene generation. The PDD shows enhanced scene quality after finetuning with SemanticKITTI data, as indicated by improved results in Table 6. This fine-tuning effectively adapts the model to the dataset's complex object distributions and scene dynamics. Notably, despite the higher training effort of the Discrete Diffusion (DD) approach, our method outperforms DD even without fine-tuning, using only coarse scenes from SemanticKITTI. This demonstrates the strong cross-data transfer capability of our approach." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "Infinite Scene Generation. Figure 7 visualizes the process of generating large-scale infinite scenes using our PDD model. As discussed in Section 3.4, we first use the small-scale model to swiftly generate a coarse infinite 3D scene (bottom level in Figure 7). We then leverage larger-scale models to progressively add intricate details (middle and top level in Figure 7), enhancing realism. This approach enables our model to produce high-quality, continuous cityscapes without additional inputs, overcoming the limitations of conventional datasets with finite scenes and supporting downstream tasks like 3D scene segmentation." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "type": "text", + "content": "Pyramid Diffusion for Fine 3D Large Scene Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 116, + 294, + 271 + ], + "blocks": [ + { + "bbox": [ + 138, + 116, + 294, + 271 + ], + "lines": [ + { + "bbox": [ + 138, + 116, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 294, + 271 + ], + "type": "image", + "image_path": "9926aa00e7c5b83b87c912a7f776e89cf588dfadbeb85dc1dac0d09d9597482a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 283, + 301, + 316 + ], + "lines": [ + { + "bbox": [ + 132, + 283, + 301, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 283, + 301, + 316 + ], + "type": "text", + "content": "Fig. 8: SemanticKITTI unconditional generation. " + }, + { + "bbox": [ + 132, + 283, + 301, + 316 + ], + "type": "inline_equation", + "content": "FT" + }, + { + "bbox": [ + 132, + 283, + 301, + 316 + ], + "type": "text", + "content": " stands for finetuning pretrained model from CarlaSC." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 317, + 116, + 476, + 273 + ], + "blocks": [ + { + "bbox": [ + 317, + 116, + 476, + 273 + ], + "lines": [ + { + "bbox": [ + 317, + 116, + 476, + 273 + ], + "spans": [ + { + "bbox": [ + 317, + 116, + 476, + 273 + ], + "type": "image", + "image_path": "22d576bfcc5a8c80fcebf39cfdc082b22581abcade1d53d2fd59527249ee997b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 283, + 482, + 316 + ], + "lines": [ + { + "bbox": [ + 311, + 283, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 311, + 283, + 482, + 316 + ], + "type": "text", + "content": "Fig.9: SemanticKITTI conditional generation. " + }, + { + "bbox": [ + 311, + 283, + 482, + 316 + ], + "type": "inline_equation", + "content": "FT" + }, + { + "bbox": [ + 311, + 283, + 482, + 316 + ], + "type": "text", + "content": " stands for finetuning from CarlaSC models." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 140, + 371, + 479, + 449 + ], + "blocks": [ + { + "bbox": [ + 132, + 327, + 482, + 361 + ], + "lines": [ + { + "bbox": [ + 132, + 327, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 482, + 361 + ], + "type": "text", + "content": "Table 6: Generation results on SemanticKITTI. Finetuned Scales set to None indicates training from scratch and others stand for finetuning corresponding pre-trained CarlaSC model." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 371, + 479, + 449 + ], + "lines": [ + { + "bbox": [ + 140, + 371, + 479, + 449 + ], + "spans": [ + { + "bbox": [ + 140, + 371, + 479, + 449 + ], + "type": "table", + "html": "
MethodFinetuned ScalesConditionedmIoU(V)mIoU(P)F3D(↓)MMD(↓)
DD [2]s4×29.116.00.460.31
PDDNone✓33.422.80.270.32
PDDs2, s4✓43.922.80.280.16
PDDs1×31.323.20.220.13
PDDs1, s2, s4×44.724.20.210.11
", + "image_path": "2410f20815ce02b9c45fd25ac6c2b81e421c65f7569d37b62cdf4f1d12e7277b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 470, + 220, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 470, + 220, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 470, + 220, + 483 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 495, + 482, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 651 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 651 + ], + "type": "text", + "content": "In this work, we introduce the Pyramid Discrete Diffusion model (PDD) to address the significant challenges associated with 3D large scene generation, particularly in the context of limited scale and available datasets. The PDD demonstrates a progressive approach to generating high-quality 3D outdoor scenes, transitioning seamlessly from coarse to fine details. Compared to the other methods, the PDD can generate high-quality scenes within limited resource constraints and does not require additional data sources. Our extensive experimental results show that PDD consistently performs favourably in both unconditional and conditional generation tasks, establishing itself as a reliable and robust solution for the creation of realistic and intricate scenes. Furthermore, the proposed PDD method has great potential in efficiently adapting models trained on synthetic data to real-world datasets and suggests a promising solution to the current challenge of limited real-world data." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 246 + ], + "type": "text", + "content": "Supported by the Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/ Interior Business Center (DOI/IBC) contract number 140D0423C0074. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of IARPA, DOI/IBC, or the U.S. Government. Y. Liu and C. Li are supported in part by the National Science Foundation of China (Grant No: 62202395)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 263, + 197, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 263, + 197, + 276 + ], + "spans": [ + { + "bbox": [ + 133, + 263, + 197, + 276 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 286, + 480, + 664 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 286, + 480, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 286, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 286, + 480, + 308 + ], + "type": "text", + "content": "1. Anvekar, T., Tabib, R.A., Hegde, D., Mudengudi, U.: Vg-vae: A venatus geometry point-cloud variational auto-encoder. In: CVPR (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 308, + 480, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 308, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 138, + 308, + 480, + 329 + ], + "type": "text", + "content": "2. Austin, J., Johnson, D.D., Ho, J., Tarlow, D., Van Den Berg, R.: Structured denoising diffusion models in discrete state-spaces. In: NeurIPS (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 331, + 480, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 480, + 361 + ], + "type": "text", + "content": "3. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: ICCV (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 363, + 480, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 363, + 480, + 383 + ], + "spans": [ + { + "bbox": [ + 138, + 363, + 480, + 383 + ], + "type": "text", + "content": "4. Chen, Z., Wang, G., Liu, Z.: Scenedreamer: Unbounded 3d scene generation from 2d image collections. In: arXiv preprint arXiv: 2302.01330 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 384, + 480, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 384, + 480, + 405 + ], + "spans": [ + { + "bbox": [ + 138, + 384, + 480, + 405 + ], + "type": "text", + "content": "5. Cheng, A.C., Li, X., Liu, S., Sun, M., Yang, M.H.: Learning 3d dense correspondence via canonical point autoencoder. In: ECCV (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 406, + 480, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 406, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 138, + 406, + 480, + 426 + ], + "type": "text", + "content": "6. Cheng, A.C., Li, X., Sun, M., Yang, M.H., Liu, S.: Learning 3d dense correspondence via canonical point autoencoder. In: NeurIPS (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 427, + 480, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 427, + 480, + 459 + ], + "spans": [ + { + "bbox": [ + 138, + 427, + 480, + 459 + ], + "type": "text", + "content": "7. Çiçek, Ö., Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3d unet: learning dense volumetric segmentation from sparse annotation. In: MICCAI (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 460, + 480, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 480, + 480 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 480, + 480 + ], + "type": "text", + "content": "8. Cong, Y., Chen, R., Ma, B., Liu, H., Hou, D., Yang, C.: A comprehensive study of 3-d vision-based robot manipulation. IEEE Transactions on Cybernetics (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 482, + 480, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 482, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 482, + 480, + 502 + ], + "type": "text", + "content": "9. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: NeurIPS (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 503, + 480, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 503, + 480, + 524 + ], + "spans": [ + { + "bbox": [ + 138, + 503, + 480, + 524 + ], + "type": "text", + "content": "0. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: CVPR (2012)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 525, + 480, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 480, + 545 + ], + "type": "text", + "content": "1. Graham, B., Engelcke, M., Van Der Maaten, L.: 3d semantic segmentation with submanifold sparse convolutional networks. In: CVPR (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 546, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 546, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 546, + 480, + 578 + ], + "type": "text", + "content": "2. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 579, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 579, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 579, + 480, + 610 + ], + "type": "text", + "content": "3. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "text", + "content": "4. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 633, + 480, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 664 + ], + "type": "text", + "content": "5. Ho, J., Saharia, C., Chan, W., Fleet, D.J., Norouzi, M., Salimans, T.: Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research (2022)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "type": "text", + "content": "Pyramid Diffusion for Fine 3D Large Scene Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "16. Huang, W., Wang, C., Zhang, R., Li, Y., Wu, J., Fei-Fei, L.: Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 481, + 171 + ], + "type": "text", + "content": "17. Lan, Z., Yew, Z.J., Lee, G.H.: Robust point cloud based reconstruction of large-scale outdoor scenes. In: CVPR (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 481, + 192 + ], + "type": "text", + "content": "18. Lee, J., Im, W., Lee, S., Yoon, S.E.: Diffusion probabilistic models for scene-scale 3d categorical data. arXiv preprint arXiv:2301.00527 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 481, + 224 + ], + "type": "text", + "content": "19. Li, X., Li, C., Tong, Z., Lim, A., Yuan, J., Wu, Y., Tang, J., Huang, R.: Campus3d: A photogrammetry point cloud benchmark for hierarchical understanding of outdoor scene. In: ACM MM (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 481, + 246 + ], + "type": "text", + "content": "20. Li, Y., Ma, L., Zhong, Z., Liu, F., Chapman, M.A., Cao, D., Li, J.: Deep learning for lidar point clouds in autonomous driving: A review. NeurIPS (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "type": "text", + "content": "21. Li, Y., Yu, A.W., Meng, T., Caine, B., Ngiam, J., Peng, D., Shen, J., Lu, Y., Zhou, D., Le, Q.V., et al.: Deepfusion: Lidar-camera deep fusion for multi-modal 3d object detection. In: CVPR (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 278, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 481, + 300 + ], + "type": "text", + "content": "22. Li, Z., Wang, Q., Snavely, N., Kanazawa, A.: Infinitenature-zero: Learning perpetual view generation of natural scenes from single images. In: ECCV (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 300, + 481, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 321 + ], + "type": "text", + "content": "23. Lin, C.H., Lee, H.Y., Menapace, W., Chai, M., Siarohin, A., Yang, M.H., Tulyakov, S.: Infinicity: Infinite-scale city synthesis. arXiv preprint arXiv:2301.09637 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 321, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 481, + 354 + ], + "type": "text", + "content": "24. Liu, M., Shi, R., Chen, L., Zhang, Z., Xu, C., Wei, X., Chen, H., Zeng, C., Gu, J., Su, H.: One-2-3-45++: Fast single image to 3d objects with consistent multi-view generation and 3d diffusion. arXiv preprint arXiv:2311.07885 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 354, + 481, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 481, + 375 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 481, + 375 + ], + "type": "text", + "content": "25. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 375, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 481, + 396 + ], + "type": "text", + "content": "26. Luo, S., Hu, W.: Diffusion probabilistic models for 3d point cloud generation. In: CVPR (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 396, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 481, + 418 + ], + "type": "text", + "content": "27. Ma, Q., Yang, J., Tang, S., Black, M.J.: The power of points for modeling humans in clothing. In: ICCV (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 418, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 439 + ], + "type": "text", + "content": "28. Mascaro, R., Teixeira, L., Chli, M.: Diffuser: Multi-view 2d-to-3d label diffusion for semantic scene segmentation. In: ICRA (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 439, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 481, + 460 + ], + "type": "text", + "content": "29. Moon, T., Choi, M., Lee, G., Ha, J.W., Lee, J.: Fine-tuning diffusion models with limited data. In: NeurIPS 2022 Workshop on Score-Based Methods (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 460, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 481, + 482 + ], + "type": "text", + "content": "30. Moro, S., Komuro, T.: Generation of virtual reality environment based on 3d scanned indoor physical space. In: ISVC (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 482, + 481, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 481, + 503 + ], + "type": "text", + "content": "31. Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: ICML (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 503, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 503, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 503, + 481, + 536 + ], + "type": "text", + "content": "32. Ögün, M.N., Kurul, R., Yasar, M.F., Turkoglu, S.A., Avci, Ş., Yildiz, N.: Effect of leap motion-based 3d immersive virtual reality usage on upper extremity function in ischemic stroke patients. Arquivos de neuro-psiquiatria (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 536, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 557 + ], + "type": "text", + "content": "33. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 557, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 481, + 578 + ], + "type": "text", + "content": "34. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: NeurIPS (2017)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "type": "text", + "content": "35. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "text", + "content": "36. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "37. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS (2022)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 479 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "38. Sohl-Dickstein, J., Weiss, E.A., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. arXiv preprint arXiv:1503.03585 (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "text", + "content": "39. Sra, M., Garrido-Jurado, S., Schmandt, C., Maes, P.: Procedurally generated virtual reality from 3d reconstructed physical space. In: Proceedings of the 22nd ACM Conference on Virtual Reality Software and Technology (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "text", + "content": "40. Su, S.Y., Bagautdinov, T., Rhodin, H.: Npc: Neural point characters from video. In: ICCV (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "41. Tang, J., Nie, Y., Markhasin, L., Dai, A., Thies, J., Nießner, M.: Diffuscene: Scene graph denoising diffusion probabilistic model for generative indoor scene synthesis. arXiv preprint arXiv:2303.14207 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 259 + ], + "type": "text", + "content": "42. Tang, Y., He, H., Wang, Y., Mao, Z., Wang, H.: Multi-modality 3d object detection in autonomous driving: A review. Neurocomputing (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 260, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 481, + 293 + ], + "type": "text", + "content": "43. Wilson, J., Song, J., Fu, Y., Zhang, A., Capodieci, A., Jayakumar, P., Barton, K., Ghaffari, M.: Motionsc: Data set and network for real-time semantic mapping in dynamic environments. IEEE Robotics and Automation Letters 7(3) (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "text", + "content": "44. Wu, P., Jia, X., Chen, L., Yan, J., Li, H., Qiao, Y.: Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. In: NeurIPS (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "type": "text", + "content": "45. Xie, H., Chen, Z., Hong, F., Liu, Z.: Citydreamer: Compositional generative model of unbounded 3d cities. arXiv preprint arXiv:2309.00610 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 348, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 481, + 369 + ], + "type": "text", + "content": "46. Xu, Z., He, Z., Wu, J., Song, S.: Learning 3d dynamic scene representations for robot manipulation. arXiv preprint arXiv:2011.01968 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 370, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 481, + 392 + ], + "type": "text", + "content": "47. Zeng, X., Vahdat, A., Williams, F., Gojcic, Z., Litany, O., Fidler, S., Kreis, K.: Lion: Latent point diffusion models for 3d shape generation. In: NeurIPS (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "text", + "content": "48. Zheng, H., Nie, W., Vahdat, A., Anandkumar, A.: Fast training of diffusion models with masked transformers. arXiv preprint arXiv:2306.09305 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 414, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 435 + ], + "type": "text", + "content": "49. Zheng, Y., Yifan, W., Wetzstein, G., Black, M.J., Hilliges, O.: Pointavatar: Deformable point-based head avatars from videos. In: CVPR (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 436, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 436, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 481, + 456 + ], + "type": "text", + "content": "50. Zhou, L., Du, Y., Wu, J.: 3d shape generation and completion through point-voxel diffusion. In: ICCV (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "type": "text", + "content": "51. Zhuang, F., Qi, Z., Duan, K., Xi, D., Zhu, Y., Zhu, H., Xiong, H., He, Q.: A comprehensive survey on transfer learning. arXiv preprint arXiv:1911.02685 (2020)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "type": "text", + "content": "Pyramid Diffusion for Fine 3D Large Scene Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Q&A Prompts_ Discovering Rich Visual Clues through Mining Question-Answer Prompts for VQA requiring Diverse World Knowledge/d6b849aa-8c9f-421c-b457-5802b3bd4aca_content_list.json b/2024/Q&A Prompts_ Discovering Rich Visual Clues through Mining Question-Answer Prompts for VQA requiring Diverse World Knowledge/d6b849aa-8c9f-421c-b457-5802b3bd4aca_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3f28553257cb0af4a7bc452c171abc7da8a2ae68 --- /dev/null +++ b/2024/Q&A Prompts_ Discovering Rich Visual Clues through Mining Question-Answer Prompts for VQA requiring Diverse World Knowledge/d6b849aa-8c9f-421c-b457-5802b3bd4aca_content_list.json @@ -0,0 +1,2125 @@ +[ + { + "type": "text", + "text": "Q&A Prompts: Discovering Rich Visual Clues through Mining Question-Answer Prompts for VQA requiring Diverse World Knowledge", + "text_level": 1, + "bbox": [ + 238, + 140, + 767, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haibo Wang $^{1,2}$ and Weifeng Ge $^{1,2\\boxtimes}$", + "bbox": [ + 359, + 233, + 643, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Nebula AI Group, School of Computer Science, Fudan University \n2 Shanghai Key Lab of Intelligent Information Processing hbwang22@m.fudan.edu.cn, wfge@fudan.edu.cn", + "bbox": [ + 274, + 261, + 725, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. With the breakthrough of multi-modal large language models (MLLMs), answering complex visual questions that demand advanced reasoning abilities and world knowledge has become a much more important testbed for developing AI models than ever. However, equipping MLLMs with robust cross-modality reasoning ability remains challenging since the cognition scheme of humans has not been understood systematically. In this paper, we believe that if we can collect rich visual clues, we will recognize the image more accurately, understand the question better, recall relevant knowledge more easily, and finally reason out the answer. We discover these rich visual clues by mining question-answer pairs in images and sending them into multi-modal large language models as prompts. We call the proposed method Q&A Prompts. Specifically, we first use the image-answer pairs and the corresponding questions in the training set as inputs and outputs to train a visual question generation (VQG) model. Then, we use an image tagging model to identify various instances and send packaged image-tag pairs into the VQG model to generate relevant questions with the extracted image tags as answers. Finally, we encode these generated question-answer pairs as prompts with a visual-aware prompting module and send them into pretrained MLLMs to reason out the final answers. Experimental results show that, compared with state-of-the-art methods, our Q&A Prompts achieves substantial improvements on the challenging visual question answering datasets requiring reasoning over diverse world knowledge, such as OK-VQA and A-OKVQA. Codes will be available at link.", + "bbox": [ + 261, + 344, + 743, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: multi-modal large language model $\\cdot$ visual question answering $\\cdot$ visual language reasoning", + "bbox": [ + 259, + 690, + 740, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 746, + 375, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual Question Answering (VQA) has longstanding been considered a Visual Turing Test [12] in the machine learning community. To solve this Turing test, artificial intelligence (AI) models are expected to be equipped with the human capabilities of visual recognition, language understanding, logical reasoning, world", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/378d7f50367610104950f29c2507ff8f41ccd40980f1e804df8a8295710bbea5.jpg", + "image_caption": [ + "VQA Tasks well-solved by MLLMs" + ], + "image_footnote": [], + "bbox": [ + 241, + 157, + 375, + 223 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/092e870ac39982361dc1d008fa0b49a7d4c9340f959388bfda9975accc98aa90.jpg", + "image_caption": [ + "VQA Tasks unsolved by MLLMs" + ], + "image_footnote": [], + "bbox": [ + 388, + 157, + 527, + 232 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/66ae73ade15069150a3b6540ba59cfc07d167dedc0194298fb820cba1a5a7e58.jpg", + "image_caption": [ + "Fig. 1: Illustration of VQA tasks that require simple perception abilities ((a) and (b)) and those demanding chains of reasoning over diverse world knowledge ((c) and (d)). The example on the right ((e) and (f)) indicates MLLMs with Q&A prompts can solve difficult VQA problems that require reasoning over diverse world knowledge." + ], + "image_footnote": [], + "bbox": [ + 240, + 234, + 370, + 309 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e5d657845a74283f272017c9a1f73e768e1da56eca27fb2f108f6faa50ad7b33.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 234, + 527, + 309 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/66bb53699443cc5b275aa8fc2e0bf0c4bacbba12d7c557a3b3269735f3e4d24f.jpg", + "image_caption": [ + "VQA Tasks solved by MLLMs with Q&A Prompts" + ], + "image_footnote": [], + "bbox": [ + 531, + 157, + 772, + 309 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "knowledge referring, etc. There are a variety of datasets focusing on different problems in VQA, such as the perception and language understanding problem [3, 13], reasoning on procedurally generated images [22], and structured factual knowledge [50, 51] or commonsense knowledge [55]. Recently, with the advent of multi-modal large language models (MLLMs) [6, 26, 30, 54], significant progress has been achieved in perception-based VQA tasks [3, 13]. Fig. 1 (a) and Fig. 1 (b) show such VQA tasks, which can be reliably solved by MLLMs.", + "bbox": [ + 212, + 411, + 784, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, for more complex VQA tasks [37, 42] as in Fig. 1 (c) and Fig. 1 (d), that feature the necessity of diverse world knowledge and complex chains of reasoning, state-of-the-art methods [14, 18, 29] even MLLM models [6, 26] fail to give the correct answers. Different from perception-based VQA [3, 13] and \"closed\" knowledge-based VQA [50, 51], visual question reasoning in OK-VQA [37] and A-OKVQA [42] requires not only diverse forms of world knowledge but also need complex reasoning logic. In Fig. 1 (d), where a woman sits on the ground with her pet dogs and holds an umbrella, the question is [Why is the woman holding the umbrella?] Simply associating the keyword [umbrella] with the word [rain] due to their frequent co-occurrence in the pre-training corpus [7] will lead to the wrong answer. To reason out the correct answer, AI models need to consider rich visual clues presented in the image. For example, in Fig. 1 (f), having discovered the clues including [the woman is camping] and [the weather is shiny] from the Q&A pairs, the model can correctly fill the logic gap and arrive at the answer [for shade]. This evidence indicates the necessity of collecting rich visual clues to achieve good results.", + "bbox": [ + 212, + 518, + 787, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the analysis above, we believe that collecting visual clues of various instances in images will help MLLM recognize the image more accurately, understand the question better, recall relevant knowledge more easily, and finally generate the reasoning results more intelligently. To investigate this hypothesis, we design a novel VQA framework called $Q\\mathcal{E}A$ Prompts, which extracts nu", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Haibo Wang, Weifeng Ge", + "bbox": [ + 271, + 114, + 441, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "merous question-answer pairs associated with different perspectives in images as prompts for MLLMs. We choose an instance in the given image as the answer and generate the related question to form a question-answer pair. We generate the question-answer pairs for almost every instance to get a bag of Q&A prompts. The reason we choose question-answer pairs as prompts comes from two folds: First, Q&A prompts can focus on diverse instances in images like objects, scenes, or actions, thereby offering multiple different perspectives to understand the given image and the target question; Second, the knowledge and reasoning insights hidden in these question-answer pairs can help to explicitly uncover more relevant world knowledge, which may be critical to filling the logic gap between perception and reasoning when answering the target question.", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Q&A Prompts, there are three key stages, which are the visual question generation (VQG) model training stage, the question-answer prompts generation stage, and the visual-language reasoning stage. In the VQG model training stage, we gather question-answer pairs in a VQA training set to train a VQG model that can map an answer together with the given image to a correlated question. Subsequently, in the question-answer prompts generation stage, we leverage a pre-trained image tagging model [19, 57] to recognize various objects, scenes, and actions in images with text tags. We employ these tags as answers and send them into the VQG model along with the corresponding images to generate questions. Finally, in the visual-language reasoning stage, we encode the generated question-answer pairs with a visual-aware prompting module and send them together with image and question features into a frozen language model to reason out the answer. Since previous VQA benchmarks [3,13] have seen saturated performance, we conduct experiments on the more challenging OK-VQA [37] and A-OKVQA [42] datasets, which are newer and larger testbeds for MLLM to investigate their abilities in accomplishing reasoning tasks requiring diverse world knowledge. Experimental results show that our Q&A Prompts can substantially improve the reasoning ability of MLLMs, such as InstructBLIP [6], LLaVA [30], and MiniGPT-4 [59]. In summary, we make the following contributions:", + "bbox": [ + 212, + 313, + 787, + 602 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel VQA framework that effectively generates and leverages question-answer pairs as prompts to improve the reasoning ability of MLLMs. It provides a new perspective for the community to design VQA solutions by explicitly collecting rich visual clues to bridge the logic gap between perception and reasoning when answering the target question.", + "- We design a novel question-answer prompts generation scheme with a VQG model and an image tagging model, which can generate Q&A prompts for recognizable objects, scenes, and activities in images. Besides, we design a new visual-aware prompting module to encode these prompts efficiently for the subsequent reasoning.", + "- We test the method on the challenging OK-VQA [37] and A-OKVQA [42] benchmarks, which study reasoning over diverse forms of world knowledge. The proposed Q&A Prompts achieves an accuracy of $68.1\\%$ and $64.3\\%$ on A-OKVQA [42] and OK-VQA [37], outperforming previous state-of-the-art methods by clear margins." + ], + "bbox": [ + 225, + 612, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Q&A Prompts", + "bbox": [ + 632, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 387, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multimodal Large Language Models (MLLMs). MLLMs [2,6,25,26,34,54, 56,59] have shown strong ability in image-language understanding and reasoning, by adapting frozen language models to frozen image encoders with trainable connection modules. For instance, Flamingo [2] incorporates visual features into the language model with gated cross-attention blocks. LLaVA [30], instead, directly projects visual features into the space of text embeddings using a linear layer. BLIP-2 and InstructBLIP [6,26] introduce a more complex Q-Former to bridge the modality gap. These models align images and language based on large-scale image-text pre-training [5,10,27] and conduct reasoning with language models. In this paper, we mine more numerous question-answer pairs and send them into the MLLM as prompts. These prompts provide rich visual clues and uncover relevant world knowledge, which is beneficial to reasoning out correct answers.", + "bbox": [ + 212, + 172, + 785, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "VQA Requiring Reasoning over Diverse World Knowledge. Different from classical VQA tasks [3, 13, 50, 51], VQA requiring reasoning over world knowledge involves inferring intents, goals, the physics of the world, and the social dynamics of individuals. More recently, A-OKVQA [42] and OK-VQA [37], which our work focuses on, present large-scale challenging VQA datasets that encompass questions requiring reasoning based on diverse world knowledge, including visual knowledge, commonsense knowledge, and factoid knowledge. Previous methods like Img2LLM [15], Prophet [44], and PromptCap [17] exploit the rich hidden knowledge and strong reasoning abilities of LLMs (e.g., GPT3 [4]) to solve such problems. These methods convert images into natural languages and construct various prompt templates for LLMs with in-context learning [4] to generate answers. Another group of works, such as KAT [14], REVIVE [29], and REVEAL [18], directly encode questions and relevant knowledge retrieved from external knowledge bases [4, 46, 49] to train end-to-end models. Unlike these methods, we design a visual-question prompts generation scheme to mine the information in images and unambiguous words to reduce confusion when answering the target questions. With Q&A prompts, different types of world knowledge are incorporated into the reasoning process to help MLLMs infer the correct answers.", + "bbox": [ + 212, + 356, + 785, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Visual Question Generation (VQG). Various works have explored Visual Question Generation (VQG) in the context of VQA tasks. SQuINT [43] and SelTDA [24] focus on dataset augmentation to enhance the generalization of VQA models. SQuINT employed manual workers to annotate questions, while SelTDA generated questions automatically using unlabeled images. Another line of research, exemplified by [48], employs a VQG model to generate informative sub-questions. These sub-questions are then used alongside the original question during inference, enabling another VQA model to answer both simultaneously. Different from previous works, our VQG model uses the tags of images [19, 57] as answers for generated questions. This approach ensures that our VQG model generates more relevant questions with additional details and perspectives. Furthermore, we introduce a novel module to encode the question-answer pairs as prompts for MLLMs, instead of directly feeding them as the formation of natural language into the model [48].", + "bbox": [ + 212, + 628, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Haibo Wang, Weifeng Ge", + "bbox": [ + 271, + 114, + 441, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a62552742be334efc773e4f01521fea1e129b587aadfed2ff84415e3404f26e9.jpg", + "image_caption": [ + "Fig. 2: (a) Firstly, we train a VQG model that can ask informative questions given the image and specified answer. (b) Then, we exploit the image tagging model to extract image tags as the answers to generated questions and obtain diverse question-answer prompts. (c) At last, we feed the image, question, and question-answer prompts into the VQA model to perform visual-language reasoning in (c)." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 787, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 443, + 328, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Q&A Prompts is a conceptually simple three-stage framework, and Fig. 2 gives an overview: a VQG model training stage, a question-answer prompts generation stage, and a visual-language reasoning stage. Specifically, in the VQG model training stage (Fig. 2 (a)), we train a vanilla MLLM with a broad range of image-answer pairs as input and the corresponding question as the target output to learn a mapping from answers to questions (detailed in Sec. 3.1). In the prompts generation stage (Fig. 2 (b)), we exploit an image tagging model [19, 57] to obtain relevant tags including attributes, objects, actions, and scenes in images. These tags serve as answers and will be fed into the VQG model trained in the first stage to generate diverse questions containing rich visual clues and world knowledge (detailed in Sec. 3.2). In the final reasoning stage (Fig. 2 (c)), we utilize our proposed visual-aware prompting module to encode these question-answer pairs into prompts, which will be fed into the MLLM such as BLIP-2 [26] or InstructBLIP [6] together with the image and the target question, to predict the answer (detailed in Sec. 3.3). The whole pipeline is flexible and effective in mining informative visual clues to correctly answer the given questions.", + "bbox": [ + 212, + 474, + 787, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Stage-1: Train the VQG Model", + "text_level": 1, + "bbox": [ + 215, + 739, + 519, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given an image $\\mathcal{V}$ and answer $\\mathcal{A}$ , the VQG model is responsible for posing an informative question $\\mathcal{Q}$ which can be properly answered by $\\mathcal{A}$ . To explore as many visual clues in the generated questions as possible, and avoid generating general questions containing limited clues (e.g., \"What is this?\"), we utilize the A-OKVQA [42] or OK-VQA [37] as the training set $D$ for our VQG model, since", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Q&A Prompts", + "bbox": [ + 632, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/029793f67d72050a57a9e7b74347ae346d883a9305c90d77a4a6e6e3ab2959ea.jpg", + "image_caption": [], + "image_footnote": [ + "A-OKVQA What sport do the people have equipment for? \nVQG Which object in this image used for winter sports? \nZero-shot Is this a snowboard?" + ], + "bbox": [ + 220, + 143, + 395, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8492208510e5bde4c93c6360c46596cc3926ad6e6b6e3869093c7d23d2f25f6b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 143, + 589, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f470664c4fe5cd2a5670c7cb9e920a8f4d6ee942ef5a24b4a19784c02d43d2a0.jpg", + "image_caption": [], + "image_footnote": [ + "A-OKVQA What is the vehicle traveling on? \nVQG What is this type of transportation called? \nZero-shot What is this?" + ], + "bbox": [ + 607, + 143, + 782, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d877037b565550b9fa92e2376ea343376f2583f821eebf7faf88b68adb28a9ce.jpg", + "image_caption": [ + " snowboard" + ], + "image_footnote": [], + "bbox": [ + 220, + 226, + 321, + 279 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/97de4a57b3897445ccd8a0a038f0ddb8f0fec304f108cfbecb4f20cc8aaa7b60.jpg", + "image_caption": [ + " rail", + "Fig. 3: Visualizations and T-SNE embedding distributions of the questions in manually annotated A-OKVQA (green), and the questions generated by our VQG model (blue) or by directly prompting a frozen MLLM (orange)." + ], + "image_footnote": [], + "bbox": [ + 516, + 227, + 616, + 279 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the questions in their annotations are diverse and specific, covering broader areas of knowledge. To create the VQG model that approximates $P(\\mathcal{Q}|\\mathcal{V},\\mathcal{A})$ , we treat the problem of learning such a model as a text-generation problem and wish to train the MLLM (e.g., InstructBLIP [6]) as our VQG model with $D$ due to its strong image-language reasoning ability.", + "bbox": [ + 212, + 375, + 784, + 449 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, as shown in Fig. 2 (a), the VQG model comprises a frozen image encoder, a trainable connection module, and a frozen LLM. We extract image embeddings $\\mathbf{E}_v\\in \\mathbb{R}^{n\\times d_v}$ of $\\nu$ with the image encoder ( $n$ is the patch number, $d_v$ is the embedding dimension), and then feed $\\mathbf{E}_v$ into the connection module to obtain fixed-length visual features $\\mathbf{F}_v\\in \\mathbb{R}^{k\\times d_q}$ , which will be sent into the LLM together with the instruction. We write the instruction as the template containing the answer $\\mathcal{A}$ , such as \"Given the image, generate a question whose answer is: $\\langle \\mathcal{A}\\rangle$ .\" The model is trained using the cross-entropy loss with parameters $\\theta$ :", + "bbox": [ + 212, + 449, + 787, + 571 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {V Q G}} = - \\sum_ {t = 1} ^ {L _ {q}} \\log P _ {\\theta} \\left(\\mathcal {Q} _ {t} \\mid \\mathcal {Q} _ {< t}, \\mathcal {V}, \\mathcal {A}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 579, + 785, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{Q}_t$ is predicted autoregressively at position $t$ , and $L_{q}$ is the sequence length of the question $\\mathcal{Q}$ . We emphasize the necessity of training the VQG model, instead of prompting a frozen MLLM in a zero-shot manner. In Fig. 3, we use sentence-bert [41] to obtain the dense embedding vectors of each question and show the T-SNE [35] embedding of them, as well as visualization examples. We can observe that the questions generated by our VQG model (blue) are more diverse and specific like questions in A-OKVQA (green), while the questions generated by MLLM without training (orange) are often general and limited with low differentiation. Our following experiments will further verify this.", + "bbox": [ + 212, + 628, + 787, + 767 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Stage-2: Generate Question-Answer Prompts", + "text_level": 1, + "bbox": [ + 214, + 786, + 633, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Once the VQG model has been obtained, the generation of question-answer pairs can proceed. In this stage, we will explicitly generate a wide range of question-", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Haibo Wang, Weifeng Ge", + "bbox": [ + 271, + 114, + 442, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3d4224a2be9d5450b8c6a0cc34ed314f6e817eb757ce5b15aa4ddd293b8aff71.jpg", + "image_caption": [ + "Fig. 4: Overview of the visual-aware prompting module, consisting of the prompt encoder, visual gated fusion, and resampler decoder." + ], + "image_footnote": [], + "bbox": [ + 223, + 143, + 781, + 236 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "answer pairs to help the model have a deeper and broader understanding of the image. In order to generate diverse questions covering various perspectives of the image, we should ensure the answers fed into the VQG model are sufficient in varied forms. Therefore, we adopt a strong image tagging model [19, 57] to produce rich tags as answers. Different from object detectors that can only recognize limited object labels, image tagging models aim to identify various open-vocabulary elements in images, including attributes, objects, actions, scenes, etc, which can provide an appropriate source of the answers we need.", + "bbox": [ + 212, + 304, + 782, + 425 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We utilize the RAM [57] model as our image tagging model due to its superior recognition ability. As shown in Fig. 2 (b), we feed the image $\\mathcal{V}$ into the tagging model and obtain various tags for each image. To make the tags of each image more distinguishable, we remove the most frequent ones containing general information, such as 'person', 'food', 'man', etc., resulting in an average of 14 tags per image. We denote this set of tags as $\\mathcal{O} = \\{o_i\\}_{i=1}^M$ , where $M$ is the number of tags associated with an image.", + "bbox": [ + 212, + 425, + 784, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Using these tags as answers, we input them into the trained VQG model along with the image, as shown in Fig. 2 (b), to obtain corresponding questions that answer these tags. To ensure that the generated questions (denoted as $\\mathcal{T} = \\{t_i\\}_{i=1}^M$ ) contribute effectively to answering the target question, we rank the generated questions $\\mathcal{T} = \\{t_i\\}_{i=1}^M$ based on their similarity with the target question $\\mathcal{Q}$ , and adopt the questions with the Top-P similarities as the final obtained question prompts $\\mathcal{U}$ :", + "bbox": [ + 214, + 530, + 784, + 636 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {U} = \\left\\{u _ {p} \\right\\} _ {p = 1} ^ {P} = \\underset {t _ {i} \\in \\mathcal {T}} {\\arg \\operatorname {T o p - P}} \\langle \\mathrm {T} (t _ {i}), \\mathrm {T} (\\mathcal {Q}) \\rangle \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 356, + 647, + 784, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathrm{T}(\\cdot)$ represents the embedding model [41], which encodes sentences into embedding vectors for similarity computation. The $\\langle \\cdot, \\cdot \\rangle$ operator is the inner product, computing the cosine similarity between two sentences. We then concatenate each $u_i$ in $\\mathcal{U}$ and its corresponding $o_i$ in $\\mathcal{O}$ using the template \"Question: $\\langle u_i \\rangle$ Answer: $\\langle o_i \\rangle$ \", resulting in the final question-answer pairs as prompts $\\mathcal{S}$ .", + "bbox": [ + 214, + 685, + 787, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Stage-3: Reasoning with Q&A Prompts", + "text_level": 1, + "bbox": [ + 214, + 784, + 586, + 799 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After obtaining these question-answer pairs involving rich visual clues and world knowledge, we use the proposed visual-aware prompting module, as shown in", + "bbox": [ + 214, + 809, + 784, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Q&A Prompts", + "bbox": [ + 630, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 4, to encode these question-answer pairs into a set of prompt embeddings $\\mathbf{F}_p\\in \\mathbb{R}^{k\\times d_q}$ for our VQA model. We still utilize the same MLLM (e.g., Instruct-BLIP [6]) as our VQA model for the final reasoning. As in Fig. 2 (c), the encoded prompt embeddings $\\mathbf{F}_p$ will be fed into the frozen LLM in MLLM, together with the instruction embeddings and visual features $\\mathbf{F}_v$ to predict the answer $\\mathcal{A}$ :", + "bbox": [ + 214, + 146, + 787, + 223 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} = \\operatorname {L L M} \\left(\\left[ \\mathbf {F} _ {v}; \\mathbf {F} _ {p}; \\operatorname {E m b e d} (\\operatorname {I n s}) \\right]\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 232, + 785, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Embed is the word embedding layer in LLM, the operator $[\\cdot]$ means concatenation, and Ins is the instruction including the target question $\\mathcal{Q}$ , represented as \"Question: $\\langle \\mathcal{Q} \\rangle$ ? Answer:\". We then introduce how the visual-aware prompting module bootstrapped the reasoning ability of MLLM within Q&A Prompts.", + "bbox": [ + 214, + 258, + 782, + 318 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prompt Encoder. The prompt encoder is responsible for obtaining the semantics representations of the prompts $S$ . Specifically, we choose encoder-based language models as our prompt encoder due to their strong ability in language understanding. To preserve the well-trained representations, we keep the parameters of the prompt encoder frozen. We feed the prompts $S$ into the prompt encoder and get the encoded $\\mathbf{F}_s \\in \\mathbb{R}^{L_s \\times d_q}$ , where $L_s$ is the sequence length of question-answer prompts and $d_q$ is the hidden size of the prompt encoder.", + "bbox": [ + 212, + 319, + 787, + 424 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visual Gated Fusion. Since the output $\\mathbf{F}_s$ from the prompt encoder is an unimodal language representation, to fully utilize the explicit information in images, the visual gated fusion adaptively incorporates image information into this representation with the visual embeddings $\\mathbf{E}_v \\in \\mathbb{R}^{n \\times d_v}$ . Specifically, in Fig. 4, we use a single-head attention mechanism, similar to MM-COT [58], to correlate the language representation with the visual embeddings. We define the query, key and value as $\\mathbf{Q} = \\mathbf{F}_s \\in \\mathbb{R}^{L_s \\times d_q}$ , $\\mathbf{K} = \\mathbf{V} = \\mathrm{MLP}(\\mathbf{E}_v) \\in \\mathbb{R}^{n \\times d_q}$ , where MLP here is projecting $d_v$ to $d_q$ . The attention output $\\mathbf{F}_v^{attn}$ is defined as:", + "bbox": [ + 214, + 424, + 787, + 546 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {v} ^ {\\text {a t t n}} = \\mathbf {Q} + \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {T}}{\\sqrt {d _ {q}}}\\right) \\mathbf {V}, \\quad \\mathbf {F} _ {v} ^ {\\text {a t t n}} \\in \\mathbb {R} ^ {L _ {s} \\times d _ {q}} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 555, + 785, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Next, we apply the gated fusion mechanism to merge $\\mathbf{F}_s$ and $\\mathbf{F}_v^{attn}$ . The fused representation $\\mathbf{F}_m \\in \\mathbb{R}^{L_s \\times d_q}$ is obtained using the following equations:", + "bbox": [ + 214, + 602, + 785, + 633 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\lambda = \\operatorname {S i g m o i d} \\left(\\mathbf {F} _ {s} \\mathbf {W} _ {s} + \\mathbf {F} _ {v} ^ {\\text {a t t n}} \\mathbf {W} _ {v}\\right) \\\\ \\mathbf {F} _ {m} = (1 - \\lambda) \\mathbf {F} _ {s} + \\lambda \\mathbf {F} _ {v} ^ {\\text {a t t n}} \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 642, + 785, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in which $\\mathbf{W}_s$ and $\\mathbf{W}_v$ are learnable parameters. The gated fusion allows the model to adaptively blend the information from the language representation and visual embeddings, resulting in the multimodal representations $\\mathbf{F}_m$ .", + "bbox": [ + 212, + 688, + 782, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Resampler Decoder. Although $\\mathbf{F}_m \\in \\mathbb{R}^{L_s \\times d_q}$ in Eq. (5) is the multimodal representation of $\\mathcal{S}$ , we cannot directly concatenate it with $\\mathbf{F}_v$ and instruction embeddings as in Eq. (3), due to the long sequence length of $L_s$ , which can result in extra computational burden and distraction to the target question $\\mathcal{Q}$ . In light of this, we employ a lightweight resampler decoder inspired by Perceiver [20], to map $\\mathbf{F}_m \\in \\mathbb{R}^{L_s \\times d_q}$ into smaller-sized prompt embeddings $\\mathbf{F}_p \\in \\mathbb{R}^{k \\times d_q}$ with fixed length $k$ , where $k \\ll L_s$ (e.g., $k = 32$ and $L_s \\geq 100$ ).", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Haibo Wang, Weifeng Ge", + "bbox": [ + 271, + 114, + 442, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Concretely, we additionally introduce the learnable embeddings $\\mathbf{F}_p\\in \\mathbb{R}^{k\\times d_q}$ to interact with $\\mathbf{F}_m$ . As in Fig. 4, a cross-attention layer is applied by taking $\\mathbf{F}_p$ as query, and $\\mathbf{F}_m$ as key and value, followed with a feed-forward network, to generate prompt embeddings with the multimodal representations from $\\mathbf{F}_m$ :", + "bbox": [ + 214, + 145, + 787, + 207 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {p} = \\text {C r o s s - A t t e n t i o n} (\\mathbf {F} _ {p}, \\mathbf {F} _ {m}) + \\mathbf {F} _ {p}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 215, + 635, + 231 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {p} = \\operatorname {F F N} \\left(\\mathbf {F} _ {p}\\right) + \\mathbf {F} _ {p}, \\mathbf {F} _ {p} = \\operatorname {M L P} \\left(\\mathbf {F} _ {p}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 227, + 784, + 251 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where MLP projects $d_{q}$ to $d_{lm}$ (the hidden size of LLM).", + "bbox": [ + 214, + 258, + 627, + 273 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "At last, the final output of $\\mathbf{F}_p\\in \\mathbb{R}^{k\\times d_q}$ will be fed into the LLM together with the question instruction embeddings and visual features as in Eq. (3), to make predictions of answers. The model is trained using the cross-entropy loss function with trainable parameters $\\theta$ :", + "bbox": [ + 212, + 273, + 785, + 335 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {V Q A}} = - \\sum_ {t = 1} ^ {L _ {a}} \\log P _ {\\theta} \\left(\\mathcal {A} _ {t} \\mid \\mathcal {A} _ {< t}, \\mathcal {V}, \\mathcal {Q}, \\mathcal {S}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 343, + 785, + 383 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathcal{A}_t$ is predicted autoregressively at position $t$ , and $L_{a}$ is the sequence length of the ground truth answer text $\\mathcal{A}$ .", + "bbox": [ + 214, + 391, + 785, + 422 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 441, + 375, + 459 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Dataset", + "text_level": 1, + "bbox": [ + 214, + 472, + 328, + 486 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We evaluate Q&A Prompts on the challenging A-OKVQA [42] and OK-VQA [37] since they are currently the largest VQA datasets requiring complex reasoning over diverse world knowledge. Specifically, the A-OKVQA dataset consists of 24,903 samples, with 17.1k samples for training, 1.1k for validation, and 6.7k for testing. The questions in this dataset require reasoning over various types of world knowledge, such as commonsense knowledge, factoid knowledge, and visual knowledge. The OK-VQA dataset includes 14k questions covering a variety of knowledge categories, with 9k samples for training, and 5k for validation. Each sample in these two datasets includes an image, a question, and 10 ground-truth answers. We use the soft accuracy [3] as the standard evaluation metric.", + "bbox": [ + 212, + 494, + 787, + 647 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 666, + 455, + 681 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We use InstructBLIP [6] as our default MLLM, to both generate Q&A pairs and perform answer predictions. RAM [57] is utilized as the image tagging model due to its strong recognition ability. The number of Q&A pairs is set to $P = 8$ . To configure the visual-aware prompting module, we use the text encoder in CLIP [40] as our prompt encoder, and a value of 32 of the $k$ for the number of prompt embeddings $\\mathbf{F}_p \\in \\mathbb{R}^{k \\times d_q}$ , the same as the number of visual tokens $\\mathbf{F}_v \\in \\mathbb{R}^{k \\times d_q}$ . We train the models using AdamW [31] as the optimizer with a learning rate of $2e^{-5}$ and the strategy of mixed precision training. It is important to note that the parameters of the image encoder, the large language model, and the prompt encoder are kept frozen to maintain their pre-trained representations.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Q&A Prompts", + "bbox": [ + 630, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/1b5f54cb964ad6f8f992a798cd0426f8218478c76cda6a5845e481f6566e472e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodImage RepresentationKnowledge SourceOK-VQA (%)A-OKVQA (%)
ValTest
ClipCap [39]FeaturePretrain-18.115.8
Pythia [21]FeaturePretrain-25.221.9
ViLBERT [32]FeaturePretrain-30.625.9
LXMERT [47]FeaturePretrain-30.725.9
GPV-2 [23]FeaturePretrain-48.640.7
Unified-IO [33]FeaturePretrain54.0-45.2
Flamingo (80B) [2]FeaturePretrain57.8--
BLIP-2 [26]FeaturePretrain59.360.058.7
Muko [60]FeatureDBPedia + ConceptNet29.2--
ConceptBERT [11]FeatureNumberBatch + ConceptNet33.7--
KRISP [36]FeatureWikipedia + ConceptNet38.933.722.1
MAVEx [52]FeatureWikipedia + ConceptNet40.3--
UniER [16]FeatureConceptNet42.1--
TRiG [9]Caption + Tags + OCRWikipedia49.4--
REVEAL [18]FeatureWIT + Wikidata59.1-52.2
RA-VQA-v2 [28]FeatureGoogle Search62.1--
PICa [53]Caption + TagsFrozen GPT-3 (175B)48.0--
KAT [14]Caption + FeatureWikidata + GPT3 (175B)53.1--
Img2LLM [15]CaptionFrozen GPT-3 (175B)-42.940.7
REVIVE [29]Caption + FeatureWikipedia + GPT3 (175B)56.6--
PromptCap [17]CaptionFrozen GPT-3 (175B)60.456.359.6
Prophet [44]Caption + TagsFrozen GPT-3 (175B)61.158.261.1
InstructBLIP (7B) [6] + Q&A PromptsFeaturePretrain62.164.062.1
FeaturePretrain64.3 (+2.2)69.4 (+5.4)68.1 (+6.0)
", + "bbox": [ + 218, + 143, + 777, + 429 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Main Results", + "text_level": 1, + "bbox": [ + 215, + 512, + 372, + 526 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Compariosn with SoTAs. The results in Tab. 1 demonstrate the superiority of our proposed Q&A Prompts. We have a significant improvement over existing SoTA methods, achieving an accuracy of $69.4\\%$ on the validation set and $68.1\\%$ on the test set in A-OKVQA, and an accuracy of $64.3\\%$ in OK-VQA. We also observe that Q&A Prompts achieves a greater improvement on A-OKVQA $(+5.4\\%)$ and $+6.0\\%$ compared to OKVQA $(+2.2\\%)$ . This can be explained that: different from questions in OK-VQA", + "bbox": [ + 212, + 537, + 516, + 734 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "that heavily rely on external knowledge, the questions in A-OKVQA emphasize visually-grounded reasoning with a comprehensive understanding of the image with diverse visual clues, rather than external knowledge retrieval. This aligns with Q&A Prompts's philosophy, that is, instead of relying on accessing any external knowledge base such as Wikipedia [49], ConceptNet [46], or GPT3 [4], our focus is on enhancing reasoning capabilities by discovering rich visual clues hidden in the image and model.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/27682390f75480912a17fd91d28ff51e4a0b0f0b1b030c60e15c7f8255503fc7.jpg", + "table_caption": [ + "Table 1: Results of A-OKVQA [42] and OK-VQA [37] comparing to standard baselines show that our method achieves state-of-the-art performance, outperforming previous methods by a large margin. The best results are highlighted." + ], + "table_footnote": [], + "table_body": "
ModelsA-OKVQA (%)OKVQA (%)
MiniGPT-454.252.5
+ Q&A Prompts61.2 (+7.0)55.9 (+3.4)
LLaVA-1.566.963.3
+ Q&A Prompts72.4 (+5.5)65.2 (+1.9)
InstructBLIP (T5)58.555.1
+ Q&A Prompts64.2 (+5.7)57.3 (+2.2)
", + "bbox": [ + 532, + 566, + 772, + 659 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 2: Comparison of performance with different MLLMs as baselines.", + "bbox": [ + 524, + 669, + 787, + 696 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Haibo Wang, Weifeng Ge", + "bbox": [ + 271, + 114, + 441, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/907393e7f0ff67c8c5b3a99e94ba24a7a4c13187fe4152a63eba77bac49448ac.jpg", + "image_caption": [ + "Question: Which room in the house is this man in?", + "Fig. 5: Representative cases with our Q&A Prompts. We denote the question, tags, Q&A prompts, predictive answers, and ground truth respectively. Our Q&A Prompts can discover rich visual clues in the image and reason out the correct answer." + ], + "image_footnote": [], + "bbox": [ + 223, + 143, + 357, + 210 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ca3702426d46d23bd5aec70da7912dfabff42387c528a2d5005bd84e3e7822ad.jpg", + "image_caption": [ + "Question: Why are the men's vests orange?", + "Image Tags: mirror, sink, reflection, tile wall, tie, bathroom accessory, stare, faucet, cabinet", + "Image Tags: constructor, train, safety vest, train track, passenger, railroad, railway station, platform" + ], + "image_footnote": [], + "bbox": [ + 364, + 143, + 500, + 210 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/bf5b532ef57dd2db96cd18d3cf7ba95d260b5f0dd3f0bdae332145382b00395e.jpg", + "image_caption": [ + "Question: What problem will the people on the ferry face?", + "Image Tags: \ncloudy, travel, ferry, barge, sail, waterway, boat, vessel, float, water, city" + ], + "image_footnote": [], + "bbox": [ + 504, + 143, + 640, + 210 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f0fe938da132948b3fbb05fc7e21b4fd670a9098fcd5e45fc6e8efa7ecbd534c.jpg", + "image_caption": [ + "Question: Why are they holding hands in front of people?", + "Image Tags: \nbible, veil, hand, couple \ngroom, ceremony, \nwedding, father, bride, \nnewlywed, church" + ], + "image_footnote": [], + "bbox": [ + 645, + 143, + 779, + 210 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Q&A Prompts:", + "text_level": 1, + "bbox": [ + 225, + 301, + 300, + 311 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Q: What is the man looking at? A: mirror", + "Q: What is the man washing hands at? A: sink" + ], + "bbox": [ + 225, + 311, + 354, + 353 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Q: Where are the bathroom accessory placed? A: cabinet", + "bbox": [ + 225, + 361, + 356, + 393 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Q&A Prompts:", + "text_level": 1, + "bbox": [ + 367, + 301, + 442, + 311 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Q: What's the man in orange vest? A: constructor Q: What is the yellow and black vehicle? A: train", + "bbox": [ + 367, + 311, + 498, + 352 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Q: What is the area where the train is located? A: platform", + "bbox": [ + 367, + 362, + 491, + 393 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Q&A Prompts:", + "text_level": 1, + "bbox": [ + 508, + 301, + 581, + 311 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Q: What type of weather is it? A: cloudy", + "Q: What is the purpose of the orange and white boat? A: travel" + ], + "bbox": [ + 508, + 311, + 638, + 362 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Q: Where are these boats in? A: city", + "bbox": [ + 508, + 369, + 627, + 393 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Q&A Prompts:", + "text_level": 1, + "bbox": [ + 648, + 301, + 720, + 311 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Q: What is the man in black suit holding? A: bible \nQ: What is the woman wearing on her head? \nA: veil", + "bbox": [ + 647, + 311, + 776, + 362 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Q: What kind of building is this? A: church", + "bbox": [ + 647, + 371, + 774, + 393 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Predictive Answer: \nInstructBLIP: bedroom \nBLIP-2: bedroom \nOurs: bathroom \nGround Truth: bathroom", + "bbox": [ + 225, + 400, + 336, + 460 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Predictive Answer: \nInstructBLIP: fashion \nBLIP-2: working \nOurs: visibility \nGround Truth: visibility", + "bbox": [ + 367, + 398, + 470, + 460 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Predictive Answer: \nInstructBLIP: sickness \nBLIP-2: fall into water \nOurs: raining \nGround Truth: \nraining", + "bbox": [ + 508, + 400, + 616, + 460 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Predictive Answer: \nInstructBLIP: praying \nBLIP-2: praying \nOurs: getting married \nGround Truth: \ngetting married", + "bbox": [ + 647, + 400, + 751, + 460 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Different baselines. To further validate the effectiveness of our method, we conduct experiments to compare various MLLMs as baseline models, including MiniGPT-4 [59], LLaVA-1.5 [30] and InstructBLIP (FLAN-T5-XL). Tab. 2 shows our Q&A Prompts method consistently improves performance on A-OKVQA by approximately $5.5\\%$ to $7.0\\%$ and OKVQA by $1.9\\%$ to $3.4\\%$ , demonstrating the generalizability of Q&A Prompts across different vision-language models.", + "bbox": [ + 212, + 546, + 799, + 637 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Qualitative results. We present qualitative results in Fig. 5. These cases demonstrate scenarios where correct answers are only possible with Q&A prompts. For instance, the important clues provided by Q&A Prompts of the [mirror], [constructor], [cloudy] and [bible], lead to the successful reasoning of correct answers [bathroom], [visibility], [raining], and [getting married] respectively. This showcases the valuable contribution of the Q&A prompts in certain types of tasks, which can explicitly mine rich visual clues and diverse world knowledge hidden in the image to help the model reason out correct answers.", + "bbox": [ + 212, + 638, + 792, + 758 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Analysis of the Generated Questions", + "text_level": 1, + "bbox": [ + 214, + 782, + 565, + 797 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To systematically evaluate the questions, we leverage GPT-4V to measure their reasonableness. Specifically, we randomly sample 250 ", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Q&A Prompts", + "bbox": [ + 632, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/86335532cabda9b9d1bc74978322bb9e4073ee5fd2a8dcf6a9fbf4ec13c2ad1d.jpg", + "image_caption": [ + "Visual Knowledge", + "Q: Where is the dog sitting? A: floor" + ], + "image_footnote": [], + "bbox": [ + 218, + 154, + 310, + 205 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/542adeaf0e3fa05ccff84ac10e717e8dc0ccc5952fa8c029f670c137e0e71cc5.jpg", + "image_caption": [ + "Q: What is in the middle of the plate? A: meat" + ], + "image_footnote": [], + "bbox": [ + 323, + 154, + 413, + 205 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8ee66fab8354b9c061e66c491d3dcf1e1b714bc25e1e527f0e13e19009570023.jpg", + "image_caption": [ + "Commonsien Knowledge", + "Q: Why are people carrying the bags? A: load" + ], + "image_footnote": [], + "bbox": [ + 426, + 152, + 516, + 205 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/28ffe4210dc656a9e678ac8ecd2ae6e58fc6da7dc5a0cee7e23580514e801dc6.jpg", + "image_caption": [ + "Q: What is the glass covered area used for? A: showcase" + ], + "image_footnote": [], + "bbox": [ + 529, + 152, + 620, + 205 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4d7b47f7bf1189f59ffbe1fb7a53cc614609adcd940f1a3c4aaaff8c49248a0a.jpg", + "image_caption": [ + "Required Knowledge Types", + "Q: How is the woman's face visible? A: reflection" + ], + "image_footnote": [], + "bbox": [ + 633, + 176, + 785, + 292 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/36524fbc4f8abf4b489bfced18244225ea842e999bdc78d6350d504377dac9ea.jpg", + "image_caption": [ + "Knowledge base", + "Q: What event is this most likely? A: Christmas", + "Fig. 6: Generated Q&A pairs with different knowledge types." + ], + "image_footnote": [], + "bbox": [ + 218, + 243, + 308, + 295 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1e7dc1283518917169742cf8022fe2d961071456170e2ac7c29a6e1e8a89bb29.jpg", + "image_caption": [ + "Q: What species of bird is this? A: heron", + "Q: How do the people feel here? A: cold" + ], + "image_footnote": [], + "bbox": [ + 323, + 243, + 413, + 295 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/e3b4dc96022c06c4fe3e18b65a7e3877c686f914efe952adb14ea814f0eb1a4f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Q&A sourceScore
Annotations in A-OKVQA9.73
VQG w/ training8.86
VQG w/o training4.12
", + "bbox": [ + 235, + 359, + 454, + 429 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/60479a7f0d9d1d9eb4238514f1ab081c317238cbd859ab1ae18b3979ba26164c.jpg", + "table_caption": [ + "Table 3: The evaluation for reasonableness of different questions." + ], + "table_footnote": [], + "table_body": "
Q&ATraining SetRouge-LA-OKVQA (%)OK-VQA (%)
×--64.062.1
✓--61.961.0
✓A-OKVQA90.869.462.9
✓OK-VQA92.467.164.3
✓Mixture89.468.763.9
", + "bbox": [ + 491, + 359, + 772, + 429 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 4: Ablation studies on training sets for VQG model training.", + "bbox": [ + 478, + 439, + 784, + 465 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "tuples from the QA pairs generated by our VQG model, and ask GPT-4V to return an integer score of reasonableness ranging from 1 to 10 for them. For comparison, we also evaluate the scores of the manually annotated tuples in the A-OKVQA and those generated by directly prompting MLLM without training (with the same number of 250 tuples). Tab. 3 shows the questions generated by our VQG model (VQG w/ training) have a close quality to the manually annotated questions in A-OKVQA and are more reasonable than directly prompting a frozen MLLM (VQG w/o training). See appendix for more details.", + "bbox": [ + 212, + 494, + 784, + 614 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We also annotated what kind of world knowledge was required to answer the 250 sampled questions generated by our VQG model. The choices are the same with A-OKVQA including Commonsense Knowledge, Visual Knowledge, Knowledge Bases, and Physical Knowledge (explained in the appendix). The examples and distribution are shown in Fig. 6. Notably, sometimes there is no clear distinction between these categories and a question may belong to either.", + "bbox": [ + 212, + 614, + 784, + 705 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 726, + 388, + 742 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We investigate the role of our question-answer prompts and the visual-aware prompting module, based on InstructBLIP-Vicuna-7B and the A-OKVQA dataset. Different training sets for generated questions. We investigate how the VQG model will influence the performance by directly prompting the MLLM without training, or using different training sets to train our VQG model, including the A-OKVQA, OK-VQA, and a mixture of them. We report the Rouge-L", + "bbox": [ + 212, + 750, + 795, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Haibo Wang, Weifeng Ge", + "bbox": [ + 271, + 114, + 441, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/cde4595775f80357841fd9c42a6f1248bec208b5f6eb40b16b0837ea9f82c4de.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Fusion strategyAccuracy (%)
Prepend + Q&A66.6
+ Captions64.2
+ Tags64.7
VPM + Q&A69.4
", + "bbox": [ + 230, + 143, + 426, + 213 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/b973c1b3f5c0063813ecfc6b6284e6e577f6d2c5ca11d64ad72ee45a4028fe1c.jpg", + "table_caption": [ + "Table 5: Ablation studies on strategies of fusing Q&A pairs." + ], + "table_footnote": [], + "table_body": "
#NumAccuracy (%)
166.2
467.9
869.4
1669.2
", + "bbox": [ + 457, + 143, + 578, + 213 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/5ef800d341d556edc684fef826f78eb17562c8d38d8746d7e64fa253fd9ffc6a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
VariantsAccuracy (%)
VPM69.4
w/ BERT68.6
w/o Fusion68.6
w/o Decoder68.1
", + "bbox": [ + 594, + 143, + 761, + 213 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 6: Ablation studies Table 7: Ablation studies on the on numbers of question-different components in the visual-answer pairs.", + "bbox": [ + 444, + 222, + 772, + 257 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "score for the generated questions. Results in Tab. 4 show that performance declines severely when using Q&A prompts generated by MLLMs without training. Besides, the best performance is only achieved when the corresponding dataset is utilized as the training set, and a mismatch or mixture of datasets could lead to a slightly weaker performance. This could be attributed to the domain gap in the question types of these datasets.", + "bbox": [ + 212, + 291, + 784, + 381 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Strategies of fusing question-answer pairs. In Tab. 5, we explore a simple approach where we directly prepend the Q&A pairs as a natural language to the instruction (denoted as Prepend + Q&A) without encoded by the visual-aware prompting module (VPM). For comparison, we also replace the question-answer pairs with the same number of image captions generated by BLIP-2 [26] and tags recognized by RAM (denoted as Prepend + Captions and Prepend + Tags). The results in Tab. 5 show that the incorporation of question-answer prompts is more effective than pure captions and tags, and both directly prepending the question-answer pairs and using the VPM to encode them can improve performance. Notably, the models with VPM perform much better, which we attributed to its ability to combat some noise in the generated question-answer pairs and avoid interference with the target question.", + "bbox": [ + 212, + 383, + 785, + 565 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The number of the question-answer prompts. We also study to figure out the effect of using different numbers of question-answer pairs to construct the Q&A Prompts. The results are displayed in Tab. 6. We denote the desired number as $P$ . It can be observed that when $P$ is set to 8, the model achieves optimal performance. We can deduce that when $P$ is too large, it may introduce noisy pairs, while a too small $P$ can result in an ignorance of many essential and informative question-answer pairs.", + "bbox": [ + 212, + 566, + 787, + 672 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Components in visual-aware prompting module. We validate the key components of the visual-aware prompting module in Tab. 7. We first study the effectiveness of using different language models as the prompt encoder, where we replace the text encoder in CLIP with BERT [8] (w/ BERT), which results in a performance drop. This result is consistent with previous findings [38, 45], which suggest that models initialized with CLIP [40] are more suitable for vision-language tasks due to their shared understanding of the visual and textual content. We then remove the visual gated fusion (w/o fusion), which also induces severe performance decline. This is expected since the unimodal language representations tend to involve ambiguity and bias. The additional incorporation of visual embeddings effectively allows for multi-modal interaction beneficial for", + "bbox": [ + 212, + 672, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Q&A Prompts", + "bbox": [ + 632, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/889498b47b68ec8e0249ae1a0cb2e73fb7b21b131e44ae8231d5dafae4f4cdfc.jpg", + "image_caption": [ + "Fig. 7: Qualitative analysis on how Q&A Prompts work." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 401, + 234 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Q&A Prompts:", + "text_level": 1, + "bbox": [ + 406, + 146, + 473, + 156 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Q1: What are the flowing vegetation? A: grass \nQ2: What are the effects of air movement on grass? A: bent \nQ3: What word describes the green grass? A: lush \nQ4: What animal in this image is used for riding? A: horse \nQ5: What is the red machine called? A: tractor \nQ6: Where is the horses standing? A: dirt field \nQ7: How many horses are there? A: couple \nQ8: What type of property is this? A: farm", + "bbox": [ + 405, + 157, + 653, + 232 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Ground Truth: wind \nPredictive Answer: \nw/ Q&A Prompts: wind \nw/o Q&A Prompts: drought", + "bbox": [ + 658, + 146, + 779, + 186 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "w/Q&A Prompts (remove Q1, Q2): drought w/Q&A Prompts (remove Q3~Q8): wind", + "bbox": [ + 658, + 190, + 763, + 231 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "vision-language reasoning. We also evaluate the importance of the resampler decoder (w/o decoder), where we directly feed the long sequence $\\mathbf{F}_m\\in \\mathbb{R}^{L_s\\times d_q}$ into the large language model in Eq. (3) and also caused a performance drop. The sequence length of target question $L_{q}$ is $\\times 10$ times smaller than the sequence length $L_{s}$ of prompts $\\mathcal{S}$ ( $L_{q}\\leq 10$ , $L_{s}\\geq 100$ here), resulting in a distraction to the understanding of the target question. This highlights the importance of resampling $\\mathbf{F}_m\\in \\mathbb{R}^{L_s\\times d_q}$ into a shorter $\\mathbf{F}_p\\in \\mathbb{R}^{k\\times d_q}$ ( $k = 32$ here).", + "bbox": [ + 212, + 287, + 787, + 393 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 A Further Qualitative Analysis of Q&A Prompts", + "text_level": 1, + "bbox": [ + 214, + 414, + 723, + 431 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Fig. 7 further reveals why Q&A Prompts are necessary and how Q&A Prompts work in particular case. To find out [what ambient effect] causes the appearance of grass, the key point is to grasp the movement of them. The Q&A pairs $Q_{1}$ and $Q_{2}$ precisely capture the visual clues that [the grass are flowing] and [air movement cause the grass bent], which are beneficial for reasoning out the answer [wind]. We also test scenarios where we remove $Q_{1}$ and $Q_{2}$ or only reserve them, leading to the incorrect answer [drought] and correct answer [wind] respectively. This indicates that the quality of Q&A Prompts is crucial, particularly when the question needs multiple reasoning steps beyond perception, and Q&A Prompts serve as the intermediate steps that fill the logical gap.", + "bbox": [ + 212, + 441, + 787, + 595 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusions and Limitations", + "text_level": 1, + "bbox": [ + 214, + 614, + 526, + 630 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper has studied the problem of VQA tasks requiring reasoning over diverse world knowledge. We introduce a novel framework, Q&A Prompts, which effectively generates a set of question-answer prompts and encodes them with a visual-aware prompting module, significantly boosting the reasoning ability of current multi-modal large language models. We conducted a series of experiments on the A-OKVQA and OK-VQA benchmarks and achieved significant results compared with previous methods. Extensive ablations and comprehensive analyses have demonstrated the effectiveness and superiority of our method. Limitations. While our method shows promising results, it is essential to acknowledge the potential presence of biases in the data as observed in previous VQA studies [1,7], as well as the lack of the ability for fine-grained counting and Optical Character Recognition. In our future work, we plan to mitigate these biases and weaknesses to enhance the reasoning ability of current models further.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Haibo Wang, Weifeng Ge", + "bbox": [ + 271, + 114, + 441, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China under Grant Nos. 62106051 and the National Key R&D Program of China 2022YFC3601405.", + "bbox": [ + 212, + 175, + 787, + 220 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 243, + 321, + 258 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Agrawal, A., Batra, D., Parikh, D., Kembhavi, A.: Don't just assume; look and answer: Overcoming priors for visual question answering. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018) 14", + "2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: Adv. Neural Inform. Process. Syst. (2022) 4, 10", + "3. Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Zitnick, C.L., Parikh, D.: Vqa: visual question answering. In: Int. Conf. Comput. Vis. (2015) 2, 3, 4, 9", + "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: Adv. Neural Inform. Process. Syst. (2020) 4, 10", + "5. Changpinyo, S., Sharma, P., Ding, N., Soricut, R.: Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021) 4", + "6. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500 (2023) 2, 3, 4, 5, 6, 8, 9, 10", + "7. Dancette, C., Cadene, R., Teney, D., Cord, M.: Beyond question-based biases: Assessing multimodal shortcut learning in visual question answering. In: Int. Conf. Comput. Vis. (2021) 2, 14", + "8. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: NAACL (2019) 13", + "9. Gao, F., Ping, Q., Thattai, G., Reganti, A., Wu, Y.N., Natarajan, P.: Transform-retrieve-produce: Natural language-centric outside-knowledge visual question answering. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) 10", + "0. Gao, L., Biderman, S., Black, S., Golding, L., Hoppe, T., Foster, C., Phang, J., He, H., Thite, A., Nabeshima, N., et al.: The pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027 (2020) 4", + "1. Gardères, F., Ziaeffard, M., Abeloos, B., Lecue, F.: ConceptBert: Concept-aware representation for visual question answering. In: EMNLP (2020) 10", + "2. Geman, D., Geman, S., Hallonquist, N., Younes, L.: Visual Turing test for computer vision systems. Proceedings of the National Academy of Sciences 112(12), 3618-3623 (2015) 1", + "3. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In: IEEE Conf. Comput. Vis. Pattern Recog. (2017) 2, 3, 4", + "4. Gui, L., Wang, B., Huang, Q., Hauptmann, A., Bisk, Y., Gao, J.: Kat: A knowledge augmented transformer for vision-and-language. In: NAACL (2022) 2, 4, 10", + "5. Guo, J., Li, J., Li, D., Tiong, A.M.H., Li, B., Tao, D., Hoi, S.C.: From images to textual prompts: Zero-shot vqa with frozen large language models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) 4, 10" + ], + "bbox": [ + 225, + 273, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Q&A Prompts", + "bbox": [ + 630, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Guo, Y., Nie, L., Wong, Y., Liu, Y., Cheng, Z., Kankanhalli, M.: A unified end-to-end retriever-reader framework for knowledge-based vqa. In: ACM Int. Conf. Multimedia (2022) 10", + "17. Hu, Y., Hua, H., Yang, Z., Shi, W., Smith, N.A., Luo, J.: Promptcap: Prompt-guided task-aware image captioning. arXiv preprint arXiv:2211.09699 (2022) 4, 10", + "18. Hu, Z., Iscen, A., Sun, C., Wang, Z., Chang, K.W., Sun, Y., Schmid, C., Ross, D.A., Fathi, A.: **Retrieval-augmented visual-language pre-training with multi-source multimodal knowledge memory. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) 2, 4, 10", + "19. Huang, X., Zhang, Y., Ma, J., Tian, W., Feng, R., Zhang, Y., Li, Y., Guo, Y., Zhang, L.: Tag2text: Guiding vision-language model via image tagging. arXiv preprint arXiv:2303.05657 (2023) 3, 4, 5, 7", + "20. Jaegle, A., Borgeaud, S., Alayrac, J.B., Doersch, C., Ionescu, C., Ding, D., Koppula, S., Zoran, D., Brock, A., Shelhamer, E., et al.: Perceiver io: A general architecture for structured inputs & outputs. In: Int. Conf. Learn. Represent. (2022) 8", + "21. Jiang, Y., Natarajan, V., Chen, X., Rohrbach, M., Batra, D., Parikh, D.: Pythia v0. 1: The winning entry to the vqa challenge 2018. arxiv 2018. arXiv preprint arXiv:1807.09956 (2018) 10", + "22. Johnson, J., Hariharan, B., Van Der Maaten, L., Fei-Fei, L., Lawrence Zitnick, C., Girshick, R.: Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2901-2910 (2017) 2", + "23. Kamath, A., Clark, C., Gupta, T., Kolve, E., Hoiem, D., Kembhavi, A.: Webly supervised concept expansion for general purpose vision models. In: Eur. Conf. Comput. Vis. (2022) 10", + "24. Khan, Z., BG, V.K., Schulter, S., Yu, X., Fu, Y., Chandraker, M.: Q: How to specialize large vision-language models to data-scarce vqa tasks? a: Self-train on unlabeled images! In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) 4", + "25. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023) 4", + "26. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. In: Int. Conf. Mach. Learn. (2023) 2, 4, 5, 10, 13", + "27. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: common objects in context. In: Eur. Conf. Comput. Vis. (2014) 4", + "28. Lin, W., Chen, J., Mei, J., Coca, A., Byrne, B.: Fine-grained late-interaction multimodal retrieval for retrieval augmented visual question answering. In: Adv. Neural Inform. Process. Syst. (2023) 10", + "29. Lin, Y., Xie, Y., Chen, D., Xu, Y., Zhu, C., Yuan, L.: Revive: Regional visual representation matters in knowledge-based visual question answering. In: Adv. Neural Inform. Process. Syst. (2022) 2, 4, 10", + "30. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: Adv. Neural Inform. Process. Syst. (2023) 2, 3, 4, 11", + "31. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: Int. Conf. Learn. Represent. (2019) 9", + "32. Lu, J., Batra, D., Parikh, D., Lee, S.: Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In: Adv. Neural Inform. Process. Syst. (2019) 10" + ], + "bbox": [ + 217, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Haibo Wang, Weifeng Ge", + "bbox": [ + 271, + 114, + 441, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Lu, J., Clark, C., Zellers, R., Mottaghi, R., Kembhavi, A.: Unified-io: A unified model for vision, language, and multi-modal tasks. arXiv preprint arXiv:2206.08916 (2022) 10", + "34. Luo, G., Zhou, Y., Ren, T., Chen, S., Sun, X., Ji, R.: Cheap and quick: Efficient vision-language instruction tuning for large language models. arXiv preprint arXiv:2305.15023 (2023) 4", + "35. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008) 6", + "36. Marino, K., Chen, X., Parikh, D., Gupta, A., Rohrbach, M.: Krisp: Integrating implicit and symbolic knowledge for open-domain knowledge-based vqa. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021) 10", + "37. Marino, K., Rastegari, M., Farhadi, A., Mottaghi, R.: Ok-vqa: A visual question answering benchmark requiring external knowledge. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019) 2, 3, 4, 5, 9, 10", + "38. Merullo, J., Castricato, L., Eickhoff, C., Pavlick, E.: Linearly mapping from image to text space. In: Int. Conf. Learn. Represent. (2023) 13", + "39. Mokady, R., Hertz, A., Bermano, A.H.: Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734 (2021) 10", + "40. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Int. Conf. Mach. Learn. (2021) 9, 13", + "41. Reimers, N., Gurevych, I.: Sentence-bert: Sentence embeddings using siamese bert-networks. In: EMNLP (2019) 6, 7", + "42. Schwenk, D., Khandelwal, A., Clark, C., Marino, K., Mottaghi, R.: A-okvqa: A benchmark for visual question answering using world knowledge. In: Eur. Conf. Comput. Vis. (2022) 2, 3, 4, 5, 9, 10", + "43. Selvaraju, R.R., Tendulkar, P., Parikh, D., Horvitz, E., Ribeiro, M.T., Nushi, B., Kamar, E.: Squinting at vqa models: Introspecting vqa models with sub-questions. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020) 4", + "44. Shao, Z., Yu, Z., Wang, M., Yu, J.: Prompting large language models with answer heuristics for knowledge-based visual question answering. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) 4, 10", + "45. Shen, S., Li, L.H., Tan, H., Bansal, M., Rohrbach, A., Chang, K.W., Yao, Z., Keutzer, K.: How much can clip benefit vision-and-language tasks? In: Int. Conf. Learn. Represent. (2022) 13", + "46. Speer, R., Chin, J., Havasi, C.: Conceptnet 5.5: An open multilingual graph of general knowledge. In: AAAI (2017) 4, 10", + "47. Tan, H., Bansal, M.: Lxmert: Learning cross-modality encoder representations from transformers. In: EMNLP (2019) 10", + "48. Uehara, K., Duan, N., Harada, T.: Learning to ask informative sub-questions for visual question answering. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) 4", + "49. Vrandecic, D., Krötzsch, M.: Wikidata: a free collaborative knowledgebase. Communications of the ACM 57(10), 78-85 (2014) 4, 10", + "50. Wang, P., Wu, Q., Shen, C., Dick, A., Van Den Hengel, A.: Fvqa: Fact-based visual question answering. IEEE transactions on pattern analysis and machine intelligence 40(10), 2413-2427 (2017) 2, 4", + "51. Wang, P., Wu, Q., Shen, C., Hengel, A.v.d., Dick, A.: Explicit knowledge-based reasoning for visual question answering. arXiv preprint arXiv:1511.02570 (2015) 2, 4", + "52. Wu, J., Lu, J., Sabharwal, A., Mottaghi, R.: Multi-modal answer validation for knowledge-based vqa. In: AAAI (2021) 10" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Q&A Prompts", + "bbox": [ + 632, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "53. Yang, Z., Gan, Z., Wang, J., Hu, X., Lu, Y., Liu, Z., Wang, L.: An empirical study of gpt-3 for few-shot knowledge-based vqa. In: AAAI (2022) 10", + "54. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023) 2, 4", + "55. Zellers, R., Bisk, Y., Farhadi, A., Choi, Y.: From recognition to cognition: Visual commonsense reasoning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019) 2", + "56. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199 (2023) 4", + "57. Zhang, Y., Huang, X., Ma, J., Li, Z., Luo, Z., Xie, Y., Qin, Y., Luo, T., Li, Y., Liu, S., et al.: Recognize anything: A strong image tagging model. arXiv preprint arXiv:2306.03514 (2023) 3, 4, 5, 7, 9", + "58. Zhang, Z., Zhang, A., Li, M., Zhao, H., Karypis, G., Smola, A.: Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923 (2023) 8", + "59. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023) 3, 4, 11", + "60. Zhu, Z., Yu, J., Wang, Y., Sun, Y., Hu, Y., Wu, Q.: Mucko: multi-layer cross-modal knowledge reasoning for fact-based visual question answering. In: IJCAI (2020) 10" + ], + "bbox": [ + 215, + 146, + 784, + 438 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Haibo Wang, Weifeng Ge", + "bbox": [ + 271, + 114, + 441, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Q&A Prompts_ Discovering Rich Visual Clues through Mining Question-Answer Prompts for VQA requiring Diverse World Knowledge/d6b849aa-8c9f-421c-b457-5802b3bd4aca_model.json b/2024/Q&A Prompts_ Discovering Rich Visual Clues through Mining Question-Answer Prompts for VQA requiring Diverse World Knowledge/d6b849aa-8c9f-421c-b457-5802b3bd4aca_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7a10a12b515f9a07752a37fa2032942905511cba --- /dev/null +++ b/2024/Q&A Prompts_ Discovering Rich Visual Clues through Mining Question-Answer Prompts for VQA requiring Diverse World Knowledge/d6b849aa-8c9f-421c-b457-5802b3bd4aca_model.json @@ -0,0 +1,3063 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.239, + 0.141, + 0.768, + 0.21 + ], + "angle": 0, + "content": "Q&A Prompts: Discovering Rich Visual Clues through Mining Question-Answer Prompts for VQA requiring Diverse World Knowledge" + }, + { + "type": "text", + "bbox": [ + 0.36, + 0.234, + 0.644, + 0.252 + ], + "angle": 0, + "content": "Haibo Wang\\(^{1,2}\\) and Weifeng Ge\\(^{1,2\\boxtimes}\\)" + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.262, + 0.726, + 0.306 + ], + "angle": 0, + "content": "1 Nebula AI Group, School of Computer Science, Fudan University \n2 Shanghai Key Lab of Intelligent Information Processing hbwang22@m.fudan.edu.cn, wfge@fudan.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.345, + 0.744, + 0.679 + ], + "angle": 0, + "content": "Abstract. With the breakthrough of multi-modal large language models (MLLMs), answering complex visual questions that demand advanced reasoning abilities and world knowledge has become a much more important testbed for developing AI models than ever. However, equipping MLLMs with robust cross-modality reasoning ability remains challenging since the cognition scheme of humans has not been understood systematically. In this paper, we believe that if we can collect rich visual clues, we will recognize the image more accurately, understand the question better, recall relevant knowledge more easily, and finally reason out the answer. We discover these rich visual clues by mining question-answer pairs in images and sending them into multi-modal large language models as prompts. We call the proposed method Q&A Prompts. Specifically, we first use the image-answer pairs and the corresponding questions in the training set as inputs and outputs to train a visual question generation (VQG) model. Then, we use an image tagging model to identify various instances and send packaged image-tag pairs into the VQG model to generate relevant questions with the extracted image tags as answers. Finally, we encode these generated question-answer pairs as prompts with a visual-aware prompting module and send them into pretrained MLLMs to reason out the final answers. Experimental results show that, compared with state-of-the-art methods, our Q&A Prompts achieves substantial improvements on the challenging visual question answering datasets requiring reasoning over diverse world knowledge, such as OK-VQA and A-OKVQA. Codes will be available at link." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.691, + 0.741, + 0.721 + ], + "angle": 0, + "content": "Keywords: multi-modal large language model \\(\\cdot\\) visual question answering \\(\\cdot\\) visual language reasoning" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.747, + 0.377, + 0.763 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Visual Question Answering (VQA) has longstanding been considered a Visual Turing Test [12] in the machine learning community. To solve this Turing test, artificial intelligence (AI) models are expected to be equipped with the human capabilities of visual recognition, language understanding, logical reasoning, world" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.442, + 0.129 + ], + "angle": 0, + "content": "Haibo Wang, Weifeng Ge" + }, + { + "type": "image_caption", + "bbox": [ + 0.229, + 0.145, + 0.373, + 0.156 + ], + "angle": 0, + "content": "VQA Tasks well-solved by MLLMs" + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.159, + 0.377, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.393, + 0.145, + 0.528, + 0.156 + ], + "angle": 0, + "content": "VQA Tasks unsolved by MLLMs" + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.158, + 0.528, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.235, + 0.372, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.235, + 0.528, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.145, + 0.76, + 0.156 + ], + "angle": 0, + "content": "VQA Tasks solved by MLLMs with Q&A Prompts" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.158, + 0.774, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.321, + 0.788, + 0.378 + ], + "angle": 0, + "content": "Fig. 1: Illustration of VQA tasks that require simple perception abilities ((a) and (b)) and those demanding chains of reasoning over diverse world knowledge ((c) and (d)). The example on the right ((e) and (f)) indicates MLLMs with Q&A prompts can solve difficult VQA problems that require reasoning over diverse world knowledge." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.785, + 0.518 + ], + "angle": 0, + "content": "knowledge referring, etc. There are a variety of datasets focusing on different problems in VQA, such as the perception and language understanding problem [3, 13], reasoning on procedurally generated images [22], and structured factual knowledge [50, 51] or commonsense knowledge [55]. Recently, with the advent of multi-modal large language models (MLLMs) [6, 26, 30, 54], significant progress has been achieved in perception-based VQA tasks [3, 13]. Fig. 1 (a) and Fig. 1 (b) show such VQA tasks, which can be reliably solved by MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.788, + 0.763 + ], + "angle": 0, + "content": "However, for more complex VQA tasks [37, 42] as in Fig. 1 (c) and Fig. 1 (d), that feature the necessity of diverse world knowledge and complex chains of reasoning, state-of-the-art methods [14, 18, 29] even MLLM models [6, 26] fail to give the correct answers. Different from perception-based VQA [3, 13] and \"closed\" knowledge-based VQA [50, 51], visual question reasoning in OK-VQA [37] and A-OKVQA [42] requires not only diverse forms of world knowledge but also need complex reasoning logic. In Fig. 1 (d), where a woman sits on the ground with her pet dogs and holds an umbrella, the question is [Why is the woman holding the umbrella?] Simply associating the keyword [umbrella] with the word [rain] due to their frequent co-occurrence in the pre-training corpus [7] will lead to the wrong answer. To reason out the correct answer, AI models need to consider rich visual clues presented in the image. For example, in Fig. 1 (f), having discovered the clues including [the woman is camping] and [the weather is shiny] from the Q&A pairs, the model can correctly fill the logic gap and arrive at the answer [for shade]. This evidence indicates the necessity of collecting rich visual clues to achieve good results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Based on the analysis above, we believe that collecting visual clues of various instances in images will help MLLM recognize the image more accurately, understand the question better, recall relevant knowledge more easily, and finally generate the reasoning results more intelligently. To investigate this hypothesis, we design a novel VQA framework called \\( Q\\mathcal{E}A \\) Prompts, which extracts nu" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.633, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Q&A Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.313 + ], + "angle": 0, + "content": "merous question-answer pairs associated with different perspectives in images as prompts for MLLMs. We choose an instance in the given image as the answer and generate the related question to form a question-answer pair. We generate the question-answer pairs for almost every instance to get a bag of Q&A prompts. The reason we choose question-answer pairs as prompts comes from two folds: First, Q&A prompts can focus on diverse instances in images like objects, scenes, or actions, thereby offering multiple different perspectives to understand the given image and the target question; Second, the knowledge and reasoning insights hidden in these question-answer pairs can help to explicitly uncover more relevant world knowledge, which may be critical to filling the logic gap between perception and reasoning when answering the target question." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.603 + ], + "angle": 0, + "content": "In Q&A Prompts, there are three key stages, which are the visual question generation (VQG) model training stage, the question-answer prompts generation stage, and the visual-language reasoning stage. In the VQG model training stage, we gather question-answer pairs in a VQA training set to train a VQG model that can map an answer together with the given image to a correlated question. Subsequently, in the question-answer prompts generation stage, we leverage a pre-trained image tagging model [19, 57] to recognize various objects, scenes, and actions in images with text tags. We employ these tags as answers and send them into the VQG model along with the corresponding images to generate questions. Finally, in the visual-language reasoning stage, we encode the generated question-answer pairs with a visual-aware prompting module and send them together with image and question features into a frozen language model to reason out the answer. Since previous VQA benchmarks [3,13] have seen saturated performance, we conduct experiments on the more challenging OK-VQA [37] and A-OKVQA [42] datasets, which are newer and larger testbeds for MLLM to investigate their abilities in accomplishing reasoning tasks requiring diverse world knowledge. Experimental results show that our Q&A Prompts can substantially improve the reasoning ability of MLLMs, such as InstructBLIP [6], LLaVA [30], and MiniGPT-4 [59]. In summary, we make the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.613, + 0.787, + 0.689 + ], + "angle": 0, + "content": "- We propose a novel VQA framework that effectively generates and leverages question-answer pairs as prompts to improve the reasoning ability of MLLMs. It provides a new perspective for the community to design VQA solutions by explicitly collecting rich visual clues to bridge the logic gap between perception and reasoning when answering the target question." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.69, + 0.787, + 0.764 + ], + "angle": 0, + "content": "- We design a novel question-answer prompts generation scheme with a VQG model and an image tagging model, which can generate Q&A prompts for recognizable objects, scenes, and activities in images. Besides, we design a new visual-aware prompting module to encode these prompts efficiently for the subsequent reasoning." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "- We test the method on the challenging OK-VQA [37] and A-OKVQA [42] benchmarks, which study reasoning over diverse forms of world knowledge. The proposed Q&A Prompts achieves an accuracy of \\(68.1\\%\\) and \\(64.3\\%\\) on A-OKVQA [42] and OK-VQA [37], outperforming previous state-of-the-art methods by clear margins." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.613, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.442, + 0.129 + ], + "angle": 0, + "content": "Haibo Wang, Weifeng Ge" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.388, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.787, + 0.356 + ], + "angle": 0, + "content": "Multimodal Large Language Models (MLLMs). MLLMs [2,6,25,26,34,54, 56,59] have shown strong ability in image-language understanding and reasoning, by adapting frozen language models to frozen image encoders with trainable connection modules. For instance, Flamingo [2] incorporates visual features into the language model with gated cross-attention blocks. LLaVA [30], instead, directly projects visual features into the space of text embeddings using a linear layer. BLIP-2 and InstructBLIP [6,26] introduce a more complex Q-Former to bridge the modality gap. These models align images and language based on large-scale image-text pre-training [5,10,27] and conduct reasoning with language models. In this paper, we mine more numerous question-answer pairs and send them into the MLLM as prompts. These prompts provide rich visual clues and uncover relevant world knowledge, which is beneficial to reasoning out correct answers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.357, + 0.787, + 0.629 + ], + "angle": 0, + "content": "VQA Requiring Reasoning over Diverse World Knowledge. Different from classical VQA tasks [3, 13, 50, 51], VQA requiring reasoning over world knowledge involves inferring intents, goals, the physics of the world, and the social dynamics of individuals. More recently, A-OKVQA [42] and OK-VQA [37], which our work focuses on, present large-scale challenging VQA datasets that encompass questions requiring reasoning based on diverse world knowledge, including visual knowledge, commonsense knowledge, and factoid knowledge. Previous methods like Img2LLM [15], Prophet [44], and PromptCap [17] exploit the rich hidden knowledge and strong reasoning abilities of LLMs (e.g., GPT3 [4]) to solve such problems. These methods convert images into natural languages and construct various prompt templates for LLMs with in-context learning [4] to generate answers. Another group of works, such as KAT [14], REVIVE [29], and REVEAL [18], directly encode questions and relevant knowledge retrieved from external knowledge bases [4, 46, 49] to train end-to-end models. Unlike these methods, we design a visual-question prompts generation scheme to mine the information in images and unambiguous words to reduce confusion when answering the target questions. With Q&A prompts, different types of world knowledge are incorporated into the reasoning process to help MLLMs infer the correct answers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Visual Question Generation (VQG). Various works have explored Visual Question Generation (VQG) in the context of VQA tasks. SQuINT [43] and SelTDA [24] focus on dataset augmentation to enhance the generalization of VQA models. SQuINT employed manual workers to annotate questions, while SelTDA generated questions automatically using unlabeled images. Another line of research, exemplified by [48], employs a VQG model to generate informative sub-questions. These sub-questions are then used alongside the original question during inference, enabling another VQA model to answer both simultaneously. Different from previous works, our VQG model uses the tags of images [19, 57] as answers for generated questions. This approach ensures that our VQG model generates more relevant questions with additional details and perspectives. Furthermore, we introduce a novel module to encode the question-answer pairs as prompts for MLLMs, instead of directly feeding them as the formation of natural language into the model [48]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.633, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Q&A Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.788, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.346, + 0.788, + 0.417 + ], + "angle": 0, + "content": "Fig. 2: (a) Firstly, we train a VQG model that can ask informative questions given the image and specified answer. (b) Then, we exploit the image tagging model to extract image tags as the answers to generated questions and obtain diverse question-answer prompts. (c) At last, we feed the image, question, and question-answer prompts into the VQA model to perform visual-language reasoning in (c)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.444, + 0.33, + 0.459 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.789, + 0.719 + ], + "angle": 0, + "content": "Q&A Prompts is a conceptually simple three-stage framework, and Fig. 2 gives an overview: a VQG model training stage, a question-answer prompts generation stage, and a visual-language reasoning stage. Specifically, in the VQG model training stage (Fig. 2 (a)), we train a vanilla MLLM with a broad range of image-answer pairs as input and the corresponding question as the target output to learn a mapping from answers to questions (detailed in Sec. 3.1). In the prompts generation stage (Fig. 2 (b)), we exploit an image tagging model [19, 57] to obtain relevant tags including attributes, objects, actions, and scenes in images. These tags serve as answers and will be fed into the VQG model trained in the first stage to generate diverse questions containing rich visual clues and world knowledge (detailed in Sec. 3.2). In the final reasoning stage (Fig. 2 (c)), we utilize our proposed visual-aware prompting module to encode these question-answer pairs into prompts, which will be fed into the MLLM such as BLIP-2 [26] or InstructBLIP [6] together with the image and the target question, to predict the answer (detailed in Sec. 3.3). The whole pipeline is flexible and effective in mining informative visual clues to correctly answer the given questions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.74, + 0.52, + 0.756 + ], + "angle": 0, + "content": "3.1 Stage-1: Train the VQG Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Given an image \\(\\mathcal{V}\\) and answer \\(\\mathcal{A}\\), the VQG model is responsible for posing an informative question \\(\\mathcal{Q}\\) which can be properly answered by \\(\\mathcal{A}\\). To explore as many visual clues in the generated questions as possible, and avoid generating general questions containing limited clues (e.g., \"What is this?\"), we utilize the A-OKVQA [42] or OK-VQA [37] as the training set \\(D\\) for our VQG model, since" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.444, + 0.129 + ], + "angle": 0, + "content": "Haibo Wang, Weifeng Ge" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.145, + 0.396, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.144, + 0.591, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.144, + 0.783, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.227, + 0.323, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.324, + 0.228, + 0.5, + 0.28 + ], + "angle": 0, + "content": "A-OKVQA What sport do the people have equipment for? \nVQG Which object in this image used for winter sports? \nZero-shot Is this a snowboard?" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.228, + 0.617, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.619, + 0.228, + 0.782, + 0.28 + ], + "angle": 0, + "content": "A-OKVQA What is the vehicle traveling on? \nVQG What is this type of transportation called? \nZero-shot What is this?" + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.282, + 0.329, + 0.293 + ], + "angle": 0, + "content": " snowboard" + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.283, + 0.587, + 0.293 + ], + "angle": 0, + "content": " rail" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.304, + 0.785, + 0.348 + ], + "angle": 0, + "content": "Fig. 3: Visualizations and T-SNE embedding distributions of the questions in manually annotated A-OKVQA (green), and the questions generated by our VQG model (blue) or by directly prompting a frozen MLLM (orange)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.785, + 0.45 + ], + "angle": 0, + "content": "the questions in their annotations are diverse and specific, covering broader areas of knowledge. To create the VQG model that approximates \\( P(\\mathcal{Q}|\\mathcal{V},\\mathcal{A}) \\), we treat the problem of learning such a model as a text-generation problem and wish to train the MLLM (e.g., InstructBLIP [6]) as our VQG model with \\( D \\) due to its strong image-language reasoning ability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.572 + ], + "angle": 0, + "content": "Specifically, as shown in Fig. 2 (a), the VQG model comprises a frozen image encoder, a trainable connection module, and a frozen LLM. We extract image embeddings \\(\\mathbf{E}_v\\in \\mathbb{R}^{n\\times d_v}\\) of \\(\\nu\\) with the image encoder (\\(n\\) is the patch number, \\(d_v\\) is the embedding dimension), and then feed \\(\\mathbf{E}_v\\) into the connection module to obtain fixed-length visual features \\(\\mathbf{F}_v\\in \\mathbb{R}^{k\\times d_q}\\), which will be sent into the LLM together with the instruction. We write the instruction as the template containing the answer \\(\\mathcal{A}\\), such as \"Given the image, generate a question whose answer is: \\(\\langle \\mathcal{A}\\rangle\\).\" The model is trained using the cross-entropy loss with parameters \\(\\theta\\):" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.58, + 0.787, + 0.625 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {V Q G}} = - \\sum_ {t = 1} ^ {L _ {q}} \\log P _ {\\theta} \\left(\\mathcal {Q} _ {t} \\mid \\mathcal {Q} _ {< t}, \\mathcal {V}, \\mathcal {A}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.768 + ], + "angle": 0, + "content": "where \\(\\mathcal{Q}_t\\) is predicted autoregressively at position \\(t\\), and \\(L_{q}\\) is the sequence length of the question \\(\\mathcal{Q}\\). We emphasize the necessity of training the VQG model, instead of prompting a frozen MLLM in a zero-shot manner. In Fig. 3, we use sentence-bert [41] to obtain the dense embedding vectors of each question and show the T-SNE [35] embedding of them, as well as visualization examples. We can observe that the questions generated by our VQG model (blue) are more diverse and specific like questions in A-OKVQA (green), while the questions generated by MLLM without training (orange) are often general and limited with low differentiation. Our following experiments will further verify this." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.787, + 0.634, + 0.803 + ], + "angle": 0, + "content": "3.2 Stage-2: Generate Question-Answer Prompts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Once the VQG model has been obtained, the generation of question-answer pairs can proceed. In this stage, we will explicitly generate a wide range of question-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.632, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Q&A Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.145, + 0.782, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.246, + 0.785, + 0.274 + ], + "angle": 0, + "content": "Fig. 4: Overview of the visual-aware prompting module, consisting of the prompt encoder, visual gated fusion, and resampler decoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.305, + 0.784, + 0.426 + ], + "angle": 0, + "content": "answer pairs to help the model have a deeper and broader understanding of the image. In order to generate diverse questions covering various perspectives of the image, we should ensure the answers fed into the VQG model are sufficient in varied forms. Therefore, we adopt a strong image tagging model [19, 57] to produce rich tags as answers. Different from object detectors that can only recognize limited object labels, image tagging models aim to identify various open-vocabulary elements in images, including attributes, objects, actions, scenes, etc, which can provide an appropriate source of the answers we need." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.426, + 0.785, + 0.531 + ], + "angle": 0, + "content": "We utilize the RAM [57] model as our image tagging model due to its superior recognition ability. As shown in Fig. 2 (b), we feed the image \\(\\mathcal{V}\\) into the tagging model and obtain various tags for each image. To make the tags of each image more distinguishable, we remove the most frequent ones containing general information, such as 'person', 'food', 'man', etc., resulting in an average of 14 tags per image. We denote this set of tags as \\(\\mathcal{O} = \\{o_i\\}_{i=1}^M\\), where \\(M\\) is the number of tags associated with an image." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.531, + 0.785, + 0.637 + ], + "angle": 0, + "content": "Using these tags as answers, we input them into the trained VQG model along with the image, as shown in Fig. 2 (b), to obtain corresponding questions that answer these tags. To ensure that the generated questions (denoted as \\(\\mathcal{T} = \\{t_i\\}_{i=1}^M\\)) contribute effectively to answering the target question, we rank the generated questions \\(\\mathcal{T} = \\{t_i\\}_{i=1}^M\\) based on their similarity with the target question \\(\\mathcal{Q}\\), and adopt the questions with the Top-P similarities as the final obtained question prompts \\(\\mathcal{U}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.357, + 0.648, + 0.785, + 0.674 + ], + "angle": 0, + "content": "\\[\n\\mathcal {U} = \\left\\{u _ {p} \\right\\} _ {p = 1} ^ {P} = \\underset {t _ {i} \\in \\mathcal {T}} {\\arg \\operatorname {T o p - P}} \\langle \\mathrm {T} (t _ {i}), \\mathrm {T} (\\mathcal {Q}) \\rangle \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.686, + 0.788, + 0.763 + ], + "angle": 0, + "content": "where \\(\\mathrm{T}(\\cdot)\\) represents the embedding model [41], which encodes sentences into embedding vectors for similarity computation. The \\(\\langle \\cdot, \\cdot \\rangle\\) operator is the inner product, computing the cosine similarity between two sentences. We then concatenate each \\(u_i\\) in \\(\\mathcal{U}\\) and its corresponding \\(o_i\\) in \\(\\mathcal{O}\\) using the template \"Question: \\(\\langle u_i \\rangle\\) Answer: \\(\\langle o_i \\rangle\\)\", resulting in the final question-answer pairs as prompts \\(\\mathcal{S}\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.785, + 0.588, + 0.8 + ], + "angle": 0, + "content": "3.3 Stage-3: Reasoning with Q&A Prompts" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "After obtaining these question-answer pairs involving rich visual clues and world knowledge, we use the proposed visual-aware prompting module, as shown in" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.444, + 0.13 + ], + "angle": 0, + "content": "Haibo Wang, Weifeng Ge" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "Fig. 4, to encode these question-answer pairs into a set of prompt embeddings \\(\\mathbf{F}_p\\in \\mathbb{R}^{k\\times d_q}\\) for our VQA model. We still utilize the same MLLM (e.g., Instruct-BLIP [6]) as our VQA model for the final reasoning. As in Fig. 2 (c), the encoded prompt embeddings \\(\\mathbf{F}_p\\) will be fed into the frozen LLM in MLLM, together with the instruction embeddings and visual features \\(\\mathbf{F}_v\\) to predict the answer \\(\\mathcal{A}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.233, + 0.786, + 0.25 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} = \\operatorname {L L M} \\left(\\left[ \\mathbf {F} _ {v}; \\mathbf {F} _ {p}; \\operatorname {E m b e d} (\\operatorname {I n s}) \\right]\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.259, + 0.784, + 0.319 + ], + "angle": 0, + "content": "Embed is the word embedding layer in LLM, the operator \\([\\cdot]\\) means concatenation, and Ins is the instruction including the target question \\(\\mathcal{Q}\\), represented as \"Question: \\(\\langle \\mathcal{Q} \\rangle\\)? Answer:\". We then introduce how the visual-aware prompting module bootstrapped the reasoning ability of MLLM within Q&A Prompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.32, + 0.788, + 0.425 + ], + "angle": 0, + "content": "Prompt Encoder. The prompt encoder is responsible for obtaining the semantics representations of the prompts \\( S \\). Specifically, we choose encoder-based language models as our prompt encoder due to their strong ability in language understanding. To preserve the well-trained representations, we keep the parameters of the prompt encoder frozen. We feed the prompts \\( S \\) into the prompt encoder and get the encoded \\( \\mathbf{F}_s \\in \\mathbb{R}^{L_s \\times d_q} \\), where \\( L_s \\) is the sequence length of question-answer prompts and \\( d_q \\) is the hidden size of the prompt encoder." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.425, + 0.788, + 0.547 + ], + "angle": 0, + "content": "Visual Gated Fusion. Since the output \\(\\mathbf{F}_s\\) from the prompt encoder is an unimodal language representation, to fully utilize the explicit information in images, the visual gated fusion adaptively incorporates image information into this representation with the visual embeddings \\(\\mathbf{E}_v \\in \\mathbb{R}^{n \\times d_v}\\). Specifically, in Fig. 4, we use a single-head attention mechanism, similar to MM-COT [58], to correlate the language representation with the visual embeddings. We define the query, key and value as \\(\\mathbf{Q} = \\mathbf{F}_s \\in \\mathbb{R}^{L_s \\times d_q}\\), \\(\\mathbf{K} = \\mathbf{V} = \\mathrm{MLP}(\\mathbf{E}_v) \\in \\mathbb{R}^{n \\times d_q}\\), where MLP here is projecting \\(d_v\\) to \\(d_q\\). The attention output \\(\\mathbf{F}_v^{attn}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.32, + 0.556, + 0.787, + 0.595 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {v} ^ {\\text {a t t n}} = \\mathbf {Q} + \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {T}}{\\sqrt {d _ {q}}}\\right) \\mathbf {V}, \\quad \\mathbf {F} _ {v} ^ {\\text {a t t n}} \\in \\mathbb {R} ^ {L _ {s} \\times d _ {q}} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.603, + 0.787, + 0.634 + ], + "angle": 0, + "content": "Next, we apply the gated fusion mechanism to merge \\(\\mathbf{F}_s\\) and \\(\\mathbf{F}_v^{attn}\\). The fused representation \\(\\mathbf{F}_m \\in \\mathbb{R}^{L_s \\times d_q}\\) is obtained using the following equations:" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.643, + 0.786, + 0.68 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\lambda = \\operatorname {S i g m o i d} \\left(\\mathbf {F} _ {s} \\mathbf {W} _ {s} + \\mathbf {F} _ {v} ^ {\\text {a t t n}} \\mathbf {W} _ {v}\\right) \\\\ \\mathbf {F} _ {m} = (1 - \\lambda) \\mathbf {F} _ {s} + \\lambda \\mathbf {F} _ {v} ^ {\\text {a t t n}} \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.784, + 0.734 + ], + "angle": 0, + "content": "in which \\(\\mathbf{W}_s\\) and \\(\\mathbf{W}_v\\) are learnable parameters. The gated fusion allows the model to adaptively blend the information from the language representation and visual embeddings, resulting in the multimodal representations \\(\\mathbf{F}_m\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Resampler Decoder. Although \\(\\mathbf{F}_m \\in \\mathbb{R}^{L_s \\times d_q}\\) in Eq. (5) is the multimodal representation of \\(\\mathcal{S}\\), we cannot directly concatenate it with \\(\\mathbf{F}_v\\) and instruction embeddings as in Eq. (3), due to the long sequence length of \\(L_s\\), which can result in extra computational burden and distraction to the target question \\(\\mathcal{Q}\\). In light of this, we employ a lightweight resampler decoder inspired by Perceiver [20], to map \\(\\mathbf{F}_m \\in \\mathbb{R}^{L_s \\times d_q}\\) into smaller-sized prompt embeddings \\(\\mathbf{F}_p \\in \\mathbb{R}^{k \\times d_q}\\) with fixed length \\(k\\), where \\(k \\ll L_s\\) (e.g., \\(k = 32\\) and \\(L_s \\geq 100\\))." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.632, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Q&A Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.146, + 0.788, + 0.208 + ], + "angle": 0, + "content": "Concretely, we additionally introduce the learnable embeddings \\(\\mathbf{F}_p\\in \\mathbb{R}^{k\\times d_q}\\) to interact with \\(\\mathbf{F}_m\\). As in Fig. 4, a cross-attention layer is applied by taking \\(\\mathbf{F}_p\\) as query, and \\(\\mathbf{F}_m\\) as key and value, followed with a feed-forward network, to generate prompt embeddings with the multimodal representations from \\(\\mathbf{F}_m\\):" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.217, + 0.636, + 0.232 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {p} = \\text {C r o s s - A t t e n t i o n} (\\mathbf {F} _ {p}, \\mathbf {F} _ {m}) + \\mathbf {F} _ {p}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.228, + 0.785, + 0.252 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {p} = \\operatorname {F F N} \\left(\\mathbf {F} _ {p}\\right) + \\mathbf {F} _ {p}, \\mathbf {F} _ {p} = \\operatorname {M L P} \\left(\\mathbf {F} _ {p}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.26, + 0.628, + 0.274 + ], + "angle": 0, + "content": "where MLP projects \\(d_{q}\\) to \\(d_{lm}\\) (the hidden size of LLM)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.275, + 0.787, + 0.336 + ], + "angle": 0, + "content": "At last, the final output of \\(\\mathbf{F}_p\\in \\mathbb{R}^{k\\times d_q}\\) will be fed into the LLM together with the question instruction embeddings and visual features as in Eq. (3), to make predictions of answers. The model is trained using the cross-entropy loss function with trainable parameters \\(\\theta\\):" + }, + { + "type": "equation", + "bbox": [ + 0.364, + 0.344, + 0.786, + 0.385 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {V Q A}} = - \\sum_ {t = 1} ^ {L _ {a}} \\log P _ {\\theta} \\left(\\mathcal {A} _ {t} \\mid \\mathcal {A} _ {< t}, \\mathcal {V}, \\mathcal {Q}, \\mathcal {S}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.392, + 0.787, + 0.423 + ], + "angle": 0, + "content": "where \\(\\mathcal{A}_t\\) is predicted autoregressively at position \\(t\\), and \\(L_{a}\\) is the sequence length of the ground truth answer text \\(\\mathcal{A}\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.443, + 0.377, + 0.46 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.473, + 0.329, + 0.487 + ], + "angle": 0, + "content": "4.1 Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.789, + 0.648 + ], + "angle": 0, + "content": "We evaluate Q&A Prompts on the challenging A-OKVQA [42] and OK-VQA [37] since they are currently the largest VQA datasets requiring complex reasoning over diverse world knowledge. Specifically, the A-OKVQA dataset consists of 24,903 samples, with 17.1k samples for training, 1.1k for validation, and 6.7k for testing. The questions in this dataset require reasoning over various types of world knowledge, such as commonsense knowledge, factoid knowledge, and visual knowledge. The OK-VQA dataset includes 14k questions covering a variety of knowledge categories, with 9k samples for training, and 5k for validation. Each sample in these two datasets includes an image, a question, and 10 ground-truth answers. We use the soft accuracy [3] as the standard evaluation metric." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.667, + 0.457, + 0.683 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.842 + ], + "angle": 0, + "content": "We use InstructBLIP [6] as our default MLLM, to both generate Q&A pairs and perform answer predictions. RAM [57] is utilized as the image tagging model due to its strong recognition ability. The number of Q&A pairs is set to \\( P = 8 \\). To configure the visual-aware prompting module, we use the text encoder in CLIP [40] as our prompt encoder, and a value of 32 of the \\( k \\) for the number of prompt embeddings \\( \\mathbf{F}_p \\in \\mathbb{R}^{k \\times d_q} \\), the same as the number of visual tokens \\( \\mathbf{F}_v \\in \\mathbb{R}^{k \\times d_q} \\). We train the models using AdamW [31] as the optimizer with a learning rate of \\( 2e^{-5} \\) and the strategy of mixed precision training. It is important to note that the parameters of the image encoder, the large language model, and the prompt encoder are kept frozen to maintain their pre-trained representations." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.442, + 0.129 + ], + "angle": 0, + "content": "Haibo Wang, Weifeng Ge" + }, + { + "type": "table", + "bbox": [ + 0.219, + 0.144, + 0.778, + 0.43 + ], + "angle": 0, + "content": "
MethodImage RepresentationKnowledge SourceOK-VQA (%)A-OKVQA (%)
ValTest
ClipCap [39]FeaturePretrain-18.115.8
Pythia [21]FeaturePretrain-25.221.9
ViLBERT [32]FeaturePretrain-30.625.9
LXMERT [47]FeaturePretrain-30.725.9
GPV-2 [23]FeaturePretrain-48.640.7
Unified-IO [33]FeaturePretrain54.0-45.2
Flamingo (80B) [2]FeaturePretrain57.8--
BLIP-2 [26]FeaturePretrain59.360.058.7
Muko [60]FeatureDBPedia + ConceptNet29.2--
ConceptBERT [11]FeatureNumberBatch + ConceptNet33.7--
KRISP [36]FeatureWikipedia + ConceptNet38.933.722.1
MAVEx [52]FeatureWikipedia + ConceptNet40.3--
UniER [16]FeatureConceptNet42.1--
TRiG [9]Caption + Tags + OCRWikipedia49.4--
REVEAL [18]FeatureWIT + Wikidata59.1-52.2
RA-VQA-v2 [28]FeatureGoogle Search62.1--
PICa [53]Caption + TagsFrozen GPT-3 (175B)48.0--
KAT [14]Caption + FeatureWikidata + GPT3 (175B)53.1--
Img2LLM [15]CaptionFrozen GPT-3 (175B)-42.940.7
REVIVE [29]Caption + FeatureWikipedia + GPT3 (175B)56.6--
PromptCap [17]CaptionFrozen GPT-3 (175B)60.456.359.6
Prophet [44]Caption + TagsFrozen GPT-3 (175B)61.158.261.1
InstructBLIP (7B) [6] + Q&A PromptsFeaturePretrain62.164.062.1
FeaturePretrain64.3 (+2.2)69.4 (+5.4)68.1 (+6.0)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.441, + 0.785, + 0.484 + ], + "angle": 0, + "content": "Table 1: Results of A-OKVQA [42] and OK-VQA [37] comparing to standard baselines show that our method achieves state-of-the-art performance, outperforming previous methods by a large margin. The best results are highlighted." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.513, + 0.373, + 0.527 + ], + "angle": 0, + "content": "4.3 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.517, + 0.735 + ], + "angle": 0, + "content": "Compariosn with SoTAs. The results in Tab. 1 demonstrate the superiority of our proposed Q&A Prompts. We have a significant improvement over existing SoTA methods, achieving an accuracy of \\(69.4\\%\\) on the validation set and \\(68.1\\%\\) on the test set in A-OKVQA, and an accuracy of \\(64.3\\%\\) in OK-VQA. We also observe that Q&A Prompts achieves a greater improvement on A-OKVQA \\((+5.4\\%)\\) and \\(+6.0\\%\\) compared to OKVQA \\((+2.2\\%)\\). This can be explained that: different from questions in OK-VQA" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "that heavily rely on external knowledge, the questions in A-OKVQA emphasize visually-grounded reasoning with a comprehensive understanding of the image with diverse visual clues, rather than external knowledge retrieval. This aligns with Q&A Prompts's philosophy, that is, instead of relying on accessing any external knowledge base such as Wikipedia [49], ConceptNet [46], or GPT3 [4], our focus is on enhancing reasoning capabilities by discovering rich visual clues hidden in the image and model." + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.568, + 0.773, + 0.66 + ], + "angle": 0, + "content": "
ModelsA-OKVQA (%)OKVQA (%)
MiniGPT-454.252.5
+ Q&A Prompts61.2 (+7.0)55.9 (+3.4)
LLaVA-1.566.963.3
+ Q&A Prompts72.4 (+5.5)65.2 (+1.9)
InstructBLIP (T5)58.555.1
+ Q&A Prompts64.2 (+5.7)57.3 (+2.2)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.526, + 0.67, + 0.788, + 0.698 + ], + "angle": 0, + "content": "Table 2: Comparison of performance with different MLLMs as baselines." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.633, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Q&A Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.145, + 0.359, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.213, + 0.347, + 0.242 + ], + "angle": 0, + "content": "Question: Which room in the house is this man in?" + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.145, + 0.501, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.369, + 0.213, + 0.484, + 0.244 + ], + "angle": 0, + "content": "Question: Why are the men's vests orange?" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.145, + 0.641, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.213, + 0.628, + 0.244 + ], + "angle": 0, + "content": "Question: What problem will the people on the ferry face?" + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.145, + 0.78, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.213, + 0.768, + 0.244 + ], + "angle": 0, + "content": "Question: Why are they holding hands in front of people?" + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.248, + 0.355, + 0.297 + ], + "angle": 0, + "content": "Image Tags: mirror, sink, reflection, tile wall, tie, bathroom accessory, stare, faucet, cabinet" + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.247, + 0.488, + 0.298 + ], + "angle": 0, + "content": "Image Tags: constructor, train, safety vest, train track, passenger, railroad, railway station, platform" + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.247, + 0.636, + 0.289 + ], + "angle": 0, + "content": "Image Tags: \ncloudy, travel, ferry, barge, sail, waterway, boat, vessel, float, water, city" + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.247, + 0.756, + 0.297 + ], + "angle": 0, + "content": "Image Tags: \nbible, veil, hand, couple \ngroom, ceremony, \nwedding, father, bride, \nnewlywed, church" + }, + { + "type": "title", + "bbox": [ + 0.226, + 0.303, + 0.301, + 0.313 + ], + "angle": 0, + "content": "Q&A Prompts:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.313, + 0.356, + 0.332 + ], + "angle": 0, + "content": "Q: What is the man looking at? A: mirror" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.333, + 0.351, + 0.354 + ], + "angle": 0, + "content": "Q: What is the man washing hands at? A: sink" + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.313, + 0.356, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.362, + 0.357, + 0.394 + ], + "angle": 0, + "content": "Q: Where are the bathroom accessory placed? A: cabinet" + }, + { + "type": "title", + "bbox": [ + 0.369, + 0.303, + 0.444, + 0.313 + ], + "angle": 0, + "content": "Q&A Prompts:" + }, + { + "type": "text", + "bbox": [ + 0.368, + 0.313, + 0.499, + 0.353 + ], + "angle": 0, + "content": "Q: What's the man in orange vest? A: constructor Q: What is the yellow and black vehicle? A: train" + }, + { + "type": "text", + "bbox": [ + 0.368, + 0.363, + 0.492, + 0.394 + ], + "angle": 0, + "content": "Q: What is the area where the train is located? A: platform" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.303, + 0.583, + 0.313 + ], + "angle": 0, + "content": "Q&A Prompts:" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.313, + 0.637, + 0.333 + ], + "angle": 0, + "content": "Q: What type of weather is it? A: cloudy" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.333, + 0.639, + 0.363 + ], + "angle": 0, + "content": "Q: What is the purpose of the orange and white boat? A: travel" + }, + { + "type": "list", + "bbox": [ + 0.509, + 0.313, + 0.639, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.37, + 0.629, + 0.394 + ], + "angle": 0, + "content": "Q: Where are these boats in? A: city" + }, + { + "type": "title", + "bbox": [ + 0.649, + 0.303, + 0.722, + 0.313 + ], + "angle": 0, + "content": "Q&A Prompts:" + }, + { + "type": "text", + "bbox": [ + 0.648, + 0.313, + 0.777, + 0.363 + ], + "angle": 0, + "content": "Q: What is the man in black suit holding? A: bible \nQ: What is the woman wearing on her head? \nA: veil" + }, + { + "type": "text", + "bbox": [ + 0.648, + 0.372, + 0.775, + 0.394 + ], + "angle": 0, + "content": "Q: What kind of building is this? A: church" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.401, + 0.338, + 0.461 + ], + "angle": 0, + "content": "Predictive Answer: \nInstructBLIP: bedroom \nBLIP-2: bedroom \nOurs: bathroom \nGround Truth: bathroom" + }, + { + "type": "text", + "bbox": [ + 0.368, + 0.4, + 0.471, + 0.462 + ], + "angle": 0, + "content": "Predictive Answer: \nInstructBLIP: fashion \nBLIP-2: working \nOurs: visibility \nGround Truth: visibility" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.401, + 0.617, + 0.462 + ], + "angle": 0, + "content": "Predictive Answer: \nInstructBLIP: sickness \nBLIP-2: fall into water \nOurs: raining \nGround Truth: \nraining" + }, + { + "type": "text", + "bbox": [ + 0.648, + 0.401, + 0.753, + 0.462 + ], + "angle": 0, + "content": "Predictive Answer: \nInstructBLIP: praying \nBLIP-2: praying \nOurs: getting married \nGround Truth: \ngetting married" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.475, + 0.789, + 0.517 + ], + "angle": 0, + "content": "Fig. 5: Representative cases with our Q&A Prompts. We denote the question, tags, Q&A prompts, predictive answers, and ground truth respectively. Our Q&A Prompts can discover rich visual clues in the image and reason out the correct answer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.547, + 0.8, + 0.638 + ], + "angle": 0, + "content": "Different baselines. To further validate the effectiveness of our method, we conduct experiments to compare various MLLMs as baseline models, including MiniGPT-4 [59], LLaVA-1.5 [30] and InstructBLIP (FLAN-T5-XL). Tab. 2 shows our Q&A Prompts method consistently improves performance on A-OKVQA by approximately \\(5.5\\%\\) to \\(7.0\\%\\) and OKVQA by \\(1.9\\%\\) to \\(3.4\\%\\), demonstrating the generalizability of Q&A Prompts across different vision-language models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.639, + 0.793, + 0.76 + ], + "angle": 0, + "content": "Qualitative results. We present qualitative results in Fig. 5. These cases demonstrate scenarios where correct answers are only possible with Q&A prompts. For instance, the important clues provided by Q&A Prompts of the [mirror], [constructor], [cloudy] and [bible], lead to the successful reasoning of correct answers [bathroom], [visibility], [raining], and [getting married] respectively. This showcases the valuable contribution of the Q&A prompts in certain types of tasks, which can explicitly mine rich visual clues and diverse world knowledge hidden in the image to help the model reason out correct answers." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.783, + 0.566, + 0.799 + ], + "angle": 0, + "content": "4.4 Analysis of the Generated Questions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "To systematically evaluate the questions, we leverage GPT-4V to measure their reasonableness. Specifically, we randomly sample 250 " + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.442, + 0.129 + ], + "angle": 0, + "content": "Haibo Wang, Weifeng Ge" + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.145, + 0.354, + 0.153 + ], + "angle": 0, + "content": "Visual Knowledge" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.155, + 0.311, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.22, + 0.208, + 0.297, + 0.233 + ], + "angle": 0, + "content": "Q: Where is the dog sitting? A: floor" + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.155, + 0.415, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.208, + 0.402, + 0.233 + ], + "angle": 0, + "content": "Q: What is in the middle of the plate? A: meat" + }, + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.145, + 0.575, + 0.153 + ], + "angle": 0, + "content": "Commonsien Knowledge" + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.154, + 0.517, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.427, + 0.207, + 0.5, + 0.232 + ], + "angle": 0, + "content": "Q: Why are people carrying the bags? A: load" + }, + { + "type": "image", + "bbox": [ + 0.531, + 0.154, + 0.621, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.531, + 0.207, + 0.619, + 0.232 + ], + "angle": 0, + "content": "Q: What is the glass covered area used for? A: showcase" + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.162, + 0.778, + 0.173 + ], + "angle": 0, + "content": "Required Knowledge Types" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.177, + 0.787, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.283, + 0.234, + 0.351, + 0.242 + ], + "angle": 0, + "content": "Knowledge base" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.244, + 0.31, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.22, + 0.297, + 0.3, + 0.323 + ], + "angle": 0, + "content": "Q: What event is this most likely? A: Christmas" + }, + { + "type": "image", + "bbox": [ + 0.324, + 0.244, + 0.415, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.297, + 0.397, + 0.323 + ], + "angle": 0, + "content": "Q: What species of bird is this? A: heron" + }, + { + "type": "image_caption", + "bbox": [ + 0.427, + 0.297, + 0.51, + 0.323 + ], + "angle": 0, + "content": "Q: How do the people feel here? A: cold" + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.297, + 0.619, + 0.323 + ], + "angle": 0, + "content": "Q: How is the woman's face visible? A: reflection" + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.333, + 0.705, + 0.348 + ], + "angle": 0, + "content": "Fig. 6: Generated Q&A pairs with different knowledge types." + }, + { + "type": "table", + "bbox": [ + 0.236, + 0.36, + 0.455, + 0.43 + ], + "angle": 0, + "content": "
Q&A sourceScore
Annotations in A-OKVQA9.73
VQG w/ training8.86
VQG w/o training4.12
" + }, + { + "type": "table_caption", + "bbox": [ + 0.22, + 0.44, + 0.468, + 0.467 + ], + "angle": 0, + "content": "Table 3: The evaluation for reasonableness of different questions." + }, + { + "type": "table", + "bbox": [ + 0.493, + 0.361, + 0.774, + 0.43 + ], + "angle": 0, + "content": "
Q&ATraining SetRouge-LA-OKVQA (%)OK-VQA (%)
×--64.062.1
✓--61.961.0
✓A-OKVQA90.869.462.9
✓OK-VQA92.467.164.3
✓Mixture89.468.763.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.48, + 0.44, + 0.785, + 0.467 + ], + "angle": 0, + "content": "Table 4: Ablation studies on training sets for VQG model training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.785, + 0.615 + ], + "angle": 0, + "content": "tuples from the QA pairs generated by our VQG model, and ask GPT-4V to return an integer score of reasonableness ranging from 1 to 10 for them. For comparison, we also evaluate the scores of the manually annotated tuples in the A-OKVQA and those generated by directly prompting MLLM without training (with the same number of 250 tuples). Tab. 3 shows the questions generated by our VQG model (VQG w/ training) have a close quality to the manually annotated questions in A-OKVQA and are more reasonable than directly prompting a frozen MLLM (VQG w/o training). See appendix for more details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.616, + 0.785, + 0.706 + ], + "angle": 0, + "content": "We also annotated what kind of world knowledge was required to answer the 250 sampled questions generated by our VQG model. The choices are the same with A-OKVQA including Commonsense Knowledge, Visual Knowledge, Knowledge Bases, and Physical Knowledge (explained in the appendix). The examples and distribution are shown in Fig. 6. Notably, sometimes there is no clear distinction between these categories and a question may belong to either." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.727, + 0.389, + 0.743 + ], + "angle": 0, + "content": "4.5 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.796, + 0.84 + ], + "angle": 0, + "content": "We investigate the role of our question-answer prompts and the visual-aware prompting module, based on InstructBLIP-Vicuna-7B and the A-OKVQA dataset. Different training sets for generated questions. We investigate how the VQG model will influence the performance by directly prompting the MLLM without training, or using different training sets to train our VQG model, including the A-OKVQA, OK-VQA, and a mixture of them. We report the Rouge-L" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.633, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Q&A Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.232, + 0.144, + 0.428, + 0.214 + ], + "angle": 0, + "content": "
Fusion strategyAccuracy (%)
Prepend + Q&A66.6
+ Captions64.2
+ Tags64.7
VPM + Q&A69.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.22, + 0.223, + 0.438, + 0.247 + ], + "angle": 0, + "content": "Table 5: Ablation studies on strategies of fusing Q&A pairs." + }, + { + "type": "table", + "bbox": [ + 0.459, + 0.144, + 0.58, + 0.214 + ], + "angle": 0, + "content": "
#NumAccuracy (%)
166.2
467.9
869.4
1669.2
" + }, + { + "type": "table", + "bbox": [ + 0.596, + 0.144, + 0.763, + 0.214 + ], + "angle": 0, + "content": "
VariantsAccuracy (%)
VPM69.4
w/ BERT68.6
w/o Fusion68.6
w/o Decoder68.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.445, + 0.223, + 0.774, + 0.258 + ], + "angle": 0, + "content": "Table 6: Ablation studies Table 7: Ablation studies on the on numbers of question-different components in the visual-answer pairs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.292, + 0.785, + 0.382 + ], + "angle": 0, + "content": "score for the generated questions. Results in Tab. 4 show that performance declines severely when using Q&A prompts generated by MLLMs without training. Besides, the best performance is only achieved when the corresponding dataset is utilized as the training set, and a mismatch or mixture of datasets could lead to a slightly weaker performance. This could be attributed to the domain gap in the question types of these datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.384, + 0.787, + 0.566 + ], + "angle": 0, + "content": "Strategies of fusing question-answer pairs. In Tab. 5, we explore a simple approach where we directly prepend the Q&A pairs as a natural language to the instruction (denoted as Prepend + Q&A) without encoded by the visual-aware prompting module (VPM). For comparison, we also replace the question-answer pairs with the same number of image captions generated by BLIP-2 [26] and tags recognized by RAM (denoted as Prepend + Captions and Prepend + Tags). The results in Tab. 5 show that the incorporation of question-answer prompts is more effective than pure captions and tags, and both directly prepending the question-answer pairs and using the VPM to encode them can improve performance. Notably, the models with VPM perform much better, which we attributed to its ability to combat some noise in the generated question-answer pairs and avoid interference with the target question." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.788, + 0.673 + ], + "angle": 0, + "content": "The number of the question-answer prompts. We also study to figure out the effect of using different numbers of question-answer pairs to construct the Q&A Prompts. The results are displayed in Tab. 6. We denote the desired number as \\( P \\). It can be observed that when \\( P \\) is set to 8, the model achieves optimal performance. We can deduce that when \\( P \\) is too large, it may introduce noisy pairs, while a too small \\( P \\) can result in an ignorance of many essential and informative question-answer pairs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Components in visual-aware prompting module. We validate the key components of the visual-aware prompting module in Tab. 7. We first study the effectiveness of using different language models as the prompt encoder, where we replace the text encoder in CLIP with BERT [8] (w/ BERT), which results in a performance drop. This result is consistent with previous findings [38, 45], which suggest that models initialized with CLIP [40] are more suitable for vision-language tasks due to their shared understanding of the visual and textual content. We then remove the visual gated fusion (w/o fusion), which also induces severe performance decline. This is expected since the unimodal language representations tend to involve ambiguity and bias. The additional incorporation of visual embeddings effectively allows for multi-modal interaction beneficial for" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.442, + 0.129 + ], + "angle": 0, + "content": "Haibo Wang, Weifeng Ge" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.402, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.408, + 0.147, + 0.475, + 0.157 + ], + "angle": 0, + "content": "Q&A Prompts:" + }, + { + "type": "text", + "bbox": [ + 0.406, + 0.158, + 0.654, + 0.233 + ], + "angle": 0, + "content": "Q1: What are the flowing vegetation? A: grass \nQ2: What are the effects of air movement on grass? A: bent \nQ3: What word describes the green grass? A: lush \nQ4: What animal in this image is used for riding? A: horse \nQ5: What is the red machine called? A: tractor \nQ6: Where is the horses standing? A: dirt field \nQ7: How many horses are there? A: couple \nQ8: What type of property is this? A: farm" + }, + { + "type": "text", + "bbox": [ + 0.66, + 0.147, + 0.781, + 0.187 + ], + "angle": 0, + "content": "Ground Truth: wind \nPredictive Answer: \nw/ Q&A Prompts: wind \nw/o Q&A Prompts: drought" + }, + { + "type": "text", + "bbox": [ + 0.66, + 0.191, + 0.764, + 0.232 + ], + "angle": 0, + "content": "w/Q&A Prompts (remove Q1, Q2): drought w/Q&A Prompts (remove Q3~Q8): wind" + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.247, + 0.691, + 0.261 + ], + "angle": 0, + "content": "Fig. 7: Qualitative analysis on how Q&A Prompts work." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.288, + 0.788, + 0.395 + ], + "angle": 0, + "content": "vision-language reasoning. We also evaluate the importance of the resampler decoder (w/o decoder), where we directly feed the long sequence \\(\\mathbf{F}_m\\in \\mathbb{R}^{L_s\\times d_q}\\) into the large language model in Eq. (3) and also caused a performance drop. The sequence length of target question \\(L_{q}\\) is \\(\\times 10\\) times smaller than the sequence length \\(L_{s}\\) of prompts \\(\\mathcal{S}\\) (\\(L_{q}\\leq 10\\), \\(L_{s}\\geq 100\\) here), resulting in a distraction to the understanding of the target question. This highlights the importance of resampling \\(\\mathbf{F}_m\\in \\mathbb{R}^{L_s\\times d_q}\\) into a shorter \\(\\mathbf{F}_p\\in \\mathbb{R}^{k\\times d_q}\\) (\\(k = 32\\) here)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.415, + 0.725, + 0.433 + ], + "angle": 0, + "content": "5 A Further Qualitative Analysis of Q&A Prompts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.788, + 0.596 + ], + "angle": 0, + "content": "Fig. 7 further reveals why Q&A Prompts are necessary and how Q&A Prompts work in particular case. To find out [what ambient effect] causes the appearance of grass, the key point is to grasp the movement of them. The Q&A pairs \\(Q_{1}\\) and \\(Q_{2}\\) precisely capture the visual clues that [the grass are flowing] and [air movement cause the grass bent], which are beneficial for reasoning out the answer [wind]. We also test scenarios where we remove \\(Q_{1}\\) and \\(Q_{2}\\) or only reserve them, leading to the incorrect answer [drought] and correct answer [wind] respectively. This indicates that the quality of Q&A Prompts is crucial, particularly when the question needs multiple reasoning steps beyond perception, and Q&A Prompts serve as the intermediate steps that fill the logical gap." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.615, + 0.527, + 0.631 + ], + "angle": 0, + "content": "6 Conclusions and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This paper has studied the problem of VQA tasks requiring reasoning over diverse world knowledge. We introduce a novel framework, Q&A Prompts, which effectively generates a set of question-answer prompts and encodes them with a visual-aware prompting module, significantly boosting the reasoning ability of current multi-modal large language models. We conducted a series of experiments on the A-OKVQA and OK-VQA benchmarks and achieved significant results compared with previous methods. Extensive ablations and comprehensive analyses have demonstrated the effectiveness and superiority of our method. Limitations. While our method shows promising results, it is essential to acknowledge the potential presence of biases in the data as observed in previous VQA studies [1,7], as well as the lack of the ability for fine-grained counting and Optical Character Recognition. In our future work, we plan to mitigate these biases and weaknesses to enhance the reasoning ability of current models further." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.632, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Q&A Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.222 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China under Grant Nos. 62106051 and the National Key R&D Program of China 2022YFC3601405." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.244, + 0.323, + 0.26 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.274, + 0.787, + 0.317 + ], + "angle": 0, + "content": "1. Agrawal, A., Batra, D., Parikh, D., Kembhavi, A.: Don't just assume; look and answer: Overcoming priors for visual question answering. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.787, + 0.358 + ], + "angle": 0, + "content": "2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: Adv. Neural Inform. Process. Syst. (2022) 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.359, + 0.786, + 0.385 + ], + "angle": 0, + "content": "3. Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Zitnick, C.L., Parikh, D.: Vqa: visual question answering. In: Int. Conf. Comput. Vis. (2015) 2, 3, 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.386, + 0.787, + 0.427 + ], + "angle": 0, + "content": "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: Adv. Neural Inform. Process. Syst. (2020) 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.428, + 0.787, + 0.468 + ], + "angle": 0, + "content": "5. Changpinyo, S., Sharma, P., Ding, N., Soricut, R.: Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.469, + 0.787, + 0.509 + ], + "angle": 0, + "content": "6. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500 (2023) 2, 3, 4, 5, 6, 8, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.51, + 0.787, + 0.551 + ], + "angle": 0, + "content": "7. Dancette, C., Cadene, R., Teney, D., Cord, M.: Beyond question-based biases: Assessing multimodal shortcut learning in visual question answering. In: Int. Conf. Comput. Vis. (2021) 2, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.552, + 0.787, + 0.579 + ], + "angle": 0, + "content": "8. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: NAACL (2019) 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.58, + 0.787, + 0.62 + ], + "angle": 0, + "content": "9. Gao, F., Ping, Q., Thattai, G., Reganti, A., Wu, Y.N., Natarajan, P.: Transform-retrieve-produce: Natural language-centric outside-knowledge visual question answering. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.621, + 0.787, + 0.661 + ], + "angle": 0, + "content": "0. Gao, L., Biderman, S., Black, S., Golding, L., Hoppe, T., Foster, C., Phang, J., He, H., Thite, A., Nabeshima, N., et al.: The pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.662, + 0.787, + 0.689 + ], + "angle": 0, + "content": "1. Gardères, F., Ziaeffard, M., Abeloos, B., Lecue, F.: ConceptBert: Concept-aware representation for visual question answering. In: EMNLP (2020) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.787, + 0.73 + ], + "angle": 0, + "content": "2. Geman, D., Geman, S., Hallonquist, N., Younes, L.: Visual Turing test for computer vision systems. Proceedings of the National Academy of Sciences 112(12), 3618-3623 (2015) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.787, + 0.771 + ], + "angle": 0, + "content": "3. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In: IEEE Conf. Comput. Vis. Pattern Recog. (2017) 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.787, + 0.799 + ], + "angle": 0, + "content": "4. Gui, L., Wang, B., Huang, Q., Hauptmann, A., Bisk, Y., Gao, J.: Kat: A knowledge augmented transformer for vision-and-language. In: NAACL (2022) 2, 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "5. Guo, J., Li, J., Li, D., Tiong, A.M.H., Li, B., Tao, D., Hoi, S.C.: From images to textual prompts: Zero-shot vqa with frozen large language models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) 4, 10" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.274, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.442, + 0.129 + ], + "angle": 0, + "content": "Haibo Wang, Weifeng Ge" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "16. Guo, Y., Nie, L., Wong, Y., Liu, Y., Cheng, Z., Kankanhalli, M.: A unified end-to-end retriever-reader framework for knowledge-based vqa. In: ACM Int. Conf. Multimedia (2022) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.228 + ], + "angle": 0, + "content": "17. Hu, Y., Hua, H., Yang, Z., Shi, W., Smith, N.A., Luo, J.: Promptcap: Prompt-guided task-aware image captioning. arXiv preprint arXiv:2211.09699 (2022) 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.228, + 0.785, + 0.284 + ], + "angle": 0, + "content": "18. Hu, Z., Iscen, A., Sun, C., Wang, Z., Chang, K.W., Sun, Y., Schmid, C., Ross, D.A., Fathi, A.: **Retrieval-augmented visual-language pre-training with multi-source multimodal knowledge memory. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) 2, 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.284, + 0.785, + 0.325 + ], + "angle": 0, + "content": "19. Huang, X., Zhang, Y., Ma, J., Tian, W., Feng, R., Zhang, Y., Li, Y., Guo, Y., Zhang, L.: Tag2text: Guiding vision-language model via image tagging. arXiv preprint arXiv:2303.05657 (2023) 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.325, + 0.785, + 0.366 + ], + "angle": 0, + "content": "20. Jaegle, A., Borgeaud, S., Alayrac, J.B., Doersch, C., Ionescu, C., Ding, D., Koppula, S., Zoran, D., Brock, A., Shelhamer, E., et al.: Perceiver io: A general architecture for structured inputs & outputs. In: Int. Conf. Learn. Represent. (2022) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.366, + 0.785, + 0.406 + ], + "angle": 0, + "content": "21. Jiang, Y., Natarajan, V., Chen, X., Rohrbach, M., Batra, D., Parikh, D.: Pythia v0. 1: The winning entry to the vqa challenge 2018. arxiv 2018. arXiv preprint arXiv:1807.09956 (2018) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.406, + 0.785, + 0.461 + ], + "angle": 0, + "content": "22. Johnson, J., Hariharan, B., Van Der Maaten, L., Fei-Fei, L., Lawrence Zitnick, C., Girshick, R.: Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2901-2910 (2017) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.461, + 0.785, + 0.502 + ], + "angle": 0, + "content": "23. Kamath, A., Clark, C., Gupta, T., Kolve, E., Hoiem, D., Kembhavi, A.: Webly supervised concept expansion for general purpose vision models. In: Eur. Conf. Comput. Vis. (2022) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.502, + 0.785, + 0.542 + ], + "angle": 0, + "content": "24. Khan, Z., BG, V.K., Schulter, S., Yu, X., Fu, Y., Chandraker, M.: Q: How to specialize large vision-language models to data-scarce vqa tasks? a: Self-train on unlabeled images! In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.542, + 0.785, + 0.581 + ], + "angle": 0, + "content": "25. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.582, + 0.785, + 0.624 + ], + "angle": 0, + "content": "26. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. In: Int. Conf. Mach. Learn. (2023) 2, 4, 5, 10, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.624, + 0.785, + 0.664 + ], + "angle": 0, + "content": "27. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: common objects in context. In: Eur. Conf. Comput. Vis. (2014) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.664, + 0.785, + 0.705 + ], + "angle": 0, + "content": "28. Lin, W., Chen, J., Mei, J., Coca, A., Byrne, B.: Fine-grained late-interaction multimodal retrieval for retrieval augmented visual question answering. In: Adv. Neural Inform. Process. Syst. (2023) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.705, + 0.785, + 0.746 + ], + "angle": 0, + "content": "29. Lin, Y., Xie, Y., Chen, D., Xu, Y., Zhu, C., Yuan, L.: Revive: Regional visual representation matters in knowledge-based visual question answering. In: Adv. Neural Inform. Process. Syst. (2022) 2, 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.746, + 0.785, + 0.772 + ], + "angle": 0, + "content": "30. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: Adv. Neural Inform. Process. Syst. (2023) 2, 3, 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.772, + 0.785, + 0.8 + ], + "angle": 0, + "content": "31. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: Int. Conf. Learn. Represent. (2019) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "32. Lu, J., Batra, D., Parikh, D., Lee, S.: Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In: Adv. Neural Inform. Process. Syst. (2019) 10" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.633, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Q&A Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "33. Lu, J., Clark, C., Zellers, R., Mottaghi, R., Kembhavi, A.: Unified-io: A unified model for vision, language, and multi-modal tasks. arXiv preprint arXiv:2206.08916 (2022) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.23 + ], + "angle": 0, + "content": "34. Luo, G., Zhou, Y., Ren, T., Chen, S., Sun, X., Ji, R.: Cheap and quick: Efficient vision-language instruction tuning for large language models. arXiv preprint arXiv:2305.15023 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.23, + 0.788, + 0.256 + ], + "angle": 0, + "content": "35. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.257, + 0.788, + 0.298 + ], + "angle": 0, + "content": "36. Marino, K., Chen, X., Parikh, D., Gupta, A., Rohrbach, M.: Krisp: Integrating implicit and symbolic knowledge for open-domain knowledge-based vqa. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.298, + 0.788, + 0.339 + ], + "angle": 0, + "content": "37. Marino, K., Rastegari, M., Farhadi, A., Mottaghi, R.: Ok-vqa: A visual question answering benchmark requiring external knowledge. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019) 2, 3, 4, 5, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.339, + 0.788, + 0.366 + ], + "angle": 0, + "content": "38. Merullo, J., Castricato, L., Eickhoff, C., Pavlick, E.: Linearly mapping from image to text space. In: Int. Conf. Learn. Represent. (2023) 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.366, + 0.788, + 0.393 + ], + "angle": 0, + "content": "39. Mokady, R., Hertz, A., Bermano, A.H.: Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734 (2021) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.393, + 0.788, + 0.434 + ], + "angle": 0, + "content": "40. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Int. Conf. Mach. Learn. (2021) 9, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.46 + ], + "angle": 0, + "content": "41. Reimers, N., Gurevych, I.: Sentence-bert: Sentence embeddings using siamese bert-networks. In: EMNLP (2019) 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.46, + 0.788, + 0.501 + ], + "angle": 0, + "content": "42. Schwenk, D., Khandelwal, A., Clark, C., Marino, K., Mottaghi, R.: A-okvqa: A benchmark for visual question answering using world knowledge. In: Eur. Conf. Comput. Vis. (2022) 2, 3, 4, 5, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.501, + 0.788, + 0.542 + ], + "angle": 0, + "content": "43. Selvaraju, R.R., Tendulkar, P., Parikh, D., Horvitz, E., Ribeiro, M.T., Nushi, B., Kamar, E.: Squinting at vqa models: Introspecting vqa models with sub-questions. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.542, + 0.788, + 0.583 + ], + "angle": 0, + "content": "44. Shao, Z., Yu, Z., Wang, M., Yu, J.: Prompting large language models with answer heuristics for knowledge-based visual question answering. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.624 + ], + "angle": 0, + "content": "45. Shen, S., Li, L.H., Tan, H., Bansal, M., Rohrbach, A., Chang, K.W., Yao, Z., Keutzer, K.: How much can clip benefit vision-and-language tasks? In: Int. Conf. Learn. Represent. (2022) 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.624, + 0.788, + 0.651 + ], + "angle": 0, + "content": "46. Speer, R., Chin, J., Havasi, C.: Conceptnet 5.5: An open multilingual graph of general knowledge. In: AAAI (2017) 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.651, + 0.788, + 0.678 + ], + "angle": 0, + "content": "47. Tan, H., Bansal, M.: Lxmert: Learning cross-modality encoder representations from transformers. In: EMNLP (2019) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.678, + 0.788, + 0.704 + ], + "angle": 0, + "content": "48. Uehara, K., Duan, N., Harada, T.: Learning to ask informative sub-questions for visual question answering. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.731 + ], + "angle": 0, + "content": "49. Vrandecic, D., Krötzsch, M.: Wikidata: a free collaborative knowledgebase. Communications of the ACM 57(10), 78-85 (2014) 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.731, + 0.788, + 0.772 + ], + "angle": 0, + "content": "50. Wang, P., Wu, Q., Shen, C., Dick, A., Van Den Hengel, A.: Fvqa: Fact-based visual question answering. IEEE transactions on pattern analysis and machine intelligence 40(10), 2413-2427 (2017) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.772, + 0.788, + 0.812 + ], + "angle": 0, + "content": "51. Wang, P., Wu, Q., Shen, C., Hengel, A.v.d., Dick, A.: Explicit knowledge-based reasoning for visual question answering. arXiv preprint arXiv:1511.02570 (2015) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.812, + 0.788, + 0.84 + ], + "angle": 0, + "content": "52. Wu, J., Lu, J., Sabharwal, A., Mottaghi, R.: Multi-modal answer validation for knowledge-based vqa. In: AAAI (2021) 10" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.442, + 0.129 + ], + "angle": 0, + "content": "Haibo Wang, Weifeng Ge" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "53. Yang, Z., Gan, Z., Wang, J., Hu, X., Lu, Y., Liu, Z., Wang, L.: An empirical study of gpt-3 for few-shot knowledge-based vqa. In: AAAI (2022) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.785, + 0.217 + ], + "angle": 0, + "content": "54. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "55. Zellers, R., Bisk, Y., Farhadi, A., Choi, Y.: From recognition to cognition: Visual commonsense reasoning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.286 + ], + "angle": 0, + "content": "56. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "57. Zhang, Y., Huang, X., Ma, J., Li, Z., Luo, Z., Xie, Y., Qin, Y., Luo, T., Li, Y., Liu, S., et al.: Recognize anything: A strong image tagging model. arXiv preprint arXiv:2306.03514 (2023) 3, 4, 5, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.368 + ], + "angle": 0, + "content": "58. Zhang, Z., Zhang, A., Li, M., Zhao, H., Karypis, G., Smola, A.: Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923 (2023) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.41 + ], + "angle": 0, + "content": "59. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023) 3, 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.439 + ], + "angle": 0, + "content": "60. Zhu, Z., Yu, J., Wang, Y., Sun, Y., Hu, Y., Wu, Q.: Mucko: multi-layer cross-modal knowledge reasoning for fact-based visual question answering. In: IJCAI (2020) 10" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.439 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Q&A Prompts_ Discovering Rich Visual Clues through Mining Question-Answer Prompts for VQA requiring Diverse World Knowledge/d6b849aa-8c9f-421c-b457-5802b3bd4aca_origin.pdf b/2024/Q&A Prompts_ Discovering Rich Visual Clues through Mining Question-Answer Prompts for VQA requiring Diverse World Knowledge/d6b849aa-8c9f-421c-b457-5802b3bd4aca_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..41e0deee3f157610561caf0dbb72d44e03560454 --- /dev/null +++ b/2024/Q&A Prompts_ Discovering Rich Visual Clues through Mining Question-Answer Prompts for VQA requiring Diverse World Knowledge/d6b849aa-8c9f-421c-b457-5802b3bd4aca_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fc0f2e6df756cee8e38e989b7a103f89acf05f6ac675ed636ef4b623c70f17cb +size 2914613 diff --git a/2024/Q&A Prompts_ Discovering Rich Visual Clues through Mining Question-Answer Prompts for VQA requiring Diverse World Knowledge/full.md b/2024/Q&A Prompts_ Discovering Rich Visual Clues through Mining Question-Answer Prompts for VQA requiring Diverse World Knowledge/full.md new file mode 100644 index 0000000000000000000000000000000000000000..26cdfc646597d44e8b6b080dd74b9f5db06d6a24 --- /dev/null +++ b/2024/Q&A Prompts_ Discovering Rich Visual Clues through Mining Question-Answer Prompts for VQA requiring Diverse World Knowledge/full.md @@ -0,0 +1,429 @@ +# Q&A Prompts: Discovering Rich Visual Clues through Mining Question-Answer Prompts for VQA requiring Diverse World Knowledge + +Haibo Wang $^{1,2}$ and Weifeng Ge $^{1,2\boxtimes}$ + +1 Nebula AI Group, School of Computer Science, Fudan University +2 Shanghai Key Lab of Intelligent Information Processing hbwang22@m.fudan.edu.cn, wfge@fudan.edu.cn + +Abstract. With the breakthrough of multi-modal large language models (MLLMs), answering complex visual questions that demand advanced reasoning abilities and world knowledge has become a much more important testbed for developing AI models than ever. However, equipping MLLMs with robust cross-modality reasoning ability remains challenging since the cognition scheme of humans has not been understood systematically. In this paper, we believe that if we can collect rich visual clues, we will recognize the image more accurately, understand the question better, recall relevant knowledge more easily, and finally reason out the answer. We discover these rich visual clues by mining question-answer pairs in images and sending them into multi-modal large language models as prompts. We call the proposed method Q&A Prompts. Specifically, we first use the image-answer pairs and the corresponding questions in the training set as inputs and outputs to train a visual question generation (VQG) model. Then, we use an image tagging model to identify various instances and send packaged image-tag pairs into the VQG model to generate relevant questions with the extracted image tags as answers. Finally, we encode these generated question-answer pairs as prompts with a visual-aware prompting module and send them into pretrained MLLMs to reason out the final answers. Experimental results show that, compared with state-of-the-art methods, our Q&A Prompts achieves substantial improvements on the challenging visual question answering datasets requiring reasoning over diverse world knowledge, such as OK-VQA and A-OKVQA. Codes will be available at link. + +Keywords: multi-modal large language model $\cdot$ visual question answering $\cdot$ visual language reasoning + +# 1 Introduction + +Visual Question Answering (VQA) has longstanding been considered a Visual Turing Test [12] in the machine learning community. To solve this Turing test, artificial intelligence (AI) models are expected to be equipped with the human capabilities of visual recognition, language understanding, logical reasoning, world + +![](images/378d7f50367610104950f29c2507ff8f41ccd40980f1e804df8a8295710bbea5.jpg) +VQA Tasks well-solved by MLLMs + +![](images/092e870ac39982361dc1d008fa0b49a7d4c9340f959388bfda9975accc98aa90.jpg) +VQA Tasks unsolved by MLLMs + +![](images/66ae73ade15069150a3b6540ba59cfc07d167dedc0194298fb820cba1a5a7e58.jpg) +Fig. 1: Illustration of VQA tasks that require simple perception abilities ((a) and (b)) and those demanding chains of reasoning over diverse world knowledge ((c) and (d)). The example on the right ((e) and (f)) indicates MLLMs with Q&A prompts can solve difficult VQA problems that require reasoning over diverse world knowledge. + +![](images/e5d657845a74283f272017c9a1f73e768e1da56eca27fb2f108f6faa50ad7b33.jpg) + +![](images/66bb53699443cc5b275aa8fc2e0bf0c4bacbba12d7c557a3b3269735f3e4d24f.jpg) +VQA Tasks solved by MLLMs with Q&A Prompts + +knowledge referring, etc. There are a variety of datasets focusing on different problems in VQA, such as the perception and language understanding problem [3, 13], reasoning on procedurally generated images [22], and structured factual knowledge [50, 51] or commonsense knowledge [55]. Recently, with the advent of multi-modal large language models (MLLMs) [6, 26, 30, 54], significant progress has been achieved in perception-based VQA tasks [3, 13]. Fig. 1 (a) and Fig. 1 (b) show such VQA tasks, which can be reliably solved by MLLMs. + +However, for more complex VQA tasks [37, 42] as in Fig. 1 (c) and Fig. 1 (d), that feature the necessity of diverse world knowledge and complex chains of reasoning, state-of-the-art methods [14, 18, 29] even MLLM models [6, 26] fail to give the correct answers. Different from perception-based VQA [3, 13] and "closed" knowledge-based VQA [50, 51], visual question reasoning in OK-VQA [37] and A-OKVQA [42] requires not only diverse forms of world knowledge but also need complex reasoning logic. In Fig. 1 (d), where a woman sits on the ground with her pet dogs and holds an umbrella, the question is [Why is the woman holding the umbrella?] Simply associating the keyword [umbrella] with the word [rain] due to their frequent co-occurrence in the pre-training corpus [7] will lead to the wrong answer. To reason out the correct answer, AI models need to consider rich visual clues presented in the image. For example, in Fig. 1 (f), having discovered the clues including [the woman is camping] and [the weather is shiny] from the Q&A pairs, the model can correctly fill the logic gap and arrive at the answer [for shade]. This evidence indicates the necessity of collecting rich visual clues to achieve good results. + +Based on the analysis above, we believe that collecting visual clues of various instances in images will help MLLM recognize the image more accurately, understand the question better, recall relevant knowledge more easily, and finally generate the reasoning results more intelligently. To investigate this hypothesis, we design a novel VQA framework called $Q\mathcal{E}A$ Prompts, which extracts nu + +merous question-answer pairs associated with different perspectives in images as prompts for MLLMs. We choose an instance in the given image as the answer and generate the related question to form a question-answer pair. We generate the question-answer pairs for almost every instance to get a bag of Q&A prompts. The reason we choose question-answer pairs as prompts comes from two folds: First, Q&A prompts can focus on diverse instances in images like objects, scenes, or actions, thereby offering multiple different perspectives to understand the given image and the target question; Second, the knowledge and reasoning insights hidden in these question-answer pairs can help to explicitly uncover more relevant world knowledge, which may be critical to filling the logic gap between perception and reasoning when answering the target question. + +In Q&A Prompts, there are three key stages, which are the visual question generation (VQG) model training stage, the question-answer prompts generation stage, and the visual-language reasoning stage. In the VQG model training stage, we gather question-answer pairs in a VQA training set to train a VQG model that can map an answer together with the given image to a correlated question. Subsequently, in the question-answer prompts generation stage, we leverage a pre-trained image tagging model [19, 57] to recognize various objects, scenes, and actions in images with text tags. We employ these tags as answers and send them into the VQG model along with the corresponding images to generate questions. Finally, in the visual-language reasoning stage, we encode the generated question-answer pairs with a visual-aware prompting module and send them together with image and question features into a frozen language model to reason out the answer. Since previous VQA benchmarks [3,13] have seen saturated performance, we conduct experiments on the more challenging OK-VQA [37] and A-OKVQA [42] datasets, which are newer and larger testbeds for MLLM to investigate their abilities in accomplishing reasoning tasks requiring diverse world knowledge. Experimental results show that our Q&A Prompts can substantially improve the reasoning ability of MLLMs, such as InstructBLIP [6], LLaVA [30], and MiniGPT-4 [59]. In summary, we make the following contributions: + +- We propose a novel VQA framework that effectively generates and leverages question-answer pairs as prompts to improve the reasoning ability of MLLMs. It provides a new perspective for the community to design VQA solutions by explicitly collecting rich visual clues to bridge the logic gap between perception and reasoning when answering the target question. +- We design a novel question-answer prompts generation scheme with a VQG model and an image tagging model, which can generate Q&A prompts for recognizable objects, scenes, and activities in images. Besides, we design a new visual-aware prompting module to encode these prompts efficiently for the subsequent reasoning. +- We test the method on the challenging OK-VQA [37] and A-OKVQA [42] benchmarks, which study reasoning over diverse forms of world knowledge. The proposed Q&A Prompts achieves an accuracy of $68.1\%$ and $64.3\%$ on A-OKVQA [42] and OK-VQA [37], outperforming previous state-of-the-art methods by clear margins. + +# 2 Related Work + +Multimodal Large Language Models (MLLMs). MLLMs [2,6,25,26,34,54, 56,59] have shown strong ability in image-language understanding and reasoning, by adapting frozen language models to frozen image encoders with trainable connection modules. For instance, Flamingo [2] incorporates visual features into the language model with gated cross-attention blocks. LLaVA [30], instead, directly projects visual features into the space of text embeddings using a linear layer. BLIP-2 and InstructBLIP [6,26] introduce a more complex Q-Former to bridge the modality gap. These models align images and language based on large-scale image-text pre-training [5,10,27] and conduct reasoning with language models. In this paper, we mine more numerous question-answer pairs and send them into the MLLM as prompts. These prompts provide rich visual clues and uncover relevant world knowledge, which is beneficial to reasoning out correct answers. + +VQA Requiring Reasoning over Diverse World Knowledge. Different from classical VQA tasks [3, 13, 50, 51], VQA requiring reasoning over world knowledge involves inferring intents, goals, the physics of the world, and the social dynamics of individuals. More recently, A-OKVQA [42] and OK-VQA [37], which our work focuses on, present large-scale challenging VQA datasets that encompass questions requiring reasoning based on diverse world knowledge, including visual knowledge, commonsense knowledge, and factoid knowledge. Previous methods like Img2LLM [15], Prophet [44], and PromptCap [17] exploit the rich hidden knowledge and strong reasoning abilities of LLMs (e.g., GPT3 [4]) to solve such problems. These methods convert images into natural languages and construct various prompt templates for LLMs with in-context learning [4] to generate answers. Another group of works, such as KAT [14], REVIVE [29], and REVEAL [18], directly encode questions and relevant knowledge retrieved from external knowledge bases [4, 46, 49] to train end-to-end models. Unlike these methods, we design a visual-question prompts generation scheme to mine the information in images and unambiguous words to reduce confusion when answering the target questions. With Q&A prompts, different types of world knowledge are incorporated into the reasoning process to help MLLMs infer the correct answers. + +Visual Question Generation (VQG). Various works have explored Visual Question Generation (VQG) in the context of VQA tasks. SQuINT [43] and SelTDA [24] focus on dataset augmentation to enhance the generalization of VQA models. SQuINT employed manual workers to annotate questions, while SelTDA generated questions automatically using unlabeled images. Another line of research, exemplified by [48], employs a VQG model to generate informative sub-questions. These sub-questions are then used alongside the original question during inference, enabling another VQA model to answer both simultaneously. Different from previous works, our VQG model uses the tags of images [19, 57] as answers for generated questions. This approach ensures that our VQG model generates more relevant questions with additional details and perspectives. Furthermore, we introduce a novel module to encode the question-answer pairs as prompts for MLLMs, instead of directly feeding them as the formation of natural language into the model [48]. + +![](images/a62552742be334efc773e4f01521fea1e129b587aadfed2ff84415e3404f26e9.jpg) +Fig. 2: (a) Firstly, we train a VQG model that can ask informative questions given the image and specified answer. (b) Then, we exploit the image tagging model to extract image tags as the answers to generated questions and obtain diverse question-answer prompts. (c) At last, we feed the image, question, and question-answer prompts into the VQA model to perform visual-language reasoning in (c). + +# 3 Method + +Q&A Prompts is a conceptually simple three-stage framework, and Fig. 2 gives an overview: a VQG model training stage, a question-answer prompts generation stage, and a visual-language reasoning stage. Specifically, in the VQG model training stage (Fig. 2 (a)), we train a vanilla MLLM with a broad range of image-answer pairs as input and the corresponding question as the target output to learn a mapping from answers to questions (detailed in Sec. 3.1). In the prompts generation stage (Fig. 2 (b)), we exploit an image tagging model [19, 57] to obtain relevant tags including attributes, objects, actions, and scenes in images. These tags serve as answers and will be fed into the VQG model trained in the first stage to generate diverse questions containing rich visual clues and world knowledge (detailed in Sec. 3.2). In the final reasoning stage (Fig. 2 (c)), we utilize our proposed visual-aware prompting module to encode these question-answer pairs into prompts, which will be fed into the MLLM such as BLIP-2 [26] or InstructBLIP [6] together with the image and the target question, to predict the answer (detailed in Sec. 3.3). The whole pipeline is flexible and effective in mining informative visual clues to correctly answer the given questions. + +# 3.1 Stage-1: Train the VQG Model + +Given an image $\mathcal{V}$ and answer $\mathcal{A}$ , the VQG model is responsible for posing an informative question $\mathcal{Q}$ which can be properly answered by $\mathcal{A}$ . To explore as many visual clues in the generated questions as possible, and avoid generating general questions containing limited clues (e.g., "What is this?"), we utilize the A-OKVQA [42] or OK-VQA [37] as the training set $D$ for our VQG model, since + +![](images/029793f67d72050a57a9e7b74347ae346d883a9305c90d77a4a6e6e3ab2959ea.jpg) +A-OKVQA What sport do the people have equipment for? +VQG Which object in this image used for winter sports? +Zero-shot Is this a snowboard? + +![](images/8492208510e5bde4c93c6360c46596cc3926ad6e6b6e3869093c7d23d2f25f6b.jpg) + +![](images/f470664c4fe5cd2a5670c7cb9e920a8f4d6ee942ef5a24b4a19784c02d43d2a0.jpg) +A-OKVQA What is the vehicle traveling on? +VQG What is this type of transportation called? +Zero-shot What is this? + +![](images/d877037b565550b9fa92e2376ea343376f2583f821eebf7faf88b68adb28a9ce.jpg) + snowboard + +![](images/97de4a57b3897445ccd8a0a038f0ddb8f0fec304f108cfbecb4f20cc8aaa7b60.jpg) + rail +Fig. 3: Visualizations and T-SNE embedding distributions of the questions in manually annotated A-OKVQA (green), and the questions generated by our VQG model (blue) or by directly prompting a frozen MLLM (orange). + +the questions in their annotations are diverse and specific, covering broader areas of knowledge. To create the VQG model that approximates $P(\mathcal{Q}|\mathcal{V},\mathcal{A})$ , we treat the problem of learning such a model as a text-generation problem and wish to train the MLLM (e.g., InstructBLIP [6]) as our VQG model with $D$ due to its strong image-language reasoning ability. + +Specifically, as shown in Fig. 2 (a), the VQG model comprises a frozen image encoder, a trainable connection module, and a frozen LLM. We extract image embeddings $\mathbf{E}_v\in \mathbb{R}^{n\times d_v}$ of $\nu$ with the image encoder ( $n$ is the patch number, $d_v$ is the embedding dimension), and then feed $\mathbf{E}_v$ into the connection module to obtain fixed-length visual features $\mathbf{F}_v\in \mathbb{R}^{k\times d_q}$ , which will be sent into the LLM together with the instruction. We write the instruction as the template containing the answer $\mathcal{A}$ , such as "Given the image, generate a question whose answer is: $\langle \mathcal{A}\rangle$ ." The model is trained using the cross-entropy loss with parameters $\theta$ : + +$$ +\mathcal {L} _ {\mathcal {V Q G}} = - \sum_ {t = 1} ^ {L _ {q}} \log P _ {\theta} \left(\mathcal {Q} _ {t} \mid \mathcal {Q} _ {< t}, \mathcal {V}, \mathcal {A}\right) \tag {1} +$$ + +where $\mathcal{Q}_t$ is predicted autoregressively at position $t$ , and $L_{q}$ is the sequence length of the question $\mathcal{Q}$ . We emphasize the necessity of training the VQG model, instead of prompting a frozen MLLM in a zero-shot manner. In Fig. 3, we use sentence-bert [41] to obtain the dense embedding vectors of each question and show the T-SNE [35] embedding of them, as well as visualization examples. We can observe that the questions generated by our VQG model (blue) are more diverse and specific like questions in A-OKVQA (green), while the questions generated by MLLM without training (orange) are often general and limited with low differentiation. Our following experiments will further verify this. + +# 3.2 Stage-2: Generate Question-Answer Prompts + +Once the VQG model has been obtained, the generation of question-answer pairs can proceed. In this stage, we will explicitly generate a wide range of question- + +![](images/3d4224a2be9d5450b8c6a0cc34ed314f6e817eb757ce5b15aa4ddd293b8aff71.jpg) +Fig. 4: Overview of the visual-aware prompting module, consisting of the prompt encoder, visual gated fusion, and resampler decoder. + +answer pairs to help the model have a deeper and broader understanding of the image. In order to generate diverse questions covering various perspectives of the image, we should ensure the answers fed into the VQG model are sufficient in varied forms. Therefore, we adopt a strong image tagging model [19, 57] to produce rich tags as answers. Different from object detectors that can only recognize limited object labels, image tagging models aim to identify various open-vocabulary elements in images, including attributes, objects, actions, scenes, etc, which can provide an appropriate source of the answers we need. + +We utilize the RAM [57] model as our image tagging model due to its superior recognition ability. As shown in Fig. 2 (b), we feed the image $\mathcal{V}$ into the tagging model and obtain various tags for each image. To make the tags of each image more distinguishable, we remove the most frequent ones containing general information, such as 'person', 'food', 'man', etc., resulting in an average of 14 tags per image. We denote this set of tags as $\mathcal{O} = \{o_i\}_{i=1}^M$ , where $M$ is the number of tags associated with an image. + +Using these tags as answers, we input them into the trained VQG model along with the image, as shown in Fig. 2 (b), to obtain corresponding questions that answer these tags. To ensure that the generated questions (denoted as $\mathcal{T} = \{t_i\}_{i=1}^M$ ) contribute effectively to answering the target question, we rank the generated questions $\mathcal{T} = \{t_i\}_{i=1}^M$ based on their similarity with the target question $\mathcal{Q}$ , and adopt the questions with the Top-P similarities as the final obtained question prompts $\mathcal{U}$ : + +$$ +\mathcal {U} = \left\{u _ {p} \right\} _ {p = 1} ^ {P} = \underset {t _ {i} \in \mathcal {T}} {\arg \operatorname {T o p - P}} \langle \mathrm {T} (t _ {i}), \mathrm {T} (\mathcal {Q}) \rangle \tag {2} +$$ + +where $\mathrm{T}(\cdot)$ represents the embedding model [41], which encodes sentences into embedding vectors for similarity computation. The $\langle \cdot, \cdot \rangle$ operator is the inner product, computing the cosine similarity between two sentences. We then concatenate each $u_i$ in $\mathcal{U}$ and its corresponding $o_i$ in $\mathcal{O}$ using the template "Question: $\langle u_i \rangle$ Answer: $\langle o_i \rangle$ ", resulting in the final question-answer pairs as prompts $\mathcal{S}$ . + +# 3.3 Stage-3: Reasoning with Q&A Prompts + +After obtaining these question-answer pairs involving rich visual clues and world knowledge, we use the proposed visual-aware prompting module, as shown in + +Fig. 4, to encode these question-answer pairs into a set of prompt embeddings $\mathbf{F}_p\in \mathbb{R}^{k\times d_q}$ for our VQA model. We still utilize the same MLLM (e.g., Instruct-BLIP [6]) as our VQA model for the final reasoning. As in Fig. 2 (c), the encoded prompt embeddings $\mathbf{F}_p$ will be fed into the frozen LLM in MLLM, together with the instruction embeddings and visual features $\mathbf{F}_v$ to predict the answer $\mathcal{A}$ : + +$$ +\mathcal {A} = \operatorname {L L M} \left(\left[ \mathbf {F} _ {v}; \mathbf {F} _ {p}; \operatorname {E m b e d} (\operatorname {I n s}) \right]\right) \tag {3} +$$ + +Embed is the word embedding layer in LLM, the operator $[\cdot]$ means concatenation, and Ins is the instruction including the target question $\mathcal{Q}$ , represented as "Question: $\langle \mathcal{Q} \rangle$ ? Answer:". We then introduce how the visual-aware prompting module bootstrapped the reasoning ability of MLLM within Q&A Prompts. + +Prompt Encoder. The prompt encoder is responsible for obtaining the semantics representations of the prompts $S$ . Specifically, we choose encoder-based language models as our prompt encoder due to their strong ability in language understanding. To preserve the well-trained representations, we keep the parameters of the prompt encoder frozen. We feed the prompts $S$ into the prompt encoder and get the encoded $\mathbf{F}_s \in \mathbb{R}^{L_s \times d_q}$ , where $L_s$ is the sequence length of question-answer prompts and $d_q$ is the hidden size of the prompt encoder. + +Visual Gated Fusion. Since the output $\mathbf{F}_s$ from the prompt encoder is an unimodal language representation, to fully utilize the explicit information in images, the visual gated fusion adaptively incorporates image information into this representation with the visual embeddings $\mathbf{E}_v \in \mathbb{R}^{n \times d_v}$ . Specifically, in Fig. 4, we use a single-head attention mechanism, similar to MM-COT [58], to correlate the language representation with the visual embeddings. We define the query, key and value as $\mathbf{Q} = \mathbf{F}_s \in \mathbb{R}^{L_s \times d_q}$ , $\mathbf{K} = \mathbf{V} = \mathrm{MLP}(\mathbf{E}_v) \in \mathbb{R}^{n \times d_q}$ , where MLP here is projecting $d_v$ to $d_q$ . The attention output $\mathbf{F}_v^{attn}$ is defined as: + +$$ +\mathbf {F} _ {v} ^ {\text {a t t n}} = \mathbf {Q} + \operatorname {S o f t m a x} \left(\frac {\mathbf {Q} \mathbf {K} ^ {T}}{\sqrt {d _ {q}}}\right) \mathbf {V}, \quad \mathbf {F} _ {v} ^ {\text {a t t n}} \in \mathbb {R} ^ {L _ {s} \times d _ {q}} \tag {4} +$$ + +Next, we apply the gated fusion mechanism to merge $\mathbf{F}_s$ and $\mathbf{F}_v^{attn}$ . The fused representation $\mathbf{F}_m \in \mathbb{R}^{L_s \times d_q}$ is obtained using the following equations: + +$$ +\begin{array}{l} \lambda = \operatorname {S i g m o i d} \left(\mathbf {F} _ {s} \mathbf {W} _ {s} + \mathbf {F} _ {v} ^ {\text {a t t n}} \mathbf {W} _ {v}\right) \\ \mathbf {F} _ {m} = (1 - \lambda) \mathbf {F} _ {s} + \lambda \mathbf {F} _ {v} ^ {\text {a t t n}} \tag {5} \\ \end{array} +$$ + +in which $\mathbf{W}_s$ and $\mathbf{W}_v$ are learnable parameters. The gated fusion allows the model to adaptively blend the information from the language representation and visual embeddings, resulting in the multimodal representations $\mathbf{F}_m$ . + +Resampler Decoder. Although $\mathbf{F}_m \in \mathbb{R}^{L_s \times d_q}$ in Eq. (5) is the multimodal representation of $\mathcal{S}$ , we cannot directly concatenate it with $\mathbf{F}_v$ and instruction embeddings as in Eq. (3), due to the long sequence length of $L_s$ , which can result in extra computational burden and distraction to the target question $\mathcal{Q}$ . In light of this, we employ a lightweight resampler decoder inspired by Perceiver [20], to map $\mathbf{F}_m \in \mathbb{R}^{L_s \times d_q}$ into smaller-sized prompt embeddings $\mathbf{F}_p \in \mathbb{R}^{k \times d_q}$ with fixed length $k$ , where $k \ll L_s$ (e.g., $k = 32$ and $L_s \geq 100$ ). + +Concretely, we additionally introduce the learnable embeddings $\mathbf{F}_p\in \mathbb{R}^{k\times d_q}$ to interact with $\mathbf{F}_m$ . As in Fig. 4, a cross-attention layer is applied by taking $\mathbf{F}_p$ as query, and $\mathbf{F}_m$ as key and value, followed with a feed-forward network, to generate prompt embeddings with the multimodal representations from $\mathbf{F}_m$ : + +$$ +\mathbf {F} _ {p} = \text {C r o s s - A t t e n t i o n} (\mathbf {F} _ {p}, \mathbf {F} _ {m}) + \mathbf {F} _ {p} +$$ + +$$ +\mathbf {F} _ {p} = \operatorname {F F N} \left(\mathbf {F} _ {p}\right) + \mathbf {F} _ {p}, \mathbf {F} _ {p} = \operatorname {M L P} \left(\mathbf {F} _ {p}\right) \tag {6} +$$ + +where MLP projects $d_{q}$ to $d_{lm}$ (the hidden size of LLM). + +At last, the final output of $\mathbf{F}_p\in \mathbb{R}^{k\times d_q}$ will be fed into the LLM together with the question instruction embeddings and visual features as in Eq. (3), to make predictions of answers. The model is trained using the cross-entropy loss function with trainable parameters $\theta$ : + +$$ +\mathcal {L} _ {\mathcal {V Q A}} = - \sum_ {t = 1} ^ {L _ {a}} \log P _ {\theta} \left(\mathcal {A} _ {t} \mid \mathcal {A} _ {< t}, \mathcal {V}, \mathcal {Q}, \mathcal {S}\right) \tag {7} +$$ + +where $\mathcal{A}_t$ is predicted autoregressively at position $t$ , and $L_{a}$ is the sequence length of the ground truth answer text $\mathcal{A}$ . + +# 4 Experiments + +# 4.1 Dataset + +We evaluate Q&A Prompts on the challenging A-OKVQA [42] and OK-VQA [37] since they are currently the largest VQA datasets requiring complex reasoning over diverse world knowledge. Specifically, the A-OKVQA dataset consists of 24,903 samples, with 17.1k samples for training, 1.1k for validation, and 6.7k for testing. The questions in this dataset require reasoning over various types of world knowledge, such as commonsense knowledge, factoid knowledge, and visual knowledge. The OK-VQA dataset includes 14k questions covering a variety of knowledge categories, with 9k samples for training, and 5k for validation. Each sample in these two datasets includes an image, a question, and 10 ground-truth answers. We use the soft accuracy [3] as the standard evaluation metric. + +# 4.2 Implementation Details + +We use InstructBLIP [6] as our default MLLM, to both generate Q&A pairs and perform answer predictions. RAM [57] is utilized as the image tagging model due to its strong recognition ability. The number of Q&A pairs is set to $P = 8$ . To configure the visual-aware prompting module, we use the text encoder in CLIP [40] as our prompt encoder, and a value of 32 of the $k$ for the number of prompt embeddings $\mathbf{F}_p \in \mathbb{R}^{k \times d_q}$ , the same as the number of visual tokens $\mathbf{F}_v \in \mathbb{R}^{k \times d_q}$ . We train the models using AdamW [31] as the optimizer with a learning rate of $2e^{-5}$ and the strategy of mixed precision training. It is important to note that the parameters of the image encoder, the large language model, and the prompt encoder are kept frozen to maintain their pre-trained representations. + +
MethodImage RepresentationKnowledge SourceOK-VQA (%)A-OKVQA (%)
ValTest
ClipCap [39]FeaturePretrain-18.115.8
Pythia [21]FeaturePretrain-25.221.9
ViLBERT [32]FeaturePretrain-30.625.9
LXMERT [47]FeaturePretrain-30.725.9
GPV-2 [23]FeaturePretrain-48.640.7
Unified-IO [33]FeaturePretrain54.0-45.2
Flamingo (80B) [2]FeaturePretrain57.8--
BLIP-2 [26]FeaturePretrain59.360.058.7
Muko [60]FeatureDBPedia + ConceptNet29.2--
ConceptBERT [11]FeatureNumberBatch + ConceptNet33.7--
KRISP [36]FeatureWikipedia + ConceptNet38.933.722.1
MAVEx [52]FeatureWikipedia + ConceptNet40.3--
UniER [16]FeatureConceptNet42.1--
TRiG [9]Caption + Tags + OCRWikipedia49.4--
REVEAL [18]FeatureWIT + Wikidata59.1-52.2
RA-VQA-v2 [28]FeatureGoogle Search62.1--
PICa [53]Caption + TagsFrozen GPT-3 (175B)48.0--
KAT [14]Caption + FeatureWikidata + GPT3 (175B)53.1--
Img2LLM [15]CaptionFrozen GPT-3 (175B)-42.940.7
REVIVE [29]Caption + FeatureWikipedia + GPT3 (175B)56.6--
PromptCap [17]CaptionFrozen GPT-3 (175B)60.456.359.6
Prophet [44]Caption + TagsFrozen GPT-3 (175B)61.158.261.1
InstructBLIP (7B) [6] + Q&A PromptsFeaturePretrain62.164.062.1
FeaturePretrain64.3 (+2.2)69.4 (+5.4)68.1 (+6.0)
+ +# 4.3 Main Results + +Compariosn with SoTAs. The results in Tab. 1 demonstrate the superiority of our proposed Q&A Prompts. We have a significant improvement over existing SoTA methods, achieving an accuracy of $69.4\%$ on the validation set and $68.1\%$ on the test set in A-OKVQA, and an accuracy of $64.3\%$ in OK-VQA. We also observe that Q&A Prompts achieves a greater improvement on A-OKVQA $(+5.4\%)$ and $+6.0\%$ compared to OKVQA $(+2.2\%)$ . This can be explained that: different from questions in OK-VQA + +that heavily rely on external knowledge, the questions in A-OKVQA emphasize visually-grounded reasoning with a comprehensive understanding of the image with diverse visual clues, rather than external knowledge retrieval. This aligns with Q&A Prompts's philosophy, that is, instead of relying on accessing any external knowledge base such as Wikipedia [49], ConceptNet [46], or GPT3 [4], our focus is on enhancing reasoning capabilities by discovering rich visual clues hidden in the image and model. + +Table 1: Results of A-OKVQA [42] and OK-VQA [37] comparing to standard baselines show that our method achieves state-of-the-art performance, outperforming previous methods by a large margin. The best results are highlighted. + +
ModelsA-OKVQA (%)OKVQA (%)
MiniGPT-454.252.5
+ Q&A Prompts61.2 (+7.0)55.9 (+3.4)
LLaVA-1.566.963.3
+ Q&A Prompts72.4 (+5.5)65.2 (+1.9)
InstructBLIP (T5)58.555.1
+ Q&A Prompts64.2 (+5.7)57.3 (+2.2)
+ +Table 2: Comparison of performance with different MLLMs as baselines. + +![](images/907393e7f0ff67c8c5b3a99e94ba24a7a4c13187fe4152a63eba77bac49448ac.jpg) +Question: Which room in the house is this man in? +Fig. 5: Representative cases with our Q&A Prompts. We denote the question, tags, Q&A prompts, predictive answers, and ground truth respectively. Our Q&A Prompts can discover rich visual clues in the image and reason out the correct answer. + +![](images/ca3702426d46d23bd5aec70da7912dfabff42387c528a2d5005bd84e3e7822ad.jpg) +Question: Why are the men's vests orange? +Image Tags: mirror, sink, reflection, tile wall, tie, bathroom accessory, stare, faucet, cabinet +Image Tags: constructor, train, safety vest, train track, passenger, railroad, railway station, platform + +![](images/bf5b532ef57dd2db96cd18d3cf7ba95d260b5f0dd3f0bdae332145382b00395e.jpg) +Question: What problem will the people on the ferry face? +Image Tags: +cloudy, travel, ferry, barge, sail, waterway, boat, vessel, float, water, city + +![](images/f0fe938da132948b3fbb05fc7e21b4fd670a9098fcd5e45fc6e8efa7ecbd534c.jpg) +Question: Why are they holding hands in front of people? +Image Tags: +bible, veil, hand, couple +groom, ceremony, +wedding, father, bride, +newlywed, church + +# Q&A Prompts: + +Q: What is the man looking at? A: mirror +Q: What is the man washing hands at? A: sink + +Q: Where are the bathroom accessory placed? A: cabinet + +# Q&A Prompts: + +Q: What's the man in orange vest? A: constructor Q: What is the yellow and black vehicle? A: train + +Q: What is the area where the train is located? A: platform + +# Q&A Prompts: + +Q: What type of weather is it? A: cloudy +Q: What is the purpose of the orange and white boat? A: travel + +Q: Where are these boats in? A: city + +# Q&A Prompts: + +Q: What is the man in black suit holding? A: bible +Q: What is the woman wearing on her head? +A: veil + +Q: What kind of building is this? A: church + +Predictive Answer: +InstructBLIP: bedroom +BLIP-2: bedroom +Ours: bathroom +Ground Truth: bathroom + +Predictive Answer: +InstructBLIP: fashion +BLIP-2: working +Ours: visibility +Ground Truth: visibility + +Predictive Answer: +InstructBLIP: sickness +BLIP-2: fall into water +Ours: raining +Ground Truth: +raining + +Predictive Answer: +InstructBLIP: praying +BLIP-2: praying +Ours: getting married +Ground Truth: +getting married + +Different baselines. To further validate the effectiveness of our method, we conduct experiments to compare various MLLMs as baseline models, including MiniGPT-4 [59], LLaVA-1.5 [30] and InstructBLIP (FLAN-T5-XL). Tab. 2 shows our Q&A Prompts method consistently improves performance on A-OKVQA by approximately $5.5\%$ to $7.0\%$ and OKVQA by $1.9\%$ to $3.4\%$ , demonstrating the generalizability of Q&A Prompts across different vision-language models. + +Qualitative results. We present qualitative results in Fig. 5. These cases demonstrate scenarios where correct answers are only possible with Q&A prompts. For instance, the important clues provided by Q&A Prompts of the [mirror], [constructor], [cloudy] and [bible], lead to the successful reasoning of correct answers [bathroom], [visibility], [raining], and [getting married] respectively. This showcases the valuable contribution of the Q&A prompts in certain types of tasks, which can explicitly mine rich visual clues and diverse world knowledge hidden in the image to help the model reason out correct answers. + +# 4.4 Analysis of the Generated Questions + +To systematically evaluate the questions, we leverage GPT-4V to measure their reasonableness. Specifically, we randomly sample 250 + +![](images/86335532cabda9b9d1bc74978322bb9e4073ee5fd2a8dcf6a9fbf4ec13c2ad1d.jpg) +Visual Knowledge +Q: Where is the dog sitting? A: floor + +![](images/542adeaf0e3fa05ccff84ac10e717e8dc0ccc5952fa8c029f670c137e0e71cc5.jpg) +Q: What is in the middle of the plate? A: meat + +![](images/8ee66fab8354b9c061e66c491d3dcf1e1b714bc25e1e527f0e13e19009570023.jpg) +Commonsien Knowledge +Q: Why are people carrying the bags? A: load + +![](images/28ffe4210dc656a9e678ac8ecd2ae6e58fc6da7dc5a0cee7e23580514e801dc6.jpg) +Q: What is the glass covered area used for? A: showcase + +![](images/4d7b47f7bf1189f59ffbe1fb7a53cc614609adcd940f1a3c4aaaff8c49248a0a.jpg) +Required Knowledge Types +Q: How is the woman's face visible? A: reflection + +![](images/36524fbc4f8abf4b489bfced18244225ea842e999bdc78d6350d504377dac9ea.jpg) +Knowledge base +Q: What event is this most likely? A: Christmas +Fig. 6: Generated Q&A pairs with different knowledge types. + +![](images/1e7dc1283518917169742cf8022fe2d961071456170e2ac7c29a6e1e8a89bb29.jpg) +Q: What species of bird is this? A: heron +Q: How do the people feel here? A: cold + +
Q&A sourceScore
Annotations in A-OKVQA9.73
VQG w/ training8.86
VQG w/o training4.12
+ +Table 3: The evaluation for reasonableness of different questions. + +
Q&ATraining SetRouge-LA-OKVQA (%)OK-VQA (%)
×--64.062.1
✓--61.961.0
✓A-OKVQA90.869.462.9
✓OK-VQA92.467.164.3
✓Mixture89.468.763.9
+ +Table 4: Ablation studies on training sets for VQG model training. + +tuples from the QA pairs generated by our VQG model, and ask GPT-4V to return an integer score of reasonableness ranging from 1 to 10 for them. For comparison, we also evaluate the scores of the manually annotated tuples in the A-OKVQA and those generated by directly prompting MLLM without training (with the same number of 250 tuples). Tab. 3 shows the questions generated by our VQG model (VQG w/ training) have a close quality to the manually annotated questions in A-OKVQA and are more reasonable than directly prompting a frozen MLLM (VQG w/o training). See appendix for more details. + +We also annotated what kind of world knowledge was required to answer the 250 sampled questions generated by our VQG model. The choices are the same with A-OKVQA including Commonsense Knowledge, Visual Knowledge, Knowledge Bases, and Physical Knowledge (explained in the appendix). The examples and distribution are shown in Fig. 6. Notably, sometimes there is no clear distinction between these categories and a question may belong to either. + +# 4.5 Ablation Study + +We investigate the role of our question-answer prompts and the visual-aware prompting module, based on InstructBLIP-Vicuna-7B and the A-OKVQA dataset. Different training sets for generated questions. We investigate how the VQG model will influence the performance by directly prompting the MLLM without training, or using different training sets to train our VQG model, including the A-OKVQA, OK-VQA, and a mixture of them. We report the Rouge-L + +
Fusion strategyAccuracy (%)
Prepend + Q&A66.6
+ Captions64.2
+ Tags64.7
VPM + Q&A69.4
+ +Table 5: Ablation studies on strategies of fusing Q&A pairs. + +
#NumAccuracy (%)
166.2
467.9
869.4
1669.2
+ +
VariantsAccuracy (%)
VPM69.4
w/ BERT68.6
w/o Fusion68.6
w/o Decoder68.1
+ +Table 6: Ablation studies Table 7: Ablation studies on the on numbers of question-different components in the visual-answer pairs. + +score for the generated questions. Results in Tab. 4 show that performance declines severely when using Q&A prompts generated by MLLMs without training. Besides, the best performance is only achieved when the corresponding dataset is utilized as the training set, and a mismatch or mixture of datasets could lead to a slightly weaker performance. This could be attributed to the domain gap in the question types of these datasets. + +Strategies of fusing question-answer pairs. In Tab. 5, we explore a simple approach where we directly prepend the Q&A pairs as a natural language to the instruction (denoted as Prepend + Q&A) without encoded by the visual-aware prompting module (VPM). For comparison, we also replace the question-answer pairs with the same number of image captions generated by BLIP-2 [26] and tags recognized by RAM (denoted as Prepend + Captions and Prepend + Tags). The results in Tab. 5 show that the incorporation of question-answer prompts is more effective than pure captions and tags, and both directly prepending the question-answer pairs and using the VPM to encode them can improve performance. Notably, the models with VPM perform much better, which we attributed to its ability to combat some noise in the generated question-answer pairs and avoid interference with the target question. + +The number of the question-answer prompts. We also study to figure out the effect of using different numbers of question-answer pairs to construct the Q&A Prompts. The results are displayed in Tab. 6. We denote the desired number as $P$ . It can be observed that when $P$ is set to 8, the model achieves optimal performance. We can deduce that when $P$ is too large, it may introduce noisy pairs, while a too small $P$ can result in an ignorance of many essential and informative question-answer pairs. + +Components in visual-aware prompting module. We validate the key components of the visual-aware prompting module in Tab. 7. We first study the effectiveness of using different language models as the prompt encoder, where we replace the text encoder in CLIP with BERT [8] (w/ BERT), which results in a performance drop. This result is consistent with previous findings [38, 45], which suggest that models initialized with CLIP [40] are more suitable for vision-language tasks due to their shared understanding of the visual and textual content. We then remove the visual gated fusion (w/o fusion), which also induces severe performance decline. This is expected since the unimodal language representations tend to involve ambiguity and bias. The additional incorporation of visual embeddings effectively allows for multi-modal interaction beneficial for + +![](images/889498b47b68ec8e0249ae1a0cb2e73fb7b21b131e44ae8231d5dafae4f4cdfc.jpg) +Fig. 7: Qualitative analysis on how Q&A Prompts work. + +# Q&A Prompts: + +Q1: What are the flowing vegetation? A: grass +Q2: What are the effects of air movement on grass? A: bent +Q3: What word describes the green grass? A: lush +Q4: What animal in this image is used for riding? A: horse +Q5: What is the red machine called? A: tractor +Q6: Where is the horses standing? A: dirt field +Q7: How many horses are there? A: couple +Q8: What type of property is this? A: farm + +Ground Truth: wind +Predictive Answer: +w/ Q&A Prompts: wind +w/o Q&A Prompts: drought + +w/Q&A Prompts (remove Q1, Q2): drought w/Q&A Prompts (remove Q3~Q8): wind + +vision-language reasoning. We also evaluate the importance of the resampler decoder (w/o decoder), where we directly feed the long sequence $\mathbf{F}_m\in \mathbb{R}^{L_s\times d_q}$ into the large language model in Eq. (3) and also caused a performance drop. The sequence length of target question $L_{q}$ is $\times 10$ times smaller than the sequence length $L_{s}$ of prompts $\mathcal{S}$ ( $L_{q}\leq 10$ , $L_{s}\geq 100$ here), resulting in a distraction to the understanding of the target question. This highlights the importance of resampling $\mathbf{F}_m\in \mathbb{R}^{L_s\times d_q}$ into a shorter $\mathbf{F}_p\in \mathbb{R}^{k\times d_q}$ ( $k = 32$ here). + +# 5 A Further Qualitative Analysis of Q&A Prompts + +Fig. 7 further reveals why Q&A Prompts are necessary and how Q&A Prompts work in particular case. To find out [what ambient effect] causes the appearance of grass, the key point is to grasp the movement of them. The Q&A pairs $Q_{1}$ and $Q_{2}$ precisely capture the visual clues that [the grass are flowing] and [air movement cause the grass bent], which are beneficial for reasoning out the answer [wind]. We also test scenarios where we remove $Q_{1}$ and $Q_{2}$ or only reserve them, leading to the incorrect answer [drought] and correct answer [wind] respectively. This indicates that the quality of Q&A Prompts is crucial, particularly when the question needs multiple reasoning steps beyond perception, and Q&A Prompts serve as the intermediate steps that fill the logical gap. + +# 6 Conclusions and Limitations + +This paper has studied the problem of VQA tasks requiring reasoning over diverse world knowledge. We introduce a novel framework, Q&A Prompts, which effectively generates a set of question-answer prompts and encodes them with a visual-aware prompting module, significantly boosting the reasoning ability of current multi-modal large language models. We conducted a series of experiments on the A-OKVQA and OK-VQA benchmarks and achieved significant results compared with previous methods. Extensive ablations and comprehensive analyses have demonstrated the effectiveness and superiority of our method. Limitations. While our method shows promising results, it is essential to acknowledge the potential presence of biases in the data as observed in previous VQA studies [1,7], as well as the lack of the ability for fine-grained counting and Optical Character Recognition. In our future work, we plan to mitigate these biases and weaknesses to enhance the reasoning ability of current models further. + +# Acknowledgements + +This work was supported by the National Natural Science Foundation of China under Grant Nos. 62106051 and the National Key R&D Program of China 2022YFC3601405. + +# References + +1. Agrawal, A., Batra, D., Parikh, D., Kembhavi, A.: Don't just assume; look and answer: Overcoming priors for visual question answering. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018) 14 +2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: Adv. Neural Inform. Process. Syst. (2022) 4, 10 +3. Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Zitnick, C.L., Parikh, D.: Vqa: visual question answering. In: Int. Conf. Comput. Vis. (2015) 2, 3, 4, 9 +4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: Adv. Neural Inform. Process. Syst. (2020) 4, 10 +5. Changpinyo, S., Sharma, P., Ding, N., Soricut, R.: Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021) 4 +6. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500 (2023) 2, 3, 4, 5, 6, 8, 9, 10 +7. Dancette, C., Cadene, R., Teney, D., Cord, M.: Beyond question-based biases: Assessing multimodal shortcut learning in visual question answering. In: Int. Conf. Comput. Vis. (2021) 2, 14 +8. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: NAACL (2019) 13 +9. Gao, F., Ping, Q., Thattai, G., Reganti, A., Wu, Y.N., Natarajan, P.: Transform-retrieve-produce: Natural language-centric outside-knowledge visual question answering. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) 10 +0. Gao, L., Biderman, S., Black, S., Golding, L., Hoppe, T., Foster, C., Phang, J., He, H., Thite, A., Nabeshima, N., et al.: The pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027 (2020) 4 +1. Gardères, F., Ziaeffard, M., Abeloos, B., Lecue, F.: ConceptBert: Concept-aware representation for visual question answering. In: EMNLP (2020) 10 +2. Geman, D., Geman, S., Hallonquist, N., Younes, L.: Visual Turing test for computer vision systems. Proceedings of the National Academy of Sciences 112(12), 3618-3623 (2015) 1 +3. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In: IEEE Conf. Comput. Vis. Pattern Recog. (2017) 2, 3, 4 +4. Gui, L., Wang, B., Huang, Q., Hauptmann, A., Bisk, Y., Gao, J.: Kat: A knowledge augmented transformer for vision-and-language. In: NAACL (2022) 2, 4, 10 +5. Guo, J., Li, J., Li, D., Tiong, A.M.H., Li, B., Tao, D., Hoi, S.C.: From images to textual prompts: Zero-shot vqa with frozen large language models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) 4, 10 + +16. Guo, Y., Nie, L., Wong, Y., Liu, Y., Cheng, Z., Kankanhalli, M.: A unified end-to-end retriever-reader framework for knowledge-based vqa. In: ACM Int. Conf. Multimedia (2022) 10 +17. Hu, Y., Hua, H., Yang, Z., Shi, W., Smith, N.A., Luo, J.: Promptcap: Prompt-guided task-aware image captioning. arXiv preprint arXiv:2211.09699 (2022) 4, 10 +18. Hu, Z., Iscen, A., Sun, C., Wang, Z., Chang, K.W., Sun, Y., Schmid, C., Ross, D.A., Fathi, A.: **Retrieval-augmented visual-language pre-training with multi-source multimodal knowledge memory. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) 2, 4, 10 +19. Huang, X., Zhang, Y., Ma, J., Tian, W., Feng, R., Zhang, Y., Li, Y., Guo, Y., Zhang, L.: Tag2text: Guiding vision-language model via image tagging. arXiv preprint arXiv:2303.05657 (2023) 3, 4, 5, 7 +20. Jaegle, A., Borgeaud, S., Alayrac, J.B., Doersch, C., Ionescu, C., Ding, D., Koppula, S., Zoran, D., Brock, A., Shelhamer, E., et al.: Perceiver io: A general architecture for structured inputs & outputs. In: Int. Conf. Learn. Represent. (2022) 8 +21. Jiang, Y., Natarajan, V., Chen, X., Rohrbach, M., Batra, D., Parikh, D.: Pythia v0. 1: The winning entry to the vqa challenge 2018. arxiv 2018. arXiv preprint arXiv:1807.09956 (2018) 10 +22. Johnson, J., Hariharan, B., Van Der Maaten, L., Fei-Fei, L., Lawrence Zitnick, C., Girshick, R.: Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2901-2910 (2017) 2 +23. Kamath, A., Clark, C., Gupta, T., Kolve, E., Hoiem, D., Kembhavi, A.: Webly supervised concept expansion for general purpose vision models. In: Eur. Conf. Comput. Vis. (2022) 10 +24. Khan, Z., BG, V.K., Schulter, S., Yu, X., Fu, Y., Chandraker, M.: Q: How to specialize large vision-language models to data-scarce vqa tasks? a: Self-train on unlabeled images! In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) 4 +25. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023) 4 +26. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. In: Int. Conf. Mach. Learn. (2023) 2, 4, 5, 10, 13 +27. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: common objects in context. In: Eur. Conf. Comput. Vis. (2014) 4 +28. Lin, W., Chen, J., Mei, J., Coca, A., Byrne, B.: Fine-grained late-interaction multimodal retrieval for retrieval augmented visual question answering. In: Adv. Neural Inform. Process. Syst. (2023) 10 +29. Lin, Y., Xie, Y., Chen, D., Xu, Y., Zhu, C., Yuan, L.: Revive: Regional visual representation matters in knowledge-based visual question answering. In: Adv. Neural Inform. Process. Syst. (2022) 2, 4, 10 +30. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: Adv. Neural Inform. Process. Syst. (2023) 2, 3, 4, 11 +31. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: Int. Conf. Learn. Represent. (2019) 9 +32. Lu, J., Batra, D., Parikh, D., Lee, S.: Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In: Adv. Neural Inform. Process. Syst. (2019) 10 + +33. Lu, J., Clark, C., Zellers, R., Mottaghi, R., Kembhavi, A.: Unified-io: A unified model for vision, language, and multi-modal tasks. arXiv preprint arXiv:2206.08916 (2022) 10 +34. Luo, G., Zhou, Y., Ren, T., Chen, S., Sun, X., Ji, R.: Cheap and quick: Efficient vision-language instruction tuning for large language models. arXiv preprint arXiv:2305.15023 (2023) 4 +35. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008) 6 +36. Marino, K., Chen, X., Parikh, D., Gupta, A., Rohrbach, M.: Krisp: Integrating implicit and symbolic knowledge for open-domain knowledge-based vqa. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021) 10 +37. Marino, K., Rastegari, M., Farhadi, A., Mottaghi, R.: Ok-vqa: A visual question answering benchmark requiring external knowledge. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019) 2, 3, 4, 5, 9, 10 +38. Merullo, J., Castricato, L., Eickhoff, C., Pavlick, E.: Linearly mapping from image to text space. In: Int. Conf. Learn. Represent. (2023) 13 +39. Mokady, R., Hertz, A., Bermano, A.H.: Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734 (2021) 10 +40. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Int. Conf. Mach. Learn. (2021) 9, 13 +41. Reimers, N., Gurevych, I.: Sentence-bert: Sentence embeddings using siamese bert-networks. In: EMNLP (2019) 6, 7 +42. Schwenk, D., Khandelwal, A., Clark, C., Marino, K., Mottaghi, R.: A-okvqa: A benchmark for visual question answering using world knowledge. In: Eur. Conf. Comput. Vis. (2022) 2, 3, 4, 5, 9, 10 +43. Selvaraju, R.R., Tendulkar, P., Parikh, D., Horvitz, E., Ribeiro, M.T., Nushi, B., Kamar, E.: Squinting at vqa models: Introspecting vqa models with sub-questions. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020) 4 +44. Shao, Z., Yu, Z., Wang, M., Yu, J.: Prompting large language models with answer heuristics for knowledge-based visual question answering. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) 4, 10 +45. Shen, S., Li, L.H., Tan, H., Bansal, M., Rohrbach, A., Chang, K.W., Yao, Z., Keutzer, K.: How much can clip benefit vision-and-language tasks? In: Int. Conf. Learn. Represent. (2022) 13 +46. Speer, R., Chin, J., Havasi, C.: Conceptnet 5.5: An open multilingual graph of general knowledge. In: AAAI (2017) 4, 10 +47. Tan, H., Bansal, M.: Lxmert: Learning cross-modality encoder representations from transformers. In: EMNLP (2019) 10 +48. Uehara, K., Duan, N., Harada, T.: Learning to ask informative sub-questions for visual question answering. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) 4 +49. Vrandecic, D., Krötzsch, M.: Wikidata: a free collaborative knowledgebase. Communications of the ACM 57(10), 78-85 (2014) 4, 10 +50. Wang, P., Wu, Q., Shen, C., Dick, A., Van Den Hengel, A.: Fvqa: Fact-based visual question answering. IEEE transactions on pattern analysis and machine intelligence 40(10), 2413-2427 (2017) 2, 4 +51. Wang, P., Wu, Q., Shen, C., Hengel, A.v.d., Dick, A.: Explicit knowledge-based reasoning for visual question answering. arXiv preprint arXiv:1511.02570 (2015) 2, 4 +52. Wu, J., Lu, J., Sabharwal, A., Mottaghi, R.: Multi-modal answer validation for knowledge-based vqa. In: AAAI (2021) 10 + +53. Yang, Z., Gan, Z., Wang, J., Hu, X., Lu, Y., Liu, Z., Wang, L.: An empirical study of gpt-3 for few-shot knowledge-based vqa. In: AAAI (2022) 10 +54. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023) 2, 4 +55. Zellers, R., Bisk, Y., Farhadi, A., Choi, Y.: From recognition to cognition: Visual commonsense reasoning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019) 2 +56. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199 (2023) 4 +57. Zhang, Y., Huang, X., Ma, J., Li, Z., Luo, Z., Xie, Y., Qin, Y., Luo, T., Li, Y., Liu, S., et al.: Recognize anything: A strong image tagging model. arXiv preprint arXiv:2306.03514 (2023) 3, 4, 5, 7, 9 +58. Zhang, Z., Zhang, A., Li, M., Zhao, H., Karypis, G., Smola, A.: Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923 (2023) 8 +59. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023) 3, 4, 11 +60. Zhu, Z., Yu, J., Wang, Y., Sun, Y., Hu, Y., Wu, Q.: Mucko: multi-layer cross-modal knowledge reasoning for fact-based visual question answering. In: IJCAI (2020) 10 \ No newline at end of file diff --git a/2024/Q&A Prompts_ Discovering Rich Visual Clues through Mining Question-Answer Prompts for VQA requiring Diverse World Knowledge/images.zip b/2024/Q&A Prompts_ Discovering Rich Visual Clues through Mining Question-Answer Prompts for VQA requiring Diverse World Knowledge/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0a61e73e8a6858cc761fc2c8aa043a275fde3555 --- /dev/null +++ b/2024/Q&A Prompts_ Discovering Rich Visual Clues through Mining Question-Answer Prompts for VQA requiring Diverse World Knowledge/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b78e8644822d1a09c515bd98aad8d251093677825be8ad1b84e2c221d5679c8 +size 563389 diff --git a/2024/Q&A Prompts_ Discovering Rich Visual Clues through Mining Question-Answer Prompts for VQA requiring Diverse World Knowledge/layout.json b/2024/Q&A Prompts_ Discovering Rich Visual Clues through Mining Question-Answer Prompts for VQA requiring Diverse World Knowledge/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2844163cbfecf53355c155b6576837a9f7766b0a --- /dev/null +++ b/2024/Q&A Prompts_ Discovering Rich Visual Clues through Mining Question-Answer Prompts for VQA requiring Diverse World Knowledge/layout.json @@ -0,0 +1,12022 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 111, + 470, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 111, + 470, + 166 + ], + "spans": [ + { + "bbox": [ + 146, + 111, + 470, + 166 + ], + "type": "text", + "content": "Q&A Prompts: Discovering Rich Visual Clues through Mining Question-Answer Prompts for VQA requiring Diverse World Knowledge" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 220, + 185, + 394, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 185, + 394, + 199 + ], + "spans": [ + { + "bbox": [ + 220, + 185, + 394, + 199 + ], + "type": "text", + "content": "Haibo Wang" + }, + { + "bbox": [ + 220, + 185, + 394, + 199 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 220, + 185, + 394, + 199 + ], + "type": "text", + "content": " and Weifeng Ge" + }, + { + "bbox": [ + 220, + 185, + 394, + 199 + ], + "type": "inline_equation", + "content": "^{1,2\\boxtimes}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 168, + 207, + 444, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 207, + 444, + 242 + ], + "spans": [ + { + "bbox": [ + 168, + 207, + 444, + 242 + ], + "type": "text", + "content": "1 Nebula AI Group, School of Computer Science, Fudan University \n2 Shanghai Key Lab of Intelligent Information Processing hbwang22@m.fudan.edu.cn, wfge@fudan.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 273, + 455, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 273, + 455, + 537 + ], + "spans": [ + { + "bbox": [ + 160, + 273, + 455, + 537 + ], + "type": "text", + "content": "Abstract. With the breakthrough of multi-modal large language models (MLLMs), answering complex visual questions that demand advanced reasoning abilities and world knowledge has become a much more important testbed for developing AI models than ever. However, equipping MLLMs with robust cross-modality reasoning ability remains challenging since the cognition scheme of humans has not been understood systematically. In this paper, we believe that if we can collect rich visual clues, we will recognize the image more accurately, understand the question better, recall relevant knowledge more easily, and finally reason out the answer. We discover these rich visual clues by mining question-answer pairs in images and sending them into multi-modal large language models as prompts. We call the proposed method Q&A Prompts. Specifically, we first use the image-answer pairs and the corresponding questions in the training set as inputs and outputs to train a visual question generation (VQG) model. Then, we use an image tagging model to identify various instances and send packaged image-tag pairs into the VQG model to generate relevant questions with the extracted image tags as answers. Finally, we encode these generated question-answer pairs as prompts with a visual-aware prompting module and send them into pretrained MLLMs to reason out the final answers. Experimental results show that, compared with state-of-the-art methods, our Q&A Prompts achieves substantial improvements on the challenging visual question answering datasets requiring reasoning over diverse world knowledge, such as OK-VQA and A-OKVQA. Codes will be available at link." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 547, + 453, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 547, + 453, + 571 + ], + "spans": [ + { + "bbox": [ + 159, + 547, + 453, + 571 + ], + "type": "text", + "content": "Keywords: multi-modal large language model " + }, + { + "bbox": [ + 159, + 547, + 453, + 571 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 547, + 453, + 571 + ], + "type": "text", + "content": " visual question answering " + }, + { + "bbox": [ + 159, + 547, + 453, + 571 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 547, + 453, + 571 + ], + "type": "text", + "content": " visual language reasoning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 591, + 230, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 591, + 230, + 604 + ], + "spans": [ + { + "bbox": [ + 132, + 591, + 230, + 604 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Visual Question Answering (VQA) has longstanding been considered a Visual Turing Test [12] in the machine learning community. To solve this Turing test, artificial intelligence (AI) models are expected to be equipped with the human capabilities of visual recognition, language understanding, logical reasoning, world" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 148, + 125, + 230, + 177 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 228, + 123 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 228, + 123 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 228, + 123 + ], + "type": "text", + "content": "VQA Tasks well-solved by MLLMs" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 148, + 125, + 230, + 177 + ], + "lines": [ + { + "bbox": [ + 148, + 125, + 230, + 177 + ], + "spans": [ + { + "bbox": [ + 148, + 125, + 230, + 177 + ], + "type": "image", + "image_path": "378d7f50367610104950f29c2507ff8f41ccd40980f1e804df8a8295710bbea5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 238, + 125, + 323, + 184 + ], + "blocks": [ + { + "bbox": [ + 240, + 114, + 323, + 123 + ], + "lines": [ + { + "bbox": [ + 240, + 114, + 323, + 123 + ], + "spans": [ + { + "bbox": [ + 240, + 114, + 323, + 123 + ], + "type": "text", + "content": "VQA Tasks unsolved by MLLMs" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 238, + 125, + 323, + 184 + ], + "lines": [ + { + "bbox": [ + 238, + 125, + 323, + 184 + ], + "spans": [ + { + "bbox": [ + 238, + 125, + 323, + 184 + ], + "type": "image", + "image_path": "092e870ac39982361dc1d008fa0b49a7d4c9340f959388bfda9975accc98aa90.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 147, + 186, + 227, + 245 + ], + "blocks": [ + { + "bbox": [ + 147, + 186, + 227, + 245 + ], + "lines": [ + { + "bbox": [ + 147, + 186, + 227, + 245 + ], + "spans": [ + { + "bbox": [ + 147, + 186, + 227, + 245 + ], + "type": "image", + "image_path": "66ae73ade15069150a3b6540ba59cfc07d167dedc0194298fb820cba1a5a7e58.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 254, + 482, + 299 + ], + "lines": [ + { + "bbox": [ + 130, + 254, + 482, + 299 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 482, + 299 + ], + "type": "text", + "content": "Fig. 1: Illustration of VQA tasks that require simple perception abilities ((a) and (b)) and those demanding chains of reasoning over diverse world knowledge ((c) and (d)). The example on the right ((e) and (f)) indicates MLLMs with Q&A prompts can solve difficult VQA problems that require reasoning over diverse world knowledge." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 238, + 186, + 323, + 245 + ], + "blocks": [ + { + "bbox": [ + 238, + 186, + 323, + 245 + ], + "lines": [ + { + "bbox": [ + 238, + 186, + 323, + 245 + ], + "spans": [ + { + "bbox": [ + 238, + 186, + 323, + 245 + ], + "type": "image", + "image_path": "e5d657845a74283f272017c9a1f73e768e1da56eca27fb2f108f6faa50ad7b33.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 325, + 125, + 473, + 245 + ], + "blocks": [ + { + "bbox": [ + 340, + 114, + 465, + 123 + ], + "lines": [ + { + "bbox": [ + 340, + 114, + 465, + 123 + ], + "spans": [ + { + "bbox": [ + 340, + 114, + 465, + 123 + ], + "type": "text", + "content": "VQA Tasks solved by MLLMs with Q&A Prompts" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 325, + 125, + 473, + 245 + ], + "lines": [ + { + "bbox": [ + 325, + 125, + 473, + 245 + ], + "spans": [ + { + "bbox": [ + 325, + 125, + 473, + 245 + ], + "type": "image", + "image_path": "66bb53699443cc5b275aa8fc2e0bf0c4bacbba12d7c557a3b3269735f3e4d24f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 326, + 480, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 480, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 480, + 410 + ], + "type": "text", + "content": "knowledge referring, etc. There are a variety of datasets focusing on different problems in VQA, such as the perception and language understanding problem [3, 13], reasoning on procedurally generated images [22], and structured factual knowledge [50, 51] or commonsense knowledge [55]. Recently, with the advent of multi-modal large language models (MLLMs) [6, 26, 30, 54], significant progress has been achieved in perception-based VQA tasks [3, 13]. Fig. 1 (a) and Fig. 1 (b) show such VQA tasks, which can be reliably solved by MLLMs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 411, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 604 + ], + "type": "text", + "content": "However, for more complex VQA tasks [37, 42] as in Fig. 1 (c) and Fig. 1 (d), that feature the necessity of diverse world knowledge and complex chains of reasoning, state-of-the-art methods [14, 18, 29] even MLLM models [6, 26] fail to give the correct answers. Different from perception-based VQA [3, 13] and \"closed\" knowledge-based VQA [50, 51], visual question reasoning in OK-VQA [37] and A-OKVQA [42] requires not only diverse forms of world knowledge but also need complex reasoning logic. In Fig. 1 (d), where a woman sits on the ground with her pet dogs and holds an umbrella, the question is [Why is the woman holding the umbrella?] Simply associating the keyword [umbrella] with the word [rain] due to their frequent co-occurrence in the pre-training corpus [7] will lead to the wrong answer. To reason out the correct answer, AI models need to consider rich visual clues presented in the image. For example, in Fig. 1 (f), having discovered the clues including [the woman is camping] and [the weather is shiny] from the Q&A pairs, the model can correctly fill the logic gap and arrive at the answer [for shade]. This evidence indicates the necessity of collecting rich visual clues to achieve good results." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Based on the analysis above, we believe that collecting visual clues of various instances in images will help MLLM recognize the image more accurately, understand the question better, recall relevant knowledge more easily, and finally generate the reasoning results more intelligently. To investigate this hypothesis, we design a novel VQA framework called " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "Q\\mathcal{E}A" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " Prompts, which extracts nu" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 270, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 270, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 270, + 102 + ], + "type": "text", + "content": "Haibo Wang, Weifeng Ge" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "merous question-answer pairs associated with different perspectives in images as prompts for MLLMs. We choose an instance in the given image as the answer and generate the related question to form a question-answer pair. We generate the question-answer pairs for almost every instance to get a bag of Q&A prompts. The reason we choose question-answer pairs as prompts comes from two folds: First, Q&A prompts can focus on diverse instances in images like objects, scenes, or actions, thereby offering multiple different perspectives to understand the given image and the target question; Second, the knowledge and reasoning insights hidden in these question-answer pairs can help to explicitly uncover more relevant world knowledge, which may be critical to filling the logic gap between perception and reasoning when answering the target question." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 248, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 477 + ], + "type": "text", + "content": "In Q&A Prompts, there are three key stages, which are the visual question generation (VQG) model training stage, the question-answer prompts generation stage, and the visual-language reasoning stage. In the VQG model training stage, we gather question-answer pairs in a VQA training set to train a VQG model that can map an answer together with the given image to a correlated question. Subsequently, in the question-answer prompts generation stage, we leverage a pre-trained image tagging model [19, 57] to recognize various objects, scenes, and actions in images with text tags. We employ these tags as answers and send them into the VQG model along with the corresponding images to generate questions. Finally, in the visual-language reasoning stage, we encode the generated question-answer pairs with a visual-aware prompting module and send them together with image and question features into a frozen language model to reason out the answer. Since previous VQA benchmarks [3,13] have seen saturated performance, we conduct experiments on the more challenging OK-VQA [37] and A-OKVQA [42] datasets, which are newer and larger testbeds for MLLM to investigate their abilities in accomplishing reasoning tasks requiring diverse world knowledge. Experimental results show that our Q&A Prompts can substantially improve the reasoning ability of MLLMs, such as InstructBLIP [6], LLaVA [30], and MiniGPT-4 [59]. In summary, we make the following contributions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 485, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 485, + 481, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 485, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 138, + 485, + 481, + 545 + ], + "type": "text", + "content": "- We propose a novel VQA framework that effectively generates and leverages question-answer pairs as prompts to improve the reasoning ability of MLLMs. It provides a new perspective for the community to design VQA solutions by explicitly collecting rich visual clues to bridge the logic gap between perception and reasoning when answering the target question." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 546, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 546, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 138, + 546, + 481, + 605 + ], + "type": "text", + "content": "- We design a novel question-answer prompts generation scheme with a VQG model and an image tagging model, which can generate Q&A prompts for recognizable objects, scenes, and activities in images. Besides, we design a new visual-aware prompting module to encode these prompts efficiently for the subsequent reasoning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 605, + 481, + 665 + ], + "type": "text", + "content": "- We test the method on the challenging OK-VQA [37] and A-OKVQA [42] benchmarks, which study reasoning over diverse forms of world knowledge. The proposed Q&A Prompts achieves an accuracy of " + }, + { + "bbox": [ + 138, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "68.1\\%" + }, + { + "bbox": [ + 138, + 605, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 138, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "64.3\\%" + }, + { + "bbox": [ + 138, + 605, + 481, + 665 + ], + "type": "text", + "content": " on A-OKVQA [42] and OK-VQA [37], outperforming previous state-of-the-art methods by clear margins." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "type": "text", + "content": "Q&A Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 481, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 481, + 281 + ], + "type": "text", + "content": "Multimodal Large Language Models (MLLMs). MLLMs [2,6,25,26,34,54, 56,59] have shown strong ability in image-language understanding and reasoning, by adapting frozen language models to frozen image encoders with trainable connection modules. For instance, Flamingo [2] incorporates visual features into the language model with gated cross-attention blocks. LLaVA [30], instead, directly projects visual features into the space of text embeddings using a linear layer. BLIP-2 and InstructBLIP [6,26] introduce a more complex Q-Former to bridge the modality gap. These models align images and language based on large-scale image-text pre-training [5,10,27] and conduct reasoning with language models. In this paper, we mine more numerous question-answer pairs and send them into the MLLM as prompts. These prompts provide rich visual clues and uncover relevant world knowledge, which is beneficial to reasoning out correct answers." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 282, + 481, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 481, + 498 + ], + "type": "text", + "content": "VQA Requiring Reasoning over Diverse World Knowledge. Different from classical VQA tasks [3, 13, 50, 51], VQA requiring reasoning over world knowledge involves inferring intents, goals, the physics of the world, and the social dynamics of individuals. More recently, A-OKVQA [42] and OK-VQA [37], which our work focuses on, present large-scale challenging VQA datasets that encompass questions requiring reasoning based on diverse world knowledge, including visual knowledge, commonsense knowledge, and factoid knowledge. Previous methods like Img2LLM [15], Prophet [44], and PromptCap [17] exploit the rich hidden knowledge and strong reasoning abilities of LLMs (e.g., GPT3 [4]) to solve such problems. These methods convert images into natural languages and construct various prompt templates for LLMs with in-context learning [4] to generate answers. Another group of works, such as KAT [14], REVIVE [29], and REVEAL [18], directly encode questions and relevant knowledge retrieved from external knowledge bases [4, 46, 49] to train end-to-end models. Unlike these methods, we design a visual-question prompts generation scheme to mine the information in images and unambiguous words to reduce confusion when answering the target questions. With Q&A prompts, different types of world knowledge are incorporated into the reasoning process to help MLLMs infer the correct answers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 498, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 666 + ], + "type": "text", + "content": "Visual Question Generation (VQG). Various works have explored Visual Question Generation (VQG) in the context of VQA tasks. SQuINT [43] and SelTDA [24] focus on dataset augmentation to enhance the generalization of VQA models. SQuINT employed manual workers to annotate questions, while SelTDA generated questions automatically using unlabeled images. Another line of research, exemplified by [48], employs a VQG model to generate informative sub-questions. These sub-questions are then used alongside the original question during inference, enabling another VQA model to answer both simultaneously. Different from previous works, our VQG model uses the tags of images [19, 57] as answers for generated questions. This approach ensures that our VQG model generates more relevant questions with additional details and perspectives. Furthermore, we introduce a novel module to encode the question-answer pairs as prompts for MLLMs, instead of directly feeding them as the formation of natural language into the model [48]." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 270, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 270, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 270, + 102 + ], + "type": "text", + "content": "Haibo Wang, Weifeng Ge" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 482, + 266 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 482, + 266 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 482, + 266 + ], + "type": "image", + "image_path": "a62552742be334efc773e4f01521fea1e129b587aadfed2ff84415e3404f26e9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 274, + 482, + 330 + ], + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 330 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 330 + ], + "type": "text", + "content": "Fig. 2: (a) Firstly, we train a VQG model that can ask informative questions given the image and specified answer. (b) Then, we exploit the image tagging model to extract image tags as the answers to generated questions and obtain diverse question-answer prompts. (c) At last, we feed the image, question, and question-answer prompts into the VQA model to perform visual-language reasoning in (c)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 351, + 201, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 351, + 201, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 351, + 201, + 363 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 376, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 569 + ], + "type": "text", + "content": "Q&A Prompts is a conceptually simple three-stage framework, and Fig. 2 gives an overview: a VQG model training stage, a question-answer prompts generation stage, and a visual-language reasoning stage. Specifically, in the VQG model training stage (Fig. 2 (a)), we train a vanilla MLLM with a broad range of image-answer pairs as input and the corresponding question as the target output to learn a mapping from answers to questions (detailed in Sec. 3.1). In the prompts generation stage (Fig. 2 (b)), we exploit an image tagging model [19, 57] to obtain relevant tags including attributes, objects, actions, and scenes in images. These tags serve as answers and will be fed into the VQG model trained in the first stage to generate diverse questions containing rich visual clues and world knowledge (detailed in Sec. 3.2). In the final reasoning stage (Fig. 2 (c)), we utilize our proposed visual-aware prompting module to encode these question-answer pairs into prompts, which will be fed into the MLLM such as BLIP-2 [26] or InstructBLIP [6] together with the image and the target question, to predict the answer (detailed in Sec. 3.3). The whole pipeline is flexible and effective in mining informative visual clues to correctly answer the given questions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 586, + 318, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 318, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 318, + 598 + ], + "type": "text", + "content": "3.1 Stage-1: Train the VQG Model" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Given an image " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " and answer " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ", the VQG model is responsible for posing an informative question " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " which can be properly answered by " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ". To explore as many visual clues in the generated questions as possible, and avoid generating general questions containing limited clues (e.g., \"What is this?\"), we utilize the A-OKVQA [42] or OK-VQA [37] as the training set " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " for our VQG model, since" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "type": "text", + "content": "Q&A Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 242, + 178 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 242, + 178 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 242, + 178 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 242, + 178 + ], + "type": "image", + "image_path": "029793f67d72050a57a9e7b74347ae346d883a9305c90d77a4a6e6e3ab2959ea.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 180, + 306, + 221 + ], + "lines": [ + { + "bbox": [ + 198, + 180, + 306, + 221 + ], + "spans": [ + { + "bbox": [ + 198, + 180, + 306, + 221 + ], + "type": "text", + "content": "A-OKVQA What sport do the people have equipment for? \nVQG Which object in this image used for winter sports? \nZero-shot Is this a snowboard?" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 255, + 114, + 361, + 178 + ], + "blocks": [ + { + "bbox": [ + 255, + 114, + 361, + 178 + ], + "lines": [ + { + "bbox": [ + 255, + 114, + 361, + 178 + ], + "spans": [ + { + "bbox": [ + 255, + 114, + 361, + 178 + ], + "type": "image", + "image_path": "8492208510e5bde4c93c6360c46596cc3926ad6e6b6e3869093c7d23d2f25f6b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 372, + 114, + 479, + 178 + ], + "blocks": [ + { + "bbox": [ + 372, + 114, + 479, + 178 + ], + "lines": [ + { + "bbox": [ + 372, + 114, + 479, + 178 + ], + "spans": [ + { + "bbox": [ + 372, + 114, + 479, + 178 + ], + "type": "image", + "image_path": "f470664c4fe5cd2a5670c7cb9e920a8f4d6ee942ef5a24b4a19784c02d43d2a0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 180, + 478, + 221 + ], + "lines": [ + { + "bbox": [ + 378, + 180, + 478, + 221 + ], + "spans": [ + { + "bbox": [ + 378, + 180, + 478, + 221 + ], + "type": "text", + "content": "A-OKVQA What is the vehicle traveling on? \nVQG What is this type of transportation called? \nZero-shot What is this?" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 135, + 179, + 197, + 221 + ], + "blocks": [ + { + "bbox": [ + 135, + 179, + 197, + 221 + ], + "lines": [ + { + "bbox": [ + 135, + 179, + 197, + 221 + ], + "spans": [ + { + "bbox": [ + 135, + 179, + 197, + 221 + ], + "type": "image", + "image_path": "d877037b565550b9fa92e2376ea343376f2583f821eebf7faf88b68adb28a9ce.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 223, + 201, + 232 + ], + "lines": [ + { + "bbox": [ + 136, + 223, + 201, + 232 + ], + "spans": [ + { + "bbox": [ + 136, + 223, + 201, + 232 + ], + "type": "text", + "content": " snowboard" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 180, + 377, + 221 + ], + "blocks": [ + { + "bbox": [ + 316, + 180, + 377, + 221 + ], + "lines": [ + { + "bbox": [ + 316, + 180, + 377, + 221 + ], + "spans": [ + { + "bbox": [ + 316, + 180, + 377, + 221 + ], + "type": "image", + "image_path": "97de4a57b3897445ccd8a0a038f0ddb8f0fec304f108cfbecb4f20cc8aaa7b60.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 224, + 359, + 232 + ], + "lines": [ + { + "bbox": [ + 317, + 224, + 359, + 232 + ], + "spans": [ + { + "bbox": [ + 317, + 224, + 359, + 232 + ], + "type": "text", + "content": " rail" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 240, + 480, + 275 + ], + "lines": [ + { + "bbox": [ + 130, + 240, + 480, + 275 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 480, + 275 + ], + "type": "text", + "content": "Fig. 3: Visualizations and T-SNE embedding distributions of the questions in manually annotated A-OKVQA (green), and the questions generated by our VQG model (blue) or by directly prompting a frozen MLLM (orange)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 297, + 480, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 480, + 356 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 480, + 356 + ], + "type": "text", + "content": "the questions in their annotations are diverse and specific, covering broader areas of knowledge. To create the VQG model that approximates " + }, + { + "bbox": [ + 130, + 297, + 480, + 356 + ], + "type": "inline_equation", + "content": "P(\\mathcal{Q}|\\mathcal{V},\\mathcal{A})" + }, + { + "bbox": [ + 130, + 297, + 480, + 356 + ], + "type": "text", + "content": ", we treat the problem of learning such a model as a text-generation problem and wish to train the MLLM (e.g., InstructBLIP [6]) as our VQG model with " + }, + { + "bbox": [ + 130, + 297, + 480, + 356 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 297, + 480, + 356 + ], + "type": "text", + "content": " due to its strong image-language reasoning ability." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "text", + "content": "Specifically, as shown in Fig. 2 (a), the VQG model comprises a frozen image encoder, a trainable connection module, and a frozen LLM. We extract image embeddings " + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_v\\in \\mathbb{R}^{n\\times d_v}" + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "text", + "content": " with the image encoder (" + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "text", + "content": " is the patch number, " + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "inline_equation", + "content": "d_v" + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "text", + "content": " is the embedding dimension), and then feed " + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_v" + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "text", + "content": " into the connection module to obtain fixed-length visual features " + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_v\\in \\mathbb{R}^{k\\times d_q}" + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "text", + "content": ", which will be sent into the LLM together with the instruction. We write the instruction as the template containing the answer " + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "text", + "content": ", such as \"Given the image, generate a question whose answer is: " + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\langle \\mathcal{A}\\rangle" + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "text", + "content": ".\" The model is trained using the cross-entropy loss with parameters " + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 228, + 459, + 481, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 459, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 228, + 459, + 481, + 495 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {V Q G}} = - \\sum_ {t = 1} ^ {L _ {q}} \\log P _ {\\theta} \\left(\\mathcal {Q} _ {t} \\mid \\mathcal {Q} _ {< t}, \\mathcal {V}, \\mathcal {A}\\right) \\tag {1}", + "image_path": "82ae2ccd511c7c91c3435ac31f5e3840fcdb03789f3b26847befe47f28453736.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 498, + 482, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 608 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 498, + 482, + 608 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_t" + }, + { + "bbox": [ + 130, + 498, + 482, + 608 + ], + "type": "text", + "content": " is predicted autoregressively at position " + }, + { + "bbox": [ + 130, + 498, + 482, + 608 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 498, + 482, + 608 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 498, + 482, + 608 + ], + "type": "inline_equation", + "content": "L_{q}" + }, + { + "bbox": [ + 130, + 498, + 482, + 608 + ], + "type": "text", + "content": " is the sequence length of the question " + }, + { + "bbox": [ + 130, + 498, + 482, + 608 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 130, + 498, + 482, + 608 + ], + "type": "text", + "content": ". We emphasize the necessity of training the VQG model, instead of prompting a frozen MLLM in a zero-shot manner. In Fig. 3, we use sentence-bert [41] to obtain the dense embedding vectors of each question and show the T-SNE [35] embedding of them, as well as visualization examples. We can observe that the questions generated by our VQG model (blue) are more diverse and specific like questions in A-OKVQA (green), while the questions generated by MLLM without training (orange) are often general and limited with low differentiation. Our following experiments will further verify this." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 623, + 388, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 623, + 388, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 623, + 388, + 635 + ], + "type": "text", + "content": "3.2 Stage-2: Generate Question-Answer Prompts" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Once the VQG model has been obtained, the generation of question-answer pairs can proceed. In this stage, we will explicitly generate a wide range of question-" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "type": "text", + "content": "Haibo Wang, Weifeng Ge" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 114, + 478, + 187 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 478, + 187 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 478, + 187 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 478, + 187 + ], + "type": "image", + "image_path": "3d4224a2be9d5450b8c6a0cc34ed314f6e817eb757ce5b15aa4ddd293b8aff71.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 194, + 480, + 217 + ], + "lines": [ + { + "bbox": [ + 131, + 194, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 131, + 194, + 480, + 217 + ], + "type": "text", + "content": "Fig. 4: Overview of the visual-aware prompting module, consisting of the prompt encoder, visual gated fusion, and resampler decoder." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 241, + 479, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 241, + 479, + 337 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 479, + 337 + ], + "type": "text", + "content": "answer pairs to help the model have a deeper and broader understanding of the image. In order to generate diverse questions covering various perspectives of the image, we should ensure the answers fed into the VQG model are sufficient in varied forms. Therefore, we adopt a strong image tagging model [19, 57] to produce rich tags as answers. Different from object detectors that can only recognize limited object labels, image tagging models aim to identify various open-vocabulary elements in images, including attributes, objects, actions, scenes, etc, which can provide an appropriate source of the answers we need." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 337, + 480, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 337, + 480, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 337, + 480, + 420 + ], + "type": "text", + "content": "We utilize the RAM [57] model as our image tagging model due to its superior recognition ability. As shown in Fig. 2 (b), we feed the image " + }, + { + "bbox": [ + 130, + 337, + 480, + 420 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 337, + 480, + 420 + ], + "type": "text", + "content": " into the tagging model and obtain various tags for each image. To make the tags of each image more distinguishable, we remove the most frequent ones containing general information, such as 'person', 'food', 'man', etc., resulting in an average of 14 tags per image. We denote this set of tags as " + }, + { + "bbox": [ + 130, + 337, + 480, + 420 + ], + "type": "inline_equation", + "content": "\\mathcal{O} = \\{o_i\\}_{i=1}^M" + }, + { + "bbox": [ + 130, + 337, + 480, + 420 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 337, + 480, + 420 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 337, + 480, + 420 + ], + "type": "text", + "content": " is the number of tags associated with an image." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 420, + 480, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 420, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 131, + 420, + 480, + 504 + ], + "type": "text", + "content": "Using these tags as answers, we input them into the trained VQG model along with the image, as shown in Fig. 2 (b), to obtain corresponding questions that answer these tags. To ensure that the generated questions (denoted as " + }, + { + "bbox": [ + 131, + 420, + 480, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{t_i\\}_{i=1}^M" + }, + { + "bbox": [ + 131, + 420, + 480, + 504 + ], + "type": "text", + "content": ") contribute effectively to answering the target question, we rank the generated questions " + }, + { + "bbox": [ + 131, + 420, + 480, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{t_i\\}_{i=1}^M" + }, + { + "bbox": [ + 131, + 420, + 480, + 504 + ], + "type": "text", + "content": " based on their similarity with the target question " + }, + { + "bbox": [ + 131, + 420, + 480, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 131, + 420, + 480, + 504 + ], + "type": "text", + "content": ", and adopt the questions with the Top-P similarities as the final obtained question prompts " + }, + { + "bbox": [ + 131, + 420, + 480, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 131, + 420, + 480, + 504 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 218, + 513, + 480, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 513, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 218, + 513, + 480, + 533 + ], + "type": "interline_equation", + "content": "\\mathcal {U} = \\left\\{u _ {p} \\right\\} _ {p = 1} ^ {P} = \\underset {t _ {i} \\in \\mathcal {T}} {\\arg \\operatorname {T o p - P}} \\langle \\mathrm {T} (t _ {i}), \\mathrm {T} (\\mathcal {Q}) \\rangle \\tag {2}", + "image_path": "575ec7df12d704987ac2679d62a59b328656615fd721e35afebc461c374dd091.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 543, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 543, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 131, + 543, + 482, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 543, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\mathrm{T}(\\cdot)" + }, + { + "bbox": [ + 131, + 543, + 482, + 604 + ], + "type": "text", + "content": " represents the embedding model [41], which encodes sentences into embedding vectors for similarity computation. The " + }, + { + "bbox": [ + 131, + 543, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\langle \\cdot, \\cdot \\rangle" + }, + { + "bbox": [ + 131, + 543, + 482, + 604 + ], + "type": "text", + "content": " operator is the inner product, computing the cosine similarity between two sentences. We then concatenate each " + }, + { + "bbox": [ + 131, + 543, + 482, + 604 + ], + "type": "inline_equation", + "content": "u_i" + }, + { + "bbox": [ + 131, + 543, + 482, + 604 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 131, + 543, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 131, + 543, + 482, + 604 + ], + "type": "text", + "content": " and its corresponding " + }, + { + "bbox": [ + 131, + 543, + 482, + 604 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 131, + 543, + 482, + 604 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 131, + 543, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 131, + 543, + 482, + 604 + ], + "type": "text", + "content": " using the template \"Question: " + }, + { + "bbox": [ + 131, + 543, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\langle u_i \\rangle" + }, + { + "bbox": [ + 131, + 543, + 482, + 604 + ], + "type": "text", + "content": " Answer: " + }, + { + "bbox": [ + 131, + 543, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\langle o_i \\rangle" + }, + { + "bbox": [ + 131, + 543, + 482, + 604 + ], + "type": "text", + "content": "\", resulting in the final question-answer pairs as prompts " + }, + { + "bbox": [ + 131, + 543, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 131, + 543, + 482, + 604 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 621, + 359, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 621, + 359, + 633 + ], + "spans": [ + { + "bbox": [ + 131, + 621, + 359, + 633 + ], + "type": "text", + "content": "3.3 Stage-3: Reasoning with Q&A Prompts" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "type": "text", + "content": "After obtaining these question-answer pairs involving rich visual clues and world knowledge, we use the proposed visual-aware prompting module, as shown in" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 386, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 386, + 91, + 447, + 102 + ], + "type": "text", + "content": "Q&A Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "text", + "content": "Fig. 4, to encode these question-answer pairs into a set of prompt embeddings " + }, + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_p\\in \\mathbb{R}^{k\\times d_q}" + }, + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "text", + "content": " for our VQA model. We still utilize the same MLLM (e.g., Instruct-BLIP [6]) as our VQA model for the final reasoning. As in Fig. 2 (c), the encoded prompt embeddings " + }, + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_p" + }, + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "text", + "content": " will be fed into the frozen LLM in MLLM, together with the instruction embeddings and visual features " + }, + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_v" + }, + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "text", + "content": " to predict the answer " + }, + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 230, + 184, + 481, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 184, + 481, + 198 + ], + "spans": [ + { + "bbox": [ + 230, + 184, + 481, + 198 + ], + "type": "interline_equation", + "content": "\\mathcal {A} = \\operatorname {L L M} \\left(\\left[ \\mathbf {F} _ {v}; \\mathbf {F} _ {p}; \\operatorname {E m b e d} (\\operatorname {I n s}) \\right]\\right) \\tag {3}", + "image_path": "dfad2e768d825ed07c562d509c0ff0951d78a2b659876e94cb9180fa7fb3c6ac.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 205, + 479, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 205, + 479, + 252 + ], + "spans": [ + { + "bbox": [ + 131, + 205, + 479, + 252 + ], + "type": "text", + "content": "Embed is the word embedding layer in LLM, the operator " + }, + { + "bbox": [ + 131, + 205, + 479, + 252 + ], + "type": "inline_equation", + "content": "[\\cdot]" + }, + { + "bbox": [ + 131, + 205, + 479, + 252 + ], + "type": "text", + "content": " means concatenation, and Ins is the instruction including the target question " + }, + { + "bbox": [ + 131, + 205, + 479, + 252 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 131, + 205, + 479, + 252 + ], + "type": "text", + "content": ", represented as \"Question: " + }, + { + "bbox": [ + 131, + 205, + 479, + 252 + ], + "type": "inline_equation", + "content": "\\langle \\mathcal{Q} \\rangle" + }, + { + "bbox": [ + 131, + 205, + 479, + 252 + ], + "type": "text", + "content": "? Answer:\". We then introduce how the visual-aware prompting module bootstrapped the reasoning ability of MLLM within Q&A Prompts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 253, + 482, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 253, + 482, + 336 + ], + "spans": [ + { + "bbox": [ + 130, + 253, + 482, + 336 + ], + "type": "text", + "content": "Prompt Encoder. The prompt encoder is responsible for obtaining the semantics representations of the prompts " + }, + { + "bbox": [ + 130, + 253, + 482, + 336 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 253, + 482, + 336 + ], + "type": "text", + "content": ". Specifically, we choose encoder-based language models as our prompt encoder due to their strong ability in language understanding. To preserve the well-trained representations, we keep the parameters of the prompt encoder frozen. We feed the prompts " + }, + { + "bbox": [ + 130, + 253, + 482, + 336 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 253, + 482, + 336 + ], + "type": "text", + "content": " into the prompt encoder and get the encoded " + }, + { + "bbox": [ + 130, + 253, + 482, + 336 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_s \\in \\mathbb{R}^{L_s \\times d_q}" + }, + { + "bbox": [ + 130, + 253, + 482, + 336 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 253, + 482, + 336 + ], + "type": "inline_equation", + "content": "L_s" + }, + { + "bbox": [ + 130, + 253, + 482, + 336 + ], + "type": "text", + "content": " is the sequence length of question-answer prompts and " + }, + { + "bbox": [ + 130, + 253, + 482, + 336 + ], + "type": "inline_equation", + "content": "d_q" + }, + { + "bbox": [ + 130, + 253, + 482, + 336 + ], + "type": "text", + "content": " is the hidden size of the prompt encoder." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 336, + 482, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 336, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 131, + 336, + 482, + 433 + ], + "type": "text", + "content": "Visual Gated Fusion. Since the output " + }, + { + "bbox": [ + 131, + 336, + 482, + 433 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_s" + }, + { + "bbox": [ + 131, + 336, + 482, + 433 + ], + "type": "text", + "content": " from the prompt encoder is an unimodal language representation, to fully utilize the explicit information in images, the visual gated fusion adaptively incorporates image information into this representation with the visual embeddings " + }, + { + "bbox": [ + 131, + 336, + 482, + 433 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_v \\in \\mathbb{R}^{n \\times d_v}" + }, + { + "bbox": [ + 131, + 336, + 482, + 433 + ], + "type": "text", + "content": ". Specifically, in Fig. 4, we use a single-head attention mechanism, similar to MM-COT [58], to correlate the language representation with the visual embeddings. We define the query, key and value as " + }, + { + "bbox": [ + 131, + 336, + 482, + 433 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} = \\mathbf{F}_s \\in \\mathbb{R}^{L_s \\times d_q}" + }, + { + "bbox": [ + 131, + 336, + 482, + 433 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 336, + 482, + 433 + ], + "type": "inline_equation", + "content": "\\mathbf{K} = \\mathbf{V} = \\mathrm{MLP}(\\mathbf{E}_v) \\in \\mathbb{R}^{n \\times d_q}" + }, + { + "bbox": [ + 131, + 336, + 482, + 433 + ], + "type": "text", + "content": ", where MLP here is projecting " + }, + { + "bbox": [ + 131, + 336, + 482, + 433 + ], + "type": "inline_equation", + "content": "d_v" + }, + { + "bbox": [ + 131, + 336, + 482, + 433 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 131, + 336, + 482, + 433 + ], + "type": "inline_equation", + "content": "d_q" + }, + { + "bbox": [ + 131, + 336, + 482, + 433 + ], + "type": "text", + "content": ". The attention output " + }, + { + "bbox": [ + 131, + 336, + 482, + 433 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_v^{attn}" + }, + { + "bbox": [ + 131, + 336, + 482, + 433 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 195, + 440, + 481, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 440, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 195, + 440, + 481, + 471 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {v} ^ {\\text {a t t n}} = \\mathbf {Q} + \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {T}}{\\sqrt {d _ {q}}}\\right) \\mathbf {V}, \\quad \\mathbf {F} _ {v} ^ {\\text {a t t n}} \\in \\mathbb {R} ^ {L _ {s} \\times d _ {q}} \\tag {4}", + "image_path": "da462dfb9197c390c3b1cd49221c3532a5a77abf5f5758e6718cd2c3917c586f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 477, + 481, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 477, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 131, + 477, + 481, + 502 + ], + "type": "text", + "content": "Next, we apply the gated fusion mechanism to merge " + }, + { + "bbox": [ + 131, + 477, + 481, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_s" + }, + { + "bbox": [ + 131, + 477, + 481, + 502 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 477, + 481, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_v^{attn}" + }, + { + "bbox": [ + 131, + 477, + 481, + 502 + ], + "type": "text", + "content": ". The fused representation " + }, + { + "bbox": [ + 131, + 477, + 481, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_m \\in \\mathbb{R}^{L_s \\times d_q}" + }, + { + "bbox": [ + 131, + 477, + 481, + 502 + ], + "type": "text", + "content": " is obtained using the following equations:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 233, + 509, + 481, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 509, + 481, + 538 + ], + "spans": [ + { + "bbox": [ + 233, + 509, + 481, + 538 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\lambda = \\operatorname {S i g m o i d} \\left(\\mathbf {F} _ {s} \\mathbf {W} _ {s} + \\mathbf {F} _ {v} ^ {\\text {a t t n}} \\mathbf {W} _ {v}\\right) \\\\ \\mathbf {F} _ {m} = (1 - \\lambda) \\mathbf {F} _ {s} + \\lambda \\mathbf {F} _ {v} ^ {\\text {a t t n}} \\tag {5} \\\\ \\end{array}", + "image_path": "84c7fdd718fc95b9e50ec01254fb1d250ae2ee51ded632da2e3dab68ea369a68.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 545, + 479, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 479, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 479, + 581 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 130, + 545, + 479, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_s" + }, + { + "bbox": [ + 130, + 545, + 479, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 545, + 479, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_v" + }, + { + "bbox": [ + 130, + 545, + 479, + 581 + ], + "type": "text", + "content": " are learnable parameters. The gated fusion allows the model to adaptively blend the information from the language representation and visual embeddings, resulting in the multimodal representations " + }, + { + "bbox": [ + 130, + 545, + 479, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_m" + }, + { + "bbox": [ + 130, + 545, + 479, + 581 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Resampler Decoder. Although " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_m \\in \\mathbb{R}^{L_s \\times d_q}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " in Eq. (5) is the multimodal representation of " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ", we cannot directly concatenate it with " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_v" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " and instruction embeddings as in Eq. (3), due to the long sequence length of " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "L_s" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ", which can result in extra computational burden and distraction to the target question " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ". In light of this, we employ a lightweight resampler decoder inspired by Perceiver [20], to map " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_m \\in \\mathbb{R}^{L_s \\times d_q}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " into smaller-sized prompt embeddings " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_p \\in \\mathbb{R}^{k \\times d_q}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " with fixed length " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "k \\ll L_s" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " (e.g., " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "k = 32" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "L_s \\geq 100" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "type": "text", + "content": "Haibo Wang, Weifeng Ge" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 115, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 115, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 131, + 115, + 482, + 164 + ], + "type": "text", + "content": "Concretely, we additionally introduce the learnable embeddings " + }, + { + "bbox": [ + 131, + 115, + 482, + 164 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_p\\in \\mathbb{R}^{k\\times d_q}" + }, + { + "bbox": [ + 131, + 115, + 482, + 164 + ], + "type": "text", + "content": " to interact with " + }, + { + "bbox": [ + 131, + 115, + 482, + 164 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_m" + }, + { + "bbox": [ + 131, + 115, + 482, + 164 + ], + "type": "text", + "content": ". As in Fig. 4, a cross-attention layer is applied by taking " + }, + { + "bbox": [ + 131, + 115, + 482, + 164 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_p" + }, + { + "bbox": [ + 131, + 115, + 482, + 164 + ], + "type": "text", + "content": " as query, and " + }, + { + "bbox": [ + 131, + 115, + 482, + 164 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_m" + }, + { + "bbox": [ + 131, + 115, + 482, + 164 + ], + "type": "text", + "content": " as key and value, followed with a feed-forward network, to generate prompt embeddings with the multimodal representations from " + }, + { + "bbox": [ + 131, + 115, + 482, + 164 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_m" + }, + { + "bbox": [ + 131, + 115, + 482, + 164 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 231, + 171, + 389, + 183 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 171, + 389, + 183 + ], + "spans": [ + { + "bbox": [ + 231, + 171, + 389, + 183 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {p} = \\text {C r o s s - A t t e n t i o n} (\\mathbf {F} _ {p}, \\mathbf {F} _ {m}) + \\mathbf {F} _ {p}", + "image_path": "d54f7f9dfc9650864a018195925020793b061c2d61c2f52206d558199bbd5fed.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 225, + 180, + 480, + 199 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 180, + 480, + 199 + ], + "spans": [ + { + "bbox": [ + 225, + 180, + 480, + 199 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {p} = \\operatorname {F F N} \\left(\\mathbf {F} _ {p}\\right) + \\mathbf {F} _ {p}, \\mathbf {F} _ {p} = \\operatorname {M L P} \\left(\\mathbf {F} _ {p}\\right) \\tag {6}", + "image_path": "b51053ad99e75a097c63ba2739b43915cee93dd16dcf0a09098f1199dfbfa9ee.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 205, + 384, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 205, + 384, + 217 + ], + "spans": [ + { + "bbox": [ + 131, + 205, + 384, + 217 + ], + "type": "text", + "content": "where MLP projects " + }, + { + "bbox": [ + 131, + 205, + 384, + 217 + ], + "type": "inline_equation", + "content": "d_{q}" + }, + { + "bbox": [ + 131, + 205, + 384, + 217 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 131, + 205, + 384, + 217 + ], + "type": "inline_equation", + "content": "d_{lm}" + }, + { + "bbox": [ + 131, + 205, + 384, + 217 + ], + "type": "text", + "content": " (the hidden size of LLM)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 217, + 481, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 217, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 481, + 266 + ], + "type": "text", + "content": "At last, the final output of " + }, + { + "bbox": [ + 130, + 217, + 481, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_p\\in \\mathbb{R}^{k\\times d_q}" + }, + { + "bbox": [ + 130, + 217, + 481, + 266 + ], + "type": "text", + "content": " will be fed into the LLM together with the question instruction embeddings and visual features as in Eq. (3), to make predictions of answers. The model is trained using the cross-entropy loss function with trainable parameters " + }, + { + "bbox": [ + 130, + 217, + 481, + 266 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 217, + 481, + 266 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 222, + 272, + 481, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 272, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 222, + 272, + 481, + 304 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {V Q A}} = - \\sum_ {t = 1} ^ {L _ {a}} \\log P _ {\\theta} \\left(\\mathcal {A} _ {t} \\mid \\mathcal {A} _ {< t}, \\mathcal {V}, \\mathcal {Q}, \\mathcal {S}\\right) \\tag {7}", + "image_path": "e896c443e798d58f0d0cba38d1184c2e0a5f4f9759a656c155887deac191ccb2.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 310, + 481, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 310, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 131, + 310, + 481, + 335 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 310, + 481, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_t" + }, + { + "bbox": [ + 131, + 310, + 481, + 335 + ], + "type": "text", + "content": " is predicted autoregressively at position " + }, + { + "bbox": [ + 131, + 310, + 481, + 335 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 310, + 481, + 335 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 310, + 481, + 335 + ], + "type": "inline_equation", + "content": "L_{a}" + }, + { + "bbox": [ + 131, + 310, + 481, + 335 + ], + "type": "text", + "content": " is the sequence length of the ground truth answer text " + }, + { + "bbox": [ + 131, + 310, + 481, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 131, + 310, + 481, + 335 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 350, + 230, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 350, + 230, + 364 + ], + "spans": [ + { + "bbox": [ + 131, + 350, + 230, + 364 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 374, + 201, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 374, + 201, + 385 + ], + "spans": [ + { + "bbox": [ + 131, + 374, + 201, + 385 + ], + "type": "text", + "content": "4.1 Dataset" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 392, + 482, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 513 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 513 + ], + "type": "text", + "content": "We evaluate Q&A Prompts on the challenging A-OKVQA [42] and OK-VQA [37] since they are currently the largest VQA datasets requiring complex reasoning over diverse world knowledge. Specifically, the A-OKVQA dataset consists of 24,903 samples, with 17.1k samples for training, 1.1k for validation, and 6.7k for testing. The questions in this dataset require reasoning over various types of world knowledge, such as commonsense knowledge, factoid knowledge, and visual knowledge. The OK-VQA dataset includes 14k questions covering a variety of knowledge categories, with 9k samples for training, and 5k for validation. Each sample in these two datasets includes an image, a question, and 10 ground-truth answers. We use the soft accuracy [3] as the standard evaluation metric." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 528, + 279, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 528, + 279, + 540 + ], + "spans": [ + { + "bbox": [ + 131, + 528, + 279, + 540 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "We use InstructBLIP [6] as our default MLLM, to both generate Q&A pairs and perform answer predictions. RAM [57] is utilized as the image tagging model due to its strong recognition ability. The number of Q&A pairs is set to " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "P = 8" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ". To configure the visual-aware prompting module, we use the text encoder in CLIP [40] as our prompt encoder, and a value of 32 of the " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " for the number of prompt embeddings " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_p \\in \\mathbb{R}^{k \\times d_q}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ", the same as the number of visual tokens " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_v \\in \\mathbb{R}^{k \\times d_q}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ". We train the models using AdamW [31] as the optimizer with a learning rate of " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "2e^{-5}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " and the strategy of mixed precision training. It is important to note that the parameters of the image encoder, the large language model, and the prompt encoder are kept frozen to maintain their pre-trained representations." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 386, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 386, + 91, + 447, + 102 + ], + "type": "text", + "content": "Q&A Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 114, + 476, + 340 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 476, + 340 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 476, + 340 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 476, + 340 + ], + "type": "table", + "html": "
MethodImage RepresentationKnowledge SourceOK-VQA (%)A-OKVQA (%)
ValTest
ClipCap [39]FeaturePretrain-18.115.8
Pythia [21]FeaturePretrain-25.221.9
ViLBERT [32]FeaturePretrain-30.625.9
LXMERT [47]FeaturePretrain-30.725.9
GPV-2 [23]FeaturePretrain-48.640.7
Unified-IO [33]FeaturePretrain54.0-45.2
Flamingo (80B) [2]FeaturePretrain57.8--
BLIP-2 [26]FeaturePretrain59.360.058.7
Muko [60]FeatureDBPedia + ConceptNet29.2--
ConceptBERT [11]FeatureNumberBatch + ConceptNet33.7--
KRISP [36]FeatureWikipedia + ConceptNet38.933.722.1
MAVEx [52]FeatureWikipedia + ConceptNet40.3--
UniER [16]FeatureConceptNet42.1--
TRiG [9]Caption + Tags + OCRWikipedia49.4--
REVEAL [18]FeatureWIT + Wikidata59.1-52.2
RA-VQA-v2 [28]FeatureGoogle Search62.1--
PICa [53]Caption + TagsFrozen GPT-3 (175B)48.0--
KAT [14]Caption + FeatureWikidata + GPT3 (175B)53.1--
Img2LLM [15]CaptionFrozen GPT-3 (175B)-42.940.7
REVIVE [29]Caption + FeatureWikipedia + GPT3 (175B)56.6--
PromptCap [17]CaptionFrozen GPT-3 (175B)60.456.359.6
Prophet [44]Caption + TagsFrozen GPT-3 (175B)61.158.261.1
InstructBLIP (7B) [6] + Q&A PromptsFeaturePretrain62.164.062.1
FeaturePretrain64.3 (+2.2)69.4 (+5.4)68.1 (+6.0)
", + "image_path": "1b5f54cb964ad6f8f992a798cd0426f8218478c76cda6a5845e481f6566e472e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 406, + 228, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 406, + 228, + 417 + ], + "spans": [ + { + "bbox": [ + 132, + 406, + 228, + 417 + ], + "type": "text", + "content": "4.3 Main Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 426, + 316, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 316, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 316, + 582 + ], + "type": "text", + "content": "Compariosn with SoTAs. The results in Tab. 1 demonstrate the superiority of our proposed Q&A Prompts. We have a significant improvement over existing SoTA methods, achieving an accuracy of " + }, + { + "bbox": [ + 130, + 426, + 316, + 582 + ], + "type": "inline_equation", + "content": "69.4\\%" + }, + { + "bbox": [ + 130, + 426, + 316, + 582 + ], + "type": "text", + "content": " on the validation set and " + }, + { + "bbox": [ + 130, + 426, + 316, + 582 + ], + "type": "inline_equation", + "content": "68.1\\%" + }, + { + "bbox": [ + 130, + 426, + 316, + 582 + ], + "type": "text", + "content": " on the test set in A-OKVQA, and an accuracy of " + }, + { + "bbox": [ + 130, + 426, + 316, + 582 + ], + "type": "inline_equation", + "content": "64.3\\%" + }, + { + "bbox": [ + 130, + 426, + 316, + 582 + ], + "type": "text", + "content": " in OK-VQA. We also observe that Q&A Prompts achieves a greater improvement on A-OKVQA " + }, + { + "bbox": [ + 130, + 426, + 316, + 582 + ], + "type": "inline_equation", + "content": "(+5.4\\%)" + }, + { + "bbox": [ + 130, + 426, + 316, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 426, + 316, + 582 + ], + "type": "inline_equation", + "content": "+6.0\\%" + }, + { + "bbox": [ + 130, + 426, + 316, + 582 + ], + "type": "text", + "content": " compared to OKVQA " + }, + { + "bbox": [ + 130, + 426, + 316, + 582 + ], + "type": "inline_equation", + "content": "(+2.2\\%)" + }, + { + "bbox": [ + 130, + 426, + 316, + 582 + ], + "type": "text", + "content": ". This can be explained that: different from questions in OK-VQA" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "that heavily rely on external knowledge, the questions in A-OKVQA emphasize visually-grounded reasoning with a comprehensive understanding of the image with diverse visual clues, rather than external knowledge retrieval. This aligns with Q&A Prompts's philosophy, that is, instead of relying on accessing any external knowledge base such as Wikipedia [49], ConceptNet [46], or GPT3 [4], our focus is on enhancing reasoning capabilities by discovering rich visual clues hidden in the image and model." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 326, + 449, + 473, + 522 + ], + "blocks": [ + { + "bbox": [ + 130, + 349, + 480, + 383 + ], + "lines": [ + { + "bbox": [ + 130, + 349, + 480, + 383 + ], + "spans": [ + { + "bbox": [ + 130, + 349, + 480, + 383 + ], + "type": "text", + "content": "Table 1: Results of A-OKVQA [42] and OK-VQA [37] comparing to standard baselines show that our method achieves state-of-the-art performance, outperforming previous methods by a large margin. The best results are highlighted." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 449, + 473, + 522 + ], + "lines": [ + { + "bbox": [ + 326, + 449, + 473, + 522 + ], + "spans": [ + { + "bbox": [ + 326, + 449, + 473, + 522 + ], + "type": "table", + "html": "
ModelsA-OKVQA (%)OKVQA (%)
MiniGPT-454.252.5
+ Q&A Prompts61.2 (+7.0)55.9 (+3.4)
LLaVA-1.566.963.3
+ Q&A Prompts72.4 (+5.5)65.2 (+1.9)
InstructBLIP (T5)58.555.1
+ Q&A Prompts64.2 (+5.7)57.3 (+2.2)
", + "image_path": "27682390f75480912a17fd91d28ff51e4a0b0f0b1b030c60e15c7f8255503fc7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 321, + 530, + 482, + 552 + ], + "lines": [ + { + "bbox": [ + 321, + 530, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 321, + 530, + 482, + 552 + ], + "type": "text", + "content": "Table 2: Comparison of performance with different MLLMs as baselines." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 270, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 270, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 270, + 102 + ], + "type": "text", + "content": "Haibo Wang, Weifeng Ge" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 114, + 219, + 167 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 219, + 167 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 219, + 167 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 219, + 167 + ], + "type": "image", + "image_path": "907393e7f0ff67c8c5b3a99e94ba24a7a4c13187fe4152a63eba77bac49448ac.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 168, + 212, + 191 + ], + "lines": [ + { + "bbox": [ + 138, + 168, + 212, + 191 + ], + "spans": [ + { + "bbox": [ + 138, + 168, + 212, + 191 + ], + "type": "text", + "content": "Question: Which room in the house is this man in?" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 376, + 482, + 409 + ], + "lines": [ + { + "bbox": [ + 131, + 376, + 482, + 409 + ], + "spans": [ + { + "bbox": [ + 131, + 376, + 482, + 409 + ], + "type": "text", + "content": "Fig. 5: Representative cases with our Q&A Prompts. We denote the question, tags, Q&A prompts, predictive answers, and ground truth respectively. Our Q&A Prompts can discover rich visual clues in the image and reason out the correct answer." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 223, + 114, + 306, + 167 + ], + "blocks": [ + { + "bbox": [ + 223, + 114, + 306, + 167 + ], + "lines": [ + { + "bbox": [ + 223, + 114, + 306, + 167 + ], + "spans": [ + { + "bbox": [ + 223, + 114, + 306, + 167 + ], + "type": "image", + "image_path": "ca3702426d46d23bd5aec70da7912dfabff42387c528a2d5005bd84e3e7822ad.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 168, + 296, + 193 + ], + "lines": [ + { + "bbox": [ + 225, + 168, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 225, + 168, + 296, + 193 + ], + "type": "text", + "content": "Question: Why are the men's vests orange?" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 138, + 196, + 217, + 235 + ], + "lines": [ + { + "bbox": [ + 138, + 196, + 217, + 235 + ], + "spans": [ + { + "bbox": [ + 138, + 196, + 217, + 235 + ], + "type": "text", + "content": "Image Tags: mirror, sink, reflection, tile wall, tie, bathroom accessory, stare, faucet, cabinet" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 225, + 195, + 298, + 236 + ], + "lines": [ + { + "bbox": [ + 225, + 195, + 298, + 236 + ], + "spans": [ + { + "bbox": [ + 225, + 195, + 298, + 236 + ], + "type": "text", + "content": "Image Tags: constructor, train, safety vest, train track, passenger, railroad, railway station, platform" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 392, + 167 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 392, + 167 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 392, + 167 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 392, + 167 + ], + "type": "image", + "image_path": "bf5b532ef57dd2db96cd18d3cf7ba95d260b5f0dd3f0bdae332145382b00395e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 168, + 384, + 193 + ], + "lines": [ + { + "bbox": [ + 311, + 168, + 384, + 193 + ], + "spans": [ + { + "bbox": [ + 311, + 168, + 384, + 193 + ], + "type": "text", + "content": "Question: What problem will the people on the ferry face?" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 195, + 389, + 228 + ], + "lines": [ + { + "bbox": [ + 310, + 195, + 389, + 228 + ], + "spans": [ + { + "bbox": [ + 310, + 195, + 389, + 228 + ], + "type": "text", + "content": "Image Tags: \ncloudy, travel, ferry, barge, sail, waterway, boat, vessel, float, water, city" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 395, + 114, + 477, + 167 + ], + "blocks": [ + { + "bbox": [ + 395, + 114, + 477, + 167 + ], + "lines": [ + { + "bbox": [ + 395, + 114, + 477, + 167 + ], + "spans": [ + { + "bbox": [ + 395, + 114, + 477, + 167 + ], + "type": "image", + "image_path": "f0fe938da132948b3fbb05fc7e21b4fd670a9098fcd5e45fc6e8efa7ecbd534c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 396, + 168, + 470, + 193 + ], + "lines": [ + { + "bbox": [ + 396, + 168, + 470, + 193 + ], + "spans": [ + { + "bbox": [ + 396, + 168, + 470, + 193 + ], + "type": "text", + "content": "Question: Why are they holding hands in front of people?" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 396, + 195, + 462, + 235 + ], + "lines": [ + { + "bbox": [ + 396, + 195, + 462, + 235 + ], + "spans": [ + { + "bbox": [ + 396, + 195, + 462, + 235 + ], + "type": "text", + "content": "Image Tags: \nbible, veil, hand, couple \ngroom, ceremony, \nwedding, father, bride, \nnewlywed, church" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 239, + 184, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 184, + 247 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 184, + 247 + ], + "type": "text", + "content": "Q&A Prompts:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 247, + 217, + 280 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 247, + 217, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 247, + 217, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 247, + 217, + 262 + ], + "type": "text", + "content": "Q: What is the man looking at? A: mirror" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 263, + 214, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 263, + 214, + 280 + ], + "spans": [ + { + "bbox": [ + 138, + 263, + 214, + 280 + ], + "type": "text", + "content": "Q: What is the man washing hands at? A: sink" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 138, + 286, + 218, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 286, + 218, + 312 + ], + "spans": [ + { + "bbox": [ + 138, + 286, + 218, + 312 + ], + "type": "text", + "content": "Q: Where are the bathroom accessory placed? A: cabinet" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 225, + 239, + 271, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 239, + 271, + 247 + ], + "spans": [ + { + "bbox": [ + 225, + 239, + 271, + 247 + ], + "type": "text", + "content": "Q&A Prompts:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 225, + 247, + 305, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 247, + 305, + 279 + ], + "spans": [ + { + "bbox": [ + 225, + 247, + 305, + 279 + ], + "type": "text", + "content": "Q: What's the man in orange vest? A: constructor Q: What is the yellow and black vehicle? A: train" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 225, + 287, + 301, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 287, + 301, + 312 + ], + "spans": [ + { + "bbox": [ + 225, + 287, + 301, + 312 + ], + "type": "text", + "content": "Q: What is the area where the train is located? A: platform" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 311, + 239, + 356, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 239, + 356, + 247 + ], + "spans": [ + { + "bbox": [ + 311, + 239, + 356, + 247 + ], + "type": "text", + "content": "Q&A Prompts:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 311, + 247, + 391, + 287 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 311, + 247, + 389, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 247, + 389, + 263 + ], + "spans": [ + { + "bbox": [ + 311, + 247, + 389, + 263 + ], + "type": "text", + "content": "Q: What type of weather is it? A: cloudy" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 311, + 263, + 391, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 263, + 391, + 287 + ], + "spans": [ + { + "bbox": [ + 311, + 263, + 391, + 287 + ], + "type": "text", + "content": "Q: What is the purpose of the orange and white boat? A: travel" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 311, + 293, + 384, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 293, + 384, + 312 + ], + "spans": [ + { + "bbox": [ + 311, + 293, + 384, + 312 + ], + "type": "text", + "content": "Q: Where are these boats in? A: city" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 397, + 239, + 441, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 239, + 441, + 247 + ], + "spans": [ + { + "bbox": [ + 397, + 239, + 441, + 247 + ], + "type": "text", + "content": "Q&A Prompts:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 396, + 247, + 475, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 247, + 475, + 287 + ], + "spans": [ + { + "bbox": [ + 396, + 247, + 475, + 287 + ], + "type": "text", + "content": "Q: What is the man in black suit holding? A: bible \nQ: What is the woman wearing on her head? \nA: veil" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 396, + 294, + 474, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 294, + 474, + 312 + ], + "spans": [ + { + "bbox": [ + 396, + 294, + 474, + 312 + ], + "type": "text", + "content": "Q: What kind of building is this? A: church" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 138, + 317, + 206, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 317, + 206, + 365 + ], + "spans": [ + { + "bbox": [ + 138, + 317, + 206, + 365 + ], + "type": "text", + "content": "Predictive Answer: \nInstructBLIP: bedroom \nBLIP-2: bedroom \nOurs: bathroom \nGround Truth: bathroom" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 225, + 316, + 288, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 316, + 288, + 365 + ], + "spans": [ + { + "bbox": [ + 225, + 316, + 288, + 365 + ], + "type": "text", + "content": "Predictive Answer: \nInstructBLIP: fashion \nBLIP-2: working \nOurs: visibility \nGround Truth: visibility" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 311, + 317, + 377, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 317, + 377, + 365 + ], + "spans": [ + { + "bbox": [ + 311, + 317, + 377, + 365 + ], + "type": "text", + "content": "Predictive Answer: \nInstructBLIP: sickness \nBLIP-2: fall into water \nOurs: raining \nGround Truth: \nraining" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 396, + 317, + 460, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 317, + 460, + 365 + ], + "spans": [ + { + "bbox": [ + 396, + 317, + 460, + 365 + ], + "type": "text", + "content": "Predictive Answer: \nInstructBLIP: praying \nBLIP-2: praying \nOurs: getting married \nGround Truth: \ngetting married" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 130, + 433, + 489, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 489, + 505 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 489, + 505 + ], + "type": "text", + "content": "Different baselines. To further validate the effectiveness of our method, we conduct experiments to compare various MLLMs as baseline models, including MiniGPT-4 [59], LLaVA-1.5 [30] and InstructBLIP (FLAN-T5-XL). Tab. 2 shows our Q&A Prompts method consistently improves performance on A-OKVQA by approximately " + }, + { + "bbox": [ + 130, + 433, + 489, + 505 + ], + "type": "inline_equation", + "content": "5.5\\%" + }, + { + "bbox": [ + 130, + 433, + 489, + 505 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 433, + 489, + 505 + ], + "type": "inline_equation", + "content": "7.0\\%" + }, + { + "bbox": [ + 130, + 433, + 489, + 505 + ], + "type": "text", + "content": " and OKVQA by " + }, + { + "bbox": [ + 130, + 433, + 489, + 505 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 130, + 433, + 489, + 505 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 433, + 489, + 505 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 130, + 433, + 489, + 505 + ], + "type": "text", + "content": ", demonstrating the generalizability of Q&A Prompts across different vision-language models." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 130, + 506, + 485, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 485, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 485, + 601 + ], + "type": "text", + "content": "Qualitative results. We present qualitative results in Fig. 5. These cases demonstrate scenarios where correct answers are only possible with Q&A prompts. For instance, the important clues provided by Q&A Prompts of the [mirror], [constructor], [cloudy] and [bible], lead to the successful reasoning of correct answers [bathroom], [visibility], [raining], and [getting married] respectively. This showcases the valuable contribution of the Q&A prompts in certain types of tasks, which can explicitly mine rich visual clues and diverse world knowledge hidden in the image to help the model reason out correct answers." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 131, + 620, + 346, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 620, + 346, + 632 + ], + "spans": [ + { + "bbox": [ + 131, + 620, + 346, + 632 + ], + "type": "text", + "content": "4.4 Analysis of the Generated Questions" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "To systematically evaluate the questions, we leverage GPT-4V to measure their reasonableness. Specifically, we randomly sample 250 " + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "type": "text", + "content": "Q&A Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 122, + 190, + 163 + ], + "blocks": [ + { + "bbox": [ + 169, + 114, + 216, + 121 + ], + "lines": [ + { + "bbox": [ + 169, + 114, + 216, + 121 + ], + "spans": [ + { + "bbox": [ + 169, + 114, + 216, + 121 + ], + "type": "text", + "content": "Visual Knowledge" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 134, + 122, + 190, + 163 + ], + "lines": [ + { + "bbox": [ + 134, + 122, + 190, + 163 + ], + "spans": [ + { + "bbox": [ + 134, + 122, + 190, + 163 + ], + "type": "image", + "image_path": "86335532cabda9b9d1bc74978322bb9e4073ee5fd2a8dcf6a9fbf4ec13c2ad1d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 134, + 164, + 181, + 184 + ], + "lines": [ + { + "bbox": [ + 134, + 164, + 181, + 184 + ], + "spans": [ + { + "bbox": [ + 134, + 164, + 181, + 184 + ], + "type": "text", + "content": "Q: Where is the dog sitting? A: floor" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 198, + 122, + 253, + 163 + ], + "blocks": [ + { + "bbox": [ + 198, + 122, + 253, + 163 + ], + "lines": [ + { + "bbox": [ + 198, + 122, + 253, + 163 + ], + "spans": [ + { + "bbox": [ + 198, + 122, + 253, + 163 + ], + "type": "image", + "image_path": "542adeaf0e3fa05ccff84ac10e717e8dc0ccc5952fa8c029f670c137e0e71cc5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 164, + 246, + 184 + ], + "lines": [ + { + "bbox": [ + 198, + 164, + 246, + 184 + ], + "spans": [ + { + "bbox": [ + 198, + 164, + 246, + 184 + ], + "type": "text", + "content": "Q: What is in the middle of the plate? A: meat" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 261, + 121, + 316, + 163 + ], + "blocks": [ + { + "bbox": [ + 290, + 114, + 351, + 121 + ], + "lines": [ + { + "bbox": [ + 290, + 114, + 351, + 121 + ], + "spans": [ + { + "bbox": [ + 290, + 114, + 351, + 121 + ], + "type": "text", + "content": "Commonsien Knowledge" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 261, + 121, + 316, + 163 + ], + "lines": [ + { + "bbox": [ + 261, + 121, + 316, + 163 + ], + "spans": [ + { + "bbox": [ + 261, + 121, + 316, + 163 + ], + "type": "image", + "image_path": "8ee66fab8354b9c061e66c491d3dcf1e1b714bc25e1e527f0e13e19009570023.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 163, + 306, + 183 + ], + "lines": [ + { + "bbox": [ + 261, + 163, + 306, + 183 + ], + "spans": [ + { + "bbox": [ + 261, + 163, + 306, + 183 + ], + "type": "text", + "content": "Q: Why are people carrying the bags? A: load" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 324, + 121, + 380, + 163 + ], + "blocks": [ + { + "bbox": [ + 324, + 121, + 380, + 163 + ], + "lines": [ + { + "bbox": [ + 324, + 121, + 380, + 163 + ], + "spans": [ + { + "bbox": [ + 324, + 121, + 380, + 163 + ], + "type": "image", + "image_path": "28ffe4210dc656a9e678ac8ecd2ae6e58fc6da7dc5a0cee7e23580514e801dc6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 163, + 378, + 183 + ], + "lines": [ + { + "bbox": [ + 324, + 163, + 378, + 183 + ], + "spans": [ + { + "bbox": [ + 324, + 163, + 378, + 183 + ], + "type": "text", + "content": "Q: What is the glass covered area used for? A: showcase" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 388, + 140, + 481, + 232 + ], + "blocks": [ + { + "bbox": [ + 394, + 128, + 476, + 137 + ], + "lines": [ + { + "bbox": [ + 394, + 128, + 476, + 137 + ], + "spans": [ + { + "bbox": [ + 394, + 128, + 476, + 137 + ], + "type": "text", + "content": "Required Knowledge Types" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 388, + 140, + 481, + 232 + ], + "lines": [ + { + "bbox": [ + 388, + 140, + 481, + 232 + ], + "spans": [ + { + "bbox": [ + 388, + 140, + 481, + 232 + ], + "type": "image", + "image_path": "4d7b47f7bf1189f59ffbe1fb7a53cc614609adcd940f1a3c4aaaff8c49248a0a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 235, + 378, + 255 + ], + "lines": [ + { + "bbox": [ + 324, + 235, + 378, + 255 + ], + "spans": [ + { + "bbox": [ + 324, + 235, + 378, + 255 + ], + "type": "text", + "content": "Q: How is the woman's face visible? A: reflection" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 134, + 193, + 189, + 234 + ], + "blocks": [ + { + "bbox": [ + 173, + 185, + 214, + 191 + ], + "lines": [ + { + "bbox": [ + 173, + 185, + 214, + 191 + ], + "spans": [ + { + "bbox": [ + 173, + 185, + 214, + 191 + ], + "type": "text", + "content": "Knowledge base" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 134, + 193, + 189, + 234 + ], + "lines": [ + { + "bbox": [ + 134, + 193, + 189, + 234 + ], + "spans": [ + { + "bbox": [ + 134, + 193, + 189, + 234 + ], + "type": "image", + "image_path": "36524fbc4f8abf4b489bfced18244225ea842e999bdc78d6350d504377dac9ea.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 134, + 235, + 183, + 255 + ], + "lines": [ + { + "bbox": [ + 134, + 235, + 183, + 255 + ], + "spans": [ + { + "bbox": [ + 134, + 235, + 183, + 255 + ], + "type": "text", + "content": "Q: What event is this most likely? A: Christmas" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 180, + 263, + 431, + 275 + ], + "lines": [ + { + "bbox": [ + 180, + 263, + 431, + 275 + ], + "spans": [ + { + "bbox": [ + 180, + 263, + 431, + 275 + ], + "type": "text", + "content": "Fig. 6: Generated Q&A pairs with different knowledge types." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 198, + 193, + 253, + 234 + ], + "blocks": [ + { + "bbox": [ + 198, + 193, + 253, + 234 + ], + "lines": [ + { + "bbox": [ + 198, + 193, + 253, + 234 + ], + "spans": [ + { + "bbox": [ + 198, + 193, + 253, + 234 + ], + "type": "image", + "image_path": "1e7dc1283518917169742cf8022fe2d961071456170e2ac7c29a6e1e8a89bb29.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 235, + 242, + 255 + ], + "lines": [ + { + "bbox": [ + 198, + 235, + 242, + 255 + ], + "spans": [ + { + "bbox": [ + 198, + 235, + 242, + 255 + ], + "type": "text", + "content": "Q: What species of bird is this? A: heron" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 261, + 235, + 312, + 255 + ], + "lines": [ + { + "bbox": [ + 261, + 235, + 312, + 255 + ], + "spans": [ + { + "bbox": [ + 261, + 235, + 312, + 255 + ], + "type": "text", + "content": "Q: How do the people feel here? A: cold" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 144, + 285, + 278, + 340 + ], + "blocks": [ + { + "bbox": [ + 144, + 285, + 278, + 340 + ], + "lines": [ + { + "bbox": [ + 144, + 285, + 278, + 340 + ], + "spans": [ + { + "bbox": [ + 144, + 285, + 278, + 340 + ], + "type": "table", + "html": "
Q&A sourceScore
Annotations in A-OKVQA9.73
VQG w/ training8.86
VQG w/o training4.12
", + "image_path": "e3b4dc96022c06c4fe3e18b65a7e3877c686f914efe952adb14ea814f0eb1a4f.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_body" + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 301, + 285, + 473, + 340 + ], + "blocks": [ + { + "bbox": [ + 134, + 348, + 286, + 369 + ], + "lines": [ + { + "bbox": [ + 134, + 348, + 286, + 369 + ], + "spans": [ + { + "bbox": [ + 134, + 348, + 286, + 369 + ], + "type": "text", + "content": "Table 3: The evaluation for reasonableness of different questions." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 301, + 285, + 473, + 340 + ], + "lines": [ + { + "bbox": [ + 301, + 285, + 473, + 340 + ], + "spans": [ + { + "bbox": [ + 301, + 285, + 473, + 340 + ], + "type": "table", + "html": "
Q&ATraining SetRouge-LA-OKVQA (%)OK-VQA (%)
×--64.062.1
✓--61.961.0
✓A-OKVQA90.869.462.9
✓OK-VQA92.467.164.3
✓Mixture89.468.763.9
", + "image_path": "60479a7f0d9d1d9eb4238514f1ab081c317238cbd859ab1ae18b3979ba26164c.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 293, + 348, + 480, + 369 + ], + "lines": [ + { + "bbox": [ + 293, + 348, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 293, + 348, + 480, + 369 + ], + "type": "text", + "content": "Table 4: Ablation studies on training sets for VQG model training." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 392, + 480, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 480, + 487 + ], + "type": "text", + "content": "tuples from the QA pairs generated by our VQG model, and ask GPT-4V to return an integer score of reasonableness ranging from 1 to 10 for them. For comparison, we also evaluate the scores of the manually annotated tuples in the A-OKVQA and those generated by directly prompting MLLM without training (with the same number of 250 tuples). Tab. 3 shows the questions generated by our VQG model (VQG w/ training) have a close quality to the manually annotated questions in A-OKVQA and are more reasonable than directly prompting a frozen MLLM (VQG w/o training). See appendix for more details." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 130, + 487, + 480, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 480, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 480, + 559 + ], + "type": "text", + "content": "We also annotated what kind of world knowledge was required to answer the 250 sampled questions generated by our VQG model. The choices are the same with A-OKVQA including Commonsense Knowledge, Visual Knowledge, Knowledge Bases, and Physical Knowledge (explained in the appendix). The examples and distribution are shown in Fig. 6. Notably, sometimes there is no clear distinction between these categories and a question may belong to either." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 131, + 575, + 238, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 575, + 238, + 588 + ], + "spans": [ + { + "bbox": [ + 131, + 575, + 238, + 588 + ], + "type": "text", + "content": "4.5 Ablation Study" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 130, + 594, + 487, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 487, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 487, + 665 + ], + "type": "text", + "content": "We investigate the role of our question-answer prompts and the visual-aware prompting module, based on InstructBLIP-Vicuna-7B and the A-OKVQA dataset. Different training sets for generated questions. We investigate how the VQG model will influence the performance by directly prompting the MLLM without training, or using different training sets to train our VQG model, including the A-OKVQA, OK-VQA, and a mixture of them. We report the Rouge-L" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 270, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 270, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 270, + 102 + ], + "type": "text", + "content": "Haibo Wang, Weifeng Ge" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 141, + 114, + 261, + 169 + ], + "blocks": [ + { + "bbox": [ + 141, + 114, + 261, + 169 + ], + "lines": [ + { + "bbox": [ + 141, + 114, + 261, + 169 + ], + "spans": [ + { + "bbox": [ + 141, + 114, + 261, + 169 + ], + "type": "table", + "html": "
Fusion strategyAccuracy (%)
Prepend + Q&A66.6
+ Captions64.2
+ Tags64.7
VPM + Q&A69.4
", + "image_path": "cde4595775f80357841fd9c42a6f1248bec208b5f6eb40b16b0837ea9f82c4de.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 280, + 114, + 354, + 169 + ], + "blocks": [ + { + "bbox": [ + 134, + 176, + 268, + 195 + ], + "lines": [ + { + "bbox": [ + 134, + 176, + 268, + 195 + ], + "spans": [ + { + "bbox": [ + 134, + 176, + 268, + 195 + ], + "type": "text", + "content": "Table 5: Ablation studies on strategies of fusing Q&A pairs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 280, + 114, + 354, + 169 + ], + "lines": [ + { + "bbox": [ + 280, + 114, + 354, + 169 + ], + "spans": [ + { + "bbox": [ + 280, + 114, + 354, + 169 + ], + "type": "table", + "html": "
#NumAccuracy (%)
166.2
467.9
869.4
1669.2
", + "image_path": "b973c1b3f5c0063813ecfc6b6284e6e577f6d2c5ca11d64ad72ee45a4028fe1c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 364, + 114, + 466, + 169 + ], + "blocks": [ + { + "bbox": [ + 364, + 114, + 466, + 169 + ], + "lines": [ + { + "bbox": [ + 364, + 114, + 466, + 169 + ], + "spans": [ + { + "bbox": [ + 364, + 114, + 466, + 169 + ], + "type": "table", + "html": "
VariantsAccuracy (%)
VPM69.4
w/ BERT68.6
w/o Fusion68.6
w/o Decoder68.1
", + "image_path": "5ef800d341d556edc684fef826f78eb17562c8d38d8746d7e64fa253fd9ffc6a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 272, + 176, + 473, + 204 + ], + "lines": [ + { + "bbox": [ + 272, + 176, + 473, + 204 + ], + "spans": [ + { + "bbox": [ + 272, + 176, + 473, + 204 + ], + "type": "text", + "content": "Table 6: Ablation studies Table 7: Ablation studies on the on numbers of question-different components in the visual-answer pairs." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 231, + 480, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 231, + 480, + 302 + ], + "spans": [ + { + "bbox": [ + 130, + 231, + 480, + 302 + ], + "type": "text", + "content": "score for the generated questions. Results in Tab. 4 show that performance declines severely when using Q&A prompts generated by MLLMs without training. Besides, the best performance is only achieved when the corresponding dataset is utilized as the training set, and a mismatch or mixture of datasets could lead to a slightly weaker performance. This could be attributed to the domain gap in the question types of these datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 304, + 481, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 481, + 448 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 481, + 448 + ], + "type": "text", + "content": "Strategies of fusing question-answer pairs. In Tab. 5, we explore a simple approach where we directly prepend the Q&A pairs as a natural language to the instruction (denoted as Prepend + Q&A) without encoded by the visual-aware prompting module (VPM). For comparison, we also replace the question-answer pairs with the same number of image captions generated by BLIP-2 [26] and tags recognized by RAM (denoted as Prepend + Captions and Prepend + Tags). The results in Tab. 5 show that the incorporation of question-answer prompts is more effective than pure captions and tags, and both directly prepending the question-answer pairs and using the VPM to encode them can improve performance. Notably, the models with VPM perform much better, which we attributed to its ability to combat some noise in the generated question-answer pairs and avoid interference with the target question." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "text", + "content": "The number of the question-answer prompts. We also study to figure out the effect of using different numbers of question-answer pairs to construct the Q&A Prompts. The results are displayed in Tab. 6. We denote the desired number as " + }, + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "text", + "content": ". It can be observed that when " + }, + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "text", + "content": " is set to 8, the model achieves optimal performance. We can deduce that when " + }, + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "text", + "content": " is too large, it may introduce noisy pairs, while a too small " + }, + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "text", + "content": " can result in an ignorance of many essential and informative question-answer pairs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "content": "Components in visual-aware prompting module. We validate the key components of the visual-aware prompting module in Tab. 7. We first study the effectiveness of using different language models as the prompt encoder, where we replace the text encoder in CLIP with BERT [8] (w/ BERT), which results in a performance drop. This result is consistent with previous findings [38, 45], which suggest that models initialized with CLIP [40] are more suitable for vision-language tasks due to their shared understanding of the visual and textual content. We then remove the visual gated fusion (w/o fusion), which also induces severe performance decline. This is expected since the unimodal language representations tend to involve ambiguity and bias. The additional incorporation of visual embeddings effectively allows for multi-modal interaction beneficial for" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "type": "text", + "content": "Q&A Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 246, + 186 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 246, + 186 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 246, + 186 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 246, + 186 + ], + "type": "image", + "image_path": "889498b47b68ec8e0249ae1a0cb2e73fb7b21b131e44ae8231d5dafae4f4cdfc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 195, + 422, + 206 + ], + "lines": [ + { + "bbox": [ + 189, + 195, + 422, + 206 + ], + "spans": [ + { + "bbox": [ + 189, + 195, + 422, + 206 + ], + "type": "text", + "content": "Fig. 7: Qualitative analysis on how Q&A Prompts work." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 249, + 116, + 290, + 124 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 116, + 290, + 124 + ], + "spans": [ + { + "bbox": [ + 249, + 116, + 290, + 124 + ], + "type": "text", + "content": "Q&A Prompts:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 248, + 125, + 400, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 125, + 400, + 184 + ], + "spans": [ + { + "bbox": [ + 248, + 125, + 400, + 184 + ], + "type": "text", + "content": "Q1: What are the flowing vegetation? A: grass \nQ2: What are the effects of air movement on grass? A: bent \nQ3: What word describes the green grass? A: lush \nQ4: What animal in this image is used for riding? A: horse \nQ5: What is the red machine called? A: tractor \nQ6: Where is the horses standing? A: dirt field \nQ7: How many horses are there? A: couple \nQ8: What type of property is this? A: farm" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 403, + 116, + 477, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 116, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 403, + 116, + 477, + 148 + ], + "type": "text", + "content": "Ground Truth: wind \nPredictive Answer: \nw/ Q&A Prompts: wind \nw/o Q&A Prompts: drought" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 403, + 151, + 467, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 151, + 467, + 183 + ], + "spans": [ + { + "bbox": [ + 403, + 151, + 467, + 183 + ], + "type": "text", + "content": "w/Q&A Prompts (remove Q1, Q2): drought w/Q&A Prompts (remove Q3~Q8): wind" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 228, + 482, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 482, + 312 + ], + "type": "text", + "content": "vision-language reasoning. We also evaluate the importance of the resampler decoder (w/o decoder), where we directly feed the long sequence " + }, + { + "bbox": [ + 130, + 228, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_m\\in \\mathbb{R}^{L_s\\times d_q}" + }, + { + "bbox": [ + 130, + 228, + 482, + 312 + ], + "type": "text", + "content": " into the large language model in Eq. (3) and also caused a performance drop. The sequence length of target question " + }, + { + "bbox": [ + 130, + 228, + 482, + 312 + ], + "type": "inline_equation", + "content": "L_{q}" + }, + { + "bbox": [ + 130, + 228, + 482, + 312 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 130, + 228, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\times 10" + }, + { + "bbox": [ + 130, + 228, + 482, + 312 + ], + "type": "text", + "content": " times smaller than the sequence length " + }, + { + "bbox": [ + 130, + 228, + 482, + 312 + ], + "type": "inline_equation", + "content": "L_{s}" + }, + { + "bbox": [ + 130, + 228, + 482, + 312 + ], + "type": "text", + "content": " of prompts " + }, + { + "bbox": [ + 130, + 228, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 228, + 482, + 312 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 228, + 482, + 312 + ], + "type": "inline_equation", + "content": "L_{q}\\leq 10" + }, + { + "bbox": [ + 130, + 228, + 482, + 312 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 228, + 482, + 312 + ], + "type": "inline_equation", + "content": "L_{s}\\geq 100" + }, + { + "bbox": [ + 130, + 228, + 482, + 312 + ], + "type": "text", + "content": " here), resulting in a distraction to the understanding of the target question. This highlights the importance of resampling " + }, + { + "bbox": [ + 130, + 228, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_m\\in \\mathbb{R}^{L_s\\times d_q}" + }, + { + "bbox": [ + 130, + 228, + 482, + 312 + ], + "type": "text", + "content": " into a shorter " + }, + { + "bbox": [ + 130, + 228, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_p\\in \\mathbb{R}^{k\\times d_q}" + }, + { + "bbox": [ + 130, + 228, + 482, + 312 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 228, + 482, + 312 + ], + "type": "inline_equation", + "content": "k = 32" + }, + { + "bbox": [ + 130, + 228, + 482, + 312 + ], + "type": "text", + "content": " here)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 328, + 443, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 328, + 443, + 342 + ], + "spans": [ + { + "bbox": [ + 131, + 328, + 443, + 342 + ], + "type": "text", + "content": "5 A Further Qualitative Analysis of Q&A Prompts" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 350, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 472 + ], + "type": "text", + "content": "Fig. 7 further reveals why Q&A Prompts are necessary and how Q&A Prompts work in particular case. To find out [what ambient effect] causes the appearance of grass, the key point is to grasp the movement of them. The Q&A pairs " + }, + { + "bbox": [ + 130, + 350, + 482, + 472 + ], + "type": "inline_equation", + "content": "Q_{1}" + }, + { + "bbox": [ + 130, + 350, + 482, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 350, + 482, + 472 + ], + "type": "inline_equation", + "content": "Q_{2}" + }, + { + "bbox": [ + 130, + 350, + 482, + 472 + ], + "type": "text", + "content": " precisely capture the visual clues that [the grass are flowing] and [air movement cause the grass bent], which are beneficial for reasoning out the answer [wind]. We also test scenarios where we remove " + }, + { + "bbox": [ + 130, + 350, + 482, + 472 + ], + "type": "inline_equation", + "content": "Q_{1}" + }, + { + "bbox": [ + 130, + 350, + 482, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 350, + 482, + 472 + ], + "type": "inline_equation", + "content": "Q_{2}" + }, + { + "bbox": [ + 130, + 350, + 482, + 472 + ], + "type": "text", + "content": " or only reserve them, leading to the incorrect answer [drought] and correct answer [wind] respectively. This indicates that the quality of Q&A Prompts is crucial, particularly when the question needs multiple reasoning steps beyond perception, and Q&A Prompts serve as the intermediate steps that fill the logical gap." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 487, + 322, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 487, + 322, + 499 + ], + "spans": [ + { + "bbox": [ + 131, + 487, + 322, + 499 + ], + "type": "text", + "content": "6 Conclusions and Limitations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "This paper has studied the problem of VQA tasks requiring reasoning over diverse world knowledge. We introduce a novel framework, Q&A Prompts, which effectively generates a set of question-answer prompts and encodes them with a visual-aware prompting module, significantly boosting the reasoning ability of current multi-modal large language models. We conducted a series of experiments on the A-OKVQA and OK-VQA benchmarks and achieved significant results compared with previous methods. Extensive ablations and comprehensive analyses have demonstrated the effectiveness and superiority of our method. Limitations. While our method shows promising results, it is essential to acknowledge the potential presence of biases in the data as observed in previous VQA studies [1,7], as well as the lack of the ability for fine-grained counting and Optical Character Recognition. In our future work, we plan to mitigate these biases and weaknesses to enhance the reasoning ability of current models further." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 270, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 270, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 270, + 102 + ], + "type": "text", + "content": "Haibo Wang, Weifeng Ge" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 175 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China under Grant Nos. 62106051 and the National Key R&D Program of China 2022YFC3601405." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 217, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 217, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 481, + 251 + ], + "type": "text", + "content": "1. Agrawal, A., Batra, D., Parikh, D., Kembhavi, A.: Don't just assume; look and answer: Overcoming priors for visual question answering. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018) 14" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 251, + 481, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 481, + 283 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 481, + 283 + ], + "type": "text", + "content": "2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: Adv. Neural Inform. Process. Syst. (2022) 4, 10" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 284, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 284, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 284, + 481, + 304 + ], + "type": "text", + "content": "3. Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Zitnick, C.L., Parikh, D.: Vqa: visual question answering. In: Int. Conf. Comput. Vis. (2015) 2, 3, 4, 9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 305, + 481, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 305, + 481, + 338 + ], + "spans": [ + { + "bbox": [ + 138, + 305, + 481, + 338 + ], + "type": "text", + "content": "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: Adv. Neural Inform. Process. Syst. (2020) 4, 10" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 338, + 481, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 338, + 481, + 370 + ], + "spans": [ + { + "bbox": [ + 138, + 338, + 481, + 370 + ], + "type": "text", + "content": "5. Changpinyo, S., Sharma, P., Ding, N., Soricut, R.: Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 371, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 371, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 371, + 481, + 403 + ], + "type": "text", + "content": "6. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500 (2023) 2, 3, 4, 5, 6, 8, 9, 10" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 403, + 481, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 481, + 436 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 481, + 436 + ], + "type": "text", + "content": "7. Dancette, C., Cadene, R., Teney, D., Cord, M.: Beyond question-based biases: Assessing multimodal shortcut learning in visual question answering. In: Int. Conf. Comput. Vis. (2021) 2, 14" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 437, + 481, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 437, + 481, + 458 + ], + "spans": [ + { + "bbox": [ + 138, + 437, + 481, + 458 + ], + "type": "text", + "content": "8. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: NAACL (2019) 13" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 459, + 481, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 459, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 459, + 481, + 491 + ], + "type": "text", + "content": "9. Gao, F., Ping, Q., Thattai, G., Reganti, A., Wu, Y.N., Natarajan, P.: Transform-retrieve-produce: Natural language-centric outside-knowledge visual question answering. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) 10" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 491, + 481, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 481, + 523 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 481, + 523 + ], + "type": "text", + "content": "0. Gao, L., Biderman, S., Black, S., Golding, L., Hoppe, T., Foster, C., Phang, J., He, H., Thite, A., Nabeshima, N., et al.: The pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027 (2020) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 524, + 481, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 524, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 138, + 524, + 481, + 545 + ], + "type": "text", + "content": "1. Gardères, F., Ziaeffard, M., Abeloos, B., Lecue, F.: ConceptBert: Concept-aware representation for visual question answering. In: EMNLP (2020) 10" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 545, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 481, + 578 + ], + "type": "text", + "content": "2. Geman, D., Geman, S., Hallonquist, N., Younes, L.: Visual Turing test for computer vision systems. Proceedings of the National Academy of Sciences 112(12), 3618-3623 (2015) 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 578, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 481, + 610 + ], + "type": "text", + "content": "3. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In: IEEE Conf. Comput. Vis. Pattern Recog. (2017) 2, 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "type": "text", + "content": "4. Gui, L., Wang, B., Huang, Q., Hauptmann, A., Bisk, Y., Gao, J.: Kat: A knowledge augmented transformer for vision-and-language. In: NAACL (2022) 2, 4, 10" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "text", + "content": "5. Guo, J., Li, J., Li, D., Tiong, A.M.H., Li, B., Tao, D., Hoi, S.C.: From images to textual prompts: Zero-shot vqa with frozen large language models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) 4, 10" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 386, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 386, + 91, + 447, + 102 + ], + "type": "text", + "content": "Q&A Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "16. Guo, Y., Nie, L., Wong, Y., Liu, Y., Cheng, Z., Kankanhalli, M.: A unified end-to-end retriever-reader framework for knowledge-based vqa. In: ACM Int. Conf. Multimedia (2022) 10" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 180 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 180 + ], + "type": "text", + "content": "17. Hu, Y., Hua, H., Yang, Z., Shi, W., Smith, N.A., Luo, J.: Promptcap: Prompt-guided task-aware image captioning. arXiv preprint arXiv:2211.09699 (2022) 4, 10" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 180, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 180, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 180, + 480, + 224 + ], + "type": "text", + "content": "18. Hu, Z., Iscen, A., Sun, C., Wang, Z., Chang, K.W., Sun, Y., Schmid, C., Ross, D.A., Fathi, A.: **Retrieval-augmented visual-language pre-training with multi-source multimodal knowledge memory. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) 2, 4, 10" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 224, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 224, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 133, + 224, + 480, + 257 + ], + "type": "text", + "content": "19. Huang, X., Zhang, Y., Ma, J., Tian, W., Feng, R., Zhang, Y., Li, Y., Guo, Y., Zhang, L.: Tag2text: Guiding vision-language model via image tagging. arXiv preprint arXiv:2303.05657 (2023) 3, 4, 5, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 257, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 257, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 133, + 257, + 480, + 289 + ], + "type": "text", + "content": "20. Jaegle, A., Borgeaud, S., Alayrac, J.B., Doersch, C., Ionescu, C., Ding, D., Koppula, S., Zoran, D., Brock, A., Shelhamer, E., et al.: Perceiver io: A general architecture for structured inputs & outputs. In: Int. Conf. Learn. Represent. (2022) 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 289, + 480, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 289, + 480, + 321 + ], + "spans": [ + { + "bbox": [ + 133, + 289, + 480, + 321 + ], + "type": "text", + "content": "21. Jiang, Y., Natarajan, V., Chen, X., Rohrbach, M., Batra, D., Parikh, D.: Pythia v0. 1: The winning entry to the vqa challenge 2018. arxiv 2018. arXiv preprint arXiv:1807.09956 (2018) 10" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 321, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 321, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 133, + 321, + 480, + 365 + ], + "type": "text", + "content": "22. Johnson, J., Hariharan, B., Van Der Maaten, L., Fei-Fei, L., Lawrence Zitnick, C., Girshick, R.: Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2901-2910 (2017) 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 365, + 480, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 365, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 133, + 365, + 480, + 397 + ], + "type": "text", + "content": "23. Kamath, A., Clark, C., Gupta, T., Kolve, E., Hoiem, D., Kembhavi, A.: Webly supervised concept expansion for general purpose vision models. In: Eur. Conf. Comput. Vis. (2022) 10" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 397, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 397, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 133, + 397, + 480, + 429 + ], + "type": "text", + "content": "24. Khan, Z., BG, V.K., Schulter, S., Yu, X., Fu, Y., Chandraker, M.: Q: How to specialize large vision-language models to data-scarce vqa tasks? a: Self-train on unlabeled images! In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 429, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 429, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 133, + 429, + 480, + 460 + ], + "type": "text", + "content": "25. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 460, + 480, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 460, + 480, + 494 + ], + "spans": [ + { + "bbox": [ + 133, + 460, + 480, + 494 + ], + "type": "text", + "content": "26. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. In: Int. Conf. Mach. Learn. (2023) 2, 4, 5, 10, 13" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 494, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 494, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 133, + 494, + 480, + 525 + ], + "type": "text", + "content": "27. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: common objects in context. In: Eur. Conf. Comput. Vis. (2014) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 525, + 480, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 525, + 480, + 558 + ], + "spans": [ + { + "bbox": [ + 133, + 525, + 480, + 558 + ], + "type": "text", + "content": "28. Lin, W., Chen, J., Mei, J., Coca, A., Byrne, B.: Fine-grained late-interaction multimodal retrieval for retrieval augmented visual question answering. In: Adv. Neural Inform. Process. Syst. (2023) 10" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 558, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 558, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 133, + 558, + 480, + 590 + ], + "type": "text", + "content": "29. Lin, Y., Xie, Y., Chen, D., Xu, Y., Zhu, C., Yuan, L.: Revive: Regional visual representation matters in knowledge-based visual question answering. In: Adv. Neural Inform. Process. Syst. (2022) 2, 4, 10" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 590, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 590, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 133, + 590, + 480, + 611 + ], + "type": "text", + "content": "30. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: Adv. Neural Inform. Process. Syst. (2023) 2, 3, 4, 11" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 611, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 611, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 133, + 611, + 480, + 633 + ], + "type": "text", + "content": "31. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: Int. Conf. Learn. Represent. (2019) 9" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "text", + "content": "32. Lu, J., Batra, D., Parikh, D., Lee, S.: Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In: Adv. Neural Inform. Process. Syst. (2019) 10" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 270, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 270, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 270, + 102 + ], + "type": "text", + "content": "Haibo Wang, Weifeng Ge" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "33. Lu, J., Clark, C., Zellers, R., Mottaghi, R., Kembhavi, A.: Unified-io: A unified model for vision, language, and multi-modal tasks. arXiv preprint arXiv:2206.08916 (2022) 10" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "text", + "content": "34. Luo, G., Zhou, Y., Ren, T., Chen, S., Sun, X., Ji, R.: Cheap and quick: Efficient vision-language instruction tuning for large language models. arXiv preprint arXiv:2305.15023 (2023) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 202 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 202 + ], + "type": "text", + "content": "35. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008) 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 203, + 482, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 203, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 203, + 482, + 236 + ], + "type": "text", + "content": "36. Marino, K., Chen, X., Parikh, D., Gupta, A., Rohrbach, M.: Krisp: Integrating implicit and symbolic knowledge for open-domain knowledge-based vqa. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021) 10" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 236, + 482, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 268 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 268 + ], + "type": "text", + "content": "37. Marino, K., Rastegari, M., Farhadi, A., Mottaghi, R.: Ok-vqa: A visual question answering benchmark requiring external knowledge. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019) 2, 3, 4, 5, 9, 10" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 268, + 482, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 482, + 289 + ], + "type": "text", + "content": "38. Merullo, J., Castricato, L., Eickhoff, C., Pavlick, E.: Linearly mapping from image to text space. In: Int. Conf. Learn. Represent. (2023) 13" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 289, + 482, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 311 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 311 + ], + "type": "text", + "content": "39. Mokady, R., Hertz, A., Bermano, A.H.: Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734 (2021) 10" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 311, + 482, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 482, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 482, + 343 + ], + "type": "text", + "content": "40. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Int. Conf. Mach. Learn. (2021) 9, 13" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 343, + 482, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 364 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 364 + ], + "type": "text", + "content": "41. Reimers, N., Gurevych, I.: Sentence-bert: Sentence embeddings using siamese bert-networks. In: EMNLP (2019) 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 364, + 482, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 482, + 396 + ], + "type": "text", + "content": "42. Schwenk, D., Khandelwal, A., Clark, C., Marino, K., Mottaghi, R.: A-okvqa: A benchmark for visual question answering using world knowledge. In: Eur. Conf. Comput. Vis. (2022) 2, 3, 4, 5, 9, 10" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 396, + 482, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 429 + ], + "type": "text", + "content": "43. Selvaraju, R.R., Tendulkar, P., Parikh, D., Horvitz, E., Ribeiro, M.T., Nushi, B., Kamar, E.: Squinting at vqa models: Introspecting vqa models with sub-questions. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 429, + 482, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 482, + 461 + ], + "type": "text", + "content": "44. Shao, Z., Yu, Z., Wang, M., Yu, J.: Prompting large language models with answer heuristics for knowledge-based visual question answering. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) 4, 10" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 461, + 482, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 494 + ], + "type": "text", + "content": "45. Shen, S., Li, L.H., Tan, H., Bansal, M., Rohrbach, A., Chang, K.W., Yao, Z., Keutzer, K.: How much can clip benefit vision-and-language tasks? In: Int. Conf. Learn. Represent. (2022) 13" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 494, + 482, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 482, + 515 + ], + "type": "text", + "content": "46. Speer, R., Chin, J., Havasi, C.: Conceptnet 5.5: An open multilingual graph of general knowledge. In: AAAI (2017) 4, 10" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 515, + 482, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 515, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 515, + 482, + 536 + ], + "type": "text", + "content": "47. Tan, H., Bansal, M.: Lxmert: Learning cross-modality encoder representations from transformers. In: EMNLP (2019) 10" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 536, + 482, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 482, + 557 + ], + "type": "text", + "content": "48. Uehara, K., Duan, N., Harada, T.: Learning to ask informative sub-questions for visual question answering. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 557, + 482, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 578 + ], + "type": "text", + "content": "49. Vrandecic, D., Krötzsch, M.: Wikidata: a free collaborative knowledgebase. Communications of the ACM 57(10), 78-85 (2014) 4, 10" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 578, + 482, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 578, + 482, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 578, + 482, + 611 + ], + "type": "text", + "content": "50. Wang, P., Wu, Q., Shen, C., Dick, A., Van Den Hengel, A.: Fvqa: Fact-based visual question answering. IEEE transactions on pattern analysis and machine intelligence 40(10), 2413-2427 (2017) 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 611, + 482, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 482, + 643 + ], + "type": "text", + "content": "51. Wang, P., Wu, Q., Shen, C., Hengel, A.v.d., Dick, A.: Explicit knowledge-based reasoning for visual question answering. arXiv preprint arXiv:1511.02570 (2015) 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 643, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 643, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 643, + 482, + 665 + ], + "type": "text", + "content": "52. Wu, J., Lu, J., Sabharwal, A., Mottaghi, R.: Multi-modal answer validation for knowledge-based vqa. In: AAAI (2021) 10" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "type": "text", + "content": "Q&A Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 347 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "53. Yang, Z., Gan, Z., Wang, J., Hu, X., Lu, Y., Liu, Z., Wang, L.: An empirical study of gpt-3 for few-shot knowledge-based vqa. In: AAAI (2022) 10" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "type": "text", + "content": "54. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023) 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "text", + "content": "55. Zellers, R., Bisk, Y., Farhadi, A., Choi, Y.: From recognition to cognition: Visual commonsense reasoning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019) 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "text", + "content": "56. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199 (2023) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "57. Zhang, Y., Huang, X., Ma, J., Li, Z., Luo, Z., Xie, Y., Qin, Y., Luo, T., Li, Y., Liu, S., et al.: Recognize anything: A strong image tagging model. arXiv preprint arXiv:2306.03514 (2023) 3, 4, 5, 7, 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 291 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 291 + ], + "type": "text", + "content": "58. Zhang, Z., Zhang, A., Li, M., Zhao, H., Karypis, G., Smola, A.: Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923 (2023) 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 324 + ], + "type": "text", + "content": "59. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023) 3, 4, 11" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 347 + ], + "type": "text", + "content": "60. Zhu, Z., Yu, J., Wang, Y., Sun, Y., Hu, Y., Wu, Q.: Mucko: multi-layer cross-modal knowledge reasoning for fact-based visual question answering. In: IJCAI (2020) 10" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 270, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 270, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 270, + 102 + ], + "type": "text", + "content": "Haibo Wang, Weifeng Ge" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/QUAR-VLA_ Vision-Language-Action Model for Quadruped Robots/7817397f-b04e-4f74-8ceb-92291da9f1da_content_list.json b/2024/QUAR-VLA_ Vision-Language-Action Model for Quadruped Robots/7817397f-b04e-4f74-8ceb-92291da9f1da_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..75d63c3b1a6f1ace2bf01feba858c26c53f73070 --- /dev/null +++ b/2024/QUAR-VLA_ Vision-Language-Action Model for Quadruped Robots/7817397f-b04e-4f74-8ceb-92291da9f1da_content_list.json @@ -0,0 +1,1486 @@ +[ + { + "type": "text", + "text": "QUAR-VLA: Vision-Language-Action Model for Quadruped Robots", + "text_level": 1, + "bbox": [ + 225, + 140, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pengxiang Ding $^{12}$ Han Zhao $^{12}$ Wenjie Zhang $^{3}$ Wenxuan Song $^{4}$ Min Zhang $^{12}$ Siteng Huang $^{12}$ Ningxi Yang $^{2}$ Donglin Wang $^{1\\star}$", + "bbox": [ + 251, + 210, + 754, + 246 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Zhejiang University 2Westlake University 3Beijing Normal University 4Monash University {dingpengxiang}@westlake.edu.cn", + "bbox": [ + 339, + 253, + 676, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The important manifestation of robot intelligence is the ability to naturally interact and autonomously make decisions. Traditional quadruped robot learning typically handles language interaction and visual autonomous perception separately, which, while simplifying system design, also limits the synergy between different information streams. This separation poses challenges in achieving seamless autonomous reasoning, decision-making, and action execution. To address these limitations, a novel paradigm, named Vision-Language-Action tasks for QUAdruped Robots (QUAR-VLA), has been introduced in this paper. This approach tightly integrates visual information and instructions to generate executable actions, effectively merging perception, planning, and decision-making. The central idea is to elevate the overall intelligence of the robot. Within this framework, a notable challenge lies in aligning fine-grained instructions with visual perception information. This emphasizes the complexity involved in ensuring that the robot accurately interprets and acts upon detailed instructions in harmony with its visual observations. Consequently, we propose QUAdruped Robotic Transformer (QUART), a VLA model to integrate visual information and instructions from diverse modalities as input and generates executable actions for real-world robots and present QUAdruped Robot Dataset (QUARD), a large-scale multi-task dataset including perception, navigation and advanced capability like whole-body manipulation tasks for training QUART model. Our extensive evaluation shows that our approach leads to performant robotic policies and enables QUART to obtain a range of generalization capabilities.", + "bbox": [ + 261, + 321, + 743, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Robotics $\\cdot$ Quadruped Robot Learning $\\cdot$ Vision-Language-Action Model", + "bbox": [ + 261, + 681, + 743, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 729, + 375, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Quadruped robots, characterized by their excellent traversability on complex terrains and agile movements, have garnered significant attention in the field of robotics [14]. Researchers have extensively employed these robots to explore tasks encompassing autonomous navigation and manipulation [16, 17, 36].", + "bbox": [ + 212, + 757, + 787, + 820 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 218, + 825, + 382, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e8766da15dc3f80c3b4234c895c71db6552a514e54fe341da2e6b7bc2e650210.jpg", + "image_caption": [ + "(a) QUAR-VA" + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 401, + 233 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4caa3042167d5cba3571d471f8b0f66ab4dd32381d27f1a51bbd1177cb1ff5c4.jpg", + "image_caption": [ + "(b) QUAR-LA" + ], + "image_footnote": [], + "bbox": [ + 408, + 143, + 593, + 229 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8d7565c4199cb5db8bc9ed69df7ff4c547878b0b524cf732423968de2e6cc4cb.jpg", + "image_caption": [ + "(c) QUAR-VLA", + "Fig. 1: Comparison of QUAR-VA, QUAR-LA, and QUAR-VLA. QUAR-VA solely utilizes coarse-grained vision information, lacking explicit instructions for handling diverse tasks. In contrast, QUAR-LA exclusively relies on language information and lacks of vision information for autonomy. Therefore, QUART-VLA combines both vision information and language instructions as inputs, enabling autonomous problem-solving across a range of tasks, revealing distinct input modalities and task capabilities." + ], + "image_footnote": [], + "bbox": [ + 602, + 143, + 787, + 233 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Broadly speaking, the quadruped tasks consist of two major specifications: Vision-Action tasks for QUAdruped Robots (QUAR-VA) and Language-Action tasks for QUAdruped Robots (QUAR-LA). As depicted in Fig. 1, in QUAR-VA approaches [31], quadruped robots {receive} perception images captured from a first-person perspective and instruction images obtained from a third-person perspective to guide their actions. However, such a task specification often relies on a single (coarse-grained) goal image instruction, making it difficult to apply in many real-world combination tasks, i.e. requiring combining multiple sub-instructions. In contrast, employing language as instructions [36], QUAR-LA formulation allows for executing more fine-grained and diverse tasks, suitable for expressing combinational commands (i.e. \"before / then\"), complex spatial relationships (i.e. \"move to the left\"), and many commonsense command priors (i.e. \"move fast\"). Nevertheless, QUAR-LA approaches, which lack the integration of visual modality, hinder the robots' ability to perceive the environment, thereby impeding their autonomous navigation capabilities [36]. To enable quadruped robots to autonomously navigate and manipulate various tasks, in this paper, we propose a new paradigm: Vision-Language-Action tasks for QUAdruped Robots (QUAR-VLA), integrating visual information and instructions from diverse modalities as input and generating executable actions for real-world robots.", + "bbox": [ + 212, + 412, + 787, + 714 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This task primarily encompasses two challenges. Firstly, there is a lack of large-scale datasets in the research community for quadruped robots performing diverse tasks. Pretraining models require abundant trajectory data from a variety of tasks. However, collecting substantial amounts of real-world data often necessitates human involvement and significant time investment in manual instruction operations.", + "bbox": [ + 212, + 715, + 787, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Secondly, different from navigation tasks [31] for general mobile robots and manipulation tasks for fixed-base systems [7, 9, 12, 34], building a VLA model", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "P. Ding et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/36671b8685ef7c299dd0f8a023a98aea7645b7abed95670b6ee402114d5c9a11.jpg", + "image_caption": [ + "Fig. 2: Overview of QUAR-VLA. Our tasks encompass a diverse range of perception, navigation, and other advanced capability. The Vision-Language-Action (VLA) model first undergoes training with a huge amount of simulation data (246K episodes) and a small amount of real-world data (3K episodes). In the inference phase, images and texts undergo tokenization, after which QUART generates 12-dimensional action tokens. These tokens are subsequently detokenized into valid robot actions and deployed on a physical quadruped robot. This methodology effectively extends the learned capabilities from a simulated environment to real-world applications." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to solve complex quadruped robot tasks is considerably more challenging due to their agile locomotion behaviors. The action space needs to be properly defined to strike a balance between movement flexibility and computational efficiency. The action generated by the model should not be too simplistic, akin to the planner base velocities output by 2-D navigation modules, nor should they require a high execution frequency like the low-level motion policy that directly controls the joint motors.", + "bbox": [ + 212, + 458, + 787, + 563 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address these two problems, we collect a large-scale multi-task dataset QUAdruped Robot Dataset (QUARD). It includes multiple tasks such as perception, navigation, and advanced capabilities like object avoidance. To the best of our knowledge, this is the first quadruped robot dataset that incorporates a significant amount of vision, language instruction, and robot command data. As collecting data on real robots is expensive and inefficient, we primarily rely on data generated in simulation, which exhibits significant differences in visual, sensor, and system dynamics. We also introduce QUAdruped Robotic Transformer (QUART), a VLA model for training QUARD. QUART takes images of the robot's first-view camera and the natural language instruction as inputs and generates rich control commands that include controlling the robot's base velocity, posture, and gait parameters. QUART leverages a pre-trained large-scale visual language model and fine-tunes it on our dataset to enable the generation of executable commands for quadruped robots. To address the simple-to-real gap caused by the data disparity, we construct a co-training pipeline to effectively distill the knowledge of simulation data for real-scene deployment. Our extensive evaluation shows that our approach leads to performant robotic policies and enables QUART to obtain a range of generalization capabilities.", + "bbox": [ + 212, + 568, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "QUAR-VLA", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions are as follows: 1) To the best of our knowledge, we first propose a new paradigm: QUAR-VLA, integrating visual and language instructions to executable actions for more autonomous and diverse quadruped robot tasks. 2) We present a large-scale multi-task dataset, QUARD, and a Vision-Language-Action model, QUART to solve the QUAR-VLA tasks. 3) Our extensive evaluation shows that our approach leads to performant robotic policies and enables QUART to obtain a range of generalization capabilities.", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 277, + 388, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Quadruped Robot Learning. Previous research [2,4,10,13,16,17,19,26,30,36, 39] has delved into single-model policies. As to vision-based methods (QUAR-VA), Kareer et al. [16] utilizes vision perception to predict a privileged terrain map, enabling the robot to perform specific locomotion based on the terrain characteristics. Karnan [17] leverages ego-motion estimates obtained through vision perception, allowing the robot to understand its relative position and direction to avoid collisions and navigate toward a goal location. While some works have succeeded in these tasks through vision perception, relying solely on vision limits the capability to handle only a specific type of task. In response to this limitation, Tang et al. [36] have started exploring language-conditioned tasks (QUAR-LA). However, the absence of vision perception in their approach results in a lack of autonomy for the robot to interact with its environment. Therefore, this paper proposes a new QUAR-VLA paradigm to integrate visual information and instructions to generate executable actions, effectively ensuring that the robot accurately interprets and acts upon detailed instructions in harmony with its visual observations.", + "bbox": [ + 212, + 310, + 787, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vision-language-action models. Visual-language-action models [1, 5, 20, 24, 28, 33, 35] integrates visual information and instructions to generate executable actions, effectively merging perception, planning, and decision-making and elevating the overall intelligence of the robot. As a result, vision-language-action models have been a popular area of research in robotics. In the realm of visual-language-action models, Shridhar et al. [33] have made significant strides by leveraging CLIP to encode text inputs, thereby enhancing the vision model's semantic comprehension and action execution capabilities. Further advancing the field, Brohan et al. [8] and Li et al. [20] have innovatively applied large language models to craft manipulation policies specifically for robotic applications. This paper delves into the domain of embodied agents, focusing on the quadruped robot as a case study. It explores the challenges and opportunities in enabling quadruped robots to autonomously navigate and perform a variety of tasks as directed by human instructions, contributing to the broader discourse on robotic autonomy and intelligence.", + "bbox": [ + 212, + 551, + 787, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Robot Dataset. The realm of robotic learning has been at the forefront of advancing open-source datasets tailored explicitly for robot learning [6,15,17,32, 36,38,40]. Within the domain of quadrupedal robots, previous works on datasets primarily focused on algorithms and leg controllers. Eckert et al. [11] introduced", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "P. Ding et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a grading system and an online dataset for the collection and distribution of agility scores. Caluwaerts et al. [10] sought to address the lack of a standardized and scalable environment and inductive metrics. Tang et al. [36] pioneered the generation of desired joint position control using LLM. Nevertheless, common limitations persist among these quadruped robot datasets: the quantity is insufficient and the tasks are limited in diversity. Hence, our data repository aims to complement these endeavors. We curate and process an extensive array of skills on various objects, across diverse scenes in both the real world and virtual environments. The dataset includes image data, robot action data, and point cloud data, providing quadruped robots with rich perceptual information to tackle increasingly intricate and demanding tasks.", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 334, + 330, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section provides a detailed exposition of our proposed methodology. Initially, we present the definition of our proposed QUAR-VLA in Section 3.1. Following that, Section 3.2 outlines the collecting process of QUARD. Lastly, we delve into the overarching architecture of QUART in Section 3.3.", + "bbox": [ + 212, + 364, + 787, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Problem Setup", + "text_level": 1, + "bbox": [ + 215, + 446, + 385, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The objective of QUAR-VLA is to construct a vision-language-action model learned from large-scale demonstration data and generate actions for closed-loop robot control.", + "bbox": [ + 212, + 469, + 785, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "An overview of our QUAR-VLA is shown in Fig. 2. Our goal is to train a conditional policy QUART that can interpret RGB image(s), denoted $s \\in S$ , together with a task instruction $w \\in \\mathcal{W}$ , which correspond to a language string. The policy is a mapping from images and instructions to actions, and can be written as $\\mu : S \\times \\mathcal{W} \\to \\mathcal{A}$ , where the action space $\\mathcal{A}$ consists of the 11-dimensional high-level commands as well as a terminate signal.", + "bbox": [ + 212, + 515, + 787, + 604 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ v _ {x}, v _ {y}, \\omega_ {z}, \\theta_ {1}, \\theta_ {2}, \\theta_ {3}, f, h _ {z}, \\phi , s _ {y}, h _ {z} ^ {f}, t \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 614, + 784, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $v_{x}, v_{y}$ , and $\\omega_{z}$ represent the velocities along the x-axis, y-axis, and z-axis respectively. $\\theta_{1}, \\theta_{2}$ , and $\\theta_{3}$ indicate the gait pattern, $f$ denotes the frequency, $h_{z}$ represents the height of the robot, $\\phi$ denotes the pitch angle, $s_{y}$ corresponds to the foot width, $h_{z}^{f}$ represents the foot height, and $t$ indicates the termination signal of the action.", + "bbox": [ + 212, + 642, + 787, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Notably, in this study, we employ the discretization method proposed by [9] to discretize all continuous dimensions into 256 uniformly sized bins. This choice of discretization is motivated by the desire to reduce the complexity of action search and improve the stability and convergence of algorithms.", + "bbox": [ + 212, + 717, + 787, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Large-scale Quadruped Robot Datasets", + "text_level": 1, + "bbox": [ + 215, + 785, + 586, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enable an imitation learning system to generalize to new tasks with zero demonstrations of said task, we must be able to easily collect a diverse dataset,", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "QUAR-VLA", + "bbox": [ + 643, + 113, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6cbd25dad11e2c5c15b1a13f11d7d26f570b79e71030e6030d48f390e78626f8.jpg", + "table_caption": [ + "Table 1: Tasks Definition. The \"Type\" means different capabilities of robots. The \"Level\" devises the difficulty into 3 levels. The \"Skill\" means different skill/task categories. The \"Episode\" signifies the number of experiments conducted for each task, which also corresponds to the number of trajectories. The \"Description\" is the illustration of the tasks." + ], + "table_footnote": [], + "table_body": "
TypeLevelSkillEpisodeDescription
PerceptionEasyDistinguish (sim)Letter10KIdentify the correct one from multiple boxes with different printed letters
BasicMediumGo to Object (sim)72KNavigate to the object and stop in front of it
NavigationMediumGo to Object (real)3KNavigate to the object and stop in front of it
Advanced CapabilityHardGo through Tunnel48KSpatial Navigation: Go through the correct tunnel from two tunnels with different colors and shapes
HardGo to Object and avoid the obstacle (sim)63KObstacle Avoidance: Navigate to the object without colliding with the obstacle
HardCrawl under Bar (sim)1KEnvironment Adaptation: Crawl under a bar with a low height
HardUnload Object into Re-ceptacle (sim)52KObject Manipulation: Move with a ball on the back and unload it into a receptacle
Total249KThe total number of episodes
", + "bbox": [ + 217, + 218, + 785, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "provide corrective feedback, and evaluate many tasks at scale. Therefore, we collect a large-scale multi-task dataset, QUAdruped Robot Dataset (QUARD), which includes multiple tasks such as perception, basic navigation, and advanced capability like object manipulation.", + "bbox": [ + 212, + 393, + 784, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Task Definition. As is shown in Table 1, there are seven tasks distributed across three levels of difficulty: easy, medium, and hard. The easy and medium tasks are designed such that the robot can accomplish them using fundamental skills. For instance, the \"Go to\" task is a basic action integral to all navigation tasks, while the \"Distinguish\" task serves as the foundational action for all perception tasks. Hard tasks necessitate the robot to execute tasks using more advanced skills. A case in point is the \"Unload object\" task, which involves the robot distinguishing the target container, navigating to the target position, adjusting body position, and considering potential collisions to successfully unload the object. This process integrates perception, navigation, and whole-body manipulation. The complexity of tasks is reflected in the average trajectory length, with more challenging tasks requiring a greater number of steps and consequently taking longer to complete. The specific trajectory lengths and task distribution for each task are detailed in Fig. 3 and Fig. 4.", + "bbox": [ + 212, + 455, + 785, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "System Setup. The robot used to collect the trajectory data is WR-2, which is a quadruped robot with 12 joints. WR-2 has around $25\\mathrm{cm}$ in standing height and $40\\mathrm{cm}$ body length. The command output is sent to the low-level command tracking controller (pre-trained command-conditioned policy in [23]) to generate the actual joint action of the robot. The simulation data was collected in Nvidia's Isaac Gym [22], a powerful simulator that allows us to collect massive robot trajectories in parallel. The real data was collected by manually manipulating a quadruped robot in the lab setting. The perception data is provided by a RealSense d435 camera installed in the front of the robot.", + "bbox": [ + 212, + 670, + 785, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data Collection. As to simulation data, $\\mathbf{A}^*$ and $\\mathbf{D}^*$ algorithms are used to plot optimal paths for a dog navigating through various objects and obstacles. The", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "P. Ding et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/945e96e3aad707c6432044216c62843e8e774f84b5812c9f4efecbb20bc7624f.jpg", + "image_caption": [ + "Fig. 3: The left figure illustrates the trajectory lengths corresponding to different tasks and the right figure illustrates the relationships between tasks. As the difficulty of the skill increases, it can be observed that the average trajectory length gradually increases. The right figure demonstrates the relationship between the types of tasks: perception as foundational capability; basic navigation ability is built on perception; object manipulation, obstacle avoidance, spatial navigation, and environment adaption extend from perception and navigation." + ], + "image_footnote": [], + "bbox": [ + 272, + 161, + 537, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1af447d773b093233f4d6693c985b85bb02d0a99a77e2dba13a4e4129c92115c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 545, + 161, + 705, + 281 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\mathrm{A}^*$ algorithm seeks the most cost-effective path, while the $\\mathrm{D}^*$ algorithm adapts to changes in real time. A Proportional-Derivative controller then converts these paths into target velocities for the dog, ensuring smooth movement. The combination of both algorithms provides a flexible and efficient method for path planning in various environments. For real data, The real data set is obtained in a laboratory environment using remote control.", + "bbox": [ + 212, + 455, + 787, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Consistency constraints. To maintain consistency between simulation and reality, we've established constraints for data collection. The robot starts at the origin, with the target randomly positioned within [2.7,3.3] meters on the x-axis and [0.9,1.1] meters on the y-axis. In obstacle scenarios, obstacles are placed 1.5 meters from the target's x-coordinate, with the same y-coordinate. Task termination criteria vary. For \"go to object\", \"go to the object and avoid obstacle\", and \"crawl under bar\" tasks, success is when the robot is less than one meter from the target. Other task success criteria are: 1) \"Unload object\": successful when the object is in the target container. 2) \"Go through tunnel\" and \"Crawl under bar\": successful when the robot reaches a specific position behind the tunnel or bar. 3) \"Distinguish letter\": successful when the robot correctly orients towards the visual target.", + "bbox": [ + 212, + 550, + 787, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "More Statistics. The data collection setup adheres to a predefined language template as depicted in Fig. 4, where the task, target object, speed, and gait are explicitly defined. To enhance diversity and generalization capabilities, the dataset includes a variety of common indoor furniture and outdoor facilities, in addition to basic shapes. The basic shapes come in four color variations: green, red, blue, and yellow. Examples of indoor furniture include bookshelves, ovens, and vases, while outdoor facilities encompass trashcans and benches. Recognizing", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "QUAR-VLA", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/508d4d28b3e1f2022b7f49ec06e00e477cd54a2c5d9f82c38e8bc1b0933bafbb.jpg", + "image_caption": [ + "Fig. 4: Statistic analysis of QUARD. 1) Simulation data accounts for a larger proportion compared to real data. 2) Trotting occupied most of the share of gait. 3) Most tasks own similar episodes while Distinguish Letter and Go to Object (in reality) owns much less. 4) 3 types of speed occupied $1/3$ separately." + ], + "image_footnote": [], + "bbox": [ + 222, + 154, + 781, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the distinct nature of motion robots as compared to manipulation robots, we also require robots to perform the same task with varying gaits and speeds to adapt to different instructions and environments. The corresponding ratios of gaits and speeds are presented in Fig. 4. More details about the datasets can be found in the supplementary material.", + "bbox": [ + 212, + 369, + 787, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Vision-Language-Action Model", + "text_level": 1, + "bbox": [ + 215, + 473, + 517, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In comparison to VLA models [7,9,12,31,34] in other field, our research stands out due to the abundance of kinematics information available in this domain. This allows the robot to not only excel at goal-oriented navigation tasks but also perform a diverse range of whole-body manipulation operations with enhanced flexibility in gaits and body control. Next, we will present our QUAdruped Robotic Transformers (QUART).", + "bbox": [ + 212, + 503, + 787, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "QUART relies on a pre-trained vision-language-model [3] to associate tokens from the model's existing tokenization with the discrete action bins. It is worth noting that training MLLMs to override existing tokens with action tokens is a form of symbol tuning [37], which has been shown to work well in prior work. For [3], integers up to 1000 each have a unique token, so we simply associate the action bins with the token representing the corresponding integer. Notably, QUART model takes a single image $s$ and a natural language instruction $w$ as input, which are first converted into corresponding tokens $t$ through a tokenizer $\\tau(t|s,w)$ and fed into a decoder-only transformer module to obtain discretized action tokens $p(a_d|t)$ .", + "bbox": [ + 212, + 595, + 787, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The policy QUART could be shown as follow:", + "bbox": [ + 238, + 748, + 584, + 763 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {Q U A R T} \\left(a _ {d} | s, w\\right) = p \\left(a _ {d} | t\\right) \\tau (t | s, w) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 374, + 782, + 785, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $w, s$ are the input images and language instruction and $\\tau$ represents the tokenizer and $p$ indicates the vision-language model to output action $a_{d}$ .", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "P. Ding et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1853b291f835ff48236e735d4a90624feb41154f530194ce97b5a5607213a1cf.jpg", + "image_caption": [ + "Fig. 5: Architecture of QUART. It is designed to leverage the scene comprehension capability of a pretrained MLLM. It receives visual information as observation, and outputs an action representing the actual action taken by the robot based on text-form instructions, and de-tokenizes it into specific action values. QUART can generate a complete action sequence at a processing rate of $2\\mathrm{Hz}$ in actual scenarios, and hand it over to the underlying low-level strategy for execution." + ], + "image_footnote": [], + "bbox": [ + 230, + 143, + 769, + 369 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Action Detokenize. To directly convert models' output to valid robot actions for downstream control, we need detokenize the discrete action token $a_{d}$ into continuous representation $a_{c}$ (except for the discrete termination command).", + "bbox": [ + 212, + 507, + 787, + 554 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\na _ {c} = \\operatorname {D e t o k e n i z e} \\left(a _ {d}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 566, + 785, + 583 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Loss. We use a standard categorical cross-entropy objective and causal masking that was utilized in prior Transformer-based controllers [18, 29].", + "bbox": [ + 212, + 589, + 785, + 619 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Inference Speed. In contrast to many applications of large models, such as natural language or image generation, one of the unique requirements for a model that needs to run on real robots in real-time is fast and consistent inference speed. It is important to highlight that our research specifically emphasizes the command control of quadruped robots. In contrast to low-level motor control, the command control outputs alleviate the strict control frequency requirements, enabling smooth integration of larger models and unlocking improved reasoning capabilities. For QUART, the inference time could get 2Hz. More details about the deployment could be found in the supplementary material.", + "bbox": [ + 212, + 619, + 787, + 755 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 777, + 375, + 795 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We concentrate our experiments on the multi-task ability and generalization. We try to address the following questions: 1. How effective is the vision-language-", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "QUAR-VLA", + "bbox": [ + 643, + 113, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "action architecture for multi-task quadruped task compared to previous VLM baselines? 2. How well do these models generalize to unseen semantic attributes like object shape, color, and unseen verbal information?", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 219, + 457, + 234 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training Details. We train a specific instantiation of QUART, derived from 8B pre-trained VLM models [3] for superior performance. And we use learning rate 2e-5 and batch size 256 to fine-tune the model for 100K gradient steps. Both models are trained with the next token prediction objective, which corresponds to the behavior cloning loss in robot learning.", + "bbox": [ + 212, + 248, + 787, + 324 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Details. We follow the standard robot evaluation metrics [7, 9], success rate (SR), to evaluate the overall performance. The signal we used for robot control is the 11-dimensional command information with the action space outlined in Section 3.1 and 1 stop terminal command means the end of action. More details for each experiment can be seen in the supplementary material.", + "bbox": [ + 212, + 325, + 787, + 400 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baselines. Considering the absence of VLA models work on quadruped robots at present, we have taken the following baselines into account for a fair comparison: CLIP [27], R3M [25], and VC-1 [21].", + "bbox": [ + 212, + 401, + 787, + 446 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. CLIP [27] uniquely encodes both textual and visual data, fusing these embeddings to create feature sets that integrate textual and visual information.", + "2. R3M [25]is a visual representation model derived from the Ego4D dataset, which acts as a unified, static perception module for policy learning applications.", + "3. VC-1 [21] pushes the boundaries of visual representation learning by amalgamating data from multiple datasets and assessing performance across a diverse range of tasks via the implementation of CortexBench." + ], + "bbox": [ + 214, + 448, + 784, + 555 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Since the aforementioned models do not include language conditioning, we include this aspect by separately embedding the language command, allowing us to compare it to our method. Specifically, we concatenate the resulting language embedding tokens with the image tokens generated by the vision model and pass the combined token sequences through a policy head to produce action outputs.", + "bbox": [ + 212, + 556, + 787, + 633 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Overall Performance", + "text_level": 1, + "bbox": [ + 215, + 659, + 433, + 672 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Multi-task Performance. To assess the overall capabilities of VLA models on multiple tasks, we conduct evaluations using instructions randomly selected from the training set. It's important to note that this evaluation introduces variations in the placement of objects and other setup factors (such as robot position). This variation demands the system's ability to generalize effectively to realistic environmental variability. In total, over 1500 episodes are tested in this evaluation, comprising 425 episodes for going to objects, 500 for going to objects without colliding with the obstacle. 150 for going through the tunnel, 100 for unloading objects, 100 for distinguishing objects, and 75 for crawling under the bar.", + "bbox": [ + 212, + 688, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "P. Ding et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/96f36b3e20974a8a02fa104f0858003730dbf647a2d6c5989abb1496d473009b.jpg", + "image_caption": [ + "Fig. 6: Tasks Successful Rate. R3M lacks alignment with text, which means that although it has some recognition capabilities, it struggles to understand actions through instructions. CLIP and the VC-1 possess the ability to align text and images, enabling them to generate and execute actions based on instructions and observations. As shown above, while R3M may have a slight advantage in easy seen tasks (distinguish), it performs poorly in more complex tasks requiring action execution, significantly lagging behind CLIP and VC-1. QUART fully leverages the advantages of large models, enabling semantic understanding of images and alignment with textual information." + ], + "image_footnote": [], + "bbox": [ + 218, + 146, + 782, + 272 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/7d94ecc7956c413873e159574bb71e6c8a8fc2fdd44f0bdf9dba9025f90e269c.jpg", + "table_caption": [ + "Table 2: Overall performance. QUART has achieved success rates far exceeding those of the baselines in tasks of all difficulty levels, especially in the most challenging crawl and unload tasks, where the baselines have no record of success." + ], + "table_footnote": [], + "table_body": "
SeenUnseen
Easy DistinguishMedium Go toHard Go avoid Go through Crawl UnloadObjectVerbal
CLIP [27]0.440.430.450.19000.110.14
R3M [25]0.580000000
VC-1 [21]0.460.430.450.31000.290.19
QUART0.660.600.530.410.320.120.350.33
", + "bbox": [ + 228, + 478, + 774, + 595 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Comparison within VLM baselines. The experiment results reveal that R3M has poor performances on tasks except distinguish. The primary reason appears to be the lack of alignment with language semantics, which hinders its ability to comprehend tasks beyond basic discrimination. In contrast, CLIP and VC-1 demonstrate satisfactory performance on fundamental perceptual tasks, such as navigation (e.g., \"go to\" commands). However, their efficacy diminishes significantly when tasks require complex mechanical movements. This observation suggests that while visual language models (VLMs) can grasp abstract principles of the world, directly applying VLMs does not readily translate to the execution of intricate physical tasks.", + "2. VLM baselines vs QUART. As is shown in Table 2, Our model has achieved optimal performance on nearly all the baseline models. In our study, we have achieved a decoder-only VLA framework. This approach diverges from the traditional single-layer MLP policy head by leveraging the sequential nature" + ], + "bbox": [ + 212, + 627, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "QUAR-VLA", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/4c3f62ca0c7afa00027d8c4e2ec78126493344173709a1dbc89f6fc349d0d549.jpg", + "image_caption": [ + "Fig. 7: Cases with unseen verbal instruction. Here are 3 verbal instructions. The words below represent the names of instructions in the training data, with the words above in bold indicating verbal instruction. When confronted with unseen instructions, the alignment between the existing language and the integration of vision and action cues within the baselines is compromised, resulting in task failure. This failure manifests in behaviors such as repetitive motion, misdirection, wrong terminate commands. Conversely, QUART, leveraging the language prowess inherited from large language models, adeptly achieves generalization under novel instructions, thereby effectuating the harmonization of vision, language, and action." + ], + "image_footnote": [], + "bbox": [ + 272, + 142, + 723, + 321 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "of action generation. This allows for the implicit learning of interdependencies between different action dimensions through the use of a transformer. Consequently, while the performance gains may be marginal in simple tasks, there is a noticeable enhancement in tasks that involve complex mechanical movements. significantly improved its perceptual capabilities by incorporating commonsense from the multi-modal large model (MLLM).", + "bbox": [ + 212, + 505, + 787, + 595 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Generalization capabilities. To examine the adaptability in unanticipated scenarios, we orchestrated two primary examinations: one concentrated on unfamiliar objects, and the other on unprecedented linguistic directives. For the unfamiliar objects, we cast our gaze upon an array of circumstances: objects belonging to an identical category but exhibiting divergent textures and colors; objects from the same category but of disparate shapes; and objects with differences in shapes and textures, absent from current datasets. Turning towards the unprecedented linguistic instructions, we characterized unfamiliar instances as linguistic directives that bear the same semantics but vary in expression. For example, within the \"go to the object\" task, we gauged the directive's adaptability by employing \"navigate to target\" as the testing instruction.", + "bbox": [ + 212, + 597, + 787, + 763 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "1. Unseen Object. As depicted in Fig. 7, we can observe the enhanced generalization ability of our QUART model. For unseen objects, both QUART and VC-1 perform well, thanks to the advantages of the pre-trained VLM model. However, when it comes to unseen verbal descriptions, apart from QUART, all other models essentially failed. This is attributed to the inherent advantage pro", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "P. Ding et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5cc5a024383e109948eef3cd9e0291c8e7d24af6f2277aecef3bb37b8c84147d.jpg", + "image_caption": [ + "Fig. 8: QUART can extend basic skills to accommodate complex customized skills with commonsense. It shows example performances of combinational command and spatial understanding towards customized command." + ], + "image_footnote": [], + "bbox": [ + 215, + 164, + 767, + 305 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "vided by the pre-trained Multimodal Large Language Model (MLLM), enabling the recognition of distinct variations in object types and the comprehension of subtle semantic nuances. Consequently, the model exhibits exceptional generalization performance.", + "bbox": [ + 212, + 407, + 787, + 468 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "2. Unseen Verbal instruction. The training directives utilized were relatively monotonous, demonstrating that minor alterations in instruction can often perplex the model, thereby revealing an imperative for generalization across language directives. Initially, the difficulties introduced by merely replacing descriptions of skill for task generalization were investigated. Evident from Table 2 and Fig. 7 is that QUART surpasses other baseline methods, an innate advantage of large-scale multimodal models. In addition, to explore more intriguing linguistic directives and fully harness the universal capabilities of the Large Language Model (LLM), Fig. 8 presents instances where the model demonstrates a clear comprehension and application of combined directives constructed via \"first\" and \"then\". It was further noted that even in the absence of explicit directional information within the dataset, the model reliably interpreted linguistic directives which, in turn, influenced task execution accuracy. The ability to utilize language for precise control of quadruped robots facilitated by scene perception, as showcased in Fig. 8, is of immense significance. More intriguing examples and analyses can be found in the supplementary materials.", + "bbox": [ + 212, + 470, + 787, + 715 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Qualitative Performance", + "text_level": 1, + "bbox": [ + 215, + 752, + 464, + 768 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Here, we offer the visualization results of seen and unseen tasks in Fig. 9 with WR-2. We also test QUART in other quadruped robots like Go2 with minor adjustments. More demos can be found in the supplementary materials.", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "QUAR-VLA", + "bbox": [ + 643, + 113, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/93c45c3339580cbf4b271fcfab73b85ea49ae10c2cea78578d15a04cc20d0c05.jpg", + "image_caption": [ + "Fig.9: Visualization of seen and unseen tasks. This figure demonstrates the excellent performance of QUART for seen tasks in both simulated and real-world environments. The left part shows 4 instances of seen tasks in both simulation environment and real-world. The right part shows performances on unseen tasks, which contain unseen verbal, object, and background." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 506, + 310 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/53696d91f86a064b825e390183bb77623289d6480f3ef03f382fc67875b02c50.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 143, + 787, + 310 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion & Future Work", + "text_level": 1, + "bbox": [ + 215, + 435, + 514, + 450 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper emphasizes the significance of deploying Vision-Language-Action models on quadruped robots. We introduce the concept of Vision-Language-Action tasks for QUAdruped Robots (QUAR-VLA), which seamlessly integrates visual information and instructions from diverse modalities to generate executable actions. Specifically, QUAR-VLA focuses on two main aspects of deploying VLA models for quadruped robots: defining the action space to balance flexibility and efficiency, and addressing the scarcity of large-scale training datasets. To tackle these two questions, we present the QUART models tailored for quadruped robots and the QUARD dataset, which includes diverse tasks such as navigation and manipulation. Our extensive evaluation shows that our approach leads to performant robotic policies and enables QUART to obtain a range of emergent capabilities. This includes generalization to novel objects, the ability to interpret commands not present in the robot training data, and the ability to perform rudimentary reasoning in response to user commands.", + "bbox": [ + 212, + 474, + 787, + 686 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Furthermore, this work has these areas requiring further enhancements:", + "bbox": [ + 240, + 689, + 754, + 704 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "1) Automated Data Collection for Larger Datasets. Future work can explore automated processes for data collection, involving techniques such as active learning and data augmentation can be employed to further expand the dataset size and diversity. 2) Better solution for sim2Real gap. Integrating domain adaptation methods like domain randomization, data augmentation, and more accurate modeling of physics engines will be a better solution compared with direct co-training. 3) Improving Inference Speed for Real-time Control. Future works will explore hardware acceleration techniques and model compression techniques to enable faster and more efficient execution of the models.", + "bbox": [ + 212, + 704, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "P. Ding et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by the National Science and Technology Innovation 2030 - Major Project (Grant No. 2022ZD0208800), and NSFC General Program (Grant No. 62176215).", + "bbox": [ + 212, + 175, + 787, + 220 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 243, + 321, + 258 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ahn, M., Dwibedi, D., Finn, C., Arenas, M.G., Gopalakrishnan, K., Hausman, K., Ichter, B., Irpan, A., Joshi, N., Julian, R., Kirmani, S., Leal, I., Lee, E., Levine, S., Lu, Y., Leal, I., Maddineni, S., Rao, K., Sadigh, D., Sanketi, P., Sermanet, P., Vuong, Q., Welker, S., Xia, F., Xiao, T., Xu, P., Xu, S., Xu, Z.: Autort: Embodied foundation models for large scale orchestration of robotic agents (2024)", + "2. Bahl, S., Mendonca, R., Chen, L., Jain, U., Pathak, D.: Affordances from human videos as a versatile representation for robotics (2023)", + "3. Bavishi, R., Olsen, E., Hawthorne, C., Nye, M., Odena, A., Somani, A., Tasirlar, S.: Introducing our multimodal models (2023), https://www.adept.ai/blog/fuyu-8b", + "4. Belkhale, S., Cui, Y., Sadigh, D.: Hydra: Hybrid robot actions for imitation learning. In: Proceedings of the 7th Conference on Robot Learning (CoRL) (2023)", + "5. Bharadhwaj, H., Vakil, J., Sharma, M., Gupta, A., Tulsiani, S., Kumar, V.: Roboagent: Generalization and efficiency in robot manipulation via semantic augmentations and action chunking (2023)", + "6. Bousmalis, K., Irpan, A., Wohlhart, P., Bai, Y., Kelcey, M., Kalakrishnan, M., Downs, L., Ibarz, J., Pastor, P., Konolige, K., et al.: Using simulation and domain adaptation to improve efficiency of deep robotic grasping. In: 2018 IEEE international conference on robotics and automation (ICRA). pp. 4243-4250. IEEE (2018)", + "7. Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Chen, X., Choromanski, K., Ding, T., Driess, D., Dubey, A., Finn, C., Florence, P., Fu, C., Arenas, M.G., Gopalakrishnan, K., Han, K., Hausman, K., Herzog, A., Hsu, J., Ichter, B., Irpan, A., Joshi, N., Julian, R., Kalashnikov, D., Kuang, Y., Leal, I., Lee, L., Lee, T.W.E., Levine, S., Lu, Y., Michalewski, H., Mordatch, I., Pertsch, K., Rao, K., Reymann, K., Ryoo, M., Salazar, G., Sanketi, P., Sermanet, P., Singh, J., Singh, A., Soricut, R., Tran, H., Vanhoucke, V., Vuong, Q., Wahid, A., Welker, S., Wohlhart, P., Wu, J., Xia, F., Xiao, T., Xu, P., Xu, S., Yu, T., Zitkovich, B.: Rt-2: Vision-language-action models transfer web knowledge to robotic control (2023)", + "8. Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Chen, X., Choromanski, K., Ding, T., Driess, D., Dubey, A., Finn, C., et al.: Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818 (2023)", + "9. Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Dabis, J., Finn, C., Gopalakrishnan, K., Hausman, K., Herzog, A., Hsu, J., Ibarz, J., Ichter, B., Irpan, A., Jackson, T., Jesmonth, S., Joshi, N.J., Julian, R., Kalashnikov, D., Kuang, Y., Leal, I., Lee, K.H., Levine, S., Lu, Y., Malla, U., Manjunath, D., Mordatch, I., Nachum, O., Parada, C., Peralta, J., Perez, E., Pertsch, K., Quiambao, J., Rao, K., Ryoo, M., Salazar, G., Sanketi, P., Sayed, K., Singh, J., Sontakke, S., Stone, A., Tan, C., Tran, H., Vanhoucke, V., Vega, S., Vuong, Q., Xia, F., Xiao, T., Xu, P., Xu, S., Yu, T., Zitkovich, B.: Rt-1: Robotics transformer for real-world control at scale (2023)" + ], + "bbox": [ + 225, + 273, + 785, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "QUAR-VLA", + "bbox": [ + 643, + 113, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "10. Caluwaerts, K., Iscen, A., Kew, J.C., Yu, W., Zhang, T., Freeman, D., Lee, K.H., Lee, L., Saliceti, S., Zhuang, V., et al.: Barkour: Benchmarking animal-level agility with quadruped robots. arXiv preprint arXiv:2305.14654 (2023)", + "11. Eckert, P., Ijspeert, A.J.: Benchmarking agility for multilegged terrestrial robots. IEEE Transactions on Robotics 35(2), 529-535 (2019)", + "12. Gu, J., Kirmani, S., Wohlhart, P., Lu, Y., Arenas, M.G., Rao, K., Yu, W., Fu, C., Gopalakrishnan, K., Xu, Z., Sundaresan, P., Xu, P., Su, H., Hausman, K., Finn, C., Vuong, Q., Xiao, T.: Rt-trajectory: Robotic task generalization via hindsight trajectory sketches (2023)", + "13. Haldar, S., Mathur, V., Yarats, D., Pinto, L.: Watch and match: Supercharging imitation with regularized optimal transport. CoRL (2022)", + "14. Hutter, M., Gehring, C., Jud, D., Lauber, A., Bellicoso, C.D., Tsounis, V., Hwangbo, J., Bodie, K., Fankhauser, P., Bloesch, M., Diethelm, R., Bachmann, S., Melzer, A., Hoepflinger, M.: Anymal - a highly mobile and dynamic quadrupedal robot. In: 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 38-44 (2016). https://doi.org/10.1109/IROS.2016.7758092", + "15. Jiang, Y., Moseson, S., Saxena, A.: Efficient grasping from rgbd images: Learning using a new rectangle representation. In: 2011 IEEE International conference on robotics and automation. pp. 3304-3311. IEEE (2011)", + "16. Kareer, S., Yokoyama, N., Batra, D., Ha, S., Truong, J.: Vinl: Visual navigation and locomotion over obstacles. 2023 IEEE International Conference on Robotics and Automation (ICRA) pp. 2018-2024 (2022), https://api-semanticscholar.org/CorpusID:253117178", + "17. Karnan, H., Nair, A., Xiao, X., Warnell, G., Pirk, S., Toshev, A., Hart, J., Biswas, J., Stone, P.: Socially compliant navigation dataset (scand): A large-scale dataset of demonstrations for social navigation. IEEE Robotics and Automation Letters 7(4), 11807-11814 (2022)", + "18. Lee, K.H., Nachum, O., Yang, M.S., Lee, L., Freeman, D., Guadarrama, S., Fischer, I., Xu, W., Jang, E., Michalewski, H., et al.: Multi-game decision transformers. Advances in Neural Information Processing Systems 35, 27921-27936 (2022)", + "19. Lee, M.A., Zhu, Y., Srinivasan, K., Shah, P., Savarese, S., Fei-Fei, L., Garg, A., Bohg, J.: Making sense of vision and touch: Self-supervised learning of multimodal representations for contact-rich tasks. In: 2019 IEEE International Conference on Robotics and Automation (ICRA) (2019), https://arxiv.org/abs/1810.10191", + "20. Li, X., Liu, M., Zhang, H., Yu, C., Xu, J., Wu, H., Cheang, C., Jing, Y., Zhang, W., Liu, H., et al.: Vision-language foundation models as effective robot imitators. arXiv preprint arXiv:2311.01378 (2023)", + "21. Majumdar, A., Yadav, K., Arnaud, S., Ma, Y.J., Chen, C., Silwal, S., Jain, A., Berges, V.P., Abbeel, P., Malik, J., Batra, D., Lin, Y., Maksymets, O., Rajeswaran, A., Meier, F.: Where are we in the search for an artificial visual cortex for embodied intelligence? (2023)", + "22. Makoviychuk, V., Wawrzyniak, L., Guo, Y., Lu, M., Storey, K., Macklin, M., Hoeller, D., Rudin, N., Allshire, A., Handa, A., State, G.: Isaac gym: High performancegpu-based physics simulation for robot learning (2021)", + "23. Margolis, G.B., Agrawal, P.: Walk these ways: Tuning robot control for generalization with multiplicity of behavior. In: Liu, K., Kulic, D., Ichnowski, J. (eds.) Proceedings of The 6th Conference on Robot Learning. Proceedings of Machine Learning Research, vol. 205, pp. 22-31. PMLR (14-18 Dec 2023)", + "24. Nair, S., Rajeswaran, A., Kumar, V., Finn, C., Gupta, A.: R3m: A universal visual representation for robot manipulation. In: Conference on Robot Learning (2022), https://api-semanticscholar.org/CorpusID:247618840" + ], + "bbox": [ + 217, + 146, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "P. Ding et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "25. Nair, S., Rajeswaran, A., Kumar, V., Finn, C., Gupta, A.: R3m: A universal visual representation for robot manipulation (2022)", + "26. Pari, J., Shafiullah, N.M., Arunachalam, S.P., Pinto, L.: The surprising effectiveness of representation learning for visual imitation (2021)", + "27. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision (2021)", + "28. Reed, S., Zolna, K., Parisotto, E., Colmenarejo, S.G., Novikov, A., Barth-Maron, G., Gimenez, M., Sulsky, Y., Kay, J., Springenberg, J.T., et al.: A generalist agent. arXiv preprint arXiv:2205.06175 (2022)", + "29. Reed, S., Zolna, K., Parisotto, E., Colmenarejo, S.G., Novikov, A., Barth-Maron, G., Gimenez, M., Sulsky, Y., Kay, J., Springenberg, J.T., et al.: A generalist agent. arXiv preprint arXiv:2205.06175 (2022)", + "30. Schiavi, G., Wulkop, P., Rizzi, G., Ott, L., Siegwart, R., Chung, J.J.: Learning agent-aware affordances for closed-loop interaction with articulated objects. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 5916-5922 (2023). https://doi.org/10.1109/ICRA48891.2023.10160747", + "31. Shah, D., Sridhar, A., Dashora, N., Stachowicz, K., Black, K., Hirose, N., Levine, S.: Vint: A foundation model for visual navigation (2023)", + "32. Shilane, P., Min, P., Kazhdan, M., Funkhouser, T.: The princeton shape benchmark. In: Proceedings Shape Modeling Applications, 2004. pp. 167-178 (2004)", + "33. Shridhar, M., Manuelli, L., Fox, D.: CIoport: What and where pathways for robotic manipulation. In: Conference on Robot Learning. pp. 894-906. PMLR (2022)", + "34. Stone, A., Xiao, T., Lu, Y., Gopalakrishnan, K., Lee, K.H., Vuong, Q., Wohlhart, P., Kirmani, S., Zitkovich, B., Xia, F., Finn, C., Hausman, K.: Open-world object manipulation using pre-trained vision-language models (2023)", + "35. Szot, A., Schwarzer, M., Agrawal, H., Mazoure, B., Talbott, W., Metcalf, K., Mackraz, N., Hjelm, D., Toshev, A.: Large language models as generalizable policies for embodied tasks. arXiv preprint arXiv:2310.17722 (2023)", + "36. Tang, Y., Yu, W., Tan, J., Zen, H., Faust, A., Harada, T.: Saytap: Language to quadrupedal locomotion. arXiv preprint arXiv:2306.07580 (2023)", + "37. Wei, J., Hou, L., Lampinen, A., Chen, X., Huang, D., Tay, Y., Chen, X., Lu, Y., Zhou, D., Ma, T., et al.: Symbol tuning improves in-context learning in language models. arXiv preprint arXiv:2305.08298 (2023)", + "38. Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., Xiao, J.: 3d shapenets: A deep representation for volumetric shapes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1912-1920 (2015)", + "39. Yang, R., Zhang, M., Hansen, N., Xu, H., Wang, X.: Learning vision-guided quadrupedal locomotion end-to-end with cross-modal transformers. ArXiv abs/2107.03996 (2021), https://api_semanticscholar.org/CorpusID: 235765481", + "40. Yu, K.T., Bauza, M., Fazeli, N., Rodriguez, A.: More than a million ways to be pushed. a high-fidelity experimental dataset of planar pushing. In: 2016 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 30-37. IEEE (2016)" + ], + "bbox": [ + 212, + 146, + 787, + 771 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "QUAR-VLA", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/QUAR-VLA_ Vision-Language-Action Model for Quadruped Robots/7817397f-b04e-4f74-8ceb-92291da9f1da_model.json b/2024/QUAR-VLA_ Vision-Language-Action Model for Quadruped Robots/7817397f-b04e-4f74-8ceb-92291da9f1da_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9e088c7e195f36b986947f95770220f2b3536915 --- /dev/null +++ b/2024/QUAR-VLA_ Vision-Language-Action Model for Quadruped Robots/7817397f-b04e-4f74-8ceb-92291da9f1da_model.json @@ -0,0 +1,2038 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.227, + 0.141, + 0.78, + 0.187 + ], + "angle": 0, + "content": "QUAR-VLA: Vision-Language-Action Model for Quadruped Robots" + }, + { + "type": "text", + "bbox": [ + 0.252, + 0.212, + 0.756, + 0.247 + ], + "angle": 0, + "content": "Pengxiang Ding\\(^{12}\\) Han Zhao\\(^{12}\\) Wenjie Zhang\\(^{3}\\) Wenxuan Song\\(^{4}\\) Min Zhang\\(^{12}\\) Siteng Huang\\(^{12}\\) Ningxi Yang\\(^{2}\\) Donglin Wang\\(^{1\\star}\\)" + }, + { + "type": "text", + "bbox": [ + 0.341, + 0.254, + 0.677, + 0.298 + ], + "angle": 0, + "content": "1Zhejiang University 2Westlake University 3Beijing Normal University 4Monash University {dingpengxiang}@westlake.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.323, + 0.744, + 0.671 + ], + "angle": 0, + "content": "Abstract. The important manifestation of robot intelligence is the ability to naturally interact and autonomously make decisions. Traditional quadruped robot learning typically handles language interaction and visual autonomous perception separately, which, while simplifying system design, also limits the synergy between different information streams. This separation poses challenges in achieving seamless autonomous reasoning, decision-making, and action execution. To address these limitations, a novel paradigm, named Vision-Language-Action tasks for QUAdruped Robots (QUAR-VLA), has been introduced in this paper. This approach tightly integrates visual information and instructions to generate executable actions, effectively merging perception, planning, and decision-making. The central idea is to elevate the overall intelligence of the robot. Within this framework, a notable challenge lies in aligning fine-grained instructions with visual perception information. This emphasizes the complexity involved in ensuring that the robot accurately interprets and acts upon detailed instructions in harmony with its visual observations. Consequently, we propose QUAdruped Robotic Transformer (QUART), a VLA model to integrate visual information and instructions from diverse modalities as input and generates executable actions for real-world robots and present QUAdruped Robot Dataset (QUARD), a large-scale multi-task dataset including perception, navigation and advanced capability like whole-body manipulation tasks for training QUART model. Our extensive evaluation shows that our approach leads to performant robotic policies and enables QUART to obtain a range of generalization capabilities." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.682, + 0.744, + 0.71 + ], + "angle": 0, + "content": "Keywords: Robotics \\(\\cdot\\) Quadruped Robot Learning \\(\\cdot\\) Vision-Language-Action Model" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.73, + 0.377, + 0.746 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.758, + 0.788, + 0.821 + ], + "angle": 0, + "content": "Quadruped robots, characterized by their excellent traversability on complex terrains and agile movements, have garnered significant attention in the field of robotics [14]. Researchers have extensively employed these robots to explore tasks encompassing autonomous navigation and manipulation [16, 17, 36]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.826, + 0.383, + 0.841 + ], + "angle": 0, + "content": "* Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "P. Ding et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.144, + 0.402, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.278, + 0.244, + 0.361, + 0.258 + ], + "angle": 0, + "content": "(a) QUAR-VA" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.145, + 0.594, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.471, + 0.244, + 0.554, + 0.258 + ], + "angle": 0, + "content": "(b) QUAR-LA" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.144, + 0.788, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.663, + 0.244, + 0.754, + 0.258 + ], + "angle": 0, + "content": "(c) QUAR-VLA" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.281, + 0.788, + 0.379 + ], + "angle": 0, + "content": "Fig. 1: Comparison of QUAR-VA, QUAR-LA, and QUAR-VLA. QUAR-VA solely utilizes coarse-grained vision information, lacking explicit instructions for handling diverse tasks. In contrast, QUAR-LA exclusively relies on language information and lacks of vision information for autonomy. Therefore, QUART-VLA combines both vision information and language instructions as inputs, enabling autonomous problem-solving across a range of tasks, revealing distinct input modalities and task capabilities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.789, + 0.715 + ], + "angle": 0, + "content": "Broadly speaking, the quadruped tasks consist of two major specifications: Vision-Action tasks for QUAdruped Robots (QUAR-VA) and Language-Action tasks for QUAdruped Robots (QUAR-LA). As depicted in Fig. 1, in QUAR-VA approaches [31], quadruped robots {receive} perception images captured from a first-person perspective and instruction images obtained from a third-person perspective to guide their actions. However, such a task specification often relies on a single (coarse-grained) goal image instruction, making it difficult to apply in many real-world combination tasks, i.e. requiring combining multiple sub-instructions. In contrast, employing language as instructions [36], QUAR-LA formulation allows for executing more fine-grained and diverse tasks, suitable for expressing combinational commands (i.e. \"before / then\"), complex spatial relationships (i.e. \"move to the left\"), and many commonsense command priors (i.e. \"move fast\"). Nevertheless, QUAR-LA approaches, which lack the integration of visual modality, hinder the robots' ability to perceive the environment, thereby impeding their autonomous navigation capabilities [36]. To enable quadruped robots to autonomously navigate and manipulate various tasks, in this paper, we propose a new paradigm: Vision-Language-Action tasks for QUAdruped Robots (QUAR-VLA), integrating visual information and instructions from diverse modalities as input and generating executable actions for real-world robots." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.717, + 0.789, + 0.808 + ], + "angle": 0, + "content": "This task primarily encompasses two challenges. Firstly, there is a lack of large-scale datasets in the research community for quadruped robots performing diverse tasks. Pretraining models require abundant trajectory data from a variety of tasks. However, collecting substantial amounts of real-world data often necessitates human involvement and significant time investment in manual instruction operations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Secondly, different from navigation tasks [31] for general mobile robots and manipulation tasks for fixed-base systems [7, 9, 12, 34], building a VLA model" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.646, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "QUAR-VLA" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.31, + 0.788, + 0.422 + ], + "angle": 0, + "content": "Fig. 2: Overview of QUAR-VLA. Our tasks encompass a diverse range of perception, navigation, and other advanced capability. The Vision-Language-Action (VLA) model first undergoes training with a huge amount of simulation data (246K episodes) and a small amount of real-world data (3K episodes). In the inference phase, images and texts undergo tokenization, after which QUART generates 12-dimensional action tokens. These tokens are subsequently detokenized into valid robot actions and deployed on a physical quadruped robot. This methodology effectively extends the learned capabilities from a simulated environment to real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.564 + ], + "angle": 0, + "content": "to solve complex quadruped robot tasks is considerably more challenging due to their agile locomotion behaviors. The action space needs to be properly defined to strike a balance between movement flexibility and computational efficiency. The action generated by the model should not be too simplistic, akin to the planner base velocities output by 2-D navigation modules, nor should they require a high execution frequency like the low-level motion policy that directly controls the joint motors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.789, + 0.842 + ], + "angle": 0, + "content": "To address these two problems, we collect a large-scale multi-task dataset QUAdruped Robot Dataset (QUARD). It includes multiple tasks such as perception, navigation, and advanced capabilities like object avoidance. To the best of our knowledge, this is the first quadruped robot dataset that incorporates a significant amount of vision, language instruction, and robot command data. As collecting data on real robots is expensive and inefficient, we primarily rely on data generated in simulation, which exhibits significant differences in visual, sensor, and system dynamics. We also introduce QUAdruped Robotic Transformer (QUART), a VLA model for training QUARD. QUART takes images of the robot's first-view camera and the natural language instruction as inputs and generates rich control commands that include controlling the robot's base velocity, posture, and gait parameters. QUART leverages a pre-trained large-scale visual language model and fine-tunes it on our dataset to enable the generation of executable commands for quadruped robots. To address the simple-to-real gap caused by the data disparity, we construct a co-training pipeline to effectively distill the knowledge of simulation data for real-scene deployment. Our extensive evaluation shows that our approach leads to performant robotic policies and enables QUART to obtain a range of generalization capabilities." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "P. Ding et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "Our contributions are as follows: 1) To the best of our knowledge, we first propose a new paradigm: QUAR-VLA, integrating visual and language instructions to executable actions for more autonomous and diverse quadruped robot tasks. 2) We present a large-scale multi-task dataset, QUARD, and a Vision-Language-Action model, QUART to solve the QUAR-VLA tasks. 3) Our extensive evaluation shows that our approach leads to performant robotic policies and enables QUART to obtain a range of generalization capabilities." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.278, + 0.389, + 0.294 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.311, + 0.789, + 0.552 + ], + "angle": 0, + "content": "Quadruped Robot Learning. Previous research [2,4,10,13,16,17,19,26,30,36, 39] has delved into single-model policies. As to vision-based methods (QUAR-VA), Kareer et al. [16] utilizes vision perception to predict a privileged terrain map, enabling the robot to perform specific locomotion based on the terrain characteristics. Karnan [17] leverages ego-motion estimates obtained through vision perception, allowing the robot to understand its relative position and direction to avoid collisions and navigate toward a goal location. While some works have succeeded in these tasks through vision perception, relying solely on vision limits the capability to handle only a specific type of task. In response to this limitation, Tang et al. [36] have started exploring language-conditioned tasks (QUAR-LA). However, the absence of vision perception in their approach results in a lack of autonomy for the robot to interact with its environment. Therefore, this paper proposes a new QUAR-VLA paradigm to integrate visual information and instructions to generate executable actions, effectively ensuring that the robot accurately interprets and acts upon detailed instructions in harmony with its visual observations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.789, + 0.78 + ], + "angle": 0, + "content": "Vision-language-action models. Visual-language-action models [1, 5, 20, 24, 28, 33, 35] integrates visual information and instructions to generate executable actions, effectively merging perception, planning, and decision-making and elevating the overall intelligence of the robot. As a result, vision-language-action models have been a popular area of research in robotics. In the realm of visual-language-action models, Shridhar et al. [33] have made significant strides by leveraging CLIP to encode text inputs, thereby enhancing the vision model's semantic comprehension and action execution capabilities. Further advancing the field, Brohan et al. [8] and Li et al. [20] have innovatively applied large language models to craft manipulation policies specifically for robotic applications. This paper delves into the domain of embodied agents, focusing on the quadruped robot as a case study. It explores the challenges and opportunities in enabling quadruped robots to autonomously navigate and perform a variety of tasks as directed by human instructions, contributing to the broader discourse on robotic autonomy and intelligence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Robot Dataset. The realm of robotic learning has been at the forefront of advancing open-source datasets tailored explicitly for robot learning [6,15,17,32, 36,38,40]. Within the domain of quadrupedal robots, previous works on datasets primarily focused on algorithms and leg controllers. Eckert et al. [11] introduced" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.645, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "QUAR-VLA" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "a grading system and an online dataset for the collection and distribution of agility scores. Caluwaerts et al. [10] sought to address the lack of a standardized and scalable environment and inductive metrics. Tang et al. [36] pioneered the generation of desired joint position control using LLM. Nevertheless, common limitations persist among these quadruped robot datasets: the quantity is insufficient and the tasks are limited in diversity. Hence, our data repository aims to complement these endeavors. We curate and process an extensive array of skills on various objects, across diverse scenes in both the real world and virtual environments. The dataset includes image data, robot action data, and point cloud data, providing quadruped robots with rich perceptual information to tackle increasingly intricate and demanding tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.335, + 0.331, + 0.351 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.426 + ], + "angle": 0, + "content": "This section provides a detailed exposition of our proposed methodology. Initially, we present the definition of our proposed QUAR-VLA in Section 3.1. Following that, Section 3.2 outlines the collecting process of QUARD. Lastly, we delve into the overarching architecture of QUART in Section 3.3." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.447, + 0.387, + 0.463 + ], + "angle": 0, + "content": "3.1 Problem Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.47, + 0.787, + 0.514 + ], + "angle": 0, + "content": "The objective of QUAR-VLA is to construct a vision-language-action model learned from large-scale demonstration data and generate actions for closed-loop robot control." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.516, + 0.788, + 0.606 + ], + "angle": 0, + "content": "An overview of our QUAR-VLA is shown in Fig. 2. Our goal is to train a conditional policy QUART that can interpret RGB image(s), denoted \\( s \\in S \\), together with a task instruction \\( w \\in \\mathcal{W} \\), which correspond to a language string. The policy is a mapping from images and instructions to actions, and can be written as \\( \\mu : S \\times \\mathcal{W} \\to \\mathcal{A} \\), where the action space \\( \\mathcal{A} \\) consists of the 11-dimensional high-level commands as well as a terminate signal." + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.616, + 0.785, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\left[ v _ {x}, v _ {y}, \\omega_ {z}, \\theta_ {1}, \\theta_ {2}, \\theta_ {3}, f, h _ {z}, \\phi , s _ {y}, h _ {z} ^ {f}, t \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.788, + 0.717 + ], + "angle": 0, + "content": "Here, \\( v_{x}, v_{y} \\), and \\( \\omega_{z} \\) represent the velocities along the x-axis, y-axis, and z-axis respectively. \\( \\theta_{1}, \\theta_{2} \\), and \\( \\theta_{3} \\) indicate the gait pattern, \\( f \\) denotes the frequency, \\( h_{z} \\) represents the height of the robot, \\( \\phi \\) denotes the pitch angle, \\( s_{y} \\) corresponds to the foot width, \\( h_{z}^{f} \\) represents the foot height, and \\( t \\) indicates the termination signal of the action." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.718, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Notably, in this study, we employ the discretization method proposed by [9] to discretize all continuous dimensions into 256 uniformly sized bins. This choice of discretization is motivated by the desire to reduce the complexity of action search and improve the stability and convergence of algorithms." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.786, + 0.587, + 0.802 + ], + "angle": 0, + "content": "3.2 Large-scale Quadruped Robot Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To enable an imitation learning system to generalize to new tasks with zero demonstrations of said task, we must be able to easily collect a diverse dataset," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "P. Ding et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.214 + ], + "angle": 0, + "content": "Table 1: Tasks Definition. The \"Type\" means different capabilities of robots. The \"Level\" devises the difficulty into 3 levels. The \"Skill\" means different skill/task categories. The \"Episode\" signifies the number of experiments conducted for each task, which also corresponds to the number of trajectories. The \"Description\" is the illustration of the tasks." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.219, + 0.787, + 0.361 + ], + "angle": 0, + "content": "
TypeLevelSkillEpisodeDescription
PerceptionEasyDistinguish (sim)Letter10KIdentify the correct one from multiple boxes with different printed letters
BasicMediumGo to Object (sim)72KNavigate to the object and stop in front of it
NavigationMediumGo to Object (real)3KNavigate to the object and stop in front of it
Advanced CapabilityHardGo through Tunnel48KSpatial Navigation: Go through the correct tunnel from two tunnels with different colors and shapes
HardGo to Object and avoid the obstacle (sim)63KObstacle Avoidance: Navigate to the object without colliding with the obstacle
HardCrawl under Bar (sim)1KEnvironment Adaptation: Crawl under a bar with a low height
HardUnload Object into Re-ceptacle (sim)52KObject Manipulation: Move with a ball on the back and unload it into a receptacle
Total249KThe total number of episodes
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.785, + 0.455 + ], + "angle": 0, + "content": "provide corrective feedback, and evaluate many tasks at scale. Therefore, we collect a large-scale multi-task dataset, QUAdruped Robot Dataset (QUARD), which includes multiple tasks such as perception, basic navigation, and advanced capability like object manipulation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.787, + 0.669 + ], + "angle": 0, + "content": "Task Definition. As is shown in Table 1, there are seven tasks distributed across three levels of difficulty: easy, medium, and hard. The easy and medium tasks are designed such that the robot can accomplish them using fundamental skills. For instance, the \"Go to\" task is a basic action integral to all navigation tasks, while the \"Distinguish\" task serves as the foundational action for all perception tasks. Hard tasks necessitate the robot to execute tasks using more advanced skills. A case in point is the \"Unload object\" task, which involves the robot distinguishing the target container, navigating to the target position, adjusting body position, and considering potential collisions to successfully unload the object. This process integrates perception, navigation, and whole-body manipulation. The complexity of tasks is reflected in the average trajectory length, with more challenging tasks requiring a greater number of steps and consequently taking longer to complete. The specific trajectory lengths and task distribution for each task are detailed in Fig. 3 and Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.671, + 0.787, + 0.807 + ], + "angle": 0, + "content": "System Setup. The robot used to collect the trajectory data is WR-2, which is a quadruped robot with 12 joints. WR-2 has around \\(25\\mathrm{cm}\\) in standing height and \\(40\\mathrm{cm}\\) body length. The command output is sent to the low-level command tracking controller (pre-trained command-conditioned policy in [23]) to generate the actual joint action of the robot. The simulation data was collected in Nvidia's Isaac Gym [22], a powerful simulator that allows us to collect massive robot trajectories in parallel. The real data was collected by manually manipulating a quadruped robot in the lab setting. The perception data is provided by a RealSense d435 camera installed in the front of the robot." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Data Collection. As to simulation data, \\(\\mathbf{A}^*\\) and \\(\\mathbf{D}^*\\) algorithms are used to plot optimal paths for a dog navigating through various objects and obstacles. The" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.646, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "QUAR-VLA" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.162, + 0.538, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.162, + 0.707, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.324, + 0.788, + 0.422 + ], + "angle": 0, + "content": "Fig. 3: The left figure illustrates the trajectory lengths corresponding to different tasks and the right figure illustrates the relationships between tasks. As the difficulty of the skill increases, it can be observed that the average trajectory length gradually increases. The right figure demonstrates the relationship between the types of tasks: perception as foundational capability; basic navigation ability is built on perception; object manipulation, obstacle avoidance, spatial navigation, and environment adaption extend from perception and navigation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.548 + ], + "angle": 0, + "content": "\\(\\mathrm{A}^*\\) algorithm seeks the most cost-effective path, while the \\(\\mathrm{D}^*\\) algorithm adapts to changes in real time. A Proportional-Derivative controller then converts these paths into target velocities for the dog, ensuring smooth movement. The combination of both algorithms provides a flexible and efficient method for path planning in various environments. For real data, The real data set is obtained in a laboratory environment using remote control." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.788, + 0.734 + ], + "angle": 0, + "content": "Consistency constraints. To maintain consistency between simulation and reality, we've established constraints for data collection. The robot starts at the origin, with the target randomly positioned within [2.7,3.3] meters on the x-axis and [0.9,1.1] meters on the y-axis. In obstacle scenarios, obstacles are placed 1.5 meters from the target's x-coordinate, with the same y-coordinate. Task termination criteria vary. For \"go to object\", \"go to the object and avoid obstacle\", and \"crawl under bar\" tasks, success is when the robot is less than one meter from the target. Other task success criteria are: 1) \"Unload object\": successful when the object is in the target container. 2) \"Go through tunnel\" and \"Crawl under bar\": successful when the robot reaches a specific position behind the tunnel or bar. 3) \"Distinguish letter\": successful when the robot correctly orients towards the visual target." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "More Statistics. The data collection setup adheres to a predefined language template as depicted in Fig. 4, where the task, target object, speed, and gait are explicitly defined. To enhance diversity and generalization capabilities, the dataset includes a variety of common indoor furniture and outdoor facilities, in addition to basic shapes. The basic shapes come in four color variations: green, red, blue, and yellow. Examples of indoor furniture include bookshelves, ovens, and vases, while outdoor facilities encompass trashcans and benches. Recognizing" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "P. Ding et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.155, + 0.782, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.282, + 0.788, + 0.34 + ], + "angle": 0, + "content": "Fig. 4: Statistic analysis of QUARD. 1) Simulation data accounts for a larger proportion compared to real data. 2) Trotting occupied most of the share of gait. 3) Most tasks own similar episodes while Distinguish Letter and Go to Object (in reality) owns much less. 4) 3 types of speed occupied \\(1/3\\) separately." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.371, + 0.788, + 0.448 + ], + "angle": 0, + "content": "the distinct nature of motion robots as compared to manipulation robots, we also require robots to perform the same task with varying gaits and speeds to adapt to different instructions and environments. The corresponding ratios of gaits and speeds are presented in Fig. 4. More details about the datasets can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.474, + 0.518, + 0.489 + ], + "angle": 0, + "content": "3.3 Vision-Language-Action Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.788, + 0.595 + ], + "angle": 0, + "content": "In comparison to VLA models [7,9,12,31,34] in other field, our research stands out due to the abundance of kinematics information available in this domain. This allows the robot to not only excel at goal-oriented navigation tasks but also perform a diverse range of whole-body manipulation operations with enhanced flexibility in gaits and body control. Next, we will present our QUAdruped Robotic Transformers (QUART)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.788, + 0.748 + ], + "angle": 0, + "content": "QUART relies on a pre-trained vision-language-model [3] to associate tokens from the model's existing tokenization with the discrete action bins. It is worth noting that training MLLMs to override existing tokens with action tokens is a form of symbol tuning [37], which has been shown to work well in prior work. For [3], integers up to 1000 each have a unique token, so we simply associate the action bins with the token representing the corresponding integer. Notably, QUART model takes a single image \\( s \\) and a natural language instruction \\( w \\) as input, which are first converted into corresponding tokens \\( t \\) through a tokenizer \\( \\tau(t|s,w) \\) and fed into a decoder-only transformer module to obtain discretized action tokens \\( p(a_d|t) \\)." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.749, + 0.585, + 0.765 + ], + "angle": 0, + "content": "The policy QUART could be shown as follow:" + }, + { + "type": "equation", + "bbox": [ + 0.375, + 0.783, + 0.787, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\operatorname {Q U A R T} \\left(a _ {d} | s, w\\right) = p \\left(a _ {d} | t\\right) \\tau (t | s, w) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "where \\( w, s \\) are the input images and language instruction and \\( \\tau \\) represents the tokenizer and \\( p \\) indicates the vision-language model to output action \\( a_{d} \\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.645, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "QUAR-VLA" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.144, + 0.77, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.394, + 0.788, + 0.478 + ], + "angle": 0, + "content": "Fig. 5: Architecture of QUART. It is designed to leverage the scene comprehension capability of a pretrained MLLM. It receives visual information as observation, and outputs an action representing the actual action taken by the robot based on text-form instructions, and de-tokenizes it into specific action values. QUART can generate a complete action sequence at a processing rate of \\(2\\mathrm{Hz}\\) in actual scenarios, and hand it over to the underlying low-level strategy for execution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.555 + ], + "angle": 0, + "content": "Action Detokenize. To directly convert models' output to valid robot actions for downstream control, we need detokenize the discrete action token \\( a_{d} \\) into continuous representation \\( a_{c} \\) (except for the discrete termination command)." + }, + { + "type": "equation", + "bbox": [ + 0.418, + 0.568, + 0.786, + 0.584 + ], + "angle": 0, + "content": "\\[\na _ {c} = \\operatorname {D e t o k e n i z e} \\left(a _ {d}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.59, + 0.787, + 0.621 + ], + "angle": 0, + "content": "Loss. We use a standard categorical cross-entropy objective and causal masking that was utilized in prior Transformer-based controllers [18, 29]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.621, + 0.788, + 0.756 + ], + "angle": 0, + "content": "Inference Speed. In contrast to many applications of large models, such as natural language or image generation, one of the unique requirements for a model that needs to run on real robots in real-time is fast and consistent inference speed. It is important to highlight that our research specifically emphasizes the command control of quadruped robots. In contrast to low-level motor control, the command control outputs alleviate the strict control frequency requirements, enabling smooth integration of larger models and unlocking improved reasoning capabilities. For QUART, the inference time could get 2Hz. More details about the deployment could be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.779, + 0.377, + 0.796 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We concentrate our experiments on the multi-task ability and generalization. We try to address the following questions: 1. How effective is the vision-language-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "P. Ding et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "action architecture for multi-task quadruped task compared to previous VLM baselines? 2. How well do these models generalize to unseen semantic attributes like object shape, color, and unseen verbal information?" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.22, + 0.458, + 0.235 + ], + "angle": 0, + "content": "4.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.249, + 0.788, + 0.325 + ], + "angle": 0, + "content": "Training Details. We train a specific instantiation of QUART, derived from 8B pre-trained VLM models [3] for superior performance. And we use learning rate 2e-5 and batch size 256 to fine-tune the model for 100K gradient steps. Both models are trained with the next token prediction objective, which corresponds to the behavior cloning loss in robot learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.326, + 0.788, + 0.401 + ], + "angle": 0, + "content": "Evaluation Details. We follow the standard robot evaluation metrics [7, 9], success rate (SR), to evaluate the overall performance. The signal we used for robot control is the 11-dimensional command information with the action space outlined in Section 3.1 and 1 stop terminal command means the end of action. More details for each experiment can be seen in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.788, + 0.447 + ], + "angle": 0, + "content": "Baselines. Considering the absence of VLA models work on quadruped robots at present, we have taken the following baselines into account for a fair comparison: CLIP [27], R3M [25], and VC-1 [21]." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.449, + 0.785, + 0.478 + ], + "angle": 0, + "content": "1. CLIP [27] uniquely encodes both textual and visual data, fusing these embeddings to create feature sets that integrate textual and visual information." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.48, + 0.785, + 0.509 + ], + "angle": 0, + "content": "2. R3M [25]is a visual representation model derived from the Ego4D dataset, which acts as a unified, static perception module for policy learning applications." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.511, + 0.785, + 0.556 + ], + "angle": 0, + "content": "3. VC-1 [21] pushes the boundaries of visual representation learning by amalgamating data from multiple datasets and assessing performance across a diverse range of tasks via the implementation of CortexBench." + }, + { + "type": "list", + "bbox": [ + 0.215, + 0.449, + 0.785, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.557, + 0.788, + 0.634 + ], + "angle": 0, + "content": "Since the aforementioned models do not include language conditioning, we include this aspect by separately embedding the language command, allowing us to compare it to our method. Specifically, we concatenate the resulting language embedding tokens with the image tokens generated by the vision model and pass the combined token sequences through a policy head to produce action outputs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.66, + 0.434, + 0.673 + ], + "angle": 0, + "content": "4.2 Overall Performance" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Multi-task Performance. To assess the overall capabilities of VLA models on multiple tasks, we conduct evaluations using instructions randomly selected from the training set. It's important to note that this evaluation introduces variations in the placement of objects and other setup factors (such as robot position). This variation demands the system's ability to generalize effectively to realistic environmental variability. In total, over 1500 episodes are tested in this evaluation, comprising 425 episodes for going to objects, 500 for going to objects without colliding with the obstacle. 150 for going through the tunnel, 100 for unloading objects, 100 for distinguishing objects, and 75 for crawling under the bar." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.646, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "QUAR-VLA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.147, + 0.784, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.296, + 0.785, + 0.407 + ], + "angle": 0, + "content": "Fig. 6: Tasks Successful Rate. R3M lacks alignment with text, which means that although it has some recognition capabilities, it struggles to understand actions through instructions. CLIP and the VC-1 possess the ability to align text and images, enabling them to generate and execute actions based on instructions and observations. As shown above, while R3M may have a slight advantage in easy seen tasks (distinguish), it performs poorly in more complex tasks requiring action execution, significantly lagging behind CLIP and VC-1. QUART fully leverages the advantages of large models, enabling semantic understanding of images and alignment with textual information." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.425, + 0.788, + 0.466 + ], + "angle": 0, + "content": "Table 2: Overall performance. QUART has achieved success rates far exceeding those of the baselines in tasks of all difficulty levels, especially in the most challenging crawl and unload tasks, where the baselines have no record of success." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.479, + 0.776, + 0.597 + ], + "angle": 0, + "content": "
SeenUnseen
Easy DistinguishMedium Go toHard Go avoid Go through Crawl UnloadObjectVerbal
CLIP [27]0.440.430.450.19000.110.14
R3M [25]0.580000000
VC-1 [21]0.460.430.450.31000.290.19
QUART0.660.600.530.410.320.120.350.33
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.779 + ], + "angle": 0, + "content": "1. Comparison within VLM baselines. The experiment results reveal that R3M has poor performances on tasks except distinguish. The primary reason appears to be the lack of alignment with language semantics, which hinders its ability to comprehend tasks beyond basic discrimination. In contrast, CLIP and VC-1 demonstrate satisfactory performance on fundamental perceptual tasks, such as navigation (e.g., \"go to\" commands). However, their efficacy diminishes significantly when tasks require complex mechanical movements. This observation suggests that while visual language models (VLMs) can grasp abstract principles of the world, directly applying VLMs does not readily translate to the execution of intricate physical tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "2. VLM baselines vs QUART. As is shown in Table 2, Our model has achieved optimal performance on nearly all the baseline models. In our study, we have achieved a decoder-only VLA framework. This approach diverges from the traditional single-layer MLP policy head by leveraging the sequential nature" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.128 + ], + "angle": 0, + "content": "P. Ding et al." + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.143, + 0.725, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.348, + 0.788, + 0.473 + ], + "angle": 0, + "content": "Fig. 7: Cases with unseen verbal instruction. Here are 3 verbal instructions. The words below represent the names of instructions in the training data, with the words above in bold indicating verbal instruction. When confronted with unseen instructions, the alignment between the existing language and the integration of vision and action cues within the baselines is compromised, resulting in task failure. This failure manifests in behaviors such as repetitive motion, misdirection, wrong terminate commands. Conversely, QUART, leveraging the language prowess inherited from large language models, adeptly achieves generalization under novel instructions, thereby effectuating the harmonization of vision, language, and action." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.788, + 0.597 + ], + "angle": 0, + "content": "of action generation. This allows for the implicit learning of interdependencies between different action dimensions through the use of a transformer. Consequently, while the performance gains may be marginal in simple tasks, there is a noticeable enhancement in tasks that involve complex mechanical movements. significantly improved its perceptual capabilities by incorporating commonsense from the multi-modal large model (MLLM)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.788, + 0.764 + ], + "angle": 0, + "content": "Generalization capabilities. To examine the adaptability in unanticipated scenarios, we orchestrated two primary examinations: one concentrated on unfamiliar objects, and the other on unprecedented linguistic directives. For the unfamiliar objects, we cast our gaze upon an array of circumstances: objects belonging to an identical category but exhibiting divergent textures and colors; objects from the same category but of disparate shapes; and objects with differences in shapes and textures, absent from current datasets. Turning towards the unprecedented linguistic instructions, we characterized unfamiliar instances as linguistic directives that bear the same semantics but vary in expression. For example, within the \"go to the object\" task, we gauged the directive's adaptability by employing \"navigate to target\" as the testing instruction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "1. Unseen Object. As depicted in Fig. 7, we can observe the enhanced generalization ability of our QUART model. For unseen objects, both QUART and VC-1 perform well, thanks to the advantages of the pre-trained VLM model. However, when it comes to unseen verbal descriptions, apart from QUART, all other models essentially failed. This is attributed to the inherent advantage pro" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.645, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "QUAR-VLA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.165, + 0.768, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.375 + ], + "angle": 0, + "content": "Fig. 8: QUART can extend basic skills to accommodate complex customized skills with commonsense. It shows example performances of combinational command and spatial understanding towards customized command." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.408, + 0.789, + 0.469 + ], + "angle": 0, + "content": "vided by the pre-trained Multimodal Large Language Model (MLLM), enabling the recognition of distinct variations in object types and the comprehension of subtle semantic nuances. Consequently, the model exhibits exceptional generalization performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.472, + 0.789, + 0.716 + ], + "angle": 0, + "content": "2. Unseen Verbal instruction. The training directives utilized were relatively monotonous, demonstrating that minor alterations in instruction can often perplex the model, thereby revealing an imperative for generalization across language directives. Initially, the difficulties introduced by merely replacing descriptions of skill for task generalization were investigated. Evident from Table 2 and Fig. 7 is that QUART surpasses other baseline methods, an innate advantage of large-scale multimodal models. In addition, to explore more intriguing linguistic directives and fully harness the universal capabilities of the Large Language Model (LLM), Fig. 8 presents instances where the model demonstrates a clear comprehension and application of combined directives constructed via \"first\" and \"then\". It was further noted that even in the absence of explicit directional information within the dataset, the model reliably interpreted linguistic directives which, in turn, influenced task execution accuracy. The ability to utilize language for precise control of quadruped robots facilitated by scene perception, as showcased in Fig. 8, is of immense significance. More intriguing examples and analyses can be found in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.753, + 0.465, + 0.769 + ], + "angle": 0, + "content": "4.3 Qualitative Performance" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Here, we offer the visualization results of seen and unseen tasks in Fig. 9 with WR-2. We also test QUART in other quadruped robots like Go2 with minor adjustments. More demos can be found in the supplementary materials." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "P. Ding et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.145, + 0.508, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.145, + 0.788, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.335, + 0.788, + 0.406 + ], + "angle": 0, + "content": "Fig.9: Visualization of seen and unseen tasks. This figure demonstrates the excellent performance of QUART for seen tasks in both simulated and real-world environments. The left part shows 4 instances of seen tasks in both simulation environment and real-world. The right part shows performances on unseen tasks, which contain unseen verbal, object, and background." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.436, + 0.515, + 0.452 + ], + "angle": 0, + "content": "5 Conclusion & Future Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.789, + 0.688 + ], + "angle": 0, + "content": "This paper emphasizes the significance of deploying Vision-Language-Action models on quadruped robots. We introduce the concept of Vision-Language-Action tasks for QUAdruped Robots (QUAR-VLA), which seamlessly integrates visual information and instructions from diverse modalities to generate executable actions. Specifically, QUAR-VLA focuses on two main aspects of deploying VLA models for quadruped robots: defining the action space to balance flexibility and efficiency, and addressing the scarcity of large-scale training datasets. To tackle these two questions, we present the QUART models tailored for quadruped robots and the QUARD dataset, which includes diverse tasks such as navigation and manipulation. Our extensive evaluation shows that our approach leads to performant robotic policies and enables QUART to obtain a range of emergent capabilities. This includes generalization to novel objects, the ability to interpret commands not present in the robot training data, and the ability to perform rudimentary reasoning in response to user commands." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.69, + 0.756, + 0.705 + ], + "angle": 0, + "content": "Furthermore, this work has these areas requiring further enhancements:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.84 + ], + "angle": 0, + "content": "1) Automated Data Collection for Larger Datasets. Future work can explore automated processes for data collection, involving techniques such as active learning and data augmentation can be employed to further expand the dataset size and diversity. 2) Better solution for sim2Real gap. Integrating domain adaptation methods like domain randomization, data augmentation, and more accurate modeling of physics engines will be a better solution compared with direct co-training. 3) Improving Inference Speed for Real-time Control. Future works will explore hardware acceleration techniques and model compression techniques to enable faster and more efficient execution of the models." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.645, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "QUAR-VLA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.222 + ], + "angle": 0, + "content": "This work was supported by the National Science and Technology Innovation 2030 - Major Project (Grant No. 2022ZD0208800), and NSFC General Program (Grant No. 62176215)." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.244, + 0.323, + 0.26 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.274, + 0.787, + 0.344 + ], + "angle": 0, + "content": "1. Ahn, M., Dwibedi, D., Finn, C., Arenas, M.G., Gopalakrishnan, K., Hausman, K., Ichter, B., Irpan, A., Joshi, N., Julian, R., Kirmani, S., Leal, I., Lee, E., Levine, S., Lu, Y., Leal, I., Maddineni, S., Rao, K., Sadigh, D., Sanketi, P., Sermanet, P., Vuong, Q., Welker, S., Xia, F., Xiao, T., Xu, P., Xu, S., Xu, Z.: Autort: Embodied foundation models for large scale orchestration of robotic agents (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.344, + 0.787, + 0.371 + ], + "angle": 0, + "content": "2. Bahl, S., Mendonca, R., Chen, L., Jain, U., Pathak, D.: Affordances from human videos as a versatile representation for robotics (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.787, + 0.399 + ], + "angle": 0, + "content": "3. Bavishi, R., Olsen, E., Hawthorne, C., Nye, M., Odena, A., Somani, A., Tasirlar, S.: Introducing our multimodal models (2023), https://www.adept.ai/blog/fuyu-8b" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.399, + 0.787, + 0.426 + ], + "angle": 0, + "content": "4. Belkhale, S., Cui, Y., Sadigh, D.: Hydra: Hybrid robot actions for imitation learning. In: Proceedings of the 7th Conference on Robot Learning (CoRL) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.426, + 0.787, + 0.468 + ], + "angle": 0, + "content": "5. Bharadhwaj, H., Vakil, J., Sharma, M., Gupta, A., Tulsiani, S., Kumar, V.: Roboagent: Generalization and efficiency in robot manipulation via semantic augmentations and action chunking (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.468, + 0.787, + 0.536 + ], + "angle": 0, + "content": "6. Bousmalis, K., Irpan, A., Wohlhart, P., Bai, Y., Kelcey, M., Kalakrishnan, M., Downs, L., Ibarz, J., Pastor, P., Konolige, K., et al.: Using simulation and domain adaptation to improve efficiency of deep robotic grasping. In: 2018 IEEE international conference on robotics and automation (ICRA). pp. 4243-4250. IEEE (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.536, + 0.787, + 0.661 + ], + "angle": 0, + "content": "7. Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Chen, X., Choromanski, K., Ding, T., Driess, D., Dubey, A., Finn, C., Florence, P., Fu, C., Arenas, M.G., Gopalakrishnan, K., Han, K., Hausman, K., Herzog, A., Hsu, J., Ichter, B., Irpan, A., Joshi, N., Julian, R., Kalashnikov, D., Kuang, Y., Leal, I., Lee, L., Lee, T.W.E., Levine, S., Lu, Y., Michalewski, H., Mordatch, I., Pertsch, K., Rao, K., Reymann, K., Ryoo, M., Salazar, G., Sanketi, P., Sermanet, P., Singh, J., Singh, A., Soricut, R., Tran, H., Vanhoucke, V., Vuong, Q., Wahid, A., Welker, S., Wohlhart, P., Wu, J., Xia, F., Xiao, T., Xu, P., Xu, S., Yu, T., Zitkovich, B.: Rt-2: Vision-language-action models transfer web knowledge to robotic control (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.787, + 0.716 + ], + "angle": 0, + "content": "8. Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Chen, X., Choromanski, K., Ding, T., Driess, D., Dubey, A., Finn, C., et al.: Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.716, + 0.787, + 0.841 + ], + "angle": 0, + "content": "9. Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Dabis, J., Finn, C., Gopalakrishnan, K., Hausman, K., Herzog, A., Hsu, J., Ibarz, J., Ichter, B., Irpan, A., Jackson, T., Jesmonth, S., Joshi, N.J., Julian, R., Kalashnikov, D., Kuang, Y., Leal, I., Lee, K.H., Levine, S., Lu, Y., Malla, U., Manjunath, D., Mordatch, I., Nachum, O., Parada, C., Peralta, J., Perez, E., Pertsch, K., Quiambao, J., Rao, K., Ryoo, M., Salazar, G., Sanketi, P., Sayed, K., Singh, J., Sontakke, S., Stone, A., Tan, C., Tran, H., Vanhoucke, V., Vega, S., Vuong, Q., Xia, F., Xiao, T., Xu, P., Xu, S., Yu, T., Zitkovich, B.: Rt-1: Robotics transformer for real-world control at scale (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.274, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "P. Ding et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "10. Caluwaerts, K., Iscen, A., Kew, J.C., Yu, W., Zhang, T., Freeman, D., Lee, K.H., Lee, L., Saliceti, S., Zhuang, V., et al.: Barkour: Benchmarking animal-level agility with quadruped robots. arXiv preprint arXiv:2305.14654 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.217 + ], + "angle": 0, + "content": "11. Eckert, P., Ijspeert, A.J.: Benchmarking agility for multilegged terrestrial robots. IEEE Transactions on Robotics 35(2), 529-535 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.217, + 0.785, + 0.271 + ], + "angle": 0, + "content": "12. Gu, J., Kirmani, S., Wohlhart, P., Lu, Y., Arenas, M.G., Rao, K., Yu, W., Fu, C., Gopalakrishnan, K., Xu, Z., Sundaresan, P., Xu, P., Su, H., Hausman, K., Finn, C., Vuong, Q., Xiao, T.: Rt-trajectory: Robotic task generalization via hindsight trajectory sketches (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.785, + 0.297 + ], + "angle": 0, + "content": "13. Haldar, S., Mathur, V., Yarats, D., Pinto, L.: Watch and match: Supercharging imitation with regularized optimal transport. CoRL (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.297, + 0.785, + 0.366 + ], + "angle": 0, + "content": "14. Hutter, M., Gehring, C., Jud, D., Lauber, A., Bellicoso, C.D., Tsounis, V., Hwangbo, J., Bodie, K., Fankhauser, P., Bloesch, M., Diethelm, R., Bachmann, S., Melzer, A., Hoepflinger, M.: Anymal - a highly mobile and dynamic quadrupedal robot. In: 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 38-44 (2016). https://doi.org/10.1109/IROS.2016.7758092" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.366, + 0.785, + 0.407 + ], + "angle": 0, + "content": "15. Jiang, Y., Moseson, S., Saxena, A.: Efficient grasping from rgbd images: Learning using a new rectangle representation. In: 2011 IEEE International conference on robotics and automation. pp. 3304-3311. IEEE (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.407, + 0.785, + 0.46 + ], + "angle": 0, + "content": "16. Kareer, S., Yokoyama, N., Batra, D., Ha, S., Truong, J.: Vinl: Visual navigation and locomotion over obstacles. 2023 IEEE International Conference on Robotics and Automation (ICRA) pp. 2018-2024 (2022), https://api-semanticscholar.org/CorpusID:253117178" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.46, + 0.785, + 0.514 + ], + "angle": 0, + "content": "17. Karnan, H., Nair, A., Xiao, X., Warnell, G., Pirk, S., Toshev, A., Hart, J., Biswas, J., Stone, P.: Socially compliant navigation dataset (scand): A large-scale dataset of demonstrations for social navigation. IEEE Robotics and Automation Letters 7(4), 11807-11814 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.514, + 0.785, + 0.556 + ], + "angle": 0, + "content": "18. Lee, K.H., Nachum, O., Yang, M.S., Lee, L., Freeman, D., Guadarrama, S., Fischer, I., Xu, W., Jang, E., Michalewski, H., et al.: Multi-game decision transformers. Advances in Neural Information Processing Systems 35, 27921-27936 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.556, + 0.785, + 0.61 + ], + "angle": 0, + "content": "19. Lee, M.A., Zhu, Y., Srinivasan, K., Shah, P., Savarese, S., Fei-Fei, L., Garg, A., Bohg, J.: Making sense of vision and touch: Self-supervised learning of multimodal representations for contact-rich tasks. In: 2019 IEEE International Conference on Robotics and Automation (ICRA) (2019), https://arxiv.org/abs/1810.10191" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.61, + 0.785, + 0.651 + ], + "angle": 0, + "content": "20. Li, X., Liu, M., Zhang, H., Yu, C., Xu, J., Wu, H., Cheang, C., Jing, Y., Zhang, W., Liu, H., et al.: Vision-language foundation models as effective robot imitators. arXiv preprint arXiv:2311.01378 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.651, + 0.785, + 0.704 + ], + "angle": 0, + "content": "21. Majumdar, A., Yadav, K., Arnaud, S., Ma, Y.J., Chen, C., Silwal, S., Jain, A., Berges, V.P., Abbeel, P., Malik, J., Batra, D., Lin, Y., Maksymets, O., Rajeswaran, A., Meier, F.: Where are we in the search for an artificial visual cortex for embodied intelligence? (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.704, + 0.785, + 0.746 + ], + "angle": 0, + "content": "22. Makoviychuk, V., Wawrzyniak, L., Guo, Y., Lu, M., Storey, K., Macklin, M., Hoeller, D., Rudin, N., Allshire, A., Handa, A., State, G.: Isaac gym: High performancegpu-based physics simulation for robot learning (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.746, + 0.785, + 0.8 + ], + "angle": 0, + "content": "23. Margolis, G.B., Agrawal, P.: Walk these ways: Tuning robot control for generalization with multiplicity of behavior. In: Liu, K., Kulic, D., Ichnowski, J. (eds.) Proceedings of The 6th Conference on Robot Learning. Proceedings of Machine Learning Research, vol. 205, pp. 22-31. PMLR (14-18 Dec 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.8, + 0.785, + 0.841 + ], + "angle": 0, + "content": "24. Nair, S., Rajeswaran, A., Kumar, V., Finn, C., Gupta, A.: R3m: A universal visual representation for robot manipulation. In: Conference on Robot Learning (2022), https://api-semanticscholar.org/CorpusID:247618840" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.646, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "QUAR-VLA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "25. Nair, S., Rajeswaran, A., Kumar, V., Finn, C., Gupta, A.: R3m: A universal visual representation for robot manipulation (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.203 + ], + "angle": 0, + "content": "26. Pari, J., Shafiullah, N.M., Arunachalam, S.P., Pinto, L.: The surprising effectiveness of representation learning for visual imitation (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.788, + 0.245 + ], + "angle": 0, + "content": "27. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.245, + 0.788, + 0.286 + ], + "angle": 0, + "content": "28. Reed, S., Zolna, K., Parisotto, E., Colmenarejo, S.G., Novikov, A., Barth-Maron, G., Gimenez, M., Sulsky, Y., Kay, J., Springenberg, J.T., et al.: A generalist agent. arXiv preprint arXiv:2205.06175 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.327 + ], + "angle": 0, + "content": "29. Reed, S., Zolna, K., Parisotto, E., Colmenarejo, S.G., Novikov, A., Barth-Maron, G., Gimenez, M., Sulsky, Y., Kay, J., Springenberg, J.T., et al.: A generalist agent. arXiv preprint arXiv:2205.06175 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.327, + 0.788, + 0.383 + ], + "angle": 0, + "content": "30. Schiavi, G., Wulkop, P., Rizzi, G., Ott, L., Siegwart, R., Chung, J.J.: Learning agent-aware affordances for closed-loop interaction with articulated objects. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 5916-5922 (2023). https://doi.org/10.1109/ICRA48891.2023.10160747" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.383, + 0.788, + 0.411 + ], + "angle": 0, + "content": "31. Shah, D., Sridhar, A., Dashora, N., Stachowicz, K., Black, K., Hirose, N., Levine, S.: Vint: A foundation model for visual navigation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.411, + 0.788, + 0.439 + ], + "angle": 0, + "content": "32. Shilane, P., Min, P., Kazhdan, M., Funkhouser, T.: The princeton shape benchmark. In: Proceedings Shape Modeling Applications, 2004. pp. 167-178 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.466 + ], + "angle": 0, + "content": "33. Shridhar, M., Manuelli, L., Fox, D.: CIoport: What and where pathways for robotic manipulation. In: Conference on Robot Learning. pp. 894-906. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.466, + 0.788, + 0.508 + ], + "angle": 0, + "content": "34. Stone, A., Xiao, T., Lu, Y., Gopalakrishnan, K., Lee, K.H., Vuong, Q., Wohlhart, P., Kirmani, S., Zitkovich, B., Xia, F., Finn, C., Hausman, K.: Open-world object manipulation using pre-trained vision-language models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.55 + ], + "angle": 0, + "content": "35. Szot, A., Schwarzer, M., Agrawal, H., Mazoure, B., Talbott, W., Metcalf, K., Mackraz, N., Hjelm, D., Toshev, A.: Large language models as generalizable policies for embodied tasks. arXiv preprint arXiv:2310.17722 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.55, + 0.788, + 0.577 + ], + "angle": 0, + "content": "36. Tang, Y., Yu, W., Tan, J., Zen, H., Faust, A., Harada, T.: Saytap: Language to quadrupedal locomotion. arXiv preprint arXiv:2306.07580 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.577, + 0.788, + 0.618 + ], + "angle": 0, + "content": "37. Wei, J., Hou, L., Lampinen, A., Chen, X., Huang, D., Tay, Y., Chen, X., Lu, Y., Zhou, D., Ma, T., et al.: Symbol tuning improves in-context learning in language models. arXiv preprint arXiv:2305.08298 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.618, + 0.788, + 0.66 + ], + "angle": 0, + "content": "38. Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., Xiao, J.: 3d shapenets: A deep representation for volumetric shapes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1912-1920 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.714 + ], + "angle": 0, + "content": "39. Yang, R., Zhang, M., Hansen, N., Xu, H., Wang, X.: Learning vision-guided quadrupedal locomotion end-to-end with cross-modal transformers. ArXiv abs/2107.03996 (2021), https://api_semanticscholar.org/CorpusID: 235765481" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.714, + 0.788, + 0.772 + ], + "angle": 0, + "content": "40. Yu, K.T., Bauza, M., Fazeli, N., Rodriguez, A.: More than a million ways to be pushed. a high-fidelity experimental dataset of planar pushing. In: 2016 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 30-37. IEEE (2016)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.772 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/QUAR-VLA_ Vision-Language-Action Model for Quadruped Robots/7817397f-b04e-4f74-8ceb-92291da9f1da_origin.pdf b/2024/QUAR-VLA_ Vision-Language-Action Model for Quadruped Robots/7817397f-b04e-4f74-8ceb-92291da9f1da_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5b42c654c6a608b7e88243d64c27835243bced38 --- /dev/null +++ b/2024/QUAR-VLA_ Vision-Language-Action Model for Quadruped Robots/7817397f-b04e-4f74-8ceb-92291da9f1da_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd38d8402381502304165b6359fbb8a6e27fbf5b97436506faa3f7eebd035bca +size 3920228 diff --git a/2024/QUAR-VLA_ Vision-Language-Action Model for Quadruped Robots/full.md b/2024/QUAR-VLA_ Vision-Language-Action Model for Quadruped Robots/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a66fbfdd3611ba5693cc5c628d862fb186f55cb1 --- /dev/null +++ b/2024/QUAR-VLA_ Vision-Language-Action Model for Quadruped Robots/full.md @@ -0,0 +1,241 @@ +# QUAR-VLA: Vision-Language-Action Model for Quadruped Robots + +Pengxiang Ding $^{12}$ Han Zhao $^{12}$ Wenjie Zhang $^{3}$ Wenxuan Song $^{4}$ Min Zhang $^{12}$ Siteng Huang $^{12}$ Ningxi Yang $^{2}$ Donglin Wang $^{1\star}$ + +1Zhejiang University 2Westlake University 3Beijing Normal University 4Monash University {dingpengxiang}@westlake.edu.cn + +Abstract. The important manifestation of robot intelligence is the ability to naturally interact and autonomously make decisions. Traditional quadruped robot learning typically handles language interaction and visual autonomous perception separately, which, while simplifying system design, also limits the synergy between different information streams. This separation poses challenges in achieving seamless autonomous reasoning, decision-making, and action execution. To address these limitations, a novel paradigm, named Vision-Language-Action tasks for QUAdruped Robots (QUAR-VLA), has been introduced in this paper. This approach tightly integrates visual information and instructions to generate executable actions, effectively merging perception, planning, and decision-making. The central idea is to elevate the overall intelligence of the robot. Within this framework, a notable challenge lies in aligning fine-grained instructions with visual perception information. This emphasizes the complexity involved in ensuring that the robot accurately interprets and acts upon detailed instructions in harmony with its visual observations. Consequently, we propose QUAdruped Robotic Transformer (QUART), a VLA model to integrate visual information and instructions from diverse modalities as input and generates executable actions for real-world robots and present QUAdruped Robot Dataset (QUARD), a large-scale multi-task dataset including perception, navigation and advanced capability like whole-body manipulation tasks for training QUART model. Our extensive evaluation shows that our approach leads to performant robotic policies and enables QUART to obtain a range of generalization capabilities. + +Keywords: Robotics $\cdot$ Quadruped Robot Learning $\cdot$ Vision-Language-Action Model + +# 1 Introduction + +Quadruped robots, characterized by their excellent traversability on complex terrains and agile movements, have garnered significant attention in the field of robotics [14]. Researchers have extensively employed these robots to explore tasks encompassing autonomous navigation and manipulation [16, 17, 36]. + +![](images/e8766da15dc3f80c3b4234c895c71db6552a514e54fe341da2e6b7bc2e650210.jpg) +(a) QUAR-VA + +![](images/4caa3042167d5cba3571d471f8b0f66ab4dd32381d27f1a51bbd1177cb1ff5c4.jpg) +(b) QUAR-LA + +![](images/8d7565c4199cb5db8bc9ed69df7ff4c547878b0b524cf732423968de2e6cc4cb.jpg) +(c) QUAR-VLA +Fig. 1: Comparison of QUAR-VA, QUAR-LA, and QUAR-VLA. QUAR-VA solely utilizes coarse-grained vision information, lacking explicit instructions for handling diverse tasks. In contrast, QUAR-LA exclusively relies on language information and lacks of vision information for autonomy. Therefore, QUART-VLA combines both vision information and language instructions as inputs, enabling autonomous problem-solving across a range of tasks, revealing distinct input modalities and task capabilities. + +Broadly speaking, the quadruped tasks consist of two major specifications: Vision-Action tasks for QUAdruped Robots (QUAR-VA) and Language-Action tasks for QUAdruped Robots (QUAR-LA). As depicted in Fig. 1, in QUAR-VA approaches [31], quadruped robots {receive} perception images captured from a first-person perspective and instruction images obtained from a third-person perspective to guide their actions. However, such a task specification often relies on a single (coarse-grained) goal image instruction, making it difficult to apply in many real-world combination tasks, i.e. requiring combining multiple sub-instructions. In contrast, employing language as instructions [36], QUAR-LA formulation allows for executing more fine-grained and diverse tasks, suitable for expressing combinational commands (i.e. "before / then"), complex spatial relationships (i.e. "move to the left"), and many commonsense command priors (i.e. "move fast"). Nevertheless, QUAR-LA approaches, which lack the integration of visual modality, hinder the robots' ability to perceive the environment, thereby impeding their autonomous navigation capabilities [36]. To enable quadruped robots to autonomously navigate and manipulate various tasks, in this paper, we propose a new paradigm: Vision-Language-Action tasks for QUAdruped Robots (QUAR-VLA), integrating visual information and instructions from diverse modalities as input and generating executable actions for real-world robots. + +This task primarily encompasses two challenges. Firstly, there is a lack of large-scale datasets in the research community for quadruped robots performing diverse tasks. Pretraining models require abundant trajectory data from a variety of tasks. However, collecting substantial amounts of real-world data often necessitates human involvement and significant time investment in manual instruction operations. + +Secondly, different from navigation tasks [31] for general mobile robots and manipulation tasks for fixed-base systems [7, 9, 12, 34], building a VLA model + +![](images/36671b8685ef7c299dd0f8a023a98aea7645b7abed95670b6ee402114d5c9a11.jpg) +Fig. 2: Overview of QUAR-VLA. Our tasks encompass a diverse range of perception, navigation, and other advanced capability. The Vision-Language-Action (VLA) model first undergoes training with a huge amount of simulation data (246K episodes) and a small amount of real-world data (3K episodes). In the inference phase, images and texts undergo tokenization, after which QUART generates 12-dimensional action tokens. These tokens are subsequently detokenized into valid robot actions and deployed on a physical quadruped robot. This methodology effectively extends the learned capabilities from a simulated environment to real-world applications. + +to solve complex quadruped robot tasks is considerably more challenging due to their agile locomotion behaviors. The action space needs to be properly defined to strike a balance between movement flexibility and computational efficiency. The action generated by the model should not be too simplistic, akin to the planner base velocities output by 2-D navigation modules, nor should they require a high execution frequency like the low-level motion policy that directly controls the joint motors. + +To address these two problems, we collect a large-scale multi-task dataset QUAdruped Robot Dataset (QUARD). It includes multiple tasks such as perception, navigation, and advanced capabilities like object avoidance. To the best of our knowledge, this is the first quadruped robot dataset that incorporates a significant amount of vision, language instruction, and robot command data. As collecting data on real robots is expensive and inefficient, we primarily rely on data generated in simulation, which exhibits significant differences in visual, sensor, and system dynamics. We also introduce QUAdruped Robotic Transformer (QUART), a VLA model for training QUARD. QUART takes images of the robot's first-view camera and the natural language instruction as inputs and generates rich control commands that include controlling the robot's base velocity, posture, and gait parameters. QUART leverages a pre-trained large-scale visual language model and fine-tunes it on our dataset to enable the generation of executable commands for quadruped robots. To address the simple-to-real gap caused by the data disparity, we construct a co-training pipeline to effectively distill the knowledge of simulation data for real-scene deployment. Our extensive evaluation shows that our approach leads to performant robotic policies and enables QUART to obtain a range of generalization capabilities. + +Our contributions are as follows: 1) To the best of our knowledge, we first propose a new paradigm: QUAR-VLA, integrating visual and language instructions to executable actions for more autonomous and diverse quadruped robot tasks. 2) We present a large-scale multi-task dataset, QUARD, and a Vision-Language-Action model, QUART to solve the QUAR-VLA tasks. 3) Our extensive evaluation shows that our approach leads to performant robotic policies and enables QUART to obtain a range of generalization capabilities. + +# 2 Related Work + +Quadruped Robot Learning. Previous research [2,4,10,13,16,17,19,26,30,36, 39] has delved into single-model policies. As to vision-based methods (QUAR-VA), Kareer et al. [16] utilizes vision perception to predict a privileged terrain map, enabling the robot to perform specific locomotion based on the terrain characteristics. Karnan [17] leverages ego-motion estimates obtained through vision perception, allowing the robot to understand its relative position and direction to avoid collisions and navigate toward a goal location. While some works have succeeded in these tasks through vision perception, relying solely on vision limits the capability to handle only a specific type of task. In response to this limitation, Tang et al. [36] have started exploring language-conditioned tasks (QUAR-LA). However, the absence of vision perception in their approach results in a lack of autonomy for the robot to interact with its environment. Therefore, this paper proposes a new QUAR-VLA paradigm to integrate visual information and instructions to generate executable actions, effectively ensuring that the robot accurately interprets and acts upon detailed instructions in harmony with its visual observations. + +Vision-language-action models. Visual-language-action models [1, 5, 20, 24, 28, 33, 35] integrates visual information and instructions to generate executable actions, effectively merging perception, planning, and decision-making and elevating the overall intelligence of the robot. As a result, vision-language-action models have been a popular area of research in robotics. In the realm of visual-language-action models, Shridhar et al. [33] have made significant strides by leveraging CLIP to encode text inputs, thereby enhancing the vision model's semantic comprehension and action execution capabilities. Further advancing the field, Brohan et al. [8] and Li et al. [20] have innovatively applied large language models to craft manipulation policies specifically for robotic applications. This paper delves into the domain of embodied agents, focusing on the quadruped robot as a case study. It explores the challenges and opportunities in enabling quadruped robots to autonomously navigate and perform a variety of tasks as directed by human instructions, contributing to the broader discourse on robotic autonomy and intelligence. + +Robot Dataset. The realm of robotic learning has been at the forefront of advancing open-source datasets tailored explicitly for robot learning [6,15,17,32, 36,38,40]. Within the domain of quadrupedal robots, previous works on datasets primarily focused on algorithms and leg controllers. Eckert et al. [11] introduced + +a grading system and an online dataset for the collection and distribution of agility scores. Caluwaerts et al. [10] sought to address the lack of a standardized and scalable environment and inductive metrics. Tang et al. [36] pioneered the generation of desired joint position control using LLM. Nevertheless, common limitations persist among these quadruped robot datasets: the quantity is insufficient and the tasks are limited in diversity. Hence, our data repository aims to complement these endeavors. We curate and process an extensive array of skills on various objects, across diverse scenes in both the real world and virtual environments. The dataset includes image data, robot action data, and point cloud data, providing quadruped robots with rich perceptual information to tackle increasingly intricate and demanding tasks. + +# 3 Method + +This section provides a detailed exposition of our proposed methodology. Initially, we present the definition of our proposed QUAR-VLA in Section 3.1. Following that, Section 3.2 outlines the collecting process of QUARD. Lastly, we delve into the overarching architecture of QUART in Section 3.3. + +# 3.1 Problem Setup + +The objective of QUAR-VLA is to construct a vision-language-action model learned from large-scale demonstration data and generate actions for closed-loop robot control. + +An overview of our QUAR-VLA is shown in Fig. 2. Our goal is to train a conditional policy QUART that can interpret RGB image(s), denoted $s \in S$ , together with a task instruction $w \in \mathcal{W}$ , which correspond to a language string. The policy is a mapping from images and instructions to actions, and can be written as $\mu : S \times \mathcal{W} \to \mathcal{A}$ , where the action space $\mathcal{A}$ consists of the 11-dimensional high-level commands as well as a terminate signal. + +$$ +\left[ v _ {x}, v _ {y}, \omega_ {z}, \theta_ {1}, \theta_ {2}, \theta_ {3}, f, h _ {z}, \phi , s _ {y}, h _ {z} ^ {f}, t \right] \tag {1} +$$ + +Here, $v_{x}, v_{y}$ , and $\omega_{z}$ represent the velocities along the x-axis, y-axis, and z-axis respectively. $\theta_{1}, \theta_{2}$ , and $\theta_{3}$ indicate the gait pattern, $f$ denotes the frequency, $h_{z}$ represents the height of the robot, $\phi$ denotes the pitch angle, $s_{y}$ corresponds to the foot width, $h_{z}^{f}$ represents the foot height, and $t$ indicates the termination signal of the action. + +Notably, in this study, we employ the discretization method proposed by [9] to discretize all continuous dimensions into 256 uniformly sized bins. This choice of discretization is motivated by the desire to reduce the complexity of action search and improve the stability and convergence of algorithms. + +# 3.2 Large-scale Quadruped Robot Datasets + +To enable an imitation learning system to generalize to new tasks with zero demonstrations of said task, we must be able to easily collect a diverse dataset, + +Table 1: Tasks Definition. The "Type" means different capabilities of robots. The "Level" devises the difficulty into 3 levels. The "Skill" means different skill/task categories. The "Episode" signifies the number of experiments conducted for each task, which also corresponds to the number of trajectories. The "Description" is the illustration of the tasks. + +
TypeLevelSkillEpisodeDescription
PerceptionEasyDistinguish (sim)Letter10KIdentify the correct one from multiple boxes with different printed letters
BasicMediumGo to Object (sim)72KNavigate to the object and stop in front of it
NavigationMediumGo to Object (real)3KNavigate to the object and stop in front of it
Advanced CapabilityHardGo through Tunnel48KSpatial Navigation: Go through the correct tunnel from two tunnels with different colors and shapes
HardGo to Object and avoid the obstacle (sim)63KObstacle Avoidance: Navigate to the object without colliding with the obstacle
HardCrawl under Bar (sim)1KEnvironment Adaptation: Crawl under a bar with a low height
HardUnload Object into Re-ceptacle (sim)52KObject Manipulation: Move with a ball on the back and unload it into a receptacle
Total249KThe total number of episodes
+ +provide corrective feedback, and evaluate many tasks at scale. Therefore, we collect a large-scale multi-task dataset, QUAdruped Robot Dataset (QUARD), which includes multiple tasks such as perception, basic navigation, and advanced capability like object manipulation. + +Task Definition. As is shown in Table 1, there are seven tasks distributed across three levels of difficulty: easy, medium, and hard. The easy and medium tasks are designed such that the robot can accomplish them using fundamental skills. For instance, the "Go to" task is a basic action integral to all navigation tasks, while the "Distinguish" task serves as the foundational action for all perception tasks. Hard tasks necessitate the robot to execute tasks using more advanced skills. A case in point is the "Unload object" task, which involves the robot distinguishing the target container, navigating to the target position, adjusting body position, and considering potential collisions to successfully unload the object. This process integrates perception, navigation, and whole-body manipulation. The complexity of tasks is reflected in the average trajectory length, with more challenging tasks requiring a greater number of steps and consequently taking longer to complete. The specific trajectory lengths and task distribution for each task are detailed in Fig. 3 and Fig. 4. + +System Setup. The robot used to collect the trajectory data is WR-2, which is a quadruped robot with 12 joints. WR-2 has around $25\mathrm{cm}$ in standing height and $40\mathrm{cm}$ body length. The command output is sent to the low-level command tracking controller (pre-trained command-conditioned policy in [23]) to generate the actual joint action of the robot. The simulation data was collected in Nvidia's Isaac Gym [22], a powerful simulator that allows us to collect massive robot trajectories in parallel. The real data was collected by manually manipulating a quadruped robot in the lab setting. The perception data is provided by a RealSense d435 camera installed in the front of the robot. + +Data Collection. As to simulation data, $\mathbf{A}^*$ and $\mathbf{D}^*$ algorithms are used to plot optimal paths for a dog navigating through various objects and obstacles. The + +![](images/945e96e3aad707c6432044216c62843e8e774f84b5812c9f4efecbb20bc7624f.jpg) +Fig. 3: The left figure illustrates the trajectory lengths corresponding to different tasks and the right figure illustrates the relationships between tasks. As the difficulty of the skill increases, it can be observed that the average trajectory length gradually increases. The right figure demonstrates the relationship between the types of tasks: perception as foundational capability; basic navigation ability is built on perception; object manipulation, obstacle avoidance, spatial navigation, and environment adaption extend from perception and navigation. + +![](images/1af447d773b093233f4d6693c985b85bb02d0a99a77e2dba13a4e4129c92115c.jpg) + +$\mathrm{A}^*$ algorithm seeks the most cost-effective path, while the $\mathrm{D}^*$ algorithm adapts to changes in real time. A Proportional-Derivative controller then converts these paths into target velocities for the dog, ensuring smooth movement. The combination of both algorithms provides a flexible and efficient method for path planning in various environments. For real data, The real data set is obtained in a laboratory environment using remote control. + +Consistency constraints. To maintain consistency between simulation and reality, we've established constraints for data collection. The robot starts at the origin, with the target randomly positioned within [2.7,3.3] meters on the x-axis and [0.9,1.1] meters on the y-axis. In obstacle scenarios, obstacles are placed 1.5 meters from the target's x-coordinate, with the same y-coordinate. Task termination criteria vary. For "go to object", "go to the object and avoid obstacle", and "crawl under bar" tasks, success is when the robot is less than one meter from the target. Other task success criteria are: 1) "Unload object": successful when the object is in the target container. 2) "Go through tunnel" and "Crawl under bar": successful when the robot reaches a specific position behind the tunnel or bar. 3) "Distinguish letter": successful when the robot correctly orients towards the visual target. + +More Statistics. The data collection setup adheres to a predefined language template as depicted in Fig. 4, where the task, target object, speed, and gait are explicitly defined. To enhance diversity and generalization capabilities, the dataset includes a variety of common indoor furniture and outdoor facilities, in addition to basic shapes. The basic shapes come in four color variations: green, red, blue, and yellow. Examples of indoor furniture include bookshelves, ovens, and vases, while outdoor facilities encompass trashcans and benches. Recognizing + +![](images/508d4d28b3e1f2022b7f49ec06e00e477cd54a2c5d9f82c38e8bc1b0933bafbb.jpg) +Fig. 4: Statistic analysis of QUARD. 1) Simulation data accounts for a larger proportion compared to real data. 2) Trotting occupied most of the share of gait. 3) Most tasks own similar episodes while Distinguish Letter and Go to Object (in reality) owns much less. 4) 3 types of speed occupied $1/3$ separately. + +the distinct nature of motion robots as compared to manipulation robots, we also require robots to perform the same task with varying gaits and speeds to adapt to different instructions and environments. The corresponding ratios of gaits and speeds are presented in Fig. 4. More details about the datasets can be found in the supplementary material. + +# 3.3 Vision-Language-Action Model + +In comparison to VLA models [7,9,12,31,34] in other field, our research stands out due to the abundance of kinematics information available in this domain. This allows the robot to not only excel at goal-oriented navigation tasks but also perform a diverse range of whole-body manipulation operations with enhanced flexibility in gaits and body control. Next, we will present our QUAdruped Robotic Transformers (QUART). + +QUART relies on a pre-trained vision-language-model [3] to associate tokens from the model's existing tokenization with the discrete action bins. It is worth noting that training MLLMs to override existing tokens with action tokens is a form of symbol tuning [37], which has been shown to work well in prior work. For [3], integers up to 1000 each have a unique token, so we simply associate the action bins with the token representing the corresponding integer. Notably, QUART model takes a single image $s$ and a natural language instruction $w$ as input, which are first converted into corresponding tokens $t$ through a tokenizer $\tau(t|s,w)$ and fed into a decoder-only transformer module to obtain discretized action tokens $p(a_d|t)$ . + +The policy QUART could be shown as follow: + +$$ +\operatorname {Q U A R T} \left(a _ {d} | s, w\right) = p \left(a _ {d} | t\right) \tau (t | s, w) \tag {2} +$$ + +where $w, s$ are the input images and language instruction and $\tau$ represents the tokenizer and $p$ indicates the vision-language model to output action $a_{d}$ . + +![](images/1853b291f835ff48236e735d4a90624feb41154f530194ce97b5a5607213a1cf.jpg) +Fig. 5: Architecture of QUART. It is designed to leverage the scene comprehension capability of a pretrained MLLM. It receives visual information as observation, and outputs an action representing the actual action taken by the robot based on text-form instructions, and de-tokenizes it into specific action values. QUART can generate a complete action sequence at a processing rate of $2\mathrm{Hz}$ in actual scenarios, and hand it over to the underlying low-level strategy for execution. + +Action Detokenize. To directly convert models' output to valid robot actions for downstream control, we need detokenize the discrete action token $a_{d}$ into continuous representation $a_{c}$ (except for the discrete termination command). + +$$ +a _ {c} = \operatorname {D e t o k e n i z e} \left(a _ {d}\right) \tag {3} +$$ + +Loss. We use a standard categorical cross-entropy objective and causal masking that was utilized in prior Transformer-based controllers [18, 29]. + +Inference Speed. In contrast to many applications of large models, such as natural language or image generation, one of the unique requirements for a model that needs to run on real robots in real-time is fast and consistent inference speed. It is important to highlight that our research specifically emphasizes the command control of quadruped robots. In contrast to low-level motor control, the command control outputs alleviate the strict control frequency requirements, enabling smooth integration of larger models and unlocking improved reasoning capabilities. For QUART, the inference time could get 2Hz. More details about the deployment could be found in the supplementary material. + +# 4 Experiments + +We concentrate our experiments on the multi-task ability and generalization. We try to address the following questions: 1. How effective is the vision-language- + +action architecture for multi-task quadruped task compared to previous VLM baselines? 2. How well do these models generalize to unseen semantic attributes like object shape, color, and unseen verbal information? + +# 4.1 Implementation Details + +Training Details. We train a specific instantiation of QUART, derived from 8B pre-trained VLM models [3] for superior performance. And we use learning rate 2e-5 and batch size 256 to fine-tune the model for 100K gradient steps. Both models are trained with the next token prediction objective, which corresponds to the behavior cloning loss in robot learning. + +Evaluation Details. We follow the standard robot evaluation metrics [7, 9], success rate (SR), to evaluate the overall performance. The signal we used for robot control is the 11-dimensional command information with the action space outlined in Section 3.1 and 1 stop terminal command means the end of action. More details for each experiment can be seen in the supplementary material. + +Baselines. Considering the absence of VLA models work on quadruped robots at present, we have taken the following baselines into account for a fair comparison: CLIP [27], R3M [25], and VC-1 [21]. + +1. CLIP [27] uniquely encodes both textual and visual data, fusing these embeddings to create feature sets that integrate textual and visual information. +2. R3M [25]is a visual representation model derived from the Ego4D dataset, which acts as a unified, static perception module for policy learning applications. +3. VC-1 [21] pushes the boundaries of visual representation learning by amalgamating data from multiple datasets and assessing performance across a diverse range of tasks via the implementation of CortexBench. + +Since the aforementioned models do not include language conditioning, we include this aspect by separately embedding the language command, allowing us to compare it to our method. Specifically, we concatenate the resulting language embedding tokens with the image tokens generated by the vision model and pass the combined token sequences through a policy head to produce action outputs. + +# 4.2 Overall Performance + +Multi-task Performance. To assess the overall capabilities of VLA models on multiple tasks, we conduct evaluations using instructions randomly selected from the training set. It's important to note that this evaluation introduces variations in the placement of objects and other setup factors (such as robot position). This variation demands the system's ability to generalize effectively to realistic environmental variability. In total, over 1500 episodes are tested in this evaluation, comprising 425 episodes for going to objects, 500 for going to objects without colliding with the obstacle. 150 for going through the tunnel, 100 for unloading objects, 100 for distinguishing objects, and 75 for crawling under the bar. + +![](images/96f36b3e20974a8a02fa104f0858003730dbf647a2d6c5989abb1496d473009b.jpg) +Fig. 6: Tasks Successful Rate. R3M lacks alignment with text, which means that although it has some recognition capabilities, it struggles to understand actions through instructions. CLIP and the VC-1 possess the ability to align text and images, enabling them to generate and execute actions based on instructions and observations. As shown above, while R3M may have a slight advantage in easy seen tasks (distinguish), it performs poorly in more complex tasks requiring action execution, significantly lagging behind CLIP and VC-1. QUART fully leverages the advantages of large models, enabling semantic understanding of images and alignment with textual information. + +Table 2: Overall performance. QUART has achieved success rates far exceeding those of the baselines in tasks of all difficulty levels, especially in the most challenging crawl and unload tasks, where the baselines have no record of success. + +
SeenUnseen
Easy DistinguishMedium Go toHard Go avoid Go through Crawl UnloadObjectVerbal
CLIP [27]0.440.430.450.19000.110.14
R3M [25]0.580000000
VC-1 [21]0.460.430.450.31000.290.19
QUART0.660.600.530.410.320.120.350.33
+ +1. Comparison within VLM baselines. The experiment results reveal that R3M has poor performances on tasks except distinguish. The primary reason appears to be the lack of alignment with language semantics, which hinders its ability to comprehend tasks beyond basic discrimination. In contrast, CLIP and VC-1 demonstrate satisfactory performance on fundamental perceptual tasks, such as navigation (e.g., "go to" commands). However, their efficacy diminishes significantly when tasks require complex mechanical movements. This observation suggests that while visual language models (VLMs) can grasp abstract principles of the world, directly applying VLMs does not readily translate to the execution of intricate physical tasks. +2. VLM baselines vs QUART. As is shown in Table 2, Our model has achieved optimal performance on nearly all the baseline models. In our study, we have achieved a decoder-only VLA framework. This approach diverges from the traditional single-layer MLP policy head by leveraging the sequential nature + +![](images/4c3f62ca0c7afa00027d8c4e2ec78126493344173709a1dbc89f6fc349d0d549.jpg) +Fig. 7: Cases with unseen verbal instruction. Here are 3 verbal instructions. The words below represent the names of instructions in the training data, with the words above in bold indicating verbal instruction. When confronted with unseen instructions, the alignment between the existing language and the integration of vision and action cues within the baselines is compromised, resulting in task failure. This failure manifests in behaviors such as repetitive motion, misdirection, wrong terminate commands. Conversely, QUART, leveraging the language prowess inherited from large language models, adeptly achieves generalization under novel instructions, thereby effectuating the harmonization of vision, language, and action. + +of action generation. This allows for the implicit learning of interdependencies between different action dimensions through the use of a transformer. Consequently, while the performance gains may be marginal in simple tasks, there is a noticeable enhancement in tasks that involve complex mechanical movements. significantly improved its perceptual capabilities by incorporating commonsense from the multi-modal large model (MLLM). + +Generalization capabilities. To examine the adaptability in unanticipated scenarios, we orchestrated two primary examinations: one concentrated on unfamiliar objects, and the other on unprecedented linguistic directives. For the unfamiliar objects, we cast our gaze upon an array of circumstances: objects belonging to an identical category but exhibiting divergent textures and colors; objects from the same category but of disparate shapes; and objects with differences in shapes and textures, absent from current datasets. Turning towards the unprecedented linguistic instructions, we characterized unfamiliar instances as linguistic directives that bear the same semantics but vary in expression. For example, within the "go to the object" task, we gauged the directive's adaptability by employing "navigate to target" as the testing instruction. + +1. Unseen Object. As depicted in Fig. 7, we can observe the enhanced generalization ability of our QUART model. For unseen objects, both QUART and VC-1 perform well, thanks to the advantages of the pre-trained VLM model. However, when it comes to unseen verbal descriptions, apart from QUART, all other models essentially failed. This is attributed to the inherent advantage pro + +![](images/5cc5a024383e109948eef3cd9e0291c8e7d24af6f2277aecef3bb37b8c84147d.jpg) +Fig. 8: QUART can extend basic skills to accommodate complex customized skills with commonsense. It shows example performances of combinational command and spatial understanding towards customized command. + +vided by the pre-trained Multimodal Large Language Model (MLLM), enabling the recognition of distinct variations in object types and the comprehension of subtle semantic nuances. Consequently, the model exhibits exceptional generalization performance. + +2. Unseen Verbal instruction. The training directives utilized were relatively monotonous, demonstrating that minor alterations in instruction can often perplex the model, thereby revealing an imperative for generalization across language directives. Initially, the difficulties introduced by merely replacing descriptions of skill for task generalization were investigated. Evident from Table 2 and Fig. 7 is that QUART surpasses other baseline methods, an innate advantage of large-scale multimodal models. In addition, to explore more intriguing linguistic directives and fully harness the universal capabilities of the Large Language Model (LLM), Fig. 8 presents instances where the model demonstrates a clear comprehension and application of combined directives constructed via "first" and "then". It was further noted that even in the absence of explicit directional information within the dataset, the model reliably interpreted linguistic directives which, in turn, influenced task execution accuracy. The ability to utilize language for precise control of quadruped robots facilitated by scene perception, as showcased in Fig. 8, is of immense significance. More intriguing examples and analyses can be found in the supplementary materials. + +# 4.3 Qualitative Performance + +Here, we offer the visualization results of seen and unseen tasks in Fig. 9 with WR-2. We also test QUART in other quadruped robots like Go2 with minor adjustments. More demos can be found in the supplementary materials. + +![](images/93c45c3339580cbf4b271fcfab73b85ea49ae10c2cea78578d15a04cc20d0c05.jpg) +Fig.9: Visualization of seen and unseen tasks. This figure demonstrates the excellent performance of QUART for seen tasks in both simulated and real-world environments. The left part shows 4 instances of seen tasks in both simulation environment and real-world. The right part shows performances on unseen tasks, which contain unseen verbal, object, and background. + +![](images/53696d91f86a064b825e390183bb77623289d6480f3ef03f382fc67875b02c50.jpg) + +# 5 Conclusion & Future Work + +This paper emphasizes the significance of deploying Vision-Language-Action models on quadruped robots. We introduce the concept of Vision-Language-Action tasks for QUAdruped Robots (QUAR-VLA), which seamlessly integrates visual information and instructions from diverse modalities to generate executable actions. Specifically, QUAR-VLA focuses on two main aspects of deploying VLA models for quadruped robots: defining the action space to balance flexibility and efficiency, and addressing the scarcity of large-scale training datasets. To tackle these two questions, we present the QUART models tailored for quadruped robots and the QUARD dataset, which includes diverse tasks such as navigation and manipulation. Our extensive evaluation shows that our approach leads to performant robotic policies and enables QUART to obtain a range of emergent capabilities. This includes generalization to novel objects, the ability to interpret commands not present in the robot training data, and the ability to perform rudimentary reasoning in response to user commands. + +Furthermore, this work has these areas requiring further enhancements: + +1) Automated Data Collection for Larger Datasets. Future work can explore automated processes for data collection, involving techniques such as active learning and data augmentation can be employed to further expand the dataset size and diversity. 2) Better solution for sim2Real gap. Integrating domain adaptation methods like domain randomization, data augmentation, and more accurate modeling of physics engines will be a better solution compared with direct co-training. 3) Improving Inference Speed for Real-time Control. Future works will explore hardware acceleration techniques and model compression techniques to enable faster and more efficient execution of the models. + +# Acknowledgements + +This work was supported by the National Science and Technology Innovation 2030 - Major Project (Grant No. 2022ZD0208800), and NSFC General Program (Grant No. 62176215). + +# References + +1. Ahn, M., Dwibedi, D., Finn, C., Arenas, M.G., Gopalakrishnan, K., Hausman, K., Ichter, B., Irpan, A., Joshi, N., Julian, R., Kirmani, S., Leal, I., Lee, E., Levine, S., Lu, Y., Leal, I., Maddineni, S., Rao, K., Sadigh, D., Sanketi, P., Sermanet, P., Vuong, Q., Welker, S., Xia, F., Xiao, T., Xu, P., Xu, S., Xu, Z.: Autort: Embodied foundation models for large scale orchestration of robotic agents (2024) +2. Bahl, S., Mendonca, R., Chen, L., Jain, U., Pathak, D.: Affordances from human videos as a versatile representation for robotics (2023) +3. Bavishi, R., Olsen, E., Hawthorne, C., Nye, M., Odena, A., Somani, A., Tasirlar, S.: Introducing our multimodal models (2023), https://www.adept.ai/blog/fuyu-8b +4. Belkhale, S., Cui, Y., Sadigh, D.: Hydra: Hybrid robot actions for imitation learning. In: Proceedings of the 7th Conference on Robot Learning (CoRL) (2023) +5. Bharadhwaj, H., Vakil, J., Sharma, M., Gupta, A., Tulsiani, S., Kumar, V.: Roboagent: Generalization and efficiency in robot manipulation via semantic augmentations and action chunking (2023) +6. Bousmalis, K., Irpan, A., Wohlhart, P., Bai, Y., Kelcey, M., Kalakrishnan, M., Downs, L., Ibarz, J., Pastor, P., Konolige, K., et al.: Using simulation and domain adaptation to improve efficiency of deep robotic grasping. In: 2018 IEEE international conference on robotics and automation (ICRA). pp. 4243-4250. IEEE (2018) +7. Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Chen, X., Choromanski, K., Ding, T., Driess, D., Dubey, A., Finn, C., Florence, P., Fu, C., Arenas, M.G., Gopalakrishnan, K., Han, K., Hausman, K., Herzog, A., Hsu, J., Ichter, B., Irpan, A., Joshi, N., Julian, R., Kalashnikov, D., Kuang, Y., Leal, I., Lee, L., Lee, T.W.E., Levine, S., Lu, Y., Michalewski, H., Mordatch, I., Pertsch, K., Rao, K., Reymann, K., Ryoo, M., Salazar, G., Sanketi, P., Sermanet, P., Singh, J., Singh, A., Soricut, R., Tran, H., Vanhoucke, V., Vuong, Q., Wahid, A., Welker, S., Wohlhart, P., Wu, J., Xia, F., Xiao, T., Xu, P., Xu, S., Yu, T., Zitkovich, B.: Rt-2: Vision-language-action models transfer web knowledge to robotic control (2023) +8. Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Chen, X., Choromanski, K., Ding, T., Driess, D., Dubey, A., Finn, C., et al.: Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818 (2023) +9. Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Dabis, J., Finn, C., Gopalakrishnan, K., Hausman, K., Herzog, A., Hsu, J., Ibarz, J., Ichter, B., Irpan, A., Jackson, T., Jesmonth, S., Joshi, N.J., Julian, R., Kalashnikov, D., Kuang, Y., Leal, I., Lee, K.H., Levine, S., Lu, Y., Malla, U., Manjunath, D., Mordatch, I., Nachum, O., Parada, C., Peralta, J., Perez, E., Pertsch, K., Quiambao, J., Rao, K., Ryoo, M., Salazar, G., Sanketi, P., Sayed, K., Singh, J., Sontakke, S., Stone, A., Tan, C., Tran, H., Vanhoucke, V., Vega, S., Vuong, Q., Xia, F., Xiao, T., Xu, P., Xu, S., Yu, T., Zitkovich, B.: Rt-1: Robotics transformer for real-world control at scale (2023) + +10. Caluwaerts, K., Iscen, A., Kew, J.C., Yu, W., Zhang, T., Freeman, D., Lee, K.H., Lee, L., Saliceti, S., Zhuang, V., et al.: Barkour: Benchmarking animal-level agility with quadruped robots. arXiv preprint arXiv:2305.14654 (2023) +11. Eckert, P., Ijspeert, A.J.: Benchmarking agility for multilegged terrestrial robots. IEEE Transactions on Robotics 35(2), 529-535 (2019) +12. Gu, J., Kirmani, S., Wohlhart, P., Lu, Y., Arenas, M.G., Rao, K., Yu, W., Fu, C., Gopalakrishnan, K., Xu, Z., Sundaresan, P., Xu, P., Su, H., Hausman, K., Finn, C., Vuong, Q., Xiao, T.: Rt-trajectory: Robotic task generalization via hindsight trajectory sketches (2023) +13. Haldar, S., Mathur, V., Yarats, D., Pinto, L.: Watch and match: Supercharging imitation with regularized optimal transport. CoRL (2022) +14. Hutter, M., Gehring, C., Jud, D., Lauber, A., Bellicoso, C.D., Tsounis, V., Hwangbo, J., Bodie, K., Fankhauser, P., Bloesch, M., Diethelm, R., Bachmann, S., Melzer, A., Hoepflinger, M.: Anymal - a highly mobile and dynamic quadrupedal robot. In: 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 38-44 (2016). https://doi.org/10.1109/IROS.2016.7758092 +15. Jiang, Y., Moseson, S., Saxena, A.: Efficient grasping from rgbd images: Learning using a new rectangle representation. In: 2011 IEEE International conference on robotics and automation. pp. 3304-3311. IEEE (2011) +16. Kareer, S., Yokoyama, N., Batra, D., Ha, S., Truong, J.: Vinl: Visual navigation and locomotion over obstacles. 2023 IEEE International Conference on Robotics and Automation (ICRA) pp. 2018-2024 (2022), https://api-semanticscholar.org/CorpusID:253117178 +17. Karnan, H., Nair, A., Xiao, X., Warnell, G., Pirk, S., Toshev, A., Hart, J., Biswas, J., Stone, P.: Socially compliant navigation dataset (scand): A large-scale dataset of demonstrations for social navigation. IEEE Robotics and Automation Letters 7(4), 11807-11814 (2022) +18. Lee, K.H., Nachum, O., Yang, M.S., Lee, L., Freeman, D., Guadarrama, S., Fischer, I., Xu, W., Jang, E., Michalewski, H., et al.: Multi-game decision transformers. Advances in Neural Information Processing Systems 35, 27921-27936 (2022) +19. Lee, M.A., Zhu, Y., Srinivasan, K., Shah, P., Savarese, S., Fei-Fei, L., Garg, A., Bohg, J.: Making sense of vision and touch: Self-supervised learning of multimodal representations for contact-rich tasks. In: 2019 IEEE International Conference on Robotics and Automation (ICRA) (2019), https://arxiv.org/abs/1810.10191 +20. Li, X., Liu, M., Zhang, H., Yu, C., Xu, J., Wu, H., Cheang, C., Jing, Y., Zhang, W., Liu, H., et al.: Vision-language foundation models as effective robot imitators. arXiv preprint arXiv:2311.01378 (2023) +21. Majumdar, A., Yadav, K., Arnaud, S., Ma, Y.J., Chen, C., Silwal, S., Jain, A., Berges, V.P., Abbeel, P., Malik, J., Batra, D., Lin, Y., Maksymets, O., Rajeswaran, A., Meier, F.: Where are we in the search for an artificial visual cortex for embodied intelligence? (2023) +22. Makoviychuk, V., Wawrzyniak, L., Guo, Y., Lu, M., Storey, K., Macklin, M., Hoeller, D., Rudin, N., Allshire, A., Handa, A., State, G.: Isaac gym: High performancegpu-based physics simulation for robot learning (2021) +23. Margolis, G.B., Agrawal, P.: Walk these ways: Tuning robot control for generalization with multiplicity of behavior. In: Liu, K., Kulic, D., Ichnowski, J. (eds.) Proceedings of The 6th Conference on Robot Learning. Proceedings of Machine Learning Research, vol. 205, pp. 22-31. PMLR (14-18 Dec 2023) +24. Nair, S., Rajeswaran, A., Kumar, V., Finn, C., Gupta, A.: R3m: A universal visual representation for robot manipulation. In: Conference on Robot Learning (2022), https://api-semanticscholar.org/CorpusID:247618840 + +25. Nair, S., Rajeswaran, A., Kumar, V., Finn, C., Gupta, A.: R3m: A universal visual representation for robot manipulation (2022) +26. Pari, J., Shafiullah, N.M., Arunachalam, S.P., Pinto, L.: The surprising effectiveness of representation learning for visual imitation (2021) +27. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision (2021) +28. Reed, S., Zolna, K., Parisotto, E., Colmenarejo, S.G., Novikov, A., Barth-Maron, G., Gimenez, M., Sulsky, Y., Kay, J., Springenberg, J.T., et al.: A generalist agent. arXiv preprint arXiv:2205.06175 (2022) +29. Reed, S., Zolna, K., Parisotto, E., Colmenarejo, S.G., Novikov, A., Barth-Maron, G., Gimenez, M., Sulsky, Y., Kay, J., Springenberg, J.T., et al.: A generalist agent. arXiv preprint arXiv:2205.06175 (2022) +30. Schiavi, G., Wulkop, P., Rizzi, G., Ott, L., Siegwart, R., Chung, J.J.: Learning agent-aware affordances for closed-loop interaction with articulated objects. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 5916-5922 (2023). https://doi.org/10.1109/ICRA48891.2023.10160747 +31. Shah, D., Sridhar, A., Dashora, N., Stachowicz, K., Black, K., Hirose, N., Levine, S.: Vint: A foundation model for visual navigation (2023) +32. Shilane, P., Min, P., Kazhdan, M., Funkhouser, T.: The princeton shape benchmark. In: Proceedings Shape Modeling Applications, 2004. pp. 167-178 (2004) +33. Shridhar, M., Manuelli, L., Fox, D.: CIoport: What and where pathways for robotic manipulation. In: Conference on Robot Learning. pp. 894-906. PMLR (2022) +34. Stone, A., Xiao, T., Lu, Y., Gopalakrishnan, K., Lee, K.H., Vuong, Q., Wohlhart, P., Kirmani, S., Zitkovich, B., Xia, F., Finn, C., Hausman, K.: Open-world object manipulation using pre-trained vision-language models (2023) +35. Szot, A., Schwarzer, M., Agrawal, H., Mazoure, B., Talbott, W., Metcalf, K., Mackraz, N., Hjelm, D., Toshev, A.: Large language models as generalizable policies for embodied tasks. arXiv preprint arXiv:2310.17722 (2023) +36. Tang, Y., Yu, W., Tan, J., Zen, H., Faust, A., Harada, T.: Saytap: Language to quadrupedal locomotion. arXiv preprint arXiv:2306.07580 (2023) +37. Wei, J., Hou, L., Lampinen, A., Chen, X., Huang, D., Tay, Y., Chen, X., Lu, Y., Zhou, D., Ma, T., et al.: Symbol tuning improves in-context learning in language models. arXiv preprint arXiv:2305.08298 (2023) +38. Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., Xiao, J.: 3d shapenets: A deep representation for volumetric shapes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1912-1920 (2015) +39. Yang, R., Zhang, M., Hansen, N., Xu, H., Wang, X.: Learning vision-guided quadrupedal locomotion end-to-end with cross-modal transformers. ArXiv abs/2107.03996 (2021), https://api_semanticscholar.org/CorpusID: 235765481 +40. Yu, K.T., Bauza, M., Fazeli, N., Rodriguez, A.: More than a million ways to be pushed. a high-fidelity experimental dataset of planar pushing. In: 2016 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 30-37. IEEE (2016) \ No newline at end of file diff --git a/2024/QUAR-VLA_ Vision-Language-Action Model for Quadruped Robots/images.zip b/2024/QUAR-VLA_ Vision-Language-Action Model for Quadruped Robots/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..870958d1aa7da7605f046194f9f01e47b60fe00f --- /dev/null +++ b/2024/QUAR-VLA_ Vision-Language-Action Model for Quadruped Robots/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d09706d9dd0ccf9b5592b94637326d2f36305045d84e8665e118ea30afc859fe +size 395207 diff --git a/2024/QUAR-VLA_ Vision-Language-Action Model for Quadruped Robots/layout.json b/2024/QUAR-VLA_ Vision-Language-Action Model for Quadruped Robots/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..05dea897631338f3bc6c5a07de26c3040ed9f742 --- /dev/null +++ b/2024/QUAR-VLA_ Vision-Language-Action Model for Quadruped Robots/layout.json @@ -0,0 +1,7129 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "type": "text", + "content": "QUAR-VLA: Vision-Language-Action Model for Quadruped Robots" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 154, + 167, + 462, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 167, + 462, + 195 + ], + "spans": [ + { + "bbox": [ + 154, + 167, + 462, + 195 + ], + "type": "text", + "content": "Pengxiang Ding" + }, + { + "bbox": [ + 154, + 167, + 462, + 195 + ], + "type": "inline_equation", + "content": "^{12}" + }, + { + "bbox": [ + 154, + 167, + 462, + 195 + ], + "type": "text", + "content": " Han Zhao" + }, + { + "bbox": [ + 154, + 167, + 462, + 195 + ], + "type": "inline_equation", + "content": "^{12}" + }, + { + "bbox": [ + 154, + 167, + 462, + 195 + ], + "type": "text", + "content": " Wenjie Zhang" + }, + { + "bbox": [ + 154, + 167, + 462, + 195 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 154, + 167, + 462, + 195 + ], + "type": "text", + "content": " Wenxuan Song" + }, + { + "bbox": [ + 154, + 167, + 462, + 195 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 154, + 167, + 462, + 195 + ], + "type": "text", + "content": " Min Zhang" + }, + { + "bbox": [ + 154, + 167, + 462, + 195 + ], + "type": "inline_equation", + "content": "^{12}" + }, + { + "bbox": [ + 154, + 167, + 462, + 195 + ], + "type": "text", + "content": " Siteng Huang" + }, + { + "bbox": [ + 154, + 167, + 462, + 195 + ], + "type": "inline_equation", + "content": "^{12}" + }, + { + "bbox": [ + 154, + 167, + 462, + 195 + ], + "type": "text", + "content": " Ningxi Yang" + }, + { + "bbox": [ + 154, + 167, + 462, + 195 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 154, + 167, + 462, + 195 + ], + "type": "text", + "content": " Donglin Wang" + }, + { + "bbox": [ + 154, + 167, + 462, + 195 + ], + "type": "inline_equation", + "content": "^{1\\star}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 208, + 201, + 414, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 201, + 414, + 236 + ], + "spans": [ + { + "bbox": [ + 208, + 201, + 414, + 236 + ], + "type": "text", + "content": "1Zhejiang University 2Westlake University 3Beijing Normal University 4Monash University {dingpengxiang}@westlake.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 255, + 455, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 255, + 455, + 531 + ], + "spans": [ + { + "bbox": [ + 160, + 255, + 455, + 531 + ], + "type": "text", + "content": "Abstract. The important manifestation of robot intelligence is the ability to naturally interact and autonomously make decisions. Traditional quadruped robot learning typically handles language interaction and visual autonomous perception separately, which, while simplifying system design, also limits the synergy between different information streams. This separation poses challenges in achieving seamless autonomous reasoning, decision-making, and action execution. To address these limitations, a novel paradigm, named Vision-Language-Action tasks for QUAdruped Robots (QUAR-VLA), has been introduced in this paper. This approach tightly integrates visual information and instructions to generate executable actions, effectively merging perception, planning, and decision-making. The central idea is to elevate the overall intelligence of the robot. Within this framework, a notable challenge lies in aligning fine-grained instructions with visual perception information. This emphasizes the complexity involved in ensuring that the robot accurately interprets and acts upon detailed instructions in harmony with its visual observations. Consequently, we propose QUAdruped Robotic Transformer (QUART), a VLA model to integrate visual information and instructions from diverse modalities as input and generates executable actions for real-world robots and present QUAdruped Robot Dataset (QUARD), a large-scale multi-task dataset including perception, navigation and advanced capability like whole-body manipulation tasks for training QUART model. Our extensive evaluation shows that our approach leads to performant robotic policies and enables QUART to obtain a range of generalization capabilities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 540, + 455, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 540, + 455, + 562 + ], + "spans": [ + { + "bbox": [ + 160, + 540, + 455, + 562 + ], + "type": "text", + "content": "Keywords: Robotics " + }, + { + "bbox": [ + 160, + 540, + 455, + 562 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 540, + 455, + 562 + ], + "type": "text", + "content": " Quadruped Robot Learning " + }, + { + "bbox": [ + 160, + 540, + 455, + 562 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 540, + 455, + 562 + ], + "type": "text", + "content": " Vision-Language-Action Model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 578, + 230, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 230, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 230, + 590 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 600, + 482, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 600, + 482, + 650 + ], + "spans": [ + { + "bbox": [ + 130, + 600, + 482, + 650 + ], + "type": "text", + "content": "Quadruped robots, characterized by their excellent traversability on complex terrains and agile movements, have garnered significant attention in the field of robotics [14]. Researchers have extensively employed these robots to explore tasks encompassing autonomous navigation and manipulation [16, 17, 36]." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 654, + 234, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 654, + 234, + 666 + ], + "spans": [ + { + "bbox": [ + 134, + 654, + 234, + 666 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 246, + 185 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 246, + 185 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 246, + 185 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 246, + 185 + ], + "type": "image", + "image_path": "e8766da15dc3f80c3b4234c895c71db6552a514e54fe341da2e6b7bc2e650210.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 193, + 220, + 204 + ], + "lines": [ + { + "bbox": [ + 170, + 193, + 220, + 204 + ], + "spans": [ + { + "bbox": [ + 170, + 193, + 220, + 204 + ], + "type": "text", + "content": "(a) QUAR-VA" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 250, + 114, + 363, + 182 + ], + "blocks": [ + { + "bbox": [ + 250, + 114, + 363, + 182 + ], + "lines": [ + { + "bbox": [ + 250, + 114, + 363, + 182 + ], + "spans": [ + { + "bbox": [ + 250, + 114, + 363, + 182 + ], + "type": "image", + "image_path": "4caa3042167d5cba3571d471f8b0f66ab4dd32381d27f1a51bbd1177cb1ff5c4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 288, + 193, + 339, + 204 + ], + "lines": [ + { + "bbox": [ + 288, + 193, + 339, + 204 + ], + "spans": [ + { + "bbox": [ + 288, + 193, + 339, + 204 + ], + "type": "text", + "content": "(b) QUAR-LA" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 369, + 114, + 482, + 185 + ], + "blocks": [ + { + "bbox": [ + 369, + 114, + 482, + 185 + ], + "lines": [ + { + "bbox": [ + 369, + 114, + 482, + 185 + ], + "spans": [ + { + "bbox": [ + 369, + 114, + 482, + 185 + ], + "type": "image", + "image_path": "8d7565c4199cb5db8bc9ed69df7ff4c547878b0b524cf732423968de2e6cc4cb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 405, + 193, + 461, + 204 + ], + "lines": [ + { + "bbox": [ + 405, + 193, + 461, + 204 + ], + "spans": [ + { + "bbox": [ + 405, + 193, + 461, + 204 + ], + "type": "text", + "content": "(c) QUAR-VLA" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 222, + 482, + 300 + ], + "lines": [ + { + "bbox": [ + 130, + 222, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 222, + 482, + 300 + ], + "type": "text", + "content": "Fig. 1: Comparison of QUAR-VA, QUAR-LA, and QUAR-VLA. QUAR-VA solely utilizes coarse-grained vision information, lacking explicit instructions for handling diverse tasks. In contrast, QUAR-LA exclusively relies on language information and lacks of vision information for autonomy. Therefore, QUART-VLA combines both vision information and language instructions as inputs, enabling autonomous problem-solving across a range of tasks, revealing distinct input modalities and task capabilities." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 327, + 482, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 566 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 566 + ], + "type": "text", + "content": "Broadly speaking, the quadruped tasks consist of two major specifications: Vision-Action tasks for QUAdruped Robots (QUAR-VA) and Language-Action tasks for QUAdruped Robots (QUAR-LA). As depicted in Fig. 1, in QUAR-VA approaches [31], quadruped robots {receive} perception images captured from a first-person perspective and instruction images obtained from a third-person perspective to guide their actions. However, such a task specification often relies on a single (coarse-grained) goal image instruction, making it difficult to apply in many real-world combination tasks, i.e. requiring combining multiple sub-instructions. In contrast, employing language as instructions [36], QUAR-LA formulation allows for executing more fine-grained and diverse tasks, suitable for expressing combinational commands (i.e. \"before / then\"), complex spatial relationships (i.e. \"move to the left\"), and many commonsense command priors (i.e. \"move fast\"). Nevertheless, QUAR-LA approaches, which lack the integration of visual modality, hinder the robots' ability to perceive the environment, thereby impeding their autonomous navigation capabilities [36]. To enable quadruped robots to autonomously navigate and manipulate various tasks, in this paper, we propose a new paradigm: Vision-Language-Action tasks for QUAdruped Robots (QUAR-VLA), integrating visual information and instructions from diverse modalities as input and generating executable actions for real-world robots." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 567, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 482, + 639 + ], + "type": "text", + "content": "This task primarily encompasses two challenges. Firstly, there is a lack of large-scale datasets in the research community for quadruped robots performing diverse tasks. Pretraining models require abundant trajectory data from a variety of tasks. However, collecting substantial amounts of real-world data often necessitates human involvement and significant time investment in manual instruction operations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Secondly, different from navigation tasks [31] for general mobile robots and manipulation tasks for fixed-base systems [7, 9, 12, 34], building a VLA model" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "P. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 228 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 228 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 228 + ], + "type": "image", + "image_path": "36671b8685ef7c299dd0f8a023a98aea7645b7abed95670b6ee402114d5c9a11.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 245, + 482, + 334 + ], + "lines": [ + { + "bbox": [ + 130, + 245, + 482, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 482, + 334 + ], + "type": "text", + "content": "Fig. 2: Overview of QUAR-VLA. Our tasks encompass a diverse range of perception, navigation, and other advanced capability. The Vision-Language-Action (VLA) model first undergoes training with a huge amount of simulation data (246K episodes) and a small amount of real-world data (3K episodes). In the inference phase, images and texts undergo tokenization, after which QUART generates 12-dimensional action tokens. These tokens are subsequently detokenized into valid robot actions and deployed on a physical quadruped robot. This methodology effectively extends the learned capabilities from a simulated environment to real-world applications." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 363, + 482, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 446 + ], + "type": "text", + "content": "to solve complex quadruped robot tasks is considerably more challenging due to their agile locomotion behaviors. The action space needs to be properly defined to strike a balance between movement flexibility and computational efficiency. The action generated by the model should not be too simplistic, akin to the planner base velocities output by 2-D navigation modules, nor should they require a high execution frequency like the low-level motion policy that directly controls the joint motors." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": "To address these two problems, we collect a large-scale multi-task dataset QUAdruped Robot Dataset (QUARD). It includes multiple tasks such as perception, navigation, and advanced capabilities like object avoidance. To the best of our knowledge, this is the first quadruped robot dataset that incorporates a significant amount of vision, language instruction, and robot command data. As collecting data on real robots is expensive and inefficient, we primarily rely on data generated in simulation, which exhibits significant differences in visual, sensor, and system dynamics. We also introduce QUAdruped Robotic Transformer (QUART), a VLA model for training QUARD. QUART takes images of the robot's first-view camera and the natural language instruction as inputs and generates rich control commands that include controlling the robot's base velocity, posture, and gait parameters. QUART leverages a pre-trained large-scale visual language model and fine-tunes it on our dataset to enable the generation of executable commands for quadruped robots. To address the simple-to-real gap caused by the data disparity, we construct a co-training pipeline to effectively distill the knowledge of simulation data for real-scene deployment. Our extensive evaluation shows that our approach leads to performant robotic policies and enables QUART to obtain a range of generalization capabilities." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "QUAR-VLA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "Our contributions are as follows: 1) To the best of our knowledge, we first propose a new paradigm: QUAR-VLA, integrating visual and language instructions to executable actions for more autonomous and diverse quadruped robot tasks. 2) We present a large-scale multi-task dataset, QUARD, and a Vision-Language-Action model, QUART to solve the QUAR-VLA tasks. 3) Our extensive evaluation shows that our approach leads to performant robotic policies and enables QUART to obtain a range of generalization capabilities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 220, + 238, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 220, + 238, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 220, + 238, + 232 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 246, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 482, + 437 + ], + "type": "text", + "content": "Quadruped Robot Learning. Previous research [2,4,10,13,16,17,19,26,30,36, 39] has delved into single-model policies. As to vision-based methods (QUAR-VA), Kareer et al. [16] utilizes vision perception to predict a privileged terrain map, enabling the robot to perform specific locomotion based on the terrain characteristics. Karnan [17] leverages ego-motion estimates obtained through vision perception, allowing the robot to understand its relative position and direction to avoid collisions and navigate toward a goal location. While some works have succeeded in these tasks through vision perception, relying solely on vision limits the capability to handle only a specific type of task. In response to this limitation, Tang et al. [36] have started exploring language-conditioned tasks (QUAR-LA). However, the absence of vision perception in their approach results in a lack of autonomy for the robot to interact with its environment. Therefore, this paper proposes a new QUAR-VLA paradigm to integrate visual information and instructions to generate executable actions, effectively ensuring that the robot accurately interprets and acts upon detailed instructions in harmony with its visual observations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "text", + "content": "Vision-language-action models. Visual-language-action models [1, 5, 20, 24, 28, 33, 35] integrates visual information and instructions to generate executable actions, effectively merging perception, planning, and decision-making and elevating the overall intelligence of the robot. As a result, vision-language-action models have been a popular area of research in robotics. In the realm of visual-language-action models, Shridhar et al. [33] have made significant strides by leveraging CLIP to encode text inputs, thereby enhancing the vision model's semantic comprehension and action execution capabilities. Further advancing the field, Brohan et al. [8] and Li et al. [20] have innovatively applied large language models to craft manipulation policies specifically for robotic applications. This paper delves into the domain of embodied agents, focusing on the quadruped robot as a case study. It explores the challenges and opportunities in enabling quadruped robots to autonomously navigate and perform a variety of tasks as directed by human instructions, contributing to the broader discourse on robotic autonomy and intelligence." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Robot Dataset. The realm of robotic learning has been at the forefront of advancing open-source datasets tailored explicitly for robot learning [6,15,17,32, 36,38,40]. Within the domain of quadrupedal robots, previous works on datasets primarily focused on algorithms and leg controllers. Eckert et al. [11] introduced" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "P. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "a grading system and an online dataset for the collection and distribution of agility scores. Caluwaerts et al. [10] sought to address the lack of a standardized and scalable environment and inductive metrics. Tang et al. [36] pioneered the generation of desired joint position control using LLM. Nevertheless, common limitations persist among these quadruped robot datasets: the quantity is insufficient and the tasks are limited in diversity. Hence, our data repository aims to complement these endeavors. We curate and process an extensive array of skills on various objects, across diverse scenes in both the real world and virtual environments. The dataset includes image data, robot action data, and point cloud data, providing quadruped robots with rich perceptual information to tackle increasingly intricate and demanding tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 265, + 202, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 265, + 202, + 277 + ], + "spans": [ + { + "bbox": [ + 132, + 265, + 202, + 277 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 289, + 482, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 337 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 337 + ], + "type": "text", + "content": "This section provides a detailed exposition of our proposed methodology. Initially, we present the definition of our proposed QUAR-VLA in Section 3.1. Following that, Section 3.2 outlines the collecting process of QUARD. Lastly, we delve into the overarching architecture of QUART in Section 3.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 354, + 236, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 236, + 366 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 236, + 366 + ], + "type": "text", + "content": "3.1 Problem Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 372, + 481, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 372, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 372, + 481, + 407 + ], + "type": "text", + "content": "The objective of QUAR-VLA is to construct a vision-language-action model learned from large-scale demonstration data and generate actions for closed-loop robot control." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 408, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 482, + 479 + ], + "type": "text", + "content": "An overview of our QUAR-VLA is shown in Fig. 2. Our goal is to train a conditional policy QUART that can interpret RGB image(s), denoted " + }, + { + "bbox": [ + 130, + 408, + 482, + 479 + ], + "type": "inline_equation", + "content": "s \\in S" + }, + { + "bbox": [ + 130, + 408, + 482, + 479 + ], + "type": "text", + "content": ", together with a task instruction " + }, + { + "bbox": [ + 130, + 408, + 482, + 479 + ], + "type": "inline_equation", + "content": "w \\in \\mathcal{W}" + }, + { + "bbox": [ + 130, + 408, + 482, + 479 + ], + "type": "text", + "content": ", which correspond to a language string. The policy is a mapping from images and instructions to actions, and can be written as " + }, + { + "bbox": [ + 130, + 408, + 482, + 479 + ], + "type": "inline_equation", + "content": "\\mu : S \\times \\mathcal{W} \\to \\mathcal{A}" + }, + { + "bbox": [ + 130, + 408, + 482, + 479 + ], + "type": "text", + "content": ", where the action space " + }, + { + "bbox": [ + 130, + 408, + 482, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 408, + 482, + 479 + ], + "type": "text", + "content": " consists of the 11-dimensional high-level commands as well as a terminate signal." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 228, + 487, + 480, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 487, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 228, + 487, + 480, + 502 + ], + "type": "interline_equation", + "content": "\\left[ v _ {x}, v _ {y}, \\omega_ {z}, \\theta_ {1}, \\theta_ {2}, \\theta_ {3}, f, h _ {z}, \\phi , s _ {y}, h _ {z} ^ {f}, t \\right] \\tag {1}", + "image_path": "b1e2b8d1a1b58fc90d12f1daebdc96285ec27364ebf530dab72b5e3cd958a3f6.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 509, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 567 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 509, + 482, + 567 + ], + "type": "inline_equation", + "content": "v_{x}, v_{y}" + }, + { + "bbox": [ + 130, + 509, + 482, + 567 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 509, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\omega_{z}" + }, + { + "bbox": [ + 130, + 509, + 482, + 567 + ], + "type": "text", + "content": " represent the velocities along the x-axis, y-axis, and z-axis respectively. " + }, + { + "bbox": [ + 130, + 509, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\theta_{1}, \\theta_{2}" + }, + { + "bbox": [ + 130, + 509, + 482, + 567 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 509, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\theta_{3}" + }, + { + "bbox": [ + 130, + 509, + 482, + 567 + ], + "type": "text", + "content": " indicate the gait pattern, " + }, + { + "bbox": [ + 130, + 509, + 482, + 567 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 509, + 482, + 567 + ], + "type": "text", + "content": " denotes the frequency, " + }, + { + "bbox": [ + 130, + 509, + 482, + 567 + ], + "type": "inline_equation", + "content": "h_{z}" + }, + { + "bbox": [ + 130, + 509, + 482, + 567 + ], + "type": "text", + "content": " represents the height of the robot, " + }, + { + "bbox": [ + 130, + 509, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 509, + 482, + 567 + ], + "type": "text", + "content": " denotes the pitch angle, " + }, + { + "bbox": [ + 130, + 509, + 482, + 567 + ], + "type": "inline_equation", + "content": "s_{y}" + }, + { + "bbox": [ + 130, + 509, + 482, + 567 + ], + "type": "text", + "content": " corresponds to the foot width, " + }, + { + "bbox": [ + 130, + 509, + 482, + 567 + ], + "type": "inline_equation", + "content": "h_{z}^{f}" + }, + { + "bbox": [ + 130, + 509, + 482, + 567 + ], + "type": "text", + "content": " represents the foot height, and " + }, + { + "bbox": [ + 130, + 509, + 482, + 567 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 509, + 482, + 567 + ], + "type": "text", + "content": " indicates the termination signal of the action." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 568, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 616 + ], + "type": "text", + "content": "Notably, in this study, we employ the discretization method proposed by [9] to discretize all continuous dimensions into 256 uniformly sized bins. This choice of discretization is motivated by the desire to reduce the complexity of action search and improve the stability and convergence of algorithms." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 622, + 359, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 359, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 359, + 635 + ], + "type": "text", + "content": "3.2 Large-scale Quadruped Robot Datasets" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "To enable an imitation learning system to generalize to new tasks with zero demonstrations of said task, we must be able to easily collect a diverse dataset," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 394, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 394, + 90, + 447, + 102 + ], + "type": "text", + "content": "QUAR-VLA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 173, + 481, + 285 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 169 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 169 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 169 + ], + "type": "text", + "content": "Table 1: Tasks Definition. The \"Type\" means different capabilities of robots. The \"Level\" devises the difficulty into 3 levels. The \"Skill\" means different skill/task categories. The \"Episode\" signifies the number of experiments conducted for each task, which also corresponds to the number of trajectories. The \"Description\" is the illustration of the tasks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 173, + 481, + 285 + ], + "lines": [ + { + "bbox": [ + 133, + 173, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 133, + 173, + 481, + 285 + ], + "type": "table", + "html": "
TypeLevelSkillEpisodeDescription
PerceptionEasyDistinguish (sim)Letter10KIdentify the correct one from multiple boxes with different printed letters
BasicMediumGo to Object (sim)72KNavigate to the object and stop in front of it
NavigationMediumGo to Object (real)3KNavigate to the object and stop in front of it
Advanced CapabilityHardGo through Tunnel48KSpatial Navigation: Go through the correct tunnel from two tunnels with different colors and shapes
HardGo to Object and avoid the obstacle (sim)63KObstacle Avoidance: Navigate to the object without colliding with the obstacle
HardCrawl under Bar (sim)1KEnvironment Adaptation: Crawl under a bar with a low height
HardUnload Object into Re-ceptacle (sim)52KObject Manipulation: Move with a ball on the back and unload it into a receptacle
Total249KThe total number of episodes
", + "image_path": "6cbd25dad11e2c5c15b1a13f11d7d26f570b79e71030e6030d48f390e78626f8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 312, + 480, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 480, + 360 + ], + "type": "text", + "content": "provide corrective feedback, and evaluate many tasks at scale. Therefore, we collect a large-scale multi-task dataset, QUAdruped Robot Dataset (QUARD), which includes multiple tasks such as perception, basic navigation, and advanced capability like object manipulation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 361, + 481, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 481, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 481, + 529 + ], + "type": "text", + "content": "Task Definition. As is shown in Table 1, there are seven tasks distributed across three levels of difficulty: easy, medium, and hard. The easy and medium tasks are designed such that the robot can accomplish them using fundamental skills. For instance, the \"Go to\" task is a basic action integral to all navigation tasks, while the \"Distinguish\" task serves as the foundational action for all perception tasks. Hard tasks necessitate the robot to execute tasks using more advanced skills. A case in point is the \"Unload object\" task, which involves the robot distinguishing the target container, navigating to the target position, adjusting body position, and considering potential collisions to successfully unload the object. This process integrates perception, navigation, and whole-body manipulation. The complexity of tasks is reflected in the average trajectory length, with more challenging tasks requiring a greater number of steps and consequently taking longer to complete. The specific trajectory lengths and task distribution for each task are detailed in Fig. 3 and Fig. 4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 531, + 481, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 481, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 481, + 639 + ], + "type": "text", + "content": "System Setup. The robot used to collect the trajectory data is WR-2, which is a quadruped robot with 12 joints. WR-2 has around " + }, + { + "bbox": [ + 130, + 531, + 481, + 639 + ], + "type": "inline_equation", + "content": "25\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 531, + 481, + 639 + ], + "type": "text", + "content": " in standing height and " + }, + { + "bbox": [ + 130, + 531, + 481, + 639 + ], + "type": "inline_equation", + "content": "40\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 531, + 481, + 639 + ], + "type": "text", + "content": " body length. The command output is sent to the low-level command tracking controller (pre-trained command-conditioned policy in [23]) to generate the actual joint action of the robot. The simulation data was collected in Nvidia's Isaac Gym [22], a powerful simulator that allows us to collect massive robot trajectories in parallel. The real data was collected by manually manipulating a quadruped robot in the lab setting. The perception data is provided by a RealSense d435 camera installed in the front of the robot." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Data Collection. As to simulation data, " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^*" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^*" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " algorithms are used to plot optimal paths for a dog navigating through various objects and obstacles. The" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "P. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 128, + 329, + 220 + ], + "blocks": [ + { + "bbox": [ + 167, + 128, + 329, + 220 + ], + "lines": [ + { + "bbox": [ + 167, + 128, + 329, + 220 + ], + "spans": [ + { + "bbox": [ + 167, + 128, + 329, + 220 + ], + "type": "image", + "image_path": "945e96e3aad707c6432044216c62843e8e774f84b5812c9f4efecbb20bc7624f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 256, + 482, + 334 + ], + "lines": [ + { + "bbox": [ + 130, + 256, + 482, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 256, + 482, + 334 + ], + "type": "text", + "content": "Fig. 3: The left figure illustrates the trajectory lengths corresponding to different tasks and the right figure illustrates the relationships between tasks. As the difficulty of the skill increases, it can be observed that the average trajectory length gradually increases. The right figure demonstrates the relationship between the types of tasks: perception as foundational capability; basic navigation ability is built on perception; object manipulation, obstacle avoidance, spatial navigation, and environment adaption extend from perception and navigation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 334, + 128, + 432, + 223 + ], + "blocks": [ + { + "bbox": [ + 334, + 128, + 432, + 223 + ], + "lines": [ + { + "bbox": [ + 334, + 128, + 432, + 223 + ], + "spans": [ + { + "bbox": [ + 334, + 128, + 432, + 223 + ], + "type": "image", + "image_path": "1af447d773b093233f4d6693c985b85bb02d0a99a77e2dba13a4e4129c92115c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 361, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\mathrm{A}^*" + }, + { + "bbox": [ + 130, + 361, + 482, + 434 + ], + "type": "text", + "content": " algorithm seeks the most cost-effective path, while the " + }, + { + "bbox": [ + 130, + 361, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^*" + }, + { + "bbox": [ + 130, + 361, + 482, + 434 + ], + "type": "text", + "content": " algorithm adapts to changes in real time. A Proportional-Derivative controller then converts these paths into target velocities for the dog, ensuring smooth movement. The combination of both algorithms provides a flexible and efficient method for path planning in various environments. For real data, The real data set is obtained in a laboratory environment using remote control." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 436, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 581 + ], + "type": "text", + "content": "Consistency constraints. To maintain consistency between simulation and reality, we've established constraints for data collection. The robot starts at the origin, with the target randomly positioned within [2.7,3.3] meters on the x-axis and [0.9,1.1] meters on the y-axis. In obstacle scenarios, obstacles are placed 1.5 meters from the target's x-coordinate, with the same y-coordinate. Task termination criteria vary. For \"go to object\", \"go to the object and avoid obstacle\", and \"crawl under bar\" tasks, success is when the robot is less than one meter from the target. Other task success criteria are: 1) \"Unload object\": successful when the object is in the target container. 2) \"Go through tunnel\" and \"Crawl under bar\": successful when the robot reaches a specific position behind the tunnel or bar. 3) \"Distinguish letter\": successful when the robot correctly orients towards the visual target." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "More Statistics. The data collection setup adheres to a predefined language template as depicted in Fig. 4, where the task, target object, speed, and gait are explicitly defined. To enhance diversity and generalization capabilities, the dataset includes a variety of common indoor furniture and outdoor facilities, in addition to basic shapes. The basic shapes come in four color variations: green, red, blue, and yellow. Examples of indoor furniture include bookshelves, ovens, and vases, while outdoor facilities encompass trashcans and benches. Recognizing" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "QUAR-VLA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 122, + 478, + 198 + ], + "blocks": [ + { + "bbox": [ + 136, + 122, + 478, + 198 + ], + "lines": [ + { + "bbox": [ + 136, + 122, + 478, + 198 + ], + "spans": [ + { + "bbox": [ + 136, + 122, + 478, + 198 + ], + "type": "image", + "image_path": "508d4d28b3e1f2022b7f49ec06e00e477cd54a2c5d9f82c38e8bc1b0933bafbb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 223, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 130, + 223, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 223, + 482, + 269 + ], + "type": "text", + "content": "Fig. 4: Statistic analysis of QUARD. 1) Simulation data accounts for a larger proportion compared to real data. 2) Trotting occupied most of the share of gait. 3) Most tasks own similar episodes while Distinguish Letter and Go to Object (in reality) owns much less. 4) 3 types of speed occupied " + }, + { + "bbox": [ + 130, + 223, + 482, + 269 + ], + "type": "inline_equation", + "content": "1/3" + }, + { + "bbox": [ + 130, + 223, + 482, + 269 + ], + "type": "text", + "content": " separately." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 293, + 482, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 354 + ], + "type": "text", + "content": "the distinct nature of motion robots as compared to manipulation robots, we also require robots to perform the same task with varying gaits and speeds to adapt to different instructions and environments. The corresponding ratios of gaits and speeds are presented in Fig. 4. More details about the datasets can be found in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 375, + 317, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 317, + 387 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 317, + 387 + ], + "type": "text", + "content": "3.3 Vision-Language-Action Model" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 399, + 482, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 471 + ], + "type": "text", + "content": "In comparison to VLA models [7,9,12,31,34] in other field, our research stands out due to the abundance of kinematics information available in this domain. This allows the robot to not only excel at goal-oriented navigation tasks but also perform a diverse range of whole-body manipulation operations with enhanced flexibility in gaits and body control. Next, we will present our QUAdruped Robotic Transformers (QUART)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "text", + "content": "QUART relies on a pre-trained vision-language-model [3] to associate tokens from the model's existing tokenization with the discrete action bins. It is worth noting that training MLLMs to override existing tokens with action tokens is a form of symbol tuning [37], which has been shown to work well in prior work. For [3], integers up to 1000 each have a unique token, so we simply associate the action bins with the token representing the corresponding integer. Notably, QUART model takes a single image " + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "text", + "content": " and a natural language instruction " + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "text", + "content": " as input, which are first converted into corresponding tokens " + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "text", + "content": " through a tokenizer " + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "inline_equation", + "content": "\\tau(t|s,w)" + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "text", + "content": " and fed into a decoder-only transformer module to obtain discretized action tokens " + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "inline_equation", + "content": "p(a_d|t)" + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 146, + 593, + 358, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 593, + 358, + 605 + ], + "spans": [ + { + "bbox": [ + 146, + 593, + 358, + 605 + ], + "type": "text", + "content": "The policy QUART could be shown as follow:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 229, + 620, + 481, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 620, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 229, + 620, + 481, + 633 + ], + "type": "interline_equation", + "content": "\\operatorname {Q U A R T} \\left(a _ {d} | s, w\\right) = p \\left(a _ {d} | t\\right) \\tau (t | s, w) \\tag {2}", + "image_path": "b30c5f5f808fa8f68de24524194d75091eb935e7f0140f649c79d89d12113c60.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "w, s" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": " are the input images and language instruction and " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": " represents the tokenizer and " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": " indicates the vision-language model to output action " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "a_{d}" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "P. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 114, + 471, + 293 + ], + "blocks": [ + { + "bbox": [ + 141, + 114, + 471, + 293 + ], + "lines": [ + { + "bbox": [ + 141, + 114, + 471, + 293 + ], + "spans": [ + { + "bbox": [ + 141, + 114, + 471, + 293 + ], + "type": "image", + "image_path": "1853b291f835ff48236e735d4a90624feb41154f530194ce97b5a5607213a1cf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 312, + 482, + 378 + ], + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 378 + ], + "type": "text", + "content": "Fig. 5: Architecture of QUART. It is designed to leverage the scene comprehension capability of a pretrained MLLM. It receives visual information as observation, and outputs an action representing the actual action taken by the robot based on text-form instructions, and de-tokenizes it into specific action values. QUART can generate a complete action sequence at a processing rate of " + }, + { + "bbox": [ + 130, + 312, + 482, + 378 + ], + "type": "inline_equation", + "content": "2\\mathrm{Hz}" + }, + { + "bbox": [ + 130, + 312, + 482, + 378 + ], + "type": "text", + "content": " in actual scenarios, and hand it over to the underlying low-level strategy for execution." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 402, + 482, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 439 + ], + "type": "text", + "content": "Action Detokenize. To directly convert models' output to valid robot actions for downstream control, we need detokenize the discrete action token " + }, + { + "bbox": [ + 130, + 402, + 482, + 439 + ], + "type": "inline_equation", + "content": "a_{d}" + }, + { + "bbox": [ + 130, + 402, + 482, + 439 + ], + "type": "text", + "content": " into continuous representation " + }, + { + "bbox": [ + 130, + 402, + 482, + 439 + ], + "type": "inline_equation", + "content": "a_{c}" + }, + { + "bbox": [ + 130, + 402, + 482, + 439 + ], + "type": "text", + "content": " (except for the discrete termination command)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 255, + 449, + 481, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 449, + 481, + 462 + ], + "spans": [ + { + "bbox": [ + 255, + 449, + 481, + 462 + ], + "type": "interline_equation", + "content": "a _ {c} = \\operatorname {D e t o k e n i z e} \\left(a _ {d}\\right) \\tag {3}", + "image_path": "79a0b77eb33ef36372677f21c70d5156824acb35668831fa264369a13ae1e007.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 467, + 481, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 467, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 467, + 481, + 491 + ], + "type": "text", + "content": "Loss. We use a standard categorical cross-entropy objective and causal masking that was utilized in prior Transformer-based controllers [18, 29]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 491, + 482, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 598 + ], + "type": "text", + "content": "Inference Speed. In contrast to many applications of large models, such as natural language or image generation, one of the unique requirements for a model that needs to run on real robots in real-time is fast and consistent inference speed. It is important to highlight that our research specifically emphasizes the command control of quadruped robots. In contrast to low-level motor control, the command control outputs alleviate the strict control frequency requirements, enabling smooth integration of larger models and unlocking improved reasoning capabilities. For QUART, the inference time could get 2Hz. More details about the deployment could be found in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 616, + 230, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 616, + 230, + 630 + ], + "spans": [ + { + "bbox": [ + 132, + 616, + 230, + 630 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "We concentrate our experiments on the multi-task ability and generalization. We try to address the following questions: 1. How effective is the vision-language-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 394, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 394, + 90, + 447, + 102 + ], + "type": "text", + "content": "QUAR-VLA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "action architecture for multi-task quadruped task compared to previous VLM baselines? 2. How well do these models generalize to unseen semantic attributes like object shape, color, and unseen verbal information?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 174, + 280, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 174, + 280, + 186 + ], + "spans": [ + { + "bbox": [ + 132, + 174, + 280, + 186 + ], + "type": "text", + "content": "4.1 Implementation Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 197, + 482, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 197, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 197, + 482, + 257 + ], + "type": "text", + "content": "Training Details. We train a specific instantiation of QUART, derived from 8B pre-trained VLM models [3] for superior performance. And we use learning rate 2e-5 and batch size 256 to fine-tune the model for 100K gradient steps. Both models are trained with the next token prediction objective, which corresponds to the behavior cloning loss in robot learning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 258, + 482, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 317 + ], + "type": "text", + "content": "Evaluation Details. We follow the standard robot evaluation metrics [7, 9], success rate (SR), to evaluate the overall performance. The signal we used for robot control is the 11-dimensional command information with the action space outlined in Section 3.1 and 1 stop terminal command means the end of action. More details for each experiment can be seen in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 318, + 482, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 482, + 354 + ], + "type": "text", + "content": "Baselines. Considering the absence of VLA models work on quadruped robots at present, we have taken the following baselines into account for a fair comparison: CLIP [27], R3M [25], and VC-1 [21]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 355, + 480, + 440 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 131, + 355, + 480, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 355, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 131, + 355, + 480, + 378 + ], + "type": "text", + "content": "1. CLIP [27] uniquely encodes both textual and visual data, fusing these embeddings to create feature sets that integrate textual and visual information." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 380, + 480, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 380, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 131, + 380, + 480, + 403 + ], + "type": "text", + "content": "2. R3M [25]is a visual representation model derived from the Ego4D dataset, which acts as a unified, static perception module for policy learning applications." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 404, + 480, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 404, + 480, + 440 + ], + "spans": [ + { + "bbox": [ + 131, + 404, + 480, + 440 + ], + "type": "text", + "content": "3. VC-1 [21] pushes the boundaries of visual representation learning by amalgamating data from multiple datasets and assessing performance across a diverse range of tasks via the implementation of CortexBench." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 441, + 482, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 482, + 502 + ], + "type": "text", + "content": "Since the aforementioned models do not include language conditioning, we include this aspect by separately embedding the language command, allowing us to compare it to our method. Specifically, we concatenate the resulting language embedding tokens with the image tokens generated by the vision model and pass the combined token sequences through a policy head to produce action outputs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 522, + 265, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 265, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 265, + 533 + ], + "type": "text", + "content": "4.2 Overall Performance" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": "Multi-task Performance. To assess the overall capabilities of VLA models on multiple tasks, we conduct evaluations using instructions randomly selected from the training set. It's important to note that this evaluation introduces variations in the placement of objects and other setup factors (such as robot position). This variation demands the system's ability to generalize effectively to realistic environmental variability. In total, over 1500 episodes are tested in this evaluation, comprising 425 episodes for going to objects, 500 for going to objects without colliding with the obstacle. 150 for going through the tunnel, 100 for unloading objects, 100 for distinguishing objects, and 75 for crawling under the bar." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "P. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 116, + 479, + 216 + ], + "blocks": [ + { + "bbox": [ + 134, + 116, + 479, + 216 + ], + "lines": [ + { + "bbox": [ + 134, + 116, + 479, + 216 + ], + "spans": [ + { + "bbox": [ + 134, + 116, + 479, + 216 + ], + "type": "image", + "image_path": "96f36b3e20974a8a02fa104f0858003730dbf647a2d6c5989abb1496d473009b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 234, + 480, + 322 + ], + "lines": [ + { + "bbox": [ + 130, + 234, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 130, + 234, + 480, + 322 + ], + "type": "text", + "content": "Fig. 6: Tasks Successful Rate. R3M lacks alignment with text, which means that although it has some recognition capabilities, it struggles to understand actions through instructions. CLIP and the VC-1 possess the ability to align text and images, enabling them to generate and execute actions based on instructions and observations. As shown above, while R3M may have a slight advantage in easy seen tasks (distinguish), it performs poorly in more complex tasks requiring action execution, significantly lagging behind CLIP and VC-1. QUART fully leverages the advantages of large models, enabling semantic understanding of images and alignment with textual information." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 140, + 379, + 474, + 472 + ], + "blocks": [ + { + "bbox": [ + 130, + 336, + 482, + 369 + ], + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 369 + ], + "type": "text", + "content": "Table 2: Overall performance. QUART has achieved success rates far exceeding those of the baselines in tasks of all difficulty levels, especially in the most challenging crawl and unload tasks, where the baselines have no record of success." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 379, + 474, + 472 + ], + "lines": [ + { + "bbox": [ + 140, + 379, + 474, + 472 + ], + "spans": [ + { + "bbox": [ + 140, + 379, + 474, + 472 + ], + "type": "table", + "html": "
SeenUnseen
Easy DistinguishMedium Go toHard Go avoid Go through Crawl UnloadObjectVerbal
CLIP [27]0.440.430.450.19000.110.14
R3M [25]0.580000000
VC-1 [21]0.460.430.450.31000.290.19
QUART0.660.600.530.410.320.120.350.33
", + "image_path": "7d94ecc7956c413873e159574bb71e6c8a8fc2fdd44f0bdf9dba9025f90e269c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 497, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "text", + "content": "1. Comparison within VLM baselines. The experiment results reveal that R3M has poor performances on tasks except distinguish. The primary reason appears to be the lack of alignment with language semantics, which hinders its ability to comprehend tasks beyond basic discrimination. In contrast, CLIP and VC-1 demonstrate satisfactory performance on fundamental perceptual tasks, such as navigation (e.g., \"go to\" commands). However, their efficacy diminishes significantly when tasks require complex mechanical movements. This observation suggests that while visual language models (VLMs) can grasp abstract principles of the world, directly applying VLMs does not readily translate to the execution of intricate physical tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "2. VLM baselines vs QUART. As is shown in Table 2, Our model has achieved optimal performance on nearly all the baseline models. In our study, we have achieved a decoder-only VLA framework. This approach diverges from the traditional single-layer MLP policy head by leveraging the sequential nature" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "QUAR-VLA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 113, + 443, + 255 + ], + "blocks": [ + { + "bbox": [ + 167, + 113, + 443, + 255 + ], + "lines": [ + { + "bbox": [ + 167, + 113, + 443, + 255 + ], + "spans": [ + { + "bbox": [ + 167, + 113, + 443, + 255 + ], + "type": "image", + "image_path": "4c3f62ca0c7afa00027d8c4e2ec78126493344173709a1dbc89f6fc349d0d549.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 275, + 482, + 374 + ], + "lines": [ + { + "bbox": [ + 130, + 275, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 482, + 374 + ], + "type": "text", + "content": "Fig. 7: Cases with unseen verbal instruction. Here are 3 verbal instructions. The words below represent the names of instructions in the training data, with the words above in bold indicating verbal instruction. When confronted with unseen instructions, the alignment between the existing language and the integration of vision and action cues within the baselines is compromised, resulting in task failure. This failure manifests in behaviors such as repetitive motion, misdirection, wrong terminate commands. Conversely, QUART, leveraging the language prowess inherited from large language models, adeptly achieves generalization under novel instructions, thereby effectuating the harmonization of vision, language, and action." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 400, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 472 + ], + "type": "text", + "content": "of action generation. This allows for the implicit learning of interdependencies between different action dimensions through the use of a transformer. Consequently, while the performance gains may be marginal in simple tasks, there is a noticeable enhancement in tasks that involve complex mechanical movements. significantly improved its perceptual capabilities by incorporating commonsense from the multi-modal large model (MLLM)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 473, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 605 + ], + "type": "text", + "content": "Generalization capabilities. To examine the adaptability in unanticipated scenarios, we orchestrated two primary examinations: one concentrated on unfamiliar objects, and the other on unprecedented linguistic directives. For the unfamiliar objects, we cast our gaze upon an array of circumstances: objects belonging to an identical category but exhibiting divergent textures and colors; objects from the same category but of disparate shapes; and objects with differences in shapes and textures, absent from current datasets. Turning towards the unprecedented linguistic instructions, we characterized unfamiliar instances as linguistic directives that bear the same semantics but vary in expression. For example, within the \"go to the object\" task, we gauged the directive's adaptability by employing \"navigate to target\" as the testing instruction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "1. Unseen Object. As depicted in Fig. 7, we can observe the enhanced generalization ability of our QUART model. For unseen objects, both QUART and VC-1 perform well, thanks to the advantages of the pre-trained VLM model. However, when it comes to unseen verbal descriptions, apart from QUART, all other models essentially failed. This is attributed to the inherent advantage pro" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "P. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 130, + 470, + 242 + ], + "blocks": [ + { + "bbox": [ + 132, + 130, + 470, + 242 + ], + "lines": [ + { + "bbox": [ + 132, + 130, + 470, + 242 + ], + "spans": [ + { + "bbox": [ + 132, + 130, + 470, + 242 + ], + "type": "image", + "image_path": "5cc5a024383e109948eef3cd9e0291c8e7d24af6f2277aecef3bb37b8c84147d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 261, + 482, + 297 + ], + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 297 + ], + "type": "text", + "content": "Fig. 8: QUART can extend basic skills to accommodate complex customized skills with commonsense. It shows example performances of combinational command and spatial understanding towards customized command." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 323, + 482, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 371 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 371 + ], + "type": "text", + "content": "vided by the pre-trained Multimodal Large Language Model (MLLM), enabling the recognition of distinct variations in object types and the comprehension of subtle semantic nuances. Consequently, the model exhibits exceptional generalization performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 373, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 567 + ], + "type": "text", + "content": "2. Unseen Verbal instruction. The training directives utilized were relatively monotonous, demonstrating that minor alterations in instruction can often perplex the model, thereby revealing an imperative for generalization across language directives. Initially, the difficulties introduced by merely replacing descriptions of skill for task generalization were investigated. Evident from Table 2 and Fig. 7 is that QUART surpasses other baseline methods, an innate advantage of large-scale multimodal models. In addition, to explore more intriguing linguistic directives and fully harness the universal capabilities of the Large Language Model (LLM), Fig. 8 presents instances where the model demonstrates a clear comprehension and application of combined directives constructed via \"first\" and \"then\". It was further noted that even in the absence of explicit directional information within the dataset, the model reliably interpreted linguistic directives which, in turn, influenced task execution accuracy. The ability to utilize language for precise control of quadruped robots facilitated by scene perception, as showcased in Fig. 8, is of immense significance. More intriguing examples and analyses can be found in the supplementary materials." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 596, + 284, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 596, + 284, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 596, + 284, + 609 + ], + "type": "text", + "content": "4.3 Qualitative Performance" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Here, we offer the visualization results of seen and unseen tasks in Fig. 9 with WR-2. We also test QUART in other quadruped robots like Go2 with minor adjustments. More demos can be found in the supplementary materials." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 394, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 394, + 90, + 447, + 102 + ], + "type": "text", + "content": "QUAR-VLA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 310, + 246 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 310, + 246 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 310, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 310, + 246 + ], + "type": "image", + "image_path": "93c45c3339580cbf4b271fcfab73b85ea49ae10c2cea78578d15a04cc20d0c05.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 265, + 482, + 321 + ], + "lines": [ + { + "bbox": [ + 130, + 265, + 482, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 482, + 321 + ], + "type": "text", + "content": "Fig.9: Visualization of seen and unseen tasks. This figure demonstrates the excellent performance of QUART for seen tasks in both simulated and real-world environments. The left part shows 4 instances of seen tasks in both simulation environment and real-world. The right part shows performances on unseen tasks, which contain unseen verbal, object, and background." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 313, + 114, + 482, + 246 + ], + "blocks": [ + { + "bbox": [ + 313, + 114, + 482, + 246 + ], + "lines": [ + { + "bbox": [ + 313, + 114, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 482, + 246 + ], + "type": "image", + "image_path": "53696d91f86a064b825e390183bb77623289d6480f3ef03f382fc67875b02c50.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 345, + 315, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 345, + 315, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 345, + 315, + 357 + ], + "type": "text", + "content": "5 Conclusion & Future Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 376, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 544 + ], + "type": "text", + "content": "This paper emphasizes the significance of deploying Vision-Language-Action models on quadruped robots. We introduce the concept of Vision-Language-Action tasks for QUAdruped Robots (QUAR-VLA), which seamlessly integrates visual information and instructions from diverse modalities to generate executable actions. Specifically, QUAR-VLA focuses on two main aspects of deploying VLA models for quadruped robots: defining the action space to balance flexibility and efficiency, and addressing the scarcity of large-scale training datasets. To tackle these two questions, we present the QUART models tailored for quadruped robots and the QUARD dataset, which includes diverse tasks such as navigation and manipulation. Our extensive evaluation shows that our approach leads to performant robotic policies and enables QUART to obtain a range of emergent capabilities. This includes generalization to novel objects, the ability to interpret commands not present in the robot training data, and the ability to perform rudimentary reasoning in response to user commands." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 147, + 546, + 462, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 546, + 462, + 558 + ], + "spans": [ + { + "bbox": [ + 147, + 546, + 462, + 558 + ], + "type": "text", + "content": "Furthermore, this work has these areas requiring further enhancements:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "1) Automated Data Collection for Larger Datasets. Future work can explore automated processes for data collection, involving techniques such as active learning and data augmentation can be employed to further expand the dataset size and diversity. 2) Better solution for sim2Real gap. Integrating domain adaptation methods like domain randomization, data augmentation, and more accurate modeling of physics engines will be a better solution compared with direct co-training. 3) Improving Inference Speed for Real-time Control. Future works will explore hardware acceleration techniques and model compression techniques to enable faster and more efficient execution of the models." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "P. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 175 + ], + "type": "text", + "content": "This work was supported by the National Science and Technology Innovation 2030 - Major Project (Grant No. 2022ZD0208800), and NSFC General Program (Grant No. 62176215)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 193, + 197, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 193, + 197, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 193, + 197, + 205 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 217, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 138, + 217, + 481, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 481, + 272 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 481, + 272 + ], + "type": "text", + "content": "1. Ahn, M., Dwibedi, D., Finn, C., Arenas, M.G., Gopalakrishnan, K., Hausman, K., Ichter, B., Irpan, A., Joshi, N., Julian, R., Kirmani, S., Leal, I., Lee, E., Levine, S., Lu, Y., Leal, I., Maddineni, S., Rao, K., Sadigh, D., Sanketi, P., Sermanet, P., Vuong, Q., Welker, S., Xia, F., Xiao, T., Xu, P., Xu, S., Xu, Z.: Autort: Embodied foundation models for large scale orchestration of robotic agents (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 272, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 272, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 272, + 481, + 293 + ], + "type": "text", + "content": "2. Bahl, S., Mendonca, R., Chen, L., Jain, U., Pathak, D.: Affordances from human videos as a versatile representation for robotics (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 293, + 481, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 481, + 316 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 481, + 316 + ], + "type": "text", + "content": "3. Bavishi, R., Olsen, E., Hawthorne, C., Nye, M., Odena, A., Somani, A., Tasirlar, S.: Introducing our multimodal models (2023), https://www.adept.ai/blog/fuyu-8b" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 316, + 481, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 481, + 337 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 481, + 337 + ], + "type": "text", + "content": "4. Belkhale, S., Cui, Y., Sadigh, D.: Hydra: Hybrid robot actions for imitation learning. In: Proceedings of the 7th Conference on Robot Learning (CoRL) (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 337, + 481, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 337, + 481, + 370 + ], + "spans": [ + { + "bbox": [ + 138, + 337, + 481, + 370 + ], + "type": "text", + "content": "5. Bharadhwaj, H., Vakil, J., Sharma, M., Gupta, A., Tulsiani, S., Kumar, V.: Roboagent: Generalization and efficiency in robot manipulation via semantic augmentations and action chunking (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 370, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 481, + 424 + ], + "type": "text", + "content": "6. Bousmalis, K., Irpan, A., Wohlhart, P., Bai, Y., Kelcey, M., Kalakrishnan, M., Downs, L., Ibarz, J., Pastor, P., Konolige, K., et al.: Using simulation and domain adaptation to improve efficiency of deep robotic grasping. In: 2018 IEEE international conference on robotics and automation (ICRA). pp. 4243-4250. IEEE (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 424, + 481, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 424, + 481, + 523 + ], + "spans": [ + { + "bbox": [ + 138, + 424, + 481, + 523 + ], + "type": "text", + "content": "7. Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Chen, X., Choromanski, K., Ding, T., Driess, D., Dubey, A., Finn, C., Florence, P., Fu, C., Arenas, M.G., Gopalakrishnan, K., Han, K., Hausman, K., Herzog, A., Hsu, J., Ichter, B., Irpan, A., Joshi, N., Julian, R., Kalashnikov, D., Kuang, Y., Leal, I., Lee, L., Lee, T.W.E., Levine, S., Lu, Y., Michalewski, H., Mordatch, I., Pertsch, K., Rao, K., Reymann, K., Ryoo, M., Salazar, G., Sanketi, P., Sermanet, P., Singh, J., Singh, A., Soricut, R., Tran, H., Vanhoucke, V., Vuong, Q., Wahid, A., Welker, S., Wohlhart, P., Wu, J., Xia, F., Xiao, T., Xu, P., Xu, S., Yu, T., Zitkovich, B.: Rt-2: Vision-language-action models transfer web knowledge to robotic control (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 523, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 481, + 567 + ], + "type": "text", + "content": "8. Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Chen, X., Choromanski, K., Ding, T., Driess, D., Dubey, A., Finn, C., et al.: Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 567, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 481, + 666 + ], + "type": "text", + "content": "9. Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Dabis, J., Finn, C., Gopalakrishnan, K., Hausman, K., Herzog, A., Hsu, J., Ibarz, J., Ichter, B., Irpan, A., Jackson, T., Jesmonth, S., Joshi, N.J., Julian, R., Kalashnikov, D., Kuang, Y., Leal, I., Lee, K.H., Levine, S., Lu, Y., Malla, U., Manjunath, D., Mordatch, I., Nachum, O., Parada, C., Peralta, J., Perez, E., Pertsch, K., Quiambao, J., Rao, K., Ryoo, M., Salazar, G., Sanketi, P., Sayed, K., Singh, J., Sontakke, S., Stone, A., Tan, C., Tran, H., Vanhoucke, V., Vega, S., Vuong, Q., Xia, F., Xiao, T., Xu, P., Xu, S., Yu, T., Zitkovich, B.: Rt-1: Robotics transformer for real-world control at scale (2023)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 394, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 394, + 90, + 447, + 102 + ], + "type": "text", + "content": "QUAR-VLA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "10. Caluwaerts, K., Iscen, A., Kew, J.C., Yu, W., Zhang, T., Freeman, D., Lee, K.H., Lee, L., Saliceti, S., Zhuang, V., et al.: Barkour: Benchmarking animal-level agility with quadruped robots. arXiv preprint arXiv:2305.14654 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 171 + ], + "type": "text", + "content": "11. Eckert, P., Ijspeert, A.J.: Benchmarking agility for multilegged terrestrial robots. IEEE Transactions on Robotics 35(2), 529-535 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 480, + 214 + ], + "type": "text", + "content": "12. Gu, J., Kirmani, S., Wohlhart, P., Lu, Y., Arenas, M.G., Rao, K., Yu, W., Fu, C., Gopalakrishnan, K., Xu, Z., Sundaresan, P., Xu, P., Su, H., Hausman, K., Finn, C., Vuong, Q., Xiao, T.: Rt-trajectory: Robotic task generalization via hindsight trajectory sketches (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 480, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 480, + 235 + ], + "type": "text", + "content": "13. Haldar, S., Mathur, V., Yarats, D., Pinto, L.: Watch and match: Supercharging imitation with regularized optimal transport. CoRL (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 235, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 235, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 133, + 235, + 480, + 289 + ], + "type": "text", + "content": "14. Hutter, M., Gehring, C., Jud, D., Lauber, A., Bellicoso, C.D., Tsounis, V., Hwangbo, J., Bodie, K., Fankhauser, P., Bloesch, M., Diethelm, R., Bachmann, S., Melzer, A., Hoepflinger, M.: Anymal - a highly mobile and dynamic quadrupedal robot. In: 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 38-44 (2016). https://doi.org/10.1109/IROS.2016.7758092" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 289, + 480, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 289, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 133, + 289, + 480, + 322 + ], + "type": "text", + "content": "15. Jiang, Y., Moseson, S., Saxena, A.: Efficient grasping from rgbd images: Learning using a new rectangle representation. In: 2011 IEEE International conference on robotics and automation. pp. 3304-3311. IEEE (2011)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 322, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 322, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 133, + 322, + 480, + 364 + ], + "type": "text", + "content": "16. Kareer, S., Yokoyama, N., Batra, D., Ha, S., Truong, J.: Vinl: Visual navigation and locomotion over obstacles. 2023 IEEE International Conference on Robotics and Automation (ICRA) pp. 2018-2024 (2022), https://api-semanticscholar.org/CorpusID:253117178" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 364, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 364, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 133, + 364, + 480, + 407 + ], + "type": "text", + "content": "17. Karnan, H., Nair, A., Xiao, X., Warnell, G., Pirk, S., Toshev, A., Hart, J., Biswas, J., Stone, P.: Socially compliant navigation dataset (scand): A large-scale dataset of demonstrations for social navigation. IEEE Robotics and Automation Letters 7(4), 11807-11814 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 407, + 480, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 407, + 480, + 440 + ], + "spans": [ + { + "bbox": [ + 133, + 407, + 480, + 440 + ], + "type": "text", + "content": "18. Lee, K.H., Nachum, O., Yang, M.S., Lee, L., Freeman, D., Guadarrama, S., Fischer, I., Xu, W., Jang, E., Michalewski, H., et al.: Multi-game decision transformers. Advances in Neural Information Processing Systems 35, 27921-27936 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 440, + 480, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 440, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 133, + 440, + 480, + 483 + ], + "type": "text", + "content": "19. Lee, M.A., Zhu, Y., Srinivasan, K., Shah, P., Savarese, S., Fei-Fei, L., Garg, A., Bohg, J.: Making sense of vision and touch: Self-supervised learning of multimodal representations for contact-rich tasks. In: 2019 IEEE International Conference on Robotics and Automation (ICRA) (2019), https://arxiv.org/abs/1810.10191" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 483, + 480, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 483, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 133, + 483, + 480, + 515 + ], + "type": "text", + "content": "20. Li, X., Liu, M., Zhang, H., Yu, C., Xu, J., Wu, H., Cheang, C., Jing, Y., Zhang, W., Liu, H., et al.: Vision-language foundation models as effective robot imitators. arXiv preprint arXiv:2311.01378 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 515, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 515, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 133, + 515, + 480, + 557 + ], + "type": "text", + "content": "21. Majumdar, A., Yadav, K., Arnaud, S., Ma, Y.J., Chen, C., Silwal, S., Jain, A., Berges, V.P., Abbeel, P., Malik, J., Batra, D., Lin, Y., Maksymets, O., Rajeswaran, A., Meier, F.: Where are we in the search for an artificial visual cortex for embodied intelligence? (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 557, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 557, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 133, + 557, + 480, + 590 + ], + "type": "text", + "content": "22. Makoviychuk, V., Wawrzyniak, L., Guo, Y., Lu, M., Storey, K., Macklin, M., Hoeller, D., Rudin, N., Allshire, A., Handa, A., State, G.: Isaac gym: High performancegpu-based physics simulation for robot learning (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 590, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 590, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 133, + 590, + 480, + 633 + ], + "type": "text", + "content": "23. Margolis, G.B., Agrawal, P.: Walk these ways: Tuning robot control for generalization with multiplicity of behavior. In: Liu, K., Kulic, D., Ichnowski, J. (eds.) Proceedings of The 6th Conference on Robot Learning. Proceedings of Machine Learning Research, vol. 205, pp. 22-31. PMLR (14-18 Dec 2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 633, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 633, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 633, + 480, + 666 + ], + "type": "text", + "content": "24. Nair, S., Rajeswaran, A., Kumar, V., Finn, C., Gupta, A.: R3m: A universal visual representation for robot manipulation. In: Conference on Robot Learning (2022), https://api-semanticscholar.org/CorpusID:247618840" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "P. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 611 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "25. Nair, S., Rajeswaran, A., Kumar, V., Finn, C., Gupta, A.: R3m: A universal visual representation for robot manipulation (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "type": "text", + "content": "26. Pari, J., Shafiullah, N.M., Arunachalam, S.P., Pinto, L.: The surprising effectiveness of representation learning for visual imitation (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 482, + 194 + ], + "type": "text", + "content": "27. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "type": "text", + "content": "28. Reed, S., Zolna, K., Parisotto, E., Colmenarejo, S.G., Novikov, A., Barth-Maron, G., Gimenez, M., Sulsky, Y., Kay, J., Springenberg, J.T., et al.: A generalist agent. arXiv preprint arXiv:2205.06175 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 226, + 482, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 258 + ], + "type": "text", + "content": "29. Reed, S., Zolna, K., Parisotto, E., Colmenarejo, S.G., Novikov, A., Barth-Maron, G., Gimenez, M., Sulsky, Y., Kay, J., Springenberg, J.T., et al.: A generalist agent. arXiv preprint arXiv:2205.06175 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 258, + 482, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 303 + ], + "type": "text", + "content": "30. Schiavi, G., Wulkop, P., Rizzi, G., Ott, L., Siegwart, R., Chung, J.J.: Learning agent-aware affordances for closed-loop interaction with articulated objects. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 5916-5922 (2023). https://doi.org/10.1109/ICRA48891.2023.10160747" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 303, + 482, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 325 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 325 + ], + "type": "text", + "content": "31. Shah, D., Sridhar, A., Dashora, N., Stachowicz, K., Black, K., Hirose, N., Levine, S.: Vint: A foundation model for visual navigation (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 325, + 482, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 482, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 482, + 347 + ], + "type": "text", + "content": "32. Shilane, P., Min, P., Kazhdan, M., Funkhouser, T.: The princeton shape benchmark. In: Proceedings Shape Modeling Applications, 2004. pp. 167-178 (2004)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 347, + 482, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 369 + ], + "type": "text", + "content": "33. Shridhar, M., Manuelli, L., Fox, D.: CIoport: What and where pathways for robotic manipulation. In: Conference on Robot Learning. pp. 894-906. PMLR (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 369, + 482, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 402 + ], + "type": "text", + "content": "34. Stone, A., Xiao, T., Lu, Y., Gopalakrishnan, K., Lee, K.H., Vuong, Q., Wohlhart, P., Kirmani, S., Zitkovich, B., Xia, F., Finn, C., Hausman, K.: Open-world object manipulation using pre-trained vision-language models (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 402, + 482, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 435 + ], + "type": "text", + "content": "35. Szot, A., Schwarzer, M., Agrawal, H., Mazoure, B., Talbott, W., Metcalf, K., Mackraz, N., Hjelm, D., Toshev, A.: Large language models as generalizable policies for embodied tasks. arXiv preprint arXiv:2310.17722 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 435, + 482, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 456 + ], + "type": "text", + "content": "36. Tang, Y., Yu, W., Tan, J., Zen, H., Faust, A., Harada, T.: Saytap: Language to quadrupedal locomotion. arXiv preprint arXiv:2306.07580 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 456, + 482, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 489 + ], + "type": "text", + "content": "37. Wei, J., Hou, L., Lampinen, A., Chen, X., Huang, D., Tay, Y., Chen, X., Lu, Y., Zhou, D., Ma, T., et al.: Symbol tuning improves in-context learning in language models. arXiv preprint arXiv:2305.08298 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 489, + 482, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 489, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 489, + 482, + 522 + ], + "type": "text", + "content": "38. Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., Xiao, J.: 3d shapenets: A deep representation for volumetric shapes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1912-1920 (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 522, + 482, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 565 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 565 + ], + "type": "text", + "content": "39. Yang, R., Zhang, M., Hansen, N., Xu, H., Wang, X.: Learning vision-guided quadrupedal locomotion end-to-end with cross-modal transformers. ArXiv abs/2107.03996 (2021), https://api_semanticscholar.org/CorpusID: 235765481" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 565, + 482, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 565, + 482, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 565, + 482, + 611 + ], + "type": "text", + "content": "40. Yu, K.T., Bauza, M., Fazeli, N., Rodriguez, A.: More than a million ways to be pushed. a high-fidelity experimental dataset of planar pushing. In: 2016 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 30-37. IEEE (2016)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "QUAR-VLA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Quality Assured_ Rethinking Annotation Strategies in Imaging AI/1dd3eeac-8808-45a9-8928-d1c207e0686a_content_list.json b/2024/Quality Assured_ Rethinking Annotation Strategies in Imaging AI/1dd3eeac-8808-45a9-8928-d1c207e0686a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..82c7516ca57c2f8f86d4be6a9fa736597c11d1e6 --- /dev/null +++ b/2024/Quality Assured_ Rethinking Annotation Strategies in Imaging AI/1dd3eeac-8808-45a9-8928-d1c207e0686a_content_list.json @@ -0,0 +1,1637 @@ +[ + { + "type": "text", + "text": "Quality Assured: Rethinking Annotation Strategies in Imaging AI", + "text_level": 1, + "bbox": [ + 269, + 140, + 736, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tim Rädsch1,2,3, Annika Reinke1,2, Vivienne Weru1, Minu D. Tizabi1, Nicholas Heller4, Fabian Isensee1,2, Annette Kopp-Schneider1*, and Lena Maier-Hein1,2,3,5*", + "bbox": [ + 220, + 218, + 781, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ German Cancer Research Center, GER $^{2}$ Helmholtz Imaging, GER $^{3}$ Heidelberg University, GER $^{4}$ University of Minnesota, USA $^{5}$ National Center for Tumor Diseases, GER $\\boxed{\\boxtimes}$ :{tim.raedsch; l.maier-hein}@dkfz-heidelberg.de", + "bbox": [ + 235, + 276, + 769, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. This paper does not describe a novel method. Instead, it studies an essential foundation for reliable benchmarking and ultimately real-world application of AI-based image analysis: generating high-quality reference annotations. Previous research has focused on crowdsourcing as a means of outsourcing annotations. However, little attention has so far been given to annotation companies, specifically regarding their internal quality assurance (QA) processes. Therefore, our aim is to evaluate the influence of QA employed by annotation companies on annotation quality and devise methodologies for maximizing data annotation efficacy. Based on a total of 57,648 instance segmented images obtained from a total of 924 annotators and 34 QA workers from four annotation companies and Amazon Mechanical Turk (MTurk), we derived the following insights: (1) Annotation companies perform better both in terms of quantity and quality compared to the widely used platform MTurk. (2) Annotation companies' internal QA only provides marginal improvements, if any. However, improving labeling instructions instead of investing in QA can substantially boost annotation performance. (3) The benefit of internal QA depends on specific image characteristics. Our work could enable researchers to derive substantially more value from a fixed annotation budget and change the way annotation companies conduct internal QA.", + "bbox": [ + 261, + 344, + 743, + 623 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Data Annotation $\\cdot$ Quality Assurance $\\cdot$ Crowdsourcing", + "bbox": [ + 261, + 633, + 715, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 669, + 375, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Following a long period of extensive focus on pure method development, proper benchmarking and validation have recently been recognized as a major challenge in AI-based image analysis [33, 43]. In this context, high-quality annotation of reference datasets is a fundamental requirement to ensure the meaningful validation of image analysis algorithms [15]. As annotation tasks may be cumbersome, crowdsourcing is often used to partially outsource image annotation. Traditionally, such outsourcing is conducted via crowdsourcing platforms [32], with Amazon Mechanical Turk (MTurk)1 as the predominant choice [9, 49].", + "bbox": [ + 212, + 696, + 787, + 820 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1 https://www.mturk.com * shared last author", + "bbox": [ + 217, + 824, + 537, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/061f4698a80d451656990b78fc9c077e4110fb5d60a98b97de5f2fb5c2e4a271.jpg", + "image_caption": [ + "Fig. 1: Research Questions (RQs) tackled in this work. Based on 57,648 instance segmentation masks annotated by 924 annotators and 34 quality assurance (QA) workers from five different annotation providers, we (1) compared the effectiveness of generating high-quality annotations between annotation companies and Amazon Mechanical Turk and (2) investigated the effects of annotation companies' internal QA and (3) real-world image characteristics on the annotation quality." + ], + "image_footnote": [], + "bbox": [ + 217, + 142, + 787, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent literature has reflected the rise of annotation companies as a new type of provider with potentially better annotation performance compared to established crowdsourcing platforms such as MTurk [28,36,42,44]. Annotation companies differ substantially from crowdsourcing platforms. First, they employ annotators directly and at shared office spaces [28,44], enabling better external control of worker conditions and stricter data security measures, while crowdsourcing platforms can be accessed by anyone, without specific knowledge or training being required. Second, annotators in annotation companies typically work in small teams, with an experienced quality assurance (QA) worker being responsible for the team's output. In the observed internal QA, QA workers make adjustments to the annotations generated by the team by reviewing each individual annotation and modifying those where they deem changes necessary. As a result, the annotation requester receives the quality-assured annotation, not the initial one (see Fig. 2a). In contrast, crowdsourcing platforms usually do not employ internal QA mechanics similar to annotation companies.", + "bbox": [ + 212, + 425, + 787, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The importance of annotation providers for real-world applications is evident at the major computer vision conferences, where over the past two years, approximately $20\\%$ of exhibitors were annotation providers (see Fig. 2b). While QA as a post-processing step for annotations from platforms such as MTurk has been researched (e.g. [10, 17, 21, 27]), little is known about the benefit of internal QA on annotation quality within annotation companies. The only related work that we are aware of pertains to external QA from the perspective of an annotation company. Unlike in internal QA, in the case of external QA, domain experts or experienced annotators working on behalf of the requester check the annotation quality as a post-processing step of the received annotations [31, 47]. Our paper, on the other hand, focuses on the internal QA processes of an annotation company.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "T. Rädsch et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/08608e7b3b91ceaa4f4d71b9628e1435ebf43201ebda0d9b769d31413438f963.jpg", + "image_caption": [ + "Fig. 2: a) In contrast to Amazon Mechanical Turk (MTurk), annotation companies commonly perform quality assurance (QA) before delivering the result to the annotation requester. b) Annotation providers are an integral part of the computer vision community. They account for around $20\\%$ of exhibitors present at major computer vision conferences (ECCV 2022: $\\mathrm{n} = 40$ , CVPR 2022: $\\mathrm{n} = 97$ , ICCV 2023: $\\mathrm{n} = 43$ , CVPR 2023: $\\mathrm{n} = 104$ ) in the last two years." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 309 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In light of the vast sums of money to be spent on QA [40], the purpose of this paper is to shed light on whether annotation companies' internal QA benefits the annotation quality. Based on more than 50,000 instance segmentation masks annotated by more than 900 annotators from five different annotation providers, we investigate the following research questions (RQs; see Fig. 1):", + "bbox": [ + 212, + 438, + 787, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "RQ 1) How does the choice of annotation provider influence the quality of annotations? As the selection of the annotation provider is among the first and key decisions in any annotation project after data collection and generation of labeling instructions, we first investigated companies' effectiveness in generating annotations, which is particularly relevant in cases of constrained annotation budgets requiring a cost-effective strategy. As both quantity and quality of annotated images for a given budget are relevant in provider selection, we compared the annotation quality and costs of five different providers, including MTurk and four annotation companies, on identical images.", + "bbox": [ + 212, + 517, + 787, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "RQ 2) Does annotation companies' quality assurance improve annotation quality? A major decision in the design of an annotation process concerns the allocation of resources. A common perception is that internal QA of annotation companies, following the initial annotation, substantially enhances the quality. Here, we challenged this assumption by investigating the improvement in annotation quality from internal QA compared to that achieved through enhanced labeling instructions.", + "bbox": [ + 212, + 656, + 787, + 761 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "RQ 3) Should quality assurance focus on specific subsets of images? In theory, the efficacy of QA may be contingent upon specific image characteristics [20, 24, 41]. For instance, underexposed images might benefit more from a secondary review compared to those with optimal illumination. Concurrently, images that present annotation challenges for human observers might be partic", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Quality Assured: Rethinking Annotation Strategies in Imaging AI", + "bbox": [ + 290, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ularly pertinent for the training and validation of algorithms [20, 41]. Here, we investigated whether and to what extent the benefit of QA depends on specific image characteristics.", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 210, + 388, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Comparison of annotation providers", + "text_level": 1, + "bbox": [ + 214, + 237, + 558, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With respect to the quality of annotations obtained from varying annotation providers, the closest work conducted are comparisons between crowdsourcing platforms and domain experts. Irshad et al. [23] and Petrovic et al. [38], among others, have shown that crowdsourced annotations can reach the annotation quality of domain experts in safety-critical applications. In recent literature, Duggan et al. [13] and Kentley et al. [25] compared domain-specialized crowds as a new type of annotation provider and found comparable performance to domain experts. Furthermore, Rädsch et al. [42] provided a comparison of annotation companies to crowdsourcing, but focused solely on raw annotation quality without QA and without relating it to annotation budgets.", + "bbox": [ + 212, + 258, + 787, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Interestingly, researchers report concerns about the annotation quality obtained from MTurk [31], such as obtaining unusable annotations in up to $67.8\\%$ of cases [6]. Other research fields, such as psychology, even report a \"substantial decrease in data quality\" [7]. However, we are not aware of work that has systematically compared the performance between annotation companies and crowdsourcing platforms such as MTurk while considering annotation budgets.", + "bbox": [ + 212, + 409, + 787, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Quality assurance", + "text_level": 1, + "bbox": [ + 214, + 517, + 410, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The main goal of QA is to enhance and ensure the quality of annotations. A comprehensive overview on QA in crowdsourcing platforms is provided by Daniel et al. [10]. However, their research does not include annotation companies. During the annotation process, quality control techniques generally fall into three categories [31]: (1) comparing to a gold standard with pre-defined reference annotations (e.g., [27]), (2) analyzing combined outputs from multiple annotators (e.g., [21]), and (3) investigating annotator behaviors (e.g., [17, 19]). Other approaches to enhancing annotation quality on crowdsourcing platforms target training of crowdworkers [12] and iterative refinement of the annotation environment [3,4,14]. It should be noted that annotation byproducts, such as mouse movement during annotation generation [16, 17] or inter-rater agreement [48], can be used to boost model performance, by incorporating this information in the training.", + "bbox": [ + 212, + 537, + 787, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The closest related work focusing on the impact of QA in annotation companies was introduced by Wang et al. [47] and Lu et al. [31]. However, their work focuses on external QA conducted by experienced personnel from the requester rather than internal QA, and features a substantially smaller number of annotators from annotation companies. To our knowledge, a systematic investigation into the effects of internal QA processes within annotation companies is thus, so far, lacking.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "T. Rädsch et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Impact of real-world image characteristics on annotation quality", + "text_level": 1, + "bbox": [ + 215, + 146, + 785, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Related research analyzing the impact of real-world image characteristics highlights their effect on annotations created by machine learning models and annotators alike. Michaelis et al. [37] demonstrated that models easily encounter failures with increasingly challenging real-world image characteristics, such as added snow or poor lighting characteristics. Similarly, annotators generate less reliable annotations in the presence of challenging real-world image characteristics. This effect can be observed both within [1,24,31] and across domains [18,26]. However, a systematic analysis of the impact of such characteristics on internal QA employed by annotation companies is lacking.", + "bbox": [ + 212, + 172, + 787, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methods", + "text_level": 1, + "bbox": [ + 215, + 332, + 338, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the following, we describe the selected data, the labeling instructions, annotation providers and annotators, the QA process, and the experimental design of our work.", + "bbox": [ + 212, + 364, + 785, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Data", + "text_level": 1, + "bbox": [ + 215, + 433, + 303, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conducted our experiments on the instance segmentation data from the Heidelberg Colorectal (HeiCo) dataset for surgical data science in the sensor operating room [34], representing a safety-critical application that crowdsourcing platforms can still handle. The task involved segmenting surgical instruments from laparoscopic videos, resulting in 57,648 instance segmentation masks on 4,050 unique frames. Furthermore, we used 57,636 metadata annotations on the unique images to quantify the effect of varying real-world image characteristics (see RQ3). Characteristics included motion blur, underexposure, object occlusion, or overlapping objects, among others, and are illustrated in Suppl. A.", + "bbox": [ + 212, + 458, + 787, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Labeling instructions", + "text_level": 1, + "bbox": [ + 215, + 617, + 434, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To investigate the benefit of QA compared to an investment in detailed labeling instructions, we created three distinct types of labeling instructions with varying levels of information density, namely, (1) minimal text, (2) extended text, and (3) extended text including exemplary pictures. Each subsequent labeling instruction type was structured to contain more detailed information than the preceding. To ensure accurate labeling of the imaging data, we established a collection of design principles applicable to all labeling instructions:", + "bbox": [ + 212, + 643, + 787, + 750 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) Labeling information is presented in a slide format to enhance the processing of information by humans.", + "(2) Each slide contains a distinct 'chunk' of information, which is a self-contained piece of content. This approach of 'chunking' information lessens the cognitive load on the working memory.", + "(3) Related information chunks are placed in close proximity to one another." + ], + "bbox": [ + 214, + 750, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Quality Assured: Rethinking Annotation Strategies in Imaging AI", + "bbox": [ + 290, + 114, + 730, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(4) A uniform design template is used throughout, featuring specific fonts, symbols, and color schemes.", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The three labeling instructions differ as follows:", + "bbox": [ + 214, + 176, + 560, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Minimal text labeling instructions: These instructions provide a concise textual outline, offering examples of typical annotations and highlighting the most frequent annotation types (see Suppl. B.1). They reflect a scenario with minimal effort invested in developing the instructions. The basic framework includes seven slides totaling 168 words.", + "bbox": [ + 212, + 191, + 784, + 267 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Extended text labeling instructions: Building upon the simplified version, the extended text instructions offer a thorough textual explanation, supplemented by both positive and negative examples (refer to Suppl. B.2). They cover both regular and some rare annotation scenarios. The enhanced version comprises ten slides with a total of 446 words.", + "bbox": [ + 212, + 267, + 785, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Extended text including pictures labeling instructions: These instructions augment the enhanced textual guidelines with images (detailed in Suppl. B.3). The visual aids feature descriptions, symbols, annotations, and color coding to effectively communicate the details on the slides. They also include illustrations of rare annotation cases, indicating a deep and well-documented understanding of the labeling process. This set contains 16 slides with a total of 961 words. We refer the reader to Suppl. B for full details of each labeling instruction type.", + "bbox": [ + 212, + 343, + 787, + 449 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Annotation providers", + "text_level": 1, + "bbox": [ + 214, + 469, + 439, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The study was conducted based on 216 annotators and 34 QA workers from four annotation companies and 708 crowdworkers from the crowdsourcing platform MTurk. Our goal was to provide the best representation for both annotation provider types.", + "bbox": [ + 212, + 492, + 784, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Consequently, we established higher standards for our MTurk crowdworkers compared to the usual research benchmarks, which generally require a $95\\%$ acceptance rate of Human Intelligence Tasks (HITs) and at least 100 completed tasks [3, 6, 17]. Our criteria included a $98\\%$ acceptance rate with over 5,000 completed HITs. We spread out the HITs over 40 days at all times to obtain a representative MTurk sample and compensated workers fairly, as recommended by Litman et al. [30]. The annotation companies we selected have a solid background in handling large-scale annotation tasks in a variety of domains and operate internationally. The annotators were allocated by the companies themselves to replicate the typical workflow of a project carried out by an annotation company. Each annotator was linked to a specific annotation provider that had been predetermined. No selection process from our side took place for either the annotators, QA workers, or the MTurk crowdworkers.", + "bbox": [ + 212, + 553, + 785, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To prevent information leakage, we conducted the labeling process with increasingly detailed labeling instructions (minimal text instruction, extended text instruction, and extended text including pictures instruction). Each annotator worked exclusively with one of the three instruction types as introduced in Sec. 3.2. To enhance security, we included a mandatory minimum gap of ten days between consecutive labeling instructions.", + "bbox": [ + 212, + 750, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "T. Rädsch et al.", + "bbox": [ + 271, + 114, + 380, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4 Quality assurance process", + "text_level": 1, + "bbox": [ + 215, + 146, + 475, + 162 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the internal QA process, annotators typically report to a small number of experienced QA workers, who are often responsible for training their team and, in our case, responsible for the QA of the annotations.", + "bbox": [ + 212, + 176, + 785, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The general workflow for all companies consisted of an annotation stage, in which images were assigned randomly to each annotator, and a QA stage. During the annotation stage, annotators received raw images with the task of annotating them until they considered the quality of annotations satisfactory. Annotators were aware that their work would be checked by QA. Once submitted, the QA worker gained access to the generated annotations, allowing them to make modifications until they were satisfied before submitting the final annotations for delivery to the requester. Throughout this process, both annotators and QA workers retained access to the respective labeling instructions at any given moment. We obtained access to the annotations both before and after QA through a software company handling a substantial number of annotation projects annually. The experiments were seamlessly integrated into the software company's pipeline of annotation projects, reducing the risk of exposing our experiments. Pricing was established to be representative of other projects of similar size.", + "bbox": [ + 212, + 223, + 787, + 434 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To maintain fairness and avoid giving any annotation provider a potential informational advantage that may have affected our statistical analysis, no specific queries from the annotation companies about the content of the labeling instructions were addressed.", + "bbox": [ + 212, + 434, + 787, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.5 Experimental design", + "text_level": 1, + "bbox": [ + 214, + 522, + 431, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "RQ 1: Each of the five labeling providers annotated the same images with each set of labeling instructions, employing separate annotators for each task. We analyzed the number of spam images, images with severe errors, i.e., images with at least one false positive and/or false negative, and the overall annotation quality with the Dice Similarity Coefficient (DSC) [11] as the overlap-based metric and the Normalized Surface Distance (NSD) [39] as the distance-based metric. We define a spam annotation as a purposefully poor-quality annotation with the intent to obtain the payment without properly conducting the annotation task. Annotation duration was used as an auxiliary measure, where accessible. For instance, if an annotator completed the task in only 17 seconds, they could not have read the instructions and properly performed annotations on several images. Of note, a spammer could wait a relevant number of seconds and then enter a poor-quality annotation to bypass the annotation duration check. Thus, each image was additionally visually inspected by an engineer with extensive labeling experience. Spam annotation examples are displayed in Suppl. D.", + "bbox": [ + 212, + 551, + 787, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "RQ 2: The annotation companies conducted their internal QA processes directly after their annotators completed the annotation job, as explained in Sec. 3.4 and Fig. 2a. The experiments began with only providing the minimal text labeling instructions, followed by the extended text labeling instructions,", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Quality Assured: Rethinking Annotation Strategies in Imaging AI", + "bbox": [ + 290, + 114, + 730, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bfbdf51154c1873a74b129cde29ab08df2292fbbea703a12bc8b83545e5fdd59.jpg", + "image_caption": [ + "Fig. 3: Annotation companies are more efficient at generating high-quality images than Amazon Mechanical Turk (MTurk). Percentage of images without severe errors (green), images with severe errors (orange), and spam (red) obtained for the five different annotation providers investigated. Annotation costs per image were lower for all companies compared to MTurk (median: $61.1\\%$ )." + ], + "image_footnote": [], + "bbox": [ + 272, + 143, + 733, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and finally, the extended text instructions with pictures, to prevent information leakage.", + "bbox": [ + 212, + 396, + 784, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For each annotated image, we calculated the DSC and NSD per instance, aggregated the scores per annotated image, and aggregated the scores over annotation company and labeling instruction. We then compared the annotation quality without and with QA across the three types of labeling instructions for the four annotation companies by computing both absolute and relative changes in DSC/NSD aggregates, the percentage of modified images, and the change in the percentage of severe errors, among other parameters.", + "bbox": [ + 212, + 426, + 784, + 531 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "RQ 3: In contrast to MTurk, the four annotation companies conducted an additional annotation round, including QA, with the extended text including pictures labeling instructions to quantify the impact of different real-world image characteristics, represented by 57,636 metadata annotations, on the probability of improvement. The data was analyzed with a logistic mixed model [35] which included random effects for the image, the annotator, and the annotation company. Further method details are provided in Suppl. C.", + "bbox": [ + 212, + 532, + 785, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Results", + "text_level": 1, + "bbox": [ + 214, + 661, + 326, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 How does the choice of annotation provider influence the quality and cost of annotations?", + "text_level": 1, + "bbox": [ + 214, + 693, + 740, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Annotated images generated by the annotation companies were of higher quality, comprising no spam (MTurk: $\\sim 20\\%$ ) and a much higher proportion of remaining images without severe annotation errors (factor of $>2$ improvement for all companies compared to MTurk), as shown in Fig. 3. In addition, the annotation companies were $61.1\\%$ cheaper in the median, despite their additional employment of QA workers. More specifically, the costs ranged from $24.5\\%$ to $81.5\\%$ of the MTurk costs for the same images and number of annotations. We have", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "T. Rädsch et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fe018066ac0287db08f641d2088b2f831e466d2d8629b14fe815d48281c48e45.jpg", + "image_caption": [ + "(a)", + "Fig. 4: QA of annotation companies only provides marginal improvements, if any. These depend highly on the company and type of labeling instruction. a: Absolute change of the median DSC resulting from performing QA. For each company and labeling instruction, the median was obtained by aggregating the DSC scores over the images. b: Corresponding absolute values of the DSC. Positive and negative changes from annotation to QA are colored according to the color scale." + ], + "image_footnote": [ + "Annotation company Company 1 Company 2 Company 3 Company 4" + ], + "bbox": [ + 215, + 155, + 444, + 277 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/70135a47f093ad8a5b103dfad0fde1b89700ec9c33ac790fa31e80e9afb72070.jpg", + "table_caption": [ + "(b)" + ], + "table_footnote": [ + "COLOR SCALE 1.00 0.75 0.50 0.25 0.00 -0.25 -0.50 0.75 -1.00" + ], + "table_body": "
CompanyLabeling instruction25th PercentileMedian75th Percentile
AnnotationQAAnnotationQAAnnotationQA
11) Min. text0.610.620.930.930.960.96
2) Ext. text0.580.580.930.930.960.96
3) Ext. text + pictures0.900.900.950.950.960.96
21) Min. text0.570.470.920.900.960.95
2) Ext. text0.620.840.930.940.960.96
3) Ext. text + pictures0.850.880.940.940.960.96
31) Min. text0.600.600.930.930.960.96
2) Ext. text0.470.640.910.930.960.96
3) Ext. text + pictures0.740.850.930.940.960.96
41) Min. text0.630.630.940.940.960.96
2) Ext. text0.880.880.940.940.960.96
3) Ext. text + pictures0.910.910.950.950.970.97
", + "bbox": [ + 460, + 155, + 785, + 301 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "not been given permission to reveal the identities and exact absolute budgets used for the annotation companies, hence we report costs relative to the MTurk budget. Examples of spam annotations are provided in Suppl. D.", + "bbox": [ + 212, + 445, + 782, + 491 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Regardless of the annotation scenario, the annotation companies generated higher quality annotations than MTurk with higher DSC scores including smaller interquartile ranges (IQRs) (see Fig. 7a), higher NSD scores including smaller IQRs (see Fig. 7b), and lower numbers of severe errors (see Fig. 7c).", + "bbox": [ + 212, + 491, + 784, + 551 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Does annotation companies' quality assurance improve annotation quality?", + "text_level": 1, + "bbox": [ + 214, + 573, + 718, + 603 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As shown in Fig. 4, the internal QA conducted by annotation companies does not substantially improve annotation quality if the QA workers only use minimal text instructions. The impact of QA when having access to extended text instructions or extended text including pictures instructions depended highly on the annotation company, while the effect was still minimal compared to not performing QA: Two of the four annotation companies show no change in the median DSC when comparing QA to non-QA outputs for extended text or extended text and picture instructions. The remaining two companies only show a very small relative improvement of below $2\\%$ .", + "bbox": [ + 212, + 613, + 784, + 748 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We obtained even smaller improvements with the NSD scores, as depicted in Suppl. E. In certain scenarios, QA was able to remove a small proportion of severe errors, however, the vast majority remains (see Suppl. F). Notably, the percentage of modified images by QA workers depended highly on the company, as depicted in Suppl. G. Fig. 5 further illustrates that providing annotators with high-quality labeling instructions including pictures (green) leads them to vastly", + "bbox": [ + 212, + 750, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Quality Assured: Rethinking Annotation Strategies in Imaging AI", + "bbox": [ + 290, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/6e4b981eabe812cf9322d05a5db34aeb5fc9d0ddce135d8bc953de53d4effa00.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 143, + 467, + 315 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/e21b511fe7d170cc6b84b0cb6c2b2f9f5a415a00e726d65b9c1f5eab932c261f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 142, + 785, + 315 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/8441f60cd3965c736c8ad62c652025e0f18cfdc2ee53735e671eabfc2dd3b52f.jpg", + "image_caption": [ + "Fig. 5: Improving the labeling instructions yields a higher effect compared to adding QA to the annotation process. The baseline performance (minimal labeling instruction, no QA) according to the DSC (a) and the percentage of severe errors (b) is shown individually per company with a red asterisk (minimal labeling instruction, no QA). While adding QA only yields marginal improvements (dark blue), major performance boosts are obtained by improving the labeling instructions provided to the annotators (green). DSC scores were aggregated for each image and per labeling type. DSC displayed as dots and box plot (band indicates the median, box indicates the first (25th percentile) and third (75th percentile) quartiles). $0 \\leq \\mathrm{DSC} \\leq 1$ ." + ], + "image_footnote": [], + "bbox": [ + 240, + 323, + 750, + 353 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "outperform annotators working with lower-quality labeling instructions, even if these are followed by additional internal QA (dark blue).", + "bbox": [ + 212, + 518, + 787, + 550 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Should quality assurance focus on specific subsets of images?", + "text_level": 1, + "bbox": [ + 214, + 569, + 763, + 585 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As shown in Fig. 6a, QA significantly increases the odds of improving the annotation quality for images with generalizable image characteristics such as underexposed objects, intersection of objects, objects occluded by background in comparison to the absence of this characteristic (vertical black line). The only exceptions are motion artifacts present in the background and object(s) overexposure.", + "bbox": [ + 212, + 592, + 787, + 681 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In contrast to the general increase of images with generalizable image characteristics, none of the domain-specific image characteristics yield a significant change in the odds of improving the annotation quality over a regular image, see Fig. 6b. The only exception for domain-specific characteristics is object(s) covered by smoke, which could be interpreted as a form of object occlusion.", + "bbox": [ + 212, + 683, + 787, + 758 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Discussion", + "text_level": 1, + "bbox": [ + 214, + 779, + 354, + 795 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The data annotation market is estimated to reach a volume of more than USD 4.1 billion by the end of 2024 [8]. As the annotation process becomes increasingly", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "T. Rädsch et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/8703522a81b1d11a71e79334a1e70ca4661eaf6523b4fa5500fb88ee5fd6a924.jpg", + "image_caption": [ + "(a) GENERALIZABLE CHARACTERISTICS" + ], + "image_footnote": [], + "bbox": [ + 218, + 160, + 787, + 287 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b1db69be1ef2add7f6d9688be19cea42fec65d9bcba3f8b654b2a3ced5c6f6ac.jpg", + "image_caption": [ + "(b) DOMAIN-SPECIFIC CHARACTERISTICS", + "Fig. 6: QA of annotation companies can be beneficial to certain images with difficult real-world generalizable image characteristics, such as underexposure or intersection of objects. We display the impact of (a) real-world generalizable and (b) domain-specific image characteristics on the odds of improvement. Improvement is defined as a positive difference in the QA, while no improvement refers to 0 or negative difference. The effect of an image characteristic is shown in the form of the log(odds ratio (OR)) representing the logarithms of the odds of improvement in the presence of the characteristic, compared to the odds of improvement occurring in the absence of that characteristic. The vertical black line refers to the point of equal odds. (Note: ***: p-value ≤ 0.001, **: p-value ≤ 0.01, *: p-value ≤ 0.05)." + ], + "image_footnote": [], + "bbox": [ + 218, + 303, + 787, + 465 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "automated and foundation models evolve, the market is undergoing drastic shifts. With lower entry barriers for annotation, more efficient data annotation tools, and a rising impact of QA, it is thought to result in an estimated USD 8.2 billion market by 2028 [40]. As a research community, we need to understand how to ensure high annotation quality throughout this evolution, as reliable reference annotations are an indispensable foundation for the benchmarking of image analysis algorithms, and thus their ultimate translation into real-world practice.", + "bbox": [ + 212, + 654, + 787, + 775 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To our knowledge, this study is the first to quantitatively and critically examine the impact of internal QA conducted by annotation companies in the context of safety-critical applications. The most important insights can be summarized as follows:", + "bbox": [ + 212, + 779, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Quality Assured: Rethinking Annotation Strategies in Imaging AI", + "bbox": [ + 290, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "1. Annotation companies perform better both in terms of quantity and quality compared to the widely used platform MTurk (Fig. 3). This implies that requesters working with crowdsourcing platforms should seriously consider annotation companies as an alternative. The organizational structure of these companies, which enables a single point of contact for requesters and eliminates spam annotations, appears to be a contributing factor to their efficiency. In light of poor working conditions being a major problem in annotation generation [22, 36], annotation companies, which can more easily be inspected than distributed crowdsourcing platforms, are thought to offer better working contracts, enabling fairer working conditions for annotators. It should be noted that not every annotation company automatically ensures adequate working conditions. The annotation companies employed in this study underwent thorough scrutiny to guarantee compliance with good working condition guidelines. With annotators working at a shared location instead of in a distributed crowd, verifying annotators' working conditions has become substantially easier. We thus view our work as an initial step towards greater accountability on the part of annotation requesters. Further dimensions of annotation provider selection are discussed in Suppl. H.", + "bbox": [ + 212, + 146, + 787, + 419 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "2. In contrast to common belief, internal QA mechanisms do not necessarily improve annotation quality (Fig. 4). Furthermore, we observe no systematic QA quality improvement difference between annotation companies depending on their region of origin, size, or price, meaning more expensive annotation companies do not translate to better QA. However, improving labeling instructions instead of investing in QA can substantially boost the annotation performance (Fig. 5). Consequently, in order to obtain high-quality annotations, annotation requesters should prioritize refining their labeling instructions, followed by implementing QA or even a different external QA process as a secondary step. Compared to QA, better labeling instructions can be generated with relatively few resources.", + "bbox": [ + 212, + 422, + 787, + 589 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "3. The benefit of QA depends on the specific image characteristics (Fig. 6). In fact, QA did improve the quality of images with certain challenging real-world image characteristics, see Fig. 6a. Consequently, it might be useful for requesters and annotation companies alike to allocate resources to identifying challenging image characteristics during a project, train classifiers to identify images containing these characteristics and focus the internal QA effort on these images.", + "bbox": [ + 212, + 593, + 787, + 700 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "One could contend that the recent rise of foundation models marginalizes the role of annotation providers. We argue, however, that it only changes their role. In this context, it is necessary to differentiate between the data quality demands for the training and testing of models. With a recent trend towards fine-tuning pre-trained models, smaller high-quality datasets can achieve similar performance with less computation [5, 29], highlighting the importance of the quality of unlabeled data in unsupervised training and its role in reducing carbon emissions [45]. In testing, especially for safety-critical applications such as autonomous driving and clinical medicine, ensuring high-quality annotated test", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "T. Rädsch et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "data is crucial for accurate model evaluation and real-world applicability [43]. Consequently, while training data annotation may lose relevance, test data annotation and its high quality demands are here to stay. Regarding foundation models, it should further be noted that annotations purely produced by foundation models introduce artifacts and biases [2] and represent an increasing rising problem for MTurk, as demonstrated by Veselovsky et al. [46].", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Our study is subject to several limitations which deserve further discussion. Comparing prices across providers is non-trivial as the costs to annotate a given data set depend on many factors including, among others, image quantity, type of annotations, location of annotation company and quality requirements. Our analysis was based on the actual costs incurred during our project. Confidentiality agreements with the software company which integrated our experiments into their annotation project pipeline prevent us from disclosing the identities of the annotation companies. Additionally, we took precautions to protect the identities of individual annotators and QA workers. This was particularly important given our findings that QA workers did not significantly enhance annotation quality, potentially exposing them to negative repercussions.", + "bbox": [ + 212, + 241, + 787, + 407 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Another limitation could be seen in the fact that our data set only represents one safety-critical domain. While this is a valid concern, we would like to highlight that we chose a dataset which combined quality, complexity, and volume. Furthermore, regarding the breadth of annotation providers and the number of annotators involved, our study is likely the largest of its kind, thereby incurring substantial costs.", + "bbox": [ + 212, + 411, + 787, + 501 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Finally, conducting multiple experiments with identical annotation companies and varying datasets could have potentially revealed our experimental design. Had the annotation companies been aware of our research, outcomes could potentially have been biased to appear more favorable, given companies' inherent interest in demonstrating their services as reliable. We thus chose to conduct our experiments such that they could be seamlessly integrated into the software company's annotation pipeline, reducing their likelihood of exposure.", + "bbox": [ + 212, + 505, + 787, + 612 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6 Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 214, + 652, + 532, + 669 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In conclusion, our study underscores the pivotal role of enhancing the initial annotation quality, for instance via allocating resources to the improvement of labeling instructions, over QA in obtaining higher reference annotation quality in high-stakes AI [43]. Future research should delve into the development and implementation of comprehensive training programs for annotators from annotation companies to further enhance initial annotation quality. Exploring the role of annotation companies in the context of increasingly automated annotation processes driven by foundation models presents another avenue for future research.", + "bbox": [ + 212, + 704, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Quality Assured: Rethinking Annotation Strategies in Imaging AI", + "bbox": [ + 290, + 114, + 730, + 130 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 127 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3e05d47b2ce47ec50a7b906ae56fde5c4e572da9e5397f0c35f7d6ecccd4bc0c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 142, + 759, + 271 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/ad0d0af2e78cbc279915f970a0c7b569e8e1266abd92bfa0c2ffd7b330945e90.jpg", + "image_caption": [ + "(a) Annotation companies generate higher DSC scores than MTurk. $0 \\leq \\mathrm{DSC} \\leq 1$ ." + ], + "image_footnote": [], + "bbox": [ + 243, + 284, + 759, + 412 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/eb8a72bf10a3bf83277f1a31f88a856424bec4c703a9f4cc685553377f49cb11.jpg", + "image_caption": [ + "(b) Annotation companies generate higher NSD scores than MTurk. $0 \\leq \\mathrm{NSD} \\leq 1$ .", + "(c) Annotation companies generate a lower number of severe errors.", + "Fig. 7: Annotation companies generate higher quality annotations with their annotation pipeline than MTurk. This is observed for (a) the DSC, (b) the NSD, and (c) number of severe errors (lower is better). Annotation companies include a QA step, in contrast to crowdsourcing on MTurk. (a) and (b) displayed as dots and box plots (band indicates the median, box indicates the first (25th percentile) and third (75th percentile) quartiles)." + ], + "image_footnote": [], + "bbox": [ + 243, + 426, + 759, + 563 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 710, + 401, + 729 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "A part of this work was funded by M. Mengler and S. Funke - during their time at understand.ai -, and Helmholtz Imaging. T. Rädsch was supported by a scholarship from the Hanns Seidel Foundation with funds from the Federal Ministry of Education and Research Germany (BMBF). We thank M. Gelz, S. Strzysch and N. Kraft for their continuous support.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "T. Rädsch et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ali, S., Zhou, F., Braden, B., Bailey, A., Yang, S., Cheng, G., Zhang, P., Li, X., Kayser, M., Soberanis-Mukul, R.D., Albarqouni, S., Wang, X., Wang, C., Watanabe, S., Oksuz, I., Ning, Q., Yang, S., Khan, M.A., Gao, X.W., Realdon, S., Loshchenov, M., Schnabel, J.A., East, J.E., Wagnieres, G., Loschenov, V.B., Grisan, E., Daul, C., Blondel, W., Rittscher, J.: An objective comparison of detection and segmentation algorithms for artefacts in clinical endoscopy. Scientific Reports 10(1), 2748 (Feb 2020). https://doi.org/10.1038/s41598-020-59413-5, https://www.nature.com/articles/s41598-020-59413-5, number: 1 Publisher: Nature Publishing Group 5", + "2. Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M.S., Bohg, J., Bosselut, A., Brunskill, E., Brynjolfsson, E., Buch, S., Card, D., Castellon, R., Chatterji, N., Chen, A., Creel, K., Davis, J.Q., Demszky, D., Donahue, C., Doumbouya, M., Durmus, E., Ermon, S., Etchemendy, J., Ethayarajh, K., Fei-Fei, L., Finn, C., Gale, T., Gillespie, L., Goel, K., Goodman, N., Grossman, S., Guha, N., Hashimoto, T., Henderson, P., Hewitt, J., Ho, D.E., Hong, J., Hsu, K., Huang, J., Icard, T., Jain, S., Jurafsky, D., Kalluri, P., Karamcheti, S., Keeling, G., Khani, F., Khattab, O., Koh, P.W., Krass, M., Krishna, R., Kuditipudi, R., Kumar, A., Ladhak, F., Lee, M., Lee, T., Leskovec, J., Levent, I., Li, X.L., Li, X., Ma, T., Malik, A., Manning, C.D., Mirchandani, S., Mitchell, E., Munyikwa, Z., Nair, S., Narayan, A., Narayanan, D., Newman, B., Nie, A., Niebles, J.C., Nilforoshan, H., Nyarko, J., Ogut, G., Orr, L., Papadimitriou, I., Park, J.S., Piech, C., Portelance, E., Potts, C., Raghunathan, A., Reich, R., Ren, H., Rong, F., Roohani, Y., Ruiz, C., Ryan, J., Re, C., Sadigh, D., Sagawa, S., Santhanam, K., Shih, A., Srinivasan, K., Tamkin, A., Taori, R., Thomas, A.W., Tramer, F., Wang, R.E., Wang, W., Wu, B., Wu, J., Wu, Y., Xie, S.M., Yasunaga, M., You, J., Zaharia, M., Zhang, M., Zhang, T., Zhang, X., Zhang, Y., Zheng, L., Zhou, K., Liang, P.: On the Opportunities and Risks of Foundation Models (Jul 2022). https://doi.org/10.48550/arXiv.2108.07258. http://arxiv.org/abs/2108.07258 13", + "3. Bragg, J., Mausam, Weld, D.S.: Sprout: Crowd-Powered Task Design for Crowdsourcing. Proceedings of the ACM on User Interface Software and Technology (2018). https://doi.org/10.1145/3242587.3242598 4, 6", + "4. Chang, J.C., Amershi, S., Kamar, E.: Revolt: Collaborative Crowdsourcing for Labeling Machine Learning Datasets. In: Proceedings of the 2017 CHI Conference on Human Factors in Computing Systems. pp. 2334-2346. CHI '17, Association for Computing Machinery, New York, NY, USA (May 2017). https://doi.org/10.1145/3025453.3026044 4", + "5. Chen, L., Li, S., Yan, J., Wang, H., Gunaratna, K., Yadav, V., Tang, Z., Srinivasan, V., Zhou, T., Huang, H., Jin, H.: AlpaGasus: Training A Better Alpaca with Fewer Data (Nov 2023). https://doi.org/10.48550/arXiv.2307.08701, http://arxiv.org/abs/2307.08701 12", + "6. Cheplygina, V., Perez-Rovira, A., Kuo, W., Tiddens, H.A.W.M., de Bruijne, M.: Early Experiences with Crowdsourcing Airway Annotations in Chest CT. In: Carneiro, G., Mateus, D., Peter, L., Bradley, A., Tavares, J.M.R.S., Belagiannis, V., Papa, J.P., Nascimento, J.C., Loog, M., Lu, Z., Cardoso, J.S., Cornebise, J. (eds.) Deep Learning and Data Labeling for Medical Applications. pp. 209-218. Lecture Notes in Computer Science, Springer International Publishing, Cham (2016). https://doi.org/10.1007/978-3-319-46976-8_22_4_6" + ], + "bbox": [ + 225, + 174, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Quality Assured: Rethinking Annotation Strategies in Imaging AI", + "bbox": [ + 290, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "7. Chmielewski, M., Kucker, S.C.: An MTurk Crisis? Shifts in Data Quality and the Impact on Study Results. Social Psychological and Personality Science 11(4), 464-473 (May 2020). https://doi.org/10.1177/1948550619875149, publisher: SAGE Publications Inc 4", + "8. Cognilytica: Data engineering, preparation, and labeling for ai 2020 (2020), https://www.cognilytica.com/2020/01/31/data-preparation-labeling-for-ai-2020/10", + "9. Créquet, P., Mansouri, G., Benchoufi, M., Vivot, A., Ravaud, P.: Mapping of Crowdsourcing in Health: Systematic Review. Journal of Medical Internet Research 20(5), 187 (2018). https://doi.org/10.2196/jmir.9330, https://www.jmir.org/2018/5/e187/1", + "10. Daniel, F., Kucherbaev, P., Cappiello, C., Benatallah, B., Allahbakhsh, M.: Quality Control in Crowdsourcing: A Survey of Quality Attributes, Assessment Techniques and Assurance Actions. ACM Computing Surveys 51, 1-40 (Jan 2019). https://doi.org/10.1145/3148148 2, 4", + "11. Dice, L.R.: Measures of the amount of ecologic association between species. Ecology 26(3), 297-302 (1945) 7", + "12. Dow, S., Kulkarni, A., Klemmer, S., Hartmann, B.: Shepherding the crowd yields better work. In: Proceedings of the ACM 2012 conference on Computer Supported Cooperative Work. pp. 1013-1022. CSCW '12, Association for Computing Machinery, New York, NY, USA (Feb 2012). https://doi.org/10.1145/2145204.2145355 4", + "13. Duggan, N.M., Jin, M., Mendicuti, M.A.D., Hallisey, S., Bernier, D., Selame, L.A., Asgari-Targhi, A., Fischetti, C.E., Lucassen, R., Samir, A.E., Duhaime+, E., Kapur, T., Goldsmith, A.J.: Gamified Crowdsourcing as a Novel Approach to Lung Ultrasound Dataset Labeling (Jun 2023). https://doi.org/10.48550/arXiv.2306.06773, http://arxiv.org/abs/2306.06773 4", + "14. Gaikwad, S., Chhibber, N., Sehgal, V., Ballav, A., Mullings, C., Nasser, A., Richmond-Fuller, A., Gilbee, A., Gamage, D., Whiting, M., et al.: Prototype tasks: improving crowdsourcing results through rapid, iterative task design (2017). https://doi.org/10.48550/arXiv.1707.05645, https://arxiv.org/abs/1707.05645 4", + "15. Halevy, A., Norvig, P., Pereira, F.: The unreasonable effectiveness of data. IEEE intelligent systems 24(2), 8-12 (2009). https://doi.org/10.1109/MIS.2009.36, ISBN: 1541-1672 Publisher: IEEE 1", + "16. Han, D., Choe, J., Chun, S., Chung, J.J.Y., Chang, M., Yun, S., Song, J.Y., Oh, S.J.: Neglected free lunch-learning image classifiers using annotation byproducts (2023). https://doi.org/10.48550/arXiv.2303.17595, https://arxiv.org/abs/2303.17595 4", + "17. Heim, E., Roß, T., Teil, A., März, K., Stieltjes, B., Eisenmann, M., Lebert, J., Metzger, J., Sommer, G., Sauter, A.W.: Large-scale medical image annotation with crowd-powered algorithms. Journal of Medical Imaging 5(3), 034002 (2018). https://doi.org/10.1117/1.JMI.5.3.034002 2, 4, 6", + "18. Hettiachchi, D., Schaekermann, M., McKinney, T.J., Lease, M.: The Challenge of Variable Effort Crowdsourcing and How Visible Gold Can Help. Proceedings of the ACM on Human-Computer Interaction 5, 332:1-332:26 (Oct 2021). https://doi.org/10.1145/3476073 5", + "19. Hutson, M., Kanzheleva, O., Taggart, C., Campana, B.J.L., Duong, Q.: Quality Control Challenges in Crowdsourcing Medical Labeling (2019) 4" + ], + "bbox": [ + 217, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "T. Rädsch et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "20. Idrissi, B.Y., Bouchacourt, D., Balestriero, R., Evtimov, I., Hazirbas, C., Ballas, N., Vincent, P., Drozdjal, M., Lopez-Paz, D., Ibrahim, M.: ImageNet-X: Understanding Model Mistakes with Factor of Variation Annotations (Nov 2022). https://doi.org/10.48550/arXiv.2211.01866, http://arxiv.org/abs/2211.01866 3, 4", + "21. Ipeirotis, P.G., Provost, F., Wang, J.: Quality management on Amazon Mechanical Turk. In: Proceedings of the ACM SIGKDD Workshop on Human Computation. pp. 64-67. ACM, Washington DC (Jul 2010). https://doi.org/10.1145/1837885.1837906, https://dl.acm.org/doi/10.1145/1837885.1837906 2, 4", + "22. Irani, L.: The cultural work of microwork. New Media & Society 17(5), 720-739 (May 2015). https://doi.org/10.1177/1461444813511926, publisher: SAGE Publications 12", + "23. Irshad, H., Montaser-Kouhsari, L., Waltz, G., Bucur, O., Nowak, J., Dong, F., Knoblauch, N., Beck, A.: Crowdsourcing image annotation for nucleus detection and segmentation in computational pathology: evaluating experts, automated methods, and the crowd. In: Biocomputing 2015. pp. 294-305. WORLD SCIENTIFIC, Kohala Coast, Hawaii, USA (Nov 2014). https://doi.org/10.1142/9789814644730_0029, http://www.worldscientific.com/doi/abs/10.1142/9789814644730_0029 4", + "24. Joskowicz, L., Cohen, D., Caplan, N., Sosna, J.: Inter-observer variability of manual contour delineation of structures in CT. European Radiology 29(3), 1391-1399 (Mar 2019). https://doi.org/10.1007/s00330-018-5695-5, http://link.springer.com/10.1007/s00330-018-5695-5 3, 5", + "25. Kentley, J., Weber, J., Liopyris, K., Braun, R.P., Marghoob, A.A., Quigley, E.A., Nelson, K., Prentice, K., Duhaime, E., Halpern, A.C., Rotemberg, V.: Agreement Between Experts and an Untrained Crowd for Identifying Dermoscopic Features Using a Gamified App: Reader Feasibility Study. JMIR Medical Informatics 11, e38412 (Jan 2023). https://doi.org/10.2196/38412, https://medinform.jmir.org/2023/1/e38412 4", + "26. Kohli, M.D., Summers, R.M., Geis, J.R.: Medical image data and datasets in the era of machine learning—whitepaper from the 2016 C-MIMI meeting dataset session. Journal of digital imaging 30(4), 392–399 (2017). https://doi.org/10.1007/s10278-017-9976-3, publisher: Springer 5", + "27. Le, J., Edmonds, A., Hester, V., Biewald, L.: Ensuring quality in crowdsourced search relevance evaluation: The effects of training question distribution. In: SIGIR 2010 workshop on crowdsourcing for search evaluation. vol. 2126, pp. 22-32 (2010) 2, 4", + "28. Le Ludec, C., Cornet, M., Casilli, A.A.: The problem with annotation. Human labour and outsourcing between France and Madagascar. Big Data & Society 10(2), 20539517231188723 (Jul 2023). https://doi.org/10.1177/20539517231188723, http://journals.sagepub.com/doi/10.1177/20539517231188723 2", + "29. Lee, A.N., Hunter, C.J., Ruiz, N.: Platypus: Quick, Cheap, and Powerful Refinement of LLMs (Aug 2023). https://doi.org/10.48550/arXiv.2308.07317, http://arxiv.org/abs/2308.07317 12", + "30. Litman, L., Robinson, J., Rosenzweig, C.: The relationship between motivation, monetary compensation, and data quality among US- and India-based workers on Mechanical Turk. Behavior Research Methods 47(2), 519-528 (Jun 2015). https://doi.org/10.3758/s13428-014-0483-x, http://link.springer.com/10.3758/s13428-014-0483-x 6" + ], + "bbox": [ + 215, + 146, + 787, + 838 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Quality Assured: Rethinking Annotation Strategies in Imaging AI", + "bbox": [ + 290, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Lu, X., Ratcliffe, D., Kao, T.T., Tikhonov, A., Litchfield, L., Rodger, C., Wang, K.: Rethinking Quality Assurance for Crowdsourced Multi-ROI Image Segmentation. Proceedings of the AAAI Conference on Human Computation and Crowdsourcing 11(1), 103–114 (Nov 2023). https://doi.org/10.1609/hcomp.v11i1.27552, https://ojs(aaai.org/index.php/HCOMP/article/view/27552, number: 1 2, 4, 5", + "32. Maier-Hein, L., Mersmann, S., Kondermann, D., Bodenstedt, S., Sanchez, A., Stock, C., Kenngott, H.G., Eisenmann, M., Speidel, S.: Can masses of non-experts train highly accurate image classifiers? In: International conference on medical image computing and computer-assisted intervention. pp. 438-445. Springer (2014). https://doi.org/10.1007/978-3-319-10470-6_55 1", + "33. Maier-Hein, L., Reinke, A., Godau, P., Tizabi, M.D., Buettner, F., Christodoulou, E., Glocker, B., Isensee, F., Kleesiek, J., Kozubek, M., et al.: Metrics reloaded: recommendations for image analysis validation. Nature Methods 21, 195–212 (2024). https://doi.org/10.1038/s41592-023-02151-z, https://www.nature.com/articles/s41592-023-02151-z 1", + "34. Maier-Hein, L., Wagner, M., Ross, T., Reinke, A., Bodenstedt, S., Full, P.M., Hempe, H., Mindroc-Filimon, D., Scholz, P., Tran, T.N., Bruno, P., Kisilenko, A., Müller, B., Davitashvili, T., Capek, M., Tizabi, M.D., Eisenmann, M., Adler, T.J., Grohl, J., Schellenberg, M., Seidlitz, S., Lai, T.Y.E., Pekdemir, B., Roethlingshoefer, V., Both, F., Bittel, S., Mengler, M., Mündermann, L., Apitz, M., Kopp-Schneider, A., Speidel, S., Nickel, F., Probst, P., Kenngott, H.G., Müller-Stich, B.P.: Heidelberg colorectal data set for surgical data science in the sensor operating room. Scientific Data 8(1), 101 (Apr 2021). https://doi.org/10.1038/s41597-021-00882-2, https://www.nature.com/articles/s41597-021-00882-2, number: 1 Publisher: Nature Publishing Group 5", + "35. McCulloch, C.E., Searle, S.R.: Generalized, linear, and mixed models. John Wiley & Sons (2004). https://doi.org/10.1002/0471722073, https://books.google.com/books?hl=en&lr=&id=bWDPukohugQC&oi=fnd&pg=PR6&dq=Generalized, +Linear, +and+Mixed+Models.&ots=sSy3HM7CGo&sig=Ss04EHKG_xkcZZCYWSEJl6paHos 8", + "36. Miceli, M., Posada, J.: The Data-Production Dispositif. Proceedings of the ACM on Human-Computer Interaction 6(CSCW2), 1-37 (Nov 2022). https://doi.org/10.1145/3555561, https://dl.acm.org/doi/10.1145/3555561 2, 12, 14", + "37. Michaelis, C., Mitzkus, B., Geirhos, R., Rusak, E., Bringmann, O., Ecker, A.S., Bethge, M., Brendel, W.: Benchmarking Robustness in Object Detection: Autonomous Driving when Winter is Coming (Mar 2020). https://doi.org/10.48550/arXiv.1907.07484, http://arxiv.org/abs/1907.07484 5", + "38. Petrovic, N., Moya-Alcover, G., Varona, J., Jaume-i Capó, A.: Crowdsourcing human-based computation for medical image analysis: A systematic literature review. Health Informatics Journal 26(4), 2446-2469 (Dec 2020). https://doi.org/10.1177/1460458220907435, http://journals.sagepub.com/doi/10.1177/1460458220907435 4", + "39. Reinke, A., Tizabi, M.D., Baumgartner, M., Eisenmann, M., Heckmann-Nötzel, D., Kavur, A.E., Rädsch, T., Sudre, C.H., Acion, L., Antonelli, M.: Understanding metric-related pitfalls in image analysis validation. Nature methods pp. 1-13 (2024). https://doi.org/doi.org/10.1038/s41592-023-02150-0, https://www.nature.com/articles/s41592-023-02150-0, publisher: Nature Publishing Group 7" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "T. Rädsch et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "40. Research, G.V.: Data Collection And Labeling Market Size & Share Report 2030 (2022), https://www.grandviewresearch.com/industry-analysis/data-collection-labeling-market-3,11", + "41. Rofs, T., Bruno, P., Reinke, A., Wiesenfarth, M., Koeppel, L., Full, P.M., Pekdemir, B., Godau, P., Trofimova, D., Isensee, F.: How can we learn (more) from challenges? A statistical approach to driving future algorithm development. arXiv preprint arXiv:2106.09302 (2021). https://doi.org/10.48550/arXiv.2106.09302, https://arxiv.org/abs/2106.09302 3, 4", + "42. Rädsch, T., Reinke, A., Weru, V., Tizabi, M.D., Schreck, N., Kavur, A.E., Pekdemir, B., Roß, T., Kopp-Schneider, A., Maier-Hein, L.: Labelling instructions matter in biomedical image analysis. Nature Machine Intelligence 5(3), 273–283 (Mar 2023). https://doi.org/10.1038/s42256-023-00625-5, https://www.nature.com/articles/s42256-023-00625-5, number: 3 Publisher: Nature Publishing Group 2, 4", + "43. Sambasivan, N., Kapania, S., Highfill, H., Akrong, D., Paritosh, P., Aroyo, L.M.: \"Everyone wants to do the model work, not the data work\": Data Cascades in High-Stakes AI. In: Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. pp. 1-15. ACM, Yokohama Japan (May 2021). https://doi.org/10.1145/3411764.3445518, https://dl.acm.org/doi/10.1145/3411764.3445518 1, 13", + "44. Schmidt, F.A.: The Planetary Stacking Order of Multilayered Crowd-AI Systems. In: Graham, M., Ferrari, F. (eds.) Digital Work in the Planetary Market, pp. 137-156. The MIT Press (Jun 2022). https://doi.org/10.7551/mitpress/13835.003.0012, https://direct.mit.edu/books/book/5319/chapter/3800159/The-Planetary-Stacking-Order-of-Multilayered-Crowd 2", + "45. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: LLaMA: Open and Efficient Foundation Language Models (Feb 2023). https://doi.org/10.48550/arXiv.2302.13971, http://arxiv.org/abs/2302.13971 12", + "46. Veselovsky, V., Ribeiro, M.H., West, R.: Artificial Artificial Artificial Intelligence: Crowd Workers Widely Use Large Language Models for Text Production Tasks (Jun 2023). https://doi.org/10.48550/arXiv.2306.07899, http://arxiv.org/abs/2306.07899 13", + "47. Wang, K., Hill, M., Knowles-Barley, S., Tikhonov, A., Litchfield, L., Bare, J.C.: Improving Segmentation of Breast Arterial Calcifications from Digital Mammography: Good Annotation is All You Need. In: Zheng, Y., Keles, H.Y., Koniusz, P. (eds.) Computer Vision - ACCV 2022 Workshops, vol. 13848, pp. 134-150. Springer Nature Switzerland, Cham (2023). https://doi.org/10.1007/978-3-031-27066-6_10, https://link.springer.com/10.1007/978-3-031-27066-6_10, series Title: Lecture Notes in Computer Science 2, 4", + "48. Yu, S., Zhou, H.Y., Ma, K., Bian, C., Chu, C., Liu, H., Zheng, Y.: Difficulty-Aware Glaucoma Classification with Multi-rater Consensus Modeling. In: Martel, A.L., Abolmaesumi, P., Stoyanov, D., Mateus, D., Zuluaga, M.A., Zhou, S.K., Racoceanu, D., Joskowicz, L. (eds.) Medical Image Computing and Computer Assisted Intervention - MICCAI 2020. pp. 741-750. Lecture Notes in Computer Science, Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-59710-8_72_4", + "49. Ørting, S.N., Doyle, A., Hilten, A.v., Hirth, M., Inel, O., Madan, C.R., Mavridis, P., Spiers, H., Cheplygina, V.: A Survey of Crowdsourcing in Medical Image Analysis." + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Quality Assured: Rethinking Annotation Strategies in Imaging AI", + "bbox": [ + 290, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Human Computation 7, 1-26 (Dec 2020). https://doi.org/10.15346/hc.v7i1.1, https://hcjournal.org/index.php/jhc/article/view/111 1", + "bbox": [ + 241, + 147, + 785, + 176 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "T. Rädsch et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/2024/Quality Assured_ Rethinking Annotation Strategies in Imaging AI/1dd3eeac-8808-45a9-8928-d1c207e0686a_model.json b/2024/Quality Assured_ Rethinking Annotation Strategies in Imaging AI/1dd3eeac-8808-45a9-8928-d1c207e0686a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..909b42015f0e306251fddc197b50521c60b989dd --- /dev/null +++ b/2024/Quality Assured_ Rethinking Annotation Strategies in Imaging AI/1dd3eeac-8808-45a9-8928-d1c207e0686a_model.json @@ -0,0 +1,2286 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.271, + 0.141, + 0.737, + 0.187 + ], + "angle": 0, + "content": "Quality Assured: Rethinking Annotation Strategies in Imaging AI" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.219, + 0.782, + 0.265 + ], + "angle": 0, + "content": "Tim Rädsch1,2,3, Annika Reinke1,2, Vivienne Weru1, Minu D. Tizabi1, Nicholas Heller4, Fabian Isensee1,2, Annette Kopp-Schneider1*, and Lena Maier-Hein1,2,3,5*" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.277, + 0.77, + 0.321 + ], + "angle": 0, + "content": "\\(^{1}\\)German Cancer Research Center, GER \\(^{2}\\)Helmholtz Imaging, GER \\(^{3}\\)Heidelberg University, GER \\(^{4}\\)University of Minnesota, USA \\(^{5}\\)National Center for Tumor Diseases, GER \\(\\boxed{\\boxtimes}\\):{tim.raedsch; l.maier-hein}@dkfz-heidelberg.de" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.345, + 0.744, + 0.624 + ], + "angle": 0, + "content": "Abstract. This paper does not describe a novel method. Instead, it studies an essential foundation for reliable benchmarking and ultimately real-world application of AI-based image analysis: generating high-quality reference annotations. Previous research has focused on crowdsourcing as a means of outsourcing annotations. However, little attention has so far been given to annotation companies, specifically regarding their internal quality assurance (QA) processes. Therefore, our aim is to evaluate the influence of QA employed by annotation companies on annotation quality and devise methodologies for maximizing data annotation efficacy. Based on a total of 57,648 instance segmented images obtained from a total of 924 annotators and 34 QA workers from four annotation companies and Amazon Mechanical Turk (MTurk), we derived the following insights: (1) Annotation companies perform better both in terms of quantity and quality compared to the widely used platform MTurk. (2) Annotation companies' internal QA only provides marginal improvements, if any. However, improving labeling instructions instead of investing in QA can substantially boost annotation performance. (3) The benefit of internal QA depends on specific image characteristics. Our work could enable researchers to derive substantially more value from a fixed annotation budget and change the way annotation companies conduct internal QA." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.635, + 0.717, + 0.65 + ], + "angle": 0, + "content": "Keywords: Data Annotation \\(\\cdot\\) Quality Assurance \\(\\cdot\\) Crowdsourcing" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.67, + 0.377, + 0.687 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.698, + 0.788, + 0.821 + ], + "angle": 0, + "content": "Following a long period of extensive focus on pure method development, proper benchmarking and validation have recently been recognized as a major challenge in AI-based image analysis [33, 43]. In this context, high-quality annotation of reference datasets is a fundamental requirement to ensure the meaningful validation of image analysis algorithms [15]. As annotation tasks may be cumbersome, crowdsourcing is often used to partially outsource image annotation. Traditionally, such outsourcing is conducted via crowdsourcing platforms [32], with Amazon Mechanical Turk (MTurk)1 as the predominant choice [9, 49]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.538, + 0.84 + ], + "angle": 0, + "content": "1 https://www.mturk.com * shared last author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "T. Rädsch et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.143, + 0.788, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.298, + 0.789, + 0.385 + ], + "angle": 0, + "content": "Fig. 1: Research Questions (RQs) tackled in this work. Based on 57,648 instance segmentation masks annotated by 924 annotators and 34 quality assurance (QA) workers from five different annotation providers, we (1) compared the effectiveness of generating high-quality annotations between annotation companies and Amazon Mechanical Turk and (2) investigated the effects of annotation companies' internal QA and (3) real-world image characteristics on the annotation quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.426, + 0.789, + 0.653 + ], + "angle": 0, + "content": "Recent literature has reflected the rise of annotation companies as a new type of provider with potentially better annotation performance compared to established crowdsourcing platforms such as MTurk [28,36,42,44]. Annotation companies differ substantially from crowdsourcing platforms. First, they employ annotators directly and at shared office spaces [28,44], enabling better external control of worker conditions and stricter data security measures, while crowdsourcing platforms can be accessed by anyone, without specific knowledge or training being required. Second, annotators in annotation companies typically work in small teams, with an experienced quality assurance (QA) worker being responsible for the team's output. In the observed internal QA, QA workers make adjustments to the annotations generated by the team by reviewing each individual annotation and modifying those where they deem changes necessary. As a result, the annotation requester receives the quality-assured annotation, not the initial one (see Fig. 2a). In contrast, crowdsourcing platforms usually do not employ internal QA mechanics similar to annotation companies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.841 + ], + "angle": 0, + "content": "The importance of annotation providers for real-world applications is evident at the major computer vision conferences, where over the past two years, approximately \\(20\\%\\) of exhibitors were annotation providers (see Fig. 2b). While QA as a post-processing step for annotations from platforms such as MTurk has been researched (e.g. [10, 17, 21, 27]), little is known about the benefit of internal QA on annotation quality within annotation companies. The only related work that we are aware of pertains to external QA from the perspective of an annotation company. Unlike in internal QA, in the case of external QA, domain experts or experienced annotators working on behalf of the requester check the annotation quality as a post-processing step of the received annotations [31, 47]. Our paper, on the other hand, focuses on the internal QA processes of an annotation company." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.292, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Quality Assured: Rethinking Annotation Strategies in Imaging AI" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.32, + 0.788, + 0.404 + ], + "angle": 0, + "content": "Fig. 2: a) In contrast to Amazon Mechanical Turk (MTurk), annotation companies commonly perform quality assurance (QA) before delivering the result to the annotation requester. b) Annotation providers are an integral part of the computer vision community. They account for around \\(20\\%\\) of exhibitors present at major computer vision conferences (ECCV 2022: \\(\\mathrm{n} = 40\\), CVPR 2022: \\(\\mathrm{n} = 97\\), ICCV 2023: \\(\\mathrm{n} = 43\\), CVPR 2023: \\(\\mathrm{n} = 104\\)) in the last two years." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.515 + ], + "angle": 0, + "content": "In light of the vast sums of money to be spent on QA [40], the purpose of this paper is to shed light on whether annotation companies' internal QA benefits the annotation quality. Based on more than 50,000 instance segmentation masks annotated by more than 900 annotators from five different annotation providers, we investigate the following research questions (RQs; see Fig. 1):" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.788, + 0.653 + ], + "angle": 0, + "content": "RQ 1) How does the choice of annotation provider influence the quality of annotations? As the selection of the annotation provider is among the first and key decisions in any annotation project after data collection and generation of labeling instructions, we first investigated companies' effectiveness in generating annotations, which is particularly relevant in cases of constrained annotation budgets requiring a cost-effective strategy. As both quantity and quality of annotated images for a given budget are relevant in provider selection, we compared the annotation quality and costs of five different providers, including MTurk and four annotation companies, on identical images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.657, + 0.788, + 0.762 + ], + "angle": 0, + "content": "RQ 2) Does annotation companies' quality assurance improve annotation quality? A major decision in the design of an annotation process concerns the allocation of resources. A common perception is that internal QA of annotation companies, following the initial annotation, substantially enhances the quality. Here, we challenged this assumption by investigating the improvement in annotation quality from internal QA compared to that achieved through enhanced labeling instructions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "RQ 3) Should quality assurance focus on specific subsets of images? In theory, the efficacy of QA may be contingent upon specific image characteristics [20, 24, 41]. For instance, underexposed images might benefit more from a secondary review compared to those with optimal illumination. Concurrently, images that present annotation challenges for human observers might be partic" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "T. Rädsch et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "ularly pertinent for the training and validation of algorithms [20, 41]. Here, we investigated whether and to what extent the benefit of QA depends on specific image characteristics." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.211, + 0.39, + 0.227 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.238, + 0.56, + 0.254 + ], + "angle": 0, + "content": "2.1 Comparison of annotation providers" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.41 + ], + "angle": 0, + "content": "With respect to the quality of annotations obtained from varying annotation providers, the closest work conducted are comparisons between crowdsourcing platforms and domain experts. Irshad et al. [23] and Petrovic et al. [38], among others, have shown that crowdsourced annotations can reach the annotation quality of domain experts in safety-critical applications. In recent literature, Duggan et al. [13] and Kentley et al. [25] compared domain-specialized crowds as a new type of annotation provider and found comparable performance to domain experts. Furthermore, Rädsch et al. [42] provided a comparison of annotation companies to crowdsourcing, but focused solely on raw annotation quality without QA and without relating it to annotation budgets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.788, + 0.502 + ], + "angle": 0, + "content": "Interestingly, researchers report concerns about the annotation quality obtained from MTurk [31], such as obtaining unusable annotations in up to \\(67.8\\%\\) of cases [6]. Other research fields, such as psychology, even report a \"substantial decrease in data quality\" [7]. However, we are not aware of work that has systematically compared the performance between annotation companies and crowdsourcing platforms such as MTurk while considering annotation budgets." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.518, + 0.411, + 0.534 + ], + "angle": 0, + "content": "2.2 Quality assurance" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.735 + ], + "angle": 0, + "content": "The main goal of QA is to enhance and ensure the quality of annotations. A comprehensive overview on QA in crowdsourcing platforms is provided by Daniel et al. [10]. However, their research does not include annotation companies. During the annotation process, quality control techniques generally fall into three categories [31]: (1) comparing to a gold standard with pre-defined reference annotations (e.g., [27]), (2) analyzing combined outputs from multiple annotators (e.g., [21]), and (3) investigating annotator behaviors (e.g., [17, 19]). Other approaches to enhancing annotation quality on crowdsourcing platforms target training of crowdworkers [12] and iterative refinement of the annotation environment [3,4,14]. It should be noted that annotation byproducts, such as mouse movement during annotation generation [16, 17] or inter-rater agreement [48], can be used to boost model performance, by incorporating this information in the training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The closest related work focusing on the impact of QA in annotation companies was introduced by Wang et al. [47] and Lu et al. [31]. However, their work focuses on external QA conducted by experienced personnel from the requester rather than internal QA, and features a substantially smaller number of annotators from annotation companies. To our knowledge, a systematic investigation into the effects of internal QA processes within annotation companies is thus, so far, lacking." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.292, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Quality Assured: Rethinking Annotation Strategies in Imaging AI" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.787, + 0.163 + ], + "angle": 0, + "content": "2.3 Impact of real-world image characteristics on annotation quality" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.788, + 0.31 + ], + "angle": 0, + "content": "Related research analyzing the impact of real-world image characteristics highlights their effect on annotations created by machine learning models and annotators alike. Michaelis et al. [37] demonstrated that models easily encounter failures with increasingly challenging real-world image characteristics, such as added snow or poor lighting characteristics. Similarly, annotators generate less reliable annotations in the presence of challenging real-world image characteristics. This effect can be observed both within [1,24,31] and across domains [18,26]. However, a systematic analysis of the impact of such characteristics on internal QA employed by annotation companies is lacking." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.333, + 0.339, + 0.348 + ], + "angle": 0, + "content": "3 Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.365, + 0.787, + 0.41 + ], + "angle": 0, + "content": "In the following, we describe the selected data, the labeling instructions, annotation providers and annotators, the QA process, and the experimental design of our work." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.434, + 0.304, + 0.448 + ], + "angle": 0, + "content": "3.1 Data" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.596 + ], + "angle": 0, + "content": "We conducted our experiments on the instance segmentation data from the Heidelberg Colorectal (HeiCo) dataset for surgical data science in the sensor operating room [34], representing a safety-critical application that crowdsourcing platforms can still handle. The task involved segmenting surgical instruments from laparoscopic videos, resulting in 57,648 instance segmentation masks on 4,050 unique frames. Furthermore, we used 57,636 metadata annotations on the unique images to quantify the effect of varying real-world image characteristics (see RQ3). Characteristics included motion blur, underexposure, object occlusion, or overlapping objects, among others, and are illustrated in Suppl. A." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.618, + 0.436, + 0.634 + ], + "angle": 0, + "content": "3.2 Labeling instructions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.75 + ], + "angle": 0, + "content": "To investigate the benefit of QA compared to an investment in detailed labeling instructions, we created three distinct types of labeling instructions with varying levels of information density, namely, (1) minimal text, (2) extended text, and (3) extended text including exemplary pictures. Each subsequent labeling instruction type was structured to contain more detailed information than the preceding. To ensure accurate labeling of the imaging data, we established a collection of design principles applicable to all labeling instructions:" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.75, + 0.787, + 0.779 + ], + "angle": 0, + "content": "(1) Labeling information is presented in a slide format to enhance the processing of information by humans." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.787, + 0.825 + ], + "angle": 0, + "content": "(2) Each slide contains a distinct 'chunk' of information, which is a self-contained piece of content. This approach of 'chunking' information lessens the cognitive load on the working memory." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.825, + 0.767, + 0.84 + ], + "angle": 0, + "content": "(3) Related information chunks are placed in close proximity to one another." + }, + { + "type": "list", + "bbox": [ + 0.215, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.128 + ], + "angle": 0, + "content": "T. Rädsch et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "(4) A uniform design template is used throughout, featuring specific fonts, symbols, and color schemes." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.178, + 0.561, + 0.192 + ], + "angle": 0, + "content": "The three labeling instructions differ as follows:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.785, + 0.268 + ], + "angle": 0, + "content": "Minimal text labeling instructions: These instructions provide a concise textual outline, offering examples of typical annotations and highlighting the most frequent annotation types (see Suppl. B.1). They reflect a scenario with minimal effort invested in developing the instructions. The basic framework includes seven slides totaling 168 words." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.787, + 0.342 + ], + "angle": 0, + "content": "Extended text labeling instructions: Building upon the simplified version, the extended text instructions offer a thorough textual explanation, supplemented by both positive and negative examples (refer to Suppl. B.2). They cover both regular and some rare annotation scenarios. The enhanced version comprises ten slides with a total of 446 words." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.45 + ], + "angle": 0, + "content": "Extended text including pictures labeling instructions: These instructions augment the enhanced textual guidelines with images (detailed in Suppl. B.3). The visual aids feature descriptions, symbols, annotations, and color coding to effectively communicate the details on the slides. They also include illustrations of rare annotation cases, indicating a deep and well-documented understanding of the labeling process. This set contains 16 slides with a total of 961 words. We refer the reader to Suppl. B for full details of each labeling instruction type." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.47, + 0.44, + 0.485 + ], + "angle": 0, + "content": "3.3 Annotation providers" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.785, + 0.553 + ], + "angle": 0, + "content": "The study was conducted based on 216 annotators and 34 QA workers from four annotation companies and 708 crowdworkers from the crowdsourcing platform MTurk. Our goal was to provide the best representation for both annotation provider types." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.787, + 0.749 + ], + "angle": 0, + "content": "Consequently, we established higher standards for our MTurk crowdworkers compared to the usual research benchmarks, which generally require a \\(95\\%\\) acceptance rate of Human Intelligence Tasks (HITs) and at least 100 completed tasks [3, 6, 17]. Our criteria included a \\(98\\%\\) acceptance rate with over 5,000 completed HITs. We spread out the HITs over 40 days at all times to obtain a representative MTurk sample and compensated workers fairly, as recommended by Litman et al. [30]. The annotation companies we selected have a solid background in handling large-scale annotation tasks in a variety of domains and operate internationally. The annotators were allocated by the companies themselves to replicate the typical workflow of a project carried out by an annotation company. Each annotator was linked to a specific annotation provider that had been predetermined. No selection process from our side took place for either the annotators, QA workers, or the MTurk crowdworkers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To prevent information leakage, we conducted the labeling process with increasingly detailed labeling instructions (minimal text instruction, extended text instruction, and extended text including pictures instruction). Each annotator worked exclusively with one of the three instruction types as introduced in Sec. 3.2. To enhance security, we included a mandatory minimum gap of ten days between consecutive labeling instructions." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.292, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Quality Assured: Rethinking Annotation Strategies in Imaging AI" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.476, + 0.163 + ], + "angle": 0, + "content": "3.4 Quality assurance process" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.786, + 0.221 + ], + "angle": 0, + "content": "In the internal QA process, annotators typically report to a small number of experienced QA workers, who are often responsible for training their team and, in our case, responsible for the QA of the annotations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.224, + 0.788, + 0.435 + ], + "angle": 0, + "content": "The general workflow for all companies consisted of an annotation stage, in which images were assigned randomly to each annotator, and a QA stage. During the annotation stage, annotators received raw images with the task of annotating them until they considered the quality of annotations satisfactory. Annotators were aware that their work would be checked by QA. Once submitted, the QA worker gained access to the generated annotations, allowing them to make modifications until they were satisfied before submitting the final annotations for delivery to the requester. Throughout this process, both annotators and QA workers retained access to the respective labeling instructions at any given moment. We obtained access to the annotations both before and after QA through a software company handling a substantial number of annotation projects annually. The experiments were seamlessly integrated into the software company's pipeline of annotation projects, reducing the risk of exposing our experiments. Pricing was established to be representative of other projects of similar size." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.788, + 0.496 + ], + "angle": 0, + "content": "To maintain fairness and avoid giving any annotation provider a potential informational advantage that may have affected our statistical analysis, no specific queries from the annotation companies about the content of the labeling instructions were addressed." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.523, + 0.433, + 0.538 + ], + "angle": 0, + "content": "3.5 Experimental design" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.78 + ], + "angle": 0, + "content": "RQ 1: Each of the five labeling providers annotated the same images with each set of labeling instructions, employing separate annotators for each task. We analyzed the number of spam images, images with severe errors, i.e., images with at least one false positive and/or false negative, and the overall annotation quality with the Dice Similarity Coefficient (DSC) [11] as the overlap-based metric and the Normalized Surface Distance (NSD) [39] as the distance-based metric. We define a spam annotation as a purposefully poor-quality annotation with the intent to obtain the payment without properly conducting the annotation task. Annotation duration was used as an auxiliary measure, where accessible. For instance, if an annotator completed the task in only 17 seconds, they could not have read the instructions and properly performed annotations on several images. Of note, a spammer could wait a relevant number of seconds and then enter a poor-quality annotation to bypass the annotation duration check. Thus, each image was additionally visually inspected by an engineer with extensive labeling experience. Spam annotation examples are displayed in Suppl. D." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "RQ 2: The annotation companies conducted their internal QA processes directly after their annotators completed the annotation job, as explained in Sec. 3.4 and Fig. 2a. The experiments began with only providing the minimal text labeling instructions, followed by the extended text labeling instructions," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "T. Rädsch et al." + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.144, + 0.734, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.297, + 0.788, + 0.368 + ], + "angle": 0, + "content": "Fig. 3: Annotation companies are more efficient at generating high-quality images than Amazon Mechanical Turk (MTurk). Percentage of images without severe errors (green), images with severe errors (orange), and spam (red) obtained for the five different annotation providers investigated. Annotation costs per image were lower for all companies compared to MTurk (median: \\(61.1\\%\\))." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.785, + 0.426 + ], + "angle": 0, + "content": "and finally, the extended text instructions with pictures, to prevent information leakage." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.428, + 0.785, + 0.532 + ], + "angle": 0, + "content": "For each annotated image, we calculated the DSC and NSD per instance, aggregated the scores per annotated image, and aggregated the scores over annotation company and labeling instruction. We then compared the annotation quality without and with QA across the three types of labeling instructions for the four annotation companies by computing both absolute and relative changes in DSC/NSD aggregates, the percentage of modified images, and the change in the percentage of severe errors, among other parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.787, + 0.64 + ], + "angle": 0, + "content": "RQ 3: In contrast to MTurk, the four annotation companies conducted an additional annotation round, including QA, with the extended text including pictures labeling instructions to quantify the impact of different real-world image characteristics, represented by 57,636 metadata annotations, on the probability of improvement. The data was analyzed with a logistic mixed model [35] which included random effects for the image, the annotator, and the annotation company. Further method details are provided in Suppl. C." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.662, + 0.327, + 0.677 + ], + "angle": 0, + "content": "4 Results" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.694, + 0.741, + 0.725 + ], + "angle": 0, + "content": "4.1 How does the choice of annotation provider influence the quality and cost of annotations?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Annotated images generated by the annotation companies were of higher quality, comprising no spam (MTurk: \\(\\sim 20\\%\\) ) and a much higher proportion of remaining images without severe annotation errors (factor of \\(>2\\) improvement for all companies compared to MTurk), as shown in Fig. 3. In addition, the annotation companies were \\(61.1\\%\\) cheaper in the median, despite their additional employment of QA workers. More specifically, the costs ranged from \\(24.5\\%\\) to \\(81.5\\%\\) of the MTurk costs for the same images and number of annotations. We have" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.292, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Quality Assured: Rethinking Annotation Strategies in Imaging AI" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.144, + 0.234, + 0.154 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.156, + 0.445, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.236, + 0.301, + 0.443, + 0.322 + ], + "angle": 0, + "content": "Annotation company Company 1 Company 2 Company 3 Company 4" + }, + { + "type": "table_caption", + "bbox": [ + 0.46, + 0.144, + 0.476, + 0.154 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "table", + "bbox": [ + 0.462, + 0.156, + 0.787, + 0.303 + ], + "angle": 0, + "content": "
CompanyLabeling instruction25th PercentileMedian75th Percentile
AnnotationQAAnnotationQAAnnotationQA
11) Min. text0.610.620.930.930.960.96
2) Ext. text0.580.580.930.930.960.96
3) Ext. text + pictures0.900.900.950.950.960.96
21) Min. text0.570.470.920.900.960.95
2) Ext. text0.620.840.930.940.960.96
3) Ext. text + pictures0.850.880.940.940.960.96
31) Min. text0.600.600.930.930.960.96
2) Ext. text0.470.640.910.930.960.96
3) Ext. text + pictures0.740.850.930.940.960.96
41) Min. text0.630.630.940.940.960.96
2) Ext. text0.880.880.940.940.960.96
3) Ext. text + pictures0.910.910.950.950.970.97
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.305, + 0.754, + 0.322 + ], + "angle": 0, + "content": "COLOR SCALE 1.00 0.75 0.50 0.25 0.00 -0.25 -0.50 0.75 -1.00" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.332, + 0.785, + 0.415 + ], + "angle": 0, + "content": "Fig. 4: QA of annotation companies only provides marginal improvements, if any. These depend highly on the company and type of labeling instruction. a: Absolute change of the median DSC resulting from performing QA. For each company and labeling instruction, the median was obtained by aggregating the DSC scores over the images. b: Corresponding absolute values of the DSC. Positive and negative changes from annotation to QA are colored according to the color scale." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.784, + 0.492 + ], + "angle": 0, + "content": "not been given permission to reveal the identities and exact absolute budgets used for the annotation companies, hence we report costs relative to the MTurk budget. Examples of spam annotations are provided in Suppl. D." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.785, + 0.552 + ], + "angle": 0, + "content": "Regardless of the annotation scenario, the annotation companies generated higher quality annotations than MTurk with higher DSC scores including smaller interquartile ranges (IQRs) (see Fig. 7a), higher NSD scores including smaller IQRs (see Fig. 7b), and lower numbers of severe errors (see Fig. 7c)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.574, + 0.72, + 0.604 + ], + "angle": 0, + "content": "4.2 Does annotation companies' quality assurance improve annotation quality?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.785, + 0.749 + ], + "angle": 0, + "content": "As shown in Fig. 4, the internal QA conducted by annotation companies does not substantially improve annotation quality if the QA workers only use minimal text instructions. The impact of QA when having access to extended text instructions or extended text including pictures instructions depended highly on the annotation company, while the effect was still minimal compared to not performing QA: Two of the four annotation companies show no change in the median DSC when comparing QA to non-QA outputs for extended text or extended text and picture instructions. The remaining two companies only show a very small relative improvement of below \\(2\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.841 + ], + "angle": 0, + "content": "We obtained even smaller improvements with the NSD scores, as depicted in Suppl. E. In certain scenarios, QA was able to remove a small proportion of severe errors, however, the vast majority remains (see Suppl. F). Notably, the percentage of modified images by QA workers depended highly on the company, as depicted in Suppl. G. Fig. 5 further illustrates that providing annotators with high-quality labeling instructions including pictures (green) leads them to vastly" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "T. Rädsch et al." + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.144, + 0.468, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.143, + 0.787, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.324, + 0.75, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.492 + ], + "angle": 0, + "content": "Fig. 5: Improving the labeling instructions yields a higher effect compared to adding QA to the annotation process. The baseline performance (minimal labeling instruction, no QA) according to the DSC (a) and the percentage of severe errors (b) is shown individually per company with a red asterisk (minimal labeling instruction, no QA). While adding QA only yields marginal improvements (dark blue), major performance boosts are obtained by improving the labeling instructions provided to the annotators (green). DSC scores were aggregated for each image and per labeling type. DSC displayed as dots and box plot (band indicates the median, box indicates the first (25th percentile) and third (75th percentile) quartiles). \\(0 \\leq \\mathrm{DSC} \\leq 1\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.519, + 0.788, + 0.551 + ], + "angle": 0, + "content": "outperform annotators working with lower-quality labeling instructions, even if these are followed by additional internal QA (dark blue)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.57, + 0.764, + 0.586 + ], + "angle": 0, + "content": "4.3 Should quality assurance focus on specific subsets of images?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.593, + 0.788, + 0.683 + ], + "angle": 0, + "content": "As shown in Fig. 6a, QA significantly increases the odds of improving the annotation quality for images with generalizable image characteristics such as underexposed objects, intersection of objects, objects occluded by background in comparison to the absence of this characteristic (vertical black line). The only exceptions are motion artifacts present in the background and object(s) overexposure." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.684, + 0.788, + 0.759 + ], + "angle": 0, + "content": "In contrast to the general increase of images with generalizable image characteristics, none of the domain-specific image characteristics yield a significant change in the odds of improving the annotation quality over a regular image, see Fig. 6b. The only exception for domain-specific characteristics is object(s) covered by smoke, which could be interpreted as a form of object occlusion." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.78, + 0.355, + 0.796 + ], + "angle": 0, + "content": "5 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "The data annotation market is estimated to reach a volume of more than USD 4.1 billion by the end of 2024 [8]. As the annotation process becomes increasingly" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.292, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Quality Assured: Rethinking Annotation Strategies in Imaging AI" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.146, + 0.422, + 0.158 + ], + "angle": 0, + "content": "(a) GENERALIZABLE CHARACTERISTICS" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.161, + 0.788, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.29, + 0.432, + 0.302 + ], + "angle": 0, + "content": "(b) DOMAIN-SPECIFIC CHARACTERISTICS" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.304, + 0.789, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.476, + 0.789, + 0.617 + ], + "angle": 0, + "content": "Fig. 6: QA of annotation companies can be beneficial to certain images with difficult real-world generalizable image characteristics, such as underexposure or intersection of objects. We display the impact of (a) real-world generalizable and (b) domain-specific image characteristics on the odds of improvement. Improvement is defined as a positive difference in the QA, while no improvement refers to 0 or negative difference. The effect of an image characteristic is shown in the form of the log(odds ratio (OR)) representing the logarithms of the odds of improvement in the presence of the characteristic, compared to the odds of improvement occurring in the absence of that characteristic. The vertical black line refers to the point of equal odds. (Note: ***: p-value ≤ 0.001, **: p-value ≤ 0.01, *: p-value ≤ 0.05)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.655, + 0.788, + 0.776 + ], + "angle": 0, + "content": "automated and foundation models evolve, the market is undergoing drastic shifts. With lower entry barriers for annotation, more efficient data annotation tools, and a rising impact of QA, it is thought to result in an estimated USD 8.2 billion market by 2028 [40]. As a research community, we need to understand how to ensure high annotation quality throughout this evolution, as reliable reference annotations are an indispensable foundation for the benchmarking of image analysis algorithms, and thus their ultimate translation into real-world practice." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To our knowledge, this study is the first to quantitatively and critically examine the impact of internal QA conducted by annotation companies in the context of safety-critical applications. The most important insights can be summarized as follows:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "T. Rädsch et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.42 + ], + "angle": 0, + "content": "1. Annotation companies perform better both in terms of quantity and quality compared to the widely used platform MTurk (Fig. 3). This implies that requesters working with crowdsourcing platforms should seriously consider annotation companies as an alternative. The organizational structure of these companies, which enables a single point of contact for requesters and eliminates spam annotations, appears to be a contributing factor to their efficiency. In light of poor working conditions being a major problem in annotation generation [22, 36], annotation companies, which can more easily be inspected than distributed crowdsourcing platforms, are thought to offer better working contracts, enabling fairer working conditions for annotators. It should be noted that not every annotation company automatically ensures adequate working conditions. The annotation companies employed in this study underwent thorough scrutiny to guarantee compliance with good working condition guidelines. With annotators working at a shared location instead of in a distributed crowd, verifying annotators' working conditions has become substantially easier. We thus view our work as an initial step towards greater accountability on the part of annotation requesters. Further dimensions of annotation provider selection are discussed in Suppl. H." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.424, + 0.788, + 0.59 + ], + "angle": 0, + "content": "2. In contrast to common belief, internal QA mechanisms do not necessarily improve annotation quality (Fig. 4). Furthermore, we observe no systematic QA quality improvement difference between annotation companies depending on their region of origin, size, or price, meaning more expensive annotation companies do not translate to better QA. However, improving labeling instructions instead of investing in QA can substantially boost the annotation performance (Fig. 5). Consequently, in order to obtain high-quality annotations, annotation requesters should prioritize refining their labeling instructions, followed by implementing QA or even a different external QA process as a secondary step. Compared to QA, better labeling instructions can be generated with relatively few resources." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.788, + 0.701 + ], + "angle": 0, + "content": "3. The benefit of QA depends on the specific image characteristics (Fig. 6). In fact, QA did improve the quality of images with certain challenging real-world image characteristics, see Fig. 6a. Consequently, it might be useful for requesters and annotation companies alike to allocate resources to identifying challenging image characteristics during a project, train classifiers to identify images containing these characteristics and focus the internal QA effort on these images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.842 + ], + "angle": 0, + "content": "One could contend that the recent rise of foundation models marginalizes the role of annotation providers. We argue, however, that it only changes their role. In this context, it is necessary to differentiate between the data quality demands for the training and testing of models. With a recent trend towards fine-tuning pre-trained models, smaller high-quality datasets can achieve similar performance with less computation [5, 29], highlighting the importance of the quality of unlabeled data in unsupervised training and its role in reducing carbon emissions [45]. In testing, especially for safety-critical applications such as autonomous driving and clinical medicine, ensuring high-quality annotated test" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.292, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Quality Assured: Rethinking Annotation Strategies in Imaging AI" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.128 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "data is crucial for accurate model evaluation and real-world applicability [43]. Consequently, while training data annotation may lose relevance, test data annotation and its high quality demands are here to stay. Regarding foundation models, it should further be noted that annotations purely produced by foundation models introduce artifacts and biases [2] and represent an increasing rising problem for MTurk, as demonstrated by Veselovsky et al. [46]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.242, + 0.788, + 0.409 + ], + "angle": 0, + "content": "Our study is subject to several limitations which deserve further discussion. Comparing prices across providers is non-trivial as the costs to annotate a given data set depend on many factors including, among others, image quantity, type of annotations, location of annotation company and quality requirements. Our analysis was based on the actual costs incurred during our project. Confidentiality agreements with the software company which integrated our experiments into their annotation project pipeline prevent us from disclosing the identities of the annotation companies. Additionally, we took precautions to protect the identities of individual annotators and QA workers. This was particularly important given our findings that QA workers did not significantly enhance annotation quality, potentially exposing them to negative repercussions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.788, + 0.502 + ], + "angle": 0, + "content": "Another limitation could be seen in the fact that our data set only represents one safety-critical domain. While this is a valid concern, we would like to highlight that we chose a dataset which combined quality, complexity, and volume. Furthermore, regarding the breadth of annotation providers and the number of annotators involved, our study is likely the largest of its kind, thereby incurring substantial costs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.788, + 0.613 + ], + "angle": 0, + "content": "Finally, conducting multiple experiments with identical annotation companies and varying datasets could have potentially revealed our experimental design. Had the annotation companies been aware of our research, outcomes could potentially have been biased to appear more favorable, given companies' inherent interest in demonstrating their services as reliable. We thus chose to conduct our experiments such that they could be seamlessly integrated into the software company's annotation pipeline, reducing their likelihood of exposure." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.654, + 0.533, + 0.67 + ], + "angle": 0, + "content": "6 Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In conclusion, our study underscores the pivotal role of enhancing the initial annotation quality, for instance via allocating resources to the improvement of labeling instructions, over QA in obtaining higher reference annotation quality in high-stakes AI [43]. Future research should delve into the development and implementation of comprehensive training programs for annotators from annotation companies to further enhance initial annotation quality. Exploring the role of annotation companies in the context of increasingly automated annotation processes driven by foundation models presents another avenue for future research." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "T. Rädsch et al." + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.143, + 0.761, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.265, + 0.274, + 0.74, + 0.285 + ], + "angle": 0, + "content": "(a) Annotation companies generate higher DSC scores than MTurk. \\(0 \\leq \\mathrm{DSC} \\leq 1\\)." + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.285, + 0.76, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.268, + 0.415, + 0.738, + 0.427 + ], + "angle": 0, + "content": "(b) Annotation companies generate higher NSD scores than MTurk. \\(0 \\leq \\mathrm{NSD} \\leq 1\\)." + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.427, + 0.76, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.305, + 0.567, + 0.698, + 0.58 + ], + "angle": 0, + "content": "(c) Annotation companies generate a lower number of severe errors." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.591, + 0.788, + 0.675 + ], + "angle": 0, + "content": "Fig. 7: Annotation companies generate higher quality annotations with their annotation pipeline than MTurk. This is observed for (a) the DSC, (b) the NSD, and (c) number of severe errors (lower is better). Annotation companies include a QA step, in contrast to crowdsourcing on MTurk. (a) and (b) displayed as dots and box plots (band indicates the median, box indicates the first (25th percentile) and third (75th percentile) quartiles)." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.712, + 0.403, + 0.73 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "A part of this work was funded by M. Mengler and S. Funke - during their time at understand.ai -, and Helmholtz Imaging. T. Rädsch was supported by a scholarship from the Hanns Seidel Foundation with funds from the Federal Ministry of Education and Research Germany (BMBF). We thank M. Gelz, S. Strzysch and N. Kraft for their continuous support." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.292, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Quality Assured: Rethinking Annotation Strategies in Imaging AI" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.175, + 0.787, + 0.3 + ], + "angle": 0, + "content": "1. Ali, S., Zhou, F., Braden, B., Bailey, A., Yang, S., Cheng, G., Zhang, P., Li, X., Kayser, M., Soberanis-Mukul, R.D., Albarqouni, S., Wang, X., Wang, C., Watanabe, S., Oksuz, I., Ning, Q., Yang, S., Khan, M.A., Gao, X.W., Realdon, S., Loshchenov, M., Schnabel, J.A., East, J.E., Wagnieres, G., Loschenov, V.B., Grisan, E., Daul, C., Blondel, W., Rittscher, J.: An objective comparison of detection and segmentation algorithms for artefacts in clinical endoscopy. Scientific Reports 10(1), 2748 (Feb 2020). https://doi.org/10.1038/s41598-020-59413-5, https://www.nature.com/articles/s41598-020-59413-5, number: 1 Publisher: Nature Publishing Group 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.301, + 0.787, + 0.577 + ], + "angle": 0, + "content": "2. Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M.S., Bohg, J., Bosselut, A., Brunskill, E., Brynjolfsson, E., Buch, S., Card, D., Castellon, R., Chatterji, N., Chen, A., Creel, K., Davis, J.Q., Demszky, D., Donahue, C., Doumbouya, M., Durmus, E., Ermon, S., Etchemendy, J., Ethayarajh, K., Fei-Fei, L., Finn, C., Gale, T., Gillespie, L., Goel, K., Goodman, N., Grossman, S., Guha, N., Hashimoto, T., Henderson, P., Hewitt, J., Ho, D.E., Hong, J., Hsu, K., Huang, J., Icard, T., Jain, S., Jurafsky, D., Kalluri, P., Karamcheti, S., Keeling, G., Khani, F., Khattab, O., Koh, P.W., Krass, M., Krishna, R., Kuditipudi, R., Kumar, A., Ladhak, F., Lee, M., Lee, T., Leskovec, J., Levent, I., Li, X.L., Li, X., Ma, T., Malik, A., Manning, C.D., Mirchandani, S., Mitchell, E., Munyikwa, Z., Nair, S., Narayan, A., Narayanan, D., Newman, B., Nie, A., Niebles, J.C., Nilforoshan, H., Nyarko, J., Ogut, G., Orr, L., Papadimitriou, I., Park, J.S., Piech, C., Portelance, E., Potts, C., Raghunathan, A., Reich, R., Ren, H., Rong, F., Roohani, Y., Ruiz, C., Ryan, J., Re, C., Sadigh, D., Sagawa, S., Santhanam, K., Shih, A., Srinivasan, K., Tamkin, A., Taori, R., Thomas, A.W., Tramer, F., Wang, R.E., Wang, W., Wu, B., Wu, J., Wu, Y., Xie, S.M., Yasunaga, M., You, J., Zaharia, M., Zhang, M., Zhang, T., Zhang, X., Zhang, Y., Zheng, L., Zhou, K., Liang, P.: On the Opportunities and Risks of Foundation Models (Jul 2022). https://doi.org/10.48550/arXiv.2108.07258. http://arxiv.org/abs/2108.07258 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.578, + 0.787, + 0.619 + ], + "angle": 0, + "content": "3. Bragg, J., Mausam, Weld, D.S.: Sprout: Crowd-Powered Task Design for Crowdsourcing. Proceedings of the ACM on User Interface Software and Technology (2018). https://doi.org/10.1145/3242587.3242598 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.787, + 0.687 + ], + "angle": 0, + "content": "4. Chang, J.C., Amershi, S., Kamar, E.: Revolt: Collaborative Crowdsourcing for Labeling Machine Learning Datasets. In: Proceedings of the 2017 CHI Conference on Human Factors in Computing Systems. pp. 2334-2346. CHI '17, Association for Computing Machinery, New York, NY, USA (May 2017). https://doi.org/10.1145/3025453.3026044 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.688, + 0.787, + 0.743 + ], + "angle": 0, + "content": "5. Chen, L., Li, S., Yan, J., Wang, H., Gunaratna, K., Yadav, V., Tang, Z., Srinivasan, V., Zhou, T., Huang, H., Jin, H.: AlpaGasus: Training A Better Alpaca with Fewer Data (Nov 2023). https://doi.org/10.48550/arXiv.2307.08701, http://arxiv.org/abs/2307.08701 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.787, + 0.84 + ], + "angle": 0, + "content": "6. Cheplygina, V., Perez-Rovira, A., Kuo, W., Tiddens, H.A.W.M., de Bruijne, M.: Early Experiences with Crowdsourcing Airway Annotations in Chest CT. In: Carneiro, G., Mateus, D., Peter, L., Bradley, A., Tavares, J.M.R.S., Belagiannis, V., Papa, J.P., Nascimento, J.C., Loog, M., Lu, Z., Cardoso, J.S., Cornebise, J. (eds.) Deep Learning and Data Labeling for Medical Applications. pp. 209-218. Lecture Notes in Computer Science, Springer International Publishing, Cham (2016). https://doi.org/10.1007/978-3-319-46976-8_22_4_6" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.175, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "T. Rädsch et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "7. Chmielewski, M., Kucker, S.C.: An MTurk Crisis? Shifts in Data Quality and the Impact on Study Results. Social Psychological and Personality Science 11(4), 464-473 (May 2020). https://doi.org/10.1177/1948550619875149, publisher: SAGE Publications Inc 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.786, + 0.245 + ], + "angle": 0, + "content": "8. Cognilytica: Data engineering, preparation, and labeling for ai 2020 (2020), https://www.cognilytica.com/2020/01/31/data-preparation-labeling-for-ai-2020/10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.248, + 0.787, + 0.303 + ], + "angle": 0, + "content": "9. Créquet, P., Mansouri, G., Benchoufi, M., Vivot, A., Ravaud, P.: Mapping of Crowdsourcing in Health: Systematic Review. Journal of Medical Internet Research 20(5), 187 (2018). https://doi.org/10.2196/jmir.9330, https://www.jmir.org/2018/5/e187/1" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.304, + 0.786, + 0.359 + ], + "angle": 0, + "content": "10. Daniel, F., Kucherbaev, P., Cappiello, C., Benatallah, B., Allahbakhsh, M.: Quality Control in Crowdsourcing: A Survey of Quality Attributes, Assessment Techniques and Assurance Actions. ACM Computing Surveys 51, 1-40 (Jan 2019). https://doi.org/10.1145/3148148 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.361, + 0.786, + 0.388 + ], + "angle": 0, + "content": "11. Dice, L.R.: Measures of the amount of ecologic association between species. Ecology 26(3), 297-302 (1945) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.39, + 0.787, + 0.457 + ], + "angle": 0, + "content": "12. Dow, S., Kulkarni, A., Klemmer, S., Hartmann, B.: Shepherding the crowd yields better work. In: Proceedings of the ACM 2012 conference on Computer Supported Cooperative Work. pp. 1013-1022. CSCW '12, Association for Computing Machinery, New York, NY, USA (Feb 2012). https://doi.org/10.1145/2145204.2145355 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.46, + 0.787, + 0.528 + ], + "angle": 0, + "content": "13. Duggan, N.M., Jin, M., Mendicuti, M.A.D., Hallisey, S., Bernier, D., Selame, L.A., Asgari-Targhi, A., Fischetti, C.E., Lucassen, R., Samir, A.E., Duhaime+, E., Kapur, T., Goldsmith, A.J.: Gamified Crowdsourcing as a Novel Approach to Lung Ultrasound Dataset Labeling (Jun 2023). https://doi.org/10.48550/arXiv.2306.06773, http://arxiv.org/abs/2306.06773 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.531, + 0.787, + 0.598 + ], + "angle": 0, + "content": "14. Gaikwad, S., Chhibber, N., Sehgal, V., Ballav, A., Mullings, C., Nasser, A., Richmond-Fuller, A., Gilbee, A., Gamage, D., Whiting, M., et al.: Prototype tasks: improving crowdsourcing results through rapid, iterative task design (2017). https://doi.org/10.48550/arXiv.1707.05645, https://arxiv.org/abs/1707.05645 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.601, + 0.787, + 0.641 + ], + "angle": 0, + "content": "15. Halevy, A., Norvig, P., Pereira, F.: The unreasonable effectiveness of data. IEEE intelligent systems 24(2), 8-12 (2009). https://doi.org/10.1109/MIS.2009.36, ISBN: 1541-1672 Publisher: IEEE 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.643, + 0.787, + 0.697 + ], + "angle": 0, + "content": "16. Han, D., Choe, J., Chun, S., Chung, J.J.Y., Chang, M., Yun, S., Song, J.Y., Oh, S.J.: Neglected free lunch-learning image classifiers using annotation byproducts (2023). https://doi.org/10.48550/arXiv.2303.17595, https://arxiv.org/abs/2303.17595 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.7, + 0.787, + 0.755 + ], + "angle": 0, + "content": "17. Heim, E., Roß, T., Teil, A., März, K., Stieltjes, B., Eisenmann, M., Lebert, J., Metzger, J., Sommer, G., Sauter, A.W.: Large-scale medical image annotation with crowd-powered algorithms. Journal of Medical Imaging 5(3), 034002 (2018). https://doi.org/10.1117/1.JMI.5.3.034002 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.757, + 0.787, + 0.811 + ], + "angle": 0, + "content": "18. Hettiachchi, D., Schaekermann, M., McKinney, T.J., Lease, M.: The Challenge of Variable Effort Crowdsourcing and How Visible Gold Can Help. Proceedings of the ACM on Human-Computer Interaction 5, 332:1-332:26 (Oct 2021). https://doi.org/10.1145/3476073 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.813, + 0.787, + 0.841 + ], + "angle": 0, + "content": "19. Hutson, M., Kanzheleva, O., Taggart, C., Campana, B.J.L., Duong, Q.: Quality Control Challenges in Crowdsourcing Medical Labeling (2019) 4" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.292, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Quality Assured: Rethinking Annotation Strategies in Imaging AI" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.215 + ], + "angle": 0, + "content": "20. Idrissi, B.Y., Bouchacourt, D., Balestriero, R., Evtimov, I., Hazirbas, C., Ballas, N., Vincent, P., Drozdjal, M., Lopez-Paz, D., Ibrahim, M.: ImageNet-X: Understanding Model Mistakes with Factor of Variation Annotations (Nov 2022). https://doi.org/10.48550/arXiv.2211.01866, http://arxiv.org/abs/2211.01866 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.272 + ], + "angle": 0, + "content": "21. Ipeirotis, P.G., Provost, F., Wang, J.: Quality management on Amazon Mechanical Turk. In: Proceedings of the ACM SIGKDD Workshop on Human Computation. pp. 64-67. ACM, Washington DC (Jul 2010). https://doi.org/10.1145/1837885.1837906, https://dl.acm.org/doi/10.1145/1837885.1837906 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.313 + ], + "angle": 0, + "content": "22. Irani, L.: The cultural work of microwork. New Media & Society 17(5), 720-739 (May 2015). https://doi.org/10.1177/1461444813511926, publisher: SAGE Publications 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.41 + ], + "angle": 0, + "content": "23. Irshad, H., Montaser-Kouhsari, L., Waltz, G., Bucur, O., Nowak, J., Dong, F., Knoblauch, N., Beck, A.: Crowdsourcing image annotation for nucleus detection and segmentation in computational pathology: evaluating experts, automated methods, and the crowd. In: Biocomputing 2015. pp. 294-305. WORLD SCIENTIFIC, Kohala Coast, Hawaii, USA (Nov 2014). https://doi.org/10.1142/9789814644730_0029, http://www.worldscientific.com/doi/abs/10.1142/9789814644730_0029 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.466 + ], + "angle": 0, + "content": "24. Joskowicz, L., Cohen, D., Caplan, N., Sosna, J.: Inter-observer variability of manual contour delineation of structures in CT. European Radiology 29(3), 1391-1399 (Mar 2019). https://doi.org/10.1007/s00330-018-5695-5, http://link.springer.com/10.1007/s00330-018-5695-5 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.787, + 0.548 + ], + "angle": 0, + "content": "25. Kentley, J., Weber, J., Liopyris, K., Braun, R.P., Marghoob, A.A., Quigley, E.A., Nelson, K., Prentice, K., Duhaime, E., Halpern, A.C., Rotemberg, V.: Agreement Between Experts and an Untrained Crowd for Identifying Dermoscopic Features Using a Gamified App: Reader Feasibility Study. JMIR Medical Informatics 11, e38412 (Jan 2023). https://doi.org/10.2196/38412, https://medinform.jmir.org/2023/1/e38412 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.605 + ], + "angle": 0, + "content": "26. Kohli, M.D., Summers, R.M., Geis, J.R.: Medical image data and datasets in the era of machine learning—whitepaper from the 2016 C-MIMI meeting dataset session. Journal of digital imaging 30(4), 392–399 (2017). https://doi.org/10.1007/s10278-017-9976-3, publisher: Springer 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.659 + ], + "angle": 0, + "content": "27. Le, J., Edmonds, A., Hester, V., Biewald, L.: Ensuring quality in crowdsourced search relevance evaluation: The effects of training question distribution. In: SIGIR 2010 workshop on crowdsourcing for search evaluation. vol. 2126, pp. 22-32 (2010) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.787, + 0.728 + ], + "angle": 0, + "content": "28. Le Ludec, C., Cornet, M., Casilli, A.A.: The problem with annotation. Human labour and outsourcing between France and Madagascar. Big Data & Society 10(2), 20539517231188723 (Jul 2023). https://doi.org/10.1177/20539517231188723, http://journals.sagepub.com/doi/10.1177/20539517231188723 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.787, + 0.77 + ], + "angle": 0, + "content": "29. Lee, A.N., Hunter, C.J., Ruiz, N.: Platypus: Quick, Cheap, and Powerful Refinement of LLMs (Aug 2023). https://doi.org/10.48550/arXiv.2308.07317, http://arxiv.org/abs/2308.07317 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.839 + ], + "angle": 0, + "content": "30. Litman, L., Robinson, J., Rosenzweig, C.: The relationship between motivation, monetary compensation, and data quality among US- and India-based workers on Mechanical Turk. Behavior Research Methods 47(2), 519-528 (Jun 2015). https://doi.org/10.3758/s13428-014-0483-x, http://link.springer.com/10.3758/s13428-014-0483-x 6" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "T. Rädsch et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.23 + ], + "angle": 0, + "content": "31. Lu, X., Ratcliffe, D., Kao, T.T., Tikhonov, A., Litchfield, L., Rodger, C., Wang, K.: Rethinking Quality Assurance for Crowdsourced Multi-ROI Image Segmentation. Proceedings of the AAAI Conference on Human Computation and Crowdsourcing 11(1), 103–114 (Nov 2023). https://doi.org/10.1609/hcomp.v11i1.27552, https://ojs(aaai.org/index.php/HCOMP/article/view/27552, number: 1 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.302 + ], + "angle": 0, + "content": "32. Maier-Hein, L., Mersmann, S., Kondermann, D., Bodenstedt, S., Sanchez, A., Stock, C., Kenngott, H.G., Eisenmann, M., Speidel, S.: Can masses of non-experts train highly accurate image classifiers? In: International conference on medical image computing and computer-assisted intervention. pp. 438-445. Springer (2014). https://doi.org/10.1007/978-3-319-10470-6_55 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.787, + 0.373 + ], + "angle": 0, + "content": "33. Maier-Hein, L., Reinke, A., Godau, P., Tizabi, M.D., Buettner, F., Christodoulou, E., Glocker, B., Isensee, F., Kleesiek, J., Kozubek, M., et al.: Metrics reloaded: recommendations for image analysis validation. Nature Methods 21, 195–212 (2024). https://doi.org/10.1038/s41592-023-02151-z, https://www.nature.com/articles/s41592-023-02151-z 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.374, + 0.787, + 0.514 + ], + "angle": 0, + "content": "34. Maier-Hein, L., Wagner, M., Ross, T., Reinke, A., Bodenstedt, S., Full, P.M., Hempe, H., Mindroc-Filimon, D., Scholz, P., Tran, T.N., Bruno, P., Kisilenko, A., Müller, B., Davitashvili, T., Capek, M., Tizabi, M.D., Eisenmann, M., Adler, T.J., Grohl, J., Schellenberg, M., Seidlitz, S., Lai, T.Y.E., Pekdemir, B., Roethlingshoefer, V., Both, F., Bittel, S., Mengler, M., Mündermann, L., Apitz, M., Kopp-Schneider, A., Speidel, S., Nickel, F., Probst, P., Kenngott, H.G., Müller-Stich, B.P.: Heidelberg colorectal data set for surgical data science in the sensor operating room. Scientific Data 8(1), 101 (Apr 2021). https://doi.org/10.1038/s41597-021-00882-2, https://www.nature.com/articles/s41597-021-00882-2, number: 1 Publisher: Nature Publishing Group 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.787, + 0.583 + ], + "angle": 0, + "content": "35. McCulloch, C.E., Searle, S.R.: Generalized, linear, and mixed models. John Wiley & Sons (2004). https://doi.org/10.1002/0471722073, https://books.google.com/books?hl=en&lr=&id=bWDPukohugQC&oi=fnd&pg=PR6&dq=Generalized, +Linear, +and+Mixed+Models.&ots=sSy3HM7CGo&sig=Ss04EHKG_xkcZZCYWSEJl6paHos 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.585, + 0.787, + 0.627 + ], + "angle": 0, + "content": "36. Miceli, M., Posada, J.: The Data-Production Dispositif. Proceedings of the ACM on Human-Computer Interaction 6(CSCW2), 1-37 (Nov 2022). https://doi.org/10.1145/3555561, https://dl.acm.org/doi/10.1145/3555561 2, 12, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.629, + 0.787, + 0.685 + ], + "angle": 0, + "content": "37. Michaelis, C., Mitzkus, B., Geirhos, R., Rusak, E., Bringmann, O., Ecker, A.S., Bethge, M., Brendel, W.: Benchmarking Robustness in Object Detection: Autonomous Driving when Winter is Coming (Mar 2020). https://doi.org/10.48550/arXiv.1907.07484, http://arxiv.org/abs/1907.07484 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.787, + 0.755 + ], + "angle": 0, + "content": "38. Petrovic, N., Moya-Alcover, G., Varona, J., Jaume-i Capó, A.: Crowdsourcing human-based computation for medical image analysis: A systematic literature review. Health Informatics Journal 26(4), 2446-2469 (Dec 2020). https://doi.org/10.1177/1460458220907435, http://journals.sagepub.com/doi/10.1177/1460458220907435 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.787, + 0.84 + ], + "angle": 0, + "content": "39. Reinke, A., Tizabi, M.D., Baumgartner, M., Eisenmann, M., Heckmann-Nötzel, D., Kavur, A.E., Rädsch, T., Sudre, C.H., Acion, L., Antonelli, M.: Understanding metric-related pitfalls in image analysis validation. Nature methods pp. 1-13 (2024). https://doi.org/doi.org/10.1038/s41592-023-02150-0, https://www.nature.com/articles/s41592-023-02150-0, publisher: Nature Publishing Group 7" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.292, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Quality Assured: Rethinking Annotation Strategies in Imaging AI" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "40. Research, G.V.: Data Collection And Labeling Market Size & Share Report 2030 (2022), https://www.grandviewresearch.com/industry-analysis/data-collection-labeling-market-3,11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.259 + ], + "angle": 0, + "content": "41. Rofs, T., Bruno, P., Reinke, A., Wiesenfarth, M., Koeppel, L., Full, P.M., Pekdemir, B., Godau, P., Trofimova, D., Isensee, F.: How can we learn (more) from challenges? A statistical approach to driving future algorithm development. arXiv preprint arXiv:2106.09302 (2021). https://doi.org/10.48550/arXiv.2106.09302, https://arxiv.org/abs/2106.09302 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.787, + 0.342 + ], + "angle": 0, + "content": "42. Rädsch, T., Reinke, A., Weru, V., Tizabi, M.D., Schreck, N., Kavur, A.E., Pekdemir, B., Roß, T., Kopp-Schneider, A., Maier-Hein, L.: Labelling instructions matter in biomedical image analysis. Nature Machine Intelligence 5(3), 273–283 (Mar 2023). https://doi.org/10.1038/s42256-023-00625-5, https://www.nature.com/articles/s42256-023-00625-5, number: 3 Publisher: Nature Publishing Group 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.787, + 0.424 + ], + "angle": 0, + "content": "43. Sambasivan, N., Kapania, S., Highfill, H., Akrong, D., Paritosh, P., Aroyo, L.M.: \"Everyone wants to do the model work, not the data work\": Data Cascades in High-Stakes AI. In: Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. pp. 1-15. ACM, Yokohama Japan (May 2021). https://doi.org/10.1145/3411764.3445518, https://dl.acm.org/doi/10.1145/3411764.3445518 1, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.424, + 0.787, + 0.495 + ], + "angle": 0, + "content": "44. Schmidt, F.A.: The Planetary Stacking Order of Multilayered Crowd-AI Systems. In: Graham, M., Ferrari, F. (eds.) Digital Work in the Planetary Market, pp. 137-156. The MIT Press (Jun 2022). https://doi.org/10.7551/mitpress/13835.003.0012, https://direct.mit.edu/books/book/5319/chapter/3800159/The-Planetary-Stacking-Order-of-Multilayered-Crowd 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.562 + ], + "angle": 0, + "content": "45. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: LLaMA: Open and Efficient Foundation Language Models (Feb 2023). https://doi.org/10.48550/arXiv.2302.13971, http://arxiv.org/abs/2302.13971 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.563, + 0.787, + 0.618 + ], + "angle": 0, + "content": "46. Veselovsky, V., Ribeiro, M.H., West, R.: Artificial Artificial Artificial Intelligence: Crowd Workers Widely Use Large Language Models for Text Production Tasks (Jun 2023). https://doi.org/10.48550/arXiv.2306.07899, http://arxiv.org/abs/2306.07899 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.787, + 0.715 + ], + "angle": 0, + "content": "47. Wang, K., Hill, M., Knowles-Barley, S., Tikhonov, A., Litchfield, L., Bare, J.C.: Improving Segmentation of Breast Arterial Calcifications from Digital Mammography: Good Annotation is All You Need. In: Zheng, Y., Keles, H.Y., Koniusz, P. (eds.) Computer Vision - ACCV 2022 Workshops, vol. 13848, pp. 134-150. Springer Nature Switzerland, Cham (2023). https://doi.org/10.1007/978-3-031-27066-6_10, https://link.springer.com/10.1007/978-3-031-27066-6_10, series Title: Lecture Notes in Computer Science 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.715, + 0.787, + 0.811 + ], + "angle": 0, + "content": "48. Yu, S., Zhou, H.Y., Ma, K., Bian, C., Chu, C., Liu, H., Zheng, Y.: Difficulty-Aware Glaucoma Classification with Multi-rater Consensus Modeling. In: Martel, A.L., Abolmaesumi, P., Stoyanov, D., Mateus, D., Zuluaga, M.A., Zhou, S.K., Racoceanu, D., Joskowicz, L. (eds.) Medical Image Computing and Computer Assisted Intervention - MICCAI 2020. pp. 741-750. Lecture Notes in Computer Science, Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-59710-8_72_4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.812, + 0.787, + 0.841 + ], + "angle": 0, + "content": "49. Ørting, S.N., Doyle, A., Hilten, A.v., Hirth, M., Inel, O., Madan, C.R., Mavridis, P., Spiers, H., Cheplygina, V.: A Survey of Crowdsourcing in Medical Image Analysis." + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "T. Rädsch et al." + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.148, + 0.786, + 0.177 + ], + "angle": 0, + "content": "Human Computation 7, 1-26 (Dec 2020). https://doi.org/10.15346/hc.v7i1.1, https://hcjournal.org/index.php/jhc/article/view/111 1" + } + ] +] \ No newline at end of file diff --git a/2024/Quality Assured_ Rethinking Annotation Strategies in Imaging AI/1dd3eeac-8808-45a9-8928-d1c207e0686a_origin.pdf b/2024/Quality Assured_ Rethinking Annotation Strategies in Imaging AI/1dd3eeac-8808-45a9-8928-d1c207e0686a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..894110eee35860446a8d204997df9660821f716e --- /dev/null +++ b/2024/Quality Assured_ Rethinking Annotation Strategies in Imaging AI/1dd3eeac-8808-45a9-8928-d1c207e0686a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb8a0bebabfe842fe462f1dcc318426fad621c8fcf7dadd18083e01b4223d69b +size 1861522 diff --git a/2024/Quality Assured_ Rethinking Annotation Strategies in Imaging AI/full.md b/2024/Quality Assured_ Rethinking Annotation Strategies in Imaging AI/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c281b96c13251807f23cc599abc91de44b098378 --- /dev/null +++ b/2024/Quality Assured_ Rethinking Annotation Strategies in Imaging AI/full.md @@ -0,0 +1,257 @@ +# Quality Assured: Rethinking Annotation Strategies in Imaging AI + +Tim Rädsch1,2,3, Annika Reinke1,2, Vivienne Weru1, Minu D. Tizabi1, Nicholas Heller4, Fabian Isensee1,2, Annette Kopp-Schneider1*, and Lena Maier-Hein1,2,3,5* + +$^{1}$ German Cancer Research Center, GER $^{2}$ Helmholtz Imaging, GER $^{3}$ Heidelberg University, GER $^{4}$ University of Minnesota, USA $^{5}$ National Center for Tumor Diseases, GER $\boxed{\boxtimes}$ :{tim.raedsch; l.maier-hein}@dkfz-heidelberg.de + +Abstract. This paper does not describe a novel method. Instead, it studies an essential foundation for reliable benchmarking and ultimately real-world application of AI-based image analysis: generating high-quality reference annotations. Previous research has focused on crowdsourcing as a means of outsourcing annotations. However, little attention has so far been given to annotation companies, specifically regarding their internal quality assurance (QA) processes. Therefore, our aim is to evaluate the influence of QA employed by annotation companies on annotation quality and devise methodologies for maximizing data annotation efficacy. Based on a total of 57,648 instance segmented images obtained from a total of 924 annotators and 34 QA workers from four annotation companies and Amazon Mechanical Turk (MTurk), we derived the following insights: (1) Annotation companies perform better both in terms of quantity and quality compared to the widely used platform MTurk. (2) Annotation companies' internal QA only provides marginal improvements, if any. However, improving labeling instructions instead of investing in QA can substantially boost annotation performance. (3) The benefit of internal QA depends on specific image characteristics. Our work could enable researchers to derive substantially more value from a fixed annotation budget and change the way annotation companies conduct internal QA. + +Keywords: Data Annotation $\cdot$ Quality Assurance $\cdot$ Crowdsourcing + +# 1 Introduction + +Following a long period of extensive focus on pure method development, proper benchmarking and validation have recently been recognized as a major challenge in AI-based image analysis [33, 43]. In this context, high-quality annotation of reference datasets is a fundamental requirement to ensure the meaningful validation of image analysis algorithms [15]. As annotation tasks may be cumbersome, crowdsourcing is often used to partially outsource image annotation. Traditionally, such outsourcing is conducted via crowdsourcing platforms [32], with Amazon Mechanical Turk (MTurk)1 as the predominant choice [9, 49]. + +![](images/061f4698a80d451656990b78fc9c077e4110fb5d60a98b97de5f2fb5c2e4a271.jpg) +Fig. 1: Research Questions (RQs) tackled in this work. Based on 57,648 instance segmentation masks annotated by 924 annotators and 34 quality assurance (QA) workers from five different annotation providers, we (1) compared the effectiveness of generating high-quality annotations between annotation companies and Amazon Mechanical Turk and (2) investigated the effects of annotation companies' internal QA and (3) real-world image characteristics on the annotation quality. + +Recent literature has reflected the rise of annotation companies as a new type of provider with potentially better annotation performance compared to established crowdsourcing platforms such as MTurk [28,36,42,44]. Annotation companies differ substantially from crowdsourcing platforms. First, they employ annotators directly and at shared office spaces [28,44], enabling better external control of worker conditions and stricter data security measures, while crowdsourcing platforms can be accessed by anyone, without specific knowledge or training being required. Second, annotators in annotation companies typically work in small teams, with an experienced quality assurance (QA) worker being responsible for the team's output. In the observed internal QA, QA workers make adjustments to the annotations generated by the team by reviewing each individual annotation and modifying those where they deem changes necessary. As a result, the annotation requester receives the quality-assured annotation, not the initial one (see Fig. 2a). In contrast, crowdsourcing platforms usually do not employ internal QA mechanics similar to annotation companies. + +The importance of annotation providers for real-world applications is evident at the major computer vision conferences, where over the past two years, approximately $20\%$ of exhibitors were annotation providers (see Fig. 2b). While QA as a post-processing step for annotations from platforms such as MTurk has been researched (e.g. [10, 17, 21, 27]), little is known about the benefit of internal QA on annotation quality within annotation companies. The only related work that we are aware of pertains to external QA from the perspective of an annotation company. Unlike in internal QA, in the case of external QA, domain experts or experienced annotators working on behalf of the requester check the annotation quality as a post-processing step of the received annotations [31, 47]. Our paper, on the other hand, focuses on the internal QA processes of an annotation company. + +![](images/08608e7b3b91ceaa4f4d71b9628e1435ebf43201ebda0d9b769d31413438f963.jpg) +Fig. 2: a) In contrast to Amazon Mechanical Turk (MTurk), annotation companies commonly perform quality assurance (QA) before delivering the result to the annotation requester. b) Annotation providers are an integral part of the computer vision community. They account for around $20\%$ of exhibitors present at major computer vision conferences (ECCV 2022: $\mathrm{n} = 40$ , CVPR 2022: $\mathrm{n} = 97$ , ICCV 2023: $\mathrm{n} = 43$ , CVPR 2023: $\mathrm{n} = 104$ ) in the last two years. + +In light of the vast sums of money to be spent on QA [40], the purpose of this paper is to shed light on whether annotation companies' internal QA benefits the annotation quality. Based on more than 50,000 instance segmentation masks annotated by more than 900 annotators from five different annotation providers, we investigate the following research questions (RQs; see Fig. 1): + +RQ 1) How does the choice of annotation provider influence the quality of annotations? As the selection of the annotation provider is among the first and key decisions in any annotation project after data collection and generation of labeling instructions, we first investigated companies' effectiveness in generating annotations, which is particularly relevant in cases of constrained annotation budgets requiring a cost-effective strategy. As both quantity and quality of annotated images for a given budget are relevant in provider selection, we compared the annotation quality and costs of five different providers, including MTurk and four annotation companies, on identical images. + +RQ 2) Does annotation companies' quality assurance improve annotation quality? A major decision in the design of an annotation process concerns the allocation of resources. A common perception is that internal QA of annotation companies, following the initial annotation, substantially enhances the quality. Here, we challenged this assumption by investigating the improvement in annotation quality from internal QA compared to that achieved through enhanced labeling instructions. + +RQ 3) Should quality assurance focus on specific subsets of images? In theory, the efficacy of QA may be contingent upon specific image characteristics [20, 24, 41]. For instance, underexposed images might benefit more from a secondary review compared to those with optimal illumination. Concurrently, images that present annotation challenges for human observers might be partic + +ularly pertinent for the training and validation of algorithms [20, 41]. Here, we investigated whether and to what extent the benefit of QA depends on specific image characteristics. + +# 2 Related Work + +# 2.1 Comparison of annotation providers + +With respect to the quality of annotations obtained from varying annotation providers, the closest work conducted are comparisons between crowdsourcing platforms and domain experts. Irshad et al. [23] and Petrovic et al. [38], among others, have shown that crowdsourced annotations can reach the annotation quality of domain experts in safety-critical applications. In recent literature, Duggan et al. [13] and Kentley et al. [25] compared domain-specialized crowds as a new type of annotation provider and found comparable performance to domain experts. Furthermore, Rädsch et al. [42] provided a comparison of annotation companies to crowdsourcing, but focused solely on raw annotation quality without QA and without relating it to annotation budgets. + +Interestingly, researchers report concerns about the annotation quality obtained from MTurk [31], such as obtaining unusable annotations in up to $67.8\%$ of cases [6]. Other research fields, such as psychology, even report a "substantial decrease in data quality" [7]. However, we are not aware of work that has systematically compared the performance between annotation companies and crowdsourcing platforms such as MTurk while considering annotation budgets. + +# 2.2 Quality assurance + +The main goal of QA is to enhance and ensure the quality of annotations. A comprehensive overview on QA in crowdsourcing platforms is provided by Daniel et al. [10]. However, their research does not include annotation companies. During the annotation process, quality control techniques generally fall into three categories [31]: (1) comparing to a gold standard with pre-defined reference annotations (e.g., [27]), (2) analyzing combined outputs from multiple annotators (e.g., [21]), and (3) investigating annotator behaviors (e.g., [17, 19]). Other approaches to enhancing annotation quality on crowdsourcing platforms target training of crowdworkers [12] and iterative refinement of the annotation environment [3,4,14]. It should be noted that annotation byproducts, such as mouse movement during annotation generation [16, 17] or inter-rater agreement [48], can be used to boost model performance, by incorporating this information in the training. + +The closest related work focusing on the impact of QA in annotation companies was introduced by Wang et al. [47] and Lu et al. [31]. However, their work focuses on external QA conducted by experienced personnel from the requester rather than internal QA, and features a substantially smaller number of annotators from annotation companies. To our knowledge, a systematic investigation into the effects of internal QA processes within annotation companies is thus, so far, lacking. + +# 2.3 Impact of real-world image characteristics on annotation quality + +Related research analyzing the impact of real-world image characteristics highlights their effect on annotations created by machine learning models and annotators alike. Michaelis et al. [37] demonstrated that models easily encounter failures with increasingly challenging real-world image characteristics, such as added snow or poor lighting characteristics. Similarly, annotators generate less reliable annotations in the presence of challenging real-world image characteristics. This effect can be observed both within [1,24,31] and across domains [18,26]. However, a systematic analysis of the impact of such characteristics on internal QA employed by annotation companies is lacking. + +# 3 Methods + +In the following, we describe the selected data, the labeling instructions, annotation providers and annotators, the QA process, and the experimental design of our work. + +# 3.1 Data + +We conducted our experiments on the instance segmentation data from the Heidelberg Colorectal (HeiCo) dataset for surgical data science in the sensor operating room [34], representing a safety-critical application that crowdsourcing platforms can still handle. The task involved segmenting surgical instruments from laparoscopic videos, resulting in 57,648 instance segmentation masks on 4,050 unique frames. Furthermore, we used 57,636 metadata annotations on the unique images to quantify the effect of varying real-world image characteristics (see RQ3). Characteristics included motion blur, underexposure, object occlusion, or overlapping objects, among others, and are illustrated in Suppl. A. + +# 3.2 Labeling instructions + +To investigate the benefit of QA compared to an investment in detailed labeling instructions, we created three distinct types of labeling instructions with varying levels of information density, namely, (1) minimal text, (2) extended text, and (3) extended text including exemplary pictures. Each subsequent labeling instruction type was structured to contain more detailed information than the preceding. To ensure accurate labeling of the imaging data, we established a collection of design principles applicable to all labeling instructions: + +(1) Labeling information is presented in a slide format to enhance the processing of information by humans. +(2) Each slide contains a distinct 'chunk' of information, which is a self-contained piece of content. This approach of 'chunking' information lessens the cognitive load on the working memory. +(3) Related information chunks are placed in close proximity to one another. + +(4) A uniform design template is used throughout, featuring specific fonts, symbols, and color schemes. + +The three labeling instructions differ as follows: + +Minimal text labeling instructions: These instructions provide a concise textual outline, offering examples of typical annotations and highlighting the most frequent annotation types (see Suppl. B.1). They reflect a scenario with minimal effort invested in developing the instructions. The basic framework includes seven slides totaling 168 words. + +Extended text labeling instructions: Building upon the simplified version, the extended text instructions offer a thorough textual explanation, supplemented by both positive and negative examples (refer to Suppl. B.2). They cover both regular and some rare annotation scenarios. The enhanced version comprises ten slides with a total of 446 words. + +Extended text including pictures labeling instructions: These instructions augment the enhanced textual guidelines with images (detailed in Suppl. B.3). The visual aids feature descriptions, symbols, annotations, and color coding to effectively communicate the details on the slides. They also include illustrations of rare annotation cases, indicating a deep and well-documented understanding of the labeling process. This set contains 16 slides with a total of 961 words. We refer the reader to Suppl. B for full details of each labeling instruction type. + +# 3.3 Annotation providers + +The study was conducted based on 216 annotators and 34 QA workers from four annotation companies and 708 crowdworkers from the crowdsourcing platform MTurk. Our goal was to provide the best representation for both annotation provider types. + +Consequently, we established higher standards for our MTurk crowdworkers compared to the usual research benchmarks, which generally require a $95\%$ acceptance rate of Human Intelligence Tasks (HITs) and at least 100 completed tasks [3, 6, 17]. Our criteria included a $98\%$ acceptance rate with over 5,000 completed HITs. We spread out the HITs over 40 days at all times to obtain a representative MTurk sample and compensated workers fairly, as recommended by Litman et al. [30]. The annotation companies we selected have a solid background in handling large-scale annotation tasks in a variety of domains and operate internationally. The annotators were allocated by the companies themselves to replicate the typical workflow of a project carried out by an annotation company. Each annotator was linked to a specific annotation provider that had been predetermined. No selection process from our side took place for either the annotators, QA workers, or the MTurk crowdworkers. + +To prevent information leakage, we conducted the labeling process with increasingly detailed labeling instructions (minimal text instruction, extended text instruction, and extended text including pictures instruction). Each annotator worked exclusively with one of the three instruction types as introduced in Sec. 3.2. To enhance security, we included a mandatory minimum gap of ten days between consecutive labeling instructions. + +# 3.4 Quality assurance process + +In the internal QA process, annotators typically report to a small number of experienced QA workers, who are often responsible for training their team and, in our case, responsible for the QA of the annotations. + +The general workflow for all companies consisted of an annotation stage, in which images were assigned randomly to each annotator, and a QA stage. During the annotation stage, annotators received raw images with the task of annotating them until they considered the quality of annotations satisfactory. Annotators were aware that their work would be checked by QA. Once submitted, the QA worker gained access to the generated annotations, allowing them to make modifications until they were satisfied before submitting the final annotations for delivery to the requester. Throughout this process, both annotators and QA workers retained access to the respective labeling instructions at any given moment. We obtained access to the annotations both before and after QA through a software company handling a substantial number of annotation projects annually. The experiments were seamlessly integrated into the software company's pipeline of annotation projects, reducing the risk of exposing our experiments. Pricing was established to be representative of other projects of similar size. + +To maintain fairness and avoid giving any annotation provider a potential informational advantage that may have affected our statistical analysis, no specific queries from the annotation companies about the content of the labeling instructions were addressed. + +# 3.5 Experimental design + +RQ 1: Each of the five labeling providers annotated the same images with each set of labeling instructions, employing separate annotators for each task. We analyzed the number of spam images, images with severe errors, i.e., images with at least one false positive and/or false negative, and the overall annotation quality with the Dice Similarity Coefficient (DSC) [11] as the overlap-based metric and the Normalized Surface Distance (NSD) [39] as the distance-based metric. We define a spam annotation as a purposefully poor-quality annotation with the intent to obtain the payment without properly conducting the annotation task. Annotation duration was used as an auxiliary measure, where accessible. For instance, if an annotator completed the task in only 17 seconds, they could not have read the instructions and properly performed annotations on several images. Of note, a spammer could wait a relevant number of seconds and then enter a poor-quality annotation to bypass the annotation duration check. Thus, each image was additionally visually inspected by an engineer with extensive labeling experience. Spam annotation examples are displayed in Suppl. D. + +RQ 2: The annotation companies conducted their internal QA processes directly after their annotators completed the annotation job, as explained in Sec. 3.4 and Fig. 2a. The experiments began with only providing the minimal text labeling instructions, followed by the extended text labeling instructions, + +![](images/bfbdf51154c1873a74b129cde29ab08df2292fbbea703a12bc8b83545e5fdd59.jpg) +Fig. 3: Annotation companies are more efficient at generating high-quality images than Amazon Mechanical Turk (MTurk). Percentage of images without severe errors (green), images with severe errors (orange), and spam (red) obtained for the five different annotation providers investigated. Annotation costs per image were lower for all companies compared to MTurk (median: $61.1\%$ ). + +and finally, the extended text instructions with pictures, to prevent information leakage. + +For each annotated image, we calculated the DSC and NSD per instance, aggregated the scores per annotated image, and aggregated the scores over annotation company and labeling instruction. We then compared the annotation quality without and with QA across the three types of labeling instructions for the four annotation companies by computing both absolute and relative changes in DSC/NSD aggregates, the percentage of modified images, and the change in the percentage of severe errors, among other parameters. + +RQ 3: In contrast to MTurk, the four annotation companies conducted an additional annotation round, including QA, with the extended text including pictures labeling instructions to quantify the impact of different real-world image characteristics, represented by 57,636 metadata annotations, on the probability of improvement. The data was analyzed with a logistic mixed model [35] which included random effects for the image, the annotator, and the annotation company. Further method details are provided in Suppl. C. + +# 4 Results + +# 4.1 How does the choice of annotation provider influence the quality and cost of annotations? + +Annotated images generated by the annotation companies were of higher quality, comprising no spam (MTurk: $\sim 20\%$ ) and a much higher proportion of remaining images without severe annotation errors (factor of $>2$ improvement for all companies compared to MTurk), as shown in Fig. 3. In addition, the annotation companies were $61.1\%$ cheaper in the median, despite their additional employment of QA workers. More specifically, the costs ranged from $24.5\%$ to $81.5\%$ of the MTurk costs for the same images and number of annotations. We have + +(a) +Fig. 4: QA of annotation companies only provides marginal improvements, if any. These depend highly on the company and type of labeling instruction. a: Absolute change of the median DSC resulting from performing QA. For each company and labeling instruction, the median was obtained by aggregating the DSC scores over the images. b: Corresponding absolute values of the DSC. Positive and negative changes from annotation to QA are colored according to the color scale. +![](images/fe018066ac0287db08f641d2088b2f831e466d2d8629b14fe815d48281c48e45.jpg) +Annotation company Company 1 Company 2 Company 3 Company 4 + +(b) + +
CompanyLabeling instruction25th PercentileMedian75th Percentile
AnnotationQAAnnotationQAAnnotationQA
11) Min. text0.610.620.930.930.960.96
2) Ext. text0.580.580.930.930.960.96
3) Ext. text + pictures0.900.900.950.950.960.96
21) Min. text0.570.470.920.900.960.95
2) Ext. text0.620.840.930.940.960.96
3) Ext. text + pictures0.850.880.940.940.960.96
31) Min. text0.600.600.930.930.960.96
2) Ext. text0.470.640.910.930.960.96
3) Ext. text + pictures0.740.850.930.940.960.96
41) Min. text0.630.630.940.940.960.96
2) Ext. text0.880.880.940.940.960.96
3) Ext. text + pictures0.910.910.950.950.970.97
+ +COLOR SCALE 1.00 0.75 0.50 0.25 0.00 -0.25 -0.50 0.75 -1.00 + +not been given permission to reveal the identities and exact absolute budgets used for the annotation companies, hence we report costs relative to the MTurk budget. Examples of spam annotations are provided in Suppl. D. + +Regardless of the annotation scenario, the annotation companies generated higher quality annotations than MTurk with higher DSC scores including smaller interquartile ranges (IQRs) (see Fig. 7a), higher NSD scores including smaller IQRs (see Fig. 7b), and lower numbers of severe errors (see Fig. 7c). + +# 4.2 Does annotation companies' quality assurance improve annotation quality? + +As shown in Fig. 4, the internal QA conducted by annotation companies does not substantially improve annotation quality if the QA workers only use minimal text instructions. The impact of QA when having access to extended text instructions or extended text including pictures instructions depended highly on the annotation company, while the effect was still minimal compared to not performing QA: Two of the four annotation companies show no change in the median DSC when comparing QA to non-QA outputs for extended text or extended text and picture instructions. The remaining two companies only show a very small relative improvement of below $2\%$ . + +We obtained even smaller improvements with the NSD scores, as depicted in Suppl. E. In certain scenarios, QA was able to remove a small proportion of severe errors, however, the vast majority remains (see Suppl. F). Notably, the percentage of modified images by QA workers depended highly on the company, as depicted in Suppl. G. Fig. 5 further illustrates that providing annotators with high-quality labeling instructions including pictures (green) leads them to vastly + +![](images/6e4b981eabe812cf9322d05a5db34aeb5fc9d0ddce135d8bc953de53d4effa00.jpg) + +![](images/e21b511fe7d170cc6b84b0cb6c2b2f9f5a415a00e726d65b9c1f5eab932c261f.jpg) + +![](images/8441f60cd3965c736c8ad62c652025e0f18cfdc2ee53735e671eabfc2dd3b52f.jpg) +Fig. 5: Improving the labeling instructions yields a higher effect compared to adding QA to the annotation process. The baseline performance (minimal labeling instruction, no QA) according to the DSC (a) and the percentage of severe errors (b) is shown individually per company with a red asterisk (minimal labeling instruction, no QA). While adding QA only yields marginal improvements (dark blue), major performance boosts are obtained by improving the labeling instructions provided to the annotators (green). DSC scores were aggregated for each image and per labeling type. DSC displayed as dots and box plot (band indicates the median, box indicates the first (25th percentile) and third (75th percentile) quartiles). $0 \leq \mathrm{DSC} \leq 1$ . + +outperform annotators working with lower-quality labeling instructions, even if these are followed by additional internal QA (dark blue). + +# 4.3 Should quality assurance focus on specific subsets of images? + +As shown in Fig. 6a, QA significantly increases the odds of improving the annotation quality for images with generalizable image characteristics such as underexposed objects, intersection of objects, objects occluded by background in comparison to the absence of this characteristic (vertical black line). The only exceptions are motion artifacts present in the background and object(s) overexposure. + +In contrast to the general increase of images with generalizable image characteristics, none of the domain-specific image characteristics yield a significant change in the odds of improving the annotation quality over a regular image, see Fig. 6b. The only exception for domain-specific characteristics is object(s) covered by smoke, which could be interpreted as a form of object occlusion. + +# 5 Discussion + +The data annotation market is estimated to reach a volume of more than USD 4.1 billion by the end of 2024 [8]. As the annotation process becomes increasingly + +![](images/8703522a81b1d11a71e79334a1e70ca4661eaf6523b4fa5500fb88ee5fd6a924.jpg) +(a) GENERALIZABLE CHARACTERISTICS + +![](images/b1db69be1ef2add7f6d9688be19cea42fec65d9bcba3f8b654b2a3ced5c6f6ac.jpg) +(b) DOMAIN-SPECIFIC CHARACTERISTICS +Fig. 6: QA of annotation companies can be beneficial to certain images with difficult real-world generalizable image characteristics, such as underexposure or intersection of objects. We display the impact of (a) real-world generalizable and (b) domain-specific image characteristics on the odds of improvement. Improvement is defined as a positive difference in the QA, while no improvement refers to 0 or negative difference. The effect of an image characteristic is shown in the form of the log(odds ratio (OR)) representing the logarithms of the odds of improvement in the presence of the characteristic, compared to the odds of improvement occurring in the absence of that characteristic. The vertical black line refers to the point of equal odds. (Note: ***: p-value ≤ 0.001, **: p-value ≤ 0.01, *: p-value ≤ 0.05). + +automated and foundation models evolve, the market is undergoing drastic shifts. With lower entry barriers for annotation, more efficient data annotation tools, and a rising impact of QA, it is thought to result in an estimated USD 8.2 billion market by 2028 [40]. As a research community, we need to understand how to ensure high annotation quality throughout this evolution, as reliable reference annotations are an indispensable foundation for the benchmarking of image analysis algorithms, and thus their ultimate translation into real-world practice. + +To our knowledge, this study is the first to quantitatively and critically examine the impact of internal QA conducted by annotation companies in the context of safety-critical applications. The most important insights can be summarized as follows: + +1. Annotation companies perform better both in terms of quantity and quality compared to the widely used platform MTurk (Fig. 3). This implies that requesters working with crowdsourcing platforms should seriously consider annotation companies as an alternative. The organizational structure of these companies, which enables a single point of contact for requesters and eliminates spam annotations, appears to be a contributing factor to their efficiency. In light of poor working conditions being a major problem in annotation generation [22, 36], annotation companies, which can more easily be inspected than distributed crowdsourcing platforms, are thought to offer better working contracts, enabling fairer working conditions for annotators. It should be noted that not every annotation company automatically ensures adequate working conditions. The annotation companies employed in this study underwent thorough scrutiny to guarantee compliance with good working condition guidelines. With annotators working at a shared location instead of in a distributed crowd, verifying annotators' working conditions has become substantially easier. We thus view our work as an initial step towards greater accountability on the part of annotation requesters. Further dimensions of annotation provider selection are discussed in Suppl. H. + +2. In contrast to common belief, internal QA mechanisms do not necessarily improve annotation quality (Fig. 4). Furthermore, we observe no systematic QA quality improvement difference between annotation companies depending on their region of origin, size, or price, meaning more expensive annotation companies do not translate to better QA. However, improving labeling instructions instead of investing in QA can substantially boost the annotation performance (Fig. 5). Consequently, in order to obtain high-quality annotations, annotation requesters should prioritize refining their labeling instructions, followed by implementing QA or even a different external QA process as a secondary step. Compared to QA, better labeling instructions can be generated with relatively few resources. + +3. The benefit of QA depends on the specific image characteristics (Fig. 6). In fact, QA did improve the quality of images with certain challenging real-world image characteristics, see Fig. 6a. Consequently, it might be useful for requesters and annotation companies alike to allocate resources to identifying challenging image characteristics during a project, train classifiers to identify images containing these characteristics and focus the internal QA effort on these images. + +One could contend that the recent rise of foundation models marginalizes the role of annotation providers. We argue, however, that it only changes their role. In this context, it is necessary to differentiate between the data quality demands for the training and testing of models. With a recent trend towards fine-tuning pre-trained models, smaller high-quality datasets can achieve similar performance with less computation [5, 29], highlighting the importance of the quality of unlabeled data in unsupervised training and its role in reducing carbon emissions [45]. In testing, especially for safety-critical applications such as autonomous driving and clinical medicine, ensuring high-quality annotated test + +data is crucial for accurate model evaluation and real-world applicability [43]. Consequently, while training data annotation may lose relevance, test data annotation and its high quality demands are here to stay. Regarding foundation models, it should further be noted that annotations purely produced by foundation models introduce artifacts and biases [2] and represent an increasing rising problem for MTurk, as demonstrated by Veselovsky et al. [46]. + +Our study is subject to several limitations which deserve further discussion. Comparing prices across providers is non-trivial as the costs to annotate a given data set depend on many factors including, among others, image quantity, type of annotations, location of annotation company and quality requirements. Our analysis was based on the actual costs incurred during our project. Confidentiality agreements with the software company which integrated our experiments into their annotation project pipeline prevent us from disclosing the identities of the annotation companies. Additionally, we took precautions to protect the identities of individual annotators and QA workers. This was particularly important given our findings that QA workers did not significantly enhance annotation quality, potentially exposing them to negative repercussions. + +Another limitation could be seen in the fact that our data set only represents one safety-critical domain. While this is a valid concern, we would like to highlight that we chose a dataset which combined quality, complexity, and volume. Furthermore, regarding the breadth of annotation providers and the number of annotators involved, our study is likely the largest of its kind, thereby incurring substantial costs. + +Finally, conducting multiple experiments with identical annotation companies and varying datasets could have potentially revealed our experimental design. Had the annotation companies been aware of our research, outcomes could potentially have been biased to appear more favorable, given companies' inherent interest in demonstrating their services as reliable. We thus chose to conduct our experiments such that they could be seamlessly integrated into the software company's annotation pipeline, reducing their likelihood of exposure. + +# 6 Conclusion and Future Work + +In conclusion, our study underscores the pivotal role of enhancing the initial annotation quality, for instance via allocating resources to the improvement of labeling instructions, over QA in obtaining higher reference annotation quality in high-stakes AI [43]. Future research should delve into the development and implementation of comprehensive training programs for annotators from annotation companies to further enhance initial annotation quality. Exploring the role of annotation companies in the context of increasingly automated annotation processes driven by foundation models presents another avenue for future research. + +![](images/3e05d47b2ce47ec50a7b906ae56fde5c4e572da9e5397f0c35f7d6ecccd4bc0c.jpg) + +![](images/ad0d0af2e78cbc279915f970a0c7b569e8e1266abd92bfa0c2ffd7b330945e90.jpg) +(a) Annotation companies generate higher DSC scores than MTurk. $0 \leq \mathrm{DSC} \leq 1$ . + +![](images/eb8a72bf10a3bf83277f1a31f88a856424bec4c703a9f4cc685553377f49cb11.jpg) +(b) Annotation companies generate higher NSD scores than MTurk. $0 \leq \mathrm{NSD} \leq 1$ . +(c) Annotation companies generate a lower number of severe errors. +Fig. 7: Annotation companies generate higher quality annotations with their annotation pipeline than MTurk. This is observed for (a) the DSC, (b) the NSD, and (c) number of severe errors (lower is better). Annotation companies include a QA step, in contrast to crowdsourcing on MTurk. (a) and (b) displayed as dots and box plots (band indicates the median, box indicates the first (25th percentile) and third (75th percentile) quartiles). + +# Acknowledgements + +A part of this work was funded by M. Mengler and S. Funke - during their time at understand.ai -, and Helmholtz Imaging. T. Rädsch was supported by a scholarship from the Hanns Seidel Foundation with funds from the Federal Ministry of Education and Research Germany (BMBF). We thank M. Gelz, S. Strzysch and N. Kraft for their continuous support. + +# References + +1. Ali, S., Zhou, F., Braden, B., Bailey, A., Yang, S., Cheng, G., Zhang, P., Li, X., Kayser, M., Soberanis-Mukul, R.D., Albarqouni, S., Wang, X., Wang, C., Watanabe, S., Oksuz, I., Ning, Q., Yang, S., Khan, M.A., Gao, X.W., Realdon, S., Loshchenov, M., Schnabel, J.A., East, J.E., Wagnieres, G., Loschenov, V.B., Grisan, E., Daul, C., Blondel, W., Rittscher, J.: An objective comparison of detection and segmentation algorithms for artefacts in clinical endoscopy. Scientific Reports 10(1), 2748 (Feb 2020). https://doi.org/10.1038/s41598-020-59413-5, https://www.nature.com/articles/s41598-020-59413-5, number: 1 Publisher: Nature Publishing Group 5 +2. Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M.S., Bohg, J., Bosselut, A., Brunskill, E., Brynjolfsson, E., Buch, S., Card, D., Castellon, R., Chatterji, N., Chen, A., Creel, K., Davis, J.Q., Demszky, D., Donahue, C., Doumbouya, M., Durmus, E., Ermon, S., Etchemendy, J., Ethayarajh, K., Fei-Fei, L., Finn, C., Gale, T., Gillespie, L., Goel, K., Goodman, N., Grossman, S., Guha, N., Hashimoto, T., Henderson, P., Hewitt, J., Ho, D.E., Hong, J., Hsu, K., Huang, J., Icard, T., Jain, S., Jurafsky, D., Kalluri, P., Karamcheti, S., Keeling, G., Khani, F., Khattab, O., Koh, P.W., Krass, M., Krishna, R., Kuditipudi, R., Kumar, A., Ladhak, F., Lee, M., Lee, T., Leskovec, J., Levent, I., Li, X.L., Li, X., Ma, T., Malik, A., Manning, C.D., Mirchandani, S., Mitchell, E., Munyikwa, Z., Nair, S., Narayan, A., Narayanan, D., Newman, B., Nie, A., Niebles, J.C., Nilforoshan, H., Nyarko, J., Ogut, G., Orr, L., Papadimitriou, I., Park, J.S., Piech, C., Portelance, E., Potts, C., Raghunathan, A., Reich, R., Ren, H., Rong, F., Roohani, Y., Ruiz, C., Ryan, J., Re, C., Sadigh, D., Sagawa, S., Santhanam, K., Shih, A., Srinivasan, K., Tamkin, A., Taori, R., Thomas, A.W., Tramer, F., Wang, R.E., Wang, W., Wu, B., Wu, J., Wu, Y., Xie, S.M., Yasunaga, M., You, J., Zaharia, M., Zhang, M., Zhang, T., Zhang, X., Zhang, Y., Zheng, L., Zhou, K., Liang, P.: On the Opportunities and Risks of Foundation Models (Jul 2022). https://doi.org/10.48550/arXiv.2108.07258. http://arxiv.org/abs/2108.07258 13 +3. Bragg, J., Mausam, Weld, D.S.: Sprout: Crowd-Powered Task Design for Crowdsourcing. Proceedings of the ACM on User Interface Software and Technology (2018). https://doi.org/10.1145/3242587.3242598 4, 6 +4. Chang, J.C., Amershi, S., Kamar, E.: Revolt: Collaborative Crowdsourcing for Labeling Machine Learning Datasets. In: Proceedings of the 2017 CHI Conference on Human Factors in Computing Systems. pp. 2334-2346. CHI '17, Association for Computing Machinery, New York, NY, USA (May 2017). https://doi.org/10.1145/3025453.3026044 4 +5. Chen, L., Li, S., Yan, J., Wang, H., Gunaratna, K., Yadav, V., Tang, Z., Srinivasan, V., Zhou, T., Huang, H., Jin, H.: AlpaGasus: Training A Better Alpaca with Fewer Data (Nov 2023). https://doi.org/10.48550/arXiv.2307.08701, http://arxiv.org/abs/2307.08701 12 +6. Cheplygina, V., Perez-Rovira, A., Kuo, W., Tiddens, H.A.W.M., de Bruijne, M.: Early Experiences with Crowdsourcing Airway Annotations in Chest CT. In: Carneiro, G., Mateus, D., Peter, L., Bradley, A., Tavares, J.M.R.S., Belagiannis, V., Papa, J.P., Nascimento, J.C., Loog, M., Lu, Z., Cardoso, J.S., Cornebise, J. (eds.) Deep Learning and Data Labeling for Medical Applications. pp. 209-218. Lecture Notes in Computer Science, Springer International Publishing, Cham (2016). https://doi.org/10.1007/978-3-319-46976-8_22_4_6 + +7. Chmielewski, M., Kucker, S.C.: An MTurk Crisis? Shifts in Data Quality and the Impact on Study Results. Social Psychological and Personality Science 11(4), 464-473 (May 2020). https://doi.org/10.1177/1948550619875149, publisher: SAGE Publications Inc 4 +8. Cognilytica: Data engineering, preparation, and labeling for ai 2020 (2020), https://www.cognilytica.com/2020/01/31/data-preparation-labeling-for-ai-2020/10 +9. Créquet, P., Mansouri, G., Benchoufi, M., Vivot, A., Ravaud, P.: Mapping of Crowdsourcing in Health: Systematic Review. Journal of Medical Internet Research 20(5), 187 (2018). https://doi.org/10.2196/jmir.9330, https://www.jmir.org/2018/5/e187/1 +10. Daniel, F., Kucherbaev, P., Cappiello, C., Benatallah, B., Allahbakhsh, M.: Quality Control in Crowdsourcing: A Survey of Quality Attributes, Assessment Techniques and Assurance Actions. ACM Computing Surveys 51, 1-40 (Jan 2019). https://doi.org/10.1145/3148148 2, 4 +11. Dice, L.R.: Measures of the amount of ecologic association between species. Ecology 26(3), 297-302 (1945) 7 +12. Dow, S., Kulkarni, A., Klemmer, S., Hartmann, B.: Shepherding the crowd yields better work. In: Proceedings of the ACM 2012 conference on Computer Supported Cooperative Work. pp. 1013-1022. CSCW '12, Association for Computing Machinery, New York, NY, USA (Feb 2012). https://doi.org/10.1145/2145204.2145355 4 +13. Duggan, N.M., Jin, M., Mendicuti, M.A.D., Hallisey, S., Bernier, D., Selame, L.A., Asgari-Targhi, A., Fischetti, C.E., Lucassen, R., Samir, A.E., Duhaime+, E., Kapur, T., Goldsmith, A.J.: Gamified Crowdsourcing as a Novel Approach to Lung Ultrasound Dataset Labeling (Jun 2023). https://doi.org/10.48550/arXiv.2306.06773, http://arxiv.org/abs/2306.06773 4 +14. Gaikwad, S., Chhibber, N., Sehgal, V., Ballav, A., Mullings, C., Nasser, A., Richmond-Fuller, A., Gilbee, A., Gamage, D., Whiting, M., et al.: Prototype tasks: improving crowdsourcing results through rapid, iterative task design (2017). https://doi.org/10.48550/arXiv.1707.05645, https://arxiv.org/abs/1707.05645 4 +15. Halevy, A., Norvig, P., Pereira, F.: The unreasonable effectiveness of data. IEEE intelligent systems 24(2), 8-12 (2009). https://doi.org/10.1109/MIS.2009.36, ISBN: 1541-1672 Publisher: IEEE 1 +16. Han, D., Choe, J., Chun, S., Chung, J.J.Y., Chang, M., Yun, S., Song, J.Y., Oh, S.J.: Neglected free lunch-learning image classifiers using annotation byproducts (2023). https://doi.org/10.48550/arXiv.2303.17595, https://arxiv.org/abs/2303.17595 4 +17. Heim, E., Roß, T., Teil, A., März, K., Stieltjes, B., Eisenmann, M., Lebert, J., Metzger, J., Sommer, G., Sauter, A.W.: Large-scale medical image annotation with crowd-powered algorithms. Journal of Medical Imaging 5(3), 034002 (2018). https://doi.org/10.1117/1.JMI.5.3.034002 2, 4, 6 +18. Hettiachchi, D., Schaekermann, M., McKinney, T.J., Lease, M.: The Challenge of Variable Effort Crowdsourcing and How Visible Gold Can Help. Proceedings of the ACM on Human-Computer Interaction 5, 332:1-332:26 (Oct 2021). https://doi.org/10.1145/3476073 5 +19. Hutson, M., Kanzheleva, O., Taggart, C., Campana, B.J.L., Duong, Q.: Quality Control Challenges in Crowdsourcing Medical Labeling (2019) 4 + +20. Idrissi, B.Y., Bouchacourt, D., Balestriero, R., Evtimov, I., Hazirbas, C., Ballas, N., Vincent, P., Drozdjal, M., Lopez-Paz, D., Ibrahim, M.: ImageNet-X: Understanding Model Mistakes with Factor of Variation Annotations (Nov 2022). https://doi.org/10.48550/arXiv.2211.01866, http://arxiv.org/abs/2211.01866 3, 4 +21. Ipeirotis, P.G., Provost, F., Wang, J.: Quality management on Amazon Mechanical Turk. In: Proceedings of the ACM SIGKDD Workshop on Human Computation. pp. 64-67. ACM, Washington DC (Jul 2010). https://doi.org/10.1145/1837885.1837906, https://dl.acm.org/doi/10.1145/1837885.1837906 2, 4 +22. Irani, L.: The cultural work of microwork. New Media & Society 17(5), 720-739 (May 2015). https://doi.org/10.1177/1461444813511926, publisher: SAGE Publications 12 +23. Irshad, H., Montaser-Kouhsari, L., Waltz, G., Bucur, O., Nowak, J., Dong, F., Knoblauch, N., Beck, A.: Crowdsourcing image annotation for nucleus detection and segmentation in computational pathology: evaluating experts, automated methods, and the crowd. In: Biocomputing 2015. pp. 294-305. WORLD SCIENTIFIC, Kohala Coast, Hawaii, USA (Nov 2014). https://doi.org/10.1142/9789814644730_0029, http://www.worldscientific.com/doi/abs/10.1142/9789814644730_0029 4 +24. Joskowicz, L., Cohen, D., Caplan, N., Sosna, J.: Inter-observer variability of manual contour delineation of structures in CT. European Radiology 29(3), 1391-1399 (Mar 2019). https://doi.org/10.1007/s00330-018-5695-5, http://link.springer.com/10.1007/s00330-018-5695-5 3, 5 +25. Kentley, J., Weber, J., Liopyris, K., Braun, R.P., Marghoob, A.A., Quigley, E.A., Nelson, K., Prentice, K., Duhaime, E., Halpern, A.C., Rotemberg, V.: Agreement Between Experts and an Untrained Crowd for Identifying Dermoscopic Features Using a Gamified App: Reader Feasibility Study. JMIR Medical Informatics 11, e38412 (Jan 2023). https://doi.org/10.2196/38412, https://medinform.jmir.org/2023/1/e38412 4 +26. Kohli, M.D., Summers, R.M., Geis, J.R.: Medical image data and datasets in the era of machine learning—whitepaper from the 2016 C-MIMI meeting dataset session. Journal of digital imaging 30(4), 392–399 (2017). https://doi.org/10.1007/s10278-017-9976-3, publisher: Springer 5 +27. Le, J., Edmonds, A., Hester, V., Biewald, L.: Ensuring quality in crowdsourced search relevance evaluation: The effects of training question distribution. In: SIGIR 2010 workshop on crowdsourcing for search evaluation. vol. 2126, pp. 22-32 (2010) 2, 4 +28. Le Ludec, C., Cornet, M., Casilli, A.A.: The problem with annotation. Human labour and outsourcing between France and Madagascar. Big Data & Society 10(2), 20539517231188723 (Jul 2023). https://doi.org/10.1177/20539517231188723, http://journals.sagepub.com/doi/10.1177/20539517231188723 2 +29. Lee, A.N., Hunter, C.J., Ruiz, N.: Platypus: Quick, Cheap, and Powerful Refinement of LLMs (Aug 2023). https://doi.org/10.48550/arXiv.2308.07317, http://arxiv.org/abs/2308.07317 12 +30. Litman, L., Robinson, J., Rosenzweig, C.: The relationship between motivation, monetary compensation, and data quality among US- and India-based workers on Mechanical Turk. Behavior Research Methods 47(2), 519-528 (Jun 2015). https://doi.org/10.3758/s13428-014-0483-x, http://link.springer.com/10.3758/s13428-014-0483-x 6 + +31. Lu, X., Ratcliffe, D., Kao, T.T., Tikhonov, A., Litchfield, L., Rodger, C., Wang, K.: Rethinking Quality Assurance for Crowdsourced Multi-ROI Image Segmentation. Proceedings of the AAAI Conference on Human Computation and Crowdsourcing 11(1), 103–114 (Nov 2023). https://doi.org/10.1609/hcomp.v11i1.27552, https://ojs(aaai.org/index.php/HCOMP/article/view/27552, number: 1 2, 4, 5 +32. Maier-Hein, L., Mersmann, S., Kondermann, D., Bodenstedt, S., Sanchez, A., Stock, C., Kenngott, H.G., Eisenmann, M., Speidel, S.: Can masses of non-experts train highly accurate image classifiers? In: International conference on medical image computing and computer-assisted intervention. pp. 438-445. Springer (2014). https://doi.org/10.1007/978-3-319-10470-6_55 1 +33. Maier-Hein, L., Reinke, A., Godau, P., Tizabi, M.D., Buettner, F., Christodoulou, E., Glocker, B., Isensee, F., Kleesiek, J., Kozubek, M., et al.: Metrics reloaded: recommendations for image analysis validation. Nature Methods 21, 195–212 (2024). https://doi.org/10.1038/s41592-023-02151-z, https://www.nature.com/articles/s41592-023-02151-z 1 +34. Maier-Hein, L., Wagner, M., Ross, T., Reinke, A., Bodenstedt, S., Full, P.M., Hempe, H., Mindroc-Filimon, D., Scholz, P., Tran, T.N., Bruno, P., Kisilenko, A., Müller, B., Davitashvili, T., Capek, M., Tizabi, M.D., Eisenmann, M., Adler, T.J., Grohl, J., Schellenberg, M., Seidlitz, S., Lai, T.Y.E., Pekdemir, B., Roethlingshoefer, V., Both, F., Bittel, S., Mengler, M., Mündermann, L., Apitz, M., Kopp-Schneider, A., Speidel, S., Nickel, F., Probst, P., Kenngott, H.G., Müller-Stich, B.P.: Heidelberg colorectal data set for surgical data science in the sensor operating room. Scientific Data 8(1), 101 (Apr 2021). https://doi.org/10.1038/s41597-021-00882-2, https://www.nature.com/articles/s41597-021-00882-2, number: 1 Publisher: Nature Publishing Group 5 +35. McCulloch, C.E., Searle, S.R.: Generalized, linear, and mixed models. John Wiley & Sons (2004). https://doi.org/10.1002/0471722073, https://books.google.com/books?hl=en&lr=&id=bWDPukohugQC&oi=fnd&pg=PR6&dq=Generalized, +Linear, +and+Mixed+Models.&ots=sSy3HM7CGo&sig=Ss04EHKG_xkcZZCYWSEJl6paHos 8 +36. Miceli, M., Posada, J.: The Data-Production Dispositif. Proceedings of the ACM on Human-Computer Interaction 6(CSCW2), 1-37 (Nov 2022). https://doi.org/10.1145/3555561, https://dl.acm.org/doi/10.1145/3555561 2, 12, 14 +37. Michaelis, C., Mitzkus, B., Geirhos, R., Rusak, E., Bringmann, O., Ecker, A.S., Bethge, M., Brendel, W.: Benchmarking Robustness in Object Detection: Autonomous Driving when Winter is Coming (Mar 2020). https://doi.org/10.48550/arXiv.1907.07484, http://arxiv.org/abs/1907.07484 5 +38. Petrovic, N., Moya-Alcover, G., Varona, J., Jaume-i Capó, A.: Crowdsourcing human-based computation for medical image analysis: A systematic literature review. Health Informatics Journal 26(4), 2446-2469 (Dec 2020). https://doi.org/10.1177/1460458220907435, http://journals.sagepub.com/doi/10.1177/1460458220907435 4 +39. Reinke, A., Tizabi, M.D., Baumgartner, M., Eisenmann, M., Heckmann-Nötzel, D., Kavur, A.E., Rädsch, T., Sudre, C.H., Acion, L., Antonelli, M.: Understanding metric-related pitfalls in image analysis validation. Nature methods pp. 1-13 (2024). https://doi.org/doi.org/10.1038/s41592-023-02150-0, https://www.nature.com/articles/s41592-023-02150-0, publisher: Nature Publishing Group 7 + +40. Research, G.V.: Data Collection And Labeling Market Size & Share Report 2030 (2022), https://www.grandviewresearch.com/industry-analysis/data-collection-labeling-market-3,11 +41. Rofs, T., Bruno, P., Reinke, A., Wiesenfarth, M., Koeppel, L., Full, P.M., Pekdemir, B., Godau, P., Trofimova, D., Isensee, F.: How can we learn (more) from challenges? A statistical approach to driving future algorithm development. arXiv preprint arXiv:2106.09302 (2021). https://doi.org/10.48550/arXiv.2106.09302, https://arxiv.org/abs/2106.09302 3, 4 +42. Rädsch, T., Reinke, A., Weru, V., Tizabi, M.D., Schreck, N., Kavur, A.E., Pekdemir, B., Roß, T., Kopp-Schneider, A., Maier-Hein, L.: Labelling instructions matter in biomedical image analysis. Nature Machine Intelligence 5(3), 273–283 (Mar 2023). https://doi.org/10.1038/s42256-023-00625-5, https://www.nature.com/articles/s42256-023-00625-5, number: 3 Publisher: Nature Publishing Group 2, 4 +43. Sambasivan, N., Kapania, S., Highfill, H., Akrong, D., Paritosh, P., Aroyo, L.M.: "Everyone wants to do the model work, not the data work": Data Cascades in High-Stakes AI. In: Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. pp. 1-15. ACM, Yokohama Japan (May 2021). https://doi.org/10.1145/3411764.3445518, https://dl.acm.org/doi/10.1145/3411764.3445518 1, 13 +44. Schmidt, F.A.: The Planetary Stacking Order of Multilayered Crowd-AI Systems. In: Graham, M., Ferrari, F. (eds.) Digital Work in the Planetary Market, pp. 137-156. The MIT Press (Jun 2022). https://doi.org/10.7551/mitpress/13835.003.0012, https://direct.mit.edu/books/book/5319/chapter/3800159/The-Planetary-Stacking-Order-of-Multilayered-Crowd 2 +45. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: LLaMA: Open and Efficient Foundation Language Models (Feb 2023). https://doi.org/10.48550/arXiv.2302.13971, http://arxiv.org/abs/2302.13971 12 +46. Veselovsky, V., Ribeiro, M.H., West, R.: Artificial Artificial Artificial Intelligence: Crowd Workers Widely Use Large Language Models for Text Production Tasks (Jun 2023). https://doi.org/10.48550/arXiv.2306.07899, http://arxiv.org/abs/2306.07899 13 +47. Wang, K., Hill, M., Knowles-Barley, S., Tikhonov, A., Litchfield, L., Bare, J.C.: Improving Segmentation of Breast Arterial Calcifications from Digital Mammography: Good Annotation is All You Need. In: Zheng, Y., Keles, H.Y., Koniusz, P. (eds.) Computer Vision - ACCV 2022 Workshops, vol. 13848, pp. 134-150. Springer Nature Switzerland, Cham (2023). https://doi.org/10.1007/978-3-031-27066-6_10, https://link.springer.com/10.1007/978-3-031-27066-6_10, series Title: Lecture Notes in Computer Science 2, 4 +48. Yu, S., Zhou, H.Y., Ma, K., Bian, C., Chu, C., Liu, H., Zheng, Y.: Difficulty-Aware Glaucoma Classification with Multi-rater Consensus Modeling. In: Martel, A.L., Abolmaesumi, P., Stoyanov, D., Mateus, D., Zuluaga, M.A., Zhou, S.K., Racoceanu, D., Joskowicz, L. (eds.) Medical Image Computing and Computer Assisted Intervention - MICCAI 2020. pp. 741-750. Lecture Notes in Computer Science, Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-59710-8_72_4 +49. Ørting, S.N., Doyle, A., Hilten, A.v., Hirth, M., Inel, O., Madan, C.R., Mavridis, P., Spiers, H., Cheplygina, V.: A Survey of Crowdsourcing in Medical Image Analysis. + +Human Computation 7, 1-26 (Dec 2020). https://doi.org/10.15346/hc.v7i1.1, https://hcjournal.org/index.php/jhc/article/view/111 1 \ No newline at end of file diff --git a/2024/Quality Assured_ Rethinking Annotation Strategies in Imaging AI/images.zip b/2024/Quality Assured_ Rethinking Annotation Strategies in Imaging AI/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2212b3e9487c0f8f5bbd2476a22d489760183b27 --- /dev/null +++ b/2024/Quality Assured_ Rethinking Annotation Strategies in Imaging AI/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:99222eafce2b44bd2734600b1429a379993d62ccbc511858173a54bbe965b32f +size 378801 diff --git a/2024/Quality Assured_ Rethinking Annotation Strategies in Imaging AI/layout.json b/2024/Quality Assured_ Rethinking Annotation Strategies in Imaging AI/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1e3a719bddaffe872798d49aba621df499a644fc --- /dev/null +++ b/2024/Quality Assured_ Rethinking Annotation Strategies in Imaging AI/layout.json @@ -0,0 +1,7531 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 165, + 111, + 451, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 111, + 451, + 148 + ], + "spans": [ + { + "bbox": [ + 165, + 111, + 451, + 148 + ], + "type": "text", + "content": "Quality Assured: Rethinking Annotation Strategies in Imaging AI" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 135, + 173, + 478, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 173, + 478, + 209 + ], + "spans": [ + { + "bbox": [ + 135, + 173, + 478, + 209 + ], + "type": "text", + "content": "Tim Rädsch1,2,3, Annika Reinke1,2, Vivienne Weru1, Minu D. Tizabi1, Nicholas Heller4, Fabian Isensee1,2, Annette Kopp-Schneider1*, and Lena Maier-Hein1,2,3,5*" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 144, + 219, + 471, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 219, + 471, + 254 + ], + "spans": [ + { + "bbox": [ + 144, + 219, + 471, + 254 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 144, + 219, + 471, + 254 + ], + "type": "text", + "content": "German Cancer Research Center, GER " + }, + { + "bbox": [ + 144, + 219, + 471, + 254 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 144, + 219, + 471, + 254 + ], + "type": "text", + "content": "Helmholtz Imaging, GER " + }, + { + "bbox": [ + 144, + 219, + 471, + 254 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 144, + 219, + 471, + 254 + ], + "type": "text", + "content": "Heidelberg University, GER " + }, + { + "bbox": [ + 144, + 219, + 471, + 254 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 144, + 219, + 471, + 254 + ], + "type": "text", + "content": "University of Minnesota, USA " + }, + { + "bbox": [ + 144, + 219, + 471, + 254 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 144, + 219, + 471, + 254 + ], + "type": "text", + "content": "National Center for Tumor Diseases, GER " + }, + { + "bbox": [ + 144, + 219, + 471, + 254 + ], + "type": "inline_equation", + "content": "\\boxed{\\boxtimes}" + }, + { + "bbox": [ + 144, + 219, + 471, + 254 + ], + "type": "text", + "content": ":{tim.raedsch; l.maier-hein}@dkfz-heidelberg.de" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 273, + 455, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 273, + 455, + 494 + ], + "spans": [ + { + "bbox": [ + 160, + 273, + 455, + 494 + ], + "type": "text", + "content": "Abstract. This paper does not describe a novel method. Instead, it studies an essential foundation for reliable benchmarking and ultimately real-world application of AI-based image analysis: generating high-quality reference annotations. Previous research has focused on crowdsourcing as a means of outsourcing annotations. However, little attention has so far been given to annotation companies, specifically regarding their internal quality assurance (QA) processes. Therefore, our aim is to evaluate the influence of QA employed by annotation companies on annotation quality and devise methodologies for maximizing data annotation efficacy. Based on a total of 57,648 instance segmented images obtained from a total of 924 annotators and 34 QA workers from four annotation companies and Amazon Mechanical Turk (MTurk), we derived the following insights: (1) Annotation companies perform better both in terms of quantity and quality compared to the widely used platform MTurk. (2) Annotation companies' internal QA only provides marginal improvements, if any. However, improving labeling instructions instead of investing in QA can substantially boost annotation performance. (3) The benefit of internal QA depends on specific image characteristics. Our work could enable researchers to derive substantially more value from a fixed annotation budget and change the way annotation companies conduct internal QA." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 502, + 438, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 502, + 438, + 514 + ], + "spans": [ + { + "bbox": [ + 160, + 502, + 438, + 514 + ], + "type": "text", + "content": "Keywords: Data Annotation " + }, + { + "bbox": [ + 160, + 502, + 438, + 514 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 502, + 438, + 514 + ], + "type": "text", + "content": " Quality Assurance " + }, + { + "bbox": [ + 160, + 502, + 438, + 514 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 502, + 438, + 514 + ], + "type": "text", + "content": " Crowdsourcing" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 530, + 230, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 530, + 230, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 530, + 230, + 544 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 552, + 482, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 482, + 650 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 482, + 650 + ], + "type": "text", + "content": "Following a long period of extensive focus on pure method development, proper benchmarking and validation have recently been recognized as a major challenge in AI-based image analysis [33, 43]. In this context, high-quality annotation of reference datasets is a fundamental requirement to ensure the meaningful validation of image analysis algorithms [15]. As annotation tasks may be cumbersome, crowdsourcing is often used to partially outsource image annotation. Traditionally, such outsourcing is conducted via crowdsourcing platforms [32], with Amazon Mechanical Turk (MTurk)1 as the predominant choice [9, 49]." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 329, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 329, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 329, + 665 + ], + "type": "text", + "content": "1 https://www.mturk.com * shared last author" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 113, + 482, + 228 + ], + "blocks": [ + { + "bbox": [ + 133, + 113, + 482, + 228 + ], + "lines": [ + { + "bbox": [ + 133, + 113, + 482, + 228 + ], + "spans": [ + { + "bbox": [ + 133, + 113, + 482, + 228 + ], + "type": "image", + "image_path": "061f4698a80d451656990b78fc9c077e4110fb5d60a98b97de5f2fb5c2e4a271.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 236, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 304 + ], + "type": "text", + "content": "Fig. 1: Research Questions (RQs) tackled in this work. Based on 57,648 instance segmentation masks annotated by 924 annotators and 34 quality assurance (QA) workers from five different annotation providers, we (1) compared the effectiveness of generating high-quality annotations between annotation companies and Amazon Mechanical Turk and (2) investigated the effects of annotation companies' internal QA and (3) real-world image characteristics on the annotation quality." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 337, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 337, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 337, + 482, + 517 + ], + "type": "text", + "content": "Recent literature has reflected the rise of annotation companies as a new type of provider with potentially better annotation performance compared to established crowdsourcing platforms such as MTurk [28,36,42,44]. Annotation companies differ substantially from crowdsourcing platforms. First, they employ annotators directly and at shared office spaces [28,44], enabling better external control of worker conditions and stricter data security measures, while crowdsourcing platforms can be accessed by anyone, without specific knowledge or training being required. Second, annotators in annotation companies typically work in small teams, with an experienced quality assurance (QA) worker being responsible for the team's output. In the observed internal QA, QA workers make adjustments to the annotations generated by the team by reviewing each individual annotation and modifying those where they deem changes necessary. As a result, the annotation requester receives the quality-assured annotation, not the initial one (see Fig. 2a). In contrast, crowdsourcing platforms usually do not employ internal QA mechanics similar to annotation companies." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "The importance of annotation providers for real-world applications is evident at the major computer vision conferences, where over the past two years, approximately " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " of exhibitors were annotation providers (see Fig. 2b). While QA as a post-processing step for annotations from platforms such as MTurk has been researched (e.g. [10, 17, 21, 27]), little is known about the benefit of internal QA on annotation quality within annotation companies. The only related work that we are aware of pertains to external QA from the perspective of an annotation company. Unlike in internal QA, in the case of external QA, domain experts or experienced annotators working on behalf of the requester check the annotation quality as a post-processing step of the received annotations [31, 47]. Our paper, on the other hand, focuses on the internal QA processes of an annotation company." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "T. Rädsch et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 245 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 245 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 245 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 245 + ], + "type": "image", + "image_path": "08608e7b3b91ceaa4f4d71b9628e1435ebf43201ebda0d9b769d31413438f963.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 253, + 482, + 319 + ], + "lines": [ + { + "bbox": [ + 130, + 253, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 253, + 482, + 319 + ], + "type": "text", + "content": "Fig. 2: a) In contrast to Amazon Mechanical Turk (MTurk), annotation companies commonly perform quality assurance (QA) before delivering the result to the annotation requester. b) Annotation providers are an integral part of the computer vision community. They account for around " + }, + { + "bbox": [ + 130, + 253, + 482, + 319 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 253, + 482, + 319 + ], + "type": "text", + "content": " of exhibitors present at major computer vision conferences (ECCV 2022: " + }, + { + "bbox": [ + 130, + 253, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\mathrm{n} = 40" + }, + { + "bbox": [ + 130, + 253, + 482, + 319 + ], + "type": "text", + "content": ", CVPR 2022: " + }, + { + "bbox": [ + 130, + 253, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\mathrm{n} = 97" + }, + { + "bbox": [ + 130, + 253, + 482, + 319 + ], + "type": "text", + "content": ", ICCV 2023: " + }, + { + "bbox": [ + 130, + 253, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\mathrm{n} = 43" + }, + { + "bbox": [ + 130, + 253, + 482, + 319 + ], + "type": "text", + "content": ", CVPR 2023: " + }, + { + "bbox": [ + 130, + 253, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\mathrm{n} = 104" + }, + { + "bbox": [ + 130, + 253, + 482, + 319 + ], + "type": "text", + "content": ") in the last two years." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 347, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 407 + ], + "type": "text", + "content": "In light of the vast sums of money to be spent on QA [40], the purpose of this paper is to shed light on whether annotation companies' internal QA benefits the annotation quality. Based on more than 50,000 instance segmentation masks annotated by more than 900 annotators from five different annotation providers, we investigate the following research questions (RQs; see Fig. 1):" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 410, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 482, + 517 + ], + "type": "text", + "content": "RQ 1) How does the choice of annotation provider influence the quality of annotations? As the selection of the annotation provider is among the first and key decisions in any annotation project after data collection and generation of labeling instructions, we first investigated companies' effectiveness in generating annotations, which is particularly relevant in cases of constrained annotation budgets requiring a cost-effective strategy. As both quantity and quality of annotated images for a given budget are relevant in provider selection, we compared the annotation quality and costs of five different providers, including MTurk and four annotation companies, on identical images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 520, + 482, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 520, + 482, + 603 + ], + "spans": [ + { + "bbox": [ + 130, + 520, + 482, + 603 + ], + "type": "text", + "content": "RQ 2) Does annotation companies' quality assurance improve annotation quality? A major decision in the design of an annotation process concerns the allocation of resources. A common perception is that internal QA of annotation companies, following the initial annotation, substantially enhances the quality. Here, we challenged this assumption by investigating the improvement in annotation quality from internal QA compared to that achieved through enhanced labeling instructions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "RQ 3) Should quality assurance focus on specific subsets of images? In theory, the efficacy of QA may be contingent upon specific image characteristics [20, 24, 41]. For instance, underexposed images might benefit more from a secondary review compared to those with optimal illumination. Concurrently, images that present annotation challenges for human observers might be partic" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 178, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 178, + 91, + 448, + 102 + ], + "type": "text", + "content": "Quality Assured: Rethinking Annotation Strategies in Imaging AI" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "ularly pertinent for the training and validation of algorithms [20, 41]. Here, we investigated whether and to what extent the benefit of QA depends on specific image characteristics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 167, + 238, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 167, + 238, + 179 + ], + "spans": [ + { + "bbox": [ + 131, + 167, + 238, + 179 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 188, + 342, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 188, + 342, + 201 + ], + "spans": [ + { + "bbox": [ + 131, + 188, + 342, + 201 + ], + "type": "text", + "content": "2.1 Comparison of annotation providers" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 205, + 482, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 324 + ], + "type": "text", + "content": "With respect to the quality of annotations obtained from varying annotation providers, the closest work conducted are comparisons between crowdsourcing platforms and domain experts. Irshad et al. [23] and Petrovic et al. [38], among others, have shown that crowdsourced annotations can reach the annotation quality of domain experts in safety-critical applications. In recent literature, Duggan et al. [13] and Kentley et al. [25] compared domain-specialized crowds as a new type of annotation provider and found comparable performance to domain experts. Furthermore, Rädsch et al. [42] provided a comparison of annotation companies to crowdsourcing, but focused solely on raw annotation quality without QA and without relating it to annotation budgets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 324, + 482, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 482, + 397 + ], + "type": "text", + "content": "Interestingly, researchers report concerns about the annotation quality obtained from MTurk [31], such as obtaining unusable annotations in up to " + }, + { + "bbox": [ + 130, + 324, + 482, + 397 + ], + "type": "inline_equation", + "content": "67.8\\%" + }, + { + "bbox": [ + 130, + 324, + 482, + 397 + ], + "type": "text", + "content": " of cases [6]. Other research fields, such as psychology, even report a \"substantial decrease in data quality\" [7]. However, we are not aware of work that has systematically compared the performance between annotation companies and crowdsourcing platforms such as MTurk while considering annotation budgets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 410, + 251, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 410, + 251, + 422 + ], + "spans": [ + { + "bbox": [ + 131, + 410, + 251, + 422 + ], + "type": "text", + "content": "2.2 Quality assurance" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 426, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 582 + ], + "type": "text", + "content": "The main goal of QA is to enhance and ensure the quality of annotations. A comprehensive overview on QA in crowdsourcing platforms is provided by Daniel et al. [10]. However, their research does not include annotation companies. During the annotation process, quality control techniques generally fall into three categories [31]: (1) comparing to a gold standard with pre-defined reference annotations (e.g., [27]), (2) analyzing combined outputs from multiple annotators (e.g., [21]), and (3) investigating annotator behaviors (e.g., [17, 19]). Other approaches to enhancing annotation quality on crowdsourcing platforms target training of crowdworkers [12] and iterative refinement of the annotation environment [3,4,14]. It should be noted that annotation byproducts, such as mouse movement during annotation generation [16, 17] or inter-rater agreement [48], can be used to boost model performance, by incorporating this information in the training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "The closest related work focusing on the impact of QA in annotation companies was introduced by Wang et al. [47] and Lu et al. [31]. However, their work focuses on external QA conducted by experienced personnel from the requester rather than internal QA, and features a substantially smaller number of annotators from annotation companies. To our knowledge, a systematic investigation into the effects of internal QA processes within annotation companies is thus, so far, lacking." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "T. Rädsch et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 129 + ], + "type": "text", + "content": "2.3 Impact of real-world image characteristics on annotation quality" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 245 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 245 + ], + "type": "text", + "content": "Related research analyzing the impact of real-world image characteristics highlights their effect on annotations created by machine learning models and annotators alike. Michaelis et al. [37] demonstrated that models easily encounter failures with increasingly challenging real-world image characteristics, such as added snow or poor lighting characteristics. Similarly, annotators generate less reliable annotations in the presence of challenging real-world image characteristics. This effect can be observed both within [1,24,31] and across domains [18,26]. However, a systematic analysis of the impact of such characteristics on internal QA employed by annotation companies is lacking." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 263, + 207, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 207, + 275 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 207, + 275 + ], + "type": "text", + "content": "3 Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 289, + 481, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 481, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 481, + 324 + ], + "type": "text", + "content": "In the following, we describe the selected data, the labeling instructions, annotation providers and annotators, the QA process, and the experimental design of our work." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 343, + 186, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 186, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 186, + 354 + ], + "type": "text", + "content": "3.1 Data" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 363, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 472 + ], + "type": "text", + "content": "We conducted our experiments on the instance segmentation data from the Heidelberg Colorectal (HeiCo) dataset for surgical data science in the sensor operating room [34], representing a safety-critical application that crowdsourcing platforms can still handle. The task involved segmenting surgical instruments from laparoscopic videos, resulting in 57,648 instance segmentation masks on 4,050 unique frames. Furthermore, we used 57,636 metadata annotations on the unique images to quantify the effect of varying real-world image characteristics (see RQ3). Characteristics included motion blur, underexposure, object occlusion, or overlapping objects, among others, and are illustrated in Suppl. A." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 489, + 266, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 489, + 266, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 489, + 266, + 502 + ], + "type": "text", + "content": "3.2 Labeling instructions" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "text", + "content": "To investigate the benefit of QA compared to an investment in detailed labeling instructions, we created three distinct types of labeling instructions with varying levels of information density, namely, (1) minimal text, (2) extended text, and (3) extended text including exemplary pictures. Each subsequent labeling instruction type was structured to contain more detailed information than the preceding. To ensure accurate labeling of the imaging data, we established a collection of design principles applicable to all labeling instructions:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 594, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 131, + 594, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 594, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 131, + 594, + 481, + 616 + ], + "type": "text", + "content": "(1) Labeling information is presented in a slide format to enhance the processing of information by humans." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 617, + 481, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 481, + 653 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 481, + 653 + ], + "type": "text", + "content": "(2) Each slide contains a distinct 'chunk' of information, which is a self-contained piece of content. This approach of 'chunking' information lessens the cognitive load on the working memory." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 653, + 469, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 653, + 469, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 653, + 469, + 665 + ], + "type": "text", + "content": "(3) Related information chunks are placed in close proximity to one another." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 178, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 178, + 91, + 447, + 103 + ], + "type": "text", + "content": "Quality Assured: Rethinking Annotation Strategies in Imaging AI" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "(4) A uniform design template is used throughout, featuring specific fonts, symbols, and color schemes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 140, + 343, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 140, + 343, + 152 + ], + "spans": [ + { + "bbox": [ + 131, + 140, + 343, + 152 + ], + "type": "text", + "content": "The three labeling instructions differ as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 152, + 480, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 212 + ], + "type": "text", + "content": "Minimal text labeling instructions: These instructions provide a concise textual outline, offering examples of typical annotations and highlighting the most frequent annotation types (see Suppl. B.1). They reflect a scenario with minimal effort invested in developing the instructions. The basic framework includes seven slides totaling 168 words." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 212, + 481, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 481, + 270 + ], + "type": "text", + "content": "Extended text labeling instructions: Building upon the simplified version, the extended text instructions offer a thorough textual explanation, supplemented by both positive and negative examples (refer to Suppl. B.2). They cover both regular and some rare annotation scenarios. The enhanced version comprises ten slides with a total of 446 words." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 272, + 482, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 356 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 356 + ], + "type": "text", + "content": "Extended text including pictures labeling instructions: These instructions augment the enhanced textual guidelines with images (detailed in Suppl. B.3). The visual aids feature descriptions, symbols, annotations, and color coding to effectively communicate the details on the slides. They also include illustrations of rare annotation cases, indicating a deep and well-documented understanding of the labeling process. This set contains 16 slides with a total of 961 words. We refer the reader to Suppl. B for full details of each labeling instruction type." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 372, + 269, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 372, + 269, + 384 + ], + "spans": [ + { + "bbox": [ + 131, + 372, + 269, + 384 + ], + "type": "text", + "content": "3.3 Annotation providers" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 390, + 480, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 480, + 437 + ], + "type": "text", + "content": "The study was conducted based on 216 annotators and 34 QA workers from four annotation companies and 708 crowdworkers from the crowdsourcing platform MTurk. Our goal was to provide the best representation for both annotation provider types." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 438, + 481, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 481, + 593 + ], + "type": "text", + "content": "Consequently, we established higher standards for our MTurk crowdworkers compared to the usual research benchmarks, which generally require a " + }, + { + "bbox": [ + 130, + 438, + 481, + 593 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 438, + 481, + 593 + ], + "type": "text", + "content": " acceptance rate of Human Intelligence Tasks (HITs) and at least 100 completed tasks [3, 6, 17]. Our criteria included a " + }, + { + "bbox": [ + 130, + 438, + 481, + 593 + ], + "type": "inline_equation", + "content": "98\\%" + }, + { + "bbox": [ + 130, + 438, + 481, + 593 + ], + "type": "text", + "content": " acceptance rate with over 5,000 completed HITs. We spread out the HITs over 40 days at all times to obtain a representative MTurk sample and compensated workers fairly, as recommended by Litman et al. [30]. The annotation companies we selected have a solid background in handling large-scale annotation tasks in a variety of domains and operate internationally. The annotators were allocated by the companies themselves to replicate the typical workflow of a project carried out by an annotation company. Each annotator was linked to a specific annotation provider that had been predetermined. No selection process from our side took place for either the annotators, QA workers, or the MTurk crowdworkers." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": "To prevent information leakage, we conducted the labeling process with increasingly detailed labeling instructions (minimal text instruction, extended text instruction, and extended text including pictures instruction). Each annotator worked exclusively with one of the three instruction types as introduced in Sec. 3.2. To enhance security, we included a mandatory minimum gap of ten days between consecutive labeling instructions." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "text", + "content": "T. Rädsch et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 291, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 291, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 291, + 129 + ], + "type": "text", + "content": "3.4 Quality assurance process" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 175 + ], + "type": "text", + "content": "In the internal QA process, annotators typically report to a small number of experienced QA workers, who are often responsible for training their team and, in our case, responsible for the QA of the annotations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 177, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 177, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 177, + 482, + 344 + ], + "type": "text", + "content": "The general workflow for all companies consisted of an annotation stage, in which images were assigned randomly to each annotator, and a QA stage. During the annotation stage, annotators received raw images with the task of annotating them until they considered the quality of annotations satisfactory. Annotators were aware that their work would be checked by QA. Once submitted, the QA worker gained access to the generated annotations, allowing them to make modifications until they were satisfied before submitting the final annotations for delivery to the requester. Throughout this process, both annotators and QA workers retained access to the respective labeling instructions at any given moment. We obtained access to the annotations both before and after QA through a software company handling a substantial number of annotation projects annually. The experiments were seamlessly integrated into the software company's pipeline of annotation projects, reducing the risk of exposing our experiments. Pricing was established to be representative of other projects of similar size." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 344, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 482, + 392 + ], + "type": "text", + "content": "To maintain fairness and avoid giving any annotation provider a potential informational advantage that may have affected our statistical analysis, no specific queries from the annotation companies about the content of the labeling instructions were addressed." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 414, + 264, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 414, + 264, + 426 + ], + "spans": [ + { + "bbox": [ + 131, + 414, + 264, + 426 + ], + "type": "text", + "content": "3.5 Experimental design" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "text", + "content": "RQ 1: Each of the five labeling providers annotated the same images with each set of labeling instructions, employing separate annotators for each task. We analyzed the number of spam images, images with severe errors, i.e., images with at least one false positive and/or false negative, and the overall annotation quality with the Dice Similarity Coefficient (DSC) [11] as the overlap-based metric and the Normalized Surface Distance (NSD) [39] as the distance-based metric. We define a spam annotation as a purposefully poor-quality annotation with the intent to obtain the payment without properly conducting the annotation task. Annotation duration was used as an auxiliary measure, where accessible. For instance, if an annotator completed the task in only 17 seconds, they could not have read the instructions and properly performed annotations on several images. Of note, a spammer could wait a relevant number of seconds and then enter a poor-quality annotation to bypass the annotation duration check. Thus, each image was additionally visually inspected by an engineer with extensive labeling experience. Spam annotation examples are displayed in Suppl. D." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "RQ 2: The annotation companies conducted their internal QA processes directly after their annotators completed the annotation job, as explained in Sec. 3.4 and Fig. 2a. The experiments began with only providing the minimal text labeling instructions, followed by the extended text labeling instructions," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 178, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 178, + 91, + 447, + 103 + ], + "type": "text", + "content": "Quality Assured: Rethinking Annotation Strategies in Imaging AI" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 114, + 449, + 228 + ], + "blocks": [ + { + "bbox": [ + 167, + 114, + 449, + 228 + ], + "lines": [ + { + "bbox": [ + 167, + 114, + 449, + 228 + ], + "spans": [ + { + "bbox": [ + 167, + 114, + 449, + 228 + ], + "type": "image", + "image_path": "bfbdf51154c1873a74b129cde29ab08df2292fbbea703a12bc8b83545e5fdd59.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 235, + 482, + 291 + ], + "lines": [ + { + "bbox": [ + 130, + 235, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 235, + 482, + 291 + ], + "type": "text", + "content": "Fig. 3: Annotation companies are more efficient at generating high-quality images than Amazon Mechanical Turk (MTurk). Percentage of images without severe errors (green), images with severe errors (orange), and spam (red) obtained for the five different annotation providers investigated. Annotation costs per image were lower for all companies compared to MTurk (median: " + }, + { + "bbox": [ + 130, + 235, + 482, + 291 + ], + "type": "inline_equation", + "content": "61.1\\%" + }, + { + "bbox": [ + 130, + 235, + 482, + 291 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 314, + 480, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 480, + 337 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 480, + 337 + ], + "type": "text", + "content": "and finally, the extended text instructions with pictures, to prevent information leakage." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 338, + 480, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 480, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 480, + 421 + ], + "type": "text", + "content": "For each annotated image, we calculated the DSC and NSD per instance, aggregated the scores per annotated image, and aggregated the scores over annotation company and labeling instruction. We then compared the annotation quality without and with QA across the three types of labeling instructions for the four annotation companies by computing both absolute and relative changes in DSC/NSD aggregates, the percentage of modified images, and the change in the percentage of severe errors, among other parameters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 422, + 481, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 481, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 481, + 506 + ], + "type": "text", + "content": "RQ 3: In contrast to MTurk, the four annotation companies conducted an additional annotation round, including QA, with the extended text including pictures labeling instructions to quantify the impact of different real-world image characteristics, represented by 57,636 metadata annotations, on the probability of improvement. The data was analyzed with a logistic mixed model [35] which included random effects for the image, the annotator, and the annotation company. Further method details are provided in Suppl. C." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 524, + 200, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 524, + 200, + 536 + ], + "spans": [ + { + "bbox": [ + 131, + 524, + 200, + 536 + ], + "type": "text", + "content": "4 Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 549, + 453, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 549, + 453, + 574 + ], + "spans": [ + { + "bbox": [ + 131, + 549, + 453, + 574 + ], + "type": "text", + "content": "4.1 How does the choice of annotation provider influence the quality and cost of annotations?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "Annotated images generated by the annotation companies were of higher quality, comprising no spam (MTurk: " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\sim 20\\%" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " ) and a much higher proportion of remaining images without severe annotation errors (factor of " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": ">2" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " improvement for all companies compared to MTurk), as shown in Fig. 3. In addition, the annotation companies were " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "61.1\\%" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " cheaper in the median, despite their additional employment of QA workers. More specifically, the costs ranged from " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "24.5\\%" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "81.5\\%" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " of the MTurk costs for the same images and number of annotations. We have" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "T. Rädsch et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 123, + 272, + 220 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 143, + 121 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 143, + 121 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 143, + 121 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 123, + 272, + 220 + ], + "lines": [ + { + "bbox": [ + 132, + 123, + 272, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 123, + 272, + 220 + ], + "type": "image", + "image_path": "fe018066ac0287db08f641d2088b2f831e466d2d8629b14fe815d48281c48e45.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 238, + 271, + 255 + ], + "lines": [ + { + "bbox": [ + 144, + 238, + 271, + 255 + ], + "spans": [ + { + "bbox": [ + 144, + 238, + 271, + 255 + ], + "type": "text", + "content": "Annotation company Company 1 Company 2 Company 3 Company 4" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 130, + 262, + 480, + 328 + ], + "lines": [ + { + "bbox": [ + 130, + 262, + 480, + 328 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 480, + 328 + ], + "type": "text", + "content": "Fig. 4: QA of annotation companies only provides marginal improvements, if any. These depend highly on the company and type of labeling instruction. a: Absolute change of the median DSC resulting from performing QA. For each company and labeling instruction, the median was obtained by aggregating the DSC scores over the images. b: Corresponding absolute values of the DSC. Positive and negative changes from annotation to QA are colored according to the color scale." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 282, + 123, + 481, + 239 + ], + "blocks": [ + { + "bbox": [ + 281, + 114, + 291, + 121 + ], + "lines": [ + { + "bbox": [ + 281, + 114, + 291, + 121 + ], + "spans": [ + { + "bbox": [ + 281, + 114, + 291, + 121 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 282, + 123, + 481, + 239 + ], + "lines": [ + { + "bbox": [ + 282, + 123, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 282, + 123, + 481, + 239 + ], + "type": "table", + "html": "
CompanyLabeling instruction25th PercentileMedian75th Percentile
AnnotationQAAnnotationQAAnnotationQA
11) Min. text0.610.620.930.930.960.96
2) Ext. text0.580.580.930.930.960.96
3) Ext. text + pictures0.900.900.950.950.960.96
21) Min. text0.570.470.920.900.960.95
2) Ext. text0.620.840.930.940.960.96
3) Ext. text + pictures0.850.880.940.940.960.96
31) Min. text0.600.600.930.930.960.96
2) Ext. text0.470.640.910.930.960.96
3) Ext. text + pictures0.740.850.930.940.960.96
41) Min. text0.630.630.940.940.960.96
2) Ext. text0.880.880.940.940.960.96
3) Ext. text + pictures0.910.910.950.950.970.97
", + "image_path": "70135a47f093ad8a5b103dfad0fde1b89700ec9c33ac790fa31e80e9afb72070.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 241, + 461, + 255 + ], + "lines": [ + { + "bbox": [ + 305, + 241, + 461, + 255 + ], + "spans": [ + { + "bbox": [ + 305, + 241, + 461, + 255 + ], + "type": "text", + "content": "COLOR SCALE 1.00 0.75 0.50 0.25 0.00 -0.25 -0.50 0.75 -1.00" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 353, + 479, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 479, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 479, + 389 + ], + "type": "text", + "content": "not been given permission to reveal the identities and exact absolute budgets used for the annotation companies, hence we report costs relative to the MTurk budget. Examples of spam annotations are provided in Suppl. D." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 389, + 480, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 480, + 437 + ], + "type": "text", + "content": "Regardless of the annotation scenario, the annotation companies generated higher quality annotations than MTurk with higher DSC scores including smaller interquartile ranges (IQRs) (see Fig. 7a), higher NSD scores including smaller IQRs (see Fig. 7b), and lower numbers of severe errors (see Fig. 7c)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 454, + 440, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 454, + 440, + 478 + ], + "spans": [ + { + "bbox": [ + 131, + 454, + 440, + 478 + ], + "type": "text", + "content": "4.2 Does annotation companies' quality assurance improve annotation quality?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 486, + 480, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 480, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 480, + 593 + ], + "type": "text", + "content": "As shown in Fig. 4, the internal QA conducted by annotation companies does not substantially improve annotation quality if the QA workers only use minimal text instructions. The impact of QA when having access to extended text instructions or extended text including pictures instructions depended highly on the annotation company, while the effect was still minimal compared to not performing QA: Two of the four annotation companies show no change in the median DSC when comparing QA to non-QA outputs for extended text or extended text and picture instructions. The remaining two companies only show a very small relative improvement of below " + }, + { + "bbox": [ + 130, + 486, + 480, + 593 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 130, + 486, + 480, + 593 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "content": "We obtained even smaller improvements with the NSD scores, as depicted in Suppl. E. In certain scenarios, QA was able to remove a small proportion of severe errors, however, the vast majority remains (see Suppl. F). Notably, the percentage of modified images by QA workers depended highly on the company, as depicted in Suppl. G. Fig. 5 further illustrates that providing annotators with high-quality labeling instructions including pictures (green) leads them to vastly" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 178, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 178, + 91, + 448, + 103 + ], + "type": "text", + "content": "Quality Assured: Rethinking Annotation Strategies in Imaging AI" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 286, + 250 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 286, + 250 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 286, + 250 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 286, + 250 + ], + "type": "image", + "image_path": "6e4b981eabe812cf9322d05a5db34aeb5fc9d0ddce135d8bc953de53d4effa00.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 298, + 113, + 481, + 250 + ], + "blocks": [ + { + "bbox": [ + 298, + 113, + 481, + 250 + ], + "lines": [ + { + "bbox": [ + 298, + 113, + 481, + 250 + ], + "spans": [ + { + "bbox": [ + 298, + 113, + 481, + 250 + ], + "type": "image", + "image_path": "e21b511fe7d170cc6b84b0cb6c2b2f9f5a415a00e726d65b9c1f5eab932c261f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 147, + 256, + 459, + 280 + ], + "blocks": [ + { + "bbox": [ + 147, + 256, + 459, + 280 + ], + "lines": [ + { + "bbox": [ + 147, + 256, + 459, + 280 + ], + "spans": [ + { + "bbox": [ + 147, + 256, + 459, + 280 + ], + "type": "image", + "image_path": "8441f60cd3965c736c8ad62c652025e0f18cfdc2ee53735e671eabfc2dd3b52f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 289, + 482, + 389 + ], + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 389 + ], + "type": "text", + "content": "Fig. 5: Improving the labeling instructions yields a higher effect compared to adding QA to the annotation process. The baseline performance (minimal labeling instruction, no QA) according to the DSC (a) and the percentage of severe errors (b) is shown individually per company with a red asterisk (minimal labeling instruction, no QA). While adding QA only yields marginal improvements (dark blue), major performance boosts are obtained by improving the labeling instructions provided to the annotators (green). DSC scores were aggregated for each image and per labeling type. DSC displayed as dots and box plot (band indicates the median, box indicates the first (25th percentile) and third (75th percentile) quartiles). " + }, + { + "bbox": [ + 130, + 289, + 482, + 389 + ], + "type": "inline_equation", + "content": "0 \\leq \\mathrm{DSC} \\leq 1" + }, + { + "bbox": [ + 130, + 289, + 482, + 389 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 411, + 482, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 436 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 436 + ], + "type": "text", + "content": "outperform annotators working with lower-quality labeling instructions, even if these are followed by additional internal QA (dark blue)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 451, + 467, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 451, + 467, + 464 + ], + "spans": [ + { + "bbox": [ + 131, + 451, + 467, + 464 + ], + "type": "text", + "content": "4.3 Should quality assurance focus on specific subsets of images?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 469, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 469, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 469, + 482, + 540 + ], + "type": "text", + "content": "As shown in Fig. 6a, QA significantly increases the odds of improving the annotation quality for images with generalizable image characteristics such as underexposed objects, intersection of objects, objects occluded by background in comparison to the absence of this characteristic (vertical black line). The only exceptions are motion artifacts present in the background and object(s) overexposure." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 541, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 541, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 541, + 482, + 601 + ], + "type": "text", + "content": "In contrast to the general increase of images with generalizable image characteristics, none of the domain-specific image characteristics yield a significant change in the odds of improving the annotation quality over a regular image, see Fig. 6b. The only exception for domain-specific characteristics is object(s) covered by smoke, which could be interpreted as a form of object occlusion." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 617, + 217, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 217, + 630 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 217, + 630 + ], + "type": "text", + "content": "5 Discussion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "The data annotation market is estimated to reach a volume of more than USD 4.1 billion by the end of 2024 [8]. As the annotation process becomes increasingly" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "T. Rädsch et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 127, + 482, + 228 + ], + "blocks": [ + { + "bbox": [ + 135, + 115, + 258, + 125 + ], + "lines": [ + { + "bbox": [ + 135, + 115, + 258, + 125 + ], + "spans": [ + { + "bbox": [ + 135, + 115, + 258, + 125 + ], + "type": "text", + "content": "(a) GENERALIZABLE CHARACTERISTICS" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 134, + 127, + 482, + 228 + ], + "lines": [ + { + "bbox": [ + 134, + 127, + 482, + 228 + ], + "spans": [ + { + "bbox": [ + 134, + 127, + 482, + 228 + ], + "type": "image", + "image_path": "8703522a81b1d11a71e79334a1e70ca4661eaf6523b4fa5500fb88ee5fd6a924.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 134, + 240, + 482, + 369 + ], + "blocks": [ + { + "bbox": [ + 135, + 229, + 264, + 239 + ], + "lines": [ + { + "bbox": [ + 135, + 229, + 264, + 239 + ], + "spans": [ + { + "bbox": [ + 135, + 229, + 264, + 239 + ], + "type": "text", + "content": "(b) DOMAIN-SPECIFIC CHARACTERISTICS" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 134, + 240, + 482, + 369 + ], + "lines": [ + { + "bbox": [ + 134, + 240, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 134, + 240, + 482, + 369 + ], + "type": "image", + "image_path": "b1db69be1ef2add7f6d9688be19cea42fec65d9bcba3f8b654b2a3ced5c6f6ac.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 376, + 482, + 488 + ], + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 488 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 488 + ], + "type": "text", + "content": "Fig. 6: QA of annotation companies can be beneficial to certain images with difficult real-world generalizable image characteristics, such as underexposure or intersection of objects. We display the impact of (a) real-world generalizable and (b) domain-specific image characteristics on the odds of improvement. Improvement is defined as a positive difference in the QA, while no improvement refers to 0 or negative difference. The effect of an image characteristic is shown in the form of the log(odds ratio (OR)) representing the logarithms of the odds of improvement in the presence of the characteristic, compared to the odds of improvement occurring in the absence of that characteristic. The vertical black line refers to the point of equal odds. (Note: ***: p-value ≤ 0.001, **: p-value ≤ 0.01, *: p-value ≤ 0.05)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 518, + 482, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 518, + 482, + 614 + ], + "spans": [ + { + "bbox": [ + 130, + 518, + 482, + 614 + ], + "type": "text", + "content": "automated and foundation models evolve, the market is undergoing drastic shifts. With lower entry barriers for annotation, more efficient data annotation tools, and a rising impact of QA, it is thought to result in an estimated USD 8.2 billion market by 2028 [40]. As a research community, we need to understand how to ensure high annotation quality throughout this evolution, as reliable reference annotations are an indispensable foundation for the benchmarking of image analysis algorithms, and thus their ultimate translation into real-world practice." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": "To our knowledge, this study is the first to quantitatively and critically examine the impact of internal QA conducted by annotation companies in the context of safety-critical applications. The most important insights can be summarized as follows:" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 178, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 178, + 91, + 448, + 102 + ], + "type": "text", + "content": "Quality Assured: Rethinking Annotation Strategies in Imaging AI" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": "1. Annotation companies perform better both in terms of quantity and quality compared to the widely used platform MTurk (Fig. 3). This implies that requesters working with crowdsourcing platforms should seriously consider annotation companies as an alternative. The organizational structure of these companies, which enables a single point of contact for requesters and eliminates spam annotations, appears to be a contributing factor to their efficiency. In light of poor working conditions being a major problem in annotation generation [22, 36], annotation companies, which can more easily be inspected than distributed crowdsourcing platforms, are thought to offer better working contracts, enabling fairer working conditions for annotators. It should be noted that not every annotation company automatically ensures adequate working conditions. The annotation companies employed in this study underwent thorough scrutiny to guarantee compliance with good working condition guidelines. With annotators working at a shared location instead of in a distributed crowd, verifying annotators' working conditions has become substantially easier. We thus view our work as an initial step towards greater accountability on the part of annotation requesters. Further dimensions of annotation provider selection are discussed in Suppl. H." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 335, + 482, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 467 + ], + "type": "text", + "content": "2. In contrast to common belief, internal QA mechanisms do not necessarily improve annotation quality (Fig. 4). Furthermore, we observe no systematic QA quality improvement difference between annotation companies depending on their region of origin, size, or price, meaning more expensive annotation companies do not translate to better QA. However, improving labeling instructions instead of investing in QA can substantially boost the annotation performance (Fig. 5). Consequently, in order to obtain high-quality annotations, annotation requesters should prioritize refining their labeling instructions, followed by implementing QA or even a different external QA process as a secondary step. Compared to QA, better labeling instructions can be generated with relatively few resources." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 470, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 482, + 555 + ], + "type": "text", + "content": "3. The benefit of QA depends on the specific image characteristics (Fig. 6). In fact, QA did improve the quality of images with certain challenging real-world image characteristics, see Fig. 6a. Consequently, it might be useful for requesters and annotation companies alike to allocate resources to identifying challenging image characteristics during a project, train classifiers to identify images containing these characteristics and focus the internal QA effort on these images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "One could contend that the recent rise of foundation models marginalizes the role of annotation providers. We argue, however, that it only changes their role. In this context, it is necessary to differentiate between the data quality demands for the training and testing of models. With a recent trend towards fine-tuning pre-trained models, smaller high-quality datasets can achieve similar performance with less computation [5, 29], highlighting the importance of the quality of unlabeled data in unsupervised training and its role in reducing carbon emissions [45]. In testing, especially for safety-critical applications such as autonomous driving and clinical medicine, ensuring high-quality annotated test" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "T. Rädsch et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "data is crucial for accurate model evaluation and real-world applicability [43]. Consequently, while training data annotation may lose relevance, test data annotation and its high quality demands are here to stay. Regarding foundation models, it should further be noted that annotations purely produced by foundation models introduce artifacts and biases [2] and represent an increasing rising problem for MTurk, as demonstrated by Veselovsky et al. [46]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 191, + 482, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 191, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 191, + 482, + 323 + ], + "type": "text", + "content": "Our study is subject to several limitations which deserve further discussion. Comparing prices across providers is non-trivial as the costs to annotate a given data set depend on many factors including, among others, image quantity, type of annotations, location of annotation company and quality requirements. Our analysis was based on the actual costs incurred during our project. Confidentiality agreements with the software company which integrated our experiments into their annotation project pipeline prevent us from disclosing the identities of the annotation companies. Additionally, we took precautions to protect the identities of individual annotators and QA workers. This was particularly important given our findings that QA workers did not significantly enhance annotation quality, potentially exposing them to negative repercussions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 326, + 482, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 482, + 397 + ], + "type": "text", + "content": "Another limitation could be seen in the fact that our data set only represents one safety-critical domain. While this is a valid concern, we would like to highlight that we chose a dataset which combined quality, complexity, and volume. Furthermore, regarding the breadth of annotation providers and the number of annotators involved, our study is likely the largest of its kind, thereby incurring substantial costs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 400, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 485 + ], + "type": "text", + "content": "Finally, conducting multiple experiments with identical annotation companies and varying datasets could have potentially revealed our experimental design. Had the annotation companies been aware of our research, outcomes could potentially have been biased to appear more favorable, given companies' inherent interest in demonstrating their services as reliable. We thus chose to conduct our experiments such that they could be seamlessly integrated into the software company's annotation pipeline, reducing their likelihood of exposure." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 517, + 326, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 517, + 326, + 530 + ], + "spans": [ + { + "bbox": [ + 131, + 517, + 326, + 530 + ], + "type": "text", + "content": "6 Conclusion and Future Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "In conclusion, our study underscores the pivotal role of enhancing the initial annotation quality, for instance via allocating resources to the improvement of labeling instructions, over QA in obtaining higher reference annotation quality in high-stakes AI [43]. Future research should delve into the development and implementation of comprehensive training programs for annotators from annotation companies to further enhance initial annotation quality. Exploring the role of annotation companies in the context of increasingly automated annotation processes driven by foundation models presents another avenue for future research." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 178, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 178, + 91, + 447, + 103 + ], + "type": "text", + "content": "Quality Assured: Rethinking Annotation Strategies in Imaging AI" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 113, + 465, + 215 + ], + "blocks": [ + { + "bbox": [ + 149, + 113, + 465, + 215 + ], + "lines": [ + { + "bbox": [ + 149, + 113, + 465, + 215 + ], + "spans": [ + { + "bbox": [ + 149, + 113, + 465, + 215 + ], + "type": "image", + "image_path": "3e05d47b2ce47ec50a7b906ae56fde5c4e572da9e5397f0c35f7d6ecccd4bc0c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 149, + 225, + 465, + 327 + ], + "blocks": [ + { + "bbox": [ + 162, + 217, + 452, + 225 + ], + "lines": [ + { + "bbox": [ + 162, + 217, + 452, + 225 + ], + "spans": [ + { + "bbox": [ + 162, + 217, + 452, + 225 + ], + "type": "text", + "content": "(a) Annotation companies generate higher DSC scores than MTurk. " + }, + { + "bbox": [ + 162, + 217, + 452, + 225 + ], + "type": "inline_equation", + "content": "0 \\leq \\mathrm{DSC} \\leq 1" + }, + { + "bbox": [ + 162, + 217, + 452, + 225 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 149, + 225, + 465, + 327 + ], + "lines": [ + { + "bbox": [ + 149, + 225, + 465, + 327 + ], + "spans": [ + { + "bbox": [ + 149, + 225, + 465, + 327 + ], + "type": "image", + "image_path": "ad0d0af2e78cbc279915f970a0c7b569e8e1266abd92bfa0c2ffd7b330945e90.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 149, + 338, + 465, + 446 + ], + "blocks": [ + { + "bbox": [ + 164, + 328, + 451, + 338 + ], + "lines": [ + { + "bbox": [ + 164, + 328, + 451, + 338 + ], + "spans": [ + { + "bbox": [ + 164, + 328, + 451, + 338 + ], + "type": "text", + "content": "(b) Annotation companies generate higher NSD scores than MTurk. " + }, + { + "bbox": [ + 164, + 328, + 451, + 338 + ], + "type": "inline_equation", + "content": "0 \\leq \\mathrm{NSD} \\leq 1" + }, + { + "bbox": [ + 164, + 328, + 451, + 338 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 149, + 338, + 465, + 446 + ], + "lines": [ + { + "bbox": [ + 149, + 338, + 465, + 446 + ], + "spans": [ + { + "bbox": [ + 149, + 338, + 465, + 446 + ], + "type": "image", + "image_path": "eb8a72bf10a3bf83277f1a31f88a856424bec4c703a9f4cc685553377f49cb11.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 186, + 449, + 427, + 459 + ], + "lines": [ + { + "bbox": [ + 186, + 449, + 427, + 459 + ], + "spans": [ + { + "bbox": [ + 186, + 449, + 427, + 459 + ], + "type": "text", + "content": "(c) Annotation companies generate a lower number of severe errors." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 468, + 482, + 534 + ], + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 534 + ], + "type": "text", + "content": "Fig. 7: Annotation companies generate higher quality annotations with their annotation pipeline than MTurk. This is observed for (a) the DSC, (b) the NSD, and (c) number of severe errors (lower is better). Annotation companies include a QA step, in contrast to crowdsourcing on MTurk. (a) and (b) displayed as dots and box plots (band indicates the median, box indicates the first (25th percentile) and third (75th percentile) quartiles)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 563, + 246, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 563, + 246, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 563, + 246, + 578 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "A part of this work was funded by M. Mengler and S. Funke - during their time at understand.ai -, and Helmholtz Imaging. T. Rädsch was supported by a scholarship from the Hanns Seidel Foundation with funds from the Federal Ministry of Education and Research Germany (BMBF). We thank M. Gelz, S. Strzysch and N. Kraft for their continuous support." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "T. Rädsch et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 138, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 138, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 138, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 138, + 138, + 481, + 237 + ], + "type": "text", + "content": "1. Ali, S., Zhou, F., Braden, B., Bailey, A., Yang, S., Cheng, G., Zhang, P., Li, X., Kayser, M., Soberanis-Mukul, R.D., Albarqouni, S., Wang, X., Wang, C., Watanabe, S., Oksuz, I., Ning, Q., Yang, S., Khan, M.A., Gao, X.W., Realdon, S., Loshchenov, M., Schnabel, J.A., East, J.E., Wagnieres, G., Loschenov, V.B., Grisan, E., Daul, C., Blondel, W., Rittscher, J.: An objective comparison of detection and segmentation algorithms for artefacts in clinical endoscopy. Scientific Reports 10(1), 2748 (Feb 2020). https://doi.org/10.1038/s41598-020-59413-5, https://www.nature.com/articles/s41598-020-59413-5, number: 1 Publisher: Nature Publishing Group 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 238, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 238, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 238, + 481, + 456 + ], + "type": "text", + "content": "2. Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M.S., Bohg, J., Bosselut, A., Brunskill, E., Brynjolfsson, E., Buch, S., Card, D., Castellon, R., Chatterji, N., Chen, A., Creel, K., Davis, J.Q., Demszky, D., Donahue, C., Doumbouya, M., Durmus, E., Ermon, S., Etchemendy, J., Ethayarajh, K., Fei-Fei, L., Finn, C., Gale, T., Gillespie, L., Goel, K., Goodman, N., Grossman, S., Guha, N., Hashimoto, T., Henderson, P., Hewitt, J., Ho, D.E., Hong, J., Hsu, K., Huang, J., Icard, T., Jain, S., Jurafsky, D., Kalluri, P., Karamcheti, S., Keeling, G., Khani, F., Khattab, O., Koh, P.W., Krass, M., Krishna, R., Kuditipudi, R., Kumar, A., Ladhak, F., Lee, M., Lee, T., Leskovec, J., Levent, I., Li, X.L., Li, X., Ma, T., Malik, A., Manning, C.D., Mirchandani, S., Mitchell, E., Munyikwa, Z., Nair, S., Narayan, A., Narayanan, D., Newman, B., Nie, A., Niebles, J.C., Nilforoshan, H., Nyarko, J., Ogut, G., Orr, L., Papadimitriou, I., Park, J.S., Piech, C., Portelance, E., Potts, C., Raghunathan, A., Reich, R., Ren, H., Rong, F., Roohani, Y., Ruiz, C., Ryan, J., Re, C., Sadigh, D., Sagawa, S., Santhanam, K., Shih, A., Srinivasan, K., Tamkin, A., Taori, R., Thomas, A.W., Tramer, F., Wang, R.E., Wang, W., Wu, B., Wu, J., Wu, Y., Xie, S.M., Yasunaga, M., You, J., Zaharia, M., Zhang, M., Zhang, T., Zhang, X., Zhang, Y., Zheng, L., Zhou, K., Liang, P.: On the Opportunities and Risks of Foundation Models (Jul 2022). https://doi.org/10.48550/arXiv.2108.07258. http://arxiv.org/abs/2108.07258 13" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 457, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 457, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 138, + 457, + 481, + 490 + ], + "type": "text", + "content": "3. Bragg, J., Mausam, Weld, D.S.: Sprout: Crowd-Powered Task Design for Crowdsourcing. Proceedings of the ACM on User Interface Software and Technology (2018). https://doi.org/10.1145/3242587.3242598 4, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 491, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 481, + 544 + ], + "type": "text", + "content": "4. Chang, J.C., Amershi, S., Kamar, E.: Revolt: Collaborative Crowdsourcing for Labeling Machine Learning Datasets. In: Proceedings of the 2017 CHI Conference on Human Factors in Computing Systems. pp. 2334-2346. CHI '17, Association for Computing Machinery, New York, NY, USA (May 2017). https://doi.org/10.1145/3025453.3026044 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 544, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 481, + 588 + ], + "type": "text", + "content": "5. Chen, L., Li, S., Yan, J., Wang, H., Gunaratna, K., Yadav, V., Tang, Z., Srinivasan, V., Zhou, T., Huang, H., Jin, H.: AlpaGasus: Training A Better Alpaca with Fewer Data (Nov 2023). https://doi.org/10.48550/arXiv.2307.08701, http://arxiv.org/abs/2307.08701 12" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 589, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 481, + 665 + ], + "type": "text", + "content": "6. Cheplygina, V., Perez-Rovira, A., Kuo, W., Tiddens, H.A.W.M., de Bruijne, M.: Early Experiences with Crowdsourcing Airway Annotations in Chest CT. In: Carneiro, G., Mateus, D., Peter, L., Bradley, A., Tavares, J.M.R.S., Belagiannis, V., Papa, J.P., Nascimento, J.C., Loog, M., Lu, Z., Cardoso, J.S., Cornebise, J. (eds.) Deep Learning and Data Labeling for Medical Applications. pp. 209-218. Lecture Notes in Computer Science, Springer International Publishing, Cham (2016). https://doi.org/10.1007/978-3-319-46976-8_22_4_6" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 178, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 178, + 91, + 447, + 102 + ], + "type": "text", + "content": "Quality Assured: Rethinking Annotation Strategies in Imaging AI" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 136, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 481, + 160 + ], + "type": "text", + "content": "7. Chmielewski, M., Kucker, S.C.: An MTurk Crisis? Shifts in Data Quality and the Impact on Study Results. Social Psychological and Personality Science 11(4), 464-473 (May 2020). https://doi.org/10.1177/1948550619875149, publisher: SAGE Publications Inc 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 162, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 481, + 194 + ], + "type": "text", + "content": "8. Cognilytica: Data engineering, preparation, and labeling for ai 2020 (2020), https://www.cognilytica.com/2020/01/31/data-preparation-labeling-for-ai-2020/10" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 196, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 196, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 138, + 196, + 481, + 239 + ], + "type": "text", + "content": "9. Créquet, P., Mansouri, G., Benchoufi, M., Vivot, A., Ravaud, P.: Mapping of Crowdsourcing in Health: Systematic Review. Journal of Medical Internet Research 20(5), 187 (2018). https://doi.org/10.2196/jmir.9330, https://www.jmir.org/2018/5/e187/1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 240, + 481, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 240, + 481, + 284 + ], + "spans": [ + { + "bbox": [ + 133, + 240, + 481, + 284 + ], + "type": "text", + "content": "10. Daniel, F., Kucherbaev, P., Cappiello, C., Benatallah, B., Allahbakhsh, M.: Quality Control in Crowdsourcing: A Survey of Quality Attributes, Assessment Techniques and Assurance Actions. ACM Computing Surveys 51, 1-40 (Jan 2019). https://doi.org/10.1145/3148148 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 285, + 481, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 285, + 481, + 307 + ], + "spans": [ + { + "bbox": [ + 133, + 285, + 481, + 307 + ], + "type": "text", + "content": "11. Dice, L.R.: Measures of the amount of ecologic association between species. Ecology 26(3), 297-302 (1945) 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 308, + 481, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 308, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 134, + 308, + 481, + 361 + ], + "type": "text", + "content": "12. Dow, S., Kulkarni, A., Klemmer, S., Hartmann, B.: Shepherding the crowd yields better work. In: Proceedings of the ACM 2012 conference on Computer Supported Cooperative Work. pp. 1013-1022. CSCW '12, Association for Computing Machinery, New York, NY, USA (Feb 2012). https://doi.org/10.1145/2145204.2145355 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 364, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 364, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 134, + 364, + 481, + 418 + ], + "type": "text", + "content": "13. Duggan, N.M., Jin, M., Mendicuti, M.A.D., Hallisey, S., Bernier, D., Selame, L.A., Asgari-Targhi, A., Fischetti, C.E., Lucassen, R., Samir, A.E., Duhaime+, E., Kapur, T., Goldsmith, A.J.: Gamified Crowdsourcing as a Novel Approach to Lung Ultrasound Dataset Labeling (Jun 2023). https://doi.org/10.48550/arXiv.2306.06773, http://arxiv.org/abs/2306.06773 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 420, + 481, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 420, + 481, + 473 + ], + "spans": [ + { + "bbox": [ + 134, + 420, + 481, + 473 + ], + "type": "text", + "content": "14. Gaikwad, S., Chhibber, N., Sehgal, V., Ballav, A., Mullings, C., Nasser, A., Richmond-Fuller, A., Gilbee, A., Gamage, D., Whiting, M., et al.: Prototype tasks: improving crowdsourcing results through rapid, iterative task design (2017). https://doi.org/10.48550/arXiv.1707.05645, https://arxiv.org/abs/1707.05645 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 475, + 481, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 475, + 481, + 507 + ], + "spans": [ + { + "bbox": [ + 134, + 475, + 481, + 507 + ], + "type": "text", + "content": "15. Halevy, A., Norvig, P., Pereira, F.: The unreasonable effectiveness of data. IEEE intelligent systems 24(2), 8-12 (2009). https://doi.org/10.1109/MIS.2009.36, ISBN: 1541-1672 Publisher: IEEE 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 509, + 481, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 509, + 481, + 552 + ], + "spans": [ + { + "bbox": [ + 134, + 509, + 481, + 552 + ], + "type": "text", + "content": "16. Han, D., Choe, J., Chun, S., Chung, J.J.Y., Chang, M., Yun, S., Song, J.Y., Oh, S.J.: Neglected free lunch-learning image classifiers using annotation byproducts (2023). https://doi.org/10.48550/arXiv.2303.17595, https://arxiv.org/abs/2303.17595 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 554, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 554, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 134, + 554, + 481, + 597 + ], + "type": "text", + "content": "17. Heim, E., Roß, T., Teil, A., März, K., Stieltjes, B., Eisenmann, M., Lebert, J., Metzger, J., Sommer, G., Sauter, A.W.: Large-scale medical image annotation with crowd-powered algorithms. Journal of Medical Imaging 5(3), 034002 (2018). https://doi.org/10.1117/1.JMI.5.3.034002 2, 4, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 599, + 481, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 599, + 481, + 642 + ], + "spans": [ + { + "bbox": [ + 134, + 599, + 481, + 642 + ], + "type": "text", + "content": "18. Hettiachchi, D., Schaekermann, M., McKinney, T.J., Lease, M.: The Challenge of Variable Effort Crowdsourcing and How Visible Gold Can Help. Proceedings of the ACM on Human-Computer Interaction 5, 332:1-332:26 (Oct 2021). https://doi.org/10.1145/3476073 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 481, + 666 + ], + "type": "text", + "content": "19. Hutson, M., Kanzheleva, O., Taggart, C., Campana, B.J.L., Duong, Q.: Quality Control Challenges in Crowdsourcing Medical Labeling (2019) 4" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "T. Rädsch et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 664 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 170 + ], + "type": "text", + "content": "20. Idrissi, B.Y., Bouchacourt, D., Balestriero, R., Evtimov, I., Hazirbas, C., Ballas, N., Vincent, P., Drozdjal, M., Lopez-Paz, D., Ibrahim, M.: ImageNet-X: Understanding Model Mistakes with Factor of Variation Annotations (Nov 2022). https://doi.org/10.48550/arXiv.2211.01866, http://arxiv.org/abs/2211.01866 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "type": "text", + "content": "21. Ipeirotis, P.G., Provost, F., Wang, J.: Quality management on Amazon Mechanical Turk. In: Proceedings of the ACM SIGKDD Workshop on Human Computation. pp. 64-67. ACM, Washington DC (Jul 2010). https://doi.org/10.1145/1837885.1837906, https://dl.acm.org/doi/10.1145/1837885.1837906 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 216, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 247 + ], + "type": "text", + "content": "22. Irani, L.: The cultural work of microwork. New Media & Society 17(5), 720-739 (May 2015). https://doi.org/10.1177/1461444813511926, publisher: SAGE Publications 12" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 249, + 481, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 324 + ], + "type": "text", + "content": "23. Irshad, H., Montaser-Kouhsari, L., Waltz, G., Bucur, O., Nowak, J., Dong, F., Knoblauch, N., Beck, A.: Crowdsourcing image annotation for nucleus detection and segmentation in computational pathology: evaluating experts, automated methods, and the crowd. In: Biocomputing 2015. pp. 294-305. WORLD SCIENTIFIC, Kohala Coast, Hawaii, USA (Nov 2014). https://doi.org/10.1142/9789814644730_0029, http://www.worldscientific.com/doi/abs/10.1142/9789814644730_0029 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "type": "text", + "content": "24. Joskowicz, L., Cohen, D., Caplan, N., Sosna, J.: Inter-observer variability of manual contour delineation of structures in CT. European Radiology 29(3), 1391-1399 (Mar 2019). https://doi.org/10.1007/s00330-018-5695-5, http://link.springer.com/10.1007/s00330-018-5695-5 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 369, + 481, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 434 + ], + "type": "text", + "content": "25. Kentley, J., Weber, J., Liopyris, K., Braun, R.P., Marghoob, A.A., Quigley, E.A., Nelson, K., Prentice, K., Duhaime, E., Halpern, A.C., Rotemberg, V.: Agreement Between Experts and an Untrained Crowd for Identifying Dermoscopic Features Using a Gamified App: Reader Feasibility Study. JMIR Medical Informatics 11, e38412 (Jan 2023). https://doi.org/10.2196/38412, https://medinform.jmir.org/2023/1/e38412 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "type": "text", + "content": "26. Kohli, M.D., Summers, R.M., Geis, J.R.: Medical image data and datasets in the era of machine learning—whitepaper from the 2016 C-MIMI meeting dataset session. Journal of digital imaging 30(4), 392–399 (2017). https://doi.org/10.1007/s10278-017-9976-3, publisher: Springer 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 479, + 481, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 521 + ], + "type": "text", + "content": "27. Le, J., Edmonds, A., Hester, V., Biewald, L.: Ensuring quality in crowdsourced search relevance evaluation: The effects of training question distribution. In: SIGIR 2010 workshop on crowdsourcing for search evaluation. vol. 2126, pp. 22-32 (2010) 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 523, + 481, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 481, + 576 + ], + "type": "text", + "content": "28. Le Ludec, C., Cornet, M., Casilli, A.A.: The problem with annotation. Human labour and outsourcing between France and Madagascar. Big Data & Society 10(2), 20539517231188723 (Jul 2023). https://doi.org/10.1177/20539517231188723, http://journals.sagepub.com/doi/10.1177/20539517231188723 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 577, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 481, + 609 + ], + "type": "text", + "content": "29. Lee, A.N., Hunter, C.J., Ruiz, N.: Platypus: Quick, Cheap, and Powerful Refinement of LLMs (Aug 2023). https://doi.org/10.48550/arXiv.2308.07317, http://arxiv.org/abs/2308.07317 12" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 611, + 481, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 664 + ], + "type": "text", + "content": "30. Litman, L., Robinson, J., Rosenzweig, C.: The relationship between motivation, monetary compensation, and data quality among US- and India-based workers on Mechanical Turk. Behavior Research Methods 47(2), 519-528 (Jun 2015). https://doi.org/10.3758/s13428-014-0483-x, http://link.springer.com/10.3758/s13428-014-0483-x 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 178, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 178, + 91, + 447, + 102 + ], + "type": "text", + "content": "Quality Assured: Rethinking Annotation Strategies in Imaging AI" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 182 + ], + "type": "text", + "content": "31. Lu, X., Ratcliffe, D., Kao, T.T., Tikhonov, A., Litchfield, L., Rodger, C., Wang, K.: Rethinking Quality Assurance for Crowdsourced Multi-ROI Image Segmentation. Proceedings of the AAAI Conference on Human Computation and Crowdsourcing 11(1), 103–114 (Nov 2023). https://doi.org/10.1609/hcomp.v11i1.27552, https://ojs(aaai.org/index.php/HCOMP/article/view/27552, number: 1 2, 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 183, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 239 + ], + "type": "text", + "content": "32. Maier-Hein, L., Mersmann, S., Kondermann, D., Bodenstedt, S., Sanchez, A., Stock, C., Kenngott, H.G., Eisenmann, M., Speidel, S.: Can masses of non-experts train highly accurate image classifiers? In: International conference on medical image computing and computer-assisted intervention. pp. 438-445. Springer (2014). https://doi.org/10.1007/978-3-319-10470-6_55 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 240, + 481, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 481, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 481, + 295 + ], + "type": "text", + "content": "33. Maier-Hein, L., Reinke, A., Godau, P., Tizabi, M.D., Buettner, F., Christodoulou, E., Glocker, B., Isensee, F., Kleesiek, J., Kozubek, M., et al.: Metrics reloaded: recommendations for image analysis validation. Nature Methods 21, 195–212 (2024). https://doi.org/10.1038/s41592-023-02151-z, https://www.nature.com/articles/s41592-023-02151-z 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 296, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 296, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 296, + 481, + 407 + ], + "type": "text", + "content": "34. Maier-Hein, L., Wagner, M., Ross, T., Reinke, A., Bodenstedt, S., Full, P.M., Hempe, H., Mindroc-Filimon, D., Scholz, P., Tran, T.N., Bruno, P., Kisilenko, A., Müller, B., Davitashvili, T., Capek, M., Tizabi, M.D., Eisenmann, M., Adler, T.J., Grohl, J., Schellenberg, M., Seidlitz, S., Lai, T.Y.E., Pekdemir, B., Roethlingshoefer, V., Both, F., Bittel, S., Mengler, M., Mündermann, L., Apitz, M., Kopp-Schneider, A., Speidel, S., Nickel, F., Probst, P., Kenngott, H.G., Müller-Stich, B.P.: Heidelberg colorectal data set for surgical data science in the sensor operating room. Scientific Data 8(1), 101 (Apr 2021). https://doi.org/10.1038/s41597-021-00882-2, https://www.nature.com/articles/s41597-021-00882-2, number: 1 Publisher: Nature Publishing Group 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 407, + 481, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 461 + ], + "type": "text", + "content": "35. McCulloch, C.E., Searle, S.R.: Generalized, linear, and mixed models. John Wiley & Sons (2004). https://doi.org/10.1002/0471722073, https://books.google.com/books?hl=en&lr=&id=bWDPukohugQC&oi=fnd&pg=PR6&dq=Generalized, +Linear, +and+Mixed+Models.&ots=sSy3HM7CGo&sig=Ss04EHKG_xkcZZCYWSEJl6paHos 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 463, + 481, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 463, + 481, + 496 + ], + "spans": [ + { + "bbox": [ + 132, + 463, + 481, + 496 + ], + "type": "text", + "content": "36. Miceli, M., Posada, J.: The Data-Production Dispositif. Proceedings of the ACM on Human-Computer Interaction 6(CSCW2), 1-37 (Nov 2022). https://doi.org/10.1145/3555561, https://dl.acm.org/doi/10.1145/3555561 2, 12, 14" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 498, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 542 + ], + "type": "text", + "content": "37. Michaelis, C., Mitzkus, B., Geirhos, R., Rusak, E., Bringmann, O., Ecker, A.S., Bethge, M., Brendel, W.: Benchmarking Robustness in Object Detection: Autonomous Driving when Winter is Coming (Mar 2020). https://doi.org/10.48550/arXiv.1907.07484, http://arxiv.org/abs/1907.07484 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 543, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 481, + 597 + ], + "type": "text", + "content": "38. Petrovic, N., Moya-Alcover, G., Varona, J., Jaume-i Capó, A.: Crowdsourcing human-based computation for medical image analysis: A systematic literature review. Health Informatics Journal 26(4), 2446-2469 (Dec 2020). https://doi.org/10.1177/1460458220907435, http://journals.sagepub.com/doi/10.1177/1460458220907435 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 599, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 665 + ], + "type": "text", + "content": "39. Reinke, A., Tizabi, M.D., Baumgartner, M., Eisenmann, M., Heckmann-Nötzel, D., Kavur, A.E., Rädsch, T., Sudre, C.H., Acion, L., Antonelli, M.: Understanding metric-related pitfalls in image analysis validation. Nature methods pp. 1-13 (2024). https://doi.org/doi.org/10.1038/s41592-023-02150-0, https://www.nature.com/articles/s41592-023-02150-0, publisher: Nature Publishing Group 7" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "T. Rädsch et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "40. Research, G.V.: Data Collection And Labeling Market Size & Share Report 2030 (2022), https://www.grandviewresearch.com/industry-analysis/data-collection-labeling-market-3,11" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 205 + ], + "type": "text", + "content": "41. Rofs, T., Bruno, P., Reinke, A., Wiesenfarth, M., Koeppel, L., Full, P.M., Pekdemir, B., Godau, P., Trofimova, D., Isensee, F.: How can we learn (more) from challenges? A statistical approach to driving future algorithm development. arXiv preprint arXiv:2106.09302 (2021). https://doi.org/10.48550/arXiv.2106.09302, https://arxiv.org/abs/2106.09302 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 270 + ], + "type": "text", + "content": "42. Rädsch, T., Reinke, A., Weru, V., Tizabi, M.D., Schreck, N., Kavur, A.E., Pekdemir, B., Roß, T., Kopp-Schneider, A., Maier-Hein, L.: Labelling instructions matter in biomedical image analysis. Nature Machine Intelligence 5(3), 273–283 (Mar 2023). https://doi.org/10.1038/s42256-023-00625-5, https://www.nature.com/articles/s42256-023-00625-5, number: 3 Publisher: Nature Publishing Group 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 270, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 335 + ], + "type": "text", + "content": "43. Sambasivan, N., Kapania, S., Highfill, H., Akrong, D., Paritosh, P., Aroyo, L.M.: \"Everyone wants to do the model work, not the data work\": Data Cascades in High-Stakes AI. In: Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. pp. 1-15. ACM, Yokohama Japan (May 2021). https://doi.org/10.1145/3411764.3445518, https://dl.acm.org/doi/10.1145/3411764.3445518 1, 13" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 335, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 335, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 335, + 481, + 392 + ], + "type": "text", + "content": "44. Schmidt, F.A.: The Planetary Stacking Order of Multilayered Crowd-AI Systems. In: Graham, M., Ferrari, F. (eds.) Digital Work in the Planetary Market, pp. 137-156. The MIT Press (Jun 2022). https://doi.org/10.7551/mitpress/13835.003.0012, https://direct.mit.edu/books/book/5319/chapter/3800159/The-Planetary-Stacking-Order-of-Multilayered-Crowd 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 392, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 445 + ], + "type": "text", + "content": "45. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: LLaMA: Open and Efficient Foundation Language Models (Feb 2023). https://doi.org/10.48550/arXiv.2302.13971, http://arxiv.org/abs/2302.13971 12" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 445, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 445, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 445, + 481, + 489 + ], + "type": "text", + "content": "46. Veselovsky, V., Ribeiro, M.H., West, R.: Artificial Artificial Artificial Intelligence: Crowd Workers Widely Use Large Language Models for Text Production Tasks (Jun 2023). https://doi.org/10.48550/arXiv.2306.07899, http://arxiv.org/abs/2306.07899 13" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 490, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 481, + 566 + ], + "type": "text", + "content": "47. Wang, K., Hill, M., Knowles-Barley, S., Tikhonov, A., Litchfield, L., Bare, J.C.: Improving Segmentation of Breast Arterial Calcifications from Digital Mammography: Good Annotation is All You Need. In: Zheng, Y., Keles, H.Y., Koniusz, P. (eds.) Computer Vision - ACCV 2022 Workshops, vol. 13848, pp. 134-150. Springer Nature Switzerland, Cham (2023). https://doi.org/10.1007/978-3-031-27066-6_10, https://link.springer.com/10.1007/978-3-031-27066-6_10, series Title: Lecture Notes in Computer Science 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 566, + 481, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 566, + 481, + 642 + ], + "spans": [ + { + "bbox": [ + 132, + 566, + 481, + 642 + ], + "type": "text", + "content": "48. Yu, S., Zhou, H.Y., Ma, K., Bian, C., Chu, C., Liu, H., Zheng, Y.: Difficulty-Aware Glaucoma Classification with Multi-rater Consensus Modeling. In: Martel, A.L., Abolmaesumi, P., Stoyanov, D., Mateus, D., Zuluaga, M.A., Zhou, S.K., Racoceanu, D., Joskowicz, L. (eds.) Medical Image Computing and Computer Assisted Intervention - MICCAI 2020. pp. 741-750. Lecture Notes in Computer Science, Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-59710-8_72_4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "49. Ørting, S.N., Doyle, A., Hilten, A.v., Hirth, M., Inel, O., Madan, C.R., Mavridis, P., Spiers, H., Cheplygina, V.: A Survey of Crowdsourcing in Medical Image Analysis." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 178, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 178, + 91, + 447, + 102 + ], + "type": "text", + "content": "Quality Assured: Rethinking Annotation Strategies in Imaging AI" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 148, + 117, + 481, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 117, + 481, + 140 + ], + "spans": [ + { + "bbox": [ + 148, + 117, + 481, + 140 + ], + "type": "text", + "content": "Human Computation 7, 1-26 (Dec 2020). https://doi.org/10.15346/hc.v7i1.1, https://hcjournal.org/index.php/jhc/article/view/111 1" + } + ] + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "T. Rädsch et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Quanta Video Restoration/c31fed1f-5357-4c1e-adf0-39121e1936f9_content_list.json b/2024/Quanta Video Restoration/c31fed1f-5357-4c1e-adf0-39121e1936f9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..69351018300103e7215b2ff535bed3e243bd91bd --- /dev/null +++ b/2024/Quanta Video Restoration/c31fed1f-5357-4c1e-adf0-39121e1936f9_content_list.json @@ -0,0 +1,1803 @@ +[ + { + "type": "text", + "text": "Quanta Video Restoration", + "text_level": 1, + "bbox": [ + 349, + 142, + 653, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prateek Chennuri1, Yiheng Chi1, Enze Jiang1, G.M. Dilshan Godaliyadda2, Abhiram Gnanasambandam2, Hamid R. Sheikh2, Istvan Gyongy3, and Stanley H. Chan1", + "bbox": [ + 228, + 188, + 774, + 236 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Purdue University", + "bbox": [ + 431, + 247, + 570, + 260 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Samsung Research America", + "bbox": [ + 401, + 261, + 599, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 University of Edinburgh", + "bbox": [ + 413, + 275, + 589, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{pchennur,chi14,jiang708,stanchan}@purdue.edu", + "bbox": [ + 328, + 290, + 674, + 303 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{dilshan.g,abhiram.g,hr.sheikh}@samsung.com", + "bbox": [ + 334, + 304, + 666, + 316 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{igyongy2}@exseed.ed.ac.uk", + "bbox": [ + 401, + 318, + 602, + 330 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/717431360766581130aafba6d78ed325ed59dbd715ae724c05202ebba57c80da.jpg", + "image_caption": [ + "(a) Ground Truth 16-bit" + ], + "image_footnote": [], + "bbox": [ + 232, + 345, + 367, + 397 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/72f8a4bfd81ddfc6cf365b6676284a7f2c4013fba754cdeef86c3d959eba390d.jpg", + "image_caption": [ + "(b) CMOS 16-bit, 60 fps at 1 lux" + ], + "image_footnote": [], + "bbox": [ + 367, + 345, + 500, + 397 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/800119ba33ffe1f34b74b36b8aab7890dcc9847ac7f3984a09ac19f94891562c.jpg", + "image_caption": [ + "(c) CMOS 16-bit, 240 fps at 1 lux" + ], + "image_footnote": [], + "bbox": [ + 500, + 345, + 633, + 397 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4f6c112ec80eed3beec144f79075088a7834e772eb0833a34e906e039b2cb13f.jpg", + "image_caption": [ + "(d) CMOS 16-bit, 2000 fps at 1 lux" + ], + "image_footnote": [], + "bbox": [ + 633, + 345, + 767, + 397 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b06a96207be44f26a7ad4b1270b5bce50ca8ce2cbc15a7bf67686639c9f03b99.jpg", + "image_caption": [ + "(e) Quanta 3-bit, 2000 fps at 1 lux", + "Fig. 1: Goal of this paper. (a) Blur-free video frame of a moving car. (b)-(d) CMOS image sensor simulations using realistic sensor parameters. The strong shot noise and read noise $(5.1\\mathrm{e}^{-}/\\mathrm{pix})$ of CMOS sensor make the signal acquisition difficult. (e) With low read noise $(0.2\\mathrm{e}^{-}/\\mathrm{pix})$ , low-bit single-photon detectors capture valuable information. (f)-(g) Existing state-of-the-art algorithm, QBP [47] cannot handle strong motion and noise. (h) The proposed algorithm, QUIVER, produces high quality results." + ], + "image_footnote": [], + "bbox": [ + 232, + 406, + 367, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e28400dbfc1bfff024ff7fe8fe45c537123c8bcc2dc0d2b2fae4933cb5cf2bca.jpg", + "image_caption": [ + "f) Quanta 3-bit, reconstructed from 11-frame avg. using Traditional Methods [47]" + ], + "image_footnote": [], + "bbox": [ + 367, + 406, + 500, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3a5136568023e9a68a1a504178093a13cb9e6dda89203ff7fb294173bac88216.jpg", + "image_caption": [ + "(g) Quanta 3-bit, reconstructed from 66-frame avg, using Traditional Methods [47]", + "(h) Quanta 3-bit, reconstructed from 11-frame avg using QUIVER" + ], + "image_footnote": [], + "bbox": [ + 500, + 406, + 633, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/67a01e4617af268a2feb53a47ff71f8049f7205a092816eb76a5ffa442fe0698.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 406, + 767, + 458 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The proliferation of single-photon image sensors has opened the door to a plethora of high-speed and low-light imaging applications. However, data collected by these sensors are often 1-bit or few-bit, and corrupted by noise and strong motion. Conventional video restoration methods are not designed to handle this situation, while specialized quanta burst algorithms have limited performance when the number of input frames is low. In this paper, we introduce Quanta Video Restoration (QUIVER), an end-to-end trainable network built on the core ideas of classical quanta restoration methods, i.e., pre-filtering, flow estimation, fusion, and refinement. We also collect and publish I2-2000FPS, a high-speed video dataset with the highest temporal resolution of 2000 frames-per-second, for training and testing. On simulated and real data, QUIVER outperforms existing quanta restoration methods by a significant margin. Code and dataset available at https://github.com/ chennuripratek/Quanta_Video_Restoration-QUIVER", + "bbox": [ + 259, + 590, + 740, + 799 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Single Photon Detectors $\\cdot$ Video Restoration $\\cdot$ High-Speed Dataset", + "bbox": [ + 261, + 811, + 740, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 143, + 375, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Over the past decade, the astonishing growth of single-photon detectors has fundamentally changed the landscape of computational imaging. With the invention and proliferation of quanta image sensors (QIS) [21] and single-photon avalanche diodes (SPAD) [52, 62], there is an unprecedented volume of new applications in low-light imaging [6, 10, 66], computer vision [27, 31, 40], high-speed videography [47, 48], time-of-flight sensing [32, 63], and 3D imaging [30, 44]. In most of these use cases, the main core question that lies is how to recover the image from the photon counts measured in the scene. Specifically, given a video stream of 1-bit or few-bit data captured from a scene involving moving objects, how do we reconstruct a gray-scale image/video while eliminating the noise without incurring motion blur?", + "bbox": [ + 212, + 209, + 787, + 376 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To give the reader a visual perspective of the problem scope, Fig. 1 depicts a blur-free video of a moving car. We simulate the captured images at 1 lux assuming 60 fps, 240 fps, and 2000 fps CMOS image sensors with realistic sensor specifications. As illustrated in the figure, the resulting CMOS outputs are either severely blurred due to strong motion or completely distorted by noise due to sparse photons. In the same figure, we demonstrate a simulated single-photon camera output (a 3-bit QIS in this case) where the content the largely preserved despite heavy noise. Upon utilizing state-of-the-art Quanta Burst Photography (QBP) [47] for reconstructing the frames, provided the motion is slow, a decent output can be obtained. However, as the temporal window narrows down, as shown in Fig. 1 (f), the noise remains. In this paper, we address this problem with a new algorithm, designed to remove the noise while avoiding distortions in the presence of fast motion while utilizing only a few frames.", + "bbox": [ + 212, + 382, + 787, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The core innovation of this paper is QUanta Video Restoration (QUIVER), a deep-learning based video restoration algorithm for quanta image data. QUIVER is specialized for few-bit data (3-bit) captured at thousands of frames-per-second (2000 fps) with an average motion range of 1 to 7 pixels per frame. The main contributions of this paper can be summarized as follows.", + "bbox": [ + 212, + 585, + 787, + 662 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose QUIVER, an end-to-end trainable quanta (video) restoration method built by embracing the core ideas from traditional quanta restoration algorithms. On a comprehensive evaluation dataset containing both simulated and real data, QUIVER outperforms all methods we compared in this paper by a significant margin.", + "- We introduce I2-2000FPS, the first high-speed video dataset with a temporal resolution of 2000 frames-per-second for training and testing image and video reconstruction neural networks. We captured a total of 280 high-speed videos covering 114 distinct scenes with ground truth and simulated 3-bit videos." + ], + "bbox": [ + 223, + 696, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "P. Chennuri et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/072ec0ae0cceaf8219c9aee1913c2fbb23e1b6ae6b49c88aee0b8831e59a9b5b.jpg", + "image_caption": [ + "Fig. 2: Motion Blur and SNR Trade-off. The effects of bit depth on SNR and motion blur are illustrated using real captures by a single-photon sensor. For the motion range we target, 3-bit single-photon detectors provide the best trade-off between blur and SNR. The images are captured using a 1-bit SPAD [17] at $10\\mathrm{k}$ fps at an average photon level of 0.51 and 0.40 photons-per-pixel (PPP) per frame, respectively. Higher bit-depth outputs are generated through temporal frame averaging." + ], + "image_footnote": [], + "bbox": [ + 272, + 143, + 723, + 280 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/f0751314eb054760e6f73b2292a750e1140bb5bc9e7f2dcaf47e99ea16358d82.jpg", + "table_caption": [ + "Table 1: Frame-rate, motion, read-noise, and data-rate statistics for various bit-depths at the same exposure level." + ], + "table_footnote": [], + "table_body": "
Bit-DepthfpsMotion (pixels/frame)σread (/pixel/sec)Data-rate (Mb/sec)
110k0 - 12000 e-96
314282 - 3285.6 e-41.13
53236 - 1264.6 e-15.5
77825 - 3015.6 e-5.24
92070 - 804 e-1.73
", + "bbox": [ + 357, + 426, + 638, + 520 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Background", + "text_level": 1, + "bbox": [ + 215, + 547, + 369, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Few-bit Single-Photon Detectors", + "text_level": 1, + "bbox": [ + 214, + 583, + 532, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "What is few-bit photon counting? Single-photon detectors (QIS and SPAD) differ from conventional CMOS pixels by their extraordinary photon counting capability. QIS uses a two-stage pump-gate technique and correlated double sampling to suppress the read noise, while SPAD uses avalanche multiplication to amplify the photocharge. In both cases, the sensors are capable of resolving photons up to a single-photon sensitivity. We refer readers interested in the sensor development of QIS and SPAD to consult, for example, [5,9,18,45].", + "bbox": [ + 212, + 612, + 787, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Along with the single-photon detectors' unique capability to count individual photons, these devices can generate data at a bit-depth as low as 1-bit to as high as 16-bit or even more. However, higher bit-depth is accompanied by longer integration time. If the scene contains motion, longer integration time will eventually result in strong motion blurs as shown in Fig. 2. On the other hand, 1-bit sensing with high frame rates will result in motion-blur-free but extremely noisy images. Therefore, from a pure data acquisition perspective, there exists an optimal bit-depth with respect to the motion that will give us mini-", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "QUIVER", + "bbox": [ + 666, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mal/no motion-blur data with a minimum per-frame signal-to-noise ratio (SNR) required for good quality reconstruction.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "How about 1-bit and reconstruct afterward? Readers familiar with single-photon counting may wonder whether we can collect as many 1-bit frames as possible and then process the data afterward. The problem is power consumption and data rate. Fixing the same level of exposure, as described in Tab. 1, a 1-bit video at $10\\mathrm{k}$ fps would require $96\\mathrm{Mb / sec}$ whereas a 9-bit video at 20 fps would only need $1.73\\mathrm{Mb / sec}$ . Another problem is read noise accumulation. For sensors with non-zero read noise (such as QIS), every frame contributes to a finite amount of read noise. The more frames we read, the more read noise we accumulate. Therefore, recording 1-bit is not always the best option.", + "bbox": [ + 212, + 176, + 787, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 333, + 377, + 347 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image and Video Denoising. Classical state-of-the-art methods utilize a non-local strategy to identify similar patches across an image/video [3, 38, 50]. Deep neural networks have been proven to be successful in producing high quality denoised outputs [13, 69-71]. Among these architectures, Vision Transformers [41, 42] have been rated the state-of-the-art in recent times. However, all these solutions make simplistic assumptions on noise statistics, thus failing to perform on real noisy images or videos [55].", + "bbox": [ + 212, + 356, + 787, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Coming to low light, Burst denoising [34, 43], where images are aligned, merged and denoised, is one of the most popular methods. However, these methods fail without robust alignment. To overcome this, a number of alternative solutions with learnable alignment modules have been proposed [29, 54, 73]. Recent solutions have focused on practical noise models that replicate real camera sensor noise, to produce visually appealing results [51, 74]. Nevertheless, the existing solutions utilize images captured using CMOS image sensors, resulting in a notably higher photon level compared to the one utilized in our study.", + "bbox": [ + 212, + 462, + 787, + 582 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SPADs, Event, and Spike Cameras. Gariepy et al. [24, 25] firstly introduced the utilization of Single Photon Avalanche Diodes (SPADs) at pico-second temporal resolution to capture light in motion. Gyongy et al. [33] demonstrated 2D motion tracking of rigid planar objects using SPADs at $10\\mathrm{k}$ fps. Recently, Ma et al. [47] and Seets et al. [64] utilized SPADs in a passive imaging setting to capture motion in low illumination. However, all these methods utilize extremely high-temporal resolutions hindering the deployment of these sensors into consumer devices where bandwidth is the bottleneck. Event [53, 60] and Spike Cameras [84, 86] also have demonstrated their effectiveness in capturing high-speed motion. However, these cameras focus on luminance/brightness variation and record a spike only when variation is above a threshold (changes based on factors like temperature, event rate, etc.) [15, 60]. Therefore, unlike single-photon detectors (QIS and SPADs), these cameras are NOT designed for single-photon counting and cannot operate in extreme low-light conditions.", + "bbox": [ + 212, + 583, + 787, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "QIS Reconstruction. Reconstructing quanta images is a challenging task due to the underlying Poisson-Gaussian statistics. Initial solutions to this problem included utilizing standard gradient descent [78], greedy algorithms [79],", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "P. Chennuri et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b6a0645154d3686245d723db34243ef0054eda139319e1bec42f459e3a01387e.jpg", + "image_caption": [ + "Fig. 3: Traditional Methods' Design. Depiction of existing classical quanta restoration algorithms' design philosophy. Best viewed in zoom." + ], + "image_footnote": [], + "bbox": [ + 232, + 143, + 772, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(ADMM) [7,8], among others [19,20,23,26,61,77]. Chan et al. [6] were the first to propose a non-iterative approach using Anscombe transform for reconstructing quanta images. Choi et al. [12] proposed the first end-to-end trainable deep neural network (DNN) for QIS reconstruction. Alternative DNN-based solutions include utilizing vision transformers [75], Dual Prior Integrated networks [80], among others [81]. Nonetheless, all these methods fail to produce good results when the scene is in motion. Chi et al. [10] is the only method which focuses on capturing dynamic scenes using QIS but only targets at slow motion (1 pixel/frame).", + "bbox": [ + 212, + 275, + 787, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 QUanta VIDEo REstoration (QUIVER)", + "text_level": 1, + "bbox": [ + 215, + 436, + 630, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Design Philosophy", + "text_level": 1, + "bbox": [ + 215, + 470, + 415, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section we present the design of our proposed algorithm. We start by briefly reviewing the design of classical methods [33, 47] which, to an extent, have been successful in restoring quanta images. As shown in Fig. 3, classical methods' algorithm design can be divided into four stages: (1) computing sum images to increase SNR, (2) optical flow (or) transformation matrix estimation for aligning the input frames, (3) warping and linear combination for generating preliminary restored output, and (4) refinement for producing the final output. While the steps seem intuitive and straightforward, existing methods are heavily vulnerable to extreme noise and strong motion in the input frames primarily due to two reasons. (1) none of the stages are designed to handle extreme noise and strong motion simultaneously (will be discussed further). (2) Since all the stages are sequential yet independent of each other, it is difficult to obtain an optimal result for a wide range of noise and motion. Our proposed algorithm QUIVER, leverages the design philosophy of existing classical methods while designing each stage to simultaneously handle both noise and motion. Moreover, QUIVER is an end-to-end trainable model making all the stages inter-dependent, thus leading to good restoration outputs.", + "bbox": [ + 212, + 500, + 787, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Design of QUIVER", + "text_level": 1, + "bbox": [ + 215, + 781, + 423, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "QUIVER is a deep-learning-based video restoration method for quanta imaging. The design philosophy of QUIVER is to adopt the intuitive thoughts behind", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "QUIVER", + "bbox": [ + 666, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5c7f37774aa0261c6b87616ddb40d180a26679556a92da9868216b68ec56eceb.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 250, + 146, + 455, + 268 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d42c30ab7da6ddb25b54d18efe25c754167206411928a384ee6f4c4199841416.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 467, + 146, + 759, + 268 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b756e46082b42a0320042036eede933e7b2c3d4978c22bdf302fa000d7b96400.jpg", + "image_caption": [ + "Fig. 4: Traditional Methods' Limitations. (a) Traditional methods' [33, 47] predenoising/temporal-averaging fails in strong motion. It is visible in the restored images that an input with strong motion between the frames results in several artifacts in the output even though SNR levels are similar. (b) Traditional methods [47] utilize a patch-based pre-trained optical flow module similar to [34]. This optical flow module fails to compensate for motion in the presence of noise.", + "Fig.5: The proposed QUIVER network. The corresponding stages of QUIVER, built by embracing the intuitive thoughts behind existing classical methods. Best viewed in zoom." + ], + "image_footnote": [], + "bbox": [ + 217, + 395, + 781, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "classical quanta restoration methods and develop an end-to-end trainable, robust to noise and motion deep learning based framework, as shown in Fig. 5. Specifically, QUIVER can be divided into four main stages:", + "bbox": [ + 212, + 628, + 784, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pre-Denoising to improve SNR: Since the input quanta frames possess extreme noise, classical methods adopt naive averaging to increase the SNR and thereby predict better optical flows or transformation matrices. However, as shown in Fig. 4(a), the simple averaging is vulnerable to motion and will negatively impact subsequent processing, ultimately leading to distorted outputs. Simply eliminating this stage is not the solution, because it leads to poor optical flow estimation, resulting in over-smoothed outputs with lack of low-level intricate details, as shown in Fig. 5 and Fig. 11. Therefore, a preliminary denoising step robust to noise and motion is crucial. In QUIVER, we leverage a computational undemanding single image denoiser built using Residual Dense Blocks [83] (RDBs) to provide minimal pre-preprocessing of the input quanta", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "P. Chennuri et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ebaaf12819d0307f87a753e9c3cd487c6c9fc5e0b48962895385d2af119a9b8c.jpg", + "image_caption": [ + "Fig. 6: Design of the proposed modules DC-GFU and RMDF." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 454, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/232f176b6f5c7320a3c9b211ba6e2d98d0d23ec67cb9a63335249cac1d1db811.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 457, + 143, + 784, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "data. A multi-frame denoiser is not an option due to it computationally demanding nature. We use RDB based network due to its history in handling noise while preserving details with its simple yet effective design.", + "bbox": [ + 212, + 306, + 782, + 352 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Optical Flow Estimation + Feature Alignment using Deformable Convolution - Gated Fusion Unit: Classical methods utilize an off-the-shelf pre-trained optical flow estimation module or predict a transformation matrix to compensate for motion between the frames. The basic assumption behind such approaches is that the motion between the frames is limited and the SNR is high enough. However, when such assumption is not met, the motion compensation is sub-optimal, as shown in Fig. 4(b). As most state-of-the-art pre-trained optical flow estimators are optimized on the CMOS RGB sensor images, it leads to sub-optimal performance when applied on quanta frames. Eliminating the flow estimation step is not recommended since experiments reveal the critical role it plays in motion compensation, as shown in Fig. 5 and Fig. 11. For QUIVER, we deploy a learnable optical flow estimation module and utilize SpyNet [57] owing to its computational efficiency while using a multi-scale approach.", + "bbox": [ + 212, + 353, + 785, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We deploy 3D convolution blocks to extract multi-scale spatio-temporal features from both the noisy and denoised quanta frames. We reuse the noisy frames to compensate for any information lost in the pre-denoising stage.", + "bbox": [ + 212, + 550, + 785, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The estimated multi-scale robust-to-noise optical flows are utilised for feature-level alignment of the extracted multi-scale spatio-temporal features. We utilize the deformable convolution with residual offsets proposed by [4] to warp the features. Inspired by the superior performance of Gated Linear Units (GLUs) in Transformers [65] we design and add a GLU based multi-layer-perceptron layer with GeLU activation for efficiently fusing the aligned features extracted from both the noisy and denoised frames. As shown in Fig. 6(a), we name this deformable convolution-GLU combination as the DC-GFU module. At this fusion-stage each frame is processed separately and the fusion is performed only along the channel dimension. Recurrence is applied for the alignment stage across all the multi-scale features of each frame. For the fusion module, we do not employ recurrence, as different scale features capture distinct long-range dependencies owing to their varying receptive fields.", + "bbox": [ + 212, + 597, + 785, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Deep Feature Fusion using Recurrent Multi-scale Residual Dense Feature Fusion Unit: Post warping we want to perform a robust-to-noise dense feature fusion while taking advantage of the temporal correlations among", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "QUIVER", + "bbox": [ + 666, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/37909717e764ae1269c14e17a918a810d8e05332b3e7b5ead6b27d6c378f1d82.jpg", + "image_caption": [ + "Fig. 7: Design of the proposed modules TCA and RFRM." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 506, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6cdfb18d9d26e7805896c65eba94777ed07ffadc0be2020ad6d22caaf23958d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 143, + 782, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the features of all the input frames and also the spatial correlations between the multi-scale features within the same frame. For this task, we design and propose a Recurrent Multi-scale Residual Dense Feature Fusion Unit (RMDF) as shown in Fig. 6(b). The recurrence comes from the fact that the same RMDF module is applied progressively to all the frames' features. For any frame $t$ , the RMDF takes in the corresponding frame's multi-scale features $\\{\\mathbf{F}_t^1, \\mathbf{F}_t^2, \\mathbf{F}_t^4\\}$ , bilinearly interpolated noisy frames $\\{\\mathbf{I}_t^1, \\mathbf{I}_t^2, \\mathbf{I}_t^4\\}$ and a hidden state $\\mathbf{h}_{t-1}$ as inputs. The multi-scale features are progressively fused in a feed-forward fashion, with Residual Dense Block (RDB) as the basic block, to effectively extract both the short and long range dependencies required for good reconstruction. As shown in Fig. 6(b), multi-scale features are extracted from the noisy frames and fused with the other corresponding input features to minimize any errors accumulated through the previous stages. While these features are utilized to exploit the spatial correlations within the frame, the hidden state $\\mathbf{h}$ captures the temporal correlations between all the input frames. Thus, the design of RMDF enables it to extract densely fused multi-scale spatio-temporal features required for enhanced quality outputs.", + "bbox": [ + 212, + 337, + 787, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multi-Scale Reconstruction using Residual Frame Refinement Module: Considering the heavy noise in the input quanta frames, this ill-posed problem's restored image subspace can be quite large. To output a restored image close to the ground truth we prefer deep supervision that lets the model preserve critical details of the scene. We opt for a multi-scale reconstruction approach where the image at each scale is reconstructed in a progressive fashion. Experiments reveal the efficacy of this approach (Fig. 5, Fig. 11). The stage starts with the lowest scale features extracted from all the frames using RMDF being concatenated and sent into a newly designed Temporal Cross Attention (TCA) module. As shown in Fig. 7(a), the TCA module is similar to the multi-head attention in vision transformers [16] in terms of generating queries, keys, and values. However, we maintain the number of heads to be one and apply attention only on the channel dimension. The cross attention is from the fact that we input features extracted from all the input frames. The extracted cross-attention features are then fed into a newly designed Residual Frame Refinement Module (RFRM). As shown in Fig. 7(b), RFRM takes in a residual frame $\\mathbf{r}_t^\\alpha$ , a", + "bbox": [ + 212, + 598, + 787, + 843 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "P. Chennuri et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "hidden state $\\mathbf{f}_t^\\alpha$ , and the features as input, concatenates the hidden state with the features to input into the channel attention block [82] to emphasize critical spatio-temporal information. Further, we divide the module into 2 branches. While the former is designed to extract multi-scale spatial correlation information and output a modified hidden state $\\mathbf{f}_t^{\\alpha /2}$ , the latter focuses on refining the residual frame to output a corresponding scale reconstructed image $\\mathbf{O}_t^\\alpha$ and a residual frame $\\mathbf{r}_t^{\\alpha /2}$ . The main purpose of this setup is to initially restore the high-level features through estimating $\\mathbf{O}_t^4$ , followed by focusing on the low-level intricate details while refining the residual frames for scales 2 and 1.", + "bbox": [ + 212, + 146, + 782, + 284 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Loss Function. We train QUIVER with a multi-scale loss. The overall loss function can be represented as", + "bbox": [ + 212, + 286, + 782, + 316 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {Q}} = \\lambda_ {1} \\cdot \\mathcal {L} (\\mathbf {I} ^ {1, \\mathrm {G T}}, \\mathbf {I} _ {d} ^ {1}) + \\lambda_ {2} \\cdot \\mathcal {L} (\\mathbf {I} _ {t} ^ {1, \\mathrm {G T}}, \\mathbf {O} _ {t} ^ {1}) + \\lambda_ {3} \\cdot \\mathcal {L} (\\mathbf {I} _ {t} ^ {2, \\mathrm {G T}}, \\mathbf {O} _ {t} ^ {2}) + \\dots \\\\ \\lambda_ {4} \\cdot \\mathcal {L} \\left(\\mathbf {I} _ {t} ^ {4, \\mathrm {G T}}, \\mathbf {O} _ {t} ^ {4}\\right), \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 241, + 325, + 784, + 364 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathbf{I}_t^{\\alpha,\\mathrm{GT}}$ is the captured $t^{\\mathrm{th}}$ ground truth frame bicubically downsampled by $\\alpha$ , and $\\mathcal{L}(\\mathbf{I}_a, \\mathbf{I}_b) = ||\\mathbf{I}_a - \\mathbf{I}_b||_1 + ||\\nabla_x \\mathbf{I}_a - \\nabla_x \\mathbf{I}_b||_1 + ||\\nabla_y \\mathbf{I}_a - \\nabla_y \\mathbf{I}_b||_1$ . Here, $\\nabla_x$ and $\\nabla_y$ represent the operations of computing horizontal and vertical gradients.", + "bbox": [ + 212, + 376, + 782, + 424 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Proposed I2-2000FPS dataset", + "text_level": 1, + "bbox": [ + 214, + 445, + 537, + 462 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "While several high-frame-rate datasets have been open-sourced in recent times [36, 49, 60, 67, 68, 72], these datasets mainly feature videos tainted by severe motion blur, making them unsuitable in our problem setting. Moreover, features such as high motion speed and sufficient number of videos are also not always guaranteed. A visual representation of existing datasets' comparison is shown in Fig. 8(a). To address the gaps, we introduce the I2-2000FPS dataset, a high frame rate video collection meticulously designed to capture high-speed motion with precision.", + "bbox": [ + 212, + 476, + 787, + 595 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The I2-2000FPS dataset has a temporal resolution of 2000 FPS and a spatial resolution of $512 \\times 1024$ pixels, comprising 280 unique videos spanning 114 diverse scenes. The videos are captured using the Chronos 1.4 high-speed CMOS sensor-based camera from Kron Technologies. Notably, I2-2000FPS incorporates dark current calibration, leveraging the camera's capabilities to mitigate dark current effects. Throughout the data collection process, analog and digital gain were consistently maintained at 0 dB to avoid amplification of noise. To minimize noise, the videos are exclusively captured outdoor with ambient lighting conditions. More details on I2-2000FPS can be found in the supplementary.", + "bbox": [ + 212, + 597, + 785, + 732 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 214, + 755, + 375, + 771 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Image Formation Model", + "text_level": 1, + "bbox": [ + 214, + 785, + 462, + 800 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For experiments involving synthetic data, we use a single-photon detector simulator based on an underlying image formation model discussed below. We build", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "QUIVER", + "bbox": [ + 666, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c60c27b47b93c3ff5cdc147408750ac71e30375e7aa8ba198fb3ba8499464a74.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 259, + 143, + 498, + 281 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/ab268fd60e08d107421860a19227ab4fe06f2f630d1a821c9220c6a03e519d59.jpg", + "image_caption": [ + "(b)", + "Fig. 8: (a) Benchmarking high-speed video datasets. Horizontal axis represents the temporal resolution and the vertical axis indicates the maximum speed captured by the dataset, assuming a fixed camera-object distance. The circles in blue and orange indicate blur and blur-free videos, respectively. (b) Benchmarking Quanta Video Restoration Models on the I2-2000FPS dataset. Horizontal axis represents the computational complexity in terms of GFLOPs and the vertical axis indicates the PSNR acquired at 3.25PPP." + ], + "image_footnote": [], + "bbox": [ + 504, + 143, + 743, + 281 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "upon the prototype initially suggested in [46] adopted in prior works [6, 10, 11, 20, 26, 28, 56].", + "bbox": [ + 212, + 434, + 782, + 463 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Given the quanta exposure [22], $\\mathbf{I}^{\\mathrm{GT}}$ , dependent on the photon flux and exposure time, the observed signal by the sensor can be represented as a Poisson-Gaussian random variable, where the Poisson represents the photon arrival process and the Gaussian models the read noise. The readout process involves various sources of distortions and an Analog-to-Digital Converter (ADC) to convert the real numbers into integers $\\{0,1,2,\\dots,L\\}$ , where $L = 2^{\\mathrm{Nbits}} - 1$ depending on the bit-depth (Nbits) allocated to the sensor. The final sensor readout, $\\mathbf{Y}$ , can be represented using the following equation,", + "bbox": [ + 212, + 464, + 787, + 585 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} \\sim \\mathrm {A D C} _ {[ 0, \\mathrm {L} ]} \\{\\mathrm {P o i s s o n (Q E \\times I ^ {G T} +} \\theta_ {\\mathrm {d a r k}}) + \\underbrace {\\mathrm {G a u s s} (0 , \\sigma_ {\\mathrm {r e a d}} ^ {2} \\mathbf {1})} _ {\\mathrm {r e a d n o i s e}} \\}. \\qquad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 272, + 595, + 785, + 632 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Akin to previous works [6, 10, 20, 26, 28], we assume our sensor to be monochromatic as we utilize monochromatic real data in our experiments. For our sensor prototype, we utilize a Quantum Efficiency (QE) of 0.80. The dark current $(\\theta_{\\mathrm{dark}})$ and read noise $(\\sigma_{\\mathrm{read}})$ are set to $1.6\\mathrm{e}^{-}/\\mathrm{pix/sec}$ and $0.2\\mathrm{e}^{-}/\\mathrm{pix}$ , respectively.", + "bbox": [ + 212, + 641, + 787, + 704 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Experimental Settings", + "text_level": 1, + "bbox": [ + 214, + 724, + 447, + 739 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training data. We curate a set of 249 videos from the I2-2000FPS collection and employ it as the training dataset for all the deep-learning models in our experiments. Each training sample is fetched on the fly from each clip. A training sample here is defined as a tuple containing the ground-truth/target frames and the 3-bit quanta frames simulated at 3.25 photons-per-pixel (PPP) ( $\\sim$ 1 lux assuming a $1.1\\mu \\mathrm{m}$ pixel pitch and a 1/2000 second exposure time) using the", + "bbox": [ + 212, + 748, + 787, + 842 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "P. Chennuri et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/70fccee96f5fb920aef792c769f603f1d94bf5a4c8563c7aec956a8243fb4551.jpg", + "image_caption": [ + "Fig. 9: Visual comparisons of the reconstructed results on test videos from the proposed I2-2000FPS dataset. For fair comparison, all methods utilize 11 3-bit quanta frames simulated at 3.25 PPP per frame ( $\\sim$ 1 lux) to produce a restored frame. Best viewed in zoom." + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 756, + 529 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "image formation model described in Sec. 5.1.", + "bbox": [ + 212, + 628, + 539, + 643 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Testing data. To effectively analyze the performance of various methods, we carefully sample 31 videos from I2-2000FPS containing various motion types, shapes, and speeds. To test the generalizability, we also test the algorithms on X4K1000FPS [67] test dataset containing 15 videos from distinct scenes. Lastly, to measure the performance on real-world data, we collect binary frames using a SPAD sensor [17] and compare the reconstructed outputs. More details will be discussed in Sec. 5.3.", + "bbox": [ + 212, + 643, + 787, + 747 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Baselines. We compare the proposed method with eight existing dynamic scene reconstruction algorithms, namely Transform Denoise [6], QBP [47], Student-Teacher [10], RVRT [42], EMVD [2], FloRNN [1], MemDeblur [35], and Spk2ImgNet [85]. We also add an off-the-shelf denoiser BM3D [14] to QBP, denoted QBP (+BM3D), as a baseline for comparison. As we will discuss in Sec. 5.3, QUIVER beats all the baselines, both quantitatively and qualitatively.", + "bbox": [ + 212, + 750, + 802, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "QUIVER", + "bbox": [ + 666, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7f356c9d60b1328faa2a85feb5ce8510853e17943abc715d35f3f19d150d5460.jpg", + "image_caption": [ + "Fig. 10: Performance on Real Quanta Data. We capture real 1-bit quanta data using a SPAD [17] and generate 3-bit frames through temporal averaging. All deep learning based models are trained using a photon level of 4.9 PPP per frame. Best viewed in zoom." + ], + "image_footnote": [], + "bbox": [ + 217, + 150, + 785, + 282 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Training QUIVER. We utilize the function mentioned in Eq. (1) as the cost function for training QUIVER with regularization parameters $\\lambda_{1} = 0.2$ , $\\lambda_{2} = 0.85$ , $\\lambda_{3} = 0.1$ , and $\\lambda_{4} = 0.05$ . The training data is extracted with patch size $128 \\times 128$ and a batch size of 4. The weights are initialized with Lecun initialization [39]. The network is trained using the Adam optimizer [37] with an initial learning rate of $2.5 \\times 10^{-5}$ . The low learning rate is driven by the inherent instability of recurrent networks, as it mitigates the risk of divergent behavior during training. We use a learning rate scheduler that reduces the learning rate by a factor of 2 when a plateau is reached. QUIVER takes approximately 1.5 days to train on a NVIDIA A100 Tensor Core GPU using Pytorch.", + "bbox": [ + 212, + 381, + 787, + 532 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Results", + "text_level": 1, + "bbox": [ + 215, + 556, + 323, + 569 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Synthetic Data Experiments. We begin with the synthetic experiments where we utilize 3-bit quanta frames, simulated using the parameters mentioned in Sec. 5.1 at 3.25, 9.75, 19.5, and 26 PPP to test the algorithms' performance. Tab. 2 and Tab. 3 demonstrate the PSNR and SSIM [76] of various methods extracted by predicting 6017 I2-2000FPS frames and 345 X4K1000FPS frames. To further substantiate the efficacy of QUIVER's design, we introduced a scaled-down variant, QUIVER-s (Refer Fig. 8(b) for complexity comparison). Quantitative results indicate that both QUIVER and QUIVER-s offer substantially better performance than all the baselines across a range of light levels. Fig. 9 depicts visual results of all the methods on the I2-2000FPS dataset. It is evident that existing methods fail to handle both motion and noise simultaneously, whereas, our proposed method, QUIVER, produces blur free high SNR outputs while preserving high-frequency details to a large extent.", + "bbox": [ + 212, + 583, + 787, + 779 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Real Data Experiments. We verify the methods' performance on real data. The real data is collected as binary frames using a SPAD sensor [17] at 10000 FPS with a spatial resolution of $240 \\times 320$ . As SPADs possess zero read noise, the binary frames are summed up to generate 3-bit frames. The average observed", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "P. Chennuri et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/cff3bba01ea1337b397e0304b92466ad35f557fb8d6c33137b0b498d0262c008.jpg", + "table_caption": [ + "Table 2: Performance comparison on the proposed I2-2000FPS dataset across various light levels. Models are trained using the I2-2000FPS dataset. QUIVER performs significantly better than the existing methods." + ], + "table_footnote": [], + "table_body": "
Photons-Per-Pixel (PPP)3.259.7519.526
MethodPSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑
Transform Denoise [6]21.31700.718423.15210.767122.77480.781222.30960.7811
QBP [47]15.94110.129319.18560.265420.40000.371320.79780.4114
QBP (+ BM3D [14])21.54760.703322.20010.689922.83510.769622.86170.7832
Student-Teacher [10]18.72000.400616.51950.247915.76360.213313.28890.0735
RVRT [42]19.41150.353921.67140.456822.08260.502121.75280.4968
EMVD [2]20.01940.587321.05590.604822.44030.559223.40530.5576
FloRNN [1]21.03410.678525.61320.709127.43220.739527.85200.7784
MemDeblur [35]19.48770.386814.49060.111216.17750.166716.00580.1712
Spk2ImgNet [85]20.39450.564219.66650.673322.93720.700814.97690.6861
QUIVER-s (Ours)24.70130.756526.86760.788327.29890.843227.86590.8408
QUIVER (Ours)26.21430.789726.80580.825027.75380.856327.93770.8446
", + "bbox": [ + 305, + 198, + 697, + 319 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/ac76287c574ff91fc7026f099a91257e06d273744203e01a778fb1009eeb1152.jpg", + "table_caption": [ + "Table 3: Performance comparison on the X4K1000FPS dataset across various light levels. Models are trained using the I2-2000FPS dataset. QUIVER performs significantly better than the existing methods." + ], + "table_footnote": [], + "table_body": "
Photons-Per-Pixel (PPP)3.259.7519.526
MethodPSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑
Transform Denoise [6]19.62550.632322.17030.704422.99380.722922.62300.7204
QBP [47]15.56340.230216.97580.323017.17980.395717.78070.4188
QBP (+ BM3D [14])17.96770.512318.53080.522618.24070.541418.79170.5586
Student-Teacher [10]18.82080.365216.15480.260814.93590.257113.97620.1186
RVRT [42]19.92030.364121.07810.447221.47800.492520.78990.4919
EMVD [2]20.51020.483621.81520.559522.94400.593622.45870.5860
FloRNN [1]20.82830.577823.58740.648424.32140.668325.24830.7170
MemDeblur [35]19.55340.364214.55950.220316.67490.311615.64960.2974
Spk2ImgNet [85]18.94240.473119.25320.572220.34420.571616.09310.6106
QUIVER-s (Ours)20.91970.595521.79900.652324.19240.731623.44110.7248
QUIVER (Ours)21.87300.652123.16540.705724.59560.764525.00860.7513
", + "bbox": [ + 305, + 387, + 697, + 507 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "light level after summation is 4.9 PPP. We generate results with networks trained at 4.9 PPP and demonstrate the visual results in Fig. 10. QUIVER, as opposed to existing state-of-the-art, effectively recovers high-frequency information while applying visually appealing smoothening effect to low-frequency regions of the scene. It is noteworthy that SPADs' image formation model is significantly different from that of the QIS' imaging model [58,59]. Therefore, the visual results also indicate that the proposed QUIVER can thoroughly generalize to various single-photon detectors.", + "bbox": [ + 212, + 536, + 787, + 657 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 684, + 401, + 702 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effect of Pre-Denoising, Optical Flow and Multi-Scale. We conduct an ablation study to evaluate the effect of the pre-denoising, the learnable optical flow, and multi-scale reconstruction on performance. Upon removal of either module, we expand the features dimension of layers in the feature extraction stage and add RDB blocks to the RMDF module, thereby maintaining a similar model capacity. We train all possible combinations and display the quantitative results in Tab. 4. Results indicate that, in the absence of either one or more modules, the network's performance is substantially worse. Visual Results in Fig. 11", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "QUIVER", + "bbox": [ + 666, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 127 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2c84702399a19cbd297424cb59aeb45ff3d7d64c55e5b4f6ec8446cfc03715c5.jpg", + "image_caption": [ + "Fig.11: Ablation Study. Visual Comparisons depicting the effectiveness of Predenoiser, Optical Flow, and Multi-Scale Reconstruction Modules. Best viewed in zoom." + ], + "image_footnote": [], + "bbox": [ + 217, + 145, + 785, + 231 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "indicate that these modules serve critical roles as they significantly contribute to model's performance.", + "bbox": [ + 212, + 297, + 782, + 325 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Does loading pre-trained optical flow module weights help? We initialize SPyNet with its pre-trained weights and finetune the same while training QUIVER, and display its quantitative results in Tab. 4. As the pre-trained SPyNet is not robust to photon shot noise and read noise, especially in low-light conditions, initializing the module with it will result in sub-optimal performance.", + "bbox": [ + 212, + 327, + 782, + 402 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/f31605512d70ecdecaee73ed5775e4c92c150a7b00d964a22448fa5ad22fdee6.jpg", + "table_caption": [ + "Table 4: Ablation study. We conduct experiments to emphasize the role of Denoiser, Optical flow, and Multi-Scale reconstruction modules. We also show the effect of loading pretrained optical flow weights $(^{*})$ on performance." + ], + "table_footnote": [], + "table_body": "
Pre-Denoising Optical Flow Multi-ScaleI2-2000FPS
PSNR↑SSIM↑
×××23.67020.7756
✓✓*✓23.94790.7709
✓✓×24.38410.7808
××✓24.74450.7755
×✓✓24.99990.7753
✓×✓25.75210.7760
✓✓✓26.21430.7897
", + "bbox": [ + 372, + 494, + 630, + 595 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 645, + 359, + 660 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we presented a methodology to reconstruct blur-free grayscale images/videos captured using 1-bit or few-bit quanta data. While adopting the ideology of classical quanta restoration methods, we proposed an end-to-end deep learning framework, QUIVER, that utilizes pre-filtering, learnable optical flow module, and a novel multi-scale reconstruction approach to produce high-visual outputs. Experiments on synthetic and real data indicate QUIVER beats state-of-the-art and can generalize across single-photon sensors. We also introduce the world's first high-speed video dataset, I2-2000FPS, that captures fast moving scenes at 2000 fps, covering wide ranges of motion. We believe that I2-2000FPS will be a valuable asset for researchers in high-speed motion analysis and other computer vision tasks.", + "bbox": [ + 212, + 672, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "P. Chennuri et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. The work is supported, in part, by Samsung Research America (SRA), DARPA / SRC CogniSense JUMP 2.0 Center, NSF IIS-2133032, and NSF ECCS-2030570. The authors would like to thank German Mora Martin for his help in recording the SPAD data.", + "bbox": [ + 215, + 146, + 787, + 205 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 229, + 321, + 244 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Li et al, J.: Unidirectional Video Denoising by Mimicking Backward Recurrent Modules with Look-Ahead Forward Ones. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) ECCV. Lecture Notes in Computer Science (2022). https://doi.org/10.1007/978-3-031-19797-0_34", + "2. Maggioni et al, M.: Efficient Multi-Stage Video Denoising With Recurrent Spatio-Temporal Fusion. In: CVPR (2021)", + "3. Arias, P., Morel, J.M.: Video Denoising via Empirical Bayesian Estimation of Space-Time Patches. Journal of Mathematical Imaging and Vision 60(1), 70-93 (Jan 2018). https://doi.org/10.1007/s10851-017-0742-4", + "4. Chan, K.C.K., Zhou, S., Xu, X., Loy, C.C.: BasicVSR++: Improving Video SuperResolution With Enhanced Propagation and Alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5972-5981 (2022)", + "5. Chan, S.H.: What Does a One-Bit Quanta Image Sensor Offer? IEEE Transactions on Computational Imaging 8, 770-783 (2022). https://doi.org/10.1109/TCI.2022.3202012", + "6. Chan, S.H., Elgendy, O.A., Wang, X.: Images from Bits: Non-Iterative Image Reconstruction for Quanta Image Sensors. Sensors 16(11), 1961 (Nov 2016). https://doi.org/10.3390/s16111961", + "7. Chan, S.H., Lu, Y.M.: Efficient image reconstruction for gigapixel quantum image sensors. In: 2014 IEEE Global Conference on Signal and Information Processing (GlobalSIP). pp. 312-316 (Dec 2014). https://doi.org/10.1109/GlobalSIP.2014.7032129", + "8. Chan, S.H., Wang, X., Elgendy, O.A.: Plug-and-Play ADMM for Image Restoration: Fixed-Point Convergence and Applications. IEEE Transactions on Computational Imaging 3(1), 84-98 (Mar 2017). https://doi.org/10.1109/TCI.2016.2629286", + "9. Charbon, E., Fishburn, M., Walker, R., Henderson, R.K., Niclass, C.: SPAD-Based Sensors. In: Remondino, F., Stoppa, D. (eds.) TOF Range-Imaging Cameras, pp. 11-38. Springer Berlin Heidelberg, Berlin, Heidelberg (2013). https://doi.org/10.1007/978-3-642-27523-4_2", + "10. Chi, Y., Gnanasambandam, A., Koltun, V., Chan, S.H.: Dynamic Low-Light Imaging with Quanta Image Sensors. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision - ECCV 2020, vol. 12366, pp. 122-138. Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-58589-1_8", + "11. Chi, Y., Zhang, X., Chan, S.H.: HDR imaging with spatially varying signal-to-noise ratios. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5724-5734. IEEE (2023). https://doi.org/10.1109/CVPR52729.2023.00554, https://doi.org/10.1109/CVPR52729.2023.00554", + "12. Choi, J.H., Elgendy, O.A., Chan, S.H.: Image Reconstruction for Quanta Image Sensors Using Deep Neural Networks. In: 2018 IEEE International Conference on" + ], + "bbox": [ + 218, + 258, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "QUIVER", + "bbox": [ + 666, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Acoustics, Speech and Signal Processing (ICASSP). pp. 6543-6547 (Apr 2018). https://doi.org/10.1109/ICASSP.2018.8461685", + "13. Claus, M., van Gemert, J.: ViDeNN: Deep Blind Video Denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019)", + "14. Dabov, K., Foi, A., Katkovnik, V., Egiazarian, K.: Image denoising by sparse 3-D transform-domain collaborative filtering. IEEE Transactions on image processing 16(8), 2080-2095 (2007)", + "15. Dong, S., Huang, T., Tian, Y.: Spike camera and its coding methods. arXiv preprint arXiv:2104.04669 (2021)", + "16. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "17. Dutton, N.A.W., Gyongy, I., Parmesan, L., Gnecchi, S., Calder, N., Rae, B.R., Pellegrini, S., Grant, L.A., Henderson, R.K.: A SPAD-Based QVGA Image Sensor for Single-Photon Counting and Quanta Imaging. IEEE Transactions on Electron Devices 63(1), 189–196 (Jan 2016). https://doi.org/10.1109/TED.2015.2464682", + "18. Dutton, N.A., Gyongy, I., Parmesan, L., Henderson, R.K.: Single photon counting performance and noise analysis of CMOS SPAD-based image sensors. Sensors 16(7), 1122 (2016)", + "19. Elgendy, O.A., Chan, S.H.: Optimal Threshold Design for Quanta Image Sensor. IEEE Transactions on Computational Imaging 4(1), 99-111 (Mar 2018). https://doi.org/10.1109/TCI.2017.2781185", + "20. Feng Yang, Lu, Y.M., Sbaiz, L., Vetterli, M.: Bits From Photons: Oversampled Image Acquisition Using Binary Poisson Statistics. IEEE Transactions on Image Processing 21(4), 1421-1436 (Apr 2012). https://doi.org/10.1109/TIP.2011.2179306", + "21. Fossum, E., Ma, J., Masoodian, S., Anzagira, L., Zizza, R.: The Quanta Image Sensor: Every Photon Counts. Sensors 16(8), 1260 (Aug 2016). https://doi.org/10.3390/s16081260", + "22. Fossum, E.R.: Modeling the performance of single-bit and multi-bit quanta image sensors. IEEE Journal of the Electron Devices Society 1(9), 166-174 (2013)", + "23. Gao, J., Shang, Z., Nie, K., Luo, T.: High dynamic range image reconstruction for multi-bit quanta image sensor. Optoelectronics Letters 18(9), 553-558 (Sep 2022). https://doi.org/10.1007/s11801-022-2014-9", + "24. Gariepy, G., Krstajic, N., Henderson, R., Li, C., Thomson, R.R., Buller, G.S., Heshmat, B., Raskar, R., Leach, J., Faccio, D.: Single-photon sensitive light-infight imaging. Nature Communications 6(1), 6021 (Jan 2015). https://doi.org/10.1038/ncomms7021", + "25. Gariepy, G., Leach, J., Warburton, R., Chan, S., Henderson, R., Faccio, D.: Picosecond time-resolved imaging using SPAD cameras. In: Lewis, K.L., Hollins, R.C. (eds.) SPIE Security + Defence. p. 99920N. Edinburgh, United Kingdom (Oct 2016). https://doi.org/10.1117/12.2241184", + "26. Gnanasambandam, A., Chan, S.H.: HDR Imaging With Quanta Image Sensors: Theoretical Limits and Optimal Reconstruction. IEEE Transactions on Computational Imaging 6, 1571-1585 (2020). https://doi.org/10.1109/TCI.2020.3041093", + "27. Gnanasambandam, A., Chan, S.H.: Image Classification in the Dark Using Quanta Image Sensors. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Proceed-" + ], + "bbox": [ + 217, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "P. Chennuri et al.", + "bbox": [ + 271, + 114, + 390, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ings of the European Conference on Computer Vision (ECCV). pp. 484-501. Lecture Notes in Computer Science, Cham (2020). https://doi.org/10.1007/978-3-030-58598-3_29", + "28. Gnanasambandam, A., Chan, S.H.: Exposure-Referred Signal-to-Noise Ratio for Digital Image Sensors. IEEE Transactions on Computational Imaging 8, 561-575 (2022). https://doi.org/10.1109/TCI.2022.3187657", + "29. Godard, C., Matzen, K., Uytendaele, M.: Deep Burst Denoising. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 538-554 (2018)", + "30. Gupta, A., Ingle, A., Velten, A., Gupta, M.: Photon-Flooded Single-Photon 3D Cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6770-6779 (2019)", + "31. Gupta, S., Gupta, M.: Eulerian Single-Photon Vision. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10465–10476 (2023)", + "32. Gutierrez-Barragan, F., Ingle, A., Seets, T., Gupta, M., Velten, A.: Compressive Single-Photon 3D Cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17854-17864 (2022)", + "33. Gyongy, I., Dutton, N., Henderson, R.: Single-Photon Tracking for High-Speed Vision. Sensors 18(2), 323 (Jan 2018). https://doi.org/10.3390/s18020323", + "34. Hasinoff, S.W., Sharlet, D., Geiss, R., Adams, A., Barron, J.T., Kainz, F., Chen, J., Levoy, M.: Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM Transactions on Graphics 35(6), 1-12 (Nov 2016). https://doi.org/10.1145/2980179.2980254", + "35. Ji, B., Yao, A.: Multi-Scale Memory-Based Video Deblurring. In: CVPR (2022)", + "36. Kiani Galoogahi, H., Fagg, A., Huang, C., Ramanan, D., Lucey, S.: Need for Speed: A Benchmark for Higher Frame Rate Object Tracking. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1125-1134 (2017)", + "37. Kingma, D., Ba, J.: Adam: A Method for Stochastic Optimization. In: International Conference on Learning Representations (ICLR) (2015). https://doi.org/10.48550/arXiv.1412.6980", + "38. Lebrun, M., Buades, A., Morel, J.M.: A Nonlocal Bayesian Image Denoising Algorithm. SIAM Journal on Imaging Sciences 6(3), 1665-1688 (Jan 2013). https://doi.org/10.1137/120874989", + "39. Lecun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (Nov/1998). https://doi.org/10.1109/5.726791", + "40. Li, C., Qu, X., Gnanasambandam, A., Elgendy, O.A., Ma, J., Chan, S.H.: Photon-limited object detection using non-local feature matching and knowledge distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3976-3987 (2021)", + "41. Liang, J., Cao, J., Fan, Y., Zhang, K., Ranjan, R., Li, Y., Timofte, R., Van Gool, L.: VRT: A Video Restoration Transformer (Jun 2022). https://doi.org/10.48550/arXiv.2201.12288", + "42. Liang, J., Fan, Y., Xiang, X., Ranjan, R., Ilg, E., Green, S., Cao, J., Zhang, K., Timofte, R., Gool, L.V.: Recurrent Video Restoration Transformer with Guided Deformable Attention. Advances in Neural Information Processing Systems 35, 378-393 (Dec 2022)", + "43. Liba, O., Murthy, K., Tsai, Y.T., Brooks, T., Xue, T., Karnad, N., He, Q., Barron, J.T., Sharlet, D., Geiss, R., Hasinoff, S.W., Pritch, Y., Levoy, M.: Handheld mobile photography in very low light. ACM Transactions on Graphics 38(6), 1-16 (Dec 2019). https://doi.org/10.1145/3355089.3356508" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "QUIVER", + "bbox": [ + 666, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "44. Lindell, D.B., O'Toole, M., Wetzstein, G.: Single-photon 3D imaging with deep sensor fusion. ACM Transactions on Graphics 37(4), 113-1 (2018)", + "45. Ma, J., Chan, S., Fossum, E.R.: Review of Quanta Image Sensors for UltralowLight Imaging. IEEE Transactions on Electron Devices 69(6), 2824-2839 (Jun 2022). https://doi.org/10.1109/TED.2022.3166716", + "46. Ma, J., Masoodian, S., Starkey, D.A., Fossum, E.R.: Photon-number-resolving megapixel image sensor at room temperature without avalanche gain. Optica 4(12), 1474-1481 (Dec 2017). https://doi.org/10.1364/OPTICA.4.001474", + "47. Ma, S., Gupta, S., Ulku, A.C., Bruschini, C., Charbon, E., Gupta, M.: Quanta burst photography. ACM Transactions on Graphics 39(4) (Aug 2020). https://doi.org/10.1145/3386569.3392470", + "48. Ma, S., Mos, P., Charbon, E., Gupta, M.: Burst Vision Using Single-Photon Cameras. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5375-5385 (2023)", + "49. Madhusudana, P.C., Yu, X., Birkbeck, N., Wang, Y., Adsumilli, B., Bovik, A.C.: Subjective and Objective Quality Assessment of High Frame Rate Videos. IEEE Access 9, 108069-108082 (2021). https://doi.org/10.1109/ACCESS.2021.3100462", + "50. Maggioni, M., Boracchi, G., Foi, A., Egiazarian, K.: Video Denoising, Deblocking, and Enhancement Through Separable 4-D Nonlocal Spatiotemporal Transforms. IEEE Transactions on Image Processing 21(9), 3952-3966 (Sep 2012). https://doi.org/10.1109/TIP.2012.2199324", + "51. Monakhova, K., Richter, S.R., Waller, L., Koltun, V.: Dancing Under the Stars: Video Denoising in Starlight. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16241-16251 (2022)", + "52. Niclass, C., Rochas, A., Besse, P.A., Charbon, E.: Design and characterization of a CMOS 3-D image sensor based on single photon avalanche diodes. IEEE Journal of Solid-State Circuits 40(9), 1847-1854 (Sep 2005). https://doi.org/10.1109/JSSC.2005.848173", + "53. Pan, L., Scheerlinck, C., Yu, X., Hartley, R., Liu, M., Dai, Y.: Bringing a Blur Frame Alive at High Frame-Rate With an Event Camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6820-6829 (2019)", + "54. Pearl, N., Treibitz, T., Korman, S.: NAN: Noise-Aware NeRFs for Burst-Denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12672-12681 (2022)", + "55. Plotz, T., Roth, S.: Benchmarking Denoising Algorithms With Real Photographs. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1586-1595 (2017)", + "56. Qu, X., Chi, Y., Chan, S.H.: Spatially varying exposure with 2-by-2 multiplexing: Optimality and universality. IEEE Transactions on Computational Imaging 10, 261-276 (2024)", + "57. Ranjan, A., Black, M.J.: Optical Flow Estimation Using a Spatial Pyramid Network. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4161-4170 (2017)", + "58. Rapp, J., Goyal, V.K.: A Few Photons Among Many: Unmixing Signal and Noise for Photon-Efficient Active Imaging. IEEE Transactions on Computational Imaging 3(3), 445-459 (Sep 2017). https://doi.org/10.1109/TCI.2017.2706028", + "59. Rapp, J., Ma, Y., Dawson, R.M.A., Goyal, V.K.: Dead Time Compensation for High-Flux Ranging. IEEE Transactions on Signal Processing 67(13), 3471-3486 (Jul 2019). https://doi.org/10.1109/TSP.2019.2914891" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "P. Chennuri et al.", + "bbox": [ + 271, + 114, + 390, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "60. Rebecq, H., Ranftl, R., Koltun, V., Scaramuzza, D.: High Speed and High Dynamic Range Video with an Event Camera. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(6), 1964-1980 (Jun 2021). https://doi.org/10.1109/TPAMI.2019.2963386", + "61. Remez, T., Litany, O., Bronstein, A.: A picture is worth a billion bits: Real-time image reconstruction from dense binary threshold pixels. In: 2016 IEEE International Conference on Computational Photography (ICCP). pp. 1-9 (May 2016). https://doi.org/10.1109/ICCPHOT.2016.7492874", + "62. Rochas, A.: Single photon avalanche diodes in CMOS technology. Tech. rep., Cite-seer (2003)", + "63. Ruget, A., Tyler, M., Mora Martin, G., Scholes, S., Zhu, F., Gyongy, I., Hearn, B., McLaughlin, S., Halimi, A., Leach, J.: Pixels2Pose: Super-resolution time-of-flight imaging for 3D pose estimation. Science Advances 8(48), eade0123 (Nov 2022). https://doi.org/10.1126/sciadv.ade0123", + "64. Seets, T., Ingle, A., Laurenzis, M., Velten, A.: Motion Adaptive Deblurring With Single-Photon Cameras. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1945–1954 (2021)", + "65. Shazeer, N.: Glu variants improve transformer. arXiv preprint arXiv:2002.05202 (2020)", + "66. Shin, D., Xu, F., Venkatraman, D., Lussana, R., Villa, F., Zappa, F., Goyal, V.K., Wong, F.N.C., Shapiro, J.H.: Photon-efficient imaging with a single-photon camera. Nature Communications 7(1), 12046 (Jun 2016). https://doi.org/10.1038/ncomms12046", + "67. Sim, H., Oh, J., Kim, M.: XVFI: eXtreme Video Frame Interpolation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14489-14498 (2021)", + "68. Su, S., Delbracio, M., Wang, J., Sapiro, G., Heidrich, W., Wang, O.: Deep Video Deblurring for Hand-Held Cameras. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1279-1288 (2017)", + "69. Tassano, M., Delon, J., Veit, T.: DVDNET: A Fast Network for Deep Video Denoising. In: 2019 IEEE International Conference on Image Processing (ICIP). pp. 1805-1809. IEEE, Taipei, Taiwan (Sep 2019). https://doi.org/10.1109/ICIP.2019.8803136", + "70. Tassano, M., Delon, J., Veit, T.: FastDVDnet: Towards Real-Time Deep Video Denoising Without Flow Estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1354-1363 (2020)", + "71. Vaksman, G., Elad, M., Milanfar, P.: Patch Craft: Video Denoising by Deep Modeling and Patch Matching. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2157-2166 (2021)", + "72. Voeikov, R., Falaleev, N., Baikulov, R.: TTNet: Real-Time Temporal and Spatial Video Analysis of Table Tennis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 884-885 (2020)", + "73. Vogels, T., Rousselle, F., Mcwilliams, B., Rothlin, G., Harvill, A., Adler, D., Meyer, M., Novak, J.: Denoising with kernel prediction and asymmetric loss functions. ACM Transactions on Graphics 37(4), 124:1-124:15 (Jul 2018). https://doi.org/10.1145/3197517.3201388", + "74. Wang, W., Chen, X., Yang, C., Li, X., Hu, X., Yue, T.: Enhancing Low Light Videos by Exploring High Sensitivity Camera Noise. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4111-4119 (2019)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "QUIVER", + "bbox": [ + 666, + 114, + 730, + 127 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "75. Wang, X.: Single-Photon Cameras Image Reconstruction Using Vision Transformer. In: 2023 IEEE 3rd International Conference on Computer Communication and Artificial Intelligence (CCAI). pp. 296-300 (May 2023). https://doi.org/10.1109/CCAI57533.2023.10201259", + "76. Wang, Z., Bovik, A., Sheikh, H., Simoncelli, E.: Image Quality Assessment: From Error Visibility to Structural Similarity. IEEE Transactions on Image Processing 13(4), 600-612 (Apr 2004). https://doi.org/10.1109/TIP.2003.819861", + "77. Wong, H.T., Leung, C.S., Ho, D.: Theoretical analysis and image reconstruction for multi-bit quanta image sensors. Signal Processing 185, 108087 (Aug 2021). https://doi.org/10.1016/j.sigpro.2021.108087", + "78. Yang, F., Lu, Y.M., Sbaiz, L., Vetterli, M.: An optimal algorithm for reconstructing images from binary measurements. In: Bouman, C.A., Pollak, I., Wolfe, P.J. (eds.) IS&T/SPIE Electronic Imaging. p. 75330K. San Jose, California (Feb 2010). https://doi.org/10.1117/12.850887", + "79. Yang, F., Sbaiz, L., Charbon, E., Susstrunk, S., Vetterli, M.: Image reconstruction in the gigavision camera. In: 2009 IEEE 12th International Conference on Computer Vision Workshops, ICCV Workshops. pp. 2212-2219 (Sep 2009). https://doi.org/10.1109/ICCVW.2009.5457554", + "80. Zhang, D., Lian, Q., Su, Y., Ren, T.: Dual-Prior Integrated Image Reconstruction for Quanta Image Sensors Using Multi-Agent Consensus Equilibrium. IEEE/CAA Journal of Automatica Sinica 10(6), 1407-1420 (Jun 2023). https://doi.org/10.1109/JAS.2023.123390", + "81. Zhang, D., Lian, Q., Yang, Y.: TwP: Two-stage projection framework with manifold constraint for image reconstruction. Digital Signal Processing 141, 104186 (Sep 2023). https://doi.org/10.1016/j.dsp.2023.104186", + "82. Zhang, Y., Li, K., Li, K., Wang, L., Zhong, B., Fu, Y.: Image Super-Resolution Using Very Deep Residual Channel Attention Networks. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 286-301 (2018)", + "83. Zhang, Y., Tian, Y., Kong, Y., Zhong, B., Fu, Y.: Residual Dense Network for Image Restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(7), 2480-2495 (Jul 2021). https://doi.org/10.1109/TPAMI.2020.2968521", + "84. Zhao, J., Xiong, R., Huang, T.: High-Speed Motion Scene Reconstruction for Spike Camera via Motion Aligned Filtering. In: 2020 IEEE International Symposium on Circuits and Systems (ISCAS). pp. 1-5. IEEE, Seville, Spain (Oct 2020). https://doi.org/10.1109/ISCAS45731.2020.9181055", + "85. Zhao, J., Xiong, R., Liu, H., Zhang, J., Huang, T.: Spk2imgnet: Learning to reconstruct dynamic scene from continuous spike stream. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11996-12005 (2021)", + "86. Zhao, J., Xiong, R., Xie, J., Shi, B., Yu, Z., Gao, W., Huang, T.: Reconstructing Clear Image for High-Speed Motion Scene With a Retina-Inspired Spike Camera. IEEE Transactions on Computational Imaging 8, 12-27 (2022). https://doi.org/10.1109/TCI.2021.3136446" + ], + "bbox": [ + 215, + 147, + 785, + 742 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "P. Chennuri et al.", + "bbox": [ + 271, + 114, + 390, + 127 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/2024/Quanta Video Restoration/c31fed1f-5357-4c1e-adf0-39121e1936f9_model.json b/2024/Quanta Video Restoration/c31fed1f-5357-4c1e-adf0-39121e1936f9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..377d69fae98f6de9143827b32d41c460a6d9f9be --- /dev/null +++ b/2024/Quanta Video Restoration/c31fed1f-5357-4c1e-adf0-39121e1936f9_model.json @@ -0,0 +1,2902 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.351, + 0.143, + 0.654, + 0.163 + ], + "angle": 0, + "content": "Quanta Video Restoration" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.189, + 0.775, + 0.237 + ], + "angle": 0, + "content": "Prateek Chennuri1, Yiheng Chi1, Enze Jiang1, G.M. Dilshan Godaliyadda2, Abhiram Gnanasambandam2, Hamid R. Sheikh2, Istvan Gyongy3, and Stanley H. Chan1" + }, + { + "type": "text", + "bbox": [ + 0.433, + 0.248, + 0.571, + 0.261 + ], + "angle": 0, + "content": "1 Purdue University" + }, + { + "type": "text", + "bbox": [ + 0.403, + 0.262, + 0.601, + 0.276 + ], + "angle": 0, + "content": "Samsung Research America" + }, + { + "type": "text", + "bbox": [ + 0.414, + 0.276, + 0.59, + 0.289 + ], + "angle": 0, + "content": "3 University of Edinburgh" + }, + { + "type": "text", + "bbox": [ + 0.33, + 0.291, + 0.675, + 0.304 + ], + "angle": 0, + "content": "{pchennur,chi14,jiang708,stanchan}@purdue.edu" + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.305, + 0.668, + 0.318 + ], + "angle": 0, + "content": "{dilshan.g,abhiram.g,hr.sheikh}@samsung.com" + }, + { + "type": "text", + "bbox": [ + 0.402, + 0.319, + 0.603, + 0.332 + ], + "angle": 0, + "content": "{igyongy2}@exseed.ed.ac.uk" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.346, + 0.368, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.399, + 0.336, + 0.406 + ], + "angle": 0, + "content": "(a) Ground Truth 16-bit" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.347, + 0.501, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.385, + 0.399, + 0.483, + 0.406 + ], + "angle": 0, + "content": "(b) CMOS 16-bit, 60 fps at 1 lux" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.347, + 0.635, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.399, + 0.618, + 0.406 + ], + "angle": 0, + "content": "(c) CMOS 16-bit, 240 fps at 1 lux" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.347, + 0.769, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.649, + 0.399, + 0.754, + 0.406 + ], + "angle": 0, + "content": "(d) CMOS 16-bit, 2000 fps at 1 lux" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.407, + 0.368, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.459, + 0.351, + 0.467 + ], + "angle": 0, + "content": "(e) Quanta 3-bit, 2000 fps at 1 lux" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.407, + 0.501, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.372, + 0.46, + 0.496, + 0.473 + ], + "angle": 0, + "content": "f) Quanta 3-bit, reconstructed from 11-frame avg. using Traditional Methods [47]" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.407, + 0.635, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.46, + 0.631, + 0.473 + ], + "angle": 0, + "content": "(g) Quanta 3-bit, reconstructed from 66-frame avg, using Traditional Methods [47]" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.407, + 0.769, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.635, + 0.46, + 0.769, + 0.473 + ], + "angle": 0, + "content": "(h) Quanta 3-bit, reconstructed from 11-frame avg using QUIVER" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.49, + 0.788, + 0.574 + ], + "angle": 0, + "content": "Fig. 1: Goal of this paper. (a) Blur-free video frame of a moving car. (b)-(d) CMOS image sensor simulations using realistic sensor parameters. The strong shot noise and read noise \\((5.1\\mathrm{e}^{-}/\\mathrm{pix})\\) of CMOS sensor make the signal acquisition difficult. (e) With low read noise \\((0.2\\mathrm{e}^{-}/\\mathrm{pix})\\), low-bit single-photon detectors capture valuable information. (f)-(g) Existing state-of-the-art algorithm, QBP [47] cannot handle strong motion and noise. (h) The proposed algorithm, QUIVER, produces high quality results." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.592, + 0.741, + 0.8 + ], + "angle": 0, + "content": "Abstract. The proliferation of single-photon image sensors has opened the door to a plethora of high-speed and low-light imaging applications. However, data collected by these sensors are often 1-bit or few-bit, and corrupted by noise and strong motion. Conventional video restoration methods are not designed to handle this situation, while specialized quanta burst algorithms have limited performance when the number of input frames is low. In this paper, we introduce Quanta Video Restoration (QUIVER), an end-to-end trainable network built on the core ideas of classical quanta restoration methods, i.e., pre-filtering, flow estimation, fusion, and refinement. We also collect and publish I2-2000FPS, a high-speed video dataset with the highest temporal resolution of 2000 frames-per-second, for training and testing. On simulated and real data, QUIVER outperforms existing quanta restoration methods by a significant margin. Code and dataset available at https://github.com/ chennuripratek/Quanta_Video_Restoration-QUIVER" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.812, + 0.741, + 0.839 + ], + "angle": 0, + "content": "Keywords: Single Photon Detectors \\(\\cdot\\) Video Restoration \\(\\cdot\\) High-Speed Dataset" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "P. Chennuri et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.377, + 0.161 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.21, + 0.788, + 0.377 + ], + "angle": 0, + "content": "Over the past decade, the astonishing growth of single-photon detectors has fundamentally changed the landscape of computational imaging. With the invention and proliferation of quanta image sensors (QIS) [21] and single-photon avalanche diodes (SPAD) [52, 62], there is an unprecedented volume of new applications in low-light imaging [6, 10, 66], computer vision [27, 31, 40], high-speed videography [47, 48], time-of-flight sensing [32, 63], and 3D imaging [30, 44]. In most of these use cases, the main core question that lies is how to recover the image from the photon counts measured in the scene. Specifically, given a video stream of 1-bit or few-bit data captured from a scene involving moving objects, how do we reconstruct a gray-scale image/video while eliminating the noise without incurring motion blur?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.383, + 0.788, + 0.581 + ], + "angle": 0, + "content": "To give the reader a visual perspective of the problem scope, Fig. 1 depicts a blur-free video of a moving car. We simulate the captured images at 1 lux assuming 60 fps, 240 fps, and 2000 fps CMOS image sensors with realistic sensor specifications. As illustrated in the figure, the resulting CMOS outputs are either severely blurred due to strong motion or completely distorted by noise due to sparse photons. In the same figure, we demonstrate a simulated single-photon camera output (a 3-bit QIS in this case) where the content the largely preserved despite heavy noise. Upon utilizing state-of-the-art Quanta Burst Photography (QBP) [47] for reconstructing the frames, provided the motion is slow, a decent output can be obtained. However, as the temporal window narrows down, as shown in Fig. 1 (f), the noise remains. In this paper, we address this problem with a new algorithm, designed to remove the noise while avoiding distortions in the presence of fast motion while utilizing only a few frames." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.586, + 0.788, + 0.664 + ], + "angle": 0, + "content": "The core innovation of this paper is QUanta Video Restoration (QUIVER), a deep-learning based video restoration algorithm for quanta image data. QUIVER is specialized for few-bit data (3-bit) captured at thousands of frames-per-second (2000 fps) with an average motion range of 1 to 7 pixels per frame. The main contributions of this paper can be summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.697, + 0.787, + 0.775 + ], + "angle": 0, + "content": "- We propose QUIVER, an end-to-end trainable quanta (video) restoration method built by embracing the core ideas from traditional quanta restoration algorithms. On a comprehensive evaluation dataset containing both simulated and real data, QUIVER outperforms all methods we compared in this paper by a significant margin." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.78, + 0.787, + 0.842 + ], + "angle": 0, + "content": "- We introduce I2-2000FPS, the first high-speed video dataset with a temporal resolution of 2000 frames-per-second for training and testing image and video reconstruction neural networks. We captured a total of 280 high-speed videos covering 114 distinct scenes with ground truth and simulated 3-bit videos." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.697, + 0.787, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "QUIVER" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.145, + 0.725, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.287, + 0.788, + 0.373 + ], + "angle": 0, + "content": "Fig. 2: Motion Blur and SNR Trade-off. The effects of bit depth on SNR and motion blur are illustrated using real captures by a single-photon sensor. For the motion range we target, 3-bit single-photon detectors provide the best trade-off between blur and SNR. The images are captured using a 1-bit SPAD [17] at \\(10\\mathrm{k}\\) fps at an average photon level of 0.51 and 0.40 photons-per-pixel (PPP) per frame, respectively. Higher bit-depth outputs are generated through temporal frame averaging." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.387, + 0.787, + 0.416 + ], + "angle": 0, + "content": "Table 1: Frame-rate, motion, read-noise, and data-rate statistics for various bit-depths at the same exposure level." + }, + { + "type": "table", + "bbox": [ + 0.359, + 0.428, + 0.64, + 0.521 + ], + "angle": 0, + "content": "
Bit-DepthfpsMotion (pixels/frame)σread (/pixel/sec)Data-rate (Mb/sec)
110k0 - 12000 e-96
314282 - 3285.6 e-41.13
53236 - 1264.6 e-15.5
77825 - 3015.6 e-5.24
92070 - 804 e-1.73
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.548, + 0.37, + 0.566 + ], + "angle": 0, + "content": "2 Background" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.584, + 0.533, + 0.6 + ], + "angle": 0, + "content": "2.1 Few-bit Single-Photon Detectors" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.719 + ], + "angle": 0, + "content": "What is few-bit photon counting? Single-photon detectors (QIS and SPAD) differ from conventional CMOS pixels by their extraordinary photon counting capability. QIS uses a two-stage pump-gate technique and correlated double sampling to suppress the read noise, while SPAD uses avalanche multiplication to amplify the photocharge. In both cases, the sensors are capable of resolving photons up to a single-photon sensitivity. We refer readers interested in the sensor development of QIS and SPAD to consult, for example, [5,9,18,45]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Along with the single-photon detectors' unique capability to count individual photons, these devices can generate data at a bit-depth as low as 1-bit to as high as 16-bit or even more. However, higher bit-depth is accompanied by longer integration time. If the scene contains motion, longer integration time will eventually result in strong motion blurs as shown in Fig. 2. On the other hand, 1-bit sensing with high frame rates will result in motion-blur-free but extremely noisy images. Therefore, from a pure data acquisition perspective, there exists an optimal bit-depth with respect to the motion that will give us mini-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "P. Chennuri et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "mal/no motion-blur data with a minimum per-frame signal-to-noise ratio (SNR) required for good quality reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.314 + ], + "angle": 0, + "content": "How about 1-bit and reconstruct afterward? Readers familiar with single-photon counting may wonder whether we can collect as many 1-bit frames as possible and then process the data afterward. The problem is power consumption and data rate. Fixing the same level of exposure, as described in Tab. 1, a 1-bit video at \\(10\\mathrm{k}\\) fps would require \\(96\\mathrm{Mb / sec}\\) whereas a 9-bit video at 20 fps would only need \\(1.73\\mathrm{Mb / sec}\\). Another problem is read noise accumulation. For sensors with non-zero read noise (such as QIS), every frame contributes to a finite amount of read noise. The more frames we read, the more read noise we accumulate. Therefore, recording 1-bit is not always the best option." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.334, + 0.379, + 0.348 + ], + "angle": 0, + "content": "2.2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.357, + 0.788, + 0.462 + ], + "angle": 0, + "content": "Image and Video Denoising. Classical state-of-the-art methods utilize a non-local strategy to identify similar patches across an image/video [3, 38, 50]. Deep neural networks have been proven to be successful in producing high quality denoised outputs [13, 69-71]. Among these architectures, Vision Transformers [41, 42] have been rated the state-of-the-art in recent times. However, all these solutions make simplistic assumptions on noise statistics, thus failing to perform on real noisy images or videos [55]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.788, + 0.583 + ], + "angle": 0, + "content": "Coming to low light, Burst denoising [34, 43], where images are aligned, merged and denoised, is one of the most popular methods. However, these methods fail without robust alignment. To overcome this, a number of alternative solutions with learnable alignment modules have been proposed [29, 54, 73]. Recent solutions have focused on practical noise models that replicate real camera sensor noise, to produce visually appealing results [51, 74]. Nevertheless, the existing solutions utilize images captured using CMOS image sensors, resulting in a notably higher photon level compared to the one utilized in our study." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.794 + ], + "angle": 0, + "content": "SPADs, Event, and Spike Cameras. Gariepy et al. [24, 25] firstly introduced the utilization of Single Photon Avalanche Diodes (SPADs) at pico-second temporal resolution to capture light in motion. Gyongy et al. [33] demonstrated 2D motion tracking of rigid planar objects using SPADs at \\(10\\mathrm{k}\\) fps. Recently, Ma et al. [47] and Seets et al. [64] utilized SPADs in a passive imaging setting to capture motion in low illumination. However, all these methods utilize extremely high-temporal resolutions hindering the deployment of these sensors into consumer devices where bandwidth is the bottleneck. Event [53, 60] and Spike Cameras [84, 86] also have demonstrated their effectiveness in capturing high-speed motion. However, these cameras focus on luminance/brightness variation and record a spike only when variation is above a threshold (changes based on factors like temperature, event rate, etc.) [15, 60]. Therefore, unlike single-photon detectors (QIS and SPADs), these cameras are NOT designed for single-photon counting and cannot operate in extreme low-light conditions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "QIS Reconstruction. Reconstructing quanta images is a challenging task due to the underlying Poisson-Gaussian statistics. Initial solutions to this problem included utilizing standard gradient descent [78], greedy algorithms [79]," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "QUIVER" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.145, + 0.773, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.217, + 0.785, + 0.245 + ], + "angle": 0, + "content": "Fig. 3: Traditional Methods' Design. Depiction of existing classical quanta restoration algorithms' design philosophy. Best viewed in zoom." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.276, + 0.788, + 0.413 + ], + "angle": 0, + "content": "(ADMM) [7,8], among others [19,20,23,26,61,77]. Chan et al. [6] were the first to propose a non-iterative approach using Anscombe transform for reconstructing quanta images. Choi et al. [12] proposed the first end-to-end trainable deep neural network (DNN) for QIS reconstruction. Alternative DNN-based solutions include utilizing vision transformers [75], Dual Prior Integrated networks [80], among others [81]. Nonetheless, all these methods fail to produce good results when the scene is in motion. Chi et al. [10] is the only method which focuses on capturing dynamic scenes using QIS but only targets at slow motion (1 pixel/frame)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.438, + 0.631, + 0.456 + ], + "angle": 0, + "content": "3 QUanta VIDEo REstoration (QUIVER)" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.472, + 0.416, + 0.489 + ], + "angle": 0, + "content": "3.1 Design Philosophy" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.501, + 0.789, + 0.759 + ], + "angle": 0, + "content": "In this section we present the design of our proposed algorithm. We start by briefly reviewing the design of classical methods [33, 47] which, to an extent, have been successful in restoring quanta images. As shown in Fig. 3, classical methods' algorithm design can be divided into four stages: (1) computing sum images to increase SNR, (2) optical flow (or) transformation matrix estimation for aligning the input frames, (3) warping and linear combination for generating preliminary restored output, and (4) refinement for producing the final output. While the steps seem intuitive and straightforward, existing methods are heavily vulnerable to extreme noise and strong motion in the input frames primarily due to two reasons. (1) none of the stages are designed to handle extreme noise and strong motion simultaneously (will be discussed further). (2) Since all the stages are sequential yet independent of each other, it is difficult to obtain an optimal result for a wide range of noise and motion. Our proposed algorithm QUIVER, leverages the design philosophy of existing classical methods while designing each stage to simultaneously handle both noise and motion. Moreover, QUIVER is an end-to-end trainable model making all the stages inter-dependent, thus leading to good restoration outputs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.782, + 0.424, + 0.798 + ], + "angle": 0, + "content": "3.2 Design of QUIVER" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "QUIVER is a deep-learning-based video restoration method for quanta imaging. The design philosophy of QUIVER is to adopt the intuitive thoughts behind" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "P. Chennuri et al." + }, + { + "type": "image", + "bbox": [ + 0.25, + 0.147, + 0.457, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.348, + 0.27, + 0.365, + 0.282 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.147, + 0.761, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.27, + 0.627, + 0.281 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.295, + 0.789, + 0.379 + ], + "angle": 0, + "content": "Fig. 4: Traditional Methods' Limitations. (a) Traditional methods' [33, 47] predenoising/temporal-averaging fails in strong motion. It is visible in the restored images that an input with strong motion between the frames results in several artifacts in the output even though SNR levels are similar. (b) Traditional methods [47] utilize a patch-based pre-trained optical flow module similar to [34]. This optical flow module fails to compensate for motion in the presence of noise." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.396, + 0.782, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.556, + 0.79, + 0.597 + ], + "angle": 0, + "content": "Fig.5: The proposed QUIVER network. The corresponding stages of QUIVER, built by embracing the intuitive thoughts behind existing classical methods. Best viewed in zoom." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.785, + 0.674 + ], + "angle": 0, + "content": "classical quanta restoration methods and develop an end-to-end trainable, robust to noise and motion deep learning based framework, as shown in Fig. 5. Specifically, QUIVER can be divided into four main stages:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Pre-Denoising to improve SNR: Since the input quanta frames possess extreme noise, classical methods adopt naive averaging to increase the SNR and thereby predict better optical flows or transformation matrices. However, as shown in Fig. 4(a), the simple averaging is vulnerable to motion and will negatively impact subsequent processing, ultimately leading to distorted outputs. Simply eliminating this stage is not the solution, because it leads to poor optical flow estimation, resulting in over-smoothed outputs with lack of low-level intricate details, as shown in Fig. 5 and Fig. 11. Therefore, a preliminary denoising step robust to noise and motion is crucial. In QUIVER, we leverage a computational undemanding single image denoiser built using Residual Dense Blocks [83] (RDBs) to provide minimal pre-preprocessing of the input quanta" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "QUIVER" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.455, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.145, + 0.785, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.292, + 0.261, + 0.71, + 0.275 + ], + "angle": 0, + "content": "Fig. 6: Design of the proposed modules DC-GFU and RMDF." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.308, + 0.784, + 0.353 + ], + "angle": 0, + "content": "data. A multi-frame denoiser is not an option due to it computationally demanding nature. We use RDB based network due to its history in handling noise while preserving details with its simple yet effective design." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.354, + 0.787, + 0.55 + ], + "angle": 0, + "content": "Optical Flow Estimation + Feature Alignment using Deformable Convolution - Gated Fusion Unit: Classical methods utilize an off-the-shelf pre-trained optical flow estimation module or predict a transformation matrix to compensate for motion between the frames. The basic assumption behind such approaches is that the motion between the frames is limited and the SNR is high enough. However, when such assumption is not met, the motion compensation is sub-optimal, as shown in Fig. 4(b). As most state-of-the-art pre-trained optical flow estimators are optimized on the CMOS RGB sensor images, it leads to sub-optimal performance when applied on quanta frames. Eliminating the flow estimation step is not recommended since experiments reveal the critical role it plays in motion compensation, as shown in Fig. 5 and Fig. 11. For QUIVER, we deploy a learnable optical flow estimation module and utilize SpyNet [57] owing to its computational efficiency while using a multi-scale approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.786, + 0.597 + ], + "angle": 0, + "content": "We deploy 3D convolution blocks to extract multi-scale spatio-temporal features from both the noisy and denoised quanta frames. We reuse the noisy frames to compensate for any information lost in the pre-denoising stage." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.787, + 0.793 + ], + "angle": 0, + "content": "The estimated multi-scale robust-to-noise optical flows are utilised for feature-level alignment of the extracted multi-scale spatio-temporal features. We utilize the deformable convolution with residual offsets proposed by [4] to warp the features. Inspired by the superior performance of Gated Linear Units (GLUs) in Transformers [65] we design and add a GLU based multi-layer-perceptron layer with GeLU activation for efficiently fusing the aligned features extracted from both the noisy and denoised frames. As shown in Fig. 6(a), we name this deformable convolution-GLU combination as the DC-GFU module. At this fusion-stage each frame is processed separately and the fusion is performed only along the channel dimension. Recurrence is applied for the alignment stage across all the multi-scale features of each frame. For the fusion module, we do not employ recurrence, as different scale features capture distinct long-range dependencies owing to their varying receptive fields." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Deep Feature Fusion using Recurrent Multi-scale Residual Dense Feature Fusion Unit: Post warping we want to perform a robust-to-noise dense feature fusion while taking advantage of the temporal correlations among" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "P. Chennuri et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.508, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.145, + 0.783, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.306, + 0.285, + 0.696, + 0.3 + ], + "angle": 0, + "content": "Fig. 7: Design of the proposed modules TCA and RFRM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.338, + 0.788, + 0.595 + ], + "angle": 0, + "content": "the features of all the input frames and also the spatial correlations between the multi-scale features within the same frame. For this task, we design and propose a Recurrent Multi-scale Residual Dense Feature Fusion Unit (RMDF) as shown in Fig. 6(b). The recurrence comes from the fact that the same RMDF module is applied progressively to all the frames' features. For any frame \\( t \\), the RMDF takes in the corresponding frame's multi-scale features \\( \\{\\mathbf{F}_t^1, \\mathbf{F}_t^2, \\mathbf{F}_t^4\\} \\), bilinearly interpolated noisy frames \\( \\{\\mathbf{I}_t^1, \\mathbf{I}_t^2, \\mathbf{I}_t^4\\} \\) and a hidden state \\( \\mathbf{h}_{t-1} \\) as inputs. The multi-scale features are progressively fused in a feed-forward fashion, with Residual Dense Block (RDB) as the basic block, to effectively extract both the short and long range dependencies required for good reconstruction. As shown in Fig. 6(b), multi-scale features are extracted from the noisy frames and fused with the other corresponding input features to minimize any errors accumulated through the previous stages. While these features are utilized to exploit the spatial correlations within the frame, the hidden state \\( \\mathbf{h} \\) captures the temporal correlations between all the input frames. Thus, the design of RMDF enables it to extract densely fused multi-scale spatio-temporal features required for enhanced quality outputs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Multi-Scale Reconstruction using Residual Frame Refinement Module: Considering the heavy noise in the input quanta frames, this ill-posed problem's restored image subspace can be quite large. To output a restored image close to the ground truth we prefer deep supervision that lets the model preserve critical details of the scene. We opt for a multi-scale reconstruction approach where the image at each scale is reconstructed in a progressive fashion. Experiments reveal the efficacy of this approach (Fig. 5, Fig. 11). The stage starts with the lowest scale features extracted from all the frames using RMDF being concatenated and sent into a newly designed Temporal Cross Attention (TCA) module. As shown in Fig. 7(a), the TCA module is similar to the multi-head attention in vision transformers [16] in terms of generating queries, keys, and values. However, we maintain the number of heads to be one and apply attention only on the channel dimension. The cross attention is from the fact that we input features extracted from all the input frames. The extracted cross-attention features are then fed into a newly designed Residual Frame Refinement Module (RFRM). As shown in Fig. 7(b), RFRM takes in a residual frame \\(\\mathbf{r}_t^\\alpha\\), a" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "QUIVER" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.285 + ], + "angle": 0, + "content": "hidden state \\(\\mathbf{f}_t^\\alpha\\), and the features as input, concatenates the hidden state with the features to input into the channel attention block [82] to emphasize critical spatio-temporal information. Further, we divide the module into 2 branches. While the former is designed to extract multi-scale spatial correlation information and output a modified hidden state \\(\\mathbf{f}_t^{\\alpha /2}\\), the latter focuses on refining the residual frame to output a corresponding scale reconstructed image \\(\\mathbf{O}_t^\\alpha\\) and a residual frame \\(\\mathbf{r}_t^{\\alpha /2}\\). The main purpose of this setup is to initially restore the high-level features through estimating \\(\\mathbf{O}_t^4\\), followed by focusing on the low-level intricate details while refining the residual frames for scales 2 and 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.287, + 0.784, + 0.317 + ], + "angle": 0, + "content": "Loss Function. We train QUIVER with a multi-scale loss. The overall loss function can be represented as" + }, + { + "type": "equation", + "bbox": [ + 0.242, + 0.327, + 0.785, + 0.366 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {Q}} = \\lambda_ {1} \\cdot \\mathcal {L} (\\mathbf {I} ^ {1, \\mathrm {G T}}, \\mathbf {I} _ {d} ^ {1}) + \\lambda_ {2} \\cdot \\mathcal {L} (\\mathbf {I} _ {t} ^ {1, \\mathrm {G T}}, \\mathbf {O} _ {t} ^ {1}) + \\lambda_ {3} \\cdot \\mathcal {L} (\\mathbf {I} _ {t} ^ {2, \\mathrm {G T}}, \\mathbf {O} _ {t} ^ {2}) + \\dots \\\\ \\lambda_ {4} \\cdot \\mathcal {L} \\left(\\mathbf {I} _ {t} ^ {4, \\mathrm {G T}}, \\mathbf {O} _ {t} ^ {4}\\right), \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.784, + 0.425 + ], + "angle": 0, + "content": "where \\(\\mathbf{I}_t^{\\alpha,\\mathrm{GT}}\\) is the captured \\(t^{\\mathrm{th}}\\) ground truth frame bicubically downsampled by \\(\\alpha\\), and \\(\\mathcal{L}(\\mathbf{I}_a, \\mathbf{I}_b) = ||\\mathbf{I}_a - \\mathbf{I}_b||_1 + ||\\nabla_x \\mathbf{I}_a - \\nabla_x \\mathbf{I}_b||_1 + ||\\nabla_y \\mathbf{I}_a - \\nabla_y \\mathbf{I}_b||_1\\). Here, \\(\\nabla_x\\) and \\(\\nabla_y\\) represent the operations of computing horizontal and vertical gradients." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.446, + 0.538, + 0.463 + ], + "angle": 0, + "content": "4 Proposed I2-2000FPS dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.789, + 0.596 + ], + "angle": 0, + "content": "While several high-frame-rate datasets have been open-sourced in recent times [36, 49, 60, 67, 68, 72], these datasets mainly feature videos tainted by severe motion blur, making them unsuitable in our problem setting. Moreover, features such as high motion speed and sufficient number of videos are also not always guaranteed. A visual representation of existing datasets' comparison is shown in Fig. 8(a). To address the gaps, we introduce the I2-2000FPS dataset, a high frame rate video collection meticulously designed to capture high-speed motion with precision." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.786, + 0.733 + ], + "angle": 0, + "content": "The I2-2000FPS dataset has a temporal resolution of 2000 FPS and a spatial resolution of \\(512 \\times 1024\\) pixels, comprising 280 unique videos spanning 114 diverse scenes. The videos are captured using the Chronos 1.4 high-speed CMOS sensor-based camera from Kron Technologies. Notably, I2-2000FPS incorporates dark current calibration, leveraging the camera's capabilities to mitigate dark current effects. Throughout the data collection process, analog and digital gain were consistently maintained at 0 dB to avoid amplification of noise. To minimize noise, the videos are exclusively captured outdoor with ambient lighting conditions. More details on I2-2000FPS can be found in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.756, + 0.377, + 0.772 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.786, + 0.463, + 0.801 + ], + "angle": 0, + "content": "5.1 Image Formation Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "For experiments involving synthetic data, we use a single-photon detector simulator based on an underlying image formation model discussed below. We build" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "P. Chennuri et al." + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.144, + 0.499, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.284, + 0.387, + 0.295 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.144, + 0.744, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.618, + 0.284, + 0.633, + 0.295 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.308, + 0.789, + 0.406 + ], + "angle": 0, + "content": "Fig. 8: (a) Benchmarking high-speed video datasets. Horizontal axis represents the temporal resolution and the vertical axis indicates the maximum speed captured by the dataset, assuming a fixed camera-object distance. The circles in blue and orange indicate blur and blur-free videos, respectively. (b) Benchmarking Quanta Video Restoration Models on the I2-2000FPS dataset. Horizontal axis represents the computational complexity in terms of GFLOPs and the vertical axis indicates the PSNR acquired at 3.25PPP." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.784, + 0.464 + ], + "angle": 0, + "content": "upon the prototype initially suggested in [46] adopted in prior works [6, 10, 11, 20, 26, 28, 56]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.788, + 0.587 + ], + "angle": 0, + "content": "Given the quanta exposure [22], \\(\\mathbf{I}^{\\mathrm{GT}}\\), dependent on the photon flux and exposure time, the observed signal by the sensor can be represented as a Poisson-Gaussian random variable, where the Poisson represents the photon arrival process and the Gaussian models the read noise. The readout process involves various sources of distortions and an Analog-to-Digital Converter (ADC) to convert the real numbers into integers \\(\\{0,1,2,\\dots,L\\}\\), where \\(L = 2^{\\mathrm{Nbits}} - 1\\) depending on the bit-depth (Nbits) allocated to the sensor. The final sensor readout, \\(\\mathbf{Y}\\), can be represented using the following equation," + }, + { + "type": "equation", + "bbox": [ + 0.273, + 0.596, + 0.787, + 0.633 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} \\sim \\mathrm {A D C} _ {[ 0, \\mathrm {L} ]} \\{\\mathrm {P o i s s o n (Q E \\times I ^ {G T} +} \\theta_ {\\mathrm {d a r k}}) + \\underbrace {\\mathrm {G a u s s} (0 , \\sigma_ {\\mathrm {r e a d}} ^ {2} \\mathbf {1})} _ {\\mathrm {r e a d n o i s e}} \\}. \\qquad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.788, + 0.705 + ], + "angle": 0, + "content": "Akin to previous works [6, 10, 20, 26, 28], we assume our sensor to be monochromatic as we utilize monochromatic real data in our experiments. For our sensor prototype, we utilize a Quantum Efficiency (QE) of 0.80. The dark current \\((\\theta_{\\mathrm{dark}})\\) and read noise \\((\\sigma_{\\mathrm{read}})\\) are set to \\(1.6\\mathrm{e}^{-}/\\mathrm{pix/sec}\\) and \\(0.2\\mathrm{e}^{-}/\\mathrm{pix}\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.725, + 0.448, + 0.74 + ], + "angle": 0, + "content": "5.2 Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Training data. We curate a set of 249 videos from the I2-2000FPS collection and employ it as the training dataset for all the deep-learning models in our experiments. Each training sample is fetched on the fly from each clip. A training sample here is defined as a tuple containing the ground-truth/target frames and the 3-bit quanta frames simulated at 3.25 photons-per-pixel (PPP) (\\(\\sim\\)1 lux assuming a \\(1.1\\mu \\mathrm{m}\\) pixel pitch and a 1/2000 second exposure time) using the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "QUIVER" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.144, + 0.758, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.539, + 0.79, + 0.594 + ], + "angle": 0, + "content": "Fig. 9: Visual comparisons of the reconstructed results on test videos from the proposed I2-2000FPS dataset. For fair comparison, all methods utilize 11 3-bit quanta frames simulated at 3.25 PPP per frame (\\(\\sim\\)1 lux) to produce a restored frame. Best viewed in zoom." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.54, + 0.644 + ], + "angle": 0, + "content": "image formation model described in Sec. 5.1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.788, + 0.748 + ], + "angle": 0, + "content": "Testing data. To effectively analyze the performance of various methods, we carefully sample 31 videos from I2-2000FPS containing various motion types, shapes, and speeds. To test the generalizability, we also test the algorithms on X4K1000FPS [67] test dataset containing 15 videos from distinct scenes. Lastly, to measure the performance on real-world data, we collect binary frames using a SPAD sensor [17] and compare the reconstructed outputs. More details will be discussed in Sec. 5.3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.803, + 0.841 + ], + "angle": 0, + "content": "Baselines. We compare the proposed method with eight existing dynamic scene reconstruction algorithms, namely Transform Denoise [6], QBP [47], Student-Teacher [10], RVRT [42], EMVD [2], FloRNN [1], MemDeblur [35], and Spk2ImgNet [85]. We also add an off-the-shelf denoiser BM3D [14] to QBP, denoted QBP (+BM3D), as a baseline for comparison. As we will discuss in Sec. 5.3, QUIVER beats all the baselines, both quantitatively and qualitatively." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "P. Chennuri et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.151, + 0.787, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.295, + 0.789, + 0.351 + ], + "angle": 0, + "content": "Fig. 10: Performance on Real Quanta Data. We capture real 1-bit quanta data using a SPAD [17] and generate 3-bit frames through temporal averaging. All deep learning based models are trained using a photon level of 4.9 PPP per frame. Best viewed in zoom." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.789, + 0.534 + ], + "angle": 0, + "content": "Training QUIVER. We utilize the function mentioned in Eq. (1) as the cost function for training QUIVER with regularization parameters \\(\\lambda_{1} = 0.2\\), \\(\\lambda_{2} = 0.85\\), \\(\\lambda_{3} = 0.1\\), and \\(\\lambda_{4} = 0.05\\). The training data is extracted with patch size \\(128 \\times 128\\) and a batch size of 4. The weights are initialized with Lecun initialization [39]. The network is trained using the Adam optimizer [37] with an initial learning rate of \\(2.5 \\times 10^{-5}\\). The low learning rate is driven by the inherent instability of recurrent networks, as it mitigates the risk of divergent behavior during training. We use a learning rate scheduler that reduces the learning rate by a factor of 2 when a plateau is reached. QUIVER takes approximately 1.5 days to train on a NVIDIA A100 Tensor Core GPU using Pytorch." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.557, + 0.325, + 0.57 + ], + "angle": 0, + "content": "5.3 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.789, + 0.78 + ], + "angle": 0, + "content": "Synthetic Data Experiments. We begin with the synthetic experiments where we utilize 3-bit quanta frames, simulated using the parameters mentioned in Sec. 5.1 at 3.25, 9.75, 19.5, and 26 PPP to test the algorithms' performance. Tab. 2 and Tab. 3 demonstrate the PSNR and SSIM [76] of various methods extracted by predicting 6017 I2-2000FPS frames and 345 X4K1000FPS frames. To further substantiate the efficacy of QUIVER's design, we introduced a scaled-down variant, QUIVER-s (Refer Fig. 8(b) for complexity comparison). Quantitative results indicate that both QUIVER and QUIVER-s offer substantially better performance than all the baselines across a range of light levels. Fig. 9 depicts visual results of all the methods on the I2-2000FPS dataset. It is evident that existing methods fail to handle both motion and noise simultaneously, whereas, our proposed method, QUIVER, produces blur free high SNR outputs while preserving high-frequency details to a large extent." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Real Data Experiments. We verify the methods' performance on real data. The real data is collected as binary frames using a SPAD sensor [17] at 10000 FPS with a spatial resolution of \\(240 \\times 320\\). As SPADs possess zero read noise, the binary frames are summed up to generate 3-bit frames. The average observed" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "QUIVER" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.128 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 2: Performance comparison on the proposed I2-2000FPS dataset across various light levels. Models are trained using the I2-2000FPS dataset. QUIVER performs significantly better than the existing methods." + }, + { + "type": "table", + "bbox": [ + 0.306, + 0.199, + 0.699, + 0.32 + ], + "angle": 0, + "content": "
Photons-Per-Pixel (PPP)3.259.7519.526
MethodPSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑
Transform Denoise [6]21.31700.718423.15210.767122.77480.781222.30960.7811
QBP [47]15.94110.129319.18560.265420.40000.371320.79780.4114
QBP (+ BM3D [14])21.54760.703322.20010.689922.83510.769622.86170.7832
Student-Teacher [10]18.72000.400616.51950.247915.76360.213313.28890.0735
RVRT [42]19.41150.353921.67140.456822.08260.502121.75280.4968
EMVD [2]20.01940.587321.05590.604822.44030.559223.40530.5576
FloRNN [1]21.03410.678525.61320.709127.43220.739527.85200.7784
MemDeblur [35]19.48770.386814.49060.111216.17750.166716.00580.1712
Spk2ImgNet [85]20.39450.564219.66650.673322.93720.700814.97690.6861
QUIVER-s (Ours)24.70130.756526.86760.788327.29890.843227.86590.8408
QUIVER (Ours)26.21430.789726.80580.825027.75380.856327.93770.8446
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.333, + 0.788, + 0.375 + ], + "angle": 0, + "content": "Table 3: Performance comparison on the X4K1000FPS dataset across various light levels. Models are trained using the I2-2000FPS dataset. QUIVER performs significantly better than the existing methods." + }, + { + "type": "table", + "bbox": [ + 0.306, + 0.388, + 0.699, + 0.508 + ], + "angle": 0, + "content": "
Photons-Per-Pixel (PPP)3.259.7519.526
MethodPSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑
Transform Denoise [6]19.62550.632322.17030.704422.99380.722922.62300.7204
QBP [47]15.56340.230216.97580.323017.17980.395717.78070.4188
QBP (+ BM3D [14])17.96770.512318.53080.522618.24070.541418.79170.5586
Student-Teacher [10]18.82080.365216.15480.260814.93590.257113.97620.1186
RVRT [42]19.92030.364121.07810.447221.47800.492520.78990.4919
EMVD [2]20.51020.483621.81520.559522.94400.593622.45870.5860
FloRNN [1]20.82830.577823.58740.648424.32140.668325.24830.7170
MemDeblur [35]19.55340.364214.55950.220316.67490.311615.64960.2974
Spk2ImgNet [85]18.94240.473119.25320.572220.34420.571616.09310.6106
QUIVER-s (Ours)20.91970.595521.79900.652324.19240.731623.44110.7248
QUIVER (Ours)21.87300.652123.16540.705724.59560.764525.00860.7513
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.788, + 0.658 + ], + "angle": 0, + "content": "light level after summation is 4.9 PPP. We generate results with networks trained at 4.9 PPP and demonstrate the visual results in Fig. 10. QUIVER, as opposed to existing state-of-the-art, effectively recovers high-frequency information while applying visually appealing smoothening effect to low-frequency regions of the scene. It is noteworthy that SPADs' image formation model is significantly different from that of the QIS' imaging model [58,59]. Therefore, the visual results also indicate that the proposed QUIVER can thoroughly generalize to various single-photon detectors." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.685, + 0.403, + 0.703 + ], + "angle": 0, + "content": "6 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Effect of Pre-Denoising, Optical Flow and Multi-Scale. We conduct an ablation study to evaluate the effect of the pre-denoising, the learnable optical flow, and multi-scale reconstruction on performance. Upon removal of either module, we expand the features dimension of layers in the feature extraction stage and add RDB blocks to the RMDF module, thereby maintaining a similar model capacity. We train all possible combinations and display the quantitative results in Tab. 4. Results indicate that, in the absence of either one or more modules, the network's performance is substantially worse. Visual Results in Fig. 11" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "P. Chennuri et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.146, + 0.787, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.242, + 0.788, + 0.271 + ], + "angle": 0, + "content": "Fig.11: Ablation Study. Visual Comparisons depicting the effectiveness of Predenoiser, Optical Flow, and Multi-Scale Reconstruction Modules. Best viewed in zoom." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.784, + 0.327 + ], + "angle": 0, + "content": "indicate that these modules serve critical roles as they significantly contribute to model's performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.328, + 0.784, + 0.404 + ], + "angle": 0, + "content": "Does loading pre-trained optical flow module weights help? We initialize SPyNet with its pre-trained weights and finetune the same while training QUIVER, and display its quantitative results in Tab. 4. As the pre-trained SPyNet is not robust to photon shot noise and read noise, especially in low-light conditions, initializing the module with it will result in sub-optimal performance." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.441, + 0.788, + 0.484 + ], + "angle": 0, + "content": "Table 4: Ablation study. We conduct experiments to emphasize the role of Denoiser, Optical flow, and Multi-Scale reconstruction modules. We also show the effect of loading pretrained optical flow weights \\((^{*})\\) on performance." + }, + { + "type": "table", + "bbox": [ + 0.374, + 0.495, + 0.631, + 0.597 + ], + "angle": 0, + "content": "
Pre-Denoising Optical Flow Multi-ScaleI2-2000FPS
PSNR↑SSIM↑
×××23.67020.7756
✓✓*✓23.94790.7709
✓✓×24.38410.7808
××✓24.74450.7755
×✓✓24.99990.7753
✓×✓25.75210.7760
✓✓✓26.21430.7897
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.646, + 0.36, + 0.661 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In this paper, we presented a methodology to reconstruct blur-free grayscale images/videos captured using 1-bit or few-bit quanta data. While adopting the ideology of classical quanta restoration methods, we proposed an end-to-end deep learning framework, QUIVER, that utilizes pre-filtering, learnable optical flow module, and a novel multi-scale reconstruction approach to produce high-visual outputs. Experiments on synthetic and real data indicate QUIVER beats state-of-the-art and can generalize across single-photon sensors. We also introduce the world's first high-speed video dataset, I2-2000FPS, that captures fast moving scenes at 2000 fps, covering wide ranges of motion. We believe that I2-2000FPS will be a valuable asset for researchers in high-speed motion analysis and other computer vision tasks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "QUIVER" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "Acknowledgements. The work is supported, in part, by Samsung Research America (SRA), DARPA / SRC CogniSense JUMP 2.0 Center, NSF IIS-2133032, and NSF ECCS-2030570. The authors would like to thank German Mora Martin for his help in recording the SPAD data." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.23, + 0.323, + 0.246 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.26, + 0.786, + 0.316 + ], + "angle": 0, + "content": "1. Li et al, J.: Unidirectional Video Denoising by Mimicking Backward Recurrent Modules with Look-Ahead Forward Ones. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) ECCV. Lecture Notes in Computer Science (2022). https://doi.org/10.1007/978-3-031-19797-0_34" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.317, + 0.786, + 0.343 + ], + "angle": 0, + "content": "2. Maggioni et al, M.: Efficient Multi-Stage Video Denoising With Recurrent Spatio-Temporal Fusion. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.344, + 0.786, + 0.385 + ], + "angle": 0, + "content": "3. Arias, P., Morel, J.M.: Video Denoising via Empirical Bayesian Estimation of Space-Time Patches. Journal of Mathematical Imaging and Vision 60(1), 70-93 (Jan 2018). https://doi.org/10.1007/s10851-017-0742-4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.385, + 0.786, + 0.44 + ], + "angle": 0, + "content": "4. Chan, K.C.K., Zhou, S., Xu, X., Loy, C.C.: BasicVSR++: Improving Video SuperResolution With Enhanced Propagation and Alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5972-5981 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.441, + 0.786, + 0.48 + ], + "angle": 0, + "content": "5. Chan, S.H.: What Does a One-Bit Quanta Image Sensor Offer? IEEE Transactions on Computational Imaging 8, 770-783 (2022). https://doi.org/10.1109/TCI.2022.3202012" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.482, + 0.786, + 0.523 + ], + "angle": 0, + "content": "6. Chan, S.H., Elgendy, O.A., Wang, X.: Images from Bits: Non-Iterative Image Reconstruction for Quanta Image Sensors. Sensors 16(11), 1961 (Nov 2016). https://doi.org/10.3390/s16111961" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.524, + 0.786, + 0.577 + ], + "angle": 0, + "content": "7. Chan, S.H., Lu, Y.M.: Efficient image reconstruction for gigapixel quantum image sensors. In: 2014 IEEE Global Conference on Signal and Information Processing (GlobalSIP). pp. 312-316 (Dec 2014). https://doi.org/10.1109/GlobalSIP.2014.7032129" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.578, + 0.786, + 0.631 + ], + "angle": 0, + "content": "8. Chan, S.H., Wang, X., Elgendy, O.A.: Plug-and-Play ADMM for Image Restoration: Fixed-Point Convergence and Applications. IEEE Transactions on Computational Imaging 3(1), 84-98 (Mar 2017). https://doi.org/10.1109/TCI.2016.2629286" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.786, + 0.687 + ], + "angle": 0, + "content": "9. Charbon, E., Fishburn, M., Walker, R., Henderson, R.K., Niclass, C.: SPAD-Based Sensors. In: Remondino, F., Stoppa, D. (eds.) TOF Range-Imaging Cameras, pp. 11-38. Springer Berlin Heidelberg, Berlin, Heidelberg (2013). https://doi.org/10.1007/978-3-642-27523-4_2" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.689, + 0.786, + 0.757 + ], + "angle": 0, + "content": "10. Chi, Y., Gnanasambandam, A., Koltun, V., Chan, S.H.: Dynamic Low-Light Imaging with Quanta Image Sensors. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision - ECCV 2020, vol. 12366, pp. 122-138. Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-58589-1_8" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.758, + 0.786, + 0.812 + ], + "angle": 0, + "content": "11. Chi, Y., Zhang, X., Chan, S.H.: HDR imaging with spatially varying signal-to-noise ratios. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5724-5734. IEEE (2023). https://doi.org/10.1109/CVPR52729.2023.00554, https://doi.org/10.1109/CVPR52729.2023.00554" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.813, + 0.786, + 0.84 + ], + "angle": 0, + "content": "12. Choi, J.H., Elgendy, O.A., Chan, S.H.: Image Reconstruction for Quanta Image Sensors Using Deep Neural Networks. In: 2018 IEEE International Conference on" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.26, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.392, + 0.128 + ], + "angle": 0, + "content": "P. Chennuri et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "Acoustics, Speech and Signal Processing (ICASSP). pp. 6543-6547 (Apr 2018). https://doi.org/10.1109/ICASSP.2018.8461685" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.786, + 0.217 + ], + "angle": 0, + "content": "13. Claus, M., van Gemert, J.: ViDeNN: Deep Blind Video Denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.786, + 0.259 + ], + "angle": 0, + "content": "14. Dabov, K., Foi, A., Katkovnik, V., Egiazarian, K.: Image denoising by sparse 3-D transform-domain collaborative filtering. IEEE Transactions on image processing 16(8), 2080-2095 (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.26, + 0.785, + 0.285 + ], + "angle": 0, + "content": "15. Dong, S., Huang, T., Tian, Y.: Spike camera and its coding methods. arXiv preprint arXiv:2104.04669 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.287, + 0.786, + 0.342 + ], + "angle": 0, + "content": "16. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.343, + 0.786, + 0.397 + ], + "angle": 0, + "content": "17. Dutton, N.A.W., Gyongy, I., Parmesan, L., Gnecchi, S., Calder, N., Rae, B.R., Pellegrini, S., Grant, L.A., Henderson, R.K.: A SPAD-Based QVGA Image Sensor for Single-Photon Counting and Quanta Imaging. IEEE Transactions on Electron Devices 63(1), 189–196 (Jan 2016). https://doi.org/10.1109/TED.2015.2464682" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.398, + 0.786, + 0.438 + ], + "angle": 0, + "content": "18. Dutton, N.A., Gyongy, I., Parmesan, L., Henderson, R.K.: Single photon counting performance and noise analysis of CMOS SPAD-based image sensors. Sensors 16(7), 1122 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.439, + 0.786, + 0.48 + ], + "angle": 0, + "content": "19. Elgendy, O.A., Chan, S.H.: Optimal Threshold Design for Quanta Image Sensor. IEEE Transactions on Computational Imaging 4(1), 99-111 (Mar 2018). https://doi.org/10.1109/TCI.2017.2781185" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.481, + 0.786, + 0.534 + ], + "angle": 0, + "content": "20. Feng Yang, Lu, Y.M., Sbaiz, L., Vetterli, M.: Bits From Photons: Oversampled Image Acquisition Using Binary Poisson Statistics. IEEE Transactions on Image Processing 21(4), 1421-1436 (Apr 2012). https://doi.org/10.1109/TIP.2011.2179306" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.536, + 0.786, + 0.577 + ], + "angle": 0, + "content": "21. Fossum, E., Ma, J., Masoodian, S., Anzagira, L., Zizza, R.: The Quanta Image Sensor: Every Photon Counts. Sensors 16(8), 1260 (Aug 2016). https://doi.org/10.3390/s16081260" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.578, + 0.786, + 0.605 + ], + "angle": 0, + "content": "22. Fossum, E.R.: Modeling the performance of single-bit and multi-bit quanta image sensors. IEEE Journal of the Electron Devices Society 1(9), 166-174 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.606, + 0.786, + 0.646 + ], + "angle": 0, + "content": "23. Gao, J., Shang, Z., Nie, K., Luo, T.: High dynamic range image reconstruction for multi-bit quanta image sensor. Optoelectronics Letters 18(9), 553-558 (Sep 2022). https://doi.org/10.1007/s11801-022-2014-9" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.647, + 0.786, + 0.7 + ], + "angle": 0, + "content": "24. Gariepy, G., Krstajic, N., Henderson, R., Li, C., Thomson, R.R., Buller, G.S., Heshmat, B., Raskar, R., Leach, J., Faccio, D.: Single-photon sensitive light-infight imaging. Nature Communications 6(1), 6021 (Jan 2015). https://doi.org/10.1038/ncomms7021" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.702, + 0.786, + 0.757 + ], + "angle": 0, + "content": "25. Gariepy, G., Leach, J., Warburton, R., Chan, S., Henderson, R., Faccio, D.: Picosecond time-resolved imaging using SPAD cameras. In: Lewis, K.L., Hollins, R.C. (eds.) SPIE Security + Defence. p. 99920N. Edinburgh, United Kingdom (Oct 2016). https://doi.org/10.1117/12.2241184" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.758, + 0.786, + 0.811 + ], + "angle": 0, + "content": "26. Gnanasambandam, A., Chan, S.H.: HDR Imaging With Quanta Image Sensors: Theoretical Limits and Optimal Reconstruction. IEEE Transactions on Computational Imaging 6, 1571-1585 (2020). https://doi.org/10.1109/TCI.2020.3041093" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.812, + 0.786, + 0.84 + ], + "angle": 0, + "content": "27. Gnanasambandam, A., Chan, S.H.: Image Classification in the Dark Using Quanta Image Sensors. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Proceed-" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "QUIVER" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.147, + 0.786, + 0.189 + ], + "angle": 0, + "content": "ings of the European Conference on Computer Vision (ECCV). pp. 484-501. Lecture Notes in Computer Science, Cham (2020). https://doi.org/10.1007/978-3-030-58598-3_29" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "28. Gnanasambandam, A., Chan, S.H.: Exposure-Referred Signal-to-Noise Ratio for Digital Image Sensors. IEEE Transactions on Computational Imaging 8, 561-575 (2022). https://doi.org/10.1109/TCI.2022.3187657" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.786, + 0.259 + ], + "angle": 0, + "content": "29. Godard, C., Matzen, K., Uytendaele, M.: Deep Burst Denoising. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 538-554 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.786, + 0.3 + ], + "angle": 0, + "content": "30. Gupta, A., Ingle, A., Velten, A., Gupta, M.: Photon-Flooded Single-Photon 3D Cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6770-6779 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.786, + 0.328 + ], + "angle": 0, + "content": "31. Gupta, S., Gupta, M.: Eulerian Single-Photon Vision. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10465–10476 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.786, + 0.37 + ], + "angle": 0, + "content": "32. Gutierrez-Barragan, F., Ingle, A., Seets, T., Gupta, M., Velten, A.: Compressive Single-Photon 3D Cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17854-17864 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.786, + 0.398 + ], + "angle": 0, + "content": "33. Gyongy, I., Dutton, N., Henderson, R.: Single-Photon Tracking for High-Speed Vision. Sensors 18(2), 323 (Jan 2018). https://doi.org/10.3390/s18020323" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.786, + 0.452 + ], + "angle": 0, + "content": "34. Hasinoff, S.W., Sharlet, D., Geiss, R., Adams, A., Barron, J.T., Kainz, F., Chen, J., Levoy, M.: Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM Transactions on Graphics 35(6), 1-12 (Nov 2016). https://doi.org/10.1145/2980179.2980254" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.786, + 0.467 + ], + "angle": 0, + "content": "35. Ji, B., Yao, A.: Multi-Scale Memory-Based Video Deblurring. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.786, + 0.508 + ], + "angle": 0, + "content": "36. Kiani Galoogahi, H., Fagg, A., Huang, C., Ramanan, D., Lucey, S.: Need for Speed: A Benchmark for Higher Frame Rate Object Tracking. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1125-1134 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.786, + 0.549 + ], + "angle": 0, + "content": "37. Kingma, D., Ba, J.: Adam: A Method for Stochastic Optimization. In: International Conference on Learning Representations (ICLR) (2015). https://doi.org/10.48550/arXiv.1412.6980" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.549, + 0.786, + 0.591 + ], + "angle": 0, + "content": "38. Lebrun, M., Buades, A., Morel, J.M.: A Nonlocal Bayesian Image Denoising Algorithm. SIAM Journal on Imaging Sciences 6(3), 1665-1688 (Jan 2013). https://doi.org/10.1137/120874989" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.786, + 0.633 + ], + "angle": 0, + "content": "39. Lecun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (Nov/1998). https://doi.org/10.1109/5.726791" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.786, + 0.688 + ], + "angle": 0, + "content": "40. Li, C., Qu, X., Gnanasambandam, A., Elgendy, O.A., Ma, J., Chan, S.H.: Photon-limited object detection using non-local feature matching and knowledge distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3976-3987 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.688, + 0.786, + 0.729 + ], + "angle": 0, + "content": "41. Liang, J., Cao, J., Fan, Y., Zhang, K., Ranjan, R., Li, Y., Timofte, R., Van Gool, L.: VRT: A Video Restoration Transformer (Jun 2022). https://doi.org/10.48550/arXiv.2201.12288" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.786, + 0.785 + ], + "angle": 0, + "content": "42. Liang, J., Fan, Y., Xiang, X., Ranjan, R., Ilg, E., Green, S., Cao, J., Zhang, K., Timofte, R., Gool, L.V.: Recurrent Video Restoration Transformer with Guided Deformable Attention. Advances in Neural Information Processing Systems 35, 378-393 (Dec 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.786, + 0.84 + ], + "angle": 0, + "content": "43. Liba, O., Murthy, K., Tsai, Y.T., Brooks, T., Xue, T., Karnad, N., He, Q., Barron, J.T., Sharlet, D., Geiss, R., Hasinoff, S.W., Pritch, Y., Levoy, M.: Handheld mobile photography in very low light. ACM Transactions on Graphics 38(6), 1-16 (Dec 2019). https://doi.org/10.1145/3355089.3356508" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.392, + 0.128 + ], + "angle": 0, + "content": "P. Chennuri et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "44. Lindell, D.B., O'Toole, M., Wetzstein, G.: Single-photon 3D imaging with deep sensor fusion. ACM Transactions on Graphics 37(4), 113-1 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.175, + 0.785, + 0.215 + ], + "angle": 0, + "content": "45. Ma, J., Chan, S., Fossum, E.R.: Review of Quanta Image Sensors for UltralowLight Imaging. IEEE Transactions on Electron Devices 69(6), 2824-2839 (Jun 2022). https://doi.org/10.1109/TED.2022.3166716" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.217, + 0.785, + 0.256 + ], + "angle": 0, + "content": "46. Ma, J., Masoodian, S., Starkey, D.A., Fossum, E.R.: Photon-number-resolving megapixel image sensor at room temperature without avalanche gain. Optica 4(12), 1474-1481 (Dec 2017). https://doi.org/10.1364/OPTICA.4.001474" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.297 + ], + "angle": 0, + "content": "47. Ma, S., Gupta, S., Ulku, A.C., Bruschini, C., Charbon, E., Gupta, M.: Quanta burst photography. ACM Transactions on Graphics 39(4) (Aug 2020). https://doi.org/10.1145/3386569.3392470" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.297, + 0.785, + 0.337 + ], + "angle": 0, + "content": "48. Ma, S., Mos, P., Charbon, E., Gupta, M.: Burst Vision Using Single-Photon Cameras. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5375-5385 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.785, + 0.39 + ], + "angle": 0, + "content": "49. Madhusudana, P.C., Yu, X., Birkbeck, N., Wang, Y., Adsumilli, B., Bovik, A.C.: Subjective and Objective Quality Assessment of High Frame Rate Videos. IEEE Access 9, 108069-108082 (2021). https://doi.org/10.1109/ACCESS.2021.3100462" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.392, + 0.785, + 0.446 + ], + "angle": 0, + "content": "50. Maggioni, M., Boracchi, G., Foi, A., Egiazarian, K.: Video Denoising, Deblocking, and Enhancement Through Separable 4-D Nonlocal Spatiotemporal Transforms. IEEE Transactions on Image Processing 21(9), 3952-3966 (Sep 2012). https://doi.org/10.1109/TIP.2012.2199324" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.785, + 0.487 + ], + "angle": 0, + "content": "51. Monakhova, K., Richter, S.R., Waller, L., Koltun, V.: Dancing Under the Stars: Video Denoising in Starlight. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16241-16251 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.54 + ], + "angle": 0, + "content": "52. Niclass, C., Rochas, A., Besse, P.A., Charbon, E.: Design and characterization of a CMOS 3-D image sensor based on single photon avalanche diodes. IEEE Journal of Solid-State Circuits 40(9), 1847-1854 (Sep 2005). https://doi.org/10.1109/JSSC.2005.848173" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.541, + 0.785, + 0.596 + ], + "angle": 0, + "content": "53. Pan, L., Scheerlinck, C., Yu, X., Hartley, R., Liu, M., Dai, Y.: Bringing a Blur Frame Alive at High Frame-Rate With an Event Camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6820-6829 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.597, + 0.785, + 0.637 + ], + "angle": 0, + "content": "54. Pearl, N., Treibitz, T., Korman, S.: NAN: Noise-Aware NeRFs for Burst-Denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12672-12681 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.677 + ], + "angle": 0, + "content": "55. Plotz, T., Roth, S.: Benchmarking Denoising Algorithms With Real Photographs. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1586-1595 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.785, + 0.717 + ], + "angle": 0, + "content": "56. Qu, X., Chi, Y., Chan, S.H.: Spatially varying exposure with 2-by-2 multiplexing: Optimality and universality. IEEE Transactions on Computational Imaging 10, 261-276 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "57. Ranjan, A., Black, M.J.: Optical Flow Estimation Using a Spatial Pyramid Network. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4161-4170 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "58. Rapp, J., Goyal, V.K.: A Few Photons Among Many: Unmixing Signal and Noise for Photon-Efficient Active Imaging. IEEE Transactions on Computational Imaging 3(3), 445-459 (Sep 2017). https://doi.org/10.1109/TCI.2017.2706028" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "59. Rapp, J., Ma, Y., Dawson, R.M.A., Goyal, V.K.: Dead Time Compensation for High-Flux Ranging. IEEE Transactions on Signal Processing 67(13), 3471-3486 (Jul 2019). https://doi.org/10.1109/TSP.2019.2914891" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "QUIVER" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "60. Rebecq, H., Ranftl, R., Koltun, V., Scaramuzza, D.: High Speed and High Dynamic Range Video with an Event Camera. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(6), 1964-1980 (Jun 2021). https://doi.org/10.1109/TPAMI.2019.2963386" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.26 + ], + "angle": 0, + "content": "61. Remez, T., Litany, O., Bronstein, A.: A picture is worth a billion bits: Real-time image reconstruction from dense binary threshold pixels. In: 2016 IEEE International Conference on Computational Photography (ICCP). pp. 1-9 (May 2016). https://doi.org/10.1109/ICCPHOT.2016.7492874" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.261, + 0.787, + 0.288 + ], + "angle": 0, + "content": "62. Rochas, A.: Single photon avalanche diodes in CMOS technology. Tech. rep., Cite-seer (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.289, + 0.787, + 0.345 + ], + "angle": 0, + "content": "63. Ruget, A., Tyler, M., Mora Martin, G., Scholes, S., Zhu, F., Gyongy, I., Hearn, B., McLaughlin, S., Halimi, A., Leach, J.: Pixels2Pose: Super-resolution time-of-flight imaging for 3D pose estimation. Science Advances 8(48), eade0123 (Nov 2022). https://doi.org/10.1126/sciadv.ade0123" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.787, + 0.387 + ], + "angle": 0, + "content": "64. Seets, T., Ingle, A., Laurenzis, M., Velten, A.: Motion Adaptive Deblurring With Single-Photon Cameras. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1945–1954 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.388, + 0.787, + 0.416 + ], + "angle": 0, + "content": "65. Shazeer, N.: Glu variants improve transformer. arXiv preprint arXiv:2002.05202 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.417, + 0.787, + 0.471 + ], + "angle": 0, + "content": "66. Shin, D., Xu, F., Venkatraman, D., Lussana, R., Villa, F., Zappa, F., Goyal, V.K., Wong, F.N.C., Shapiro, J.H.: Photon-efficient imaging with a single-photon camera. Nature Communications 7(1), 12046 (Jun 2016). https://doi.org/10.1038/ncomms12046" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.515 + ], + "angle": 0, + "content": "67. Sim, H., Oh, J., Kim, M.: XVFI: eXtreme Video Frame Interpolation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14489-14498 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.787, + 0.558 + ], + "angle": 0, + "content": "68. Su, S., Delbracio, M., Wang, J., Sapiro, G., Heidrich, W., Wang, O.: Deep Video Deblurring for Hand-Held Cameras. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1279-1288 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.787, + 0.613 + ], + "angle": 0, + "content": "69. Tassano, M., Delon, J., Veit, T.: DVDNET: A Fast Network for Deep Video Denoising. In: 2019 IEEE International Conference on Image Processing (ICIP). pp. 1805-1809. IEEE, Taipei, Taiwan (Sep 2019). https://doi.org/10.1109/ICIP.2019.8803136" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.614, + 0.787, + 0.657 + ], + "angle": 0, + "content": "70. Tassano, M., Delon, J., Veit, T.: FastDVDnet: Towards Real-Time Deep Video Denoising Without Flow Estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1354-1363 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.787, + 0.699 + ], + "angle": 0, + "content": "71. Vaksman, G., Elad, M., Milanfar, P.: Patch Craft: Video Denoising by Deep Modeling and Patch Matching. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2157-2166 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.787, + 0.742 + ], + "angle": 0, + "content": "72. Voeikov, R., Falaleev, N., Baikulov, R.: TTNet: Real-Time Temporal and Spatial Video Analysis of Table Tennis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 884-885 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.787, + 0.798 + ], + "angle": 0, + "content": "73. Vogels, T., Rousselle, F., Mcwilliams, B., Rothlin, G., Harvill, A., Adler, D., Meyer, M., Novak, J.: Denoising with kernel prediction and asymmetric loss functions. ACM Transactions on Graphics 37(4), 124:1-124:15 (Jul 2018). https://doi.org/10.1145/3197517.3201388" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "74. Wang, W., Chen, X., Yang, C., Li, X., Hu, X., Yue, T.: Enhancing Low Light Videos by Exploring High Sensitivity Camera Noise. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4111-4119 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.392, + 0.128 + ], + "angle": 0, + "content": "P. Chennuri et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.203 + ], + "angle": 0, + "content": "75. Wang, X.: Single-Photon Cameras Image Reconstruction Using Vision Transformer. In: 2023 IEEE 3rd International Conference on Computer Communication and Artificial Intelligence (CCAI). pp. 296-300 (May 2023). https://doi.org/10.1109/CCAI57533.2023.10201259" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "76. Wang, Z., Bovik, A., Sheikh, H., Simoncelli, E.: Image Quality Assessment: From Error Visibility to Structural Similarity. IEEE Transactions on Image Processing 13(4), 600-612 (Apr 2004). https://doi.org/10.1109/TIP.2003.819861" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.786, + 0.286 + ], + "angle": 0, + "content": "77. Wong, H.T., Leung, C.S., Ho, D.: Theoretical analysis and image reconstruction for multi-bit quanta image sensors. Signal Processing 185, 108087 (Aug 2021). https://doi.org/10.1016/j.sigpro.2021.108087" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.342 + ], + "angle": 0, + "content": "78. Yang, F., Lu, Y.M., Sbaiz, L., Vetterli, M.: An optimal algorithm for reconstructing images from binary measurements. In: Bouman, C.A., Pollak, I., Wolfe, P.J. (eds.) IS&T/SPIE Electronic Imaging. p. 75330K. San Jose, California (Feb 2010). https://doi.org/10.1117/12.850887" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.787, + 0.397 + ], + "angle": 0, + "content": "79. Yang, F., Sbaiz, L., Charbon, E., Susstrunk, S., Vetterli, M.: Image reconstruction in the gigavision camera. In: 2009 IEEE 12th International Conference on Computer Vision Workshops, ICCV Workshops. pp. 2212-2219 (Sep 2009). https://doi.org/10.1109/ICCVW.2009.5457554" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.452 + ], + "angle": 0, + "content": "80. Zhang, D., Lian, Q., Su, Y., Ren, T.: Dual-Prior Integrated Image Reconstruction for Quanta Image Sensors Using Multi-Agent Consensus Equilibrium. IEEE/CAA Journal of Automatica Sinica 10(6), 1407-1420 (Jun 2023). https://doi.org/10.1109/JAS.2023.123390" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.493 + ], + "angle": 0, + "content": "81. Zhang, D., Lian, Q., Yang, Y.: TwP: Two-stage projection framework with manifold constraint for image reconstruction. Digital Signal Processing 141, 104186 (Sep 2023). https://doi.org/10.1016/j.dsp.2023.104186" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.787, + 0.536 + ], + "angle": 0, + "content": "82. Zhang, Y., Li, K., Li, K., Wang, L., Zhong, B., Fu, Y.: Image Super-Resolution Using Very Deep Residual Channel Attention Networks. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 286-301 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.787, + 0.577 + ], + "angle": 0, + "content": "83. Zhang, Y., Tian, Y., Kong, Y., Zhong, B., Fu, Y.: Residual Dense Network for Image Restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(7), 2480-2495 (Jul 2021). https://doi.org/10.1109/TPAMI.2020.2968521" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.632 + ], + "angle": 0, + "content": "84. Zhao, J., Xiong, R., Huang, T.: High-Speed Motion Scene Reconstruction for Spike Camera via Motion Aligned Filtering. In: 2020 IEEE International Symposium on Circuits and Systems (ISCAS). pp. 1-5. IEEE, Seville, Spain (Oct 2020). https://doi.org/10.1109/ISCAS45731.2020.9181055" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.787, + 0.687 + ], + "angle": 0, + "content": "85. Zhao, J., Xiong, R., Liu, H., Zhang, J., Huang, T.: Spk2imgnet: Learning to reconstruct dynamic scene from continuous spike stream. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11996-12005 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.688, + 0.787, + 0.743 + ], + "angle": 0, + "content": "86. Zhao, J., Xiong, R., Xie, J., Shi, B., Yu, Z., Gao, W., Huang, T.: Reconstructing Clear Image for High-Speed Motion Scene With a Retina-Inspired Spike Camera. IEEE Transactions on Computational Imaging 8, 12-27 (2022). https://doi.org/10.1109/TCI.2021.3136446" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.743 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Quanta Video Restoration/c31fed1f-5357-4c1e-adf0-39121e1936f9_origin.pdf b/2024/Quanta Video Restoration/c31fed1f-5357-4c1e-adf0-39121e1936f9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..51b7d1ec8472aaf4218dfa2fd2566b05ea41bc2a --- /dev/null +++ b/2024/Quanta Video Restoration/c31fed1f-5357-4c1e-adf0-39121e1936f9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eeacd31e30d988a4d09305bd7dd6ba0a9db2196562b690cac701bd7a87b78805 +size 9758694 diff --git a/2024/Quanta Video Restoration/full.md b/2024/Quanta Video Restoration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0f8c206465b3d081ae7f1fb22e2422102a70ca79 --- /dev/null +++ b/2024/Quanta Video Restoration/full.md @@ -0,0 +1,327 @@ +# Quanta Video Restoration + +Prateek Chennuri1, Yiheng Chi1, Enze Jiang1, G.M. Dilshan Godaliyadda2, Abhiram Gnanasambandam2, Hamid R. Sheikh2, Istvan Gyongy3, and Stanley H. Chan1 + +1 Purdue University + +Samsung Research America + +3 University of Edinburgh + +{pchennur,chi14,jiang708,stanchan}@purdue.edu + +{dilshan.g,abhiram.g,hr.sheikh}@samsung.com + +{igyongy2}@exseed.ed.ac.uk + +![](images/717431360766581130aafba6d78ed325ed59dbd715ae724c05202ebba57c80da.jpg) +(a) Ground Truth 16-bit + +![](images/72f8a4bfd81ddfc6cf365b6676284a7f2c4013fba754cdeef86c3d959eba390d.jpg) +(b) CMOS 16-bit, 60 fps at 1 lux + +![](images/800119ba33ffe1f34b74b36b8aab7890dcc9847ac7f3984a09ac19f94891562c.jpg) +(c) CMOS 16-bit, 240 fps at 1 lux + +![](images/4f6c112ec80eed3beec144f79075088a7834e772eb0833a34e906e039b2cb13f.jpg) +(d) CMOS 16-bit, 2000 fps at 1 lux + +![](images/b06a96207be44f26a7ad4b1270b5bce50ca8ce2cbc15a7bf67686639c9f03b99.jpg) +(e) Quanta 3-bit, 2000 fps at 1 lux +Fig. 1: Goal of this paper. (a) Blur-free video frame of a moving car. (b)-(d) CMOS image sensor simulations using realistic sensor parameters. The strong shot noise and read noise $(5.1\mathrm{e}^{-}/\mathrm{pix})$ of CMOS sensor make the signal acquisition difficult. (e) With low read noise $(0.2\mathrm{e}^{-}/\mathrm{pix})$ , low-bit single-photon detectors capture valuable information. (f)-(g) Existing state-of-the-art algorithm, QBP [47] cannot handle strong motion and noise. (h) The proposed algorithm, QUIVER, produces high quality results. + +![](images/e28400dbfc1bfff024ff7fe8fe45c537123c8bcc2dc0d2b2fae4933cb5cf2bca.jpg) +f) Quanta 3-bit, reconstructed from 11-frame avg. using Traditional Methods [47] + +![](images/3a5136568023e9a68a1a504178093a13cb9e6dda89203ff7fb294173bac88216.jpg) +(g) Quanta 3-bit, reconstructed from 66-frame avg, using Traditional Methods [47] +(h) Quanta 3-bit, reconstructed from 11-frame avg using QUIVER + +![](images/67a01e4617af268a2feb53a47ff71f8049f7205a092816eb76a5ffa442fe0698.jpg) + +Abstract. The proliferation of single-photon image sensors has opened the door to a plethora of high-speed and low-light imaging applications. However, data collected by these sensors are often 1-bit or few-bit, and corrupted by noise and strong motion. Conventional video restoration methods are not designed to handle this situation, while specialized quanta burst algorithms have limited performance when the number of input frames is low. In this paper, we introduce Quanta Video Restoration (QUIVER), an end-to-end trainable network built on the core ideas of classical quanta restoration methods, i.e., pre-filtering, flow estimation, fusion, and refinement. We also collect and publish I2-2000FPS, a high-speed video dataset with the highest temporal resolution of 2000 frames-per-second, for training and testing. On simulated and real data, QUIVER outperforms existing quanta restoration methods by a significant margin. Code and dataset available at https://github.com/ chennuripratek/Quanta_Video_Restoration-QUIVER + +Keywords: Single Photon Detectors $\cdot$ Video Restoration $\cdot$ High-Speed Dataset + +# 1 Introduction + +Over the past decade, the astonishing growth of single-photon detectors has fundamentally changed the landscape of computational imaging. With the invention and proliferation of quanta image sensors (QIS) [21] and single-photon avalanche diodes (SPAD) [52, 62], there is an unprecedented volume of new applications in low-light imaging [6, 10, 66], computer vision [27, 31, 40], high-speed videography [47, 48], time-of-flight sensing [32, 63], and 3D imaging [30, 44]. In most of these use cases, the main core question that lies is how to recover the image from the photon counts measured in the scene. Specifically, given a video stream of 1-bit or few-bit data captured from a scene involving moving objects, how do we reconstruct a gray-scale image/video while eliminating the noise without incurring motion blur? + +To give the reader a visual perspective of the problem scope, Fig. 1 depicts a blur-free video of a moving car. We simulate the captured images at 1 lux assuming 60 fps, 240 fps, and 2000 fps CMOS image sensors with realistic sensor specifications. As illustrated in the figure, the resulting CMOS outputs are either severely blurred due to strong motion or completely distorted by noise due to sparse photons. In the same figure, we demonstrate a simulated single-photon camera output (a 3-bit QIS in this case) where the content the largely preserved despite heavy noise. Upon utilizing state-of-the-art Quanta Burst Photography (QBP) [47] for reconstructing the frames, provided the motion is slow, a decent output can be obtained. However, as the temporal window narrows down, as shown in Fig. 1 (f), the noise remains. In this paper, we address this problem with a new algorithm, designed to remove the noise while avoiding distortions in the presence of fast motion while utilizing only a few frames. + +The core innovation of this paper is QUanta Video Restoration (QUIVER), a deep-learning based video restoration algorithm for quanta image data. QUIVER is specialized for few-bit data (3-bit) captured at thousands of frames-per-second (2000 fps) with an average motion range of 1 to 7 pixels per frame. The main contributions of this paper can be summarized as follows. + +- We propose QUIVER, an end-to-end trainable quanta (video) restoration method built by embracing the core ideas from traditional quanta restoration algorithms. On a comprehensive evaluation dataset containing both simulated and real data, QUIVER outperforms all methods we compared in this paper by a significant margin. +- We introduce I2-2000FPS, the first high-speed video dataset with a temporal resolution of 2000 frames-per-second for training and testing image and video reconstruction neural networks. We captured a total of 280 high-speed videos covering 114 distinct scenes with ground truth and simulated 3-bit videos. + +![](images/072ec0ae0cceaf8219c9aee1913c2fbb23e1b6ae6b49c88aee0b8831e59a9b5b.jpg) +Fig. 2: Motion Blur and SNR Trade-off. The effects of bit depth on SNR and motion blur are illustrated using real captures by a single-photon sensor. For the motion range we target, 3-bit single-photon detectors provide the best trade-off between blur and SNR. The images are captured using a 1-bit SPAD [17] at $10\mathrm{k}$ fps at an average photon level of 0.51 and 0.40 photons-per-pixel (PPP) per frame, respectively. Higher bit-depth outputs are generated through temporal frame averaging. + +Table 1: Frame-rate, motion, read-noise, and data-rate statistics for various bit-depths at the same exposure level. + +
Bit-DepthfpsMotion (pixels/frame)σread (/pixel/sec)Data-rate (Mb/sec)
110k0 - 12000 e-96
314282 - 3285.6 e-41.13
53236 - 1264.6 e-15.5
77825 - 3015.6 e-5.24
92070 - 804 e-1.73
+ +# 2 Background + +# 2.1 Few-bit Single-Photon Detectors + +What is few-bit photon counting? Single-photon detectors (QIS and SPAD) differ from conventional CMOS pixels by their extraordinary photon counting capability. QIS uses a two-stage pump-gate technique and correlated double sampling to suppress the read noise, while SPAD uses avalanche multiplication to amplify the photocharge. In both cases, the sensors are capable of resolving photons up to a single-photon sensitivity. We refer readers interested in the sensor development of QIS and SPAD to consult, for example, [5,9,18,45]. + +Along with the single-photon detectors' unique capability to count individual photons, these devices can generate data at a bit-depth as low as 1-bit to as high as 16-bit or even more. However, higher bit-depth is accompanied by longer integration time. If the scene contains motion, longer integration time will eventually result in strong motion blurs as shown in Fig. 2. On the other hand, 1-bit sensing with high frame rates will result in motion-blur-free but extremely noisy images. Therefore, from a pure data acquisition perspective, there exists an optimal bit-depth with respect to the motion that will give us mini- + +mal/no motion-blur data with a minimum per-frame signal-to-noise ratio (SNR) required for good quality reconstruction. + +How about 1-bit and reconstruct afterward? Readers familiar with single-photon counting may wonder whether we can collect as many 1-bit frames as possible and then process the data afterward. The problem is power consumption and data rate. Fixing the same level of exposure, as described in Tab. 1, a 1-bit video at $10\mathrm{k}$ fps would require $96\mathrm{Mb / sec}$ whereas a 9-bit video at 20 fps would only need $1.73\mathrm{Mb / sec}$ . Another problem is read noise accumulation. For sensors with non-zero read noise (such as QIS), every frame contributes to a finite amount of read noise. The more frames we read, the more read noise we accumulate. Therefore, recording 1-bit is not always the best option. + +# 2.2 Related Work + +Image and Video Denoising. Classical state-of-the-art methods utilize a non-local strategy to identify similar patches across an image/video [3, 38, 50]. Deep neural networks have been proven to be successful in producing high quality denoised outputs [13, 69-71]. Among these architectures, Vision Transformers [41, 42] have been rated the state-of-the-art in recent times. However, all these solutions make simplistic assumptions on noise statistics, thus failing to perform on real noisy images or videos [55]. + +Coming to low light, Burst denoising [34, 43], where images are aligned, merged and denoised, is one of the most popular methods. However, these methods fail without robust alignment. To overcome this, a number of alternative solutions with learnable alignment modules have been proposed [29, 54, 73]. Recent solutions have focused on practical noise models that replicate real camera sensor noise, to produce visually appealing results [51, 74]. Nevertheless, the existing solutions utilize images captured using CMOS image sensors, resulting in a notably higher photon level compared to the one utilized in our study. + +SPADs, Event, and Spike Cameras. Gariepy et al. [24, 25] firstly introduced the utilization of Single Photon Avalanche Diodes (SPADs) at pico-second temporal resolution to capture light in motion. Gyongy et al. [33] demonstrated 2D motion tracking of rigid planar objects using SPADs at $10\mathrm{k}$ fps. Recently, Ma et al. [47] and Seets et al. [64] utilized SPADs in a passive imaging setting to capture motion in low illumination. However, all these methods utilize extremely high-temporal resolutions hindering the deployment of these sensors into consumer devices where bandwidth is the bottleneck. Event [53, 60] and Spike Cameras [84, 86] also have demonstrated their effectiveness in capturing high-speed motion. However, these cameras focus on luminance/brightness variation and record a spike only when variation is above a threshold (changes based on factors like temperature, event rate, etc.) [15, 60]. Therefore, unlike single-photon detectors (QIS and SPADs), these cameras are NOT designed for single-photon counting and cannot operate in extreme low-light conditions. + +QIS Reconstruction. Reconstructing quanta images is a challenging task due to the underlying Poisson-Gaussian statistics. Initial solutions to this problem included utilizing standard gradient descent [78], greedy algorithms [79], + +![](images/b6a0645154d3686245d723db34243ef0054eda139319e1bec42f459e3a01387e.jpg) +Fig. 3: Traditional Methods' Design. Depiction of existing classical quanta restoration algorithms' design philosophy. Best viewed in zoom. + +(ADMM) [7,8], among others [19,20,23,26,61,77]. Chan et al. [6] were the first to propose a non-iterative approach using Anscombe transform for reconstructing quanta images. Choi et al. [12] proposed the first end-to-end trainable deep neural network (DNN) for QIS reconstruction. Alternative DNN-based solutions include utilizing vision transformers [75], Dual Prior Integrated networks [80], among others [81]. Nonetheless, all these methods fail to produce good results when the scene is in motion. Chi et al. [10] is the only method which focuses on capturing dynamic scenes using QIS but only targets at slow motion (1 pixel/frame). + +# 3 QUanta VIDEo REstoration (QUIVER) + +# 3.1 Design Philosophy + +In this section we present the design of our proposed algorithm. We start by briefly reviewing the design of classical methods [33, 47] which, to an extent, have been successful in restoring quanta images. As shown in Fig. 3, classical methods' algorithm design can be divided into four stages: (1) computing sum images to increase SNR, (2) optical flow (or) transformation matrix estimation for aligning the input frames, (3) warping and linear combination for generating preliminary restored output, and (4) refinement for producing the final output. While the steps seem intuitive and straightforward, existing methods are heavily vulnerable to extreme noise and strong motion in the input frames primarily due to two reasons. (1) none of the stages are designed to handle extreme noise and strong motion simultaneously (will be discussed further). (2) Since all the stages are sequential yet independent of each other, it is difficult to obtain an optimal result for a wide range of noise and motion. Our proposed algorithm QUIVER, leverages the design philosophy of existing classical methods while designing each stage to simultaneously handle both noise and motion. Moreover, QUIVER is an end-to-end trainable model making all the stages inter-dependent, thus leading to good restoration outputs. + +# 3.2 Design of QUIVER + +QUIVER is a deep-learning-based video restoration method for quanta imaging. The design philosophy of QUIVER is to adopt the intuitive thoughts behind + +![](images/5c7f37774aa0261c6b87616ddb40d180a26679556a92da9868216b68ec56eceb.jpg) +(a) + +![](images/d42c30ab7da6ddb25b54d18efe25c754167206411928a384ee6f4c4199841416.jpg) +(b) + +![](images/b756e46082b42a0320042036eede933e7b2c3d4978c22bdf302fa000d7b96400.jpg) +Fig. 4: Traditional Methods' Limitations. (a) Traditional methods' [33, 47] predenoising/temporal-averaging fails in strong motion. It is visible in the restored images that an input with strong motion between the frames results in several artifacts in the output even though SNR levels are similar. (b) Traditional methods [47] utilize a patch-based pre-trained optical flow module similar to [34]. This optical flow module fails to compensate for motion in the presence of noise. +Fig.5: The proposed QUIVER network. The corresponding stages of QUIVER, built by embracing the intuitive thoughts behind existing classical methods. Best viewed in zoom. + +classical quanta restoration methods and develop an end-to-end trainable, robust to noise and motion deep learning based framework, as shown in Fig. 5. Specifically, QUIVER can be divided into four main stages: + +Pre-Denoising to improve SNR: Since the input quanta frames possess extreme noise, classical methods adopt naive averaging to increase the SNR and thereby predict better optical flows or transformation matrices. However, as shown in Fig. 4(a), the simple averaging is vulnerable to motion and will negatively impact subsequent processing, ultimately leading to distorted outputs. Simply eliminating this stage is not the solution, because it leads to poor optical flow estimation, resulting in over-smoothed outputs with lack of low-level intricate details, as shown in Fig. 5 and Fig. 11. Therefore, a preliminary denoising step robust to noise and motion is crucial. In QUIVER, we leverage a computational undemanding single image denoiser built using Residual Dense Blocks [83] (RDBs) to provide minimal pre-preprocessing of the input quanta + +![](images/ebaaf12819d0307f87a753e9c3cd487c6c9fc5e0b48962895385d2af119a9b8c.jpg) +Fig. 6: Design of the proposed modules DC-GFU and RMDF. + +![](images/232f176b6f5c7320a3c9b211ba6e2d98d0d23ec67cb9a63335249cac1d1db811.jpg) + +data. A multi-frame denoiser is not an option due to it computationally demanding nature. We use RDB based network due to its history in handling noise while preserving details with its simple yet effective design. + +Optical Flow Estimation + Feature Alignment using Deformable Convolution - Gated Fusion Unit: Classical methods utilize an off-the-shelf pre-trained optical flow estimation module or predict a transformation matrix to compensate for motion between the frames. The basic assumption behind such approaches is that the motion between the frames is limited and the SNR is high enough. However, when such assumption is not met, the motion compensation is sub-optimal, as shown in Fig. 4(b). As most state-of-the-art pre-trained optical flow estimators are optimized on the CMOS RGB sensor images, it leads to sub-optimal performance when applied on quanta frames. Eliminating the flow estimation step is not recommended since experiments reveal the critical role it plays in motion compensation, as shown in Fig. 5 and Fig. 11. For QUIVER, we deploy a learnable optical flow estimation module and utilize SpyNet [57] owing to its computational efficiency while using a multi-scale approach. + +We deploy 3D convolution blocks to extract multi-scale spatio-temporal features from both the noisy and denoised quanta frames. We reuse the noisy frames to compensate for any information lost in the pre-denoising stage. + +The estimated multi-scale robust-to-noise optical flows are utilised for feature-level alignment of the extracted multi-scale spatio-temporal features. We utilize the deformable convolution with residual offsets proposed by [4] to warp the features. Inspired by the superior performance of Gated Linear Units (GLUs) in Transformers [65] we design and add a GLU based multi-layer-perceptron layer with GeLU activation for efficiently fusing the aligned features extracted from both the noisy and denoised frames. As shown in Fig. 6(a), we name this deformable convolution-GLU combination as the DC-GFU module. At this fusion-stage each frame is processed separately and the fusion is performed only along the channel dimension. Recurrence is applied for the alignment stage across all the multi-scale features of each frame. For the fusion module, we do not employ recurrence, as different scale features capture distinct long-range dependencies owing to their varying receptive fields. + +Deep Feature Fusion using Recurrent Multi-scale Residual Dense Feature Fusion Unit: Post warping we want to perform a robust-to-noise dense feature fusion while taking advantage of the temporal correlations among + +![](images/37909717e764ae1269c14e17a918a810d8e05332b3e7b5ead6b27d6c378f1d82.jpg) +Fig. 7: Design of the proposed modules TCA and RFRM. + +![](images/6cdfb18d9d26e7805896c65eba94777ed07ffadc0be2020ad6d22caaf23958d8.jpg) + +the features of all the input frames and also the spatial correlations between the multi-scale features within the same frame. For this task, we design and propose a Recurrent Multi-scale Residual Dense Feature Fusion Unit (RMDF) as shown in Fig. 6(b). The recurrence comes from the fact that the same RMDF module is applied progressively to all the frames' features. For any frame $t$ , the RMDF takes in the corresponding frame's multi-scale features $\{\mathbf{F}_t^1, \mathbf{F}_t^2, \mathbf{F}_t^4\}$ , bilinearly interpolated noisy frames $\{\mathbf{I}_t^1, \mathbf{I}_t^2, \mathbf{I}_t^4\}$ and a hidden state $\mathbf{h}_{t-1}$ as inputs. The multi-scale features are progressively fused in a feed-forward fashion, with Residual Dense Block (RDB) as the basic block, to effectively extract both the short and long range dependencies required for good reconstruction. As shown in Fig. 6(b), multi-scale features are extracted from the noisy frames and fused with the other corresponding input features to minimize any errors accumulated through the previous stages. While these features are utilized to exploit the spatial correlations within the frame, the hidden state $\mathbf{h}$ captures the temporal correlations between all the input frames. Thus, the design of RMDF enables it to extract densely fused multi-scale spatio-temporal features required for enhanced quality outputs. + +Multi-Scale Reconstruction using Residual Frame Refinement Module: Considering the heavy noise in the input quanta frames, this ill-posed problem's restored image subspace can be quite large. To output a restored image close to the ground truth we prefer deep supervision that lets the model preserve critical details of the scene. We opt for a multi-scale reconstruction approach where the image at each scale is reconstructed in a progressive fashion. Experiments reveal the efficacy of this approach (Fig. 5, Fig. 11). The stage starts with the lowest scale features extracted from all the frames using RMDF being concatenated and sent into a newly designed Temporal Cross Attention (TCA) module. As shown in Fig. 7(a), the TCA module is similar to the multi-head attention in vision transformers [16] in terms of generating queries, keys, and values. However, we maintain the number of heads to be one and apply attention only on the channel dimension. The cross attention is from the fact that we input features extracted from all the input frames. The extracted cross-attention features are then fed into a newly designed Residual Frame Refinement Module (RFRM). As shown in Fig. 7(b), RFRM takes in a residual frame $\mathbf{r}_t^\alpha$ , a + +hidden state $\mathbf{f}_t^\alpha$ , and the features as input, concatenates the hidden state with the features to input into the channel attention block [82] to emphasize critical spatio-temporal information. Further, we divide the module into 2 branches. While the former is designed to extract multi-scale spatial correlation information and output a modified hidden state $\mathbf{f}_t^{\alpha /2}$ , the latter focuses on refining the residual frame to output a corresponding scale reconstructed image $\mathbf{O}_t^\alpha$ and a residual frame $\mathbf{r}_t^{\alpha /2}$ . The main purpose of this setup is to initially restore the high-level features through estimating $\mathbf{O}_t^4$ , followed by focusing on the low-level intricate details while refining the residual frames for scales 2 and 1. + +Loss Function. We train QUIVER with a multi-scale loss. The overall loss function can be represented as + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {Q}} = \lambda_ {1} \cdot \mathcal {L} (\mathbf {I} ^ {1, \mathrm {G T}}, \mathbf {I} _ {d} ^ {1}) + \lambda_ {2} \cdot \mathcal {L} (\mathbf {I} _ {t} ^ {1, \mathrm {G T}}, \mathbf {O} _ {t} ^ {1}) + \lambda_ {3} \cdot \mathcal {L} (\mathbf {I} _ {t} ^ {2, \mathrm {G T}}, \mathbf {O} _ {t} ^ {2}) + \dots \\ \lambda_ {4} \cdot \mathcal {L} \left(\mathbf {I} _ {t} ^ {4, \mathrm {G T}}, \mathbf {O} _ {t} ^ {4}\right), \tag {1} \\ \end{array} +$$ + +where $\mathbf{I}_t^{\alpha,\mathrm{GT}}$ is the captured $t^{\mathrm{th}}$ ground truth frame bicubically downsampled by $\alpha$ , and $\mathcal{L}(\mathbf{I}_a, \mathbf{I}_b) = ||\mathbf{I}_a - \mathbf{I}_b||_1 + ||\nabla_x \mathbf{I}_a - \nabla_x \mathbf{I}_b||_1 + ||\nabla_y \mathbf{I}_a - \nabla_y \mathbf{I}_b||_1$ . Here, $\nabla_x$ and $\nabla_y$ represent the operations of computing horizontal and vertical gradients. + +# 4 Proposed I2-2000FPS dataset + +While several high-frame-rate datasets have been open-sourced in recent times [36, 49, 60, 67, 68, 72], these datasets mainly feature videos tainted by severe motion blur, making them unsuitable in our problem setting. Moreover, features such as high motion speed and sufficient number of videos are also not always guaranteed. A visual representation of existing datasets' comparison is shown in Fig. 8(a). To address the gaps, we introduce the I2-2000FPS dataset, a high frame rate video collection meticulously designed to capture high-speed motion with precision. + +The I2-2000FPS dataset has a temporal resolution of 2000 FPS and a spatial resolution of $512 \times 1024$ pixels, comprising 280 unique videos spanning 114 diverse scenes. The videos are captured using the Chronos 1.4 high-speed CMOS sensor-based camera from Kron Technologies. Notably, I2-2000FPS incorporates dark current calibration, leveraging the camera's capabilities to mitigate dark current effects. Throughout the data collection process, analog and digital gain were consistently maintained at 0 dB to avoid amplification of noise. To minimize noise, the videos are exclusively captured outdoor with ambient lighting conditions. More details on I2-2000FPS can be found in the supplementary. + +# 5 Experiments + +# 5.1 Image Formation Model + +For experiments involving synthetic data, we use a single-photon detector simulator based on an underlying image formation model discussed below. We build + +![](images/c60c27b47b93c3ff5cdc147408750ac71e30375e7aa8ba198fb3ba8499464a74.jpg) +(a) + +![](images/ab268fd60e08d107421860a19227ab4fe06f2f630d1a821c9220c6a03e519d59.jpg) +(b) +Fig. 8: (a) Benchmarking high-speed video datasets. Horizontal axis represents the temporal resolution and the vertical axis indicates the maximum speed captured by the dataset, assuming a fixed camera-object distance. The circles in blue and orange indicate blur and blur-free videos, respectively. (b) Benchmarking Quanta Video Restoration Models on the I2-2000FPS dataset. Horizontal axis represents the computational complexity in terms of GFLOPs and the vertical axis indicates the PSNR acquired at 3.25PPP. + +upon the prototype initially suggested in [46] adopted in prior works [6, 10, 11, 20, 26, 28, 56]. + +Given the quanta exposure [22], $\mathbf{I}^{\mathrm{GT}}$ , dependent on the photon flux and exposure time, the observed signal by the sensor can be represented as a Poisson-Gaussian random variable, where the Poisson represents the photon arrival process and the Gaussian models the read noise. The readout process involves various sources of distortions and an Analog-to-Digital Converter (ADC) to convert the real numbers into integers $\{0,1,2,\dots,L\}$ , where $L = 2^{\mathrm{Nbits}} - 1$ depending on the bit-depth (Nbits) allocated to the sensor. The final sensor readout, $\mathbf{Y}$ , can be represented using the following equation, + +$$ +\mathbf {Y} \sim \mathrm {A D C} _ {[ 0, \mathrm {L} ]} \{\mathrm {P o i s s o n (Q E \times I ^ {G T} +} \theta_ {\mathrm {d a r k}}) + \underbrace {\mathrm {G a u s s} (0 , \sigma_ {\mathrm {r e a d}} ^ {2} \mathbf {1})} _ {\mathrm {r e a d n o i s e}} \}. \qquad (2) +$$ + +Akin to previous works [6, 10, 20, 26, 28], we assume our sensor to be monochromatic as we utilize monochromatic real data in our experiments. For our sensor prototype, we utilize a Quantum Efficiency (QE) of 0.80. The dark current $(\theta_{\mathrm{dark}})$ and read noise $(\sigma_{\mathrm{read}})$ are set to $1.6\mathrm{e}^{-}/\mathrm{pix/sec}$ and $0.2\mathrm{e}^{-}/\mathrm{pix}$ , respectively. + +# 5.2 Experimental Settings + +Training data. We curate a set of 249 videos from the I2-2000FPS collection and employ it as the training dataset for all the deep-learning models in our experiments. Each training sample is fetched on the fly from each clip. A training sample here is defined as a tuple containing the ground-truth/target frames and the 3-bit quanta frames simulated at 3.25 photons-per-pixel (PPP) ( $\sim$ 1 lux assuming a $1.1\mu \mathrm{m}$ pixel pitch and a 1/2000 second exposure time) using the + +![](images/70fccee96f5fb920aef792c769f603f1d94bf5a4c8563c7aec956a8243fb4551.jpg) +Fig. 9: Visual comparisons of the reconstructed results on test videos from the proposed I2-2000FPS dataset. For fair comparison, all methods utilize 11 3-bit quanta frames simulated at 3.25 PPP per frame ( $\sim$ 1 lux) to produce a restored frame. Best viewed in zoom. + +image formation model described in Sec. 5.1. + +Testing data. To effectively analyze the performance of various methods, we carefully sample 31 videos from I2-2000FPS containing various motion types, shapes, and speeds. To test the generalizability, we also test the algorithms on X4K1000FPS [67] test dataset containing 15 videos from distinct scenes. Lastly, to measure the performance on real-world data, we collect binary frames using a SPAD sensor [17] and compare the reconstructed outputs. More details will be discussed in Sec. 5.3. + +Baselines. We compare the proposed method with eight existing dynamic scene reconstruction algorithms, namely Transform Denoise [6], QBP [47], Student-Teacher [10], RVRT [42], EMVD [2], FloRNN [1], MemDeblur [35], and Spk2ImgNet [85]. We also add an off-the-shelf denoiser BM3D [14] to QBP, denoted QBP (+BM3D), as a baseline for comparison. As we will discuss in Sec. 5.3, QUIVER beats all the baselines, both quantitatively and qualitatively. + +![](images/7f356c9d60b1328faa2a85feb5ce8510853e17943abc715d35f3f19d150d5460.jpg) +Fig. 10: Performance on Real Quanta Data. We capture real 1-bit quanta data using a SPAD [17] and generate 3-bit frames through temporal averaging. All deep learning based models are trained using a photon level of 4.9 PPP per frame. Best viewed in zoom. + +Training QUIVER. We utilize the function mentioned in Eq. (1) as the cost function for training QUIVER with regularization parameters $\lambda_{1} = 0.2$ , $\lambda_{2} = 0.85$ , $\lambda_{3} = 0.1$ , and $\lambda_{4} = 0.05$ . The training data is extracted with patch size $128 \times 128$ and a batch size of 4. The weights are initialized with Lecun initialization [39]. The network is trained using the Adam optimizer [37] with an initial learning rate of $2.5 \times 10^{-5}$ . The low learning rate is driven by the inherent instability of recurrent networks, as it mitigates the risk of divergent behavior during training. We use a learning rate scheduler that reduces the learning rate by a factor of 2 when a plateau is reached. QUIVER takes approximately 1.5 days to train on a NVIDIA A100 Tensor Core GPU using Pytorch. + +# 5.3 Results + +Synthetic Data Experiments. We begin with the synthetic experiments where we utilize 3-bit quanta frames, simulated using the parameters mentioned in Sec. 5.1 at 3.25, 9.75, 19.5, and 26 PPP to test the algorithms' performance. Tab. 2 and Tab. 3 demonstrate the PSNR and SSIM [76] of various methods extracted by predicting 6017 I2-2000FPS frames and 345 X4K1000FPS frames. To further substantiate the efficacy of QUIVER's design, we introduced a scaled-down variant, QUIVER-s (Refer Fig. 8(b) for complexity comparison). Quantitative results indicate that both QUIVER and QUIVER-s offer substantially better performance than all the baselines across a range of light levels. Fig. 9 depicts visual results of all the methods on the I2-2000FPS dataset. It is evident that existing methods fail to handle both motion and noise simultaneously, whereas, our proposed method, QUIVER, produces blur free high SNR outputs while preserving high-frequency details to a large extent. + +Real Data Experiments. We verify the methods' performance on real data. The real data is collected as binary frames using a SPAD sensor [17] at 10000 FPS with a spatial resolution of $240 \times 320$ . As SPADs possess zero read noise, the binary frames are summed up to generate 3-bit frames. The average observed + +Table 2: Performance comparison on the proposed I2-2000FPS dataset across various light levels. Models are trained using the I2-2000FPS dataset. QUIVER performs significantly better than the existing methods. + +
Photons-Per-Pixel (PPP)3.259.7519.526
MethodPSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑
Transform Denoise [6]21.31700.718423.15210.767122.77480.781222.30960.7811
QBP [47]15.94110.129319.18560.265420.40000.371320.79780.4114
QBP (+ BM3D [14])21.54760.703322.20010.689922.83510.769622.86170.7832
Student-Teacher [10]18.72000.400616.51950.247915.76360.213313.28890.0735
RVRT [42]19.41150.353921.67140.456822.08260.502121.75280.4968
EMVD [2]20.01940.587321.05590.604822.44030.559223.40530.5576
FloRNN [1]21.03410.678525.61320.709127.43220.739527.85200.7784
MemDeblur [35]19.48770.386814.49060.111216.17750.166716.00580.1712
Spk2ImgNet [85]20.39450.564219.66650.673322.93720.700814.97690.6861
QUIVER-s (Ours)24.70130.756526.86760.788327.29890.843227.86590.8408
QUIVER (Ours)26.21430.789726.80580.825027.75380.856327.93770.8446
+ +Table 3: Performance comparison on the X4K1000FPS dataset across various light levels. Models are trained using the I2-2000FPS dataset. QUIVER performs significantly better than the existing methods. + +
Photons-Per-Pixel (PPP)3.259.7519.526
MethodPSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑
Transform Denoise [6]19.62550.632322.17030.704422.99380.722922.62300.7204
QBP [47]15.56340.230216.97580.323017.17980.395717.78070.4188
QBP (+ BM3D [14])17.96770.512318.53080.522618.24070.541418.79170.5586
Student-Teacher [10]18.82080.365216.15480.260814.93590.257113.97620.1186
RVRT [42]19.92030.364121.07810.447221.47800.492520.78990.4919
EMVD [2]20.51020.483621.81520.559522.94400.593622.45870.5860
FloRNN [1]20.82830.577823.58740.648424.32140.668325.24830.7170
MemDeblur [35]19.55340.364214.55950.220316.67490.311615.64960.2974
Spk2ImgNet [85]18.94240.473119.25320.572220.34420.571616.09310.6106
QUIVER-s (Ours)20.91970.595521.79900.652324.19240.731623.44110.7248
QUIVER (Ours)21.87300.652123.16540.705724.59560.764525.00860.7513
+ +light level after summation is 4.9 PPP. We generate results with networks trained at 4.9 PPP and demonstrate the visual results in Fig. 10. QUIVER, as opposed to existing state-of-the-art, effectively recovers high-frequency information while applying visually appealing smoothening effect to low-frequency regions of the scene. It is noteworthy that SPADs' image formation model is significantly different from that of the QIS' imaging model [58,59]. Therefore, the visual results also indicate that the proposed QUIVER can thoroughly generalize to various single-photon detectors. + +# 6 Ablation Study + +Effect of Pre-Denoising, Optical Flow and Multi-Scale. We conduct an ablation study to evaluate the effect of the pre-denoising, the learnable optical flow, and multi-scale reconstruction on performance. Upon removal of either module, we expand the features dimension of layers in the feature extraction stage and add RDB blocks to the RMDF module, thereby maintaining a similar model capacity. We train all possible combinations and display the quantitative results in Tab. 4. Results indicate that, in the absence of either one or more modules, the network's performance is substantially worse. Visual Results in Fig. 11 + +![](images/2c84702399a19cbd297424cb59aeb45ff3d7d64c55e5b4f6ec8446cfc03715c5.jpg) +Fig.11: Ablation Study. Visual Comparisons depicting the effectiveness of Predenoiser, Optical Flow, and Multi-Scale Reconstruction Modules. Best viewed in zoom. + +indicate that these modules serve critical roles as they significantly contribute to model's performance. + +Does loading pre-trained optical flow module weights help? We initialize SPyNet with its pre-trained weights and finetune the same while training QUIVER, and display its quantitative results in Tab. 4. As the pre-trained SPyNet is not robust to photon shot noise and read noise, especially in low-light conditions, initializing the module with it will result in sub-optimal performance. + +Table 4: Ablation study. We conduct experiments to emphasize the role of Denoiser, Optical flow, and Multi-Scale reconstruction modules. We also show the effect of loading pretrained optical flow weights $(^{*})$ on performance. + +
Pre-Denoising Optical Flow Multi-ScaleI2-2000FPS
PSNR↑SSIM↑
×××23.67020.7756
✓✓*✓23.94790.7709
✓✓×24.38410.7808
××✓24.74450.7755
×✓✓24.99990.7753
✓×✓25.75210.7760
✓✓✓26.21430.7897
+ +# 7 Conclusion + +In this paper, we presented a methodology to reconstruct blur-free grayscale images/videos captured using 1-bit or few-bit quanta data. While adopting the ideology of classical quanta restoration methods, we proposed an end-to-end deep learning framework, QUIVER, that utilizes pre-filtering, learnable optical flow module, and a novel multi-scale reconstruction approach to produce high-visual outputs. Experiments on synthetic and real data indicate QUIVER beats state-of-the-art and can generalize across single-photon sensors. We also introduce the world's first high-speed video dataset, I2-2000FPS, that captures fast moving scenes at 2000 fps, covering wide ranges of motion. We believe that I2-2000FPS will be a valuable asset for researchers in high-speed motion analysis and other computer vision tasks. + +Acknowledgements. The work is supported, in part, by Samsung Research America (SRA), DARPA / SRC CogniSense JUMP 2.0 Center, NSF IIS-2133032, and NSF ECCS-2030570. The authors would like to thank German Mora Martin for his help in recording the SPAD data. + +# References + +1. Li et al, J.: Unidirectional Video Denoising by Mimicking Backward Recurrent Modules with Look-Ahead Forward Ones. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) ECCV. Lecture Notes in Computer Science (2022). https://doi.org/10.1007/978-3-031-19797-0_34 +2. Maggioni et al, M.: Efficient Multi-Stage Video Denoising With Recurrent Spatio-Temporal Fusion. In: CVPR (2021) +3. Arias, P., Morel, J.M.: Video Denoising via Empirical Bayesian Estimation of Space-Time Patches. Journal of Mathematical Imaging and Vision 60(1), 70-93 (Jan 2018). https://doi.org/10.1007/s10851-017-0742-4 +4. Chan, K.C.K., Zhou, S., Xu, X., Loy, C.C.: BasicVSR++: Improving Video SuperResolution With Enhanced Propagation and Alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5972-5981 (2022) +5. Chan, S.H.: What Does a One-Bit Quanta Image Sensor Offer? IEEE Transactions on Computational Imaging 8, 770-783 (2022). https://doi.org/10.1109/TCI.2022.3202012 +6. Chan, S.H., Elgendy, O.A., Wang, X.: Images from Bits: Non-Iterative Image Reconstruction for Quanta Image Sensors. Sensors 16(11), 1961 (Nov 2016). https://doi.org/10.3390/s16111961 +7. Chan, S.H., Lu, Y.M.: Efficient image reconstruction for gigapixel quantum image sensors. In: 2014 IEEE Global Conference on Signal and Information Processing (GlobalSIP). pp. 312-316 (Dec 2014). https://doi.org/10.1109/GlobalSIP.2014.7032129 +8. Chan, S.H., Wang, X., Elgendy, O.A.: Plug-and-Play ADMM for Image Restoration: Fixed-Point Convergence and Applications. IEEE Transactions on Computational Imaging 3(1), 84-98 (Mar 2017). https://doi.org/10.1109/TCI.2016.2629286 +9. Charbon, E., Fishburn, M., Walker, R., Henderson, R.K., Niclass, C.: SPAD-Based Sensors. In: Remondino, F., Stoppa, D. (eds.) TOF Range-Imaging Cameras, pp. 11-38. Springer Berlin Heidelberg, Berlin, Heidelberg (2013). https://doi.org/10.1007/978-3-642-27523-4_2 +10. Chi, Y., Gnanasambandam, A., Koltun, V., Chan, S.H.: Dynamic Low-Light Imaging with Quanta Image Sensors. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision - ECCV 2020, vol. 12366, pp. 122-138. Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-58589-1_8 +11. Chi, Y., Zhang, X., Chan, S.H.: HDR imaging with spatially varying signal-to-noise ratios. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5724-5734. IEEE (2023). https://doi.org/10.1109/CVPR52729.2023.00554, https://doi.org/10.1109/CVPR52729.2023.00554 +12. Choi, J.H., Elgendy, O.A., Chan, S.H.: Image Reconstruction for Quanta Image Sensors Using Deep Neural Networks. In: 2018 IEEE International Conference on + +Acoustics, Speech and Signal Processing (ICASSP). pp. 6543-6547 (Apr 2018). https://doi.org/10.1109/ICASSP.2018.8461685 +13. Claus, M., van Gemert, J.: ViDeNN: Deep Blind Video Denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019) +14. Dabov, K., Foi, A., Katkovnik, V., Egiazarian, K.: Image denoising by sparse 3-D transform-domain collaborative filtering. IEEE Transactions on image processing 16(8), 2080-2095 (2007) +15. Dong, S., Huang, T., Tian, Y.: Spike camera and its coding methods. arXiv preprint arXiv:2104.04669 (2021) +16. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +17. Dutton, N.A.W., Gyongy, I., Parmesan, L., Gnecchi, S., Calder, N., Rae, B.R., Pellegrini, S., Grant, L.A., Henderson, R.K.: A SPAD-Based QVGA Image Sensor for Single-Photon Counting and Quanta Imaging. IEEE Transactions on Electron Devices 63(1), 189–196 (Jan 2016). https://doi.org/10.1109/TED.2015.2464682 +18. Dutton, N.A., Gyongy, I., Parmesan, L., Henderson, R.K.: Single photon counting performance and noise analysis of CMOS SPAD-based image sensors. Sensors 16(7), 1122 (2016) +19. Elgendy, O.A., Chan, S.H.: Optimal Threshold Design for Quanta Image Sensor. IEEE Transactions on Computational Imaging 4(1), 99-111 (Mar 2018). https://doi.org/10.1109/TCI.2017.2781185 +20. Feng Yang, Lu, Y.M., Sbaiz, L., Vetterli, M.: Bits From Photons: Oversampled Image Acquisition Using Binary Poisson Statistics. IEEE Transactions on Image Processing 21(4), 1421-1436 (Apr 2012). https://doi.org/10.1109/TIP.2011.2179306 +21. Fossum, E., Ma, J., Masoodian, S., Anzagira, L., Zizza, R.: The Quanta Image Sensor: Every Photon Counts. Sensors 16(8), 1260 (Aug 2016). https://doi.org/10.3390/s16081260 +22. Fossum, E.R.: Modeling the performance of single-bit and multi-bit quanta image sensors. IEEE Journal of the Electron Devices Society 1(9), 166-174 (2013) +23. Gao, J., Shang, Z., Nie, K., Luo, T.: High dynamic range image reconstruction for multi-bit quanta image sensor. Optoelectronics Letters 18(9), 553-558 (Sep 2022). https://doi.org/10.1007/s11801-022-2014-9 +24. Gariepy, G., Krstajic, N., Henderson, R., Li, C., Thomson, R.R., Buller, G.S., Heshmat, B., Raskar, R., Leach, J., Faccio, D.: Single-photon sensitive light-infight imaging. Nature Communications 6(1), 6021 (Jan 2015). https://doi.org/10.1038/ncomms7021 +25. Gariepy, G., Leach, J., Warburton, R., Chan, S., Henderson, R., Faccio, D.: Picosecond time-resolved imaging using SPAD cameras. In: Lewis, K.L., Hollins, R.C. (eds.) SPIE Security + Defence. p. 99920N. Edinburgh, United Kingdom (Oct 2016). https://doi.org/10.1117/12.2241184 +26. Gnanasambandam, A., Chan, S.H.: HDR Imaging With Quanta Image Sensors: Theoretical Limits and Optimal Reconstruction. IEEE Transactions on Computational Imaging 6, 1571-1585 (2020). https://doi.org/10.1109/TCI.2020.3041093 +27. Gnanasambandam, A., Chan, S.H.: Image Classification in the Dark Using Quanta Image Sensors. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Proceed- + +ings of the European Conference on Computer Vision (ECCV). pp. 484-501. Lecture Notes in Computer Science, Cham (2020). https://doi.org/10.1007/978-3-030-58598-3_29 +28. Gnanasambandam, A., Chan, S.H.: Exposure-Referred Signal-to-Noise Ratio for Digital Image Sensors. IEEE Transactions on Computational Imaging 8, 561-575 (2022). https://doi.org/10.1109/TCI.2022.3187657 +29. Godard, C., Matzen, K., Uytendaele, M.: Deep Burst Denoising. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 538-554 (2018) +30. Gupta, A., Ingle, A., Velten, A., Gupta, M.: Photon-Flooded Single-Photon 3D Cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6770-6779 (2019) +31. Gupta, S., Gupta, M.: Eulerian Single-Photon Vision. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10465–10476 (2023) +32. Gutierrez-Barragan, F., Ingle, A., Seets, T., Gupta, M., Velten, A.: Compressive Single-Photon 3D Cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17854-17864 (2022) +33. Gyongy, I., Dutton, N., Henderson, R.: Single-Photon Tracking for High-Speed Vision. Sensors 18(2), 323 (Jan 2018). https://doi.org/10.3390/s18020323 +34. Hasinoff, S.W., Sharlet, D., Geiss, R., Adams, A., Barron, J.T., Kainz, F., Chen, J., Levoy, M.: Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM Transactions on Graphics 35(6), 1-12 (Nov 2016). https://doi.org/10.1145/2980179.2980254 +35. Ji, B., Yao, A.: Multi-Scale Memory-Based Video Deblurring. In: CVPR (2022) +36. Kiani Galoogahi, H., Fagg, A., Huang, C., Ramanan, D., Lucey, S.: Need for Speed: A Benchmark for Higher Frame Rate Object Tracking. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1125-1134 (2017) +37. Kingma, D., Ba, J.: Adam: A Method for Stochastic Optimization. In: International Conference on Learning Representations (ICLR) (2015). https://doi.org/10.48550/arXiv.1412.6980 +38. Lebrun, M., Buades, A., Morel, J.M.: A Nonlocal Bayesian Image Denoising Algorithm. SIAM Journal on Imaging Sciences 6(3), 1665-1688 (Jan 2013). https://doi.org/10.1137/120874989 +39. Lecun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (Nov/1998). https://doi.org/10.1109/5.726791 +40. Li, C., Qu, X., Gnanasambandam, A., Elgendy, O.A., Ma, J., Chan, S.H.: Photon-limited object detection using non-local feature matching and knowledge distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3976-3987 (2021) +41. Liang, J., Cao, J., Fan, Y., Zhang, K., Ranjan, R., Li, Y., Timofte, R., Van Gool, L.: VRT: A Video Restoration Transformer (Jun 2022). https://doi.org/10.48550/arXiv.2201.12288 +42. Liang, J., Fan, Y., Xiang, X., Ranjan, R., Ilg, E., Green, S., Cao, J., Zhang, K., Timofte, R., Gool, L.V.: Recurrent Video Restoration Transformer with Guided Deformable Attention. Advances in Neural Information Processing Systems 35, 378-393 (Dec 2022) +43. Liba, O., Murthy, K., Tsai, Y.T., Brooks, T., Xue, T., Karnad, N., He, Q., Barron, J.T., Sharlet, D., Geiss, R., Hasinoff, S.W., Pritch, Y., Levoy, M.: Handheld mobile photography in very low light. ACM Transactions on Graphics 38(6), 1-16 (Dec 2019). https://doi.org/10.1145/3355089.3356508 + +44. Lindell, D.B., O'Toole, M., Wetzstein, G.: Single-photon 3D imaging with deep sensor fusion. ACM Transactions on Graphics 37(4), 113-1 (2018) +45. Ma, J., Chan, S., Fossum, E.R.: Review of Quanta Image Sensors for UltralowLight Imaging. IEEE Transactions on Electron Devices 69(6), 2824-2839 (Jun 2022). https://doi.org/10.1109/TED.2022.3166716 +46. Ma, J., Masoodian, S., Starkey, D.A., Fossum, E.R.: Photon-number-resolving megapixel image sensor at room temperature without avalanche gain. Optica 4(12), 1474-1481 (Dec 2017). https://doi.org/10.1364/OPTICA.4.001474 +47. Ma, S., Gupta, S., Ulku, A.C., Bruschini, C., Charbon, E., Gupta, M.: Quanta burst photography. ACM Transactions on Graphics 39(4) (Aug 2020). https://doi.org/10.1145/3386569.3392470 +48. Ma, S., Mos, P., Charbon, E., Gupta, M.: Burst Vision Using Single-Photon Cameras. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5375-5385 (2023) +49. Madhusudana, P.C., Yu, X., Birkbeck, N., Wang, Y., Adsumilli, B., Bovik, A.C.: Subjective and Objective Quality Assessment of High Frame Rate Videos. IEEE Access 9, 108069-108082 (2021). https://doi.org/10.1109/ACCESS.2021.3100462 +50. Maggioni, M., Boracchi, G., Foi, A., Egiazarian, K.: Video Denoising, Deblocking, and Enhancement Through Separable 4-D Nonlocal Spatiotemporal Transforms. IEEE Transactions on Image Processing 21(9), 3952-3966 (Sep 2012). https://doi.org/10.1109/TIP.2012.2199324 +51. Monakhova, K., Richter, S.R., Waller, L., Koltun, V.: Dancing Under the Stars: Video Denoising in Starlight. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16241-16251 (2022) +52. Niclass, C., Rochas, A., Besse, P.A., Charbon, E.: Design and characterization of a CMOS 3-D image sensor based on single photon avalanche diodes. IEEE Journal of Solid-State Circuits 40(9), 1847-1854 (Sep 2005). https://doi.org/10.1109/JSSC.2005.848173 +53. Pan, L., Scheerlinck, C., Yu, X., Hartley, R., Liu, M., Dai, Y.: Bringing a Blur Frame Alive at High Frame-Rate With an Event Camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6820-6829 (2019) +54. Pearl, N., Treibitz, T., Korman, S.: NAN: Noise-Aware NeRFs for Burst-Denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12672-12681 (2022) +55. Plotz, T., Roth, S.: Benchmarking Denoising Algorithms With Real Photographs. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1586-1595 (2017) +56. Qu, X., Chi, Y., Chan, S.H.: Spatially varying exposure with 2-by-2 multiplexing: Optimality and universality. IEEE Transactions on Computational Imaging 10, 261-276 (2024) +57. Ranjan, A., Black, M.J.: Optical Flow Estimation Using a Spatial Pyramid Network. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4161-4170 (2017) +58. Rapp, J., Goyal, V.K.: A Few Photons Among Many: Unmixing Signal and Noise for Photon-Efficient Active Imaging. IEEE Transactions on Computational Imaging 3(3), 445-459 (Sep 2017). https://doi.org/10.1109/TCI.2017.2706028 +59. Rapp, J., Ma, Y., Dawson, R.M.A., Goyal, V.K.: Dead Time Compensation for High-Flux Ranging. IEEE Transactions on Signal Processing 67(13), 3471-3486 (Jul 2019). https://doi.org/10.1109/TSP.2019.2914891 + +60. Rebecq, H., Ranftl, R., Koltun, V., Scaramuzza, D.: High Speed and High Dynamic Range Video with an Event Camera. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(6), 1964-1980 (Jun 2021). https://doi.org/10.1109/TPAMI.2019.2963386 +61. Remez, T., Litany, O., Bronstein, A.: A picture is worth a billion bits: Real-time image reconstruction from dense binary threshold pixels. In: 2016 IEEE International Conference on Computational Photography (ICCP). pp. 1-9 (May 2016). https://doi.org/10.1109/ICCPHOT.2016.7492874 +62. Rochas, A.: Single photon avalanche diodes in CMOS technology. Tech. rep., Cite-seer (2003) +63. Ruget, A., Tyler, M., Mora Martin, G., Scholes, S., Zhu, F., Gyongy, I., Hearn, B., McLaughlin, S., Halimi, A., Leach, J.: Pixels2Pose: Super-resolution time-of-flight imaging for 3D pose estimation. Science Advances 8(48), eade0123 (Nov 2022). https://doi.org/10.1126/sciadv.ade0123 +64. Seets, T., Ingle, A., Laurenzis, M., Velten, A.: Motion Adaptive Deblurring With Single-Photon Cameras. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1945–1954 (2021) +65. Shazeer, N.: Glu variants improve transformer. arXiv preprint arXiv:2002.05202 (2020) +66. Shin, D., Xu, F., Venkatraman, D., Lussana, R., Villa, F., Zappa, F., Goyal, V.K., Wong, F.N.C., Shapiro, J.H.: Photon-efficient imaging with a single-photon camera. Nature Communications 7(1), 12046 (Jun 2016). https://doi.org/10.1038/ncomms12046 +67. Sim, H., Oh, J., Kim, M.: XVFI: eXtreme Video Frame Interpolation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14489-14498 (2021) +68. Su, S., Delbracio, M., Wang, J., Sapiro, G., Heidrich, W., Wang, O.: Deep Video Deblurring for Hand-Held Cameras. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1279-1288 (2017) +69. Tassano, M., Delon, J., Veit, T.: DVDNET: A Fast Network for Deep Video Denoising. In: 2019 IEEE International Conference on Image Processing (ICIP). pp. 1805-1809. IEEE, Taipei, Taiwan (Sep 2019). https://doi.org/10.1109/ICIP.2019.8803136 +70. Tassano, M., Delon, J., Veit, T.: FastDVDnet: Towards Real-Time Deep Video Denoising Without Flow Estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1354-1363 (2020) +71. Vaksman, G., Elad, M., Milanfar, P.: Patch Craft: Video Denoising by Deep Modeling and Patch Matching. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2157-2166 (2021) +72. Voeikov, R., Falaleev, N., Baikulov, R.: TTNet: Real-Time Temporal and Spatial Video Analysis of Table Tennis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 884-885 (2020) +73. Vogels, T., Rousselle, F., Mcwilliams, B., Rothlin, G., Harvill, A., Adler, D., Meyer, M., Novak, J.: Denoising with kernel prediction and asymmetric loss functions. ACM Transactions on Graphics 37(4), 124:1-124:15 (Jul 2018). https://doi.org/10.1145/3197517.3201388 +74. Wang, W., Chen, X., Yang, C., Li, X., Hu, X., Yue, T.: Enhancing Low Light Videos by Exploring High Sensitivity Camera Noise. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4111-4119 (2019) + +75. Wang, X.: Single-Photon Cameras Image Reconstruction Using Vision Transformer. In: 2023 IEEE 3rd International Conference on Computer Communication and Artificial Intelligence (CCAI). pp. 296-300 (May 2023). https://doi.org/10.1109/CCAI57533.2023.10201259 +76. Wang, Z., Bovik, A., Sheikh, H., Simoncelli, E.: Image Quality Assessment: From Error Visibility to Structural Similarity. IEEE Transactions on Image Processing 13(4), 600-612 (Apr 2004). https://doi.org/10.1109/TIP.2003.819861 +77. Wong, H.T., Leung, C.S., Ho, D.: Theoretical analysis and image reconstruction for multi-bit quanta image sensors. Signal Processing 185, 108087 (Aug 2021). https://doi.org/10.1016/j.sigpro.2021.108087 +78. Yang, F., Lu, Y.M., Sbaiz, L., Vetterli, M.: An optimal algorithm for reconstructing images from binary measurements. In: Bouman, C.A., Pollak, I., Wolfe, P.J. (eds.) IS&T/SPIE Electronic Imaging. p. 75330K. San Jose, California (Feb 2010). https://doi.org/10.1117/12.850887 +79. Yang, F., Sbaiz, L., Charbon, E., Susstrunk, S., Vetterli, M.: Image reconstruction in the gigavision camera. In: 2009 IEEE 12th International Conference on Computer Vision Workshops, ICCV Workshops. pp. 2212-2219 (Sep 2009). https://doi.org/10.1109/ICCVW.2009.5457554 +80. Zhang, D., Lian, Q., Su, Y., Ren, T.: Dual-Prior Integrated Image Reconstruction for Quanta Image Sensors Using Multi-Agent Consensus Equilibrium. IEEE/CAA Journal of Automatica Sinica 10(6), 1407-1420 (Jun 2023). https://doi.org/10.1109/JAS.2023.123390 +81. Zhang, D., Lian, Q., Yang, Y.: TwP: Two-stage projection framework with manifold constraint for image reconstruction. Digital Signal Processing 141, 104186 (Sep 2023). https://doi.org/10.1016/j.dsp.2023.104186 +82. Zhang, Y., Li, K., Li, K., Wang, L., Zhong, B., Fu, Y.: Image Super-Resolution Using Very Deep Residual Channel Attention Networks. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 286-301 (2018) +83. Zhang, Y., Tian, Y., Kong, Y., Zhong, B., Fu, Y.: Residual Dense Network for Image Restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(7), 2480-2495 (Jul 2021). https://doi.org/10.1109/TPAMI.2020.2968521 +84. Zhao, J., Xiong, R., Huang, T.: High-Speed Motion Scene Reconstruction for Spike Camera via Motion Aligned Filtering. In: 2020 IEEE International Symposium on Circuits and Systems (ISCAS). pp. 1-5. IEEE, Seville, Spain (Oct 2020). https://doi.org/10.1109/ISCAS45731.2020.9181055 +85. Zhao, J., Xiong, R., Liu, H., Zhang, J., Huang, T.: Spk2imgnet: Learning to reconstruct dynamic scene from continuous spike stream. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11996-12005 (2021) +86. Zhao, J., Xiong, R., Xie, J., Shi, B., Yu, Z., Gao, W., Huang, T.: Reconstructing Clear Image for High-Speed Motion Scene With a Retina-Inspired Spike Camera. IEEE Transactions on Computational Imaging 8, 12-27 (2022). https://doi.org/10.1109/TCI.2021.3136446 \ No newline at end of file diff --git a/2024/Quanta Video Restoration/images.zip b/2024/Quanta Video Restoration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..27976a49a659a9ac48afd5b6de5157f16026d11d --- /dev/null +++ b/2024/Quanta Video Restoration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69ccb08fc957e3a3e2ba6686b7229a510952a9bd2c4d5c48d072a90bf57fc7e1 +size 777741 diff --git a/2024/Quanta Video Restoration/layout.json b/2024/Quanta Video Restoration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ac4769f9da3d8af86db8dfe494b9a5f8e5ffe9af --- /dev/null +++ b/2024/Quanta Video Restoration/layout.json @@ -0,0 +1,9910 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 214, + 113, + 400, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 113, + 400, + 129 + ], + "spans": [ + { + "bbox": [ + 214, + 113, + 400, + 129 + ], + "type": "text", + "content": "Quanta Video Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 140, + 149, + 474, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 149, + 474, + 187 + ], + "spans": [ + { + "bbox": [ + 140, + 149, + 474, + 187 + ], + "type": "text", + "content": "Prateek Chennuri1, Yiheng Chi1, Enze Jiang1, G.M. Dilshan Godaliyadda2, Abhiram Gnanasambandam2, Hamid R. Sheikh2, Istvan Gyongy3, and Stanley H. Chan1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 264, + 196, + 349, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 196, + 349, + 206 + ], + "spans": [ + { + "bbox": [ + 264, + 196, + 349, + 206 + ], + "type": "text", + "content": "1 Purdue University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 246, + 207, + 367, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 207, + 367, + 218 + ], + "spans": [ + { + "bbox": [ + 246, + 207, + 367, + 218 + ], + "type": "text", + "content": "Samsung Research America" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 253, + 218, + 361, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 218, + 361, + 228 + ], + "spans": [ + { + "bbox": [ + 253, + 218, + 361, + 228 + ], + "type": "text", + "content": "3 University of Edinburgh" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 201, + 230, + 413, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 230, + 413, + 240 + ], + "spans": [ + { + "bbox": [ + 201, + 230, + 413, + 240 + ], + "type": "text", + "content": "{pchennur,chi14,jiang708,stanchan}@purdue.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 205, + 241, + 408, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 241, + 408, + 251 + ], + "spans": [ + { + "bbox": [ + 205, + 241, + 408, + 251 + ], + "type": "text", + "content": "{dilshan.g,abhiram.g,hr.sheikh}@samsung.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 246, + 252, + 369, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 252, + 369, + 262 + ], + "spans": [ + { + "bbox": [ + 246, + 252, + 369, + 262 + ], + "type": "text", + "content": "{igyongy2}@exseed.ed.ac.uk" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 142, + 274, + 225, + 315 + ], + "blocks": [ + { + "bbox": [ + 142, + 274, + 225, + 315 + ], + "lines": [ + { + "bbox": [ + 142, + 274, + 225, + 315 + ], + "spans": [ + { + "bbox": [ + 142, + 274, + 225, + 315 + ], + "type": "image", + "image_path": "717431360766581130aafba6d78ed325ed59dbd715ae724c05202ebba57c80da.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 316, + 205, + 321 + ], + "lines": [ + { + "bbox": [ + 162, + 316, + 205, + 321 + ], + "spans": [ + { + "bbox": [ + 162, + 316, + 205, + 321 + ], + "type": "text", + "content": "(a) Ground Truth 16-bit" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 225, + 274, + 306, + 315 + ], + "blocks": [ + { + "bbox": [ + 225, + 274, + 306, + 315 + ], + "lines": [ + { + "bbox": [ + 225, + 274, + 306, + 315 + ], + "spans": [ + { + "bbox": [ + 225, + 274, + 306, + 315 + ], + "type": "image", + "image_path": "72f8a4bfd81ddfc6cf365b6676284a7f2c4013fba754cdeef86c3d959eba390d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 316, + 295, + 321 + ], + "lines": [ + { + "bbox": [ + 235, + 316, + 295, + 321 + ], + "spans": [ + { + "bbox": [ + 235, + 316, + 295, + 321 + ], + "type": "text", + "content": "(b) CMOS 16-bit, 60 fps at 1 lux" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 306, + 274, + 388, + 315 + ], + "blocks": [ + { + "bbox": [ + 306, + 274, + 388, + 315 + ], + "lines": [ + { + "bbox": [ + 306, + 274, + 388, + 315 + ], + "spans": [ + { + "bbox": [ + 306, + 274, + 388, + 315 + ], + "type": "image", + "image_path": "800119ba33ffe1f34b74b36b8aab7890dcc9847ac7f3984a09ac19f94891562c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 316, + 378, + 321 + ], + "lines": [ + { + "bbox": [ + 316, + 316, + 378, + 321 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 378, + 321 + ], + "type": "text", + "content": "(c) CMOS 16-bit, 240 fps at 1 lux" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 388, + 274, + 470, + 315 + ], + "blocks": [ + { + "bbox": [ + 388, + 274, + 470, + 315 + ], + "lines": [ + { + "bbox": [ + 388, + 274, + 470, + 315 + ], + "spans": [ + { + "bbox": [ + 388, + 274, + 470, + 315 + ], + "type": "image", + "image_path": "4f6c112ec80eed3beec144f79075088a7834e772eb0833a34e906e039b2cb13f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 316, + 461, + 321 + ], + "lines": [ + { + "bbox": [ + 397, + 316, + 461, + 321 + ], + "spans": [ + { + "bbox": [ + 397, + 316, + 461, + 321 + ], + "type": "text", + "content": "(d) CMOS 16-bit, 2000 fps at 1 lux" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 142, + 322, + 225, + 363 + ], + "blocks": [ + { + "bbox": [ + 142, + 322, + 225, + 363 + ], + "lines": [ + { + "bbox": [ + 142, + 322, + 225, + 363 + ], + "spans": [ + { + "bbox": [ + 142, + 322, + 225, + 363 + ], + "type": "image", + "image_path": "b06a96207be44f26a7ad4b1270b5bce50ca8ce2cbc15a7bf67686639c9f03b99.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 363, + 214, + 369 + ], + "lines": [ + { + "bbox": [ + 153, + 363, + 214, + 369 + ], + "spans": [ + { + "bbox": [ + 153, + 363, + 214, + 369 + ], + "type": "text", + "content": "(e) Quanta 3-bit, 2000 fps at 1 lux" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 388, + 482, + 454 + ], + "lines": [ + { + "bbox": [ + 131, + 388, + 482, + 454 + ], + "spans": [ + { + "bbox": [ + 131, + 388, + 482, + 454 + ], + "type": "text", + "content": "Fig. 1: Goal of this paper. (a) Blur-free video frame of a moving car. (b)-(d) CMOS image sensor simulations using realistic sensor parameters. The strong shot noise and read noise " + }, + { + "bbox": [ + 131, + 388, + 482, + 454 + ], + "type": "inline_equation", + "content": "(5.1\\mathrm{e}^{-}/\\mathrm{pix})" + }, + { + "bbox": [ + 131, + 388, + 482, + 454 + ], + "type": "text", + "content": " of CMOS sensor make the signal acquisition difficult. (e) With low read noise " + }, + { + "bbox": [ + 131, + 388, + 482, + 454 + ], + "type": "inline_equation", + "content": "(0.2\\mathrm{e}^{-}/\\mathrm{pix})" + }, + { + "bbox": [ + 131, + 388, + 482, + 454 + ], + "type": "text", + "content": ", low-bit single-photon detectors capture valuable information. (f)-(g) Existing state-of-the-art algorithm, QBP [47] cannot handle strong motion and noise. (h) The proposed algorithm, QUIVER, produces high quality results." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 225, + 322, + 306, + 363 + ], + "blocks": [ + { + "bbox": [ + 225, + 322, + 306, + 363 + ], + "lines": [ + { + "bbox": [ + 225, + 322, + 306, + 363 + ], + "spans": [ + { + "bbox": [ + 225, + 322, + 306, + 363 + ], + "type": "image", + "image_path": "e28400dbfc1bfff024ff7fe8fe45c537123c8bcc2dc0d2b2fae4933cb5cf2bca.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 364, + 303, + 374 + ], + "lines": [ + { + "bbox": [ + 227, + 364, + 303, + 374 + ], + "spans": [ + { + "bbox": [ + 227, + 364, + 303, + 374 + ], + "type": "text", + "content": "f) Quanta 3-bit, reconstructed from 11-frame avg. using Traditional Methods [47]" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 306, + 322, + 388, + 363 + ], + "blocks": [ + { + "bbox": [ + 306, + 322, + 388, + 363 + ], + "lines": [ + { + "bbox": [ + 306, + 322, + 388, + 363 + ], + "spans": [ + { + "bbox": [ + 306, + 322, + 388, + 363 + ], + "type": "image", + "image_path": "3a5136568023e9a68a1a504178093a13cb9e6dda89203ff7fb294173bac88216.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 364, + 386, + 374 + ], + "lines": [ + { + "bbox": [ + 309, + 364, + 386, + 374 + ], + "spans": [ + { + "bbox": [ + 309, + 364, + 386, + 374 + ], + "type": "text", + "content": "(g) Quanta 3-bit, reconstructed from 66-frame avg, using Traditional Methods [47]" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 388, + 364, + 470, + 374 + ], + "lines": [ + { + "bbox": [ + 388, + 364, + 470, + 374 + ], + "spans": [ + { + "bbox": [ + 388, + 364, + 470, + 374 + ], + "type": "text", + "content": "(h) Quanta 3-bit, reconstructed from 11-frame avg using QUIVER" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 388, + 322, + 470, + 363 + ], + "blocks": [ + { + "bbox": [ + 388, + 322, + 470, + 363 + ], + "lines": [ + { + "bbox": [ + 388, + 322, + 470, + 363 + ], + "spans": [ + { + "bbox": [ + 388, + 322, + 470, + 363 + ], + "type": "image", + "image_path": "67a01e4617af268a2feb53a47ff71f8049f7205a092816eb76a5ffa442fe0698.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 159, + 468, + 453, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 468, + 453, + 633 + ], + "spans": [ + { + "bbox": [ + 159, + 468, + 453, + 633 + ], + "type": "text", + "content": "Abstract. The proliferation of single-photon image sensors has opened the door to a plethora of high-speed and low-light imaging applications. However, data collected by these sensors are often 1-bit or few-bit, and corrupted by noise and strong motion. Conventional video restoration methods are not designed to handle this situation, while specialized quanta burst algorithms have limited performance when the number of input frames is low. In this paper, we introduce Quanta Video Restoration (QUIVER), an end-to-end trainable network built on the core ideas of classical quanta restoration methods, i.e., pre-filtering, flow estimation, fusion, and refinement. We also collect and publish I2-2000FPS, a high-speed video dataset with the highest temporal resolution of 2000 frames-per-second, for training and testing. On simulated and real data, QUIVER outperforms existing quanta restoration methods by a significant margin. Code and dataset available at https://github.com/ chennuripratek/Quanta_Video_Restoration-QUIVER" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 160, + 643, + 453, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 643, + 453, + 664 + ], + "spans": [ + { + "bbox": [ + 160, + 643, + 453, + 664 + ], + "type": "text", + "content": "Keywords: Single Photon Detectors " + }, + { + "bbox": [ + 160, + 643, + 453, + 664 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 643, + 453, + 664 + ], + "type": "text", + "content": " Video Restoration " + }, + { + "bbox": [ + 160, + 643, + 453, + 664 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 643, + 453, + 664 + ], + "type": "text", + "content": " High-Speed Dataset" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 230, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 230, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 230, + 127 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 166, + 482, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 166, + 482, + 298 + ], + "spans": [ + { + "bbox": [ + 130, + 166, + 482, + 298 + ], + "type": "text", + "content": "Over the past decade, the astonishing growth of single-photon detectors has fundamentally changed the landscape of computational imaging. With the invention and proliferation of quanta image sensors (QIS) [21] and single-photon avalanche diodes (SPAD) [52, 62], there is an unprecedented volume of new applications in low-light imaging [6, 10, 66], computer vision [27, 31, 40], high-speed videography [47, 48], time-of-flight sensing [32, 63], and 3D imaging [30, 44]. In most of these use cases, the main core question that lies is how to recover the image from the photon counts measured in the scene. Specifically, given a video stream of 1-bit or few-bit data captured from a scene involving moving objects, how do we reconstruct a gray-scale image/video while eliminating the noise without incurring motion blur?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 303, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 460 + ], + "type": "text", + "content": "To give the reader a visual perspective of the problem scope, Fig. 1 depicts a blur-free video of a moving car. We simulate the captured images at 1 lux assuming 60 fps, 240 fps, and 2000 fps CMOS image sensors with realistic sensor specifications. As illustrated in the figure, the resulting CMOS outputs are either severely blurred due to strong motion or completely distorted by noise due to sparse photons. In the same figure, we demonstrate a simulated single-photon camera output (a 3-bit QIS in this case) where the content the largely preserved despite heavy noise. Upon utilizing state-of-the-art Quanta Burst Photography (QBP) [47] for reconstructing the frames, provided the motion is slow, a decent output can be obtained. However, as the temporal window narrows down, as shown in Fig. 1 (f), the noise remains. In this paper, we address this problem with a new algorithm, designed to remove the noise while avoiding distortions in the presence of fast motion while utilizing only a few frames." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 464, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 525 + ], + "type": "text", + "content": "The core innovation of this paper is QUanta Video Restoration (QUIVER), a deep-learning based video restoration algorithm for quanta image data. QUIVER is specialized for few-bit data (3-bit) captured at thousands of frames-per-second (2000 fps) with an average motion range of 1 to 7 pixels per frame. The main contributions of this paper can be summarized as follows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 552, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 137, + 552, + 481, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 552, + 481, + 613 + ], + "spans": [ + { + "bbox": [ + 137, + 552, + 481, + 613 + ], + "type": "text", + "content": "- We propose QUIVER, an end-to-end trainable quanta (video) restoration method built by embracing the core ideas from traditional quanta restoration algorithms. On a comprehensive evaluation dataset containing both simulated and real data, QUIVER outperforms all methods we compared in this paper by a significant margin." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 137, + 617, + 481, + 666 + ], + "type": "text", + "content": "- We introduce I2-2000FPS, the first high-speed video dataset with a temporal resolution of 2000 frames-per-second for training and testing image and video reconstruction neural networks. We captured a total of 280 high-speed videos covering 114 distinct scenes with ground truth and simulated 3-bit videos." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "P. Chennuri et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 114, + 443, + 222 + ], + "blocks": [ + { + "bbox": [ + 167, + 114, + 443, + 222 + ], + "lines": [ + { + "bbox": [ + 167, + 114, + 443, + 222 + ], + "spans": [ + { + "bbox": [ + 167, + 114, + 443, + 222 + ], + "type": "image", + "image_path": "072ec0ae0cceaf8219c9aee1913c2fbb23e1b6ae6b49c88aee0b8831e59a9b5b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 227, + 482, + 295 + ], + "lines": [ + { + "bbox": [ + 130, + 227, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 482, + 295 + ], + "type": "text", + "content": "Fig. 2: Motion Blur and SNR Trade-off. The effects of bit depth on SNR and motion blur are illustrated using real captures by a single-photon sensor. For the motion range we target, 3-bit single-photon detectors provide the best trade-off between blur and SNR. The images are captured using a 1-bit SPAD [17] at " + }, + { + "bbox": [ + 130, + 227, + 482, + 295 + ], + "type": "inline_equation", + "content": "10\\mathrm{k}" + }, + { + "bbox": [ + 130, + 227, + 482, + 295 + ], + "type": "text", + "content": " fps at an average photon level of 0.51 and 0.40 photons-per-pixel (PPP) per frame, respectively. Higher bit-depth outputs are generated through temporal frame averaging." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 219, + 338, + 391, + 412 + ], + "blocks": [ + { + "bbox": [ + 131, + 306, + 481, + 329 + ], + "lines": [ + { + "bbox": [ + 131, + 306, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 131, + 306, + 481, + 329 + ], + "type": "text", + "content": "Table 1: Frame-rate, motion, read-noise, and data-rate statistics for various bit-depths at the same exposure level." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 219, + 338, + 391, + 412 + ], + "lines": [ + { + "bbox": [ + 219, + 338, + 391, + 412 + ], + "spans": [ + { + "bbox": [ + 219, + 338, + 391, + 412 + ], + "type": "table", + "html": "
Bit-DepthfpsMotion (pixels/frame)σread (/pixel/sec)Data-rate (Mb/sec)
110k0 - 12000 e-96
314282 - 3285.6 e-41.13
53236 - 1264.6 e-15.5
77825 - 3015.6 e-5.24
92070 - 804 e-1.73
", + "image_path": "f0751314eb054760e6f73b2292a750e1140bb5bc9e7f2dcaf47e99ea16358d82.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 434, + 226, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 226, + 448 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 226, + 448 + ], + "type": "text", + "content": "2 Background" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 462, + 326, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 462, + 326, + 475 + ], + "spans": [ + { + "bbox": [ + 131, + 462, + 326, + 475 + ], + "type": "text", + "content": "2.1 Few-bit Single-Photon Detectors" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 485, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 569 + ], + "type": "text", + "content": "What is few-bit photon counting? Single-photon detectors (QIS and SPAD) differ from conventional CMOS pixels by their extraordinary photon counting capability. QIS uses a two-stage pump-gate technique and correlated double sampling to suppress the read noise, while SPAD uses avalanche multiplication to amplify the photocharge. In both cases, the sensors are capable of resolving photons up to a single-photon sensitivity. We refer readers interested in the sensor development of QIS and SPAD to consult, for example, [5,9,18,45]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Along with the single-photon detectors' unique capability to count individual photons, these devices can generate data at a bit-depth as low as 1-bit to as high as 16-bit or even more. However, higher bit-depth is accompanied by longer integration time. If the scene contains motion, longer integration time will eventually result in strong motion blurs as shown in Fig. 2. On the other hand, 1-bit sensing with high frame rates will result in motion-blur-free but extremely noisy images. Therefore, from a pure data acquisition perspective, there exists an optimal bit-depth with respect to the motion that will give us mini-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "text", + "content": "QUIVER" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "mal/no motion-blur data with a minimum per-frame signal-to-noise ratio (SNR) required for good quality reconstruction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": "How about 1-bit and reconstruct afterward? Readers familiar with single-photon counting may wonder whether we can collect as many 1-bit frames as possible and then process the data afterward. The problem is power consumption and data rate. Fixing the same level of exposure, as described in Tab. 1, a 1-bit video at " + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "inline_equation", + "content": "10\\mathrm{k}" + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": " fps would require " + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "inline_equation", + "content": "96\\mathrm{Mb / sec}" + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": " whereas a 9-bit video at 20 fps would only need " + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "inline_equation", + "content": "1.73\\mathrm{Mb / sec}" + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": ". Another problem is read noise accumulation. For sensors with non-zero read noise (such as QIS), every frame contributes to a finite amount of read noise. The more frames we read, the more read noise we accumulate. Therefore, recording 1-bit is not always the best option." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 264, + 231, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 264, + 231, + 275 + ], + "spans": [ + { + "bbox": [ + 131, + 264, + 231, + 275 + ], + "type": "text", + "content": "2.2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 282, + 482, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 482, + 365 + ], + "type": "text", + "content": "Image and Video Denoising. Classical state-of-the-art methods utilize a non-local strategy to identify similar patches across an image/video [3, 38, 50]. Deep neural networks have been proven to be successful in producing high quality denoised outputs [13, 69-71]. Among these architectures, Vision Transformers [41, 42] have been rated the state-of-the-art in recent times. However, all these solutions make simplistic assumptions on noise statistics, thus failing to perform on real noisy images or videos [55]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 366, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 461 + ], + "type": "text", + "content": "Coming to low light, Burst denoising [34, 43], where images are aligned, merged and denoised, is one of the most popular methods. However, these methods fail without robust alignment. To overcome this, a number of alternative solutions with learnable alignment modules have been proposed [29, 54, 73]. Recent solutions have focused on practical noise models that replicate real camera sensor noise, to produce visually appealing results [51, 74]. Nevertheless, the existing solutions utilize images captured using CMOS image sensors, resulting in a notably higher photon level compared to the one utilized in our study." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 462, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 628 + ], + "type": "text", + "content": "SPADs, Event, and Spike Cameras. Gariepy et al. [24, 25] firstly introduced the utilization of Single Photon Avalanche Diodes (SPADs) at pico-second temporal resolution to capture light in motion. Gyongy et al. [33] demonstrated 2D motion tracking of rigid planar objects using SPADs at " + }, + { + "bbox": [ + 130, + 462, + 482, + 628 + ], + "type": "inline_equation", + "content": "10\\mathrm{k}" + }, + { + "bbox": [ + 130, + 462, + 482, + 628 + ], + "type": "text", + "content": " fps. Recently, Ma et al. [47] and Seets et al. [64] utilized SPADs in a passive imaging setting to capture motion in low illumination. However, all these methods utilize extremely high-temporal resolutions hindering the deployment of these sensors into consumer devices where bandwidth is the bottleneck. Event [53, 60] and Spike Cameras [84, 86] also have demonstrated their effectiveness in capturing high-speed motion. However, these cameras focus on luminance/brightness variation and record a spike only when variation is above a threshold (changes based on factors like temperature, event rate, etc.) [15, 60]. Therefore, unlike single-photon detectors (QIS and SPADs), these cameras are NOT designed for single-photon counting and cannot operate in extreme low-light conditions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "QIS Reconstruction. Reconstructing quanta images is a challenging task due to the underlying Poisson-Gaussian statistics. Initial solutions to this problem included utilizing standard gradient descent [78], greedy algorithms [79]," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "P. Chennuri et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 114, + 473, + 163 + ], + "blocks": [ + { + "bbox": [ + 142, + 114, + 473, + 163 + ], + "lines": [ + { + "bbox": [ + 142, + 114, + 473, + 163 + ], + "spans": [ + { + "bbox": [ + 142, + 114, + 473, + 163 + ], + "type": "image", + "image_path": "b6a0645154d3686245d723db34243ef0054eda139319e1bec42f459e3a01387e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 171, + 480, + 194 + ], + "lines": [ + { + "bbox": [ + 132, + 171, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 480, + 194 + ], + "type": "text", + "content": "Fig. 3: Traditional Methods' Design. Depiction of existing classical quanta restoration algorithms' design philosophy. Best viewed in zoom." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 218, + 482, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 218, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 218, + 482, + 327 + ], + "type": "text", + "content": "(ADMM) [7,8], among others [19,20,23,26,61,77]. Chan et al. [6] were the first to propose a non-iterative approach using Anscombe transform for reconstructing quanta images. Choi et al. [12] proposed the first end-to-end trainable deep neural network (DNN) for QIS reconstruction. Alternative DNN-based solutions include utilizing vision transformers [75], Dual Prior Integrated networks [80], among others [81]. Nonetheless, all these methods fail to produce good results when the scene is in motion. Chi et al. [10] is the only method which focuses on capturing dynamic scenes using QIS but only targets at slow motion (1 pixel/frame)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 346, + 386, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 346, + 386, + 361 + ], + "spans": [ + { + "bbox": [ + 132, + 346, + 386, + 361 + ], + "type": "text", + "content": "3 QUanta VIDEo REstoration (QUIVER)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 373, + 254, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 373, + 254, + 387 + ], + "spans": [ + { + "bbox": [ + 132, + 373, + 254, + 387 + ], + "type": "text", + "content": "3.1 Design Philosophy" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 396, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 601 + ], + "type": "text", + "content": "In this section we present the design of our proposed algorithm. We start by briefly reviewing the design of classical methods [33, 47] which, to an extent, have been successful in restoring quanta images. As shown in Fig. 3, classical methods' algorithm design can be divided into four stages: (1) computing sum images to increase SNR, (2) optical flow (or) transformation matrix estimation for aligning the input frames, (3) warping and linear combination for generating preliminary restored output, and (4) refinement for producing the final output. While the steps seem intuitive and straightforward, existing methods are heavily vulnerable to extreme noise and strong motion in the input frames primarily due to two reasons. (1) none of the stages are designed to handle extreme noise and strong motion simultaneously (will be discussed further). (2) Since all the stages are sequential yet independent of each other, it is difficult to obtain an optimal result for a wide range of noise and motion. Our proposed algorithm QUIVER, leverages the design philosophy of existing classical methods while designing each stage to simultaneously handle both noise and motion. Moreover, QUIVER is an end-to-end trainable model making all the stages inter-dependent, thus leading to good restoration outputs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 619, + 259, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 619, + 259, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 619, + 259, + 632 + ], + "type": "text", + "content": "3.2 Design of QUIVER" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "QUIVER is a deep-learning-based video restoration method for quanta imaging. The design philosophy of QUIVER is to adopt the intuitive thoughts behind" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "text", + "content": "QUIVER" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 116, + 279, + 213 + ], + "blocks": [ + { + "bbox": [ + 153, + 116, + 279, + 213 + ], + "lines": [ + { + "bbox": [ + 153, + 116, + 279, + 213 + ], + "spans": [ + { + "bbox": [ + 153, + 116, + 279, + 213 + ], + "type": "image", + "image_path": "5c7f37774aa0261c6b87616ddb40d180a26679556a92da9868216b68ec56eceb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 213, + 223, + 223 + ], + "lines": [ + { + "bbox": [ + 212, + 213, + 223, + 223 + ], + "spans": [ + { + "bbox": [ + 212, + 213, + 223, + 223 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 286, + 116, + 465, + 213 + ], + "blocks": [ + { + "bbox": [ + 286, + 116, + 465, + 213 + ], + "lines": [ + { + "bbox": [ + 286, + 116, + 465, + 213 + ], + "spans": [ + { + "bbox": [ + 286, + 116, + 465, + 213 + ], + "type": "image", + "image_path": "d42c30ab7da6ddb25b54d18efe25c754167206411928a384ee6f4c4199841416.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 213, + 383, + 222 + ], + "lines": [ + { + "bbox": [ + 372, + 213, + 383, + 222 + ], + "spans": [ + { + "bbox": [ + 372, + 213, + 383, + 222 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 133, + 313, + 478, + 430 + ], + "blocks": [ + { + "bbox": [ + 130, + 233, + 482, + 300 + ], + "lines": [ + { + "bbox": [ + 130, + 233, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 482, + 300 + ], + "type": "text", + "content": "Fig. 4: Traditional Methods' Limitations. (a) Traditional methods' [33, 47] predenoising/temporal-averaging fails in strong motion. It is visible in the restored images that an input with strong motion between the frames results in several artifacts in the output even though SNR levels are similar. (b) Traditional methods [47] utilize a patch-based pre-trained optical flow module similar to [34]. This optical flow module fails to compensate for motion in the presence of noise." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 313, + 478, + 430 + ], + "lines": [ + { + "bbox": [ + 133, + 313, + 478, + 430 + ], + "spans": [ + { + "bbox": [ + 133, + 313, + 478, + 430 + ], + "type": "image", + "image_path": "b756e46082b42a0320042036eede933e7b2c3d4978c22bdf302fa000d7b96400.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 440, + 483, + 472 + ], + "lines": [ + { + "bbox": [ + 130, + 440, + 483, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 483, + 472 + ], + "type": "text", + "content": "Fig.5: The proposed QUIVER network. The corresponding stages of QUIVER, built by embracing the intuitive thoughts behind existing classical methods. Best viewed in zoom." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 480, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 480, + 533 + ], + "type": "text", + "content": "classical quanta restoration methods and develop an end-to-end trainable, robust to noise and motion deep learning based framework, as shown in Fig. 5. Specifically, QUIVER can be divided into four main stages:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "Pre-Denoising to improve SNR: Since the input quanta frames possess extreme noise, classical methods adopt naive averaging to increase the SNR and thereby predict better optical flows or transformation matrices. However, as shown in Fig. 4(a), the simple averaging is vulnerable to motion and will negatively impact subsequent processing, ultimately leading to distorted outputs. Simply eliminating this stage is not the solution, because it leads to poor optical flow estimation, resulting in over-smoothed outputs with lack of low-level intricate details, as shown in Fig. 5 and Fig. 11. Therefore, a preliminary denoising step robust to noise and motion is crucial. In QUIVER, we leverage a computational undemanding single image denoiser built using Residual Dense Blocks [83] (RDBs) to provide minimal pre-preprocessing of the input quanta" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "P. Chennuri et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 278, + 198 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 278, + 198 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 278, + 198 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 278, + 198 + ], + "type": "image", + "image_path": "ebaaf12819d0307f87a753e9c3cd487c6c9fc5e0b48962895385d2af119a9b8c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 206, + 434, + 217 + ], + "lines": [ + { + "bbox": [ + 178, + 206, + 434, + 217 + ], + "spans": [ + { + "bbox": [ + 178, + 206, + 434, + 217 + ], + "type": "text", + "content": "Fig. 6: Design of the proposed modules DC-GFU and RMDF." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 280, + 114, + 480, + 198 + ], + "blocks": [ + { + "bbox": [ + 280, + 114, + 480, + 198 + ], + "lines": [ + { + "bbox": [ + 280, + 114, + 480, + 198 + ], + "spans": [ + { + "bbox": [ + 280, + 114, + 480, + 198 + ], + "type": "image", + "image_path": "232f176b6f5c7320a3c9b211ba6e2d98d0d23ec67cb9a63335249cac1d1db811.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 243, + 479, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 479, + 279 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 479, + 279 + ], + "type": "text", + "content": "data. A multi-frame denoiser is not an option due to it computationally demanding nature. We use RDB based network due to its history in handling noise while preserving details with its simple yet effective design." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 280, + 481, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 280, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 481, + 435 + ], + "type": "text", + "content": "Optical Flow Estimation + Feature Alignment using Deformable Convolution - Gated Fusion Unit: Classical methods utilize an off-the-shelf pre-trained optical flow estimation module or predict a transformation matrix to compensate for motion between the frames. The basic assumption behind such approaches is that the motion between the frames is limited and the SNR is high enough. However, when such assumption is not met, the motion compensation is sub-optimal, as shown in Fig. 4(b). As most state-of-the-art pre-trained optical flow estimators are optimized on the CMOS RGB sensor images, it leads to sub-optimal performance when applied on quanta frames. Eliminating the flow estimation step is not recommended since experiments reveal the critical role it plays in motion compensation, as shown in Fig. 5 and Fig. 11. For QUIVER, we deploy a learnable optical flow estimation module and utilize SpyNet [57] owing to its computational efficiency while using a multi-scale approach." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 436, + 481, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 481, + 472 + ], + "type": "text", + "content": "We deploy 3D convolution blocks to extract multi-scale spatio-temporal features from both the noisy and denoised quanta frames. We reuse the noisy frames to compensate for any information lost in the pre-denoising stage." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 473, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 481, + 628 + ], + "type": "text", + "content": "The estimated multi-scale robust-to-noise optical flows are utilised for feature-level alignment of the extracted multi-scale spatio-temporal features. We utilize the deformable convolution with residual offsets proposed by [4] to warp the features. Inspired by the superior performance of Gated Linear Units (GLUs) in Transformers [65] we design and add a GLU based multi-layer-perceptron layer with GeLU activation for efficiently fusing the aligned features extracted from both the noisy and denoised frames. As shown in Fig. 6(a), we name this deformable convolution-GLU combination as the DC-GFU module. At this fusion-stage each frame is processed separately and the fusion is performed only along the channel dimension. Recurrence is applied for the alignment stage across all the multi-scale features of each frame. For the fusion module, we do not employ recurrence, as different scale features capture distinct long-range dependencies owing to their varying receptive fields." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "Deep Feature Fusion using Recurrent Multi-scale Residual Dense Feature Fusion Unit: Post warping we want to perform a robust-to-noise dense feature fusion while taking advantage of the temporal correlations among" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "text", + "content": "QUIVER" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 310, + 217 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 310, + 217 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 310, + 217 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 310, + 217 + ], + "type": "image", + "image_path": "37909717e764ae1269c14e17a918a810d8e05332b3e7b5ead6b27d6c378f1d82.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 225, + 425, + 237 + ], + "lines": [ + { + "bbox": [ + 187, + 225, + 425, + 237 + ], + "spans": [ + { + "bbox": [ + 187, + 225, + 425, + 237 + ], + "type": "text", + "content": "Fig. 7: Design of the proposed modules TCA and RFRM." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 313, + 114, + 479, + 216 + ], + "blocks": [ + { + "bbox": [ + 313, + 114, + 479, + 216 + ], + "lines": [ + { + "bbox": [ + 313, + 114, + 479, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 479, + 216 + ], + "type": "image", + "image_path": "6cdfb18d9d26e7805896c65eba94777ed07ffadc0be2020ad6d22caaf23958d8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 267, + 482, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 482, + 471 + ], + "type": "text", + "content": "the features of all the input frames and also the spatial correlations between the multi-scale features within the same frame. For this task, we design and propose a Recurrent Multi-scale Residual Dense Feature Fusion Unit (RMDF) as shown in Fig. 6(b). The recurrence comes from the fact that the same RMDF module is applied progressively to all the frames' features. For any frame " + }, + { + "bbox": [ + 130, + 267, + 482, + 471 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 267, + 482, + 471 + ], + "type": "text", + "content": ", the RMDF takes in the corresponding frame's multi-scale features " + }, + { + "bbox": [ + 130, + 267, + 482, + 471 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{F}_t^1, \\mathbf{F}_t^2, \\mathbf{F}_t^4\\}" + }, + { + "bbox": [ + 130, + 267, + 482, + 471 + ], + "type": "text", + "content": ", bilinearly interpolated noisy frames " + }, + { + "bbox": [ + 130, + 267, + 482, + 471 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_t^1, \\mathbf{I}_t^2, \\mathbf{I}_t^4\\}" + }, + { + "bbox": [ + 130, + 267, + 482, + 471 + ], + "type": "text", + "content": " and a hidden state " + }, + { + "bbox": [ + 130, + 267, + 482, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{t-1}" + }, + { + "bbox": [ + 130, + 267, + 482, + 471 + ], + "type": "text", + "content": " as inputs. The multi-scale features are progressively fused in a feed-forward fashion, with Residual Dense Block (RDB) as the basic block, to effectively extract both the short and long range dependencies required for good reconstruction. As shown in Fig. 6(b), multi-scale features are extracted from the noisy frames and fused with the other corresponding input features to minimize any errors accumulated through the previous stages. While these features are utilized to exploit the spatial correlations within the frame, the hidden state " + }, + { + "bbox": [ + 130, + 267, + 482, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 130, + 267, + 482, + 471 + ], + "type": "text", + "content": " captures the temporal correlations between all the input frames. Thus, the design of RMDF enables it to extract densely fused multi-scale spatio-temporal features required for enhanced quality outputs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "content": "Multi-Scale Reconstruction using Residual Frame Refinement Module: Considering the heavy noise in the input quanta frames, this ill-posed problem's restored image subspace can be quite large. To output a restored image close to the ground truth we prefer deep supervision that lets the model preserve critical details of the scene. We opt for a multi-scale reconstruction approach where the image at each scale is reconstructed in a progressive fashion. Experiments reveal the efficacy of this approach (Fig. 5, Fig. 11). The stage starts with the lowest scale features extracted from all the frames using RMDF being concatenated and sent into a newly designed Temporal Cross Attention (TCA) module. As shown in Fig. 7(a), the TCA module is similar to the multi-head attention in vision transformers [16] in terms of generating queries, keys, and values. However, we maintain the number of heads to be one and apply attention only on the channel dimension. The cross attention is from the fact that we input features extracted from all the input frames. The extracted cross-attention features are then fed into a newly designed Residual Frame Refinement Module (RFRM). As shown in Fig. 7(b), RFRM takes in a residual frame " + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_t^\\alpha" + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "content": ", a" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "P. Chennuri et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 225 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 225 + ], + "type": "text", + "content": "hidden state " + }, + { + "bbox": [ + 130, + 116, + 479, + 225 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t^\\alpha" + }, + { + "bbox": [ + 130, + 116, + 479, + 225 + ], + "type": "text", + "content": ", and the features as input, concatenates the hidden state with the features to input into the channel attention block [82] to emphasize critical spatio-temporal information. Further, we divide the module into 2 branches. While the former is designed to extract multi-scale spatial correlation information and output a modified hidden state " + }, + { + "bbox": [ + 130, + 116, + 479, + 225 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t^{\\alpha /2}" + }, + { + "bbox": [ + 130, + 116, + 479, + 225 + ], + "type": "text", + "content": ", the latter focuses on refining the residual frame to output a corresponding scale reconstructed image " + }, + { + "bbox": [ + 130, + 116, + 479, + 225 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_t^\\alpha" + }, + { + "bbox": [ + 130, + 116, + 479, + 225 + ], + "type": "text", + "content": " and a residual frame " + }, + { + "bbox": [ + 130, + 116, + 479, + 225 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_t^{\\alpha /2}" + }, + { + "bbox": [ + 130, + 116, + 479, + 225 + ], + "type": "text", + "content": ". The main purpose of this setup is to initially restore the high-level features through estimating " + }, + { + "bbox": [ + 130, + 116, + 479, + 225 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_t^4" + }, + { + "bbox": [ + 130, + 116, + 479, + 225 + ], + "type": "text", + "content": ", followed by focusing on the low-level intricate details while refining the residual frames for scales 2 and 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 227, + 479, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 227, + 479, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 479, + 251 + ], + "type": "text", + "content": "Loss Function. We train QUIVER with a multi-scale loss. The overall loss function can be represented as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 148, + 258, + 480, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 258, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 148, + 258, + 480, + 289 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {Q}} = \\lambda_ {1} \\cdot \\mathcal {L} (\\mathbf {I} ^ {1, \\mathrm {G T}}, \\mathbf {I} _ {d} ^ {1}) + \\lambda_ {2} \\cdot \\mathcal {L} (\\mathbf {I} _ {t} ^ {1, \\mathrm {G T}}, \\mathbf {O} _ {t} ^ {1}) + \\lambda_ {3} \\cdot \\mathcal {L} (\\mathbf {I} _ {t} ^ {2, \\mathrm {G T}}, \\mathbf {O} _ {t} ^ {2}) + \\dots \\\\ \\lambda_ {4} \\cdot \\mathcal {L} \\left(\\mathbf {I} _ {t} ^ {4, \\mathrm {G T}}, \\mathbf {O} _ {t} ^ {4}\\right), \\tag {1} \\\\ \\end{array}", + "image_path": "505103d9929626840866936eef4c75b8173c0a4bc5a09f1e77f6507dc2f7159a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 298, + 479, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 479, + 336 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 479, + 336 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 298, + 479, + 336 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_t^{\\alpha,\\mathrm{GT}}" + }, + { + "bbox": [ + 130, + 298, + 479, + 336 + ], + "type": "text", + "content": " is the captured " + }, + { + "bbox": [ + 130, + 298, + 479, + 336 + ], + "type": "inline_equation", + "content": "t^{\\mathrm{th}}" + }, + { + "bbox": [ + 130, + 298, + 479, + 336 + ], + "type": "text", + "content": " ground truth frame bicubically downsampled by " + }, + { + "bbox": [ + 130, + 298, + 479, + 336 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 298, + 479, + 336 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 298, + 479, + 336 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\mathbf{I}_a, \\mathbf{I}_b) = ||\\mathbf{I}_a - \\mathbf{I}_b||_1 + ||\\nabla_x \\mathbf{I}_a - \\nabla_x \\mathbf{I}_b||_1 + ||\\nabla_y \\mathbf{I}_a - \\nabla_y \\mathbf{I}_b||_1" + }, + { + "bbox": [ + 130, + 298, + 479, + 336 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 298, + 479, + 336 + ], + "type": "inline_equation", + "content": "\\nabla_x" + }, + { + "bbox": [ + 130, + 298, + 479, + 336 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 298, + 479, + 336 + ], + "type": "inline_equation", + "content": "\\nabla_y" + }, + { + "bbox": [ + 130, + 298, + 479, + 336 + ], + "type": "text", + "content": " represent the operations of computing horizontal and vertical gradients." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 353, + 329, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 353, + 329, + 366 + ], + "spans": [ + { + "bbox": [ + 131, + 353, + 329, + 366 + ], + "type": "text", + "content": "4 Proposed I2-2000FPS dataset" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 377, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 482, + 472 + ], + "type": "text", + "content": "While several high-frame-rate datasets have been open-sourced in recent times [36, 49, 60, 67, 68, 72], these datasets mainly feature videos tainted by severe motion blur, making them unsuitable in our problem setting. Moreover, features such as high motion speed and sufficient number of videos are also not always guaranteed. A visual representation of existing datasets' comparison is shown in Fig. 8(a). To address the gaps, we introduce the I2-2000FPS dataset, a high frame rate video collection meticulously designed to capture high-speed motion with precision." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 473, + 481, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 481, + 580 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 481, + 580 + ], + "type": "text", + "content": "The I2-2000FPS dataset has a temporal resolution of 2000 FPS and a spatial resolution of " + }, + { + "bbox": [ + 130, + 473, + 481, + 580 + ], + "type": "inline_equation", + "content": "512 \\times 1024" + }, + { + "bbox": [ + 130, + 473, + 481, + 580 + ], + "type": "text", + "content": " pixels, comprising 280 unique videos spanning 114 diverse scenes. The videos are captured using the Chronos 1.4 high-speed CMOS sensor-based camera from Kron Technologies. Notably, I2-2000FPS incorporates dark current calibration, leveraging the camera's capabilities to mitigate dark current effects. Throughout the data collection process, analog and digital gain were consistently maintained at 0 dB to avoid amplification of noise. To minimize noise, the videos are exclusively captured outdoor with ambient lighting conditions. More details on I2-2000FPS can be found in the supplementary." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 598, + 230, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 598, + 230, + 611 + ], + "spans": [ + { + "bbox": [ + 131, + 598, + 230, + 611 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 622, + 283, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 622, + 283, + 634 + ], + "spans": [ + { + "bbox": [ + 131, + 622, + 283, + 634 + ], + "type": "text", + "content": "5.1 Image Formation Model" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "For experiments involving synthetic data, we use a single-photon detector simulator based on an underlying image formation model discussed below. We build" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 101 + ], + "type": "text", + "content": "QUIVER" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 159, + 114, + 305, + 223 + ], + "blocks": [ + { + "bbox": [ + 159, + 114, + 305, + 223 + ], + "lines": [ + { + "bbox": [ + 159, + 114, + 305, + 223 + ], + "spans": [ + { + "bbox": [ + 159, + 114, + 305, + 223 + ], + "type": "image", + "image_path": "c60c27b47b93c3ff5cdc147408750ac71e30375e7aa8ba198fb3ba8499464a74.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 224, + 236, + 233 + ], + "lines": [ + { + "bbox": [ + 228, + 224, + 236, + 233 + ], + "spans": [ + { + "bbox": [ + 228, + 224, + 236, + 233 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 455, + 223 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 455, + 223 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 455, + 223 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 455, + 223 + ], + "type": "image", + "image_path": "ab268fd60e08d107421860a19227ab4fe06f2f630d1a821c9220c6a03e519d59.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 224, + 387, + 233 + ], + "lines": [ + { + "bbox": [ + 378, + 224, + 387, + 233 + ], + "spans": [ + { + "bbox": [ + 378, + 224, + 387, + 233 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 243, + 482, + 321 + ], + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 321 + ], + "type": "text", + "content": "Fig. 8: (a) Benchmarking high-speed video datasets. Horizontal axis represents the temporal resolution and the vertical axis indicates the maximum speed captured by the dataset, assuming a fixed camera-object distance. The circles in blue and orange indicate blur and blur-free videos, respectively. (b) Benchmarking Quanta Video Restoration Models on the I2-2000FPS dataset. Horizontal axis represents the computational complexity in terms of GFLOPs and the vertical axis indicates the PSNR acquired at 3.25PPP." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 344, + 479, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 479, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 479, + 367 + ], + "type": "text", + "content": "upon the prototype initially suggested in [46] adopted in prior works [6, 10, 11, 20, 26, 28, 56]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 368, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 482, + 464 + ], + "type": "text", + "content": "Given the quanta exposure [22], " + }, + { + "bbox": [ + 130, + 368, + 482, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{GT}}" + }, + { + "bbox": [ + 130, + 368, + 482, + 464 + ], + "type": "text", + "content": ", dependent on the photon flux and exposure time, the observed signal by the sensor can be represented as a Poisson-Gaussian random variable, where the Poisson represents the photon arrival process and the Gaussian models the read noise. The readout process involves various sources of distortions and an Analog-to-Digital Converter (ADC) to convert the real numbers into integers " + }, + { + "bbox": [ + 130, + 368, + 482, + 464 + ], + "type": "inline_equation", + "content": "\\{0,1,2,\\dots,L\\}" + }, + { + "bbox": [ + 130, + 368, + 482, + 464 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 368, + 482, + 464 + ], + "type": "inline_equation", + "content": "L = 2^{\\mathrm{Nbits}} - 1" + }, + { + "bbox": [ + 130, + 368, + 482, + 464 + ], + "type": "text", + "content": " depending on the bit-depth (Nbits) allocated to the sensor. The final sensor readout, " + }, + { + "bbox": [ + 130, + 368, + 482, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 130, + 368, + 482, + 464 + ], + "type": "text", + "content": ", can be represented using the following equation," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 167, + 472, + 481, + 501 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 472, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 167, + 472, + 481, + 501 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} \\sim \\mathrm {A D C} _ {[ 0, \\mathrm {L} ]} \\{\\mathrm {P o i s s o n (Q E \\times I ^ {G T} +} \\theta_ {\\mathrm {d a r k}}) + \\underbrace {\\mathrm {G a u s s} (0 , \\sigma_ {\\mathrm {r e a d}} ^ {2} \\mathbf {1})} _ {\\mathrm {r e a d n o i s e}} \\}. \\qquad (2)", + "image_path": "0841143251434040d68595092257bfe98e32846928d1205a64417d71e8e9c55b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 508, + 482, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 482, + 558 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 482, + 558 + ], + "type": "text", + "content": "Akin to previous works [6, 10, 20, 26, 28], we assume our sensor to be monochromatic as we utilize monochromatic real data in our experiments. For our sensor prototype, we utilize a Quantum Efficiency (QE) of 0.80. The dark current " + }, + { + "bbox": [ + 130, + 508, + 482, + 558 + ], + "type": "inline_equation", + "content": "(\\theta_{\\mathrm{dark}})" + }, + { + "bbox": [ + 130, + 508, + 482, + 558 + ], + "type": "text", + "content": " and read noise " + }, + { + "bbox": [ + 130, + 508, + 482, + 558 + ], + "type": "inline_equation", + "content": "(\\sigma_{\\mathrm{read}})" + }, + { + "bbox": [ + 130, + 508, + 482, + 558 + ], + "type": "text", + "content": " are set to " + }, + { + "bbox": [ + 130, + 508, + 482, + 558 + ], + "type": "inline_equation", + "content": "1.6\\mathrm{e}^{-}/\\mathrm{pix/sec}" + }, + { + "bbox": [ + 130, + 508, + 482, + 558 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 508, + 482, + 558 + ], + "type": "inline_equation", + "content": "0.2\\mathrm{e}^{-}/\\mathrm{pix}" + }, + { + "bbox": [ + 130, + 508, + 482, + 558 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 574, + 274, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 574, + 274, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 574, + 274, + 586 + ], + "type": "text", + "content": "5.2 Experimental Settings" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": "Training data. We curate a set of 249 videos from the I2-2000FPS collection and employ it as the training dataset for all the deep-learning models in our experiments. Each training sample is fetched on the fly from each clip. A training sample here is defined as a tuple containing the ground-truth/target frames and the 3-bit quanta frames simulated at 3.25 photons-per-pixel (PPP) (" + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": "1 lux assuming a " + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "inline_equation", + "content": "1.1\\mu \\mathrm{m}" + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": " pixel pitch and a 1/2000 second exposure time) using the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "P. Chennuri et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 463, + 419 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 463, + 419 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 463, + 419 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 463, + 419 + ], + "type": "image", + "image_path": "70fccee96f5fb920aef792c769f603f1d94bf5a4c8563c7aec956a8243fb4551.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 426, + 483, + 470 + ], + "lines": [ + { + "bbox": [ + 132, + 426, + 483, + 470 + ], + "spans": [ + { + "bbox": [ + 132, + 426, + 483, + 470 + ], + "type": "text", + "content": "Fig. 9: Visual comparisons of the reconstructed results on test videos from the proposed I2-2000FPS dataset. For fair comparison, all methods utilize 11 3-bit quanta frames simulated at 3.25 PPP per frame (" + }, + { + "bbox": [ + 132, + 426, + 483, + 470 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 132, + 426, + 483, + 470 + ], + "type": "text", + "content": "1 lux) to produce a restored frame. Best viewed in zoom." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 498, + 330, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 330, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 330, + 510 + ], + "type": "text", + "content": "image formation model described in Sec. 5.1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 510, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 592 + ], + "type": "text", + "content": "Testing data. To effectively analyze the performance of various methods, we carefully sample 31 videos from I2-2000FPS containing various motion types, shapes, and speeds. To test the generalizability, we also test the algorithms on X4K1000FPS [67] test dataset containing 15 videos from distinct scenes. Lastly, to measure the performance on real-world data, we collect binary frames using a SPAD sensor [17] and compare the reconstructed outputs. More details will be discussed in Sec. 5.3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 491, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 491, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 491, + 666 + ], + "type": "text", + "content": "Baselines. We compare the proposed method with eight existing dynamic scene reconstruction algorithms, namely Transform Denoise [6], QBP [47], Student-Teacher [10], RVRT [42], EMVD [2], FloRNN [1], MemDeblur [35], and Spk2ImgNet [85]. We also add an off-the-shelf denoiser BM3D [14] to QBP, denoted QBP (+BM3D), as a baseline for comparison. As we will discuss in Sec. 5.3, QUIVER beats all the baselines, both quantitatively and qualitatively." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "text", + "content": "QUIVER" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 119, + 481, + 224 + ], + "blocks": [ + { + "bbox": [ + 133, + 119, + 481, + 224 + ], + "lines": [ + { + "bbox": [ + 133, + 119, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 119, + 481, + 224 + ], + "type": "image", + "image_path": "7f356c9d60b1328faa2a85feb5ce8510853e17943abc715d35f3f19d150d5460.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 233, + 482, + 277 + ], + "lines": [ + { + "bbox": [ + 130, + 233, + 482, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 482, + 277 + ], + "type": "text", + "content": "Fig. 10: Performance on Real Quanta Data. We capture real 1-bit quanta data using a SPAD [17] and generate 3-bit frames through temporal averaging. All deep learning based models are trained using a photon level of 4.9 PPP per frame. Best viewed in zoom." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "text", + "content": "Training QUIVER. We utilize the function mentioned in Eq. (1) as the cost function for training QUIVER with regularization parameters " + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = 0.2" + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\lambda_{2} = 0.85" + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\lambda_{3} = 0.1" + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\lambda_{4} = 0.05" + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "text", + "content": ". The training data is extracted with patch size " + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "text", + "content": " and a batch size of 4. The weights are initialized with Lecun initialization [39]. The network is trained using the Adam optimizer [37] with an initial learning rate of " + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "inline_equation", + "content": "2.5 \\times 10^{-5}" + }, + { + "bbox": [ + 130, + 302, + 482, + 422 + ], + "type": "text", + "content": ". The low learning rate is driven by the inherent instability of recurrent networks, as it mitigates the risk of divergent behavior during training. We use a learning rate scheduler that reduces the learning rate by a factor of 2 when a plateau is reached. QUIVER takes approximately 1.5 days to train on a NVIDIA A100 Tensor Core GPU using Pytorch." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 441, + 198, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 198, + 451 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 198, + 451 + ], + "type": "text", + "content": "5.3 Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 462, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 617 + ], + "type": "text", + "content": "Synthetic Data Experiments. We begin with the synthetic experiments where we utilize 3-bit quanta frames, simulated using the parameters mentioned in Sec. 5.1 at 3.25, 9.75, 19.5, and 26 PPP to test the algorithms' performance. Tab. 2 and Tab. 3 demonstrate the PSNR and SSIM [76] of various methods extracted by predicting 6017 I2-2000FPS frames and 345 X4K1000FPS frames. To further substantiate the efficacy of QUIVER's design, we introduced a scaled-down variant, QUIVER-s (Refer Fig. 8(b) for complexity comparison). Quantitative results indicate that both QUIVER and QUIVER-s offer substantially better performance than all the baselines across a range of light levels. Fig. 9 depicts visual results of all the methods on the I2-2000FPS dataset. It is evident that existing methods fail to handle both motion and noise simultaneously, whereas, our proposed method, QUIVER, produces blur free high SNR outputs while preserving high-frequency details to a large extent." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Real Data Experiments. We verify the methods' performance on real data. The real data is collected as binary frames using a SPAD sensor [17] at 10000 FPS with a spatial resolution of " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "240 \\times 320" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ". As SPADs possess zero read noise, the binary frames are summed up to generate 3-bit frames. The average observed" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "P. Chennuri et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 187, + 157, + 427, + 253 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 2: Performance comparison on the proposed I2-2000FPS dataset across various light levels. Models are trained using the I2-2000FPS dataset. QUIVER performs significantly better than the existing methods." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 187, + 157, + 427, + 253 + ], + "lines": [ + { + "bbox": [ + 187, + 157, + 427, + 253 + ], + "spans": [ + { + "bbox": [ + 187, + 157, + 427, + 253 + ], + "type": "table", + "html": "
Photons-Per-Pixel (PPP)3.259.7519.526
MethodPSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑
Transform Denoise [6]21.31700.718423.15210.767122.77480.781222.30960.7811
QBP [47]15.94110.129319.18560.265420.40000.371320.79780.4114
QBP (+ BM3D [14])21.54760.703322.20010.689922.83510.769622.86170.7832
Student-Teacher [10]18.72000.400616.51950.247915.76360.213313.28890.0735
RVRT [42]19.41150.353921.67140.456822.08260.502121.75280.4968
EMVD [2]20.01940.587321.05590.604822.44030.559223.40530.5576
FloRNN [1]21.03410.678525.61320.709127.43220.739527.85200.7784
MemDeblur [35]19.48770.386814.49060.111216.17750.166716.00580.1712
Spk2ImgNet [85]20.39450.564219.66650.673322.93720.700814.97690.6861
QUIVER-s (Ours)24.70130.756526.86760.788327.29890.843227.86590.8408
QUIVER (Ours)26.21430.789726.80580.825027.75380.856327.93770.8446
", + "image_path": "cff3bba01ea1337b397e0304b92466ad35f557fb8d6c33137b0b498d0262c008.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 187, + 307, + 427, + 402 + ], + "blocks": [ + { + "bbox": [ + 130, + 263, + 482, + 297 + ], + "lines": [ + { + "bbox": [ + 130, + 263, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 263, + 482, + 297 + ], + "type": "text", + "content": "Table 3: Performance comparison on the X4K1000FPS dataset across various light levels. Models are trained using the I2-2000FPS dataset. QUIVER performs significantly better than the existing methods." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 187, + 307, + 427, + 402 + ], + "lines": [ + { + "bbox": [ + 187, + 307, + 427, + 402 + ], + "spans": [ + { + "bbox": [ + 187, + 307, + 427, + 402 + ], + "type": "table", + "html": "
Photons-Per-Pixel (PPP)3.259.7519.526
MethodPSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑
Transform Denoise [6]19.62550.632322.17030.704422.99380.722922.62300.7204
QBP [47]15.56340.230216.97580.323017.17980.395717.78070.4188
QBP (+ BM3D [14])17.96770.512318.53080.522618.24070.541418.79170.5586
Student-Teacher [10]18.82080.365216.15480.260814.93590.257113.97620.1186
RVRT [42]19.92030.364121.07810.447221.47800.492520.78990.4919
EMVD [2]20.51020.483621.81520.559522.94400.593622.45870.5860
FloRNN [1]20.82830.577823.58740.648424.32140.668325.24830.7170
MemDeblur [35]19.55340.364214.55950.220316.67490.311615.64960.2974
Spk2ImgNet [85]18.94240.473119.25320.572220.34420.571616.09310.6106
QUIVER-s (Ours)20.91970.595521.79900.652324.19240.731623.44110.7248
QUIVER (Ours)21.87300.652123.16540.705724.59560.764525.00860.7513
", + "image_path": "ac76287c574ff91fc7026f099a91257e06d273744203e01a778fb1009eeb1152.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 425, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 521 + ], + "type": "text", + "content": "light level after summation is 4.9 PPP. We generate results with networks trained at 4.9 PPP and demonstrate the visual results in Fig. 10. QUIVER, as opposed to existing state-of-the-art, effectively recovers high-frequency information while applying visually appealing smoothening effect to low-frequency regions of the scene. It is noteworthy that SPADs' image formation model is significantly different from that of the QIS' imaging model [58,59]. Therefore, the visual results also indicate that the proposed QUIVER can thoroughly generalize to various single-photon detectors." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 542, + 246, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 542, + 246, + 556 + ], + "spans": [ + { + "bbox": [ + 131, + 542, + 246, + 556 + ], + "type": "text", + "content": "6 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Effect of Pre-Denoising, Optical Flow and Multi-Scale. We conduct an ablation study to evaluate the effect of the pre-denoising, the learnable optical flow, and multi-scale reconstruction on performance. Upon removal of either module, we expand the features dimension of layers in the feature extraction stage and add RDB blocks to the RMDF module, thereby maintaining a similar model capacity. We train all possible combinations and display the quantitative results in Tab. 4. Results indicate that, in the absence of either one or more modules, the network's performance is substantially worse. Visual Results in Fig. 11" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "text", + "content": "QUIVER" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 115, + 481, + 183 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 481, + 183 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 481, + 183 + ], + "type": "image", + "image_path": "2c84702399a19cbd297424cb59aeb45ff3d7d64c55e5b4f6ec8446cfc03715c5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 191, + 482, + 214 + ], + "lines": [ + { + "bbox": [ + 132, + 191, + 482, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 191, + 482, + 214 + ], + "type": "text", + "content": "Fig.11: Ablation Study. Visual Comparisons depicting the effectiveness of Predenoiser, Optical Flow, and Multi-Scale Reconstruction Modules. Best viewed in zoom." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 479, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 479, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 479, + 258 + ], + "type": "text", + "content": "indicate that these modules serve critical roles as they significantly contribute to model's performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 259, + 479, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 479, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 479, + 319 + ], + "type": "text", + "content": "Does loading pre-trained optical flow module weights help? We initialize SPyNet with its pre-trained weights and finetune the same while training QUIVER, and display its quantitative results in Tab. 4. As the pre-trained SPyNet is not robust to photon shot noise and read noise, especially in low-light conditions, initializing the module with it will result in sub-optimal performance." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 228, + 392, + 386, + 472 + ], + "blocks": [ + { + "bbox": [ + 130, + 349, + 482, + 383 + ], + "lines": [ + { + "bbox": [ + 130, + 349, + 482, + 383 + ], + "spans": [ + { + "bbox": [ + 130, + 349, + 482, + 383 + ], + "type": "text", + "content": "Table 4: Ablation study. We conduct experiments to emphasize the role of Denoiser, Optical flow, and Multi-Scale reconstruction modules. We also show the effect of loading pretrained optical flow weights " + }, + { + "bbox": [ + 130, + 349, + 482, + 383 + ], + "type": "inline_equation", + "content": "(^{*})" + }, + { + "bbox": [ + 130, + 349, + 482, + 383 + ], + "type": "text", + "content": " on performance." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 228, + 392, + 386, + 472 + ], + "lines": [ + { + "bbox": [ + 228, + 392, + 386, + 472 + ], + "spans": [ + { + "bbox": [ + 228, + 392, + 386, + 472 + ], + "type": "table", + "html": "
Pre-Denoising Optical Flow Multi-ScaleI2-2000FPS
PSNR↑SSIM↑
×××23.67020.7756
✓✓*✓23.94790.7709
✓✓×24.38410.7808
××✓24.74450.7755
×✓✓24.99990.7753
✓×✓25.75210.7760
✓✓✓26.21430.7897
", + "image_path": "f31605512d70ecdecaee73ed5775e4c92c150a7b00d964a22448fa5ad22fdee6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 511, + 220, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 511, + 220, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 511, + 220, + 523 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "content": "In this paper, we presented a methodology to reconstruct blur-free grayscale images/videos captured using 1-bit or few-bit quanta data. While adopting the ideology of classical quanta restoration methods, we proposed an end-to-end deep learning framework, QUIVER, that utilizes pre-filtering, learnable optical flow module, and a novel multi-scale reconstruction approach to produce high-visual outputs. Experiments on synthetic and real data indicate QUIVER beats state-of-the-art and can generalize across single-photon sensors. We also introduce the world's first high-speed video dataset, I2-2000FPS, that captures fast moving scenes at 2000 fps, covering wide ranges of motion. We believe that I2-2000FPS will be a valuable asset for researchers in high-speed motion analysis and other computer vision tasks." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "P. Chennuri et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 163 + ], + "type": "text", + "content": "Acknowledgements. The work is supported, in part, by Samsung Research America (SRA), DARPA / SRC CogniSense JUMP 2.0 Center, NSF IIS-2133032, and NSF ECCS-2030570. The authors would like to thank German Mora Martin for his help in recording the SPAD data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 205, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 205, + 481, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 481, + 250 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 481, + 250 + ], + "type": "text", + "content": "1. Li et al, J.: Unidirectional Video Denoising by Mimicking Backward Recurrent Modules with Look-Ahead Forward Ones. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) ECCV. Lecture Notes in Computer Science (2022). https://doi.org/10.1007/978-3-031-19797-0_34" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 251, + 481, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 481, + 271 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 481, + 271 + ], + "type": "text", + "content": "2. Maggioni et al, M.: Efficient Multi-Stage Video Denoising With Recurrent Spatio-Temporal Fusion. In: CVPR (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 272, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 272, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 272, + 481, + 304 + ], + "type": "text", + "content": "3. Arias, P., Morel, J.M.: Video Denoising via Empirical Bayesian Estimation of Space-Time Patches. Journal of Mathematical Imaging and Vision 60(1), 70-93 (Jan 2018). https://doi.org/10.1007/s10851-017-0742-4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 304, + 481, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 481, + 348 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 481, + 348 + ], + "type": "text", + "content": "4. Chan, K.C.K., Zhou, S., Xu, X., Loy, C.C.: BasicVSR++: Improving Video SuperResolution With Enhanced Propagation and Alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5972-5981 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 349, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 349, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 349, + 481, + 380 + ], + "type": "text", + "content": "5. Chan, S.H.: What Does a One-Bit Quanta Image Sensor Offer? IEEE Transactions on Computational Imaging 8, 770-783 (2022). https://doi.org/10.1109/TCI.2022.3202012" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 381, + 481, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 381, + 481, + 414 + ], + "spans": [ + { + "bbox": [ + 138, + 381, + 481, + 414 + ], + "type": "text", + "content": "6. Chan, S.H., Elgendy, O.A., Wang, X.: Images from Bits: Non-Iterative Image Reconstruction for Quanta Image Sensors. Sensors 16(11), 1961 (Nov 2016). https://doi.org/10.3390/s16111961" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 415, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 415, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 415, + 481, + 456 + ], + "type": "text", + "content": "7. Chan, S.H., Lu, Y.M.: Efficient image reconstruction for gigapixel quantum image sensors. In: 2014 IEEE Global Conference on Signal and Information Processing (GlobalSIP). pp. 312-316 (Dec 2014). https://doi.org/10.1109/GlobalSIP.2014.7032129" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 457, + 481, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 457, + 481, + 499 + ], + "spans": [ + { + "bbox": [ + 138, + 457, + 481, + 499 + ], + "type": "text", + "content": "8. Chan, S.H., Wang, X., Elgendy, O.A.: Plug-and-Play ADMM for Image Restoration: Fixed-Point Convergence and Applications. IEEE Transactions on Computational Imaging 3(1), 84-98 (Mar 2017). https://doi.org/10.1109/TCI.2016.2629286" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 501, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 481, + 544 + ], + "type": "text", + "content": "9. Charbon, E., Fishburn, M., Walker, R., Henderson, R.K., Niclass, C.: SPAD-Based Sensors. In: Remondino, F., Stoppa, D. (eds.) TOF Range-Imaging Cameras, pp. 11-38. Springer Berlin Heidelberg, Berlin, Heidelberg (2013). https://doi.org/10.1007/978-3-642-27523-4_2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 545, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 545, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 134, + 545, + 481, + 599 + ], + "type": "text", + "content": "10. Chi, Y., Gnanasambandam, A., Koltun, V., Chan, S.H.: Dynamic Low-Light Imaging with Quanta Image Sensors. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision - ECCV 2020, vol. 12366, pp. 122-138. Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-58589-1_8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 600, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 481, + 643 + ], + "type": "text", + "content": "11. Chi, Y., Zhang, X., Chan, S.H.: HDR imaging with spatially varying signal-to-noise ratios. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5724-5734. IEEE (2023). https://doi.org/10.1109/CVPR52729.2023.00554, https://doi.org/10.1109/CVPR52729.2023.00554" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "text", + "content": "12. Choi, J.H., Elgendy, O.A., Chan, S.H.: Image Reconstruction for Quanta Image Sensors Using Deep Neural Networks. In: 2018 IEEE International Conference on" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 101 + ], + "type": "text", + "content": "QUIVER" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 147, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 480, + 138 + ], + "type": "text", + "content": "Acoustics, Speech and Signal Processing (ICASSP). pp. 6543-6547 (Apr 2018). https://doi.org/10.1109/ICASSP.2018.8461685" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 171 + ], + "type": "text", + "content": "13. Claus, M., van Gemert, J.: ViDeNN: Deep Blind Video Denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "type": "text", + "content": "14. Dabov, K., Foi, A., Katkovnik, V., Egiazarian, K.: Image denoising by sparse 3-D transform-domain collaborative filtering. IEEE Transactions on image processing 16(8), 2080-2095 (2007)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 205, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 205, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 134, + 205, + 480, + 225 + ], + "type": "text", + "content": "15. Dong, S., Huang, T., Tian, Y.: Spike camera and its coding methods. arXiv preprint arXiv:2104.04669 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 227, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 227, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 134, + 227, + 481, + 270 + ], + "type": "text", + "content": "16. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 271, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 271, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 134, + 271, + 481, + 314 + ], + "type": "text", + "content": "17. Dutton, N.A.W., Gyongy, I., Parmesan, L., Gnecchi, S., Calder, N., Rae, B.R., Pellegrini, S., Grant, L.A., Henderson, R.K.: A SPAD-Based QVGA Image Sensor for Single-Photon Counting and Quanta Imaging. IEEE Transactions on Electron Devices 63(1), 189–196 (Jan 2016). https://doi.org/10.1109/TED.2015.2464682" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 315, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 315, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 134, + 315, + 481, + 346 + ], + "type": "text", + "content": "18. Dutton, N.A., Gyongy, I., Parmesan, L., Henderson, R.K.: Single photon counting performance and noise analysis of CMOS SPAD-based image sensors. Sensors 16(7), 1122 (2016)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 134, + 347, + 481, + 380 + ], + "type": "text", + "content": "19. Elgendy, O.A., Chan, S.H.: Optimal Threshold Design for Quanta Image Sensor. IEEE Transactions on Computational Imaging 4(1), 99-111 (Mar 2018). https://doi.org/10.1109/TCI.2017.2781185" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 380, + 481, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 380, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 134, + 380, + 481, + 422 + ], + "type": "text", + "content": "20. Feng Yang, Lu, Y.M., Sbaiz, L., Vetterli, M.: Bits From Photons: Oversampled Image Acquisition Using Binary Poisson Statistics. IEEE Transactions on Image Processing 21(4), 1421-1436 (Apr 2012). https://doi.org/10.1109/TIP.2011.2179306" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 424, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 424, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 134, + 424, + 481, + 456 + ], + "type": "text", + "content": "21. Fossum, E., Ma, J., Masoodian, S., Anzagira, L., Zizza, R.: The Quanta Image Sensor: Every Photon Counts. Sensors 16(8), 1260 (Aug 2016). https://doi.org/10.3390/s16081260" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 457, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 457, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 134, + 457, + 481, + 479 + ], + "type": "text", + "content": "22. Fossum, E.R.: Modeling the performance of single-bit and multi-bit quanta image sensors. IEEE Journal of the Electron Devices Society 1(9), 166-174 (2013)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 134, + 479, + 481, + 511 + ], + "type": "text", + "content": "23. Gao, J., Shang, Z., Nie, K., Luo, T.: High dynamic range image reconstruction for multi-bit quanta image sensor. Optoelectronics Letters 18(9), 553-558 (Sep 2022). https://doi.org/10.1007/s11801-022-2014-9" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 512, + 481, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 512, + 481, + 554 + ], + "spans": [ + { + "bbox": [ + 134, + 512, + 481, + 554 + ], + "type": "text", + "content": "24. Gariepy, G., Krstajic, N., Henderson, R., Li, C., Thomson, R.R., Buller, G.S., Heshmat, B., Raskar, R., Leach, J., Faccio, D.: Single-photon sensitive light-infight imaging. Nature Communications 6(1), 6021 (Jan 2015). https://doi.org/10.1038/ncomms7021" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 555, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 555, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 134, + 555, + 481, + 599 + ], + "type": "text", + "content": "25. Gariepy, G., Leach, J., Warburton, R., Chan, S., Henderson, R., Faccio, D.: Picosecond time-resolved imaging using SPAD cameras. In: Lewis, K.L., Hollins, R.C. (eds.) SPIE Security + Defence. p. 99920N. Edinburgh, United Kingdom (Oct 2016). https://doi.org/10.1117/12.2241184" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 600, + 481, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 481, + 642 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 481, + 642 + ], + "type": "text", + "content": "26. Gnanasambandam, A., Chan, S.H.: HDR Imaging With Quanta Image Sensors: Theoretical Limits and Optimal Reconstruction. IEEE Transactions on Computational Imaging 6, 1571-1585 (2020). https://doi.org/10.1109/TCI.2020.3041093" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "text", + "content": "27. Gnanasambandam, A., Chan, S.H.: Image Classification in the Dark Using Quanta Image Sensors. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Proceed-" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 239, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 239, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 239, + 101 + ], + "type": "text", + "content": "P. Chennuri et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 149, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 149, + 116, + 481, + 149 + ], + "type": "text", + "content": "ings of the European Conference on Computer Vision (ECCV). pp. 484-501. Lecture Notes in Computer Science, Cham (2020). https://doi.org/10.1007/978-3-030-58598-3_29" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "text", + "content": "28. Gnanasambandam, A., Chan, S.H.: Exposure-Referred Signal-to-Noise Ratio for Digital Image Sensors. IEEE Transactions on Computational Imaging 8, 561-575 (2022). https://doi.org/10.1109/TCI.2022.3187657" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "text", + "content": "29. Godard, C., Matzen, K., Uytendaele, M.: Deep Burst Denoising. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 538-554 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "30. Gupta, A., Ingle, A., Velten, A., Gupta, M.: Photon-Flooded Single-Photon 3D Cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6770-6779 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "type": "text", + "content": "31. Gupta, S., Gupta, M.: Eulerian Single-Photon Vision. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10465–10476 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "32. Gutierrez-Barragan, F., Ingle, A., Seets, T., Gupta, M., Velten, A.: Compressive Single-Photon 3D Cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17854-17864 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 315 + ], + "type": "text", + "content": "33. Gyongy, I., Dutton, N., Henderson, R.: Single-Photon Tracking for High-Speed Vision. Sensors 18(2), 323 (Jan 2018). https://doi.org/10.3390/s18020323" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 357 + ], + "type": "text", + "content": "34. Hasinoff, S.W., Sharlet, D., Geiss, R., Adams, A., Barron, J.T., Kainz, F., Chen, J., Levoy, M.: Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM Transactions on Graphics 35(6), 1-12 (Nov 2016). https://doi.org/10.1145/2980179.2980254" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 369 + ], + "type": "text", + "content": "35. Ji, B., Yao, A.: Multi-Scale Memory-Based Video Deblurring. In: CVPR (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "text", + "content": "36. Kiani Galoogahi, H., Fagg, A., Huang, C., Ramanan, D., Lucey, S.: Need for Speed: A Benchmark for Higher Frame Rate Object Tracking. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1125-1134 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 402, + 481, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 481, + 434 + ], + "type": "text", + "content": "37. Kingma, D., Ba, J.: Adam: A Method for Stochastic Optimization. In: International Conference on Learning Representations (ICLR) (2015). https://doi.org/10.48550/arXiv.1412.6980" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 434, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 481, + 468 + ], + "type": "text", + "content": "38. Lebrun, M., Buades, A., Morel, J.M.: A Nonlocal Bayesian Image Denoising Algorithm. SIAM Journal on Imaging Sciences 6(3), 1665-1688 (Jan 2013). https://doi.org/10.1137/120874989" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "text", + "content": "39. Lecun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (Nov/1998). https://doi.org/10.1109/5.726791" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 501, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 544 + ], + "type": "text", + "content": "40. Li, C., Qu, X., Gnanasambandam, A., Elgendy, O.A., Ma, J., Chan, S.H.: Photon-limited object detection using non-local feature matching and knowledge distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3976-3987 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 544, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 481, + 577 + ], + "type": "text", + "content": "41. Liang, J., Cao, J., Fan, Y., Zhang, K., Ranjan, R., Li, Y., Timofte, R., Van Gool, L.: VRT: A Video Restoration Transformer (Jun 2022). https://doi.org/10.48550/arXiv.2201.12288" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 577, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 481, + 621 + ], + "type": "text", + "content": "42. Liang, J., Fan, Y., Xiang, X., Ranjan, R., Ilg, E., Green, S., Cao, J., Zhang, K., Timofte, R., Gool, L.V.: Recurrent Video Restoration Transformer with Guided Deformable Attention. Advances in Neural Information Processing Systems 35, 378-393 (Dec 2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "43. Liba, O., Murthy, K., Tsai, Y.T., Brooks, T., Xue, T., Karnad, N., He, Q., Barron, J.T., Sharlet, D., Geiss, R., Hasinoff, S.W., Pritch, Y., Levoy, M.: Handheld mobile photography in very low light. ACM Transactions on Graphics 38(6), 1-16 (Dec 2019). https://doi.org/10.1145/3355089.3356508" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 101 + ], + "type": "text", + "content": "QUIVER" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "44. Lindell, D.B., O'Toole, M., Wetzstein, G.: Single-photon 3D imaging with deep sensor fusion. ACM Transactions on Graphics 37(4), 113-1 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 138, + 480, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 138, + 480, + 170 + ], + "spans": [ + { + "bbox": [ + 133, + 138, + 480, + 170 + ], + "type": "text", + "content": "45. Ma, J., Chan, S., Fossum, E.R.: Review of Quanta Image Sensors for UltralowLight Imaging. IEEE Transactions on Electron Devices 69(6), 2824-2839 (Jun 2022). https://doi.org/10.1109/TED.2022.3166716" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 480, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 480, + 202 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 480, + 202 + ], + "type": "text", + "content": "46. Ma, J., Masoodian, S., Starkey, D.A., Fossum, E.R.: Photon-number-resolving megapixel image sensor at room temperature without avalanche gain. Optica 4(12), 1474-1481 (Dec 2017). https://doi.org/10.1364/OPTICA.4.001474" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 480, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 235 + ], + "type": "text", + "content": "47. Ma, S., Gupta, S., Ulku, A.C., Bruschini, C., Charbon, E., Gupta, M.: Quanta burst photography. ACM Transactions on Graphics 39(4) (Aug 2020). https://doi.org/10.1145/3386569.3392470" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 235, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 235, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 235, + 480, + 266 + ], + "type": "text", + "content": "48. Ma, S., Mos, P., Charbon, E., Gupta, M.: Burst Vision Using Single-Photon Cameras. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5375-5385 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 267, + 480, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 480, + 308 + ], + "type": "text", + "content": "49. Madhusudana, P.C., Yu, X., Birkbeck, N., Wang, Y., Adsumilli, B., Bovik, A.C.: Subjective and Objective Quality Assessment of High Frame Rate Videos. IEEE Access 9, 108069-108082 (2021). https://doi.org/10.1109/ACCESS.2021.3100462" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 310, + 480, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 310, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 480, + 353 + ], + "type": "text", + "content": "50. Maggioni, M., Boracchi, G., Foi, A., Egiazarian, K.: Video Denoising, Deblocking, and Enhancement Through Separable 4-D Nonlocal Spatiotemporal Transforms. IEEE Transactions on Image Processing 21(9), 3952-3966 (Sep 2012). https://doi.org/10.1109/TIP.2012.2199324" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 354, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 385 + ], + "type": "text", + "content": "51. Monakhova, K., Richter, S.R., Waller, L., Koltun, V.: Dancing Under the Stars: Video Denoising in Starlight. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16241-16251 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 386, + 480, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 427 + ], + "type": "text", + "content": "52. Niclass, C., Rochas, A., Besse, P.A., Charbon, E.: Design and characterization of a CMOS 3-D image sensor based on single photon avalanche diodes. IEEE Journal of Solid-State Circuits 40(9), 1847-1854 (Sep 2005). https://doi.org/10.1109/JSSC.2005.848173" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 428, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 428, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 428, + 480, + 472 + ], + "type": "text", + "content": "53. Pan, L., Scheerlinck, C., Yu, X., Hartley, R., Liu, M., Dai, Y.: Bringing a Blur Frame Alive at High Frame-Rate With an Event Camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6820-6829 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "type": "text", + "content": "54. Pearl, N., Treibitz, T., Korman, S.: NAN: Noise-Aware NeRFs for Burst-Denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12672-12681 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "text", + "content": "55. Plotz, T., Roth, S.: Benchmarking Denoising Algorithms With Real Photographs. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1586-1595 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 536, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 567 + ], + "type": "text", + "content": "56. Qu, X., Chi, Y., Chan, S.H.: Spatially varying exposure with 2-by-2 multiplexing: Optimality and universality. IEEE Transactions on Computational Imaging 10, 261-276 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "text", + "content": "57. Ranjan, A., Black, M.J.: Optical Flow Estimation Using a Spatial Pyramid Network. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4161-4170 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "text", + "content": "58. Rapp, J., Goyal, V.K.: A Few Photons Among Many: Unmixing Signal and Noise for Photon-Efficient Active Imaging. IEEE Transactions on Computational Imaging 3(3), 445-459 (Sep 2017). https://doi.org/10.1109/TCI.2017.2706028" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "59. Rapp, J., Ma, Y., Dawson, R.M.A., Goyal, V.K.: Dead Time Compensation for High-Flux Ranging. IEEE Transactions on Signal Processing 67(13), 3471-3486 (Jul 2019). https://doi.org/10.1109/TSP.2019.2914891" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 239, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 239, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 239, + 101 + ], + "type": "text", + "content": "P. Chennuri et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "60. Rebecq, H., Ranftl, R., Koltun, V., Scaramuzza, D.: High Speed and High Dynamic Range Video with an Event Camera. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(6), 1964-1980 (Jun 2021). https://doi.org/10.1109/TPAMI.2019.2963386" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 205 + ], + "type": "text", + "content": "61. Remez, T., Litany, O., Bronstein, A.: A picture is worth a billion bits: Real-time image reconstruction from dense binary threshold pixels. In: 2016 IEEE International Conference on Computational Photography (ICCP). pp. 1-9 (May 2016). https://doi.org/10.1109/ICCPHOT.2016.7492874" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 206, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 481, + 228 + ], + "type": "text", + "content": "62. Rochas, A.: Single photon avalanche diodes in CMOS technology. Tech. rep., Cite-seer (2003)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 228, + 481, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 481, + 273 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 481, + 273 + ], + "type": "text", + "content": "63. Ruget, A., Tyler, M., Mora Martin, G., Scholes, S., Zhu, F., Gyongy, I., Hearn, B., McLaughlin, S., Halimi, A., Leach, J.: Pixels2Pose: Super-resolution time-of-flight imaging for 3D pose estimation. Science Advances 8(48), eade0123 (Nov 2022). https://doi.org/10.1126/sciadv.ade0123" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 274, + 481, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 306 + ], + "type": "text", + "content": "64. Seets, T., Ingle, A., Laurenzis, M., Velten, A.: Motion Adaptive Deblurring With Single-Photon Cameras. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1945–1954 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 307, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 307, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 307, + 481, + 329 + ], + "type": "text", + "content": "65. Shazeer, N.: Glu variants improve transformer. arXiv preprint arXiv:2002.05202 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 330, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 481, + 373 + ], + "type": "text", + "content": "66. Shin, D., Xu, F., Venkatraman, D., Lussana, R., Villa, F., Zappa, F., Goyal, V.K., Wong, F.N.C., Shapiro, J.H.: Photon-efficient imaging with a single-photon camera. Nature Communications 7(1), 12046 (Jun 2016). https://doi.org/10.1038/ncomms12046" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "text", + "content": "67. Sim, H., Oh, J., Kim, M.: XVFI: eXtreme Video Frame Interpolation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14489-14498 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 408, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 481, + 441 + ], + "type": "text", + "content": "68. Su, S., Delbracio, M., Wang, J., Sapiro, G., Heidrich, W., Wang, O.: Deep Video Deblurring for Hand-Held Cameras. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1279-1288 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 442, + 481, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 481, + 485 + ], + "type": "text", + "content": "69. Tassano, M., Delon, J., Veit, T.: DVDNET: A Fast Network for Deep Video Denoising. In: 2019 IEEE International Conference on Image Processing (ICIP). pp. 1805-1809. IEEE, Taipei, Taiwan (Sep 2019). https://doi.org/10.1109/ICIP.2019.8803136" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 486, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 481, + 520 + ], + "type": "text", + "content": "70. Tassano, M., Delon, J., Veit, T.: FastDVDnet: Towards Real-Time Deep Video Denoising Without Flow Estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1354-1363 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "type": "text", + "content": "71. Vaksman, G., Elad, M., Milanfar, P.: Patch Craft: Video Denoising by Deep Modeling and Patch Matching. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2157-2166 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 554, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 481, + 587 + ], + "type": "text", + "content": "72. Voeikov, R., Falaleev, N., Baikulov, R.: TTNet: Real-Time Temporal and Spatial Video Analysis of Table Tennis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 884-885 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 587, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 481, + 632 + ], + "type": "text", + "content": "73. Vogels, T., Rousselle, F., Mcwilliams, B., Rothlin, G., Harvill, A., Adler, D., Meyer, M., Novak, J.: Denoising with kernel prediction and asymmetric loss functions. ACM Transactions on Graphics 37(4), 124:1-124:15 (Jul 2018). https://doi.org/10.1145/3197517.3201388" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "74. Wang, W., Chen, X., Yang, C., Li, X., Hu, X., Yue, T.: Enhancing Low Light Videos by Exploring High Sensitivity Camera Noise. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4111-4119 (2019)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 101 + ], + "type": "text", + "content": "QUIVER" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 588 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 481, + 160 + ], + "type": "text", + "content": "75. Wang, X.: Single-Photon Cameras Image Reconstruction Using Vision Transformer. In: 2023 IEEE 3rd International Conference on Computer Communication and Artificial Intelligence (CCAI). pp. 296-300 (May 2023). https://doi.org/10.1109/CCAI57533.2023.10201259" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "text", + "content": "76. Wang, Z., Bovik, A., Sheikh, H., Simoncelli, E.: Image Quality Assessment: From Error Visibility to Structural Similarity. IEEE Transactions on Image Processing 13(4), 600-612 (Apr 2004). https://doi.org/10.1109/TIP.2003.819861" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "77. Wong, H.T., Leung, C.S., Ho, D.: Theoretical analysis and image reconstruction for multi-bit quanta image sensors. Signal Processing 185, 108087 (Aug 2021). https://doi.org/10.1016/j.sigpro.2021.108087" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "type": "text", + "content": "78. Yang, F., Lu, Y.M., Sbaiz, L., Vetterli, M.: An optimal algorithm for reconstructing images from binary measurements. In: Bouman, C.A., Pollak, I., Wolfe, P.J. (eds.) IS&T/SPIE Electronic Imaging. p. 75330K. San Jose, California (Feb 2010). https://doi.org/10.1117/12.850887" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 314 + ], + "type": "text", + "content": "79. Yang, F., Sbaiz, L., Charbon, E., Susstrunk, S., Vetterli, M.: Image reconstruction in the gigavision camera. In: 2009 IEEE 12th International Conference on Computer Vision Workshops, ICCV Workshops. pp. 2212-2219 (Sep 2009). https://doi.org/10.1109/ICCVW.2009.5457554" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 315, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 357 + ], + "type": "text", + "content": "80. Zhang, D., Lian, Q., Su, Y., Ren, T.: Dual-Prior Integrated Image Reconstruction for Quanta Image Sensors Using Multi-Agent Consensus Equilibrium. IEEE/CAA Journal of Automatica Sinica 10(6), 1407-1420 (Jun 2023). https://doi.org/10.1109/JAS.2023.123390" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "text", + "content": "81. Zhang, D., Lian, Q., Yang, Y.: TwP: Two-stage projection framework with manifold constraint for image reconstruction. Digital Signal Processing 141, 104186 (Sep 2023). https://doi.org/10.1016/j.dsp.2023.104186" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 391, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 481, + 424 + ], + "type": "text", + "content": "82. Zhang, Y., Li, K., Li, K., Wang, L., Zhong, B., Fu, Y.: Image Super-Resolution Using Very Deep Residual Channel Attention Networks. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 286-301 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 425, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 481, + 456 + ], + "type": "text", + "content": "83. Zhang, Y., Tian, Y., Kong, Y., Zhong, B., Fu, Y.: Residual Dense Network for Image Restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(7), 2480-2495 (Jul 2021). https://doi.org/10.1109/TPAMI.2020.2968521" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 457, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 500 + ], + "type": "text", + "content": "84. Zhao, J., Xiong, R., Huang, T.: High-Speed Motion Scene Reconstruction for Spike Camera via Motion Aligned Filtering. In: 2020 IEEE International Symposium on Circuits and Systems (ISCAS). pp. 1-5. IEEE, Seville, Spain (Oct 2020). https://doi.org/10.1109/ISCAS45731.2020.9181055" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 501, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 544 + ], + "type": "text", + "content": "85. Zhao, J., Xiong, R., Liu, H., Zhang, J., Huang, T.: Spk2imgnet: Learning to reconstruct dynamic scene from continuous spike stream. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11996-12005 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 544, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 481, + 588 + ], + "type": "text", + "content": "86. Zhao, J., Xiong, R., Xie, J., Shi, B., Yu, Z., Gao, W., Huang, T.: Reconstructing Clear Image for High-Speed Motion Scene With a Retina-Inspired Spike Camera. IEEE Transactions on Computational Imaging 8, 12-27 (2022). https://doi.org/10.1109/TCI.2021.3136446" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 239, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 239, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 239, + 101 + ], + "type": "text", + "content": "P. Chennuri et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Quantization-Friendly Winograd Transformations for Convolutional Neural Networks/90e7434b-cd12-4e62-9976-08dc6035e0cd_content_list.json b/2024/Quantization-Friendly Winograd Transformations for Convolutional Neural Networks/90e7434b-cd12-4e62-9976-08dc6035e0cd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..93b42e144767ae1abc29d46258eca2d9b655b393 --- /dev/null +++ b/2024/Quantization-Friendly Winograd Transformations for Convolutional Neural Networks/90e7434b-cd12-4e62-9976-08dc6035e0cd_content_list.json @@ -0,0 +1,1616 @@ +[ + { + "type": "text", + "text": "Quantization-Friendly Winograd Transformations for Convolutional Neural Networks", + "text_level": 1, + "bbox": [ + 218, + 140, + 782, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vladimir Protsenko, Vladimir Kryzhanovskiy, and Alexander Filippov", + "bbox": [ + 228, + 212, + 772, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Huawei Noah's Ark Lab \nprotsenko.vladimir@huawei-partners.com, {kryzhanovskiy.vladimir, filippov alexander}@huawei.com", + "bbox": [ + 251, + 239, + 746, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Efficient deployment of modern deep convolutional neural networks on resource-constrained devices suffers from demanding computational requirements of convolution operations. Quantization and use of Winograd convolutions operating on sufficiently large-tile inputs are two powerful strategies to speed up convolution operations. However, their combination results in numerical instability, which manifests itself in a strong quality performance degradation. We present an efficient learning scenario that either completely overcomes or strongly reduces the accuracy degradation of full 8-bit quantized $F(4,3)$ and $F(6,3)$ Winograd convolutions. Within the global particle swarm optimization (PSO), we derived a set of quantization-friendly Winograd transformations. Following the state-of-the-art (SOTA) training pipeline [J. Fernandez-Marques et al., Proc. Mach. Learn. Syst. 2, 14-29 (2020)], we treat Winograd transformations as learnable parameters during network training. Evolving transformations starting from our PSO-derived ones rather than the standard Winograd transformations results in significant numerical error reduction and accuracy improvement. As a consequence, our approach significantly outperforms SOTA methods on various tasks.", + "bbox": [ + 261, + 321, + 743, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Winograd convolutions $\\cdot$ Quantization $\\cdot$ Efficient inference", + "bbox": [ + 261, + 584, + 740, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 625, + 375, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To date, deep convolutional neural networks (CNN) provide state-of-the-art performance in a wide variety of machine learning domains, including image processing, computer vision, video-based workflow analysis tasks, etc. The convolution operations, being the core block of CNNs, represent their most time-consuming and computationally intensive part [13]. Since modern CNNs typically include deep stacks of convolution layers both their training and inference are highly computationally demanding. As a consequence, despite recent advances in hardware technology, efficient inference of modern CNNs on resource-constrained mobile and embedded devices remains to be a challenging task.", + "bbox": [ + 212, + 657, + 787, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the recent explosion of support for low-precision arithmetic in modern hardware, quantization [18, 24, 36] has become one of the standard approaches for compression and acceleration of CNNs. The quantization process is defined", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "by mapping the network parameters to their low-precision representation. This allows to operate within low-bit arithmetic, having significantly lower computational and storage requirements than the full precision counterpart. Quantization to 8-bit integers generally produces only a moderate accuracy degradation and is widely acceptable in mobile devices.", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "An alternative way to achieve efficient inference is to speed up convolution operations at the algorithm level. Several efficient convolution techniques have been developed, including the fast Fourier transform (FFT) [31, 43] and Winograd minimal filtering algorithm [25, 47]. The latter is known to minimize computational costs for convolutions with kernels of sizes $3 \\times 3$ [4, 22], which are most often used in modern CNNs. The $F(m, r)$ Winograd algorithm transforms tiles of input feature map and $r \\times r$ filter into the Winograd domain, where $m^2$ outputs can be simultaneously computed with a single element-wise matrix multiplication. Contrary to the FFT, the original Winograd algorithm does not involve complex number multiplications, which are at least $3 \\times$ computationally costly compared to that of real numbers. Because of its cutting-edge performance, the Winograd algorithm is integrated into various deep-learning libraries.", + "bbox": [ + 212, + 227, + 789, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As the quantization and Winograd algorithms are not mutually exclusive but complementary, their mixing is attracting considerable attention from both academic and industrial fields. Theoretically, the larger the tile size $k_{\\mathrm{W}} = m + r - 1$ of input feature maps, the more speedup can be expected from using the $F(m,r)$ Winograd algorithm. However, in practice, large-tile Winograd algorithms turn out to be highly sensitive to numerical errors, which grow exponentially with tile size [7,8]. Quantization, lowering arithmetic precision, further amplifies numerical errors [19] and causes numerical instability of the Winograd algorithm, which results in severe degradation of network performance. This strongly inhibits the industrial use of large-tile ( $k_{\\mathrm{W}} \\gtrsim 4$ ) low-bit Winograd-based convolutions, including the perspective $F(4,3)$ and $F(6,3)$ . Despite all efforts to mitigate this problem (see Sec. 2), the performance of large-tile Winograd algorithms quantized to the widely supported 8-bit integer format remains far from satisfactory.", + "bbox": [ + 212, + 412, + 789, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present an efficient learning scenario that either completely overcomes or strongly reduces the accuracy degradation of full 8-bit quantized $F(4,3)$ and $F(6,3)$ Winograd convolutions. As a first step, we perform a global particle swarm optimization (PSO) search over the general family of Winograd algorithms, which are defined by specifying a set of transformation matrices $\\Theta \\equiv \\{A,B,G\\}$ , to select one that is more robust to quantization errors. This step is prior to network training and yields a model- and data-free set of quantization-friendly Winograd transformation matrices $\\Theta$ . We demonstrate that replacing the standard Winograd transformation matrices with our PSO-derived ones already results in significant accuracy improvements. To further tackle the performance degradation challenge, we integrated our transformations into the SOTA learning pipeline of Ref. [17], which treats Winograd transformation matrices as trainable parameters during network training. We show that allowing the evolution of Winograd transformations $\\Theta$ starting from our PSO-derived matrices, rather than from standard Winograd matrices as in the original pipeline of", + "bbox": [ + 212, + 613, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "V. Protsenko, V. Kryzhanovskiy et al.", + "bbox": [ + 271, + 114, + 524, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Ref. [17], has a direct impact on increasing networks performance. As a result, our learning strategy allows to outperform existing approaches (and the 8-bit SOTA results) by demonstrating superior accuracy in various tasks, including classification, super-resolution, and semantic segmentation.", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 231, + 397, + 248 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Toom-Cook minimal filtering algorithms [14, 42], generalized by Winograd [47], were first applied to CNNs by Lavin and Gray [25] to accelerate convolutions with $3 \\times 3$ filters. Barabasz et al. [7] presented a theoretical analysis of the accuracy of Winograd algorithms and revealed that the error bound for the Toom-Cook algorithm grows exponentially with tile size.", + "bbox": [ + 212, + 265, + 787, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To reduce poor numerical properties of Winograd convolutions, approaches to a more efficient selection of polynomial points have been proposed [3, 7, 26]. While most of them have empirical grounds, Ref. [26] leverages reinforcement learning to construct an optimal polynomial point selection policy. This reinforcement learning-based framework significantly improves the accuracies of quantized $F(m, 3)$ Winograd convolutions with $m \\leq 5$ and leads to only a small accuracy gap between Winograd convolutions and the direct convolutions. However, the performance benefits of this approach are only demonstrated for mixed 8/16-bit quantization, which is a significantly softer restriction compared to considering all Winograd convolution steps quantized below 8-bit. The accuracy improvement with mixed-bit precision or quantization only in the Winograd domain has been shown in [5, 27, 28, 40]. For quantized networks, Meng and Brothers [32] extended the Winograd algorithm to the complex number domain. But, the shift to complex transformations substantially reduces the multiplication saving and hence the speedup of the Winograd algorithm.", + "bbox": [ + 212, + 340, + 787, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fernandez-Marques et al. [16, 17] first introduced the idea of considering the Winograd transformation matrices as learnable parameters while training quantized networks. Including the Winograd transformations in the training loop introduces additional degrees of freedom that are aware of numerical errors and can be adjusted to reduce them. On $F(m \\leq 6,3)$ , it was demonstrated that training the Winograd transformation matrices, as opposed to keeping them static (non-trainable), causes a dramatic reduction of the accuracy gap between 8-bit quantized Winograd and direct convolutions. It is important to emphasize that, despite it is not clearly stated in the paper by Fernandez-Marques et al., all the parts of Winograd convolutions (including the Winograd transformation matrices) are maintained in 8-bit precision. This is clear from the official source code provided by the authors (https://github.com/jafermarq/WinogradAwareNets). We additionally verified that indeed the above official implementation strictly reproduces all the results of the corresponding study.", + "bbox": [ + 212, + 568, + 787, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite the significant accuracy improvements achieved by Ref. [17], there are still accuracy losses for the perspective $F(4,3)$ and $F(6,3)$ algorithms, which hinder their industrial potential. The study [6] continues the work of [17] and applies a base change technique for quantized Winograd convolutions. This tech-", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Quantization-Friendly Winograd Transformations for CNN", + "bbox": [ + 339, + 113, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "nique was tested for $F(4,3)$ ResNet-18 model on CIFAR-10 dataset. It is shown that transformations into Legendre polynomials base for 8-bit quantized trainable Winograd convolutions further reduce the errors of the original pipeline of Ref. [17]. However, the proposed approach can achieve the same accuracy as for direct convolutions only when the element-wise (Hadamard) product is extended to 9-bit. These improvements after modification of Winograd matrices yield the important hint that the selection of the appropriate starting values for these matrices in the training pipeline of Ref. [17] is of critical importance.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Besides the works [6,17], there are other studies on the use of full 8-bit quantized Winograd convolutions [11, 12, 30, 34, 35]. Each of these studies, however, compared to the approach of Ref. [17] comes with its own limitations on either performance or resulting acceleration. Ref. [35] performs quantization-aware training (QAT) to mitigate the numerical error of quantized Winograd inference. While preserving multiplication savings of the Winograd algorithm, this strategy loses in performance relative to Ref. [17]. Ref. [30] adopts a residue number system to eliminate accuracy degradation, but as a consequence, the theoretical multiplication savings are dramatically reduced by this approach. Ref. [12] (and the most accurate scheme of Ref. [34]) uses complex numbers based Winograd algorithms, which heavily reduces the speedup.", + "bbox": [ + 212, + 267, + 787, + 433 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In light of the above, the training scheme of Ref. [17] remains the SOTA in terms of tackling the problem of accuracy degradation while preserving the speedup of the standard Winograd algorithm. In the following, we refer to the pipeline of Ref. [17] as the SOTA approach. We emphasize that by this we mean only the case of full 8-bit quantization and real number arithmetic of the Winograd convolutions, since mixed-precision quantization has the potential to achieve better performance (e.g., see [26] for the SOTA results for mix 8/16-bit quantization).", + "bbox": [ + 212, + 433, + 787, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Our Method", + "text_level": 1, + "bbox": [ + 215, + 577, + 375, + 593 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 609, + 372, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Winograd Convolution. The 2D Winograd convolution $F(m, r)$ computes $m \\times m$ output by element-wise matrix multiplication of the Winograd domain-transformed filter $g$ of size $r \\times r$ and input feature map $d$ of size $k_W \\times k_W$ , where $k_W = m + r - 1$ . The whole procedure can be expressed in matrix form as follows:", + "bbox": [ + 212, + 635, + 787, + 696 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nY = A ^ {T} \\left[ \\left(G g G ^ {T}\\right) \\odot \\left(B ^ {T} d B\\right) \\right] A, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 709, + 785, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $A \\in \\mathbb{R}^{k_W \\times m}$ , $B \\in \\mathbb{R}^{k_W \\times k_W}$ , $G \\in \\mathbb{R}^{k_W \\times r}$ are constant real Winograd transformation matrices and $\\odot$ is the Hadamard (element-wise) multiplication. For brevity, we use the notation $\\Theta \\equiv \\{A, B, G\\}$ for the set of the Winograd matrices hereafter. The use of this algorithm reduces arithmetic complexity by simultaneously allowing $m^2$ outputs to be computed by performing $(m + r - 1)^2$ multiplications, instead of the $(rm)^2$ required for a naive convolution operation. This reduction of computationally costly multiplication operations together with the", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "V. Protsenko, V. Kryzhanovskiy et al.", + "bbox": [ + 271, + 114, + 524, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ability to compute multiple outputs results in a significant acceleration compared using of the direct convolution (e.g., theoretically $4 \\times$ speedup on $F(4,3)$ ). In this study, we consider $3 \\times 3$ filters, and for concise notation we denote $F(m,r = 3)$ as $F_{\\mathrm{m}}$ (e.g., $F(4,3) \\equiv F4$ in our notations).", + "bbox": [ + 212, + 146, + 784, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As can be seen from Eq. (1), the Winograd convolution is defined by specifying the transformation matrices $\\Theta$ . The particular form of $\\Theta$ depends on the size of the filter $r$ and the desired number of outputs. In the case of $3 \\times 3$ filters, which are widely used in modern CNNs, for $m \\leq 6$ there is a general agreement on the form of the transformation matrices $\\Theta$ . The corresponding matrices are given in [25] (see also Sec. 1 of the Supplemental Material) and we refer to this set of matrices as the standard Winograd transformations $\\Theta^{\\mathrm{std}}$ .", + "bbox": [ + 212, + 207, + 785, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantization. Following Ref. [17], we quantize all steps of the Winograd algorithm (including the transformations $\\Theta$ ) to 8-bit. Therefore, the full 8-bit quantized Winograd convolution can be written as", + "bbox": [ + 212, + 311, + 785, + 357 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nY = q _ {8} \\left(\\hat {Q} _ {8} ^ {A} \\left(\\hat {Q} _ {8} ^ {G ^ {T}} (g) \\odot \\hat {Q} _ {8} ^ {B} (d)\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 366, + 785, + 392 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{Q}_8^A (x) = q_8(A^T)q_8(x)q_8(A)$ is the transformation operator to/from the quantized Winograd domain and $q_{8}$ stands for 8-bit quantization. As in [17], we use the signed symmetric uniform scalar quantization scheme [36]. Note that for a proper comparison with the SOTA pipeline of Ref. [17], as in that study, both input and output are quantized. However, this quantization granularity is clearly redundant and may result in additional loss of accuracy. Therefore, we expect that within the conventional quantization schemes when either the input or the output is quantized, the proposed in this study approach provides further accuracy improvements.", + "bbox": [ + 212, + 401, + 787, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Quantization-Friendly Transformations", + "text_level": 1, + "bbox": [ + 214, + 559, + 581, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motivation. As discussed in Refs. [25,27,34], the magnitude of matrix elements of the Winograd transformations $\\Theta \\equiv \\{A,B,G\\}$ tends to increase with tile size $k_{W}$ , which considerably broadens numerical ranges and effectively reduces computation accuracy. In [7] Barabasz et al. analytically derived the floating-point error bound for the Winograd algorithm. It is shown that the error is proportional to the norms of the Winograd transform matrices. One can naturally expect the same relation between the quantization errors and matrix elements (and norms) of the Winograd transformations $\\Theta$ . Indeed, the error coming from each quantization step in Eq. (2) is amplified proportionally to magnitudes of the following matrix transformations (if not taking into account the rare self errors cancellation effects). Another point is that transformation matrices with high-magnitude matrix elements considerably extend the quantization interval [27,34], whereas most of it is very sparsely used [34]. This leads to strong rounding errors during quantization [34]. Here, to improve the numerical properties of quantized Winograd convolutions, we exploit the fact that transformations $\\Theta$ are not uniquely defined and construct a representation having as small matrix elements as possible. We demonstrate (see Sec. 4) that the resulting set of matrices $\\Theta^{\\mathrm{pso}}$ is", + "bbox": [ + 212, + 582, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Quantization-Friendly Winograd Transformations for CNN", + "bbox": [ + 338, + 113, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "quantization-friendly and yields a significant improvement in accuracy compared to using the standard Winograd transformations $\\Theta^{\\mathrm{std}}$ .", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "PSO Optimization. To reduce the magnitudes of the Winograd transformations we consider the following optimization problem", + "bbox": [ + 212, + 176, + 784, + 207 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\Theta^ {\\mathrm {p s o}} = \\underset {\\Theta} {\\arg \\min } f (\\Theta) = \\underset {\\Theta} {\\arg \\min } \\left(\\gamma_ {A} \\| A \\| _ {F} ^ {2} + \\gamma_ {B} \\| B \\| _ {F} ^ {2} + \\gamma_ {G} \\| G \\| _ {F} ^ {2}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 267, + 219, + 785, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\gamma_A, \\gamma_B, \\gamma_G > 0$ are weight factors, $\\| \\cdot \\|_F$ denotes the Frobenius norm of a matrix and the optimization is restricted to the set of transformations $\\Theta$ that fulfill Eq. (1) to exactly preserve the property to be convolution operation. Although mainly the matrices $A$ and $B$ contain high-value matrix elements (see Supplemental Material Sec. 1), one should also be aware of preventing the possible growth of the matrix elements of $G$ in a new representation. Given this $G$ is also included in Eq. (3). To reveal main the effect of minimizing $f(\\Theta)$ , we do not fine-tune the parameters $\\gamma_p$ , $(p \\in \\{A, B, G\\})$ , and simply set $\\gamma_p = 1$ . However, we note that a more careful choice of these parameters (e.g., with taking into account quantization granularity) represents one potential way to further improve the quantization properties of transformation matrices $\\Theta$ .", + "bbox": [ + 212, + 255, + 784, + 421 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To deal with Eq. (3), we start from the extended form of the Winograd transformation matrices $\\Theta$ of Ref. [44]", + "bbox": [ + 212, + 422, + 782, + 454 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nA ^ {T} = \\left(V _ {(m + n - 1) \\times m}\\right) ^ {T} S _ {A}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 478, + 784, + 497 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nB ^ {T} = S _ {B} \\left(V _ {(m + n - 1) \\times (m + n - 1)}\\right) ^ {- T}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 501, + 784, + 521 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nG = S _ {G} \\left(V _ {(m + n - 1) \\times n}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 379, + 525, + 784, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $V_{x\\times y}$ is a trimmed Vandermonde matrix given by", + "bbox": [ + 214, + 551, + 619, + 570 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nV = \\left[ \\begin{array}{c c c c} \\xi_ {0} ^ {0} \\eta_ {0} ^ {y - 1} & \\xi_ {0} ^ {1} \\eta_ {0} ^ {y - 2} & \\dots \\dots & \\xi_ {0} ^ {y - 1} \\eta_ {0} ^ {0} \\\\ \\xi_ {1} ^ {0} \\eta_ {1} ^ {y - 1} & \\xi_ {1} ^ {1} \\eta_ {1} ^ {y - 2} & \\dots \\dots & \\xi_ {1} ^ {y - 1} \\eta_ {1} ^ {0} \\\\ \\vdots & \\vdots & & \\vdots \\\\ \\xi_ {x - 1} ^ {0} \\eta_ {x - 1} ^ {y - 1} & \\xi_ {x - 1} ^ {1} \\eta_ {x - 1} ^ {y - 2} & \\dots \\dots & \\xi_ {x - 1} ^ {y - 1} \\eta_ {x - 1} ^ {0} \\end{array} \\right], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 287, + 582, + 784, + 661 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\{(\\xi_i,\\eta_i)\\}$ is the set of the homogenous coordinates (polynomial points) and $S_{p\\in \\{A,B,G\\}}$ are the diagonal scaling matrices fulfilling the $S_{A}S_{B}S_{G} = 1$ condition. Within this representation, 6 and 8 polynomial points need to be given to construct $F4$ and $F6$ , respectively. The polynomial points and scalings which generate the standard sets of the Winograd matrices $\\Theta^{\\mathrm{std}}$ for $F4$ and $F6$ algorithms are given in Supplemental Material Sec. 1.", + "bbox": [ + 212, + 672, + 784, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We consider polynomial points $\\{(\\xi_i,\\eta_i)\\}$ as optimization parameters and perform minimization of the objective function $f(\\Theta)$ using the particle swarm optimization (PSO) method [23, 45, 49]. PSO is a powerful swarm intelligence technique for solving global optimization problems. In this method, a set of $N$ candidate solutions, called swarm particles, are moved through a search-space.", + "bbox": [ + 212, + 765, + 784, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "V. Protsenko, V. Kryzhanovskiy et al.", + "bbox": [ + 271, + 114, + 524, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Each $i$ th, $i \\in (1 \\dots N)$ , particle is assigned with a position $\\mathbf{x}_i$ and a velocity $\\mathbf{v}_i$ , which evolve during unity time steps $t$ (or iteration numbers) according to the following update rules", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {i} (t + 1) \\leftarrow \\omega \\mathbf {v} _ {i} (t) + \\phi_ {1} r _ {1} [ \\mathbf {p} _ {i} - \\mathbf {x} _ {i} (t) ] + \\phi_ {2} r _ {2} [ \\mathbf {p} _ {g} - \\mathbf {x} _ {i} (t) ], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 292, + 205, + 785, + 220 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {i} (t + 1) \\leftarrow \\mathbf {x} _ {i} (t) + \\mathbf {v} _ {i} (t + 1), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 248, + 785, + 263 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{p}_i$ is the best solution that $i$ th particle has found so far, $\\mathbf{p}_g$ is the best solution revealed by the whole swarm, $r_{1(2)}$ are random numbers sampled at each time step from a uniform $[0,1]$ distribution. Finally, $\\omega$ , $\\phi_1$ and $\\phi_2$ are PSO hyperparameters, which we set to 0.9, 0.55 and 0.3. The PSO optimization is implemented within PySwarms toolkit [33].", + "bbox": [ + 212, + 270, + 782, + 345 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The evolution of the objective function $f(\\Theta)$ for $10^{4}$ iterations of the PSO algorithm with 2000 particles is shown in Fig. 1a. Compared to the value for the standard transformations $\\Theta^{\\mathrm{std}}$ (indicated by circled number 0), the objective function after the PSO optimization (circled number 6) is reduced by about 2.6 and 5.2 times for $F4$ and $F6$ , respectively. By varying the hyperparameters $(\\omega, \\phi_1, \\phi_2)$ , we verified that the PSO optimization leads either to the same value of the objective function (see inset of Fig. 1a) or to sub-optimal solutions (not shown). In the following, we denote the Winograd transformation matrices, which represent the PSO solution to the optimization problem given by Eq. (3), as $\\Theta^{\\mathrm{pso}}$ . The corresponding polynomial points $\\{(\\xi_i, \\eta_i)\\}$ from which they can be constructed according to Eqs. (4) to (7) and explicit expressions for $\\Theta^{\\mathrm{pso}}$ are given in the Supplemental Material Sec. 1. Note that in addition to the polynomial points, scaling matrices $S_{p \\in \\{A, B, G\\}}$ can also be considered as optimization parameters. However, the polynomial points already provide certain degrees of freedom for the scaling variation of the transformation matrices. In addition, the inclusion of $S_{p \\in \\{A, B, G\\}}$ in the PSO procedure leads to a significant expansion of the number of parameters to be optimized, which makes optimization $f(\\Theta)$ inefficient due to getting stuck in local minima. In view of this, we do not optimize $S_{p \\in \\{A, B, G\\}}$ and use their standard values (see Ref. [44], Secs. 1 and 2 of the Supplemental Material).", + "bbox": [ + 212, + 345, + 787, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experimental Results", + "text_level": 1, + "bbox": [ + 215, + 671, + 460, + 690 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our experiments we use two different training policies: (i) keep the Winograd transformations $\\Theta$ fixed, which we refer to as Static $\\Theta$ ; (ii) treat $\\Theta$ as learnable parameters during training, which we denote as Trainable $\\Theta$ . The scheme (i) is used to show that replacing the standard Winograd matrices $\\Theta^{\\mathrm{std}}$ by our PSO-derived ones $\\Theta^{\\mathrm{pso}}$ already leads to a significant improvement in accuracy, which demonstrates a quantization-friendly nature of the latter. Within the second scheme (ii), we use the SOTA training approach of Ref. [17] (see discussion in Sec. 2). We replace the standard matrices $\\Theta^{\\mathrm{std}}$ used in this pipeline by our $\\Theta^{\\mathrm{pso}}$ . We show that this modification allows us to outperform the SOTA results.", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Quantization-Friendly Winograd Transformations for CNN", + "bbox": [ + 338, + 113, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To ensure a fair comparison of different training scenarios, we extensively evaluate their performance on three different types of tasks: (i) classification, (ii) super-resolution, and (iii) semantic segmentation. Unless otherwise stated, all convolution layers are quantized to 8-bit, except for the first and last convolution layers. For experiments with Winograd convolutions, we replace the direct convolutions having stride 1 and kernel size $3 \\times 3$ with $F4$ or $F6$ algorithms. Details on datasets and training protocols are presented in Secs. 3 and 4 of the Supplemental Material.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Classification", + "text_level": 1, + "bbox": [ + 215, + 287, + 372, + 301 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We begin by considering the ResNet- $N$ family of models, $18 \\leq N \\leq 50$ . As in Ref. [17], in some cases we vary a channel multiplier (or \"width multiplier\" as used for the MobileNets) factor, which scales the number of output channels produced by each convolution layer. A channel multiplier equal to 1.0 corresponds to the full ResNet- $N$ model.", + "bbox": [ + 212, + 311, + 787, + 385 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To clearly demonstrate that minimization of the objective function $f(\\Theta)$ (see Eq. (3)) indeed leads to the reduction of quantization errors, we consider the accuracy of intermediate steps of the PSO-optimization in Fig. 1. Each $i$ th step (circled number) represents transformations $\\Theta^i$ corresponding to a particular value of the objective function (see panels a). The 0th step corresponds to", + "bbox": [ + 212, + 386, + 787, + 463 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/897cfdd43f65e4cfb2ea06f2e017e388adb51e0ccba094c0efe053a1345a2f22.jpg", + "image_caption": [ + "Fig. 1: a) The convergence history of the objective function $f(\\Theta)$ (see Eq. (3)) for $F4$ Winograd algorithm within the PSO optimization. Inset shows five independent runs of the PSO algorithm. b, c) Accuracy performance of the 8-bit $F4$ ResNet-18 model on the CIFAR-10 dataset at different steps of the PSO optimization (circled numbers in Fig. 1a and on the horizontal axis of Fig. 1b,c) for static (b) and trainable (c) $\\Theta$ . Red and black dots correspond to channel multipliers of 1.0 and 0.5, respectively. The dashed lines in Fig. 1c represent the accuracy of the FP direct convolution model. On F6 we obtained a qualitatively similar convergence and accuracy evolution (see Sec. 5 of the Supplemental Material)." + ], + "image_footnote": [], + "bbox": [ + 215, + 488, + 493, + 698 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f276b9be4ac85e81f5a6e0ecdd55da88d3951a16aa029f9ab8e00c0e09f81c9b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 488, + 640, + 693 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f009d19f3516ef1a59dacb60c015f8fc17f8c9ebffb86ad810e317f8568bb5e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 488, + 785, + 693 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "V. Protsenko, V. Kryzhanovskiy et al.", + "bbox": [ + 271, + 114, + 524, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f9159249a0a525cc74b8bc82629bb961a8f764b09b488494ef2d09bda1f70e32.jpg", + "image_caption": [ + "Fig. 2: Distribution of the accuracy of 8-bit $F4$ ResNet-18 model (chan. mult. is 1.0) on the CIFAR-10 dataset over 15 training runs for different training scenarios." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 491, + 242 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/91629ae99bb399bd7ad57449f9f76cf4f730bcc20d02fbdf18135c98cc76a17e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 143, + 784, + 242 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "the standard Winograd transformations $\\Theta^{\\mathrm{std}}$ , while the 6th step represents our transformations $\\Theta^{\\mathrm{pso}}$ . One can see that the PSO optimization of $f(\\Theta)$ results in a monotonic increase of the accuracy. In the case of the $F4$ algorithm with trainable $\\Theta$ (Fig. 1c), by starting training from $\\Theta = \\Theta^{\\mathrm{pso}}$ we not only outperform the SOTA $\\Theta^{\\mathrm{std}}$ -based results of Ref [17], but also achieve the accuracy of full-precision (FP) direct convolution models (denoted by dashed lines). Note that in some cases we observe the same effects as in Refs. [17,39], when quantization yields somewhat better accuracy compared to one of FP model. This is because quantization can provide an additional mechanism of regularization [2,39].", + "bbox": [ + 212, + 311, + 784, + 446 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The characteristic improvements from using $\\Theta^{\\mathrm{pseudo}}$ are not sensitive to both the precision of PSO convergence and random factors during network initialization. To show this we perform three training runs of the $F4$ ResNet-18 model with different random seeds for each of the five converged configurations shown in the inset of Fig. 1a. The corresponding accuracy distribution is compared in Fig. 2 to that obtained for 15 random seeds using the standard transformations $\\Theta^{\\mathrm{std}}$ . It can be seen that the obtained distributions are strongly/substantially separated in the case of static/trainable Winograd transformations.", + "bbox": [ + 212, + 449, + 784, + 568 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We find that replacing the standard transformations $\\Theta^{\\mathrm{std}}$ with $\\Theta^{\\mathrm{pseudo}}$ matrices, even without including them in a training loop (static $\\Theta$ scheme), already results in a considerable accuracy improvement (Tab. 1 and Fig. 3). This improvement is robust to variations of both network depth (Tab. 1) and channel scaling (dashed lines in Fig. 3). Learning Winograd transformations during training allows to", + "bbox": [ + 212, + 569, + 785, + 646 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/2a8334a65cb4bf98cd3f3e6e8d6aaa741a1739ec296259b3e50014b98efd42c0.jpg", + "table_caption": [ + "Table 1: The accuracy $(\\%)$ of 8-bit $F4-$ and $F6$ -based ResNet- $N$ models (chan. mult. is 1.0) on CIFAR-10 when $\\Theta$ is static. The best results are shown in bold. The values in parentheses give the improvement (to the first decimal place) over $\\Theta^{\\mathrm{std}}$ case, and the results for the $F4$ ResNet-18 are averaged over the experiments of Fig. 2." + ], + "table_footnote": [], + "table_body": "
ModelF4 Static ΘF6 Static Θ
Standard ΘOur Θpso (↑)Standard ΘOur Θpso (↑)
ResNet-1883.6591.54 (+7.9)83.2387.51 (+4.3)
ResNet-3484.2290.48 (+6.3)69.4887.09 (+17.6)
ResNet-5082.5791.52 (+8.9)70.1582.87 (+12.7)
", + "bbox": [ + 227, + 741, + 774, + 838 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Quantization-Friendly Winograd Transformations for CNN", + "bbox": [ + 338, + 113, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/5aee53d945fd2fa51b8f8a4fa116659dfebac915d0ddc6f25a62e1d80a153dda.jpg", + "image_caption": [ + "Fig. 3: The accuracy of the ResNet-18 on CIFAR-10 with various channel multipliers utilizing full 8-bit $F4$ (a) and $F6$ (b) Winograd convolutions. The SOTA data are taken from Fig. 4 of Ref. [17]." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 787, + 310 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "further adjust them to minimize accuracy errors (see solid lines in Fig. 3, Tabs. 2 and 3). In all experiments, the initialization of the Winograd transformations when considering them as trainable parameters via our PSO-derived matrices $\\Theta^{\\mathrm{pso}}$ is found to be highly beneficial. One can see that this training scenario not only considerably outperforms the SOTA $\\Theta^{\\mathrm{std}}$ -based training approach of Ref. [17], but also either completely closes (in the case of CIFAR-10) or strongly reduces (for CIFAR-100 and ImageNet) the error gap between the $F4$ Winograd convolution and the FP direct convolution.", + "bbox": [ + 212, + 401, + 787, + 521 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Finally, Tab. 2 shows a competitive comparison of the performance of using trainable transformations starting from $\\Theta = \\Theta^{\\mathrm{pseudo}}$ (column \"Our\") with the current SOTA methods. In most cases, our approach significantly outperforms the SOTA results in terms of accuracy. Note that for a fair comparison, in Tab. 2 we consider only the case of the scalar quantization and real-arithmetic full 8-bit quantized Winograd algorithm. However, our method is compatible with more", + "bbox": [ + 212, + 527, + 789, + 619 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/ec3a4fb3f8c65eeaab36f90df481dcda9a1136a1dfae6af68dced68a164d4b90.jpg", + "table_caption": [ + "Table 2: An accuracy (%) comparison of Winograd-aware quantization schemes." + ], + "table_footnote": [], + "table_body": "
Model & DatasetMethodFmRef.Ours
CIFAR-10ResNet-18Winograd-aware [17]F492.4693.26
Winograd-aware [17]F687.7489.57
ResNet-18 (mult. 0.5)Legendre [6]F491.892.53
ResNet-18 (mult. 0.25)Legendre [6]F489.790.29
ResNet-20WinoTrain [35]F489.792.31
WinoVidiVici [34]90.89
ResNet-18 & ImageNetWinoTrain [35]F461.2869.15
WinoVidiVici [34]69.14
", + "bbox": [ + 230, + 681, + 772, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "V. Protsenko, V. Kryzhanovskiy et al.", + "bbox": [ + 269, + 114, + 524, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/915304316f737dbc52f029f44074844493c804b04161cfa1c8f8736bb8acfd73.jpg", + "table_caption": [ + "Table 3: Performance of the full 8-bit quantized $F4-$ and $F6$ -based models (channel multiplier is 1.0) on the CIFAR-10 dataset when $\\Theta$ is trainable. Note that here SOTA refers to the results of the training pipeline of Ref. [17] obtained using the official code released by the authors. The abbreviation FP stands for the results of the full precision direct convolution models. Other notations are the same as in Tab. 1." + ], + "table_footnote": [], + "table_body": "
Task (val.)Model (Test set)Direct conv.8-bit Fm with trainable Θ
FP8-bitFmSOTAOur Θpso (↑)
Classification (Accuracy, %)ResNet-18 (CIFAR-10)93.1793.20F492.5093.26 (+0.8)
F687.7489.57 (+1.8)
ResNet-18 (CIFAR-100)74.6074.18F472.3874.10 (+1.7)
F666.1572.90 (+6.7)
ResNet-18 (ImageNet)69.7969.61F467.0569.15 (+2.1)
F653.9062.87 (+9.0)
ResNet-34 (CIFAR-10)93.4093.12F492.5493.29 (+0.7)
F680.0590.16 (+10.1)
ResNet-50 (CIFAR-10)93.3593.22F493.4693.68 (+0.2)
F686.3792.16 (+5.8)
×3 Super-resolution (PSNR, dB)ESPCN [41] (Set5)32.5332.38F430.8632.11 (+1.2)
F627.3329.82 (+2.5)
ESPCN [41] (Set14)29.2129.09F428.0528.87 (+0.8)
F625.6227.30 (+1.7)
ESPCN-6 (Set5)33.2133.12F430.4532.98 (+2.5)
F625.5330.16 (+4.6)
ESPCN-6 (Set14)29.5929.50F427.7529.44 (+1.7)
F624.0227.50 (+3.5)
RFDN [29] (Set5)34.1333.68F430.6733.36 (+2.7)
RFDN [29] (Set14)30.1729.81F427.9429.62 (+1.7)
Segm IoUENet [38] (CamVid)60.6160.17F456.1260.00 (+3.9)
F652.4356.39 (+4.0)
", + "bbox": [ + 228, + 224, + 776, + 676 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "sophisticated quantization schemes (e.g., the tapwise quantization of Ref. [5]) and further performance improvement is expected from their application.", + "bbox": [ + 212, + 703, + 782, + 734 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Super-Resolution", + "text_level": 1, + "bbox": [ + 215, + 755, + 405, + 770 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Here we examine the performance on super-resolution (SR) tasks considering the original and modified ESPCN [41] as well as the RFDN [29] models. We set the upscaling factor between low- and high-resolution images to $\\times 3$ . Since the SOTA approach of Ref. [17] (training $\\Theta$ starting from $\\Theta^{\\mathrm{std}}$ ) has only been", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Quantization-Friendly Winograd Transformations for CNN", + "bbox": [ + 339, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "implemented for classification scenarios, we extend it to SR networks based on its official implementation. All models are trained on the DIV2K dataset [1] and their performance is evaluated on the Set5 and Set14 benchmark data (more details on the SR architectures, training, and dataset preparation can be found in the Supplemental Material). As seen from the previous discussion, treating the Winograd transformations $\\Theta$ as trainable parameters is essential to overcome the quality degradation caused by quantization errors. For this reason, in the following, we focus only on the case of trainable $\\Theta$ . The mean PSNR values of the considered models on the benchmark datasets are summarized in Tab. 3.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The ESPCN model is the less-depth model under consideration. It consists of three convolutional layers followed by a pixel-shuffle operation. In our implementation, the last two convolution layers are replaced by either $F4$ or $F6$ Winograd convolutions and are quantized to 8-bit. On F4, training of the transformations from $\\Theta = \\Theta^{\\mathrm{pso}}$ gains 1.25dB and 0.82dB higher PSNR than the SOTA approach of Ref. [17] on Set5 and Set14, respectively. On $F6$ , the PSNR gap from the FP direct model is reduced by about 2 times within our training approach.", + "bbox": [ + 212, + 285, + 787, + 391 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "It can be expected that proper initialization of $\\Theta$ at the beginning of training is less critical for such lightweight networks. Indeed, due to the relatively small number of the transformation parameters to train, it is possible that even a local gradient-based optimization during network training may be enough to learn appropriate Winograd transformations. Thus, quality degradation when using standard matrices $\\Theta^{\\mathrm{std}}$ is expected to be more pronounced with increas", + "bbox": [ + 212, + 395, + 787, + 486 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5e25e33bcca89537b62ce751c595f19af0866db4c9bb77c9986e84c2a8ddf1bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 518, + 356, + 676 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7052acbee1c77cd69d0a461282920f1b300e089aef6e06bb2dda81c3cc1e7dd3.jpg", + "image_caption": [ + "(a) Direct / FP", + "Fig. 4: $\\times 3$ Super-resolution results of the RFDN model utilizing direct and $F4$ Winograd convolutions for \"Comic\" from Set14." + ], + "image_footnote": [], + "bbox": [ + 218, + 678, + 352, + 782 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2c0651d8c3112ab1aa47b80b5537ae263bc7bc3ab0ad8e44ca39748514de8a27.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 518, + 500, + 678 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5ad8d492dd2d0066425d9e50aa35e4acbe2a305bf33c67792ed711098bc4563c.jpg", + "image_caption": [ + "(b) Direct / 8-bit" + ], + "image_footnote": [], + "bbox": [ + 364, + 678, + 496, + 782 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/13ad549d98acad5ed397b442e05347e1c1a0f9ec5a015cfdbb08f78a7434c288.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 518, + 640, + 678 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/803af0da1a2da6bfae6a370334ad5d29ac5847e1c250e2f9cc8d165f9a807b5d.jpg", + "image_caption": [ + "(c) $F4$ SOTA / 8-bit" + ], + "image_footnote": [], + "bbox": [ + 506, + 678, + 637, + 782 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/585a94bc3b422de38b8f831e7b222de8c8f040bde7f99883a62e7a251325eaf1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 518, + 784, + 676 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ab0afd774726e721cf3842af3a7d9ee959b3ac3f3af6f140704a8024530c3cf3.jpg", + "image_caption": [ + "(d) $F4$ Ours / 8-bit" + ], + "image_footnote": [], + "bbox": [ + 650, + 676, + 779, + 782 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "V. Protsenko, V. Kryzhanovskiy et al.", + "bbox": [ + 271, + 114, + 524, + 128 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "ing network depth, as in the case of the classification tasks (see Sec. 4.1). To prove this, we consider the six-layered ESPCN-6 model, which is obtained by including three additional convolution layers in the ESPCN model. While the PSNR values obtained by the $\\Theta^{\\mathrm{pseudo}}$ -based training increase compared to the results of the original ESPCN model, the SOTA training approach, using $\\Theta^{\\mathrm{std}}$ , on the contrary, tends to PSNR degradation. On $F4$ , $\\Theta^{\\mathrm{pseudo}}$ -based training shows an improvement of more than 2.5 dB for Set5 and 1.6 dB for Set14 relative to the SOTA training approach. With this improvement, the quality drop when using the 8-bit quantized $F4$ algorithm is strongly reduced, and the resulting quality is close to that of FP direct convolutions.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Finally, we perform experiments with the residual feature distillation network (RFDN) [29]. It is based on the revised information distillation mechanism of the IMDN network [21] and the shallow residual blocks, which allows to achieve a better trade-off between model performance and complexity. Utilizing $\\Theta^{\\mathrm{pseudo}}$ leads to remarkably better performance compared to the SOTA training based on the standard matrices $\\Theta^{\\mathrm{std}}$ . However, we find that the RFDN model is quite sensitive to quantization noise. This results in a noticeable drop in the quality when the model is quantized to 8-bit, even in the case of direct convolutions. When the Winograd algorithm is used, this quality drop is somewhat worsened. One of the reasons for this is that the quantization scheme adapted from Ref. [17] includes redundant quantization steps (see discussion in Sec. 3.1). We also revealed a strong effect of replacing the $3\\times 3$ direct convolutions nearest to the contrast aware channel attention (CCA) layer [21, 29] with the $F4$ convolutions on the enhancement of the PSNR drop. The latter requires searching for a more careful strategy to replace direct convolutions with the Winograd ones, and, a possible solution is to use for some layers the $F2$ Winograd algorithm, which is not prone to strong quantization errors. Despite the above-mentioned drop of PSNR, our method produces super-resolution images that are visually competitive with ones", + "bbox": [ + 212, + 305, + 787, + 579 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a23a3e3e6baa137235ca21618f69e4336c2bbe7199d6404c0e8d24c7c4abf8f4.jpg", + "image_caption": [ + "(a) ESPCN F6 SOTA", + "Fig. 5: $\\times 3$ Super-resolution results for \"Comic\" from Set14 within 8-bit ESPCN (first two panels) and ESPCN-6 (last two panels) model using the $F6$ Winograd convolutions." + ], + "image_footnote": [], + "bbox": [ + 217, + 618, + 356, + 782 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c4473c1fd6faffe3c7aadd45833277d65805669eb4cb92424e7e2ac38a0484c9.jpg", + "image_caption": [ + "(b) ESPCN F6 Our" + ], + "image_footnote": [], + "bbox": [ + 359, + 618, + 500, + 782 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c44f1fb0b683444595771d76de3a7dc328b4bfb4f5a3d52e19153c0923bda86d.jpg", + "image_caption": [ + "(c) ESPCN-6 F6 SOTA" + ], + "image_footnote": [], + "bbox": [ + 504, + 618, + 642, + 782 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/96dc48826f75214e6f0689587632e384ac07f8d691bd12de2172c9ccbf18103a.jpg", + "image_caption": [ + "(d) ESPCN-6 F6 Ours" + ], + "image_footnote": [], + "bbox": [ + 645, + 619, + 784, + 782 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Quantization-Friendly Winograd Transformations for CNN", + "bbox": [ + 339, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "of the FP model (see Fig. 4) and are free of strong noise or gridding artifacts that typically occur in the case of standard Winograd transformations.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "For $F6$ , despite the considerable improvements achieved by using the transformations $\\Theta^{\\mathrm{pso}}$ (see Tab. 3 and Fig. 5), the quality drop is still pronounced. In particular, we find that the training of the RFDN model is unstable for $F6$ (for this reason, the corresponding results are not reported here). However, we emphasize that here we considered the extreme case when all direct convolutions of kernel size $3 \\times 3$ are replaced by the $F6$ convolutions. We believe it is possible to selectively replace error-tolerant direct convolutions with $F6$ ones based on our $\\Theta^{\\mathrm{pso}}$ with minimal quality losses. This can be achieved, for example, by adapting frameworks for automatic design of neural network architectures [10, 15, 17, 46].", + "bbox": [ + 212, + 176, + 787, + 313 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.3 Segmentation", + "text_level": 1, + "bbox": [ + 215, + 330, + 375, + 347 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Finally, we briefly discuss segmentation tasks. To this end, we consider the ENet model [38] which is the optimized semantic segmentation network designed to run on mobile devices. We replace all $3 \\times 3$ kernel direct convolutions with the Winograd convolutions, except ones for the last and first layers, and convolutions with a stride equal to 2. The latter is because the standard Winograd algorithm cannot be directly applied to convolutions with non-unit stride. However, we note that several adaptations of the Winograd algorithm to non-unit stride cases have recently been proposed [20,37,48]. Their quantization-friendly formulation within an optimization procedure similar to the one developed in the present study is of great interest for further research. The results of training the ENet model on the CamVid dataset [9] are shown in Tab. 3. As for SR tasks, we find that our $\\Theta^{\\mathrm{pso}}$ -based approach is considerably more accurate compared to the SOTA approach utilizing the standard Winograd transformations $\\Theta^{\\mathrm{std}}$ .", + "bbox": [ + 212, + 353, + 787, + 549 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 569, + 369, + 585 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We have introduced an efficient way of training 8-bit quantized networks utilizing $F4$ and $F6$ Winograd convolutions. To this end, using the global partial swarm optimization (PSO) method, we have derived a model- and data-free form of the Winograd transformation matrices $\\Theta \\equiv \\{A,B,G\\}$ , which has significantly better accuracy performance compared to the standard ones.", + "bbox": [ + 212, + 598, + 785, + 672 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "To further improve accuracy we have modified the previous SOTA training pipeline of Ref. [17], which allows the transformation matrices $\\Theta$ to evolve during training, by replacing the standard Winograd transformations used in this pipeline with our PSO-derived matrices. By treating all the steps of Winograd convolution as 8-bit quantized, we have considered classification, superresolution, and segmentation tasks to reveal the benefits of the proposed approach. In all of the experiments, we found that our modification results in a substantial improvement of accuracy and allows to either completely overcome or highly reduce quality drop. As a consequence, our approach significantly outperforms the SOTA method in terms of accuracy, while requiring the same number of operations during network training and retaining the same inference speedup.", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "V. Protsenko, V. Kryzhanovskiy et al.", + "bbox": [ + 271, + 114, + 524, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Agustsson, E., Timofte, R.: Ntire 2017 challenge on single image super-resolution: Dataset and study. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 126-135 (2017)", + "2. Alam, M.S., Misba, W.A., Atulasimha, J.: Quantized non-volatile nanomagnetic synapse based autoencoder for efficient unsupervised network anomaly detection. arXiv preprint arXiv:2309.06449 (2023)", + "3. Alam, S.A., Anderson, A., Barabasz, B., Gregg, D.: Winograd convolution for deep neural networks: Efficient point selection. ACM Transactions on Embedded Computing Systems 21(6), 1-28 (2022)", + "4. Anderson, A., Gregg, D.: Optimal dnn primitive selection with partitioned boolean quadratic programming. In: Proceedings of the 2018 International Symposium on Code Generation and Optimization. pp. 340-351 (2018)", + "5. Andri, R., Bussolino, B., Cipolletta, A., Cavigelli, L., Wang, Z.: Going further with winograd convolutions: Tap-wise quantization for efficient inference on 4x4 tiles. In: 2022 55th IEEE/ACM International Symposium on Microarchitecture (MICRO). pp. 582-598. IEEE (2022)", + "6. Barabasz, B.: Quantaized winograd/toom-cook convolution for dnns: Beyond canonical polynomials base. arXiv preprint arXiv:2004.11077 (2020)", + "7. Barabasz, B., Anderson, A., Soodhalter, K.M., Gregg, D.: Error analysis and improving the accuracy of winograd convolution for deep neural networks. ACM Transactions on Mathematical Software (TOMS) 46(4), 1-33 (2020)", + "8. Barabasz, B., Gregg, D.: Winograd convolution for dnns: Beyond linear polynomials. In: International Conference of the Italian Association for Artificial Intelligence. pp. 307-320. Springer (2019)", + "9. Brostow, G.J., Fauqueur, J., Cipolla, R.: Semantic object classes in video: A high-definition ground truth database. Pattern Recognition Letters 30(2), 88-97 (2009)", + "10. Cai, H., Zhu, L., Han, S.: Proxylessnas: Direct neural architecture search on target task and hardware. arXiv preprint arXiv:1812.00332 (2018)", + "1. Chen, T., Xu, W., Chen, W., Wang, P., Cheng, J.: Towards efficient and accurate winograd convolution via full quantization. Advances in Neural Information Processing Systems 36 (2024)", + "2. Chen, Y.T., Ou, Y.F., Huang, C.T.: A winograd-based highly-parallel convolution engine for 8-bit cnn acceleration. In: 2022 IEEE 4th International Conference on Artificial Intelligence Circuits and Systems (AICAS). pp. 395–398. IEEE (2022)", + "3. Chen, Y.H., Krishna, T., Emer, J.S., Sze, V.: Eyeriss: An energy-efficient reconfigurable accelerator for deep convolutional neural networks. IEEE journal of solid-state circuits 52(1), 127-138 (2016)", + "4. Cook, S.: On the minimum computation time for multiplication. Doctoral diss., Harvard U., Cambridge, Mass 1 (1966)", + "5. Elsken, T., Metzen, J.H., Hutter, F.: Neural architecture search: A survey. The Journal of Machine Learning Research 20(1), 1997-2017 (2019)", + "6. Fernández-Marqués, J.: High performance deep learning on resource constrained platforms. Ph.D. thesis, University of Oxford (2021)", + "7. Fernandez-Marques, J., Whatmough, P., Mundy, A., Mattina, M.: Searching for winograd-aware quantized networks. Proceedings of Machine Learning and Systems 2, 14-29 (2020)", + "8. Gholami, A., Kim, S., Dong, Z., Yao, Z., Mahoney, M.W., Keutzer, K.: A survey of quantization methods for efficient neural network inference. In: Low-Power Computer Vision, pp. 291-326. Chapman and Hall/CRC (2022)" + ], + "bbox": [ + 225, + 172, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Quantization-Friendly Winograd Transformations for CNN", + "bbox": [ + 339, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Holt, J.L., Hwang, J.N.: Finite precision error analysis of neural network hardware implementations. IEEE Transactions on Computers 42(3), 281-290 (1993)", + "20. Huang, C., Dong, X., Li, Z., Song, T., Liu, Z., Dong, L.: Efficient stride 2 winograd convolution method using unified transformation matrices on fpga. In: 2021 International Conference on Field-Programmable Technology (ICFPT). pp. 1-9. IEEE (2021)", + "21. Hui, Z., Gao, X., Yang, Y., Wang, X.: Lightweight image super-resolution with information multi-distillation network. In: Proceedings of the 27th acm international conference on multimedia. pp. 2024-2032 (2019)", + "22. Jorda, M., Valero-Lara, P., Pena, A.J.: Performance evaluation of cudnn convolution algorithms on nvidia volta gpus. IEEE Access 7, 70461-70473 (2019)", + "23. Kennedy, J., Eberhart, R.: Particle swarm optimization. In: Proceedings of ICNN'95-international conference on neural networks. vol. 4, pp. 1942-1948. IEEE (1995)", + "24. Krishnamoorthi, R.: Quantizing deep convolutional networks for efficient inference: A whitepaper. arXiv preprint arXiv:1806.08342 (2018)", + "25. Lavin, A., Gray, S.: Fast algorithms for convolutional neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4013-4021 (2016)", + "26. Li, C., Jiang, P., Zhou, H., Wang, X., Zhao, X.: Haw: Hardware-aware point selection for efficient winograd convolution. IEEE Signal Processing Letters 30, 269-273 (2023)", + "27. Li, G., Jia, Z., Feng, X., Wang, Y.: Lowino: Towards efficient low-precision winograd convolutions on modern cpus. In: Proceedings of the 50th International Conference on Parallel Processing. pp. 1-11 (2021)", + "28. Li, G., Liu, L., Wang, X., Ma, X., Feng, X.: Lance: efficient low-precision quantized winograd convolution for neural networks based on graphics processing units. In: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 3842-3846. IEEE (2020)", + "29. Liu, J., Tang, J., Wu, G.: Residual feature distillation network for lightweight image super-resolution. In: Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 41-55. Springer (2020)", + "30. Liu, Z.G., Mattina, M.: Efficient residue number system based winograd convolution. In: European Conference on Computer Vision. pp. 53-68. Springer (2020)", + "31. Mathieu, M., Henaff, M., LeCun, Y.: Fast training of convolutional networks through fts. arXiv preprint arXiv:1312.5851 (2013)", + "32. Meng, L., Brothers, J.: Efficient winograd convolution via integer arithmetic. arXiv preprint arXiv:1901.01965 (2019)", + "33. Miranda, L.J.: Pyswarms: a research toolkit for particle swarm optimization in python. Journal of Open Source Software 3(21), 433 (2018)", + "34. Mori, P., Frickenstein, L., Sampath, S.B., Thoma, M., Faspous, N., Vemparala, M.R., Frickenstein, A., Unger, C., Stechele, W., Mueller-Gritschneider, D., et al.: Wino vidi vici: Conquering numerical instability of 8-bit winograd convolution for accurate inference acceleration on edge. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 53-62 (2024)", + "35. Mori, P., Sampath, S.B., Frickenstein, L., Vemparala, M.R., Faspous, N., Frickenstein, A., Stechele, W., Passerone, C.: Winotrain: Winograd-aware training for accurate full 8-bit convolution acceleration. In: 2023 60th ACM/IEEE Design Automation Conference (DAC). pp. 1-6. IEEE (2023)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "V. Protsenko, V. Kryzhanovskiy et al.", + "bbox": [ + 271, + 114, + 524, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "36. Nagel, M., Fournarakis, M., Amjad, R.A., Bondarenko, Y., Van Baalen, M., Blankevoort, T.: A white paper on neural network quantization. arXiv preprint arXiv:2106.08295 (2021)", + "37. Pan, J., Chen, D.: Accelerate non-unit stride convolutions with winograd algorithms. In: Proceedings of the 26th Asia and South Pacific Design Automation Conference. pp. 358-364 (2021)", + "38. Paszke, A., Chaurasia, A., Kim, S., Culurciello, E.: Enet: A deep neural network architecture for real-time semantic segmentation. arXiv preprint arXiv:1606.02147 (2016)", + "39. Shafique, M.A., Munir, A., Kong, J.: Deep learning performance characterization on gpus for various quantization frameworks. AI 4(4), 926-948 (2023)", + "40. Shi, B., Zhang, J., He, Z., Wei, X., Li, S., Luo, G., Zheng, H., Xie, Y.: Efficient super-resolution system with block-wise hybridization and quantized winograd on fpga. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems (2023)", + "41. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1874-1883 (2016)", + "42. Toom, A.L.: The complexity of a scheme of functional elements simulating the multiplication of integers. In: Doklady Akademii Nauk. vol. 150, pp. 496-498. Russian Academy of Sciences (1963)", + "43. Vasilache, N., Johnson, J., Mathieu, M., Chintala, S., Piantino, S., LeCun, Y.: Fast convolutional nets with fbfft: Agpu performance evaluation. arXiv preprint arXiv:1412.7580 (2014)", + "44. Vincent, K., Stephano, K., Frumkin, M., Ginsburg, B., Demouth, J.: On improving the numerical stability of winograd convolutions (2017)", + "45. Wang, D., Tan, D., Liu, L.: Particle swarm optimization algorithm: an overview. Soft computing 22, 387-408 (2018)", + "46. Wang, K., Liu, Z., Lin, Y., Lin, J., Han, S.: Haq: Hardware-aware automated quantization with mixed precision. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8612-8620 (2019)", + "47. Winograd, S.: Arithmetic complexity of computations, vol. 33. Siam (1980)", + "48. Yepez, J., Ko, S.B.: Stride 2 1-d, 2-d, and 3-d winograd for convolutional neural networks. IEEE Transactions on Very Large Scale Integration (VLSI) Systems 28(4), 853-863 (2020)", + "49. Zhang, Y., Wang, S., Ji, G., et al.: A comprehensive survey on particle swarm optimization algorithm and its applications. Mathematical problems in engineering 2015 (2015)" + ], + "bbox": [ + 212, + 146, + 787, + 686 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Quantization-Friendly Winograd Transformations for CNN", + "bbox": [ + 339, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Quantization-Friendly Winograd Transformations for Convolutional Neural Networks/90e7434b-cd12-4e62-9976-08dc6035e0cd_model.json b/2024/Quantization-Friendly Winograd Transformations for Convolutional Neural Networks/90e7434b-cd12-4e62-9976-08dc6035e0cd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c8d16bda30555e5d4344d2293883b1b49c60a793 --- /dev/null +++ b/2024/Quantization-Friendly Winograd Transformations for Convolutional Neural Networks/90e7434b-cd12-4e62-9976-08dc6035e0cd_model.json @@ -0,0 +1,2214 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.22, + 0.141, + 0.783, + 0.186 + ], + "angle": 0, + "content": "Quantization-Friendly Winograd Transformations for Convolutional Neural Networks" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.213, + 0.773, + 0.229 + ], + "angle": 0, + "content": "Vladimir Protsenko, Vladimir Kryzhanovskiy, and Alexander Filippov" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.241, + 0.747, + 0.283 + ], + "angle": 0, + "content": "Huawei Noah's Ark Lab \nprotsenko.vladimir@huawei-partners.com, {kryzhanovskiy.vladimir, filippov alexander}@huawei.com" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.322, + 0.744, + 0.571 + ], + "angle": 0, + "content": "Abstract. Efficient deployment of modern deep convolutional neural networks on resource-constrained devices suffers from demanding computational requirements of convolution operations. Quantization and use of Winograd convolutions operating on sufficiently large-tile inputs are two powerful strategies to speed up convolution operations. However, their combination results in numerical instability, which manifests itself in a strong quality performance degradation. We present an efficient learning scenario that either completely overcomes or strongly reduces the accuracy degradation of full 8-bit quantized \\( F(4,3) \\) and \\( F(6,3) \\) Winograd convolutions. Within the global particle swarm optimization (PSO), we derived a set of quantization-friendly Winograd transformations. Following the state-of-the-art (SOTA) training pipeline [J. Fernandez-Marques et al., Proc. Mach. Learn. Syst. 2, 14-29 (2020)], we treat Winograd transformations as learnable parameters during network training. Evolving transformations starting from our PSO-derived ones rather than the standard Winograd transformations results in significant numerical error reduction and accuracy improvement. As a consequence, our approach significantly outperforms SOTA methods on various tasks." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.585, + 0.741, + 0.6 + ], + "angle": 0, + "content": "Keywords: Winograd convolutions \\(\\cdot\\) Quantization \\(\\cdot\\) Efficient inference" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.626, + 0.377, + 0.643 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.795 + ], + "angle": 0, + "content": "To date, deep convolutional neural networks (CNN) provide state-of-the-art performance in a wide variety of machine learning domains, including image processing, computer vision, video-based workflow analysis tasks, etc. The convolution operations, being the core block of CNNs, represent their most time-consuming and computationally intensive part [13]. Since modern CNNs typically include deep stacks of convolution layers both their training and inference are highly computationally demanding. As a consequence, despite recent advances in hardware technology, efficient inference of modern CNNs on resource-constrained mobile and embedded devices remains to be a challenging task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "With the recent explosion of support for low-precision arithmetic in modern hardware, quantization [18, 24, 36] has become one of the standard approaches for compression and acceleration of CNNs. The quantization process is defined" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.526, + 0.129 + ], + "angle": 0, + "content": "V. Protsenko, V. Kryzhanovskiy et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "by mapping the network parameters to their low-precision representation. This allows to operate within low-bit arithmetic, having significantly lower computational and storage requirements than the full precision counterpart. Quantization to 8-bit integers generally produces only a moderate accuracy degradation and is widely acceptable in mobile devices." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.228, + 0.79, + 0.409 + ], + "angle": 0, + "content": "An alternative way to achieve efficient inference is to speed up convolution operations at the algorithm level. Several efficient convolution techniques have been developed, including the fast Fourier transform (FFT) [31, 43] and Winograd minimal filtering algorithm [25, 47]. The latter is known to minimize computational costs for convolutions with kernels of sizes \\(3 \\times 3\\) [4, 22], which are most often used in modern CNNs. The \\(F(m, r)\\) Winograd algorithm transforms tiles of input feature map and \\(r \\times r\\) filter into the Winograd domain, where \\(m^2\\) outputs can be simultaneously computed with a single element-wise matrix multiplication. Contrary to the FFT, the original Winograd algorithm does not involve complex number multiplications, which are at least \\(3 \\times\\) computationally costly compared to that of real numbers. Because of its cutting-edge performance, the Winograd algorithm is integrated into various deep-learning libraries." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.79, + 0.61 + ], + "angle": 0, + "content": "As the quantization and Winograd algorithms are not mutually exclusive but complementary, their mixing is attracting considerable attention from both academic and industrial fields. Theoretically, the larger the tile size \\( k_{\\mathrm{W}} = m + r - 1 \\) of input feature maps, the more speedup can be expected from using the \\( F(m,r) \\) Winograd algorithm. However, in practice, large-tile Winograd algorithms turn out to be highly sensitive to numerical errors, which grow exponentially with tile size [7,8]. Quantization, lowering arithmetic precision, further amplifies numerical errors [19] and causes numerical instability of the Winograd algorithm, which results in severe degradation of network performance. This strongly inhibits the industrial use of large-tile (\\( k_{\\mathrm{W}} \\gtrsim 4 \\)) low-bit Winograd-based convolutions, including the perspective \\( F(4,3) \\) and \\( F(6,3) \\). Despite all efforts to mitigate this problem (see Sec. 2), the performance of large-tile Winograd algorithms quantized to the widely supported 8-bit integer format remains far from satisfactory." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.79, + 0.842 + ], + "angle": 0, + "content": "In this paper, we present an efficient learning scenario that either completely overcomes or strongly reduces the accuracy degradation of full 8-bit quantized \\( F(4,3) \\) and \\( F(6,3) \\) Winograd convolutions. As a first step, we perform a global particle swarm optimization (PSO) search over the general family of Winograd algorithms, which are defined by specifying a set of transformation matrices \\( \\Theta \\equiv \\{A,B,G\\} \\), to select one that is more robust to quantization errors. This step is prior to network training and yields a model- and data-free set of quantization-friendly Winograd transformation matrices \\( \\Theta \\). We demonstrate that replacing the standard Winograd transformation matrices with our PSO-derived ones already results in significant accuracy improvements. To further tackle the performance degradation challenge, we integrated our transformations into the SOTA learning pipeline of Ref. [17], which treats Winograd transformation matrices as trainable parameters during network training. We show that allowing the evolution of Winograd transformations \\( \\Theta \\) starting from our PSO-derived matrices, rather than from standard Winograd matrices as in the original pipeline of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.34, + 0.114, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Quantization-Friendly Winograd Transformations for CNN" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "Ref. [17], has a direct impact on increasing networks performance. As a result, our learning strategy allows to outperform existing approaches (and the 8-bit SOTA results) by demonstrating superior accuracy in various tasks, including classification, super-resolution, and semantic segmentation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.232, + 0.398, + 0.249 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.266, + 0.788, + 0.341 + ], + "angle": 0, + "content": "The Toom-Cook minimal filtering algorithms [14, 42], generalized by Winograd [47], were first applied to CNNs by Lavin and Gray [25] to accelerate convolutions with \\(3 \\times 3\\) filters. Barabasz et al. [7] presented a theoretical analysis of the accuracy of Winograd algorithms and revealed that the error bound for the Toom-Cook algorithm grows exponentially with tile size." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.789, + 0.568 + ], + "angle": 0, + "content": "To reduce poor numerical properties of Winograd convolutions, approaches to a more efficient selection of polynomial points have been proposed [3, 7, 26]. While most of them have empirical grounds, Ref. [26] leverages reinforcement learning to construct an optimal polynomial point selection policy. This reinforcement learning-based framework significantly improves the accuracies of quantized \\( F(m, 3) \\) Winograd convolutions with \\( m \\leq 5 \\) and leads to only a small accuracy gap between Winograd convolutions and the direct convolutions. However, the performance benefits of this approach are only demonstrated for mixed 8/16-bit quantization, which is a significantly softer restriction compared to considering all Winograd convolution steps quantized below 8-bit. The accuracy improvement with mixed-bit precision or quantization only in the Winograd domain has been shown in [5, 27, 28, 40]. For quantized networks, Meng and Brothers [32] extended the Winograd algorithm to the complex number domain. But, the shift to complex transformations substantially reduces the multiplication saving and hence the speedup of the Winograd algorithm." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.789, + 0.78 + ], + "angle": 0, + "content": "Fernandez-Marques et al. [16, 17] first introduced the idea of considering the Winograd transformation matrices as learnable parameters while training quantized networks. Including the Winograd transformations in the training loop introduces additional degrees of freedom that are aware of numerical errors and can be adjusted to reduce them. On \\( F(m \\leq 6,3) \\), it was demonstrated that training the Winograd transformation matrices, as opposed to keeping them static (non-trainable), causes a dramatic reduction of the accuracy gap between 8-bit quantized Winograd and direct convolutions. It is important to emphasize that, despite it is not clearly stated in the paper by Fernandez-Marques et al., all the parts of Winograd convolutions (including the Winograd transformation matrices) are maintained in 8-bit precision. This is clear from the official source code provided by the authors (https://github.com/jafermarq/WinogradAwareNets). We additionally verified that indeed the above official implementation strictly reproduces all the results of the corresponding study." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Despite the significant accuracy improvements achieved by Ref. [17], there are still accuracy losses for the perspective \\( F(4,3) \\) and \\( F(6,3) \\) algorithms, which hinder their industrial potential. The study [6] continues the work of [17] and applies a base change technique for quantized Winograd convolutions. This tech-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.526, + 0.129 + ], + "angle": 0, + "content": "V. Protsenko, V. Kryzhanovskiy et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "nique was tested for \\( F(4,3) \\) ResNet-18 model on CIFAR-10 dataset. It is shown that transformations into Legendre polynomials base for 8-bit quantized trainable Winograd convolutions further reduce the errors of the original pipeline of Ref. [17]. However, the proposed approach can achieve the same accuracy as for direct convolutions only when the element-wise (Hadamard) product is extended to 9-bit. These improvements after modification of Winograd matrices yield the important hint that the selection of the appropriate starting values for these matrices in the training pipeline of Ref. [17] is of critical importance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.788, + 0.434 + ], + "angle": 0, + "content": "Besides the works [6,17], there are other studies on the use of full 8-bit quantized Winograd convolutions [11, 12, 30, 34, 35]. Each of these studies, however, compared to the approach of Ref. [17] comes with its own limitations on either performance or resulting acceleration. Ref. [35] performs quantization-aware training (QAT) to mitigate the numerical error of quantized Winograd inference. While preserving multiplication savings of the Winograd algorithm, this strategy loses in performance relative to Ref. [17]. Ref. [30] adopts a residue number system to eliminate accuracy degradation, but as a consequence, the theoretical multiplication savings are dramatically reduced by this approach. Ref. [12] (and the most accurate scheme of Ref. [34]) uses complex numbers based Winograd algorithms, which heavily reduces the speedup." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.555 + ], + "angle": 0, + "content": "In light of the above, the training scheme of Ref. [17] remains the SOTA in terms of tackling the problem of accuracy degradation while preserving the speedup of the standard Winograd algorithm. In the following, we refer to the pipeline of Ref. [17] as the SOTA approach. We emphasize that by this we mean only the case of full 8-bit quantization and real number arithmetic of the Winograd convolutions, since mixed-precision quantization has the potential to achieve better performance (e.g., see [26] for the SOTA results for mix 8/16-bit quantization)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.578, + 0.376, + 0.594 + ], + "angle": 0, + "content": "3 Our Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.61, + 0.373, + 0.625 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.636, + 0.788, + 0.698 + ], + "angle": 0, + "content": "Winograd Convolution. The 2D Winograd convolution \\( F(m, r) \\) computes \\( m \\times m \\) output by element-wise matrix multiplication of the Winograd domain-transformed filter \\( g \\) of size \\( r \\times r \\) and input feature map \\( d \\) of size \\( k_W \\times k_W \\), where \\( k_W = m + r - 1 \\). The whole procedure can be expressed in matrix form as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.385, + 0.71, + 0.786, + 0.731 + ], + "angle": 0, + "content": "\\[\nY = A ^ {T} \\left[ \\left(G g G ^ {T}\\right) \\odot \\left(B ^ {T} d B\\right) \\right] A, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where \\(A \\in \\mathbb{R}^{k_W \\times m}\\), \\(B \\in \\mathbb{R}^{k_W \\times k_W}\\), \\(G \\in \\mathbb{R}^{k_W \\times r}\\) are constant real Winograd transformation matrices and \\(\\odot\\) is the Hadamard (element-wise) multiplication. For brevity, we use the notation \\(\\Theta \\equiv \\{A, B, G\\}\\) for the set of the Winograd matrices hereafter. The use of this algorithm reduces arithmetic complexity by simultaneously allowing \\(m^2\\) outputs to be computed by performing \\((m + r - 1)^2\\) multiplications, instead of the \\((rm)^2\\) required for a naive convolution operation. This reduction of computationally costly multiplication operations together with the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.339, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Quantization-Friendly Winograd Transformations for CNN" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.207 + ], + "angle": 0, + "content": "ability to compute multiple outputs results in a significant acceleration compared using of the direct convolution (e.g., theoretically \\(4 \\times\\) speedup on \\(F(4,3)\\)). In this study, we consider \\(3 \\times 3\\) filters, and for concise notation we denote \\(F(m,r = 3)\\) as \\(F_{\\mathrm{m}}\\) (e.g., \\(F(4,3) \\equiv F4\\) in our notations)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.787, + 0.312 + ], + "angle": 0, + "content": "As can be seen from Eq. (1), the Winograd convolution is defined by specifying the transformation matrices \\(\\Theta\\). The particular form of \\(\\Theta\\) depends on the size of the filter \\(r\\) and the desired number of outputs. In the case of \\(3 \\times 3\\) filters, which are widely used in modern CNNs, for \\(m \\leq 6\\) there is a general agreement on the form of the transformation matrices \\(\\Theta\\). The corresponding matrices are given in [25] (see also Sec. 1 of the Supplemental Material) and we refer to this set of matrices as the standard Winograd transformations \\(\\Theta^{\\mathrm{std}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.787, + 0.358 + ], + "angle": 0, + "content": "Quantization. Following Ref. [17], we quantize all steps of the Winograd algorithm (including the transformations \\(\\Theta\\)) to 8-bit. Therefore, the full 8-bit quantized Winograd convolution can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.367, + 0.786, + 0.393 + ], + "angle": 0, + "content": "\\[\nY = q _ {8} \\left(\\hat {Q} _ {8} ^ {A} \\left(\\hat {Q} _ {8} ^ {G ^ {T}} (g) \\odot \\hat {Q} _ {8} ^ {B} (d)\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.789, + 0.54 + ], + "angle": 0, + "content": "where \\(\\hat{Q}_8^A (x) = q_8(A^T)q_8(x)q_8(A)\\) is the transformation operator to/from the quantized Winograd domain and \\(q_{8}\\) stands for 8-bit quantization. As in [17], we use the signed symmetric uniform scalar quantization scheme [36]. Note that for a proper comparison with the SOTA pipeline of Ref. [17], as in that study, both input and output are quantized. However, this quantization granularity is clearly redundant and may result in additional loss of accuracy. Therefore, we expect that within the conventional quantization schemes when either the input or the output is quantized, the proposed in this study approach provides further accuracy improvements." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.56, + 0.583, + 0.575 + ], + "angle": 0, + "content": "3.2 Quantization-Friendly Transformations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Motivation. As discussed in Refs. [25,27,34], the magnitude of matrix elements of the Winograd transformations \\(\\Theta \\equiv \\{A,B,G\\}\\) tends to increase with tile size \\(k_{W}\\), which considerably broadens numerical ranges and effectively reduces computation accuracy. In [7] Barabasz et al. analytically derived the floating-point error bound for the Winograd algorithm. It is shown that the error is proportional to the norms of the Winograd transform matrices. One can naturally expect the same relation between the quantization errors and matrix elements (and norms) of the Winograd transformations \\(\\Theta\\). Indeed, the error coming from each quantization step in Eq. (2) is amplified proportionally to magnitudes of the following matrix transformations (if not taking into account the rare self errors cancellation effects). Another point is that transformation matrices with high-magnitude matrix elements considerably extend the quantization interval [27,34], whereas most of it is very sparsely used [34]. This leads to strong rounding errors during quantization [34]. Here, to improve the numerical properties of quantized Winograd convolutions, we exploit the fact that transformations \\(\\Theta\\) are not uniquely defined and construct a representation having as small matrix elements as possible. We demonstrate (see Sec. 4) that the resulting set of matrices \\(\\Theta^{\\mathrm{pso}}\\) is" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.526, + 0.129 + ], + "angle": 0, + "content": "V. Protsenko, V. Kryzhanovskiy et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "quantization-friendly and yields a significant improvement in accuracy compared to using the standard Winograd transformations \\(\\Theta^{\\mathrm{std}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.208 + ], + "angle": 0, + "content": "PSO Optimization. To reduce the magnitudes of the Winograd transformations we consider the following optimization problem" + }, + { + "type": "equation", + "bbox": [ + 0.269, + 0.22, + 0.786, + 0.245 + ], + "angle": 0, + "content": "\\[\n\\Theta^ {\\mathrm {p s o}} = \\underset {\\Theta} {\\arg \\min } f (\\Theta) = \\underset {\\Theta} {\\arg \\min } \\left(\\gamma_ {A} \\| A \\| _ {F} ^ {2} + \\gamma_ {B} \\| B \\| _ {F} ^ {2} + \\gamma_ {G} \\| G \\| _ {F} ^ {2}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.256, + 0.785, + 0.422 + ], + "angle": 0, + "content": "where \\(\\gamma_A, \\gamma_B, \\gamma_G > 0\\) are weight factors, \\(\\| \\cdot \\|_F\\) denotes the Frobenius norm of a matrix and the optimization is restricted to the set of transformations \\(\\Theta\\) that fulfill Eq. (1) to exactly preserve the property to be convolution operation. Although mainly the matrices \\(A\\) and \\(B\\) contain high-value matrix elements (see Supplemental Material Sec. 1), one should also be aware of preventing the possible growth of the matrix elements of \\(G\\) in a new representation. Given this \\(G\\) is also included in Eq. (3). To reveal main the effect of minimizing \\(f(\\Theta)\\), we do not fine-tune the parameters \\(\\gamma_p\\), \\((p \\in \\{A, B, G\\})\\), and simply set \\(\\gamma_p = 1\\). However, we note that a more careful choice of these parameters (e.g., with taking into account quantization granularity) represents one potential way to further improve the quantization properties of transformation matrices \\(\\Theta\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.423, + 0.784, + 0.455 + ], + "angle": 0, + "content": "To deal with Eq. (3), we start from the extended form of the Winograd transformation matrices \\(\\Theta\\) of Ref. [44]" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.479, + 0.785, + 0.498 + ], + "angle": 0, + "content": "\\[\nA ^ {T} = \\left(V _ {(m + n - 1) \\times m}\\right) ^ {T} S _ {A}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.502, + 0.785, + 0.522 + ], + "angle": 0, + "content": "\\[\nB ^ {T} = S _ {B} \\left(V _ {(m + n - 1) \\times (m + n - 1)}\\right) ^ {- T}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.38, + 0.526, + 0.785, + 0.542 + ], + "angle": 0, + "content": "\\[\nG = S _ {G} \\left(V _ {(m + n - 1) \\times n}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.553, + 0.62, + 0.571 + ], + "angle": 0, + "content": "where \\(V_{x\\times y}\\) is a trimmed Vandermonde matrix given by" + }, + { + "type": "equation", + "bbox": [ + 0.288, + 0.583, + 0.785, + 0.662 + ], + "angle": 0, + "content": "\\[\nV = \\left[ \\begin{array}{c c c c} \\xi_ {0} ^ {0} \\eta_ {0} ^ {y - 1} & \\xi_ {0} ^ {1} \\eta_ {0} ^ {y - 2} & \\dots \\dots & \\xi_ {0} ^ {y - 1} \\eta_ {0} ^ {0} \\\\ \\xi_ {1} ^ {0} \\eta_ {1} ^ {y - 1} & \\xi_ {1} ^ {1} \\eta_ {1} ^ {y - 2} & \\dots \\dots & \\xi_ {1} ^ {y - 1} \\eta_ {1} ^ {0} \\\\ \\vdots & \\vdots & & \\vdots \\\\ \\xi_ {x - 1} ^ {0} \\eta_ {x - 1} ^ {y - 1} & \\xi_ {x - 1} ^ {1} \\eta_ {x - 1} ^ {y - 2} & \\dots \\dots & \\xi_ {x - 1} ^ {y - 1} \\eta_ {x - 1} ^ {0} \\end{array} \\right], \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.785, + 0.764 + ], + "angle": 0, + "content": "where \\(\\{(\\xi_i,\\eta_i)\\}\\) is the set of the homogenous coordinates (polynomial points) and \\(S_{p\\in \\{A,B,G\\}}\\) are the diagonal scaling matrices fulfilling the \\(S_{A}S_{B}S_{G} = 1\\) condition. Within this representation, 6 and 8 polynomial points need to be given to construct \\(F4\\) and \\(F6\\), respectively. The polynomial points and scalings which generate the standard sets of the Winograd matrices \\(\\Theta^{\\mathrm{std}}\\) for \\(F4\\) and \\(F6\\) algorithms are given in Supplemental Material Sec. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.785, + 0.84 + ], + "angle": 0, + "content": "We consider polynomial points \\(\\{(\\xi_i,\\eta_i)\\}\\) as optimization parameters and perform minimization of the objective function \\(f(\\Theta)\\) using the particle swarm optimization (PSO) method [23, 45, 49]. PSO is a powerful swarm intelligence technique for solving global optimization problems. In this method, a set of \\(N\\) candidate solutions, called swarm particles, are moved through a search-space." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.339, + 0.114, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Quantization-Friendly Winograd Transformations for CNN" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "Each \\(i\\)th, \\(i \\in (1 \\dots N)\\), particle is assigned with a position \\(\\mathbf{x}_i\\) and a velocity \\(\\mathbf{v}_i\\), which evolve during unity time steps \\(t\\) (or iteration numbers) according to the following update rules" + }, + { + "type": "equation", + "bbox": [ + 0.293, + 0.206, + 0.786, + 0.222 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {i} (t + 1) \\leftarrow \\omega \\mathbf {v} _ {i} (t) + \\phi_ {1} r _ {1} [ \\mathbf {p} _ {i} - \\mathbf {x} _ {i} (t) ] + \\phi_ {2} r _ {2} [ \\mathbf {p} _ {g} - \\mathbf {x} _ {i} (t) ], \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.249, + 0.786, + 0.265 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {i} (t + 1) \\leftarrow \\mathbf {x} _ {i} (t) + \\mathbf {v} _ {i} (t + 1), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.271, + 0.784, + 0.346 + ], + "angle": 0, + "content": "where \\(\\mathbf{p}_i\\) is the best solution that \\(i\\)th particle has found so far, \\(\\mathbf{p}_g\\) is the best solution revealed by the whole swarm, \\(r_{1(2)}\\) are random numbers sampled at each time step from a uniform \\([0,1]\\) distribution. Finally, \\(\\omega\\), \\(\\phi_1\\) and \\(\\phi_2\\) are PSO hyperparameters, which we set to 0.9, 0.55 and 0.3. The PSO optimization is implemented within PySwarms toolkit [33]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.347, + 0.789, + 0.65 + ], + "angle": 0, + "content": "The evolution of the objective function \\( f(\\Theta) \\) for \\( 10^{4} \\) iterations of the PSO algorithm with 2000 particles is shown in Fig. 1a. Compared to the value for the standard transformations \\( \\Theta^{\\mathrm{std}} \\) (indicated by circled number 0), the objective function after the PSO optimization (circled number 6) is reduced by about 2.6 and 5.2 times for \\( F4 \\) and \\( F6 \\), respectively. By varying the hyperparameters \\( (\\omega, \\phi_1, \\phi_2) \\), we verified that the PSO optimization leads either to the same value of the objective function (see inset of Fig. 1a) or to sub-optimal solutions (not shown). In the following, we denote the Winograd transformation matrices, which represent the PSO solution to the optimization problem given by Eq. (3), as \\( \\Theta^{\\mathrm{pso}} \\). The corresponding polynomial points \\( \\{(\\xi_i, \\eta_i)\\} \\) from which they can be constructed according to Eqs. (4) to (7) and explicit expressions for \\( \\Theta^{\\mathrm{pso}} \\) are given in the Supplemental Material Sec. 1. Note that in addition to the polynomial points, scaling matrices \\( S_{p \\in \\{A, B, G\\}} \\) can also be considered as optimization parameters. However, the polynomial points already provide certain degrees of freedom for the scaling variation of the transformation matrices. In addition, the inclusion of \\( S_{p \\in \\{A, B, G\\}} \\) in the PSO procedure leads to a significant expansion of the number of parameters to be optimized, which makes optimization \\( f(\\Theta) \\) inefficient due to getting stuck in local minima. In view of this, we do not optimize \\( S_{p \\in \\{A, B, G\\}} \\) and use their standard values (see Ref. [44], Secs. 1 and 2 of the Supplemental Material)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.672, + 0.462, + 0.691 + ], + "angle": 0, + "content": "4 Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In our experiments we use two different training policies: (i) keep the Winograd transformations \\(\\Theta\\) fixed, which we refer to as Static \\(\\Theta\\); (ii) treat \\(\\Theta\\) as learnable parameters during training, which we denote as Trainable \\(\\Theta\\). The scheme (i) is used to show that replacing the standard Winograd matrices \\(\\Theta^{\\mathrm{std}}\\) by our PSO-derived ones \\(\\Theta^{\\mathrm{pso}}\\) already leads to a significant improvement in accuracy, which demonstrates a quantization-friendly nature of the latter. Within the second scheme (ii), we use the SOTA training approach of Ref. [17] (see discussion in Sec. 2). We replace the standard matrices \\(\\Theta^{\\mathrm{std}}\\) used in this pipeline by our \\(\\Theta^{\\mathrm{pso}}\\). We show that this modification allows us to outperform the SOTA results." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.526, + 0.129 + ], + "angle": 0, + "content": "V. Protsenko, V. Kryzhanovskiy et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "To ensure a fair comparison of different training scenarios, we extensively evaluate their performance on three different types of tasks: (i) classification, (ii) super-resolution, and (iii) semantic segmentation. Unless otherwise stated, all convolution layers are quantized to 8-bit, except for the first and last convolution layers. For experiments with Winograd convolutions, we replace the direct convolutions having stride 1 and kernel size \\(3 \\times 3\\) with \\(F4\\) or \\(F6\\) algorithms. Details on datasets and training protocols are presented in Secs. 3 and 4 of the Supplemental Material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.289, + 0.373, + 0.303 + ], + "angle": 0, + "content": "4.1 Classification" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.788, + 0.386 + ], + "angle": 0, + "content": "We begin by considering the ResNet- \\(N\\) family of models, \\(18 \\leq N \\leq 50\\). As in Ref. [17], in some cases we vary a channel multiplier (or \"width multiplier\" as used for the MobileNets) factor, which scales the number of output channels produced by each convolution layer. A channel multiplier equal to 1.0 corresponds to the full ResNet- \\(N\\) model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.788, + 0.464 + ], + "angle": 0, + "content": "To clearly demonstrate that minimization of the objective function \\( f(\\Theta) \\) (see Eq. (3)) indeed leads to the reduction of quantization errors, we consider the accuracy of intermediate steps of the PSO-optimization in Fig. 1. Each \\( i \\)th step (circled number) represents transformations \\( \\Theta^i \\) corresponding to a particular value of the objective function (see panels a). The 0th step corresponds to" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.489, + 0.495, + 0.699 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.489, + 0.642, + 0.694 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.489, + 0.787, + 0.694 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.711, + 0.788, + 0.837 + ], + "angle": 0, + "content": "Fig. 1: a) The convergence history of the objective function \\( f(\\Theta) \\) (see Eq. (3)) for \\( F4 \\) Winograd algorithm within the PSO optimization. Inset shows five independent runs of the PSO algorithm. b, c) Accuracy performance of the 8-bit \\( F4 \\) ResNet-18 model on the CIFAR-10 dataset at different steps of the PSO optimization (circled numbers in Fig. 1a and on the horizontal axis of Fig. 1b,c) for static (b) and trainable (c) \\( \\Theta \\). Red and black dots correspond to channel multipliers of 1.0 and 0.5, respectively. The dashed lines in Fig. 1c represent the accuracy of the FP direct convolution model. On F6 we obtained a qualitatively similar convergence and accuracy evolution (see Sec. 5 of the Supplemental Material)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.339, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Quantization-Friendly Winograd Transformations for CNN" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.493, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.145, + 0.785, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.253, + 0.785, + 0.281 + ], + "angle": 0, + "content": "Fig. 2: Distribution of the accuracy of 8-bit \\( F4 \\) ResNet-18 model (chan. mult. is 1.0) on the CIFAR-10 dataset over 15 training runs for different training scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.785, + 0.448 + ], + "angle": 0, + "content": "the standard Winograd transformations \\(\\Theta^{\\mathrm{std}}\\), while the 6th step represents our transformations \\(\\Theta^{\\mathrm{pso}}\\). One can see that the PSO optimization of \\(f(\\Theta)\\) results in a monotonic increase of the accuracy. In the case of the \\(F4\\) algorithm with trainable \\(\\Theta\\) (Fig. 1c), by starting training from \\(\\Theta = \\Theta^{\\mathrm{pso}}\\) we not only outperform the SOTA \\(\\Theta^{\\mathrm{std}}\\)-based results of Ref [17], but also achieve the accuracy of full-precision (FP) direct convolution models (denoted by dashed lines). Note that in some cases we observe the same effects as in Refs. [17,39], when quantization yields somewhat better accuracy compared to one of FP model. This is because quantization can provide an additional mechanism of regularization [2,39]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.785, + 0.569 + ], + "angle": 0, + "content": "The characteristic improvements from using \\(\\Theta^{\\mathrm{pseudo}}\\) are not sensitive to both the precision of PSO convergence and random factors during network initialization. To show this we perform three training runs of the \\(F4\\) ResNet-18 model with different random seeds for each of the five converged configurations shown in the inset of Fig. 1a. The corresponding accuracy distribution is compared in Fig. 2 to that obtained for 15 random seeds using the standard transformations \\(\\Theta^{\\mathrm{std}}\\). It can be seen that the obtained distributions are strongly/substantially separated in the case of static/trainable Winograd transformations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.787, + 0.647 + ], + "angle": 0, + "content": "We find that replacing the standard transformations \\(\\Theta^{\\mathrm{std}}\\) with \\(\\Theta^{\\mathrm{pseudo}}\\) matrices, even without including them in a training loop (static \\(\\Theta\\) scheme), already results in a considerable accuracy improvement (Tab. 1 and Fig. 3). This improvement is robust to variations of both network depth (Tab. 1) and channel scaling (dashed lines in Fig. 3). Learning Winograd transformations during training allows to" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.731 + ], + "angle": 0, + "content": "Table 1: The accuracy \\((\\%)\\) of 8-bit \\(F4-\\) and \\(F6\\) -based ResNet- \\(N\\) models (chan. mult. is 1.0) on CIFAR-10 when \\(\\Theta\\) is static. The best results are shown in bold. The values in parentheses give the improvement (to the first decimal place) over \\(\\Theta^{\\mathrm{std}}\\) case, and the results for the \\(F4\\) ResNet-18 are averaged over the experiments of Fig. 2." + }, + { + "type": "table", + "bbox": [ + 0.228, + 0.742, + 0.776, + 0.839 + ], + "angle": 0, + "content": "
ModelF4 Static ΘF6 Static Θ
Standard ΘOur Θpso (↑)Standard ΘOur Θpso (↑)
ResNet-1883.6591.54 (+7.9)83.2387.51 (+4.3)
ResNet-3484.2290.48 (+6.3)69.4887.09 (+17.6)
ResNet-5082.5791.52 (+8.9)70.1582.87 (+12.7)
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.526, + 0.129 + ], + "angle": 0, + "content": "V. Protsenko, V. Kryzhanovskiy et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.788, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.32, + 0.788, + 0.364 + ], + "angle": 0, + "content": "Fig. 3: The accuracy of the ResNet-18 on CIFAR-10 with various channel multipliers utilizing full 8-bit \\( F4 \\) (a) and \\( F6 \\) (b) Winograd convolutions. The SOTA data are taken from Fig. 4 of Ref. [17]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.789, + 0.522 + ], + "angle": 0, + "content": "further adjust them to minimize accuracy errors (see solid lines in Fig. 3, Tabs. 2 and 3). In all experiments, the initialization of the Winograd transformations when considering them as trainable parameters via our PSO-derived matrices \\(\\Theta^{\\mathrm{pso}}\\) is found to be highly beneficial. One can see that this training scenario not only considerably outperforms the SOTA \\(\\Theta^{\\mathrm{std}}\\)-based training approach of Ref. [17], but also either completely closes (in the case of CIFAR-10) or strongly reduces (for CIFAR-100 and ImageNet) the error gap between the \\(F4\\) Winograd convolution and the FP direct convolution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.528, + 0.79, + 0.62 + ], + "angle": 0, + "content": "Finally, Tab. 2 shows a competitive comparison of the performance of using trainable transformations starting from \\(\\Theta = \\Theta^{\\mathrm{pseudo}}\\) (column \"Our\") with the current SOTA methods. In most cases, our approach significantly outperforms the SOTA results in terms of accuracy. Note that for a fair comparison, in Tab. 2 we consider only the case of the scalar quantization and real-arithmetic full 8-bit quantized Winograd algorithm. However, our method is compatible with more" + }, + { + "type": "table_caption", + "bbox": [ + 0.231, + 0.655, + 0.771, + 0.671 + ], + "angle": 0, + "content": "Table 2: An accuracy (%) comparison of Winograd-aware quantization schemes." + }, + { + "type": "table", + "bbox": [ + 0.232, + 0.682, + 0.773, + 0.84 + ], + "angle": 0, + "content": "
Model & DatasetMethodFmRef.Ours
CIFAR-10ResNet-18Winograd-aware [17]F492.4693.26
Winograd-aware [17]F687.7489.57
ResNet-18 (mult. 0.5)Legendre [6]F491.892.53
ResNet-18 (mult. 0.25)Legendre [6]F489.790.29
ResNet-20WinoTrain [35]F489.792.31
WinoVidiVici [34]90.89
ResNet-18 & ImageNetWinoTrain [35]F461.2869.15
WinoVidiVici [34]69.14
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.34, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Quantization-Friendly Winograd Transformations for CNN" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.214 + ], + "angle": 0, + "content": "Table 3: Performance of the full 8-bit quantized \\( F4- \\) and \\( F6 \\)-based models (channel multiplier is 1.0) on the CIFAR-10 dataset when \\( \\Theta \\) is trainable. Note that here SOTA refers to the results of the training pipeline of Ref. [17] obtained using the official code released by the authors. The abbreviation FP stands for the results of the full precision direct convolution models. Other notations are the same as in Tab. 1." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.226, + 0.777, + 0.678 + ], + "angle": 0, + "content": "
Task (val.)Model (Test set)Direct conv.8-bit Fm with trainable Θ
FP8-bitFmSOTAOur Θpso (↑)
Classification (Accuracy, %)ResNet-18 (CIFAR-10)93.1793.20F492.5093.26 (+0.8)
F687.7489.57 (+1.8)
ResNet-18 (CIFAR-100)74.6074.18F472.3874.10 (+1.7)
F666.1572.90 (+6.7)
ResNet-18 (ImageNet)69.7969.61F467.0569.15 (+2.1)
F653.9062.87 (+9.0)
ResNet-34 (CIFAR-10)93.4093.12F492.5493.29 (+0.7)
F680.0590.16 (+10.1)
ResNet-50 (CIFAR-10)93.3593.22F493.4693.68 (+0.2)
F686.3792.16 (+5.8)
×3 Super-resolution (PSNR, dB)ESPCN [41] (Set5)32.5332.38F430.8632.11 (+1.2)
F627.3329.82 (+2.5)
ESPCN [41] (Set14)29.2129.09F428.0528.87 (+0.8)
F625.6227.30 (+1.7)
ESPCN-6 (Set5)33.2133.12F430.4532.98 (+2.5)
F625.5330.16 (+4.6)
ESPCN-6 (Set14)29.5929.50F427.7529.44 (+1.7)
F624.0227.50 (+3.5)
RFDN [29] (Set5)34.1333.68F430.6733.36 (+2.7)
RFDN [29] (Set14)30.1729.81F427.9429.62 (+1.7)
Segm IoUENet [38] (CamVid)60.6160.17F456.1260.00 (+3.9)
F652.4356.39 (+4.0)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.784, + 0.736 + ], + "angle": 0, + "content": "sophisticated quantization schemes (e.g., the tapwise quantization of Ref. [5]) and further performance improvement is expected from their application." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.756, + 0.406, + 0.771 + ], + "angle": 0, + "content": "4.2 Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Here we examine the performance on super-resolution (SR) tasks considering the original and modified ESPCN [41] as well as the RFDN [29] models. We set the upscaling factor between low- and high-resolution images to \\(\\times 3\\). Since the SOTA approach of Ref. [17] (training \\(\\Theta\\) starting from \\(\\Theta^{\\mathrm{std}}\\)) has only been" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.526, + 0.129 + ], + "angle": 0, + "content": "V. Protsenko, V. Kryzhanovskiy et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "implemented for classification scenarios, we extend it to SR networks based on its official implementation. All models are trained on the DIV2K dataset [1] and their performance is evaluated on the Set5 and Set14 benchmark data (more details on the SR architectures, training, and dataset preparation can be found in the Supplemental Material). As seen from the previous discussion, treating the Winograd transformations \\(\\Theta\\) as trainable parameters is essential to overcome the quality degradation caused by quantization errors. For this reason, in the following, we focus only on the case of trainable \\(\\Theta\\). The mean PSNR values of the considered models on the benchmark datasets are summarized in Tab. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.392 + ], + "angle": 0, + "content": "The ESPCN model is the less-depth model under consideration. It consists of three convolutional layers followed by a pixel-shuffle operation. In our implementation, the last two convolution layers are replaced by either \\( F4 \\) or \\( F6 \\) Winograd convolutions and are quantized to 8-bit. On F4, training of the transformations from \\( \\Theta = \\Theta^{\\mathrm{pso}} \\) gains 1.25dB and 0.82dB higher PSNR than the SOTA approach of Ref. [17] on Set5 and Set14, respectively. On \\( F6 \\), the PSNR gap from the FP direct model is reduced by about 2 times within our training approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.396, + 0.788, + 0.487 + ], + "angle": 0, + "content": "It can be expected that proper initialization of \\(\\Theta\\) at the beginning of training is less critical for such lightweight networks. Indeed, due to the relatively small number of the transformation parameters to train, it is possible that even a local gradient-based optimization during network training may be enough to learn appropriate Winograd transformations. Thus, quality degradation when using standard matrices \\(\\Theta^{\\mathrm{std}}\\) is expected to be more pronounced with increas" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.519, + 0.357, + 0.678 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.679, + 0.353, + 0.783 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.784, + 0.334, + 0.796 + ], + "angle": 0, + "content": "(a) Direct / FP" + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.519, + 0.5, + 0.679 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.679, + 0.497, + 0.783 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.379, + 0.784, + 0.482, + 0.796 + ], + "angle": 0, + "content": "(b) Direct / 8-bit" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.519, + 0.642, + 0.679 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.679, + 0.638, + 0.783 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.784, + 0.635, + 0.796 + ], + "angle": 0, + "content": "(c) \\(F4\\) SOTA / 8-bit" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.519, + 0.785, + 0.678 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.678, + 0.781, + 0.783 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.784, + 0.775, + 0.796 + ], + "angle": 0, + "content": "(d) \\(F4\\) Ours / 8-bit" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.808, + 0.787, + 0.836 + ], + "angle": 0, + "content": "Fig. 4: \\(\\times 3\\) Super-resolution results of the RFDN model utilizing direct and \\(F4\\) Winograd convolutions for \"Comic\" from Set14." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.34, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Quantization-Friendly Winograd Transformations for CNN" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "ing network depth, as in the case of the classification tasks (see Sec. 4.1). To prove this, we consider the six-layered ESPCN-6 model, which is obtained by including three additional convolution layers in the ESPCN model. While the PSNR values obtained by the \\(\\Theta^{\\mathrm{pseudo}}\\)-based training increase compared to the results of the original ESPCN model, the SOTA training approach, using \\(\\Theta^{\\mathrm{std}}\\), on the contrary, tends to PSNR degradation. On \\(F4\\), \\(\\Theta^{\\mathrm{pseudo}}\\)-based training shows an improvement of more than 2.5 dB for Set5 and 1.6 dB for Set14 relative to the SOTA training approach. With this improvement, the quality drop when using the 8-bit quantized \\(F4\\) algorithm is strongly reduced, and the resulting quality is close to that of FP direct convolutions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.306, + 0.789, + 0.58 + ], + "angle": 0, + "content": "Finally, we perform experiments with the residual feature distillation network (RFDN) [29]. It is based on the revised information distillation mechanism of the IMDN network [21] and the shallow residual blocks, which allows to achieve a better trade-off between model performance and complexity. Utilizing \\(\\Theta^{\\mathrm{pseudo}}\\) leads to remarkably better performance compared to the SOTA training based on the standard matrices \\(\\Theta^{\\mathrm{std}}\\). However, we find that the RFDN model is quite sensitive to quantization noise. This results in a noticeable drop in the quality when the model is quantized to 8-bit, even in the case of direct convolutions. When the Winograd algorithm is used, this quality drop is somewhat worsened. One of the reasons for this is that the quantization scheme adapted from Ref. [17] includes redundant quantization steps (see discussion in Sec. 3.1). We also revealed a strong effect of replacing the \\(3\\times 3\\) direct convolutions nearest to the contrast aware channel attention (CCA) layer [21, 29] with the \\(F4\\) convolutions on the enhancement of the PSNR drop. The latter requires searching for a more careful strategy to replace direct convolutions with the Winograd ones, and, a possible solution is to use for some layers the \\(F2\\) Winograd algorithm, which is not prone to strong quantization errors. Despite the above-mentioned drop of PSNR, our method produces super-resolution images that are visually competitive with ones" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.619, + 0.357, + 0.783 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.784, + 0.352, + 0.796 + ], + "angle": 0, + "content": "(a) ESPCN F6 SOTA" + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.619, + 0.5, + 0.783 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.784, + 0.49, + 0.796 + ], + "angle": 0, + "content": "(b) ESPCN F6 Our" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.619, + 0.643, + 0.783 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.784, + 0.642, + 0.796 + ], + "angle": 0, + "content": "(c) ESPCN-6 F6 SOTA" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.62, + 0.785, + 0.783 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.784, + 0.785, + 0.796 + ], + "angle": 0, + "content": "(d) ESPCN-6 F6 Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.808, + 0.788, + 0.837 + ], + "angle": 0, + "content": "Fig. 5: \\(\\times 3\\) Super-resolution results for \"Comic\" from Set14 within 8-bit ESPCN (first two panels) and ESPCN-6 (last two panels) model using the \\(F6\\) Winograd convolutions." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.526, + 0.129 + ], + "angle": 0, + "content": "V. Protsenko, V. Kryzhanovskiy et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "of the FP model (see Fig. 4) and are free of strong noise or gridding artifacts that typically occur in the case of standard Winograd transformations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.314 + ], + "angle": 0, + "content": "For \\( F6 \\), despite the considerable improvements achieved by using the transformations \\( \\Theta^{\\mathrm{pso}} \\) (see Tab. 3 and Fig. 5), the quality drop is still pronounced. In particular, we find that the training of the RFDN model is unstable for \\( F6 \\) (for this reason, the corresponding results are not reported here). However, we emphasize that here we considered the extreme case when all direct convolutions of kernel size \\( 3 \\times 3 \\) are replaced by the \\( F6 \\) convolutions. We believe it is possible to selectively replace error-tolerant direct convolutions with \\( F6 \\) ones based on our \\( \\Theta^{\\mathrm{pso}} \\) with minimal quality losses. This can be achieved, for example, by adapting frameworks for automatic design of neural network architectures [10, 15, 17, 46]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.332, + 0.376, + 0.348 + ], + "angle": 0, + "content": "4.3 Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.354, + 0.788, + 0.55 + ], + "angle": 0, + "content": "Finally, we briefly discuss segmentation tasks. To this end, we consider the ENet model [38] which is the optimized semantic segmentation network designed to run on mobile devices. We replace all \\( 3 \\times 3 \\) kernel direct convolutions with the Winograd convolutions, except ones for the last and first layers, and convolutions with a stride equal to 2. The latter is because the standard Winograd algorithm cannot be directly applied to convolutions with non-unit stride. However, we note that several adaptations of the Winograd algorithm to non-unit stride cases have recently been proposed [20,37,48]. Their quantization-friendly formulation within an optimization procedure similar to the one developed in the present study is of great interest for further research. The results of training the ENet model on the CamVid dataset [9] are shown in Tab. 3. As for SR tasks, we find that our \\( \\Theta^{\\mathrm{pso}} \\)-based approach is considerably more accurate compared to the SOTA approach utilizing the standard Winograd transformations \\( \\Theta^{\\mathrm{std}} \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.57, + 0.37, + 0.587 + ], + "angle": 0, + "content": "5 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.674 + ], + "angle": 0, + "content": "We have introduced an efficient way of training 8-bit quantized networks utilizing \\( F4 \\) and \\( F6 \\) Winograd convolutions. To this end, using the global partial swarm optimization (PSO) method, we have derived a model- and data-free form of the Winograd transformation matrices \\( \\Theta \\equiv \\{A,B,G\\} \\), which has significantly better accuracy performance compared to the standard ones." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To further improve accuracy we have modified the previous SOTA training pipeline of Ref. [17], which allows the transformation matrices \\(\\Theta\\) to evolve during training, by replacing the standard Winograd transformations used in this pipeline with our PSO-derived matrices. By treating all the steps of Winograd convolution as 8-bit quantized, we have considered classification, superresolution, and segmentation tasks to reveal the benefits of the proposed approach. In all of the experiments, we found that our modification results in a substantial improvement of accuracy and allows to either completely overcome or highly reduce quality drop. As a consequence, our approach significantly outperforms the SOTA method in terms of accuracy, while requiring the same number of operations during network training and retaining the same inference speedup." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.34, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Quantization-Friendly Winograd Transformations for CNN" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.174, + 0.785, + 0.214 + ], + "angle": 0, + "content": "1. Agustsson, E., Timofte, R.: Ntire 2017 challenge on single image super-resolution: Dataset and study. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 126-135 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.215, + 0.785, + 0.255 + ], + "angle": 0, + "content": "2. Alam, M.S., Misba, W.A., Atulasimha, J.: Quantized non-volatile nanomagnetic synapse based autoencoder for efficient unsupervised network anomaly detection. arXiv preprint arXiv:2309.06449 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.256, + 0.785, + 0.296 + ], + "angle": 0, + "content": "3. Alam, S.A., Anderson, A., Barabasz, B., Gregg, D.: Winograd convolution for deep neural networks: Efficient point selection. ACM Transactions on Embedded Computing Systems 21(6), 1-28 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.297, + 0.785, + 0.337 + ], + "angle": 0, + "content": "4. Anderson, A., Gregg, D.: Optimal dnn primitive selection with partitioned boolean quadratic programming. In: Proceedings of the 2018 International Symposium on Code Generation and Optimization. pp. 340-351 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.338, + 0.785, + 0.392 + ], + "angle": 0, + "content": "5. Andri, R., Bussolino, B., Cipolletta, A., Cavigelli, L., Wang, Z.: Going further with winograd convolutions: Tap-wise quantization for efficient inference on 4x4 tiles. In: 2022 55th IEEE/ACM International Symposium on Microarchitecture (MICRO). pp. 582-598. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.393, + 0.785, + 0.419 + ], + "angle": 0, + "content": "6. Barabasz, B.: Quantaized winograd/toom-cook convolution for dnns: Beyond canonical polynomials base. arXiv preprint arXiv:2004.11077 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.42, + 0.785, + 0.46 + ], + "angle": 0, + "content": "7. Barabasz, B., Anderson, A., Soodhalter, K.M., Gregg, D.: Error analysis and improving the accuracy of winograd convolution for deep neural networks. ACM Transactions on Mathematical Software (TOMS) 46(4), 1-33 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.461, + 0.785, + 0.5 + ], + "angle": 0, + "content": "8. Barabasz, B., Gregg, D.: Winograd convolution for dnns: Beyond linear polynomials. In: International Conference of the Italian Association for Artificial Intelligence. pp. 307-320. Springer (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.501, + 0.785, + 0.528 + ], + "angle": 0, + "content": "9. Brostow, G.J., Fauqueur, J., Cipolla, R.: Semantic object classes in video: A high-definition ground truth database. Pattern Recognition Letters 30(2), 88-97 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.529, + 0.785, + 0.554 + ], + "angle": 0, + "content": "10. Cai, H., Zhu, L., Han, S.: Proxylessnas: Direct neural architecture search on target task and hardware. arXiv preprint arXiv:1812.00332 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.555, + 0.785, + 0.595 + ], + "angle": 0, + "content": "1. Chen, T., Xu, W., Chen, W., Wang, P., Cheng, J.: Towards efficient and accurate winograd convolution via full quantization. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.596, + 0.785, + 0.637 + ], + "angle": 0, + "content": "2. Chen, Y.T., Ou, Y.F., Huang, C.T.: A winograd-based highly-parallel convolution engine for 8-bit cnn acceleration. In: 2022 IEEE 4th International Conference on Artificial Intelligence Circuits and Systems (AICAS). pp. 395–398. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.637, + 0.785, + 0.677 + ], + "angle": 0, + "content": "3. Chen, Y.H., Krishna, T., Emer, J.S., Sze, V.: Eyeriss: An energy-efficient reconfigurable accelerator for deep convolutional neural networks. IEEE journal of solid-state circuits 52(1), 127-138 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.678, + 0.785, + 0.704 + ], + "angle": 0, + "content": "4. Cook, S.: On the minimum computation time for multiplication. Doctoral diss., Harvard U., Cambridge, Mass 1 (1966)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.705, + 0.785, + 0.731 + ], + "angle": 0, + "content": "5. Elsken, T., Metzen, J.H., Hutter, F.: Neural architecture search: A survey. The Journal of Machine Learning Research 20(1), 1997-2017 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.732, + 0.785, + 0.758 + ], + "angle": 0, + "content": "6. Fernández-Marqués, J.: High performance deep learning on resource constrained platforms. Ph.D. thesis, University of Oxford (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "7. Fernandez-Marques, J., Whatmough, P., Mundy, A., Mattina, M.: Searching for winograd-aware quantized networks. Proceedings of Machine Learning and Systems 2, 14-29 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "8. Gholami, A., Kim, S., Dong, Z., Yao, Z., Mahoney, M.W., Keutzer, K.: A survey of quantization methods for efficient neural network inference. In: Low-Power Computer Vision, pp. 291-326. Chapman and Hall/CRC (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.174, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.525, + 0.128 + ], + "angle": 0, + "content": "V. Protsenko, V. Kryzhanovskiy et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "19. Holt, J.L., Hwang, J.N.: Finite precision error analysis of neural network hardware implementations. IEEE Transactions on Computers 42(3), 281-290 (1993)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.232 + ], + "angle": 0, + "content": "20. Huang, C., Dong, X., Li, Z., Song, T., Liu, Z., Dong, L.: Efficient stride 2 winograd convolution method using unified transformation matrices on fpga. In: 2021 International Conference on Field-Programmable Technology (ICFPT). pp. 1-9. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.785, + 0.274 + ], + "angle": 0, + "content": "21. Hui, Z., Gao, X., Yang, Y., Wang, X.: Lightweight image super-resolution with information multi-distillation network. In: Proceedings of the 27th acm international conference on multimedia. pp. 2024-2032 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.785, + 0.303 + ], + "angle": 0, + "content": "22. Jorda, M., Valero-Lara, P., Pena, A.J.: Performance evaluation of cudnn convolution algorithms on nvidia volta gpus. IEEE Access 7, 70461-70473 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.785, + 0.346 + ], + "angle": 0, + "content": "23. Kennedy, J., Eberhart, R.: Particle swarm optimization. In: Proceedings of ICNN'95-international conference on neural networks. vol. 4, pp. 1942-1948. IEEE (1995)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.785, + 0.374 + ], + "angle": 0, + "content": "24. Krishnamoorthi, R.: Quantizing deep convolutional networks for efficient inference: A whitepaper. arXiv preprint arXiv:1806.08342 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.785, + 0.416 + ], + "angle": 0, + "content": "25. Lavin, A., Gray, S.: Fast algorithms for convolutional neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4013-4021 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.418, + 0.785, + 0.459 + ], + "angle": 0, + "content": "26. Li, C., Jiang, P., Zhou, H., Wang, X., Zhao, X.: Haw: Hardware-aware point selection for efficient winograd convolution. IEEE Signal Processing Letters 30, 269-273 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.501 + ], + "angle": 0, + "content": "27. Li, G., Jia, Z., Feng, X., Wang, Y.: Lowino: Towards efficient low-precision winograd convolutions on modern cpus. In: Proceedings of the 50th International Conference on Parallel Processing. pp. 1-11 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.557 + ], + "angle": 0, + "content": "28. Li, G., Liu, L., Wang, X., Ma, X., Feng, X.: Lance: efficient low-precision quantized winograd convolution for neural networks based on graphics processing units. In: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 3842-3846. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.785, + 0.6 + ], + "angle": 0, + "content": "29. Liu, J., Tang, J., Wu, G.: Residual feature distillation network for lightweight image super-resolution. In: Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 41-55. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.785, + 0.628 + ], + "angle": 0, + "content": "30. Liu, Z.G., Mattina, M.: Efficient residue number system based winograd convolution. In: European Conference on Computer Vision. pp. 53-68. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.63, + 0.785, + 0.656 + ], + "angle": 0, + "content": "31. Mathieu, M., Henaff, M., LeCun, Y.: Fast training of convolutional networks through fts. arXiv preprint arXiv:1312.5851 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.658, + 0.785, + 0.685 + ], + "angle": 0, + "content": "32. Meng, L., Brothers, J.: Efficient winograd convolution via integer arithmetic. arXiv preprint arXiv:1901.01965 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.687, + 0.785, + 0.714 + ], + "angle": 0, + "content": "33. Miranda, L.J.: Pyswarms: a research toolkit for particle swarm optimization in python. Journal of Open Source Software 3(21), 433 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.715, + 0.785, + 0.784 + ], + "angle": 0, + "content": "34. Mori, P., Frickenstein, L., Sampath, S.B., Thoma, M., Faspous, N., Vemparala, M.R., Frickenstein, A., Unger, C., Stechele, W., Mueller-Gritschneider, D., et al.: Wino vidi vici: Conquering numerical instability of 8-bit winograd convolution for accurate inference acceleration on edge. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 53-62 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "35. Mori, P., Sampath, S.B., Frickenstein, L., Vemparala, M.R., Faspous, N., Frickenstein, A., Stechele, W., Passerone, C.: Winotrain: Winograd-aware training for accurate full 8-bit convolution acceleration. In: 2023 60th ACM/IEEE Design Automation Conference (DAC). pp. 1-6. IEEE (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.34, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Quantization-Friendly Winograd Transformations for CNN" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "36. Nagel, M., Fournarakis, M., Amjad, R.A., Bondarenko, Y., Van Baalen, M., Blankevoort, T.: A white paper on neural network quantization. arXiv preprint arXiv:2106.08295 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "37. Pan, J., Chen, D.: Accelerate non-unit stride convolutions with winograd algorithms. In: Proceedings of the 26th Asia and South Pacific Design Automation Conference. pp. 358-364 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.272 + ], + "angle": 0, + "content": "38. Paszke, A., Chaurasia, A., Kim, S., Culurciello, E.: Enet: A deep neural network architecture for real-time semantic segmentation. arXiv preprint arXiv:1606.02147 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.3 + ], + "angle": 0, + "content": "39. Shafique, M.A., Munir, A., Kong, J.: Deep learning performance characterization on gpus for various quantization frameworks. AI 4(4), 926-948 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.355 + ], + "angle": 0, + "content": "40. Shi, B., Zhang, J., He, Z., Wei, X., Li, S., Luo, G., Zheng, H., Xie, Y.: Efficient super-resolution system with block-wise hybridization and quantized winograd on fpga. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.412 + ], + "angle": 0, + "content": "41. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1874-1883 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.412, + 0.788, + 0.452 + ], + "angle": 0, + "content": "42. Toom, A.L.: The complexity of a scheme of functional elements simulating the multiplication of integers. In: Doklady Akademii Nauk. vol. 150, pp. 496-498. Russian Academy of Sciences (1963)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.453, + 0.788, + 0.493 + ], + "angle": 0, + "content": "43. Vasilache, N., Johnson, J., Mathieu, M., Chintala, S., Piantino, S., LeCun, Y.: Fast convolutional nets with fbfft: Agpu performance evaluation. arXiv preprint arXiv:1412.7580 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.494, + 0.788, + 0.522 + ], + "angle": 0, + "content": "44. Vincent, K., Stephano, K., Frumkin, M., Ginsburg, B., Demouth, J.: On improving the numerical stability of winograd convolutions (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.55 + ], + "angle": 0, + "content": "45. Wang, D., Tan, D., Liu, L.: Particle swarm optimization algorithm: an overview. Soft computing 22, 387-408 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.55, + 0.788, + 0.591 + ], + "angle": 0, + "content": "46. Wang, K., Liu, Z., Lin, Y., Lin, J., Han, S.: Haq: Hardware-aware automated quantization with mixed precision. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8612-8620 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.591, + 0.745, + 0.606 + ], + "angle": 0, + "content": "47. Winograd, S.: Arithmetic complexity of computations, vol. 33. Siam (1980)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.606, + 0.788, + 0.646 + ], + "angle": 0, + "content": "48. Yepez, J., Ko, S.B.: Stride 2 1-d, 2-d, and 3-d winograd for convolutional neural networks. IEEE Transactions on Very Large Scale Integration (VLSI) Systems 28(4), 853-863 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.646, + 0.788, + 0.688 + ], + "angle": 0, + "content": "49. Zhang, Y., Wang, S., Ji, G., et al.: A comprehensive survey on particle swarm optimization algorithm and its applications. Mathematical problems in engineering 2015 (2015)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.688 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Quantization-Friendly Winograd Transformations for Convolutional Neural Networks/90e7434b-cd12-4e62-9976-08dc6035e0cd_origin.pdf b/2024/Quantization-Friendly Winograd Transformations for Convolutional Neural Networks/90e7434b-cd12-4e62-9976-08dc6035e0cd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3a0e64a0532d1fa67fddc1d7dd86ef463c16495b --- /dev/null +++ b/2024/Quantization-Friendly Winograd Transformations for Convolutional Neural Networks/90e7434b-cd12-4e62-9976-08dc6035e0cd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:72037ebb69b73335ed993e3e63cb1847827a7650c7c1312168933b46c8b70e69 +size 7590078 diff --git a/2024/Quantization-Friendly Winograd Transformations for Convolutional Neural Networks/full.md b/2024/Quantization-Friendly Winograd Transformations for Convolutional Neural Networks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4e75f176e020c611b642639ee84fd1c5d57852d9 --- /dev/null +++ b/2024/Quantization-Friendly Winograd Transformations for Convolutional Neural Networks/full.md @@ -0,0 +1,285 @@ +# Quantization-Friendly Winograd Transformations for Convolutional Neural Networks + +Vladimir Protsenko, Vladimir Kryzhanovskiy, and Alexander Filippov + +Huawei Noah's Ark Lab +protsenko.vladimir@huawei-partners.com, {kryzhanovskiy.vladimir, filippov alexander}@huawei.com + +Abstract. Efficient deployment of modern deep convolutional neural networks on resource-constrained devices suffers from demanding computational requirements of convolution operations. Quantization and use of Winograd convolutions operating on sufficiently large-tile inputs are two powerful strategies to speed up convolution operations. However, their combination results in numerical instability, which manifests itself in a strong quality performance degradation. We present an efficient learning scenario that either completely overcomes or strongly reduces the accuracy degradation of full 8-bit quantized $F(4,3)$ and $F(6,3)$ Winograd convolutions. Within the global particle swarm optimization (PSO), we derived a set of quantization-friendly Winograd transformations. Following the state-of-the-art (SOTA) training pipeline [J. Fernandez-Marques et al., Proc. Mach. Learn. Syst. 2, 14-29 (2020)], we treat Winograd transformations as learnable parameters during network training. Evolving transformations starting from our PSO-derived ones rather than the standard Winograd transformations results in significant numerical error reduction and accuracy improvement. As a consequence, our approach significantly outperforms SOTA methods on various tasks. + +Keywords: Winograd convolutions $\cdot$ Quantization $\cdot$ Efficient inference + +# 1 Introduction + +To date, deep convolutional neural networks (CNN) provide state-of-the-art performance in a wide variety of machine learning domains, including image processing, computer vision, video-based workflow analysis tasks, etc. The convolution operations, being the core block of CNNs, represent their most time-consuming and computationally intensive part [13]. Since modern CNNs typically include deep stacks of convolution layers both their training and inference are highly computationally demanding. As a consequence, despite recent advances in hardware technology, efficient inference of modern CNNs on resource-constrained mobile and embedded devices remains to be a challenging task. + +With the recent explosion of support for low-precision arithmetic in modern hardware, quantization [18, 24, 36] has become one of the standard approaches for compression and acceleration of CNNs. The quantization process is defined + +by mapping the network parameters to their low-precision representation. This allows to operate within low-bit arithmetic, having significantly lower computational and storage requirements than the full precision counterpart. Quantization to 8-bit integers generally produces only a moderate accuracy degradation and is widely acceptable in mobile devices. + +An alternative way to achieve efficient inference is to speed up convolution operations at the algorithm level. Several efficient convolution techniques have been developed, including the fast Fourier transform (FFT) [31, 43] and Winograd minimal filtering algorithm [25, 47]. The latter is known to minimize computational costs for convolutions with kernels of sizes $3 \times 3$ [4, 22], which are most often used in modern CNNs. The $F(m, r)$ Winograd algorithm transforms tiles of input feature map and $r \times r$ filter into the Winograd domain, where $m^2$ outputs can be simultaneously computed with a single element-wise matrix multiplication. Contrary to the FFT, the original Winograd algorithm does not involve complex number multiplications, which are at least $3 \times$ computationally costly compared to that of real numbers. Because of its cutting-edge performance, the Winograd algorithm is integrated into various deep-learning libraries. + +As the quantization and Winograd algorithms are not mutually exclusive but complementary, their mixing is attracting considerable attention from both academic and industrial fields. Theoretically, the larger the tile size $k_{\mathrm{W}} = m + r - 1$ of input feature maps, the more speedup can be expected from using the $F(m,r)$ Winograd algorithm. However, in practice, large-tile Winograd algorithms turn out to be highly sensitive to numerical errors, which grow exponentially with tile size [7,8]. Quantization, lowering arithmetic precision, further amplifies numerical errors [19] and causes numerical instability of the Winograd algorithm, which results in severe degradation of network performance. This strongly inhibits the industrial use of large-tile ( $k_{\mathrm{W}} \gtrsim 4$ ) low-bit Winograd-based convolutions, including the perspective $F(4,3)$ and $F(6,3)$ . Despite all efforts to mitigate this problem (see Sec. 2), the performance of large-tile Winograd algorithms quantized to the widely supported 8-bit integer format remains far from satisfactory. + +In this paper, we present an efficient learning scenario that either completely overcomes or strongly reduces the accuracy degradation of full 8-bit quantized $F(4,3)$ and $F(6,3)$ Winograd convolutions. As a first step, we perform a global particle swarm optimization (PSO) search over the general family of Winograd algorithms, which are defined by specifying a set of transformation matrices $\Theta \equiv \{A,B,G\}$ , to select one that is more robust to quantization errors. This step is prior to network training and yields a model- and data-free set of quantization-friendly Winograd transformation matrices $\Theta$ . We demonstrate that replacing the standard Winograd transformation matrices with our PSO-derived ones already results in significant accuracy improvements. To further tackle the performance degradation challenge, we integrated our transformations into the SOTA learning pipeline of Ref. [17], which treats Winograd transformation matrices as trainable parameters during network training. We show that allowing the evolution of Winograd transformations $\Theta$ starting from our PSO-derived matrices, rather than from standard Winograd matrices as in the original pipeline of + +Ref. [17], has a direct impact on increasing networks performance. As a result, our learning strategy allows to outperform existing approaches (and the 8-bit SOTA results) by demonstrating superior accuracy in various tasks, including classification, super-resolution, and semantic segmentation. + +# 2 Related Works + +The Toom-Cook minimal filtering algorithms [14, 42], generalized by Winograd [47], were first applied to CNNs by Lavin and Gray [25] to accelerate convolutions with $3 \times 3$ filters. Barabasz et al. [7] presented a theoretical analysis of the accuracy of Winograd algorithms and revealed that the error bound for the Toom-Cook algorithm grows exponentially with tile size. + +To reduce poor numerical properties of Winograd convolutions, approaches to a more efficient selection of polynomial points have been proposed [3, 7, 26]. While most of them have empirical grounds, Ref. [26] leverages reinforcement learning to construct an optimal polynomial point selection policy. This reinforcement learning-based framework significantly improves the accuracies of quantized $F(m, 3)$ Winograd convolutions with $m \leq 5$ and leads to only a small accuracy gap between Winograd convolutions and the direct convolutions. However, the performance benefits of this approach are only demonstrated for mixed 8/16-bit quantization, which is a significantly softer restriction compared to considering all Winograd convolution steps quantized below 8-bit. The accuracy improvement with mixed-bit precision or quantization only in the Winograd domain has been shown in [5, 27, 28, 40]. For quantized networks, Meng and Brothers [32] extended the Winograd algorithm to the complex number domain. But, the shift to complex transformations substantially reduces the multiplication saving and hence the speedup of the Winograd algorithm. + +Fernandez-Marques et al. [16, 17] first introduced the idea of considering the Winograd transformation matrices as learnable parameters while training quantized networks. Including the Winograd transformations in the training loop introduces additional degrees of freedom that are aware of numerical errors and can be adjusted to reduce them. On $F(m \leq 6,3)$ , it was demonstrated that training the Winograd transformation matrices, as opposed to keeping them static (non-trainable), causes a dramatic reduction of the accuracy gap between 8-bit quantized Winograd and direct convolutions. It is important to emphasize that, despite it is not clearly stated in the paper by Fernandez-Marques et al., all the parts of Winograd convolutions (including the Winograd transformation matrices) are maintained in 8-bit precision. This is clear from the official source code provided by the authors (https://github.com/jafermarq/WinogradAwareNets). We additionally verified that indeed the above official implementation strictly reproduces all the results of the corresponding study. + +Despite the significant accuracy improvements achieved by Ref. [17], there are still accuracy losses for the perspective $F(4,3)$ and $F(6,3)$ algorithms, which hinder their industrial potential. The study [6] continues the work of [17] and applies a base change technique for quantized Winograd convolutions. This tech- + +nique was tested for $F(4,3)$ ResNet-18 model on CIFAR-10 dataset. It is shown that transformations into Legendre polynomials base for 8-bit quantized trainable Winograd convolutions further reduce the errors of the original pipeline of Ref. [17]. However, the proposed approach can achieve the same accuracy as for direct convolutions only when the element-wise (Hadamard) product is extended to 9-bit. These improvements after modification of Winograd matrices yield the important hint that the selection of the appropriate starting values for these matrices in the training pipeline of Ref. [17] is of critical importance. + +Besides the works [6,17], there are other studies on the use of full 8-bit quantized Winograd convolutions [11, 12, 30, 34, 35]. Each of these studies, however, compared to the approach of Ref. [17] comes with its own limitations on either performance or resulting acceleration. Ref. [35] performs quantization-aware training (QAT) to mitigate the numerical error of quantized Winograd inference. While preserving multiplication savings of the Winograd algorithm, this strategy loses in performance relative to Ref. [17]. Ref. [30] adopts a residue number system to eliminate accuracy degradation, but as a consequence, the theoretical multiplication savings are dramatically reduced by this approach. Ref. [12] (and the most accurate scheme of Ref. [34]) uses complex numbers based Winograd algorithms, which heavily reduces the speedup. + +In light of the above, the training scheme of Ref. [17] remains the SOTA in terms of tackling the problem of accuracy degradation while preserving the speedup of the standard Winograd algorithm. In the following, we refer to the pipeline of Ref. [17] as the SOTA approach. We emphasize that by this we mean only the case of full 8-bit quantization and real number arithmetic of the Winograd convolutions, since mixed-precision quantization has the potential to achieve better performance (e.g., see [26] for the SOTA results for mix 8/16-bit quantization). + +# 3 Our Method + +# 3.1 Preliminaries + +Winograd Convolution. The 2D Winograd convolution $F(m, r)$ computes $m \times m$ output by element-wise matrix multiplication of the Winograd domain-transformed filter $g$ of size $r \times r$ and input feature map $d$ of size $k_W \times k_W$ , where $k_W = m + r - 1$ . The whole procedure can be expressed in matrix form as follows: + +$$ +Y = A ^ {T} \left[ \left(G g G ^ {T}\right) \odot \left(B ^ {T} d B\right) \right] A, \tag {1} +$$ + +where $A \in \mathbb{R}^{k_W \times m}$ , $B \in \mathbb{R}^{k_W \times k_W}$ , $G \in \mathbb{R}^{k_W \times r}$ are constant real Winograd transformation matrices and $\odot$ is the Hadamard (element-wise) multiplication. For brevity, we use the notation $\Theta \equiv \{A, B, G\}$ for the set of the Winograd matrices hereafter. The use of this algorithm reduces arithmetic complexity by simultaneously allowing $m^2$ outputs to be computed by performing $(m + r - 1)^2$ multiplications, instead of the $(rm)^2$ required for a naive convolution operation. This reduction of computationally costly multiplication operations together with the + +ability to compute multiple outputs results in a significant acceleration compared using of the direct convolution (e.g., theoretically $4 \times$ speedup on $F(4,3)$ ). In this study, we consider $3 \times 3$ filters, and for concise notation we denote $F(m,r = 3)$ as $F_{\mathrm{m}}$ (e.g., $F(4,3) \equiv F4$ in our notations). + +As can be seen from Eq. (1), the Winograd convolution is defined by specifying the transformation matrices $\Theta$ . The particular form of $\Theta$ depends on the size of the filter $r$ and the desired number of outputs. In the case of $3 \times 3$ filters, which are widely used in modern CNNs, for $m \leq 6$ there is a general agreement on the form of the transformation matrices $\Theta$ . The corresponding matrices are given in [25] (see also Sec. 1 of the Supplemental Material) and we refer to this set of matrices as the standard Winograd transformations $\Theta^{\mathrm{std}}$ . + +Quantization. Following Ref. [17], we quantize all steps of the Winograd algorithm (including the transformations $\Theta$ ) to 8-bit. Therefore, the full 8-bit quantized Winograd convolution can be written as + +$$ +Y = q _ {8} \left(\hat {Q} _ {8} ^ {A} \left(\hat {Q} _ {8} ^ {G ^ {T}} (g) \odot \hat {Q} _ {8} ^ {B} (d)\right)\right), \tag {2} +$$ + +where $\hat{Q}_8^A (x) = q_8(A^T)q_8(x)q_8(A)$ is the transformation operator to/from the quantized Winograd domain and $q_{8}$ stands for 8-bit quantization. As in [17], we use the signed symmetric uniform scalar quantization scheme [36]. Note that for a proper comparison with the SOTA pipeline of Ref. [17], as in that study, both input and output are quantized. However, this quantization granularity is clearly redundant and may result in additional loss of accuracy. Therefore, we expect that within the conventional quantization schemes when either the input or the output is quantized, the proposed in this study approach provides further accuracy improvements. + +# 3.2 Quantization-Friendly Transformations + +Motivation. As discussed in Refs. [25,27,34], the magnitude of matrix elements of the Winograd transformations $\Theta \equiv \{A,B,G\}$ tends to increase with tile size $k_{W}$ , which considerably broadens numerical ranges and effectively reduces computation accuracy. In [7] Barabasz et al. analytically derived the floating-point error bound for the Winograd algorithm. It is shown that the error is proportional to the norms of the Winograd transform matrices. One can naturally expect the same relation between the quantization errors and matrix elements (and norms) of the Winograd transformations $\Theta$ . Indeed, the error coming from each quantization step in Eq. (2) is amplified proportionally to magnitudes of the following matrix transformations (if not taking into account the rare self errors cancellation effects). Another point is that transformation matrices with high-magnitude matrix elements considerably extend the quantization interval [27,34], whereas most of it is very sparsely used [34]. This leads to strong rounding errors during quantization [34]. Here, to improve the numerical properties of quantized Winograd convolutions, we exploit the fact that transformations $\Theta$ are not uniquely defined and construct a representation having as small matrix elements as possible. We demonstrate (see Sec. 4) that the resulting set of matrices $\Theta^{\mathrm{pso}}$ is + +quantization-friendly and yields a significant improvement in accuracy compared to using the standard Winograd transformations $\Theta^{\mathrm{std}}$ . + +PSO Optimization. To reduce the magnitudes of the Winograd transformations we consider the following optimization problem + +$$ +\Theta^ {\mathrm {p s o}} = \underset {\Theta} {\arg \min } f (\Theta) = \underset {\Theta} {\arg \min } \left(\gamma_ {A} \| A \| _ {F} ^ {2} + \gamma_ {B} \| B \| _ {F} ^ {2} + \gamma_ {G} \| G \| _ {F} ^ {2}\right), \tag {3} +$$ + +where $\gamma_A, \gamma_B, \gamma_G > 0$ are weight factors, $\| \cdot \|_F$ denotes the Frobenius norm of a matrix and the optimization is restricted to the set of transformations $\Theta$ that fulfill Eq. (1) to exactly preserve the property to be convolution operation. Although mainly the matrices $A$ and $B$ contain high-value matrix elements (see Supplemental Material Sec. 1), one should also be aware of preventing the possible growth of the matrix elements of $G$ in a new representation. Given this $G$ is also included in Eq. (3). To reveal main the effect of minimizing $f(\Theta)$ , we do not fine-tune the parameters $\gamma_p$ , $(p \in \{A, B, G\})$ , and simply set $\gamma_p = 1$ . However, we note that a more careful choice of these parameters (e.g., with taking into account quantization granularity) represents one potential way to further improve the quantization properties of transformation matrices $\Theta$ . + +To deal with Eq. (3), we start from the extended form of the Winograd transformation matrices $\Theta$ of Ref. [44] + +$$ +A ^ {T} = \left(V _ {(m + n - 1) \times m}\right) ^ {T} S _ {A}, \tag {4} +$$ + +$$ +B ^ {T} = S _ {B} \left(V _ {(m + n - 1) \times (m + n - 1)}\right) ^ {- T}, \tag {5} +$$ + +$$ +G = S _ {G} \left(V _ {(m + n - 1) \times n}\right), \tag {6} +$$ + +where $V_{x\times y}$ is a trimmed Vandermonde matrix given by + +$$ +V = \left[ \begin{array}{c c c c} \xi_ {0} ^ {0} \eta_ {0} ^ {y - 1} & \xi_ {0} ^ {1} \eta_ {0} ^ {y - 2} & \dots \dots & \xi_ {0} ^ {y - 1} \eta_ {0} ^ {0} \\ \xi_ {1} ^ {0} \eta_ {1} ^ {y - 1} & \xi_ {1} ^ {1} \eta_ {1} ^ {y - 2} & \dots \dots & \xi_ {1} ^ {y - 1} \eta_ {1} ^ {0} \\ \vdots & \vdots & & \vdots \\ \xi_ {x - 1} ^ {0} \eta_ {x - 1} ^ {y - 1} & \xi_ {x - 1} ^ {1} \eta_ {x - 1} ^ {y - 2} & \dots \dots & \xi_ {x - 1} ^ {y - 1} \eta_ {x - 1} ^ {0} \end{array} \right], \tag {7} +$$ + +where $\{(\xi_i,\eta_i)\}$ is the set of the homogenous coordinates (polynomial points) and $S_{p\in \{A,B,G\}}$ are the diagonal scaling matrices fulfilling the $S_{A}S_{B}S_{G} = 1$ condition. Within this representation, 6 and 8 polynomial points need to be given to construct $F4$ and $F6$ , respectively. The polynomial points and scalings which generate the standard sets of the Winograd matrices $\Theta^{\mathrm{std}}$ for $F4$ and $F6$ algorithms are given in Supplemental Material Sec. 1. + +We consider polynomial points $\{(\xi_i,\eta_i)\}$ as optimization parameters and perform minimization of the objective function $f(\Theta)$ using the particle swarm optimization (PSO) method [23, 45, 49]. PSO is a powerful swarm intelligence technique for solving global optimization problems. In this method, a set of $N$ candidate solutions, called swarm particles, are moved through a search-space. + +Each $i$ th, $i \in (1 \dots N)$ , particle is assigned with a position $\mathbf{x}_i$ and a velocity $\mathbf{v}_i$ , which evolve during unity time steps $t$ (or iteration numbers) according to the following update rules + +$$ +\mathbf {v} _ {i} (t + 1) \leftarrow \omega \mathbf {v} _ {i} (t) + \phi_ {1} r _ {1} [ \mathbf {p} _ {i} - \mathbf {x} _ {i} (t) ] + \phi_ {2} r _ {2} [ \mathbf {p} _ {g} - \mathbf {x} _ {i} (t) ], \tag {8} +$$ + +$$ +\mathbf {x} _ {i} (t + 1) \leftarrow \mathbf {x} _ {i} (t) + \mathbf {v} _ {i} (t + 1), \tag {9} +$$ + +where $\mathbf{p}_i$ is the best solution that $i$ th particle has found so far, $\mathbf{p}_g$ is the best solution revealed by the whole swarm, $r_{1(2)}$ are random numbers sampled at each time step from a uniform $[0,1]$ distribution. Finally, $\omega$ , $\phi_1$ and $\phi_2$ are PSO hyperparameters, which we set to 0.9, 0.55 and 0.3. The PSO optimization is implemented within PySwarms toolkit [33]. + +The evolution of the objective function $f(\Theta)$ for $10^{4}$ iterations of the PSO algorithm with 2000 particles is shown in Fig. 1a. Compared to the value for the standard transformations $\Theta^{\mathrm{std}}$ (indicated by circled number 0), the objective function after the PSO optimization (circled number 6) is reduced by about 2.6 and 5.2 times for $F4$ and $F6$ , respectively. By varying the hyperparameters $(\omega, \phi_1, \phi_2)$ , we verified that the PSO optimization leads either to the same value of the objective function (see inset of Fig. 1a) or to sub-optimal solutions (not shown). In the following, we denote the Winograd transformation matrices, which represent the PSO solution to the optimization problem given by Eq. (3), as $\Theta^{\mathrm{pso}}$ . The corresponding polynomial points $\{(\xi_i, \eta_i)\}$ from which they can be constructed according to Eqs. (4) to (7) and explicit expressions for $\Theta^{\mathrm{pso}}$ are given in the Supplemental Material Sec. 1. Note that in addition to the polynomial points, scaling matrices $S_{p \in \{A, B, G\}}$ can also be considered as optimization parameters. However, the polynomial points already provide certain degrees of freedom for the scaling variation of the transformation matrices. In addition, the inclusion of $S_{p \in \{A, B, G\}}$ in the PSO procedure leads to a significant expansion of the number of parameters to be optimized, which makes optimization $f(\Theta)$ inefficient due to getting stuck in local minima. In view of this, we do not optimize $S_{p \in \{A, B, G\}}$ and use their standard values (see Ref. [44], Secs. 1 and 2 of the Supplemental Material). + +# 4 Experimental Results + +In our experiments we use two different training policies: (i) keep the Winograd transformations $\Theta$ fixed, which we refer to as Static $\Theta$ ; (ii) treat $\Theta$ as learnable parameters during training, which we denote as Trainable $\Theta$ . The scheme (i) is used to show that replacing the standard Winograd matrices $\Theta^{\mathrm{std}}$ by our PSO-derived ones $\Theta^{\mathrm{pso}}$ already leads to a significant improvement in accuracy, which demonstrates a quantization-friendly nature of the latter. Within the second scheme (ii), we use the SOTA training approach of Ref. [17] (see discussion in Sec. 2). We replace the standard matrices $\Theta^{\mathrm{std}}$ used in this pipeline by our $\Theta^{\mathrm{pso}}$ . We show that this modification allows us to outperform the SOTA results. + +To ensure a fair comparison of different training scenarios, we extensively evaluate their performance on three different types of tasks: (i) classification, (ii) super-resolution, and (iii) semantic segmentation. Unless otherwise stated, all convolution layers are quantized to 8-bit, except for the first and last convolution layers. For experiments with Winograd convolutions, we replace the direct convolutions having stride 1 and kernel size $3 \times 3$ with $F4$ or $F6$ algorithms. Details on datasets and training protocols are presented in Secs. 3 and 4 of the Supplemental Material. + +# 4.1 Classification + +We begin by considering the ResNet- $N$ family of models, $18 \leq N \leq 50$ . As in Ref. [17], in some cases we vary a channel multiplier (or "width multiplier" as used for the MobileNets) factor, which scales the number of output channels produced by each convolution layer. A channel multiplier equal to 1.0 corresponds to the full ResNet- $N$ model. + +To clearly demonstrate that minimization of the objective function $f(\Theta)$ (see Eq. (3)) indeed leads to the reduction of quantization errors, we consider the accuracy of intermediate steps of the PSO-optimization in Fig. 1. Each $i$ th step (circled number) represents transformations $\Theta^i$ corresponding to a particular value of the objective function (see panels a). The 0th step corresponds to + +![](images/897cfdd43f65e4cfb2ea06f2e017e388adb51e0ccba094c0efe053a1345a2f22.jpg) +Fig. 1: a) The convergence history of the objective function $f(\Theta)$ (see Eq. (3)) for $F4$ Winograd algorithm within the PSO optimization. Inset shows five independent runs of the PSO algorithm. b, c) Accuracy performance of the 8-bit $F4$ ResNet-18 model on the CIFAR-10 dataset at different steps of the PSO optimization (circled numbers in Fig. 1a and on the horizontal axis of Fig. 1b,c) for static (b) and trainable (c) $\Theta$ . Red and black dots correspond to channel multipliers of 1.0 and 0.5, respectively. The dashed lines in Fig. 1c represent the accuracy of the FP direct convolution model. On F6 we obtained a qualitatively similar convergence and accuracy evolution (see Sec. 5 of the Supplemental Material). + +![](images/f276b9be4ac85e81f5a6e0ecdd55da88d3951a16aa029f9ab8e00c0e09f81c9b.jpg) + +![](images/f009d19f3516ef1a59dacb60c015f8fc17f8c9ebffb86ad810e317f8568bb5e5.jpg) + +![](images/f9159249a0a525cc74b8bc82629bb961a8f764b09b488494ef2d09bda1f70e32.jpg) +Fig. 2: Distribution of the accuracy of 8-bit $F4$ ResNet-18 model (chan. mult. is 1.0) on the CIFAR-10 dataset over 15 training runs for different training scenarios. + +![](images/91629ae99bb399bd7ad57449f9f76cf4f730bcc20d02fbdf18135c98cc76a17e.jpg) + +the standard Winograd transformations $\Theta^{\mathrm{std}}$ , while the 6th step represents our transformations $\Theta^{\mathrm{pso}}$ . One can see that the PSO optimization of $f(\Theta)$ results in a monotonic increase of the accuracy. In the case of the $F4$ algorithm with trainable $\Theta$ (Fig. 1c), by starting training from $\Theta = \Theta^{\mathrm{pso}}$ we not only outperform the SOTA $\Theta^{\mathrm{std}}$ -based results of Ref [17], but also achieve the accuracy of full-precision (FP) direct convolution models (denoted by dashed lines). Note that in some cases we observe the same effects as in Refs. [17,39], when quantization yields somewhat better accuracy compared to one of FP model. This is because quantization can provide an additional mechanism of regularization [2,39]. + +The characteristic improvements from using $\Theta^{\mathrm{pseudo}}$ are not sensitive to both the precision of PSO convergence and random factors during network initialization. To show this we perform three training runs of the $F4$ ResNet-18 model with different random seeds for each of the five converged configurations shown in the inset of Fig. 1a. The corresponding accuracy distribution is compared in Fig. 2 to that obtained for 15 random seeds using the standard transformations $\Theta^{\mathrm{std}}$ . It can be seen that the obtained distributions are strongly/substantially separated in the case of static/trainable Winograd transformations. + +We find that replacing the standard transformations $\Theta^{\mathrm{std}}$ with $\Theta^{\mathrm{pseudo}}$ matrices, even without including them in a training loop (static $\Theta$ scheme), already results in a considerable accuracy improvement (Tab. 1 and Fig. 3). This improvement is robust to variations of both network depth (Tab. 1) and channel scaling (dashed lines in Fig. 3). Learning Winograd transformations during training allows to + +Table 1: The accuracy $(\%)$ of 8-bit $F4-$ and $F6$ -based ResNet- $N$ models (chan. mult. is 1.0) on CIFAR-10 when $\Theta$ is static. The best results are shown in bold. The values in parentheses give the improvement (to the first decimal place) over $\Theta^{\mathrm{std}}$ case, and the results for the $F4$ ResNet-18 are averaged over the experiments of Fig. 2. + +
ModelF4 Static ΘF6 Static Θ
Standard ΘOur Θpso (↑)Standard ΘOur Θpso (↑)
ResNet-1883.6591.54 (+7.9)83.2387.51 (+4.3)
ResNet-3484.2290.48 (+6.3)69.4887.09 (+17.6)
ResNet-5082.5791.52 (+8.9)70.1582.87 (+12.7)
+ +![](images/5aee53d945fd2fa51b8f8a4fa116659dfebac915d0ddc6f25a62e1d80a153dda.jpg) +Fig. 3: The accuracy of the ResNet-18 on CIFAR-10 with various channel multipliers utilizing full 8-bit $F4$ (a) and $F6$ (b) Winograd convolutions. The SOTA data are taken from Fig. 4 of Ref. [17]. + +further adjust them to minimize accuracy errors (see solid lines in Fig. 3, Tabs. 2 and 3). In all experiments, the initialization of the Winograd transformations when considering them as trainable parameters via our PSO-derived matrices $\Theta^{\mathrm{pso}}$ is found to be highly beneficial. One can see that this training scenario not only considerably outperforms the SOTA $\Theta^{\mathrm{std}}$ -based training approach of Ref. [17], but also either completely closes (in the case of CIFAR-10) or strongly reduces (for CIFAR-100 and ImageNet) the error gap between the $F4$ Winograd convolution and the FP direct convolution. + +Finally, Tab. 2 shows a competitive comparison of the performance of using trainable transformations starting from $\Theta = \Theta^{\mathrm{pseudo}}$ (column "Our") with the current SOTA methods. In most cases, our approach significantly outperforms the SOTA results in terms of accuracy. Note that for a fair comparison, in Tab. 2 we consider only the case of the scalar quantization and real-arithmetic full 8-bit quantized Winograd algorithm. However, our method is compatible with more + +Table 2: An accuracy (%) comparison of Winograd-aware quantization schemes. + +
Model & DatasetMethodFmRef.Ours
CIFAR-10ResNet-18Winograd-aware [17]F492.4693.26
Winograd-aware [17]F687.7489.57
ResNet-18 (mult. 0.5)Legendre [6]F491.892.53
ResNet-18 (mult. 0.25)Legendre [6]F489.790.29
ResNet-20WinoTrain [35]F489.792.31
WinoVidiVici [34]90.89
ResNet-18 & ImageNetWinoTrain [35]F461.2869.15
WinoVidiVici [34]69.14
+ +Table 3: Performance of the full 8-bit quantized $F4-$ and $F6$ -based models (channel multiplier is 1.0) on the CIFAR-10 dataset when $\Theta$ is trainable. Note that here SOTA refers to the results of the training pipeline of Ref. [17] obtained using the official code released by the authors. The abbreviation FP stands for the results of the full precision direct convolution models. Other notations are the same as in Tab. 1. + +
Task (val.)Model (Test set)Direct conv.8-bit Fm with trainable Θ
FP8-bitFmSOTAOur Θpso (↑)
Classification (Accuracy, %)ResNet-18 (CIFAR-10)93.1793.20F492.5093.26 (+0.8)
F687.7489.57 (+1.8)
ResNet-18 (CIFAR-100)74.6074.18F472.3874.10 (+1.7)
F666.1572.90 (+6.7)
ResNet-18 (ImageNet)69.7969.61F467.0569.15 (+2.1)
F653.9062.87 (+9.0)
ResNet-34 (CIFAR-10)93.4093.12F492.5493.29 (+0.7)
F680.0590.16 (+10.1)
ResNet-50 (CIFAR-10)93.3593.22F493.4693.68 (+0.2)
F686.3792.16 (+5.8)
×3 Super-resolution (PSNR, dB)ESPCN [41] (Set5)32.5332.38F430.8632.11 (+1.2)
F627.3329.82 (+2.5)
ESPCN [41] (Set14)29.2129.09F428.0528.87 (+0.8)
F625.6227.30 (+1.7)
ESPCN-6 (Set5)33.2133.12F430.4532.98 (+2.5)
F625.5330.16 (+4.6)
ESPCN-6 (Set14)29.5929.50F427.7529.44 (+1.7)
F624.0227.50 (+3.5)
RFDN [29] (Set5)34.1333.68F430.6733.36 (+2.7)
RFDN [29] (Set14)30.1729.81F427.9429.62 (+1.7)
Segm IoUENet [38] (CamVid)60.6160.17F456.1260.00 (+3.9)
F652.4356.39 (+4.0)
+ +sophisticated quantization schemes (e.g., the tapwise quantization of Ref. [5]) and further performance improvement is expected from their application. + +# 4.2 Super-Resolution + +Here we examine the performance on super-resolution (SR) tasks considering the original and modified ESPCN [41] as well as the RFDN [29] models. We set the upscaling factor between low- and high-resolution images to $\times 3$ . Since the SOTA approach of Ref. [17] (training $\Theta$ starting from $\Theta^{\mathrm{std}}$ ) has only been + +implemented for classification scenarios, we extend it to SR networks based on its official implementation. All models are trained on the DIV2K dataset [1] and their performance is evaluated on the Set5 and Set14 benchmark data (more details on the SR architectures, training, and dataset preparation can be found in the Supplemental Material). As seen from the previous discussion, treating the Winograd transformations $\Theta$ as trainable parameters is essential to overcome the quality degradation caused by quantization errors. For this reason, in the following, we focus only on the case of trainable $\Theta$ . The mean PSNR values of the considered models on the benchmark datasets are summarized in Tab. 3. + +The ESPCN model is the less-depth model under consideration. It consists of three convolutional layers followed by a pixel-shuffle operation. In our implementation, the last two convolution layers are replaced by either $F4$ or $F6$ Winograd convolutions and are quantized to 8-bit. On F4, training of the transformations from $\Theta = \Theta^{\mathrm{pso}}$ gains 1.25dB and 0.82dB higher PSNR than the SOTA approach of Ref. [17] on Set5 and Set14, respectively. On $F6$ , the PSNR gap from the FP direct model is reduced by about 2 times within our training approach. + +It can be expected that proper initialization of $\Theta$ at the beginning of training is less critical for such lightweight networks. Indeed, due to the relatively small number of the transformation parameters to train, it is possible that even a local gradient-based optimization during network training may be enough to learn appropriate Winograd transformations. Thus, quality degradation when using standard matrices $\Theta^{\mathrm{std}}$ is expected to be more pronounced with increas + +![](images/5e25e33bcca89537b62ce751c595f19af0866db4c9bb77c9986e84c2a8ddf1bf.jpg) + +![](images/7052acbee1c77cd69d0a461282920f1b300e089aef6e06bb2dda81c3cc1e7dd3.jpg) +(a) Direct / FP +Fig. 4: $\times 3$ Super-resolution results of the RFDN model utilizing direct and $F4$ Winograd convolutions for "Comic" from Set14. + +![](images/2c0651d8c3112ab1aa47b80b5537ae263bc7bc3ab0ad8e44ca39748514de8a27.jpg) + +![](images/5ad8d492dd2d0066425d9e50aa35e4acbe2a305bf33c67792ed711098bc4563c.jpg) +(b) Direct / 8-bit + +![](images/13ad549d98acad5ed397b442e05347e1c1a0f9ec5a015cfdbb08f78a7434c288.jpg) + +![](images/803af0da1a2da6bfae6a370334ad5d29ac5847e1c250e2f9cc8d165f9a807b5d.jpg) +(c) $F4$ SOTA / 8-bit + +![](images/585a94bc3b422de38b8f831e7b222de8c8f040bde7f99883a62e7a251325eaf1.jpg) + +![](images/ab0afd774726e721cf3842af3a7d9ee959b3ac3f3af6f140704a8024530c3cf3.jpg) +(d) $F4$ Ours / 8-bit + +ing network depth, as in the case of the classification tasks (see Sec. 4.1). To prove this, we consider the six-layered ESPCN-6 model, which is obtained by including three additional convolution layers in the ESPCN model. While the PSNR values obtained by the $\Theta^{\mathrm{pseudo}}$ -based training increase compared to the results of the original ESPCN model, the SOTA training approach, using $\Theta^{\mathrm{std}}$ , on the contrary, tends to PSNR degradation. On $F4$ , $\Theta^{\mathrm{pseudo}}$ -based training shows an improvement of more than 2.5 dB for Set5 and 1.6 dB for Set14 relative to the SOTA training approach. With this improvement, the quality drop when using the 8-bit quantized $F4$ algorithm is strongly reduced, and the resulting quality is close to that of FP direct convolutions. + +Finally, we perform experiments with the residual feature distillation network (RFDN) [29]. It is based on the revised information distillation mechanism of the IMDN network [21] and the shallow residual blocks, which allows to achieve a better trade-off between model performance and complexity. Utilizing $\Theta^{\mathrm{pseudo}}$ leads to remarkably better performance compared to the SOTA training based on the standard matrices $\Theta^{\mathrm{std}}$ . However, we find that the RFDN model is quite sensitive to quantization noise. This results in a noticeable drop in the quality when the model is quantized to 8-bit, even in the case of direct convolutions. When the Winograd algorithm is used, this quality drop is somewhat worsened. One of the reasons for this is that the quantization scheme adapted from Ref. [17] includes redundant quantization steps (see discussion in Sec. 3.1). We also revealed a strong effect of replacing the $3\times 3$ direct convolutions nearest to the contrast aware channel attention (CCA) layer [21, 29] with the $F4$ convolutions on the enhancement of the PSNR drop. The latter requires searching for a more careful strategy to replace direct convolutions with the Winograd ones, and, a possible solution is to use for some layers the $F2$ Winograd algorithm, which is not prone to strong quantization errors. Despite the above-mentioned drop of PSNR, our method produces super-resolution images that are visually competitive with ones + +![](images/a23a3e3e6baa137235ca21618f69e4336c2bbe7199d6404c0e8d24c7c4abf8f4.jpg) +(a) ESPCN F6 SOTA +Fig. 5: $\times 3$ Super-resolution results for "Comic" from Set14 within 8-bit ESPCN (first two panels) and ESPCN-6 (last two panels) model using the $F6$ Winograd convolutions. + +![](images/c4473c1fd6faffe3c7aadd45833277d65805669eb4cb92424e7e2ac38a0484c9.jpg) +(b) ESPCN F6 Our + +![](images/c44f1fb0b683444595771d76de3a7dc328b4bfb4f5a3d52e19153c0923bda86d.jpg) +(c) ESPCN-6 F6 SOTA + +![](images/96dc48826f75214e6f0689587632e384ac07f8d691bd12de2172c9ccbf18103a.jpg) +(d) ESPCN-6 F6 Ours + +of the FP model (see Fig. 4) and are free of strong noise or gridding artifacts that typically occur in the case of standard Winograd transformations. + +For $F6$ , despite the considerable improvements achieved by using the transformations $\Theta^{\mathrm{pso}}$ (see Tab. 3 and Fig. 5), the quality drop is still pronounced. In particular, we find that the training of the RFDN model is unstable for $F6$ (for this reason, the corresponding results are not reported here). However, we emphasize that here we considered the extreme case when all direct convolutions of kernel size $3 \times 3$ are replaced by the $F6$ convolutions. We believe it is possible to selectively replace error-tolerant direct convolutions with $F6$ ones based on our $\Theta^{\mathrm{pso}}$ with minimal quality losses. This can be achieved, for example, by adapting frameworks for automatic design of neural network architectures [10, 15, 17, 46]. + +# 4.3 Segmentation + +Finally, we briefly discuss segmentation tasks. To this end, we consider the ENet model [38] which is the optimized semantic segmentation network designed to run on mobile devices. We replace all $3 \times 3$ kernel direct convolutions with the Winograd convolutions, except ones for the last and first layers, and convolutions with a stride equal to 2. The latter is because the standard Winograd algorithm cannot be directly applied to convolutions with non-unit stride. However, we note that several adaptations of the Winograd algorithm to non-unit stride cases have recently been proposed [20,37,48]. Their quantization-friendly formulation within an optimization procedure similar to the one developed in the present study is of great interest for further research. The results of training the ENet model on the CamVid dataset [9] are shown in Tab. 3. As for SR tasks, we find that our $\Theta^{\mathrm{pso}}$ -based approach is considerably more accurate compared to the SOTA approach utilizing the standard Winograd transformations $\Theta^{\mathrm{std}}$ . + +# 5 Conclusions + +We have introduced an efficient way of training 8-bit quantized networks utilizing $F4$ and $F6$ Winograd convolutions. To this end, using the global partial swarm optimization (PSO) method, we have derived a model- and data-free form of the Winograd transformation matrices $\Theta \equiv \{A,B,G\}$ , which has significantly better accuracy performance compared to the standard ones. + +To further improve accuracy we have modified the previous SOTA training pipeline of Ref. [17], which allows the transformation matrices $\Theta$ to evolve during training, by replacing the standard Winograd transformations used in this pipeline with our PSO-derived matrices. By treating all the steps of Winograd convolution as 8-bit quantized, we have considered classification, superresolution, and segmentation tasks to reveal the benefits of the proposed approach. In all of the experiments, we found that our modification results in a substantial improvement of accuracy and allows to either completely overcome or highly reduce quality drop. As a consequence, our approach significantly outperforms the SOTA method in terms of accuracy, while requiring the same number of operations during network training and retaining the same inference speedup. + +# References + +1. Agustsson, E., Timofte, R.: Ntire 2017 challenge on single image super-resolution: Dataset and study. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 126-135 (2017) +2. Alam, M.S., Misba, W.A., Atulasimha, J.: Quantized non-volatile nanomagnetic synapse based autoencoder for efficient unsupervised network anomaly detection. arXiv preprint arXiv:2309.06449 (2023) +3. Alam, S.A., Anderson, A., Barabasz, B., Gregg, D.: Winograd convolution for deep neural networks: Efficient point selection. ACM Transactions on Embedded Computing Systems 21(6), 1-28 (2022) +4. Anderson, A., Gregg, D.: Optimal dnn primitive selection with partitioned boolean quadratic programming. In: Proceedings of the 2018 International Symposium on Code Generation and Optimization. pp. 340-351 (2018) +5. Andri, R., Bussolino, B., Cipolletta, A., Cavigelli, L., Wang, Z.: Going further with winograd convolutions: Tap-wise quantization for efficient inference on 4x4 tiles. In: 2022 55th IEEE/ACM International Symposium on Microarchitecture (MICRO). pp. 582-598. IEEE (2022) +6. Barabasz, B.: Quantaized winograd/toom-cook convolution for dnns: Beyond canonical polynomials base. arXiv preprint arXiv:2004.11077 (2020) +7. Barabasz, B., Anderson, A., Soodhalter, K.M., Gregg, D.: Error analysis and improving the accuracy of winograd convolution for deep neural networks. ACM Transactions on Mathematical Software (TOMS) 46(4), 1-33 (2020) +8. Barabasz, B., Gregg, D.: Winograd convolution for dnns: Beyond linear polynomials. In: International Conference of the Italian Association for Artificial Intelligence. pp. 307-320. Springer (2019) +9. Brostow, G.J., Fauqueur, J., Cipolla, R.: Semantic object classes in video: A high-definition ground truth database. Pattern Recognition Letters 30(2), 88-97 (2009) +10. Cai, H., Zhu, L., Han, S.: Proxylessnas: Direct neural architecture search on target task and hardware. arXiv preprint arXiv:1812.00332 (2018) +1. Chen, T., Xu, W., Chen, W., Wang, P., Cheng, J.: Towards efficient and accurate winograd convolution via full quantization. Advances in Neural Information Processing Systems 36 (2024) +2. Chen, Y.T., Ou, Y.F., Huang, C.T.: A winograd-based highly-parallel convolution engine for 8-bit cnn acceleration. In: 2022 IEEE 4th International Conference on Artificial Intelligence Circuits and Systems (AICAS). pp. 395–398. IEEE (2022) +3. Chen, Y.H., Krishna, T., Emer, J.S., Sze, V.: Eyeriss: An energy-efficient reconfigurable accelerator for deep convolutional neural networks. IEEE journal of solid-state circuits 52(1), 127-138 (2016) +4. Cook, S.: On the minimum computation time for multiplication. Doctoral diss., Harvard U., Cambridge, Mass 1 (1966) +5. Elsken, T., Metzen, J.H., Hutter, F.: Neural architecture search: A survey. The Journal of Machine Learning Research 20(1), 1997-2017 (2019) +6. Fernández-Marqués, J.: High performance deep learning on resource constrained platforms. Ph.D. thesis, University of Oxford (2021) +7. Fernandez-Marques, J., Whatmough, P., Mundy, A., Mattina, M.: Searching for winograd-aware quantized networks. Proceedings of Machine Learning and Systems 2, 14-29 (2020) +8. Gholami, A., Kim, S., Dong, Z., Yao, Z., Mahoney, M.W., Keutzer, K.: A survey of quantization methods for efficient neural network inference. In: Low-Power Computer Vision, pp. 291-326. Chapman and Hall/CRC (2022) + +19. Holt, J.L., Hwang, J.N.: Finite precision error analysis of neural network hardware implementations. IEEE Transactions on Computers 42(3), 281-290 (1993) +20. Huang, C., Dong, X., Li, Z., Song, T., Liu, Z., Dong, L.: Efficient stride 2 winograd convolution method using unified transformation matrices on fpga. In: 2021 International Conference on Field-Programmable Technology (ICFPT). pp. 1-9. IEEE (2021) +21. Hui, Z., Gao, X., Yang, Y., Wang, X.: Lightweight image super-resolution with information multi-distillation network. In: Proceedings of the 27th acm international conference on multimedia. pp. 2024-2032 (2019) +22. Jorda, M., Valero-Lara, P., Pena, A.J.: Performance evaluation of cudnn convolution algorithms on nvidia volta gpus. IEEE Access 7, 70461-70473 (2019) +23. Kennedy, J., Eberhart, R.: Particle swarm optimization. In: Proceedings of ICNN'95-international conference on neural networks. vol. 4, pp. 1942-1948. IEEE (1995) +24. Krishnamoorthi, R.: Quantizing deep convolutional networks for efficient inference: A whitepaper. arXiv preprint arXiv:1806.08342 (2018) +25. Lavin, A., Gray, S.: Fast algorithms for convolutional neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4013-4021 (2016) +26. Li, C., Jiang, P., Zhou, H., Wang, X., Zhao, X.: Haw: Hardware-aware point selection for efficient winograd convolution. IEEE Signal Processing Letters 30, 269-273 (2023) +27. Li, G., Jia, Z., Feng, X., Wang, Y.: Lowino: Towards efficient low-precision winograd convolutions on modern cpus. In: Proceedings of the 50th International Conference on Parallel Processing. pp. 1-11 (2021) +28. Li, G., Liu, L., Wang, X., Ma, X., Feng, X.: Lance: efficient low-precision quantized winograd convolution for neural networks based on graphics processing units. In: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 3842-3846. IEEE (2020) +29. Liu, J., Tang, J., Wu, G.: Residual feature distillation network for lightweight image super-resolution. In: Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 41-55. Springer (2020) +30. Liu, Z.G., Mattina, M.: Efficient residue number system based winograd convolution. In: European Conference on Computer Vision. pp. 53-68. Springer (2020) +31. Mathieu, M., Henaff, M., LeCun, Y.: Fast training of convolutional networks through fts. arXiv preprint arXiv:1312.5851 (2013) +32. Meng, L., Brothers, J.: Efficient winograd convolution via integer arithmetic. arXiv preprint arXiv:1901.01965 (2019) +33. Miranda, L.J.: Pyswarms: a research toolkit for particle swarm optimization in python. Journal of Open Source Software 3(21), 433 (2018) +34. Mori, P., Frickenstein, L., Sampath, S.B., Thoma, M., Faspous, N., Vemparala, M.R., Frickenstein, A., Unger, C., Stechele, W., Mueller-Gritschneider, D., et al.: Wino vidi vici: Conquering numerical instability of 8-bit winograd convolution for accurate inference acceleration on edge. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 53-62 (2024) +35. Mori, P., Sampath, S.B., Frickenstein, L., Vemparala, M.R., Faspous, N., Frickenstein, A., Stechele, W., Passerone, C.: Winotrain: Winograd-aware training for accurate full 8-bit convolution acceleration. In: 2023 60th ACM/IEEE Design Automation Conference (DAC). pp. 1-6. IEEE (2023) + +36. Nagel, M., Fournarakis, M., Amjad, R.A., Bondarenko, Y., Van Baalen, M., Blankevoort, T.: A white paper on neural network quantization. arXiv preprint arXiv:2106.08295 (2021) +37. Pan, J., Chen, D.: Accelerate non-unit stride convolutions with winograd algorithms. In: Proceedings of the 26th Asia and South Pacific Design Automation Conference. pp. 358-364 (2021) +38. Paszke, A., Chaurasia, A., Kim, S., Culurciello, E.: Enet: A deep neural network architecture for real-time semantic segmentation. arXiv preprint arXiv:1606.02147 (2016) +39. Shafique, M.A., Munir, A., Kong, J.: Deep learning performance characterization on gpus for various quantization frameworks. AI 4(4), 926-948 (2023) +40. Shi, B., Zhang, J., He, Z., Wei, X., Li, S., Luo, G., Zheng, H., Xie, Y.: Efficient super-resolution system with block-wise hybridization and quantized winograd on fpga. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems (2023) +41. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1874-1883 (2016) +42. Toom, A.L.: The complexity of a scheme of functional elements simulating the multiplication of integers. In: Doklady Akademii Nauk. vol. 150, pp. 496-498. Russian Academy of Sciences (1963) +43. Vasilache, N., Johnson, J., Mathieu, M., Chintala, S., Piantino, S., LeCun, Y.: Fast convolutional nets with fbfft: Agpu performance evaluation. arXiv preprint arXiv:1412.7580 (2014) +44. Vincent, K., Stephano, K., Frumkin, M., Ginsburg, B., Demouth, J.: On improving the numerical stability of winograd convolutions (2017) +45. Wang, D., Tan, D., Liu, L.: Particle swarm optimization algorithm: an overview. Soft computing 22, 387-408 (2018) +46. Wang, K., Liu, Z., Lin, Y., Lin, J., Han, S.: Haq: Hardware-aware automated quantization with mixed precision. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8612-8620 (2019) +47. Winograd, S.: Arithmetic complexity of computations, vol. 33. Siam (1980) +48. Yepez, J., Ko, S.B.: Stride 2 1-d, 2-d, and 3-d winograd for convolutional neural networks. IEEE Transactions on Very Large Scale Integration (VLSI) Systems 28(4), 853-863 (2020) +49. Zhang, Y., Wang, S., Ji, G., et al.: A comprehensive survey on particle swarm optimization algorithm and its applications. Mathematical problems in engineering 2015 (2015) \ No newline at end of file diff --git a/2024/Quantization-Friendly Winograd Transformations for Convolutional Neural Networks/images.zip b/2024/Quantization-Friendly Winograd Transformations for Convolutional Neural Networks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..34570355038e5169517c46f6ff2614a8a1bb0310 --- /dev/null +++ b/2024/Quantization-Friendly Winograd Transformations for Convolutional Neural Networks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0731e0d10d50eeb3523b8f77123c6ed5452b18bc528b5ebc01c01d27c7ce7422 +size 573607 diff --git a/2024/Quantization-Friendly Winograd Transformations for Convolutional Neural Networks/layout.json b/2024/Quantization-Friendly Winograd Transformations for Convolutional Neural Networks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0b3a2bbb5a88bccc583a8e23438c99814e91209f --- /dev/null +++ b/2024/Quantization-Friendly Winograd Transformations for Convolutional Neural Networks/layout.json @@ -0,0 +1,11053 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 111, + 479, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 111, + 479, + 147 + ], + "spans": [ + { + "bbox": [ + 134, + 111, + 479, + 147 + ], + "type": "text", + "content": "Quantization-Friendly Winograd Transformations for Convolutional Neural Networks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 140, + 168, + 473, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 168, + 473, + 181 + ], + "spans": [ + { + "bbox": [ + 140, + 168, + 473, + 181 + ], + "type": "text", + "content": "Vladimir Protsenko, Vladimir Kryzhanovskiy, and Alexander Filippov" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 154, + 190, + 457, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 190, + 457, + 224 + ], + "spans": [ + { + "bbox": [ + 154, + 190, + 457, + 224 + ], + "type": "text", + "content": "Huawei Noah's Ark Lab \nprotsenko.vladimir@huawei-partners.com, {kryzhanovskiy.vladimir, filippov alexander}@huawei.com" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 255, + 455, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 255, + 455, + 452 + ], + "spans": [ + { + "bbox": [ + 160, + 255, + 455, + 452 + ], + "type": "text", + "content": "Abstract. Efficient deployment of modern deep convolutional neural networks on resource-constrained devices suffers from demanding computational requirements of convolution operations. Quantization and use of Winograd convolutions operating on sufficiently large-tile inputs are two powerful strategies to speed up convolution operations. However, their combination results in numerical instability, which manifests itself in a strong quality performance degradation. We present an efficient learning scenario that either completely overcomes or strongly reduces the accuracy degradation of full 8-bit quantized " + }, + { + "bbox": [ + 160, + 255, + 455, + 452 + ], + "type": "inline_equation", + "content": "F(4,3)" + }, + { + "bbox": [ + 160, + 255, + 455, + 452 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 160, + 255, + 455, + 452 + ], + "type": "inline_equation", + "content": "F(6,3)" + }, + { + "bbox": [ + 160, + 255, + 455, + 452 + ], + "type": "text", + "content": " Winograd convolutions. Within the global particle swarm optimization (PSO), we derived a set of quantization-friendly Winograd transformations. Following the state-of-the-art (SOTA) training pipeline [J. Fernandez-Marques et al., Proc. Mach. Learn. Syst. 2, 14-29 (2020)], we treat Winograd transformations as learnable parameters during network training. Evolving transformations starting from our PSO-derived ones rather than the standard Winograd transformations results in significant numerical error reduction and accuracy improvement. As a consequence, our approach significantly outperforms SOTA methods on various tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 463, + 453, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 463, + 453, + 475 + ], + "spans": [ + { + "bbox": [ + 160, + 463, + 453, + 475 + ], + "type": "text", + "content": "Keywords: Winograd convolutions " + }, + { + "bbox": [ + 160, + 463, + 453, + 475 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 463, + 453, + 475 + ], + "type": "text", + "content": " Quantization " + }, + { + "bbox": [ + 160, + 463, + 453, + 475 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 463, + 453, + 475 + ], + "type": "text", + "content": " Efficient inference" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 495, + 230, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 495, + 230, + 509 + ], + "spans": [ + { + "bbox": [ + 132, + 495, + 230, + 509 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 629 + ], + "type": "text", + "content": "To date, deep convolutional neural networks (CNN) provide state-of-the-art performance in a wide variety of machine learning domains, including image processing, computer vision, video-based workflow analysis tasks, etc. The convolution operations, being the core block of CNNs, represent their most time-consuming and computationally intensive part [13]. Since modern CNNs typically include deep stacks of convolution layers both their training and inference are highly computationally demanding. As a consequence, despite recent advances in hardware technology, efficient inference of modern CNNs on resource-constrained mobile and embedded devices remains to be a challenging task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "With the recent explosion of support for low-precision arithmetic in modern hardware, quantization [18, 24, 36] has become one of the standard approaches for compression and acceleration of CNNs. The quantization process is defined" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "by mapping the network parameters to their low-precision representation. This allows to operate within low-bit arithmetic, having significantly lower computational and storage requirements than the full precision counterpart. Quantization to 8-bit integers generally produces only a moderate accuracy degradation and is widely acceptable in mobile devices." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 180, + 483, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 180, + 483, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 180, + 483, + 323 + ], + "type": "text", + "content": "An alternative way to achieve efficient inference is to speed up convolution operations at the algorithm level. Several efficient convolution techniques have been developed, including the fast Fourier transform (FFT) [31, 43] and Winograd minimal filtering algorithm [25, 47]. The latter is known to minimize computational costs for convolutions with kernels of sizes " + }, + { + "bbox": [ + 130, + 180, + 483, + 323 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 180, + 483, + 323 + ], + "type": "text", + "content": " [4, 22], which are most often used in modern CNNs. The " + }, + { + "bbox": [ + 130, + 180, + 483, + 323 + ], + "type": "inline_equation", + "content": "F(m, r)" + }, + { + "bbox": [ + 130, + 180, + 483, + 323 + ], + "type": "text", + "content": " Winograd algorithm transforms tiles of input feature map and " + }, + { + "bbox": [ + 130, + 180, + 483, + 323 + ], + "type": "inline_equation", + "content": "r \\times r" + }, + { + "bbox": [ + 130, + 180, + 483, + 323 + ], + "type": "text", + "content": " filter into the Winograd domain, where " + }, + { + "bbox": [ + 130, + 180, + 483, + 323 + ], + "type": "inline_equation", + "content": "m^2" + }, + { + "bbox": [ + 130, + 180, + 483, + 323 + ], + "type": "text", + "content": " outputs can be simultaneously computed with a single element-wise matrix multiplication. Contrary to the FFT, the original Winograd algorithm does not involve complex number multiplications, which are at least " + }, + { + "bbox": [ + 130, + 180, + 483, + 323 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 130, + 180, + 483, + 323 + ], + "type": "text", + "content": " computationally costly compared to that of real numbers. Because of its cutting-edge performance, the Winograd algorithm is integrated into various deep-learning libraries." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 327, + 483, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 483, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 483, + 483 + ], + "type": "text", + "content": "As the quantization and Winograd algorithms are not mutually exclusive but complementary, their mixing is attracting considerable attention from both academic and industrial fields. Theoretically, the larger the tile size " + }, + { + "bbox": [ + 130, + 327, + 483, + 483 + ], + "type": "inline_equation", + "content": "k_{\\mathrm{W}} = m + r - 1" + }, + { + "bbox": [ + 130, + 327, + 483, + 483 + ], + "type": "text", + "content": " of input feature maps, the more speedup can be expected from using the " + }, + { + "bbox": [ + 130, + 327, + 483, + 483 + ], + "type": "inline_equation", + "content": "F(m,r)" + }, + { + "bbox": [ + 130, + 327, + 483, + 483 + ], + "type": "text", + "content": " Winograd algorithm. However, in practice, large-tile Winograd algorithms turn out to be highly sensitive to numerical errors, which grow exponentially with tile size [7,8]. Quantization, lowering arithmetic precision, further amplifies numerical errors [19] and causes numerical instability of the Winograd algorithm, which results in severe degradation of network performance. This strongly inhibits the industrial use of large-tile (" + }, + { + "bbox": [ + 130, + 327, + 483, + 483 + ], + "type": "inline_equation", + "content": "k_{\\mathrm{W}} \\gtrsim 4" + }, + { + "bbox": [ + 130, + 327, + 483, + 483 + ], + "type": "text", + "content": ") low-bit Winograd-based convolutions, including the perspective " + }, + { + "bbox": [ + 130, + 327, + 483, + 483 + ], + "type": "inline_equation", + "content": "F(4,3)" + }, + { + "bbox": [ + 130, + 327, + 483, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 327, + 483, + 483 + ], + "type": "inline_equation", + "content": "F(6,3)" + }, + { + "bbox": [ + 130, + 327, + 483, + 483 + ], + "type": "text", + "content": ". Despite all efforts to mitigate this problem (see Sec. 2), the performance of large-tile Winograd algorithms quantized to the widely supported 8-bit integer format remains far from satisfactory." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 486, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 483, + 666 + ], + "type": "text", + "content": "In this paper, we present an efficient learning scenario that either completely overcomes or strongly reduces the accuracy degradation of full 8-bit quantized " + }, + { + "bbox": [ + 130, + 486, + 483, + 666 + ], + "type": "inline_equation", + "content": "F(4,3)" + }, + { + "bbox": [ + 130, + 486, + 483, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 486, + 483, + 666 + ], + "type": "inline_equation", + "content": "F(6,3)" + }, + { + "bbox": [ + 130, + 486, + 483, + 666 + ], + "type": "text", + "content": " Winograd convolutions. As a first step, we perform a global particle swarm optimization (PSO) search over the general family of Winograd algorithms, which are defined by specifying a set of transformation matrices " + }, + { + "bbox": [ + 130, + 486, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\Theta \\equiv \\{A,B,G\\}" + }, + { + "bbox": [ + 130, + 486, + 483, + 666 + ], + "type": "text", + "content": ", to select one that is more robust to quantization errors. This step is prior to network training and yields a model- and data-free set of quantization-friendly Winograd transformation matrices " + }, + { + "bbox": [ + 130, + 486, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 486, + 483, + 666 + ], + "type": "text", + "content": ". We demonstrate that replacing the standard Winograd transformation matrices with our PSO-derived ones already results in significant accuracy improvements. To further tackle the performance degradation challenge, we integrated our transformations into the SOTA learning pipeline of Ref. [17], which treats Winograd transformation matrices as trainable parameters during network training. We show that allowing the evolution of Winograd transformations " + }, + { + "bbox": [ + 130, + 486, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 486, + 483, + 666 + ], + "type": "text", + "content": " starting from our PSO-derived matrices, rather than from standard Winograd matrices as in the original pipeline of" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 321, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 321, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 321, + 102 + ], + "type": "text", + "content": "V. Protsenko, V. Kryzhanovskiy et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "Ref. [17], has a direct impact on increasing networks performance. As a result, our learning strategy allows to outperform existing approaches (and the 8-bit SOTA results) by demonstrating superior accuracy in various tasks, including classification, super-resolution, and semantic segmentation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 183, + 243, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 243, + 197 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 243, + 197 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 210, + 482, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 210, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 210, + 482, + 270 + ], + "type": "text", + "content": "The Toom-Cook minimal filtering algorithms [14, 42], generalized by Winograd [47], were first applied to CNNs by Lavin and Gray [25] to accelerate convolutions with " + }, + { + "bbox": [ + 130, + 210, + 482, + 270 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 210, + 482, + 270 + ], + "type": "text", + "content": " filters. Barabasz et al. [7] presented a theoretical analysis of the accuracy of Winograd algorithms and revealed that the error bound for the Toom-Cook algorithm grows exponentially with tile size." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 270, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 449 + ], + "type": "text", + "content": "To reduce poor numerical properties of Winograd convolutions, approaches to a more efficient selection of polynomial points have been proposed [3, 7, 26]. While most of them have empirical grounds, Ref. [26] leverages reinforcement learning to construct an optimal polynomial point selection policy. This reinforcement learning-based framework significantly improves the accuracies of quantized " + }, + { + "bbox": [ + 130, + 270, + 482, + 449 + ], + "type": "inline_equation", + "content": "F(m, 3)" + }, + { + "bbox": [ + 130, + 270, + 482, + 449 + ], + "type": "text", + "content": " Winograd convolutions with " + }, + { + "bbox": [ + 130, + 270, + 482, + 449 + ], + "type": "inline_equation", + "content": "m \\leq 5" + }, + { + "bbox": [ + 130, + 270, + 482, + 449 + ], + "type": "text", + "content": " and leads to only a small accuracy gap between Winograd convolutions and the direct convolutions. However, the performance benefits of this approach are only demonstrated for mixed 8/16-bit quantization, which is a significantly softer restriction compared to considering all Winograd convolution steps quantized below 8-bit. The accuracy improvement with mixed-bit precision or quantization only in the Winograd domain has been shown in [5, 27, 28, 40]. For quantized networks, Meng and Brothers [32] extended the Winograd algorithm to the complex number domain. But, the shift to complex transformations substantially reduces the multiplication saving and hence the speedup of the Winograd algorithm." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 450, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 617 + ], + "type": "text", + "content": "Fernandez-Marques et al. [16, 17] first introduced the idea of considering the Winograd transformation matrices as learnable parameters while training quantized networks. Including the Winograd transformations in the training loop introduces additional degrees of freedom that are aware of numerical errors and can be adjusted to reduce them. On " + }, + { + "bbox": [ + 130, + 450, + 482, + 617 + ], + "type": "inline_equation", + "content": "F(m \\leq 6,3)" + }, + { + "bbox": [ + 130, + 450, + 482, + 617 + ], + "type": "text", + "content": ", it was demonstrated that training the Winograd transformation matrices, as opposed to keeping them static (non-trainable), causes a dramatic reduction of the accuracy gap between 8-bit quantized Winograd and direct convolutions. It is important to emphasize that, despite it is not clearly stated in the paper by Fernandez-Marques et al., all the parts of Winograd convolutions (including the Winograd transformation matrices) are maintained in 8-bit precision. This is clear from the official source code provided by the authors (https://github.com/jafermarq/WinogradAwareNets). We additionally verified that indeed the above official implementation strictly reproduces all the results of the corresponding study." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Despite the significant accuracy improvements achieved by Ref. [17], there are still accuracy losses for the perspective " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "F(4,3)" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "F(6,3)" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " algorithms, which hinder their industrial potential. The study [6] continues the work of [17] and applies a base change technique for quantized Winograd convolutions. This tech-" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 208, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 208, + 90, + 447, + 102 + ], + "type": "text", + "content": "Quantization-Friendly Winograd Transformations for CNN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "nique was tested for " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "F(4,3)" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " ResNet-18 model on CIFAR-10 dataset. It is shown that transformations into Legendre polynomials base for 8-bit quantized trainable Winograd convolutions further reduce the errors of the original pipeline of Ref. [17]. However, the proposed approach can achieve the same accuracy as for direct convolutions only when the element-wise (Hadamard) product is extended to 9-bit. These improvements after modification of Winograd matrices yield the important hint that the selection of the appropriate starting values for these matrices in the training pipeline of Ref. [17] is of critical importance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 482, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 343 + ], + "type": "text", + "content": "Besides the works [6,17], there are other studies on the use of full 8-bit quantized Winograd convolutions [11, 12, 30, 34, 35]. Each of these studies, however, compared to the approach of Ref. [17] comes with its own limitations on either performance or resulting acceleration. Ref. [35] performs quantization-aware training (QAT) to mitigate the numerical error of quantized Winograd inference. While preserving multiplication savings of the Winograd algorithm, this strategy loses in performance relative to Ref. [17]. Ref. [30] adopts a residue number system to eliminate accuracy degradation, but as a consequence, the theoretical multiplication savings are dramatically reduced by this approach. Ref. [12] (and the most accurate scheme of Ref. [34]) uses complex numbers based Winograd algorithms, which heavily reduces the speedup." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 343, + 482, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 439 + ], + "type": "text", + "content": "In light of the above, the training scheme of Ref. [17] remains the SOTA in terms of tackling the problem of accuracy degradation while preserving the speedup of the standard Winograd algorithm. In the following, we refer to the pipeline of Ref. [17] as the SOTA approach. We emphasize that by this we mean only the case of full 8-bit quantization and real number arithmetic of the Winograd convolutions, since mixed-precision quantization has the potential to achieve better performance (e.g., see [26] for the SOTA results for mix 8/16-bit quantization)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 457, + 230, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 230, + 470 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 230, + 470 + ], + "type": "text", + "content": "3 Our Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 483, + 228, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 228, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 228, + 495 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 503, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 482, + 552 + ], + "type": "text", + "content": "Winograd Convolution. The 2D Winograd convolution " + }, + { + "bbox": [ + 130, + 503, + 482, + 552 + ], + "type": "inline_equation", + "content": "F(m, r)" + }, + { + "bbox": [ + 130, + 503, + 482, + 552 + ], + "type": "text", + "content": " computes " + }, + { + "bbox": [ + 130, + 503, + 482, + 552 + ], + "type": "inline_equation", + "content": "m \\times m" + }, + { + "bbox": [ + 130, + 503, + 482, + 552 + ], + "type": "text", + "content": " output by element-wise matrix multiplication of the Winograd domain-transformed filter " + }, + { + "bbox": [ + 130, + 503, + 482, + 552 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 503, + 482, + 552 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 130, + 503, + 482, + 552 + ], + "type": "inline_equation", + "content": "r \\times r" + }, + { + "bbox": [ + 130, + 503, + 482, + 552 + ], + "type": "text", + "content": " and input feature map " + }, + { + "bbox": [ + 130, + 503, + 482, + 552 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 503, + 482, + 552 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 130, + 503, + 482, + 552 + ], + "type": "inline_equation", + "content": "k_W \\times k_W" + }, + { + "bbox": [ + 130, + 503, + 482, + 552 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 503, + 482, + 552 + ], + "type": "inline_equation", + "content": "k_W = m + r - 1" + }, + { + "bbox": [ + 130, + 503, + 482, + 552 + ], + "type": "text", + "content": ". The whole procedure can be expressed in matrix form as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 235, + 562, + 481, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 562, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 235, + 562, + 481, + 578 + ], + "type": "interline_equation", + "content": "Y = A ^ {T} \\left[ \\left(G g G ^ {T}\\right) \\odot \\left(B ^ {T} d B\\right) \\right] A, \\tag {1}", + "image_path": "a1a5d2765c01066b290d8d8b8ffe231dab269ff465729a5a47f0b6281c6146ca.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "A \\in \\mathbb{R}^{k_W \\times m}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "B \\in \\mathbb{R}^{k_W \\times k_W}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "G \\in \\mathbb{R}^{k_W \\times r}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " are constant real Winograd transformation matrices and " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " is the Hadamard (element-wise) multiplication. For brevity, we use the notation " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Theta \\equiv \\{A, B, G\\}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " for the set of the Winograd matrices hereafter. The use of this algorithm reduces arithmetic complexity by simultaneously allowing " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "m^2" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " outputs to be computed by performing " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "(m + r - 1)^2" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " multiplications, instead of the " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "(rm)^2" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " required for a naive convolution operation. This reduction of computationally costly multiplication operations together with the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 321, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 321, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 321, + 102 + ], + "type": "text", + "content": "V. Protsenko, V. Kryzhanovskiy et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "content": "ability to compute multiple outputs results in a significant acceleration compared using of the direct convolution (e.g., theoretically " + }, + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "content": " speedup on " + }, + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "inline_equation", + "content": "F(4,3)" + }, + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "content": "). In this study, we consider " + }, + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "content": " filters, and for concise notation we denote " + }, + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "inline_equation", + "content": "F(m,r = 3)" + }, + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{m}}" + }, + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "content": " (e.g., " + }, + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "inline_equation", + "content": "F(4,3) \\equiv F4" + }, + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "content": " in our notations)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 481, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 481, + 247 + ], + "type": "text", + "content": "As can be seen from Eq. (1), the Winograd convolution is defined by specifying the transformation matrices " + }, + { + "bbox": [ + 130, + 164, + 481, + 247 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 164, + 481, + 247 + ], + "type": "text", + "content": ". The particular form of " + }, + { + "bbox": [ + 130, + 164, + 481, + 247 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 164, + 481, + 247 + ], + "type": "text", + "content": " depends on the size of the filter " + }, + { + "bbox": [ + 130, + 164, + 481, + 247 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 164, + 481, + 247 + ], + "type": "text", + "content": " and the desired number of outputs. In the case of " + }, + { + "bbox": [ + 130, + 164, + 481, + 247 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 164, + 481, + 247 + ], + "type": "text", + "content": " filters, which are widely used in modern CNNs, for " + }, + { + "bbox": [ + 130, + 164, + 481, + 247 + ], + "type": "inline_equation", + "content": "m \\leq 6" + }, + { + "bbox": [ + 130, + 164, + 481, + 247 + ], + "type": "text", + "content": " there is a general agreement on the form of the transformation matrices " + }, + { + "bbox": [ + 130, + 164, + 481, + 247 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 164, + 481, + 247 + ], + "type": "text", + "content": ". The corresponding matrices are given in [25] (see also Sec. 1 of the Supplemental Material) and we refer to this set of matrices as the standard Winograd transformations " + }, + { + "bbox": [ + 130, + 164, + 481, + 247 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{std}}" + }, + { + "bbox": [ + 130, + 164, + 481, + 247 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 247, + 481, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 481, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 481, + 283 + ], + "type": "text", + "content": "Quantization. Following Ref. [17], we quantize all steps of the Winograd algorithm (including the transformations " + }, + { + "bbox": [ + 130, + 247, + 481, + 283 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 247, + 481, + 283 + ], + "type": "text", + "content": ") to 8-bit. Therefore, the full 8-bit quantized Winograd convolution can be written as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 230, + 290, + 481, + 311 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 290, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 230, + 290, + 481, + 311 + ], + "type": "interline_equation", + "content": "Y = q _ {8} \\left(\\hat {Q} _ {8} ^ {A} \\left(\\hat {Q} _ {8} ^ {G ^ {T}} (g) \\odot \\hat {Q} _ {8} ^ {B} (d)\\right)\\right), \\tag {2}", + "image_path": "ae0054a79948517a47830e430a388ca8c5424a05375a667d00a765d8f3943353.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 318, + 482, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 482, + 427 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 482, + 427 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 318, + 482, + 427 + ], + "type": "inline_equation", + "content": "\\hat{Q}_8^A (x) = q_8(A^T)q_8(x)q_8(A)" + }, + { + "bbox": [ + 130, + 318, + 482, + 427 + ], + "type": "text", + "content": " is the transformation operator to/from the quantized Winograd domain and " + }, + { + "bbox": [ + 130, + 318, + 482, + 427 + ], + "type": "inline_equation", + "content": "q_{8}" + }, + { + "bbox": [ + 130, + 318, + 482, + 427 + ], + "type": "text", + "content": " stands for 8-bit quantization. As in [17], we use the signed symmetric uniform scalar quantization scheme [36]. Note that for a proper comparison with the SOTA pipeline of Ref. [17], as in that study, both input and output are quantized. However, this quantization granularity is clearly redundant and may result in additional loss of accuracy. Therefore, we expect that within the conventional quantization schemes when either the input or the output is quantized, the proposed in this study approach provides further accuracy improvements." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 443, + 356, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 443, + 356, + 455 + ], + "spans": [ + { + "bbox": [ + 131, + 443, + 356, + 455 + ], + "type": "text", + "content": "3.2 Quantization-Friendly Transformations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "text", + "content": "Motivation. As discussed in Refs. [25,27,34], the magnitude of matrix elements of the Winograd transformations " + }, + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Theta \\equiv \\{A,B,G\\}" + }, + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "text", + "content": " tends to increase with tile size " + }, + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "inline_equation", + "content": "k_{W}" + }, + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "text", + "content": ", which considerably broadens numerical ranges and effectively reduces computation accuracy. In [7] Barabasz et al. analytically derived the floating-point error bound for the Winograd algorithm. It is shown that the error is proportional to the norms of the Winograd transform matrices. One can naturally expect the same relation between the quantization errors and matrix elements (and norms) of the Winograd transformations " + }, + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "text", + "content": ". Indeed, the error coming from each quantization step in Eq. (2) is amplified proportionally to magnitudes of the following matrix transformations (if not taking into account the rare self errors cancellation effects). Another point is that transformation matrices with high-magnitude matrix elements considerably extend the quantization interval [27,34], whereas most of it is very sparsely used [34]. This leads to strong rounding errors during quantization [34]. Here, to improve the numerical properties of quantized Winograd convolutions, we exploit the fact that transformations " + }, + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "text", + "content": " are not uniquely defined and construct a representation having as small matrix elements as possible. We demonstrate (see Sec. 4) that the resulting set of matrices " + }, + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{pso}}" + }, + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 207, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 207, + 90, + 448, + 103 + ], + "type": "text", + "content": "Quantization-Friendly Winograd Transformations for CNN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "quantization-friendly and yields a significant improvement in accuracy compared to using the standard Winograd transformations " + }, + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{std}}" + }, + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 164 + ], + "type": "text", + "content": "PSO Optimization. To reduce the magnitudes of the Winograd transformations we consider the following optimization problem" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 164, + 174, + 481, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 174, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 164, + 174, + 481, + 194 + ], + "type": "interline_equation", + "content": "\\Theta^ {\\mathrm {p s o}} = \\underset {\\Theta} {\\arg \\min } f (\\Theta) = \\underset {\\Theta} {\\arg \\min } \\left(\\gamma_ {A} \\| A \\| _ {F} ^ {2} + \\gamma_ {B} \\| B \\| _ {F} ^ {2} + \\gamma_ {G} \\| G \\| _ {F} ^ {2}\\right), \\tag {3}", + "image_path": "bafff1f5632a9a50e4033290fa5c0277fb5a60e69e2585868296f16f4c2389bb.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 202, + 480, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 202, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 202, + 480, + 334 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 202, + 480, + 334 + ], + "type": "inline_equation", + "content": "\\gamma_A, \\gamma_B, \\gamma_G > 0" + }, + { + "bbox": [ + 130, + 202, + 480, + 334 + ], + "type": "text", + "content": " are weight factors, " + }, + { + "bbox": [ + 130, + 202, + 480, + 334 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_F" + }, + { + "bbox": [ + 130, + 202, + 480, + 334 + ], + "type": "text", + "content": " denotes the Frobenius norm of a matrix and the optimization is restricted to the set of transformations " + }, + { + "bbox": [ + 130, + 202, + 480, + 334 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 202, + 480, + 334 + ], + "type": "text", + "content": " that fulfill Eq. (1) to exactly preserve the property to be convolution operation. Although mainly the matrices " + }, + { + "bbox": [ + 130, + 202, + 480, + 334 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 202, + 480, + 334 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 202, + 480, + 334 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 202, + 480, + 334 + ], + "type": "text", + "content": " contain high-value matrix elements (see Supplemental Material Sec. 1), one should also be aware of preventing the possible growth of the matrix elements of " + }, + { + "bbox": [ + 130, + 202, + 480, + 334 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 202, + 480, + 334 + ], + "type": "text", + "content": " in a new representation. Given this " + }, + { + "bbox": [ + 130, + 202, + 480, + 334 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 202, + 480, + 334 + ], + "type": "text", + "content": " is also included in Eq. (3). To reveal main the effect of minimizing " + }, + { + "bbox": [ + 130, + 202, + 480, + 334 + ], + "type": "inline_equation", + "content": "f(\\Theta)" + }, + { + "bbox": [ + 130, + 202, + 480, + 334 + ], + "type": "text", + "content": ", we do not fine-tune the parameters " + }, + { + "bbox": [ + 130, + 202, + 480, + 334 + ], + "type": "inline_equation", + "content": "\\gamma_p" + }, + { + "bbox": [ + 130, + 202, + 480, + 334 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 202, + 480, + 334 + ], + "type": "inline_equation", + "content": "(p \\in \\{A, B, G\\})" + }, + { + "bbox": [ + 130, + 202, + 480, + 334 + ], + "type": "text", + "content": ", and simply set " + }, + { + "bbox": [ + 130, + 202, + 480, + 334 + ], + "type": "inline_equation", + "content": "\\gamma_p = 1" + }, + { + "bbox": [ + 130, + 202, + 480, + 334 + ], + "type": "text", + "content": ". However, we note that a more careful choice of these parameters (e.g., with taking into account quantization granularity) represents one potential way to further improve the quantization properties of transformation matrices " + }, + { + "bbox": [ + 130, + 202, + 480, + 334 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 202, + 480, + 334 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 335, + 479, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 479, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 479, + 360 + ], + "type": "text", + "content": "To deal with Eq. (3), we start from the extended form of the Winograd transformation matrices " + }, + { + "bbox": [ + 130, + 335, + 479, + 360 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 335, + 479, + 360 + ], + "type": "text", + "content": " of Ref. [44]" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 233, + 379, + 480, + 394 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 379, + 480, + 394 + ], + "spans": [ + { + "bbox": [ + 233, + 379, + 480, + 394 + ], + "type": "interline_equation", + "content": "A ^ {T} = \\left(V _ {(m + n - 1) \\times m}\\right) ^ {T} S _ {A}, \\tag {4}", + "image_path": "eafcef9c9b17fb9b4dc6cc433c98168ef1fe9f681c812ba2fb1d4514bf5b7840.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 233, + 397, + 480, + 413 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 397, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 233, + 397, + 480, + 413 + ], + "type": "interline_equation", + "content": "B ^ {T} = S _ {B} \\left(V _ {(m + n - 1) \\times (m + n - 1)}\\right) ^ {- T}, \\tag {5}", + "image_path": "ff741d8f9a9ffb7960b9b0760c8baac5d7cf3418e4abea1761f68d657438591b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 232, + 416, + 480, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 416, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 232, + 416, + 480, + 429 + ], + "type": "interline_equation", + "content": "G = S _ {G} \\left(V _ {(m + n - 1) \\times n}\\right), \\tag {6}", + "image_path": "2a55d43bb7a01c789dcc5d9687736fa9a2c2ef56a0c9f2b8215e8d12b63525ea.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 437, + 379, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 437, + 379, + 452 + ], + "spans": [ + { + "bbox": [ + 131, + 437, + 379, + 452 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 437, + 379, + 452 + ], + "type": "inline_equation", + "content": "V_{x\\times y}" + }, + { + "bbox": [ + 131, + 437, + 379, + 452 + ], + "type": "text", + "content": " is a trimmed Vandermonde matrix given by" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 176, + 461, + 480, + 524 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 461, + 480, + 524 + ], + "spans": [ + { + "bbox": [ + 176, + 461, + 480, + 524 + ], + "type": "interline_equation", + "content": "V = \\left[ \\begin{array}{c c c c} \\xi_ {0} ^ {0} \\eta_ {0} ^ {y - 1} & \\xi_ {0} ^ {1} \\eta_ {0} ^ {y - 2} & \\dots \\dots & \\xi_ {0} ^ {y - 1} \\eta_ {0} ^ {0} \\\\ \\xi_ {1} ^ {0} \\eta_ {1} ^ {y - 1} & \\xi_ {1} ^ {1} \\eta_ {1} ^ {y - 2} & \\dots \\dots & \\xi_ {1} ^ {y - 1} \\eta_ {1} ^ {0} \\\\ \\vdots & \\vdots & & \\vdots \\\\ \\xi_ {x - 1} ^ {0} \\eta_ {x - 1} ^ {y - 1} & \\xi_ {x - 1} ^ {1} \\eta_ {x - 1} ^ {y - 2} & \\dots \\dots & \\xi_ {x - 1} ^ {y - 1} \\eta_ {x - 1} ^ {0} \\end{array} \\right], \\tag {7}", + "image_path": "cd8b12f3fd35fd03be9c8c0cd014e93d95c46d0b9e8020fbd1ad411081a4f8f8.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "inline_equation", + "content": "\\{(\\xi_i,\\eta_i)\\}" + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "text", + "content": " is the set of the homogenous coordinates (polynomial points) and " + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "inline_equation", + "content": "S_{p\\in \\{A,B,G\\}}" + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "text", + "content": " are the diagonal scaling matrices fulfilling the " + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "inline_equation", + "content": "S_{A}S_{B}S_{G} = 1" + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "text", + "content": " condition. Within this representation, 6 and 8 polynomial points need to be given to construct " + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "inline_equation", + "content": "F4" + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "inline_equation", + "content": "F6" + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "text", + "content": ", respectively. The polynomial points and scalings which generate the standard sets of the Winograd matrices " + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{std}}" + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "inline_equation", + "content": "F4" + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "inline_equation", + "content": "F6" + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "text", + "content": " algorithms are given in Supplemental Material Sec. 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "type": "text", + "content": "We consider polynomial points " + }, + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\{(\\xi_i,\\eta_i)\\}" + }, + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "type": "text", + "content": " as optimization parameters and perform minimization of the objective function " + }, + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "type": "inline_equation", + "content": "f(\\Theta)" + }, + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "type": "text", + "content": " using the particle swarm optimization (PSO) method [23, 45, 49]. PSO is a powerful swarm intelligence technique for solving global optimization problems. In this method, a set of " + }, + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "type": "text", + "content": " candidate solutions, called swarm particles, are moved through a search-space." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 321, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 321, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 321, + 102 + ], + "type": "text", + "content": "V. Protsenko, V. Kryzhanovskiy et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "Each " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "th, " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "i \\in (1 \\dots N)" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": ", particle is assigned with a position " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " and a velocity " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": ", which evolve during unity time steps " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " (or iteration numbers) according to the following update rules" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 163, + 481, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 163, + 481, + 175 + ], + "spans": [ + { + "bbox": [ + 179, + 163, + 481, + 175 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {i} (t + 1) \\leftarrow \\omega \\mathbf {v} _ {i} (t) + \\phi_ {1} r _ {1} [ \\mathbf {p} _ {i} - \\mathbf {x} _ {i} (t) ] + \\phi_ {2} r _ {2} [ \\mathbf {p} _ {g} - \\mathbf {x} _ {i} (t) ], \\tag {8}", + "image_path": "0c44dd800fe43e962785b19ec0a11bb30c5a7f9d4a42f3a66db5032bbac68c59.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 240, + 197, + 481, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 197, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 240, + 197, + 481, + 209 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {i} (t + 1) \\leftarrow \\mathbf {x} _ {i} (t) + \\mathbf {v} _ {i} (t + 1), \\tag {9}", + "image_path": "79db8e4ddbea1b74cf5f66958a36e86f50c36834ef6cea0bcdb70945bf3dfe3f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 214, + 479, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 479, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 479, + 274 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 214, + 479, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 130, + 214, + 479, + 274 + ], + "type": "text", + "content": " is the best solution that " + }, + { + "bbox": [ + 130, + 214, + 479, + 274 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 214, + 479, + 274 + ], + "type": "text", + "content": "th particle has found so far, " + }, + { + "bbox": [ + 130, + 214, + 479, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_g" + }, + { + "bbox": [ + 130, + 214, + 479, + 274 + ], + "type": "text", + "content": " is the best solution revealed by the whole swarm, " + }, + { + "bbox": [ + 130, + 214, + 479, + 274 + ], + "type": "inline_equation", + "content": "r_{1(2)}" + }, + { + "bbox": [ + 130, + 214, + 479, + 274 + ], + "type": "text", + "content": " are random numbers sampled at each time step from a uniform " + }, + { + "bbox": [ + 130, + 214, + 479, + 274 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 130, + 214, + 479, + 274 + ], + "type": "text", + "content": " distribution. Finally, " + }, + { + "bbox": [ + 130, + 214, + 479, + 274 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 214, + 479, + 274 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 214, + 479, + 274 + ], + "type": "inline_equation", + "content": "\\phi_1" + }, + { + "bbox": [ + 130, + 214, + 479, + 274 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 214, + 479, + 274 + ], + "type": "inline_equation", + "content": "\\phi_2" + }, + { + "bbox": [ + 130, + 214, + 479, + 274 + ], + "type": "text", + "content": " are PSO hyperparameters, which we set to 0.9, 0.55 and 0.3. The PSO optimization is implemented within PySwarms toolkit [33]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 274, + 482, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 514 + ], + "type": "text", + "content": "The evolution of the objective function " + }, + { + "bbox": [ + 130, + 274, + 482, + 514 + ], + "type": "inline_equation", + "content": "f(\\Theta)" + }, + { + "bbox": [ + 130, + 274, + 482, + 514 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 274, + 482, + 514 + ], + "type": "inline_equation", + "content": "10^{4}" + }, + { + "bbox": [ + 130, + 274, + 482, + 514 + ], + "type": "text", + "content": " iterations of the PSO algorithm with 2000 particles is shown in Fig. 1a. Compared to the value for the standard transformations " + }, + { + "bbox": [ + 130, + 274, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{std}}" + }, + { + "bbox": [ + 130, + 274, + 482, + 514 + ], + "type": "text", + "content": " (indicated by circled number 0), the objective function after the PSO optimization (circled number 6) is reduced by about 2.6 and 5.2 times for " + }, + { + "bbox": [ + 130, + 274, + 482, + 514 + ], + "type": "inline_equation", + "content": "F4" + }, + { + "bbox": [ + 130, + 274, + 482, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 274, + 482, + 514 + ], + "type": "inline_equation", + "content": "F6" + }, + { + "bbox": [ + 130, + 274, + 482, + 514 + ], + "type": "text", + "content": ", respectively. By varying the hyperparameters " + }, + { + "bbox": [ + 130, + 274, + 482, + 514 + ], + "type": "inline_equation", + "content": "(\\omega, \\phi_1, \\phi_2)" + }, + { + "bbox": [ + 130, + 274, + 482, + 514 + ], + "type": "text", + "content": ", we verified that the PSO optimization leads either to the same value of the objective function (see inset of Fig. 1a) or to sub-optimal solutions (not shown). In the following, we denote the Winograd transformation matrices, which represent the PSO solution to the optimization problem given by Eq. (3), as " + }, + { + "bbox": [ + 130, + 274, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{pso}}" + }, + { + "bbox": [ + 130, + 274, + 482, + 514 + ], + "type": "text", + "content": ". The corresponding polynomial points " + }, + { + "bbox": [ + 130, + 274, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\{(\\xi_i, \\eta_i)\\}" + }, + { + "bbox": [ + 130, + 274, + 482, + 514 + ], + "type": "text", + "content": " from which they can be constructed according to Eqs. (4) to (7) and explicit expressions for " + }, + { + "bbox": [ + 130, + 274, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{pso}}" + }, + { + "bbox": [ + 130, + 274, + 482, + 514 + ], + "type": "text", + "content": " are given in the Supplemental Material Sec. 1. Note that in addition to the polynomial points, scaling matrices " + }, + { + "bbox": [ + 130, + 274, + 482, + 514 + ], + "type": "inline_equation", + "content": "S_{p \\in \\{A, B, G\\}}" + }, + { + "bbox": [ + 130, + 274, + 482, + 514 + ], + "type": "text", + "content": " can also be considered as optimization parameters. However, the polynomial points already provide certain degrees of freedom for the scaling variation of the transformation matrices. In addition, the inclusion of " + }, + { + "bbox": [ + 130, + 274, + 482, + 514 + ], + "type": "inline_equation", + "content": "S_{p \\in \\{A, B, G\\}}" + }, + { + "bbox": [ + 130, + 274, + 482, + 514 + ], + "type": "text", + "content": " in the PSO procedure leads to a significant expansion of the number of parameters to be optimized, which makes optimization " + }, + { + "bbox": [ + 130, + 274, + 482, + 514 + ], + "type": "inline_equation", + "content": "f(\\Theta)" + }, + { + "bbox": [ + 130, + 274, + 482, + 514 + ], + "type": "text", + "content": " inefficient due to getting stuck in local minima. In view of this, we do not optimize " + }, + { + "bbox": [ + 130, + 274, + 482, + 514 + ], + "type": "inline_equation", + "content": "S_{p \\in \\{A, B, G\\}}" + }, + { + "bbox": [ + 130, + 274, + 482, + 514 + ], + "type": "text", + "content": " and use their standard values (see Ref. [44], Secs. 1 and 2 of the Supplemental Material)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 532, + 282, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 282, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 282, + 547 + ], + "type": "text", + "content": "4 Experimental Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "In our experiments we use two different training policies: (i) keep the Winograd transformations " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " fixed, which we refer to as Static " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "; (ii) treat " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " as learnable parameters during training, which we denote as Trainable " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ". The scheme (i) is used to show that replacing the standard Winograd matrices " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{std}}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " by our PSO-derived ones " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{pso}}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " already leads to a significant improvement in accuracy, which demonstrates a quantization-friendly nature of the latter. Within the second scheme (ii), we use the SOTA training approach of Ref. [17] (see discussion in Sec. 2). We replace the standard matrices " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{std}}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " used in this pipeline by our " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{pso}}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ". We show that this modification allows us to outperform the SOTA results." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 207, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 207, + 90, + 448, + 102 + ], + "type": "text", + "content": "Quantization-Friendly Winograd Transformations for CNN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "To ensure a fair comparison of different training scenarios, we extensively evaluate their performance on three different types of tasks: (i) classification, (ii) super-resolution, and (iii) semantic segmentation. Unless otherwise stated, all convolution layers are quantized to 8-bit, except for the first and last convolution layers. For experiments with Winograd convolutions, we replace the direct convolutions having stride 1 and kernel size " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "F4" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "F6" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " algorithms. Details on datasets and training protocols are presented in Secs. 3 and 4 of the Supplemental Material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 228, + 228, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 228, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 228, + 239 + ], + "type": "text", + "content": "4.1 Classification" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 247, + 482, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 305 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 305 + ], + "type": "text", + "content": "We begin by considering the ResNet- " + }, + { + "bbox": [ + 130, + 247, + 482, + 305 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 247, + 482, + 305 + ], + "type": "text", + "content": " family of models, " + }, + { + "bbox": [ + 130, + 247, + 482, + 305 + ], + "type": "inline_equation", + "content": "18 \\leq N \\leq 50" + }, + { + "bbox": [ + 130, + 247, + 482, + 305 + ], + "type": "text", + "content": ". As in Ref. [17], in some cases we vary a channel multiplier (or \"width multiplier\" as used for the MobileNets) factor, which scales the number of output channels produced by each convolution layer. A channel multiplier equal to 1.0 corresponds to the full ResNet- " + }, + { + "bbox": [ + 130, + 247, + 482, + 305 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 247, + 482, + 305 + ], + "type": "text", + "content": " model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 306, + 482, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 482, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 482, + 367 + ], + "type": "text", + "content": "To clearly demonstrate that minimization of the objective function " + }, + { + "bbox": [ + 130, + 306, + 482, + 367 + ], + "type": "inline_equation", + "content": "f(\\Theta)" + }, + { + "bbox": [ + 130, + 306, + 482, + 367 + ], + "type": "text", + "content": " (see Eq. (3)) indeed leads to the reduction of quantization errors, we consider the accuracy of intermediate steps of the PSO-optimization in Fig. 1. Each " + }, + { + "bbox": [ + 130, + 306, + 482, + 367 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 306, + 482, + 367 + ], + "type": "text", + "content": "th step (circled number) represents transformations " + }, + { + "bbox": [ + 130, + 306, + 482, + 367 + ], + "type": "inline_equation", + "content": "\\Theta^i" + }, + { + "bbox": [ + 130, + 306, + 482, + 367 + ], + "type": "text", + "content": " corresponding to a particular value of the objective function (see panels a). The 0th step corresponds to" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 132, + 387, + 302, + 553 + ], + "blocks": [ + { + "bbox": [ + 132, + 387, + 302, + 553 + ], + "lines": [ + { + "bbox": [ + 132, + 387, + 302, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 302, + 553 + ], + "type": "image", + "image_path": "897cfdd43f65e4cfb2ea06f2e017e388adb51e0ccba094c0efe053a1345a2f22.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 563, + 482, + 662 + ], + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 662 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 662 + ], + "type": "text", + "content": "Fig. 1: a) The convergence history of the objective function " + }, + { + "bbox": [ + 130, + 563, + 482, + 662 + ], + "type": "inline_equation", + "content": "f(\\Theta)" + }, + { + "bbox": [ + 130, + 563, + 482, + 662 + ], + "type": "text", + "content": " (see Eq. (3)) for " + }, + { + "bbox": [ + 130, + 563, + 482, + 662 + ], + "type": "inline_equation", + "content": "F4" + }, + { + "bbox": [ + 130, + 563, + 482, + 662 + ], + "type": "text", + "content": " Winograd algorithm within the PSO optimization. Inset shows five independent runs of the PSO algorithm. b, c) Accuracy performance of the 8-bit " + }, + { + "bbox": [ + 130, + 563, + 482, + 662 + ], + "type": "inline_equation", + "content": "F4" + }, + { + "bbox": [ + 130, + 563, + 482, + 662 + ], + "type": "text", + "content": " ResNet-18 model on the CIFAR-10 dataset at different steps of the PSO optimization (circled numbers in Fig. 1a and on the horizontal axis of Fig. 1b,c) for static (b) and trainable (c) " + }, + { + "bbox": [ + 130, + 563, + 482, + 662 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 563, + 482, + 662 + ], + "type": "text", + "content": ". Red and black dots correspond to channel multipliers of 1.0 and 0.5, respectively. The dashed lines in Fig. 1c represent the accuracy of the FP direct convolution model. On F6 we obtained a qualitatively similar convergence and accuracy evolution (see Sec. 5 of the Supplemental Material)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 303, + 387, + 392, + 549 + ], + "blocks": [ + { + "bbox": [ + 303, + 387, + 392, + 549 + ], + "lines": [ + { + "bbox": [ + 303, + 387, + 392, + 549 + ], + "spans": [ + { + "bbox": [ + 303, + 387, + 392, + 549 + ], + "type": "image", + "image_path": "f276b9be4ac85e81f5a6e0ecdd55da88d3951a16aa029f9ab8e00c0e09f81c9b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 392, + 387, + 481, + 549 + ], + "blocks": [ + { + "bbox": [ + 392, + 387, + 481, + 549 + ], + "lines": [ + { + "bbox": [ + 392, + 387, + 481, + 549 + ], + "spans": [ + { + "bbox": [ + 392, + 387, + 481, + 549 + ], + "type": "image", + "image_path": "f009d19f3516ef1a59dacb60c015f8fc17f8c9ebffb86ad810e317f8568bb5e5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 321, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 321, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 321, + 102 + ], + "type": "text", + "content": "V. Protsenko, V. Kryzhanovskiy et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 301, + 192 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 301, + 192 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 301, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 301, + 192 + ], + "type": "image", + "image_path": "f9159249a0a525cc74b8bc82629bb961a8f764b09b488494ef2d09bda1f70e32.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 200, + 480, + 222 + ], + "lines": [ + { + "bbox": [ + 130, + 200, + 480, + 222 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 480, + 222 + ], + "type": "text", + "content": "Fig. 2: Distribution of the accuracy of 8-bit " + }, + { + "bbox": [ + 130, + 200, + 480, + 222 + ], + "type": "inline_equation", + "content": "F4" + }, + { + "bbox": [ + 130, + 200, + 480, + 222 + ], + "type": "text", + "content": " ResNet-18 model (chan. mult. is 1.0) on the CIFAR-10 dataset over 15 training runs for different training scenarios." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 310, + 114, + 480, + 192 + ], + "blocks": [ + { + "bbox": [ + 310, + 114, + 480, + 192 + ], + "lines": [ + { + "bbox": [ + 310, + 114, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 310, + 114, + 480, + 192 + ], + "type": "image", + "image_path": "91629ae99bb399bd7ad57449f9f76cf4f730bcc20d02fbdf18135c98cc76a17e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 247, + 480, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 480, + 354 + ], + "type": "text", + "content": "the standard Winograd transformations " + }, + { + "bbox": [ + 130, + 247, + 480, + 354 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{std}}" + }, + { + "bbox": [ + 130, + 247, + 480, + 354 + ], + "type": "text", + "content": ", while the 6th step represents our transformations " + }, + { + "bbox": [ + 130, + 247, + 480, + 354 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{pso}}" + }, + { + "bbox": [ + 130, + 247, + 480, + 354 + ], + "type": "text", + "content": ". One can see that the PSO optimization of " + }, + { + "bbox": [ + 130, + 247, + 480, + 354 + ], + "type": "inline_equation", + "content": "f(\\Theta)" + }, + { + "bbox": [ + 130, + 247, + 480, + 354 + ], + "type": "text", + "content": " results in a monotonic increase of the accuracy. In the case of the " + }, + { + "bbox": [ + 130, + 247, + 480, + 354 + ], + "type": "inline_equation", + "content": "F4" + }, + { + "bbox": [ + 130, + 247, + 480, + 354 + ], + "type": "text", + "content": " algorithm with trainable " + }, + { + "bbox": [ + 130, + 247, + 480, + 354 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 247, + 480, + 354 + ], + "type": "text", + "content": " (Fig. 1c), by starting training from " + }, + { + "bbox": [ + 130, + 247, + 480, + 354 + ], + "type": "inline_equation", + "content": "\\Theta = \\Theta^{\\mathrm{pso}}" + }, + { + "bbox": [ + 130, + 247, + 480, + 354 + ], + "type": "text", + "content": " we not only outperform the SOTA " + }, + { + "bbox": [ + 130, + 247, + 480, + 354 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{std}}" + }, + { + "bbox": [ + 130, + 247, + 480, + 354 + ], + "type": "text", + "content": "-based results of Ref [17], but also achieve the accuracy of full-precision (FP) direct convolution models (denoted by dashed lines). Note that in some cases we observe the same effects as in Refs. [17,39], when quantization yields somewhat better accuracy compared to one of FP model. This is because quantization can provide an additional mechanism of regularization [2,39]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 356, + 480, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 480, + 450 + ], + "type": "text", + "content": "The characteristic improvements from using " + }, + { + "bbox": [ + 130, + 356, + 480, + 450 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{pseudo}}" + }, + { + "bbox": [ + 130, + 356, + 480, + 450 + ], + "type": "text", + "content": " are not sensitive to both the precision of PSO convergence and random factors during network initialization. To show this we perform three training runs of the " + }, + { + "bbox": [ + 130, + 356, + 480, + 450 + ], + "type": "inline_equation", + "content": "F4" + }, + { + "bbox": [ + 130, + 356, + 480, + 450 + ], + "type": "text", + "content": " ResNet-18 model with different random seeds for each of the five converged configurations shown in the inset of Fig. 1a. The corresponding accuracy distribution is compared in Fig. 2 to that obtained for 15 random seeds using the standard transformations " + }, + { + "bbox": [ + 130, + 356, + 480, + 450 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{std}}" + }, + { + "bbox": [ + 130, + 356, + 480, + 450 + ], + "type": "text", + "content": ". It can be seen that the obtained distributions are strongly/substantially separated in the case of static/trainable Winograd transformations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 451, + 481, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 481, + 512 + ], + "type": "text", + "content": "We find that replacing the standard transformations " + }, + { + "bbox": [ + 130, + 451, + 481, + 512 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{std}}" + }, + { + "bbox": [ + 130, + 451, + 481, + 512 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 451, + 481, + 512 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{pseudo}}" + }, + { + "bbox": [ + 130, + 451, + 481, + 512 + ], + "type": "text", + "content": " matrices, even without including them in a training loop (static " + }, + { + "bbox": [ + 130, + 451, + 481, + 512 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 451, + 481, + 512 + ], + "type": "text", + "content": " scheme), already results in a considerable accuracy improvement (Tab. 1 and Fig. 3). This improvement is robust to variations of both network depth (Tab. 1) and channel scaling (dashed lines in Fig. 3). Learning Winograd transformations during training allows to" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 139, + 587, + 474, + 664 + ], + "blocks": [ + { + "bbox": [ + 130, + 533, + 482, + 578 + ], + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 578 + ], + "type": "text", + "content": "Table 1: The accuracy " + }, + { + "bbox": [ + 130, + 533, + 482, + 578 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 130, + 533, + 482, + 578 + ], + "type": "text", + "content": " of 8-bit " + }, + { + "bbox": [ + 130, + 533, + 482, + 578 + ], + "type": "inline_equation", + "content": "F4-" + }, + { + "bbox": [ + 130, + 533, + 482, + 578 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 482, + 578 + ], + "type": "inline_equation", + "content": "F6" + }, + { + "bbox": [ + 130, + 533, + 482, + 578 + ], + "type": "text", + "content": " -based ResNet- " + }, + { + "bbox": [ + 130, + 533, + 482, + 578 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 533, + 482, + 578 + ], + "type": "text", + "content": " models (chan. mult. is 1.0) on CIFAR-10 when " + }, + { + "bbox": [ + 130, + 533, + 482, + 578 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 533, + 482, + 578 + ], + "type": "text", + "content": " is static. The best results are shown in bold. The values in parentheses give the improvement (to the first decimal place) over " + }, + { + "bbox": [ + 130, + 533, + 482, + 578 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{std}}" + }, + { + "bbox": [ + 130, + 533, + 482, + 578 + ], + "type": "text", + "content": " case, and the results for the " + }, + { + "bbox": [ + 130, + 533, + 482, + 578 + ], + "type": "inline_equation", + "content": "F4" + }, + { + "bbox": [ + 130, + 533, + 482, + 578 + ], + "type": "text", + "content": " ResNet-18 are averaged over the experiments of Fig. 2." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 139, + 587, + 474, + 664 + ], + "lines": [ + { + "bbox": [ + 139, + 587, + 474, + 664 + ], + "spans": [ + { + "bbox": [ + 139, + 587, + 474, + 664 + ], + "type": "table", + "html": "
ModelF4 Static ΘF6 Static Θ
Standard ΘOur Θpso (↑)Standard ΘOur Θpso (↑)
ResNet-1883.6591.54 (+7.9)83.2387.51 (+4.3)
ResNet-3484.2290.48 (+6.3)69.4887.09 (+17.6)
ResNet-5082.5791.52 (+8.9)70.1582.87 (+12.7)
", + "image_path": "2a8334a65cb4bf98cd3f3e6e8d6aaa741a1739ec296259b3e50014b98efd42c0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 207, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 207, + 90, + 448, + 102 + ], + "type": "text", + "content": "Quantization-Friendly Winograd Transformations for CNN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 482, + 246 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 482, + 246 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 482, + 246 + ], + "type": "image", + "image_path": "5aee53d945fd2fa51b8f8a4fa116659dfebac915d0ddc6f25a62e1d80a153dda.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 253, + 482, + 288 + ], + "lines": [ + { + "bbox": [ + 130, + 253, + 482, + 288 + ], + "spans": [ + { + "bbox": [ + 130, + 253, + 482, + 288 + ], + "type": "text", + "content": "Fig. 3: The accuracy of the ResNet-18 on CIFAR-10 with various channel multipliers utilizing full 8-bit " + }, + { + "bbox": [ + 130, + 253, + 482, + 288 + ], + "type": "inline_equation", + "content": "F4" + }, + { + "bbox": [ + 130, + 253, + 482, + 288 + ], + "type": "text", + "content": " (a) and " + }, + { + "bbox": [ + 130, + 253, + 482, + 288 + ], + "type": "inline_equation", + "content": "F6" + }, + { + "bbox": [ + 130, + 253, + 482, + 288 + ], + "type": "text", + "content": " (b) Winograd convolutions. The SOTA data are taken from Fig. 4 of Ref. [17]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 318, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 482, + 413 + ], + "type": "text", + "content": "further adjust them to minimize accuracy errors (see solid lines in Fig. 3, Tabs. 2 and 3). In all experiments, the initialization of the Winograd transformations when considering them as trainable parameters via our PSO-derived matrices " + }, + { + "bbox": [ + 130, + 318, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{pso}}" + }, + { + "bbox": [ + 130, + 318, + 482, + 413 + ], + "type": "text", + "content": " is found to be highly beneficial. One can see that this training scenario not only considerably outperforms the SOTA " + }, + { + "bbox": [ + 130, + 318, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{std}}" + }, + { + "bbox": [ + 130, + 318, + 482, + 413 + ], + "type": "text", + "content": "-based training approach of Ref. [17], but also either completely closes (in the case of CIFAR-10) or strongly reduces (for CIFAR-100 and ImageNet) the error gap between the " + }, + { + "bbox": [ + 130, + 318, + 482, + 413 + ], + "type": "inline_equation", + "content": "F4" + }, + { + "bbox": [ + 130, + 318, + 482, + 413 + ], + "type": "text", + "content": " Winograd convolution and the FP direct convolution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 418, + 483, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 483, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 483, + 491 + ], + "type": "text", + "content": "Finally, Tab. 2 shows a competitive comparison of the performance of using trainable transformations starting from " + }, + { + "bbox": [ + 130, + 418, + 483, + 491 + ], + "type": "inline_equation", + "content": "\\Theta = \\Theta^{\\mathrm{pseudo}}" + }, + { + "bbox": [ + 130, + 418, + 483, + 491 + ], + "type": "text", + "content": " (column \"Our\") with the current SOTA methods. In most cases, our approach significantly outperforms the SOTA results in terms of accuracy. Note that for a fair comparison, in Tab. 2 we consider only the case of the scalar quantization and real-arithmetic full 8-bit quantized Winograd algorithm. However, our method is compatible with more" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 141, + 540, + 473, + 665 + ], + "blocks": [ + { + "bbox": [ + 141, + 518, + 471, + 531 + ], + "lines": [ + { + "bbox": [ + 141, + 518, + 471, + 531 + ], + "spans": [ + { + "bbox": [ + 141, + 518, + 471, + 531 + ], + "type": "text", + "content": "Table 2: An accuracy (%) comparison of Winograd-aware quantization schemes." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 540, + 473, + 665 + ], + "lines": [ + { + "bbox": [ + 141, + 540, + 473, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 540, + 473, + 665 + ], + "type": "table", + "html": "
Model & DatasetMethodFmRef.Ours
CIFAR-10ResNet-18Winograd-aware [17]F492.4693.26
Winograd-aware [17]F687.7489.57
ResNet-18 (mult. 0.5)Legendre [6]F491.892.53
ResNet-18 (mult. 0.25)Legendre [6]F489.790.29
ResNet-20WinoTrain [35]F489.792.31
WinoVidiVici [34]90.89
ResNet-18 & ImageNetWinoTrain [35]F461.2869.15
WinoVidiVici [34]69.14
", + "image_path": "ec3a4fb3f8c65eeaab36f90df481dcda9a1136a1dfae6af68dced68a164d4b90.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 321, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 321, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 321, + 102 + ], + "type": "text", + "content": "V. Protsenko, V. Kryzhanovskiy et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 178, + 475, + 536 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": "Table 3: Performance of the full 8-bit quantized " + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "inline_equation", + "content": "F4-" + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "inline_equation", + "content": "F6" + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": "-based models (channel multiplier is 1.0) on the CIFAR-10 dataset when " + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": " is trainable. Note that here SOTA refers to the results of the training pipeline of Ref. [17] obtained using the official code released by the authors. The abbreviation FP stands for the results of the full precision direct convolution models. Other notations are the same as in Tab. 1." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 178, + 475, + 536 + ], + "lines": [ + { + "bbox": [ + 140, + 178, + 475, + 536 + ], + "spans": [ + { + "bbox": [ + 140, + 178, + 475, + 536 + ], + "type": "table", + "html": "
Task (val.)Model (Test set)Direct conv.8-bit Fm with trainable Θ
FP8-bitFmSOTAOur Θpso (↑)
Classification (Accuracy, %)ResNet-18 (CIFAR-10)93.1793.20F492.5093.26 (+0.8)
F687.7489.57 (+1.8)
ResNet-18 (CIFAR-100)74.6074.18F472.3874.10 (+1.7)
F666.1572.90 (+6.7)
ResNet-18 (ImageNet)69.7969.61F467.0569.15 (+2.1)
F653.9062.87 (+9.0)
ResNet-34 (CIFAR-10)93.4093.12F492.5493.29 (+0.7)
F680.0590.16 (+10.1)
ResNet-50 (CIFAR-10)93.3593.22F493.4693.68 (+0.2)
F686.3792.16 (+5.8)
×3 Super-resolution (PSNR, dB)ESPCN [41] (Set5)32.5332.38F430.8632.11 (+1.2)
F627.3329.82 (+2.5)
ESPCN [41] (Set14)29.2129.09F428.0528.87 (+0.8)
F625.6227.30 (+1.7)
ESPCN-6 (Set5)33.2133.12F430.4532.98 (+2.5)
F625.5330.16 (+4.6)
ESPCN-6 (Set14)29.5929.50F427.7529.44 (+1.7)
F624.0227.50 (+3.5)
RFDN [29] (Set5)34.1333.68F430.6733.36 (+2.7)
RFDN [29] (Set14)30.1729.81F427.9429.62 (+1.7)
Segm IoUENet [38] (CamVid)60.6160.17F456.1260.00 (+3.9)
F652.4356.39 (+4.0)
", + "image_path": "915304316f737dbc52f029f44074844493c804b04161cfa1c8f8736bb8acfd73.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 557, + 479, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 479, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 479, + 582 + ], + "type": "text", + "content": "sophisticated quantization schemes (e.g., the tapwise quantization of Ref. [5]) and further performance improvement is expected from their application." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 598, + 248, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 248, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 248, + 610 + ], + "type": "text", + "content": "4.2 Super-Resolution" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Here we examine the performance on super-resolution (SR) tasks considering the original and modified ESPCN [41] as well as the RFDN [29] models. We set the upscaling factor between low- and high-resolution images to " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\times 3" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ". Since the SOTA approach of Ref. [17] (training " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " starting from " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{std}}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ") has only been" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "type": "text", + "content": "Quantization-Friendly Winograd Transformations for CNN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "implemented for classification scenarios, we extend it to SR networks based on its official implementation. All models are trained on the DIV2K dataset [1] and their performance is evaluated on the Set5 and Set14 benchmark data (more details on the SR architectures, training, and dataset preparation can be found in the Supplemental Material). As seen from the previous discussion, treating the Winograd transformations " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " as trainable parameters is essential to overcome the quality degradation caused by quantization errors. For this reason, in the following, we focus only on the case of trainable " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ". The mean PSNR values of the considered models on the benchmark datasets are summarized in Tab. 3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 226, + 482, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 310 + ], + "type": "text", + "content": "The ESPCN model is the less-depth model under consideration. It consists of three convolutional layers followed by a pixel-shuffle operation. In our implementation, the last two convolution layers are replaced by either " + }, + { + "bbox": [ + 130, + 226, + 482, + 310 + ], + "type": "inline_equation", + "content": "F4" + }, + { + "bbox": [ + 130, + 226, + 482, + 310 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 226, + 482, + 310 + ], + "type": "inline_equation", + "content": "F6" + }, + { + "bbox": [ + 130, + 226, + 482, + 310 + ], + "type": "text", + "content": " Winograd convolutions and are quantized to 8-bit. On F4, training of the transformations from " + }, + { + "bbox": [ + 130, + 226, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\Theta = \\Theta^{\\mathrm{pso}}" + }, + { + "bbox": [ + 130, + 226, + 482, + 310 + ], + "type": "text", + "content": " gains 1.25dB and 0.82dB higher PSNR than the SOTA approach of Ref. [17] on Set5 and Set14, respectively. On " + }, + { + "bbox": [ + 130, + 226, + 482, + 310 + ], + "type": "inline_equation", + "content": "F6" + }, + { + "bbox": [ + 130, + 226, + 482, + 310 + ], + "type": "text", + "content": ", the PSNR gap from the FP direct model is reduced by about 2 times within our training approach." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 313, + 482, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 313, + 482, + 385 + ], + "spans": [ + { + "bbox": [ + 130, + 313, + 482, + 385 + ], + "type": "text", + "content": "It can be expected that proper initialization of " + }, + { + "bbox": [ + 130, + 313, + 482, + 385 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 313, + 482, + 385 + ], + "type": "text", + "content": " at the beginning of training is less critical for such lightweight networks. Indeed, due to the relatively small number of the transformation parameters to train, it is possible that even a local gradient-based optimization during network training may be enough to learn appropriate Winograd transformations. Thus, quality degradation when using standard matrices " + }, + { + "bbox": [ + 130, + 313, + 482, + 385 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{std}}" + }, + { + "bbox": [ + 130, + 313, + 482, + 385 + ], + "type": "text", + "content": " is expected to be more pronounced with increas" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 133, + 411, + 218, + 536 + ], + "blocks": [ + { + "bbox": [ + 133, + 411, + 218, + 536 + ], + "lines": [ + { + "bbox": [ + 133, + 411, + 218, + 536 + ], + "spans": [ + { + "bbox": [ + 133, + 411, + 218, + 536 + ], + "type": "image", + "image_path": "5e25e33bcca89537b62ce751c595f19af0866db4c9bb77c9986e84c2a8ddf1bf.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 134, + 537, + 216, + 620 + ], + "blocks": [ + { + "bbox": [ + 134, + 537, + 216, + 620 + ], + "lines": [ + { + "bbox": [ + 134, + 537, + 216, + 620 + ], + "spans": [ + { + "bbox": [ + 134, + 537, + 216, + 620 + ], + "type": "image", + "image_path": "7052acbee1c77cd69d0a461282920f1b300e089aef6e06bb2dda81c3cc1e7dd3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 620, + 204, + 630 + ], + "lines": [ + { + "bbox": [ + 146, + 620, + 204, + 630 + ], + "spans": [ + { + "bbox": [ + 146, + 620, + 204, + 630 + ], + "type": "text", + "content": "(a) Direct / FP" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 639, + 481, + 662 + ], + "lines": [ + { + "bbox": [ + 130, + 639, + 481, + 662 + ], + "spans": [ + { + "bbox": [ + 130, + 639, + 481, + 662 + ], + "type": "text", + "content": "Fig. 4: " + }, + { + "bbox": [ + 130, + 639, + 481, + 662 + ], + "type": "inline_equation", + "content": "\\times 3" + }, + { + "bbox": [ + 130, + 639, + 481, + 662 + ], + "type": "text", + "content": " Super-resolution results of the RFDN model utilizing direct and " + }, + { + "bbox": [ + 130, + 639, + 481, + 662 + ], + "type": "inline_equation", + "content": "F4" + }, + { + "bbox": [ + 130, + 639, + 481, + 662 + ], + "type": "text", + "content": " Winograd convolutions for \"Comic\" from Set14." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 220, + 411, + 306, + 537 + ], + "blocks": [ + { + "bbox": [ + 220, + 411, + 306, + 537 + ], + "lines": [ + { + "bbox": [ + 220, + 411, + 306, + 537 + ], + "spans": [ + { + "bbox": [ + 220, + 411, + 306, + 537 + ], + "type": "image", + "image_path": "2c0651d8c3112ab1aa47b80b5537ae263bc7bc3ab0ad8e44ca39748514de8a27.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 223, + 537, + 304, + 620 + ], + "blocks": [ + { + "bbox": [ + 223, + 537, + 304, + 620 + ], + "lines": [ + { + "bbox": [ + 223, + 537, + 304, + 620 + ], + "spans": [ + { + "bbox": [ + 223, + 537, + 304, + 620 + ], + "type": "image", + "image_path": "5ad8d492dd2d0066425d9e50aa35e4acbe2a305bf33c67792ed711098bc4563c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 620, + 294, + 630 + ], + "lines": [ + { + "bbox": [ + 231, + 620, + 294, + 630 + ], + "spans": [ + { + "bbox": [ + 231, + 620, + 294, + 630 + ], + "type": "text", + "content": "(b) Direct / 8-bit" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 411, + 392, + 537 + ], + "blocks": [ + { + "bbox": [ + 309, + 411, + 392, + 537 + ], + "lines": [ + { + "bbox": [ + 309, + 411, + 392, + 537 + ], + "spans": [ + { + "bbox": [ + 309, + 411, + 392, + 537 + ], + "type": "image", + "image_path": "13ad549d98acad5ed397b442e05347e1c1a0f9ec5a015cfdbb08f78a7434c288.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 310, + 537, + 390, + 620 + ], + "blocks": [ + { + "bbox": [ + 310, + 537, + 390, + 620 + ], + "lines": [ + { + "bbox": [ + 310, + 537, + 390, + 620 + ], + "spans": [ + { + "bbox": [ + 310, + 537, + 390, + 620 + ], + "type": "image", + "image_path": "803af0da1a2da6bfae6a370334ad5d29ac5847e1c250e2f9cc8d165f9a807b5d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 620, + 388, + 630 + ], + "lines": [ + { + "bbox": [ + 313, + 620, + 388, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 620, + 388, + 630 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 313, + 620, + 388, + 630 + ], + "type": "inline_equation", + "content": "F4" + }, + { + "bbox": [ + 313, + 620, + 388, + 630 + ], + "type": "text", + "content": " SOTA / 8-bit" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 395, + 411, + 480, + 536 + ], + "blocks": [ + { + "bbox": [ + 395, + 411, + 480, + 536 + ], + "lines": [ + { + "bbox": [ + 395, + 411, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 395, + 411, + 480, + 536 + ], + "type": "image", + "image_path": "585a94bc3b422de38b8f831e7b222de8c8f040bde7f99883a62e7a251325eaf1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 398, + 536, + 477, + 620 + ], + "blocks": [ + { + "bbox": [ + 398, + 536, + 477, + 620 + ], + "lines": [ + { + "bbox": [ + 398, + 536, + 477, + 620 + ], + "spans": [ + { + "bbox": [ + 398, + 536, + 477, + 620 + ], + "type": "image", + "image_path": "ab0afd774726e721cf3842af3a7d9ee959b3ac3f3af6f140704a8024530c3cf3.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 620, + 474, + 630 + ], + "lines": [ + { + "bbox": [ + 403, + 620, + 474, + 630 + ], + "spans": [ + { + "bbox": [ + 403, + 620, + 474, + 630 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 403, + 620, + 474, + 630 + ], + "type": "inline_equation", + "content": "F4" + }, + { + "bbox": [ + 403, + 620, + 474, + 630 + ], + "type": "text", + "content": " Ours / 8-bit" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 321, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 321, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 321, + 102 + ], + "type": "text", + "content": "V. Protsenko, V. Kryzhanovskiy et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "ing network depth, as in the case of the classification tasks (see Sec. 4.1). To prove this, we consider the six-layered ESPCN-6 model, which is obtained by including three additional convolution layers in the ESPCN model. While the PSNR values obtained by the " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{pseudo}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "-based training increase compared to the results of the original ESPCN model, the SOTA training approach, using " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{std}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": ", on the contrary, tends to PSNR degradation. On " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "F4" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{pseudo}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "-based training shows an improvement of more than 2.5 dB for Set5 and 1.6 dB for Set14 relative to the SOTA training approach. With this improvement, the quality drop when using the 8-bit quantized " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "F4" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " algorithm is strongly reduced, and the resulting quality is close to that of FP direct convolutions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 242, + 482, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 242, + 482, + 459 + ], + "spans": [ + { + "bbox": [ + 130, + 242, + 482, + 459 + ], + "type": "text", + "content": "Finally, we perform experiments with the residual feature distillation network (RFDN) [29]. It is based on the revised information distillation mechanism of the IMDN network [21] and the shallow residual blocks, which allows to achieve a better trade-off between model performance and complexity. Utilizing " + }, + { + "bbox": [ + 130, + 242, + 482, + 459 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{pseudo}}" + }, + { + "bbox": [ + 130, + 242, + 482, + 459 + ], + "type": "text", + "content": " leads to remarkably better performance compared to the SOTA training based on the standard matrices " + }, + { + "bbox": [ + 130, + 242, + 482, + 459 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{std}}" + }, + { + "bbox": [ + 130, + 242, + 482, + 459 + ], + "type": "text", + "content": ". However, we find that the RFDN model is quite sensitive to quantization noise. This results in a noticeable drop in the quality when the model is quantized to 8-bit, even in the case of direct convolutions. When the Winograd algorithm is used, this quality drop is somewhat worsened. One of the reasons for this is that the quantization scheme adapted from Ref. [17] includes redundant quantization steps (see discussion in Sec. 3.1). We also revealed a strong effect of replacing the " + }, + { + "bbox": [ + 130, + 242, + 482, + 459 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 130, + 242, + 482, + 459 + ], + "type": "text", + "content": " direct convolutions nearest to the contrast aware channel attention (CCA) layer [21, 29] with the " + }, + { + "bbox": [ + 130, + 242, + 482, + 459 + ], + "type": "inline_equation", + "content": "F4" + }, + { + "bbox": [ + 130, + 242, + 482, + 459 + ], + "type": "text", + "content": " convolutions on the enhancement of the PSNR drop. The latter requires searching for a more careful strategy to replace direct convolutions with the Winograd ones, and, a possible solution is to use for some layers the " + }, + { + "bbox": [ + 130, + 242, + 482, + 459 + ], + "type": "inline_equation", + "content": "F2" + }, + { + "bbox": [ + 130, + 242, + 482, + 459 + ], + "type": "text", + "content": " Winograd algorithm, which is not prone to strong quantization errors. Despite the above-mentioned drop of PSNR, our method produces super-resolution images that are visually competitive with ones" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 490, + 218, + 620 + ], + "blocks": [ + { + "bbox": [ + 133, + 490, + 218, + 620 + ], + "lines": [ + { + "bbox": [ + 133, + 490, + 218, + 620 + ], + "spans": [ + { + "bbox": [ + 133, + 490, + 218, + 620 + ], + "type": "image", + "image_path": "a23a3e3e6baa137235ca21618f69e4336c2bbe7199d6404c0e8d24c7c4abf8f4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 620, + 215, + 630 + ], + "lines": [ + { + "bbox": [ + 136, + 620, + 215, + 630 + ], + "spans": [ + { + "bbox": [ + 136, + 620, + 215, + 630 + ], + "type": "text", + "content": "(a) ESPCN F6 SOTA" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 639, + 482, + 662 + ], + "lines": [ + { + "bbox": [ + 131, + 639, + 482, + 662 + ], + "spans": [ + { + "bbox": [ + 131, + 639, + 482, + 662 + ], + "type": "text", + "content": "Fig. 5: " + }, + { + "bbox": [ + 131, + 639, + 482, + 662 + ], + "type": "inline_equation", + "content": "\\times 3" + }, + { + "bbox": [ + 131, + 639, + 482, + 662 + ], + "type": "text", + "content": " Super-resolution results for \"Comic\" from Set14 within 8-bit ESPCN (first two panels) and ESPCN-6 (last two panels) model using the " + }, + { + "bbox": [ + 131, + 639, + 482, + 662 + ], + "type": "inline_equation", + "content": "F6" + }, + { + "bbox": [ + 131, + 639, + 482, + 662 + ], + "type": "text", + "content": " Winograd convolutions." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 220, + 490, + 306, + 620 + ], + "blocks": [ + { + "bbox": [ + 220, + 490, + 306, + 620 + ], + "lines": [ + { + "bbox": [ + 220, + 490, + 306, + 620 + ], + "spans": [ + { + "bbox": [ + 220, + 490, + 306, + 620 + ], + "type": "image", + "image_path": "c4473c1fd6faffe3c7aadd45833277d65805669eb4cb92424e7e2ac38a0484c9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 620, + 299, + 630 + ], + "lines": [ + { + "bbox": [ + 228, + 620, + 299, + 630 + ], + "spans": [ + { + "bbox": [ + 228, + 620, + 299, + 630 + ], + "type": "text", + "content": "(b) ESPCN F6 Our" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 490, + 393, + 620 + ], + "blocks": [ + { + "bbox": [ + 309, + 490, + 393, + 620 + ], + "lines": [ + { + "bbox": [ + 309, + 490, + 393, + 620 + ], + "spans": [ + { + "bbox": [ + 309, + 490, + 393, + 620 + ], + "type": "image", + "image_path": "c44f1fb0b683444595771d76de3a7dc328b4bfb4f5a3d52e19153c0923bda86d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 620, + 392, + 630 + ], + "lines": [ + { + "bbox": [ + 309, + 620, + 392, + 630 + ], + "spans": [ + { + "bbox": [ + 309, + 620, + 392, + 630 + ], + "type": "text", + "content": "(c) ESPCN-6 F6 SOTA" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 395, + 491, + 480, + 620 + ], + "blocks": [ + { + "bbox": [ + 395, + 491, + 480, + 620 + ], + "lines": [ + { + "bbox": [ + 395, + 491, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 395, + 491, + 480, + 620 + ], + "type": "image", + "image_path": "96dc48826f75214e6f0689587632e384ac07f8d691bd12de2172c9ccbf18103a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 398, + 620, + 480, + 630 + ], + "lines": [ + { + "bbox": [ + 398, + 620, + 480, + 630 + ], + "spans": [ + { + "bbox": [ + 398, + 620, + 480, + 630 + ], + "type": "text", + "content": "(d) ESPCN-6 F6 Ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "type": "text", + "content": "Quantization-Friendly Winograd Transformations for CNN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "of the FP model (see Fig. 4) and are free of strong noise or gridding artifacts that typically occur in the case of standard Winograd transformations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": "For " + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "inline_equation", + "content": "F6" + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": ", despite the considerable improvements achieved by using the transformations " + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{pso}}" + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": " (see Tab. 3 and Fig. 5), the quality drop is still pronounced. In particular, we find that the training of the RFDN model is unstable for " + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "inline_equation", + "content": "F6" + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": " (for this reason, the corresponding results are not reported here). However, we emphasize that here we considered the extreme case when all direct convolutions of kernel size " + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": " are replaced by the " + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "inline_equation", + "content": "F6" + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": " convolutions. We believe it is possible to selectively replace error-tolerant direct convolutions with " + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "inline_equation", + "content": "F6" + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": " ones based on our " + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{pso}}" + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": " with minimal quality losses. This can be achieved, for example, by adapting frameworks for automatic design of neural network architectures [10, 15, 17, 46]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 262, + 230, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 230, + 275 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 230, + 275 + ], + "type": "text", + "content": "4.3 Segmentation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 280, + 482, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 280, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 482, + 435 + ], + "type": "text", + "content": "Finally, we briefly discuss segmentation tasks. To this end, we consider the ENet model [38] which is the optimized semantic segmentation network designed to run on mobile devices. We replace all " + }, + { + "bbox": [ + 130, + 280, + 482, + 435 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 280, + 482, + 435 + ], + "type": "text", + "content": " kernel direct convolutions with the Winograd convolutions, except ones for the last and first layers, and convolutions with a stride equal to 2. The latter is because the standard Winograd algorithm cannot be directly applied to convolutions with non-unit stride. However, we note that several adaptations of the Winograd algorithm to non-unit stride cases have recently been proposed [20,37,48]. Their quantization-friendly formulation within an optimization procedure similar to the one developed in the present study is of great interest for further research. The results of training the ENet model on the CamVid dataset [9] are shown in Tab. 3. As for SR tasks, we find that our " + }, + { + "bbox": [ + 130, + 280, + 482, + 435 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{pso}}" + }, + { + "bbox": [ + 130, + 280, + 482, + 435 + ], + "type": "text", + "content": "-based approach is considerably more accurate compared to the SOTA approach utilizing the standard Winograd transformations " + }, + { + "bbox": [ + 130, + 280, + 482, + 435 + ], + "type": "inline_equation", + "content": "\\Theta^{\\mathrm{std}}" + }, + { + "bbox": [ + 130, + 280, + 482, + 435 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 451, + 226, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 451, + 226, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 451, + 226, + 464 + ], + "type": "text", + "content": "5 Conclusions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 533 + ], + "type": "text", + "content": "We have introduced an efficient way of training 8-bit quantized networks utilizing " + }, + { + "bbox": [ + 130, + 474, + 481, + 533 + ], + "type": "inline_equation", + "content": "F4" + }, + { + "bbox": [ + 130, + 474, + 481, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 474, + 481, + 533 + ], + "type": "inline_equation", + "content": "F6" + }, + { + "bbox": [ + 130, + 474, + 481, + 533 + ], + "type": "text", + "content": " Winograd convolutions. To this end, using the global partial swarm optimization (PSO) method, we have derived a model- and data-free form of the Winograd transformation matrices " + }, + { + "bbox": [ + 130, + 474, + 481, + 533 + ], + "type": "inline_equation", + "content": "\\Theta \\equiv \\{A,B,G\\}" + }, + { + "bbox": [ + 130, + 474, + 481, + 533 + ], + "type": "text", + "content": ", which has significantly better accuracy performance compared to the standard ones." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "To further improve accuracy we have modified the previous SOTA training pipeline of Ref. [17], which allows the transformation matrices " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": " to evolve during training, by replacing the standard Winograd transformations used in this pipeline with our PSO-derived matrices. By treating all the steps of Winograd convolution as 8-bit quantized, we have considered classification, superresolution, and segmentation tasks to reveal the benefits of the proposed approach. In all of the experiments, we found that our modification results in a substantial improvement of accuracy and allows to either completely overcome or highly reduce quality drop. As a consequence, our approach significantly outperforms the SOTA method in terms of accuracy, while requiring the same number of operations during network training and retaining the same inference speedup." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 321, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 321, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 321, + 102 + ], + "type": "text", + "content": "V. Protsenko, V. Kryzhanovskiy et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "type": "text", + "content": "1. Agustsson, E., Timofte, R.: Ntire 2017 challenge on single image super-resolution: Dataset and study. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 126-135 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 170, + 480, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 170, + 480, + 201 + ], + "spans": [ + { + "bbox": [ + 138, + 170, + 480, + 201 + ], + "type": "text", + "content": "2. Alam, M.S., Misba, W.A., Atulasimha, J.: Quantized non-volatile nanomagnetic synapse based autoencoder for efficient unsupervised network anomaly detection. arXiv preprint arXiv:2309.06449 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 202, + 480, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 202, + 480, + 234 + ], + "spans": [ + { + "bbox": [ + 138, + 202, + 480, + 234 + ], + "type": "text", + "content": "3. Alam, S.A., Anderson, A., Barabasz, B., Gregg, D.: Winograd convolution for deep neural networks: Efficient point selection. ACM Transactions on Embedded Computing Systems 21(6), 1-28 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 235, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 235, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 235, + 480, + 266 + ], + "type": "text", + "content": "4. Anderson, A., Gregg, D.: Optimal dnn primitive selection with partitioned boolean quadratic programming. In: Proceedings of the 2018 International Symposium on Code Generation and Optimization. pp. 340-351 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 267, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 480, + 310 + ], + "type": "text", + "content": "5. Andri, R., Bussolino, B., Cipolletta, A., Cavigelli, L., Wang, Z.: Going further with winograd convolutions: Tap-wise quantization for efficient inference on 4x4 tiles. In: 2022 55th IEEE/ACM International Symposium on Microarchitecture (MICRO). pp. 582-598. IEEE (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 311, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 311, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 311, + 480, + 331 + ], + "type": "text", + "content": "6. Barabasz, B.: Quantaized winograd/toom-cook convolution for dnns: Beyond canonical polynomials base. arXiv preprint arXiv:2004.11077 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 332, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 480, + 364 + ], + "type": "text", + "content": "7. Barabasz, B., Anderson, A., Soodhalter, K.M., Gregg, D.: Error analysis and improving the accuracy of winograd convolution for deep neural networks. ACM Transactions on Mathematical Software (TOMS) 46(4), 1-33 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 365, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 480, + 396 + ], + "type": "text", + "content": "8. Barabasz, B., Gregg, D.: Winograd convolution for dnns: Beyond linear polynomials. In: International Conference of the Italian Association for Artificial Intelligence. pp. 307-320. Springer (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 396, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 480, + 418 + ], + "type": "text", + "content": "9. Brostow, G.J., Fauqueur, J., Cipolla, R.: Semantic object classes in video: A high-definition ground truth database. Pattern Recognition Letters 30(2), 88-97 (2009)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 418, + 480, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 480, + 438 + ], + "type": "text", + "content": "10. Cai, H., Zhu, L., Han, S.: Proxylessnas: Direct neural architecture search on target task and hardware. arXiv preprint arXiv:1812.00332 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 439, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 439, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 138, + 439, + 480, + 471 + ], + "type": "text", + "content": "1. Chen, T., Xu, W., Chen, W., Wang, P., Cheng, J.: Towards efficient and accurate winograd convolution via full quantization. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 472, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 472, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 480, + 504 + ], + "type": "text", + "content": "2. Chen, Y.T., Ou, Y.F., Huang, C.T.: A winograd-based highly-parallel convolution engine for 8-bit cnn acceleration. In: 2022 IEEE 4th International Conference on Artificial Intelligence Circuits and Systems (AICAS). pp. 395–398. IEEE (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 138, + 504, + 480, + 536 + ], + "type": "text", + "content": "3. Chen, Y.H., Krishna, T., Emer, J.S., Sze, V.: Eyeriss: An energy-efficient reconfigurable accelerator for deep convolutional neural networks. IEEE journal of solid-state circuits 52(1), 127-138 (2016)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 536, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 480, + 557 + ], + "type": "text", + "content": "4. Cook, S.: On the minimum computation time for multiplication. Doctoral diss., Harvard U., Cambridge, Mass 1 (1966)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 558, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 558, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 558, + 480, + 578 + ], + "type": "text", + "content": "5. Elsken, T., Metzen, J.H., Hutter, F.: Neural architecture search: A survey. The Journal of Machine Learning Research 20(1), 1997-2017 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 579, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 579, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 579, + 480, + 600 + ], + "type": "text", + "content": "6. Fernández-Marqués, J.: High performance deep learning on resource constrained platforms. Ph.D. thesis, University of Oxford (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "text", + "content": "7. Fernandez-Marques, J., Whatmough, P., Mundy, A., Mattina, M.: Searching for winograd-aware quantized networks. Proceedings of Machine Learning and Systems 2, 14-29 (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "8. Gholami, A., Kim, S., Dong, Z., Yao, Z., Mahoney, M.W., Keutzer, K.: A survey of quantization methods for efficient neural network inference. In: Low-Power Computer Vision, pp. 291-326. Chapman and Hall/CRC (2022)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "type": "text", + "content": "Quantization-Friendly Winograd Transformations for CNN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "19. Holt, J.L., Hwang, J.N.: Finite precision error analysis of neural network hardware implementations. IEEE Transactions on Computers 42(3), 281-290 (1993)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 183 + ], + "type": "text", + "content": "20. Huang, C., Dong, X., Li, Z., Song, T., Liu, Z., Dong, L.: Efficient stride 2 winograd convolution method using unified transformation matrices on fpga. In: 2021 International Conference on Field-Programmable Technology (ICFPT). pp. 1-9. IEEE (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 480, + 217 + ], + "type": "text", + "content": "21. Hui, Z., Gao, X., Yang, Y., Wang, X.: Lightweight image super-resolution with information multi-distillation network. In: Proceedings of the 27th acm international conference on multimedia. pp. 2024-2032 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 480, + 239 + ], + "type": "text", + "content": "22. Jorda, M., Valero-Lara, P., Pena, A.J.: Performance evaluation of cudnn convolution algorithms on nvidia volta gpus. IEEE Access 7, 70461-70473 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 480, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 480, + 274 + ], + "type": "text", + "content": "23. Kennedy, J., Eberhart, R.: Particle swarm optimization. In: Proceedings of ICNN'95-international conference on neural networks. vol. 4, pp. 1942-1948. IEEE (1995)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 480, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 480, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 480, + 296 + ], + "type": "text", + "content": "24. Krishnamoorthi, R.: Quantizing deep convolutional networks for efficient inference: A whitepaper. arXiv preprint arXiv:1806.08342 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 297, + 480, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 480, + 329 + ], + "type": "text", + "content": "25. Lavin, A., Gray, S.: Fast algorithms for convolutional neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4013-4021 (2016)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 331, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 480, + 363 + ], + "type": "text", + "content": "26. Li, C., Jiang, P., Zhou, H., Wang, X., Zhao, X.: Haw: Hardware-aware point selection for efficient winograd convolution. IEEE Signal Processing Letters 30, 269-273 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "text", + "content": "27. Li, G., Jia, Z., Feng, X., Wang, Y.: Lowino: Towards efficient low-precision winograd convolutions on modern cpus. In: Proceedings of the 50th International Conference on Parallel Processing. pp. 1-11 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 397, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 441 + ], + "type": "text", + "content": "28. Li, G., Liu, L., Wang, X., Ma, X., Feng, X.: Lance: efficient low-precision quantized winograd convolution for neural networks based on graphics processing units. In: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 3842-3846. IEEE (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 442, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 480, + 475 + ], + "type": "text", + "content": "29. Liu, J., Tang, J., Wu, G.: Residual feature distillation network for lightweight image super-resolution. In: Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 41-55. Springer (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 475, + 480, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 480, + 497 + ], + "type": "text", + "content": "30. Liu, Z.G., Mattina, M.: Efficient residue number system based winograd convolution. In: European Conference on Computer Vision. pp. 53-68. Springer (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 498, + 480, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 480, + 519 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 480, + 519 + ], + "type": "text", + "content": "31. Mathieu, M., Henaff, M., LeCun, Y.: Fast training of convolutional networks through fts. arXiv preprint arXiv:1312.5851 (2013)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 521, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 480, + 542 + ], + "type": "text", + "content": "32. Meng, L., Brothers, J.: Efficient winograd convolution via integer arithmetic. arXiv preprint arXiv:1901.01965 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 544, + 480, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 480, + 565 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 480, + 565 + ], + "type": "text", + "content": "33. Miranda, L.J.: Pyswarms: a research toolkit for particle swarm optimization in python. Journal of Open Source Software 3(21), 433 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 566, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 566, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 566, + 480, + 620 + ], + "type": "text", + "content": "34. Mori, P., Frickenstein, L., Sampath, S.B., Thoma, M., Faspous, N., Vemparala, M.R., Frickenstein, A., Unger, C., Stechele, W., Mueller-Gritschneider, D., et al.: Wino vidi vici: Conquering numerical instability of 8-bit winograd convolution for accurate inference acceleration on edge. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 53-62 (2024)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 480, + 665 + ], + "type": "text", + "content": "35. Mori, P., Sampath, S.B., Frickenstein, L., Vemparala, M.R., Faspous, N., Frickenstein, A., Stechele, W., Passerone, C.: Winotrain: Winograd-aware training for accurate full 8-bit convolution acceleration. In: 2023 60th ACM/IEEE Design Automation Conference (DAC). pp. 1-6. IEEE (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 321, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 321, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 321, + 101 + ], + "type": "text", + "content": "V. Protsenko, V. Kryzhanovskiy et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 544 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "36. Nagel, M., Fournarakis, M., Amjad, R.A., Bondarenko, Y., Van Baalen, M., Blankevoort, T.: A white paper on neural network quantization. arXiv preprint arXiv:2106.08295 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "37. Pan, J., Chen, D.: Accelerate non-unit stride convolutions with winograd algorithms. In: Proceedings of the 26th Asia and South Pacific Design Automation Conference. pp. 358-364 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "type": "text", + "content": "38. Paszke, A., Chaurasia, A., Kim, S., Culurciello, E.: Enet: A deep neural network architecture for real-time semantic segmentation. arXiv preprint arXiv:1606.02147 (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 237 + ], + "type": "text", + "content": "39. Shafique, M.A., Munir, A., Kong, J.: Deep learning performance characterization on gpus for various quantization frameworks. AI 4(4), 926-948 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 237, + 482, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 281 + ], + "type": "text", + "content": "40. Shi, B., Zhang, J., He, Z., Wei, X., Li, S., Luo, G., Zheng, H., Xie, Y.: Efficient super-resolution system with block-wise hybridization and quantized winograd on fpga. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 281, + 482, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 326 + ], + "type": "text", + "content": "41. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1874-1883 (2016)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 326, + 482, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 482, + 357 + ], + "type": "text", + "content": "42. Toom, A.L.: The complexity of a scheme of functional elements simulating the multiplication of integers. In: Doklady Akademii Nauk. vol. 150, pp. 496-498. Russian Academy of Sciences (1963)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 358, + 482, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 390 + ], + "type": "text", + "content": "43. Vasilache, N., Johnson, J., Mathieu, M., Chintala, S., Piantino, S., LeCun, Y.: Fast convolutional nets with fbfft: Agpu performance evaluation. arXiv preprint arXiv:1412.7580 (2014)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 391, + 482, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 413 + ], + "type": "text", + "content": "44. Vincent, K., Stephano, K., Frumkin, M., Ginsburg, B., Demouth, J.: On improving the numerical stability of winograd convolutions (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 413, + 482, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 435 + ], + "type": "text", + "content": "45. Wang, D., Tan, D., Liu, L.: Particle swarm optimization algorithm: an overview. Soft computing 22, 387-408 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 435, + 482, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 468 + ], + "type": "text", + "content": "46. Wang, K., Liu, Z., Lin, Y., Lin, J., Han, S.: Haq: Hardware-aware automated quantization with mixed precision. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8612-8620 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 468, + 455, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 455, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 455, + 479 + ], + "type": "text", + "content": "47. Winograd, S.: Arithmetic complexity of computations, vol. 33. Siam (1980)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 479, + 482, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 511 + ], + "type": "text", + "content": "48. Yepez, J., Ko, S.B.: Stride 2 1-d, 2-d, and 3-d winograd for convolutional neural networks. IEEE Transactions on Very Large Scale Integration (VLSI) Systems 28(4), 853-863 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 511, + 482, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 511, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 511, + 482, + 544 + ], + "type": "text", + "content": "49. Zhang, Y., Wang, S., Ji, G., et al.: A comprehensive survey on particle swarm optimization algorithm and its applications. Mathematical problems in engineering 2015 (2015)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "type": "text", + "content": "Quantization-Friendly Winograd Transformations for CNN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Quantized Prompt for Efficient Generalization of Vision-Language Models/d406d327-feb0-4aee-91eb-de3562f63166_content_list.json b/2024/Quantized Prompt for Efficient Generalization of Vision-Language Models/d406d327-feb0-4aee-91eb-de3562f63166_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5dbbf229a93548575473c8d147a3ce25a70108bb --- /dev/null +++ b/2024/Quantized Prompt for Efficient Generalization of Vision-Language Models/d406d327-feb0-4aee-91eb-de3562f63166_content_list.json @@ -0,0 +1,1833 @@ +[ + { + "type": "text", + "text": "Quantized Prompt for Efficient Generalization of Vision-Language Models", + "text_level": 1, + "bbox": [ + 222, + 140, + 782, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tianxiang Hao $^{1,2}$ , Xiaohan Ding $^{3(\\boxtimes)}$ , Juexiao Feng $^{1,2}$ , Yuhong Yang $^{1,2}$ , Hui Chen $^{2}$ , and Guiguang Ding $^{1,2(\\boxtimes)}$", + "bbox": [ + 243, + 210, + 759, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 School of Software, Tsinghua University, Beijing, China \n2 BNrist, Beijing, China \n3 Bytedance", + "bbox": [ + 307, + 253, + 694, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{beyondtx,xiaohding,suise.con,jichenhui}@gmail.com \nfjx20@mails.tsinghua.edu.cn, dinggg@tsinghua.edu.cn", + "bbox": [ + 297, + 297, + 702, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In the past few years, large-scale pre-trained vision-language models like CLIP have achieved tremendous success in various fields. Naturally, how to transfer the rich knowledge in such huge pre-trained models to downstream tasks and datasets becomes a hot topic. During downstream adaptation, the most challenging problems are overfitting and catastrophic forgetting, which can cause the model to overly focus on the current data and lose more crucial domain-general knowledge. Existing works use classic regularization techniques to solve the problems. As solutions become increasingly complex, the ever-growing storage and inference costs are also a significant problem that urgently needs to be addressed. While in this paper, we start from an observation that proper random noise can suppress overfitting and catastrophic forgetting. Then we regard quantization error as a kind of noise, and explore quantization for regularizing vision-language model, which is quite efficiency and effective. Furthermore, to improve the model's generalization capability while maintaining its specialization capacity at minimal cost, we deeply analyze the characteristics of the weight distribution in prompts, conclude several principles for quantization module design and follow such principles to create several competitive baselines. The proposed method is significantly efficient due to its inherent lightweight nature, making it possible to adapt on extremely resource-limited devices. Our method can be fruitfully integrated into many existing approaches like MaPLe, enhancing accuracy while reducing storage overhead, making it more powerful yet versatile. Extensive experiments on 11 datasets shows great superiority of our method sufficiently. Code is available at github", + "bbox": [ + 261, + 354, + 743, + 703 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Quantization $\\cdot$ Generalization $\\cdot$ Vision-language model", + "bbox": [ + 261, + 713, + 715, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 750, + 377, + 766 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, deep learning models and related technologies have seen rapid development [9, 11-13, 22-24, 49, 50, 65, 66]. Vision-language model (VLM) like CLIP [58] is one of the hottest research topics and leads to huge success. The excellent generalization ability is a crucial cornerstone of such achievement [26, 53, 58, 59, 68].", + "bbox": [ + 212, + 779, + 787, + 842 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c3201a6a53517b9e1d48b08e43469486061b1f3fbbebf67fdcf5733f50b11212.jpg", + "image_caption": [ + "Fig. 1: Overall performance comparison with existing vision-language tuning methods. Our method outperforms all of the state-of-the-art competitors with significantly fewer storage space. Based on the proposed quantization algorithm, our method could be integrated into many of the existing methods and bring consistent improvements with excellent efficiency." + ], + "image_footnote": [], + "bbox": [ + 323, + 143, + 669, + 368 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "When people have access to downstream data, it is better to tune the pretrained VLM on the target dataset for higher accuracy. However, such full finetuning could easily cause the model to overfit the small downstream dataset and face catastrophic forgetting problem, leading to severe performance drop.", + "bbox": [ + 212, + 468, + 787, + 530 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To solve the problem, in this paper we will start from rethinking the relationship between noise and generalization. We propose to think noise as a kind of regularization techniques, which may be helpful for alleviating overfitting and catastrophic forgetting problem. As in Sec. 3.2 and Fig. 2, we then find that directly adding some random Gaussian noise to the tunable prompts of vision-language models would result in a performance gain in several cases, which partly validates our conjecture about using noise as a form of regularization. In particular, excessive noise diminishes the model's adaptation capability, while insufficient noise fails to provide effective regularization. Only noise of moderate intensity is beneficial for the model's generalization.", + "bbox": [ + 212, + 532, + 787, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, Gaussian noise is absolutely random and hard to control, and thus it is difficult for us to take full advantage of such type of noise to benefit generalization. Instead, we point out that quantization error is also a form of noise, and therefore, it is also possible to leverage quantization error to enhance the model's generalization performance. With this idea in mind, we thoroughly analyzed the distribution pattern of prompt weights in Sec. 3.3 and derived several design principles for quantization algorithms based on the observed phenomena. In Sec. 4, following the principles we summarized, we successfully designed an efficient quantization-aware training algorithm, which largely enhances the model's generalization ability while quantizing.", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "T. Hao et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since our design is general, we could fruitfully integrate our quantization strategy into many existing methods and reach a higher accuracy with much smaller model size due to the lightweight nature of quantization. For example, we integrate our method onto an existing popular method MaPLe, and our QMaPLe earns $0.57\\%$ accuracy improvement with only $0.25 \\times$ size.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In conclusion, we summarize our contribution as follows:", + "bbox": [ + 238, + 222, + 647, + 234 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We deeply analyze the effect of noise and rethink the relationship between noise and generalization in Sec. 3.2 for vision-language models. As a result, we confirm that moderate noise would promote the model generalization", + "- We are the first to propose to quantize prompts. By detailed observation and hard thinking, we conclude several principles about how to effectively design a quantization method for the prompts of vision-language models.", + "- Following the principles we concluded, we build our method and successfully quantize the prompts as well as some other weights, and our method significantly outperforms existing ones as in Fig. 1. Extensive experiments show great power of our method. In base-to-new generalization, domain generalization, cross-dataset transfer and few-shot learning settings, we consistently reach competitive results, winning many state-of-the-art tuning methods with a much smaller size of model." + ], + "bbox": [ + 225, + 248, + 782, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 469, + 395, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Vision-Language Models", + "text_level": 1, + "bbox": [ + 215, + 503, + 462, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In recent times, large-scale vision-language models have demonstrated remarkable performance across various tasks. Seminal works such as [26, 58, 68, 70, 73]. Classic works have focused on learning multimodal representations through self-supervised methods using extensive sets of image-text pairs. Among these, CLIP [58] stands out as a milestone achievement, employing contrastive learning to align vision and language representations and achieving exceptional performance. A well-trained vision-language model is invaluable, offering substantial support to various fields. Successful applications of these robust models include few-shot recognition [76, 77], detection tasks [15, 51, 60, 72], and segmentation tasks [7, 36, 47, 59]. Furthermore, for video data, research efforts have emerged in video classification [57] and video understanding [30].", + "bbox": [ + 212, + 531, + 782, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Parameter-Efficient Fine-Tuning", + "text_level": 1, + "bbox": [ + 215, + 720, + 527, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, a series of works [19, 21, 23, 24, 27, 29, 38, 41, 44, 48, 64, 76, 77] have been proposed to help transfer the learnt knowledge, where one of the most popular field is parameter-efficient fine-tuning (PEFT). PEFT aims to transfer a pre-trained model to downstream tasks by a minimum number of parameters. Originating from natural language processing tasks, classic methods like adapter [23], prompt tuning [28, 35, 38, 42, 43, 61] and LoRA [24] follows a similar", + "bbox": [ + 212, + 750, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Quantized Prompt for Efficient Generalization of Vision-Language Models", + "bbox": [ + 236, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "principle to add extra modules with a small number of parameters into the backbone model, freeze the original parameters and only tune and store the newly added parameters. Inspired by the success of PEFT in language field, researchers have extended such kind of approaches to adapt visual models in a similar fashion [6,19,27,41,74]. In the field of vision-language modeling, several explorations have been made as well. Bahng et al. [1] exclusively apply prompt tuning to the image encoder. CoOp [77] replaces the fixed template in CLIP [58] with tunable text prompts. CoCoOp [76] leverages image features to guide the optimization of tunable text prompts in CoOp. Other works [32, 34] optimize both image and text prompts simultaneously and establish additional connections between different modalities. To mitigate overfitting and catastrophic forgetting, various works [5,33,67,75] integrate regularization modules or losses into prompt tuning.", + "bbox": [ + 212, + 146, + 787, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Quantization", + "text_level": 1, + "bbox": [ + 215, + 351, + 372, + 367 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Quantization is one of the most effective compression methods [4,8,10,20,37, 46,69] for deep learning models. Generally, parameters such as weights and activations are typically stored as 32-bit floating-point numbers, which consume a significant amount of memory and require intensive computation during inference. Quantization [17, 18, 55] involves representing these parameters with reduced precision, such as 8-bit integers or even lower bit-widths. By doing so, quantization can significantly reduce the memory footprint and computational complexity of the model without significantly sacrificing accuracy. Quantization methods can be divided into two groups, Post-Training Quantization (PTQ) [2,16,25,40,45,52,54,71] that consumes few resources but suffers higher accuracy loss, and Quantization-Aware Training (QAT) [3,14,31,39,56] that relies on plenty of resources for training and shows better accuracy. Existing works aim to minimize quantization error to improve accuracy, while our work demonstrates that both excessive and insufficient errors are detrimental to model generalization. To achieve optimal generalization performance, a moderate error is required.", + "bbox": [ + 212, + 378, + 787, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Exploring Quantization in Model Generalization", + "text_level": 1, + "bbox": [ + 215, + 643, + 722, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminaries: Prompt Tuning of Vision-Language Models", + "text_level": 1, + "bbox": [ + 215, + 676, + 740, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CLIP comprises a text encoder $\\mathcal{L}$ and an image encoder $\\mathcal{V}$ . Typically, $\\mathcal{L}$ is implemented as a language transformer, whereas $\\mathcal{V}$ may be realized using either a convolutional neural network or a vision transformer. In this study, following the methodologies outlined by [76, 77], we employ a ViT-B/16 model [13] as the image encoder $\\mathcal{V}$ , except where otherwise specified. The subsequent sections will provide a brief overview of the methods used to prompt CLIP for prediction tasks.", + "bbox": [ + 212, + 703, + 787, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Text Encoder Consider a text encoder composed of $M$ layers. For the $k$ -th layer, denoted as $\\mathcal{L}_k$ , the inputs consist of a sequence of prompt tokens $P_{k - 1}^{l}$", + "bbox": [ + 212, + 809, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "T. Hao et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and a $[CLS]$ token $c_{k-1}^{l}$ , while the outputs are represented by $P_{k}^{l}$ and $c_{k}^{l}$ . The initial inputs, $P_{0}^{l}$ and $c_{0}^{l}$ , correspond to the word embeddings of the prompts combined with the label, such as \"A photo of a $[CLS]''$ or alternatively, some randomly initialized vectors. Formally, we denote $P_{k}^{l} \\in \\mathbb{R}^{n^{l} \\times d^{l}}$ and $c_{k}^{l} \\in \\mathbb{R}^{d^{l}}$ , where $n^{l}$ signifies the length of the text prompts and $d^{l}$ represents the dimension of the word embeddings. For each layer, $1 \\leq k \\leq M$ , the relationship is given by $[P_{k}^{l}, c_{k}^{l}] = \\mathcal{L}_{k}([P_{k-1}^{l}, c_{k-1}^{l}])$ . The output feature of the text encoder $f^{l} \\in \\mathbb{R}^{d^{v}}$ , where $d^{v}$ is the dimension of the visual feature space, is calculated by projecting the $[CLS]$ token of its last layer to the visual latent space through a linear transformation, i.e. $f^{l} = \\mathrm{Proj}(c_{M}^{l})$ .", + "bbox": [ + 212, + 145, + 787, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image Encoder Suppose there are $N$ layers in the image encoder. For $k$ -th layer $\\mathcal{V}_k$ , the inputs are a series of image tokens $I_{k-1}$ , a classification token $c_{k-1}^v$ and prompt tokens $P_{k-1}^v$ , and the outputs are $I_k$ , $c_k^v$ and $P_{k}^{v}$ . The inputs of the first layer $I_0$ and $c_0^v$ are the patch embeddings of the input image and the pre-trained class token. $P_0^v$ is randomly initialized. Formally, $I_k \\in \\mathbb{R}^{p \\times d^v}$ , $c_k^v \\in \\mathbb{R}^{d^v}$ and $P_k^v \\in \\mathbb{R}^{n^v \\times d^v}$ , where $p$ denotes the number of image patches and $d^v$ denotes the dimension of visual embedding. $\\forall 1 \\leq k \\leq N$ , $[P_k^v, c_k^v, I_k] = \\mathcal{V}_k([P_{k-1}^v, c_{k-1}^v, I_{k-1}])$ . The output feature of the image encoder is $f^v = c_N^v$ .", + "bbox": [ + 212, + 303, + 787, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prediction For image classification, suppose there are $C$ classes, and $\\{f_c^l\\}_{c = 1}^C$ are the text features. Label $y$ 's probability is $p(y|f^v) = \\frac{\\exp(\\sin(f^v,f_l^i) / \\tau)}{\\sum_{c = 1}^{C}\\exp(\\sin(f^v,f_c^l) / \\tau)}$ where $\\sin (\\cdot ,\\cdot)$ denotes cosine similarity function and $\\tau$ is temperature. The final prediction is $\\hat{z} = \\arg \\max_{1\\leq y < C}(p(y|f^v))$ .", + "bbox": [ + 212, + 422, + 787, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We have introduced shallow prompts. There are also different types of prompts. Several works [27, 32] use them for improve performance. They directly add and tune the prompt in each layer in the feature encoder, instead of inheriting the output prompt calculated by the last encoder. Now we have $[\\_ ,c_k^l ] = \\mathcal{L}_k([P_{k - 1}^l,c_{k - 1}^l])$ and $[\\_ ,c_k^v,I_k] = \\mathcal{V}_k([P_{k - 1}^v,c_{k - 1}^v,I_{k - 1}])$ . Note that $P^l$ and $P^v$ are independent tunable parameters. They are no longer determined by the previous layer.", + "bbox": [ + 212, + 502, + 792, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Rethinking the Relationship between Noise and Generalization", + "text_level": 1, + "bbox": [ + 212, + 632, + 781, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Some existing work has explored how to use the addition of noise to suppress model overfitting, e.g., techniques like Dropout [62] that randomly drops connections, and random jitter that directly introduce perturbations in the input data. However, the impact of directly adding noise to model weights has been less explored, especially in the context of prompt structures, which have become popular only in recent years, in Transformer architectures. The variations of the specialization capability represented by the test accuracy on base, seen classes and the generalization capability represented by the test accuracy on new, unseen classes. The curves of different colors represent the data under the influence of random Gaussian noise of different intensities, e.g. \"Noise_0.01\" adds random noise with a distribution of $\\mathcal{N}(0,0.01^2)$ to the prompt. \"Noise_0\" denotes the baseline prompt tuning. As training progresses, the generalization capability", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Quantized Prompt for Efficient Generalization of Vision-Language Models", + "bbox": [ + 236, + 113, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/795d95e07398067757fe3533b9326a893b3088234d8b65bd463502db3a4d4768.jpg", + "image_caption": [ + "Fig. 2: The variations of the specialization capability represented by the test accuracy on base, seen classes and the generalization capability represented by the test accuracy on new, unseen classes on average of the ten datasets in base-to-new generalization setting in Sec. 5 except for time-consuming ImageNet. The curves of different colors represent the data under the influence of random Gaussian noise of different intensities, e.g. \"Noise_0.01\" adds random noise with a distribution of $\\mathcal{N}(0,0.01^2)$ to the prompt. \"Noise_0\" denotes the baseline prompt tuning. As training progresses, the generalization capability of baseline prompt tuning continuously decreases while the specialization capability improves. Therefore, we expect that adding noise can achieve a better balance between generalization and specialization. However, excessive noise, e.g. 0.1, greatly diminishes the model's specialization capability, while insufficient noise, e.g. 0.001, fails to provide effective regularization. Only noise of moderate intensity outperforms baseline in specialization-generalization trade-off, effectively enhancing the unseen class accuracy without significantly compromising seen class accuracy." + ], + "image_footnote": [], + "bbox": [ + 250, + 159, + 472, + 290 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a10f97832116346084f99cc7315ae08e98e1238496ca61e2de859f49045f0004.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 159, + 733, + 290 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of baseline continuously decreases while the specialization capability improves. Therefore, we expect that adding noise can achieve a better balance between generalization and specialization. However, excessive noise, e.g. 0.1, greatly diminishes the model's specialization capability, while insufficient noise, e.g. 0.001, fails to provide effective regularization. Only noise of moderate intensity outperforms baseline in specialization-generalization trade-off, effectively enhancing the unseen class accuracy without significantly compromising seen class accuracy.", + "bbox": [ + 212, + 505, + 787, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantization, the technique of mapping parameter values from high precision to low precision, can also be viewed as introducing some form of noise into the parameters, and thus can possibly improve the generalizability as the noise did in Fig. 2. Compared to Gaussian noise, quantization error is more controllable. Better still, quantization also has another major advantage: it can significantly reduce the storage required for parameters. Therefore, using quantization algorithms to generalize vision-language models is a very promising direction.", + "bbox": [ + 212, + 611, + 787, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Characteristics of Prompts in Vision-Language Models", + "text_level": 1, + "bbox": [ + 214, + 739, + 714, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this subsection, we will demonstrate the characteristics of prompts in the vision-language model to facilitate targeted design of quantization schemes.", + "bbox": [ + 212, + 763, + 785, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We start from analyzing the training procedure of CoOp [77], which freezes all the parameters in backbone and tunes the added prompts only. The histogram about the frequency of weights of prompts for different training epochs is shown", + "bbox": [ + 212, + 795, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "T. Hao et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/59081e3f35601fe0e14fbb29ce3f8e076fc28a0420491ae46b6b8133466ef4e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 146, + 321, + 208 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ee31c05f39048038f87beb9c76025c27ad8e3253677f963621a4f4cd9f886232.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 146, + 433, + 208 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0cf7047e2eaa60eec58a465de3d9f36f6d7b6dcf67c9c92396088a36aa2db1e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 446, + 146, + 547, + 208 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4de6c6cf6f2747070b02fb9ca1b9f14a279485613c55e3d7986e4ac1522bf200.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 146, + 660, + 208 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4bbc9d7a2ee54cd2750e0c5bce3f4d178bb00f64776556f4e55093f79571ce50.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 673, + 146, + 774, + 208 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1df42b99b9e7efb8e5bcd7c80663a1d48ca5312d5463ab5b671cabffe9ab2275.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 210, + 321, + 271 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/454cd5847e87043494fec38d3880f7a9ee648d62b74d6b8a181602bc574b86d9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 333, + 210, + 433, + 271 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/66bf6a9ede6a2b896b2ece0985627f28319370c1269d79491dafcd28c98b3c92.jpg", + "image_caption": [ + "Fig. 3: The histogram about the frequency of weights of prompts during training of CoOp [77] on eurosat dataset. We find the shape of the prompt's weight distribution remains largely unchanged throughout the entire training, but the variance of the weight distribution increases rapidly at the beginning of training. Additionally, we also notice that there are almost no outliers in the prompt's weights throughout the entire training process, and apart from the unstable initial training phase, the changes in weights between adjacent phases are not significant, indicating a very gentle overall updating trend." + ], + "image_footnote": [], + "bbox": [ + 446, + 210, + 547, + 272 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/583972c5c62d0a30dbcda92ce9ce68fd94e26db29ed99bed01c2483148f9cf8e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 210, + 660, + 272 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/90e6e9dc8d9477c7e68bad3002e0f806846119b45a69f4a7ccfbfb5defca68f7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 673, + 210, + 774, + 272 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in Fig. 3. We observe that the shape of the prompt's weight distribution remains mostly consistent throughout the entire training process. However, the variance of the weight distribution increases rapidly at the initial stages of training. Refer to Fig. 6 for more details. Moreover, we notice minimal presence of outliers in the prompt's weights throughout the training process. Apart from the unstable initial training phase, the changes in weights between consecutive phases are not substantial, indicating a very gentle overall updating trend.", + "bbox": [ + 212, + 402, + 784, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Taking into account the observations from Secs. 3.2 and 3.3, we can draw the following conclusions:", + "bbox": [ + 212, + 508, + 785, + 539 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Because the prompt weights are not sensitive to noise, and moderate noise in the current scenario not only does not degrade the model's performance but actually enhances its generalization ability, we can possibly adopt some quantization strategies that are considered very aggressive and highly likely to significantly degrade performance, such as 1-bit quantization, without causing destructive effects on the model's generalization ability.", + "- Given that the parameters targeted for quantization in the current scenario constitute only a tiny fraction of the total model parameters, we speculate that QAT may have more advantages over PTQ, as any additional training costs incurred by QAT which is considered a major drawback in traditional situations would naturally be kept at a low level due to the low proportion of parameters targeted for quantization. Due to the observation in Fig. 2 that both excessive and insufficient noise are detrimental to generalization, in designing the QAT algorithm, we no longer need to always minimize quantization error as the objective, as in classic algorithms. Instead, we aim to maintain the quantization error at a moderate level. Furthermore, as seen in Fig. 3, the changes in prompt weight distribution during most of the training time are gentle, indicating that we can appropriately continue to use past-time algorithm state without the need for real-time updates." + ], + "bbox": [ + 225, + 551, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Quantized Prompt for Efficient Generalization of Vision-Language Models", + "bbox": [ + 238, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/85258a1dff8fc578d8697103b18340416c413713e8e0e96b55961cd5383fae1a.jpg", + "image_caption": [ + "Fig. 4: Overview of Quantized Prompt. We set $b = 2$ for a clear explanation. Normalization and denormalization processes are not shown here." + ], + "image_footnote": [], + "bbox": [ + 271, + 143, + 718, + 330 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Throughout the training process, the shape of the prompt weight distribution remains roughly unchanged, with the main variation coming from the distribution's variance. Therefore, it may be beneficial to attempt to eliminate the translation and scaling transformations of the distribution by normalizing the distribution before quantization and denormalizing it after quantization, which might help improve quantization accuracy.", + "- Since there are almost no outliers throughout the entire training process, some classic algorithms are revitalized. Specifically, clustering algorithms like K-Means are good choices, because one of the major drawbacks of them in traditional usage is that they are greatly affected by outliers. In addition, as we have analyzed before, there is not a high demand for dynamically minimizing quantization error in vision-language generalization, opting for a more efficient non-parametric clustering approach seems a better choice compared with some complex parametric approaches." + ], + "bbox": [ + 223, + 383, + 784, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Prompt Quantization", + "text_level": 1, + "bbox": [ + 215, + 628, + 459, + 646 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Preliminaries: Quantization Basis", + "text_level": 1, + "bbox": [ + 215, + 667, + 537, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Suppose there are $m$ parameters in total to be quantized, which are denoted by $W \\in \\mathbb{R}^m$ . Each $w_i$ here is a high-precision float-type number. A $b$ -bit model quantization algorithm divides the value range of parameters into $2^b$ intervals $\\{\\mathcal{U}_i, 1 \\leq i \\leq 2^b\\}$ , aiming to find a mapping $Q$ from intervals to points, mapping all values within an interval $\\mathcal{U}_i$ to the same quantized value $q_i$ , i.e. $Q(x) = q_i, \\forall x \\in \\mathcal{U}_i$ . We define the quantization error $E = \\sum_{i=1}^{N} ||Q(W_i) - W_i||^2$ . Such error serves as the objective of K-Means algorithm as well, so it can be ensured that each time clustering is redone using the K-Means algorithm, the quantization error will most likely decrease.", + "bbox": [ + 212, + 702, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "T. Hao et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Overview of Quantized Prompt", + "text_level": 1, + "bbox": [ + 215, + 146, + 519, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Based on the observation and analysis in Sec. 3.3, we build a simple yet quite proper quantization algorithm for generalizing vision-language models. As stated before, we choose a widely-used clustering method K-Means, to construct the mapping $Q$ . Initially, K-Means algorithm is run to fit the pre-trained prompt weights, and record the codebook, i.e. cluster centers. Normalization and denormalization are conducted beyond quantization to alleviate the influence of varying variance. Formally, given the original prompt $W$ with $N$ parameters, we first calculate $\\mu = \\frac{1}{N}\\sum_{i=1}^{N}W_i$ and $\\sigma = \\sqrt{\\frac{\\sum_{1\\leq i\\leq N}(W_i - \\mu)^2}{N}}$ , then normalize $W$ and get $\\hat{W} = \\frac{W - \\mu}{\\sigma}$ . K-Means quantization is applied on $\\hat{W}$ , and the quantized prompt $W_q = \\sigma Q(\\hat{W}) + \\mu$ . Since $Q$ is not differentiable, we adopt a common practice to directly propagate the gradient across the quantization function by Straight-Through Estimator (STE), i.e. $\\frac{\\partial Q(x)}{\\partial x} = x$ .", + "bbox": [ + 212, + 172, + 787, + 366 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The overall framework is shown in Fig. 4. In training, we keep tuning the weight by the fake gradient propagated from quantization operation. For storage, we first convert the fp16 parameters in prompt to $b$ -bit indexes, which could be used to search for the corresponding cluster center in the codebook. In Fig. 4, we set $b = 2$ for a clear explanation. Compared with baseline method, our quantized method could save a lot more storage space. Specifically, for storage, an ordinary method needs $16N$ bits, while ours only needs $bN + 2^b \\times 16$ . In experiments, we usually set $b$ to 1, 2 or 4, which is far more smaller than $N$ . For example, with $b = 1$ , the storage space of our method is roughly $16 \\times$ smaller than baseline.", + "bbox": [ + 212, + 367, + 787, + 502 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Note that the codebook here is updated by a rule called \"constrained adaptive clustering\". We will give a detailed description of it in the following subsection.", + "bbox": [ + 212, + 503, + 785, + 532 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Constrained Adaptive Clustering for Quantization", + "text_level": 1, + "bbox": [ + 214, + 556, + 674, + 571 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In Fig. 4, re-clustering the prompt every iteration like classic QAT algorithms did to keep the codebook updated at all times is not a good choice. There are three reasons: 1. K-Means algorithm is not quite efficient, and thus if we run it every iteration then the training efficiency would decrease; 2. Only moderate noise promotes generalization, so keep updating the clusters to minimize the quantization error may be not helpful but harmful; 3. From Fig. 3, we can observe that the weight changes are very gentle for most of the training time, and the weight distributions of adjacent stages are highly similar. Therefore, even if we update the internal clustering state of K-Means with these already very similar data at each iteration, it's difficult to generate a better clustering solution, which is just futile effort.", + "bbox": [ + 212, + 583, + 787, + 748 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We propose constrained adaptive clustering to instruct the update of parameters inside K-Means which would not updated by gradient at all. Intuitively, first, we do not want too often update, so we set a minimum cluster update interval $t$ . Second, to avoid K-Means meaninglessly handling similar weights with current cached state, i.e. the weights of prompt that triggers the last re-clustering, we plan to only do re-clustering when current weight distribution is far different", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Quantized Prompt for Efficient Generalization of Vision-Language Models", + "bbox": [ + 236, + 113, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "from the cached weight distribution. To reach the goal, Kullback-Leibler divergence is a good metric. However, if we want to compute the Kullback-Leibler divergence between two sets of discrete random variables, they must be defined on the same event space. However, that's not the case. So we first project each of them into the same event space spanning by the K-Means clustering algorithm. Specifically, given a current weight $W^{cur}$ , cached weight $W^{old}$ , quantization function $Q$ , cluster centers $\\mathcal{C} = \\{c_i, 1 \\leq i \\leq 2^b\\}$ , we will compute the $W_{index}^{cur}$ and $W_{index}^{old}$ as in Fig. 4. Then, we can obtain the probability distributions of the indices $p^{cur}$ and $p^{old}$ , respectively, implied by $W_{index}^{cur}$ and $W_{index}^{old}$ , thus successfully transforming two originally unrelated discrete random variables into the same event space. The Kullback-Leibler divergence could be computed as follows:", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nK L \\left(p ^ {c u r} | | p ^ {o l d}\\right) = \\sum_ {i = 1} ^ {2 ^ {b}} p ^ {c u r} (i) \\log \\frac {p ^ {c u r} (i)}{p ^ {o l d} (i)} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 323, + 785, + 367 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The update would continue only if Kullback-Leibler divergence exceeds a certain threshold $T_{KL}$ which shows the distribution difference is significant.", + "bbox": [ + 214, + 372, + 785, + 402 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 425, + 375, + 443 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We evaluate the method and make comparisons with the latest state-of-the-art methods in terms of the following settings across a wide range:", + "bbox": [ + 214, + 455, + 785, + 488 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Base-to-new generalization, which models are trained with base classes and evaluated on both base and new classes.", + "2. Domain generalization, which aims to show generalization to the domain shift, especially for out-of-distribution data.", + "3. Cross-dataset transfer, which aims to see if the method has the potential to transfer beyond a single dataset. It is a much more challenging problem because the fundamentals can be totally changed across different datasets.", + "4. Few-shot learning, which aims to evaluate the adaptation performance of the model to extract knowledge from a dataset whose samples are very few, e.g. 1, 2, 4, 8 or 16 samples." + ], + "bbox": [ + 220, + 496, + 784, + 648 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "All methods are initialized with the same CLIP weights provided by the open-source CLIP [58]. In Appendix, we provide more details of datasets, experimental setting and competitors introduction. Due to page limit, we also put some detailed experimental results into Appendix, e.g. the performance of each method on each dataset in base-to-new generalization setting.", + "bbox": [ + 212, + 656, + 787, + 734 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Main Results", + "text_level": 1, + "bbox": [ + 215, + 753, + 372, + 768 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Base-to-new Generalization The average results over 11 datasets are shown in Tab. 1. For complete results, please refer to the Appendix. Our proposed QCoOp reaches $77.43\\%$ average harmonic mean accuracy within merely 0.26KB size. QCoOp outperforms all kinds of lightweight state-of-the-art methods with", + "bbox": [ + 212, + 779, + 789, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "T. Hao et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/a297cabd252a9de203c54efb6e0f479c6b28b54bddc608685f67584f06a908e6.jpg", + "table_caption": [ + "Table 1: Comparisons with latest methods in base-to-new generalization. H: harmonic mean [63]. QCoOp and QMaPLe are significantly better than other competitors, which prove that our quantization design could be fruitfully integrated into many existing approaches to further improve the performance and efficiency." + ], + "table_footnote": [], + "table_body": "
SizeBaseNewH
CLIP0KB69.3474.2271.70
CoOp4.1KB82.6963.2271.66
CoCoOp70.8KB80.4771.6975.83
Adapter1051KB82.6270.9776.35
LoRA258KB84.3067.3374.86
ProGrad16.4KB82.7968.5575.00
QCoOp0.26KB80.6874.4477.43
MaPLe7096KB82.2875.1478.55
QMaPLe1774KB83.0275.5779.12
", + "bbox": [ + 362, + 199, + 640, + 316 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/d1f832615b95b605ad681dbab8b6836fd06d73a4f1d35aefeab803208e395ab8.jpg", + "table_caption": [ + "Table 2: Comparisons with latest methods in domain generalization. QCoOp gets comparable or even better results with the latest state-of-the-art methods with much fewer parameters, showing excellent robustness for domain shift." + ], + "table_footnote": [], + "table_body": "
SizeSourceImageNetTargetAverage
-V2-Sketch-Adversarial-Rendition
CLIP0KB66.7360.8346.1547.7773.9657.18
CoOp4.1KB71.5164.2047.9949.7175.2159.28
CoCoOp70.8KB71.0264.0748.7550.6376.1859.91
Adapter1051KB69.3362.5347.6749.1775.4258.70
LoRA258KB70.3062.3742.4338.4068.9753.04
ProGrad16.4KB72.2464.7347.6149.3974.5859.07
QCoOp0.26KB70.6763.8748.9351.1076.9060.20
", + "bbox": [ + 217, + 371, + 784, + 477 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "much more efficiency and higher accuracy. For heavier methods like MaPLe, our method can be fruitfully integrated into existing solutions. Besides prompts, we also perform a similar quantization operation on the other weights of MaPLe that are in the linear layers. As a result, QMaPLe not only shows stronger generalization and adaptation capability and gives $0.57\\%$ higher accuracy, but also enjoys a much more smaller model size compared with the original MaPLe.", + "bbox": [ + 212, + 506, + 784, + 595 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Notably, when comparing QCoOp and a lightweight method ProGrad, even if QCoOp is $63 \\times$ smaller than ProGrad, QCoOp still outperforms it by a clear margin, demonstrating the outstanding efficiency and effectiveness.", + "bbox": [ + 212, + 597, + 785, + 642 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Domain Generalization In this paragraph, ImageNet, ImageNet-A, ImageNet-R, ImageNet-v2, and ImageNet-S are used to construct domain generalization experiments. As shown in Tab. 2, on downstream target datasets, QCoOp gets better average accuracy compared with the other methods with significantly better efficiency. For CLIP, CoOp, CLIP-Adapter and ProGrad, there is a clear performance gap between our method and them.", + "bbox": [ + 212, + 664, + 787, + 756 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Cross-dataset Transfer In this paragraph, we do cross-dataset transfer evaluation to further verify our QPrompt. Results are shown in Tab. 3. CoOp is good on source domain but fails on target domains. Probably because it focuses too much on the dataset shown to its eyes and face overfitting and catastrophic", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Quantized Prompt for Efficient Generalization of Vision-Language Models", + "bbox": [ + 236, + 113, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/7c69785b5ec4919a88347084f5a7147916edb5f9514dfa91c5d840d455ef2ea9.jpg", + "table_caption": [ + "Table 3: Results in the cross-dataset transfer setting. QCoOp gives the highest accuracy on 5 of 10 datasets, which well demonstrates that QCoOp could maximally extract general and data-agnostic knowledge from input images." + ], + "table_footnote": [], + "table_body": "
SizeSourceTargetAverage
ImageNetCaltech101PetsCarsFlowersFood101AircraftSun397DTDEuroSATUCF101
CoOp4.1KB71.5193.7089.1464.5168.7185.3018.4764.1541.9246.3966.5563.88
CoCoOp70.8KB71.0294.4390.1465.3271.8886.0622.9467.3645.7345.3768.2165.74
Adapter1051KB69.3393.4388.8764.4070.2785.6324.6765.8044.9047.7066.0065.17
QCoOp0.26KB70.6394.0790.5365.9771.3386.2322.7366.8044.2048.2369.1765.93
", + "bbox": [ + 217, + 186, + 787, + 303 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8b65cd08a49e60b2a3556308655af1b59224075e9b15641d33174225a8b076f6.jpg", + "image_caption": [ + "Fig. 5: Average few-shot learning results on 11 datasets." + ], + "image_footnote": [], + "bbox": [ + 230, + 324, + 468, + 467 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e73a42a7e4c61104bd9f3503316a71777aff14598a86e8124426444a136919b3.jpg", + "image_caption": [ + "Fig. 6: KLD trend of prompt weights under the same experiment with Fig. 3." + ], + "image_footnote": [], + "bbox": [ + 493, + 324, + 736, + 468 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "forgetting problems, which leads to a severe performance drop on unseen objects. QCoOp wins on 5 of 10 datasets and its average accuracy is also slightly better than the best competitor CoCoOp, showing that QCoOp could maximally extract both general and data-agnostic knowledge from given images.", + "bbox": [ + 212, + 523, + 787, + 585 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Few-shot Learning Here we will show the experiment results of QCoOp in the few-shot learning setting that is originated from CoOp. Seen from Fig. 5, QCoOp consistently outperforms zero-shot CLIP, CoOp, and CLIP-Adapter across all the shot numbers. Such results demonstrate the superiority of QCoOp in adaptation ability when there are few samples in downstream tasks.", + "bbox": [ + 212, + 603, + 787, + 678 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Overall, in base-to-new generalization, domain generalization, cross-dataset transfer and few-shot learning, the proposed method can consistently accomplish state-of-the-art performance while enjoying extremely high parameter efficiency, fruitfully demonstrating the effectiveness and efficiency of the proposed method.", + "bbox": [ + 212, + 679, + 787, + 739 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "6 Analysis", + "text_level": 1, + "bbox": [ + 214, + 758, + 334, + 776 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "6.1 Ablation Studies", + "text_level": 1, + "bbox": [ + 214, + 787, + 401, + 801 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Component In this subsection, we decompose the method into pieces and show the influence of each component. As in Tab. 4, we clearly show how", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "T. Hao et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/287a23776cc2cbc1732a434a4799cfcb38a37dc11c5eabd09f5f3240f9f29c7c.jpg", + "table_caption": [ + "Table 4: Ablation study on separate component of QCoOp. Norm: Normalization/Denormalization. CAC: Constrained Adaptive Clustering." + ], + "table_footnote": [], + "table_body": "
K-MeansNormCACbasenewH
QCoOp✓✗✗78.7172.5575.50
✓✓✗78.8473.0975.85
✓✓✓78.2474.0276.07
", + "bbox": [ + 303, + 172, + 689, + 223 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/cd1517ffd452c624abb2a5fb92662394b84a1425be87c4ebc5ffc130adb62a90.jpg", + "table_caption": [ + "Table 5: Comparisons between QAT and PTQ based on CoOp. Both QAT and CoOp are trained with the same hyper-parameters. Here the quantization bit is 1." + ], + "table_footnote": [], + "table_body": "
basenewH
QAT80.7272.3576.31
PTQ82.2168.5074.73
", + "bbox": [ + 377, + 265, + 614, + 304 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "much the K-Means algorithm, normalization/denormalization and constrained adaptive clustering influence the final performance. One interesting point is that K-Means+Norm+CAC only improves the new accuracy compared with K-Means+Norm, showing the superiority of our constrained adaptive clustering.", + "bbox": [ + 214, + 334, + 782, + 393 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Trend of the distributions of prompt between adjacent epochs. To verify the opinions we proposed at the end of Sec. 6, we show the KLD of prompt distributions between adjacent epochs during CoOp's training in Fig. 6. Clearly, this trend is consistent with what we summarized before.", + "bbox": [ + 214, + 402, + 782, + 462 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "QAT v.s. PTQ In Tab. 5, we study the choice of QAT or PTQ following the same strategy, K-Means clustering, in the paper. Results show that QAT consistently outperforms PTQ with the same hyper-parameters. The accuracy on unseen new classes of PTQ is significantly lower than QAT, again proving our opinion that quantization helps generalization by alleviating overfitting as well as catastrophic forgetting. Quantization error is not always undesirable.", + "bbox": [ + 214, + 470, + 782, + 561 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Quantization bit In Tab. 6 and Fig. 7, we show the results across multiple quantization bits. In conclusions, more bits did not always lead to good results, and new accuracy continues decreasing as the training goes on.", + "bbox": [ + 214, + 571, + 782, + 616 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6.2 Performance on Self-Supervised Vision-Language Model", + "text_level": 1, + "bbox": [ + 214, + 641, + 723, + 656 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this paragraph, we explore the usage of QCoOp among other different backbones besides CLIP. We choose a self-supervised vision-language model, SLIP [53] to further verify the universality and robustness of our proposed method. The experimental results are shown in Tab. 7. We could see that the base accuracies of QCoOp and CoOp are similar, but the new accuracies of QCoOp is much", + "bbox": [ + 214, + 667, + 782, + 743 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/c5495eac07bf1f594ad42f6860c1082094d7ad9bf51df64caebd4717a71c30c8.jpg", + "table_caption": [ + "Table 6: Results across different quantization bits." + ], + "table_footnote": [], + "table_body": "
SizebasenewH
QCoOp (b=1)0.26KB79.4972.6575.92
QCoOp (b=2)0.52KB80.3071.9875.91
QCoOp (b=4)1.05KB80.9271.2275.76
", + "bbox": [ + 321, + 784, + 671, + 835 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Quantized Prompt for Efficient Generalization of Vision-Language Models", + "bbox": [ + 236, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0d118ce17f079fe7974dd01be1f854a44b3ee9131443371cc8a04cc25aded9c0.jpg", + "image_caption": [ + "Fig. 7: The same training curves with Fig. 2. Clearly, increasing quantization noise leads to the same phenomenon with increasing gaussian noise, i.e. more generalizability (new accuracy) and less adaptability (base accuracy)." + ], + "image_footnote": [], + "bbox": [ + 243, + 159, + 472, + 290 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/233c5cf195cce4aabca49cfa8561d3cdf849d18b4353cf4e2ce97bced1439ac6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 508, + 159, + 733, + 290 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/26236f5e18e5be315a25b2a86f048093ab76d7223cee06cbf724535c8f3d68a3.jpg", + "table_caption": [ + "Table 7: Results based on SLIP [53] model." + ], + "table_footnote": [], + "table_body": "
SizebasenewH
CoOp4.1KB68.6546.6055.51
QCoOp0.26KB68.3374.0471.07
", + "bbox": [ + 338, + 349, + 656, + 388 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "higher than CoOp. Such observation verifies our assumption that quantization is quite helpful for generalization again.", + "bbox": [ + 212, + 417, + 784, + 448 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 470, + 359, + 486 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "With the development of huge vision-language models, how to effectively and efficiently adapt such huge models to downstream tasks becomes a challenging problem. Much effort has been made to leverage the potential of prompt tuning in adapting vision-language models. However, existing methods suffer from inefficiency. To reach extremely efficient generalization, we propose QPrompt based on the detailed analysis and deep understanding of the characteristics of the prompt weight distribution. Following the several principles concluded by us, we use K-Means clustering algorithm as the base of our quantization method. To adaptively control the quantization error and minimize the number of re-clustering operations, we propose to do a dynamic check every few iterations. If the Kullback-Leibler divergence between the current weights and the original weights used in the last clustering exceeds a predefined threshold and the quantization error also increases, we will let the model re-cluster the weights and update the new centers and weights. Similarly, continue this process until completion. The proposed method could be simply integrated into many of the existing vision-language tuning methods like CoOp and MaPLe and reach good performance. Importantly, our proposed method is significantly effective and efficient. Extensive experiments show that the designed quantization algorithm indeed improves the generalizability of vision-language model, and save a lot more storage space as well.", + "bbox": [ + 212, + 501, + 787, + 803 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by National Natural Science Foundation of China (Nos. 61925107, 62271281, 62021002)", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "T. Hao et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bahng, H., Jahanian, A., Sankaranarayanan, S., Isola, P.: Visual prompting: Modifying pixel space to adapt pre-trained models. arXiv preprint arXiv:2203.17274 (2022) 4", + "2. Banner, R., Nahshan, Y., Soudry, D.: Post training 4-bit quantization of convolutional networks for rapid-deployment. Advances in Neural Information Processing Systems 32 (2019) 4", + "3. Bhalgat, Y., Lee, J., Nagel, M., Blankevoort, T., Kwak, N.: Lsq+: Improving low-bit quantization through learnable offsets and better initialization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 696-697 (2020) 4", + "4. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. In: The Eleventh International Conference on Learning Representations (2023) 4", + "5. Bulat, A., Tzimiropoulos, G.: Lasp: Text-to-text optimization for language-aware soft prompting of vision & language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23232-23241 (2023) 4", + "6. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022) 4", + "7. Ding, J., Xue, N., Xia, G.S., Dai, D.: Decoupling zero-shot semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11583-11592 (2022) 3", + "8. Ding, X., Ding, G., Guo, Y., Han, J.: Centripetal sgd for pruning very deep convolutional networks with complicated structure. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4943-4953 (2019) 4", + "9. Ding, X., Guo, Y., Ding, G., Han, J.: Acnet: Strengthening the kernel skeletons for powerful cnn via asymmetric convolution blocks. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1911-1920 (2019) 1", + "10. Ding, X., Hao, T., Tan, J., Liu, J., Han, J., Guo, Y., Ding, G.: Resrep: Lossless cnn pruning via decoupling remembering and forgetting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4510-4520 (2021) 4", + "11. Ding, X., Zhang, X., Han, J., Ding, G.: Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11963-11975 (2022) 1", + "12. Ding, X., Zhang, X., Ma, N., Han, J., Ding, G., Sun, J.: Repvgg: Making vgg-style convnets great again. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13733-13742 (2021) 1", + "13. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020) 1, 4", + "14. Esser, S.K., McKinstry, J.L., Bablani, D., Appuswamy, R., Modha, D.S.: Learned step size quantization. In: International Conference on Learning Representations (2019) 4", + "15. Feng, C., Zhong, Y., Jie, Z., Chu, X., Ren, H., Wei, X., Xie, W., Ma, L.: Promptdet: Towards open-vocabulary detection using uncurated images. In: European Conference on Computer Vision. pp. 701-717. Springer (2022) 3" + ], + "bbox": [ + 218, + 172, + 785, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Quantized Prompt for Efficient Generalization of Vision-Language Models", + "bbox": [ + 238, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Finkelstein, A., Almog, U., Grobman, M.: Fighting quantization bias with bias. arXiv preprint arXiv:1906.03193 (2019) 4", + "17. Gholami, A., Kim, S., Dong, Z., Yao, Z., Mahoney, M.W., Keutzer, K.: A survey of quantization methods for efficient neural network inference. In: Low-Power Computer Vision, pp. 291-326. Chapman and Hall/CRC (2022) 4", + "18. Han, S., Mao, H., Dally, W.J.: Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149 (2015) 4", + "19. Hao, T., Chen, H., Guo, Y., Ding, G.: Consolidator: Mergeable adapter with grouped connections for visual adaptation. arXiv preprint arXiv:2305.00603 (2023) 3, 4", + "20. Hao, T., Ding, X., Han, J., Guo, Y., Ding, G.: Manipulating identical filter redundancy for efficient pruning on deep and complicated cnn. IEEE Transactions on Neural Networks and Learning Systems (2023) 4", + "21. Hao, T., Lyu, M., Chen, H., Zhao, S., Han, J., Ding, G.: Re-parameterized low-rank prompt: Generalize a vision-language model within $0.5\\mathrm{k}$ parameters. arXiv preprint arXiv:2312.10813 (2023) 3", + "22. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 1", + "23. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019) 1, 3", + "24. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021) 1, 3", + "25. Hubara, I., Nahshan, Y., Hanani, Y., Banner, R., Soudry, D.: Accurate post training quantization with small calibration sets. In: International Conference on Machine Learning. pp. 4466-4475. PMLR (2021) 4", + "26. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International Conference on Machine Learning. pp. 4904-4916. PMLR (2021) 1, 3", + "27. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. arXiv preprint arXiv:2203.12119 (2022) 3, 4, 5", + "28. Jiang, Z., Araki, J., Ding, H., Neubig, G.: How can we know when language models know? on the calibration of language models for question answering. Transactions of the Association for Computational Linguistics 9, 962-977 (2021) 3", + "29. Jie, S., Wang, H., Deng, Z.H.: Revisiting the parameter efficiency of adapters from the perspective of precision redundancy. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17217-17226 (2023) 3", + "30. Ju, C., Han, T., Zheng, K., Zhang, Y., Xie, W.: Prompting visual-language models for efficient video understanding. In: European Conference on Computer Vision. pp. 105-124. Springer (2022) 3", + "31. Jung, S., Son, C., Lee, S., Son, J., Han, J.J., Kwak, Y., Hwang, S.J., Choi, C.: Learning to quantize deep networks by optimizing quantization intervals with task loss. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4350-4359 (2019) 4" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "T. Hao et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Khattak, M.U., Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Maple: Multi-modal prompt learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19113-19122 (2023) 4, 5", + "33. Khattak, M.U., Wasim, S.T., Naseer, M., Khan, S., Yang, M.H., Khan, F.S.: Self-regulating prompts: Foundational model adaptation without forgetting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15190-15200 (2023) 4", + "34. Lee, D., Song, S., Suh, J., Choi, J., Lee, S., Kim, H.J.: Read-only prompt optimization for vision-language few-shot learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1401-1411 (2023) 4", + "35. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 (2021) 3", + "36. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=RriDjddCLN 3", + "37. Li, H., Kadav, A., Durdanovic, I., Samet, H., Graf, H.P.: Pruning filters for efficient convnets. arXiv preprint arXiv:1608.08710 (2016) 4", + "38. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 (2021) 3", + "39. Li, Y., Xu, S., Zhang, B., Cao, X., Gao, P., Guo, G.: Q-vit: Accurate and fully quantized low-bit vision transformer. Advances in Neural Information Processing Systems 35, 34451-34463 (2022) 4", + "40. Li, Z., Xiao, J., Yang, L., Gu, Q.: Repq-vit: Scale reparameterization for posttraining quantization of vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17227-17236 (2023) 4", + "41. Lian, D., Zhou, D., Feng, J., Wang, X.: Scaling & shifting your features: A new baseline for efficient model tuning. In: Advances in Neural Information Processing Systems (NeurIPS) (2022) 3, 4", + "42. Liu, P., Yuan, W., Fu, J., Jiang, Z., Hayashi, H., Neubig, G.: Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. arXiv preprint arXiv:2107.13586 (2021) 3", + "43. Liu, X., Ji, K., Fu, Y., Du, Z., Yang, Z., Tang, J.: P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv preprint arXiv:2110.07602 (2021) 3", + "44. Liu, X., Zheng, Y., Du, Z., Ding, M., Qian, Y., Yang, Z., Tang, J.: Gpt understands, too. arXiv preprint arXiv:2103.10385 (2021) 3", + "45. Liu, Z., Wang, Y., Han, K., Zhang, W., Ma, S., Gao, W.: Post-training quantization for vision transformer. Advances in Neural Information Processing Systems 34, 28092-28103 (2021) 4", + "46. Liu, Z., Li, J., Shen, Z., Huang, G., Yan, S., Zhang, C.: Learning efficient convolutional networks through network slimming. In: Proceedings of the IEEE international conference on computer vision. pp. 2736-2744 (2017) 4", + "47. Lüddecke, T., Ecker, A.: Image segmentation using text and image prompts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7086-7096 (2022) 3", + "48. Lyu, M., Hao, T., Xu, X., Chen, H., Han, J., Ding, G.: Learn from the learnt: Source-free active domain adaptation via contrastive sampling and visual persistence. In: European Conference on Computer Vision (ECCV). Springer (2024) 3", + "49. Lyu, M., Yang, Y., Hong, H., Chen, H., Jin, X., He, Y., Xue, H., Han, J., Ding, G.: One-dimensional adapter to rule them all: Concepts diffusion models and erasing" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Quantized Prompt for Efficient Generalization of Vision-Language Models", + "bbox": [ + 238, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "applications. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7559-7568 (2024) 1", + "50. Lyu, M., Zhou, J., Chen, H., Huang, Y., Yu, D., Li, Y., Guo, Y., Guo, Y., Xiang, L., Ding, G.: Box-level active detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23766-23775 (2023) 1", + "51. Maaz, M., Rasheed, H., Khan, S., Khan, F.S., Anwer, R.M., Yang, M.H.: Class-agnostic object detection with multi-modal transformer. In: The European Conference on Computer Vision. Springer (2022) 3", + "52. Meller, E., Finkelstein, A., Almog, U., Grobman, M.: Same, same but different: Recovering neural network quantization error through weight factorization. In: International Conference on Machine Learning. pp. 4486-4495. PMLR (2019) 4", + "53. Mu, N., Kirillov, A., Wagner, D., Xie, S.: Slip: Self-supervision meets language-image pre-training. In: European Conference on Computer Vision. pp. 529-544. Springer (2022) 1, 13, 14", + "54. Nagel, M., Baalen, M.v., Blankevoort, T., Welling, M.: Data-free quantization through weight equalization and bias correction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1325-1334 (2019) 4", + "55. Nagel, M., Fournarakis, M., Amjad, R.A., Bondarenko, Y., Van Baalen, M., Blankevoort, T.: A white paper on neural network quantization. arXiv preprint arXiv:2106.08295 (2021) 4", + "56. Nagel, M., Fournarakis, M., Bondarenko, Y., Blankevoort, T.: Overcoming oscillations in quantization-aware training. In: International Conference on Machine Learning. pp. 16318-16330. PMLR (2022) 4", + "57. Qian, R., Li, Y., Xu, Z., Yang, M.H., Belongie, S., Cui, Y.: Multimodal open-vocabulary video classification via pre-trained vision and language models. arXiv preprint arXiv:2207.07646 (2022) 3", + "58. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021) 1, 3, 4, 10", + "59. Rao, Y., Zhao, W., Chen, G., Tang, Y., Zhu, Z., Huang, G., Zhou, J., Lu, J.: Denseclip: Language-guided dense prediction with context-aware prompting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18082-18091 (2022) 1, 3", + "60. Rasheed, H.A., Maaz, M., Khattak, M.U., Khan, S., Khan, F.: Bridging the gap between object and image-level representations for open-vocabulary detection. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=aKXBrjODHm 3", + "61. Shin, T., Razeghi, Y., Logan IV, R.L., Wallace, E., Singh, S.: Autoprompt: Eliciting knowledge from language models with automatically generated prompts. In: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). pp. 4222-4235 (2020) 3", + "62. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research 15(1), 1929-1958 (2014) 5", + "63. Xian, Y., Schiele, B., Akata, Z.: Zero-shot learning-the good, the bad and the ugly. In: CVPR (2017) 11", + "64. Xiong, Y., Chen, H., Hao, T., Lin, Z., Han, J., Zhang, Y., Wang, G., Bao, Y., Ding, G.: Pyra: Parallel yielding re-activation for training-inference efficient task adaptation. arXiv preprint arXiv:2403.09192 (2024) 3" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "T. Hao et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "65. Xiong, Y., Chen, H., Lin, Z., Zhao, S., Ding, G.: Confidence-based visual dispersal for few-shot unsupervised domain adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11621-11631 (2023) 1", + "66. Xiong, Y., Chen, X., Ye, X., Chen, H., Lin, Z., Lian, H., Niu, J., Ding, G.: Temporal scaling law for large language models. arXiv preprint arXiv:2404.17785 (2024) 1", + "67. Yao, H., Zhang, R., Xu, C.: Visual-language prompt tuning with knowledge-guided context optimization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6757-6767 (2023) 4", + "68. Yao, L., Huang, R., Hou, L., Lu, G., Niu, M., Xu, H., Liang, X., Li, Z., Jiang, X., Xu, C.: FILIP: Fine-grained interactive language-image pre-training. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=cpDhcsEDC2 1, 3", + "69. Yu, S., Chen, T., Shen, J., Yuan, H., Tan, J., Yang, S., Liu, J., Wang, Z.: Unified visual transformer compression. In: International Conference on Learning Representations (2022) 4", + "70. Yuan, L., Chen, D., Chen, Y.L., Codella, N., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., et al.: Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432 (2021) 3", + "71. Yuan, Z., Xue, C., Chen, Y., Wu, Q., Sun, G.: Ptg4vit: Post-training quantization for vision transformers with twin uniform quantization. In: European Conference on Computer Vision. pp. 191-207. Springer (2022) 4", + "72. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary detr with conditional matching. arXiv preprint arXiv:2203.11876 (2022) 3", + "73. Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., Beyer, L.: Lit: Zero-shot transfer with locked-image text tuning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18123-18133 (2022) 3", + "74. Zhang, Y., Zhou, K., Liu, Z.: Neural prompt search. arXiv preprint arXiv:2206.04673 (2022) 4", + "75. Zheng, K., Wu, W., Feng, R., Zhu, K., Liu, J., Zhao, D., Zha, Z.J., Chen, W., Shen, Y.: Regularized mask tuning: Uncovering hidden knowledge in pre-trained vision-language models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11663-11673 (2023) 4", + "76. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022) 3, 4", + "77. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022) 3, 4, 6, 7" + ], + "bbox": [ + 212, + 146, + 787, + 686 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Quantized Prompt for Efficient Generalization of Vision-Language Models", + "bbox": [ + 238, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Quantized Prompt for Efficient Generalization of Vision-Language Models/d406d327-feb0-4aee-91eb-de3562f63166_model.json b/2024/Quantized Prompt for Efficient Generalization of Vision-Language Models/d406d327-feb0-4aee-91eb-de3562f63166_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a74ffe240010c18ed4d1705449af3d6d5d90ecc8 --- /dev/null +++ b/2024/Quantized Prompt for Efficient Generalization of Vision-Language Models/d406d327-feb0-4aee-91eb-de3562f63166_model.json @@ -0,0 +1,2790 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.223, + 0.141, + 0.783, + 0.187 + ], + "angle": 0, + "content": "Quantized Prompt for Efficient Generalization of Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.244, + 0.211, + 0.761, + 0.245 + ], + "angle": 0, + "content": "Tianxiang Hao\\(^{1,2}\\), Xiaohan Ding\\(^{3(\\boxtimes)}\\), Juexiao Feng\\(^{1,2}\\), Yuhong Yang\\(^{1,2}\\), Hui Chen\\(^{2}\\), and Guiguang Ding\\(^{1,2(\\boxtimes)}\\)" + }, + { + "type": "text", + "bbox": [ + 0.308, + 0.255, + 0.696, + 0.297 + ], + "angle": 0, + "content": "1 School of Software, Tsinghua University, Beijing, China \n2 BNrist, Beijing, China \n3 Bytedance" + }, + { + "type": "text", + "bbox": [ + 0.299, + 0.298, + 0.703, + 0.326 + ], + "angle": 0, + "content": "{beyondtx,xiaohding,suise.con,jichenhui}@gmail.com \nfjx20@mails.tsinghua.edu.cn, dinggg@tsinghua.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.355, + 0.744, + 0.704 + ], + "angle": 0, + "content": "Abstract. In the past few years, large-scale pre-trained vision-language models like CLIP have achieved tremendous success in various fields. Naturally, how to transfer the rich knowledge in such huge pre-trained models to downstream tasks and datasets becomes a hot topic. During downstream adaptation, the most challenging problems are overfitting and catastrophic forgetting, which can cause the model to overly focus on the current data and lose more crucial domain-general knowledge. Existing works use classic regularization techniques to solve the problems. As solutions become increasingly complex, the ever-growing storage and inference costs are also a significant problem that urgently needs to be addressed. While in this paper, we start from an observation that proper random noise can suppress overfitting and catastrophic forgetting. Then we regard quantization error as a kind of noise, and explore quantization for regularizing vision-language model, which is quite efficiency and effective. Furthermore, to improve the model's generalization capability while maintaining its specialization capacity at minimal cost, we deeply analyze the characteristics of the weight distribution in prompts, conclude several principles for quantization module design and follow such principles to create several competitive baselines. The proposed method is significantly efficient due to its inherent lightweight nature, making it possible to adapt on extremely resource-limited devices. Our method can be fruitfully integrated into many existing approaches like MaPLe, enhancing accuracy while reducing storage overhead, making it more powerful yet versatile. Extensive experiments on 11 datasets shows great superiority of our method sufficiently. Code is available at github" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.714, + 0.717, + 0.73 + ], + "angle": 0, + "content": "Keywords: Quantization \\(\\cdot\\) Generalization \\(\\cdot\\) Vision-language model" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.751, + 0.378, + 0.767 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Recently, deep learning models and related technologies have seen rapid development [9, 11-13, 22-24, 49, 50, 65, 66]. Vision-language model (VLM) like CLIP [58] is one of the hottest research topics and leads to huge success. The excellent generalization ability is a crucial cornerstone of such achievement [26, 53, 58, 59, 68]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "T. Hao et al." + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.144, + 0.67, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.371, + 0.788, + 0.442 + ], + "angle": 0, + "content": "Fig. 1: Overall performance comparison with existing vision-language tuning methods. Our method outperforms all of the state-of-the-art competitors with significantly fewer storage space. Based on the proposed quantization algorithm, our method could be integrated into many of the existing methods and bring consistent improvements with excellent efficiency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.469, + 0.788, + 0.531 + ], + "angle": 0, + "content": "When people have access to downstream data, it is better to tune the pretrained VLM on the target dataset for higher accuracy. However, such full finetuning could easily cause the model to overfit the small downstream dataset and face catastrophic forgetting problem, leading to severe performance drop." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.788, + 0.686 + ], + "angle": 0, + "content": "To solve the problem, in this paper we will start from rethinking the relationship between noise and generalization. We propose to think noise as a kind of regularization techniques, which may be helpful for alleviating overfitting and catastrophic forgetting problem. As in Sec. 3.2 and Fig. 2, we then find that directly adding some random Gaussian noise to the tunable prompts of vision-language models would result in a performance gain in several cases, which partly validates our conjecture about using noise as a form of regularization. In particular, excessive noise diminishes the model's adaptation capability, while insufficient noise fails to provide effective regularization. Only noise of moderate intensity is beneficial for the model's generalization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.842 + ], + "angle": 0, + "content": "However, Gaussian noise is absolutely random and hard to control, and thus it is difficult for us to take full advantage of such type of noise to benefit generalization. Instead, we point out that quantization error is also a form of noise, and therefore, it is also possible to leverage quantization error to enhance the model's generalization performance. With this idea in mind, we thoroughly analyzed the distribution pattern of prompt weights in Sec. 3.3 and derived several design principles for quantization algorithms based on the observed phenomena. In Sec. 4, following the principles we summarized, we successfully designed an efficient quantization-aware training algorithm, which largely enhances the model's generalization ability while quantizing." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.238, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Quantized Prompt for Efficient Generalization of Vision-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.221 + ], + "angle": 0, + "content": "Since our design is general, we could fruitfully integrate our quantization strategy into many existing methods and reach a higher accuracy with much smaller model size due to the lightweight nature of quantization. For example, we integrate our method onto an existing popular method MaPLe, and our QMaPLe earns \\(0.57\\%\\) accuracy improvement with only \\(0.25 \\times\\) size." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.223, + 0.648, + 0.236 + ], + "angle": 0, + "content": "In conclusion, we summarize our contribution as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.249, + 0.784, + 0.293 + ], + "angle": 0, + "content": "- We deeply analyze the effect of noise and rethink the relationship between noise and generalization in Sec. 3.2 for vision-language models. As a result, we confirm that moderate noise would promote the model generalization" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.295, + 0.784, + 0.338 + ], + "angle": 0, + "content": "- We are the first to propose to quantize prompts. By detailed observation and hard thinking, we conclude several principles about how to effectively design a quantization method for the prompts of vision-language models." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.34, + 0.784, + 0.442 + ], + "angle": 0, + "content": "- Following the principles we concluded, we build our method and successfully quantize the prompts as well as some other weights, and our method significantly outperforms existing ones as in Fig. 1. Extensive experiments show great power of our method. In base-to-new generalization, domain generalization, cross-dataset transfer and few-shot learning settings, we consistently reach competitive results, winning many state-of-the-art tuning methods with a much smaller size of model." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.249, + 0.784, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.47, + 0.396, + 0.485 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.505, + 0.464, + 0.519 + ], + "angle": 0, + "content": "2.1 Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.532, + 0.784, + 0.698 + ], + "angle": 0, + "content": "In recent times, large-scale vision-language models have demonstrated remarkable performance across various tasks. Seminal works such as [26, 58, 68, 70, 73]. Classic works have focused on learning multimodal representations through self-supervised methods using extensive sets of image-text pairs. Among these, CLIP [58] stands out as a milestone achievement, employing contrastive learning to align vision and language representations and achieving exceptional performance. A well-trained vision-language model is invaluable, offering substantial support to various fields. Successful applications of these robust models include few-shot recognition [76, 77], detection tasks [15, 51, 60, 72], and segmentation tasks [7, 36, 47, 59]. Furthermore, for video data, research efforts have emerged in video classification [57] and video understanding [30]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.722, + 0.529, + 0.737 + ], + "angle": 0, + "content": "2.2 Parameter-Efficient Fine-Tuning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Recently, a series of works [19, 21, 23, 24, 27, 29, 38, 41, 44, 48, 64, 76, 77] have been proposed to help transfer the learnt knowledge, where one of the most popular field is parameter-efficient fine-tuning (PEFT). PEFT aims to transfer a pre-trained model to downstream tasks by a minimum number of parameters. Originating from natural language processing tasks, classic methods like adapter [23], prompt tuning [28, 35, 38, 42, 43, 61] and LoRA [24] follows a similar" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "T. Hao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.33 + ], + "angle": 0, + "content": "principle to add extra modules with a small number of parameters into the backbone model, freeze the original parameters and only tune and store the newly added parameters. Inspired by the success of PEFT in language field, researchers have extended such kind of approaches to adapt visual models in a similar fashion [6,19,27,41,74]. In the field of vision-language modeling, several explorations have been made as well. Bahng et al. [1] exclusively apply prompt tuning to the image encoder. CoOp [77] replaces the fixed template in CLIP [58] with tunable text prompts. CoCoOp [76] leverages image features to guide the optimization of tunable text prompts in CoOp. Other works [32, 34] optimize both image and text prompts simultaneously and establish additional connections between different modalities. To mitigate overfitting and catastrophic forgetting, various works [5,33,67,75] integrate regularization modules or losses into prompt tuning." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.352, + 0.373, + 0.368 + ], + "angle": 0, + "content": "2.3 Quantization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.379, + 0.789, + 0.621 + ], + "angle": 0, + "content": "Quantization is one of the most effective compression methods [4,8,10,20,37, 46,69] for deep learning models. Generally, parameters such as weights and activations are typically stored as 32-bit floating-point numbers, which consume a significant amount of memory and require intensive computation during inference. Quantization [17, 18, 55] involves representing these parameters with reduced precision, such as 8-bit integers or even lower bit-widths. By doing so, quantization can significantly reduce the memory footprint and computational complexity of the model without significantly sacrificing accuracy. Quantization methods can be divided into two groups, Post-Training Quantization (PTQ) [2,16,25,40,45,52,54,71] that consumes few resources but suffers higher accuracy loss, and Quantization-Aware Training (QAT) [3,14,31,39,56] that relies on plenty of resources for training and shows better accuracy. Existing works aim to minimize quantization error to improve accuracy, while our work demonstrates that both excessive and insufficient errors are detrimental to model generalization. To achieve optimal generalization performance, a moderate error is required." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.645, + 0.723, + 0.663 + ], + "angle": 0, + "content": "3 Exploring Quantization in Model Generalization" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.678, + 0.741, + 0.695 + ], + "angle": 0, + "content": "3.1 Preliminaries: Prompt Tuning of Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.809 + ], + "angle": 0, + "content": "CLIP comprises a text encoder \\(\\mathcal{L}\\) and an image encoder \\(\\mathcal{V}\\). Typically, \\(\\mathcal{L}\\) is implemented as a language transformer, whereas \\(\\mathcal{V}\\) may be realized using either a convolutional neural network or a vision transformer. In this study, following the methodologies outlined by [76, 77], we employ a ViT-B/16 model [13] as the image encoder \\(\\mathcal{V}\\), except where otherwise specified. The subsequent sections will provide a brief overview of the methods used to prompt CLIP for prediction tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Text Encoder Consider a text encoder composed of \\(M\\) layers. For the \\(k\\)-th layer, denoted as \\(\\mathcal{L}_k\\), the inputs consist of a sequence of prompt tokens \\(P_{k - 1}^{l}\\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.238, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Quantized Prompt for Efficient Generalization of Vision-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.304 + ], + "angle": 0, + "content": "and a \\([CLS]\\) token \\(c_{k-1}^{l}\\), while the outputs are represented by \\(P_{k}^{l}\\) and \\(c_{k}^{l}\\). The initial inputs, \\(P_{0}^{l}\\) and \\(c_{0}^{l}\\), correspond to the word embeddings of the prompts combined with the label, such as \"A photo of a \\([CLS]''\\) or alternatively, some randomly initialized vectors. Formally, we denote \\(P_{k}^{l} \\in \\mathbb{R}^{n^{l} \\times d^{l}}\\) and \\(c_{k}^{l} \\in \\mathbb{R}^{d^{l}}\\), where \\(n^{l}\\) signifies the length of the text prompts and \\(d^{l}\\) represents the dimension of the word embeddings. For each layer, \\(1 \\leq k \\leq M\\), the relationship is given by \\([P_{k}^{l}, c_{k}^{l}] = \\mathcal{L}_{k}([P_{k-1}^{l}, c_{k-1}^{l}])\\). The output feature of the text encoder \\(f^{l} \\in \\mathbb{R}^{d^{v}}\\), where \\(d^{v}\\) is the dimension of the visual feature space, is calculated by projecting the \\([CLS]\\) token of its last layer to the visual latent space through a linear transformation, i.e. \\(f^{l} = \\mathrm{Proj}(c_{M}^{l})\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.304, + 0.788, + 0.424 + ], + "angle": 0, + "content": "Image Encoder Suppose there are \\(N\\) layers in the image encoder. For \\(k\\)-th layer \\(\\mathcal{V}_k\\), the inputs are a series of image tokens \\(I_{k-1}\\), a classification token \\(c_{k-1}^v\\) and prompt tokens \\(P_{k-1}^v\\), and the outputs are \\(I_k\\), \\(c_k^v\\) and \\(P_{k}^{v}\\). The inputs of the first layer \\(I_0\\) and \\(c_0^v\\) are the patch embeddings of the input image and the pre-trained class token. \\(P_0^v\\) is randomly initialized. Formally, \\(I_k \\in \\mathbb{R}^{p \\times d^v}\\), \\(c_k^v \\in \\mathbb{R}^{d^v}\\) and \\(P_k^v \\in \\mathbb{R}^{n^v \\times d^v}\\), where \\(p\\) denotes the number of image patches and \\(d^v\\) denotes the dimension of visual embedding. \\(\\forall 1 \\leq k \\leq N\\), \\([P_k^v, c_k^v, I_k] = \\mathcal{V}_k([P_{k-1}^v, c_{k-1}^v, I_{k-1}])\\). The output feature of the image encoder is \\(f^v = c_N^v\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.424, + 0.788, + 0.502 + ], + "angle": 0, + "content": "Prediction For image classification, suppose there are \\(C\\) classes, and \\(\\{f_c^l\\}_{c = 1}^C\\) are the text features. Label \\(y\\)'s probability is \\(p(y|f^v) = \\frac{\\exp(\\sin(f^v,f_l^i) / \\tau)}{\\sum_{c = 1}^{C}\\exp(\\sin(f^v,f_c^l) / \\tau)}\\) where \\(\\sin (\\cdot ,\\cdot)\\) denotes cosine similarity function and \\(\\tau\\) is temperature. The final prediction is \\(\\hat{z} = \\arg \\max_{1\\leq y < C}(p(y|f^v))\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.794, + 0.609 + ], + "angle": 0, + "content": "We have introduced shallow prompts. There are also different types of prompts. Several works [27, 32] use them for improve performance. They directly add and tune the prompt in each layer in the feature encoder, instead of inheriting the output prompt calculated by the last encoder. Now we have \\([\\_ ,c_k^l ] = \\mathcal{L}_k([P_{k - 1}^l,c_{k - 1}^l])\\) and \\([\\_ ,c_k^v,I_k] = \\mathcal{V}_k([P_{k - 1}^v,c_{k - 1}^v,I_{k - 1}])\\). Note that \\(P^l\\) and \\(P^v\\) are independent tunable parameters. They are no longer determined by the previous layer." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.633, + 0.782, + 0.649 + ], + "angle": 0, + "content": "3.2 Rethinking the Relationship between Noise and Generalization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Some existing work has explored how to use the addition of noise to suppress model overfitting, e.g., techniques like Dropout [62] that randomly drops connections, and random jitter that directly introduce perturbations in the input data. However, the impact of directly adding noise to model weights has been less explored, especially in the context of prompt structures, which have become popular only in recent years, in Transformer architectures. The variations of the specialization capability represented by the test accuracy on base, seen classes and the generalization capability represented by the test accuracy on new, unseen classes. The curves of different colors represent the data under the influence of random Gaussian noise of different intensities, e.g. \"Noise_0.01\" adds random noise with a distribution of \\(\\mathcal{N}(0,0.01^2)\\) to the prompt. \"Noise_0\" denotes the baseline prompt tuning. As training progresses, the generalization capability" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "T. Hao et al." + }, + { + "type": "image", + "bbox": [ + 0.25, + 0.16, + 0.473, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.16, + 0.734, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.293, + 0.788, + 0.486 + ], + "angle": 0, + "content": "Fig. 2: The variations of the specialization capability represented by the test accuracy on base, seen classes and the generalization capability represented by the test accuracy on new, unseen classes on average of the ten datasets in base-to-new generalization setting in Sec. 5 except for time-consuming ImageNet. The curves of different colors represent the data under the influence of random Gaussian noise of different intensities, e.g. \"Noise_0.01\" adds random noise with a distribution of \\(\\mathcal{N}(0,0.01^2)\\) to the prompt. \"Noise_0\" denotes the baseline prompt tuning. As training progresses, the generalization capability of baseline prompt tuning continuously decreases while the specialization capability improves. Therefore, we expect that adding noise can achieve a better balance between generalization and specialization. However, excessive noise, e.g. 0.1, greatly diminishes the model's specialization capability, while insufficient noise, e.g. 0.001, fails to provide effective regularization. Only noise of moderate intensity outperforms baseline in specialization-generalization trade-off, effectively enhancing the unseen class accuracy without significantly compromising seen class accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.788, + 0.611 + ], + "angle": 0, + "content": "of baseline continuously decreases while the specialization capability improves. Therefore, we expect that adding noise can achieve a better balance between generalization and specialization. However, excessive noise, e.g. 0.1, greatly diminishes the model's specialization capability, while insufficient noise, e.g. 0.001, fails to provide effective regularization. Only noise of moderate intensity outperforms baseline in specialization-generalization trade-off, effectively enhancing the unseen class accuracy without significantly compromising seen class accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Quantization, the technique of mapping parameter values from high precision to low precision, can also be viewed as introducing some form of noise into the parameters, and thus can possibly improve the generalizability as the noise did in Fig. 2. Compared to Gaussian noise, quantization error is more controllable. Better still, quantization also has another major advantage: it can significantly reduce the storage required for parameters. Therefore, using quantization algorithms to generalize vision-language models is a very promising direction." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.74, + 0.715, + 0.756 + ], + "angle": 0, + "content": "3.3 Characteristics of Prompts in Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.794 + ], + "angle": 0, + "content": "In this subsection, we will demonstrate the characteristics of prompts in the vision-language model to facilitate targeted design of quantization schemes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.841 + ], + "angle": 0, + "content": "We start from analyzing the training procedure of CoOp [77], which freezes all the parameters in backbone and tunes the added prompts only. The histogram about the frequency of weights of prompts for different training epochs is shown" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.239, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Quantized Prompt for Efficient Generalization of Vision-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.147, + 0.323, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.147, + 0.434, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.147, + 0.549, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.147, + 0.661, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.147, + 0.776, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.212, + 0.322, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.212, + 0.434, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.212, + 0.548, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.212, + 0.661, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.212, + 0.776, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.273, + 0.787, + 0.384 + ], + "angle": 0, + "content": "Fig. 3: The histogram about the frequency of weights of prompts during training of CoOp [77] on eurosat dataset. We find the shape of the prompt's weight distribution remains largely unchanged throughout the entire training, but the variance of the weight distribution increases rapidly at the beginning of training. Additionally, we also notice that there are almost no outliers in the prompt's weights throughout the entire training process, and apart from the unstable initial training phase, the changes in weights between adjacent phases are not significant, indicating a very gentle overall updating trend." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.785, + 0.509 + ], + "angle": 0, + "content": "in Fig. 3. We observe that the shape of the prompt's weight distribution remains mostly consistent throughout the entire training process. However, the variance of the weight distribution increases rapidly at the initial stages of training. Refer to Fig. 6 for more details. Moreover, we notice minimal presence of outliers in the prompt's weights throughout the training process. Apart from the unstable initial training phase, the changes in weights between consecutive phases are not substantial, indicating a very gentle overall updating trend." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.787, + 0.54 + ], + "angle": 0, + "content": "Taking into account the observations from Secs. 3.2 and 3.3, we can draw the following conclusions:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.553, + 0.788, + 0.644 + ], + "angle": 0, + "content": "- Because the prompt weights are not sensitive to noise, and moderate noise in the current scenario not only does not degrade the model's performance but actually enhances its generalization ability, we can possibly adopt some quantization strategies that are considered very aggressive and highly likely to significantly degrade performance, such as 1-bit quantization, without causing destructive effects on the model's generalization ability." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.644, + 0.789, + 0.841 + ], + "angle": 0, + "content": "- Given that the parameters targeted for quantization in the current scenario constitute only a tiny fraction of the total model parameters, we speculate that QAT may have more advantages over PTQ, as any additional training costs incurred by QAT which is considered a major drawback in traditional situations would naturally be kept at a low level due to the low proportion of parameters targeted for quantization. Due to the observation in Fig. 2 that both excessive and insufficient noise are detrimental to generalization, in designing the QAT algorithm, we no longer need to always minimize quantization error as the objective, as in classic algorithms. Instead, we aim to maintain the quantization error at a moderate level. Furthermore, as seen in Fig. 3, the changes in prompt weight distribution during most of the training time are gentle, indicating that we can appropriately continue to use past-time algorithm state without the need for real-time updates." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.553, + 0.789, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "T. Hao et al." + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.145, + 0.72, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.333, + 0.785, + 0.361 + ], + "angle": 0, + "content": "Fig. 4: Overview of Quantized Prompt. We set \\( b = 2 \\) for a clear explanation. Normalization and denormalization processes are not shown here." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.384, + 0.784, + 0.475 + ], + "angle": 0, + "content": "- Throughout the training process, the shape of the prompt weight distribution remains roughly unchanged, with the main variation coming from the distribution's variance. Therefore, it may be beneficial to attempt to eliminate the translation and scaling transformations of the distribution by normalizing the distribution before quantization and denormalizing it after quantization, which might help improve quantization accuracy." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.476, + 0.785, + 0.597 + ], + "angle": 0, + "content": "- Since there are almost no outliers throughout the entire training process, some classic algorithms are revitalized. Specifically, clustering algorithms like K-Means are good choices, because one of the major drawbacks of them in traditional usage is that they are greatly affected by outliers. In addition, as we have analyzed before, there is not a high demand for dynamically minimizing quantization error in vision-language generalization, opting for a more efficient non-parametric clustering approach seems a better choice compared with some complex parametric approaches." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.384, + 0.785, + 0.597 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.629, + 0.46, + 0.647 + ], + "angle": 0, + "content": "4 Prompt Quantization" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.669, + 0.538, + 0.685 + ], + "angle": 0, + "content": "4.1 Preliminaries: Quantization Basis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.703, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Suppose there are \\(m\\) parameters in total to be quantized, which are denoted by \\(W \\in \\mathbb{R}^m\\). Each \\(w_i\\) here is a high-precision float-type number. A \\(b\\)-bit model quantization algorithm divides the value range of parameters into \\(2^b\\) intervals \\(\\{\\mathcal{U}_i, 1 \\leq i \\leq 2^b\\}\\), aiming to find a mapping \\(Q\\) from intervals to points, mapping all values within an interval \\(\\mathcal{U}_i\\) to the same quantized value \\(q_i\\), i.e. \\(Q(x) = q_i, \\forall x \\in \\mathcal{U}_i\\). We define the quantization error \\(E = \\sum_{i=1}^{N} ||Q(W_i) - W_i||^2\\). Such error serves as the objective of K-Means algorithm as well, so it can be ensured that each time clustering is redone using the K-Means algorithm, the quantization error will most likely decrease." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.238, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Quantized Prompt for Efficient Generalization of Vision-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.52, + 0.162 + ], + "angle": 0, + "content": "4.2 Overview of Quantized Prompt" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.788, + 0.367 + ], + "angle": 0, + "content": "Based on the observation and analysis in Sec. 3.3, we build a simple yet quite proper quantization algorithm for generalizing vision-language models. As stated before, we choose a widely-used clustering method K-Means, to construct the mapping \\( Q \\). Initially, K-Means algorithm is run to fit the pre-trained prompt weights, and record the codebook, i.e. cluster centers. Normalization and denormalization are conducted beyond quantization to alleviate the influence of varying variance. Formally, given the original prompt \\( W \\) with \\( N \\) parameters, we first calculate \\( \\mu = \\frac{1}{N}\\sum_{i=1}^{N}W_i \\) and \\( \\sigma = \\sqrt{\\frac{\\sum_{1\\leq i\\leq N}(W_i - \\mu)^2}{N}} \\), then normalize \\( W \\) and get \\( \\hat{W} = \\frac{W - \\mu}{\\sigma} \\). K-Means quantization is applied on \\( \\hat{W} \\), and the quantized prompt \\( W_q = \\sigma Q(\\hat{W}) + \\mu \\). Since \\( Q \\) is not differentiable, we adopt a common practice to directly propagate the gradient across the quantization function by Straight-Through Estimator (STE), i.e. \\( \\frac{\\partial Q(x)}{\\partial x} = x \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.368, + 0.788, + 0.503 + ], + "angle": 0, + "content": "The overall framework is shown in Fig. 4. In training, we keep tuning the weight by the fake gradient propagated from quantization operation. For storage, we first convert the fp16 parameters in prompt to \\( b \\)-bit indexes, which could be used to search for the corresponding cluster center in the codebook. In Fig. 4, we set \\( b = 2 \\) for a clear explanation. Compared with baseline method, our quantized method could save a lot more storage space. Specifically, for storage, an ordinary method needs \\( 16N \\) bits, while ours only needs \\( bN + 2^b \\times 16 \\). In experiments, we usually set \\( b \\) to 1, 2 or 4, which is far more smaller than \\( N \\). For example, with \\( b = 1 \\), the storage space of our method is roughly \\( 16 \\times \\) smaller than baseline." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.787, + 0.534 + ], + "angle": 0, + "content": "Note that the codebook here is updated by a rule called \"constrained adaptive clustering\". We will give a detailed description of it in the following subsection." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.558, + 0.676, + 0.573 + ], + "angle": 0, + "content": "4.3 Constrained Adaptive Clustering for Quantization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.749 + ], + "angle": 0, + "content": "In Fig. 4, re-clustering the prompt every iteration like classic QAT algorithms did to keep the codebook updated at all times is not a good choice. There are three reasons: 1. K-Means algorithm is not quite efficient, and thus if we run it every iteration then the training efficiency would decrease; 2. Only moderate noise promotes generalization, so keep updating the clusters to minimize the quantization error may be not helpful but harmful; 3. From Fig. 3, we can observe that the weight changes are very gentle for most of the training time, and the weight distributions of adjacent stages are highly similar. Therefore, even if we update the internal clustering state of K-Means with these already very similar data at each iteration, it's difficult to generate a better clustering solution, which is just futile effort." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We propose constrained adaptive clustering to instruct the update of parameters inside K-Means which would not updated by gradient at all. Intuitively, first, we do not want too often update, so we set a minimum cluster update interval \\( t \\). Second, to avoid K-Means meaninglessly handling similar weights with current cached state, i.e. the weights of prompt that triggers the last re-clustering, we plan to only do re-clustering when current weight distribution is far different" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "T. Hao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "from the cached weight distribution. To reach the goal, Kullback-Leibler divergence is a good metric. However, if we want to compute the Kullback-Leibler divergence between two sets of discrete random variables, they must be defined on the same event space. However, that's not the case. So we first project each of them into the same event space spanning by the K-Means clustering algorithm. Specifically, given a current weight \\( W^{cur} \\), cached weight \\( W^{old} \\), quantization function \\( Q \\), cluster centers \\( \\mathcal{C} = \\{c_i, 1 \\leq i \\leq 2^b\\} \\), we will compute the \\( W_{index}^{cur} \\) and \\( W_{index}^{old} \\) as in Fig. 4. Then, we can obtain the probability distributions of the indices \\( p^{cur} \\) and \\( p^{old} \\), respectively, implied by \\( W_{index}^{cur} \\) and \\( W_{index}^{old} \\), thus successfully transforming two originally unrelated discrete random variables into the same event space. The Kullback-Leibler divergence could be computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.36, + 0.324, + 0.786, + 0.368 + ], + "angle": 0, + "content": "\\[\nK L \\left(p ^ {c u r} | | p ^ {o l d}\\right) = \\sum_ {i = 1} ^ {2 ^ {b}} p ^ {c u r} (i) \\log \\frac {p ^ {c u r} (i)}{p ^ {o l d} (i)} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.373, + 0.786, + 0.403 + ], + "angle": 0, + "content": "The update would continue only if Kullback-Leibler divergence exceeds a certain threshold \\( T_{KL} \\) which shows the distribution difference is significant." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.426, + 0.377, + 0.444 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.457, + 0.786, + 0.489 + ], + "angle": 0, + "content": "We evaluate the method and make comparisons with the latest state-of-the-art methods in terms of the following settings across a wide range:" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.497, + 0.785, + 0.526 + ], + "angle": 0, + "content": "1. Base-to-new generalization, which models are trained with base classes and evaluated on both base and new classes." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.528, + 0.785, + 0.556 + ], + "angle": 0, + "content": "2. Domain generalization, which aims to show generalization to the domain shift, especially for out-of-distribution data." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.558, + 0.785, + 0.603 + ], + "angle": 0, + "content": "3. Cross-dataset transfer, which aims to see if the method has the potential to transfer beyond a single dataset. It is a much more challenging problem because the fundamentals can be totally changed across different datasets." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.604, + 0.785, + 0.649 + ], + "angle": 0, + "content": "4. Few-shot learning, which aims to evaluate the adaptation performance of the model to extract knowledge from a dataset whose samples are very few, e.g. 1, 2, 4, 8 or 16 samples." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.497, + 0.785, + 0.649 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.657, + 0.788, + 0.735 + ], + "angle": 0, + "content": "All methods are initialized with the same CLIP weights provided by the open-source CLIP [58]. In Appendix, we provide more details of datasets, experimental setting and competitors introduction. Due to page limit, we also put some detailed experimental results into Appendix, e.g. the performance of each method on each dataset in base-to-new generalization setting." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.755, + 0.373, + 0.769 + ], + "angle": 0, + "content": "5.1 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.79, + 0.842 + ], + "angle": 0, + "content": "Base-to-new Generalization The average results over 11 datasets are shown in Tab. 1. For complete results, please refer to the Appendix. Our proposed QCoOp reaches \\(77.43\\%\\) average harmonic mean accuracy within merely 0.26KB size. QCoOp outperforms all kinds of lightweight state-of-the-art methods with" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.238, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Quantized Prompt for Efficient Generalization of Vision-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.2 + ], + "angle": 0, + "content": "Table 1: Comparisons with latest methods in base-to-new generalization. H: harmonic mean [63]. QCoOp and QMaPLe are significantly better than other competitors, which prove that our quantization design could be fruitfully integrated into many existing approaches to further improve the performance and efficiency." + }, + { + "type": "table", + "bbox": [ + 0.364, + 0.2, + 0.642, + 0.317 + ], + "angle": 0, + "content": "
SizeBaseNewH
CLIP0KB69.3474.2271.70
CoOp4.1KB82.6963.2271.66
CoCoOp70.8KB80.4771.6975.83
Adapter1051KB82.6270.9776.35
LoRA258KB84.3067.3374.86
ProGrad16.4KB82.7968.5575.00
QCoOp0.26KB80.6874.4477.43
MaPLe7096KB82.2875.1478.55
QMaPLe1774KB83.0275.5779.12
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.33, + 0.785, + 0.371 + ], + "angle": 0, + "content": "Table 2: Comparisons with latest methods in domain generalization. QCoOp gets comparable or even better results with the latest state-of-the-art methods with much fewer parameters, showing excellent robustness for domain shift." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.372, + 0.785, + 0.478 + ], + "angle": 0, + "content": "
SizeSourceImageNetTargetAverage
-V2-Sketch-Adversarial-Rendition
CLIP0KB66.7360.8346.1547.7773.9657.18
CoOp4.1KB71.5164.2047.9949.7175.2159.28
CoCoOp70.8KB71.0264.0748.7550.6376.1859.91
Adapter1051KB69.3362.5347.6749.1775.4258.70
LoRA258KB70.3062.3742.4338.4068.9753.04
ProGrad16.4KB72.2464.7347.6149.3974.5859.07
QCoOp0.26KB70.6763.8748.9351.1076.9060.20
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.785, + 0.597 + ], + "angle": 0, + "content": "much more efficiency and higher accuracy. For heavier methods like MaPLe, our method can be fruitfully integrated into existing solutions. Besides prompts, we also perform a similar quantization operation on the other weights of MaPLe that are in the linear layers. As a result, QMaPLe not only shows stronger generalization and adaptation capability and gives \\(0.57\\%\\) higher accuracy, but also enjoys a much more smaller model size compared with the original MaPLe." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.787, + 0.643 + ], + "angle": 0, + "content": "Notably, when comparing QCoOp and a lightweight method ProGrad, even if QCoOp is \\(63 \\times\\) smaller than ProGrad, QCoOp still outperforms it by a clear margin, demonstrating the outstanding efficiency and effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.665, + 0.788, + 0.757 + ], + "angle": 0, + "content": "Domain Generalization In this paragraph, ImageNet, ImageNet-A, ImageNet-R, ImageNet-v2, and ImageNet-S are used to construct domain generalization experiments. As shown in Tab. 2, on downstream target datasets, QCoOp gets better average accuracy compared with the other methods with significantly better efficiency. For CLIP, CoOp, CLIP-Adapter and ProGrad, there is a clear performance gap between our method and them." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Cross-dataset Transfer In this paragraph, we do cross-dataset transfer evaluation to further verify our QPrompt. Results are shown in Tab. 3. CoOp is good on source domain but fails on target domains. Probably because it focuses too much on the dataset shown to its eyes and face overfitting and catastrophic" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "T. Hao et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 3: Results in the cross-dataset transfer setting. QCoOp gives the highest accuracy on 5 of 10 datasets, which well demonstrates that QCoOp could maximally extract general and data-agnostic knowledge from input images." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.187, + 0.789, + 0.304 + ], + "angle": 0, + "content": "
SizeSourceTargetAverage
ImageNetCaltech101PetsCarsFlowersFood101AircraftSun397DTDEuroSATUCF101
CoOp4.1KB71.5193.7089.1464.5168.7185.3018.4764.1541.9246.3966.5563.88
CoCoOp70.8KB71.0294.4390.1465.3271.8886.0622.9467.3645.7345.3768.2165.74
Adapter1051KB69.3393.4388.8764.4070.2785.6324.6765.8044.9047.7066.0065.17
QCoOp0.26KB70.6394.0790.5365.9771.3386.2322.7366.8044.2048.2369.1765.93
" + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.325, + 0.47, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.47, + 0.486, + 0.496 + ], + "angle": 0, + "content": "Fig. 5: Average few-shot learning results on 11 datasets." + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.325, + 0.738, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.487, + 0.47, + 0.774, + 0.498 + ], + "angle": 0, + "content": "Fig. 6: KLD trend of prompt weights under the same experiment with Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.587 + ], + "angle": 0, + "content": "forgetting problems, which leads to a severe performance drop on unseen objects. QCoOp wins on 5 of 10 datasets and its average accuracy is also slightly better than the best competitor CoCoOp, showing that QCoOp could maximally extract both general and data-agnostic knowledge from given images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.604, + 0.788, + 0.679 + ], + "angle": 0, + "content": "Few-shot Learning Here we will show the experiment results of QCoOp in the few-shot learning setting that is originated from CoOp. Seen from Fig. 5, QCoOp consistently outperforms zero-shot CLIP, CoOp, and CLIP-Adapter across all the shot numbers. Such results demonstrate the superiority of QCoOp in adaptation ability when there are few samples in downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.68, + 0.788, + 0.741 + ], + "angle": 0, + "content": "Overall, in base-to-new generalization, domain generalization, cross-dataset transfer and few-shot learning, the proposed method can consistently accomplish state-of-the-art performance while enjoying extremely high parameter efficiency, fruitfully demonstrating the effectiveness and efficiency of the proposed method." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.76, + 0.336, + 0.777 + ], + "angle": 0, + "content": "6 Analysis" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.788, + 0.402, + 0.803 + ], + "angle": 0, + "content": "6.1 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Component In this subsection, we decompose the method into pieces and show the influence of each component. As in Tab. 4, we clearly show how" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.238, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Quantized Prompt for Efficient Generalization of Vision-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.784, + 0.172 + ], + "angle": 0, + "content": "Table 4: Ablation study on separate component of QCoOp. Norm: Normalization/Denormalization. CAC: Constrained Adaptive Clustering." + }, + { + "type": "table", + "bbox": [ + 0.304, + 0.173, + 0.691, + 0.224 + ], + "angle": 0, + "content": "
K-MeansNormCACbasenewH
QCoOp✓✗✗78.7172.5575.50
✓✓✗78.8473.0975.85
✓✓✓78.2474.0276.07
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.238, + 0.784, + 0.265 + ], + "angle": 0, + "content": "Table 5: Comparisons between QAT and PTQ based on CoOp. Both QAT and CoOp are trained with the same hyper-parameters. Here the quantization bit is 1." + }, + { + "type": "table", + "bbox": [ + 0.379, + 0.266, + 0.615, + 0.305 + ], + "angle": 0, + "content": "
basenewH
QAT80.7272.3576.31
PTQ82.2168.5074.73
" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.335, + 0.784, + 0.395 + ], + "angle": 0, + "content": "much the K-Means algorithm, normalization/denormalization and constrained adaptive clustering influence the final performance. One interesting point is that K-Means+Norm+CAC only improves the new accuracy compared with K-Means+Norm, showing the superiority of our constrained adaptive clustering." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.404, + 0.784, + 0.463 + ], + "angle": 0, + "content": "Trend of the distributions of prompt between adjacent epochs. To verify the opinions we proposed at the end of Sec. 6, we show the KLD of prompt distributions between adjacent epochs during CoOp's training in Fig. 6. Clearly, this trend is consistent with what we summarized before." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.472, + 0.784, + 0.563 + ], + "angle": 0, + "content": "QAT v.s. PTQ In Tab. 5, we study the choice of QAT or PTQ following the same strategy, K-Means clustering, in the paper. Results show that QAT consistently outperforms PTQ with the same hyper-parameters. The accuracy on unseen new classes of PTQ is significantly lower than QAT, again proving our opinion that quantization helps generalization by alleviating overfitting as well as catastrophic forgetting. Quantization error is not always undesirable." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.572, + 0.784, + 0.617 + ], + "angle": 0, + "content": "Quantization bit In Tab. 6 and Fig. 7, we show the results across multiple quantization bits. In conclusions, more bits did not always lead to good results, and new accuracy continues decreasing as the training goes on." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.642, + 0.724, + 0.657 + ], + "angle": 0, + "content": "6.2 Performance on Self-Supervised Vision-Language Model" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.669, + 0.784, + 0.744 + ], + "angle": 0, + "content": "In this paragraph, we explore the usage of QCoOp among other different backbones besides CLIP. We choose a self-supervised vision-language model, SLIP [53] to further verify the universality and robustness of our proposed method. The experimental results are shown in Tab. 7. We could see that the base accuracies of QCoOp and CoOp are similar, but the new accuracies of QCoOp is much" + }, + { + "type": "table_caption", + "bbox": [ + 0.33, + 0.772, + 0.671, + 0.784 + ], + "angle": 0, + "content": "Table 6: Results across different quantization bits." + }, + { + "type": "table", + "bbox": [ + 0.322, + 0.785, + 0.672, + 0.837 + ], + "angle": 0, + "content": "
SizebasenewH
QCoOp (b=1)0.26KB79.4972.6575.92
QCoOp (b=2)0.52KB80.3071.9875.91
QCoOp (b=4)1.05KB80.9271.2275.76
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "T. Hao et al." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.16, + 0.473, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.16, + 0.734, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.292, + 0.788, + 0.334 + ], + "angle": 0, + "content": "Fig. 7: The same training curves with Fig. 2. Clearly, increasing quantization noise leads to the same phenomenon with increasing gaussian noise, i.e. more generalizability (new accuracy) and less adaptability (base accuracy)." + }, + { + "type": "table_caption", + "bbox": [ + 0.352, + 0.338, + 0.649, + 0.351 + ], + "angle": 0, + "content": "Table 7: Results based on SLIP [53] model." + }, + { + "type": "table", + "bbox": [ + 0.339, + 0.351, + 0.657, + 0.39 + ], + "angle": 0, + "content": "
SizebasenewH
CoOp4.1KB68.6546.6055.51
QCoOp0.26KB68.3374.0471.07
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.785, + 0.449 + ], + "angle": 0, + "content": "higher than CoOp. Such observation verifies our assumption that quantization is quite helpful for generalization again." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.471, + 0.36, + 0.487 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.502, + 0.789, + 0.804 + ], + "angle": 0, + "content": "With the development of huge vision-language models, how to effectively and efficiently adapt such huge models to downstream tasks becomes a challenging problem. Much effort has been made to leverage the potential of prompt tuning in adapting vision-language models. However, existing methods suffer from inefficiency. To reach extremely efficient generalization, we propose QPrompt based on the detailed analysis and deep understanding of the characteristics of the prompt weight distribution. Following the several principles concluded by us, we use K-Means clustering algorithm as the base of our quantization method. To adaptively control the quantization error and minimize the number of re-clustering operations, we propose to do a dynamic check every few iterations. If the Kullback-Leibler divergence between the current weights and the original weights used in the last clustering exceeds a predefined threshold and the quantization error also increases, we will let the model re-cluster the weights and update the new centers and weights. Similarly, continue this process until completion. The proposed method could be simply integrated into many of the existing vision-language tuning methods like CoOp and MaPLe and reach good performance. Importantly, our proposed method is significantly effective and efficient. Extensive experiments show that the designed quantization algorithm indeed improves the generalizability of vision-language model, and save a lot more storage space as well." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by National Natural Science Foundation of China (Nos. 61925107, 62271281, 62021002)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.239, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Quantized Prompt for Efficient Generalization of Vision-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.173, + 0.786, + 0.214 + ], + "angle": 0, + "content": "1. Bahng, H., Jahanian, A., Sankaranarayanan, S., Isola, P.: Visual prompting: Modifying pixel space to adapt pre-trained models. arXiv preprint arXiv:2203.17274 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.215, + 0.787, + 0.255 + ], + "angle": 0, + "content": "2. Banner, R., Nahshan, Y., Soudry, D.: Post training 4-bit quantization of convolutional networks for rapid-deployment. Advances in Neural Information Processing Systems 32 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.256, + 0.786, + 0.31 + ], + "angle": 0, + "content": "3. Bhalgat, Y., Lee, J., Nagel, M., Blankevoort, T., Kwak, N.: Lsq+: Improving low-bit quantization through learnable offsets and better initialization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 696-697 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.786, + 0.351 + ], + "angle": 0, + "content": "4. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. In: The Eleventh International Conference on Learning Representations (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.351, + 0.786, + 0.403 + ], + "angle": 0, + "content": "5. Bulat, A., Tzimiropoulos, G.: Lasp: Text-to-text optimization for language-aware soft prompting of vision & language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23232-23241 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.786, + 0.445 + ], + "angle": 0, + "content": "6. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.446, + 0.786, + 0.487 + ], + "angle": 0, + "content": "7. Ding, J., Xue, N., Xia, G.S., Dai, D.: Decoupling zero-shot semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11583-11592 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.487, + 0.786, + 0.528 + ], + "angle": 0, + "content": "8. Ding, X., Ding, G., Guo, Y., Han, J.: Centripetal sgd for pruning very deep convolutional networks with complicated structure. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4943-4953 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.528, + 0.786, + 0.569 + ], + "angle": 0, + "content": "9. Ding, X., Guo, Y., Ding, G., Han, J.: Acnet: Strengthening the kernel skeletons for powerful cnn via asymmetric convolution blocks. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1911-1920 (2019) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.569, + 0.786, + 0.621 + ], + "angle": 0, + "content": "10. Ding, X., Hao, T., Tan, J., Liu, J., Han, J., Guo, Y., Ding, G.: Resrep: Lossless cnn pruning via decoupling remembering and forgetting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4510-4520 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.622, + 0.786, + 0.663 + ], + "angle": 0, + "content": "11. Ding, X., Zhang, X., Han, J., Ding, G.: Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11963-11975 (2022) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.663, + 0.786, + 0.704 + ], + "angle": 0, + "content": "12. Ding, X., Zhang, X., Ma, N., Han, J., Ding, G., Sun, J.: Repvgg: Making vgg-style convnets great again. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13733-13742 (2021) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.705, + 0.786, + 0.759 + ], + "angle": 0, + "content": "13. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020) 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.759, + 0.786, + 0.8 + ], + "angle": 0, + "content": "14. Esser, S.K., McKinstry, J.L., Bablani, D., Appuswamy, R., Modha, D.S.: Learned step size quantization. In: International Conference on Learning Representations (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.8, + 0.786, + 0.841 + ], + "angle": 0, + "content": "15. Feng, C., Zhong, Y., Jie, Z., Chu, X., Ren, H., Wei, X., Xie, W., Ma, L.: Promptdet: Towards open-vocabulary detection using uncurated images. In: European Conference on Computer Vision. pp. 701-717. Springer (2022) 3" + }, + { + "type": "list", + "bbox": [ + 0.219, + 0.173, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "T. Hao et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "16. Finkelstein, A., Almog, U., Grobman, M.: Fighting quantization bias with bias. arXiv preprint arXiv:1906.03193 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "17. Gholami, A., Kim, S., Dong, Z., Yao, Z., Mahoney, M.W., Keutzer, K.: A survey of quantization methods for efficient neural network inference. In: Low-Power Computer Vision, pp. 291-326. Chapman and Hall/CRC (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.219, + 0.785, + 0.26 + ], + "angle": 0, + "content": "18. Han, S., Mao, H., Dally, W.J.: Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149 (2015) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.261, + 0.785, + 0.301 + ], + "angle": 0, + "content": "19. Hao, T., Chen, H., Guo, Y., Ding, G.: Consolidator: Mergeable adapter with grouped connections for visual adaptation. arXiv preprint arXiv:2305.00603 (2023) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.304, + 0.785, + 0.346 + ], + "angle": 0, + "content": "20. Hao, T., Ding, X., Han, J., Guo, Y., Ding, G.: Manipulating identical filter redundancy for efficient pruning on deep and complicated cnn. IEEE Transactions on Neural Networks and Learning Systems (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.785, + 0.389 + ], + "angle": 0, + "content": "21. Hao, T., Lyu, M., Chen, H., Zhao, S., Han, J., Ding, G.: Re-parameterized low-rank prompt: Generalize a vision-language model within \\(0.5\\mathrm{k}\\) parameters. arXiv preprint arXiv:2312.10813 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.389, + 0.785, + 0.431 + ], + "angle": 0, + "content": "22. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.432, + 0.785, + 0.485 + ], + "angle": 0, + "content": "23. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.487, + 0.785, + 0.529 + ], + "angle": 0, + "content": "24. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.785, + 0.572 + ], + "angle": 0, + "content": "25. Hubara, I., Nahshan, Y., Hanani, Y., Banner, R., Soudry, D.: Accurate post training quantization with small calibration sets. In: International Conference on Machine Learning. pp. 4466-4475. PMLR (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.573, + 0.785, + 0.627 + ], + "angle": 0, + "content": "26. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International Conference on Machine Learning. pp. 4904-4916. PMLR (2021) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.629, + 0.785, + 0.656 + ], + "angle": 0, + "content": "27. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. arXiv preprint arXiv:2203.12119 (2022) 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.785, + 0.699 + ], + "angle": 0, + "content": "28. Jiang, Z., Araki, J., Ding, H., Neubig, G.: How can we know when language models know? on the calibration of language models for question answering. Transactions of the Association for Computational Linguistics 9, 962-977 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.785, + 0.741 + ], + "angle": 0, + "content": "29. Jie, S., Wang, H., Deng, Z.H.: Revisiting the parameter efficiency of adapters from the perspective of precision redundancy. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17217-17226 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.785, + 0.784 + ], + "angle": 0, + "content": "30. Ju, C., Han, T., Zheng, K., Zhang, Y., Xie, W.: Prompting visual-language models for efficient video understanding. In: European Conference on Computer Vision. pp. 105-124. Springer (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "31. Jung, S., Son, C., Lee, S., Son, J., Han, J.J., Kwak, Y., Hwang, S.J., Choi, C.: Learning to quantize deep networks by optimizing quantization intervals with task loss. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4350-4359 (2019) 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.239, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Quantized Prompt for Efficient Generalization of Vision-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.189 + ], + "angle": 0, + "content": "32. Khattak, M.U., Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Maple: Multi-modal prompt learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19113-19122 (2023) 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.244 + ], + "angle": 0, + "content": "33. Khattak, M.U., Wasim, S.T., Naseer, M., Khan, S., Yang, M.H., Khan, F.S.: Self-regulating prompts: Foundational model adaptation without forgetting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15190-15200 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.786, + 0.286 + ], + "angle": 0, + "content": "34. Lee, D., Song, S., Suh, J., Choi, J., Lee, S., Kim, H.J.: Read-only prompt optimization for vision-language few-shot learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1401-1411 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.786, + 0.314 + ], + "angle": 0, + "content": "35. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.786, + 0.355 + ], + "angle": 0, + "content": "36. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=RriDjddCLN 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.786, + 0.384 + ], + "angle": 0, + "content": "37. Li, H., Kadav, A., Durdanovic, I., Samet, H., Graf, H.P.: Pruning filters for efficient convnets. arXiv preprint arXiv:1608.08710 (2016) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.786, + 0.411 + ], + "angle": 0, + "content": "38. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.786, + 0.452 + ], + "angle": 0, + "content": "39. Li, Y., Xu, S., Zhang, B., Cao, X., Gao, P., Guo, G.: Q-vit: Accurate and fully quantized low-bit vision transformer. Advances in Neural Information Processing Systems 35, 34451-34463 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.786, + 0.493 + ], + "angle": 0, + "content": "40. Li, Z., Xiao, J., Yang, L., Gu, Q.: Repq-vit: Scale reparameterization for posttraining quantization of vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17227-17236 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.786, + 0.535 + ], + "angle": 0, + "content": "41. Lian, D., Zhou, D., Feng, J., Wang, X.: Scaling & shifting your features: A new baseline for efficient model tuning. In: Advances in Neural Information Processing Systems (NeurIPS) (2022) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.786, + 0.577 + ], + "angle": 0, + "content": "42. Liu, P., Yuan, W., Fu, J., Jiang, Z., Hayashi, H., Neubig, G.: Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. arXiv preprint arXiv:2107.13586 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.786, + 0.618 + ], + "angle": 0, + "content": "43. Liu, X., Ji, K., Fu, Y., Du, Z., Yang, Z., Tang, J.: P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv preprint arXiv:2110.07602 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.786, + 0.646 + ], + "angle": 0, + "content": "44. Liu, X., Zheng, Y., Du, Z., Ding, M., Qian, Y., Yang, Z., Tang, J.: Gpt understands, too. arXiv preprint arXiv:2103.10385 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.786, + 0.688 + ], + "angle": 0, + "content": "45. Liu, Z., Wang, Y., Han, K., Zhang, W., Ma, S., Gao, W.: Post-training quantization for vision transformer. Advances in Neural Information Processing Systems 34, 28092-28103 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.786, + 0.73 + ], + "angle": 0, + "content": "46. Liu, Z., Li, J., Shen, Z., Huang, G., Yan, S., Zhang, C.: Learning efficient convolutional networks through network slimming. In: Proceedings of the IEEE international conference on computer vision. pp. 2736-2744 (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.786, + 0.771 + ], + "angle": 0, + "content": "47. Lüddecke, T., Ecker, A.: Image segmentation using text and image prompts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7086-7096 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.786, + 0.813 + ], + "angle": 0, + "content": "48. Lyu, M., Hao, T., Xu, X., Chen, H., Han, J., Ding, G.: Learn from the learnt: Source-free active domain adaptation via contrastive sampling and visual persistence. In: European Conference on Computer Vision (ECCV). Springer (2024) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.786, + 0.84 + ], + "angle": 0, + "content": "49. Lyu, M., Yang, Y., Hong, H., Chen, H., Jin, X., He, Y., Xue, H., Han, J., Ding, G.: One-dimensional adapter to rule them all: Concepts diffusion models and erasing" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "T. Hao et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.147, + 0.786, + 0.175 + ], + "angle": 0, + "content": "applications. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7559-7568 (2024) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.787, + 0.217 + ], + "angle": 0, + "content": "50. Lyu, M., Zhou, J., Chen, H., Huang, Y., Yu, D., Li, Y., Guo, Y., Guo, Y., Xiang, L., Ding, G.: Box-level active detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23766-23775 (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.786, + 0.259 + ], + "angle": 0, + "content": "51. Maaz, M., Rasheed, H., Khan, S., Khan, F.S., Anwer, R.M., Yang, M.H.: Class-agnostic object detection with multi-modal transformer. In: The European Conference on Computer Vision. Springer (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.26, + 0.785, + 0.3 + ], + "angle": 0, + "content": "52. Meller, E., Finkelstein, A., Almog, U., Grobman, M.: Same, same but different: Recovering neural network quantization error through weight factorization. In: International Conference on Machine Learning. pp. 4486-4495. PMLR (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.301, + 0.786, + 0.342 + ], + "angle": 0, + "content": "53. Mu, N., Kirillov, A., Wagner, D., Xie, S.: Slip: Self-supervision meets language-image pre-training. In: European Conference on Computer Vision. pp. 529-544. Springer (2022) 1, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.786, + 0.383 + ], + "angle": 0, + "content": "54. Nagel, M., Baalen, M.v., Blankevoort, T., Welling, M.: Data-free quantization through weight equalization and bias correction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1325-1334 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.384, + 0.786, + 0.424 + ], + "angle": 0, + "content": "55. Nagel, M., Fournarakis, M., Amjad, R.A., Bondarenko, Y., Van Baalen, M., Blankevoort, T.: A white paper on neural network quantization. arXiv preprint arXiv:2106.08295 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.786, + 0.466 + ], + "angle": 0, + "content": "56. Nagel, M., Fournarakis, M., Bondarenko, Y., Blankevoort, T.: Overcoming oscillations in quantization-aware training. In: International Conference on Machine Learning. pp. 16318-16330. PMLR (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.786, + 0.508 + ], + "angle": 0, + "content": "57. Qian, R., Li, Y., Xu, Z., Yang, M.H., Belongie, S., Cui, Y.: Multimodal open-vocabulary video classification via pre-trained vision and language models. arXiv preprint arXiv:2207.07646 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.509, + 0.786, + 0.563 + ], + "angle": 0, + "content": "58. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021) 1, 3, 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.565, + 0.786, + 0.618 + ], + "angle": 0, + "content": "59. Rao, Y., Zhao, W., Chen, G., Tang, Y., Zhu, Z., Huang, G., Zhou, J., Lu, J.: Denseclip: Language-guided dense prediction with context-aware prompting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18082-18091 (2022) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.62, + 0.786, + 0.674 + ], + "angle": 0, + "content": "60. Rasheed, H.A., Maaz, M., Khattak, M.U., Khan, S., Khan, F.: Bridging the gap between object and image-level representations for open-vocabulary detection. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=aKXBrjODHm 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.675, + 0.786, + 0.729 + ], + "angle": 0, + "content": "61. Shin, T., Razeghi, Y., Logan IV, R.L., Wallace, E., Singh, S.: Autoprompt: Eliciting knowledge from language models with automatically generated prompts. In: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). pp. 4222-4235 (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.73, + 0.786, + 0.771 + ], + "angle": 0, + "content": "62. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research 15(1), 1929-1958 (2014) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.772, + 0.786, + 0.799 + ], + "angle": 0, + "content": "63. Xian, Y., Schiele, B., Akata, Z.: Zero-shot learning-the good, the bad and the ugly. In: CVPR (2017) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "64. Xiong, Y., Chen, H., Hao, T., Lin, Z., Han, J., Zhang, Y., Wang, G., Bao, Y., Ding, G.: Pyra: Parallel yielding re-activation for training-inference efficient task adaptation. arXiv preprint arXiv:2403.09192 (2024) 3" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.239, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Quantized Prompt for Efficient Generalization of Vision-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "65. Xiong, Y., Chen, H., Lin, Z., Zhao, S., Ding, G.: Confidence-based visual dispersal for few-shot unsupervised domain adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11621-11631 (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "66. Xiong, Y., Chen, X., Ye, X., Chen, H., Lin, Z., Lian, H., Niu, J., Ding, G.: Temporal scaling law for large language models. arXiv preprint arXiv:2404.17785 (2024) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.218, + 0.787, + 0.259 + ], + "angle": 0, + "content": "67. Yao, H., Zhang, R., Xu, C.: Visual-language prompt tuning with knowledge-guided context optimization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6757-6767 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.787, + 0.313 + ], + "angle": 0, + "content": "68. Yao, L., Huang, R., Hou, L., Lu, G., Niu, M., Xu, H., Liang, X., Li, Z., Jiang, X., Xu, C.: FILIP: Fine-grained interactive language-image pre-training. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=cpDhcsEDC2 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.314, + 0.787, + 0.355 + ], + "angle": 0, + "content": "69. Yu, S., Chen, T., Shen, J., Yuan, H., Tan, J., Yang, S., Liu, J., Wang, Z.: Unified visual transformer compression. In: International Conference on Learning Representations (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.356, + 0.787, + 0.397 + ], + "angle": 0, + "content": "70. Yuan, L., Chen, D., Chen, Y.L., Codella, N., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., et al.: Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.398, + 0.787, + 0.439 + ], + "angle": 0, + "content": "71. Yuan, Z., Xue, C., Chen, Y., Wu, Q., Sun, G.: Ptg4vit: Post-training quantization for vision transformers with twin uniform quantization. In: European Conference on Computer Vision. pp. 191-207. Springer (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.439, + 0.787, + 0.467 + ], + "angle": 0, + "content": "72. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary detr with conditional matching. arXiv preprint arXiv:2203.11876 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.467, + 0.787, + 0.522 + ], + "angle": 0, + "content": "73. Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., Beyer, L.: Lit: Zero-shot transfer with locked-image text tuning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18123-18133 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.787, + 0.55 + ], + "angle": 0, + "content": "74. Zhang, Y., Zhou, K., Liu, Z.: Neural prompt search. arXiv preprint arXiv:2206.04673 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.55, + 0.787, + 0.605 + ], + "angle": 0, + "content": "75. Zheng, K., Wu, W., Feng, R., Zhu, K., Liu, J., Zhao, D., Zha, Z.J., Chen, W., Shen, Y.: Regularized mask tuning: Uncovering hidden knowledge in pre-trained vision-language models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11663-11673 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.606, + 0.787, + 0.646 + ], + "angle": 0, + "content": "76. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.646, + 0.787, + 0.687 + ], + "angle": 0, + "content": "77. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022) 3, 4, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.687 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Quantized Prompt for Efficient Generalization of Vision-Language Models/d406d327-feb0-4aee-91eb-de3562f63166_origin.pdf b/2024/Quantized Prompt for Efficient Generalization of Vision-Language Models/d406d327-feb0-4aee-91eb-de3562f63166_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..847aee0cd87f599782a5a9ddface3bd9a49e3836 --- /dev/null +++ b/2024/Quantized Prompt for Efficient Generalization of Vision-Language Models/d406d327-feb0-4aee-91eb-de3562f63166_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c97f3be57359121c46e62bac1f82b951000bb3dcc0df3321f66ca10f1e24292 +size 1308358 diff --git a/2024/Quantized Prompt for Efficient Generalization of Vision-Language Models/full.md b/2024/Quantized Prompt for Efficient Generalization of Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f9199f9295244326490e534a6585bbca5553d633 --- /dev/null +++ b/2024/Quantized Prompt for Efficient Generalization of Vision-Language Models/full.md @@ -0,0 +1,327 @@ +# Quantized Prompt for Efficient Generalization of Vision-Language Models + +Tianxiang Hao $^{1,2}$ , Xiaohan Ding $^{3(\boxtimes)}$ , Juexiao Feng $^{1,2}$ , Yuhong Yang $^{1,2}$ , Hui Chen $^{2}$ , and Guiguang Ding $^{1,2(\boxtimes)}$ + +1 School of Software, Tsinghua University, Beijing, China +2 BNrist, Beijing, China +3 Bytedance + +{beyondtx,xiaohding,suise.con,jichenhui}@gmail.com +fjx20@mails.tsinghua.edu.cn, dinggg@tsinghua.edu.cn + +Abstract. In the past few years, large-scale pre-trained vision-language models like CLIP have achieved tremendous success in various fields. Naturally, how to transfer the rich knowledge in such huge pre-trained models to downstream tasks and datasets becomes a hot topic. During downstream adaptation, the most challenging problems are overfitting and catastrophic forgetting, which can cause the model to overly focus on the current data and lose more crucial domain-general knowledge. Existing works use classic regularization techniques to solve the problems. As solutions become increasingly complex, the ever-growing storage and inference costs are also a significant problem that urgently needs to be addressed. While in this paper, we start from an observation that proper random noise can suppress overfitting and catastrophic forgetting. Then we regard quantization error as a kind of noise, and explore quantization for regularizing vision-language model, which is quite efficiency and effective. Furthermore, to improve the model's generalization capability while maintaining its specialization capacity at minimal cost, we deeply analyze the characteristics of the weight distribution in prompts, conclude several principles for quantization module design and follow such principles to create several competitive baselines. The proposed method is significantly efficient due to its inherent lightweight nature, making it possible to adapt on extremely resource-limited devices. Our method can be fruitfully integrated into many existing approaches like MaPLe, enhancing accuracy while reducing storage overhead, making it more powerful yet versatile. Extensive experiments on 11 datasets shows great superiority of our method sufficiently. Code is available at github + +Keywords: Quantization $\cdot$ Generalization $\cdot$ Vision-language model + +# 1 Introduction + +Recently, deep learning models and related technologies have seen rapid development [9, 11-13, 22-24, 49, 50, 65, 66]. Vision-language model (VLM) like CLIP [58] is one of the hottest research topics and leads to huge success. The excellent generalization ability is a crucial cornerstone of such achievement [26, 53, 58, 59, 68]. + +![](images/c3201a6a53517b9e1d48b08e43469486061b1f3fbbebf67fdcf5733f50b11212.jpg) +Fig. 1: Overall performance comparison with existing vision-language tuning methods. Our method outperforms all of the state-of-the-art competitors with significantly fewer storage space. Based on the proposed quantization algorithm, our method could be integrated into many of the existing methods and bring consistent improvements with excellent efficiency. + +When people have access to downstream data, it is better to tune the pretrained VLM on the target dataset for higher accuracy. However, such full finetuning could easily cause the model to overfit the small downstream dataset and face catastrophic forgetting problem, leading to severe performance drop. + +To solve the problem, in this paper we will start from rethinking the relationship between noise and generalization. We propose to think noise as a kind of regularization techniques, which may be helpful for alleviating overfitting and catastrophic forgetting problem. As in Sec. 3.2 and Fig. 2, we then find that directly adding some random Gaussian noise to the tunable prompts of vision-language models would result in a performance gain in several cases, which partly validates our conjecture about using noise as a form of regularization. In particular, excessive noise diminishes the model's adaptation capability, while insufficient noise fails to provide effective regularization. Only noise of moderate intensity is beneficial for the model's generalization. + +However, Gaussian noise is absolutely random and hard to control, and thus it is difficult for us to take full advantage of such type of noise to benefit generalization. Instead, we point out that quantization error is also a form of noise, and therefore, it is also possible to leverage quantization error to enhance the model's generalization performance. With this idea in mind, we thoroughly analyzed the distribution pattern of prompt weights in Sec. 3.3 and derived several design principles for quantization algorithms based on the observed phenomena. In Sec. 4, following the principles we summarized, we successfully designed an efficient quantization-aware training algorithm, which largely enhances the model's generalization ability while quantizing. + +Since our design is general, we could fruitfully integrate our quantization strategy into many existing methods and reach a higher accuracy with much smaller model size due to the lightweight nature of quantization. For example, we integrate our method onto an existing popular method MaPLe, and our QMaPLe earns $0.57\%$ accuracy improvement with only $0.25 \times$ size. + +In conclusion, we summarize our contribution as follows: + +- We deeply analyze the effect of noise and rethink the relationship between noise and generalization in Sec. 3.2 for vision-language models. As a result, we confirm that moderate noise would promote the model generalization +- We are the first to propose to quantize prompts. By detailed observation and hard thinking, we conclude several principles about how to effectively design a quantization method for the prompts of vision-language models. +- Following the principles we concluded, we build our method and successfully quantize the prompts as well as some other weights, and our method significantly outperforms existing ones as in Fig. 1. Extensive experiments show great power of our method. In base-to-new generalization, domain generalization, cross-dataset transfer and few-shot learning settings, we consistently reach competitive results, winning many state-of-the-art tuning methods with a much smaller size of model. + +# 2 Related Works + +# 2.1 Vision-Language Models + +In recent times, large-scale vision-language models have demonstrated remarkable performance across various tasks. Seminal works such as [26, 58, 68, 70, 73]. Classic works have focused on learning multimodal representations through self-supervised methods using extensive sets of image-text pairs. Among these, CLIP [58] stands out as a milestone achievement, employing contrastive learning to align vision and language representations and achieving exceptional performance. A well-trained vision-language model is invaluable, offering substantial support to various fields. Successful applications of these robust models include few-shot recognition [76, 77], detection tasks [15, 51, 60, 72], and segmentation tasks [7, 36, 47, 59]. Furthermore, for video data, research efforts have emerged in video classification [57] and video understanding [30]. + +# 2.2 Parameter-Efficient Fine-Tuning + +Recently, a series of works [19, 21, 23, 24, 27, 29, 38, 41, 44, 48, 64, 76, 77] have been proposed to help transfer the learnt knowledge, where one of the most popular field is parameter-efficient fine-tuning (PEFT). PEFT aims to transfer a pre-trained model to downstream tasks by a minimum number of parameters. Originating from natural language processing tasks, classic methods like adapter [23], prompt tuning [28, 35, 38, 42, 43, 61] and LoRA [24] follows a similar + +principle to add extra modules with a small number of parameters into the backbone model, freeze the original parameters and only tune and store the newly added parameters. Inspired by the success of PEFT in language field, researchers have extended such kind of approaches to adapt visual models in a similar fashion [6,19,27,41,74]. In the field of vision-language modeling, several explorations have been made as well. Bahng et al. [1] exclusively apply prompt tuning to the image encoder. CoOp [77] replaces the fixed template in CLIP [58] with tunable text prompts. CoCoOp [76] leverages image features to guide the optimization of tunable text prompts in CoOp. Other works [32, 34] optimize both image and text prompts simultaneously and establish additional connections between different modalities. To mitigate overfitting and catastrophic forgetting, various works [5,33,67,75] integrate regularization modules or losses into prompt tuning. + +# 2.3 Quantization + +Quantization is one of the most effective compression methods [4,8,10,20,37, 46,69] for deep learning models. Generally, parameters such as weights and activations are typically stored as 32-bit floating-point numbers, which consume a significant amount of memory and require intensive computation during inference. Quantization [17, 18, 55] involves representing these parameters with reduced precision, such as 8-bit integers or even lower bit-widths. By doing so, quantization can significantly reduce the memory footprint and computational complexity of the model without significantly sacrificing accuracy. Quantization methods can be divided into two groups, Post-Training Quantization (PTQ) [2,16,25,40,45,52,54,71] that consumes few resources but suffers higher accuracy loss, and Quantization-Aware Training (QAT) [3,14,31,39,56] that relies on plenty of resources for training and shows better accuracy. Existing works aim to minimize quantization error to improve accuracy, while our work demonstrates that both excessive and insufficient errors are detrimental to model generalization. To achieve optimal generalization performance, a moderate error is required. + +# 3 Exploring Quantization in Model Generalization + +# 3.1 Preliminaries: Prompt Tuning of Vision-Language Models + +CLIP comprises a text encoder $\mathcal{L}$ and an image encoder $\mathcal{V}$ . Typically, $\mathcal{L}$ is implemented as a language transformer, whereas $\mathcal{V}$ may be realized using either a convolutional neural network or a vision transformer. In this study, following the methodologies outlined by [76, 77], we employ a ViT-B/16 model [13] as the image encoder $\mathcal{V}$ , except where otherwise specified. The subsequent sections will provide a brief overview of the methods used to prompt CLIP for prediction tasks. + +Text Encoder Consider a text encoder composed of $M$ layers. For the $k$ -th layer, denoted as $\mathcal{L}_k$ , the inputs consist of a sequence of prompt tokens $P_{k - 1}^{l}$ + +and a $[CLS]$ token $c_{k-1}^{l}$ , while the outputs are represented by $P_{k}^{l}$ and $c_{k}^{l}$ . The initial inputs, $P_{0}^{l}$ and $c_{0}^{l}$ , correspond to the word embeddings of the prompts combined with the label, such as "A photo of a $[CLS]''$ or alternatively, some randomly initialized vectors. Formally, we denote $P_{k}^{l} \in \mathbb{R}^{n^{l} \times d^{l}}$ and $c_{k}^{l} \in \mathbb{R}^{d^{l}}$ , where $n^{l}$ signifies the length of the text prompts and $d^{l}$ represents the dimension of the word embeddings. For each layer, $1 \leq k \leq M$ , the relationship is given by $[P_{k}^{l}, c_{k}^{l}] = \mathcal{L}_{k}([P_{k-1}^{l}, c_{k-1}^{l}])$ . The output feature of the text encoder $f^{l} \in \mathbb{R}^{d^{v}}$ , where $d^{v}$ is the dimension of the visual feature space, is calculated by projecting the $[CLS]$ token of its last layer to the visual latent space through a linear transformation, i.e. $f^{l} = \mathrm{Proj}(c_{M}^{l})$ . + +Image Encoder Suppose there are $N$ layers in the image encoder. For $k$ -th layer $\mathcal{V}_k$ , the inputs are a series of image tokens $I_{k-1}$ , a classification token $c_{k-1}^v$ and prompt tokens $P_{k-1}^v$ , and the outputs are $I_k$ , $c_k^v$ and $P_{k}^{v}$ . The inputs of the first layer $I_0$ and $c_0^v$ are the patch embeddings of the input image and the pre-trained class token. $P_0^v$ is randomly initialized. Formally, $I_k \in \mathbb{R}^{p \times d^v}$ , $c_k^v \in \mathbb{R}^{d^v}$ and $P_k^v \in \mathbb{R}^{n^v \times d^v}$ , where $p$ denotes the number of image patches and $d^v$ denotes the dimension of visual embedding. $\forall 1 \leq k \leq N$ , $[P_k^v, c_k^v, I_k] = \mathcal{V}_k([P_{k-1}^v, c_{k-1}^v, I_{k-1}])$ . The output feature of the image encoder is $f^v = c_N^v$ . + +Prediction For image classification, suppose there are $C$ classes, and $\{f_c^l\}_{c = 1}^C$ are the text features. Label $y$ 's probability is $p(y|f^v) = \frac{\exp(\sin(f^v,f_l^i) / \tau)}{\sum_{c = 1}^{C}\exp(\sin(f^v,f_c^l) / \tau)}$ where $\sin (\cdot ,\cdot)$ denotes cosine similarity function and $\tau$ is temperature. The final prediction is $\hat{z} = \arg \max_{1\leq y < C}(p(y|f^v))$ . + +We have introduced shallow prompts. There are also different types of prompts. Several works [27, 32] use them for improve performance. They directly add and tune the prompt in each layer in the feature encoder, instead of inheriting the output prompt calculated by the last encoder. Now we have $[\_ ,c_k^l ] = \mathcal{L}_k([P_{k - 1}^l,c_{k - 1}^l])$ and $[\_ ,c_k^v,I_k] = \mathcal{V}_k([P_{k - 1}^v,c_{k - 1}^v,I_{k - 1}])$ . Note that $P^l$ and $P^v$ are independent tunable parameters. They are no longer determined by the previous layer. + +# 3.2 Rethinking the Relationship between Noise and Generalization + +Some existing work has explored how to use the addition of noise to suppress model overfitting, e.g., techniques like Dropout [62] that randomly drops connections, and random jitter that directly introduce perturbations in the input data. However, the impact of directly adding noise to model weights has been less explored, especially in the context of prompt structures, which have become popular only in recent years, in Transformer architectures. The variations of the specialization capability represented by the test accuracy on base, seen classes and the generalization capability represented by the test accuracy on new, unseen classes. The curves of different colors represent the data under the influence of random Gaussian noise of different intensities, e.g. "Noise_0.01" adds random noise with a distribution of $\mathcal{N}(0,0.01^2)$ to the prompt. "Noise_0" denotes the baseline prompt tuning. As training progresses, the generalization capability + +![](images/795d95e07398067757fe3533b9326a893b3088234d8b65bd463502db3a4d4768.jpg) +Fig. 2: The variations of the specialization capability represented by the test accuracy on base, seen classes and the generalization capability represented by the test accuracy on new, unseen classes on average of the ten datasets in base-to-new generalization setting in Sec. 5 except for time-consuming ImageNet. The curves of different colors represent the data under the influence of random Gaussian noise of different intensities, e.g. "Noise_0.01" adds random noise with a distribution of $\mathcal{N}(0,0.01^2)$ to the prompt. "Noise_0" denotes the baseline prompt tuning. As training progresses, the generalization capability of baseline prompt tuning continuously decreases while the specialization capability improves. Therefore, we expect that adding noise can achieve a better balance between generalization and specialization. However, excessive noise, e.g. 0.1, greatly diminishes the model's specialization capability, while insufficient noise, e.g. 0.001, fails to provide effective regularization. Only noise of moderate intensity outperforms baseline in specialization-generalization trade-off, effectively enhancing the unseen class accuracy without significantly compromising seen class accuracy. + +![](images/a10f97832116346084f99cc7315ae08e98e1238496ca61e2de859f49045f0004.jpg) + +of baseline continuously decreases while the specialization capability improves. Therefore, we expect that adding noise can achieve a better balance between generalization and specialization. However, excessive noise, e.g. 0.1, greatly diminishes the model's specialization capability, while insufficient noise, e.g. 0.001, fails to provide effective regularization. Only noise of moderate intensity outperforms baseline in specialization-generalization trade-off, effectively enhancing the unseen class accuracy without significantly compromising seen class accuracy. + +Quantization, the technique of mapping parameter values from high precision to low precision, can also be viewed as introducing some form of noise into the parameters, and thus can possibly improve the generalizability as the noise did in Fig. 2. Compared to Gaussian noise, quantization error is more controllable. Better still, quantization also has another major advantage: it can significantly reduce the storage required for parameters. Therefore, using quantization algorithms to generalize vision-language models is a very promising direction. + +# 3.3 Characteristics of Prompts in Vision-Language Models + +In this subsection, we will demonstrate the characteristics of prompts in the vision-language model to facilitate targeted design of quantization schemes. + +We start from analyzing the training procedure of CoOp [77], which freezes all the parameters in backbone and tunes the added prompts only. The histogram about the frequency of weights of prompts for different training epochs is shown + +![](images/59081e3f35601fe0e14fbb29ce3f8e076fc28a0420491ae46b6b8133466ef4e9.jpg) + +![](images/ee31c05f39048038f87beb9c76025c27ad8e3253677f963621a4f4cd9f886232.jpg) + +![](images/0cf7047e2eaa60eec58a465de3d9f36f6d7b6dcf67c9c92396088a36aa2db1e2.jpg) + +![](images/4de6c6cf6f2747070b02fb9ca1b9f14a279485613c55e3d7986e4ac1522bf200.jpg) + +![](images/4bbc9d7a2ee54cd2750e0c5bce3f4d178bb00f64776556f4e55093f79571ce50.jpg) + +![](images/1df42b99b9e7efb8e5bcd7c80663a1d48ca5312d5463ab5b671cabffe9ab2275.jpg) + +![](images/454cd5847e87043494fec38d3880f7a9ee648d62b74d6b8a181602bc574b86d9.jpg) + +![](images/66bf6a9ede6a2b896b2ece0985627f28319370c1269d79491dafcd28c98b3c92.jpg) +Fig. 3: The histogram about the frequency of weights of prompts during training of CoOp [77] on eurosat dataset. We find the shape of the prompt's weight distribution remains largely unchanged throughout the entire training, but the variance of the weight distribution increases rapidly at the beginning of training. Additionally, we also notice that there are almost no outliers in the prompt's weights throughout the entire training process, and apart from the unstable initial training phase, the changes in weights between adjacent phases are not significant, indicating a very gentle overall updating trend. + +![](images/583972c5c62d0a30dbcda92ce9ce68fd94e26db29ed99bed01c2483148f9cf8e.jpg) + +![](images/90e6e9dc8d9477c7e68bad3002e0f806846119b45a69f4a7ccfbfb5defca68f7.jpg) + +in Fig. 3. We observe that the shape of the prompt's weight distribution remains mostly consistent throughout the entire training process. However, the variance of the weight distribution increases rapidly at the initial stages of training. Refer to Fig. 6 for more details. Moreover, we notice minimal presence of outliers in the prompt's weights throughout the training process. Apart from the unstable initial training phase, the changes in weights between consecutive phases are not substantial, indicating a very gentle overall updating trend. + +Taking into account the observations from Secs. 3.2 and 3.3, we can draw the following conclusions: + +- Because the prompt weights are not sensitive to noise, and moderate noise in the current scenario not only does not degrade the model's performance but actually enhances its generalization ability, we can possibly adopt some quantization strategies that are considered very aggressive and highly likely to significantly degrade performance, such as 1-bit quantization, without causing destructive effects on the model's generalization ability. +- Given that the parameters targeted for quantization in the current scenario constitute only a tiny fraction of the total model parameters, we speculate that QAT may have more advantages over PTQ, as any additional training costs incurred by QAT which is considered a major drawback in traditional situations would naturally be kept at a low level due to the low proportion of parameters targeted for quantization. Due to the observation in Fig. 2 that both excessive and insufficient noise are detrimental to generalization, in designing the QAT algorithm, we no longer need to always minimize quantization error as the objective, as in classic algorithms. Instead, we aim to maintain the quantization error at a moderate level. Furthermore, as seen in Fig. 3, the changes in prompt weight distribution during most of the training time are gentle, indicating that we can appropriately continue to use past-time algorithm state without the need for real-time updates. + +![](images/85258a1dff8fc578d8697103b18340416c413713e8e0e96b55961cd5383fae1a.jpg) +Fig. 4: Overview of Quantized Prompt. We set $b = 2$ for a clear explanation. Normalization and denormalization processes are not shown here. + +- Throughout the training process, the shape of the prompt weight distribution remains roughly unchanged, with the main variation coming from the distribution's variance. Therefore, it may be beneficial to attempt to eliminate the translation and scaling transformations of the distribution by normalizing the distribution before quantization and denormalizing it after quantization, which might help improve quantization accuracy. +- Since there are almost no outliers throughout the entire training process, some classic algorithms are revitalized. Specifically, clustering algorithms like K-Means are good choices, because one of the major drawbacks of them in traditional usage is that they are greatly affected by outliers. In addition, as we have analyzed before, there is not a high demand for dynamically minimizing quantization error in vision-language generalization, opting for a more efficient non-parametric clustering approach seems a better choice compared with some complex parametric approaches. + +# 4 Prompt Quantization + +# 4.1 Preliminaries: Quantization Basis + +Suppose there are $m$ parameters in total to be quantized, which are denoted by $W \in \mathbb{R}^m$ . Each $w_i$ here is a high-precision float-type number. A $b$ -bit model quantization algorithm divides the value range of parameters into $2^b$ intervals $\{\mathcal{U}_i, 1 \leq i \leq 2^b\}$ , aiming to find a mapping $Q$ from intervals to points, mapping all values within an interval $\mathcal{U}_i$ to the same quantized value $q_i$ , i.e. $Q(x) = q_i, \forall x \in \mathcal{U}_i$ . We define the quantization error $E = \sum_{i=1}^{N} ||Q(W_i) - W_i||^2$ . Such error serves as the objective of K-Means algorithm as well, so it can be ensured that each time clustering is redone using the K-Means algorithm, the quantization error will most likely decrease. + +# 4.2 Overview of Quantized Prompt + +Based on the observation and analysis in Sec. 3.3, we build a simple yet quite proper quantization algorithm for generalizing vision-language models. As stated before, we choose a widely-used clustering method K-Means, to construct the mapping $Q$ . Initially, K-Means algorithm is run to fit the pre-trained prompt weights, and record the codebook, i.e. cluster centers. Normalization and denormalization are conducted beyond quantization to alleviate the influence of varying variance. Formally, given the original prompt $W$ with $N$ parameters, we first calculate $\mu = \frac{1}{N}\sum_{i=1}^{N}W_i$ and $\sigma = \sqrt{\frac{\sum_{1\leq i\leq N}(W_i - \mu)^2}{N}}$ , then normalize $W$ and get $\hat{W} = \frac{W - \mu}{\sigma}$ . K-Means quantization is applied on $\hat{W}$ , and the quantized prompt $W_q = \sigma Q(\hat{W}) + \mu$ . Since $Q$ is not differentiable, we adopt a common practice to directly propagate the gradient across the quantization function by Straight-Through Estimator (STE), i.e. $\frac{\partial Q(x)}{\partial x} = x$ . + +The overall framework is shown in Fig. 4. In training, we keep tuning the weight by the fake gradient propagated from quantization operation. For storage, we first convert the fp16 parameters in prompt to $b$ -bit indexes, which could be used to search for the corresponding cluster center in the codebook. In Fig. 4, we set $b = 2$ for a clear explanation. Compared with baseline method, our quantized method could save a lot more storage space. Specifically, for storage, an ordinary method needs $16N$ bits, while ours only needs $bN + 2^b \times 16$ . In experiments, we usually set $b$ to 1, 2 or 4, which is far more smaller than $N$ . For example, with $b = 1$ , the storage space of our method is roughly $16 \times$ smaller than baseline. + +Note that the codebook here is updated by a rule called "constrained adaptive clustering". We will give a detailed description of it in the following subsection. + +# 4.3 Constrained Adaptive Clustering for Quantization + +In Fig. 4, re-clustering the prompt every iteration like classic QAT algorithms did to keep the codebook updated at all times is not a good choice. There are three reasons: 1. K-Means algorithm is not quite efficient, and thus if we run it every iteration then the training efficiency would decrease; 2. Only moderate noise promotes generalization, so keep updating the clusters to minimize the quantization error may be not helpful but harmful; 3. From Fig. 3, we can observe that the weight changes are very gentle for most of the training time, and the weight distributions of adjacent stages are highly similar. Therefore, even if we update the internal clustering state of K-Means with these already very similar data at each iteration, it's difficult to generate a better clustering solution, which is just futile effort. + +We propose constrained adaptive clustering to instruct the update of parameters inside K-Means which would not updated by gradient at all. Intuitively, first, we do not want too often update, so we set a minimum cluster update interval $t$ . Second, to avoid K-Means meaninglessly handling similar weights with current cached state, i.e. the weights of prompt that triggers the last re-clustering, we plan to only do re-clustering when current weight distribution is far different + +from the cached weight distribution. To reach the goal, Kullback-Leibler divergence is a good metric. However, if we want to compute the Kullback-Leibler divergence between two sets of discrete random variables, they must be defined on the same event space. However, that's not the case. So we first project each of them into the same event space spanning by the K-Means clustering algorithm. Specifically, given a current weight $W^{cur}$ , cached weight $W^{old}$ , quantization function $Q$ , cluster centers $\mathcal{C} = \{c_i, 1 \leq i \leq 2^b\}$ , we will compute the $W_{index}^{cur}$ and $W_{index}^{old}$ as in Fig. 4. Then, we can obtain the probability distributions of the indices $p^{cur}$ and $p^{old}$ , respectively, implied by $W_{index}^{cur}$ and $W_{index}^{old}$ , thus successfully transforming two originally unrelated discrete random variables into the same event space. The Kullback-Leibler divergence could be computed as follows: + +$$ +K L \left(p ^ {c u r} | | p ^ {o l d}\right) = \sum_ {i = 1} ^ {2 ^ {b}} p ^ {c u r} (i) \log \frac {p ^ {c u r} (i)}{p ^ {o l d} (i)} \tag {1} +$$ + +The update would continue only if Kullback-Leibler divergence exceeds a certain threshold $T_{KL}$ which shows the distribution difference is significant. + +# 5 Experiments + +We evaluate the method and make comparisons with the latest state-of-the-art methods in terms of the following settings across a wide range: + +1. Base-to-new generalization, which models are trained with base classes and evaluated on both base and new classes. +2. Domain generalization, which aims to show generalization to the domain shift, especially for out-of-distribution data. +3. Cross-dataset transfer, which aims to see if the method has the potential to transfer beyond a single dataset. It is a much more challenging problem because the fundamentals can be totally changed across different datasets. +4. Few-shot learning, which aims to evaluate the adaptation performance of the model to extract knowledge from a dataset whose samples are very few, e.g. 1, 2, 4, 8 or 16 samples. + +All methods are initialized with the same CLIP weights provided by the open-source CLIP [58]. In Appendix, we provide more details of datasets, experimental setting and competitors introduction. Due to page limit, we also put some detailed experimental results into Appendix, e.g. the performance of each method on each dataset in base-to-new generalization setting. + +# 5.1 Main Results + +Base-to-new Generalization The average results over 11 datasets are shown in Tab. 1. For complete results, please refer to the Appendix. Our proposed QCoOp reaches $77.43\%$ average harmonic mean accuracy within merely 0.26KB size. QCoOp outperforms all kinds of lightweight state-of-the-art methods with + +Table 1: Comparisons with latest methods in base-to-new generalization. H: harmonic mean [63]. QCoOp and QMaPLe are significantly better than other competitors, which prove that our quantization design could be fruitfully integrated into many existing approaches to further improve the performance and efficiency. + +
SizeBaseNewH
CLIP0KB69.3474.2271.70
CoOp4.1KB82.6963.2271.66
CoCoOp70.8KB80.4771.6975.83
Adapter1051KB82.6270.9776.35
LoRA258KB84.3067.3374.86
ProGrad16.4KB82.7968.5575.00
QCoOp0.26KB80.6874.4477.43
MaPLe7096KB82.2875.1478.55
QMaPLe1774KB83.0275.5779.12
+ +Table 2: Comparisons with latest methods in domain generalization. QCoOp gets comparable or even better results with the latest state-of-the-art methods with much fewer parameters, showing excellent robustness for domain shift. + +
SizeSourceImageNetTargetAverage
-V2-Sketch-Adversarial-Rendition
CLIP0KB66.7360.8346.1547.7773.9657.18
CoOp4.1KB71.5164.2047.9949.7175.2159.28
CoCoOp70.8KB71.0264.0748.7550.6376.1859.91
Adapter1051KB69.3362.5347.6749.1775.4258.70
LoRA258KB70.3062.3742.4338.4068.9753.04
ProGrad16.4KB72.2464.7347.6149.3974.5859.07
QCoOp0.26KB70.6763.8748.9351.1076.9060.20
+ +much more efficiency and higher accuracy. For heavier methods like MaPLe, our method can be fruitfully integrated into existing solutions. Besides prompts, we also perform a similar quantization operation on the other weights of MaPLe that are in the linear layers. As a result, QMaPLe not only shows stronger generalization and adaptation capability and gives $0.57\%$ higher accuracy, but also enjoys a much more smaller model size compared with the original MaPLe. + +Notably, when comparing QCoOp and a lightweight method ProGrad, even if QCoOp is $63 \times$ smaller than ProGrad, QCoOp still outperforms it by a clear margin, demonstrating the outstanding efficiency and effectiveness. + +Domain Generalization In this paragraph, ImageNet, ImageNet-A, ImageNet-R, ImageNet-v2, and ImageNet-S are used to construct domain generalization experiments. As shown in Tab. 2, on downstream target datasets, QCoOp gets better average accuracy compared with the other methods with significantly better efficiency. For CLIP, CoOp, CLIP-Adapter and ProGrad, there is a clear performance gap between our method and them. + +Cross-dataset Transfer In this paragraph, we do cross-dataset transfer evaluation to further verify our QPrompt. Results are shown in Tab. 3. CoOp is good on source domain but fails on target domains. Probably because it focuses too much on the dataset shown to its eyes and face overfitting and catastrophic + +Table 3: Results in the cross-dataset transfer setting. QCoOp gives the highest accuracy on 5 of 10 datasets, which well demonstrates that QCoOp could maximally extract general and data-agnostic knowledge from input images. + +
SizeSourceTargetAverage
ImageNetCaltech101PetsCarsFlowersFood101AircraftSun397DTDEuroSATUCF101
CoOp4.1KB71.5193.7089.1464.5168.7185.3018.4764.1541.9246.3966.5563.88
CoCoOp70.8KB71.0294.4390.1465.3271.8886.0622.9467.3645.7345.3768.2165.74
Adapter1051KB69.3393.4388.8764.4070.2785.6324.6765.8044.9047.7066.0065.17
QCoOp0.26KB70.6394.0790.5365.9771.3386.2322.7366.8044.2048.2369.1765.93
+ +![](images/8b65cd08a49e60b2a3556308655af1b59224075e9b15641d33174225a8b076f6.jpg) +Fig. 5: Average few-shot learning results on 11 datasets. + +![](images/e73a42a7e4c61104bd9f3503316a71777aff14598a86e8124426444a136919b3.jpg) +Fig. 6: KLD trend of prompt weights under the same experiment with Fig. 3. + +forgetting problems, which leads to a severe performance drop on unseen objects. QCoOp wins on 5 of 10 datasets and its average accuracy is also slightly better than the best competitor CoCoOp, showing that QCoOp could maximally extract both general and data-agnostic knowledge from given images. + +Few-shot Learning Here we will show the experiment results of QCoOp in the few-shot learning setting that is originated from CoOp. Seen from Fig. 5, QCoOp consistently outperforms zero-shot CLIP, CoOp, and CLIP-Adapter across all the shot numbers. Such results demonstrate the superiority of QCoOp in adaptation ability when there are few samples in downstream tasks. + +Overall, in base-to-new generalization, domain generalization, cross-dataset transfer and few-shot learning, the proposed method can consistently accomplish state-of-the-art performance while enjoying extremely high parameter efficiency, fruitfully demonstrating the effectiveness and efficiency of the proposed method. + +# 6 Analysis + +# 6.1 Ablation Studies + +Component In this subsection, we decompose the method into pieces and show the influence of each component. As in Tab. 4, we clearly show how + +Table 4: Ablation study on separate component of QCoOp. Norm: Normalization/Denormalization. CAC: Constrained Adaptive Clustering. + +
K-MeansNormCACbasenewH
QCoOp✓✗✗78.7172.5575.50
✓✓✗78.8473.0975.85
✓✓✓78.2474.0276.07
+ +Table 5: Comparisons between QAT and PTQ based on CoOp. Both QAT and CoOp are trained with the same hyper-parameters. Here the quantization bit is 1. + +
basenewH
QAT80.7272.3576.31
PTQ82.2168.5074.73
+ +much the K-Means algorithm, normalization/denormalization and constrained adaptive clustering influence the final performance. One interesting point is that K-Means+Norm+CAC only improves the new accuracy compared with K-Means+Norm, showing the superiority of our constrained adaptive clustering. + +Trend of the distributions of prompt between adjacent epochs. To verify the opinions we proposed at the end of Sec. 6, we show the KLD of prompt distributions between adjacent epochs during CoOp's training in Fig. 6. Clearly, this trend is consistent with what we summarized before. + +QAT v.s. PTQ In Tab. 5, we study the choice of QAT or PTQ following the same strategy, K-Means clustering, in the paper. Results show that QAT consistently outperforms PTQ with the same hyper-parameters. The accuracy on unseen new classes of PTQ is significantly lower than QAT, again proving our opinion that quantization helps generalization by alleviating overfitting as well as catastrophic forgetting. Quantization error is not always undesirable. + +Quantization bit In Tab. 6 and Fig. 7, we show the results across multiple quantization bits. In conclusions, more bits did not always lead to good results, and new accuracy continues decreasing as the training goes on. + +# 6.2 Performance on Self-Supervised Vision-Language Model + +In this paragraph, we explore the usage of QCoOp among other different backbones besides CLIP. We choose a self-supervised vision-language model, SLIP [53] to further verify the universality and robustness of our proposed method. The experimental results are shown in Tab. 7. We could see that the base accuracies of QCoOp and CoOp are similar, but the new accuracies of QCoOp is much + +Table 6: Results across different quantization bits. + +
SizebasenewH
QCoOp (b=1)0.26KB79.4972.6575.92
QCoOp (b=2)0.52KB80.3071.9875.91
QCoOp (b=4)1.05KB80.9271.2275.76
+ +![](images/0d118ce17f079fe7974dd01be1f854a44b3ee9131443371cc8a04cc25aded9c0.jpg) +Fig. 7: The same training curves with Fig. 2. Clearly, increasing quantization noise leads to the same phenomenon with increasing gaussian noise, i.e. more generalizability (new accuracy) and less adaptability (base accuracy). + +![](images/233c5cf195cce4aabca49cfa8561d3cdf849d18b4353cf4e2ce97bced1439ac6.jpg) + +Table 7: Results based on SLIP [53] model. + +
SizebasenewH
CoOp4.1KB68.6546.6055.51
QCoOp0.26KB68.3374.0471.07
+ +higher than CoOp. Such observation verifies our assumption that quantization is quite helpful for generalization again. + +# 7 Conclusion + +With the development of huge vision-language models, how to effectively and efficiently adapt such huge models to downstream tasks becomes a challenging problem. Much effort has been made to leverage the potential of prompt tuning in adapting vision-language models. However, existing methods suffer from inefficiency. To reach extremely efficient generalization, we propose QPrompt based on the detailed analysis and deep understanding of the characteristics of the prompt weight distribution. Following the several principles concluded by us, we use K-Means clustering algorithm as the base of our quantization method. To adaptively control the quantization error and minimize the number of re-clustering operations, we propose to do a dynamic check every few iterations. If the Kullback-Leibler divergence between the current weights and the original weights used in the last clustering exceeds a predefined threshold and the quantization error also increases, we will let the model re-cluster the weights and update the new centers and weights. Similarly, continue this process until completion. The proposed method could be simply integrated into many of the existing vision-language tuning methods like CoOp and MaPLe and reach good performance. Importantly, our proposed method is significantly effective and efficient. Extensive experiments show that the designed quantization algorithm indeed improves the generalizability of vision-language model, and save a lot more storage space as well. + +Acknowledgements. This work was supported by National Natural Science Foundation of China (Nos. 61925107, 62271281, 62021002) + +# References + +1. Bahng, H., Jahanian, A., Sankaranarayanan, S., Isola, P.: Visual prompting: Modifying pixel space to adapt pre-trained models. arXiv preprint arXiv:2203.17274 (2022) 4 +2. Banner, R., Nahshan, Y., Soudry, D.: Post training 4-bit quantization of convolutional networks for rapid-deployment. Advances in Neural Information Processing Systems 32 (2019) 4 +3. Bhalgat, Y., Lee, J., Nagel, M., Blankevoort, T., Kwak, N.: Lsq+: Improving low-bit quantization through learnable offsets and better initialization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 696-697 (2020) 4 +4. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. In: The Eleventh International Conference on Learning Representations (2023) 4 +5. Bulat, A., Tzimiropoulos, G.: Lasp: Text-to-text optimization for language-aware soft prompting of vision & language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23232-23241 (2023) 4 +6. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022) 4 +7. Ding, J., Xue, N., Xia, G.S., Dai, D.: Decoupling zero-shot semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11583-11592 (2022) 3 +8. Ding, X., Ding, G., Guo, Y., Han, J.: Centripetal sgd for pruning very deep convolutional networks with complicated structure. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4943-4953 (2019) 4 +9. Ding, X., Guo, Y., Ding, G., Han, J.: Acnet: Strengthening the kernel skeletons for powerful cnn via asymmetric convolution blocks. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1911-1920 (2019) 1 +10. Ding, X., Hao, T., Tan, J., Liu, J., Han, J., Guo, Y., Ding, G.: Resrep: Lossless cnn pruning via decoupling remembering and forgetting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4510-4520 (2021) 4 +11. Ding, X., Zhang, X., Han, J., Ding, G.: Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11963-11975 (2022) 1 +12. Ding, X., Zhang, X., Ma, N., Han, J., Ding, G., Sun, J.: Repvgg: Making vgg-style convnets great again. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13733-13742 (2021) 1 +13. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020) 1, 4 +14. Esser, S.K., McKinstry, J.L., Bablani, D., Appuswamy, R., Modha, D.S.: Learned step size quantization. In: International Conference on Learning Representations (2019) 4 +15. Feng, C., Zhong, Y., Jie, Z., Chu, X., Ren, H., Wei, X., Xie, W., Ma, L.: Promptdet: Towards open-vocabulary detection using uncurated images. In: European Conference on Computer Vision. pp. 701-717. Springer (2022) 3 + +16. Finkelstein, A., Almog, U., Grobman, M.: Fighting quantization bias with bias. arXiv preprint arXiv:1906.03193 (2019) 4 +17. Gholami, A., Kim, S., Dong, Z., Yao, Z., Mahoney, M.W., Keutzer, K.: A survey of quantization methods for efficient neural network inference. In: Low-Power Computer Vision, pp. 291-326. Chapman and Hall/CRC (2022) 4 +18. Han, S., Mao, H., Dally, W.J.: Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149 (2015) 4 +19. Hao, T., Chen, H., Guo, Y., Ding, G.: Consolidator: Mergeable adapter with grouped connections for visual adaptation. arXiv preprint arXiv:2305.00603 (2023) 3, 4 +20. Hao, T., Ding, X., Han, J., Guo, Y., Ding, G.: Manipulating identical filter redundancy for efficient pruning on deep and complicated cnn. IEEE Transactions on Neural Networks and Learning Systems (2023) 4 +21. Hao, T., Lyu, M., Chen, H., Zhao, S., Han, J., Ding, G.: Re-parameterized low-rank prompt: Generalize a vision-language model within $0.5\mathrm{k}$ parameters. arXiv preprint arXiv:2312.10813 (2023) 3 +22. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 1 +23. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019) 1, 3 +24. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021) 1, 3 +25. Hubara, I., Nahshan, Y., Hanani, Y., Banner, R., Soudry, D.: Accurate post training quantization with small calibration sets. In: International Conference on Machine Learning. pp. 4466-4475. PMLR (2021) 4 +26. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International Conference on Machine Learning. pp. 4904-4916. PMLR (2021) 1, 3 +27. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. arXiv preprint arXiv:2203.12119 (2022) 3, 4, 5 +28. Jiang, Z., Araki, J., Ding, H., Neubig, G.: How can we know when language models know? on the calibration of language models for question answering. Transactions of the Association for Computational Linguistics 9, 962-977 (2021) 3 +29. Jie, S., Wang, H., Deng, Z.H.: Revisiting the parameter efficiency of adapters from the perspective of precision redundancy. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17217-17226 (2023) 3 +30. Ju, C., Han, T., Zheng, K., Zhang, Y., Xie, W.: Prompting visual-language models for efficient video understanding. In: European Conference on Computer Vision. pp. 105-124. Springer (2022) 3 +31. Jung, S., Son, C., Lee, S., Son, J., Han, J.J., Kwak, Y., Hwang, S.J., Choi, C.: Learning to quantize deep networks by optimizing quantization intervals with task loss. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4350-4359 (2019) 4 + +32. Khattak, M.U., Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Maple: Multi-modal prompt learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19113-19122 (2023) 4, 5 +33. Khattak, M.U., Wasim, S.T., Naseer, M., Khan, S., Yang, M.H., Khan, F.S.: Self-regulating prompts: Foundational model adaptation without forgetting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15190-15200 (2023) 4 +34. Lee, D., Song, S., Suh, J., Choi, J., Lee, S., Kim, H.J.: Read-only prompt optimization for vision-language few-shot learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1401-1411 (2023) 4 +35. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 (2021) 3 +36. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=RriDjddCLN 3 +37. Li, H., Kadav, A., Durdanovic, I., Samet, H., Graf, H.P.: Pruning filters for efficient convnets. arXiv preprint arXiv:1608.08710 (2016) 4 +38. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 (2021) 3 +39. Li, Y., Xu, S., Zhang, B., Cao, X., Gao, P., Guo, G.: Q-vit: Accurate and fully quantized low-bit vision transformer. Advances in Neural Information Processing Systems 35, 34451-34463 (2022) 4 +40. Li, Z., Xiao, J., Yang, L., Gu, Q.: Repq-vit: Scale reparameterization for posttraining quantization of vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17227-17236 (2023) 4 +41. Lian, D., Zhou, D., Feng, J., Wang, X.: Scaling & shifting your features: A new baseline for efficient model tuning. In: Advances in Neural Information Processing Systems (NeurIPS) (2022) 3, 4 +42. Liu, P., Yuan, W., Fu, J., Jiang, Z., Hayashi, H., Neubig, G.: Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. arXiv preprint arXiv:2107.13586 (2021) 3 +43. Liu, X., Ji, K., Fu, Y., Du, Z., Yang, Z., Tang, J.: P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv preprint arXiv:2110.07602 (2021) 3 +44. Liu, X., Zheng, Y., Du, Z., Ding, M., Qian, Y., Yang, Z., Tang, J.: Gpt understands, too. arXiv preprint arXiv:2103.10385 (2021) 3 +45. Liu, Z., Wang, Y., Han, K., Zhang, W., Ma, S., Gao, W.: Post-training quantization for vision transformer. Advances in Neural Information Processing Systems 34, 28092-28103 (2021) 4 +46. Liu, Z., Li, J., Shen, Z., Huang, G., Yan, S., Zhang, C.: Learning efficient convolutional networks through network slimming. In: Proceedings of the IEEE international conference on computer vision. pp. 2736-2744 (2017) 4 +47. Lüddecke, T., Ecker, A.: Image segmentation using text and image prompts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7086-7096 (2022) 3 +48. Lyu, M., Hao, T., Xu, X., Chen, H., Han, J., Ding, G.: Learn from the learnt: Source-free active domain adaptation via contrastive sampling and visual persistence. In: European Conference on Computer Vision (ECCV). Springer (2024) 3 +49. Lyu, M., Yang, Y., Hong, H., Chen, H., Jin, X., He, Y., Xue, H., Han, J., Ding, G.: One-dimensional adapter to rule them all: Concepts diffusion models and erasing + +applications. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7559-7568 (2024) 1 +50. Lyu, M., Zhou, J., Chen, H., Huang, Y., Yu, D., Li, Y., Guo, Y., Guo, Y., Xiang, L., Ding, G.: Box-level active detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23766-23775 (2023) 1 +51. Maaz, M., Rasheed, H., Khan, S., Khan, F.S., Anwer, R.M., Yang, M.H.: Class-agnostic object detection with multi-modal transformer. In: The European Conference on Computer Vision. Springer (2022) 3 +52. Meller, E., Finkelstein, A., Almog, U., Grobman, M.: Same, same but different: Recovering neural network quantization error through weight factorization. In: International Conference on Machine Learning. pp. 4486-4495. PMLR (2019) 4 +53. Mu, N., Kirillov, A., Wagner, D., Xie, S.: Slip: Self-supervision meets language-image pre-training. In: European Conference on Computer Vision. pp. 529-544. Springer (2022) 1, 13, 14 +54. Nagel, M., Baalen, M.v., Blankevoort, T., Welling, M.: Data-free quantization through weight equalization and bias correction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1325-1334 (2019) 4 +55. Nagel, M., Fournarakis, M., Amjad, R.A., Bondarenko, Y., Van Baalen, M., Blankevoort, T.: A white paper on neural network quantization. arXiv preprint arXiv:2106.08295 (2021) 4 +56. Nagel, M., Fournarakis, M., Bondarenko, Y., Blankevoort, T.: Overcoming oscillations in quantization-aware training. In: International Conference on Machine Learning. pp. 16318-16330. PMLR (2022) 4 +57. Qian, R., Li, Y., Xu, Z., Yang, M.H., Belongie, S., Cui, Y.: Multimodal open-vocabulary video classification via pre-trained vision and language models. arXiv preprint arXiv:2207.07646 (2022) 3 +58. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021) 1, 3, 4, 10 +59. Rao, Y., Zhao, W., Chen, G., Tang, Y., Zhu, Z., Huang, G., Zhou, J., Lu, J.: Denseclip: Language-guided dense prediction with context-aware prompting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18082-18091 (2022) 1, 3 +60. Rasheed, H.A., Maaz, M., Khattak, M.U., Khan, S., Khan, F.: Bridging the gap between object and image-level representations for open-vocabulary detection. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=aKXBrjODHm 3 +61. Shin, T., Razeghi, Y., Logan IV, R.L., Wallace, E., Singh, S.: Autoprompt: Eliciting knowledge from language models with automatically generated prompts. In: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). pp. 4222-4235 (2020) 3 +62. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research 15(1), 1929-1958 (2014) 5 +63. Xian, Y., Schiele, B., Akata, Z.: Zero-shot learning-the good, the bad and the ugly. In: CVPR (2017) 11 +64. Xiong, Y., Chen, H., Hao, T., Lin, Z., Han, J., Zhang, Y., Wang, G., Bao, Y., Ding, G.: Pyra: Parallel yielding re-activation for training-inference efficient task adaptation. arXiv preprint arXiv:2403.09192 (2024) 3 + +65. Xiong, Y., Chen, H., Lin, Z., Zhao, S., Ding, G.: Confidence-based visual dispersal for few-shot unsupervised domain adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11621-11631 (2023) 1 +66. Xiong, Y., Chen, X., Ye, X., Chen, H., Lin, Z., Lian, H., Niu, J., Ding, G.: Temporal scaling law for large language models. arXiv preprint arXiv:2404.17785 (2024) 1 +67. Yao, H., Zhang, R., Xu, C.: Visual-language prompt tuning with knowledge-guided context optimization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6757-6767 (2023) 4 +68. Yao, L., Huang, R., Hou, L., Lu, G., Niu, M., Xu, H., Liang, X., Li, Z., Jiang, X., Xu, C.: FILIP: Fine-grained interactive language-image pre-training. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=cpDhcsEDC2 1, 3 +69. Yu, S., Chen, T., Shen, J., Yuan, H., Tan, J., Yang, S., Liu, J., Wang, Z.: Unified visual transformer compression. In: International Conference on Learning Representations (2022) 4 +70. Yuan, L., Chen, D., Chen, Y.L., Codella, N., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., et al.: Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432 (2021) 3 +71. Yuan, Z., Xue, C., Chen, Y., Wu, Q., Sun, G.: Ptg4vit: Post-training quantization for vision transformers with twin uniform quantization. In: European Conference on Computer Vision. pp. 191-207. Springer (2022) 4 +72. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary detr with conditional matching. arXiv preprint arXiv:2203.11876 (2022) 3 +73. Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., Beyer, L.: Lit: Zero-shot transfer with locked-image text tuning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18123-18133 (2022) 3 +74. Zhang, Y., Zhou, K., Liu, Z.: Neural prompt search. arXiv preprint arXiv:2206.04673 (2022) 4 +75. Zheng, K., Wu, W., Feng, R., Zhu, K., Liu, J., Zhao, D., Zha, Z.J., Chen, W., Shen, Y.: Regularized mask tuning: Uncovering hidden knowledge in pre-trained vision-language models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11663-11673 (2023) 4 +76. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022) 3, 4 +77. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022) 3, 4, 6, 7 \ No newline at end of file diff --git a/2024/Quantized Prompt for Efficient Generalization of Vision-Language Models/images.zip b/2024/Quantized Prompt for Efficient Generalization of Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7888d4a733234756e96db8b5ba60bd5a01bb6d17 --- /dev/null +++ b/2024/Quantized Prompt for Efficient Generalization of Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d865d090e09b38feb4d29a704860b5d736109d923bd6463466f8f5ac7f93e35e +size 343690 diff --git a/2024/Quantized Prompt for Efficient Generalization of Vision-Language Models/layout.json b/2024/Quantized Prompt for Efficient Generalization of Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..86dc5641031f90381b3c63fd459486b85daa1d5b --- /dev/null +++ b/2024/Quantized Prompt for Efficient Generalization of Vision-Language Models/layout.json @@ -0,0 +1,10901 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 111, + 479, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 111, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 136, + 111, + 479, + 148 + ], + "type": "text", + "content": "Quantized Prompt for Efficient Generalization of Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "spans": [ + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "text", + "content": "Tianxiang Hao" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Xiaohan Ding" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{3(\\boxtimes)}" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Juexiao Feng" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Yuhong Yang" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Hui Chen" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "text", + "content": ", and Guiguang Ding" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1,2(\\boxtimes)}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 188, + 201, + 425, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 201, + 425, + 235 + ], + "spans": [ + { + "bbox": [ + 188, + 201, + 425, + 235 + ], + "type": "text", + "content": "1 School of Software, Tsinghua University, Beijing, China \n2 BNrist, Beijing, China \n3 Bytedance" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 182, + 236, + 430, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 236, + 430, + 258 + ], + "spans": [ + { + "bbox": [ + 182, + 236, + 430, + 258 + ], + "type": "text", + "content": "{beyondtx,xiaohding,suise.con,jichenhui}@gmail.com \nfjx20@mails.tsinghua.edu.cn, dinggg@tsinghua.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 281, + 455, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 281, + 455, + 557 + ], + "spans": [ + { + "bbox": [ + 160, + 281, + 455, + 557 + ], + "type": "text", + "content": "Abstract. In the past few years, large-scale pre-trained vision-language models like CLIP have achieved tremendous success in various fields. Naturally, how to transfer the rich knowledge in such huge pre-trained models to downstream tasks and datasets becomes a hot topic. During downstream adaptation, the most challenging problems are overfitting and catastrophic forgetting, which can cause the model to overly focus on the current data and lose more crucial domain-general knowledge. Existing works use classic regularization techniques to solve the problems. As solutions become increasingly complex, the ever-growing storage and inference costs are also a significant problem that urgently needs to be addressed. While in this paper, we start from an observation that proper random noise can suppress overfitting and catastrophic forgetting. Then we regard quantization error as a kind of noise, and explore quantization for regularizing vision-language model, which is quite efficiency and effective. Furthermore, to improve the model's generalization capability while maintaining its specialization capacity at minimal cost, we deeply analyze the characteristics of the weight distribution in prompts, conclude several principles for quantization module design and follow such principles to create several competitive baselines. The proposed method is significantly efficient due to its inherent lightweight nature, making it possible to adapt on extremely resource-limited devices. Our method can be fruitfully integrated into many existing approaches like MaPLe, enhancing accuracy while reducing storage overhead, making it more powerful yet versatile. Extensive experiments on 11 datasets shows great superiority of our method sufficiently. Code is available at github" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 565, + 438, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 565, + 438, + 578 + ], + "spans": [ + { + "bbox": [ + 160, + 565, + 438, + 578 + ], + "type": "text", + "content": "Keywords: Quantization " + }, + { + "bbox": [ + 160, + 565, + 438, + 578 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 565, + 438, + 578 + ], + "type": "text", + "content": " Generalization " + }, + { + "bbox": [ + 160, + 565, + 438, + 578 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 565, + 438, + 578 + ], + "type": "text", + "content": " Vision-language model" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 594, + 231, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 231, + 607 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 231, + 607 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": "Recently, deep learning models and related technologies have seen rapid development [9, 11-13, 22-24, 49, 50, 65, 66]. Vision-language model (VLM) like CLIP [58] is one of the hottest research topics and leads to huge success. The excellent generalization ability is a crucial cornerstone of such achievement [26, 53, 58, 59, 68]." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 198, + 114, + 410, + 292 + ], + "blocks": [ + { + "bbox": [ + 198, + 114, + 410, + 292 + ], + "lines": [ + { + "bbox": [ + 198, + 114, + 410, + 292 + ], + "spans": [ + { + "bbox": [ + 198, + 114, + 410, + 292 + ], + "type": "image", + "image_path": "c3201a6a53517b9e1d48b08e43469486061b1f3fbbebf67fdcf5733f50b11212.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 293, + 482, + 350 + ], + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 350 + ], + "type": "text", + "content": "Fig. 1: Overall performance comparison with existing vision-language tuning methods. Our method outperforms all of the state-of-the-art competitors with significantly fewer storage space. Based on the proposed quantization algorithm, our method could be integrated into many of the existing methods and bring consistent improvements with excellent efficiency." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 371, + 482, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 371, + 482, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 482, + 420 + ], + "type": "text", + "content": "When people have access to downstream data, it is better to tune the pretrained VLM on the target dataset for higher accuracy. However, such full finetuning could easily cause the model to overfit the small downstream dataset and face catastrophic forgetting problem, leading to severe performance drop." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 422, + 482, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 543 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 543 + ], + "type": "text", + "content": "To solve the problem, in this paper we will start from rethinking the relationship between noise and generalization. We propose to think noise as a kind of regularization techniques, which may be helpful for alleviating overfitting and catastrophic forgetting problem. As in Sec. 3.2 and Fig. 2, we then find that directly adding some random Gaussian noise to the tunable prompts of vision-language models would result in a performance gain in several cases, which partly validates our conjecture about using noise as a form of regularization. In particular, excessive noise diminishes the model's adaptation capability, while insufficient noise fails to provide effective regularization. Only noise of moderate intensity is beneficial for the model's generalization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "However, Gaussian noise is absolutely random and hard to control, and thus it is difficult for us to take full advantage of such type of noise to benefit generalization. Instead, we point out that quantization error is also a form of noise, and therefore, it is also possible to leverage quantization error to enhance the model's generalization performance. With this idea in mind, we thoroughly analyzed the distribution pattern of prompt weights in Sec. 3.3 and derived several design principles for quantization algorithms based on the observed phenomena. In Sec. 4, following the principles we summarized, we successfully designed an efficient quantization-aware training algorithm, which largely enhances the model's generalization ability while quantizing." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "T. Hao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "Since our design is general, we could fruitfully integrate our quantization strategy into many existing methods and reach a higher accuracy with much smaller model size due to the lightweight nature of quantization. For example, we integrate our method onto an existing popular method MaPLe, and our QMaPLe earns " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "0.57\\%" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " accuracy improvement with only " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "0.25 \\times" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " size." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 176, + 396, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 176, + 396, + 186 + ], + "spans": [ + { + "bbox": [ + 146, + 176, + 396, + 186 + ], + "type": "text", + "content": "In conclusion, we summarize our contribution as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 197, + 479, + 350 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 197, + 479, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 197, + 479, + 232 + ], + "spans": [ + { + "bbox": [ + 138, + 197, + 479, + 232 + ], + "type": "text", + "content": "- We deeply analyze the effect of noise and rethink the relationship between noise and generalization in Sec. 3.2 for vision-language models. As a result, we confirm that moderate noise would promote the model generalization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 233, + 479, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 233, + 479, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 233, + 479, + 267 + ], + "type": "text", + "content": "- We are the first to propose to quantize prompts. By detailed observation and hard thinking, we conclude several principles about how to effectively design a quantization method for the prompts of vision-language models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 269, + 479, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 269, + 479, + 350 + ], + "spans": [ + { + "bbox": [ + 138, + 269, + 479, + 350 + ], + "type": "text", + "content": "- Following the principles we concluded, we build our method and successfully quantize the prompts as well as some other weights, and our method significantly outperforms existing ones as in Fig. 1. Extensive experiments show great power of our method. In base-to-new generalization, domain generalization, cross-dataset transfer and few-shot learning settings, we consistently reach competitive results, winning many state-of-the-art tuning methods with a much smaller size of model." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 372, + 242, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 372, + 242, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 372, + 242, + 384 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 399, + 283, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 399, + 283, + 411 + ], + "spans": [ + { + "bbox": [ + 132, + 399, + 283, + 411 + ], + "type": "text", + "content": "2.1 Vision-Language Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 421, + 479, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 421, + 479, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 421, + 479, + 552 + ], + "type": "text", + "content": "In recent times, large-scale vision-language models have demonstrated remarkable performance across various tasks. Seminal works such as [26, 58, 68, 70, 73]. Classic works have focused on learning multimodal representations through self-supervised methods using extensive sets of image-text pairs. Among these, CLIP [58] stands out as a milestone achievement, employing contrastive learning to align vision and language representations and achieving exceptional performance. A well-trained vision-language model is invaluable, offering substantial support to various fields. Successful applications of these robust models include few-shot recognition [76, 77], detection tasks [15, 51, 60, 72], and segmentation tasks [7, 36, 47, 59]. Furthermore, for video data, research efforts have emerged in video classification [57] and video understanding [30]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 571, + 323, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 571, + 323, + 583 + ], + "spans": [ + { + "bbox": [ + 132, + 571, + 323, + 583 + ], + "type": "text", + "content": "2.2 Parameter-Efficient Fine-Tuning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "text", + "content": "Recently, a series of works [19, 21, 23, 24, 27, 29, 38, 41, 44, 48, 64, 76, 77] have been proposed to help transfer the learnt knowledge, where one of the most popular field is parameter-efficient fine-tuning (PEFT). PEFT aims to transfer a pre-trained model to downstream tasks by a minimum number of parameters. Originating from natural language processing tasks, classic methods like adapter [23], prompt tuning [28, 35, 38, 42, 43, 61] and LoRA [24] follows a similar" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "type": "text", + "content": "Quantized Prompt for Efficient Generalization of Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "type": "text", + "content": "principle to add extra modules with a small number of parameters into the backbone model, freeze the original parameters and only tune and store the newly added parameters. Inspired by the success of PEFT in language field, researchers have extended such kind of approaches to adapt visual models in a similar fashion [6,19,27,41,74]. In the field of vision-language modeling, several explorations have been made as well. Bahng et al. [1] exclusively apply prompt tuning to the image encoder. CoOp [77] replaces the fixed template in CLIP [58] with tunable text prompts. CoCoOp [76] leverages image features to guide the optimization of tunable text prompts in CoOp. Other works [32, 34] optimize both image and text prompts simultaneously and establish additional connections between different modalities. To mitigate overfitting and catastrophic forgetting, various works [5,33,67,75] integrate regularization modules or losses into prompt tuning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 278, + 228, + 291 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 228, + 291 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 228, + 291 + ], + "type": "text", + "content": "2.3 Quantization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 300, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 491 + ], + "type": "text", + "content": "Quantization is one of the most effective compression methods [4,8,10,20,37, 46,69] for deep learning models. Generally, parameters such as weights and activations are typically stored as 32-bit floating-point numbers, which consume a significant amount of memory and require intensive computation during inference. Quantization [17, 18, 55] involves representing these parameters with reduced precision, such as 8-bit integers or even lower bit-widths. By doing so, quantization can significantly reduce the memory footprint and computational complexity of the model without significantly sacrificing accuracy. Quantization methods can be divided into two groups, Post-Training Quantization (PTQ) [2,16,25,40,45,52,54,71] that consumes few resources but suffers higher accuracy loss, and Quantization-Aware Training (QAT) [3,14,31,39,56] that relies on plenty of resources for training and shows better accuracy. Existing works aim to minimize quantization error to improve accuracy, while our work demonstrates that both excessive and insufficient errors are detrimental to model generalization. To achieve optimal generalization performance, a moderate error is required." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 510, + 442, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 442, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 442, + 525 + ], + "type": "text", + "content": "3 Exploring Quantization in Model Generalization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 536, + 453, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 453, + 550 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 453, + 550 + ], + "type": "text", + "content": "3.1 Preliminaries: Prompt Tuning of Vision-Language Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 557, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 640 + ], + "type": "text", + "content": "CLIP comprises a text encoder " + }, + { + "bbox": [ + 130, + 557, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 130, + 557, + 482, + 640 + ], + "type": "text", + "content": " and an image encoder " + }, + { + "bbox": [ + 130, + 557, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 557, + 482, + 640 + ], + "type": "text", + "content": ". Typically, " + }, + { + "bbox": [ + 130, + 557, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 130, + 557, + 482, + 640 + ], + "type": "text", + "content": " is implemented as a language transformer, whereas " + }, + { + "bbox": [ + 130, + 557, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 557, + 482, + 640 + ], + "type": "text", + "content": " may be realized using either a convolutional neural network or a vision transformer. In this study, following the methodologies outlined by [76, 77], we employ a ViT-B/16 model [13] as the image encoder " + }, + { + "bbox": [ + 130, + 557, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 557, + 482, + 640 + ], + "type": "text", + "content": ", except where otherwise specified. The subsequent sections will provide a brief overview of the methods used to prompt CLIP for prediction tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 667 + ], + "type": "text", + "content": "Text Encoder Consider a text encoder composed of " + }, + { + "bbox": [ + 130, + 641, + 482, + 667 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 641, + 482, + 667 + ], + "type": "text", + "content": " layers. For the " + }, + { + "bbox": [ + 130, + 641, + 482, + 667 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 641, + 482, + 667 + ], + "type": "text", + "content": "-th layer, denoted as " + }, + { + "bbox": [ + 130, + 641, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_k" + }, + { + "bbox": [ + 130, + 641, + 482, + 667 + ], + "type": "text", + "content": ", the inputs consist of a sequence of prompt tokens " + }, + { + "bbox": [ + 130, + 641, + 482, + 667 + ], + "type": "inline_equation", + "content": "P_{k - 1}^{l}" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "T. Hao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "text", + "content": "and a " + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "inline_equation", + "content": "[CLS]" + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "text", + "content": " token " + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "inline_equation", + "content": "c_{k-1}^{l}" + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "text", + "content": ", while the outputs are represented by " + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "inline_equation", + "content": "P_{k}^{l}" + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "inline_equation", + "content": "c_{k}^{l}" + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "text", + "content": ". The initial inputs, " + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "inline_equation", + "content": "P_{0}^{l}" + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "inline_equation", + "content": "c_{0}^{l}" + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "text", + "content": ", correspond to the word embeddings of the prompts combined with the label, such as \"A photo of a " + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "inline_equation", + "content": "[CLS]''" + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "text", + "content": " or alternatively, some randomly initialized vectors. Formally, we denote " + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "inline_equation", + "content": "P_{k}^{l} \\in \\mathbb{R}^{n^{l} \\times d^{l}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "inline_equation", + "content": "c_{k}^{l} \\in \\mathbb{R}^{d^{l}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "inline_equation", + "content": "n^{l}" + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "text", + "content": " signifies the length of the text prompts and " + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "inline_equation", + "content": "d^{l}" + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "text", + "content": " represents the dimension of the word embeddings. For each layer, " + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "inline_equation", + "content": "1 \\leq k \\leq M" + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "text", + "content": ", the relationship is given by " + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "inline_equation", + "content": "[P_{k}^{l}, c_{k}^{l}] = \\mathcal{L}_{k}([P_{k-1}^{l}, c_{k-1}^{l}])" + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "text", + "content": ". The output feature of the text encoder " + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "inline_equation", + "content": "f^{l} \\in \\mathbb{R}^{d^{v}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "inline_equation", + "content": "d^{v}" + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "text", + "content": " is the dimension of the visual feature space, is calculated by projecting the " + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "inline_equation", + "content": "[CLS]" + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "text", + "content": " token of its last layer to the visual latent space through a linear transformation, i.e. " + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "inline_equation", + "content": "f^{l} = \\mathrm{Proj}(c_{M}^{l})" + }, + { + "bbox": [ + 130, + 115, + 482, + 240 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "text", + "content": "Image Encoder Suppose there are " + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "text", + "content": " layers in the image encoder. For " + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "text", + "content": "-th layer " + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_k" + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "text", + "content": ", the inputs are a series of image tokens " + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "inline_equation", + "content": "I_{k-1}" + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "text", + "content": ", a classification token " + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "inline_equation", + "content": "c_{k-1}^v" + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "text", + "content": " and prompt tokens " + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "inline_equation", + "content": "P_{k-1}^v" + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "text", + "content": ", and the outputs are " + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "inline_equation", + "content": "I_k" + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "inline_equation", + "content": "c_k^v" + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "inline_equation", + "content": "P_{k}^{v}" + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "text", + "content": ". The inputs of the first layer " + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "inline_equation", + "content": "c_0^v" + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "text", + "content": " are the patch embeddings of the input image and the pre-trained class token. " + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "inline_equation", + "content": "P_0^v" + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "text", + "content": " is randomly initialized. Formally, " + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "inline_equation", + "content": "I_k \\in \\mathbb{R}^{p \\times d^v}" + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "inline_equation", + "content": "c_k^v \\in \\mathbb{R}^{d^v}" + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "inline_equation", + "content": "P_k^v \\in \\mathbb{R}^{n^v \\times d^v}" + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "text", + "content": " denotes the number of image patches and " + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "inline_equation", + "content": "d^v" + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "text", + "content": " denotes the dimension of visual embedding. " + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "inline_equation", + "content": "\\forall 1 \\leq k \\leq N" + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "inline_equation", + "content": "[P_k^v, c_k^v, I_k] = \\mathcal{V}_k([P_{k-1}^v, c_{k-1}^v, I_{k-1}])" + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "text", + "content": ". The output feature of the image encoder is " + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "inline_equation", + "content": "f^v = c_N^v" + }, + { + "bbox": [ + 130, + 240, + 482, + 335 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 335, + 482, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 397 + ], + "type": "text", + "content": "Prediction For image classification, suppose there are " + }, + { + "bbox": [ + 130, + 335, + 482, + 397 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 335, + 482, + 397 + ], + "type": "text", + "content": " classes, and " + }, + { + "bbox": [ + 130, + 335, + 482, + 397 + ], + "type": "inline_equation", + "content": "\\{f_c^l\\}_{c = 1}^C" + }, + { + "bbox": [ + 130, + 335, + 482, + 397 + ], + "type": "text", + "content": " are the text features. Label " + }, + { + "bbox": [ + 130, + 335, + 482, + 397 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 335, + 482, + 397 + ], + "type": "text", + "content": "'s probability is " + }, + { + "bbox": [ + 130, + 335, + 482, + 397 + ], + "type": "inline_equation", + "content": "p(y|f^v) = \\frac{\\exp(\\sin(f^v,f_l^i) / \\tau)}{\\sum_{c = 1}^{C}\\exp(\\sin(f^v,f_c^l) / \\tau)}" + }, + { + "bbox": [ + 130, + 335, + 482, + 397 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 335, + 482, + 397 + ], + "type": "inline_equation", + "content": "\\sin (\\cdot ,\\cdot)" + }, + { + "bbox": [ + 130, + 335, + 482, + 397 + ], + "type": "text", + "content": " denotes cosine similarity function and " + }, + { + "bbox": [ + 130, + 335, + 482, + 397 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 335, + 482, + 397 + ], + "type": "text", + "content": " is temperature. The final prediction is " + }, + { + "bbox": [ + 130, + 335, + 482, + 397 + ], + "type": "inline_equation", + "content": "\\hat{z} = \\arg \\max_{1\\leq y < C}(p(y|f^v))" + }, + { + "bbox": [ + 130, + 335, + 482, + 397 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 398, + 485, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 485, + 482 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 485, + 482 + ], + "type": "text", + "content": "We have introduced shallow prompts. There are also different types of prompts. Several works [27, 32] use them for improve performance. They directly add and tune the prompt in each layer in the feature encoder, instead of inheriting the output prompt calculated by the last encoder. Now we have " + }, + { + "bbox": [ + 130, + 398, + 485, + 482 + ], + "type": "inline_equation", + "content": "[\\_ ,c_k^l ] = \\mathcal{L}_k([P_{k - 1}^l,c_{k - 1}^l])" + }, + { + "bbox": [ + 130, + 398, + 485, + 482 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 398, + 485, + 482 + ], + "type": "inline_equation", + "content": "[\\_ ,c_k^v,I_k] = \\mathcal{V}_k([P_{k - 1}^v,c_{k - 1}^v,I_{k - 1}])" + }, + { + "bbox": [ + 130, + 398, + 485, + 482 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 130, + 398, + 485, + 482 + ], + "type": "inline_equation", + "content": "P^l" + }, + { + "bbox": [ + 130, + 398, + 485, + 482 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 398, + 485, + 482 + ], + "type": "inline_equation", + "content": "P^v" + }, + { + "bbox": [ + 130, + 398, + 485, + 482 + ], + "type": "text", + "content": " are independent tunable parameters. They are no longer determined by the previous layer." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 501, + 478, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 478, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 478, + 514 + ], + "type": "text", + "content": "3.2 Rethinking the Relationship between Noise and Generalization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Some existing work has explored how to use the addition of noise to suppress model overfitting, e.g., techniques like Dropout [62] that randomly drops connections, and random jitter that directly introduce perturbations in the input data. However, the impact of directly adding noise to model weights has been less explored, especially in the context of prompt structures, which have become popular only in recent years, in Transformer architectures. The variations of the specialization capability represented by the test accuracy on base, seen classes and the generalization capability represented by the test accuracy on new, unseen classes. The curves of different colors represent the data under the influence of random Gaussian noise of different intensities, e.g. \"Noise_0.01\" adds random noise with a distribution of " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,0.01^2)" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " to the prompt. \"Noise_0\" denotes the baseline prompt tuning. As training progresses, the generalization capability" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "type": "text", + "content": "Quantized Prompt for Efficient Generalization of Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 126, + 289, + 230 + ], + "blocks": [ + { + "bbox": [ + 153, + 126, + 289, + 230 + ], + "lines": [ + { + "bbox": [ + 153, + 126, + 289, + 230 + ], + "spans": [ + { + "bbox": [ + 153, + 126, + 289, + 230 + ], + "type": "image", + "image_path": "795d95e07398067757fe3533b9326a893b3088234d8b65bd463502db3a4d4768.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 232, + 482, + 384 + ], + "lines": [ + { + "bbox": [ + 130, + 232, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 482, + 384 + ], + "type": "text", + "content": "Fig. 2: The variations of the specialization capability represented by the test accuracy on base, seen classes and the generalization capability represented by the test accuracy on new, unseen classes on average of the ten datasets in base-to-new generalization setting in Sec. 5 except for time-consuming ImageNet. The curves of different colors represent the data under the influence of random Gaussian noise of different intensities, e.g. \"Noise_0.01\" adds random noise with a distribution of " + }, + { + "bbox": [ + 130, + 232, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,0.01^2)" + }, + { + "bbox": [ + 130, + 232, + 482, + 384 + ], + "type": "text", + "content": " to the prompt. \"Noise_0\" denotes the baseline prompt tuning. As training progresses, the generalization capability of baseline prompt tuning continuously decreases while the specialization capability improves. Therefore, we expect that adding noise can achieve a better balance between generalization and specialization. However, excessive noise, e.g. 0.1, greatly diminishes the model's specialization capability, while insufficient noise, e.g. 0.001, fails to provide effective regularization. Only noise of moderate intensity outperforms baseline in specialization-generalization trade-off, effectively enhancing the unseen class accuracy without significantly compromising seen class accuracy." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 308, + 126, + 449, + 230 + ], + "blocks": [ + { + "bbox": [ + 308, + 126, + 449, + 230 + ], + "lines": [ + { + "bbox": [ + 308, + 126, + 449, + 230 + ], + "spans": [ + { + "bbox": [ + 308, + 126, + 449, + 230 + ], + "type": "image", + "image_path": "a10f97832116346084f99cc7315ae08e98e1238496ca61e2de859f49045f0004.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 400, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 483 + ], + "type": "text", + "content": "of baseline continuously decreases while the specialization capability improves. Therefore, we expect that adding noise can achieve a better balance between generalization and specialization. However, excessive noise, e.g. 0.1, greatly diminishes the model's specialization capability, while insufficient noise, e.g. 0.001, fails to provide effective regularization. Only noise of moderate intensity outperforms baseline in specialization-generalization trade-off, effectively enhancing the unseen class accuracy without significantly compromising seen class accuracy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 484, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 482, + 568 + ], + "type": "text", + "content": "Quantization, the technique of mapping parameter values from high precision to low precision, can also be viewed as introducing some form of noise into the parameters, and thus can possibly improve the generalizability as the noise did in Fig. 2. Compared to Gaussian noise, quantization error is more controllable. Better still, quantization also has another major advantage: it can significantly reduce the storage required for parameters. Therefore, using quantization algorithms to generalize vision-language models is a very promising direction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 586, + 437, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 437, + 598 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 437, + 598 + ], + "type": "text", + "content": "3.3 Characteristics of Prompts in Vision-Language Models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 628 + ], + "type": "text", + "content": "In this subsection, we will demonstrate the characteristics of prompts in the vision-language model to facilitate targeted design of quantization schemes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "content": "We start from analyzing the training procedure of CoOp [77], which freezes all the parameters in backbone and tunes the added prompts only. The histogram about the frequency of weights of prompts for different training epochs is shown" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "T. Hao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 116, + 197, + 165 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 197, + 165 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 197, + 165 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 197, + 165 + ], + "type": "image", + "image_path": "59081e3f35601fe0e14fbb29ce3f8e076fc28a0420491ae46b6b8133466ef4e9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 203, + 116, + 265, + 165 + ], + "blocks": [ + { + "bbox": [ + 203, + 116, + 265, + 165 + ], + "lines": [ + { + "bbox": [ + 203, + 116, + 265, + 165 + ], + "spans": [ + { + "bbox": [ + 203, + 116, + 265, + 165 + ], + "type": "image", + "image_path": "ee31c05f39048038f87beb9c76025c27ad8e3253677f963621a4f4cd9f886232.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 273, + 116, + 335, + 165 + ], + "blocks": [ + { + "bbox": [ + 273, + 116, + 335, + 165 + ], + "lines": [ + { + "bbox": [ + 273, + 116, + 335, + 165 + ], + "spans": [ + { + "bbox": [ + 273, + 116, + 335, + 165 + ], + "type": "image", + "image_path": "0cf7047e2eaa60eec58a465de3d9f36f6d7b6dcf67c9c92396088a36aa2db1e2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 342, + 116, + 404, + 165 + ], + "blocks": [ + { + "bbox": [ + 342, + 116, + 404, + 165 + ], + "lines": [ + { + "bbox": [ + 342, + 116, + 404, + 165 + ], + "spans": [ + { + "bbox": [ + 342, + 116, + 404, + 165 + ], + "type": "image", + "image_path": "4de6c6cf6f2747070b02fb9ca1b9f14a279485613c55e3d7986e4ac1522bf200.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 412, + 116, + 474, + 165 + ], + "blocks": [ + { + "bbox": [ + 412, + 116, + 474, + 165 + ], + "lines": [ + { + "bbox": [ + 412, + 116, + 474, + 165 + ], + "spans": [ + { + "bbox": [ + 412, + 116, + 474, + 165 + ], + "type": "image", + "image_path": "4bbc9d7a2ee54cd2750e0c5bce3f4d178bb00f64776556f4e55093f79571ce50.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 133, + 167, + 197, + 215 + ], + "blocks": [ + { + "bbox": [ + 133, + 167, + 197, + 215 + ], + "lines": [ + { + "bbox": [ + 133, + 167, + 197, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 167, + 197, + 215 + ], + "type": "image", + "image_path": "1df42b99b9e7efb8e5bcd7c80663a1d48ca5312d5463ab5b671cabffe9ab2275.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 204, + 167, + 265, + 215 + ], + "blocks": [ + { + "bbox": [ + 204, + 167, + 265, + 215 + ], + "lines": [ + { + "bbox": [ + 204, + 167, + 265, + 215 + ], + "spans": [ + { + "bbox": [ + 204, + 167, + 265, + 215 + ], + "type": "image", + "image_path": "454cd5847e87043494fec38d3880f7a9ee648d62b74d6b8a181602bc574b86d9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 273, + 167, + 335, + 216 + ], + "blocks": [ + { + "bbox": [ + 273, + 167, + 335, + 216 + ], + "lines": [ + { + "bbox": [ + 273, + 167, + 335, + 216 + ], + "spans": [ + { + "bbox": [ + 273, + 167, + 335, + 216 + ], + "type": "image", + "image_path": "66bf6a9ede6a2b896b2ece0985627f28319370c1269d79491dafcd28c98b3c92.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 216, + 481, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 216, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 481, + 304 + ], + "type": "text", + "content": "Fig. 3: The histogram about the frequency of weights of prompts during training of CoOp [77] on eurosat dataset. We find the shape of the prompt's weight distribution remains largely unchanged throughout the entire training, but the variance of the weight distribution increases rapidly at the beginning of training. Additionally, we also notice that there are almost no outliers in the prompt's weights throughout the entire training process, and apart from the unstable initial training phase, the changes in weights between adjacent phases are not significant, indicating a very gentle overall updating trend." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 343, + 167, + 404, + 216 + ], + "blocks": [ + { + "bbox": [ + 343, + 167, + 404, + 216 + ], + "lines": [ + { + "bbox": [ + 343, + 167, + 404, + 216 + ], + "spans": [ + { + "bbox": [ + 343, + 167, + 404, + 216 + ], + "type": "image", + "image_path": "583972c5c62d0a30dbcda92ce9ce68fd94e26db29ed99bed01c2483148f9cf8e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 412, + 167, + 474, + 216 + ], + "blocks": [ + { + "bbox": [ + 412, + 167, + 474, + 216 + ], + "lines": [ + { + "bbox": [ + 412, + 167, + 474, + 216 + ], + "spans": [ + { + "bbox": [ + 412, + 167, + 474, + 216 + ], + "type": "image", + "image_path": "90e6e9dc8d9477c7e68bad3002e0f806846119b45a69f4a7ccfbfb5defca68f7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 319, + 480, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 480, + 403 + ], + "type": "text", + "content": "in Fig. 3. We observe that the shape of the prompt's weight distribution remains mostly consistent throughout the entire training process. However, the variance of the weight distribution increases rapidly at the initial stages of training. Refer to Fig. 6 for more details. Moreover, we notice minimal presence of outliers in the prompt's weights throughout the training process. Apart from the unstable initial training phase, the changes in weights between consecutive phases are not substantial, indicating a very gentle overall updating trend." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 403, + 481, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 481, + 427 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 481, + 427 + ], + "type": "text", + "content": "Taking into account the observations from Secs. 3.2 and 3.3, we can draw the following conclusions:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 437, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 437, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 437, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 437, + 482, + 510 + ], + "type": "text", + "content": "- Because the prompt weights are not sensitive to noise, and moderate noise in the current scenario not only does not degrade the model's performance but actually enhances its generalization ability, we can possibly adopt some quantization strategies that are considered very aggressive and highly likely to significantly degrade performance, such as 1-bit quantization, without causing destructive effects on the model's generalization ability." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 510, + 482, + 666 + ], + "type": "text", + "content": "- Given that the parameters targeted for quantization in the current scenario constitute only a tiny fraction of the total model parameters, we speculate that QAT may have more advantages over PTQ, as any additional training costs incurred by QAT which is considered a major drawback in traditional situations would naturally be kept at a low level due to the low proportion of parameters targeted for quantization. Due to the observation in Fig. 2 that both excessive and insufficient noise are detrimental to generalization, in designing the QAT algorithm, we no longer need to always minimize quantization error as the objective, as in classic algorithms. Instead, we aim to maintain the quantization error at a moderate level. Furthermore, as seen in Fig. 3, the changes in prompt weight distribution during most of the training time are gentle, indicating that we can appropriately continue to use past-time algorithm state without the need for real-time updates." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 146, + 91, + 448, + 102 + ], + "type": "text", + "content": "Quantized Prompt for Efficient Generalization of Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 166, + 114, + 440, + 262 + ], + "blocks": [ + { + "bbox": [ + 166, + 114, + 440, + 262 + ], + "lines": [ + { + "bbox": [ + 166, + 114, + 440, + 262 + ], + "spans": [ + { + "bbox": [ + 166, + 114, + 440, + 262 + ], + "type": "image", + "image_path": "85258a1dff8fc578d8697103b18340416c413713e8e0e96b55961cd5383fae1a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 263, + 480, + 285 + ], + "lines": [ + { + "bbox": [ + 130, + 263, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 263, + 480, + 285 + ], + "type": "text", + "content": "Fig. 4: Overview of Quantized Prompt. We set " + }, + { + "bbox": [ + 130, + 263, + 480, + 285 + ], + "type": "inline_equation", + "content": "b = 2" + }, + { + "bbox": [ + 130, + 263, + 480, + 285 + ], + "type": "text", + "content": " for a clear explanation. Normalization and denormalization processes are not shown here." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 304, + 480, + 472 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 137, + 304, + 479, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 304, + 479, + 376 + ], + "spans": [ + { + "bbox": [ + 137, + 304, + 479, + 376 + ], + "type": "text", + "content": "- Throughout the training process, the shape of the prompt weight distribution remains roughly unchanged, with the main variation coming from the distribution's variance. Therefore, it may be beneficial to attempt to eliminate the translation and scaling transformations of the distribution by normalizing the distribution before quantization and denormalizing it after quantization, which might help improve quantization accuracy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 376, + 480, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 480, + 472 + ], + "type": "text", + "content": "- Since there are almost no outliers throughout the entire training process, some classic algorithms are revitalized. Specifically, clustering algorithms like K-Means are good choices, because one of the major drawbacks of them in traditional usage is that they are greatly affected by outliers. In addition, as we have analyzed before, there is not a high demand for dynamically minimizing quantization error in vision-language generalization, opting for a more efficient non-parametric clustering approach seems a better choice compared with some complex parametric approaches." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 498, + 281, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 281, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 281, + 512 + ], + "type": "text", + "content": "4 Prompt Quantization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 529, + 329, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 529, + 329, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 529, + 329, + 542 + ], + "type": "text", + "content": "4.1 Preliminaries: Quantization Basis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "text", + "content": "Suppose there are " + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "text", + "content": " parameters in total to be quantized, which are denoted by " + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "inline_equation", + "content": "W \\in \\mathbb{R}^m" + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "text", + "content": ". Each " + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "inline_equation", + "content": "w_i" + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "text", + "content": " here is a high-precision float-type number. A " + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "text", + "content": "-bit model quantization algorithm divides the value range of parameters into " + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "inline_equation", + "content": "2^b" + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "text", + "content": " intervals " + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{U}_i, 1 \\leq i \\leq 2^b\\}" + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "text", + "content": ", aiming to find a mapping " + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "text", + "content": " from intervals to points, mapping all values within an interval " + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_i" + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "text", + "content": " to the same quantized value " + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "inline_equation", + "content": "q_i" + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "inline_equation", + "content": "Q(x) = q_i, \\forall x \\in \\mathcal{U}_i" + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "text", + "content": ". We define the quantization error " + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "inline_equation", + "content": "E = \\sum_{i=1}^{N} ||Q(W_i) - W_i||^2" + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "text", + "content": ". Such error serves as the objective of K-Means algorithm as well, so it can be ensured that each time clustering is redone using the K-Means algorithm, the quantization error will most likely decrease." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "T. Hao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 318, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 318, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 318, + 128 + ], + "type": "text", + "content": "4.2 Overview of Quantized Prompt" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 290 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 290 + ], + "type": "text", + "content": "Based on the observation and analysis in Sec. 3.3, we build a simple yet quite proper quantization algorithm for generalizing vision-language models. As stated before, we choose a widely-used clustering method K-Means, to construct the mapping " + }, + { + "bbox": [ + 130, + 137, + 482, + 290 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 137, + 482, + 290 + ], + "type": "text", + "content": ". Initially, K-Means algorithm is run to fit the pre-trained prompt weights, and record the codebook, i.e. cluster centers. Normalization and denormalization are conducted beyond quantization to alleviate the influence of varying variance. Formally, given the original prompt " + }, + { + "bbox": [ + 130, + 137, + 482, + 290 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 137, + 482, + 290 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 137, + 482, + 290 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 137, + 482, + 290 + ], + "type": "text", + "content": " parameters, we first calculate " + }, + { + "bbox": [ + 130, + 137, + 482, + 290 + ], + "type": "inline_equation", + "content": "\\mu = \\frac{1}{N}\\sum_{i=1}^{N}W_i" + }, + { + "bbox": [ + 130, + 137, + 482, + 290 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 137, + 482, + 290 + ], + "type": "inline_equation", + "content": "\\sigma = \\sqrt{\\frac{\\sum_{1\\leq i\\leq N}(W_i - \\mu)^2}{N}}" + }, + { + "bbox": [ + 130, + 137, + 482, + 290 + ], + "type": "text", + "content": ", then normalize " + }, + { + "bbox": [ + 130, + 137, + 482, + 290 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 137, + 482, + 290 + ], + "type": "text", + "content": " and get " + }, + { + "bbox": [ + 130, + 137, + 482, + 290 + ], + "type": "inline_equation", + "content": "\\hat{W} = \\frac{W - \\mu}{\\sigma}" + }, + { + "bbox": [ + 130, + 137, + 482, + 290 + ], + "type": "text", + "content": ". K-Means quantization is applied on " + }, + { + "bbox": [ + 130, + 137, + 482, + 290 + ], + "type": "inline_equation", + "content": "\\hat{W}" + }, + { + "bbox": [ + 130, + 137, + 482, + 290 + ], + "type": "text", + "content": ", and the quantized prompt " + }, + { + "bbox": [ + 130, + 137, + 482, + 290 + ], + "type": "inline_equation", + "content": "W_q = \\sigma Q(\\hat{W}) + \\mu" + }, + { + "bbox": [ + 130, + 137, + 482, + 290 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 130, + 137, + 482, + 290 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 137, + 482, + 290 + ], + "type": "text", + "content": " is not differentiable, we adopt a common practice to directly propagate the gradient across the quantization function by Straight-Through Estimator (STE), i.e. " + }, + { + "bbox": [ + 130, + 137, + 482, + 290 + ], + "type": "inline_equation", + "content": "\\frac{\\partial Q(x)}{\\partial x} = x" + }, + { + "bbox": [ + 130, + 137, + 482, + 290 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 291, + 482, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 291, + 482, + 398 + ], + "spans": [ + { + "bbox": [ + 130, + 291, + 482, + 398 + ], + "type": "text", + "content": "The overall framework is shown in Fig. 4. In training, we keep tuning the weight by the fake gradient propagated from quantization operation. For storage, we first convert the fp16 parameters in prompt to " + }, + { + "bbox": [ + 130, + 291, + 482, + 398 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 291, + 482, + 398 + ], + "type": "text", + "content": "-bit indexes, which could be used to search for the corresponding cluster center in the codebook. In Fig. 4, we set " + }, + { + "bbox": [ + 130, + 291, + 482, + 398 + ], + "type": "inline_equation", + "content": "b = 2" + }, + { + "bbox": [ + 130, + 291, + 482, + 398 + ], + "type": "text", + "content": " for a clear explanation. Compared with baseline method, our quantized method could save a lot more storage space. Specifically, for storage, an ordinary method needs " + }, + { + "bbox": [ + 130, + 291, + 482, + 398 + ], + "type": "inline_equation", + "content": "16N" + }, + { + "bbox": [ + 130, + 291, + 482, + 398 + ], + "type": "text", + "content": " bits, while ours only needs " + }, + { + "bbox": [ + 130, + 291, + 482, + 398 + ], + "type": "inline_equation", + "content": "bN + 2^b \\times 16" + }, + { + "bbox": [ + 130, + 291, + 482, + 398 + ], + "type": "text", + "content": ". In experiments, we usually set " + }, + { + "bbox": [ + 130, + 291, + 482, + 398 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 291, + 482, + 398 + ], + "type": "text", + "content": " to 1, 2 or 4, which is far more smaller than " + }, + { + "bbox": [ + 130, + 291, + 482, + 398 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 291, + 482, + 398 + ], + "type": "text", + "content": ". For example, with " + }, + { + "bbox": [ + 130, + 291, + 482, + 398 + ], + "type": "inline_equation", + "content": "b = 1" + }, + { + "bbox": [ + 130, + 291, + 482, + 398 + ], + "type": "text", + "content": ", the storage space of our method is roughly " + }, + { + "bbox": [ + 130, + 291, + 482, + 398 + ], + "type": "inline_equation", + "content": "16 \\times" + }, + { + "bbox": [ + 130, + 291, + 482, + 398 + ], + "type": "text", + "content": " smaller than baseline." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 399, + 481, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 481, + 422 + ], + "type": "text", + "content": "Note that the codebook here is updated by a rule called \"constrained adaptive clustering\". We will give a detailed description of it in the following subsection." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 441, + 413, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 441, + 413, + 453 + ], + "spans": [ + { + "bbox": [ + 131, + 441, + 413, + 453 + ], + "type": "text", + "content": "4.3 Constrained Adaptive Clustering for Quantization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 462, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 593 + ], + "type": "text", + "content": "In Fig. 4, re-clustering the prompt every iteration like classic QAT algorithms did to keep the codebook updated at all times is not a good choice. There are three reasons: 1. K-Means algorithm is not quite efficient, and thus if we run it every iteration then the training efficiency would decrease; 2. Only moderate noise promotes generalization, so keep updating the clusters to minimize the quantization error may be not helpful but harmful; 3. From Fig. 3, we can observe that the weight changes are very gentle for most of the training time, and the weight distributions of adjacent stages are highly similar. Therefore, even if we update the internal clustering state of K-Means with these already very similar data at each iteration, it's difficult to generate a better clustering solution, which is just futile effort." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "We propose constrained adaptive clustering to instruct the update of parameters inside K-Means which would not updated by gradient at all. Intuitively, first, we do not want too often update, so we set a minimum cluster update interval " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": ". Second, to avoid K-Means meaninglessly handling similar weights with current cached state, i.e. the weights of prompt that triggers the last re-clustering, we plan to only do re-clustering when current weight distribution is far different" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "type": "text", + "content": "Quantized Prompt for Efficient Generalization of Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "from the cached weight distribution. To reach the goal, Kullback-Leibler divergence is a good metric. However, if we want to compute the Kullback-Leibler divergence between two sets of discrete random variables, they must be defined on the same event space. However, that's not the case. So we first project each of them into the same event space spanning by the K-Means clustering algorithm. Specifically, given a current weight " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "W^{cur}" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": ", cached weight " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "W^{old}" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": ", quantization function " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": ", cluster centers " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{c_i, 1 \\leq i \\leq 2^b\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": ", we will compute the " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "W_{index}^{cur}" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "W_{index}^{old}" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": " as in Fig. 4. Then, we can obtain the probability distributions of the indices " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "p^{cur}" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "p^{old}" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": ", respectively, implied by " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "W_{index}^{cur}" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "W_{index}^{old}" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": ", thus successfully transforming two originally unrelated discrete random variables into the same event space. The Kullback-Leibler divergence could be computed as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 220, + 256, + 481, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 256, + 481, + 291 + ], + "spans": [ + { + "bbox": [ + 220, + 256, + 481, + 291 + ], + "type": "interline_equation", + "content": "K L \\left(p ^ {c u r} | | p ^ {o l d}\\right) = \\sum_ {i = 1} ^ {2 ^ {b}} p ^ {c u r} (i) \\log \\frac {p ^ {c u r} (i)}{p ^ {o l d} (i)} \\tag {1}", + "image_path": "8fc94ac7b7896bce33d4c73f31c145d0bdf060673dabe80c40c9de5ef179f5c6.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 295, + 481, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 295, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 131, + 295, + 481, + 319 + ], + "type": "text", + "content": "The update would continue only if Kullback-Leibler divergence exceeds a certain threshold " + }, + { + "bbox": [ + 131, + 295, + 481, + 319 + ], + "type": "inline_equation", + "content": "T_{KL}" + }, + { + "bbox": [ + 131, + 295, + 481, + 319 + ], + "type": "text", + "content": " which shows the distribution difference is significant." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 337, + 230, + 351 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 337, + 230, + 351 + ], + "spans": [ + { + "bbox": [ + 132, + 337, + 230, + 351 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 361, + 481, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 361, + 481, + 387 + ], + "spans": [ + { + "bbox": [ + 131, + 361, + 481, + 387 + ], + "type": "text", + "content": "We evaluate the method and make comparisons with the latest state-of-the-art methods in terms of the following settings across a wide range:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 135, + 393, + 480, + 514 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 135, + 393, + 480, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 393, + 480, + 416 + ], + "spans": [ + { + "bbox": [ + 135, + 393, + 480, + 416 + ], + "type": "text", + "content": "1. Base-to-new generalization, which models are trained with base classes and evaluated on both base and new classes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 135, + 418, + 480, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 418, + 480, + 440 + ], + "spans": [ + { + "bbox": [ + 135, + 418, + 480, + 440 + ], + "type": "text", + "content": "2. Domain generalization, which aims to show generalization to the domain shift, especially for out-of-distribution data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 135, + 441, + 480, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 441, + 480, + 477 + ], + "spans": [ + { + "bbox": [ + 135, + 441, + 480, + 477 + ], + "type": "text", + "content": "3. Cross-dataset transfer, which aims to see if the method has the potential to transfer beyond a single dataset. It is a much more challenging problem because the fundamentals can be totally changed across different datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 135, + 478, + 480, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 478, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 135, + 478, + 480, + 514 + ], + "type": "text", + "content": "4. Few-shot learning, which aims to evaluate the adaptation performance of the model to extract knowledge from a dataset whose samples are very few, e.g. 1, 2, 4, 8 or 16 samples." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 520, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 520, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 520, + 482, + 582 + ], + "type": "text", + "content": "All methods are initialized with the same CLIP weights provided by the open-source CLIP [58]. In Appendix, we provide more details of datasets, experimental setting and competitors introduction. Due to page limit, we also put some detailed experimental results into Appendix, e.g. the performance of each method on each dataset in base-to-new generalization setting." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 597, + 228, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 228, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 228, + 609 + ], + "type": "text", + "content": "5.1 Main Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "text", + "content": "Base-to-new Generalization The average results over 11 datasets are shown in Tab. 1. For complete results, please refer to the Appendix. Our proposed QCoOp reaches " + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "inline_equation", + "content": "77.43\\%" + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "text", + "content": " average harmonic mean accuracy within merely 0.26KB size. QCoOp outperforms all kinds of lightweight state-of-the-art methods with" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "T. Hao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 222, + 158, + 392, + 251 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "text", + "content": "Table 1: Comparisons with latest methods in base-to-new generalization. H: harmonic mean [63]. QCoOp and QMaPLe are significantly better than other competitors, which prove that our quantization design could be fruitfully integrated into many existing approaches to further improve the performance and efficiency." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 222, + 158, + 392, + 251 + ], + "lines": [ + { + "bbox": [ + 222, + 158, + 392, + 251 + ], + "spans": [ + { + "bbox": [ + 222, + 158, + 392, + 251 + ], + "type": "table", + "html": "
SizeBaseNewH
CLIP0KB69.3474.2271.70
CoOp4.1KB82.6963.2271.66
CoCoOp70.8KB80.4771.6975.83
Adapter1051KB82.6270.9776.35
LoRA258KB84.3067.3374.86
ProGrad16.4KB82.7968.5575.00
QCoOp0.26KB80.6874.4477.43
MaPLe7096KB82.2875.1478.55
QMaPLe1774KB83.0275.5779.12
", + "image_path": "a297cabd252a9de203c54efb6e0f479c6b28b54bddc608685f67584f06a908e6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 133, + 294, + 480, + 378 + ], + "blocks": [ + { + "bbox": [ + 130, + 261, + 480, + 293 + ], + "lines": [ + { + "bbox": [ + 130, + 261, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 480, + 293 + ], + "type": "text", + "content": "Table 2: Comparisons with latest methods in domain generalization. QCoOp gets comparable or even better results with the latest state-of-the-art methods with much fewer parameters, showing excellent robustness for domain shift." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 294, + 480, + 378 + ], + "lines": [ + { + "bbox": [ + 133, + 294, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 133, + 294, + 480, + 378 + ], + "type": "table", + "html": "
SizeSourceImageNetTargetAverage
-V2-Sketch-Adversarial-Rendition
CLIP0KB66.7360.8346.1547.7773.9657.18
CoOp4.1KB71.5164.2047.9949.7175.2159.28
CoCoOp70.8KB71.0264.0748.7550.6376.1859.91
Adapter1051KB69.3362.5347.6749.1775.4258.70
LoRA258KB70.3062.3742.4338.4068.9753.04
ProGrad16.4KB72.2464.7347.6149.3974.5859.07
QCoOp0.26KB70.6763.8748.9351.1076.9060.20
", + "image_path": "d1f832615b95b605ad681dbab8b6836fd06d73a4f1d35aefeab803208e395ab8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 401, + 480, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 480, + 472 + ], + "type": "text", + "content": "much more efficiency and higher accuracy. For heavier methods like MaPLe, our method can be fruitfully integrated into existing solutions. Besides prompts, we also perform a similar quantization operation on the other weights of MaPLe that are in the linear layers. As a result, QMaPLe not only shows stronger generalization and adaptation capability and gives " + }, + { + "bbox": [ + 130, + 401, + 480, + 472 + ], + "type": "inline_equation", + "content": "0.57\\%" + }, + { + "bbox": [ + 130, + 401, + 480, + 472 + ], + "type": "text", + "content": " higher accuracy, but also enjoys a much more smaller model size compared with the original MaPLe." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 473, + 481, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 481, + 509 + ], + "type": "text", + "content": "Notably, when comparing QCoOp and a lightweight method ProGrad, even if QCoOp is " + }, + { + "bbox": [ + 130, + 473, + 481, + 509 + ], + "type": "inline_equation", + "content": "63 \\times" + }, + { + "bbox": [ + 130, + 473, + 481, + 509 + ], + "type": "text", + "content": " smaller than ProGrad, QCoOp still outperforms it by a clear margin, demonstrating the outstanding efficiency and effectiveness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 526, + 482, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 526, + 482, + 599 + ], + "spans": [ + { + "bbox": [ + 130, + 526, + 482, + 599 + ], + "type": "text", + "content": "Domain Generalization In this paragraph, ImageNet, ImageNet-A, ImageNet-R, ImageNet-v2, and ImageNet-S are used to construct domain generalization experiments. As shown in Tab. 2, on downstream target datasets, QCoOp gets better average accuracy compared with the other methods with significantly better efficiency. For CLIP, CoOp, CLIP-Adapter and ProGrad, there is a clear performance gap between our method and them." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Cross-dataset Transfer In this paragraph, we do cross-dataset transfer evaluation to further verify our QPrompt. Results are shown in Tab. 3. CoOp is good on source domain but fails on target domains. Probably because it focuses too much on the dataset shown to its eyes and face overfitting and catastrophic" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "type": "text", + "content": "Quantized Prompt for Efficient Generalization of Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 148, + 482, + 240 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 3: Results in the cross-dataset transfer setting. QCoOp gives the highest accuracy on 5 of 10 datasets, which well demonstrates that QCoOp could maximally extract general and data-agnostic knowledge from input images." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 148, + 482, + 240 + ], + "lines": [ + { + "bbox": [ + 133, + 148, + 482, + 240 + ], + "spans": [ + { + "bbox": [ + 133, + 148, + 482, + 240 + ], + "type": "table", + "html": "
SizeSourceTargetAverage
ImageNetCaltech101PetsCarsFlowersFood101AircraftSun397DTDEuroSATUCF101
CoOp4.1KB71.5193.7089.1464.5168.7185.3018.4764.1541.9246.3966.5563.88
CoCoOp70.8KB71.0294.4390.1465.3271.8886.0622.9467.3645.7345.3768.2165.74
Adapter1051KB69.3393.4388.8764.4070.2785.6324.6765.8044.9047.7066.0065.17
QCoOp0.26KB70.6394.0790.5365.9771.3386.2322.7366.8044.2048.2369.1765.93
", + "image_path": "7c69785b5ec4919a88347084f5a7147916edb5f9514dfa91c5d840d455ef2ea9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 141, + 257, + 287, + 370 + ], + "blocks": [ + { + "bbox": [ + 141, + 257, + 287, + 370 + ], + "lines": [ + { + "bbox": [ + 141, + 257, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 141, + 257, + 287, + 370 + ], + "type": "image", + "image_path": "8b65cd08a49e60b2a3556308655af1b59224075e9b15641d33174225a8b076f6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 372, + 297, + 392 + ], + "lines": [ + { + "bbox": [ + 138, + 372, + 297, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 372, + 297, + 392 + ], + "type": "text", + "content": "Fig. 5: Average few-shot learning results on 11 datasets." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 302, + 257, + 451, + 371 + ], + "blocks": [ + { + "bbox": [ + 302, + 257, + 451, + 371 + ], + "lines": [ + { + "bbox": [ + 302, + 257, + 451, + 371 + ], + "spans": [ + { + "bbox": [ + 302, + 257, + 451, + 371 + ], + "type": "image", + "image_path": "e73a42a7e4c61104bd9f3503316a71777aff14598a86e8124426444a136919b3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 298, + 372, + 473, + 394 + ], + "lines": [ + { + "bbox": [ + 298, + 372, + 473, + 394 + ], + "spans": [ + { + "bbox": [ + 298, + 372, + 473, + 394 + ], + "type": "text", + "content": "Fig. 6: KLD trend of prompt weights under the same experiment with Fig. 3." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 415, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 464 + ], + "type": "text", + "content": "forgetting problems, which leads to a severe performance drop on unseen objects. QCoOp wins on 5 of 10 datasets and its average accuracy is also slightly better than the best competitor CoCoOp, showing that QCoOp could maximally extract both general and data-agnostic knowledge from given images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 478, + 482, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 478, + 482, + 537 + ], + "spans": [ + { + "bbox": [ + 130, + 478, + 482, + 537 + ], + "type": "text", + "content": "Few-shot Learning Here we will show the experiment results of QCoOp in the few-shot learning setting that is originated from CoOp. Seen from Fig. 5, QCoOp consistently outperforms zero-shot CLIP, CoOp, and CLIP-Adapter across all the shot numbers. Such results demonstrate the superiority of QCoOp in adaptation ability when there are few samples in downstream tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 538, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 538, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 538, + 482, + 586 + ], + "type": "text", + "content": "Overall, in base-to-new generalization, domain generalization, cross-dataset transfer and few-shot learning, the proposed method can consistently accomplish state-of-the-art performance while enjoying extremely high parameter efficiency, fruitfully demonstrating the effectiveness and efficiency of the proposed method." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 601, + 205, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 601, + 205, + 615 + ], + "spans": [ + { + "bbox": [ + 131, + 601, + 205, + 615 + ], + "type": "text", + "content": "6 Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 624, + 246, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 246, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 246, + 635 + ], + "type": "text", + "content": "6.1 Ablation Studies" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Component In this subsection, we decompose the method into pieces and show the influence of each component. As in Tab. 4, we clearly show how" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "T. Hao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 186, + 137, + 422, + 177 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 479, + 136 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 479, + 136 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 479, + 136 + ], + "type": "text", + "content": "Table 4: Ablation study on separate component of QCoOp. Norm: Normalization/Denormalization. CAC: Constrained Adaptive Clustering." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 186, + 137, + 422, + 177 + ], + "lines": [ + { + "bbox": [ + 186, + 137, + 422, + 177 + ], + "spans": [ + { + "bbox": [ + 186, + 137, + 422, + 177 + ], + "type": "table", + "html": "
K-MeansNormCACbasenewH
QCoOp✓✗✗78.7172.5575.50
✓✓✗78.8473.0975.85
✓✓✓78.2474.0276.07
", + "image_path": "287a23776cc2cbc1732a434a4799cfcb38a37dc11c5eabd09f5f3240f9f29c7c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 231, + 210, + 376, + 241 + ], + "blocks": [ + { + "bbox": [ + 131, + 188, + 479, + 209 + ], + "lines": [ + { + "bbox": [ + 131, + 188, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 131, + 188, + 479, + 209 + ], + "type": "text", + "content": "Table 5: Comparisons between QAT and PTQ based on CoOp. Both QAT and CoOp are trained with the same hyper-parameters. Here the quantization bit is 1." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 231, + 210, + 376, + 241 + ], + "lines": [ + { + "bbox": [ + 231, + 210, + 376, + 241 + ], + "spans": [ + { + "bbox": [ + 231, + 210, + 376, + 241 + ], + "type": "table", + "html": "
basenewH
QAT80.7272.3576.31
PTQ82.2168.5074.73
", + "image_path": "cd1517ffd452c624abb2a5fb92662394b84a1425be87c4ebc5ffc130adb62a90.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 265, + 479, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 265, + 479, + 312 + ], + "spans": [ + { + "bbox": [ + 131, + 265, + 479, + 312 + ], + "type": "text", + "content": "much the K-Means algorithm, normalization/denormalization and constrained adaptive clustering influence the final performance. One interesting point is that K-Means+Norm+CAC only improves the new accuracy compared with K-Means+Norm, showing the superiority of our constrained adaptive clustering." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 319, + 479, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 319, + 479, + 366 + ], + "spans": [ + { + "bbox": [ + 131, + 319, + 479, + 366 + ], + "type": "text", + "content": "Trend of the distributions of prompt between adjacent epochs. To verify the opinions we proposed at the end of Sec. 6, we show the KLD of prompt distributions between adjacent epochs during CoOp's training in Fig. 6. Clearly, this trend is consistent with what we summarized before." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 373, + 479, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 373, + 479, + 445 + ], + "spans": [ + { + "bbox": [ + 131, + 373, + 479, + 445 + ], + "type": "text", + "content": "QAT v.s. PTQ In Tab. 5, we study the choice of QAT or PTQ following the same strategy, K-Means clustering, in the paper. Results show that QAT consistently outperforms PTQ with the same hyper-parameters. The accuracy on unseen new classes of PTQ is significantly lower than QAT, again proving our opinion that quantization helps generalization by alleviating overfitting as well as catastrophic forgetting. Quantization error is not always undesirable." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 453, + 479, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 453, + 479, + 488 + ], + "spans": [ + { + "bbox": [ + 131, + 453, + 479, + 488 + ], + "type": "text", + "content": "Quantization bit In Tab. 6 and Fig. 7, we show the results across multiple quantization bits. In conclusions, more bits did not always lead to good results, and new accuracy continues decreasing as the training goes on." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 508, + 443, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 508, + 443, + 520 + ], + "spans": [ + { + "bbox": [ + 131, + 508, + 443, + 520 + ], + "type": "text", + "content": "6.2 Performance on Self-Supervised Vision-Language Model" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 529, + 479, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 529, + 479, + 589 + ], + "spans": [ + { + "bbox": [ + 131, + 529, + 479, + 589 + ], + "type": "text", + "content": "In this paragraph, we explore the usage of QCoOp among other different backbones besides CLIP. We choose a self-supervised vision-language model, SLIP [53] to further verify the universality and robustness of our proposed method. The experimental results are shown in Tab. 7. We could see that the base accuracies of QCoOp and CoOp are similar, but the new accuracies of QCoOp is much" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 197, + 621, + 411, + 662 + ], + "blocks": [ + { + "bbox": [ + 201, + 611, + 410, + 620 + ], + "lines": [ + { + "bbox": [ + 201, + 611, + 410, + 620 + ], + "spans": [ + { + "bbox": [ + 201, + 611, + 410, + 620 + ], + "type": "text", + "content": "Table 6: Results across different quantization bits." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 197, + 621, + 411, + 662 + ], + "lines": [ + { + "bbox": [ + 197, + 621, + 411, + 662 + ], + "spans": [ + { + "bbox": [ + 197, + 621, + 411, + 662 + ], + "type": "table", + "html": "
SizebasenewH
QCoOp (b=1)0.26KB79.4972.6575.92
QCoOp (b=2)0.52KB80.3071.9875.91
QCoOp (b=4)1.05KB80.9271.2275.76
", + "image_path": "c5495eac07bf1f594ad42f6860c1082094d7ad9bf51df64caebd4717a71c30c8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "type": "text", + "content": "Quantized Prompt for Efficient Generalization of Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 126, + 289, + 230 + ], + "blocks": [ + { + "bbox": [ + 149, + 126, + 289, + 230 + ], + "lines": [ + { + "bbox": [ + 149, + 126, + 289, + 230 + ], + "spans": [ + { + "bbox": [ + 149, + 126, + 289, + 230 + ], + "type": "image", + "image_path": "0d118ce17f079fe7974dd01be1f854a44b3ee9131443371cc8a04cc25aded9c0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 231, + 482, + 264 + ], + "lines": [ + { + "bbox": [ + 130, + 231, + 482, + 264 + ], + "spans": [ + { + "bbox": [ + 130, + 231, + 482, + 264 + ], + "type": "text", + "content": "Fig. 7: The same training curves with Fig. 2. Clearly, increasing quantization noise leads to the same phenomenon with increasing gaussian noise, i.e. more generalizability (new accuracy) and less adaptability (base accuracy)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 311, + 126, + 449, + 230 + ], + "blocks": [ + { + "bbox": [ + 311, + 126, + 449, + 230 + ], + "lines": [ + { + "bbox": [ + 311, + 126, + 449, + 230 + ], + "spans": [ + { + "bbox": [ + 311, + 126, + 449, + 230 + ], + "type": "image", + "image_path": "233c5cf195cce4aabca49cfa8561d3cdf849d18b4353cf4e2ce97bced1439ac6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 207, + 277, + 402, + 308 + ], + "blocks": [ + { + "bbox": [ + 215, + 267, + 397, + 277 + ], + "lines": [ + { + "bbox": [ + 215, + 267, + 397, + 277 + ], + "spans": [ + { + "bbox": [ + 215, + 267, + 397, + 277 + ], + "type": "text", + "content": "Table 7: Results based on SLIP [53] model." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 207, + 277, + 402, + 308 + ], + "lines": [ + { + "bbox": [ + 207, + 277, + 402, + 308 + ], + "spans": [ + { + "bbox": [ + 207, + 277, + 402, + 308 + ], + "type": "table", + "html": "
SizebasenewH
CoOp4.1KB68.6546.6055.51
QCoOp0.26KB68.3374.0471.07
", + "image_path": "26236f5e18e5be315a25b2a86f048093ab76d7223cee06cbf724535c8f3d68a3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 331, + 480, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 480, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 480, + 355 + ], + "type": "text", + "content": "higher than CoOp. Such observation verifies our assumption that quantization is quite helpful for generalization again." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 373, + 220, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 373, + 220, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 373, + 220, + 385 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 397, + 482, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 482, + 636 + ], + "type": "text", + "content": "With the development of huge vision-language models, how to effectively and efficiently adapt such huge models to downstream tasks becomes a challenging problem. Much effort has been made to leverage the potential of prompt tuning in adapting vision-language models. However, existing methods suffer from inefficiency. To reach extremely efficient generalization, we propose QPrompt based on the detailed analysis and deep understanding of the characteristics of the prompt weight distribution. Following the several principles concluded by us, we use K-Means clustering algorithm as the base of our quantization method. To adaptively control the quantization error and minimize the number of re-clustering operations, we propose to do a dynamic check every few iterations. If the Kullback-Leibler divergence between the current weights and the original weights used in the last clustering exceeds a predefined threshold and the quantization error also increases, we will let the model re-cluster the weights and update the new centers and weights. Similarly, continue this process until completion. The proposed method could be simply integrated into many of the existing vision-language tuning methods like CoOp and MaPLe and reach good performance. Importantly, our proposed method is significantly effective and efficient. Extensive experiments show that the designed quantization algorithm indeed improves the generalizability of vision-language model, and save a lot more storage space as well." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by National Natural Science Foundation of China (Nos. 61925107, 62271281, 62021002)" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "T. Hao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 137, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 137, + 137, + 481, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 137, + 481, + 169 + ], + "spans": [ + { + "bbox": [ + 137, + 137, + 481, + 169 + ], + "type": "text", + "content": "1. Bahng, H., Jahanian, A., Sankaranarayanan, S., Isola, P.: Visual prompting: Modifying pixel space to adapt pre-trained models. arXiv preprint arXiv:2203.17274 (2022) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 170, + 481, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 170, + 481, + 201 + ], + "spans": [ + { + "bbox": [ + 137, + 170, + 481, + 201 + ], + "type": "text", + "content": "2. Banner, R., Nahshan, Y., Soudry, D.: Post training 4-bit quantization of convolutional networks for rapid-deployment. Advances in Neural Information Processing Systems 32 (2019) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 202, + 481, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 202, + 481, + 245 + ], + "spans": [ + { + "bbox": [ + 137, + 202, + 481, + 245 + ], + "type": "text", + "content": "3. Bhalgat, Y., Lee, J., Nagel, M., Blankevoort, T., Kwak, N.: Lsq+: Improving low-bit quantization through learnable offsets and better initialization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 696-697 (2020) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 245, + 481, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 481, + 277 + ], + "type": "text", + "content": "4. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. In: The Eleventh International Conference on Learning Representations (2023) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 277, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 481, + 319 + ], + "type": "text", + "content": "5. Bulat, A., Tzimiropoulos, G.: Lasp: Text-to-text optimization for language-aware soft prompting of vision & language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23232-23241 (2023) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 320, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 481, + 352 + ], + "type": "text", + "content": "6. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 353, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 353, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 353, + 481, + 385 + ], + "type": "text", + "content": "7. Ding, J., Xue, N., Xia, G.S., Dai, D.: Decoupling zero-shot semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11583-11592 (2022) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 385, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 385, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 385, + 481, + 418 + ], + "type": "text", + "content": "8. Ding, X., Ding, G., Guo, Y., Han, J.: Centripetal sgd for pruning very deep convolutional networks with complicated structure. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4943-4953 (2019) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 418, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 481, + 450 + ], + "type": "text", + "content": "9. Ding, X., Guo, Y., Ding, G., Han, J.: Acnet: Strengthening the kernel skeletons for powerful cnn via asymmetric convolution blocks. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1911-1920 (2019) 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 450, + 481, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 450, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 134, + 450, + 481, + 491 + ], + "type": "text", + "content": "10. Ding, X., Hao, T., Tan, J., Liu, J., Han, J., Guo, Y., Ding, G.: Resrep: Lossless cnn pruning via decoupling remembering and forgetting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4510-4520 (2021) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 492, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 492, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 134, + 492, + 481, + 525 + ], + "type": "text", + "content": "11. Ding, X., Zhang, X., Han, J., Ding, G.: Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11963-11975 (2022) 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 525, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 525, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 134, + 525, + 481, + 557 + ], + "type": "text", + "content": "12. Ding, X., Zhang, X., Ma, N., Han, J., Ding, G., Sun, J.: Repvgg: Making vgg-style convnets great again. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13733-13742 (2021) 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 558, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 558, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 134, + 558, + 481, + 601 + ], + "type": "text", + "content": "13. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020) 1, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 134, + 601, + 481, + 633 + ], + "type": "text", + "content": "14. Esser, S.K., McKinstry, J.L., Bablani, D., Appuswamy, R., Modha, D.S.: Learned step size quantization. In: International Conference on Learning Representations (2019) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 481, + 666 + ], + "type": "text", + "content": "15. Feng, C., Zhong, Y., Jie, Z., Chu, X., Ren, H., Wei, X., Xie, W., Ma, L.: Promptdet: Towards open-vocabulary detection using uncurated images. In: European Conference on Computer Vision. pp. 701-717. Springer (2022) 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "text", + "content": "Quantized Prompt for Efficient Generalization of Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "16. Finkelstein, A., Almog, U., Grobman, M.: Fighting quantization bias with bias. arXiv preprint arXiv:1906.03193 (2019) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 172 + ], + "type": "text", + "content": "17. Gholami, A., Kim, S., Dong, Z., Yao, Z., Mahoney, M.W., Keutzer, K.: A survey of quantization methods for efficient neural network inference. In: Low-Power Computer Vision, pp. 291-326. Chapman and Hall/CRC (2022) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 173, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 173, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 173, + 480, + 205 + ], + "type": "text", + "content": "18. Han, S., Mao, H., Dally, W.J.: Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149 (2015) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 206, + 480, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 206, + 480, + 238 + ], + "spans": [ + { + "bbox": [ + 134, + 206, + 480, + 238 + ], + "type": "text", + "content": "19. Hao, T., Chen, H., Guo, Y., Ding, G.: Consolidator: Mergeable adapter with grouped connections for visual adaptation. arXiv preprint arXiv:2305.00603 (2023) 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 240, + 480, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 240, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 133, + 240, + 480, + 274 + ], + "type": "text", + "content": "20. Hao, T., Ding, X., Han, J., Guo, Y., Ding, G.: Manipulating identical filter redundancy for efficient pruning on deep and complicated cnn. IEEE Transactions on Neural Networks and Learning Systems (2023) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 480, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 480, + 308 + ], + "type": "text", + "content": "21. Hao, T., Lyu, M., Chen, H., Zhao, S., Han, J., Ding, G.: Re-parameterized low-rank prompt: Generalize a vision-language model within " + }, + { + "bbox": [ + 132, + 274, + 480, + 308 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{k}" + }, + { + "bbox": [ + 132, + 274, + 480, + 308 + ], + "type": "text", + "content": " parameters. arXiv preprint arXiv:2312.10813 (2023) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 308, + 480, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 480, + 341 + ], + "type": "text", + "content": "22. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 342, + 480, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 384 + ], + "type": "text", + "content": "23. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019) 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 385, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 480, + 418 + ], + "type": "text", + "content": "24. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021) 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 419, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 480, + 453 + ], + "type": "text", + "content": "25. Hubara, I., Nahshan, Y., Hanani, Y., Banner, R., Soudry, D.: Accurate post training quantization with small calibration sets. In: International Conference on Machine Learning. pp. 4466-4475. PMLR (2021) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 453, + 480, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 453, + 480, + 496 + ], + "spans": [ + { + "bbox": [ + 133, + 453, + 480, + 496 + ], + "type": "text", + "content": "26. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International Conference on Machine Learning. pp. 4904-4916. PMLR (2021) 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 498, + 480, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 480, + 519 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 480, + 519 + ], + "type": "text", + "content": "27. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. arXiv preprint arXiv:2203.12119 (2022) 3, 4, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 520, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 480, + 553 + ], + "type": "text", + "content": "28. Jiang, Z., Araki, J., Ding, H., Neubig, G.: How can we know when language models know? on the calibration of language models for question answering. Transactions of the Association for Computational Linguistics 9, 962-977 (2021) 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 554, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 480, + 586 + ], + "type": "text", + "content": "29. Jie, S., Wang, H., Deng, Z.H.: Revisiting the parameter efficiency of adapters from the perspective of precision redundancy. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17217-17226 (2023) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 587, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 480, + 620 + ], + "type": "text", + "content": "30. Ju, C., Han, T., Zheng, K., Zhang, Y., Xie, W.: Prompting visual-language models for efficient video understanding. In: European Conference on Computer Vision. pp. 105-124. Springer (2022) 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 480, + 665 + ], + "type": "text", + "content": "31. Jung, S., Son, C., Lee, S., Son, J., Han, J.J., Kwak, Y., Hwang, S.J., Choi, C.: Learning to quantize deep networks by optimizing quantization intervals with task loss. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4350-4359 (2019) 4" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "T. Hao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "text", + "content": "32. Khattak, M.U., Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Maple: Multi-modal prompt learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19113-19122 (2023) 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 193 + ], + "type": "text", + "content": "33. Khattak, M.U., Wasim, S.T., Naseer, M., Khan, S., Yang, M.H., Khan, F.S.: Self-regulating prompts: Foundational model adaptation without forgetting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15190-15200 (2023) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "34. Lee, D., Song, S., Suh, J., Choi, J., Lee, S., Kim, H.J.: Read-only prompt optimization for vision-language few-shot learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1401-1411 (2023) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "text", + "content": "35. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 (2021) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "36. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=RriDjddCLN 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 304 + ], + "type": "text", + "content": "37. Li, H., Kadav, A., Durdanovic, I., Samet, H., Graf, H.P.: Pruning filters for efficient convnets. arXiv preprint arXiv:1608.08710 (2016) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "text", + "content": "38. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 (2021) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "type": "text", + "content": "39. Li, Y., Xu, S., Zhang, B., Cao, X., Gao, P., Guo, G.: Q-vit: Accurate and fully quantized low-bit vision transformer. Advances in Neural Information Processing Systems 35, 34451-34463 (2022) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "text", + "content": "40. Li, Z., Xiao, J., Yang, L., Gu, Q.: Repq-vit: Scale reparameterization for posttraining quantization of vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17227-17236 (2023) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 391, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 481, + 423 + ], + "type": "text", + "content": "41. Lian, D., Zhou, D., Feng, J., Wang, X.: Scaling & shifting your features: A new baseline for efficient model tuning. In: Advances in Neural Information Processing Systems (NeurIPS) (2022) 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "text", + "content": "42. Liu, P., Yuan, W., Fu, J., Jiang, Z., Hayashi, H., Neubig, G.: Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. arXiv preprint arXiv:2107.13586 (2021) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 489 + ], + "type": "text", + "content": "43. Liu, X., Ji, K., Fu, Y., Du, Z., Yang, Z., Tang, J.: P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv preprint arXiv:2110.07602 (2021) 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 490, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 481, + 511 + ], + "type": "text", + "content": "44. Liu, X., Zheng, Y., Du, Z., Ding, M., Qian, Y., Yang, Z., Tang, J.: Gpt understands, too. arXiv preprint arXiv:2103.10385 (2021) 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "text", + "content": "45. Liu, Z., Wang, Y., Han, K., Zhang, W., Ma, S., Gao, W.: Post-training quantization for vision transformer. Advances in Neural Information Processing Systems 34, 28092-28103 (2021) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "text", + "content": "46. Liu, Z., Li, J., Shen, Z., Huang, G., Yan, S., Zhang, C.: Learning efficient convolutional networks through network slimming. In: Proceedings of the IEEE international conference on computer vision. pp. 2736-2744 (2017) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "type": "text", + "content": "47. Lüddecke, T., Ecker, A.: Image segmentation using text and image prompts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7086-7096 (2022) 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "text", + "content": "48. Lyu, M., Hao, T., Xu, X., Chen, H., Han, J., Ding, G.: Learn from the learnt: Source-free active domain adaptation via contrastive sampling and visual persistence. In: European Conference on Computer Vision (ECCV). Springer (2024) 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "text", + "content": "49. Lyu, M., Yang, Y., Hong, H., Chen, H., Jin, X., He, Y., Xue, H., Han, J., Ding, G.: One-dimensional adapter to rule them all: Concepts diffusion models and erasing" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "text", + "content": "Quantized Prompt for Efficient Generalization of Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "type": "text", + "content": "applications. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7559-7568 (2024) 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "type": "text", + "content": "50. Lyu, M., Zhou, J., Chen, H., Huang, Y., Yu, D., Li, Y., Guo, Y., Guo, Y., Xiang, L., Ding, G.: Box-level active detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23766-23775 (2023) 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "type": "text", + "content": "51. Maaz, M., Rasheed, H., Khan, S., Khan, F.S., Anwer, R.M., Yang, M.H.: Class-agnostic object detection with multi-modal transformer. In: The European Conference on Computer Vision. Springer (2022) 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "type": "text", + "content": "52. Meller, E., Finkelstein, A., Almog, U., Grobman, M.: Same, same but different: Recovering neural network quantization error through weight factorization. In: International Conference on Machine Learning. pp. 4486-4495. PMLR (2019) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 238, + 481, + 270 + ], + "type": "text", + "content": "53. Mu, N., Kirillov, A., Wagner, D., Xie, S.: Slip: Self-supervision meets language-image pre-training. In: European Conference on Computer Vision. pp. 529-544. Springer (2022) 1, 13, 14" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "text", + "content": "54. Nagel, M., Baalen, M.v., Blankevoort, T., Welling, M.: Data-free quantization through weight equalization and bias correction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1325-1334 (2019) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 304, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 304, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 133, + 304, + 481, + 335 + ], + "type": "text", + "content": "55. Nagel, M., Fournarakis, M., Amjad, R.A., Bondarenko, Y., Van Baalen, M., Blankevoort, T.: A white paper on neural network quantization. arXiv preprint arXiv:2106.08295 (2021) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "56. Nagel, M., Fournarakis, M., Bondarenko, Y., Blankevoort, T.: Overcoming oscillations in quantization-aware training. In: International Conference on Machine Learning. pp. 16318-16330. PMLR (2022) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "text", + "content": "57. Qian, R., Li, Y., Xu, Z., Yang, M.H., Belongie, S., Cui, Y.: Multimodal open-vocabulary video classification via pre-trained vision and language models. arXiv preprint arXiv:2207.07646 (2022) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 403, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 403, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 133, + 403, + 481, + 445 + ], + "type": "text", + "content": "58. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021) 1, 3, 4, 10" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 447, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 447, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 133, + 447, + 481, + 489 + ], + "type": "text", + "content": "59. Rao, Y., Zhao, W., Chen, G., Tang, Y., Zhu, Z., Huang, G., Zhou, J., Lu, J.: Denseclip: Language-guided dense prediction with context-aware prompting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18082-18091 (2022) 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 491, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 491, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 133, + 491, + 481, + 533 + ], + "type": "text", + "content": "60. Rasheed, H.A., Maaz, M., Khattak, M.U., Khan, S., Khan, F.: Bridging the gap between object and image-level representations for open-vocabulary detection. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=aKXBrjODHm 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 534, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 534, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 133, + 534, + 481, + 577 + ], + "type": "text", + "content": "61. Shin, T., Razeghi, Y., Logan IV, R.L., Wallace, E., Singh, S.: Autoprompt: Eliciting knowledge from language models with automatically generated prompts. In: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). pp. 4222-4235 (2020) 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 578, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 578, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 133, + 578, + 481, + 610 + ], + "type": "text", + "content": "62. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research 15(1), 1929-1958 (2014) 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 611, + 481, + 632 + ], + "type": "text", + "content": "63. Xian, Y., Schiele, B., Akata, Z.: Zero-shot learning-the good, the bad and the ugly. In: CVPR (2017) 11" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 481, + 665 + ], + "type": "text", + "content": "64. Xiong, Y., Chen, H., Hao, T., Lin, Z., Han, J., Zhang, Y., Wang, G., Bao, Y., Ding, G.: Pyra: Parallel yielding re-activation for training-inference efficient task adaptation. arXiv preprint arXiv:2403.09192 (2024) 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "T. Hao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 544 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "65. Xiong, Y., Chen, H., Lin, Z., Zhao, S., Ding, G.: Confidence-based visual dispersal for few-shot unsupervised domain adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11621-11631 (2023) 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "text", + "content": "66. Xiong, Y., Chen, X., Ye, X., Chen, H., Lin, Z., Lian, H., Niu, J., Ding, G.: Temporal scaling law for large language models. arXiv preprint arXiv:2404.17785 (2024) 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 172, + 481, + 205 + ], + "type": "text", + "content": "67. Yao, H., Zhang, R., Xu, C.: Visual-language prompt tuning with knowledge-guided context optimization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6757-6767 (2023) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 205, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 481, + 247 + ], + "type": "text", + "content": "68. Yao, L., Huang, R., Hou, L., Lu, G., Niu, M., Xu, H., Liang, X., Li, Z., Jiang, X., Xu, C.: FILIP: Fine-grained interactive language-image pre-training. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=cpDhcsEDC2 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 248, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 481, + 281 + ], + "type": "text", + "content": "69. Yu, S., Chen, T., Shen, J., Yuan, H., Tan, J., Yang, S., Liu, J., Wang, Z.: Unified visual transformer compression. In: International Conference on Learning Representations (2022) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 281, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 481, + 314 + ], + "type": "text", + "content": "70. Yuan, L., Chen, D., Chen, Y.L., Codella, N., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., et al.: Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432 (2021) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 315, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 481, + 347 + ], + "type": "text", + "content": "71. Yuan, Z., Xue, C., Chen, Y., Wu, Q., Sun, G.: Ptg4vit: Post-training quantization for vision transformers with twin uniform quantization. In: European Conference on Computer Vision. pp. 191-207. Springer (2022) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 347, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 481, + 369 + ], + "type": "text", + "content": "72. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary detr with conditional matching. arXiv preprint arXiv:2203.11876 (2022) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 369, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 481, + 413 + ], + "type": "text", + "content": "73. Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., Beyer, L.: Lit: Zero-shot transfer with locked-image text tuning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18123-18133 (2022) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 413, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 481, + 435 + ], + "type": "text", + "content": "74. Zhang, Y., Zhou, K., Liu, Z.: Neural prompt search. arXiv preprint arXiv:2206.04673 (2022) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 435, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 481, + 479 + ], + "type": "text", + "content": "75. Zheng, K., Wu, W., Feng, R., Zhu, K., Liu, J., Zhao, D., Zha, Z.J., Chen, W., Shen, Y.: Regularized mask tuning: Uncovering hidden knowledge in pre-trained vision-language models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11663-11673 (2023) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 481, + 511 + ], + "type": "text", + "content": "76. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022) 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 511, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 511, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 511, + 481, + 544 + ], + "type": "text", + "content": "77. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022) 3, 4, 6, 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "text", + "content": "Quantized Prompt for Efficient Generalization of Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/QueryCDR_ Query-based Controllable Distortion Rectification Network for Fisheye Images/8182781d-9f7c-4e2a-86d5-0c2e6b9982a2_content_list.json b/2024/QueryCDR_ Query-based Controllable Distortion Rectification Network for Fisheye Images/8182781d-9f7c-4e2a-86d5-0c2e6b9982a2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fdeac3a4b0e4b10f9609e9e43f06d4940ff5a9f7 --- /dev/null +++ b/2024/QueryCDR_ Query-based Controllable Distortion Rectification Network for Fisheye Images/8182781d-9f7c-4e2a-86d5-0c2e6b9982a2_content_list.json @@ -0,0 +1,1809 @@ +[ + { + "type": "text", + "text": "QueryCDR: Query-based Controllable Distortion Rectification Network for Fisheye Images", + "text_level": 1, + "bbox": [ + 222, + 140, + 781, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pengbo Guo $^{1,2}$ , Chengxu Liu $^{2,3}$ , Xingsong Hou $^{2(\\boxtimes)}$ , and Xueming Qian $^{2,3}$", + "bbox": [ + 279, + 210, + 720, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Software Engineering, Xi'an Jiaotong University, Xi'an, China \n $^{2}$ Xi'an Jiaotong University, Xi'an, China", + "bbox": [ + 251, + 253, + 746, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Shaanxi Yulan Jiuzhou Intelligent Optoelectronic Tech. Co., Ltd, Xi'an, China {guopengbo,chengxuliu}@stu.xjtu.edu.cn, {houxs,qianxm}@mail.xjtu.edu.cn", + "bbox": [ + 225, + 282, + 774, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Fisheye image rectification aims to correct distortions in images taken with fisheye cameras. Although current models show promising results on images with a similar degree of distortion as the training data, they will produce sub-optimal results when the degree of distortion changes and without retraining. The lack of generalization ability for dealing with varying degrees of distortion limits their practical application. In this paper, we take one step further to enable effective distortion rectification for images with varying degrees of distortion without retraining. We propose a novel Query-based Controllable Distortion Rectification network for fisheye images (QueryCDR). In particular, we first present the Distortion-aware Learnable Query Mechanism (DLQM), which defines the latent spatial relationships for different distortion degrees as a series of learnable queries. Each query can be learned to obtain position-dependent rectification control conditions, providing control over the rectification process. Then, we propose two kinds of controllable modulating blocks to enable the control conditions to guide the modulation of the distortion features better. These core components cooperate with each other to effectively boost the generalization ability of the model at varying degrees of distortion. Extensive experiments on fisheye image datasets with different distortion degrees demonstrate our approach achieves high-quality and controllable distortion rectification. Code is available at https://github.com/PbGuo/QueryCDR.", + "bbox": [ + 259, + 345, + 743, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Fisheye image $\\cdot$ Distortion rectification $\\cdot$ Controllable", + "bbox": [ + 261, + 662, + 705, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 703, + 375, + 719 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Benefiting from the huge field-of-view (FoV), fisheye cameras are widely utilized in various fields, including security surveillance [26, 34] and autonomous driving [11, 38]. However, the distortion brought by the fisheye lenses greatly limits the performance of downstream vision tasks [9, 21, 36, 60]. How to eliminate the distortion in fisheye images has attracted great attention in recent years.", + "bbox": [ + 212, + 733, + 787, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Early methods [3,5,14,18,37,40,46,59] primarily relied on identifying matching feature points or curves for automatic rectification. However, constrained by", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/411c437a400b17158787ed04a95c04437841e28263fb6a1ccba12b2d37dc9705.jpg", + "image_caption": [ + "Fig. 1: Different approaches to fisheye image distortion rectification. (a) Regression-Based: Using a neural network to predict distortion-related parameters, then apply rectification algorithms $R$ for rectification. (b) Generation-Based: Input the distorted fisheye image and directly generate the rectified image end-to-end. (c) Control-Based: Users provide control conditions to guide the rectification process, resulting in promising rectified images of various distortion degrees." + ], + "image_footnote": [], + "bbox": [ + 215, + 142, + 784, + 281 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the instability of feature detection, both the generalization of the algorithm and the quality of the rectified images were unsatisfactory [10, 24, 52]. In recent years, owing to the robust learning and generalization capabilities of neural networks, deep learning-based fisheye image rectification methods have become mainstream. These methods can be divided into two categories: regression-based rectification [4, 23, 39, 50, 56] and generation-based rectification [10, 24, 25, 52, 54]. The former uses deep regression models to predict distortion parameters for image reconstruction, as shown in Fig. 1(a). The latter uses an encoder-decoder structure to generate well-rectified images directly, as shown in Fig. 1(b).", + "bbox": [ + 212, + 411, + 787, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "It is worth noting that these methods only achieve satisfactory results on a similar degree of distortion as the training data. This means that when handling different degrees of distortion without retraining, there will be a significant decrease in the quality of the rectified images [25]. This is because the model tends to learn fixed position mapping relationships during training. When the distortion degree changes, this relationship does not work for the new distortion distribution. Therefore, these models usually need to be retrained on images with different distortion degrees. In addition, fisheye image acquisition is difficult, and re-collecting new datasets with varying degrees of distortion would incur significant costs [52, 54]. Thus, it is essential to explore a model that can handle different degrees of distortion simultaneously.", + "bbox": [ + 212, + 550, + 787, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, some methods have been proposed to achieve controllable image restoration [7,15,16,33,47,48,55,61]. Typically, CFSNet [47], MM-RealSR [33], and Yao et al. [55] introduce scalars as control conditions to effectively restore degradation of varying degrees. It would be a promising solution to introduce an effectively controllable mechanism to deal with all degrees of distortion, as shown in Fig. 1(c). However, directly applying existing controllable mechanisms in restoration tasks to the fisheye rectification model suffers from the following challenges: 1) There is a gap between optimization objectives (i.e. the roles", + "bbox": [ + 212, + 719, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "P. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of the controllable mechanisms). Restoration models learn the pixel-level detail restoration, whereas fisheye rectification models learn the spatial-level positional mapping relationships [10,25]. The control mechanism lacking position information in restoration tasks is ineffective in controlling distortion rectification networks. 2) There is a gap between optimization difficulties (i.e. the roles of the control conditions). The distortion of a fisheye image increases gradually from the image center to boundary [51]. Therefore, it is not suitable to use a single scalar control condition in restoration tasks to deal with such spatially varying distortions. These challenges restrict the application of controllable mechanisms and control conditions in fisheye image rectification.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address these issues, we propose the Query-based Controllable Distortion Rectification network (QueryCDR), as shown in Fig. 2. By introducing a series of learnable queries as control conditions, QueryCDR allows the users to achieve fisheye rectification with different distortion degrees. Specifically, to incorporate positional mapping relationships into control conditions, we introduce the Distortion-aware Learnable Query Mechanism (DLQM), which defines a series of queries representing different rectification control conditions. During inference, DLQM extracts position-dependent control conditions from the user-given query and feeds them into the network for controlling the rectification process. Furthermore, to enable the control conditions to guide the rectification efficiently, we propose two types of controllable modulating blocks: the Controllable Convolution Modulating Block (CCMB) based on CNN [17], and the Controllable Attention Modulating Block (CAMB) based on Transformer [45]. They are good at extracting local texture features and learning long-range distortion mapping relationships, respectively. By combining CCMB and CAMB, we construct a robust controllable rectification network. Our QueryCDR can handle various distortions without retraining, enhancing the generalization ability of the fisheye rectification model.", + "bbox": [ + 212, + 297, + 787, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We summarize our contributions as follows:", + "bbox": [ + 240, + 569, + 553, + 583 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose QueryCDR, a Query-based Controllable Distortion Rectification network for fisheye images. Extensive experiments demonstrate that our QueryCDR can deliver superior results on a variety of distortion degrees.", + "- We propose the Distortion-aware Learnable Query Mechanism (DLQM), which effectively introduces the latent spatial relationships to control conditions for fisheye image rectification.", + "- We propose two kinds of blocks for modulating features using control conditions: the Controllable Convolution Modulating Block (CCMB) and the Controllable Attention Modulating Block (CAMB). They can effectively utilize control conditions to guide the rectification process." + ], + "bbox": [ + 225, + 589, + 784, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 758, + 387, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Traditional Fisheye Image Rectification", + "text_level": 1, + "bbox": [ + 215, + 787, + 584, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Traditional rectification methods can be divided into two types, multi-view-based and line-based methods. Multi-view-based methods [3,14,18,20,37,41,43]", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "QueryCDR: Query-based Controllable Distortion Rectification Network", + "bbox": [ + 254, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "calibrated fisheye images by finding corresponding feature points from multiple viewpoints. Line-based methods [2,5,8,32,40,44,46,59] employed line detection to rectify the curved lines, thereby achieving distortion rectification. However, these methods require manual intervention or handcrafted feature extractors, and the rectification process is unstable, failing to achieve satisfactory results.", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Deep Learning Based Fisheye Image Rectification", + "text_level": 1, + "bbox": [ + 214, + 243, + 671, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rapid advances in deep learning have allowed them to shine in low-level vision tasks [6, 22, 28-31, 49, 58]. As a core task in low-level vision, distortion image rectification has received increasing attention [10, 24, 39, 52, 54, 56]. According to the network architecture, deep learning-based distortion image rectification methods can be categorized into two types, regression-based methods [4, 23, 39, 50, 56] and generation-based methods [10, 24, 25, 52, 54].", + "bbox": [ + 212, + 266, + 785, + 356 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Regression-based methods [4, 23, 39, 50, 56] utilize neural networks to predict distortion-related coefficients for rectifying the image. Rong et al. [39] were the first to use CNNs for fisheye image rectification. They employed the network to predict across multiple distortion intervals, achieving preliminary rectification results. Yin et al. [56] integrated semantics as prior information to guide rectification. While these methods have shown some performance improvement, they face the challenge of non-end-to-end design, thus requiring a large number of additional labels, and increasing operational costs. Therefore, to reduce models' complexity, Generation-based methods [10, 24, 25, 52, 54] employ generative networks to take distorted images as input and directly generate rectified images. DR-GAN [24] was the first to apply a generative adversarial network (GAN) [12] to fisheye image rectification, enabling the network to directly generate rectified images without estimating additional parameters. DDM [25] introduced distortion maps to help the model better learn the distortion distribution. PCN [52] designed a flow estimation module to predict appearance flows in fisheye images, using it to assist the rectification progress. SimFIR [10] introduced a self-supervised rectification module, allowing the network to better learn the distortion representations at different locations. However, both paradigms require retraining when handling images with different distortion degrees, failing to address the issue of weak model generalization.", + "bbox": [ + 212, + 357, + 787, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Controllable Low-level Vision", + "text_level": 1, + "bbox": [ + 214, + 680, + 504, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address the issue of varying degradation levels in low-level vision tasks, an increasing number of studies [7,15,16,33,47,48,55,61] propose controllable network architectures to tackle this challenge. DNI [48] interpolated all parameters of different restoration networks, which were trained with different degradation levels. By adjusting the interpolation coefficients, a smooth control of the image can be achieved. AdaFM [15] achieved better results by inserting AdaFM layers after each convolution layer to change the filters' statistics, thus the users can interactively manipulate the restoration results by tuning a control coefficient. CFS-Net [47] introduced the tuning branch to adaptively learn the control coefficients,", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "P. Guo et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/acb5866890a2dd5ad447ffb26a49a8749b43d0162695c1893cd98db73fc9af47.jpg", + "image_caption": [ + "Fig. 2: Overview of our proposed Query-based Controllable Distortion Rectification network (QueryCDR). The Distortion-aware Learnable Query Mechanism (DLQM) extracts control conditions from user-given queries and feeds them layer by layer into the rectification network. The rectification network is composed of Controllable Convolution Modulating Blocks (CCMB) and Controllable Attention Modulating Blocks (CAMB), which modulate the input features $F_{in}$ with control conditions $F_{c}$ , enabling controllable rectification process." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 787, + 351 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and then use them to couple the features with the main branch. MM-RealSR [33] proposed a metric learning strategy to map unquantifiable degradation levels to a metric space as control conditions. However, due to the different optimization objectives and complexity of the control mechanisms, directly applying these methods to fisheye image distortion rectification tasks will not achieve effective control over the rectification process.", + "bbox": [ + 212, + 488, + 787, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 602, + 380, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Overview", + "text_level": 1, + "bbox": [ + 215, + 633, + 341, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overview of our proposed Query-based Controllable Distortion Rectification network (QueryCDR) is shown in Fig. 2. First, following existing work [52], the input fisheye image is fed into the flow estimation module to obtain the appearance flow, which performs a coarse-grained rectification for image features (i.e., warping $\\omega(\\cdot)$ ). Then, the Distortion-aware Learnable Query Mechanism (DLQM) (in Sec. 3.2) extracts control conditions from the user-given query and feeds them layer by layer into the rectification network. Finally, a U-shaped hierarchical network composed of several controllable modulating blocks (in Sec. 3.3) is used to rectify the distorted input image. These blocks modulate features with the control conditions given by DLQM, to get the final output.", + "bbox": [ + 212, + 657, + 787, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the following, we will provide a comprehensive description of each module and its corresponding role in the QueryCDR.", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "QueryCDR: Query-based Controllable Distortion Rectification Network", + "bbox": [ + 254, + 113, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ce06cff48d09e005246042699b655d3fa7a47e92a20038f2ae663b985c4385dd.jpg", + "image_caption": [ + "Fig. 3: Given an input image, QueryCDR can accurately produce results with different rectification degrees by feeding different queries. Moreover, by interpolating between different queries, we can achieve smooth continuous rectification for any distortion degree. For examples, $Q_{1.25} = 0.75Q_1 + 0.25Q_2$ , and $Q_{8.5} = 0.5Q_8 + 0.5Q_9$ ." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 200 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Distortion-aware Learnable Query Mechanism (DLQM)", + "text_level": 1, + "bbox": [ + 214, + 300, + 718, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To achieve control over the rectification process, most existing controllable methods [33,47,55] introduce scalars as control conditions to represent the degradation degrees for image reconstruction. However, due to the peculiarities of distortion in fisheye images [10, 42, 53], these methods fail to achieve effective control over the rectification process. To address this issue, we propose the DLQM. It maintains a learnable query set, providing diverse effective control conditions for the rectification network. During training, DLQM projects the learned high-dimensional positional mapping relationship into a low-dimensional latent space of queries. During inference, DLQM extracts the control information from the queries and converts it into corresponding control conditions for each layer of the rectification network. Simply by providing different queries, DLQM can effectively control the rectification process, as shown in Fig. 3. At the same time, our QueryCDR can also achieve smooth continuous rectification for any distortion degrees by simply interpolating between different queries.", + "bbox": [ + 212, + 329, + 787, + 540 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, we construct a query set $\\mathbf{Q}_s = \\{\\bar{Q}_i\\mid Q_i\\in \\mathbb{R}^{C_{in}\\times H_{in}\\times W_{in}},i = 1,2,\\ldots ,N\\}$ , which comprises $N$ queries representing different distortion degrees. $H_{in}$ , $W_{in}$ , and $C_{in}$ represent the query's height, width, and channel, respectively. Each query has the same size as the input image $I_{in}\\in \\mathbb{R}^{C_{in}\\times H_{in}\\times W_{in}}$ . During inference, users select a query $Q_{i}$ from $\\mathbf{Q}_s$ and feed it into the DLQM. In DLQM, the control extracting part $\\mathrm{CE}(\\cdot)$ is firstly used to extract the features in $Q_{i}$ . The input processing part can be expressed as follows,", + "bbox": [ + 214, + 539, + 787, + 647 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {e x} = \\operatorname {C E} \\left(Q _ {i}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 444, + 660, + 785, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $Q_{ex} \\in \\mathbb{R}^{C_{in} \\times H_{in} \\times W_{in}}$ is the extracted feature, and serves as the input to the first control layer of DLQM. To minimize the computational costs and effectively extract the control information, $\\mathrm{CE}(\\cdot)$ is composed of three convolution layers with a kernel size of $3 \\times 3$ .", + "bbox": [ + 212, + 688, + 787, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Subsequently, DLQM provides corresponding control conditions to the respective layers of the rectification network. To provide appropriate control conditions, each control layer of DLQM comprises two fully connected layers $\\mathrm{FC}_1(\\cdot)$ , $\\mathrm{FC}_2(\\cdot)$ . The $l$ -th layer of DLQM can be represented as follows,", + "bbox": [ + 212, + 750, + 787, + 811 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {c} ^ {l} = \\mathrm {F C} _ {2} ^ {l} \\left(\\mathrm {F C} _ {1} ^ {l} \\left(Q _ {c} ^ {l - 1}\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 415, + 821, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "P. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $Q_{c}^{l-1}$ is the output control condition from the $(l-1)$ -th control layer. For $l = 1$ , $Q_{c}^{0}$ represents the extracted feature $Q_{ex}$ . The output $Q_{c}^{l}$ of the $l$ -th control layer is fed into the $l$ -th rectification layer as a control condition, and simultaneously serves as input to the $(l+1)$ -th control layer of DLQM.", + "bbox": [ + 212, + 145, + 782, + 205 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Afterward, we feed the control conditions into the rectification network. Each layer's controllable modulating block modulates the input feature with the control condition, guiding the modulated features toward the desired direction.", + "bbox": [ + 212, + 207, + 784, + 252 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Controllable Modulating Block", + "text_level": 1, + "bbox": [ + 214, + 273, + 516, + 289 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "When controlling the rectification process, to avoid obtaining results with blurred texture details or residual distortions, we introduce two types of controllable modulating blocks tailored for reconstructing local texture details and learning continuous distortion patterns. They are denoted as the Controllable Convolution Modulating Block (CCMB) based on CNN [17] (in Fig. 2(a)) and the Controllable Attention Modulating Block (CAMB) based on Transformer [45] (in Fig. 2(b)). The CCMB can adaptively fuse the original features and the controlled features, preserving more local details. The CAMB can better capture the spatial distortion information, guaranteeing the integrity of the recovered content [10,35,53]. To balance performance and computational cost, our QueryCDR constructs a U-shaped rectification network [52] composed of CCMB and CAMB. In the first three layers with larger feature maps, i.e., $l = \\{1,2,3,9,10,11\\}$ , CCMB is employed to learn more local texture details. In the remaining layers with smaller feature maps, i.e., $l = \\{4,5,6,7,8\\}$ , CAMB is utilized to capture more global dependencies within the images.", + "bbox": [ + 212, + 299, + 784, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Controllable Convolution Modulating Block (CCMB). When incorporating control conditions into the rectification network, existing methods [47,55] fail to effectively balance the fusion ratio between original features and controlled features. Directly using the controlled features or fusing them with original features at a fixed ratio will degrade the quality of the rectified images. To harmoniously incorporate control conditions into the rectification process, we design the CCMB, as shown in Fig. 2(a). CCMB can dynamically find an optimal ratio to modulate the distorted features with control conditions.", + "bbox": [ + 212, + 546, + 782, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, CCMB receives an input feature $F_{in} \\in \\mathbb{R}^{C \\times H \\times W}$ and a control condition $Q_{c} \\in \\mathbb{R}^{C \\times H \\times W}$ . Both of them are used to predict the fusion ratio. Here we omit the layer information for brevity,", + "bbox": [ + 212, + 667, + 782, + 712 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\theta = \\operatorname {C P} \\left(F _ {i n}, Q _ {c}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 439, + 726, + 784, + 741 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathrm{CP}(\\cdot)$ is the coefficient predictor composed of two fully connected layers, takes the concatenation of $F_{in}$ and $Q_{c}$ as input and predicts the fusion ratio $\\theta$ .", + "bbox": [ + 212, + 752, + 782, + 781 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Then, we perform element-wise multiplication $\\otimes$ on $F_{in}$ and $Q_{c}$ to yield the controlled features $F_{c} \\in \\mathbb{R}^{C \\times H \\times W}$ , which can be expressed as,", + "bbox": [ + 212, + 782, + 782, + 811 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nF _ {c} = F _ {i n} \\otimes Q _ {c}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 447, + 825, + 784, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "QueryCDR: Query-based Controllable Distortion Rectification Network", + "bbox": [ + 254, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, the input features $F_{in}$ and controlled features $F_{c}$ are combined in a weighted sum according to the fusion ratio $\\theta$ , illustrated as $\\psi(\\cdot)$ in Fig. 2(a),", + "bbox": [ + 215, + 146, + 787, + 176 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nF _ {o u t} = \\psi \\left(F _ {c}, F _ {i n}, \\theta\\right) = \\theta F _ {c} \\oplus (1 - \\theta) F _ {i n}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 185, + 785, + 203 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\oplus$ represents the element-wise addition, and $F_{out} \\in \\mathbb{R}^{C \\times H \\times W}$ is the final output of CCMB.", + "bbox": [ + 212, + 210, + 785, + 239 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Due to this dynamic modulation mechanism, CCMB achieves effective control over the rectification process while preserving richer texture details.", + "bbox": [ + 212, + 242, + 785, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Controllable Attention Modulating Block (CAMB). Due to the lack of perception of global information, CNN-based networks struggle to learn long-range dependencies, particularly the continuous and amorphous distortions prevalent in fisheye distortions [10, 42, 53]. In contrast, Transformer-based networks effectively compensate for this, with their global attention mechanism [13, 35].", + "bbox": [ + 212, + 292, + 787, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Therefore, we propose the CAMB, as illustrated in Fig. 2(b). To optimally leverage the control conditions given by DLQM, we designed the control-attention mechanism, enabling CAMB to perceive the global spatial relationships in the control conditions effectively. Specifically, we unfold and project the controlled feature $F_{c}$ (in Eq. (4)) as the query $\\mathcal{Q} \\in \\mathbb{R}^{m \\times L}$ , with the input feature $F_{in}$ as the key $\\mathcal{K} \\in \\mathbb{R}^{m \\times L}$ and value $\\mathcal{V} \\in \\mathbb{R}^{m \\times L}$ , where $L = H \\times W$ represents the sequence length, and $m$ denotes the dimensions of the sequences. The control-attention is described by the following,", + "bbox": [ + 212, + 368, + 787, + 489 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {Q} = W _ {\\mathcal {Q}} F _ {c}, \\mathcal {K} = W _ {\\mathcal {K}} F _ {i n}, \\mathcal {V} = W _ {\\mathcal {V}} F _ {i n}, \\\\ \\operatorname {C T R L - A T T N} (\\mathcal {Q}, \\mathcal {K}, \\mathcal {V}) = \\operatorname {s o f t m a x} \\left(\\frac {\\mathcal {Q} \\mathcal {K} ^ {T}}{\\sqrt {m}}\\right) \\mathcal {V}, \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 498, + 784, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $W_{\\mathcal{Q}}, W_{\\mathcal{K}}, W_{\\mathcal{V}} \\in \\mathbb{R}^{m \\times C}$ represent the projection matrices of the queries, keys, and values, respectively. CTRL-ATTN( $\\cdot$ ) represents the control-attention we proposed. The overall computation of CAMB can be formulated as follows,", + "bbox": [ + 212, + 559, + 785, + 606 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} F _ {a} = \\operatorname {C T R L - A T T N} (\\mathrm {L N} (\\mathcal {Q}, \\mathcal {K}, \\mathcal {V})) \\oplus F _ {i n}, \\\\ F _ {o u t} = \\operatorname {C o n v} _ {1 \\times 1} (\\operatorname {F F N} (\\operatorname {L N} (F _ {a})) \\oplus F _ {a}), \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 614, + 784, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $F_{in}$ means the input feature, $F_{a} \\in \\mathbb{R}^{m \\times H \\times W}$ and $F_{out} \\in \\mathbb{R}^{C \\times H \\times W}$ are the outputs of CTRL-ATTN( $\\cdot$ ) and CAMB, respectively. $\\mathrm{LN}(\\cdot)$ denotes the layer normalization [1]. $\\mathrm{FFN}(\\cdot)$ stands for the feed-forward network composed of three fully connected layers, which helps CAMB to focus on the global dependencies.", + "bbox": [ + 212, + 659, + 785, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CAMB can discern the global mapping relationships within fisheye images, ensuring rectification uniformity compared to the CNN-based networks.", + "bbox": [ + 212, + 720, + 785, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Training Strategy", + "text_level": 1, + "bbox": [ + 214, + 771, + 408, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To guarantee the robust and stable controllable distortion rectification of Query-CDR, we design a two-stage training strategy that combines coarse-grained distortion pre-training and fine-grained distortion fine-tuning.", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "P. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During the coarse-grained distortion pre-training phase, we choose the most commonly used dataset [10,24,52,54] that only contains one degree of distortion to train our QueryCDR. For clarity, we denote the distortion degree in this phase as $d$ . Correspondingly, only one query, denoted as $Q$ , is used for training. The optimization objective can be expressed as follows,", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p r e} = \\mathcal {L} _ {r} + \\mathcal {L} _ {m}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 441, + 232, + 784, + 247 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{pre}$ is the overall loss function for the pre-training phase. $\\mathcal{L}_r$ denotes the reconstruction loss,", + "bbox": [ + 212, + 253, + 782, + 282 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r} = \\left\\| I _ {o u t} ^ {d} - I _ {g t} ^ {d} \\right\\| _ {1}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 429, + 282, + 784, + 303 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $I_{out}^{d}$ and $I_{gt}^{d}$ signifie the output result and ground truth, respectively. $\\mathcal{L}_m$ denotes the multi-scale loss,", + "bbox": [ + 212, + 306, + 782, + 338 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m} = \\sum_ {j = 1} ^ {Z - 1} \\| S \\left(I _ {g t} ^ {d}, j\\right) - C \\left(F _ {\\text {o u t}} ^ {j}\\right) \\| _ {1}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 345, + 784, + 388 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $S(\\cdot)$ represents the operation that down-samples the input $I_{gt}^{d}$ by a factor of $1/2^{j}$ . $Z$ represents the number of decoder's layers, and we set $Z$ to 6. $F_{out}^{j}$ denotes the feature in $j$ -th decoder layer. $C(\\cdot)$ is $3 \\times 3$ convolution for decoding the features into 3-channel RGB images. In this way, each feature map on the decoder can be effectively supervised. The $I_{out}^{d}$ can be obtained as,", + "bbox": [ + 212, + 396, + 784, + 476 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nI _ {o u t} ^ {d} = \\operatorname {Q u e r y C D R} \\left(I _ {i n} ^ {d}, Q\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 484, + 784, + 501 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $I_{in}^{d}$ represents fisheye images with distortion degree $d$ as input. This way can effectively boost the model stability and accelerate the convergence.", + "bbox": [ + 212, + 507, + 782, + 537 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "During the fine-grained distortion fine-tuning phase, we use the varying distortion degrees datasets to fine-tune our QueryCDR. Before training, We replicate the weight of the query $Q$ to the other queries to accelerate convergence on other distortion degrees. Subsequently, we fine-tune our QueryCDR using fisheye images with varying distortion degrees, and feeding corresponding queries into QueryCDR for training at the same time. This allows the query set to efficiently acquire diverse latent spatial relationships. The optimization objective can be expressed as follows,", + "bbox": [ + 212, + 539, + 784, + 657 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {f i n e}} ^ {d _ {i}} = \\mathcal {L} _ {r} ^ {d _ {i}} + \\mathcal {L} _ {m} ^ {d _ {i}}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 434, + 657, + 784, + 676 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{fine}^{d_i}$ is the overall fine-tuning loss function for distortion degree $d_{i}, i \\in \\{1,2,\\dots,9\\}$ , which is similar to the $\\mathcal{L}_{pre}$ but calculated across different distortion degrees. The $I_{out}^{d_i}$ can be obtained as,", + "bbox": [ + 214, + 681, + 782, + 731 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\text {o u t}} ^ {d _ {i}} = \\operatorname {Q u e r y C D R} \\left(I _ {\\text {i n}} ^ {d _ {i}}, Q _ {i}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 405, + 738, + 784, + 755 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $I_{in}^{d_i}$ denotes fisheye images with a distortion degree of $d_i$ as input, $Q_i$ denotes the query corresponding to $d_i, i \\in \\{1, 2, \\dots, 9\\}$ .", + "bbox": [ + 212, + 763, + 782, + 794 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "With this two-stage training strategy, our QueryCDR can effectively utilize only a small amount of varying distortion degrees data that is difficult to obtain, to rapidly finish the training of our network.", + "bbox": [ + 212, + 795, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "QueryCDR: Query-based Controllable Distortion Rectification Network", + "bbox": [ + 254, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/b3a3911e45c37d1c5551a67707ae878006fe6b4bbe061533d6f8e867cd285b29.jpg", + "table_caption": [ + "Table 1: Quantitative comparison (PSNR (dB)↑, SSIM ↑) on COCO [27] fisheye image dataset with varying distortion degrees. Red indicates the best and blue indicates the second best performance (best viewed in color)." + ], + "table_footnote": [], + "table_body": "
MethodPSNR
d1d2d3d4d5d6d7d8d9Avg
SC [40]10.0511.5911.8111.0311.5010.1911.1110.199.1410.73
DeepCalib [4]10.0010.6911.0111.1911.2811.4611.4511.1311.1111.04
Blind [23]12.9811.2410.309.629.9911.7512.6912.7512.8011.57
DR-GAN [24]15.6817.4017.9718.3418.5018.4417.9517.9417.4717.74
PCN [52]14.9317.4318.4318.8618.8618.8818.7417.3518.2617.97
DDA [54]16.3917.4117.4319.4820.1218.9018.8518.1718.2218.33
SimFIR [10]16.5717.8818.4318.9719.3119.2819.1918.6518.4818.53
QueryCDR20.0120.2920.3920.4120.7220.8120.5819.1120.5320.32
MethodSSIM
d1d2d3d4d5d6d7d8d9Avg
SC [40]0.1010.1130.1490.1820.2830.1750.1410.1260.0930.151
DeepCalib [4]0.1840.2100.2230.2300.2340.2460.2500.2450.2460.229
Blind [23]0.3080.2440.1990.1760.1940.2960.3670.3950.4200.289
DR-GAN [24]0.2950.3300.3390.3440.3440.3320.3140.3120.2990.323
PCN [52]0.4200.5470.5890.6070.6080.6100.6150.5760.6030.575
DDA [54]0.4550.5890.5920.6200.6750.6260.6190.5640.5810.591
SimFIR [10]0.4920.5810.6260.6350.6400.6280.6220.5910.5950.601
QueryCDR0.6430.6650.6680.6770.6880.6990.6920.6560.6930.676
", + "bbox": [ + 256, + 198, + 746, + 530 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 215, + 553, + 366, + 569 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experimental Settings", + "text_level": 1, + "bbox": [ + 215, + 580, + 444, + 597 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of our proposed QueryCDR, we followed the existing works [10, 24, 52, 54], employed the four-parameter polynomial model to synthesize fisheye images. We constructed synthetic datasets based on the original images of COCO [27] and Places2 [62] datasets, respectively. Specifically, for images with distortion degree $d$ , 40,000 images were used for the pre-training stage. And for images with varying distortion degrees $d_{i}$ , $i \\in \\{1,2,\\dots ,9\\}$ , 18,000 images were used for the fine-tuning stage, and 3,600 images for testing. The images were resized to $256 \\times 256$ when fed into the network. For a fair comparison, we followed existing works [10, 24, 52, 54] utilizing a batch size of 16 and the Adam [19] optimizer with a learning rate of 1e-4.", + "bbox": [ + 212, + 602, + 787, + 755 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Performance Evaluation", + "text_level": 1, + "bbox": [ + 215, + 772, + 460, + 787 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To evaluate the performance of our method, we retrained and validated existing fisheye image rectification methods on synthetic fisheye datasets with 9 distortion degrees. The methods can be summarized into three categories: traditional", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "P. Guo et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/1fe983ba68f6777d12ebc469e5cfda3a4141d0d3ad7cef8eedd2c622d3b9a149.jpg", + "image_caption": [ + "Fig. 4: Qualitative results on synthetic fisheye images." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 588 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "method SC [40], regression-based methods DeepCalib [4] and Blind [23], and generation-based DR-GAN [24], PCN [52], DDA [54] and SimFIR [10]. For a fair comparison, we followed existing works [10,24,52,54] employing Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) to quantify performance.", + "bbox": [ + 212, + 643, + 787, + 705 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Quantitative Results. Our performance comparisons are shown in Tab. 1. Compared to existing methods, QueryCDR achieves the best performance across all distortion degrees without retraining. It is because DLQM effectively utilizes the control information in queries to guide the network in achieving varying degrees of rectification. Additionally, the capability of CCMB and CAMB allows for controlled rectification while obtaining high-quality results. Moreover, we observe that even on images with distortion degree $d_{5}$ , which the existing", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "QueryCDR: Query-based Controllable Distortion Rectification Network", + "bbox": [ + 256, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6378eef9c633bf24a573b34d50f19ace4590a5bce813bf929c82cc2d99278a98.jpg", + "image_caption": [ + "Fig. 5: Qualitative results on real-world fisheye images." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 262 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "methods perform well [10, 24, 52, 54], QueryCDR still outperforms the second best method [54] by 0.60 dB of PSNR and 0.013 of SSIM. This further demonstrates the outstanding generalization ability of QueryCDR.", + "bbox": [ + 212, + 314, + 784, + 359 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qualitative Results. To further compare the visual qualities of different methods, we show the results rectified by QueryCDR and other rectification methods in Fig. 4. It can be observed intuitively that other methods fail to effectively rectify images with varying degrees of distortion. Particularly the objects in images, our QueryCDR maintains the accurate structure of objects across different distortions. Furthermore, benefiting from the modulation capability of CCMB and CAMB, QueryCDR preserves richer texture details after rectification.", + "bbox": [ + 212, + 381, + 784, + 486 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In addition, to further demonstrate the generalization ability of QueryCDR, we conducted experiments on real-world fisheye image datasets [57] in Fig. 5. Despite the disparities between synthetic and real-world datasets, our QueryCDR still shows robust rectification capabilities, and preservers richer texture details. These results further validate its practicality in real-world scenarios.", + "bbox": [ + 212, + 487, + 784, + 561 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 584, + 388, + 599 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this section, we conduct ablation for different control conditions, and study the effect of the controllable modulating blocks and the architecture setting.", + "bbox": [ + 212, + 607, + 784, + 638 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Distortion-aware Learnable Query Mechanism (DLQM). To validate the effectiveness of DLQM in controlling the rectification process, we implemented different rectification networks controlled by scalar, fixed query, and our learnable query in Tab. 2. When using scalar to control, the value of the scalar increments with the distortion of the entire image. When using fixed query to control, the parameters of the query are set to incremental values from 0 to 1 with the degree of distortion from the center to the edges. Following the settings in Sec. 4.1, we use the learnable query $Q_{i}, i \\in \\{1,2,\\dots ,9\\}$ to learn the distortion distribution corresponding to $d_{i}, i \\in \\{1,2,\\dots ,9\\}$ .", + "bbox": [ + 212, + 657, + 785, + 794 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We trained different methods using the same strategy as illustrated in Sec. 3.4, the results are shown in Tab. 2 and Fig. 6. All three controllable methods outperform the uncontrollable method significantly at various degrees of distortion.", + "bbox": [ + 212, + 795, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "P. Guo et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/73a291ec711c09b0aa102fa71c9817f5a5b244a1d68f7f295bd2f2e162a0d0d1.jpg", + "table_caption": [ + "Table 2: Performance comparison of different control methods." + ], + "table_footnote": [], + "table_body": "
MethodPSNR
d1d2d3d4d5d6d7d8d9Avg
W/o Control14.9317.4318.4318.8618.8618.8818.7417.3518.2617.97
Scalar19.5220.2519.9719.4319.8920.0720.0318.7620.1519.78
Fixed Query20.1320.0320.1419.8820.1620.2520.1618.7419.8419.93
Learnable Query20.0120.2920.3920.4120.7220.8120.5819.1120.5320.32
", + "bbox": [ + 220, + 181, + 785, + 284 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d922fd891d1c1c34348f1b1205873cb07c3cb862fbbf9aeb9cf7a4a7ca9b1350.jpg", + "image_caption": [ + "Fig. 6: Visual results of different control mechanisms." + ], + "image_footnote": [], + "bbox": [ + 223, + 311, + 444, + 405 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/332ab2ce251515723329b0833b584c06fb71ed31c35b01002ee8aca053da6309.jpg", + "image_caption": [ + "Fig.7: Visual results of different modulation methods." + ], + "image_footnote": [], + "bbox": [ + 450, + 311, + 777, + 405 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "This is because the controllable mechanism effectively assists the rectification network in distinguishing between different degrees of distortion, thereby enhancing the generalization capability of the rectification network. The improvement brought by scalar demonstrates the crucial role of controllable mechanisms in handling fisheye images with different degrees of distortion. The further enhancement with fixed query validates the importance of employing higher-dimensional control conditions for fisheye image rectification. Lastly, our tailored learnable query significantly outperforms all other controllable methods, verifying the unique superiority of our approach in controllable rectification.", + "bbox": [ + 212, + 476, + 787, + 612 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Controllable Modulating Block. To demonstrate the capability of the CCMB and CAMB in modulating input features with control conditions, we implemented two comparison methods with different modulating approaches: one directly uses the controlled feature (in Eq. (4)), and the other adds controlled feature and original feature in a fixed 1:1 ratio. We constructed different controllable networks using each of these methods.", + "bbox": [ + 212, + 642, + 787, + 733 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "As shown in Tab. 3 and Fig. 7, retaining original features and fusing them with a 1:1 ratio improved by $0.03\\mathrm{dB}$ , highlighting the importance of preserving original features. The dynamic modulation mechanism boosted performance by $0.06\\mathrm{dB}$ , verifying its effectiveness in dynamic feature modulation. Furthermore, leveraging the control-attention mechanism led to a performance improvement of $0.12\\mathrm{dB}$ , demonstrating the superiority of CAMB for perceiving global distortion distribution, and achieving better uniform global rectification.", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "QueryCDR: Query-based Controllable Distortion Rectification Network", + "bbox": [ + 254, + 113, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/1b32ce70ebb36fac3e942378c48aad45f2da0a516e0ebba1ff5df72cbc9d33ad.jpg", + "table_caption": [ + "Table 3: Ablation study of different modulation methods. Where Direct Use $F_{c}$ means directly using the controlled feature $F_{c}$ , Fixed Ratio means adding controlled features and original features in a 1:1 ratio, Dynamic Mechanism means the CCMB and Attention Mechanism means the CAMB." + ], + "table_footnote": [], + "table_body": "
MethodPSNRSSIM
Direct UseFc20.140.655
Fixed Ratio20.170.658
Dynamic Mechanism20.200.669
Attention Mechanism20.260.671
", + "bbox": [ + 228, + 267, + 470, + 354 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/736db10e4f78782afcbdfbadba4c7658601b80fcc769b2eb85c69cb17c747330.jpg", + "table_caption": [ + "Table 4: Ablation study of different network architectures. Where $x\\mathrm{C} + \\left( {{11} - x}\\right) \\mathrm{A}$ denotes that the network uses CCMB in the $x$ layers with larger feature maps and CAMB in the remaining ${11} - x$ layers with smaller feature maps." + ], + "table_footnote": [], + "table_body": "
MethodFlops(G)Param(M)PSNRSSIM
PCN [52]12.30535.63717.970.575
11C+0A12.73637.70120.200.669
8C+3A13.38346.39820.270.665
6C+5A12.35343.24420.320.676
4C+7A12.53846.99420.310.670
0C+11A15.19051.79520.260.671
", + "bbox": [ + 500, + 239, + 795, + 354 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Controllable Rectification Network Architecture. As described in Sec. 3.3, rectification networks based on CCMB may ignore the long-range dependencies, while those based on CAMB may ignore the texture details and also increase computational costs. To strike a balance between these two architectures, We incorporate CCMB into the layers with larger feature map size, and integrate CAMB into the layers with smaller feature map size. To validate the effectiveness of this hybrid architecture and find an optimal combination, we conducted performance and computational cost comparisons between pure CCMB, pure CAMB, and multiple hybrid networks, as shown in Tab. 4. Pure CAMB network (i.e., Row 6) outperforms pure CCMB network (i.e., Row 2) in rectification performance but incurs higher computational costs. Hybrid network architectures effectively solve this problem, with higher performance and fewer parameters. After a trade-off between performance and parameters, we empirically choose the $6\\mathrm{C} + 5\\mathrm{A}$ hybrid architecture as our final model.", + "bbox": [ + 212, + 388, + 789, + 601 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 640, + 359, + 656 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we propose the Query-based Controllable Distortion Rectification network for fisheye images (QueryCDR), which achieves controllable rectification at different distortion degrees without retraining. In particular, we design a series of learnable queries as control conditions to guide the rectification process. Additionally, we design two different controllable modulating blocks, achieving controllable rectification while improving image quality. Extensive experiments have demonstrated the robustness and effectiveness of our QueryCDR. In the future, it is expected to further rectify distortions of different degrees by autocontrolling mechanisms to avoid human involvement. We believe our work provides an effective solution for fisheye camera applications.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "P. Guo et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported in part by the NSFC under Grant 62272376, 62272380 and 62103317, the Key Research and Development Program of Shaanxi Province under Grant 2020ZDLGY04-05, the Fundamental Research Funds for the Central Universities, China (xzy022023051), the Innovative Leading Talents Scholarship of Xi'an Jiaotong University (Corresponding author: Xingsong Hou).", + "bbox": [ + 212, + 175, + 787, + 252 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 273, + 321, + 287 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016)", + "2. Barreto, J.P., Araujo, H.: Geometric properties of central catadioptric line images and their application in calibration. IEEE Transactions on Pattern Analysis and Machine Intelligence 27(8), 1327-1333 (2005)", + "3. Barreto, J.P., Daniilidis, K.: Fundamental matrix for cameras with radial distortion. In: Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1. vol. 1, pp. 625-632. IEEE (2005)", + "4. Bogdan, O., Eckstein, V., Rameau, F., Bazin, J.C.: Deepcalib: A deep learning approach for automatic intrinsic calibration of wide field-of-view cameras. In: Proceedings of the 15th ACM SIGGRAPH European Conference on Visual Media Production. pp. 1-10 (2018)", + "5. Bukhari, F., Dailey, M.N.: Automatic radial distortion estimation from a single image. Journal of mathematical imaging and vision 45, 31-45 (2013)", + "6. Cai, B., Xu, X., Jia, K., Qing, C., Tao, D.: Dehazenet: An end-to-end system for single image haze removal. IEEE transactions on image processing 25(11), 5187-5198 (2016)", + "7. Cai, H., He, J., Qiao, Y., Dong, C.: Toward interactive modulation for photorealistic image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 294-303 (2021)", + "8. Devernay, F., Faugeras, O.: Straight lines have to be straight. Machine vision and applications 13, 14-24 (2001)", + "9. Duan, Z., Tezcan, O., Nakamura, H., Ishwar, P., Konrad, J.: Rapid: rotation-aware people detection in overhead fisheye images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 636-637 (2020)", + "10. Feng, H., Wang, W., Deng, J., Zhou, W., Li, L., Li, H.: Simfir: A simple framework for fisheye image rectification with self-supervised representation learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12418-12427 (2023)", + "11. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: 2012 IEEE conference on computer vision and pattern recognition. pp. 3354-3361. IEEE (2012)", + "12. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020)", + "13. Guo, J., Han, K., Wu, H., Tang, Y., Chen, X., Wang, Y., Xu, C.: Cmt: Convolutional neural networks meet vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12175-12185 (2022)" + ], + "bbox": [ + 218, + 301, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "QueryCDR: Query-based Controllable Distortion Rectification Network", + "bbox": [ + 254, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Hartley, R., Kang, S.B.: Parameter-free radial distortion correction with center of distortion estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence 29(8), 1309-1321 (2007)", + "15. He, J., Dong, C., Qiao, Y.: Modulating image restoration with continual levels via adaptive feature modification layers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11056-11064 (2019)", + "16. He, J., Dong, C., Qiao, Y.: Interactive multi-dimension modulation with dynamic controllable residual learning for image restoration. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16. pp. 53-68. Springer (2020)", + "17. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "18. Henrique Brito, J., Angst, R., Koser, K., Pollefeys, M.: Radial distortion self-calibration. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1368-1375 (2013)", + "19. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)", + "20. Kukelova, Z., Pajdla, T.: A minimal solution to radial distortion autocalibration. IEEE Transactions on Pattern Analysis and Machine Intelligence 33(12), 2410-2422 (2011)", + "21. Kumar, V.R., Klingner, M., Yogamani, S., Milz, S., Fingscheidt, T., Mader, P.: Syndistnet: Self-supervised monocular fisheye camera distance estimation synergized with semantic segmentation for autonomous driving. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 61-71 (2021)", + "22. Ledig, C., Theis, L., Huszár, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., et al.: Photo-realistic single image superresolution using a generative adversarial network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4681-4690 (2017)", + "23. Li, X., Zhang, B., Sander, P.V., Liao, J.: Blind geometric distortion correction on images through deep learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4855-4864 (2019)", + "24. Liao, K., Lin, C., Zhao, Y., Gabbouj, M.: Dr-gan: Automatic radial distortion rectification using conditional gan in real-time. IEEE Transactions on Circuits and Systems for Video Technology 30(3), 725-733 (2019)", + "25. Liao, K., Lin, C., Zhao, Y., Xu, M.: Model-free distortion rectification framework bridged by distortion distribution map. IEEE Transactions on Image Processing 29, 3707-3718 (2020)", + "26. Lin, L., Lu, Y., Pan, Y., Chen, X.: Integrating graph partitioning and matching for trajectory analysis in video surveillance. IEEE transactions on Image Processing 21(12), 4844-4857 (2012)", + "27. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "28. Liu, C., Wang, X., Li, S., Wang, Y., Qian, X.: Fsi: Frequency and spatial interactive learning for image restoration in under-display cameras. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12537-12546 (2023)", + "29. Liu, C., Wang, X., Xu, X., Tian, R., Li, S., Qian, X., Yang, M.H.: Motion-adaptive separable collaborative filters for blind motion deblurring. In: Proceedings of the" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "P. Guo et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 25595-25605 (2024)", + "30. Liu, C., Yang, H., Fu, J., Qian, X.: Learning trajectory-aware transformer for video super-resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5687-5696 (2022)", + "31. Liu, C., Yang, H., Fu, J., Qian, X.: 4d lut: learnable context-aware 4d lookup table for image enhancement. IEEE Transactions on Image Processing 32, 4742-4756 (2023)", + "32. Mei, C., Rives, P.: Single view point omnidirectional camera calibration from planar grids. In: Proceedings 2007 IEEE International Conference on Robotics and Automation. pp. 3945-3950. IEEE (2007)", + "33. Mou, C., Wu, Y., Wang, X., Dong, C., Zhang, J., Shan, Y.: Metric learning based interactive modulation for real-world super-resolution. In: European Conference on Computer Vision. pp. 723-740. Springer (2022)", + "34. Muhammad, K., Ahmad, J., Lv, Z., Bellavista, P., Yang, P., Baik, S.W.: Efficient deep cnn-based fire detection and localization in video surveillance applications. IEEE Transactions on Systems, Man, and Cybernetics: Systems 49(7), 1419-1434 (2018)", + "35. Peng, Z., Huang, W., Gu, S., Xie, L., Wang, Y., Jiao, J., Ye, Q.: Conformer: Local features coupling global representations for visual recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 367-376 (2021)", + "36. Plaut, E., Ben Yaacov, E., El Shlomo, B.: 3d object detection from a single fisheye image without a single fisheye training image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3659-3667 (2021)", + "37. Puig, L., Bastanlar, Y., Sturm, P., Guerrero, J.J., Barreto, J.: Calibration of central catadioptric cameras using a dlt-like approach. International Journal of Computer Vision 93, 101-114 (2011)", + "38. Rashed, H., Mohamed, E., Sistu, G., Kumar, V.R., Eising, C., El-Sallab, A., Yogamani, S.: Generalized object detection on fisheye cameras for autonomous driving: Dataset, representations and baseline. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2272-2280 (2021)", + "39. Rong, J., Huang, S., Shang, Z., Ying, X.: Radial lens distortion correction using convolutional neural networks trained with synthesized images. In: Computer Vision-ACCV 2016: 13th Asian Conference on Computer Vision, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part III 13. pp. 35-49. Springer (2017)", + "40. Santana-Cedrés, D., Gomez, L., Aleman-Flores, M., Salgado, A., Esclarín, J., Mazorra, L., Alvarez, L.: An iterative optimization algorithm for lens distortion correction using two-parameter models. Image Processing On Line 6, 326-364 (2016)", + "41. Scaramuzza, D., Martinelli, A., Siegwart, R.: A flexible technique for accurate omnidirectional camera calibration and structure from motion. In: Fourth IEEE International Conference on Computer Vision Systems (ICVS'06). pp. 45-45. IEEE (2006)", + "42. Shen, Z., Lin, C., Liao, K., Nie, L., Zheng, Z., Zhao, Y.: Panoformer: Panorama transformer for indoor 360 depth estimation. In: European Conference on Computer Vision. pp. 195-211. Springer (2022)", + "43. Sturm, P., Ramalingam, S.: A generic concept for camera calibration. In: Computer Vision-ECCV 2004: 8th European Conference on Computer Vision, Prague, Czech Republic, May 11-14, 2004. Proceedings, Part II 8. pp. 1-13. Springer (2004)", + "44. Thormahlen, T., Broszio, H., Wassermann, I.: Robust line-based calibration of lens distortion from a single view. Mirage 2003 pp. 105-112 (2003)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "QueryCDR: Query-based Controllable Distortion Rectification Network", + "bbox": [ + 256, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "45. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)", + "46. Wang, A., Qiu, T., Shao, L.: A simple method of radial distortion correction with centre of distortion estimation. Journal of Mathematical Imaging and Vision 35, 165-172 (2009)", + "47. Wang, W., Guo, R., Tian, Y., Yang, W.: Cfsnet: Toward a controllable feature space for image restoration. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4140-4149 (2019)", + "48. Wang, X., Yu, K., Dong, C., Tang, X., Loy, C.C.: Deep network interpolation for continuous imagery effect transition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1692-1701 (2019)", + "49. Wang, X., Yu, K., Wu, S., Gu, J., Liu, Y., Dong, C., Qiao, Y., Change Loy, C.: Esrgan: Enhanced super-resolution generative adversarial networks. In: Proceedings of the European conference on computer vision (ECCV) workshops. pp. 0-0 (2018)", + "50. Xue, Z., Xue, N., Xia, G.S., Shen, W.: Learning to calibrate straight lines for fisheye image rectification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1643-1651 (2019)", + "51. Yang, C.Y., Chen, H.H.: Efficient face detection in the fisheye image domain. IEEE Transactions on Image Processing 30, 5641-5651 (2021)", + "52. Yang, S., Lin, C., Liao, K., Zhang, C., Zhao, Y.: Progressively complementary network for fisheye image rectification using appearance flow. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6348-6357 (2021)", + "53. Yang, S., Lin, C., Liao, K., Zhao, Y.: Fishformer: Annulus slicing-based transformer for fisheye rectification with efficacy domain exploration. arXiv preprint arXiv:2207.01925 (2022)", + "54. Yang, S., Lin, C., Liao, K., Zhao, Y.: Dual diffusion architecture for fisheye image rectification: Synthetic-to-real generalization. arXiv preprint arXiv:2301.11785 (2023)", + "55. Yao, M., He, D., Li, X., Li, F., Xiong, Z.: Towards interactive self-supervised denoising. IEEE Transactions on Circuits and Systems for Video Technology (2023)", + "56. Yin, X., Wang, X., Yu, J., Zhang, M., Fua, P., Tao, D.: Fisheyerecnet: A multicontext collaborative deep network for fisheye image rectification. In: Proceedings of the European conference on computer vision (ECCV). pp. 469-484 (2018)", + "57. Yogamani, S., Hughes, C., Horgan, J., Sistu, G., Varley, P., O'Dea, D., Uricár, M., Milz, S., Simon, M., Amende, K., et al.: Woodscape: A multi-task, multicamera fisheye dataset for autonomous driving. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9308-9318 (2019)", + "58. Zhang, K., Zuo, W., Chen, Y., Meng, D., Zhang, L.: Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing 26(7), 3142-3155 (2017)", + "59. Zhang, M., Yao, J., Xia, M., Li, K., Zhang, Y., Liu, Y.: Line-based multi-label energy optimization for fisheye image rectification and calibration. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4137-4145 (2015)", + "60. Zhang, Y., You, S., Gevers, T.: Automatic calibration of the fisheye camera for egocentric 3d human pose estimation from a single image. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1772-1781 (2021)" + ], + "bbox": [ + 217, + 147, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "P. Guo et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "61. Zhang, Z., Jiang, Y., Shao, W., Wang, X., Luo, P., Lin, K., Gu, J.: Real-time controllable denoising for image and video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14028-14038 (2023)", + "62. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence 40(6), 1452-1464 (2017)" + ], + "bbox": [ + 215, + 146, + 787, + 231 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "QueryCDR: Query-based Controllable Distortion Rectification Network", + "bbox": [ + 256, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/QueryCDR_ Query-based Controllable Distortion Rectification Network for Fisheye Images/8182781d-9f7c-4e2a-86d5-0c2e6b9982a2_model.json b/2024/QueryCDR_ Query-based Controllable Distortion Rectification Network for Fisheye Images/8182781d-9f7c-4e2a-86d5-0c2e6b9982a2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..aa17f34a9bf4f887e3f3a01541f68d77dd423d98 --- /dev/null +++ b/2024/QueryCDR_ Query-based Controllable Distortion Rectification Network for Fisheye Images/8182781d-9f7c-4e2a-86d5-0c2e6b9982a2_model.json @@ -0,0 +1,2537 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.223, + 0.141, + 0.782, + 0.187 + ], + "angle": 0, + "content": "QueryCDR: Query-based Controllable Distortion Rectification Network for Fisheye Images" + }, + { + "type": "text", + "bbox": [ + 0.281, + 0.212, + 0.721, + 0.244 + ], + "angle": 0, + "content": "Pengbo Guo\\(^{1,2}\\), Chengxu Liu\\(^{2,3}\\), Xingsong Hou\\(^{2(\\boxtimes)}\\), and Xueming Qian\\(^{2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.255, + 0.747, + 0.282 + ], + "angle": 0, + "content": "\\(^{1}\\) School of Software Engineering, Xi'an Jiaotong University, Xi'an, China \n\\(^{2}\\) Xi'an Jiaotong University, Xi'an, China" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.283, + 0.776, + 0.311 + ], + "angle": 0, + "content": "3 Shaanxi Yulan Jiuzhou Intelligent Optoelectronic Tech. Co., Ltd, Xi'an, China {guopengbo,chengxuliu}@stu.xjtu.edu.cn, {houxs,qianxm}@mail.xjtu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.346, + 0.744, + 0.653 + ], + "angle": 0, + "content": "Abstract. Fisheye image rectification aims to correct distortions in images taken with fisheye cameras. Although current models show promising results on images with a similar degree of distortion as the training data, they will produce sub-optimal results when the degree of distortion changes and without retraining. The lack of generalization ability for dealing with varying degrees of distortion limits their practical application. In this paper, we take one step further to enable effective distortion rectification for images with varying degrees of distortion without retraining. We propose a novel Query-based Controllable Distortion Rectification network for fisheye images (QueryCDR). In particular, we first present the Distortion-aware Learnable Query Mechanism (DLQM), which defines the latent spatial relationships for different distortion degrees as a series of learnable queries. Each query can be learned to obtain position-dependent rectification control conditions, providing control over the rectification process. Then, we propose two kinds of controllable modulating blocks to enable the control conditions to guide the modulation of the distortion features better. These core components cooperate with each other to effectively boost the generalization ability of the model at varying degrees of distortion. Extensive experiments on fisheye image datasets with different distortion degrees demonstrate our approach achieves high-quality and controllable distortion rectification. Code is available at https://github.com/PbGuo/QueryCDR." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.664, + 0.707, + 0.679 + ], + "angle": 0, + "content": "Keywords: Fisheye image \\(\\cdot\\) Distortion rectification \\(\\cdot\\) Controllable" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.704, + 0.376, + 0.72 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Benefiting from the huge field-of-view (FoV), fisheye cameras are widely utilized in various fields, including security surveillance [26, 34] and autonomous driving [11, 38]. However, the distortion brought by the fisheye lenses greatly limits the performance of downstream vision tasks [9, 21, 36, 60]. How to eliminate the distortion in fisheye images has attracted great attention in recent years." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Early methods [3,5,14,18,37,40,46,59] primarily relied on identifying matching feature points or curves for automatic rectification. However, constrained by" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "P. Guo et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.143, + 0.785, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.293, + 0.788, + 0.378 + ], + "angle": 0, + "content": "Fig. 1: Different approaches to fisheye image distortion rectification. (a) Regression-Based: Using a neural network to predict distortion-related parameters, then apply rectification algorithms \\( R \\) for rectification. (b) Generation-Based: Input the distorted fisheye image and directly generate the rectified image end-to-end. (c) Control-Based: Users provide control conditions to guide the rectification process, resulting in promising rectified images of various distortion degrees." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.788, + 0.548 + ], + "angle": 0, + "content": "the instability of feature detection, both the generalization of the algorithm and the quality of the rectified images were unsatisfactory [10, 24, 52]. In recent years, owing to the robust learning and generalization capabilities of neural networks, deep learning-based fisheye image rectification methods have become mainstream. These methods can be divided into two categories: regression-based rectification [4, 23, 39, 50, 56] and generation-based rectification [10, 24, 25, 52, 54]. The former uses deep regression models to predict distortion parameters for image reconstruction, as shown in Fig. 1(a). The latter uses an encoder-decoder structure to generate well-rectified images directly, as shown in Fig. 1(b)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.788, + 0.717 + ], + "angle": 0, + "content": "It is worth noting that these methods only achieve satisfactory results on a similar degree of distortion as the training data. This means that when handling different degrees of distortion without retraining, there will be a significant decrease in the quality of the rectified images [25]. This is because the model tends to learn fixed position mapping relationships during training. When the distortion degree changes, this relationship does not work for the new distortion distribution. Therefore, these models usually need to be retrained on images with different distortion degrees. In addition, fisheye image acquisition is difficult, and re-collecting new datasets with varying degrees of distortion would incur significant costs [52, 54]. Thus, it is essential to explore a model that can handle different degrees of distortion simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.79, + 0.842 + ], + "angle": 0, + "content": "In recent years, some methods have been proposed to achieve controllable image restoration [7,15,16,33,47,48,55,61]. Typically, CFSNet [47], MM-RealSR [33], and Yao et al. [55] introduce scalars as control conditions to effectively restore degradation of varying degrees. It would be a promising solution to introduce an effectively controllable mechanism to deal with all degrees of distortion, as shown in Fig. 1(c). However, directly applying existing controllable mechanisms in restoration tasks to the fisheye rectification model suffers from the following challenges: 1) There is a gap between optimization objectives (i.e. the roles" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.256, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "QueryCDR: Query-based Controllable Distortion Rectification Network" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "of the controllable mechanisms). Restoration models learn the pixel-level detail restoration, whereas fisheye rectification models learn the spatial-level positional mapping relationships [10,25]. The control mechanism lacking position information in restoration tasks is ineffective in controlling distortion rectification networks. 2) There is a gap between optimization difficulties (i.e. the roles of the control conditions). The distortion of a fisheye image increases gradually from the image center to boundary [51]. Therefore, it is not suitable to use a single scalar control condition in restoration tasks to deal with such spatially varying distortions. These challenges restrict the application of controllable mechanisms and control conditions in fisheye image rectification." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.789, + 0.569 + ], + "angle": 0, + "content": "To address these issues, we propose the Query-based Controllable Distortion Rectification network (QueryCDR), as shown in Fig. 2. By introducing a series of learnable queries as control conditions, QueryCDR allows the users to achieve fisheye rectification with different distortion degrees. Specifically, to incorporate positional mapping relationships into control conditions, we introduce the Distortion-aware Learnable Query Mechanism (DLQM), which defines a series of queries representing different rectification control conditions. During inference, DLQM extracts position-dependent control conditions from the user-given query and feeds them into the network for controlling the rectification process. Furthermore, to enable the control conditions to guide the rectification efficiently, we propose two types of controllable modulating blocks: the Controllable Convolution Modulating Block (CCMB) based on CNN [17], and the Controllable Attention Modulating Block (CAMB) based on Transformer [45]. They are good at extracting local texture features and learning long-range distortion mapping relationships, respectively. By combining CCMB and CAMB, we construct a robust controllable rectification network. Our QueryCDR can handle various distortions without retraining, enhancing the generalization ability of the fisheye rectification model." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.57, + 0.555, + 0.584 + ], + "angle": 0, + "content": "We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.59, + 0.785, + 0.635 + ], + "angle": 0, + "content": "- We propose QueryCDR, a Query-based Controllable Distortion Rectification network for fisheye images. Extensive experiments demonstrate that our QueryCDR can deliver superior results on a variety of distortion degrees." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.636, + 0.785, + 0.679 + ], + "angle": 0, + "content": "- We propose the Distortion-aware Learnable Query Mechanism (DLQM), which effectively introduces the latent spatial relationships to control conditions for fisheye image rectification." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.68, + 0.785, + 0.739 + ], + "angle": 0, + "content": "- We propose two kinds of blocks for modulating features using control conditions: the Controllable Convolution Modulating Block (CCMB) and the Controllable Attention Modulating Block (CAMB). They can effectively utilize control conditions to guide the rectification process." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.59, + 0.785, + 0.739 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.759, + 0.388, + 0.775 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.586, + 0.803 + ], + "angle": 0, + "content": "2.1 Traditional Fisheye Image Rectification" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Traditional rectification methods can be divided into two types, multi-view-based and line-based methods. Multi-view-based methods [3,14,18,20,37,41,43]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.128 + ], + "angle": 0, + "content": "P. Guo et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "calibrated fisheye images by finding corresponding feature points from multiple viewpoints. Line-based methods [2,5,8,32,40,44,46,59] employed line detection to rectify the curved lines, thereby achieving distortion rectification. However, these methods require manual intervention or handcrafted feature extractors, and the rectification process is unstable, failing to achieve satisfactory results." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.244, + 0.673, + 0.26 + ], + "angle": 0, + "content": "2.2 Deep Learning Based Fisheye Image Rectification" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.267, + 0.787, + 0.357 + ], + "angle": 0, + "content": "Rapid advances in deep learning have allowed them to shine in low-level vision tasks [6, 22, 28-31, 49, 58]. As a core task in low-level vision, distortion image rectification has received increasing attention [10, 24, 39, 52, 54, 56]. According to the network architecture, deep learning-based distortion image rectification methods can be categorized into two types, regression-based methods [4, 23, 39, 50, 56] and generation-based methods [10, 24, 25, 52, 54]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.789, + 0.661 + ], + "angle": 0, + "content": "Regression-based methods [4, 23, 39, 50, 56] utilize neural networks to predict distortion-related coefficients for rectifying the image. Rong et al. [39] were the first to use CNNs for fisheye image rectification. They employed the network to predict across multiple distortion intervals, achieving preliminary rectification results. Yin et al. [56] integrated semantics as prior information to guide rectification. While these methods have shown some performance improvement, they face the challenge of non-end-to-end design, thus requiring a large number of additional labels, and increasing operational costs. Therefore, to reduce models' complexity, Generation-based methods [10, 24, 25, 52, 54] employ generative networks to take distorted images as input and directly generate rectified images. DR-GAN [24] was the first to apply a generative adversarial network (GAN) [12] to fisheye image rectification, enabling the network to directly generate rectified images without estimating additional parameters. DDM [25] introduced distortion maps to help the model better learn the distortion distribution. PCN [52] designed a flow estimation module to predict appearance flows in fisheye images, using it to assist the rectification progress. SimFIR [10] introduced a self-supervised rectification module, allowing the network to better learn the distortion representations at different locations. However, both paradigms require retraining when handling images with different distortion degrees, failing to address the issue of weak model generalization." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.681, + 0.506, + 0.696 + ], + "angle": 0, + "content": "2.3 Controllable Low-level Vision" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To address the issue of varying degradation levels in low-level vision tasks, an increasing number of studies [7,15,16,33,47,48,55,61] propose controllable network architectures to tackle this challenge. DNI [48] interpolated all parameters of different restoration networks, which were trained with different degradation levels. By adjusting the interpolation coefficients, a smooth control of the image can be achieved. AdaFM [15] achieved better results by inserting AdaFM layers after each convolution layer to change the filters' statistics, thus the users can interactively manipulate the restoration results by tuning a control coefficient. CFS-Net [47] introduced the tuning branch to adaptively learn the control coefficients," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.256, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "QueryCDR: Query-based Controllable Distortion Rectification Network" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.144, + 0.789, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.362, + 0.788, + 0.46 + ], + "angle": 0, + "content": "Fig. 2: Overview of our proposed Query-based Controllable Distortion Rectification network (QueryCDR). The Distortion-aware Learnable Query Mechanism (DLQM) extracts control conditions from user-given queries and feeds them layer by layer into the rectification network. The rectification network is composed of Controllable Convolution Modulating Blocks (CCMB) and Controllable Attention Modulating Blocks (CAMB), which modulate the input features \\( F_{in} \\) with control conditions \\( F_{c} \\), enabling controllable rectification process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.788, + 0.581 + ], + "angle": 0, + "content": "and then use them to couple the features with the main branch. MM-RealSR [33] proposed a metric learning strategy to map unquantifiable degradation levels to a metric space as control conditions. However, due to the different optimization objectives and complexity of the control mechanisms, directly applying these methods to fisheye image distortion rectification tasks will not achieve effective control over the rectification process." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.603, + 0.381, + 0.621 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.635, + 0.342, + 0.648 + ], + "angle": 0, + "content": "3.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.81 + ], + "angle": 0, + "content": "The overview of our proposed Query-based Controllable Distortion Rectification network (QueryCDR) is shown in Fig. 2. First, following existing work [52], the input fisheye image is fed into the flow estimation module to obtain the appearance flow, which performs a coarse-grained rectification for image features (i.e., warping \\(\\omega(\\cdot)\\)). Then, the Distortion-aware Learnable Query Mechanism (DLQM) (in Sec. 3.2) extracts control conditions from the user-given query and feeds them layer by layer into the rectification network. Finally, a U-shaped hierarchical network composed of several controllable modulating blocks (in Sec. 3.3) is used to rectify the distorted input image. These blocks modulate features with the control conditions given by DLQM, to get the final output." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In the following, we will provide a comprehensive description of each module and its corresponding role in the QueryCDR." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "P. Guo et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.214, + 0.788, + 0.271 + ], + "angle": 0, + "content": "Fig. 3: Given an input image, QueryCDR can accurately produce results with different rectification degrees by feeding different queries. Moreover, by interpolating between different queries, we can achieve smooth continuous rectification for any distortion degree. For examples, \\( Q_{1.25} = 0.75Q_1 + 0.25Q_2 \\), and \\( Q_{8.5} = 0.5Q_8 + 0.5Q_9 \\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.301, + 0.72, + 0.317 + ], + "angle": 0, + "content": "3.2 Distortion-aware Learnable Query Mechanism (DLQM)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.541 + ], + "angle": 0, + "content": "To achieve control over the rectification process, most existing controllable methods [33,47,55] introduce scalars as control conditions to represent the degradation degrees for image reconstruction. However, due to the peculiarities of distortion in fisheye images [10, 42, 53], these methods fail to achieve effective control over the rectification process. To address this issue, we propose the DLQM. It maintains a learnable query set, providing diverse effective control conditions for the rectification network. During training, DLQM projects the learned high-dimensional positional mapping relationship into a low-dimensional latent space of queries. During inference, DLQM extracts the control information from the queries and converts it into corresponding control conditions for each layer of the rectification network. Simply by providing different queries, DLQM can effectively control the rectification process, as shown in Fig. 3. At the same time, our QueryCDR can also achieve smooth continuous rectification for any distortion degrees by simply interpolating between different queries." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.54, + 0.788, + 0.648 + ], + "angle": 0, + "content": "Specifically, we construct a query set \\(\\mathbf{Q}_s = \\{\\bar{Q}_i\\mid Q_i\\in \\mathbb{R}^{C_{in}\\times H_{in}\\times W_{in}},i = 1,2,\\ldots ,N\\}\\), which comprises \\(N\\) queries representing different distortion degrees. \\(H_{in}\\), \\(W_{in}\\), and \\(C_{in}\\) represent the query's height, width, and channel, respectively. Each query has the same size as the input image \\(I_{in}\\in \\mathbb{R}^{C_{in}\\times H_{in}\\times W_{in}}\\). During inference, users select a query \\(Q_{i}\\) from \\(\\mathbf{Q}_s\\) and feed it into the DLQM. In DLQM, the control extracting part \\(\\mathrm{CE}(\\cdot)\\) is firstly used to extract the features in \\(Q_{i}\\). The input processing part can be expressed as follows," + }, + { + "type": "equation", + "bbox": [ + 0.445, + 0.661, + 0.786, + 0.677 + ], + "angle": 0, + "content": "\\[\nQ _ {e x} = \\operatorname {C E} \\left(Q _ {i}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.75 + ], + "angle": 0, + "content": "where \\( Q_{ex} \\in \\mathbb{R}^{C_{in} \\times H_{in} \\times W_{in}} \\) is the extracted feature, and serves as the input to the first control layer of DLQM. To minimize the computational costs and effectively extract the control information, \\( \\mathrm{CE}(\\cdot) \\) is composed of three convolution layers with a kernel size of \\( 3 \\times 3 \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.751, + 0.788, + 0.813 + ], + "angle": 0, + "content": "Subsequently, DLQM provides corresponding control conditions to the respective layers of the rectification network. To provide appropriate control conditions, each control layer of DLQM comprises two fully connected layers \\(\\mathrm{FC}_1(\\cdot)\\), \\(\\mathrm{FC}_2(\\cdot)\\). The \\(l\\)-th layer of DLQM can be represented as follows," + }, + { + "type": "equation", + "bbox": [ + 0.416, + 0.823, + 0.786, + 0.842 + ], + "angle": 0, + "content": "\\[\nQ _ {c} ^ {l} = \\mathrm {F C} _ {2} ^ {l} \\left(\\mathrm {F C} _ {1} ^ {l} \\left(Q _ {c} ^ {l - 1}\\right)\\right), \\tag {2}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.256, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "QueryCDR: Query-based Controllable Distortion Rectification Network" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.207 + ], + "angle": 0, + "content": "where \\( Q_{c}^{l-1} \\) is the output control condition from the \\( (l-1) \\)-th control layer. For \\( l = 1 \\), \\( Q_{c}^{0} \\) represents the extracted feature \\( Q_{ex} \\). The output \\( Q_{c}^{l} \\) of the \\( l \\)-th control layer is fed into the \\( l \\)-th rectification layer as a control condition, and simultaneously serves as input to the \\( (l+1) \\)-th control layer of DLQM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.253 + ], + "angle": 0, + "content": "Afterward, we feed the control conditions into the rectification network. Each layer's controllable modulating block modulates the input feature with the control condition, guiding the modulated features toward the desired direction." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.275, + 0.517, + 0.29 + ], + "angle": 0, + "content": "3.3 Controllable Modulating Block" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.3, + 0.785, + 0.526 + ], + "angle": 0, + "content": "When controlling the rectification process, to avoid obtaining results with blurred texture details or residual distortions, we introduce two types of controllable modulating blocks tailored for reconstructing local texture details and learning continuous distortion patterns. They are denoted as the Controllable Convolution Modulating Block (CCMB) based on CNN [17] (in Fig. 2(a)) and the Controllable Attention Modulating Block (CAMB) based on Transformer [45] (in Fig. 2(b)). The CCMB can adaptively fuse the original features and the controlled features, preserving more local details. The CAMB can better capture the spatial distortion information, guaranteeing the integrity of the recovered content [10,35,53]. To balance performance and computational cost, our QueryCDR constructs a U-shaped rectification network [52] composed of CCMB and CAMB. In the first three layers with larger feature maps, i.e., \\( l = \\{1,2,3,9,10,11\\} \\), CCMB is employed to learn more local texture details. In the remaining layers with smaller feature maps, i.e., \\( l = \\{4,5,6,7,8\\} \\), CAMB is utilized to capture more global dependencies within the images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.547, + 0.784, + 0.667 + ], + "angle": 0, + "content": "Controllable Convolution Modulating Block (CCMB). When incorporating control conditions into the rectification network, existing methods [47,55] fail to effectively balance the fusion ratio between original features and controlled features. Directly using the controlled features or fusing them with original features at a fixed ratio will degrade the quality of the rectified images. To harmoniously incorporate control conditions into the rectification process, we design the CCMB, as shown in Fig. 2(a). CCMB can dynamically find an optimal ratio to modulate the distorted features with control conditions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.668, + 0.784, + 0.713 + ], + "angle": 0, + "content": "Specifically, CCMB receives an input feature \\( F_{in} \\in \\mathbb{R}^{C \\times H \\times W} \\) and a control condition \\( Q_{c} \\in \\mathbb{R}^{C \\times H \\times W} \\). Both of them are used to predict the fusion ratio. Here we omit the layer information for brevity," + }, + { + "type": "equation", + "bbox": [ + 0.441, + 0.727, + 0.785, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\theta = \\operatorname {C P} \\left(F _ {i n}, Q _ {c}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.753, + 0.783, + 0.782 + ], + "angle": 0, + "content": "where \\(\\mathrm{CP}(\\cdot)\\) is the coefficient predictor composed of two fully connected layers, takes the concatenation of \\(F_{in}\\) and \\(Q_{c}\\) as input and predicts the fusion ratio \\(\\theta\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.784, + 0.784, + 0.813 + ], + "angle": 0, + "content": "Then, we perform element-wise multiplication \\(\\otimes\\) on \\(F_{in}\\) and \\(Q_{c}\\) to yield the controlled features \\(F_{c} \\in \\mathbb{R}^{C \\times H \\times W}\\), which can be expressed as," + }, + { + "type": "equation", + "bbox": [ + 0.449, + 0.827, + 0.785, + 0.841 + ], + "angle": 0, + "content": "\\[\nF _ {c} = F _ {i n} \\otimes Q _ {c}. \\tag {4}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.36, + 0.128 + ], + "angle": 0, + "content": "P. Guo et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.178 + ], + "angle": 0, + "content": "Finally, the input features \\( F_{in} \\) and controlled features \\( F_{c} \\) are combined in a weighted sum according to the fusion ratio \\( \\theta \\), illustrated as \\( \\psi(\\cdot) \\) in Fig. 2(a)," + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.186, + 0.786, + 0.204 + ], + "angle": 0, + "content": "\\[\nF _ {o u t} = \\psi \\left(F _ {c}, F _ {i n}, \\theta\\right) = \\theta F _ {c} \\oplus (1 - \\theta) F _ {i n}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.211, + 0.786, + 0.241 + ], + "angle": 0, + "content": "where \\(\\oplus\\) represents the element-wise addition, and \\(F_{out} \\in \\mathbb{R}^{C \\times H \\times W}\\) is the final output of CCMB." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.243, + 0.786, + 0.273 + ], + "angle": 0, + "content": "Due to this dynamic modulation mechanism, CCMB achieves effective control over the rectification process while preserving richer texture details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.293, + 0.788, + 0.368 + ], + "angle": 0, + "content": "Controllable Attention Modulating Block (CAMB). Due to the lack of perception of global information, CNN-based networks struggle to learn long-range dependencies, particularly the continuous and amorphous distortions prevalent in fisheye distortions [10, 42, 53]. In contrast, Transformer-based networks effectively compensate for this, with their global attention mechanism [13, 35]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.369, + 0.788, + 0.49 + ], + "angle": 0, + "content": "Therefore, we propose the CAMB, as illustrated in Fig. 2(b). To optimally leverage the control conditions given by DLQM, we designed the control-attention mechanism, enabling CAMB to perceive the global spatial relationships in the control conditions effectively. Specifically, we unfold and project the controlled feature \\( F_{c} \\) (in Eq. (4)) as the query \\( \\mathcal{Q} \\in \\mathbb{R}^{m \\times L} \\), with the input feature \\( F_{in} \\) as the key \\( \\mathcal{K} \\in \\mathbb{R}^{m \\times L} \\) and value \\( \\mathcal{V} \\in \\mathbb{R}^{m \\times L} \\), where \\( L = H \\times W \\) represents the sequence length, and \\( m \\) denotes the dimensions of the sequences. The control-attention is described by the following," + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.499, + 0.785, + 0.552 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {Q} = W _ {\\mathcal {Q}} F _ {c}, \\mathcal {K} = W _ {\\mathcal {K}} F _ {i n}, \\mathcal {V} = W _ {\\mathcal {V}} F _ {i n}, \\\\ \\operatorname {C T R L - A T T N} (\\mathcal {Q}, \\mathcal {K}, \\mathcal {V}) = \\operatorname {s o f t m a x} \\left(\\frac {\\mathcal {Q} \\mathcal {K} ^ {T}}{\\sqrt {m}}\\right) \\mathcal {V}, \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.56, + 0.786, + 0.607 + ], + "angle": 0, + "content": "where \\( W_{\\mathcal{Q}}, W_{\\mathcal{K}}, W_{\\mathcal{V}} \\in \\mathbb{R}^{m \\times C} \\) represent the projection matrices of the queries, keys, and values, respectively. CTRL-ATTN(\\(\\cdot\\)) represents the control-attention we proposed. The overall computation of CAMB can be formulated as follows," + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.616, + 0.785, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} F _ {a} = \\operatorname {C T R L - A T T N} (\\mathrm {L N} (\\mathcal {Q}, \\mathcal {K}, \\mathcal {V})) \\oplus F _ {i n}, \\\\ F _ {o u t} = \\operatorname {C o n v} _ {1 \\times 1} (\\operatorname {F F N} (\\operatorname {L N} (F _ {a})) \\oplus F _ {a}), \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.786, + 0.72 + ], + "angle": 0, + "content": "where \\(F_{in}\\) means the input feature, \\(F_{a} \\in \\mathbb{R}^{m \\times H \\times W}\\) and \\(F_{out} \\in \\mathbb{R}^{C \\times H \\times W}\\) are the outputs of CTRL-ATTN(\\(\\cdot\\)) and CAMB, respectively. \\(\\mathrm{LN}(\\cdot)\\) denotes the layer normalization [1]. \\(\\mathrm{FFN}(\\cdot)\\) stands for the feed-forward network composed of three fully connected layers, which helps CAMB to focus on the global dependencies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.722, + 0.786, + 0.751 + ], + "angle": 0, + "content": "CAMB can discern the global mapping relationships within fisheye images, ensuring rectification uniformity compared to the CNN-based networks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.772, + 0.41, + 0.788 + ], + "angle": 0, + "content": "3.4 Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.786, + 0.841 + ], + "angle": 0, + "content": "To guarantee the robust and stable controllable distortion rectification of Query-CDR, we design a two-stage training strategy that combines coarse-grained distortion pre-training and fine-grained distortion fine-tuning." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.256, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "QueryCDR: Query-based Controllable Distortion Rectification Network" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.222 + ], + "angle": 0, + "content": "During the coarse-grained distortion pre-training phase, we choose the most commonly used dataset [10,24,52,54] that only contains one degree of distortion to train our QueryCDR. For clarity, we denote the distortion degree in this phase as \\( d \\). Correspondingly, only one query, denoted as \\( Q \\), is used for training. The optimization objective can be expressed as follows," + }, + { + "type": "equation", + "bbox": [ + 0.442, + 0.233, + 0.785, + 0.248 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p r e} = \\mathcal {L} _ {r} + \\mathcal {L} _ {m}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.255, + 0.784, + 0.284 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{pre}\\) is the overall loss function for the pre-training phase. \\(\\mathcal{L}_r\\) denotes the reconstruction loss," + }, + { + "type": "equation", + "bbox": [ + 0.431, + 0.284, + 0.785, + 0.304 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r} = \\left\\| I _ {o u t} ^ {d} - I _ {g t} ^ {d} \\right\\| _ {1}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.307, + 0.784, + 0.339 + ], + "angle": 0, + "content": "where \\( I_{out}^{d} \\) and \\( I_{gt}^{d} \\) signifie the output result and ground truth, respectively. \\( \\mathcal{L}_m \\) denotes the multi-scale loss," + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.347, + 0.785, + 0.389 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m} = \\sum_ {j = 1} ^ {Z - 1} \\| S \\left(I _ {g t} ^ {d}, j\\right) - C \\left(F _ {\\text {o u t}} ^ {j}\\right) \\| _ {1}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.785, + 0.477 + ], + "angle": 0, + "content": "where \\( S(\\cdot) \\) represents the operation that down-samples the input \\( I_{gt}^{d} \\) by a factor of \\( 1/2^{j} \\). \\( Z \\) represents the number of decoder's layers, and we set \\( Z \\) to 6. \\( F_{out}^{j} \\) denotes the feature in \\( j \\)-th decoder layer. \\( C(\\cdot) \\) is \\( 3 \\times 3 \\) convolution for decoding the features into 3-channel RGB images. In this way, each feature map on the decoder can be effectively supervised. The \\( I_{out}^{d} \\) can be obtained as," + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.485, + 0.785, + 0.502 + ], + "angle": 0, + "content": "\\[\nI _ {o u t} ^ {d} = \\operatorname {Q u e r y C D R} \\left(I _ {i n} ^ {d}, Q\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.784, + 0.539 + ], + "angle": 0, + "content": "where \\( I_{in}^{d} \\) represents fisheye images with distortion degree \\( d \\) as input. This way can effectively boost the model stability and accelerate the convergence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.54, + 0.785, + 0.659 + ], + "angle": 0, + "content": "During the fine-grained distortion fine-tuning phase, we use the varying distortion degrees datasets to fine-tune our QueryCDR. Before training, We replicate the weight of the query \\( Q \\) to the other queries to accelerate convergence on other distortion degrees. Subsequently, we fine-tune our QueryCDR using fisheye images with varying distortion degrees, and feeding corresponding queries into QueryCDR for training at the same time. This allows the query set to efficiently acquire diverse latent spatial relationships. The optimization objective can be expressed as follows," + }, + { + "type": "equation", + "bbox": [ + 0.436, + 0.659, + 0.785, + 0.678 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {f i n e}} ^ {d _ {i}} = \\mathcal {L} _ {r} ^ {d _ {i}} + \\mathcal {L} _ {m} ^ {d _ {i}}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.683, + 0.784, + 0.732 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{fine}^{d_i}\\) is the overall fine-tuning loss function for distortion degree \\(d_{i}, i \\in \\{1,2,\\dots,9\\}\\), which is similar to the \\(\\mathcal{L}_{pre}\\) but calculated across different distortion degrees. The \\(I_{out}^{d_i}\\) can be obtained as," + }, + { + "type": "equation", + "bbox": [ + 0.406, + 0.739, + 0.785, + 0.756 + ], + "angle": 0, + "content": "\\[\nI _ {\\text {o u t}} ^ {d _ {i}} = \\operatorname {Q u e r y C D R} \\left(I _ {\\text {i n}} ^ {d _ {i}}, Q _ {i}\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.764, + 0.784, + 0.795 + ], + "angle": 0, + "content": "where \\( I_{in}^{d_i} \\) denotes fisheye images with a distortion degree of \\( d_i \\) as input, \\( Q_i \\) denotes the query corresponding to \\( d_i, i \\in \\{1, 2, \\dots, 9\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.784, + 0.84 + ], + "angle": 0, + "content": "With this two-stage training strategy, our QueryCDR can effectively utilize only a small amount of varying distortion degrees data that is difficult to obtain, to rapidly finish the training of our network." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.128 + ], + "angle": 0, + "content": "P. Guo et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison (PSNR (dB)↑, SSIM ↑) on COCO [27] fisheye image dataset with varying distortion degrees. Red indicates the best and blue indicates the second best performance (best viewed in color)." + }, + { + "type": "table", + "bbox": [ + 0.258, + 0.199, + 0.747, + 0.531 + ], + "angle": 0, + "content": "
MethodPSNR
d1d2d3d4d5d6d7d8d9Avg
SC [40]10.0511.5911.8111.0311.5010.1911.1110.199.1410.73
DeepCalib [4]10.0010.6911.0111.1911.2811.4611.4511.1311.1111.04
Blind [23]12.9811.2410.309.629.9911.7512.6912.7512.8011.57
DR-GAN [24]15.6817.4017.9718.3418.5018.4417.9517.9417.4717.74
PCN [52]14.9317.4318.4318.8618.8618.8818.7417.3518.2617.97
DDA [54]16.3917.4117.4319.4820.1218.9018.8518.1718.2218.33
SimFIR [10]16.5717.8818.4318.9719.3119.2819.1918.6518.4818.53
QueryCDR20.0120.2920.3920.4120.7220.8120.5819.1120.5320.32
MethodSSIM
d1d2d3d4d5d6d7d8d9Avg
SC [40]0.1010.1130.1490.1820.2830.1750.1410.1260.0930.151
DeepCalib [4]0.1840.2100.2230.2300.2340.2460.2500.2450.2460.229
Blind [23]0.3080.2440.1990.1760.1940.2960.3670.3950.4200.289
DR-GAN [24]0.2950.3300.3390.3440.3440.3320.3140.3120.2990.323
PCN [52]0.4200.5470.5890.6070.6080.6100.6150.5760.6030.575
DDA [54]0.4550.5890.5920.6200.6750.6260.6190.5640.5810.591
SimFIR [10]0.4920.5810.6260.6350.6400.6280.6220.5910.5950.601
QueryCDR0.6430.6650.6680.6770.6880.6990.6920.6560.6930.676
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.554, + 0.367, + 0.57 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.582, + 0.446, + 0.598 + ], + "angle": 0, + "content": "4.1 Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.603, + 0.788, + 0.756 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of our proposed QueryCDR, we followed the existing works [10, 24, 52, 54], employed the four-parameter polynomial model to synthesize fisheye images. We constructed synthetic datasets based on the original images of COCO [27] and Places2 [62] datasets, respectively. Specifically, for images with distortion degree \\( d \\), 40,000 images were used for the pre-training stage. And for images with varying distortion degrees \\( d_{i} \\), \\( i \\in \\{1,2,\\dots ,9\\} \\), 18,000 images were used for the fine-tuning stage, and 3,600 images for testing. The images were resized to \\( 256 \\times 256 \\) when fed into the network. For a fair comparison, we followed existing works [10, 24, 52, 54] utilizing a batch size of 16 and the Adam [19] optimizer with a learning rate of 1e-4." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.773, + 0.461, + 0.789 + ], + "angle": 0, + "content": "4.2 Performance Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To evaluate the performance of our method, we retrained and validated existing fisheye image rectification methods on synthetic fisheye datasets with 9 distortion degrees. The methods can be summarized into three categories: traditional" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.257, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "QueryCDR: Query-based Controllable Distortion Rectification Network" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.597, + 0.684, + 0.612 + ], + "angle": 0, + "content": "Fig. 4: Qualitative results on synthetic fisheye images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.707 + ], + "angle": 0, + "content": "method SC [40], regression-based methods DeepCalib [4] and Blind [23], and generation-based DR-GAN [24], PCN [52], DDA [54] and SimFIR [10]. For a fair comparison, we followed existing works [10,24,52,54] employing Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) to quantify performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Quantitative Results. Our performance comparisons are shown in Tab. 1. Compared to existing methods, QueryCDR achieves the best performance across all distortion degrees without retraining. It is because DLQM effectively utilizes the control information in queries to guide the network in achieving varying degrees of rectification. Additionally, the capability of CCMB and CAMB allows for controlled rectification while obtaining high-quality results. Moreover, we observe that even on images with distortion degree \\( d_{5} \\), which the existing" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "P. Guo et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.273, + 0.687, + 0.288 + ], + "angle": 0, + "content": "Fig. 5: Qualitative results on real-world fisheye images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.315, + 0.785, + 0.361 + ], + "angle": 0, + "content": "methods perform well [10, 24, 52, 54], QueryCDR still outperforms the second best method [54] by 0.60 dB of PSNR and 0.013 of SSIM. This further demonstrates the outstanding generalization ability of QueryCDR." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.785, + 0.487 + ], + "angle": 0, + "content": "Qualitative Results. To further compare the visual qualities of different methods, we show the results rectified by QueryCDR and other rectification methods in Fig. 4. It can be observed intuitively that other methods fail to effectively rectify images with varying degrees of distortion. Particularly the objects in images, our QueryCDR maintains the accurate structure of objects across different distortions. Furthermore, benefiting from the modulation capability of CCMB and CAMB, QueryCDR preserves richer texture details after rectification." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.785, + 0.563 + ], + "angle": 0, + "content": "In addition, to further demonstrate the generalization ability of QueryCDR, we conducted experiments on real-world fisheye image datasets [57] in Fig. 5. Despite the disparities between synthetic and real-world datasets, our QueryCDR still shows robust rectification capabilities, and preservers richer texture details. These results further validate its practicality in real-world scenarios." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.585, + 0.389, + 0.6 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.785, + 0.639 + ], + "angle": 0, + "content": "In this section, we conduct ablation for different control conditions, and study the effect of the controllable modulating blocks and the architecture setting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.786, + 0.795 + ], + "angle": 0, + "content": "Distortion-aware Learnable Query Mechanism (DLQM). To validate the effectiveness of DLQM in controlling the rectification process, we implemented different rectification networks controlled by scalar, fixed query, and our learnable query in Tab. 2. When using scalar to control, the value of the scalar increments with the distortion of the entire image. When using fixed query to control, the parameters of the query are set to incremental values from 0 to 1 with the degree of distortion from the center to the edges. Following the settings in Sec. 4.1, we use the learnable query \\( Q_{i}, i \\in \\{1,2,\\dots ,9\\} \\) to learn the distortion distribution corresponding to \\( d_{i}, i \\in \\{1,2,\\dots ,9\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We trained different methods using the same strategy as illustrated in Sec. 3.4, the results are shown in Tab. 2 and Fig. 6. All three controllable methods outperform the uncontrollable method significantly at various degrees of distortion." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.256, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "QueryCDR: Query-based Controllable Distortion Rectification Network" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.289, + 0.145, + 0.713, + 0.159 + ], + "angle": 0, + "content": "Table 2: Performance comparison of different control methods." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.182, + 0.787, + 0.285 + ], + "angle": 0, + "content": "
MethodPSNR
d1d2d3d4d5d6d7d8d9Avg
W/o Control14.9317.4318.4318.8618.8618.8818.7417.3518.2617.97
Scalar19.5220.2519.9719.4319.8920.0720.0318.7620.1519.78
Fixed Query20.1320.0320.1419.8820.1620.2520.1618.7419.8419.93
Learnable Query20.0120.2920.3920.4120.7220.8120.5819.1120.5320.32
" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.313, + 0.446, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.416, + 0.443, + 0.443 + ], + "angle": 0, + "content": "Fig. 6: Visual results of different control mechanisms." + }, + { + "type": "image", + "bbox": [ + 0.452, + 0.313, + 0.779, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.451, + 0.416, + 0.778, + 0.443 + ], + "angle": 0, + "content": "Fig.7: Visual results of different modulation methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.788, + 0.613 + ], + "angle": 0, + "content": "This is because the controllable mechanism effectively assists the rectification network in distinguishing between different degrees of distortion, thereby enhancing the generalization capability of the rectification network. The improvement brought by scalar demonstrates the crucial role of controllable mechanisms in handling fisheye images with different degrees of distortion. The further enhancement with fixed query validates the importance of employing higher-dimensional control conditions for fisheye image rectification. Lastly, our tailored learnable query significantly outperforms all other controllable methods, verifying the unique superiority of our approach in controllable rectification." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.788, + 0.734 + ], + "angle": 0, + "content": "Controllable Modulating Block. To demonstrate the capability of the CCMB and CAMB in modulating input features with control conditions, we implemented two comparison methods with different modulating approaches: one directly uses the controlled feature (in Eq. (4)), and the other adds controlled feature and original feature in a fixed 1:1 ratio. We constructed different controllable networks using each of these methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "As shown in Tab. 3 and Fig. 7, retaining original features and fusing them with a 1:1 ratio improved by \\(0.03\\mathrm{dB}\\), highlighting the importance of preserving original features. The dynamic modulation mechanism boosted performance by \\(0.06\\mathrm{dB}\\), verifying its effectiveness in dynamic feature modulation. Furthermore, leveraging the control-attention mechanism led to a performance improvement of \\(0.12\\mathrm{dB}\\), demonstrating the superiority of CAMB for perceiving global distortion distribution, and achieving better uniform global rectification." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.128 + ], + "angle": 0, + "content": "P. Guo et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.482, + 0.256 + ], + "angle": 0, + "content": "Table 3: Ablation study of different modulation methods. Where Direct Use \\(F_{c}\\) means directly using the controlled feature \\(F_{c}\\), Fixed Ratio means adding controlled features and original features in a 1:1 ratio, Dynamic Mechanism means the CCMB and Attention Mechanism means the CAMB." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.268, + 0.472, + 0.356 + ], + "angle": 0, + "content": "
MethodPSNRSSIM
Direct UseFc20.140.655
Fixed Ratio20.170.658
Dynamic Mechanism20.200.669
Attention Mechanism20.260.671
" + }, + { + "type": "table_caption", + "bbox": [ + 0.497, + 0.145, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Table 4: Ablation study of different network architectures. Where \\( x\\mathrm{C} + \\left( {{11} - x}\\right) \\mathrm{A} \\) denotes that the network uses CCMB in the \\( x \\) layers with larger feature maps and CAMB in the remaining \\( {11} - x \\) layers with smaller feature maps." + }, + { + "type": "table", + "bbox": [ + 0.5, + 0.241, + 0.797, + 0.356 + ], + "angle": 0, + "content": "
MethodFlops(G)Param(M)PSNRSSIM
PCN [52]12.30535.63717.970.575
11C+0A12.73637.70120.200.669
8C+3A13.38346.39820.270.665
6C+5A12.35343.24420.320.676
4C+7A12.53846.99420.310.670
0C+11A15.19051.79520.260.671
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.79, + 0.602 + ], + "angle": 0, + "content": "Controllable Rectification Network Architecture. As described in Sec. 3.3, rectification networks based on CCMB may ignore the long-range dependencies, while those based on CAMB may ignore the texture details and also increase computational costs. To strike a balance between these two architectures, We incorporate CCMB into the layers with larger feature map size, and integrate CAMB into the layers with smaller feature map size. To validate the effectiveness of this hybrid architecture and find an optimal combination, we conducted performance and computational cost comparisons between pure CCMB, pure CAMB, and multiple hybrid networks, as shown in Tab. 4. Pure CAMB network (i.e., Row 6) outperforms pure CCMB network (i.e., Row 2) in rectification performance but incurs higher computational costs. Hybrid network architectures effectively solve this problem, with higher performance and fewer parameters. After a trade-off between performance and parameters, we empirically choose the \\(6\\mathrm{C} + 5\\mathrm{A}\\) hybrid architecture as our final model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.641, + 0.36, + 0.657 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we propose the Query-based Controllable Distortion Rectification network for fisheye images (QueryCDR), which achieves controllable rectification at different distortion degrees without retraining. In particular, we design a series of learnable queries as control conditions to guide the rectification process. Additionally, we design two different controllable modulating blocks, achieving controllable rectification while improving image quality. Extensive experiments have demonstrated the robustness and effectiveness of our QueryCDR. In the future, it is expected to further rectify distortions of different degrees by autocontrolling mechanisms to avoid human involvement. We believe our work provides an effective solution for fisheye camera applications." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.256, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "QueryCDR: Query-based Controllable Distortion Rectification Network" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.253 + ], + "angle": 0, + "content": "This work was supported in part by the NSFC under Grant 62272376, 62272380 and 62103317, the Key Research and Development Program of Shaanxi Province under Grant 2020ZDLGY04-05, the Fundamental Research Funds for the Central Universities, China (xzy022023051), the Innovative Leading Talents Scholarship of Xi'an Jiaotong University (Corresponding author: Xingsong Hou)." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.274, + 0.323, + 0.289 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.303, + 0.785, + 0.331 + ], + "angle": 0, + "content": "1. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.333, + 0.785, + 0.373 + ], + "angle": 0, + "content": "2. Barreto, J.P., Araujo, H.: Geometric properties of central catadioptric line images and their application in calibration. IEEE Transactions on Pattern Analysis and Machine Intelligence 27(8), 1327-1333 (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.374, + 0.785, + 0.413 + ], + "angle": 0, + "content": "3. Barreto, J.P., Daniilidis, K.: Fundamental matrix for cameras with radial distortion. In: Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1. vol. 1, pp. 625-632. IEEE (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.415, + 0.785, + 0.469 + ], + "angle": 0, + "content": "4. Bogdan, O., Eckstein, V., Rameau, F., Bazin, J.C.: Deepcalib: A deep learning approach for automatic intrinsic calibration of wide field-of-view cameras. In: Proceedings of the 15th ACM SIGGRAPH European Conference on Visual Media Production. pp. 1-10 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.47, + 0.785, + 0.496 + ], + "angle": 0, + "content": "5. Bukhari, F., Dailey, M.N.: Automatic radial distortion estimation from a single image. Journal of mathematical imaging and vision 45, 31-45 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.498, + 0.785, + 0.537 + ], + "angle": 0, + "content": "6. Cai, B., Xu, X., Jia, K., Qing, C., Tao, D.: Dehazenet: An end-to-end system for single image haze removal. IEEE transactions on image processing 25(11), 5187-5198 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.539, + 0.785, + 0.579 + ], + "angle": 0, + "content": "7. Cai, H., He, J., Qiao, Y., Dong, C.: Toward interactive modulation for photorealistic image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 294-303 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.58, + 0.785, + 0.606 + ], + "angle": 0, + "content": "8. Devernay, F., Faugeras, O.: Straight lines have to be straight. Machine vision and applications 13, 14-24 (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.608, + 0.785, + 0.661 + ], + "angle": 0, + "content": "9. Duan, Z., Tezcan, O., Nakamura, H., Ishwar, P., Konrad, J.: Rapid: rotation-aware people detection in overhead fisheye images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 636-637 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.662, + 0.785, + 0.717 + ], + "angle": 0, + "content": "10. Feng, H., Wang, W., Deng, J., Zhou, W., Li, L., Li, H.: Simfir: A simple framework for fisheye image rectification with self-supervised representation learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12418-12427 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.718, + 0.785, + 0.757 + ], + "angle": 0, + "content": "11. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: 2012 IEEE conference on computer vision and pattern recognition. pp. 3354-3361. IEEE (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "12. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "13. Guo, J., Han, K., Wu, H., Tang, Y., Chen, X., Wang, Y., Xu, C.: Cmt: Convolutional neural networks meet vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12175-12185 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.303, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "P. Guo et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.189 + ], + "angle": 0, + "content": "14. Hartley, R., Kang, S.B.: Parameter-free radial distortion correction with center of distortion estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence 29(8), 1309-1321 (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "15. He, J., Dong, C., Qiao, Y.: Modulating image restoration with continual levels via adaptive feature modification layers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11056-11064 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.787, + 0.286 + ], + "angle": 0, + "content": "16. He, J., Dong, C., Qiao, Y.: Interactive multi-dimension modulation with dynamic controllable residual learning for image restoration. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16. pp. 53-68. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.328 + ], + "angle": 0, + "content": "17. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "18. Henrique Brito, J., Angst, R., Koser, K., Pollefeys, M.: Radial distortion self-calibration. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1368-1375 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.397 + ], + "angle": 0, + "content": "19. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.438 + ], + "angle": 0, + "content": "20. Kukelova, Z., Pajdla, T.: A minimal solution to radial distortion autocalibration. IEEE Transactions on Pattern Analysis and Machine Intelligence 33(12), 2410-2422 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.495 + ], + "angle": 0, + "content": "21. Kumar, V.R., Klingner, M., Yogamani, S., Milz, S., Fingscheidt, T., Mader, P.: Syndistnet: Self-supervised monocular fisheye camera distance estimation synergized with semantic segmentation for autonomous driving. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 61-71 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.55 + ], + "angle": 0, + "content": "22. Ledig, C., Theis, L., Huszár, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., et al.: Photo-realistic single image superresolution using a generative adversarial network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4681-4690 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.592 + ], + "angle": 0, + "content": "23. Li, X., Zhang, B., Sander, P.V., Liao, J.: Blind geometric distortion correction on images through deep learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4855-4864 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.787, + 0.633 + ], + "angle": 0, + "content": "24. Liao, K., Lin, C., Zhao, Y., Gabbouj, M.: Dr-gan: Automatic radial distortion rectification using conditional gan in real-time. IEEE Transactions on Circuits and Systems for Video Technology 30(3), 725-733 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.787, + 0.675 + ], + "angle": 0, + "content": "25. Liao, K., Lin, C., Zhao, Y., Xu, M.: Model-free distortion rectification framework bridged by distortion distribution map. IEEE Transactions on Image Processing 29, 3707-3718 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.716 + ], + "angle": 0, + "content": "26. Lin, L., Lu, Y., Pan, Y., Chen, X.: Integrating graph partitioning and matching for trajectory analysis in video surveillance. IEEE transactions on Image Processing 21(12), 4844-4857 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.771 + ], + "angle": 0, + "content": "27. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.787, + 0.813 + ], + "angle": 0, + "content": "28. Liu, C., Wang, X., Li, S., Wang, Y., Qian, X.: Fsi: Frequency and spatial interactive learning for image restoration in under-display cameras. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12537-12546 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "29. Liu, C., Wang, X., Xu, X., Tian, R., Li, S., Qian, X., Yang, M.H.: Motion-adaptive separable collaborative filters for blind motion deblurring. In: Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.257, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "QueryCDR: Query-based Controllable Distortion Rectification Network" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.147, + 0.786, + 0.175 + ], + "angle": 0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 25595-25605 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.175, + 0.787, + 0.215 + ], + "angle": 0, + "content": "30. Liu, C., Yang, H., Fu, J., Qian, X.: Learning trajectory-aware transformer for video super-resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5687-5696 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.216, + 0.787, + 0.257 + ], + "angle": 0, + "content": "31. Liu, C., Yang, H., Fu, J., Qian, X.: 4d lut: learnable context-aware 4d lookup table for image enhancement. IEEE Transactions on Image Processing 32, 4742-4756 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.787, + 0.298 + ], + "angle": 0, + "content": "32. Mei, C., Rives, P.: Single view point omnidirectional camera calibration from planar grids. In: Proceedings 2007 IEEE International Conference on Robotics and Automation. pp. 3945-3950. IEEE (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.787, + 0.338 + ], + "angle": 0, + "content": "33. Mou, C., Wu, Y., Wang, X., Dong, C., Zhang, J., Shan, Y.: Metric learning based interactive modulation for real-world super-resolution. In: European Conference on Computer Vision. pp. 723-740. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.787, + 0.392 + ], + "angle": 0, + "content": "34. Muhammad, K., Ahmad, J., Lv, Z., Bellavista, P., Yang, P., Baik, S.W.: Efficient deep cnn-based fire detection and localization in video surveillance applications. IEEE Transactions on Systems, Man, and Cybernetics: Systems 49(7), 1419-1434 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.392, + 0.787, + 0.433 + ], + "angle": 0, + "content": "35. Peng, Z., Huang, W., Gu, S., Xie, L., Wang, Y., Jiao, J., Ye, Q.: Conformer: Local features coupling global representations for visual recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 367-376 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.787, + 0.473 + ], + "angle": 0, + "content": "36. Plaut, E., Ben Yaacov, E., El Shlomo, B.: 3d object detection from a single fisheye image without a single fisheye training image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3659-3667 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.514 + ], + "angle": 0, + "content": "37. Puig, L., Bastanlar, Y., Sturm, P., Guerrero, J.J., Barreto, J.: Calibration of central catadioptric cameras using a dlt-like approach. International Journal of Computer Vision 93, 101-114 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.787, + 0.569 + ], + "angle": 0, + "content": "38. Rashed, H., Mohamed, E., Sistu, G., Kumar, V.R., Eising, C., El-Sallab, A., Yogamani, S.: Generalized object detection on fisheye cameras for autonomous driving: Dataset, representations and baseline. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2272-2280 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.787, + 0.637 + ], + "angle": 0, + "content": "39. Rong, J., Huang, S., Shang, Z., Ying, X.: Radial lens distortion correction using convolutional neural networks trained with synthesized images. In: Computer Vision-ACCV 2016: 13th Asian Conference on Computer Vision, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part III 13. pp. 35-49. Springer (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.787, + 0.678 + ], + "angle": 0, + "content": "40. Santana-Cedrés, D., Gomez, L., Aleman-Flores, M., Salgado, A., Esclarín, J., Mazorra, L., Alvarez, L.: An iterative optimization algorithm for lens distortion correction using two-parameter models. Image Processing On Line 6, 326-364 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.787, + 0.732 + ], + "angle": 0, + "content": "41. Scaramuzza, D., Martinelli, A., Siegwart, R.: A flexible technique for accurate omnidirectional camera calibration and structure from motion. In: Fourth IEEE International Conference on Computer Vision Systems (ICVS'06). pp. 45-45. IEEE (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.787, + 0.772 + ], + "angle": 0, + "content": "42. Shen, Z., Lin, C., Liao, K., Nie, L., Zheng, Z., Zhao, Y.: Panoformer: Panorama transformer for indoor 360 depth estimation. In: European Conference on Computer Vision. pp. 195-211. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.813 + ], + "angle": 0, + "content": "43. Sturm, P., Ramalingam, S.: A generic concept for camera calibration. In: Computer Vision-ECCV 2004: 8th European Conference on Computer Vision, Prague, Czech Republic, May 11-14, 2004. Proceedings, Part II 8. pp. 1-13. Springer (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.841 + ], + "angle": 0, + "content": "44. Thormahlen, T., Broszio, H., Wassermann, I.: Robust line-based calibration of lens distortion from a single view. Mirage 2003 pp. 105-112 (2003)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.128 + ], + "angle": 0, + "content": "P. Guo et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "45. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.23 + ], + "angle": 0, + "content": "46. Wang, A., Qiu, T., Shao, L.: A simple method of radial distortion correction with centre of distortion estimation. Journal of Mathematical Imaging and Vision 35, 165-172 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.785, + 0.271 + ], + "angle": 0, + "content": "47. Wang, W., Guo, R., Tian, Y., Yang, W.: Cfsnet: Toward a controllable feature space for image restoration. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4140-4149 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.785, + 0.311 + ], + "angle": 0, + "content": "48. Wang, X., Yu, K., Dong, C., Tang, X., Loy, C.C.: Deep network interpolation for continuous imagery effect transition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1692-1701 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.311, + 0.785, + 0.366 + ], + "angle": 0, + "content": "49. Wang, X., Yu, K., Wu, S., Gu, J., Liu, Y., Dong, C., Qiao, Y., Change Loy, C.: Esrgan: Enhanced super-resolution generative adversarial networks. In: Proceedings of the European conference on computer vision (ECCV) workshops. pp. 0-0 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.366, + 0.785, + 0.407 + ], + "angle": 0, + "content": "50. Xue, Z., Xue, N., Xia, G.S., Shen, W.: Learning to calibrate straight lines for fisheye image rectification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1643-1651 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.407, + 0.785, + 0.433 + ], + "angle": 0, + "content": "51. Yang, C.Y., Chen, H.H.: Efficient face detection in the fisheye image domain. IEEE Transactions on Image Processing 30, 5641-5651 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.433, + 0.785, + 0.488 + ], + "angle": 0, + "content": "52. Yang, S., Lin, C., Liao, K., Zhang, C., Zhao, Y.: Progressively complementary network for fisheye image rectification using appearance flow. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6348-6357 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.488, + 0.785, + 0.528 + ], + "angle": 0, + "content": "53. Yang, S., Lin, C., Liao, K., Zhao, Y.: Fishformer: Annulus slicing-based transformer for fisheye rectification with efficacy domain exploration. arXiv preprint arXiv:2207.01925 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.528, + 0.785, + 0.569 + ], + "angle": 0, + "content": "54. Yang, S., Lin, C., Liao, K., Zhao, Y.: Dual diffusion architecture for fisheye image rectification: Synthetic-to-real generalization. arXiv preprint arXiv:2301.11785 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.569, + 0.785, + 0.596 + ], + "angle": 0, + "content": "55. Yao, M., He, D., Li, X., Li, F., Xiong, Z.: Towards interactive self-supervised denoising. IEEE Transactions on Circuits and Systems for Video Technology (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.596, + 0.785, + 0.637 + ], + "angle": 0, + "content": "56. Yin, X., Wang, X., Yu, J., Zhang, M., Fua, P., Tao, D.: Fisheyerecnet: A multicontext collaborative deep network for fisheye image rectification. In: Proceedings of the European conference on computer vision (ECCV). pp. 469-484 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.637, + 0.785, + 0.691 + ], + "angle": 0, + "content": "57. Yogamani, S., Hughes, C., Horgan, J., Sistu, G., Varley, P., O'Dea, D., Uricár, M., Milz, S., Simon, M., Amende, K., et al.: Woodscape: A multi-task, multicamera fisheye dataset for autonomous driving. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9308-9318 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.691, + 0.785, + 0.732 + ], + "angle": 0, + "content": "58. Zhang, K., Zuo, W., Chen, Y., Meng, D., Zhang, L.: Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing 26(7), 3142-3155 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.732, + 0.785, + 0.786 + ], + "angle": 0, + "content": "59. Zhang, M., Yao, J., Xia, M., Li, K., Zhang, Y., Liu, Y.: Line-based multi-label energy optimization for fisheye image rectification and calibration. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4137-4145 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "60. Zhang, Y., You, S., Gevers, T.: Automatic calibration of the fisheye camera for egocentric 3d human pose estimation from a single image. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1772-1781 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.258, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "QueryCDR: Query-based Controllable Distortion Rectification Network" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.19 + ], + "angle": 0, + "content": "61. Zhang, Z., Jiang, Y., Shao, W., Wang, X., Luo, P., Lin, K., Gu, J.: Real-time controllable denoising for image and video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14028-14038 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "62. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence 40(6), 1452-1464 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.232 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/QueryCDR_ Query-based Controllable Distortion Rectification Network for Fisheye Images/8182781d-9f7c-4e2a-86d5-0c2e6b9982a2_origin.pdf b/2024/QueryCDR_ Query-based Controllable Distortion Rectification Network for Fisheye Images/8182781d-9f7c-4e2a-86d5-0c2e6b9982a2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b1eef3469c40e62fa4428cde9a078e271a8796fe --- /dev/null +++ b/2024/QueryCDR_ Query-based Controllable Distortion Rectification Network for Fisheye Images/8182781d-9f7c-4e2a-86d5-0c2e6b9982a2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c496a60e27bef51bab283fcff9dbb359babf7b9678168de2c304f672607c8976 +size 2349795 diff --git a/2024/QueryCDR_ Query-based Controllable Distortion Rectification Network for Fisheye Images/full.md b/2024/QueryCDR_ Query-based Controllable Distortion Rectification Network for Fisheye Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a8c0e73b2def45615f988e0726e782bb5e9fa111 --- /dev/null +++ b/2024/QueryCDR_ Query-based Controllable Distortion Rectification Network for Fisheye Images/full.md @@ -0,0 +1,329 @@ +# QueryCDR: Query-based Controllable Distortion Rectification Network for Fisheye Images + +Pengbo Guo $^{1,2}$ , Chengxu Liu $^{2,3}$ , Xingsong Hou $^{2(\boxtimes)}$ , and Xueming Qian $^{2,3}$ + +$^{1}$ School of Software Engineering, Xi'an Jiaotong University, Xi'an, China + $^{2}$ Xi'an Jiaotong University, Xi'an, China + +3 Shaanxi Yulan Jiuzhou Intelligent Optoelectronic Tech. Co., Ltd, Xi'an, China {guopengbo,chengxuliu}@stu.xjtu.edu.cn, {houxs,qianxm}@mail.xjtu.edu.cn + +Abstract. Fisheye image rectification aims to correct distortions in images taken with fisheye cameras. Although current models show promising results on images with a similar degree of distortion as the training data, they will produce sub-optimal results when the degree of distortion changes and without retraining. The lack of generalization ability for dealing with varying degrees of distortion limits their practical application. In this paper, we take one step further to enable effective distortion rectification for images with varying degrees of distortion without retraining. We propose a novel Query-based Controllable Distortion Rectification network for fisheye images (QueryCDR). In particular, we first present the Distortion-aware Learnable Query Mechanism (DLQM), which defines the latent spatial relationships for different distortion degrees as a series of learnable queries. Each query can be learned to obtain position-dependent rectification control conditions, providing control over the rectification process. Then, we propose two kinds of controllable modulating blocks to enable the control conditions to guide the modulation of the distortion features better. These core components cooperate with each other to effectively boost the generalization ability of the model at varying degrees of distortion. Extensive experiments on fisheye image datasets with different distortion degrees demonstrate our approach achieves high-quality and controllable distortion rectification. Code is available at https://github.com/PbGuo/QueryCDR. + +Keywords: Fisheye image $\cdot$ Distortion rectification $\cdot$ Controllable + +# 1 Introduction + +Benefiting from the huge field-of-view (FoV), fisheye cameras are widely utilized in various fields, including security surveillance [26, 34] and autonomous driving [11, 38]. However, the distortion brought by the fisheye lenses greatly limits the performance of downstream vision tasks [9, 21, 36, 60]. How to eliminate the distortion in fisheye images has attracted great attention in recent years. + +Early methods [3,5,14,18,37,40,46,59] primarily relied on identifying matching feature points or curves for automatic rectification. However, constrained by + +![](images/411c437a400b17158787ed04a95c04437841e28263fb6a1ccba12b2d37dc9705.jpg) +Fig. 1: Different approaches to fisheye image distortion rectification. (a) Regression-Based: Using a neural network to predict distortion-related parameters, then apply rectification algorithms $R$ for rectification. (b) Generation-Based: Input the distorted fisheye image and directly generate the rectified image end-to-end. (c) Control-Based: Users provide control conditions to guide the rectification process, resulting in promising rectified images of various distortion degrees. + +the instability of feature detection, both the generalization of the algorithm and the quality of the rectified images were unsatisfactory [10, 24, 52]. In recent years, owing to the robust learning and generalization capabilities of neural networks, deep learning-based fisheye image rectification methods have become mainstream. These methods can be divided into two categories: regression-based rectification [4, 23, 39, 50, 56] and generation-based rectification [10, 24, 25, 52, 54]. The former uses deep regression models to predict distortion parameters for image reconstruction, as shown in Fig. 1(a). The latter uses an encoder-decoder structure to generate well-rectified images directly, as shown in Fig. 1(b). + +It is worth noting that these methods only achieve satisfactory results on a similar degree of distortion as the training data. This means that when handling different degrees of distortion without retraining, there will be a significant decrease in the quality of the rectified images [25]. This is because the model tends to learn fixed position mapping relationships during training. When the distortion degree changes, this relationship does not work for the new distortion distribution. Therefore, these models usually need to be retrained on images with different distortion degrees. In addition, fisheye image acquisition is difficult, and re-collecting new datasets with varying degrees of distortion would incur significant costs [52, 54]. Thus, it is essential to explore a model that can handle different degrees of distortion simultaneously. + +In recent years, some methods have been proposed to achieve controllable image restoration [7,15,16,33,47,48,55,61]. Typically, CFSNet [47], MM-RealSR [33], and Yao et al. [55] introduce scalars as control conditions to effectively restore degradation of varying degrees. It would be a promising solution to introduce an effectively controllable mechanism to deal with all degrees of distortion, as shown in Fig. 1(c). However, directly applying existing controllable mechanisms in restoration tasks to the fisheye rectification model suffers from the following challenges: 1) There is a gap between optimization objectives (i.e. the roles + +of the controllable mechanisms). Restoration models learn the pixel-level detail restoration, whereas fisheye rectification models learn the spatial-level positional mapping relationships [10,25]. The control mechanism lacking position information in restoration tasks is ineffective in controlling distortion rectification networks. 2) There is a gap between optimization difficulties (i.e. the roles of the control conditions). The distortion of a fisheye image increases gradually from the image center to boundary [51]. Therefore, it is not suitable to use a single scalar control condition in restoration tasks to deal with such spatially varying distortions. These challenges restrict the application of controllable mechanisms and control conditions in fisheye image rectification. + +To address these issues, we propose the Query-based Controllable Distortion Rectification network (QueryCDR), as shown in Fig. 2. By introducing a series of learnable queries as control conditions, QueryCDR allows the users to achieve fisheye rectification with different distortion degrees. Specifically, to incorporate positional mapping relationships into control conditions, we introduce the Distortion-aware Learnable Query Mechanism (DLQM), which defines a series of queries representing different rectification control conditions. During inference, DLQM extracts position-dependent control conditions from the user-given query and feeds them into the network for controlling the rectification process. Furthermore, to enable the control conditions to guide the rectification efficiently, we propose two types of controllable modulating blocks: the Controllable Convolution Modulating Block (CCMB) based on CNN [17], and the Controllable Attention Modulating Block (CAMB) based on Transformer [45]. They are good at extracting local texture features and learning long-range distortion mapping relationships, respectively. By combining CCMB and CAMB, we construct a robust controllable rectification network. Our QueryCDR can handle various distortions without retraining, enhancing the generalization ability of the fisheye rectification model. + +We summarize our contributions as follows: + +- We propose QueryCDR, a Query-based Controllable Distortion Rectification network for fisheye images. Extensive experiments demonstrate that our QueryCDR can deliver superior results on a variety of distortion degrees. +- We propose the Distortion-aware Learnable Query Mechanism (DLQM), which effectively introduces the latent spatial relationships to control conditions for fisheye image rectification. +- We propose two kinds of blocks for modulating features using control conditions: the Controllable Convolution Modulating Block (CCMB) and the Controllable Attention Modulating Block (CAMB). They can effectively utilize control conditions to guide the rectification process. + +# 2 Related Work + +# 2.1 Traditional Fisheye Image Rectification + +Traditional rectification methods can be divided into two types, multi-view-based and line-based methods. Multi-view-based methods [3,14,18,20,37,41,43] + +calibrated fisheye images by finding corresponding feature points from multiple viewpoints. Line-based methods [2,5,8,32,40,44,46,59] employed line detection to rectify the curved lines, thereby achieving distortion rectification. However, these methods require manual intervention or handcrafted feature extractors, and the rectification process is unstable, failing to achieve satisfactory results. + +# 2.2 Deep Learning Based Fisheye Image Rectification + +Rapid advances in deep learning have allowed them to shine in low-level vision tasks [6, 22, 28-31, 49, 58]. As a core task in low-level vision, distortion image rectification has received increasing attention [10, 24, 39, 52, 54, 56]. According to the network architecture, deep learning-based distortion image rectification methods can be categorized into two types, regression-based methods [4, 23, 39, 50, 56] and generation-based methods [10, 24, 25, 52, 54]. + +Regression-based methods [4, 23, 39, 50, 56] utilize neural networks to predict distortion-related coefficients for rectifying the image. Rong et al. [39] were the first to use CNNs for fisheye image rectification. They employed the network to predict across multiple distortion intervals, achieving preliminary rectification results. Yin et al. [56] integrated semantics as prior information to guide rectification. While these methods have shown some performance improvement, they face the challenge of non-end-to-end design, thus requiring a large number of additional labels, and increasing operational costs. Therefore, to reduce models' complexity, Generation-based methods [10, 24, 25, 52, 54] employ generative networks to take distorted images as input and directly generate rectified images. DR-GAN [24] was the first to apply a generative adversarial network (GAN) [12] to fisheye image rectification, enabling the network to directly generate rectified images without estimating additional parameters. DDM [25] introduced distortion maps to help the model better learn the distortion distribution. PCN [52] designed a flow estimation module to predict appearance flows in fisheye images, using it to assist the rectification progress. SimFIR [10] introduced a self-supervised rectification module, allowing the network to better learn the distortion representations at different locations. However, both paradigms require retraining when handling images with different distortion degrees, failing to address the issue of weak model generalization. + +# 2.3 Controllable Low-level Vision + +To address the issue of varying degradation levels in low-level vision tasks, an increasing number of studies [7,15,16,33,47,48,55,61] propose controllable network architectures to tackle this challenge. DNI [48] interpolated all parameters of different restoration networks, which were trained with different degradation levels. By adjusting the interpolation coefficients, a smooth control of the image can be achieved. AdaFM [15] achieved better results by inserting AdaFM layers after each convolution layer to change the filters' statistics, thus the users can interactively manipulate the restoration results by tuning a control coefficient. CFS-Net [47] introduced the tuning branch to adaptively learn the control coefficients, + +![](images/acb5866890a2dd5ad447ffb26a49a8749b43d0162695c1893cd98db73fc9af47.jpg) +Fig. 2: Overview of our proposed Query-based Controllable Distortion Rectification network (QueryCDR). The Distortion-aware Learnable Query Mechanism (DLQM) extracts control conditions from user-given queries and feeds them layer by layer into the rectification network. The rectification network is composed of Controllable Convolution Modulating Blocks (CCMB) and Controllable Attention Modulating Blocks (CAMB), which modulate the input features $F_{in}$ with control conditions $F_{c}$ , enabling controllable rectification process. + +and then use them to couple the features with the main branch. MM-RealSR [33] proposed a metric learning strategy to map unquantifiable degradation levels to a metric space as control conditions. However, due to the different optimization objectives and complexity of the control mechanisms, directly applying these methods to fisheye image distortion rectification tasks will not achieve effective control over the rectification process. + +# 3 Methodology + +# 3.1 Overview + +The overview of our proposed Query-based Controllable Distortion Rectification network (QueryCDR) is shown in Fig. 2. First, following existing work [52], the input fisheye image is fed into the flow estimation module to obtain the appearance flow, which performs a coarse-grained rectification for image features (i.e., warping $\omega(\cdot)$ ). Then, the Distortion-aware Learnable Query Mechanism (DLQM) (in Sec. 3.2) extracts control conditions from the user-given query and feeds them layer by layer into the rectification network. Finally, a U-shaped hierarchical network composed of several controllable modulating blocks (in Sec. 3.3) is used to rectify the distorted input image. These blocks modulate features with the control conditions given by DLQM, to get the final output. + +In the following, we will provide a comprehensive description of each module and its corresponding role in the QueryCDR. + +![](images/ce06cff48d09e005246042699b655d3fa7a47e92a20038f2ae663b985c4385dd.jpg) +Fig. 3: Given an input image, QueryCDR can accurately produce results with different rectification degrees by feeding different queries. Moreover, by interpolating between different queries, we can achieve smooth continuous rectification for any distortion degree. For examples, $Q_{1.25} = 0.75Q_1 + 0.25Q_2$ , and $Q_{8.5} = 0.5Q_8 + 0.5Q_9$ . + +# 3.2 Distortion-aware Learnable Query Mechanism (DLQM) + +To achieve control over the rectification process, most existing controllable methods [33,47,55] introduce scalars as control conditions to represent the degradation degrees for image reconstruction. However, due to the peculiarities of distortion in fisheye images [10, 42, 53], these methods fail to achieve effective control over the rectification process. To address this issue, we propose the DLQM. It maintains a learnable query set, providing diverse effective control conditions for the rectification network. During training, DLQM projects the learned high-dimensional positional mapping relationship into a low-dimensional latent space of queries. During inference, DLQM extracts the control information from the queries and converts it into corresponding control conditions for each layer of the rectification network. Simply by providing different queries, DLQM can effectively control the rectification process, as shown in Fig. 3. At the same time, our QueryCDR can also achieve smooth continuous rectification for any distortion degrees by simply interpolating between different queries. + +Specifically, we construct a query set $\mathbf{Q}_s = \{\bar{Q}_i\mid Q_i\in \mathbb{R}^{C_{in}\times H_{in}\times W_{in}},i = 1,2,\ldots ,N\}$ , which comprises $N$ queries representing different distortion degrees. $H_{in}$ , $W_{in}$ , and $C_{in}$ represent the query's height, width, and channel, respectively. Each query has the same size as the input image $I_{in}\in \mathbb{R}^{C_{in}\times H_{in}\times W_{in}}$ . During inference, users select a query $Q_{i}$ from $\mathbf{Q}_s$ and feed it into the DLQM. In DLQM, the control extracting part $\mathrm{CE}(\cdot)$ is firstly used to extract the features in $Q_{i}$ . The input processing part can be expressed as follows, + +$$ +Q _ {e x} = \operatorname {C E} \left(Q _ {i}\right), \tag {1} +$$ + +where $Q_{ex} \in \mathbb{R}^{C_{in} \times H_{in} \times W_{in}}$ is the extracted feature, and serves as the input to the first control layer of DLQM. To minimize the computational costs and effectively extract the control information, $\mathrm{CE}(\cdot)$ is composed of three convolution layers with a kernel size of $3 \times 3$ . + +Subsequently, DLQM provides corresponding control conditions to the respective layers of the rectification network. To provide appropriate control conditions, each control layer of DLQM comprises two fully connected layers $\mathrm{FC}_1(\cdot)$ , $\mathrm{FC}_2(\cdot)$ . The $l$ -th layer of DLQM can be represented as follows, + +$$ +Q _ {c} ^ {l} = \mathrm {F C} _ {2} ^ {l} \left(\mathrm {F C} _ {1} ^ {l} \left(Q _ {c} ^ {l - 1}\right)\right), \tag {2} +$$ + +where $Q_{c}^{l-1}$ is the output control condition from the $(l-1)$ -th control layer. For $l = 1$ , $Q_{c}^{0}$ represents the extracted feature $Q_{ex}$ . The output $Q_{c}^{l}$ of the $l$ -th control layer is fed into the $l$ -th rectification layer as a control condition, and simultaneously serves as input to the $(l+1)$ -th control layer of DLQM. + +Afterward, we feed the control conditions into the rectification network. Each layer's controllable modulating block modulates the input feature with the control condition, guiding the modulated features toward the desired direction. + +# 3.3 Controllable Modulating Block + +When controlling the rectification process, to avoid obtaining results with blurred texture details or residual distortions, we introduce two types of controllable modulating blocks tailored for reconstructing local texture details and learning continuous distortion patterns. They are denoted as the Controllable Convolution Modulating Block (CCMB) based on CNN [17] (in Fig. 2(a)) and the Controllable Attention Modulating Block (CAMB) based on Transformer [45] (in Fig. 2(b)). The CCMB can adaptively fuse the original features and the controlled features, preserving more local details. The CAMB can better capture the spatial distortion information, guaranteeing the integrity of the recovered content [10,35,53]. To balance performance and computational cost, our QueryCDR constructs a U-shaped rectification network [52] composed of CCMB and CAMB. In the first three layers with larger feature maps, i.e., $l = \{1,2,3,9,10,11\}$ , CCMB is employed to learn more local texture details. In the remaining layers with smaller feature maps, i.e., $l = \{4,5,6,7,8\}$ , CAMB is utilized to capture more global dependencies within the images. + +Controllable Convolution Modulating Block (CCMB). When incorporating control conditions into the rectification network, existing methods [47,55] fail to effectively balance the fusion ratio between original features and controlled features. Directly using the controlled features or fusing them with original features at a fixed ratio will degrade the quality of the rectified images. To harmoniously incorporate control conditions into the rectification process, we design the CCMB, as shown in Fig. 2(a). CCMB can dynamically find an optimal ratio to modulate the distorted features with control conditions. + +Specifically, CCMB receives an input feature $F_{in} \in \mathbb{R}^{C \times H \times W}$ and a control condition $Q_{c} \in \mathbb{R}^{C \times H \times W}$ . Both of them are used to predict the fusion ratio. Here we omit the layer information for brevity, + +$$ +\theta = \operatorname {C P} \left(F _ {i n}, Q _ {c}\right), \tag {3} +$$ + +where $\mathrm{CP}(\cdot)$ is the coefficient predictor composed of two fully connected layers, takes the concatenation of $F_{in}$ and $Q_{c}$ as input and predicts the fusion ratio $\theta$ . + +Then, we perform element-wise multiplication $\otimes$ on $F_{in}$ and $Q_{c}$ to yield the controlled features $F_{c} \in \mathbb{R}^{C \times H \times W}$ , which can be expressed as, + +$$ +F _ {c} = F _ {i n} \otimes Q _ {c}. \tag {4} +$$ + +Finally, the input features $F_{in}$ and controlled features $F_{c}$ are combined in a weighted sum according to the fusion ratio $\theta$ , illustrated as $\psi(\cdot)$ in Fig. 2(a), + +$$ +F _ {o u t} = \psi \left(F _ {c}, F _ {i n}, \theta\right) = \theta F _ {c} \oplus (1 - \theta) F _ {i n}, \tag {5} +$$ + +where $\oplus$ represents the element-wise addition, and $F_{out} \in \mathbb{R}^{C \times H \times W}$ is the final output of CCMB. + +Due to this dynamic modulation mechanism, CCMB achieves effective control over the rectification process while preserving richer texture details. + +Controllable Attention Modulating Block (CAMB). Due to the lack of perception of global information, CNN-based networks struggle to learn long-range dependencies, particularly the continuous and amorphous distortions prevalent in fisheye distortions [10, 42, 53]. In contrast, Transformer-based networks effectively compensate for this, with their global attention mechanism [13, 35]. + +Therefore, we propose the CAMB, as illustrated in Fig. 2(b). To optimally leverage the control conditions given by DLQM, we designed the control-attention mechanism, enabling CAMB to perceive the global spatial relationships in the control conditions effectively. Specifically, we unfold and project the controlled feature $F_{c}$ (in Eq. (4)) as the query $\mathcal{Q} \in \mathbb{R}^{m \times L}$ , with the input feature $F_{in}$ as the key $\mathcal{K} \in \mathbb{R}^{m \times L}$ and value $\mathcal{V} \in \mathbb{R}^{m \times L}$ , where $L = H \times W$ represents the sequence length, and $m$ denotes the dimensions of the sequences. The control-attention is described by the following, + +$$ +\begin{array}{l} \mathcal {Q} = W _ {\mathcal {Q}} F _ {c}, \mathcal {K} = W _ {\mathcal {K}} F _ {i n}, \mathcal {V} = W _ {\mathcal {V}} F _ {i n}, \\ \operatorname {C T R L - A T T N} (\mathcal {Q}, \mathcal {K}, \mathcal {V}) = \operatorname {s o f t m a x} \left(\frac {\mathcal {Q} \mathcal {K} ^ {T}}{\sqrt {m}}\right) \mathcal {V}, \tag {6} \\ \end{array} +$$ + +where $W_{\mathcal{Q}}, W_{\mathcal{K}}, W_{\mathcal{V}} \in \mathbb{R}^{m \times C}$ represent the projection matrices of the queries, keys, and values, respectively. CTRL-ATTN( $\cdot$ ) represents the control-attention we proposed. The overall computation of CAMB can be formulated as follows, + +$$ +\begin{array}{l} F _ {a} = \operatorname {C T R L - A T T N} (\mathrm {L N} (\mathcal {Q}, \mathcal {K}, \mathcal {V})) \oplus F _ {i n}, \\ F _ {o u t} = \operatorname {C o n v} _ {1 \times 1} (\operatorname {F F N} (\operatorname {L N} (F _ {a})) \oplus F _ {a}), \tag {7} \\ \end{array} +$$ + +where $F_{in}$ means the input feature, $F_{a} \in \mathbb{R}^{m \times H \times W}$ and $F_{out} \in \mathbb{R}^{C \times H \times W}$ are the outputs of CTRL-ATTN( $\cdot$ ) and CAMB, respectively. $\mathrm{LN}(\cdot)$ denotes the layer normalization [1]. $\mathrm{FFN}(\cdot)$ stands for the feed-forward network composed of three fully connected layers, which helps CAMB to focus on the global dependencies. + +CAMB can discern the global mapping relationships within fisheye images, ensuring rectification uniformity compared to the CNN-based networks. + +# 3.4 Training Strategy + +To guarantee the robust and stable controllable distortion rectification of Query-CDR, we design a two-stage training strategy that combines coarse-grained distortion pre-training and fine-grained distortion fine-tuning. + +During the coarse-grained distortion pre-training phase, we choose the most commonly used dataset [10,24,52,54] that only contains one degree of distortion to train our QueryCDR. For clarity, we denote the distortion degree in this phase as $d$ . Correspondingly, only one query, denoted as $Q$ , is used for training. The optimization objective can be expressed as follows, + +$$ +\mathcal {L} _ {p r e} = \mathcal {L} _ {r} + \mathcal {L} _ {m}, \tag {8} +$$ + +where $\mathcal{L}_{pre}$ is the overall loss function for the pre-training phase. $\mathcal{L}_r$ denotes the reconstruction loss, + +$$ +\mathcal {L} _ {r} = \left\| I _ {o u t} ^ {d} - I _ {g t} ^ {d} \right\| _ {1}, \tag {9} +$$ + +where $I_{out}^{d}$ and $I_{gt}^{d}$ signifie the output result and ground truth, respectively. $\mathcal{L}_m$ denotes the multi-scale loss, + +$$ +\mathcal {L} _ {m} = \sum_ {j = 1} ^ {Z - 1} \| S \left(I _ {g t} ^ {d}, j\right) - C \left(F _ {\text {o u t}} ^ {j}\right) \| _ {1}, \tag {10} +$$ + +where $S(\cdot)$ represents the operation that down-samples the input $I_{gt}^{d}$ by a factor of $1/2^{j}$ . $Z$ represents the number of decoder's layers, and we set $Z$ to 6. $F_{out}^{j}$ denotes the feature in $j$ -th decoder layer. $C(\cdot)$ is $3 \times 3$ convolution for decoding the features into 3-channel RGB images. In this way, each feature map on the decoder can be effectively supervised. The $I_{out}^{d}$ can be obtained as, + +$$ +I _ {o u t} ^ {d} = \operatorname {Q u e r y C D R} \left(I _ {i n} ^ {d}, Q\right), \tag {11} +$$ + +where $I_{in}^{d}$ represents fisheye images with distortion degree $d$ as input. This way can effectively boost the model stability and accelerate the convergence. + +During the fine-grained distortion fine-tuning phase, we use the varying distortion degrees datasets to fine-tune our QueryCDR. Before training, We replicate the weight of the query $Q$ to the other queries to accelerate convergence on other distortion degrees. Subsequently, we fine-tune our QueryCDR using fisheye images with varying distortion degrees, and feeding corresponding queries into QueryCDR for training at the same time. This allows the query set to efficiently acquire diverse latent spatial relationships. The optimization objective can be expressed as follows, + +$$ +\mathcal {L} _ {\text {f i n e}} ^ {d _ {i}} = \mathcal {L} _ {r} ^ {d _ {i}} + \mathcal {L} _ {m} ^ {d _ {i}}, \tag {12} +$$ + +where $\mathcal{L}_{fine}^{d_i}$ is the overall fine-tuning loss function for distortion degree $d_{i}, i \in \{1,2,\dots,9\}$ , which is similar to the $\mathcal{L}_{pre}$ but calculated across different distortion degrees. The $I_{out}^{d_i}$ can be obtained as, + +$$ +I _ {\text {o u t}} ^ {d _ {i}} = \operatorname {Q u e r y C D R} \left(I _ {\text {i n}} ^ {d _ {i}}, Q _ {i}\right), \tag {13} +$$ + +where $I_{in}^{d_i}$ denotes fisheye images with a distortion degree of $d_i$ as input, $Q_i$ denotes the query corresponding to $d_i, i \in \{1, 2, \dots, 9\}$ . + +With this two-stage training strategy, our QueryCDR can effectively utilize only a small amount of varying distortion degrees data that is difficult to obtain, to rapidly finish the training of our network. + +Table 1: Quantitative comparison (PSNR (dB)↑, SSIM ↑) on COCO [27] fisheye image dataset with varying distortion degrees. Red indicates the best and blue indicates the second best performance (best viewed in color). + +
MethodPSNR
d1d2d3d4d5d6d7d8d9Avg
SC [40]10.0511.5911.8111.0311.5010.1911.1110.199.1410.73
DeepCalib [4]10.0010.6911.0111.1911.2811.4611.4511.1311.1111.04
Blind [23]12.9811.2410.309.629.9911.7512.6912.7512.8011.57
DR-GAN [24]15.6817.4017.9718.3418.5018.4417.9517.9417.4717.74
PCN [52]14.9317.4318.4318.8618.8618.8818.7417.3518.2617.97
DDA [54]16.3917.4117.4319.4820.1218.9018.8518.1718.2218.33
SimFIR [10]16.5717.8818.4318.9719.3119.2819.1918.6518.4818.53
QueryCDR20.0120.2920.3920.4120.7220.8120.5819.1120.5320.32
MethodSSIM
d1d2d3d4d5d6d7d8d9Avg
SC [40]0.1010.1130.1490.1820.2830.1750.1410.1260.0930.151
DeepCalib [4]0.1840.2100.2230.2300.2340.2460.2500.2450.2460.229
Blind [23]0.3080.2440.1990.1760.1940.2960.3670.3950.4200.289
DR-GAN [24]0.2950.3300.3390.3440.3440.3320.3140.3120.2990.323
PCN [52]0.4200.5470.5890.6070.6080.6100.6150.5760.6030.575
DDA [54]0.4550.5890.5920.6200.6750.6260.6190.5640.5810.591
SimFIR [10]0.4920.5810.6260.6350.6400.6280.6220.5910.5950.601
QueryCDR0.6430.6650.6680.6770.6880.6990.6920.6560.6930.676
+ +# 4 Experiment + +# 4.1 Experimental Settings + +To demonstrate the effectiveness of our proposed QueryCDR, we followed the existing works [10, 24, 52, 54], employed the four-parameter polynomial model to synthesize fisheye images. We constructed synthetic datasets based on the original images of COCO [27] and Places2 [62] datasets, respectively. Specifically, for images with distortion degree $d$ , 40,000 images were used for the pre-training stage. And for images with varying distortion degrees $d_{i}$ , $i \in \{1,2,\dots ,9\}$ , 18,000 images were used for the fine-tuning stage, and 3,600 images for testing. The images were resized to $256 \times 256$ when fed into the network. For a fair comparison, we followed existing works [10, 24, 52, 54] utilizing a batch size of 16 and the Adam [19] optimizer with a learning rate of 1e-4. + +# 4.2 Performance Evaluation + +To evaluate the performance of our method, we retrained and validated existing fisheye image rectification methods on synthetic fisheye datasets with 9 distortion degrees. The methods can be summarized into three categories: traditional + +![](images/1fe983ba68f6777d12ebc469e5cfda3a4141d0d3ad7cef8eedd2c622d3b9a149.jpg) +Fig. 4: Qualitative results on synthetic fisheye images. + +method SC [40], regression-based methods DeepCalib [4] and Blind [23], and generation-based DR-GAN [24], PCN [52], DDA [54] and SimFIR [10]. For a fair comparison, we followed existing works [10,24,52,54] employing Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) to quantify performance. + +Quantitative Results. Our performance comparisons are shown in Tab. 1. Compared to existing methods, QueryCDR achieves the best performance across all distortion degrees without retraining. It is because DLQM effectively utilizes the control information in queries to guide the network in achieving varying degrees of rectification. Additionally, the capability of CCMB and CAMB allows for controlled rectification while obtaining high-quality results. Moreover, we observe that even on images with distortion degree $d_{5}$ , which the existing + +![](images/6378eef9c633bf24a573b34d50f19ace4590a5bce813bf929c82cc2d99278a98.jpg) +Fig. 5: Qualitative results on real-world fisheye images. + +methods perform well [10, 24, 52, 54], QueryCDR still outperforms the second best method [54] by 0.60 dB of PSNR and 0.013 of SSIM. This further demonstrates the outstanding generalization ability of QueryCDR. + +Qualitative Results. To further compare the visual qualities of different methods, we show the results rectified by QueryCDR and other rectification methods in Fig. 4. It can be observed intuitively that other methods fail to effectively rectify images with varying degrees of distortion. Particularly the objects in images, our QueryCDR maintains the accurate structure of objects across different distortions. Furthermore, benefiting from the modulation capability of CCMB and CAMB, QueryCDR preserves richer texture details after rectification. + +In addition, to further demonstrate the generalization ability of QueryCDR, we conducted experiments on real-world fisheye image datasets [57] in Fig. 5. Despite the disparities between synthetic and real-world datasets, our QueryCDR still shows robust rectification capabilities, and preservers richer texture details. These results further validate its practicality in real-world scenarios. + +# 4.3 Ablation Study + +In this section, we conduct ablation for different control conditions, and study the effect of the controllable modulating blocks and the architecture setting. + +Distortion-aware Learnable Query Mechanism (DLQM). To validate the effectiveness of DLQM in controlling the rectification process, we implemented different rectification networks controlled by scalar, fixed query, and our learnable query in Tab. 2. When using scalar to control, the value of the scalar increments with the distortion of the entire image. When using fixed query to control, the parameters of the query are set to incremental values from 0 to 1 with the degree of distortion from the center to the edges. Following the settings in Sec. 4.1, we use the learnable query $Q_{i}, i \in \{1,2,\dots ,9\}$ to learn the distortion distribution corresponding to $d_{i}, i \in \{1,2,\dots ,9\}$ . + +We trained different methods using the same strategy as illustrated in Sec. 3.4, the results are shown in Tab. 2 and Fig. 6. All three controllable methods outperform the uncontrollable method significantly at various degrees of distortion. + +Table 2: Performance comparison of different control methods. + +
MethodPSNR
d1d2d3d4d5d6d7d8d9Avg
W/o Control14.9317.4318.4318.8618.8618.8818.7417.3518.2617.97
Scalar19.5220.2519.9719.4319.8920.0720.0318.7620.1519.78
Fixed Query20.1320.0320.1419.8820.1620.2520.1618.7419.8419.93
Learnable Query20.0120.2920.3920.4120.7220.8120.5819.1120.5320.32
+ +![](images/d922fd891d1c1c34348f1b1205873cb07c3cb862fbbf9aeb9cf7a4a7ca9b1350.jpg) +Fig. 6: Visual results of different control mechanisms. + +![](images/332ab2ce251515723329b0833b584c06fb71ed31c35b01002ee8aca053da6309.jpg) +Fig.7: Visual results of different modulation methods. + +This is because the controllable mechanism effectively assists the rectification network in distinguishing between different degrees of distortion, thereby enhancing the generalization capability of the rectification network. The improvement brought by scalar demonstrates the crucial role of controllable mechanisms in handling fisheye images with different degrees of distortion. The further enhancement with fixed query validates the importance of employing higher-dimensional control conditions for fisheye image rectification. Lastly, our tailored learnable query significantly outperforms all other controllable methods, verifying the unique superiority of our approach in controllable rectification. + +Controllable Modulating Block. To demonstrate the capability of the CCMB and CAMB in modulating input features with control conditions, we implemented two comparison methods with different modulating approaches: one directly uses the controlled feature (in Eq. (4)), and the other adds controlled feature and original feature in a fixed 1:1 ratio. We constructed different controllable networks using each of these methods. + +As shown in Tab. 3 and Fig. 7, retaining original features and fusing them with a 1:1 ratio improved by $0.03\mathrm{dB}$ , highlighting the importance of preserving original features. The dynamic modulation mechanism boosted performance by $0.06\mathrm{dB}$ , verifying its effectiveness in dynamic feature modulation. Furthermore, leveraging the control-attention mechanism led to a performance improvement of $0.12\mathrm{dB}$ , demonstrating the superiority of CAMB for perceiving global distortion distribution, and achieving better uniform global rectification. + +Table 3: Ablation study of different modulation methods. Where Direct Use $F_{c}$ means directly using the controlled feature $F_{c}$ , Fixed Ratio means adding controlled features and original features in a 1:1 ratio, Dynamic Mechanism means the CCMB and Attention Mechanism means the CAMB. + +
MethodPSNRSSIM
Direct UseFc20.140.655
Fixed Ratio20.170.658
Dynamic Mechanism20.200.669
Attention Mechanism20.260.671
+ +Table 4: Ablation study of different network architectures. Where $x\mathrm{C} + \left( {{11} - x}\right) \mathrm{A}$ denotes that the network uses CCMB in the $x$ layers with larger feature maps and CAMB in the remaining ${11} - x$ layers with smaller feature maps. + +
MethodFlops(G)Param(M)PSNRSSIM
PCN [52]12.30535.63717.970.575
11C+0A12.73637.70120.200.669
8C+3A13.38346.39820.270.665
6C+5A12.35343.24420.320.676
4C+7A12.53846.99420.310.670
0C+11A15.19051.79520.260.671
+ +Controllable Rectification Network Architecture. As described in Sec. 3.3, rectification networks based on CCMB may ignore the long-range dependencies, while those based on CAMB may ignore the texture details and also increase computational costs. To strike a balance between these two architectures, We incorporate CCMB into the layers with larger feature map size, and integrate CAMB into the layers with smaller feature map size. To validate the effectiveness of this hybrid architecture and find an optimal combination, we conducted performance and computational cost comparisons between pure CCMB, pure CAMB, and multiple hybrid networks, as shown in Tab. 4. Pure CAMB network (i.e., Row 6) outperforms pure CCMB network (i.e., Row 2) in rectification performance but incurs higher computational costs. Hybrid network architectures effectively solve this problem, with higher performance and fewer parameters. After a trade-off between performance and parameters, we empirically choose the $6\mathrm{C} + 5\mathrm{A}$ hybrid architecture as our final model. + +# 5 Conclusion + +In this paper, we propose the Query-based Controllable Distortion Rectification network for fisheye images (QueryCDR), which achieves controllable rectification at different distortion degrees without retraining. In particular, we design a series of learnable queries as control conditions to guide the rectification process. Additionally, we design two different controllable modulating blocks, achieving controllable rectification while improving image quality. Extensive experiments have demonstrated the robustness and effectiveness of our QueryCDR. In the future, it is expected to further rectify distortions of different degrees by autocontrolling mechanisms to avoid human involvement. We believe our work provides an effective solution for fisheye camera applications. + +# Acknowledgements + +This work was supported in part by the NSFC under Grant 62272376, 62272380 and 62103317, the Key Research and Development Program of Shaanxi Province under Grant 2020ZDLGY04-05, the Fundamental Research Funds for the Central Universities, China (xzy022023051), the Innovative Leading Talents Scholarship of Xi'an Jiaotong University (Corresponding author: Xingsong Hou). + +# References + +1. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016) +2. Barreto, J.P., Araujo, H.: Geometric properties of central catadioptric line images and their application in calibration. IEEE Transactions on Pattern Analysis and Machine Intelligence 27(8), 1327-1333 (2005) +3. Barreto, J.P., Daniilidis, K.: Fundamental matrix for cameras with radial distortion. In: Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1. vol. 1, pp. 625-632. IEEE (2005) +4. Bogdan, O., Eckstein, V., Rameau, F., Bazin, J.C.: Deepcalib: A deep learning approach for automatic intrinsic calibration of wide field-of-view cameras. In: Proceedings of the 15th ACM SIGGRAPH European Conference on Visual Media Production. pp. 1-10 (2018) +5. Bukhari, F., Dailey, M.N.: Automatic radial distortion estimation from a single image. Journal of mathematical imaging and vision 45, 31-45 (2013) +6. Cai, B., Xu, X., Jia, K., Qing, C., Tao, D.: Dehazenet: An end-to-end system for single image haze removal. IEEE transactions on image processing 25(11), 5187-5198 (2016) +7. Cai, H., He, J., Qiao, Y., Dong, C.: Toward interactive modulation for photorealistic image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 294-303 (2021) +8. Devernay, F., Faugeras, O.: Straight lines have to be straight. Machine vision and applications 13, 14-24 (2001) +9. Duan, Z., Tezcan, O., Nakamura, H., Ishwar, P., Konrad, J.: Rapid: rotation-aware people detection in overhead fisheye images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 636-637 (2020) +10. Feng, H., Wang, W., Deng, J., Zhou, W., Li, L., Li, H.: Simfir: A simple framework for fisheye image rectification with self-supervised representation learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12418-12427 (2023) +11. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: 2012 IEEE conference on computer vision and pattern recognition. pp. 3354-3361. IEEE (2012) +12. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020) +13. Guo, J., Han, K., Wu, H., Tang, Y., Chen, X., Wang, Y., Xu, C.: Cmt: Convolutional neural networks meet vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12175-12185 (2022) + +14. Hartley, R., Kang, S.B.: Parameter-free radial distortion correction with center of distortion estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence 29(8), 1309-1321 (2007) +15. He, J., Dong, C., Qiao, Y.: Modulating image restoration with continual levels via adaptive feature modification layers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11056-11064 (2019) +16. He, J., Dong, C., Qiao, Y.: Interactive multi-dimension modulation with dynamic controllable residual learning for image restoration. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16. pp. 53-68. Springer (2020) +17. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +18. Henrique Brito, J., Angst, R., Koser, K., Pollefeys, M.: Radial distortion self-calibration. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1368-1375 (2013) +19. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) +20. Kukelova, Z., Pajdla, T.: A minimal solution to radial distortion autocalibration. IEEE Transactions on Pattern Analysis and Machine Intelligence 33(12), 2410-2422 (2011) +21. Kumar, V.R., Klingner, M., Yogamani, S., Milz, S., Fingscheidt, T., Mader, P.: Syndistnet: Self-supervised monocular fisheye camera distance estimation synergized with semantic segmentation for autonomous driving. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 61-71 (2021) +22. Ledig, C., Theis, L., Huszár, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., et al.: Photo-realistic single image superresolution using a generative adversarial network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4681-4690 (2017) +23. Li, X., Zhang, B., Sander, P.V., Liao, J.: Blind geometric distortion correction on images through deep learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4855-4864 (2019) +24. Liao, K., Lin, C., Zhao, Y., Gabbouj, M.: Dr-gan: Automatic radial distortion rectification using conditional gan in real-time. IEEE Transactions on Circuits and Systems for Video Technology 30(3), 725-733 (2019) +25. Liao, K., Lin, C., Zhao, Y., Xu, M.: Model-free distortion rectification framework bridged by distortion distribution map. IEEE Transactions on Image Processing 29, 3707-3718 (2020) +26. Lin, L., Lu, Y., Pan, Y., Chen, X.: Integrating graph partitioning and matching for trajectory analysis in video surveillance. IEEE transactions on Image Processing 21(12), 4844-4857 (2012) +27. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +28. Liu, C., Wang, X., Li, S., Wang, Y., Qian, X.: Fsi: Frequency and spatial interactive learning for image restoration in under-display cameras. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12537-12546 (2023) +29. Liu, C., Wang, X., Xu, X., Tian, R., Li, S., Qian, X., Yang, M.H.: Motion-adaptive separable collaborative filters for blind motion deblurring. In: Proceedings of the + +IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 25595-25605 (2024) +30. Liu, C., Yang, H., Fu, J., Qian, X.: Learning trajectory-aware transformer for video super-resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5687-5696 (2022) +31. Liu, C., Yang, H., Fu, J., Qian, X.: 4d lut: learnable context-aware 4d lookup table for image enhancement. IEEE Transactions on Image Processing 32, 4742-4756 (2023) +32. Mei, C., Rives, P.: Single view point omnidirectional camera calibration from planar grids. In: Proceedings 2007 IEEE International Conference on Robotics and Automation. pp. 3945-3950. IEEE (2007) +33. Mou, C., Wu, Y., Wang, X., Dong, C., Zhang, J., Shan, Y.: Metric learning based interactive modulation for real-world super-resolution. In: European Conference on Computer Vision. pp. 723-740. Springer (2022) +34. Muhammad, K., Ahmad, J., Lv, Z., Bellavista, P., Yang, P., Baik, S.W.: Efficient deep cnn-based fire detection and localization in video surveillance applications. IEEE Transactions on Systems, Man, and Cybernetics: Systems 49(7), 1419-1434 (2018) +35. Peng, Z., Huang, W., Gu, S., Xie, L., Wang, Y., Jiao, J., Ye, Q.: Conformer: Local features coupling global representations for visual recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 367-376 (2021) +36. Plaut, E., Ben Yaacov, E., El Shlomo, B.: 3d object detection from a single fisheye image without a single fisheye training image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3659-3667 (2021) +37. Puig, L., Bastanlar, Y., Sturm, P., Guerrero, J.J., Barreto, J.: Calibration of central catadioptric cameras using a dlt-like approach. International Journal of Computer Vision 93, 101-114 (2011) +38. Rashed, H., Mohamed, E., Sistu, G., Kumar, V.R., Eising, C., El-Sallab, A., Yogamani, S.: Generalized object detection on fisheye cameras for autonomous driving: Dataset, representations and baseline. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2272-2280 (2021) +39. Rong, J., Huang, S., Shang, Z., Ying, X.: Radial lens distortion correction using convolutional neural networks trained with synthesized images. In: Computer Vision-ACCV 2016: 13th Asian Conference on Computer Vision, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part III 13. pp. 35-49. Springer (2017) +40. Santana-Cedrés, D., Gomez, L., Aleman-Flores, M., Salgado, A., Esclarín, J., Mazorra, L., Alvarez, L.: An iterative optimization algorithm for lens distortion correction using two-parameter models. Image Processing On Line 6, 326-364 (2016) +41. Scaramuzza, D., Martinelli, A., Siegwart, R.: A flexible technique for accurate omnidirectional camera calibration and structure from motion. In: Fourth IEEE International Conference on Computer Vision Systems (ICVS'06). pp. 45-45. IEEE (2006) +42. Shen, Z., Lin, C., Liao, K., Nie, L., Zheng, Z., Zhao, Y.: Panoformer: Panorama transformer for indoor 360 depth estimation. In: European Conference on Computer Vision. pp. 195-211. Springer (2022) +43. Sturm, P., Ramalingam, S.: A generic concept for camera calibration. In: Computer Vision-ECCV 2004: 8th European Conference on Computer Vision, Prague, Czech Republic, May 11-14, 2004. Proceedings, Part II 8. pp. 1-13. Springer (2004) +44. Thormahlen, T., Broszio, H., Wassermann, I.: Robust line-based calibration of lens distortion from a single view. Mirage 2003 pp. 105-112 (2003) + +45. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) +46. Wang, A., Qiu, T., Shao, L.: A simple method of radial distortion correction with centre of distortion estimation. Journal of Mathematical Imaging and Vision 35, 165-172 (2009) +47. Wang, W., Guo, R., Tian, Y., Yang, W.: Cfsnet: Toward a controllable feature space for image restoration. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4140-4149 (2019) +48. Wang, X., Yu, K., Dong, C., Tang, X., Loy, C.C.: Deep network interpolation for continuous imagery effect transition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1692-1701 (2019) +49. Wang, X., Yu, K., Wu, S., Gu, J., Liu, Y., Dong, C., Qiao, Y., Change Loy, C.: Esrgan: Enhanced super-resolution generative adversarial networks. In: Proceedings of the European conference on computer vision (ECCV) workshops. pp. 0-0 (2018) +50. Xue, Z., Xue, N., Xia, G.S., Shen, W.: Learning to calibrate straight lines for fisheye image rectification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1643-1651 (2019) +51. Yang, C.Y., Chen, H.H.: Efficient face detection in the fisheye image domain. IEEE Transactions on Image Processing 30, 5641-5651 (2021) +52. Yang, S., Lin, C., Liao, K., Zhang, C., Zhao, Y.: Progressively complementary network for fisheye image rectification using appearance flow. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6348-6357 (2021) +53. Yang, S., Lin, C., Liao, K., Zhao, Y.: Fishformer: Annulus slicing-based transformer for fisheye rectification with efficacy domain exploration. arXiv preprint arXiv:2207.01925 (2022) +54. Yang, S., Lin, C., Liao, K., Zhao, Y.: Dual diffusion architecture for fisheye image rectification: Synthetic-to-real generalization. arXiv preprint arXiv:2301.11785 (2023) +55. Yao, M., He, D., Li, X., Li, F., Xiong, Z.: Towards interactive self-supervised denoising. IEEE Transactions on Circuits and Systems for Video Technology (2023) +56. Yin, X., Wang, X., Yu, J., Zhang, M., Fua, P., Tao, D.: Fisheyerecnet: A multicontext collaborative deep network for fisheye image rectification. In: Proceedings of the European conference on computer vision (ECCV). pp. 469-484 (2018) +57. Yogamani, S., Hughes, C., Horgan, J., Sistu, G., Varley, P., O'Dea, D., Uricár, M., Milz, S., Simon, M., Amende, K., et al.: Woodscape: A multi-task, multicamera fisheye dataset for autonomous driving. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9308-9318 (2019) +58. Zhang, K., Zuo, W., Chen, Y., Meng, D., Zhang, L.: Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing 26(7), 3142-3155 (2017) +59. Zhang, M., Yao, J., Xia, M., Li, K., Zhang, Y., Liu, Y.: Line-based multi-label energy optimization for fisheye image rectification and calibration. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4137-4145 (2015) +60. Zhang, Y., You, S., Gevers, T.: Automatic calibration of the fisheye camera for egocentric 3d human pose estimation from a single image. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1772-1781 (2021) + +61. Zhang, Z., Jiang, Y., Shao, W., Wang, X., Luo, P., Lin, K., Gu, J.: Real-time controllable denoising for image and video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14028-14038 (2023) +62. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence 40(6), 1452-1464 (2017) \ No newline at end of file diff --git a/2024/QueryCDR_ Query-based Controllable Distortion Rectification Network for Fisheye Images/images.zip b/2024/QueryCDR_ Query-based Controllable Distortion Rectification Network for Fisheye Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1ec36aa485af324ca885cf5ef161f6280a4ffe78 --- /dev/null +++ b/2024/QueryCDR_ Query-based Controllable Distortion Rectification Network for Fisheye Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4b03174f1857bd3c7a9e9848c1dcb937780be489c702d4e8660d24a22641cb22 +size 764085 diff --git a/2024/QueryCDR_ Query-based Controllable Distortion Rectification Network for Fisheye Images/layout.json b/2024/QueryCDR_ Query-based Controllable Distortion Rectification Network for Fisheye Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8075288a5b8511504c915b781c61c862b492649c --- /dev/null +++ b/2024/QueryCDR_ Query-based Controllable Distortion Rectification Network for Fisheye Images/layout.json @@ -0,0 +1,10040 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 111, + 478, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 111, + 478, + 148 + ], + "spans": [ + { + "bbox": [ + 136, + 111, + 478, + 148 + ], + "type": "text", + "content": "QueryCDR: Query-based Controllable Distortion Rectification Network for Fisheye Images" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 171, + 167, + 441, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 167, + 441, + 193 + ], + "spans": [ + { + "bbox": [ + 171, + 167, + 441, + 193 + ], + "type": "text", + "content": "Pengbo Guo" + }, + { + "bbox": [ + 171, + 167, + 441, + 193 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 171, + 167, + 441, + 193 + ], + "type": "text", + "content": ", Chengxu Liu" + }, + { + "bbox": [ + 171, + 167, + 441, + 193 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 171, + 167, + 441, + 193 + ], + "type": "text", + "content": ", Xingsong Hou" + }, + { + "bbox": [ + 171, + 167, + 441, + 193 + ], + "type": "inline_equation", + "content": "^{2(\\boxtimes)}" + }, + { + "bbox": [ + 171, + 167, + 441, + 193 + ], + "type": "text", + "content": ", and Xueming Qian" + }, + { + "bbox": [ + 171, + 167, + 441, + 193 + ], + "type": "inline_equation", + "content": "^{2,3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 154, + 201, + 457, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 201, + 457, + 223 + ], + "spans": [ + { + "bbox": [ + 154, + 201, + 457, + 223 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 154, + 201, + 457, + 223 + ], + "type": "text", + "content": " School of Software Engineering, Xi'an Jiaotong University, Xi'an, China \n" + }, + { + "bbox": [ + 154, + 201, + 457, + 223 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 154, + 201, + 457, + 223 + ], + "type": "text", + "content": " Xi'an Jiaotong University, Xi'an, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 224, + 474, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 474, + 246 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 474, + 246 + ], + "type": "text", + "content": "3 Shaanxi Yulan Jiuzhou Intelligent Optoelectronic Tech. Co., Ltd, Xi'an, China {guopengbo,chengxuliu}@stu.xjtu.edu.cn, {houxs,qianxm}@mail.xjtu.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 274, + 455, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 274, + 455, + 517 + ], + "spans": [ + { + "bbox": [ + 159, + 274, + 455, + 517 + ], + "type": "text", + "content": "Abstract. Fisheye image rectification aims to correct distortions in images taken with fisheye cameras. Although current models show promising results on images with a similar degree of distortion as the training data, they will produce sub-optimal results when the degree of distortion changes and without retraining. The lack of generalization ability for dealing with varying degrees of distortion limits their practical application. In this paper, we take one step further to enable effective distortion rectification for images with varying degrees of distortion without retraining. We propose a novel Query-based Controllable Distortion Rectification network for fisheye images (QueryCDR). In particular, we first present the Distortion-aware Learnable Query Mechanism (DLQM), which defines the latent spatial relationships for different distortion degrees as a series of learnable queries. Each query can be learned to obtain position-dependent rectification control conditions, providing control over the rectification process. Then, we propose two kinds of controllable modulating blocks to enable the control conditions to guide the modulation of the distortion features better. These core components cooperate with each other to effectively boost the generalization ability of the model at varying degrees of distortion. Extensive experiments on fisheye image datasets with different distortion degrees demonstrate our approach achieves high-quality and controllable distortion rectification. Code is available at https://github.com/PbGuo/QueryCDR." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 525, + 432, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 525, + 432, + 537 + ], + "spans": [ + { + "bbox": [ + 160, + 525, + 432, + 537 + ], + "type": "text", + "content": "Keywords: Fisheye image " + }, + { + "bbox": [ + 160, + 525, + 432, + 537 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 525, + 432, + 537 + ], + "type": "text", + "content": " Distortion rectification " + }, + { + "bbox": [ + 160, + 525, + 432, + 537 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 525, + 432, + 537 + ], + "type": "text", + "content": " Controllable" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 557, + 230, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 230, + 570 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 230, + 570 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "type": "text", + "content": "Benefiting from the huge field-of-view (FoV), fisheye cameras are widely utilized in various fields, including security surveillance [26, 34] and autonomous driving [11, 38]. However, the distortion brought by the fisheye lenses greatly limits the performance of downstream vision tasks [9, 21, 36, 60]. How to eliminate the distortion in fisheye images has attracted great attention in recent years." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Early methods [3,5,14,18,37,40,46,59] primarily relied on identifying matching feature points or curves for automatic rectification. However, constrained by" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 113, + 480, + 223 + ], + "blocks": [ + { + "bbox": [ + 132, + 113, + 480, + 223 + ], + "lines": [ + { + "bbox": [ + 132, + 113, + 480, + 223 + ], + "spans": [ + { + "bbox": [ + 132, + 113, + 480, + 223 + ], + "type": "image", + "image_path": "411c437a400b17158787ed04a95c04437841e28263fb6a1ccba12b2d37dc9705.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 232, + 482, + 299 + ], + "lines": [ + { + "bbox": [ + 130, + 232, + 482, + 299 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 482, + 299 + ], + "type": "text", + "content": "Fig. 1: Different approaches to fisheye image distortion rectification. (a) Regression-Based: Using a neural network to predict distortion-related parameters, then apply rectification algorithms " + }, + { + "bbox": [ + 130, + 232, + 482, + 299 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 232, + 482, + 299 + ], + "type": "text", + "content": " for rectification. (b) Generation-Based: Input the distorted fisheye image and directly generate the rectified image end-to-end. (c) Control-Based: Users provide control conditions to guide the rectification process, resulting in promising rectified images of various distortion degrees." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 326, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 482, + 434 + ], + "type": "text", + "content": "the instability of feature detection, both the generalization of the algorithm and the quality of the rectified images were unsatisfactory [10, 24, 52]. In recent years, owing to the robust learning and generalization capabilities of neural networks, deep learning-based fisheye image rectification methods have become mainstream. These methods can be divided into two categories: regression-based rectification [4, 23, 39, 50, 56] and generation-based rectification [10, 24, 25, 52, 54]. The former uses deep regression models to predict distortion parameters for image reconstruction, as shown in Fig. 1(a). The latter uses an encoder-decoder structure to generate well-rectified images directly, as shown in Fig. 1(b)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 436, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 567 + ], + "type": "text", + "content": "It is worth noting that these methods only achieve satisfactory results on a similar degree of distortion as the training data. This means that when handling different degrees of distortion without retraining, there will be a significant decrease in the quality of the rectified images [25]. This is because the model tends to learn fixed position mapping relationships during training. When the distortion degree changes, this relationship does not work for the new distortion distribution. Therefore, these models usually need to be retrained on images with different distortion degrees. In addition, fisheye image acquisition is difficult, and re-collecting new datasets with varying degrees of distortion would incur significant costs [52, 54]. Thus, it is essential to explore a model that can handle different degrees of distortion simultaneously." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "type": "text", + "content": "In recent years, some methods have been proposed to achieve controllable image restoration [7,15,16,33,47,48,55,61]. Typically, CFSNet [47], MM-RealSR [33], and Yao et al. [55] introduce scalars as control conditions to effectively restore degradation of varying degrees. It would be a promising solution to introduce an effectively controllable mechanism to deal with all degrees of distortion, as shown in Fig. 1(c). However, directly applying existing controllable mechanisms in restoration tasks to the fisheye rectification model suffers from the following challenges: 1) There is a gap between optimization objectives (i.e. the roles" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "P. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "of the controllable mechanisms). Restoration models learn the pixel-level detail restoration, whereas fisheye rectification models learn the spatial-level positional mapping relationships [10,25]. The control mechanism lacking position information in restoration tasks is ineffective in controlling distortion rectification networks. 2) There is a gap between optimization difficulties (i.e. the roles of the control conditions). The distortion of a fisheye image increases gradually from the image center to boundary [51]. Therefore, it is not suitable to use a single scalar control condition in restoration tasks to deal with such spatially varying distortions. These challenges restrict the application of controllable mechanisms and control conditions in fisheye image rectification." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 450 + ], + "type": "text", + "content": "To address these issues, we propose the Query-based Controllable Distortion Rectification network (QueryCDR), as shown in Fig. 2. By introducing a series of learnable queries as control conditions, QueryCDR allows the users to achieve fisheye rectification with different distortion degrees. Specifically, to incorporate positional mapping relationships into control conditions, we introduce the Distortion-aware Learnable Query Mechanism (DLQM), which defines a series of queries representing different rectification control conditions. During inference, DLQM extracts position-dependent control conditions from the user-given query and feeds them into the network for controlling the rectification process. Furthermore, to enable the control conditions to guide the rectification efficiently, we propose two types of controllable modulating blocks: the Controllable Convolution Modulating Block (CCMB) based on CNN [17], and the Controllable Attention Modulating Block (CAMB) based on Transformer [45]. They are good at extracting local texture features and learning long-range distortion mapping relationships, respectively. By combining CCMB and CAMB, we construct a robust controllable rectification network. Our QueryCDR can handle various distortions without retraining, enhancing the generalization ability of the fisheye rectification model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 147, + 451, + 339, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 451, + 339, + 462 + ], + "spans": [ + { + "bbox": [ + 147, + 451, + 339, + 462 + ], + "type": "text", + "content": "We summarize our contributions as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 467, + 480, + 585 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 467, + 480, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 467, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 467, + 480, + 502 + ], + "type": "text", + "content": "- We propose QueryCDR, a Query-based Controllable Distortion Rectification network for fisheye images. Extensive experiments demonstrate that our QueryCDR can deliver superior results on a variety of distortion degrees." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 503, + 480, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 503, + 480, + 537 + ], + "spans": [ + { + "bbox": [ + 138, + 503, + 480, + 537 + ], + "type": "text", + "content": "- We propose the Distortion-aware Learnable Query Mechanism (DLQM), which effectively introduces the latent spatial relationships to control conditions for fisheye image rectification." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 538, + 480, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 538, + 480, + 585 + ], + "spans": [ + { + "bbox": [ + 138, + 538, + 480, + 585 + ], + "type": "text", + "content": "- We propose two kinds of blocks for modulating features using control conditions: the Controllable Convolution Modulating Block (CCMB) and the Controllable Attention Modulating Block (CAMB). They can effectively utilize control conditions to guide the rectification process." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 601, + 237, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 237, + 613 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 237, + 613 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 624, + 358, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 358, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 358, + 635 + ], + "type": "text", + "content": "2.1 Traditional Fisheye Image Rectification" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Traditional rectification methods can be divided into two types, multi-view-based and line-based methods. Multi-view-based methods [3,14,18,20,37,41,43]" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 156, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 156, + 91, + 447, + 102 + ], + "type": "text", + "content": "QueryCDR: Query-based Controllable Distortion Rectification Network" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "calibrated fisheye images by finding corresponding feature points from multiple viewpoints. Line-based methods [2,5,8,32,40,44,46,59] employed line detection to rectify the curved lines, thereby achieving distortion rectification. However, these methods require manual intervention or handcrafted feature extractors, and the rectification process is unstable, failing to achieve satisfactory results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 193, + 411, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 193, + 411, + 205 + ], + "spans": [ + { + "bbox": [ + 131, + 193, + 411, + 205 + ], + "type": "text", + "content": "2.2 Deep Learning Based Fisheye Image Rectification" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 211, + 481, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 211, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 211, + 481, + 282 + ], + "type": "text", + "content": "Rapid advances in deep learning have allowed them to shine in low-level vision tasks [6, 22, 28-31, 49, 58]. As a core task in low-level vision, distortion image rectification has received increasing attention [10, 24, 39, 52, 54, 56]. According to the network architecture, deep learning-based distortion image rectification methods can be categorized into two types, regression-based methods [4, 23, 39, 50, 56] and generation-based methods [10, 24, 25, 52, 54]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 283, + 482, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 523 + ], + "type": "text", + "content": "Regression-based methods [4, 23, 39, 50, 56] utilize neural networks to predict distortion-related coefficients for rectifying the image. Rong et al. [39] were the first to use CNNs for fisheye image rectification. They employed the network to predict across multiple distortion intervals, achieving preliminary rectification results. Yin et al. [56] integrated semantics as prior information to guide rectification. While these methods have shown some performance improvement, they face the challenge of non-end-to-end design, thus requiring a large number of additional labels, and increasing operational costs. Therefore, to reduce models' complexity, Generation-based methods [10, 24, 25, 52, 54] employ generative networks to take distorted images as input and directly generate rectified images. DR-GAN [24] was the first to apply a generative adversarial network (GAN) [12] to fisheye image rectification, enabling the network to directly generate rectified images without estimating additional parameters. DDM [25] introduced distortion maps to help the model better learn the distortion distribution. PCN [52] designed a flow estimation module to predict appearance flows in fisheye images, using it to assist the rectification progress. SimFIR [10] introduced a self-supervised rectification module, allowing the network to better learn the distortion representations at different locations. However, both paradigms require retraining when handling images with different distortion degrees, failing to address the issue of weak model generalization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 539, + 309, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 539, + 309, + 551 + ], + "spans": [ + { + "bbox": [ + 131, + 539, + 309, + 551 + ], + "type": "text", + "content": "2.3 Controllable Low-level Vision" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "To address the issue of varying degradation levels in low-level vision tasks, an increasing number of studies [7,15,16,33,47,48,55,61] propose controllable network architectures to tackle this challenge. DNI [48] interpolated all parameters of different restoration networks, which were trained with different degradation levels. By adjusting the interpolation coefficients, a smooth control of the image can be achieved. AdaFM [15] achieved better results by inserting AdaFM layers after each convolution layer to change the filters' statistics, thus the users can interactively manipulate the restoration results by tuning a control coefficient. CFS-Net [47] introduced the tuning branch to adaptively learn the control coefficients," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "P. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 482, + 278 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 278 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 278 + ], + "type": "image", + "image_path": "acb5866890a2dd5ad447ffb26a49a8749b43d0162695c1893cd98db73fc9af47.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 286, + 482, + 364 + ], + "lines": [ + { + "bbox": [ + 130, + 286, + 482, + 364 + ], + "spans": [ + { + "bbox": [ + 130, + 286, + 482, + 364 + ], + "type": "text", + "content": "Fig. 2: Overview of our proposed Query-based Controllable Distortion Rectification network (QueryCDR). The Distortion-aware Learnable Query Mechanism (DLQM) extracts control conditions from user-given queries and feeds them layer by layer into the rectification network. The rectification network is composed of Controllable Convolution Modulating Blocks (CCMB) and Controllable Attention Modulating Blocks (CAMB), which modulate the input features " + }, + { + "bbox": [ + 130, + 286, + 482, + 364 + ], + "type": "inline_equation", + "content": "F_{in}" + }, + { + "bbox": [ + 130, + 286, + 482, + 364 + ], + "type": "text", + "content": " with control conditions " + }, + { + "bbox": [ + 130, + 286, + 482, + 364 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 130, + 286, + 482, + 364 + ], + "type": "text", + "content": ", enabling controllable rectification process." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 387, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 460 + ], + "type": "text", + "content": "and then use them to couple the features with the main branch. MM-RealSR [33] proposed a metric learning strategy to map unquantifiable degradation levels to a metric space as control conditions. However, due to the different optimization objectives and complexity of the control mechanisms, directly applying these methods to fisheye image distortion rectification tasks will not achieve effective control over the rectification process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 477, + 233, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 477, + 233, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 477, + 233, + 491 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 502, + 209, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 209, + 513 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 209, + 513 + ], + "type": "text", + "content": "3.1 Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "type": "text", + "content": "The overview of our proposed Query-based Controllable Distortion Rectification network (QueryCDR) is shown in Fig. 2. First, following existing work [52], the input fisheye image is fed into the flow estimation module to obtain the appearance flow, which performs a coarse-grained rectification for image features (i.e., warping " + }, + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\omega(\\cdot)" + }, + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "type": "text", + "content": "). Then, the Distortion-aware Learnable Query Mechanism (DLQM) (in Sec. 3.2) extracts control conditions from the user-given query and feeds them layer by layer into the rectification network. Finally, a U-shaped hierarchical network composed of several controllable modulating blocks (in Sec. 3.3) is used to rectify the distorted input image. These blocks modulate features with the control conditions given by DLQM, to get the final output." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "In the following, we will provide a comprehensive description of each module and its corresponding role in the QueryCDR." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 156, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 156, + 90, + 448, + 102 + ], + "type": "text", + "content": "QueryCDR: Query-based Controllable Distortion Rectification Network" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 159 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 159 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 159 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 159 + ], + "type": "image", + "image_path": "ce06cff48d09e005246042699b655d3fa7a47e92a20038f2ae663b985c4385dd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 169, + 482, + 214 + ], + "lines": [ + { + "bbox": [ + 130, + 169, + 482, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 169, + 482, + 214 + ], + "type": "text", + "content": "Fig. 3: Given an input image, QueryCDR can accurately produce results with different rectification degrees by feeding different queries. Moreover, by interpolating between different queries, we can achieve smooth continuous rectification for any distortion degree. For examples, " + }, + { + "bbox": [ + 130, + 169, + 482, + 214 + ], + "type": "inline_equation", + "content": "Q_{1.25} = 0.75Q_1 + 0.25Q_2" + }, + { + "bbox": [ + 130, + 169, + 482, + 214 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 169, + 482, + 214 + ], + "type": "inline_equation", + "content": "Q_{8.5} = 0.5Q_8 + 0.5Q_9" + }, + { + "bbox": [ + 130, + 169, + 482, + 214 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 238, + 440, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 238, + 440, + 251 + ], + "spans": [ + { + "bbox": [ + 131, + 238, + 440, + 251 + ], + "type": "text", + "content": "3.2 Distortion-aware Learnable Query Mechanism (DLQM)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 261, + 482, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 428 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 428 + ], + "type": "text", + "content": "To achieve control over the rectification process, most existing controllable methods [33,47,55] introduce scalars as control conditions to represent the degradation degrees for image reconstruction. However, due to the peculiarities of distortion in fisheye images [10, 42, 53], these methods fail to achieve effective control over the rectification process. To address this issue, we propose the DLQM. It maintains a learnable query set, providing diverse effective control conditions for the rectification network. During training, DLQM projects the learned high-dimensional positional mapping relationship into a low-dimensional latent space of queries. During inference, DLQM extracts the control information from the queries and converts it into corresponding control conditions for each layer of the rectification network. Simply by providing different queries, DLQM can effectively control the rectification process, as shown in Fig. 3. At the same time, our QueryCDR can also achieve smooth continuous rectification for any distortion degrees by simply interpolating between different queries." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 427, + 482, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 427, + 482, + 513 + ], + "spans": [ + { + "bbox": [ + 131, + 427, + 482, + 513 + ], + "type": "text", + "content": "Specifically, we construct a query set " + }, + { + "bbox": [ + 131, + 427, + 482, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_s = \\{\\bar{Q}_i\\mid Q_i\\in \\mathbb{R}^{C_{in}\\times H_{in}\\times W_{in}},i = 1,2,\\ldots ,N\\}" + }, + { + "bbox": [ + 131, + 427, + 482, + 513 + ], + "type": "text", + "content": ", which comprises " + }, + { + "bbox": [ + 131, + 427, + 482, + 513 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 427, + 482, + 513 + ], + "type": "text", + "content": " queries representing different distortion degrees. " + }, + { + "bbox": [ + 131, + 427, + 482, + 513 + ], + "type": "inline_equation", + "content": "H_{in}" + }, + { + "bbox": [ + 131, + 427, + 482, + 513 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 427, + 482, + 513 + ], + "type": "inline_equation", + "content": "W_{in}" + }, + { + "bbox": [ + 131, + 427, + 482, + 513 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 427, + 482, + 513 + ], + "type": "inline_equation", + "content": "C_{in}" + }, + { + "bbox": [ + 131, + 427, + 482, + 513 + ], + "type": "text", + "content": " represent the query's height, width, and channel, respectively. Each query has the same size as the input image " + }, + { + "bbox": [ + 131, + 427, + 482, + 513 + ], + "type": "inline_equation", + "content": "I_{in}\\in \\mathbb{R}^{C_{in}\\times H_{in}\\times W_{in}}" + }, + { + "bbox": [ + 131, + 427, + 482, + 513 + ], + "type": "text", + "content": ". During inference, users select a query " + }, + { + "bbox": [ + 131, + 427, + 482, + 513 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 131, + 427, + 482, + 513 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 131, + 427, + 482, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_s" + }, + { + "bbox": [ + 131, + 427, + 482, + 513 + ], + "type": "text", + "content": " and feed it into the DLQM. In DLQM, the control extracting part " + }, + { + "bbox": [ + 131, + 427, + 482, + 513 + ], + "type": "inline_equation", + "content": "\\mathrm{CE}(\\cdot)" + }, + { + "bbox": [ + 131, + 427, + 482, + 513 + ], + "type": "text", + "content": " is firstly used to extract the features in " + }, + { + "bbox": [ + 131, + 427, + 482, + 513 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 131, + 427, + 482, + 513 + ], + "type": "text", + "content": ". The input processing part can be expressed as follows," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 272, + 523, + 481, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 523, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 272, + 523, + 481, + 536 + ], + "type": "interline_equation", + "content": "Q _ {e x} = \\operatorname {C E} \\left(Q _ {i}\\right), \\tag {1}", + "image_path": "73fac09b23ef7c5da22c04deb86933fa58b1e03d92496a914b8600d02664be85.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 545, + 482, + 594 + ], + "type": "inline_equation", + "content": "Q_{ex} \\in \\mathbb{R}^{C_{in} \\times H_{in} \\times W_{in}}" + }, + { + "bbox": [ + 130, + 545, + 482, + 594 + ], + "type": "text", + "content": " is the extracted feature, and serves as the input to the first control layer of DLQM. To minimize the computational costs and effectively extract the control information, " + }, + { + "bbox": [ + 130, + 545, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\mathrm{CE}(\\cdot)" + }, + { + "bbox": [ + 130, + 545, + 482, + 594 + ], + "type": "text", + "content": " is composed of three convolution layers with a kernel size of " + }, + { + "bbox": [ + 130, + 545, + 482, + 594 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 545, + 482, + 594 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 643 + ], + "type": "text", + "content": "Subsequently, DLQM provides corresponding control conditions to the respective layers of the rectification network. To provide appropriate control conditions, each control layer of DLQM comprises two fully connected layers " + }, + { + "bbox": [ + 130, + 594, + 482, + 643 + ], + "type": "inline_equation", + "content": "\\mathrm{FC}_1(\\cdot)" + }, + { + "bbox": [ + 130, + 594, + 482, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 594, + 482, + 643 + ], + "type": "inline_equation", + "content": "\\mathrm{FC}_2(\\cdot)" + }, + { + "bbox": [ + 130, + 594, + 482, + 643 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 130, + 594, + 482, + 643 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 594, + 482, + 643 + ], + "type": "text", + "content": "-th layer of DLQM can be represented as follows," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 254, + 651, + 481, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 651, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 254, + 651, + 481, + 666 + ], + "type": "interline_equation", + "content": "Q _ {c} ^ {l} = \\mathrm {F C} _ {2} ^ {l} \\left(\\mathrm {F C} _ {1} ^ {l} \\left(Q _ {c} ^ {l - 1}\\right)\\right), \\tag {2}", + "image_path": "807a658483622445b58bdb9ecea3132c2dbd52af11d290624fc3703c3d04cdb6.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "P. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "Q_{c}^{l-1}" + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": " is the output control condition from the " + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "(l-1)" + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": "-th control layer. For " + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "l = 1" + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "Q_{c}^{0}" + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": " represents the extracted feature " + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "Q_{ex}" + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": ". The output " + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "Q_{c}^{l}" + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": "-th control layer is fed into the " + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": "-th rectification layer as a control condition, and simultaneously serves as input to the " + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "(l+1)" + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": "-th control layer of DLQM." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 480, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 200 + ], + "type": "text", + "content": "Afterward, we feed the control conditions into the rectification network. Each layer's controllable modulating block modulates the input feature with the control condition, guiding the modulated features toward the desired direction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 217, + 316, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 217, + 316, + 229 + ], + "spans": [ + { + "bbox": [ + 131, + 217, + 316, + 229 + ], + "type": "text", + "content": "3.3 Controllable Modulating Block" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 237, + 480, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 480, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 480, + 416 + ], + "type": "text", + "content": "When controlling the rectification process, to avoid obtaining results with blurred texture details or residual distortions, we introduce two types of controllable modulating blocks tailored for reconstructing local texture details and learning continuous distortion patterns. They are denoted as the Controllable Convolution Modulating Block (CCMB) based on CNN [17] (in Fig. 2(a)) and the Controllable Attention Modulating Block (CAMB) based on Transformer [45] (in Fig. 2(b)). The CCMB can adaptively fuse the original features and the controlled features, preserving more local details. The CAMB can better capture the spatial distortion information, guaranteeing the integrity of the recovered content [10,35,53]. To balance performance and computational cost, our QueryCDR constructs a U-shaped rectification network [52] composed of CCMB and CAMB. In the first three layers with larger feature maps, i.e., " + }, + { + "bbox": [ + 130, + 237, + 480, + 416 + ], + "type": "inline_equation", + "content": "l = \\{1,2,3,9,10,11\\}" + }, + { + "bbox": [ + 130, + 237, + 480, + 416 + ], + "type": "text", + "content": ", CCMB is employed to learn more local texture details. In the remaining layers with smaller feature maps, i.e., " + }, + { + "bbox": [ + 130, + 237, + 480, + 416 + ], + "type": "inline_equation", + "content": "l = \\{4,5,6,7,8\\}" + }, + { + "bbox": [ + 130, + 237, + 480, + 416 + ], + "type": "text", + "content": ", CAMB is utilized to capture more global dependencies within the images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 433, + 479, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 479, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 479, + 528 + ], + "type": "text", + "content": "Controllable Convolution Modulating Block (CCMB). When incorporating control conditions into the rectification network, existing methods [47,55] fail to effectively balance the fusion ratio between original features and controlled features. Directly using the controlled features or fusing them with original features at a fixed ratio will degrade the quality of the rectified images. To harmoniously incorporate control conditions into the rectification process, we design the CCMB, as shown in Fig. 2(a). CCMB can dynamically find an optimal ratio to modulate the distorted features with control conditions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 529, + 479, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 479, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 479, + 564 + ], + "type": "text", + "content": "Specifically, CCMB receives an input feature " + }, + { + "bbox": [ + 130, + 529, + 479, + 564 + ], + "type": "inline_equation", + "content": "F_{in} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 130, + 529, + 479, + 564 + ], + "type": "text", + "content": " and a control condition " + }, + { + "bbox": [ + 130, + 529, + 479, + 564 + ], + "type": "inline_equation", + "content": "Q_{c} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 130, + 529, + 479, + 564 + ], + "type": "text", + "content": ". Both of them are used to predict the fusion ratio. Here we omit the layer information for brevity," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 269, + 575, + 480, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 575, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 269, + 575, + 480, + 587 + ], + "type": "interline_equation", + "content": "\\theta = \\operatorname {C P} \\left(F _ {i n}, Q _ {c}\\right), \\tag {3}", + "image_path": "d3a49c39d5e41512e5546fd91b48ada2bd43e2ce4e85fb8a45c97c1e1b9fcef6.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 596, + 479, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 596, + 479, + 619 + ], + "spans": [ + { + "bbox": [ + 130, + 596, + 479, + 619 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 596, + 479, + 619 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}(\\cdot)" + }, + { + "bbox": [ + 130, + 596, + 479, + 619 + ], + "type": "text", + "content": " is the coefficient predictor composed of two fully connected layers, takes the concatenation of " + }, + { + "bbox": [ + 130, + 596, + 479, + 619 + ], + "type": "inline_equation", + "content": "F_{in}" + }, + { + "bbox": [ + 130, + 596, + 479, + 619 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 596, + 479, + 619 + ], + "type": "inline_equation", + "content": "Q_{c}" + }, + { + "bbox": [ + 130, + 596, + 479, + 619 + ], + "type": "text", + "content": " as input and predicts the fusion ratio " + }, + { + "bbox": [ + 130, + 596, + 479, + 619 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 596, + 479, + 619 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 620, + 479, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 620, + 479, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 620, + 479, + 643 + ], + "type": "text", + "content": "Then, we perform element-wise multiplication " + }, + { + "bbox": [ + 130, + 620, + 479, + 643 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 130, + 620, + 479, + 643 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 130, + 620, + 479, + 643 + ], + "type": "inline_equation", + "content": "F_{in}" + }, + { + "bbox": [ + 130, + 620, + 479, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 620, + 479, + 643 + ], + "type": "inline_equation", + "content": "Q_{c}" + }, + { + "bbox": [ + 130, + 620, + 479, + 643 + ], + "type": "text", + "content": " to yield the controlled features " + }, + { + "bbox": [ + 130, + 620, + 479, + 643 + ], + "type": "inline_equation", + "content": "F_{c} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 130, + 620, + 479, + 643 + ], + "type": "text", + "content": ", which can be expressed as," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 274, + 654, + 480, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 654, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 274, + 654, + 480, + 666 + ], + "type": "interline_equation", + "content": "F _ {c} = F _ {i n} \\otimes Q _ {c}. \\tag {4}", + "image_path": "0ce84cfb5f2096b18a6ca3c7d0316315d96643aee0f6eab4db9d5f7db8c8f030.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 156, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 156, + 91, + 448, + 102 + ], + "type": "text", + "content": "QueryCDR: Query-based Controllable Distortion Rectification Network" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "text", + "content": "Finally, the input features " + }, + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "inline_equation", + "content": "F_{in}" + }, + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "text", + "content": " and controlled features " + }, + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "text", + "content": " are combined in a weighted sum according to the fusion ratio " + }, + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "text", + "content": ", illustrated as " + }, + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "inline_equation", + "content": "\\psi(\\cdot)" + }, + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "text", + "content": " in Fig. 2(a)," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 219, + 147, + 481, + 161 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 147, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 219, + 147, + 481, + 161 + ], + "type": "interline_equation", + "content": "F _ {o u t} = \\psi \\left(F _ {c}, F _ {i n}, \\theta\\right) = \\theta F _ {c} \\oplus (1 - \\theta) F _ {i n}, \\tag {5}", + "image_path": "52b9ce90388e5fc63768280ce57a143b6f485962080986ff388c905abd1d20fd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 167, + 481, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 167, + 481, + 190 + ], + "spans": [ + { + "bbox": [ + 130, + 167, + 481, + 190 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 167, + 481, + 190 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 130, + 167, + 481, + 190 + ], + "type": "text", + "content": " represents the element-wise addition, and " + }, + { + "bbox": [ + 130, + 167, + 481, + 190 + ], + "type": "inline_equation", + "content": "F_{out} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 130, + 167, + 481, + 190 + ], + "type": "text", + "content": " is the final output of CCMB." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 192, + 481, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 192, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 192, + 481, + 216 + ], + "type": "text", + "content": "Due to this dynamic modulation mechanism, CCMB achieves effective control over the rectification process while preserving richer texture details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 232, + 482, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 232, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 482, + 291 + ], + "type": "text", + "content": "Controllable Attention Modulating Block (CAMB). Due to the lack of perception of global information, CNN-based networks struggle to learn long-range dependencies, particularly the continuous and amorphous distortions prevalent in fisheye distortions [10, 42, 53]. In contrast, Transformer-based networks effectively compensate for this, with their global attention mechanism [13, 35]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 292, + 482, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 292, + 482, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 482, + 388 + ], + "type": "text", + "content": "Therefore, we propose the CAMB, as illustrated in Fig. 2(b). To optimally leverage the control conditions given by DLQM, we designed the control-attention mechanism, enabling CAMB to perceive the global spatial relationships in the control conditions effectively. Specifically, we unfold and project the controlled feature " + }, + { + "bbox": [ + 130, + 292, + 482, + 388 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 130, + 292, + 482, + 388 + ], + "type": "text", + "content": " (in Eq. (4)) as the query " + }, + { + "bbox": [ + 130, + 292, + 482, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} \\in \\mathbb{R}^{m \\times L}" + }, + { + "bbox": [ + 130, + 292, + 482, + 388 + ], + "type": "text", + "content": ", with the input feature " + }, + { + "bbox": [ + 130, + 292, + 482, + 388 + ], + "type": "inline_equation", + "content": "F_{in}" + }, + { + "bbox": [ + 130, + 292, + 482, + 388 + ], + "type": "text", + "content": " as the key " + }, + { + "bbox": [ + 130, + 292, + 482, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{K} \\in \\mathbb{R}^{m \\times L}" + }, + { + "bbox": [ + 130, + 292, + 482, + 388 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 130, + 292, + 482, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{V} \\in \\mathbb{R}^{m \\times L}" + }, + { + "bbox": [ + 130, + 292, + 482, + 388 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 292, + 482, + 388 + ], + "type": "inline_equation", + "content": "L = H \\times W" + }, + { + "bbox": [ + 130, + 292, + 482, + 388 + ], + "type": "text", + "content": " represents the sequence length, and " + }, + { + "bbox": [ + 130, + 292, + 482, + 388 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 292, + 482, + 388 + ], + "type": "text", + "content": " denotes the dimensions of the sequences. The control-attention is described by the following," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 212, + 395, + 480, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 395, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 212, + 395, + 480, + 437 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {Q} = W _ {\\mathcal {Q}} F _ {c}, \\mathcal {K} = W _ {\\mathcal {K}} F _ {i n}, \\mathcal {V} = W _ {\\mathcal {V}} F _ {i n}, \\\\ \\operatorname {C T R L - A T T N} (\\mathcal {Q}, \\mathcal {K}, \\mathcal {V}) = \\operatorname {s o f t m a x} \\left(\\frac {\\mathcal {Q} \\mathcal {K} ^ {T}}{\\sqrt {m}}\\right) \\mathcal {V}, \\tag {6} \\\\ \\end{array}", + "image_path": "d27ddc4a81f845afc05c0c2e799a73655e12c001b3c526978ce1167440393b69.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 443, + 481, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 443, + 481, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 443, + 481, + 480 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 443, + 481, + 480 + ], + "type": "inline_equation", + "content": "W_{\\mathcal{Q}}, W_{\\mathcal{K}}, W_{\\mathcal{V}} \\in \\mathbb{R}^{m \\times C}" + }, + { + "bbox": [ + 130, + 443, + 481, + 480 + ], + "type": "text", + "content": " represent the projection matrices of the queries, keys, and values, respectively. CTRL-ATTN(" + }, + { + "bbox": [ + 130, + 443, + 481, + 480 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 130, + 443, + 481, + 480 + ], + "type": "text", + "content": ") represents the control-attention we proposed. The overall computation of CAMB can be formulated as follows," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 219, + 487, + 480, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 487, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 219, + 487, + 480, + 515 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} F _ {a} = \\operatorname {C T R L - A T T N} (\\mathrm {L N} (\\mathcal {Q}, \\mathcal {K}, \\mathcal {V})) \\oplus F _ {i n}, \\\\ F _ {o u t} = \\operatorname {C o n v} _ {1 \\times 1} (\\operatorname {F F N} (\\operatorname {L N} (F _ {a})) \\oplus F _ {a}), \\tag {7} \\\\ \\end{array}", + "image_path": "881fbafe3f61c40d671edb52130f386b4ae7255b62a9a79c9e87f47e8fd675f8.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 522, + 481, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 481, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 481, + 570 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 522, + 481, + 570 + ], + "type": "inline_equation", + "content": "F_{in}" + }, + { + "bbox": [ + 130, + 522, + 481, + 570 + ], + "type": "text", + "content": " means the input feature, " + }, + { + "bbox": [ + 130, + 522, + 481, + 570 + ], + "type": "inline_equation", + "content": "F_{a} \\in \\mathbb{R}^{m \\times H \\times W}" + }, + { + "bbox": [ + 130, + 522, + 481, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 522, + 481, + 570 + ], + "type": "inline_equation", + "content": "F_{out} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 130, + 522, + 481, + 570 + ], + "type": "text", + "content": " are the outputs of CTRL-ATTN(" + }, + { + "bbox": [ + 130, + 522, + 481, + 570 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 130, + 522, + 481, + 570 + ], + "type": "text", + "content": ") and CAMB, respectively. " + }, + { + "bbox": [ + 130, + 522, + 481, + 570 + ], + "type": "inline_equation", + "content": "\\mathrm{LN}(\\cdot)" + }, + { + "bbox": [ + 130, + 522, + 481, + 570 + ], + "type": "text", + "content": " denotes the layer normalization [1]. " + }, + { + "bbox": [ + 130, + 522, + 481, + 570 + ], + "type": "inline_equation", + "content": "\\mathrm{FFN}(\\cdot)" + }, + { + "bbox": [ + 130, + 522, + 481, + 570 + ], + "type": "text", + "content": " stands for the feed-forward network composed of three fully connected layers, which helps CAMB to focus on the global dependencies." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 571, + 481, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 571, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 571, + 481, + 594 + ], + "type": "text", + "content": "CAMB can discern the global mapping relationships within fisheye images, ensuring rectification uniformity compared to the CNN-based networks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 611, + 250, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 611, + 250, + 624 + ], + "spans": [ + { + "bbox": [ + 131, + 611, + 250, + 624 + ], + "type": "text", + "content": "3.4 Training Strategy" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "To guarantee the robust and stable controllable distortion rectification of Query-CDR, we design a two-stage training strategy that combines coarse-grained distortion pre-training and fine-grained distortion fine-tuning." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "P. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "During the coarse-grained distortion pre-training phase, we choose the most commonly used dataset [10,24,52,54] that only contains one degree of distortion to train our QueryCDR. For clarity, we denote the distortion degree in this phase as " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": ". Correspondingly, only one query, denoted as " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": ", is used for training. The optimization objective can be expressed as follows," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 270, + 184, + 480, + 196 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 184, + 480, + 196 + ], + "spans": [ + { + "bbox": [ + 270, + 184, + 480, + 196 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p r e} = \\mathcal {L} _ {r} + \\mathcal {L} _ {m}, \\tag {8}", + "image_path": "80f39aa725f56e3a0ac752030260bce7bca759410ceb99b5bb667a450da66655.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 201, + 479, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 479, + 224 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 201, + 479, + 224 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pre}" + }, + { + "bbox": [ + 130, + 201, + 479, + 224 + ], + "type": "text", + "content": " is the overall loss function for the pre-training phase. " + }, + { + "bbox": [ + 130, + 201, + 479, + 224 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 130, + 201, + 479, + 224 + ], + "type": "text", + "content": " denotes the reconstruction loss," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 263, + 224, + 480, + 240 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 224, + 480, + 240 + ], + "spans": [ + { + "bbox": [ + 263, + 224, + 480, + 240 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r} = \\left\\| I _ {o u t} ^ {d} - I _ {g t} ^ {d} \\right\\| _ {1}, \\tag {9}", + "image_path": "62cf4546b90de1b042e2d6f5113197870d616c7714be04e7c159ce0c384fcdc7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 243, + 479, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 479, + 268 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 479, + 268 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 243, + 479, + 268 + ], + "type": "inline_equation", + "content": "I_{out}^{d}" + }, + { + "bbox": [ + 130, + 243, + 479, + 268 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 243, + 479, + 268 + ], + "type": "inline_equation", + "content": "I_{gt}^{d}" + }, + { + "bbox": [ + 130, + 243, + 479, + 268 + ], + "type": "text", + "content": " signifie the output result and ground truth, respectively. " + }, + { + "bbox": [ + 130, + 243, + 479, + 268 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_m" + }, + { + "bbox": [ + 130, + 243, + 479, + 268 + ], + "type": "text", + "content": " denotes the multi-scale loss," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 235, + 274, + 480, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 274, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 235, + 274, + 480, + 308 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m} = \\sum_ {j = 1} ^ {Z - 1} \\| S \\left(I _ {g t} ^ {d}, j\\right) - C \\left(F _ {\\text {o u t}} ^ {j}\\right) \\| _ {1}, \\tag {10}", + "image_path": "a018314c6cbcefc9280cc71a41024abbeae1b0b79f7d552ca428430907053ba7.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 314, + 480, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 480, + 377 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 480, + 377 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 314, + 480, + 377 + ], + "type": "inline_equation", + "content": "S(\\cdot)" + }, + { + "bbox": [ + 130, + 314, + 480, + 377 + ], + "type": "text", + "content": " represents the operation that down-samples the input " + }, + { + "bbox": [ + 130, + 314, + 480, + 377 + ], + "type": "inline_equation", + "content": "I_{gt}^{d}" + }, + { + "bbox": [ + 130, + 314, + 480, + 377 + ], + "type": "text", + "content": " by a factor of " + }, + { + "bbox": [ + 130, + 314, + 480, + 377 + ], + "type": "inline_equation", + "content": "1/2^{j}" + }, + { + "bbox": [ + 130, + 314, + 480, + 377 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 314, + 480, + 377 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 314, + 480, + 377 + ], + "type": "text", + "content": " represents the number of decoder's layers, and we set " + }, + { + "bbox": [ + 130, + 314, + 480, + 377 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 314, + 480, + 377 + ], + "type": "text", + "content": " to 6. " + }, + { + "bbox": [ + 130, + 314, + 480, + 377 + ], + "type": "inline_equation", + "content": "F_{out}^{j}" + }, + { + "bbox": [ + 130, + 314, + 480, + 377 + ], + "type": "text", + "content": " denotes the feature in " + }, + { + "bbox": [ + 130, + 314, + 480, + 377 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 314, + 480, + 377 + ], + "type": "text", + "content": "-th decoder layer. " + }, + { + "bbox": [ + 130, + 314, + 480, + 377 + ], + "type": "inline_equation", + "content": "C(\\cdot)" + }, + { + "bbox": [ + 130, + 314, + 480, + 377 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 130, + 314, + 480, + 377 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 314, + 480, + 377 + ], + "type": "text", + "content": " convolution for decoding the features into 3-channel RGB images. In this way, each feature map on the decoder can be effectively supervised. The " + }, + { + "bbox": [ + 130, + 314, + 480, + 377 + ], + "type": "inline_equation", + "content": "I_{out}^{d}" + }, + { + "bbox": [ + 130, + 314, + 480, + 377 + ], + "type": "text", + "content": " can be obtained as," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 249, + 384, + 480, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 384, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 249, + 384, + 480, + 397 + ], + "type": "interline_equation", + "content": "I _ {o u t} ^ {d} = \\operatorname {Q u e r y C D R} \\left(I _ {i n} ^ {d}, Q\\right), \\tag {11}", + "image_path": "70979b4a6d2dcbfd731344a7d2a60186b1c3f2d0b9d7f31613a62a6d28a57923.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 402, + 479, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 479, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 479, + 426 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 402, + 479, + 426 + ], + "type": "inline_equation", + "content": "I_{in}^{d}" + }, + { + "bbox": [ + 130, + 402, + 479, + 426 + ], + "type": "text", + "content": " represents fisheye images with distortion degree " + }, + { + "bbox": [ + 130, + 402, + 479, + 426 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 402, + 479, + 426 + ], + "type": "text", + "content": " as input. This way can effectively boost the model stability and accelerate the convergence." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 427, + 480, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 427, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 427, + 480, + 521 + ], + "type": "text", + "content": "During the fine-grained distortion fine-tuning phase, we use the varying distortion degrees datasets to fine-tune our QueryCDR. Before training, We replicate the weight of the query " + }, + { + "bbox": [ + 130, + 427, + 480, + 521 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 427, + 480, + 521 + ], + "type": "text", + "content": " to the other queries to accelerate convergence on other distortion degrees. Subsequently, we fine-tune our QueryCDR using fisheye images with varying distortion degrees, and feeding corresponding queries into QueryCDR for training at the same time. This allows the query set to efficiently acquire diverse latent spatial relationships. The optimization objective can be expressed as follows," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 266, + 521, + 480, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 521, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 266, + 521, + 480, + 536 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {f i n e}} ^ {d _ {i}} = \\mathcal {L} _ {r} ^ {d _ {i}} + \\mathcal {L} _ {m} ^ {d _ {i}}, \\tag {12}", + "image_path": "13fb9145596ab2ad8dd874baec4e082797a7f53fc479c3ae429d1d32df23f244.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 540, + 479, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 540, + 479, + 579 + ], + "spans": [ + { + "bbox": [ + 131, + 540, + 479, + 579 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 540, + 479, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{fine}^{d_i}" + }, + { + "bbox": [ + 131, + 540, + 479, + 579 + ], + "type": "text", + "content": " is the overall fine-tuning loss function for distortion degree " + }, + { + "bbox": [ + 131, + 540, + 479, + 579 + ], + "type": "inline_equation", + "content": "d_{i}, i \\in \\{1,2,\\dots,9\\}" + }, + { + "bbox": [ + 131, + 540, + 479, + 579 + ], + "type": "text", + "content": ", which is similar to the " + }, + { + "bbox": [ + 131, + 540, + 479, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pre}" + }, + { + "bbox": [ + 131, + 540, + 479, + 579 + ], + "type": "text", + "content": " but calculated across different distortion degrees. The " + }, + { + "bbox": [ + 131, + 540, + 479, + 579 + ], + "type": "inline_equation", + "content": "I_{out}^{d_i}" + }, + { + "bbox": [ + 131, + 540, + 479, + 579 + ], + "type": "text", + "content": " can be obtained as," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 248, + 585, + 480, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 585, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 248, + 585, + 480, + 598 + ], + "type": "interline_equation", + "content": "I _ {\\text {o u t}} ^ {d _ {i}} = \\operatorname {Q u e r y C D R} \\left(I _ {\\text {i n}} ^ {d _ {i}}, Q _ {i}\\right), \\tag {13}", + "image_path": "9c2a16cce95e68675cb5d391a9bd870e4fe7a47eeb7afa4dd997beed1e5d4ca4.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 605, + 479, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 479, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 479, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 605, + 479, + 629 + ], + "type": "inline_equation", + "content": "I_{in}^{d_i}" + }, + { + "bbox": [ + 130, + 605, + 479, + 629 + ], + "type": "text", + "content": " denotes fisheye images with a distortion degree of " + }, + { + "bbox": [ + 130, + 605, + 479, + 629 + ], + "type": "inline_equation", + "content": "d_i" + }, + { + "bbox": [ + 130, + 605, + 479, + 629 + ], + "type": "text", + "content": " as input, " + }, + { + "bbox": [ + 130, + 605, + 479, + 629 + ], + "type": "inline_equation", + "content": "Q_i" + }, + { + "bbox": [ + 130, + 605, + 479, + 629 + ], + "type": "text", + "content": " denotes the query corresponding to " + }, + { + "bbox": [ + 130, + 605, + 479, + 629 + ], + "type": "inline_equation", + "content": "d_i, i \\in \\{1, 2, \\dots, 9\\}" + }, + { + "bbox": [ + 130, + 605, + 479, + 629 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 630, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 479, + 665 + ], + "type": "text", + "content": "With this two-stage training strategy, our QueryCDR can effectively utilize only a small amount of varying distortion degrees data that is difficult to obtain, to rapidly finish the training of our network." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 156, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 156, + 91, + 448, + 102 + ], + "type": "text", + "content": "QueryCDR: Query-based Controllable Distortion Rectification Network" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 157, + 157, + 457, + 420 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Quantitative comparison (PSNR (dB)↑, SSIM ↑) on COCO [27] fisheye image dataset with varying distortion degrees. Red indicates the best and blue indicates the second best performance (best viewed in color)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 157, + 157, + 457, + 420 + ], + "lines": [ + { + "bbox": [ + 157, + 157, + 457, + 420 + ], + "spans": [ + { + "bbox": [ + 157, + 157, + 457, + 420 + ], + "type": "table", + "html": "
MethodPSNR
d1d2d3d4d5d6d7d8d9Avg
SC [40]10.0511.5911.8111.0311.5010.1911.1110.199.1410.73
DeepCalib [4]10.0010.6911.0111.1911.2811.4611.4511.1311.1111.04
Blind [23]12.9811.2410.309.629.9911.7512.6912.7512.8011.57
DR-GAN [24]15.6817.4017.9718.3418.5018.4417.9517.9417.4717.74
PCN [52]14.9317.4318.4318.8618.8618.8818.7417.3518.2617.97
DDA [54]16.3917.4117.4319.4820.1218.9018.8518.1718.2218.33
SimFIR [10]16.5717.8818.4318.9719.3119.2819.1918.6518.4818.53
QueryCDR20.0120.2920.3920.4120.7220.8120.5819.1120.5320.32
MethodSSIM
d1d2d3d4d5d6d7d8d9Avg
SC [40]0.1010.1130.1490.1820.2830.1750.1410.1260.0930.151
DeepCalib [4]0.1840.2100.2230.2300.2340.2460.2500.2450.2460.229
Blind [23]0.3080.2440.1990.1760.1940.2960.3670.3950.4200.289
DR-GAN [24]0.2950.3300.3390.3440.3440.3320.3140.3120.2990.323
PCN [52]0.4200.5470.5890.6070.6080.6100.6150.5760.6030.575
DDA [54]0.4550.5890.5920.6200.6750.6260.6190.5640.5810.591
SimFIR [10]0.4920.5810.6260.6350.6400.6280.6220.5910.5950.601
QueryCDR0.6430.6650.6680.6770.6880.6990.6920.6560.6930.676
", + "image_path": "b3a3911e45c37d1c5551a67707ae878006fe6b4bbe061533d6f8e867cd285b29.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 438, + 224, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 438, + 224, + 451 + ], + "spans": [ + { + "bbox": [ + 132, + 438, + 224, + 451 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 460, + 272, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 272, + 473 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 272, + 473 + ], + "type": "text", + "content": "4.1 Experimental Settings" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 477, + 482, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 477, + 482, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 477, + 482, + 598 + ], + "type": "text", + "content": "To demonstrate the effectiveness of our proposed QueryCDR, we followed the existing works [10, 24, 52, 54], employed the four-parameter polynomial model to synthesize fisheye images. We constructed synthetic datasets based on the original images of COCO [27] and Places2 [62] datasets, respectively. Specifically, for images with distortion degree " + }, + { + "bbox": [ + 130, + 477, + 482, + 598 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 477, + 482, + 598 + ], + "type": "text", + "content": ", 40,000 images were used for the pre-training stage. And for images with varying distortion degrees " + }, + { + "bbox": [ + 130, + 477, + 482, + 598 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 130, + 477, + 482, + 598 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 477, + 482, + 598 + ], + "type": "inline_equation", + "content": "i \\in \\{1,2,\\dots ,9\\}" + }, + { + "bbox": [ + 130, + 477, + 482, + 598 + ], + "type": "text", + "content": ", 18,000 images were used for the fine-tuning stage, and 3,600 images for testing. The images were resized to " + }, + { + "bbox": [ + 130, + 477, + 482, + 598 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 477, + 482, + 598 + ], + "type": "text", + "content": " when fed into the network. For a fair comparison, we followed existing works [10, 24, 52, 54] utilizing a batch size of 16 and the Adam [19] optimizer with a learning rate of 1e-4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 612, + 282, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 612, + 282, + 624 + ], + "spans": [ + { + "bbox": [ + 132, + 612, + 282, + 624 + ], + "type": "text", + "content": "4.2 Performance Evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "To evaluate the performance of our method, we retrained and validated existing fisheye image rectification methods on synthetic fisheye datasets with 9 distortion degrees. The methods can be summarized into three categories: traditional" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "P. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 466 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 466 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 466 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 466 + ], + "type": "image", + "image_path": "1fe983ba68f6777d12ebc469e5cfda3a4141d0d3ad7cef8eedd2c622d3b9a149.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 472, + 418, + 484 + ], + "lines": [ + { + "bbox": [ + 194, + 472, + 418, + 484 + ], + "spans": [ + { + "bbox": [ + 194, + 472, + 418, + 484 + ], + "type": "text", + "content": "Fig. 4: Qualitative results on synthetic fisheye images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 510, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 559 + ], + "type": "text", + "content": "method SC [40], regression-based methods DeepCalib [4] and Blind [23], and generation-based DR-GAN [24], PCN [52], DDA [54] and SimFIR [10]. For a fair comparison, we followed existing works [10,24,52,54] employing Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) to quantify performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Quantitative Results. Our performance comparisons are shown in Tab. 1. Compared to existing methods, QueryCDR achieves the best performance across all distortion degrees without retraining. It is because DLQM effectively utilizes the control information in queries to guide the network in achieving varying degrees of rectification. Additionally, the capability of CCMB and CAMB allows for controlled rectification while obtaining high-quality results. Moreover, we observe that even on images with distortion degree " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "d_{5}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ", which the existing" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "text", + "content": "QueryCDR: Query-based Controllable Distortion Rectification Network" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 208 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 208 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 208 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 208 + ], + "type": "image", + "image_path": "6378eef9c633bf24a573b34d50f19ace4590a5bce813bf929c82cc2d99278a98.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 216, + 420, + 228 + ], + "lines": [ + { + "bbox": [ + 192, + 216, + 420, + 228 + ], + "spans": [ + { + "bbox": [ + 192, + 216, + 420, + 228 + ], + "type": "text", + "content": "Fig. 5: Qualitative results on real-world fisheye images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 249, + 480, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 480, + 285 + ], + "type": "text", + "content": "methods perform well [10, 24, 52, 54], QueryCDR still outperforms the second best method [54] by 0.60 dB of PSNR and 0.013 of SSIM. This further demonstrates the outstanding generalization ability of QueryCDR." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 302, + 480, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 480, + 385 + ], + "type": "text", + "content": "Qualitative Results. To further compare the visual qualities of different methods, we show the results rectified by QueryCDR and other rectification methods in Fig. 4. It can be observed intuitively that other methods fail to effectively rectify images with varying degrees of distortion. Particularly the objects in images, our QueryCDR maintains the accurate structure of objects across different distortions. Furthermore, benefiting from the modulation capability of CCMB and CAMB, QueryCDR preserves richer texture details after rectification." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 386, + 480, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 480, + 445 + ], + "type": "text", + "content": "In addition, to further demonstrate the generalization ability of QueryCDR, we conducted experiments on real-world fisheye image datasets [57] in Fig. 5. Despite the disparities between synthetic and real-world datasets, our QueryCDR still shows robust rectification capabilities, and preservers richer texture details. These results further validate its practicality in real-world scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 463, + 238, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 463, + 238, + 475 + ], + "spans": [ + { + "bbox": [ + 131, + 463, + 238, + 475 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 481, + 480, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 480, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 480, + 506 + ], + "type": "text", + "content": "In this section, we conduct ablation for different control conditions, and study the effect of the controllable modulating blocks and the architecture setting." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 481, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 629 + ], + "type": "text", + "content": "Distortion-aware Learnable Query Mechanism (DLQM). To validate the effectiveness of DLQM in controlling the rectification process, we implemented different rectification networks controlled by scalar, fixed query, and our learnable query in Tab. 2. When using scalar to control, the value of the scalar increments with the distortion of the entire image. When using fixed query to control, the parameters of the query are set to incremental values from 0 to 1 with the degree of distortion from the center to the edges. Following the settings in Sec. 4.1, we use the learnable query " + }, + { + "bbox": [ + 130, + 521, + 481, + 629 + ], + "type": "inline_equation", + "content": "Q_{i}, i \\in \\{1,2,\\dots ,9\\}" + }, + { + "bbox": [ + 130, + 521, + 481, + 629 + ], + "type": "text", + "content": " to learn the distortion distribution corresponding to " + }, + { + "bbox": [ + 130, + 521, + 481, + 629 + ], + "type": "inline_equation", + "content": "d_{i}, i \\in \\{1,2,\\dots ,9\\}" + }, + { + "bbox": [ + 130, + 521, + 481, + 629 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "content": "We trained different methods using the same strategy as illustrated in Sec. 3.4, the results are shown in Tab. 2 and Fig. 6. All three controllable methods outperform the uncontrollable method significantly at various degrees of distortion." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "P. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 144, + 481, + 225 + ], + "blocks": [ + { + "bbox": [ + 176, + 114, + 436, + 125 + ], + "lines": [ + { + "bbox": [ + 176, + 114, + 436, + 125 + ], + "spans": [ + { + "bbox": [ + 176, + 114, + 436, + 125 + ], + "type": "text", + "content": "Table 2: Performance comparison of different control methods." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 144, + 481, + 225 + ], + "lines": [ + { + "bbox": [ + 135, + 144, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 135, + 144, + 481, + 225 + ], + "type": "table", + "html": "
MethodPSNR
d1d2d3d4d5d6d7d8d9Avg
W/o Control14.9317.4318.4318.8618.8618.8818.7417.3518.2617.97
Scalar19.5220.2519.9719.4319.8920.0720.0318.7620.1519.78
Fixed Query20.1320.0320.1419.8820.1620.2520.1618.7419.8419.93
Learnable Query20.0120.2920.3920.4120.7220.8120.5819.1120.5320.32
", + "image_path": "73a291ec711c09b0aa102fa71c9817f5a5b244a1d68f7f295bd2f2e162a0d0d1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 137, + 247, + 272, + 321 + ], + "blocks": [ + { + "bbox": [ + 137, + 247, + 272, + 321 + ], + "lines": [ + { + "bbox": [ + 137, + 247, + 272, + 321 + ], + "spans": [ + { + "bbox": [ + 137, + 247, + 272, + 321 + ], + "type": "image", + "image_path": "d922fd891d1c1c34348f1b1205873cb07c3cb862fbbf9aeb9cf7a4a7ca9b1350.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 329, + 271, + 350 + ], + "lines": [ + { + "bbox": [ + 136, + 329, + 271, + 350 + ], + "spans": [ + { + "bbox": [ + 136, + 329, + 271, + 350 + ], + "type": "text", + "content": "Fig. 6: Visual results of different control mechanisms." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 276, + 247, + 476, + 321 + ], + "blocks": [ + { + "bbox": [ + 276, + 247, + 476, + 321 + ], + "lines": [ + { + "bbox": [ + 276, + 247, + 476, + 321 + ], + "spans": [ + { + "bbox": [ + 276, + 247, + 476, + 321 + ], + "type": "image", + "image_path": "332ab2ce251515723329b0833b584c06fb71ed31c35b01002ee8aca053da6309.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 329, + 476, + 350 + ], + "lines": [ + { + "bbox": [ + 276, + 329, + 476, + 350 + ], + "spans": [ + { + "bbox": [ + 276, + 329, + 476, + 350 + ], + "type": "text", + "content": "Fig.7: Visual results of different modulation methods." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 377, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 482, + 485 + ], + "type": "text", + "content": "This is because the controllable mechanism effectively assists the rectification network in distinguishing between different degrees of distortion, thereby enhancing the generalization capability of the rectification network. The improvement brought by scalar demonstrates the crucial role of controllable mechanisms in handling fisheye images with different degrees of distortion. The further enhancement with fixed query validates the importance of employing higher-dimensional control conditions for fisheye image rectification. Lastly, our tailored learnable query significantly outperforms all other controllable methods, verifying the unique superiority of our approach in controllable rectification." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 509, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 581 + ], + "type": "text", + "content": "Controllable Modulating Block. To demonstrate the capability of the CCMB and CAMB in modulating input features with control conditions, we implemented two comparison methods with different modulating approaches: one directly uses the controlled feature (in Eq. (4)), and the other adds controlled feature and original feature in a fixed 1:1 ratio. We constructed different controllable networks using each of these methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "As shown in Tab. 3 and Fig. 7, retaining original features and fusing them with a 1:1 ratio improved by " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "0.03\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": ", highlighting the importance of preserving original features. The dynamic modulation mechanism boosted performance by " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "0.06\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": ", verifying its effectiveness in dynamic feature modulation. Furthermore, leveraging the control-attention mechanism led to a performance improvement of " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "0.12\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": ", demonstrating the superiority of CAMB for perceiving global distortion distribution, and achieving better uniform global rectification." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 156, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 156, + 90, + 447, + 102 + ], + "type": "text", + "content": "QueryCDR: Query-based Controllable Distortion Rectification Network" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 212, + 288, + 281 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 294, + 202 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 294, + 202 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 294, + 202 + ], + "type": "text", + "content": "Table 3: Ablation study of different modulation methods. Where Direct Use " + }, + { + "bbox": [ + 132, + 114, + 294, + 202 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 132, + 114, + 294, + 202 + ], + "type": "text", + "content": " means directly using the controlled feature " + }, + { + "bbox": [ + 132, + 114, + 294, + 202 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 132, + 114, + 294, + 202 + ], + "type": "text", + "content": ", Fixed Ratio means adding controlled features and original features in a 1:1 ratio, Dynamic Mechanism means the CCMB and Attention Mechanism means the CAMB." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 212, + 288, + 281 + ], + "lines": [ + { + "bbox": [ + 140, + 212, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 140, + 212, + 288, + 281 + ], + "type": "table", + "html": "
MethodPSNRSSIM
Direct UseFc20.140.655
Fixed Ratio20.170.658
Dynamic Mechanism20.200.669
Attention Mechanism20.260.671
", + "image_path": "1b32ce70ebb36fac3e942378c48aad45f2da0a516e0ebba1ff5df72cbc9d33ad.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 306, + 190, + 487, + 281 + ], + "blocks": [ + { + "bbox": [ + 304, + 114, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 304, + 114, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 114, + 482, + 180 + ], + "type": "text", + "content": "Table 4: Ablation study of different network architectures. Where " + }, + { + "bbox": [ + 304, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "x\\mathrm{C} + \\left( {{11} - x}\\right) \\mathrm{A}" + }, + { + "bbox": [ + 304, + 114, + 482, + 180 + ], + "type": "text", + "content": " denotes that the network uses CCMB in the " + }, + { + "bbox": [ + 304, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 114, + 482, + 180 + ], + "type": "text", + "content": " layers with larger feature maps and CAMB in the remaining " + }, + { + "bbox": [ + 304, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "{11} - x" + }, + { + "bbox": [ + 304, + 114, + 482, + 180 + ], + "type": "text", + "content": " layers with smaller feature maps." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 190, + 487, + 281 + ], + "lines": [ + { + "bbox": [ + 306, + 190, + 487, + 281 + ], + "spans": [ + { + "bbox": [ + 306, + 190, + 487, + 281 + ], + "type": "table", + "html": "
MethodFlops(G)Param(M)PSNRSSIM
PCN [52]12.30535.63717.970.575
11C+0A12.73637.70120.200.669
8C+3A13.38346.39820.270.665
6C+5A12.35343.24420.320.676
4C+7A12.53846.99420.310.670
0C+11A15.19051.79520.260.671
", + "image_path": "736db10e4f78782afcbdfbadba4c7658601b80fcc769b2eb85c69cb17c747330.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 308, + 483, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 483, + 476 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 483, + 476 + ], + "type": "text", + "content": "Controllable Rectification Network Architecture. As described in Sec. 3.3, rectification networks based on CCMB may ignore the long-range dependencies, while those based on CAMB may ignore the texture details and also increase computational costs. To strike a balance between these two architectures, We incorporate CCMB into the layers with larger feature map size, and integrate CAMB into the layers with smaller feature map size. To validate the effectiveness of this hybrid architecture and find an optimal combination, we conducted performance and computational cost comparisons between pure CCMB, pure CAMB, and multiple hybrid networks, as shown in Tab. 4. Pure CAMB network (i.e., Row 6) outperforms pure CCMB network (i.e., Row 2) in rectification performance but incurs higher computational costs. Hybrid network architectures effectively solve this problem, with higher performance and fewer parameters. After a trade-off between performance and parameters, we empirically choose the " + }, + { + "bbox": [ + 130, + 308, + 483, + 476 + ], + "type": "inline_equation", + "content": "6\\mathrm{C} + 5\\mathrm{A}" + }, + { + "bbox": [ + 130, + 308, + 483, + 476 + ], + "type": "text", + "content": " hybrid architecture as our final model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 507, + 220, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 507, + 220, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 507, + 220, + 520 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we propose the Query-based Controllable Distortion Rectification network for fisheye images (QueryCDR), which achieves controllable rectification at different distortion degrees without retraining. In particular, we design a series of learnable queries as control conditions to guide the rectification process. Additionally, we design two different controllable modulating blocks, achieving controllable rectification while improving image quality. Extensive experiments have demonstrated the robustness and effectiveness of our QueryCDR. In the future, it is expected to further rectify distortions of different degrees by autocontrolling mechanisms to avoid human involvement. We believe our work provides an effective solution for fisheye camera applications." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "P. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 200 + ], + "type": "text", + "content": "This work was supported in part by the NSFC under Grant 62272376, 62272380 and 62103317, the Key Research and Development Program of Shaanxi Province under Grant 2020ZDLGY04-05, the Fundamental Research Funds for the Central Universities, China (xzy022023051), the Innovative Leading Talents Scholarship of Xi'an Jiaotong University (Corresponding author: Xingsong Hou)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 217, + 197, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 217, + 197, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 197, + 228 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 239, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 239, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 480, + 262 + ], + "type": "text", + "content": "1. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 263, + 480, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 263, + 480, + 295 + ], + "spans": [ + { + "bbox": [ + 138, + 263, + 480, + 295 + ], + "type": "text", + "content": "2. Barreto, J.P., Araujo, H.: Geometric properties of central catadioptric line images and their application in calibration. IEEE Transactions on Pattern Analysis and Machine Intelligence 27(8), 1327-1333 (2005)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 296, + 480, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 296, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 138, + 296, + 480, + 327 + ], + "type": "text", + "content": "3. Barreto, J.P., Daniilidis, K.: Fundamental matrix for cameras with radial distortion. In: Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1. vol. 1, pp. 625-632. IEEE (2005)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 328, + 480, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 328, + 480, + 371 + ], + "spans": [ + { + "bbox": [ + 138, + 328, + 480, + 371 + ], + "type": "text", + "content": "4. Bogdan, O., Eckstein, V., Rameau, F., Bazin, J.C.: Deepcalib: A deep learning approach for automatic intrinsic calibration of wide field-of-view cameras. In: Proceedings of the 15th ACM SIGGRAPH European Conference on Visual Media Production. pp. 1-10 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 372, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 372, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 372, + 480, + 392 + ], + "type": "text", + "content": "5. Bukhari, F., Dailey, M.N.: Automatic radial distortion estimation from a single image. Journal of mathematical imaging and vision 45, 31-45 (2013)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 394, + 480, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 394, + 480, + 425 + ], + "spans": [ + { + "bbox": [ + 138, + 394, + 480, + 425 + ], + "type": "text", + "content": "6. Cai, B., Xu, X., Jia, K., Qing, C., Tao, D.: Dehazenet: An end-to-end system for single image haze removal. IEEE transactions on image processing 25(11), 5187-5198 (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 426, + 480, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 426, + 480, + 458 + ], + "spans": [ + { + "bbox": [ + 138, + 426, + 480, + 458 + ], + "type": "text", + "content": "7. Cai, H., He, J., Qiao, Y., Dong, C.: Toward interactive modulation for photorealistic image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 294-303 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 459, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 459, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 459, + 480, + 479 + ], + "type": "text", + "content": "8. Devernay, F., Faugeras, O.: Straight lines have to be straight. Machine vision and applications 13, 14-24 (2001)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 481, + 480, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 481, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 138, + 481, + 480, + 523 + ], + "type": "text", + "content": "9. Duan, Z., Tezcan, O., Nakamura, H., Ishwar, P., Konrad, J.: Rapid: rotation-aware people detection in overhead fisheye images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 636-637 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 524, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 524, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 134, + 524, + 480, + 567 + ], + "type": "text", + "content": "10. Feng, H., Wang, W., Deng, J., Zhou, W., Li, L., Li, H.: Simfir: A simple framework for fisheye image rectification with self-supervised representation learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12418-12427 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 568, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 568, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 134, + 568, + 480, + 599 + ], + "type": "text", + "content": "11. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: 2012 IEEE conference on computer vision and pattern recognition. pp. 3354-3361. IEEE (2012)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "type": "text", + "content": "12. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "text", + "content": "13. Guo, J., Han, K., Wu, H., Tang, Y., Chen, X., Wang, Y., Xu, C.: Cmt: Convolutional neural networks meet vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12175-12185 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 156, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 156, + 91, + 447, + 102 + ], + "type": "text", + "content": "QueryCDR: Query-based Controllable Distortion Rectification Network" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "type": "text", + "content": "14. Hartley, R., Kang, S.B.: Parameter-free radial distortion correction with center of distortion estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence 29(8), 1309-1321 (2007)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "text", + "content": "15. He, J., Dong, C., Qiao, Y.: Modulating image restoration with continual levels via adaptive feature modification layers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11056-11064 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 183, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 481, + 226 + ], + "type": "text", + "content": "16. He, J., Dong, C., Qiao, Y.: Interactive multi-dimension modulation with dynamic controllable residual learning for image restoration. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16. pp. 53-68. Springer (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "text", + "content": "17. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "18. Henrique Brito, J., Angst, R., Koser, K., Pollefeys, M.: Radial distortion self-calibration. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1368-1375 (2013)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "text", + "content": "19. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "text", + "content": "20. Kukelova, Z., Pajdla, T.: A minimal solution to radial distortion autocalibration. IEEE Transactions on Pattern Analysis and Machine Intelligence 33(12), 2410-2422 (2011)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 392 + ], + "type": "text", + "content": "21. Kumar, V.R., Klingner, M., Yogamani, S., Milz, S., Fingscheidt, T., Mader, P.: Syndistnet: Self-supervised monocular fisheye camera distance estimation synergized with semantic segmentation for autonomous driving. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 61-71 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 435 + ], + "type": "text", + "content": "22. Ledig, C., Theis, L., Huszár, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., et al.: Photo-realistic single image superresolution using a generative adversarial network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4681-4690 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "text", + "content": "23. Li, X., Zhang, B., Sander, P.V., Liao, J.: Blind geometric distortion correction on images through deep learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4855-4864 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "text", + "content": "24. Liao, K., Lin, C., Zhao, Y., Gabbouj, M.: Dr-gan: Automatic radial distortion rectification using conditional gan in real-time. IEEE Transactions on Circuits and Systems for Video Technology 30(3), 725-733 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 501, + 481, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 534 + ], + "type": "text", + "content": "25. Liao, K., Lin, C., Zhao, Y., Xu, M.: Model-free distortion rectification framework bridged by distortion distribution map. IEEE Transactions on Image Processing 29, 3707-3718 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "type": "text", + "content": "26. Lin, L., Lu, Y., Pan, Y., Chen, X.: Integrating graph partitioning and matching for trajectory analysis in video surveillance. IEEE transactions on Image Processing 21(12), 4844-4857 (2012)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 567, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 610 + ], + "type": "text", + "content": "27. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "text", + "content": "28. Liu, C., Wang, X., Li, S., Wang, Y., Qian, X.: Fsi: Frequency and spatial interactive learning for image restoration in under-display cameras. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12537-12546 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "29. Liu, C., Wang, X., Xu, X., Tian, R., Li, S., Qian, X., Yang, M.H.: Motion-adaptive separable collaborative filters for blind motion deblurring. In: Proceedings of the" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "P. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "type": "text", + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 25595-25605 (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 138, + 481, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 138, + 481, + 170 + ], + "spans": [ + { + "bbox": [ + 133, + 138, + 481, + 170 + ], + "type": "text", + "content": "30. Liu, C., Yang, H., Fu, J., Qian, X.: Learning trajectory-aware transformer for video super-resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5687-5696 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 481, + 203 + ], + "type": "text", + "content": "31. Liu, C., Yang, H., Fu, J., Qian, X.: 4d lut: learnable context-aware 4d lookup table for image enhancement. IEEE Transactions on Image Processing 32, 4742-4756 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "type": "text", + "content": "32. Mei, C., Rives, P.: Single view point omnidirectional camera calibration from planar grids. In: Proceedings 2007 IEEE International Conference on Robotics and Automation. pp. 3945-3950. IEEE (2007)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 236, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 481, + 267 + ], + "type": "text", + "content": "33. Mou, C., Wu, Y., Wang, X., Dong, C., Zhang, J., Shan, Y.: Metric learning based interactive modulation for real-world super-resolution. In: European Conference on Computer Vision. pp. 723-740. Springer (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 267, + 481, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 481, + 310 + ], + "type": "text", + "content": "34. Muhammad, K., Ahmad, J., Lv, Z., Bellavista, P., Yang, P., Baik, S.W.: Efficient deep cnn-based fire detection and localization in video surveillance applications. IEEE Transactions on Systems, Man, and Cybernetics: Systems 49(7), 1419-1434 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 310, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 310, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 481, + 342 + ], + "type": "text", + "content": "35. Peng, Z., Huang, W., Gu, S., Xie, L., Wang, Y., Jiao, J., Ye, Q.: Conformer: Local features coupling global representations for visual recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 367-376 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "type": "text", + "content": "36. Plaut, E., Ben Yaacov, E., El Shlomo, B.: 3d object detection from a single fisheye image without a single fisheye training image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3659-3667 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "text", + "content": "37. Puig, L., Bastanlar, Y., Sturm, P., Guerrero, J.J., Barreto, J.: Calibration of central catadioptric cameras using a dlt-like approach. International Journal of Computer Vision 93, 101-114 (2011)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 407, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 450 + ], + "type": "text", + "content": "38. Rashed, H., Mohamed, E., Sistu, G., Kumar, V.R., Eising, C., El-Sallab, A., Yogamani, S.: Generalized object detection on fisheye cameras for autonomous driving: Dataset, representations and baseline. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2272-2280 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 450, + 481, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 481, + 504 + ], + "type": "text", + "content": "39. Rong, J., Huang, S., Shang, Z., Ying, X.: Radial lens distortion correction using convolutional neural networks trained with synthesized images. In: Computer Vision-ACCV 2016: 13th Asian Conference on Computer Vision, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part III 13. pp. 35-49. Springer (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "type": "text", + "content": "40. Santana-Cedrés, D., Gomez, L., Aleman-Flores, M., Salgado, A., Esclarín, J., Mazorra, L., Alvarez, L.: An iterative optimization algorithm for lens distortion correction using two-parameter models. Image Processing On Line 6, 326-364 (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 536, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 579 + ], + "type": "text", + "content": "41. Scaramuzza, D., Martinelli, A., Siegwart, R.: A flexible technique for accurate omnidirectional camera calibration and structure from motion. In: Fourth IEEE International Conference on Computer Vision Systems (ICVS'06). pp. 45-45. IEEE (2006)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 579, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 481, + 611 + ], + "type": "text", + "content": "42. Shen, Z., Lin, C., Liao, K., Nie, L., Zheng, Z., Zhao, Y.: Panoformer: Panorama transformer for indoor 360 depth estimation. In: European Conference on Computer Vision. pp. 195-211. Springer (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "text", + "content": "43. Sturm, P., Ramalingam, S.: A generic concept for camera calibration. In: Computer Vision-ECCV 2004: 8th European Conference on Computer Vision, Prague, Czech Republic, May 11-14, 2004. Proceedings, Part II 8. pp. 1-13. Springer (2004)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "44. Thormahlen, T., Broszio, H., Wassermann, I.: Robust line-based calibration of lens distortion from a single view. Mirage 2003 pp. 105-112 (2003)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "text", + "content": "QueryCDR: Query-based Controllable Distortion Rectification Network" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "45. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "type": "text", + "content": "46. Wang, A., Qiu, T., Shao, L.: A simple method of radial distortion correction with centre of distortion estimation. Journal of Mathematical Imaging and Vision 35, 165-172 (2009)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 480, + 214 + ], + "type": "text", + "content": "47. Wang, W., Guo, R., Tian, Y., Yang, W.: Cfsnet: Toward a controllable feature space for image restoration. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4140-4149 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 480, + 246 + ], + "type": "text", + "content": "48. Wang, X., Yu, K., Dong, C., Tang, X., Loy, C.C.: Deep network interpolation for continuous imagery effect transition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1692-1701 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 246, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 246, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 133, + 246, + 480, + 289 + ], + "type": "text", + "content": "49. Wang, X., Yu, K., Wu, S., Gu, J., Liu, Y., Dong, C., Qiao, Y., Change Loy, C.: Esrgan: Enhanced super-resolution generative adversarial networks. In: Proceedings of the European conference on computer vision (ECCV) workshops. pp. 0-0 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 289, + 480, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 289, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 133, + 289, + 480, + 322 + ], + "type": "text", + "content": "50. Xue, Z., Xue, N., Xia, G.S., Shen, W.: Learning to calibrate straight lines for fisheye image rectification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1643-1651 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 322, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 322, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 133, + 322, + 480, + 342 + ], + "type": "text", + "content": "51. Yang, C.Y., Chen, H.H.: Efficient face detection in the fisheye image domain. IEEE Transactions on Image Processing 30, 5641-5651 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 342, + 480, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 342, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 133, + 342, + 480, + 386 + ], + "type": "text", + "content": "52. Yang, S., Lin, C., Liao, K., Zhang, C., Zhao, Y.: Progressively complementary network for fisheye image rectification using appearance flow. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6348-6357 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 386, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 386, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 133, + 386, + 480, + 418 + ], + "type": "text", + "content": "53. Yang, S., Lin, C., Liao, K., Zhao, Y.: Fishformer: Annulus slicing-based transformer for fisheye rectification with efficacy domain exploration. arXiv preprint arXiv:2207.01925 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 418, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 418, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 133, + 418, + 480, + 450 + ], + "type": "text", + "content": "54. Yang, S., Lin, C., Liao, K., Zhao, Y.: Dual diffusion architecture for fisheye image rectification: Synthetic-to-real generalization. arXiv preprint arXiv:2301.11785 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 450, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 450, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 133, + 450, + 480, + 472 + ], + "type": "text", + "content": "55. Yao, M., He, D., Li, X., Li, F., Xiong, Z.: Towards interactive self-supervised denoising. IEEE Transactions on Circuits and Systems for Video Technology (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 472, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 472, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 133, + 472, + 480, + 504 + ], + "type": "text", + "content": "56. Yin, X., Wang, X., Yu, J., Zhang, M., Fua, P., Tao, D.: Fisheyerecnet: A multicontext collaborative deep network for fisheye image rectification. In: Proceedings of the European conference on computer vision (ECCV). pp. 469-484 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 504, + 480, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 504, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 133, + 504, + 480, + 547 + ], + "type": "text", + "content": "57. Yogamani, S., Hughes, C., Horgan, J., Sistu, G., Varley, P., O'Dea, D., Uricár, M., Milz, S., Simon, M., Amende, K., et al.: Woodscape: A multi-task, multicamera fisheye dataset for autonomous driving. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9308-9318 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 547, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 547, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 133, + 547, + 480, + 579 + ], + "type": "text", + "content": "58. Zhang, K., Zuo, W., Chen, Y., Meng, D., Zhang, L.: Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing 26(7), 3142-3155 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 579, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 579, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 133, + 579, + 480, + 622 + ], + "type": "text", + "content": "59. Zhang, M., Yao, J., Xia, M., Li, K., Zhang, Y., Liu, Y.: Line-based multi-label energy optimization for fisheye image rectification and calibration. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4137-4145 (2015)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 622, + 480, + 665 + ], + "type": "text", + "content": "60. Zhang, Y., You, S., Gevers, T.: Automatic calibration of the fisheye camera for egocentric 3d human pose estimation from a single image. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1772-1781 (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "P. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 183 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 150 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 150 + ], + "type": "text", + "content": "61. Zhang, Z., Jiang, Y., Shao, W., Wang, X., Luo, P., Lin, K., Gu, J.: Real-time controllable denoising for image and video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14028-14038 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "type": "text", + "content": "62. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence 40(6), 1452-1464 (2017)" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "text", + "content": "QueryCDR: Query-based Controllable Distortion Rectification Network" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/R.A.C.E._ Robust Adversarial Concept Erasure for Secure Text-to-Image Diffusion Model/547862c1-28d0-440c-98c1-6204b0fe03dd_content_list.json b/2024/R.A.C.E._ Robust Adversarial Concept Erasure for Secure Text-to-Image Diffusion Model/547862c1-28d0-440c-98c1-6204b0fe03dd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b5e98b6dd088d4ebf8b27d191d58068b390dac6f --- /dev/null +++ b/2024/R.A.C.E._ Robust Adversarial Concept Erasure for Secure Text-to-Image Diffusion Model/547862c1-28d0-440c-98c1-6204b0fe03dd_content_list.json @@ -0,0 +1,1932 @@ +[ + { + "type": "text", + "text": "R.A.C.E.: Robust Adversarial Concept Erasure for Secure Text-to-Image Diffusion Model", + "text_level": 1, + "bbox": [ + 227, + 140, + 776, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Changhoon $\\mathbf{Kim}^{\\star 1}$ Kyle Min\\*2 Yezhou Yang $^{1}$", + "bbox": [ + 302, + 210, + 699, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Arizona State University", + "bbox": [ + 410, + 238, + 591, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Intel Labs", + "bbox": [ + 459, + 253, + 544, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{kch,yz.yang}@asu.edu kyle.min@intel.com", + "bbox": [ + 341, + 268, + 661, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In the evolving landscape of text-to-image (T2I) diffusion models, the remarkable capability to generate high-quality images from textual descriptions faces challenges with the potential misuse of reproducing sensitive content. To address this critical issue, we introduce Robust Adversarial Concept Erase (RACE), a novel approach designed to mitigate these risks by enhancing the robustness of concept erasure method for T2I models. RACE utilizes a sophisticated adversarial training framework to identify and mitigate adversarial text embeddings, significantly reducing the Attack Success Rate (ASR). Impressively, RACE achieves a $30\\%$ reduction in ASR for the \"nudity\" concept against the leading white-box attack method. Our extensive evaluations demonstrate RACE's effectiveness in defending against both white-box and black-box attacks, marking a significant advancement in protecting T2I diffusion models from generating inappropriate or misleading imagery. This work underlines the essential need for proactive defense measures in adapting to the rapidly advancing field of adversarial challenges. Our code is publicly available: https://github.com/chkimmmmm/R.A.C.E.", + "bbox": [ + 259, + 309, + 743, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Concept Erasure $\\cdot$ Responsible Image Generative Models $\\cdot$ Secure T2I Diffusion Models", + "bbox": [ + 259, + 558, + 740, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 608, + 375, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The field of text-to-image (T2I) diffusion models has garnered significant attention for their ability to produce high-quality images that can be adaptively generated from textual descriptions [36, 38]. This advancement is predicated on the training of T2I models with extensive datasets, often encompassing a range of content including copyrighted, explicit, and private materials [44-46]. Consequently, these models possess the capacity to inadvertently replicate protected images, potentially without user awareness [2, 45, 46]. The misuse of T2I models by malicious actors for misinformation or public opinion manipulation presents a significant concern [5, 27].", + "bbox": [ + 212, + 636, + 787, + 772 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In response to the challenges posed by the malicious exploitation of generative models, Stable Diffusion (SD) [38] has integrated a safety checker [37] and advocates for the utilization of a watermarking module [57]. Despite these initiatives,", + "bbox": [ + 212, + 772, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* These authors contributed equally to this work.", + "bbox": [ + 217, + 825, + 550, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/205f6a9c0c8b3e58ed77835d4547748584f97ea6246b8e4c7e1b161d6422bc5e.jpg", + "image_caption": [ + "Original SD" + ], + "image_footnote": [], + "bbox": [ + 295, + 167, + 395, + 243 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e3670e011dcfe5c24ebf75f4d7cb5399969d9b16cbd52d615c6613d0b66bffbb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 244, + 393, + 320 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2ab93595de143ec01edda69c1c4e55ab1da1f39ac614221f38505e037b364767.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 321, + 393, + 398 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8b524239b697e9c2a0775e76c3fda24933548c9629b4bf48c5ef80e170e34ee2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 398, + 393, + 472 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ab35abca0f2f9a885e54996965f1022704bdec022c9dd6363526892752069ba8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 474, + 393, + 550 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f63810ed8dc1cc22ad52d01ecb3c4cb8635ba0bcf9de0e6fdc6478154269870b.jpg", + "image_caption": [ + "\\* Added by authors" + ], + "image_footnote": [], + "bbox": [ + 295, + 551, + 393, + 626 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8d0a896ea6206157a3dfb5c120ac02e20d2cadb809a0071aada2538ca54340f5.jpg", + "image_caption": [ + "ESD" + ], + "image_footnote": [], + "bbox": [ + 398, + 169, + 517, + 244 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0edae4884de40bdd35286cfdc748a2a39a11da8492086a63adcebb2f4e198b23.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 246, + 517, + 320 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/63ce50bf4386cf8019c1ff4b23025c32858c77ef7d56fa6bf8a718cbff588a71.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 321, + 517, + 398 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0e40f06f06dd62cdabb2d14b7eb2f580c3f79193d477ea7227c34d722bb63265.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 400, + 517, + 472 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8314b795122a05433515e882e5161075270f60ca8af49b3690a3b376f45475d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 474, + 517, + 551 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/16301b4bf1a3395f993f42ba48cf79eb37f3a691e6d81a681eee7817f7b3dac6.jpg", + "image_caption": [ + "Fig. 1: Comparative demonstration of concept erasure, red teaming, and robust erasure within T2I diffusion models. The ESD method [9] removes targeted concepts from the original SD outputs, yet these concepts can be reconstructed using Unlearn-Diff [58]. Our proposed R.A.C.E. method showcases enhanced robustness against such red teaming reconstruction efforts." + ], + "image_footnote": [], + "bbox": [ + 398, + 551, + 517, + 626 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/53fec3f3a579efac278179478c008e9af430c132ce00ab5e4a247a6785921c49.jpg", + "image_caption": [ + "UnlearnDiff" + ], + "image_footnote": [], + "bbox": [ + 524, + 169, + 614, + 244 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c6472838582e434d0514bc2c8704e9ef14c99384fcdfe4261136e6408ae228e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 246, + 614, + 320 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/199929d779bde906d6917a3fad44789375cb9bba75127d4514af487d89185a8f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 321, + 614, + 398 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/de8c9cf6bff8ab750563d9234bc89b46869640a0213a1fee2e33a53f3de37797.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 400, + 614, + 472 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0f3cece5680d995eadf5147b60ad97736a4b90b2f7868c966d6658264d953a4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 474, + 614, + 551 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9c3b02b0e780b06185c14ef62bb61a32cc0b3efa7203f53b0570b21d91eb3819.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 551, + 614, + 626 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f1dba7589a490df884268a15e22a41faf6dabe3c7161af82e2169ddadd0dcc47.jpg", + "image_caption": [ + "R.A.C.E." + ], + "image_footnote": [], + "bbox": [ + 616, + 169, + 712, + 244 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5a7cf3bae1ed4f9968efde1d13dcd98d26c197b091f5628e39919c3f018f5877.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 246, + 712, + 320 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/779d885d2df40145b9f9aaae6cd7a6f2a0a5b126fc3481ed6ff8a89d9598a4a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 321, + 712, + 398 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ff1b1d8b9af8d737b933e34e1bb7580175a2e9ee0c0f51e1c7eee67aa77ef45c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 400, + 712, + 472 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dc46c7a4e0580e1a80d2a929ba3bae21974625f8d0b22e150d26fd91ebbad718.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 474, + 712, + 551 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7eb090393e3bb2b78909bb39ca71de1df5e293ce8c536e9881faaefb73871b68.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 551, + 712, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the reliance on post-hoc interventions presents limitations due to their potential for circumvention [7, 18]. Consequently, the research community is pivoting towards formulating methodologies that embed safety protocols directly within the image generation pipeline. This approach ensures that content security is an intrinsic aspect of image creation [7, 18, 19, 30, 50, 54, 55]. Although these methods are valuable for identifying the source of content after an incident, their reactive", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "C. Kim*, K. Min*, Y. Yang", + "bbox": [ + 271, + 113, + 457, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "nature highlights the necessity for proactive strategies. This includes pioneering techniques attempting early removal of sensitive content from T2I models [9, 21], thereby preempting the production of inappropriate or harmful material.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the challenge of removing undesirable content from T2I models, even when users attempt to circumvent content restrictions, recent research has focused on the development of concept erasure techniques within T2I diffusion models [9, 11, 21, 56]. These techniques primarily aim to eliminate specific concepts (e.g., \"nudity\") by altering the text embeddings associated with these concepts to neutral representations. Despite these efforts, there remains a vulnerability wherein erased concepts can be reconstructed. This is achieved by identifying text tokens that closely align with the visual embeddings of the targeted concepts, thus enabling the regeneration of prohibited content [3, 47, 49, 53, 58]. This issue is illustrated in Fig. 1, demonstrating that even with concept erasure, T2I models can be manipulated through prompt modification to regenerate the restricted content. This underscores the imperative for a more robust concept erasure methodology that can withstand such reconstruction attempts, ensuring the integrity of content generation within T2I models.", + "bbox": [ + 212, + 196, + 787, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Acknowledging the imperative for enhanced concept erasure methodologies within T2I models, we pose a critical question: \"Is it feasible to develop a concept erasure approach that is resilient against reconstruction efforts?\" In pursuit of this, we introduce R.A.C.E. (Robust Adversarial Concept Erase), a novel strategy aimed at bolstering the resilience of concept erasure techniques against adversarial manipulations, as delineated in Algorithm 1. At the heart of RACE lies an adversarial training framework, leveraging insights from the effort of adversarial robustness [26]. Our method effectively identifies adversarial text embeddings capable of reconstructing erased concepts and then facilitates their integration into the T2I concept erasure workflow.", + "bbox": [ + 212, + 414, + 787, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A pivotal aspect of RACE is its ability to efficiently uncover adversarial text embeddings within a single time step of the diffusion process, an approach elaborated in Section 3.2. This efficiency not only streamlines the process of identifying adversarial examples but also facilitates the integration of our adversarial attack mechanism into the concept erasure workflow. To demonstrate the robustness of RACE, we have carried out an extensive array of experiments. These experiments validate the effectiveness of RACE in countering diverse red teaming strategies, with detailed results presented in Section 4. Our empirical investigations underscore the capacity of RACE to significantly enhance the robustness of T2I models against both white-box and black-box attacks across a broad spectrum of target concepts, including artistic, explicit, and object categories.", + "bbox": [ + 212, + 571, + 787, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We summarize the three main contributions here:", + "bbox": [ + 238, + 743, + 598, + 757 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We are the first to present, to our knowledge, an adversarial training approach specifically designed to fortify concept erasure methods against prompt-based adversarial attacks without introducing additional modules.", + "- Our method, RACE, implements a computationally efficient adversarial attack method that can be plugged into the concept erasing method." + ], + "bbox": [ + 225, + 763, + 795, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "R.A.C.E. for Secure T2I Diffusion", + "bbox": [ + 503, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We show RACE significantly improves T2I models' robustness against prompts based on white/black box attacks.", + "bbox": [ + 227, + 146, + 795, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 200, + 395, + 217 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a comprehensive overview of additional related works, please refer to the supplementary material.", + "bbox": [ + 212, + 234, + 784, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Text-to-Image Synthesis. The field of generative models has seen remarkable advancements, notably extending their capabilities beyond generating photorealistic images [6, 17] to include Text-to-Image (T2I) synthesis [29, 33, 36, 38, 41]. This progress has led to the development of fine-tuning techniques that allow for the customization of T2I models to user-specific needs [8, 22, 32, 40, 48], thereby enabling the creation of highly realistic images that align closely with textual prompts. However, the potential for misuse by malicious entities, using these models for purposes such as spreading misinformation [5, 27], raises significant concerns. This underscores the urgency of devising protective measures to mitigate the risk of such exploitations.", + "bbox": [ + 212, + 280, + 787, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Advanced Techniques in Concept Erasure for T2I Diffusion Models. Within the realm of machine unlearning, concept erasure for T2I Diffusion models has recently emerged as a critical area of research, focusing on the removal of sensitive or copyrighted concepts from T2I models. Methods to achieve this include guiding the image generation process or adjusting the model's weights to exclude these elements [9,11,21,24,28,43,56]. Notably, techniques by Gandikota et al. [9] and Kumari et al. [21] involve mapping sensitive concepts to null entities or benign equivalents by fine-tuning the weights of Stable Diffusion (SD) [38] models, effectively preventing the generation of undesirable content. Despite these advancements, red teaming methods have exposed potential loopholes, indicating that erased concepts might be regenerated through meticulously designed text prompts. Addressing this issue, our work contributes an adversarial training strategy aimed at bolstering the resilience of Stable Diffusion models against such text-prompt-based attacks [3,49,58], thereby enhancing the security and integrity of the content generation.", + "bbox": [ + 212, + 446, + 787, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Robustness Evaluation via Red Teaming in T2I Models. While various safety measures have been proposed to shield SD models from misuse, red teaming strategies reveal vulnerabilities that still allow for circumvention. Research has demonstrated that techniques like Textual Inversion [8] can be exploited to regenerate content previously erased from SD models [34], prompting the development of countermeasures aimed at safeguarding against such inversions [52,59]. In real-world applications, T2I services such as Midjourney predominantly rely on user-provided text prompts, making them susceptible to prompt-based red teaming attacks [3, 47, 49, 53, 58]. These methods employ sophisticated prompt optimization techniques to restore images containing erased content, with their", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "C. Kim*, K. Min*, Y. Yang", + "bbox": [ + 271, + 114, + 457, + 128 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Robust Adversarial Concept Erasure: RACE Algorithm" + ], + "code_body": "Input: Diffusion Model $\\varPhi_{\\theta}$ , frozen diffusion model $\\varPhi_{\\theta*}$ , scheduler $S$ target concept $c$ training steps $M$ , adversarial steps $N$ perturbation limit $\\epsilon$ attack step size $\\alpha$ for $i = 0,\\dots,M$ do Sample noise $n\\sim \\mathcal{N}(0,1)$ , timestep $t\\sim \\mathcal{U}(1,1000)$ Initialize $\\delta \\sim \\mathcal{U}(-\\epsilon ,\\epsilon)$ Denoise $z_{t} = S(n,t,c)$ for $j = 0,\\ldots ,N$ do Perform targeted attack $\\delta = \\delta +\\alpha \\cdot sign(\\nabla_{\\delta} - L_{SD}(\\varPhi_{\\theta},z_{t},t,c,\\delta))$ Clamp $\\delta$ within $[- \\epsilon ,\\epsilon ]$ end for $\\theta = \\theta -\\nabla_{\\theta}L_{RACE}(\\varPhi_{\\theta},\\varPhi_{\\theta^{*}},z_{t},t,c,\\delta)$ \nend for \nreturn $\\varPhi_{\\theta}$", + "bbox": [ + 215, + 164, + 787, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "efficacy contingent upon the level of model access—categorized into white-box approaches, which utilize SD's U-Net [39] for prompt optimization [3, 58], and black-box strategies, where such access is restricted [47, 49, 53]. Both approaches establish formidable benchmarks in attack success rates, as detailed in Tab. 2. However, the landscape lacks robust defense mechanisms against prompt-based red teaming, primarily due to the prohibitive computational demands associated with identifying adversarial prompts—a challenge that renders traditional adversarial training approaches impractical. Addressing this gap, our work introduces a novel defense strategy tailored to counteract prompt-based red teaming attacks, marking a significant step forward in fortifying T2I diffusion models against adversarial threats.", + "bbox": [ + 212, + 378, + 787, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 571, + 330, + 587 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our methodology aims to expunge target concepts from T2I diffusion models through an adversarial training framework [26]. Initially, in Section 3.1, we establish the foundation by formally introducing the notations and the rationale underpinning our approach. Following this, Section 3.2 details our proposed adversarial attack, specifically designed for robust concept erasure, and delineates its integration into the adversarial training regime.", + "bbox": [ + 212, + 604, + 787, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 720, + 372, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stable Diffusion Models. Our method is built upon the Stable Diffusion Model (SD) [38], which operates as the foundational architecture for our concept erasure technique. The SD model is composed of two primary elements: firstly, an image autoencoder that has been pre-trained on a diverse and extensive image dataset [6]. Within this autoencoder, an encoder function $\\mathcal{E}(\\cdot)$ transforms an input image $x$ into a latent representation $z = \\mathcal{E}(x)$ . Conversely, a decoder", + "bbox": [ + 212, + 748, + 787, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "R.A.C.E. for Secure T2I Diffusion", + "bbox": [ + 503, + 114, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3de98e383fa876bf1af1398071b4dce0d2fa828873eda862604549597c3717bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 143, + 733, + 311 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4b974e44b7904758df9875f1592b6c7943e1edc22eb5b198c4329583f9186bf8.jpg", + "image_caption": [ + "Fig. 2: Single-Timestep Adversarial Attack Efficacy. This figure illustrates the Attack Success Rate (ASR) across various timesteps, alongside representative images. Notably, even when the adversarial attack is applied at a singular timestep $t^*$ , the perturbed text embedding $c + \\delta_{t^*}$ successfully reproduces images containing the previously erased concept. For method details, see Sec. 3.2." + ], + "image_footnote": [], + "bbox": [ + 282, + 311, + 728, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "function $\\mathcal{D}(\\cdot)$ aims to reconstruct the input image from its latent form, where $\\mathcal{D}(z) = \\hat{x}\\approx x$", + "bbox": [ + 212, + 561, + 787, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The second element is a U-Net [39]-based diffusion model trained to craft latent representations within the acquired latent space. This model facilitates the conditioning on either class labels or text embeddings derived from training data. Let us denote by $c = \\mathcal{E}_{txt}(y)$ the textual embedding encoded from a conditioning text prompt $y$ , where $\\mathcal{E}_{txt}$ symbolizes the text encoder, such as CLIP [35]. Under these constructs, the SD training objective is encapsulated by the loss function:", + "bbox": [ + 212, + 590, + 787, + 684 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {S D} := \\mathbb {E} _ {n \\sim \\mathcal {N} (0, 1), z, c, t} \\left[ | | n - \\Phi_ {\\theta} \\left(z _ {t}, t, c\\right) | | _ {2} ^ {2} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 695, + 785, + 714 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $t$ indexes the time step, $n$ represents a noise sample drawn from a standard Gaussian distribution, $z_{t}$ is the perturbed version of $z$ up to time step $t$ , and $\\varPhi_{\\theta}$ is the denoising network based on a U-Net architecture. During inference, a random noise sample is procured from a Gaussian distribution and denoised using $\\varPhi_{\\theta}$ following a scheduler $\\mathcal{S}$ , operating over a sequence of predetermined time steps $T$ . The resulting denoised latent, $z_{0}$ , is then decoded to produce the final image, $\\hat{x} = \\mathcal{D}(z_0)$ .", + "bbox": [ + 212, + 719, + 787, + 825 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "C. Kim*, K. Min*, Y. Yang", + "bbox": [ + 271, + 114, + 457, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Erase and Reconstruction of Target Concept. The objective of concept erasure is to remove specific concepts, such as \"nudity\", from the latent space of a pre-trained diffusion model. The Erased Stable Diffusion (ESD) approach [9] introduces a method for excising these concepts from the latent representations within the Stable Diffusion framework. The erasure loss is formalized as follows:", + "bbox": [ + 214, + 146, + 787, + 220 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {e r a s e} := \\left\\| \\Phi_ {\\theta} \\left(z _ {t}, t, c\\right) - \\left(\\Phi_ {\\theta^ {*}} \\left(z _ {t}, t\\right) - \\eta \\left(\\Phi_ {\\theta^ {*}} \\left(z _ {t}, t, c\\right) - \\Phi_ {\\theta^ {*}} \\left(z _ {t}, t\\right)\\right)\\right) \\right\\| _ {2} ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 264, + 238, + 785, + 256 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\varPhi_{\\theta^{*}}$ represents the frozen denoising U-Net, and $\\eta$ denotes the guidance scale associated with classifier-free guidance [15]. ESD fine-tunes $L_{erase}$ with respect to $\\theta$ , guiding $\\varPhi_{\\theta}$ to produce outputs where the target concept is effectively nullified. Crucially, this process does not necessitate additional datasets; it operates successfully with only a concise textual description.", + "bbox": [ + 212, + 263, + 782, + 338 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Conversely, red teaming efforts aim to counteract concept erasure by crafting adversarial text prompts $\\tilde{y}$ capable of resurrecting the erased concept in the generated image. White-box methods [3,58] engage in this adversarial prompt optimization by leveraging the gradients of $\\varPhi_{\\theta}$ . Meanwhile, black-box approaches [47, 49, 53] aim to achieve comparable outcomes without reliance on the gradients of $\\varPhi_{\\theta}$ . Both approaches pose significant computational demands, which presents challenges for their integration into an adversarial training framework.", + "bbox": [ + 212, + 339, + 787, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Adversarial Training on Concept Erased Diffusion Models", + "text_level": 1, + "bbox": [ + 214, + 470, + 738, + 486 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Motivation. The primary aim of T2I diffusion models is to generate high-quality images conditioned on specific prompts. Intriguingly, the SD model's loss function, as depicted in Eq.(1), can also facilitate image classification tasks [23]. This classification capability is derived by applying Bayes' Theorem to the model's predictions $p_{\\theta}(x|c_i)$ and the prior distribution $p(c)$ across a set of conditions $c_i$ , where each $c_i = \\mathcal{E}_{txt}(y_i)$ represents a textual embedding of the prompt $y_i$ :", + "bbox": [ + 214, + 498, + 787, + 590 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} (c _ {i} | x) = \\frac {p (c _ {i}) p _ {\\theta} (x | c _ {i})}{\\sum_ {j} p (c _ {j}) p _ {\\theta} (x | c _ {j})}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 604, + 785, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Notably, the prior terms $p(c)$ can be disregarded when they are uniformly distributed over the prompts $c_i$ (i.e., $p(c_i) = \\frac{1}{N}$ ). In the context of diffusion models, directly computing $p_{\\theta}(x|c_i)$ is computationally challenging, leading to the reliance on the computation of $\\log p_{\\theta}(x|c_i)$ and the utilization of the Evidence Lower Bound (ELBO) for optimization purposes. Leveraging approximations introduced in [14], we can approximate the posterior distribution over prompts $c_i$ as follows:", + "bbox": [ + 214, + 645, + 787, + 750 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(c _ {i} \\mid x\\right) = \\frac {\\exp \\left\\{- \\mathbb {E} _ {z , n , t} \\left[ | | n - \\Phi_ {\\theta} \\left(z _ {t} , t , c _ {i}\\right) | | ^ {2} \\right] \\right\\}}{\\sum_ {j} \\exp \\left\\{- \\mathbb {E} _ {z , n , t} \\left[ | | n - \\Phi_ {\\theta} \\left(z _ {t} , t , c _ {j}\\right) | | ^ {2} \\right] \\right\\}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 325, + 761, + 785, + 797 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Building on this foundation, the Diffusion Classifier [23] proposes a method for estimating the class of a given image $x$ by finding argmin of the following", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "R.A.C.E. for Secure T2I Diffusion", + "bbox": [ + 503, + 114, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "expression:", + "bbox": [ + 215, + 147, + 299, + 161 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {c} {\\operatorname {a r g m i n}} \\sum_ {t = 1} ^ {T} \\left[ \\left| \\left| n - \\Phi_ {\\theta} \\left(z _ {t}, t, c\\right) \\right| \\right| ^ {2} \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 387, + 161, + 785, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $z_{t} = \\mathcal{E}(x)$ represents the latent encoding of the image $x$ and the goal is to identify the label $c$ from the available set of classes $c_{i}$ . A notable insight from the Diffusion Classifier is the feasibility of classifying image $x$ even when computations are performed using a single time step $t^{*}$ .", + "bbox": [ + 212, + 205, + 787, + 267 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Single-Timestep Adversarial Attacks on Erased T2I Models. Leveraging insights from the Diffusion Classifier, we re-conceptualize the SD model's loss function in Eq.(1) as a classification mechanism. This perspective allows us to view Textual Inversion (TI) [8] as a form of targeted adversarial attack, where the objective is to optimize the conditioning text embedding $c$ to regenerate the image $x$ [34, 58]. Notably, TI is computationally intensive as it necessitates optimization across all time steps.", + "bbox": [ + 212, + 281, + 787, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prompted by these considerations, our investigation centers around a critical inquiry: Can adversarial text embeddings be identified with just a single timestep? We investigate whether adversarial text embeddings can be effectively identified at a singular timestep $t^*$ . Our approach is geared towards nullifying the embedding of a target concept $c$ and its proximate embeddings that might facilitate the regeneration of an erased concept image $\\tilde{x}$ , such as an explicit image. To this end, we devise a targeted adversarial attack to produce $\\tilde{x}$ from a concept-erased diffusion model $\\varPhi_{\\theta}$ by introducing an adversarial perturbation $\\delta$ . The perturbation $\\delta$ is determined through the optimization:", + "bbox": [ + 212, + 387, + 787, + 523 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {a r g m i n} \\left| | n - \\Phi_ {\\theta} \\left(\\tilde {z} _ {t}, t, c + \\delta\\right) | | _ {2} ^ {2}, \\right. \\tag {6} \\\\ | | \\delta | | _ {\\infty} \\leq \\epsilon \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 532, + 785, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\tilde{z} = \\mathcal{E}(\\tilde{x})$ denotes the latent representation of image $\\tilde{x}$ , $\\epsilon$ is a small number, and $c = \\mathcal{E}_{txt}(y)$ encodes the textual embedding of the targeted concept, for instance, $y =$ \"nudity\". The Projected Gradient Descent (PGD) algorithm [26] is employed to address this optimization problem. Specifically, when selecting timestep $t^* = 500$ as the critical adversarial point, $\\tilde{z}_t$ undergoes denoising via $\\Phi_{\\theta}(\\tilde{z}_t,t,c)$ transitioning from timesteps $t = 1000$ to $t = 500$ . Subsequently, the targeted adversarial approach outlined in Eq. (6) is executed to determine $\\delta_{t^*}$ at $t = 500$ . In subsequent denoising steps, $\\tilde{z}_t$ is denoised considering the introduced perturbation $c + \\delta_{t^*}$ .", + "bbox": [ + 212, + 566, + 787, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method introduces a distinctive single-timestep adversarial attack, contrasting with prior approaches that required optimization over a wide range of timesteps [3,58]. This approach enables the seamless incorporation of our attack strategy into the adversarial training process specifically for T2I concept erasure.", + "bbox": [ + 212, + 704, + 787, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To assess the efficacy of our method, we execute tests on a $\\varPhi_{\\theta}$ model trained for removing the \"nudity\" concept via the ESD method. Utilizing 142 nuditycentric prompts from the I2P dataset [43], we systematically select timesteps $t^{*}$ at 100-step intervals within the [1, 1000] range and compute the Attack Success Rate (ASR), as formulated in Eq.(8), for each $t^{*}$ . The results, depicted in", + "bbox": [ + 212, + 765, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "C. Kim*, K. Min*, Y. Yang", + "bbox": [ + 271, + 113, + 457, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 2, unveil a notable capability: the reconstruction of the erased concept is feasible even with the attack confined to a single timestep $t^*$ . Interestingly, after $t^* = 500$ , we observe an increasing trend in the ASR. The modified images resulting from the attack visually convey the transition back to the erased concept and reflect the ASR patterns. For additional insights into various concepts and their corresponding ASR trends across different timesteps $t^*$ , the supplementary material offers further information.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Adversarial Training for T2I Concept", + "text_level": 1, + "bbox": [ + 215, + 268, + 542, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Erasure. Motivated by the findings from our single-timestep adversarial attack experiments, we explore the potential of such attacks to enhance the robustness of concept erasure in T2I models, posing the question: Can adversarial attacks improve the resilience of concept erasure mechanisms?", + "bbox": [ + 212, + 284, + 542, + 387 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/ef7f6312a7d0e849762c9e06bdbb86da6b7bc500a9da64283bf985c0fa5e8b53.jpg", + "table_caption": [ + "Table 1: Performance comparison of $L_{\\text{erase}}$ and $L_{\\text{RACE}}$" + ], + "table_footnote": [], + "table_body": "
LeraseLRACE
I2P [43]0.080.05
FID [13]33.1225.16
CLIP-Score [12]0.7260.745
", + "bbox": [ + 555, + 313, + 785, + 381 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "RACE distinguishes itself from existing approaches [9,21] by aiming to eliminate not only the targeted concept's embedding but also its adjacent embedding within the model's latent space, which could otherwise lead to the inadvertent generation of the erased concept by $\\varPhi_{\\theta}$ . We incorporate our adversarial attack into the erasure loss function ( $L_{erase}$ ), yielding an enhanced adversarial training loss:", + "bbox": [ + 212, + 388, + 787, + 479 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {R A C E} := \\left| \\left| \\Phi_ {\\theta} \\left(z _ {t}, t, c + \\delta\\right) - \\left(\\Phi_ {\\theta^ {*}} \\left(z _ {t}, t\\right) - \\eta \\left(\\Phi_ {\\theta^ {*}} \\left(z _ {t}, t, c\\right) - \\Phi_ {\\theta^ {*}} \\left(z _ {t}, t\\right)\\right)\\right) \\right| \\right| _ {2} ^ {2}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 493, + 785, + 511 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This method is deliberate, substituting the concept embedding $c$ with $c + \\delta$ within the trainable parameters of $\\varPhi_{\\theta}$ . This precise adjustment ensures enhanced fidelity of the generated images by mapping the $\\epsilon$ -neighborhood of the concept embedding to its null representation. Comparative metrics between the direct substitution in $L_{erase}$ and the strategic use of $L_{RACE}$ are provided in Tab. 1. The latter approach demonstrates promising reductions in ASR for \"nudity\" prompts within I2P dataset [43] and improvements in image quality metrics, as assessed on the MS-COCO [25]. The RACE methodology is comprehensively detailed in Algorithm 1. To validate RACE's effectiveness, we evaluate the ASR against both white-box and black-box attacks, as elaborated in Sec. 4.", + "bbox": [ + 212, + 522, + 787, + 674 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 700, + 375, + 718 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental Setting", + "text_level": 1, + "bbox": [ + 215, + 734, + 439, + 752 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. Our assessment of the RACE framework spans various domains, including artistic styles, explicit concepts, and identifiable objects, in line with established benchmarks [3,9,58]. To ensure a uniform image generation process, we standardize key hyperparameters such as the scale of classifier-free guidance and random seeds. Artistic style evaluations leverage shared text prompts from", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "R.A.C.E. for Secure T2I Diffusion", + "bbox": [ + 503, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ESD [9] and UnlearnDiff [58]. For explicit content, we draw from the inappropriate image prompt benchmark [43], selecting a diverse set of prompts encompassing 142 nudity, 98 illegal acts, and 101 violence instances. Objects are curated from a subset of Imagenette [16], known for its distinct and recognizable classes, with text prompts synthesized via ChatGPT [31] following the approach by Kumari et al. [21]. It's crucial to note that these datasets serve solely to gauge the effectiveness of adversarial attacks; the adversarial training component of RACE does not necessitate the use of these specific prompts.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training Details. In validating RACE, we choose the Erased Stable Diffusion (ESD) model [9] for its ability to erase a wide range of concepts, serving as an ideal testbed to showcase RACE's efficacy. We integrate RACE with ESD, optimizing with attack parameters $\\epsilon = 0.1$ and $\\alpha = \\epsilon /4$ . The optimization of model parameters $\\theta$ utilizes the Adam optimizer [20] at a learning rate of $1e - 5$ , consistent with ESD. Additional details on attack parameter selection are available in the supplementary material.", + "bbox": [ + 212, + 282, + 787, + 388 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Red Teaming Methods. To rigorously test RACE's robustness, we deploy a comprehensive suite of adversarial attacks, spanning both white-box and black-box approaches. Initial assessments utilize the I2P red teaming prompt dataset [43]. In black-box scenarios, we employ PEZ [49], which crafts adversarial prompts via CLIP [35]. In the white-box scenario, methods like P4D [3] and UnlearnDiff [58] are used, which generate adversarial prompts by leveraging gradients from the SD.", + "bbox": [ + 212, + 402, + 797, + 508 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation. To gauge the robustness of RACE, we employ domain-specific classifiers: a ViT-base model [51] pre-trained on ImageNet [4] and fine-tuned on WikiArt [42] for artistic styles, Nudenet [1] for explicit content, and ResNet-50 [10] trained on ImageNet for object removal. We measure robustness using the Attack Success Rate (ASR):", + "bbox": [ + 212, + 525, + 787, + 602 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A S R} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathbb {1} \\left(f \\left(S D \\left(\\tilde {y} _ {i}\\right)\\right) = \\tilde {y} _ {i}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 613, + 785, + 655 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\tilde{y}_i$ is the adversarial prompt, $f$ is the classifier, and $N$ is the number of prompts. Additionally, we assess image quality after applying RACE by generating 5,000 images from the MS-COCO [25] test set, computing the Frechet Inception Distance (FID) score [13] and the CLIP score [12] to evaluate RACE's impact on image fidelity while ensuring concept erasure.", + "bbox": [ + 212, + 666, + 787, + 743 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Robust Concept Erase against Red Teaming", + "text_level": 1, + "bbox": [ + 214, + 766, + 627, + 782 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In our comprehensive analysis, RACE undergoes a series of red teaming evaluations [3, 43, 49, 58], encompassing both white-box and black-box techniques aimed at regenerating concepts targeted by RACE for erasure, as depicted in", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "C. Kim*, K. Min*, Y. Yang", + "bbox": [ + 271, + 114, + 457, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/d6f45bec7d9a4297ef2a3d4b8324e15041e8849f1ab497415b676c08a2e93b15.jpg", + "table_caption": [ + "Table 2: Attack Success Rate (ASR) against white $(\\bigcirc)$ /black $(\\bullet)$ box attacks. We conduct experiments on artistic style (Van Gogh), explicit concepts (nudity, violence, and illegal acts), and objects (church, golf ball, and parachute). \"ESD/RACE-Concept\" denotes the concept erased from the model. We also measure the quality of T2I models. We can observe that RACE reduces ASR by over $30\\%$ (-0.30 or below) in absolute value on Van Gogh, nudity, and church for UnlearnDiff [58], which is the previous SOTA attack method." + ], + "table_footnote": [], + "table_body": "
PromptsPEZ [49]P4D [3]UnlearnDiff [58]CLIP-Score [12]FID [13]
White/Black Box●●○○--
ESD [9]-VanGogh0.040.000.260.360.799719.16
ESD [9]-Nudity0.140.080.750.800.793118.88
ESD [9]-Violence0.270.130.840.790.783421.55
ESD [9]-Illegal0.290.200.890.850.785421.50
ESD [9]-Church0.160.000.580.680.789619.68
ESD [9]-GolfBall0.040.000.160.160.773820.64
ESD [9]-Parachute0.060.040.480.600.786519.72
RACE-VanGogh0.00 (-0.04)0.00 (-0.00)0.00 (-0.26)0.04 (-0.32)0.802420.65
RACE-Nudity0.05 (-0.09)0.02 (-0.06)0.49 (-0.26)0.47 (-0.33)0.745225.16
RACE-Violence0.11 (-0.16)0.08 (-0.05)0.75 (-0.09)0.68 (-0.11)0.737428.71
RACE-Illlegal0.20 (-0.09)0.13 (-0.07)0.85 (-0.04)0.80 (-0.05)0.759124.87
RACE-Church0.02 (-0.14)0.00 (-0.00)0.26 (-0.32)0.38 (-0.30)0.773023.92
RACE-GolfBall0.00 (-0.04)0.00 (-0.00)0.10 (-0.06)0.06 (-0.10)0.748025.38
RACE-Parachute0.02 (-0.04)0.00 (-0.04)0.24 (-0.24)0.38 (-0.22)0.757026.42
", + "bbox": [ + 215, + 253, + 785, + 449 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Fig.1. The comparative analysis of ASR presented in Tab.2 spans diverse conceptual domains, from \"Van Gogh\"-inspired artistry to explicit content such as \"nudity\" and \"violence\", extending to tangible objects like \"churches\", \"golf balls\", and \"parachutes\". Remarkably, RACE consistently diminishes ASR, notably surpassing $30\\%$ for \"Van Gogh\" styles, \"nudity\", and \"church\" categories, particularly outperforming UnlearnDiff [58], the current state-of-the-art in white-box adversarial methodologies. This marked decline in ASR underscores RACE's heightened robustness and delineates its capability as a potent, computationally efficient defense mechanism for T2I diffusion models against intricate adversarial attacks. Crucially, RACE's methodological advantage stems from its independence from external imagery or prompts, diverging from traditional red teaming techniques reliant on such data for prompt generation.", + "bbox": [ + 212, + 460, + 787, + 643 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As demonstrated in Tab.2, RACE achieves a significant $33\\%$ reduction in ASR for the \"nudity\" concept, underscoring its effectiveness. To further elucidate how RACE enhances ASR, we analyze the specific categories or elements it targets for removal. Utilizing Nudenet, we enumerate the body parts generated by various models—original SD, ESD, UnlearnDiff, and RACE—when prompted with nudity-related inputs from the I2P dataset. Illustrated in Fig.3, the analysis reveals that while ESD enhances the original SD's resilience to such prompts, UnlearnDiff manages to bypass ESD's defenses, reconstructing explicit content. In contrast, RACE maintains its robustness even against the sophisticated UnlearnDiff attacks, showcasing the advanced protective capabilities of our approach in safeguarding against the regeneration of sensitive content.", + "bbox": [ + 212, + 643, + 787, + 808 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "One caveat to mention is that our experiments reveal a nuanced trade-off between robustness and image quality. While artistic style erasures maintain", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "R.A.C.E. for Secure T2I Diffusion", + "bbox": [ + 503, + 114, + 730, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/8109cb659eb06404b45ab762e449479b0fe4ce0bd162f8321df64716e87e71a3.jpg", + "image_caption": [ + "Fig. 3: Although ESD significantly reduces the chance of generating images with exposed body parts, state-of-the-art red teaming methods, such as UnlearnDiff, can be used to bypass ESD's defense and reconstruct explicit content. RACE and its variant can effectively defend the malicious attempts to reconstruct explicit content from the ESD model that erased the concept of nudity." + ], + "image_footnote": [], + "bbox": [ + 220, + 145, + 782, + 320 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "quality metrics, erasures of other concepts inadvertently degrade image quality. This divergence could be attributed to methodological differences; erasing artistic styles predominantly involves fine-tuning the cross-attention layers of SD as per ESD guidelines, whereas erasing other concepts necessitates adjustments in non-cross-attention layers [9]. Another plausible explanation for the observed trade-off between robust concept erasure and overall image quality could relate to the inherent complexity of differentiating between closely related or overlapping concepts within the model's latent space. As RACE intensifies adversarial robustness, it may inadvertently alter the delicate equilibrium within these conceptual overlaps, leading to unintended modifications in adjacent, non-targeted conceptual representations. This issue highlights the complex tension between precise concept erasure and maintaining the model's overall integrity against attacks. To address the quality concerns arising from this trade-off, we explore a potential strategy for improvement in Sec. 4.4, aiming to reduce the trade-off between targeted erasure with high quality and the model's defensive robustness.", + "bbox": [ + 212, + 412, + 787, + 641 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Disentanglement", + "text_level": 1, + "bbox": [ + 215, + 661, + 401, + 676 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Investigating RACE's disentanglement performance, our study focuses on its capability to precisely erase intended concepts without impacting other elements. The evaluation spans both qualitative and quantitative measures. On the qualitative front, we configure separate RACE-enhanced models to specifically erase artistic imprints such as \"Van Gogh\", \"Thomas Kinkade\", and \"Kilian Eng\" from the SD. The illustrative outcomes, showcased in Fig. 4, underscore RACE's erasure precision, ensuring that the excision of one artistic style doesn't lead to the collateral removal of others. This meticulous erasure extends to discrete object concepts like \"church\", \"golf ball\", and \"parachute\", with the generated images further affirming the method's discernment.", + "bbox": [ + 212, + 686, + 787, + 838 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "C. Kim*, K. Min*, Y. Yang", + "bbox": [ + 271, + 114, + 457, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/21077ca7c798cbb9362adaf70208ccc27f451b96eaa0539861ebd4937fecec10.jpg", + "image_caption": [ + "Fig. 4: RACE's Disentanglement in Concept Erasure. This figure highlights RACE's precision in erasing specific concepts, as shown in diagonal images, while preserving unrelated concepts, which is evident in off-diagonal images. For reference, baseline images generated by the original Stable Diffusion (SD) model are also presented." + ], + "image_footnote": [], + "bbox": [ + 222, + 145, + 787, + 306 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/e0c3753aefc178cf1d9f91f982a0d6e813757c418db372e57569ed1a038f9fc3.jpg", + "table_caption": [ + "Table 4: Ablation Studies for Performance Improvement. Using the \"nudity\" concept, we evaluate the effectiveness of adding weight regularization and close concept keywords. ASR and quality metrics are measured to assess performance." + ], + "table_footnote": [], + "table_body": "
I2P [43]PEZ [49]P4D [3]UnleranDiff [58]CLIP-Score [12]FID [13]
ESD0.140.080.750.800.793118.88
RACE0.050.020.490.470.745225.16
RACE+Reg.0.070.020.600.620.759324.42
RACE+keywords0.020.010.420.460.720130.97
", + "bbox": [ + 215, + 438, + 785, + 513 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Quantitatively, we generate 5,000 prompts with varied random seeds, such as \"an image of a [class name]\", to produce a diverse set of images, subsequently evaluated using a pre-trained ResNet-50 classifier for top-1 accuracy. As shown in Tab. 3, RACE showcases an improved capability for object concept erasure compared to ESD. Despite this, there exists a slight decrement in classification accuracy for non-target classes. This effect", + "bbox": [ + 212, + 540, + 509, + 705 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "likely stems from RACE's method of targeting the $\\epsilon$ -neighborhood surrounding the intended concept, potentially influencing proximate concepts. Nevertheless, we can observe from Fig. 4 that RACE can precisely erase the target concept with minimal visual impacts on other concepts.", + "bbox": [ + 212, + 705, + 782, + 767 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/ed72e5d29478760d90c0002d54b14515404616f7b8a3647bc40b24120586944e.jpg", + "table_caption": [ + "Table 3: Accuracy of erased and non-erased classes in Imagenet [4]. We evaluate the classification accuracy of ESD and RACE for the target erased concepts and other non-target concepts." + ], + "table_footnote": [], + "table_body": "
Erased ConceptAcc. ErasedAcc. Others
ESDRACEESDRACE
Church0.160.020.570.53
Golf Ball0.040.000.450.35
Parachute0.060.020.570.41
", + "bbox": [ + 521, + 609, + 785, + 695 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Discussion", + "text_level": 1, + "bbox": [ + 215, + 786, + 349, + 801 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Potential Strategy to Improve the Robustness-Quality Trade-off. Our findings underscore the ability of RACE to significantly bolster the SD model's", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "R.A.C.E. for Secure T2I Diffusion", + "bbox": [ + 504, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "defense against prompt-based adversarial attacks across a variety of concepts. However, as illustrated in Tab.2, enhancing robustness appears to inversely impact image quality. To address this dichotomy, we test a refined version of the RACE loss function incorporating a regularization term:", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 13 + }, + { + "type": "equation", + "text": "\n$$\nL _ {R A C E + R e g.} := L _ {R A C E} + \\lambda \\| \\theta - \\theta^ {*} \\| _ {1}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 214, + 785, + 229 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "where $\\theta$ and $\\theta^{*}$ represent the parameters of the RACE and original Stable Diffusion models, respectively, and $\\lambda$ is the regularization strength, set to 0.1 in our experiment. This regularization approach, as evidenced in Tab.4, helps to partially reconcile the robustness-quality trade-off, enhancing image quality while maintaining improved robustness over ESD.", + "bbox": [ + 212, + 234, + 782, + 310 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Enhancing Concept Erasure. In pursuit of further reducing the ASR, we explore strategies for more comprehensive concept erasure. Recognizing that a target concept may manifest in various synonymous forms, we extend RACE's erasure scope to include semantically related concepts. Leveraging the CLIP text encoder embedded within Stable Diffusion, we identify and subsequently erase concepts closely related to the target, based on their proximity in the CLIP embedding space. For instance, alongside \"nudity\", we also target synonymous concepts like \"nude\", \"nsfw\", and \"bare\", identified as the top-3 semantically similar terms. In Fig. 3, we can observe that our method equipped with this expanded erasure strategy (denoted as RACE+keywords) is more effective in defending the malicious attempts to bypass the ESD by further reducing the number of exposed body parts. Tab. 4 also indicates that our strategy indeed fortifies the model's robustness against red teaming tactics. Nonetheless, this broadened concept removal spectrum reaffirms the robustness-quality trade-off, manifesting as a decrement in image quality. This aspect opens an intriguing avenue for future enhancements to the RACE methodology, balancing the twin objectives of robust concept erasure and preserved image fidelity.", + "bbox": [ + 212, + 325, + 787, + 583 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 601, + 359, + 617 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we present RACE, a novel defense approach designed to protect the Text-to-Image Stable Diffusion models from prompt-based red teaming attacks. RACE effectively strengthens the model's concept erasure capabilities while maintaining computational efficiency, offering a valuable enhancement to the current erasure framework and bolstering defenses against various adversarial techniques. We also observe the robustness-quality trade-off and discuss possible future directions to improve it. This initial contribution lays the groundwork for further exploration, underscoring the critical importance of developing sophisticated defenses in the rapidly evolving domain of generative AI.", + "bbox": [ + 212, + 628, + 787, + 763 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. We would like to express our gratitude to Sangmin Jung for assisting with the experiments. This work is partially supported by the National Science Foundation under Grant No. 2038666 and No. 2101052. The views and opinions of the authors expressed herein do not necessarily state or reflect those of the funding agencies and employers.", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "C. Kim*, K. Min*, Y. Yang", + "bbox": [ + 271, + 113, + 457, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bedapudi, P.: Nudenet: Neural nets for nudity classification, detection and selective censoring (12 2019) 10", + "2. Carlini, N., Hayes, J., Nasr, M., Jagielski, M., Sehwag, V., Tramér, F., Balle, B., Ippolito, D., Wallace, E.: Extracting training data from diffusion models. ArXiv abs/2301.13188 (2023), https://api-semanticscholar.org/CorpusID: 256389993 1", + "3. Chin, Z.Y., Jiang, C.M., Huang, C.C., Chen, P.Y., Chiu, W.C.: Prompting4debugging: Red-teaming text-to-image diffusion models by finding problematic prompts. arXiv preprint arXiv:2309.06135 (2023) 3, 4, 5, 7, 8, 9, 10, 11, 13", + "4. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) 10, 13", + "5. Devlin, K., Cheetham, J.: Fake trump arrest photos: How to spot an ai-generated image. BBC News (March 2023), https://www.bbc.com/news/technology-68981525 1, 4", + "6. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12873-12883 (2021) 4, 5", + "7. Fernandez, P., Couairon, G., Jégou, H., Douze, M., Furon, T.: The stable signature: Rooting watermarks in latent diffusion models. arXiv preprint arXiv:2303.15435 (2023) 2", + "8. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. ArXiv abs/2208.01618 (2022), https://apisemantic scholar.org/CorpusID:251253049 4, 8", + "9. Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., Bau, D.: Erasing concepts from diffusion models. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 2426-2436 (2023), https://apisemantic scholar.org/CorpusID:257495777 2, 3, 4, 7, 9, 10, 11, 12", + "0. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 10", + "1. Heng, A., Soh, H.: Selective amnesia: A continual learning approach to forgetting in deep generative models. ArXiv abs/2305.10120 (2023), https://apisemantic scholar.org/CorpusID:258740988 3, 4", + "2. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021) 9, 10, 11, 13", + "3. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Advances in Neural Information Processing Systems. pp. 6626-6637 (2017) 9, 10, 11, 13", + "4. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020) 7", + "5. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022) 7", + "6. Howard, J., Gugger, S.: fastai: A layered api for deep learning. Inf. 11, 108 (2020), https://api-semanticscholar.org/CorpusID:211082837 10" + ], + "bbox": [ + 225, + 179, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "R.A.C.E. for Secure T2I Diffusion", + "bbox": [ + 503, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8110-8119 (2020) 4", + "18. Kim, C., Min, K., Patel, M., Cheng, S., Yang, Y.: Wouaf: Weight modulation for user attribution and fingerprinting in text-to-image diffusion models. arXiv preprint arXiv:2306.04744 (2023) 2", + "19. Kim, C., Ren, Y., Yang, Y.: Decentralized attribution of generative models. In: International Conference on Learning Representations (2021) 2", + "20. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization (2014), http://arxiv.org/abs/1412.6980, cite arxiv:1412.6980Comment: Published as a conference paper at the 3rd International Conference for Learning Representations, San Diego, 2015 10", + "21. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating concepts in text-to-image diffusion models. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 22634-22645 (2023), https://apisemanticscholar.org/CorpusID:257687839 3, 4, 9, 10", + "22. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 1931-1941 (2022), https://api-semanticscholar.org/CorpusID:254408780 4", + "23. Li, A.C., Prabhudesai, M., Duggal, S., Brown, E., Pathak, D.: Your diffusion model is secretly a zero-shot classifier. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2206-2217 (October 2023) 7", + "24. Li, S., van de Weijer, J., Hu, T., Khan, F.S., Hou, Q., Wang, Y., Yang, J.: Get what you want, not what you don't: Image content suppression for text-to-image diffusion models. ArXiv abs/2402.05375 (2024), https://api.sementicscholar.org/CorpusID:267547985 4", + "25. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) 9, 10", + "26. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083 (2017) 3, 5, 8", + "27. Marcelo, P.: Fact focus: Fake image of pentagon explosion briefly sends jitters through stock market. Associated Press (May 2023), https://apnews.com/article/pentagon-explosion-fake-image-stock-market-jitters-78d5913603bdf9f17cc7a8c77a72e59b 1, 4", + "28. Ni, M., Wu, C., Wang, X., Yin, S.S., Wang, L., Liu, Z., Duan, N.: Ores: Open-vocabulary responsible visual synthesis. ArXiv abs/2308.13785 (2023), https://api-semanticscholar.org/CorpusID:261243073 4", + "29. Nichol, A.Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., Mcgrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In: International Conference on Machine Learning. pp. 16784-16804. PMLR (2022) 4", + "30. Nie, G., Kim, C., Yang, Y., Ren, Y.: Attributing image generative models using latent fingerprints. arXiv preprint arXiv:2304.09752 (2023) 2", + "31. OpenAI: Chatgpt. Online (2022), https://chat.openai.com/chat, accessed on February 24, 2024 10" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "C. Kim*, K. Min*, Y. Yang", + "bbox": [ + 271, + 114, + 457, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Patel, M., Jung, S., Baral, C., Yang, Y.: $\\lambda$ -eclipse: Multi-concept personalized text-to-image diffusion models by leveraging clip latent space. ArXiv abs/2402.05195 (2024), https://api_semanticscholar.org/CorpusID:267547418 4", + "33. Patel, M., Kim, C.S., Cheng, S., Baral, C., Yang, Y.: Eclipse: A resource-efficient text-to-image prior for image generations. ArXiv abs/2312.04655 (2023), https://api-semanticscholar.org/CorpusID:266149498 4", + "34. Pham, M., Marshall, K.O., Cohen, N., Mittal, G., Hegde, C.: Circumventing concept erasure methods for text-to-image generative models. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=ag3o2T51Ht 4, 8", + "35. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 6, 10", + "36. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022) 1, 4", + "37. Rando, J., Paleka, D., Lindner, D., Heim, L., Tramér, F.: Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610 (2022) 1", + "38. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684-10695 (June 2022) 1, 4, 5", + "39. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015) 5, 6", + "40. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 22500-22510 (2022), https://api_semanticscholar.org/CorpusID:251800180 4", + "41. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022) 4", + "42. Saleh, B., Elgammal, A.: Large-scale classification of fine-art paintings: Learning the right metric on the right feature. ArXiv abs/1505.00855 (2015), https://api-semanticscholar.org/CorpusID:14168099 10", + "43. Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 22522-22531 (2022), https://apisemanticscholar.org/CorpusID:253420366 4, 8, 9, 10, 13", + "44. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402 (2022) 1", + "45. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Diffusion art or digital forgery? investigating data replication in diffusion models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 6048-6058 (2022), https://apisemanticscholar.org/CorpusID:2543666341" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "R.A.C.E. for Secure T2I Diffusion", + "bbox": [ + 503, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "46. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Understanding and mitigating copying in diffusion models. ArXiv abs/2305.20086 (2023), https://api.sementicscholar.org/CorpusID:258987384 1", + "47. Tsai, Y.L., Hsu, C.Y., Xie, C., Lin, C.H., Chen, J.Y., Li, B., Chen, P.Y., Yu, C.M., ying Huang, C.: Ring-a-bell! how reliable are concept removal methods for diffusion models? ArXiv abs/2310.10012 (2023), https://api_semanticscholar.org/CorpusID:264146485 3, 4, 5, 7", + "48. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 15897-15907 (2023), https://apisemantic scholar.org/CorpusID:257219968 4", + "49. Wen, Y., Jain, N., Kirchenbauer, J., Goldblum, M., Geiping, J., Goldstein, T.: Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery. Advances in Neural Information Processing Systems 36 (2024) 3, 4, 5, 7, 10, 11, 13", + "50. Wen, Y., Kirchenbauer, J., Geiping, J., Goldstein, T.: Tree-ring watermarks: Fingerprints for diffusion images that are invisible and robust. arXiv preprint arXiv:2305.20030 (2023) 2", + "51. Wu, B., Xu, C., Dai, X., Wan, A., Zhang, P., Tomizuka, M., Keutzer, K., Vajda, P.: Visual transformers: Token-based image representation and processing for computer vision. ArXiv abs/2006.03677 (2020), https://api_semanticscholar.org/CorpusID:219531480 10", + "52. Wu, Y., Zhang, J., Kerschbaum, F., Zhang, T.: Backdooring textual inversion for concept censorship. ArXiv abs/2308.10718 (2023), https://apisemantic scholar.org/CorpusID:261049298 4", + "53. Yang, Y., Gao, R., Wang, X., Xu, N., Xu, Q.: Mma-diffusion: Multimodal attack on diffusion models. ArXiv abs/2311.17516 (2023), https://apisemanticscholar.org/CorpusID:265498727 3, 4, 5, 7", + "54. Yu, N., Skripniuk, V., Abdelnabi, S., Fritz, M.: Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In: Proceedings of the IEEE/CVF International conference on computer vision. pp. 14448-14457 (2021) 2", + "55. Yu, N., Skripniuk, V., Chen, D., Davis, L., Fritz, M.: Responsible disclosure of generative models using scalable fingerprinting. arXiv preprint arXiv:2012.08726 (2020) 2", + "56. Zhang, E., Wang, K., Xu, X., Wang, Z., Shi, H.: Forget-me-not: Learning to forget in text-to-image diffusion models. ArXiv abs/2303.17591 (2023), https://api.sementicscholar.org/CorpusID:257833863 3, 4", + "57. Zhang, K.A., Xu, L., Cuesta-Infante, A., Veeramachaneni, K.: Robust invisible video watermarking with attention (2019) 1", + "58. Zhang, Y., Jia, J., Chen, X., Chen, A., Zhang, Y., Liu, J., Ding, K., Liu, S.: To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images ... for now. ArXiv abs/2310.11868 (2023), https://api.sementicscholar.org/CorpusID:264289091 2, 3, 4, 5, 7, 8, 9, 10, 11, 13", + "59. Zheng, Y., Yeh, R.A.: Imma: Immunizing text-to-image models against malicious adaptation. ArXiv abs/2311.18815 (2023), https://api_semanticscholar.org/CorpusID:265506125 4" + ], + "bbox": [ + 215, + 146, + 785, + 797 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "C. Kim*, K. Min*, Y. Yang", + "bbox": [ + 271, + 114, + 457, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/R.A.C.E._ Robust Adversarial Concept Erasure for Secure Text-to-Image Diffusion Model/547862c1-28d0-440c-98c1-6204b0fe03dd_model.json b/2024/R.A.C.E._ Robust Adversarial Concept Erasure for Secure Text-to-Image Diffusion Model/547862c1-28d0-440c-98c1-6204b0fe03dd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3766a74ee3a7855fac330db9c750b0695991123e --- /dev/null +++ b/2024/R.A.C.E._ Robust Adversarial Concept Erasure for Secure Text-to-Image Diffusion Model/547862c1-28d0-440c-98c1-6204b0fe03dd_model.json @@ -0,0 +1,2601 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.228, + 0.141, + 0.777, + 0.187 + ], + "angle": 0, + "content": "R.A.C.E.: Robust Adversarial Concept Erasure for Secure Text-to-Image Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.303, + 0.212, + 0.7, + 0.228 + ], + "angle": 0, + "content": "Changhoon \\(\\mathbf{Kim}^{\\star 1}\\) Kyle Min\\*2 Yezhou Yang\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.411, + 0.239, + 0.593, + 0.254 + ], + "angle": 0, + "content": "1 Arizona State University" + }, + { + "type": "text", + "bbox": [ + 0.46, + 0.254, + 0.545, + 0.267 + ], + "angle": 0, + "content": "Intel Labs" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.269, + 0.663, + 0.282 + ], + "angle": 0, + "content": "{kch,yz.yang}@asu.edu kyle.min@intel.com" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.31, + 0.744, + 0.548 + ], + "angle": 0, + "content": "Abstract. In the evolving landscape of text-to-image (T2I) diffusion models, the remarkable capability to generate high-quality images from textual descriptions faces challenges with the potential misuse of reproducing sensitive content. To address this critical issue, we introduce Robust Adversarial Concept Erase (RACE), a novel approach designed to mitigate these risks by enhancing the robustness of concept erasure method for T2I models. RACE utilizes a sophisticated adversarial training framework to identify and mitigate adversarial text embeddings, significantly reducing the Attack Success Rate (ASR). Impressively, RACE achieves a \\(30\\%\\) reduction in ASR for the \"nudity\" concept against the leading white-box attack method. Our extensive evaluations demonstrate RACE's effectiveness in defending against both white-box and black-box attacks, marking a significant advancement in protecting T2I diffusion models from generating inappropriate or misleading imagery. This work underlines the essential need for proactive defense measures in adapting to the rapidly advancing field of adversarial challenges. Our code is publicly available: https://github.com/chkimmmmm/R.A.C.E." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.559, + 0.741, + 0.587 + ], + "angle": 0, + "content": "Keywords: Concept Erasure \\(\\cdot\\) Responsible Image Generative Models \\(\\cdot\\) Secure T2I Diffusion Models" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.609, + 0.376, + 0.625 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.637, + 0.788, + 0.773 + ], + "angle": 0, + "content": "The field of text-to-image (T2I) diffusion models has garnered significant attention for their ability to produce high-quality images that can be adaptively generated from textual descriptions [36, 38]. This advancement is predicated on the training of T2I models with extensive datasets, often encompassing a range of content including copyrighted, explicit, and private materials [44-46]. Consequently, these models possess the capacity to inadvertently replicate protected images, potentially without user awareness [2, 45, 46]. The misuse of T2I models by malicious actors for misinformation or public opinion manipulation presents a significant concern [5, 27]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.773, + 0.788, + 0.819 + ], + "angle": 0, + "content": "In response to the challenges posed by the malicious exploitation of generative models, Stable Diffusion (SD) [38] has integrated a safety checker [37] and advocates for the utilization of a watermarking module [57]. Despite these initiatives," + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.826, + 0.551, + 0.84 + ], + "angle": 0, + "content": "* These authors contributed equally to this work." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.114, + 0.458, + 0.129 + ], + "angle": 0, + "content": "C. Kim*, K. Min*, Y. Yang" + }, + { + "type": "image_caption", + "bbox": [ + 0.304, + 0.151, + 0.389, + 0.166 + ], + "angle": 0, + "content": "Original SD" + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.169, + 0.396, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.245, + 0.395, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.322, + 0.395, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.4, + 0.395, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.476, + 0.395, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.552, + 0.395, + 0.627 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.297, + 0.628, + 0.41, + 0.641 + ], + "angle": 0, + "content": "\\* Added by authors" + }, + { + "type": "image_caption", + "bbox": [ + 0.454, + 0.151, + 0.488, + 0.164 + ], + "angle": 0, + "content": "ESD" + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.17, + 0.518, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.247, + 0.518, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.323, + 0.518, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.401, + 0.518, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.476, + 0.518, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.553, + 0.518, + 0.627 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.151, + 0.612, + 0.165 + ], + "angle": 0, + "content": "UnlearnDiff" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.17, + 0.615, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.247, + 0.615, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.323, + 0.615, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.401, + 0.615, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.476, + 0.615, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.553, + 0.615, + 0.627 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.633, + 0.151, + 0.699, + 0.165 + ], + "angle": 0, + "content": "R.A.C.E." + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.17, + 0.714, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.247, + 0.714, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.323, + 0.714, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.401, + 0.714, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.476, + 0.714, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.553, + 0.714, + 0.627 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.661, + 0.788, + 0.732 + ], + "angle": 0, + "content": "Fig. 1: Comparative demonstration of concept erasure, red teaming, and robust erasure within T2I diffusion models. The ESD method [9] removes targeted concepts from the original SD outputs, yet these concepts can be reconstructed using Unlearn-Diff [58]. Our proposed R.A.C.E. method showcases enhanced robustness against such red teaming reconstruction efforts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "the reliance on post-hoc interventions presents limitations due to their potential for circumvention [7, 18]. Consequently, the research community is pivoting towards formulating methodologies that embed safety protocols directly within the image generation pipeline. This approach ensures that content security is an intrinsic aspect of image creation [7, 18, 19, 30, 50, 54, 55]. Although these methods are valuable for identifying the source of content after an incident, their reactive" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.504, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "R.A.C.E. for Secure T2I Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "nature highlights the necessity for proactive strategies. This includes pioneering techniques attempting early removal of sensitive content from T2I models [9, 21], thereby preempting the production of inappropriate or harmful material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.198, + 0.788, + 0.41 + ], + "angle": 0, + "content": "To address the challenge of removing undesirable content from T2I models, even when users attempt to circumvent content restrictions, recent research has focused on the development of concept erasure techniques within T2I diffusion models [9, 11, 21, 56]. These techniques primarily aim to eliminate specific concepts (e.g., \"nudity\") by altering the text embeddings associated with these concepts to neutral representations. Despite these efforts, there remains a vulnerability wherein erased concepts can be reconstructed. This is achieved by identifying text tokens that closely align with the visual embeddings of the targeted concepts, thus enabling the regeneration of prohibited content [3, 47, 49, 53, 58]. This issue is illustrated in Fig. 1, demonstrating that even with concept erasure, T2I models can be manipulated through prompt modification to regenerate the restricted content. This underscores the imperative for a more robust concept erasure methodology that can withstand such reconstruction attempts, ensuring the integrity of content generation within T2I models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.788, + 0.567 + ], + "angle": 0, + "content": "Acknowledging the imperative for enhanced concept erasure methodologies within T2I models, we pose a critical question: \"Is it feasible to develop a concept erasure approach that is resilient against reconstruction efforts?\" In pursuit of this, we introduce R.A.C.E. (Robust Adversarial Concept Erase), a novel strategy aimed at bolstering the resilience of concept erasure techniques against adversarial manipulations, as delineated in Algorithm 1. At the heart of RACE lies an adversarial training framework, leveraging insights from the effort of adversarial robustness [26]. Our method effectively identifies adversarial text embeddings capable of reconstructing erased concepts and then facilitates their integration into the T2I concept erasure workflow." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.572, + 0.788, + 0.739 + ], + "angle": 0, + "content": "A pivotal aspect of RACE is its ability to efficiently uncover adversarial text embeddings within a single time step of the diffusion process, an approach elaborated in Section 3.2. This efficiency not only streamlines the process of identifying adversarial examples but also facilitates the integration of our adversarial attack mechanism into the concept erasure workflow. To demonstrate the robustness of RACE, we have carried out an extensive array of experiments. These experiments validate the effectiveness of RACE in countering diverse red teaming strategies, with detailed results presented in Section 4. Our empirical investigations underscore the capacity of RACE to significantly enhance the robustness of T2I models against both white-box and black-box attacks across a broad spectrum of target concepts, including artistic, explicit, and object categories." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.744, + 0.599, + 0.758 + ], + "angle": 0, + "content": "We summarize the three main contributions here:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.765, + 0.796, + 0.809 + ], + "angle": 0, + "content": "- We are the first to present, to our knowledge, an adversarial training approach specifically designed to fortify concept erasure methods against prompt-based adversarial attacks without introducing additional modules." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "- Our method, RACE, implements a computationally efficient adversarial attack method that can be plugged into the concept erasing method." + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.765, + 0.796, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.458, + 0.129 + ], + "angle": 0, + "content": "C. Kim*, K. Min*, Y. Yang" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.147, + 0.797, + 0.178 + ], + "angle": 0, + "content": "- We show RACE significantly improves T2I models' robustness against prompts based on white/black box attacks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.202, + 0.396, + 0.218 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.235, + 0.785, + 0.266 + ], + "angle": 0, + "content": "For a comprehensive overview of additional related works, please refer to the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.281, + 0.788, + 0.432 + ], + "angle": 0, + "content": "Text-to-Image Synthesis. The field of generative models has seen remarkable advancements, notably extending their capabilities beyond generating photorealistic images [6, 17] to include Text-to-Image (T2I) synthesis [29, 33, 36, 38, 41]. This progress has led to the development of fine-tuning techniques that allow for the customization of T2I models to user-specific needs [8, 22, 32, 40, 48], thereby enabling the creation of highly realistic images that align closely with textual prompts. However, the potential for misuse by malicious entities, using these models for purposes such as spreading misinformation [5, 27], raises significant concerns. This underscores the urgency of devising protective measures to mitigate the risk of such exploitations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.675 + ], + "angle": 0, + "content": "Advanced Techniques in Concept Erasure for T2I Diffusion Models. Within the realm of machine unlearning, concept erasure for T2I Diffusion models has recently emerged as a critical area of research, focusing on the removal of sensitive or copyrighted concepts from T2I models. Methods to achieve this include guiding the image generation process or adjusting the model's weights to exclude these elements [9,11,21,24,28,43,56]. Notably, techniques by Gandikota et al. [9] and Kumari et al. [21] involve mapping sensitive concepts to null entities or benign equivalents by fine-tuning the weights of Stable Diffusion (SD) [38] models, effectively preventing the generation of undesirable content. Despite these advancements, red teaming methods have exposed potential loopholes, indicating that erased concepts might be regenerated through meticulously designed text prompts. Addressing this issue, our work contributes an adversarial training strategy aimed at bolstering the resilience of Stable Diffusion models against such text-prompt-based attacks [3,49,58], thereby enhancing the security and integrity of the content generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Robustness Evaluation via Red Teaming in T2I Models. While various safety measures have been proposed to shield SD models from misuse, red teaming strategies reveal vulnerabilities that still allow for circumvention. Research has demonstrated that techniques like Textual Inversion [8] can be exploited to regenerate content previously erased from SD models [34], prompting the development of countermeasures aimed at safeguarding against such inversions [52,59]. In real-world applications, T2I services such as Midjourney predominantly rely on user-provided text prompts, making them susceptible to prompt-based red teaming attacks [3, 47, 49, 53, 58]. These methods employ sophisticated prompt optimization techniques to restore images containing erased content, with their" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.504, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "R.A.C.E. for Secure T2I Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.146, + 0.731, + 0.163 + ], + "angle": 0, + "content": "Algorithm 1 Robust Adversarial Concept Erasure: RACE Algorithm" + }, + { + "type": "algorithm", + "bbox": [ + 0.217, + 0.165, + 0.789, + 0.349 + ], + "angle": 0, + "content": "Input: Diffusion Model \\(\\varPhi_{\\theta}\\) , frozen diffusion model \\(\\varPhi_{\\theta*}\\) , scheduler \\(S\\) target concept \\(c\\) training steps \\(M\\) , adversarial steps \\(N\\) perturbation limit \\(\\epsilon\\) attack step size \\(\\alpha\\) for \\(i = 0,\\dots,M\\) do Sample noise \\(n\\sim \\mathcal{N}(0,1)\\) , timestep \\(t\\sim \\mathcal{U}(1,1000)\\) Initialize \\(\\delta \\sim \\mathcal{U}(-\\epsilon ,\\epsilon)\\) Denoise \\(z_{t} = S(n,t,c)\\) for \\(j = 0,\\ldots ,N\\) do Perform targeted attack \\(\\delta = \\delta +\\alpha \\cdot sign(\\nabla_{\\delta} - L_{SD}(\\varPhi_{\\theta},z_{t},t,c,\\delta))\\) Clamp \\(\\delta\\) within \\([- \\epsilon ,\\epsilon ]\\) end for \\(\\theta = \\theta -\\nabla_{\\theta}L_{RACE}(\\varPhi_{\\theta},\\varPhi_{\\theta^{*}},z_{t},t,c,\\delta)\\) \nend for \nreturn \\(\\varPhi_{\\theta}\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.789, + 0.547 + ], + "angle": 0, + "content": "efficacy contingent upon the level of model access—categorized into white-box approaches, which utilize SD's U-Net [39] for prompt optimization [3, 58], and black-box strategies, where such access is restricted [47, 49, 53]. Both approaches establish formidable benchmarks in attack success rates, as detailed in Tab. 2. However, the landscape lacks robust defense mechanisms against prompt-based red teaming, primarily due to the prohibitive computational demands associated with identifying adversarial prompts—a challenge that renders traditional adversarial training approaches impractical. Addressing this gap, our work introduces a novel defense strategy tailored to counteract prompt-based red teaming attacks, marking a significant step forward in fortifying T2I diffusion models against adversarial threats." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.572, + 0.331, + 0.588 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.606, + 0.789, + 0.699 + ], + "angle": 0, + "content": "Our methodology aims to expunge target concepts from T2I diffusion models through an adversarial training framework [26]. Initially, in Section 3.1, we establish the foundation by formally introducing the notations and the rationale underpinning our approach. Following this, Section 3.2 details our proposed adversarial attack, specifically designed for robust concept erasure, and delineates its integration into the adversarial training regime." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.722, + 0.373, + 0.736 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Stable Diffusion Models. Our method is built upon the Stable Diffusion Model (SD) [38], which operates as the foundational architecture for our concept erasure technique. The SD model is composed of two primary elements: firstly, an image autoencoder that has been pre-trained on a diverse and extensive image dataset [6]. Within this autoencoder, an encoder function \\(\\mathcal{E}(\\cdot)\\) transforms an input image \\(x\\) into a latent representation \\(z = \\mathcal{E}(x)\\). Conversely, a decoder" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.458, + 0.129 + ], + "angle": 0, + "content": "C. Kim*, K. Min*, Y. Yang" + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.145, + 0.734, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.313, + 0.73, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.481, + 0.789, + 0.552 + ], + "angle": 0, + "content": "Fig. 2: Single-Timestep Adversarial Attack Efficacy. This figure illustrates the Attack Success Rate (ASR) across various timesteps, alongside representative images. Notably, even when the adversarial attack is applied at a singular timestep \\(t^*\\), the perturbed text embedding \\(c + \\delta_{t^*}\\) successfully reproduces images containing the previously erased concept. For method details, see Sec. 3.2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.562, + 0.789, + 0.592 + ], + "angle": 0, + "content": "function \\(\\mathcal{D}(\\cdot)\\) aims to reconstruct the input image from its latent form, where \\(\\mathcal{D}(z) = \\hat{x}\\approx x\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.789, + 0.685 + ], + "angle": 0, + "content": "The second element is a U-Net [39]-based diffusion model trained to craft latent representations within the acquired latent space. This model facilitates the conditioning on either class labels or text embeddings derived from training data. Let us denote by \\( c = \\mathcal{E}_{txt}(y) \\) the textual embedding encoded from a conditioning text prompt \\( y \\), where \\( \\mathcal{E}_{txt} \\) symbolizes the text encoder, such as CLIP [35]. Under these constructs, the SD training objective is encapsulated by the loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.696, + 0.787, + 0.715 + ], + "angle": 0, + "content": "\\[\nL _ {S D} := \\mathbb {E} _ {n \\sim \\mathcal {N} (0, 1), z, c, t} \\left[ | | n - \\Phi_ {\\theta} \\left(z _ {t}, t, c\\right) | | _ {2} ^ {2} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.827 + ], + "angle": 0, + "content": "where \\(t\\) indexes the time step, \\(n\\) represents a noise sample drawn from a standard Gaussian distribution, \\(z_{t}\\) is the perturbed version of \\(z\\) up to time step \\(t\\), and \\(\\varPhi_{\\theta}\\) is the denoising network based on a U-Net architecture. During inference, a random noise sample is procured from a Gaussian distribution and denoised using \\(\\varPhi_{\\theta}\\) following a scheduler \\(\\mathcal{S}\\), operating over a sequence of predetermined time steps \\(T\\). The resulting denoised latent, \\(z_{0}\\), is then decoded to produce the final image, \\(\\hat{x} = \\mathcal{D}(z_0)\\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.504, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "R.A.C.E. for Secure T2I Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "Erase and Reconstruction of Target Concept. The objective of concept erasure is to remove specific concepts, such as \"nudity\", from the latent space of a pre-trained diffusion model. The Erased Stable Diffusion (ESD) approach [9] introduces a method for excising these concepts from the latent representations within the Stable Diffusion framework. The erasure loss is formalized as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.266, + 0.239, + 0.786, + 0.257 + ], + "angle": 0, + "content": "\\[\nL _ {e r a s e} := \\left\\| \\Phi_ {\\theta} \\left(z _ {t}, t, c\\right) - \\left(\\Phi_ {\\theta^ {*}} \\left(z _ {t}, t\\right) - \\eta \\left(\\Phi_ {\\theta^ {*}} \\left(z _ {t}, t, c\\right) - \\Phi_ {\\theta^ {*}} \\left(z _ {t}, t\\right)\\right)\\right) \\right\\| _ {2} ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.264, + 0.784, + 0.339 + ], + "angle": 0, + "content": "where \\(\\varPhi_{\\theta^{*}}\\) represents the frozen denoising U-Net, and \\(\\eta\\) denotes the guidance scale associated with classifier-free guidance [15]. ESD fine-tunes \\(L_{erase}\\) with respect to \\(\\theta\\), guiding \\(\\varPhi_{\\theta}\\) to produce outputs where the target concept is effectively nullified. Crucially, this process does not necessitate additional datasets; it operates successfully with only a concise textual description." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.788, + 0.446 + ], + "angle": 0, + "content": "Conversely, red teaming efforts aim to counteract concept erasure by crafting adversarial text prompts \\(\\tilde{y}\\) capable of resurrecting the erased concept in the generated image. White-box methods [3,58] engage in this adversarial prompt optimization by leveraging the gradients of \\(\\varPhi_{\\theta}\\). Meanwhile, black-box approaches [47, 49, 53] aim to achieve comparable outcomes without reliance on the gradients of \\(\\varPhi_{\\theta}\\). Both approaches pose significant computational demands, which presents challenges for their integration into an adversarial training framework." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.471, + 0.74, + 0.487 + ], + "angle": 0, + "content": "3.2 Adversarial Training on Concept Erased Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.499, + 0.788, + 0.592 + ], + "angle": 0, + "content": "Motivation. The primary aim of T2I diffusion models is to generate high-quality images conditioned on specific prompts. Intriguingly, the SD model's loss function, as depicted in Eq.(1), can also facilitate image classification tasks [23]. This classification capability is derived by applying Bayes' Theorem to the model's predictions \\( p_{\\theta}(x|c_i) \\) and the prior distribution \\( p(c) \\) across a set of conditions \\( c_i \\), where each \\( c_i = \\mathcal{E}_{txt}(y_i) \\) represents a textual embedding of the prompt \\( y_i \\):" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.605, + 0.786, + 0.639 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} (c _ {i} | x) = \\frac {p (c _ {i}) p _ {\\theta} (x | c _ {i})}{\\sum_ {j} p (c _ {j}) p _ {\\theta} (x | c _ {j})}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.646, + 0.788, + 0.751 + ], + "angle": 0, + "content": "Notably, the prior terms \\( p(c) \\) can be disregarded when they are uniformly distributed over the prompts \\( c_i \\) (i.e., \\( p(c_i) = \\frac{1}{N} \\)). In the context of diffusion models, directly computing \\( p_{\\theta}(x|c_i) \\) is computationally challenging, leading to the reliance on the computation of \\( \\log p_{\\theta}(x|c_i) \\) and the utilization of the Evidence Lower Bound (ELBO) for optimization purposes. Leveraging approximations introduced in [14], we can approximate the posterior distribution over prompts \\( c_i \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.326, + 0.762, + 0.786, + 0.799 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(c _ {i} \\mid x\\right) = \\frac {\\exp \\left\\{- \\mathbb {E} _ {z , n , t} \\left[ | | n - \\Phi_ {\\theta} \\left(z _ {t} , t , c _ {i}\\right) | | ^ {2} \\right] \\right\\}}{\\sum_ {j} \\exp \\left\\{- \\mathbb {E} _ {z , n , t} \\left[ | | n - \\Phi_ {\\theta} \\left(z _ {t} , t , c _ {j}\\right) | | ^ {2} \\right] \\right\\}}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Building on this foundation, the Diffusion Classifier [23] proposes a method for estimating the class of a given image \\( x \\) by finding argmin of the following" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.114, + 0.458, + 0.129 + ], + "angle": 0, + "content": "C. Kim*, K. Min*, Y. Yang" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.148, + 0.3, + 0.162 + ], + "angle": 0, + "content": "expression:" + }, + { + "type": "equation", + "bbox": [ + 0.388, + 0.162, + 0.786, + 0.202 + ], + "angle": 0, + "content": "\\[\n\\underset {c} {\\operatorname {a r g m i n}} \\sum_ {t = 1} ^ {T} \\left[ \\left| \\left| n - \\Phi_ {\\theta} \\left(z _ {t}, t, c\\right) \\right| \\right| ^ {2} \\right], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.206, + 0.788, + 0.268 + ], + "angle": 0, + "content": "where \\( z_{t} = \\mathcal{E}(x) \\) represents the latent encoding of the image \\( x \\) and the goal is to identify the label \\( c \\) from the available set of classes \\( c_{i} \\). A notable insight from the Diffusion Classifier is the feasibility of classifying image \\( x \\) even when computations are performed using a single time step \\( t^{*} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.282, + 0.788, + 0.387 + ], + "angle": 0, + "content": "Single-Timestep Adversarial Attacks on Erased T2I Models. Leveraging insights from the Diffusion Classifier, we re-conceptualize the SD model's loss function in Eq.(1) as a classification mechanism. This perspective allows us to view Textual Inversion (TI) [8] as a form of targeted adversarial attack, where the objective is to optimize the conditioning text embedding \\( c \\) to regenerate the image \\( x \\) [34, 58]. Notably, TI is computationally intensive as it necessitates optimization across all time steps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.788, + 0.525 + ], + "angle": 0, + "content": "Prompted by these considerations, our investigation centers around a critical inquiry: Can adversarial text embeddings be identified with just a single timestep? We investigate whether adversarial text embeddings can be effectively identified at a singular timestep \\( t^* \\). Our approach is geared towards nullifying the embedding of a target concept \\( c \\) and its proximate embeddings that might facilitate the regeneration of an erased concept image \\( \\tilde{x} \\), such as an explicit image. To this end, we devise a targeted adversarial attack to produce \\( \\tilde{x} \\) from a concept-erased diffusion model \\( \\varPhi_{\\theta} \\) by introducing an adversarial perturbation \\( \\delta \\). The perturbation \\( \\delta \\) is determined through the optimization:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.533, + 0.786, + 0.559 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {a r g m i n} \\left| | n - \\Phi_ {\\theta} \\left(\\tilde {z} _ {t}, t, c + \\delta\\right) | | _ {2} ^ {2}, \\right. \\tag {6} \\\\ | | \\delta | | _ {\\infty} \\leq \\epsilon \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.704 + ], + "angle": 0, + "content": "where \\(\\tilde{z} = \\mathcal{E}(\\tilde{x})\\) denotes the latent representation of image \\(\\tilde{x}\\), \\(\\epsilon\\) is a small number, and \\(c = \\mathcal{E}_{txt}(y)\\) encodes the textual embedding of the targeted concept, for instance, \\(y =\\) \"nudity\". The Projected Gradient Descent (PGD) algorithm [26] is employed to address this optimization problem. Specifically, when selecting timestep \\(t^* = 500\\) as the critical adversarial point, \\(\\tilde{z}_t\\) undergoes denoising via \\(\\Phi_{\\theta}(\\tilde{z}_t,t,c)\\) transitioning from timesteps \\(t = 1000\\) to \\(t = 500\\). Subsequently, the targeted adversarial approach outlined in Eq. (6) is executed to determine \\(\\delta_{t^*}\\) at \\(t = 500\\). In subsequent denoising steps, \\(\\tilde{z}_t\\) is denoised considering the introduced perturbation \\(c + \\delta_{t^*}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.764 + ], + "angle": 0, + "content": "Our method introduces a distinctive single-timestep adversarial attack, contrasting with prior approaches that required optimization over a wide range of timesteps [3,58]. This approach enables the seamless incorporation of our attack strategy into the adversarial training process specifically for T2I concept erasure." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To assess the efficacy of our method, we execute tests on a \\(\\varPhi_{\\theta}\\) model trained for removing the \"nudity\" concept via the ESD method. Utilizing 142 nuditycentric prompts from the I2P dataset [43], we systematically select timesteps \\(t^{*}\\) at 100-step intervals within the [1, 1000] range and compute the Attack Success Rate (ASR), as formulated in Eq.(8), for each \\(t^{*}\\). The results, depicted in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.504, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "R.A.C.E. for Secure T2I Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "Fig. 2, unveil a notable capability: the reconstruction of the erased concept is feasible even with the attack confined to a single timestep \\( t^* \\). Interestingly, after \\( t^* = 500 \\), we observe an increasing trend in the ASR. The modified images resulting from the attack visually convey the transition back to the erased concept and reflect the ASR patterns. For additional insights into various concepts and their corresponding ASR trends across different timesteps \\( t^* \\), the supplementary material offers further information." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.269, + 0.544, + 0.284 + ], + "angle": 0, + "content": "Adversarial Training for T2I Concept" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.285, + 0.544, + 0.388 + ], + "angle": 0, + "content": "Erasure. Motivated by the findings from our single-timestep adversarial attack experiments, we explore the potential of such attacks to enhance the robustness of concept erasure in T2I models, posing the question: Can adversarial attacks improve the resilience of concept erasure mechanisms?" + }, + { + "type": "table_caption", + "bbox": [ + 0.555, + 0.285, + 0.786, + 0.313 + ], + "angle": 0, + "content": "Table 1: Performance comparison of \\(L_{\\text{erase}}\\) and \\(L_{\\text{RACE}}\\)" + }, + { + "type": "table", + "bbox": [ + 0.556, + 0.314, + 0.787, + 0.382 + ], + "angle": 0, + "content": "
LeraseLRACE
I2P [43]0.080.05
FID [13]33.1225.16
CLIP-Score [12]0.7260.745
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.788, + 0.48 + ], + "angle": 0, + "content": "RACE distinguishes itself from existing approaches [9,21] by aiming to eliminate not only the targeted concept's embedding but also its adjacent embedding within the model's latent space, which could otherwise lead to the inadvertent generation of the erased concept by \\(\\varPhi_{\\theta}\\). We incorporate our adversarial attack into the erasure loss function (\\(L_{erase}\\)), yielding an enhanced adversarial training loss:" + }, + { + "type": "equation", + "bbox": [ + 0.235, + 0.494, + 0.787, + 0.512 + ], + "angle": 0, + "content": "\\[\nL _ {R A C E} := \\left| \\left| \\Phi_ {\\theta} \\left(z _ {t}, t, c + \\delta\\right) - \\left(\\Phi_ {\\theta^ {*}} \\left(z _ {t}, t\\right) - \\eta \\left(\\Phi_ {\\theta^ {*}} \\left(z _ {t}, t, c\\right) - \\Phi_ {\\theta^ {*}} \\left(z _ {t}, t\\right)\\right)\\right) \\right| \\right| _ {2} ^ {2}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.788, + 0.675 + ], + "angle": 0, + "content": "This method is deliberate, substituting the concept embedding \\( c \\) with \\( c + \\delta \\) within the trainable parameters of \\( \\varPhi_{\\theta} \\). This precise adjustment ensures enhanced fidelity of the generated images by mapping the \\( \\epsilon \\)-neighborhood of the concept embedding to its null representation. Comparative metrics between the direct substitution in \\( L_{erase} \\) and the strategic use of \\( L_{RACE} \\) are provided in Tab. 1. The latter approach demonstrates promising reductions in ASR for \"nudity\" prompts within I2P dataset [43] and improvements in image quality metrics, as assessed on the MS-COCO [25]. The RACE methodology is comprehensively detailed in Algorithm 1. To validate RACE's effectiveness, we evaluate the ASR against both white-box and black-box attacks, as elaborated in Sec. 4." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.702, + 0.377, + 0.719 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.736, + 0.441, + 0.753 + ], + "angle": 0, + "content": "4.1 Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Datasets. Our assessment of the RACE framework spans various domains, including artistic styles, explicit concepts, and identifiable objects, in line with established benchmarks [3,9,58]. To ensure a uniform image generation process, we standardize key hyperparameters such as the scale of classifier-free guidance and random seeds. Artistic style evaluations leverage shared text prompts from" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.458, + 0.129 + ], + "angle": 0, + "content": "C. Kim*, K. Min*, Y. Yang" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "ESD [9] and UnlearnDiff [58]. For explicit content, we draw from the inappropriate image prompt benchmark [43], selecting a diverse set of prompts encompassing 142 nudity, 98 illegal acts, and 101 violence instances. Objects are curated from a subset of Imagenette [16], known for its distinct and recognizable classes, with text prompts synthesized via ChatGPT [31] following the approach by Kumari et al. [21]. It's crucial to note that these datasets serve solely to gauge the effectiveness of adversarial attacks; the adversarial training component of RACE does not necessitate the use of these specific prompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.39 + ], + "angle": 0, + "content": "Training Details. In validating RACE, we choose the Erased Stable Diffusion (ESD) model [9] for its ability to erase a wide range of concepts, serving as an ideal testbed to showcase RACE's efficacy. We integrate RACE with ESD, optimizing with attack parameters \\(\\epsilon = 0.1\\) and \\(\\alpha = \\epsilon /4\\). The optimization of model parameters \\(\\theta\\) utilizes the Adam optimizer [20] at a learning rate of \\(1e - 5\\), consistent with ESD. Additional details on attack parameter selection are available in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.798, + 0.51 + ], + "angle": 0, + "content": "Red Teaming Methods. To rigorously test RACE's robustness, we deploy a comprehensive suite of adversarial attacks, spanning both white-box and black-box approaches. Initial assessments utilize the I2P red teaming prompt dataset [43]. In black-box scenarios, we employ PEZ [49], which crafts adversarial prompts via CLIP [35]. In the white-box scenario, methods like P4D [3] and UnlearnDiff [58] are used, which generate adversarial prompts by leveraging gradients from the SD." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.526, + 0.788, + 0.603 + ], + "angle": 0, + "content": "Evaluation. To gauge the robustness of RACE, we employ domain-specific classifiers: a ViT-base model [51] pre-trained on ImageNet [4] and fine-tuned on WikiArt [42] for artistic styles, Nudenet [1] for explicit content, and ResNet-50 [10] trained on ImageNet for object removal. We measure robustness using the Attack Success Rate (ASR):" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.614, + 0.787, + 0.656 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A S R} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathbb {1} \\left(f \\left(S D \\left(\\tilde {y} _ {i}\\right)\\right) = \\tilde {y} _ {i}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.667, + 0.788, + 0.744 + ], + "angle": 0, + "content": "where \\(\\tilde{y}_i\\) is the adversarial prompt, \\(f\\) is the classifier, and \\(N\\) is the number of prompts. Additionally, we assess image quality after applying RACE by generating 5,000 images from the MS-COCO [25] test set, computing the Frechet Inception Distance (FID) score [13] and the CLIP score [12] to evaluate RACE's impact on image fidelity while ensuring concept erasure." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.767, + 0.629, + 0.784 + ], + "angle": 0, + "content": "4.2 Robust Concept Erase against Red Teaming" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "In our comprehensive analysis, RACE undergoes a series of red teaming evaluations [3, 43, 49, 58], encompassing both white-box and black-box techniques aimed at regenerating concepts targeted by RACE for erasure, as depicted in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.504, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "R.A.C.E. for Secure T2I Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.243 + ], + "angle": 0, + "content": "Table 2: Attack Success Rate (ASR) against white \\((\\bigcirc)\\)/black \\((\\bullet)\\) box attacks. We conduct experiments on artistic style (Van Gogh), explicit concepts (nudity, violence, and illegal acts), and objects (church, golf ball, and parachute). \"ESD/RACE-Concept\" denotes the concept erased from the model. We also measure the quality of T2I models. We can observe that RACE reduces ASR by over \\(30\\%\\) (-0.30 or below) in absolute value on Van Gogh, nudity, and church for UnlearnDiff [58], which is the previous SOTA attack method." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.255, + 0.787, + 0.45 + ], + "angle": 0, + "content": "
PromptsPEZ [49]P4D [3]UnlearnDiff [58]CLIP-Score [12]FID [13]
White/Black Box●●○○--
ESD [9]-VanGogh0.040.000.260.360.799719.16
ESD [9]-Nudity0.140.080.750.800.793118.88
ESD [9]-Violence0.270.130.840.790.783421.55
ESD [9]-Illegal0.290.200.890.850.785421.50
ESD [9]-Church0.160.000.580.680.789619.68
ESD [9]-GolfBall0.040.000.160.160.773820.64
ESD [9]-Parachute0.060.040.480.600.786519.72
RACE-VanGogh0.00 (-0.04)0.00 (-0.00)0.00 (-0.26)0.04 (-0.32)0.802420.65
RACE-Nudity0.05 (-0.09)0.02 (-0.06)0.49 (-0.26)0.47 (-0.33)0.745225.16
RACE-Violence0.11 (-0.16)0.08 (-0.05)0.75 (-0.09)0.68 (-0.11)0.737428.71
RACE-Illlegal0.20 (-0.09)0.13 (-0.07)0.85 (-0.04)0.80 (-0.05)0.759124.87
RACE-Church0.02 (-0.14)0.00 (-0.00)0.26 (-0.32)0.38 (-0.30)0.773023.92
RACE-GolfBall0.00 (-0.04)0.00 (-0.00)0.10 (-0.06)0.06 (-0.10)0.748025.38
RACE-Parachute0.02 (-0.04)0.00 (-0.04)0.24 (-0.24)0.38 (-0.22)0.757026.42
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.644 + ], + "angle": 0, + "content": "Fig.1. The comparative analysis of ASR presented in Tab.2 spans diverse conceptual domains, from \"Van Gogh\"-inspired artistry to explicit content such as \"nudity\" and \"violence\", extending to tangible objects like \"churches\", \"golf balls\", and \"parachutes\". Remarkably, RACE consistently diminishes ASR, notably surpassing \\(30\\%\\) for \"Van Gogh\" styles, \"nudity\", and \"church\" categories, particularly outperforming UnlearnDiff [58], the current state-of-the-art in white-box adversarial methodologies. This marked decline in ASR underscores RACE's heightened robustness and delineates its capability as a potent, computationally efficient defense mechanism for T2I diffusion models against intricate adversarial attacks. Crucially, RACE's methodological advantage stems from its independence from external imagery or prompts, diverging from traditional red teaming techniques reliant on such data for prompt generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.809 + ], + "angle": 0, + "content": "As demonstrated in Tab.2, RACE achieves a significant \\(33\\%\\) reduction in ASR for the \"nudity\" concept, underscoring its effectiveness. To further elucidate how RACE enhances ASR, we analyze the specific categories or elements it targets for removal. Utilizing Nudenet, we enumerate the body parts generated by various models—original SD, ESD, UnlearnDiff, and RACE—when prompted with nudity-related inputs from the I2P dataset. Illustrated in Fig.3, the analysis reveals that while ESD enhances the original SD's resilience to such prompts, UnlearnDiff manages to bypass ESD's defenses, reconstructing explicit content. In contrast, RACE maintains its robustness even against the sophisticated UnlearnDiff attacks, showcasing the advanced protective capabilities of our approach in safeguarding against the regeneration of sensitive content." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "One caveat to mention is that our experiments reveal a nuanced trade-off between robustness and image quality. While artistic style erasures maintain" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.458, + 0.129 + ], + "angle": 0, + "content": "C. Kim*, K. Min*, Y. Yang" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.146, + 0.784, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.332, + 0.788, + 0.404 + ], + "angle": 0, + "content": "Fig. 3: Although ESD significantly reduces the chance of generating images with exposed body parts, state-of-the-art red teaming methods, such as UnlearnDiff, can be used to bypass ESD's defense and reconstruct explicit content. RACE and its variant can effectively defend the malicious attempts to reconstruct explicit content from the ESD model that erased the concept of nudity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.789, + 0.642 + ], + "angle": 0, + "content": "quality metrics, erasures of other concepts inadvertently degrade image quality. This divergence could be attributed to methodological differences; erasing artistic styles predominantly involves fine-tuning the cross-attention layers of SD as per ESD guidelines, whereas erasing other concepts necessitates adjustments in non-cross-attention layers [9]. Another plausible explanation for the observed trade-off between robust concept erasure and overall image quality could relate to the inherent complexity of differentiating between closely related or overlapping concepts within the model's latent space. As RACE intensifies adversarial robustness, it may inadvertently alter the delicate equilibrium within these conceptual overlaps, leading to unintended modifications in adjacent, non-targeted conceptual representations. This issue highlights the complex tension between precise concept erasure and maintaining the model's overall integrity against attacks. To address the quality concerns arising from this trade-off, we explore a potential strategy for improvement in Sec. 4.4, aiming to reduce the trade-off between targeted erasure with high quality and the model's defensive robustness." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.662, + 0.402, + 0.677 + ], + "angle": 0, + "content": "4.3 Disentanglement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.789, + 0.839 + ], + "angle": 0, + "content": "Investigating RACE's disentanglement performance, our study focuses on its capability to precisely erase intended concepts without impacting other elements. The evaluation spans both qualitative and quantitative measures. On the qualitative front, we configure separate RACE-enhanced models to specifically erase artistic imprints such as \"Van Gogh\", \"Thomas Kinkade\", and \"Kilian Eng\" from the SD. The illustrative outcomes, showcased in Fig. 4, underscore RACE's erasure precision, ensuring that the excision of one artistic style doesn't lead to the collateral removal of others. This meticulous erasure extends to discrete object concepts like \"church\", \"golf ball\", and \"parachute\", with the generated images further affirming the method's discernment." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.505, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "R.A.C.E. for Secure T2I Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.146, + 0.788, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.317, + 0.785, + 0.375 + ], + "angle": 0, + "content": "Fig. 4: RACE's Disentanglement in Concept Erasure. This figure highlights RACE's precision in erasing specific concepts, as shown in diagonal images, while preserving unrelated concepts, which is evident in off-diagonal images. For reference, baseline images generated by the original Stable Diffusion (SD) model are also presented." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.384, + 0.784, + 0.427 + ], + "angle": 0, + "content": "Table 4: Ablation Studies for Performance Improvement. Using the \"nudity\" concept, we evaluate the effectiveness of adding weight regularization and close concept keywords. ASR and quality metrics are measured to assess performance." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.515 + ], + "angle": 0, + "content": "
I2P [43]PEZ [49]P4D [3]UnleranDiff [58]CLIP-Score [12]FID [13]
ESD0.140.080.750.800.793118.88
RACE0.050.020.490.470.745225.16
RACE+Reg.0.070.020.600.620.759324.42
RACE+keywords0.020.010.420.460.720130.97
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.541, + 0.511, + 0.706 + ], + "angle": 0, + "content": "Quantitatively, we generate 5,000 prompts with varied random seeds, such as \"an image of a [class name]\", to produce a diverse set of images, subsequently evaluated using a pre-trained ResNet-50 classifier for top-1 accuracy. As shown in Tab. 3, RACE showcases an improved capability for object concept erasure compared to ESD. Despite this, there exists a slight decrement in classification accuracy for non-target classes. This effect" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.707, + 0.784, + 0.768 + ], + "angle": 0, + "content": "likely stems from RACE's method of targeting the \\(\\epsilon\\)-neighborhood surrounding the intended concept, potentially influencing proximate concepts. Nevertheless, we can observe from Fig. 4 that RACE can precisely erase the target concept with minimal visual impacts on other concepts." + }, + { + "type": "table_caption", + "bbox": [ + 0.52, + 0.54, + 0.788, + 0.609 + ], + "angle": 0, + "content": "Table 3: Accuracy of erased and non-erased classes in Imagenet [4]. We evaluate the classification accuracy of ESD and RACE for the target erased concepts and other non-target concepts." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.61, + 0.787, + 0.696 + ], + "angle": 0, + "content": "
Erased ConceptAcc. ErasedAcc. Others
ESDRACEESDRACE
Church0.160.020.570.53
Golf Ball0.040.000.450.35
Parachute0.060.020.570.41
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.787, + 0.35, + 0.802 + ], + "angle": 0, + "content": "4.4 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Potential Strategy to Improve the Robustness-Quality Trade-off. Our findings underscore the ability of RACE to significantly bolster the SD model's" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.114, + 0.458, + 0.129 + ], + "angle": 0, + "content": "C. Kim*, K. Min*, Y. Yang" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "defense against prompt-based adversarial attacks across a variety of concepts. However, as illustrated in Tab.2, enhancing robustness appears to inversely impact image quality. To address this dichotomy, we test a refined version of the RACE loss function incorporating a regularization term:" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.215, + 0.786, + 0.231 + ], + "angle": 0, + "content": "\\[\nL _ {R A C E + R e g.} := L _ {R A C E} + \\lambda \\| \\theta - \\theta^ {*} \\| _ {1}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.235, + 0.784, + 0.311 + ], + "angle": 0, + "content": "where \\(\\theta\\) and \\(\\theta^{*}\\) represent the parameters of the RACE and original Stable Diffusion models, respectively, and \\(\\lambda\\) is the regularization strength, set to 0.1 in our experiment. This regularization approach, as evidenced in Tab.4, helps to partially reconcile the robustness-quality trade-off, enhancing image quality while maintaining improved robustness over ESD." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.326, + 0.788, + 0.584 + ], + "angle": 0, + "content": "Enhancing Concept Erasure. In pursuit of further reducing the ASR, we explore strategies for more comprehensive concept erasure. Recognizing that a target concept may manifest in various synonymous forms, we extend RACE's erasure scope to include semantically related concepts. Leveraging the CLIP text encoder embedded within Stable Diffusion, we identify and subsequently erase concepts closely related to the target, based on their proximity in the CLIP embedding space. For instance, alongside \"nudity\", we also target synonymous concepts like \"nude\", \"nsfw\", and \"bare\", identified as the top-3 semantically similar terms. In Fig. 3, we can observe that our method equipped with this expanded erasure strategy (denoted as RACE+keywords) is more effective in defending the malicious attempts to bypass the ESD by further reducing the number of exposed body parts. Tab. 4 also indicates that our strategy indeed fortifies the model's robustness against red teaming tactics. Nonetheless, this broadened concept removal spectrum reaffirms the robustness-quality trade-off, manifesting as a decrement in image quality. This aspect opens an intriguing avenue for future enhancements to the RACE methodology, balancing the twin objectives of robust concept erasure and preserved image fidelity." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.602, + 0.36, + 0.618 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.764 + ], + "angle": 0, + "content": "In this work, we present RACE, a novel defense approach designed to protect the Text-to-Image Stable Diffusion models from prompt-based red teaming attacks. RACE effectively strengthens the model's concept erasure capabilities while maintaining computational efficiency, offering a valuable enhancement to the current erasure framework and bolstering defenses against various adversarial techniques. We also observe the robustness-quality trade-off and discuss possible future directions to improve it. This initial contribution lays the groundwork for further exploration, underscoring the critical importance of developing sophisticated defenses in the rapidly evolving domain of generative AI." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Acknowledgements. We would like to express our gratitude to Sangmin Jung for assisting with the experiments. This work is partially supported by the National Science Foundation under Grant No. 2038666 and No. 2101052. The views and opinions of the authors expressed herein do not necessarily state or reflect those of the funding agencies and employers." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.504, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "R.A.C.E. for Secure T2I Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.18, + 0.785, + 0.208 + ], + "angle": 0, + "content": "1. Bedapudi, P.: Nudenet: Neural nets for nudity classification, detection and selective censoring (12 2019) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.209, + 0.785, + 0.263 + ], + "angle": 0, + "content": "2. Carlini, N., Hayes, J., Nasr, M., Jagielski, M., Sehwag, V., Tramér, F., Balle, B., Ippolito, D., Wallace, E.: Extracting training data from diffusion models. ArXiv abs/2301.13188 (2023), https://api-semanticscholar.org/CorpusID: 256389993 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.264, + 0.785, + 0.307 + ], + "angle": 0, + "content": "3. Chin, Z.Y., Jiang, C.M., Huang, C.C., Chen, P.Y., Chiu, W.C.: Prompting4debugging: Red-teaming text-to-image diffusion models by finding problematic prompts. arXiv preprint arXiv:2309.06135 (2023) 3, 4, 5, 7, 8, 9, 10, 11, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.307, + 0.785, + 0.349 + ], + "angle": 0, + "content": "4. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) 10, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.349, + 0.785, + 0.39 + ], + "angle": 0, + "content": "5. Devlin, K., Cheetham, J.: Fake trump arrest photos: How to spot an ai-generated image. BBC News (March 2023), https://www.bbc.com/news/technology-68981525 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.391, + 0.785, + 0.433 + ], + "angle": 0, + "content": "6. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12873-12883 (2021) 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.785, + 0.475 + ], + "angle": 0, + "content": "7. Fernandez, P., Couairon, G., Jégou, H., Douze, M., Furon, T.: The stable signature: Rooting watermarks in latent diffusion models. arXiv preprint arXiv:2303.15435 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.476, + 0.785, + 0.531 + ], + "angle": 0, + "content": "8. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. ArXiv abs/2208.01618 (2022), https://apisemantic scholar.org/CorpusID:251253049 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.531, + 0.785, + 0.587 + ], + "angle": 0, + "content": "9. Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., Bau, D.: Erasing concepts from diffusion models. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 2426-2436 (2023), https://apisemantic scholar.org/CorpusID:257495777 2, 3, 4, 7, 9, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.588, + 0.785, + 0.629 + ], + "angle": 0, + "content": "0. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.63, + 0.785, + 0.671 + ], + "angle": 0, + "content": "1. Heng, A., Soh, H.: Selective amnesia: A continual learning approach to forgetting in deep generative models. ArXiv abs/2305.10120 (2023), https://apisemantic scholar.org/CorpusID:258740988 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.672, + 0.785, + 0.713 + ], + "angle": 0, + "content": "2. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021) 9, 10, 11, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.714, + 0.785, + 0.756 + ], + "angle": 0, + "content": "3. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Advances in Neural Information Processing Systems. pp. 6626-6637 (2017) 9, 10, 11, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.785, + 0.784 + ], + "angle": 0, + "content": "4. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.812 + ], + "angle": 0, + "content": "5. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "6. Howard, J., Gugger, S.: fastai: A layered api for deep learning. Inf. 11, 108 (2020), https://api-semanticscholar.org/CorpusID:211082837 10" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.18, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.458, + 0.129 + ], + "angle": 0, + "content": "C. Kim*, K. Min*, Y. Yang" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "17. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8110-8119 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.191, + 0.787, + 0.232 + ], + "angle": 0, + "content": "18. Kim, C., Min, K., Patel, M., Cheng, S., Yang, Y.: Wouaf: Weight modulation for user attribution and fingerprinting in text-to-image diffusion models. arXiv preprint arXiv:2306.04744 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.786, + 0.261 + ], + "angle": 0, + "content": "19. Kim, C., Ren, Y., Yang, Y.: Decentralized attribution of generative models. In: International Conference on Learning Representations (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.262, + 0.787, + 0.317 + ], + "angle": 0, + "content": "20. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization (2014), http://arxiv.org/abs/1412.6980, cite arxiv:1412.6980Comment: Published as a conference paper at the 3rd International Conference for Learning Representations, San Diego, 2015 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.787, + 0.374 + ], + "angle": 0, + "content": "21. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating concepts in text-to-image diffusion models. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 22634-22645 (2023), https://apisemanticscholar.org/CorpusID:257687839 3, 4, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.787, + 0.431 + ], + "angle": 0, + "content": "22. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 1931-1941 (2022), https://api-semanticscholar.org/CorpusID:254408780 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.431, + 0.787, + 0.472 + ], + "angle": 0, + "content": "23. Li, A.C., Prabhudesai, M., Duggal, S., Brown, E., Pathak, D.: Your diffusion model is secretly a zero-shot classifier. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2206-2217 (October 2023) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.528 + ], + "angle": 0, + "content": "24. Li, S., van de Weijer, J., Hu, T., Khan, F.S., Hou, Q., Wang, Y., Yang, J.: Get what you want, not what you don't: Image content suppression for text-to-image diffusion models. ArXiv abs/2402.05375 (2024), https://api.sementicscholar.org/CorpusID:267547985 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.529, + 0.787, + 0.585 + ], + "angle": 0, + "content": "25. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.787, + 0.626 + ], + "angle": 0, + "content": "26. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083 (2017) 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.628, + 0.787, + 0.683 + ], + "angle": 0, + "content": "27. Marcelo, P.: Fact focus: Fake image of pentagon explosion briefly sends jitters through stock market. Associated Press (May 2023), https://apnews.com/article/pentagon-explosion-fake-image-stock-market-jitters-78d5913603bdf9f17cc7a8c77a72e59b 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.685, + 0.787, + 0.727 + ], + "angle": 0, + "content": "28. Ni, M., Wu, C., Wang, X., Yin, S.S., Wang, L., Liu, Z., Duan, N.: Ores: Open-vocabulary responsible visual synthesis. ArXiv abs/2308.13785 (2023), https://api-semanticscholar.org/CorpusID:261243073 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.728, + 0.787, + 0.783 + ], + "angle": 0, + "content": "29. Nichol, A.Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., Mcgrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In: International Conference on Machine Learning. pp. 16784-16804. PMLR (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.784, + 0.787, + 0.812 + ], + "angle": 0, + "content": "30. Nie, G., Kim, C., Yang, Y., Ren, Y.: Attributing image generative models using latent fingerprints. arXiv preprint arXiv:2304.09752 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "31. OpenAI: Chatgpt. Online (2022), https://chat.openai.com/chat, accessed on February 24, 2024 10" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.504, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "R.A.C.E. for Secure T2I Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "32. Patel, M., Jung, S., Baral, C., Yang, Y.: \\(\\lambda\\)-eclipse: Multi-concept personalized text-to-image diffusion models by leveraging clip latent space. ArXiv abs/2402.05195 (2024), https://api_semanticscholar.org/CorpusID:267547418 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.229 + ], + "angle": 0, + "content": "33. Patel, M., Kim, C.S., Cheng, S., Baral, C., Yang, Y.: Eclipse: A resource-efficient text-to-image prior for image generations. ArXiv abs/2312.04655 (2023), https://api-semanticscholar.org/CorpusID:266149498 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.229, + 0.788, + 0.284 + ], + "angle": 0, + "content": "34. Pham, M., Marshall, K.O., Cohen, N., Mittal, G., Hegde, C.: Circumventing concept erasure methods for text-to-image generative models. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=ag3o2T51Ht 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.284, + 0.788, + 0.338 + ], + "angle": 0, + "content": "35. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 6, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.338, + 0.788, + 0.379 + ], + "angle": 0, + "content": "36. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022) 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.379, + 0.788, + 0.405 + ], + "angle": 0, + "content": "37. Rando, J., Paleka, D., Lindner, D., Heim, L., Tramér, F.: Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610 (2022) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.405, + 0.788, + 0.459 + ], + "angle": 0, + "content": "38. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684-10695 (June 2022) 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.514 + ], + "angle": 0, + "content": "39. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015) 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.514, + 0.788, + 0.581 + ], + "angle": 0, + "content": "40. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 22500-22510 (2022), https://api_semanticscholar.org/CorpusID:251800180 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.581, + 0.788, + 0.636 + ], + "angle": 0, + "content": "41. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.636, + 0.788, + 0.677 + ], + "angle": 0, + "content": "42. Saleh, B., Elgammal, A.: Large-scale classification of fine-art paintings: Learning the right metric on the right feature. ArXiv abs/1505.00855 (2015), https://api-semanticscholar.org/CorpusID:14168099 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.677, + 0.788, + 0.732 + ], + "angle": 0, + "content": "43. Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 22522-22531 (2022), https://apisemanticscholar.org/CorpusID:253420366 4, 8, 9, 10, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.732, + 0.788, + 0.786 + ], + "angle": 0, + "content": "44. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402 (2022) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.786, + 0.788, + 0.84 + ], + "angle": 0, + "content": "45. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Diffusion art or digital forgery? investigating data replication in diffusion models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 6048-6058 (2022), https://apisemanticscholar.org/CorpusID:2543666341" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.458, + 0.129 + ], + "angle": 0, + "content": "C. Kim*, K. Min*, Y. Yang" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "46. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Understanding and mitigating copying in diffusion models. ArXiv abs/2305.20086 (2023), https://api.sementicscholar.org/CorpusID:258987384 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "47. Tsai, Y.L., Hsu, C.Y., Xie, C., Lin, C.H., Chen, J.Y., Li, B., Chen, P.Y., Yu, C.M., ying Huang, C.: Ring-a-bell! how reliable are concept removal methods for diffusion models? ArXiv abs/2310.10012 (2023), https://api_semanticscholar.org/CorpusID:264146485 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.787, + 0.3 + ], + "angle": 0, + "content": "48. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 15897-15907 (2023), https://apisemantic scholar.org/CorpusID:257219968 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.787, + 0.356 + ], + "angle": 0, + "content": "49. Wen, Y., Jain, N., Kirchenbauer, J., Goldblum, M., Geiping, J., Goldstein, T.: Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery. Advances in Neural Information Processing Systems 36 (2024) 3, 4, 5, 7, 10, 11, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.397 + ], + "angle": 0, + "content": "50. Wen, Y., Kirchenbauer, J., Geiping, J., Goldstein, T.: Tree-ring watermarks: Fingerprints for diffusion images that are invisible and robust. arXiv preprint arXiv:2305.20030 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.453 + ], + "angle": 0, + "content": "51. Wu, B., Xu, C., Dai, X., Wan, A., Zhang, P., Tomizuka, M., Keutzer, K., Vajda, P.: Visual transformers: Token-based image representation and processing for computer vision. ArXiv abs/2006.03677 (2020), https://api_semanticscholar.org/CorpusID:219531480 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.495 + ], + "angle": 0, + "content": "52. Wu, Y., Zhang, J., Kerschbaum, F., Zhang, T.: Backdooring textual inversion for concept censorship. ArXiv abs/2308.10718 (2023), https://apisemantic scholar.org/CorpusID:261049298 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.535 + ], + "angle": 0, + "content": "53. Yang, Y., Gao, R., Wang, X., Xu, N., Xu, Q.: Mma-diffusion: Multimodal attack on diffusion models. ArXiv abs/2311.17516 (2023), https://apisemanticscholar.org/CorpusID:265498727 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.787, + 0.588 + ], + "angle": 0, + "content": "54. Yu, N., Skripniuk, V., Abdelnabi, S., Fritz, M.: Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In: Proceedings of the IEEE/CVF International conference on computer vision. pp. 14448-14457 (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.787, + 0.632 + ], + "angle": 0, + "content": "55. Yu, N., Skripniuk, V., Chen, D., Davis, L., Fritz, M.: Responsible disclosure of generative models using scalable fingerprinting. arXiv preprint arXiv:2012.08726 (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.787, + 0.674 + ], + "angle": 0, + "content": "56. Zhang, E., Wang, K., Xu, X., Wang, Z., Shi, H.: Forget-me-not: Learning to forget in text-to-image diffusion models. ArXiv abs/2303.17591 (2023), https://api.sementicscholar.org/CorpusID:257833863 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.702 + ], + "angle": 0, + "content": "57. Zhang, K.A., Xu, L., Cuesta-Infante, A., Veeramachaneni, K.: Robust invisible video watermarking with attention (2019) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.787, + 0.757 + ], + "angle": 0, + "content": "58. Zhang, Y., Jia, J., Chen, X., Chen, A., Zhang, Y., Liu, J., Ding, K., Liu, S.: To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images ... for now. ArXiv abs/2310.11868 (2023), https://api.sementicscholar.org/CorpusID:264289091 2, 3, 4, 5, 7, 8, 9, 10, 11, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "59. Zheng, Y., Yeh, R.A.: Imma: Immunizing text-to-image models against malicious adaptation. ArXiv abs/2311.18815 (2023), https://api_semanticscholar.org/CorpusID:265506125 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.799 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/R.A.C.E._ Robust Adversarial Concept Erasure for Secure Text-to-Image Diffusion Model/547862c1-28d0-440c-98c1-6204b0fe03dd_origin.pdf b/2024/R.A.C.E._ Robust Adversarial Concept Erasure for Secure Text-to-Image Diffusion Model/547862c1-28d0-440c-98c1-6204b0fe03dd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1322bdfaa30d5d46be19637081c6ea2ffde68254 --- /dev/null +++ b/2024/R.A.C.E._ Robust Adversarial Concept Erasure for Secure Text-to-Image Diffusion Model/547862c1-28d0-440c-98c1-6204b0fe03dd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d4e34a2ef3e7fa5e7151676c4f1e05bb8c274df82838d89d2e51f32f8605014d +size 10158819 diff --git a/2024/R.A.C.E._ Robust Adversarial Concept Erasure for Secure Text-to-Image Diffusion Model/full.md b/2024/R.A.C.E._ Robust Adversarial Concept Erasure for Secure Text-to-Image Diffusion Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5017d7963c159dd01f743516ed84cb265efcffeb --- /dev/null +++ b/2024/R.A.C.E._ Robust Adversarial Concept Erasure for Secure Text-to-Image Diffusion Model/full.md @@ -0,0 +1,339 @@ +# R.A.C.E.: Robust Adversarial Concept Erasure for Secure Text-to-Image Diffusion Model + +Changhoon $\mathbf{Kim}^{\star 1}$ Kyle Min\*2 Yezhou Yang $^{1}$ + +1 Arizona State University + +Intel Labs + +{kch,yz.yang}@asu.edu kyle.min@intel.com + +Abstract. In the evolving landscape of text-to-image (T2I) diffusion models, the remarkable capability to generate high-quality images from textual descriptions faces challenges with the potential misuse of reproducing sensitive content. To address this critical issue, we introduce Robust Adversarial Concept Erase (RACE), a novel approach designed to mitigate these risks by enhancing the robustness of concept erasure method for T2I models. RACE utilizes a sophisticated adversarial training framework to identify and mitigate adversarial text embeddings, significantly reducing the Attack Success Rate (ASR). Impressively, RACE achieves a $30\%$ reduction in ASR for the "nudity" concept against the leading white-box attack method. Our extensive evaluations demonstrate RACE's effectiveness in defending against both white-box and black-box attacks, marking a significant advancement in protecting T2I diffusion models from generating inappropriate or misleading imagery. This work underlines the essential need for proactive defense measures in adapting to the rapidly advancing field of adversarial challenges. Our code is publicly available: https://github.com/chkimmmmm/R.A.C.E. + +Keywords: Concept Erasure $\cdot$ Responsible Image Generative Models $\cdot$ Secure T2I Diffusion Models + +# 1 Introduction + +The field of text-to-image (T2I) diffusion models has garnered significant attention for their ability to produce high-quality images that can be adaptively generated from textual descriptions [36, 38]. This advancement is predicated on the training of T2I models with extensive datasets, often encompassing a range of content including copyrighted, explicit, and private materials [44-46]. Consequently, these models possess the capacity to inadvertently replicate protected images, potentially without user awareness [2, 45, 46]. The misuse of T2I models by malicious actors for misinformation or public opinion manipulation presents a significant concern [5, 27]. + +In response to the challenges posed by the malicious exploitation of generative models, Stable Diffusion (SD) [38] has integrated a safety checker [37] and advocates for the utilization of a watermarking module [57]. Despite these initiatives, + +![](images/205f6a9c0c8b3e58ed77835d4547748584f97ea6246b8e4c7e1b161d6422bc5e.jpg) +Original SD + +![](images/e3670e011dcfe5c24ebf75f4d7cb5399969d9b16cbd52d615c6613d0b66bffbb.jpg) + +![](images/2ab93595de143ec01edda69c1c4e55ab1da1f39ac614221f38505e037b364767.jpg) + +![](images/8b524239b697e9c2a0775e76c3fda24933548c9629b4bf48c5ef80e170e34ee2.jpg) + +![](images/ab35abca0f2f9a885e54996965f1022704bdec022c9dd6363526892752069ba8.jpg) + +![](images/f63810ed8dc1cc22ad52d01ecb3c4cb8635ba0bcf9de0e6fdc6478154269870b.jpg) +\* Added by authors + +![](images/8d0a896ea6206157a3dfb5c120ac02e20d2cadb809a0071aada2538ca54340f5.jpg) +ESD + +![](images/0edae4884de40bdd35286cfdc748a2a39a11da8492086a63adcebb2f4e198b23.jpg) + +![](images/63ce50bf4386cf8019c1ff4b23025c32858c77ef7d56fa6bf8a718cbff588a71.jpg) + +![](images/0e40f06f06dd62cdabb2d14b7eb2f580c3f79193d477ea7227c34d722bb63265.jpg) + +![](images/8314b795122a05433515e882e5161075270f60ca8af49b3690a3b376f45475d8.jpg) + +![](images/16301b4bf1a3395f993f42ba48cf79eb37f3a691e6d81a681eee7817f7b3dac6.jpg) +Fig. 1: Comparative demonstration of concept erasure, red teaming, and robust erasure within T2I diffusion models. The ESD method [9] removes targeted concepts from the original SD outputs, yet these concepts can be reconstructed using Unlearn-Diff [58]. Our proposed R.A.C.E. method showcases enhanced robustness against such red teaming reconstruction efforts. + +![](images/53fec3f3a579efac278179478c008e9af430c132ce00ab5e4a247a6785921c49.jpg) +UnlearnDiff + +![](images/c6472838582e434d0514bc2c8704e9ef14c99384fcdfe4261136e6408ae228e0.jpg) + +![](images/199929d779bde906d6917a3fad44789375cb9bba75127d4514af487d89185a8f.jpg) + +![](images/de8c9cf6bff8ab750563d9234bc89b46869640a0213a1fee2e33a53f3de37797.jpg) + +![](images/0f3cece5680d995eadf5147b60ad97736a4b90b2f7868c966d6658264d953a4d.jpg) + +![](images/9c3b02b0e780b06185c14ef62bb61a32cc0b3efa7203f53b0570b21d91eb3819.jpg) + +![](images/f1dba7589a490df884268a15e22a41faf6dabe3c7161af82e2169ddadd0dcc47.jpg) +R.A.C.E. + +![](images/5a7cf3bae1ed4f9968efde1d13dcd98d26c197b091f5628e39919c3f018f5877.jpg) + +![](images/779d885d2df40145b9f9aaae6cd7a6f2a0a5b126fc3481ed6ff8a89d9598a4a3.jpg) + +![](images/ff1b1d8b9af8d737b933e34e1bb7580175a2e9ee0c0f51e1c7eee67aa77ef45c.jpg) + +![](images/dc46c7a4e0580e1a80d2a929ba3bae21974625f8d0b22e150d26fd91ebbad718.jpg) + +![](images/7eb090393e3bb2b78909bb39ca71de1df5e293ce8c536e9881faaefb73871b68.jpg) + +the reliance on post-hoc interventions presents limitations due to their potential for circumvention [7, 18]. Consequently, the research community is pivoting towards formulating methodologies that embed safety protocols directly within the image generation pipeline. This approach ensures that content security is an intrinsic aspect of image creation [7, 18, 19, 30, 50, 54, 55]. Although these methods are valuable for identifying the source of content after an incident, their reactive + +nature highlights the necessity for proactive strategies. This includes pioneering techniques attempting early removal of sensitive content from T2I models [9, 21], thereby preempting the production of inappropriate or harmful material. + +To address the challenge of removing undesirable content from T2I models, even when users attempt to circumvent content restrictions, recent research has focused on the development of concept erasure techniques within T2I diffusion models [9, 11, 21, 56]. These techniques primarily aim to eliminate specific concepts (e.g., "nudity") by altering the text embeddings associated with these concepts to neutral representations. Despite these efforts, there remains a vulnerability wherein erased concepts can be reconstructed. This is achieved by identifying text tokens that closely align with the visual embeddings of the targeted concepts, thus enabling the regeneration of prohibited content [3, 47, 49, 53, 58]. This issue is illustrated in Fig. 1, demonstrating that even with concept erasure, T2I models can be manipulated through prompt modification to regenerate the restricted content. This underscores the imperative for a more robust concept erasure methodology that can withstand such reconstruction attempts, ensuring the integrity of content generation within T2I models. + +Acknowledging the imperative for enhanced concept erasure methodologies within T2I models, we pose a critical question: "Is it feasible to develop a concept erasure approach that is resilient against reconstruction efforts?" In pursuit of this, we introduce R.A.C.E. (Robust Adversarial Concept Erase), a novel strategy aimed at bolstering the resilience of concept erasure techniques against adversarial manipulations, as delineated in Algorithm 1. At the heart of RACE lies an adversarial training framework, leveraging insights from the effort of adversarial robustness [26]. Our method effectively identifies adversarial text embeddings capable of reconstructing erased concepts and then facilitates their integration into the T2I concept erasure workflow. + +A pivotal aspect of RACE is its ability to efficiently uncover adversarial text embeddings within a single time step of the diffusion process, an approach elaborated in Section 3.2. This efficiency not only streamlines the process of identifying adversarial examples but also facilitates the integration of our adversarial attack mechanism into the concept erasure workflow. To demonstrate the robustness of RACE, we have carried out an extensive array of experiments. These experiments validate the effectiveness of RACE in countering diverse red teaming strategies, with detailed results presented in Section 4. Our empirical investigations underscore the capacity of RACE to significantly enhance the robustness of T2I models against both white-box and black-box attacks across a broad spectrum of target concepts, including artistic, explicit, and object categories. + +We summarize the three main contributions here: + +- We are the first to present, to our knowledge, an adversarial training approach specifically designed to fortify concept erasure methods against prompt-based adversarial attacks without introducing additional modules. +- Our method, RACE, implements a computationally efficient adversarial attack method that can be plugged into the concept erasing method. + +- We show RACE significantly improves T2I models' robustness against prompts based on white/black box attacks. + +# 2 Related Works + +For a comprehensive overview of additional related works, please refer to the supplementary material. + +Text-to-Image Synthesis. The field of generative models has seen remarkable advancements, notably extending their capabilities beyond generating photorealistic images [6, 17] to include Text-to-Image (T2I) synthesis [29, 33, 36, 38, 41]. This progress has led to the development of fine-tuning techniques that allow for the customization of T2I models to user-specific needs [8, 22, 32, 40, 48], thereby enabling the creation of highly realistic images that align closely with textual prompts. However, the potential for misuse by malicious entities, using these models for purposes such as spreading misinformation [5, 27], raises significant concerns. This underscores the urgency of devising protective measures to mitigate the risk of such exploitations. + +Advanced Techniques in Concept Erasure for T2I Diffusion Models. Within the realm of machine unlearning, concept erasure for T2I Diffusion models has recently emerged as a critical area of research, focusing on the removal of sensitive or copyrighted concepts from T2I models. Methods to achieve this include guiding the image generation process or adjusting the model's weights to exclude these elements [9,11,21,24,28,43,56]. Notably, techniques by Gandikota et al. [9] and Kumari et al. [21] involve mapping sensitive concepts to null entities or benign equivalents by fine-tuning the weights of Stable Diffusion (SD) [38] models, effectively preventing the generation of undesirable content. Despite these advancements, red teaming methods have exposed potential loopholes, indicating that erased concepts might be regenerated through meticulously designed text prompts. Addressing this issue, our work contributes an adversarial training strategy aimed at bolstering the resilience of Stable Diffusion models against such text-prompt-based attacks [3,49,58], thereby enhancing the security and integrity of the content generation. + +Robustness Evaluation via Red Teaming in T2I Models. While various safety measures have been proposed to shield SD models from misuse, red teaming strategies reveal vulnerabilities that still allow for circumvention. Research has demonstrated that techniques like Textual Inversion [8] can be exploited to regenerate content previously erased from SD models [34], prompting the development of countermeasures aimed at safeguarding against such inversions [52,59]. In real-world applications, T2I services such as Midjourney predominantly rely on user-provided text prompts, making them susceptible to prompt-based red teaming attacks [3, 47, 49, 53, 58]. These methods employ sophisticated prompt optimization techniques to restore images containing erased content, with their + +Algorithm 1 Robust Adversarial Concept Erasure: RACE Algorithm +Input: Diffusion Model $\varPhi_{\theta}$ , frozen diffusion model $\varPhi_{\theta*}$ , scheduler $S$ target concept $c$ training steps $M$ , adversarial steps $N$ perturbation limit $\epsilon$ attack step size $\alpha$ for $i = 0,\dots,M$ do Sample noise $n\sim \mathcal{N}(0,1)$ , timestep $t\sim \mathcal{U}(1,1000)$ Initialize $\delta \sim \mathcal{U}(-\epsilon ,\epsilon)$ Denoise $z_{t} = S(n,t,c)$ for $j = 0,\ldots ,N$ do Perform targeted attack $\delta = \delta +\alpha \cdot sign(\nabla_{\delta} - L_{SD}(\varPhi_{\theta},z_{t},t,c,\delta))$ Clamp $\delta$ within $[- \epsilon ,\epsilon ]$ end for $\theta = \theta -\nabla_{\theta}L_{RACE}(\varPhi_{\theta},\varPhi_{\theta^{*}},z_{t},t,c,\delta)$ +end for +return $\varPhi_{\theta}$ + +efficacy contingent upon the level of model access—categorized into white-box approaches, which utilize SD's U-Net [39] for prompt optimization [3, 58], and black-box strategies, where such access is restricted [47, 49, 53]. Both approaches establish formidable benchmarks in attack success rates, as detailed in Tab. 2. However, the landscape lacks robust defense mechanisms against prompt-based red teaming, primarily due to the prohibitive computational demands associated with identifying adversarial prompts—a challenge that renders traditional adversarial training approaches impractical. Addressing this gap, our work introduces a novel defense strategy tailored to counteract prompt-based red teaming attacks, marking a significant step forward in fortifying T2I diffusion models against adversarial threats. + +# 3 Method + +Our methodology aims to expunge target concepts from T2I diffusion models through an adversarial training framework [26]. Initially, in Section 3.1, we establish the foundation by formally introducing the notations and the rationale underpinning our approach. Following this, Section 3.2 details our proposed adversarial attack, specifically designed for robust concept erasure, and delineates its integration into the adversarial training regime. + +# 3.1 Preliminaries + +Stable Diffusion Models. Our method is built upon the Stable Diffusion Model (SD) [38], which operates as the foundational architecture for our concept erasure technique. The SD model is composed of two primary elements: firstly, an image autoencoder that has been pre-trained on a diverse and extensive image dataset [6]. Within this autoencoder, an encoder function $\mathcal{E}(\cdot)$ transforms an input image $x$ into a latent representation $z = \mathcal{E}(x)$ . Conversely, a decoder + +![](images/3de98e383fa876bf1af1398071b4dce0d2fa828873eda862604549597c3717bb.jpg) + +![](images/4b974e44b7904758df9875f1592b6c7943e1edc22eb5b198c4329583f9186bf8.jpg) +Fig. 2: Single-Timestep Adversarial Attack Efficacy. This figure illustrates the Attack Success Rate (ASR) across various timesteps, alongside representative images. Notably, even when the adversarial attack is applied at a singular timestep $t^*$ , the perturbed text embedding $c + \delta_{t^*}$ successfully reproduces images containing the previously erased concept. For method details, see Sec. 3.2. + +function $\mathcal{D}(\cdot)$ aims to reconstruct the input image from its latent form, where $\mathcal{D}(z) = \hat{x}\approx x$ + +The second element is a U-Net [39]-based diffusion model trained to craft latent representations within the acquired latent space. This model facilitates the conditioning on either class labels or text embeddings derived from training data. Let us denote by $c = \mathcal{E}_{txt}(y)$ the textual embedding encoded from a conditioning text prompt $y$ , where $\mathcal{E}_{txt}$ symbolizes the text encoder, such as CLIP [35]. Under these constructs, the SD training objective is encapsulated by the loss function: + +$$ +L _ {S D} := \mathbb {E} _ {n \sim \mathcal {N} (0, 1), z, c, t} \left[ | | n - \Phi_ {\theta} \left(z _ {t}, t, c\right) | | _ {2} ^ {2} \right], \tag {1} +$$ + +where $t$ indexes the time step, $n$ represents a noise sample drawn from a standard Gaussian distribution, $z_{t}$ is the perturbed version of $z$ up to time step $t$ , and $\varPhi_{\theta}$ is the denoising network based on a U-Net architecture. During inference, a random noise sample is procured from a Gaussian distribution and denoised using $\varPhi_{\theta}$ following a scheduler $\mathcal{S}$ , operating over a sequence of predetermined time steps $T$ . The resulting denoised latent, $z_{0}$ , is then decoded to produce the final image, $\hat{x} = \mathcal{D}(z_0)$ . + +Erase and Reconstruction of Target Concept. The objective of concept erasure is to remove specific concepts, such as "nudity", from the latent space of a pre-trained diffusion model. The Erased Stable Diffusion (ESD) approach [9] introduces a method for excising these concepts from the latent representations within the Stable Diffusion framework. The erasure loss is formalized as follows: + +$$ +L _ {e r a s e} := \left\| \Phi_ {\theta} \left(z _ {t}, t, c\right) - \left(\Phi_ {\theta^ {*}} \left(z _ {t}, t\right) - \eta \left(\Phi_ {\theta^ {*}} \left(z _ {t}, t, c\right) - \Phi_ {\theta^ {*}} \left(z _ {t}, t\right)\right)\right) \right\| _ {2} ^ {2}, \tag {2} +$$ + +where $\varPhi_{\theta^{*}}$ represents the frozen denoising U-Net, and $\eta$ denotes the guidance scale associated with classifier-free guidance [15]. ESD fine-tunes $L_{erase}$ with respect to $\theta$ , guiding $\varPhi_{\theta}$ to produce outputs where the target concept is effectively nullified. Crucially, this process does not necessitate additional datasets; it operates successfully with only a concise textual description. + +Conversely, red teaming efforts aim to counteract concept erasure by crafting adversarial text prompts $\tilde{y}$ capable of resurrecting the erased concept in the generated image. White-box methods [3,58] engage in this adversarial prompt optimization by leveraging the gradients of $\varPhi_{\theta}$ . Meanwhile, black-box approaches [47, 49, 53] aim to achieve comparable outcomes without reliance on the gradients of $\varPhi_{\theta}$ . Both approaches pose significant computational demands, which presents challenges for their integration into an adversarial training framework. + +# 3.2 Adversarial Training on Concept Erased Diffusion Models + +Motivation. The primary aim of T2I diffusion models is to generate high-quality images conditioned on specific prompts. Intriguingly, the SD model's loss function, as depicted in Eq.(1), can also facilitate image classification tasks [23]. This classification capability is derived by applying Bayes' Theorem to the model's predictions $p_{\theta}(x|c_i)$ and the prior distribution $p(c)$ across a set of conditions $c_i$ , where each $c_i = \mathcal{E}_{txt}(y_i)$ represents a textual embedding of the prompt $y_i$ : + +$$ +p _ {\theta} (c _ {i} | x) = \frac {p (c _ {i}) p _ {\theta} (x | c _ {i})}{\sum_ {j} p (c _ {j}) p _ {\theta} (x | c _ {j})}. \tag {3} +$$ + +Notably, the prior terms $p(c)$ can be disregarded when they are uniformly distributed over the prompts $c_i$ (i.e., $p(c_i) = \frac{1}{N}$ ). In the context of diffusion models, directly computing $p_{\theta}(x|c_i)$ is computationally challenging, leading to the reliance on the computation of $\log p_{\theta}(x|c_i)$ and the utilization of the Evidence Lower Bound (ELBO) for optimization purposes. Leveraging approximations introduced in [14], we can approximate the posterior distribution over prompts $c_i$ as follows: + +$$ +p _ {\theta} \left(c _ {i} \mid x\right) = \frac {\exp \left\{- \mathbb {E} _ {z , n , t} \left[ | | n - \Phi_ {\theta} \left(z _ {t} , t , c _ {i}\right) | | ^ {2} \right] \right\}}{\sum_ {j} \exp \left\{- \mathbb {E} _ {z , n , t} \left[ | | n - \Phi_ {\theta} \left(z _ {t} , t , c _ {j}\right) | | ^ {2} \right] \right\}}. \tag {4} +$$ + +Building on this foundation, the Diffusion Classifier [23] proposes a method for estimating the class of a given image $x$ by finding argmin of the following + +expression: + +$$ +\underset {c} {\operatorname {a r g m i n}} \sum_ {t = 1} ^ {T} \left[ \left| \left| n - \Phi_ {\theta} \left(z _ {t}, t, c\right) \right| \right| ^ {2} \right], \tag {5} +$$ + +where $z_{t} = \mathcal{E}(x)$ represents the latent encoding of the image $x$ and the goal is to identify the label $c$ from the available set of classes $c_{i}$ . A notable insight from the Diffusion Classifier is the feasibility of classifying image $x$ even when computations are performed using a single time step $t^{*}$ . + +Single-Timestep Adversarial Attacks on Erased T2I Models. Leveraging insights from the Diffusion Classifier, we re-conceptualize the SD model's loss function in Eq.(1) as a classification mechanism. This perspective allows us to view Textual Inversion (TI) [8] as a form of targeted adversarial attack, where the objective is to optimize the conditioning text embedding $c$ to regenerate the image $x$ [34, 58]. Notably, TI is computationally intensive as it necessitates optimization across all time steps. + +Prompted by these considerations, our investigation centers around a critical inquiry: Can adversarial text embeddings be identified with just a single timestep? We investigate whether adversarial text embeddings can be effectively identified at a singular timestep $t^*$ . Our approach is geared towards nullifying the embedding of a target concept $c$ and its proximate embeddings that might facilitate the regeneration of an erased concept image $\tilde{x}$ , such as an explicit image. To this end, we devise a targeted adversarial attack to produce $\tilde{x}$ from a concept-erased diffusion model $\varPhi_{\theta}$ by introducing an adversarial perturbation $\delta$ . The perturbation $\delta$ is determined through the optimization: + +$$ +\begin{array}{l} \operatorname {a r g m i n} \left| | n - \Phi_ {\theta} \left(\tilde {z} _ {t}, t, c + \delta\right) | | _ {2} ^ {2}, \right. \tag {6} \\ | | \delta | | _ {\infty} \leq \epsilon \\ \end{array} +$$ + +where $\tilde{z} = \mathcal{E}(\tilde{x})$ denotes the latent representation of image $\tilde{x}$ , $\epsilon$ is a small number, and $c = \mathcal{E}_{txt}(y)$ encodes the textual embedding of the targeted concept, for instance, $y =$ "nudity". The Projected Gradient Descent (PGD) algorithm [26] is employed to address this optimization problem. Specifically, when selecting timestep $t^* = 500$ as the critical adversarial point, $\tilde{z}_t$ undergoes denoising via $\Phi_{\theta}(\tilde{z}_t,t,c)$ transitioning from timesteps $t = 1000$ to $t = 500$ . Subsequently, the targeted adversarial approach outlined in Eq. (6) is executed to determine $\delta_{t^*}$ at $t = 500$ . In subsequent denoising steps, $\tilde{z}_t$ is denoised considering the introduced perturbation $c + \delta_{t^*}$ . + +Our method introduces a distinctive single-timestep adversarial attack, contrasting with prior approaches that required optimization over a wide range of timesteps [3,58]. This approach enables the seamless incorporation of our attack strategy into the adversarial training process specifically for T2I concept erasure. + +To assess the efficacy of our method, we execute tests on a $\varPhi_{\theta}$ model trained for removing the "nudity" concept via the ESD method. Utilizing 142 nuditycentric prompts from the I2P dataset [43], we systematically select timesteps $t^{*}$ at 100-step intervals within the [1, 1000] range and compute the Attack Success Rate (ASR), as formulated in Eq.(8), for each $t^{*}$ . The results, depicted in + +Fig. 2, unveil a notable capability: the reconstruction of the erased concept is feasible even with the attack confined to a single timestep $t^*$ . Interestingly, after $t^* = 500$ , we observe an increasing trend in the ASR. The modified images resulting from the attack visually convey the transition back to the erased concept and reflect the ASR patterns. For additional insights into various concepts and their corresponding ASR trends across different timesteps $t^*$ , the supplementary material offers further information. + +# Adversarial Training for T2I Concept + +Erasure. Motivated by the findings from our single-timestep adversarial attack experiments, we explore the potential of such attacks to enhance the robustness of concept erasure in T2I models, posing the question: Can adversarial attacks improve the resilience of concept erasure mechanisms? + +Table 1: Performance comparison of $L_{\text{erase}}$ and $L_{\text{RACE}}$ + +
LeraseLRACE
I2P [43]0.080.05
FID [13]33.1225.16
CLIP-Score [12]0.7260.745
+ +RACE distinguishes itself from existing approaches [9,21] by aiming to eliminate not only the targeted concept's embedding but also its adjacent embedding within the model's latent space, which could otherwise lead to the inadvertent generation of the erased concept by $\varPhi_{\theta}$ . We incorporate our adversarial attack into the erasure loss function ( $L_{erase}$ ), yielding an enhanced adversarial training loss: + +$$ +L _ {R A C E} := \left| \left| \Phi_ {\theta} \left(z _ {t}, t, c + \delta\right) - \left(\Phi_ {\theta^ {*}} \left(z _ {t}, t\right) - \eta \left(\Phi_ {\theta^ {*}} \left(z _ {t}, t, c\right) - \Phi_ {\theta^ {*}} \left(z _ {t}, t\right)\right)\right) \right| \right| _ {2} ^ {2}. \tag {7} +$$ + +This method is deliberate, substituting the concept embedding $c$ with $c + \delta$ within the trainable parameters of $\varPhi_{\theta}$ . This precise adjustment ensures enhanced fidelity of the generated images by mapping the $\epsilon$ -neighborhood of the concept embedding to its null representation. Comparative metrics between the direct substitution in $L_{erase}$ and the strategic use of $L_{RACE}$ are provided in Tab. 1. The latter approach demonstrates promising reductions in ASR for "nudity" prompts within I2P dataset [43] and improvements in image quality metrics, as assessed on the MS-COCO [25]. The RACE methodology is comprehensively detailed in Algorithm 1. To validate RACE's effectiveness, we evaluate the ASR against both white-box and black-box attacks, as elaborated in Sec. 4. + +# 4 Experiments + +# 4.1 Experimental Setting + +Datasets. Our assessment of the RACE framework spans various domains, including artistic styles, explicit concepts, and identifiable objects, in line with established benchmarks [3,9,58]. To ensure a uniform image generation process, we standardize key hyperparameters such as the scale of classifier-free guidance and random seeds. Artistic style evaluations leverage shared text prompts from + +ESD [9] and UnlearnDiff [58]. For explicit content, we draw from the inappropriate image prompt benchmark [43], selecting a diverse set of prompts encompassing 142 nudity, 98 illegal acts, and 101 violence instances. Objects are curated from a subset of Imagenette [16], known for its distinct and recognizable classes, with text prompts synthesized via ChatGPT [31] following the approach by Kumari et al. [21]. It's crucial to note that these datasets serve solely to gauge the effectiveness of adversarial attacks; the adversarial training component of RACE does not necessitate the use of these specific prompts. + +Training Details. In validating RACE, we choose the Erased Stable Diffusion (ESD) model [9] for its ability to erase a wide range of concepts, serving as an ideal testbed to showcase RACE's efficacy. We integrate RACE with ESD, optimizing with attack parameters $\epsilon = 0.1$ and $\alpha = \epsilon /4$ . The optimization of model parameters $\theta$ utilizes the Adam optimizer [20] at a learning rate of $1e - 5$ , consistent with ESD. Additional details on attack parameter selection are available in the supplementary material. + +Red Teaming Methods. To rigorously test RACE's robustness, we deploy a comprehensive suite of adversarial attacks, spanning both white-box and black-box approaches. Initial assessments utilize the I2P red teaming prompt dataset [43]. In black-box scenarios, we employ PEZ [49], which crafts adversarial prompts via CLIP [35]. In the white-box scenario, methods like P4D [3] and UnlearnDiff [58] are used, which generate adversarial prompts by leveraging gradients from the SD. + +Evaluation. To gauge the robustness of RACE, we employ domain-specific classifiers: a ViT-base model [51] pre-trained on ImageNet [4] and fine-tuned on WikiArt [42] for artistic styles, Nudenet [1] for explicit content, and ResNet-50 [10] trained on ImageNet for object removal. We measure robustness using the Attack Success Rate (ASR): + +$$ +\operatorname {A S R} = \frac {1}{N} \sum_ {i = 1} ^ {N} \mathbb {1} \left(f \left(S D \left(\tilde {y} _ {i}\right)\right) = \tilde {y} _ {i}\right), \tag {8} +$$ + +where $\tilde{y}_i$ is the adversarial prompt, $f$ is the classifier, and $N$ is the number of prompts. Additionally, we assess image quality after applying RACE by generating 5,000 images from the MS-COCO [25] test set, computing the Frechet Inception Distance (FID) score [13] and the CLIP score [12] to evaluate RACE's impact on image fidelity while ensuring concept erasure. + +# 4.2 Robust Concept Erase against Red Teaming + +In our comprehensive analysis, RACE undergoes a series of red teaming evaluations [3, 43, 49, 58], encompassing both white-box and black-box techniques aimed at regenerating concepts targeted by RACE for erasure, as depicted in + +Table 2: Attack Success Rate (ASR) against white $(\bigcirc)$ /black $(\bullet)$ box attacks. We conduct experiments on artistic style (Van Gogh), explicit concepts (nudity, violence, and illegal acts), and objects (church, golf ball, and parachute). "ESD/RACE-Concept" denotes the concept erased from the model. We also measure the quality of T2I models. We can observe that RACE reduces ASR by over $30\%$ (-0.30 or below) in absolute value on Van Gogh, nudity, and church for UnlearnDiff [58], which is the previous SOTA attack method. + +
PromptsPEZ [49]P4D [3]UnlearnDiff [58]CLIP-Score [12]FID [13]
White/Black Box●●○○--
ESD [9]-VanGogh0.040.000.260.360.799719.16
ESD [9]-Nudity0.140.080.750.800.793118.88
ESD [9]-Violence0.270.130.840.790.783421.55
ESD [9]-Illegal0.290.200.890.850.785421.50
ESD [9]-Church0.160.000.580.680.789619.68
ESD [9]-GolfBall0.040.000.160.160.773820.64
ESD [9]-Parachute0.060.040.480.600.786519.72
RACE-VanGogh0.00 (-0.04)0.00 (-0.00)0.00 (-0.26)0.04 (-0.32)0.802420.65
RACE-Nudity0.05 (-0.09)0.02 (-0.06)0.49 (-0.26)0.47 (-0.33)0.745225.16
RACE-Violence0.11 (-0.16)0.08 (-0.05)0.75 (-0.09)0.68 (-0.11)0.737428.71
RACE-Illlegal0.20 (-0.09)0.13 (-0.07)0.85 (-0.04)0.80 (-0.05)0.759124.87
RACE-Church0.02 (-0.14)0.00 (-0.00)0.26 (-0.32)0.38 (-0.30)0.773023.92
RACE-GolfBall0.00 (-0.04)0.00 (-0.00)0.10 (-0.06)0.06 (-0.10)0.748025.38
RACE-Parachute0.02 (-0.04)0.00 (-0.04)0.24 (-0.24)0.38 (-0.22)0.757026.42
+ +Fig.1. The comparative analysis of ASR presented in Tab.2 spans diverse conceptual domains, from "Van Gogh"-inspired artistry to explicit content such as "nudity" and "violence", extending to tangible objects like "churches", "golf balls", and "parachutes". Remarkably, RACE consistently diminishes ASR, notably surpassing $30\%$ for "Van Gogh" styles, "nudity", and "church" categories, particularly outperforming UnlearnDiff [58], the current state-of-the-art in white-box adversarial methodologies. This marked decline in ASR underscores RACE's heightened robustness and delineates its capability as a potent, computationally efficient defense mechanism for T2I diffusion models against intricate adversarial attacks. Crucially, RACE's methodological advantage stems from its independence from external imagery or prompts, diverging from traditional red teaming techniques reliant on such data for prompt generation. + +As demonstrated in Tab.2, RACE achieves a significant $33\%$ reduction in ASR for the "nudity" concept, underscoring its effectiveness. To further elucidate how RACE enhances ASR, we analyze the specific categories or elements it targets for removal. Utilizing Nudenet, we enumerate the body parts generated by various models—original SD, ESD, UnlearnDiff, and RACE—when prompted with nudity-related inputs from the I2P dataset. Illustrated in Fig.3, the analysis reveals that while ESD enhances the original SD's resilience to such prompts, UnlearnDiff manages to bypass ESD's defenses, reconstructing explicit content. In contrast, RACE maintains its robustness even against the sophisticated UnlearnDiff attacks, showcasing the advanced protective capabilities of our approach in safeguarding against the regeneration of sensitive content. + +One caveat to mention is that our experiments reveal a nuanced trade-off between robustness and image quality. While artistic style erasures maintain + +![](images/8109cb659eb06404b45ab762e449479b0fe4ce0bd162f8321df64716e87e71a3.jpg) +Fig. 3: Although ESD significantly reduces the chance of generating images with exposed body parts, state-of-the-art red teaming methods, such as UnlearnDiff, can be used to bypass ESD's defense and reconstruct explicit content. RACE and its variant can effectively defend the malicious attempts to reconstruct explicit content from the ESD model that erased the concept of nudity. + +quality metrics, erasures of other concepts inadvertently degrade image quality. This divergence could be attributed to methodological differences; erasing artistic styles predominantly involves fine-tuning the cross-attention layers of SD as per ESD guidelines, whereas erasing other concepts necessitates adjustments in non-cross-attention layers [9]. Another plausible explanation for the observed trade-off between robust concept erasure and overall image quality could relate to the inherent complexity of differentiating between closely related or overlapping concepts within the model's latent space. As RACE intensifies adversarial robustness, it may inadvertently alter the delicate equilibrium within these conceptual overlaps, leading to unintended modifications in adjacent, non-targeted conceptual representations. This issue highlights the complex tension between precise concept erasure and maintaining the model's overall integrity against attacks. To address the quality concerns arising from this trade-off, we explore a potential strategy for improvement in Sec. 4.4, aiming to reduce the trade-off between targeted erasure with high quality and the model's defensive robustness. + +# 4.3 Disentanglement + +Investigating RACE's disentanglement performance, our study focuses on its capability to precisely erase intended concepts without impacting other elements. The evaluation spans both qualitative and quantitative measures. On the qualitative front, we configure separate RACE-enhanced models to specifically erase artistic imprints such as "Van Gogh", "Thomas Kinkade", and "Kilian Eng" from the SD. The illustrative outcomes, showcased in Fig. 4, underscore RACE's erasure precision, ensuring that the excision of one artistic style doesn't lead to the collateral removal of others. This meticulous erasure extends to discrete object concepts like "church", "golf ball", and "parachute", with the generated images further affirming the method's discernment. + +![](images/21077ca7c798cbb9362adaf70208ccc27f451b96eaa0539861ebd4937fecec10.jpg) +Fig. 4: RACE's Disentanglement in Concept Erasure. This figure highlights RACE's precision in erasing specific concepts, as shown in diagonal images, while preserving unrelated concepts, which is evident in off-diagonal images. For reference, baseline images generated by the original Stable Diffusion (SD) model are also presented. + +Table 4: Ablation Studies for Performance Improvement. Using the "nudity" concept, we evaluate the effectiveness of adding weight regularization and close concept keywords. ASR and quality metrics are measured to assess performance. + +
I2P [43]PEZ [49]P4D [3]UnleranDiff [58]CLIP-Score [12]FID [13]
ESD0.140.080.750.800.793118.88
RACE0.050.020.490.470.745225.16
RACE+Reg.0.070.020.600.620.759324.42
RACE+keywords0.020.010.420.460.720130.97
+ +Quantitatively, we generate 5,000 prompts with varied random seeds, such as "an image of a [class name]", to produce a diverse set of images, subsequently evaluated using a pre-trained ResNet-50 classifier for top-1 accuracy. As shown in Tab. 3, RACE showcases an improved capability for object concept erasure compared to ESD. Despite this, there exists a slight decrement in classification accuracy for non-target classes. This effect + +likely stems from RACE's method of targeting the $\epsilon$ -neighborhood surrounding the intended concept, potentially influencing proximate concepts. Nevertheless, we can observe from Fig. 4 that RACE can precisely erase the target concept with minimal visual impacts on other concepts. + +Table 3: Accuracy of erased and non-erased classes in Imagenet [4]. We evaluate the classification accuracy of ESD and RACE for the target erased concepts and other non-target concepts. + +
Erased ConceptAcc. ErasedAcc. Others
ESDRACEESDRACE
Church0.160.020.570.53
Golf Ball0.040.000.450.35
Parachute0.060.020.570.41
+ +# 4.4 Discussion + +Potential Strategy to Improve the Robustness-Quality Trade-off. Our findings underscore the ability of RACE to significantly bolster the SD model's + +defense against prompt-based adversarial attacks across a variety of concepts. However, as illustrated in Tab.2, enhancing robustness appears to inversely impact image quality. To address this dichotomy, we test a refined version of the RACE loss function incorporating a regularization term: + +$$ +L _ {R A C E + R e g.} := L _ {R A C E} + \lambda \| \theta - \theta^ {*} \| _ {1}, \tag {9} +$$ + +where $\theta$ and $\theta^{*}$ represent the parameters of the RACE and original Stable Diffusion models, respectively, and $\lambda$ is the regularization strength, set to 0.1 in our experiment. This regularization approach, as evidenced in Tab.4, helps to partially reconcile the robustness-quality trade-off, enhancing image quality while maintaining improved robustness over ESD. + +Enhancing Concept Erasure. In pursuit of further reducing the ASR, we explore strategies for more comprehensive concept erasure. Recognizing that a target concept may manifest in various synonymous forms, we extend RACE's erasure scope to include semantically related concepts. Leveraging the CLIP text encoder embedded within Stable Diffusion, we identify and subsequently erase concepts closely related to the target, based on their proximity in the CLIP embedding space. For instance, alongside "nudity", we also target synonymous concepts like "nude", "nsfw", and "bare", identified as the top-3 semantically similar terms. In Fig. 3, we can observe that our method equipped with this expanded erasure strategy (denoted as RACE+keywords) is more effective in defending the malicious attempts to bypass the ESD by further reducing the number of exposed body parts. Tab. 4 also indicates that our strategy indeed fortifies the model's robustness against red teaming tactics. Nonetheless, this broadened concept removal spectrum reaffirms the robustness-quality trade-off, manifesting as a decrement in image quality. This aspect opens an intriguing avenue for future enhancements to the RACE methodology, balancing the twin objectives of robust concept erasure and preserved image fidelity. + +# 5 Conclusion + +In this work, we present RACE, a novel defense approach designed to protect the Text-to-Image Stable Diffusion models from prompt-based red teaming attacks. RACE effectively strengthens the model's concept erasure capabilities while maintaining computational efficiency, offering a valuable enhancement to the current erasure framework and bolstering defenses against various adversarial techniques. We also observe the robustness-quality trade-off and discuss possible future directions to improve it. This initial contribution lays the groundwork for further exploration, underscoring the critical importance of developing sophisticated defenses in the rapidly evolving domain of generative AI. + +Acknowledgements. We would like to express our gratitude to Sangmin Jung for assisting with the experiments. This work is partially supported by the National Science Foundation under Grant No. 2038666 and No. 2101052. The views and opinions of the authors expressed herein do not necessarily state or reflect those of the funding agencies and employers. + +# References + +1. Bedapudi, P.: Nudenet: Neural nets for nudity classification, detection and selective censoring (12 2019) 10 +2. Carlini, N., Hayes, J., Nasr, M., Jagielski, M., Sehwag, V., Tramér, F., Balle, B., Ippolito, D., Wallace, E.: Extracting training data from diffusion models. ArXiv abs/2301.13188 (2023), https://api-semanticscholar.org/CorpusID: 256389993 1 +3. Chin, Z.Y., Jiang, C.M., Huang, C.C., Chen, P.Y., Chiu, W.C.: Prompting4debugging: Red-teaming text-to-image diffusion models by finding problematic prompts. arXiv preprint arXiv:2309.06135 (2023) 3, 4, 5, 7, 8, 9, 10, 11, 13 +4. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) 10, 13 +5. Devlin, K., Cheetham, J.: Fake trump arrest photos: How to spot an ai-generated image. BBC News (March 2023), https://www.bbc.com/news/technology-68981525 1, 4 +6. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12873-12883 (2021) 4, 5 +7. Fernandez, P., Couairon, G., Jégou, H., Douze, M., Furon, T.: The stable signature: Rooting watermarks in latent diffusion models. arXiv preprint arXiv:2303.15435 (2023) 2 +8. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. ArXiv abs/2208.01618 (2022), https://apisemantic scholar.org/CorpusID:251253049 4, 8 +9. Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., Bau, D.: Erasing concepts from diffusion models. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 2426-2436 (2023), https://apisemantic scholar.org/CorpusID:257495777 2, 3, 4, 7, 9, 10, 11, 12 +0. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 10 +1. Heng, A., Soh, H.: Selective amnesia: A continual learning approach to forgetting in deep generative models. ArXiv abs/2305.10120 (2023), https://apisemantic scholar.org/CorpusID:258740988 3, 4 +2. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021) 9, 10, 11, 13 +3. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Advances in Neural Information Processing Systems. pp. 6626-6637 (2017) 9, 10, 11, 13 +4. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020) 7 +5. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022) 7 +6. Howard, J., Gugger, S.: fastai: A layered api for deep learning. Inf. 11, 108 (2020), https://api-semanticscholar.org/CorpusID:211082837 10 + +17. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8110-8119 (2020) 4 +18. Kim, C., Min, K., Patel, M., Cheng, S., Yang, Y.: Wouaf: Weight modulation for user attribution and fingerprinting in text-to-image diffusion models. arXiv preprint arXiv:2306.04744 (2023) 2 +19. Kim, C., Ren, Y., Yang, Y.: Decentralized attribution of generative models. In: International Conference on Learning Representations (2021) 2 +20. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization (2014), http://arxiv.org/abs/1412.6980, cite arxiv:1412.6980Comment: Published as a conference paper at the 3rd International Conference for Learning Representations, San Diego, 2015 10 +21. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating concepts in text-to-image diffusion models. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 22634-22645 (2023), https://apisemanticscholar.org/CorpusID:257687839 3, 4, 9, 10 +22. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 1931-1941 (2022), https://api-semanticscholar.org/CorpusID:254408780 4 +23. Li, A.C., Prabhudesai, M., Duggal, S., Brown, E., Pathak, D.: Your diffusion model is secretly a zero-shot classifier. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2206-2217 (October 2023) 7 +24. Li, S., van de Weijer, J., Hu, T., Khan, F.S., Hou, Q., Wang, Y., Yang, J.: Get what you want, not what you don't: Image content suppression for text-to-image diffusion models. ArXiv abs/2402.05375 (2024), https://api.sementicscholar.org/CorpusID:267547985 4 +25. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) 9, 10 +26. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083 (2017) 3, 5, 8 +27. Marcelo, P.: Fact focus: Fake image of pentagon explosion briefly sends jitters through stock market. Associated Press (May 2023), https://apnews.com/article/pentagon-explosion-fake-image-stock-market-jitters-78d5913603bdf9f17cc7a8c77a72e59b 1, 4 +28. Ni, M., Wu, C., Wang, X., Yin, S.S., Wang, L., Liu, Z., Duan, N.: Ores: Open-vocabulary responsible visual synthesis. ArXiv abs/2308.13785 (2023), https://api-semanticscholar.org/CorpusID:261243073 4 +29. Nichol, A.Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., Mcgrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In: International Conference on Machine Learning. pp. 16784-16804. PMLR (2022) 4 +30. Nie, G., Kim, C., Yang, Y., Ren, Y.: Attributing image generative models using latent fingerprints. arXiv preprint arXiv:2304.09752 (2023) 2 +31. OpenAI: Chatgpt. Online (2022), https://chat.openai.com/chat, accessed on February 24, 2024 10 + +32. Patel, M., Jung, S., Baral, C., Yang, Y.: $\lambda$ -eclipse: Multi-concept personalized text-to-image diffusion models by leveraging clip latent space. ArXiv abs/2402.05195 (2024), https://api_semanticscholar.org/CorpusID:267547418 4 +33. Patel, M., Kim, C.S., Cheng, S., Baral, C., Yang, Y.: Eclipse: A resource-efficient text-to-image prior for image generations. ArXiv abs/2312.04655 (2023), https://api-semanticscholar.org/CorpusID:266149498 4 +34. Pham, M., Marshall, K.O., Cohen, N., Mittal, G., Hegde, C.: Circumventing concept erasure methods for text-to-image generative models. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=ag3o2T51Ht 4, 8 +35. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 6, 10 +36. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022) 1, 4 +37. Rando, J., Paleka, D., Lindner, D., Heim, L., Tramér, F.: Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610 (2022) 1 +38. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684-10695 (June 2022) 1, 4, 5 +39. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015) 5, 6 +40. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 22500-22510 (2022), https://api_semanticscholar.org/CorpusID:251800180 4 +41. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022) 4 +42. Saleh, B., Elgammal, A.: Large-scale classification of fine-art paintings: Learning the right metric on the right feature. ArXiv abs/1505.00855 (2015), https://api-semanticscholar.org/CorpusID:14168099 10 +43. Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 22522-22531 (2022), https://apisemanticscholar.org/CorpusID:253420366 4, 8, 9, 10, 13 +44. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402 (2022) 1 +45. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Diffusion art or digital forgery? investigating data replication in diffusion models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 6048-6058 (2022), https://apisemanticscholar.org/CorpusID:2543666341 + +46. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Understanding and mitigating copying in diffusion models. ArXiv abs/2305.20086 (2023), https://api.sementicscholar.org/CorpusID:258987384 1 +47. Tsai, Y.L., Hsu, C.Y., Xie, C., Lin, C.H., Chen, J.Y., Li, B., Chen, P.Y., Yu, C.M., ying Huang, C.: Ring-a-bell! how reliable are concept removal methods for diffusion models? ArXiv abs/2310.10012 (2023), https://api_semanticscholar.org/CorpusID:264146485 3, 4, 5, 7 +48. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 15897-15907 (2023), https://apisemantic scholar.org/CorpusID:257219968 4 +49. Wen, Y., Jain, N., Kirchenbauer, J., Goldblum, M., Geiping, J., Goldstein, T.: Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery. Advances in Neural Information Processing Systems 36 (2024) 3, 4, 5, 7, 10, 11, 13 +50. Wen, Y., Kirchenbauer, J., Geiping, J., Goldstein, T.: Tree-ring watermarks: Fingerprints for diffusion images that are invisible and robust. arXiv preprint arXiv:2305.20030 (2023) 2 +51. Wu, B., Xu, C., Dai, X., Wan, A., Zhang, P., Tomizuka, M., Keutzer, K., Vajda, P.: Visual transformers: Token-based image representation and processing for computer vision. ArXiv abs/2006.03677 (2020), https://api_semanticscholar.org/CorpusID:219531480 10 +52. Wu, Y., Zhang, J., Kerschbaum, F., Zhang, T.: Backdooring textual inversion for concept censorship. ArXiv abs/2308.10718 (2023), https://apisemantic scholar.org/CorpusID:261049298 4 +53. Yang, Y., Gao, R., Wang, X., Xu, N., Xu, Q.: Mma-diffusion: Multimodal attack on diffusion models. ArXiv abs/2311.17516 (2023), https://apisemanticscholar.org/CorpusID:265498727 3, 4, 5, 7 +54. Yu, N., Skripniuk, V., Abdelnabi, S., Fritz, M.: Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In: Proceedings of the IEEE/CVF International conference on computer vision. pp. 14448-14457 (2021) 2 +55. Yu, N., Skripniuk, V., Chen, D., Davis, L., Fritz, M.: Responsible disclosure of generative models using scalable fingerprinting. arXiv preprint arXiv:2012.08726 (2020) 2 +56. Zhang, E., Wang, K., Xu, X., Wang, Z., Shi, H.: Forget-me-not: Learning to forget in text-to-image diffusion models. ArXiv abs/2303.17591 (2023), https://api.sementicscholar.org/CorpusID:257833863 3, 4 +57. Zhang, K.A., Xu, L., Cuesta-Infante, A., Veeramachaneni, K.: Robust invisible video watermarking with attention (2019) 1 +58. Zhang, Y., Jia, J., Chen, X., Chen, A., Zhang, Y., Liu, J., Ding, K., Liu, S.: To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images ... for now. ArXiv abs/2310.11868 (2023), https://api.sementicscholar.org/CorpusID:264289091 2, 3, 4, 5, 7, 8, 9, 10, 11, 13 +59. Zheng, Y., Yeh, R.A.: Imma: Immunizing text-to-image models against malicious adaptation. ArXiv abs/2311.18815 (2023), https://api_semanticscholar.org/CorpusID:265506125 4 \ No newline at end of file diff --git a/2024/R.A.C.E._ Robust Adversarial Concept Erasure for Secure Text-to-Image Diffusion Model/images.zip b/2024/R.A.C.E._ Robust Adversarial Concept Erasure for Secure Text-to-Image Diffusion Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..50125fe23c3b46f3b6bb86ab042fa8a78c3be75e --- /dev/null +++ b/2024/R.A.C.E._ Robust Adversarial Concept Erasure for Secure Text-to-Image Diffusion Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9ca895a0efad8cd1ebc5a3afb0a844826817b7c2621783f96036d036ed61ac0a +size 598522 diff --git a/2024/R.A.C.E._ Robust Adversarial Concept Erasure for Secure Text-to-Image Diffusion Model/layout.json b/2024/R.A.C.E._ Robust Adversarial Concept Erasure for Secure Text-to-Image Diffusion Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..51f839b76f3f100667feb3e4ce8c3c898f11e67c --- /dev/null +++ b/2024/R.A.C.E._ Robust Adversarial Concept Erasure for Secure Text-to-Image Diffusion Model/layout.json @@ -0,0 +1,10708 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 139, + 111, + 475, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 111, + 475, + 148 + ], + "spans": [ + { + "bbox": [ + 139, + 111, + 475, + 148 + ], + "type": "text", + "content": "R.A.C.E.: Robust Adversarial Concept Erasure for Secure Text-to-Image Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 185, + 167, + 428, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 167, + 428, + 180 + ], + "spans": [ + { + "bbox": [ + 185, + 167, + 428, + 180 + ], + "type": "text", + "content": "Changhoon " + }, + { + "bbox": [ + 185, + 167, + 428, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{Kim}^{\\star 1}" + }, + { + "bbox": [ + 185, + 167, + 428, + 180 + ], + "type": "text", + "content": " Kyle Min\\*2 Yezhou Yang" + }, + { + "bbox": [ + 185, + 167, + 428, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 251, + 189, + 362, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 189, + 362, + 201 + ], + "spans": [ + { + "bbox": [ + 251, + 189, + 362, + 201 + ], + "type": "text", + "content": "1 Arizona State University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 281, + 201, + 333, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 201, + 333, + 211 + ], + "spans": [ + { + "bbox": [ + 281, + 201, + 333, + 211 + ], + "type": "text", + "content": "Intel Labs" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 209, + 213, + 405, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 213, + 405, + 223 + ], + "spans": [ + { + "bbox": [ + 209, + 213, + 405, + 223 + ], + "type": "text", + "content": "{kch,yz.yang}@asu.edu kyle.min@intel.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 245, + 455, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 245, + 455, + 434 + ], + "spans": [ + { + "bbox": [ + 159, + 245, + 455, + 434 + ], + "type": "text", + "content": "Abstract. In the evolving landscape of text-to-image (T2I) diffusion models, the remarkable capability to generate high-quality images from textual descriptions faces challenges with the potential misuse of reproducing sensitive content. To address this critical issue, we introduce Robust Adversarial Concept Erase (RACE), a novel approach designed to mitigate these risks by enhancing the robustness of concept erasure method for T2I models. RACE utilizes a sophisticated adversarial training framework to identify and mitigate adversarial text embeddings, significantly reducing the Attack Success Rate (ASR). Impressively, RACE achieves a " + }, + { + "bbox": [ + 159, + 245, + 455, + 434 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 159, + 245, + 455, + 434 + ], + "type": "text", + "content": " reduction in ASR for the \"nudity\" concept against the leading white-box attack method. Our extensive evaluations demonstrate RACE's effectiveness in defending against both white-box and black-box attacks, marking a significant advancement in protecting T2I diffusion models from generating inappropriate or misleading imagery. This work underlines the essential need for proactive defense measures in adapting to the rapidly advancing field of adversarial challenges. Our code is publicly available: https://github.com/chkimmmmm/R.A.C.E." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 442, + 453, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 442, + 453, + 464 + ], + "spans": [ + { + "bbox": [ + 159, + 442, + 453, + 464 + ], + "type": "text", + "content": "Keywords: Concept Erasure " + }, + { + "bbox": [ + 159, + 442, + 453, + 464 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 442, + 453, + 464 + ], + "type": "text", + "content": " Responsible Image Generative Models " + }, + { + "bbox": [ + 159, + 442, + 453, + 464 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 442, + 453, + 464 + ], + "type": "text", + "content": " Secure T2I Diffusion Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 482, + 230, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 230, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 230, + 495 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 504, + 482, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 504, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 504, + 482, + 612 + ], + "type": "text", + "content": "The field of text-to-image (T2I) diffusion models has garnered significant attention for their ability to produce high-quality images that can be adaptively generated from textual descriptions [36, 38]. This advancement is predicated on the training of T2I models with extensive datasets, often encompassing a range of content including copyrighted, explicit, and private materials [44-46]. Consequently, these models possess the capacity to inadvertently replicate protected images, potentially without user awareness [2, 45, 46]. The misuse of T2I models by malicious actors for misinformation or public opinion manipulation presents a significant concern [5, 27]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 612, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 612, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 612, + 482, + 648 + ], + "type": "text", + "content": "In response to the challenges posed by the malicious exploitation of generative models, Stable Diffusion (SD) [38] has integrated a safety checker [37] and advocates for the utilization of a watermarking module [57]. Despite these initiatives," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 654, + 337, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 654, + 337, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 654, + 337, + 665 + ], + "type": "text", + "content": "* These authors contributed equally to this work." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 181, + 133, + 242, + 193 + ], + "blocks": [ + { + "bbox": [ + 186, + 119, + 238, + 131 + ], + "lines": [ + { + "bbox": [ + 186, + 119, + 238, + 131 + ], + "spans": [ + { + "bbox": [ + 186, + 119, + 238, + 131 + ], + "type": "text", + "content": "Original SD" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 181, + 133, + 242, + 193 + ], + "lines": [ + { + "bbox": [ + 181, + 133, + 242, + 193 + ], + "spans": [ + { + "bbox": [ + 181, + 133, + 242, + 193 + ], + "type": "image", + "image_path": "205f6a9c0c8b3e58ed77835d4547748584f97ea6246b8e4c7e1b161d6422bc5e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 181, + 194, + 241, + 254 + ], + "blocks": [ + { + "bbox": [ + 181, + 194, + 241, + 254 + ], + "lines": [ + { + "bbox": [ + 181, + 194, + 241, + 254 + ], + "spans": [ + { + "bbox": [ + 181, + 194, + 241, + 254 + ], + "type": "image", + "image_path": "e3670e011dcfe5c24ebf75f4d7cb5399969d9b16cbd52d615c6613d0b66bffbb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 181, + 255, + 241, + 316 + ], + "blocks": [ + { + "bbox": [ + 181, + 255, + 241, + 316 + ], + "lines": [ + { + "bbox": [ + 181, + 255, + 241, + 316 + ], + "spans": [ + { + "bbox": [ + 181, + 255, + 241, + 316 + ], + "type": "image", + "image_path": "2ab93595de143ec01edda69c1c4e55ab1da1f39ac614221f38505e037b364767.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 181, + 316, + 241, + 374 + ], + "blocks": [ + { + "bbox": [ + 181, + 316, + 241, + 374 + ], + "lines": [ + { + "bbox": [ + 181, + 316, + 241, + 374 + ], + "spans": [ + { + "bbox": [ + 181, + 316, + 241, + 374 + ], + "type": "image", + "image_path": "8b524239b697e9c2a0775e76c3fda24933548c9629b4bf48c5ef80e170e34ee2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 181, + 376, + 241, + 436 + ], + "blocks": [ + { + "bbox": [ + 181, + 376, + 241, + 436 + ], + "lines": [ + { + "bbox": [ + 181, + 376, + 241, + 436 + ], + "spans": [ + { + "bbox": [ + 181, + 376, + 241, + 436 + ], + "type": "image", + "image_path": "ab35abca0f2f9a885e54996965f1022704bdec022c9dd6363526892752069ba8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 181, + 437, + 241, + 496 + ], + "blocks": [ + { + "bbox": [ + 181, + 437, + 241, + 496 + ], + "lines": [ + { + "bbox": [ + 181, + 437, + 241, + 496 + ], + "spans": [ + { + "bbox": [ + 181, + 437, + 241, + 496 + ], + "type": "image", + "image_path": "f63810ed8dc1cc22ad52d01ecb3c4cb8635ba0bcf9de0e6fdc6478154269870b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 497, + 250, + 507 + ], + "lines": [ + { + "bbox": [ + 181, + 497, + 250, + 507 + ], + "spans": [ + { + "bbox": [ + 181, + 497, + 250, + 507 + ], + "type": "text", + "content": "\\* Added by authors" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 244, + 134, + 317, + 194 + ], + "blocks": [ + { + "bbox": [ + 277, + 119, + 298, + 129 + ], + "lines": [ + { + "bbox": [ + 277, + 119, + 298, + 129 + ], + "spans": [ + { + "bbox": [ + 277, + 119, + 298, + 129 + ], + "type": "text", + "content": "ESD" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 244, + 134, + 317, + 194 + ], + "lines": [ + { + "bbox": [ + 244, + 134, + 317, + 194 + ], + "spans": [ + { + "bbox": [ + 244, + 134, + 317, + 194 + ], + "type": "image", + "image_path": "8d0a896ea6206157a3dfb5c120ac02e20d2cadb809a0071aada2538ca54340f5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 244, + 195, + 317, + 254 + ], + "blocks": [ + { + "bbox": [ + 244, + 195, + 317, + 254 + ], + "lines": [ + { + "bbox": [ + 244, + 195, + 317, + 254 + ], + "spans": [ + { + "bbox": [ + 244, + 195, + 317, + 254 + ], + "type": "image", + "image_path": "0edae4884de40bdd35286cfdc748a2a39a11da8492086a63adcebb2f4e198b23.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 244, + 255, + 317, + 316 + ], + "blocks": [ + { + "bbox": [ + 244, + 255, + 317, + 316 + ], + "lines": [ + { + "bbox": [ + 244, + 255, + 317, + 316 + ], + "spans": [ + { + "bbox": [ + 244, + 255, + 317, + 316 + ], + "type": "image", + "image_path": "63ce50bf4386cf8019c1ff4b23025c32858c77ef7d56fa6bf8a718cbff588a71.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 244, + 317, + 317, + 374 + ], + "blocks": [ + { + "bbox": [ + 244, + 317, + 317, + 374 + ], + "lines": [ + { + "bbox": [ + 244, + 317, + 317, + 374 + ], + "spans": [ + { + "bbox": [ + 244, + 317, + 317, + 374 + ], + "type": "image", + "image_path": "0e40f06f06dd62cdabb2d14b7eb2f580c3f79193d477ea7227c34d722bb63265.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 244, + 376, + 317, + 437 + ], + "blocks": [ + { + "bbox": [ + 244, + 376, + 317, + 437 + ], + "lines": [ + { + "bbox": [ + 244, + 376, + 317, + 437 + ], + "spans": [ + { + "bbox": [ + 244, + 376, + 317, + 437 + ], + "type": "image", + "image_path": "8314b795122a05433515e882e5161075270f60ca8af49b3690a3b376f45475d8.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 244, + 437, + 317, + 496 + ], + "blocks": [ + { + "bbox": [ + 244, + 437, + 317, + 496 + ], + "lines": [ + { + "bbox": [ + 244, + 437, + 317, + 496 + ], + "spans": [ + { + "bbox": [ + 244, + 437, + 317, + 496 + ], + "type": "image", + "image_path": "16301b4bf1a3395f993f42ba48cf79eb37f3a691e6d81a681eee7817f7b3dac6.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 523, + 482, + 579 + ], + "lines": [ + { + "bbox": [ + 130, + 523, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 523, + 482, + 579 + ], + "type": "text", + "content": "Fig. 1: Comparative demonstration of concept erasure, red teaming, and robust erasure within T2I diffusion models. The ESD method [9] removes targeted concepts from the original SD outputs, yet these concepts can be reconstructed using Unlearn-Diff [58]. Our proposed R.A.C.E. method showcases enhanced robustness against such red teaming reconstruction efforts." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 321, + 134, + 376, + 194 + ], + "blocks": [ + { + "bbox": [ + 321, + 119, + 374, + 130 + ], + "lines": [ + { + "bbox": [ + 321, + 119, + 374, + 130 + ], + "spans": [ + { + "bbox": [ + 321, + 119, + 374, + 130 + ], + "type": "text", + "content": "UnlearnDiff" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 321, + 134, + 376, + 194 + ], + "lines": [ + { + "bbox": [ + 321, + 134, + 376, + 194 + ], + "spans": [ + { + "bbox": [ + 321, + 134, + 376, + 194 + ], + "type": "image", + "image_path": "53fec3f3a579efac278179478c008e9af430c132ce00ab5e4a247a6785921c49.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 321, + 195, + 376, + 254 + ], + "blocks": [ + { + "bbox": [ + 321, + 195, + 376, + 254 + ], + "lines": [ + { + "bbox": [ + 321, + 195, + 376, + 254 + ], + "spans": [ + { + "bbox": [ + 321, + 195, + 376, + 254 + ], + "type": "image", + "image_path": "c6472838582e434d0514bc2c8704e9ef14c99384fcdfe4261136e6408ae228e0.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 321, + 255, + 376, + 316 + ], + "blocks": [ + { + "bbox": [ + 321, + 255, + 376, + 316 + ], + "lines": [ + { + "bbox": [ + 321, + 255, + 376, + 316 + ], + "spans": [ + { + "bbox": [ + 321, + 255, + 376, + 316 + ], + "type": "image", + "image_path": "199929d779bde906d6917a3fad44789375cb9bba75127d4514af487d89185a8f.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 321, + 317, + 376, + 374 + ], + "blocks": [ + { + "bbox": [ + 321, + 317, + 376, + 374 + ], + "lines": [ + { + "bbox": [ + 321, + 317, + 376, + 374 + ], + "spans": [ + { + "bbox": [ + 321, + 317, + 376, + 374 + ], + "type": "image", + "image_path": "de8c9cf6bff8ab750563d9234bc89b46869640a0213a1fee2e33a53f3de37797.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 321, + 376, + 376, + 437 + ], + "blocks": [ + { + "bbox": [ + 321, + 376, + 376, + 437 + ], + "lines": [ + { + "bbox": [ + 321, + 376, + 376, + 437 + ], + "spans": [ + { + "bbox": [ + 321, + 376, + 376, + 437 + ], + "type": "image", + "image_path": "0f3cece5680d995eadf5147b60ad97736a4b90b2f7868c966d6658264d953a4d.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 321, + 437, + 376, + 496 + ], + "blocks": [ + { + "bbox": [ + 321, + 437, + 376, + 496 + ], + "lines": [ + { + "bbox": [ + 321, + 437, + 376, + 496 + ], + "spans": [ + { + "bbox": [ + 321, + 437, + 376, + 496 + ], + "type": "image", + "image_path": "9c3b02b0e780b06185c14ef62bb61a32cc0b3efa7203f53b0570b21d91eb3819.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 377, + 134, + 436, + 194 + ], + "blocks": [ + { + "bbox": [ + 387, + 119, + 427, + 130 + ], + "lines": [ + { + "bbox": [ + 387, + 119, + 427, + 130 + ], + "spans": [ + { + "bbox": [ + 387, + 119, + 427, + 130 + ], + "type": "text", + "content": "R.A.C.E." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 377, + 134, + 436, + 194 + ], + "lines": [ + { + "bbox": [ + 377, + 134, + 436, + 194 + ], + "spans": [ + { + "bbox": [ + 377, + 134, + 436, + 194 + ], + "type": "image", + "image_path": "f1dba7589a490df884268a15e22a41faf6dabe3c7161af82e2169ddadd0dcc47.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 377, + 195, + 436, + 254 + ], + "blocks": [ + { + "bbox": [ + 377, + 195, + 436, + 254 + ], + "lines": [ + { + "bbox": [ + 377, + 195, + 436, + 254 + ], + "spans": [ + { + "bbox": [ + 377, + 195, + 436, + 254 + ], + "type": "image", + "image_path": "5a7cf3bae1ed4f9968efde1d13dcd98d26c197b091f5628e39919c3f018f5877.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 377, + 255, + 436, + 316 + ], + "blocks": [ + { + "bbox": [ + 377, + 255, + 436, + 316 + ], + "lines": [ + { + "bbox": [ + 377, + 255, + 436, + 316 + ], + "spans": [ + { + "bbox": [ + 377, + 255, + 436, + 316 + ], + "type": "image", + "image_path": "779d885d2df40145b9f9aaae6cd7a6f2a0a5b126fc3481ed6ff8a89d9598a4a3.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 377, + 317, + 436, + 374 + ], + "blocks": [ + { + "bbox": [ + 377, + 317, + 436, + 374 + ], + "lines": [ + { + "bbox": [ + 377, + 317, + 436, + 374 + ], + "spans": [ + { + "bbox": [ + 377, + 317, + 436, + 374 + ], + "type": "image", + "image_path": "ff1b1d8b9af8d737b933e34e1bb7580175a2e9ee0c0f51e1c7eee67aa77ef45c.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 377, + 376, + 436, + 437 + ], + "blocks": [ + { + "bbox": [ + 377, + 376, + 436, + 437 + ], + "lines": [ + { + "bbox": [ + 377, + 376, + 436, + 437 + ], + "spans": [ + { + "bbox": [ + 377, + 376, + 436, + 437 + ], + "type": "image", + "image_path": "dc46c7a4e0580e1a80d2a929ba3bae21974625f8d0b22e150d26fd91ebbad718.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 377, + 437, + 436, + 496 + ], + "blocks": [ + { + "bbox": [ + 377, + 437, + 436, + 496 + ], + "lines": [ + { + "bbox": [ + 377, + 437, + 436, + 496 + ], + "spans": [ + { + "bbox": [ + 377, + 437, + 436, + 496 + ], + "type": "image", + "image_path": "7eb090393e3bb2b78909bb39ca71de1df5e293ce8c536e9881faaefb73871b68.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "the reliance on post-hoc interventions presents limitations due to their potential for circumvention [7, 18]. Consequently, the research community is pivoting towards formulating methodologies that embed safety protocols directly within the image generation pipeline. This approach ensures that content security is an intrinsic aspect of image creation [7, 18, 19, 30, 50, 54, 55]. Although these methods are valuable for identifying the source of content after an incident, their reactive" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 90, + 280, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 90, + 280, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 90, + 280, + 102 + ], + "type": "text", + "content": "C. Kim*, K. Min*, Y. Yang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "nature highlights the necessity for proactive strategies. This includes pioneering techniques attempting early removal of sensitive content from T2I models [9, 21], thereby preempting the production of inappropriate or harmful material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 156, + 482, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 156, + 482, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 156, + 482, + 324 + ], + "type": "text", + "content": "To address the challenge of removing undesirable content from T2I models, even when users attempt to circumvent content restrictions, recent research has focused on the development of concept erasure techniques within T2I diffusion models [9, 11, 21, 56]. These techniques primarily aim to eliminate specific concepts (e.g., \"nudity\") by altering the text embeddings associated with these concepts to neutral representations. Despite these efforts, there remains a vulnerability wherein erased concepts can be reconstructed. This is achieved by identifying text tokens that closely align with the visual embeddings of the targeted concepts, thus enabling the regeneration of prohibited content [3, 47, 49, 53, 58]. This issue is illustrated in Fig. 1, demonstrating that even with concept erasure, T2I models can be manipulated through prompt modification to regenerate the restricted content. This underscores the imperative for a more robust concept erasure methodology that can withstand such reconstruction attempts, ensuring the integrity of content generation within T2I models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 328, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 482, + 449 + ], + "type": "text", + "content": "Acknowledging the imperative for enhanced concept erasure methodologies within T2I models, we pose a critical question: \"Is it feasible to develop a concept erasure approach that is resilient against reconstruction efforts?\" In pursuit of this, we introduce R.A.C.E. (Robust Adversarial Concept Erase), a novel strategy aimed at bolstering the resilience of concept erasure techniques against adversarial manipulations, as delineated in Algorithm 1. At the heart of RACE lies an adversarial training framework, leveraging insights from the effort of adversarial robustness [26]. Our method effectively identifies adversarial text embeddings capable of reconstructing erased concepts and then facilitates their integration into the T2I concept erasure workflow." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 453, + 482, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 585 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 585 + ], + "type": "text", + "content": "A pivotal aspect of RACE is its ability to efficiently uncover adversarial text embeddings within a single time step of the diffusion process, an approach elaborated in Section 3.2. This efficiency not only streamlines the process of identifying adversarial examples but also facilitates the integration of our adversarial attack mechanism into the concept erasure workflow. To demonstrate the robustness of RACE, we have carried out an extensive array of experiments. These experiments validate the effectiveness of RACE in countering diverse red teaming strategies, with detailed results presented in Section 4. Our empirical investigations underscore the capacity of RACE to significantly enhance the robustness of T2I models against both white-box and black-box attacks across a broad spectrum of target concepts, including artistic, explicit, and object categories." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 146, + 589, + 366, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 589, + 366, + 600 + ], + "spans": [ + { + "bbox": [ + 146, + 589, + 366, + 600 + ], + "type": "text", + "content": "We summarize the three main contributions here:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 605, + 487, + 665 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 605, + 487, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 605, + 487, + 640 + ], + "spans": [ + { + "bbox": [ + 138, + 605, + 487, + 640 + ], + "type": "text", + "content": "- We are the first to present, to our knowledge, an adversarial training approach specifically designed to fortify concept erasure methods against prompt-based adversarial attacks without introducing additional modules." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 139, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 139, + 641, + 480, + 665 + ], + "type": "text", + "content": "- Our method, RACE, implements a computationally efficient adversarial attack method that can be plugged into the concept erasing method." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 308, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 308, + 91, + 447, + 101 + ], + "type": "text", + "content": "R.A.C.E. for Secure T2I Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 139, + 116, + 487, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 116, + 487, + 140 + ], + "spans": [ + { + "bbox": [ + 139, + 116, + 487, + 140 + ], + "type": "text", + "content": "- We show RACE significantly improves T2I models' robustness against prompts based on white/black box attacks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 159, + 242, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 159, + 242, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 159, + 242, + 172 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 186, + 480, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 186, + 480, + 210 + ], + "spans": [ + { + "bbox": [ + 130, + 186, + 480, + 210 + ], + "type": "text", + "content": "For a comprehensive overview of additional related works, please refer to the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 222, + 482, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 222, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 222, + 482, + 342 + ], + "type": "text", + "content": "Text-to-Image Synthesis. The field of generative models has seen remarkable advancements, notably extending their capabilities beyond generating photorealistic images [6, 17] to include Text-to-Image (T2I) synthesis [29, 33, 36, 38, 41]. This progress has led to the development of fine-tuning techniques that allow for the customization of T2I models to user-specific needs [8, 22, 32, 40, 48], thereby enabling the creation of highly realistic images that align closely with textual prompts. However, the potential for misuse by malicious entities, using these models for purposes such as spreading misinformation [5, 27], raises significant concerns. This underscores the urgency of devising protective measures to mitigate the risk of such exploitations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 354, + 482, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 534 + ], + "type": "text", + "content": "Advanced Techniques in Concept Erasure for T2I Diffusion Models. Within the realm of machine unlearning, concept erasure for T2I Diffusion models has recently emerged as a critical area of research, focusing on the removal of sensitive or copyrighted concepts from T2I models. Methods to achieve this include guiding the image generation process or adjusting the model's weights to exclude these elements [9,11,21,24,28,43,56]. Notably, techniques by Gandikota et al. [9] and Kumari et al. [21] involve mapping sensitive concepts to null entities or benign equivalents by fine-tuning the weights of Stable Diffusion (SD) [38] models, effectively preventing the generation of undesirable content. Despite these advancements, red teaming methods have exposed potential loopholes, indicating that erased concepts might be regenerated through meticulously designed text prompts. Addressing this issue, our work contributes an adversarial training strategy aimed at bolstering the resilience of Stable Diffusion models against such text-prompt-based attacks [3,49,58], thereby enhancing the security and integrity of the content generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Robustness Evaluation via Red Teaming in T2I Models. While various safety measures have been proposed to shield SD models from misuse, red teaming strategies reveal vulnerabilities that still allow for circumvention. Research has demonstrated that techniques like Textual Inversion [8] can be exploited to regenerate content previously erased from SD models [34], prompting the development of countermeasures aimed at safeguarding against such inversions [52,59]. In real-world applications, T2I services such as Midjourney predominantly rely on user-provided text prompts, making them susceptible to prompt-based red teaming attacks [3, 47, 49, 53, 58]. These methods employ sophisticated prompt optimization techniques to restore images containing erased content, with their" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 280, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 280, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 280, + 102 + ], + "type": "text", + "content": "C. Kim*, K. Min*, Y. Yang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 132, + 130, + 482, + 276 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 447, + 129 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 447, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 447, + 129 + ], + "type": "text", + "content": "Algorithm 1 Robust Adversarial Concept Erasure: RACE Algorithm" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "lines": [ + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "spans": [ + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "text", + "content": "Input: Diffusion Model " + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\theta}" + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "text", + "content": " , frozen diffusion model " + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\theta*}" + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "text", + "content": " , scheduler " + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "text", + "content": " target concept " + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "text", + "content": " training steps " + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "text", + "content": " , adversarial steps " + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "text", + "content": " perturbation limit " + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "text", + "content": " attack step size " + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "inline_equation", + "content": "i = 0,\\dots,M" + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "text", + "content": " do Sample noise " + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "inline_equation", + "content": "n\\sim \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "text", + "content": " , timestep " + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "inline_equation", + "content": "t\\sim \\mathcal{U}(1,1000)" + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "text", + "content": " Initialize " + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "inline_equation", + "content": "\\delta \\sim \\mathcal{U}(-\\epsilon ,\\epsilon)" + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "text", + "content": " Denoise " + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "inline_equation", + "content": "z_{t} = S(n,t,c)" + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "inline_equation", + "content": "j = 0,\\ldots ,N" + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "text", + "content": " do Perform targeted attack " + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "inline_equation", + "content": "\\delta = \\delta +\\alpha \\cdot sign(\\nabla_{\\delta} - L_{SD}(\\varPhi_{\\theta},z_{t},t,c,\\delta))" + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "text", + "content": " Clamp " + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "text", + "content": " within " + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "inline_equation", + "content": "[- \\epsilon ,\\epsilon ]" + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "text", + "content": " end for " + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "inline_equation", + "content": "\\theta = \\theta -\\nabla_{\\theta}L_{RACE}(\\varPhi_{\\theta},\\varPhi_{\\theta^{*}},z_{t},t,c,\\delta)" + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "text", + "content": " \nend for \nreturn " + }, + { + "bbox": [ + 132, + 130, + 482, + 276 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\theta}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 300, + 482, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 433 + ], + "type": "text", + "content": "efficacy contingent upon the level of model access—categorized into white-box approaches, which utilize SD's U-Net [39] for prompt optimization [3, 58], and black-box strategies, where such access is restricted [47, 49, 53]. Both approaches establish formidable benchmarks in attack success rates, as detailed in Tab. 2. However, the landscape lacks robust defense mechanisms against prompt-based red teaming, primarily due to the prohibitive computational demands associated with identifying adversarial prompts—a challenge that renders traditional adversarial training approaches impractical. Addressing this gap, our work introduces a novel defense strategy tailored to counteract prompt-based red teaming attacks, marking a significant step forward in fortifying T2I diffusion models against adversarial threats." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 453, + 202, + 465 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 202, + 465 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 202, + 465 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 479, + 482, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 553 + ], + "type": "text", + "content": "Our methodology aims to expunge target concepts from T2I diffusion models through an adversarial training framework [26]. Initially, in Section 3.1, we establish the foundation by formally introducing the notations and the rationale underpinning our approach. Following this, Section 3.2 details our proposed adversarial attack, specifically designed for robust concept erasure, and delineates its integration into the adversarial training regime." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 571, + 228, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 571, + 228, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 571, + 228, + 582 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": "Stable Diffusion Models. Our method is built upon the Stable Diffusion Model (SD) [38], which operates as the foundational architecture for our concept erasure technique. The SD model is composed of two primary elements: firstly, an image autoencoder that has been pre-trained on a diverse and extensive image dataset [6]. Within this autoencoder, an encoder function " + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\cdot)" + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": " transforms an input image " + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": " into a latent representation " + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "inline_equation", + "content": "z = \\mathcal{E}(x)" + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": ". Conversely, a decoder" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 308, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 308, + 91, + 447, + 101 + ], + "type": "text", + "content": "R.A.C.E. for Secure T2I Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 173, + 114, + 449, + 247 + ], + "blocks": [ + { + "bbox": [ + 173, + 114, + 449, + 247 + ], + "lines": [ + { + "bbox": [ + 173, + 114, + 449, + 247 + ], + "spans": [ + { + "bbox": [ + 173, + 114, + 449, + 247 + ], + "type": "image", + "image_path": "3de98e383fa876bf1af1398071b4dce0d2fa828873eda862604549597c3717bb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 173, + 247, + 446, + 372 + ], + "blocks": [ + { + "bbox": [ + 173, + 247, + 446, + 372 + ], + "lines": [ + { + "bbox": [ + 173, + 247, + 446, + 372 + ], + "spans": [ + { + "bbox": [ + 173, + 247, + 446, + 372 + ], + "type": "image", + "image_path": "4b974e44b7904758df9875f1592b6c7943e1edc22eb5b198c4329583f9186bf8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 380, + 482, + 437 + ], + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 437 + ], + "type": "text", + "content": "Fig. 2: Single-Timestep Adversarial Attack Efficacy. This figure illustrates the Attack Success Rate (ASR) across various timesteps, alongside representative images. Notably, even when the adversarial attack is applied at a singular timestep " + }, + { + "bbox": [ + 130, + 380, + 482, + 437 + ], + "type": "inline_equation", + "content": "t^*" + }, + { + "bbox": [ + 130, + 380, + 482, + 437 + ], + "type": "text", + "content": ", the perturbed text embedding " + }, + { + "bbox": [ + 130, + 380, + 482, + 437 + ], + "type": "inline_equation", + "content": "c + \\delta_{t^*}" + }, + { + "bbox": [ + 130, + 380, + 482, + 437 + ], + "type": "text", + "content": " successfully reproduces images containing the previously erased concept. For method details, see Sec. 3.2." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 445, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 468 + ], + "type": "text", + "content": "function " + }, + { + "bbox": [ + 130, + 445, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\cdot)" + }, + { + "bbox": [ + 130, + 445, + 482, + 468 + ], + "type": "text", + "content": " aims to reconstruct the input image from its latent form, where " + }, + { + "bbox": [ + 130, + 445, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(z) = \\hat{x}\\approx x" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 468, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 542 + ], + "type": "text", + "content": "The second element is a U-Net [39]-based diffusion model trained to craft latent representations within the acquired latent space. This model facilitates the conditioning on either class labels or text embeddings derived from training data. Let us denote by " + }, + { + "bbox": [ + 130, + 468, + 482, + 542 + ], + "type": "inline_equation", + "content": "c = \\mathcal{E}_{txt}(y)" + }, + { + "bbox": [ + 130, + 468, + 482, + 542 + ], + "type": "text", + "content": " the textual embedding encoded from a conditioning text prompt " + }, + { + "bbox": [ + 130, + 468, + 482, + 542 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 468, + 482, + 542 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 468, + 482, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{txt}" + }, + { + "bbox": [ + 130, + 468, + 482, + 542 + ], + "type": "text", + "content": " symbolizes the text encoder, such as CLIP [35]. Under these constructs, the SD training objective is encapsulated by the loss function:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 211, + 551, + 481, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 551, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 211, + 551, + 481, + 566 + ], + "type": "interline_equation", + "content": "L _ {S D} := \\mathbb {E} _ {n \\sim \\mathcal {N} (0, 1), z, c, t} \\left[ | | n - \\Phi_ {\\theta} \\left(z _ {t}, t, c\\right) | | _ {2} ^ {2} \\right], \\tag {1}", + "image_path": "5552244c567d963cc413e9558cf0b62a3ab99790472bbf54f27caee28a487a61.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 654 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 570, + 482, + 654 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 570, + 482, + 654 + ], + "type": "text", + "content": " indexes the time step, " + }, + { + "bbox": [ + 130, + 570, + 482, + 654 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 570, + 482, + 654 + ], + "type": "text", + "content": " represents a noise sample drawn from a standard Gaussian distribution, " + }, + { + "bbox": [ + 130, + 570, + 482, + 654 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 570, + 482, + 654 + ], + "type": "text", + "content": " is the perturbed version of " + }, + { + "bbox": [ + 130, + 570, + 482, + 654 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 570, + 482, + 654 + ], + "type": "text", + "content": " up to time step " + }, + { + "bbox": [ + 130, + 570, + 482, + 654 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 570, + 482, + 654 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 570, + 482, + 654 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\theta}" + }, + { + "bbox": [ + 130, + 570, + 482, + 654 + ], + "type": "text", + "content": " is the denoising network based on a U-Net architecture. During inference, a random noise sample is procured from a Gaussian distribution and denoised using " + }, + { + "bbox": [ + 130, + 570, + 482, + 654 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\theta}" + }, + { + "bbox": [ + 130, + 570, + 482, + 654 + ], + "type": "text", + "content": " following a scheduler " + }, + { + "bbox": [ + 130, + 570, + 482, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 570, + 482, + 654 + ], + "type": "text", + "content": ", operating over a sequence of predetermined time steps " + }, + { + "bbox": [ + 130, + 570, + 482, + 654 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 570, + 482, + 654 + ], + "type": "text", + "content": ". The resulting denoised latent, " + }, + { + "bbox": [ + 130, + 570, + 482, + 654 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 130, + 570, + 482, + 654 + ], + "type": "text", + "content": ", is then decoded to produce the final image, " + }, + { + "bbox": [ + 130, + 570, + 482, + 654 + ], + "type": "inline_equation", + "content": "\\hat{x} = \\mathcal{D}(z_0)" + }, + { + "bbox": [ + 130, + 570, + 482, + 654 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 280, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 280, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 280, + 102 + ], + "type": "text", + "content": "C. Kim*, K. Min*, Y. Yang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 175 + ], + "type": "text", + "content": "Erase and Reconstruction of Target Concept. The objective of concept erasure is to remove specific concepts, such as \"nudity\", from the latent space of a pre-trained diffusion model. The Erased Stable Diffusion (ESD) approach [9] introduces a method for excising these concepts from the latent representations within the Stable Diffusion framework. The erasure loss is formalized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 162, + 189, + 481, + 203 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 189, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 162, + 189, + 481, + 203 + ], + "type": "interline_equation", + "content": "L _ {e r a s e} := \\left\\| \\Phi_ {\\theta} \\left(z _ {t}, t, c\\right) - \\left(\\Phi_ {\\theta^ {*}} \\left(z _ {t}, t\\right) - \\eta \\left(\\Phi_ {\\theta^ {*}} \\left(z _ {t}, t, c\\right) - \\Phi_ {\\theta^ {*}} \\left(z _ {t}, t\\right)\\right)\\right) \\right\\| _ {2} ^ {2}, \\tag {2}", + "image_path": "8594b676a1627b1a8b96800d95acf25151a155f6dcf228f71c4ea0afb7897d6d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 209, + 479, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 209, + 479, + 268 + ], + "spans": [ + { + "bbox": [ + 130, + 209, + 479, + 268 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 209, + 479, + 268 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\theta^{*}}" + }, + { + "bbox": [ + 130, + 209, + 479, + 268 + ], + "type": "text", + "content": " represents the frozen denoising U-Net, and " + }, + { + "bbox": [ + 130, + 209, + 479, + 268 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 209, + 479, + 268 + ], + "type": "text", + "content": " denotes the guidance scale associated with classifier-free guidance [15]. ESD fine-tunes " + }, + { + "bbox": [ + 130, + 209, + 479, + 268 + ], + "type": "inline_equation", + "content": "L_{erase}" + }, + { + "bbox": [ + 130, + 209, + 479, + 268 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 130, + 209, + 479, + 268 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 209, + 479, + 268 + ], + "type": "text", + "content": ", guiding " + }, + { + "bbox": [ + 130, + 209, + 479, + 268 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\theta}" + }, + { + "bbox": [ + 130, + 209, + 479, + 268 + ], + "type": "text", + "content": " to produce outputs where the target concept is effectively nullified. Crucially, this process does not necessitate additional datasets; it operates successfully with only a concise textual description." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 269, + 482, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 482, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 482, + 353 + ], + "type": "text", + "content": "Conversely, red teaming efforts aim to counteract concept erasure by crafting adversarial text prompts " + }, + { + "bbox": [ + 130, + 269, + 482, + 353 + ], + "type": "inline_equation", + "content": "\\tilde{y}" + }, + { + "bbox": [ + 130, + 269, + 482, + 353 + ], + "type": "text", + "content": " capable of resurrecting the erased concept in the generated image. White-box methods [3,58] engage in this adversarial prompt optimization by leveraging the gradients of " + }, + { + "bbox": [ + 130, + 269, + 482, + 353 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\theta}" + }, + { + "bbox": [ + 130, + 269, + 482, + 353 + ], + "type": "text", + "content": ". Meanwhile, black-box approaches [47, 49, 53] aim to achieve comparable outcomes without reliance on the gradients of " + }, + { + "bbox": [ + 130, + 269, + 482, + 353 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\theta}" + }, + { + "bbox": [ + 130, + 269, + 482, + 353 + ], + "type": "text", + "content": ". Both approaches pose significant computational demands, which presents challenges for their integration into an adversarial training framework." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 373, + 452, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 373, + 452, + 385 + ], + "spans": [ + { + "bbox": [ + 131, + 373, + 452, + 385 + ], + "type": "text", + "content": "3.2 Adversarial Training on Concept Erased Diffusion Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 395, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 395, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 131, + 395, + 482, + 468 + ], + "type": "text", + "content": "Motivation. The primary aim of T2I diffusion models is to generate high-quality images conditioned on specific prompts. Intriguingly, the SD model's loss function, as depicted in Eq.(1), can also facilitate image classification tasks [23]. This classification capability is derived by applying Bayes' Theorem to the model's predictions " + }, + { + "bbox": [ + 131, + 395, + 482, + 468 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x|c_i)" + }, + { + "bbox": [ + 131, + 395, + 482, + 468 + ], + "type": "text", + "content": " and the prior distribution " + }, + { + "bbox": [ + 131, + 395, + 482, + 468 + ], + "type": "inline_equation", + "content": "p(c)" + }, + { + "bbox": [ + 131, + 395, + 482, + 468 + ], + "type": "text", + "content": " across a set of conditions " + }, + { + "bbox": [ + 131, + 395, + 482, + 468 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 131, + 395, + 482, + 468 + ], + "type": "text", + "content": ", where each " + }, + { + "bbox": [ + 131, + 395, + 482, + 468 + ], + "type": "inline_equation", + "content": "c_i = \\mathcal{E}_{txt}(y_i)" + }, + { + "bbox": [ + 131, + 395, + 482, + 468 + ], + "type": "text", + "content": " represents a textual embedding of the prompt " + }, + { + "bbox": [ + 131, + 395, + 482, + 468 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 131, + 395, + 482, + 468 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 244, + 479, + 481, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 479, + 481, + 506 + ], + "spans": [ + { + "bbox": [ + 244, + 479, + 481, + 506 + ], + "type": "interline_equation", + "content": "p _ {\\theta} (c _ {i} | x) = \\frac {p (c _ {i}) p _ {\\theta} (x | c _ {i})}{\\sum_ {j} p (c _ {j}) p _ {\\theta} (x | c _ {j})}. \\tag {3}", + "image_path": "af774ddc6929bae5143947ddd4a7d27cb691bfded32cf2762e7d0475ae3337e1.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 511, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 511, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 131, + 511, + 482, + 594 + ], + "type": "text", + "content": "Notably, the prior terms " + }, + { + "bbox": [ + 131, + 511, + 482, + 594 + ], + "type": "inline_equation", + "content": "p(c)" + }, + { + "bbox": [ + 131, + 511, + 482, + 594 + ], + "type": "text", + "content": " can be disregarded when they are uniformly distributed over the prompts " + }, + { + "bbox": [ + 131, + 511, + 482, + 594 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 131, + 511, + 482, + 594 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 131, + 511, + 482, + 594 + ], + "type": "inline_equation", + "content": "p(c_i) = \\frac{1}{N}" + }, + { + "bbox": [ + 131, + 511, + 482, + 594 + ], + "type": "text", + "content": "). In the context of diffusion models, directly computing " + }, + { + "bbox": [ + 131, + 511, + 482, + 594 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x|c_i)" + }, + { + "bbox": [ + 131, + 511, + 482, + 594 + ], + "type": "text", + "content": " is computationally challenging, leading to the reliance on the computation of " + }, + { + "bbox": [ + 131, + 511, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\log p_{\\theta}(x|c_i)" + }, + { + "bbox": [ + 131, + 511, + 482, + 594 + ], + "type": "text", + "content": " and the utilization of the Evidence Lower Bound (ELBO) for optimization purposes. Leveraging approximations introduced in [14], we can approximate the posterior distribution over prompts " + }, + { + "bbox": [ + 131, + 511, + 482, + 594 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 131, + 511, + 482, + 594 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 199, + 603, + 481, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 603, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 199, + 603, + 481, + 632 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(c _ {i} \\mid x\\right) = \\frac {\\exp \\left\\{- \\mathbb {E} _ {z , n , t} \\left[ | | n - \\Phi_ {\\theta} \\left(z _ {t} , t , c _ {i}\\right) | | ^ {2} \\right] \\right\\}}{\\sum_ {j} \\exp \\left\\{- \\mathbb {E} _ {z , n , t} \\left[ | | n - \\Phi_ {\\theta} \\left(z _ {t} , t , c _ {j}\\right) | | ^ {2} \\right] \\right\\}}. \\tag {4}", + "image_path": "0fd544e8ef2091f37df1b9c85fceb829cfb58e4e34c269edeb10eb200d69034a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "Building on this foundation, the Diffusion Classifier [23] proposes a method for estimating the class of a given image " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": " by finding argmin of the following" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 308, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 308, + 91, + 447, + 101 + ], + "type": "text", + "content": "R.A.C.E. for Secure T2I Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 183, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 183, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 183, + 128 + ], + "type": "text", + "content": "expression:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 237, + 128, + 481, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 128, + 481, + 159 + ], + "spans": [ + { + "bbox": [ + 237, + 128, + 481, + 159 + ], + "type": "interline_equation", + "content": "\\underset {c} {\\operatorname {a r g m i n}} \\sum_ {t = 1} ^ {T} \\left[ \\left| \\left| n - \\Phi_ {\\theta} \\left(z _ {t}, t, c\\right) \\right| \\right| ^ {2} \\right], \\tag {5}", + "image_path": "b7461cf33454ea3cb5c36b6239bae6efa264a2258887d40866955ba82f5480bf.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 163, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 163, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 163, + 482, + 212 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 163, + 482, + 212 + ], + "type": "inline_equation", + "content": "z_{t} = \\mathcal{E}(x)" + }, + { + "bbox": [ + 130, + 163, + 482, + 212 + ], + "type": "text", + "content": " represents the latent encoding of the image " + }, + { + "bbox": [ + 130, + 163, + 482, + 212 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 163, + 482, + 212 + ], + "type": "text", + "content": " and the goal is to identify the label " + }, + { + "bbox": [ + 130, + 163, + 482, + 212 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 163, + 482, + 212 + ], + "type": "text", + "content": " from the available set of classes " + }, + { + "bbox": [ + 130, + 163, + 482, + 212 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 163, + 482, + 212 + ], + "type": "text", + "content": ". A notable insight from the Diffusion Classifier is the feasibility of classifying image " + }, + { + "bbox": [ + 130, + 163, + 482, + 212 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 163, + 482, + 212 + ], + "type": "text", + "content": " even when computations are performed using a single time step " + }, + { + "bbox": [ + 130, + 163, + 482, + 212 + ], + "type": "inline_equation", + "content": "t^{*}" + }, + { + "bbox": [ + 130, + 163, + 482, + 212 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 223, + 482, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 223, + 482, + 306 + ], + "spans": [ + { + "bbox": [ + 130, + 223, + 482, + 306 + ], + "type": "text", + "content": "Single-Timestep Adversarial Attacks on Erased T2I Models. Leveraging insights from the Diffusion Classifier, we re-conceptualize the SD model's loss function in Eq.(1) as a classification mechanism. This perspective allows us to view Textual Inversion (TI) [8] as a form of targeted adversarial attack, where the objective is to optimize the conditioning text embedding " + }, + { + "bbox": [ + 130, + 223, + 482, + 306 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 223, + 482, + 306 + ], + "type": "text", + "content": " to regenerate the image " + }, + { + "bbox": [ + 130, + 223, + 482, + 306 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 223, + 482, + 306 + ], + "type": "text", + "content": " [34, 58]. Notably, TI is computationally intensive as it necessitates optimization across all time steps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 307, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 482, + 415 + ], + "type": "text", + "content": "Prompted by these considerations, our investigation centers around a critical inquiry: Can adversarial text embeddings be identified with just a single timestep? We investigate whether adversarial text embeddings can be effectively identified at a singular timestep " + }, + { + "bbox": [ + 130, + 307, + 482, + 415 + ], + "type": "inline_equation", + "content": "t^*" + }, + { + "bbox": [ + 130, + 307, + 482, + 415 + ], + "type": "text", + "content": ". Our approach is geared towards nullifying the embedding of a target concept " + }, + { + "bbox": [ + 130, + 307, + 482, + 415 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 307, + 482, + 415 + ], + "type": "text", + "content": " and its proximate embeddings that might facilitate the regeneration of an erased concept image " + }, + { + "bbox": [ + 130, + 307, + 482, + 415 + ], + "type": "inline_equation", + "content": "\\tilde{x}" + }, + { + "bbox": [ + 130, + 307, + 482, + 415 + ], + "type": "text", + "content": ", such as an explicit image. To this end, we devise a targeted adversarial attack to produce " + }, + { + "bbox": [ + 130, + 307, + 482, + 415 + ], + "type": "inline_equation", + "content": "\\tilde{x}" + }, + { + "bbox": [ + 130, + 307, + 482, + 415 + ], + "type": "text", + "content": " from a concept-erased diffusion model " + }, + { + "bbox": [ + 130, + 307, + 482, + 415 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\theta}" + }, + { + "bbox": [ + 130, + 307, + 482, + 415 + ], + "type": "text", + "content": " by introducing an adversarial perturbation " + }, + { + "bbox": [ + 130, + 307, + 482, + 415 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 130, + 307, + 482, + 415 + ], + "type": "text", + "content": ". The perturbation " + }, + { + "bbox": [ + 130, + 307, + 482, + 415 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 130, + 307, + 482, + 415 + ], + "type": "text", + "content": " is determined through the optimization:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 242, + 422, + 481, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 422, + 481, + 442 + ], + "spans": [ + { + "bbox": [ + 242, + 422, + 481, + 442 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {a r g m i n} \\left| | n - \\Phi_ {\\theta} \\left(\\tilde {z} _ {t}, t, c + \\delta\\right) | | _ {2} ^ {2}, \\right. \\tag {6} \\\\ | | \\delta | | _ {\\infty} \\leq \\epsilon \\\\ \\end{array}", + "image_path": "f5385af1ddcccbe126b7df185cc31064448fad91714cb31211295d158046c150.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\tilde{z} = \\mathcal{E}(\\tilde{x})" + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "content": " denotes the latent representation of image " + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\tilde{x}" + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "content": " is a small number, and " + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "inline_equation", + "content": "c = \\mathcal{E}_{txt}(y)" + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "content": " encodes the textual embedding of the targeted concept, for instance, " + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "inline_equation", + "content": "y =" + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "content": " \"nudity\". The Projected Gradient Descent (PGD) algorithm [26] is employed to address this optimization problem. Specifically, when selecting timestep " + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "inline_equation", + "content": "t^* = 500" + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "content": " as the critical adversarial point, " + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\tilde{z}_t" + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "content": " undergoes denoising via " + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\Phi_{\\theta}(\\tilde{z}_t,t,c)" + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "content": " transitioning from timesteps " + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "inline_equation", + "content": "t = 1000" + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "inline_equation", + "content": "t = 500" + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "content": ". Subsequently, the targeted adversarial approach outlined in Eq. (6) is executed to determine " + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\delta_{t^*}" + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "inline_equation", + "content": "t = 500" + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "content": ". In subsequent denoising steps, " + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\tilde{z}_t" + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "content": " is denoised considering the introduced perturbation " + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "inline_equation", + "content": "c + \\delta_{t^*}" + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 605 + ], + "type": "text", + "content": "Our method introduces a distinctive single-timestep adversarial attack, contrasting with prior approaches that required optimization over a wide range of timesteps [3,58]. This approach enables the seamless incorporation of our attack strategy into the adversarial training process specifically for T2I concept erasure." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": "To assess the efficacy of our method, we execute tests on a " + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\theta}" + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": " model trained for removing the \"nudity\" concept via the ESD method. Utilizing 142 nuditycentric prompts from the I2P dataset [43], we systematically select timesteps " + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "inline_equation", + "content": "t^{*}" + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": " at 100-step intervals within the [1, 1000] range and compute the Attack Success Rate (ASR), as formulated in Eq.(8), for each " + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "inline_equation", + "content": "t^{*}" + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": ". The results, depicted in" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 90, + 280, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 90, + 280, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 90, + 280, + 102 + ], + "type": "text", + "content": "C. Kim*, K. Min*, Y. Yang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "Fig. 2, unveil a notable capability: the reconstruction of the erased concept is feasible even with the attack confined to a single timestep " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "t^*" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ". Interestingly, after " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "t^* = 500" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ", we observe an increasing trend in the ASR. The modified images resulting from the attack visually convey the transition back to the erased concept and reflect the ASR patterns. For additional insights into various concepts and their corresponding ASR trends across different timesteps " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "t^*" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ", the supplementary material offers further information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 213, + 332, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 213, + 332, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 213, + 332, + 224 + ], + "type": "text", + "content": "Adversarial Training for T2I Concept" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 225, + 332, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 225, + 332, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 332, + 307 + ], + "type": "text", + "content": "Erasure. Motivated by the findings from our single-timestep adversarial attack experiments, we explore the potential of such attacks to enhance the robustness of concept erasure in T2I models, posing the question: Can adversarial attacks improve the resilience of concept erasure mechanisms?" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 340, + 248, + 481, + 302 + ], + "blocks": [ + { + "bbox": [ + 339, + 225, + 481, + 247 + ], + "lines": [ + { + "bbox": [ + 339, + 225, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 339, + 225, + 481, + 247 + ], + "type": "text", + "content": "Table 1: Performance comparison of " + }, + { + "bbox": [ + 339, + 225, + 481, + 247 + ], + "type": "inline_equation", + "content": "L_{\\text{erase}}" + }, + { + "bbox": [ + 339, + 225, + 481, + 247 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 339, + 225, + 481, + 247 + ], + "type": "inline_equation", + "content": "L_{\\text{RACE}}" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 340, + 248, + 481, + 302 + ], + "lines": [ + { + "bbox": [ + 340, + 248, + 481, + 302 + ], + "spans": [ + { + "bbox": [ + 340, + 248, + 481, + 302 + ], + "type": "table", + "html": "
LeraseLRACE
I2P [43]0.080.05
FID [13]33.1225.16
CLIP-Score [12]0.7260.745
", + "image_path": "ef7f6312a7d0e849762c9e06bdbb86da6b7bc500a9da64283bf985c0fa5e8b53.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 308, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 380 + ], + "type": "text", + "content": "RACE distinguishes itself from existing approaches [9,21] by aiming to eliminate not only the targeted concept's embedding but also its adjacent embedding within the model's latent space, which could otherwise lead to the inadvertent generation of the erased concept by " + }, + { + "bbox": [ + 130, + 308, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\theta}" + }, + { + "bbox": [ + 130, + 308, + 482, + 380 + ], + "type": "text", + "content": ". We incorporate our adversarial attack into the erasure loss function (" + }, + { + "bbox": [ + 130, + 308, + 482, + 380 + ], + "type": "inline_equation", + "content": "L_{erase}" + }, + { + "bbox": [ + 130, + 308, + 482, + 380 + ], + "type": "text", + "content": "), yielding an enhanced adversarial training loss:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 391, + 481, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 391, + 481, + 405 + ], + "spans": [ + { + "bbox": [ + 143, + 391, + 481, + 405 + ], + "type": "interline_equation", + "content": "L _ {R A C E} := \\left| \\left| \\Phi_ {\\theta} \\left(z _ {t}, t, c + \\delta\\right) - \\left(\\Phi_ {\\theta^ {*}} \\left(z _ {t}, t\\right) - \\eta \\left(\\Phi_ {\\theta^ {*}} \\left(z _ {t}, t, c\\right) - \\Phi_ {\\theta^ {*}} \\left(z _ {t}, t\\right)\\right)\\right) \\right| \\right| _ {2} ^ {2}. \\tag {7}", + "image_path": "d2ca0403aa187e65242e04df03e159a6599da188429683c6f466467f66d9f176.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 414, + 482, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 534 + ], + "type": "text", + "content": "This method is deliberate, substituting the concept embedding " + }, + { + "bbox": [ + 130, + 414, + 482, + 534 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 414, + 482, + 534 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 414, + 482, + 534 + ], + "type": "inline_equation", + "content": "c + \\delta" + }, + { + "bbox": [ + 130, + 414, + 482, + 534 + ], + "type": "text", + "content": " within the trainable parameters of " + }, + { + "bbox": [ + 130, + 414, + 482, + 534 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\theta}" + }, + { + "bbox": [ + 130, + 414, + 482, + 534 + ], + "type": "text", + "content": ". This precise adjustment ensures enhanced fidelity of the generated images by mapping the " + }, + { + "bbox": [ + 130, + 414, + 482, + 534 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 414, + 482, + 534 + ], + "type": "text", + "content": "-neighborhood of the concept embedding to its null representation. Comparative metrics between the direct substitution in " + }, + { + "bbox": [ + 130, + 414, + 482, + 534 + ], + "type": "inline_equation", + "content": "L_{erase}" + }, + { + "bbox": [ + 130, + 414, + 482, + 534 + ], + "type": "text", + "content": " and the strategic use of " + }, + { + "bbox": [ + 130, + 414, + 482, + 534 + ], + "type": "inline_equation", + "content": "L_{RACE}" + }, + { + "bbox": [ + 130, + 414, + 482, + 534 + ], + "type": "text", + "content": " are provided in Tab. 1. The latter approach demonstrates promising reductions in ASR for \"nudity\" prompts within I2P dataset [43] and improvements in image quality metrics, as assessed on the MS-COCO [25]. The RACE methodology is comprehensively detailed in Algorithm 1. To validate RACE's effectiveness, we evaluate the ASR against both white-box and black-box attacks, as elaborated in Sec. 4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 555, + 230, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 230, + 569 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 230, + 569 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 582, + 269, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 582, + 269, + 596 + ], + "spans": [ + { + "bbox": [ + 132, + 582, + 269, + 596 + ], + "type": "text", + "content": "4.1 Experimental Setting" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Datasets. Our assessment of the RACE framework spans various domains, including artistic styles, explicit concepts, and identifiable objects, in line with established benchmarks [3,9,58]. To ensure a uniform image generation process, we standardize key hyperparameters such as the scale of classifier-free guidance and random seeds. Artistic style evaluations leverage shared text prompts from" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 308, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 308, + 91, + 447, + 101 + ], + "type": "text", + "content": "R.A.C.E. for Secure T2I Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "ESD [9] and UnlearnDiff [58]. For explicit content, we draw from the inappropriate image prompt benchmark [43], selecting a diverse set of prompts encompassing 142 nudity, 98 illegal acts, and 101 violence instances. Objects are curated from a subset of Imagenette [16], known for its distinct and recognizable classes, with text prompts synthesized via ChatGPT [31] following the approach by Kumari et al. [21]. It's crucial to note that these datasets serve solely to gauge the effectiveness of adversarial attacks; the adversarial training component of RACE does not necessitate the use of these specific prompts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "text", + "content": "Training Details. In validating RACE, we choose the Erased Stable Diffusion (ESD) model [9] for its ability to erase a wide range of concepts, serving as an ideal testbed to showcase RACE's efficacy. We integrate RACE with ESD, optimizing with attack parameters " + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.1" + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\alpha = \\epsilon /4" + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "text", + "content": ". The optimization of model parameters " + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "text", + "content": " utilizes the Adam optimizer [20] at a learning rate of " + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "inline_equation", + "content": "1e - 5" + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "text", + "content": ", consistent with ESD. Additional details on attack parameter selection are available in the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 319, + 488, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 488, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 488, + 403 + ], + "type": "text", + "content": "Red Teaming Methods. To rigorously test RACE's robustness, we deploy a comprehensive suite of adversarial attacks, spanning both white-box and black-box approaches. Initial assessments utilize the I2P red teaming prompt dataset [43]. In black-box scenarios, we employ PEZ [49], which crafts adversarial prompts via CLIP [35]. In the white-box scenario, methods like P4D [3] and UnlearnDiff [58] are used, which generate adversarial prompts by leveraging gradients from the SD." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 416, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 416, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 416, + 482, + 477 + ], + "type": "text", + "content": "Evaluation. To gauge the robustness of RACE, we employ domain-specific classifiers: a ViT-base model [51] pre-trained on ImageNet [4] and fine-tuned on WikiArt [42] for artistic styles, Nudenet [1] for explicit content, and ResNet-50 [10] trained on ImageNet for object removal. We measure robustness using the Attack Success Rate (ASR):" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 228, + 486, + 481, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 486, + 481, + 519 + ], + "spans": [ + { + "bbox": [ + 228, + 486, + 481, + 519 + ], + "type": "interline_equation", + "content": "\\operatorname {A S R} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathbb {1} \\left(f \\left(S D \\left(\\tilde {y} _ {i}\\right)\\right) = \\tilde {y} _ {i}\\right), \\tag {8}", + "image_path": "390e16d6e7d15a06e0872a09a92e64f2a57889eb13270f4ea81dfb27db27134e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 528, + 482, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 482, + 589 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 528, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\tilde{y}_i" + }, + { + "bbox": [ + 130, + 528, + 482, + 589 + ], + "type": "text", + "content": " is the adversarial prompt, " + }, + { + "bbox": [ + 130, + 528, + 482, + 589 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 528, + 482, + 589 + ], + "type": "text", + "content": " is the classifier, and " + }, + { + "bbox": [ + 130, + 528, + 482, + 589 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 528, + 482, + 589 + ], + "type": "text", + "content": " is the number of prompts. Additionally, we assess image quality after applying RACE by generating 5,000 images from the MS-COCO [25] test set, computing the Frechet Inception Distance (FID) score [13] and the CLIP score [12] to evaluate RACE's impact on image fidelity while ensuring concept erasure." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 607, + 384, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 607, + 384, + 620 + ], + "spans": [ + { + "bbox": [ + 131, + 607, + 384, + 620 + ], + "type": "text", + "content": "4.2 Robust Concept Erase against Red Teaming" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "In our comprehensive analysis, RACE undergoes a series of red teaming evaluations [3, 43, 49, 58], encompassing both white-box and black-box techniques aimed at regenerating concepts targeted by RACE for erasure, as depicted in" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 280, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 280, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 280, + 102 + ], + "type": "text", + "content": "C. Kim*, K. Min*, Y. Yang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 201, + 481, + 356 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "type": "text", + "content": "Table 2: Attack Success Rate (ASR) against white " + }, + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "type": "inline_equation", + "content": "(\\bigcirc)" + }, + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "type": "text", + "content": "/black " + }, + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "type": "inline_equation", + "content": "(\\bullet)" + }, + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "type": "text", + "content": " box attacks. We conduct experiments on artistic style (Van Gogh), explicit concepts (nudity, violence, and illegal acts), and objects (church, golf ball, and parachute). \"ESD/RACE-Concept\" denotes the concept erased from the model. We also measure the quality of T2I models. We can observe that RACE reduces ASR by over " + }, + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "type": "text", + "content": " (-0.30 or below) in absolute value on Van Gogh, nudity, and church for UnlearnDiff [58], which is the previous SOTA attack method." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 201, + 481, + 356 + ], + "lines": [ + { + "bbox": [ + 132, + 201, + 481, + 356 + ], + "spans": [ + { + "bbox": [ + 132, + 201, + 481, + 356 + ], + "type": "table", + "html": "
PromptsPEZ [49]P4D [3]UnlearnDiff [58]CLIP-Score [12]FID [13]
White/Black Box●●○○--
ESD [9]-VanGogh0.040.000.260.360.799719.16
ESD [9]-Nudity0.140.080.750.800.793118.88
ESD [9]-Violence0.270.130.840.790.783421.55
ESD [9]-Illegal0.290.200.890.850.785421.50
ESD [9]-Church0.160.000.580.680.789619.68
ESD [9]-GolfBall0.040.000.160.160.773820.64
ESD [9]-Parachute0.060.040.480.600.786519.72
RACE-VanGogh0.00 (-0.04)0.00 (-0.00)0.00 (-0.26)0.04 (-0.32)0.802420.65
RACE-Nudity0.05 (-0.09)0.02 (-0.06)0.49 (-0.26)0.47 (-0.33)0.745225.16
RACE-Violence0.11 (-0.16)0.08 (-0.05)0.75 (-0.09)0.68 (-0.11)0.737428.71
RACE-Illlegal0.20 (-0.09)0.13 (-0.07)0.85 (-0.04)0.80 (-0.05)0.759124.87
RACE-Church0.02 (-0.14)0.00 (-0.00)0.26 (-0.32)0.38 (-0.30)0.773023.92
RACE-GolfBall0.00 (-0.04)0.00 (-0.00)0.10 (-0.06)0.06 (-0.10)0.748025.38
RACE-Parachute0.02 (-0.04)0.00 (-0.04)0.24 (-0.24)0.38 (-0.22)0.757026.42
", + "image_path": "d6f45bec7d9a4297ef2a3d4b8324e15041e8849f1ab497415b676c08a2e93b15.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "text", + "content": "Fig.1. The comparative analysis of ASR presented in Tab.2 spans diverse conceptual domains, from \"Van Gogh\"-inspired artistry to explicit content such as \"nudity\" and \"violence\", extending to tangible objects like \"churches\", \"golf balls\", and \"parachutes\". Remarkably, RACE consistently diminishes ASR, notably surpassing " + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "text", + "content": " for \"Van Gogh\" styles, \"nudity\", and \"church\" categories, particularly outperforming UnlearnDiff [58], the current state-of-the-art in white-box adversarial methodologies. This marked decline in ASR underscores RACE's heightened robustness and delineates its capability as a potent, computationally efficient defense mechanism for T2I diffusion models against intricate adversarial attacks. Crucially, RACE's methodological advantage stems from its independence from external imagery or prompts, diverging from traditional red teaming techniques reliant on such data for prompt generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 510, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 640 + ], + "type": "text", + "content": "As demonstrated in Tab.2, RACE achieves a significant " + }, + { + "bbox": [ + 130, + 510, + 482, + 640 + ], + "type": "inline_equation", + "content": "33\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 640 + ], + "type": "text", + "content": " reduction in ASR for the \"nudity\" concept, underscoring its effectiveness. To further elucidate how RACE enhances ASR, we analyze the specific categories or elements it targets for removal. Utilizing Nudenet, we enumerate the body parts generated by various models—original SD, ESD, UnlearnDiff, and RACE—when prompted with nudity-related inputs from the I2P dataset. Illustrated in Fig.3, the analysis reveals that while ESD enhances the original SD's resilience to such prompts, UnlearnDiff manages to bypass ESD's defenses, reconstructing explicit content. In contrast, RACE maintains its robustness even against the sophisticated UnlearnDiff attacks, showcasing the advanced protective capabilities of our approach in safeguarding against the regeneration of sensitive content." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "One caveat to mention is that our experiments reveal a nuanced trade-off between robustness and image quality. While artistic style erasures maintain" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 308, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 308, + 91, + 447, + 101 + ], + "type": "text", + "content": "R.A.C.E. for Secure T2I Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 115, + 479, + 254 + ], + "blocks": [ + { + "bbox": [ + 135, + 115, + 479, + 254 + ], + "lines": [ + { + "bbox": [ + 135, + 115, + 479, + 254 + ], + "spans": [ + { + "bbox": [ + 135, + 115, + 479, + 254 + ], + "type": "image", + "image_path": "8109cb659eb06404b45ab762e449479b0fe4ce0bd162f8321df64716e87e71a3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "type": "text", + "content": "Fig. 3: Although ESD significantly reduces the chance of generating images with exposed body parts, state-of-the-art red teaming methods, such as UnlearnDiff, can be used to bypass ESD's defense and reconstruct explicit content. RACE and its variant can effectively defend the malicious attempts to reconstruct explicit content from the ESD model that erased the concept of nudity." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 327, + 482, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 508 + ], + "type": "text", + "content": "quality metrics, erasures of other concepts inadvertently degrade image quality. This divergence could be attributed to methodological differences; erasing artistic styles predominantly involves fine-tuning the cross-attention layers of SD as per ESD guidelines, whereas erasing other concepts necessitates adjustments in non-cross-attention layers [9]. Another plausible explanation for the observed trade-off between robust concept erasure and overall image quality could relate to the inherent complexity of differentiating between closely related or overlapping concepts within the model's latent space. As RACE intensifies adversarial robustness, it may inadvertently alter the delicate equilibrium within these conceptual overlaps, leading to unintended modifications in adjacent, non-targeted conceptual representations. This issue highlights the complex tension between precise concept erasure and maintaining the model's overall integrity against attacks. To address the quality concerns arising from this trade-off, we explore a potential strategy for improvement in Sec. 4.4, aiming to reduce the trade-off between targeted erasure with high quality and the model's defensive robustness." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 524, + 246, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 524, + 246, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 524, + 246, + 536 + ], + "type": "text", + "content": "4.3 Disentanglement" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 544, + 482, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 664 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 664 + ], + "type": "text", + "content": "Investigating RACE's disentanglement performance, our study focuses on its capability to precisely erase intended concepts without impacting other elements. The evaluation spans both qualitative and quantitative measures. On the qualitative front, we configure separate RACE-enhanced models to specifically erase artistic imprints such as \"Van Gogh\", \"Thomas Kinkade\", and \"Kilian Eng\" from the SD. The illustrative outcomes, showcased in Fig. 4, underscore RACE's erasure precision, ensuring that the excision of one artistic style doesn't lead to the collateral removal of others. This meticulous erasure extends to discrete object concepts like \"church\", \"golf ball\", and \"parachute\", with the generated images further affirming the method's discernment." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 280, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 280, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 280, + 102 + ], + "type": "text", + "content": "C. Kim*, K. Min*, Y. Yang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 115, + 482, + 243 + ], + "blocks": [ + { + "bbox": [ + 136, + 115, + 482, + 243 + ], + "lines": [ + { + "bbox": [ + 136, + 115, + 482, + 243 + ], + "spans": [ + { + "bbox": [ + 136, + 115, + 482, + 243 + ], + "type": "image", + "image_path": "21077ca7c798cbb9362adaf70208ccc27f451b96eaa0539861ebd4937fecec10.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 251, + 480, + 297 + ], + "lines": [ + { + "bbox": [ + 131, + 251, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 131, + 251, + 480, + 297 + ], + "type": "text", + "content": "Fig. 4: RACE's Disentanglement in Concept Erasure. This figure highlights RACE's precision in erasing specific concepts, as shown in diagonal images, while preserving unrelated concepts, which is evident in off-diagonal images. For reference, baseline images generated by the original Stable Diffusion (SD) model are also presented." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 132, + 347, + 481, + 407 + ], + "blocks": [ + { + "bbox": [ + 131, + 304, + 479, + 338 + ], + "lines": [ + { + "bbox": [ + 131, + 304, + 479, + 338 + ], + "spans": [ + { + "bbox": [ + 131, + 304, + 479, + 338 + ], + "type": "text", + "content": "Table 4: Ablation Studies for Performance Improvement. Using the \"nudity\" concept, we evaluate the effectiveness of adding weight regularization and close concept keywords. ASR and quality metrics are measured to assess performance." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 347, + 481, + 407 + ], + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 407 + ], + "type": "table", + "html": "
I2P [43]PEZ [49]P4D [3]UnleranDiff [58]CLIP-Score [12]FID [13]
ESD0.140.080.750.800.793118.88
RACE0.050.020.490.470.745225.16
RACE+Reg.0.070.020.600.620.759324.42
RACE+keywords0.020.010.420.460.720130.97
", + "image_path": "e0c3753aefc178cf1d9f91f982a0d6e813757c418db372e57569ed1a038f9fc3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 428, + 312, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 428, + 312, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 428, + 312, + 559 + ], + "type": "text", + "content": "Quantitatively, we generate 5,000 prompts with varied random seeds, such as \"an image of a [class name]\", to produce a diverse set of images, subsequently evaluated using a pre-trained ResNet-50 classifier for top-1 accuracy. As shown in Tab. 3, RACE showcases an improved capability for object concept erasure compared to ESD. Despite this, there exists a slight decrement in classification accuracy for non-target classes. This effect" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 559, + 479, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 559, + 479, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 559, + 479, + 608 + ], + "type": "text", + "content": "likely stems from RACE's method of targeting the " + }, + { + "bbox": [ + 130, + 559, + 479, + 608 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 559, + 479, + 608 + ], + "type": "text", + "content": "-neighborhood surrounding the intended concept, potentially influencing proximate concepts. Nevertheless, we can observe from Fig. 4 that RACE can precisely erase the target concept with minimal visual impacts on other concepts." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 319, + 483, + 481, + 551 + ], + "blocks": [ + { + "bbox": [ + 318, + 427, + 482, + 482 + ], + "lines": [ + { + "bbox": [ + 318, + 427, + 482, + 482 + ], + "spans": [ + { + "bbox": [ + 318, + 427, + 482, + 482 + ], + "type": "text", + "content": "Table 3: Accuracy of erased and non-erased classes in Imagenet [4]. We evaluate the classification accuracy of ESD and RACE for the target erased concepts and other non-target concepts." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 483, + 481, + 551 + ], + "lines": [ + { + "bbox": [ + 319, + 483, + 481, + 551 + ], + "spans": [ + { + "bbox": [ + 319, + 483, + 481, + 551 + ], + "type": "table", + "html": "
Erased ConceptAcc. ErasedAcc. Others
ESDRACEESDRACE
Church0.160.020.570.53
Golf Ball0.040.000.450.35
Parachute0.060.020.570.41
", + "image_path": "ed72e5d29478760d90c0002d54b14515404616f7b8a3647bc40b24120586944e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 623, + 214, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 623, + 214, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 623, + 214, + 635 + ], + "type": "text", + "content": "4.4 Discussion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "Potential Strategy to Improve the Robustness-Quality Trade-off. Our findings underscore the ability of RACE to significantly bolster the SD model's" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 309, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 309, + 91, + 447, + 101 + ], + "type": "text", + "content": "R.A.C.E. for Secure T2I Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "defense against prompt-based adversarial attacks across a variety of concepts. However, as illustrated in Tab.2, enhancing robustness appears to inversely impact image quality. To address this dichotomy, we test a refined version of the RACE loss function incorporating a regularization term:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 224, + 170, + 481, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 170, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 224, + 170, + 481, + 182 + ], + "type": "interline_equation", + "content": "L _ {R A C E + R e g.} := L _ {R A C E} + \\lambda \\| \\theta - \\theta^ {*} \\| _ {1}, \\tag {9}", + "image_path": "775c84421a2739f5fd71df18b81fbff75498c5c7f6d45fd625b5a4c5ede4fc42.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 186, + 479, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 186, + 479, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 186, + 479, + 246 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 186, + 479, + 246 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 186, + 479, + 246 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 186, + 479, + 246 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 130, + 186, + 479, + 246 + ], + "type": "text", + "content": " represent the parameters of the RACE and original Stable Diffusion models, respectively, and " + }, + { + "bbox": [ + 130, + 186, + 479, + 246 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 186, + 479, + 246 + ], + "type": "text", + "content": " is the regularization strength, set to 0.1 in our experiment. This regularization approach, as evidenced in Tab.4, helps to partially reconcile the robustness-quality trade-off, enhancing image quality while maintaining improved robustness over ESD." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 258, + 482, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 462 + ], + "type": "text", + "content": "Enhancing Concept Erasure. In pursuit of further reducing the ASR, we explore strategies for more comprehensive concept erasure. Recognizing that a target concept may manifest in various synonymous forms, we extend RACE's erasure scope to include semantically related concepts. Leveraging the CLIP text encoder embedded within Stable Diffusion, we identify and subsequently erase concepts closely related to the target, based on their proximity in the CLIP embedding space. For instance, alongside \"nudity\", we also target synonymous concepts like \"nude\", \"nsfw\", and \"bare\", identified as the top-3 semantically similar terms. In Fig. 3, we can observe that our method equipped with this expanded erasure strategy (denoted as RACE+keywords) is more effective in defending the malicious attempts to bypass the ESD by further reducing the number of exposed body parts. Tab. 4 also indicates that our strategy indeed fortifies the model's robustness against red teaming tactics. Nonetheless, this broadened concept removal spectrum reaffirms the robustness-quality trade-off, manifesting as a decrement in image quality. This aspect opens an intriguing avenue for future enhancements to the RACE methodology, balancing the twin objectives of robust concept erasure and preserved image fidelity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 476, + 220, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 476, + 220, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 220, + 489 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "content": "In this work, we present RACE, a novel defense approach designed to protect the Text-to-Image Stable Diffusion models from prompt-based red teaming attacks. RACE effectively strengthens the model's concept erasure capabilities while maintaining computational efficiency, offering a valuable enhancement to the current erasure framework and bolstering defenses against various adversarial techniques. We also observe the robustness-quality trade-off and discuss possible future directions to improve it. This initial contribution lays the groundwork for further exploration, underscoring the critical importance of developing sophisticated defenses in the rapidly evolving domain of generative AI." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Acknowledgements. We would like to express our gratitude to Sangmin Jung for assisting with the experiments. This work is partially supported by the National Science Foundation under Grant No. 2038666 and No. 2101052. The views and opinions of the authors expressed herein do not necessarily state or reflect those of the funding agencies and employers." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 90, + 280, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 90, + 280, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 90, + 280, + 102 + ], + "type": "text", + "content": "C. Kim*, K. Min*, Y. Yang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 142, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 142, + 480, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 142, + 480, + 164 + ], + "spans": [ + { + "bbox": [ + 138, + 142, + 480, + 164 + ], + "type": "text", + "content": "1. Bedapudi, P.: Nudenet: Neural nets for nudity classification, detection and selective censoring (12 2019) 10" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 165, + 480, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 165, + 480, + 208 + ], + "spans": [ + { + "bbox": [ + 138, + 165, + 480, + 208 + ], + "type": "text", + "content": "2. Carlini, N., Hayes, J., Nasr, M., Jagielski, M., Sehwag, V., Tramér, F., Balle, B., Ippolito, D., Wallace, E.: Extracting training data from diffusion models. ArXiv abs/2301.13188 (2023), https://api-semanticscholar.org/CorpusID: 256389993 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 209, + 480, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 209, + 480, + 243 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 480, + 243 + ], + "type": "text", + "content": "3. Chin, Z.Y., Jiang, C.M., Huang, C.C., Chen, P.Y., Chiu, W.C.: Prompting4debugging: Red-teaming text-to-image diffusion models by finding problematic prompts. arXiv preprint arXiv:2309.06135 (2023) 3, 4, 5, 7, 8, 9, 10, 11, 13" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 243, + 480, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 243, + 480, + 276 + ], + "spans": [ + { + "bbox": [ + 138, + 243, + 480, + 276 + ], + "type": "text", + "content": "4. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) 10, 13" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 276, + 480, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 276, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 276, + 480, + 308 + ], + "type": "text", + "content": "5. Devlin, K., Cheetham, J.: Fake trump arrest photos: How to spot an ai-generated image. BBC News (March 2023), https://www.bbc.com/news/technology-68981525 1, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 309, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 309, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 309, + 480, + 342 + ], + "type": "text", + "content": "6. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12873-12883 (2021) 4, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 343, + 480, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 480, + 376 + ], + "type": "text", + "content": "7. Fernandez, P., Couairon, G., Jégou, H., Douze, M., Furon, T.: The stable signature: Rooting watermarks in latent diffusion models. arXiv preprint arXiv:2303.15435 (2023) 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 376, + 480, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 480, + 420 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 480, + 420 + ], + "type": "text", + "content": "8. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. ArXiv abs/2208.01618 (2022), https://apisemantic scholar.org/CorpusID:251253049 4, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 420, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 420, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 420, + 480, + 464 + ], + "type": "text", + "content": "9. Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., Bau, D.: Erasing concepts from diffusion models. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 2426-2436 (2023), https://apisemantic scholar.org/CorpusID:257495777 2, 3, 4, 7, 9, 10, 11, 12" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 465, + 480, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 465, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 465, + 480, + 498 + ], + "type": "text", + "content": "0. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 10" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 498, + 480, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 498, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 138, + 498, + 480, + 531 + ], + "type": "text", + "content": "1. Heng, A., Soh, H.: Selective amnesia: A continual learning approach to forgetting in deep generative models. ArXiv abs/2305.10120 (2023), https://apisemantic scholar.org/CorpusID:258740988 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 532, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 532, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 138, + 532, + 480, + 564 + ], + "type": "text", + "content": "2. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021) 9, 10, 11, 13" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 565, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 565, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 138, + 565, + 480, + 598 + ], + "type": "text", + "content": "3. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Advances in Neural Information Processing Systems. pp. 6626-6637 (2017) 9, 10, 11, 13" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "type": "text", + "content": "4. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020) 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "text", + "content": "5. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022) 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "6. Howard, J., Gugger, S.: fastai: A layered api for deep learning. Inf. 11, 108 (2020), https://api-semanticscholar.org/CorpusID:211082837 10" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 308, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 308, + 91, + 447, + 101 + ], + "type": "text", + "content": "R.A.C.E. for Secure T2I Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "17. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8110-8119 (2020) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 151, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 151, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 151, + 481, + 183 + ], + "type": "text", + "content": "18. Kim, C., Min, K., Patel, M., Cheng, S., Yang, Y.: Wouaf: Weight modulation for user attribution and fingerprinting in text-to-image diffusion models. arXiv preprint arXiv:2306.04744 (2023) 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 481, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 481, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 481, + 206 + ], + "type": "text", + "content": "19. Kim, C., Ren, Y., Yang, Y.: Decentralized attribution of generative models. In: International Conference on Learning Representations (2021) 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 207, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 481, + 251 + ], + "type": "text", + "content": "20. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization (2014), http://arxiv.org/abs/1412.6980, cite arxiv:1412.6980Comment: Published as a conference paper at the 3rd International Conference for Learning Representations, San Diego, 2015 10" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 251, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 481, + 296 + ], + "type": "text", + "content": "21. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating concepts in text-to-image diffusion models. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 22634-22645 (2023), https://apisemanticscholar.org/CorpusID:257687839 3, 4, 9, 10" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 297, + 481, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 481, + 341 + ], + "type": "text", + "content": "22. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 1931-1941 (2022), https://api-semanticscholar.org/CorpusID:254408780 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 341, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 481, + 373 + ], + "type": "text", + "content": "23. Li, A.C., Prabhudesai, M., Duggal, S., Brown, E., Pathak, D.: Your diffusion model is secretly a zero-shot classifier. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2206-2217 (October 2023) 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 374, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 418 + ], + "type": "text", + "content": "24. Li, S., van de Weijer, J., Hu, T., Khan, F.S., Hou, Q., Wang, Y., Yang, J.: Get what you want, not what you don't: Image content suppression for text-to-image diffusion models. ArXiv abs/2402.05375 (2024), https://api.sementicscholar.org/CorpusID:267547985 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 418, + 481, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 463 + ], + "type": "text", + "content": "25. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) 9, 10" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 464, + 481, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 481, + 495 + ], + "type": "text", + "content": "26. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083 (2017) 3, 5, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 497, + 481, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 497, + 481, + 540 + ], + "spans": [ + { + "bbox": [ + 132, + 497, + 481, + 540 + ], + "type": "text", + "content": "27. Marcelo, P.: Fact focus: Fake image of pentagon explosion briefly sends jitters through stock market. Associated Press (May 2023), https://apnews.com/article/pentagon-explosion-fake-image-stock-market-jitters-78d5913603bdf9f17cc7a8c77a72e59b 1, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 542, + 481, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 542, + 481, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 542, + 481, + 575 + ], + "type": "text", + "content": "28. Ni, M., Wu, C., Wang, X., Yin, S.S., Wang, L., Liu, Z., Duan, N.: Ores: Open-vocabulary responsible visual synthesis. ArXiv abs/2308.13785 (2023), https://api-semanticscholar.org/CorpusID:261243073 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 576, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 481, + 620 + ], + "type": "text", + "content": "29. Nichol, A.Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., Mcgrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In: International Conference on Machine Learning. pp. 16784-16804. PMLR (2022) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 620, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 481, + 643 + ], + "type": "text", + "content": "30. Nie, G., Kim, C., Yang, Y., Ren, Y.: Attributing image generative models using latent fingerprints. arXiv preprint arXiv:2304.09752 (2023) 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "31. OpenAI: Chatgpt. Online (2022), https://chat.openai.com/chat, accessed on February 24, 2024 10" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 280, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 280, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 280, + 102 + ], + "type": "text", + "content": "C. Kim*, K. Min*, Y. Yang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "32. Patel, M., Jung, S., Baral, C., Yang, Y.: " + }, + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "-eclipse: Multi-concept personalized text-to-image diffusion models by leveraging clip latent space. ArXiv abs/2402.05195 (2024), https://api_semanticscholar.org/CorpusID:267547418 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 181 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 181 + ], + "type": "text", + "content": "33. Patel, M., Kim, C.S., Cheng, S., Baral, C., Yang, Y.: Eclipse: A resource-efficient text-to-image prior for image generations. ArXiv abs/2312.04655 (2023), https://api-semanticscholar.org/CorpusID:266149498 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 181, + 482, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 181, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 181, + 482, + 224 + ], + "type": "text", + "content": "34. Pham, M., Marshall, K.O., Cohen, N., Mittal, G., Hegde, C.: Circumventing concept erasure methods for text-to-image generative models. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=ag3o2T51Ht 4, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 224, + 482, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 267 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 267 + ], + "type": "text", + "content": "35. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 6, 10" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 267, + 482, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 482, + 300 + ], + "type": "text", + "content": "36. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022) 1, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 300, + 482, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 320 + ], + "type": "text", + "content": "37. Rando, J., Paleka, D., Lindner, D., Heim, L., Tramér, F.: Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610 (2022) 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 320, + 482, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 482, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 482, + 363 + ], + "type": "text", + "content": "38. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684-10695 (June 2022) 1, 4, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 363, + 482, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 407 + ], + "type": "text", + "content": "39. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015) 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 407, + 482, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 460 + ], + "type": "text", + "content": "40. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 22500-22510 (2022), https://api_semanticscholar.org/CorpusID:251800180 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 460, + 482, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 503 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 503 + ], + "type": "text", + "content": "41. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 503, + 482, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 482, + 536 + ], + "type": "text", + "content": "42. Saleh, B., Elgammal, A.: Large-scale classification of fine-art paintings: Learning the right metric on the right feature. ArXiv abs/1505.00855 (2015), https://api-semanticscholar.org/CorpusID:14168099 10" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 536, + 482, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 482, + 579 + ], + "type": "text", + "content": "43. Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 22522-22531 (2022), https://apisemanticscholar.org/CorpusID:253420366 4, 8, 9, 10, 13" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 579, + 482, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 579, + 482, + 622 + ], + "spans": [ + { + "bbox": [ + 130, + 579, + 482, + 622 + ], + "type": "text", + "content": "44. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402 (2022) 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 622, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 622, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 622, + 482, + 665 + ], + "type": "text", + "content": "45. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Diffusion art or digital forgery? investigating data replication in diffusion models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 6048-6058 (2022), https://apisemanticscholar.org/CorpusID:2543666341" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 308, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 308, + 91, + 447, + 101 + ], + "type": "text", + "content": "R.A.C.E. for Secure T2I Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 632 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "46. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Understanding and mitigating copying in diffusion models. ArXiv abs/2305.20086 (2023), https://api.sementicscholar.org/CorpusID:258987384 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "text", + "content": "47. Tsai, Y.L., Hsu, C.Y., Xie, C., Lin, C.H., Chen, J.Y., Li, B., Chen, P.Y., Yu, C.M., ying Huang, C.: Ring-a-bell! how reliable are concept removal methods for diffusion models? ArXiv abs/2310.10012 (2023), https://api_semanticscholar.org/CorpusID:264146485 3, 4, 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 237 + ], + "type": "text", + "content": "48. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 15897-15907 (2023), https://apisemantic scholar.org/CorpusID:257219968 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 238, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 281 + ], + "type": "text", + "content": "49. Wen, Y., Jain, N., Kirchenbauer, J., Goldblum, M., Geiping, J., Goldstein, T.: Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery. Advances in Neural Information Processing Systems 36 (2024) 3, 4, 5, 7, 10, 11, 13" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 282, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 314 + ], + "type": "text", + "content": "50. Wen, Y., Kirchenbauer, J., Geiping, J., Goldstein, T.: Tree-ring watermarks: Fingerprints for diffusion images that are invisible and robust. arXiv preprint arXiv:2305.20030 (2023) 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 315, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 358 + ], + "type": "text", + "content": "51. Wu, B., Xu, C., Dai, X., Wan, A., Zhang, P., Tomizuka, M., Keutzer, K., Vajda, P.: Visual transformers: Token-based image representation and processing for computer vision. ArXiv abs/2006.03677 (2020), https://api_semanticscholar.org/CorpusID:219531480 10" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "text", + "content": "52. Wu, Y., Zhang, J., Kerschbaum, F., Zhang, T.: Backdooring textual inversion for concept censorship. ArXiv abs/2308.10718 (2023), https://apisemantic scholar.org/CorpusID:261049298 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "type": "text", + "content": "53. Yang, Y., Gao, R., Wang, X., Xu, N., Xu, Q.: Mma-diffusion: Multimodal attack on diffusion models. ArXiv abs/2311.17516 (2023), https://apisemanticscholar.org/CorpusID:265498727 3, 4, 5, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 424, + 481, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 465 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 465 + ], + "type": "text", + "content": "54. Yu, N., Skripniuk, V., Abdelnabi, S., Fritz, M.: Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In: Proceedings of the IEEE/CVF International conference on computer vision. pp. 14448-14457 (2021) 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "type": "text", + "content": "55. Yu, N., Skripniuk, V., Chen, D., Davis, L., Fritz, M.: Responsible disclosure of generative models using scalable fingerprinting. arXiv preprint arXiv:2012.08726 (2020) 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "text", + "content": "56. Zhang, E., Wang, K., Xu, X., Wang, Z., Shi, H.: Forget-me-not: Learning to forget in text-to-image diffusion models. ArXiv abs/2303.17591 (2023), https://api.sementicscholar.org/CorpusID:257833863 3, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "type": "text", + "content": "57. Zhang, K.A., Xu, L., Cuesta-Infante, A., Veeramachaneni, K.: Robust invisible video watermarking with attention (2019) 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 555, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 599 + ], + "type": "text", + "content": "58. Zhang, Y., Jia, J., Chen, X., Chen, A., Zhang, Y., Liu, J., Ding, K., Liu, S.: To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images ... for now. ArXiv abs/2310.11868 (2023), https://api.sementicscholar.org/CorpusID:264289091 2, 3, 4, 5, 7, 8, 9, 10, 11, 13" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "text", + "content": "59. Zheng, Y., Yeh, R.A.: Imma: Immunizing text-to-image models against malicious adaptation. ArXiv abs/2311.18815 (2023), https://api_semanticscholar.org/CorpusID:265506125 4" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 280, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 280, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 280, + 102 + ], + "type": "text", + "content": "C. Kim*, K. Min*, Y. Yang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/R3D-AD_ Reconstruction via Diffusion for 3D Anomaly Detection/a722039e-753c-4cd1-9dab-7069c8f25f9e_content_list.json b/2024/R3D-AD_ Reconstruction via Diffusion for 3D Anomaly Detection/a722039e-753c-4cd1-9dab-7069c8f25f9e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..52340e3cf0b7f6a48a67c11e86c2d441d94a354f --- /dev/null +++ b/2024/R3D-AD_ Reconstruction via Diffusion for 3D Anomaly Detection/a722039e-753c-4cd1-9dab-7069c8f25f9e_content_list.json @@ -0,0 +1,1694 @@ +[ + { + "type": "text", + "text": "R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection", + "text_level": 1, + "bbox": [ + 240, + 140, + 763, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zheyuan Zhou $^{1*}$ , Le Wang $^{1*}$ , Naiyu Fang $^{1,2}$ , Zili Wang $^{1}$ , Lemiao Qiu $^{1\\boxtimes}$ , and Shuyou Zhang $^{1}$", + "bbox": [ + 321, + 210, + 676, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ State Key Laboratory of Fluid Power & Mechatronic System, Zhejiang University, Hangzhou, China", + "bbox": [ + 285, + 253, + 712, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ S-Lab, Nanyang Technological University, Singapore, Singapore", + "bbox": [ + 281, + 282, + 717, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://zhouzheyuan.github.io/r3d-ad", + "bbox": [ + 357, + 310, + 640, + 323 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. 3D anomaly detection plays a crucial role in monitoring parts for localized inherent defects in precision manufacturing. Embedding-based and reconstruction-based approaches are among the most popular and successful methods. However, there are two major challenges to the practical application of the current approaches: 1) the embedded models suffer the prohibitive computational and storage due to the memory bank structure; 2) the reconstructive models based on the MAE mechanism fail to detect anomalies in the unmasked regions. In this paper, we propose R3D-AD, reconstructing anomalous point clouds by diffusion model for precise 3D anomaly detection. Our approach capitalizes on the data distribution conversion of the diffusion process to entirely obscure the input's anomalous geometry. It step-wisely learns a strict point-level displacement behavior, which methodically corrects the aberrant points. To increase the generalization of the model, we further present a novel 3D anomaly simulation strategy named Patch-Gen to generate realistic and diverse defect shapes, which narrows the domain gap between training and testing. Our R3D-AD ensures a uniform spatial transformation, which allows straightforwardly generating anomaly results by distance comparison. Extensive experiments show that our R3D-AD outperforms previous state-of-the-art methods, achieving $73.4\\%$ Image-level AUROC on the Real3D-AD dataset and $74.9\\%$ Image-level AUROC on the Anomaly-ShapeNet dataset with an exceptional efficiency.", + "bbox": [ + 261, + 357, + 738, + 661 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D anomaly detection, industrial applications, 3D reconstruction, self-supervised learning", + "bbox": [ + 259, + 674, + 740, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 726, + 375, + 742 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anomaly detection aims to identify instances containing anomalies and to precisely locate the specific positions of defects. This task is extensively applied across multiple fields and plays a crucial role in quality control within industrial production [28]. 3D anomaly detection [19] has emerged due to its intrinsic", + "bbox": [ + 212, + 756, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution.", + "bbox": [ + 217, + 825, + 364, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3f369856b866b0810a85c9855efae2baf59879e5aeab2e0a8a889bc771d06180.jpg", + "image_caption": [ + "(a) Embedded model" + ], + "image_footnote": [], + "bbox": [ + 282, + 148, + 483, + 243 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0fd877a4c795486bb0d559134332b65e7c05ea94fdd0ea8529ca7ec254083f45.jpg", + "image_caption": [ + "(b) Reconstructive model", + "Fig. 1: Comparison of architectures. (a) Embedded model encodes the input $\\mathcal{X}$ into features and stores them in the memory bank during training. The anomaly map $\\mathcal{M}$ is obtained by comparing the test features with all the features in the memory bank. (b) Reconstructive model is trained by minimizing the loss between its input $\\mathcal{X}$ and the output $\\widehat{\\mathcal{X}}$ . The anomaly map $\\mathcal{M}$ is obtained by comparing the test phase input with its corresponding reconstruction target." + ], + "image_footnote": [], + "bbox": [ + 537, + 148, + 717, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "modality superior for avoiding blind spots in advanced processing and precision manufacturing. However, the discrete and disordered data form of point clouds makes it more difficult to acquire features compared to images. With the scarcity of anomalies, 3D anomaly detection also faces the problem of domain shift while only normal data are presented during training. The presence of these issues underscores the necessity and urgency of devising an efficient framework for the 3D anomaly detection task.", + "bbox": [ + 212, + 393, + 787, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Similar to traditional 2D anomaly detection [28,42], current 3D anomaly detection can be primarily categorized into embedding-based and reconstruction-based, as illustrated in Fig. 1. The embedding-based methods involve mapping features extracted with a pre-trained encoder onto a normal distribution for learning. Distributions that do not fall within the interval are classified as anomalies. Most existing 3D anomaly detection methods are based on a memory bank mechanism [3, 11, 19, 36], which stores some representative features during the training phase to implicitly construct a feature distribution. In the testing phase, the presence of anomalies is determined by calculating the Euclidean distance between the input test object and all template point clouds stored in memory. The reconstruction-based methods train a network capable of accurately reconstructing normal point clouds, under the presumption that anomalous point clouds will not be effectively reconstructed since they are not included during training. The anomaly map is produced through the comparison of discrepancies between the input point cloud and its reconstruction. IMRNet [18] employs PointMAE [25] to reconstruct the input in several iterations, getting the final anomaly map by calculating the explicit spatial coordinate differences and implicit deep feature differences of the point cloud, respectively.", + "bbox": [ + 212, + 503, + 787, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, existing methods face two key issues, high resource cost and irreparable reconstruction. Firstly, methods based on the memory bank [3,11,19, 36] store all features from the training phase, each test point cloud needs to be compared with all samples in the memory bank, significantly increasing memory", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z. Zhou et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "overhead and inference time costs. This makes such methods almost inapplicable in real industrial production lines due to their inefficiency. Secondly, masked autoencoder (MAE) mechanism [9,25,38] only reconstructs the masked portions of the input, defects within unmasked portions may be preserved. This contradicts the fundamental assumption of detecting anomalies by comparing the original defect-containing point cloud with a reconstructed anomaly-free version. These methods inevitably lead to incorrect reconstructions, undermining their effectiveness in accurately localizing defects.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose R3D-AD, a novel 3D anomaly detection method that does not suffer from the space burden and time endurance in memory-based embedded models nor the anomaly unmasking probability in the MAE-based reconstructive models. In contrast to PointMAE, one of our key insights is to perform undifferentiated masking for 3D objects via the noise diffusion mechanism, which maximizes the preservation of anomaly-free shapes and reconstructs abnormal regions. In the reparameterized diffusion process, one-step full mask and reconstruction are achieved by converting the point cloud distribution, instead of the multiple iterative method [18]. We hypothesize that anomaly detection verifies the gap between the reconstructed shapes and the positive samples by learning point movement. Specifically, for input models with arbitrary anomalies, we encode them as latent shape embeddings as decoding conditions and explicitly control the point cloud reconstruction process by step-wise displacements (SWD) decoding. The shape embedding harbors abundant global features and makes it easier to train the network without dwelling on the introduction of local anomaly details. Another key to our approach is to implement a controllable method of point-wise displacement during the diffusion process to refine the point cloud deformation iteratively. We propose to inject latent shape embedding into each step of the inverse denoising process, which drives the anomalous regions to converge to a smooth surface. We further adopted a 3D anomaly simulation strategy Patch-Gen to address the limitations of the dataset, which generates abundant defectives by producing spatial irregularity that is faithful to the real scene, including bulges, sinks, etc. This point cloud data augmentation encourages the self-supervised model to reconstruct more realistic anomaly-free shapes when facing the actual anomaly.", + "bbox": [ + 212, + 281, + 787, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To the best of our knowledge, this is the very first attempt at exploring diffusion in reconstruction-based 3D anomaly detection. Our main contributions are summarized as follows: (i) We introduce a novel framework, termed R3D-AD, which performs a one-step full mask and anomaly-free reconstruction for fast and accurate 3D anomaly detection. (ii) We propose to learn the step-wise displacement in the reverse diffusion process to explicitly control the reconstruction of anomalous shapes. (iii) We introduce a 3D anomaly simulation strategy named Patch-Gen to address the limitation of the data anomaly patterns and improve the reconstruction performance in a supervised setting. (iv) Extensive experiments demonstrate that our R3D-AD has achieved state-of-the-art performance on both Real3D-AD and Anomaly-ShapeNet datasets.", + "bbox": [ + 212, + 672, + 789, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection", + "bbox": [ + 292, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 215, + 143, + 382, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 2D Anomaly Detection", + "text_level": 1, + "bbox": [ + 215, + 176, + 452, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Anomaly detection has received increasing attention from researchers in recent years, and many new methods have been proposed to address the problem. Flow-based methods [8, 29, 35, 39] use learned distributions and flow's bijective properties to spot defects, while Memory-based approaches [1, 14, 28] gauge anomaly scores by contrasting test sample features with memory bank-stored norms. Reconstruction-based models [2, 41, 42] flag anomalies by comparing inputs to their online reconstructions. Recent works [12, 16, 32, 43] augment the anomaly detection datasets with generated synthetic anomalies to compensate for the negative example scarcity problem.", + "bbox": [ + 212, + 200, + 787, + 338 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 3D Anomaly Detection", + "text_level": 1, + "bbox": [ + 215, + 359, + 452, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This field lags behind the development of 2D anomaly detection since 3D data are harder to obtain, while point cloud data are sparser and contain more noise than image data. BTF [11] integration of handcrafted 3D descriptors with classic 2D method PatchCore [28], constructing a basic framework for 3D anomaly detection. M3DM [36] advances the field by separately analyzing features from point clouds and RGB images, then merging these for improved decision-making. CPMF [3] converts point clouds into two-dimensional images from multiple angles, extracting additional features from these images with a pre-trained network, and enhancing detection capabilities through information fusion. Reg3D-AD [19] develops a registration-based method, the RANSAC algorithm was used to align each sample before comparing it to the stored template during the test phase. IMRNet [18] trains a PointMAE [25] to reconstruct anomaly-free samples and identifies anomalies by juxtaposing the reconstructed point cloud against the initial input. Many of these use memory banks to store the features of the training samples or require multiple iterations to restore points. Unlike previous methods, our approach requires only one step of reconstruction and has significant advantages in both time and space efficiency.", + "bbox": [ + 212, + 383, + 787, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Diffusion Models", + "text_level": 1, + "bbox": [ + 215, + 662, + 401, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion models have proven their effectiveness in several generative tasks, such as image generation [31], speech generation [15], and video generation [10]. Denoising Diffusion Probabilistic Models (DDPMs) [13, 33, 34] employ a forward noisng mechanism, incrementally integrating Gaussian noise into images, alongside a reverse process meticulously trained to counteract the forward mechanism. Denoise AD [22] conducts DDPM for reconstructing within the features space, generating images that contain less noise. In recent years, many studies [6, 17, 20, 24] have attempted to use the diffusion model to explore the 3D reconstruction task. DPM [23] incorporates a shape latent variable to encapsulate the geometric intricacies of 3D shapes, it distinctively models this variable's", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z. Zhou et al.", + "bbox": [ + 271, + 114, + 364, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "distribution utilizing Normalizing Flows [7, 27]. PVD [44] utilizes PVCNNs [21] for the point-voxel representation of 3D shapes and integrates structured locality into point clouds. This innovative approach leverages the strengths of both point and voxel representations, optimizing the model's ability to capture the intricate spatial hierarchies and local geometries within 3D objects. Since diffusion-based reconstruction recovers the target shape from complete noise, the dilemma of reconstructing only the masked region in the MAE [9] mechanism does not exist.", + "bbox": [ + 212, + 146, + 782, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 276, + 330, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Overview", + "text_level": 1, + "bbox": [ + 215, + 308, + 339, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We model the anomaly detection problem as mapping an anomalous point cloud $\\mathcal{P}_{\\mathrm{a}} \\in \\mathbb{R}^{N \\times 3}$ to a positive shape with which it is aligned. The framework of R3D-AD is shown in Fig. 2, where the simulated anomalous shapes are reconstructed in a self-supervised setting in the training phase and then compared with the original input to detect anomalies. The reconstructed anomaly-free model is aligned with the input, thus allowing direct computation of anomaly scores and segmentation of anomalous regions by conditioned distance functions. Simultaneously, the anomaly simulation strategy faithfully generates realistic defects and randomly synthesizes diverse anomaly shapes on normal samples, improving the generalization ability of the network in the case of limited anomaly samples.", + "bbox": [ + 212, + 334, + 784, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Preliminary of denoising diffusion probabilistic models", + "text_level": 1, + "bbox": [ + 215, + 508, + 710, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A DDPM is inspired by the thermal diffusion process in an evolving thermodynamic system, which consists of a diffusion process and a reverse process.", + "bbox": [ + 212, + 534, + 782, + 564 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The forward Markovian process gradually adds Gaussian noise to a clean sample $\\pmb{x}^{(0)}$ from a data distribution $q(\\pmb{x}^{(0)})$ and turns it into a Gaussian noise $\\pmb{x}^{(T)}$ , which is defined as", + "bbox": [ + 212, + 565, + 784, + 609 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\boldsymbol {x} ^ {(0)}, \\dots , \\boldsymbol {x} ^ {(T)}\\right) = \\prod_ {t = 1} ^ {T} q \\left(\\boldsymbol {x} ^ {(t)} \\mid \\boldsymbol {x} ^ {(t - 1)}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 623, + 784, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $q(\\pmb{x}^{(t)}|\\pmb{x}^{(t-1)}) = \\mathcal{N}(\\pmb{x}^{(t)};\\sqrt{1 - \\beta_t}\\pmb{x}^{(t-1)},\\beta_t\\pmb{I})$ is the Markov diffusion kernel, $t = 1,\\dots,T$ , $T$ is the number of diffusion steps, and $\\beta_{t}$ is a variance schedule. We have $q(\\pmb{x}^{(t)}|\\pmb{x}^{(0)}) = \\mathcal{N}(\\pmb{x}^{(t)};\\sqrt{\\overline{\\alpha}_t}\\pmb{x}^{(0)},(1 - \\overline{\\alpha}_t)\\pmb{I})$ by reparameterization with $\\alpha_{t} = 1 - \\beta_{t}$ , $\\overline{\\alpha}_{t} = \\prod_{s=0}^{t}\\alpha_{s}$ . $\\pmb{x}^{(t)}$ can be sampled by", + "bbox": [ + 214, + 674, + 782, + 737 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} ^ {(t)} = \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} ^ {(0)} + \\epsilon \\sqrt {\\left(1 - \\bar {\\alpha} _ {t}\\right)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 750, + 784, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{\\epsilon}$ is a standard Gaussian noise and $\\epsilon \\sim \\mathcal{N}(0, I)$ . When $T$ is large enough, $\\pmb{x}^{(T)}$ will eventually become a Gaussian noise.", + "bbox": [ + 212, + 779, + 782, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The reverse process is also a Markovian process that denoises over a series of steps to generate meaningful data from the target distribution $q(\\pmb{x}^{(0)})$ . The", + "bbox": [ + 212, + 810, + 784, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection", + "bbox": [ + 292, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e69d491b539e63f2664d9ac55a9a05d526ce672ebe2881b232452ca0d303e8ab.jpg", + "image_caption": [ + "Fig. 2: Overall architecture of R3D-AD for shape reconstruction and anomaly detection of point cloud objects. Reconstruction training phase: The simulated anomalous $\\mathcal{P}_{\\mathrm{a}}^{(0)}$ is generated resort to Patch-Gen from the input point cloud. It further fully masked as $\\mathcal{P}_{\\mathrm{a}}^{(T)}$ while also encoded to latent shape embedding. The SWD decoder then explicitly reconstructs the anomaly-free object $\\mathcal{P}_{\\mathrm{r}}^{(0)}$ with consistent spatial transform by conditionally generating point-level displacements $\\Delta^{(t)}$ at each step of the inverse process. Detection testing phase: The test point cloud $\\mathcal{P}_{\\mathrm{a}}^{(0)}$ is reconstructed to $\\mathcal{P}_{\\mathrm{r}}^{(0)}$ with normal shape, and compared at a distance level to detect the anomalous region." + ], + "image_footnote": [], + "bbox": [ + 225, + 146, + 645, + 348 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c912d14ec7b9122444326834473d856d05cf95cf04711dca06a9231d400e89b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 146, + 781, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "inverse process denoises the noise $\\pmb{x}^{(T)}$ from a distribution $p(\\pmb{x}^{(T)})$ , which is defined as", + "bbox": [ + 212, + 501, + 787, + 532 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\boldsymbol {x} ^ {(0)}, \\dots , \\boldsymbol {x} ^ {(T - 1)} \\mid \\boldsymbol {x} ^ {(T)}, \\boldsymbol {c}\\right) = \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(\\boldsymbol {x} ^ {(t - 1)} \\mid \\boldsymbol {x} ^ {(t)}, \\boldsymbol {c}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 542, + 785, + 584 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $p_{\\theta}(\\boldsymbol{x}^{(t-1)}|\\boldsymbol{x}^{(t)})$ , $\\boldsymbol{c} = \\mathcal{N}(\\boldsymbol{x}^{(t-1)};\\mu_{\\theta}(\\boldsymbol{x}^{(t)},t,\\boldsymbol{c}),\\sigma_t^2\\boldsymbol{I})$ , the mean $\\mu_{\\theta}(\\boldsymbol{x}^{(t)},t,\\boldsymbol{c})$ is estimated by a neural network parameterized by $\\boldsymbol{\\theta}$ , $\\boldsymbol{c}$ is the latent condition encoding, and $\\sigma_t^2$ is a step-dependent variance. $\\mu_{\\theta}$ can be reparameterized as", + "bbox": [ + 214, + 595, + 787, + 643 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {\\theta} \\left(\\boldsymbol {x} ^ {(t)}, t, \\boldsymbol {c}\\right) = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\boldsymbol {x} ^ {(t)} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} ^ {(t)}, t, \\boldsymbol {c}\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 656, + 785, + 689 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\theta}(\\pmb{x}^{(t)}, t, \\pmb{c})$ is a neural network utilized to denoise the Gaussian noise from $\\pmb{x}^{(T)}$ .", + "bbox": [ + 212, + 702, + 787, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The training objective is minimized by training $\\pmb{\\epsilon}_{\\pmb{\\theta}}(\\pmb{x}^{(t)}, t, \\pmb{c})$ to approximate $\\pmb{\\epsilon}$ . The training objective is defined as", + "bbox": [ + 212, + 733, + 787, + 763 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} _ {t \\sim [ 1: T ], \\boldsymbol {x} ^ {(0)} \\sim q (\\boldsymbol {x} ^ {(0)})}, \\epsilon \\sim \\mathcal {N} (0, I) \\| \\epsilon - \\epsilon_ {\\boldsymbol {\\theta}} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} ^ {(0)} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , t, \\boldsymbol {c}\\right) \\|, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 241, + 776, + 785, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $t$ is sampled from the uniform distribution over $1,2,\\dots,T$ , $q(\\pmb{x}^{(0)})$ is the distribution of $\\pmb{x}^{(0)}$ , and $\\epsilon$ is the Gaussian noise.", + "bbox": [ + 212, + 808, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z. Zhou et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Diffusion-based 3D anomaly reconstruction", + "text_level": 1, + "bbox": [ + 215, + 146, + 617, + 162 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We formulate the point cloud reconstruction task of the anomaly-free model as the conditional generation, which decodes the explicit displacement with the target distribution $q(\\mathcal{P}_{\\mathrm{r}}|\\mathcal{P}_{\\mathrm{a}}^{(T)},c)$ , where $c$ is the decoding condition. The essential question of anomaly detection in this paper is how to conditional reconstruct anomaly-free shapes on the reference of input point clouds with different spatial transformations. Since there is a high similarity of global features between abnormal and normal samples during self-supervised reconstruction, the most immediate approach is to extract an efficient global feature from input to serve as an auxiliary conditional embedding for the denoising function $\\epsilon_{\\theta}$ . We implement the encoding of latent shape embedding $c$ as a conditional input to guide reconstruction in the reverse diffusion process.", + "bbox": [ + 212, + 176, + 787, + 345 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Latent shape embedding The feature encoder aims to encode the point cloud to the latent shape embedding $c$ with high-level features for the conditional generation process. Different from other global-local extracting methods [37,40], we focus more on extracting global features, which characterize the semantic information of shape and pose of most anomaly-free regions in the point cloud. The feature encoder mainly consists of cascaded multi-layer perceptions (MLP) based on PointNet [5]. It implements max-pooling after mapping $\\mathcal{P}_{\\mathrm{a}}^{(0)}$ to different dimensions and then compresses them to extract the global shape embedding.", + "bbox": [ + 212, + 372, + 787, + 497 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Step-wise displacement decoding To achieve point cloud reconstruction with transformation consistency while preserving the structure of non-anomalous regions, our method injects latent shape embedding $\\pmb{c}$ to the decoder at each step of the reverse diffusion process, as shown in Fig. 2. In principle, in the training phase, $\\epsilon_{\\theta}$ learns the added Gaussian noise in the forward diffusion process by the decoder to model the conditional probability distribution. Conditionally generating target shapes from $N \\times 3$ Gaussian noise is a straightforward approach, but it is afflicted by the issues of reconstructing the point cloud details and transform consistency. Learning the relative deformation of points for anomalous objects is more efficient. Considering the mapping degradation of the vanilla autoencoder in the reconstruction training phase [22], we utilize the Gaussian noise of the forward process Eq. 2 to fully mask the point cloud object directly without blind spots, preventing the decoding process from receiving negative state shapes. The masked points $\\mathcal{P}_{\\mathrm{a}}^{(T)}$ and latent shape embedding $\\pmb{c} \\in \\mathbb{R}^{256}$ are as the inputs of the SWD decoder. The point-wise displacement vector $\\Delta^{(t)}$ is generated at each step of the iterative process thus disentangling the prediction noise and the desired anomaly-free shape. The reverse process can be defined according to Eq. 3 and the displacement vector $\\Delta^{(t)}$ can be represented by", + "bbox": [ + 212, + 523, + 787, + 800 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta^ {(t - 1)} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\Delta^ {(t)} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\theta}} \\left(\\Delta^ {(t)}, \\beta_ {t}, \\boldsymbol {c}\\right)\\right) + \\sigma \\boldsymbol {\\epsilon}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 811, + 785, + 844 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection", + "bbox": [ + 292, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e024759dbdd519ebdbac73a91ea4561af0c2d82ba80c259a9a900488aea45250.jpg", + "image_caption": [ + "Fig. 3: Illustration of Patch-Gen, the 3D anomaly simulation strategy. The input normal point cloud is first randomly rotated. On the surface of the normalized cube, we randomly select viewpoints to find the nearest patch of points. The selected points are then transformed into irregular defects according to the specific deformation solution." + ], + "image_footnote": [], + "bbox": [ + 277, + 143, + 727, + 296 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\sigma$ is the variance. A PointwiseNet is adopted for $\\epsilon_{\\theta}$ to decode the $\\Delta^{(t-1)}$ from the previous step and $c$ . $\\beta_{t}$ is used to generate trigonometric position embedding $\\pmb{e}_{\\mathrm{p}} = (\\beta_{t}, \\sin(\\beta_{t}), \\cos(\\beta_{t}))$ . $\\pmb{e}_{\\mathrm{p}}$ is concatenated with $c$ and then fed into the concatenate-squash linear module of PointwiseNet with a residual function. The output reconstructed point cloud at the $t$ step is $\\mathcal{P}_{\\mathrm{r}}^{(t)} = \\mathcal{P}_{\\mathrm{r}}^{(t+1)} + \\Delta^{(t)}$ . The registered original and reconstructed objects are distinguished from the anomalous shape by the anomaly scores based on the conditioned distance function.", + "bbox": [ + 212, + 391, + 787, + 501 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 3D anomaly simulation strategy", + "text_level": 1, + "bbox": [ + 214, + 521, + 526, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Given that a small number of normal samples is not conducive for the model to learn diverse and essential features, we propose the Patch-Gen strategy to simulate the defects from anomaly-free shapes for training data augmentation. Patch-Gen encourages the reconstruction model to learn to detect irregularity, where the anomaly-free point clouds and their diverse anomaly patterns are integrated into training pairs and are utilized to learn the discrimination feature between normal and anomalous surfaces. The intuition is that the diversity of simulated negative samples forces our network to learn how to reconstruct anomaly-free shapes instead of memorizing their complete outfits.", + "bbox": [ + 212, + 545, + 787, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Fig. 3, the input normal sample is first randomly rotated. The random spatial rotation is designed to improve the generalization capability for test samples with very different spatial transformations, as defined by:", + "bbox": [ + 212, + 681, + 787, + 726 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} _ {\\mathrm {a}} = \\mathcal {P} \\cdot \\mathcal {R}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 455, + 737, + 785, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{P}$ is input normal sample point cloud and $\\mathcal{R} \\in \\mathbb{R}^{3 \\times 3}$ is obtained by randomly selecting rotation angles for all three axes. In addition to global shape awareness of the model by the random rotation, we further perform a fine granularity of the anomaly simulation. We randomly take a viewpoint $\\mathcal{P}_v$ from the surface of the cube. Therefore, the patch of nearest $N$ points $\\mathcal{P}_n$ from $\\mathcal{P}_a$ can be", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z. Zhou et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "determined according to the $\\mathcal{P}_v$ . The shape augmentation scheme Patch-Gen is defined as follows:", + "bbox": [ + 212, + 146, + 785, + 175 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} _ {n} = \\mathcal {P} _ {n} + S \\cdot \\text {n o r m a l i z e} \\left(\\mathcal {P} _ {n} - \\mathcal {P} _ {v}\\right) \\odot \\mathcal {T}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 189, + 785, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\text{nomalize}$ represents a normalization operation on a vector, $S$ is a predefined hyper-parameter that controls the scaling of the patch points, and $\\mathcal{T}$ is the translation matrix originating from a Gaussian distribution. The $\\mathcal{P}_{\\mathrm{a}}$ is finally obtained by only updating the patch region $\\mathcal{P}_n$ .", + "bbox": [ + 212, + 215, + 782, + 276 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "With the proposed Patch-Gen, we can simulate the generation of multiple anomalies, which is mainly done by controlling $\\mathcal{T}$ . Bulge or sink can be generated by sorting $\\mathcal{T}$ after sampling from the distribution, while damage can be generated by direct overlaying without manipulation.", + "bbox": [ + 212, + 277, + 784, + 338 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.5 Training objective", + "text_level": 1, + "bbox": [ + 214, + 359, + 413, + 376 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In the reconstruction task of the object with $N$ points, the network learns a diffusion model with an $\\mathbb{R}^{N\\times 3}\\to \\mathbb{R}^{N\\times 3}$ mapping relation. Iterative denoising under the semantic condition of point embedding realizes the prediction of point offsets. Concretely, the network is trained to learn the noise that needs to be eliminated to recover the anomaly-free shape with the $L_{2}$ distance between the ground truth and the denoised reconstructed points. We make use of the mean squared error (MSE) loss as the primary reconstruction loss which evaluates the mean squared error of the element-wise distances between $\\mathcal{P}_{\\mathrm{a}}^{(0)}$ and $\\mathcal{P}_{\\mathrm{r}}^{(0)}$ . The MSE training loss is formulated as:", + "bbox": [ + 212, + 383, + 787, + 523 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {P} _ {\\mathrm {a}}, \\mathcal {P} _ {\\mathrm {r}}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} p _ {\\mathrm {a}} \\in \\mathcal {P} _ {\\mathrm {a}}, p _ {\\mathrm {r}} \\in \\mathcal {P} _ {\\mathrm {r}} \\| p _ {\\mathrm {a}} - p _ {\\mathrm {r}} \\| ^ {2}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 534, + 785, + 575 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 595, + 375, + 612 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 214, + 626, + 334, + 640 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Real3D-AD [19] is a 3D anomaly detection dataset based on real samples, exhibiting a higher point precision and spatial distance per point cloud. Each category contains 4 training samples and 100 test samples. The training set contains $360^{\\circ}$ complete surface point clouds of the objects, which are obtained by manually calibrating and stitching the scans of multiple sides of the objects. The test samples are scans only one side with a huge difference from the training set. The distribution of the point clouds also varies among the total 12 categories, further deepening the detection difficulty compared to 2D scenes.", + "bbox": [ + 212, + 651, + 787, + 773 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Anomaly-ShapeNet [18] is a 3D anomaly detection, crafted through modifications to the synthetic samples found in ShapeNetCorev2 [4]. It contains 40 diverse categories, featuring over 1600 samples of its complete surface point", + "bbox": [ + 214, + 794, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection", + "bbox": [ + 292, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/5f8b31c05850a13cbb655442ed458fa6320a530222d6eb2c77c905c4e226f90f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBTF [11]M3DM [36]PatchCore [28]CPMF [3]Reg3D-AD [19]IMRNet [18]Ours
Feat.RawFPFHPointMAEFPFHPointMAEResNetPointMAEPointMAERaw
Airplane0.7300.5200.4340.8820.7260.7010.7160.7620.772
Candybar0.5390.6300.5520.5410.6630.5520.6850.7550.696
Car0.6470.5600.5410.5900.4980.5510.6970.7110.713
Chicken0.7890.4320.6830.8370.8270.5040.8520.7800.714
Diamond0.7070.5450.6020.5740.7830.5230.9000.9050.685
Duck0.6910.7840.4330.5460.4890.5820.5840.5170.909
Fish0.6020.5490.5400.6750.6300.5580.9150.8800.692
Gemstone0.6860.6480.6440.3700.3740.5890.4170.6740.665
Seahorse0.5960.7790.4950.5050.5390.7290.7620.6040.720
Shell0.3960.7540.6940.5890.5010.6530.5830.6650.840
Starfish0.5300.5750.5510.4410.5190.7000.5060.6740.701
Toffees0.7030.4620.4500.5650.5850.3900.8270.7740.703
Average0.6350.6030.5520.5930.5950.5860.7040.7250.734
", + "bbox": [ + 218, + 143, + 782, + 335 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 1: Image-level anomaly detection AUROC on Real3D-AD dataset. We highlight the best result in bold and the second best result with an underline.", + "bbox": [ + 215, + 338, + 784, + 364 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "clouds. Each category's training set contains merely 4 samples, while the test sets are designed to assess the model's performance across both normal and a spectrum of abnormal samples. It widely increases the anomaly types while keeping the number of points the same as the previous studies, which places higher demands on the robustness and generality of the proposed algorithms.", + "bbox": [ + 212, + 411, + 784, + 488 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Evaluation metrics", + "text_level": 1, + "bbox": [ + 215, + 517, + 416, + 531 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For image-level anomaly detection, the Area Under the Receiver Operating Curve (AUROC) is utilized in line with established practices. For the evaluation of pixel-level anomalies, the AUROC metric is similarly applied in the context of point segmentation accuracy. A value of 0.5 of the AUROC score denotes no discriminative capability (equivalent to random guessing), whereas a score of 1.0 indicates perfect discrimination between positive and negative classes.", + "bbox": [ + 212, + 550, + 785, + 642 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Implementation details", + "text_level": 1, + "bbox": [ + 215, + 671, + 454, + 686 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Our methodology is implemented using PyTorch [26] with end-to-end training across the network. The optimization is performed using the Adam optimizer, starting at an initial learning of 0.001. The training process involves a total batch size of 128 across 40,000 iterations for comprehensive learning. All input point clouds undergo a preprocessing step where they are randomly downsampled to a fixed size of 4096 and 2048 points on Real3D-AD and Anomaly-ShapeNet, respectively. Additionally, we normalized these point clouds by setting their center of gravity as the origin of coordinates and scaling their dimensions to fall within the range of -1 to 1, optimizing for the diffusion process.", + "bbox": [ + 212, + 703, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z. Zhou et al.", + "bbox": [ + 271, + 114, + 364, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/5dec720dc96e48350668acfbd4ee1313c7b5defae52b6f60ee1c219fd44b3093.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBTF [11]M3DM [36]PatchCore [28]CPMF [3]Reg3D-AD [19]IMRNet [18]Ours
Feat.RawFPFHPointMAEFPFHPointMAEResNetPointMAEPointMAERaw
Ashtray0.5780.4200.5770.5870.5910.3530.5970.6710.833
Bag0.4100.5460.5370.5710.6010.6430.7060.6600.719
Bottle0.5580.4040.5840.6140.5880.4690.5690.6310.750
Bowl0.4700.5810.5790.5580.5470.6790.5480.6760.751
Bucket0.4690.5170.4050.5100.5770.5420.6810.6760.719
Cap0.5090.5620.5990.6450.5830.6010.6320.7040.726
Cup0.4620.5980.5480.5930.5830.4980.5240.7000.767
Eraser0.5250.7190.6270.6570.6770.6890.3430.5480.890
Headset0.4470.5050.5970.6100.6090.5510.5740.6980.767
Helmet0.5080.5690.4880.4650.4950.5320.4910.6030.704
Jar0.4200.4240.4410.4720.4830.6100.5920.7800.838
Microphone0.5630.6710.3570.3880.4880.5090.4140.7550.762
Shelf0.1640.6090.5640.4940.5230.6850.6880.6030.696
Tap0.5490.5530.7470.7600.4980.5280.6590.6860.818
Vase0.5170.4640.5340.5540.5820.5140.5760.6290.734
Average0.4930.5280.5520.5680.5620.5590.5720.6590.749
", + "bbox": [ + 218, + 143, + 782, + 368 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 2: Image-level anomaly detection AUROC on Anomaly-ShapeNet dataset. We highlight the best result in bold and the second best result with an underline.", + "bbox": [ + 215, + 369, + 784, + 398 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Main results", + "text_level": 1, + "bbox": [ + 215, + 450, + 366, + 464 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We conduct experiments on Real3D-AD [19] based on real sampling and AnomalyShapeNet [18] based on simulation.", + "bbox": [ + 212, + 497, + 790, + 527 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As shown in Table 1, we first compare the image-level AUROC metric with current cutting-edge 3D anomaly detection models on Real3D-AD. It shows that our method achieves the best performance using only raw point cloud data, while most of the existing methods use Fast Point Feature Histograms (FPFH) operator [30] or ShapeNet [4] pre-trained PointMAE [25] as feature extractor. Due to significant disparities in quantity, size, and distribution among different categories of point clouds in Real3D-AD, scoring variations across categories are more pronounced with other methods. For instance, numerous methods perform under 0.5 in certain categories, indicating their inadequacy in extracting meaningful features while facing challenging samples. In contrast, our method not only exhibits superior performance in 3D anomaly detection across the majority of categories but also achieves the best overall average across all categories. This demonstrates the strong generalizability and robustness of our approach.", + "bbox": [ + 212, + 532, + 787, + 731 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We further evaluate our method on Anomaly-ShapeNet in Table 2, which encompasses a broader array of categories and a greater diversity of defect types. Compared to Real3D-AD, Anomaly-ShapeNet significantly enhances the diversity of defects, wherein the increased variety of defect types further escalates the complexity of detection tasks. The results highlight the exceptional performance of our method across all evaluated categories, demonstrating an average improvement of $9\\%$ on AUROC relative to the approaches previously utilized.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection", + "bbox": [ + 292, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/c596b9bfa2fcc9ce7a2f3f5ccb0b98ad697bcb151c52feaca07b95578da4735b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelDiffusionConditionRelativePatch-GenI-AUROCP-AUROC
A✓✗✗✗0.5860.524
B✓✓✗✗0.6670.513
C✓✓✓✗0.7120.573
D✓✓✓✓0.7340.592
", + "bbox": [ + 277, + 143, + 723, + 231 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 3: Ablation studies for 3D adaptation components on Real3D-AD dataset.", + "bbox": [ + 228, + 232, + 771, + 244 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Ablation study", + "text_level": 1, + "bbox": [ + 215, + 287, + 385, + 303 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To delve into the effect of individual components, we conduct ablation experiments on the Real3D-AD dataset. To fully demonstrate and compare the performance of the models, we report both image-level and pixel-level results with I-AUROC and P-AUROC, respectively.", + "bbox": [ + 212, + 311, + 785, + 375 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Main component Table 3 compares the performance of different variants from R3D-AD, which includes the influence of the denoising condition embedding, displacement-based reconstruction way, and the data augmentation strategy of Patch-Gen. Model A is denoted as our baseline, which is a vanilla DDPM model for point cloud reconstruction. Introducing a condition into the DDPM (Model B) significantly boosts performance, particularly in terms of I-AUROC, which sees a $13.8\\%$ increase to 0.667. Model C, which predicts point displacements based on conditional DDPM, preserving detailed structural information while accommodating the relative displacement of points contributes to a notable $6.0\\%$ gain in P-AUROC over Model B. Model D is trained under the conditions of shape embedding with the Patch-Gen strategy. Considering that the defective portion contains only a small portion of the original point cloud, we try to reconstruct the relative displacement in a way that preserves as much detail as possible, which is effective for both 3D anomaly detection and segmentation.", + "bbox": [ + 212, + 393, + 787, + 608 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Patch-Gen Table. 4 analyzes the influence of two key parameters in Patch-Gen: the selection points ratio and the scaling points factor.", + "bbox": [ + 212, + 628, + 785, + 657 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The selection points ratio from Table. 4a determines the proportion of points in the point cloud that are selected for transformation. Our findings suggest that a selection ratio of $1/32$ achieves the best performance. It appears that this ratio provides a balanced trade-off between maintaining sufficient structure for anomaly detection and introducing enough variation to simulate anomalies effectively. Notably, as the ratio increases beyond $1/16$ , both I-AUROC and P-AUROC scores decrease in severity, since real defects only account for a small portion of the overall point cloud, a wide selection of points not only destroys the structure of the original point cloud, but also makes the distribution of the training and test sets inconsistent.", + "bbox": [ + 212, + 659, + 787, + 808 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The scaling points factor is the intensity of the random transformation applied to the selected points, as detailed in Table 4b. The optimal performance is", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z. Zhou et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/a91c5ed4ddf9ced017aa5607e8a2fe1adeb799c55f72b04dc5fdaf52b1231fc9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ratioI-AUROCP-AUROC
1/640.7160.584
1/320.7340.592
1/160.7270.579
1/80.6830.528
", + "bbox": [ + 310, + 142, + 496, + 228 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/41a4ac21da48477f1a60024c075876390d2a4a42b9b9bbd67fc9773b55662eed.jpg", + "table_caption": [ + "(a) Selection points ratio." + ], + "table_footnote": [], + "table_body": "
factorI-AUROCP-AUROC
0.10.7340.592
0.20.7270.572
0.40.7150.554
0.80.6610.517
", + "bbox": [ + 503, + 143, + 697, + 229 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "(b) Scaling points factor.", + "bbox": [ + 514, + 229, + 678, + 242 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 4: Ablation studies for Patch-Gen implementation on Real3D-AD dataset. Default settings are marked in gray.", + "bbox": [ + 214, + 253, + 782, + 282 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e6ff16748f29b8ed19dce6e60f2b302acbe66932d90b7223ee0b54508a5bb848.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 223, + 318, + 493, + 434 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8e47002cd3a71ddd50d9a1c6e5d39e3739396e30da9e3e4b36a840d07aebe135.jpg", + "image_caption": [ + "(b)", + "Fig. 4: Memory and time cost during inference on Real3D-AD dataset. (a) Memory usage comparison between different models. (b) 3D anomaly detection performance vs. frames per second on an NVIDIA RTX 3090 GPU. Our R3D-AD outperforms all previous methods on both accuracy and efficiency by a significant margin." + ], + "image_footnote": [], + "bbox": [ + 504, + 321, + 774, + 441 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "observed at a scaling factor of 0.1, which implies that minor transformations are more effective for simulating anomalies without significantly altering the original data distribution. Larger scaling factors lead to a consistent decline in performance, underscoring the importance of subtle transformations for preserving the utility of the simulated anomalies for detection tasks.", + "bbox": [ + 212, + 554, + 787, + 630 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Memory and time cost As depicted in Figure 4, we evaluate the disparity in both storage consumption and inference time of our model under identical experimental conditions, compared to existing methods. Regarding memory usage, our approach demonstrates a marked superiority by employing raw coordinate features instead of FPFH or PointMAE features, significantly reducing the memory footprint. Since no memory bank exists, our method is also more space-efficient compared to BTF which also uses raw features. Moreover, our method eliminates the necessity to compare all the features in memory, substantially increasing operational efficiency. The implementation of Patch-Gen inherently bestows our model with exceptional robustness, enabling precise reconstruction of point clouds from various angles without the need for the time-intensive RANSAC alignment process required by Reg3D-AD.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection", + "bbox": [ + 292, + 113, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/4bfac1c3112a59a58b7e6a096b999e1658417378b8cbe516422c3f7190ea7e2e.jpg", + "image_caption": [ + "Fig. 5: Qualitative analysis on Real3D-AD dataset and Anomaly-ShapeNet dataset. The anomaly map is obtained directly by calculating the differences between the input and reconstructed point clouds, where deeper colors represent more confidence." + ], + "image_footnote": [], + "bbox": [ + 223, + 143, + 777, + 325 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.6 Qualitative results", + "text_level": 1, + "bbox": [ + 215, + 409, + 415, + 425 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Figure 5 presents some qualitative outcomes, with varying shades of color indicating different levels of anomaly scores. We select several representative defective samples to demonstrate the robustness of our algorithm. The left four columns display samples from Real3D-AD, while the right four columns samples are from Anomaly-ShapeNet. The illustration reveals that our R3D-AD algorithm has precisely reconstructed the defective portions of the point cloud across various samples: the deep sink in the Seahorse sample, the concavity in the Bag sample, and the bulge in the Jar sample. Leveraging the accurately reconstructed point clouds, final point cloud segmentation maps are also produced, further evidencing the efficacy of our approach.", + "bbox": [ + 212, + 435, + 787, + 587 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 609, + 359, + 626 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we presented R3D-AD, a novel reconstructive 3D anomaly detection model based on conditional diffusion. Our goal is to overcome the limitations faced by current 3D anomaly detection methods, such as the inefficiencies due to the memory bank module and low performance caused by incorrect rebuilds with MAE. To address these challenges, we leverage the diffusion process for full reconstruction, followed by a direct comparison between the input and the reconstructed point cloud to obtain the final anomaly score. The embedded latent variable that spans the decoding process, step-wisely generating point-level displacements from the noise to the target anomaly-free sample. We also propose Patch-Gen, a data augmentation tailored for point cloud anomaly simulation. Extensive experiments conducted on 3D anomaly benchmarks validate the superiority of our R3D-AD in comparison to state-of-the-art alternatives in terms of both accuracy and versatility.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z. Zhou et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported in part by the Pioneer and Leading Goose R&D Program of Zhejiang (Grant No. 2022C01051), in part by the National Natural Science Foundation of China (Grant No. 52375271, 52275274), and in part by the Natural Science Foundation of Zhejiang Province (Grant No. LY23E050011).", + "bbox": [ + 212, + 172, + 787, + 236 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 253, + 321, + 271 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bae, J., Lee, J.H., Kim, S.: Pni: Industrial anomaly detection using position and neighborhood information. In: ICCV (2023)", + "2. Bergmann, P., Löwe, S., Fauser, M., Sattlegger, D., Steger, C.: Improving unsupervised defect segmentation by applying structural similarity to autoencoders. In: VISIGRAPP (2019)", + "3. Cao, Y., Xu, X., Shen, W.: Complementary pseudo multimodal feature for point cloud anomaly detection. arXiv preprint (2023)", + "4. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., Xiao, J., Yi, L., Yu, F.: ShapeNet: An Information-Rich 3D Model Repository. arXiv preprint (2015)", + "5. Charles, R.Q., Su, H., Kaichun, M., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: CVPR (2017)", + "6. Chu, R., Xie, E., Mo, S., Li, Z., Nießner, M., Fu, C.W., Jia, J.: Diffcomplete: Diffusion-based generative 3d shape completion. In: NeurIPS (2023)", + "7. Dinh, L., Sohl-Dickstein, J., Bengio, S.: Density estimation using real nvp. In: ICLR (2017)", + "8. Gudovskiy, D., Ishizaka, S., Kozuka, K.: Cflow-ad: Real-time unsupervised anomaly detection with localization via conditional normalizing flows. In: WACV (2022)", + "9. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022)", + "10. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., Salimans, T.: Imagen video: High definition video generation with diffusion models (2022)", + "11. Horwitz, E., Hoshen, Y.: Back to the feature: Classical 3d features are (almost) all you need for 3d anomaly detection. In: CVPRW (2023)", + "12. Hu, T., Zhang, J., Yi, R., Du, Y., Chen, X., Liu, L., Wang, Y., Wang, C.: Anomalydiffusion: Few-shot anomaly image generation with diffusion model. In: AAAI (2024)", + "13. Jonathan Ho, Ajay Jain, and Pieter Abbeel: Denoising diffusion probabilistic models. In: NeurIPS (2020)", + "14. Kim, D., Park, C., Cho, S., Lee, S.: Fapm: Fast adaptive patch memory for real-time industrial anomaly detection. In: ICASSP (2023)", + "15. Kong, Z., Ping, W., Huang, J., Zhao, K., Catanzaro, B.: Diffwave: A versatile diffusion model for audio synthesis. In: ICLR (2021)", + "16. Li, C.L., Sohn, K., Yoon, J., Pfister, T.: Cutpaste: Self-supervised learning for anomaly detection and localization. In: CVPR (2021)", + "17. Li, M., Duan, Y., Zhou, J., Lu, J.: Diffusion-sdf: Text-to-shape via voxelized diffusion. In: CVPR (2023)", + "18. Li, W., Xu, X., Gu, Y., Zheng, B., Gao, S., Wu, Y.: Towards scalable 3d anomaly detection and localization: A benchmark via 3d anomaly synthesis and a self-supervised learning network. arXiv preprint (2023)" + ], + "bbox": [ + 218, + 282, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection", + "bbox": [ + 292, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Liu, J., Xie, G., Li, X., Wang, J., Liu, Y., Wang, C., Zheng, F., et al.: Real3d-ad: A dataset of point cloud anomaly detection. In: NeurIPS (2023)", + "20. Liu, Z., Feng, Y., Black, M.J., Nowrouzezahrai, D., Paull, L., Liu, W.: Meshdiffusion: Score-based generative 3d mesh modeling. In: ICLR (2023)", + "21. Liu, Z., Tang, H., Lin, Y., Han, S.: Point-voxel cnn for efficient 3d deep learning. In: NeurIPS (2019)", + "22. Lu, F., Yao, X., Fu, C., Jia, J.: Removing anomalies as noises for industrial defect localization. In: ICCV (2023)", + "23. Luo, S., Hu, W.: Diffusion probabilistic models for 3d point cloud generation. In: CVPR (2021)", + "24. Mo, S., Xie, E., Chu, R., Hong, L., Nießner, M., Li, Z.: Dit-3d: Exploring plain diffusion transformers for 3d shape generation. In: NeurIPS (2023)", + "25. Pang, Y., Wang, W., Tay, F.E., Liu, W., Tian, Y., Yuan, L.: Masked autoencoders for point cloud self-supervised learning. In: ECCV (2022)", + "26. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. NeurIPS (2019)", + "27. Rezende, D.J., Mohamed, S.: Variational inference with normalizing flows. In: ICML (2015)", + "28. Roth, K., Pemula, L., Zepeda, J., Scholkopf, B., Brox, T., Gehler, P.: Towards total recall in industrial anomaly detection. In: CVPR (2022)", + "29. Rudolph, M., Wandt, B., Rosenhahn, B.: Same same but differnet: Semi-supervised defect detection with normalizing flows. In: WACV (2021)", + "30. Rusu, R.B., Blodow, N., Beetz, M.: Fast point feature histograms (fpfh) for 3d registration. In: ICRA (2009)", + "31. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Ghasemipour, S.K.S., Ayan, B.K., Mahdavi, S.S., Lopes, R.G., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint (2022)", + "32. Schlüter, H.M., Tan, J., Hou, B., Kainz, B.: Natural synthetic anomalies for self-supervised anomaly detection and localization. In: ECCV (2022)", + "33. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint (2020)", + "34. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: ICLR (2021)", + "35. Tailanian, M., Pardo, Å., Muse, P.: U-flow: A u-shaped normalizing flow for anomaly detection with unsupervised threshold. arXiv preprint (2022)", + "36. Wang, Y., Peng, J., Zhang, J., Yi, R., Wang, Y., Wang, C.: Multimodal industrial anomaly detection via hybrid fusion. In: CVPR (2023)", + "37. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph cnn for learning on point clouds. TOG (2019)", + "38. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: Simmim: A simple framework for masked image modeling. In: CVPR (2022)", + "39. Yu, J., Zheng, Y., Wang, X., Li, W., Wu, Y., Zhao, R., Wu, L.: Fastflow: Unsupervised anomaly detection and localization via 2d normalizing flows. arXiv preprint (2021)", + "40. Yu, X., Tang, L., Rao, Y., Huang, T., Zhou, J., Lu, J.: Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In: CVPR (2022)", + "41. Zavrtanik, V., Kristan, M., Skočaj, D.: Draem-a discriminatively trained reconstruction embedding for surface anomaly detection. In: ICCV (2021)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z. Zhou et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "42. Zavrtanik, V., Kristan, M., Skočaj, D.: Reconstruction by inpainting for visual anomaly detection. Pattern Recognition (2021)", + "43. Zhang, X., Xu, M., Zhou, X.: Realnet: A feature selection network with realistic synthetic anomaly for anomaly detection. In: CVPR (2024)", + "44. Zhou, L., Du, Y., Wu, J.: 3d shape generation and completion through point-voxel diffusion. In: ICCV (2021)" + ], + "bbox": [ + 215, + 146, + 784, + 229 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection", + "bbox": [ + 292, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/R3D-AD_ Reconstruction via Diffusion for 3D Anomaly Detection/a722039e-753c-4cd1-9dab-7069c8f25f9e_model.json b/2024/R3D-AD_ Reconstruction via Diffusion for 3D Anomaly Detection/a722039e-753c-4cd1-9dab-7069c8f25f9e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..51ec232df815cd77848037ad78fdbee832cee5f9 --- /dev/null +++ b/2024/R3D-AD_ Reconstruction via Diffusion for 3D Anomaly Detection/a722039e-753c-4cd1-9dab-7069c8f25f9e_model.json @@ -0,0 +1,2192 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.241, + 0.141, + 0.764, + 0.187 + ], + "angle": 0, + "content": "R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection" + }, + { + "type": "text", + "bbox": [ + 0.323, + 0.212, + 0.677, + 0.244 + ], + "angle": 0, + "content": "Zheyuan Zhou\\(^{1*}\\), Le Wang\\(^{1*}\\), Naiyu Fang\\(^{1,2}\\), Zili Wang\\(^{1}\\), Lemiao Qiu\\(^{1\\boxtimes}\\), and Shuyou Zhang\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.255, + 0.713, + 0.282 + ], + "angle": 0, + "content": "\\(^{1}\\) State Key Laboratory of Fluid Power & Mechatronic System, Zhejiang University, Hangzhou, China" + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.283, + 0.718, + 0.297 + ], + "angle": 0, + "content": "\\(^{2}\\) S-Lab, Nanyang Technological University, Singapore, Singapore" + }, + { + "type": "text", + "bbox": [ + 0.359, + 0.311, + 0.642, + 0.324 + ], + "angle": 0, + "content": "https://zhouzheyuan.github.io/r3d-ad" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.358, + 0.74, + 0.662 + ], + "angle": 0, + "content": "Abstract. 3D anomaly detection plays a crucial role in monitoring parts for localized inherent defects in precision manufacturing. Embedding-based and reconstruction-based approaches are among the most popular and successful methods. However, there are two major challenges to the practical application of the current approaches: 1) the embedded models suffer the prohibitive computational and storage due to the memory bank structure; 2) the reconstructive models based on the MAE mechanism fail to detect anomalies in the unmasked regions. In this paper, we propose R3D-AD, reconstructing anomalous point clouds by diffusion model for precise 3D anomaly detection. Our approach capitalizes on the data distribution conversion of the diffusion process to entirely obscure the input's anomalous geometry. It step-wisely learns a strict point-level displacement behavior, which methodically corrects the aberrant points. To increase the generalization of the model, we further present a novel 3D anomaly simulation strategy named Patch-Gen to generate realistic and diverse defect shapes, which narrows the domain gap between training and testing. Our R3D-AD ensures a uniform spatial transformation, which allows straightforwardly generating anomaly results by distance comparison. Extensive experiments show that our R3D-AD outperforms previous state-of-the-art methods, achieving \\(73.4\\%\\) Image-level AUROC on the Real3D-AD dataset and \\(74.9\\%\\) Image-level AUROC on the Anomaly-ShapeNet dataset with an exceptional efficiency." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.675, + 0.741, + 0.703 + ], + "angle": 0, + "content": "Keywords: 3D anomaly detection, industrial applications, 3D reconstruction, self-supervised learning" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.727, + 0.377, + 0.743 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.757, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Anomaly detection aims to identify instances containing anomalies and to precisely locate the specific positions of defects. This task is extensively applied across multiple fields and plays a crucial role in quality control within industrial production [28]. 3D anomaly detection [19] has emerged due to its intrinsic" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.826, + 0.365, + 0.84 + ], + "angle": 0, + "content": "* Equal contribution." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "Z. Zhou et al." + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.149, + 0.484, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.332, + 0.247, + 0.425, + 0.257 + ], + "angle": 0, + "content": "(a) Embedded model" + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.149, + 0.718, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.246, + 0.682, + 0.256 + ], + "angle": 0, + "content": "(b) Reconstructive model" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.274, + 0.788, + 0.357 + ], + "angle": 0, + "content": "Fig. 1: Comparison of architectures. (a) Embedded model encodes the input \\(\\mathcal{X}\\) into features and stores them in the memory bank during training. The anomaly map \\(\\mathcal{M}\\) is obtained by comparing the test features with all the features in the memory bank. (b) Reconstructive model is trained by minimizing the loss between its input \\(\\mathcal{X}\\) and the output \\(\\widehat{\\mathcal{X}}\\). The anomaly map \\(\\mathcal{M}\\) is obtained by comparing the test phase input with its corresponding reconstruction target." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.395, + 0.788, + 0.501 + ], + "angle": 0, + "content": "modality superior for avoiding blind spots in advanced processing and precision manufacturing. However, the discrete and disordered data form of point clouds makes it more difficult to acquire features compared to images. With the scarcity of anomalies, 3D anomaly detection also faces the problem of domain shift while only normal data are presented during training. The presence of these issues underscores the necessity and urgency of devising an efficient framework for the 3D anomaly detection task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.788, + 0.777 + ], + "angle": 0, + "content": "Similar to traditional 2D anomaly detection [28,42], current 3D anomaly detection can be primarily categorized into embedding-based and reconstruction-based, as illustrated in Fig. 1. The embedding-based methods involve mapping features extracted with a pre-trained encoder onto a normal distribution for learning. Distributions that do not fall within the interval are classified as anomalies. Most existing 3D anomaly detection methods are based on a memory bank mechanism [3, 11, 19, 36], which stores some representative features during the training phase to implicitly construct a feature distribution. In the testing phase, the presence of anomalies is determined by calculating the Euclidean distance between the input test object and all template point clouds stored in memory. The reconstruction-based methods train a network capable of accurately reconstructing normal point clouds, under the presumption that anomalous point clouds will not be effectively reconstructed since they are not included during training. The anomaly map is produced through the comparison of discrepancies between the input point cloud and its reconstruction. IMRNet [18] employs PointMAE [25] to reconstruct the input in several iterations, getting the final anomaly map by calculating the explicit spatial coordinate differences and implicit deep feature differences of the point cloud, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "However, existing methods face two key issues, high resource cost and irreparable reconstruction. Firstly, methods based on the memory bank [3,11,19, 36] store all features from the training phase, each test point cloud needs to be compared with all samples in the memory bank, significantly increasing memory" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "overhead and inference time costs. This makes such methods almost inapplicable in real industrial production lines due to their inefficiency. Secondly, masked autoencoder (MAE) mechanism [9,25,38] only reconstructs the masked portions of the input, defects within unmasked portions may be preserved. This contradicts the fundamental assumption of detecting anomalies by comparing the original defect-containing point cloud with a reconstructed anomaly-free version. These methods inevitably lead to incorrect reconstructions, undermining their effectiveness in accurately localizing defects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.282, + 0.789, + 0.661 + ], + "angle": 0, + "content": "We propose R3D-AD, a novel 3D anomaly detection method that does not suffer from the space burden and time endurance in memory-based embedded models nor the anomaly unmasking probability in the MAE-based reconstructive models. In contrast to PointMAE, one of our key insights is to perform undifferentiated masking for 3D objects via the noise diffusion mechanism, which maximizes the preservation of anomaly-free shapes and reconstructs abnormal regions. In the reparameterized diffusion process, one-step full mask and reconstruction are achieved by converting the point cloud distribution, instead of the multiple iterative method [18]. We hypothesize that anomaly detection verifies the gap between the reconstructed shapes and the positive samples by learning point movement. Specifically, for input models with arbitrary anomalies, we encode them as latent shape embeddings as decoding conditions and explicitly control the point cloud reconstruction process by step-wise displacements (SWD) decoding. The shape embedding harbors abundant global features and makes it easier to train the network without dwelling on the introduction of local anomaly details. Another key to our approach is to implement a controllable method of point-wise displacement during the diffusion process to refine the point cloud deformation iteratively. We propose to inject latent shape embedding into each step of the inverse denoising process, which drives the anomalous regions to converge to a smooth surface. We further adopted a 3D anomaly simulation strategy Patch-Gen to address the limitations of the dataset, which generates abundant defectives by producing spatial irregularity that is faithful to the real scene, including bulges, sinks, etc. This point cloud data augmentation encourages the self-supervised model to reconstruct more realistic anomaly-free shapes when facing the actual anomaly." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.79, + 0.841 + ], + "angle": 0, + "content": "To the best of our knowledge, this is the very first attempt at exploring diffusion in reconstruction-based 3D anomaly detection. Our main contributions are summarized as follows: (i) We introduce a novel framework, termed R3D-AD, which performs a one-step full mask and anomaly-free reconstruction for fast and accurate 3D anomaly detection. (ii) We propose to learn the step-wise displacement in the reverse diffusion process to explicitly control the reconstruction of anomalous shapes. (iii) We introduce a 3D anomaly simulation strategy named Patch-Gen to address the limitation of the data anomaly patterns and improve the reconstruction performance in a supervised setting. (iv) Extensive experiments demonstrate that our R3D-AD has achieved state-of-the-art performance on both Real3D-AD and Anomaly-ShapeNet datasets." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.128 + ], + "angle": 0, + "content": "Z. Zhou et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.383, + 0.161 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.177, + 0.454, + 0.192 + ], + "angle": 0, + "content": "2.1 2D Anomaly Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.202, + 0.788, + 0.339 + ], + "angle": 0, + "content": "Anomaly detection has received increasing attention from researchers in recent years, and many new methods have been proposed to address the problem. Flow-based methods [8, 29, 35, 39] use learned distributions and flow's bijective properties to spot defects, while Memory-based approaches [1, 14, 28] gauge anomaly scores by contrasting test sample features with memory bank-stored norms. Reconstruction-based models [2, 41, 42] flag anomalies by comparing inputs to their online reconstructions. Recent works [12, 16, 32, 43] augment the anomaly detection datasets with generated synthetic anomalies to compensate for the negative example scarcity problem." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.36, + 0.454, + 0.375 + ], + "angle": 0, + "content": "2.2 3D Anomaly Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.789, + 0.643 + ], + "angle": 0, + "content": "This field lags behind the development of 2D anomaly detection since 3D data are harder to obtain, while point cloud data are sparser and contain more noise than image data. BTF [11] integration of handcrafted 3D descriptors with classic 2D method PatchCore [28], constructing a basic framework for 3D anomaly detection. M3DM [36] advances the field by separately analyzing features from point clouds and RGB images, then merging these for improved decision-making. CPMF [3] converts point clouds into two-dimensional images from multiple angles, extracting additional features from these images with a pre-trained network, and enhancing detection capabilities through information fusion. Reg3D-AD [19] develops a registration-based method, the RANSAC algorithm was used to align each sample before comparing it to the stored template during the test phase. IMRNet [18] trains a PointMAE [25] to reconstruct anomaly-free samples and identifies anomalies by juxtaposing the reconstructed point cloud against the initial input. Many of these use memory banks to store the features of the training samples or require multiple iterations to restore points. Unlike previous methods, our approach requires only one step of reconstruction and has significant advantages in both time and space efficiency." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.664, + 0.403, + 0.679 + ], + "angle": 0, + "content": "2.3 Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Diffusion models have proven their effectiveness in several generative tasks, such as image generation [31], speech generation [15], and video generation [10]. Denoising Diffusion Probabilistic Models (DDPMs) [13, 33, 34] employ a forward noisng mechanism, incrementally integrating Gaussian noise into images, alongside a reverse process meticulously trained to counteract the forward mechanism. Denoise AD [22] conducts DDPM for reconstructing within the features space, generating images that contain less noise. In recent years, many studies [6, 17, 20, 24] have attempted to use the diffusion model to explore the 3D reconstruction task. DPM [23] incorporates a shape latent variable to encapsulate the geometric intricacies of 3D shapes, it distinctively models this variable's" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.253 + ], + "angle": 0, + "content": "distribution utilizing Normalizing Flows [7, 27]. PVD [44] utilizes PVCNNs [21] for the point-voxel representation of 3D shapes and integrates structured locality into point clouds. This innovative approach leverages the strengths of both point and voxel representations, optimizing the model's ability to capture the intricate spatial hierarchies and local geometries within 3D objects. Since diffusion-based reconstruction recovers the target shape from complete noise, the dilemma of reconstructing only the masked region in the MAE [9] mechanism does not exist." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.277, + 0.331, + 0.293 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.309, + 0.341, + 0.323 + ], + "angle": 0, + "content": "3.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.335, + 0.785, + 0.486 + ], + "angle": 0, + "content": "We model the anomaly detection problem as mapping an anomalous point cloud \\(\\mathcal{P}_{\\mathrm{a}} \\in \\mathbb{R}^{N \\times 3}\\) to a positive shape with which it is aligned. The framework of R3D-AD is shown in Fig. 2, where the simulated anomalous shapes are reconstructed in a self-supervised setting in the training phase and then compared with the original input to detect anomalies. The reconstructed anomaly-free model is aligned with the input, thus allowing direct computation of anomaly scores and segmentation of anomalous regions by conditioned distance functions. Simultaneously, the anomaly simulation strategy faithfully generates realistic defects and randomly synthesizes diverse anomaly shapes on normal samples, improving the generalization ability of the network in the case of limited anomaly samples." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.509, + 0.712, + 0.525 + ], + "angle": 0, + "content": "3.2 Preliminary of denoising diffusion probabilistic models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.784, + 0.565 + ], + "angle": 0, + "content": "A DDPM is inspired by the thermal diffusion process in an evolving thermodynamic system, which consists of a diffusion process and a reverse process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.785, + 0.61 + ], + "angle": 0, + "content": "The forward Markovian process gradually adds Gaussian noise to a clean sample \\(\\pmb{x}^{(0)}\\) from a data distribution \\(q(\\pmb{x}^{(0)})\\) and turns it into a Gaussian noise \\(\\pmb{x}^{(T)}\\), which is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.376, + 0.624, + 0.785, + 0.663 + ], + "angle": 0, + "content": "\\[\nq \\left(\\boldsymbol {x} ^ {(0)}, \\dots , \\boldsymbol {x} ^ {(T)}\\right) = \\prod_ {t = 1} ^ {T} q \\left(\\boldsymbol {x} ^ {(t)} \\mid \\boldsymbol {x} ^ {(t - 1)}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.675, + 0.784, + 0.738 + ], + "angle": 0, + "content": "where \\( q(\\pmb{x}^{(t)}|\\pmb{x}^{(t-1)}) = \\mathcal{N}(\\pmb{x}^{(t)};\\sqrt{1 - \\beta_t}\\pmb{x}^{(t-1)},\\beta_t\\pmb{I}) \\) is the Markov diffusion kernel, \\( t = 1,\\dots,T \\), \\( T \\) is the number of diffusion steps, and \\( \\beta_{t} \\) is a variance schedule. We have \\( q(\\pmb{x}^{(t)}|\\pmb{x}^{(0)}) = \\mathcal{N}(\\pmb{x}^{(t)};\\sqrt{\\overline{\\alpha}_t}\\pmb{x}^{(0)},(1 - \\overline{\\alpha}_t)\\pmb{I}) \\) by reparameterization with \\( \\alpha_{t} = 1 - \\beta_{t} \\), \\( \\overline{\\alpha}_{t} = \\prod_{s=0}^{t}\\alpha_{s} \\). \\( \\pmb{x}^{(t)} \\) can be sampled by" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.751, + 0.785, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} ^ {(t)} = \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} ^ {(0)} + \\epsilon \\sqrt {\\left(1 - \\bar {\\alpha} _ {t}\\right)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.784, + 0.809 + ], + "angle": 0, + "content": "where \\(\\pmb{\\epsilon}\\) is a standard Gaussian noise and \\(\\epsilon \\sim \\mathcal{N}(0, I)\\). When \\(T\\) is large enough, \\(\\pmb{x}^{(T)}\\) will eventually become a Gaussian noise." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.785, + 0.841 + ], + "angle": 0, + "content": "The reverse process is also a Markovian process that denoises over a series of steps to generate meaningful data from the target distribution \\( q(\\pmb{x}^{(0)}) \\). The" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "Z. Zhou et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.147, + 0.647, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.147, + 0.782, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.36, + 0.789, + 0.473 + ], + "angle": 0, + "content": "Fig. 2: Overall architecture of R3D-AD for shape reconstruction and anomaly detection of point cloud objects. Reconstruction training phase: The simulated anomalous \\(\\mathcal{P}_{\\mathrm{a}}^{(0)}\\) is generated resort to Patch-Gen from the input point cloud. It further fully masked as \\(\\mathcal{P}_{\\mathrm{a}}^{(T)}\\) while also encoded to latent shape embedding. The SWD decoder then explicitly reconstructs the anomaly-free object \\(\\mathcal{P}_{\\mathrm{r}}^{(0)}\\) with consistent spatial transform by conditionally generating point-level displacements \\(\\Delta^{(t)}\\) at each step of the inverse process. Detection testing phase: The test point cloud \\(\\mathcal{P}_{\\mathrm{a}}^{(0)}\\) is reconstructed to \\(\\mathcal{P}_{\\mathrm{r}}^{(0)}\\) with normal shape, and compared at a distance level to detect the anomalous region." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.502, + 0.789, + 0.533 + ], + "angle": 0, + "content": "inverse process denoises the noise \\(\\pmb{x}^{(T)}\\) from a distribution \\(p(\\pmb{x}^{(T)})\\), which is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.32, + 0.543, + 0.787, + 0.585 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\boldsymbol {x} ^ {(0)}, \\dots , \\boldsymbol {x} ^ {(T - 1)} \\mid \\boldsymbol {x} ^ {(T)}, \\boldsymbol {c}\\right) = \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(\\boldsymbol {x} ^ {(t - 1)} \\mid \\boldsymbol {x} ^ {(t)}, \\boldsymbol {c}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.597, + 0.789, + 0.645 + ], + "angle": 0, + "content": "where \\(p_{\\theta}(\\boldsymbol{x}^{(t-1)}|\\boldsymbol{x}^{(t)})\\), \\(\\boldsymbol{c} = \\mathcal{N}(\\boldsymbol{x}^{(t-1)};\\mu_{\\theta}(\\boldsymbol{x}^{(t)},t,\\boldsymbol{c}),\\sigma_t^2\\boldsymbol{I})\\), the mean \\(\\mu_{\\theta}(\\boldsymbol{x}^{(t)},t,\\boldsymbol{c})\\) is estimated by a neural network parameterized by \\(\\boldsymbol{\\theta}\\), \\(\\boldsymbol{c}\\) is the latent condition encoding, and \\(\\sigma_t^2\\) is a step-dependent variance. \\(\\mu_{\\theta}\\) can be reparameterized as" + }, + { + "type": "equation", + "bbox": [ + 0.322, + 0.657, + 0.787, + 0.69 + ], + "angle": 0, + "content": "\\[\n\\mu_ {\\theta} \\left(\\boldsymbol {x} ^ {(t)}, t, \\boldsymbol {c}\\right) = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\boldsymbol {x} ^ {(t)} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} ^ {(t)}, t, \\boldsymbol {c}\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.703, + 0.789, + 0.732 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\theta}(\\pmb{x}^{(t)}, t, \\pmb{c})\\) is a neural network utilized to denoise the Gaussian noise from \\(\\pmb{x}^{(T)}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.789, + 0.765 + ], + "angle": 0, + "content": "The training objective is minimized by training \\(\\pmb{\\epsilon}_{\\pmb{\\theta}}(\\pmb{x}^{(t)}, t, \\pmb{c})\\) to approximate \\(\\pmb{\\epsilon}\\). The training objective is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.243, + 0.777, + 0.787, + 0.798 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} _ {t \\sim [ 1: T ], \\boldsymbol {x} ^ {(0)} \\sim q (\\boldsymbol {x} ^ {(0)})}, \\epsilon \\sim \\mathcal {N} (0, I) \\| \\epsilon - \\epsilon_ {\\boldsymbol {\\theta}} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} ^ {(0)} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , t, \\boldsymbol {c}\\right) \\|, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.809, + 0.789, + 0.84 + ], + "angle": 0, + "content": "where \\( t \\) is sampled from the uniform distribution over \\( 1,2,\\dots,T \\), \\( q(\\pmb{x}^{(0)}) \\) is the distribution of \\( \\pmb{x}^{(0)} \\), and \\( \\epsilon \\) is the Gaussian noise." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.618, + 0.163 + ], + "angle": 0, + "content": "3.3 Diffusion-based 3D anomaly reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.346 + ], + "angle": 0, + "content": "We formulate the point cloud reconstruction task of the anomaly-free model as the conditional generation, which decodes the explicit displacement with the target distribution \\( q(\\mathcal{P}_{\\mathrm{r}}|\\mathcal{P}_{\\mathrm{a}}^{(T)},c) \\), where \\( c \\) is the decoding condition. The essential question of anomaly detection in this paper is how to conditional reconstruct anomaly-free shapes on the reference of input point clouds with different spatial transformations. Since there is a high similarity of global features between abnormal and normal samples during self-supervised reconstruction, the most immediate approach is to extract an efficient global feature from input to serve as an auxiliary conditional embedding for the denoising function \\( \\epsilon_{\\theta} \\). We implement the encoding of latent shape embedding \\( c \\) as a conditional input to guide reconstruction in the reverse diffusion process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.788, + 0.498 + ], + "angle": 0, + "content": "Latent shape embedding The feature encoder aims to encode the point cloud to the latent shape embedding \\( c \\) with high-level features for the conditional generation process. Different from other global-local extracting methods [37,40], we focus more on extracting global features, which characterize the semantic information of shape and pose of most anomaly-free regions in the point cloud. The feature encoder mainly consists of cascaded multi-layer perceptions (MLP) based on PointNet [5]. It implements max-pooling after mapping \\( \\mathcal{P}_{\\mathrm{a}}^{(0)} \\) to different dimensions and then compresses them to extract the global shape embedding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.525, + 0.788, + 0.801 + ], + "angle": 0, + "content": "Step-wise displacement decoding To achieve point cloud reconstruction with transformation consistency while preserving the structure of non-anomalous regions, our method injects latent shape embedding \\( \\pmb{c} \\) to the decoder at each step of the reverse diffusion process, as shown in Fig. 2. In principle, in the training phase, \\( \\epsilon_{\\theta} \\) learns the added Gaussian noise in the forward diffusion process by the decoder to model the conditional probability distribution. Conditionally generating target shapes from \\( N \\times 3 \\) Gaussian noise is a straightforward approach, but it is afflicted by the issues of reconstructing the point cloud details and transform consistency. Learning the relative deformation of points for anomalous objects is more efficient. Considering the mapping degradation of the vanilla autoencoder in the reconstruction training phase [22], we utilize the Gaussian noise of the forward process Eq. 2 to fully mask the point cloud object directly without blind spots, preventing the decoding process from receiving negative state shapes. The masked points \\( \\mathcal{P}_{\\mathrm{a}}^{(T)} \\) and latent shape embedding \\( \\pmb{c} \\in \\mathbb{R}^{256} \\) are as the inputs of the SWD decoder. The point-wise displacement vector \\( \\Delta^{(t)} \\) is generated at each step of the iterative process thus disentangling the prediction noise and the desired anomaly-free shape. The reverse process can be defined according to Eq. 3 and the displacement vector \\( \\Delta^{(t)} \\) can be represented by" + }, + { + "type": "equation", + "bbox": [ + 0.32, + 0.812, + 0.786, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\Delta^ {(t - 1)} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\Delta^ {(t)} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\theta}} \\left(\\Delta^ {(t)}, \\beta_ {t}, \\boldsymbol {c}\\right)\\right) + \\sigma \\boldsymbol {\\epsilon}, \\tag {6}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "Z. Zhou et al." + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.145, + 0.728, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.307, + 0.788, + 0.364 + ], + "angle": 0, + "content": "Fig. 3: Illustration of Patch-Gen, the 3D anomaly simulation strategy. The input normal point cloud is first randomly rotated. On the surface of the normalized cube, we randomly select viewpoints to find the nearest patch of points. The selected points are then transformed into irregular defects according to the specific deformation solution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.392, + 0.789, + 0.502 + ], + "angle": 0, + "content": "where \\(\\sigma\\) is the variance. A PointwiseNet is adopted for \\(\\epsilon_{\\theta}\\) to decode the \\(\\Delta^{(t-1)}\\) from the previous step and \\(c\\). \\(\\beta_{t}\\) is used to generate trigonometric position embedding \\(\\pmb{e}_{\\mathrm{p}} = (\\beta_{t}, \\sin(\\beta_{t}), \\cos(\\beta_{t}))\\). \\(\\pmb{e}_{\\mathrm{p}}\\) is concatenated with \\(c\\) and then fed into the concatenate-squash linear module of PointwiseNet with a residual function. The output reconstructed point cloud at the \\(t\\) step is \\(\\mathcal{P}_{\\mathrm{r}}^{(t)} = \\mathcal{P}_{\\mathrm{r}}^{(t+1)} + \\Delta^{(t)}\\). The registered original and reconstructed objects are distinguished from the anomalous shape by the anomaly scores based on the conditioned distance function." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.522, + 0.527, + 0.537 + ], + "angle": 0, + "content": "3.4 3D anomaly simulation strategy" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.546, + 0.788, + 0.681 + ], + "angle": 0, + "content": "Given that a small number of normal samples is not conducive for the model to learn diverse and essential features, we propose the Patch-Gen strategy to simulate the defects from anomaly-free shapes for training data augmentation. Patch-Gen encourages the reconstruction model to learn to detect irregularity, where the anomaly-free point clouds and their diverse anomaly patterns are integrated into training pairs and are utilized to learn the discrimination feature between normal and anomalous surfaces. The intuition is that the diversity of simulated negative samples forces our network to learn how to reconstruct anomaly-free shapes instead of memorizing their complete outfits." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.682, + 0.788, + 0.727 + ], + "angle": 0, + "content": "As shown in Fig. 3, the input normal sample is first randomly rotated. The random spatial rotation is designed to improve the generalization capability for test samples with very different spatial transformations, as defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.457, + 0.738, + 0.786, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} _ {\\mathrm {a}} = \\mathcal {P} \\cdot \\mathcal {R}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "where \\(\\mathcal{P}\\) is input normal sample point cloud and \\(\\mathcal{R} \\in \\mathbb{R}^{3 \\times 3}\\) is obtained by randomly selecting rotation angles for all three axes. In addition to global shape awareness of the model by the random rotation, we further perform a fine granularity of the anomaly simulation. We randomly take a viewpoint \\(\\mathcal{P}_v\\) from the surface of the cube. Therefore, the patch of nearest \\(N\\) points \\(\\mathcal{P}_n\\) from \\(\\mathcal{P}_a\\) can be" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.786, + 0.176 + ], + "angle": 0, + "content": "determined according to the \\(\\mathcal{P}_v\\). The shape augmentation scheme Patch-Gen is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.359, + 0.19, + 0.786, + 0.207 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} _ {n} = \\mathcal {P} _ {n} + S \\cdot \\text {n o r m a l i z e} \\left(\\mathcal {P} _ {n} - \\mathcal {P} _ {v}\\right) \\odot \\mathcal {T}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.217, + 0.784, + 0.277 + ], + "angle": 0, + "content": "where \\( \\text{nomalize} \\) represents a normalization operation on a vector, \\( S \\) is a predefined hyper-parameter that controls the scaling of the patch points, and \\( \\mathcal{T} \\) is the translation matrix originating from a Gaussian distribution. The \\( \\mathcal{P}_{\\mathrm{a}} \\) is finally obtained by only updating the patch region \\( \\mathcal{P}_n \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.278, + 0.785, + 0.339 + ], + "angle": 0, + "content": "With the proposed Patch-Gen, we can simulate the generation of multiple anomalies, which is mainly done by controlling \\(\\mathcal{T}\\). Bulge or sink can be generated by sorting \\(\\mathcal{T}\\) after sampling from the distribution, while damage can be generated by direct overlaying without manipulation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.36, + 0.414, + 0.377 + ], + "angle": 0, + "content": "3.5 Training objective" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.788, + 0.525 + ], + "angle": 0, + "content": "In the reconstruction task of the object with \\(N\\) points, the network learns a diffusion model with an \\(\\mathbb{R}^{N\\times 3}\\to \\mathbb{R}^{N\\times 3}\\) mapping relation. Iterative denoising under the semantic condition of point embedding realizes the prediction of point offsets. Concretely, the network is trained to learn the noise that needs to be eliminated to recover the anomaly-free shape with the \\(L_{2}\\) distance between the ground truth and the denoised reconstructed points. We make use of the mean squared error (MSE) loss as the primary reconstruction loss which evaluates the mean squared error of the element-wise distances between \\(\\mathcal{P}_{\\mathrm{a}}^{(0)}\\) and \\(\\mathcal{P}_{\\mathrm{r}}^{(0)}\\). The MSE training loss is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.364, + 0.535, + 0.787, + 0.577 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {P} _ {\\mathrm {a}}, \\mathcal {P} _ {\\mathrm {r}}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} p _ {\\mathrm {a}} \\in \\mathcal {P} _ {\\mathrm {a}}, p _ {\\mathrm {r}} \\in \\mathcal {P} _ {\\mathrm {r}} \\| p _ {\\mathrm {a}} - p _ {\\mathrm {r}} \\| ^ {2}. \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.596, + 0.377, + 0.613 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.627, + 0.336, + 0.641 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.652, + 0.788, + 0.775 + ], + "angle": 0, + "content": "Real3D-AD [19] is a 3D anomaly detection dataset based on real samples, exhibiting a higher point precision and spatial distance per point cloud. Each category contains 4 training samples and 100 test samples. The training set contains \\(360^{\\circ}\\) complete surface point clouds of the objects, which are obtained by manually calibrating and stitching the scans of multiple sides of the objects. The test samples are scans only one side with a huge difference from the training set. The distribution of the point clouds also varies among the total 12 categories, further deepening the detection difficulty compared to 2D scenes." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Anomaly-ShapeNet [18] is a 3D anomaly detection, crafted through modifications to the synthetic samples found in ShapeNetCorev2 [4]. It contains 40 diverse categories, featuring over 1600 samples of its complete surface point" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.128 + ], + "angle": 0, + "content": "Z. Zhou et al." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.145, + 0.783, + 0.337 + ], + "angle": 0, + "content": "
MethodBTF [11]M3DM [36]PatchCore [28]CPMF [3]Reg3D-AD [19]IMRNet [18]Ours
Feat.RawFPFHPointMAEFPFHPointMAEResNetPointMAEPointMAERaw
Airplane0.7300.5200.4340.8820.7260.7010.7160.7620.772
Candybar0.5390.6300.5520.5410.6630.5520.6850.7550.696
Car0.6470.5600.5410.5900.4980.5510.6970.7110.713
Chicken0.7890.4320.6830.8370.8270.5040.8520.7800.714
Diamond0.7070.5450.6020.5740.7830.5230.9000.9050.685
Duck0.6910.7840.4330.5460.4890.5820.5840.5170.909
Fish0.6020.5490.5400.6750.6300.5580.9150.8800.692
Gemstone0.6860.6480.6440.3700.3740.5890.4170.6740.665
Seahorse0.5960.7790.4950.5050.5390.7290.7620.6040.720
Shell0.3960.7540.6940.5890.5010.6530.5830.6650.840
Starfish0.5300.5750.5510.4410.5190.7000.5060.6740.701
Toffees0.7030.4620.4500.5650.5850.3900.8270.7740.703
Average0.6350.6030.5520.5930.5950.5860.7040.7250.734
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.339, + 0.785, + 0.366 + ], + "angle": 0, + "content": "Table 1: Image-level anomaly detection AUROC on Real3D-AD dataset. We highlight the best result in bold and the second best result with an underline." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.785, + 0.489 + ], + "angle": 0, + "content": "clouds. Each category's training set contains merely 4 samples, while the test sets are designed to assess the model's performance across both normal and a spectrum of abnormal samples. It widely increases the anomaly types while keeping the number of points the same as the previous studies, which places higher demands on the robustness and generality of the proposed algorithms." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.518, + 0.418, + 0.532 + ], + "angle": 0, + "content": "4.2 Evaluation metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.787, + 0.643 + ], + "angle": 0, + "content": "For image-level anomaly detection, the Area Under the Receiver Operating Curve (AUROC) is utilized in line with established practices. For the evaluation of pixel-level anomalies, the AUROC metric is similarly applied in the context of point segmentation accuracy. A value of 0.5 of the AUROC score denotes no discriminative capability (equivalent to random guessing), whereas a score of 1.0 indicates perfect discrimination between positive and negative classes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.672, + 0.455, + 0.687 + ], + "angle": 0, + "content": "4.3 Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Our methodology is implemented using PyTorch [26] with end-to-end training across the network. The optimization is performed using the Adam optimizer, starting at an initial learning of 0.001. The training process involves a total batch size of 128 across 40,000 iterations for comprehensive learning. All input point clouds undergo a preprocessing step where they are randomly downsampled to a fixed size of 4096 and 2048 points on Real3D-AD and Anomaly-ShapeNet, respectively. Additionally, we normalized these point clouds by setting their center of gravity as the origin of coordinates and scaling their dimensions to fall within the range of -1 to 1, optimizing for the diffusion process." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.144, + 0.784, + 0.369 + ], + "angle": 0, + "content": "
MethodBTF [11]M3DM [36]PatchCore [28]CPMF [3]Reg3D-AD [19]IMRNet [18]Ours
Feat.RawFPFHPointMAEFPFHPointMAEResNetPointMAEPointMAERaw
Ashtray0.5780.4200.5770.5870.5910.3530.5970.6710.833
Bag0.4100.5460.5370.5710.6010.6430.7060.6600.719
Bottle0.5580.4040.5840.6140.5880.4690.5690.6310.750
Bowl0.4700.5810.5790.5580.5470.6790.5480.6760.751
Bucket0.4690.5170.4050.5100.5770.5420.6810.6760.719
Cap0.5090.5620.5990.6450.5830.6010.6320.7040.726
Cup0.4620.5980.5480.5930.5830.4980.5240.7000.767
Eraser0.5250.7190.6270.6570.6770.6890.3430.5480.890
Headset0.4470.5050.5970.6100.6090.5510.5740.6980.767
Helmet0.5080.5690.4880.4650.4950.5320.4910.6030.704
Jar0.4200.4240.4410.4720.4830.6100.5920.7800.838
Microphone0.5630.6710.3570.3880.4880.5090.4140.7550.762
Shelf0.1640.6090.5640.4940.5230.6850.6880.6030.696
Tap0.5490.5530.7470.7600.4980.5280.6590.6860.818
Vase0.5170.4640.5340.5540.5820.5140.5760.6290.734
Average0.4930.5280.5520.5680.5620.5590.5720.6590.749
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.371, + 0.785, + 0.399 + ], + "angle": 0, + "content": "Table 2: Image-level anomaly detection AUROC on Anomaly-ShapeNet dataset. We highlight the best result in bold and the second best result with an underline." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.451, + 0.367, + 0.465 + ], + "angle": 0, + "content": "4.4 Main results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.498, + 0.791, + 0.529 + ], + "angle": 0, + "content": "We conduct experiments on Real3D-AD [19] based on real sampling and AnomalyShapeNet [18] based on simulation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.788, + 0.732 + ], + "angle": 0, + "content": "As shown in Table 1, we first compare the image-level AUROC metric with current cutting-edge 3D anomaly detection models on Real3D-AD. It shows that our method achieves the best performance using only raw point cloud data, while most of the existing methods use Fast Point Feature Histograms (FPFH) operator [30] or ShapeNet [4] pre-trained PointMAE [25] as feature extractor. Due to significant disparities in quantity, size, and distribution among different categories of point clouds in Real3D-AD, scoring variations across categories are more pronounced with other methods. For instance, numerous methods perform under 0.5 in certain categories, indicating their inadequacy in extracting meaningful features while facing challenging samples. In contrast, our method not only exhibits superior performance in 3D anomaly detection across the majority of categories but also achieves the best overall average across all categories. This demonstrates the strong generalizability and robustness of our approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We further evaluate our method on Anomaly-ShapeNet in Table 2, which encompasses a broader array of categories and a greater diversity of defect types. Compared to Real3D-AD, Anomaly-ShapeNet significantly enhances the diversity of defects, wherein the increased variety of defect types further escalates the complexity of detection tasks. The results highlight the exceptional performance of our method across all evaluated categories, demonstrating an average improvement of \\(9\\%\\) on AUROC relative to the approaches previously utilized." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "Z. Zhou et al." + }, + { + "type": "table", + "bbox": [ + 0.279, + 0.144, + 0.725, + 0.232 + ], + "angle": 0, + "content": "
ModelDiffusionConditionRelativePatch-GenI-AUROCP-AUROC
A✓✗✗✗0.5860.524
B✓✓✗✗0.6670.513
C✓✓✓✗0.7120.573
D✓✓✓✓0.7340.592
" + }, + { + "type": "table_caption", + "bbox": [ + 0.229, + 0.233, + 0.772, + 0.246 + ], + "angle": 0, + "content": "Table 3: Ablation studies for 3D adaptation components on Real3D-AD dataset." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.288, + 0.387, + 0.304 + ], + "angle": 0, + "content": "4.5 Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.787, + 0.375 + ], + "angle": 0, + "content": "To delve into the effect of individual components, we conduct ablation experiments on the Real3D-AD dataset. To fully demonstrate and compare the performance of the models, we report both image-level and pixel-level results with I-AUROC and P-AUROC, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.395, + 0.789, + 0.609 + ], + "angle": 0, + "content": "Main component Table 3 compares the performance of different variants from R3D-AD, which includes the influence of the denoising condition embedding, displacement-based reconstruction way, and the data augmentation strategy of Patch-Gen. Model A is denoted as our baseline, which is a vanilla DDPM model for point cloud reconstruction. Introducing a condition into the DDPM (Model B) significantly boosts performance, particularly in terms of I-AUROC, which sees a \\(13.8\\%\\) increase to 0.667. Model C, which predicts point displacements based on conditional DDPM, preserving detailed structural information while accommodating the relative displacement of points contributes to a notable \\(6.0\\%\\) gain in P-AUROC over Model B. Model D is trained under the conditions of shape embedding with the Patch-Gen strategy. Considering that the defective portion contains only a small portion of the original point cloud, we try to reconstruct the relative displacement in a way that preserves as much detail as possible, which is effective for both 3D anomaly detection and segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.786, + 0.659 + ], + "angle": 0, + "content": "Patch-Gen Table. 4 analyzes the influence of two key parameters in Patch-Gen: the selection points ratio and the scaling points factor." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.809 + ], + "angle": 0, + "content": "The selection points ratio from Table. 4a determines the proportion of points in the point cloud that are selected for transformation. Our findings suggest that a selection ratio of \\(1/32\\) achieves the best performance. It appears that this ratio provides a balanced trade-off between maintaining sufficient structure for anomaly detection and introducing enough variation to simulate anomalies effectively. Notably, as the ratio increases beyond \\(1/16\\), both I-AUROC and P-AUROC scores decrease in severity, since real defects only account for a small portion of the overall point cloud, a wide selection of points not only destroys the structure of the original point cloud, but also makes the distribution of the training and test sets inconsistent." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "The scaling points factor is the intensity of the random transformation applied to the selected points, as detailed in Table 4b. The optimal performance is" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.114, + 0.731, + 0.129 + ], + "angle": 0, + "content": "R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.312, + 0.143, + 0.498, + 0.229 + ], + "angle": 0, + "content": "
ratioI-AUROCP-AUROC
1/640.7160.584
1/320.7340.592
1/160.7270.579
1/80.6830.528
" + }, + { + "type": "table_caption", + "bbox": [ + 0.319, + 0.231, + 0.487, + 0.243 + ], + "angle": 0, + "content": "(a) Selection points ratio." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.144, + 0.698, + 0.23 + ], + "angle": 0, + "content": "
factorI-AUROCP-AUROC
0.10.7340.592
0.20.7270.572
0.40.7150.554
0.80.6610.517
" + }, + { + "type": "table_caption", + "bbox": [ + 0.516, + 0.231, + 0.679, + 0.243 + ], + "angle": 0, + "content": "(b) Scaling points factor." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.254, + 0.784, + 0.283 + ], + "angle": 0, + "content": "Table 4: Ablation studies for Patch-Gen implementation on Real3D-AD dataset. Default settings are marked in gray." + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.319, + 0.495, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.348, + 0.444, + 0.362, + 0.454 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.322, + 0.776, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.444, + 0.665, + 0.454 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.466, + 0.788, + 0.523 + ], + "angle": 0, + "content": "Fig. 4: Memory and time cost during inference on Real3D-AD dataset. (a) Memory usage comparison between different models. (b) 3D anomaly detection performance vs. frames per second on an NVIDIA RTX 3090 GPU. Our R3D-AD outperforms all previous methods on both accuracy and efficiency by a significant margin." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.555, + 0.788, + 0.631 + ], + "angle": 0, + "content": "observed at a scaling factor of 0.1, which implies that minor transformations are more effective for simulating anomalies without significantly altering the original data distribution. Larger scaling factors lead to a consistent decline in performance, underscoring the importance of subtle transformations for preserving the utility of the simulated anomalies for detection tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Memory and time cost As depicted in Figure 4, we evaluate the disparity in both storage consumption and inference time of our model under identical experimental conditions, compared to existing methods. Regarding memory usage, our approach demonstrates a marked superiority by employing raw coordinate features instead of FPFH or PointMAE features, significantly reducing the memory footprint. Since no memory bank exists, our method is also more space-efficient compared to BTF which also uses raw features. Moreover, our method eliminates the necessity to compare all the features in memory, substantially increasing operational efficiency. The implementation of Patch-Gen inherently bestows our model with exceptional robustness, enabling precise reconstruction of point clouds from various angles without the need for the time-intensive RANSAC alignment process required by Reg3D-AD." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "Z. Zhou et al." + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.145, + 0.778, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.337, + 0.788, + 0.38 + ], + "angle": 0, + "content": "Fig. 5: Qualitative analysis on Real3D-AD dataset and Anomaly-ShapeNet dataset. The anomaly map is obtained directly by calculating the differences between the input and reconstructed point clouds, where deeper colors represent more confidence." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.41, + 0.416, + 0.426 + ], + "angle": 0, + "content": "4.6 Qualitative results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.789, + 0.588 + ], + "angle": 0, + "content": "Figure 5 presents some qualitative outcomes, with varying shades of color indicating different levels of anomaly scores. We select several representative defective samples to demonstrate the robustness of our algorithm. The left four columns display samples from Real3D-AD, while the right four columns samples are from Anomaly-ShapeNet. The illustration reveals that our R3D-AD algorithm has precisely reconstructed the defective portions of the point cloud across various samples: the deep sink in the Seahorse sample, the concavity in the Bag sample, and the bulge in the Jar sample. Leveraging the accurately reconstructed point clouds, final point cloud segmentation maps are also produced, further evidencing the efficacy of our approach." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.611, + 0.36, + 0.627 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In this work, we presented R3D-AD, a novel reconstructive 3D anomaly detection model based on conditional diffusion. Our goal is to overcome the limitations faced by current 3D anomaly detection methods, such as the inefficiencies due to the memory bank module and low performance caused by incorrect rebuilds with MAE. To address these challenges, we leverage the diffusion process for full reconstruction, followed by a direct comparison between the input and the reconstructed point cloud to obtain the final anomaly score. The embedded latent variable that spans the decoding process, step-wisely generating point-level displacements from the noise to the target anomaly-free sample. We also propose Patch-Gen, a data augmentation tailored for point cloud anomaly simulation. Extensive experiments conducted on 3D anomaly benchmarks validate the superiority of our R3D-AD in comparison to state-of-the-art alternatives in terms of both accuracy and versatility." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.788, + 0.237 + ], + "angle": 0, + "content": "This work was supported in part by the Pioneer and Leading Goose R&D Program of Zhejiang (Grant No. 2022C01051), in part by the National Natural Science Foundation of China (Grant No. 52375271, 52275274), and in part by the Natural Science Foundation of Zhejiang Province (Grant No. LY23E050011)." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.255, + 0.323, + 0.272 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.284, + 0.785, + 0.311 + ], + "angle": 0, + "content": "1. Bae, J., Lee, J.H., Kim, S.: Pni: Industrial anomaly detection using position and neighborhood information. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.312, + 0.785, + 0.351 + ], + "angle": 0, + "content": "2. Bergmann, P., Löwe, S., Fauser, M., Sattlegger, D., Steger, C.: Improving unsupervised defect segmentation by applying structural similarity to autoencoders. In: VISIGRAPP (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.353, + 0.785, + 0.379 + ], + "angle": 0, + "content": "3. Cao, Y., Xu, X., Shen, W.: Complementary pseudo multimodal feature for point cloud anomaly detection. arXiv preprint (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.379, + 0.785, + 0.42 + ], + "angle": 0, + "content": "4. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., Xiao, J., Yi, L., Yu, F.: ShapeNet: An Information-Rich 3D Model Repository. arXiv preprint (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.421, + 0.785, + 0.447 + ], + "angle": 0, + "content": "5. Charles, R.Q., Su, H., Kaichun, M., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.448, + 0.785, + 0.473 + ], + "angle": 0, + "content": "6. Chu, R., Xie, E., Mo, S., Li, Z., Nießner, M., Fu, C.W., Jia, J.: Diffcomplete: Diffusion-based generative 3d shape completion. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.475, + 0.785, + 0.5 + ], + "angle": 0, + "content": "7. Dinh, L., Sohl-Dickstein, J., Bengio, S.: Density estimation using real nvp. In: ICLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.501, + 0.785, + 0.528 + ], + "angle": 0, + "content": "8. Gudovskiy, D., Ishizaka, S., Kozuka, K.: Cflow-ad: Real-time unsupervised anomaly detection with localization via conditional normalizing flows. In: WACV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.529, + 0.785, + 0.554 + ], + "angle": 0, + "content": "9. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.555, + 0.785, + 0.596 + ], + "angle": 0, + "content": "10. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., Salimans, T.: Imagen video: High definition video generation with diffusion models (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.597, + 0.785, + 0.623 + ], + "angle": 0, + "content": "11. Horwitz, E., Hoshen, Y.: Back to the feature: Classical 3d features are (almost) all you need for 3d anomaly detection. In: CVPRW (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.624, + 0.785, + 0.663 + ], + "angle": 0, + "content": "12. Hu, T., Zhang, J., Yi, R., Du, Y., Chen, X., Liu, L., Wang, Y., Wang, C.: Anomalydiffusion: Few-shot anomaly image generation with diffusion model. In: AAAI (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.664, + 0.785, + 0.69 + ], + "angle": 0, + "content": "13. Jonathan Ho, Ajay Jain, and Pieter Abbeel: Denoising diffusion probabilistic models. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.691, + 0.785, + 0.718 + ], + "angle": 0, + "content": "14. Kim, D., Park, C., Cho, S., Lee, S.: Fapm: Fast adaptive patch memory for real-time industrial anomaly detection. In: ICASSP (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.719, + 0.785, + 0.745 + ], + "angle": 0, + "content": "15. Kong, Z., Ping, W., Huang, J., Zhao, K., Catanzaro, B.: Diffwave: A versatile diffusion model for audio synthesis. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.746, + 0.785, + 0.772 + ], + "angle": 0, + "content": "16. Li, C.L., Sohn, K., Yoon, J., Pfister, T.: Cutpaste: Self-supervised learning for anomaly detection and localization. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.773, + 0.785, + 0.799 + ], + "angle": 0, + "content": "17. Li, M., Duan, Y., Zhou, J., Lu, J.: Diffusion-sdf: Text-to-shape via voxelized diffusion. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "18. Li, W., Xu, X., Gu, Y., Zheng, B., Gao, S., Wu, Y.: Towards scalable 3d anomaly detection and localization: A benchmark via 3d anomaly synthesis and a self-supervised learning network. arXiv preprint (2023)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.284, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "Z. Zhou et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "19. Liu, J., Xie, G., Li, X., Wang, J., Liu, Y., Wang, C., Zheng, F., et al.: Real3d-ad: A dataset of point cloud anomaly detection. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.202 + ], + "angle": 0, + "content": "20. Liu, Z., Feng, Y., Black, M.J., Nowrouzezahrai, D., Paull, L., Liu, W.: Meshdiffusion: Score-based generative 3d mesh modeling. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.229 + ], + "angle": 0, + "content": "21. Liu, Z., Tang, H., Lin, Y., Han, S.: Point-voxel cnn for efficient 3d deep learning. In: NeurIPS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.256 + ], + "angle": 0, + "content": "22. Lu, F., Yao, X., Fu, C., Jia, J.: Removing anomalies as noises for industrial defect localization. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.283 + ], + "angle": 0, + "content": "23. Luo, S., Hu, W.: Diffusion probabilistic models for 3d point cloud generation. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.785, + 0.31 + ], + "angle": 0, + "content": "24. Mo, S., Xie, E., Chu, R., Hong, L., Nießner, M., Li, Z.: Dit-3d: Exploring plain diffusion transformers for 3d shape generation. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.785, + 0.338 + ], + "angle": 0, + "content": "25. Pang, Y., Wang, W., Tay, F.E., Liu, W., Tian, Y., Yuan, L.: Masked autoencoders for point cloud self-supervised learning. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.785, + 0.379 + ], + "angle": 0, + "content": "26. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. NeurIPS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.405 + ], + "angle": 0, + "content": "27. Rezende, D.J., Mohamed, S.: Variational inference with normalizing flows. In: ICML (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.406, + 0.785, + 0.433 + ], + "angle": 0, + "content": "28. Roth, K., Pemula, L., Zepeda, J., Scholkopf, B., Brox, T., Gehler, P.: Towards total recall in industrial anomaly detection. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.785, + 0.459 + ], + "angle": 0, + "content": "29. Rudolph, M., Wandt, B., Rosenhahn, B.: Same same but differnet: Semi-supervised defect detection with normalizing flows. In: WACV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.487 + ], + "angle": 0, + "content": "30. Rusu, R.B., Blodow, N., Beetz, M.: Fast point feature histograms (fpfh) for 3d registration. In: ICRA (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.542 + ], + "angle": 0, + "content": "31. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Ghasemipour, S.K.S., Ayan, B.K., Mahdavi, S.S., Lopes, R.G., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.543, + 0.785, + 0.569 + ], + "angle": 0, + "content": "32. Schlüter, H.M., Tan, J., Hou, B., Kainz, B.: Natural synthetic anomalies for self-supervised anomaly detection and localization. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.785, + 0.595 + ], + "angle": 0, + "content": "33. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.597, + 0.785, + 0.636 + ], + "angle": 0, + "content": "34. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.664 + ], + "angle": 0, + "content": "35. Tailanian, M., Pardo, Å., Muse, P.: U-flow: A u-shaped normalizing flow for anomaly detection with unsupervised threshold. arXiv preprint (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.785, + 0.69 + ], + "angle": 0, + "content": "36. Wang, Y., Peng, J., Zhang, J., Yi, R., Wang, Y., Wang, C.: Multimodal industrial anomaly detection via hybrid fusion. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.785, + 0.718 + ], + "angle": 0, + "content": "37. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph cnn for learning on point clouds. TOG (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.719, + 0.785, + 0.745 + ], + "angle": 0, + "content": "38. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: Simmim: A simple framework for masked image modeling. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.746, + 0.785, + 0.785 + ], + "angle": 0, + "content": "39. Yu, J., Zheng, Y., Wang, X., Li, W., Wu, Y., Zhao, R., Wu, L.: Fastflow: Unsupervised anomaly detection and localization via 2d normalizing flows. arXiv preprint (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "40. Yu, X., Tang, L., Rao, Y., Huang, T., Zhou, J., Lu, J.: Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "41. Zavrtanik, V., Kristan, M., Skočaj, D.: Draem-a discriminatively trained reconstruction embedding for surface anomaly detection. In: ICCV (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "42. Zavrtanik, V., Kristan, M., Skočaj, D.: Reconstruction by inpainting for visual anomaly detection. Pattern Recognition (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.785, + 0.203 + ], + "angle": 0, + "content": "43. Zhang, X., Xu, M., Zhou, X.: Realnet: A feature selection network with realistic synthetic anomaly for anomaly detection. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.231 + ], + "angle": 0, + "content": "44. Zhou, L., Du, Y., Wu, J.: 3d shape generation and completion through point-voxel diffusion. In: ICCV (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.231 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/R3D-AD_ Reconstruction via Diffusion for 3D Anomaly Detection/a722039e-753c-4cd1-9dab-7069c8f25f9e_origin.pdf b/2024/R3D-AD_ Reconstruction via Diffusion for 3D Anomaly Detection/a722039e-753c-4cd1-9dab-7069c8f25f9e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b6963b0bcb9cd2277664674f6c25fd5bca987ce9 --- /dev/null +++ b/2024/R3D-AD_ Reconstruction via Diffusion for 3D Anomaly Detection/a722039e-753c-4cd1-9dab-7069c8f25f9e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:23c4241fc5df0a9c01e7d46c1fc6d5fab0f042d2bf97e8d6da1506b2d7bd6965 +size 5658502 diff --git a/2024/R3D-AD_ Reconstruction via Diffusion for 3D Anomaly Detection/full.md b/2024/R3D-AD_ Reconstruction via Diffusion for 3D Anomaly Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7faf66e49a13f9f1e82d3d18b0dffe1a5e023684 --- /dev/null +++ b/2024/R3D-AD_ Reconstruction via Diffusion for 3D Anomaly Detection/full.md @@ -0,0 +1,287 @@ +# R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection + +Zheyuan Zhou $^{1*}$ , Le Wang $^{1*}$ , Naiyu Fang $^{1,2}$ , Zili Wang $^{1}$ , Lemiao Qiu $^{1\boxtimes}$ , and Shuyou Zhang $^{1}$ + +$^{1}$ State Key Laboratory of Fluid Power & Mechatronic System, Zhejiang University, Hangzhou, China + +$^{2}$ S-Lab, Nanyang Technological University, Singapore, Singapore + +https://zhouzheyuan.github.io/r3d-ad + +Abstract. 3D anomaly detection plays a crucial role in monitoring parts for localized inherent defects in precision manufacturing. Embedding-based and reconstruction-based approaches are among the most popular and successful methods. However, there are two major challenges to the practical application of the current approaches: 1) the embedded models suffer the prohibitive computational and storage due to the memory bank structure; 2) the reconstructive models based on the MAE mechanism fail to detect anomalies in the unmasked regions. In this paper, we propose R3D-AD, reconstructing anomalous point clouds by diffusion model for precise 3D anomaly detection. Our approach capitalizes on the data distribution conversion of the diffusion process to entirely obscure the input's anomalous geometry. It step-wisely learns a strict point-level displacement behavior, which methodically corrects the aberrant points. To increase the generalization of the model, we further present a novel 3D anomaly simulation strategy named Patch-Gen to generate realistic and diverse defect shapes, which narrows the domain gap between training and testing. Our R3D-AD ensures a uniform spatial transformation, which allows straightforwardly generating anomaly results by distance comparison. Extensive experiments show that our R3D-AD outperforms previous state-of-the-art methods, achieving $73.4\%$ Image-level AUROC on the Real3D-AD dataset and $74.9\%$ Image-level AUROC on the Anomaly-ShapeNet dataset with an exceptional efficiency. + +Keywords: 3D anomaly detection, industrial applications, 3D reconstruction, self-supervised learning + +# 1 Introduction + +Anomaly detection aims to identify instances containing anomalies and to precisely locate the specific positions of defects. This task is extensively applied across multiple fields and plays a crucial role in quality control within industrial production [28]. 3D anomaly detection [19] has emerged due to its intrinsic + +![](images/3f369856b866b0810a85c9855efae2baf59879e5aeab2e0a8a889bc771d06180.jpg) +(a) Embedded model + +![](images/0fd877a4c795486bb0d559134332b65e7c05ea94fdd0ea8529ca7ec254083f45.jpg) +(b) Reconstructive model +Fig. 1: Comparison of architectures. (a) Embedded model encodes the input $\mathcal{X}$ into features and stores them in the memory bank during training. The anomaly map $\mathcal{M}$ is obtained by comparing the test features with all the features in the memory bank. (b) Reconstructive model is trained by minimizing the loss between its input $\mathcal{X}$ and the output $\widehat{\mathcal{X}}$ . The anomaly map $\mathcal{M}$ is obtained by comparing the test phase input with its corresponding reconstruction target. + +modality superior for avoiding blind spots in advanced processing and precision manufacturing. However, the discrete and disordered data form of point clouds makes it more difficult to acquire features compared to images. With the scarcity of anomalies, 3D anomaly detection also faces the problem of domain shift while only normal data are presented during training. The presence of these issues underscores the necessity and urgency of devising an efficient framework for the 3D anomaly detection task. + +Similar to traditional 2D anomaly detection [28,42], current 3D anomaly detection can be primarily categorized into embedding-based and reconstruction-based, as illustrated in Fig. 1. The embedding-based methods involve mapping features extracted with a pre-trained encoder onto a normal distribution for learning. Distributions that do not fall within the interval are classified as anomalies. Most existing 3D anomaly detection methods are based on a memory bank mechanism [3, 11, 19, 36], which stores some representative features during the training phase to implicitly construct a feature distribution. In the testing phase, the presence of anomalies is determined by calculating the Euclidean distance between the input test object and all template point clouds stored in memory. The reconstruction-based methods train a network capable of accurately reconstructing normal point clouds, under the presumption that anomalous point clouds will not be effectively reconstructed since they are not included during training. The anomaly map is produced through the comparison of discrepancies between the input point cloud and its reconstruction. IMRNet [18] employs PointMAE [25] to reconstruct the input in several iterations, getting the final anomaly map by calculating the explicit spatial coordinate differences and implicit deep feature differences of the point cloud, respectively. + +However, existing methods face two key issues, high resource cost and irreparable reconstruction. Firstly, methods based on the memory bank [3,11,19, 36] store all features from the training phase, each test point cloud needs to be compared with all samples in the memory bank, significantly increasing memory + +overhead and inference time costs. This makes such methods almost inapplicable in real industrial production lines due to their inefficiency. Secondly, masked autoencoder (MAE) mechanism [9,25,38] only reconstructs the masked portions of the input, defects within unmasked portions may be preserved. This contradicts the fundamental assumption of detecting anomalies by comparing the original defect-containing point cloud with a reconstructed anomaly-free version. These methods inevitably lead to incorrect reconstructions, undermining their effectiveness in accurately localizing defects. + +We propose R3D-AD, a novel 3D anomaly detection method that does not suffer from the space burden and time endurance in memory-based embedded models nor the anomaly unmasking probability in the MAE-based reconstructive models. In contrast to PointMAE, one of our key insights is to perform undifferentiated masking for 3D objects via the noise diffusion mechanism, which maximizes the preservation of anomaly-free shapes and reconstructs abnormal regions. In the reparameterized diffusion process, one-step full mask and reconstruction are achieved by converting the point cloud distribution, instead of the multiple iterative method [18]. We hypothesize that anomaly detection verifies the gap between the reconstructed shapes and the positive samples by learning point movement. Specifically, for input models with arbitrary anomalies, we encode them as latent shape embeddings as decoding conditions and explicitly control the point cloud reconstruction process by step-wise displacements (SWD) decoding. The shape embedding harbors abundant global features and makes it easier to train the network without dwelling on the introduction of local anomaly details. Another key to our approach is to implement a controllable method of point-wise displacement during the diffusion process to refine the point cloud deformation iteratively. We propose to inject latent shape embedding into each step of the inverse denoising process, which drives the anomalous regions to converge to a smooth surface. We further adopted a 3D anomaly simulation strategy Patch-Gen to address the limitations of the dataset, which generates abundant defectives by producing spatial irregularity that is faithful to the real scene, including bulges, sinks, etc. This point cloud data augmentation encourages the self-supervised model to reconstruct more realistic anomaly-free shapes when facing the actual anomaly. + +To the best of our knowledge, this is the very first attempt at exploring diffusion in reconstruction-based 3D anomaly detection. Our main contributions are summarized as follows: (i) We introduce a novel framework, termed R3D-AD, which performs a one-step full mask and anomaly-free reconstruction for fast and accurate 3D anomaly detection. (ii) We propose to learn the step-wise displacement in the reverse diffusion process to explicitly control the reconstruction of anomalous shapes. (iii) We introduce a 3D anomaly simulation strategy named Patch-Gen to address the limitation of the data anomaly patterns and improve the reconstruction performance in a supervised setting. (iv) Extensive experiments demonstrate that our R3D-AD has achieved state-of-the-art performance on both Real3D-AD and Anomaly-ShapeNet datasets. + +# 2 Related work + +# 2.1 2D Anomaly Detection + +Anomaly detection has received increasing attention from researchers in recent years, and many new methods have been proposed to address the problem. Flow-based methods [8, 29, 35, 39] use learned distributions and flow's bijective properties to spot defects, while Memory-based approaches [1, 14, 28] gauge anomaly scores by contrasting test sample features with memory bank-stored norms. Reconstruction-based models [2, 41, 42] flag anomalies by comparing inputs to their online reconstructions. Recent works [12, 16, 32, 43] augment the anomaly detection datasets with generated synthetic anomalies to compensate for the negative example scarcity problem. + +# 2.2 3D Anomaly Detection + +This field lags behind the development of 2D anomaly detection since 3D data are harder to obtain, while point cloud data are sparser and contain more noise than image data. BTF [11] integration of handcrafted 3D descriptors with classic 2D method PatchCore [28], constructing a basic framework for 3D anomaly detection. M3DM [36] advances the field by separately analyzing features from point clouds and RGB images, then merging these for improved decision-making. CPMF [3] converts point clouds into two-dimensional images from multiple angles, extracting additional features from these images with a pre-trained network, and enhancing detection capabilities through information fusion. Reg3D-AD [19] develops a registration-based method, the RANSAC algorithm was used to align each sample before comparing it to the stored template during the test phase. IMRNet [18] trains a PointMAE [25] to reconstruct anomaly-free samples and identifies anomalies by juxtaposing the reconstructed point cloud against the initial input. Many of these use memory banks to store the features of the training samples or require multiple iterations to restore points. Unlike previous methods, our approach requires only one step of reconstruction and has significant advantages in both time and space efficiency. + +# 2.3 Diffusion Models + +Diffusion models have proven their effectiveness in several generative tasks, such as image generation [31], speech generation [15], and video generation [10]. Denoising Diffusion Probabilistic Models (DDPMs) [13, 33, 34] employ a forward noisng mechanism, incrementally integrating Gaussian noise into images, alongside a reverse process meticulously trained to counteract the forward mechanism. Denoise AD [22] conducts DDPM for reconstructing within the features space, generating images that contain less noise. In recent years, many studies [6, 17, 20, 24] have attempted to use the diffusion model to explore the 3D reconstruction task. DPM [23] incorporates a shape latent variable to encapsulate the geometric intricacies of 3D shapes, it distinctively models this variable's + +distribution utilizing Normalizing Flows [7, 27]. PVD [44] utilizes PVCNNs [21] for the point-voxel representation of 3D shapes and integrates structured locality into point clouds. This innovative approach leverages the strengths of both point and voxel representations, optimizing the model's ability to capture the intricate spatial hierarchies and local geometries within 3D objects. Since diffusion-based reconstruction recovers the target shape from complete noise, the dilemma of reconstructing only the masked region in the MAE [9] mechanism does not exist. + +# 3 Method + +# 3.1 Overview + +We model the anomaly detection problem as mapping an anomalous point cloud $\mathcal{P}_{\mathrm{a}} \in \mathbb{R}^{N \times 3}$ to a positive shape with which it is aligned. The framework of R3D-AD is shown in Fig. 2, where the simulated anomalous shapes are reconstructed in a self-supervised setting in the training phase and then compared with the original input to detect anomalies. The reconstructed anomaly-free model is aligned with the input, thus allowing direct computation of anomaly scores and segmentation of anomalous regions by conditioned distance functions. Simultaneously, the anomaly simulation strategy faithfully generates realistic defects and randomly synthesizes diverse anomaly shapes on normal samples, improving the generalization ability of the network in the case of limited anomaly samples. + +# 3.2 Preliminary of denoising diffusion probabilistic models + +A DDPM is inspired by the thermal diffusion process in an evolving thermodynamic system, which consists of a diffusion process and a reverse process. + +The forward Markovian process gradually adds Gaussian noise to a clean sample $\pmb{x}^{(0)}$ from a data distribution $q(\pmb{x}^{(0)})$ and turns it into a Gaussian noise $\pmb{x}^{(T)}$ , which is defined as + +$$ +q \left(\boldsymbol {x} ^ {(0)}, \dots , \boldsymbol {x} ^ {(T)}\right) = \prod_ {t = 1} ^ {T} q \left(\boldsymbol {x} ^ {(t)} \mid \boldsymbol {x} ^ {(t - 1)}\right), \tag {1} +$$ + +where $q(\pmb{x}^{(t)}|\pmb{x}^{(t-1)}) = \mathcal{N}(\pmb{x}^{(t)};\sqrt{1 - \beta_t}\pmb{x}^{(t-1)},\beta_t\pmb{I})$ is the Markov diffusion kernel, $t = 1,\dots,T$ , $T$ is the number of diffusion steps, and $\beta_{t}$ is a variance schedule. We have $q(\pmb{x}^{(t)}|\pmb{x}^{(0)}) = \mathcal{N}(\pmb{x}^{(t)};\sqrt{\overline{\alpha}_t}\pmb{x}^{(0)},(1 - \overline{\alpha}_t)\pmb{I})$ by reparameterization with $\alpha_{t} = 1 - \beta_{t}$ , $\overline{\alpha}_{t} = \prod_{s=0}^{t}\alpha_{s}$ . $\pmb{x}^{(t)}$ can be sampled by + +$$ +\boldsymbol {x} ^ {(t)} = \sqrt {\bar {\alpha} _ {t}} \boldsymbol {x} ^ {(0)} + \epsilon \sqrt {\left(1 - \bar {\alpha} _ {t}\right)}, \tag {2} +$$ + +where $\pmb{\epsilon}$ is a standard Gaussian noise and $\epsilon \sim \mathcal{N}(0, I)$ . When $T$ is large enough, $\pmb{x}^{(T)}$ will eventually become a Gaussian noise. + +The reverse process is also a Markovian process that denoises over a series of steps to generate meaningful data from the target distribution $q(\pmb{x}^{(0)})$ . The + +![](images/e69d491b539e63f2664d9ac55a9a05d526ce672ebe2881b232452ca0d303e8ab.jpg) +Fig. 2: Overall architecture of R3D-AD for shape reconstruction and anomaly detection of point cloud objects. Reconstruction training phase: The simulated anomalous $\mathcal{P}_{\mathrm{a}}^{(0)}$ is generated resort to Patch-Gen from the input point cloud. It further fully masked as $\mathcal{P}_{\mathrm{a}}^{(T)}$ while also encoded to latent shape embedding. The SWD decoder then explicitly reconstructs the anomaly-free object $\mathcal{P}_{\mathrm{r}}^{(0)}$ with consistent spatial transform by conditionally generating point-level displacements $\Delta^{(t)}$ at each step of the inverse process. Detection testing phase: The test point cloud $\mathcal{P}_{\mathrm{a}}^{(0)}$ is reconstructed to $\mathcal{P}_{\mathrm{r}}^{(0)}$ with normal shape, and compared at a distance level to detect the anomalous region. + +![](images/c912d14ec7b9122444326834473d856d05cf95cf04711dca06a9231d400e89b1.jpg) + +inverse process denoises the noise $\pmb{x}^{(T)}$ from a distribution $p(\pmb{x}^{(T)})$ , which is defined as + +$$ +p _ {\theta} \left(\boldsymbol {x} ^ {(0)}, \dots , \boldsymbol {x} ^ {(T - 1)} \mid \boldsymbol {x} ^ {(T)}, \boldsymbol {c}\right) = \prod_ {t = 1} ^ {T} p _ {\theta} \left(\boldsymbol {x} ^ {(t - 1)} \mid \boldsymbol {x} ^ {(t)}, \boldsymbol {c}\right), \tag {3} +$$ + +where $p_{\theta}(\boldsymbol{x}^{(t-1)}|\boldsymbol{x}^{(t)})$ , $\boldsymbol{c} = \mathcal{N}(\boldsymbol{x}^{(t-1)};\mu_{\theta}(\boldsymbol{x}^{(t)},t,\boldsymbol{c}),\sigma_t^2\boldsymbol{I})$ , the mean $\mu_{\theta}(\boldsymbol{x}^{(t)},t,\boldsymbol{c})$ is estimated by a neural network parameterized by $\boldsymbol{\theta}$ , $\boldsymbol{c}$ is the latent condition encoding, and $\sigma_t^2$ is a step-dependent variance. $\mu_{\theta}$ can be reparameterized as + +$$ +\mu_ {\theta} \left(\boldsymbol {x} ^ {(t)}, t, \boldsymbol {c}\right) = \frac {1}{\sqrt {\alpha_ {t}}} \left(\boldsymbol {x} ^ {(t)} - \frac {\beta_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \boldsymbol {\epsilon} _ {\boldsymbol {\theta}} \left(\boldsymbol {x} ^ {(t)}, t, \boldsymbol {c}\right)\right), \tag {4} +$$ + +where $\epsilon_{\theta}(\pmb{x}^{(t)}, t, \pmb{c})$ is a neural network utilized to denoise the Gaussian noise from $\pmb{x}^{(T)}$ . + +The training objective is minimized by training $\pmb{\epsilon}_{\pmb{\theta}}(\pmb{x}^{(t)}, t, \pmb{c})$ to approximate $\pmb{\epsilon}$ . The training objective is defined as + +$$ +\mathcal {L} = \mathbb {E} _ {t \sim [ 1: T ], \boldsymbol {x} ^ {(0)} \sim q (\boldsymbol {x} ^ {(0)})}, \epsilon \sim \mathcal {N} (0, I) \| \epsilon - \epsilon_ {\boldsymbol {\theta}} \left(\sqrt {\bar {\alpha} _ {t}} \boldsymbol {x} ^ {(0)} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon , t, \boldsymbol {c}\right) \|, \tag {5} +$$ + +where $t$ is sampled from the uniform distribution over $1,2,\dots,T$ , $q(\pmb{x}^{(0)})$ is the distribution of $\pmb{x}^{(0)}$ , and $\epsilon$ is the Gaussian noise. + +# 3.3 Diffusion-based 3D anomaly reconstruction + +We formulate the point cloud reconstruction task of the anomaly-free model as the conditional generation, which decodes the explicit displacement with the target distribution $q(\mathcal{P}_{\mathrm{r}}|\mathcal{P}_{\mathrm{a}}^{(T)},c)$ , where $c$ is the decoding condition. The essential question of anomaly detection in this paper is how to conditional reconstruct anomaly-free shapes on the reference of input point clouds with different spatial transformations. Since there is a high similarity of global features between abnormal and normal samples during self-supervised reconstruction, the most immediate approach is to extract an efficient global feature from input to serve as an auxiliary conditional embedding for the denoising function $\epsilon_{\theta}$ . We implement the encoding of latent shape embedding $c$ as a conditional input to guide reconstruction in the reverse diffusion process. + +Latent shape embedding The feature encoder aims to encode the point cloud to the latent shape embedding $c$ with high-level features for the conditional generation process. Different from other global-local extracting methods [37,40], we focus more on extracting global features, which characterize the semantic information of shape and pose of most anomaly-free regions in the point cloud. The feature encoder mainly consists of cascaded multi-layer perceptions (MLP) based on PointNet [5]. It implements max-pooling after mapping $\mathcal{P}_{\mathrm{a}}^{(0)}$ to different dimensions and then compresses them to extract the global shape embedding. + +Step-wise displacement decoding To achieve point cloud reconstruction with transformation consistency while preserving the structure of non-anomalous regions, our method injects latent shape embedding $\pmb{c}$ to the decoder at each step of the reverse diffusion process, as shown in Fig. 2. In principle, in the training phase, $\epsilon_{\theta}$ learns the added Gaussian noise in the forward diffusion process by the decoder to model the conditional probability distribution. Conditionally generating target shapes from $N \times 3$ Gaussian noise is a straightforward approach, but it is afflicted by the issues of reconstructing the point cloud details and transform consistency. Learning the relative deformation of points for anomalous objects is more efficient. Considering the mapping degradation of the vanilla autoencoder in the reconstruction training phase [22], we utilize the Gaussian noise of the forward process Eq. 2 to fully mask the point cloud object directly without blind spots, preventing the decoding process from receiving negative state shapes. The masked points $\mathcal{P}_{\mathrm{a}}^{(T)}$ and latent shape embedding $\pmb{c} \in \mathbb{R}^{256}$ are as the inputs of the SWD decoder. The point-wise displacement vector $\Delta^{(t)}$ is generated at each step of the iterative process thus disentangling the prediction noise and the desired anomaly-free shape. The reverse process can be defined according to Eq. 3 and the displacement vector $\Delta^{(t)}$ can be represented by + +$$ +\Delta^ {(t - 1)} = \frac {1}{\sqrt {\alpha_ {t}}} \left(\Delta^ {(t)} - \frac {\beta_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \boldsymbol {\epsilon} _ {\boldsymbol {\theta}} \left(\Delta^ {(t)}, \beta_ {t}, \boldsymbol {c}\right)\right) + \sigma \boldsymbol {\epsilon}, \tag {6} +$$ + +![](images/e024759dbdd519ebdbac73a91ea4561af0c2d82ba80c259a9a900488aea45250.jpg) +Fig. 3: Illustration of Patch-Gen, the 3D anomaly simulation strategy. The input normal point cloud is first randomly rotated. On the surface of the normalized cube, we randomly select viewpoints to find the nearest patch of points. The selected points are then transformed into irregular defects according to the specific deformation solution. + +where $\sigma$ is the variance. A PointwiseNet is adopted for $\epsilon_{\theta}$ to decode the $\Delta^{(t-1)}$ from the previous step and $c$ . $\beta_{t}$ is used to generate trigonometric position embedding $\pmb{e}_{\mathrm{p}} = (\beta_{t}, \sin(\beta_{t}), \cos(\beta_{t}))$ . $\pmb{e}_{\mathrm{p}}$ is concatenated with $c$ and then fed into the concatenate-squash linear module of PointwiseNet with a residual function. The output reconstructed point cloud at the $t$ step is $\mathcal{P}_{\mathrm{r}}^{(t)} = \mathcal{P}_{\mathrm{r}}^{(t+1)} + \Delta^{(t)}$ . The registered original and reconstructed objects are distinguished from the anomalous shape by the anomaly scores based on the conditioned distance function. + +# 3.4 3D anomaly simulation strategy + +Given that a small number of normal samples is not conducive for the model to learn diverse and essential features, we propose the Patch-Gen strategy to simulate the defects from anomaly-free shapes for training data augmentation. Patch-Gen encourages the reconstruction model to learn to detect irregularity, where the anomaly-free point clouds and their diverse anomaly patterns are integrated into training pairs and are utilized to learn the discrimination feature between normal and anomalous surfaces. The intuition is that the diversity of simulated negative samples forces our network to learn how to reconstruct anomaly-free shapes instead of memorizing their complete outfits. + +As shown in Fig. 3, the input normal sample is first randomly rotated. The random spatial rotation is designed to improve the generalization capability for test samples with very different spatial transformations, as defined by: + +$$ +\mathcal {P} _ {\mathrm {a}} = \mathcal {P} \cdot \mathcal {R}, \tag {7} +$$ + +where $\mathcal{P}$ is input normal sample point cloud and $\mathcal{R} \in \mathbb{R}^{3 \times 3}$ is obtained by randomly selecting rotation angles for all three axes. In addition to global shape awareness of the model by the random rotation, we further perform a fine granularity of the anomaly simulation. We randomly take a viewpoint $\mathcal{P}_v$ from the surface of the cube. Therefore, the patch of nearest $N$ points $\mathcal{P}_n$ from $\mathcal{P}_a$ can be + +determined according to the $\mathcal{P}_v$ . The shape augmentation scheme Patch-Gen is defined as follows: + +$$ +\mathcal {P} _ {n} = \mathcal {P} _ {n} + S \cdot \text {n o r m a l i z e} \left(\mathcal {P} _ {n} - \mathcal {P} _ {v}\right) \odot \mathcal {T}, \tag {8} +$$ + +where $\text{nomalize}$ represents a normalization operation on a vector, $S$ is a predefined hyper-parameter that controls the scaling of the patch points, and $\mathcal{T}$ is the translation matrix originating from a Gaussian distribution. The $\mathcal{P}_{\mathrm{a}}$ is finally obtained by only updating the patch region $\mathcal{P}_n$ . + +With the proposed Patch-Gen, we can simulate the generation of multiple anomalies, which is mainly done by controlling $\mathcal{T}$ . Bulge or sink can be generated by sorting $\mathcal{T}$ after sampling from the distribution, while damage can be generated by direct overlaying without manipulation. + +# 3.5 Training objective + +In the reconstruction task of the object with $N$ points, the network learns a diffusion model with an $\mathbb{R}^{N\times 3}\to \mathbb{R}^{N\times 3}$ mapping relation. Iterative denoising under the semantic condition of point embedding realizes the prediction of point offsets. Concretely, the network is trained to learn the noise that needs to be eliminated to recover the anomaly-free shape with the $L_{2}$ distance between the ground truth and the denoised reconstructed points. We make use of the mean squared error (MSE) loss as the primary reconstruction loss which evaluates the mean squared error of the element-wise distances between $\mathcal{P}_{\mathrm{a}}^{(0)}$ and $\mathcal{P}_{\mathrm{r}}^{(0)}$ . The MSE training loss is formulated as: + +$$ +\mathcal {L} _ {\mathcal {P} _ {\mathrm {a}}, \mathcal {P} _ {\mathrm {r}}} = \frac {1}{N} \sum_ {i = 1} ^ {N} p _ {\mathrm {a}} \in \mathcal {P} _ {\mathrm {a}}, p _ {\mathrm {r}} \in \mathcal {P} _ {\mathrm {r}} \| p _ {\mathrm {a}} - p _ {\mathrm {r}} \| ^ {2}. \tag {9} +$$ + +# 4 Experiments + +# 4.1 Datasets + +Real3D-AD [19] is a 3D anomaly detection dataset based on real samples, exhibiting a higher point precision and spatial distance per point cloud. Each category contains 4 training samples and 100 test samples. The training set contains $360^{\circ}$ complete surface point clouds of the objects, which are obtained by manually calibrating and stitching the scans of multiple sides of the objects. The test samples are scans only one side with a huge difference from the training set. The distribution of the point clouds also varies among the total 12 categories, further deepening the detection difficulty compared to 2D scenes. + +Anomaly-ShapeNet [18] is a 3D anomaly detection, crafted through modifications to the synthetic samples found in ShapeNetCorev2 [4]. It contains 40 diverse categories, featuring over 1600 samples of its complete surface point + +
MethodBTF [11]M3DM [36]PatchCore [28]CPMF [3]Reg3D-AD [19]IMRNet [18]Ours
Feat.RawFPFHPointMAEFPFHPointMAEResNetPointMAEPointMAERaw
Airplane0.7300.5200.4340.8820.7260.7010.7160.7620.772
Candybar0.5390.6300.5520.5410.6630.5520.6850.7550.696
Car0.6470.5600.5410.5900.4980.5510.6970.7110.713
Chicken0.7890.4320.6830.8370.8270.5040.8520.7800.714
Diamond0.7070.5450.6020.5740.7830.5230.9000.9050.685
Duck0.6910.7840.4330.5460.4890.5820.5840.5170.909
Fish0.6020.5490.5400.6750.6300.5580.9150.8800.692
Gemstone0.6860.6480.6440.3700.3740.5890.4170.6740.665
Seahorse0.5960.7790.4950.5050.5390.7290.7620.6040.720
Shell0.3960.7540.6940.5890.5010.6530.5830.6650.840
Starfish0.5300.5750.5510.4410.5190.7000.5060.6740.701
Toffees0.7030.4620.4500.5650.5850.3900.8270.7740.703
Average0.6350.6030.5520.5930.5950.5860.7040.7250.734
+ +Table 1: Image-level anomaly detection AUROC on Real3D-AD dataset. We highlight the best result in bold and the second best result with an underline. + +clouds. Each category's training set contains merely 4 samples, while the test sets are designed to assess the model's performance across both normal and a spectrum of abnormal samples. It widely increases the anomaly types while keeping the number of points the same as the previous studies, which places higher demands on the robustness and generality of the proposed algorithms. + +# 4.2 Evaluation metrics + +For image-level anomaly detection, the Area Under the Receiver Operating Curve (AUROC) is utilized in line with established practices. For the evaluation of pixel-level anomalies, the AUROC metric is similarly applied in the context of point segmentation accuracy. A value of 0.5 of the AUROC score denotes no discriminative capability (equivalent to random guessing), whereas a score of 1.0 indicates perfect discrimination between positive and negative classes. + +# 4.3 Implementation details + +Our methodology is implemented using PyTorch [26] with end-to-end training across the network. The optimization is performed using the Adam optimizer, starting at an initial learning of 0.001. The training process involves a total batch size of 128 across 40,000 iterations for comprehensive learning. All input point clouds undergo a preprocessing step where they are randomly downsampled to a fixed size of 4096 and 2048 points on Real3D-AD and Anomaly-ShapeNet, respectively. Additionally, we normalized these point clouds by setting their center of gravity as the origin of coordinates and scaling their dimensions to fall within the range of -1 to 1, optimizing for the diffusion process. + +
MethodBTF [11]M3DM [36]PatchCore [28]CPMF [3]Reg3D-AD [19]IMRNet [18]Ours
Feat.RawFPFHPointMAEFPFHPointMAEResNetPointMAEPointMAERaw
Ashtray0.5780.4200.5770.5870.5910.3530.5970.6710.833
Bag0.4100.5460.5370.5710.6010.6430.7060.6600.719
Bottle0.5580.4040.5840.6140.5880.4690.5690.6310.750
Bowl0.4700.5810.5790.5580.5470.6790.5480.6760.751
Bucket0.4690.5170.4050.5100.5770.5420.6810.6760.719
Cap0.5090.5620.5990.6450.5830.6010.6320.7040.726
Cup0.4620.5980.5480.5930.5830.4980.5240.7000.767
Eraser0.5250.7190.6270.6570.6770.6890.3430.5480.890
Headset0.4470.5050.5970.6100.6090.5510.5740.6980.767
Helmet0.5080.5690.4880.4650.4950.5320.4910.6030.704
Jar0.4200.4240.4410.4720.4830.6100.5920.7800.838
Microphone0.5630.6710.3570.3880.4880.5090.4140.7550.762
Shelf0.1640.6090.5640.4940.5230.6850.6880.6030.696
Tap0.5490.5530.7470.7600.4980.5280.6590.6860.818
Vase0.5170.4640.5340.5540.5820.5140.5760.6290.734
Average0.4930.5280.5520.5680.5620.5590.5720.6590.749
+ +Table 2: Image-level anomaly detection AUROC on Anomaly-ShapeNet dataset. We highlight the best result in bold and the second best result with an underline. + +# 4.4 Main results + +We conduct experiments on Real3D-AD [19] based on real sampling and AnomalyShapeNet [18] based on simulation. + +As shown in Table 1, we first compare the image-level AUROC metric with current cutting-edge 3D anomaly detection models on Real3D-AD. It shows that our method achieves the best performance using only raw point cloud data, while most of the existing methods use Fast Point Feature Histograms (FPFH) operator [30] or ShapeNet [4] pre-trained PointMAE [25] as feature extractor. Due to significant disparities in quantity, size, and distribution among different categories of point clouds in Real3D-AD, scoring variations across categories are more pronounced with other methods. For instance, numerous methods perform under 0.5 in certain categories, indicating their inadequacy in extracting meaningful features while facing challenging samples. In contrast, our method not only exhibits superior performance in 3D anomaly detection across the majority of categories but also achieves the best overall average across all categories. This demonstrates the strong generalizability and robustness of our approach. + +We further evaluate our method on Anomaly-ShapeNet in Table 2, which encompasses a broader array of categories and a greater diversity of defect types. Compared to Real3D-AD, Anomaly-ShapeNet significantly enhances the diversity of defects, wherein the increased variety of defect types further escalates the complexity of detection tasks. The results highlight the exceptional performance of our method across all evaluated categories, demonstrating an average improvement of $9\%$ on AUROC relative to the approaches previously utilized. + +
ModelDiffusionConditionRelativePatch-GenI-AUROCP-AUROC
A✓✗✗✗0.5860.524
B✓✓✗✗0.6670.513
C✓✓✓✗0.7120.573
D✓✓✓✓0.7340.592
+ +Table 3: Ablation studies for 3D adaptation components on Real3D-AD dataset. + +# 4.5 Ablation study + +To delve into the effect of individual components, we conduct ablation experiments on the Real3D-AD dataset. To fully demonstrate and compare the performance of the models, we report both image-level and pixel-level results with I-AUROC and P-AUROC, respectively. + +Main component Table 3 compares the performance of different variants from R3D-AD, which includes the influence of the denoising condition embedding, displacement-based reconstruction way, and the data augmentation strategy of Patch-Gen. Model A is denoted as our baseline, which is a vanilla DDPM model for point cloud reconstruction. Introducing a condition into the DDPM (Model B) significantly boosts performance, particularly in terms of I-AUROC, which sees a $13.8\%$ increase to 0.667. Model C, which predicts point displacements based on conditional DDPM, preserving detailed structural information while accommodating the relative displacement of points contributes to a notable $6.0\%$ gain in P-AUROC over Model B. Model D is trained under the conditions of shape embedding with the Patch-Gen strategy. Considering that the defective portion contains only a small portion of the original point cloud, we try to reconstruct the relative displacement in a way that preserves as much detail as possible, which is effective for both 3D anomaly detection and segmentation. + +Patch-Gen Table. 4 analyzes the influence of two key parameters in Patch-Gen: the selection points ratio and the scaling points factor. + +The selection points ratio from Table. 4a determines the proportion of points in the point cloud that are selected for transformation. Our findings suggest that a selection ratio of $1/32$ achieves the best performance. It appears that this ratio provides a balanced trade-off between maintaining sufficient structure for anomaly detection and introducing enough variation to simulate anomalies effectively. Notably, as the ratio increases beyond $1/16$ , both I-AUROC and P-AUROC scores decrease in severity, since real defects only account for a small portion of the overall point cloud, a wide selection of points not only destroys the structure of the original point cloud, but also makes the distribution of the training and test sets inconsistent. + +The scaling points factor is the intensity of the random transformation applied to the selected points, as detailed in Table 4b. The optimal performance is + +
ratioI-AUROCP-AUROC
1/640.7160.584
1/320.7340.592
1/160.7270.579
1/80.6830.528
+ +(a) Selection points ratio. + +
factorI-AUROCP-AUROC
0.10.7340.592
0.20.7270.572
0.40.7150.554
0.80.6610.517
+ +(b) Scaling points factor. + +Table 4: Ablation studies for Patch-Gen implementation on Real3D-AD dataset. Default settings are marked in gray. + +![](images/e6ff16748f29b8ed19dce6e60f2b302acbe66932d90b7223ee0b54508a5bb848.jpg) +(a) + +![](images/8e47002cd3a71ddd50d9a1c6e5d39e3739396e30da9e3e4b36a840d07aebe135.jpg) +(b) +Fig. 4: Memory and time cost during inference on Real3D-AD dataset. (a) Memory usage comparison between different models. (b) 3D anomaly detection performance vs. frames per second on an NVIDIA RTX 3090 GPU. Our R3D-AD outperforms all previous methods on both accuracy and efficiency by a significant margin. + +observed at a scaling factor of 0.1, which implies that minor transformations are more effective for simulating anomalies without significantly altering the original data distribution. Larger scaling factors lead to a consistent decline in performance, underscoring the importance of subtle transformations for preserving the utility of the simulated anomalies for detection tasks. + +Memory and time cost As depicted in Figure 4, we evaluate the disparity in both storage consumption and inference time of our model under identical experimental conditions, compared to existing methods. Regarding memory usage, our approach demonstrates a marked superiority by employing raw coordinate features instead of FPFH or PointMAE features, significantly reducing the memory footprint. Since no memory bank exists, our method is also more space-efficient compared to BTF which also uses raw features. Moreover, our method eliminates the necessity to compare all the features in memory, substantially increasing operational efficiency. The implementation of Patch-Gen inherently bestows our model with exceptional robustness, enabling precise reconstruction of point clouds from various angles without the need for the time-intensive RANSAC alignment process required by Reg3D-AD. + +![](images/4bfac1c3112a59a58b7e6a096b999e1658417378b8cbe516422c3f7190ea7e2e.jpg) +Fig. 5: Qualitative analysis on Real3D-AD dataset and Anomaly-ShapeNet dataset. The anomaly map is obtained directly by calculating the differences between the input and reconstructed point clouds, where deeper colors represent more confidence. + +# 4.6 Qualitative results + +Figure 5 presents some qualitative outcomes, with varying shades of color indicating different levels of anomaly scores. We select several representative defective samples to demonstrate the robustness of our algorithm. The left four columns display samples from Real3D-AD, while the right four columns samples are from Anomaly-ShapeNet. The illustration reveals that our R3D-AD algorithm has precisely reconstructed the defective portions of the point cloud across various samples: the deep sink in the Seahorse sample, the concavity in the Bag sample, and the bulge in the Jar sample. Leveraging the accurately reconstructed point clouds, final point cloud segmentation maps are also produced, further evidencing the efficacy of our approach. + +# 5 Conclusion + +In this work, we presented R3D-AD, a novel reconstructive 3D anomaly detection model based on conditional diffusion. Our goal is to overcome the limitations faced by current 3D anomaly detection methods, such as the inefficiencies due to the memory bank module and low performance caused by incorrect rebuilds with MAE. To address these challenges, we leverage the diffusion process for full reconstruction, followed by a direct comparison between the input and the reconstructed point cloud to obtain the final anomaly score. The embedded latent variable that spans the decoding process, step-wisely generating point-level displacements from the noise to the target anomaly-free sample. We also propose Patch-Gen, a data augmentation tailored for point cloud anomaly simulation. Extensive experiments conducted on 3D anomaly benchmarks validate the superiority of our R3D-AD in comparison to state-of-the-art alternatives in terms of both accuracy and versatility. + +# Acknowledgements + +This work was supported in part by the Pioneer and Leading Goose R&D Program of Zhejiang (Grant No. 2022C01051), in part by the National Natural Science Foundation of China (Grant No. 52375271, 52275274), and in part by the Natural Science Foundation of Zhejiang Province (Grant No. LY23E050011). + +# References + +1. Bae, J., Lee, J.H., Kim, S.: Pni: Industrial anomaly detection using position and neighborhood information. In: ICCV (2023) +2. Bergmann, P., Löwe, S., Fauser, M., Sattlegger, D., Steger, C.: Improving unsupervised defect segmentation by applying structural similarity to autoencoders. In: VISIGRAPP (2019) +3. Cao, Y., Xu, X., Shen, W.: Complementary pseudo multimodal feature for point cloud anomaly detection. arXiv preprint (2023) +4. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., Xiao, J., Yi, L., Yu, F.: ShapeNet: An Information-Rich 3D Model Repository. arXiv preprint (2015) +5. Charles, R.Q., Su, H., Kaichun, M., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: CVPR (2017) +6. Chu, R., Xie, E., Mo, S., Li, Z., Nießner, M., Fu, C.W., Jia, J.: Diffcomplete: Diffusion-based generative 3d shape completion. In: NeurIPS (2023) +7. Dinh, L., Sohl-Dickstein, J., Bengio, S.: Density estimation using real nvp. In: ICLR (2017) +8. Gudovskiy, D., Ishizaka, S., Kozuka, K.: Cflow-ad: Real-time unsupervised anomaly detection with localization via conditional normalizing flows. In: WACV (2022) +9. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022) +10. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., Salimans, T.: Imagen video: High definition video generation with diffusion models (2022) +11. Horwitz, E., Hoshen, Y.: Back to the feature: Classical 3d features are (almost) all you need for 3d anomaly detection. In: CVPRW (2023) +12. Hu, T., Zhang, J., Yi, R., Du, Y., Chen, X., Liu, L., Wang, Y., Wang, C.: Anomalydiffusion: Few-shot anomaly image generation with diffusion model. In: AAAI (2024) +13. Jonathan Ho, Ajay Jain, and Pieter Abbeel: Denoising diffusion probabilistic models. In: NeurIPS (2020) +14. Kim, D., Park, C., Cho, S., Lee, S.: Fapm: Fast adaptive patch memory for real-time industrial anomaly detection. In: ICASSP (2023) +15. Kong, Z., Ping, W., Huang, J., Zhao, K., Catanzaro, B.: Diffwave: A versatile diffusion model for audio synthesis. In: ICLR (2021) +16. Li, C.L., Sohn, K., Yoon, J., Pfister, T.: Cutpaste: Self-supervised learning for anomaly detection and localization. In: CVPR (2021) +17. Li, M., Duan, Y., Zhou, J., Lu, J.: Diffusion-sdf: Text-to-shape via voxelized diffusion. In: CVPR (2023) +18. Li, W., Xu, X., Gu, Y., Zheng, B., Gao, S., Wu, Y.: Towards scalable 3d anomaly detection and localization: A benchmark via 3d anomaly synthesis and a self-supervised learning network. arXiv preprint (2023) + +19. Liu, J., Xie, G., Li, X., Wang, J., Liu, Y., Wang, C., Zheng, F., et al.: Real3d-ad: A dataset of point cloud anomaly detection. In: NeurIPS (2023) +20. Liu, Z., Feng, Y., Black, M.J., Nowrouzezahrai, D., Paull, L., Liu, W.: Meshdiffusion: Score-based generative 3d mesh modeling. In: ICLR (2023) +21. Liu, Z., Tang, H., Lin, Y., Han, S.: Point-voxel cnn for efficient 3d deep learning. In: NeurIPS (2019) +22. Lu, F., Yao, X., Fu, C., Jia, J.: Removing anomalies as noises for industrial defect localization. In: ICCV (2023) +23. Luo, S., Hu, W.: Diffusion probabilistic models for 3d point cloud generation. In: CVPR (2021) +24. Mo, S., Xie, E., Chu, R., Hong, L., Nießner, M., Li, Z.: Dit-3d: Exploring plain diffusion transformers for 3d shape generation. In: NeurIPS (2023) +25. Pang, Y., Wang, W., Tay, F.E., Liu, W., Tian, Y., Yuan, L.: Masked autoencoders for point cloud self-supervised learning. In: ECCV (2022) +26. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. NeurIPS (2019) +27. Rezende, D.J., Mohamed, S.: Variational inference with normalizing flows. In: ICML (2015) +28. Roth, K., Pemula, L., Zepeda, J., Scholkopf, B., Brox, T., Gehler, P.: Towards total recall in industrial anomaly detection. In: CVPR (2022) +29. Rudolph, M., Wandt, B., Rosenhahn, B.: Same same but differnet: Semi-supervised defect detection with normalizing flows. In: WACV (2021) +30. Rusu, R.B., Blodow, N., Beetz, M.: Fast point feature histograms (fpfh) for 3d registration. In: ICRA (2009) +31. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Ghasemipour, S.K.S., Ayan, B.K., Mahdavi, S.S., Lopes, R.G., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint (2022) +32. Schlüter, H.M., Tan, J., Hou, B., Kainz, B.: Natural synthetic anomalies for self-supervised anomaly detection and localization. In: ECCV (2022) +33. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint (2020) +34. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: ICLR (2021) +35. Tailanian, M., Pardo, Å., Muse, P.: U-flow: A u-shaped normalizing flow for anomaly detection with unsupervised threshold. arXiv preprint (2022) +36. Wang, Y., Peng, J., Zhang, J., Yi, R., Wang, Y., Wang, C.: Multimodal industrial anomaly detection via hybrid fusion. In: CVPR (2023) +37. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph cnn for learning on point clouds. TOG (2019) +38. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: Simmim: A simple framework for masked image modeling. In: CVPR (2022) +39. Yu, J., Zheng, Y., Wang, X., Li, W., Wu, Y., Zhao, R., Wu, L.: Fastflow: Unsupervised anomaly detection and localization via 2d normalizing flows. arXiv preprint (2021) +40. Yu, X., Tang, L., Rao, Y., Huang, T., Zhou, J., Lu, J.: Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In: CVPR (2022) +41. Zavrtanik, V., Kristan, M., Skočaj, D.: Draem-a discriminatively trained reconstruction embedding for surface anomaly detection. In: ICCV (2021) + +42. Zavrtanik, V., Kristan, M., Skočaj, D.: Reconstruction by inpainting for visual anomaly detection. Pattern Recognition (2021) +43. Zhang, X., Xu, M., Zhou, X.: Realnet: A feature selection network with realistic synthetic anomaly for anomaly detection. In: CVPR (2024) +44. Zhou, L., Du, Y., Wu, J.: 3d shape generation and completion through point-voxel diffusion. In: ICCV (2021) \ No newline at end of file diff --git a/2024/R3D-AD_ Reconstruction via Diffusion for 3D Anomaly Detection/images.zip b/2024/R3D-AD_ Reconstruction via Diffusion for 3D Anomaly Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8dca263ee56c5b1d7482f64ef498a41813f76b5a --- /dev/null +++ b/2024/R3D-AD_ Reconstruction via Diffusion for 3D Anomaly Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0d1a4f167095f1aa655f2144ef201f074c4702f27f00d33c277c73ff319f449a +size 498331 diff --git a/2024/R3D-AD_ Reconstruction via Diffusion for 3D Anomaly Detection/layout.json b/2024/R3D-AD_ Reconstruction via Diffusion for 3D Anomaly Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b08ceed5fbf15398bf2c0c77ffd49e9de0c296f5 --- /dev/null +++ b/2024/R3D-AD_ Reconstruction via Diffusion for 3D Anomaly Detection/layout.json @@ -0,0 +1,8824 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 147, + 111, + 467, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 111, + 467, + 148 + ], + "spans": [ + { + "bbox": [ + 147, + 111, + 467, + 148 + ], + "type": "text", + "content": "R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 197, + 167, + 414, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 167, + 414, + 193 + ], + "spans": [ + { + "bbox": [ + 197, + 167, + 414, + 193 + ], + "type": "text", + "content": "Zheyuan Zhou" + }, + { + "bbox": [ + 197, + 167, + 414, + 193 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 197, + 167, + 414, + 193 + ], + "type": "text", + "content": ", Le Wang" + }, + { + "bbox": [ + 197, + 167, + 414, + 193 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 197, + 167, + 414, + 193 + ], + "type": "text", + "content": ", Naiyu Fang" + }, + { + "bbox": [ + 197, + 167, + 414, + 193 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 197, + 167, + 414, + 193 + ], + "type": "text", + "content": ", Zili Wang" + }, + { + "bbox": [ + 197, + 167, + 414, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 197, + 167, + 414, + 193 + ], + "type": "text", + "content": ", Lemiao Qiu" + }, + { + "bbox": [ + 197, + 167, + 414, + 193 + ], + "type": "inline_equation", + "content": "^{1\\boxtimes}" + }, + { + "bbox": [ + 197, + 167, + 414, + 193 + ], + "type": "text", + "content": ", and Shuyou Zhang" + }, + { + "bbox": [ + 197, + 167, + 414, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 175, + 201, + 436, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 201, + 436, + 223 + ], + "spans": [ + { + "bbox": [ + 175, + 201, + 436, + 223 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 175, + 201, + 436, + 223 + ], + "type": "text", + "content": " State Key Laboratory of Fluid Power & Mechatronic System, Zhejiang University, Hangzhou, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 172, + 224, + 439, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 224, + 439, + 235 + ], + "spans": [ + { + "bbox": [ + 172, + 224, + 439, + 235 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 172, + 224, + 439, + 235 + ], + "type": "text", + "content": " S-Lab, Nanyang Technological University, Singapore, Singapore" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 219, + 246, + 392, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 246, + 392, + 256 + ], + "spans": [ + { + "bbox": [ + 219, + 246, + 392, + 256 + ], + "type": "text", + "content": "https://zhouzheyuan.github.io/r3d-ad" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 283, + 452, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 283, + 452, + 524 + ], + "spans": [ + { + "bbox": [ + 160, + 283, + 452, + 524 + ], + "type": "text", + "content": "Abstract. 3D anomaly detection plays a crucial role in monitoring parts for localized inherent defects in precision manufacturing. Embedding-based and reconstruction-based approaches are among the most popular and successful methods. However, there are two major challenges to the practical application of the current approaches: 1) the embedded models suffer the prohibitive computational and storage due to the memory bank structure; 2) the reconstructive models based on the MAE mechanism fail to detect anomalies in the unmasked regions. In this paper, we propose R3D-AD, reconstructing anomalous point clouds by diffusion model for precise 3D anomaly detection. Our approach capitalizes on the data distribution conversion of the diffusion process to entirely obscure the input's anomalous geometry. It step-wisely learns a strict point-level displacement behavior, which methodically corrects the aberrant points. To increase the generalization of the model, we further present a novel 3D anomaly simulation strategy named Patch-Gen to generate realistic and diverse defect shapes, which narrows the domain gap between training and testing. Our R3D-AD ensures a uniform spatial transformation, which allows straightforwardly generating anomaly results by distance comparison. Extensive experiments show that our R3D-AD outperforms previous state-of-the-art methods, achieving " + }, + { + "bbox": [ + 160, + 283, + 452, + 524 + ], + "type": "inline_equation", + "content": "73.4\\%" + }, + { + "bbox": [ + 160, + 283, + 452, + 524 + ], + "type": "text", + "content": " Image-level AUROC on the Real3D-AD dataset and " + }, + { + "bbox": [ + 160, + 283, + 452, + 524 + ], + "type": "inline_equation", + "content": "74.9\\%" + }, + { + "bbox": [ + 160, + 283, + 452, + 524 + ], + "type": "text", + "content": " Image-level AUROC on the Anomaly-ShapeNet dataset with an exceptional efficiency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 534, + 453, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 534, + 453, + 556 + ], + "spans": [ + { + "bbox": [ + 159, + 534, + 453, + 556 + ], + "type": "text", + "content": "Keywords: 3D anomaly detection, industrial applications, 3D reconstruction, self-supervised learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 575, + 230, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 575, + 230, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 575, + 230, + 588 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 599, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 599, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 599, + 482, + 647 + ], + "type": "text", + "content": "Anomaly detection aims to identify instances containing anomalies and to precisely locate the specific positions of defects. This task is extensively applied across multiple fields and plays a crucial role in quality control within industrial production [28]. 3D anomaly detection [19] has emerged due to its intrinsic" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 654, + 223, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 654, + 223, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 654, + 223, + 665 + ], + "type": "text", + "content": "* Equal contribution." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 173, + 118, + 296, + 193 + ], + "blocks": [ + { + "bbox": [ + 173, + 118, + 296, + 193 + ], + "lines": [ + { + "bbox": [ + 173, + 118, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 173, + 118, + 296, + 193 + ], + "type": "image", + "image_path": "3f369856b866b0810a85c9855efae2baf59879e5aeab2e0a8a889bc771d06180.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 195, + 260, + 203 + ], + "lines": [ + { + "bbox": [ + 203, + 195, + 260, + 203 + ], + "spans": [ + { + "bbox": [ + 203, + 195, + 260, + 203 + ], + "type": "text", + "content": "(a) Embedded model" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 329, + 118, + 439, + 193 + ], + "blocks": [ + { + "bbox": [ + 329, + 118, + 439, + 193 + ], + "lines": [ + { + "bbox": [ + 329, + 118, + 439, + 193 + ], + "spans": [ + { + "bbox": [ + 329, + 118, + 439, + 193 + ], + "type": "image", + "image_path": "0fd877a4c795486bb0d559134332b65e7c05ea94fdd0ea8529ca7ec254083f45.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 194, + 417, + 202 + ], + "lines": [ + { + "bbox": [ + 349, + 194, + 417, + 202 + ], + "spans": [ + { + "bbox": [ + 349, + 194, + 417, + 202 + ], + "type": "text", + "content": "(b) Reconstructive model" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 217, + 482, + 282 + ], + "lines": [ + { + "bbox": [ + 130, + 217, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 482, + 282 + ], + "type": "text", + "content": "Fig. 1: Comparison of architectures. (a) Embedded model encodes the input " + }, + { + "bbox": [ + 130, + 217, + 482, + 282 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 130, + 217, + 482, + 282 + ], + "type": "text", + "content": " into features and stores them in the memory bank during training. The anomaly map " + }, + { + "bbox": [ + 130, + 217, + 482, + 282 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 217, + 482, + 282 + ], + "type": "text", + "content": " is obtained by comparing the test features with all the features in the memory bank. (b) Reconstructive model is trained by minimizing the loss between its input " + }, + { + "bbox": [ + 130, + 217, + 482, + 282 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 130, + 217, + 482, + 282 + ], + "type": "text", + "content": " and the output " + }, + { + "bbox": [ + 130, + 217, + 482, + 282 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{X}}" + }, + { + "bbox": [ + 130, + 217, + 482, + 282 + ], + "type": "text", + "content": ". The anomaly map " + }, + { + "bbox": [ + 130, + 217, + 482, + 282 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 217, + 482, + 282 + ], + "type": "text", + "content": " is obtained by comparing the test phase input with its corresponding reconstruction target." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 312, + 482, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 396 + ], + "type": "text", + "content": "modality superior for avoiding blind spots in advanced processing and precision manufacturing. However, the discrete and disordered data form of point clouds makes it more difficult to acquire features compared to images. With the scarcity of anomalies, 3D anomaly detection also faces the problem of domain shift while only normal data are presented during training. The presence of these issues underscores the necessity and urgency of devising an efficient framework for the 3D anomaly detection task." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 399, + 482, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 615 + ], + "type": "text", + "content": "Similar to traditional 2D anomaly detection [28,42], current 3D anomaly detection can be primarily categorized into embedding-based and reconstruction-based, as illustrated in Fig. 1. The embedding-based methods involve mapping features extracted with a pre-trained encoder onto a normal distribution for learning. Distributions that do not fall within the interval are classified as anomalies. Most existing 3D anomaly detection methods are based on a memory bank mechanism [3, 11, 19, 36], which stores some representative features during the training phase to implicitly construct a feature distribution. In the testing phase, the presence of anomalies is determined by calculating the Euclidean distance between the input test object and all template point clouds stored in memory. The reconstruction-based methods train a network capable of accurately reconstructing normal point clouds, under the presumption that anomalous point clouds will not be effectively reconstructed since they are not included during training. The anomaly map is produced through the comparison of discrepancies between the input point cloud and its reconstruction. IMRNet [18] employs PointMAE [25] to reconstruct the input in several iterations, getting the final anomaly map by calculating the explicit spatial coordinate differences and implicit deep feature differences of the point cloud, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "However, existing methods face two key issues, high resource cost and irreparable reconstruction. Firstly, methods based on the memory bank [3,11,19, 36] store all features from the training phase, each test point cloud needs to be compared with all samples in the memory bank, significantly increasing memory" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "Z. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "overhead and inference time costs. This makes such methods almost inapplicable in real industrial production lines due to their inefficiency. Secondly, masked autoencoder (MAE) mechanism [9,25,38] only reconstructs the masked portions of the input, defects within unmasked portions may be preserved. This contradicts the fundamental assumption of detecting anomalies by comparing the original defect-containing point cloud with a reconstructed anomaly-free version. These methods inevitably lead to incorrect reconstructions, undermining their effectiveness in accurately localizing defects." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 223, + 482, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 223, + 482, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 223, + 482, + 523 + ], + "type": "text", + "content": "We propose R3D-AD, a novel 3D anomaly detection method that does not suffer from the space burden and time endurance in memory-based embedded models nor the anomaly unmasking probability in the MAE-based reconstructive models. In contrast to PointMAE, one of our key insights is to perform undifferentiated masking for 3D objects via the noise diffusion mechanism, which maximizes the preservation of anomaly-free shapes and reconstructs abnormal regions. In the reparameterized diffusion process, one-step full mask and reconstruction are achieved by converting the point cloud distribution, instead of the multiple iterative method [18]. We hypothesize that anomaly detection verifies the gap between the reconstructed shapes and the positive samples by learning point movement. Specifically, for input models with arbitrary anomalies, we encode them as latent shape embeddings as decoding conditions and explicitly control the point cloud reconstruction process by step-wise displacements (SWD) decoding. The shape embedding harbors abundant global features and makes it easier to train the network without dwelling on the introduction of local anomaly details. Another key to our approach is to implement a controllable method of point-wise displacement during the diffusion process to refine the point cloud deformation iteratively. We propose to inject latent shape embedding into each step of the inverse denoising process, which drives the anomalous regions to converge to a smooth surface. We further adopted a 3D anomaly simulation strategy Patch-Gen to address the limitations of the dataset, which generates abundant defectives by producing spatial irregularity that is faithful to the real scene, including bulges, sinks, etc. This point cloud data augmentation encourages the self-supervised model to reconstruct more realistic anomaly-free shapes when facing the actual anomaly." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "content": "To the best of our knowledge, this is the very first attempt at exploring diffusion in reconstruction-based 3D anomaly detection. Our main contributions are summarized as follows: (i) We introduce a novel framework, termed R3D-AD, which performs a one-step full mask and anomaly-free reconstruction for fast and accurate 3D anomaly detection. (ii) We propose to learn the step-wise displacement in the reverse diffusion process to explicitly control the reconstruction of anomalous shapes. (iii) We introduce a 3D anomaly simulation strategy named Patch-Gen to address the limitation of the data anomaly patterns and improve the reconstruction performance in a supervised setting. (iv) Extensive experiments demonstrate that our R3D-AD has achieved state-of-the-art performance on both Real3D-AD and Anomaly-ShapeNet datasets." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "text", + "content": "R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 234, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 234, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 234, + 127 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 277, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 277, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 277, + 152 + ], + "type": "text", + "content": "2.1 2D Anomaly Detection" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 159, + 482, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 159, + 482, + 268 + ], + "spans": [ + { + "bbox": [ + 130, + 159, + 482, + 268 + ], + "type": "text", + "content": "Anomaly detection has received increasing attention from researchers in recent years, and many new methods have been proposed to address the problem. Flow-based methods [8, 29, 35, 39] use learned distributions and flow's bijective properties to spot defects, while Memory-based approaches [1, 14, 28] gauge anomaly scores by contrasting test sample features with memory bank-stored norms. Reconstruction-based models [2, 41, 42] flag anomalies by comparing inputs to their online reconstructions. Recent works [12, 16, 32, 43] augment the anomaly detection datasets with generated synthetic anomalies to compensate for the negative example scarcity problem." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 285, + 277, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 277, + 297 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 277, + 297 + ], + "type": "text", + "content": "2.2 3D Anomaly Detection" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 304, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 509 + ], + "type": "text", + "content": "This field lags behind the development of 2D anomaly detection since 3D data are harder to obtain, while point cloud data are sparser and contain more noise than image data. BTF [11] integration of handcrafted 3D descriptors with classic 2D method PatchCore [28], constructing a basic framework for 3D anomaly detection. M3DM [36] advances the field by separately analyzing features from point clouds and RGB images, then merging these for improved decision-making. CPMF [3] converts point clouds into two-dimensional images from multiple angles, extracting additional features from these images with a pre-trained network, and enhancing detection capabilities through information fusion. Reg3D-AD [19] develops a registration-based method, the RANSAC algorithm was used to align each sample before comparing it to the stored template during the test phase. IMRNet [18] trains a PointMAE [25] to reconstruct anomaly-free samples and identifies anomalies by juxtaposing the reconstructed point cloud against the initial input. Many of these use memory banks to store the features of the training samples or require multiple iterations to restore points. Unlike previous methods, our approach requires only one step of reconstruction and has significant advantages in both time and space efficiency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 525, + 246, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 246, + 537 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 246, + 537 + ], + "type": "text", + "content": "2.3 Diffusion Models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Diffusion models have proven their effectiveness in several generative tasks, such as image generation [31], speech generation [15], and video generation [10]. Denoising Diffusion Probabilistic Models (DDPMs) [13, 33, 34] employ a forward noisng mechanism, incrementally integrating Gaussian noise into images, alongside a reverse process meticulously trained to counteract the forward mechanism. Denoise AD [22] conducts DDPM for reconstructing within the features space, generating images that contain less noise. In recent years, many studies [6, 17, 20, 24] have attempted to use the diffusion model to explore the 3D reconstruction task. DPM [23] incorporates a shape latent variable to encapsulate the geometric intricacies of 3D shapes, it distinctively models this variable's" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "text", + "content": "Z. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": "distribution utilizing Normalizing Flows [7, 27]. PVD [44] utilizes PVCNNs [21] for the point-voxel representation of 3D shapes and integrates structured locality into point clouds. This innovative approach leverages the strengths of both point and voxel representations, optimizing the model's ability to capture the intricate spatial hierarchies and local geometries within 3D objects. Since diffusion-based reconstruction recovers the target shape from complete noise, the dilemma of reconstructing only the masked region in the MAE [9] mechanism does not exist." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 219, + 202, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 219, + 202, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 219, + 202, + 232 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 244, + 208, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 244, + 208, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 244, + 208, + 255 + ], + "type": "text", + "content": "3.1 Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 265, + 480, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 265, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 480, + 384 + ], + "type": "text", + "content": "We model the anomaly detection problem as mapping an anomalous point cloud " + }, + { + "bbox": [ + 130, + 265, + 480, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{a}} \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 130, + 265, + 480, + 384 + ], + "type": "text", + "content": " to a positive shape with which it is aligned. The framework of R3D-AD is shown in Fig. 2, where the simulated anomalous shapes are reconstructed in a self-supervised setting in the training phase and then compared with the original input to detect anomalies. The reconstructed anomaly-free model is aligned with the input, thus allowing direct computation of anomaly scores and segmentation of anomalous regions by conditioned distance functions. Simultaneously, the anomaly simulation strategy faithfully generates realistic defects and randomly synthesizes diverse anomaly shapes on normal samples, improving the generalization ability of the network in the case of limited anomaly samples." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 403, + 435, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 435, + 415 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 435, + 415 + ], + "type": "text", + "content": "3.2 Preliminary of denoising diffusion probabilistic models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 423, + 479, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 479, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 479, + 447 + ], + "type": "text", + "content": "A DDPM is inspired by the thermal diffusion process in an evolving thermodynamic system, which consists of a diffusion process and a reverse process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 448, + 480, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 480, + 483 + ], + "type": "text", + "content": "The forward Markovian process gradually adds Gaussian noise to a clean sample " + }, + { + "bbox": [ + 130, + 448, + 480, + 483 + ], + "type": "inline_equation", + "content": "\\pmb{x}^{(0)}" + }, + { + "bbox": [ + 130, + 448, + 480, + 483 + ], + "type": "text", + "content": " from a data distribution " + }, + { + "bbox": [ + 130, + 448, + 480, + 483 + ], + "type": "inline_equation", + "content": "q(\\pmb{x}^{(0)})" + }, + { + "bbox": [ + 130, + 448, + 480, + 483 + ], + "type": "text", + "content": " and turns it into a Gaussian noise " + }, + { + "bbox": [ + 130, + 448, + 480, + 483 + ], + "type": "inline_equation", + "content": "\\pmb{x}^{(T)}" + }, + { + "bbox": [ + 130, + 448, + 480, + 483 + ], + "type": "text", + "content": ", which is defined as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 230, + 494, + 480, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 494, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 230, + 494, + 480, + 525 + ], + "type": "interline_equation", + "content": "q \\left(\\boldsymbol {x} ^ {(0)}, \\dots , \\boldsymbol {x} ^ {(T)}\\right) = \\prod_ {t = 1} ^ {T} q \\left(\\boldsymbol {x} ^ {(t)} \\mid \\boldsymbol {x} ^ {(t - 1)}\\right), \\tag {1}", + "image_path": "a98a808840ce7db5040d7d115f946c51ef8b50fb90a5b02f769469adfad083ad.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 534, + 479, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 534, + 479, + 584 + ], + "spans": [ + { + "bbox": [ + 131, + 534, + 479, + 584 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 534, + 479, + 584 + ], + "type": "inline_equation", + "content": "q(\\pmb{x}^{(t)}|\\pmb{x}^{(t-1)}) = \\mathcal{N}(\\pmb{x}^{(t)};\\sqrt{1 - \\beta_t}\\pmb{x}^{(t-1)},\\beta_t\\pmb{I})" + }, + { + "bbox": [ + 131, + 534, + 479, + 584 + ], + "type": "text", + "content": " is the Markov diffusion kernel, " + }, + { + "bbox": [ + 131, + 534, + 479, + 584 + ], + "type": "inline_equation", + "content": "t = 1,\\dots,T" + }, + { + "bbox": [ + 131, + 534, + 479, + 584 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 534, + 479, + 584 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 534, + 479, + 584 + ], + "type": "text", + "content": " is the number of diffusion steps, and " + }, + { + "bbox": [ + 131, + 534, + 479, + 584 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 131, + 534, + 479, + 584 + ], + "type": "text", + "content": " is a variance schedule. We have " + }, + { + "bbox": [ + 131, + 534, + 479, + 584 + ], + "type": "inline_equation", + "content": "q(\\pmb{x}^{(t)}|\\pmb{x}^{(0)}) = \\mathcal{N}(\\pmb{x}^{(t)};\\sqrt{\\overline{\\alpha}_t}\\pmb{x}^{(0)},(1 - \\overline{\\alpha}_t)\\pmb{I})" + }, + { + "bbox": [ + 131, + 534, + 479, + 584 + ], + "type": "text", + "content": " by reparameterization with " + }, + { + "bbox": [ + 131, + 534, + 479, + 584 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = 1 - \\beta_{t}" + }, + { + "bbox": [ + 131, + 534, + 479, + 584 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 534, + 479, + 584 + ], + "type": "inline_equation", + "content": "\\overline{\\alpha}_{t} = \\prod_{s=0}^{t}\\alpha_{s}" + }, + { + "bbox": [ + 131, + 534, + 479, + 584 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 534, + 479, + 584 + ], + "type": "inline_equation", + "content": "\\pmb{x}^{(t)}" + }, + { + "bbox": [ + 131, + 534, + 479, + 584 + ], + "type": "text", + "content": " can be sampled by" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 242, + 594, + 480, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 594, + 480, + 608 + ], + "spans": [ + { + "bbox": [ + 242, + 594, + 480, + 608 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} ^ {(t)} = \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} ^ {(0)} + \\epsilon \\sqrt {\\left(1 - \\bar {\\alpha} _ {t}\\right)}, \\tag {2}", + "image_path": "a8d9b7e7334da0c667a253c199a3127632252609f83a3e9216b2a1d130ca4c89.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 617, + 479, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 479, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 479, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 617, + 479, + 640 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}" + }, + { + "bbox": [ + 130, + 617, + 479, + 640 + ], + "type": "text", + "content": " is a standard Gaussian noise and " + }, + { + "bbox": [ + 130, + 617, + 479, + 640 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 130, + 617, + 479, + 640 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 130, + 617, + 479, + 640 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 617, + 479, + 640 + ], + "type": "text", + "content": " is large enough, " + }, + { + "bbox": [ + 130, + 617, + 479, + 640 + ], + "type": "inline_equation", + "content": "\\pmb{x}^{(T)}" + }, + { + "bbox": [ + 130, + 617, + 479, + 640 + ], + "type": "text", + "content": " will eventually become a Gaussian noise." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 642, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 480, + 666 + ], + "type": "text", + "content": "The reverse process is also a Markovian process that denoises over a series of steps to generate meaningful data from the target distribution " + }, + { + "bbox": [ + 130, + 642, + 480, + 666 + ], + "type": "inline_equation", + "content": "q(\\pmb{x}^{(0)})" + }, + { + "bbox": [ + 130, + 642, + 480, + 666 + ], + "type": "text", + "content": ". The" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 179, + 91, + 448, + 102 + ], + "type": "text", + "content": "R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 116, + 395, + 276 + ], + "blocks": [ + { + "bbox": [ + 138, + 116, + 395, + 276 + ], + "lines": [ + { + "bbox": [ + 138, + 116, + 395, + 276 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 395, + 276 + ], + "type": "image", + "image_path": "e69d491b539e63f2664d9ac55a9a05d526ce672ebe2881b232452ca0d303e8ab.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 285, + 482, + 374 + ], + "lines": [ + { + "bbox": [ + 131, + 285, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 131, + 285, + 482, + 374 + ], + "type": "text", + "content": "Fig. 2: Overall architecture of R3D-AD for shape reconstruction and anomaly detection of point cloud objects. Reconstruction training phase: The simulated anomalous " + }, + { + "bbox": [ + 131, + 285, + 482, + 374 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{a}}^{(0)}" + }, + { + "bbox": [ + 131, + 285, + 482, + 374 + ], + "type": "text", + "content": " is generated resort to Patch-Gen from the input point cloud. It further fully masked as " + }, + { + "bbox": [ + 131, + 285, + 482, + 374 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{a}}^{(T)}" + }, + { + "bbox": [ + 131, + 285, + 482, + 374 + ], + "type": "text", + "content": " while also encoded to latent shape embedding. The SWD decoder then explicitly reconstructs the anomaly-free object " + }, + { + "bbox": [ + 131, + 285, + 482, + 374 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{r}}^{(0)}" + }, + { + "bbox": [ + 131, + 285, + 482, + 374 + ], + "type": "text", + "content": " with consistent spatial transform by conditionally generating point-level displacements " + }, + { + "bbox": [ + 131, + 285, + 482, + 374 + ], + "type": "inline_equation", + "content": "\\Delta^{(t)}" + }, + { + "bbox": [ + 131, + 285, + 482, + 374 + ], + "type": "text", + "content": " at each step of the inverse process. Detection testing phase: The test point cloud " + }, + { + "bbox": [ + 131, + 285, + 482, + 374 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{a}}^{(0)}" + }, + { + "bbox": [ + 131, + 285, + 482, + 374 + ], + "type": "text", + "content": " is reconstructed to " + }, + { + "bbox": [ + 131, + 285, + 482, + 374 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{r}}^{(0)}" + }, + { + "bbox": [ + 131, + 285, + 482, + 374 + ], + "type": "text", + "content": " with normal shape, and compared at a distance level to detect the anomalous region." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 399, + 116, + 478, + 276 + ], + "blocks": [ + { + "bbox": [ + 399, + 116, + 478, + 276 + ], + "lines": [ + { + "bbox": [ + 399, + 116, + 478, + 276 + ], + "spans": [ + { + "bbox": [ + 399, + 116, + 478, + 276 + ], + "type": "image", + "image_path": "c912d14ec7b9122444326834473d856d05cf95cf04711dca06a9231d400e89b1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 397, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 482, + 422 + ], + "type": "text", + "content": "inverse process denoises the noise " + }, + { + "bbox": [ + 130, + 397, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\pmb{x}^{(T)}" + }, + { + "bbox": [ + 130, + 397, + 482, + 422 + ], + "type": "text", + "content": " from a distribution " + }, + { + "bbox": [ + 130, + 397, + 482, + 422 + ], + "type": "inline_equation", + "content": "p(\\pmb{x}^{(T)})" + }, + { + "bbox": [ + 130, + 397, + 482, + 422 + ], + "type": "text", + "content": ", which is defined as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 195, + 430, + 481, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 430, + 481, + 463 + ], + "spans": [ + { + "bbox": [ + 195, + 430, + 481, + 463 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\boldsymbol {x} ^ {(0)}, \\dots , \\boldsymbol {x} ^ {(T - 1)} \\mid \\boldsymbol {x} ^ {(T)}, \\boldsymbol {c}\\right) = \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(\\boldsymbol {x} ^ {(t - 1)} \\mid \\boldsymbol {x} ^ {(t)}, \\boldsymbol {c}\\right), \\tag {3}", + "image_path": "28dd235314cc5aaae2f0290e3d50f54e2c733ec9d340925a6e37f328263d67cf.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 472, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 472, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 131, + 472, + 482, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 472, + 482, + 510 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\boldsymbol{x}^{(t-1)}|\\boldsymbol{x}^{(t)})" + }, + { + "bbox": [ + 131, + 472, + 482, + 510 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 472, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\boldsymbol{c} = \\mathcal{N}(\\boldsymbol{x}^{(t-1)};\\mu_{\\theta}(\\boldsymbol{x}^{(t)},t,\\boldsymbol{c}),\\sigma_t^2\\boldsymbol{I})" + }, + { + "bbox": [ + 131, + 472, + 482, + 510 + ], + "type": "text", + "content": ", the mean " + }, + { + "bbox": [ + 131, + 472, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}(\\boldsymbol{x}^{(t)},t,\\boldsymbol{c})" + }, + { + "bbox": [ + 131, + 472, + 482, + 510 + ], + "type": "text", + "content": " is estimated by a neural network parameterized by " + }, + { + "bbox": [ + 131, + 472, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\theta}" + }, + { + "bbox": [ + 131, + 472, + 482, + 510 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 472, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\boldsymbol{c}" + }, + { + "bbox": [ + 131, + 472, + 482, + 510 + ], + "type": "text", + "content": " is the latent condition encoding, and " + }, + { + "bbox": [ + 131, + 472, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\sigma_t^2" + }, + { + "bbox": [ + 131, + 472, + 482, + 510 + ], + "type": "text", + "content": " is a step-dependent variance. " + }, + { + "bbox": [ + 131, + 472, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}" + }, + { + "bbox": [ + 131, + 472, + 482, + 510 + ], + "type": "text", + "content": " can be reparameterized as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 197, + 520, + 481, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 520, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 197, + 520, + 481, + 546 + ], + "type": "interline_equation", + "content": "\\mu_ {\\theta} \\left(\\boldsymbol {x} ^ {(t)}, t, \\boldsymbol {c}\\right) = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\boldsymbol {x} ^ {(t)} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} ^ {(t)}, t, \\boldsymbol {c}\\right)\\right), \\tag {4}", + "image_path": "50c2b5a94cdf170d630e596b1a5b5240eea185253551f26d828e7e901313cbc6.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 556, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 556, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 556, + 482, + 579 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 556, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\pmb{x}^{(t)}, t, \\pmb{c})" + }, + { + "bbox": [ + 130, + 556, + 482, + 579 + ], + "type": "text", + "content": " is a neural network utilized to denoise the Gaussian noise from " + }, + { + "bbox": [ + 130, + 556, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\pmb{x}^{(T)}" + }, + { + "bbox": [ + 130, + 556, + 482, + 579 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 581, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 605 + ], + "type": "text", + "content": "The training objective is minimized by training " + }, + { + "bbox": [ + 130, + 581, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}_{\\pmb{\\theta}}(\\pmb{x}^{(t)}, t, \\pmb{c})" + }, + { + "bbox": [ + 130, + 581, + 482, + 605 + ], + "type": "text", + "content": " to approximate " + }, + { + "bbox": [ + 130, + 581, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}" + }, + { + "bbox": [ + 130, + 581, + 482, + 605 + ], + "type": "text", + "content": ". The training objective is defined as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 148, + 615, + 481, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 615, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 148, + 615, + 481, + 632 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} _ {t \\sim [ 1: T ], \\boldsymbol {x} ^ {(0)} \\sim q (\\boldsymbol {x} ^ {(0)})}, \\epsilon \\sim \\mathcal {N} (0, I) \\| \\epsilon - \\epsilon_ {\\boldsymbol {\\theta}} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} ^ {(0)} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , t, \\boldsymbol {c}\\right) \\|, \\tag {5}", + "image_path": "b146970882c8b35e871012256abdd8f30204f3676824f2c73bf6a0405c8a3544.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 640, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 640, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 640, + 482, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 640, + 482, + 665 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 640, + 482, + 665 + ], + "type": "text", + "content": " is sampled from the uniform distribution over " + }, + { + "bbox": [ + 130, + 640, + 482, + 665 + ], + "type": "inline_equation", + "content": "1,2,\\dots,T" + }, + { + "bbox": [ + 130, + 640, + 482, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 640, + 482, + 665 + ], + "type": "inline_equation", + "content": "q(\\pmb{x}^{(0)})" + }, + { + "bbox": [ + 130, + 640, + 482, + 665 + ], + "type": "text", + "content": " is the distribution of " + }, + { + "bbox": [ + 130, + 640, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\pmb{x}^{(0)}" + }, + { + "bbox": [ + 130, + 640, + 482, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 640, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 640, + 482, + 665 + ], + "type": "text", + "content": " is the Gaussian noise." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "Z. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 378, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 378, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 378, + 129 + ], + "type": "text", + "content": "3.3 Diffusion-based 3D anomaly reconstruction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "text", + "content": "We formulate the point cloud reconstruction task of the anomaly-free model as the conditional generation, which decodes the explicit displacement with the target distribution " + }, + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "inline_equation", + "content": "q(\\mathcal{P}_{\\mathrm{r}}|\\mathcal{P}_{\\mathrm{a}}^{(T)},c)" + }, + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "text", + "content": " is the decoding condition. The essential question of anomaly detection in this paper is how to conditional reconstruct anomaly-free shapes on the reference of input point clouds with different spatial transformations. Since there is a high similarity of global features between abnormal and normal samples during self-supervised reconstruction, the most immediate approach is to extract an efficient global feature from input to serve as an auxiliary conditional embedding for the denoising function " + }, + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "text", + "content": ". We implement the encoding of latent shape embedding " + }, + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 140, + 482, + 274 + ], + "type": "text", + "content": " as a conditional input to guide reconstruction in the reverse diffusion process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 295, + 482, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 482, + 394 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 482, + 394 + ], + "type": "text", + "content": "Latent shape embedding The feature encoder aims to encode the point cloud to the latent shape embedding " + }, + { + "bbox": [ + 130, + 295, + 482, + 394 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 295, + 482, + 394 + ], + "type": "text", + "content": " with high-level features for the conditional generation process. Different from other global-local extracting methods [37,40], we focus more on extracting global features, which characterize the semantic information of shape and pose of most anomaly-free regions in the point cloud. The feature encoder mainly consists of cascaded multi-layer perceptions (MLP) based on PointNet [5]. It implements max-pooling after mapping " + }, + { + "bbox": [ + 130, + 295, + 482, + 394 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{a}}^{(0)}" + }, + { + "bbox": [ + 130, + 295, + 482, + 394 + ], + "type": "text", + "content": " to different dimensions and then compresses them to extract the global shape embedding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 415, + 482, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 634 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 634 + ], + "type": "text", + "content": "Step-wise displacement decoding To achieve point cloud reconstruction with transformation consistency while preserving the structure of non-anomalous regions, our method injects latent shape embedding " + }, + { + "bbox": [ + 130, + 415, + 482, + 634 + ], + "type": "inline_equation", + "content": "\\pmb{c}" + }, + { + "bbox": [ + 130, + 415, + 482, + 634 + ], + "type": "text", + "content": " to the decoder at each step of the reverse diffusion process, as shown in Fig. 2. In principle, in the training phase, " + }, + { + "bbox": [ + 130, + 415, + 482, + 634 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 415, + 482, + 634 + ], + "type": "text", + "content": " learns the added Gaussian noise in the forward diffusion process by the decoder to model the conditional probability distribution. Conditionally generating target shapes from " + }, + { + "bbox": [ + 130, + 415, + 482, + 634 + ], + "type": "inline_equation", + "content": "N \\times 3" + }, + { + "bbox": [ + 130, + 415, + 482, + 634 + ], + "type": "text", + "content": " Gaussian noise is a straightforward approach, but it is afflicted by the issues of reconstructing the point cloud details and transform consistency. Learning the relative deformation of points for anomalous objects is more efficient. Considering the mapping degradation of the vanilla autoencoder in the reconstruction training phase [22], we utilize the Gaussian noise of the forward process Eq. 2 to fully mask the point cloud object directly without blind spots, preventing the decoding process from receiving negative state shapes. The masked points " + }, + { + "bbox": [ + 130, + 415, + 482, + 634 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{a}}^{(T)}" + }, + { + "bbox": [ + 130, + 415, + 482, + 634 + ], + "type": "text", + "content": " and latent shape embedding " + }, + { + "bbox": [ + 130, + 415, + 482, + 634 + ], + "type": "inline_equation", + "content": "\\pmb{c} \\in \\mathbb{R}^{256}" + }, + { + "bbox": [ + 130, + 415, + 482, + 634 + ], + "type": "text", + "content": " are as the inputs of the SWD decoder. The point-wise displacement vector " + }, + { + "bbox": [ + 130, + 415, + 482, + 634 + ], + "type": "inline_equation", + "content": "\\Delta^{(t)}" + }, + { + "bbox": [ + 130, + 415, + 482, + 634 + ], + "type": "text", + "content": " is generated at each step of the iterative process thus disentangling the prediction noise and the desired anomaly-free shape. The reverse process can be defined according to Eq. 3 and the displacement vector " + }, + { + "bbox": [ + 130, + 415, + 482, + 634 + ], + "type": "inline_equation", + "content": "\\Delta^{(t)}" + }, + { + "bbox": [ + 130, + 415, + 482, + 634 + ], + "type": "text", + "content": " can be represented by" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 195, + 643, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 643, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 195, + 643, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\Delta^ {(t - 1)} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\Delta^ {(t)} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\theta}} \\left(\\Delta^ {(t)}, \\beta_ {t}, \\boldsymbol {c}\\right)\\right) + \\sigma \\boldsymbol {\\epsilon}, \\tag {6}", + "image_path": "7ccd1047208c977d3f4106e00a0f06205424cbe7f81a230846a2df9e4339efaa.jpg" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 179, + 91, + 448, + 102 + ], + "type": "text", + "content": "R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 170, + 114, + 445, + 235 + ], + "blocks": [ + { + "bbox": [ + 170, + 114, + 445, + 235 + ], + "lines": [ + { + "bbox": [ + 170, + 114, + 445, + 235 + ], + "spans": [ + { + "bbox": [ + 170, + 114, + 445, + 235 + ], + "type": "image", + "image_path": "e024759dbdd519ebdbac73a91ea4561af0c2d82ba80c259a9a900488aea45250.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 243, + 482, + 288 + ], + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 288 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 288 + ], + "type": "text", + "content": "Fig. 3: Illustration of Patch-Gen, the 3D anomaly simulation strategy. The input normal point cloud is first randomly rotated. On the surface of the normalized cube, we randomly select viewpoints to find the nearest patch of points. The selected points are then transformed into irregular defects according to the specific deformation solution." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 310, + 482, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 482, + 397 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 310, + 482, + 397 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 310, + 482, + 397 + ], + "type": "text", + "content": " is the variance. A PointwiseNet is adopted for " + }, + { + "bbox": [ + 130, + 310, + 482, + 397 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 310, + 482, + 397 + ], + "type": "text", + "content": " to decode the " + }, + { + "bbox": [ + 130, + 310, + 482, + 397 + ], + "type": "inline_equation", + "content": "\\Delta^{(t-1)}" + }, + { + "bbox": [ + 130, + 310, + 482, + 397 + ], + "type": "text", + "content": " from the previous step and " + }, + { + "bbox": [ + 130, + 310, + 482, + 397 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 310, + 482, + 397 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 310, + 482, + 397 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 130, + 310, + 482, + 397 + ], + "type": "text", + "content": " is used to generate trigonometric position embedding " + }, + { + "bbox": [ + 130, + 310, + 482, + 397 + ], + "type": "inline_equation", + "content": "\\pmb{e}_{\\mathrm{p}} = (\\beta_{t}, \\sin(\\beta_{t}), \\cos(\\beta_{t}))" + }, + { + "bbox": [ + 130, + 310, + 482, + 397 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 310, + 482, + 397 + ], + "type": "inline_equation", + "content": "\\pmb{e}_{\\mathrm{p}}" + }, + { + "bbox": [ + 130, + 310, + 482, + 397 + ], + "type": "text", + "content": " is concatenated with " + }, + { + "bbox": [ + 130, + 310, + 482, + 397 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 310, + 482, + 397 + ], + "type": "text", + "content": " and then fed into the concatenate-squash linear module of PointwiseNet with a residual function. The output reconstructed point cloud at the " + }, + { + "bbox": [ + 130, + 310, + 482, + 397 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 310, + 482, + 397 + ], + "type": "text", + "content": " step is " + }, + { + "bbox": [ + 130, + 310, + 482, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{r}}^{(t)} = \\mathcal{P}_{\\mathrm{r}}^{(t+1)} + \\Delta^{(t)}" + }, + { + "bbox": [ + 130, + 310, + 482, + 397 + ], + "type": "text", + "content": ". The registered original and reconstructed objects are distinguished from the anomalous shape by the anomaly scores based on the conditioned distance function." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 413, + 322, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 413, + 322, + 425 + ], + "spans": [ + { + "bbox": [ + 131, + 413, + 322, + 425 + ], + "type": "text", + "content": "3.4 3D anomaly simulation strategy" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 432, + 482, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 432, + 482, + 539 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 482, + 539 + ], + "type": "text", + "content": "Given that a small number of normal samples is not conducive for the model to learn diverse and essential features, we propose the Patch-Gen strategy to simulate the defects from anomaly-free shapes for training data augmentation. Patch-Gen encourages the reconstruction model to learn to detect irregularity, where the anomaly-free point clouds and their diverse anomaly patterns are integrated into training pairs and are utilized to learn the discrimination feature between normal and anomalous surfaces. The intuition is that the diversity of simulated negative samples forces our network to learn how to reconstruct anomaly-free shapes instead of memorizing their complete outfits." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 540, + 482, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 540, + 482, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 540, + 482, + 575 + ], + "type": "text", + "content": "As shown in Fig. 3, the input normal sample is first randomly rotated. The random spatial rotation is designed to improve the generalization capability for test samples with very different spatial transformations, as defined by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 279, + 584, + 481, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 584, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 279, + 584, + 481, + 597 + ], + "type": "interline_equation", + "content": "\\mathcal {P} _ {\\mathrm {a}} = \\mathcal {P} \\cdot \\mathcal {R}, \\tag {7}", + "image_path": "8853db2c3243b3ead913889af950bf9b1ab11272e2edac4f0cf749fb95fef294.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " is input normal sample point cloud and " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{R} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " is obtained by randomly selecting rotation angles for all three axes. In addition to global shape awareness of the model by the random rotation, we further perform a fine granularity of the anomaly simulation. We randomly take a viewpoint " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_v" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " from the surface of the cube. Therefore, the patch of nearest " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " points " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_n" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_a" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " can be" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "Z. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 481, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 481, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 481, + 139 + ], + "type": "text", + "content": "determined according to the " + }, + { + "bbox": [ + 130, + 116, + 481, + 139 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_v" + }, + { + "bbox": [ + 130, + 116, + 481, + 139 + ], + "type": "text", + "content": ". The shape augmentation scheme Patch-Gen is defined as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 219, + 150, + 481, + 163 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 150, + 481, + 163 + ], + "spans": [ + { + "bbox": [ + 219, + 150, + 481, + 163 + ], + "type": "interline_equation", + "content": "\\mathcal {P} _ {n} = \\mathcal {P} _ {n} + S \\cdot \\text {n o r m a l i z e} \\left(\\mathcal {P} _ {n} - \\mathcal {P} _ {v}\\right) \\odot \\mathcal {T}, \\tag {8}", + "image_path": "4b5b84056fbe98ed49c6261e44efc608709d5912f5e4d0d7ba433312fee55374.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 171, + 479, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 171, + 479, + 219 + ], + "spans": [ + { + "bbox": [ + 130, + 171, + 479, + 219 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 171, + 479, + 219 + ], + "type": "inline_equation", + "content": "\\text{nomalize}" + }, + { + "bbox": [ + 130, + 171, + 479, + 219 + ], + "type": "text", + "content": " represents a normalization operation on a vector, " + }, + { + "bbox": [ + 130, + 171, + 479, + 219 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 171, + 479, + 219 + ], + "type": "text", + "content": " is a predefined hyper-parameter that controls the scaling of the patch points, and " + }, + { + "bbox": [ + 130, + 171, + 479, + 219 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 171, + 479, + 219 + ], + "type": "text", + "content": " is the translation matrix originating from a Gaussian distribution. The " + }, + { + "bbox": [ + 130, + 171, + 479, + 219 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{a}}" + }, + { + "bbox": [ + 130, + 171, + 479, + 219 + ], + "type": "text", + "content": " is finally obtained by only updating the patch region " + }, + { + "bbox": [ + 130, + 171, + 479, + 219 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_n" + }, + { + "bbox": [ + 130, + 171, + 479, + 219 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 220, + 480, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 220, + 480, + 268 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 480, + 268 + ], + "type": "text", + "content": "With the proposed Patch-Gen, we can simulate the generation of multiple anomalies, which is mainly done by controlling " + }, + { + "bbox": [ + 130, + 220, + 480, + 268 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 220, + 480, + 268 + ], + "type": "text", + "content": ". Bulge or sink can be generated by sorting " + }, + { + "bbox": [ + 130, + 220, + 480, + 268 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 220, + 480, + 268 + ], + "type": "text", + "content": " after sampling from the distribution, while damage can be generated by direct overlaying without manipulation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 285, + 253, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 285, + 253, + 298 + ], + "spans": [ + { + "bbox": [ + 131, + 285, + 253, + 298 + ], + "type": "text", + "content": "3.5 Training objective" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 304, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 415 + ], + "type": "text", + "content": "In the reconstruction task of the object with " + }, + { + "bbox": [ + 130, + 304, + 482, + 415 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 304, + 482, + 415 + ], + "type": "text", + "content": " points, the network learns a diffusion model with an " + }, + { + "bbox": [ + 130, + 304, + 482, + 415 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{N\\times 3}\\to \\mathbb{R}^{N\\times 3}" + }, + { + "bbox": [ + 130, + 304, + 482, + 415 + ], + "type": "text", + "content": " mapping relation. Iterative denoising under the semantic condition of point embedding realizes the prediction of point offsets. Concretely, the network is trained to learn the noise that needs to be eliminated to recover the anomaly-free shape with the " + }, + { + "bbox": [ + 130, + 304, + 482, + 415 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 130, + 304, + 482, + 415 + ], + "type": "text", + "content": " distance between the ground truth and the denoised reconstructed points. We make use of the mean squared error (MSE) loss as the primary reconstruction loss which evaluates the mean squared error of the element-wise distances between " + }, + { + "bbox": [ + 130, + 304, + 482, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{a}}^{(0)}" + }, + { + "bbox": [ + 130, + 304, + 482, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 304, + 482, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{r}}^{(0)}" + }, + { + "bbox": [ + 130, + 304, + 482, + 415 + ], + "type": "text", + "content": ". The MSE training loss is formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 222, + 423, + 481, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 423, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 222, + 423, + 481, + 456 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {P} _ {\\mathrm {a}}, \\mathcal {P} _ {\\mathrm {r}}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} p _ {\\mathrm {a}} \\in \\mathcal {P} _ {\\mathrm {a}}, p _ {\\mathrm {r}} \\in \\mathcal {P} _ {\\mathrm {r}} \\| p _ {\\mathrm {a}} - p _ {\\mathrm {r}} \\| ^ {2}. \\tag {9}", + "image_path": "efae48d8b7a19cf5536b160110869e590ec23e439a23e30e897f576a0ad9ee76.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 472, + 230, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 472, + 230, + 485 + ], + "spans": [ + { + "bbox": [ + 131, + 472, + 230, + 485 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 496, + 205, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 496, + 205, + 507 + ], + "spans": [ + { + "bbox": [ + 131, + 496, + 205, + 507 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 516, + 482, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 516, + 482, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 516, + 482, + 613 + ], + "type": "text", + "content": "Real3D-AD [19] is a 3D anomaly detection dataset based on real samples, exhibiting a higher point precision and spatial distance per point cloud. Each category contains 4 training samples and 100 test samples. The training set contains " + }, + { + "bbox": [ + 130, + 516, + 482, + 613 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 516, + 482, + 613 + ], + "type": "text", + "content": " complete surface point clouds of the objects, which are obtained by manually calibrating and stitching the scans of multiple sides of the objects. The test samples are scans only one side with a huge difference from the training set. The distribution of the point clouds also varies among the total 12 categories, further deepening the detection difficulty compared to 2D scenes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "type": "text", + "content": "Anomaly-ShapeNet [18] is a 3D anomaly detection, crafted through modifications to the synthetic samples found in ShapeNetCorev2 [4]. It contains 40 diverse categories, featuring over 1600 samples of its complete surface point" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 179, + 91, + 448, + 102 + ], + "type": "text", + "content": "R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 114, + 479, + 266 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 479, + 266 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 479, + 266 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 479, + 266 + ], + "type": "table", + "html": "
MethodBTF [11]M3DM [36]PatchCore [28]CPMF [3]Reg3D-AD [19]IMRNet [18]Ours
Feat.RawFPFHPointMAEFPFHPointMAEResNetPointMAEPointMAERaw
Airplane0.7300.5200.4340.8820.7260.7010.7160.7620.772
Candybar0.5390.6300.5520.5410.6630.5520.6850.7550.696
Car0.6470.5600.5410.5900.4980.5510.6970.7110.713
Chicken0.7890.4320.6830.8370.8270.5040.8520.7800.714
Diamond0.7070.5450.6020.5740.7830.5230.9000.9050.685
Duck0.6910.7840.4330.5460.4890.5820.5840.5170.909
Fish0.6020.5490.5400.6750.6300.5580.9150.8800.692
Gemstone0.6860.6480.6440.3700.3740.5890.4170.6740.665
Seahorse0.5960.7790.4950.5050.5390.7290.7620.6040.720
Shell0.3960.7540.6940.5890.5010.6530.5830.6650.840
Starfish0.5300.5750.5510.4410.5190.7000.5060.6740.701
Toffees0.7030.4620.4500.5650.5850.3900.8270.7740.703
Average0.6350.6030.5520.5930.5950.5860.7040.7250.734
", + "image_path": "5f8b31c05850a13cbb655442ed458fa6320a530222d6eb2c77c905c4e226f90f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "lines": [ + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "type": "text", + "content": "Table 1: Image-level anomaly detection AUROC on Real3D-AD dataset. We highlight the best result in bold and the second best result with an underline." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 326, + 480, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 480, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 480, + 387 + ], + "type": "text", + "content": "clouds. Each category's training set contains merely 4 samples, while the test sets are designed to assess the model's performance across both normal and a spectrum of abnormal samples. It widely increases the anomaly types while keeping the number of points the same as the previous studies, which places higher demands on the robustness and generality of the proposed algorithms." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 410, + 255, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 410, + 255, + 421 + ], + "spans": [ + { + "bbox": [ + 132, + 410, + 255, + 421 + ], + "type": "text", + "content": "4.2 Evaluation metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 436, + 481, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 481, + 509 + ], + "type": "text", + "content": "For image-level anomaly detection, the Area Under the Receiver Operating Curve (AUROC) is utilized in line with established practices. For the evaluation of pixel-level anomalies, the AUROC metric is similarly applied in the context of point segmentation accuracy. A value of 0.5 of the AUROC score denotes no discriminative capability (equivalent to random guessing), whereas a score of 1.0 indicates perfect discrimination between positive and negative classes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 532, + 278, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 278, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 278, + 544 + ], + "type": "text", + "content": "4.3 Implementation details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 666 + ], + "type": "text", + "content": "Our methodology is implemented using PyTorch [26] with end-to-end training across the network. The optimization is performed using the Adam optimizer, starting at an initial learning of 0.001. The training process involves a total batch size of 128 across 40,000 iterations for comprehensive learning. All input point clouds undergo a preprocessing step where they are randomly downsampled to a fixed size of 4096 and 2048 points on Real3D-AD and Anomaly-ShapeNet, respectively. Additionally, we normalized these point clouds by setting their center of gravity as the origin of coordinates and scaling their dimensions to fall within the range of -1 to 1, optimizing for the diffusion process." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "text", + "content": "Z. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 114, + 479, + 292 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 479, + 292 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 479, + 292 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 479, + 292 + ], + "type": "table", + "html": "
MethodBTF [11]M3DM [36]PatchCore [28]CPMF [3]Reg3D-AD [19]IMRNet [18]Ours
Feat.RawFPFHPointMAEFPFHPointMAEResNetPointMAEPointMAERaw
Ashtray0.5780.4200.5770.5870.5910.3530.5970.6710.833
Bag0.4100.5460.5370.5710.6010.6430.7060.6600.719
Bottle0.5580.4040.5840.6140.5880.4690.5690.6310.750
Bowl0.4700.5810.5790.5580.5470.6790.5480.6760.751
Bucket0.4690.5170.4050.5100.5770.5420.6810.6760.719
Cap0.5090.5620.5990.6450.5830.6010.6320.7040.726
Cup0.4620.5980.5480.5930.5830.4980.5240.7000.767
Eraser0.5250.7190.6270.6570.6770.6890.3430.5480.890
Headset0.4470.5050.5970.6100.6090.5510.5740.6980.767
Helmet0.5080.5690.4880.4650.4950.5320.4910.6030.704
Jar0.4200.4240.4410.4720.4830.6100.5920.7800.838
Microphone0.5630.6710.3570.3880.4880.5090.4140.7550.762
Shelf0.1640.6090.5640.4940.5230.6850.6880.6030.696
Tap0.5490.5530.7470.7600.4980.5280.6590.6860.818
Vase0.5170.4640.5340.5540.5820.5140.5760.6290.734
Average0.4930.5280.5520.5680.5620.5590.5720.6590.749
", + "image_path": "5dec720dc96e48350668acfbd4ee1313c7b5defae52b6f60ee1c219fd44b3093.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 293, + 480, + 316 + ], + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 316 + ], + "type": "text", + "content": "Table 2: Image-level anomaly detection AUROC on Anomaly-ShapeNet dataset. We highlight the best result in bold and the second best result with an underline." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 357, + 224, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 357, + 224, + 368 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 224, + 368 + ], + "type": "text", + "content": "4.4 Main results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 394, + 484, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 484, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 484, + 418 + ], + "type": "text", + "content": "We conduct experiments on Real3D-AD [19] based on real sampling and AnomalyShapeNet [18] based on simulation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 422, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 579 + ], + "type": "text", + "content": "As shown in Table 1, we first compare the image-level AUROC metric with current cutting-edge 3D anomaly detection models on Real3D-AD. It shows that our method achieves the best performance using only raw point cloud data, while most of the existing methods use Fast Point Feature Histograms (FPFH) operator [30] or ShapeNet [4] pre-trained PointMAE [25] as feature extractor. Due to significant disparities in quantity, size, and distribution among different categories of point clouds in Real3D-AD, scoring variations across categories are more pronounced with other methods. For instance, numerous methods perform under 0.5 in certain categories, indicating their inadequacy in extracting meaningful features while facing challenging samples. In contrast, our method not only exhibits superior performance in 3D anomaly detection across the majority of categories but also achieves the best overall average across all categories. This demonstrates the strong generalizability and robustness of our approach." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "We further evaluate our method on Anomaly-ShapeNet in Table 2, which encompasses a broader array of categories and a greater diversity of defect types. Compared to Real3D-AD, Anomaly-ShapeNet significantly enhances the diversity of defects, wherein the increased variety of defect types further escalates the complexity of detection tasks. The results highlight the exceptional performance of our method across all evaluated categories, demonstrating an average improvement of " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " on AUROC relative to the approaches previously utilized." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "text", + "content": "R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 170, + 114, + 443, + 183 + ], + "blocks": [ + { + "bbox": [ + 170, + 114, + 443, + 183 + ], + "lines": [ + { + "bbox": [ + 170, + 114, + 443, + 183 + ], + "spans": [ + { + "bbox": [ + 170, + 114, + 443, + 183 + ], + "type": "table", + "html": "
ModelDiffusionConditionRelativePatch-GenI-AUROCP-AUROC
A✓✗✗✗0.5860.524
B✓✓✗✗0.6670.513
C✓✓✓✗0.7120.573
D✓✓✓✓0.7340.592
", + "image_path": "c596b9bfa2fcc9ce7a2f3f5ccb0b98ad697bcb151c52feaca07b95578da4735b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 184, + 472, + 194 + ], + "lines": [ + { + "bbox": [ + 140, + 184, + 472, + 194 + ], + "spans": [ + { + "bbox": [ + 140, + 184, + 472, + 194 + ], + "type": "text", + "content": "Table 3: Ablation studies for 3D adaptation components on Real3D-AD dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 228, + 236, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 236, + 240 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 236, + 240 + ], + "type": "text", + "content": "4.5 Ablation study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 247, + 481, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 481, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 481, + 297 + ], + "type": "text", + "content": "To delve into the effect of individual components, we conduct ablation experiments on the Real3D-AD dataset. To fully demonstrate and compare the performance of the models, we report both image-level and pixel-level results with I-AUROC and P-AUROC, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 312, + 482, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 482 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 482 + ], + "type": "text", + "content": "Main component Table 3 compares the performance of different variants from R3D-AD, which includes the influence of the denoising condition embedding, displacement-based reconstruction way, and the data augmentation strategy of Patch-Gen. Model A is denoted as our baseline, which is a vanilla DDPM model for point cloud reconstruction. Introducing a condition into the DDPM (Model B) significantly boosts performance, particularly in terms of I-AUROC, which sees a " + }, + { + "bbox": [ + 130, + 312, + 482, + 482 + ], + "type": "inline_equation", + "content": "13.8\\%" + }, + { + "bbox": [ + 130, + 312, + 482, + 482 + ], + "type": "text", + "content": " increase to 0.667. Model C, which predicts point displacements based on conditional DDPM, preserving detailed structural information while accommodating the relative displacement of points contributes to a notable " + }, + { + "bbox": [ + 130, + 312, + 482, + 482 + ], + "type": "inline_equation", + "content": "6.0\\%" + }, + { + "bbox": [ + 130, + 312, + 482, + 482 + ], + "type": "text", + "content": " gain in P-AUROC over Model B. Model D is trained under the conditions of shape embedding with the Patch-Gen strategy. Considering that the defective portion contains only a small portion of the original point cloud, we try to reconstruct the relative displacement in a way that preserves as much detail as possible, which is effective for both 3D anomaly detection and segmentation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 481, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 521 + ], + "type": "text", + "content": "Patch-Gen Table. 4 analyzes the influence of two key parameters in Patch-Gen: the selection points ratio and the scaling points factor." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 522, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 640 + ], + "type": "text", + "content": "The selection points ratio from Table. 4a determines the proportion of points in the point cloud that are selected for transformation. Our findings suggest that a selection ratio of " + }, + { + "bbox": [ + 130, + 522, + 482, + 640 + ], + "type": "inline_equation", + "content": "1/32" + }, + { + "bbox": [ + 130, + 522, + 482, + 640 + ], + "type": "text", + "content": " achieves the best performance. It appears that this ratio provides a balanced trade-off between maintaining sufficient structure for anomaly detection and introducing enough variation to simulate anomalies effectively. Notably, as the ratio increases beyond " + }, + { + "bbox": [ + 130, + 522, + 482, + 640 + ], + "type": "inline_equation", + "content": "1/16" + }, + { + "bbox": [ + 130, + 522, + 482, + 640 + ], + "type": "text", + "content": ", both I-AUROC and P-AUROC scores decrease in severity, since real defects only account for a small portion of the overall point cloud, a wide selection of points not only destroys the structure of the original point cloud, but also makes the distribution of the training and test sets inconsistent." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "The scaling points factor is the intensity of the random transformation applied to the selected points, as detailed in Table 4b. The optimal performance is" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "Z. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 190, + 113, + 304, + 181 + ], + "blocks": [ + { + "bbox": [ + 190, + 113, + 304, + 181 + ], + "lines": [ + { + "bbox": [ + 190, + 113, + 304, + 181 + ], + "spans": [ + { + "bbox": [ + 190, + 113, + 304, + 181 + ], + "type": "table", + "html": "
ratioI-AUROCP-AUROC
1/640.7160.584
1/320.7340.592
1/160.7270.579
1/80.6830.528
", + "image_path": "a91c5ed4ddf9ced017aa5607e8a2fe1adeb799c55f72b04dc5fdaf52b1231fc9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 308, + 114, + 427, + 182 + ], + "blocks": [ + { + "bbox": [ + 195, + 182, + 298, + 192 + ], + "lines": [ + { + "bbox": [ + 195, + 182, + 298, + 192 + ], + "spans": [ + { + "bbox": [ + 195, + 182, + 298, + 192 + ], + "type": "text", + "content": "(a) Selection points ratio." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 114, + 427, + 182 + ], + "lines": [ + { + "bbox": [ + 308, + 114, + 427, + 182 + ], + "spans": [ + { + "bbox": [ + 308, + 114, + 427, + 182 + ], + "type": "table", + "html": "
factorI-AUROCP-AUROC
0.10.7340.592
0.20.7270.572
0.40.7150.554
0.80.6610.517
", + "image_path": "41a4ac21da48477f1a60024c075876390d2a4a42b9b9bbd67fc9773b55662eed.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 315, + 182, + 415, + 192 + ], + "lines": [ + { + "bbox": [ + 315, + 182, + 415, + 192 + ], + "spans": [ + { + "bbox": [ + 315, + 182, + 415, + 192 + ], + "type": "text", + "content": "(b) Scaling points factor." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 131, + 201, + 479, + 224 + ], + "lines": [ + { + "bbox": [ + 131, + 201, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 131, + 201, + 479, + 224 + ], + "type": "text", + "content": "Table 4: Ablation studies for Patch-Gen implementation on Real3D-AD dataset. Default settings are marked in gray." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 137, + 252, + 302, + 344 + ], + "blocks": [ + { + "bbox": [ + 137, + 252, + 302, + 344 + ], + "lines": [ + { + "bbox": [ + 137, + 252, + 302, + 344 + ], + "spans": [ + { + "bbox": [ + 137, + 252, + 302, + 344 + ], + "type": "image", + "image_path": "e6ff16748f29b8ed19dce6e60f2b302acbe66932d90b7223ee0b54508a5bb848.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 351, + 221, + 359 + ], + "lines": [ + { + "bbox": [ + 212, + 351, + 221, + 359 + ], + "spans": [ + { + "bbox": [ + 212, + 351, + 221, + 359 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 255, + 474, + 350 + ], + "blocks": [ + { + "bbox": [ + 309, + 255, + 474, + 350 + ], + "lines": [ + { + "bbox": [ + 309, + 255, + 474, + 350 + ], + "spans": [ + { + "bbox": [ + 309, + 255, + 474, + 350 + ], + "type": "image", + "image_path": "8e47002cd3a71ddd50d9a1c6e5d39e3739396e30da9e3e4b36a840d07aebe135.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 398, + 351, + 406, + 359 + ], + "lines": [ + { + "bbox": [ + 398, + 351, + 406, + 359 + ], + "spans": [ + { + "bbox": [ + 398, + 351, + 406, + 359 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 369, + 482, + 414 + ], + "lines": [ + { + "bbox": [ + 131, + 369, + 482, + 414 + ], + "spans": [ + { + "bbox": [ + 131, + 369, + 482, + 414 + ], + "type": "text", + "content": "Fig. 4: Memory and time cost during inference on Real3D-AD dataset. (a) Memory usage comparison between different models. (b) 3D anomaly detection performance vs. frames per second on an NVIDIA RTX 3090 GPU. Our R3D-AD outperforms all previous methods on both accuracy and efficiency by a significant margin." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 439, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 439, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 482, + 499 + ], + "type": "text", + "content": "observed at a scaling factor of 0.1, which implies that minor transformations are more effective for simulating anomalies without significantly altering the original data distribution. Larger scaling factors lead to a consistent decline in performance, underscoring the importance of subtle transformations for preserving the utility of the simulated anomalies for detection tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Memory and time cost As depicted in Figure 4, we evaluate the disparity in both storage consumption and inference time of our model under identical experimental conditions, compared to existing methods. Regarding memory usage, our approach demonstrates a marked superiority by employing raw coordinate features instead of FPFH or PointMAE features, significantly reducing the memory footprint. Since no memory bank exists, our method is also more space-efficient compared to BTF which also uses raw features. Moreover, our method eliminates the necessity to compare all the features in memory, substantially increasing operational efficiency. The implementation of Patch-Gen inherently bestows our model with exceptional robustness, enabling precise reconstruction of point clouds from various angles without the need for the time-intensive RANSAC alignment process required by Reg3D-AD." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 179, + 90, + 447, + 102 + ], + "type": "text", + "content": "R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 114, + 476, + 258 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 476, + 258 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 476, + 258 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 476, + 258 + ], + "type": "image", + "image_path": "4bfac1c3112a59a58b7e6a096b999e1658417378b8cbe516422c3f7190ea7e2e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 266, + 482, + 300 + ], + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 300 + ], + "type": "text", + "content": "Fig. 5: Qualitative analysis on Real3D-AD dataset and Anomaly-ShapeNet dataset. The anomaly map is obtained directly by calculating the differences between the input and reconstructed point clouds, where deeper colors represent more confidence." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 324, + 254, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 324, + 254, + 337 + ], + "spans": [ + { + "bbox": [ + 132, + 324, + 254, + 337 + ], + "type": "text", + "content": "4.6 Qualitative results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 345, + 482, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 482, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 482, + 465 + ], + "type": "text", + "content": "Figure 5 presents some qualitative outcomes, with varying shades of color indicating different levels of anomaly scores. We select several representative defective samples to demonstrate the robustness of our algorithm. The left four columns display samples from Real3D-AD, while the right four columns samples are from Anomaly-ShapeNet. The illustration reveals that our R3D-AD algorithm has precisely reconstructed the defective portions of the point cloud across various samples: the deep sink in the Seahorse sample, the concavity in the Bag sample, and the bulge in the Jar sample. Leveraging the accurately reconstructed point clouds, final point cloud segmentation maps are also produced, further evidencing the efficacy of our approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 483, + 220, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 220, + 496 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 220, + 496 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "In this work, we presented R3D-AD, a novel reconstructive 3D anomaly detection model based on conditional diffusion. Our goal is to overcome the limitations faced by current 3D anomaly detection methods, such as the inefficiencies due to the memory bank module and low performance caused by incorrect rebuilds with MAE. To address these challenges, we leverage the diffusion process for full reconstruction, followed by a direct comparison between the input and the reconstructed point cloud to obtain the final anomaly score. The embedded latent variable that spans the decoding process, step-wisely generating point-level displacements from the noise to the target anomaly-free sample. We also propose Patch-Gen, a data augmentation tailored for point cloud anomaly simulation. Extensive experiments conducted on 3D anomaly benchmarks validate the superiority of our R3D-AD in comparison to state-of-the-art alternatives in terms of both accuracy and versatility." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "Z. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 187 + ], + "type": "text", + "content": "This work was supported in part by the Pioneer and Leading Goose R&D Program of Zhejiang (Grant No. 2022C01051), in part by the National Natural Science Foundation of China (Grant No. 52375271, 52275274), and in part by the Natural Science Foundation of Zhejiang Province (Grant No. LY23E050011)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 201, + 197, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 201, + 197, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 201, + 197, + 215 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 224, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 138, + 224, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 480, + 246 + ], + "type": "text", + "content": "1. Bae, J., Lee, J.H., Kim, S.: Pni: Industrial anomaly detection using position and neighborhood information. In: ICCV (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 247, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 247, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 247, + 480, + 277 + ], + "type": "text", + "content": "2. Bergmann, P., Löwe, S., Fauser, M., Sattlegger, D., Steger, C.: Improving unsupervised defect segmentation by applying structural similarity to autoencoders. In: VISIGRAPP (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 279, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 279, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 279, + 480, + 300 + ], + "type": "text", + "content": "3. Cao, Y., Xu, X., Shen, W.: Complementary pseudo multimodal feature for point cloud anomaly detection. arXiv preprint (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 300, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 300, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 480, + 332 + ], + "type": "text", + "content": "4. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., Xiao, J., Yi, L., Yu, F.: ShapeNet: An Information-Rich 3D Model Repository. arXiv preprint (2015)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 333, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 333, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 138, + 333, + 480, + 354 + ], + "type": "text", + "content": "5. Charles, R.Q., Su, H., Kaichun, M., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: CVPR (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 354, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 480, + 374 + ], + "type": "text", + "content": "6. Chu, R., Xie, E., Mo, S., Li, Z., Nießner, M., Fu, C.W., Jia, J.: Diffcomplete: Diffusion-based generative 3d shape completion. In: NeurIPS (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 376, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 480, + 396 + ], + "type": "text", + "content": "7. Dinh, L., Sohl-Dickstein, J., Bengio, S.: Density estimation using real nvp. In: ICLR (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 396, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 480, + 418 + ], + "type": "text", + "content": "8. Gudovskiy, D., Ishizaka, S., Kozuka, K.: Cflow-ad: Real-time unsupervised anomaly detection with localization via conditional normalizing flows. In: WACV (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 418, + 480, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 480, + 438 + ], + "type": "text", + "content": "9. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 439, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 439, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 134, + 439, + 480, + 472 + ], + "type": "text", + "content": "10. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., Salimans, T.: Imagen video: High definition video generation with diffusion models (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 472, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 472, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 134, + 472, + 480, + 493 + ], + "type": "text", + "content": "11. Horwitz, E., Hoshen, Y.: Back to the feature: Classical 3d features are (almost) all you need for 3d anomaly detection. In: CVPRW (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 494, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 494, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 134, + 494, + 480, + 525 + ], + "type": "text", + "content": "12. Hu, T., Zhang, J., Yi, R., Du, Y., Chen, X., Liu, L., Wang, Y., Wang, C.: Anomalydiffusion: Few-shot anomaly image generation with diffusion model. In: AAAI (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 525, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 525, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 134, + 525, + 480, + 546 + ], + "type": "text", + "content": "13. Jonathan Ho, Ajay Jain, and Pieter Abbeel: Denoising diffusion probabilistic models. In: NeurIPS (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 547, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 547, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 134, + 547, + 480, + 568 + ], + "type": "text", + "content": "14. Kim, D., Park, C., Cho, S., Lee, S.: Fapm: Fast adaptive patch memory for real-time industrial anomaly detection. In: ICASSP (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 569, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 569, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 134, + 569, + 480, + 590 + ], + "type": "text", + "content": "15. Kong, Z., Ping, W., Huang, J., Zhao, K., Catanzaro, B.: Diffwave: A versatile diffusion model for audio synthesis. In: ICLR (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 134, + 590, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 590, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 134, + 590, + 480, + 611 + ], + "type": "text", + "content": "16. Li, C.L., Sohn, K., Yoon, J., Pfister, T.: Cutpaste: Self-supervised learning for anomaly detection and localization. In: CVPR (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 134, + 612, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 612, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 612, + 480, + 632 + ], + "type": "text", + "content": "17. Li, M., Duan, Y., Zhou, J., Lu, J.: Diffusion-sdf: Text-to-shape via voxelized diffusion. In: CVPR (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "text", + "content": "18. Li, W., Xu, X., Gu, Y., Zheng, B., Gao, S., Wu, Y.: Towards scalable 3d anomaly detection and localization: A benchmark via 3d anomaly synthesis and a self-supervised learning network. arXiv preprint (2023)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "text", + "content": "R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "19. Liu, J., Xie, G., Li, X., Wang, J., Liu, Y., Wang, C., Zheng, F., et al.: Real3d-ad: A dataset of point cloud anomaly detection. In: NeurIPS (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "type": "text", + "content": "20. Liu, Z., Feng, Y., Black, M.J., Nowrouzezahrai, D., Paull, L., Liu, W.: Meshdiffusion: Score-based generative 3d mesh modeling. In: ICLR (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "type": "text", + "content": "21. Liu, Z., Tang, H., Lin, Y., Han, S.: Point-voxel cnn for efficient 3d deep learning. In: NeurIPS (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "type": "text", + "content": "22. Lu, F., Yao, X., Fu, C., Jia, J.: Removing anomalies as noises for industrial defect localization. In: ICCV (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "text", + "content": "23. Luo, S., Hu, W.: Diffusion probabilistic models for 3d point cloud generation. In: CVPR (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 224, + 480, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 245 + ], + "type": "text", + "content": "24. Mo, S., Xie, E., Chu, R., Hong, L., Nießner, M., Li, Z.: Dit-3d: Exploring plain diffusion transformers for 3d shape generation. In: NeurIPS (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "type": "text", + "content": "25. Pang, Y., Wang, W., Tay, F.E., Liu, W., Tian, Y., Yuan, L.: Masked autoencoders for point cloud self-supervised learning. In: ECCV (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "type": "text", + "content": "26. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. NeurIPS (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 300, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 320 + ], + "type": "text", + "content": "27. Rezende, D.J., Mohamed, S.: Variational inference with normalizing flows. In: ICML (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 321, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 480, + 342 + ], + "type": "text", + "content": "28. Roth, K., Pemula, L., Zepeda, J., Scholkopf, B., Brox, T., Gehler, P.: Towards total recall in industrial anomaly detection. In: CVPR (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 343, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 480, + 363 + ], + "type": "text", + "content": "29. Rudolph, M., Wandt, B., Rosenhahn, B.: Same same but differnet: Semi-supervised defect detection with normalizing flows. In: WACV (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "type": "text", + "content": "30. Rusu, R.B., Blodow, N., Beetz, M.: Fast point feature histograms (fpfh) for 3d registration. In: ICRA (2009)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 386, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 429 + ], + "type": "text", + "content": "31. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Ghasemipour, S.K.S., Ayan, B.K., Mahdavi, S.S., Lopes, R.G., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 430, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 480, + 450 + ], + "type": "text", + "content": "32. Schlüter, H.M., Tan, J., Hou, B., Kainz, B.: Natural synthetic anomalies for self-supervised anomaly detection and localization. In: ECCV (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 450, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 471 + ], + "type": "text", + "content": "33. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 472, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 503 + ], + "type": "text", + "content": "34. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: ICLR (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "type": "text", + "content": "35. Tailanian, M., Pardo, Å., Muse, P.: U-flow: A u-shaped normalizing flow for anomaly detection with unsupervised threshold. arXiv preprint (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 525, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 546 + ], + "type": "text", + "content": "36. Wang, Y., Peng, J., Zhang, J., Yi, R., Wang, Y., Wang, C.: Multimodal industrial anomaly detection via hybrid fusion. In: CVPR (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 547, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 480, + 568 + ], + "type": "text", + "content": "37. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph cnn for learning on point clouds. TOG (2019)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 569, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 480, + 590 + ], + "type": "text", + "content": "38. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: Simmim: A simple framework for masked image modeling. In: CVPR (2022)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 590, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 621 + ], + "type": "text", + "content": "39. Yu, J., Zheng, Y., Wang, X., Li, W., Wu, Y., Zhao, R., Wu, L.: Fastflow: Unsupervised anomaly detection and localization via 2d normalizing flows. arXiv preprint (2021)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "text", + "content": "40. Yu, X., Tang, L., Rao, Y., Huang, T., Zhou, J., Lu, J.: Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In: CVPR (2022)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "text", + "content": "41. Zavrtanik, V., Kristan, M., Skočaj, D.: Draem-a discriminatively trained reconstruction embedding for surface anomaly detection. In: ICCV (2021)" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "Z. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 182 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "42. Zavrtanik, V., Kristan, M., Skočaj, D.: Reconstruction by inpainting for visual anomaly detection. Pattern Recognition (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 480, + 160 + ], + "type": "text", + "content": "43. Zhang, X., Xu, M., Zhou, X.: Realnet: A feature selection network with realistic synthetic anomaly for anomaly detection. In: CVPR (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 182 + ], + "type": "text", + "content": "44. Zhou, L., Du, Y., Wu, J.: 3d shape generation and completion through point-voxel diffusion. In: ICCV (2021)" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "text", + "content": "R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/R3DS_ Reality-linked 3D Scenes for Panoramic Scene Understanding/7708992f-16b8-4792-b8c3-10a97f869f8d_content_list.json b/2024/R3DS_ Reality-linked 3D Scenes for Panoramic Scene Understanding/7708992f-16b8-4792-b8c3-10a97f869f8d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7f05516ea5a3b9f88d6df7c0f32f802d492cbbf3 --- /dev/null +++ b/2024/R3DS_ Reality-linked 3D Scenes for Panoramic Scene Understanding/7708992f-16b8-4792-b8c3-10a97f869f8d_content_list.json @@ -0,0 +1,1575 @@ +[ + { + "type": "text", + "text": "R3DS: Reality-linked 3D Scenes for Panoramic Scene Understanding", + "text_level": 1, + "bbox": [ + 294, + 141, + 709, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qirui Wu $^{1}$ , Sonia Raychaudhuri $^{1}$ , Daniel Ritchie $^{2}$ , Manolis Savva $^{1}$ , and Angel X. Chang $^{1,3}$", + "bbox": [ + 243, + 210, + 759, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Simon Fraser University", + "bbox": [ + 411, + 253, + 589, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Brown University", + "bbox": [ + 434, + 268, + 566, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Alberta Machine Intelligence Institute (Amii)", + "bbox": [ + 341, + 282, + 658, + 297 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ff7acf139001c077257d8e97088718c8a412a9fa3ad7cf529f28962eaddf32c9.jpg", + "image_caption": [ + "Fig. 1: Left: the Reality-linked 3D Scenes dataset (R3DS) fills a gap between synthetic 3D scenes and reconstructions of real-world environments by providing 3D scene proxies linked to real-world panoramas from Matterport3D (three example panoramas and 3D scenes shown). Right: our dataset contains scenes with higher density and completeness compared to prior datasets, and provides additional annotations such as object support (what objects or architectural elements support other objects), and matching object sets (e.g., pairs of the same nightstand). We use our dataset for the panoramic scene understanding task and demonstrate its value for research on room layout estimation, as well as 2D and 3D object detection." + ], + "image_footnote": [], + "bbox": [ + 218, + 324, + 782, + 498 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We introduce the Reality-linked 3D Scenes (R3DS) dataset of synthetic 3D scenes mirroring the real-world scene arrangements from Matterport3D panoramas. Compared to prior work, R3DS has more complete and densely populated scenes with objects linked to real-world observations in panoramas. R3DS also provides an object support hierarchy, and matching object sets (e.g., same chairs around a dining table) for each scene. Overall, R3DS contains 19K objects represented by 3,784 distinct CAD models from over 100 object categories. We demonstrate the effectiveness of R3DS on the Panoramic Scene Understanding task. We find that: 1) training on R3DS enables better generalization; 2) support relation prediction trained with R3DS improves performance compared to heuristically calculated support; and 3) R3DS offers a challenging benchmark for future work on panoramic scene understanding.", + "bbox": [ + 259, + 659, + 743, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 143, + 375, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Datasets of 3D indoor environments are increasingly used for research on scene understanding [1, 3, 28], embodied AI [2, 19, 23], and scene generation [15, 25]. There are two strategies for constructing 3D scene datasets: reconstruction of real-world spaces [3], or authoring scenes using synthetic 3D objects [8] (CAD models). Reconstruction captures real spaces but is hard to scale, and the resulting scenes exhibit imperfections and artifacts. On the other hand, synthetic 3D scenes are complete and easy to manipulate but often do not match the statistics of real-world spaces and are artificially \"clean\". Moreover, both strategies are time-consuming and require expertise. There have been some attempts to create \"synthetic\" replicas of real environments by matching CAD models to objects in scans [19, 23]. These efforts have been limited in scale and often result in partial and sparsely populated synthetic counterparts of the real environments.", + "bbox": [ + 212, + 183, + 787, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We design a framework that allows users to create 3D scenes from RGB panoramas and use it to create R3DS: a dataset of 'Reality-linked' 3D Scenes. Each 3D scene in our dataset is a complete proxy of an environment from the Matterport3D [3] dataset, representing both the 3D architecture and the objects. Thus, each scene is linked to a real space, with correspondences established between panorama observations of each object and the synthetic object. These reality-linked scenes reflect denser real-world arrangements of objects.", + "bbox": [ + 212, + 364, + 787, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The use of panoramas for reference is advantageous compared to either perspective images or 3D reconstructions. Panoramas are not limited by the field of view unlike perspective images, enabling more complete 3D synthetic proxies. Panoramas also better capture relatively small objects and objects with challenging materials or illumination conditions compared to reconstructions. Additionally, there is a scarcity of synthetic 3D scenes coupled with real-world panoramas, with only one relatively small algorithmically constructed dataset provided by Zhang et al. [28] being available to the community.", + "bbox": [ + 212, + 473, + 787, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compared to prior efforts such as Scan2CAD [1] and CAD-Estate [14], our dataset provides more complete scenes, with salient observed objects being captured in the layout. Moreover, we provide a support hierarchy defining what objects are placed on other objects and specify sets of identical objects such as dining chairs around a table, allowing for creating realistic variations of the scene by swapping the entire set to a different chair design.", + "bbox": [ + 212, + 594, + 787, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We demonstrate the value of our dataset by using it for the Panoramic Scene Understanding task. We show that leveraging the denser layouts and support hierarchy information in our scenes leads to improved object detection performance and better generalization compared to training using other datasets previously used for this task. In summary, we make the following contributions:", + "bbox": [ + 212, + 686, + 787, + 763 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We design a framework for efficient construction of synthetic scenes from real panoramas and use it to create R3DS: a dataset of reality-linked 3D scenes.", + "- R3DS provides more complete and realistic scenes with correspondences between real and synthetic objects, and object-object support relations." + ], + "bbox": [ + 223, + 777, + 782, + 837 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Q. Wu et al.", + "bbox": [ + 271, + 114, + 357, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "R3DS (Ours): linked to panoramicas, densely populated scenes, matched object sets, objects supported", + "bbox": [ + 223, + 148, + 750, + 160 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/13caafa1cafea1e0f7f812fec3dcbfb2055dd8bb5c1972f496effc9bb7d6bc79.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 161, + 782, + 258 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8f160698894748f4a793518d06989ec6f99b5f51bfe62065b613eecca100ea01.jpg", + "image_caption": [ + "CAD-Estate: sparse, floating", + "Fig. 2: Dataset comparison. (Top) shows different views of a scene annotated in R3DS. Comparison with previous datasets (bottom) shows (1) R3DS has more complete scenes than the previous datasets; (2) Objects in R3DS are properly supported by either architecture or other objects unlike the others (e.g. floating objects with no proper support); (3) R3DS is annotated using the same 3D model for objects arranged together (chairs by the dining table, couches arranged together)." + ], + "image_footnote": [], + "bbox": [ + 222, + 260, + 411, + 345 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b3a9e4b37345c3acd518271a6033b3db202843053028d532bdb00b9aaec8d748.jpg", + "image_caption": [ + "Scan2CAD: sparse, floating, mismatched" + ], + "image_footnote": [], + "bbox": [ + 418, + 260, + 616, + 345 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bea44ff58d17b393c906c9f9715fda51b081eb836a52841492dcbfd6359c31a5.jpg", + "image_caption": [ + "iGibson DPC: floating" + ], + "image_footnote": [], + "bbox": [ + 620, + 261, + 782, + 344 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We show that the more complete layouts and support relations in our dataset enable better performance and generalization in the Panoramic Scene Understanding task, and that our dataset offers a challenging benchmark for future work in scene understanding.", + "bbox": [ + 225, + 487, + 787, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 580, + 387, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D scene datasets. A spectrum of scene datasets have been used for scene understanding tasks. One type provides annotated 3D reconstructions of real scenes based on RGB-D videos [3, 6, 10, 16, 20, 27]. These datasets are usually subject to the limitations of RGB-D reconstruction, typically containing noise, artifacts such as holes, and poor reconstructions of thin structures, shiny objects, or light sources. Another type of 3D datasets is authored by manually designing 3D object assets [5, 9] and inserting them into synthetic 3D scenes [8]. However, such datasets lack the realism of real-world reconstructions and demand expert knowledge, making them expensive to create. A third, hybrid approach which is closest to our work creates 3D scene datasets by aligning existing object CAD models to real world data.", + "bbox": [ + 212, + 613, + 787, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Datasets that align CAD models to real world. There have been a number of recent efforts in aligning CAD models with real-world data. Prior work [13, 22, 26] has annotated object images with 3D models, typically using keypoint correspondences to perspective images. These perspective images usually do not", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "R3DS", + "bbox": [ + 689, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "depict a complete scene; they typically focus on one or two objects and are limited in field of view, resulting in a sparse proxy of the real scene.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Another line of work aligns 3D CAD models to RGB-D scans either through annotation as in Scan2CAD [1], or automated heuristics as in iGibson [19]. OpenRooms [12] extends Scan2CAD [1] with photorealistic material annotations and focuses on inverse rendering tasks. Conceptually, these allow for more complete synthetic scene proxies. However, statistics from these datasets show that they are still relatively sparse (see Tab. 1). In addition, the poor quality of reconstruction makes aligning CAD models challenging without referring to the original RGB images. A prominent exception is Replica [20] which has fairly high-quality reconstructions and the artist-created Replica-CAD [23]. However, creating such high quality \"replicas\" is labor intensive and costly. Szot et al. [23] report $900+$ work hours required to model approximately 90 objects, resulting in a dataset of limited scale with 105 different layouts of what is effectively a single room.", + "bbox": [ + 212, + 176, + 785, + 356 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "More recently, Maninis et al. [14] introduced CAD-Estate, which aligns CAD models to RGB videos for over 19K spaces. Because the data is based on monocular video, the coverage of the spaces is incomplete. In addition, the annotation is relatively sparse, with an average of only 6 objects per scene.", + "bbox": [ + 212, + 357, + 785, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets for panoramic scene understanding. There have been relatively few datasets introduced for Panoramic Scene Understanding [7, 28, 29]. In the initial PanoContext dataset [29], the data did not have aligned CAD models and only included object cuboids. The ground truth data was collected on 2D panorama images by annotating visible cuboid vertices; 3D cuboids were obtained by minimizing the re-projection error from the annotated 2D vertices. Moreover, these 3D cuboids and the room layout are obtained with the assumption that the room layout is a cuboid and that the objects are vertically aligned. Thus, the resulting object layout may deviate from the real arrangement of objects. More recently, datasets for Panoramic Scene Understanding have been built by taking 3D scans, aligning CAD objects to them, and then generating panoramas [7, 28]. Compared to these datasets, our R3DS is manually curated for a larger number of distinct regions and provides support hierarchy and matching object set annotations.", + "bbox": [ + 212, + 417, + 787, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 The R3DS Dataset", + "text_level": 1, + "bbox": [ + 215, + 648, + 434, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We describe the construction of the R3DS dataset and present a statistical analysis of the scenes it contains. Compared to previous datasets [1, 14, 28] (Fig. 2), our scenes are more densely populated, and objects are annotated with a hierarchy of support relations. Moreover, our dataset specifies matching object instances in furniture arrangements. Figure 1 shows example annotations from our dataset.", + "bbox": [ + 212, + 676, + 785, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Dataset construction", + "text_level": 1, + "bbox": [ + 215, + 787, + 434, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We developed a 3D annotation interface (Fig. 3) showing a panorama of a room from Matterport3D and allowing users to insert 3D CAD objects into a 3D scene", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Q. Wu et al.", + "bbox": [ + 271, + 114, + 357, + 128 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/598ba0463e6e92e4268e3ef92dbca919995f53465674820db96ea540fe02a979.jpg", + "table_caption": [ + "Table 1: Comparison with 3D indoor scene datasets aligned with real-world images, videos, or scans. Our R3DS dataset contains more densely populated annotations compared to other datasets, with objects from 110 different categories. We report the unique models (#CAD), object categories (#Cat), object instances (#Objects) as well as average number of objects and object categories per annotation. For Scan2Cad [1] and ReplicaCAD [23] the average is per scan. Note that ReplicaCAD consists of 105 different layouts (arrangements) of effectively one room. CAD-Estate has partial views into 19K spaces, many of which are 1-2 rooms. Of the datasets used for panoramic scene understanding, our R3DS dataset covers more rooms with 842 panoramicas over 22 room types. Replica-Pano (in gray) was not released, so we report statistics from the paper. Our annotations per panorama are more complete and our dataset has both support relations (Sup) and matching object instance sets (Match)." + ], + "table_footnote": [], + "table_body": "
DatasetSourceCAD AlignmentTypeHouses/RoomsPanos#CAD#Objects#CatAveObjAveCatSupMatch
Scan2CAD [1]ScanNet [6]Annotatorscan- / 1506✘3,04914,225359.44.1✘✘
OpenRooms [12]ScanNet [6]Scan2CAD [1]scan- / 1288✘2,65116,0143812.46.3✘✘
ReplicaCAD [23]Replica [20]Artist recreationscan- / 105*✘922,2934421.814.4✘✘
CAD-Estate [14]RealEstate10K [31]Annotatorvideo19,512✘12,024100,882496.33.4✘✘
Replica-Pano [7]Replica [20]Heuristicpano- / 272700--25-✘✘
iGibson-DPC [28]iGibson [19]Heuristicpano15 / 100150050026,9985717.910.2✘✘
R3DS (Ours)Matterport3D [3]Annotatorpano20 / 3708423,78419,05011022.910.4✓✓
", + "bbox": [ + 217, + 323, + 787, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which is visually overlaid on the panorama. The 3D scene is initially empty, consisting only of 3D architectural geometry which specifies the walls, floor, ceiling as well as the placement of openings (e.g. doors, windows, and other openings) on the walls. We create this 3D architecture by taking 20 houses from Matterport3D, constructing an initial architecture based on the region and object annotations for the windows and doors, and manually refining the placement of walls and openings. By combining panoramas and 3D architectures, users can see through openings and annotate objects located in other rooms.", + "bbox": [ + 212, + 441, + 787, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We ask annotators to select and place 3D object models to best match the panoramic image. We use CAD models from Wayfair [17] and ShapeNet [4] models collected from 3D Trimble warehouse. Wayfair provides a large collection of furniture CAD models that match real-world products and are sized based on real-world dimensions. However, it does not include bathroom fittings, electronic equipment and kitchen appliances, for which we manually scale and align CAD models from ShapeNet. Compared with ShapeNetCore, the CAD models we use are already sized to real-world sizes (instead of normalized to a unit cube).", + "bbox": [ + 212, + 566, + 787, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To assist the annotators, we provide segmented masks of objects visible in the panorama. Since Matterport3D has annotated 3D object masks on the scans we use those annotations, but it is also possible to run an instance segmentation on the panorama. When the user clicks one of these masks, a search panel automatically opens and shows objects matching the clicked mask category label. For each mask, the annotator selects a matching object and positions and aligns it to match the mask. Annotators are instructed to choose objects which match the shape of the corresponding object in the panorama (rather than its color or texture). To help annotators focus on shape, we render all 3D objects in a neutral gray color. Annotators are also explicitly asked to select the same", + "bbox": [ + 212, + 688, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "R3DS", + "bbox": [ + 689, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7d9a40641378deb27cc912483930a1134192010421f790ad506aed5c2e0cba91.jpg", + "image_caption": [ + "Fig. 3: R3DS annotation pipeline. Annotators see an empty scene (architecture only). They then insert and manipulate 3D object models from a panorama viewpoint to create a populated 3D scene proxy corresponding to the panorama." + ], + "image_footnote": [], + "bbox": [ + 289, + 148, + 712, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D asset for objects that should be the same; our interface provides a list of recently selected assets to make this process easier. In addition, annotators are instructed to add annotations for any objects that are not segmented (due to errors in Matterport3D) through simple clicks. These additional objects provide a more complete annotation that covers poorly reconstructed objects such as glass tables, lamps, and other small objects. The interface enforces that each object is placed on a support surface (either an architecture element or another object). The annotator can review their work by toggling off the panorama overlay or by switching to a perspective view of the 3D scene. For more annotated scene examples and details on the annotation process please refer to the supplement.", + "bbox": [ + 212, + 515, + 787, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Dataset analysis and statistics", + "text_level": 1, + "bbox": [ + 214, + 691, + 511, + 707 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We collect annotations for 20 Matterport3D houses with 808 panoramas in total. We discard panoramas taken on stairs or outside a house, since they have a limited number of objects that can be placed. After filtering we have 769 panoramas for our analysis and experiments. For 73 panoramas, we collect two sets of annotations for each, to obtain a total of 842 annotated object arrangements across 22 different Matterport3D region types. The panoramas with two annotations serve as a test of annotator consistency and add diversity. In total, R3DS contains 19,050 object instances from 3,784 unique 3D CAD models spanning", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Q. Wu et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fb06118c490c651e0c9ab388dc1f3fc796a2703aa2a609cf399a9452a64ea020.jpg", + "image_caption": [ + "Scan2CAD", + "Floor warped, no arch", + "Fig. 4: Architecture comparison. Compared to Scan2CAD (no architecture) and CAD-Estate (partial architecture), R3DS provides complete architecture with door/window portals." + ], + "image_footnote": [], + "bbox": [ + 245, + 167, + 416, + 241 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e4674f4724b0f8db038006ff740d63989cb02799e85284258465603dd3387ae9.jpg", + "image_caption": [ + "CAD-Estate", + "Partial arch" + ], + "image_footnote": [], + "bbox": [ + 437, + 166, + 606, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/65c45845923779202792ad615589436d7ef23755a7f9ae77441cc5b0ace51604.jpg", + "image_caption": [ + "R3DS (ours)", + "Complete arch with portals" + ], + "image_footnote": [], + "bbox": [ + 625, + 161, + 733, + 238 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "over 110 fine-grained object categories. Table 1 shows a comparison of overall statistics with previous 3D indoor scene datasets. See the supplement for more statistics about annotated objects.", + "bbox": [ + 212, + 349, + 784, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compared to prior datasets that align CAD models to real-world scenes, R3DS is more complete, providing annotated object support hierarchies and matching object instances. CAD-Estate [14] annotates RGB videos with 3D objects and architectural room layouts. However, the architecture is partial as the videos have limited view (Fig. 4), and not all objects in the scenes are annotated (Fig. 2). This results in annotations of objects floating mid-air and not properly supported (e.g. lamp Fig. 2). Scan2CAD [1] also lacks support structure (e.g., lamp not supported by cabinet in Fig. 2). In addition, because Scan2CAD does not provide clean 3D architecture on which objects are placed (Fig. 4), objects on the floor are not always placed such that their bottom face is parallel with a horizontal plane. In contrast, our R3DS scenes have an accurate support hierarchy by construction. OpenRooms [12] augments Scan2CAD with room layouts representing the architecture. However, the architecture in R3DS is more complex and realistic, especially due to inclusion of more doors (1.92 doors per room in R3DS vs 0.67 in OpenRooms).", + "bbox": [ + 212, + 396, + 785, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation of CAD object annotation quality is non-trivial as the 'ground truth' from the semantically annotated 3D reconstructions is itself imperfect. We measured how closely our annotated CAD objects conform to the real objects using the average 2D IoU between CAD object mask and ground truth 2D mask. R3DS is at $42.6\\%$ vs $38.5\\%$ for Scan2CAD, across 8 common object categories (bed, sofa, chair, cabinet, tv/monitor, table, shelving, bathtub).", + "bbox": [ + 212, + 626, + 784, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Of the datasets previously used for Panoramic Scene Understanding, ReplicaPano [7] has not been released, and iGibson-DPC [28] is the only dataset with synthetic panoramic images annotated with 3D objects and room layout. iGibson-DPC is built on scenes from iGibson [18] by randomly replacing objects with different models from the same category and rendering using the iGibson simulator to render panoramas. The selection and placement of objects in iGibson-DPC is based on heuristic algorithms, while our R3DS is manually annotated and placed 3D models are verified in terms of match and alignment to the object masks.", + "bbox": [ + 212, + 719, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "R3DS", + "bbox": [ + 689, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Moreover, iGibson-DPC contains unrealistic object arrangements (e.g., floating TV in Fig. 2).", + "bbox": [ + 215, + 146, + 784, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 200, + 374, + 217 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We showcase the value of R3DS on the Panoramic Scene Understanding (PanoSun) task [7, 28, 29]. Given an input RGB panorama, the goal is to estimate the room layout, detect objects in 2D, estimate their 3D oriented bounding boxes and also reconstruct 3D object meshes. Our experiments show that methods trained on R3DS data benefit from its realism and generalize better when evaluated on photorealistic images. We also investigate the role of object support hierarchy information in improving performance.", + "bbox": [ + 215, + 232, + 792, + 338 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Task setup", + "text_level": 1, + "bbox": [ + 215, + 361, + 352, + 376 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Method. DeepPanoContext (DPC) [28] predicts the room layout, detects objects in 3D and recovers object meshes from a panorama image using a relation-based graph convolutional network and a differentiable relation optimization procedure. Since DPC has a publicly-available implementation, we use it to benchmark the R3DS data on the PanoSun task. We keep all hyperparameters unchanged except lowering the relation optimization loss weight of 3D bounding box back-projection from 10 to 1, since the ground truth 2D masks are noisy.", + "bbox": [ + 215, + 386, + 785, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets. We train and evaluate DPC on the iGibson-DPC (IG) [7, 18, 28], Structured3D (S3D) [30], and R3DS datasets. Zhang et al. [28] render 1,500 panoramas from 15 iGibson houses composed of $500+$ objects spanning 57 object categories. We use the same data and splits for IG. Structured3D consists of 3500 houses and around 18K photo-realistic rendered panoramas in total. We use 14K for training and the remaining 4K for testing. Note that Structured3D does not provide ground truth object meshes.", + "bbox": [ + 215, + 492, + 785, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To prepare R3DS for this task, we generate the ground truth room layout from the 3D architecture based on the camera viewpoint and obtain 3D oriented bounding boxes (OBBs) from all objects. We use 2D object masks from the Matterport3D mesh instance segmentation. We consider three variants of R3DS based on the input panorama: $R3DS$ -real where we use the Matterport3D panoramas, $R3DS$ -syn where we use rendered panoramas (at the same camera poses) from the annotated synthetic scenes, and $R3DS$ -mix where we combine the two types of panoramas and double the available data. We follow the MP3D house split and merge the train and val sets to obtain a disjoint split of 15 train and 5 test houses. Based on the split, we have 696 annotated panoramas for train and 146 for test. To fairly evaluate methods trained on different datasets, we curate a list of 25 object classes common to all datasets.", + "bbox": [ + 215, + 598, + 785, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Metrics. Following Zhang et al. [28] we use separate metrics for room layout estimation, 3D object detection, and scene relation prediction. For room layout estimation, we use 2D IoU for predicted 2D floorplan, 3D IoU for lifted 3D room geometry, and dRMSE for predicted depth with respect to the camera location.", + "bbox": [ + 215, + 779, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Q. Wu et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/99614648e7ed69f5fba01d7e83d9fbc1193841977a68d0094089d3a0e632af20.jpg", + "table_caption": [ + "Table 2: Room layout estimation on R3DS-real test set. DPC [28] was pretrained on IG and S3D. For the last three rows, we fine-tune the pretrained weights on variants of R3DS." + ], + "table_footnote": [], + "table_body": "
Train2D IoU ↑3D IoU ↑dRMSE ↓
DPC [28]53.450.30.682
R3DS-real55.153.10.610
R3DS-syn59.056.10.629
R3DS-mix59.657.00.572
", + "bbox": [ + 359, + 198, + 633, + 285 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/e367b9c8d5c3f7ae81f5fe74b32202629c607a0a0927727d629cfcd24b9350c9.jpg", + "table_caption": [ + "Table 3: Cross-dataset evaluation for the Panoramic Scene Understanding task. We evaluate 3D detections with class-agnostic IoU and mAP at IoU of 0.15, and report object collisions. The highlighted rows indicate the most challenging scenario." + ], + "table_footnote": [], + "table_body": "
TestTrain3D detection ↑Collision ↓Attachment F1 ↑
IoUmAPmesharchobjwallfloor
IG27.530.31.6622.59453.176.895.086.2
IG+R3DS24.030.21.4042.25459.764.194.62.7
R3DS-real17.313.40.2421.45638.864.092.80.0
R3DS-syn23.214.20.4801.93848.546.793.828.6
R3DS-mix21.615.60.4341.24843.167.290.19.8
S3D19.53.51.0162.65150.968.790.811.6
IG+R3DS19.77.00.8682.08952.067.491.21.8
R3DS-real18.47.10.6002.59845.061.689.70.7
R3DS-syn19.04.80.6442.56149.349.791.22.4
R3DS-mix19.67.50.4631.67347.864.187.20.9
R3DS15.65.90.5751.95953.850.751.20.0
IG+R3DS17.514.10.2811.26749.561.658.60.0
R3DS-real16.415.00.2261.56244.057.358.90.0
R3DS-syn14.08.40.3901.66454.140.649.10.0
R3DS-mix17.615.80.1711.00748.558.360.10.0
", + "bbox": [ + 285, + 371, + 707, + 642 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For 3D object detection, we report bounding box-based class-agnostic 3D IoU as well as mean average precision (mAP) across the 25 object classes, where an IoU greater than 0.15 counts as a \"true\" result. For scene relation prediction, we report F1 scores for relation classification. We also report the average number of objects colliding with each other or with architectural structures (wall, floor, ceiling). Specifically, we follow Zhang et al. [28] and measure collisions using the Separating Axis Theorem (SAT) to test whether object bounding boxes overlap. Since bounding box-based collision is a poor proxy for real-world physical collision, we also compute mesh-based collision by checking if the meshes for object pairs have any interpenetrating triangles [11, 24].", + "bbox": [ + 215, + 688, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "R3DS", + "bbox": [ + 689, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/6d998a1d6b58433b80cea888a734c8aa9d4097684a20ec35bf88c28f72426518.jpg", + "image_caption": [ + "Fig. 5: Qualitative results for cross-dataset Panoramic Scene Understanding task. Correct and incorrect object detections shown in green and red boxes. Ground truth room layout and meshes shown in gray color, while layout prediction is in yellow. Training on R3DS leads to fewer errors compared to training on other datasets, especially when mixed with real data." + ], + "image_footnote": [], + "bbox": [ + 217, + 147, + 787, + 420 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Results", + "text_level": 1, + "bbox": [ + 215, + 532, + 323, + 546 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "1) Does R3DS help DPC generalize to real images? Since the original DPC work only trained and evaluated on synthetic data, it is unclear how well it performs on realistic panoramic imagery. We hypothesize that training on R3DS will lead to better performance. We separately show results on room layout estimation (Tab. 2) and 3D object detection.", + "bbox": [ + 212, + 565, + 787, + 642 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Room layout estimation. For room layout estimation, DPC uses Horizon-Net [21] pretrained on iGibson (IG) and Structured3D (S3D) panoramas. This model achieves good performance on IG data (91.0 3D IoU). When directly testing the official pretrained model on $R3DS$ -real panoramas, we notice a significant performance drop compared to results on the rendered panoramas from iGibson (Tab. 2 shows that the 3D IoU drops to 50.3). By finetuning the pretrained model with $R3DS$ -real, we can predict more precise room layouts for real cluttered scenes. Even only trained on $R3DS$ -syn, we outperform the original DPC model by $5.6\\%$ and $5.8\\%$ on 2D and 3D IoU, respectively. This is likely due to renderings from $R3DS$ -syn reflecting more realistic object arrangements in a room instead of pushing all objects against walls. Best performance on 2D IoU (59.6), 3D IoU (57.0) and depth RMSE (0.572) is achieved by fine tuning on $R3DS$ -mix.", + "bbox": [ + 212, + 643, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Q. Wu et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/b7b45d83d5ac5a112aedff829bd2408f7e134a9f579c5029115dc4efe6af2a4a.jpg", + "table_caption": [ + "Table 4: Performance of models trained on three synthetic datasets (IG, S3D, and IG+S3D) evaluated on the R3DS-full dataset, where \"full\" indicates all 840 panoramicas are used for testing." + ], + "table_footnote": [], + "table_body": "
Train3D detectionCollision ↓Support F1 ↑
IoU↑mAP↑mesharchobj floorceil
IG14.25.20.7031.6394.185.3
S3D16.410.00.1121.2263.186.8
IG+S3D17.19.70.1331.1623.384.8
", + "bbox": [ + 328, + 198, + 663, + 292 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Object detection. For 3D object detection, we train DPC on different data settings and conduct a cross-dataset evaluation (see Tab. 3). To investigate how models perform on out-of-distribution scenes, we evaluate models on Structured3D, as its images are near-realistic. To explore whether DPC training benefits from R3DS given the same amount of data, we create a special data input $IG + R3DS$ that combines iGibson and R3DS panoramas by randomly replacing half (500) of iGibson data with $R3DS$ -real data. The results show that $IG + R3DS$ performs almost the same as $IG$ with fewer collisions on iGibson, but it remarkably outperforms $IG$ on the test set of $R3DS$ and $S3D$ by 8.2 and 3.5 improvements on 3D mAP, respectively. It also averages 0.221 fewer mesh collisions. There are noticeable performance gaps on iGibson for models trained on R3DS data likely due to the data domain shift. Among the three variants of R3DS data, $R3DS$ -mix outperforms the others on all three test sets regarding 3D IoU and 3D mAP with the fewest mesh and architecture collisions. Although $R3DS$ -syn underperforms on $R3DS$ and $S3D$ test sets, it achieves better performance than $IG$ with even less data.", + "bbox": [ + 212, + 321, + 787, + 561 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Scene relation classification. We report F1 scores for identifying attachment relationships of objects to other objects and architecture elements (see Tab. 3). We note that models trained with synthetic renderings perform better than those trained on real images. That is because synthetic renderings present cleaner and simpler scenes with fewer objects than real world and simpler illumination such that DPC finds it easier to learn object-object and object-architecture relations. Also, note that the predictions of object-ceiling attachments can be extremely low because few objects are attached to the ceiling in the ground truth data. We show qualitative examples in Figure 5.", + "bbox": [ + 212, + 566, + 787, + 700 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "2) Is R3DS a challenging, high-quality test set? How would a model trained on pure synthetic data perform on complex real data (R3DS)? Due to its modest scale, we propose using R3DS as a challenging, high-quality test set rather than a train set. Specifically, we evaluate the synthetic-to-real performance of DPC by training on iGibson and/or Structured3D and testing on all panoramas in R3DS-real. Table 4 shows that a model trained with Structured3D performs the best (10.0 3D mAP and 0.112 mesh collision) as it observes the most photo-realistic images. DPC benefits from the synthetic data for higher bounding box IoUs, since it possesses accurately aligned 3D bounding box and", + "bbox": [ + 212, + 703, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "R3DS", + "bbox": [ + 689, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/9b5bba0ece1b1c2ef146afa2e3c9b4cdb4f8fed4eb9c9cfa099902f16bbe2847.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 143, + 357, + 205 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a9664d032672092e7f57a467f56a7953fff909237f7ec9ddf77b9985f25df455.jpg", + "image_caption": [ + "(a) Object appearance gap $(60\\%)$" + ], + "image_footnote": [], + "bbox": [ + 230, + 205, + 349, + 255 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/af2b5338e6663aae7efe41d48e274464aa8dad559ebd929a77d08be854d0fd42.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 143, + 498, + 205 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ac9a00bfd5a6f7027d94e75047d7b78b48be61cfa2c1629c5575d53bfd9ebf0f.jpg", + "image_caption": [ + "(b) Common occlusions (76.7%)" + ], + "image_footnote": [], + "bbox": [ + 357, + 205, + 498, + 255 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f1209d65168c46fc11e3a01cbe261163def8ebc142c1ec0b0c8bf2c89f9c1e36.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 143, + 638, + 205 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b88b0680aeb84e234cf6f8400cbda72a61ff6e309f4ee3d902175d61346a7730.jpg", + "image_caption": [ + "(c) Wrong 3D prediction $(65\\%)$" + ], + "image_footnote": [], + "bbox": [ + 504, + 205, + 640, + 253 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5bcc3a638f5c4300acc3018e8a9f30f76652bf422af6081d25b382bbbc681e14.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 143, + 781, + 205 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/826ed6e7d75414a4cdda1ec61f1d1e3d65f913f855118a62fad8dd3c9fe3f865.jpg", + "image_caption": [ + "(d) Objects in mirrors (15.6%)", + "Fig. 6: Error analysis on R3DS panoramas using DPC pretrained on S3D. Typical prediction errors are categorized into 4 groups: (a) Appearance gap between real and synthetic images. (b) Detection failure due to common occlusions. (c) Bad performance of monocular 3D prediction. (d) Perception error for objects in mirrors. Numbers in parentheses mean the error rates." + ], + "image_footnote": [], + "bbox": [ + 651, + 205, + 779, + 255 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/eda72c7838d71b3a9d0a4a0e3344f8a6eb1a8667be8c2df1715077c63d2fc584.jpg", + "image_caption": [ + "Panorama" + ], + "image_footnote": [], + "bbox": [ + 217, + 378, + 403, + 452 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/55e6b6e47c630c4f3c949b1d21ab86b15d074c3d90ec835db3e70d803eb0bc44.jpg", + "image_caption": [ + "Fig. 7: Comparison between heuristic and R3DS support relations. Dark lines between objects show support relations. Heuristic support often mistakenly assigns relations to spatially close objects. In contrast, R3DS support relations are annotated." + ], + "image_footnote": [], + "bbox": [ + 217, + 454, + 403, + 525 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3d6d39fd0b8998887e4edcb3f56eb6972c290eae77ff140de3f5e26c0e05cd56.jpg", + "image_caption": [ + "Heuristic Support" + ], + "image_footnote": [], + "bbox": [ + 408, + 378, + 594, + 450 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f35887575e5067bf1ce1c6f326266669a49d3d36e7fce4fa11abf4e3de870dca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 454, + 594, + 525 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/921e3c29ca87d8b788bf8410ff2af6d632dc69be794505a03a6ea0acb3543302.jpg", + "image_caption": [ + "R3DS Support" + ], + "image_footnote": [], + "bbox": [ + 599, + 378, + 784, + 452 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/606970567abdc77fe6603563009f963acaa8200aff9c26db499add1c4c2bfc57.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 454, + 784, + 525 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "more unoccluded objects. However, mAP performance is lower due to worse object recognition ability. All models struggle to predict correct object-wise support relations but do a better job of predicting object-floor support relations.", + "bbox": [ + 212, + 611, + 784, + 656 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conduct error analysis on 120 randomly sampled panoramas using the model pretrained on $S3D$ to identify typical errors (see Fig. 6). Errors are categorized into 4 groups: (a) $60\\%$ panoramas have 2D perception errors due to the synthetic-to-real appearance gap; (b) $76.7\\%$ panoramas show detection failures due to occlusions; (c) $65\\%$ panoramas exhibit correct 2D detections but fail to correctly perform 3D predictions; and (d) $15.6\\%$ out of 45 panoramas with mirrors mistakenly predict virtual objects in mirrors.", + "bbox": [ + 212, + 657, + 785, + 763 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "3) Are R3DS support relations helpful for PanoSun? We investigate whether the support relationships between objects provided in our R3DS scene hierarchy help boost performance of holistic scene understanding. We augment DPC's Relation Scene-GCN module with additional support relation prediction branches. Besides obtaining explicitly annotated scene support relations from", + "bbox": [ + 212, + 763, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Q. Wu et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/2ab00e23b3e3b222fac516f51ca6d27fb174e0b59a778ffbd66c87f06b4b6d5e.jpg", + "table_caption": [ + "Table 5: Performance on R3DS-real of DPC models trained on variants of R3DS with different support relation settings. We compare the original model without support (none) against models supervised with support that is heuristically computed (heur) or annotated from R3DS (anno). Classification results are evaluated on annotated ground-truth scene hierarchy." + ], + "table_footnote": [], + "table_body": "
TrainSupp.3D detectionCollision ↓Support F1 ↑
IoU↑mAP↑mesharchobj floorceil
R3DS-realnone16.415.00.2261.562--
heur16.214.10.2191.3293.269.9
anno16.615.90.3491.40438.594.4
R3DS-synnone14.08.40.3901.664--
heur14.67.80.2811.3014.682.9
anno14.38.20.3491.21932.095.0
R3DS-mixnone17.615.80.1711.007--
heur19.217.70.1511.2673.683.7
anno18.618.20.1581.30812.096.2
", + "bbox": [ + 295, + 224, + 697, + 415 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/b9cf6b6f1cd5d5aa7a57758e4d93fbff55efdf05ccd935dd32489d36e8987aa1.jpg", + "table_caption": [ + "Table 6: Comparison of the average number of bounding box and mesh-based object collisions per scene in IG and R3DS. R3DS exhibits more bounding box-based collisions, but almost none of these are actual physical collisions between object meshes. Measuring collisions between bounding boxes is a poor collision measure for fully-populated, real-world scenes." + ], + "table_footnote": [], + "table_body": "
DatasetsMeshCollisionsBox Collisions
obj-objobj-wallobj-floorobj-ceil
IG-1.1850.0750.0000.790
R3DS0.00062.8230.3880.0350.064
", + "bbox": [ + 299, + 521, + 694, + 602 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "R3DS, it is also possible to compute heuristic support relations from object bounding boxes. Specifically, an object is supported by another if their bounding boxes intersect within tolerance distance of $0.1\\mathrm{m}$ and the centroid of the former object is higher than that of the latter. Support by wall/floor/ceiling is calculated in the same way without the height judgment. This definition is similar to how DPC defines object attachment. Figure 7 compares these two ways of computing support relations, showing that heuristic computation can mistakenly designate support relations to two nearby objects. Table 5 shows that incorporating support relation prediction indeed influences the performance of DPC. Heuristic support information may worsen 3D object detection (mAP in R3DS-real and R3DS-syn), but it eliminates mesh collisions the most. Learning support relations from R3DS annotations leads to a 2.4 improvement on mAP in R3DS-mix, although the classification F1 score is low.", + "bbox": [ + 212, + 643, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "R3DS", + "bbox": [ + 689, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/1e9e3c30bf3392d725d9e9e422459d60d5d67c79d2c0e9bbc9a66dd67f641b0f.jpg", + "table_caption": [ + "Table 7: Ablation of relation optimization (RO) on R3DS-real. The 2nd and 3rd row remove optimization terms in RO. The last row replaces bounding-box collisions with mesh collisions." + ], + "table_footnote": [], + "table_body": "
Rel. Opt.3D mAP ↑ MeshCollisions ↓Box Collisions ↓
obj-objobj-arch
DPC18.20.1580.0621.308
w/o obj col18.91.3421.1301.301
w/o obj col+tch19.61.2191.0621.295
w/ mesh col19.71.0270.8561.394
", + "bbox": [ + 284, + 198, + 707, + 306 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4) Is relation optimization (RO) effective on R3DS? Test-time relation optimization (RO) was introduced by Zhang et al. [28] to reduce physical violations, floating objects, and misalignment between objects and architecture. The original cost function based on bounding box collisions succeeds in optimizing object poses, since there are few such collisions in the IG data originally used for evaluation (see in Tab. 6). However, the same data assumption does not hold for R3DS, which has more bounding-box-based collisions but nearly zero mesh-based collisions. We ablate the design of RO on R3DS-real in Table 7. By removing two optimization terms (bounding-box-based object-wise collision and touching step-by-step), the model outperforms the original one in 3D mAP (+1.4) but degrades in mesh-based and box-based collisions. We show that using mesh-based collision optimization leads to the best performance. The increase in collisions is unsurprising as the R3DS data reflects more cluttered real interiors. Limitations. Our dataset construction relied on 3D architectures for each Matterport3D scan which are simplifications of the geometry of the real environment. One issue is imperfect wall positions, resulting in objects attached to these virtual walls being offset from the true surface. In addition, objects in our 3D scenes were placed without regard to the materials, meaning that the detailed surface appearance does not match that of the observed object. Future work can investigate transfer of surface appearance to the synthetic objects by projecting textures from the RGB-D data and 3D reconstructed meshes.", + "bbox": [ + 212, + 334, + 787, + 654 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 681, + 359, + 698 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We introduced the R3DS dataset. R3DS provides more complete, densely populated, and richly annotated synthetic 3D scene proxies of real-world environments with linked panoramic images. We showed the usefulness of R3DS on the Panoramic Scene Understanding task. Our experiments demonstrate the value of realistic synthetic recreations in this task, in particular through the use of object support information. While we focused on the PanoSun task, R3DS can also be useful for other tasks such as single-view shape retrieval, single-view object pose estimation, and panoramic scene graph prediction.", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Q. Wu et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was funded in part by a CIFAR AI Chair, a Canada Research Chair, NSERC Discovery Grant, NSF award #2016532, and enabled by support from WestGrid and Compute Canada. Daniel Ritchie is an advisor to Geope and owns equity in the company. Geope is a start-up that is developing 3D technology to build immersive virtual copies of the real world with applications in various fields, including games and architecture. We thank Madhawa Vidanapathirana, Weijie Lin, and David Han for help with development of the annotation tool, and Denys Iliash, Mrinal Goshalia, Brandon Robles, Paul Brown, Chloe Ye, Coco Kaleel, Elizabeth Wu and Hannah Julius for data annotation, and Ivan Tam, Austin Wang, and Ning Wang for feedback on the paper draft.", + "bbox": [ + 212, + 176, + 787, + 329 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 352, + 321, + 368 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Avetisyan, A., Dahnert, M., Dai, A., Savva, M., Chang, A.X., Nießner, M.: Scan2CAD: Learning CAD model alignment in RGB-D scans. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019)", + "2. Batra, D., Chang, A.X., Chernova, S., Davison, A.J., Deng, J., Koltun, V., Levine, S., Malik, J., Mordatch, I., Mottaghi, R., Savva, M., Su, H.: Rearrangement: A challenge for embodied AI. arXiv preprint arXiv:2011.01975 (2020)", + "3. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niebner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3D: Learning from RGB-D data in indoor environments. In: Proceedings of the International Conference on 3D Vision (3DV). pp. 667-676. IEEE (2017)", + "4. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015)", + "5. Collins, J., Goel, S., Deng, K., Luthra, A., Xu, L., Gundogdu, E., Zhang, X., Vicente, T.F.Y., Dideriksen, T., Arora, H., et al.: Abo: Dataset and benchmarks for real-world 3d object understanding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21126-21136 (2022)", + "6. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5828-5839 (2017)", + "7. Dong, Y., Fang, C., Dong, Z., Bo, L., Tan, P.: PanoContext-Former: Panoramic total scene understanding with a transformer. arXiv preprint arXiv:2305.12497 (2023)", + "8. Fu, H., Cai, B., Gao, L., Zhang, L., Li, C., Xun, Z., Sun, C., Fei, Y., Zheng, Y., Li, Y., et al.: 3D-FRONT: 3D Furnished Rooms with layOuts and semaNTics. arXiv preprint arXiv:2011.09127 (2020)" + ], + "bbox": [ + 225, + 385, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "R3DS", + "bbox": [ + 689, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "9. Fu, H., Jia, R., Gao, L., Gong, M., Zhao, B., Maybank, S., Tao, D.: 3D-Future: 3D Furniture shape with TextURE. arXiv preprint arXiv:2009.09633 (2020)", + "10. Hua, B.S., Pham, Q.H., Nguyen, D.T., Tran, M.K., Yu, L.F., Yeung, S.K.: SceneNN: A scene meshes dataset with annotations. In: Proceedings of the International Conference on 3D Vision (3DV). pp. 92-101. IEEE (2016)", + "11. Karras, T.: Maximizing parallelism in the construction of bvhs, octrees, and k-d trees. In: Proceedings of the Fourth ACM SIGGRAPH / Eurographics Conference on High-Performance Graphics. pp. 33-37. Eurographics Association (2012). https://doi.org/10.2312/EGGH/HPG12/033-037, https://doi.org/10.2312/EGGH/HPG12/033-037", + "12. Li, Z., Yu, T.W., Sang, S., Wang, S., Song, M., Liu, Y., Yeh, Y.Y., Zhu, R., Gundavarapu, N., Shi, J., Bi, S., Xu, Z., Yu, H.X., Sunkavalli, K., Hasan, M., Ramamoorthi, R., Chandraker, M.: OpenRooms: An end-to-end open framework for photorealistic indoor scene datasets. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "13. Lim, J.J., Pirsiavash, H., Torralba, A.: Parsing intelia objects: Fine pose estimation. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 2992-2999 (2013)", + "14. Maninis, K.K., Popov, S., Nießner, M., Ferrari, V.: CAD-estate: Large-scale CAD model annotation in RGB videos. arXiv preprint arXiv:2306.09011 (2023)", + "15. Paschalidou, D., Kar, A., Shugrina, M., Kreis, K., Geiger, A., Fidler, S.: Atiss: Autoregressive transformers for indoor scene synthesis. Advances in Neural Information Processing Systems 34, 12013-12026 (2021)", + "16. Ramakrishnan, S.K., Gokaslan, A., Wijmans, E., Maksymets, O., Clegg, A., Turner, J., Undersander, E., Galuba, W., Westbury, A., Chang, A.X., et al.: Habitat-Matterport 3D dataset (hm3d): 1000 large-scale 3D environments for embodied AI. arXiv preprint arXiv:2109.08238 (2021)", + "17. Sadalgi, S.: Wayfair's 3D Model API. https://www/aboutwayfair.com/tech-innovation/wayfairs-3d-model-api (2016), [Online; accessed 15-Nov-2023]", + "18. Shen, B., Xia, F., Li, C., Martin-Martin, R., Fan, L., Wang, G., Buch, S., D'Arpino, C., Srivastava, S., Tchapmi, L.P., Vainio, K., Fei-Fei, L., Savarese, S.: iGibson, a simulation environment for interactive tasks in large realistic scenes. In: Proceedings of the International Conference on Intelligent Robots and Systems (IROS) (2021)", + "19. Shen, B., Xia, F., Li, C., Martin-Martin, R., Fan, L., Wang, G., Pérez-D'Arpino, C., Buch, S., Srivastava, S., Tchapmi, L., et al.: iGibson 1.0: A simulation environment for interactive tasks in large realistic scenes. In: Proceedings of the International Conference on Intelligent Robots and Systems (IROS). pp. 7520-7527. IEEE (2021)", + "20. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019)" + ], + "bbox": [ + 215, + 145, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Q. Wu et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "21. Sun, C., Hsiao, C.W., Sun, M., Chen, H.T.: Horizonnet: Learning room layout with 1d representation and pano stretch data augmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1047-1056 (2019)", + "22. Sun, X., Wu, J., Zhang, X., Zhang, Z., Zhang, C., Xue, T., Tenenbaum, J.B., Freeman, W.T.: Pix3D: Dataset and methods for single-image 3D shape modeling. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2974-2983 (2018)", + "23. Szot, A., Clegg, A., Undersander, E., Wijmans, E., Zhao, Y., Turner, J., Maestre, N., Mukadam, M., Chaplot, D.S., Maksymets, O., et al.: Habitat 2.0: Training home assistants to rearrange their habitat. Advances in Neural Information Processing Systems 34, 251-266 (2021)", + "24. Tzionas, D., Ballan, L., Srikantha, A., Aponte, P., Pollefeys, M., Gall, J.: Capturing hands in action using discriminative salient points and physics simulation. International Journal of Computer Vision (IJCV) 118(2), 172-193 (Jun 2016), https://doi.org/10.1007/s11263-016-0895-4", + "25. Wang, K., Lin, Y.A., Weissmann, B., Savva, M., Chang, A.X., Ritchie, D.: Planit: Planning and instantiating indoor scenes with relation graph and spatial prior networks. ACM Transactions on Graphics (TOG) 38(4), 1-15 (2019)", + "26. Xiang, Y., Kim, W., Chen, W., Ji, J., Choy, C., Su, H., Mottaghi, R., Guibas, L., Savarese, S.: Objectnet3D: A large scale database for 3D object recognition. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 160-176. Springer (2016)", + "27. Yadav, K., Ramrakhya, R., Ramakrishnan, S.K., Gervet, T., Turner, J., Gokaslan, A., Maestre, N., Chang, A.X., Batra, D., Savva, M., et al.: Habitat-Matterport 3D semantics dataset. arXiv preprint arXiv:2210.05633 (2022)", + "28. Zhang, C., Cui, Z., Chen, C., Liu, S., Zeng, B., Bao, H., Zhang, Y.: Deep-PanoContext: Panoramic 3D scene understanding with holistic scene context graph and relation-based optimization. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 12632-12641 (2021)", + "29. Zhang, Y., Song, S., Tan, P., Xiao, J.: PanoContext: A whole-room 3D context model for panoramic scene understanding. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 668-686. Springer (2014)", + "30. Zheng, J., Zhang, J., Li, J., Tang, R., Gao, S., Zhou, Z.: Structured3D: A large photo-realistic dataset for structured 3D modeling. arXiv preprint arXiv:1908.00222 (2019)", + "31. Zhou, T., Tucker, R., Flynn, J., Fyffe, G., Snavely, N.: Stereo magnification: learning view synthesis using multiplane images. ACM Transactions on Graphics (TOG) 37(4), 1-12 (2018)" + ], + "bbox": [ + 212, + 145, + 787, + 766 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "R3DS", + "bbox": [ + 689, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/R3DS_ Reality-linked 3D Scenes for Panoramic Scene Understanding/7708992f-16b8-4792-b8c3-10a97f869f8d_model.json b/2024/R3DS_ Reality-linked 3D Scenes for Panoramic Scene Understanding/7708992f-16b8-4792-b8c3-10a97f869f8d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ed740280f8060ee6fa0c46cc853c5358a9848dbb --- /dev/null +++ b/2024/R3DS_ Reality-linked 3D Scenes for Panoramic Scene Understanding/7708992f-16b8-4792-b8c3-10a97f869f8d_model.json @@ -0,0 +1,2126 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.295, + 0.142, + 0.71, + 0.187 + ], + "angle": 0, + "content": "R3DS: Reality-linked 3D Scenes for Panoramic Scene Understanding" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.212, + 0.761, + 0.245 + ], + "angle": 0, + "content": "Qirui Wu\\(^{1}\\), Sonia Raychaudhuri\\(^{1}\\), Daniel Ritchie\\(^{2}\\), Manolis Savva\\(^{1}\\), and Angel X. Chang\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.413, + 0.255, + 0.59, + 0.269 + ], + "angle": 0, + "content": "\\(^{1}\\) Simon Fraser University" + }, + { + "type": "text", + "bbox": [ + 0.435, + 0.269, + 0.567, + 0.284 + ], + "angle": 0, + "content": "2 Brown University" + }, + { + "type": "text", + "bbox": [ + 0.343, + 0.283, + 0.66, + 0.298 + ], + "angle": 0, + "content": "3 Alberta Machine Intelligence Institute (Amii)" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.325, + 0.784, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.509, + 0.788, + 0.634 + ], + "angle": 0, + "content": "Fig. 1: Left: the Reality-linked 3D Scenes dataset (R3DS) fills a gap between synthetic 3D scenes and reconstructions of real-world environments by providing 3D scene proxies linked to real-world panoramas from Matterport3D (three example panoramas and 3D scenes shown). Right: our dataset contains scenes with higher density and completeness compared to prior datasets, and provides additional annotations such as object support (what objects or architectural elements support other objects), and matching object sets (e.g., pairs of the same nightstand). We use our dataset for the panoramic scene understanding task and demonstrate its value for research on room layout estimation, as well as 2D and 3D object detection." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.66, + 0.744, + 0.841 + ], + "angle": 0, + "content": "Abstract. We introduce the Reality-linked 3D Scenes (R3DS) dataset of synthetic 3D scenes mirroring the real-world scene arrangements from Matterport3D panoramas. Compared to prior work, R3DS has more complete and densely populated scenes with objects linked to real-world observations in panoramas. R3DS also provides an object support hierarchy, and matching object sets (e.g., same chairs around a dining table) for each scene. Overall, R3DS contains 19K objects represented by 3,784 distinct CAD models from over 100 object categories. We demonstrate the effectiveness of R3DS on the Panoramic Scene Understanding task. We find that: 1) training on R3DS enables better generalization; 2) support relation prediction trained with R3DS improves performance compared to heuristically calculated support; and 3) R3DS offers a challenging benchmark for future work on panoramic scene understanding." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.129 + ], + "angle": 0, + "content": "Q. Wu et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.377, + 0.161 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.184, + 0.788, + 0.365 + ], + "angle": 0, + "content": "Datasets of 3D indoor environments are increasingly used for research on scene understanding [1, 3, 28], embodied AI [2, 19, 23], and scene generation [15, 25]. There are two strategies for constructing 3D scene datasets: reconstruction of real-world spaces [3], or authoring scenes using synthetic 3D objects [8] (CAD models). Reconstruction captures real spaces but is hard to scale, and the resulting scenes exhibit imperfections and artifacts. On the other hand, synthetic 3D scenes are complete and easy to manipulate but often do not match the statistics of real-world spaces and are artificially \"clean\". Moreover, both strategies are time-consuming and require expertise. There have been some attempts to create \"synthetic\" replicas of real environments by matching CAD models to objects in scans [19, 23]. These efforts have been limited in scale and often result in partial and sparsely populated synthetic counterparts of the real environments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.366, + 0.788, + 0.473 + ], + "angle": 0, + "content": "We design a framework that allows users to create 3D scenes from RGB panoramas and use it to create R3DS: a dataset of 'Reality-linked' 3D Scenes. Each 3D scene in our dataset is a complete proxy of an environment from the Matterport3D [3] dataset, representing both the 3D architecture and the objects. Thus, each scene is linked to a real space, with correspondences established between panorama observations of each object and the synthetic object. These reality-linked scenes reflect denser real-world arrangements of objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.788, + 0.595 + ], + "angle": 0, + "content": "The use of panoramas for reference is advantageous compared to either perspective images or 3D reconstructions. Panoramas are not limited by the field of view unlike perspective images, enabling more complete 3D synthetic proxies. Panoramas also better capture relatively small objects and objects with challenging materials or illumination conditions compared to reconstructions. Additionally, there is a scarcity of synthetic 3D scenes coupled with real-world panoramas, with only one relatively small algorithmically constructed dataset provided by Zhang et al. [28] being available to the community." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.788, + 0.687 + ], + "angle": 0, + "content": "Compared to prior efforts such as Scan2CAD [1] and CAD-Estate [14], our dataset provides more complete scenes, with salient observed objects being captured in the layout. Moreover, we provide a support hierarchy defining what objects are placed on other objects and specify sets of identical objects such as dining chairs around a table, allowing for creating realistic variations of the scene by swapping the entire set to a different chair design." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.788, + 0.765 + ], + "angle": 0, + "content": "We demonstrate the value of our dataset by using it for the Panoramic Scene Understanding task. We show that leveraging the denser layouts and support hierarchy information in our scenes leads to improved object detection performance and better generalization compared to training using other datasets previously used for this task. In summary, we make the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.779, + 0.784, + 0.805 + ], + "angle": 0, + "content": "- We design a framework for efficient construction of synthetic scenes from real panoramas and use it to create R3DS: a dataset of reality-linked 3D scenes." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.81, + 0.784, + 0.838 + ], + "angle": 0, + "content": "- R3DS provides more complete and realistic scenes with correspondences between real and synthetic objects, and object-object support relations." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.779, + 0.784, + 0.838 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.69, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "R3DS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.149, + 0.751, + 0.161 + ], + "angle": 0, + "content": "R3DS (Ours): linked to panoramicas, densely populated scenes, matched object sets, objects supported" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.162, + 0.784, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.261, + 0.413, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.347, + 0.383, + 0.358 + ], + "angle": 0, + "content": "CAD-Estate: sparse, floating" + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.261, + 0.617, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.427, + 0.348, + 0.614, + 0.358 + ], + "angle": 0, + "content": "Scan2CAD: sparse, floating, mismatched" + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.262, + 0.784, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.347, + 0.753, + 0.358 + ], + "angle": 0, + "content": "iGibson DPC: floating" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.373, + 0.788, + 0.457 + ], + "angle": 0, + "content": "Fig. 2: Dataset comparison. (Top) shows different views of a scene annotated in R3DS. Comparison with previous datasets (bottom) shows (1) R3DS has more complete scenes than the previous datasets; (2) Objects in R3DS are properly supported by either architecture or other objects unlike the others (e.g. floating objects with no proper support); (3) R3DS is annotated using the same 3D model for objects arranged together (chairs by the dining table, couches arranged together)." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.488, + 0.788, + 0.548 + ], + "angle": 0, + "content": "- We show that the more complete layouts and support relations in our dataset enable better performance and generalization in the Panoramic Scene Understanding task, and that our dataset offers a challenging benchmark for future work in scene understanding." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.582, + 0.388, + 0.598 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.779 + ], + "angle": 0, + "content": "3D scene datasets. A spectrum of scene datasets have been used for scene understanding tasks. One type provides annotated 3D reconstructions of real scenes based on RGB-D videos [3, 6, 10, 16, 20, 27]. These datasets are usually subject to the limitations of RGB-D reconstruction, typically containing noise, artifacts such as holes, and poor reconstructions of thin structures, shiny objects, or light sources. Another type of 3D datasets is authored by manually designing 3D object assets [5, 9] and inserting them into synthetic 3D scenes [8]. However, such datasets lack the realism of real-world reconstructions and demand expert knowledge, making them expensive to create. A third, hybrid approach which is closest to our work creates 3D scene datasets by aligning existing object CAD models to real world data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Datasets that align CAD models to real world. There have been a number of recent efforts in aligning CAD models with real-world data. Prior work [13, 22, 26] has annotated object images with 3D models, typically using keypoint correspondences to perspective images. These perspective images usually do not" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.129 + ], + "angle": 0, + "content": "Q. Wu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "depict a complete scene; they typically focus on one or two objects and are limited in field of view, resulting in a sparse proxy of the real scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.787, + 0.357 + ], + "angle": 0, + "content": "Another line of work aligns 3D CAD models to RGB-D scans either through annotation as in Scan2CAD [1], or automated heuristics as in iGibson [19]. OpenRooms [12] extends Scan2CAD [1] with photorealistic material annotations and focuses on inverse rendering tasks. Conceptually, these allow for more complete synthetic scene proxies. However, statistics from these datasets show that they are still relatively sparse (see Tab. 1). In addition, the poor quality of reconstruction makes aligning CAD models challenging without referring to the original RGB images. A prominent exception is Replica [20] which has fairly high-quality reconstructions and the artist-created Replica-CAD [23]. However, creating such high quality \"replicas\" is labor intensive and costly. Szot et al. [23] report \\(900+\\) work hours required to model approximately 90 objects, resulting in a dataset of limited scale with 105 different layouts of what is effectively a single room." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.787, + 0.418 + ], + "angle": 0, + "content": "More recently, Maninis et al. [14] introduced CAD-Estate, which aligns CAD models to RGB videos for over 19K spaces. Because the data is based on monocular video, the coverage of the spaces is incomplete. In addition, the annotation is relatively sparse, with an average of only 6 objects per scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.788, + 0.631 + ], + "angle": 0, + "content": "Datasets for panoramic scene understanding. There have been relatively few datasets introduced for Panoramic Scene Understanding [7, 28, 29]. In the initial PanoContext dataset [29], the data did not have aligned CAD models and only included object cuboids. The ground truth data was collected on 2D panorama images by annotating visible cuboid vertices; 3D cuboids were obtained by minimizing the re-projection error from the annotated 2D vertices. Moreover, these 3D cuboids and the room layout are obtained with the assumption that the room layout is a cuboid and that the objects are vertically aligned. Thus, the resulting object layout may deviate from the real arrangement of objects. More recently, datasets for Panoramic Scene Understanding have been built by taking 3D scans, aligning CAD objects to them, and then generating panoramas [7, 28]. Compared to these datasets, our R3DS is manually curated for a larger number of distinct regions and provides support hierarchy and matching object set annotations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.65, + 0.436, + 0.667 + ], + "angle": 0, + "content": "3 The R3DS Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.678, + 0.787, + 0.769 + ], + "angle": 0, + "content": "We describe the construction of the R3DS dataset and present a statistical analysis of the scenes it contains. Compared to previous datasets [1, 14, 28] (Fig. 2), our scenes are more densely populated, and objects are annotated with a hierarchy of support relations. Moreover, our dataset specifies matching object instances in furniture arrangements. Figure 1 shows example annotations from our dataset." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.436, + 0.803 + ], + "angle": 0, + "content": "3.1 Dataset construction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We developed a 3D annotation interface (Fig. 3) showing a panorama of a room from Matterport3D and allowing users to insert 3D CAD objects into a 3D scene" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.69, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "R3DS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Table 1: Comparison with 3D indoor scene datasets aligned with real-world images, videos, or scans. Our R3DS dataset contains more densely populated annotations compared to other datasets, with objects from 110 different categories. We report the unique models (#CAD), object categories (#Cat), object instances (#Objects) as well as average number of objects and object categories per annotation. For Scan2Cad [1] and ReplicaCAD [23] the average is per scan. Note that ReplicaCAD consists of 105 different layouts (arrangements) of effectively one room. CAD-Estate has partial views into 19K spaces, many of which are 1-2 rooms. Of the datasets used for panoramic scene understanding, our R3DS dataset covers more rooms with 842 panoramicas over 22 room types. Replica-Pano (in gray) was not released, so we report statistics from the paper. Our annotations per panorama are more complete and our dataset has both support relations (Sup) and matching object instance sets (Match)." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.324, + 0.789, + 0.41 + ], + "angle": 0, + "content": "
DatasetSourceCAD AlignmentTypeHouses/RoomsPanos#CAD#Objects#CatAveObjAveCatSupMatch
Scan2CAD [1]ScanNet [6]Annotatorscan- / 1506✘3,04914,225359.44.1✘✘
OpenRooms [12]ScanNet [6]Scan2CAD [1]scan- / 1288✘2,65116,0143812.46.3✘✘
ReplicaCAD [23]Replica [20]Artist recreationscan- / 105*✘922,2934421.814.4✘✘
CAD-Estate [14]RealEstate10K [31]Annotatorvideo19,512✘12,024100,882496.33.4✘✘
Replica-Pano [7]Replica [20]Heuristicpano- / 272700--25-✘✘
iGibson-DPC [28]iGibson [19]Heuristicpano15 / 100150050026,9985717.910.2✘✘
R3DS (Ours)Matterport3D [3]Annotatorpano20 / 3708423,78419,05011022.910.4✓✓
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.788, + 0.563 + ], + "angle": 0, + "content": "which is visually overlaid on the panorama. The 3D scene is initially empty, consisting only of 3D architectural geometry which specifies the walls, floor, ceiling as well as the placement of openings (e.g. doors, windows, and other openings) on the walls. We create this 3D architecture by taking 20 houses from Matterport3D, constructing an initial architecture based on the region and object annotations for the windows and doors, and manually refining the placement of walls and openings. By combining panoramas and 3D architectures, users can see through openings and annotate objects located in other rooms." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.788, + 0.687 + ], + "angle": 0, + "content": "We ask annotators to select and place 3D object models to best match the panoramic image. We use CAD models from Wayfair [17] and ShapeNet [4] models collected from 3D Trimble warehouse. Wayfair provides a large collection of furniture CAD models that match real-world products and are sized based on real-world dimensions. However, it does not include bathroom fittings, electronic equipment and kitchen appliances, for which we manually scale and align CAD models from ShapeNet. Compared with ShapeNetCore, the CAD models we use are already sized to real-world sizes (instead of normalized to a unit cube)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To assist the annotators, we provide segmented masks of objects visible in the panorama. Since Matterport3D has annotated 3D object masks on the scans we use those annotations, but it is also possible to run an instance segmentation on the panorama. When the user clicks one of these masks, a search panel automatically opens and shows objects matching the clicked mask category label. For each mask, the annotator selects a matching object and positions and aligns it to match the mask. Annotators are instructed to choose objects which match the shape of the corresponding object in the panorama (rather than its color or texture). To help annotators focus on shape, we render all 3D objects in a neutral gray color. Annotators are also explicitly asked to select the same" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Q. Wu et al." + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.149, + 0.714, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.443, + 0.788, + 0.486 + ], + "angle": 0, + "content": "Fig. 3: R3DS annotation pipeline. Annotators see an empty scene (architecture only). They then insert and manipulate 3D object models from a panorama viewpoint to create a populated 3D scene proxy corresponding to the panorama." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.516, + 0.788, + 0.669 + ], + "angle": 0, + "content": "3D asset for objects that should be the same; our interface provides a list of recently selected assets to make this process easier. In addition, annotators are instructed to add annotations for any objects that are not segmented (due to errors in Matterport3D) through simple clicks. These additional objects provide a more complete annotation that covers poorly reconstructed objects such as glass tables, lamps, and other small objects. The interface enforces that each object is placed on a support surface (either an architecture element or another object). The annotator can review their work by toggling off the panorama overlay or by switching to a perspective view of the 3D scene. For more annotated scene examples and details on the annotation process please refer to the supplement." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.692, + 0.513, + 0.708 + ], + "angle": 0, + "content": "3.2 Dataset analysis and statistics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We collect annotations for 20 Matterport3D houses with 808 panoramas in total. We discard panoramas taken on stairs or outside a house, since they have a limited number of objects that can be placed. After filtering we have 769 panoramas for our analysis and experiments. For 73 panoramas, we collect two sets of annotations for each, to obtain a total of 842 annotated object arrangements across 22 different Matterport3D region types. The panoramas with two annotations serve as a test of annotator consistency and add diversity. In total, R3DS contains 19,050 object instances from 3,784 unique 3D CAD models spanning" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.69, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "R3DS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image_caption", + "bbox": [ + 0.303, + 0.15, + 0.361, + 0.16 + ], + "angle": 0, + "content": "Scan2CAD" + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.169, + 0.417, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.274, + 0.242, + 0.39, + 0.253 + ], + "angle": 0, + "content": "Floor warped, no arch" + }, + { + "type": "image_caption", + "bbox": [ + 0.481, + 0.15, + 0.543, + 0.16 + ], + "angle": 0, + "content": "CAD-Estate" + }, + { + "type": "image", + "bbox": [ + 0.439, + 0.167, + 0.607, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.489, + 0.243, + 0.55, + 0.253 + ], + "angle": 0, + "content": "Partial arch" + }, + { + "type": "image_caption", + "bbox": [ + 0.649, + 0.15, + 0.712, + 0.161 + ], + "angle": 0, + "content": "R3DS (ours)" + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.162, + 0.735, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.243, + 0.749, + 0.254 + ], + "angle": 0, + "content": "Complete arch with portals" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.272, + 0.785, + 0.314 + ], + "angle": 0, + "content": "Fig. 4: Architecture comparison. Compared to Scan2CAD (no architecture) and CAD-Estate (partial architecture), R3DS provides complete architecture with door/window portals." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.35, + 0.785, + 0.395 + ], + "angle": 0, + "content": "over 110 fine-grained object categories. Table 1 shows a comparison of overall statistics with previous 3D indoor scene datasets. See the supplement for more statistics about annotated objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.787, + 0.625 + ], + "angle": 0, + "content": "Compared to prior datasets that align CAD models to real-world scenes, R3DS is more complete, providing annotated object support hierarchies and matching object instances. CAD-Estate [14] annotates RGB videos with 3D objects and architectural room layouts. However, the architecture is partial as the videos have limited view (Fig. 4), and not all objects in the scenes are annotated (Fig. 2). This results in annotations of objects floating mid-air and not properly supported (e.g. lamp Fig. 2). Scan2CAD [1] also lacks support structure (e.g., lamp not supported by cabinet in Fig. 2). In addition, because Scan2CAD does not provide clean 3D architecture on which objects are placed (Fig. 4), objects on the floor are not always placed such that their bottom face is parallel with a horizontal plane. In contrast, our R3DS scenes have an accurate support hierarchy by construction. OpenRooms [12] augments Scan2CAD with room layouts representing the architecture. However, the architecture in R3DS is more complex and realistic, especially due to inclusion of more doors (1.92 doors per room in R3DS vs 0.67 in OpenRooms)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.785, + 0.717 + ], + "angle": 0, + "content": "Evaluation of CAD object annotation quality is non-trivial as the 'ground truth' from the semantically annotated 3D reconstructions is itself imperfect. We measured how closely our annotated CAD objects conform to the real objects using the average 2D IoU between CAD object mask and ground truth 2D mask. R3DS is at \\(42.6\\%\\) vs \\(38.5\\%\\) for Scan2CAD, across 8 common object categories (bed, sofa, chair, cabinet, tv/monitor, table, shelving, bathtub)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Of the datasets previously used for Panoramic Scene Understanding, ReplicaPano [7] has not been released, and iGibson-DPC [28] is the only dataset with synthetic panoramic images annotated with 3D objects and room layout. iGibson-DPC is built on scenes from iGibson [18] by randomly replacing objects with different models from the same category and rendering using the iGibson simulator to render panoramas. The selection and placement of objects in iGibson-DPC is based on heuristic algorithms, while our R3DS is manually annotated and placed 3D models are verified in terms of match and alignment to the object masks." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Q. Wu et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.178 + ], + "angle": 0, + "content": "Moreover, iGibson-DPC contains unrealistic object arrangements (e.g., floating TV in Fig. 2)." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.201, + 0.375, + 0.218 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.233, + 0.794, + 0.339 + ], + "angle": 0, + "content": "We showcase the value of R3DS on the Panoramic Scene Understanding (PanoSun) task [7, 28, 29]. Given an input RGB panorama, the goal is to estimate the room layout, detect objects in 2D, estimate their 3D oriented bounding boxes and also reconstruct 3D object meshes. Our experiments show that methods trained on R3DS data benefit from its realism and generalize better when evaluated on photorealistic images. We also investigate the role of object support hierarchy information in improving performance." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.362, + 0.353, + 0.377 + ], + "angle": 0, + "content": "4.1 Task setup" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.387, + 0.787, + 0.493 + ], + "angle": 0, + "content": "Method. DeepPanoContext (DPC) [28] predicts the room layout, detects objects in 3D and recovers object meshes from a panorama image using a relation-based graph convolutional network and a differentiable relation optimization procedure. Since DPC has a publicly-available implementation, we use it to benchmark the R3DS data on the PanoSun task. We keep all hyperparameters unchanged except lowering the relation optimization loss weight of 3D bounding box back-projection from 10 to 1, since the ground truth 2D masks are noisy." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.493, + 0.787, + 0.599 + ], + "angle": 0, + "content": "Datasets. We train and evaluate DPC on the iGibson-DPC (IG) [7, 18, 28], Structured3D (S3D) [30], and R3DS datasets. Zhang et al. [28] render 1,500 panoramas from 15 iGibson houses composed of \\(500+\\) objects spanning 57 object categories. We use the same data and splits for IG. Structured3D consists of 3500 houses and around 18K photo-realistic rendered panoramas in total. We use 14K for training and the remaining 4K for testing. Note that Structured3D does not provide ground truth object meshes." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.599, + 0.787, + 0.78 + ], + "angle": 0, + "content": "To prepare R3DS for this task, we generate the ground truth room layout from the 3D architecture based on the camera viewpoint and obtain 3D oriented bounding boxes (OBBs) from all objects. We use 2D object masks from the Matterport3D mesh instance segmentation. We consider three variants of R3DS based on the input panorama: \\(R3DS\\)-real where we use the Matterport3D panoramas, \\(R3DS\\)-syn where we use rendered panoramas (at the same camera poses) from the annotated synthetic scenes, and \\(R3DS\\)-mix where we combine the two types of panoramas and double the available data. We follow the MP3D house split and merge the train and val sets to obtain a disjoint split of 15 train and 5 test houses. Based on the split, we have 696 annotated panoramas for train and 146 for test. To fairly evaluate methods trained on different datasets, we curate a list of 25 object classes common to all datasets." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Metrics. Following Zhang et al. [28] we use separate metrics for room layout estimation, 3D object detection, and scene relation prediction. For room layout estimation, we use 2D IoU for predicted 2D floorplan, 3D IoU for lifted 3D room geometry, and dRMSE for predicted depth with respect to the camera location." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.69, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "R3DS" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.788, + 0.186 + ], + "angle": 0, + "content": "Table 2: Room layout estimation on R3DS-real test set. DPC [28] was pretrained on IG and S3D. For the last three rows, we fine-tune the pretrained weights on variants of R3DS." + }, + { + "type": "table", + "bbox": [ + 0.36, + 0.199, + 0.635, + 0.286 + ], + "angle": 0, + "content": "
Train2D IoU ↑3D IoU ↑dRMSE ↓
DPC [28]53.450.30.682
R3DS-real55.153.10.610
R3DS-syn59.056.10.629
R3DS-mix59.657.00.572
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.316, + 0.788, + 0.358 + ], + "angle": 0, + "content": "Table 3: Cross-dataset evaluation for the Panoramic Scene Understanding task. We evaluate 3D detections with class-agnostic IoU and mAP at IoU of 0.15, and report object collisions. The highlighted rows indicate the most challenging scenario." + }, + { + "type": "table", + "bbox": [ + 0.286, + 0.372, + 0.709, + 0.643 + ], + "angle": 0, + "content": "
TestTrain3D detection ↑Collision ↓Attachment F1 ↑
IoUmAPmesharchobjwallfloor
IG27.530.31.6622.59453.176.895.086.2
IG+R3DS24.030.21.4042.25459.764.194.62.7
R3DS-real17.313.40.2421.45638.864.092.80.0
R3DS-syn23.214.20.4801.93848.546.793.828.6
R3DS-mix21.615.60.4341.24843.167.290.19.8
S3D19.53.51.0162.65150.968.790.811.6
IG+R3DS19.77.00.8682.08952.067.491.21.8
R3DS-real18.47.10.6002.59845.061.689.70.7
R3DS-syn19.04.80.6442.56149.349.791.22.4
R3DS-mix19.67.50.4631.67347.864.187.20.9
R3DS15.65.90.5751.95953.850.751.20.0
IG+R3DS17.514.10.2811.26749.561.658.60.0
R3DS-real16.415.00.2261.56244.057.358.90.0
R3DS-syn14.08.40.3901.66454.140.649.10.0
R3DS-mix17.615.80.1711.00748.558.360.10.0
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "For 3D object detection, we report bounding box-based class-agnostic 3D IoU as well as mean average precision (mAP) across the 25 object classes, where an IoU greater than 0.15 counts as a \"true\" result. For scene relation prediction, we report F1 scores for relation classification. We also report the average number of objects colliding with each other or with architectural structures (wall, floor, ceiling). Specifically, we follow Zhang et al. [28] and measure collisions using the Separating Axis Theorem (SAT) to test whether object bounding boxes overlap. Since bounding box-based collision is a poor proxy for real-world physical collision, we also compute mesh-based collision by checking if the meshes for object pairs have any interpenetrating triangles [11, 24]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Q. Wu et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.148, + 0.788, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.5 + ], + "angle": 0, + "content": "Fig. 5: Qualitative results for cross-dataset Panoramic Scene Understanding task. Correct and incorrect object detections shown in green and red boxes. Ground truth room layout and meshes shown in gray color, while layout prediction is in yellow. Training on R3DS leads to fewer errors compared to training on other datasets, especially when mixed with real data." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.533, + 0.325, + 0.547 + ], + "angle": 0, + "content": "4.2 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.788, + 0.643 + ], + "angle": 0, + "content": "1) Does R3DS help DPC generalize to real images? Since the original DPC work only trained and evaluated on synthetic data, it is unclear how well it performs on realistic panoramic imagery. We hypothesize that training on R3DS will lead to better performance. We separately show results on room layout estimation (Tab. 2) and 3D object detection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Room layout estimation. For room layout estimation, DPC uses Horizon-Net [21] pretrained on iGibson (IG) and Structured3D (S3D) panoramas. This model achieves good performance on IG data (91.0 3D IoU). When directly testing the official pretrained model on \\( R3DS \\)-real panoramas, we notice a significant performance drop compared to results on the rendered panoramas from iGibson (Tab. 2 shows that the 3D IoU drops to 50.3). By finetuning the pretrained model with \\( R3DS \\)-real, we can predict more precise room layouts for real cluttered scenes. Even only trained on \\( R3DS \\)-syn, we outperform the original DPC model by \\( 5.6\\% \\) and \\( 5.8\\% \\) on 2D and 3D IoU, respectively. This is likely due to renderings from \\( R3DS \\)-syn reflecting more realistic object arrangements in a room instead of pushing all objects against walls. Best performance on 2D IoU (59.6), 3D IoU (57.0) and depth RMSE (0.572) is achieved by fine tuning on \\( R3DS \\)-mix." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.69, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "R3DS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 4: Performance of models trained on three synthetic datasets (IG, S3D, and IG+S3D) evaluated on the R3DS-full dataset, where \"full\" indicates all 840 panoramicas are used for testing." + }, + { + "type": "table", + "bbox": [ + 0.33, + 0.199, + 0.665, + 0.293 + ], + "angle": 0, + "content": "
Train3D detectionCollision ↓Support F1 ↑
IoU↑mAP↑mesharchobj floorceil
IG14.25.20.7031.6394.185.3
S3D16.410.00.1121.2263.186.8
IG+S3D17.19.70.1331.1623.384.8
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.323, + 0.789, + 0.563 + ], + "angle": 0, + "content": "Object detection. For 3D object detection, we train DPC on different data settings and conduct a cross-dataset evaluation (see Tab. 3). To investigate how models perform on out-of-distribution scenes, we evaluate models on Structured3D, as its images are near-realistic. To explore whether DPC training benefits from R3DS given the same amount of data, we create a special data input \\(IG + R3DS\\) that combines iGibson and R3DS panoramas by randomly replacing half (500) of iGibson data with \\(R3DS\\)-real data. The results show that \\(IG + R3DS\\) performs almost the same as \\(IG\\) with fewer collisions on iGibson, but it remarkably outperforms \\(IG\\) on the test set of \\(R3DS\\) and \\(S3D\\) by 8.2 and 3.5 improvements on 3D mAP, respectively. It also averages 0.221 fewer mesh collisions. There are noticeable performance gaps on iGibson for models trained on R3DS data likely due to the data domain shift. Among the three variants of R3DS data, \\(R3DS\\)-mix outperforms the others on all three test sets regarding 3D IoU and 3D mAP with the fewest mesh and architecture collisions. Although \\(R3DS\\)-syn underperforms on \\(R3DS\\) and \\(S3D\\) test sets, it achieves better performance than \\(IG\\) with even less data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.789, + 0.702 + ], + "angle": 0, + "content": "Scene relation classification. We report F1 scores for identifying attachment relationships of objects to other objects and architecture elements (see Tab. 3). We note that models trained with synthetic renderings perform better than those trained on real images. That is because synthetic renderings present cleaner and simpler scenes with fewer objects than real world and simpler illumination such that DPC finds it easier to learn object-object and object-architecture relations. Also, note that the predictions of object-ceiling attachments can be extremely low because few objects are attached to the ceiling in the ground truth data. We show qualitative examples in Figure 5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.84 + ], + "angle": 0, + "content": "2) Is R3DS a challenging, high-quality test set? How would a model trained on pure synthetic data perform on complex real data (R3DS)? Due to its modest scale, we propose using R3DS as a challenging, high-quality test set rather than a train set. Specifically, we evaluate the synthetic-to-real performance of DPC by training on iGibson and/or Structured3D and testing on all panoramas in R3DS-real. Table 4 shows that a model trained with Structured3D performs the best (10.0 3D mAP and 0.112 mesh collision) as it observes the most photo-realistic images. DPC benefits from the synthetic data for higher bounding box IoUs, since it possesses accurately aligned 3D bounding box and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Q. Wu et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.358, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.207, + 0.35, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.243, + 0.257, + 0.34, + 0.264 + ], + "angle": 0, + "content": "(a) Object appearance gap \\((60\\%)\\)" + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.145, + 0.499, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.206, + 0.499, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.387, + 0.257, + 0.482, + 0.264 + ], + "angle": 0, + "content": "(b) Common occlusions (76.7%)" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.145, + 0.64, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.206, + 0.641, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.257, + 0.624, + 0.264 + ], + "angle": 0, + "content": "(c) Wrong 3D prediction \\((65\\%)\\)" + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.145, + 0.782, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.206, + 0.78, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.257, + 0.764, + 0.264 + ], + "angle": 0, + "content": "(d) Objects in mirrors (15.6%)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.279, + 0.787, + 0.35 + ], + "angle": 0, + "content": "Fig. 6: Error analysis on R3DS panoramas using DPC pretrained on S3D. Typical prediction errors are categorized into 4 groups: (a) Appearance gap between real and synthetic images. (b) Detection failure due to common occlusions. (c) Bad performance of monocular 3D prediction. (d) Perception error for objects in mirrors. Numbers in parentheses mean the error rates." + }, + { + "type": "image_caption", + "bbox": [ + 0.287, + 0.369, + 0.334, + 0.378 + ], + "angle": 0, + "content": "Panorama" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.38, + 0.404, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.455, + 0.404, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.456, + 0.37, + 0.536, + 0.379 + ], + "angle": 0, + "content": "Heuristic Support" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.38, + 0.596, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.455, + 0.596, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.658, + 0.37, + 0.722, + 0.379 + ], + "angle": 0, + "content": "R3DS Support" + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.38, + 0.785, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.455, + 0.785, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.537, + 0.788, + 0.581 + ], + "angle": 0, + "content": "Fig. 7: Comparison between heuristic and R3DS support relations. Dark lines between objects show support relations. Heuristic support often mistakenly assigns relations to spatially close objects. In contrast, R3DS support relations are annotated." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.785, + 0.657 + ], + "angle": 0, + "content": "more unoccluded objects. However, mAP performance is lower due to worse object recognition ability. All models struggle to predict correct object-wise support relations but do a better job of predicting object-floor support relations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.787, + 0.764 + ], + "angle": 0, + "content": "We conduct error analysis on 120 randomly sampled panoramas using the model pretrained on \\( S3D \\) to identify typical errors (see Fig. 6). Errors are categorized into 4 groups: (a) \\( 60\\% \\) panoramas have 2D perception errors due to the synthetic-to-real appearance gap; (b) \\( 76.7\\% \\) panoramas show detection failures due to occlusions; (c) \\( 65\\% \\) panoramas exhibit correct 2D detections but fail to correctly perform 3D predictions; and (d) \\( 15.6\\% \\) out of 45 panoramas with mirrors mistakenly predict virtual objects in mirrors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.841 + ], + "angle": 0, + "content": "3) Are R3DS support relations helpful for PanoSun? We investigate whether the support relationships between objects provided in our R3DS scene hierarchy help boost performance of holistic scene understanding. We augment DPC's Relation Scene-GCN module with additional support relation prediction branches. Besides obtaining explicitly annotated scene support relations from" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.69, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "R3DS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 5: Performance on R3DS-real of DPC models trained on variants of R3DS with different support relation settings. We compare the original model without support (none) against models supervised with support that is heuristically computed (heur) or annotated from R3DS (anno). Classification results are evaluated on annotated ground-truth scene hierarchy." + }, + { + "type": "table", + "bbox": [ + 0.297, + 0.226, + 0.699, + 0.416 + ], + "angle": 0, + "content": "
TrainSupp.3D detectionCollision ↓Support F1 ↑
IoU↑mAP↑mesharchobj floorceil
R3DS-realnone16.415.00.2261.562--
heur16.214.10.2191.3293.269.9
anno16.615.90.3491.40438.594.4
R3DS-synnone14.08.40.3901.664--
heur14.67.80.2811.3014.682.9
anno14.38.20.3491.21932.095.0
R3DS-mixnone17.615.80.1711.007--
heur19.217.70.1511.2673.683.7
anno18.618.20.1581.30812.096.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.441, + 0.788, + 0.512 + ], + "angle": 0, + "content": "Table 6: Comparison of the average number of bounding box and mesh-based object collisions per scene in IG and R3DS. R3DS exhibits more bounding box-based collisions, but almost none of these are actual physical collisions between object meshes. Measuring collisions between bounding boxes is a poor collision measure for fully-populated, real-world scenes." + }, + { + "type": "table", + "bbox": [ + 0.3, + 0.522, + 0.695, + 0.603 + ], + "angle": 0, + "content": "
DatasetsMeshCollisionsBox Collisions
obj-objobj-wallobj-floorobj-ceil
IG-1.1850.0750.0000.790
R3DS0.00062.8230.3880.0350.064
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.84 + ], + "angle": 0, + "content": "R3DS, it is also possible to compute heuristic support relations from object bounding boxes. Specifically, an object is supported by another if their bounding boxes intersect within tolerance distance of \\(0.1\\mathrm{m}\\) and the centroid of the former object is higher than that of the latter. Support by wall/floor/ceiling is calculated in the same way without the height judgment. This definition is similar to how DPC defines object attachment. Figure 7 compares these two ways of computing support relations, showing that heuristic computation can mistakenly designate support relations to two nearby objects. Table 5 shows that incorporating support relation prediction indeed influences the performance of DPC. Heuristic support information may worsen 3D object detection (mAP in R3DS-real and R3DS-syn), but it eliminates mesh collisions the most. Learning support relations from R3DS annotations leads to a 2.4 improvement on mAP in R3DS-mix, although the classification F1 score is low." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Q. Wu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.144, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 7: Ablation of relation optimization (RO) on R3DS-real. The 2nd and 3rd row remove optimization terms in RO. The last row replaces bounding-box collisions with mesh collisions." + }, + { + "type": "table", + "bbox": [ + 0.285, + 0.199, + 0.709, + 0.307 + ], + "angle": 0, + "content": "
Rel. Opt.3D mAP ↑ MeshCollisions ↓Box Collisions ↓
obj-objobj-arch
DPC18.20.1580.0621.308
w/o obj col18.91.3421.1301.301
w/o obj col+tch19.61.2191.0621.295
w/ mesh col19.71.0270.8561.394
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.335, + 0.789, + 0.655 + ], + "angle": 0, + "content": "4) Is relation optimization (RO) effective on R3DS? Test-time relation optimization (RO) was introduced by Zhang et al. [28] to reduce physical violations, floating objects, and misalignment between objects and architecture. The original cost function based on bounding box collisions succeeds in optimizing object poses, since there are few such collisions in the IG data originally used for evaluation (see in Tab. 6). However, the same data assumption does not hold for R3DS, which has more bounding-box-based collisions but nearly zero mesh-based collisions. We ablate the design of RO on R3DS-real in Table 7. By removing two optimization terms (bounding-box-based object-wise collision and touching step-by-step), the model outperforms the original one in 3D mAP (+1.4) but degrades in mesh-based and box-based collisions. We show that using mesh-based collision optimization leads to the best performance. The increase in collisions is unsurprising as the R3DS data reflects more cluttered real interiors. Limitations. Our dataset construction relied on 3D architectures for each Matterport3D scan which are simplifications of the geometry of the real environment. One issue is imperfect wall positions, resulting in objects attached to these virtual walls being offset from the true surface. In addition, objects in our 3D scenes were placed without regard to the materials, meaning that the detailed surface appearance does not match that of the observed object. Future work can investigate transfer of surface appearance to the synthetic objects by projecting textures from the RGB-D data and 3D reconstructed meshes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.682, + 0.36, + 0.699 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We introduced the R3DS dataset. R3DS provides more complete, densely populated, and richly annotated synthetic 3D scene proxies of real-world environments with linked panoramic images. We showed the usefulness of R3DS on the Panoramic Scene Understanding task. Our experiments demonstrate the value of realistic synthetic recreations in this task, in particular through the use of object support information. While we focused on the PanoSun task, R3DS can also be useful for other tasks such as single-view shape retrieval, single-view object pose estimation, and panoramic scene graph prediction." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.69, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "R3DS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.33 + ], + "angle": 0, + "content": "This work was funded in part by a CIFAR AI Chair, a Canada Research Chair, NSERC Discovery Grant, NSF award #2016532, and enabled by support from WestGrid and Compute Canada. Daniel Ritchie is an advisor to Geope and owns equity in the company. Geope is a start-up that is developing 3D technology to build immersive virtual copies of the real world with applications in various fields, including games and architecture. We thank Madhawa Vidanapathirana, Weijie Lin, and David Han for help with development of the annotation tool, and Denys Iliash, Mrinal Goshalia, Brandon Robles, Paul Brown, Chloe Ye, Coco Kaleel, Elizabeth Wu and Hannah Julius for data annotation, and Ivan Tam, Austin Wang, and Ning Wang for feedback on the paper draft." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.353, + 0.323, + 0.369 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.386, + 0.785, + 0.446 + ], + "angle": 0, + "content": "1. Avetisyan, A., Dahnert, M., Dai, A., Savva, M., Chang, A.X., Nießner, M.: Scan2CAD: Learning CAD model alignment in RGB-D scans. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.447, + 0.785, + 0.506 + ], + "angle": 0, + "content": "2. Batra, D., Chang, A.X., Chernova, S., Davison, A.J., Deng, J., Koltun, V., Levine, S., Malik, J., Mordatch, I., Mottaghi, R., Savva, M., Su, H.: Rearrangement: A challenge for embodied AI. arXiv preprint arXiv:2011.01975 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.508, + 0.785, + 0.568 + ], + "angle": 0, + "content": "3. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niebner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3D: Learning from RGB-D data in indoor environments. In: Proceedings of the International Conference on 3D Vision (3DV). pp. 667-676. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.569, + 0.785, + 0.613 + ], + "angle": 0, + "content": "4. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.614, + 0.785, + 0.689 + ], + "angle": 0, + "content": "5. Collins, J., Goel, S., Deng, K., Luthra, A., Xu, L., Gundogdu, E., Zhang, X., Vicente, T.F.Y., Dideriksen, T., Arora, H., et al.: Abo: Dataset and benchmarks for real-world 3d object understanding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21126-21136 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.749 + ], + "angle": 0, + "content": "6. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5828-5839 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.75, + 0.785, + 0.794 + ], + "angle": 0, + "content": "7. Dong, Y., Fang, C., Dong, Z., Bo, L., Tan, P.: PanoContext-Former: Panoramic total scene understanding with a transformer. arXiv preprint arXiv:2305.12497 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "8. Fu, H., Cai, B., Gao, L., Zhang, L., Li, C., Xun, Z., Sun, C., Fei, Y., Zheng, Y., Li, Y., et al.: 3D-FRONT: 3D Furnished Rooms with layOuts and semaNTics. arXiv preprint arXiv:2011.09127 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.386, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Q. Wu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.146, + 0.787, + 0.192 + ], + "angle": 0, + "content": "9. Fu, H., Jia, R., Gao, L., Gong, M., Zhao, B., Maybank, S., Tao, D.: 3D-Future: 3D Furniture shape with TextURE. arXiv preprint arXiv:2009.09633 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.192, + 0.787, + 0.237 + ], + "angle": 0, + "content": "10. Hua, B.S., Pham, Q.H., Nguyen, D.T., Tran, M.K., Yu, L.F., Yeung, S.K.: SceneNN: A scene meshes dataset with annotations. In: Proceedings of the International Conference on 3D Vision (3DV). pp. 92-101. IEEE (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.239, + 0.787, + 0.312 + ], + "angle": 0, + "content": "11. Karras, T.: Maximizing parallelism in the construction of bvhs, octrees, and k-d trees. In: Proceedings of the Fourth ACM SIGGRAPH / Eurographics Conference on High-Performance Graphics. pp. 33-37. Eurographics Association (2012). https://doi.org/10.2312/EGGH/HPG12/033-037, https://doi.org/10.2312/EGGH/HPG12/033-037" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.313, + 0.787, + 0.403 + ], + "angle": 0, + "content": "12. Li, Z., Yu, T.W., Sang, S., Wang, S., Song, M., Liu, Y., Yeh, Y.Y., Zhu, R., Gundavarapu, N., Shi, J., Bi, S., Xu, Z., Yu, H.X., Sunkavalli, K., Hasan, M., Ramamoorthi, R., Chandraker, M.: OpenRooms: An end-to-end open framework for photorealistic indoor scene datasets. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.403, + 0.787, + 0.449 + ], + "angle": 0, + "content": "13. Lim, J.J., Pirsiavash, H., Torralba, A.: Parsing intelia objects: Fine pose estimation. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 2992-2999 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.449, + 0.787, + 0.495 + ], + "angle": 0, + "content": "14. Maninis, K.K., Popov, S., Nießner, M., Ferrari, V.: CAD-estate: Large-scale CAD model annotation in RGB videos. arXiv preprint arXiv:2306.09011 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.539 + ], + "angle": 0, + "content": "15. Paschalidou, D., Kar, A., Shugrina, M., Kreis, K., Geiger, A., Fidler, S.: Atiss: Autoregressive transformers for indoor scene synthesis. Advances in Neural Information Processing Systems 34, 12013-12026 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.539, + 0.787, + 0.6 + ], + "angle": 0, + "content": "16. Ramakrishnan, S.K., Gokaslan, A., Wijmans, E., Maksymets, O., Clegg, A., Turner, J., Undersander, E., Galuba, W., Westbury, A., Chang, A.X., et al.: Habitat-Matterport 3D dataset (hm3d): 1000 large-scale 3D environments for embodied AI. arXiv preprint arXiv:2109.08238 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.6, + 0.787, + 0.645 + ], + "angle": 0, + "content": "17. Sadalgi, S.: Wayfair's 3D Model API. https://www/aboutwayfair.com/tech-innovation/wayfairs-3d-model-api (2016), [Online; accessed 15-Nov-2023]" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.645, + 0.787, + 0.72 + ], + "angle": 0, + "content": "18. Shen, B., Xia, F., Li, C., Martin-Martin, R., Fan, L., Wang, G., Buch, S., D'Arpino, C., Srivastava, S., Tchapmi, L.P., Vainio, K., Fei-Fei, L., Savarese, S.: iGibson, a simulation environment for interactive tasks in large realistic scenes. In: Proceedings of the International Conference on Intelligent Robots and Systems (IROS) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.72, + 0.787, + 0.796 + ], + "angle": 0, + "content": "19. Shen, B., Xia, F., Li, C., Martin-Martin, R., Fan, L., Wang, G., Pérez-D'Arpino, C., Buch, S., Srivastava, S., Tchapmi, L., et al.: iGibson 1.0: A simulation environment for interactive tasks in large realistic scenes. In: Proceedings of the International Conference on Intelligent Robots and Systems (IROS). pp. 7520-7527. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.796, + 0.787, + 0.842 + ], + "angle": 0, + "content": "20. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.146, + 0.787, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.69, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "R3DS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.208 + ], + "angle": 0, + "content": "21. Sun, C., Hsiao, C.W., Sun, M., Chen, H.T.: Horizonnet: Learning room layout with 1d representation and pano stretch data augmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1047-1056 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.209, + 0.788, + 0.268 + ], + "angle": 0, + "content": "22. Sun, X., Wu, J., Zhang, X., Zhang, Z., Zhang, C., Xue, T., Tenenbaum, J.B., Freeman, W.T.: Pix3D: Dataset and methods for single-image 3D shape modeling. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2974-2983 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.269, + 0.788, + 0.328 + ], + "angle": 0, + "content": "23. Szot, A., Clegg, A., Undersander, E., Wijmans, E., Zhao, Y., Turner, J., Maestre, N., Mukadam, M., Chaplot, D.S., Maksymets, O., et al.: Habitat 2.0: Training home assistants to rearrange their habitat. Advances in Neural Information Processing Systems 34, 251-266 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.389 + ], + "angle": 0, + "content": "24. Tzionas, D., Ballan, L., Srikantha, A., Aponte, P., Pollefeys, M., Gall, J.: Capturing hands in action using discriminative salient points and physics simulation. International Journal of Computer Vision (IJCV) 118(2), 172-193 (Jun 2016), https://doi.org/10.1007/s11263-016-0895-4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.39, + 0.788, + 0.449 + ], + "angle": 0, + "content": "25. Wang, K., Lin, Y.A., Weissmann, B., Savva, M., Chang, A.X., Ritchie, D.: Planit: Planning and instantiating indoor scenes with relation graph and spatial prior networks. ACM Transactions on Graphics (TOG) 38(4), 1-15 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.509 + ], + "angle": 0, + "content": "26. Xiang, Y., Kim, W., Chen, W., Ji, J., Choy, C., Su, H., Mottaghi, R., Guibas, L., Savarese, S.: Objectnet3D: A large scale database for 3D object recognition. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 160-176. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.51, + 0.788, + 0.569 + ], + "angle": 0, + "content": "27. Yadav, K., Ramrakhya, R., Ramakrishnan, S.K., Gervet, T., Turner, J., Gokaslan, A., Maestre, N., Chang, A.X., Batra, D., Savva, M., et al.: Habitat-Matterport 3D semantics dataset. arXiv preprint arXiv:2210.05633 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.57, + 0.788, + 0.631 + ], + "angle": 0, + "content": "28. Zhang, C., Cui, Z., Chen, C., Liu, S., Zeng, B., Bao, H., Zhang, Y.: Deep-PanoContext: Panoramic 3D scene understanding with holistic scene context graph and relation-based optimization. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 12632-12641 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.631, + 0.788, + 0.676 + ], + "angle": 0, + "content": "29. Zhang, Y., Song, S., Tan, P., Xiao, J.: PanoContext: A whole-room 3D context model for panoramic scene understanding. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 668-686. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.676, + 0.788, + 0.72 + ], + "angle": 0, + "content": "30. Zheng, J., Zhang, J., Li, J., Tang, R., Gao, S., Zhou, Z.: Structured3D: A large photo-realistic dataset for structured 3D modeling. arXiv preprint arXiv:1908.00222 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.721, + 0.788, + 0.767 + ], + "angle": 0, + "content": "31. Zhou, T., Tucker, R., Flynn, J., Fyffe, G., Snavely, N.: Stereo magnification: learning view synthesis using multiplane images. ACM Transactions on Graphics (TOG) 37(4), 1-12 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.767 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/R3DS_ Reality-linked 3D Scenes for Panoramic Scene Understanding/7708992f-16b8-4792-b8c3-10a97f869f8d_origin.pdf b/2024/R3DS_ Reality-linked 3D Scenes for Panoramic Scene Understanding/7708992f-16b8-4792-b8c3-10a97f869f8d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..55a4f487b3615c2828d50615937f206c25283812 --- /dev/null +++ b/2024/R3DS_ Reality-linked 3D Scenes for Panoramic Scene Understanding/7708992f-16b8-4792-b8c3-10a97f869f8d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d7843adaf985b9fffcc66bd050fc9a241c8235e0a1e3633e60137e43ee915ac +size 8465701 diff --git a/2024/R3DS_ Reality-linked 3D Scenes for Panoramic Scene Understanding/full.md b/2024/R3DS_ Reality-linked 3D Scenes for Panoramic Scene Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c7800e1e5c807f09d5795093d80ffc639af69d9e --- /dev/null +++ b/2024/R3DS_ Reality-linked 3D Scenes for Panoramic Scene Understanding/full.md @@ -0,0 +1,255 @@ +# R3DS: Reality-linked 3D Scenes for Panoramic Scene Understanding + +Qirui Wu $^{1}$ , Sonia Raychaudhuri $^{1}$ , Daniel Ritchie $^{2}$ , Manolis Savva $^{1}$ , and Angel X. Chang $^{1,3}$ + +$^{1}$ Simon Fraser University + +2 Brown University + +3 Alberta Machine Intelligence Institute (Amii) + +![](images/ff7acf139001c077257d8e97088718c8a412a9fa3ad7cf529f28962eaddf32c9.jpg) +Fig. 1: Left: the Reality-linked 3D Scenes dataset (R3DS) fills a gap between synthetic 3D scenes and reconstructions of real-world environments by providing 3D scene proxies linked to real-world panoramas from Matterport3D (three example panoramas and 3D scenes shown). Right: our dataset contains scenes with higher density and completeness compared to prior datasets, and provides additional annotations such as object support (what objects or architectural elements support other objects), and matching object sets (e.g., pairs of the same nightstand). We use our dataset for the panoramic scene understanding task and demonstrate its value for research on room layout estimation, as well as 2D and 3D object detection. + +Abstract. We introduce the Reality-linked 3D Scenes (R3DS) dataset of synthetic 3D scenes mirroring the real-world scene arrangements from Matterport3D panoramas. Compared to prior work, R3DS has more complete and densely populated scenes with objects linked to real-world observations in panoramas. R3DS also provides an object support hierarchy, and matching object sets (e.g., same chairs around a dining table) for each scene. Overall, R3DS contains 19K objects represented by 3,784 distinct CAD models from over 100 object categories. We demonstrate the effectiveness of R3DS on the Panoramic Scene Understanding task. We find that: 1) training on R3DS enables better generalization; 2) support relation prediction trained with R3DS improves performance compared to heuristically calculated support; and 3) R3DS offers a challenging benchmark for future work on panoramic scene understanding. + +# 1 Introduction + +Datasets of 3D indoor environments are increasingly used for research on scene understanding [1, 3, 28], embodied AI [2, 19, 23], and scene generation [15, 25]. There are two strategies for constructing 3D scene datasets: reconstruction of real-world spaces [3], or authoring scenes using synthetic 3D objects [8] (CAD models). Reconstruction captures real spaces but is hard to scale, and the resulting scenes exhibit imperfections and artifacts. On the other hand, synthetic 3D scenes are complete and easy to manipulate but often do not match the statistics of real-world spaces and are artificially "clean". Moreover, both strategies are time-consuming and require expertise. There have been some attempts to create "synthetic" replicas of real environments by matching CAD models to objects in scans [19, 23]. These efforts have been limited in scale and often result in partial and sparsely populated synthetic counterparts of the real environments. + +We design a framework that allows users to create 3D scenes from RGB panoramas and use it to create R3DS: a dataset of 'Reality-linked' 3D Scenes. Each 3D scene in our dataset is a complete proxy of an environment from the Matterport3D [3] dataset, representing both the 3D architecture and the objects. Thus, each scene is linked to a real space, with correspondences established between panorama observations of each object and the synthetic object. These reality-linked scenes reflect denser real-world arrangements of objects. + +The use of panoramas for reference is advantageous compared to either perspective images or 3D reconstructions. Panoramas are not limited by the field of view unlike perspective images, enabling more complete 3D synthetic proxies. Panoramas also better capture relatively small objects and objects with challenging materials or illumination conditions compared to reconstructions. Additionally, there is a scarcity of synthetic 3D scenes coupled with real-world panoramas, with only one relatively small algorithmically constructed dataset provided by Zhang et al. [28] being available to the community. + +Compared to prior efforts such as Scan2CAD [1] and CAD-Estate [14], our dataset provides more complete scenes, with salient observed objects being captured in the layout. Moreover, we provide a support hierarchy defining what objects are placed on other objects and specify sets of identical objects such as dining chairs around a table, allowing for creating realistic variations of the scene by swapping the entire set to a different chair design. + +We demonstrate the value of our dataset by using it for the Panoramic Scene Understanding task. We show that leveraging the denser layouts and support hierarchy information in our scenes leads to improved object detection performance and better generalization compared to training using other datasets previously used for this task. In summary, we make the following contributions: + +- We design a framework for efficient construction of synthetic scenes from real panoramas and use it to create R3DS: a dataset of reality-linked 3D scenes. +- R3DS provides more complete and realistic scenes with correspondences between real and synthetic objects, and object-object support relations. + +R3DS (Ours): linked to panoramicas, densely populated scenes, matched object sets, objects supported + +![](images/13caafa1cafea1e0f7f812fec3dcbfb2055dd8bb5c1972f496effc9bb7d6bc79.jpg) + +![](images/8f160698894748f4a793518d06989ec6f99b5f51bfe62065b613eecca100ea01.jpg) +CAD-Estate: sparse, floating +Fig. 2: Dataset comparison. (Top) shows different views of a scene annotated in R3DS. Comparison with previous datasets (bottom) shows (1) R3DS has more complete scenes than the previous datasets; (2) Objects in R3DS are properly supported by either architecture or other objects unlike the others (e.g. floating objects with no proper support); (3) R3DS is annotated using the same 3D model for objects arranged together (chairs by the dining table, couches arranged together). + +![](images/b3a9e4b37345c3acd518271a6033b3db202843053028d532bdb00b9aaec8d748.jpg) +Scan2CAD: sparse, floating, mismatched + +![](images/bea44ff58d17b393c906c9f9715fda51b081eb836a52841492dcbfd6359c31a5.jpg) +iGibson DPC: floating + +- We show that the more complete layouts and support relations in our dataset enable better performance and generalization in the Panoramic Scene Understanding task, and that our dataset offers a challenging benchmark for future work in scene understanding. + +# 2 Related Work + +3D scene datasets. A spectrum of scene datasets have been used for scene understanding tasks. One type provides annotated 3D reconstructions of real scenes based on RGB-D videos [3, 6, 10, 16, 20, 27]. These datasets are usually subject to the limitations of RGB-D reconstruction, typically containing noise, artifacts such as holes, and poor reconstructions of thin structures, shiny objects, or light sources. Another type of 3D datasets is authored by manually designing 3D object assets [5, 9] and inserting them into synthetic 3D scenes [8]. However, such datasets lack the realism of real-world reconstructions and demand expert knowledge, making them expensive to create. A third, hybrid approach which is closest to our work creates 3D scene datasets by aligning existing object CAD models to real world data. + +Datasets that align CAD models to real world. There have been a number of recent efforts in aligning CAD models with real-world data. Prior work [13, 22, 26] has annotated object images with 3D models, typically using keypoint correspondences to perspective images. These perspective images usually do not + +depict a complete scene; they typically focus on one or two objects and are limited in field of view, resulting in a sparse proxy of the real scene. + +Another line of work aligns 3D CAD models to RGB-D scans either through annotation as in Scan2CAD [1], or automated heuristics as in iGibson [19]. OpenRooms [12] extends Scan2CAD [1] with photorealistic material annotations and focuses on inverse rendering tasks. Conceptually, these allow for more complete synthetic scene proxies. However, statistics from these datasets show that they are still relatively sparse (see Tab. 1). In addition, the poor quality of reconstruction makes aligning CAD models challenging without referring to the original RGB images. A prominent exception is Replica [20] which has fairly high-quality reconstructions and the artist-created Replica-CAD [23]. However, creating such high quality "replicas" is labor intensive and costly. Szot et al. [23] report $900+$ work hours required to model approximately 90 objects, resulting in a dataset of limited scale with 105 different layouts of what is effectively a single room. + +More recently, Maninis et al. [14] introduced CAD-Estate, which aligns CAD models to RGB videos for over 19K spaces. Because the data is based on monocular video, the coverage of the spaces is incomplete. In addition, the annotation is relatively sparse, with an average of only 6 objects per scene. + +Datasets for panoramic scene understanding. There have been relatively few datasets introduced for Panoramic Scene Understanding [7, 28, 29]. In the initial PanoContext dataset [29], the data did not have aligned CAD models and only included object cuboids. The ground truth data was collected on 2D panorama images by annotating visible cuboid vertices; 3D cuboids were obtained by minimizing the re-projection error from the annotated 2D vertices. Moreover, these 3D cuboids and the room layout are obtained with the assumption that the room layout is a cuboid and that the objects are vertically aligned. Thus, the resulting object layout may deviate from the real arrangement of objects. More recently, datasets for Panoramic Scene Understanding have been built by taking 3D scans, aligning CAD objects to them, and then generating panoramas [7, 28]. Compared to these datasets, our R3DS is manually curated for a larger number of distinct regions and provides support hierarchy and matching object set annotations. + +# 3 The R3DS Dataset + +We describe the construction of the R3DS dataset and present a statistical analysis of the scenes it contains. Compared to previous datasets [1, 14, 28] (Fig. 2), our scenes are more densely populated, and objects are annotated with a hierarchy of support relations. Moreover, our dataset specifies matching object instances in furniture arrangements. Figure 1 shows example annotations from our dataset. + +# 3.1 Dataset construction + +We developed a 3D annotation interface (Fig. 3) showing a panorama of a room from Matterport3D and allowing users to insert 3D CAD objects into a 3D scene + +Table 1: Comparison with 3D indoor scene datasets aligned with real-world images, videos, or scans. Our R3DS dataset contains more densely populated annotations compared to other datasets, with objects from 110 different categories. We report the unique models (#CAD), object categories (#Cat), object instances (#Objects) as well as average number of objects and object categories per annotation. For Scan2Cad [1] and ReplicaCAD [23] the average is per scan. Note that ReplicaCAD consists of 105 different layouts (arrangements) of effectively one room. CAD-Estate has partial views into 19K spaces, many of which are 1-2 rooms. Of the datasets used for panoramic scene understanding, our R3DS dataset covers more rooms with 842 panoramicas over 22 room types. Replica-Pano (in gray) was not released, so we report statistics from the paper. Our annotations per panorama are more complete and our dataset has both support relations (Sup) and matching object instance sets (Match). + +
DatasetSourceCAD AlignmentTypeHouses/RoomsPanos#CAD#Objects#CatAveObjAveCatSupMatch
Scan2CAD [1]ScanNet [6]Annotatorscan- / 1506✘3,04914,225359.44.1✘✘
OpenRooms [12]ScanNet [6]Scan2CAD [1]scan- / 1288✘2,65116,0143812.46.3✘✘
ReplicaCAD [23]Replica [20]Artist recreationscan- / 105*✘922,2934421.814.4✘✘
CAD-Estate [14]RealEstate10K [31]Annotatorvideo19,512✘12,024100,882496.33.4✘✘
Replica-Pano [7]Replica [20]Heuristicpano- / 272700--25-✘✘
iGibson-DPC [28]iGibson [19]Heuristicpano15 / 100150050026,9985717.910.2✘✘
R3DS (Ours)Matterport3D [3]Annotatorpano20 / 3708423,78419,05011022.910.4✓✓
+ +which is visually overlaid on the panorama. The 3D scene is initially empty, consisting only of 3D architectural geometry which specifies the walls, floor, ceiling as well as the placement of openings (e.g. doors, windows, and other openings) on the walls. We create this 3D architecture by taking 20 houses from Matterport3D, constructing an initial architecture based on the region and object annotations for the windows and doors, and manually refining the placement of walls and openings. By combining panoramas and 3D architectures, users can see through openings and annotate objects located in other rooms. + +We ask annotators to select and place 3D object models to best match the panoramic image. We use CAD models from Wayfair [17] and ShapeNet [4] models collected from 3D Trimble warehouse. Wayfair provides a large collection of furniture CAD models that match real-world products and are sized based on real-world dimensions. However, it does not include bathroom fittings, electronic equipment and kitchen appliances, for which we manually scale and align CAD models from ShapeNet. Compared with ShapeNetCore, the CAD models we use are already sized to real-world sizes (instead of normalized to a unit cube). + +To assist the annotators, we provide segmented masks of objects visible in the panorama. Since Matterport3D has annotated 3D object masks on the scans we use those annotations, but it is also possible to run an instance segmentation on the panorama. When the user clicks one of these masks, a search panel automatically opens and shows objects matching the clicked mask category label. For each mask, the annotator selects a matching object and positions and aligns it to match the mask. Annotators are instructed to choose objects which match the shape of the corresponding object in the panorama (rather than its color or texture). To help annotators focus on shape, we render all 3D objects in a neutral gray color. Annotators are also explicitly asked to select the same + +![](images/7d9a40641378deb27cc912483930a1134192010421f790ad506aed5c2e0cba91.jpg) +Fig. 3: R3DS annotation pipeline. Annotators see an empty scene (architecture only). They then insert and manipulate 3D object models from a panorama viewpoint to create a populated 3D scene proxy corresponding to the panorama. + +3D asset for objects that should be the same; our interface provides a list of recently selected assets to make this process easier. In addition, annotators are instructed to add annotations for any objects that are not segmented (due to errors in Matterport3D) through simple clicks. These additional objects provide a more complete annotation that covers poorly reconstructed objects such as glass tables, lamps, and other small objects. The interface enforces that each object is placed on a support surface (either an architecture element or another object). The annotator can review their work by toggling off the panorama overlay or by switching to a perspective view of the 3D scene. For more annotated scene examples and details on the annotation process please refer to the supplement. + +# 3.2 Dataset analysis and statistics + +We collect annotations for 20 Matterport3D houses with 808 panoramas in total. We discard panoramas taken on stairs or outside a house, since they have a limited number of objects that can be placed. After filtering we have 769 panoramas for our analysis and experiments. For 73 panoramas, we collect two sets of annotations for each, to obtain a total of 842 annotated object arrangements across 22 different Matterport3D region types. The panoramas with two annotations serve as a test of annotator consistency and add diversity. In total, R3DS contains 19,050 object instances from 3,784 unique 3D CAD models spanning + +![](images/fb06118c490c651e0c9ab388dc1f3fc796a2703aa2a609cf399a9452a64ea020.jpg) +Scan2CAD +Floor warped, no arch +Fig. 4: Architecture comparison. Compared to Scan2CAD (no architecture) and CAD-Estate (partial architecture), R3DS provides complete architecture with door/window portals. + +![](images/e4674f4724b0f8db038006ff740d63989cb02799e85284258465603dd3387ae9.jpg) +CAD-Estate +Partial arch + +![](images/65c45845923779202792ad615589436d7ef23755a7f9ae77441cc5b0ace51604.jpg) +R3DS (ours) +Complete arch with portals + +over 110 fine-grained object categories. Table 1 shows a comparison of overall statistics with previous 3D indoor scene datasets. See the supplement for more statistics about annotated objects. + +Compared to prior datasets that align CAD models to real-world scenes, R3DS is more complete, providing annotated object support hierarchies and matching object instances. CAD-Estate [14] annotates RGB videos with 3D objects and architectural room layouts. However, the architecture is partial as the videos have limited view (Fig. 4), and not all objects in the scenes are annotated (Fig. 2). This results in annotations of objects floating mid-air and not properly supported (e.g. lamp Fig. 2). Scan2CAD [1] also lacks support structure (e.g., lamp not supported by cabinet in Fig. 2). In addition, because Scan2CAD does not provide clean 3D architecture on which objects are placed (Fig. 4), objects on the floor are not always placed such that their bottom face is parallel with a horizontal plane. In contrast, our R3DS scenes have an accurate support hierarchy by construction. OpenRooms [12] augments Scan2CAD with room layouts representing the architecture. However, the architecture in R3DS is more complex and realistic, especially due to inclusion of more doors (1.92 doors per room in R3DS vs 0.67 in OpenRooms). + +Evaluation of CAD object annotation quality is non-trivial as the 'ground truth' from the semantically annotated 3D reconstructions is itself imperfect. We measured how closely our annotated CAD objects conform to the real objects using the average 2D IoU between CAD object mask and ground truth 2D mask. R3DS is at $42.6\%$ vs $38.5\%$ for Scan2CAD, across 8 common object categories (bed, sofa, chair, cabinet, tv/monitor, table, shelving, bathtub). + +Of the datasets previously used for Panoramic Scene Understanding, ReplicaPano [7] has not been released, and iGibson-DPC [28] is the only dataset with synthetic panoramic images annotated with 3D objects and room layout. iGibson-DPC is built on scenes from iGibson [18] by randomly replacing objects with different models from the same category and rendering using the iGibson simulator to render panoramas. The selection and placement of objects in iGibson-DPC is based on heuristic algorithms, while our R3DS is manually annotated and placed 3D models are verified in terms of match and alignment to the object masks. + +Moreover, iGibson-DPC contains unrealistic object arrangements (e.g., floating TV in Fig. 2). + +# 4 Experiments + +We showcase the value of R3DS on the Panoramic Scene Understanding (PanoSun) task [7, 28, 29]. Given an input RGB panorama, the goal is to estimate the room layout, detect objects in 2D, estimate their 3D oriented bounding boxes and also reconstruct 3D object meshes. Our experiments show that methods trained on R3DS data benefit from its realism and generalize better when evaluated on photorealistic images. We also investigate the role of object support hierarchy information in improving performance. + +# 4.1 Task setup + +Method. DeepPanoContext (DPC) [28] predicts the room layout, detects objects in 3D and recovers object meshes from a panorama image using a relation-based graph convolutional network and a differentiable relation optimization procedure. Since DPC has a publicly-available implementation, we use it to benchmark the R3DS data on the PanoSun task. We keep all hyperparameters unchanged except lowering the relation optimization loss weight of 3D bounding box back-projection from 10 to 1, since the ground truth 2D masks are noisy. + +Datasets. We train and evaluate DPC on the iGibson-DPC (IG) [7, 18, 28], Structured3D (S3D) [30], and R3DS datasets. Zhang et al. [28] render 1,500 panoramas from 15 iGibson houses composed of $500+$ objects spanning 57 object categories. We use the same data and splits for IG. Structured3D consists of 3500 houses and around 18K photo-realistic rendered panoramas in total. We use 14K for training and the remaining 4K for testing. Note that Structured3D does not provide ground truth object meshes. + +To prepare R3DS for this task, we generate the ground truth room layout from the 3D architecture based on the camera viewpoint and obtain 3D oriented bounding boxes (OBBs) from all objects. We use 2D object masks from the Matterport3D mesh instance segmentation. We consider three variants of R3DS based on the input panorama: $R3DS$ -real where we use the Matterport3D panoramas, $R3DS$ -syn where we use rendered panoramas (at the same camera poses) from the annotated synthetic scenes, and $R3DS$ -mix where we combine the two types of panoramas and double the available data. We follow the MP3D house split and merge the train and val sets to obtain a disjoint split of 15 train and 5 test houses. Based on the split, we have 696 annotated panoramas for train and 146 for test. To fairly evaluate methods trained on different datasets, we curate a list of 25 object classes common to all datasets. + +Metrics. Following Zhang et al. [28] we use separate metrics for room layout estimation, 3D object detection, and scene relation prediction. For room layout estimation, we use 2D IoU for predicted 2D floorplan, 3D IoU for lifted 3D room geometry, and dRMSE for predicted depth with respect to the camera location. + +Table 2: Room layout estimation on R3DS-real test set. DPC [28] was pretrained on IG and S3D. For the last three rows, we fine-tune the pretrained weights on variants of R3DS. + +
Train2D IoU ↑3D IoU ↑dRMSE ↓
DPC [28]53.450.30.682
R3DS-real55.153.10.610
R3DS-syn59.056.10.629
R3DS-mix59.657.00.572
+ +Table 3: Cross-dataset evaluation for the Panoramic Scene Understanding task. We evaluate 3D detections with class-agnostic IoU and mAP at IoU of 0.15, and report object collisions. The highlighted rows indicate the most challenging scenario. + +
TestTrain3D detection ↑Collision ↓Attachment F1 ↑
IoUmAPmesharchobjwallfloor
IG27.530.31.6622.59453.176.895.086.2
IG+R3DS24.030.21.4042.25459.764.194.62.7
R3DS-real17.313.40.2421.45638.864.092.80.0
R3DS-syn23.214.20.4801.93848.546.793.828.6
R3DS-mix21.615.60.4341.24843.167.290.19.8
S3D19.53.51.0162.65150.968.790.811.6
IG+R3DS19.77.00.8682.08952.067.491.21.8
R3DS-real18.47.10.6002.59845.061.689.70.7
R3DS-syn19.04.80.6442.56149.349.791.22.4
R3DS-mix19.67.50.4631.67347.864.187.20.9
R3DS15.65.90.5751.95953.850.751.20.0
IG+R3DS17.514.10.2811.26749.561.658.60.0
R3DS-real16.415.00.2261.56244.057.358.90.0
R3DS-syn14.08.40.3901.66454.140.649.10.0
R3DS-mix17.615.80.1711.00748.558.360.10.0
+ +For 3D object detection, we report bounding box-based class-agnostic 3D IoU as well as mean average precision (mAP) across the 25 object classes, where an IoU greater than 0.15 counts as a "true" result. For scene relation prediction, we report F1 scores for relation classification. We also report the average number of objects colliding with each other or with architectural structures (wall, floor, ceiling). Specifically, we follow Zhang et al. [28] and measure collisions using the Separating Axis Theorem (SAT) to test whether object bounding boxes overlap. Since bounding box-based collision is a poor proxy for real-world physical collision, we also compute mesh-based collision by checking if the meshes for object pairs have any interpenetrating triangles [11, 24]. + +![](images/6d998a1d6b58433b80cea888a734c8aa9d4097684a20ec35bf88c28f72426518.jpg) +Fig. 5: Qualitative results for cross-dataset Panoramic Scene Understanding task. Correct and incorrect object detections shown in green and red boxes. Ground truth room layout and meshes shown in gray color, while layout prediction is in yellow. Training on R3DS leads to fewer errors compared to training on other datasets, especially when mixed with real data. + +# 4.2 Results + +1) Does R3DS help DPC generalize to real images? Since the original DPC work only trained and evaluated on synthetic data, it is unclear how well it performs on realistic panoramic imagery. We hypothesize that training on R3DS will lead to better performance. We separately show results on room layout estimation (Tab. 2) and 3D object detection. + +Room layout estimation. For room layout estimation, DPC uses Horizon-Net [21] pretrained on iGibson (IG) and Structured3D (S3D) panoramas. This model achieves good performance on IG data (91.0 3D IoU). When directly testing the official pretrained model on $R3DS$ -real panoramas, we notice a significant performance drop compared to results on the rendered panoramas from iGibson (Tab. 2 shows that the 3D IoU drops to 50.3). By finetuning the pretrained model with $R3DS$ -real, we can predict more precise room layouts for real cluttered scenes. Even only trained on $R3DS$ -syn, we outperform the original DPC model by $5.6\%$ and $5.8\%$ on 2D and 3D IoU, respectively. This is likely due to renderings from $R3DS$ -syn reflecting more realistic object arrangements in a room instead of pushing all objects against walls. Best performance on 2D IoU (59.6), 3D IoU (57.0) and depth RMSE (0.572) is achieved by fine tuning on $R3DS$ -mix. + +Table 4: Performance of models trained on three synthetic datasets (IG, S3D, and IG+S3D) evaluated on the R3DS-full dataset, where "full" indicates all 840 panoramicas are used for testing. + +
Train3D detectionCollision ↓Support F1 ↑
IoU↑mAP↑mesharchobj floorceil
IG14.25.20.7031.6394.185.3
S3D16.410.00.1121.2263.186.8
IG+S3D17.19.70.1331.1623.384.8
+ +Object detection. For 3D object detection, we train DPC on different data settings and conduct a cross-dataset evaluation (see Tab. 3). To investigate how models perform on out-of-distribution scenes, we evaluate models on Structured3D, as its images are near-realistic. To explore whether DPC training benefits from R3DS given the same amount of data, we create a special data input $IG + R3DS$ that combines iGibson and R3DS panoramas by randomly replacing half (500) of iGibson data with $R3DS$ -real data. The results show that $IG + R3DS$ performs almost the same as $IG$ with fewer collisions on iGibson, but it remarkably outperforms $IG$ on the test set of $R3DS$ and $S3D$ by 8.2 and 3.5 improvements on 3D mAP, respectively. It also averages 0.221 fewer mesh collisions. There are noticeable performance gaps on iGibson for models trained on R3DS data likely due to the data domain shift. Among the three variants of R3DS data, $R3DS$ -mix outperforms the others on all three test sets regarding 3D IoU and 3D mAP with the fewest mesh and architecture collisions. Although $R3DS$ -syn underperforms on $R3DS$ and $S3D$ test sets, it achieves better performance than $IG$ with even less data. + +Scene relation classification. We report F1 scores for identifying attachment relationships of objects to other objects and architecture elements (see Tab. 3). We note that models trained with synthetic renderings perform better than those trained on real images. That is because synthetic renderings present cleaner and simpler scenes with fewer objects than real world and simpler illumination such that DPC finds it easier to learn object-object and object-architecture relations. Also, note that the predictions of object-ceiling attachments can be extremely low because few objects are attached to the ceiling in the ground truth data. We show qualitative examples in Figure 5. + +2) Is R3DS a challenging, high-quality test set? How would a model trained on pure synthetic data perform on complex real data (R3DS)? Due to its modest scale, we propose using R3DS as a challenging, high-quality test set rather than a train set. Specifically, we evaluate the synthetic-to-real performance of DPC by training on iGibson and/or Structured3D and testing on all panoramas in R3DS-real. Table 4 shows that a model trained with Structured3D performs the best (10.0 3D mAP and 0.112 mesh collision) as it observes the most photo-realistic images. DPC benefits from the synthetic data for higher bounding box IoUs, since it possesses accurately aligned 3D bounding box and + +![](images/9b5bba0ece1b1c2ef146afa2e3c9b4cdb4f8fed4eb9c9cfa099902f16bbe2847.jpg) + +![](images/a9664d032672092e7f57a467f56a7953fff909237f7ec9ddf77b9985f25df455.jpg) +(a) Object appearance gap $(60\%)$ + +![](images/af2b5338e6663aae7efe41d48e274464aa8dad559ebd929a77d08be854d0fd42.jpg) + +![](images/ac9a00bfd5a6f7027d94e75047d7b78b48be61cfa2c1629c5575d53bfd9ebf0f.jpg) +(b) Common occlusions (76.7%) + +![](images/f1209d65168c46fc11e3a01cbe261163def8ebc142c1ec0b0c8bf2c89f9c1e36.jpg) + +![](images/b88b0680aeb84e234cf6f8400cbda72a61ff6e309f4ee3d902175d61346a7730.jpg) +(c) Wrong 3D prediction $(65\%)$ + +![](images/5bcc3a638f5c4300acc3018e8a9f30f76652bf422af6081d25b382bbbc681e14.jpg) + +![](images/826ed6e7d75414a4cdda1ec61f1d1e3d65f913f855118a62fad8dd3c9fe3f865.jpg) +(d) Objects in mirrors (15.6%) +Fig. 6: Error analysis on R3DS panoramas using DPC pretrained on S3D. Typical prediction errors are categorized into 4 groups: (a) Appearance gap between real and synthetic images. (b) Detection failure due to common occlusions. (c) Bad performance of monocular 3D prediction. (d) Perception error for objects in mirrors. Numbers in parentheses mean the error rates. + +![](images/eda72c7838d71b3a9d0a4a0e3344f8a6eb1a8667be8c2df1715077c63d2fc584.jpg) +Panorama + +![](images/55e6b6e47c630c4f3c949b1d21ab86b15d074c3d90ec835db3e70d803eb0bc44.jpg) +Fig. 7: Comparison between heuristic and R3DS support relations. Dark lines between objects show support relations. Heuristic support often mistakenly assigns relations to spatially close objects. In contrast, R3DS support relations are annotated. + +![](images/3d6d39fd0b8998887e4edcb3f56eb6972c290eae77ff140de3f5e26c0e05cd56.jpg) +Heuristic Support + +![](images/f35887575e5067bf1ce1c6f326266669a49d3d36e7fce4fa11abf4e3de870dca.jpg) + +![](images/921e3c29ca87d8b788bf8410ff2af6d632dc69be794505a03a6ea0acb3543302.jpg) +R3DS Support + +![](images/606970567abdc77fe6603563009f963acaa8200aff9c26db499add1c4c2bfc57.jpg) + +more unoccluded objects. However, mAP performance is lower due to worse object recognition ability. All models struggle to predict correct object-wise support relations but do a better job of predicting object-floor support relations. + +We conduct error analysis on 120 randomly sampled panoramas using the model pretrained on $S3D$ to identify typical errors (see Fig. 6). Errors are categorized into 4 groups: (a) $60\%$ panoramas have 2D perception errors due to the synthetic-to-real appearance gap; (b) $76.7\%$ panoramas show detection failures due to occlusions; (c) $65\%$ panoramas exhibit correct 2D detections but fail to correctly perform 3D predictions; and (d) $15.6\%$ out of 45 panoramas with mirrors mistakenly predict virtual objects in mirrors. + +3) Are R3DS support relations helpful for PanoSun? We investigate whether the support relationships between objects provided in our R3DS scene hierarchy help boost performance of holistic scene understanding. We augment DPC's Relation Scene-GCN module with additional support relation prediction branches. Besides obtaining explicitly annotated scene support relations from + +Table 5: Performance on R3DS-real of DPC models trained on variants of R3DS with different support relation settings. We compare the original model without support (none) against models supervised with support that is heuristically computed (heur) or annotated from R3DS (anno). Classification results are evaluated on annotated ground-truth scene hierarchy. + +
TrainSupp.3D detectionCollision ↓Support F1 ↑
IoU↑mAP↑mesharchobj floorceil
R3DS-realnone16.415.00.2261.562--
heur16.214.10.2191.3293.269.9
anno16.615.90.3491.40438.594.4
R3DS-synnone14.08.40.3901.664--
heur14.67.80.2811.3014.682.9
anno14.38.20.3491.21932.095.0
R3DS-mixnone17.615.80.1711.007--
heur19.217.70.1511.2673.683.7
anno18.618.20.1581.30812.096.2
+ +Table 6: Comparison of the average number of bounding box and mesh-based object collisions per scene in IG and R3DS. R3DS exhibits more bounding box-based collisions, but almost none of these are actual physical collisions between object meshes. Measuring collisions between bounding boxes is a poor collision measure for fully-populated, real-world scenes. + +
DatasetsMeshCollisionsBox Collisions
obj-objobj-wallobj-floorobj-ceil
IG-1.1850.0750.0000.790
R3DS0.00062.8230.3880.0350.064
+ +R3DS, it is also possible to compute heuristic support relations from object bounding boxes. Specifically, an object is supported by another if their bounding boxes intersect within tolerance distance of $0.1\mathrm{m}$ and the centroid of the former object is higher than that of the latter. Support by wall/floor/ceiling is calculated in the same way without the height judgment. This definition is similar to how DPC defines object attachment. Figure 7 compares these two ways of computing support relations, showing that heuristic computation can mistakenly designate support relations to two nearby objects. Table 5 shows that incorporating support relation prediction indeed influences the performance of DPC. Heuristic support information may worsen 3D object detection (mAP in R3DS-real and R3DS-syn), but it eliminates mesh collisions the most. Learning support relations from R3DS annotations leads to a 2.4 improvement on mAP in R3DS-mix, although the classification F1 score is low. + +Table 7: Ablation of relation optimization (RO) on R3DS-real. The 2nd and 3rd row remove optimization terms in RO. The last row replaces bounding-box collisions with mesh collisions. + +
Rel. Opt.3D mAP ↑ MeshCollisions ↓Box Collisions ↓
obj-objobj-arch
DPC18.20.1580.0621.308
w/o obj col18.91.3421.1301.301
w/o obj col+tch19.61.2191.0621.295
w/ mesh col19.71.0270.8561.394
+ +4) Is relation optimization (RO) effective on R3DS? Test-time relation optimization (RO) was introduced by Zhang et al. [28] to reduce physical violations, floating objects, and misalignment between objects and architecture. The original cost function based on bounding box collisions succeeds in optimizing object poses, since there are few such collisions in the IG data originally used for evaluation (see in Tab. 6). However, the same data assumption does not hold for R3DS, which has more bounding-box-based collisions but nearly zero mesh-based collisions. We ablate the design of RO on R3DS-real in Table 7. By removing two optimization terms (bounding-box-based object-wise collision and touching step-by-step), the model outperforms the original one in 3D mAP (+1.4) but degrades in mesh-based and box-based collisions. We show that using mesh-based collision optimization leads to the best performance. The increase in collisions is unsurprising as the R3DS data reflects more cluttered real interiors. Limitations. Our dataset construction relied on 3D architectures for each Matterport3D scan which are simplifications of the geometry of the real environment. One issue is imperfect wall positions, resulting in objects attached to these virtual walls being offset from the true surface. In addition, objects in our 3D scenes were placed without regard to the materials, meaning that the detailed surface appearance does not match that of the observed object. Future work can investigate transfer of surface appearance to the synthetic objects by projecting textures from the RGB-D data and 3D reconstructed meshes. + +# 5 Conclusion + +We introduced the R3DS dataset. R3DS provides more complete, densely populated, and richly annotated synthetic 3D scene proxies of real-world environments with linked panoramic images. We showed the usefulness of R3DS on the Panoramic Scene Understanding task. Our experiments demonstrate the value of realistic synthetic recreations in this task, in particular through the use of object support information. While we focused on the PanoSun task, R3DS can also be useful for other tasks such as single-view shape retrieval, single-view object pose estimation, and panoramic scene graph prediction. + +# Acknowledgements + +This work was funded in part by a CIFAR AI Chair, a Canada Research Chair, NSERC Discovery Grant, NSF award #2016532, and enabled by support from WestGrid and Compute Canada. Daniel Ritchie is an advisor to Geope and owns equity in the company. Geope is a start-up that is developing 3D technology to build immersive virtual copies of the real world with applications in various fields, including games and architecture. We thank Madhawa Vidanapathirana, Weijie Lin, and David Han for help with development of the annotation tool, and Denys Iliash, Mrinal Goshalia, Brandon Robles, Paul Brown, Chloe Ye, Coco Kaleel, Elizabeth Wu and Hannah Julius for data annotation, and Ivan Tam, Austin Wang, and Ning Wang for feedback on the paper draft. + +# References + +1. Avetisyan, A., Dahnert, M., Dai, A., Savva, M., Chang, A.X., Nießner, M.: Scan2CAD: Learning CAD model alignment in RGB-D scans. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019) +2. Batra, D., Chang, A.X., Chernova, S., Davison, A.J., Deng, J., Koltun, V., Levine, S., Malik, J., Mordatch, I., Mottaghi, R., Savva, M., Su, H.: Rearrangement: A challenge for embodied AI. arXiv preprint arXiv:2011.01975 (2020) +3. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niebner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3D: Learning from RGB-D data in indoor environments. In: Proceedings of the International Conference on 3D Vision (3DV). pp. 667-676. IEEE (2017) +4. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015) +5. Collins, J., Goel, S., Deng, K., Luthra, A., Xu, L., Gundogdu, E., Zhang, X., Vicente, T.F.Y., Dideriksen, T., Arora, H., et al.: Abo: Dataset and benchmarks for real-world 3d object understanding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21126-21136 (2022) +6. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5828-5839 (2017) +7. Dong, Y., Fang, C., Dong, Z., Bo, L., Tan, P.: PanoContext-Former: Panoramic total scene understanding with a transformer. arXiv preprint arXiv:2305.12497 (2023) +8. Fu, H., Cai, B., Gao, L., Zhang, L., Li, C., Xun, Z., Sun, C., Fei, Y., Zheng, Y., Li, Y., et al.: 3D-FRONT: 3D Furnished Rooms with layOuts and semaNTics. arXiv preprint arXiv:2011.09127 (2020) + +9. Fu, H., Jia, R., Gao, L., Gong, M., Zhao, B., Maybank, S., Tao, D.: 3D-Future: 3D Furniture shape with TextURE. arXiv preprint arXiv:2009.09633 (2020) +10. Hua, B.S., Pham, Q.H., Nguyen, D.T., Tran, M.K., Yu, L.F., Yeung, S.K.: SceneNN: A scene meshes dataset with annotations. In: Proceedings of the International Conference on 3D Vision (3DV). pp. 92-101. IEEE (2016) +11. Karras, T.: Maximizing parallelism in the construction of bvhs, octrees, and k-d trees. In: Proceedings of the Fourth ACM SIGGRAPH / Eurographics Conference on High-Performance Graphics. pp. 33-37. Eurographics Association (2012). https://doi.org/10.2312/EGGH/HPG12/033-037, https://doi.org/10.2312/EGGH/HPG12/033-037 +12. Li, Z., Yu, T.W., Sang, S., Wang, S., Song, M., Liu, Y., Yeh, Y.Y., Zhu, R., Gundavarapu, N., Shi, J., Bi, S., Xu, Z., Yu, H.X., Sunkavalli, K., Hasan, M., Ramamoorthi, R., Chandraker, M.: OpenRooms: An end-to-end open framework for photorealistic indoor scene datasets. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +13. Lim, J.J., Pirsiavash, H., Torralba, A.: Parsing intelia objects: Fine pose estimation. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 2992-2999 (2013) +14. Maninis, K.K., Popov, S., Nießner, M., Ferrari, V.: CAD-estate: Large-scale CAD model annotation in RGB videos. arXiv preprint arXiv:2306.09011 (2023) +15. Paschalidou, D., Kar, A., Shugrina, M., Kreis, K., Geiger, A., Fidler, S.: Atiss: Autoregressive transformers for indoor scene synthesis. Advances in Neural Information Processing Systems 34, 12013-12026 (2021) +16. Ramakrishnan, S.K., Gokaslan, A., Wijmans, E., Maksymets, O., Clegg, A., Turner, J., Undersander, E., Galuba, W., Westbury, A., Chang, A.X., et al.: Habitat-Matterport 3D dataset (hm3d): 1000 large-scale 3D environments for embodied AI. arXiv preprint arXiv:2109.08238 (2021) +17. Sadalgi, S.: Wayfair's 3D Model API. https://www/aboutwayfair.com/tech-innovation/wayfairs-3d-model-api (2016), [Online; accessed 15-Nov-2023] +18. Shen, B., Xia, F., Li, C., Martin-Martin, R., Fan, L., Wang, G., Buch, S., D'Arpino, C., Srivastava, S., Tchapmi, L.P., Vainio, K., Fei-Fei, L., Savarese, S.: iGibson, a simulation environment for interactive tasks in large realistic scenes. In: Proceedings of the International Conference on Intelligent Robots and Systems (IROS) (2021) +19. Shen, B., Xia, F., Li, C., Martin-Martin, R., Fan, L., Wang, G., Pérez-D'Arpino, C., Buch, S., Srivastava, S., Tchapmi, L., et al.: iGibson 1.0: A simulation environment for interactive tasks in large realistic scenes. In: Proceedings of the International Conference on Intelligent Robots and Systems (IROS). pp. 7520-7527. IEEE (2021) +20. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019) + +21. Sun, C., Hsiao, C.W., Sun, M., Chen, H.T.: Horizonnet: Learning room layout with 1d representation and pano stretch data augmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1047-1056 (2019) +22. Sun, X., Wu, J., Zhang, X., Zhang, Z., Zhang, C., Xue, T., Tenenbaum, J.B., Freeman, W.T.: Pix3D: Dataset and methods for single-image 3D shape modeling. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2974-2983 (2018) +23. Szot, A., Clegg, A., Undersander, E., Wijmans, E., Zhao, Y., Turner, J., Maestre, N., Mukadam, M., Chaplot, D.S., Maksymets, O., et al.: Habitat 2.0: Training home assistants to rearrange their habitat. Advances in Neural Information Processing Systems 34, 251-266 (2021) +24. Tzionas, D., Ballan, L., Srikantha, A., Aponte, P., Pollefeys, M., Gall, J.: Capturing hands in action using discriminative salient points and physics simulation. International Journal of Computer Vision (IJCV) 118(2), 172-193 (Jun 2016), https://doi.org/10.1007/s11263-016-0895-4 +25. Wang, K., Lin, Y.A., Weissmann, B., Savva, M., Chang, A.X., Ritchie, D.: Planit: Planning and instantiating indoor scenes with relation graph and spatial prior networks. ACM Transactions on Graphics (TOG) 38(4), 1-15 (2019) +26. Xiang, Y., Kim, W., Chen, W., Ji, J., Choy, C., Su, H., Mottaghi, R., Guibas, L., Savarese, S.: Objectnet3D: A large scale database for 3D object recognition. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 160-176. Springer (2016) +27. Yadav, K., Ramrakhya, R., Ramakrishnan, S.K., Gervet, T., Turner, J., Gokaslan, A., Maestre, N., Chang, A.X., Batra, D., Savva, M., et al.: Habitat-Matterport 3D semantics dataset. arXiv preprint arXiv:2210.05633 (2022) +28. Zhang, C., Cui, Z., Chen, C., Liu, S., Zeng, B., Bao, H., Zhang, Y.: Deep-PanoContext: Panoramic 3D scene understanding with holistic scene context graph and relation-based optimization. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 12632-12641 (2021) +29. Zhang, Y., Song, S., Tan, P., Xiao, J.: PanoContext: A whole-room 3D context model for panoramic scene understanding. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 668-686. Springer (2014) +30. Zheng, J., Zhang, J., Li, J., Tang, R., Gao, S., Zhou, Z.: Structured3D: A large photo-realistic dataset for structured 3D modeling. arXiv preprint arXiv:1908.00222 (2019) +31. Zhou, T., Tucker, R., Flynn, J., Fyffe, G., Snavely, N.: Stereo magnification: learning view synthesis using multiplane images. ACM Transactions on Graphics (TOG) 37(4), 1-12 (2018) \ No newline at end of file diff --git a/2024/R3DS_ Reality-linked 3D Scenes for Panoramic Scene Understanding/images.zip b/2024/R3DS_ Reality-linked 3D Scenes for Panoramic Scene Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..df6c6f1c79886dd613893ca018f49c8e656b0963 --- /dev/null +++ b/2024/R3DS_ Reality-linked 3D Scenes for Panoramic Scene Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a1b01a2928e9ecc96107501e9f36a111f7ac8f636c75ae7e549369646d2c9942 +size 671646 diff --git a/2024/R3DS_ Reality-linked 3D Scenes for Panoramic Scene Understanding/layout.json b/2024/R3DS_ Reality-linked 3D Scenes for Panoramic Scene Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ec5b26ba96d114e3ccd2b5103cce32def37fd548 --- /dev/null +++ b/2024/R3DS_ Reality-linked 3D Scenes for Panoramic Scene Understanding/layout.json @@ -0,0 +1,7525 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 180, + 112, + 434, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 112, + 434, + 148 + ], + "spans": [ + { + "bbox": [ + 180, + 112, + 434, + 148 + ], + "type": "text", + "content": "R3DS: Reality-linked 3D Scenes for Panoramic Scene Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "spans": [ + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "text", + "content": "Qirui Wu" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Sonia Raychaudhuri" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Daniel Ritchie" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Manolis Savva" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "text", + "content": ", and Angel X. Chang" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 252, + 201, + 361, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 201, + 361, + 213 + ], + "spans": [ + { + "bbox": [ + 252, + 201, + 361, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 252, + 201, + 361, + 213 + ], + "type": "text", + "content": " Simon Fraser University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 266, + 213, + 347, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 213, + 347, + 224 + ], + "spans": [ + { + "bbox": [ + 266, + 213, + 347, + 224 + ], + "type": "text", + "content": "2 Brown University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 209, + 224, + 403, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 224, + 403, + 236 + ], + "spans": [ + { + "bbox": [ + 209, + 224, + 403, + 236 + ], + "type": "text", + "content": "3 Alberta Machine Intelligence Institute (Amii)" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 134, + 257, + 479, + 395 + ], + "blocks": [ + { + "bbox": [ + 134, + 257, + 479, + 395 + ], + "lines": [ + { + "bbox": [ + 134, + 257, + 479, + 395 + ], + "spans": [ + { + "bbox": [ + 134, + 257, + 479, + 395 + ], + "type": "image", + "image_path": "ff7acf139001c077257d8e97088718c8a412a9fa3ad7cf529f28962eaddf32c9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 403, + 482, + 502 + ], + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 502 + ], + "type": "text", + "content": "Fig. 1: Left: the Reality-linked 3D Scenes dataset (R3DS) fills a gap between synthetic 3D scenes and reconstructions of real-world environments by providing 3D scene proxies linked to real-world panoramas from Matterport3D (three example panoramas and 3D scenes shown). Right: our dataset contains scenes with higher density and completeness compared to prior datasets, and provides additional annotations such as object support (what objects or architectural elements support other objects), and matching object sets (e.g., pairs of the same nightstand). We use our dataset for the panoramic scene understanding task and demonstrate its value for research on room layout estimation, as well as 2D and 3D object detection." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 522, + 455, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 522, + 455, + 666 + ], + "spans": [ + { + "bbox": [ + 159, + 522, + 455, + 666 + ], + "type": "text", + "content": "Abstract. We introduce the Reality-linked 3D Scenes (R3DS) dataset of synthetic 3D scenes mirroring the real-world scene arrangements from Matterport3D panoramas. Compared to prior work, R3DS has more complete and densely populated scenes with objects linked to real-world observations in panoramas. R3DS also provides an object support hierarchy, and matching object sets (e.g., same chairs around a dining table) for each scene. Overall, R3DS contains 19K objects represented by 3,784 distinct CAD models from over 100 object categories. We demonstrate the effectiveness of R3DS on the Panoramic Scene Understanding task. We find that: 1) training on R3DS enables better generalization; 2) support relation prediction trained with R3DS improves performance compared to heuristically calculated support; and 3) R3DS offers a challenging benchmark for future work on panoramic scene understanding." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 230, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 230, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 230, + 127 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 145, + 482, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 145, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 145, + 482, + 289 + ], + "type": "text", + "content": "Datasets of 3D indoor environments are increasingly used for research on scene understanding [1, 3, 28], embodied AI [2, 19, 23], and scene generation [15, 25]. There are two strategies for constructing 3D scene datasets: reconstruction of real-world spaces [3], or authoring scenes using synthetic 3D objects [8] (CAD models). Reconstruction captures real spaces but is hard to scale, and the resulting scenes exhibit imperfections and artifacts. On the other hand, synthetic 3D scenes are complete and easy to manipulate but often do not match the statistics of real-world spaces and are artificially \"clean\". Moreover, both strategies are time-consuming and require expertise. There have been some attempts to create \"synthetic\" replicas of real environments by matching CAD models to objects in scans [19, 23]. These efforts have been limited in scale and often result in partial and sparsely populated synthetic counterparts of the real environments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 289, + 482, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 374 + ], + "type": "text", + "content": "We design a framework that allows users to create 3D scenes from RGB panoramas and use it to create R3DS: a dataset of 'Reality-linked' 3D Scenes. Each 3D scene in our dataset is a complete proxy of an environment from the Matterport3D [3] dataset, representing both the 3D architecture and the objects. Thus, each scene is linked to a real space, with correspondences established between panorama observations of each object and the synthetic object. These reality-linked scenes reflect denser real-world arrangements of objects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 375, + 482, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 482, + 471 + ], + "type": "text", + "content": "The use of panoramas for reference is advantageous compared to either perspective images or 3D reconstructions. Panoramas are not limited by the field of view unlike perspective images, enabling more complete 3D synthetic proxies. Panoramas also better capture relatively small objects and objects with challenging materials or illumination conditions compared to reconstructions. Additionally, there is a scarcity of synthetic 3D scenes coupled with real-world panoramas, with only one relatively small algorithmically constructed dataset provided by Zhang et al. [28] being available to the community." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 471, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 544 + ], + "type": "text", + "content": "Compared to prior efforts such as Scan2CAD [1] and CAD-Estate [14], our dataset provides more complete scenes, with salient observed objects being captured in the layout. Moreover, we provide a support hierarchy defining what objects are placed on other objects and specify sets of identical objects such as dining chairs around a table, allowing for creating realistic variations of the scene by swapping the entire set to a different chair design." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 544, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 605 + ], + "type": "text", + "content": "We demonstrate the value of our dataset by using it for the Panoramic Scene Understanding task. We show that leveraging the denser layouts and support hierarchy information in our scenes leads to improved object detection performance and better generalization compared to training using other datasets previously used for this task. In summary, we make the following contributions:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 616, + 479, + 663 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 137, + 616, + 479, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 616, + 479, + 637 + ], + "spans": [ + { + "bbox": [ + 137, + 616, + 479, + 637 + ], + "type": "text", + "content": "- We design a framework for efficient construction of synthetic scenes from real panoramas and use it to create R3DS: a dataset of reality-linked 3D scenes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 641, + 479, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 641, + 479, + 663 + ], + "spans": [ + { + "bbox": [ + 137, + 641, + 479, + 663 + ], + "type": "text", + "content": "- R3DS provides more complete and realistic scenes with correspondences between real and synthetic objects, and object-object support relations." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "text", + "content": "Q. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 118, + 459, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 118, + 459, + 127 + ], + "spans": [ + { + "bbox": [ + 137, + 118, + 459, + 127 + ], + "type": "text", + "content": "R3DS (Ours): linked to panoramicas, densely populated scenes, matched object sets, objects supported" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 135, + 128, + 479, + 205 + ], + "blocks": [ + { + "bbox": [ + 135, + 128, + 479, + 205 + ], + "lines": [ + { + "bbox": [ + 135, + 128, + 479, + 205 + ], + "spans": [ + { + "bbox": [ + 135, + 128, + 479, + 205 + ], + "type": "image", + "image_path": "13caafa1cafea1e0f7f812fec3dcbfb2055dd8bb5c1972f496effc9bb7d6bc79.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 136, + 206, + 252, + 274 + ], + "blocks": [ + { + "bbox": [ + 136, + 206, + 252, + 274 + ], + "lines": [ + { + "bbox": [ + 136, + 206, + 252, + 274 + ], + "spans": [ + { + "bbox": [ + 136, + 206, + 252, + 274 + ], + "type": "image", + "image_path": "8f160698894748f4a793518d06989ec6f99b5f51bfe62065b613eecca100ea01.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 274, + 234, + 283 + ], + "lines": [ + { + "bbox": [ + 155, + 274, + 234, + 283 + ], + "spans": [ + { + "bbox": [ + 155, + 274, + 234, + 283 + ], + "type": "text", + "content": "CAD-Estate: sparse, floating" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 295, + 482, + 361 + ], + "lines": [ + { + "bbox": [ + 130, + 295, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 482, + 361 + ], + "type": "text", + "content": "Fig. 2: Dataset comparison. (Top) shows different views of a scene annotated in R3DS. Comparison with previous datasets (bottom) shows (1) R3DS has more complete scenes than the previous datasets; (2) Objects in R3DS are properly supported by either architecture or other objects unlike the others (e.g. floating objects with no proper support); (3) R3DS is annotated using the same 3D model for objects arranged together (chairs by the dining table, couches arranged together)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 256, + 206, + 377, + 274 + ], + "blocks": [ + { + "bbox": [ + 256, + 206, + 377, + 274 + ], + "lines": [ + { + "bbox": [ + 256, + 206, + 377, + 274 + ], + "spans": [ + { + "bbox": [ + 256, + 206, + 377, + 274 + ], + "type": "image", + "image_path": "b3a9e4b37345c3acd518271a6033b3db202843053028d532bdb00b9aaec8d748.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 275, + 375, + 283 + ], + "lines": [ + { + "bbox": [ + 261, + 275, + 375, + 283 + ], + "spans": [ + { + "bbox": [ + 261, + 275, + 375, + 283 + ], + "type": "text", + "content": "Scan2CAD: sparse, floating, mismatched" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 380, + 207, + 479, + 273 + ], + "blocks": [ + { + "bbox": [ + 380, + 207, + 479, + 273 + ], + "lines": [ + { + "bbox": [ + 380, + 207, + 479, + 273 + ], + "spans": [ + { + "bbox": [ + 380, + 207, + 479, + 273 + ], + "type": "image", + "image_path": "bea44ff58d17b393c906c9f9715fda51b081eb836a52841492dcbfd6359c31a5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 274, + 460, + 283 + ], + "lines": [ + { + "bbox": [ + 399, + 274, + 460, + 283 + ], + "spans": [ + { + "bbox": [ + 399, + 274, + 460, + 283 + ], + "type": "text", + "content": "iGibson DPC: floating" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 386, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 386, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 138, + 386, + 482, + 434 + ], + "type": "text", + "content": "- We show that the more complete layouts and support relations in our dataset enable better performance and generalization in the Panoramic Scene Understanding task, and that our dataset offers a challenging benchmark for future work in scene understanding." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 460, + 237, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 237, + 473 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 237, + 473 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 486, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 616 + ], + "type": "text", + "content": "3D scene datasets. A spectrum of scene datasets have been used for scene understanding tasks. One type provides annotated 3D reconstructions of real scenes based on RGB-D videos [3, 6, 10, 16, 20, 27]. These datasets are usually subject to the limitations of RGB-D reconstruction, typically containing noise, artifacts such as holes, and poor reconstructions of thin structures, shiny objects, or light sources. Another type of 3D datasets is authored by manually designing 3D object assets [5, 9] and inserting them into synthetic 3D scenes [8]. However, such datasets lack the realism of real-world reconstructions and demand expert knowledge, making them expensive to create. A third, hybrid approach which is closest to our work creates 3D scene datasets by aligning existing object CAD models to real world data." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Datasets that align CAD models to real world. There have been a number of recent efforts in aligning CAD models with real-world data. Prior work [13, 22, 26] has annotated object images with 3D models, typically using keypoint correspondences to perspective images. These perspective images usually do not" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "text", + "content": "R3DS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "depict a complete scene; they typically focus on one or two objects and are limited in field of view, resulting in a sparse proxy of the real scene." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 282 + ], + "type": "text", + "content": "Another line of work aligns 3D CAD models to RGB-D scans either through annotation as in Scan2CAD [1], or automated heuristics as in iGibson [19]. OpenRooms [12] extends Scan2CAD [1] with photorealistic material annotations and focuses on inverse rendering tasks. Conceptually, these allow for more complete synthetic scene proxies. However, statistics from these datasets show that they are still relatively sparse (see Tab. 1). In addition, the poor quality of reconstruction makes aligning CAD models challenging without referring to the original RGB images. A prominent exception is Replica [20] which has fairly high-quality reconstructions and the artist-created Replica-CAD [23]. However, creating such high quality \"replicas\" is labor intensive and costly. Szot et al. [23] report " + }, + { + "bbox": [ + 130, + 140, + 481, + 282 + ], + "type": "inline_equation", + "content": "900+" + }, + { + "bbox": [ + 130, + 140, + 481, + 282 + ], + "type": "text", + "content": " work hours required to model approximately 90 objects, resulting in a dataset of limited scale with 105 different layouts of what is effectively a single room." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 283, + 481, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 481, + 331 + ], + "type": "text", + "content": "More recently, Maninis et al. [14] introduced CAD-Estate, which aligns CAD models to RGB videos for over 19K spaces. Because the data is based on monocular video, the coverage of the spaces is incomplete. In addition, the annotation is relatively sparse, with an average of only 6 objects per scene." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 331, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 499 + ], + "type": "text", + "content": "Datasets for panoramic scene understanding. There have been relatively few datasets introduced for Panoramic Scene Understanding [7, 28, 29]. In the initial PanoContext dataset [29], the data did not have aligned CAD models and only included object cuboids. The ground truth data was collected on 2D panorama images by annotating visible cuboid vertices; 3D cuboids were obtained by minimizing the re-projection error from the annotated 2D vertices. Moreover, these 3D cuboids and the room layout are obtained with the assumption that the room layout is a cuboid and that the objects are vertically aligned. Thus, the resulting object layout may deviate from the real arrangement of objects. More recently, datasets for Panoramic Scene Understanding have been built by taking 3D scans, aligning CAD objects to them, and then generating panoramas [7, 28]. Compared to these datasets, our R3DS is manually curated for a larger number of distinct regions and provides support hierarchy and matching object set annotations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 514, + 266, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 266, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 266, + 528 + ], + "type": "text", + "content": "3 The R3DS Dataset" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 536, + 481, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 481, + 609 + ], + "type": "text", + "content": "We describe the construction of the R3DS dataset and present a statistical analysis of the scenes it contains. Compared to previous datasets [1, 14, 28] (Fig. 2), our scenes are more densely populated, and objects are annotated with a hierarchy of support relations. Moreover, our dataset specifies matching object instances in furniture arrangements. Figure 1 shows example annotations from our dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 624, + 266, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 266, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 266, + 635 + ], + "type": "text", + "content": "3.1 Dataset construction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "We developed a 3D annotation interface (Fig. 3) showing a panorama of a room from Matterport3D and allowing users to insert 3D CAD objects into a 3D scene" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "text", + "content": "Q. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 256, + 482, + 324 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 248 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 248 + ], + "type": "text", + "content": "Table 1: Comparison with 3D indoor scene datasets aligned with real-world images, videos, or scans. Our R3DS dataset contains more densely populated annotations compared to other datasets, with objects from 110 different categories. We report the unique models (#CAD), object categories (#Cat), object instances (#Objects) as well as average number of objects and object categories per annotation. For Scan2Cad [1] and ReplicaCAD [23] the average is per scan. Note that ReplicaCAD consists of 105 different layouts (arrangements) of effectively one room. CAD-Estate has partial views into 19K spaces, many of which are 1-2 rooms. Of the datasets used for panoramic scene understanding, our R3DS dataset covers more rooms with 842 panoramicas over 22 room types. Replica-Pano (in gray) was not released, so we report statistics from the paper. Our annotations per panorama are more complete and our dataset has both support relations (Sup) and matching object instance sets (Match)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 256, + 482, + 324 + ], + "lines": [ + { + "bbox": [ + 133, + 256, + 482, + 324 + ], + "spans": [ + { + "bbox": [ + 133, + 256, + 482, + 324 + ], + "type": "table", + "html": "
DatasetSourceCAD AlignmentTypeHouses/RoomsPanos#CAD#Objects#CatAveObjAveCatSupMatch
Scan2CAD [1]ScanNet [6]Annotatorscan- / 1506✘3,04914,225359.44.1✘✘
OpenRooms [12]ScanNet [6]Scan2CAD [1]scan- / 1288✘2,65116,0143812.46.3✘✘
ReplicaCAD [23]Replica [20]Artist recreationscan- / 105*✘922,2934421.814.4✘✘
CAD-Estate [14]RealEstate10K [31]Annotatorvideo19,512✘12,024100,882496.33.4✘✘
Replica-Pano [7]Replica [20]Heuristicpano- / 272700--25-✘✘
iGibson-DPC [28]iGibson [19]Heuristicpano15 / 100150050026,9985717.910.2✘✘
R3DS (Ours)Matterport3D [3]Annotatorpano20 / 3708423,78419,05011022.910.4✓✓
", + "image_path": "598ba0463e6e92e4268e3ef92dbca919995f53465674820db96ea540fe02a979.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 350, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 445 + ], + "type": "text", + "content": "which is visually overlaid on the panorama. The 3D scene is initially empty, consisting only of 3D architectural geometry which specifies the walls, floor, ceiling as well as the placement of openings (e.g. doors, windows, and other openings) on the walls. We create this 3D architecture by taking 20 houses from Matterport3D, constructing an initial architecture based on the region and object annotations for the windows and doors, and manually refining the placement of walls and openings. By combining panoramas and 3D architectures, users can see through openings and annotate objects located in other rooms." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "type": "text", + "content": "We ask annotators to select and place 3D object models to best match the panoramic image. We use CAD models from Wayfair [17] and ShapeNet [4] models collected from 3D Trimble warehouse. Wayfair provides a large collection of furniture CAD models that match real-world products and are sized based on real-world dimensions. However, it does not include bathroom fittings, electronic equipment and kitchen appliances, for which we manually scale and align CAD models from ShapeNet. Compared with ShapeNetCore, the CAD models we use are already sized to real-world sizes (instead of normalized to a unit cube)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": "To assist the annotators, we provide segmented masks of objects visible in the panorama. Since Matterport3D has annotated 3D object masks on the scans we use those annotations, but it is also possible to run an instance segmentation on the panorama. When the user clicks one of these masks, a search panel automatically opens and shows objects matching the clicked mask category label. For each mask, the annotator selects a matching object and positions and aligns it to match the mask. Annotators are instructed to choose objects which match the shape of the corresponding object in the panorama (rather than its color or texture). To help annotators focus on shape, we render all 3D objects in a neutral gray color. Annotators are also explicitly asked to select the same" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "text", + "content": "R3DS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 177, + 118, + 436, + 342 + ], + "blocks": [ + { + "bbox": [ + 177, + 118, + 436, + 342 + ], + "lines": [ + { + "bbox": [ + 177, + 118, + 436, + 342 + ], + "spans": [ + { + "bbox": [ + 177, + 118, + 436, + 342 + ], + "type": "image", + "image_path": "7d9a40641378deb27cc912483930a1134192010421f790ad506aed5c2e0cba91.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 350, + 482, + 384 + ], + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 384 + ], + "type": "text", + "content": "Fig. 3: R3DS annotation pipeline. Annotators see an empty scene (architecture only). They then insert and manipulate 3D object models from a panorama viewpoint to create a populated 3D scene proxy corresponding to the panorama." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 408, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 482, + 529 + ], + "type": "text", + "content": "3D asset for objects that should be the same; our interface provides a list of recently selected assets to make this process easier. In addition, annotators are instructed to add annotations for any objects that are not segmented (due to errors in Matterport3D) through simple clicks. These additional objects provide a more complete annotation that covers poorly reconstructed objects such as glass tables, lamps, and other small objects. The interface enforces that each object is placed on a support surface (either an architecture element or another object). The annotator can review their work by toggling off the panorama overlay or by switching to a perspective view of the 3D scene. For more annotated scene examples and details on the annotation process please refer to the supplement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 548, + 313, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 548, + 313, + 560 + ], + "spans": [ + { + "bbox": [ + 131, + 548, + 313, + 560 + ], + "type": "text", + "content": "3.2 Dataset analysis and statistics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "We collect annotations for 20 Matterport3D houses with 808 panoramas in total. We discard panoramas taken on stairs or outside a house, since they have a limited number of objects that can be placed. After filtering we have 769 panoramas for our analysis and experiments. For 73 panoramas, we collect two sets of annotations for each, to obtain a total of 842 annotated object arrangements across 22 different Matterport3D region types. The panoramas with two annotations serve as a test of annotator consistency and add diversity. In total, R3DS contains 19,050 object instances from 3,784 unique 3D CAD models spanning" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 150, + 133, + 255, + 191 + ], + "blocks": [ + { + "bbox": [ + 185, + 118, + 220, + 126 + ], + "lines": [ + { + "bbox": [ + 185, + 118, + 220, + 126 + ], + "spans": [ + { + "bbox": [ + 185, + 118, + 220, + 126 + ], + "type": "text", + "content": "Scan2CAD" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 150, + 133, + 255, + 191 + ], + "lines": [ + { + "bbox": [ + 150, + 133, + 255, + 191 + ], + "spans": [ + { + "bbox": [ + 150, + 133, + 255, + 191 + ], + "type": "image", + "image_path": "fb06118c490c651e0c9ab388dc1f3fc796a2703aa2a609cf399a9452a64ea020.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 191, + 238, + 200 + ], + "lines": [ + { + "bbox": [ + 167, + 191, + 238, + 200 + ], + "spans": [ + { + "bbox": [ + 167, + 191, + 238, + 200 + ], + "type": "text", + "content": "Floor warped, no arch" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 215, + 480, + 248 + ], + "lines": [ + { + "bbox": [ + 130, + 215, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 480, + 248 + ], + "type": "text", + "content": "Fig. 4: Architecture comparison. Compared to Scan2CAD (no architecture) and CAD-Estate (partial architecture), R3DS provides complete architecture with door/window portals." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 268, + 132, + 371, + 178 + ], + "blocks": [ + { + "bbox": [ + 294, + 118, + 332, + 126 + ], + "lines": [ + { + "bbox": [ + 294, + 118, + 332, + 126 + ], + "spans": [ + { + "bbox": [ + 294, + 118, + 332, + 126 + ], + "type": "text", + "content": "CAD-Estate" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 268, + 132, + 371, + 178 + ], + "lines": [ + { + "bbox": [ + 268, + 132, + 371, + 178 + ], + "spans": [ + { + "bbox": [ + 268, + 132, + 371, + 178 + ], + "type": "image", + "image_path": "e4674f4724b0f8db038006ff740d63989cb02799e85284258465603dd3387ae9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 299, + 192, + 336, + 200 + ], + "lines": [ + { + "bbox": [ + 299, + 192, + 336, + 200 + ], + "spans": [ + { + "bbox": [ + 299, + 192, + 336, + 200 + ], + "type": "text", + "content": "Partial arch" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 383, + 128, + 449, + 189 + ], + "blocks": [ + { + "bbox": [ + 397, + 118, + 435, + 127 + ], + "lines": [ + { + "bbox": [ + 397, + 118, + 435, + 127 + ], + "spans": [ + { + "bbox": [ + 397, + 118, + 435, + 127 + ], + "type": "text", + "content": "R3DS (ours)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 383, + 128, + 449, + 189 + ], + "lines": [ + { + "bbox": [ + 383, + 128, + 449, + 189 + ], + "spans": [ + { + "bbox": [ + 383, + 128, + 449, + 189 + ], + "type": "image", + "image_path": "65c45845923779202792ad615589436d7ef23755a7f9ae77441cc5b0ace51604.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 192, + 458, + 201 + ], + "lines": [ + { + "bbox": [ + 370, + 192, + 458, + 201 + ], + "spans": [ + { + "bbox": [ + 370, + 192, + 458, + 201 + ], + "type": "text", + "content": "Complete arch with portals" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 277, + 480, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 480, + 312 + ], + "type": "text", + "content": "over 110 fine-grained object categories. Table 1 shows a comparison of overall statistics with previous 3D indoor scene datasets. See the supplement for more statistics about annotated objects." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 314, + 481, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 481, + 495 + ], + "type": "text", + "content": "Compared to prior datasets that align CAD models to real-world scenes, R3DS is more complete, providing annotated object support hierarchies and matching object instances. CAD-Estate [14] annotates RGB videos with 3D objects and architectural room layouts. However, the architecture is partial as the videos have limited view (Fig. 4), and not all objects in the scenes are annotated (Fig. 2). This results in annotations of objects floating mid-air and not properly supported (e.g. lamp Fig. 2). Scan2CAD [1] also lacks support structure (e.g., lamp not supported by cabinet in Fig. 2). In addition, because Scan2CAD does not provide clean 3D architecture on which objects are placed (Fig. 4), objects on the floor are not always placed such that their bottom face is parallel with a horizontal plane. In contrast, our R3DS scenes have an accurate support hierarchy by construction. OpenRooms [12] augments Scan2CAD with room layouts representing the architecture. However, the architecture in R3DS is more complex and realistic, especially due to inclusion of more doors (1.92 doors per room in R3DS vs 0.67 in OpenRooms)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 496, + 480, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 480, + 567 + ], + "type": "text", + "content": "Evaluation of CAD object annotation quality is non-trivial as the 'ground truth' from the semantically annotated 3D reconstructions is itself imperfect. We measured how closely our annotated CAD objects conform to the real objects using the average 2D IoU between CAD object mask and ground truth 2D mask. R3DS is at " + }, + { + "bbox": [ + 130, + 496, + 480, + 567 + ], + "type": "inline_equation", + "content": "42.6\\%" + }, + { + "bbox": [ + 130, + 496, + 480, + 567 + ], + "type": "text", + "content": " vs " + }, + { + "bbox": [ + 130, + 496, + 480, + 567 + ], + "type": "inline_equation", + "content": "38.5\\%" + }, + { + "bbox": [ + 130, + 496, + 480, + 567 + ], + "type": "text", + "content": " for Scan2CAD, across 8 common object categories (bed, sofa, chair, cabinet, tv/monitor, table, shelving, bathtub)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": "Of the datasets previously used for Panoramic Scene Understanding, ReplicaPano [7] has not been released, and iGibson-DPC [28] is the only dataset with synthetic panoramic images annotated with 3D objects and room layout. iGibson-DPC is built on scenes from iGibson [18] by randomly replacing objects with different models from the same category and rendering using the iGibson simulator to render panoramas. The selection and placement of objects in iGibson-DPC is based on heuristic algorithms, while our R3DS is manually annotated and placed 3D models are verified in terms of match and alignment to the object masks." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "text", + "content": "R3DS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": "Moreover, iGibson-DPC contains unrealistic object arrangements (e.g., floating TV in Fig. 2)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 159, + 229, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 159, + 229, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 159, + 229, + 172 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 485, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 485, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 485, + 268 + ], + "type": "text", + "content": "We showcase the value of R3DS on the Panoramic Scene Understanding (PanoSun) task [7, 28, 29]. Given an input RGB panorama, the goal is to estimate the room layout, detect objects in 2D, estimate their 3D oriented bounding boxes and also reconstruct 3D object meshes. Our experiments show that methods trained on R3DS data benefit from its realism and generalize better when evaluated on photorealistic images. We also investigate the role of object support hierarchy information in improving performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 286, + 216, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 286, + 216, + 298 + ], + "spans": [ + { + "bbox": [ + 132, + 286, + 216, + 298 + ], + "type": "text", + "content": "4.1 Task setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 306, + 481, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 306, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 306, + 481, + 390 + ], + "type": "text", + "content": "Method. DeepPanoContext (DPC) [28] predicts the room layout, detects objects in 3D and recovers object meshes from a panorama image using a relation-based graph convolutional network and a differentiable relation optimization procedure. Since DPC has a publicly-available implementation, we use it to benchmark the R3DS data on the PanoSun task. We keep all hyperparameters unchanged except lowering the relation optimization loss weight of 3D bounding box back-projection from 10 to 1, since the ground truth 2D masks are noisy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 390, + 481, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 390, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 390, + 481, + 474 + ], + "type": "text", + "content": "Datasets. We train and evaluate DPC on the iGibson-DPC (IG) [7, 18, 28], Structured3D (S3D) [30], and R3DS datasets. Zhang et al. [28] render 1,500 panoramas from 15 iGibson houses composed of " + }, + { + "bbox": [ + 132, + 390, + 481, + 474 + ], + "type": "inline_equation", + "content": "500+" + }, + { + "bbox": [ + 132, + 390, + 481, + 474 + ], + "type": "text", + "content": " objects spanning 57 object categories. We use the same data and splits for IG. Structured3D consists of 3500 houses and around 18K photo-realistic rendered panoramas in total. We use 14K for training and the remaining 4K for testing. Note that Structured3D does not provide ground truth object meshes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 474, + 481, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 474, + 481, + 617 + ], + "spans": [ + { + "bbox": [ + 132, + 474, + 481, + 617 + ], + "type": "text", + "content": "To prepare R3DS for this task, we generate the ground truth room layout from the 3D architecture based on the camera viewpoint and obtain 3D oriented bounding boxes (OBBs) from all objects. We use 2D object masks from the Matterport3D mesh instance segmentation. We consider three variants of R3DS based on the input panorama: " + }, + { + "bbox": [ + 132, + 474, + 481, + 617 + ], + "type": "inline_equation", + "content": "R3DS" + }, + { + "bbox": [ + 132, + 474, + 481, + 617 + ], + "type": "text", + "content": "-real where we use the Matterport3D panoramas, " + }, + { + "bbox": [ + 132, + 474, + 481, + 617 + ], + "type": "inline_equation", + "content": "R3DS" + }, + { + "bbox": [ + 132, + 474, + 481, + 617 + ], + "type": "text", + "content": "-syn where we use rendered panoramas (at the same camera poses) from the annotated synthetic scenes, and " + }, + { + "bbox": [ + 132, + 474, + 481, + 617 + ], + "type": "inline_equation", + "content": "R3DS" + }, + { + "bbox": [ + 132, + 474, + 481, + 617 + ], + "type": "text", + "content": "-mix where we combine the two types of panoramas and double the available data. We follow the MP3D house split and merge the train and val sets to obtain a disjoint split of 15 train and 5 test houses. Based on the split, we have 696 annotated panoramas for train and 146 for test. To fairly evaluate methods trained on different datasets, we curate a list of 25 object classes common to all datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 481, + 666 + ], + "type": "text", + "content": "Metrics. Following Zhang et al. [28] we use separate metrics for room layout estimation, 3D object detection, and scene relation prediction. For room layout estimation, we use 2D IoU for predicted 2D floorplan, 3D IoU for lifted 3D room geometry, and dRMSE for predicted depth with respect to the camera location." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 220, + 157, + 388, + 226 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 147 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 147 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 147 + ], + "type": "text", + "content": "Table 2: Room layout estimation on R3DS-real test set. DPC [28] was pretrained on IG and S3D. For the last three rows, we fine-tune the pretrained weights on variants of R3DS." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 220, + 157, + 388, + 226 + ], + "lines": [ + { + "bbox": [ + 220, + 157, + 388, + 226 + ], + "spans": [ + { + "bbox": [ + 220, + 157, + 388, + 226 + ], + "type": "table", + "html": "
Train2D IoU ↑3D IoU ↑dRMSE ↓
DPC [28]53.450.30.682
R3DS-real55.153.10.610
R3DS-syn59.056.10.629
R3DS-mix59.657.00.572
", + "image_path": "99614648e7ed69f5fba01d7e83d9fbc1193841977a68d0094089d3a0e632af20.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 175, + 294, + 433, + 509 + ], + "blocks": [ + { + "bbox": [ + 132, + 250, + 482, + 283 + ], + "lines": [ + { + "bbox": [ + 132, + 250, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 132, + 250, + 482, + 283 + ], + "type": "text", + "content": "Table 3: Cross-dataset evaluation for the Panoramic Scene Understanding task. We evaluate 3D detections with class-agnostic IoU and mAP at IoU of 0.15, and report object collisions. The highlighted rows indicate the most challenging scenario." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 175, + 294, + 433, + 509 + ], + "lines": [ + { + "bbox": [ + 175, + 294, + 433, + 509 + ], + "spans": [ + { + "bbox": [ + 175, + 294, + 433, + 509 + ], + "type": "table", + "html": "
TestTrain3D detection ↑Collision ↓Attachment F1 ↑
IoUmAPmesharchobjwallfloor
IG27.530.31.6622.59453.176.895.086.2
IG+R3DS24.030.21.4042.25459.764.194.62.7
R3DS-real17.313.40.2421.45638.864.092.80.0
R3DS-syn23.214.20.4801.93848.546.793.828.6
R3DS-mix21.615.60.4341.24843.167.290.19.8
S3D19.53.51.0162.65150.968.790.811.6
IG+R3DS19.77.00.8682.08952.067.491.21.8
R3DS-real18.47.10.6002.59845.061.689.70.7
R3DS-syn19.04.80.6442.56149.349.791.22.4
R3DS-mix19.67.50.4631.67347.864.187.20.9
R3DS15.65.90.5751.95953.850.751.20.0
IG+R3DS17.514.10.2811.26749.561.658.60.0
R3DS-real16.415.00.2261.56244.057.358.90.0
R3DS-syn14.08.40.3901.66454.140.649.10.0
R3DS-mix17.615.80.1711.00748.558.360.10.0
", + "image_path": "e367b9c8d5c3f7ae81f5fe74b32202629c607a0a0927727d629cfcd24b9350c9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "text", + "content": "For 3D object detection, we report bounding box-based class-agnostic 3D IoU as well as mean average precision (mAP) across the 25 object classes, where an IoU greater than 0.15 counts as a \"true\" result. For scene relation prediction, we report F1 scores for relation classification. We also report the average number of objects colliding with each other or with architectural structures (wall, floor, ceiling). Specifically, we follow Zhang et al. [28] and measure collisions using the Separating Axis Theorem (SAT) to test whether object bounding boxes overlap. Since bounding box-based collision is a poor proxy for real-world physical collision, we also compute mesh-based collision by checking if the meshes for object pairs have any interpenetrating triangles [11, 24]." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "text", + "content": "R3DS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 117, + 482, + 333 + ], + "blocks": [ + { + "bbox": [ + 133, + 117, + 482, + 333 + ], + "lines": [ + { + "bbox": [ + 133, + 117, + 482, + 333 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 482, + 333 + ], + "type": "image", + "image_path": "6d998a1d6b58433b80cea888a734c8aa9d4097684a20ec35bf88c28f72426518.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 340, + 482, + 396 + ], + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 396 + ], + "type": "text", + "content": "Fig. 5: Qualitative results for cross-dataset Panoramic Scene Understanding task. Correct and incorrect object detections shown in green and red boxes. Ground truth room layout and meshes shown in gray color, while layout prediction is in yellow. Training on R3DS leads to fewer errors compared to training on other datasets, especially when mixed with real data." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 422, + 198, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 422, + 198, + 433 + ], + "spans": [ + { + "bbox": [ + 132, + 422, + 198, + 433 + ], + "type": "text", + "content": "4.2 Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 448, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 509 + ], + "type": "text", + "content": "1) Does R3DS help DPC generalize to real images? Since the original DPC work only trained and evaluated on synthetic data, it is unclear how well it performs on realistic panoramic imagery. We hypothesize that training on R3DS will lead to better performance. We separately show results on room layout estimation (Tab. 2) and 3D object detection." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": "Room layout estimation. For room layout estimation, DPC uses Horizon-Net [21] pretrained on iGibson (IG) and Structured3D (S3D) panoramas. This model achieves good performance on IG data (91.0 3D IoU). When directly testing the official pretrained model on " + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "inline_equation", + "content": "R3DS" + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": "-real panoramas, we notice a significant performance drop compared to results on the rendered panoramas from iGibson (Tab. 2 shows that the 3D IoU drops to 50.3). By finetuning the pretrained model with " + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "inline_equation", + "content": "R3DS" + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": "-real, we can predict more precise room layouts for real cluttered scenes. Even only trained on " + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "inline_equation", + "content": "R3DS" + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": "-syn, we outperform the original DPC model by " + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "inline_equation", + "content": "5.6\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "inline_equation", + "content": "5.8\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": " on 2D and 3D IoU, respectively. This is likely due to renderings from " + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "inline_equation", + "content": "R3DS" + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": "-syn reflecting more realistic object arrangements in a room instead of pushing all objects against walls. Best performance on 2D IoU (59.6), 3D IoU (57.0) and depth RMSE (0.572) is achieved by fine tuning on " + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "inline_equation", + "content": "R3DS" + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": "-mix." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 201, + 157, + 406, + 232 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 4: Performance of models trained on three synthetic datasets (IG, S3D, and IG+S3D) evaluated on the R3DS-full dataset, where \"full\" indicates all 840 panoramicas are used for testing." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 201, + 157, + 406, + 232 + ], + "lines": [ + { + "bbox": [ + 201, + 157, + 406, + 232 + ], + "spans": [ + { + "bbox": [ + 201, + 157, + 406, + 232 + ], + "type": "table", + "html": "
Train3D detectionCollision ↓Support F1 ↑
IoU↑mAP↑mesharchobj floorceil
IG14.25.20.7031.6394.185.3
S3D16.410.00.1121.2263.186.8
IG+S3D17.19.70.1331.1623.384.8
", + "image_path": "b7b45d83d5ac5a112aedff829bd2408f7e134a9f579c5029115dc4efe6af2a4a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 255, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 482, + 445 + ], + "type": "text", + "content": "Object detection. For 3D object detection, we train DPC on different data settings and conduct a cross-dataset evaluation (see Tab. 3). To investigate how models perform on out-of-distribution scenes, we evaluate models on Structured3D, as its images are near-realistic. To explore whether DPC training benefits from R3DS given the same amount of data, we create a special data input " + }, + { + "bbox": [ + 130, + 255, + 482, + 445 + ], + "type": "inline_equation", + "content": "IG + R3DS" + }, + { + "bbox": [ + 130, + 255, + 482, + 445 + ], + "type": "text", + "content": " that combines iGibson and R3DS panoramas by randomly replacing half (500) of iGibson data with " + }, + { + "bbox": [ + 130, + 255, + 482, + 445 + ], + "type": "inline_equation", + "content": "R3DS" + }, + { + "bbox": [ + 130, + 255, + 482, + 445 + ], + "type": "text", + "content": "-real data. The results show that " + }, + { + "bbox": [ + 130, + 255, + 482, + 445 + ], + "type": "inline_equation", + "content": "IG + R3DS" + }, + { + "bbox": [ + 130, + 255, + 482, + 445 + ], + "type": "text", + "content": " performs almost the same as " + }, + { + "bbox": [ + 130, + 255, + 482, + 445 + ], + "type": "inline_equation", + "content": "IG" + }, + { + "bbox": [ + 130, + 255, + 482, + 445 + ], + "type": "text", + "content": " with fewer collisions on iGibson, but it remarkably outperforms " + }, + { + "bbox": [ + 130, + 255, + 482, + 445 + ], + "type": "inline_equation", + "content": "IG" + }, + { + "bbox": [ + 130, + 255, + 482, + 445 + ], + "type": "text", + "content": " on the test set of " + }, + { + "bbox": [ + 130, + 255, + 482, + 445 + ], + "type": "inline_equation", + "content": "R3DS" + }, + { + "bbox": [ + 130, + 255, + 482, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 255, + 482, + 445 + ], + "type": "inline_equation", + "content": "S3D" + }, + { + "bbox": [ + 130, + 255, + 482, + 445 + ], + "type": "text", + "content": " by 8.2 and 3.5 improvements on 3D mAP, respectively. It also averages 0.221 fewer mesh collisions. There are noticeable performance gaps on iGibson for models trained on R3DS data likely due to the data domain shift. Among the three variants of R3DS data, " + }, + { + "bbox": [ + 130, + 255, + 482, + 445 + ], + "type": "inline_equation", + "content": "R3DS" + }, + { + "bbox": [ + 130, + 255, + 482, + 445 + ], + "type": "text", + "content": "-mix outperforms the others on all three test sets regarding 3D IoU and 3D mAP with the fewest mesh and architecture collisions. Although " + }, + { + "bbox": [ + 130, + 255, + 482, + 445 + ], + "type": "inline_equation", + "content": "R3DS" + }, + { + "bbox": [ + 130, + 255, + 482, + 445 + ], + "type": "text", + "content": "-syn underperforms on " + }, + { + "bbox": [ + 130, + 255, + 482, + 445 + ], + "type": "inline_equation", + "content": "R3DS" + }, + { + "bbox": [ + 130, + 255, + 482, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 255, + 482, + 445 + ], + "type": "inline_equation", + "content": "S3D" + }, + { + "bbox": [ + 130, + 255, + 482, + 445 + ], + "type": "text", + "content": " test sets, it achieves better performance than " + }, + { + "bbox": [ + 130, + 255, + 482, + 445 + ], + "type": "inline_equation", + "content": "IG" + }, + { + "bbox": [ + 130, + 255, + 482, + 445 + ], + "type": "text", + "content": " with even less data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 449, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 555 + ], + "type": "text", + "content": "Scene relation classification. We report F1 scores for identifying attachment relationships of objects to other objects and architecture elements (see Tab. 3). We note that models trained with synthetic renderings perform better than those trained on real images. That is because synthetic renderings present cleaner and simpler scenes with fewer objects than real world and simpler illumination such that DPC finds it easier to learn object-object and object-architecture relations. Also, note that the predictions of object-ceiling attachments can be extremely low because few objects are attached to the ceiling in the ground truth data. We show qualitative examples in Figure 5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "content": "2) Is R3DS a challenging, high-quality test set? How would a model trained on pure synthetic data perform on complex real data (R3DS)? Due to its modest scale, we propose using R3DS as a challenging, high-quality test set rather than a train set. Specifically, we evaluate the synthetic-to-real performance of DPC by training on iGibson and/or Structured3D and testing on all panoramas in R3DS-real. Table 4 shows that a model trained with Structured3D performs the best (10.0 3D mAP and 0.112 mesh collision) as it observes the most photo-realistic images. DPC benefits from the synthetic data for higher bounding box IoUs, since it possesses accurately aligned 3D bounding box and" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "text", + "content": "R3DS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 219, + 163 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 219, + 163 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 219, + 163 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 219, + 163 + ], + "type": "image", + "image_path": "9b5bba0ece1b1c2ef146afa2e3c9b4cdb4f8fed4eb9c9cfa099902f16bbe2847.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 141, + 163, + 214, + 202 + ], + "blocks": [ + { + "bbox": [ + 141, + 163, + 214, + 202 + ], + "lines": [ + { + "bbox": [ + 141, + 163, + 214, + 202 + ], + "spans": [ + { + "bbox": [ + 141, + 163, + 214, + 202 + ], + "type": "image", + "image_path": "a9664d032672092e7f57a467f56a7953fff909237f7ec9ddf77b9985f25df455.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 203, + 208, + 209 + ], + "lines": [ + { + "bbox": [ + 148, + 203, + 208, + 209 + ], + "spans": [ + { + "bbox": [ + 148, + 203, + 208, + 209 + ], + "type": "text", + "content": "(a) Object appearance gap " + }, + { + "bbox": [ + 148, + 203, + 208, + 209 + ], + "type": "inline_equation", + "content": "(60\\%)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 219, + 114, + 305, + 163 + ], + "blocks": [ + { + "bbox": [ + 219, + 114, + 305, + 163 + ], + "lines": [ + { + "bbox": [ + 219, + 114, + 305, + 163 + ], + "spans": [ + { + "bbox": [ + 219, + 114, + 305, + 163 + ], + "type": "image", + "image_path": "af2b5338e6663aae7efe41d48e274464aa8dad559ebd929a77d08be854d0fd42.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 219, + 163, + 305, + 202 + ], + "blocks": [ + { + "bbox": [ + 219, + 163, + 305, + 202 + ], + "lines": [ + { + "bbox": [ + 219, + 163, + 305, + 202 + ], + "spans": [ + { + "bbox": [ + 219, + 163, + 305, + 202 + ], + "type": "image", + "image_path": "ac9a00bfd5a6f7027d94e75047d7b78b48be61cfa2c1629c5575d53bfd9ebf0f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 236, + 203, + 294, + 209 + ], + "lines": [ + { + "bbox": [ + 236, + 203, + 294, + 209 + ], + "spans": [ + { + "bbox": [ + 236, + 203, + 294, + 209 + ], + "type": "text", + "content": "(b) Common occlusions (76.7%)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 114, + 391, + 163 + ], + "blocks": [ + { + "bbox": [ + 306, + 114, + 391, + 163 + ], + "lines": [ + { + "bbox": [ + 306, + 114, + 391, + 163 + ], + "spans": [ + { + "bbox": [ + 306, + 114, + 391, + 163 + ], + "type": "image", + "image_path": "f1209d65168c46fc11e3a01cbe261163def8ebc142c1ec0b0c8bf2c89f9c1e36.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 163, + 392, + 201 + ], + "blocks": [ + { + "bbox": [ + 309, + 163, + 392, + 201 + ], + "lines": [ + { + "bbox": [ + 309, + 163, + 392, + 201 + ], + "spans": [ + { + "bbox": [ + 309, + 163, + 392, + 201 + ], + "type": "image", + "image_path": "b88b0680aeb84e234cf6f8400cbda72a61ff6e309f4ee3d902175d61346a7730.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 203, + 381, + 209 + ], + "lines": [ + { + "bbox": [ + 326, + 203, + 381, + 209 + ], + "spans": [ + { + "bbox": [ + 326, + 203, + 381, + 209 + ], + "type": "text", + "content": "(c) Wrong 3D prediction " + }, + { + "bbox": [ + 326, + 203, + 381, + 209 + ], + "type": "inline_equation", + "content": "(65\\%)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 392, + 114, + 478, + 163 + ], + "blocks": [ + { + "bbox": [ + 392, + 114, + 478, + 163 + ], + "lines": [ + { + "bbox": [ + 392, + 114, + 478, + 163 + ], + "spans": [ + { + "bbox": [ + 392, + 114, + 478, + 163 + ], + "type": "image", + "image_path": "5bcc3a638f5c4300acc3018e8a9f30f76652bf422af6081d25b382bbbc681e14.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 399, + 163, + 477, + 202 + ], + "blocks": [ + { + "bbox": [ + 399, + 163, + 477, + 202 + ], + "lines": [ + { + "bbox": [ + 399, + 163, + 477, + 202 + ], + "spans": [ + { + "bbox": [ + 399, + 163, + 477, + 202 + ], + "type": "image", + "image_path": "826ed6e7d75414a4cdda1ec61f1d1e3d65f913f855118a62fad8dd3c9fe3f865.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 203, + 467, + 209 + ], + "lines": [ + { + "bbox": [ + 413, + 203, + 467, + 209 + ], + "spans": [ + { + "bbox": [ + 413, + 203, + 467, + 209 + ], + "type": "text", + "content": "(d) Objects in mirrors (15.6%)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 220, + 481, + 277 + ], + "lines": [ + { + "bbox": [ + 131, + 220, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 131, + 220, + 481, + 277 + ], + "type": "text", + "content": "Fig. 6: Error analysis on R3DS panoramas using DPC pretrained on S3D. Typical prediction errors are categorized into 4 groups: (a) Appearance gap between real and synthetic images. (b) Detection failure due to common occlusions. (c) Bad performance of monocular 3D prediction. (d) Perception error for objects in mirrors. Numbers in parentheses mean the error rates." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 133, + 300, + 247, + 358 + ], + "blocks": [ + { + "bbox": [ + 175, + 292, + 204, + 299 + ], + "lines": [ + { + "bbox": [ + 175, + 292, + 204, + 299 + ], + "spans": [ + { + "bbox": [ + 175, + 292, + 204, + 299 + ], + "type": "text", + "content": "Panorama" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 300, + 247, + 358 + ], + "lines": [ + { + "bbox": [ + 133, + 300, + 247, + 358 + ], + "spans": [ + { + "bbox": [ + 133, + 300, + 247, + 358 + ], + "type": "image", + "image_path": "eda72c7838d71b3a9d0a4a0e3344f8a6eb1a8667be8c2df1715077c63d2fc584.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 133, + 360, + 247, + 416 + ], + "blocks": [ + { + "bbox": [ + 133, + 360, + 247, + 416 + ], + "lines": [ + { + "bbox": [ + 133, + 360, + 247, + 416 + ], + "spans": [ + { + "bbox": [ + 133, + 360, + 247, + 416 + ], + "type": "image", + "image_path": "55e6b6e47c630c4f3c949b1d21ab86b15d074c3d90ec835db3e70d803eb0bc44.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 425, + 482, + 460 + ], + "lines": [ + { + "bbox": [ + 131, + 425, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 131, + 425, + 482, + 460 + ], + "type": "text", + "content": "Fig. 7: Comparison between heuristic and R3DS support relations. Dark lines between objects show support relations. Heuristic support often mistakenly assigns relations to spatially close objects. In contrast, R3DS support relations are annotated." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 250, + 300, + 364, + 357 + ], + "blocks": [ + { + "bbox": [ + 279, + 293, + 328, + 300 + ], + "lines": [ + { + "bbox": [ + 279, + 293, + 328, + 300 + ], + "spans": [ + { + "bbox": [ + 279, + 293, + 328, + 300 + ], + "type": "text", + "content": "Heuristic Support" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 250, + 300, + 364, + 357 + ], + "lines": [ + { + "bbox": [ + 250, + 300, + 364, + 357 + ], + "spans": [ + { + "bbox": [ + 250, + 300, + 364, + 357 + ], + "type": "image", + "image_path": "3d6d39fd0b8998887e4edcb3f56eb6972c290eae77ff140de3f5e26c0e05cd56.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 250, + 360, + 364, + 416 + ], + "blocks": [ + { + "bbox": [ + 250, + 360, + 364, + 416 + ], + "lines": [ + { + "bbox": [ + 250, + 360, + 364, + 416 + ], + "spans": [ + { + "bbox": [ + 250, + 360, + 364, + 416 + ], + "type": "image", + "image_path": "f35887575e5067bf1ce1c6f326266669a49d3d36e7fce4fa11abf4e3de870dca.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 367, + 300, + 480, + 358 + ], + "blocks": [ + { + "bbox": [ + 402, + 293, + 441, + 300 + ], + "lines": [ + { + "bbox": [ + 402, + 293, + 441, + 300 + ], + "spans": [ + { + "bbox": [ + 402, + 293, + 441, + 300 + ], + "type": "text", + "content": "R3DS Support" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 367, + 300, + 480, + 358 + ], + "lines": [ + { + "bbox": [ + 367, + 300, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 367, + 300, + 480, + 358 + ], + "type": "image", + "image_path": "921e3c29ca87d8b788bf8410ff2af6d632dc69be794505a03a6ea0acb3543302.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 367, + 360, + 480, + 416 + ], + "blocks": [ + { + "bbox": [ + 367, + 360, + 480, + 416 + ], + "lines": [ + { + "bbox": [ + 367, + 360, + 480, + 416 + ], + "spans": [ + { + "bbox": [ + 367, + 360, + 480, + 416 + ], + "type": "image", + "image_path": "606970567abdc77fe6603563009f963acaa8200aff9c26db499add1c4c2bfc57.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 130, + 484, + 480, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 480, + 520 + ], + "type": "text", + "content": "more unoccluded objects. However, mAP performance is lower due to worse object recognition ability. All models struggle to predict correct object-wise support relations but do a better job of predicting object-floor support relations." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 130, + 521, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 605 + ], + "type": "text", + "content": "We conduct error analysis on 120 randomly sampled panoramas using the model pretrained on " + }, + { + "bbox": [ + 130, + 521, + 481, + 605 + ], + "type": "inline_equation", + "content": "S3D" + }, + { + "bbox": [ + 130, + 521, + 481, + 605 + ], + "type": "text", + "content": " to identify typical errors (see Fig. 6). Errors are categorized into 4 groups: (a) " + }, + { + "bbox": [ + 130, + 521, + 481, + 605 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 130, + 521, + 481, + 605 + ], + "type": "text", + "content": " panoramas have 2D perception errors due to the synthetic-to-real appearance gap; (b) " + }, + { + "bbox": [ + 130, + 521, + 481, + 605 + ], + "type": "inline_equation", + "content": "76.7\\%" + }, + { + "bbox": [ + 130, + 521, + 481, + 605 + ], + "type": "text", + "content": " panoramas show detection failures due to occlusions; (c) " + }, + { + "bbox": [ + 130, + 521, + 481, + 605 + ], + "type": "inline_equation", + "content": "65\\%" + }, + { + "bbox": [ + 130, + 521, + 481, + 605 + ], + "type": "text", + "content": " panoramas exhibit correct 2D detections but fail to correctly perform 3D predictions; and (d) " + }, + { + "bbox": [ + 130, + 521, + 481, + 605 + ], + "type": "inline_equation", + "content": "15.6\\%" + }, + { + "bbox": [ + 130, + 521, + 481, + 605 + ], + "type": "text", + "content": " out of 45 panoramas with mirrors mistakenly predict virtual objects in mirrors." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "3) Are R3DS support relations helpful for PanoSun? We investigate whether the support relationships between objects provided in our R3DS scene hierarchy help boost performance of holistic scene understanding. We augment DPC's Relation Scene-GCN module with additional support relation prediction branches. Besides obtaining explicitly annotated scene support relations from" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 181, + 178, + 427, + 329 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 5: Performance on R3DS-real of DPC models trained on variants of R3DS with different support relation settings. We compare the original model without support (none) against models supervised with support that is heuristically computed (heur) or annotated from R3DS (anno). Classification results are evaluated on annotated ground-truth scene hierarchy." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 181, + 178, + 427, + 329 + ], + "lines": [ + { + "bbox": [ + 181, + 178, + 427, + 329 + ], + "spans": [ + { + "bbox": [ + 181, + 178, + 427, + 329 + ], + "type": "table", + "html": "
TrainSupp.3D detectionCollision ↓Support F1 ↑
IoU↑mAP↑mesharchobj floorceil
R3DS-realnone16.415.00.2261.562--
heur16.214.10.2191.3293.269.9
anno16.615.90.3491.40438.594.4
R3DS-synnone14.08.40.3901.664--
heur14.67.80.2811.3014.682.9
anno14.38.20.3491.21932.095.0
R3DS-mixnone17.615.80.1711.007--
heur19.217.70.1511.2673.683.7
anno18.618.20.1581.30812.096.2
", + "image_path": "2ab00e23b3e3b222fac516f51ca6d27fb174e0b59a778ffbd66c87f06b4b6d5e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 183, + 413, + 425, + 477 + ], + "blocks": [ + { + "bbox": [ + 130, + 349, + 482, + 405 + ], + "lines": [ + { + "bbox": [ + 130, + 349, + 482, + 405 + ], + "spans": [ + { + "bbox": [ + 130, + 349, + 482, + 405 + ], + "type": "text", + "content": "Table 6: Comparison of the average number of bounding box and mesh-based object collisions per scene in IG and R3DS. R3DS exhibits more bounding box-based collisions, but almost none of these are actual physical collisions between object meshes. Measuring collisions between bounding boxes is a poor collision measure for fully-populated, real-world scenes." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 183, + 413, + 425, + 477 + ], + "lines": [ + { + "bbox": [ + 183, + 413, + 425, + 477 + ], + "spans": [ + { + "bbox": [ + 183, + 413, + 425, + 477 + ], + "type": "table", + "html": "
DatasetsMeshCollisionsBox Collisions
obj-objobj-wallobj-floorobj-ceil
IG-1.1850.0750.0000.790
R3DS0.00062.8230.3880.0350.064
", + "image_path": "b9cf6b6f1cd5d5aa7a57758e4d93fbff55efdf05ccd935dd32489d36e8987aa1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": "R3DS, it is also possible to compute heuristic support relations from object bounding boxes. Specifically, an object is supported by another if their bounding boxes intersect within tolerance distance of " + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "inline_equation", + "content": "0.1\\mathrm{m}" + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": " and the centroid of the former object is higher than that of the latter. Support by wall/floor/ceiling is calculated in the same way without the height judgment. This definition is similar to how DPC defines object attachment. Figure 7 compares these two ways of computing support relations, showing that heuristic computation can mistakenly designate support relations to two nearby objects. Table 5 shows that incorporating support relation prediction indeed influences the performance of DPC. Heuristic support information may worsen 3D object detection (mAP in R3DS-real and R3DS-syn), but it eliminates mesh collisions the most. Learning support relations from R3DS annotations leads to a 2.4 improvement on mAP in R3DS-mix, although the classification F1 score is low." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "text", + "content": "R3DS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 174, + 157, + 433, + 243 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 7: Ablation of relation optimization (RO) on R3DS-real. The 2nd and 3rd row remove optimization terms in RO. The last row replaces bounding-box collisions with mesh collisions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 174, + 157, + 433, + 243 + ], + "lines": [ + { + "bbox": [ + 174, + 157, + 433, + 243 + ], + "spans": [ + { + "bbox": [ + 174, + 157, + 433, + 243 + ], + "type": "table", + "html": "
Rel. Opt.3D mAP ↑ MeshCollisions ↓Box Collisions ↓
obj-objobj-arch
DPC18.20.1580.0621.308
w/o obj col18.91.3421.1301.301
w/o obj col+tch19.61.2191.0621.295
w/ mesh col19.71.0270.8561.394
", + "image_path": "1e9e3c30bf3392d725d9e9e422459d60d5d67c79d2c0e9bbc9a66dd67f641b0f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 265, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 265, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 482, + 518 + ], + "type": "text", + "content": "4) Is relation optimization (RO) effective on R3DS? Test-time relation optimization (RO) was introduced by Zhang et al. [28] to reduce physical violations, floating objects, and misalignment between objects and architecture. The original cost function based on bounding box collisions succeeds in optimizing object poses, since there are few such collisions in the IG data originally used for evaluation (see in Tab. 6). However, the same data assumption does not hold for R3DS, which has more bounding-box-based collisions but nearly zero mesh-based collisions. We ablate the design of RO on R3DS-real in Table 7. By removing two optimization terms (bounding-box-based object-wise collision and touching step-by-step), the model outperforms the original one in 3D mAP (+1.4) but degrades in mesh-based and box-based collisions. We show that using mesh-based collision optimization leads to the best performance. The increase in collisions is unsurprising as the R3DS data reflects more cluttered real interiors. Limitations. Our dataset construction relied on 3D architectures for each Matterport3D scan which are simplifications of the geometry of the real environment. One issue is imperfect wall positions, resulting in objects attached to these virtual walls being offset from the true surface. In addition, objects in our 3D scenes were placed without regard to the materials, meaning that the detailed surface appearance does not match that of the observed object. Future work can investigate transfer of surface appearance to the synthetic objects by projecting textures from the RGB-D data and 3D reconstructed meshes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 540, + 220, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 540, + 220, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 540, + 220, + 553 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "We introduced the R3DS dataset. R3DS provides more complete, densely populated, and richly annotated synthetic 3D scene proxies of real-world environments with linked panoramic images. We showed the usefulness of R3DS on the Panoramic Scene Understanding task. Our experiments demonstrate the value of realistic synthetic recreations in this task, in particular through the use of object support information. While we focused on the PanoSun task, R3DS can also be useful for other tasks such as single-view shape retrieval, single-view object pose estimation, and panoramic scene graph prediction." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 261 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 261 + ], + "type": "text", + "content": "This work was funded in part by a CIFAR AI Chair, a Canada Research Chair, NSERC Discovery Grant, NSF award #2016532, and enabled by support from WestGrid and Compute Canada. Daniel Ritchie is an advisor to Geope and owns equity in the company. Geope is a start-up that is developing 3D technology to build immersive virtual copies of the real world with applications in various fields, including games and architecture. We thank Madhawa Vidanapathirana, Weijie Lin, and David Han for help with development of the annotation tool, and Denys Iliash, Mrinal Goshalia, Brandon Robles, Paul Brown, Chloe Ye, Coco Kaleel, Elizabeth Wu and Hannah Julius for data annotation, and Ivan Tam, Austin Wang, and Ning Wang for feedback on the paper draft." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 279, + 197, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 279, + 197, + 292 + ], + "spans": [ + { + "bbox": [ + 133, + 279, + 197, + 292 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 305, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 138, + 305, + 480, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 305, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 138, + 305, + 480, + 353 + ], + "type": "text", + "content": "1. Avetisyan, A., Dahnert, M., Dai, A., Savva, M., Chang, A.X., Nießner, M.: Scan2CAD: Learning CAD model alignment in RGB-D scans. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 354, + 480, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 480, + 400 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 480, + 400 + ], + "type": "text", + "content": "2. Batra, D., Chang, A.X., Chernova, S., Davison, A.J., Deng, J., Koltun, V., Levine, S., Malik, J., Mordatch, I., Mottaghi, R., Savva, M., Su, H.: Rearrangement: A challenge for embodied AI. arXiv preprint arXiv:2011.01975 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 402, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 402, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 138, + 402, + 480, + 449 + ], + "type": "text", + "content": "3. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niebner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3D: Learning from RGB-D data in indoor environments. In: Proceedings of the International Conference on 3D Vision (3DV). pp. 667-676. IEEE (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 450, + 480, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 450, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 138, + 450, + 480, + 485 + ], + "type": "text", + "content": "4. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 486, + 480, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 486, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 138, + 486, + 480, + 545 + ], + "type": "text", + "content": "5. Collins, J., Goel, S., Deng, K., Luthra, A., Xu, L., Gundogdu, E., Zhang, X., Vicente, T.F.Y., Dideriksen, T., Arora, H., et al.: Abo: Dataset and benchmarks for real-world 3d object understanding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21126-21136 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 545, + 480, + 593 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 593 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 593 + ], + "type": "text", + "content": "6. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5828-5839 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 594, + 480, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 594, + 480, + 628 + ], + "spans": [ + { + "bbox": [ + 138, + 594, + 480, + 628 + ], + "type": "text", + "content": "7. Dong, Y., Fang, C., Dong, Z., Bo, L., Tan, P.: PanoContext-Former: Panoramic total scene understanding with a transformer. arXiv preprint arXiv:2305.12497 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 629, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 629, + 480, + 665 + ], + "type": "text", + "content": "8. Fu, H., Cai, B., Gao, L., Zhang, L., Li, C., Xun, Z., Sun, C., Fei, Y., Zheng, Y., Li, Y., et al.: 3D-FRONT: 3D Furnished Rooms with layOuts and semaNTics. arXiv preprint arXiv:2011.09127 (2020)" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "text", + "content": "R3DS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 136, + 115, + 481, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 115, + 481, + 152 + ], + "spans": [ + { + "bbox": [ + 136, + 115, + 481, + 152 + ], + "type": "text", + "content": "9. Fu, H., Jia, R., Gao, L., Gong, M., Zhao, B., Maybank, S., Tao, D.: 3D-Future: 3D Furniture shape with TextURE. arXiv preprint arXiv:2009.09633 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 152, + 481, + 187 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 152, + 481, + 187 + ], + "spans": [ + { + "bbox": [ + 133, + 152, + 481, + 187 + ], + "type": "text", + "content": "10. Hua, B.S., Pham, Q.H., Nguyen, D.T., Tran, M.K., Yu, L.F., Yeung, S.K.: SceneNN: A scene meshes dataset with annotations. In: Proceedings of the International Conference on 3D Vision (3DV). pp. 92-101. IEEE (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 189, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 189, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 133, + 189, + 481, + 247 + ], + "type": "text", + "content": "11. Karras, T.: Maximizing parallelism in the construction of bvhs, octrees, and k-d trees. In: Proceedings of the Fourth ACM SIGGRAPH / Eurographics Conference on High-Performance Graphics. pp. 33-37. Eurographics Association (2012). https://doi.org/10.2312/EGGH/HPG12/033-037, https://doi.org/10.2312/EGGH/HPG12/033-037" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 247, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 481, + 319 + ], + "type": "text", + "content": "12. Li, Z., Yu, T.W., Sang, S., Wang, S., Song, M., Liu, Y., Yeh, Y.Y., Zhu, R., Gundavarapu, N., Shi, J., Bi, S., Xu, Z., Yu, H.X., Sunkavalli, K., Hasan, M., Ramamoorthi, R., Chandraker, M.: OpenRooms: An end-to-end open framework for photorealistic indoor scene datasets. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 319, + 481, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 481, + 355 + ], + "type": "text", + "content": "13. Lim, J.J., Pirsiavash, H., Torralba, A.: Parsing intelia objects: Fine pose estimation. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 2992-2999 (2013)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 355, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 355, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 355, + 481, + 392 + ], + "type": "text", + "content": "14. Maninis, K.K., Popov, S., Nießner, M., Ferrari, V.: CAD-estate: Large-scale CAD model annotation in RGB videos. arXiv preprint arXiv:2306.09011 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 392, + 481, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 426 + ], + "type": "text", + "content": "15. Paschalidou, D., Kar, A., Shugrina, M., Kreis, K., Geiger, A., Fidler, S.: Atiss: Autoregressive transformers for indoor scene synthesis. Advances in Neural Information Processing Systems 34, 12013-12026 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 426, + 481, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 426, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 426, + 481, + 475 + ], + "type": "text", + "content": "16. Ramakrishnan, S.K., Gokaslan, A., Wijmans, E., Maksymets, O., Clegg, A., Turner, J., Undersander, E., Galuba, W., Westbury, A., Chang, A.X., et al.: Habitat-Matterport 3D dataset (hm3d): 1000 large-scale 3D environments for embodied AI. arXiv preprint arXiv:2109.08238 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 475, + 481, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 481, + 510 + ], + "type": "text", + "content": "17. Sadalgi, S.: Wayfair's 3D Model API. https://www/aboutwayfair.com/tech-innovation/wayfairs-3d-model-api (2016), [Online; accessed 15-Nov-2023]" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 510, + 481, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 481, + 570 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 481, + 570 + ], + "type": "text", + "content": "18. Shen, B., Xia, F., Li, C., Martin-Martin, R., Fan, L., Wang, G., Buch, S., D'Arpino, C., Srivastava, S., Tchapmi, L.P., Vainio, K., Fei-Fei, L., Savarese, S.: iGibson, a simulation environment for interactive tasks in large realistic scenes. In: Proceedings of the International Conference on Intelligent Robots and Systems (IROS) (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 570, + 481, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 570, + 481, + 630 + ], + "spans": [ + { + "bbox": [ + 132, + 570, + 481, + 630 + ], + "type": "text", + "content": "19. Shen, B., Xia, F., Li, C., Martin-Martin, R., Fan, L., Wang, G., Pérez-D'Arpino, C., Buch, S., Srivastava, S., Tchapmi, L., et al.: iGibson 1.0: A simulation environment for interactive tasks in large realistic scenes. In: Proceedings of the International Conference on Intelligent Robots and Systems (IROS). pp. 7520-7527. IEEE (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 630, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 630, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 630, + 481, + 666 + ], + "type": "text", + "content": "20. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 607 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 164 + ], + "type": "text", + "content": "21. Sun, C., Hsiao, C.W., Sun, M., Chen, H.T.: Horizonnet: Learning room layout with 1d representation and pano stretch data augmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1047-1056 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 165, + 482, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 165, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 165, + 482, + 212 + ], + "type": "text", + "content": "22. Sun, X., Wu, J., Zhang, X., Zhang, Z., Zhang, C., Xue, T., Tenenbaum, J.B., Freeman, W.T.: Pix3D: Dataset and methods for single-image 3D shape modeling. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2974-2983 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 213, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 259 + ], + "type": "text", + "content": "23. Szot, A., Clegg, A., Undersander, E., Wijmans, E., Zhao, Y., Turner, J., Maestre, N., Mukadam, M., Chaplot, D.S., Maksymets, O., et al.: Habitat 2.0: Training home assistants to rearrange their habitat. Advances in Neural Information Processing Systems 34, 251-266 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 259, + 482, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 308 + ], + "type": "text", + "content": "24. Tzionas, D., Ballan, L., Srikantha, A., Aponte, P., Pollefeys, M., Gall, J.: Capturing hands in action using discriminative salient points and physics simulation. International Journal of Computer Vision (IJCV) 118(2), 172-193 (Jun 2016), https://doi.org/10.1007/s11263-016-0895-4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 308, + 482, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 355 + ], + "type": "text", + "content": "25. Wang, K., Lin, Y.A., Weissmann, B., Savva, M., Chang, A.X., Ritchie, D.: Planit: Planning and instantiating indoor scenes with relation graph and spatial prior networks. ACM Transactions on Graphics (TOG) 38(4), 1-15 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 356, + 482, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 403 + ], + "type": "text", + "content": "26. Xiang, Y., Kim, W., Chen, W., Ji, J., Choy, C., Su, H., Mottaghi, R., Guibas, L., Savarese, S.: Objectnet3D: A large scale database for 3D object recognition. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 160-176. Springer (2016)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 403, + 482, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 450 + ], + "type": "text", + "content": "27. Yadav, K., Ramrakhya, R., Ramakrishnan, S.K., Gervet, T., Turner, J., Gokaslan, A., Maestre, N., Chang, A.X., Batra, D., Savva, M., et al.: Habitat-Matterport 3D semantics dataset. arXiv preprint arXiv:2210.05633 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 451, + 482, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 482, + 499 + ], + "type": "text", + "content": "28. Zhang, C., Cui, Z., Chen, C., Liu, S., Zeng, B., Bao, H., Zhang, Y.: Deep-PanoContext: Panoramic 3D scene understanding with holistic scene context graph and relation-based optimization. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 12632-12641 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 499, + 482, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 499, + 482, + 535 + ], + "spans": [ + { + "bbox": [ + 130, + 499, + 482, + 535 + ], + "type": "text", + "content": "29. Zhang, Y., Song, S., Tan, P., Xiao, J.: PanoContext: A whole-room 3D context model for panoramic scene understanding. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 668-686. Springer (2014)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 535, + 482, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 535, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 535, + 482, + 570 + ], + "type": "text", + "content": "30. Zheng, J., Zhang, J., Li, J., Tang, R., Gao, S., Zhou, Z.: Structured3D: A large photo-realistic dataset for structured 3D modeling. arXiv preprint arXiv:1908.00222 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 571, + 482, + 607 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 571, + 482, + 607 + ], + "spans": [ + { + "bbox": [ + 130, + 571, + 482, + 607 + ], + "type": "text", + "content": "31. Zhou, T., Tucker, R., Flynn, J., Fyffe, G., Snavely, N.: Stereo magnification: learning view synthesis using multiplane images. ACM Transactions on Graphics (TOG) 37(4), 1-12 (2018)" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 422, + 91, + 447, + 100 + ], + "type": "text", + "content": "R3DS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/RANRAC_ Robust Neural Scene Representations via Random Ray Consensus/de9512b1-9e5d-48ed-a48f-d5995f4754ec_content_list.json b/2024/RANRAC_ Robust Neural Scene Representations via Random Ray Consensus/de9512b1-9e5d-48ed-a48f-d5995f4754ec_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..933729a2dd5e981eb2d3197260436f4280c88a25 --- /dev/null +++ b/2024/RANRAC_ Robust Neural Scene Representations via Random Ray Consensus/de9512b1-9e5d-48ed-a48f-d5995f4754ec_content_list.json @@ -0,0 +1,1546 @@ +[ + { + "type": "text", + "text": "RANRAC: Robust Neural Scene Representations via Random Ray Consensus", + "text_level": 1, + "bbox": [ + 220, + 140, + 782, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Benno Buschmann1,2, Andreea Dogaru2, Elmar Eisemann1, Michael Weinmann1, and Bernhard Egger2", + "bbox": [ + 256, + 210, + 746, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Delft University of Technology, Netherlands", + "bbox": [ + 346, + 253, + 653, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Friedrich-Alexander-University Erlangen-Nürnberg, Germany", + "bbox": [ + 289, + 268, + 710, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Learning-based scene representations such as neural radiance fields or light field networks, that rely on fitting a scene model to image observations, commonly encounter challenges in the presence of inconsistencies within the images caused by occlusions, inaccurately estimated camera parameters or effects like lens flare. To address this challenge, we introduce RANdom RAy Consensus (RANRAC), an efficient approach to eliminate the effect of inconsistent data, thereby taking inspiration from classical RANSAC based outlier detection for model fitting. In contrast to the down-weighting of the effect of outliers based on robust loss formulations, our approach reliably detects and excludes inconsistent perspectives, resulting in clean images without floating artifacts. For this purpose, we formulate a fuzzy adaption of the RANSAC paradigm, enabling its application to large scale models. We interpret the minimal number of samples to determine the model parameters as a tunable hyperparameter, investigate the generation of hypotheses with data-driven models, and analyse the validation of hypotheses in noisy environments. We demonstrate the compatibility and potential of our solution for both photo-realistic robust multi-view reconstruction from real-world images based on neural radiance fields and for single-shot reconstruction based on light-field networks. In particular, the results indicate significant improvements compared to state-of-the-art robust methods for novel-view synthesis on both synthetic and captured scenes with various inconsistencies including occlusions, noisy camera pose estimates, and unfocused perspectives. The results further indicate significant improvements for single-shot reconstruction from occluded images.", + "bbox": [ + 261, + 319, + 738, + 665 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Neural scene representations $\\cdot$ neural rendering $\\cdot$ RANSAC $\\cdot$ robust estimation $\\cdot$ neural radiance fields $\\cdot$ light-field networks", + "bbox": [ + 261, + 679, + 738, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 732, + 374, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D reconstruction is a classical task in computer vision and computer graphics, which has attracted research for decades. It offers numerous applications, including autonomous systems, entertainment, design, advertisement, cultural heritage, VR/AR experiences or medical scenarios. In recent years, neural scene representations and rendering techniques [44,51], including light field networks", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2aee6ce2116d1ebbcb2915d1807117e82963fb404ad9dcd39d13887a5e328f6b.jpg", + "image_caption": [ + "Fig. 1: We propose a robust algorithm for 3D reconstruction from occluded input perspectives that is based on the random sampling of hypotheses. Our algorithm is general and we demonstrate the use for single-shot reconstruction using light field networks or multi-view reconstruction using NeRF. In these cases, it successfully removes the artifacts that normally occur due to occluded input perspectives." + ], + "image_footnote": [], + "bbox": [ + 303, + 143, + 700, + 233 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(LFN) [40] and neural radiance fields (NeRF) [29] have demonstrated great performance in single-view and multi-view reconstruction tasks. The key to the success of such techniques is the coupling of differentiable rendering methods with custom neural field parametrizations of scene properties. However, a common limitation of neural scene reconstruction methods is their sensitivity to inconsistencies among input images induced by occlusions, inaccurately estimated camera parameters or effects like lens flares. Despite the use of view-dependent radiance representations to address view-dependent appearance changes, these inconsistencies severely impact local density estimation, resulting in a poor generalization to novel views.", + "bbox": [ + 212, + 352, + 787, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To increase the robustness to potential distractors within the training data, Sabour et al. [38] recently introduced the use of robust losses in the context of training unconditioned NeRF, where distractors in the training data were modeled as outliers of an optimization problem. However, the adaptation of this approach to conditioned neural fields (e.g., pixelNeRF [52]) is not obvious, as no optimization takes place during inference, and the input data is constrained to only a few views. Achieving robustness to data inconsistencies is a well-analyzed problem in computer vision, covered not only by the aforementioned robust loss functions [1], but also other strategies, like the random sample consensus (RANSAC) paradigm [15]. The latter is widely employed for fitting models to outlier-heavy data. The underlying idea is to randomly select subsets of the data to form potential models, evaluating these models against the entire dataset, and identifying the subset that best fits the majority of the data, while disregarding outliers. Despite being the state-of-the-art solution to many challenges, RANSAC-based schemes are particularly favoured for the fitting of analytical models with a relatively small amount of tuneable parameters. In this paper, we direct our attention to achieving robustness against inconsistencies and occlusions in the observations by using a novel combination of neural scene representation and rendering techniques with dedicated outlier removal techniques such as RANSAC [15]. While downweighting the influence of distractors based on robust losses [1, 38] can affect clean samples, representing details, we aim at improving robustness to distractors by only removing the influence of", + "bbox": [ + 212, + 507, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "B. Buschmann et al.", + "bbox": [ + 271, + 114, + 410, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "outliers. To this extent, we integrate a RANSAC-based scheme to distinguish inliers and outliers in the data and the inlier-based optimization of the neural fields (Fig. 1); a stochastic scenario characterized by a large-scale, data-driven model that exceeds RANSAC's classical convergence expectations. Instead of guaranteeing convergence to a clean sample set based on a minimal number of samples, we aim for a feasible (cleaner) sample set using a tuneable amount of samples. The proposed algorithm exhibits robustness and versatility, accommodating a wide range of neural fields-based reconstruction methods.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method inherits the strengths of RANSAC, such as the ability to handle various classes of outliers without relying on semantics. Yet, it also inherits the need for sufficiently clean samples and the reliance on an iterative scheme. In practice, the first condition is often fulfilled because typically only some of the perspectives are affected by inconsistencies. We validate our approach using synthetic data, focusing on the task of multi-class single-shot reconstruction with LFNs [40], and observe significant quality improvements over the baseline in the presence of occlusions. Furthermore, we showcase robust photo-realistic reconstructions of 3D objects using unconditioned NeRFs from sequences of real-world images in the presence of distractors. In comparison to RobustNeRF [38], we use all available clean data, hence improving the reconstruction quality for single-object scenes. Code and data are available under https://bennobuschmann.com/ranrac/. Our key contributions are:", + "bbox": [ + 212, + 268, + 787, + 464 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- a general, robust RANSAC-based reconstruction method applicable to a variety of neural-fields and handling diverse inconsistencies", + "- an analysis of the implication to RANSAC's hyperparameters and theoretical convergence expectations, and the experimental study of their effect", + "- a method for robust photo-realistic object reconstruction using NeRF and for robust single-shot multi-class reconstruction using LFNs", + "- a qualitative/quantitative evaluation of our method on both synthetic and real-world data with different inconsistencies (occlusions, invalid calibrations, ...) indicating the state-of-the-art performance of our approach" + ], + "bbox": [ + 225, + 476, + 784, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 638, + 387, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Among the vast literature on neural fields, the seminal work of Mildenhall et al. [29] opened many avenues in the computer vision community. It contributed to state-of-the-art solutions for novel view synthesis and 3D reconstruction that have been covered in respective surveys [44, 51]. Noteworthy is the more recent contribution of instant neural graphics primitives (iNGP) [31], which uses a hash table of trainable feature vectors alongside a small network for representing the scene. iNGP achieved major run-time improvements, thereby enhancing the feasibility of practical applications for neural fields.", + "bbox": [ + 212, + 672, + 787, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Baseline models are highly sensitive to imperfections in the input data, which led to many works on robustness enhancements of neural fields; addressing a reduced amount of input views [17,21,32,52], errors in camera parameters [4,19,", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "RANRAC", + "bbox": [ + 661, + 113, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[25, 53], variations in illumination conditions across observations [28, 42], multiscale image data [2, 3, 50], and the targeted removal of floating artifacts [33, 47, 48].", + "bbox": [ + 215, + 145, + 784, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fewer works solve the reconstruction task in the presence of inconsistencies between observations. Bayes' Rays [16] provides a framework to quantify uncertainty of a pretrained NeRF by approximating a spatial uncertainty field. It handles missing information due to self-occlusion or missing perspectives well, but cannot deal with inconsistencies caused by noise or distractors. Similarly, NeuRay [27] only supports missing, but not inconsistent information. Naive occlusion handling via semantic segmentation requires the occluding object types to be known in advance [28,36,37,43,46]. Solutions to learn semantic priors on transience exist [23] but separating occlusions via semantic segmentation without manual guidance is ill posed. Occ-NeRF [54] considers any foreground element as occlusion and removes them via depth reasoning, but their removal leaves behind blurry artifacts. Alternatively, some methods do not remove dynamic distractors, but reconstruct them together with the rest of the scene using time-conditioned representations [9,26,34,49]. Closely related to our work, RobustNeRF [38] considers input-image distractors as outliers of the model optimization task. The authors employ robust losses improved via patching to preserve high-frequency details. RobustNeRF does not rely on prior assumptions about the nature of distractors, nor does it require preprocessing of the input data or postprocessing of the trained model. Nevertheless, their method comes at the cost of loosing view-dependent details and a reduced reconstruction quality in undistracted scenes. Furthermore, their method is limited to unconditioned models that overfit to a single scene. A generalization to conditioned NeRFs, such as pixelNeRF [52], is not obvious, as no further optimization takes place during inference.", + "bbox": [ + 215, + 181, + 787, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Conditioned neural fields offer a distinct advantage in their ability to generalize to novel scenes by leveraging knowledge acquired from diverse scenes during learning. This results in a more robust model that requires as few as one input view for inference, showcasing the efficiency and adaptability of the approach. Contrary to PixelNeRF [52], which relies on a volumetric parametrization of the scene, demanding multiple network evaluations along the ray, Light Field Networks (LFNs) [40], which succeed Scene Representation Networks [41], take a different approach. LFNs represent the scene as a 4D light field, enabling a more efficient single evaluation per ray for inference. The network takes as input a ray represented in Plücker coordinates and maps it to an observed radiance, all within an autoencoder framework used for conditioning.", + "bbox": [ + 215, + 532, + 787, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "None of the mentioned methods can deal with occlusions in single-shot reconstruction and no prior work exists on robust LFNs or robustness of other conditioned neural fields for single-shot reconstruction, which we address via the RANSAC paradigm [15]. Since its introduction in 1981, RANSAC has gained attention for fitting analytical models with a small number of parameters, such as homography estimation in panorama stitching [7]. Among the few direct applications of classical RANSAC to larger models is robust morphable face reconstruction [13, 14]. Other common expansions and applications include differentiable RANSAC [5] for camera parameter estimation in a deep learning pipeline, locally", + "bbox": [ + 215, + 704, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "B. Buschmann et al.", + "bbox": [ + 271, + 114, + 408, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/92f72ab285e6d711cd62058e8ff29bf00bd4f204d2952d6b976d9b2b5395fa69.jpg", + "image_caption": [ + "Fig. 2: The RANRAC algorithm for LFNs samples random hypotheses by selecting a set of random samples from the given perspective (a), and inferring the latent representation of these rays using the autodecoder of a pretrained LFN (b). The obtained light field is then used to predict an image from the input perspective (c). Based on this prediction, confidence in the random hypothesis is evaluated via the Euclidean distance between the predicted ray colors and the remaining color samples of the input image. The amount of samples which are explained by each hypothesis up to some margin are used to determine the best hypothesis (d). All samples explained by the selected hypothesis are used for a final inference with the LFN to obtain the final model and latent representation (e)." + ], + "image_footnote": [], + "bbox": [ + 217, + 146, + 782, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "optimized RANSAC [11] to account for the requirement of a descriptive sample set, and adaptive real-time RANSAC [35].", + "bbox": [ + 212, + 439, + 787, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 491, + 328, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we present our approach to increase the robustness of neural scene representations to inconsistencies in the input data. First, we recap RANSAC, its theoretical convergence and hyperparameters, and the required adaptions for its application to high-dimensional data-driven models. We then introduce a general scheme for the random sampling and validation of neural fields. Based hereon, we formulate a robust algorithm using LFNs for 3D reconstruction from a single image with occlusions. Finally, using NeRF, we formulate an algorithm for robust photo-realistic reconstruction from multiple views in the presence of common sources of inconsistencies.", + "bbox": [ + 212, + 522, + 787, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 RANSAC Convergence on Complex Models", + "text_level": 1, + "bbox": [ + 215, + 679, + 624, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Classical RANSAC [10,12,15] follows an iterative process. Initially, a minimal set of samples is randomly selected to determine the model parameters, known as the hypothesis generation phase. Then, the hypothesis is evaluated by assessing the number of observations it explains, within a specified margin. These steps are repeated until the best hypothesis is chosen to constitute the consensus set, which comprises all of its inliers.", + "bbox": [ + 212, + 703, + 787, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This paradigm cannot be directly applied to complex models such as neural fields, as a significant amount of samples is required to obtain decent initial model parameters and additional clean samples improve the quality further.", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "RANRAC", + "bbox": [ + 661, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This imposes a challenge regarding the expected amount of clean initial sample sets, $S_{clean}$ :", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} [ \\# S _ {\\text {c l e a n}} ] = N * \\prod_ {m = 1} ^ {M} \\frac {s _ {i m g} - s _ {i m g} ^ {\\text {o c c}} - m}{s _ {i m g} - m}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 176, + 785, + 217 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $s_{img}$ denotes the total amount of samples (e.g., image pixels), $s_{img}^{occ}$ represents the occluded samples, and $N / M$ are the number of iterations/samples. The expected amount exponentially decreases with the initial number of samples.", + "bbox": [ + 212, + 222, + 782, + 267 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The samples and respective requirements for analytic and data-driven models vary a lot. The effect of individual samples is less traceable in data-driven models and the information entropy varies more significantly across samples. When using a model that projects onto a latent space, some very atypical outliers do not show an effect at all if the latent space is not expressive enough to explain them in an overall loss-reducing way, yet, outliers close to the object or its color, or larger chunks of outliers, will usually be distracting. At the same time, samples of small-scale high-frequency details are important for the reconstruction and contain a lot of information, whereas multiple samples of larger-scale lower-frequency details contribute much less. The amount of initial samples for the hypothesis generation becomes a tunable hyperparameter trading initial reconstruction quality for likelihood of finding desired sample sets. This invalidates the classical convergence idea [10, 12] where the RANSAC iterations $N$ with", + "bbox": [ + 212, + 268, + 784, + 464 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nN \\geq \\frac {\\log (1 - p)}{\\log (1 - t ^ {M})} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 437, + 470, + 785, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "are chosen such that at least one clean sample set is found with a probability $p$ , given the expected ratio of clean samples $t$ and the amount of initial samples $M$ . There is not only the need to find a clean sample set, but one that captures all important details. At the same time, a completely clean sample set is not required at all, as long as the contained outliers are not represented by the local minimum of the latent space or the model itself, depending on the concrete scene representation.", + "bbox": [ + 212, + 506, + 787, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Random Sampling Neural Fields", + "text_level": 1, + "bbox": [ + 214, + 633, + 532, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We propose a general strategy for RANSAC-like, iterative, robust reconstruction with neural fields, before formulating respective algorithms for LFNs and NeRF.", + "bbox": [ + 212, + 657, + 782, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sampling Hypotheses Based on the application a feasible sampling domain is chosen e.g. pixels/rays, or observations. Depending on the requirements of the neural field an appropriate sample size is determined. The initial sample size is chosen as small as possible to allow for a reasonable convergence expectation to cleaner sets, while being large enough for a coarse fit of the model expressive enough to enable the discrimination of outliers.", + "bbox": [ + 212, + 688, + 782, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Determine Model Parameters The determination of the model parameters in a classical RANSAC application corresponds to the inference of the sample sets with a neural field: In case of unconditioned neural fields (such as NeRF) this corresponds to (over-)fitting the model to the scene, in case of conditioned fields", + "bbox": [ + 212, + 779, + 784, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "B. Buschmann et al.", + "bbox": [ + 271, + 114, + 408, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "this corresponds to obtaining a latent representation. Full convergence is not required, the reconstruction is only used to evaluate the outlier contamination.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Validation of Hypotheses For the validation, the rays/perspectives not used for the inference are rendered using the obtained model and compared to the input. The quality of each hypothesis is evaluated based on the amount of other samples explained by it up to some margin. Depending on the sampling/inference strategy, different similarity metrics can be used to distinguish outliers from other sources of noise such as the coarse inference. The initial sample set of the best hypothesis together with all its inliers is used for a final complete model fit.", + "bbox": [ + 212, + 176, + 782, + 281 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As will be demonstrated in our evaluation in Section 5, the major benefit of our RANSAC-like neural field approach over formulations based on robust loss functions [1] is the possibility to reliably filter outliers and inconsistencies in the input data in comparison to the down-weighting of their influence.", + "bbox": [ + 212, + 282, + 782, + 342 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Robust Light Field Networks", + "text_level": 1, + "bbox": [ + 214, + 364, + 501, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the following, we propose a novel fast and robust single-shot multi-class reconstruction algorithm based on LFNs [40]. LFNs are globally conditioned, meaning that the supported subset of 3D consistent scenes is represented by a single global latent vector. Therefore, when the latent space is not expressive enough to represent object and distractor correctly, large inconsistencies cause global damage instead of local artifacts. LFNs intrinsically support parallel inference, allowing to jointly process all hypotheses rather than following an iterative scheme. Our algorithm consists of the following steps (Fig. 2):", + "bbox": [ + 212, + 390, + 782, + 510 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "1. Hypothesis Consensus Set: Given the input image $I$ as a set of pixel color values $\\pmb{c}_i$ , and the intrinsic and extrinsic camera parameters, the set of rays $R$ - one ray $\\pmb{r}_i$ for every pixel - represented by Plücker coordinates, is generated. In the first step, $N$ initial consensus sets $S_{n}$ are drawn using a uniform distribution, where each consensus set consists of $M$ random samples:", + "bbox": [ + 218, + 523, + 782, + 597 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\boldsymbol {c} _ {n} ^ {m}, \\boldsymbol {r} _ {n} ^ {m}\\right) \\in_ {R} \\left\\{\\left(\\boldsymbol {c} _ {i}, \\boldsymbol {r} _ {i}\\right) \\mid \\boldsymbol {c} _ {i} \\in I, \\boldsymbol {r} _ {i} \\in R \\right\\} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 374, + 598, + 784, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $n \\in \\{1, \\dots, N\\}$ , $m \\in \\{1, \\dots, M\\}$ , and $\\in_R$ denotes a sample randomly drawn from the set without replacement according to a uniform distribution.", + "bbox": [ + 236, + 614, + 782, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "2. Hypothesis Inference: The autodecoder of an LFN $\\varPhi$ with hypernetwork $\\varPsi$ and pretrained hypernetwork weights $\\psi$ , is used to infer the latent codes $z_{n}$ for each of the initial sample sets in parallel.", + "bbox": [ + 218, + 643, + 782, + 688 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\boldsymbol {z} _ {n} \\right\\} = \\underset {\\left\\{\\boldsymbol {z} _ {n} \\right\\}} {\\operatorname {a r g m i n}} \\sum_ {n} \\sum_ {m} \\| \\Phi \\left(\\boldsymbol {r} _ {n} ^ {m} \\mid \\Psi_ {\\psi} (\\boldsymbol {z} _ {n})\\right) - \\boldsymbol {c} _ {n} ^ {m} \\| _ {2} ^ {2} + \\lambda_ {l a t} \\| \\boldsymbol {z} _ {n} \\| _ {2} ^ {2} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 299, + 688, + 784, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\lambda_{lat}$ determines the strength of the Gaussian prior on the latent space [40]. An exponential learning rate schedule speeds up the inference. The inferred latent codes form the hypotheses.", + "bbox": [ + 236, + 718, + 782, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3. Hypothesis Prediction: Each of the hypotheses is used to render an entire image $I_{n}^{pred}$ from the perspective of the input image, each consisting of the pixel color values $\\pmb{c}_{n,i}^{pred} = \\Phi(\\pmb{r}_i \\mid \\Psi_\\psi(\\pmb{z}_n))$ , using again the set of rays $R = \\{\\pmb{r}_i\\}$ obtained from the camera parameters. The rendered pixels resemble the predictions for the remaining observations under each hypothesis.", + "bbox": [ + 218, + 762, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "RANRAC", + "bbox": [ + 661, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "4. Hypothesis Validation: The obtained predictions are compared to the input image to validate the hypothesis. For each pixel in each predicted image, we calculate the Euclidean distance in color space: $e_{n,i} = \\| \\pmb{c}_{n,i}^{pred} - \\pmb{c}_i\\| _2$ . For each image, using these distances, we collect, up to some margin $\\epsilon$ , the observations explained by the model (inliers): $S_{n}^{inlier} = \\{(c_{i},r_{i})\\mid e_{n,i} < \\epsilon \\}$", + "5. Model Selection: We select the best hypothesis sample set $S_{best}$ based on the number of inliers $\\# S_n^{inlier}$ . The model is inferred once more, similar to the second step, to obtain the final latent code $z_{cons}$ . The inference is based on the final consensus set $S_{cons} = S_{best} \\cup S_{best}^{inlier}$ , the initial sample set of the strongest hypothesis $S_{best}$ together with all its inliers $S_{best}^{inlier}$ . The final output consists of both the latent code $z_{cons}$ and the final consensus set $S_{cons}$ of the selected model, and can be used to render arbitrary new perspectives." + ], + "bbox": [ + 218, + 146, + 787, + 328 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Robust Neural Radiance Fields", + "text_level": 1, + "bbox": [ + 215, + 353, + 519, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the following, we propose a novel robust multi-view reconstruction approach based on NeRFs. NeRFs are fit to a single scene based on a set of observations. As they support view-dependent radiance, one might expect inconsistencies in the input observations to only have an effect on specific perspectives. However, the density is only spatially parametrized, hence, inconsistencies lead to significant ghosting and smearing artifacts in more than just the inconsistent perspective. This can be leveraged in the hypothesis validation. The following steps are performed for N iterations:", + "bbox": [ + 212, + 381, + 787, + 501 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "1. Hypothesis Consensus Set: A NeRF fit requires a large set of rays, making a sampling in ray space infeasible, as even obtaining significantly cleaner sets becomes unlikely. In order to obtain a sampling domain with reasonably sized initial consensus sets $S_{n}$ , we sample $M$ observations from the given sets of images $\\mathcal{I}$ and corresponding camera poses $\\mathcal{C}$ in every iteration:", + "bbox": [ + 218, + 517, + 787, + 593 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\left(I _ {n} ^ {m}, E _ {n} ^ {m}\\right) \\in_ {R} \\left\\{\\left(I _ {i}, E _ {i}\\right) \\mid I _ {i} \\in \\mathcal {I}, E _ {i} \\in \\mathcal {C} \\right\\} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 593, + 785, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $n \\in \\{1, \\dots, N\\}$ , $m \\in \\{1, \\dots, M\\}$ , and $\\in_R$ denotes a sample randomly drawn from the set without replacement according to a uniform distribution.", + "bbox": [ + 238, + 609, + 785, + 638 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2. Hypothesis Inference: The sampled observations are used to fit a hypothesis neural radiance field $F_{n}$ .", + "3. Hypothesis Prediction: The obtained NeRF $F_{n}$ is used to render predictions $I_{n,i}^{pred}$ with pixel colors $\\pmb{p}_{n,i,x}^{pred}$ under the hypothesis for all unseen input perspectives $E_{i}$ .", + "4. Hypothesis Validation: The chosen sample space requires a careful evaluation of the hypotheses. We propose a two-step evaluation, where, first, the pixels inliers $P_{n,i}$ up to some margin $\\epsilon_{pix}$ are determined for every observation based on the Euclidean distance in color space to the pixels $\\pmb{p}_{i,x}$ of the input images $I_i$ : $P_{n,i} = \\{\\pmb{p}_{n,i,x}^{pred} \\mid \\| \\pmb{p}_{n,i,x}^{pred} - \\pmb{p}_{i,x} \\|_2 < \\epsilon_{pix}\\}$ and, second, the observations themselves are labeled as inliers or outliers based on the amount of pixel inliers, again up to some margin $\\epsilon_{img}$ , to obtain the consensus set $S_n^{inlier}$ under the hypothesis: $S_n^{inlier} = \\{(I_i, E_i) \\mid \\# P_{n,i} > \\epsilon_{img}\\}$ ." + ], + "bbox": [ + 218, + 638, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "B. Buschmann et al.", + "bbox": [ + 271, + 114, + 408, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The binary metric for pixels ensures that smaller mispredictions (due to, e.g., view-dependent lighting effects) do not introduce noise into the evaluation.", + "bbox": [ + 236, + 146, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Model Selection: The strongest hypothesis, with its initial sample set $S_{best}$ , is selected based on the number of inliers $\\# S_n^{inlier}$ to obtain the final consensus set $S_{cons} = S_{best} \\cup S_{best}^{inlier}$ . The final model is obtained with one more NeRF fit of the consensus set.", + "bbox": [ + 218, + 176, + 787, + 237 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.5 Hyperparameters", + "text_level": 1, + "bbox": [ + 215, + 258, + 408, + 273 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For LFNs, the amount of initial samples and random hypotheses to evaluate (iterations), are determined experimentally. Without fine-tuning per class, the experimentally determined parameters are 90 initial samples and 2000 iterations, which supersedes the theoretical value for a convergence because the latent space introduces an intrinsic robustness. The inlier margin balances the amount of slight high-frequency variations that are being captured and the capability of separating outliers that are similar to the object. A margin of 0.25 in terms of the Euclidean distance of the predicted colors to the input samples in an RGB color space normalized to the range $(-1, 1)$ has been found to be optimal. Please refer to the supplementary material for further experimental results.", + "bbox": [ + 212, + 285, + 787, + 435 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For NeRFs, the parameters behave more natural and the amount of perspectives required for a meaningful, not completely artifact-free fit of the model lies around 25 observations [29, Table 2]. With fewer samples, more artifacts are introduced that get harder to separate from the ones caused by inconsistencies, and the samples get more dependent on being evenly spaced. For real-world captures with $10\\%$ inconsistent perspectives, as few as 50 iterations are sufficient. With the color space normalized to $(0,1)$ , a pixel margin between around 0.15 in terms of Euclidean distance worked well for the determination of actual artifacts. We consider an observation an inlier based on a margin of $90\\% - 98\\%$ pixel inliers, which proved to be a good choice to separate minor artifacts (due to the sparse sampling) from artifacts caused by actual inconsistencies. For different datasets or inconsistencies, these values can be adapted.", + "bbox": [ + 212, + 436, + 787, + 618 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Implementation & Preprocessing", + "text_level": 1, + "bbox": [ + 215, + 641, + 571, + 659 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For LFNs, we build on top of the original implementation [40], with a slight adaptation to enable a parallel sub-sampled inference. We furthermore use the provided pretrained multi-class model. The camera parameters are known. For efficiency reasons, the steps of the algorithm are not performed iteratively, but multiple hypotheses are validated in parallel. To further speed up the inference, an exponential learning rate schedule is used for the auto-decoding, leading to a total runtime of about a minute on a single GPU.", + "bbox": [ + 212, + 672, + 787, + 777 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For the robust reconstruction of objects from lazily captured real-world data, one has to estimate the camera parameters and extract foreground masks before applying the algorithm. For the estimation of the camera parameters, we used the COLMAP structure-from-motion package [39]. We extracted foreground masks", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "RANRAC", + "bbox": [ + 661, + 113, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "using Segment Anything [22]. However, only foreground masks, containing the objects and the occlusions, are extracted. Segment Anything is not capable of removing arbitrary occlusions in an automatized way. After these preprocessing steps, the robust reconstruction algorithm can be applied as described. Erroneous estimates of the camera parameters or foreground masks are excluded by our algorithm, thus making the entire reconstruction pipeline robust.", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Our algorithm is not limited to a specific NeRF implementation. The chosen sampling domain eases integration into arbitrary existing NeRF implementations, which commonly expect images instead of unstructured ray sets. However, using a fast NeRF variant is advantageous when applying an iterative scheme. We used the instant NGP implementation [31] of the instant NSR repository [18], which includes some accelerations [24, 30]. Other, (specifically fast) variants are likely good choices as well. Antialiased and unbounded, but slow variants, such as MipNeRF360 [3], are not feasible. For further implementation details please refer to the supplementary.", + "bbox": [ + 212, + 238, + 787, + 375 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 398, + 375, + 417 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Inconsistencies,Baselines & Datasets", + "text_level": 1, + "bbox": [ + 215, + 433, + 568, + 448 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Multi-View Reconstruction (NeRF) We provide a comprehensive quantitative and qualitative analysis of RANRAC's multi-view reconstructions compared to NeRF without any method of robustness (baseline) and RobustNeRF [38] (state of the art). We conduct experiments for various common sources of inconsistencies such as occluded perspectives, noisy camera parameter estimates, and blurred perspectives. RobustNeRF [38] targets unbounded scenes with multiple objects and small amounts of distractors in every perspective. In comparison, our RANSAC-based approach deals well with single-object reconstruction, even with heavy occlusions, as long as enough clean perspectives are available. As their dataset reflects the algorithm's properties, we cannot provide a fair comparison. Instead, we demonstrate the applicability using a custom dataset of a single object with a controlled amount of deliberately occluded perspectives. For the other inconsistencies we use of-the-shelf datasets and add noise to the camera parameters and blur to the images. Furthermore, we implement the robust losses [38] on the same NeRF variant as RANRAC to provide a fair comparison of the method's robustness, independent of the NeRF variant. For further details on the implementation, please refer to the supplementary material.", + "bbox": [ + 212, + 460, + 787, + 718 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Single-Shot Reconstruction (LFN) We benchmark against the original LFN implementation of Sitzmann et al. [40] as baseline, as there are no other robust methods for LFNs or conditioned neural fields, nor are there methods for robust single-shot multi-class reconstruction in general. Furthermore, we use the same pretrained LFN for the baseline and for the application of our method. The LFN is pretrained on the thirteen largest ShapeNet classes [8].", + "bbox": [ + 212, + 718, + 787, + 809 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We provide a detailed qualitative and quantitative performance comparison under different amounts of occlusion for three representative classes (plane, car,", + "bbox": [ + 215, + 809, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "B. Buschmann et al.", + "bbox": [ + 271, + 114, + 408, + 127 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "and chair), while just stating reconstruction performance in a fixed environment without additional tuning of the hyperparameters for the others. The plane class is mostly challenging due to the low-frequency shape, while the car class contains a lot of high-frequency color details. The chair class represents shapes that are generally problematic for vanilla LFNs, even without occlusions. We provide a complementing analysis of the hyperparameters in the supplementary material. If not stated otherwise, we evaluate using 50 randomly selected images of the corresponding class. All comparisons use the same images.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The occlusions are created synthetically. They consist of randomly generated patches while controlling two metrics of occlusion: Image occlusion and object occlusion. The former is the naive ratio of occluded over total pixels. The latter are the occluded pixels on the object compared to the total pixels covered by the object. We use both metrics to take the vastly different information entropy of samples across the image into account. For further details on the generation of the synthetic occlusions, please refer to the supplemental.", + "bbox": [ + 212, + 267, + 787, + 373 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Evaluation", + "text_level": 1, + "bbox": [ + 215, + 393, + 351, + 407 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/3bed12e2e029c354e129405521e84d64e8b395ae40c59dda98bc8d7891795a77.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MetricModelBenchBoatCabin.Displ.LampPhoneRifleSofaSpeak.Table
PSNR↑RANRAC19.2122.9221.9217.6519.9918.4521.3521.6120.720.44
LFN17.8919.220.518.8519.0917.8118.4620.1219.9920.29
SSIM↑RANRAC0.7670.8580.8010.6990.7640.750.8530.8050.7610.784
LFN0.7240.7910.7670.730.7480.7260.7950.7750.7430.777
", + "bbox": [ + 225, + 422, + 777, + 478 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 1: RANRAC obtains a significant quantitative improvement in PSNR and SSIM (higher is better) compared to the baseline. We compare RANRAC to vanilla LFNs for the 13 largest ShapeNet classes (find plane, car, and chair with more detail in Fig. 3). The results are based on a moderate amount of occlusion of approximately $25\\%$ object occlusion and about $5\\%$ image occlusion. The reported results are conservative, as higher amounts of image occlusion result in a more significant performance increase (Fig. 3). No hyperparameter tuning has been performed for these classes; the same configuration obtained from the analysis of the other three classes is used.", + "bbox": [ + 212, + 479, + 787, + 589 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For LFNs, our approach leads to a significant improvement in occluded scenarios of up to 8dB in PSNR and a similarly strong improvement for the SSIM. The improvement is most significant in heavily distracted scenarios (Fig. 3). In clean scenarios a slight performance penalty can be observed, but even with small amounts of object occlusion (information loss), our algorithm outperforms the baseline, leading to numerically better results up to $50\\%$ information loss (Fig. 3). The effect is not only measurable, but also well visible (Fig. 4). Increasing amounts of occlusion slowly introduce local artifacts into our reconstruction while preserving a reasonable shape estimate even for larger amounts of occlusion. In contrast, the reconstruction of LFNs breaks rather early in a global fashion. Still, our consensus set (Fig. 4), reveals that some high-frequency details were wrongfully excluded, explaining the slight performance decrease on clean images. In general, the benefit of RANRAC is best observable for classes that can be well described by LFNs, as evidenced by the lower improvements the chair class exhibits, compared to the plane and car class. The same effect is visible in the quantitative evaluation on the other classes without additional", + "bbox": [ + 212, + 598, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "RANRAC", + "bbox": [ + 661, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/67c07b3b1a3d8dbe282989f20132f533f8508aec7cd161c7b0ad171a7fd958d0.jpg", + "image_caption": [ + "Fig. 3: RANRAC (solid lines) leads to a quantitative improvement in PSNR and SSIM (higher is better) for occluded inputs compared to vanilla LFNs (dashed lines). The same hyperparameter configuration and LFN is used for all classes. On the left and in the middle, the amount of image occlusion is increased, while the object occlusion is constant at $25\\%$ . On the right, the amount of object occlusion is increased while the image occlusion is kept low. For the car class, a large improvement is observed over the entire occlusion spectrum. For the plane class the improvement is similarly significant, but absolute performance degenerates a bit sooner. This stems from the smaller object size and the related faster occlusion-to-object increase when increasing image occlusions. For the chair class, the improvement is less significant but the structural similarity is preserved for much longer. For the plane and car class the reconstruction quality is resilient to information loss (right) up to $\\sim 50\\%$ , where the decrease gains momentum. With the low amounts of image occlusion, the improvement is not significant for the chair class (consistent with left and middle)." + ], + "image_footnote": [], + "bbox": [ + 218, + 150, + 785, + 290 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/79686c1d75bc7e7e181fe8e489ab0db2b4cf96e9f2fcfd30764ed3fc85bc3deb.jpg", + "image_caption": [ + "Fig. 4: On the left, we show the qualitative effect of increasing occlusion on the same observation for the reconstruction of a novel view. Reconstructions of LFNs break early globally whereas RANRAC still provides a very decent reconstruction, only slowly introducing minor local (and natural/comprehensible) artifacts for completely hidden object parts. We further show the obtained consensus set, used for the final reconstruction (green inliers, red outliers). On the right, we show more qualitative results for novel view synthesis on different classes and the corresponding consensus sets." + ], + "image_footnote": [], + "bbox": [ + 236, + 516, + 772, + 722 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "B. Buschmann et al.", + "bbox": [ + 271, + 114, + 408, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/25a1557f50a85e6137cf682d89827c485cd7f03fa3be508e6ad29c12f8e5c6e5.jpg", + "image_caption": [ + "Fig. 5: The occlusions lead to well-visible artifacts in the reconstructions of NeRF, these artifacts are completely removed by RANRAC. While RobustNeRF struggles with view-dependent and high-frequency details, RANRAC reliably reconstructs them." + ], + "image_footnote": [], + "bbox": [ + 230, + 157, + 769, + 448 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/510e3e55331ca81aac990b6deee12b237bebf28a7d864a5e32056f9226ae3c16.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Inconsistency DatasetMild Occ.Strong Occ.Watering PotBlurred Perspectives
LegoShipChairFicusMic
RANRAC27.1125.9926.1224.9434.7929.9129.7616.1435.2531.26
RobustNeRF26.8325.5825.9324.7929.1423.6923.3120.1931.1127.12
NeRF26.6522.6125.3618.4731.1519.0028.4820.3533.2122.10
", + "bbox": [ + 236, + 540, + 764, + 583 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/8f7cdf596133d6621b6a70c477567db6d181a657be275f9aaeab458edfbd9f25.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Inconsistency \nDatasetNoisy Camera Parameters
LegoDrumsMicShipFicusHotdogMaterials
RANRAC34.95|31.2225.88|22.7735.85|33.5530.77|22.0531.41|28.8237.16|30.5328.93|25.39
RobustNeRF29.83|26.8223.91|22.1029.36|27.5524.21|19.9823.17|21.8532.65|25.1624.90|22.30
NeRF24.67|15.9523.22|17.9628.82|22.1523.68|14.8028.12|22.7728.16|18.3724.73|19.89
", + "bbox": [ + 236, + 587, + 764, + 628 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 2: RANRAC outperforms the state of the art RobustNeRF [38] and the baseline NeRF (without a method of robustness) for scenes contaminated with occlusions, blur and noisy camera parameters. We report PSNR $\\uparrow$ averaged across perspectives and the 5th percentile $(\\mathrm{Avg.}|P_5)$ as artifacts introduced by inconsistencies only contaminate some views. We compare on the captured watering pot dataset with milder $(10\\%)$ and heavier $(17.5\\%)$ amounts of occluded perspectives and off-the-shelf datasets [29] with blurred perspectives and additive Gaussian noise $\\mathcal{N}(5^{\\circ},1^{\\circ})$ on the camera parameters of $10\\%$ of the perspectives. All three variants are built on top of instant-nsr for an isolated fair comparison of the robustness method. Note that both robust approaches struggle separating the strong view-dependent effects of the ship scene from blur, leading to the exclusion of some perspectives (lower $P_5$ PSNR), while RANRAC still improves the overall reconstruction. For all other scenes and inconsistencies, RANRAC reliably separates inconsistent observations from clean ones.", + "bbox": [ + 212, + 628, + 787, + 808 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "RANRAC", + "bbox": [ + 661, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "tuning (Tab. 1). The only outlier is the display class, on which LFNs struggle the most on [40], even in unoccluded scenarios. It is not consistently represented in the latent space and robust reconstruction amplifies this effect by reducing the samples to a consistent set.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The application to NeRF shows the significantly improved reconstructions of RANRAC (Tab. 2) for different types of inconsistencies such as occlusions and blur, as well as noisy camera parameters on captured and synthetic datasets. RANRAC consistently outperforms RobustNeRF [38] for object reconstruction from inconsistent inputs. RobustNeRF specifically struggles in capturing view-dependent appearance and at concavities while RANRAC seamlessly reconstructs them. The effects are not only measurable but also well-visible (Fig. 5).", + "bbox": [ + 212, + 208, + 785, + 313 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Limitations & Future Work", + "text_level": 1, + "bbox": [ + 214, + 342, + 517, + 358 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "For the application to NeRF, the iterations imply the use of a fast inferable variant, ruling out MipNeRF360 [3] and other high-quality variants for unbounded scenes. Foreground separation is a must, limiting this application of RANRAC to object-centric scenes. Novel fast, unbounded methods might resolve this [20].", + "bbox": [ + 212, + 378, + 782, + 439 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "By following a RANSAC-like approach, we inherit the requirement of sufficient clean perspectives, which could be lifted via NeRF variants that require fewer perspectives [32, 45] or by using different sampling domains. In return, our method is not limited to specific kinds of inconsistencies, and is robust to arbitrarily heavy distractions or inconsistencies in the impure perspectives.", + "bbox": [ + 212, + 441, + 782, + 515 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Whereas our robust LFN approach led to significant improvements for single-shot reconstruction, our concept might also be applied to other conditioned neural fields, based on a smart choice of the sampling domain, targeting photorealism. One could also use importance sampling based on a prior, instead of a uniform sampling, leveraging the unevenly distributed information entropy. Neural sampling priors [6], semantic segmentation and more advanced schemes (e.g. LO-RANSAC [11], DSAC [5]) might prove useful here.", + "bbox": [ + 212, + 517, + 785, + 623 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 651, + 359, + 667 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We introduced a novel approach to increase the robustness of neural fields, inspired by the RANSAC paradigm. Following this concept, we introduced a novel robust approach for single-shot reconstruction from occluded views based on LFNs which achieves a significant improvement in reconstruction quality for distracted and occluded scenarios, even for extreme cases. Furthermore, we introduced a respective RANRAC-based NeRF variant that allows robust photorealistic reconstruction from multiple views with typical inconsistencies such as occlusions, noisy camera parameters, or blurred images - resulting in significant improvements compared to state-of-the-art methods - without relying on assumptions about the distractions.", + "bbox": [ + 212, + 688, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "B. Buschmann et al.", + "bbox": [ + 271, + 114, + 408, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements - Benno Buschmann was at FAU funded through a gift by Mitsubishi Electric Research Laboratories (MERL). Andreea Dogaru was funded by the German Federal Ministry of Education and Research (BMBF), FKZ: 01IS22082 (IRRW). The authors are responsible for the content of this publication.", + "bbox": [ + 215, + 147, + 787, + 205 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 229, + 321, + 244 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Barron, J.T.: A general and adaptive robust loss function. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4331-4339 (2019) 2, 7", + "2. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: ICCV (2021) 4", + "3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In: CVPR (2022) 4, 10, 14", + "4. Bian, W., Wang, Z., Li, K., Bian, J.W., Prisacariu, V.A.: Nope-nerf: Optimising neural radiance field with no pose prior. In: CVPR (2023) 3", + "5. Brachmann, E., Krull, A., Nowozin, S., Shotton, J., Michel, F., Gumhold, S., Rother, C.: Dsac-differentiable ransac for camera localization. In: CVPR (2017) 4, 14", + "6. Brachmann, E., Rother, C.: Neural-guided ransac: Learning where to sample model hypotheses. In: ICCV (2019) 14", + "7. Brown, M., Lowe, D.G.: Automatic panoramic image stitching using invariant features. International Journal of Computer Vision 74, 59-73 (2007) 4", + "8. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., Xiao, J., Yi, L., Yu, F.: ShapeNet: An Information-Rich 3D Model Repository. Tech. Rep. arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago (2015) 10", + "9. Chen, X., Zhang, Q., Li, X., Chen, Y., Feng, Y., Wang, X., Wang, J.: Hallucinated neural radiance fields in the wild. In: CVPR (2022) 4", + "0. Choi, S., Kim, T., Yu, W.: Performance evaluation of ransac family. Journal of Computer Vision 24(3), 271-300 (1997) 5, 6", + "1. Chum, O., Matas, J., Kittler, J.: Locally optimized ransac. In: Pattern Recognition: 25th DAGM Symposium, Magdeburg, Germany, September 10-12, 2003. Proceedings 25. pp. 236-243. Springer (2003) 5, 14", + "2. Derpanis, K.G.: Overview of the ransac algorithm. Image Rochester NY 4(1), 2-3 (2010) 5, 6", + "3. Egger, B., Schneider, A., Blumer, C., Forster, A., Schonborn, S., Vetter, T.: Occlusion-aware 3d morphable face models. In: BMVC. vol. 2, p. 4 (2016) 4", + "4. Egger, B., Schonborn, S., Schneider, A., Kortylewski, A., Morel-Forster, A., Blumer, C., Vetter, T.: Occlusion-aware 3d morphable models and an illumination prior for face image analysis. International Journal of Computer Vision 126, 1269-1287 (2018) 4", + "5. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM 24(6), 381-395 (1981) 2, 4, 5", + "6. Goli, L., Reading, C., Sellan, S., Jacobson, A., Tagliasacchi, A.: Bayes' Rays: Uncertainty quantification in neural radiance fields. arXiv (2023) 4" + ], + "bbox": [ + 225, + 258, + 785, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "RANRAC", + "bbox": [ + 661, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Guangcong, Chen, Z., Loy, C.C., Liu, Z.: Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. ICCV (2023) 3", + "18. Guo, Y.C.: Instant neural surface reconstruction (2022), https://github.com/bennyguo/instant-nsr-pl 10", + "19. Jeong, Y., Ahn, S., Choy, C., Anandkumar, A., Cho, M., Park, J.: Self-calibrating neural radiance fields. In: ICCV (2021) 3", + "20. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering (2023) 14", + "21. Kim, M., Seo, S., Han, B.: Infonerf: Ray entropy minimization for few-shot neural volume rendering. In: CVPR (2022) 3", + "22. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023) 10", + "23. Lee, J., Kim, I., Heo, H., Kim, H.J.: Semantic-aware occlusion filtering neural radiance fields in the wild. arXiv preprint arXiv:2303.03966 (2023) 4", + "24. Li, R., Gao, H., Tancik, M., Kanazawa, A.: Nerfacc: Efficient sampling accelerates nerfs. arXiv preprint arXiv:2305.04966 (2023) 10", + "25. Lin, C.H., Ma, W.C., Torralba, A., Lucey, S.: Barf: Bundle-adjusting neural radiance fields. In: ICCV (2021) 3", + "26. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust dynamic radiance fields. In: CVPR (2023) 4", + "27. Liu, Y., Peng, S., Liu, L., Wang, Q., Wang, P., Christian, T., Zhou, X., Wang, W.: Neural rays for occlusion-aware image-based rendering. In: CVPR (2022) 4", + "28. Martin-Brualla, R., Radwan, N., Sajjadi, M.S.M., Barron, J.T., Dosovitskiy, A., Duckworth, D.: NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In: CVPR (2021) 4", + "29. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) 2, 3, 9, 13", + "30. Müller, T.: tiny-cuda-nn (4 2021), https://github.com/NVlabs/tiny-cuda-nn 10", + "31. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. In: SIGGRAPH (2022) 3, 10", + "32. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S.M., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: CVPR (2022) 3, 14", + "33. Philip, J., Deschaintre, V.: Floaters no more: Radiance field gradient scaling for improved near-camera training. In: Eurographics Symposium on Rendering. The Eurographics Association (2023) 4", + "34. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: CVPR (2021) 4", + "35. Raguram, R., Frahm, J.M., Pollefeys, M.: A comparative analysis of ransac techniques leading to adaptive real-time random sample consensus. In: ECCV (2008) 5", + "36. Rebain, D., Matthews, M., Yi, K.M., Lagun, D., Tagliasacchi, A.: Lolnerf: Learn from one look. In: CVPR (2022) 4", + "37. Rematas, K., Liu, A., Srinivasan, P.P., Barron, J.T., Tagliasacchi, A., Funkhouser, T., Ferrari, V.: Urban radiance fields. CVPR (2022) 4", + "38. Sabour, S., Vora, S., Duckworth, D., Krasin, I., Fleet, D.J., Tagliasacchi, A.: Robustnerf: Ignoring distractors with robust losses. In: CVPR (2023) 2, 3, 4, 10, 13, 14" + ], + "bbox": [ + 215, + 146, + 784, + 837 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "B. Buschmann et al.", + "bbox": [ + 271, + 114, + 408, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "39. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: CVPR (2016) 9", + "40. Sitzmann, V., Rezchikov, S., Freeman, W.T., Tenenbaum, J.B., Durand, F.: Light field networks: Neural scene representations with single-evaluation rendering. In: NeurIPS (2021) 2, 3, 4, 7, 9, 10, 14", + "41. Sitzmann, V., Zollhöfer, M., Wetzstein, G.: Scene representation networks: Continuous 3d-structure-aware neural scene representations. In: NeurIPS (2019) 4", + "42. Sun, J., Chen, X., Wang, Q., Li, Z., Averbuch-Elor, H., Zhou, X., Snavely, N.: Neural 3D reconstruction in the wild. In: SIGGRAPH (2022) 4", + "43. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: CVPR (2022) 4", + "44. Tewari, A., Thies, J., Mildenhall, B., et al.: Advances in neural rendering. In: Comput. Graph. Forum. vol. 41, pp. 703-735. Wiley Online Library, Wiley (5 2022) 1, 3", + "45. Truong, P., Rakotosaona, M.J., Manhardt, F., Tombari, F.: Sparf: Neural radiance fields from sparse and noisy poses. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4190-4200 (2023) 14", + "46. Turki, H., Ramanan, D., Satyanarayanan, M.: Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In: CVPR (2022) 4", + "47. Warburg*, F., Weber*, E., Tancik, M., Holski, A., Kanazawa, A.: Nerfbusters: Removing ghostly artifacts from casually captured nerfs. In: ICCV (2023) 4", + "48. Wirth, T., Rak, A., Knauthe, V., Fellner, D.W.: A post processing technique to automatically remove floater artifacts in neural radiance fields. In: Computer Graphics Forum (2023) 4", + "49. Wu, T., Zhong, F., Tagliasacchi, A., Cole, F., Oztireli, C.: D\\~ 2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. NeurIPS (2022) 4", + "50. Xiangli, Y., Xu, L., Pan, X., Zhao, N., Rao, A., Theobalt, C., Dai, B., Lin, D.: Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In: ECCV (2022) 4", + "51. Xie, Y., Takikawa, T., Saito, S., Litany, O., Yan, S., Khan, N., Tombari, F., Tompkin, J., Sitzmann, V., Sridhar, S.: Neural fields in visual computing and beyond. Computer Graphics Forum 41 (2021) 1, 3", + "52. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: CVPR (2021) 2, 3, 4", + "53. Zhang, J.Y., Ramanan, D., Tulsiani, S.: Relpose: Predicting probabilistic relative rotation for single objects in the wild. In: ECCV (2022) 3", + "54. Zhu, C., Wan, R., Tang, Y., Shi, B.: Occlusion-free scene recovery via neural radiance fields. In: CVPR (2023) 4" + ], + "bbox": [ + 212, + 146, + 785, + 700 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "RANRAC", + "bbox": [ + 661, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/RANRAC_ Robust Neural Scene Representations via Random Ray Consensus/de9512b1-9e5d-48ed-a48f-d5995f4754ec_model.json b/2024/RANRAC_ Robust Neural Scene Representations via Random Ray Consensus/de9512b1-9e5d-48ed-a48f-d5995f4754ec_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2f5a8969fba0c0f5114fde6fd03b1a3ec81a2819 --- /dev/null +++ b/2024/RANRAC_ Robust Neural Scene Representations via Random Ray Consensus/de9512b1-9e5d-48ed-a48f-d5995f4754ec_model.json @@ -0,0 +1,2203 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.221, + 0.141, + 0.783, + 0.187 + ], + "angle": 0, + "content": "RANRAC: Robust Neural Scene Representations via Random Ray Consensus" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.212, + 0.747, + 0.244 + ], + "angle": 0, + "content": "Benno Buschmann1,2, Andreea Dogaru2, Elmar Eisemann1, Michael Weinmann1, and Bernhard Egger2" + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.255, + 0.654, + 0.269 + ], + "angle": 0, + "content": "1 Delft University of Technology, Netherlands" + }, + { + "type": "text", + "bbox": [ + 0.29, + 0.27, + 0.711, + 0.283 + ], + "angle": 0, + "content": "\\(^{2}\\) Friedrich-Alexander-University Erlangen-Nürnberg, Germany" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.32, + 0.74, + 0.666 + ], + "angle": 0, + "content": "Abstract. Learning-based scene representations such as neural radiance fields or light field networks, that rely on fitting a scene model to image observations, commonly encounter challenges in the presence of inconsistencies within the images caused by occlusions, inaccurately estimated camera parameters or effects like lens flare. To address this challenge, we introduce RANdom RAy Consensus (RANRAC), an efficient approach to eliminate the effect of inconsistent data, thereby taking inspiration from classical RANSAC based outlier detection for model fitting. In contrast to the down-weighting of the effect of outliers based on robust loss formulations, our approach reliably detects and excludes inconsistent perspectives, resulting in clean images without floating artifacts. For this purpose, we formulate a fuzzy adaption of the RANSAC paradigm, enabling its application to large scale models. We interpret the minimal number of samples to determine the model parameters as a tunable hyperparameter, investigate the generation of hypotheses with data-driven models, and analyse the validation of hypotheses in noisy environments. We demonstrate the compatibility and potential of our solution for both photo-realistic robust multi-view reconstruction from real-world images based on neural radiance fields and for single-shot reconstruction based on light-field networks. In particular, the results indicate significant improvements compared to state-of-the-art robust methods for novel-view synthesis on both synthetic and captured scenes with various inconsistencies including occlusions, noisy camera pose estimates, and unfocused perspectives. The results further indicate significant improvements for single-shot reconstruction from occluded images." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.68, + 0.74, + 0.707 + ], + "angle": 0, + "content": "Keywords: Neural scene representations \\(\\cdot\\) neural rendering \\(\\cdot\\) RANSAC \\(\\cdot\\) robust estimation \\(\\cdot\\) neural radiance fields \\(\\cdot\\) light-field networks" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.733, + 0.375, + 0.75 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "3D reconstruction is a classical task in computer vision and computer graphics, which has attracted research for decades. It offers numerous applications, including autonomous systems, entertainment, design, advertisement, cultural heritage, VR/AR experiences or medical scenarios. In recent years, neural scene representations and rendering techniques [44,51], including light field networks" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.411, + 0.128 + ], + "angle": 0, + "content": "B. Buschmann et al." + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.145, + 0.701, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.246, + 0.788, + 0.318 + ], + "angle": 0, + "content": "Fig. 1: We propose a robust algorithm for 3D reconstruction from occluded input perspectives that is based on the random sampling of hypotheses. Our algorithm is general and we demonstrate the use for single-shot reconstruction using light field networks or multi-view reconstruction using NeRF. In these cases, it successfully removes the artifacts that normally occur due to occluded input perspectives." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.353, + 0.788, + 0.504 + ], + "angle": 0, + "content": "(LFN) [40] and neural radiance fields (NeRF) [29] have demonstrated great performance in single-view and multi-view reconstruction tasks. The key to the success of such techniques is the coupling of differentiable rendering methods with custom neural field parametrizations of scene properties. However, a common limitation of neural scene reconstruction methods is their sensitivity to inconsistencies among input images induced by occlusions, inaccurately estimated camera parameters or effects like lens flares. Despite the use of view-dependent radiance representations to address view-dependent appearance changes, these inconsistencies severely impact local density estimation, resulting in a poor generalization to novel views." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.789, + 0.842 + ], + "angle": 0, + "content": "To increase the robustness to potential distractors within the training data, Sabour et al. [38] recently introduced the use of robust losses in the context of training unconditioned NeRF, where distractors in the training data were modeled as outliers of an optimization problem. However, the adaptation of this approach to conditioned neural fields (e.g., pixelNeRF [52]) is not obvious, as no optimization takes place during inference, and the input data is constrained to only a few views. Achieving robustness to data inconsistencies is a well-analyzed problem in computer vision, covered not only by the aforementioned robust loss functions [1], but also other strategies, like the random sample consensus (RANSAC) paradigm [15]. The latter is widely employed for fitting models to outlier-heavy data. The underlying idea is to randomly select subsets of the data to form potential models, evaluating these models against the entire dataset, and identifying the subset that best fits the majority of the data, while disregarding outliers. Despite being the state-of-the-art solution to many challenges, RANSAC-based schemes are particularly favoured for the fitting of analytical models with a relatively small amount of tuneable parameters. In this paper, we direct our attention to achieving robustness against inconsistencies and occlusions in the observations by using a novel combination of neural scene representation and rendering techniques with dedicated outlier removal techniques such as RANSAC [15]. While downweighting the influence of distractors based on robust losses [1, 38] can affect clean samples, representing details, we aim at improving robustness to distractors by only removing the influence of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.662, + 0.114, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RANRAC" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "outliers. To this extent, we integrate a RANSAC-based scheme to distinguish inliers and outliers in the data and the inlier-based optimization of the neural fields (Fig. 1); a stochastic scenario characterized by a large-scale, data-driven model that exceeds RANSAC's classical convergence expectations. Instead of guaranteeing convergence to a clean sample set based on a minimal number of samples, we aim for a feasible (cleaner) sample set using a tuneable amount of samples. The proposed algorithm exhibits robustness and versatility, accommodating a wide range of neural fields-based reconstruction methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.789, + 0.465 + ], + "angle": 0, + "content": "Our method inherits the strengths of RANSAC, such as the ability to handle various classes of outliers without relying on semantics. Yet, it also inherits the need for sufficiently clean samples and the reliance on an iterative scheme. In practice, the first condition is often fulfilled because typically only some of the perspectives are affected by inconsistencies. We validate our approach using synthetic data, focusing on the task of multi-class single-shot reconstruction with LFNs [40], and observe significant quality improvements over the baseline in the presence of occlusions. Furthermore, we showcase robust photo-realistic reconstructions of 3D objects using unconditioned NeRFs from sequences of real-world images in the presence of distractors. In comparison to RobustNeRF [38], we use all available clean data, hence improving the reconstruction quality for single-object scenes. Code and data are available under https://bennobuschmann.com/ranrac/. Our key contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.477, + 0.784, + 0.505 + ], + "angle": 0, + "content": "- a general, robust RANSAC-based reconstruction method applicable to a variety of neural-fields and handling diverse inconsistencies" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.508, + 0.785, + 0.536 + ], + "angle": 0, + "content": "- an analysis of the implication to RANSAC's hyperparameters and theoretical convergence expectations, and the experimental study of their effect" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.538, + 0.784, + 0.566 + ], + "angle": 0, + "content": "- a method for robust photo-realistic object reconstruction using NeRF and for robust single-shot multi-class reconstruction using LFNs" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.569, + 0.784, + 0.613 + ], + "angle": 0, + "content": "- a qualitative/quantitative evaluation of our method on both synthetic and real-world data with different inconsistencies (occlusions, invalid calibrations, ...) indicating the state-of-the-art performance of our approach" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.477, + 0.785, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.639, + 0.388, + 0.656 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Among the vast literature on neural fields, the seminal work of Mildenhall et al. [29] opened many avenues in the computer vision community. It contributed to state-of-the-art solutions for novel view synthesis and 3D reconstruction that have been covered in respective surveys [44, 51]. Noteworthy is the more recent contribution of instant neural graphics primitives (iNGP) [31], which uses a hash table of trainable feature vectors alongside a small network for representing the scene. iNGP achieved major run-time improvements, thereby enhancing the feasibility of practical applications for neural fields." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Baseline models are highly sensitive to imperfections in the input data, which led to many works on robustness enhancements of neural fields; addressing a reduced amount of input views [17,21,32,52], errors in camera parameters [4,19," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.128 + ], + "angle": 0, + "content": "B. Buschmann et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.146, + 0.785, + 0.178 + ], + "angle": 0, + "content": "[25, 53], variations in illumination conditions across observations [28, 42], multiscale image data [2, 3, 50], and the targeted removal of floating artifacts [33, 47, 48]." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.182, + 0.788, + 0.53 + ], + "angle": 0, + "content": "Fewer works solve the reconstruction task in the presence of inconsistencies between observations. Bayes' Rays [16] provides a framework to quantify uncertainty of a pretrained NeRF by approximating a spatial uncertainty field. It handles missing information due to self-occlusion or missing perspectives well, but cannot deal with inconsistencies caused by noise or distractors. Similarly, NeuRay [27] only supports missing, but not inconsistent information. Naive occlusion handling via semantic segmentation requires the occluding object types to be known in advance [28,36,37,43,46]. Solutions to learn semantic priors on transience exist [23] but separating occlusions via semantic segmentation without manual guidance is ill posed. Occ-NeRF [54] considers any foreground element as occlusion and removes them via depth reasoning, but their removal leaves behind blurry artifacts. Alternatively, some methods do not remove dynamic distractors, but reconstruct them together with the rest of the scene using time-conditioned representations [9,26,34,49]. Closely related to our work, RobustNeRF [38] considers input-image distractors as outliers of the model optimization task. The authors employ robust losses improved via patching to preserve high-frequency details. RobustNeRF does not rely on prior assumptions about the nature of distractors, nor does it require preprocessing of the input data or postprocessing of the trained model. Nevertheless, their method comes at the cost of loosing view-dependent details and a reduced reconstruction quality in undistracted scenes. Furthermore, their method is limited to unconditioned models that overfit to a single scene. A generalization to conditioned NeRFs, such as pixelNeRF [52], is not obvious, as no further optimization takes place during inference." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.534, + 0.788, + 0.701 + ], + "angle": 0, + "content": "Conditioned neural fields offer a distinct advantage in their ability to generalize to novel scenes by leveraging knowledge acquired from diverse scenes during learning. This results in a more robust model that requires as few as one input view for inference, showcasing the efficiency and adaptability of the approach. Contrary to PixelNeRF [52], which relies on a volumetric parametrization of the scene, demanding multiple network evaluations along the ray, Light Field Networks (LFNs) [40], which succeed Scene Representation Networks [41], take a different approach. LFNs represent the scene as a 4D light field, enabling a more efficient single evaluation per ray for inference. The network takes as input a ray represented in Plücker coordinates and maps it to an observed radiance, all within an autoencoder framework used for conditioning." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.705, + 0.788, + 0.842 + ], + "angle": 0, + "content": "None of the mentioned methods can deal with occlusions in single-shot reconstruction and no prior work exists on robust LFNs or robustness of other conditioned neural fields for single-shot reconstruction, which we address via the RANSAC paradigm [15]. Since its introduction in 1981, RANSAC has gained attention for fitting analytical models with a small number of parameters, such as homography estimation in panorama stitching [7]. Among the few direct applications of classical RANSAC to larger models is robust morphable face reconstruction [13, 14]. Other common expansions and applications include differentiable RANSAC [5] for camera parameter estimation in a deep learning pipeline, locally" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.663, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RANRAC" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.147, + 0.784, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.412 + ], + "angle": 0, + "content": "Fig. 2: The RANRAC algorithm for LFNs samples random hypotheses by selecting a set of random samples from the given perspective (a), and inferring the latent representation of these rays using the autodecoder of a pretrained LFN (b). The obtained light field is then used to predict an image from the input perspective (c). Based on this prediction, confidence in the random hypothesis is evaluated via the Euclidean distance between the predicted ray colors and the remaining color samples of the input image. The amount of samples which are explained by each hypothesis up to some margin are used to determine the best hypothesis (d). All samples explained by the selected hypothesis are used for a final inference with the LFN to obtain the final model and latent representation (e)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.44, + 0.788, + 0.471 + ], + "angle": 0, + "content": "optimized RANSAC [11] to account for the requirement of a descriptive sample set, and adaptive real-time RANSAC [35]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.492, + 0.33, + 0.508 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.788, + 0.658 + ], + "angle": 0, + "content": "In this section, we present our approach to increase the robustness of neural scene representations to inconsistencies in the input data. First, we recap RANSAC, its theoretical convergence and hyperparameters, and the required adaptions for its application to high-dimensional data-driven models. We then introduce a general scheme for the random sampling and validation of neural fields. Based hereon, we formulate a robust algorithm using LFNs for 3D reconstruction from a single image with occlusions. Finally, using NeRF, we formulate an algorithm for robust photo-realistic reconstruction from multiple views in the presence of common sources of inconsistencies." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.68, + 0.625, + 0.696 + ], + "angle": 0, + "content": "3.1 RANSAC Convergence on Complex Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Classical RANSAC [10,12,15] follows an iterative process. Initially, a minimal set of samples is randomly selected to determine the model parameters, known as the hypothesis generation phase. Then, the hypothesis is evaluated by assessing the number of observations it explains, within a specified margin. These steps are repeated until the best hypothesis is chosen to constitute the consensus set, which comprises all of its inliers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "This paradigm cannot be directly applied to complex models such as neural fields, as a significant amount of samples is required to obtain decent initial model parameters and additional clean samples improve the quality further." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.41, + 0.128 + ], + "angle": 0, + "content": "B. Buschmann et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "This imposes a challenge regarding the expected amount of clean initial sample sets, \\( S_{clean} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.178, + 0.786, + 0.218 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} [ \\# S _ {\\text {c l e a n}} ] = N * \\prod_ {m = 1} ^ {M} \\frac {s _ {i m g} - s _ {i m g} ^ {\\text {o c c}} - m}{s _ {i m g} - m}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.784, + 0.268 + ], + "angle": 0, + "content": "where \\( s_{img} \\) denotes the total amount of samples (e.g., image pixels), \\( s_{img}^{occ} \\) represents the occluded samples, and \\( N / M \\) are the number of iterations/samples. The expected amount exponentially decreases with the initial number of samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.785, + 0.465 + ], + "angle": 0, + "content": "The samples and respective requirements for analytic and data-driven models vary a lot. The effect of individual samples is less traceable in data-driven models and the information entropy varies more significantly across samples. When using a model that projects onto a latent space, some very atypical outliers do not show an effect at all if the latent space is not expressive enough to explain them in an overall loss-reducing way, yet, outliers close to the object or its color, or larger chunks of outliers, will usually be distracting. At the same time, samples of small-scale high-frequency details are important for the reconstruction and contain a lot of information, whereas multiple samples of larger-scale lower-frequency details contribute much less. The amount of initial samples for the hypothesis generation becomes a tunable hyperparameter trading initial reconstruction quality for likelihood of finding desired sample sets. This invalidates the classical convergence idea [10, 12] where the RANSAC iterations \\(N\\) with" + }, + { + "type": "equation", + "bbox": [ + 0.439, + 0.471, + 0.786, + 0.502 + ], + "angle": 0, + "content": "\\[\nN \\geq \\frac {\\log (1 - p)}{\\log (1 - t ^ {M})} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.788, + 0.613 + ], + "angle": 0, + "content": "are chosen such that at least one clean sample set is found with a probability \\( p \\), given the expected ratio of clean samples \\( t \\) and the amount of initial samples \\( M \\). There is not only the need to find a clean sample set, but one that captures all important details. At the same time, a completely clean sample set is not required at all, as long as the contained outliers are not represented by the local minimum of the latent space or the model itself, depending on the concrete scene representation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.634, + 0.533, + 0.65 + ], + "angle": 0, + "content": "3.2 Random Sampling Neural Fields" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.784, + 0.688 + ], + "angle": 0, + "content": "We propose a general strategy for RANSAC-like, iterative, robust reconstruction with neural fields, before formulating respective algorithms for LFNs and NeRF." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.784, + 0.779 + ], + "angle": 0, + "content": "Sampling Hypotheses Based on the application a feasible sampling domain is chosen e.g. pixels/rays, or observations. Depending on the requirements of the neural field an appropriate sample size is determined. The initial sample size is chosen as small as possible to allow for a reasonable convergence expectation to cleaner sets, while being large enough for a coarse fit of the model expressive enough to enable the discrimination of outliers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Determine Model Parameters The determination of the model parameters in a classical RANSAC application corresponds to the inference of the sample sets with a neural field: In case of unconditioned neural fields (such as NeRF) this corresponds to (over-)fitting the model to the scene, in case of conditioned fields" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.662, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RANRAC" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "this corresponds to obtaining a latent representation. Full convergence is not required, the reconstruction is only used to evaluate the outlier contamination." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.784, + 0.282 + ], + "angle": 0, + "content": "Validation of Hypotheses For the validation, the rays/perspectives not used for the inference are rendered using the obtained model and compared to the input. The quality of each hypothesis is evaluated based on the amount of other samples explained by it up to some margin. Depending on the sampling/inference strategy, different similarity metrics can be used to distinguish outliers from other sources of noise such as the coarse inference. The initial sample set of the best hypothesis together with all its inliers is used for a final complete model fit." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.784, + 0.343 + ], + "angle": 0, + "content": "As will be demonstrated in our evaluation in Section 5, the major benefit of our RANSAC-like neural field approach over formulations based on robust loss functions [1] is the possibility to reliably filter outliers and inconsistencies in the input data in comparison to the down-weighting of their influence." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.366, + 0.502, + 0.381 + ], + "angle": 0, + "content": "3.3 Robust Light Field Networks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.784, + 0.511 + ], + "angle": 0, + "content": "In the following, we propose a novel fast and robust single-shot multi-class reconstruction algorithm based on LFNs [40]. LFNs are globally conditioned, meaning that the supported subset of 3D consistent scenes is represented by a single global latent vector. Therefore, when the latent space is not expressive enough to represent object and distractor correctly, large inconsistencies cause global damage instead of local artifacts. LFNs intrinsically support parallel inference, allowing to jointly process all hypotheses rather than following an iterative scheme. Our algorithm consists of the following steps (Fig. 2):" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.524, + 0.784, + 0.598 + ], + "angle": 0, + "content": "1. Hypothesis Consensus Set: Given the input image \\(I\\) as a set of pixel color values \\(\\pmb{c}_i\\), and the intrinsic and extrinsic camera parameters, the set of rays \\(R\\) - one ray \\(\\pmb{r}_i\\) for every pixel - represented by Plücker coordinates, is generated. In the first step, \\(N\\) initial consensus sets \\(S_{n}\\) are drawn using a uniform distribution, where each consensus set consists of \\(M\\) random samples:" + }, + { + "type": "equation", + "bbox": [ + 0.375, + 0.599, + 0.785, + 0.615 + ], + "angle": 0, + "content": "\\[\n\\left(\\boldsymbol {c} _ {n} ^ {m}, \\boldsymbol {r} _ {n} ^ {m}\\right) \\in_ {R} \\left\\{\\left(\\boldsymbol {c} _ {i}, \\boldsymbol {r} _ {i}\\right) \\mid \\boldsymbol {c} _ {i} \\in I, \\boldsymbol {r} _ {i} \\in R \\right\\} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.615, + 0.784, + 0.644 + ], + "angle": 0, + "content": "where \\( n \\in \\{1, \\dots, N\\} \\), \\( m \\in \\{1, \\dots, M\\} \\), and \\( \\in_R \\) denotes a sample randomly drawn from the set without replacement according to a uniform distribution." + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.645, + 0.784, + 0.689 + ], + "angle": 0, + "content": "2. Hypothesis Inference: The autodecoder of an LFN \\(\\varPhi\\) with hypernetwork \\(\\varPsi\\) and pretrained hypernetwork weights \\(\\psi\\), is used to infer the latent codes \\(z_{n}\\) for each of the initial sample sets in parallel." + }, + { + "type": "equation", + "bbox": [ + 0.3, + 0.689, + 0.785, + 0.719 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\boldsymbol {z} _ {n} \\right\\} = \\underset {\\left\\{\\boldsymbol {z} _ {n} \\right\\}} {\\operatorname {a r g m i n}} \\sum_ {n} \\sum_ {m} \\| \\Phi \\left(\\boldsymbol {r} _ {n} ^ {m} \\mid \\Psi_ {\\psi} (\\boldsymbol {z} _ {n})\\right) - \\boldsymbol {c} _ {n} ^ {m} \\| _ {2} ^ {2} + \\lambda_ {l a t} \\| \\boldsymbol {z} _ {n} \\| _ {2} ^ {2} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.719, + 0.784, + 0.763 + ], + "angle": 0, + "content": "\\(\\lambda_{lat}\\) determines the strength of the Gaussian prior on the latent space [40]. An exponential learning rate schedule speeds up the inference. The inferred latent codes form the hypotheses." + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.763, + 0.784, + 0.84 + ], + "angle": 0, + "content": "3. Hypothesis Prediction: Each of the hypotheses is used to render an entire image \\( I_{n}^{pred} \\) from the perspective of the input image, each consisting of the pixel color values \\( \\pmb{c}_{n,i}^{pred} = \\Phi(\\pmb{r}_i \\mid \\Psi_\\psi(\\pmb{z}_n)) \\), using again the set of rays \\( R = \\{\\pmb{r}_i\\} \\) obtained from the camera parameters. The rendered pixels resemble the predictions for the remaining observations under each hypothesis." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.41, + 0.128 + ], + "angle": 0, + "content": "B. Buschmann et al." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "4. Hypothesis Validation: The obtained predictions are compared to the input image to validate the hypothesis. For each pixel in each predicted image, we calculate the Euclidean distance in color space: \\( e_{n,i} = \\| \\pmb{c}_{n,i}^{pred} - \\pmb{c}_i\\| _2 \\). For each image, using these distances, we collect, up to some margin \\( \\epsilon \\), the observations explained by the model (inliers): \\( S_{n}^{inlier} = \\{(c_{i},r_{i})\\mid e_{n,i} < \\epsilon \\} \\)" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.224, + 0.788, + 0.329 + ], + "angle": 0, + "content": "5. Model Selection: We select the best hypothesis sample set \\( S_{best} \\) based on the number of inliers \\( \\# S_n^{inlier} \\). The model is inferred once more, similar to the second step, to obtain the final latent code \\( z_{cons} \\). The inference is based on the final consensus set \\( S_{cons} = S_{best} \\cup S_{best}^{inlier} \\), the initial sample set of the strongest hypothesis \\( S_{best} \\) together with all its inliers \\( S_{best}^{inlier} \\). The final output consists of both the latent code \\( z_{cons} \\) and the final consensus set \\( S_{cons} \\) of the selected model, and can be used to render arbitrary new perspectives." + }, + { + "type": "list", + "bbox": [ + 0.219, + 0.147, + 0.788, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.354, + 0.521, + 0.368 + ], + "angle": 0, + "content": "3.4 Robust Neural Radiance Fields" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.788, + 0.502 + ], + "angle": 0, + "content": "In the following, we propose a novel robust multi-view reconstruction approach based on NeRFs. NeRFs are fit to a single scene based on a set of observations. As they support view-dependent radiance, one might expect inconsistencies in the input observations to only have an effect on specific perspectives. However, the density is only spatially parametrized, hence, inconsistencies lead to significant ghosting and smearing artifacts in more than just the inconsistent perspective. This can be leveraged in the hypothesis validation. The following steps are performed for N iterations:" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.518, + 0.788, + 0.594 + ], + "angle": 0, + "content": "1. Hypothesis Consensus Set: A NeRF fit requires a large set of rays, making a sampling in ray space infeasible, as even obtaining significantly cleaner sets becomes unlikely. In order to obtain a sampling domain with reasonably sized initial consensus sets \\( S_{n} \\), we sample \\( M \\) observations from the given sets of images \\( \\mathcal{I} \\) and corresponding camera poses \\( \\mathcal{C} \\) in every iteration:" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.594, + 0.786, + 0.61 + ], + "angle": 0, + "content": "\\[\n\\left(I _ {n} ^ {m}, E _ {n} ^ {m}\\right) \\in_ {R} \\left\\{\\left(I _ {i}, E _ {i}\\right) \\mid I _ {i} \\in \\mathcal {I}, E _ {i} \\in \\mathcal {C} \\right\\} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.61, + 0.786, + 0.639 + ], + "angle": 0, + "content": "where \\( n \\in \\{1, \\dots, N\\} \\), \\( m \\in \\{1, \\dots, M\\} \\), and \\( \\in_R \\) denotes a sample randomly drawn from the set without replacement according to a uniform distribution." + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.64, + 0.787, + 0.669 + ], + "angle": 0, + "content": "2. Hypothesis Inference: The sampled observations are used to fit a hypothesis neural radiance field \\( F_{n} \\)." + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.67, + 0.787, + 0.716 + ], + "angle": 0, + "content": "3. Hypothesis Prediction: The obtained NeRF \\( F_{n} \\) is used to render predictions \\( I_{n,i}^{pred} \\) with pixel colors \\( \\pmb{p}_{n,i,x}^{pred} \\) under the hypothesis for all unseen input perspectives \\( E_{i} \\)." + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.717, + 0.788, + 0.843 + ], + "angle": 0, + "content": "4. Hypothesis Validation: The chosen sample space requires a careful evaluation of the hypotheses. We propose a two-step evaluation, where, first, the pixels inliers \\( P_{n,i} \\) up to some margin \\( \\epsilon_{pix} \\) are determined for every observation based on the Euclidean distance in color space to the pixels \\( \\pmb{p}_{i,x} \\) of the input images \\( I_i \\): \\( P_{n,i} = \\{\\pmb{p}_{n,i,x}^{pred} \\mid \\| \\pmb{p}_{n,i,x}^{pred} - \\pmb{p}_{i,x} \\|_2 < \\epsilon_{pix}\\} \\) and, second, the observations themselves are labeled as inliers or outliers based on the amount of pixel inliers, again up to some margin \\( \\epsilon_{img} \\), to obtain the consensus set \\( S_n^{inlier} \\) under the hypothesis: \\( S_n^{inlier} = \\{(I_i, E_i) \\mid \\# P_{n,i} > \\epsilon_{img}\\} \\)." + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.64, + 0.788, + 0.843 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.662, + 0.114, + 0.732, + 0.128 + ], + "angle": 0, + "content": "RANRAC" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "The binary metric for pixels ensures that smaller mispredictions (due to, e.g., view-dependent lighting effects) do not introduce noise into the evaluation." + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.178, + 0.788, + 0.238 + ], + "angle": 0, + "content": "5. Model Selection: The strongest hypothesis, with its initial sample set \\( S_{best} \\), is selected based on the number of inliers \\( \\# S_n^{inlier} \\) to obtain the final consensus set \\( S_{cons} = S_{best} \\cup S_{best}^{inlier} \\). The final model is obtained with one more NeRF fit of the consensus set." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.26, + 0.409, + 0.275 + ], + "angle": 0, + "content": "3.5 Hyperparameters" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.286, + 0.789, + 0.436 + ], + "angle": 0, + "content": "For LFNs, the amount of initial samples and random hypotheses to evaluate (iterations), are determined experimentally. Without fine-tuning per class, the experimentally determined parameters are 90 initial samples and 2000 iterations, which supersedes the theoretical value for a convergence because the latent space introduces an intrinsic robustness. The inlier margin balances the amount of slight high-frequency variations that are being captured and the capability of separating outliers that are similar to the object. A margin of 0.25 in terms of the Euclidean distance of the predicted colors to the input samples in an RGB color space normalized to the range \\((-1, 1)\\) has been found to be optimal. Please refer to the supplementary material for further experimental results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.789, + 0.619 + ], + "angle": 0, + "content": "For NeRFs, the parameters behave more natural and the amount of perspectives required for a meaningful, not completely artifact-free fit of the model lies around 25 observations [29, Table 2]. With fewer samples, more artifacts are introduced that get harder to separate from the ones caused by inconsistencies, and the samples get more dependent on being evenly spaced. For real-world captures with \\(10\\%\\) inconsistent perspectives, as few as 50 iterations are sufficient. With the color space normalized to \\((0,1)\\), a pixel margin between around 0.15 in terms of Euclidean distance worked well for the determination of actual artifacts. We consider an observation an inlier based on a margin of \\(90\\% - 98\\%\\) pixel inliers, which proved to be a good choice to separate minor artifacts (due to the sparse sampling) from artifacts caused by actual inconsistencies. For different datasets or inconsistencies, these values can be adapted." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.642, + 0.573, + 0.66 + ], + "angle": 0, + "content": "4 Implementation & Preprocessing" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.779 + ], + "angle": 0, + "content": "For LFNs, we build on top of the original implementation [40], with a slight adaptation to enable a parallel sub-sampled inference. We furthermore use the provided pretrained multi-class model. The camera parameters are known. For efficiency reasons, the steps of the algorithm are not performed iteratively, but multiple hypotheses are validated in parallel. To further speed up the inference, an exponential learning rate schedule is used for the auto-decoding, leading to a total runtime of about a minute on a single GPU." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "For the robust reconstruction of objects from lazily captured real-world data, one has to estimate the camera parameters and extract foreground masks before applying the algorithm. For the estimation of the camera parameters, we used the COLMAP structure-from-motion package [39]. We extracted foreground masks" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.128 + ], + "angle": 0, + "content": "B. Buschmann et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "using Segment Anything [22]. However, only foreground masks, containing the objects and the occlusions, are extracted. Segment Anything is not capable of removing arbitrary occlusions in an automatized way. After these preprocessing steps, the robust reconstruction algorithm can be applied as described. Erroneous estimates of the camera parameters or foreground masks are excluded by our algorithm, thus making the entire reconstruction pipeline robust." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.239, + 0.788, + 0.375 + ], + "angle": 0, + "content": "Our algorithm is not limited to a specific NeRF implementation. The chosen sampling domain eases integration into arbitrary existing NeRF implementations, which commonly expect images instead of unstructured ray sets. However, using a fast NeRF variant is advantageous when applying an iterative scheme. We used the instant NGP implementation [31] of the instant NSR repository [18], which includes some accelerations [24, 30]. Other, (specifically fast) variants are likely good choices as well. Antialiased and unbounded, but slow variants, such as MipNeRF360 [3], are not feasible. For further implementation details please refer to the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.4, + 0.377, + 0.418 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.434, + 0.57, + 0.449 + ], + "angle": 0, + "content": "5.1 Inconsistencies,Baselines & Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.788, + 0.719 + ], + "angle": 0, + "content": "Multi-View Reconstruction (NeRF) We provide a comprehensive quantitative and qualitative analysis of RANRAC's multi-view reconstructions compared to NeRF without any method of robustness (baseline) and RobustNeRF [38] (state of the art). We conduct experiments for various common sources of inconsistencies such as occluded perspectives, noisy camera parameter estimates, and blurred perspectives. RobustNeRF [38] targets unbounded scenes with multiple objects and small amounts of distractors in every perspective. In comparison, our RANSAC-based approach deals well with single-object reconstruction, even with heavy occlusions, as long as enough clean perspectives are available. As their dataset reflects the algorithm's properties, we cannot provide a fair comparison. Instead, we demonstrate the applicability using a custom dataset of a single object with a controlled amount of deliberately occluded perspectives. For the other inconsistencies we use of-the-shelf datasets and add noise to the camera parameters and blur to the images. Furthermore, we implement the robust losses [38] on the same NeRF variant as RANRAC to provide a fair comparison of the method's robustness, independent of the NeRF variant. For further details on the implementation, please refer to the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Single-Shot Reconstruction (LFN) We benchmark against the original LFN implementation of Sitzmann et al. [40] as baseline, as there are no other robust methods for LFNs or conditioned neural fields, nor are there methods for robust single-shot multi-class reconstruction in general. Furthermore, we use the same pretrained LFN for the baseline and for the application of our method. The LFN is pretrained on the thirteen largest ShapeNet classes [8]." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "We provide a detailed qualitative and quantitative performance comparison under different amounts of occlusion for three representative classes (plane, car," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.662, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RANRAC" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "and chair), while just stating reconstruction performance in a fixed environment without additional tuning of the hyperparameters for the others. The plane class is mostly challenging due to the low-frequency shape, while the car class contains a lot of high-frequency color details. The chair class represents shapes that are generally problematic for vanilla LFNs, even without occlusions. We provide a complementing analysis of the hyperparameters in the supplementary material. If not stated otherwise, we evaluate using 50 randomly selected images of the corresponding class. All comparisons use the same images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.788, + 0.374 + ], + "angle": 0, + "content": "The occlusions are created synthetically. They consist of randomly generated patches while controlling two metrics of occlusion: Image occlusion and object occlusion. The former is the naive ratio of occluded over total pixels. The latter are the occluded pixels on the object compared to the total pixels covered by the object. We use both metrics to take the vastly different information entropy of samples across the image into account. For further details on the generation of the synthetic occlusions, please refer to the supplemental." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.395, + 0.352, + 0.409 + ], + "angle": 0, + "content": "5.2 Evaluation" + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.424, + 0.779, + 0.479 + ], + "angle": 0, + "content": "
MetricModelBenchBoatCabin.Displ.LampPhoneRifleSofaSpeak.Table
PSNR↑RANRAC19.2122.9221.9217.6519.9918.4521.3521.6120.720.44
LFN17.8919.220.518.8519.0917.8118.4620.1219.9920.29
SSIM↑RANRAC0.7670.8580.8010.6990.7640.750.8530.8050.7610.784
LFN0.7240.7910.7670.730.7480.7260.7950.7750.7430.777
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.48, + 0.788, + 0.59 + ], + "angle": 0, + "content": "Table 1: RANRAC obtains a significant quantitative improvement in PSNR and SSIM (higher is better) compared to the baseline. We compare RANRAC to vanilla LFNs for the 13 largest ShapeNet classes (find plane, car, and chair with more detail in Fig. 3). The results are based on a moderate amount of occlusion of approximately \\(25\\%\\) object occlusion and about \\(5\\%\\) image occlusion. The reported results are conservative, as higher amounts of image occlusion result in a more significant performance increase (Fig. 3). No hyperparameter tuning has been performed for these classes; the same configuration obtained from the analysis of the other three classes is used." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.789, + 0.841 + ], + "angle": 0, + "content": "For LFNs, our approach leads to a significant improvement in occluded scenarios of up to 8dB in PSNR and a similarly strong improvement for the SSIM. The improvement is most significant in heavily distracted scenarios (Fig. 3). In clean scenarios a slight performance penalty can be observed, but even with small amounts of object occlusion (information loss), our algorithm outperforms the baseline, leading to numerically better results up to \\(50\\%\\) information loss (Fig. 3). The effect is not only measurable, but also well visible (Fig. 4). Increasing amounts of occlusion slowly introduce local artifacts into our reconstruction while preserving a reasonable shape estimate even for larger amounts of occlusion. In contrast, the reconstruction of LFNs breaks rather early in a global fashion. Still, our consensus set (Fig. 4), reveals that some high-frequency details were wrongfully excluded, explaining the slight performance decrease on clean images. In general, the benefit of RANRAC is best observable for classes that can be well described by LFNs, as evidenced by the lower improvements the chair class exhibits, compared to the plane and car class. The same effect is visible in the quantitative evaluation on the other classes without additional" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.127 + ], + "angle": 0, + "content": "B. Buschmann et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.151, + 0.787, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.304, + 0.789, + 0.499 + ], + "angle": 0, + "content": "Fig. 3: RANRAC (solid lines) leads to a quantitative improvement in PSNR and SSIM (higher is better) for occluded inputs compared to vanilla LFNs (dashed lines). The same hyperparameter configuration and LFN is used for all classes. On the left and in the middle, the amount of image occlusion is increased, while the object occlusion is constant at \\(25\\%\\). On the right, the amount of object occlusion is increased while the image occlusion is kept low. For the car class, a large improvement is observed over the entire occlusion spectrum. For the plane class the improvement is similarly significant, but absolute performance degenerates a bit sooner. This stems from the smaller object size and the related faster occlusion-to-object increase when increasing image occlusions. For the chair class, the improvement is less significant but the structural similarity is preserved for much longer. For the plane and car class the reconstruction quality is resilient to information loss (right) up to \\(\\sim 50\\%\\), where the decrease gains momentum. With the low amounts of image occlusion, the improvement is not significant for the chair class (consistent with left and middle)." + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.517, + 0.773, + 0.723 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.733, + 0.788, + 0.832 + ], + "angle": 0, + "content": "Fig. 4: On the left, we show the qualitative effect of increasing occlusion on the same observation for the reconstruction of a novel view. Reconstructions of LFNs break early globally whereas RANRAC still provides a very decent reconstruction, only slowly introducing minor local (and natural/comprehensible) artifacts for completely hidden object parts. We further show the obtained consensus set, used for the final reconstruction (green inliers, red outliers). On the right, we show more qualitative results for novel view synthesis on different classes and the corresponding consensus sets." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.663, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RANRAC" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.158, + 0.77, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.459, + 0.788, + 0.503 + ], + "angle": 0, + "content": "Fig. 5: The occlusions lead to well-visible artifacts in the reconstructions of NeRF, these artifacts are completely removed by RANRAC. While RobustNeRF struggles with view-dependent and high-frequency details, RANRAC reliably reconstructs them." + }, + { + "type": "table", + "bbox": [ + 0.237, + 0.541, + 0.765, + 0.584 + ], + "angle": 0, + "content": "
Inconsistency DatasetMild Occ.Strong Occ.Watering PotBlurred Perspectives
LegoShipChairFicusMic
RANRAC27.1125.9926.1224.9434.7929.9129.7616.1435.2531.26
RobustNeRF26.8325.5825.9324.7929.1423.6923.3120.1931.1127.12
NeRF26.6522.6125.3618.4731.1519.0028.4820.3533.2122.10
" + }, + { + "type": "table", + "bbox": [ + 0.237, + 0.588, + 0.765, + 0.629 + ], + "angle": 0, + "content": "
Inconsistency \nDatasetNoisy Camera Parameters
LegoDrumsMicShipFicusHotdogMaterials
RANRAC34.95|31.2225.88|22.7735.85|33.5530.77|22.0531.41|28.8237.16|30.5328.93|25.39
RobustNeRF29.83|26.8223.91|22.1029.36|27.5524.21|19.9823.17|21.8532.65|25.1624.90|22.30
NeRF24.67|15.9523.22|17.9628.82|22.1523.68|14.8028.12|22.7728.16|18.3724.73|19.89
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Table 2: RANRAC outperforms the state of the art RobustNeRF [38] and the baseline NeRF (without a method of robustness) for scenes contaminated with occlusions, blur and noisy camera parameters. We report PSNR\\(\\uparrow\\) averaged across perspectives and the 5th percentile \\((\\mathrm{Avg.}|P_5)\\) as artifacts introduced by inconsistencies only contaminate some views. We compare on the captured watering pot dataset with milder \\((10\\%)\\) and heavier \\((17.5\\%)\\) amounts of occluded perspectives and off-the-shelf datasets [29] with blurred perspectives and additive Gaussian noise \\(\\mathcal{N}(5^{\\circ},1^{\\circ})\\) on the camera parameters of \\(10\\%\\) of the perspectives. All three variants are built on top of instant-nsr for an isolated fair comparison of the robustness method. Note that both robust approaches struggle separating the strong view-dependent effects of the ship scene from blur, leading to the exclusion of some perspectives (lower \\(P_5\\) PSNR), while RANRAC still improves the overall reconstruction. For all other scenes and inconsistencies, RANRAC reliably separates inconsistent observations from clean ones." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.128 + ], + "angle": 0, + "content": "B. Buschmann et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.207 + ], + "angle": 0, + "content": "tuning (Tab. 1). The only outlier is the display class, on which LFNs struggle the most on [40], even in unoccluded scenarios. It is not consistently represented in the latent space and robust reconstruction amplifies this effect by reducing the samples to a consistent set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.209, + 0.787, + 0.314 + ], + "angle": 0, + "content": "The application to NeRF shows the significantly improved reconstructions of RANRAC (Tab. 2) for different types of inconsistencies such as occlusions and blur, as well as noisy camera parameters on captured and synthetic datasets. RANRAC consistently outperforms RobustNeRF [38] for object reconstruction from inconsistent inputs. RobustNeRF specifically struggles in capturing view-dependent appearance and at concavities while RANRAC seamlessly reconstructs them. The effects are not only measurable but also well-visible (Fig. 5)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.343, + 0.519, + 0.359 + ], + "angle": 0, + "content": "6 Limitations & Future Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.783, + 0.44 + ], + "angle": 0, + "content": "For the application to NeRF, the iterations imply the use of a fast inferable variant, ruling out MipNeRF360 [3] and other high-quality variants for unbounded scenes. Foreground separation is a must, limiting this application of RANRAC to object-centric scenes. Novel fast, unbounded methods might resolve this [20]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.442, + 0.784, + 0.516 + ], + "angle": 0, + "content": "By following a RANSAC-like approach, we inherit the requirement of sufficient clean perspectives, which could be lifted via NeRF variants that require fewer perspectives [32, 45] or by using different sampling domains. In return, our method is not limited to specific kinds of inconsistencies, and is robust to arbitrarily heavy distractions or inconsistencies in the impure perspectives." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.787, + 0.624 + ], + "angle": 0, + "content": "Whereas our robust LFN approach led to significant improvements for single-shot reconstruction, our concept might also be applied to other conditioned neural fields, based on a smart choice of the sampling domain, targeting photorealism. One could also use importance sampling based on a prior, instead of a uniform sampling, leveraging the unevenly distributed information entropy. Neural sampling priors [6], semantic segmentation and more advanced schemes (e.g. LO-RANSAC [11], DSAC [5]) might prove useful here." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.652, + 0.36, + 0.669 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We introduced a novel approach to increase the robustness of neural fields, inspired by the RANSAC paradigm. Following this concept, we introduced a novel robust approach for single-shot reconstruction from occluded views based on LFNs which achieves a significant improvement in reconstruction quality for distracted and occluded scenarios, even for extreme cases. Furthermore, we introduced a respective RANRAC-based NeRF variant that allows robust photorealistic reconstruction from multiple views with typical inconsistencies such as occlusions, noisy camera parameters, or blurred images - resulting in significant improvements compared to state-of-the-art methods - without relying on assumptions about the distractions." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.662, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RANRAC" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.148, + 0.788, + 0.207 + ], + "angle": 0, + "content": "Acknowledgements - Benno Buschmann was at FAU funded through a gift by Mitsubishi Electric Research Laboratories (MERL). Andreea Dogaru was funded by the German Federal Ministry of Education and Research (BMBF), FKZ: 01IS22082 (IRRW). The authors are responsible for the content of this publication." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.23, + 0.323, + 0.246 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.26, + 0.786, + 0.302 + ], + "angle": 0, + "content": "1. Barron, J.T.: A general and adaptive robust loss function. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4331-4339 (2019) 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.302, + 0.787, + 0.343 + ], + "angle": 0, + "content": "2. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: ICCV (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.344, + 0.787, + 0.372 + ], + "angle": 0, + "content": "3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In: CVPR (2022) 4, 10, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.372, + 0.787, + 0.399 + ], + "angle": 0, + "content": "4. Bian, W., Wang, Z., Li, K., Bian, J.W., Prisacariu, V.A.: Nope-nerf: Optimising neural radiance field with no pose prior. In: CVPR (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.399, + 0.787, + 0.44 + ], + "angle": 0, + "content": "5. Brachmann, E., Krull, A., Nowozin, S., Shotton, J., Michel, F., Gumhold, S., Rother, C.: Dsac-differentiable ransac for camera localization. In: CVPR (2017) 4, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.44, + 0.787, + 0.469 + ], + "angle": 0, + "content": "6. Brachmann, E., Rother, C.: Neural-guided ransac: Learning where to sample model hypotheses. In: ICCV (2019) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.469, + 0.787, + 0.496 + ], + "angle": 0, + "content": "7. Brown, M., Lowe, D.G.: Automatic panoramic image stitching using invariant features. International Journal of Computer Vision 74, 59-73 (2007) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.496, + 0.787, + 0.564 + ], + "angle": 0, + "content": "8. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., Xiao, J., Yi, L., Yu, F.: ShapeNet: An Information-Rich 3D Model Repository. Tech. Rep. arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago (2015) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.787, + 0.592 + ], + "angle": 0, + "content": "9. Chen, X., Zhang, Q., Li, X., Chen, Y., Feng, Y., Wang, X., Wang, J.: Hallucinated neural radiance fields in the wild. In: CVPR (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.787, + 0.62 + ], + "angle": 0, + "content": "0. Choi, S., Kim, T., Yu, W.: Performance evaluation of ransac family. Journal of Computer Vision 24(3), 271-300 (1997) 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.787, + 0.661 + ], + "angle": 0, + "content": "1. Chum, O., Matas, J., Kittler, J.: Locally optimized ransac. In: Pattern Recognition: 25th DAGM Symposium, Magdeburg, Germany, September 10-12, 2003. Proceedings 25. pp. 236-243. Springer (2003) 5, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.787, + 0.689 + ], + "angle": 0, + "content": "2. Derpanis, K.G.: Overview of the ransac algorithm. Image Rochester NY 4(1), 2-3 (2010) 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.787, + 0.716 + ], + "angle": 0, + "content": "3. Egger, B., Schneider, A., Blumer, C., Forster, A., Schonborn, S., Vetter, T.: Occlusion-aware 3d morphable face models. In: BMVC. vol. 2, p. 4 (2016) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.787, + 0.771 + ], + "angle": 0, + "content": "4. Egger, B., Schonborn, S., Schneider, A., Kortylewski, A., Morel-Forster, A., Blumer, C., Vetter, T.: Occlusion-aware 3d morphable models and an illumination prior for face image analysis. International Journal of Computer Vision 126, 1269-1287 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.787, + 0.812 + ], + "angle": 0, + "content": "5. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM 24(6), 381-395 (1981) 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.812, + 0.787, + 0.841 + ], + "angle": 0, + "content": "6. Goli, L., Reading, C., Sellan, S., Jacobson, A., Tagliasacchi, A.: Bayes' Rays: Uncertainty quantification in neural radiance fields. arXiv (2023) 4" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.26, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.128 + ], + "angle": 0, + "content": "B. Buschmann et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "17. Guangcong, Chen, Z., Loy, C.C., Liu, Z.: Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. ICCV (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.203 + ], + "angle": 0, + "content": "18. Guo, Y.C.: Instant neural surface reconstruction (2022), https://github.com/bennyguo/instant-nsr-pl 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "19. Jeong, Y., Ahn, S., Choy, C., Anandkumar, A., Cho, M., Park, J.: Self-calibrating neural radiance fields. In: ICCV (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "20. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering (2023) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.286 + ], + "angle": 0, + "content": "21. Kim, M., Seo, S., Han, B.: Infonerf: Ray entropy minimization for few-shot neural volume rendering. In: CVPR (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "22. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.356 + ], + "angle": 0, + "content": "23. Lee, J., Kim, I., Heo, H., Kim, H.J.: Semantic-aware occlusion filtering neural radiance fields in the wild. arXiv preprint arXiv:2303.03966 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.383 + ], + "angle": 0, + "content": "24. Li, R., Gao, H., Tancik, M., Kanazawa, A.: Nerfacc: Efficient sampling accelerates nerfs. arXiv preprint arXiv:2305.04966 (2023) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "25. Lin, C.H., Ma, W.C., Torralba, A., Lucey, S.: Barf: Bundle-adjusting neural radiance fields. In: ICCV (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.438 + ], + "angle": 0, + "content": "26. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust dynamic radiance fields. In: CVPR (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.466 + ], + "angle": 0, + "content": "27. Liu, Y., Peng, S., Liu, L., Wang, Q., Wang, P., Christian, T., Zhou, X., Wang, W.: Neural rays for occlusion-aware image-based rendering. In: CVPR (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "28. Martin-Brualla, R., Radwan, N., Sajjadi, M.S.M., Barron, J.T., Dosovitskiy, A., Duckworth, D.: NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In: CVPR (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "29. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) 2, 3, 9, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.551, + 0.785, + 0.563 + ], + "angle": 0, + "content": "30. Müller, T.: tiny-cuda-nn (4 2021), https://github.com/NVlabs/tiny-cuda-nn 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.591 + ], + "angle": 0, + "content": "31. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. In: SIGGRAPH (2022) 3, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.632 + ], + "angle": 0, + "content": "32. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S.M., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: CVPR (2022) 3, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.674 + ], + "angle": 0, + "content": "33. Philip, J., Deschaintre, V.: Floaters no more: Radiance field gradient scaling for improved near-camera training. In: Eurographics Symposium on Rendering. The Eurographics Association (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "34. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: CVPR (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.741 + ], + "angle": 0, + "content": "35. Raguram, R., Frahm, J.M., Pollefeys, M.: A comparative analysis of ransac techniques leading to adaptive real-time random sample consensus. In: ECCV (2008) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.771 + ], + "angle": 0, + "content": "36. Rebain, D., Matthews, M., Yi, K.M., Lagun, D., Tagliasacchi, A.: Lolnerf: Learn from one look. In: CVPR (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.798 + ], + "angle": 0, + "content": "37. Rematas, K., Liu, A., Srinivasan, P.P., Barron, J.T., Tagliasacchi, A., Funkhouser, T., Ferrari, V.: Urban radiance fields. CVPR (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.838 + ], + "angle": 0, + "content": "38. Sabour, S., Vora, S., Duckworth, D., Krasin, I., Fleet, D.J., Tagliasacchi, A.: Robustnerf: Ignoring distractors with robust losses. In: CVPR (2023) 2, 3, 4, 10, 13, 14" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.838 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.662, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RANRAC" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.174 + ], + "angle": 0, + "content": "39. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: CVPR (2016) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.787, + 0.217 + ], + "angle": 0, + "content": "40. Sitzmann, V., Rezchikov, S., Freeman, W.T., Tenenbaum, J.B., Durand, F.: Light field networks: Neural scene representations with single-evaluation rendering. In: NeurIPS (2021) 2, 3, 4, 7, 9, 10, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.245 + ], + "angle": 0, + "content": "41. Sitzmann, V., Zollhöfer, M., Wetzstein, G.: Scene representation networks: Continuous 3d-structure-aware neural scene representations. In: NeurIPS (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.273 + ], + "angle": 0, + "content": "42. Sun, J., Chen, X., Wang, Q., Li, Z., Averbuch-Elor, H., Zhou, X., Snavely, N.: Neural 3D reconstruction in the wild. In: SIGGRAPH (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.314 + ], + "angle": 0, + "content": "43. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: CVPR (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "44. Tewari, A., Thies, J., Mildenhall, B., et al.: Advances in neural rendering. In: Comput. Graph. Forum. vol. 41, pp. 703-735. Wiley Online Library, Wiley (5 2022) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.397 + ], + "angle": 0, + "content": "45. Truong, P., Rakotosaona, M.J., Manhardt, F., Tombari, F.: Sparf: Neural radiance fields from sparse and noisy poses. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4190-4200 (2023) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "46. Turki, H., Ramanan, D., Satyanarayanan, M.: Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In: CVPR (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.452 + ], + "angle": 0, + "content": "47. Warburg*, F., Weber*, E., Tancik, M., Holski, A., Kanazawa, A.: Nerfbusters: Removing ghostly artifacts from casually captured nerfs. In: ICCV (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.493 + ], + "angle": 0, + "content": "48. Wirth, T., Rak, A., Knauthe, V., Fellner, D.W.: A post processing technique to automatically remove floater artifacts in neural radiance fields. In: Computer Graphics Forum (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.785, + 0.534 + ], + "angle": 0, + "content": "49. Wu, T., Zhong, F., Tagliasacchi, A., Cole, F., Oztireli, C.: D\\~ 2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. NeurIPS (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.535, + 0.785, + 0.577 + ], + "angle": 0, + "content": "50. Xiangli, Y., Xu, L., Pan, X., Zhao, N., Rao, A., Theobalt, C., Dai, B., Lin, D.: Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In: ECCV (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.618 + ], + "angle": 0, + "content": "51. Xie, Y., Takikawa, T., Saito, S., Litany, O., Yan, S., Khan, N., Tombari, F., Tompkin, J., Sitzmann, V., Sridhar, S.: Neural fields in visual computing and beyond. Computer Graphics Forum 41 (2021) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.646 + ], + "angle": 0, + "content": "52. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: CVPR (2021) 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.674 + ], + "angle": 0, + "content": "53. Zhang, J.Y., Ramanan, D., Tulsiani, S.: Relpose: Predicting probabilistic relative rotation for single objects in the wild. In: ECCV (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "54. Zhu, C., Wan, R., Tang, Y., Shi, B.: Occlusion-free scene recovery via neural radiance fields. In: CVPR (2023) 4" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.787, + 0.702 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/RANRAC_ Robust Neural Scene Representations via Random Ray Consensus/de9512b1-9e5d-48ed-a48f-d5995f4754ec_origin.pdf b/2024/RANRAC_ Robust Neural Scene Representations via Random Ray Consensus/de9512b1-9e5d-48ed-a48f-d5995f4754ec_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0c2754248e0f803bbcc51141a2372fa0cc2515f9 --- /dev/null +++ b/2024/RANRAC_ Robust Neural Scene Representations via Random Ray Consensus/de9512b1-9e5d-48ed-a48f-d5995f4754ec_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:29e560ad430e7d9fff3817ef6163e61e49c08381adc2865e759903e2519a6d92 +size 8075002 diff --git a/2024/RANRAC_ Robust Neural Scene Representations via Random Ray Consensus/full.md b/2024/RANRAC_ Robust Neural Scene Representations via Random Ray Consensus/full.md new file mode 100644 index 0000000000000000000000000000000000000000..47fb539d7d90b23d6dbfee92a23e6129d16efba4 --- /dev/null +++ b/2024/RANRAC_ Robust Neural Scene Representations via Random Ray Consensus/full.md @@ -0,0 +1,266 @@ +# RANRAC: Robust Neural Scene Representations via Random Ray Consensus + +Benno Buschmann1,2, Andreea Dogaru2, Elmar Eisemann1, Michael Weinmann1, and Bernhard Egger2 + +1 Delft University of Technology, Netherlands + +$^{2}$ Friedrich-Alexander-University Erlangen-Nürnberg, Germany + +Abstract. Learning-based scene representations such as neural radiance fields or light field networks, that rely on fitting a scene model to image observations, commonly encounter challenges in the presence of inconsistencies within the images caused by occlusions, inaccurately estimated camera parameters or effects like lens flare. To address this challenge, we introduce RANdom RAy Consensus (RANRAC), an efficient approach to eliminate the effect of inconsistent data, thereby taking inspiration from classical RANSAC based outlier detection for model fitting. In contrast to the down-weighting of the effect of outliers based on robust loss formulations, our approach reliably detects and excludes inconsistent perspectives, resulting in clean images without floating artifacts. For this purpose, we formulate a fuzzy adaption of the RANSAC paradigm, enabling its application to large scale models. We interpret the minimal number of samples to determine the model parameters as a tunable hyperparameter, investigate the generation of hypotheses with data-driven models, and analyse the validation of hypotheses in noisy environments. We demonstrate the compatibility and potential of our solution for both photo-realistic robust multi-view reconstruction from real-world images based on neural radiance fields and for single-shot reconstruction based on light-field networks. In particular, the results indicate significant improvements compared to state-of-the-art robust methods for novel-view synthesis on both synthetic and captured scenes with various inconsistencies including occlusions, noisy camera pose estimates, and unfocused perspectives. The results further indicate significant improvements for single-shot reconstruction from occluded images. + +Keywords: Neural scene representations $\cdot$ neural rendering $\cdot$ RANSAC $\cdot$ robust estimation $\cdot$ neural radiance fields $\cdot$ light-field networks + +# 1 Introduction + +3D reconstruction is a classical task in computer vision and computer graphics, which has attracted research for decades. It offers numerous applications, including autonomous systems, entertainment, design, advertisement, cultural heritage, VR/AR experiences or medical scenarios. In recent years, neural scene representations and rendering techniques [44,51], including light field networks + +![](images/2aee6ce2116d1ebbcb2915d1807117e82963fb404ad9dcd39d13887a5e328f6b.jpg) +Fig. 1: We propose a robust algorithm for 3D reconstruction from occluded input perspectives that is based on the random sampling of hypotheses. Our algorithm is general and we demonstrate the use for single-shot reconstruction using light field networks or multi-view reconstruction using NeRF. In these cases, it successfully removes the artifacts that normally occur due to occluded input perspectives. + +(LFN) [40] and neural radiance fields (NeRF) [29] have demonstrated great performance in single-view and multi-view reconstruction tasks. The key to the success of such techniques is the coupling of differentiable rendering methods with custom neural field parametrizations of scene properties. However, a common limitation of neural scene reconstruction methods is their sensitivity to inconsistencies among input images induced by occlusions, inaccurately estimated camera parameters or effects like lens flares. Despite the use of view-dependent radiance representations to address view-dependent appearance changes, these inconsistencies severely impact local density estimation, resulting in a poor generalization to novel views. + +To increase the robustness to potential distractors within the training data, Sabour et al. [38] recently introduced the use of robust losses in the context of training unconditioned NeRF, where distractors in the training data were modeled as outliers of an optimization problem. However, the adaptation of this approach to conditioned neural fields (e.g., pixelNeRF [52]) is not obvious, as no optimization takes place during inference, and the input data is constrained to only a few views. Achieving robustness to data inconsistencies is a well-analyzed problem in computer vision, covered not only by the aforementioned robust loss functions [1], but also other strategies, like the random sample consensus (RANSAC) paradigm [15]. The latter is widely employed for fitting models to outlier-heavy data. The underlying idea is to randomly select subsets of the data to form potential models, evaluating these models against the entire dataset, and identifying the subset that best fits the majority of the data, while disregarding outliers. Despite being the state-of-the-art solution to many challenges, RANSAC-based schemes are particularly favoured for the fitting of analytical models with a relatively small amount of tuneable parameters. In this paper, we direct our attention to achieving robustness against inconsistencies and occlusions in the observations by using a novel combination of neural scene representation and rendering techniques with dedicated outlier removal techniques such as RANSAC [15]. While downweighting the influence of distractors based on robust losses [1, 38] can affect clean samples, representing details, we aim at improving robustness to distractors by only removing the influence of + +outliers. To this extent, we integrate a RANSAC-based scheme to distinguish inliers and outliers in the data and the inlier-based optimization of the neural fields (Fig. 1); a stochastic scenario characterized by a large-scale, data-driven model that exceeds RANSAC's classical convergence expectations. Instead of guaranteeing convergence to a clean sample set based on a minimal number of samples, we aim for a feasible (cleaner) sample set using a tuneable amount of samples. The proposed algorithm exhibits robustness and versatility, accommodating a wide range of neural fields-based reconstruction methods. + +Our method inherits the strengths of RANSAC, such as the ability to handle various classes of outliers without relying on semantics. Yet, it also inherits the need for sufficiently clean samples and the reliance on an iterative scheme. In practice, the first condition is often fulfilled because typically only some of the perspectives are affected by inconsistencies. We validate our approach using synthetic data, focusing on the task of multi-class single-shot reconstruction with LFNs [40], and observe significant quality improvements over the baseline in the presence of occlusions. Furthermore, we showcase robust photo-realistic reconstructions of 3D objects using unconditioned NeRFs from sequences of real-world images in the presence of distractors. In comparison to RobustNeRF [38], we use all available clean data, hence improving the reconstruction quality for single-object scenes. Code and data are available under https://bennobuschmann.com/ranrac/. Our key contributions are: + +- a general, robust RANSAC-based reconstruction method applicable to a variety of neural-fields and handling diverse inconsistencies +- an analysis of the implication to RANSAC's hyperparameters and theoretical convergence expectations, and the experimental study of their effect +- a method for robust photo-realistic object reconstruction using NeRF and for robust single-shot multi-class reconstruction using LFNs +- a qualitative/quantitative evaluation of our method on both synthetic and real-world data with different inconsistencies (occlusions, invalid calibrations, ...) indicating the state-of-the-art performance of our approach + +# 2 Related Work + +Among the vast literature on neural fields, the seminal work of Mildenhall et al. [29] opened many avenues in the computer vision community. It contributed to state-of-the-art solutions for novel view synthesis and 3D reconstruction that have been covered in respective surveys [44, 51]. Noteworthy is the more recent contribution of instant neural graphics primitives (iNGP) [31], which uses a hash table of trainable feature vectors alongside a small network for representing the scene. iNGP achieved major run-time improvements, thereby enhancing the feasibility of practical applications for neural fields. + +Baseline models are highly sensitive to imperfections in the input data, which led to many works on robustness enhancements of neural fields; addressing a reduced amount of input views [17,21,32,52], errors in camera parameters [4,19, + +[25, 53], variations in illumination conditions across observations [28, 42], multiscale image data [2, 3, 50], and the targeted removal of floating artifacts [33, 47, 48]. + +Fewer works solve the reconstruction task in the presence of inconsistencies between observations. Bayes' Rays [16] provides a framework to quantify uncertainty of a pretrained NeRF by approximating a spatial uncertainty field. It handles missing information due to self-occlusion or missing perspectives well, but cannot deal with inconsistencies caused by noise or distractors. Similarly, NeuRay [27] only supports missing, but not inconsistent information. Naive occlusion handling via semantic segmentation requires the occluding object types to be known in advance [28,36,37,43,46]. Solutions to learn semantic priors on transience exist [23] but separating occlusions via semantic segmentation without manual guidance is ill posed. Occ-NeRF [54] considers any foreground element as occlusion and removes them via depth reasoning, but their removal leaves behind blurry artifacts. Alternatively, some methods do not remove dynamic distractors, but reconstruct them together with the rest of the scene using time-conditioned representations [9,26,34,49]. Closely related to our work, RobustNeRF [38] considers input-image distractors as outliers of the model optimization task. The authors employ robust losses improved via patching to preserve high-frequency details. RobustNeRF does not rely on prior assumptions about the nature of distractors, nor does it require preprocessing of the input data or postprocessing of the trained model. Nevertheless, their method comes at the cost of loosing view-dependent details and a reduced reconstruction quality in undistracted scenes. Furthermore, their method is limited to unconditioned models that overfit to a single scene. A generalization to conditioned NeRFs, such as pixelNeRF [52], is not obvious, as no further optimization takes place during inference. + +Conditioned neural fields offer a distinct advantage in their ability to generalize to novel scenes by leveraging knowledge acquired from diverse scenes during learning. This results in a more robust model that requires as few as one input view for inference, showcasing the efficiency and adaptability of the approach. Contrary to PixelNeRF [52], which relies on a volumetric parametrization of the scene, demanding multiple network evaluations along the ray, Light Field Networks (LFNs) [40], which succeed Scene Representation Networks [41], take a different approach. LFNs represent the scene as a 4D light field, enabling a more efficient single evaluation per ray for inference. The network takes as input a ray represented in Plücker coordinates and maps it to an observed radiance, all within an autoencoder framework used for conditioning. + +None of the mentioned methods can deal with occlusions in single-shot reconstruction and no prior work exists on robust LFNs or robustness of other conditioned neural fields for single-shot reconstruction, which we address via the RANSAC paradigm [15]. Since its introduction in 1981, RANSAC has gained attention for fitting analytical models with a small number of parameters, such as homography estimation in panorama stitching [7]. Among the few direct applications of classical RANSAC to larger models is robust morphable face reconstruction [13, 14]. Other common expansions and applications include differentiable RANSAC [5] for camera parameter estimation in a deep learning pipeline, locally + +![](images/92f72ab285e6d711cd62058e8ff29bf00bd4f204d2952d6b976d9b2b5395fa69.jpg) +Fig. 2: The RANRAC algorithm for LFNs samples random hypotheses by selecting a set of random samples from the given perspective (a), and inferring the latent representation of these rays using the autodecoder of a pretrained LFN (b). The obtained light field is then used to predict an image from the input perspective (c). Based on this prediction, confidence in the random hypothesis is evaluated via the Euclidean distance between the predicted ray colors and the remaining color samples of the input image. The amount of samples which are explained by each hypothesis up to some margin are used to determine the best hypothesis (d). All samples explained by the selected hypothesis are used for a final inference with the LFN to obtain the final model and latent representation (e). + +optimized RANSAC [11] to account for the requirement of a descriptive sample set, and adaptive real-time RANSAC [35]. + +# 3 Method + +In this section, we present our approach to increase the robustness of neural scene representations to inconsistencies in the input data. First, we recap RANSAC, its theoretical convergence and hyperparameters, and the required adaptions for its application to high-dimensional data-driven models. We then introduce a general scheme for the random sampling and validation of neural fields. Based hereon, we formulate a robust algorithm using LFNs for 3D reconstruction from a single image with occlusions. Finally, using NeRF, we formulate an algorithm for robust photo-realistic reconstruction from multiple views in the presence of common sources of inconsistencies. + +# 3.1 RANSAC Convergence on Complex Models + +Classical RANSAC [10,12,15] follows an iterative process. Initially, a minimal set of samples is randomly selected to determine the model parameters, known as the hypothesis generation phase. Then, the hypothesis is evaluated by assessing the number of observations it explains, within a specified margin. These steps are repeated until the best hypothesis is chosen to constitute the consensus set, which comprises all of its inliers. + +This paradigm cannot be directly applied to complex models such as neural fields, as a significant amount of samples is required to obtain decent initial model parameters and additional clean samples improve the quality further. + +This imposes a challenge regarding the expected amount of clean initial sample sets, $S_{clean}$ : + +$$ +\mathbb {E} [ \# S _ {\text {c l e a n}} ] = N * \prod_ {m = 1} ^ {M} \frac {s _ {i m g} - s _ {i m g} ^ {\text {o c c}} - m}{s _ {i m g} - m}, \tag {1} +$$ + +where $s_{img}$ denotes the total amount of samples (e.g., image pixels), $s_{img}^{occ}$ represents the occluded samples, and $N / M$ are the number of iterations/samples. The expected amount exponentially decreases with the initial number of samples. + +The samples and respective requirements for analytic and data-driven models vary a lot. The effect of individual samples is less traceable in data-driven models and the information entropy varies more significantly across samples. When using a model that projects onto a latent space, some very atypical outliers do not show an effect at all if the latent space is not expressive enough to explain them in an overall loss-reducing way, yet, outliers close to the object or its color, or larger chunks of outliers, will usually be distracting. At the same time, samples of small-scale high-frequency details are important for the reconstruction and contain a lot of information, whereas multiple samples of larger-scale lower-frequency details contribute much less. The amount of initial samples for the hypothesis generation becomes a tunable hyperparameter trading initial reconstruction quality for likelihood of finding desired sample sets. This invalidates the classical convergence idea [10, 12] where the RANSAC iterations $N$ with + +$$ +N \geq \frac {\log (1 - p)}{\log (1 - t ^ {M})} \tag {2} +$$ + +are chosen such that at least one clean sample set is found with a probability $p$ , given the expected ratio of clean samples $t$ and the amount of initial samples $M$ . There is not only the need to find a clean sample set, but one that captures all important details. At the same time, a completely clean sample set is not required at all, as long as the contained outliers are not represented by the local minimum of the latent space or the model itself, depending on the concrete scene representation. + +# 3.2 Random Sampling Neural Fields + +We propose a general strategy for RANSAC-like, iterative, robust reconstruction with neural fields, before formulating respective algorithms for LFNs and NeRF. + +Sampling Hypotheses Based on the application a feasible sampling domain is chosen e.g. pixels/rays, or observations. Depending on the requirements of the neural field an appropriate sample size is determined. The initial sample size is chosen as small as possible to allow for a reasonable convergence expectation to cleaner sets, while being large enough for a coarse fit of the model expressive enough to enable the discrimination of outliers. + +Determine Model Parameters The determination of the model parameters in a classical RANSAC application corresponds to the inference of the sample sets with a neural field: In case of unconditioned neural fields (such as NeRF) this corresponds to (over-)fitting the model to the scene, in case of conditioned fields + +this corresponds to obtaining a latent representation. Full convergence is not required, the reconstruction is only used to evaluate the outlier contamination. + +Validation of Hypotheses For the validation, the rays/perspectives not used for the inference are rendered using the obtained model and compared to the input. The quality of each hypothesis is evaluated based on the amount of other samples explained by it up to some margin. Depending on the sampling/inference strategy, different similarity metrics can be used to distinguish outliers from other sources of noise such as the coarse inference. The initial sample set of the best hypothesis together with all its inliers is used for a final complete model fit. + +As will be demonstrated in our evaluation in Section 5, the major benefit of our RANSAC-like neural field approach over formulations based on robust loss functions [1] is the possibility to reliably filter outliers and inconsistencies in the input data in comparison to the down-weighting of their influence. + +# 3.3 Robust Light Field Networks + +In the following, we propose a novel fast and robust single-shot multi-class reconstruction algorithm based on LFNs [40]. LFNs are globally conditioned, meaning that the supported subset of 3D consistent scenes is represented by a single global latent vector. Therefore, when the latent space is not expressive enough to represent object and distractor correctly, large inconsistencies cause global damage instead of local artifacts. LFNs intrinsically support parallel inference, allowing to jointly process all hypotheses rather than following an iterative scheme. Our algorithm consists of the following steps (Fig. 2): + +1. Hypothesis Consensus Set: Given the input image $I$ as a set of pixel color values $\pmb{c}_i$ , and the intrinsic and extrinsic camera parameters, the set of rays $R$ - one ray $\pmb{r}_i$ for every pixel - represented by Plücker coordinates, is generated. In the first step, $N$ initial consensus sets $S_{n}$ are drawn using a uniform distribution, where each consensus set consists of $M$ random samples: + +$$ +\left(\boldsymbol {c} _ {n} ^ {m}, \boldsymbol {r} _ {n} ^ {m}\right) \in_ {R} \left\{\left(\boldsymbol {c} _ {i}, \boldsymbol {r} _ {i}\right) \mid \boldsymbol {c} _ {i} \in I, \boldsymbol {r} _ {i} \in R \right\} \tag {3} +$$ + +where $n \in \{1, \dots, N\}$ , $m \in \{1, \dots, M\}$ , and $\in_R$ denotes a sample randomly drawn from the set without replacement according to a uniform distribution. + +2. Hypothesis Inference: The autodecoder of an LFN $\varPhi$ with hypernetwork $\varPsi$ and pretrained hypernetwork weights $\psi$ , is used to infer the latent codes $z_{n}$ for each of the initial sample sets in parallel. + +$$ +\left\{\boldsymbol {z} _ {n} \right\} = \underset {\left\{\boldsymbol {z} _ {n} \right\}} {\operatorname {a r g m i n}} \sum_ {n} \sum_ {m} \| \Phi \left(\boldsymbol {r} _ {n} ^ {m} \mid \Psi_ {\psi} (\boldsymbol {z} _ {n})\right) - \boldsymbol {c} _ {n} ^ {m} \| _ {2} ^ {2} + \lambda_ {l a t} \| \boldsymbol {z} _ {n} \| _ {2} ^ {2} \tag {4} +$$ + +$\lambda_{lat}$ determines the strength of the Gaussian prior on the latent space [40]. An exponential learning rate schedule speeds up the inference. The inferred latent codes form the hypotheses. + +3. Hypothesis Prediction: Each of the hypotheses is used to render an entire image $I_{n}^{pred}$ from the perspective of the input image, each consisting of the pixel color values $\pmb{c}_{n,i}^{pred} = \Phi(\pmb{r}_i \mid \Psi_\psi(\pmb{z}_n))$ , using again the set of rays $R = \{\pmb{r}_i\}$ obtained from the camera parameters. The rendered pixels resemble the predictions for the remaining observations under each hypothesis. + +4. Hypothesis Validation: The obtained predictions are compared to the input image to validate the hypothesis. For each pixel in each predicted image, we calculate the Euclidean distance in color space: $e_{n,i} = \| \pmb{c}_{n,i}^{pred} - \pmb{c}_i\| _2$ . For each image, using these distances, we collect, up to some margin $\epsilon$ , the observations explained by the model (inliers): $S_{n}^{inlier} = \{(c_{i},r_{i})\mid e_{n,i} < \epsilon \}$ +5. Model Selection: We select the best hypothesis sample set $S_{best}$ based on the number of inliers $\# S_n^{inlier}$ . The model is inferred once more, similar to the second step, to obtain the final latent code $z_{cons}$ . The inference is based on the final consensus set $S_{cons} = S_{best} \cup S_{best}^{inlier}$ , the initial sample set of the strongest hypothesis $S_{best}$ together with all its inliers $S_{best}^{inlier}$ . The final output consists of both the latent code $z_{cons}$ and the final consensus set $S_{cons}$ of the selected model, and can be used to render arbitrary new perspectives. + +# 3.4 Robust Neural Radiance Fields + +In the following, we propose a novel robust multi-view reconstruction approach based on NeRFs. NeRFs are fit to a single scene based on a set of observations. As they support view-dependent radiance, one might expect inconsistencies in the input observations to only have an effect on specific perspectives. However, the density is only spatially parametrized, hence, inconsistencies lead to significant ghosting and smearing artifacts in more than just the inconsistent perspective. This can be leveraged in the hypothesis validation. The following steps are performed for N iterations: + +1. Hypothesis Consensus Set: A NeRF fit requires a large set of rays, making a sampling in ray space infeasible, as even obtaining significantly cleaner sets becomes unlikely. In order to obtain a sampling domain with reasonably sized initial consensus sets $S_{n}$ , we sample $M$ observations from the given sets of images $\mathcal{I}$ and corresponding camera poses $\mathcal{C}$ in every iteration: + +$$ +\left(I _ {n} ^ {m}, E _ {n} ^ {m}\right) \in_ {R} \left\{\left(I _ {i}, E _ {i}\right) \mid I _ {i} \in \mathcal {I}, E _ {i} \in \mathcal {C} \right\} \tag {5} +$$ + +where $n \in \{1, \dots, N\}$ , $m \in \{1, \dots, M\}$ , and $\in_R$ denotes a sample randomly drawn from the set without replacement according to a uniform distribution. + +2. Hypothesis Inference: The sampled observations are used to fit a hypothesis neural radiance field $F_{n}$ . +3. Hypothesis Prediction: The obtained NeRF $F_{n}$ is used to render predictions $I_{n,i}^{pred}$ with pixel colors $\pmb{p}_{n,i,x}^{pred}$ under the hypothesis for all unseen input perspectives $E_{i}$ . +4. Hypothesis Validation: The chosen sample space requires a careful evaluation of the hypotheses. We propose a two-step evaluation, where, first, the pixels inliers $P_{n,i}$ up to some margin $\epsilon_{pix}$ are determined for every observation based on the Euclidean distance in color space to the pixels $\pmb{p}_{i,x}$ of the input images $I_i$ : $P_{n,i} = \{\pmb{p}_{n,i,x}^{pred} \mid \| \pmb{p}_{n,i,x}^{pred} - \pmb{p}_{i,x} \|_2 < \epsilon_{pix}\}$ and, second, the observations themselves are labeled as inliers or outliers based on the amount of pixel inliers, again up to some margin $\epsilon_{img}$ , to obtain the consensus set $S_n^{inlier}$ under the hypothesis: $S_n^{inlier} = \{(I_i, E_i) \mid \# P_{n,i} > \epsilon_{img}\}$ . + +The binary metric for pixels ensures that smaller mispredictions (due to, e.g., view-dependent lighting effects) do not introduce noise into the evaluation. + +5. Model Selection: The strongest hypothesis, with its initial sample set $S_{best}$ , is selected based on the number of inliers $\# S_n^{inlier}$ to obtain the final consensus set $S_{cons} = S_{best} \cup S_{best}^{inlier}$ . The final model is obtained with one more NeRF fit of the consensus set. + +# 3.5 Hyperparameters + +For LFNs, the amount of initial samples and random hypotheses to evaluate (iterations), are determined experimentally. Without fine-tuning per class, the experimentally determined parameters are 90 initial samples and 2000 iterations, which supersedes the theoretical value for a convergence because the latent space introduces an intrinsic robustness. The inlier margin balances the amount of slight high-frequency variations that are being captured and the capability of separating outliers that are similar to the object. A margin of 0.25 in terms of the Euclidean distance of the predicted colors to the input samples in an RGB color space normalized to the range $(-1, 1)$ has been found to be optimal. Please refer to the supplementary material for further experimental results. + +For NeRFs, the parameters behave more natural and the amount of perspectives required for a meaningful, not completely artifact-free fit of the model lies around 25 observations [29, Table 2]. With fewer samples, more artifacts are introduced that get harder to separate from the ones caused by inconsistencies, and the samples get more dependent on being evenly spaced. For real-world captures with $10\%$ inconsistent perspectives, as few as 50 iterations are sufficient. With the color space normalized to $(0,1)$ , a pixel margin between around 0.15 in terms of Euclidean distance worked well for the determination of actual artifacts. We consider an observation an inlier based on a margin of $90\% - 98\%$ pixel inliers, which proved to be a good choice to separate minor artifacts (due to the sparse sampling) from artifacts caused by actual inconsistencies. For different datasets or inconsistencies, these values can be adapted. + +# 4 Implementation & Preprocessing + +For LFNs, we build on top of the original implementation [40], with a slight adaptation to enable a parallel sub-sampled inference. We furthermore use the provided pretrained multi-class model. The camera parameters are known. For efficiency reasons, the steps of the algorithm are not performed iteratively, but multiple hypotheses are validated in parallel. To further speed up the inference, an exponential learning rate schedule is used for the auto-decoding, leading to a total runtime of about a minute on a single GPU. + +For the robust reconstruction of objects from lazily captured real-world data, one has to estimate the camera parameters and extract foreground masks before applying the algorithm. For the estimation of the camera parameters, we used the COLMAP structure-from-motion package [39]. We extracted foreground masks + +using Segment Anything [22]. However, only foreground masks, containing the objects and the occlusions, are extracted. Segment Anything is not capable of removing arbitrary occlusions in an automatized way. After these preprocessing steps, the robust reconstruction algorithm can be applied as described. Erroneous estimates of the camera parameters or foreground masks are excluded by our algorithm, thus making the entire reconstruction pipeline robust. + +Our algorithm is not limited to a specific NeRF implementation. The chosen sampling domain eases integration into arbitrary existing NeRF implementations, which commonly expect images instead of unstructured ray sets. However, using a fast NeRF variant is advantageous when applying an iterative scheme. We used the instant NGP implementation [31] of the instant NSR repository [18], which includes some accelerations [24, 30]. Other, (specifically fast) variants are likely good choices as well. Antialiased and unbounded, but slow variants, such as MipNeRF360 [3], are not feasible. For further implementation details please refer to the supplementary. + +# 5 Experiments + +# 5.1 Inconsistencies,Baselines & Datasets + +Multi-View Reconstruction (NeRF) We provide a comprehensive quantitative and qualitative analysis of RANRAC's multi-view reconstructions compared to NeRF without any method of robustness (baseline) and RobustNeRF [38] (state of the art). We conduct experiments for various common sources of inconsistencies such as occluded perspectives, noisy camera parameter estimates, and blurred perspectives. RobustNeRF [38] targets unbounded scenes with multiple objects and small amounts of distractors in every perspective. In comparison, our RANSAC-based approach deals well with single-object reconstruction, even with heavy occlusions, as long as enough clean perspectives are available. As their dataset reflects the algorithm's properties, we cannot provide a fair comparison. Instead, we demonstrate the applicability using a custom dataset of a single object with a controlled amount of deliberately occluded perspectives. For the other inconsistencies we use of-the-shelf datasets and add noise to the camera parameters and blur to the images. Furthermore, we implement the robust losses [38] on the same NeRF variant as RANRAC to provide a fair comparison of the method's robustness, independent of the NeRF variant. For further details on the implementation, please refer to the supplementary material. + +Single-Shot Reconstruction (LFN) We benchmark against the original LFN implementation of Sitzmann et al. [40] as baseline, as there are no other robust methods for LFNs or conditioned neural fields, nor are there methods for robust single-shot multi-class reconstruction in general. Furthermore, we use the same pretrained LFN for the baseline and for the application of our method. The LFN is pretrained on the thirteen largest ShapeNet classes [8]. + +We provide a detailed qualitative and quantitative performance comparison under different amounts of occlusion for three representative classes (plane, car, + +and chair), while just stating reconstruction performance in a fixed environment without additional tuning of the hyperparameters for the others. The plane class is mostly challenging due to the low-frequency shape, while the car class contains a lot of high-frequency color details. The chair class represents shapes that are generally problematic for vanilla LFNs, even without occlusions. We provide a complementing analysis of the hyperparameters in the supplementary material. If not stated otherwise, we evaluate using 50 randomly selected images of the corresponding class. All comparisons use the same images. + +The occlusions are created synthetically. They consist of randomly generated patches while controlling two metrics of occlusion: Image occlusion and object occlusion. The former is the naive ratio of occluded over total pixels. The latter are the occluded pixels on the object compared to the total pixels covered by the object. We use both metrics to take the vastly different information entropy of samples across the image into account. For further details on the generation of the synthetic occlusions, please refer to the supplemental. + +# 5.2 Evaluation + +
MetricModelBenchBoatCabin.Displ.LampPhoneRifleSofaSpeak.Table
PSNR↑RANRAC19.2122.9221.9217.6519.9918.4521.3521.6120.720.44
LFN17.8919.220.518.8519.0917.8118.4620.1219.9920.29
SSIM↑RANRAC0.7670.8580.8010.6990.7640.750.8530.8050.7610.784
LFN0.7240.7910.7670.730.7480.7260.7950.7750.7430.777
+ +Table 1: RANRAC obtains a significant quantitative improvement in PSNR and SSIM (higher is better) compared to the baseline. We compare RANRAC to vanilla LFNs for the 13 largest ShapeNet classes (find plane, car, and chair with more detail in Fig. 3). The results are based on a moderate amount of occlusion of approximately $25\%$ object occlusion and about $5\%$ image occlusion. The reported results are conservative, as higher amounts of image occlusion result in a more significant performance increase (Fig. 3). No hyperparameter tuning has been performed for these classes; the same configuration obtained from the analysis of the other three classes is used. + +For LFNs, our approach leads to a significant improvement in occluded scenarios of up to 8dB in PSNR and a similarly strong improvement for the SSIM. The improvement is most significant in heavily distracted scenarios (Fig. 3). In clean scenarios a slight performance penalty can be observed, but even with small amounts of object occlusion (information loss), our algorithm outperforms the baseline, leading to numerically better results up to $50\%$ information loss (Fig. 3). The effect is not only measurable, but also well visible (Fig. 4). Increasing amounts of occlusion slowly introduce local artifacts into our reconstruction while preserving a reasonable shape estimate even for larger amounts of occlusion. In contrast, the reconstruction of LFNs breaks rather early in a global fashion. Still, our consensus set (Fig. 4), reveals that some high-frequency details were wrongfully excluded, explaining the slight performance decrease on clean images. In general, the benefit of RANRAC is best observable for classes that can be well described by LFNs, as evidenced by the lower improvements the chair class exhibits, compared to the plane and car class. The same effect is visible in the quantitative evaluation on the other classes without additional + +![](images/67c07b3b1a3d8dbe282989f20132f533f8508aec7cd161c7b0ad171a7fd958d0.jpg) +Fig. 3: RANRAC (solid lines) leads to a quantitative improvement in PSNR and SSIM (higher is better) for occluded inputs compared to vanilla LFNs (dashed lines). The same hyperparameter configuration and LFN is used for all classes. On the left and in the middle, the amount of image occlusion is increased, while the object occlusion is constant at $25\%$ . On the right, the amount of object occlusion is increased while the image occlusion is kept low. For the car class, a large improvement is observed over the entire occlusion spectrum. For the plane class the improvement is similarly significant, but absolute performance degenerates a bit sooner. This stems from the smaller object size and the related faster occlusion-to-object increase when increasing image occlusions. For the chair class, the improvement is less significant but the structural similarity is preserved for much longer. For the plane and car class the reconstruction quality is resilient to information loss (right) up to $\sim 50\%$ , where the decrease gains momentum. With the low amounts of image occlusion, the improvement is not significant for the chair class (consistent with left and middle). + +![](images/79686c1d75bc7e7e181fe8e489ab0db2b4cf96e9f2fcfd30764ed3fc85bc3deb.jpg) +Fig. 4: On the left, we show the qualitative effect of increasing occlusion on the same observation for the reconstruction of a novel view. Reconstructions of LFNs break early globally whereas RANRAC still provides a very decent reconstruction, only slowly introducing minor local (and natural/comprehensible) artifacts for completely hidden object parts. We further show the obtained consensus set, used for the final reconstruction (green inliers, red outliers). On the right, we show more qualitative results for novel view synthesis on different classes and the corresponding consensus sets. + +![](images/25a1557f50a85e6137cf682d89827c485cd7f03fa3be508e6ad29c12f8e5c6e5.jpg) +Fig. 5: The occlusions lead to well-visible artifacts in the reconstructions of NeRF, these artifacts are completely removed by RANRAC. While RobustNeRF struggles with view-dependent and high-frequency details, RANRAC reliably reconstructs them. + +
Inconsistency DatasetMild Occ.Strong Occ.Watering PotBlurred Perspectives
LegoShipChairFicusMic
RANRAC27.1125.9926.1224.9434.7929.9129.7616.1435.2531.26
RobustNeRF26.8325.5825.9324.7929.1423.6923.3120.1931.1127.12
NeRF26.6522.6125.3618.4731.1519.0028.4820.3533.2122.10
+ +
Inconsistency +DatasetNoisy Camera Parameters
LegoDrumsMicShipFicusHotdogMaterials
RANRAC34.95|31.2225.88|22.7735.85|33.5530.77|22.0531.41|28.8237.16|30.5328.93|25.39
RobustNeRF29.83|26.8223.91|22.1029.36|27.5524.21|19.9823.17|21.8532.65|25.1624.90|22.30
NeRF24.67|15.9523.22|17.9628.82|22.1523.68|14.8028.12|22.7728.16|18.3724.73|19.89
+ +Table 2: RANRAC outperforms the state of the art RobustNeRF [38] and the baseline NeRF (without a method of robustness) for scenes contaminated with occlusions, blur and noisy camera parameters. We report PSNR $\uparrow$ averaged across perspectives and the 5th percentile $(\mathrm{Avg.}|P_5)$ as artifacts introduced by inconsistencies only contaminate some views. We compare on the captured watering pot dataset with milder $(10\%)$ and heavier $(17.5\%)$ amounts of occluded perspectives and off-the-shelf datasets [29] with blurred perspectives and additive Gaussian noise $\mathcal{N}(5^{\circ},1^{\circ})$ on the camera parameters of $10\%$ of the perspectives. All three variants are built on top of instant-nsr for an isolated fair comparison of the robustness method. Note that both robust approaches struggle separating the strong view-dependent effects of the ship scene from blur, leading to the exclusion of some perspectives (lower $P_5$ PSNR), while RANRAC still improves the overall reconstruction. For all other scenes and inconsistencies, RANRAC reliably separates inconsistent observations from clean ones. + +tuning (Tab. 1). The only outlier is the display class, on which LFNs struggle the most on [40], even in unoccluded scenarios. It is not consistently represented in the latent space and robust reconstruction amplifies this effect by reducing the samples to a consistent set. + +The application to NeRF shows the significantly improved reconstructions of RANRAC (Tab. 2) for different types of inconsistencies such as occlusions and blur, as well as noisy camera parameters on captured and synthetic datasets. RANRAC consistently outperforms RobustNeRF [38] for object reconstruction from inconsistent inputs. RobustNeRF specifically struggles in capturing view-dependent appearance and at concavities while RANRAC seamlessly reconstructs them. The effects are not only measurable but also well-visible (Fig. 5). + +# 6 Limitations & Future Work + +For the application to NeRF, the iterations imply the use of a fast inferable variant, ruling out MipNeRF360 [3] and other high-quality variants for unbounded scenes. Foreground separation is a must, limiting this application of RANRAC to object-centric scenes. Novel fast, unbounded methods might resolve this [20]. + +By following a RANSAC-like approach, we inherit the requirement of sufficient clean perspectives, which could be lifted via NeRF variants that require fewer perspectives [32, 45] or by using different sampling domains. In return, our method is not limited to specific kinds of inconsistencies, and is robust to arbitrarily heavy distractions or inconsistencies in the impure perspectives. + +Whereas our robust LFN approach led to significant improvements for single-shot reconstruction, our concept might also be applied to other conditioned neural fields, based on a smart choice of the sampling domain, targeting photorealism. One could also use importance sampling based on a prior, instead of a uniform sampling, leveraging the unevenly distributed information entropy. Neural sampling priors [6], semantic segmentation and more advanced schemes (e.g. LO-RANSAC [11], DSAC [5]) might prove useful here. + +# 7 Conclusion + +We introduced a novel approach to increase the robustness of neural fields, inspired by the RANSAC paradigm. Following this concept, we introduced a novel robust approach for single-shot reconstruction from occluded views based on LFNs which achieves a significant improvement in reconstruction quality for distracted and occluded scenarios, even for extreme cases. Furthermore, we introduced a respective RANRAC-based NeRF variant that allows robust photorealistic reconstruction from multiple views with typical inconsistencies such as occlusions, noisy camera parameters, or blurred images - resulting in significant improvements compared to state-of-the-art methods - without relying on assumptions about the distractions. + +Acknowledgements - Benno Buschmann was at FAU funded through a gift by Mitsubishi Electric Research Laboratories (MERL). Andreea Dogaru was funded by the German Federal Ministry of Education and Research (BMBF), FKZ: 01IS22082 (IRRW). The authors are responsible for the content of this publication. + +# References + +1. Barron, J.T.: A general and adaptive robust loss function. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4331-4339 (2019) 2, 7 +2. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: ICCV (2021) 4 +3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In: CVPR (2022) 4, 10, 14 +4. Bian, W., Wang, Z., Li, K., Bian, J.W., Prisacariu, V.A.: Nope-nerf: Optimising neural radiance field with no pose prior. In: CVPR (2023) 3 +5. Brachmann, E., Krull, A., Nowozin, S., Shotton, J., Michel, F., Gumhold, S., Rother, C.: Dsac-differentiable ransac for camera localization. In: CVPR (2017) 4, 14 +6. Brachmann, E., Rother, C.: Neural-guided ransac: Learning where to sample model hypotheses. In: ICCV (2019) 14 +7. Brown, M., Lowe, D.G.: Automatic panoramic image stitching using invariant features. International Journal of Computer Vision 74, 59-73 (2007) 4 +8. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., Xiao, J., Yi, L., Yu, F.: ShapeNet: An Information-Rich 3D Model Repository. Tech. Rep. arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago (2015) 10 +9. Chen, X., Zhang, Q., Li, X., Chen, Y., Feng, Y., Wang, X., Wang, J.: Hallucinated neural radiance fields in the wild. In: CVPR (2022) 4 +0. Choi, S., Kim, T., Yu, W.: Performance evaluation of ransac family. Journal of Computer Vision 24(3), 271-300 (1997) 5, 6 +1. Chum, O., Matas, J., Kittler, J.: Locally optimized ransac. In: Pattern Recognition: 25th DAGM Symposium, Magdeburg, Germany, September 10-12, 2003. Proceedings 25. pp. 236-243. Springer (2003) 5, 14 +2. Derpanis, K.G.: Overview of the ransac algorithm. Image Rochester NY 4(1), 2-3 (2010) 5, 6 +3. Egger, B., Schneider, A., Blumer, C., Forster, A., Schonborn, S., Vetter, T.: Occlusion-aware 3d morphable face models. In: BMVC. vol. 2, p. 4 (2016) 4 +4. Egger, B., Schonborn, S., Schneider, A., Kortylewski, A., Morel-Forster, A., Blumer, C., Vetter, T.: Occlusion-aware 3d morphable models and an illumination prior for face image analysis. International Journal of Computer Vision 126, 1269-1287 (2018) 4 +5. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM 24(6), 381-395 (1981) 2, 4, 5 +6. Goli, L., Reading, C., Sellan, S., Jacobson, A., Tagliasacchi, A.: Bayes' Rays: Uncertainty quantification in neural radiance fields. arXiv (2023) 4 + +17. Guangcong, Chen, Z., Loy, C.C., Liu, Z.: Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. ICCV (2023) 3 +18. Guo, Y.C.: Instant neural surface reconstruction (2022), https://github.com/bennyguo/instant-nsr-pl 10 +19. Jeong, Y., Ahn, S., Choy, C., Anandkumar, A., Cho, M., Park, J.: Self-calibrating neural radiance fields. In: ICCV (2021) 3 +20. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering (2023) 14 +21. Kim, M., Seo, S., Han, B.: Infonerf: Ray entropy minimization for few-shot neural volume rendering. In: CVPR (2022) 3 +22. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023) 10 +23. Lee, J., Kim, I., Heo, H., Kim, H.J.: Semantic-aware occlusion filtering neural radiance fields in the wild. arXiv preprint arXiv:2303.03966 (2023) 4 +24. Li, R., Gao, H., Tancik, M., Kanazawa, A.: Nerfacc: Efficient sampling accelerates nerfs. arXiv preprint arXiv:2305.04966 (2023) 10 +25. Lin, C.H., Ma, W.C., Torralba, A., Lucey, S.: Barf: Bundle-adjusting neural radiance fields. In: ICCV (2021) 3 +26. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust dynamic radiance fields. In: CVPR (2023) 4 +27. Liu, Y., Peng, S., Liu, L., Wang, Q., Wang, P., Christian, T., Zhou, X., Wang, W.: Neural rays for occlusion-aware image-based rendering. In: CVPR (2022) 4 +28. Martin-Brualla, R., Radwan, N., Sajjadi, M.S.M., Barron, J.T., Dosovitskiy, A., Duckworth, D.: NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In: CVPR (2021) 4 +29. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) 2, 3, 9, 13 +30. Müller, T.: tiny-cuda-nn (4 2021), https://github.com/NVlabs/tiny-cuda-nn 10 +31. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. In: SIGGRAPH (2022) 3, 10 +32. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S.M., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: CVPR (2022) 3, 14 +33. Philip, J., Deschaintre, V.: Floaters no more: Radiance field gradient scaling for improved near-camera training. In: Eurographics Symposium on Rendering. The Eurographics Association (2023) 4 +34. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: CVPR (2021) 4 +35. Raguram, R., Frahm, J.M., Pollefeys, M.: A comparative analysis of ransac techniques leading to adaptive real-time random sample consensus. In: ECCV (2008) 5 +36. Rebain, D., Matthews, M., Yi, K.M., Lagun, D., Tagliasacchi, A.: Lolnerf: Learn from one look. In: CVPR (2022) 4 +37. Rematas, K., Liu, A., Srinivasan, P.P., Barron, J.T., Tagliasacchi, A., Funkhouser, T., Ferrari, V.: Urban radiance fields. CVPR (2022) 4 +38. Sabour, S., Vora, S., Duckworth, D., Krasin, I., Fleet, D.J., Tagliasacchi, A.: Robustnerf: Ignoring distractors with robust losses. In: CVPR (2023) 2, 3, 4, 10, 13, 14 + +39. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: CVPR (2016) 9 +40. Sitzmann, V., Rezchikov, S., Freeman, W.T., Tenenbaum, J.B., Durand, F.: Light field networks: Neural scene representations with single-evaluation rendering. In: NeurIPS (2021) 2, 3, 4, 7, 9, 10, 14 +41. Sitzmann, V., Zollhöfer, M., Wetzstein, G.: Scene representation networks: Continuous 3d-structure-aware neural scene representations. In: NeurIPS (2019) 4 +42. Sun, J., Chen, X., Wang, Q., Li, Z., Averbuch-Elor, H., Zhou, X., Snavely, N.: Neural 3D reconstruction in the wild. In: SIGGRAPH (2022) 4 +43. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: CVPR (2022) 4 +44. Tewari, A., Thies, J., Mildenhall, B., et al.: Advances in neural rendering. In: Comput. Graph. Forum. vol. 41, pp. 703-735. Wiley Online Library, Wiley (5 2022) 1, 3 +45. Truong, P., Rakotosaona, M.J., Manhardt, F., Tombari, F.: Sparf: Neural radiance fields from sparse and noisy poses. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4190-4200 (2023) 14 +46. Turki, H., Ramanan, D., Satyanarayanan, M.: Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In: CVPR (2022) 4 +47. Warburg*, F., Weber*, E., Tancik, M., Holski, A., Kanazawa, A.: Nerfbusters: Removing ghostly artifacts from casually captured nerfs. In: ICCV (2023) 4 +48. Wirth, T., Rak, A., Knauthe, V., Fellner, D.W.: A post processing technique to automatically remove floater artifacts in neural radiance fields. In: Computer Graphics Forum (2023) 4 +49. Wu, T., Zhong, F., Tagliasacchi, A., Cole, F., Oztireli, C.: D\~ 2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. NeurIPS (2022) 4 +50. Xiangli, Y., Xu, L., Pan, X., Zhao, N., Rao, A., Theobalt, C., Dai, B., Lin, D.: Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In: ECCV (2022) 4 +51. Xie, Y., Takikawa, T., Saito, S., Litany, O., Yan, S., Khan, N., Tombari, F., Tompkin, J., Sitzmann, V., Sridhar, S.: Neural fields in visual computing and beyond. Computer Graphics Forum 41 (2021) 1, 3 +52. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: CVPR (2021) 2, 3, 4 +53. Zhang, J.Y., Ramanan, D., Tulsiani, S.: Relpose: Predicting probabilistic relative rotation for single objects in the wild. In: ECCV (2022) 3 +54. Zhu, C., Wan, R., Tang, Y., Shi, B.: Occlusion-free scene recovery via neural radiance fields. In: CVPR (2023) 4 \ No newline at end of file diff --git a/2024/RANRAC_ Robust Neural Scene Representations via Random Ray Consensus/images.zip b/2024/RANRAC_ Robust Neural Scene Representations via Random Ray Consensus/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2c5220948bf4ede6bbd9c276fd5210ff52beabf3 --- /dev/null +++ b/2024/RANRAC_ Robust Neural Scene Representations via Random Ray Consensus/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a46f19afededc1bc68ea0dad257f077100ccb034eda8c314db1d39c50c535643 +size 338800 diff --git a/2024/RANRAC_ Robust Neural Scene Representations via Random Ray Consensus/layout.json b/2024/RANRAC_ Robust Neural Scene Representations via Random Ray Consensus/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5726cb8faaed95dfb2f801accf6bb626322a1af0 --- /dev/null +++ b/2024/RANRAC_ Robust Neural Scene Representations via Random Ray Consensus/layout.json @@ -0,0 +1,8293 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 135, + 111, + 479, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 111, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 135, + 111, + 479, + 148 + ], + "type": "text", + "content": "RANRAC: Robust Neural Scene Representations via Random Ray Consensus" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 157, + 167, + 457, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 167, + 457, + 193 + ], + "spans": [ + { + "bbox": [ + 157, + 167, + 457, + 193 + ], + "type": "text", + "content": "Benno Buschmann1,2, Andreea Dogaru2, Elmar Eisemann1, Michael Weinmann1, and Bernhard Egger2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 212, + 201, + 400, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 201, + 400, + 213 + ], + "spans": [ + { + "bbox": [ + 212, + 201, + 400, + 213 + ], + "type": "text", + "content": "1 Delft University of Technology, Netherlands" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 177, + 213, + 435, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 213, + 435, + 224 + ], + "spans": [ + { + "bbox": [ + 177, + 213, + 435, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 177, + 213, + 435, + 224 + ], + "type": "text", + "content": " Friedrich-Alexander-University Erlangen-Nürnberg, Germany" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 253, + 452, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 253, + 452, + 527 + ], + "spans": [ + { + "bbox": [ + 160, + 253, + 452, + 527 + ], + "type": "text", + "content": "Abstract. Learning-based scene representations such as neural radiance fields or light field networks, that rely on fitting a scene model to image observations, commonly encounter challenges in the presence of inconsistencies within the images caused by occlusions, inaccurately estimated camera parameters or effects like lens flare. To address this challenge, we introduce RANdom RAy Consensus (RANRAC), an efficient approach to eliminate the effect of inconsistent data, thereby taking inspiration from classical RANSAC based outlier detection for model fitting. In contrast to the down-weighting of the effect of outliers based on robust loss formulations, our approach reliably detects and excludes inconsistent perspectives, resulting in clean images without floating artifacts. For this purpose, we formulate a fuzzy adaption of the RANSAC paradigm, enabling its application to large scale models. We interpret the minimal number of samples to determine the model parameters as a tunable hyperparameter, investigate the generation of hypotheses with data-driven models, and analyse the validation of hypotheses in noisy environments. We demonstrate the compatibility and potential of our solution for both photo-realistic robust multi-view reconstruction from real-world images based on neural radiance fields and for single-shot reconstruction based on light-field networks. In particular, the results indicate significant improvements compared to state-of-the-art robust methods for novel-view synthesis on both synthetic and captured scenes with various inconsistencies including occlusions, noisy camera pose estimates, and unfocused perspectives. The results further indicate significant improvements for single-shot reconstruction from occluded images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 538, + 452, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 538, + 452, + 559 + ], + "spans": [ + { + "bbox": [ + 160, + 538, + 452, + 559 + ], + "type": "text", + "content": "Keywords: Neural scene representations " + }, + { + "bbox": [ + 160, + 538, + 452, + 559 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 538, + 452, + 559 + ], + "type": "text", + "content": " neural rendering " + }, + { + "bbox": [ + 160, + 538, + 452, + 559 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 538, + 452, + 559 + ], + "type": "text", + "content": " RANSAC " + }, + { + "bbox": [ + 160, + 538, + 452, + 559 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 538, + 452, + 559 + ], + "type": "text", + "content": " robust estimation " + }, + { + "bbox": [ + 160, + 538, + 452, + 559 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 538, + 452, + 559 + ], + "type": "text", + "content": " neural radiance fields " + }, + { + "bbox": [ + 160, + 538, + 452, + 559 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 538, + 452, + 559 + ], + "type": "text", + "content": " light-field networks" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 580, + 229, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 580, + 229, + 594 + ], + "spans": [ + { + "bbox": [ + 133, + 580, + 229, + 594 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "3D reconstruction is a classical task in computer vision and computer graphics, which has attracted research for decades. It offers numerous applications, including autonomous systems, entertainment, design, advertisement, cultural heritage, VR/AR experiences or medical scenarios. In recent years, neural scene representations and rendering techniques [44,51], including light field networks" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 186, + 114, + 429, + 185 + ], + "blocks": [ + { + "bbox": [ + 186, + 114, + 429, + 185 + ], + "lines": [ + { + "bbox": [ + 186, + 114, + 429, + 185 + ], + "spans": [ + { + "bbox": [ + 186, + 114, + 429, + 185 + ], + "type": "image", + "image_path": "2aee6ce2116d1ebbcb2915d1807117e82963fb404ad9dcd39d13887a5e328f6b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 194, + 482, + 251 + ], + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 251 + ], + "type": "text", + "content": "Fig. 1: We propose a robust algorithm for 3D reconstruction from occluded input perspectives that is based on the random sampling of hypotheses. Our algorithm is general and we demonstrate the use for single-shot reconstruction using light field networks or multi-view reconstruction using NeRF. In these cases, it successfully removes the artifacts that normally occur due to occluded input perspectives." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 279, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 482, + 399 + ], + "type": "text", + "content": "(LFN) [40] and neural radiance fields (NeRF) [29] have demonstrated great performance in single-view and multi-view reconstruction tasks. The key to the success of such techniques is the coupling of differentiable rendering methods with custom neural field parametrizations of scene properties. However, a common limitation of neural scene reconstruction methods is their sensitivity to inconsistencies among input images induced by occlusions, inaccurately estimated camera parameters or effects like lens flares. Despite the use of view-dependent radiance representations to address view-dependent appearance changes, these inconsistencies severely impact local density estimation, resulting in a poor generalization to novel views." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 402, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 666 + ], + "type": "text", + "content": "To increase the robustness to potential distractors within the training data, Sabour et al. [38] recently introduced the use of robust losses in the context of training unconditioned NeRF, where distractors in the training data were modeled as outliers of an optimization problem. However, the adaptation of this approach to conditioned neural fields (e.g., pixelNeRF [52]) is not obvious, as no optimization takes place during inference, and the input data is constrained to only a few views. Achieving robustness to data inconsistencies is a well-analyzed problem in computer vision, covered not only by the aforementioned robust loss functions [1], but also other strategies, like the random sample consensus (RANSAC) paradigm [15]. The latter is widely employed for fitting models to outlier-heavy data. The underlying idea is to randomly select subsets of the data to form potential models, evaluating these models against the entire dataset, and identifying the subset that best fits the majority of the data, while disregarding outliers. Despite being the state-of-the-art solution to many challenges, RANSAC-based schemes are particularly favoured for the fitting of analytical models with a relatively small amount of tuneable parameters. In this paper, we direct our attention to achieving robustness against inconsistencies and occlusions in the observations by using a novel combination of neural scene representation and rendering techniques with dedicated outlier removal techniques such as RANSAC [15]. While downweighting the influence of distractors based on robust losses [1, 38] can affect clean samples, representing details, we aim at improving robustness to distractors by only removing the influence of" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "type": "text", + "content": "B. Buschmann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "outliers. To this extent, we integrate a RANSAC-based scheme to distinguish inliers and outliers in the data and the inlier-based optimization of the neural fields (Fig. 1); a stochastic scenario characterized by a large-scale, data-driven model that exceeds RANSAC's classical convergence expectations. Instead of guaranteeing convergence to a clean sample set based on a minimal number of samples, we aim for a feasible (cleaner) sample set using a tuneable amount of samples. The proposed algorithm exhibits robustness and versatility, accommodating a wide range of neural fields-based reconstruction methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 213, + 482, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 368 + ], + "type": "text", + "content": "Our method inherits the strengths of RANSAC, such as the ability to handle various classes of outliers without relying on semantics. Yet, it also inherits the need for sufficiently clean samples and the reliance on an iterative scheme. In practice, the first condition is often fulfilled because typically only some of the perspectives are affected by inconsistencies. We validate our approach using synthetic data, focusing on the task of multi-class single-shot reconstruction with LFNs [40], and observe significant quality improvements over the baseline in the presence of occlusions. Furthermore, we showcase robust photo-realistic reconstructions of 3D objects using unconditioned NeRFs from sequences of real-world images in the presence of distractors. In comparison to RobustNeRF [38], we use all available clean data, hence improving the reconstruction quality for single-object scenes. Code and data are available under https://bennobuschmann.com/ranrac/. Our key contributions are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 377, + 480, + 485 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 377, + 479, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 377, + 479, + 399 + ], + "spans": [ + { + "bbox": [ + 138, + 377, + 479, + 399 + ], + "type": "text", + "content": "- a general, robust RANSAC-based reconstruction method applicable to a variety of neural-fields and handling diverse inconsistencies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 402, + 480, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 402, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 402, + 480, + 424 + ], + "type": "text", + "content": "- an analysis of the implication to RANSAC's hyperparameters and theoretical convergence expectations, and the experimental study of their effect" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 426, + 479, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 426, + 479, + 448 + ], + "spans": [ + { + "bbox": [ + 138, + 426, + 479, + 448 + ], + "type": "text", + "content": "- a method for robust photo-realistic object reconstruction using NeRF and for robust single-shot multi-class reconstruction using LFNs" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 450, + 479, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 450, + 479, + 485 + ], + "spans": [ + { + "bbox": [ + 138, + 450, + 479, + 485 + ], + "type": "text", + "content": "- a qualitative/quantitative evaluation of our method on both synthetic and real-world data with different inconsistencies (occlusions, invalid calibrations, ...) indicating the state-of-the-art performance of our approach" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 506, + 237, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 506, + 237, + 519 + ], + "spans": [ + { + "bbox": [ + 132, + 506, + 237, + 519 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "text", + "content": "Among the vast literature on neural fields, the seminal work of Mildenhall et al. [29] opened many avenues in the computer vision community. It contributed to state-of-the-art solutions for novel view synthesis and 3D reconstruction that have been covered in respective surveys [44, 51]. Noteworthy is the more recent contribution of instant neural graphics primitives (iNGP) [31], which uses a hash table of trainable feature vectors alongside a small network for representing the scene. iNGP achieved major run-time improvements, thereby enhancing the feasibility of practical applications for neural fields." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Baseline models are highly sensitive to imperfections in the input data, which led to many works on robustness enhancements of neural fields; addressing a reduced amount of input views [17,21,32,52], errors in camera parameters [4,19," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 90, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 90, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 405, + 90, + 447, + 100 + ], + "type": "text", + "content": "RANRAC" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 480, + 140 + ], + "type": "text", + "content": "[25, 53], variations in illumination conditions across observations [28, 42], multiscale image data [2, 3, 50], and the targeted removal of floating artifacts [33, 47, 48]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 144, + 482, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 144, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 144, + 482, + 419 + ], + "type": "text", + "content": "Fewer works solve the reconstruction task in the presence of inconsistencies between observations. Bayes' Rays [16] provides a framework to quantify uncertainty of a pretrained NeRF by approximating a spatial uncertainty field. It handles missing information due to self-occlusion or missing perspectives well, but cannot deal with inconsistencies caused by noise or distractors. Similarly, NeuRay [27] only supports missing, but not inconsistent information. Naive occlusion handling via semantic segmentation requires the occluding object types to be known in advance [28,36,37,43,46]. Solutions to learn semantic priors on transience exist [23] but separating occlusions via semantic segmentation without manual guidance is ill posed. Occ-NeRF [54] considers any foreground element as occlusion and removes them via depth reasoning, but their removal leaves behind blurry artifacts. Alternatively, some methods do not remove dynamic distractors, but reconstruct them together with the rest of the scene using time-conditioned representations [9,26,34,49]. Closely related to our work, RobustNeRF [38] considers input-image distractors as outliers of the model optimization task. The authors employ robust losses improved via patching to preserve high-frequency details. RobustNeRF does not rely on prior assumptions about the nature of distractors, nor does it require preprocessing of the input data or postprocessing of the trained model. Nevertheless, their method comes at the cost of loosing view-dependent details and a reduced reconstruction quality in undistracted scenes. Furthermore, their method is limited to unconditioned models that overfit to a single scene. A generalization to conditioned NeRFs, such as pixelNeRF [52], is not obvious, as no further optimization takes place during inference." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 422, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 422, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 422, + 482, + 555 + ], + "type": "text", + "content": "Conditioned neural fields offer a distinct advantage in their ability to generalize to novel scenes by leveraging knowledge acquired from diverse scenes during learning. This results in a more robust model that requires as few as one input view for inference, showcasing the efficiency and adaptability of the approach. Contrary to PixelNeRF [52], which relies on a volumetric parametrization of the scene, demanding multiple network evaluations along the ray, Light Field Networks (LFNs) [40], which succeed Scene Representation Networks [41], take a different approach. LFNs represent the scene as a 4D light field, enabling a more efficient single evaluation per ray for inference. The network takes as input a ray represented in Plücker coordinates and maps it to an observed radiance, all within an autoencoder framework used for conditioning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 482, + 666 + ], + "type": "text", + "content": "None of the mentioned methods can deal with occlusions in single-shot reconstruction and no prior work exists on robust LFNs or robustness of other conditioned neural fields for single-shot reconstruction, which we address via the RANSAC paradigm [15]. Since its introduction in 1981, RANSAC has gained attention for fitting analytical models with a small number of parameters, such as homography estimation in panorama stitching [7]. Among the few direct applications of classical RANSAC to larger models is robust morphable face reconstruction [13, 14]. Other common expansions and applications include differentiable RANSAC [5] for camera parameter estimation in a deep learning pipeline, locally" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "text", + "content": "B. Buschmann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 116, + 479, + 205 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 479, + 205 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 479, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 479, + 205 + ], + "type": "image", + "image_path": "92f72ab285e6d711cd62058e8ff29bf00bd4f204d2952d6b976d9b2b5395fa69.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 215, + 482, + 326 + ], + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 326 + ], + "type": "text", + "content": "Fig. 2: The RANRAC algorithm for LFNs samples random hypotheses by selecting a set of random samples from the given perspective (a), and inferring the latent representation of these rays using the autodecoder of a pretrained LFN (b). The obtained light field is then used to predict an image from the input perspective (c). Based on this prediction, confidence in the random hypothesis is evaluated via the Euclidean distance between the predicted ray colors and the remaining color samples of the input image. The amount of samples which are explained by each hypothesis up to some margin are used to determine the best hypothesis (d). All samples explained by the selected hypothesis are used for a final inference with the LFN to obtain the final model and latent representation (e)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 348, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 482, + 373 + ], + "type": "text", + "content": "optimized RANSAC [11] to account for the requirement of a descriptive sample set, and adaptive real-time RANSAC [35]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 389, + 201, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 389, + 201, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 389, + 201, + 402 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 414, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 521 + ], + "type": "text", + "content": "In this section, we present our approach to increase the robustness of neural scene representations to inconsistencies in the input data. First, we recap RANSAC, its theoretical convergence and hyperparameters, and the required adaptions for its application to high-dimensional data-driven models. We then introduce a general scheme for the random sampling and validation of neural fields. Based hereon, we formulate a robust algorithm using LFNs for 3D reconstruction from a single image with occlusions. Finally, using NeRF, we formulate an algorithm for robust photo-realistic reconstruction from multiple views in the presence of common sources of inconsistencies." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 538, + 382, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 538, + 382, + 551 + ], + "spans": [ + { + "bbox": [ + 132, + 538, + 382, + 551 + ], + "type": "text", + "content": "3.1 RANSAC Convergence on Complex Models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "text", + "content": "Classical RANSAC [10,12,15] follows an iterative process. Initially, a minimal set of samples is randomly selected to determine the model parameters, known as the hypothesis generation phase. Then, the hypothesis is evaluated by assessing the number of observations it explains, within a specified margin. These steps are repeated until the best hypothesis is chosen to constitute the consensus set, which comprises all of its inliers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "This paradigm cannot be directly applied to complex models such as neural fields, as a significant amount of samples is required to obtain decent initial model parameters and additional clean samples improve the quality further." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "text", + "content": "RANRAC" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "This imposes a challenge regarding the expected amount of clean initial sample sets, " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "S_{clean}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 219, + 140, + 481, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 140, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 219, + 140, + 481, + 172 + ], + "type": "interline_equation", + "content": "\\mathbb {E} [ \\# S _ {\\text {c l e a n}} ] = N * \\prod_ {m = 1} ^ {M} \\frac {s _ {i m g} - s _ {i m g} ^ {\\text {o c c}} - m}{s _ {i m g} - m}, \\tag {1}", + "image_path": "628c9f4363ba0a7c0227b103b275f7629e58ad9b6e6fed6a1ba603947b037ce5.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 176, + 479, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 479, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 479, + 212 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 176, + 479, + 212 + ], + "type": "inline_equation", + "content": "s_{img}" + }, + { + "bbox": [ + 130, + 176, + 479, + 212 + ], + "type": "text", + "content": " denotes the total amount of samples (e.g., image pixels), " + }, + { + "bbox": [ + 130, + 176, + 479, + 212 + ], + "type": "inline_equation", + "content": "s_{img}^{occ}" + }, + { + "bbox": [ + 130, + 176, + 479, + 212 + ], + "type": "text", + "content": " represents the occluded samples, and " + }, + { + "bbox": [ + 130, + 176, + 479, + 212 + ], + "type": "inline_equation", + "content": "N / M" + }, + { + "bbox": [ + 130, + 176, + 479, + 212 + ], + "type": "text", + "content": " are the number of iterations/samples. The expected amount exponentially decreases with the initial number of samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 213, + 480, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 480, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 480, + 368 + ], + "type": "text", + "content": "The samples and respective requirements for analytic and data-driven models vary a lot. The effect of individual samples is less traceable in data-driven models and the information entropy varies more significantly across samples. When using a model that projects onto a latent space, some very atypical outliers do not show an effect at all if the latent space is not expressive enough to explain them in an overall loss-reducing way, yet, outliers close to the object or its color, or larger chunks of outliers, will usually be distracting. At the same time, samples of small-scale high-frequency details are important for the reconstruction and contain a lot of information, whereas multiple samples of larger-scale lower-frequency details contribute much less. The amount of initial samples for the hypothesis generation becomes a tunable hyperparameter trading initial reconstruction quality for likelihood of finding desired sample sets. This invalidates the classical convergence idea [10, 12] where the RANSAC iterations " + }, + { + "bbox": [ + 130, + 213, + 480, + 368 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 213, + 480, + 368 + ], + "type": "text", + "content": " with" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 268, + 373, + 481, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 373, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 268, + 373, + 481, + 397 + ], + "type": "interline_equation", + "content": "N \\geq \\frac {\\log (1 - p)}{\\log (1 - t ^ {M})} \\tag {2}", + "image_path": "762f4ff4eb933afe8d3e33271ed4b5aa4d1c6fa80469712d620149506d45a0d2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 401, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 485 + ], + "type": "text", + "content": "are chosen such that at least one clean sample set is found with a probability " + }, + { + "bbox": [ + 130, + 401, + 482, + 485 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 401, + 482, + 485 + ], + "type": "text", + "content": ", given the expected ratio of clean samples " + }, + { + "bbox": [ + 130, + 401, + 482, + 485 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 401, + 482, + 485 + ], + "type": "text", + "content": " and the amount of initial samples " + }, + { + "bbox": [ + 130, + 401, + 482, + 485 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 401, + 482, + 485 + ], + "type": "text", + "content": ". There is not only the need to find a clean sample set, but one that captures all important details. At the same time, a completely clean sample set is not required at all, as long as the contained outliers are not represented by the local minimum of the latent space or the model itself, depending on the concrete scene representation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 502, + 326, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 502, + 326, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 502, + 326, + 514 + ], + "type": "text", + "content": "3.2 Random Sampling Neural Fields" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 479, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 479, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 479, + 544 + ], + "type": "text", + "content": "We propose a general strategy for RANSAC-like, iterative, robust reconstruction with neural fields, before formulating respective algorithms for LFNs and NeRF." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 545, + 479, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 479, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 479, + 616 + ], + "type": "text", + "content": "Sampling Hypotheses Based on the application a feasible sampling domain is chosen e.g. pixels/rays, or observations. Depending on the requirements of the neural field an appropriate sample size is determined. The initial sample size is chosen as small as possible to allow for a reasonable convergence expectation to cleaner sets, while being large enough for a coarse fit of the model expressive enough to enable the discrimination of outliers." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "content": "Determine Model Parameters The determination of the model parameters in a classical RANSAC application corresponds to the inference of the sample sets with a neural field: In case of unconditioned neural fields (such as NeRF) this corresponds to (over-)fitting the model to the scene, in case of conditioned fields" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "text", + "content": "B. Buschmann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "this corresponds to obtaining a latent representation. Full convergence is not required, the reconstruction is only used to evaluate the outlier contamination." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 479, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 479, + 223 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 479, + 223 + ], + "type": "text", + "content": "Validation of Hypotheses For the validation, the rays/perspectives not used for the inference are rendered using the obtained model and compared to the input. The quality of each hypothesis is evaluated based on the amount of other samples explained by it up to some margin. Depending on the sampling/inference strategy, different similarity metrics can be used to distinguish outliers from other sources of noise such as the coarse inference. The initial sample set of the best hypothesis together with all its inliers is used for a final complete model fit." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 224, + 479, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 479, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 479, + 271 + ], + "type": "text", + "content": "As will be demonstrated in our evaluation in Section 5, the major benefit of our RANSAC-like neural field approach over formulations based on robust loss functions [1] is the possibility to reliably filter outliers and inconsistencies in the input data in comparison to the down-weighting of their influence." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 289, + 307, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 289, + 307, + 301 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 307, + 301 + ], + "type": "text", + "content": "3.3 Robust Light Field Networks" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 309, + 479, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 479, + 404 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 479, + 404 + ], + "type": "text", + "content": "In the following, we propose a novel fast and robust single-shot multi-class reconstruction algorithm based on LFNs [40]. LFNs are globally conditioned, meaning that the supported subset of 3D consistent scenes is represented by a single global latent vector. Therefore, when the latent space is not expressive enough to represent object and distractor correctly, large inconsistencies cause global damage instead of local artifacts. LFNs intrinsically support parallel inference, allowing to jointly process all hypotheses rather than following an iterative scheme. Our algorithm consists of the following steps (Fig. 2):" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 415, + 479, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 415, + 479, + 473 + ], + "spans": [ + { + "bbox": [ + 134, + 415, + 479, + 473 + ], + "type": "text", + "content": "1. Hypothesis Consensus Set: Given the input image " + }, + { + "bbox": [ + 134, + 415, + 479, + 473 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 134, + 415, + 479, + 473 + ], + "type": "text", + "content": " as a set of pixel color values " + }, + { + "bbox": [ + 134, + 415, + 479, + 473 + ], + "type": "inline_equation", + "content": "\\pmb{c}_i" + }, + { + "bbox": [ + 134, + 415, + 479, + 473 + ], + "type": "text", + "content": ", and the intrinsic and extrinsic camera parameters, the set of rays " + }, + { + "bbox": [ + 134, + 415, + 479, + 473 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 134, + 415, + 479, + 473 + ], + "type": "text", + "content": " - one ray " + }, + { + "bbox": [ + 134, + 415, + 479, + 473 + ], + "type": "inline_equation", + "content": "\\pmb{r}_i" + }, + { + "bbox": [ + 134, + 415, + 479, + 473 + ], + "type": "text", + "content": " for every pixel - represented by Plücker coordinates, is generated. In the first step, " + }, + { + "bbox": [ + 134, + 415, + 479, + 473 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 134, + 415, + 479, + 473 + ], + "type": "text", + "content": " initial consensus sets " + }, + { + "bbox": [ + 134, + 415, + 479, + 473 + ], + "type": "inline_equation", + "content": "S_{n}" + }, + { + "bbox": [ + 134, + 415, + 479, + 473 + ], + "type": "text", + "content": " are drawn using a uniform distribution, where each consensus set consists of " + }, + { + "bbox": [ + 134, + 415, + 479, + 473 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 134, + 415, + 479, + 473 + ], + "type": "text", + "content": " random samples:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 229, + 474, + 480, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 474, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 229, + 474, + 480, + 487 + ], + "type": "interline_equation", + "content": "\\left(\\boldsymbol {c} _ {n} ^ {m}, \\boldsymbol {r} _ {n} ^ {m}\\right) \\in_ {R} \\left\\{\\left(\\boldsymbol {c} _ {i}, \\boldsymbol {r} _ {i}\\right) \\mid \\boldsymbol {c} _ {i} \\in I, \\boldsymbol {r} _ {i} \\in R \\right\\} \\tag {3}", + "image_path": "3c5f995c265a0f7c7e2916239fd78c5c27391b8202eb678a67bf6b440eb9f8f2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 145, + 487, + 479, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 487, + 479, + 510 + ], + "spans": [ + { + "bbox": [ + 145, + 487, + 479, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 145, + 487, + 479, + 510 + ], + "type": "inline_equation", + "content": "n \\in \\{1, \\dots, N\\}" + }, + { + "bbox": [ + 145, + 487, + 479, + 510 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 145, + 487, + 479, + 510 + ], + "type": "inline_equation", + "content": "m \\in \\{1, \\dots, M\\}" + }, + { + "bbox": [ + 145, + 487, + 479, + 510 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 145, + 487, + 479, + 510 + ], + "type": "inline_equation", + "content": "\\in_R" + }, + { + "bbox": [ + 145, + 487, + 479, + 510 + ], + "type": "text", + "content": " denotes a sample randomly drawn from the set without replacement according to a uniform distribution." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 510, + 479, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 510, + 479, + 545 + ], + "spans": [ + { + "bbox": [ + 134, + 510, + 479, + 545 + ], + "type": "text", + "content": "2. Hypothesis Inference: The autodecoder of an LFN " + }, + { + "bbox": [ + 134, + 510, + 479, + 545 + ], + "type": "inline_equation", + "content": "\\varPhi" + }, + { + "bbox": [ + 134, + 510, + 479, + 545 + ], + "type": "text", + "content": " with hypernetwork " + }, + { + "bbox": [ + 134, + 510, + 479, + 545 + ], + "type": "inline_equation", + "content": "\\varPsi" + }, + { + "bbox": [ + 134, + 510, + 479, + 545 + ], + "type": "text", + "content": " and pretrained hypernetwork weights " + }, + { + "bbox": [ + 134, + 510, + 479, + 545 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 134, + 510, + 479, + 545 + ], + "type": "text", + "content": ", is used to infer the latent codes " + }, + { + "bbox": [ + 134, + 510, + 479, + 545 + ], + "type": "inline_equation", + "content": "z_{n}" + }, + { + "bbox": [ + 134, + 510, + 479, + 545 + ], + "type": "text", + "content": " for each of the initial sample sets in parallel." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 183, + 545, + 480, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 545, + 480, + 569 + ], + "spans": [ + { + "bbox": [ + 183, + 545, + 480, + 569 + ], + "type": "interline_equation", + "content": "\\left\\{\\boldsymbol {z} _ {n} \\right\\} = \\underset {\\left\\{\\boldsymbol {z} _ {n} \\right\\}} {\\operatorname {a r g m i n}} \\sum_ {n} \\sum_ {m} \\| \\Phi \\left(\\boldsymbol {r} _ {n} ^ {m} \\mid \\Psi_ {\\psi} (\\boldsymbol {z} _ {n})\\right) - \\boldsymbol {c} _ {n} ^ {m} \\| _ {2} ^ {2} + \\lambda_ {l a t} \\| \\boldsymbol {z} _ {n} \\| _ {2} ^ {2} \\tag {4}", + "image_path": "1cadafe35401868af16258d55b0dda6d6b2ff54211124400957bb27d41fa4c40.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 145, + 569, + 479, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 569, + 479, + 604 + ], + "spans": [ + { + "bbox": [ + 145, + 569, + 479, + 604 + ], + "type": "inline_equation", + "content": "\\lambda_{lat}" + }, + { + "bbox": [ + 145, + 569, + 479, + 604 + ], + "type": "text", + "content": " determines the strength of the Gaussian prior on the latent space [40]. An exponential learning rate schedule speeds up the inference. The inferred latent codes form the hypotheses." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 604, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 604, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 604, + 479, + 665 + ], + "type": "text", + "content": "3. Hypothesis Prediction: Each of the hypotheses is used to render an entire image " + }, + { + "bbox": [ + 134, + 604, + 479, + 665 + ], + "type": "inline_equation", + "content": "I_{n}^{pred}" + }, + { + "bbox": [ + 134, + 604, + 479, + 665 + ], + "type": "text", + "content": " from the perspective of the input image, each consisting of the pixel color values " + }, + { + "bbox": [ + 134, + 604, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\pmb{c}_{n,i}^{pred} = \\Phi(\\pmb{r}_i \\mid \\Psi_\\psi(\\pmb{z}_n))" + }, + { + "bbox": [ + 134, + 604, + 479, + 665 + ], + "type": "text", + "content": ", using again the set of rays " + }, + { + "bbox": [ + 134, + 604, + 479, + 665 + ], + "type": "inline_equation", + "content": "R = \\{\\pmb{r}_i\\}" + }, + { + "bbox": [ + 134, + 604, + 479, + 665 + ], + "type": "text", + "content": " obtained from the camera parameters. The rendered pixels resemble the predictions for the remaining observations under each hypothesis." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "text", + "content": "RANRAC" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 134, + 116, + 482, + 260 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 134, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 134, + 116, + 482, + 177 + ], + "type": "text", + "content": "4. Hypothesis Validation: The obtained predictions are compared to the input image to validate the hypothesis. For each pixel in each predicted image, we calculate the Euclidean distance in color space: " + }, + { + "bbox": [ + 134, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "e_{n,i} = \\| \\pmb{c}_{n,i}^{pred} - \\pmb{c}_i\\| _2" + }, + { + "bbox": [ + 134, + 116, + 482, + 177 + ], + "type": "text", + "content": ". For each image, using these distances, we collect, up to some margin " + }, + { + "bbox": [ + 134, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 134, + 116, + 482, + 177 + ], + "type": "text", + "content": ", the observations explained by the model (inliers): " + }, + { + "bbox": [ + 134, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "S_{n}^{inlier} = \\{(c_{i},r_{i})\\mid e_{n,i} < \\epsilon \\}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 177, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 177, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 134, + 177, + 482, + 260 + ], + "type": "text", + "content": "5. Model Selection: We select the best hypothesis sample set " + }, + { + "bbox": [ + 134, + 177, + 482, + 260 + ], + "type": "inline_equation", + "content": "S_{best}" + }, + { + "bbox": [ + 134, + 177, + 482, + 260 + ], + "type": "text", + "content": " based on the number of inliers " + }, + { + "bbox": [ + 134, + 177, + 482, + 260 + ], + "type": "inline_equation", + "content": "\\# S_n^{inlier}" + }, + { + "bbox": [ + 134, + 177, + 482, + 260 + ], + "type": "text", + "content": ". The model is inferred once more, similar to the second step, to obtain the final latent code " + }, + { + "bbox": [ + 134, + 177, + 482, + 260 + ], + "type": "inline_equation", + "content": "z_{cons}" + }, + { + "bbox": [ + 134, + 177, + 482, + 260 + ], + "type": "text", + "content": ". The inference is based on the final consensus set " + }, + { + "bbox": [ + 134, + 177, + 482, + 260 + ], + "type": "inline_equation", + "content": "S_{cons} = S_{best} \\cup S_{best}^{inlier}" + }, + { + "bbox": [ + 134, + 177, + 482, + 260 + ], + "type": "text", + "content": ", the initial sample set of the strongest hypothesis " + }, + { + "bbox": [ + 134, + 177, + 482, + 260 + ], + "type": "inline_equation", + "content": "S_{best}" + }, + { + "bbox": [ + 134, + 177, + 482, + 260 + ], + "type": "text", + "content": " together with all its inliers " + }, + { + "bbox": [ + 134, + 177, + 482, + 260 + ], + "type": "inline_equation", + "content": "S_{best}^{inlier}" + }, + { + "bbox": [ + 134, + 177, + 482, + 260 + ], + "type": "text", + "content": ". The final output consists of both the latent code " + }, + { + "bbox": [ + 134, + 177, + 482, + 260 + ], + "type": "inline_equation", + "content": "z_{cons}" + }, + { + "bbox": [ + 134, + 177, + 482, + 260 + ], + "type": "text", + "content": " and the final consensus set " + }, + { + "bbox": [ + 134, + 177, + 482, + 260 + ], + "type": "inline_equation", + "content": "S_{cons}" + }, + { + "bbox": [ + 134, + 177, + 482, + 260 + ], + "type": "text", + "content": " of the selected model, and can be used to render arbitrary new perspectives." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 280, + 318, + 291 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 280, + 318, + 291 + ], + "spans": [ + { + "bbox": [ + 132, + 280, + 318, + 291 + ], + "type": "text", + "content": "3.4 Robust Neural Radiance Fields" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 302, + 482, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 482, + 397 + ], + "type": "text", + "content": "In the following, we propose a novel robust multi-view reconstruction approach based on NeRFs. NeRFs are fit to a single scene based on a set of observations. As they support view-dependent radiance, one might expect inconsistencies in the input observations to only have an effect on specific perspectives. However, the density is only spatially parametrized, hence, inconsistencies lead to significant ghosting and smearing artifacts in more than just the inconsistent perspective. This can be leveraged in the hypothesis validation. The following steps are performed for N iterations:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 410, + 482, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 410, + 482, + 470 + ], + "spans": [ + { + "bbox": [ + 134, + 410, + 482, + 470 + ], + "type": "text", + "content": "1. Hypothesis Consensus Set: A NeRF fit requires a large set of rays, making a sampling in ray space infeasible, as even obtaining significantly cleaner sets becomes unlikely. In order to obtain a sampling domain with reasonably sized initial consensus sets " + }, + { + "bbox": [ + 134, + 410, + 482, + 470 + ], + "type": "inline_equation", + "content": "S_{n}" + }, + { + "bbox": [ + 134, + 410, + 482, + 470 + ], + "type": "text", + "content": ", we sample " + }, + { + "bbox": [ + 134, + 410, + 482, + 470 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 134, + 410, + 482, + 470 + ], + "type": "text", + "content": " observations from the given sets of images " + }, + { + "bbox": [ + 134, + 410, + 482, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 134, + 410, + 482, + 470 + ], + "type": "text", + "content": " and corresponding camera poses " + }, + { + "bbox": [ + 134, + 410, + 482, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 134, + 410, + 482, + 470 + ], + "type": "text", + "content": " in every iteration:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 230, + 470, + 481, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 470, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 230, + 470, + 481, + 483 + ], + "type": "interline_equation", + "content": "\\left(I _ {n} ^ {m}, E _ {n} ^ {m}\\right) \\in_ {R} \\left\\{\\left(I _ {i}, E _ {i}\\right) \\mid I _ {i} \\in \\mathcal {I}, E _ {i} \\in \\mathcal {C} \\right\\} \\tag {5}", + "image_path": "31facd7dae5bc550a6ed3fc5d63b9581deab513ee3c781a062f1d98218381aa6.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 146, + 483, + 481, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 483, + 481, + 506 + ], + "spans": [ + { + "bbox": [ + 146, + 483, + 481, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 146, + 483, + 481, + 506 + ], + "type": "inline_equation", + "content": "n \\in \\{1, \\dots, N\\}" + }, + { + "bbox": [ + 146, + 483, + 481, + 506 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 146, + 483, + 481, + 506 + ], + "type": "inline_equation", + "content": "m \\in \\{1, \\dots, M\\}" + }, + { + "bbox": [ + 146, + 483, + 481, + 506 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 146, + 483, + 481, + 506 + ], + "type": "inline_equation", + "content": "\\in_R" + }, + { + "bbox": [ + 146, + 483, + 481, + 506 + ], + "type": "text", + "content": " denotes a sample randomly drawn from the set without replacement according to a uniform distribution." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 506, + 482, + 667 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 134, + 506, + 481, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 506, + 481, + 529 + ], + "spans": [ + { + "bbox": [ + 134, + 506, + 481, + 529 + ], + "type": "text", + "content": "2. Hypothesis Inference: The sampled observations are used to fit a hypothesis neural radiance field " + }, + { + "bbox": [ + 134, + 506, + 481, + 529 + ], + "type": "inline_equation", + "content": "F_{n}" + }, + { + "bbox": [ + 134, + 506, + 481, + 529 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 530, + 481, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 530, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 134, + 530, + 481, + 567 + ], + "type": "text", + "content": "3. Hypothesis Prediction: The obtained NeRF " + }, + { + "bbox": [ + 134, + 530, + 481, + 567 + ], + "type": "inline_equation", + "content": "F_{n}" + }, + { + "bbox": [ + 134, + 530, + 481, + 567 + ], + "type": "text", + "content": " is used to render predictions " + }, + { + "bbox": [ + 134, + 530, + 481, + 567 + ], + "type": "inline_equation", + "content": "I_{n,i}^{pred}" + }, + { + "bbox": [ + 134, + 530, + 481, + 567 + ], + "type": "text", + "content": " with pixel colors " + }, + { + "bbox": [ + 134, + 530, + 481, + 567 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{n,i,x}^{pred}" + }, + { + "bbox": [ + 134, + 530, + 481, + 567 + ], + "type": "text", + "content": " under the hypothesis for all unseen input perspectives " + }, + { + "bbox": [ + 134, + 530, + 481, + 567 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 134, + 530, + 481, + 567 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 567, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 567, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 134, + 567, + 482, + 667 + ], + "type": "text", + "content": "4. Hypothesis Validation: The chosen sample space requires a careful evaluation of the hypotheses. We propose a two-step evaluation, where, first, the pixels inliers " + }, + { + "bbox": [ + 134, + 567, + 482, + 667 + ], + "type": "inline_equation", + "content": "P_{n,i}" + }, + { + "bbox": [ + 134, + 567, + 482, + 667 + ], + "type": "text", + "content": " up to some margin " + }, + { + "bbox": [ + 134, + 567, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\epsilon_{pix}" + }, + { + "bbox": [ + 134, + 567, + 482, + 667 + ], + "type": "text", + "content": " are determined for every observation based on the Euclidean distance in color space to the pixels " + }, + { + "bbox": [ + 134, + 567, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{i,x}" + }, + { + "bbox": [ + 134, + 567, + 482, + 667 + ], + "type": "text", + "content": " of the input images " + }, + { + "bbox": [ + 134, + 567, + 482, + 667 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 134, + 567, + 482, + 667 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 134, + 567, + 482, + 667 + ], + "type": "inline_equation", + "content": "P_{n,i} = \\{\\pmb{p}_{n,i,x}^{pred} \\mid \\| \\pmb{p}_{n,i,x}^{pred} - \\pmb{p}_{i,x} \\|_2 < \\epsilon_{pix}\\}" + }, + { + "bbox": [ + 134, + 567, + 482, + 667 + ], + "type": "text", + "content": " and, second, the observations themselves are labeled as inliers or outliers based on the amount of pixel inliers, again up to some margin " + }, + { + "bbox": [ + 134, + 567, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\epsilon_{img}" + }, + { + "bbox": [ + 134, + 567, + 482, + 667 + ], + "type": "text", + "content": ", to obtain the consensus set " + }, + { + "bbox": [ + 134, + 567, + 482, + 667 + ], + "type": "inline_equation", + "content": "S_n^{inlier}" + }, + { + "bbox": [ + 134, + 567, + 482, + 667 + ], + "type": "text", + "content": " under the hypothesis: " + }, + { + "bbox": [ + 134, + 567, + 482, + 667 + ], + "type": "inline_equation", + "content": "S_n^{inlier} = \\{(I_i, E_i) \\mid \\# P_{n,i} > \\epsilon_{img}\\}" + }, + { + "bbox": [ + 134, + 567, + 482, + 667 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "text", + "content": "B. Buschmann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 145, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 145, + 116, + 479, + 140 + ], + "type": "text", + "content": "The binary metric for pixels ensures that smaller mispredictions (due to, e.g., view-dependent lighting effects) do not introduce noise into the evaluation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 140, + 482, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 140, + 482, + 188 + ], + "spans": [ + { + "bbox": [ + 134, + 140, + 482, + 188 + ], + "type": "text", + "content": "5. Model Selection: The strongest hypothesis, with its initial sample set " + }, + { + "bbox": [ + 134, + 140, + 482, + 188 + ], + "type": "inline_equation", + "content": "S_{best}" + }, + { + "bbox": [ + 134, + 140, + 482, + 188 + ], + "type": "text", + "content": ", is selected based on the number of inliers " + }, + { + "bbox": [ + 134, + 140, + 482, + 188 + ], + "type": "inline_equation", + "content": "\\# S_n^{inlier}" + }, + { + "bbox": [ + 134, + 140, + 482, + 188 + ], + "type": "text", + "content": " to obtain the final consensus set " + }, + { + "bbox": [ + 134, + 140, + 482, + 188 + ], + "type": "inline_equation", + "content": "S_{cons} = S_{best} \\cup S_{best}^{inlier}" + }, + { + "bbox": [ + 134, + 140, + 482, + 188 + ], + "type": "text", + "content": ". The final model is obtained with one more NeRF fit of the consensus set." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 250, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 250, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 250, + 217 + ], + "type": "text", + "content": "3.5 Hyperparameters" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 226, + 482, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 345 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 345 + ], + "type": "text", + "content": "For LFNs, the amount of initial samples and random hypotheses to evaluate (iterations), are determined experimentally. Without fine-tuning per class, the experimentally determined parameters are 90 initial samples and 2000 iterations, which supersedes the theoretical value for a convergence because the latent space introduces an intrinsic robustness. The inlier margin balances the amount of slight high-frequency variations that are being captured and the capability of separating outliers that are similar to the object. A margin of 0.25 in terms of the Euclidean distance of the predicted colors to the input samples in an RGB color space normalized to the range " + }, + { + "bbox": [ + 130, + 226, + 482, + 345 + ], + "type": "inline_equation", + "content": "(-1, 1)" + }, + { + "bbox": [ + 130, + 226, + 482, + 345 + ], + "type": "text", + "content": " has been found to be optimal. Please refer to the supplementary material for further experimental results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 346, + 482, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 490 + ], + "type": "text", + "content": "For NeRFs, the parameters behave more natural and the amount of perspectives required for a meaningful, not completely artifact-free fit of the model lies around 25 observations [29, Table 2]. With fewer samples, more artifacts are introduced that get harder to separate from the ones caused by inconsistencies, and the samples get more dependent on being evenly spaced. For real-world captures with " + }, + { + "bbox": [ + 130, + 346, + 482, + 490 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 346, + 482, + 490 + ], + "type": "text", + "content": " inconsistent perspectives, as few as 50 iterations are sufficient. With the color space normalized to " + }, + { + "bbox": [ + 130, + 346, + 482, + 490 + ], + "type": "inline_equation", + "content": "(0,1)" + }, + { + "bbox": [ + 130, + 346, + 482, + 490 + ], + "type": "text", + "content": ", a pixel margin between around 0.15 in terms of Euclidean distance worked well for the determination of actual artifacts. We consider an observation an inlier based on a margin of " + }, + { + "bbox": [ + 130, + 346, + 482, + 490 + ], + "type": "inline_equation", + "content": "90\\% - 98\\%" + }, + { + "bbox": [ + 130, + 346, + 482, + 490 + ], + "type": "text", + "content": " pixel inliers, which proved to be a good choice to separate minor artifacts (due to the sparse sampling) from artifacts caused by actual inconsistencies. For different datasets or inconsistencies, these values can be adapted." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 508, + 350, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 508, + 350, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 508, + 350, + 522 + ], + "type": "text", + "content": "4 Implementation & Preprocessing" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "text", + "content": "For LFNs, we build on top of the original implementation [40], with a slight adaptation to enable a parallel sub-sampled inference. We furthermore use the provided pretrained multi-class model. The camera parameters are known. For efficiency reasons, the steps of the algorithm are not performed iteratively, but multiple hypotheses are validated in parallel. To further speed up the inference, an exponential learning rate schedule is used for the auto-decoding, leading to a total runtime of about a minute on a single GPU." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "For the robust reconstruction of objects from lazily captured real-world data, one has to estimate the camera parameters and extract foreground masks before applying the algorithm. For the estimation of the camera parameters, we used the COLMAP structure-from-motion package [39]. We extracted foreground masks" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 90, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 90, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 405, + 90, + 447, + 101 + ], + "type": "text", + "content": "RANRAC" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "using Segment Anything [22]. However, only foreground masks, containing the objects and the occlusions, are extracted. Segment Anything is not capable of removing arbitrary occlusions in an automatized way. After these preprocessing steps, the robust reconstruction algorithm can be applied as described. Erroneous estimates of the camera parameters or foreground masks are excluded by our algorithm, thus making the entire reconstruction pipeline robust." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 189, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 189, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 482, + 297 + ], + "type": "text", + "content": "Our algorithm is not limited to a specific NeRF implementation. The chosen sampling domain eases integration into arbitrary existing NeRF implementations, which commonly expect images instead of unstructured ray sets. However, using a fast NeRF variant is advantageous when applying an iterative scheme. We used the instant NGP implementation [31] of the instant NSR repository [18], which includes some accelerations [24, 30]. Other, (specifically fast) variants are likely good choices as well. Antialiased and unbounded, but slow variants, such as MipNeRF360 [3], are not feasible. For further implementation details please refer to the supplementary." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 316, + 230, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 316, + 230, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 316, + 230, + 331 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 343, + 348, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 348, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 348, + 355 + ], + "type": "text", + "content": "5.1 Inconsistencies,Baselines & Datasets" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 365, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 569 + ], + "type": "text", + "content": "Multi-View Reconstruction (NeRF) We provide a comprehensive quantitative and qualitative analysis of RANRAC's multi-view reconstructions compared to NeRF without any method of robustness (baseline) and RobustNeRF [38] (state of the art). We conduct experiments for various common sources of inconsistencies such as occluded perspectives, noisy camera parameter estimates, and blurred perspectives. RobustNeRF [38] targets unbounded scenes with multiple objects and small amounts of distractors in every perspective. In comparison, our RANSAC-based approach deals well with single-object reconstruction, even with heavy occlusions, as long as enough clean perspectives are available. As their dataset reflects the algorithm's properties, we cannot provide a fair comparison. Instead, we demonstrate the applicability using a custom dataset of a single object with a controlled amount of deliberately occluded perspectives. For the other inconsistencies we use of-the-shelf datasets and add noise to the camera parameters and blur to the images. Furthermore, we implement the robust losses [38] on the same NeRF variant as RANRAC to provide a fair comparison of the method's robustness, independent of the NeRF variant. For further details on the implementation, please refer to the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 569, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 641 + ], + "type": "text", + "content": "Single-Shot Reconstruction (LFN) We benchmark against the original LFN implementation of Sitzmann et al. [40] as baseline, as there are no other robust methods for LFNs or conditioned neural fields, nor are there methods for robust single-shot multi-class reconstruction in general. Furthermore, we use the same pretrained LFN for the baseline and for the application of our method. The LFN is pretrained on the thirteen largest ShapeNet classes [8]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 482, + 666 + ], + "type": "text", + "content": "We provide a detailed qualitative and quantitative performance comparison under different amounts of occlusion for three representative classes (plane, car," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "text", + "content": "B. Buschmann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "and chair), while just stating reconstruction performance in a fixed environment without additional tuning of the hyperparameters for the others. The plane class is mostly challenging due to the low-frequency shape, while the car class contains a lot of high-frequency color details. The chair class represents shapes that are generally problematic for vanilla LFNs, even without occlusions. We provide a complementing analysis of the hyperparameters in the supplementary material. If not stated otherwise, we evaluate using 50 randomly selected images of the corresponding class. All comparisons use the same images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 482, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 296 + ], + "type": "text", + "content": "The occlusions are created synthetically. They consist of randomly generated patches while controlling two metrics of occlusion: Image occlusion and object occlusion. The former is the naive ratio of occluded over total pixels. The latter are the occluded pixels on the object compared to the total pixels covered by the object. We use both metrics to take the vastly different information entropy of samples across the image into account. For further details on the generation of the synthetic occlusions, please refer to the supplemental." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 312, + 215, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 215, + 323 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 215, + 323 + ], + "type": "text", + "content": "5.2 Evaluation" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 138, + 335, + 476, + 379 + ], + "blocks": [ + { + "bbox": [ + 138, + 335, + 476, + 379 + ], + "lines": [ + { + "bbox": [ + 138, + 335, + 476, + 379 + ], + "spans": [ + { + "bbox": [ + 138, + 335, + 476, + 379 + ], + "type": "table", + "html": "
MetricModelBenchBoatCabin.Displ.LampPhoneRifleSofaSpeak.Table
PSNR↑RANRAC19.2122.9221.9217.6519.9918.4521.3521.6120.720.44
LFN17.8919.220.518.8519.0917.8118.4620.1219.9920.29
SSIM↑RANRAC0.7670.8580.8010.6990.7640.750.8530.8050.7610.784
LFN0.7240.7910.7670.730.7480.7260.7950.7750.7430.777
", + "image_path": "3bed12e2e029c354e129405521e84d64e8b395ae40c59dda98bc8d7891795a77.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 380, + 482, + 467 + ], + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 467 + ], + "type": "text", + "content": "Table 1: RANRAC obtains a significant quantitative improvement in PSNR and SSIM (higher is better) compared to the baseline. We compare RANRAC to vanilla LFNs for the 13 largest ShapeNet classes (find plane, car, and chair with more detail in Fig. 3). The results are based on a moderate amount of occlusion of approximately " + }, + { + "bbox": [ + 130, + 380, + 482, + 467 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 130, + 380, + 482, + 467 + ], + "type": "text", + "content": " object occlusion and about " + }, + { + "bbox": [ + 130, + 380, + 482, + 467 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 380, + 482, + 467 + ], + "type": "text", + "content": " image occlusion. The reported results are conservative, as higher amounts of image occlusion result in a more significant performance increase (Fig. 3). No hyperparameter tuning has been performed for these classes; the same configuration obtained from the analysis of the other three classes is used." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": "For LFNs, our approach leads to a significant improvement in occluded scenarios of up to 8dB in PSNR and a similarly strong improvement for the SSIM. The improvement is most significant in heavily distracted scenarios (Fig. 3). In clean scenarios a slight performance penalty can be observed, but even with small amounts of object occlusion (information loss), our algorithm outperforms the baseline, leading to numerically better results up to " + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": " information loss (Fig. 3). The effect is not only measurable, but also well visible (Fig. 4). Increasing amounts of occlusion slowly introduce local artifacts into our reconstruction while preserving a reasonable shape estimate even for larger amounts of occlusion. In contrast, the reconstruction of LFNs breaks rather early in a global fashion. Still, our consensus set (Fig. 4), reveals that some high-frequency details were wrongfully excluded, explaining the slight performance decrease on clean images. In general, the benefit of RANRAC is best observable for classes that can be well described by LFNs, as evidenced by the lower improvements the chair class exhibits, compared to the plane and car class. The same effect is visible in the quantitative evaluation on the other classes without additional" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "text", + "content": "RANRAC" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 119, + 481, + 230 + ], + "blocks": [ + { + "bbox": [ + 134, + 119, + 481, + 230 + ], + "lines": [ + { + "bbox": [ + 134, + 119, + 481, + 230 + ], + "spans": [ + { + "bbox": [ + 134, + 119, + 481, + 230 + ], + "type": "image", + "image_path": "67c07b3b1a3d8dbe282989f20132f533f8508aec7cd161c7b0ad171a7fd958d0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 240, + 482, + 395 + ], + "lines": [ + { + "bbox": [ + 130, + 240, + 482, + 395 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 482, + 395 + ], + "type": "text", + "content": "Fig. 3: RANRAC (solid lines) leads to a quantitative improvement in PSNR and SSIM (higher is better) for occluded inputs compared to vanilla LFNs (dashed lines). The same hyperparameter configuration and LFN is used for all classes. On the left and in the middle, the amount of image occlusion is increased, while the object occlusion is constant at " + }, + { + "bbox": [ + 130, + 240, + 482, + 395 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 130, + 240, + 482, + 395 + ], + "type": "text", + "content": ". On the right, the amount of object occlusion is increased while the image occlusion is kept low. For the car class, a large improvement is observed over the entire occlusion spectrum. For the plane class the improvement is similarly significant, but absolute performance degenerates a bit sooner. This stems from the smaller object size and the related faster occlusion-to-object increase when increasing image occlusions. For the chair class, the improvement is less significant but the structural similarity is preserved for much longer. For the plane and car class the reconstruction quality is resilient to information loss (right) up to " + }, + { + "bbox": [ + 130, + 240, + 482, + 395 + ], + "type": "inline_equation", + "content": "\\sim 50\\%" + }, + { + "bbox": [ + 130, + 240, + 482, + 395 + ], + "type": "text", + "content": ", where the decrease gains momentum. With the low amounts of image occlusion, the improvement is not significant for the chair class (consistent with left and middle)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 145, + 409, + 473, + 572 + ], + "blocks": [ + { + "bbox": [ + 145, + 409, + 473, + 572 + ], + "lines": [ + { + "bbox": [ + 145, + 409, + 473, + 572 + ], + "spans": [ + { + "bbox": [ + 145, + 409, + 473, + 572 + ], + "type": "image", + "image_path": "79686c1d75bc7e7e181fe8e489ab0db2b4cf96e9f2fcfd30764ed3fc85bc3deb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 580, + 482, + 658 + ], + "lines": [ + { + "bbox": [ + 130, + 580, + 482, + 658 + ], + "spans": [ + { + "bbox": [ + 130, + 580, + 482, + 658 + ], + "type": "text", + "content": "Fig. 4: On the left, we show the qualitative effect of increasing occlusion on the same observation for the reconstruction of a novel view. Reconstructions of LFNs break early globally whereas RANRAC still provides a very decent reconstruction, only slowly introducing minor local (and natural/comprehensible) artifacts for completely hidden object parts. We further show the obtained consensus set, used for the final reconstruction (green inliers, red outliers). On the right, we show more qualitative results for novel view synthesis on different classes and the corresponding consensus sets." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 100 + ], + "type": "text", + "content": "B. Buschmann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 125, + 471, + 355 + ], + "blocks": [ + { + "bbox": [ + 141, + 125, + 471, + 355 + ], + "lines": [ + { + "bbox": [ + 141, + 125, + 471, + 355 + ], + "spans": [ + { + "bbox": [ + 141, + 125, + 471, + 355 + ], + "type": "image", + "image_path": "25a1557f50a85e6137cf682d89827c485cd7f03fa3be508e6ad29c12f8e5c6e5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 363, + 482, + 398 + ], + "lines": [ + { + "bbox": [ + 132, + 363, + 482, + 398 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 482, + 398 + ], + "type": "text", + "content": "Fig. 5: The occlusions lead to well-visible artifacts in the reconstructions of NeRF, these artifacts are completely removed by RANRAC. While RobustNeRF struggles with view-dependent and high-frequency details, RANRAC reliably reconstructs them." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 145, + 428, + 468, + 462 + ], + "blocks": [ + { + "bbox": [ + 145, + 428, + 468, + 462 + ], + "lines": [ + { + "bbox": [ + 145, + 428, + 468, + 462 + ], + "spans": [ + { + "bbox": [ + 145, + 428, + 468, + 462 + ], + "type": "table", + "html": "
Inconsistency DatasetMild Occ.Strong Occ.Watering PotBlurred Perspectives
LegoShipChairFicusMic
RANRAC27.1125.9926.1224.9434.7929.9129.7616.1435.2531.26
RobustNeRF26.8325.5825.9324.7929.1423.6923.3120.1931.1127.12
NeRF26.6522.6125.3618.4731.1519.0028.4820.3533.2122.10
", + "image_path": "510e3e55331ca81aac990b6deee12b237bebf28a7d864a5e32056f9226ae3c16.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 145, + 465, + 468, + 498 + ], + "blocks": [ + { + "bbox": [ + 145, + 465, + 468, + 498 + ], + "lines": [ + { + "bbox": [ + 145, + 465, + 468, + 498 + ], + "spans": [ + { + "bbox": [ + 145, + 465, + 468, + 498 + ], + "type": "table", + "html": "
Inconsistency \nDatasetNoisy Camera Parameters
LegoDrumsMicShipFicusHotdogMaterials
RANRAC34.95|31.2225.88|22.7735.85|33.5530.77|22.0531.41|28.8237.16|30.5328.93|25.39
RobustNeRF29.83|26.8223.91|22.1029.36|27.5524.21|19.9823.17|21.8532.65|25.1624.90|22.30
NeRF24.67|15.9523.22|17.9628.82|22.1523.68|14.8028.12|22.7728.16|18.3724.73|19.89
", + "image_path": "8f7cdf596133d6621b6a70c477567db6d181a657be275f9aaeab458edfbd9f25.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 640 + ], + "type": "text", + "content": "Table 2: RANRAC outperforms the state of the art RobustNeRF [38] and the baseline NeRF (without a method of robustness) for scenes contaminated with occlusions, blur and noisy camera parameters. We report PSNR" + }, + { + "bbox": [ + 130, + 498, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 130, + 498, + 482, + 640 + ], + "type": "text", + "content": " averaged across perspectives and the 5th percentile " + }, + { + "bbox": [ + 130, + 498, + 482, + 640 + ], + "type": "inline_equation", + "content": "(\\mathrm{Avg.}|P_5)" + }, + { + "bbox": [ + 130, + 498, + 482, + 640 + ], + "type": "text", + "content": " as artifacts introduced by inconsistencies only contaminate some views. We compare on the captured watering pot dataset with milder " + }, + { + "bbox": [ + 130, + 498, + 482, + 640 + ], + "type": "inline_equation", + "content": "(10\\%)" + }, + { + "bbox": [ + 130, + 498, + 482, + 640 + ], + "type": "text", + "content": " and heavier " + }, + { + "bbox": [ + 130, + 498, + 482, + 640 + ], + "type": "inline_equation", + "content": "(17.5\\%)" + }, + { + "bbox": [ + 130, + 498, + 482, + 640 + ], + "type": "text", + "content": " amounts of occluded perspectives and off-the-shelf datasets [29] with blurred perspectives and additive Gaussian noise " + }, + { + "bbox": [ + 130, + 498, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(5^{\\circ},1^{\\circ})" + }, + { + "bbox": [ + 130, + 498, + 482, + 640 + ], + "type": "text", + "content": " on the camera parameters of " + }, + { + "bbox": [ + 130, + 498, + 482, + 640 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 498, + 482, + 640 + ], + "type": "text", + "content": " of the perspectives. All three variants are built on top of instant-nsr for an isolated fair comparison of the robustness method. Note that both robust approaches struggle separating the strong view-dependent effects of the ship scene from blur, leading to the exclusion of some perspectives (lower " + }, + { + "bbox": [ + 130, + 498, + 482, + 640 + ], + "type": "inline_equation", + "content": "P_5" + }, + { + "bbox": [ + 130, + 498, + 482, + 640 + ], + "type": "text", + "content": " PSNR), while RANRAC still improves the overall reconstruction. For all other scenes and inconsistencies, RANRAC reliably separates inconsistent observations from clean ones." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "text", + "content": "RANRAC" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "tuning (Tab. 1). The only outlier is the display class, on which LFNs struggle the most on [40], even in unoccluded scenarios. It is not consistently represented in the latent space and robust reconstruction amplifies this effect by reducing the samples to a consistent set." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 165, + 481, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 165, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 165, + 481, + 248 + ], + "type": "text", + "content": "The application to NeRF shows the significantly improved reconstructions of RANRAC (Tab. 2) for different types of inconsistencies such as occlusions and blur, as well as noisy camera parameters on captured and synthetic datasets. RANRAC consistently outperforms RobustNeRF [38] for object reconstruction from inconsistent inputs. RobustNeRF specifically struggles in capturing view-dependent appearance and at concavities while RANRAC seamlessly reconstructs them. The effects are not only measurable but also well-visible (Fig. 5)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 271, + 317, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 271, + 317, + 284 + ], + "spans": [ + { + "bbox": [ + 131, + 271, + 317, + 284 + ], + "type": "text", + "content": "6 Limitations & Future Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 300, + 479, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 479, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 479, + 348 + ], + "type": "text", + "content": "For the application to NeRF, the iterations imply the use of a fast inferable variant, ruling out MipNeRF360 [3] and other high-quality variants for unbounded scenes. Foreground separation is a must, limiting this application of RANRAC to object-centric scenes. Novel fast, unbounded methods might resolve this [20]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 350, + 479, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 479, + 408 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 479, + 408 + ], + "type": "text", + "content": "By following a RANSAC-like approach, we inherit the requirement of sufficient clean perspectives, which could be lifted via NeRF variants that require fewer perspectives [32, 45] or by using different sampling domains. In return, our method is not limited to specific kinds of inconsistencies, and is robust to arbitrarily heavy distractions or inconsistencies in the impure perspectives." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 410, + 481, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 481, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 481, + 494 + ], + "type": "text", + "content": "Whereas our robust LFN approach led to significant improvements for single-shot reconstruction, our concept might also be applied to other conditioned neural fields, based on a smart choice of the sampling domain, targeting photorealism. One could also use importance sampling based on a prior, instead of a uniform sampling, leveraging the unevenly distributed information entropy. Neural sampling priors [6], semantic segmentation and more advanced schemes (e.g. LO-RANSAC [11], DSAC [5]) might prove useful here." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 516, + 220, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 516, + 220, + 529 + ], + "spans": [ + { + "bbox": [ + 131, + 516, + 220, + 529 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "type": "text", + "content": "We introduced a novel approach to increase the robustness of neural fields, inspired by the RANSAC paradigm. Following this concept, we introduced a novel robust approach for single-shot reconstruction from occluded views based on LFNs which achieves a significant improvement in reconstruction quality for distracted and occluded scenarios, even for extreme cases. Furthermore, we introduced a respective RANRAC-based NeRF variant that allows robust photorealistic reconstruction from multiple views with typical inconsistencies such as occlusions, noisy camera parameters, or blurred images - resulting in significant improvements compared to state-of-the-art methods - without relying on assumptions about the distractions." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "text", + "content": "B. Buschmann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 482, + 163 + ], + "type": "text", + "content": "Acknowledgements - Benno Buschmann was at FAU funded through a gift by Mitsubishi Electric Research Laboratories (MERL). Andreea Dogaru was funded by the German Federal Ministry of Education and Research (BMBF), FKZ: 01IS22082 (IRRW). The authors are responsible for the content of this publication." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 182, + 197, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 197, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 197, + 194 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 205, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 205, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 481, + 239 + ], + "type": "text", + "content": "1. Barron, J.T.: A general and adaptive robust loss function. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4331-4339 (2019) 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 239, + 481, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 481, + 271 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 481, + 271 + ], + "type": "text", + "content": "2. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: ICCV (2021) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 272, + 481, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 272, + 481, + 294 + ], + "spans": [ + { + "bbox": [ + 138, + 272, + 481, + 294 + ], + "type": "text", + "content": "3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In: CVPR (2022) 4, 10, 14" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 294, + 481, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 294, + 481, + 316 + ], + "spans": [ + { + "bbox": [ + 138, + 294, + 481, + 316 + ], + "type": "text", + "content": "4. Bian, W., Wang, Z., Li, K., Bian, J.W., Prisacariu, V.A.: Nope-nerf: Optimising neural radiance field with no pose prior. In: CVPR (2023) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 316, + 481, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 481, + 348 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 481, + 348 + ], + "type": "text", + "content": "5. Brachmann, E., Krull, A., Nowozin, S., Shotton, J., Michel, F., Gumhold, S., Rother, C.: Dsac-differentiable ransac for camera localization. In: CVPR (2017) 4, 14" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 348, + 481, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 481, + 371 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 481, + 371 + ], + "type": "text", + "content": "6. Brachmann, E., Rother, C.: Neural-guided ransac: Learning where to sample model hypotheses. In: ICCV (2019) 14" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 371, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 371, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 371, + 481, + 392 + ], + "type": "text", + "content": "7. Brown, M., Lowe, D.G.: Automatic panoramic image stitching using invariant features. International Journal of Computer Vision 74, 59-73 (2007) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 392, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 481, + 446 + ], + "type": "text", + "content": "8. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., Xiao, J., Yi, L., Yu, F.: ShapeNet: An Information-Rich 3D Model Repository. Tech. Rep. arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago (2015) 10" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 447, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 481, + 468 + ], + "type": "text", + "content": "9. Chen, X., Zhang, Q., Li, X., Chen, Y., Feng, Y., Wang, X., Wang, J.: Hallucinated neural radiance fields in the wild. In: CVPR (2022) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 468, + 481, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 481, + 491 + ], + "type": "text", + "content": "0. Choi, S., Kim, T., Yu, W.: Performance evaluation of ransac family. Journal of Computer Vision 24(3), 271-300 (1997) 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 491, + 481, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 481, + 523 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 481, + 523 + ], + "type": "text", + "content": "1. Chum, O., Matas, J., Kittler, J.: Locally optimized ransac. In: Pattern Recognition: 25th DAGM Symposium, Magdeburg, Germany, September 10-12, 2003. Proceedings 25. pp. 236-243. Springer (2003) 5, 14" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 523, + 481, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 481, + 545 + ], + "type": "text", + "content": "2. Derpanis, K.G.: Overview of the ransac algorithm. Image Rochester NY 4(1), 2-3 (2010) 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 545, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 481, + 567 + ], + "type": "text", + "content": "3. Egger, B., Schneider, A., Blumer, C., Forster, A., Schonborn, S., Vetter, T.: Occlusion-aware 3d morphable face models. In: BMVC. vol. 2, p. 4 (2016) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 567, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 481, + 610 + ], + "type": "text", + "content": "4. Egger, B., Schonborn, S., Schneider, A., Kortylewski, A., Morel-Forster, A., Blumer, C., Vetter, T.: Occlusion-aware 3d morphable models and an illumination prior for face image analysis. International Journal of Computer Vision 126, 1269-1287 (2018) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 481, + 643 + ], + "type": "text", + "content": "5. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM 24(6), 381-395 (1981) 2, 4, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 481, + 666 + ], + "type": "text", + "content": "6. Goli, L., Reading, C., Sellan, S., Jacobson, A., Tagliasacchi, A.: Bayes' Rays: Uncertainty quantification in neural radiance fields. arXiv (2023) 4" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "text", + "content": "RANRAC" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 663 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "17. Guangcong, Chen, Z., Loy, C.C., Liu, Z.: Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. ICCV (2023) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 160 + ], + "type": "text", + "content": "18. Guo, Y.C.: Instant neural surface reconstruction (2022), https://github.com/bennyguo/instant-nsr-pl 10" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "text", + "content": "19. Jeong, Y., Ahn, S., Choy, C., Anandkumar, A., Cho, M., Park, J.: Self-calibrating neural radiance fields. In: ICCV (2021) 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "text", + "content": "20. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering (2023) 14" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "text", + "content": "21. Kim, M., Seo, S., Han, B.: Infonerf: Ray entropy minimization for few-shot neural volume rendering. In: CVPR (2022) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "22. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023) 10" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "23. Lee, J., Kim, I., Heo, H., Kim, H.J.: Semantic-aware occlusion filtering neural radiance fields in the wild. arXiv preprint arXiv:2303.03966 (2023) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "type": "text", + "content": "24. Li, R., Gao, H., Tancik, M., Kanazawa, A.: Nerfacc: Efficient sampling accelerates nerfs. arXiv preprint arXiv:2305.04966 (2023) 10" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "25. Lin, C.H., Ma, W.C., Torralba, A., Lucey, S.: Barf: Bundle-adjusting neural radiance fields. In: ICCV (2021) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "text", + "content": "26. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust dynamic radiance fields. In: CVPR (2023) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "text", + "content": "27. Liu, Y., Peng, S., Liu, L., Wang, Q., Wang, P., Christian, T., Zhou, X., Wang, W.: Neural rays for occlusion-aware image-based rendering. In: CVPR (2022) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "28. Martin-Brualla, R., Radwan, N., Sajjadi, M.S.M., Barron, J.T., Dosovitskiy, A., Duckworth, D.: NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In: CVPR (2021) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "text", + "content": "29. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) 2, 3, 9, 13" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 436, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 436, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 480, + 445 + ], + "type": "text", + "content": "30. Müller, T.: tiny-cuda-nn (4 2021), https://github.com/NVlabs/tiny-cuda-nn 10" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "text", + "content": "31. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. In: SIGGRAPH (2022) 3, 10" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "text", + "content": "32. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S.M., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: CVPR (2022) 3, 14" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "text", + "content": "33. Philip, J., Deschaintre, V.: Floaters no more: Radiance field gradient scaling for improved near-camera training. In: Eurographics Symposium on Rendering. The Eurographics Association (2023) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "text", + "content": "34. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: CVPR (2021) 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 555, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 586 + ], + "type": "text", + "content": "35. Raguram, R., Frahm, J.M., Pollefeys, M.: A comparative analysis of ransac techniques leading to adaptive real-time random sample consensus. In: ECCV (2008) 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 588, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 610 + ], + "type": "text", + "content": "36. Rebain, D., Matthews, M., Yi, K.M., Lagun, D., Tagliasacchi, A.: Lolnerf: Learn from one look. In: CVPR (2022) 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "text", + "content": "37. Rematas, K., Liu, A., Srinivasan, P.P., Barron, J.T., Tagliasacchi, A., Funkhouser, T., Ferrari, V.: Urban radiance fields. CVPR (2022) 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 632, + 480, + 663 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 663 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 663 + ], + "type": "text", + "content": "38. Sabour, S., Vora, S., Duckworth, D., Krasin, I., Fleet, D.J., Tagliasacchi, A.: Robustnerf: Ignoring distractors with robust losses. In: CVPR (2023) 2, 3, 4, 10, 13, 14" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "text", + "content": "B. Buschmann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 481, + 555 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 137 + ], + "type": "text", + "content": "39. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: CVPR (2016) 9" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 481, + 171 + ], + "type": "text", + "content": "40. Sitzmann, V., Rezchikov, S., Freeman, W.T., Tenenbaum, J.B., Durand, F.: Light field networks: Neural scene representations with single-evaluation rendering. In: NeurIPS (2021) 2, 3, 4, 7, 9, 10, 14" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 194 + ], + "type": "text", + "content": "41. Sitzmann, V., Zollhöfer, M., Wetzstein, G.: Scene representation networks: Continuous 3d-structure-aware neural scene representations. In: NeurIPS (2019) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "type": "text", + "content": "42. Sun, J., Chen, X., Wang, Q., Li, Z., Averbuch-Elor, H., Zhou, X., Snavely, N.: Neural 3D reconstruction in the wild. In: SIGGRAPH (2022) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "text", + "content": "43. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: CVPR (2022) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "text", + "content": "44. Tewari, A., Thies, J., Mildenhall, B., et al.: Advances in neural rendering. In: Comput. Graph. Forum. vol. 41, pp. 703-735. Wiley Online Library, Wiley (5 2022) 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 282, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 314 + ], + "type": "text", + "content": "45. Truong, P., Rakotosaona, M.J., Manhardt, F., Tombari, F.: Sparf: Neural radiance fields from sparse and noisy poses. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4190-4200 (2023) 14" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "text", + "content": "46. Turki, H., Ramanan, D., Satyanarayanan, M.: Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In: CVPR (2022) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "text", + "content": "47. Warburg*, F., Weber*, E., Tancik, M., Holski, A., Kanazawa, A.: Nerfbusters: Removing ghostly artifacts from casually captured nerfs. In: ICCV (2023) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "text", + "content": "48. Wirth, T., Rak, A., Knauthe, V., Fellner, D.W.: A post processing technique to automatically remove floater artifacts in neural radiance fields. In: Computer Graphics Forum (2023) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 391, + 480, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 480, + 422 + ], + "type": "text", + "content": "49. Wu, T., Zhong, F., Tagliasacchi, A., Cole, F., Oztireli, C.: D\\~ 2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. NeurIPS (2022) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 423, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 423, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 423, + 480, + 456 + ], + "type": "text", + "content": "50. Xiangli, Y., Xu, L., Pan, X., Zhao, N., Rao, A., Theobalt, C., Dai, B., Lin, D.: Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In: ECCV (2022) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "text", + "content": "51. Xie, Y., Takikawa, T., Saito, S., Litany, O., Yan, S., Khan, N., Tombari, F., Tompkin, J., Sitzmann, V., Sridhar, S.: Neural fields in visual computing and beyond. Computer Graphics Forum 41 (2021) 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 490, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 511 + ], + "type": "text", + "content": "52. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: CVPR (2021) 2, 3, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "text", + "content": "53. Zhang, J.Y., Ramanan, D., Tulsiani, S.: Relpose: Predicting probabilistic relative rotation for single objects in the wild. In: ECCV (2022) 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "text", + "content": "54. Zhu, C., Wan, R., Tang, Y., Shi, B.: Occlusion-free scene recovery via neural radiance fields. In: CVPR (2023) 4" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "text", + "content": "RANRAC" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/RAP_ Retrieval-Augmented Planner for Adaptive Procedure Planning in Instructional Videos/8baf20b2-0efa-428f-a4d5-15cb7958ae91_content_list.json b/2024/RAP_ Retrieval-Augmented Planner for Adaptive Procedure Planning in Instructional Videos/8baf20b2-0efa-428f-a4d5-15cb7958ae91_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bf512aea9732cb33f005102a794791a03e714528 --- /dev/null +++ b/2024/RAP_ Retrieval-Augmented Planner for Adaptive Procedure Planning in Instructional Videos/8baf20b2-0efa-428f-a4d5-15cb7958ae91_content_list.json @@ -0,0 +1,1571 @@ +[ + { + "type": "text", + "text": "RAP: Retrieval-Augmented Planner for Adaptive Procedure Planning in Instructional Videos", + "text_level": 1, + "bbox": [ + 218, + 140, + 782, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ali Zare, Yulei Niu*, Hammad Ayyubi, and Shih-Fu Chang", + "bbox": [ + 287, + 212, + 715, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Columbia University, New York, NY 10027, USA {az2584,yn2338,ha2578,sc250}@columbia.edu,", + "bbox": [ + 336, + 239, + 665, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Procedure Planning in instructional videos entails generating a sequence of action steps based on visual observations of the initial and target states. Despite the rapid progress in this task, there remain several critical challenges to be solved: (1) Adaptive procedures: Prior works hold an unrealistic assumption that the number of action steps is known and fixed, leading to non-generalizable models in real-world scenarios where the sequence length varies. (2) Temporal relation: Understanding the step temporal relation knowledge is essential in producing reasonable and executable plans. (3) Annotation cost: Annotating instructional videos with step-level labels (i.e., timestamp) or sequence-level labels (i.e., action category) is demanding and labor-intensive, limiting its generalizability to large-scale datasets. In this work, we propose a new and practical setting, called adaptive procedure planning in instructional videos, where the procedure length is not fixed or pre-determined. To address these challenges, we introduce Retrieval-Augmented Planner (RAP) model. Specifically, for adaptive procedures, RAP adaptively determines the conclusion of actions using an auto-regressive model architecture. For temporal relation, RAP establishes an external memory module to explicitly retrieve the most relevant state-action pairs from the training videos and revises the generated procedures. To tackle high annotation cost, RAP utilizes a weakly-supervised learning manner to expand the training dataset to other task-relevant, unannotated videos by generating pseudo labels for action steps. Experiments on CrossTask and COIN benchmarks show the superiority of RAP over traditional fixed-length models, establishing it as a strong baseline solution for adaptive procedure planning.", + "bbox": [ + 261, + 301, + 743, + 661 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Procedure Planning, Instructional Video, Retrieval Augmentation, Weak Supervision", + "bbox": [ + 261, + 674, + 740, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 726, + 375, + 742 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Procedure planning in instructional videos aims to generate an executive series of actions for completing tasks, transitioning from initial visual state to final visual state [2, 4, 12, 16, 19, 22-24]. While humans excel at such tasks, machine learning agents face challenges in achieving similar proficiency. Success in procedure", + "bbox": [ + 212, + 756, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 218, + 825, + 382, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e7068740273b45094a058d2c55ec89fba3f9992dabac1717baaf041c2aef99bc.jpg", + "image_caption": [ + "Fig. 1: Examples of adaptive procedures in instructional videos. Procedure length varies with differences between initial and goal visual states, even for the same task. Previous works assume a fixed, known sequence length, requiring a separate model for each action-horizon. This figure contrasts traditional procedure planning (A) with our adaptive approach (B), which trains a unified model without knowing the procedure length in advance." + ], + "image_footnote": [], + "bbox": [ + 220, + 145, + 781, + 334 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "planning requires an agent's deep understanding of actions, procedure states, involved objects, and their temporal relationships. Narrowing this performance gap is key to advancing autonomous systems, equipping them to adeptly perform complex, instruction-based tasks like cooking or repair.", + "bbox": [ + 212, + 472, + 787, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although recent works have made tremendous strides, there remain several challenges to real-world applications. First, previous works assume that the number of action steps is known and fixed. As depicted in Fig. 1, the visual goal may be achieved through a small number of steps or a long procedure, depending on the differences between the initial and goal states. For example, it takes three steps to transform a visual of adding butter to a visual of flipped french toast in example b (scenario b implies pre-prepared mixture and pre-dipped bread, but visually requires actions like \"melt butter,\" \"put bread in pan,\" and \"flip bread\"). but one needs six and five steps to finish the cooking process based on the visuals of the mixture in scenarios a and c respectively. Previous models, trained on the unrealistic assumption of fixed prediction horizon, cannot figure out the complexity of state understanding and are hard to generalize to real-world applications. Secondly, there is a high correlation between consecutive action steps. For example, \"add\" is often followed by \"mix\", and \"wash\" is often followed by \"rinse\". Existing methods model this relationship implicitly via intermediate states, leading to sub-par results. Third, prior methods focused on fully annotated data [2,4,19] - action steps and intermediate visual states - for supervision. However, this approach poses challenges in scaling up the training data due to its costliness. Following works [12, 16, 22-24] proposed to leverage text representations of action steps as language supervision, which still necessitate", + "bbox": [ + 212, + 537, + 787, + 843 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "A. Zare et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sequence-level annotations, a time-consuming and expensive process involving watching entire instructional videos.", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To tackle these challenges, we introduce the setting of adaptive procedure planning in instructional videos and propose the Retrieval-Augmented Planner (RAP) model for this problem. Overall, RAP is designed to (1) adaptively predict procedures with variable lengths according to the complexity of visual states, (2) explicitly exploit step temporal relation via knowledge retrieval from the training data, and (3) learn procedural knowledge from additional unannotated videos. Specifically, RAP consists of an auto-regressive model as the base planner for adaptive procedure prediction, and a retrieval-augmented memory module to refine the generated procedures. The architecture is supplemented by weakly-supervised learning method that leverages both annotated video-procedure pairs and unannotated videos.", + "bbox": [ + 212, + 178, + 787, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition, existing metrics for procedure planning, such as accuracy, fail to fairly evaluate variable length procedures (as described in Sec. 4.2). We address this by introducing a new metric edit-score. We evaluate our model on two widely used datasets for instructional videos, CrossTask [26] and COIN [20]. Our model demonstrates superior performance in the conventional setting and establishes a strong baseline in the new adaptive procedure planning setting.", + "bbox": [ + 212, + 345, + 787, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our main contributions are:", + "bbox": [ + 238, + 436, + 444, + 450 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We argue the unrealistic assumption of fixed-length plans in previous procedure planning works, and introduce a new setting adaptive procedure planning in instructional videos.", + "- We propose Retrieval-Augmented Planner that can generate adaptive procedures of variable lengths and retrieve step temporal relation knowledge from the training data.", + "- We introduce a weakly-supervised training approach to learn procedural knowledge from unannotated task-relevant videos." + ], + "bbox": [ + 225, + 454, + 784, + 577 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 215, + 606, + 382, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Procedure Planning", + "text_level": 1, + "bbox": [ + 215, + 643, + 426, + 659 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Procedure planning from instructional videos, involves generating effective plans for task completion based on start and final visual observations. Earlier methods utilized a two-branch architecture, alternating between action and state prediction, with models like recurrent neural networks [4], transformers [19], and adversarial networks [2]. Recent approaches, such as P3IV's [24] memory-augmented Transformer and E3P, a prompting-based model [22], adopt one-branch transformer models. However, these methods are limited to predicting procedures of fixed lengths, lacking adaptability to tasks with varying action sequence length requirements. Our work introduces RAP, a novel one-branch approach designed to predict plans with variable action-horizons, addressing this limitation in the literature.", + "bbox": [ + 212, + 672, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "RAP", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Sequence Generation with Auto-regressive Transformers", + "text_level": 1, + "bbox": [ + 215, + 146, + 725, + 161 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recent advancements in sequence generation using Transformer-based models [17,21] have established remarkable results in tasks involving sequence modeling, e.g. natural language modeling [17]. Procedure planning in its core, is a goal-conditioned sequence generation problem, thereby directly benefiting from the state-of-the-art sequence generation models. Leveraging these developments, our work incorporates an auto-regressive Transformer [17] into the RAP architecture.", + "bbox": [ + 212, + 167, + 787, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Retrieval-Augmented Methods", + "text_level": 1, + "bbox": [ + 215, + 275, + 517, + 289 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Augmenting language models (LM) with external memory [3, 7, 10, 11, 25] has proved to significantly improve language sequence generation. P3IV [24] uses a memory component to enhance long-range sequence modeling in procedure planning. Similarly, we adopt a retrieval-based mechanism, drawing inspiration from Khandelwal et al. [10]'s augmentation of a neural language model with a k-nearest neighbors (kNN) retrieval component, in our auto-regressive model to effectively leverage temporal prior information for predicting action sequences.", + "bbox": [ + 212, + 295, + 785, + 401 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.4 Video Grounding", + "text_level": 1, + "bbox": [ + 215, + 419, + 405, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Procedure planning suffers from a lack of annotated data. Earlier works heavily relied on fully annotated data (action-state timestamps) [2, 4, 19]. Recent approaches [22, 24] have used language supervision during training which relies on less-costly, yet still resource-intensive language-level annotations. In this work, we introduce a weak language supervision framework that employs a video-language grounding algorithm to create pseudo-annotations from unannotated data. This approach is inspired by Dvornik et al. [5] in sequence-to-sequence alignment research, utilizing a grounding algorithm that aligns video frame and action step representations.", + "bbox": [ + 212, + 439, + 787, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Technical approach", + "text_level": 1, + "bbox": [ + 215, + 593, + 437, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we introduce our proposed framework Retrieval-Augmented Planner for adaptive procedure planning in instructional videos, and expand on a weakly-supervised learning method to leverage unannotated but task-relevant videos.", + "bbox": [ + 212, + 619, + 785, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Setting: Adaptive Procedure Planning", + "text_level": 1, + "bbox": [ + 215, + 698, + 578, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In conventional formulation of procedure planning from instructional videos, a model is required to generate a plan $p = \\{a_1, a_2, \\dots, a_T\\}$ , given visual observations of the start $o_s$ , and the goal $o_g$ states, where $T$ is the plan's length (i.e., action-horizon). Previous methods assume a fixed, known action-horizon, which is unrealistic for real-world applicability and limits the generalizability of planning models. As shown in Fig. 1, we consider a more general and practical setting, adaptive procedure planning in instructional videos, where the procedure length is not fixed.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "A. Zare et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a56bf2ea0c49957863420635fb13fb856d817bb74dedb0a7bbbc45528e66ffad.jpg", + "image_caption": [ + "Fig.2: Framework overview: Visual observations $o_s$ and $o_g$ are encoded to $v_s$ and $v_g$ . A Task-Classifier predicts a task class and its representation $\\hat{c}$ using $v_s$ and $v_g$ . These, along with a learnable START token, are fed into the auto-regressive base planner $F_{base}$ for sequential action embedding prediction. For each prediction at position $t$ , the retrieval component uses the predicted context vector $s_t$ to estimate a probability distribution $P_{\\mathrm{retr},t}$ over the closest $K$ key-value pairs in memory for kNN linear interpolation with $P_{\\mathrm{base},t}$ to estimate the final probability distribution $P_{\\mathrm{all},t}$ ." + ], + "image_footnote": [], + "bbox": [ + 246, + 145, + 725, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Model: Retrieval-Augmented Planner", + "text_level": 1, + "bbox": [ + 215, + 510, + 571, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose a new framework, Retrieval-Augmented Planner (RAP), to tackle the adaptive procedure planning problem. Fig. 2 illustrates our framework. As the first step, we obtain state $v_{i}$ , step $l_{j}$ , and task $c$ representations from visual observation $o_{i}$ , action step $a_{j}$ , and task title $tsk$ (e.g. \"make pancakes\") respectively. This is done by inputting $o_{i}$ , $a_{j}$ , and $tsk$ to a pretrained vision and language encoder respectively ([14]), similar to P3IV [24].", + "bbox": [ + 212, + 537, + 787, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Task Classifier. Recent studies [22, 23] have demonstrated the effectiveness of integrating task information into training to enhance model performance in procedure planning. In alignment with previous work [23], we adopt a similar strategy by first training an MLP task classifier. This classifier takes initial and final visual observations as inputs and predicts the task class $t\\hat{s}k$ . The training is conducted using a cross-entropy loss function with the ground-truth task class $tsk$ . Subsequently, we incorporate the task representation $\\hat{c}$ corresponding to the predicted task class $t\\hat{s}k$ as an additional input to both the Base-Planner and the Retrieval-Planner.", + "bbox": [ + 212, + 642, + 787, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Base Planner. To facilitate the adaptive procedure planning setting, we utilize an auto-regressive transformer [17] as the base planner $F_{\\mathrm{base}}$ . The base planner takes the representations of the initial state $v_{s}$ , the goal state $v_{g}$ , the predicted", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "RAP", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bf4372253c13069322c35a168a1a1279c7652df37e7ffa7d94deff8326ca5a89.jpg", + "image_caption": [ + "Fig. 3: The Retrieval-Augmented Planner module uses the predicted context vector $\\varsigma_{t}$ to retrieve the $K$ closest keys and their values from memory, estimating a probability distribution $P_{\\mathrm{retr},\\mathrm{t}}$ over these entries for next action probability estimation $P_{\\mathrm{all},\\mathrm{t}}$ ." + ], + "image_footnote": [], + "bbox": [ + 303, + 146, + 697, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "task $\\hat{c}$ , and the current action $a_{t}$ , as input, to predict the next action $a_{t+1}$ , i.e., $P_{\\mathrm{base}}(y|x) = P(a_{t+1}|a_{t}, v_{s}, v_{g}, \\hat{c})$ . The sequence starts with $a_{0} = \\mathrm{START}$ and completes upon predicting $a_{T+1} = \\mathrm{END}$ . Its implementation includes a learnable positional embedding $p_{t}$ for sequence generation. Therefore, the input sequence to predict $t$ -th action step $a_{t}$ is formulated as in Eq. (1),", + "bbox": [ + 214, + 371, + 787, + 446 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ v _ {s}, v _ {g}, \\hat {c}, l _ {\\text {S T A R T}}, l _ {1} + p _ {1}, \\dots , l _ {t - 2} + p _ {t - 2}, l _ {t - 1} + p _ {t - 1} \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 457, + 785, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and the output sequence for supervision is Eq. (2),", + "bbox": [ + 214, + 484, + 584, + 500 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nl = \\left[ l _ {1}, l _ {2}, \\dots , l _ {T}, l _ {\\text {E N D}} \\right] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 421, + 510, + 785, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $l_{t}$ represents the language representation for step $a_{t}$ , $l_{\\mathrm{START}}$ and $l_{\\mathrm{END}}$ are the embeddings for START and END tokens. The predicted action step $\\hat{a}_{t}$ is determined by mapping to a probability distribution over the action vocabulary $V_{a}$ (which includes the END token), which is achieved by comparing the language similarity between predicted $\\hat{l}_{t}$ and those within $V_{a}$ , enabling us to assign probabilities to each action within a plan sequence. During the training stage, we use the ground-truth action representation $l_{t}$ as input, while we use the predicted step representation $\\hat{l}_{t}$ during the test stage.", + "bbox": [ + 214, + 536, + 787, + 659 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Retrieval Planner. To improve long-range sequence modeling, our model integrates a direct retrieval mechanism based on a training datastore, utilizing its state-action temporal dependency information, explicitly. This enhancement is achieved by augmenting the base planner in a manner similar to [7,10], by adding a retrieval-augmented memory module, which learns the state-action pairs in the training data as memory, by modeling states, as a context for the actions. Fig. 3 illustrated the workings of this component.", + "bbox": [ + 214, + 672, + 787, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The memory module consists of key-value pairs $\\{s_n:d_n\\}_{n = 1}^D$ , where $D$ is the memory size. The keys $(s_n)$ , are discrete learnable context vectors and the values $(d_{n})$ are their subsequent action step labels, conditioned on the context. We estimate a context vector $\\varsigma_t$ , given representations of the initial visual state", + "bbox": [ + 214, + 779, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "A. Zare et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$v_{s}^{\\prime}$ , the target state $v_{g}^{\\prime}$ , the current action $l_{t}$ , and a learnable positional embedding $q_{t}$ as in Eq. (3)", + "bbox": [ + 215, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\varsigma_ {t} = f _ {\\text {P r o j}} \\left(\\operatorname {C o n c a t} \\left(v _ {s}, v _ {g}, \\hat {c}, l _ {t} + q _ {t}\\right)\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 178, + 784, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With $f_{\\mathrm{Proj}}$ as a projection layer and $\\mathrm{Concat}(\\cdot)$ concatenating inputs along the embedding dimension. In training, we learn the keys $(s_n)$ by comparing them to inputs' context vectors $\\varsigma_t$ . We construct the memory values $(d_n)$ by uniformly sampling actions from the training dataset (including the END token), making sure of a balanced representation of actions. At inference, the model retrieves the $K$ nearest (i.e., most similar) keys $\\{s_k\\}_{k=1}^K$ from the memory to a given current context $\\varsigma_t$ , returning their corresponding values $\\{d_k\\}_{k=1}^K$ . The nearest neighbors $(\\{s_k\\}_{k=1}^K)$ get selected by calculating the cosine distance between $\\varsigma_t$ and the keys of the datastore $\\{s_n\\}_{n=1}^D$ . Upon identifying the nearest neighbors, a probability distribution across the retrieved values $(\\{d_k\\}_{k=1}^K)$ can be calculated by applying a softmax function to the similarity scores. The retrieval process can be formulated as $P_{\\mathrm{retr}}(y|x) = P(a_{t+1}|\\varsigma_t, a_{t+1} \\in \\{d_k\\}_{k=1}^K)$ . The final prediction ensembles the predictions from base planner and retrieval planner. Following [7, 10], we obtain the final probability distribution for the next action step as:", + "bbox": [ + 215, + 198, + 784, + 409 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nP _ {\\text {a l l}} (y | x) = \\lambda P _ {\\text {r e t r}} (y | x) + (1 - \\lambda) P _ {\\text {b a s e}} (y | x) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 421, + 784, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\lambda$ is a tuned hyper-parameter.", + "bbox": [ + 215, + 446, + 478, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Training", + "text_level": 1, + "bbox": [ + 215, + 483, + 333, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "RAP is trained in two stages: First, we train the base planner, then we augment and train the model including the retrieval planner. This approach ensures effective learning of the retrieval planner's keys, building upon the base planner's good initialization obtained from the initial training stage.", + "bbox": [ + 215, + 507, + 782, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Stage 1: Training the Base Planner. During the training phase, we first supervise the base model's output, the textual representation of each action step in the procedure sequence, through language supervision combined with contrastive learning [8]: For each feature $\\hat{l}_t$ generated by the transformer's head, we take its ground truth language embedding $l_t$ as positive sample and all other embeddings in the action vocabulary $V_a$ as negative examples. The loss is formulated as:", + "bbox": [ + 215, + 568, + 784, + 657 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {b a s e}} = - \\sum_ {t = 1} ^ {T + 1} \\log \\frac {\\exp \\hat {l} _ {t} \\cdot l _ {t}}{\\sum_ {j} \\exp \\hat {l} _ {t} \\cdot l _ {j}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 670, + 784, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Where $(T + 1)$ -th step denotes the END token.", + "bbox": [ + 215, + 719, + 537, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Stage 2: Training the Base and Retrieval Planners. In the second stage, the model is augmented with the nearest neighbor retrieval-based memory. The retrieval planner advances learning by juxtaposing the one-hot ground-truth labels $a_{t} \\in p$ , with the model's interpolated probability estimates $P_{all,t}$ , for each action at each time-step $t$ . It refines the training by assessing how likely each action is, based on the model's predictions, instead of solely relying on matching", + "bbox": [ + 215, + 750, + 784, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "RAP", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the correct action. We achieve this, using a cross-entropy minimization in Eq. (6),", + "bbox": [ + 215, + 146, + 782, + 162 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l l}} = - \\sum_ {t = 1} ^ {T + 1} a _ {t} \\log P _ {\\text {a l l , t}} (y | x) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 400, + 172, + 785, + 213 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $P_{\\mathrm{all,t}}(y|x)$ is defined based on Eq. (4).", + "bbox": [ + 215, + 219, + 537, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Addition of Unannotated Videos with Weak Supervision", + "text_level": 1, + "bbox": [ + 215, + 256, + 725, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A significant challenge in procedure planning for instructional videos is the scarcity of annotated data. This is mainly because annotating instructional videos with step-level (i.e., timestamps) or sequence-level labels (i.e., action sequences) is resource-intensive. To address this issue, we propose a weakly-supervised training approach aimed at expanding the annotated training data to include additional unannotated task-related videos in the training process.", + "bbox": [ + 212, + 281, + 785, + 371 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Specifically, we align instructional videos with their corresponding topics, such as \"making an omelet\" or \"making pancakes\". This alignment links each video subset to a subset of instructional action plans sharing a similar topic. Crucially, this approach does not require detailed annotations that map specific action steps to individual video segments or frames. Instead, it relies on the broader thematic association between the video content and the instructional plans, allowing for a more general form of supervision during training. However, the acquired pairs of instructional videos and action plans may exhibit certain discrepancies. First, the presence of action steps in a plan that do not appear in the related video. Second, a mismatch in the sequence of action steps between the plan and their occurrence in the instructional video. Third, the inclusion of additional action steps in the video that are absent from the plan.", + "bbox": [ + 212, + 372, + 787, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To tackle these challenges, we applied a video grounding algorithm, taking inspiration from Dvornik et al. [5] to generate pseudo-annotations of descriptions in instructional videos. In short, our grounding algorithm processes an instructional video alongside an ordered sequence of action steps to produce a sequence of grounding results ( $t_{start}, t_{end}, a$ ), where $t_{start}$ and $t_{end}$ denote the beginning and end times of action step $a$ . The grounding model offers flexible sequence alignment and the ability to exclude irrelevant actions or frames, effectively addressing observed discrepancies. Further details on the grounding model are provided in the appendix. We utilize these pseudo-annotations, extracted for each relevant video, which serve as structurally and format-wise equivalent to manually annotated videos. These pseudo-annotations are integrated into Stage 1 training alongside original annotations, where the model is tasked with predicting the complete sequence of actions based on the starting frame of the first action and the ending frame of the last action outlined in the plan.", + "bbox": [ + 212, + 553, + 787, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This weak supervision of unannotated data, is an extension of Stage 1 training by using both original annotations of training videos and pseudo-annotations of task-relevant videos as the new training set. This training approach is a step towards a cost-effective solution for generalizing and scaling procedure planning to abundant unannotated videos.", + "bbox": [ + 212, + 765, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "A. Zare et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 145, + 374, + 162 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We experiment on two benchmark datasets and employ four metrics to validate the efficacy of our proposed model.", + "bbox": [ + 212, + 176, + 782, + 205 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/360de421978424c0e5446f04d3867e03a38349d712fd0ec021f793109550e5fc.jpg", + "image_caption": [ + "Fig. 4: Illustration of the advantage of edit score, in evaluation of long varied-length sequences. As it can be seen, despite the prediction being plausible and acceptable, accuracy and success rate, falsely flag this scenario as a miss." + ], + "image_footnote": [], + "bbox": [ + 276, + 233, + 725, + 354 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 215, + 464, + 333, + 477 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our methodology is evaluated using two instructional video datasets: CrossTask [26] and COIN [20]. CrossTask contains 2750 videos spanning 83 different topics, such as \"How to make french toast?\", with 7.6 actions per video on average. These topics are categorized into 18 primary and 65 related events. The primary topics are detailed with precise timestamps for each action step, providing a clear sequence of instructional activities. In contrast, the related instructional topics, gathered through automatic collection methods, lack annotations for action-step timestamps, sequence order, and the specific subset of action steps each video covers. Each topic within this related set is associated with a designated instructional plan as well as a set of unannotated instructional videos of the same topic. These videos may exhibit variations from the outlined plan, as detailed in Sec. 3.3, due to the previously mentioned discrepancies. We use CrossTask's related-set data, during weakly-supervised training with language supervision. COIN, includes 11827 videos, covering 778 different actions, with an average action-horizon of 3.6 actions per video. We follow [4,24] by creating train/test splits with a $70\\% / 30\\%$ ratio as well as implementing a moving window approach to organize the datasets into plans with various action-horizons.", + "bbox": [ + 212, + 488, + 787, + 744 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Metrics", + "text_level": 1, + "bbox": [ + 215, + 767, + 326, + 781 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our methodology's performance is evaluated using three standard metrics [2, 4, 19, 24]. First, mean Intersection over Union ( $mIoU$ ) assesses overlap between predicted and ground truth action sequences, calculated by $\\frac{|a_t \\cap \\hat{a}_t|}{|a_t \\cup \\hat{a}_t|}$ . This metric", + "bbox": [ + 212, + 792, + 784, + 843 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "RAP", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "evaluates whether the model identifies correct steps but does not consider action order or repetitions. Second, mean Accuracy ( $mAcc$ ) measures action alignment at each step, balancing the impact of repeated actions. Lastly, the Success Rate ( $SR$ ) evaluates a plan as successful if it exactly matches the ground truth.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "However, Accuracy and Success Rate have limitations, particularly with long or variable-length sequences, often leading to false negatives due to their sensitivity to action order. For example, in a pancake-making plan, if the predicted sequence adds a plausible step like \"Add sugar\" at the beginning, accuracy may incorrectly deem it inaccurate. This issue is more pronounced with sequences that vary in length from the ground truth, indicating the need for a more nuanced evaluation metric in procedure planning that accommodates plausible variations in actions.", + "bbox": [ + 212, + 209, + 784, + 329 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To address the limitations of the current metrics, we introduce a new metric, mean edit-score $(mES)$ , based on the minimum edit distance (e.g. Levenshtein edit distance), measuring the similarity between two sequences. This function, $L(s_{1},s_{2})|c_{sub},c_{ins},c_{del}$ , calculates the minimum number of substitutions $(c_{sub})$ , insertions $(c_{ins})$ , and deletions $(c_{del})$ needed to transform sequence $s_1$ into $s_2$ . The mean edit-score, is derived from the expected value of sample edit-scores. The edit-score is thus defined based on the optimal edit distance as:", + "bbox": [ + 214, + 333, + 782, + 439 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nE S _ {s _ {1}, s _ {2}} = 1 0 0 \\cdot \\frac {\\operatorname* {m a x} \\left(\\left| s _ {1} \\right| , \\left| s _ {2} \\right|\\right) - L \\left(s _ {1} , s _ {2}\\right) | _ {1 , 1 , 1}}{\\operatorname* {m a x} \\left(\\left| s _ {1} \\right| , \\left| s _ {2} \\right|\\right)}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 460, + 784, + 494 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Where the operator $|\\cdot|$ measures the length of a sequence. The mean edit-score offers a more flexible and forgiving assessment compared to strict metrics like success rate and mean accuracy, while still maintaining sensitivity to the order of action steps, hence complimenting $mIoU$ (i.e., if the ground-truth is add curry leaves, stir mixture, add chili powder and the prediction is add chili powder, stir mixture, add curry leaves, we face a high $IoU = 100\\%$ and a low $ES = 33.33\\%$ ). This metric provides a more comprehensive evaluation, especially in scenarios where the predicted sequence is plausible yet varies in actions or length from the ground truth (Fig. 4).", + "bbox": [ + 214, + 507, + 787, + 643 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 679, + 455, + 695 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Consistent with earlier approaches, we encode the start and goal visual observations using the S3D network [14] pre-trained on the HowTo100M [15] dataset. For optimization, we use the AdamW optimizer [13], setting the weight decay to 0.2. The learning rate is initially set at $lr = 0.001$ , and we apply a reduce-on-plateau scheduler to adjust this rate based on performance metrics. We experiment with an autoregressive transformer with 7 layers and 8 heads. We choose weight $\\lambda = 0.1$ and volume $D = 1000$ , for the memory module. Further implementation details can be found in the supplementary section.", + "bbox": [ + 212, + 719, + 782, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "A. Zare et al.", + "bbox": [ + 271, + 114, + 361, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/e1b991b92279f56a8f9111e9201d2e6942cf41099811bc46849354a490972e63.jpg", + "table_caption": [ + "Table 1: Assessment of procedure planning on CrossTask for action-horizons $T = 3$ and $T = 4$ . Vision supervision is denoted by V, language supervision by L, and weakly-supervised training by +wL." + ], + "table_footnote": [], + "table_body": "
ModelsSupervisionT=3T=4
SR↑mAcc↑mIoU↑SR↑mAcc↑mIoU↑
Random-<0.010.941.66<0.010.831.66
Retrieval-BasedV8.0523.3032.063.9522.2236.97
WLTDO [6]V1.8721.6431.700.7717.9226.43
UAAA [1]V2.1520.2130.870.9819.8627.09
UPN [18]V2.8924.3931.561.1921.5927.85
DDN [4]V12.1831.2947.485.9727.1048.46
Ext-GAIL w/o Aug. [2]V18.0143.8657.16---
Ext-GAIL [2]V21.2749.4661.7016.4143.0560.93
P3IV w/o Adv. [24]L22.1245.5767.40---
P3IV [24]L23.3449.9673.8913.4044.1670.01
E3P (baseline) [22]L22.5646.17-12.9743.97-
E3P [22]L26.4053.0274.0516.4948.0070.16
Skip-Plan [12]L28.8561.1874.9815.5655.6470.30
RAP (ours w/o unannotated data)L29.2362.3575.6316.9655.7771.49
RAP (ours with unannotated data)L+wL30.5162.6875.9217.1055.9171.33
", + "bbox": [ + 243, + 198, + 759, + 393 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/bc98c27ba8dc3c061c4e734d5e5834d2bce31cd0e541ee9164dcae33544a165b.jpg", + "table_caption": [ + "Table 2: Comparison of RAP with PDPP, when we align our state definition with PDPP's for CrossTask and $T = 3$ ." + ], + "table_footnote": [], + "table_body": "
ModelsSupervisionSR ↑mAcc ↑mIoU ↑
PDPP [23]♣L37.2064.6766.57
RAP (ours w/o unannotated data)♣L38.5165.1877.44
", + "bbox": [ + 287, + 445, + 715, + 486 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Comparison with State-of-The-Art Baselines", + "text_level": 1, + "bbox": [ + 214, + 513, + 630, + 529 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Our study conducts a comprehensive comparison of our new framework, RAP, against established models in procedure planning, using CrossTask and COIN datasets. We refer to our framework without the retrieval-based memory augmentation, i.e., base planner, as BP (as opposed to RAP). We categorize previous approaches based on their supervision: language (L) or vision (V). Our proposed model's performance compared to the previous works, is detailed in Tabs. 1 and 2 for CrossTask, and Tab. 3 for COIN.", + "bbox": [ + 212, + 537, + 787, + 643 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The inherent uncertainty in variable action-horizon procedure planning models complicates direct and fair comparisons with traditional fixed-length models, due to the added complexity of predicting the conclusion of a sequence as well as the broader range of acceptable sequences of variable length other than the ground-truth. To ensure a fair evaluation, we compare our variable action-horizon architecture, trained and evaluated exclusively on fixed-horizon data (of length $T$ ), to the fixed-horizon models (of length $T$ ) from prior studies in procedure planning [1,2,4,6,12,18,22-24]. Tab. 1 further shows the performance boost achieved by the addition of CrossTask's unannotated data (i.e., CrossTask's related video set) through weakly-supervised training (supervision: $\\mathrm{L} + \\mathrm{wL}$ ).", + "bbox": [ + 212, + 643, + 787, + 794 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Our findings demonstrate that RAP exhibits superior performance over existing vision-supervised models. Notably, RAP, without the addition of unannotated data, surpasses Ext-GAIL [2] with a $7.97\\%$ increase in Success Rate (SR),", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "RAP", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/47b60a9113d87110c4b4457d1a711474de8d1b55830c21a9aaefb6f899e4b26d.jpg", + "table_caption": [ + "Table 3: Comparison of the procedure planing models' performance on the COIN dataset." + ], + "table_footnote": [], + "table_body": "
ModelsT=3T=4
SR↑mAcc↑mIoU↑SR↑mAcc↑mIoU↑
Random (V)<0.01<0.012.47<0.01<0.012.32
Retrieval-Based (V)4.3817.4032.062.7114.2936.97
DDN (V)13.9020.1964.7811.1317.7168.06
P3IV (L)15.4021.6776.3111.3218.8570.53
E3P (L)19.5731.4284.9513.5926.7284.72
PDPP (L)21.3345.6251.8214.4144.1051.39
Skip-plan (L)23.6547.1278.4416.0443.1977.07
RAP (w/o unannotated) (L)24.4747.5985.2416.7642.1184.64
", + "bbox": [ + 287, + 184, + 718, + 305 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/35951e494cb90f987016867ac3f959cd9a19989bfb4fa9fe932c20bfe23d85c2.jpg", + "table_caption": [ + "Table 4: Evaluation of RAP versus a retrieval-based (RB) model in variable-length sequence prediction on CrossTask and COIN." + ], + "table_footnote": [], + "table_body": "
ModelsCrossTaskCOIN
SR↑mAcc↑mIoU↑mES↑SR↑mAcc↑mIoU↑mES↑
Random0.020.561.521.210.020.090.180.17
RB8.0121.3238.8033.437.1112.9420.1717.69
RAP (w/o unannotated data)20.2849.4169.9562.7919.1433.2675.8257.48
RAP (with unannotated data)21.7950.5871.4463.86----
", + "bbox": [ + 241, + 358, + 756, + 431 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "12.89% in Mean Accuracy (mAcc), and 13.93% in Mean Intersection over Union (mIoU) for sequence lengths of $T = 3$ in CrossTask.", + "bbox": [ + 212, + 460, + 782, + 491 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "When compared to language-supervised models like P3IV, E3P, and SkipPlan, RAP demonstrates superior performance on CrossTask. For $\\mathrm{T} = 3$ , RAP exhibits improvements of $5.99\\%$ , $2.83\\%$ , and $0.38\\%$ respectively, and for $\\mathrm{T} = 4$ , it shows enhancements of $3.56\\%$ , $0.47\\%$ , and $1.4\\%$ in SR. Tab. 2 compares our model with PDPP [23], which employs a different procedure state definition. For fairness, we evaluate our model under the PDPP setting, marked with $\\clubsuit$ , showcasing its superior performance. PDPP uses a 3-second window following the onset and preceding the offset of action steps to capture visual states, giving it an advantage in accessing action-related information compared to previous models. Traditionally, a 1-second window centered around the start and end of actions is used. By aligning our state definition with PDPP's, our RAP model outperforms PDPP on all metrics, as shown in the table. We observe a similar superiority in performance, when evaluating the model on the COIN dataset as shown in Tab. 3.", + "bbox": [ + 212, + 491, + 785, + 700 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We assessed the efficacy of RAP in handling tasks with variable lengths by conducting experiments against a retrieval-based model, which operates under a similar variable action-horizon setting. The results of this comparison are presented in Tab. 4. As demonstrated in the table, RAP proves a remarkable superiority over the retrieval-based model across all evaluated metrics. This is particularly evident in the mES metric, where RAP (w/o unannotated data) shows a significant improvement by $29.36\\%$ .", + "bbox": [ + 212, + 703, + 785, + 808 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "On CrossTask, RAP is further trained using the weak language supervision framework outlined in Sec. 3.3, utilizing the unannotated data from the dataset's", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "A. Zare et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/684b24b83b27f9996977c7231ba3e2778b0aa6cd4cf3c3386136ecfc790bbd22.jpg", + "table_caption": [ + "Table 5: A study on the effect of retrieval augmentation (BP vs RAP), for variable and fixed action-horizon settings." + ], + "table_footnote": [], + "table_body": "
DatasetModelsT=3T=4Variable Length T
SR↑mAcc↑mIoU↑mES↑SR↑mAcc↑mIoU↑mES↑SR↑mAcc↑mIoU↑mES↑
CrossTask (w/o unannotated)BP25.4258.7073.9165.5915.2352.1670.5361.6318.5447.8569.1662.64
RAP29.2362.3575.6367.3316.9655.7771.4964.5120.2849.4169.9562.79
COINBP23.3846.2180.4962.2814.2941.6778.6257.2617.8631.3472.4456.24
RAP24.4747.5985.2463.1516.7642.1184.6460.4719.1433.2675.8257.48
", + "bbox": [ + 217, + 184, + 785, + 247 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e9fb284d490254a8c076638e4ba694c576929103779744433c78025910d4458e.jpg", + "image_caption": [ + "Fig. 5: Sample plausible predicted plan sequences generated by RAP vs the ground-truth (GT)." + ], + "image_footnote": [], + "bbox": [ + 250, + 263, + 751, + 385 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "related-tasks. As shown in Tab. 1 and Tab. 4 when training RAP with the addition of the unannotated data, we see significant improvement in performance across all metrics.", + "bbox": [ + 212, + 457, + 784, + 501 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Fig. 5 displays two qualitative results of RAP's performance in predicting plans of varying lengths compared to the ground truths. In the top row, it accurately identifies the addition of fish, other ingredients, and the plausible action \"add onions\". In the bottom row, it recognizes the completed bread, its removal, and suggests the reasonable action \"melt butter\", before the new bread is put in the pan, showcasing the model's planning abilities.", + "bbox": [ + 212, + 503, + 784, + 593 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 617, + 388, + 633 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Impact of retrieval augmentation. The introduction of a retrieval-based memory component significantly enhances our framework's performance. Tab. 5 demonstrates improvements for both fixed and variable action-horizon settings. Notably, enhancements include increases of $3.81\\%$ , $1.73\\%$ , and $1.74\\%$ in Success Rate (SR) for $\\mathrm{T} = 3$ and $\\mathrm{T} = 4$ in the fixed setting, and in the variable setting, respectively, on CrossTask. We observe a similar boost in performance for the COIN dataset, as evidenced in the Tab. 5. Tab. 6 details the enhancements in SR, categorized by the ground-truth (GT) sequences' action-horizon in the variable-length sequence prediction setting.", + "bbox": [ + 212, + 643, + 785, + 779 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Impact of weak language supervision. The influence of weak language supervision on our model's performance is evaluated through an ablation study on CrossTask, where we incrementally increased the unannotated data used in weakly-supervised training. Tab. 7 displays the varying percentages of unan-", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "RAP", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/239677a551f172f3f40b1e4d5fafc07fe81b870a73312def807b9a6550175a22.jpg", + "table_caption": [ + "Table 6: Impact of retrieval augmentation (BP vs RAP) in $SR$ , on CrossTask, categorized by ground-truth (GT) action-horizon in the variable action-horizon setting." + ], + "table_footnote": [], + "table_body": "
Model (w/o unannotated data)GT action-horizon
123456
BP49.5827.8416.218.375.893.34
RAP52.6029.2018.0210.526.165.18
", + "bbox": [ + 315, + 183, + 689, + 239 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/0599516968a28c5c46aaa7e7d74ae1b37ba89bfde0a0f244cd0e06d9387d8471.jpg", + "table_caption": [ + "Table 7: Weak language supervision impact: Table shows RAP's performance gains on CrossTask with increasing unannotated data use." + ], + "table_footnote": [], + "table_body": "
Used unannotated data portionSR↑mAcc↑mIoU↑mES↑
No unannotated data20.2849.4169.9562.79
25%19.2047.8869.0561.96
50%20.5749.7269.7563.12
75%21.3150.2270.8663.40
100%21.7950.5871.4463.86
", + "bbox": [ + 328, + 291, + 674, + 391 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "notated data utilized. The results show consistent performance gains for RAP under weak language supervision, particularly as more data is introduced (except at the $25\\%$ data level). This suggests potential for further enhancements with increased unannotated data. Notably, using $100\\%$ of unannotated data yields improvements in SR, mAcc, mIoU, and mES by $1.51\\%$ , $1.17\\%$ , $1.49\\%$ , and $1.07\\%$ , respectively.", + "bbox": [ + 212, + 417, + 787, + 510 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 532, + 359, + 549 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this study, we addressed the challenge of procedure planning in instructional videos by developing the Retrieval-Augmented Planner (RAP) model. This model overcomes the limitations of previous approaches that were restricted to fixed action-step predictions. RAP excels in both fixed and varied action horizon scenarios, outperforming existing models. A key innovation is our weak language supervision method, which uses thematic links between unannotated videos and plans for pseudo-annotation generation, enhancing model training. Our model also incorporates a learnable retrieval-based memory component for improved long-range sequence predictions and efficient use of temporal data. Additionally, we introduced \"edit-score\", a new metric for assessing variable-length procedures. Looking forward, we see potential in applying RAP to more diverse scenarios and in generating probabilistic, varied-length plans. We also envisage that expanding our weak supervision approach to include an even wider range of unannotated data could markedly elevate the performance of procedure planning models, paving the way for more advanced applications in this field.", + "bbox": [ + 212, + 565, + 787, + 792 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "A. Zare et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abu Farha, Y., Gall, J.: Uncertainty-aware anticipation of activities. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. pp. 0-0 (2019)", + "2. Bi, J., Luo, J., Xu, C.: Procedure planning in instructional videos via contextual modeling and model-based policy learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15611-15620 (2021)", + "3. Borgeaud, S., Mensch, A., Hoffmann, J., Cai, T., Rutherford, E., Millican, K., Van Den Driessche, G.B., Lespiau, J.B., Damoc, B., Clark, A., et al.: Improving language models by retrieving from trillions of tokens. In: International conference on machine learning. pp. 2206-2240. PMLR (2022)", + "4. Chang, C.Y., Huang, D.A., Xu, D., Adeli, E., Fei-Fei, L., Niebles, J.C.: Procedure planning in instructional videos. In: European Conference on Computer Vision. pp. 334-350. Springer (2020)", + "5. Dvornik, M., Hadji, I., Derpanis, K.G., Garg, A., Jepson, A.: Drop-dtw: Aligning common signal between sequences while dropping outliers. Advances in Neural Information Processing Systems 34, 13782-13793 (2021)", + "6. Ehsani, K., Bagherinezhad, H., Redmon, J., Mottaghi, R., Farhadi, A.: Who let the dogs out? modeling dog behavior from visual data. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4051-4060 (2018)", + "7. Grave, E., Cisse, M.M., Joulin, A.: Unbounded cache model for online language modeling with open vocabulary. Advances in neural information processing systems 30 (2017)", + "8. Gutmann, M., Hyvarinen, A.: Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In: Proceedings of the thirteenth international conference on artificial intelligence and statistics. pp. 297-304. JMLR Workshop and Conference Proceedings (2010)", + "9. Hadji, I., Derpanis, K.G., Jepson, A.D.: Representation learning via global temporal alignment and cycle-consistency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11068-11077 (2021)", + "0. Khandelwal, U., Levy, O., Jurafsky, D., Zettlemoyer, L., Lewis, M.: Generalization through memorization: Nearest neighbor language models. arXiv preprint arXiv:1911.00172 (2019)", + "1. Lei, J., Wang, L., Shen, Y., Yu, D., Berg, T.L., Bansal, M.: Mart: Memory-augmented recurrent transformer for coherent video paragraph captioning. arXiv preprint arXiv:2005.05402 (2020)", + "2. Li, Z., Geng, W., Li, M., Chen, L., Tang, Y., Lu, J., Zhou, J.: Skip-plan: Procedure planning in instructional videos via condensed action space learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10297-10306 (2023)", + "3. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)", + "4. Miech, A., Alayrac, J.B., Smaira, L., Laptev, I., Sivic, J., Zisserman, A.: End-to-end learning of visual representations from uncurated instructional videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9879-9889 (2020)", + "5. Miech, A., Zhukov, D., Alayrac, J.B., Tapaswi, M., Laptev, I., Sivic, J.: Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2630-2640 (2019)" + ], + "bbox": [ + 225, + 172, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "RAP", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Niu, Y., Guo, W., Chen, L., Lin, X., Chang, S.F.: Schema: State changes matter for procedure planning in instructional videos. arXiv preprint arXiv:2403.01599 (2024)", + "17. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019)", + "18. Srinivas, A., Jabri, A., Abbeel, P., Levine, S., Finn, C.: Universal planning networks: Learning generalizable representations for visuomotor control. In: International Conference on Machine Learning. pp. 4732-4741. PMLR (2018)", + "19. Sun, J., Huang, D.A., Lu, B., Liu, Y.H., Zhou, B., Garg, A.: Plate: Visually-grounded planning with transformers in procedural tasks. IEEE Robotics and Automation Letters 7(2), 4924-4930 (2022)", + "20. Tang, Y., Ding, D., Rao, Y., Zheng, Y., Zhang, D., Zhao, L., Lu, J., Zhou, J.: Coin: A large-scale dataset for comprehensive instructional video analysis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1207-1216 (2019)", + "21. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)", + "22. Wang, A.L., Lin, K.Y., Du, J.R., Meng, J., Zheng, W.S.: Event-guided procedure planning from instructional videos with text supervision. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13565-13575 (2023)", + "23. Wang, H., Wu, Y., Guo, S., Wang, L.: Pdpp: Projected diffusion for procedure planning in instructional videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14836-14845 (2023)", + "24. Zhao, H., Hadji, I., Dvornik, N., Derpanis, K.G., Wildes, R.P., Jepson, A.D.: P3iv: Probabilistic procedure planning from instructional videos with weak supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2938-2948 (2022)", + "25. Zhong, Z., Lei, T., Chen, D.: Training language models with memory augmentation (2022)", + "26. Zhukov, D., Alayrac, J.B., Cinbis, R.G., Fouhey, D., Laptev, I., Sivic, J.: Cross-task weakly supervised learning from instructional videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3537-3545 (2019)" + ], + "bbox": [ + 215, + 146, + 785, + 618 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "A. Zare et al.", + "bbox": [ + 271, + 114, + 361, + 126 + ], + "page_idx": 15 + } +] \ No newline at end of file diff --git a/2024/RAP_ Retrieval-Augmented Planner for Adaptive Procedure Planning in Instructional Videos/8baf20b2-0efa-428f-a4d5-15cb7958ae91_model.json b/2024/RAP_ Retrieval-Augmented Planner for Adaptive Procedure Planning in Instructional Videos/8baf20b2-0efa-428f-a4d5-15cb7958ae91_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fe86fbc6cb29d1e9287fe0d4da67cb4be4e11def --- /dev/null +++ b/2024/RAP_ Retrieval-Augmented Planner for Adaptive Procedure Planning in Instructional Videos/8baf20b2-0efa-428f-a4d5-15cb7958ae91_model.json @@ -0,0 +1,1937 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.22, + 0.141, + 0.784, + 0.187 + ], + "angle": 0, + "content": "RAP: Retrieval-Augmented Planner for Adaptive Procedure Planning in Instructional Videos" + }, + { + "type": "text", + "bbox": [ + 0.289, + 0.213, + 0.717, + 0.229 + ], + "angle": 0, + "content": "Ali Zare, Yulei Niu*, Hammad Ayyubi, and Shih-Fu Chang" + }, + { + "type": "text", + "bbox": [ + 0.337, + 0.24, + 0.666, + 0.269 + ], + "angle": 0, + "content": "Columbia University, New York, NY 10027, USA {az2584,yn2338,ha2578,sc250}@columbia.edu," + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.302, + 0.744, + 0.662 + ], + "angle": 0, + "content": "Abstract. Procedure Planning in instructional videos entails generating a sequence of action steps based on visual observations of the initial and target states. Despite the rapid progress in this task, there remain several critical challenges to be solved: (1) Adaptive procedures: Prior works hold an unrealistic assumption that the number of action steps is known and fixed, leading to non-generalizable models in real-world scenarios where the sequence length varies. (2) Temporal relation: Understanding the step temporal relation knowledge is essential in producing reasonable and executable plans. (3) Annotation cost: Annotating instructional videos with step-level labels (i.e., timestamp) or sequence-level labels (i.e., action category) is demanding and labor-intensive, limiting its generalizability to large-scale datasets. In this work, we propose a new and practical setting, called adaptive procedure planning in instructional videos, where the procedure length is not fixed or pre-determined. To address these challenges, we introduce Retrieval-Augmented Planner (RAP) model. Specifically, for adaptive procedures, RAP adaptively determines the conclusion of actions using an auto-regressive model architecture. For temporal relation, RAP establishes an external memory module to explicitly retrieve the most relevant state-action pairs from the training videos and revises the generated procedures. To tackle high annotation cost, RAP utilizes a weakly-supervised learning manner to expand the training dataset to other task-relevant, unannotated videos by generating pseudo labels for action steps. Experiments on CrossTask and COIN benchmarks show the superiority of RAP over traditional fixed-length models, establishing it as a strong baseline solution for adaptive procedure planning." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.675, + 0.741, + 0.703 + ], + "angle": 0, + "content": "Keywords: Procedure Planning, Instructional Video, Retrieval Augmentation, Weak Supervision" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.727, + 0.377, + 0.743 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.757, + 0.788, + 0.819 + ], + "angle": 0, + "content": "Procedure planning in instructional videos aims to generate an executive series of actions for completing tasks, transitioning from initial visual state to final visual state [2, 4, 12, 16, 19, 22-24]. While humans excel at such tasks, machine learning agents face challenges in achieving similar proficiency. Success in procedure" + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.826, + 0.383, + 0.841 + ], + "angle": 0, + "content": "* Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "A. Zare et al." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.146, + 0.782, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.348, + 0.788, + 0.433 + ], + "angle": 0, + "content": "Fig. 1: Examples of adaptive procedures in instructional videos. Procedure length varies with differences between initial and goal visual states, even for the same task. Previous works assume a fixed, known sequence length, requiring a separate model for each action-horizon. This figure contrasts traditional procedure planning (A) with our adaptive approach (B), which trains a unified model without knowing the procedure length in advance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.534 + ], + "angle": 0, + "content": "planning requires an agent's deep understanding of actions, procedure states, involved objects, and their temporal relationships. Narrowing this performance gap is key to advancing autonomous systems, equipping them to adeptly perform complex, instruction-based tasks like cooking or repair." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Although recent works have made tremendous strides, there remain several challenges to real-world applications. First, previous works assume that the number of action steps is known and fixed. As depicted in Fig. 1, the visual goal may be achieved through a small number of steps or a long procedure, depending on the differences between the initial and goal states. For example, it takes three steps to transform a visual of adding butter to a visual of flipped french toast in example b (scenario b implies pre-prepared mixture and pre-dipped bread, but visually requires actions like \"melt butter,\" \"put bread in pan,\" and \"flip bread\"). but one needs six and five steps to finish the cooking process based on the visuals of the mixture in scenarios a and c respectively. Previous models, trained on the unrealistic assumption of fixed prediction horizon, cannot figure out the complexity of state understanding and are hard to generalize to real-world applications. Secondly, there is a high correlation between consecutive action steps. For example, \"add\" is often followed by \"mix\", and \"wash\" is often followed by \"rinse\". Existing methods model this relationship implicitly via intermediate states, leading to sub-par results. Third, prior methods focused on fully annotated data [2,4,19] - action steps and intermediate visual states - for supervision. However, this approach poses challenges in scaling up the training data due to its costliness. Following works [12, 16, 22-24] proposed to leverage text representations of action steps as language supervision, which still necessitate" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.696, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RAP" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "sequence-level annotations, a time-consuming and expensive process involving watching entire instructional videos." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.344 + ], + "angle": 0, + "content": "To tackle these challenges, we introduce the setting of adaptive procedure planning in instructional videos and propose the Retrieval-Augmented Planner (RAP) model for this problem. Overall, RAP is designed to (1) adaptively predict procedures with variable lengths according to the complexity of visual states, (2) explicitly exploit step temporal relation via knowledge retrieval from the training data, and (3) learn procedural knowledge from additional unannotated videos. Specifically, RAP consists of an auto-regressive model as the base planner for adaptive procedure prediction, and a retrieval-augmented memory module to refine the generated procedures. The architecture is supplemented by weakly-supervised learning method that leverages both annotated video-procedure pairs and unannotated videos." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.346, + 0.788, + 0.435 + ], + "angle": 0, + "content": "In addition, existing metrics for procedure planning, such as accuracy, fail to fairly evaluate variable length procedures (as described in Sec. 4.2). We address this by introducing a new metric edit-score. We evaluate our model on two widely used datasets for instructional videos, CrossTask [26] and COIN [20]. Our model demonstrates superior performance in the conventional setting and establishes a strong baseline in the new adaptive procedure planning setting." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.438, + 0.445, + 0.451 + ], + "angle": 0, + "content": "Our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.455, + 0.784, + 0.5 + ], + "angle": 0, + "content": "- We argue the unrealistic assumption of fixed-length plans in previous procedure planning works, and introduce a new setting adaptive procedure planning in instructional videos." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.502, + 0.785, + 0.546 + ], + "angle": 0, + "content": "- We propose Retrieval-Augmented Planner that can generate adaptive procedures of variable lengths and retrieve step temporal relation knowledge from the training data." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.548, + 0.785, + 0.578 + ], + "angle": 0, + "content": "- We introduce a weakly-supervised training approach to learn procedural knowledge from unannotated task-relevant videos." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.455, + 0.785, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.607, + 0.383, + 0.623 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.644, + 0.428, + 0.66 + ], + "angle": 0, + "content": "2.1 Procedure Planning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Procedure planning from instructional videos, involves generating effective plans for task completion based on start and final visual observations. Earlier methods utilized a two-branch architecture, alternating between action and state prediction, with models like recurrent neural networks [4], transformers [19], and adversarial networks [2]. Recent approaches, such as P3IV's [24] memory-augmented Transformer and E3P, a prompting-based model [22], adopt one-branch transformer models. However, these methods are limited to predicting procedures of fixed lengths, lacking adaptability to tasks with varying action sequence length requirements. Our work introduces RAP, a novel one-branch approach designed to predict plans with variable action-horizons, addressing this limitation in the literature." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.363, + 0.128 + ], + "angle": 0, + "content": "A. Zare et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.727, + 0.162 + ], + "angle": 0, + "content": "2.2 Sequence Generation with Auto-regressive Transformers" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.168, + 0.788, + 0.259 + ], + "angle": 0, + "content": "Recent advancements in sequence generation using Transformer-based models [17,21] have established remarkable results in tasks involving sequence modeling, e.g. natural language modeling [17]. Procedure planning in its core, is a goal-conditioned sequence generation problem, thereby directly benefiting from the state-of-the-art sequence generation models. Leveraging these developments, our work incorporates an auto-regressive Transformer [17] into the RAP architecture." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.276, + 0.518, + 0.29 + ], + "angle": 0, + "content": "2.3 Retrieval-Augmented Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.296, + 0.787, + 0.402 + ], + "angle": 0, + "content": "Augmenting language models (LM) with external memory [3, 7, 10, 11, 25] has proved to significantly improve language sequence generation. P3IV [24] uses a memory component to enhance long-range sequence modeling in procedure planning. Similarly, we adopt a retrieval-based mechanism, drawing inspiration from Khandelwal et al. [10]'s augmentation of a neural language model with a k-nearest neighbors (kNN) retrieval component, in our auto-regressive model to effectively leverage temporal prior information for predicting action sequences." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.42, + 0.406, + 0.435 + ], + "angle": 0, + "content": "2.4 Video Grounding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.44, + 0.788, + 0.577 + ], + "angle": 0, + "content": "Procedure planning suffers from a lack of annotated data. Earlier works heavily relied on fully annotated data (action-state timestamps) [2, 4, 19]. Recent approaches [22, 24] have used language supervision during training which relies on less-costly, yet still resource-intensive language-level annotations. In this work, we introduce a weak language supervision framework that employs a video-language grounding algorithm to create pseudo-annotations from unannotated data. This approach is inspired by Dvornik et al. [5] in sequence-to-sequence alignment research, utilizing a grounding algorithm that aligns video frame and action step representations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.594, + 0.439, + 0.612 + ], + "angle": 0, + "content": "3 Technical approach" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.621, + 0.787, + 0.681 + ], + "angle": 0, + "content": "In this section, we introduce our proposed framework Retrieval-Augmented Planner for adaptive procedure planning in instructional videos, and expand on a weakly-supervised learning method to leverage unannotated but task-relevant videos." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.699, + 0.579, + 0.715 + ], + "angle": 0, + "content": "3.1 Setting: Adaptive Procedure Planning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In conventional formulation of procedure planning from instructional videos, a model is required to generate a plan \\( p = \\{a_1, a_2, \\dots, a_T\\} \\), given visual observations of the start \\( o_s \\), and the goal \\( o_g \\) states, where \\( T \\) is the plan's length (i.e., action-horizon). Previous methods assume a fixed, known action-horizon, which is unrealistic for real-world applicability and limits the generalizability of planning models. As shown in Fig. 1, we consider a more general and practical setting, adaptive procedure planning in instructional videos, where the procedure length is not fixed." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.696, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RAP" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.146, + 0.726, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.383, + 0.788, + 0.482 + ], + "angle": 0, + "content": "Fig.2: Framework overview: Visual observations \\( o_s \\) and \\( o_g \\) are encoded to \\( v_s \\) and \\( v_g \\). A Task-Classifier predicts a task class and its representation \\( \\hat{c} \\) using \\( v_s \\) and \\( v_g \\). These, along with a learnable START token, are fed into the auto-regressive base planner \\( F_{base} \\) for sequential action embedding prediction. For each prediction at position \\( t \\), the retrieval component uses the predicted context vector \\( s_t \\) to estimate a probability distribution \\( P_{\\mathrm{retr},t} \\) over the closest \\( K \\) key-value pairs in memory for kNN linear interpolation with \\( P_{\\mathrm{base},t} \\) to estimate the final probability distribution \\( P_{\\mathrm{all},t} \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.511, + 0.573, + 0.527 + ], + "angle": 0, + "content": "3.2 Model: Retrieval-Augmented Planner" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.631 + ], + "angle": 0, + "content": "We propose a new framework, Retrieval-Augmented Planner (RAP), to tackle the adaptive procedure planning problem. Fig. 2 illustrates our framework. As the first step, we obtain state \\( v_{i} \\), step \\( l_{j} \\), and task \\( c \\) representations from visual observation \\( o_{i} \\), action step \\( a_{j} \\), and task title \\( tsk \\) (e.g. \"make pancakes\") respectively. This is done by inputting \\( o_{i} \\), \\( a_{j} \\), and \\( tsk \\) to a pretrained vision and language encoder respectively ([14]), similar to P3IV [24]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Task Classifier. Recent studies [22, 23] have demonstrated the effectiveness of integrating task information into training to enhance model performance in procedure planning. In alignment with previous work [23], we adopt a similar strategy by first training an MLP task classifier. This classifier takes initial and final visual observations as inputs and predicts the task class \\( t\\hat{s}k \\). The training is conducted using a cross-entropy loss function with the ground-truth task class \\( tsk \\). Subsequently, we incorporate the task representation \\( \\hat{c} \\) corresponding to the predicted task class \\( t\\hat{s}k \\) as an additional input to both the Base-Planner and the Retrieval-Planner." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Base Planner. To facilitate the adaptive procedure planning setting, we utilize an auto-regressive transformer [17] as the base planner \\( F_{\\mathrm{base}} \\). The base planner takes the representations of the initial state \\( v_{s} \\), the goal state \\( v_{g} \\), the predicted" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "A. Zare et al." + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.147, + 0.699, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.301, + 0.788, + 0.345 + ], + "angle": 0, + "content": "Fig. 3: The Retrieval-Augmented Planner module uses the predicted context vector \\(\\varsigma_{t}\\) to retrieve the \\(K\\) closest keys and their values from memory, estimating a probability distribution \\(P_{\\mathrm{retr},\\mathrm{t}}\\) over these entries for next action probability estimation \\(P_{\\mathrm{all},\\mathrm{t}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.372, + 0.788, + 0.448 + ], + "angle": 0, + "content": "task \\(\\hat{c}\\), and the current action \\(a_{t}\\), as input, to predict the next action \\(a_{t+1}\\), i.e., \\(P_{\\mathrm{base}}(y|x) = P(a_{t+1}|a_{t}, v_{s}, v_{g}, \\hat{c})\\). The sequence starts with \\(a_{0} = \\mathrm{START}\\) and completes upon predicting \\(a_{T+1} = \\mathrm{END}\\). Its implementation includes a learnable positional embedding \\(p_{t}\\) for sequence generation. Therefore, the input sequence to predict \\(t\\)-th action step \\(a_{t}\\) is formulated as in Eq. (1)," + }, + { + "type": "equation", + "bbox": [ + 0.319, + 0.458, + 0.786, + 0.475 + ], + "angle": 0, + "content": "\\[\n\\left[ v _ {s}, v _ {g}, \\hat {c}, l _ {\\text {S T A R T}}, l _ {1} + p _ {1}, \\dots , l _ {t - 2} + p _ {t - 2}, l _ {t - 1} + p _ {t - 1} \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.485, + 0.585, + 0.501 + ], + "angle": 0, + "content": "and the output sequence for supervision is Eq. (2)," + }, + { + "type": "equation", + "bbox": [ + 0.422, + 0.511, + 0.786, + 0.528 + ], + "angle": 0, + "content": "\\[\nl = \\left[ l _ {1}, l _ {2}, \\dots , l _ {T}, l _ {\\text {E N D}} \\right] \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.537, + 0.788, + 0.66 + ], + "angle": 0, + "content": "where \\( l_{t} \\) represents the language representation for step \\( a_{t} \\), \\( l_{\\mathrm{START}} \\) and \\( l_{\\mathrm{END}} \\) are the embeddings for START and END tokens. The predicted action step \\( \\hat{a}_{t} \\) is determined by mapping to a probability distribution over the action vocabulary \\( V_{a} \\) (which includes the END token), which is achieved by comparing the language similarity between predicted \\( \\hat{l}_{t} \\) and those within \\( V_{a} \\), enabling us to assign probabilities to each action within a plan sequence. During the training stage, we use the ground-truth action representation \\( l_{t} \\) as input, while we use the predicted step representation \\( \\hat{l}_{t} \\) during the test stage." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.674, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Retrieval Planner. To improve long-range sequence modeling, our model integrates a direct retrieval mechanism based on a training datastore, utilizing its state-action temporal dependency information, explicitly. This enhancement is achieved by augmenting the base planner in a manner similar to [7,10], by adding a retrieval-augmented memory module, which learns the state-action pairs in the training data as memory, by modeling states, as a context for the actions. Fig. 3 illustrated the workings of this component." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The memory module consists of key-value pairs \\(\\{s_n:d_n\\}_{n = 1}^D\\), where \\(D\\) is the memory size. The keys \\((s_n)\\), are discrete learnable context vectors and the values \\((d_{n})\\) are their subsequent action step labels, conditioned on the context. We estimate a context vector \\(\\varsigma_t\\), given representations of the initial visual state" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.696, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RAP" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "\\(v_{s}^{\\prime}\\), the target state \\(v_{g}^{\\prime}\\), the current action \\(l_{t}\\), and a learnable positional embedding \\(q_{t}\\) as in Eq. (3)" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.179, + 0.785, + 0.193 + ], + "angle": 0, + "content": "\\[\n\\varsigma_ {t} = f _ {\\text {P r o j}} \\left(\\operatorname {C o n c a t} \\left(v _ {s}, v _ {g}, \\hat {c}, l _ {t} + q _ {t}\\right)\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.199, + 0.785, + 0.41 + ], + "angle": 0, + "content": "With \\( f_{\\mathrm{Proj}} \\) as a projection layer and \\( \\mathrm{Concat}(\\cdot) \\) concatenating inputs along the embedding dimension. In training, we learn the keys \\( (s_n) \\) by comparing them to inputs' context vectors \\( \\varsigma_t \\). We construct the memory values \\( (d_n) \\) by uniformly sampling actions from the training dataset (including the END token), making sure of a balanced representation of actions. At inference, the model retrieves the \\( K \\) nearest (i.e., most similar) keys \\( \\{s_k\\}_{k=1}^K \\) from the memory to a given current context \\( \\varsigma_t \\), returning their corresponding values \\( \\{d_k\\}_{k=1}^K \\). The nearest neighbors \\( (\\{s_k\\}_{k=1}^K) \\) get selected by calculating the cosine distance between \\( \\varsigma_t \\) and the keys of the datastore \\( \\{s_n\\}_{n=1}^D \\). Upon identifying the nearest neighbors, a probability distribution across the retrieved values \\( (\\{d_k\\}_{k=1}^K) \\) can be calculated by applying a softmax function to the similarity scores. The retrieval process can be formulated as \\( P_{\\mathrm{retr}}(y|x) = P(a_{t+1}|\\varsigma_t, a_{t+1} \\in \\{d_k\\}_{k=1}^K) \\). The final prediction ensembles the predictions from base planner and retrieval planner. Following [7, 10], we obtain the final probability distribution for the next action step as:" + }, + { + "type": "equation", + "bbox": [ + 0.353, + 0.422, + 0.785, + 0.437 + ], + "angle": 0, + "content": "\\[\nP _ {\\text {a l l}} (y | x) = \\lambda P _ {\\text {r e t r}} (y | x) + (1 - \\lambda) P _ {\\text {b a s e}} (y | x) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.448, + 0.48, + 0.462 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a tuned hyper-parameter." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.484, + 0.334, + 0.499 + ], + "angle": 0, + "content": "3.3 Training" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.508, + 0.784, + 0.568 + ], + "angle": 0, + "content": "RAP is trained in two stages: First, we train the base planner, then we augment and train the model including the retrieval planner. This approach ensures effective learning of the retrieval planner's keys, building upon the base planner's good initialization obtained from the initial training stage." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.569, + 0.785, + 0.659 + ], + "angle": 0, + "content": "Stage 1: Training the Base Planner. During the training phase, we first supervise the base model's output, the textual representation of each action step in the procedure sequence, through language supervision combined with contrastive learning [8]: For each feature \\(\\hat{l}_t\\) generated by the transformer's head, we take its ground truth language embedding \\(l_t\\) as positive sample and all other embeddings in the action vocabulary \\(V_a\\) as negative examples. The loss is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.671, + 0.785, + 0.71 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {b a s e}} = - \\sum_ {t = 1} ^ {T + 1} \\log \\frac {\\exp \\hat {l} _ {t} \\cdot l _ {t}}{\\sum_ {j} \\exp \\hat {l} _ {t} \\cdot l _ {j}} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.72, + 0.539, + 0.735 + ], + "angle": 0, + "content": "Where \\((T + 1)\\)-th step denotes the END token." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.75, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Stage 2: Training the Base and Retrieval Planners. In the second stage, the model is augmented with the nearest neighbor retrieval-based memory. The retrieval planner advances learning by juxtaposing the one-hot ground-truth labels \\( a_{t} \\in p \\), with the model's interpolated probability estimates \\( P_{all,t} \\), for each action at each time-step \\( t \\). It refines the training by assessing how likely each action is, based on the model's predictions, instead of solely relying on matching" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "A. Zare et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.784, + 0.163 + ], + "angle": 0, + "content": "the correct action. We achieve this, using a cross-entropy minimization in Eq. (6)," + }, + { + "type": "equation", + "bbox": [ + 0.401, + 0.174, + 0.786, + 0.214 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l l}} = - \\sum_ {t = 1} ^ {T + 1} a _ {t} \\log P _ {\\text {a l l , t}} (y | x) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.22, + 0.538, + 0.236 + ], + "angle": 0, + "content": "where \\(P_{\\mathrm{all,t}}(y|x)\\) is defined based on Eq. (4)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.257, + 0.727, + 0.273 + ], + "angle": 0, + "content": "3.4 Addition of Unannotated Videos with Weak Supervision" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.282, + 0.787, + 0.372 + ], + "angle": 0, + "content": "A significant challenge in procedure planning for instructional videos is the scarcity of annotated data. This is mainly because annotating instructional videos with step-level (i.e., timestamps) or sequence-level labels (i.e., action sequences) is resource-intensive. To address this issue, we propose a weakly-supervised training approach aimed at expanding the annotated training data to include additional unannotated task-related videos in the training process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.789, + 0.553 + ], + "angle": 0, + "content": "Specifically, we align instructional videos with their corresponding topics, such as \"making an omelet\" or \"making pancakes\". This alignment links each video subset to a subset of instructional action plans sharing a similar topic. Crucially, this approach does not require detailed annotations that map specific action steps to individual video segments or frames. Instead, it relies on the broader thematic association between the video content and the instructional plans, allowing for a more general form of supervision during training. However, the acquired pairs of instructional videos and action plans may exhibit certain discrepancies. First, the presence of action steps in a plan that do not appear in the related video. Second, a mismatch in the sequence of action steps between the plan and their occurrence in the instructional video. Third, the inclusion of additional action steps in the video that are absent from the plan." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.789, + 0.765 + ], + "angle": 0, + "content": "To tackle these challenges, we applied a video grounding algorithm, taking inspiration from Dvornik et al. [5] to generate pseudo-annotations of descriptions in instructional videos. In short, our grounding algorithm processes an instructional video alongside an ordered sequence of action steps to produce a sequence of grounding results (\\( t_{start}, t_{end}, a \\)), where \\( t_{start} \\) and \\( t_{end} \\) denote the beginning and end times of action step \\( a \\). The grounding model offers flexible sequence alignment and the ability to exclude irrelevant actions or frames, effectively addressing observed discrepancies. Further details on the grounding model are provided in the appendix. We utilize these pseudo-annotations, extracted for each relevant video, which serve as structurally and format-wise equivalent to manually annotated videos. These pseudo-annotations are integrated into Stage 1 training alongside original annotations, where the model is tasked with predicting the complete sequence of actions based on the starting frame of the first action and the ending frame of the last action outlined in the plan." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.789, + 0.84 + ], + "angle": 0, + "content": "This weak supervision of unannotated data, is an extension of Stage 1 training by using both original annotations of training videos and pseudo-annotations of task-relevant videos as the new training set. This training approach is a step towards a cost-effective solution for generalizing and scaling procedure planning to abundant unannotated videos." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.696, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RAP" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.146, + 0.375, + 0.163 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.784, + 0.207 + ], + "angle": 0, + "content": "We experiment on two benchmark datasets and employ four metrics to validate the efficacy of our proposed model." + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.234, + 0.727, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.37, + 0.785, + 0.412 + ], + "angle": 0, + "content": "Fig. 4: Illustration of the advantage of edit score, in evaluation of long varied-length sequences. As it can be seen, despite the prediction being plausible and acceptable, accuracy and success rate, falsely flag this scenario as a miss." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.465, + 0.334, + 0.478 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.788, + 0.745 + ], + "angle": 0, + "content": "Our methodology is evaluated using two instructional video datasets: CrossTask [26] and COIN [20]. CrossTask contains 2750 videos spanning 83 different topics, such as \"How to make french toast?\", with 7.6 actions per video on average. These topics are categorized into 18 primary and 65 related events. The primary topics are detailed with precise timestamps for each action step, providing a clear sequence of instructional activities. In contrast, the related instructional topics, gathered through automatic collection methods, lack annotations for action-step timestamps, sequence order, and the specific subset of action steps each video covers. Each topic within this related set is associated with a designated instructional plan as well as a set of unannotated instructional videos of the same topic. These videos may exhibit variations from the outlined plan, as detailed in Sec. 3.3, due to the previously mentioned discrepancies. We use CrossTask's related-set data, during weakly-supervised training with language supervision. COIN, includes 11827 videos, covering 778 different actions, with an average action-horizon of 3.6 actions per video. We follow [4,24] by creating train/test splits with a \\(70\\% / 30\\%\\) ratio as well as implementing a moving window approach to organize the datasets into plans with various action-horizons." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.768, + 0.327, + 0.782 + ], + "angle": 0, + "content": "4.2 Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.793, + 0.785, + 0.844 + ], + "angle": 0, + "content": "Our methodology's performance is evaluated using three standard metrics [2, 4, 19, 24]. First, mean Intersection over Union (\\(mIoU\\)) assesses overlap between predicted and ground truth action sequences, calculated by \\(\\frac{|a_t \\cap \\hat{a}_t|}{|a_t \\cup \\hat{a}_t|}\\). This metric" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.127 + ], + "angle": 0, + "content": "A. Zare et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "evaluates whether the model identifies correct steps but does not consider action order or repetitions. Second, mean Accuracy (\\( mAcc \\)) measures action alignment at each step, balancing the impact of repeated actions. Lastly, the Success Rate (\\( SR \\)) evaluates a plan as successful if it exactly matches the ground truth." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.21, + 0.785, + 0.33 + ], + "angle": 0, + "content": "However, Accuracy and Success Rate have limitations, particularly with long or variable-length sequences, often leading to false negatives due to their sensitivity to action order. For example, in a pancake-making plan, if the predicted sequence adds a plausible step like \"Add sugar\" at the beginning, accuracy may incorrectly deem it inaccurate. This issue is more pronounced with sequences that vary in length from the ground truth, indicating the need for a more nuanced evaluation metric in procedure planning that accommodates plausible variations in actions." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.334, + 0.784, + 0.44 + ], + "angle": 0, + "content": "To address the limitations of the current metrics, we introduce a new metric, mean edit-score \\((mES)\\), based on the minimum edit distance (e.g. Levenshtein edit distance), measuring the similarity between two sequences. This function, \\(L(s_{1},s_{2})|c_{sub},c_{ins},c_{del}\\), calculates the minimum number of substitutions \\((c_{sub})\\), insertions \\((c_{ins})\\), and deletions \\((c_{del})\\) needed to transform sequence \\(s_1\\) into \\(s_2\\). The mean edit-score, is derived from the expected value of sample edit-scores. The edit-score is thus defined based on the optimal edit distance as:" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.462, + 0.785, + 0.495 + ], + "angle": 0, + "content": "\\[\nE S _ {s _ {1}, s _ {2}} = 1 0 0 \\cdot \\frac {\\operatorname* {m a x} \\left(\\left| s _ {1} \\right| , \\left| s _ {2} \\right|\\right) - L \\left(s _ {1} , s _ {2}\\right) | _ {1 , 1 , 1}}{\\operatorname* {m a x} \\left(\\left| s _ {1} \\right| , \\left| s _ {2} \\right|\\right)}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.508, + 0.788, + 0.645 + ], + "angle": 0, + "content": "Where the operator \\(|\\cdot|\\) measures the length of a sequence. The mean edit-score offers a more flexible and forgiving assessment compared to strict metrics like success rate and mean accuracy, while still maintaining sensitivity to the order of action steps, hence complimenting \\(mIoU\\) (i.e., if the ground-truth is add curry leaves, stir mixture, add chili powder and the prediction is add chili powder, stir mixture, add curry leaves, we face a high \\(IoU = 100\\%\\) and a low \\(ES = 33.33\\%\\)). This metric provides a more comprehensive evaluation, especially in scenarios where the predicted sequence is plausible yet varies in actions or length from the ground truth (Fig. 4)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.68, + 0.457, + 0.696 + ], + "angle": 0, + "content": "4.3 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Consistent with earlier approaches, we encode the start and goal visual observations using the S3D network [14] pre-trained on the HowTo100M [15] dataset. For optimization, we use the AdamW optimizer [13], setting the weight decay to 0.2. The learning rate is initially set at \\( lr = 0.001 \\), and we apply a reduce-on-plateau scheduler to adjust this rate based on performance metrics. We experiment with an autoregressive transformer with 7 layers and 8 heads. We choose weight \\( \\lambda = 0.1 \\) and volume \\( D = 1000 \\), for the memory module. Further implementation details can be found in the supplementary section." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.695, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RAP" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Assessment of procedure planning on CrossTask for action-horizons \\( T = 3 \\) and \\( T = 4 \\). Vision supervision is denoted by V, language supervision by L, and weakly-supervised training by +wL." + }, + { + "type": "table", + "bbox": [ + 0.245, + 0.199, + 0.761, + 0.394 + ], + "angle": 0, + "content": "
ModelsSupervisionT=3T=4
SR↑mAcc↑mIoU↑SR↑mAcc↑mIoU↑
Random-<0.010.941.66<0.010.831.66
Retrieval-BasedV8.0523.3032.063.9522.2236.97
WLTDO [6]V1.8721.6431.700.7717.9226.43
UAAA [1]V2.1520.2130.870.9819.8627.09
UPN [18]V2.8924.3931.561.1921.5927.85
DDN [4]V12.1831.2947.485.9727.1048.46
Ext-GAIL w/o Aug. [2]V18.0143.8657.16---
Ext-GAIL [2]V21.2749.4661.7016.4143.0560.93
P3IV w/o Adv. [24]L22.1245.5767.40---
P3IV [24]L23.3449.9673.8913.4044.1670.01
E3P (baseline) [22]L22.5646.17-12.9743.97-
E3P [22]L26.4053.0274.0516.4948.0070.16
Skip-Plan [12]L28.8561.1874.9815.5655.6470.30
RAP (ours w/o unannotated data)L29.2362.3575.6316.9655.7771.49
RAP (ours with unannotated data)L+wL30.5162.6875.9217.1055.9171.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.406, + 0.788, + 0.434 + ], + "angle": 0, + "content": "Table 2: Comparison of RAP with PDPP, when we align our state definition with PDPP's for CrossTask and \\( T = 3 \\)." + }, + { + "type": "table", + "bbox": [ + 0.288, + 0.446, + 0.717, + 0.487 + ], + "angle": 0, + "content": "
ModelsSupervisionSR ↑mAcc ↑mIoU ↑
PDPP [23]♣L37.2064.6766.57
RAP (ours w/o unannotated data)♣L38.5165.1877.44
" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.514, + 0.631, + 0.53 + ], + "angle": 0, + "content": "4.4 Comparison with State-of-The-Art Baselines" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.644 + ], + "angle": 0, + "content": "Our study conducts a comprehensive comparison of our new framework, RAP, against established models in procedure planning, using CrossTask and COIN datasets. We refer to our framework without the retrieval-based memory augmentation, i.e., base planner, as BP (as opposed to RAP). We categorize previous approaches based on their supervision: language (L) or vision (V). Our proposed model's performance compared to the previous works, is detailed in Tabs. 1 and 2 for CrossTask, and Tab. 3 for COIN." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.788, + 0.795 + ], + "angle": 0, + "content": "The inherent uncertainty in variable action-horizon procedure planning models complicates direct and fair comparisons with traditional fixed-length models, due to the added complexity of predicting the conclusion of a sequence as well as the broader range of acceptable sequences of variable length other than the ground-truth. To ensure a fair evaluation, we compare our variable action-horizon architecture, trained and evaluated exclusively on fixed-horizon data (of length \\( T \\)), to the fixed-horizon models (of length \\( T \\)) from prior studies in procedure planning [1,2,4,6,12,18,22-24]. Tab. 1 further shows the performance boost achieved by the addition of CrossTask's unannotated data (i.e., CrossTask's related video set) through weakly-supervised training (supervision: \\( \\mathrm{L} + \\mathrm{wL} \\))." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Our findings demonstrate that RAP exhibits superior performance over existing vision-supervised models. Notably, RAP, without the addition of unannotated data, surpasses Ext-GAIL [2] with a \\(7.97\\%\\) increase in Success Rate (SR)," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "A. Zare et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.784, + 0.172 + ], + "angle": 0, + "content": "Table 3: Comparison of the procedure planing models' performance on the COIN dataset." + }, + { + "type": "table", + "bbox": [ + 0.288, + 0.185, + 0.719, + 0.306 + ], + "angle": 0, + "content": "
ModelsT=3T=4
SR↑mAcc↑mIoU↑SR↑mAcc↑mIoU↑
Random (V)<0.01<0.012.47<0.01<0.012.32
Retrieval-Based (V)4.3817.4032.062.7114.2936.97
DDN (V)13.9020.1964.7811.1317.7168.06
P3IV (L)15.4021.6776.3111.3218.8570.53
E3P (L)19.5731.4284.9513.5926.7284.72
PDPP (L)21.3345.6251.8214.4144.1051.39
Skip-plan (L)23.6547.1278.4416.0443.1977.07
RAP (w/o unannotated) (L)24.4747.5985.2416.7642.1184.64
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.318, + 0.784, + 0.348 + ], + "angle": 0, + "content": "Table 4: Evaluation of RAP versus a retrieval-based (RB) model in variable-length sequence prediction on CrossTask and COIN." + }, + { + "type": "table", + "bbox": [ + 0.243, + 0.359, + 0.758, + 0.432 + ], + "angle": 0, + "content": "
ModelsCrossTaskCOIN
SR↑mAcc↑mIoU↑mES↑SR↑mAcc↑mIoU↑mES↑
Random0.020.561.521.210.020.090.180.17
RB8.0121.3238.8033.437.1112.9420.1717.69
RAP (w/o unannotated data)20.2849.4169.9562.7919.1433.2675.8257.48
RAP (with unannotated data)21.7950.5871.4463.86----
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.784, + 0.492 + ], + "angle": 0, + "content": "12.89% in Mean Accuracy (mAcc), and 13.93% in Mean Intersection over Union (mIoU) for sequence lengths of \\( T = 3 \\) in CrossTask." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.787, + 0.702 + ], + "angle": 0, + "content": "When compared to language-supervised models like P3IV, E3P, and SkipPlan, RAP demonstrates superior performance on CrossTask. For \\( \\mathrm{T} = 3 \\), RAP exhibits improvements of \\( 5.99\\% \\), \\( 2.83\\% \\), and \\( 0.38\\% \\) respectively, and for \\( \\mathrm{T} = 4 \\), it shows enhancements of \\( 3.56\\% \\), \\( 0.47\\% \\), and \\( 1.4\\% \\) in SR. Tab. 2 compares our model with PDPP [23], which employs a different procedure state definition. For fairness, we evaluate our model under the PDPP setting, marked with \\( \\clubsuit \\), showcasing its superior performance. PDPP uses a 3-second window following the onset and preceding the offset of action steps to capture visual states, giving it an advantage in accessing action-related information compared to previous models. Traditionally, a 1-second window centered around the start and end of actions is used. By aligning our state definition with PDPP's, our RAP model outperforms PDPP on all metrics, as shown in the table. We observe a similar superiority in performance, when evaluating the model on the COIN dataset as shown in Tab. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.787, + 0.809 + ], + "angle": 0, + "content": "We assessed the efficacy of RAP in handling tasks with variable lengths by conducting experiments against a retrieval-based model, which operates under a similar variable action-horizon setting. The results of this comparison are presented in Tab. 4. As demonstrated in the table, RAP proves a remarkable superiority over the retrieval-based model across all evaluated metrics. This is particularly evident in the mES metric, where RAP (w/o unannotated data) shows a significant improvement by \\(29.36\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "On CrossTask, RAP is further trained using the weak language supervision framework outlined in Sec. 3.3, utilizing the unannotated data from the dataset's" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.696, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RAP" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 5: A study on the effect of retrieval augmentation (BP vs RAP), for variable and fixed action-horizon settings." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.185, + 0.787, + 0.248 + ], + "angle": 0, + "content": "
DatasetModelsT=3T=4Variable Length T
SR↑mAcc↑mIoU↑mES↑SR↑mAcc↑mIoU↑mES↑SR↑mAcc↑mIoU↑mES↑
CrossTask (w/o unannotated)BP25.4258.7073.9165.5915.2352.1670.5361.6318.5447.8569.1662.64
RAP29.2362.3575.6367.3316.9655.7771.4964.5120.2849.4169.9562.79
COINBP23.3846.2180.4962.2814.2941.6778.6257.2617.8631.3472.4456.24
RAP24.4747.5985.2463.1516.7642.1184.6460.4719.1433.2675.8257.48
" + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.264, + 0.752, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.399, + 0.785, + 0.427 + ], + "angle": 0, + "content": "Fig. 5: Sample plausible predicted plan sequences generated by RAP vs the ground-truth (GT)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.458, + 0.785, + 0.502 + ], + "angle": 0, + "content": "related-tasks. As shown in Tab. 1 and Tab. 4 when training RAP with the addition of the unannotated data, we see significant improvement in performance across all metrics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.785, + 0.594 + ], + "angle": 0, + "content": "Fig. 5 displays two qualitative results of RAP's performance in predicting plans of varying lengths compared to the ground truths. In the top row, it accurately identifies the addition of fish, other ingredients, and the plausible action \"add onions\". In the bottom row, it recognizes the completed bread, its removal, and suggests the reasonable action \"melt butter\", before the new bread is put in the pan, showcasing the model's planning abilities." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.618, + 0.389, + 0.634 + ], + "angle": 0, + "content": "4.5 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.787, + 0.78 + ], + "angle": 0, + "content": "Impact of retrieval augmentation. The introduction of a retrieval-based memory component significantly enhances our framework's performance. Tab. 5 demonstrates improvements for both fixed and variable action-horizon settings. Notably, enhancements include increases of \\(3.81\\%\\), \\(1.73\\%\\), and \\(1.74\\%\\) in Success Rate (SR) for \\(\\mathrm{T} = 3\\) and \\(\\mathrm{T} = 4\\) in the fixed setting, and in the variable setting, respectively, on CrossTask. We observe a similar boost in performance for the COIN dataset, as evidenced in the Tab. 5. Tab. 6 details the enhancements in SR, categorized by the ground-truth (GT) sequences' action-horizon in the variable-length sequence prediction setting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Impact of weak language supervision. The influence of weak language supervision on our model's performance is evaluated through an ablation study on CrossTask, where we incrementally increased the unannotated data used in weakly-supervised training. Tab. 7 displays the varying percentages of unan-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "A. Zare et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.175 + ], + "angle": 0, + "content": "Table 6: Impact of retrieval augmentation (BP vs RAP) in \\( SR \\), on CrossTask, categorized by ground-truth (GT) action-horizon in the variable action-horizon setting." + }, + { + "type": "table", + "bbox": [ + 0.316, + 0.184, + 0.69, + 0.24 + ], + "angle": 0, + "content": "
Model (w/o unannotated data)GT action-horizon
123456
BP49.5827.8416.218.375.893.34
RAP52.6029.2018.0210.526.165.18
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.252, + 0.787, + 0.281 + ], + "angle": 0, + "content": "Table 7: Weak language supervision impact: Table shows RAP's performance gains on CrossTask with increasing unannotated data use." + }, + { + "type": "table", + "bbox": [ + 0.33, + 0.292, + 0.675, + 0.392 + ], + "angle": 0, + "content": "
Used unannotated data portionSR↑mAcc↑mIoU↑mES↑
No unannotated data20.2849.4169.9562.79
25%19.2047.8869.0561.96
50%20.5749.7269.7563.12
75%21.3150.2270.8663.40
100%21.7950.5871.4463.86
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.788, + 0.511 + ], + "angle": 0, + "content": "notated data utilized. The results show consistent performance gains for RAP under weak language supervision, particularly as more data is introduced (except at the \\(25\\%\\) data level). This suggests potential for further enhancements with increased unannotated data. Notably, using \\(100\\%\\) of unannotated data yields improvements in SR, mAcc, mIoU, and mES by \\(1.51\\%\\), \\(1.17\\%\\), \\(1.49\\%\\), and \\(1.07\\%\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.534, + 0.36, + 0.55 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.789, + 0.793 + ], + "angle": 0, + "content": "In this study, we addressed the challenge of procedure planning in instructional videos by developing the Retrieval-Augmented Planner (RAP) model. This model overcomes the limitations of previous approaches that were restricted to fixed action-step predictions. RAP excels in both fixed and varied action horizon scenarios, outperforming existing models. A key innovation is our weak language supervision method, which uses thematic links between unannotated videos and plans for pseudo-annotation generation, enhancing model training. Our model also incorporates a learnable retrieval-based memory component for improved long-range sequence predictions and efficient use of temporal data. Additionally, we introduced \"edit-score\", a new metric for assessing variable-length procedures. Looking forward, we see potential in applying RAP to more diverse scenarios and in generating probabilistic, varied-length plans. We also envisage that expanding our weak supervision approach to include an even wider range of unannotated data could markedly elevate the performance of procedure planning models, paving the way for more advanced applications in this field." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.696, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RAP" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.214 + ], + "angle": 0, + "content": "1. Abu Farha, Y., Gall, J.: Uncertainty-aware anticipation of activities. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. pp. 0-0 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.215, + 0.785, + 0.255 + ], + "angle": 0, + "content": "2. Bi, J., Luo, J., Xu, C.: Procedure planning in instructional videos via contextual modeling and model-based policy learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15611-15620 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.256, + 0.785, + 0.31 + ], + "angle": 0, + "content": "3. Borgeaud, S., Mensch, A., Hoffmann, J., Cai, T., Rutherford, E., Millican, K., Van Den Driessche, G.B., Lespiau, J.B., Damoc, B., Clark, A., et al.: Improving language models by retrieving from trillions of tokens. In: International conference on machine learning. pp. 2206-2240. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.785, + 0.351 + ], + "angle": 0, + "content": "4. Chang, C.Y., Huang, D.A., Xu, D., Adeli, E., Fei-Fei, L., Niebles, J.C.: Procedure planning in instructional videos. In: European Conference on Computer Vision. pp. 334-350. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.351, + 0.785, + 0.392 + ], + "angle": 0, + "content": "5. Dvornik, M., Hadji, I., Derpanis, K.G., Garg, A., Jepson, A.: Drop-dtw: Aligning common signal between sequences while dropping outliers. Advances in Neural Information Processing Systems 34, 13782-13793 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.392, + 0.785, + 0.433 + ], + "angle": 0, + "content": "6. Ehsani, K., Bagherinezhad, H., Redmon, J., Mottaghi, R., Farhadi, A.: Who let the dogs out? modeling dog behavior from visual data. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4051-4060 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.433, + 0.785, + 0.472 + ], + "angle": 0, + "content": "7. Grave, E., Cisse, M.M., Joulin, A.: Unbounded cache model for online language modeling with open vocabulary. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.473, + 0.785, + 0.528 + ], + "angle": 0, + "content": "8. Gutmann, M., Hyvarinen, A.: Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In: Proceedings of the thirteenth international conference on artificial intelligence and statistics. pp. 297-304. JMLR Workshop and Conference Proceedings (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.528, + 0.785, + 0.568 + ], + "angle": 0, + "content": "9. Hadji, I., Derpanis, K.G., Jepson, A.D.: Representation learning via global temporal alignment and cycle-consistency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11068-11077 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.568, + 0.785, + 0.609 + ], + "angle": 0, + "content": "0. Khandelwal, U., Levy, O., Jurafsky, D., Zettlemoyer, L., Lewis, M.: Generalization through memorization: Nearest neighbor language models. arXiv preprint arXiv:1911.00172 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.609, + 0.785, + 0.65 + ], + "angle": 0, + "content": "1. Lei, J., Wang, L., Shen, Y., Yu, D., Berg, T.L., Bansal, M.: Mart: Memory-augmented recurrent transformer for coherent video paragraph captioning. arXiv preprint arXiv:2005.05402 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.65, + 0.785, + 0.704 + ], + "angle": 0, + "content": "2. Li, Z., Geng, W., Li, M., Chen, L., Tang, Y., Lu, J., Zhou, J.: Skip-plan: Procedure planning in instructional videos via condensed action space learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10297-10306 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.705, + 0.785, + 0.731 + ], + "angle": 0, + "content": "3. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.785, + 0.786 + ], + "angle": 0, + "content": "4. Miech, A., Alayrac, J.B., Smaira, L., Laptev, I., Sivic, J., Zisserman, A.: End-to-end learning of visual representations from uncurated instructional videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9879-9889 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "5. Miech, A., Zhukov, D., Alayrac, J.B., Tapaswi, M., Laptev, I., Sivic, J.: Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2630-2640 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.127 + ], + "angle": 0, + "content": "A. Zare et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "16. Niu, Y., Guo, W., Chen, L., Lin, X., Chang, S.F.: Schema: State changes matter for procedure planning in instructional videos. arXiv preprint arXiv:2403.01599 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.786, + 0.218 + ], + "angle": 0, + "content": "17. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.259 + ], + "angle": 0, + "content": "18. Srinivas, A., Jabri, A., Abbeel, P., Levine, S., Finn, C.: Universal planning networks: Learning generalizable representations for visuomotor control. In: International Conference on Machine Learning. pp. 4732-4741. PMLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.787, + 0.3 + ], + "angle": 0, + "content": "19. Sun, J., Huang, D.A., Lu, B., Liu, Y.H., Zhou, B., Garg, A.: Plate: Visually-grounded planning with transformers in procedural tasks. IEEE Robotics and Automation Letters 7(2), 4924-4930 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.357 + ], + "angle": 0, + "content": "20. Tang, Y., Ding, D., Rao, Y., Zheng, Y., Zhang, D., Zhao, L., Lu, J., Zhou, J.: Coin: A large-scale dataset for comprehensive instructional video analysis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1207-1216 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.398 + ], + "angle": 0, + "content": "21. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.439 + ], + "angle": 0, + "content": "22. Wang, A.L., Lin, K.Y., Du, J.R., Meng, J., Zheng, W.S.: Event-guided procedure planning from instructional videos with text supervision. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13565-13575 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "23. Wang, H., Wu, Y., Guo, S., Wang, L.: Pdpp: Projected diffusion for procedure planning in instructional videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14836-14845 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.536 + ], + "angle": 0, + "content": "24. Zhao, H., Hadji, I., Dvornik, N., Derpanis, K.G., Wildes, R.P., Jepson, A.D.: P3iv: Probabilistic procedure planning from instructional videos with weak supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2938-2948 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.787, + 0.563 + ], + "angle": 0, + "content": "25. Zhong, Z., Lei, T., Chen, D.: Training language models with memory augmentation (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.787, + 0.619 + ], + "angle": 0, + "content": "26. Zhukov, D., Alayrac, J.B., Cinbis, R.G., Fouhey, D., Laptev, I., Sivic, J.: Cross-task weakly supervised learning from instructional videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3537-3545 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.619 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/RAP_ Retrieval-Augmented Planner for Adaptive Procedure Planning in Instructional Videos/8baf20b2-0efa-428f-a4d5-15cb7958ae91_origin.pdf b/2024/RAP_ Retrieval-Augmented Planner for Adaptive Procedure Planning in Instructional Videos/8baf20b2-0efa-428f-a4d5-15cb7958ae91_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9f6ca3605fc8a82184c511f0b25cb53a31461520 --- /dev/null +++ b/2024/RAP_ Retrieval-Augmented Planner for Adaptive Procedure Planning in Instructional Videos/8baf20b2-0efa-428f-a4d5-15cb7958ae91_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:424e2952b21820122d0bcb72e847bb05785d04137a1375f4798e1e75852600e4 +size 7757398 diff --git a/2024/RAP_ Retrieval-Augmented Planner for Adaptive Procedure Planning in Instructional Videos/full.md b/2024/RAP_ Retrieval-Augmented Planner for Adaptive Procedure Planning in Instructional Videos/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a102549279096d45a58fc9aa7fc331e0be7a9f86 --- /dev/null +++ b/2024/RAP_ Retrieval-Augmented Planner for Adaptive Procedure Planning in Instructional Videos/full.md @@ -0,0 +1,260 @@ +# RAP: Retrieval-Augmented Planner for Adaptive Procedure Planning in Instructional Videos + +Ali Zare, Yulei Niu*, Hammad Ayyubi, and Shih-Fu Chang + +Columbia University, New York, NY 10027, USA {az2584,yn2338,ha2578,sc250}@columbia.edu, + +Abstract. Procedure Planning in instructional videos entails generating a sequence of action steps based on visual observations of the initial and target states. Despite the rapid progress in this task, there remain several critical challenges to be solved: (1) Adaptive procedures: Prior works hold an unrealistic assumption that the number of action steps is known and fixed, leading to non-generalizable models in real-world scenarios where the sequence length varies. (2) Temporal relation: Understanding the step temporal relation knowledge is essential in producing reasonable and executable plans. (3) Annotation cost: Annotating instructional videos with step-level labels (i.e., timestamp) or sequence-level labels (i.e., action category) is demanding and labor-intensive, limiting its generalizability to large-scale datasets. In this work, we propose a new and practical setting, called adaptive procedure planning in instructional videos, where the procedure length is not fixed or pre-determined. To address these challenges, we introduce Retrieval-Augmented Planner (RAP) model. Specifically, for adaptive procedures, RAP adaptively determines the conclusion of actions using an auto-regressive model architecture. For temporal relation, RAP establishes an external memory module to explicitly retrieve the most relevant state-action pairs from the training videos and revises the generated procedures. To tackle high annotation cost, RAP utilizes a weakly-supervised learning manner to expand the training dataset to other task-relevant, unannotated videos by generating pseudo labels for action steps. Experiments on CrossTask and COIN benchmarks show the superiority of RAP over traditional fixed-length models, establishing it as a strong baseline solution for adaptive procedure planning. + +Keywords: Procedure Planning, Instructional Video, Retrieval Augmentation, Weak Supervision + +# 1 Introduction + +Procedure planning in instructional videos aims to generate an executive series of actions for completing tasks, transitioning from initial visual state to final visual state [2, 4, 12, 16, 19, 22-24]. While humans excel at such tasks, machine learning agents face challenges in achieving similar proficiency. Success in procedure + +![](images/e7068740273b45094a058d2c55ec89fba3f9992dabac1717baaf041c2aef99bc.jpg) +Fig. 1: Examples of adaptive procedures in instructional videos. Procedure length varies with differences between initial and goal visual states, even for the same task. Previous works assume a fixed, known sequence length, requiring a separate model for each action-horizon. This figure contrasts traditional procedure planning (A) with our adaptive approach (B), which trains a unified model without knowing the procedure length in advance. + +planning requires an agent's deep understanding of actions, procedure states, involved objects, and their temporal relationships. Narrowing this performance gap is key to advancing autonomous systems, equipping them to adeptly perform complex, instruction-based tasks like cooking or repair. + +Although recent works have made tremendous strides, there remain several challenges to real-world applications. First, previous works assume that the number of action steps is known and fixed. As depicted in Fig. 1, the visual goal may be achieved through a small number of steps or a long procedure, depending on the differences between the initial and goal states. For example, it takes three steps to transform a visual of adding butter to a visual of flipped french toast in example b (scenario b implies pre-prepared mixture and pre-dipped bread, but visually requires actions like "melt butter," "put bread in pan," and "flip bread"). but one needs six and five steps to finish the cooking process based on the visuals of the mixture in scenarios a and c respectively. Previous models, trained on the unrealistic assumption of fixed prediction horizon, cannot figure out the complexity of state understanding and are hard to generalize to real-world applications. Secondly, there is a high correlation between consecutive action steps. For example, "add" is often followed by "mix", and "wash" is often followed by "rinse". Existing methods model this relationship implicitly via intermediate states, leading to sub-par results. Third, prior methods focused on fully annotated data [2,4,19] - action steps and intermediate visual states - for supervision. However, this approach poses challenges in scaling up the training data due to its costliness. Following works [12, 16, 22-24] proposed to leverage text representations of action steps as language supervision, which still necessitate + +sequence-level annotations, a time-consuming and expensive process involving watching entire instructional videos. + +To tackle these challenges, we introduce the setting of adaptive procedure planning in instructional videos and propose the Retrieval-Augmented Planner (RAP) model for this problem. Overall, RAP is designed to (1) adaptively predict procedures with variable lengths according to the complexity of visual states, (2) explicitly exploit step temporal relation via knowledge retrieval from the training data, and (3) learn procedural knowledge from additional unannotated videos. Specifically, RAP consists of an auto-regressive model as the base planner for adaptive procedure prediction, and a retrieval-augmented memory module to refine the generated procedures. The architecture is supplemented by weakly-supervised learning method that leverages both annotated video-procedure pairs and unannotated videos. + +In addition, existing metrics for procedure planning, such as accuracy, fail to fairly evaluate variable length procedures (as described in Sec. 4.2). We address this by introducing a new metric edit-score. We evaluate our model on two widely used datasets for instructional videos, CrossTask [26] and COIN [20]. Our model demonstrates superior performance in the conventional setting and establishes a strong baseline in the new adaptive procedure planning setting. + +Our main contributions are: + +- We argue the unrealistic assumption of fixed-length plans in previous procedure planning works, and introduce a new setting adaptive procedure planning in instructional videos. +- We propose Retrieval-Augmented Planner that can generate adaptive procedures of variable lengths and retrieve step temporal relation knowledge from the training data. +- We introduce a weakly-supervised training approach to learn procedural knowledge from unannotated task-relevant videos. + +# 2 Related work + +# 2.1 Procedure Planning + +Procedure planning from instructional videos, involves generating effective plans for task completion based on start and final visual observations. Earlier methods utilized a two-branch architecture, alternating between action and state prediction, with models like recurrent neural networks [4], transformers [19], and adversarial networks [2]. Recent approaches, such as P3IV's [24] memory-augmented Transformer and E3P, a prompting-based model [22], adopt one-branch transformer models. However, these methods are limited to predicting procedures of fixed lengths, lacking adaptability to tasks with varying action sequence length requirements. Our work introduces RAP, a novel one-branch approach designed to predict plans with variable action-horizons, addressing this limitation in the literature. + +# 2.2 Sequence Generation with Auto-regressive Transformers + +Recent advancements in sequence generation using Transformer-based models [17,21] have established remarkable results in tasks involving sequence modeling, e.g. natural language modeling [17]. Procedure planning in its core, is a goal-conditioned sequence generation problem, thereby directly benefiting from the state-of-the-art sequence generation models. Leveraging these developments, our work incorporates an auto-regressive Transformer [17] into the RAP architecture. + +# 2.3 Retrieval-Augmented Methods + +Augmenting language models (LM) with external memory [3, 7, 10, 11, 25] has proved to significantly improve language sequence generation. P3IV [24] uses a memory component to enhance long-range sequence modeling in procedure planning. Similarly, we adopt a retrieval-based mechanism, drawing inspiration from Khandelwal et al. [10]'s augmentation of a neural language model with a k-nearest neighbors (kNN) retrieval component, in our auto-regressive model to effectively leverage temporal prior information for predicting action sequences. + +# 2.4 Video Grounding + +Procedure planning suffers from a lack of annotated data. Earlier works heavily relied on fully annotated data (action-state timestamps) [2, 4, 19]. Recent approaches [22, 24] have used language supervision during training which relies on less-costly, yet still resource-intensive language-level annotations. In this work, we introduce a weak language supervision framework that employs a video-language grounding algorithm to create pseudo-annotations from unannotated data. This approach is inspired by Dvornik et al. [5] in sequence-to-sequence alignment research, utilizing a grounding algorithm that aligns video frame and action step representations. + +# 3 Technical approach + +In this section, we introduce our proposed framework Retrieval-Augmented Planner for adaptive procedure planning in instructional videos, and expand on a weakly-supervised learning method to leverage unannotated but task-relevant videos. + +# 3.1 Setting: Adaptive Procedure Planning + +In conventional formulation of procedure planning from instructional videos, a model is required to generate a plan $p = \{a_1, a_2, \dots, a_T\}$ , given visual observations of the start $o_s$ , and the goal $o_g$ states, where $T$ is the plan's length (i.e., action-horizon). Previous methods assume a fixed, known action-horizon, which is unrealistic for real-world applicability and limits the generalizability of planning models. As shown in Fig. 1, we consider a more general and practical setting, adaptive procedure planning in instructional videos, where the procedure length is not fixed. + +![](images/a56bf2ea0c49957863420635fb13fb856d817bb74dedb0a7bbbc45528e66ffad.jpg) +Fig.2: Framework overview: Visual observations $o_s$ and $o_g$ are encoded to $v_s$ and $v_g$ . A Task-Classifier predicts a task class and its representation $\hat{c}$ using $v_s$ and $v_g$ . These, along with a learnable START token, are fed into the auto-regressive base planner $F_{base}$ for sequential action embedding prediction. For each prediction at position $t$ , the retrieval component uses the predicted context vector $s_t$ to estimate a probability distribution $P_{\mathrm{retr},t}$ over the closest $K$ key-value pairs in memory for kNN linear interpolation with $P_{\mathrm{base},t}$ to estimate the final probability distribution $P_{\mathrm{all},t}$ . + +# 3.2 Model: Retrieval-Augmented Planner + +We propose a new framework, Retrieval-Augmented Planner (RAP), to tackle the adaptive procedure planning problem. Fig. 2 illustrates our framework. As the first step, we obtain state $v_{i}$ , step $l_{j}$ , and task $c$ representations from visual observation $o_{i}$ , action step $a_{j}$ , and task title $tsk$ (e.g. "make pancakes") respectively. This is done by inputting $o_{i}$ , $a_{j}$ , and $tsk$ to a pretrained vision and language encoder respectively ([14]), similar to P3IV [24]. + +Task Classifier. Recent studies [22, 23] have demonstrated the effectiveness of integrating task information into training to enhance model performance in procedure planning. In alignment with previous work [23], we adopt a similar strategy by first training an MLP task classifier. This classifier takes initial and final visual observations as inputs and predicts the task class $t\hat{s}k$ . The training is conducted using a cross-entropy loss function with the ground-truth task class $tsk$ . Subsequently, we incorporate the task representation $\hat{c}$ corresponding to the predicted task class $t\hat{s}k$ as an additional input to both the Base-Planner and the Retrieval-Planner. + +Base Planner. To facilitate the adaptive procedure planning setting, we utilize an auto-regressive transformer [17] as the base planner $F_{\mathrm{base}}$ . The base planner takes the representations of the initial state $v_{s}$ , the goal state $v_{g}$ , the predicted + +![](images/bf4372253c13069322c35a168a1a1279c7652df37e7ffa7d94deff8326ca5a89.jpg) +Fig. 3: The Retrieval-Augmented Planner module uses the predicted context vector $\varsigma_{t}$ to retrieve the $K$ closest keys and their values from memory, estimating a probability distribution $P_{\mathrm{retr},\mathrm{t}}$ over these entries for next action probability estimation $P_{\mathrm{all},\mathrm{t}}$ . + +task $\hat{c}$ , and the current action $a_{t}$ , as input, to predict the next action $a_{t+1}$ , i.e., $P_{\mathrm{base}}(y|x) = P(a_{t+1}|a_{t}, v_{s}, v_{g}, \hat{c})$ . The sequence starts with $a_{0} = \mathrm{START}$ and completes upon predicting $a_{T+1} = \mathrm{END}$ . Its implementation includes a learnable positional embedding $p_{t}$ for sequence generation. Therefore, the input sequence to predict $t$ -th action step $a_{t}$ is formulated as in Eq. (1), + +$$ +\left[ v _ {s}, v _ {g}, \hat {c}, l _ {\text {S T A R T}}, l _ {1} + p _ {1}, \dots , l _ {t - 2} + p _ {t - 2}, l _ {t - 1} + p _ {t - 1} \right] \tag {1} +$$ + +and the output sequence for supervision is Eq. (2), + +$$ +l = \left[ l _ {1}, l _ {2}, \dots , l _ {T}, l _ {\text {E N D}} \right] \tag {2} +$$ + +where $l_{t}$ represents the language representation for step $a_{t}$ , $l_{\mathrm{START}}$ and $l_{\mathrm{END}}$ are the embeddings for START and END tokens. The predicted action step $\hat{a}_{t}$ is determined by mapping to a probability distribution over the action vocabulary $V_{a}$ (which includes the END token), which is achieved by comparing the language similarity between predicted $\hat{l}_{t}$ and those within $V_{a}$ , enabling us to assign probabilities to each action within a plan sequence. During the training stage, we use the ground-truth action representation $l_{t}$ as input, while we use the predicted step representation $\hat{l}_{t}$ during the test stage. + +Retrieval Planner. To improve long-range sequence modeling, our model integrates a direct retrieval mechanism based on a training datastore, utilizing its state-action temporal dependency information, explicitly. This enhancement is achieved by augmenting the base planner in a manner similar to [7,10], by adding a retrieval-augmented memory module, which learns the state-action pairs in the training data as memory, by modeling states, as a context for the actions. Fig. 3 illustrated the workings of this component. + +The memory module consists of key-value pairs $\{s_n:d_n\}_{n = 1}^D$ , where $D$ is the memory size. The keys $(s_n)$ , are discrete learnable context vectors and the values $(d_{n})$ are their subsequent action step labels, conditioned on the context. We estimate a context vector $\varsigma_t$ , given representations of the initial visual state + +$v_{s}^{\prime}$ , the target state $v_{g}^{\prime}$ , the current action $l_{t}$ , and a learnable positional embedding $q_{t}$ as in Eq. (3) + +$$ +\varsigma_ {t} = f _ {\text {P r o j}} \left(\operatorname {C o n c a t} \left(v _ {s}, v _ {g}, \hat {c}, l _ {t} + q _ {t}\right)\right) \tag {3} +$$ + +With $f_{\mathrm{Proj}}$ as a projection layer and $\mathrm{Concat}(\cdot)$ concatenating inputs along the embedding dimension. In training, we learn the keys $(s_n)$ by comparing them to inputs' context vectors $\varsigma_t$ . We construct the memory values $(d_n)$ by uniformly sampling actions from the training dataset (including the END token), making sure of a balanced representation of actions. At inference, the model retrieves the $K$ nearest (i.e., most similar) keys $\{s_k\}_{k=1}^K$ from the memory to a given current context $\varsigma_t$ , returning their corresponding values $\{d_k\}_{k=1}^K$ . The nearest neighbors $(\{s_k\}_{k=1}^K)$ get selected by calculating the cosine distance between $\varsigma_t$ and the keys of the datastore $\{s_n\}_{n=1}^D$ . Upon identifying the nearest neighbors, a probability distribution across the retrieved values $(\{d_k\}_{k=1}^K)$ can be calculated by applying a softmax function to the similarity scores. The retrieval process can be formulated as $P_{\mathrm{retr}}(y|x) = P(a_{t+1}|\varsigma_t, a_{t+1} \in \{d_k\}_{k=1}^K)$ . The final prediction ensembles the predictions from base planner and retrieval planner. Following [7, 10], we obtain the final probability distribution for the next action step as: + +$$ +P _ {\text {a l l}} (y | x) = \lambda P _ {\text {r e t r}} (y | x) + (1 - \lambda) P _ {\text {b a s e}} (y | x) \tag {4} +$$ + +where $\lambda$ is a tuned hyper-parameter. + +# 3.3 Training + +RAP is trained in two stages: First, we train the base planner, then we augment and train the model including the retrieval planner. This approach ensures effective learning of the retrieval planner's keys, building upon the base planner's good initialization obtained from the initial training stage. + +Stage 1: Training the Base Planner. During the training phase, we first supervise the base model's output, the textual representation of each action step in the procedure sequence, through language supervision combined with contrastive learning [8]: For each feature $\hat{l}_t$ generated by the transformer's head, we take its ground truth language embedding $l_t$ as positive sample and all other embeddings in the action vocabulary $V_a$ as negative examples. The loss is formulated as: + +$$ +\mathcal {L} _ {\text {b a s e}} = - \sum_ {t = 1} ^ {T + 1} \log \frac {\exp \hat {l} _ {t} \cdot l _ {t}}{\sum_ {j} \exp \hat {l} _ {t} \cdot l _ {j}} \tag {5} +$$ + +Where $(T + 1)$ -th step denotes the END token. + +Stage 2: Training the Base and Retrieval Planners. In the second stage, the model is augmented with the nearest neighbor retrieval-based memory. The retrieval planner advances learning by juxtaposing the one-hot ground-truth labels $a_{t} \in p$ , with the model's interpolated probability estimates $P_{all,t}$ , for each action at each time-step $t$ . It refines the training by assessing how likely each action is, based on the model's predictions, instead of solely relying on matching + +the correct action. We achieve this, using a cross-entropy minimization in Eq. (6), + +$$ +\mathcal {L} _ {\text {a l l}} = - \sum_ {t = 1} ^ {T + 1} a _ {t} \log P _ {\text {a l l , t}} (y | x) \tag {6} +$$ + +where $P_{\mathrm{all,t}}(y|x)$ is defined based on Eq. (4). + +# 3.4 Addition of Unannotated Videos with Weak Supervision + +A significant challenge in procedure planning for instructional videos is the scarcity of annotated data. This is mainly because annotating instructional videos with step-level (i.e., timestamps) or sequence-level labels (i.e., action sequences) is resource-intensive. To address this issue, we propose a weakly-supervised training approach aimed at expanding the annotated training data to include additional unannotated task-related videos in the training process. + +Specifically, we align instructional videos with their corresponding topics, such as "making an omelet" or "making pancakes". This alignment links each video subset to a subset of instructional action plans sharing a similar topic. Crucially, this approach does not require detailed annotations that map specific action steps to individual video segments or frames. Instead, it relies on the broader thematic association between the video content and the instructional plans, allowing for a more general form of supervision during training. However, the acquired pairs of instructional videos and action plans may exhibit certain discrepancies. First, the presence of action steps in a plan that do not appear in the related video. Second, a mismatch in the sequence of action steps between the plan and their occurrence in the instructional video. Third, the inclusion of additional action steps in the video that are absent from the plan. + +To tackle these challenges, we applied a video grounding algorithm, taking inspiration from Dvornik et al. [5] to generate pseudo-annotations of descriptions in instructional videos. In short, our grounding algorithm processes an instructional video alongside an ordered sequence of action steps to produce a sequence of grounding results ( $t_{start}, t_{end}, a$ ), where $t_{start}$ and $t_{end}$ denote the beginning and end times of action step $a$ . The grounding model offers flexible sequence alignment and the ability to exclude irrelevant actions or frames, effectively addressing observed discrepancies. Further details on the grounding model are provided in the appendix. We utilize these pseudo-annotations, extracted for each relevant video, which serve as structurally and format-wise equivalent to manually annotated videos. These pseudo-annotations are integrated into Stage 1 training alongside original annotations, where the model is tasked with predicting the complete sequence of actions based on the starting frame of the first action and the ending frame of the last action outlined in the plan. + +This weak supervision of unannotated data, is an extension of Stage 1 training by using both original annotations of training videos and pseudo-annotations of task-relevant videos as the new training set. This training approach is a step towards a cost-effective solution for generalizing and scaling procedure planning to abundant unannotated videos. + +# 4 Experiments + +We experiment on two benchmark datasets and employ four metrics to validate the efficacy of our proposed model. + +![](images/360de421978424c0e5446f04d3867e03a38349d712fd0ec021f793109550e5fc.jpg) +Fig. 4: Illustration of the advantage of edit score, in evaluation of long varied-length sequences. As it can be seen, despite the prediction being plausible and acceptable, accuracy and success rate, falsely flag this scenario as a miss. + +# 4.1 Datasets + +Our methodology is evaluated using two instructional video datasets: CrossTask [26] and COIN [20]. CrossTask contains 2750 videos spanning 83 different topics, such as "How to make french toast?", with 7.6 actions per video on average. These topics are categorized into 18 primary and 65 related events. The primary topics are detailed with precise timestamps for each action step, providing a clear sequence of instructional activities. In contrast, the related instructional topics, gathered through automatic collection methods, lack annotations for action-step timestamps, sequence order, and the specific subset of action steps each video covers. Each topic within this related set is associated with a designated instructional plan as well as a set of unannotated instructional videos of the same topic. These videos may exhibit variations from the outlined plan, as detailed in Sec. 3.3, due to the previously mentioned discrepancies. We use CrossTask's related-set data, during weakly-supervised training with language supervision. COIN, includes 11827 videos, covering 778 different actions, with an average action-horizon of 3.6 actions per video. We follow [4,24] by creating train/test splits with a $70\% / 30\%$ ratio as well as implementing a moving window approach to organize the datasets into plans with various action-horizons. + +# 4.2 Metrics + +Our methodology's performance is evaluated using three standard metrics [2, 4, 19, 24]. First, mean Intersection over Union ( $mIoU$ ) assesses overlap between predicted and ground truth action sequences, calculated by $\frac{|a_t \cap \hat{a}_t|}{|a_t \cup \hat{a}_t|}$ . This metric + +evaluates whether the model identifies correct steps but does not consider action order or repetitions. Second, mean Accuracy ( $mAcc$ ) measures action alignment at each step, balancing the impact of repeated actions. Lastly, the Success Rate ( $SR$ ) evaluates a plan as successful if it exactly matches the ground truth. + +However, Accuracy and Success Rate have limitations, particularly with long or variable-length sequences, often leading to false negatives due to their sensitivity to action order. For example, in a pancake-making plan, if the predicted sequence adds a plausible step like "Add sugar" at the beginning, accuracy may incorrectly deem it inaccurate. This issue is more pronounced with sequences that vary in length from the ground truth, indicating the need for a more nuanced evaluation metric in procedure planning that accommodates plausible variations in actions. + +To address the limitations of the current metrics, we introduce a new metric, mean edit-score $(mES)$ , based on the minimum edit distance (e.g. Levenshtein edit distance), measuring the similarity between two sequences. This function, $L(s_{1},s_{2})|c_{sub},c_{ins},c_{del}$ , calculates the minimum number of substitutions $(c_{sub})$ , insertions $(c_{ins})$ , and deletions $(c_{del})$ needed to transform sequence $s_1$ into $s_2$ . The mean edit-score, is derived from the expected value of sample edit-scores. The edit-score is thus defined based on the optimal edit distance as: + +$$ +E S _ {s _ {1}, s _ {2}} = 1 0 0 \cdot \frac {\operatorname* {m a x} \left(\left| s _ {1} \right| , \left| s _ {2} \right|\right) - L \left(s _ {1} , s _ {2}\right) | _ {1 , 1 , 1}}{\operatorname* {m a x} \left(\left| s _ {1} \right| , \left| s _ {2} \right|\right)}, \tag {7} +$$ + +Where the operator $|\cdot|$ measures the length of a sequence. The mean edit-score offers a more flexible and forgiving assessment compared to strict metrics like success rate and mean accuracy, while still maintaining sensitivity to the order of action steps, hence complimenting $mIoU$ (i.e., if the ground-truth is add curry leaves, stir mixture, add chili powder and the prediction is add chili powder, stir mixture, add curry leaves, we face a high $IoU = 100\%$ and a low $ES = 33.33\%$ ). This metric provides a more comprehensive evaluation, especially in scenarios where the predicted sequence is plausible yet varies in actions or length from the ground truth (Fig. 4). + +# 4.3 Implementation Details + +Consistent with earlier approaches, we encode the start and goal visual observations using the S3D network [14] pre-trained on the HowTo100M [15] dataset. For optimization, we use the AdamW optimizer [13], setting the weight decay to 0.2. The learning rate is initially set at $lr = 0.001$ , and we apply a reduce-on-plateau scheduler to adjust this rate based on performance metrics. We experiment with an autoregressive transformer with 7 layers and 8 heads. We choose weight $\lambda = 0.1$ and volume $D = 1000$ , for the memory module. Further implementation details can be found in the supplementary section. + +Table 1: Assessment of procedure planning on CrossTask for action-horizons $T = 3$ and $T = 4$ . Vision supervision is denoted by V, language supervision by L, and weakly-supervised training by +wL. + +
ModelsSupervisionT=3T=4
SR↑mAcc↑mIoU↑SR↑mAcc↑mIoU↑
Random-<0.010.941.66<0.010.831.66
Retrieval-BasedV8.0523.3032.063.9522.2236.97
WLTDO [6]V1.8721.6431.700.7717.9226.43
UAAA [1]V2.1520.2130.870.9819.8627.09
UPN [18]V2.8924.3931.561.1921.5927.85
DDN [4]V12.1831.2947.485.9727.1048.46
Ext-GAIL w/o Aug. [2]V18.0143.8657.16---
Ext-GAIL [2]V21.2749.4661.7016.4143.0560.93
P3IV w/o Adv. [24]L22.1245.5767.40---
P3IV [24]L23.3449.9673.8913.4044.1670.01
E3P (baseline) [22]L22.5646.17-12.9743.97-
E3P [22]L26.4053.0274.0516.4948.0070.16
Skip-Plan [12]L28.8561.1874.9815.5655.6470.30
RAP (ours w/o unannotated data)L29.2362.3575.6316.9655.7771.49
RAP (ours with unannotated data)L+wL30.5162.6875.9217.1055.9171.33
+ +Table 2: Comparison of RAP with PDPP, when we align our state definition with PDPP's for CrossTask and $T = 3$ . + +
ModelsSupervisionSR ↑mAcc ↑mIoU ↑
PDPP [23]♣L37.2064.6766.57
RAP (ours w/o unannotated data)♣L38.5165.1877.44
+ +# 4.4 Comparison with State-of-The-Art Baselines + +Our study conducts a comprehensive comparison of our new framework, RAP, against established models in procedure planning, using CrossTask and COIN datasets. We refer to our framework without the retrieval-based memory augmentation, i.e., base planner, as BP (as opposed to RAP). We categorize previous approaches based on their supervision: language (L) or vision (V). Our proposed model's performance compared to the previous works, is detailed in Tabs. 1 and 2 for CrossTask, and Tab. 3 for COIN. + +The inherent uncertainty in variable action-horizon procedure planning models complicates direct and fair comparisons with traditional fixed-length models, due to the added complexity of predicting the conclusion of a sequence as well as the broader range of acceptable sequences of variable length other than the ground-truth. To ensure a fair evaluation, we compare our variable action-horizon architecture, trained and evaluated exclusively on fixed-horizon data (of length $T$ ), to the fixed-horizon models (of length $T$ ) from prior studies in procedure planning [1,2,4,6,12,18,22-24]. Tab. 1 further shows the performance boost achieved by the addition of CrossTask's unannotated data (i.e., CrossTask's related video set) through weakly-supervised training (supervision: $\mathrm{L} + \mathrm{wL}$ ). + +Our findings demonstrate that RAP exhibits superior performance over existing vision-supervised models. Notably, RAP, without the addition of unannotated data, surpasses Ext-GAIL [2] with a $7.97\%$ increase in Success Rate (SR), + +Table 3: Comparison of the procedure planing models' performance on the COIN dataset. + +
ModelsT=3T=4
SR↑mAcc↑mIoU↑SR↑mAcc↑mIoU↑
Random (V)<0.01<0.012.47<0.01<0.012.32
Retrieval-Based (V)4.3817.4032.062.7114.2936.97
DDN (V)13.9020.1964.7811.1317.7168.06
P3IV (L)15.4021.6776.3111.3218.8570.53
E3P (L)19.5731.4284.9513.5926.7284.72
PDPP (L)21.3345.6251.8214.4144.1051.39
Skip-plan (L)23.6547.1278.4416.0443.1977.07
RAP (w/o unannotated) (L)24.4747.5985.2416.7642.1184.64
+ +Table 4: Evaluation of RAP versus a retrieval-based (RB) model in variable-length sequence prediction on CrossTask and COIN. + +
ModelsCrossTaskCOIN
SR↑mAcc↑mIoU↑mES↑SR↑mAcc↑mIoU↑mES↑
Random0.020.561.521.210.020.090.180.17
RB8.0121.3238.8033.437.1112.9420.1717.69
RAP (w/o unannotated data)20.2849.4169.9562.7919.1433.2675.8257.48
RAP (with unannotated data)21.7950.5871.4463.86----
+ +12.89% in Mean Accuracy (mAcc), and 13.93% in Mean Intersection over Union (mIoU) for sequence lengths of $T = 3$ in CrossTask. + +When compared to language-supervised models like P3IV, E3P, and SkipPlan, RAP demonstrates superior performance on CrossTask. For $\mathrm{T} = 3$ , RAP exhibits improvements of $5.99\%$ , $2.83\%$ , and $0.38\%$ respectively, and for $\mathrm{T} = 4$ , it shows enhancements of $3.56\%$ , $0.47\%$ , and $1.4\%$ in SR. Tab. 2 compares our model with PDPP [23], which employs a different procedure state definition. For fairness, we evaluate our model under the PDPP setting, marked with $\clubsuit$ , showcasing its superior performance. PDPP uses a 3-second window following the onset and preceding the offset of action steps to capture visual states, giving it an advantage in accessing action-related information compared to previous models. Traditionally, a 1-second window centered around the start and end of actions is used. By aligning our state definition with PDPP's, our RAP model outperforms PDPP on all metrics, as shown in the table. We observe a similar superiority in performance, when evaluating the model on the COIN dataset as shown in Tab. 3. + +We assessed the efficacy of RAP in handling tasks with variable lengths by conducting experiments against a retrieval-based model, which operates under a similar variable action-horizon setting. The results of this comparison are presented in Tab. 4. As demonstrated in the table, RAP proves a remarkable superiority over the retrieval-based model across all evaluated metrics. This is particularly evident in the mES metric, where RAP (w/o unannotated data) shows a significant improvement by $29.36\%$ . + +On CrossTask, RAP is further trained using the weak language supervision framework outlined in Sec. 3.3, utilizing the unannotated data from the dataset's + +Table 5: A study on the effect of retrieval augmentation (BP vs RAP), for variable and fixed action-horizon settings. + +
DatasetModelsT=3T=4Variable Length T
SR↑mAcc↑mIoU↑mES↑SR↑mAcc↑mIoU↑mES↑SR↑mAcc↑mIoU↑mES↑
CrossTask (w/o unannotated)BP25.4258.7073.9165.5915.2352.1670.5361.6318.5447.8569.1662.64
RAP29.2362.3575.6367.3316.9655.7771.4964.5120.2849.4169.9562.79
COINBP23.3846.2180.4962.2814.2941.6778.6257.2617.8631.3472.4456.24
RAP24.4747.5985.2463.1516.7642.1184.6460.4719.1433.2675.8257.48
+ +![](images/e9fb284d490254a8c076638e4ba694c576929103779744433c78025910d4458e.jpg) +Fig. 5: Sample plausible predicted plan sequences generated by RAP vs the ground-truth (GT). + +related-tasks. As shown in Tab. 1 and Tab. 4 when training RAP with the addition of the unannotated data, we see significant improvement in performance across all metrics. + +Fig. 5 displays two qualitative results of RAP's performance in predicting plans of varying lengths compared to the ground truths. In the top row, it accurately identifies the addition of fish, other ingredients, and the plausible action "add onions". In the bottom row, it recognizes the completed bread, its removal, and suggests the reasonable action "melt butter", before the new bread is put in the pan, showcasing the model's planning abilities. + +# 4.5 Ablation Study + +Impact of retrieval augmentation. The introduction of a retrieval-based memory component significantly enhances our framework's performance. Tab. 5 demonstrates improvements for both fixed and variable action-horizon settings. Notably, enhancements include increases of $3.81\%$ , $1.73\%$ , and $1.74\%$ in Success Rate (SR) for $\mathrm{T} = 3$ and $\mathrm{T} = 4$ in the fixed setting, and in the variable setting, respectively, on CrossTask. We observe a similar boost in performance for the COIN dataset, as evidenced in the Tab. 5. Tab. 6 details the enhancements in SR, categorized by the ground-truth (GT) sequences' action-horizon in the variable-length sequence prediction setting. + +Impact of weak language supervision. The influence of weak language supervision on our model's performance is evaluated through an ablation study on CrossTask, where we incrementally increased the unannotated data used in weakly-supervised training. Tab. 7 displays the varying percentages of unan- + +Table 6: Impact of retrieval augmentation (BP vs RAP) in $SR$ , on CrossTask, categorized by ground-truth (GT) action-horizon in the variable action-horizon setting. + +
Model (w/o unannotated data)GT action-horizon
123456
BP49.5827.8416.218.375.893.34
RAP52.6029.2018.0210.526.165.18
+ +Table 7: Weak language supervision impact: Table shows RAP's performance gains on CrossTask with increasing unannotated data use. + +
Used unannotated data portionSR↑mAcc↑mIoU↑mES↑
No unannotated data20.2849.4169.9562.79
25%19.2047.8869.0561.96
50%20.5749.7269.7563.12
75%21.3150.2270.8663.40
100%21.7950.5871.4463.86
+ +notated data utilized. The results show consistent performance gains for RAP under weak language supervision, particularly as more data is introduced (except at the $25\%$ data level). This suggests potential for further enhancements with increased unannotated data. Notably, using $100\%$ of unannotated data yields improvements in SR, mAcc, mIoU, and mES by $1.51\%$ , $1.17\%$ , $1.49\%$ , and $1.07\%$ , respectively. + +# 5 Conclusion + +In this study, we addressed the challenge of procedure planning in instructional videos by developing the Retrieval-Augmented Planner (RAP) model. This model overcomes the limitations of previous approaches that were restricted to fixed action-step predictions. RAP excels in both fixed and varied action horizon scenarios, outperforming existing models. A key innovation is our weak language supervision method, which uses thematic links between unannotated videos and plans for pseudo-annotation generation, enhancing model training. Our model also incorporates a learnable retrieval-based memory component for improved long-range sequence predictions and efficient use of temporal data. Additionally, we introduced "edit-score", a new metric for assessing variable-length procedures. Looking forward, we see potential in applying RAP to more diverse scenarios and in generating probabilistic, varied-length plans. We also envisage that expanding our weak supervision approach to include an even wider range of unannotated data could markedly elevate the performance of procedure planning models, paving the way for more advanced applications in this field. + +# References + +1. Abu Farha, Y., Gall, J.: Uncertainty-aware anticipation of activities. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. pp. 0-0 (2019) +2. Bi, J., Luo, J., Xu, C.: Procedure planning in instructional videos via contextual modeling and model-based policy learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15611-15620 (2021) +3. Borgeaud, S., Mensch, A., Hoffmann, J., Cai, T., Rutherford, E., Millican, K., Van Den Driessche, G.B., Lespiau, J.B., Damoc, B., Clark, A., et al.: Improving language models by retrieving from trillions of tokens. In: International conference on machine learning. pp. 2206-2240. PMLR (2022) +4. Chang, C.Y., Huang, D.A., Xu, D., Adeli, E., Fei-Fei, L., Niebles, J.C.: Procedure planning in instructional videos. In: European Conference on Computer Vision. pp. 334-350. Springer (2020) +5. Dvornik, M., Hadji, I., Derpanis, K.G., Garg, A., Jepson, A.: Drop-dtw: Aligning common signal between sequences while dropping outliers. Advances in Neural Information Processing Systems 34, 13782-13793 (2021) +6. Ehsani, K., Bagherinezhad, H., Redmon, J., Mottaghi, R., Farhadi, A.: Who let the dogs out? modeling dog behavior from visual data. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4051-4060 (2018) +7. Grave, E., Cisse, M.M., Joulin, A.: Unbounded cache model for online language modeling with open vocabulary. Advances in neural information processing systems 30 (2017) +8. Gutmann, M., Hyvarinen, A.: Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In: Proceedings of the thirteenth international conference on artificial intelligence and statistics. pp. 297-304. JMLR Workshop and Conference Proceedings (2010) +9. Hadji, I., Derpanis, K.G., Jepson, A.D.: Representation learning via global temporal alignment and cycle-consistency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11068-11077 (2021) +0. Khandelwal, U., Levy, O., Jurafsky, D., Zettlemoyer, L., Lewis, M.: Generalization through memorization: Nearest neighbor language models. arXiv preprint arXiv:1911.00172 (2019) +1. Lei, J., Wang, L., Shen, Y., Yu, D., Berg, T.L., Bansal, M.: Mart: Memory-augmented recurrent transformer for coherent video paragraph captioning. arXiv preprint arXiv:2005.05402 (2020) +2. Li, Z., Geng, W., Li, M., Chen, L., Tang, Y., Lu, J., Zhou, J.: Skip-plan: Procedure planning in instructional videos via condensed action space learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10297-10306 (2023) +3. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017) +4. Miech, A., Alayrac, J.B., Smaira, L., Laptev, I., Sivic, J., Zisserman, A.: End-to-end learning of visual representations from uncurated instructional videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9879-9889 (2020) +5. Miech, A., Zhukov, D., Alayrac, J.B., Tapaswi, M., Laptev, I., Sivic, J.: Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2630-2640 (2019) + +16. Niu, Y., Guo, W., Chen, L., Lin, X., Chang, S.F.: Schema: State changes matter for procedure planning in instructional videos. arXiv preprint arXiv:2403.01599 (2024) +17. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019) +18. Srinivas, A., Jabri, A., Abbeel, P., Levine, S., Finn, C.: Universal planning networks: Learning generalizable representations for visuomotor control. In: International Conference on Machine Learning. pp. 4732-4741. PMLR (2018) +19. Sun, J., Huang, D.A., Lu, B., Liu, Y.H., Zhou, B., Garg, A.: Plate: Visually-grounded planning with transformers in procedural tasks. IEEE Robotics and Automation Letters 7(2), 4924-4930 (2022) +20. Tang, Y., Ding, D., Rao, Y., Zheng, Y., Zhang, D., Zhao, L., Lu, J., Zhou, J.: Coin: A large-scale dataset for comprehensive instructional video analysis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1207-1216 (2019) +21. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) +22. Wang, A.L., Lin, K.Y., Du, J.R., Meng, J., Zheng, W.S.: Event-guided procedure planning from instructional videos with text supervision. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13565-13575 (2023) +23. Wang, H., Wu, Y., Guo, S., Wang, L.: Pdpp: Projected diffusion for procedure planning in instructional videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14836-14845 (2023) +24. Zhao, H., Hadji, I., Dvornik, N., Derpanis, K.G., Wildes, R.P., Jepson, A.D.: P3iv: Probabilistic procedure planning from instructional videos with weak supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2938-2948 (2022) +25. Zhong, Z., Lei, T., Chen, D.: Training language models with memory augmentation (2022) +26. Zhukov, D., Alayrac, J.B., Cinbis, R.G., Fouhey, D., Laptev, I., Sivic, J.: Cross-task weakly supervised learning from instructional videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3537-3545 (2019) \ No newline at end of file diff --git a/2024/RAP_ Retrieval-Augmented Planner for Adaptive Procedure Planning in Instructional Videos/images.zip b/2024/RAP_ Retrieval-Augmented Planner for Adaptive Procedure Planning in Instructional Videos/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..70680fbc6e5b8a0940dade2fd10a8b9704d4c3f3 --- /dev/null +++ b/2024/RAP_ Retrieval-Augmented Planner for Adaptive Procedure Planning in Instructional Videos/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d1f9aa78ac172eb158137ac5fec88148cfaa56992fb2257457058ebf037b32d5 +size 449357 diff --git a/2024/RAP_ Retrieval-Augmented Planner for Adaptive Procedure Planning in Instructional Videos/layout.json b/2024/RAP_ Retrieval-Augmented Planner for Adaptive Procedure Planning in Instructional Videos/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d8c31f8dc4b12c7db4199a0f270037cc9c34ded0 --- /dev/null +++ b/2024/RAP_ Retrieval-Augmented Planner for Adaptive Procedure Planning in Instructional Videos/layout.json @@ -0,0 +1,8822 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "type": "text", + "content": "RAP: Retrieval-Augmented Planner for Adaptive Procedure Planning in Instructional Videos" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 176, + 168, + 438, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 168, + 438, + 181 + ], + "spans": [ + { + "bbox": [ + 176, + 168, + 438, + 181 + ], + "type": "text", + "content": "Ali Zare, Yulei Niu*, Hammad Ayyubi, and Shih-Fu Chang" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 206, + 190, + 407, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 190, + 407, + 213 + ], + "spans": [ + { + "bbox": [ + 206, + 190, + 407, + 213 + ], + "type": "text", + "content": "Columbia University, New York, NY 10027, USA {az2584,yn2338,ha2578,sc250}@columbia.edu," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 239, + 455, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 239, + 455, + 524 + ], + "spans": [ + { + "bbox": [ + 160, + 239, + 455, + 524 + ], + "type": "text", + "content": "Abstract. Procedure Planning in instructional videos entails generating a sequence of action steps based on visual observations of the initial and target states. Despite the rapid progress in this task, there remain several critical challenges to be solved: (1) Adaptive procedures: Prior works hold an unrealistic assumption that the number of action steps is known and fixed, leading to non-generalizable models in real-world scenarios where the sequence length varies. (2) Temporal relation: Understanding the step temporal relation knowledge is essential in producing reasonable and executable plans. (3) Annotation cost: Annotating instructional videos with step-level labels (i.e., timestamp) or sequence-level labels (i.e., action category) is demanding and labor-intensive, limiting its generalizability to large-scale datasets. In this work, we propose a new and practical setting, called adaptive procedure planning in instructional videos, where the procedure length is not fixed or pre-determined. To address these challenges, we introduce Retrieval-Augmented Planner (RAP) model. Specifically, for adaptive procedures, RAP adaptively determines the conclusion of actions using an auto-regressive model architecture. For temporal relation, RAP establishes an external memory module to explicitly retrieve the most relevant state-action pairs from the training videos and revises the generated procedures. To tackle high annotation cost, RAP utilizes a weakly-supervised learning manner to expand the training dataset to other task-relevant, unannotated videos by generating pseudo labels for action steps. Experiments on CrossTask and COIN benchmarks show the superiority of RAP over traditional fixed-length models, establishing it as a strong baseline solution for adaptive procedure planning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 534, + 453, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 534, + 453, + 556 + ], + "spans": [ + { + "bbox": [ + 160, + 534, + 453, + 556 + ], + "type": "text", + "content": "Keywords: Procedure Planning, Instructional Video, Retrieval Augmentation, Weak Supervision" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 575, + 230, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 575, + 230, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 575, + 230, + 588 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 599, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 599, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 599, + 482, + 648 + ], + "type": "text", + "content": "Procedure planning in instructional videos aims to generate an executive series of actions for completing tasks, transitioning from initial visual state to final visual state [2, 4, 12, 16, 19, 22-24]. While humans excel at such tasks, machine learning agents face challenges in achieving similar proficiency. Success in procedure" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 654, + 234, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 654, + 234, + 666 + ], + "spans": [ + { + "bbox": [ + 134, + 654, + 234, + 666 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 115, + 478, + 265 + ], + "blocks": [ + { + "bbox": [ + 135, + 115, + 478, + 265 + ], + "lines": [ + { + "bbox": [ + 135, + 115, + 478, + 265 + ], + "spans": [ + { + "bbox": [ + 135, + 115, + 478, + 265 + ], + "type": "image", + "image_path": "e7068740273b45094a058d2c55ec89fba3f9992dabac1717baaf041c2aef99bc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 275, + 482, + 342 + ], + "lines": [ + { + "bbox": [ + 130, + 275, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 482, + 342 + ], + "type": "text", + "content": "Fig. 1: Examples of adaptive procedures in instructional videos. Procedure length varies with differences between initial and goal visual states, even for the same task. Previous works assume a fixed, known sequence length, requiring a separate model for each action-horizon. This figure contrasts traditional procedure planning (A) with our adaptive approach (B), which trains a unified model without knowing the procedure length in advance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 374, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 422 + ], + "type": "text", + "content": "planning requires an agent's deep understanding of actions, procedure states, involved objects, and their temporal relationships. Narrowing this performance gap is key to advancing autonomous systems, equipping them to adeptly perform complex, instruction-based tasks like cooking or repair." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": "Although recent works have made tremendous strides, there remain several challenges to real-world applications. First, previous works assume that the number of action steps is known and fixed. As depicted in Fig. 1, the visual goal may be achieved through a small number of steps or a long procedure, depending on the differences between the initial and goal states. For example, it takes three steps to transform a visual of adding butter to a visual of flipped french toast in example b (scenario b implies pre-prepared mixture and pre-dipped bread, but visually requires actions like \"melt butter,\" \"put bread in pan,\" and \"flip bread\"). but one needs six and five steps to finish the cooking process based on the visuals of the mixture in scenarios a and c respectively. Previous models, trained on the unrealistic assumption of fixed prediction horizon, cannot figure out the complexity of state understanding and are hard to generalize to real-world applications. Secondly, there is a high correlation between consecutive action steps. For example, \"add\" is often followed by \"mix\", and \"wash\" is often followed by \"rinse\". Existing methods model this relationship implicitly via intermediate states, leading to sub-par results. Third, prior methods focused on fully annotated data [2,4,19] - action steps and intermediate visual states - for supervision. However, this approach poses challenges in scaling up the training data due to its costliness. Following works [12, 16, 22-24] proposed to leverage text representations of action steps as language supervision, which still necessitate" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "A. Zare et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "sequence-level annotations, a time-consuming and expensive process involving watching entire instructional videos." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 272 + ], + "type": "text", + "content": "To tackle these challenges, we introduce the setting of adaptive procedure planning in instructional videos and propose the Retrieval-Augmented Planner (RAP) model for this problem. Overall, RAP is designed to (1) adaptively predict procedures with variable lengths according to the complexity of visual states, (2) explicitly exploit step temporal relation via knowledge retrieval from the training data, and (3) learn procedural knowledge from additional unannotated videos. Specifically, RAP consists of an auto-regressive model as the base planner for adaptive procedure prediction, and a retrieval-augmented memory module to refine the generated procedures. The architecture is supplemented by weakly-supervised learning method that leverages both annotated video-procedure pairs and unannotated videos." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 274, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 344 + ], + "type": "text", + "content": "In addition, existing metrics for procedure planning, such as accuracy, fail to fairly evaluate variable length procedures (as described in Sec. 4.2). We address this by introducing a new metric edit-score. We evaluate our model on two widely used datasets for instructional videos, CrossTask [26] and COIN [20]. Our model demonstrates superior performance in the conventional setting and establishes a strong baseline in the new adaptive procedure planning setting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 346, + 272, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 346, + 272, + 357 + ], + "spans": [ + { + "bbox": [ + 146, + 346, + 272, + 357 + ], + "type": "text", + "content": "Our main contributions are:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 360, + 480, + 457 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 360, + 479, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 360, + 479, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 360, + 479, + 396 + ], + "type": "text", + "content": "- We argue the unrealistic assumption of fixed-length plans in previous procedure planning works, and introduce a new setting adaptive procedure planning in instructional videos." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 397, + 480, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 397, + 480, + 432 + ], + "spans": [ + { + "bbox": [ + 138, + 397, + 480, + 432 + ], + "type": "text", + "content": "- We propose Retrieval-Augmented Planner that can generate adaptive procedures of variable lengths and retrieve step temporal relation knowledge from the training data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 434, + 480, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 480, + 457 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 480, + 457 + ], + "type": "text", + "content": "- We introduce a weakly-supervised training approach to learn procedural knowledge from unannotated task-relevant videos." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 480, + 234, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 480, + 234, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 480, + 234, + 493 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 510, + 261, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 261, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 261, + 522 + ], + "type": "text", + "content": "2.1 Procedure Planning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "content": "Procedure planning from instructional videos, involves generating effective plans for task completion based on start and final visual observations. Earlier methods utilized a two-branch architecture, alternating between action and state prediction, with models like recurrent neural networks [4], transformers [19], and adversarial networks [2]. Recent approaches, such as P3IV's [24] memory-augmented Transformer and E3P, a prompting-based model [22], adopt one-branch transformer models. However, these methods are limited to predicting procedures of fixed lengths, lacking adaptability to tasks with varying action sequence length requirements. Our work introduces RAP, a novel one-branch approach designed to predict plans with variable action-horizons, addressing this limitation in the literature." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "RAP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 444, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 444, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 444, + 128 + ], + "type": "text", + "content": "2.2 Sequence Generation with Auto-regressive Transformers" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 133, + 482, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 133, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 133, + 482, + 205 + ], + "type": "text", + "content": "Recent advancements in sequence generation using Transformer-based models [17,21] have established remarkable results in tasks involving sequence modeling, e.g. natural language modeling [17]. Procedure planning in its core, is a goal-conditioned sequence generation problem, thereby directly benefiting from the state-of-the-art sequence generation models. Leveraging these developments, our work incorporates an auto-regressive Transformer [17] into the RAP architecture." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 218, + 317, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 317, + 229 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 317, + 229 + ], + "type": "text", + "content": "2.3 Retrieval-Augmented Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 234, + 481, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 234, + 481, + 318 + ], + "spans": [ + { + "bbox": [ + 130, + 234, + 481, + 318 + ], + "type": "text", + "content": "Augmenting language models (LM) with external memory [3, 7, 10, 11, 25] has proved to significantly improve language sequence generation. P3IV [24] uses a memory component to enhance long-range sequence modeling in procedure planning. Similarly, we adopt a retrieval-based mechanism, drawing inspiration from Khandelwal et al. [10]'s augmentation of a neural language model with a k-nearest neighbors (kNN) retrieval component, in our auto-regressive model to effectively leverage temporal prior information for predicting action sequences." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 332, + 248, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 248, + 344 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 248, + 344 + ], + "type": "text", + "content": "2.4 Video Grounding" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 348, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 482, + 456 + ], + "type": "text", + "content": "Procedure planning suffers from a lack of annotated data. Earlier works heavily relied on fully annotated data (action-state timestamps) [2, 4, 19]. Recent approaches [22, 24] have used language supervision during training which relies on less-costly, yet still resource-intensive language-level annotations. In this work, we introduce a weak language supervision framework that employs a video-language grounding algorithm to create pseudo-annotations from unannotated data. This approach is inspired by Dvornik et al. [5] in sequence-to-sequence alignment research, utilizing a grounding algorithm that aligns video frame and action step representations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 470, + 268, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 470, + 268, + 484 + ], + "spans": [ + { + "bbox": [ + 132, + 470, + 268, + 484 + ], + "type": "text", + "content": "3 Technical approach" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 491, + 481, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 481, + 539 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 481, + 539 + ], + "type": "text", + "content": "In this section, we introduce our proposed framework Retrieval-Augmented Planner for adaptive procedure planning in instructional videos, and expand on a weakly-supervised learning method to leverage unannotated but task-relevant videos." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 553, + 354, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 553, + 354, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 553, + 354, + 566 + ], + "type": "text", + "content": "3.1 Setting: Adaptive Procedure Planning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "In conventional formulation of procedure planning from instructional videos, a model is required to generate a plan " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "p = \\{a_1, a_2, \\dots, a_T\\}" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": ", given visual observations of the start " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "o_s" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": ", and the goal " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "o_g" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " states, where " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " is the plan's length (i.e., action-horizon). Previous methods assume a fixed, known action-horizon, which is unrealistic for real-world applicability and limits the generalizability of planning models. As shown in Fig. 1, we consider a more general and practical setting, adaptive procedure planning in instructional videos, where the procedure length is not fixed." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "A. Zare et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 115, + 444, + 293 + ], + "blocks": [ + { + "bbox": [ + 151, + 115, + 444, + 293 + ], + "lines": [ + { + "bbox": [ + 151, + 115, + 444, + 293 + ], + "spans": [ + { + "bbox": [ + 151, + 115, + 444, + 293 + ], + "type": "image", + "image_path": "a56bf2ea0c49957863420635fb13fb856d817bb74dedb0a7bbbc45528e66ffad.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 303, + 482, + 381 + ], + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 381 + ], + "type": "text", + "content": "Fig.2: Framework overview: Visual observations " + }, + { + "bbox": [ + 130, + 303, + 482, + 381 + ], + "type": "inline_equation", + "content": "o_s" + }, + { + "bbox": [ + 130, + 303, + 482, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 303, + 482, + 381 + ], + "type": "inline_equation", + "content": "o_g" + }, + { + "bbox": [ + 130, + 303, + 482, + 381 + ], + "type": "text", + "content": " are encoded to " + }, + { + "bbox": [ + 130, + 303, + 482, + 381 + ], + "type": "inline_equation", + "content": "v_s" + }, + { + "bbox": [ + 130, + 303, + 482, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 303, + 482, + 381 + ], + "type": "inline_equation", + "content": "v_g" + }, + { + "bbox": [ + 130, + 303, + 482, + 381 + ], + "type": "text", + "content": ". A Task-Classifier predicts a task class and its representation " + }, + { + "bbox": [ + 130, + 303, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\hat{c}" + }, + { + "bbox": [ + 130, + 303, + 482, + 381 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 130, + 303, + 482, + 381 + ], + "type": "inline_equation", + "content": "v_s" + }, + { + "bbox": [ + 130, + 303, + 482, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 303, + 482, + 381 + ], + "type": "inline_equation", + "content": "v_g" + }, + { + "bbox": [ + 130, + 303, + 482, + 381 + ], + "type": "text", + "content": ". These, along with a learnable START token, are fed into the auto-regressive base planner " + }, + { + "bbox": [ + 130, + 303, + 482, + 381 + ], + "type": "inline_equation", + "content": "F_{base}" + }, + { + "bbox": [ + 130, + 303, + 482, + 381 + ], + "type": "text", + "content": " for sequential action embedding prediction. For each prediction at position " + }, + { + "bbox": [ + 130, + 303, + 482, + 381 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 303, + 482, + 381 + ], + "type": "text", + "content": ", the retrieval component uses the predicted context vector " + }, + { + "bbox": [ + 130, + 303, + 482, + 381 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 130, + 303, + 482, + 381 + ], + "type": "text", + "content": " to estimate a probability distribution " + }, + { + "bbox": [ + 130, + 303, + 482, + 381 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{retr},t}" + }, + { + "bbox": [ + 130, + 303, + 482, + 381 + ], + "type": "text", + "content": " over the closest " + }, + { + "bbox": [ + 130, + 303, + 482, + 381 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 303, + 482, + 381 + ], + "type": "text", + "content": " key-value pairs in memory for kNN linear interpolation with " + }, + { + "bbox": [ + 130, + 303, + 482, + 381 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{base},t}" + }, + { + "bbox": [ + 130, + 303, + 482, + 381 + ], + "type": "text", + "content": " to estimate the final probability distribution " + }, + { + "bbox": [ + 130, + 303, + 482, + 381 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{all},t}" + }, + { + "bbox": [ + 130, + 303, + 482, + 381 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 404, + 350, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 404, + 350, + 417 + ], + "spans": [ + { + "bbox": [ + 132, + 404, + 350, + 417 + ], + "type": "text", + "content": "3.2 Model: Retrieval-Augmented Planner" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "text", + "content": "We propose a new framework, Retrieval-Augmented Planner (RAP), to tackle the adaptive procedure planning problem. Fig. 2 illustrates our framework. As the first step, we obtain state " + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "text", + "content": ", step " + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "inline_equation", + "content": "l_{j}" + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "text", + "content": ", and task " + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "text", + "content": " representations from visual observation " + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "inline_equation", + "content": "o_{i}" + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "text", + "content": ", action step " + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "inline_equation", + "content": "a_{j}" + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "text", + "content": ", and task title " + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "inline_equation", + "content": "tsk" + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "text", + "content": " (e.g. \"make pancakes\") respectively. This is done by inputting " + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "inline_equation", + "content": "o_{i}" + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "inline_equation", + "content": "a_{j}" + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "inline_equation", + "content": "tsk" + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "text", + "content": " to a pretrained vision and language encoder respectively ([14]), similar to P3IV [24]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 509, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 617 + ], + "type": "text", + "content": "Task Classifier. Recent studies [22, 23] have demonstrated the effectiveness of integrating task information into training to enhance model performance in procedure planning. In alignment with previous work [23], we adopt a similar strategy by first training an MLP task classifier. This classifier takes initial and final visual observations as inputs and predicts the task class " + }, + { + "bbox": [ + 130, + 509, + 482, + 617 + ], + "type": "inline_equation", + "content": "t\\hat{s}k" + }, + { + "bbox": [ + 130, + 509, + 482, + 617 + ], + "type": "text", + "content": ". The training is conducted using a cross-entropy loss function with the ground-truth task class " + }, + { + "bbox": [ + 130, + 509, + 482, + 617 + ], + "type": "inline_equation", + "content": "tsk" + }, + { + "bbox": [ + 130, + 509, + 482, + 617 + ], + "type": "text", + "content": ". Subsequently, we incorporate the task representation " + }, + { + "bbox": [ + 130, + 509, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\hat{c}" + }, + { + "bbox": [ + 130, + 509, + 482, + 617 + ], + "type": "text", + "content": " corresponding to the predicted task class " + }, + { + "bbox": [ + 130, + 509, + 482, + 617 + ], + "type": "inline_equation", + "content": "t\\hat{s}k" + }, + { + "bbox": [ + 130, + 509, + 482, + 617 + ], + "type": "text", + "content": " as an additional input to both the Base-Planner and the Retrieval-Planner." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "Base Planner. To facilitate the adaptive procedure planning setting, we utilize an auto-regressive transformer [17] as the base planner " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{base}}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": ". The base planner takes the representations of the initial state " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "v_{s}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": ", the goal state " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "v_{g}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": ", the predicted" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "RAP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 186, + 116, + 427, + 228 + ], + "blocks": [ + { + "bbox": [ + 186, + 116, + 427, + 228 + ], + "lines": [ + { + "bbox": [ + 186, + 116, + 427, + 228 + ], + "spans": [ + { + "bbox": [ + 186, + 116, + 427, + 228 + ], + "type": "image", + "image_path": "bf4372253c13069322c35a168a1a1279c7652df37e7ffa7d94deff8326ca5a89.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 238, + 482, + 273 + ], + "lines": [ + { + "bbox": [ + 131, + 238, + 482, + 273 + ], + "spans": [ + { + "bbox": [ + 131, + 238, + 482, + 273 + ], + "type": "text", + "content": "Fig. 3: The Retrieval-Augmented Planner module uses the predicted context vector " + }, + { + "bbox": [ + 131, + 238, + 482, + 273 + ], + "type": "inline_equation", + "content": "\\varsigma_{t}" + }, + { + "bbox": [ + 131, + 238, + 482, + 273 + ], + "type": "text", + "content": " to retrieve the " + }, + { + "bbox": [ + 131, + 238, + 482, + 273 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 238, + 482, + 273 + ], + "type": "text", + "content": " closest keys and their values from memory, estimating a probability distribution " + }, + { + "bbox": [ + 131, + 238, + 482, + 273 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{retr},\\mathrm{t}}" + }, + { + "bbox": [ + 131, + 238, + 482, + 273 + ], + "type": "text", + "content": " over these entries for next action probability estimation " + }, + { + "bbox": [ + 131, + 238, + 482, + 273 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{all},\\mathrm{t}}" + }, + { + "bbox": [ + 131, + 238, + 482, + 273 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 294, + 482, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 294, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 131, + 294, + 482, + 354 + ], + "type": "text", + "content": "task " + }, + { + "bbox": [ + 131, + 294, + 482, + 354 + ], + "type": "inline_equation", + "content": "\\hat{c}" + }, + { + "bbox": [ + 131, + 294, + 482, + 354 + ], + "type": "text", + "content": ", and the current action " + }, + { + "bbox": [ + 131, + 294, + 482, + 354 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 131, + 294, + 482, + 354 + ], + "type": "text", + "content": ", as input, to predict the next action " + }, + { + "bbox": [ + 131, + 294, + 482, + 354 + ], + "type": "inline_equation", + "content": "a_{t+1}" + }, + { + "bbox": [ + 131, + 294, + 482, + 354 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 131, + 294, + 482, + 354 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{base}}(y|x) = P(a_{t+1}|a_{t}, v_{s}, v_{g}, \\hat{c})" + }, + { + "bbox": [ + 131, + 294, + 482, + 354 + ], + "type": "text", + "content": ". The sequence starts with " + }, + { + "bbox": [ + 131, + 294, + 482, + 354 + ], + "type": "inline_equation", + "content": "a_{0} = \\mathrm{START}" + }, + { + "bbox": [ + 131, + 294, + 482, + 354 + ], + "type": "text", + "content": " and completes upon predicting " + }, + { + "bbox": [ + 131, + 294, + 482, + 354 + ], + "type": "inline_equation", + "content": "a_{T+1} = \\mathrm{END}" + }, + { + "bbox": [ + 131, + 294, + 482, + 354 + ], + "type": "text", + "content": ". Its implementation includes a learnable positional embedding " + }, + { + "bbox": [ + 131, + 294, + 482, + 354 + ], + "type": "inline_equation", + "content": "p_{t}" + }, + { + "bbox": [ + 131, + 294, + 482, + 354 + ], + "type": "text", + "content": " for sequence generation. Therefore, the input sequence to predict " + }, + { + "bbox": [ + 131, + 294, + 482, + 354 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 294, + 482, + 354 + ], + "type": "text", + "content": "-th action step " + }, + { + "bbox": [ + 131, + 294, + 482, + 354 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 131, + 294, + 482, + 354 + ], + "type": "text", + "content": " is formulated as in Eq. (1)," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 195, + 362, + 481, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 362, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 195, + 362, + 481, + 376 + ], + "type": "interline_equation", + "content": "\\left[ v _ {s}, v _ {g}, \\hat {c}, l _ {\\text {S T A R T}}, l _ {1} + p _ {1}, \\dots , l _ {t - 2} + p _ {t - 2}, l _ {t - 1} + p _ {t - 1} \\right] \\tag {1}", + "image_path": "3946d42f31060eed824bf99d35f8ec61f8b2d8c7bc530c6822a51aecaf0e7fa4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 384, + 358, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 384, + 358, + 396 + ], + "spans": [ + { + "bbox": [ + 131, + 384, + 358, + 396 + ], + "type": "text", + "content": "and the output sequence for supervision is Eq. (2)," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 258, + 404, + 481, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 404, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 258, + 404, + 481, + 418 + ], + "type": "interline_equation", + "content": "l = \\left[ l _ {1}, l _ {2}, \\dots , l _ {T}, l _ {\\text {E N D}} \\right] \\tag {2}", + "image_path": "81d080ef4dfae5222282e9dfea0d29370ed54a9d4d1d8c8c3878ccaccc02a0bf.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 425, + 482, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 425, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 131, + 425, + 482, + 522 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 425, + 482, + 522 + ], + "type": "inline_equation", + "content": "l_{t}" + }, + { + "bbox": [ + 131, + 425, + 482, + 522 + ], + "type": "text", + "content": " represents the language representation for step " + }, + { + "bbox": [ + 131, + 425, + 482, + 522 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 131, + 425, + 482, + 522 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 425, + 482, + 522 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{START}}" + }, + { + "bbox": [ + 131, + 425, + 482, + 522 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 425, + 482, + 522 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{END}}" + }, + { + "bbox": [ + 131, + 425, + 482, + 522 + ], + "type": "text", + "content": " are the embeddings for START and END tokens. The predicted action step " + }, + { + "bbox": [ + 131, + 425, + 482, + 522 + ], + "type": "inline_equation", + "content": "\\hat{a}_{t}" + }, + { + "bbox": [ + 131, + 425, + 482, + 522 + ], + "type": "text", + "content": " is determined by mapping to a probability distribution over the action vocabulary " + }, + { + "bbox": [ + 131, + 425, + 482, + 522 + ], + "type": "inline_equation", + "content": "V_{a}" + }, + { + "bbox": [ + 131, + 425, + 482, + 522 + ], + "type": "text", + "content": " (which includes the END token), which is achieved by comparing the language similarity between predicted " + }, + { + "bbox": [ + 131, + 425, + 482, + 522 + ], + "type": "inline_equation", + "content": "\\hat{l}_{t}" + }, + { + "bbox": [ + 131, + 425, + 482, + 522 + ], + "type": "text", + "content": " and those within " + }, + { + "bbox": [ + 131, + 425, + 482, + 522 + ], + "type": "inline_equation", + "content": "V_{a}" + }, + { + "bbox": [ + 131, + 425, + 482, + 522 + ], + "type": "text", + "content": ", enabling us to assign probabilities to each action within a plan sequence. During the training stage, we use the ground-truth action representation " + }, + { + "bbox": [ + 131, + 425, + 482, + 522 + ], + "type": "inline_equation", + "content": "l_{t}" + }, + { + "bbox": [ + 131, + 425, + 482, + 522 + ], + "type": "text", + "content": " as input, while we use the predicted step representation " + }, + { + "bbox": [ + 131, + 425, + 482, + 522 + ], + "type": "inline_equation", + "content": "\\hat{l}_{t}" + }, + { + "bbox": [ + 131, + 425, + 482, + 522 + ], + "type": "text", + "content": " during the test stage." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 533, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 533, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 131, + 533, + 482, + 616 + ], + "type": "text", + "content": "Retrieval Planner. To improve long-range sequence modeling, our model integrates a direct retrieval mechanism based on a training datastore, utilizing its state-action temporal dependency information, explicitly. This enhancement is achieved by augmenting the base planner in a manner similar to [7,10], by adding a retrieval-augmented memory module, which learns the state-action pairs in the training data as memory, by modeling states, as a context for the actions. Fig. 3 illustrated the workings of this component." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "content": "The memory module consists of key-value pairs " + }, + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{s_n:d_n\\}_{n = 1}^D" + }, + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "content": " is the memory size. The keys " + }, + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "(s_n)" + }, + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "content": ", are discrete learnable context vectors and the values " + }, + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "(d_{n})" + }, + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "content": " are their subsequent action step labels, conditioned on the context. We estimate a context vector " + }, + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\varsigma_t" + }, + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "content": ", given representations of the initial visual state" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "A. Zare et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "v_{s}^{\\prime}" + }, + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "content": ", the target state " + }, + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "v_{g}^{\\prime}" + }, + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "content": ", the current action " + }, + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "l_{t}" + }, + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "content": ", and a learnable positional embedding " + }, + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "q_{t}" + }, + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "content": " as in Eq. (3)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 231, + 141, + 480, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 141, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 231, + 141, + 480, + 152 + ], + "type": "interline_equation", + "content": "\\varsigma_ {t} = f _ {\\text {P r o j}} \\left(\\operatorname {C o n c a t} \\left(v _ {s}, v _ {g}, \\hat {c}, l _ {t} + q _ {t}\\right)\\right) \\tag {3}", + "image_path": "3f6febf2aceadfab2c9c5e36d4f97738a41bd533cb935df6aa3f09ad90eb7138.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 157, + 480, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 157, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 157, + 480, + 324 + ], + "type": "text", + "content": "With " + }, + { + "bbox": [ + 132, + 157, + 480, + 324 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{Proj}}" + }, + { + "bbox": [ + 132, + 157, + 480, + 324 + ], + "type": "text", + "content": " as a projection layer and " + }, + { + "bbox": [ + 132, + 157, + 480, + 324 + ], + "type": "inline_equation", + "content": "\\mathrm{Concat}(\\cdot)" + }, + { + "bbox": [ + 132, + 157, + 480, + 324 + ], + "type": "text", + "content": " concatenating inputs along the embedding dimension. In training, we learn the keys " + }, + { + "bbox": [ + 132, + 157, + 480, + 324 + ], + "type": "inline_equation", + "content": "(s_n)" + }, + { + "bbox": [ + 132, + 157, + 480, + 324 + ], + "type": "text", + "content": " by comparing them to inputs' context vectors " + }, + { + "bbox": [ + 132, + 157, + 480, + 324 + ], + "type": "inline_equation", + "content": "\\varsigma_t" + }, + { + "bbox": [ + 132, + 157, + 480, + 324 + ], + "type": "text", + "content": ". We construct the memory values " + }, + { + "bbox": [ + 132, + 157, + 480, + 324 + ], + "type": "inline_equation", + "content": "(d_n)" + }, + { + "bbox": [ + 132, + 157, + 480, + 324 + ], + "type": "text", + "content": " by uniformly sampling actions from the training dataset (including the END token), making sure of a balanced representation of actions. At inference, the model retrieves the " + }, + { + "bbox": [ + 132, + 157, + 480, + 324 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 132, + 157, + 480, + 324 + ], + "type": "text", + "content": " nearest (i.e., most similar) keys " + }, + { + "bbox": [ + 132, + 157, + 480, + 324 + ], + "type": "inline_equation", + "content": "\\{s_k\\}_{k=1}^K" + }, + { + "bbox": [ + 132, + 157, + 480, + 324 + ], + "type": "text", + "content": " from the memory to a given current context " + }, + { + "bbox": [ + 132, + 157, + 480, + 324 + ], + "type": "inline_equation", + "content": "\\varsigma_t" + }, + { + "bbox": [ + 132, + 157, + 480, + 324 + ], + "type": "text", + "content": ", returning their corresponding values " + }, + { + "bbox": [ + 132, + 157, + 480, + 324 + ], + "type": "inline_equation", + "content": "\\{d_k\\}_{k=1}^K" + }, + { + "bbox": [ + 132, + 157, + 480, + 324 + ], + "type": "text", + "content": ". The nearest neighbors " + }, + { + "bbox": [ + 132, + 157, + 480, + 324 + ], + "type": "inline_equation", + "content": "(\\{s_k\\}_{k=1}^K)" + }, + { + "bbox": [ + 132, + 157, + 480, + 324 + ], + "type": "text", + "content": " get selected by calculating the cosine distance between " + }, + { + "bbox": [ + 132, + 157, + 480, + 324 + ], + "type": "inline_equation", + "content": "\\varsigma_t" + }, + { + "bbox": [ + 132, + 157, + 480, + 324 + ], + "type": "text", + "content": " and the keys of the datastore " + }, + { + "bbox": [ + 132, + 157, + 480, + 324 + ], + "type": "inline_equation", + "content": "\\{s_n\\}_{n=1}^D" + }, + { + "bbox": [ + 132, + 157, + 480, + 324 + ], + "type": "text", + "content": ". Upon identifying the nearest neighbors, a probability distribution across the retrieved values " + }, + { + "bbox": [ + 132, + 157, + 480, + 324 + ], + "type": "inline_equation", + "content": "(\\{d_k\\}_{k=1}^K)" + }, + { + "bbox": [ + 132, + 157, + 480, + 324 + ], + "type": "text", + "content": " can be calculated by applying a softmax function to the similarity scores. The retrieval process can be formulated as " + }, + { + "bbox": [ + 132, + 157, + 480, + 324 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{retr}}(y|x) = P(a_{t+1}|\\varsigma_t, a_{t+1} \\in \\{d_k\\}_{k=1}^K)" + }, + { + "bbox": [ + 132, + 157, + 480, + 324 + ], + "type": "text", + "content": ". The final prediction ensembles the predictions from base planner and retrieval planner. Following [7, 10], we obtain the final probability distribution for the next action step as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 216, + 334, + 480, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 334, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 216, + 334, + 480, + 346 + ], + "type": "interline_equation", + "content": "P _ {\\text {a l l}} (y | x) = \\lambda P _ {\\text {r e t r}} (y | x) + (1 - \\lambda) P _ {\\text {b a s e}} (y | x) \\tag {4}", + "image_path": "aa60658e26d79f98bbe976c3e83df7c9251715d648bf951009bb7dca170521a0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 354, + 293, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 293, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 293, + 365 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 354, + 293, + 365 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 132, + 354, + 293, + 365 + ], + "type": "text", + "content": " is a tuned hyper-parameter." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 383, + 204, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 383, + 204, + 395 + ], + "spans": [ + { + "bbox": [ + 132, + 383, + 204, + 395 + ], + "type": "text", + "content": "3.3 Training" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 402, + 479, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 479, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 479, + 449 + ], + "type": "text", + "content": "RAP is trained in two stages: First, we train the base planner, then we augment and train the model including the retrieval planner. This approach ensures effective learning of the retrieval planner's keys, building upon the base planner's good initialization obtained from the initial training stage." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 450, + 480, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 521 + ], + "type": "text", + "content": "Stage 1: Training the Base Planner. During the training phase, we first supervise the base model's output, the textual representation of each action step in the procedure sequence, through language supervision combined with contrastive learning [8]: For each feature " + }, + { + "bbox": [ + 132, + 450, + 480, + 521 + ], + "type": "inline_equation", + "content": "\\hat{l}_t" + }, + { + "bbox": [ + 132, + 450, + 480, + 521 + ], + "type": "text", + "content": " generated by the transformer's head, we take its ground truth language embedding " + }, + { + "bbox": [ + 132, + 450, + 480, + 521 + ], + "type": "inline_equation", + "content": "l_t" + }, + { + "bbox": [ + 132, + 450, + 480, + 521 + ], + "type": "text", + "content": " as positive sample and all other embeddings in the action vocabulary " + }, + { + "bbox": [ + 132, + 450, + 480, + 521 + ], + "type": "inline_equation", + "content": "V_a" + }, + { + "bbox": [ + 132, + 450, + 480, + 521 + ], + "type": "text", + "content": " as negative examples. The loss is formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 240, + 531, + 480, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 531, + 480, + 562 + ], + "spans": [ + { + "bbox": [ + 240, + 531, + 480, + 562 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {b a s e}} = - \\sum_ {t = 1} ^ {T + 1} \\log \\frac {\\exp \\hat {l} _ {t} \\cdot l _ {t}}{\\sum_ {j} \\exp \\hat {l} _ {t} \\cdot l _ {j}} \\tag {5}", + "image_path": "15736fabb97c9da22c8df3a2531cf7d17a5be9db205944c4be654ad6329c7afc.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 570, + 329, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 570, + 329, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 570, + 329, + 582 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 132, + 570, + 329, + 582 + ], + "type": "inline_equation", + "content": "(T + 1)" + }, + { + "bbox": [ + 132, + 570, + 329, + 582 + ], + "type": "text", + "content": "-th step denotes the END token." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 594, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 480, + 665 + ], + "type": "text", + "content": "Stage 2: Training the Base and Retrieval Planners. In the second stage, the model is augmented with the nearest neighbor retrieval-based memory. The retrieval planner advances learning by juxtaposing the one-hot ground-truth labels " + }, + { + "bbox": [ + 132, + 594, + 480, + 665 + ], + "type": "inline_equation", + "content": "a_{t} \\in p" + }, + { + "bbox": [ + 132, + 594, + 480, + 665 + ], + "type": "text", + "content": ", with the model's interpolated probability estimates " + }, + { + "bbox": [ + 132, + 594, + 480, + 665 + ], + "type": "inline_equation", + "content": "P_{all,t}" + }, + { + "bbox": [ + 132, + 594, + 480, + 665 + ], + "type": "text", + "content": ", for each action at each time-step " + }, + { + "bbox": [ + 132, + 594, + 480, + 665 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 132, + 594, + 480, + 665 + ], + "type": "text", + "content": ". It refines the training by assessing how likely each action is, based on the model's predictions, instead of solely relying on matching" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "RAP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 479, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 479, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 479, + 129 + ], + "type": "text", + "content": "the correct action. We achieve this, using a cross-entropy minimization in Eq. (6)," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 245, + 137, + 481, + 169 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 137, + 481, + 169 + ], + "spans": [ + { + "bbox": [ + 245, + 137, + 481, + 169 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l l}} = - \\sum_ {t = 1} ^ {T + 1} a _ {t} \\log P _ {\\text {a l l , t}} (y | x) \\tag {6}", + "image_path": "e133c2c1aa8bd664cf5987fb8a6c0455f8bb0fddd88f0d4be3d10a02ccdd26be.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 174, + 329, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 174, + 329, + 186 + ], + "spans": [ + { + "bbox": [ + 132, + 174, + 329, + 186 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 174, + 329, + 186 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{all,t}}(y|x)" + }, + { + "bbox": [ + 132, + 174, + 329, + 186 + ], + "type": "text", + "content": " is defined based on Eq. (4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 444, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 444, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 444, + 216 + ], + "type": "text", + "content": "3.4 Addition of Unannotated Videos with Weak Supervision" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 223, + 481, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 223, + 481, + 294 + ], + "spans": [ + { + "bbox": [ + 130, + 223, + 481, + 294 + ], + "type": "text", + "content": "A significant challenge in procedure planning for instructional videos is the scarcity of annotated data. This is mainly because annotating instructional videos with step-level (i.e., timestamps) or sequence-level labels (i.e., action sequences) is resource-intensive. To address this issue, we propose a weakly-supervised training approach aimed at expanding the annotated training data to include additional unannotated task-related videos in the training process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 295, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 482, + 437 + ], + "type": "text", + "content": "Specifically, we align instructional videos with their corresponding topics, such as \"making an omelet\" or \"making pancakes\". This alignment links each video subset to a subset of instructional action plans sharing a similar topic. Crucially, this approach does not require detailed annotations that map specific action steps to individual video segments or frames. Instead, it relies on the broader thematic association between the video content and the instructional plans, allowing for a more general form of supervision during training. However, the acquired pairs of instructional videos and action plans may exhibit certain discrepancies. First, the presence of action steps in a plan that do not appear in the related video. Second, a mismatch in the sequence of action steps between the plan and their occurrence in the instructional video. Third, the inclusion of additional action steps in the video that are absent from the plan." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 438, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 605 + ], + "type": "text", + "content": "To tackle these challenges, we applied a video grounding algorithm, taking inspiration from Dvornik et al. [5] to generate pseudo-annotations of descriptions in instructional videos. In short, our grounding algorithm processes an instructional video alongside an ordered sequence of action steps to produce a sequence of grounding results (" + }, + { + "bbox": [ + 130, + 438, + 482, + 605 + ], + "type": "inline_equation", + "content": "t_{start}, t_{end}, a" + }, + { + "bbox": [ + 130, + 438, + 482, + 605 + ], + "type": "text", + "content": "), where " + }, + { + "bbox": [ + 130, + 438, + 482, + 605 + ], + "type": "inline_equation", + "content": "t_{start}" + }, + { + "bbox": [ + 130, + 438, + 482, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 438, + 482, + 605 + ], + "type": "inline_equation", + "content": "t_{end}" + }, + { + "bbox": [ + 130, + 438, + 482, + 605 + ], + "type": "text", + "content": " denote the beginning and end times of action step " + }, + { + "bbox": [ + 130, + 438, + 482, + 605 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 130, + 438, + 482, + 605 + ], + "type": "text", + "content": ". The grounding model offers flexible sequence alignment and the ability to exclude irrelevant actions or frames, effectively addressing observed discrepancies. Further details on the grounding model are provided in the appendix. We utilize these pseudo-annotations, extracted for each relevant video, which serve as structurally and format-wise equivalent to manually annotated videos. These pseudo-annotations are integrated into Stage 1 training alongside original annotations, where the model is tasked with predicting the complete sequence of actions based on the starting frame of the first action and the ending frame of the last action outlined in the plan." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "content": "This weak supervision of unannotated data, is an extension of Stage 1 training by using both original annotations of training videos and pseudo-annotations of task-relevant videos as the new training set. This training approach is a step towards a cost-effective solution for generalizing and scaling procedure planning to abundant unannotated videos." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "A. Zare et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 229, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 229, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 229, + 129 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 479, + 163 + ], + "type": "text", + "content": "We experiment on two benchmark datasets and employ four metrics to validate the efficacy of our proposed model." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 169, + 185, + 444, + 281 + ], + "blocks": [ + { + "bbox": [ + 169, + 185, + 444, + 281 + ], + "lines": [ + { + "bbox": [ + 169, + 185, + 444, + 281 + ], + "spans": [ + { + "bbox": [ + 169, + 185, + 444, + 281 + ], + "type": "image", + "image_path": "360de421978424c0e5446f04d3867e03a38349d712fd0ec021f793109550e5fc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 293, + 480, + 326 + ], + "lines": [ + { + "bbox": [ + 130, + 293, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 480, + 326 + ], + "type": "text", + "content": "Fig. 4: Illustration of the advantage of edit score, in evaluation of long varied-length sequences. As it can be seen, despite the prediction being plausible and acceptable, accuracy and success rate, falsely flag this scenario as a miss." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 368, + 204, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 368, + 204, + 378 + ], + "spans": [ + { + "bbox": [ + 132, + 368, + 204, + 378 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 387, + 482, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 590 + ], + "type": "text", + "content": "Our methodology is evaluated using two instructional video datasets: CrossTask [26] and COIN [20]. CrossTask contains 2750 videos spanning 83 different topics, such as \"How to make french toast?\", with 7.6 actions per video on average. These topics are categorized into 18 primary and 65 related events. The primary topics are detailed with precise timestamps for each action step, providing a clear sequence of instructional activities. In contrast, the related instructional topics, gathered through automatic collection methods, lack annotations for action-step timestamps, sequence order, and the specific subset of action steps each video covers. Each topic within this related set is associated with a designated instructional plan as well as a set of unannotated instructional videos of the same topic. These videos may exhibit variations from the outlined plan, as detailed in Sec. 3.3, due to the previously mentioned discrepancies. We use CrossTask's related-set data, during weakly-supervised training with language supervision. COIN, includes 11827 videos, covering 778 different actions, with an average action-horizon of 3.6 actions per video. We follow [4,24] by creating train/test splits with a " + }, + { + "bbox": [ + 130, + 387, + 482, + 590 + ], + "type": "inline_equation", + "content": "70\\% / 30\\%" + }, + { + "bbox": [ + 130, + 387, + 482, + 590 + ], + "type": "text", + "content": " ratio as well as implementing a moving window approach to organize the datasets into plans with various action-horizons." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 608, + 200, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 608, + 200, + 619 + ], + "spans": [ + { + "bbox": [ + 132, + 608, + 200, + 619 + ], + "type": "text", + "content": "4.2 Metrics" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 628, + 480, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 628, + 480, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 628, + 480, + 668 + ], + "type": "text", + "content": "Our methodology's performance is evaluated using three standard metrics [2, 4, 19, 24]. First, mean Intersection over Union (" + }, + { + "bbox": [ + 130, + 628, + 480, + 668 + ], + "type": "inline_equation", + "content": "mIoU" + }, + { + "bbox": [ + 130, + 628, + 480, + 668 + ], + "type": "text", + "content": ") assesses overlap between predicted and ground truth action sequences, calculated by " + }, + { + "bbox": [ + 130, + 628, + 480, + 668 + ], + "type": "inline_equation", + "content": "\\frac{|a_t \\cap \\hat{a}_t|}{|a_t \\cup \\hat{a}_t|}" + }, + { + "bbox": [ + 130, + 628, + 480, + 668 + ], + "type": "text", + "content": ". This metric" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "RAP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "evaluates whether the model identifies correct steps but does not consider action order or repetitions. Second, mean Accuracy (" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "mAcc" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": ") measures action alignment at each step, balancing the impact of repeated actions. Lastly, the Success Rate (" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "SR" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": ") evaluates a plan as successful if it exactly matches the ground truth." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 166, + 480, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 166, + 480, + 261 + ], + "spans": [ + { + "bbox": [ + 130, + 166, + 480, + 261 + ], + "type": "text", + "content": "However, Accuracy and Success Rate have limitations, particularly with long or variable-length sequences, often leading to false negatives due to their sensitivity to action order. For example, in a pancake-making plan, if the predicted sequence adds a plausible step like \"Add sugar\" at the beginning, accuracy may incorrectly deem it inaccurate. This issue is more pronounced with sequences that vary in length from the ground truth, indicating the need for a more nuanced evaluation metric in procedure planning that accommodates plausible variations in actions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 264, + 479, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 264, + 479, + 348 + ], + "spans": [ + { + "bbox": [ + 131, + 264, + 479, + 348 + ], + "type": "text", + "content": "To address the limitations of the current metrics, we introduce a new metric, mean edit-score " + }, + { + "bbox": [ + 131, + 264, + 479, + 348 + ], + "type": "inline_equation", + "content": "(mES)" + }, + { + "bbox": [ + 131, + 264, + 479, + 348 + ], + "type": "text", + "content": ", based on the minimum edit distance (e.g. Levenshtein edit distance), measuring the similarity between two sequences. This function, " + }, + { + "bbox": [ + 131, + 264, + 479, + 348 + ], + "type": "inline_equation", + "content": "L(s_{1},s_{2})|c_{sub},c_{ins},c_{del}" + }, + { + "bbox": [ + 131, + 264, + 479, + 348 + ], + "type": "text", + "content": ", calculates the minimum number of substitutions " + }, + { + "bbox": [ + 131, + 264, + 479, + 348 + ], + "type": "inline_equation", + "content": "(c_{sub})" + }, + { + "bbox": [ + 131, + 264, + 479, + 348 + ], + "type": "text", + "content": ", insertions " + }, + { + "bbox": [ + 131, + 264, + 479, + 348 + ], + "type": "inline_equation", + "content": "(c_{ins})" + }, + { + "bbox": [ + 131, + 264, + 479, + 348 + ], + "type": "text", + "content": ", and deletions " + }, + { + "bbox": [ + 131, + 264, + 479, + 348 + ], + "type": "inline_equation", + "content": "(c_{del})" + }, + { + "bbox": [ + 131, + 264, + 479, + 348 + ], + "type": "text", + "content": " needed to transform sequence " + }, + { + "bbox": [ + 131, + 264, + 479, + 348 + ], + "type": "inline_equation", + "content": "s_1" + }, + { + "bbox": [ + 131, + 264, + 479, + 348 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 131, + 264, + 479, + 348 + ], + "type": "inline_equation", + "content": "s_2" + }, + { + "bbox": [ + 131, + 264, + 479, + 348 + ], + "type": "text", + "content": ". The mean edit-score, is derived from the expected value of sample edit-scores. The edit-score is thus defined based on the optimal edit distance as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 200, + 365, + 480, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 365, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 200, + 365, + 480, + 392 + ], + "type": "interline_equation", + "content": "E S _ {s _ {1}, s _ {2}} = 1 0 0 \\cdot \\frac {\\operatorname* {m a x} \\left(\\left| s _ {1} \\right| , \\left| s _ {2} \\right|\\right) - L \\left(s _ {1} , s _ {2}\\right) | _ {1 , 1 , 1}}{\\operatorname* {m a x} \\left(\\left| s _ {1} \\right| , \\left| s _ {2} \\right|\\right)}, \\tag {7}", + "image_path": "57dccb8ae9b4cee2f837f727c57f980eef560b5e11d980e5eac30d9e7c9c9300.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 402, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 402, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 131, + 402, + 482, + 510 + ], + "type": "text", + "content": "Where the operator " + }, + { + "bbox": [ + 131, + 402, + 482, + 510 + ], + "type": "inline_equation", + "content": "|\\cdot|" + }, + { + "bbox": [ + 131, + 402, + 482, + 510 + ], + "type": "text", + "content": " measures the length of a sequence. The mean edit-score offers a more flexible and forgiving assessment compared to strict metrics like success rate and mean accuracy, while still maintaining sensitivity to the order of action steps, hence complimenting " + }, + { + "bbox": [ + 131, + 402, + 482, + 510 + ], + "type": "inline_equation", + "content": "mIoU" + }, + { + "bbox": [ + 131, + 402, + 482, + 510 + ], + "type": "text", + "content": " (i.e., if the ground-truth is add curry leaves, stir mixture, add chili powder and the prediction is add chili powder, stir mixture, add curry leaves, we face a high " + }, + { + "bbox": [ + 131, + 402, + 482, + 510 + ], + "type": "inline_equation", + "content": "IoU = 100\\%" + }, + { + "bbox": [ + 131, + 402, + 482, + 510 + ], + "type": "text", + "content": " and a low " + }, + { + "bbox": [ + 131, + 402, + 482, + 510 + ], + "type": "inline_equation", + "content": "ES = 33.33\\%" + }, + { + "bbox": [ + 131, + 402, + 482, + 510 + ], + "type": "text", + "content": "). This metric provides a more comprehensive evaluation, especially in scenarios where the predicted sequence is plausible yet varies in actions or length from the ground truth (Fig. 4)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 538, + 279, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 538, + 279, + 551 + ], + "spans": [ + { + "bbox": [ + 132, + 538, + 279, + 551 + ], + "type": "text", + "content": "4.3 Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": "Consistent with earlier approaches, we encode the start and goal visual observations using the S3D network [14] pre-trained on the HowTo100M [15] dataset. For optimization, we use the AdamW optimizer [13], setting the weight decay to 0.2. The learning rate is initially set at " + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "inline_equation", + "content": "lr = 0.001" + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": ", and we apply a reduce-on-plateau scheduler to adjust this rate based on performance metrics. We experiment with an autoregressive transformer with 7 layers and 8 heads. We choose weight " + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\lambda = 0.1" + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": " and volume " + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "inline_equation", + "content": "D = 1000" + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": ", for the memory module. Further implementation details can be found in the supplementary section." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "text", + "content": "A. Zare et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 149, + 157, + 465, + 312 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Assessment of procedure planning on CrossTask for action-horizons " + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "T = 3" + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "T = 4" + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": ". Vision supervision is denoted by V, language supervision by L, and weakly-supervised training by +wL." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 157, + 465, + 312 + ], + "lines": [ + { + "bbox": [ + 149, + 157, + 465, + 312 + ], + "spans": [ + { + "bbox": [ + 149, + 157, + 465, + 312 + ], + "type": "table", + "html": "
ModelsSupervisionT=3T=4
SR↑mAcc↑mIoU↑SR↑mAcc↑mIoU↑
Random-<0.010.941.66<0.010.831.66
Retrieval-BasedV8.0523.3032.063.9522.2236.97
WLTDO [6]V1.8721.6431.700.7717.9226.43
UAAA [1]V2.1520.2130.870.9819.8627.09
UPN [18]V2.8924.3931.561.1921.5927.85
DDN [4]V12.1831.2947.485.9727.1048.46
Ext-GAIL w/o Aug. [2]V18.0143.8657.16---
Ext-GAIL [2]V21.2749.4661.7016.4143.0560.93
P3IV w/o Adv. [24]L22.1245.5767.40---
P3IV [24]L23.3449.9673.8913.4044.1670.01
E3P (baseline) [22]L22.5646.17-12.9743.97-
E3P [22]L26.4053.0274.0516.4948.0070.16
Skip-Plan [12]L28.8561.1874.9815.5655.6470.30
RAP (ours w/o unannotated data)L29.2362.3575.6316.9655.7771.49
RAP (ours with unannotated data)L+wL30.5162.6875.9217.1055.9171.33
", + "image_path": "e1b991b92279f56a8f9111e9201d2e6942cf41099811bc46849354a490972e63.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 176, + 353, + 438, + 385 + ], + "blocks": [ + { + "bbox": [ + 130, + 321, + 482, + 343 + ], + "lines": [ + { + "bbox": [ + 130, + 321, + 482, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 482, + 343 + ], + "type": "text", + "content": "Table 2: Comparison of RAP with PDPP, when we align our state definition with PDPP's for CrossTask and " + }, + { + "bbox": [ + 130, + 321, + 482, + 343 + ], + "type": "inline_equation", + "content": "T = 3" + }, + { + "bbox": [ + 130, + 321, + 482, + 343 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 176, + 353, + 438, + 385 + ], + "lines": [ + { + "bbox": [ + 176, + 353, + 438, + 385 + ], + "spans": [ + { + "bbox": [ + 176, + 353, + 438, + 385 + ], + "type": "table", + "html": "
ModelsSupervisionSR ↑mAcc ↑mIoU ↑
PDPP [23]♣L37.2064.6766.57
RAP (ours w/o unannotated data)♣L38.5165.1877.44
", + "image_path": "bc98c27ba8dc3c061c4e734d5e5834d2bce31cd0e541ee9164dcae33544a165b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 407, + 386, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 407, + 386, + 419 + ], + "spans": [ + { + "bbox": [ + 131, + 407, + 386, + 419 + ], + "type": "text", + "content": "4.4 Comparison with State-of-The-Art Baselines" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "text", + "content": "Our study conducts a comprehensive comparison of our new framework, RAP, against established models in procedure planning, using CrossTask and COIN datasets. We refer to our framework without the retrieval-based memory augmentation, i.e., base planner, as BP (as opposed to RAP). We categorize previous approaches based on their supervision: language (L) or vision (V). Our proposed model's performance compared to the previous works, is detailed in Tabs. 1 and 2 for CrossTask, and Tab. 3 for COIN." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "text", + "content": "The inherent uncertainty in variable action-horizon procedure planning models complicates direct and fair comparisons with traditional fixed-length models, due to the added complexity of predicting the conclusion of a sequence as well as the broader range of acceptable sequences of variable length other than the ground-truth. To ensure a fair evaluation, we compare our variable action-horizon architecture, trained and evaluated exclusively on fixed-horizon data (of length " + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "text", + "content": "), to the fixed-horizon models (of length " + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "text", + "content": ") from prior studies in procedure planning [1,2,4,6,12,18,22-24]. Tab. 1 further shows the performance boost achieved by the addition of CrossTask's unannotated data (i.e., CrossTask's related video set) through weakly-supervised training (supervision: " + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\mathrm{L} + \\mathrm{wL}" + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Our findings demonstrate that RAP exhibits superior performance over existing vision-supervised models. Notably, RAP, without the addition of unannotated data, surpasses Ext-GAIL [2] with a " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "7.97\\%" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": " increase in Success Rate (SR)," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "RAP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 176, + 146, + 440, + 242 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 479, + 136 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 479, + 136 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 479, + 136 + ], + "type": "text", + "content": "Table 3: Comparison of the procedure planing models' performance on the COIN dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 176, + 146, + 440, + 242 + ], + "lines": [ + { + "bbox": [ + 176, + 146, + 440, + 242 + ], + "spans": [ + { + "bbox": [ + 176, + 146, + 440, + 242 + ], + "type": "table", + "html": "
ModelsT=3T=4
SR↑mAcc↑mIoU↑SR↑mAcc↑mIoU↑
Random (V)<0.01<0.012.47<0.01<0.012.32
Retrieval-Based (V)4.3817.4032.062.7114.2936.97
DDN (V)13.9020.1964.7811.1317.7168.06
P3IV (L)15.4021.6776.3111.3218.8570.53
E3P (L)19.5731.4284.9513.5926.7284.72
PDPP (L)21.3345.6251.8214.4144.1051.39
Skip-plan (L)23.6547.1278.4416.0443.1977.07
RAP (w/o unannotated) (L)24.4747.5985.2416.7642.1184.64
", + "image_path": "47b60a9113d87110c4b4457d1a711474de8d1b55830c21a9aaefb6f899e4b26d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 148, + 284, + 463, + 342 + ], + "blocks": [ + { + "bbox": [ + 130, + 251, + 479, + 275 + ], + "lines": [ + { + "bbox": [ + 130, + 251, + 479, + 275 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 479, + 275 + ], + "type": "text", + "content": "Table 4: Evaluation of RAP versus a retrieval-based (RB) model in variable-length sequence prediction on CrossTask and COIN." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 148, + 284, + 463, + 342 + ], + "lines": [ + { + "bbox": [ + 148, + 284, + 463, + 342 + ], + "spans": [ + { + "bbox": [ + 148, + 284, + 463, + 342 + ], + "type": "table", + "html": "
ModelsCrossTaskCOIN
SR↑mAcc↑mIoU↑mES↑SR↑mAcc↑mIoU↑mES↑
Random0.020.561.521.210.020.090.180.17
RB8.0121.3238.8033.437.1112.9420.1717.69
RAP (w/o unannotated data)20.2849.4169.9562.7919.1433.2675.8257.48
RAP (with unannotated data)21.7950.5871.4463.86----
", + "image_path": "35951e494cb90f987016867ac3f959cd9a19989bfb4fa9fe932c20bfe23d85c2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 365, + 479, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 479, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 479, + 389 + ], + "type": "text", + "content": "12.89% in Mean Accuracy (mAcc), and 13.93% in Mean Intersection over Union (mIoU) for sequence lengths of " + }, + { + "bbox": [ + 130, + 365, + 479, + 389 + ], + "type": "inline_equation", + "content": "T = 3" + }, + { + "bbox": [ + 130, + 365, + 479, + 389 + ], + "type": "text", + "content": " in CrossTask." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 389, + 481, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 481, + 555 + ], + "type": "text", + "content": "When compared to language-supervised models like P3IV, E3P, and SkipPlan, RAP demonstrates superior performance on CrossTask. For " + }, + { + "bbox": [ + 130, + 389, + 481, + 555 + ], + "type": "inline_equation", + "content": "\\mathrm{T} = 3" + }, + { + "bbox": [ + 130, + 389, + 481, + 555 + ], + "type": "text", + "content": ", RAP exhibits improvements of " + }, + { + "bbox": [ + 130, + 389, + 481, + 555 + ], + "type": "inline_equation", + "content": "5.99\\%" + }, + { + "bbox": [ + 130, + 389, + 481, + 555 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 389, + 481, + 555 + ], + "type": "inline_equation", + "content": "2.83\\%" + }, + { + "bbox": [ + 130, + 389, + 481, + 555 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 389, + 481, + 555 + ], + "type": "inline_equation", + "content": "0.38\\%" + }, + { + "bbox": [ + 130, + 389, + 481, + 555 + ], + "type": "text", + "content": " respectively, and for " + }, + { + "bbox": [ + 130, + 389, + 481, + 555 + ], + "type": "inline_equation", + "content": "\\mathrm{T} = 4" + }, + { + "bbox": [ + 130, + 389, + 481, + 555 + ], + "type": "text", + "content": ", it shows enhancements of " + }, + { + "bbox": [ + 130, + 389, + 481, + 555 + ], + "type": "inline_equation", + "content": "3.56\\%" + }, + { + "bbox": [ + 130, + 389, + 481, + 555 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 389, + 481, + 555 + ], + "type": "inline_equation", + "content": "0.47\\%" + }, + { + "bbox": [ + 130, + 389, + 481, + 555 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 389, + 481, + 555 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 130, + 389, + 481, + 555 + ], + "type": "text", + "content": " in SR. Tab. 2 compares our model with PDPP [23], which employs a different procedure state definition. For fairness, we evaluate our model under the PDPP setting, marked with " + }, + { + "bbox": [ + 130, + 389, + 481, + 555 + ], + "type": "inline_equation", + "content": "\\clubsuit" + }, + { + "bbox": [ + 130, + 389, + 481, + 555 + ], + "type": "text", + "content": ", showcasing its superior performance. PDPP uses a 3-second window following the onset and preceding the offset of action steps to capture visual states, giving it an advantage in accessing action-related information compared to previous models. Traditionally, a 1-second window centered around the start and end of actions is used. By aligning our state definition with PDPP's, our RAP model outperforms PDPP on all metrics, as shown in the table. We observe a similar superiority in performance, when evaluating the model on the COIN dataset as shown in Tab. 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 640 + ], + "type": "text", + "content": "We assessed the efficacy of RAP in handling tasks with variable lengths by conducting experiments against a retrieval-based model, which operates under a similar variable action-horizon setting. The results of this comparison are presented in Tab. 4. As demonstrated in the table, RAP proves a remarkable superiority over the retrieval-based model across all evaluated metrics. This is particularly evident in the mES metric, where RAP (w/o unannotated data) shows a significant improvement by " + }, + { + "bbox": [ + 130, + 557, + 481, + 640 + ], + "type": "inline_equation", + "content": "29.36\\%" + }, + { + "bbox": [ + 130, + 557, + 481, + 640 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "On CrossTask, RAP is further trained using the weak language supervision framework outlined in Sec. 3.3, utilizing the unannotated data from the dataset's" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "A. Zare et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 146, + 481, + 196 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 5: A study on the effect of retrieval augmentation (BP vs RAP), for variable and fixed action-horizon settings." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 146, + 481, + 196 + ], + "lines": [ + { + "bbox": [ + 133, + 146, + 481, + 196 + ], + "spans": [ + { + "bbox": [ + 133, + 146, + 481, + 196 + ], + "type": "table", + "html": "
DatasetModelsT=3T=4Variable Length T
SR↑mAcc↑mIoU↑mES↑SR↑mAcc↑mIoU↑mES↑SR↑mAcc↑mIoU↑mES↑
CrossTask (w/o unannotated)BP25.4258.7073.9165.5915.2352.1670.5361.6318.5447.8569.1662.64
RAP29.2362.3575.6367.3316.9655.7771.4964.5120.2849.4169.9562.79
COINBP23.3846.2180.4962.2814.2941.6778.6257.2617.8631.3472.4456.24
RAP24.4747.5985.2463.1516.7642.1184.6460.4719.1433.2675.8257.48
", + "image_path": "684b24b83b27f9996977c7231ba3e2778b0aa6cd4cf3c3386136ecfc790bbd22.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 153, + 209, + 460, + 305 + ], + "blocks": [ + { + "bbox": [ + 153, + 209, + 460, + 305 + ], + "lines": [ + { + "bbox": [ + 153, + 209, + 460, + 305 + ], + "spans": [ + { + "bbox": [ + 153, + 209, + 460, + 305 + ], + "type": "image", + "image_path": "e9fb284d490254a8c076638e4ba694c576929103779744433c78025910d4458e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 316, + 480, + 338 + ], + "lines": [ + { + "bbox": [ + 130, + 316, + 480, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 480, + 338 + ], + "type": "text", + "content": "Fig. 5: Sample plausible predicted plan sequences generated by RAP vs the ground-truth (GT)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 362, + 480, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 362, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 362, + 480, + 397 + ], + "type": "text", + "content": "related-tasks. As shown in Tab. 1 and Tab. 4 when training RAP with the addition of the unannotated data, we see significant improvement in performance across all metrics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 399, + 480, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 480, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 480, + 470 + ], + "type": "text", + "content": "Fig. 5 displays two qualitative results of RAP's performance in predicting plans of varying lengths compared to the ground truths. In the top row, it accurately identifies the addition of fish, other ingredients, and the plausible action \"add onions\". In the bottom row, it recognizes the completed bread, its removal, and suggests the reasonable action \"melt butter\", before the new bread is put in the pan, showcasing the model's planning abilities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 489, + 238, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 489, + 238, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 489, + 238, + 502 + ], + "type": "text", + "content": "4.5 Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 481, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 481, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 481, + 617 + ], + "type": "text", + "content": "Impact of retrieval augmentation. The introduction of a retrieval-based memory component significantly enhances our framework's performance. Tab. 5 demonstrates improvements for both fixed and variable action-horizon settings. Notably, enhancements include increases of " + }, + { + "bbox": [ + 130, + 510, + 481, + 617 + ], + "type": "inline_equation", + "content": "3.81\\%" + }, + { + "bbox": [ + 130, + 510, + 481, + 617 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 510, + 481, + 617 + ], + "type": "inline_equation", + "content": "1.73\\%" + }, + { + "bbox": [ + 130, + 510, + 481, + 617 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 510, + 481, + 617 + ], + "type": "inline_equation", + "content": "1.74\\%" + }, + { + "bbox": [ + 130, + 510, + 481, + 617 + ], + "type": "text", + "content": " in Success Rate (SR) for " + }, + { + "bbox": [ + 130, + 510, + 481, + 617 + ], + "type": "inline_equation", + "content": "\\mathrm{T} = 3" + }, + { + "bbox": [ + 130, + 510, + 481, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 510, + 481, + 617 + ], + "type": "inline_equation", + "content": "\\mathrm{T} = 4" + }, + { + "bbox": [ + 130, + 510, + 481, + 617 + ], + "type": "text", + "content": " in the fixed setting, and in the variable setting, respectively, on CrossTask. We observe a similar boost in performance for the COIN dataset, as evidenced in the Tab. 5. Tab. 6 details the enhancements in SR, categorized by the ground-truth (GT) sequences' action-horizon in the variable-length sequence prediction setting." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Impact of weak language supervision. The influence of weak language supervision on our model's performance is evaluated through an ablation study on CrossTask, where we incrementally increased the unannotated data used in weakly-supervised training. Tab. 7 displays the varying percentages of unan-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "RAP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 193, + 145, + 422, + 190 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 138 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 138 + ], + "type": "text", + "content": "Table 6: Impact of retrieval augmentation (BP vs RAP) in " + }, + { + "bbox": [ + 132, + 114, + 482, + 138 + ], + "type": "inline_equation", + "content": "SR" + }, + { + "bbox": [ + 132, + 114, + 482, + 138 + ], + "type": "text", + "content": ", on CrossTask, categorized by ground-truth (GT) action-horizon in the variable action-horizon setting." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 193, + 145, + 422, + 190 + ], + "lines": [ + { + "bbox": [ + 193, + 145, + 422, + 190 + ], + "spans": [ + { + "bbox": [ + 193, + 145, + 422, + 190 + ], + "type": "table", + "html": "
Model (w/o unannotated data)GT action-horizon
123456
BP49.5827.8416.218.375.893.34
RAP52.6029.2018.0210.526.165.18
", + "image_path": "239677a551f172f3f40b1e4d5fafc07fe81b870a73312def807b9a6550175a22.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 201, + 231, + 413, + 310 + ], + "blocks": [ + { + "bbox": [ + 132, + 199, + 481, + 222 + ], + "lines": [ + { + "bbox": [ + 132, + 199, + 481, + 222 + ], + "spans": [ + { + "bbox": [ + 132, + 199, + 481, + 222 + ], + "type": "text", + "content": "Table 7: Weak language supervision impact: Table shows RAP's performance gains on CrossTask with increasing unannotated data use." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 201, + 231, + 413, + 310 + ], + "lines": [ + { + "bbox": [ + 201, + 231, + 413, + 310 + ], + "spans": [ + { + "bbox": [ + 201, + 231, + 413, + 310 + ], + "type": "table", + "html": "
Used unannotated data portionSR↑mAcc↑mIoU↑mES↑
No unannotated data20.2849.4169.9562.79
25%19.2047.8869.0561.96
50%20.5749.7269.7563.12
75%21.3150.2270.8663.40
100%21.7950.5871.4463.86
", + "image_path": "0599516968a28c5c46aaa7e7d74ae1b37ba89bfde0a0f244cd0e06d9387d8471.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 331, + 482, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 404 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 404 + ], + "type": "text", + "content": "notated data utilized. The results show consistent performance gains for RAP under weak language supervision, particularly as more data is introduced (except at the " + }, + { + "bbox": [ + 130, + 331, + 482, + 404 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 130, + 331, + 482, + 404 + ], + "type": "text", + "content": " data level). This suggests potential for further enhancements with increased unannotated data. Notably, using " + }, + { + "bbox": [ + 130, + 331, + 482, + 404 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 331, + 482, + 404 + ], + "type": "text", + "content": " of unannotated data yields improvements in SR, mAcc, mIoU, and mES by " + }, + { + "bbox": [ + 130, + 331, + 482, + 404 + ], + "type": "inline_equation", + "content": "1.51\\%" + }, + { + "bbox": [ + 130, + 331, + 482, + 404 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 331, + 482, + 404 + ], + "type": "inline_equation", + "content": "1.17\\%" + }, + { + "bbox": [ + 130, + 331, + 482, + 404 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 331, + 482, + 404 + ], + "type": "inline_equation", + "content": "1.49\\%" + }, + { + "bbox": [ + 130, + 331, + 482, + 404 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 331, + 482, + 404 + ], + "type": "inline_equation", + "content": "1.07\\%" + }, + { + "bbox": [ + 130, + 331, + 482, + 404 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 422, + 220, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 422, + 220, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 422, + 220, + 435 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 448, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 628 + ], + "type": "text", + "content": "In this study, we addressed the challenge of procedure planning in instructional videos by developing the Retrieval-Augmented Planner (RAP) model. This model overcomes the limitations of previous approaches that were restricted to fixed action-step predictions. RAP excels in both fixed and varied action horizon scenarios, outperforming existing models. A key innovation is our weak language supervision method, which uses thematic links between unannotated videos and plans for pseudo-annotation generation, enhancing model training. Our model also incorporates a learnable retrieval-based memory component for improved long-range sequence predictions and efficient use of temporal data. Additionally, we introduced \"edit-score\", a new metric for assessing variable-length procedures. Looking forward, we see potential in applying RAP to more diverse scenarios and in generating probabilistic, varied-length plans. We also envisage that expanding our weak supervision approach to include an even wider range of unannotated data could markedly elevate the performance of procedure planning models, paving the way for more advanced applications in this field." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "A. Zare et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "type": "text", + "content": "1. Abu Farha, Y., Gall, J.: Uncertainty-aware anticipation of activities. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. pp. 0-0 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 170, + 480, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 170, + 480, + 201 + ], + "spans": [ + { + "bbox": [ + 138, + 170, + 480, + 201 + ], + "type": "text", + "content": "2. Bi, J., Luo, J., Xu, C.: Procedure planning in instructional videos via contextual modeling and model-based policy learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15611-15620 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 202, + 480, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 202, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 202, + 480, + 245 + ], + "type": "text", + "content": "3. Borgeaud, S., Mensch, A., Hoffmann, J., Cai, T., Rutherford, E., Millican, K., Van Den Driessche, G.B., Lespiau, J.B., Damoc, B., Clark, A., et al.: Improving language models by retrieving from trillions of tokens. In: International conference on machine learning. pp. 2206-2240. PMLR (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 245, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 480, + 277 + ], + "type": "text", + "content": "4. Chang, C.Y., Huang, D.A., Xu, D., Adeli, E., Fei-Fei, L., Niebles, J.C.: Procedure planning in instructional videos. In: European Conference on Computer Vision. pp. 334-350. Springer (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 277, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 480, + 310 + ], + "type": "text", + "content": "5. Dvornik, M., Hadji, I., Derpanis, K.G., Garg, A., Jepson, A.: Drop-dtw: Aligning common signal between sequences while dropping outliers. Advances in Neural Information Processing Systems 34, 13782-13793 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 310, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 310, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 310, + 480, + 342 + ], + "type": "text", + "content": "6. Ehsani, K., Bagherinezhad, H., Redmon, J., Mottaghi, R., Farhadi, A.: Who let the dogs out? modeling dog behavior from visual data. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4051-4060 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 342, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 480, + 373 + ], + "type": "text", + "content": "7. Grave, E., Cisse, M.M., Joulin, A.: Unbounded cache model for online language modeling with open vocabulary. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 374, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 374, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 374, + 480, + 418 + ], + "type": "text", + "content": "8. Gutmann, M., Hyvarinen, A.: Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In: Proceedings of the thirteenth international conference on artificial intelligence and statistics. pp. 297-304. JMLR Workshop and Conference Proceedings (2010)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 418, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 480, + 449 + ], + "type": "text", + "content": "9. Hadji, I., Derpanis, K.G., Jepson, A.D.: Representation learning via global temporal alignment and cycle-consistency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11068-11077 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 449, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 480, + 482 + ], + "type": "text", + "content": "0. Khandelwal, U., Levy, O., Jurafsky, D., Zettlemoyer, L., Lewis, M.: Generalization through memorization: Nearest neighbor language models. arXiv preprint arXiv:1911.00172 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 482, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 482, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 138, + 482, + 480, + 514 + ], + "type": "text", + "content": "1. Lei, J., Wang, L., Shen, Y., Yu, D., Berg, T.L., Bansal, M.: Mart: Memory-augmented recurrent transformer for coherent video paragraph captioning. arXiv preprint arXiv:2005.05402 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 514, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 480, + 557 + ], + "type": "text", + "content": "2. Li, Z., Geng, W., Li, M., Chen, L., Tang, Y., Lu, J., Zhou, J.: Skip-plan: Procedure planning in instructional videos via condensed action space learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10297-10306 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 558, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 558, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 558, + 480, + 578 + ], + "type": "text", + "content": "3. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 578, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 622 + ], + "type": "text", + "content": "4. Miech, A., Alayrac, J.B., Smaira, L., Laptev, I., Sivic, J., Zisserman, A.: End-to-end learning of visual representations from uncurated instructional videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9879-9889 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "type": "text", + "content": "5. Miech, A., Zhukov, D., Alayrac, J.B., Tapaswi, M., Laptev, I., Sivic, J.: Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2630-2640 (2019)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "RAP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 490 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "16. Niu, Y., Guo, W., Chen, L., Lin, X., Chang, S.F.: Schema: State changes matter for procedure planning in instructional videos. arXiv preprint arXiv:2403.01599 (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "text", + "content": "17. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "text", + "content": "18. Srinivas, A., Jabri, A., Abbeel, P., Levine, S., Finn, C.: Universal planning networks: Learning generalizable representations for visuomotor control. In: International Conference on Machine Learning. pp. 4732-4741. PMLR (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "19. Sun, J., Huang, D.A., Lu, B., Liu, Y.H., Zhou, B., Garg, A.: Plate: Visually-grounded planning with transformers in procedural tasks. IEEE Robotics and Automation Letters 7(2), 4924-4930 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 282 + ], + "type": "text", + "content": "20. Tang, Y., Ding, D., Rao, Y., Zheng, Y., Zhang, D., Zhao, L., Lu, J., Zhou, J.: Coin: A large-scale dataset for comprehensive instructional video analysis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1207-1216 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "type": "text", + "content": "21. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "type": "text", + "content": "22. Wang, A.L., Lin, K.Y., Du, J.R., Meng, J., Zheng, W.S.: Event-guided procedure planning from instructional videos with text supervision. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13565-13575 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "23. Wang, H., Wu, Y., Guo, S., Wang, L.: Pdpp: Projected diffusion for procedure planning in instructional videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14836-14845 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 424 + ], + "type": "text", + "content": "24. Zhao, H., Hadji, I., Dvornik, N., Derpanis, K.G., Wildes, R.P., Jepson, A.D.: P3iv: Probabilistic procedure planning from instructional videos with weak supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2938-2948 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 424, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 445 + ], + "type": "text", + "content": "25. Zhong, Z., Lei, T., Chen, D.: Training language models with memory augmentation (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 446, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 490 + ], + "type": "text", + "content": "26. Zhukov, D., Alayrac, J.B., Cinbis, R.G., Fouhey, D., Laptev, I., Sivic, J.: Cross-task weakly supervised learning from instructional videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3537-3545 (2019)" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "text", + "content": "A. Zare et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/RAPiD-Seg_ Range-Aware Pointwise Distance Distribution Networks for 3D LiDAR Segmentation/b19a59c8-eca1-4aae-977c-6a71eef1622b_content_list.json b/2024/RAPiD-Seg_ Range-Aware Pointwise Distance Distribution Networks for 3D LiDAR Segmentation/b19a59c8-eca1-4aae-977c-6a71eef1622b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3fa043c615648d845baf31c6769b95036deb435d --- /dev/null +++ b/2024/RAPiD-Seg_ Range-Aware Pointwise Distance Distribution Networks for 3D LiDAR Segmentation/b19a59c8-eca1-4aae-977c-6a71eef1622b_content_list.json @@ -0,0 +1,2000 @@ +[ + { + "type": "text", + "text": "RAPiD-Seg: Range-Aware Pointwise Distance Distribution Networks for 3D LiDAR Segmentation", + "text_level": 1, + "bbox": [ + 233, + 141, + 767, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Li Li $^{1}$ , Hubert P. H. Shum $^{1}$ , and Toby P. Breckon $^{1,2}$", + "bbox": [ + 292, + 210, + 707, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Department of $\\{\\mathrm{Computer~Science}^1\\mid$ Engineering2}, Durham University, UK {li.li4, hubert.shum, toby.breckon}@durham.ac.uk", + "bbox": [ + 243, + 239, + 756, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. 3D point clouds play a pivotal role in outdoor scene perception, especially in the context of autonomous driving. Recent advancements in 3D LiDAR segmentation often focus intensely on the spatial positioning and distribution of points for accurate segmentation. However, these methods, while robust in variable conditions, encounter challenges due to sole reliance on coordinates and point intensity, leading to poor isometric invariance and suboptimal segmentation. To tackle this challenge, our work introduces Range-Aware Pointwise Distance Distribution (RAPiD) features and the associated RAPiD-Seg architecture. Our RAPiD features exhibit rigid transformation invariance and effectively adapt to variations in point density, with a design focus on capturing the localized geometry of neighboring structures. They utilize inherent LiDAR isotropic radiation and semantic categorization for enhanced local representation and computational efficiency, while incorporating a 4D distance metric that integrates geometric and surface material reflectivity for improved semantic segmentation. To effectively embed high-dimensional RAPiD features, we propose a double-nested autoencoder structure with a novel class-aware embedding objective to encode high-dimensional features into manageable voxel-wise embeddings. Additionally, we propose RAPiD-Seg which incorporates a channel-wise attention fusion and two effective RAPiD-Seg variants, further optimizing the embedding for enhanced performance and generalization. Our method outperforms contemporary LiDAR segmentation work in terms of mIoU on SemanticKITTI (76.1) and nuScenes (83.6) datasets.", + "bbox": [ + 261, + 277, + 743, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Semantic Segmentation $\\cdot$ Point Cloud $\\cdot$ Invariance Features", + "bbox": [ + 261, + 622, + 740, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 645, + 375, + 660 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "LiDAR-based point clouds, characterized by pointwise 3D positions and LiDAR intensity/reflectivity [3,11,29,48], play a pivotal role in outdoor scene understanding, particularly in perception systems for autonomous driving. Recent literature in the domain of 3D LiDAR segmentation has witnessed significant strides, with myriad approaches attempting pointwise 3D semantic scene labeling [47]. Leveraging color [7,37,38], range imagery [22,34], and Birds Eye View (BEV) projections [34,60], multi-modal methods integrate diverse data streams from LiDAR and other sensors to enhance feature representation and segmentation accuracy. Single-modal (LiDAR-only) methods [17,30,56,64] effectively exploit spatial data interpretation and avoid multi-modal issues, e.g., modality heterogeneity and limited sensor field of view intersection [5].", + "bbox": [ + 212, + 672, + 789, + 843 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7ae4750af076c4f446545f4f228ce856c0601e8017fad9cd304cbe0044e475cf.jpg", + "image_caption": [ + "Fig. 1: Left: RAPiD exhibits excellent viewpoint invariance and geometric stability, visualizing comparable features around the vehicle door structure at varying ranges and viewpoints (feature matrix plots inset). Middle: RAPiD is distinctive in different semantic classes, as visualized by the matrices. Embedded RAPiD patterns corresponding to different points are visualized using a spectrum of colors, showcasing their capacity to represent different classes. Right: Our RAPiD-Seg achieves superior results over SoTA methods on nuScenes [3] and SemanticKITTI [2]." + ], + "image_footnote": [], + "bbox": [ + 217, + 171, + 401, + 263 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7d83ae89b7fde62c769474d221a118e7062d275753d69257cdd064fce416cb6f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 170, + 642, + 263 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/18dd574e87204f7f2dc16320e42d5696fc0be013651de4211c47c9508c79309c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 171, + 787, + 262 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite these advancements, a common challenge inherent to these methods is their poor isometric invariance, characterized by heavy reliance on solely coordinates and intensity data, which often results in suboptimal segmentation outcomes [30,64]; this is primarily due to poor translational invariance and visibility (e.g., occlusions), or sparse observations (e.g., at long range) [26], affecting data spatial distribution.", + "bbox": [ + 212, + 383, + 787, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we seek features that are (1) capable of capturing the localized geometric structure of neighboring points, (2) invariant to rotation and translation, and (3) applicable in noisy LiDAR outdoor environments. While numerous methods fulfill some of these requirements individually [21, 32, 40], they fall short of addressing them all [10]. Recognizing the need for higher-level features capable of capturing local geometry while potentially incorporating LiDAR point-specific attributes (e.g., intensity and reflectivity), we instead turn our attention to pointwise Distance Distribution (PDD) features [52, 53]. They are recognized for their remarkable efficacy in providing a robust and informative geometric representation of point clouds, excelling in both rotational and translational invariance while including intricate local geometry details.", + "bbox": [ + 212, + 462, + 787, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, directly employing PDD features is impractical due to their high-dimensional nature, and their significant memory and storage requirements for large-scale point clouds. PDD also overlooks local features, as including distant points subsequently dilutes the focus on immediate neighborhoods.", + "bbox": [ + 212, + 628, + 787, + 689 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As an enabler, we propose the Range-Aware Pointwise Distance Distribution (RAPiD-Seg) solution for LiDAR segmentation. It utilizes invariance of RAPiD features to rigid transformations and point cloud sparsity variations, while concentrating on compact features within specific local neighborhoods. Specifically, our method leverages inherent LiDAR isotropic radiation and semantic categorization for enhanced local representations while reducing computational burdens. Moreover, our formulation computes a 4D distance, incorporating both 3D geometric and reflectivity differences to enhance semantic segmentation fidelity. To compress high-dimensional RAPiD features into tractable voxel-wise embeddings, we propose a novel embedding approach with our class-aware double-nested AutoEn", + "bbox": [ + 212, + 689, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "L. Li et al.", + "bbox": [ + 271, + 114, + 344, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "coder (AE) module. We further incorporate a channel-wise attention fusion and a two-stage training strategy to optimize the AE independently before integrating into the full network, enhancing performance and generalization capability.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We conduct extensive experiments on SemanticKITTI [2] and nuScenes [3] datasets, upon which our approach surpasses the state-of-the-art (SoTA) segmentation performance.", + "bbox": [ + 212, + 191, + 784, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall, as shown in Fig. 2, our contributions can be summarized as follows:", + "bbox": [ + 238, + 237, + 785, + 252 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "A novel Range-Aware Pointwise Distance Distribution feature (RAPiD) for 3D LiDAR segmentation that ensures robustness to rigid transformations and viewpoints through isometry-invariant metrics within specific Regions of Interest, i.e., intra-ring (R-RAPiD) and intra-class (C-RAPiD).", + "A novel method for embedding RAPiD with class-aware double nested AE (RAPiD AE) to optimize the embedding of high-dimensional features, balancing efficiency and fidelity.", + "A novel open-source network architecture RAPiD-Seg $^{1}$ and supporting training methodology, to enable modular LiDAR segmentation that achieves SoTA performance on SemanticKITTI (mIoU: 76.1) and nuScenes (mIoU: 83.6)." + ], + "bbox": [ + 225, + 253, + 785, + 407 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d94a63e051fc0935ca7cf572716fc202e2be2bbf220b3bdd9c553111b078c401.jpg", + "image_caption": [ + "Fig. 2: Our proposed architecture for 3D segmentation framework leverages RAPiD features from the point cloud. We encode pointwise features into voxel-wise embeddings via the voxel encoder and multiple RAPiD AutoEncoders (RAPiD AE). After attention-based feature fusion, these fused embeddings go through the backbone network for segmentation results." + ], + "image_footnote": [], + "bbox": [ + 217, + 426, + 782, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 652, + 387, + 669 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LiDAR-Based Semantic Segmentation [7,12,17,20,22,23,27,30,34,37,38,49, 50,56-58,60,61,64,65] is fundamental for LiDAR-driven scene perception, aiming to label each point in a point cloud sequence. The majority of the approaches [17, 38,49,57,64] solely rely on the point-based features of the point cloud, such as SPVCNN [49] which introduces a point-to-voxel branch, using combined pointvoxel features for segmentation. Cylinder3D [64] proposes cylindrical partitioning with a UNet [9] backbone variant. LiM3D [30] utilizes coordinates combined with surface reflectivity attributes. Overall, such prior work relies solely on pointbased features such as coordinates and intensity of the points, lacking an effective", + "bbox": [ + 212, + 680, + 787, + 816 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "RAPiD-Seg Networks for 3D LiDAR Segmentation", + "bbox": [ + 395, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1 The code is publicly available at: https://github.com/l1997i/rapid_seg.", + "bbox": [ + 217, + 824, + 720, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "fusion mechanism, resulting in suboptimal performance [6,28,51]. They are also susceptible to changes in viewpoint, distance, and point sparsity [26] due to the lack of isometry across all inter-point distances.", + "bbox": [ + 212, + 146, + 784, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pointwise Distance Distribution (PDD) captures the local context of each point in a unit cell by enumerating distances to neighboring points in order. It is an isometry invariant proposed by Widdowson & Kurlin [52] to resolve the data ambiguity for periodic crystals, demonstrated through extensive pairwise comparisons across atomic 3D clouds from high-level periodic crystals of periodic structures [52-54]. Though the effectiveness of PDD in periodic crystals and atomic clouds has been proved by the aforementioned studies, to date, no work has applied PDD features to outdoor 3D point clouds. In outdoor settings, common invariant features [21,31,32,40] and Azimuth-normalized method [4] often face issues due to the irregular and sparse nature of the data, which is compounded by increased noise and environmental complexities [31,32]. Furthermore, the computational demands make them less suitable for the vast scale of outdoor settings [21,40]. For instance, Melia et al. [40] introduce a rotation-invariant feature that struggles with generalization across diverse point cloud densities and scales due to its computational cost and susceptibility to outdoor noise. Recognizing these limitations, we identify an opportunity to leverage the PDD features in a new domain, where representing the local context of neighboring points in a transformation invariant and geometrically robust manner is crucial. Drawing from the advantages of the PDD design, we propose the RAPiD feature, tailored specifically for LiDAR-based point clouds, to capture the localized geometry of neighboring structures.", + "bbox": [ + 212, + 191, + 787, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Point Cloud Representations and Embeddings are fundamental for LiDAR-driven tasks. 3D point cloud can be represented either as point-based [44,45,55,63] or voxel-based [17,35,50,64], each serving different purposes rooted in their unique characteristics. Voxel-based representation involves dividing the 3D space into grids of voxels, which provides a structured way of handling sparse and irregular LiDAR data [9, 13, 14, 64]. Previous work often utilizes Feedforward Neural Network (FNN) [17, 50, 64] to facilitate the conversion of pointwise features into structured voxel-wise features by projecting them onto voxel grids and generating dense embeddings. The primary drawback of FNN conversion is the potential loss of fine-grained details inherent in the original point cloud [15]. FNN may also struggle with the high dimensionality and sparsity of voxelized data, leading to computational inefficiency [46]. Our method addresses these issues by partitioning the point cloud into voxel grids and applying a novel class-ware double nested AE with self-attention mechanisms [15]. This approach not only retains more detailed information by focusing on local structures but also introduces high-dimensional feature compression, thereby improving computational efficiency.", + "bbox": [ + 212, + 508, + 787, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Range-Aware Pointwise Distances (RAPiD)", + "text_level": 1, + "bbox": [ + 214, + 766, + 669, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the context of LiDAR segmentation, a principal challenge lies in the lack of robustness against rigid transformations such as rotation and translation, as well as against variations in viewpoint, points sparsity, and occlusion [5, 26].", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "L. Li et al.", + "bbox": [ + 271, + 114, + 344, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Traditional methods predominantly leverage 3D point coordinates to furnish spatial information [17,50,64]; however, they may be inadequate in the scenes with poor visibility (e.g., occlusions) or sparse observations (e.g., at long range) [26]. Such reliance on coordinates alone could lead to inaccuracies in recognizing object transformations or occlusions [26]. Data augmentation, e.g., random geometric transformation can improve robustness under rigid transformations [50,64], but fail to guarantee comprehensive coverage of all potential transformations, resulting in vulnerability to previously unseen variations.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Overview", + "text_level": 1, + "bbox": [ + 215, + 287, + 339, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To achieve a transformation-invariant 3D data representation, we observe that distances within rigid bodies (e.g., vehicles, roads, and buildings) remain constant under rigid transformations. In light of this, we focus on assimilating the principle of the isometric invariant into the LiDAR-driven point cloud perception. Specifically, we delve into the PDD [52] — an isometry invariant that quantifies distances between adjacent points. For outdoor point clouds, vanilla PDD features are computationally intensive and susceptible to noise and sparsity. Our Range-Aware Pointwise Distance Distribution (RAPiD), specifically designed for LiDAR data, instead calculates the PDD features for each point within specific Regions of Interest (RoI), which are typically associated with the intrinsic structure of LiDAR data.", + "bbox": [ + 212, + 310, + 787, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mathematical Formulation: Given a fixed number $k > 0$ representing the fixed number of point neighbors and a $u$ -point cluster $P_{\\mathrm{RoI}}$ comprising no fewer than $k$ points based on RoI, the RAPiD is a $u \\times k$ matrix, which retains both spatial distances and LiDAR reflectivity [30] disparities between points. RAPiD are adapted to LiDAR sparsity at different distances by using range-specific parameters $k_{\\mathrm{close}}$ , $k_{\\mathrm{mid}}$ , $k_{\\mathrm{far}}$ for close, mid, and far ranges, respectively.", + "bbox": [ + 212, + 460, + 787, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The $k$ -point RAPiD in region $P_{\\mathrm{RoI}}$ is defined as:", + "bbox": [ + 238, + 551, + 586, + 566 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {R A P i D} \\left(P _ {\\text {R o I}}; k\\right) = \\operatorname {s o r t} \\left(\\left[ \\operatorname {s o r t} \\left(\\left[ \\boldsymbol {\\rho} _ {j, 1}, \\dots , \\boldsymbol {\\rho} _ {j, k} \\right]\\right) \\right] _ {j = 1} ^ {u}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 568, + 785, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\forall l \\in \\{1, \\dots, k\\}, j \\in \\{1, \\dots, u\\}, \\pmb{\\rho}_{j,l}$ is given by:", + "bbox": [ + 214, + 592, + 557, + 609 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\rho} _ {j, l} = \\left\\| \\left[ \\boldsymbol {p} _ {j} - \\boldsymbol {p} _ {j, l}, \\quad g (r _ {j}) - g (r _ {j, l}) \\right] \\right\\| _ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 616, + 785, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{p}_j$ and $\\pmb{p}_{j,l}$ denote the 3D coordinates of the $j$ -th point and its $l$ -th nearest neighbor within $P_{\\mathrm{RoI}}$ , respectively; $r_j$ and $r_{j,l}$ represent the reflectivity values of $\\pmb{p}_j$ and $\\pmb{p}_{j,l}$ , correspondingly; $\\| \\cdot \\| _2$ is the Euclidean norm. $g:\\mathbb{R}\\to [D_{\\min},D_{\\max}]$ is the reflectivity mapping function that maps the numerical range of reflectivity onto a consistent scale with the range of Euclidean distances between points:", + "bbox": [ + 212, + 642, + 787, + 719 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ng (r) = \\left(\\frac {r - r _ {\\operatorname* {m i n}}}{r _ {\\operatorname* {m a x}} - r _ {\\operatorname* {m i n}}}\\right) \\left(D _ {\\operatorname* {m a x}} - D _ {\\operatorname* {m i n}}\\right) + D _ {\\operatorname* {m i n}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 726, + 785, + 758 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD _ {\\min } = \\min _ {j, l} \\| \\boldsymbol {p} _ {j} - \\boldsymbol {p} _ {j, l} \\| _ {2}, \\quad D _ {\\max } = \\max _ {j, l} \\| \\boldsymbol {p} _ {j} - \\boldsymbol {p} _ {j, l} \\| _ {2}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 762, + 785, + 784 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $[D_{\\min}, D_{\\max}]$ is the range of the Euclidean norms of coordinate differences for all considered point pairs.", + "bbox": [ + 212, + 792, + 785, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The 4D distance in Eq. (2) integrates material reflectivity [30] into RAPiD features, enhancing feature representation and aiding in the discrimination of various materials and surfaces, which is crucial for accurate semantic segmentation.", + "bbox": [ + 212, + 823, + 785, + 868 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "RAPiD-Seg Networks for 3D LiDAR Segmentation", + "bbox": [ + 395, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For convenience to facilitate feature normalization and data alignment, we arrange RAPiD lexicographically by sorting Eq. (1), where $\\mathrm{sort}(\\cdot)$ on the inner and outer brackets sorts the elements $\\rho_{j,l}$ within each row $j$ , and the sorted rows based on their first differing elements, both in ascending order [52]. It treats RAPiD matrices with the same geometric structure but different orders as identical.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To improve the data robustness, we normalize $\\rho_{j,l}$ after excluding outliers exceeding a distance threshold $\\delta$ , substituting these entries with the maximum of the normalized distribution to represent significant inter-point distances.", + "bbox": [ + 212, + 222, + 784, + 267 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Intra-Ring and Intra-Class RAPiD", + "text_level": 1, + "bbox": [ + 214, + 276, + 542, + 290 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To enhance the RAPiD ability to capture local context, we propose R-RAPiD and C-RAPiD, based on the inherent structure of LiDAR data. The key benefit of C-RAPiD is its capacity to underscore the inherent traits of points within the same semantic class, achieved based on semantic labels before distance calculation. As a complementary, R-RAPiD is versatile, regardless of semantic label availability. Intra-Ring RAPiD (R-RAPiD) is a specialized variant of RAPiD, which confines the RoI to the ring encompassing the anchor points, thereby serving to economize on computational overhead while capitalizing on the inherent structural characteristics of LiDAR data (Fig. 2 ①).", + "bbox": [ + 212, + 301, + 784, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given LiDAR sensors with fixed beam counts [2,3,29,48], e.g., $B \\in \\{32,64,128\\}$ , beams radiate isotropically around the vehicle at set angles. We segment the $m$ -point LiDAR point cloud $C_{B,s}$ by beams, where $m = s \\times B$ , with $s$ representing measurements within a scan cycle and $B$ the number of laser beams. Suppose $\\pmb{p}_i$ with Cartesian coordinate $(x_i, y_i, z_i)$ is a point within $C_{B,s}$ , its transformation to cylindrical coordinates $(\\theta_i, \\phi_i)$ is formulated as:", + "bbox": [ + 212, + 438, + 790, + 529 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {i} = \\left\\lfloor \\arctan 2 \\left(y _ {i}, x _ {i}\\right) / \\Delta \\theta \\right\\rfloor , \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 351, + 542, + 785, + 558 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {i} = \\left\\lfloor \\arcsin \\left[ z _ {i} \\left(x _ {i} ^ {2} + y _ {i} ^ {2} + z _ {i} ^ {2}\\right) ^ {- 1 / 2} \\right] / \\Delta \\phi \\right\\rfloor , \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 561, + 785, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\varDelta\\theta$ and $\\varDelta\\phi$ denote the mean angular resolutions horizontally and vertically between adjacent beams, and $\\arctan 2(\\cdot)$ is the 2-argument $\\arctan (\\cdot)$ function.", + "bbox": [ + 212, + 597, + 784, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$C_{B,s}$ is then partitioned to $B$ rings based on $\\phi_b$ , i.e., $C_{B,s} = \\bigcup_{b=0}^{B-1} R_b$ , where each ring $R_b = \\{(\\theta_{b,i}, \\phi_{b,i}) \\mid \\phi_{b,i} = \\phi_b, \\forall i \\leq m\\}$ ; $\\bigcup$ signifies the cumulative union R-RAPiD is thus defined as the cumulative concatenate $\\oplus$ of RAPiD in each ring $R_b$ .", + "bbox": [ + 212, + 626, + 784, + 672 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\text {R - R A P i D} \\left(C _ {B, s}; k\\right) := \\bigoplus_ {b = 0} ^ {B - 1} \\text {R A P i D} \\left(R _ {b}; k\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 680, + 785, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Intra-Class RAPiD (C-RAPiD) concentrates on extracting point features within the confines of each semantic class (Fig. 2 ②). It is complementary to R-RAPiD, preserving the embedding fidelity within individual semantic class.", + "bbox": [ + 212, + 708, + 782, + 752 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, C-RAPiD concatenates RAPiD for each point $\\pmb{p}_j$ and its corresponding semantic label $y_{j}$ within semantic class $i$ , across $N_{c}$ total classes:", + "bbox": [ + 212, + 753, + 784, + 782 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\text {C - R A P i D} \\left(\\left\\{C _ {B, s}, \\mathcal {Y} \\right\\}; k\\right) := \\bigoplus_ {i = 0} ^ {N _ {c} - 1} \\operatorname {R A P i D} \\left(S _ {i}; k\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 792, + 785, + 813 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $S_{i} = \\{\\pmb{p}_{j}\\mid y_{j} = i,\\forall j\\leq m\\}$ $\\mathcal{V} = [y_j]_{j = 1}^m$", + "bbox": [ + 218, + 825, + 570, + 843 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "L. Li et al.", + "bbox": [ + 271, + 114, + 344, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 RAPiD Embedding", + "text_level": 1, + "bbox": [ + 215, + 143, + 434, + 162 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The direct utilization of high-dimensional RAPiD features imposes a substantial computational burden. Efficient processing of large data volumes requires methods that can condense dimensionality while preserving data integrity [62].", + "bbox": [ + 212, + 176, + 782, + 222 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We explore data embedding methods to reduce computational costs while providing feature learning capacity. AE is efficient in embedding high-dimensional features utilizing the reconstruction loss [16, 42]. However, conventional AE, limited by their sensitivity to input order and reliance on fixed-size misinputs [24], struggle to align with the unordered and variable size inherent in point clouds [15].", + "bbox": [ + 212, + 223, + 785, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We propose a double nested AE structure with a novel class-aware embedding objective, using the Voxel-based Set Attention (VSA) module [15] as a building block. This facilitates superior contextual awareness and adaptability to unordered and variably-sized point cloud data.", + "bbox": [ + 212, + 299, + 785, + 358 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Nested RAPiD AE, illustrated in Fig. 2 ③ and Fig. 3, is composed of two modules: the outer module is constituted by a VSA AE, primarily responsible for the point-to-voxel conversion. Within the inner module, which focuses on voxel-wise representation, we employ an additional AE specifically designed for dimensionality reduction. Specifically, considering an input of $m$ -point RAPiD features $G \\in \\mathbb{R}^{m \\times d}$ , where each point encompasses $d$ features. We compress pointwise $G$ into a voxel-wise representation $H^v \\in \\mathbb{R}^{c \\times l \\times d}$ by an outer VSA Voxel Encoder (with $c$ reduced voxel-wise dimension and $l$ latent codes), then further reduces it to a compressed embedding $\\hbar \\in \\mathbb{R}^{c \\times l \\times d'}$ with $d'$ features via an inner Encoder. The inner and outer VSA Point Decoders then reconstruct the output feature set $\\hat{G}$ akin to $G$ .", + "bbox": [ + 212, + 359, + 787, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Inner Voxel-Wise Representation is conducted via an inner AE, which takes $H^v$ from outer VSA AE as the input, yielding a lower-dimensional embedding $\\hbar$ (dimension reduced from $d$ to $d'$ ). This process involves $u \\times$ convolutional layers and a batch normalization layer [19]:", + "bbox": [ + 212, + 526, + 787, + 587 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hbar^ {(i)} = \\operatorname {B a t c h N o r m} \\circ \\operatorname {C o n v} \\left(\\hbar^ {(i - 1)}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 599, + 785, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\hbar^{(0)} = H^v$ , $\\hbar^{(u - 1)} = \\hbar$ . The Convolutional Feed-Forward Network (ConvFFN) then promotes voxel-level information exchange, enhancing spatial feature interactivity. It maps reduced hidden features $\\hbar$ to a 3D sparse tensor, indexed by voxel coordinates $X^v$ , and employs dual depth-wise convolutions (DwConv) for spatial interactivity:", + "bbox": [ + 212, + 628, + 787, + 705 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {h} = \\operatorname {D w C o n v} ^ {(2)} \\circ \\zeta \\circ \\operatorname {D w C o n v} ^ {(1)} \\left(\\operatorname {S p T} \\left(\\hbar , X ^ {v}\\right)\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 717, + 785, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\zeta$ and $\\mathrm{SpT}(\\cdot)$ represents the non-linear activation and sparse tensor construction. Subsequently, $\\hat{\\hbar}\\in \\mathbb{R}^{c\\times l\\times d'}$ is reconstructed into $\\hat{H}^v\\in \\mathbb{R}^{c\\times l\\times d}$ through an inner Decoder, consisting of multiple DeConv Layers.", + "bbox": [ + 212, + 747, + 787, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Outer VSA AE consists of a voxel encoder and point encoder. The voxel encoder project RAPiD features into key-value spaces, forming $K$ and $V$ , followed by a cross-attention mechanism with a latent query $L$ , producing an attention", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "RAPiD-Seg Networks for 3D LiDAR Segmentation", + "bbox": [ + 395, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bd43c21d0b93992e957c04b3a21a936e2d971f12f33688afff80e8c5bb26aac8.jpg", + "image_caption": [ + "Fig. 3: Our RAPiD AE consists of an Encoder, Convolution Layer, and Decoder module, aiming to reproduce the input features and generate the compressed voxel-wise RAPiD representation $\\hbar$ ." + ], + "image_footnote": [], + "bbox": [ + 289, + 145, + 715, + 277 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "matrix $A$ . The voxel-wise representation $H^v$ is the scatter sum [43] of pointwise $H$ to aggregate the value vectors $V$ :", + "bbox": [ + 214, + 333, + 787, + 364 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nH ^ {v} = \\operatorname {S u m} _ {\\text {s c a t t e r}} (H, I ^ {v}), \\quad H = \\tilde {A} ^ {\\top} V, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 375, + 787, + 392 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {A} = \\operatorname {S o f t m a x} _ {\\text {s c a t t e r}} \\left(K L ^ {\\top}, I ^ {v}\\right), \\quad (Q, V) = \\operatorname {P r o j} (G), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 395, + 787, + 414 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\operatorname{Proj}(\\cdot)$ represents the linear projection, and $I^v$ is the voxel indices. The Point Decoder reconstructs the output set from the enriched hidden features $\\hat{H}^v$ . We start by broadcasting $\\hat{H}^v$ based on $I^v$ , resulting in $\\hat{H} \\in \\mathbb{R}^{m \\times l \\times d}$ . The output $\\hat{G}$ is analogous to the operational paradigm of Voxel Encoder:", + "bbox": [ + 214, + 422, + 789, + 484 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {G} = \\tilde {A} ^ {*} ^ {\\top} V ^ {*}, \\quad \\tilde {A} ^ {*} = \\operatorname {S o f t m a x} \\left(A ^ {*}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 494, + 787, + 516 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nA ^ {*} = \\left[ K _ {i} ^ {*} Q _ {i} ^ {\\top} \\right] _ {i = 1} ^ {m}, \\quad \\left(K ^ {*}, V ^ {*}\\right) = \\operatorname {P r o j} (\\hat {\\mathcal {H}}). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 518, + 787, + 540 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Class-Aware Embedding Objective addresses the issue of non-uniqueness in embeddings produced by the AE, where various distinct inputs yield approximately the same embedding, leading to inaccurate representation [1, 25]. Our objective aims to facilitate the generation of AE embeddings that demonstrate robust semantic class discriminability.", + "bbox": [ + 214, + 547, + 787, + 623 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Specifically, we introduce a novel class-aware contrastive loss $\\mathcal{L}_{\\mathrm{contr}}$ in Eq. (15), which aims to maximize the distance between embeddings of different semantic classes while minimizing the distance of the same class.", + "bbox": [ + 214, + 623, + 787, + 665 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {c o n t r}} = \\frac {1}{m} \\sum_ {i = 1} ^ {m} \\left[ \\frac {1}{| P (i) |} \\sum_ {p \\in P (i)} \\mathrm {R e L U} (\\alpha - \\mathrm {s i m} (H _ {i}, H _ {p})) \\right. \\\\ \\left. + \\frac {1}{| N (i) |} \\sum_ {n \\in N (i)} \\operatorname {R e L U} \\left(\\sin \\left(H _ {i}, H _ {n}\\right) - \\alpha\\right) \\right], \\tag {15} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 320, + 670, + 785, + 760 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c o n}} = \\frac {1}{m \\cdot d} \\sum_ {i = 1} ^ {m} \\sum_ {j = 1} ^ {d} \\left(G _ {i, j} - \\hat {G} _ {i, j}\\right) ^ {2}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 320, + 763, + 785, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $i$ represents the point index; $P(i)$ and $N(i)$ are the indices of the nearest point in the same (Positive) and different class (Negative) relative to point $i$ ,", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "L. Li et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "respectively; $\\mathrm{sim}(\\cdot)$ is a similarity measure; $\\alpha$ is a margin controlling separation between classes. The pointwise representations $H_{i}$ , $H_{p}$ , and $H_{n}$ are broadcasted from $\\hbar$ based on voxel indices $I^{v}$ .", + "bbox": [ + 212, + 146, + 782, + 190 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We further combine this with the MSE [39] reconstruction loss in Eq. (16). The overall loss function, with $\\lambda$ balancing the fidelity of reconstruction with the distinctiveness of the embeddings, is formulated as: $\\mathcal{L}_{\\mathrm{total}} = \\mathcal{L}_{\\mathrm{recon}} + \\lambda \\mathcal{L}_{\\mathrm{contr}}$ .", + "bbox": [ + 212, + 191, + 784, + 238 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 RAPiD-Seg for 3D LiDAR Segmentation", + "text_level": 1, + "bbox": [ + 214, + 252, + 643, + 270 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We present RAPiD-Seg, a 3D LiDAR segmentation network leveraging RAPiD features. It incorporates multiple complementary features via the channel-wise fusion mechanism [18]. RAPiD-Seg takes point cloud as input and performs single-modal (LiDAR-only) 3D semantic segmentation. Specifically, the input point cloud has three types of point-wise features: (a) coordinates-based features $F_{C} = \\{p_{C} \\mid p_{C} = (x,y,z)\\}$ ; (b) intensity-based features $F_{I} = \\{p_{I} \\mid p_{I} = (I,J)\\}$ , where $I$ and $J$ are intensity and reflectivity; (c) RAPiD features $F_{R} = \\{p_{R} \\mid p_{R} = \\mathrm{RAPiD}(P_{\\mathrm{RoI}};k)\\}$ . The RoI varies depending on whether it is R- or C-RAPiD.", + "bbox": [ + 212, + 282, + 784, + 404 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As shown in Fig. 2, $F_{C} \\oplus F_{I}$ and $F_{R}$ are fed to a VSA voxel encoder 4 and RAPiD AE 6 for voxelization, resulting in voxel-wise representation $E_{C}$ , $E_{I}$ and $E_{R}$ . Subsequently, we incorporate the complementary voxel-wise features via the channel-wise fusion mechanism (Sec. 5.1). The backbone net takes in them for LiDAR segmentation prediction. Additionally, we introduce two effective RAPiD-Seg variants to expedite the convergence of the extensive segmentation network towards an optimal solution.", + "bbox": [ + 212, + 404, + 784, + 508 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 RAPiD Channel-Wise Fusion with Attention", + "text_level": 1, + "bbox": [ + 214, + 530, + 625, + 544 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To combine various LiDAR point attributes, a highly efficient method for feature fusion is still a research topic [30, 50]. Current methods, which typically concatenate different LiDAR attributes (e.g., intensity [38, 41, 59], reflectivity [30], PLS [50]) face issues with increased dimensionality and complexity, and the need for balancing weights to avoid biased training towards dominant attributes.", + "bbox": [ + 212, + 551, + 784, + 627 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thus propose a novel RAPiD Fusion with Attention module (FuAtten), pioneering application of channel-wise attention mechanism [18] in LiDAR point attribute fusion. With FuAtten, we effectively fuse $E_{C}$ , $E_{I}$ and $E_{R}$ , emphasizing informative features adaptively while suppressing less relevant ones across different channels.", + "bbox": [ + 212, + 628, + 784, + 688 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As shown in Fig. 2 ⑤ and Fig. 4, we initially concatenate $E_{C}$ , $E_{I}$ and $E_{R}$ along the voxel dimension, resulting in $\\mathbf{E}$ . In the squeeze operation, Global Average Pooling (GAP) [33] condenses each channel $f'$ of the tensor into a representative channel descriptor. These descriptors form a channel-level embedding $\\mathbf{z} = [z_{f'}]_{f' = 1}^{f^*}$ .", + "bbox": [ + 212, + 688, + 784, + 753 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} = \\left[ \\operatorname {G A P} \\left(\\mathbf {E} _ {:,:: f ^ {\\prime}}\\right) \\right] _ {f ^ {\\prime} = 1} ^ {f ^ {*}} = \\frac {1}{c \\times l} \\left[ \\sum_ {c ^ {\\prime} = 0} ^ {c - 1} \\sum_ {l ^ {\\prime} = 0} ^ {l - 1} \\mathbf {E} _ {c ^ {\\prime}, l ^ {\\prime}, f ^ {\\prime}} \\right] _ {f ^ {\\prime} = 1} ^ {f ^ {*}}, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 763, + 785, + 806 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $f^{*}$ is the dimension of the concatenated feature. A channel-wise attention $\\pmb{a}_{\\mathbf{z}} \\in \\mathbb{R}^{1 \\times 1 \\times f^{*}}$ is then computed through the excitation step: $\\pmb{a}_{\\mathbf{z}} =$", + "bbox": [ + 212, + 809, + 784, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "RAPiD-Seg Networks for 3D LiDAR Segmentation", + "bbox": [ + 395, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "$\\sigma (\\mathbf{W}_2\\delta (\\mathbf{W}_1\\mathbf{z}))$ , where $\\delta$ and $\\sigma$ represent the ReLU and Sigmoid activation. With each element in $a_{\\mathbf{z}}$ reflects the attention allocated to the corresponding feature of the voxel, the channel-wise fused features $\\mathbf{E}^{\\prime} = a_{\\mathbf{z}}\\cdot \\mathbf{E}$", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/dc886fcf06578cc3e62a634b3443e1019ce6c77fee60ca35399c435489a758fb.jpg", + "image_caption": [ + "Fig. 4: Our feature fusion module with channel-wise attention." + ], + "image_footnote": [], + "bbox": [ + 333, + 199, + 671, + 276 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 RAPiD-Seg Architectures for 3D Segmentation", + "text_level": 1, + "bbox": [ + 214, + 319, + 648, + 335 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The complexity inherent in 3D LiDAR-driven networks typically exacerbates the end-to-end training process [17,30,50], since the extensive parameter space increases the propensity for overfitting, slow convergence, and the possibility of settling into local minima [30].", + "bbox": [ + 212, + 344, + 787, + 406 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/b7e9f5e0bdcc3325367e44cee323ec1fd258f2afb3c56e1104922b348d5c9b2d.jpg", + "image_caption": [ + "Fig. 5: Two variants of RAPiD-Seg. (a) R-RAPiD-Seg utilizes R-RAPiD features, and (b) C-RAPiD-Seg utilizes both R- and C-RAPiD features for better performance." + ], + "image_footnote": [], + "bbox": [ + 261, + 419, + 743, + 510 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As shown in Fig. 5, we propose two novel and effective variants of RAPiD-Seg for quicker and better performance. R-RAPiD-Seg involves AE training from scratch, followed by C-RAPiD-Seg utilizing C-RAPiD features with the pre-trained AE and backbone for enhanced performance.", + "bbox": [ + 212, + 566, + 787, + 628 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "R-RAPiD-Seg. We first construct the lightweight R-RAPiD-Seg (Fig. 5 (a)) for fast 3D segmentation. R-RAPiD-Seg adopts an early fusion scheme utilizing LiDAR original points and R-RAPiD features. Specifically, 3D coordinate $F_{C}$ and intensity-based features $F_{I}$ are voxelized to voxel-wise representations $R_{v} = E_{C} \\oplus E_{I}$ based on VSA voxelization [15]. We get the compressed $d'$ -dimension voxel-wise R-RAPiD representations $\\hbar$ (Sec. 4) from RAPiD AE. $\\hbar$ is then fused into the voxel-wise representations $R_{v*} = \\mathrm{FuAtten}(R_v,\\hbar)$ , where $\\mathrm{FuAtten}(\\cdot)$ is Fusion with Attention in Sec. 5.1. The fused voxel-wise features are taken into the backbone network for 3D segmentation.", + "bbox": [ + 212, + 628, + 787, + 763 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "C-RAPiD-Seg. To facilitate the embedding fidelity within individual semantic classes, we also design a class-aware framework with C-RAPiD features, i.e., C-RAPiD-Seg (Fig. 5 (b)). Specifically, we fuse the voxel-wise C-RAPiD $\\hbar_C$ into the representations $R_{v*} = \\mathrm{FuAtten}(R_v,\\hbar \\oplus \\hbar_C)$ . C-RAPiD requires class labels to compute the features regarding the semantic categories. Since the ground-truth", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "L. Li et al.", + "bbox": [ + 271, + 114, + 344, + 127 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "labels are missing during the test time, we generate the reliable pseudo labels $\\tilde{\\mathcal{V}}$ with a pretrained R-RAPiD-Seg based on confidence (Fig. 5 (b) blue dotted arrow). Subsequently, the pseudo voxel-wise C-RAPiD $\\hbar_{C}$ are generated through $\\tilde{\\mathcal{V}}$ . We fuse them into the voxel-wise representations for 3D segmentation with the bonebone network.", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6 Evaluation", + "text_level": 1, + "bbox": [ + 214, + 233, + 357, + 248 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Following the popular practice of LiDAR segmentation methods [17,37,64], we evaluate our proposed RAPiD-Seg network against SoTA 3D LiDAR segmentation approaches on the SemanticKITTI [2] and nuScenes [3] datasets.", + "bbox": [ + 212, + 263, + 785, + 310 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 214, + 320, + 428, + 335 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Datasets: SemanticKITTI [2] comprises 22 point cloud sequences, with sequences 00-10, 08, and 11-21 for training, validation, and testing. 19 classes are chosen for training and evaluation by merging classes with similar motion statuses and discarding sparsely represented ones. Meanwhile, nuScenes [3] has 1000 driving scenes; 850 are for training and validation, with the remaining 150 for testing. 16 classes are used for LiDAR semantic segmentation, following the amalgamation of akin classes and the removal of rare ones.", + "bbox": [ + 212, + 342, + 787, + 446 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation Protocol: Following the popular practice of [2,36,64], we adopt the Intersection-over-Union (IoU) of each class and mean IoU (mIoU) of all classes as the evaluation metric. The IoU of class $i$ is defined as $\\mathrm{IoU}_i = \\mathrm{TP}_i / (\\mathrm{TP}_i + \\mathrm{FP}_i + \\mathrm{FN}_i)$ , where $\\mathrm{TP}_i$ , $\\mathrm{FP}_i$ and $\\mathrm{FN}_i$ denote the true positive, false positive and false negative of class $i$ , respectively.", + "bbox": [ + 212, + 449, + 789, + 523 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation Details: We construct the point-voxel backbone based on the Minkowski-UNet34 [8] (re-implemented by PCSeg [36] codebase), which is the open-access backbone with SoTA results to date. Before the AE training stage, we first generate pointwise RAPiD features (Sec. 3), which yield three RAPiD outputs for each frame, corresponding to different $k$ values based on range. Notably, this stage does not impose additional burdens on the subsequent overall training. The number of maximum training epochs for AE and the whole network is set as 100 and the initial learning rate is set as $10^{-3}$ with SGD optimizer. We use 2 epochs to warm up the network and adopt the cosine learning rate schedule for the remaining epochs. All experiments are conducted on $4 \\times$ NVIDIA A100 GPUs ( $1 \\times$ for inference).", + "bbox": [ + 212, + 525, + 789, + 691 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6.2 Experimental Results", + "text_level": 1, + "bbox": [ + 214, + 709, + 441, + 723 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In Tab. 1 and Tab. 2, we showcase the performance of our RAPiD-Seg LiDAR segmentation method on SemanticKITTI and nuScenes test set in comparison with published leading contemporary SoTA approaches to demonstrate its superior efficacy (top 10 results shown; full rankings in Supplementary Material).", + "bbox": [ + 212, + 733, + 785, + 794 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Our method significantly outperforms others, especially for rigid object categories (e.g., truck, o.veh, park, etc.), primarily due to our fusion of both localized 3D geometry and material reflectivity within RAPiD features, which", + "bbox": [ + 212, + 795, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "RAPiD-Seg Networks for 3D LiDAR Segmentation", + "bbox": [ + 395, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/559b08fb2a4b23d3f6463f08d84d64310133446303cba847ff0ec81d5a90f0bb.jpg", + "table_caption": [ + "Table 1: Quantitative results of RAPiD-Seg and top-10 SoTA segmentation methods on SemanticKITTI [2] test set; Best/2nd best highlighted; ▲ for multi-modal methods." + ], + "table_footnote": [], + "table_body": "
MethodmIoUcarbicymotortruco.vehpedb.listm.listroadparkwalko.grobuildfencvegtrunterrpolesign
Cylinder3D [64]68.997.167.663.850.858.573.769.248.092.265.077.032.390.766.585.672.569.862.466.2
AF2S3Net [7]69.794.565.486.839.241.180.780.474.391.368.872.553.587.963.270.268.553.761.571.0
RPVNet [56]70.397.668.468.744.261.175.974.473.493.470.380.733.393.572.186.575.171.764.861.4
SDSeg3D [27]70.497.458.754.254.965.270.274.452.290.969.476.741.993.271.186.174.371.165.470.6
GASN [61]70.796.965.858.059.361.080.482.746.389.866.274.630.192.369.687.373.072.566.171.6
PVKD [17]71.297.067.969.353.560.275.173.550.591.870.977.541.092.469.486.573.871.964.965.8
2DPASS [58]72.997.063.663.461.161.577.981.374.189.767.474.740.093.572.986.273.971.065.070.4
PCSeg [36]72.997.551.267.658.668.678.380.975.692.571.578.336.993.171.485.473.669.966.168.7
RangeFormer [22]73.396.769.473.759.966.278.175.958.192.473.078.842.492.370.186.673.372.866.466.6
▲UniSeg [37]75.297.971.975.263.674.178.974.860.692.674.079.546.193.472.787.576.373.168.368.5
RAPiD-Seg (Ours)76.197.771.176.272.580.779.979.159.891.878.278.646.093.672.186.974.672.365.968.5
", + "bbox": [ + 215, + 174, + 785, + 287 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/05a9e076bb593d34450fd5d365b3b1624a7cca05e26bee56bf23279836d69439.jpg", + "table_caption": [ + "Table 2: Quantitative results of RAPiD-Seg and top-10 SoTA segmentation methods on nuScenes [3] test set; Best/2nd best highlighted; ▲ for multi-modal methods." + ], + "table_footnote": [], + "table_body": "
MethodmIoUbarrbicybuscarconstmotorpedconetrailtruckdrivotherwalkterrmadeveg
▲PMF [65]77.082.040.081.088.064.079.080.076.081.067.097.068.078.074.090.088.0
Cylinder3D [64]77.282.829.884.389.463.079.377.273.484.669.197.770.280.375.590.487.6
AMVNet [34]77.380.632.081.788.967.184.376.173.584.967.397.567.479.475.591.588.7
SPVCNN [49]77.480.030.091.990.864.779.075.670.981.074.697.469.280.076.189.387.1
AF2S3Net [7]78.378.952.289.984.277.474.377.372.083.973.897.166.577.574.087.786.8
▲2D3DNet [12]80.083.059.488.085.163.784.482.076.084.871.996.967.479.876.092.189.2
GASN [61]80.485.543.290.592.164.786.083.073.383.975.897.071.081.077.791.690.2
2DPASS [58]80.881.755.392.091.873.386.578.572.584.775.597.669.179.975.590.288.0
▲LidarMultiNet [60]81.480.448.494.390.071.587.285.280.486.974.897.867.380.776.592.189.6
▲UniSeg [37]83.585.971.292.191.680.588.080.976.086.376.797.771.880.776.791.388.8
RAPiD-Seg (Ours)83.684.864.395.092.284.687.981.876.888.579.097.866.681.276.792.588.4
", + "bbox": [ + 217, + 328, + 785, + 434 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "enable segmentation based on material properties and local rigid structures. Remarkably, our single-modal methodology outperforms multi-modal approaches [22, 34,37,56,58,60], suggesting superior efficacy of our RAPiD features over alternative modalities like RGB and range images. Our inference time (105ms per frame) is comparable to other contemporary approaches [37,58,64]. Furthermore, we present supporting qualitative results in Fig. 6. Whereas the baseline method struggles with accurate vehicle type differentiation, ours achieves consistent segmentation (more visualization results in the supplementary materials).", + "bbox": [ + 212, + 450, + 787, + 571 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/d5e7100fe03ddaa78d8fb659f468ad926e6009087a354c606108f673fc58cdad.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
RAPiD FeaturesAttentionmIoUΔ
GeometricReflectivityEmbedding
70.04(Baseline)
✓70.46(+0.42)
✓71.21(+1.17)
✓✓71.93(+1.89)
✓✓✓72.15(+2.11)
✓✓✓71.80(+1.76)
✓✓✓72.32(+2.28)
✓✓✓72.78(+2.74)
✓✓✓✓73.02(+2.98)
", + "bbox": [ + 223, + 580, + 486, + 686 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 3: Component-wise ablation of RAPiD-Seg on the SemanticKITTI validation set: Geometric (Eq. (1)), Reflectivity (Eq. (3)), (RAPiD) Embedding (Sec. 4), (channel) Attention (Sec. 5.1).", + "bbox": [ + 215, + 691, + 488, + 762 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7cb6032cfe40dba5254819957fa3c7df89b71a72f571df82c85e6ba790569fdd.jpg", + "image_caption": [ + "Fig.6: Comparing our results and PCSeg (baseline) under multi-scan visualization, showing improved segmentation results." + ], + "image_footnote": [], + "bbox": [ + 506, + 580, + 782, + 691 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "6.3 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 772, + 401, + 786 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effectiveness of Components: In Tab. 3, we ablate each component of RAPiD-Seg step by step and report the performance on the SemanticKITTI validation set. We start from a baseline model, achieving an mIoU of 70.04 on validation set.", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "L. Li et al.", + "bbox": [ + 271, + 114, + 344, + 127 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Incorporating the RAPiD features leads to a notable increase in mIoU (+1.17), underscoring the efficacy of RAPiD in enhancing segmentation performance. Building upon the RAPiD framework, the integration of reflectivity further elevates our mIoU to 71.93, with an overall improvement of +1.89 from the baseline. This significant gain illustrates the critical role of reflectivity in capturing different object materials. Utilizing RAPiD AE to get the voxel-wise representation, we observe an additional improvement, taking the mIoU to 72.15 (+2.11 compared to the baseline), which demonstrates its capability in processing feature representations. We subsequently performed a module-by-module reduction to analyze the model performance. Specifically, without RAPiD-related modules (Geometric/Reflectivity/Embedding), mIoU drops from 73.02 to 70.46 - attributable to the lack of RAPiD features, which are robust to transformation and noise. Our ablation studies show that while disabling individual components leads to some improvement over the baseline, the peak performance is only attained when all components function together. This highlights the indispensable contribution of each component in maximizing segmentation accuracy.", + "bbox": [ + 212, + 146, + 787, + 388 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/dc6c5d064c78177a7d34649b75b3614996844b0e65dd51c41d88da345a6fc525.jpg", + "table_caption": [ + "Table 4: Effects of RAPiD and Reflectivity features compared to other configurations on SemanticKITTI validation set. PDD: the compressed PDD embeddings; RAPiD-R: only RAPiD, without reflectivity feature fusion; RAPiD-R: both RAPiD and reflectivity." + ], + "table_footnote": [], + "table_body": "
Conf.mIoUtrucko.vehparkwalko.grobuildfence
BaseL [36]70.059.870.369.276.936.293.769.6
PDD [52]66.240.365.867.574.633.192.868.2
RAPiD-R71.862.569.070.377.435.893.670.8
RAPiD+R73.070.478.575.878.944.294.273.1
", + "bbox": [ + 217, + 444, + 785, + 516 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effectiveness of RAPiD and Reflectivity Features: In Tab. 4, we validate the efficacy of the proposed RAPiD features. The direct application of PDD results in a substantial mIoU decrease of 3.8, with most category IoUs falling below those of the baseline method. This indicates vanilla PDD features are not well-suited for LiDAR segmentation. While our proposed RAPiD (without reflectivity disparities in Eq. (2)) demonstrates an improvement of 1.4 in mIoU over the baseline. Moreover, the integration of reflectivity with RAPiD features significantly enhances performance, yielding 2.6 mIoU increase over the baseline. Notably, our approach exhibits superior performance in the segmentation of most rigid object categories compared to other configurations. Tab. 6 further shows the RAPiD outperforms PDD in varying $k$ values at different ranges on SemanticKITTI validation set, with non-uniform $k$ configurations yielding the most significant improvements in mIoU for both datasets, highlighting the effectiveness of the range-aware design of RAPiD.", + "bbox": [ + 212, + 535, + 787, + 747 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effectiveness of RAPiD AE and Architecture Variants: In Tab. 3, we replace our RAPiD AE with a commonly-used ConvFNN [17,50,64] to generate voxel-wise embeddings. Our AE demonstrated a $0.24\\mathrm{mIoU}$ improvement over the FNN. In Tab. 5, we investigate the impact of the variant architectures, i.e., R-RAPiD-Seg and C-RAPiD-Seg. Utilizing R-RAPiD features, R-RAPiD-Seg yields an mIoU of 72.3, which outperforms the baseline method by $+2.3\\mathrm{mIoU}$ .", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "RAPiD-Seg Networks for 3D LiDAR Segmentation", + "bbox": [ + 395, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/b9e633235092bc6fb522274be80493a7f77dae1ef341c92b1361c6631a57d6ac.jpg", + "table_caption": [ + "Table 5: 3D segmentation results of different variants of RAPiD-Seg (ours) on SemanticKITTI validation set." + ], + "table_footnote": [], + "table_body": "
MethodmIoU %carpedo.gropole
Baseline70.097.278.135.463.5
R-RAPiD-Seg72.3 (+2.3)97.477.445.062.4
C-RAPiD-Seg73.0 (+3.0)97.779.344.666.4
", + "bbox": [ + 220, + 189, + 491, + 250 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/f966758ae9c0c52f27d74a067727e55eee921cbe8b047b0040751a88a4f2aef3.jpg", + "table_caption": [ + "Table 6: Effects of $k$ at various ranges." + ], + "table_footnote": [], + "table_body": "
PDD [53]RAPiD (ours)
knearkmidkfarmIoUknearkmidkfarmIoU
SumK77764.74 (-7.1)77772.04 (+0.2)
55565.18 (-6.6)55572.28 (+0.5)
107566.23 (-5.6)107573.02 (+1.2)
lnuScene66672.19 (-6.5)66678.76 (+0.1)
33373.68 (-5.0)33379.43 (+0.8)
86372.24 (-6.4)86379.91 (+1.3)
", + "bbox": [ + 516, + 161, + 781, + 250 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Concurrent fusing of both R- and C-RAPiD-Seg features in R-RAPiD-Seg improves performance by $+0.7$ mIoU, which shows a $+3.0$ overall mIoU enhancement.", + "bbox": [ + 212, + 263, + 782, + 292 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Effectiveness of Feature Fusion with Channel-wise Attention: In Tab. 3, we assess the prevalent approach of direct feature concatenation. Our method, employing feature fusion with channel-wise attention, enhances mIoU by 0.87, thereby conclusively demonstrating its effectiveness.", + "bbox": [ + 212, + 292, + 785, + 353 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Effectiveness of Backbone Networks: In Tab. 7, we evaluate the performance across multiple backbone networks to demonstrate the versatility and ease of integration of our modules. The results highlight the adaptability and effectiveness of our approach on both pointwise (P) and Voxel-wise (V) backbone architectures. Specifically, the point-based PTv2 backbone achieves an mIoU of 72.6, showing strong performance, particularly in the car and other ground categories. Among the voxel-based methods, Minkowski-UNet stands out with the highest overall mIoU of 73.0, excelling in the pedestrian and pole categories.", + "bbox": [ + 212, + 354, + 787, + 474 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/75e317ece834e83fefbcaf0d91e273a55747253c1d649be67f99a8173c4bdab5.jpg", + "table_caption": [ + "Table 7: Effects of using different backbones on SemanticKITTI validation set, where P and V for Point- and Voxel-wsie methods." + ], + "table_footnote": [], + "table_body": "
Repr.BackbonemIoU %carpedo.gropole
PPTv2 [55]72.6 (+2.8)97.477.445.062.4
VCylinder3D [64]69.896.974.237.963.0
VMinkowski-UNet [8]73.0 (+3.2)97.278.144.265.9
", + "bbox": [ + 222, + 526, + 776, + 590 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 215, + 618, + 508, + 633 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper presents a novel Range-Aware Pointwise Distance Distribution (RAPiD) feature and the RAPiD-Seg network for LiDAR segmentation, adeptly overcoming the constraints of single-modal LiDAR methods. The rigid transformation invariance and enhanced focus on local details of RAPiD significantly boost segmentation accuracy. RAPiD-Seg integrates a two-stage training approach with reflectivity-guided 4D distance metrics and a class-aware nested AE, achieving SoTA results on the SemanticKITTI and nuScenes datasets. Notably, our single-modal method surpasses the performance of multi-modal methods, indicating the superior efficacy of RAPiD features even compared to other modalities, including RGB and range images.", + "bbox": [ + 212, + 643, + 787, + 794 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our RAPiD features hold significant potential for application in various tasks such as point cloud registration, semi-/weakly- supervised learning, and extend to promising applications in multi-modal research in the future.", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "L. Li et al.", + "bbox": [ + 271, + 114, + 344, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Badger, B.L.: Depth and Representation in Vision Models (2023)", + "2. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences. In: Int. Conf. Comput. Vis. pp. 9296-9306 (2019)", + "3. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: Nuscenes: A Multimodal Dataset for Autonomous Driving. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 11618-11628 (2020)", + "4. Chen, S., Wang, X., Cheng, T., Zhang, W., Zhang, Q., Huang, C., Liu, W.: Azinorm: Exploiting the Radial Symmetry of Point Cloud for Azimuth-Normalized 3D Perception. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6387-6396 (2022)", + "5. Chen, X., Xu, S., Zou, X., Cao, T., Yeung, D.Y., Fang, L.: SVQNet: Sparse Voxel-Adjacent Query Network for 4D Spatio-Temporal LiDAR Semantic Segmentation. In: Int. Conf. Comput. Vis. pp. 8569–8578 (2023)", + "6. Chen, X., Shi, S., Zhu, B., Cheung, K.C., Xu, H., Li, H.: MPPNet: Multi-frame Feature Intertwining with Proxy Points for 3D Temporal Object Detection. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Eur. Conf. Comput. Vis. pp. 680-697. Springer Nature Switzerland (2022)", + "7. Cheng, R., Razani, R., Taghavi, E., Li, E., Liu, B.: (AF)2-S3Net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12547-12556 (2021)", + "8. Choy, C., Gwak, J., Savarese, S.: 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3070-3079 (2019)", + "9. Çiçek, Ö., Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation. In: Ourselin, S., Joskowicz, L., Sabuncu, M.R., Unal, G., Wells, W. (eds.) Med. Image Comput. Comput. Assist. Interv. pp. 424–432. Springer International Publishing (2016)", + "10. Flitton, G., Breckon, T.P., Megherbi, N.: A Comparison of 3D Interest Point Descriptors with Application to Airport Baggage Object Detection in Complex CT Imagery. Pattern Recognition 46(9), 2420-2436 (2013)", + "11. Geiger, A., Lenz, P., Urtasun, R.: Are We Ready for Autonomous Driving? The KITTI Vision Benchmark Suite. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3354-3361 (2012)", + "12. Genova, K., Yin, X., Kundu, A., Pantofaru, C., Cole, F., Sud, A., Brewington, B., Shucker, B., Funkhouser, T.: Learning 3D Semantic Segmentation with only 2D Image Supervision. In: Int. Conf. 3D Vis. pp. 361-372 (2021)", + "13. Graham, B., Engelcke, M., van der Maaten, L.: 3D Semantic Segmentation with Submanifold Sparse Convolutional Networks. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018)", + "14. Han, L., Zheng, T., Xu, L., Fang, L.: OccuSeg: Occupancy-Aware 3D Instance Segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2940-2949 (2020)", + "15. He, C., Li, R., Li, S., Zhang, L.: Voxel Set Transformer: A Set-to-Set Approach to 3D Object Detection from Point Clouds. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 8407-8417. IEEE (2022)", + "16. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked Autoencoders Are Scalable Vision Learners. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 16000-16009 (2022)" + ], + "bbox": [ + 220, + 179, + 785, + 838 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "RAPiD-Seg Networks for 3D LiDAR Segmentation", + "bbox": [ + 395, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Hou, Y., Zhu, X., Ma, Y., Loy, C.C., Li, Y.: Point-to-Voxel Knowledge Distillation for LiDAR Semantic Segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)", + "18. Hu, J., Shen, L., Sun, G.: Squeeze-and-Eccitation Networks. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018)", + "19. Ioffe, S., Szegedy, C.: Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. In: Int. Conf. Mach. Learn. pp. 448-456 (2015)", + "20. Jiang, L., Shi, S., Tian, Z., Lai, X., Liu, S., Fu, C.W., Jia, J.: Guided Point Contrastive Learning for Semi-supervised Point Cloud Semantic Segmentation. In: Int. Conf. Comput. Vis. (2021)", + "21. Jiang, M., Wu, Y., Zhao, T., Zhao, Z., Lu, C.: PointSIFT: A SIFT-like Network Module for 3D Point Cloud Semantic Segmentation (2018)", + "22. Kong, L., Liu, Y., Chen, R., Ma, Y., Zhu, X., Li, Y., Hou, Y., Qiao, Y., Liu, Z.: Rethinking Range View Representation for LiDAR Segmentation. In: Int. Conf. Comput. Vis. (2023)", + "23. Kong, L., Ren, J., Pan, L., Liu, Z.: LaserMix for Semi-Supervised LiDAR Semantic Segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 21705-21715 (2023)", + "24. Kortvelesy, R., Morad, S., Prorok, A.: Permutation-Invariant Set Autoencoders with Fixed-Size Embeddings for Multi-Agent Learning. In: Int. Conf. Auton. Agents Multiagent Syst. pp. 1661-1669. International Foundation for Autonomous Agents and Multiagent Systems (2023)", + "25. Kumar, A., Poole, B.: On Implicit Regularization in $\\beta$ -VAEs. In: Int. Conf. Mach. Learn. pp. 5480-5490 (2020)", + "26. Li, E., Casas, S., Urtasun, R.: MemorySeg: Online LiDAR Semantic Segmentation with a Latent Memory. In: Int. Conf. Comput. Vis. pp. 745-754 (2023)", + "27. Li, J., Dai, H., Ding, Y.: Self-Distillation for Robust LiDAR Semantic Segmentation in Autonomous Driving. In: Eur. Conf. Comput. Vis. pp. 659-676 (2022)", + "28. Li, J., Dai, H., Han, H., Ding, Y.: MSeg3D: Multi-Modal 3D Semantic Segmentation for Autonomous Driving. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 21694-21704 (2023)", + "29. Li, L., Ismail, K.N., Shum, H.P.H., Breckon, T.P.: DurLAR: A High-Fidelity 128-Channel LiDAR Dataset with Panoramic Ambient and Reflectivity Imagery for Multi-Modal Autonomous Driving Applications. In: Int. Conf. 3D Vis. pp. 1227-1237 (2021)", + "30. Li, L., Shum, H.P.H., Breckon, T.P.: Less is More: Reducing Task and Model Complexity for Semi-Supervised 3D Point Cloud Semantic Segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)", + "31. Li, X., Li, R., Chen, G., Fu, C.W., Cohen-Or, D., Heng, P.A.: A Rotation-Invariant Framework for Deep Point Cloud Analysis. IEEE Trans. Vis. Comput. Graph. 28(12), 4503-4514 (2022)", + "32. Liang, J., Lai, R., Wong, T.W., Zhao, H.: Geometric Understanding of Point Clouds Using Laplace-Beltrami Operator. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 214-221 (2012)", + "33. Lin, M., Chen, Q., Yan, S.: Network In Network. In: Int. Conf. Learn. Represent. (2014)", + "34. Liong, V.E., Nguyen, T.N.T., Widjaja, S., Sharma, D., Chong, Z.J.: AMVNet: Assertion-based Multi-View Fusion Network for LiDAR Semantic Segmentation (2020)", + "35. Liu, J., Yu, Z., Breckon, T.P., Shum, H.P.H.: U3DS3: Unsupervised 3D Semantic Scene Segmentation. In: IEEE Winter Conf. Appl. Comput. Vis. pp. 3759-3768 (2024)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "L. Li et al.", + "bbox": [ + 271, + 114, + 344, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "36. Liu, Y., Bai, Y., Chen, R., Hou, Y., Shi, B., Li, Y., Kong, L.: PCSeg: An Open Source Point Cloud Segmentation Codebase. https://github.com/PJLab-ADG/PCSeg (2023)", + "37. Liu, Y., Chen, R., Li, X., Kong, L., Yang, Y., Xia, Z., Bai, Y., Zhu, X., Ma, Y., Li, Y., Qiao, Y., Hou, Y.: UniSeg: A Unified Multi-Modal LiDAR Segmentation Network and the OpenPCSeg Codebase. In: Int. Conf. Comput. Vis. pp. 21662–21673 (2023)", + "38. Liu, Y., Kong, L., Cen, J., Chen, R., Zhang, W., Pan, L., Chen, K., Liu, Z.: Segment Any Point Cloud Sequences by Distilling Vision Foundation Models. In: Adv. Neural Inform. Process. Syst. (2023)", + "39. Marmolin, H.: Subjective MSE measures. IEEE Trans. Syst. Man Cybern. 16(3), 486-489 (1986)", + "40. Melia, O., Jonas, E., Willett, R.: Rotation-Invariant Random Features Provide a Strong Baseline for Machine Learning on 3D Point Clouds. Trans. Mach. Learn. Res. (2023)", + "41. Nunes, L., Wiesmann, L., Marcuzzi, R., Chen, X., Behley, J., Stachniss, C.: Temporal Consistent 3D LiDAR Representation Learning for Semantic Perception in Autonomous Driving. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)", + "42. Preechakul, K., Chathee, N., Wizadwongsa, S., Suwajanakorn, S.: Diffusion Autoencoders: Toward a Meaningful and Decodable Representation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 10619-10629 (2022)", + "43. PyTorch: Scatter: PyTorch Scatter Documentation. https://pytorch-scatter.readthedocs.io/en/latest/functions/scatter.html (2023)", + "44. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 652-660 (2017)", + "45. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space. In: Adv. Neural Inform. Process. Syst. vol. 30 (2017)", + "46. Shao, Y., Tan, A., Yan, T., Sun, Z., Zhang, Y., Liu, J.: MS23D: A 3D Object Detection Method Using Multi-Scale Semantic Feature Points to Construct 3D Feature Layer (2023)", + "47. Sun, J., Xu, X., Kong, L., Liu, Y., Li, L., Zhu, C., Zhang, J., Xiao, Z., Chen, R., Wang, T., et al.: An Empirical Study of Training State-of-the-Art LiDAR Segmentation Models. arXiv:2405.14870 (2024)", + "48. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in Perception for Autonomous Driving: Waymo Open Dataset. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2446-2454 (2020)", + "49. Tang, H., Liu, Z., Zhao, S., Lin, Y., Lin, J., Wang, H., Han, S.: Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution. In: Eur. Conf. Comput. Vis. (2020)", + "50. Unal, O., Dai, D., Van Gool, L.: Scribble-Supervised LiDAR Semantic Segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)", + "51. Wang, S., Zhu, J., Zhang, R.: Meta-RangeSeg: LiDAR Sequence Semantic Segmentation Using Multiple Feature Aggregation. IEEE Robot. Autom. Lett. 7(4), 9739-9746 (2022)", + "52. Widdowson, D., Kurlin, V.: Resolving the Data Ambiguity for Periodic Crystals. In: Adv. Neural Inform. Process. Syst. vol. 35, pp. 24625-24638 (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "RAPiD-Seg Networks for 3D LiDAR Segmentation", + "bbox": [ + 395, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "53. Widdowson, D., Kurlin, V.: Recognizing Rigid Patterns of Unlabeled Point Clouds by Complete and Continuous Isometry Invariants With No False Negatives and No False Positives. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1275-1284 (2023)", + "54. Widdowson, D., Mosca, M.M., Pulido, A., Cooper, A.I., Kurlin, V.: Average Minimum Distances of Periodic Point Sets - Foundational Invariants for Mapping Periodic Crystals. MATCH Commun. Math. Comput. Chem. 87(3), 529-559 (2022)", + "55. Wu, X., Lao, Y., Jiang, L., Liu, X., Zhao, H.: Point Transformer V2: Grouped Vector Attention and Partition-Based Pooling. In: Adv. Neural Inform. Process. Syst. (2022)", + "56. Xu, J., Zhang, R., Dou, J., Zhu, Y., Sun, J., Pu, S.: RPVNet: A Deep and Efficient Range-Point-Voxel Fusion Network for LiDAR Point Cloud Segmentation. In: Int. Conf. Comput. Vis. pp. 16024–16033 (2021)", + "57. Yan, X., Gao, J., Li, J., Zhang, R., Li, Z., Huang, R., Cui, S.: Sparse Single Sweep LiDAR Point Cloud Segmentation via Learning Contextual Shape Priors from Scene Completion. In: Conf. AAAI Artif. Intell. vol. 35, pp. 3101-3109 (2021)", + "58. Yan, X., Gao, J., Zheng, C., Zheng, C., Zhang, R., Cui, S., Li, Z.: 2DPASS: 2D Priors Assisted Semantic Segmentation on LiDAR Point Clouds. In: Eur. Conf. Comput. Vis. (2022)", + "59. Ye, D., Chen, W., Zhou, Z., Xie, Y., Wang, Y., Wang, P., Foroosh, H.: LidarMultiNet: Unifying LiDAR Semantic Segmentation, 3D Object Detection, and Panoptic Segmentation in a Single Multi-task Network. In: IEEE Conf. Comput. Vis. Pattern Recog. Worksh. (2022)", + "60. Ye, D., Zhou, Z., Chen, W., Xie, Y., Wang, Y., Wang, P., Foroosh, H.: LidarMultiNet: Towards a Unified Multi-Task Network for Lidar Perception. In: Conf. AAAI Artif. Intell. vol. 37, pp. 3231-3240 (2023)", + "61. Ye, M., Wan, R., Xu, S., Cao, T., Chen, Q.: Efficient Point Cloud Segmentation with Geometry-Aware Sparse Networks. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Eur. Conf. Comput. Vis. pp. 196-212. Springer (2022)", + "62. Zhang, C., Duc, K.D., Condon, A.: Visualizing DNA Reaction Trajectories with Deep Graph Embedding Approaches. In: Adv. Neural Inform. Process. Syst. Worksh. arXiv (2022)", + "63. Zhao, H., Jiang, L., Jia, J., Torr, P., Koltun, V.: Point Transformer. In: Int. Conf. Comput. Vis. (2021)", + "64. Zhu, X., Zhou, H., Wang, T., Hong, F., Ma, Y., Li, W., Li, H., Lin, D.: Cylindrical and Asymmetrical 3D Convolution Networks for LiDAR Segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)", + "65. Zhuang, Z., Li, R., Jia, K., Wang, Q., Li, Y., Tan, M.: Perception-Aware Multi-Sensor Fusion for 3D LiDAR Semantic Segmentation. In: Int. Conf. Comput. Vis. pp. 16280-16290 (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 700 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "L. Li et al.", + "bbox": [ + 271, + 114, + 344, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/RAPiD-Seg_ Range-Aware Pointwise Distance Distribution Networks for 3D LiDAR Segmentation/b19a59c8-eca1-4aae-977c-6a71eef1622b_model.json b/2024/RAPiD-Seg_ Range-Aware Pointwise Distance Distribution Networks for 3D LiDAR Segmentation/b19a59c8-eca1-4aae-977c-6a71eef1622b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c538da7bb8e1f72986a5ea26323e4c900032102d --- /dev/null +++ b/2024/RAPiD-Seg_ Range-Aware Pointwise Distance Distribution Networks for 3D LiDAR Segmentation/b19a59c8-eca1-4aae-977c-6a71eef1622b_model.json @@ -0,0 +1,2744 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.235, + 0.142, + 0.768, + 0.187 + ], + "angle": 0, + "content": "RAPiD-Seg: Range-Aware Pointwise Distance Distribution Networks for 3D LiDAR Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.294, + 0.212, + 0.709, + 0.228 + ], + "angle": 0, + "content": "Li Li\\(^{1}\\), Hubert P. H. Shum\\(^{1}\\), and Toby P. Breckon\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.24, + 0.757, + 0.269 + ], + "angle": 0, + "content": "Department of \\(\\{\\mathrm{Computer~Science}^1\\mid\\) Engineering2}, Durham University, UK {li.li4, hubert.shum, toby.breckon}@durham.ac.uk" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.278, + 0.744, + 0.61 + ], + "angle": 0, + "content": "Abstract. 3D point clouds play a pivotal role in outdoor scene perception, especially in the context of autonomous driving. Recent advancements in 3D LiDAR segmentation often focus intensely on the spatial positioning and distribution of points for accurate segmentation. However, these methods, while robust in variable conditions, encounter challenges due to sole reliance on coordinates and point intensity, leading to poor isometric invariance and suboptimal segmentation. To tackle this challenge, our work introduces Range-Aware Pointwise Distance Distribution (RAPiD) features and the associated RAPiD-Seg architecture. Our RAPiD features exhibit rigid transformation invariance and effectively adapt to variations in point density, with a design focus on capturing the localized geometry of neighboring structures. They utilize inherent LiDAR isotropic radiation and semantic categorization for enhanced local representation and computational efficiency, while incorporating a 4D distance metric that integrates geometric and surface material reflectivity for improved semantic segmentation. To effectively embed high-dimensional RAPiD features, we propose a double-nested autoencoder structure with a novel class-aware embedding objective to encode high-dimensional features into manageable voxel-wise embeddings. Additionally, we propose RAPiD-Seg which incorporates a channel-wise attention fusion and two effective RAPiD-Seg variants, further optimizing the embedding for enhanced performance and generalization. Our method outperforms contemporary LiDAR segmentation work in terms of mIoU on SemanticKITTI (76.1) and nuScenes (83.6) datasets." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.623, + 0.741, + 0.637 + ], + "angle": 0, + "content": "Keywords: Semantic Segmentation \\(\\cdot\\) Point Cloud \\(\\cdot\\) Invariance Features" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.646, + 0.377, + 0.661 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.79, + 0.844 + ], + "angle": 0, + "content": "LiDAR-based point clouds, characterized by pointwise 3D positions and LiDAR intensity/reflectivity [3,11,29,48], play a pivotal role in outdoor scene understanding, particularly in perception systems for autonomous driving. Recent literature in the domain of 3D LiDAR segmentation has witnessed significant strides, with myriad approaches attempting pointwise 3D semantic scene labeling [47]. Leveraging color [7,37,38], range imagery [22,34], and Birds Eye View (BEV) projections [34,60], multi-modal methods integrate diverse data streams from LiDAR and other sensors to enhance feature representation and segmentation accuracy. Single-modal (LiDAR-only) methods [17,30,56,64] effectively exploit spatial data interpretation and avoid multi-modal issues, e.g., modality heterogeneity and limited sensor field of view intersection [5]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.346, + 0.128 + ], + "angle": 0, + "content": "L. Li et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.172, + 0.402, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.171, + 0.643, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.172, + 0.788, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.267, + 0.788, + 0.366 + ], + "angle": 0, + "content": "Fig. 1: Left: RAPiD exhibits excellent viewpoint invariance and geometric stability, visualizing comparable features around the vehicle door structure at varying ranges and viewpoints (feature matrix plots inset). Middle: RAPiD is distinctive in different semantic classes, as visualized by the matrices. Embedded RAPiD patterns corresponding to different points are visualized using a spectrum of colors, showcasing their capacity to represent different classes. Right: Our RAPiD-Seg achieves superior results over SoTA methods on nuScenes [3] and SemanticKITTI [2]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.384, + 0.789, + 0.461 + ], + "angle": 0, + "content": "Despite these advancements, a common challenge inherent to these methods is their poor isometric invariance, characterized by heavy reliance on solely coordinates and intensity data, which often results in suboptimal segmentation outcomes [30,64]; this is primarily due to poor translational invariance and visibility (e.g., occlusions), or sparse observations (e.g., at long range) [26], affecting data spatial distribution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.789, + 0.629 + ], + "angle": 0, + "content": "In this work, we seek features that are (1) capable of capturing the localized geometric structure of neighboring points, (2) invariant to rotation and translation, and (3) applicable in noisy LiDAR outdoor environments. While numerous methods fulfill some of these requirements individually [21, 32, 40], they fall short of addressing them all [10]. Recognizing the need for higher-level features capable of capturing local geometry while potentially incorporating LiDAR point-specific attributes (e.g., intensity and reflectivity), we instead turn our attention to pointwise Distance Distribution (PDD) features [52, 53]. They are recognized for their remarkable efficacy in providing a robust and informative geometric representation of point clouds, excelling in both rotational and translational invariance while including intricate local geometry details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.789, + 0.69 + ], + "angle": 0, + "content": "However, directly employing PDD features is impractical due to their high-dimensional nature, and their significant memory and storage requirements for large-scale point clouds. PDD also overlooks local features, as including distant points subsequently dilutes the focus on immediate neighborhoods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.79, + 0.842 + ], + "angle": 0, + "content": "As an enabler, we propose the Range-Aware Pointwise Distance Distribution (RAPiD-Seg) solution for LiDAR segmentation. It utilizes invariance of RAPiD features to rigid transformations and point cloud sparsity variations, while concentrating on compact features within specific local neighborhoods. Specifically, our method leverages inherent LiDAR isotropic radiation and semantic categorization for enhanced local representations while reducing computational burdens. Moreover, our formulation computes a 4D distance, incorporating both 3D geometric and reflectivity differences to enhance semantic segmentation fidelity. To compress high-dimensional RAPiD features into tractable voxel-wise embeddings, we propose a novel embedding approach with our class-aware double-nested AutoEn" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.396, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "RAPiD-Seg Networks for 3D LiDAR Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "coder (AE) module. We further incorporate a channel-wise attention fusion and a two-stage training strategy to optimize the AE independently before integrating into the full network, enhancing performance and generalization capability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.785, + 0.237 + ], + "angle": 0, + "content": "We conduct extensive experiments on SemanticKITTI [2] and nuScenes [3] datasets, upon which our approach surpasses the state-of-the-art (SoTA) segmentation performance." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.238, + 0.786, + 0.253 + ], + "angle": 0, + "content": "Overall, as shown in Fig. 2, our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.255, + 0.786, + 0.315 + ], + "angle": 0, + "content": "A novel Range-Aware Pointwise Distance Distribution feature (RAPiD) for 3D LiDAR segmentation that ensures robustness to rigid transformations and viewpoints through isometry-invariant metrics within specific Regions of Interest, i.e., intra-ring (R-RAPiD) and intra-class (C-RAPiD)." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.317, + 0.786, + 0.361 + ], + "angle": 0, + "content": "A novel method for embedding RAPiD with class-aware double nested AE (RAPiD AE) to optimize the embedding of high-dimensional features, balancing efficiency and fidelity." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.362, + 0.786, + 0.409 + ], + "angle": 0, + "content": "A novel open-source network architecture RAPiD-Seg \\(^{1}\\) and supporting training methodology, to enable modular LiDAR segmentation that achieves SoTA performance on SemanticKITTI (mIoU: 76.1) and nuScenes (mIoU: 83.6)." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.255, + 0.786, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.428, + 0.784, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.575, + 0.785, + 0.631 + ], + "angle": 0, + "content": "Fig. 2: Our proposed architecture for 3D segmentation framework leverages RAPiD features from the point cloud. We encode pointwise features into voxel-wise embeddings via the voxel encoder and multiple RAPiD AutoEncoders (RAPiD AE). After attention-based feature fusion, these fused embeddings go through the backbone network for segmentation results." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.654, + 0.388, + 0.67 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.681, + 0.788, + 0.818 + ], + "angle": 0, + "content": "LiDAR-Based Semantic Segmentation [7,12,17,20,22,23,27,30,34,37,38,49, 50,56-58,60,61,64,65] is fundamental for LiDAR-driven scene perception, aiming to label each point in a point cloud sequence. The majority of the approaches [17, 38,49,57,64] solely rely on the point-based features of the point cloud, such as SPVCNN [49] which introduces a point-to-voxel branch, using combined pointvoxel features for segmentation. Cylinder3D [64] proposes cylindrical partitioning with a UNet [9] backbone variant. LiM3D [30] utilizes coordinates combined with surface reflectivity attributes. Overall, such prior work relies solely on pointbased features such as coordinates and intensity of the points, lacking an effective" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.721, + 0.84 + ], + "angle": 0, + "content": "1 The code is publicly available at: https://github.com/l1997i/rapid_seg." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.346, + 0.128 + ], + "angle": 0, + "content": "L. Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.192 + ], + "angle": 0, + "content": "fusion mechanism, resulting in suboptimal performance [6,28,51]. They are also susceptible to changes in viewpoint, distance, and point sparsity [26] due to the lack of isometry across all inter-point distances." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.788, + 0.508 + ], + "angle": 0, + "content": "Pointwise Distance Distribution (PDD) captures the local context of each point in a unit cell by enumerating distances to neighboring points in order. It is an isometry invariant proposed by Widdowson & Kurlin [52] to resolve the data ambiguity for periodic crystals, demonstrated through extensive pairwise comparisons across atomic 3D clouds from high-level periodic crystals of periodic structures [52-54]. Though the effectiveness of PDD in periodic crystals and atomic clouds has been proved by the aforementioned studies, to date, no work has applied PDD features to outdoor 3D point clouds. In outdoor settings, common invariant features [21,31,32,40] and Azimuth-normalized method [4] often face issues due to the irregular and sparse nature of the data, which is compounded by increased noise and environmental complexities [31,32]. Furthermore, the computational demands make them less suitable for the vast scale of outdoor settings [21,40]. For instance, Melia et al. [40] introduce a rotation-invariant feature that struggles with generalization across diverse point cloud densities and scales due to its computational cost and susceptibility to outdoor noise. Recognizing these limitations, we identify an opportunity to leverage the PDD features in a new domain, where representing the local context of neighboring points in a transformation invariant and geometrically robust manner is crucial. Drawing from the advantages of the PDD design, we propose the RAPiD feature, tailored specifically for LiDAR-based point clouds, to capture the localized geometry of neighboring structures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.788, + 0.751 + ], + "angle": 0, + "content": "Point Cloud Representations and Embeddings are fundamental for LiDAR-driven tasks. 3D point cloud can be represented either as point-based [44,45,55,63] or voxel-based [17,35,50,64], each serving different purposes rooted in their unique characteristics. Voxel-based representation involves dividing the 3D space into grids of voxels, which provides a structured way of handling sparse and irregular LiDAR data [9, 13, 14, 64]. Previous work often utilizes Feedforward Neural Network (FNN) [17, 50, 64] to facilitate the conversion of pointwise features into structured voxel-wise features by projecting them onto voxel grids and generating dense embeddings. The primary drawback of FNN conversion is the potential loss of fine-grained details inherent in the original point cloud [15]. FNN may also struggle with the high dimensionality and sparsity of voxelized data, leading to computational inefficiency [46]. Our method addresses these issues by partitioning the point cloud into voxel grids and applying a novel class-ware double nested AE with self-attention mechanisms [15]. This approach not only retains more detailed information by focusing on local structures but also introduces high-dimensional feature compression, thereby improving computational efficiency." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.767, + 0.67, + 0.786 + ], + "angle": 0, + "content": "3 Range-Aware Pointwise Distances (RAPiD)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In the context of LiDAR segmentation, a principal challenge lies in the lack of robustness against rigid transformations such as rotation and translation, as well as against variations in viewpoint, points sparsity, and occlusion [5, 26]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.396, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "RAPiD-Seg Networks for 3D LiDAR Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "Traditional methods predominantly leverage 3D point coordinates to furnish spatial information [17,50,64]; however, they may be inadequate in the scenes with poor visibility (e.g., occlusions) or sparse observations (e.g., at long range) [26]. Such reliance on coordinates alone could lead to inaccuracies in recognizing object transformations or occlusions [26]. Data augmentation, e.g., random geometric transformation can improve robustness under rigid transformations [50,64], but fail to guarantee comprehensive coverage of all potential transformations, resulting in vulnerability to previously unseen variations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.288, + 0.341, + 0.302 + ], + "angle": 0, + "content": "3.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.311, + 0.789, + 0.461 + ], + "angle": 0, + "content": "To achieve a transformation-invariant 3D data representation, we observe that distances within rigid bodies (e.g., vehicles, roads, and buildings) remain constant under rigid transformations. In light of this, we focus on assimilating the principle of the isometric invariant into the LiDAR-driven point cloud perception. Specifically, we delve into the PDD [52] — an isometry invariant that quantifies distances between adjacent points. For outdoor point clouds, vanilla PDD features are computationally intensive and susceptible to noise and sparsity. Our Range-Aware Pointwise Distance Distribution (RAPiD), specifically designed for LiDAR data, instead calculates the PDD features for each point within specific Regions of Interest (RoI), which are typically associated with the intrinsic structure of LiDAR data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.553 + ], + "angle": 0, + "content": "Mathematical Formulation: Given a fixed number \\( k > 0 \\) representing the fixed number of point neighbors and a \\( u \\)-point cluster \\( P_{\\mathrm{RoI}} \\) comprising no fewer than \\( k \\) points based on RoI, the RAPiD is a \\( u \\times k \\) matrix, which retains both spatial distances and LiDAR reflectivity [30] disparities between points. RAPiD are adapted to LiDAR sparsity at different distances by using range-specific parameters \\( k_{\\mathrm{close}} \\), \\( k_{\\mathrm{mid}} \\), \\( k_{\\mathrm{far}} \\) for close, mid, and far ranges, respectively." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.553, + 0.588, + 0.568 + ], + "angle": 0, + "content": "The \\(k\\)-point RAPiD in region \\(P_{\\mathrm{RoI}}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.323, + 0.569, + 0.787, + 0.592 + ], + "angle": 0, + "content": "\\[\n\\operatorname {R A P i D} \\left(P _ {\\text {R o I}}; k\\right) = \\operatorname {s o r t} \\left(\\left[ \\operatorname {s o r t} \\left(\\left[ \\boldsymbol {\\rho} _ {j, 1}, \\dots , \\boldsymbol {\\rho} _ {j, k} \\right]\\right) \\right] _ {j = 1} ^ {u}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.593, + 0.558, + 0.611 + ], + "angle": 0, + "content": "\\(\\forall l \\in \\{1, \\dots, k\\}, j \\in \\{1, \\dots, u\\}, \\pmb{\\rho}_{j,l}\\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.363, + 0.617, + 0.787, + 0.637 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\rho} _ {j, l} = \\left\\| \\left[ \\boldsymbol {p} _ {j} - \\boldsymbol {p} _ {j, l}, \\quad g (r _ {j}) - g (r _ {j, l}) \\right] \\right\\| _ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.788, + 0.72 + ], + "angle": 0, + "content": "where \\(\\pmb{p}_j\\) and \\(\\pmb{p}_{j,l}\\) denote the 3D coordinates of the \\(j\\)-th point and its \\(l\\)-th nearest neighbor within \\(P_{\\mathrm{RoI}}\\), respectively; \\(r_j\\) and \\(r_{j,l}\\) represent the reflectivity values of \\(\\pmb{p}_j\\) and \\(\\pmb{p}_{j,l}\\), correspondingly; \\(\\| \\cdot \\| _2\\) is the Euclidean norm. \\(g:\\mathbb{R}\\to [D_{\\min},D_{\\max}]\\) is the reflectivity mapping function that maps the numerical range of reflectivity onto a consistent scale with the range of Euclidean distances between points:" + }, + { + "type": "equation", + "bbox": [ + 0.32, + 0.727, + 0.787, + 0.759 + ], + "angle": 0, + "content": "\\[\ng (r) = \\left(\\frac {r - r _ {\\operatorname* {m i n}}}{r _ {\\operatorname* {m a x}} - r _ {\\operatorname* {m i n}}}\\right) \\left(D _ {\\operatorname* {m a x}} - D _ {\\operatorname* {m i n}}\\right) + D _ {\\operatorname* {m i n}}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.314, + 0.763, + 0.787, + 0.785 + ], + "angle": 0, + "content": "\\[\nD _ {\\min } = \\min _ {j, l} \\| \\boldsymbol {p} _ {j} - \\boldsymbol {p} _ {j, l} \\| _ {2}, \\quad D _ {\\max } = \\max _ {j, l} \\| \\boldsymbol {p} _ {j} - \\boldsymbol {p} _ {j, l} \\| _ {2}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.793, + 0.787, + 0.824 + ], + "angle": 0, + "content": "where \\([D_{\\min}, D_{\\max}]\\) is the range of the Euclidean norms of coordinate differences for all considered point pairs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.824, + 0.787, + 0.869 + ], + "angle": 0, + "content": "The 4D distance in Eq. (2) integrates material reflectivity [30] into RAPiD features, enhancing feature representation and aiding in the discrimination of various materials and surfaces, which is crucial for accurate semantic segmentation." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.345, + 0.127 + ], + "angle": 0, + "content": "L. Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.221 + ], + "angle": 0, + "content": "For convenience to facilitate feature normalization and data alignment, we arrange RAPiD lexicographically by sorting Eq. (1), where \\(\\mathrm{sort}(\\cdot)\\) on the inner and outer brackets sorts the elements \\(\\rho_{j,l}\\) within each row \\(j\\), and the sorted rows based on their first differing elements, both in ascending order [52]. It treats RAPiD matrices with the same geometric structure but different orders as identical." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.785, + 0.268 + ], + "angle": 0, + "content": "To improve the data robustness, we normalize \\(\\rho_{j,l}\\) after excluding outliers exceeding a distance threshold \\(\\delta\\), substituting these entries with the maximum of the normalized distribution to represent significant inter-point distances." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.277, + 0.544, + 0.291 + ], + "angle": 0, + "content": "3.2 Intra-Ring and Intra-Class RAPiD" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.303, + 0.785, + 0.438 + ], + "angle": 0, + "content": "To enhance the RAPiD ability to capture local context, we propose R-RAPiD and C-RAPiD, based on the inherent structure of LiDAR data. The key benefit of C-RAPiD is its capacity to underscore the inherent traits of points within the same semantic class, achieved based on semantic labels before distance calculation. As a complementary, R-RAPiD is versatile, regardless of semantic label availability. Intra-Ring RAPiD (R-RAPiD) is a specialized variant of RAPiD, which confines the RoI to the ring encompassing the anchor points, thereby serving to economize on computational overhead while capitalizing on the inherent structural characteristics of LiDAR data (Fig. 2 ①)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.791, + 0.53 + ], + "angle": 0, + "content": "Given LiDAR sensors with fixed beam counts [2,3,29,48], e.g., \\(B \\in \\{32,64,128\\}\\), beams radiate isotropically around the vehicle at set angles. We segment the \\(m\\)-point LiDAR point cloud \\(C_{B,s}\\) by beams, where \\(m = s \\times B\\), with \\(s\\) representing measurements within a scan cycle and \\(B\\) the number of laser beams. Suppose \\(\\pmb{p}_i\\) with Cartesian coordinate \\((x_i, y_i, z_i)\\) is a point within \\(C_{B,s}\\), its transformation to cylindrical coordinates \\((\\theta_i, \\phi_i)\\) is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.352, + 0.543, + 0.786, + 0.559 + ], + "angle": 0, + "content": "\\[\n\\theta_ {i} = \\left\\lfloor \\arctan 2 \\left(y _ {i}, x _ {i}\\right) / \\Delta \\theta \\right\\rfloor , \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.35, + 0.562, + 0.786, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\phi_ {i} = \\left\\lfloor \\arcsin \\left[ z _ {i} \\left(x _ {i} ^ {2} + y _ {i} ^ {2} + z _ {i} ^ {2}\\right) ^ {- 1 / 2} \\right] / \\Delta \\phi \\right\\rfloor , \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.785, + 0.627 + ], + "angle": 0, + "content": "where \\(\\varDelta\\theta\\) and \\(\\varDelta\\phi\\) denote the mean angular resolutions horizontally and vertically between adjacent beams, and \\(\\arctan 2(\\cdot)\\) is the 2-argument \\(\\arctan (\\cdot)\\) function." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.785, + 0.674 + ], + "angle": 0, + "content": "\\(C_{B,s}\\) is then partitioned to \\(B\\) rings based on \\(\\phi_b\\), i.e., \\(C_{B,s} = \\bigcup_{b=0}^{B-1} R_b\\), where each ring \\(R_b = \\{(\\theta_{b,i}, \\phi_{b,i}) \\mid \\phi_{b,i} = \\phi_b, \\forall i \\leq m\\}\\); \\(\\bigcup\\) signifies the cumulative union R-RAPiD is thus defined as the cumulative concatenate \\(\\oplus\\) of RAPiD in each ring \\(R_b\\)." + }, + { + "type": "equation", + "bbox": [ + 0.341, + 0.681, + 0.786, + 0.701 + ], + "angle": 0, + "content": "\\[\n\\text {R - R A P i D} \\left(C _ {B, s}; k\\right) := \\bigoplus_ {b = 0} ^ {B - 1} \\text {R A P i D} \\left(R _ {b}; k\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.709, + 0.784, + 0.753 + ], + "angle": 0, + "content": "Intra-Class RAPiD (C-RAPiD) concentrates on extracting point features within the confines of each semantic class (Fig. 2 ②). It is complementary to R-RAPiD, preserving the embedding fidelity within individual semantic class." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.754, + 0.785, + 0.784 + ], + "angle": 0, + "content": "Specifically, C-RAPiD concatenates RAPiD for each point \\(\\pmb{p}_j\\) and its corresponding semantic label \\(y_{j}\\) within semantic class \\(i\\), across \\(N_{c}\\) total classes:" + }, + { + "type": "equation", + "bbox": [ + 0.322, + 0.794, + 0.786, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\text {C - R A P i D} \\left(\\left\\{C _ {B, s}, \\mathcal {Y} \\right\\}; k\\right) := \\bigoplus_ {i = 0} ^ {N _ {c} - 1} \\operatorname {R A P i D} \\left(S _ {i}; k\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.826, + 0.571, + 0.844 + ], + "angle": 0, + "content": "where \\(S_{i} = \\{\\pmb{p}_{j}\\mid y_{j} = i,\\forall j\\leq m\\}\\) \\(\\mathcal{V} = [y_j]_{j = 1}^m\\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.396, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "RAPiD-Seg Networks for 3D LiDAR Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.436, + 0.163 + ], + "angle": 0, + "content": "4 RAPiD Embedding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.784, + 0.223 + ], + "angle": 0, + "content": "The direct utilization of high-dimensional RAPiD features imposes a substantial computational burden. Efficient processing of large data volumes requires methods that can condense dimensionality while preserving data integrity [62]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.224, + 0.787, + 0.299 + ], + "angle": 0, + "content": "We explore data embedding methods to reduce computational costs while providing feature learning capacity. AE is efficient in embedding high-dimensional features utilizing the reconstruction loss [16, 42]. However, conventional AE, limited by their sensitivity to input order and reliance on fixed-size misinputs [24], struggle to align with the unordered and variable size inherent in point clouds [15]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.3, + 0.786, + 0.359 + ], + "angle": 0, + "content": "We propose a double nested AE structure with a novel class-aware embedding objective, using the Voxel-based Set Attention (VSA) module [15] as a building block. This facilitates superior contextual awareness and adaptability to unordered and variably-sized point cloud data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.361, + 0.788, + 0.525 + ], + "angle": 0, + "content": "Nested RAPiD AE, illustrated in Fig. 2 ③ and Fig. 3, is composed of two modules: the outer module is constituted by a VSA AE, primarily responsible for the point-to-voxel conversion. Within the inner module, which focuses on voxel-wise representation, we employ an additional AE specifically designed for dimensionality reduction. Specifically, considering an input of \\( m \\)-point RAPiD features \\( G \\in \\mathbb{R}^{m \\times d} \\), where each point encompasses \\( d \\) features. We compress pointwise \\( G \\) into a voxel-wise representation \\( H^v \\in \\mathbb{R}^{c \\times l \\times d} \\) by an outer VSA Voxel Encoder (with \\( c \\) reduced voxel-wise dimension and \\( l \\) latent codes), then further reduces it to a compressed embedding \\( \\hbar \\in \\mathbb{R}^{c \\times l \\times d'} \\) with \\( d' \\) features via an inner Encoder. The inner and outer VSA Point Decoders then reconstruct the output feature set \\( \\hat{G} \\) akin to \\( G \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.527, + 0.788, + 0.588 + ], + "angle": 0, + "content": "Inner Voxel-Wise Representation is conducted via an inner AE, which takes \\( H^v \\) from outer VSA AE as the input, yielding a lower-dimensional embedding \\( \\hbar \\) (dimension reduced from \\( d \\) to \\( d' \\)). This process involves \\( u \\times \\) convolutional layers and a batch normalization layer [19]:" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.6, + 0.787, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\hbar^ {(i)} = \\operatorname {B a t c h N o r m} \\circ \\operatorname {C o n v} \\left(\\hbar^ {(i - 1)}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.706 + ], + "angle": 0, + "content": "where \\(\\hbar^{(0)} = H^v\\), \\(\\hbar^{(u - 1)} = \\hbar\\). The Convolutional Feed-Forward Network (ConvFFN) then promotes voxel-level information exchange, enhancing spatial feature interactivity. It maps reduced hidden features \\(\\hbar\\) to a 3D sparse tensor, indexed by voxel coordinates \\(X^v\\), and employs dual depth-wise convolutions (DwConv) for spatial interactivity:" + }, + { + "type": "equation", + "bbox": [ + 0.336, + 0.718, + 0.787, + 0.736 + ], + "angle": 0, + "content": "\\[\n\\hat {h} = \\operatorname {D w C o n v} ^ {(2)} \\circ \\zeta \\circ \\operatorname {D w C o n v} ^ {(1)} \\left(\\operatorname {S p T} \\left(\\hbar , X ^ {v}\\right)\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.748, + 0.788, + 0.794 + ], + "angle": 0, + "content": "where \\(\\zeta\\) and \\(\\mathrm{SpT}(\\cdot)\\) represents the non-linear activation and sparse tensor construction. Subsequently, \\(\\hat{\\hbar}\\in \\mathbb{R}^{c\\times l\\times d'}\\) is reconstructed into \\(\\hat{H}^v\\in \\mathbb{R}^{c\\times l\\times d}\\) through an inner Decoder, consisting of multiple DeConv Layers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Outer VSA AE consists of a voxel encoder and point encoder. The voxel encoder project RAPiD features into key-value spaces, forming \\( K \\) and \\( V \\), followed by a cross-attention mechanism with a latent query \\( L \\), producing an attention" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.128 + ], + "angle": 0, + "content": "L. Li et al." + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.146, + 0.716, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.281, + 0.789, + 0.325 + ], + "angle": 0, + "content": "Fig. 3: Our RAPiD AE consists of an Encoder, Convolution Layer, and Decoder module, aiming to reproduce the input features and generate the compressed voxel-wise RAPiD representation \\( \\hbar \\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.334, + 0.789, + 0.365 + ], + "angle": 0, + "content": "matrix \\(A\\). The voxel-wise representation \\(H^v\\) is the scatter sum [43] of pointwise \\(H\\) to aggregate the value vectors \\(V\\):" + }, + { + "type": "equation", + "bbox": [ + 0.314, + 0.375, + 0.788, + 0.393 + ], + "angle": 0, + "content": "\\[\nH ^ {v} = \\operatorname {S u m} _ {\\text {s c a t t e r}} (H, I ^ {v}), \\quad H = \\tilde {A} ^ {\\top} V, \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.325, + 0.396, + 0.788, + 0.415 + ], + "angle": 0, + "content": "\\[\n\\tilde {A} = \\operatorname {S o f t m a x} _ {\\text {s c a t t e r}} \\left(K L ^ {\\top}, I ^ {v}\\right), \\quad (Q, V) = \\operatorname {P r o j} (G), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.424, + 0.79, + 0.485 + ], + "angle": 0, + "content": "where \\(\\operatorname{Proj}(\\cdot)\\) represents the linear projection, and \\(I^v\\) is the voxel indices. The Point Decoder reconstructs the output set from the enriched hidden features \\(\\hat{H}^v\\). We start by broadcasting \\(\\hat{H}^v\\) based on \\(I^v\\), resulting in \\(\\hat{H} \\in \\mathbb{R}^{m \\times l \\times d}\\). The output \\(\\hat{G}\\) is analogous to the operational paradigm of Voxel Encoder:" + }, + { + "type": "equation", + "bbox": [ + 0.356, + 0.495, + 0.788, + 0.517 + ], + "angle": 0, + "content": "\\[\n\\hat {G} = \\tilde {A} ^ {*} ^ {\\top} V ^ {*}, \\quad \\tilde {A} ^ {*} = \\operatorname {S o f t m a x} \\left(A ^ {*}\\right), \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.351, + 0.519, + 0.788, + 0.541 + ], + "angle": 0, + "content": "\\[\nA ^ {*} = \\left[ K _ {i} ^ {*} Q _ {i} ^ {\\top} \\right] _ {i = 1} ^ {m}, \\quad \\left(K ^ {*}, V ^ {*}\\right) = \\operatorname {P r o j} (\\hat {\\mathcal {H}}). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.549, + 0.789, + 0.624 + ], + "angle": 0, + "content": "Class-Aware Embedding Objective addresses the issue of non-uniqueness in embeddings produced by the AE, where various distinct inputs yield approximately the same embedding, leading to inaccurate representation [1, 25]. Our objective aims to facilitate the generation of AE embeddings that demonstrate robust semantic class discriminability." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.624, + 0.789, + 0.666 + ], + "angle": 0, + "content": "Specifically, we introduce a novel class-aware contrastive loss \\(\\mathcal{L}_{\\mathrm{contr}}\\) in Eq. (15), which aims to maximize the distance between embeddings of different semantic classes while minimizing the distance of the same class." + }, + { + "type": "equation", + "bbox": [ + 0.321, + 0.671, + 0.787, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {c o n t r}} = \\frac {1}{m} \\sum_ {i = 1} ^ {m} \\left[ \\frac {1}{| P (i) |} \\sum_ {p \\in P (i)} \\mathrm {R e L U} (\\alpha - \\mathrm {s i m} (H _ {i}, H _ {p})) \\right. \\\\ \\left. + \\frac {1}{| N (i) |} \\sum_ {n \\in N (i)} \\operatorname {R e L U} \\left(\\sin \\left(H _ {i}, H _ {n}\\right) - \\alpha\\right) \\right], \\tag {15} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.321, + 0.764, + 0.787, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c o n}} = \\frac {1}{m \\cdot d} \\sum_ {i = 1} ^ {m} \\sum_ {j = 1} ^ {d} \\left(G _ {i, j} - \\hat {G} _ {i, j}\\right) ^ {2}, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "where \\( i \\) represents the point index; \\( P(i) \\) and \\( N(i) \\) are the indices of the nearest point in the same (Positive) and different class (Negative) relative to point \\( i \\)," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.396, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "RAPiD-Seg Networks for 3D LiDAR Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.191 + ], + "angle": 0, + "content": "respectively; \\(\\mathrm{sim}(\\cdot)\\) is a similarity measure; \\(\\alpha\\) is a margin controlling separation between classes. The pointwise representations \\(H_{i}\\), \\(H_{p}\\), and \\(H_{n}\\) are broadcasted from \\(\\hbar\\) based on voxel indices \\(I^{v}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.785, + 0.239 + ], + "angle": 0, + "content": "We further combine this with the MSE [39] reconstruction loss in Eq. (16). The overall loss function, with \\(\\lambda\\) balancing the fidelity of reconstruction with the distinctiveness of the embeddings, is formulated as: \\(\\mathcal{L}_{\\mathrm{total}} = \\mathcal{L}_{\\mathrm{recon}} + \\lambda \\mathcal{L}_{\\mathrm{contr}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.253, + 0.644, + 0.271 + ], + "angle": 0, + "content": "5 RAPiD-Seg for 3D LiDAR Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.785, + 0.405 + ], + "angle": 0, + "content": "We present RAPiD-Seg, a 3D LiDAR segmentation network leveraging RAPiD features. It incorporates multiple complementary features via the channel-wise fusion mechanism [18]. RAPiD-Seg takes point cloud as input and performs single-modal (LiDAR-only) 3D semantic segmentation. Specifically, the input point cloud has three types of point-wise features: (a) coordinates-based features \\( F_{C} = \\{p_{C} \\mid p_{C} = (x,y,z)\\} \\); (b) intensity-based features \\( F_{I} = \\{p_{I} \\mid p_{I} = (I,J)\\} \\), where \\( I \\) and \\( J \\) are intensity and reflectivity; (c) RAPiD features \\( F_{R} = \\{p_{R} \\mid p_{R} = \\mathrm{RAPiD}(P_{\\mathrm{RoI}};k)\\} \\). The RoI varies depending on whether it is R- or C-RAPiD." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.785, + 0.51 + ], + "angle": 0, + "content": "As shown in Fig. 2, \\( F_{C} \\oplus F_{I} \\) and \\( F_{R} \\) are fed to a VSA voxel encoder 4 and RAPiD AE 6 for voxelization, resulting in voxel-wise representation \\( E_{C} \\), \\( E_{I} \\) and \\( E_{R} \\). Subsequently, we incorporate the complementary voxel-wise features via the channel-wise fusion mechanism (Sec. 5.1). The backbone net takes in them for LiDAR segmentation prediction. Additionally, we introduce two effective RAPiD-Seg variants to expedite the convergence of the extensive segmentation network towards an optimal solution." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.531, + 0.627, + 0.545 + ], + "angle": 0, + "content": "5.1 RAPiD Channel-Wise Fusion with Attention" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.785, + 0.628 + ], + "angle": 0, + "content": "To combine various LiDAR point attributes, a highly efficient method for feature fusion is still a research topic [30, 50]. Current methods, which typically concatenate different LiDAR attributes (e.g., intensity [38, 41, 59], reflectivity [30], PLS [50]) face issues with increased dimensionality and complexity, and the need for balancing weights to avoid biased training towards dominant attributes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.785, + 0.689 + ], + "angle": 0, + "content": "We thus propose a novel RAPiD Fusion with Attention module (FuAtten), pioneering application of channel-wise attention mechanism [18] in LiDAR point attribute fusion. With FuAtten, we effectively fuse \\( E_{C} \\), \\( E_{I} \\) and \\( E_{R} \\), emphasizing informative features adaptively while suppressing less relevant ones across different channels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.785, + 0.755 + ], + "angle": 0, + "content": "As shown in Fig. 2 ⑤ and Fig. 4, we initially concatenate \\( E_{C} \\), \\( E_{I} \\) and \\( E_{R} \\) along the voxel dimension, resulting in \\( \\mathbf{E} \\). In the squeeze operation, Global Average Pooling (GAP) [33] condenses each channel \\( f' \\) of the tensor into a representative channel descriptor. These descriptors form a channel-level embedding \\( \\mathbf{z} = [z_{f'}]_{f' = 1}^{f^*} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.323, + 0.765, + 0.786, + 0.808 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} = \\left[ \\operatorname {G A P} \\left(\\mathbf {E} _ {:,:: f ^ {\\prime}}\\right) \\right] _ {f ^ {\\prime} = 1} ^ {f ^ {*}} = \\frac {1}{c \\times l} \\left[ \\sum_ {c ^ {\\prime} = 0} ^ {c - 1} \\sum_ {l ^ {\\prime} = 0} ^ {l - 1} \\mathbf {E} _ {c ^ {\\prime}, l ^ {\\prime}, f ^ {\\prime}} \\right] _ {f ^ {\\prime} = 1} ^ {f ^ {*}}, \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.841 + ], + "angle": 0, + "content": "where \\( f^{*} \\) is the dimension of the concatenated feature. A channel-wise attention \\( \\pmb{a}_{\\mathbf{z}} \\in \\mathbb{R}^{1 \\times 1 \\times f^{*}} \\) is then computed through the excitation step: \\( \\pmb{a}_{\\mathbf{z}} = \\)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.346, + 0.128 + ], + "angle": 0, + "content": "L. Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "\\(\\sigma (\\mathbf{W}_2\\delta (\\mathbf{W}_1\\mathbf{z}))\\) , where \\(\\delta\\) and \\(\\sigma\\) represent the ReLU and Sigmoid activation. With each element in \\(a_{\\mathbf{z}}\\) reflects the attention allocated to the corresponding feature of the voxel, the channel-wise fused features \\(\\mathbf{E}^{\\prime} = a_{\\mathbf{z}}\\cdot \\mathbf{E}\\)" + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.2, + 0.672, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.29, + 0.288, + 0.712, + 0.302 + ], + "angle": 0, + "content": "Fig. 4: Our feature fusion module with channel-wise attention." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.32, + 0.649, + 0.336 + ], + "angle": 0, + "content": "5.2 RAPiD-Seg Architectures for 3D Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.345, + 0.788, + 0.407 + ], + "angle": 0, + "content": "The complexity inherent in 3D LiDAR-driven networks typically exacerbates the end-to-end training process [17,30,50], since the extensive parameter space increases the propensity for overfitting, slow convergence, and the possibility of settling into local minima [30]." + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.42, + 0.744, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.52, + 0.788, + 0.551 + ], + "angle": 0, + "content": "Fig. 5: Two variants of RAPiD-Seg. (a) R-RAPiD-Seg utilizes R-RAPiD features, and (b) C-RAPiD-Seg utilizes both R- and C-RAPiD features for better performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.629 + ], + "angle": 0, + "content": "As shown in Fig. 5, we propose two novel and effective variants of RAPiD-Seg for quicker and better performance. R-RAPiD-Seg involves AE training from scratch, followed by C-RAPiD-Seg utilizing C-RAPiD features with the pre-trained AE and backbone for enhanced performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.765 + ], + "angle": 0, + "content": "R-RAPiD-Seg. We first construct the lightweight R-RAPiD-Seg (Fig. 5 (a)) for fast 3D segmentation. R-RAPiD-Seg adopts an early fusion scheme utilizing LiDAR original points and R-RAPiD features. Specifically, 3D coordinate \\( F_{C} \\) and intensity-based features \\( F_{I} \\) are voxelized to voxel-wise representations \\( R_{v} = E_{C} \\oplus E_{I} \\) based on VSA voxelization [15]. We get the compressed \\( d' \\)-dimension voxel-wise R-RAPiD representations \\( \\hbar \\) (Sec. 4) from RAPiD AE. \\( \\hbar \\) is then fused into the voxel-wise representations \\( R_{v*} = \\mathrm{FuAtten}(R_v,\\hbar) \\), where \\( \\mathrm{FuAtten}(\\cdot) \\) is Fusion with Attention in Sec. 5.1. The fused voxel-wise features are taken into the backbone network for 3D segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "C-RAPiD-Seg. To facilitate the embedding fidelity within individual semantic classes, we also design a class-aware framework with C-RAPiD features, i.e., C-RAPiD-Seg (Fig. 5 (b)). Specifically, we fuse the voxel-wise C-RAPiD \\(\\hbar_C\\) into the representations \\(R_{v*} = \\mathrm{FuAtten}(R_v,\\hbar \\oplus \\hbar_C)\\). C-RAPiD requires class labels to compute the features regarding the semantic categories. Since the ground-truth" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.396, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RAPiD-Seg Networks for 3D LiDAR Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "labels are missing during the test time, we generate the reliable pseudo labels \\(\\tilde{\\mathcal{V}}\\) with a pretrained R-RAPiD-Seg based on confidence (Fig. 5 (b) blue dotted arrow). Subsequently, the pseudo voxel-wise C-RAPiD \\(\\hbar_{C}\\) are generated through \\(\\tilde{\\mathcal{V}}\\). We fuse them into the voxel-wise representations for 3D segmentation with the bonebone network." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.234, + 0.358, + 0.249 + ], + "angle": 0, + "content": "6 Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.265, + 0.787, + 0.311 + ], + "angle": 0, + "content": "Following the popular practice of LiDAR segmentation methods [17,37,64], we evaluate our proposed RAPiD-Seg network against SoTA 3D LiDAR segmentation approaches on the SemanticKITTI [2] and nuScenes [3] datasets." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.321, + 0.429, + 0.336 + ], + "angle": 0, + "content": "6.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.788, + 0.448 + ], + "angle": 0, + "content": "Datasets: SemanticKITTI [2] comprises 22 point cloud sequences, with sequences 00-10, 08, and 11-21 for training, validation, and testing. 19 classes are chosen for training and evaluation by merging classes with similar motion statuses and discarding sparsely represented ones. Meanwhile, nuScenes [3] has 1000 driving scenes; 850 are for training and validation, with the remaining 150 for testing. 16 classes are used for LiDAR semantic segmentation, following the amalgamation of akin classes and the removal of rare ones." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.79, + 0.525 + ], + "angle": 0, + "content": "Evaluation Protocol: Following the popular practice of [2,36,64], we adopt the Intersection-over-Union (IoU) of each class and mean IoU (mIoU) of all classes as the evaluation metric. The IoU of class \\( i \\) is defined as \\( \\mathrm{IoU}_i = \\mathrm{TP}_i / (\\mathrm{TP}_i + \\mathrm{FP}_i + \\mathrm{FN}_i) \\), where \\( \\mathrm{TP}_i \\), \\( \\mathrm{FP}_i \\) and \\( \\mathrm{FN}_i \\) denote the true positive, false positive and false negative of class \\( i \\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.526, + 0.79, + 0.692 + ], + "angle": 0, + "content": "Implementation Details: We construct the point-voxel backbone based on the Minkowski-UNet34 [8] (re-implemented by PCSeg [36] codebase), which is the open-access backbone with SoTA results to date. Before the AE training stage, we first generate pointwise RAPiD features (Sec. 3), which yield three RAPiD outputs for each frame, corresponding to different \\( k \\) values based on range. Notably, this stage does not impose additional burdens on the subsequent overall training. The number of maximum training epochs for AE and the whole network is set as 100 and the initial learning rate is set as \\( 10^{-3} \\) with SGD optimizer. We use 2 epochs to warm up the network and adopt the cosine learning rate schedule for the remaining epochs. All experiments are conducted on \\( 4 \\times \\) NVIDIA A100 GPUs (\\( 1 \\times \\) for inference)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.71, + 0.442, + 0.724 + ], + "angle": 0, + "content": "6.2 Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.787, + 0.795 + ], + "angle": 0, + "content": "In Tab. 1 and Tab. 2, we showcase the performance of our RAPiD-Seg LiDAR segmentation method on SemanticKITTI and nuScenes test set in comparison with published leading contemporary SoTA approaches to demonstrate its superior efficacy (top 10 results shown; full rankings in Supplementary Material)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Our method significantly outperforms others, especially for rigid object categories (e.g., truck, o.veh, park, etc.), primarily due to our fusion of both localized 3D geometry and material reflectivity within RAPiD features, which" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.346, + 0.128 + ], + "angle": 0, + "content": "L. Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.173 + ], + "angle": 0, + "content": "Table 1: Quantitative results of RAPiD-Seg and top-10 SoTA segmentation methods on SemanticKITTI [2] test set; Best/2nd best highlighted; ▲ for multi-modal methods." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.175, + 0.787, + 0.289 + ], + "angle": 0, + "content": "
MethodmIoUcarbicymotortruco.vehpedb.listm.listroadparkwalko.grobuildfencvegtrunterrpolesign
Cylinder3D [64]68.997.167.663.850.858.573.769.248.092.265.077.032.390.766.585.672.569.862.466.2
AF2S3Net [7]69.794.565.486.839.241.180.780.474.391.368.872.553.587.963.270.268.553.761.571.0
RPVNet [56]70.397.668.468.744.261.175.974.473.493.470.380.733.393.572.186.575.171.764.861.4
SDSeg3D [27]70.497.458.754.254.965.270.274.452.290.969.476.741.993.271.186.174.371.165.470.6
GASN [61]70.796.965.858.059.361.080.482.746.389.866.274.630.192.369.687.373.072.566.171.6
PVKD [17]71.297.067.969.353.560.275.173.550.591.870.977.541.092.469.486.573.871.964.965.8
2DPASS [58]72.997.063.663.461.161.577.981.374.189.767.474.740.093.572.986.273.971.065.070.4
PCSeg [36]72.997.551.267.658.668.678.380.975.692.571.578.336.993.171.485.473.669.966.168.7
RangeFormer [22]73.396.769.473.759.966.278.175.958.192.473.078.842.492.370.186.673.372.866.466.6
▲UniSeg [37]75.297.971.975.263.674.178.974.860.692.674.079.546.193.472.787.576.373.168.368.5
RAPiD-Seg (Ours)76.197.771.176.272.580.779.979.159.891.878.278.646.093.672.186.974.672.365.968.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.298, + 0.787, + 0.326 + ], + "angle": 0, + "content": "Table 2: Quantitative results of RAPiD-Seg and top-10 SoTA segmentation methods on nuScenes [3] test set; Best/2nd best highlighted; ▲ for multi-modal methods." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.329, + 0.787, + 0.435 + ], + "angle": 0, + "content": "
MethodmIoUbarrbicybuscarconstmotorpedconetrailtruckdrivotherwalkterrmadeveg
▲PMF [65]77.082.040.081.088.064.079.080.076.081.067.097.068.078.074.090.088.0
Cylinder3D [64]77.282.829.884.389.463.079.377.273.484.669.197.770.280.375.590.487.6
AMVNet [34]77.380.632.081.788.967.184.376.173.584.967.397.567.479.475.591.588.7
SPVCNN [49]77.480.030.091.990.864.779.075.670.981.074.697.469.280.076.189.387.1
AF2S3Net [7]78.378.952.289.984.277.474.377.372.083.973.897.166.577.574.087.786.8
▲2D3DNet [12]80.083.059.488.085.163.784.482.076.084.871.996.967.479.876.092.189.2
GASN [61]80.485.543.290.592.164.786.083.073.383.975.897.071.081.077.791.690.2
2DPASS [58]80.881.755.392.091.873.386.578.572.584.775.597.669.179.975.590.288.0
▲LidarMultiNet [60]81.480.448.494.390.071.587.285.280.486.974.897.867.380.776.592.189.6
▲UniSeg [37]83.585.971.292.191.680.588.080.976.086.376.797.771.880.776.791.388.8
RAPiD-Seg (Ours)83.684.864.395.092.284.687.981.876.888.579.097.866.681.276.792.588.4
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.451, + 0.789, + 0.572 + ], + "angle": 0, + "content": "enable segmentation based on material properties and local rigid structures. Remarkably, our single-modal methodology outperforms multi-modal approaches [22, 34,37,56,58,60], suggesting superior efficacy of our RAPiD features over alternative modalities like RGB and range images. Our inference time (105ms per frame) is comparable to other contemporary approaches [37,58,64]. Furthermore, we present supporting qualitative results in Fig. 6. Whereas the baseline method struggles with accurate vehicle type differentiation, ours achieves consistent segmentation (more visualization results in the supplementary materials)." + }, + { + "type": "table", + "bbox": [ + 0.225, + 0.582, + 0.487, + 0.687 + ], + "angle": 0, + "content": "
RAPiD FeaturesAttentionmIoUΔ
GeometricReflectivityEmbedding
70.04(Baseline)
✓70.46(+0.42)
✓71.21(+1.17)
✓✓71.93(+1.89)
✓✓✓72.15(+2.11)
✓✓✓71.80(+1.76)
✓✓✓72.32(+2.28)
✓✓✓72.78(+2.74)
✓✓✓✓73.02(+2.98)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.693, + 0.49, + 0.763 + ], + "angle": 0, + "content": "Table 3: Component-wise ablation of RAPiD-Seg on the SemanticKITTI validation set: Geometric (Eq. (1)), Reflectivity (Eq. (3)), (RAPiD) Embedding (Sec. 4), (channel) Attention (Sec. 5.1)." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.581, + 0.784, + 0.693 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.712, + 0.788, + 0.754 + ], + "angle": 0, + "content": "Fig.6: Comparing our results and PCSeg (baseline) under multi-scan visualization, showing improved segmentation results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.773, + 0.402, + 0.787 + ], + "angle": 0, + "content": "6.3 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Effectiveness of Components: In Tab. 3, we ablate each component of RAPiD-Seg step by step and report the performance on the SemanticKITTI validation set. We start from a baseline model, achieving an mIoU of 70.04 on validation set." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.396, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RAPiD-Seg Networks for 3D LiDAR Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.389 + ], + "angle": 0, + "content": "Incorporating the RAPiD features leads to a notable increase in mIoU (+1.17), underscoring the efficacy of RAPiD in enhancing segmentation performance. Building upon the RAPiD framework, the integration of reflectivity further elevates our mIoU to 71.93, with an overall improvement of +1.89 from the baseline. This significant gain illustrates the critical role of reflectivity in capturing different object materials. Utilizing RAPiD AE to get the voxel-wise representation, we observe an additional improvement, taking the mIoU to 72.15 (+2.11 compared to the baseline), which demonstrates its capability in processing feature representations. We subsequently performed a module-by-module reduction to analyze the model performance. Specifically, without RAPiD-related modules (Geometric/Reflectivity/Embedding), mIoU drops from 73.02 to 70.46 - attributable to the lack of RAPiD features, which are robust to transformation and noise. Our ablation studies show that while disabling individual components leads to some improvement over the baseline, the peak performance is only attained when all components function together. This highlights the indispensable contribution of each component in maximizing segmentation accuracy." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.401, + 0.788, + 0.443 + ], + "angle": 0, + "content": "Table 4: Effects of RAPiD and Reflectivity features compared to other configurations on SemanticKITTI validation set. PDD: the compressed PDD embeddings; RAPiD-R: only RAPiD, without reflectivity feature fusion; RAPiD-R: both RAPiD and reflectivity." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.445, + 0.787, + 0.517 + ], + "angle": 0, + "content": "
Conf.mIoUtrucko.vehparkwalko.grobuildfence
BaseL [36]70.059.870.369.276.936.293.769.6
PDD [52]66.240.365.867.574.633.192.868.2
RAPiD-R71.862.569.070.377.435.893.670.8
RAPiD+R73.070.478.575.878.944.294.273.1
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.748 + ], + "angle": 0, + "content": "Effectiveness of RAPiD and Reflectivity Features: In Tab. 4, we validate the efficacy of the proposed RAPiD features. The direct application of PDD results in a substantial mIoU decrease of 3.8, with most category IoUs falling below those of the baseline method. This indicates vanilla PDD features are not well-suited for LiDAR segmentation. While our proposed RAPiD (without reflectivity disparities in Eq. (2)) demonstrates an improvement of 1.4 in mIoU over the baseline. Moreover, the integration of reflectivity with RAPiD features significantly enhances performance, yielding 2.6 mIoU increase over the baseline. Notably, our approach exhibits superior performance in the segmentation of most rigid object categories compared to other configurations. Tab. 6 further shows the RAPiD outperforms PDD in varying \\( k \\) values at different ranges on SemanticKITTI validation set, with non-uniform \\( k \\) configurations yielding the most significant improvements in mIoU for both datasets, highlighting the effectiveness of the range-aware design of RAPiD." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Effectiveness of RAPiD AE and Architecture Variants: In Tab. 3, we replace our RAPiD AE with a commonly-used ConvFNN [17,50,64] to generate voxel-wise embeddings. Our AE demonstrated a \\(0.24\\mathrm{mIoU}\\) improvement over the FNN. In Tab. 5, we investigate the impact of the variant architectures, i.e., R-RAPiD-Seg and C-RAPiD-Seg. Utilizing R-RAPiD features, R-RAPiD-Seg yields an mIoU of 72.3, which outperforms the baseline method by \\(+2.3\\mathrm{mIoU}\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.346, + 0.127 + ], + "angle": 0, + "content": "L. Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.142, + 0.497, + 0.184 + ], + "angle": 0, + "content": "Table 5: 3D segmentation results of different variants of RAPiD-Seg (ours) on SemanticKITTI validation set." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.19, + 0.492, + 0.25 + ], + "angle": 0, + "content": "
MethodmIoU %carpedo.gropole
Baseline70.097.278.135.463.5
R-RAPiD-Seg72.3 (+2.3)97.477.445.062.4
C-RAPiD-Seg73.0 (+3.0)97.779.344.666.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.516, + 0.141, + 0.782, + 0.156 + ], + "angle": 0, + "content": "Table 6: Effects of \\( k \\) at various ranges." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.162, + 0.782, + 0.25 + ], + "angle": 0, + "content": "
PDD [53]RAPiD (ours)
knearkmidkfarmIoUknearkmidkfarmIoU
SumK77764.74 (-7.1)77772.04 (+0.2)
55565.18 (-6.6)55572.28 (+0.5)
107566.23 (-5.6)107573.02 (+1.2)
lnuScene66672.19 (-6.5)66678.76 (+0.1)
33373.68 (-5.0)33379.43 (+0.8)
86372.24 (-6.4)86379.91 (+1.3)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.264, + 0.784, + 0.293 + ], + "angle": 0, + "content": "Concurrent fusing of both R- and C-RAPiD-Seg features in R-RAPiD-Seg improves performance by \\(+0.7\\) mIoU, which shows a \\(+3.0\\) overall mIoU enhancement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.294, + 0.787, + 0.354 + ], + "angle": 0, + "content": "Effectiveness of Feature Fusion with Channel-wise Attention: In Tab. 3, we assess the prevalent approach of direct feature concatenation. Our method, employing feature fusion with channel-wise attention, enhances mIoU by 0.87, thereby conclusively demonstrating its effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.355, + 0.788, + 0.476 + ], + "angle": 0, + "content": "Effectiveness of Backbone Networks: In Tab. 7, we evaluate the performance across multiple backbone networks to demonstrate the versatility and ease of integration of our modules. The results highlight the adaptability and effectiveness of our approach on both pointwise (P) and Voxel-wise (V) backbone architectures. Specifically, the point-based PTv2 backbone achieves an mIoU of 72.6, showing strong performance, particularly in the car and other ground categories. Among the voxel-based methods, Minkowski-UNet stands out with the highest overall mIoU of 73.0, excelling in the pedestrian and pole categories." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.493, + 0.787, + 0.521 + ], + "angle": 0, + "content": "Table 7: Effects of using different backbones on SemanticKITTI validation set, where P and V for Point- and Voxel-wsie methods." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.527, + 0.777, + 0.591 + ], + "angle": 0, + "content": "
Repr.BackbonemIoU %carpedo.gropole
PPTv2 [55]72.6 (+2.8)97.477.445.062.4
VCylinder3D [64]69.896.974.237.963.0
VMinkowski-UNet [8]73.0 (+3.2)97.278.144.265.9
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.619, + 0.509, + 0.635 + ], + "angle": 0, + "content": "7 Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.795 + ], + "angle": 0, + "content": "This paper presents a novel Range-Aware Pointwise Distance Distribution (RAPiD) feature and the RAPiD-Seg network for LiDAR segmentation, adeptly overcoming the constraints of single-modal LiDAR methods. The rigid transformation invariance and enhanced focus on local details of RAPiD significantly boost segmentation accuracy. RAPiD-Seg integrates a two-stage training approach with reflectivity-guided 4D distance metrics and a class-aware nested AE, achieving SoTA results on the SemanticKITTI and nuScenes datasets. Notably, our single-modal method surpasses the performance of multi-modal methods, indicating the superior efficacy of RAPiD features even compared to other modalities, including RGB and range images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Our RAPiD features hold significant potential for application in various tasks such as point cloud registration, semi-/weakly- supervised learning, and extend to promising applications in multi-modal research in the future." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.396, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RAPiD-Seg Networks for 3D LiDAR Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.18, + 0.677, + 0.194 + ], + "angle": 0, + "content": "1. Badger, B.L.: Depth and Representation in Vision Models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.196, + 0.787, + 0.235 + ], + "angle": 0, + "content": "2. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences. In: Int. Conf. Comput. Vis. pp. 9296-9306 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.238, + 0.786, + 0.278 + ], + "angle": 0, + "content": "3. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: Nuscenes: A Multimodal Dataset for Autonomous Driving. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 11618-11628 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.28, + 0.787, + 0.32 + ], + "angle": 0, + "content": "4. Chen, S., Wang, X., Cheng, T., Zhang, W., Zhang, Q., Huang, C., Liu, W.: Azinorm: Exploiting the Radial Symmetry of Point Cloud for Azimuth-Normalized 3D Perception. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6387-6396 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.322, + 0.787, + 0.362 + ], + "angle": 0, + "content": "5. Chen, X., Xu, S., Zou, X., Cao, T., Yeung, D.Y., Fang, L.: SVQNet: Sparse Voxel-Adjacent Query Network for 4D Spatio-Temporal LiDAR Semantic Segmentation. In: Int. Conf. Comput. Vis. pp. 8569–8578 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.364, + 0.787, + 0.418 + ], + "angle": 0, + "content": "6. Chen, X., Shi, S., Zhu, B., Cheung, K.C., Xu, H., Li, H.: MPPNet: Multi-frame Feature Intertwining with Proxy Points for 3D Temporal Object Detection. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Eur. Conf. Comput. Vis. pp. 680-697. Springer Nature Switzerland (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.42, + 0.787, + 0.46 + ], + "angle": 0, + "content": "7. Cheng, R., Razani, R., Taghavi, E., Li, E., Liu, B.: (AF)2-S3Net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12547-12556 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.462, + 0.787, + 0.502 + ], + "angle": 0, + "content": "8. Choy, C., Gwak, J., Savarese, S.: 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3070-3079 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.504, + 0.787, + 0.559 + ], + "angle": 0, + "content": "9. Çiçek, Ö., Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation. In: Ourselin, S., Joskowicz, L., Sabuncu, M.R., Unal, G., Wells, W. (eds.) Med. Image Comput. Comput. Assist. Interv. pp. 424–432. Springer International Publishing (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.561, + 0.786, + 0.601 + ], + "angle": 0, + "content": "10. Flitton, G., Breckon, T.P., Megherbi, N.: A Comparison of 3D Interest Point Descriptors with Application to Airport Baggage Object Detection in Complex CT Imagery. Pattern Recognition 46(9), 2420-2436 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.602, + 0.786, + 0.642 + ], + "angle": 0, + "content": "11. Geiger, A., Lenz, P., Urtasun, R.: Are We Ready for Autonomous Driving? The KITTI Vision Benchmark Suite. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3354-3361 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.644, + 0.786, + 0.685 + ], + "angle": 0, + "content": "12. Genova, K., Yin, X., Kundu, A., Pantofaru, C., Cole, F., Sud, A., Brewington, B., Shucker, B., Funkhouser, T.: Learning 3D Semantic Segmentation with only 2D Image Supervision. In: Int. Conf. 3D Vis. pp. 361-372 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.687, + 0.786, + 0.727 + ], + "angle": 0, + "content": "13. Graham, B., Engelcke, M., van der Maaten, L.: 3D Semantic Segmentation with Submanifold Sparse Convolutional Networks. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.729, + 0.786, + 0.755 + ], + "angle": 0, + "content": "14. Han, L., Zheng, T., Xu, L., Fang, L.: OccuSeg: Occupancy-Aware 3D Instance Segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2940-2949 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.757, + 0.786, + 0.797 + ], + "angle": 0, + "content": "15. He, C., Li, R., Li, S., Zhang, L.: Voxel Set Transformer: A Set-to-Set Approach to 3D Object Detection from Point Clouds. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 8407-8417. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.799, + 0.786, + 0.839 + ], + "angle": 0, + "content": "16. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked Autoencoders Are Scalable Vision Learners. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 16000-16009 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.221, + 0.18, + 0.787, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.346, + 0.128 + ], + "angle": 0, + "content": "L. Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "17. Hou, Y., Zhu, X., Ma, Y., Loy, C.C., Li, Y.: Point-to-Voxel Knowledge Distillation for LiDAR Semantic Segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.217 + ], + "angle": 0, + "content": "18. Hu, J., Shen, L., Sun, G.: Squeeze-and-Eccitation Networks. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.217, + 0.787, + 0.243 + ], + "angle": 0, + "content": "19. Ioffe, S., Szegedy, C.: Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. In: Int. Conf. Mach. Learn. pp. 448-456 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.787, + 0.284 + ], + "angle": 0, + "content": "20. Jiang, L., Shi, S., Tian, Z., Lai, X., Liu, S., Fu, C.W., Jia, J.: Guided Point Contrastive Learning for Semi-supervised Point Cloud Semantic Segmentation. In: Int. Conf. Comput. Vis. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.787, + 0.311 + ], + "angle": 0, + "content": "21. Jiang, M., Wu, Y., Zhao, T., Zhao, Z., Lu, C.: PointSIFT: A SIFT-like Network Module for 3D Point Cloud Semantic Segmentation (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.787, + 0.352 + ], + "angle": 0, + "content": "22. Kong, L., Liu, Y., Chen, R., Ma, Y., Zhu, X., Li, Y., Hou, Y., Qiao, Y., Liu, Z.: Rethinking Range View Representation for LiDAR Segmentation. In: Int. Conf. Comput. Vis. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.787, + 0.379 + ], + "angle": 0, + "content": "23. Kong, L., Ren, J., Pan, L., Liu, Z.: LaserMix for Semi-Supervised LiDAR Semantic Segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 21705-21715 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.787, + 0.434 + ], + "angle": 0, + "content": "24. Kortvelesy, R., Morad, S., Prorok, A.: Permutation-Invariant Set Autoencoders with Fixed-Size Embeddings for Multi-Agent Learning. In: Int. Conf. Auton. Agents Multiagent Syst. pp. 1661-1669. International Foundation for Autonomous Agents and Multiagent Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.787, + 0.461 + ], + "angle": 0, + "content": "25. Kumar, A., Poole, B.: On Implicit Regularization in \\(\\beta\\)-VAEs. In: Int. Conf. Mach. Learn. pp. 5480-5490 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.787, + 0.487 + ], + "angle": 0, + "content": "26. Li, E., Casas, S., Urtasun, R.: MemorySeg: Online LiDAR Semantic Segmentation with a Latent Memory. In: Int. Conf. Comput. Vis. pp. 745-754 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.487, + 0.787, + 0.514 + ], + "angle": 0, + "content": "27. Li, J., Dai, H., Ding, Y.: Self-Distillation for Robust LiDAR Semantic Segmentation in Autonomous Driving. In: Eur. Conf. Comput. Vis. pp. 659-676 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.787, + 0.555 + ], + "angle": 0, + "content": "28. Li, J., Dai, H., Han, H., Ding, Y.: MSeg3D: Multi-Modal 3D Semantic Segmentation for Autonomous Driving. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 21694-21704 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.787, + 0.609 + ], + "angle": 0, + "content": "29. Li, L., Ismail, K.N., Shum, H.P.H., Breckon, T.P.: DurLAR: A High-Fidelity 128-Channel LiDAR Dataset with Panoramic Ambient and Reflectivity Imagery for Multi-Modal Autonomous Driving Applications. In: Int. Conf. 3D Vis. pp. 1227-1237 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.787, + 0.651 + ], + "angle": 0, + "content": "30. Li, L., Shum, H.P.H., Breckon, T.P.: Less is More: Reducing Task and Model Complexity for Semi-Supervised 3D Point Cloud Semantic Segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.787, + 0.691 + ], + "angle": 0, + "content": "31. Li, X., Li, R., Chen, G., Fu, C.W., Cohen-Or, D., Heng, P.A.: A Rotation-Invariant Framework for Deep Point Cloud Analysis. IEEE Trans. Vis. Comput. Graph. 28(12), 4503-4514 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.787, + 0.732 + ], + "angle": 0, + "content": "32. Liang, J., Lai, R., Wong, T.W., Zhao, H.: Geometric Understanding of Point Clouds Using Laplace-Beltrami Operator. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 214-221 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.787, + 0.759 + ], + "angle": 0, + "content": "33. Lin, M., Chen, Q., Yan, S.: Network In Network. In: Int. Conf. Learn. Represent. (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.787, + 0.8 + ], + "angle": 0, + "content": "34. Liong, V.E., Nguyen, T.N.T., Widjaja, S., Sharma, D., Chong, Z.J.: AMVNet: Assertion-based Multi-View Fusion Network for LiDAR Semantic Segmentation (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.841 + ], + "angle": 0, + "content": "35. Liu, J., Yu, Z., Breckon, T.P., Shum, H.P.H.: U3DS3: Unsupervised 3D Semantic Scene Segmentation. In: IEEE Winter Conf. Appl. Comput. Vis. pp. 3759-3768 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.396, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RAPiD-Seg Networks for 3D LiDAR Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.19 + ], + "angle": 0, + "content": "36. Liu, Y., Bai, Y., Chen, R., Hou, Y., Shi, B., Li, Y., Kong, L.: PCSeg: An Open Source Point Cloud Segmentation Codebase. https://github.com/PJLab-ADG/PCSeg (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.233 + ], + "angle": 0, + "content": "37. Liu, Y., Chen, R., Li, X., Kong, L., Yang, Y., Xia, Z., Bai, Y., Zhu, X., Ma, Y., Li, Y., Qiao, Y., Hou, Y.: UniSeg: A Unified Multi-Modal LiDAR Segmentation Network and the OpenPCSeg Codebase. In: Int. Conf. Comput. Vis. pp. 21662–21673 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.787, + 0.274 + ], + "angle": 0, + "content": "38. Liu, Y., Kong, L., Cen, J., Chen, R., Zhang, W., Pan, L., Chen, K., Liu, Z.: Segment Any Point Cloud Sequences by Distilling Vision Foundation Models. In: Adv. Neural Inform. Process. Syst. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.274, + 0.787, + 0.303 + ], + "angle": 0, + "content": "39. Marmolin, H.: Subjective MSE measures. IEEE Trans. Syst. Man Cybern. 16(3), 486-489 (1986)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.303, + 0.787, + 0.346 + ], + "angle": 0, + "content": "40. Melia, O., Jonas, E., Willett, R.: Rotation-Invariant Random Features Provide a Strong Baseline for Machine Learning on 3D Point Clouds. Trans. Mach. Learn. Res. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.787, + 0.389 + ], + "angle": 0, + "content": "41. Nunes, L., Wiesmann, L., Marcuzzi, R., Chen, X., Behley, J., Stachniss, C.: Temporal Consistent 3D LiDAR Representation Learning for Semantic Perception in Autonomous Driving. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.389, + 0.787, + 0.431 + ], + "angle": 0, + "content": "42. Preechakul, K., Chathee, N., Wizadwongsa, S., Suwajanakorn, S.: Diffusion Autoencoders: Toward a Meaningful and Decodable Representation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 10619-10629 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.431, + 0.787, + 0.459 + ], + "angle": 0, + "content": "43. PyTorch: Scatter: PyTorch Scatter Documentation. https://pytorch-scatter.readthedocs.io/en/latest/functions/scatter.html (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.459, + 0.787, + 0.502 + ], + "angle": 0, + "content": "44. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 652-660 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.787, + 0.544 + ], + "angle": 0, + "content": "45. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space. In: Adv. Neural Inform. Process. Syst. vol. 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.544, + 0.787, + 0.587 + ], + "angle": 0, + "content": "46. Shao, Y., Tan, A., Yan, T., Sun, Z., Zhang, Y., Liu, J.: MS23D: A 3D Object Detection Method Using Multi-Scale Semantic Feature Points to Construct 3D Feature Layer (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.787, + 0.628 + ], + "angle": 0, + "content": "47. Sun, J., Xu, X., Kong, L., Liu, Y., Li, L., Zhu, C., Zhang, J., Xiao, Z., Chen, R., Wang, T., et al.: An Empirical Study of Training State-of-the-Art LiDAR Segmentation Models. arXiv:2405.14870 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.628, + 0.787, + 0.699 + ], + "angle": 0, + "content": "48. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in Perception for Autonomous Driving: Waymo Open Dataset. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2446-2454 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.699, + 0.787, + 0.741 + ], + "angle": 0, + "content": "49. Tang, H., Liu, Z., Zhao, S., Lin, Y., Lin, J., Wang, H., Han, S.: Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution. In: Eur. Conf. Comput. Vis. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.741, + 0.787, + 0.77 + ], + "angle": 0, + "content": "50. Unal, O., Dai, D., Van Gool, L.: Scribble-Supervised LiDAR Semantic Segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.77, + 0.787, + 0.812 + ], + "angle": 0, + "content": "51. Wang, S., Zhu, J., Zhang, R.: Meta-RangeSeg: LiDAR Sequence Semantic Segmentation Using Multiple Feature Aggregation. IEEE Robot. Autom. Lett. 7(4), 9739-9746 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.812, + 0.787, + 0.841 + ], + "angle": 0, + "content": "52. Widdowson, D., Kurlin, V.: Resolving the Data Ambiguity for Periodic Crystals. In: Adv. Neural Inform. Process. Syst. vol. 35, pp. 24625-24638 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.346, + 0.128 + ], + "angle": 0, + "content": "L. Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "53. Widdowson, D., Kurlin, V.: Recognizing Rigid Patterns of Unlabeled Point Clouds by Complete and Continuous Isometry Invariants With No False Negatives and No False Positives. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1275-1284 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.231 + ], + "angle": 0, + "content": "54. Widdowson, D., Mosca, M.M., Pulido, A., Cooper, A.I., Kurlin, V.: Average Minimum Distances of Periodic Point Sets - Foundational Invariants for Mapping Periodic Crystals. MATCH Commun. Math. Comput. Chem. 87(3), 529-559 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.272 + ], + "angle": 0, + "content": "55. Wu, X., Lao, Y., Jiang, L., Liu, X., Zhao, H.: Point Transformer V2: Grouped Vector Attention and Partition-Based Pooling. In: Adv. Neural Inform. Process. Syst. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "56. Xu, J., Zhang, R., Dou, J., Zhu, Y., Sun, J., Pu, S.: RPVNet: A Deep and Efficient Range-Point-Voxel Fusion Network for LiDAR Point Cloud Segmentation. In: Int. Conf. Comput. Vis. pp. 16024–16033 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "57. Yan, X., Gao, J., Li, J., Zhang, R., Li, Z., Huang, R., Cui, S.: Sparse Single Sweep LiDAR Point Cloud Segmentation via Learning Contextual Shape Priors from Scene Completion. In: Conf. AAAI Artif. Intell. vol. 35, pp. 3101-3109 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.787, + 0.396 + ], + "angle": 0, + "content": "58. Yan, X., Gao, J., Zheng, C., Zheng, C., Zhang, R., Cui, S., Li, Z.: 2DPASS: 2D Priors Assisted Semantic Segmentation on LiDAR Point Clouds. In: Eur. Conf. Comput. Vis. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.452 + ], + "angle": 0, + "content": "59. Ye, D., Chen, W., Zhou, Z., Xie, Y., Wang, Y., Wang, P., Foroosh, H.: LidarMultiNet: Unifying LiDAR Semantic Segmentation, 3D Object Detection, and Panoptic Segmentation in a Single Multi-task Network. In: IEEE Conf. Comput. Vis. Pattern Recog. Worksh. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.493 + ], + "angle": 0, + "content": "60. Ye, D., Zhou, Z., Chen, W., Xie, Y., Wang, Y., Wang, P., Foroosh, H.: LidarMultiNet: Towards a Unified Multi-Task Network for Lidar Perception. In: Conf. AAAI Artif. Intell. vol. 37, pp. 3231-3240 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.787, + 0.549 + ], + "angle": 0, + "content": "61. Ye, M., Wan, R., Xu, S., Cao, T., Chen, Q.: Efficient Point Cloud Segmentation with Geometry-Aware Sparse Networks. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Eur. Conf. Comput. Vis. pp. 196-212. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.59 + ], + "angle": 0, + "content": "62. Zhang, C., Duc, K.D., Condon, A.: Visualizing DNA Reaction Trajectories with Deep Graph Embedding Approaches. In: Adv. Neural Inform. Process. Syst. Worksh. arXiv (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.787, + 0.618 + ], + "angle": 0, + "content": "63. Zhao, H., Jiang, L., Jia, J., Torr, P., Koltun, V.: Point Transformer. In: Int. Conf. Comput. Vis. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.787, + 0.66 + ], + "angle": 0, + "content": "64. Zhu, X., Zhou, H., Wang, T., Hong, F., Ma, Y., Li, W., Li, H., Lin, D.: Cylindrical and Asymmetrical 3D Convolution Networks for LiDAR Segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.787, + 0.702 + ], + "angle": 0, + "content": "65. Zhuang, Z., Li, R., Jia, K., Wang, Q., Li, Y., Tan, M.: Perception-Aware Multi-Sensor Fusion for 3D LiDAR Semantic Segmentation. In: Int. Conf. Comput. Vis. pp. 16280-16290 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.702 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/RAPiD-Seg_ Range-Aware Pointwise Distance Distribution Networks for 3D LiDAR Segmentation/b19a59c8-eca1-4aae-977c-6a71eef1622b_origin.pdf b/2024/RAPiD-Seg_ Range-Aware Pointwise Distance Distribution Networks for 3D LiDAR Segmentation/b19a59c8-eca1-4aae-977c-6a71eef1622b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2fd56de2f7051d80145e1a8fceed1be8bb8b93c4 --- /dev/null +++ b/2024/RAPiD-Seg_ Range-Aware Pointwise Distance Distribution Networks for 3D LiDAR Segmentation/b19a59c8-eca1-4aae-977c-6a71eef1622b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d3e5ae18dcea69074cf8006dc4ff5c156d3da00a1bf26969efdaf4c7be2b53d5 +size 5591761 diff --git a/2024/RAPiD-Seg_ Range-Aware Pointwise Distance Distribution Networks for 3D LiDAR Segmentation/full.md b/2024/RAPiD-Seg_ Range-Aware Pointwise Distance Distribution Networks for 3D LiDAR Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ed12fef73999b807a5eebaa29b815a59e2d6c994 --- /dev/null +++ b/2024/RAPiD-Seg_ Range-Aware Pointwise Distance Distribution Networks for 3D LiDAR Segmentation/full.md @@ -0,0 +1,376 @@ +# RAPiD-Seg: Range-Aware Pointwise Distance Distribution Networks for 3D LiDAR Segmentation + +Li Li $^{1}$ , Hubert P. H. Shum $^{1}$ , and Toby P. Breckon $^{1,2}$ + +Department of $\{\mathrm{Computer~Science}^1\mid$ Engineering2}, Durham University, UK {li.li4, hubert.shum, toby.breckon}@durham.ac.uk + +Abstract. 3D point clouds play a pivotal role in outdoor scene perception, especially in the context of autonomous driving. Recent advancements in 3D LiDAR segmentation often focus intensely on the spatial positioning and distribution of points for accurate segmentation. However, these methods, while robust in variable conditions, encounter challenges due to sole reliance on coordinates and point intensity, leading to poor isometric invariance and suboptimal segmentation. To tackle this challenge, our work introduces Range-Aware Pointwise Distance Distribution (RAPiD) features and the associated RAPiD-Seg architecture. Our RAPiD features exhibit rigid transformation invariance and effectively adapt to variations in point density, with a design focus on capturing the localized geometry of neighboring structures. They utilize inherent LiDAR isotropic radiation and semantic categorization for enhanced local representation and computational efficiency, while incorporating a 4D distance metric that integrates geometric and surface material reflectivity for improved semantic segmentation. To effectively embed high-dimensional RAPiD features, we propose a double-nested autoencoder structure with a novel class-aware embedding objective to encode high-dimensional features into manageable voxel-wise embeddings. Additionally, we propose RAPiD-Seg which incorporates a channel-wise attention fusion and two effective RAPiD-Seg variants, further optimizing the embedding for enhanced performance and generalization. Our method outperforms contemporary LiDAR segmentation work in terms of mIoU on SemanticKITTI (76.1) and nuScenes (83.6) datasets. + +Keywords: Semantic Segmentation $\cdot$ Point Cloud $\cdot$ Invariance Features + +# 1 Introduction + +LiDAR-based point clouds, characterized by pointwise 3D positions and LiDAR intensity/reflectivity [3,11,29,48], play a pivotal role in outdoor scene understanding, particularly in perception systems for autonomous driving. Recent literature in the domain of 3D LiDAR segmentation has witnessed significant strides, with myriad approaches attempting pointwise 3D semantic scene labeling [47]. Leveraging color [7,37,38], range imagery [22,34], and Birds Eye View (BEV) projections [34,60], multi-modal methods integrate diverse data streams from LiDAR and other sensors to enhance feature representation and segmentation accuracy. Single-modal (LiDAR-only) methods [17,30,56,64] effectively exploit spatial data interpretation and avoid multi-modal issues, e.g., modality heterogeneity and limited sensor field of view intersection [5]. + +![](images/7ae4750af076c4f446545f4f228ce856c0601e8017fad9cd304cbe0044e475cf.jpg) +Fig. 1: Left: RAPiD exhibits excellent viewpoint invariance and geometric stability, visualizing comparable features around the vehicle door structure at varying ranges and viewpoints (feature matrix plots inset). Middle: RAPiD is distinctive in different semantic classes, as visualized by the matrices. Embedded RAPiD patterns corresponding to different points are visualized using a spectrum of colors, showcasing their capacity to represent different classes. Right: Our RAPiD-Seg achieves superior results over SoTA methods on nuScenes [3] and SemanticKITTI [2]. + +![](images/7d83ae89b7fde62c769474d221a118e7062d275753d69257cdd064fce416cb6f.jpg) + +![](images/18dd574e87204f7f2dc16320e42d5696fc0be013651de4211c47c9508c79309c.jpg) + +Despite these advancements, a common challenge inherent to these methods is their poor isometric invariance, characterized by heavy reliance on solely coordinates and intensity data, which often results in suboptimal segmentation outcomes [30,64]; this is primarily due to poor translational invariance and visibility (e.g., occlusions), or sparse observations (e.g., at long range) [26], affecting data spatial distribution. + +In this work, we seek features that are (1) capable of capturing the localized geometric structure of neighboring points, (2) invariant to rotation and translation, and (3) applicable in noisy LiDAR outdoor environments. While numerous methods fulfill some of these requirements individually [21, 32, 40], they fall short of addressing them all [10]. Recognizing the need for higher-level features capable of capturing local geometry while potentially incorporating LiDAR point-specific attributes (e.g., intensity and reflectivity), we instead turn our attention to pointwise Distance Distribution (PDD) features [52, 53]. They are recognized for their remarkable efficacy in providing a robust and informative geometric representation of point clouds, excelling in both rotational and translational invariance while including intricate local geometry details. + +However, directly employing PDD features is impractical due to their high-dimensional nature, and their significant memory and storage requirements for large-scale point clouds. PDD also overlooks local features, as including distant points subsequently dilutes the focus on immediate neighborhoods. + +As an enabler, we propose the Range-Aware Pointwise Distance Distribution (RAPiD-Seg) solution for LiDAR segmentation. It utilizes invariance of RAPiD features to rigid transformations and point cloud sparsity variations, while concentrating on compact features within specific local neighborhoods. Specifically, our method leverages inherent LiDAR isotropic radiation and semantic categorization for enhanced local representations while reducing computational burdens. Moreover, our formulation computes a 4D distance, incorporating both 3D geometric and reflectivity differences to enhance semantic segmentation fidelity. To compress high-dimensional RAPiD features into tractable voxel-wise embeddings, we propose a novel embedding approach with our class-aware double-nested AutoEn + +coder (AE) module. We further incorporate a channel-wise attention fusion and a two-stage training strategy to optimize the AE independently before integrating into the full network, enhancing performance and generalization capability. + +We conduct extensive experiments on SemanticKITTI [2] and nuScenes [3] datasets, upon which our approach surpasses the state-of-the-art (SoTA) segmentation performance. + +Overall, as shown in Fig. 2, our contributions can be summarized as follows: + +A novel Range-Aware Pointwise Distance Distribution feature (RAPiD) for 3D LiDAR segmentation that ensures robustness to rigid transformations and viewpoints through isometry-invariant metrics within specific Regions of Interest, i.e., intra-ring (R-RAPiD) and intra-class (C-RAPiD). +A novel method for embedding RAPiD with class-aware double nested AE (RAPiD AE) to optimize the embedding of high-dimensional features, balancing efficiency and fidelity. +A novel open-source network architecture RAPiD-Seg $^{1}$ and supporting training methodology, to enable modular LiDAR segmentation that achieves SoTA performance on SemanticKITTI (mIoU: 76.1) and nuScenes (mIoU: 83.6). + +![](images/d94a63e051fc0935ca7cf572716fc202e2be2bbf220b3bdd9c553111b078c401.jpg) +Fig. 2: Our proposed architecture for 3D segmentation framework leverages RAPiD features from the point cloud. We encode pointwise features into voxel-wise embeddings via the voxel encoder and multiple RAPiD AutoEncoders (RAPiD AE). After attention-based feature fusion, these fused embeddings go through the backbone network for segmentation results. + +# 2 Related Work + +LiDAR-Based Semantic Segmentation [7,12,17,20,22,23,27,30,34,37,38,49, 50,56-58,60,61,64,65] is fundamental for LiDAR-driven scene perception, aiming to label each point in a point cloud sequence. The majority of the approaches [17, 38,49,57,64] solely rely on the point-based features of the point cloud, such as SPVCNN [49] which introduces a point-to-voxel branch, using combined pointvoxel features for segmentation. Cylinder3D [64] proposes cylindrical partitioning with a UNet [9] backbone variant. LiM3D [30] utilizes coordinates combined with surface reflectivity attributes. Overall, such prior work relies solely on pointbased features such as coordinates and intensity of the points, lacking an effective + +fusion mechanism, resulting in suboptimal performance [6,28,51]. They are also susceptible to changes in viewpoint, distance, and point sparsity [26] due to the lack of isometry across all inter-point distances. + +Pointwise Distance Distribution (PDD) captures the local context of each point in a unit cell by enumerating distances to neighboring points in order. It is an isometry invariant proposed by Widdowson & Kurlin [52] to resolve the data ambiguity for periodic crystals, demonstrated through extensive pairwise comparisons across atomic 3D clouds from high-level periodic crystals of periodic structures [52-54]. Though the effectiveness of PDD in periodic crystals and atomic clouds has been proved by the aforementioned studies, to date, no work has applied PDD features to outdoor 3D point clouds. In outdoor settings, common invariant features [21,31,32,40] and Azimuth-normalized method [4] often face issues due to the irregular and sparse nature of the data, which is compounded by increased noise and environmental complexities [31,32]. Furthermore, the computational demands make them less suitable for the vast scale of outdoor settings [21,40]. For instance, Melia et al. [40] introduce a rotation-invariant feature that struggles with generalization across diverse point cloud densities and scales due to its computational cost and susceptibility to outdoor noise. Recognizing these limitations, we identify an opportunity to leverage the PDD features in a new domain, where representing the local context of neighboring points in a transformation invariant and geometrically robust manner is crucial. Drawing from the advantages of the PDD design, we propose the RAPiD feature, tailored specifically for LiDAR-based point clouds, to capture the localized geometry of neighboring structures. + +Point Cloud Representations and Embeddings are fundamental for LiDAR-driven tasks. 3D point cloud can be represented either as point-based [44,45,55,63] or voxel-based [17,35,50,64], each serving different purposes rooted in their unique characteristics. Voxel-based representation involves dividing the 3D space into grids of voxels, which provides a structured way of handling sparse and irregular LiDAR data [9, 13, 14, 64]. Previous work often utilizes Feedforward Neural Network (FNN) [17, 50, 64] to facilitate the conversion of pointwise features into structured voxel-wise features by projecting them onto voxel grids and generating dense embeddings. The primary drawback of FNN conversion is the potential loss of fine-grained details inherent in the original point cloud [15]. FNN may also struggle with the high dimensionality and sparsity of voxelized data, leading to computational inefficiency [46]. Our method addresses these issues by partitioning the point cloud into voxel grids and applying a novel class-ware double nested AE with self-attention mechanisms [15]. This approach not only retains more detailed information by focusing on local structures but also introduces high-dimensional feature compression, thereby improving computational efficiency. + +# 3 Range-Aware Pointwise Distances (RAPiD) + +In the context of LiDAR segmentation, a principal challenge lies in the lack of robustness against rigid transformations such as rotation and translation, as well as against variations in viewpoint, points sparsity, and occlusion [5, 26]. + +Traditional methods predominantly leverage 3D point coordinates to furnish spatial information [17,50,64]; however, they may be inadequate in the scenes with poor visibility (e.g., occlusions) or sparse observations (e.g., at long range) [26]. Such reliance on coordinates alone could lead to inaccuracies in recognizing object transformations or occlusions [26]. Data augmentation, e.g., random geometric transformation can improve robustness under rigid transformations [50,64], but fail to guarantee comprehensive coverage of all potential transformations, resulting in vulnerability to previously unseen variations. + +# 3.1 Overview + +To achieve a transformation-invariant 3D data representation, we observe that distances within rigid bodies (e.g., vehicles, roads, and buildings) remain constant under rigid transformations. In light of this, we focus on assimilating the principle of the isometric invariant into the LiDAR-driven point cloud perception. Specifically, we delve into the PDD [52] — an isometry invariant that quantifies distances between adjacent points. For outdoor point clouds, vanilla PDD features are computationally intensive and susceptible to noise and sparsity. Our Range-Aware Pointwise Distance Distribution (RAPiD), specifically designed for LiDAR data, instead calculates the PDD features for each point within specific Regions of Interest (RoI), which are typically associated with the intrinsic structure of LiDAR data. + +Mathematical Formulation: Given a fixed number $k > 0$ representing the fixed number of point neighbors and a $u$ -point cluster $P_{\mathrm{RoI}}$ comprising no fewer than $k$ points based on RoI, the RAPiD is a $u \times k$ matrix, which retains both spatial distances and LiDAR reflectivity [30] disparities between points. RAPiD are adapted to LiDAR sparsity at different distances by using range-specific parameters $k_{\mathrm{close}}$ , $k_{\mathrm{mid}}$ , $k_{\mathrm{far}}$ for close, mid, and far ranges, respectively. + +The $k$ -point RAPiD in region $P_{\mathrm{RoI}}$ is defined as: + +$$ +\operatorname {R A P i D} \left(P _ {\text {R o I}}; k\right) = \operatorname {s o r t} \left(\left[ \operatorname {s o r t} \left(\left[ \boldsymbol {\rho} _ {j, 1}, \dots , \boldsymbol {\rho} _ {j, k} \right]\right) \right] _ {j = 1} ^ {u}\right), \tag {1} +$$ + +$\forall l \in \{1, \dots, k\}, j \in \{1, \dots, u\}, \pmb{\rho}_{j,l}$ is given by: + +$$ +\boldsymbol {\rho} _ {j, l} = \left\| \left[ \boldsymbol {p} _ {j} - \boldsymbol {p} _ {j, l}, \quad g (r _ {j}) - g (r _ {j, l}) \right] \right\| _ {2}, \tag {2} +$$ + +where $\pmb{p}_j$ and $\pmb{p}_{j,l}$ denote the 3D coordinates of the $j$ -th point and its $l$ -th nearest neighbor within $P_{\mathrm{RoI}}$ , respectively; $r_j$ and $r_{j,l}$ represent the reflectivity values of $\pmb{p}_j$ and $\pmb{p}_{j,l}$ , correspondingly; $\| \cdot \| _2$ is the Euclidean norm. $g:\mathbb{R}\to [D_{\min},D_{\max}]$ is the reflectivity mapping function that maps the numerical range of reflectivity onto a consistent scale with the range of Euclidean distances between points: + +$$ +g (r) = \left(\frac {r - r _ {\operatorname* {m i n}}}{r _ {\operatorname* {m a x}} - r _ {\operatorname* {m i n}}}\right) \left(D _ {\operatorname* {m a x}} - D _ {\operatorname* {m i n}}\right) + D _ {\operatorname* {m i n}}, \tag {3} +$$ + +$$ +D _ {\min } = \min _ {j, l} \| \boldsymbol {p} _ {j} - \boldsymbol {p} _ {j, l} \| _ {2}, \quad D _ {\max } = \max _ {j, l} \| \boldsymbol {p} _ {j} - \boldsymbol {p} _ {j, l} \| _ {2}. \tag {4} +$$ + +where $[D_{\min}, D_{\max}]$ is the range of the Euclidean norms of coordinate differences for all considered point pairs. + +The 4D distance in Eq. (2) integrates material reflectivity [30] into RAPiD features, enhancing feature representation and aiding in the discrimination of various materials and surfaces, which is crucial for accurate semantic segmentation. + +For convenience to facilitate feature normalization and data alignment, we arrange RAPiD lexicographically by sorting Eq. (1), where $\mathrm{sort}(\cdot)$ on the inner and outer brackets sorts the elements $\rho_{j,l}$ within each row $j$ , and the sorted rows based on their first differing elements, both in ascending order [52]. It treats RAPiD matrices with the same geometric structure but different orders as identical. + +To improve the data robustness, we normalize $\rho_{j,l}$ after excluding outliers exceeding a distance threshold $\delta$ , substituting these entries with the maximum of the normalized distribution to represent significant inter-point distances. + +# 3.2 Intra-Ring and Intra-Class RAPiD + +To enhance the RAPiD ability to capture local context, we propose R-RAPiD and C-RAPiD, based on the inherent structure of LiDAR data. The key benefit of C-RAPiD is its capacity to underscore the inherent traits of points within the same semantic class, achieved based on semantic labels before distance calculation. As a complementary, R-RAPiD is versatile, regardless of semantic label availability. Intra-Ring RAPiD (R-RAPiD) is a specialized variant of RAPiD, which confines the RoI to the ring encompassing the anchor points, thereby serving to economize on computational overhead while capitalizing on the inherent structural characteristics of LiDAR data (Fig. 2 ①). + +Given LiDAR sensors with fixed beam counts [2,3,29,48], e.g., $B \in \{32,64,128\}$ , beams radiate isotropically around the vehicle at set angles. We segment the $m$ -point LiDAR point cloud $C_{B,s}$ by beams, where $m = s \times B$ , with $s$ representing measurements within a scan cycle and $B$ the number of laser beams. Suppose $\pmb{p}_i$ with Cartesian coordinate $(x_i, y_i, z_i)$ is a point within $C_{B,s}$ , its transformation to cylindrical coordinates $(\theta_i, \phi_i)$ is formulated as: + +$$ +\theta_ {i} = \left\lfloor \arctan 2 \left(y _ {i}, x _ {i}\right) / \Delta \theta \right\rfloor , \tag {5} +$$ + +$$ +\phi_ {i} = \left\lfloor \arcsin \left[ z _ {i} \left(x _ {i} ^ {2} + y _ {i} ^ {2} + z _ {i} ^ {2}\right) ^ {- 1 / 2} \right] / \Delta \phi \right\rfloor , \tag {6} +$$ + +where $\varDelta\theta$ and $\varDelta\phi$ denote the mean angular resolutions horizontally and vertically between adjacent beams, and $\arctan 2(\cdot)$ is the 2-argument $\arctan (\cdot)$ function. + +$C_{B,s}$ is then partitioned to $B$ rings based on $\phi_b$ , i.e., $C_{B,s} = \bigcup_{b=0}^{B-1} R_b$ , where each ring $R_b = \{(\theta_{b,i}, \phi_{b,i}) \mid \phi_{b,i} = \phi_b, \forall i \leq m\}$ ; $\bigcup$ signifies the cumulative union R-RAPiD is thus defined as the cumulative concatenate $\oplus$ of RAPiD in each ring $R_b$ . + +$$ +\text {R - R A P i D} \left(C _ {B, s}; k\right) := \bigoplus_ {b = 0} ^ {B - 1} \text {R A P i D} \left(R _ {b}; k\right). \tag {7} +$$ + +Intra-Class RAPiD (C-RAPiD) concentrates on extracting point features within the confines of each semantic class (Fig. 2 ②). It is complementary to R-RAPiD, preserving the embedding fidelity within individual semantic class. + +Specifically, C-RAPiD concatenates RAPiD for each point $\pmb{p}_j$ and its corresponding semantic label $y_{j}$ within semantic class $i$ , across $N_{c}$ total classes: + +$$ +\text {C - R A P i D} \left(\left\{C _ {B, s}, \mathcal {Y} \right\}; k\right) := \bigoplus_ {i = 0} ^ {N _ {c} - 1} \operatorname {R A P i D} \left(S _ {i}; k\right), \tag {8} +$$ + +where $S_{i} = \{\pmb{p}_{j}\mid y_{j} = i,\forall j\leq m\}$ $\mathcal{V} = [y_j]_{j = 1}^m$ + +# 4 RAPiD Embedding + +The direct utilization of high-dimensional RAPiD features imposes a substantial computational burden. Efficient processing of large data volumes requires methods that can condense dimensionality while preserving data integrity [62]. + +We explore data embedding methods to reduce computational costs while providing feature learning capacity. AE is efficient in embedding high-dimensional features utilizing the reconstruction loss [16, 42]. However, conventional AE, limited by their sensitivity to input order and reliance on fixed-size misinputs [24], struggle to align with the unordered and variable size inherent in point clouds [15]. + +We propose a double nested AE structure with a novel class-aware embedding objective, using the Voxel-based Set Attention (VSA) module [15] as a building block. This facilitates superior contextual awareness and adaptability to unordered and variably-sized point cloud data. + +Nested RAPiD AE, illustrated in Fig. 2 ③ and Fig. 3, is composed of two modules: the outer module is constituted by a VSA AE, primarily responsible for the point-to-voxel conversion. Within the inner module, which focuses on voxel-wise representation, we employ an additional AE specifically designed for dimensionality reduction. Specifically, considering an input of $m$ -point RAPiD features $G \in \mathbb{R}^{m \times d}$ , where each point encompasses $d$ features. We compress pointwise $G$ into a voxel-wise representation $H^v \in \mathbb{R}^{c \times l \times d}$ by an outer VSA Voxel Encoder (with $c$ reduced voxel-wise dimension and $l$ latent codes), then further reduces it to a compressed embedding $\hbar \in \mathbb{R}^{c \times l \times d'}$ with $d'$ features via an inner Encoder. The inner and outer VSA Point Decoders then reconstruct the output feature set $\hat{G}$ akin to $G$ . + +Inner Voxel-Wise Representation is conducted via an inner AE, which takes $H^v$ from outer VSA AE as the input, yielding a lower-dimensional embedding $\hbar$ (dimension reduced from $d$ to $d'$ ). This process involves $u \times$ convolutional layers and a batch normalization layer [19]: + +$$ +\hbar^ {(i)} = \operatorname {B a t c h N o r m} \circ \operatorname {C o n v} \left(\hbar^ {(i - 1)}\right), \tag {9} +$$ + +where $\hbar^{(0)} = H^v$ , $\hbar^{(u - 1)} = \hbar$ . The Convolutional Feed-Forward Network (ConvFFN) then promotes voxel-level information exchange, enhancing spatial feature interactivity. It maps reduced hidden features $\hbar$ to a 3D sparse tensor, indexed by voxel coordinates $X^v$ , and employs dual depth-wise convolutions (DwConv) for spatial interactivity: + +$$ +\hat {h} = \operatorname {D w C o n v} ^ {(2)} \circ \zeta \circ \operatorname {D w C o n v} ^ {(1)} \left(\operatorname {S p T} \left(\hbar , X ^ {v}\right)\right), \tag {10} +$$ + +where $\zeta$ and $\mathrm{SpT}(\cdot)$ represents the non-linear activation and sparse tensor construction. Subsequently, $\hat{\hbar}\in \mathbb{R}^{c\times l\times d'}$ is reconstructed into $\hat{H}^v\in \mathbb{R}^{c\times l\times d}$ through an inner Decoder, consisting of multiple DeConv Layers. + +Outer VSA AE consists of a voxel encoder and point encoder. The voxel encoder project RAPiD features into key-value spaces, forming $K$ and $V$ , followed by a cross-attention mechanism with a latent query $L$ , producing an attention + +![](images/bd43c21d0b93992e957c04b3a21a936e2d971f12f33688afff80e8c5bb26aac8.jpg) +Fig. 3: Our RAPiD AE consists of an Encoder, Convolution Layer, and Decoder module, aiming to reproduce the input features and generate the compressed voxel-wise RAPiD representation $\hbar$ . + +matrix $A$ . The voxel-wise representation $H^v$ is the scatter sum [43] of pointwise $H$ to aggregate the value vectors $V$ : + +$$ +H ^ {v} = \operatorname {S u m} _ {\text {s c a t t e r}} (H, I ^ {v}), \quad H = \tilde {A} ^ {\top} V, \tag {11} +$$ + +$$ +\tilde {A} = \operatorname {S o f t m a x} _ {\text {s c a t t e r}} \left(K L ^ {\top}, I ^ {v}\right), \quad (Q, V) = \operatorname {P r o j} (G), \tag {12} +$$ + +where $\operatorname{Proj}(\cdot)$ represents the linear projection, and $I^v$ is the voxel indices. The Point Decoder reconstructs the output set from the enriched hidden features $\hat{H}^v$ . We start by broadcasting $\hat{H}^v$ based on $I^v$ , resulting in $\hat{H} \in \mathbb{R}^{m \times l \times d}$ . The output $\hat{G}$ is analogous to the operational paradigm of Voxel Encoder: + +$$ +\hat {G} = \tilde {A} ^ {*} ^ {\top} V ^ {*}, \quad \tilde {A} ^ {*} = \operatorname {S o f t m a x} \left(A ^ {*}\right), \tag {13} +$$ + +$$ +A ^ {*} = \left[ K _ {i} ^ {*} Q _ {i} ^ {\top} \right] _ {i = 1} ^ {m}, \quad \left(K ^ {*}, V ^ {*}\right) = \operatorname {P r o j} (\hat {\mathcal {H}}). \tag {14} +$$ + +Class-Aware Embedding Objective addresses the issue of non-uniqueness in embeddings produced by the AE, where various distinct inputs yield approximately the same embedding, leading to inaccurate representation [1, 25]. Our objective aims to facilitate the generation of AE embeddings that demonstrate robust semantic class discriminability. + +Specifically, we introduce a novel class-aware contrastive loss $\mathcal{L}_{\mathrm{contr}}$ in Eq. (15), which aims to maximize the distance between embeddings of different semantic classes while minimizing the distance of the same class. + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {c o n t r}} = \frac {1}{m} \sum_ {i = 1} ^ {m} \left[ \frac {1}{| P (i) |} \sum_ {p \in P (i)} \mathrm {R e L U} (\alpha - \mathrm {s i m} (H _ {i}, H _ {p})) \right. \\ \left. + \frac {1}{| N (i) |} \sum_ {n \in N (i)} \operatorname {R e L U} \left(\sin \left(H _ {i}, H _ {n}\right) - \alpha\right) \right], \tag {15} \\ \end{array} +$$ + +$$ +\mathcal {L} _ {\text {r e c o n}} = \frac {1}{m \cdot d} \sum_ {i = 1} ^ {m} \sum_ {j = 1} ^ {d} \left(G _ {i, j} - \hat {G} _ {i, j}\right) ^ {2}, \tag {16} +$$ + +where $i$ represents the point index; $P(i)$ and $N(i)$ are the indices of the nearest point in the same (Positive) and different class (Negative) relative to point $i$ , + +respectively; $\mathrm{sim}(\cdot)$ is a similarity measure; $\alpha$ is a margin controlling separation between classes. The pointwise representations $H_{i}$ , $H_{p}$ , and $H_{n}$ are broadcasted from $\hbar$ based on voxel indices $I^{v}$ . + +We further combine this with the MSE [39] reconstruction loss in Eq. (16). The overall loss function, with $\lambda$ balancing the fidelity of reconstruction with the distinctiveness of the embeddings, is formulated as: $\mathcal{L}_{\mathrm{total}} = \mathcal{L}_{\mathrm{recon}} + \lambda \mathcal{L}_{\mathrm{contr}}$ . + +# 5 RAPiD-Seg for 3D LiDAR Segmentation + +We present RAPiD-Seg, a 3D LiDAR segmentation network leveraging RAPiD features. It incorporates multiple complementary features via the channel-wise fusion mechanism [18]. RAPiD-Seg takes point cloud as input and performs single-modal (LiDAR-only) 3D semantic segmentation. Specifically, the input point cloud has three types of point-wise features: (a) coordinates-based features $F_{C} = \{p_{C} \mid p_{C} = (x,y,z)\}$ ; (b) intensity-based features $F_{I} = \{p_{I} \mid p_{I} = (I,J)\}$ , where $I$ and $J$ are intensity and reflectivity; (c) RAPiD features $F_{R} = \{p_{R} \mid p_{R} = \mathrm{RAPiD}(P_{\mathrm{RoI}};k)\}$ . The RoI varies depending on whether it is R- or C-RAPiD. + +As shown in Fig. 2, $F_{C} \oplus F_{I}$ and $F_{R}$ are fed to a VSA voxel encoder 4 and RAPiD AE 6 for voxelization, resulting in voxel-wise representation $E_{C}$ , $E_{I}$ and $E_{R}$ . Subsequently, we incorporate the complementary voxel-wise features via the channel-wise fusion mechanism (Sec. 5.1). The backbone net takes in them for LiDAR segmentation prediction. Additionally, we introduce two effective RAPiD-Seg variants to expedite the convergence of the extensive segmentation network towards an optimal solution. + +# 5.1 RAPiD Channel-Wise Fusion with Attention + +To combine various LiDAR point attributes, a highly efficient method for feature fusion is still a research topic [30, 50]. Current methods, which typically concatenate different LiDAR attributes (e.g., intensity [38, 41, 59], reflectivity [30], PLS [50]) face issues with increased dimensionality and complexity, and the need for balancing weights to avoid biased training towards dominant attributes. + +We thus propose a novel RAPiD Fusion with Attention module (FuAtten), pioneering application of channel-wise attention mechanism [18] in LiDAR point attribute fusion. With FuAtten, we effectively fuse $E_{C}$ , $E_{I}$ and $E_{R}$ , emphasizing informative features adaptively while suppressing less relevant ones across different channels. + +As shown in Fig. 2 ⑤ and Fig. 4, we initially concatenate $E_{C}$ , $E_{I}$ and $E_{R}$ along the voxel dimension, resulting in $\mathbf{E}$ . In the squeeze operation, Global Average Pooling (GAP) [33] condenses each channel $f'$ of the tensor into a representative channel descriptor. These descriptors form a channel-level embedding $\mathbf{z} = [z_{f'}]_{f' = 1}^{f^*}$ . + +$$ +\mathbf {z} = \left[ \operatorname {G A P} \left(\mathbf {E} _ {:,:: f ^ {\prime}}\right) \right] _ {f ^ {\prime} = 1} ^ {f ^ {*}} = \frac {1}{c \times l} \left[ \sum_ {c ^ {\prime} = 0} ^ {c - 1} \sum_ {l ^ {\prime} = 0} ^ {l - 1} \mathbf {E} _ {c ^ {\prime}, l ^ {\prime}, f ^ {\prime}} \right] _ {f ^ {\prime} = 1} ^ {f ^ {*}}, \tag {17} +$$ + +where $f^{*}$ is the dimension of the concatenated feature. A channel-wise attention $\pmb{a}_{\mathbf{z}} \in \mathbb{R}^{1 \times 1 \times f^{*}}$ is then computed through the excitation step: $\pmb{a}_{\mathbf{z}} =$ + +$\sigma (\mathbf{W}_2\delta (\mathbf{W}_1\mathbf{z}))$ , where $\delta$ and $\sigma$ represent the ReLU and Sigmoid activation. With each element in $a_{\mathbf{z}}$ reflects the attention allocated to the corresponding feature of the voxel, the channel-wise fused features $\mathbf{E}^{\prime} = a_{\mathbf{z}}\cdot \mathbf{E}$ + +![](images/dc886fcf06578cc3e62a634b3443e1019ce6c77fee60ca35399c435489a758fb.jpg) +Fig. 4: Our feature fusion module with channel-wise attention. + +# 5.2 RAPiD-Seg Architectures for 3D Segmentation + +The complexity inherent in 3D LiDAR-driven networks typically exacerbates the end-to-end training process [17,30,50], since the extensive parameter space increases the propensity for overfitting, slow convergence, and the possibility of settling into local minima [30]. + +![](images/b7e9f5e0bdcc3325367e44cee323ec1fd258f2afb3c56e1104922b348d5c9b2d.jpg) +Fig. 5: Two variants of RAPiD-Seg. (a) R-RAPiD-Seg utilizes R-RAPiD features, and (b) C-RAPiD-Seg utilizes both R- and C-RAPiD features for better performance. + +As shown in Fig. 5, we propose two novel and effective variants of RAPiD-Seg for quicker and better performance. R-RAPiD-Seg involves AE training from scratch, followed by C-RAPiD-Seg utilizing C-RAPiD features with the pre-trained AE and backbone for enhanced performance. + +R-RAPiD-Seg. We first construct the lightweight R-RAPiD-Seg (Fig. 5 (a)) for fast 3D segmentation. R-RAPiD-Seg adopts an early fusion scheme utilizing LiDAR original points and R-RAPiD features. Specifically, 3D coordinate $F_{C}$ and intensity-based features $F_{I}$ are voxelized to voxel-wise representations $R_{v} = E_{C} \oplus E_{I}$ based on VSA voxelization [15]. We get the compressed $d'$ -dimension voxel-wise R-RAPiD representations $\hbar$ (Sec. 4) from RAPiD AE. $\hbar$ is then fused into the voxel-wise representations $R_{v*} = \mathrm{FuAtten}(R_v,\hbar)$ , where $\mathrm{FuAtten}(\cdot)$ is Fusion with Attention in Sec. 5.1. The fused voxel-wise features are taken into the backbone network for 3D segmentation. + +C-RAPiD-Seg. To facilitate the embedding fidelity within individual semantic classes, we also design a class-aware framework with C-RAPiD features, i.e., C-RAPiD-Seg (Fig. 5 (b)). Specifically, we fuse the voxel-wise C-RAPiD $\hbar_C$ into the representations $R_{v*} = \mathrm{FuAtten}(R_v,\hbar \oplus \hbar_C)$ . C-RAPiD requires class labels to compute the features regarding the semantic categories. Since the ground-truth + +labels are missing during the test time, we generate the reliable pseudo labels $\tilde{\mathcal{V}}$ with a pretrained R-RAPiD-Seg based on confidence (Fig. 5 (b) blue dotted arrow). Subsequently, the pseudo voxel-wise C-RAPiD $\hbar_{C}$ are generated through $\tilde{\mathcal{V}}$ . We fuse them into the voxel-wise representations for 3D segmentation with the bonebone network. + +# 6 Evaluation + +Following the popular practice of LiDAR segmentation methods [17,37,64], we evaluate our proposed RAPiD-Seg network against SoTA 3D LiDAR segmentation approaches on the SemanticKITTI [2] and nuScenes [3] datasets. + +# 6.1 Experimental Setup + +Datasets: SemanticKITTI [2] comprises 22 point cloud sequences, with sequences 00-10, 08, and 11-21 for training, validation, and testing. 19 classes are chosen for training and evaluation by merging classes with similar motion statuses and discarding sparsely represented ones. Meanwhile, nuScenes [3] has 1000 driving scenes; 850 are for training and validation, with the remaining 150 for testing. 16 classes are used for LiDAR semantic segmentation, following the amalgamation of akin classes and the removal of rare ones. + +Evaluation Protocol: Following the popular practice of [2,36,64], we adopt the Intersection-over-Union (IoU) of each class and mean IoU (mIoU) of all classes as the evaluation metric. The IoU of class $i$ is defined as $\mathrm{IoU}_i = \mathrm{TP}_i / (\mathrm{TP}_i + \mathrm{FP}_i + \mathrm{FN}_i)$ , where $\mathrm{TP}_i$ , $\mathrm{FP}_i$ and $\mathrm{FN}_i$ denote the true positive, false positive and false negative of class $i$ , respectively. + +Implementation Details: We construct the point-voxel backbone based on the Minkowski-UNet34 [8] (re-implemented by PCSeg [36] codebase), which is the open-access backbone with SoTA results to date. Before the AE training stage, we first generate pointwise RAPiD features (Sec. 3), which yield three RAPiD outputs for each frame, corresponding to different $k$ values based on range. Notably, this stage does not impose additional burdens on the subsequent overall training. The number of maximum training epochs for AE and the whole network is set as 100 and the initial learning rate is set as $10^{-3}$ with SGD optimizer. We use 2 epochs to warm up the network and adopt the cosine learning rate schedule for the remaining epochs. All experiments are conducted on $4 \times$ NVIDIA A100 GPUs ( $1 \times$ for inference). + +# 6.2 Experimental Results + +In Tab. 1 and Tab. 2, we showcase the performance of our RAPiD-Seg LiDAR segmentation method on SemanticKITTI and nuScenes test set in comparison with published leading contemporary SoTA approaches to demonstrate its superior efficacy (top 10 results shown; full rankings in Supplementary Material). + +Our method significantly outperforms others, especially for rigid object categories (e.g., truck, o.veh, park, etc.), primarily due to our fusion of both localized 3D geometry and material reflectivity within RAPiD features, which + +Table 1: Quantitative results of RAPiD-Seg and top-10 SoTA segmentation methods on SemanticKITTI [2] test set; Best/2nd best highlighted; ▲ for multi-modal methods. + +
MethodmIoUcarbicymotortruco.vehpedb.listm.listroadparkwalko.grobuildfencvegtrunterrpolesign
Cylinder3D [64]68.997.167.663.850.858.573.769.248.092.265.077.032.390.766.585.672.569.862.466.2
AF2S3Net [7]69.794.565.486.839.241.180.780.474.391.368.872.553.587.963.270.268.553.761.571.0
RPVNet [56]70.397.668.468.744.261.175.974.473.493.470.380.733.393.572.186.575.171.764.861.4
SDSeg3D [27]70.497.458.754.254.965.270.274.452.290.969.476.741.993.271.186.174.371.165.470.6
GASN [61]70.796.965.858.059.361.080.482.746.389.866.274.630.192.369.687.373.072.566.171.6
PVKD [17]71.297.067.969.353.560.275.173.550.591.870.977.541.092.469.486.573.871.964.965.8
2DPASS [58]72.997.063.663.461.161.577.981.374.189.767.474.740.093.572.986.273.971.065.070.4
PCSeg [36]72.997.551.267.658.668.678.380.975.692.571.578.336.993.171.485.473.669.966.168.7
RangeFormer [22]73.396.769.473.759.966.278.175.958.192.473.078.842.492.370.186.673.372.866.466.6
▲UniSeg [37]75.297.971.975.263.674.178.974.860.692.674.079.546.193.472.787.576.373.168.368.5
RAPiD-Seg (Ours)76.197.771.176.272.580.779.979.159.891.878.278.646.093.672.186.974.672.365.968.5
+ +Table 2: Quantitative results of RAPiD-Seg and top-10 SoTA segmentation methods on nuScenes [3] test set; Best/2nd best highlighted; ▲ for multi-modal methods. + +
MethodmIoUbarrbicybuscarconstmotorpedconetrailtruckdrivotherwalkterrmadeveg
▲PMF [65]77.082.040.081.088.064.079.080.076.081.067.097.068.078.074.090.088.0
Cylinder3D [64]77.282.829.884.389.463.079.377.273.484.669.197.770.280.375.590.487.6
AMVNet [34]77.380.632.081.788.967.184.376.173.584.967.397.567.479.475.591.588.7
SPVCNN [49]77.480.030.091.990.864.779.075.670.981.074.697.469.280.076.189.387.1
AF2S3Net [7]78.378.952.289.984.277.474.377.372.083.973.897.166.577.574.087.786.8
▲2D3DNet [12]80.083.059.488.085.163.784.482.076.084.871.996.967.479.876.092.189.2
GASN [61]80.485.543.290.592.164.786.083.073.383.975.897.071.081.077.791.690.2
2DPASS [58]80.881.755.392.091.873.386.578.572.584.775.597.669.179.975.590.288.0
▲LidarMultiNet [60]81.480.448.494.390.071.587.285.280.486.974.897.867.380.776.592.189.6
▲UniSeg [37]83.585.971.292.191.680.588.080.976.086.376.797.771.880.776.791.388.8
RAPiD-Seg (Ours)83.684.864.395.092.284.687.981.876.888.579.097.866.681.276.792.588.4
+ +enable segmentation based on material properties and local rigid structures. Remarkably, our single-modal methodology outperforms multi-modal approaches [22, 34,37,56,58,60], suggesting superior efficacy of our RAPiD features over alternative modalities like RGB and range images. Our inference time (105ms per frame) is comparable to other contemporary approaches [37,58,64]. Furthermore, we present supporting qualitative results in Fig. 6. Whereas the baseline method struggles with accurate vehicle type differentiation, ours achieves consistent segmentation (more visualization results in the supplementary materials). + +
RAPiD FeaturesAttentionmIoUΔ
GeometricReflectivityEmbedding
70.04(Baseline)
✓70.46(+0.42)
✓71.21(+1.17)
✓✓71.93(+1.89)
✓✓✓72.15(+2.11)
✓✓✓71.80(+1.76)
✓✓✓72.32(+2.28)
✓✓✓72.78(+2.74)
✓✓✓✓73.02(+2.98)
+ +Table 3: Component-wise ablation of RAPiD-Seg on the SemanticKITTI validation set: Geometric (Eq. (1)), Reflectivity (Eq. (3)), (RAPiD) Embedding (Sec. 4), (channel) Attention (Sec. 5.1). + +![](images/7cb6032cfe40dba5254819957fa3c7df89b71a72f571df82c85e6ba790569fdd.jpg) +Fig.6: Comparing our results and PCSeg (baseline) under multi-scan visualization, showing improved segmentation results. + +# 6.3 Ablation Studies + +Effectiveness of Components: In Tab. 3, we ablate each component of RAPiD-Seg step by step and report the performance on the SemanticKITTI validation set. We start from a baseline model, achieving an mIoU of 70.04 on validation set. + +Incorporating the RAPiD features leads to a notable increase in mIoU (+1.17), underscoring the efficacy of RAPiD in enhancing segmentation performance. Building upon the RAPiD framework, the integration of reflectivity further elevates our mIoU to 71.93, with an overall improvement of +1.89 from the baseline. This significant gain illustrates the critical role of reflectivity in capturing different object materials. Utilizing RAPiD AE to get the voxel-wise representation, we observe an additional improvement, taking the mIoU to 72.15 (+2.11 compared to the baseline), which demonstrates its capability in processing feature representations. We subsequently performed a module-by-module reduction to analyze the model performance. Specifically, without RAPiD-related modules (Geometric/Reflectivity/Embedding), mIoU drops from 73.02 to 70.46 - attributable to the lack of RAPiD features, which are robust to transformation and noise. Our ablation studies show that while disabling individual components leads to some improvement over the baseline, the peak performance is only attained when all components function together. This highlights the indispensable contribution of each component in maximizing segmentation accuracy. + +Table 4: Effects of RAPiD and Reflectivity features compared to other configurations on SemanticKITTI validation set. PDD: the compressed PDD embeddings; RAPiD-R: only RAPiD, without reflectivity feature fusion; RAPiD-R: both RAPiD and reflectivity. + +
Conf.mIoUtrucko.vehparkwalko.grobuildfence
BaseL [36]70.059.870.369.276.936.293.769.6
PDD [52]66.240.365.867.574.633.192.868.2
RAPiD-R71.862.569.070.377.435.893.670.8
RAPiD+R73.070.478.575.878.944.294.273.1
+ +Effectiveness of RAPiD and Reflectivity Features: In Tab. 4, we validate the efficacy of the proposed RAPiD features. The direct application of PDD results in a substantial mIoU decrease of 3.8, with most category IoUs falling below those of the baseline method. This indicates vanilla PDD features are not well-suited for LiDAR segmentation. While our proposed RAPiD (without reflectivity disparities in Eq. (2)) demonstrates an improvement of 1.4 in mIoU over the baseline. Moreover, the integration of reflectivity with RAPiD features significantly enhances performance, yielding 2.6 mIoU increase over the baseline. Notably, our approach exhibits superior performance in the segmentation of most rigid object categories compared to other configurations. Tab. 6 further shows the RAPiD outperforms PDD in varying $k$ values at different ranges on SemanticKITTI validation set, with non-uniform $k$ configurations yielding the most significant improvements in mIoU for both datasets, highlighting the effectiveness of the range-aware design of RAPiD. + +Effectiveness of RAPiD AE and Architecture Variants: In Tab. 3, we replace our RAPiD AE with a commonly-used ConvFNN [17,50,64] to generate voxel-wise embeddings. Our AE demonstrated a $0.24\mathrm{mIoU}$ improvement over the FNN. In Tab. 5, we investigate the impact of the variant architectures, i.e., R-RAPiD-Seg and C-RAPiD-Seg. Utilizing R-RAPiD features, R-RAPiD-Seg yields an mIoU of 72.3, which outperforms the baseline method by $+2.3\mathrm{mIoU}$ . + +Table 5: 3D segmentation results of different variants of RAPiD-Seg (ours) on SemanticKITTI validation set. + +
MethodmIoU %carpedo.gropole
Baseline70.097.278.135.463.5
R-RAPiD-Seg72.3 (+2.3)97.477.445.062.4
C-RAPiD-Seg73.0 (+3.0)97.779.344.666.4
+ +Table 6: Effects of $k$ at various ranges. + +
PDD [53]RAPiD (ours)
knearkmidkfarmIoUknearkmidkfarmIoU
SumK77764.74 (-7.1)77772.04 (+0.2)
55565.18 (-6.6)55572.28 (+0.5)
107566.23 (-5.6)107573.02 (+1.2)
lnuScene66672.19 (-6.5)66678.76 (+0.1)
33373.68 (-5.0)33379.43 (+0.8)
86372.24 (-6.4)86379.91 (+1.3)
+ +Concurrent fusing of both R- and C-RAPiD-Seg features in R-RAPiD-Seg improves performance by $+0.7$ mIoU, which shows a $+3.0$ overall mIoU enhancement. + +Effectiveness of Feature Fusion with Channel-wise Attention: In Tab. 3, we assess the prevalent approach of direct feature concatenation. Our method, employing feature fusion with channel-wise attention, enhances mIoU by 0.87, thereby conclusively demonstrating its effectiveness. + +Effectiveness of Backbone Networks: In Tab. 7, we evaluate the performance across multiple backbone networks to demonstrate the versatility and ease of integration of our modules. The results highlight the adaptability and effectiveness of our approach on both pointwise (P) and Voxel-wise (V) backbone architectures. Specifically, the point-based PTv2 backbone achieves an mIoU of 72.6, showing strong performance, particularly in the car and other ground categories. Among the voxel-based methods, Minkowski-UNet stands out with the highest overall mIoU of 73.0, excelling in the pedestrian and pole categories. + +Table 7: Effects of using different backbones on SemanticKITTI validation set, where P and V for Point- and Voxel-wsie methods. + +
Repr.BackbonemIoU %carpedo.gropole
PPTv2 [55]72.6 (+2.8)97.477.445.062.4
VCylinder3D [64]69.896.974.237.963.0
VMinkowski-UNet [8]73.0 (+3.2)97.278.144.265.9
+ +# 7 Conclusion and Discussion + +This paper presents a novel Range-Aware Pointwise Distance Distribution (RAPiD) feature and the RAPiD-Seg network for LiDAR segmentation, adeptly overcoming the constraints of single-modal LiDAR methods. The rigid transformation invariance and enhanced focus on local details of RAPiD significantly boost segmentation accuracy. RAPiD-Seg integrates a two-stage training approach with reflectivity-guided 4D distance metrics and a class-aware nested AE, achieving SoTA results on the SemanticKITTI and nuScenes datasets. Notably, our single-modal method surpasses the performance of multi-modal methods, indicating the superior efficacy of RAPiD features even compared to other modalities, including RGB and range images. + +Our RAPiD features hold significant potential for application in various tasks such as point cloud registration, semi-/weakly- supervised learning, and extend to promising applications in multi-modal research in the future. + +# References + +1. Badger, B.L.: Depth and Representation in Vision Models (2023) +2. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences. In: Int. Conf. Comput. Vis. pp. 9296-9306 (2019) +3. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: Nuscenes: A Multimodal Dataset for Autonomous Driving. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 11618-11628 (2020) +4. Chen, S., Wang, X., Cheng, T., Zhang, W., Zhang, Q., Huang, C., Liu, W.: Azinorm: Exploiting the Radial Symmetry of Point Cloud for Azimuth-Normalized 3D Perception. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6387-6396 (2022) +5. Chen, X., Xu, S., Zou, X., Cao, T., Yeung, D.Y., Fang, L.: SVQNet: Sparse Voxel-Adjacent Query Network for 4D Spatio-Temporal LiDAR Semantic Segmentation. In: Int. Conf. Comput. Vis. pp. 8569–8578 (2023) +6. Chen, X., Shi, S., Zhu, B., Cheung, K.C., Xu, H., Li, H.: MPPNet: Multi-frame Feature Intertwining with Proxy Points for 3D Temporal Object Detection. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Eur. Conf. Comput. Vis. pp. 680-697. Springer Nature Switzerland (2022) +7. Cheng, R., Razani, R., Taghavi, E., Li, E., Liu, B.: (AF)2-S3Net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12547-12556 (2021) +8. Choy, C., Gwak, J., Savarese, S.: 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3070-3079 (2019) +9. Çiçek, Ö., Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation. In: Ourselin, S., Joskowicz, L., Sabuncu, M.R., Unal, G., Wells, W. (eds.) Med. Image Comput. Comput. Assist. Interv. pp. 424–432. Springer International Publishing (2016) +10. Flitton, G., Breckon, T.P., Megherbi, N.: A Comparison of 3D Interest Point Descriptors with Application to Airport Baggage Object Detection in Complex CT Imagery. Pattern Recognition 46(9), 2420-2436 (2013) +11. Geiger, A., Lenz, P., Urtasun, R.: Are We Ready for Autonomous Driving? The KITTI Vision Benchmark Suite. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3354-3361 (2012) +12. Genova, K., Yin, X., Kundu, A., Pantofaru, C., Cole, F., Sud, A., Brewington, B., Shucker, B., Funkhouser, T.: Learning 3D Semantic Segmentation with only 2D Image Supervision. In: Int. Conf. 3D Vis. pp. 361-372 (2021) +13. Graham, B., Engelcke, M., van der Maaten, L.: 3D Semantic Segmentation with Submanifold Sparse Convolutional Networks. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018) +14. Han, L., Zheng, T., Xu, L., Fang, L.: OccuSeg: Occupancy-Aware 3D Instance Segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2940-2949 (2020) +15. He, C., Li, R., Li, S., Zhang, L.: Voxel Set Transformer: A Set-to-Set Approach to 3D Object Detection from Point Clouds. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 8407-8417. IEEE (2022) +16. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked Autoencoders Are Scalable Vision Learners. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 16000-16009 (2022) + +17. Hou, Y., Zhu, X., Ma, Y., Loy, C.C., Li, Y.: Point-to-Voxel Knowledge Distillation for LiDAR Semantic Segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) +18. Hu, J., Shen, L., Sun, G.: Squeeze-and-Eccitation Networks. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018) +19. Ioffe, S., Szegedy, C.: Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. In: Int. Conf. Mach. Learn. pp. 448-456 (2015) +20. Jiang, L., Shi, S., Tian, Z., Lai, X., Liu, S., Fu, C.W., Jia, J.: Guided Point Contrastive Learning for Semi-supervised Point Cloud Semantic Segmentation. In: Int. Conf. Comput. Vis. (2021) +21. Jiang, M., Wu, Y., Zhao, T., Zhao, Z., Lu, C.: PointSIFT: A SIFT-like Network Module for 3D Point Cloud Semantic Segmentation (2018) +22. Kong, L., Liu, Y., Chen, R., Ma, Y., Zhu, X., Li, Y., Hou, Y., Qiao, Y., Liu, Z.: Rethinking Range View Representation for LiDAR Segmentation. In: Int. Conf. Comput. Vis. (2023) +23. Kong, L., Ren, J., Pan, L., Liu, Z.: LaserMix for Semi-Supervised LiDAR Semantic Segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 21705-21715 (2023) +24. Kortvelesy, R., Morad, S., Prorok, A.: Permutation-Invariant Set Autoencoders with Fixed-Size Embeddings for Multi-Agent Learning. In: Int. Conf. Auton. Agents Multiagent Syst. pp. 1661-1669. International Foundation for Autonomous Agents and Multiagent Systems (2023) +25. Kumar, A., Poole, B.: On Implicit Regularization in $\beta$ -VAEs. In: Int. Conf. Mach. Learn. pp. 5480-5490 (2020) +26. Li, E., Casas, S., Urtasun, R.: MemorySeg: Online LiDAR Semantic Segmentation with a Latent Memory. In: Int. Conf. Comput. Vis. pp. 745-754 (2023) +27. Li, J., Dai, H., Ding, Y.: Self-Distillation for Robust LiDAR Semantic Segmentation in Autonomous Driving. In: Eur. Conf. Comput. Vis. pp. 659-676 (2022) +28. Li, J., Dai, H., Han, H., Ding, Y.: MSeg3D: Multi-Modal 3D Semantic Segmentation for Autonomous Driving. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 21694-21704 (2023) +29. Li, L., Ismail, K.N., Shum, H.P.H., Breckon, T.P.: DurLAR: A High-Fidelity 128-Channel LiDAR Dataset with Panoramic Ambient and Reflectivity Imagery for Multi-Modal Autonomous Driving Applications. In: Int. Conf. 3D Vis. pp. 1227-1237 (2021) +30. Li, L., Shum, H.P.H., Breckon, T.P.: Less is More: Reducing Task and Model Complexity for Semi-Supervised 3D Point Cloud Semantic Segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) +31. Li, X., Li, R., Chen, G., Fu, C.W., Cohen-Or, D., Heng, P.A.: A Rotation-Invariant Framework for Deep Point Cloud Analysis. IEEE Trans. Vis. Comput. Graph. 28(12), 4503-4514 (2022) +32. Liang, J., Lai, R., Wong, T.W., Zhao, H.: Geometric Understanding of Point Clouds Using Laplace-Beltrami Operator. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 214-221 (2012) +33. Lin, M., Chen, Q., Yan, S.: Network In Network. In: Int. Conf. Learn. Represent. (2014) +34. Liong, V.E., Nguyen, T.N.T., Widjaja, S., Sharma, D., Chong, Z.J.: AMVNet: Assertion-based Multi-View Fusion Network for LiDAR Semantic Segmentation (2020) +35. Liu, J., Yu, Z., Breckon, T.P., Shum, H.P.H.: U3DS3: Unsupervised 3D Semantic Scene Segmentation. In: IEEE Winter Conf. Appl. Comput. Vis. pp. 3759-3768 (2024) + +36. Liu, Y., Bai, Y., Chen, R., Hou, Y., Shi, B., Li, Y., Kong, L.: PCSeg: An Open Source Point Cloud Segmentation Codebase. https://github.com/PJLab-ADG/PCSeg (2023) +37. Liu, Y., Chen, R., Li, X., Kong, L., Yang, Y., Xia, Z., Bai, Y., Zhu, X., Ma, Y., Li, Y., Qiao, Y., Hou, Y.: UniSeg: A Unified Multi-Modal LiDAR Segmentation Network and the OpenPCSeg Codebase. In: Int. Conf. Comput. Vis. pp. 21662–21673 (2023) +38. Liu, Y., Kong, L., Cen, J., Chen, R., Zhang, W., Pan, L., Chen, K., Liu, Z.: Segment Any Point Cloud Sequences by Distilling Vision Foundation Models. In: Adv. Neural Inform. Process. Syst. (2023) +39. Marmolin, H.: Subjective MSE measures. IEEE Trans. Syst. Man Cybern. 16(3), 486-489 (1986) +40. Melia, O., Jonas, E., Willett, R.: Rotation-Invariant Random Features Provide a Strong Baseline for Machine Learning on 3D Point Clouds. Trans. Mach. Learn. Res. (2023) +41. Nunes, L., Wiesmann, L., Marcuzzi, R., Chen, X., Behley, J., Stachniss, C.: Temporal Consistent 3D LiDAR Representation Learning for Semantic Perception in Autonomous Driving. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) +42. Preechakul, K., Chathee, N., Wizadwongsa, S., Suwajanakorn, S.: Diffusion Autoencoders: Toward a Meaningful and Decodable Representation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 10619-10629 (2022) +43. PyTorch: Scatter: PyTorch Scatter Documentation. https://pytorch-scatter.readthedocs.io/en/latest/functions/scatter.html (2023) +44. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 652-660 (2017) +45. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space. In: Adv. Neural Inform. Process. Syst. vol. 30 (2017) +46. Shao, Y., Tan, A., Yan, T., Sun, Z., Zhang, Y., Liu, J.: MS23D: A 3D Object Detection Method Using Multi-Scale Semantic Feature Points to Construct 3D Feature Layer (2023) +47. Sun, J., Xu, X., Kong, L., Liu, Y., Li, L., Zhu, C., Zhang, J., Xiao, Z., Chen, R., Wang, T., et al.: An Empirical Study of Training State-of-the-Art LiDAR Segmentation Models. arXiv:2405.14870 (2024) +48. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in Perception for Autonomous Driving: Waymo Open Dataset. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2446-2454 (2020) +49. Tang, H., Liu, Z., Zhao, S., Lin, Y., Lin, J., Wang, H., Han, S.: Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution. In: Eur. Conf. Comput. Vis. (2020) +50. Unal, O., Dai, D., Van Gool, L.: Scribble-Supervised LiDAR Semantic Segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) +51. Wang, S., Zhu, J., Zhang, R.: Meta-RangeSeg: LiDAR Sequence Semantic Segmentation Using Multiple Feature Aggregation. IEEE Robot. Autom. Lett. 7(4), 9739-9746 (2022) +52. Widdowson, D., Kurlin, V.: Resolving the Data Ambiguity for Periodic Crystals. In: Adv. Neural Inform. Process. Syst. vol. 35, pp. 24625-24638 (2022) + +53. Widdowson, D., Kurlin, V.: Recognizing Rigid Patterns of Unlabeled Point Clouds by Complete and Continuous Isometry Invariants With No False Negatives and No False Positives. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1275-1284 (2023) +54. Widdowson, D., Mosca, M.M., Pulido, A., Cooper, A.I., Kurlin, V.: Average Minimum Distances of Periodic Point Sets - Foundational Invariants for Mapping Periodic Crystals. MATCH Commun. Math. Comput. Chem. 87(3), 529-559 (2022) +55. Wu, X., Lao, Y., Jiang, L., Liu, X., Zhao, H.: Point Transformer V2: Grouped Vector Attention and Partition-Based Pooling. In: Adv. Neural Inform. Process. Syst. (2022) +56. Xu, J., Zhang, R., Dou, J., Zhu, Y., Sun, J., Pu, S.: RPVNet: A Deep and Efficient Range-Point-Voxel Fusion Network for LiDAR Point Cloud Segmentation. In: Int. Conf. Comput. Vis. pp. 16024–16033 (2021) +57. Yan, X., Gao, J., Li, J., Zhang, R., Li, Z., Huang, R., Cui, S.: Sparse Single Sweep LiDAR Point Cloud Segmentation via Learning Contextual Shape Priors from Scene Completion. In: Conf. AAAI Artif. Intell. vol. 35, pp. 3101-3109 (2021) +58. Yan, X., Gao, J., Zheng, C., Zheng, C., Zhang, R., Cui, S., Li, Z.: 2DPASS: 2D Priors Assisted Semantic Segmentation on LiDAR Point Clouds. In: Eur. Conf. Comput. Vis. (2022) +59. Ye, D., Chen, W., Zhou, Z., Xie, Y., Wang, Y., Wang, P., Foroosh, H.: LidarMultiNet: Unifying LiDAR Semantic Segmentation, 3D Object Detection, and Panoptic Segmentation in a Single Multi-task Network. In: IEEE Conf. Comput. Vis. Pattern Recog. Worksh. (2022) +60. Ye, D., Zhou, Z., Chen, W., Xie, Y., Wang, Y., Wang, P., Foroosh, H.: LidarMultiNet: Towards a Unified Multi-Task Network for Lidar Perception. In: Conf. AAAI Artif. Intell. vol. 37, pp. 3231-3240 (2023) +61. Ye, M., Wan, R., Xu, S., Cao, T., Chen, Q.: Efficient Point Cloud Segmentation with Geometry-Aware Sparse Networks. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Eur. Conf. Comput. Vis. pp. 196-212. Springer (2022) +62. Zhang, C., Duc, K.D., Condon, A.: Visualizing DNA Reaction Trajectories with Deep Graph Embedding Approaches. In: Adv. Neural Inform. Process. Syst. Worksh. arXiv (2022) +63. Zhao, H., Jiang, L., Jia, J., Torr, P., Koltun, V.: Point Transformer. In: Int. Conf. Comput. Vis. (2021) +64. Zhu, X., Zhou, H., Wang, T., Hong, F., Ma, Y., Li, W., Li, H., Lin, D.: Cylindrical and Asymmetrical 3D Convolution Networks for LiDAR Segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021) +65. Zhuang, Z., Li, R., Jia, K., Wang, Q., Li, Y., Tan, M.: Perception-Aware Multi-Sensor Fusion for 3D LiDAR Semantic Segmentation. In: Int. Conf. Comput. Vis. pp. 16280-16290 (2021) \ No newline at end of file diff --git a/2024/RAPiD-Seg_ Range-Aware Pointwise Distance Distribution Networks for 3D LiDAR Segmentation/images.zip b/2024/RAPiD-Seg_ Range-Aware Pointwise Distance Distribution Networks for 3D LiDAR Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..37dfb0ca0752bf7a264ad09988c72092dc1aeab5 --- /dev/null +++ b/2024/RAPiD-Seg_ Range-Aware Pointwise Distance Distribution Networks for 3D LiDAR Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:055eea779f74195f8bd963aad56b403194f828bc5fff62806b8c38daa043a4bb +size 595491 diff --git a/2024/RAPiD-Seg_ Range-Aware Pointwise Distance Distribution Networks for 3D LiDAR Segmentation/layout.json b/2024/RAPiD-Seg_ Range-Aware Pointwise Distance Distribution Networks for 3D LiDAR Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e7993530de708b3c29e8b0baf758cd320be54301 --- /dev/null +++ b/2024/RAPiD-Seg_ Range-Aware Pointwise Distance Distribution Networks for 3D LiDAR Segmentation/layout.json @@ -0,0 +1,11830 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 112, + 470, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 112, + 470, + 148 + ], + "spans": [ + { + "bbox": [ + 143, + 112, + 470, + 148 + ], + "type": "text", + "content": "RAPiD-Seg: Range-Aware Pointwise Distance Distribution Networks for 3D LiDAR Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 179, + 167, + 433, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 167, + 433, + 180 + ], + "spans": [ + { + "bbox": [ + 179, + 167, + 433, + 180 + ], + "type": "text", + "content": "Li Li" + }, + { + "bbox": [ + 179, + 167, + 433, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 179, + 167, + 433, + 180 + ], + "type": "text", + "content": ", Hubert P. H. Shum" + }, + { + "bbox": [ + 179, + 167, + 433, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 179, + 167, + 433, + 180 + ], + "type": "text", + "content": ", and Toby P. Breckon" + }, + { + "bbox": [ + 179, + 167, + 433, + 180 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 149, + 190, + 463, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 190, + 463, + 213 + ], + "spans": [ + { + "bbox": [ + 149, + 190, + 463, + 213 + ], + "type": "text", + "content": "Department of " + }, + { + "bbox": [ + 149, + 190, + 463, + 213 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{Computer~Science}^1\\mid" + }, + { + "bbox": [ + 149, + 190, + 463, + 213 + ], + "type": "text", + "content": " Engineering2}, Durham University, UK {li.li4, hubert.shum, toby.breckon}@durham.ac.uk" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 220, + 455, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 220, + 455, + 483 + ], + "spans": [ + { + "bbox": [ + 160, + 220, + 455, + 483 + ], + "type": "text", + "content": "Abstract. 3D point clouds play a pivotal role in outdoor scene perception, especially in the context of autonomous driving. Recent advancements in 3D LiDAR segmentation often focus intensely on the spatial positioning and distribution of points for accurate segmentation. However, these methods, while robust in variable conditions, encounter challenges due to sole reliance on coordinates and point intensity, leading to poor isometric invariance and suboptimal segmentation. To tackle this challenge, our work introduces Range-Aware Pointwise Distance Distribution (RAPiD) features and the associated RAPiD-Seg architecture. Our RAPiD features exhibit rigid transformation invariance and effectively adapt to variations in point density, with a design focus on capturing the localized geometry of neighboring structures. They utilize inherent LiDAR isotropic radiation and semantic categorization for enhanced local representation and computational efficiency, while incorporating a 4D distance metric that integrates geometric and surface material reflectivity for improved semantic segmentation. To effectively embed high-dimensional RAPiD features, we propose a double-nested autoencoder structure with a novel class-aware embedding objective to encode high-dimensional features into manageable voxel-wise embeddings. Additionally, we propose RAPiD-Seg which incorporates a channel-wise attention fusion and two effective RAPiD-Seg variants, further optimizing the embedding for enhanced performance and generalization. Our method outperforms contemporary LiDAR segmentation work in terms of mIoU on SemanticKITTI (76.1) and nuScenes (83.6) datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 493, + 453, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 493, + 453, + 504 + ], + "spans": [ + { + "bbox": [ + 160, + 493, + 453, + 504 + ], + "type": "text", + "content": "Keywords: Semantic Segmentation " + }, + { + "bbox": [ + 160, + 493, + 453, + 504 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 493, + 453, + 504 + ], + "type": "text", + "content": " Point Cloud " + }, + { + "bbox": [ + 160, + 493, + 453, + 504 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 493, + 453, + 504 + ], + "type": "text", + "content": " Invariance Features" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 511, + 230, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 511, + 230, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 511, + 230, + 523 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 483, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 483, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 483, + 668 + ], + "type": "text", + "content": "LiDAR-based point clouds, characterized by pointwise 3D positions and LiDAR intensity/reflectivity [3,11,29,48], play a pivotal role in outdoor scene understanding, particularly in perception systems for autonomous driving. Recent literature in the domain of 3D LiDAR segmentation has witnessed significant strides, with myriad approaches attempting pointwise 3D semantic scene labeling [47]. Leveraging color [7,37,38], range imagery [22,34], and Birds Eye View (BEV) projections [34,60], multi-modal methods integrate diverse data streams from LiDAR and other sensors to enhance feature representation and segmentation accuracy. Single-modal (LiDAR-only) methods [17,30,56,64] effectively exploit spatial data interpretation and avoid multi-modal issues, e.g., modality heterogeneity and limited sensor field of view intersection [5]." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 136, + 246, + 209 + ], + "blocks": [ + { + "bbox": [ + 133, + 136, + 246, + 209 + ], + "lines": [ + { + "bbox": [ + 133, + 136, + 246, + 209 + ], + "spans": [ + { + "bbox": [ + 133, + 136, + 246, + 209 + ], + "type": "image", + "image_path": "7ae4750af076c4f446545f4f228ce856c0601e8017fad9cd304cbe0044e475cf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 211, + 482, + 289 + ], + "lines": [ + { + "bbox": [ + 130, + 211, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 211, + 482, + 289 + ], + "type": "text", + "content": "Fig. 1: Left: RAPiD exhibits excellent viewpoint invariance and geometric stability, visualizing comparable features around the vehicle door structure at varying ranges and viewpoints (feature matrix plots inset). Middle: RAPiD is distinctive in different semantic classes, as visualized by the matrices. Embedded RAPiD patterns corresponding to different points are visualized using a spectrum of colors, showcasing their capacity to represent different classes. Right: Our RAPiD-Seg achieves superior results over SoTA methods on nuScenes [3] and SemanticKITTI [2]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 250, + 135, + 393, + 209 + ], + "blocks": [ + { + "bbox": [ + 250, + 135, + 393, + 209 + ], + "lines": [ + { + "bbox": [ + 250, + 135, + 393, + 209 + ], + "spans": [ + { + "bbox": [ + 250, + 135, + 393, + 209 + ], + "type": "image", + "image_path": "7d83ae89b7fde62c769474d221a118e7062d275753d69257cdd064fce416cb6f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 395, + 136, + 482, + 208 + ], + "blocks": [ + { + "bbox": [ + 395, + 136, + 482, + 208 + ], + "lines": [ + { + "bbox": [ + 395, + 136, + 482, + 208 + ], + "spans": [ + { + "bbox": [ + 395, + 136, + 482, + 208 + ], + "type": "image", + "image_path": "18dd574e87204f7f2dc16320e42d5696fc0be013651de4211c47c9508c79309c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 304, + 482, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 365 + ], + "type": "text", + "content": "Despite these advancements, a common challenge inherent to these methods is their poor isometric invariance, characterized by heavy reliance on solely coordinates and intensity data, which often results in suboptimal segmentation outcomes [30,64]; this is primarily due to poor translational invariance and visibility (e.g., occlusions), or sparse observations (e.g., at long range) [26], affecting data spatial distribution." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 366, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 498 + ], + "type": "text", + "content": "In this work, we seek features that are (1) capable of capturing the localized geometric structure of neighboring points, (2) invariant to rotation and translation, and (3) applicable in noisy LiDAR outdoor environments. While numerous methods fulfill some of these requirements individually [21, 32, 40], they fall short of addressing them all [10]. Recognizing the need for higher-level features capable of capturing local geometry while potentially incorporating LiDAR point-specific attributes (e.g., intensity and reflectivity), we instead turn our attention to pointwise Distance Distribution (PDD) features [52, 53]. They are recognized for their remarkable efficacy in providing a robust and informative geometric representation of point clouds, excelling in both rotational and translational invariance while including intricate local geometry details." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 482, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 546 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 546 + ], + "type": "text", + "content": "However, directly employing PDD features is impractical due to their high-dimensional nature, and their significant memory and storage requirements for large-scale point clouds. PDD also overlooks local features, as including distant points subsequently dilutes the focus on immediate neighborhoods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 546, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 483, + 666 + ], + "type": "text", + "content": "As an enabler, we propose the Range-Aware Pointwise Distance Distribution (RAPiD-Seg) solution for LiDAR segmentation. It utilizes invariance of RAPiD features to rigid transformations and point cloud sparsity variations, while concentrating on compact features within specific local neighborhoods. Specifically, our method leverages inherent LiDAR isotropic radiation and semantic categorization for enhanced local representations while reducing computational burdens. Moreover, our formulation computes a 4D distance, incorporating both 3D geometric and reflectivity differences to enhance semantic segmentation fidelity. To compress high-dimensional RAPiD features into tractable voxel-wise embeddings, we propose a novel embedding approach with our class-aware double-nested AutoEn" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "type": "text", + "content": "L. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "coder (AE) module. We further incorporate a channel-wise attention fusion and a two-stage training strategy to optimize the AE independently before integrating into the full network, enhancing performance and generalization capability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 187 + ], + "type": "text", + "content": "We conduct extensive experiments on SemanticKITTI [2] and nuScenes [3] datasets, upon which our approach surpasses the state-of-the-art (SoTA) segmentation performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 188, + 481, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 188, + 481, + 200 + ], + "spans": [ + { + "bbox": [ + 146, + 188, + 481, + 200 + ], + "type": "text", + "content": "Overall, as shown in Fig. 2, our contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 201, + 481, + 323 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 201, + 481, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 201, + 481, + 249 + ], + "spans": [ + { + "bbox": [ + 138, + 201, + 481, + 249 + ], + "type": "text", + "content": "A novel Range-Aware Pointwise Distance Distribution feature (RAPiD) for 3D LiDAR segmentation that ensures robustness to rigid transformations and viewpoints through isometry-invariant metrics within specific Regions of Interest, i.e., intra-ring (R-RAPiD) and intra-class (C-RAPiD)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 251, + 481, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 481, + 285 + ], + "type": "text", + "content": "A novel method for embedding RAPiD with class-aware double nested AE (RAPiD AE) to optimize the embedding of high-dimensional features, balancing efficiency and fidelity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 286, + 481, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 286, + 481, + 323 + ], + "spans": [ + { + "bbox": [ + 138, + 286, + 481, + 323 + ], + "type": "text", + "content": "A novel open-source network architecture RAPiD-Seg " + }, + { + "bbox": [ + 138, + 286, + 481, + 323 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 286, + 481, + 323 + ], + "type": "text", + "content": " and supporting training methodology, to enable modular LiDAR segmentation that achieves SoTA performance on SemanticKITTI (mIoU: 76.1) and nuScenes (mIoU: 83.6)." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 133, + 338, + 479, + 450 + ], + "blocks": [ + { + "bbox": [ + 133, + 338, + 479, + 450 + ], + "lines": [ + { + "bbox": [ + 133, + 338, + 479, + 450 + ], + "spans": [ + { + "bbox": [ + 133, + 338, + 479, + 450 + ], + "type": "image", + "image_path": "d94a63e051fc0935ca7cf572716fc202e2be2bbf220b3bdd9c553111b078c401.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 455, + 480, + 499 + ], + "lines": [ + { + "bbox": [ + 130, + 455, + 480, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 455, + 480, + 499 + ], + "type": "text", + "content": "Fig. 2: Our proposed architecture for 3D segmentation framework leverages RAPiD features from the point cloud. We encode pointwise features into voxel-wise embeddings via the voxel encoder and multiple RAPiD AutoEncoders (RAPiD AE). After attention-based feature fusion, these fused embeddings go through the backbone network for segmentation results." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 517, + 237, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 517, + 237, + 530 + ], + "spans": [ + { + "bbox": [ + 131, + 517, + 237, + 530 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 539, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 539, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 539, + 482, + 647 + ], + "type": "text", + "content": "LiDAR-Based Semantic Segmentation [7,12,17,20,22,23,27,30,34,37,38,49, 50,56-58,60,61,64,65] is fundamental for LiDAR-driven scene perception, aiming to label each point in a point cloud sequence. The majority of the approaches [17, 38,49,57,64] solely rely on the point-based features of the point cloud, such as SPVCNN [49] which introduces a point-to-voxel branch, using combined pointvoxel features for segmentation. Cylinder3D [64] proposes cylindrical partitioning with a UNet [9] backbone variant. LiM3D [30] utilizes coordinates combined with surface reflectivity attributes. Overall, such prior work relies solely on pointbased features such as coordinates and intensity of the points, lacking an effective" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "type": "text", + "content": "RAPiD-Seg Networks for 3D LiDAR Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 441, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 441, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 441, + 665 + ], + "type": "text", + "content": "1 The code is publicly available at: https://github.com/l1997i/rapid_seg." + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "content": "fusion mechanism, resulting in suboptimal performance [6,28,51]. They are also susceptible to changes in viewpoint, distance, and point sparsity [26] due to the lack of isometry across all inter-point distances." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 402 + ], + "type": "text", + "content": "Pointwise Distance Distribution (PDD) captures the local context of each point in a unit cell by enumerating distances to neighboring points in order. It is an isometry invariant proposed by Widdowson & Kurlin [52] to resolve the data ambiguity for periodic crystals, demonstrated through extensive pairwise comparisons across atomic 3D clouds from high-level periodic crystals of periodic structures [52-54]. Though the effectiveness of PDD in periodic crystals and atomic clouds has been proved by the aforementioned studies, to date, no work has applied PDD features to outdoor 3D point clouds. In outdoor settings, common invariant features [21,31,32,40] and Azimuth-normalized method [4] often face issues due to the irregular and sparse nature of the data, which is compounded by increased noise and environmental complexities [31,32]. Furthermore, the computational demands make them less suitable for the vast scale of outdoor settings [21,40]. For instance, Melia et al. [40] introduce a rotation-invariant feature that struggles with generalization across diverse point cloud densities and scales due to its computational cost and susceptibility to outdoor noise. Recognizing these limitations, we identify an opportunity to leverage the PDD features in a new domain, where representing the local context of neighboring points in a transformation invariant and geometrically robust manner is crucial. Drawing from the advantages of the PDD design, we propose the RAPiD feature, tailored specifically for LiDAR-based point clouds, to capture the localized geometry of neighboring structures." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 403, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 594 + ], + "type": "text", + "content": "Point Cloud Representations and Embeddings are fundamental for LiDAR-driven tasks. 3D point cloud can be represented either as point-based [44,45,55,63] or voxel-based [17,35,50,64], each serving different purposes rooted in their unique characteristics. Voxel-based representation involves dividing the 3D space into grids of voxels, which provides a structured way of handling sparse and irregular LiDAR data [9, 13, 14, 64]. Previous work often utilizes Feedforward Neural Network (FNN) [17, 50, 64] to facilitate the conversion of pointwise features into structured voxel-wise features by projecting them onto voxel grids and generating dense embeddings. The primary drawback of FNN conversion is the potential loss of fine-grained details inherent in the original point cloud [15]. FNN may also struggle with the high dimensionality and sparsity of voxelized data, leading to computational inefficiency [46]. Our method addresses these issues by partitioning the point cloud into voxel grids and applying a novel class-ware double nested AE with self-attention mechanisms [15]. This approach not only retains more detailed information by focusing on local structures but also introduces high-dimensional feature compression, thereby improving computational efficiency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 607, + 410, + 622 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 607, + 410, + 622 + ], + "spans": [ + { + "bbox": [ + 131, + 607, + 410, + 622 + ], + "type": "text", + "content": "3 Range-Aware Pointwise Distances (RAPiD)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "In the context of LiDAR segmentation, a principal challenge lies in the lack of robustness against rigid transformations such as rotation and translation, as well as against variations in viewpoint, points sparsity, and occlusion [5, 26]." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "type": "text", + "content": "L. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "Traditional methods predominantly leverage 3D point coordinates to furnish spatial information [17,50,64]; however, they may be inadequate in the scenes with poor visibility (e.g., occlusions) or sparse observations (e.g., at long range) [26]. Such reliance on coordinates alone could lead to inaccuracies in recognizing object transformations or occlusions [26]. Data augmentation, e.g., random geometric transformation can improve robustness under rigid transformations [50,64], but fail to guarantee comprehensive coverage of all potential transformations, resulting in vulnerability to previously unseen variations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 228, + 208, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 208, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 208, + 239 + ], + "type": "text", + "content": "3.1 Overview" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 246, + 482, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 482, + 365 + ], + "type": "text", + "content": "To achieve a transformation-invariant 3D data representation, we observe that distances within rigid bodies (e.g., vehicles, roads, and buildings) remain constant under rigid transformations. In light of this, we focus on assimilating the principle of the isometric invariant into the LiDAR-driven point cloud perception. Specifically, we delve into the PDD [52] — an isometry invariant that quantifies distances between adjacent points. For outdoor point clouds, vanilla PDD features are computationally intensive and susceptible to noise and sparsity. Our Range-Aware Pointwise Distance Distribution (RAPiD), specifically designed for LiDAR data, instead calculates the PDD features for each point within specific Regions of Interest (RoI), which are typically associated with the intrinsic structure of LiDAR data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "text", + "content": "Mathematical Formulation: Given a fixed number " + }, + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "inline_equation", + "content": "k > 0" + }, + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "text", + "content": " representing the fixed number of point neighbors and a " + }, + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "text", + "content": "-point cluster " + }, + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{RoI}}" + }, + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "text", + "content": " comprising no fewer than " + }, + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "text", + "content": " points based on RoI, the RAPiD is a " + }, + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "inline_equation", + "content": "u \\times k" + }, + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "text", + "content": " matrix, which retains both spatial distances and LiDAR reflectivity [30] disparities between points. RAPiD are adapted to LiDAR sparsity at different distances by using range-specific parameters " + }, + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "inline_equation", + "content": "k_{\\mathrm{close}}" + }, + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "inline_equation", + "content": "k_{\\mathrm{mid}}" + }, + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "inline_equation", + "content": "k_{\\mathrm{far}}" + }, + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "text", + "content": " for close, mid, and far ranges, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 146, + 437, + 359, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 437, + 359, + 449 + ], + "spans": [ + { + "bbox": [ + 146, + 437, + 359, + 449 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 146, + 437, + 359, + 449 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 146, + 437, + 359, + 449 + ], + "type": "text", + "content": "-point RAPiD in region " + }, + { + "bbox": [ + 146, + 437, + 359, + 449 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{RoI}}" + }, + { + "bbox": [ + 146, + 437, + 359, + 449 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 197, + 450, + 481, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 450, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 197, + 450, + 481, + 468 + ], + "type": "interline_equation", + "content": "\\operatorname {R A P i D} \\left(P _ {\\text {R o I}}; k\\right) = \\operatorname {s o r t} \\left(\\left[ \\operatorname {s o r t} \\left(\\left[ \\boldsymbol {\\rho} _ {j, 1}, \\dots , \\boldsymbol {\\rho} _ {j, k} \\right]\\right) \\right] _ {j = 1} ^ {u}\\right), \\tag {1}", + "image_path": "d674025b5ce431da9af860f2f7ba9f1f965e610de29ef1be1f147285ccb98b7e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 469, + 341, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 469, + 341, + 483 + ], + "spans": [ + { + "bbox": [ + 131, + 469, + 341, + 483 + ], + "type": "inline_equation", + "content": "\\forall l \\in \\{1, \\dots, k\\}, j \\in \\{1, \\dots, u\\}, \\pmb{\\rho}_{j,l}" + }, + { + "bbox": [ + 131, + 469, + 341, + 483 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 222, + 488, + 481, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 488, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 222, + 488, + 481, + 504 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\rho} _ {j, l} = \\left\\| \\left[ \\boldsymbol {p} _ {j} - \\boldsymbol {p} _ {j, l}, \\quad g (r _ {j}) - g (r _ {j, l}) \\right] \\right\\| _ {2}, \\tag {2}", + "image_path": "ced7ada0a689f6868e8f834cca1e53ff28192991dcbdd71cee2237be4b4d7f3f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 509, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 570 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 509, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\pmb{p}_j" + }, + { + "bbox": [ + 130, + 509, + 482, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 509, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{j,l}" + }, + { + "bbox": [ + 130, + 509, + 482, + 570 + ], + "type": "text", + "content": " denote the 3D coordinates of the " + }, + { + "bbox": [ + 130, + 509, + 482, + 570 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 509, + 482, + 570 + ], + "type": "text", + "content": "-th point and its " + }, + { + "bbox": [ + 130, + 509, + 482, + 570 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 509, + 482, + 570 + ], + "type": "text", + "content": "-th nearest neighbor within " + }, + { + "bbox": [ + 130, + 509, + 482, + 570 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{RoI}}" + }, + { + "bbox": [ + 130, + 509, + 482, + 570 + ], + "type": "text", + "content": ", respectively; " + }, + { + "bbox": [ + 130, + 509, + 482, + 570 + ], + "type": "inline_equation", + "content": "r_j" + }, + { + "bbox": [ + 130, + 509, + 482, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 509, + 482, + 570 + ], + "type": "inline_equation", + "content": "r_{j,l}" + }, + { + "bbox": [ + 130, + 509, + 482, + 570 + ], + "type": "text", + "content": " represent the reflectivity values of " + }, + { + "bbox": [ + 130, + 509, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\pmb{p}_j" + }, + { + "bbox": [ + 130, + 509, + 482, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 509, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{j,l}" + }, + { + "bbox": [ + 130, + 509, + 482, + 570 + ], + "type": "text", + "content": ", correspondingly; " + }, + { + "bbox": [ + 130, + 509, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _2" + }, + { + "bbox": [ + 130, + 509, + 482, + 570 + ], + "type": "text", + "content": " is the Euclidean norm. " + }, + { + "bbox": [ + 130, + 509, + 482, + 570 + ], + "type": "inline_equation", + "content": "g:\\mathbb{R}\\to [D_{\\min},D_{\\max}]" + }, + { + "bbox": [ + 130, + 509, + 482, + 570 + ], + "type": "text", + "content": " is the reflectivity mapping function that maps the numerical range of reflectivity onto a consistent scale with the range of Euclidean distances between points:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 195, + 575, + 481, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 575, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 195, + 575, + 481, + 601 + ], + "type": "interline_equation", + "content": "g (r) = \\left(\\frac {r - r _ {\\operatorname* {m i n}}}{r _ {\\operatorname* {m a x}} - r _ {\\operatorname* {m i n}}}\\right) \\left(D _ {\\operatorname* {m a x}} - D _ {\\operatorname* {m i n}}\\right) + D _ {\\operatorname* {m i n}}, \\tag {3}", + "image_path": "c08d39202143707bc0844a2932d1340ff86e526cdd46cf543a3c5c5e8eda42e9.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 192, + 604, + 481, + 621 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 604, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 192, + 604, + 481, + 621 + ], + "type": "interline_equation", + "content": "D _ {\\min } = \\min _ {j, l} \\| \\boldsymbol {p} _ {j} - \\boldsymbol {p} _ {j, l} \\| _ {2}, \\quad D _ {\\max } = \\max _ {j, l} \\| \\boldsymbol {p} _ {j} - \\boldsymbol {p} _ {j, l} \\| _ {2}. \\tag {4}", + "image_path": "427882c67b2033864493ac19070c23d62e78f325d3be1e424c1d1e8cb03822db.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 628, + 481, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 628, + 481, + 652 + ], + "spans": [ + { + "bbox": [ + 130, + 628, + 481, + 652 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 628, + 481, + 652 + ], + "type": "inline_equation", + "content": "[D_{\\min}, D_{\\max}]" + }, + { + "bbox": [ + 130, + 628, + 481, + 652 + ], + "type": "text", + "content": " is the range of the Euclidean norms of coordinate differences for all considered point pairs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 652, + 481, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 652, + 481, + 688 + ], + "spans": [ + { + "bbox": [ + 130, + 652, + 481, + 688 + ], + "type": "text", + "content": "The 4D distance in Eq. (2) integrates material reflectivity [30] into RAPiD features, enhancing feature representation and aiding in the discrimination of various materials and surfaces, which is crucial for accurate semantic segmentation." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "type": "text", + "content": "RAPiD-Seg Networks for 3D LiDAR Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "For convenience to facilitate feature normalization and data alignment, we arrange RAPiD lexicographically by sorting Eq. (1), where " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\mathrm{sort}(\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " on the inner and outer brackets sorts the elements " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\rho_{j,l}" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " within each row " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": ", and the sorted rows based on their first differing elements, both in ascending order [52]. It treats RAPiD matrices with the same geometric structure but different orders as identical." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 480, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 480, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 480, + 212 + ], + "type": "text", + "content": "To improve the data robustness, we normalize " + }, + { + "bbox": [ + 130, + 176, + 480, + 212 + ], + "type": "inline_equation", + "content": "\\rho_{j,l}" + }, + { + "bbox": [ + 130, + 176, + 480, + 212 + ], + "type": "text", + "content": " after excluding outliers exceeding a distance threshold " + }, + { + "bbox": [ + 130, + 176, + 480, + 212 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 130, + 176, + 480, + 212 + ], + "type": "text", + "content": ", substituting these entries with the maximum of the normalized distribution to represent significant inter-point distances." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 219, + 332, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 219, + 332, + 230 + ], + "spans": [ + { + "bbox": [ + 131, + 219, + 332, + 230 + ], + "type": "text", + "content": "3.2 Intra-Ring and Intra-Class RAPiD" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 239, + 480, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 480, + 346 + ], + "type": "text", + "content": "To enhance the RAPiD ability to capture local context, we propose R-RAPiD and C-RAPiD, based on the inherent structure of LiDAR data. The key benefit of C-RAPiD is its capacity to underscore the inherent traits of points within the same semantic class, achieved based on semantic labels before distance calculation. As a complementary, R-RAPiD is versatile, regardless of semantic label availability. Intra-Ring RAPiD (R-RAPiD) is a specialized variant of RAPiD, which confines the RoI to the ring encompassing the anchor points, thereby serving to economize on computational overhead while capitalizing on the inherent structural characteristics of LiDAR data (Fig. 2 ①)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 347, + 484, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 484, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 484, + 419 + ], + "type": "text", + "content": "Given LiDAR sensors with fixed beam counts [2,3,29,48], e.g., " + }, + { + "bbox": [ + 130, + 347, + 484, + 419 + ], + "type": "inline_equation", + "content": "B \\in \\{32,64,128\\}" + }, + { + "bbox": [ + 130, + 347, + 484, + 419 + ], + "type": "text", + "content": ", beams radiate isotropically around the vehicle at set angles. We segment the " + }, + { + "bbox": [ + 130, + 347, + 484, + 419 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 347, + 484, + 419 + ], + "type": "text", + "content": "-point LiDAR point cloud " + }, + { + "bbox": [ + 130, + 347, + 484, + 419 + ], + "type": "inline_equation", + "content": "C_{B,s}" + }, + { + "bbox": [ + 130, + 347, + 484, + 419 + ], + "type": "text", + "content": " by beams, where " + }, + { + "bbox": [ + 130, + 347, + 484, + 419 + ], + "type": "inline_equation", + "content": "m = s \\times B" + }, + { + "bbox": [ + 130, + 347, + 484, + 419 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 347, + 484, + 419 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 347, + 484, + 419 + ], + "type": "text", + "content": " representing measurements within a scan cycle and " + }, + { + "bbox": [ + 130, + 347, + 484, + 419 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 347, + 484, + 419 + ], + "type": "text", + "content": " the number of laser beams. Suppose " + }, + { + "bbox": [ + 130, + 347, + 484, + 419 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i" + }, + { + "bbox": [ + 130, + 347, + 484, + 419 + ], + "type": "text", + "content": " with Cartesian coordinate " + }, + { + "bbox": [ + 130, + 347, + 484, + 419 + ], + "type": "inline_equation", + "content": "(x_i, y_i, z_i)" + }, + { + "bbox": [ + 130, + 347, + 484, + 419 + ], + "type": "text", + "content": " is a point within " + }, + { + "bbox": [ + 130, + 347, + 484, + 419 + ], + "type": "inline_equation", + "content": "C_{B,s}" + }, + { + "bbox": [ + 130, + 347, + 484, + 419 + ], + "type": "text", + "content": ", its transformation to cylindrical coordinates " + }, + { + "bbox": [ + 130, + 347, + 484, + 419 + ], + "type": "inline_equation", + "content": "(\\theta_i, \\phi_i)" + }, + { + "bbox": [ + 130, + 347, + 484, + 419 + ], + "type": "text", + "content": " is formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 215, + 430, + 481, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 430, + 481, + 442 + ], + "spans": [ + { + "bbox": [ + 215, + 430, + 481, + 442 + ], + "type": "interline_equation", + "content": "\\theta_ {i} = \\left\\lfloor \\arctan 2 \\left(y _ {i}, x _ {i}\\right) / \\Delta \\theta \\right\\rfloor , \\tag {5}", + "image_path": "12394200b51011cc692ddf5f55bbdcb65e7903938ad67f70361315651d194802.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 214, + 445, + 481, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 445, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 214, + 445, + 481, + 464 + ], + "type": "interline_equation", + "content": "\\phi_ {i} = \\left\\lfloor \\arcsin \\left[ z _ {i} \\left(x _ {i} ^ {2} + y _ {i} ^ {2} + z _ {i} ^ {2}\\right) ^ {- 1 / 2} \\right] / \\Delta \\phi \\right\\rfloor , \\tag {6}", + "image_path": "ee1870f5327dd933efa004b82f280911b3befe8b7b0814e6a0558846666e8cd2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 473, + 480, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 480, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 480, + 496 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 473, + 480, + 496 + ], + "type": "inline_equation", + "content": "\\varDelta\\theta" + }, + { + "bbox": [ + 130, + 473, + 480, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 473, + 480, + 496 + ], + "type": "inline_equation", + "content": "\\varDelta\\phi" + }, + { + "bbox": [ + 130, + 473, + 480, + 496 + ], + "type": "text", + "content": " denote the mean angular resolutions horizontally and vertically between adjacent beams, and " + }, + { + "bbox": [ + 130, + 473, + 480, + 496 + ], + "type": "inline_equation", + "content": "\\arctan 2(\\cdot)" + }, + { + "bbox": [ + 130, + 473, + 480, + 496 + ], + "type": "text", + "content": " is the 2-argument " + }, + { + "bbox": [ + 130, + 473, + 480, + 496 + ], + "type": "inline_equation", + "content": "\\arctan (\\cdot)" + }, + { + "bbox": [ + 130, + 473, + 480, + 496 + ], + "type": "text", + "content": " function." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 496, + 480, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 480, + 533 + ], + "type": "inline_equation", + "content": "C_{B,s}" + }, + { + "bbox": [ + 130, + 496, + 480, + 533 + ], + "type": "text", + "content": " is then partitioned to " + }, + { + "bbox": [ + 130, + 496, + 480, + 533 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 496, + 480, + 533 + ], + "type": "text", + "content": " rings based on " + }, + { + "bbox": [ + 130, + 496, + 480, + 533 + ], + "type": "inline_equation", + "content": "\\phi_b" + }, + { + "bbox": [ + 130, + 496, + 480, + 533 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 496, + 480, + 533 + ], + "type": "inline_equation", + "content": "C_{B,s} = \\bigcup_{b=0}^{B-1} R_b" + }, + { + "bbox": [ + 130, + 496, + 480, + 533 + ], + "type": "text", + "content": ", where each ring " + }, + { + "bbox": [ + 130, + 496, + 480, + 533 + ], + "type": "inline_equation", + "content": "R_b = \\{(\\theta_{b,i}, \\phi_{b,i}) \\mid \\phi_{b,i} = \\phi_b, \\forall i \\leq m\\}" + }, + { + "bbox": [ + 130, + 496, + 480, + 533 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 130, + 496, + 480, + 533 + ], + "type": "inline_equation", + "content": "\\bigcup" + }, + { + "bbox": [ + 130, + 496, + 480, + 533 + ], + "type": "text", + "content": " signifies the cumulative union R-RAPiD is thus defined as the cumulative concatenate " + }, + { + "bbox": [ + 130, + 496, + 480, + 533 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 130, + 496, + 480, + 533 + ], + "type": "text", + "content": " of RAPiD in each ring " + }, + { + "bbox": [ + 130, + 496, + 480, + 533 + ], + "type": "inline_equation", + "content": "R_b" + }, + { + "bbox": [ + 130, + 496, + 480, + 533 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 208, + 539, + 481, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 539, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 208, + 539, + 481, + 555 + ], + "type": "interline_equation", + "content": "\\text {R - R A P i D} \\left(C _ {B, s}; k\\right) := \\bigoplus_ {b = 0} ^ {B - 1} \\text {R A P i D} \\left(R _ {b}; k\\right). \\tag {7}", + "image_path": "992cae46f73877c85e6a5b3b348f3bde1a39c59a5d20f8a0fe45364f5ac239ac.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 561, + 479, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 561, + 479, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 561, + 479, + 596 + ], + "type": "text", + "content": "Intra-Class RAPiD (C-RAPiD) concentrates on extracting point features within the confines of each semantic class (Fig. 2 ②). It is complementary to R-RAPiD, preserving the embedding fidelity within individual semantic class." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 597, + 480, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 597, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 130, + 597, + 480, + 620 + ], + "type": "text", + "content": "Specifically, C-RAPiD concatenates RAPiD for each point " + }, + { + "bbox": [ + 130, + 597, + 480, + 620 + ], + "type": "inline_equation", + "content": "\\pmb{p}_j" + }, + { + "bbox": [ + 130, + 597, + 480, + 620 + ], + "type": "text", + "content": " and its corresponding semantic label " + }, + { + "bbox": [ + 130, + 597, + 480, + 620 + ], + "type": "inline_equation", + "content": "y_{j}" + }, + { + "bbox": [ + 130, + 597, + 480, + 620 + ], + "type": "text", + "content": " within semantic class " + }, + { + "bbox": [ + 130, + 597, + 480, + 620 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 597, + 480, + 620 + ], + "type": "text", + "content": ", across " + }, + { + "bbox": [ + 130, + 597, + 480, + 620 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 130, + 597, + 480, + 620 + ], + "type": "text", + "content": " total classes:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 197, + 628, + 481, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 628, + 481, + 644 + ], + "spans": [ + { + "bbox": [ + 197, + 628, + 481, + 644 + ], + "type": "interline_equation", + "content": "\\text {C - R A P i D} \\left(\\left\\{C _ {B, s}, \\mathcal {Y} \\right\\}; k\\right) := \\bigoplus_ {i = 0} ^ {N _ {c} - 1} \\operatorname {R A P i D} \\left(S _ {i}; k\\right), \\tag {8}", + "image_path": "725957b85f7c02df1b6dc4edaca2e582f9fac4f52c14ab8cf6e91497b357f3cf.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 654, + 349, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 654, + 349, + 668 + ], + "spans": [ + { + "bbox": [ + 134, + 654, + 349, + 668 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 134, + 654, + 349, + 668 + ], + "type": "inline_equation", + "content": "S_{i} = \\{\\pmb{p}_{j}\\mid y_{j} = i,\\forall j\\leq m\\}" + }, + { + "bbox": [ + 134, + 654, + 349, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = [y_j]_{j = 1}^m" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "text", + "content": "L. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 266, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 266, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 266, + 129 + ], + "type": "text", + "content": "4 RAPiD Embedding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 479, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 479, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 479, + 176 + ], + "type": "text", + "content": "The direct utilization of high-dimensional RAPiD features imposes a substantial computational burden. Efficient processing of large data volumes requires methods that can condense dimensionality while preserving data integrity [62]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 177, + 481, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 177, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 177, + 481, + 236 + ], + "type": "text", + "content": "We explore data embedding methods to reduce computational costs while providing feature learning capacity. AE is efficient in embedding high-dimensional features utilizing the reconstruction loss [16, 42]. However, conventional AE, limited by their sensitivity to input order and reliance on fixed-size misinputs [24], struggle to align with the unordered and variable size inherent in point clouds [15]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 237, + 481, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 481, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 481, + 284 + ], + "type": "text", + "content": "We propose a double nested AE structure with a novel class-aware embedding objective, using the Voxel-based Set Attention (VSA) module [15] as a building block. This facilitates superior contextual awareness and adaptability to unordered and variably-sized point cloud data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 285, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 415 + ], + "type": "text", + "content": "Nested RAPiD AE, illustrated in Fig. 2 ③ and Fig. 3, is composed of two modules: the outer module is constituted by a VSA AE, primarily responsible for the point-to-voxel conversion. Within the inner module, which focuses on voxel-wise representation, we employ an additional AE specifically designed for dimensionality reduction. Specifically, considering an input of " + }, + { + "bbox": [ + 130, + 285, + 482, + 415 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 285, + 482, + 415 + ], + "type": "text", + "content": "-point RAPiD features " + }, + { + "bbox": [ + 130, + 285, + 482, + 415 + ], + "type": "inline_equation", + "content": "G \\in \\mathbb{R}^{m \\times d}" + }, + { + "bbox": [ + 130, + 285, + 482, + 415 + ], + "type": "text", + "content": ", where each point encompasses " + }, + { + "bbox": [ + 130, + 285, + 482, + 415 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 285, + 482, + 415 + ], + "type": "text", + "content": " features. We compress pointwise " + }, + { + "bbox": [ + 130, + 285, + 482, + 415 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 285, + 482, + 415 + ], + "type": "text", + "content": " into a voxel-wise representation " + }, + { + "bbox": [ + 130, + 285, + 482, + 415 + ], + "type": "inline_equation", + "content": "H^v \\in \\mathbb{R}^{c \\times l \\times d}" + }, + { + "bbox": [ + 130, + 285, + 482, + 415 + ], + "type": "text", + "content": " by an outer VSA Voxel Encoder (with " + }, + { + "bbox": [ + 130, + 285, + 482, + 415 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 285, + 482, + 415 + ], + "type": "text", + "content": " reduced voxel-wise dimension and " + }, + { + "bbox": [ + 130, + 285, + 482, + 415 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 285, + 482, + 415 + ], + "type": "text", + "content": " latent codes), then further reduces it to a compressed embedding " + }, + { + "bbox": [ + 130, + 285, + 482, + 415 + ], + "type": "inline_equation", + "content": "\\hbar \\in \\mathbb{R}^{c \\times l \\times d'}" + }, + { + "bbox": [ + 130, + 285, + 482, + 415 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 285, + 482, + 415 + ], + "type": "inline_equation", + "content": "d'" + }, + { + "bbox": [ + 130, + 285, + 482, + 415 + ], + "type": "text", + "content": " features via an inner Encoder. The inner and outer VSA Point Decoders then reconstruct the output feature set " + }, + { + "bbox": [ + 130, + 285, + 482, + 415 + ], + "type": "inline_equation", + "content": "\\hat{G}" + }, + { + "bbox": [ + 130, + 285, + 482, + 415 + ], + "type": "text", + "content": " akin to " + }, + { + "bbox": [ + 130, + 285, + 482, + 415 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 285, + 482, + 415 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 417, + 482, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 417, + 482, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 417, + 482, + 465 + ], + "type": "text", + "content": "Inner Voxel-Wise Representation is conducted via an inner AE, which takes " + }, + { + "bbox": [ + 130, + 417, + 482, + 465 + ], + "type": "inline_equation", + "content": "H^v" + }, + { + "bbox": [ + 130, + 417, + 482, + 465 + ], + "type": "text", + "content": " from outer VSA AE as the input, yielding a lower-dimensional embedding " + }, + { + "bbox": [ + 130, + 417, + 482, + 465 + ], + "type": "inline_equation", + "content": "\\hbar" + }, + { + "bbox": [ + 130, + 417, + 482, + 465 + ], + "type": "text", + "content": " (dimension reduced from " + }, + { + "bbox": [ + 130, + 417, + 482, + 465 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 417, + 482, + 465 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 417, + 482, + 465 + ], + "type": "inline_equation", + "content": "d'" + }, + { + "bbox": [ + 130, + 417, + 482, + 465 + ], + "type": "text", + "content": "). This process involves " + }, + { + "bbox": [ + 130, + 417, + 482, + 465 + ], + "type": "inline_equation", + "content": "u \\times" + }, + { + "bbox": [ + 130, + 417, + 482, + 465 + ], + "type": "text", + "content": " convolutional layers and a batch normalization layer [19]:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 233, + 475, + 481, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 475, + 481, + 488 + ], + "spans": [ + { + "bbox": [ + 233, + 475, + 481, + 488 + ], + "type": "interline_equation", + "content": "\\hbar^ {(i)} = \\operatorname {B a t c h N o r m} \\circ \\operatorname {C o n v} \\left(\\hbar^ {(i - 1)}\\right), \\tag {9}", + "image_path": "ac1a2e321de01a5e95a88ddbc7a1afc22a4dfd2920b5ad84dfa3009f14495b36.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 498, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 559 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 498, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\hbar^{(0)} = H^v" + }, + { + "bbox": [ + 130, + 498, + 482, + 559 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 498, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\hbar^{(u - 1)} = \\hbar" + }, + { + "bbox": [ + 130, + 498, + 482, + 559 + ], + "type": "text", + "content": ". The Convolutional Feed-Forward Network (ConvFFN) then promotes voxel-level information exchange, enhancing spatial feature interactivity. It maps reduced hidden features " + }, + { + "bbox": [ + 130, + 498, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\hbar" + }, + { + "bbox": [ + 130, + 498, + 482, + 559 + ], + "type": "text", + "content": " to a 3D sparse tensor, indexed by voxel coordinates " + }, + { + "bbox": [ + 130, + 498, + 482, + 559 + ], + "type": "inline_equation", + "content": "X^v" + }, + { + "bbox": [ + 130, + 498, + 482, + 559 + ], + "type": "text", + "content": ", and employs dual depth-wise convolutions (DwConv) for spatial interactivity:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 205, + 568, + 481, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 568, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 205, + 568, + 481, + 582 + ], + "type": "interline_equation", + "content": "\\hat {h} = \\operatorname {D w C o n v} ^ {(2)} \\circ \\zeta \\circ \\operatorname {D w C o n v} ^ {(1)} \\left(\\operatorname {S p T} \\left(\\hbar , X ^ {v}\\right)\\right), \\tag {10}", + "image_path": "b60d160f1e6fa3b1d420f85993926d2722924d26ad0053a47f5aeed7f9fb4099.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 592, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 592, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 592, + 482, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 592, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 130, + 592, + 482, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 592, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{SpT}(\\cdot)" + }, + { + "bbox": [ + 130, + 592, + 482, + 628 + ], + "type": "text", + "content": " represents the non-linear activation and sparse tensor construction. Subsequently, " + }, + { + "bbox": [ + 130, + 592, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\hat{\\hbar}\\in \\mathbb{R}^{c\\times l\\times d'}" + }, + { + "bbox": [ + 130, + 592, + 482, + 628 + ], + "type": "text", + "content": " is reconstructed into " + }, + { + "bbox": [ + 130, + 592, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\hat{H}^v\\in \\mathbb{R}^{c\\times l\\times d}" + }, + { + "bbox": [ + 130, + 592, + 482, + 628 + ], + "type": "text", + "content": " through an inner Decoder, consisting of multiple DeConv Layers." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Outer VSA AE consists of a voxel encoder and point encoder. The voxel encoder project RAPiD features into key-value spaces, forming " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": ", followed by a cross-attention mechanism with a latent query " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": ", producing an attention" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "type": "text", + "content": "RAPiD-Seg Networks for 3D LiDAR Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 177, + 115, + 438, + 220 + ], + "blocks": [ + { + "bbox": [ + 177, + 115, + 438, + 220 + ], + "lines": [ + { + "bbox": [ + 177, + 115, + 438, + 220 + ], + "spans": [ + { + "bbox": [ + 177, + 115, + 438, + 220 + ], + "type": "image", + "image_path": "bd43c21d0b93992e957c04b3a21a936e2d971f12f33688afff80e8c5bb26aac8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 222, + 482, + 257 + ], + "lines": [ + { + "bbox": [ + 131, + 222, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 131, + 222, + 482, + 257 + ], + "type": "text", + "content": "Fig. 3: Our RAPiD AE consists of an Encoder, Convolution Layer, and Decoder module, aiming to reproduce the input features and generate the compressed voxel-wise RAPiD representation " + }, + { + "bbox": [ + 131, + 222, + 482, + 257 + ], + "type": "inline_equation", + "content": "\\hbar" + }, + { + "bbox": [ + 131, + 222, + 482, + 257 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 264, + 482, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 264, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 131, + 264, + 482, + 289 + ], + "type": "text", + "content": "matrix " + }, + { + "bbox": [ + 131, + 264, + 482, + 289 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 131, + 264, + 482, + 289 + ], + "type": "text", + "content": ". The voxel-wise representation " + }, + { + "bbox": [ + 131, + 264, + 482, + 289 + ], + "type": "inline_equation", + "content": "H^v" + }, + { + "bbox": [ + 131, + 264, + 482, + 289 + ], + "type": "text", + "content": " is the scatter sum [43] of pointwise " + }, + { + "bbox": [ + 131, + 264, + 482, + 289 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 131, + 264, + 482, + 289 + ], + "type": "text", + "content": " to aggregate the value vectors " + }, + { + "bbox": [ + 131, + 264, + 482, + 289 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 131, + 264, + 482, + 289 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 192, + 297, + 482, + 311 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 297, + 482, + 311 + ], + "spans": [ + { + "bbox": [ + 192, + 297, + 482, + 311 + ], + "type": "interline_equation", + "content": "H ^ {v} = \\operatorname {S u m} _ {\\text {s c a t t e r}} (H, I ^ {v}), \\quad H = \\tilde {A} ^ {\\top} V, \\tag {11}", + "image_path": "0a3c414bacc05919032f17b1fbad9d9f563136a82c001759345b6cc403ab922f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 198, + 313, + 482, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 313, + 482, + 328 + ], + "spans": [ + { + "bbox": [ + 198, + 313, + 482, + 328 + ], + "type": "interline_equation", + "content": "\\tilde {A} = \\operatorname {S o f t m a x} _ {\\text {s c a t t e r}} \\left(K L ^ {\\top}, I ^ {v}\\right), \\quad (Q, V) = \\operatorname {P r o j} (G), \\tag {12}", + "image_path": "fd801258f89d275cab6f0f01f3a2f46a3e1d83b4694faf2f84137b2296eabac7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 335, + 483, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 335, + 483, + 384 + ], + "spans": [ + { + "bbox": [ + 131, + 335, + 483, + 384 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 335, + 483, + 384 + ], + "type": "inline_equation", + "content": "\\operatorname{Proj}(\\cdot)" + }, + { + "bbox": [ + 131, + 335, + 483, + 384 + ], + "type": "text", + "content": " represents the linear projection, and " + }, + { + "bbox": [ + 131, + 335, + 483, + 384 + ], + "type": "inline_equation", + "content": "I^v" + }, + { + "bbox": [ + 131, + 335, + 483, + 384 + ], + "type": "text", + "content": " is the voxel indices. The Point Decoder reconstructs the output set from the enriched hidden features " + }, + { + "bbox": [ + 131, + 335, + 483, + 384 + ], + "type": "inline_equation", + "content": "\\hat{H}^v" + }, + { + "bbox": [ + 131, + 335, + 483, + 384 + ], + "type": "text", + "content": ". We start by broadcasting " + }, + { + "bbox": [ + 131, + 335, + 483, + 384 + ], + "type": "inline_equation", + "content": "\\hat{H}^v" + }, + { + "bbox": [ + 131, + 335, + 483, + 384 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 131, + 335, + 483, + 384 + ], + "type": "inline_equation", + "content": "I^v" + }, + { + "bbox": [ + 131, + 335, + 483, + 384 + ], + "type": "text", + "content": ", resulting in " + }, + { + "bbox": [ + 131, + 335, + 483, + 384 + ], + "type": "inline_equation", + "content": "\\hat{H} \\in \\mathbb{R}^{m \\times l \\times d}" + }, + { + "bbox": [ + 131, + 335, + 483, + 384 + ], + "type": "text", + "content": ". The output " + }, + { + "bbox": [ + 131, + 335, + 483, + 384 + ], + "type": "inline_equation", + "content": "\\hat{G}" + }, + { + "bbox": [ + 131, + 335, + 483, + 384 + ], + "type": "text", + "content": " is analogous to the operational paradigm of Voxel Encoder:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 217, + 392, + 482, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 392, + 482, + 409 + ], + "spans": [ + { + "bbox": [ + 217, + 392, + 482, + 409 + ], + "type": "interline_equation", + "content": "\\hat {G} = \\tilde {A} ^ {*} ^ {\\top} V ^ {*}, \\quad \\tilde {A} ^ {*} = \\operatorname {S o f t m a x} \\left(A ^ {*}\\right), \\tag {13}", + "image_path": "301c67f7735d734675c80bcb477018a818540893c010725a201b95c8568900a2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 214, + 411, + 482, + 428 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 411, + 482, + 428 + ], + "spans": [ + { + "bbox": [ + 214, + 411, + 482, + 428 + ], + "type": "interline_equation", + "content": "A ^ {*} = \\left[ K _ {i} ^ {*} Q _ {i} ^ {\\top} \\right] _ {i = 1} ^ {m}, \\quad \\left(K ^ {*}, V ^ {*}\\right) = \\operatorname {P r o j} (\\hat {\\mathcal {H}}). \\tag {14}", + "image_path": "42c602498725c5f8f0b86f75870ef5a654f8b0084338eca885f461b5bd38324a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 434, + 482, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 434, + 482, + 494 + ], + "spans": [ + { + "bbox": [ + 131, + 434, + 482, + 494 + ], + "type": "text", + "content": "Class-Aware Embedding Objective addresses the issue of non-uniqueness in embeddings produced by the AE, where various distinct inputs yield approximately the same embedding, leading to inaccurate representation [1, 25]. Our objective aims to facilitate the generation of AE embeddings that demonstrate robust semantic class discriminability." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 494, + 482, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 494, + 482, + 527 + ], + "spans": [ + { + "bbox": [ + 131, + 494, + 482, + 527 + ], + "type": "text", + "content": "Specifically, we introduce a novel class-aware contrastive loss " + }, + { + "bbox": [ + 131, + 494, + 482, + 527 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{contr}}" + }, + { + "bbox": [ + 131, + 494, + 482, + 527 + ], + "type": "text", + "content": " in Eq. (15), which aims to maximize the distance between embeddings of different semantic classes while minimizing the distance of the same class." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 196, + 531, + 481, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 531, + 481, + 602 + ], + "spans": [ + { + "bbox": [ + 196, + 531, + 481, + 602 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {c o n t r}} = \\frac {1}{m} \\sum_ {i = 1} ^ {m} \\left[ \\frac {1}{| P (i) |} \\sum_ {p \\in P (i)} \\mathrm {R e L U} (\\alpha - \\mathrm {s i m} (H _ {i}, H _ {p})) \\right. \\\\ \\left. + \\frac {1}{| N (i) |} \\sum_ {n \\in N (i)} \\operatorname {R e L U} \\left(\\sin \\left(H _ {i}, H _ {n}\\right) - \\alpha\\right) \\right], \\tag {15} \\\\ \\end{array}", + "image_path": "3519bef1080f67d9a1052ac3d5657665df2c04e37dc08d4b96dfac3e03073eb3.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 196, + 605, + 481, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 605, + 481, + 634 + ], + "spans": [ + { + "bbox": [ + 196, + 605, + 481, + 634 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c o n}} = \\frac {1}{m \\cdot d} \\sum_ {i = 1} ^ {m} \\sum_ {j = 1} ^ {d} \\left(G _ {i, j} - \\hat {G} _ {i, j}\\right) ^ {2}, \\tag {16}", + "image_path": "a4f6e9ac683ec7e13a2c00ba76a1c2921e8812ee162523c6a9720e20f29fd1d2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": " represents the point index; " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "P(i)" + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "N(i)" + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": " are the indices of the nearest point in the same (Positive) and different class (Negative) relative to point " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "L. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": "respectively; " + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "inline_equation", + "content": "\\mathrm{sim}(\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": " is a similarity measure; " + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": " is a margin controlling separation between classes. The pointwise representations " + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "inline_equation", + "content": "H_{i}" + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "inline_equation", + "content": "H_{p}" + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "inline_equation", + "content": "H_{n}" + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": " are broadcasted from " + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "inline_equation", + "content": "\\hbar" + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": " based on voxel indices " + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "inline_equation", + "content": "I^{v}" + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 189 + ], + "type": "text", + "content": "We further combine this with the MSE [39] reconstruction loss in Eq. (16). The overall loss function, with " + }, + { + "bbox": [ + 130, + 152, + 480, + 189 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 152, + 480, + 189 + ], + "type": "text", + "content": " balancing the fidelity of reconstruction with the distinctiveness of the embeddings, is formulated as: " + }, + { + "bbox": [ + 130, + 152, + 480, + 189 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{total}} = \\mathcal{L}_{\\mathrm{recon}} + \\lambda \\mathcal{L}_{\\mathrm{contr}}" + }, + { + "bbox": [ + 130, + 152, + 480, + 189 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 200, + 394, + 214 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 200, + 394, + 214 + ], + "spans": [ + { + "bbox": [ + 131, + 200, + 394, + 214 + ], + "type": "text", + "content": "5 RAPiD-Seg for 3D LiDAR Segmentation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 224, + 480, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 480, + 320 + ], + "type": "text", + "content": "We present RAPiD-Seg, a 3D LiDAR segmentation network leveraging RAPiD features. It incorporates multiple complementary features via the channel-wise fusion mechanism [18]. RAPiD-Seg takes point cloud as input and performs single-modal (LiDAR-only) 3D semantic segmentation. Specifically, the input point cloud has three types of point-wise features: (a) coordinates-based features " + }, + { + "bbox": [ + 130, + 224, + 480, + 320 + ], + "type": "inline_equation", + "content": "F_{C} = \\{p_{C} \\mid p_{C} = (x,y,z)\\}" + }, + { + "bbox": [ + 130, + 224, + 480, + 320 + ], + "type": "text", + "content": "; (b) intensity-based features " + }, + { + "bbox": [ + 130, + 224, + 480, + 320 + ], + "type": "inline_equation", + "content": "F_{I} = \\{p_{I} \\mid p_{I} = (I,J)\\}" + }, + { + "bbox": [ + 130, + 224, + 480, + 320 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 224, + 480, + 320 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 224, + 480, + 320 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 224, + 480, + 320 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 130, + 224, + 480, + 320 + ], + "type": "text", + "content": " are intensity and reflectivity; (c) RAPiD features " + }, + { + "bbox": [ + 130, + 224, + 480, + 320 + ], + "type": "inline_equation", + "content": "F_{R} = \\{p_{R} \\mid p_{R} = \\mathrm{RAPiD}(P_{\\mathrm{RoI}};k)\\}" + }, + { + "bbox": [ + 130, + 224, + 480, + 320 + ], + "type": "text", + "content": ". The RoI varies depending on whether it is R- or C-RAPiD." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 320, + 480, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 480, + 403 + ], + "type": "text", + "content": "As shown in Fig. 2, " + }, + { + "bbox": [ + 130, + 320, + 480, + 403 + ], + "type": "inline_equation", + "content": "F_{C} \\oplus F_{I}" + }, + { + "bbox": [ + 130, + 320, + 480, + 403 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 320, + 480, + 403 + ], + "type": "inline_equation", + "content": "F_{R}" + }, + { + "bbox": [ + 130, + 320, + 480, + 403 + ], + "type": "text", + "content": " are fed to a VSA voxel encoder 4 and RAPiD AE 6 for voxelization, resulting in voxel-wise representation " + }, + { + "bbox": [ + 130, + 320, + 480, + 403 + ], + "type": "inline_equation", + "content": "E_{C}" + }, + { + "bbox": [ + 130, + 320, + 480, + 403 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 320, + 480, + 403 + ], + "type": "inline_equation", + "content": "E_{I}" + }, + { + "bbox": [ + 130, + 320, + 480, + 403 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 320, + 480, + 403 + ], + "type": "inline_equation", + "content": "E_{R}" + }, + { + "bbox": [ + 130, + 320, + 480, + 403 + ], + "type": "text", + "content": ". Subsequently, we incorporate the complementary voxel-wise features via the channel-wise fusion mechanism (Sec. 5.1). The backbone net takes in them for LiDAR segmentation prediction. Additionally, we introduce two effective RAPiD-Seg variants to expedite the convergence of the extensive segmentation network towards an optimal solution." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 420, + 383, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 420, + 383, + 431 + ], + "spans": [ + { + "bbox": [ + 131, + 420, + 383, + 431 + ], + "type": "text", + "content": "5.1 RAPiD Channel-Wise Fusion with Attention" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 437, + 480, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 480, + 497 + ], + "type": "text", + "content": "To combine various LiDAR point attributes, a highly efficient method for feature fusion is still a research topic [30, 50]. Current methods, which typically concatenate different LiDAR attributes (e.g., intensity [38, 41, 59], reflectivity [30], PLS [50]) face issues with increased dimensionality and complexity, and the need for balancing weights to avoid biased training towards dominant attributes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 498, + 480, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 480, + 545 + ], + "type": "text", + "content": "We thus propose a novel RAPiD Fusion with Attention module (FuAtten), pioneering application of channel-wise attention mechanism [18] in LiDAR point attribute fusion. With FuAtten, we effectively fuse " + }, + { + "bbox": [ + 130, + 498, + 480, + 545 + ], + "type": "inline_equation", + "content": "E_{C}" + }, + { + "bbox": [ + 130, + 498, + 480, + 545 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 498, + 480, + 545 + ], + "type": "inline_equation", + "content": "E_{I}" + }, + { + "bbox": [ + 130, + 498, + 480, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 498, + 480, + 545 + ], + "type": "inline_equation", + "content": "E_{R}" + }, + { + "bbox": [ + 130, + 498, + 480, + 545 + ], + "type": "text", + "content": ", emphasizing informative features adaptively while suppressing less relevant ones across different channels." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 545, + 480, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 480, + 597 + ], + "type": "text", + "content": "As shown in Fig. 2 ⑤ and Fig. 4, we initially concatenate " + }, + { + "bbox": [ + 130, + 545, + 480, + 597 + ], + "type": "inline_equation", + "content": "E_{C}" + }, + { + "bbox": [ + 130, + 545, + 480, + 597 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 545, + 480, + 597 + ], + "type": "inline_equation", + "content": "E_{I}" + }, + { + "bbox": [ + 130, + 545, + 480, + 597 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 545, + 480, + 597 + ], + "type": "inline_equation", + "content": "E_{R}" + }, + { + "bbox": [ + 130, + 545, + 480, + 597 + ], + "type": "text", + "content": " along the voxel dimension, resulting in " + }, + { + "bbox": [ + 130, + 545, + 480, + 597 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 130, + 545, + 480, + 597 + ], + "type": "text", + "content": ". In the squeeze operation, Global Average Pooling (GAP) [33] condenses each channel " + }, + { + "bbox": [ + 130, + 545, + 480, + 597 + ], + "type": "inline_equation", + "content": "f'" + }, + { + "bbox": [ + 130, + 545, + 480, + 597 + ], + "type": "text", + "content": " of the tensor into a representative channel descriptor. These descriptors form a channel-level embedding " + }, + { + "bbox": [ + 130, + 545, + 480, + 597 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = [z_{f'}]_{f' = 1}^{f^*}" + }, + { + "bbox": [ + 130, + 545, + 480, + 597 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 197, + 605, + 481, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 605, + 481, + 639 + ], + "spans": [ + { + "bbox": [ + 197, + 605, + 481, + 639 + ], + "type": "interline_equation", + "content": "\\mathbf {z} = \\left[ \\operatorname {G A P} \\left(\\mathbf {E} _ {:,:: f ^ {\\prime}}\\right) \\right] _ {f ^ {\\prime} = 1} ^ {f ^ {*}} = \\frac {1}{c \\times l} \\left[ \\sum_ {c ^ {\\prime} = 0} ^ {c - 1} \\sum_ {l ^ {\\prime} = 0} ^ {l - 1} \\mathbf {E} _ {c ^ {\\prime}, l ^ {\\prime}, f ^ {\\prime}} \\right] _ {f ^ {\\prime} = 1} ^ {f ^ {*}}, \\tag {17}", + "image_path": "79a89ee949a6e30aa7e9328eaab2adb7e14ffbc66516023f5c6dcde79f81a63d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "inline_equation", + "content": "f^{*}" + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "text", + "content": " is the dimension of the concatenated feature. A channel-wise attention " + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{a}_{\\mathbf{z}} \\in \\mathbb{R}^{1 \\times 1 \\times f^{*}}" + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "text", + "content": " is then computed through the excitation step: " + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{a}_{\\mathbf{z}} =" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "type": "text", + "content": "RAPiD-Seg Networks for 3D LiDAR Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\sigma (\\mathbf{W}_2\\delta (\\mathbf{W}_1\\mathbf{z}))" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " represent the ReLU and Sigmoid activation. With each element in " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "a_{\\mathbf{z}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " reflects the attention allocated to the corresponding feature of the voxel, the channel-wise fused features " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^{\\prime} = a_{\\mathbf{z}}\\cdot \\mathbf{E}" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 204, + 158, + 411, + 219 + ], + "blocks": [ + { + "bbox": [ + 204, + 158, + 411, + 219 + ], + "lines": [ + { + "bbox": [ + 204, + 158, + 411, + 219 + ], + "spans": [ + { + "bbox": [ + 204, + 158, + 411, + 219 + ], + "type": "image", + "image_path": "dc886fcf06578cc3e62a634b3443e1019ce6c77fee60ca35399c435489a758fb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 177, + 228, + 435, + 239 + ], + "lines": [ + { + "bbox": [ + 177, + 228, + 435, + 239 + ], + "spans": [ + { + "bbox": [ + 177, + 228, + 435, + 239 + ], + "type": "text", + "content": "Fig. 4: Our feature fusion module with channel-wise attention." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 253, + 397, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 253, + 397, + 266 + ], + "spans": [ + { + "bbox": [ + 131, + 253, + 397, + 266 + ], + "type": "text", + "content": "5.2 RAPiD-Seg Architectures for 3D Segmentation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 273, + 482, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 273, + 482, + 322 + ], + "spans": [ + { + "bbox": [ + 130, + 273, + 482, + 322 + ], + "type": "text", + "content": "The complexity inherent in 3D LiDAR-driven networks typically exacerbates the end-to-end training process [17,30,50], since the extensive parameter space increases the propensity for overfitting, slow convergence, and the possibility of settling into local minima [30]." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 160, + 332, + 455, + 404 + ], + "blocks": [ + { + "bbox": [ + 160, + 332, + 455, + 404 + ], + "lines": [ + { + "bbox": [ + 160, + 332, + 455, + 404 + ], + "spans": [ + { + "bbox": [ + 160, + 332, + 455, + 404 + ], + "type": "image", + "image_path": "b7e9f5e0bdcc3325367e44cee323ec1fd258f2afb3c56e1104922b348d5c9b2d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 411, + 482, + 436 + ], + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 436 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 436 + ], + "type": "text", + "content": "Fig. 5: Two variants of RAPiD-Seg. (a) R-RAPiD-Seg utilizes R-RAPiD features, and (b) C-RAPiD-Seg utilizes both R- and C-RAPiD features for better performance." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 449, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 498 + ], + "type": "text", + "content": "As shown in Fig. 5, we propose two novel and effective variants of RAPiD-Seg for quicker and better performance. R-RAPiD-Seg involves AE training from scratch, followed by C-RAPiD-Seg utilizing C-RAPiD features with the pre-trained AE and backbone for enhanced performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "content": "R-RAPiD-Seg. We first construct the lightweight R-RAPiD-Seg (Fig. 5 (a)) for fast 3D segmentation. R-RAPiD-Seg adopts an early fusion scheme utilizing LiDAR original points and R-RAPiD features. Specifically, 3D coordinate " + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "inline_equation", + "content": "F_{C}" + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "content": " and intensity-based features " + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "inline_equation", + "content": "F_{I}" + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "content": " are voxelized to voxel-wise representations " + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "inline_equation", + "content": "R_{v} = E_{C} \\oplus E_{I}" + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "content": " based on VSA voxelization [15]. We get the compressed " + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "inline_equation", + "content": "d'" + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "content": "-dimension voxel-wise R-RAPiD representations " + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\hbar" + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "content": " (Sec. 4) from RAPiD AE. " + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\hbar" + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "content": " is then fused into the voxel-wise representations " + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "inline_equation", + "content": "R_{v*} = \\mathrm{FuAtten}(R_v,\\hbar)" + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{FuAtten}(\\cdot)" + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "content": " is Fusion with Attention in Sec. 5.1. The fused voxel-wise features are taken into the backbone network for 3D segmentation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "C-RAPiD-Seg. To facilitate the embedding fidelity within individual semantic classes, we also design a class-aware framework with C-RAPiD features, i.e., C-RAPiD-Seg (Fig. 5 (b)). Specifically, we fuse the voxel-wise C-RAPiD " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\hbar_C" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": " into the representations " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "R_{v*} = \\mathrm{FuAtten}(R_v,\\hbar \\oplus \\hbar_C)" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": ". C-RAPiD requires class labels to compute the features regarding the semantic categories. Since the ground-truth" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "type": "text", + "content": "L. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "labels are missing during the test time, we generate the reliable pseudo labels " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{V}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": " with a pretrained R-RAPiD-Seg based on confidence (Fig. 5 (b) blue dotted arrow). Subsequently, the pseudo voxel-wise C-RAPiD " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "\\hbar_{C}" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": " are generated through " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{V}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": ". We fuse them into the voxel-wise representations for 3D segmentation with the bonebone network." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 185, + 219, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 185, + 219, + 197 + ], + "spans": [ + { + "bbox": [ + 131, + 185, + 219, + 197 + ], + "type": "text", + "content": "6 Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 209, + 481, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 209, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 209, + 481, + 246 + ], + "type": "text", + "content": "Following the popular practice of LiDAR segmentation methods [17,37,64], we evaluate our proposed RAPiD-Seg network against SoTA 3D LiDAR segmentation approaches on the SemanticKITTI [2] and nuScenes [3] datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 254, + 262, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 254, + 262, + 266 + ], + "spans": [ + { + "bbox": [ + 131, + 254, + 262, + 266 + ], + "type": "text", + "content": "6.1 Experimental Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 271, + 482, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 482, + 354 + ], + "type": "text", + "content": "Datasets: SemanticKITTI [2] comprises 22 point cloud sequences, with sequences 00-10, 08, and 11-21 for training, validation, and testing. 19 classes are chosen for training and evaluation by merging classes with similar motion statuses and discarding sparsely represented ones. Meanwhile, nuScenes [3] has 1000 driving scenes; 850 are for training and validation, with the remaining 150 for testing. 16 classes are used for LiDAR semantic segmentation, following the amalgamation of akin classes and the removal of rare ones." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 356, + 483, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 483, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 483, + 415 + ], + "type": "text", + "content": "Evaluation Protocol: Following the popular practice of [2,36,64], we adopt the Intersection-over-Union (IoU) of each class and mean IoU (mIoU) of all classes as the evaluation metric. The IoU of class " + }, + { + "bbox": [ + 130, + 356, + 483, + 415 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 356, + 483, + 415 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 130, + 356, + 483, + 415 + ], + "type": "inline_equation", + "content": "\\mathrm{IoU}_i = \\mathrm{TP}_i / (\\mathrm{TP}_i + \\mathrm{FP}_i + \\mathrm{FN}_i)" + }, + { + "bbox": [ + 130, + 356, + 483, + 415 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 356, + 483, + 415 + ], + "type": "inline_equation", + "content": "\\mathrm{TP}_i" + }, + { + "bbox": [ + 130, + 356, + 483, + 415 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 356, + 483, + 415 + ], + "type": "inline_equation", + "content": "\\mathrm{FP}_i" + }, + { + "bbox": [ + 130, + 356, + 483, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 356, + 483, + 415 + ], + "type": "inline_equation", + "content": "\\mathrm{FN}_i" + }, + { + "bbox": [ + 130, + 356, + 483, + 415 + ], + "type": "text", + "content": " denote the true positive, false positive and false negative of class " + }, + { + "bbox": [ + 130, + 356, + 483, + 415 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 356, + 483, + 415 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 416, + 483, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 416, + 483, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 416, + 483, + 548 + ], + "type": "text", + "content": "Implementation Details: We construct the point-voxel backbone based on the Minkowski-UNet34 [8] (re-implemented by PCSeg [36] codebase), which is the open-access backbone with SoTA results to date. Before the AE training stage, we first generate pointwise RAPiD features (Sec. 3), which yield three RAPiD outputs for each frame, corresponding to different " + }, + { + "bbox": [ + 130, + 416, + 483, + 548 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 416, + 483, + 548 + ], + "type": "text", + "content": " values based on range. Notably, this stage does not impose additional burdens on the subsequent overall training. The number of maximum training epochs for AE and the whole network is set as 100 and the initial learning rate is set as " + }, + { + "bbox": [ + 130, + 416, + 483, + 548 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 130, + 416, + 483, + 548 + ], + "type": "text", + "content": " with SGD optimizer. We use 2 epochs to warm up the network and adopt the cosine learning rate schedule for the remaining epochs. All experiments are conducted on " + }, + { + "bbox": [ + 130, + 416, + 483, + 548 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 130, + 416, + 483, + 548 + ], + "type": "text", + "content": " NVIDIA A100 GPUs (" + }, + { + "bbox": [ + 130, + 416, + 483, + 548 + ], + "type": "inline_equation", + "content": "1 \\times" + }, + { + "bbox": [ + 130, + 416, + 483, + 548 + ], + "type": "text", + "content": " for inference)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 562, + 270, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 562, + 270, + 573 + ], + "spans": [ + { + "bbox": [ + 131, + 562, + 270, + 573 + ], + "type": "text", + "content": "6.2 Experimental Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 581, + 481, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 481, + 629 + ], + "type": "text", + "content": "In Tab. 1 and Tab. 2, we showcase the performance of our RAPiD-Seg LiDAR segmentation method on SemanticKITTI and nuScenes test set in comparison with published leading contemporary SoTA approaches to demonstrate its superior efficacy (top 10 results shown; full rankings in Supplementary Material)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": "Our method significantly outperforms others, especially for rigid object categories (e.g., truck, o.veh, park, etc.), primarily due to our fusion of both localized 3D geometry and material reflectivity within RAPiD features, which" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "type": "text", + "content": "RAPiD-Seg Networks for 3D LiDAR Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 138, + 481, + 228 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 1: Quantitative results of RAPiD-Seg and top-10 SoTA segmentation methods on SemanticKITTI [2] test set; Best/2nd best highlighted; ▲ for multi-modal methods." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 138, + 481, + 228 + ], + "lines": [ + { + "bbox": [ + 132, + 138, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 481, + 228 + ], + "type": "table", + "html": "
MethodmIoUcarbicymotortruco.vehpedb.listm.listroadparkwalko.grobuildfencvegtrunterrpolesign
Cylinder3D [64]68.997.167.663.850.858.573.769.248.092.265.077.032.390.766.585.672.569.862.466.2
AF2S3Net [7]69.794.565.486.839.241.180.780.474.391.368.872.553.587.963.270.268.553.761.571.0
RPVNet [56]70.397.668.468.744.261.175.974.473.493.470.380.733.393.572.186.575.171.764.861.4
SDSeg3D [27]70.497.458.754.254.965.270.274.452.290.969.476.741.993.271.186.174.371.165.470.6
GASN [61]70.796.965.858.059.361.080.482.746.389.866.274.630.192.369.687.373.072.566.171.6
PVKD [17]71.297.067.969.353.560.275.173.550.591.870.977.541.092.469.486.573.871.964.965.8
2DPASS [58]72.997.063.663.461.161.577.981.374.189.767.474.740.093.572.986.273.971.065.070.4
PCSeg [36]72.997.551.267.658.668.678.380.975.692.571.578.336.993.171.485.473.669.966.168.7
RangeFormer [22]73.396.769.473.759.966.278.175.958.192.473.078.842.492.370.186.673.372.866.466.6
▲UniSeg [37]75.297.971.975.263.674.178.974.860.692.674.079.546.193.472.787.576.373.168.368.5
RAPiD-Seg (Ours)76.197.771.176.272.580.779.979.159.891.878.278.646.093.672.186.974.672.365.968.5
", + "image_path": "559b08fb2a4b23d3f6463f08d84d64310133446303cba847ff0ec81d5a90f0bb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 133, + 260, + 481, + 344 + ], + "blocks": [ + { + "bbox": [ + 132, + 236, + 481, + 258 + ], + "lines": [ + { + "bbox": [ + 132, + 236, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 481, + 258 + ], + "type": "text", + "content": "Table 2: Quantitative results of RAPiD-Seg and top-10 SoTA segmentation methods on nuScenes [3] test set; Best/2nd best highlighted; ▲ for multi-modal methods." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 260, + 481, + 344 + ], + "lines": [ + { + "bbox": [ + 133, + 260, + 481, + 344 + ], + "spans": [ + { + "bbox": [ + 133, + 260, + 481, + 344 + ], + "type": "table", + "html": "
MethodmIoUbarrbicybuscarconstmotorpedconetrailtruckdrivotherwalkterrmadeveg
▲PMF [65]77.082.040.081.088.064.079.080.076.081.067.097.068.078.074.090.088.0
Cylinder3D [64]77.282.829.884.389.463.079.377.273.484.669.197.770.280.375.590.487.6
AMVNet [34]77.380.632.081.788.967.184.376.173.584.967.397.567.479.475.591.588.7
SPVCNN [49]77.480.030.091.990.864.779.075.670.981.074.697.469.280.076.189.387.1
AF2S3Net [7]78.378.952.289.984.277.474.377.372.083.973.897.166.577.574.087.786.8
▲2D3DNet [12]80.083.059.488.085.163.784.482.076.084.871.996.967.479.876.092.189.2
GASN [61]80.485.543.290.592.164.786.083.073.383.975.897.071.081.077.791.690.2
2DPASS [58]80.881.755.392.091.873.386.578.572.584.775.597.669.179.975.590.288.0
▲LidarMultiNet [60]81.480.448.494.390.071.587.285.280.486.974.897.867.380.776.592.189.6
▲UniSeg [37]83.585.971.292.191.680.588.080.976.086.376.797.771.880.776.791.388.8
RAPiD-Seg (Ours)83.684.864.395.092.284.687.981.876.888.579.097.866.681.276.792.588.4
", + "image_path": "05a9e076bb593d34450fd5d365b3b1624a7cca05e26bee56bf23279836d69439.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 357, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 453 + ], + "type": "text", + "content": "enable segmentation based on material properties and local rigid structures. Remarkably, our single-modal methodology outperforms multi-modal approaches [22, 34,37,56,58,60], suggesting superior efficacy of our RAPiD features over alternative modalities like RGB and range images. Our inference time (105ms per frame) is comparable to other contemporary approaches [37,58,64]. Furthermore, we present supporting qualitative results in Fig. 6. Whereas the baseline method struggles with accurate vehicle type differentiation, ours achieves consistent segmentation (more visualization results in the supplementary materials)." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 137, + 460, + 298, + 544 + ], + "blocks": [ + { + "bbox": [ + 137, + 460, + 298, + 544 + ], + "lines": [ + { + "bbox": [ + 137, + 460, + 298, + 544 + ], + "spans": [ + { + "bbox": [ + 137, + 460, + 298, + 544 + ], + "type": "table", + "html": "
RAPiD FeaturesAttentionmIoUΔ
GeometricReflectivityEmbedding
70.04(Baseline)
✓70.46(+0.42)
✓71.21(+1.17)
✓✓71.93(+1.89)
✓✓✓72.15(+2.11)
✓✓✓71.80(+1.76)
✓✓✓72.32(+2.28)
✓✓✓72.78(+2.74)
✓✓✓✓73.02(+2.98)
", + "image_path": "d5e7100fe03ddaa78d8fb659f468ad926e6009087a354c606108f673fc58cdad.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 548, + 299, + 604 + ], + "lines": [ + { + "bbox": [ + 132, + 548, + 299, + 604 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 299, + 604 + ], + "type": "text", + "content": "Table 3: Component-wise ablation of RAPiD-Seg on the SemanticKITTI validation set: Geometric (Eq. (1)), Reflectivity (Eq. (3)), (RAPiD) Embedding (Sec. 4), (channel) Attention (Sec. 5.1)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 310, + 460, + 479, + 548 + ], + "blocks": [ + { + "bbox": [ + 310, + 460, + 479, + 548 + ], + "lines": [ + { + "bbox": [ + 310, + 460, + 479, + 548 + ], + "spans": [ + { + "bbox": [ + 310, + 460, + 479, + 548 + ], + "type": "image", + "image_path": "7cb6032cfe40dba5254819957fa3c7df89b71a72f571df82c85e6ba790569fdd.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 563, + 482, + 597 + ], + "lines": [ + { + "bbox": [ + 309, + 563, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 309, + 563, + 482, + 597 + ], + "type": "text", + "content": "Fig.6: Comparing our results and PCSeg (baseline) under multi-scan visualization, showing improved segmentation results." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 612, + 246, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 612, + 246, + 623 + ], + "spans": [ + { + "bbox": [ + 132, + 612, + 246, + 623 + ], + "type": "text", + "content": "6.3 Ablation Studies" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "Effectiveness of Components: In Tab. 3, we ablate each component of RAPiD-Seg step by step and report the performance on the SemanticKITTI validation set. We start from a baseline model, achieving an mIoU of 70.04 on validation set." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "type": "text", + "content": "L. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": "Incorporating the RAPiD features leads to a notable increase in mIoU (+1.17), underscoring the efficacy of RAPiD in enhancing segmentation performance. Building upon the RAPiD framework, the integration of reflectivity further elevates our mIoU to 71.93, with an overall improvement of +1.89 from the baseline. This significant gain illustrates the critical role of reflectivity in capturing different object materials. Utilizing RAPiD AE to get the voxel-wise representation, we observe an additional improvement, taking the mIoU to 72.15 (+2.11 compared to the baseline), which demonstrates its capability in processing feature representations. We subsequently performed a module-by-module reduction to analyze the model performance. Specifically, without RAPiD-related modules (Geometric/Reflectivity/Embedding), mIoU drops from 73.02 to 70.46 - attributable to the lack of RAPiD features, which are robust to transformation and noise. Our ablation studies show that while disabling individual components leads to some improvement over the baseline, the peak performance is only attained when all components function together. This highlights the indispensable contribution of each component in maximizing segmentation accuracy." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 133, + 352, + 481, + 409 + ], + "blocks": [ + { + "bbox": [ + 130, + 317, + 482, + 350 + ], + "lines": [ + { + "bbox": [ + 130, + 317, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 317, + 482, + 350 + ], + "type": "text", + "content": "Table 4: Effects of RAPiD and Reflectivity features compared to other configurations on SemanticKITTI validation set. PDD: the compressed PDD embeddings; RAPiD-R: only RAPiD, without reflectivity feature fusion; RAPiD-R: both RAPiD and reflectivity." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 352, + 481, + 409 + ], + "lines": [ + { + "bbox": [ + 133, + 352, + 481, + 409 + ], + "spans": [ + { + "bbox": [ + 133, + 352, + 481, + 409 + ], + "type": "table", + "html": "
Conf.mIoUtrucko.vehparkwalko.grobuildfence
BaseL [36]70.059.870.369.276.936.293.769.6
PDD [52]66.240.365.867.574.633.192.868.2
RAPiD-R71.862.569.070.377.435.893.670.8
RAPiD+R73.070.478.575.878.944.294.273.1
", + "image_path": "dc6c5d064c78177a7d34649b75b3614996844b0e65dd51c41d88da345a6fc525.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 424, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 592 + ], + "type": "text", + "content": "Effectiveness of RAPiD and Reflectivity Features: In Tab. 4, we validate the efficacy of the proposed RAPiD features. The direct application of PDD results in a substantial mIoU decrease of 3.8, with most category IoUs falling below those of the baseline method. This indicates vanilla PDD features are not well-suited for LiDAR segmentation. While our proposed RAPiD (without reflectivity disparities in Eq. (2)) demonstrates an improvement of 1.4 in mIoU over the baseline. Moreover, the integration of reflectivity with RAPiD features significantly enhances performance, yielding 2.6 mIoU increase over the baseline. Notably, our approach exhibits superior performance in the segmentation of most rigid object categories compared to other configurations. Tab. 6 further shows the RAPiD outperforms PDD in varying " + }, + { + "bbox": [ + 130, + 424, + 482, + 592 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 424, + 482, + 592 + ], + "type": "text", + "content": " values at different ranges on SemanticKITTI validation set, with non-uniform " + }, + { + "bbox": [ + 130, + 424, + 482, + 592 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 424, + 482, + 592 + ], + "type": "text", + "content": " configurations yielding the most significant improvements in mIoU for both datasets, highlighting the effectiveness of the range-aware design of RAPiD." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Effectiveness of RAPiD AE and Architecture Variants: In Tab. 3, we replace our RAPiD AE with a commonly-used ConvFNN [17,50,64] to generate voxel-wise embeddings. Our AE demonstrated a " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.24\\mathrm{mIoU}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " improvement over the FNN. In Tab. 5, we investigate the impact of the variant architectures, i.e., R-RAPiD-Seg and C-RAPiD-Seg. Utilizing R-RAPiD features, R-RAPiD-Seg yields an mIoU of 72.3, which outperforms the baseline method by " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "+2.3\\mathrm{mIoU}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "type": "text", + "content": "RAPiD-Seg Networks for 3D LiDAR Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 150, + 301, + 198 + ], + "blocks": [ + { + "bbox": [ + 133, + 112, + 304, + 145 + ], + "lines": [ + { + "bbox": [ + 133, + 112, + 304, + 145 + ], + "spans": [ + { + "bbox": [ + 133, + 112, + 304, + 145 + ], + "type": "text", + "content": "Table 5: 3D segmentation results of different variants of RAPiD-Seg (ours) on SemanticKITTI validation set." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 150, + 301, + 198 + ], + "lines": [ + { + "bbox": [ + 135, + 150, + 301, + 198 + ], + "spans": [ + { + "bbox": [ + 135, + 150, + 301, + 198 + ], + "type": "table", + "html": "
MethodmIoU %carpedo.gropole
Baseline70.097.278.135.463.5
R-RAPiD-Seg72.3 (+2.3)97.477.445.062.4
C-RAPiD-Seg73.0 (+3.0)97.779.344.666.4
", + "image_path": "b9e633235092bc6fb522274be80493a7f77dae1ef341c92b1361c6631a57d6ac.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 316, + 128, + 478, + 198 + ], + "blocks": [ + { + "bbox": [ + 315, + 111, + 478, + 123 + ], + "lines": [ + { + "bbox": [ + 315, + 111, + 478, + 123 + ], + "spans": [ + { + "bbox": [ + 315, + 111, + 478, + 123 + ], + "type": "text", + "content": "Table 6: Effects of " + }, + { + "bbox": [ + 315, + 111, + 478, + 123 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 315, + 111, + 478, + 123 + ], + "type": "text", + "content": " at various ranges." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 128, + 478, + 198 + ], + "lines": [ + { + "bbox": [ + 316, + 128, + 478, + 198 + ], + "spans": [ + { + "bbox": [ + 316, + 128, + 478, + 198 + ], + "type": "table", + "html": "
PDD [53]RAPiD (ours)
knearkmidkfarmIoUknearkmidkfarmIoU
SumK77764.74 (-7.1)77772.04 (+0.2)
55565.18 (-6.6)55572.28 (+0.5)
107566.23 (-5.6)107573.02 (+1.2)
lnuScene66672.19 (-6.5)66678.76 (+0.1)
33373.68 (-5.0)33379.43 (+0.8)
86372.24 (-6.4)86379.91 (+1.3)
", + "image_path": "f966758ae9c0c52f27d74a067727e55eee921cbe8b047b0040751a88a4f2aef3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 209, + 479, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 209, + 479, + 232 + ], + "spans": [ + { + "bbox": [ + 130, + 209, + 479, + 232 + ], + "type": "text", + "content": "Concurrent fusing of both R- and C-RAPiD-Seg features in R-RAPiD-Seg improves performance by " + }, + { + "bbox": [ + 130, + 209, + 479, + 232 + ], + "type": "inline_equation", + "content": "+0.7" + }, + { + "bbox": [ + 130, + 209, + 479, + 232 + ], + "type": "text", + "content": " mIoU, which shows a " + }, + { + "bbox": [ + 130, + 209, + 479, + 232 + ], + "type": "inline_equation", + "content": "+3.0" + }, + { + "bbox": [ + 130, + 209, + 479, + 232 + ], + "type": "text", + "content": " overall mIoU enhancement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 232, + 481, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 232, + 481, + 280 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 481, + 280 + ], + "type": "text", + "content": "Effectiveness of Feature Fusion with Channel-wise Attention: In Tab. 3, we assess the prevalent approach of direct feature concatenation. Our method, employing feature fusion with channel-wise attention, enhances mIoU by 0.87, thereby conclusively demonstrating its effectiveness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 281, + 482, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 376 + ], + "type": "text", + "content": "Effectiveness of Backbone Networks: In Tab. 7, we evaluate the performance across multiple backbone networks to demonstrate the versatility and ease of integration of our modules. The results highlight the adaptability and effectiveness of our approach on both pointwise (P) and Voxel-wise (V) backbone architectures. Specifically, the point-based PTv2 backbone achieves an mIoU of 72.6, showing strong performance, particularly in the car and other ground categories. Among the voxel-based methods, Minkowski-UNet stands out with the highest overall mIoU of 73.0, excelling in the pedestrian and pole categories." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 136, + 417, + 475, + 468 + ], + "blocks": [ + { + "bbox": [ + 130, + 390, + 481, + 412 + ], + "lines": [ + { + "bbox": [ + 130, + 390, + 481, + 412 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 481, + 412 + ], + "type": "text", + "content": "Table 7: Effects of using different backbones on SemanticKITTI validation set, where P and V for Point- and Voxel-wsie methods." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 417, + 475, + 468 + ], + "lines": [ + { + "bbox": [ + 136, + 417, + 475, + 468 + ], + "spans": [ + { + "bbox": [ + 136, + 417, + 475, + 468 + ], + "type": "table", + "html": "
Repr.BackbonemIoU %carpedo.gropole
PPTv2 [55]72.6 (+2.8)97.477.445.062.4
VCylinder3D [64]69.896.974.237.963.0
VMinkowski-UNet [8]73.0 (+3.2)97.278.144.265.9
", + "image_path": "75e317ece834e83fefbcaf0d91e273a55747253c1d649be67f99a8173c4bdab5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 490, + 311, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 311, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 311, + 502 + ], + "type": "text", + "content": "7 Conclusion and Discussion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "text", + "content": "This paper presents a novel Range-Aware Pointwise Distance Distribution (RAPiD) feature and the RAPiD-Seg network for LiDAR segmentation, adeptly overcoming the constraints of single-modal LiDAR methods. The rigid transformation invariance and enhanced focus on local details of RAPiD significantly boost segmentation accuracy. RAPiD-Seg integrates a two-stage training approach with reflectivity-guided 4D distance metrics and a class-aware nested AE, achieving SoTA results on the SemanticKITTI and nuScenes datasets. Notably, our single-modal method surpasses the performance of multi-modal methods, indicating the superior efficacy of RAPiD features even compared to other modalities, including RGB and range images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "Our RAPiD features hold significant potential for application in various tasks such as point cloud registration, semi-/weakly- supervised learning, and extend to promising applications in multi-modal research in the future." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "text", + "content": "L. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 142, + 481, + 664 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 142, + 414, + 153 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 142, + 414, + 153 + ], + "spans": [ + { + "bbox": [ + 138, + 142, + 414, + 153 + ], + "type": "text", + "content": "1. Badger, B.L.: Depth and Representation in Vision Models (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 155, + 481, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 155, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 138, + 155, + 481, + 186 + ], + "type": "text", + "content": "2. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences. In: Int. Conf. Comput. Vis. pp. 9296-9306 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 188, + 481, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 188, + 481, + 220 + ], + "spans": [ + { + "bbox": [ + 138, + 188, + 481, + 220 + ], + "type": "text", + "content": "3. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: Nuscenes: A Multimodal Dataset for Autonomous Driving. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 11618-11628 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 221, + 481, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 221, + 481, + 253 + ], + "spans": [ + { + "bbox": [ + 138, + 221, + 481, + 253 + ], + "type": "text", + "content": "4. Chen, S., Wang, X., Cheng, T., Zhang, W., Zhang, Q., Huang, C., Liu, W.: Azinorm: Exploiting the Radial Symmetry of Point Cloud for Azimuth-Normalized 3D Perception. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6387-6396 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 255, + 481, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 255, + 481, + 286 + ], + "spans": [ + { + "bbox": [ + 138, + 255, + 481, + 286 + ], + "type": "text", + "content": "5. Chen, X., Xu, S., Zou, X., Cao, T., Yeung, D.Y., Fang, L.: SVQNet: Sparse Voxel-Adjacent Query Network for 4D Spatio-Temporal LiDAR Semantic Segmentation. In: Int. Conf. Comput. Vis. pp. 8569–8578 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 288, + 481, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 288, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 288, + 481, + 331 + ], + "type": "text", + "content": "6. Chen, X., Shi, S., Zhu, B., Cheung, K.C., Xu, H., Li, H.: MPPNet: Multi-frame Feature Intertwining with Proxy Points for 3D Temporal Object Detection. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Eur. Conf. Comput. Vis. pp. 680-697. Springer Nature Switzerland (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 332, + 481, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 481, + 364 + ], + "type": "text", + "content": "7. Cheng, R., Razani, R., Taghavi, E., Li, E., Liu, B.: (AF)2-S3Net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12547-12556 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 365, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 481, + 397 + ], + "type": "text", + "content": "8. Choy, C., Gwak, J., Savarese, S.: 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3070-3079 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 399, + 481, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 481, + 442 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 481, + 442 + ], + "type": "text", + "content": "9. Çiçek, Ö., Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation. In: Ourselin, S., Joskowicz, L., Sabuncu, M.R., Unal, G., Wells, W. (eds.) Med. Image Comput. Comput. Assist. Interv. pp. 424–432. Springer International Publishing (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 135, + 444, + 481, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 444, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 135, + 444, + 481, + 475 + ], + "type": "text", + "content": "10. Flitton, G., Breckon, T.P., Megherbi, N.: A Comparison of 3D Interest Point Descriptors with Application to Airport Baggage Object Detection in Complex CT Imagery. Pattern Recognition 46(9), 2420-2436 (2013)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 135, + 476, + 481, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 476, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 135, + 476, + 481, + 508 + ], + "type": "text", + "content": "11. Geiger, A., Lenz, P., Urtasun, R.: Are We Ready for Autonomous Driving? The KITTI Vision Benchmark Suite. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3354-3361 (2012)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 135, + 510, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 510, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 135, + 510, + 481, + 542 + ], + "type": "text", + "content": "12. Genova, K., Yin, X., Kundu, A., Pantofaru, C., Cole, F., Sud, A., Brewington, B., Shucker, B., Funkhouser, T.: Learning 3D Semantic Segmentation with only 2D Image Supervision. In: Int. Conf. 3D Vis. pp. 361-372 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 135, + 544, + 481, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 544, + 481, + 575 + ], + "spans": [ + { + "bbox": [ + 135, + 544, + 481, + 575 + ], + "type": "text", + "content": "13. Graham, B., Engelcke, M., van der Maaten, L.: 3D Semantic Segmentation with Submanifold Sparse Convolutional Networks. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 135, + 577, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 577, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 135, + 577, + 481, + 597 + ], + "type": "text", + "content": "14. Han, L., Zheng, T., Xu, L., Fang, L.: OccuSeg: Occupancy-Aware 3D Instance Segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2940-2949 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 135, + 599, + 481, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 599, + 481, + 631 + ], + "spans": [ + { + "bbox": [ + 135, + 599, + 481, + 631 + ], + "type": "text", + "content": "15. He, C., Li, R., Li, S., Zhang, L.: Voxel Set Transformer: A Set-to-Set Approach to 3D Object Detection from Point Clouds. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 8407-8417. IEEE (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 135, + 632, + 481, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 632, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 135, + 632, + 481, + 664 + ], + "type": "text", + "content": "16. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked Autoencoders Are Scalable Vision Learners. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 16000-16009 (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "type": "text", + "content": "RAPiD-Seg Networks for 3D LiDAR Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "17. Hou, Y., Zhu, X., Ma, Y., Loy, C.C., Li, Y.: Point-to-Voxel Knowledge Distillation for LiDAR Semantic Segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 171 + ], + "type": "text", + "content": "18. Hu, J., Shen, L., Sun, G.: Squeeze-and-Eccitation Networks. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 481, + 192 + ], + "type": "text", + "content": "19. Ioffe, S., Szegedy, C.: Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. In: Int. Conf. Mach. Learn. pp. 448-456 (2015)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 481, + 224 + ], + "type": "text", + "content": "20. Jiang, L., Shi, S., Tian, Z., Lai, X., Liu, S., Fu, C.W., Jia, J.: Guided Point Contrastive Learning for Semi-supervised Point Cloud Semantic Segmentation. In: Int. Conf. Comput. Vis. (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 481, + 246 + ], + "type": "text", + "content": "21. Jiang, M., Wu, Y., Zhao, T., Zhao, Z., Lu, C.: PointSIFT: A SIFT-like Network Module for 3D Point Cloud Semantic Segmentation (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "type": "text", + "content": "22. Kong, L., Liu, Y., Chen, R., Ma, Y., Zhu, X., Li, Y., Hou, Y., Qiao, Y., Liu, Z.: Rethinking Range View Representation for LiDAR Segmentation. In: Int. Conf. Comput. Vis. (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 278, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 481, + 300 + ], + "type": "text", + "content": "23. Kong, L., Ren, J., Pan, L., Liu, Z.: LaserMix for Semi-Supervised LiDAR Semantic Segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 21705-21715 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 300, + 481, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 343 + ], + "type": "text", + "content": "24. Kortvelesy, R., Morad, S., Prorok, A.: Permutation-Invariant Set Autoencoders with Fixed-Size Embeddings for Multi-Agent Learning. In: Int. Conf. Auton. Agents Multiagent Syst. pp. 1661-1669. International Foundation for Autonomous Agents and Multiagent Systems (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 343, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 481, + 365 + ], + "type": "text", + "content": "25. Kumar, A., Poole, B.: On Implicit Regularization in " + }, + { + "bbox": [ + 132, + 343, + 481, + 365 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 132, + 343, + 481, + 365 + ], + "type": "text", + "content": "-VAEs. In: Int. Conf. Mach. Learn. pp. 5480-5490 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 365, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 481, + 385 + ], + "type": "text", + "content": "26. Li, E., Casas, S., Urtasun, R.: MemorySeg: Online LiDAR Semantic Segmentation with a Latent Memory. In: Int. Conf. Comput. Vis. pp. 745-754 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 385, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 481, + 407 + ], + "type": "text", + "content": "27. Li, J., Dai, H., Ding, Y.: Self-Distillation for Robust LiDAR Semantic Segmentation in Autonomous Driving. In: Eur. Conf. Comput. Vis. pp. 659-676 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "type": "text", + "content": "28. Li, J., Dai, H., Han, H., Ding, Y.: MSeg3D: Multi-Modal 3D Semantic Segmentation for Autonomous Driving. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 21694-21704 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 439, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 481, + 482 + ], + "type": "text", + "content": "29. Li, L., Ismail, K.N., Shum, H.P.H., Breckon, T.P.: DurLAR: A High-Fidelity 128-Channel LiDAR Dataset with Panoramic Ambient and Reflectivity Imagery for Multi-Modal Autonomous Driving Applications. In: Int. Conf. 3D Vis. pp. 1227-1237 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 482, + 481, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 481, + 515 + ], + "type": "text", + "content": "30. Li, L., Shum, H.P.H., Breckon, T.P.: Less is More: Reducing Task and Model Complexity for Semi-Supervised 3D Point Cloud Semantic Segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 515, + 481, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 481, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 481, + 547 + ], + "type": "text", + "content": "31. Li, X., Li, R., Chen, G., Fu, C.W., Cohen-Or, D., Heng, P.A.: A Rotation-Invariant Framework for Deep Point Cloud Analysis. IEEE Trans. Vis. Comput. Graph. 28(12), 4503-4514 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 547, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 481, + 579 + ], + "type": "text", + "content": "32. Liang, J., Lai, R., Wong, T.W., Zhao, H.: Geometric Understanding of Point Clouds Using Laplace-Beltrami Operator. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 214-221 (2012)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 579, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 481, + 601 + ], + "type": "text", + "content": "33. Lin, M., Chen, Q., Yan, S.: Network In Network. In: Int. Conf. Learn. Represent. (2014)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "text", + "content": "34. Liong, V.E., Nguyen, T.N.T., Widjaja, S., Sharma, D., Chong, Z.J.: AMVNet: Assertion-based Multi-View Fusion Network for LiDAR Semantic Segmentation (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "text", + "content": "35. Liu, J., Yu, Z., Breckon, T.P., Shum, H.P.H.: U3DS3: Unsupervised 3D Semantic Scene Segmentation. In: IEEE Winter Conf. Appl. Comput. Vis. pp. 3759-3768 (2024)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "type": "text", + "content": "L. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "text", + "content": "36. Liu, Y., Bai, Y., Chen, R., Hou, Y., Shi, B., Li, Y., Kong, L.: PCSeg: An Open Source Point Cloud Segmentation Codebase. https://github.com/PJLab-ADG/PCSeg (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 184 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 184 + ], + "type": "text", + "content": "37. Liu, Y., Chen, R., Li, X., Kong, L., Yang, Y., Xia, Z., Bai, Y., Zhu, X., Ma, Y., Li, Y., Qiao, Y., Hou, Y.: UniSeg: A Unified Multi-Modal LiDAR Segmentation Network and the OpenPCSeg Codebase. In: Int. Conf. Comput. Vis. pp. 21662–21673 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "type": "text", + "content": "38. Liu, Y., Kong, L., Cen, J., Chen, R., Zhang, W., Pan, L., Chen, K., Liu, Z.: Segment Any Point Cloud Sequences by Distilling Vision Foundation Models. In: Adv. Neural Inform. Process. Syst. (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 217, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 217, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 481, + 239 + ], + "type": "text", + "content": "39. Marmolin, H.: Subjective MSE measures. IEEE Trans. Syst. Man Cybern. 16(3), 486-489 (1986)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 239, + 481, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 239, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 239, + 481, + 274 + ], + "type": "text", + "content": "40. Melia, O., Jonas, E., Willett, R.: Rotation-Invariant Random Features Provide a Strong Baseline for Machine Learning on 3D Point Clouds. Trans. Mach. Learn. Res. (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "type": "text", + "content": "41. Nunes, L., Wiesmann, L., Marcuzzi, R., Chen, X., Behley, J., Stachniss, C.: Temporal Consistent 3D LiDAR Representation Learning for Semantic Perception in Autonomous Driving. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "type": "text", + "content": "42. Preechakul, K., Chathee, N., Wizadwongsa, S., Suwajanakorn, S.: Diffusion Autoencoders: Toward a Meaningful and Decodable Representation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 10619-10629 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 341, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 481, + 363 + ], + "type": "text", + "content": "43. PyTorch: Scatter: PyTorch Scatter Documentation. https://pytorch-scatter.readthedocs.io/en/latest/functions/scatter.html (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 363, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 481, + 397 + ], + "type": "text", + "content": "44. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 652-660 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 397, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 481, + 430 + ], + "type": "text", + "content": "45. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space. In: Adv. Neural Inform. Process. Syst. vol. 30 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 430, + 481, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 481, + 464 + ], + "type": "text", + "content": "46. Shao, Y., Tan, A., Yan, T., Sun, Z., Zhang, Y., Liu, J.: MS23D: A 3D Object Detection Method Using Multi-Scale Semantic Feature Points to Construct 3D Feature Layer (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 464, + 481, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 481, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 481, + 497 + ], + "type": "text", + "content": "47. Sun, J., Xu, X., Kong, L., Liu, Y., Li, L., Zhu, C., Zhang, J., Xiao, Z., Chen, R., Wang, T., et al.: An Empirical Study of Training State-of-the-Art LiDAR Segmentation Models. arXiv:2405.14870 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 497, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 497, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 497, + 481, + 553 + ], + "type": "text", + "content": "48. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in Perception for Autonomous Driving: Waymo Open Dataset. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2446-2454 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 553, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 553, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 553, + 481, + 586 + ], + "type": "text", + "content": "49. Tang, H., Liu, Z., Zhao, S., Lin, Y., Lin, J., Wang, H., Han, S.: Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution. In: Eur. Conf. Comput. Vis. (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 586, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 481, + 609 + ], + "type": "text", + "content": "50. Unal, O., Dai, D., Van Gool, L.: Scribble-Supervised LiDAR Semantic Segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 609, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 481, + 643 + ], + "type": "text", + "content": "51. Wang, S., Zhu, J., Zhang, R.: Meta-RangeSeg: LiDAR Sequence Semantic Segmentation Using Multiple Feature Aggregation. IEEE Robot. Autom. Lett. 7(4), 9739-9746 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "52. Widdowson, D., Kurlin, V.: Resolving the Data Ambiguity for Periodic Crystals. In: Adv. Neural Inform. Process. Syst. vol. 35, pp. 24625-24638 (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "type": "text", + "content": "RAPiD-Seg Networks for 3D LiDAR Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 555 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "53. Widdowson, D., Kurlin, V.: Recognizing Rigid Patterns of Unlabeled Point Clouds by Complete and Continuous Isometry Invariants With No False Negatives and No False Positives. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1275-1284 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 182 + ], + "type": "text", + "content": "54. Widdowson, D., Mosca, M.M., Pulido, A., Cooper, A.I., Kurlin, V.: Average Minimum Distances of Periodic Point Sets - Foundational Invariants for Mapping Periodic Crystals. MATCH Commun. Math. Comput. Chem. 87(3), 529-559 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "text", + "content": "55. Wu, X., Lao, Y., Jiang, L., Liu, X., Zhao, H.: Point Transformer V2: Grouped Vector Attention and Partition-Based Pooling. In: Adv. Neural Inform. Process. Syst. (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "56. Xu, J., Zhang, R., Dou, J., Zhu, Y., Sun, J., Pu, S.: RPVNet: A Deep and Efficient Range-Point-Voxel Fusion Network for LiDAR Point Cloud Segmentation. In: Int. Conf. Comput. Vis. pp. 16024–16033 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "57. Yan, X., Gao, J., Li, J., Zhang, R., Li, Z., Huang, R., Cui, S.: Sparse Single Sweep LiDAR Point Cloud Segmentation via Learning Contextual Shape Priors from Scene Completion. In: Conf. AAAI Artif. Intell. vol. 35, pp. 3101-3109 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 481, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 313 + ], + "type": "text", + "content": "58. Yan, X., Gao, J., Zheng, C., Zheng, C., Zhang, R., Cui, S., Li, Z.: 2DPASS: 2D Priors Assisted Semantic Segmentation on LiDAR Point Clouds. In: Eur. Conf. Comput. Vis. (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 357 + ], + "type": "text", + "content": "59. Ye, D., Chen, W., Zhou, Z., Xie, Y., Wang, Y., Wang, P., Foroosh, H.: LidarMultiNet: Unifying LiDAR Semantic Segmentation, 3D Object Detection, and Panoptic Segmentation in a Single Multi-task Network. In: IEEE Conf. Comput. Vis. Pattern Recog. Worksh. (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "text", + "content": "60. Ye, D., Zhou, Z., Chen, W., Xie, Y., Wang, Y., Wang, P., Foroosh, H.: LidarMultiNet: Towards a Unified Multi-Task Network for Lidar Perception. In: Conf. AAAI Artif. Intell. vol. 37, pp. 3231-3240 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 391, + 481, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 481, + 434 + ], + "type": "text", + "content": "61. Ye, M., Wan, R., Xu, S., Cao, T., Chen, Q.: Efficient Point Cloud Segmentation with Geometry-Aware Sparse Networks. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Eur. Conf. Comput. Vis. pp. 196-212. Springer (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 435, + 481, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 467 + ], + "type": "text", + "content": "62. Zhang, C., Duc, K.D., Condon, A.: Visualizing DNA Reaction Trajectories with Deep Graph Embedding Approaches. In: Adv. Neural Inform. Process. Syst. Worksh. arXiv (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 468, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 489 + ], + "type": "text", + "content": "63. Zhao, H., Jiang, L., Jia, J., Torr, P., Koltun, V.: Point Transformer. In: Int. Conf. Comput. Vis. (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "type": "text", + "content": "64. Zhu, X., Zhou, H., Wang, T., Hong, F., Ma, Y., Li, W., Li, H., Lin, D.: Cylindrical and Asymmetrical 3D Convolution Networks for LiDAR Segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "type": "text", + "content": "65. Zhuang, Z., Li, R., Jia, K., Wang, Q., Li, Y., Tan, M.: Perception-Aware Multi-Sensor Fusion for 3D LiDAR Semantic Segmentation. In: Int. Conf. Comput. Vis. pp. 16280-16290 (2021)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "type": "text", + "content": "L. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/RAVE_ Residual Vector Embedding for CLIP-Guided Backlit Image Enhancement/7f0c9e4e-7b51-4d42-aac8-de63b12e0ec8_content_list.json b/2024/RAVE_ Residual Vector Embedding for CLIP-Guided Backlit Image Enhancement/7f0c9e4e-7b51-4d42-aac8-de63b12e0ec8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..26c2547d0a5e89f9d0d3ca03aadc6b8aa1a9eef4 --- /dev/null +++ b/2024/RAVE_ Residual Vector Embedding for CLIP-Guided Backlit Image Enhancement/7f0c9e4e-7b51-4d42-aac8-de63b12e0ec8_content_list.json @@ -0,0 +1,1959 @@ +[ + { + "type": "text", + "text": "RAVE: Residual Vector Embedding for CLIP-Guided Backlit Image Enhancement", + "text_level": 1, + "bbox": [ + 259, + 140, + 743, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tatiana Gaintseva1,2, Martin Benning3, and Gregory Slabaugh1,2", + "bbox": [ + 248, + 210, + 754, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Digital Environment Research Institute, Queen Mary University of London, UK", + "bbox": [ + 215, + 238, + 777, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Electronic Engineering and Computer Science, Queen Mary University of London, UK {t.gaintseva, g.slabaugh}@qmul.ac.uk", + "bbox": [ + 215, + 252, + 784, + 280 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Department of Computer Science, University College London, UK \nmartin.benning@ucl.ac.uk", + "bbox": [ + 272, + 280, + 728, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In this paper we propose a novel modification of Contrastive Language-Image Pre-Training (CLIP) guidance for the task of backlit image enhancement. Our work builds on the state-of-the-art CLIP-LIT approach, which learns a prompt pair by constraining the text-image similarity between a prompt (negative/positive sample) and a corresponding image (backlit image/well-lit image) in the CLIP embedding space. Learned prompts then guide an image enhancement network. Based on the CLIP-LIT framework, we propose two novel methods for CLIP guidance. First, we show that instead of tuning prompts in the space of text embeddings, it is possible to directly tune their embeddings in the latent space without any loss in quality. This accelerates training and potentially enables the use of additional encoders that do not have a text encoder. Second, we propose a novel approach that does not require any prompt tuning. Instead, based on CLIP embeddings of backlit and well-lit images from training data, we compute the residual vector in the embedding space as a simple difference between the mean embeddings of the well-lit and backlit images. This vector then guides the enhancement network during training, pushing a backlit image towards the space of well-lit images. This approach further dramatically reduces training time, stabilizes training and produces high quality enhanced images without artifacts. Additionally, we show that residual vectors can be interpreted, revealing biases in training data, and thereby enabling potential bias correction. Code is available at https://github.com/Atmyre/RAVE", + "bbox": [ + 261, + 345, + 743, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Backlit image enhancement $\\cdot$ Vision-language models $\\cdot$ Residual vector embedding", + "bbox": [ + 259, + 678, + 750, + 707 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 732, + 375, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Backlit image enhancement aims to improve images that suffer from backlighting issues. Backlighting is a phenomenon when the light source is situated behind the photographed object, often resulting in a loss of detail and contrast of some areas due to underexposure, diminishing the overall visual quality of the image. The goal of backlit image enhancement is to correct these exposure discrepancies and", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4c56d42defe9e54821bd923b385114375cbc9c9d19f5e4d519c205150b67e0df.jpg", + "image_caption": [ + "CLIP-LIT" + ], + "image_footnote": [], + "bbox": [ + 250, + 152, + 493, + 239 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bc4de61b474a377884e6caada4a335757951b525ca3b7afd6dd4b7dc2690ab7d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 152, + 743, + 239 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2be1c07027c5f1c05064f6f0efc139829124c31266cce7e4ce4f25ded599013c.jpg", + "image_caption": [ + "RAVE" + ], + "image_footnote": [], + "bbox": [ + 248, + 262, + 491, + 354 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2503d5a1fff65847e7ffbf956bb9c362f83d195ce80ae1c6c0e59ee4ca5558d3.jpg", + "image_caption": [ + "Fig. 1: Visual comparison of results obtained by RAVE vs CLIP-LIT. RAVE produces well-lit images with fewer artifacts (note the dark green color of underexposed regions in CLIP-LIT), as shown in the zoomed-in sections. Further comparisons can be found in supplementary material." + ], + "image_footnote": [], + "bbox": [ + 504, + 262, + 743, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "bring the details in the darkened areas out without overexposing the well-exposed parts of the image. Backlit image enhancement significantly improves photograph quality in uncontrolled lighting conditions. It is also vital for fields requiring clear image visibility, such as surveillance, security, and scientific imaging applications.", + "bbox": [ + 212, + 476, + 787, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, correcting backlit images is not an easy task. Manual correction requires skill using photo enhancement software, and often substantial time and effort. Automated solutions are preferred, but also face challenges. Global adjustment of the brightness level is not sufficient, as this typically overexposes regions of the image which were well lit prior to enhancement [11]. Spatially adaptive methods have appeared in the literature including early methods that rely on assumptions such as average luminance and a gray world model [8, 12] or Retinex theory [10]. More recent methods approximate well-lit image distributions via end-to-end training of deep networks [3, 26]. Despite progress in the field, considerable room for improvement exists for high quality automated backlit image enhancement.", + "bbox": [ + 212, + 537, + 787, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another challenge to high quality automatic backlit image enhancement is the lack of paired data, i.e. backlit images each with a corresponding well-lit image. It is complicated to collect such data, so approaches that work with unpaired data are imperative. Recently, CLIP-LIT [14] was proposed, which successfully utilizes CLIP [19] model guidance for training a backlit enhancement model with unpaired data. CLIP-LIT uses prompt learning techniques to provide CLIP guidance. More specifically, it constructs two learnable text prompts, which are trained to have CLIP embeddings close to the well-lit and backlit images, respectively. These learnt prompts then guide the image enhancement model", + "bbox": [ + 212, + 704, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "T. Gaintseva et al.", + "bbox": [ + 271, + 114, + 397, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "during training. Next, several iterations of additional prompt correction followed by enhancement model fine-tuning using the updated prompts are required to produce the final image enhancement model. This approach is the current state-of-the-art in backlit image enhancement with training on unpaired data.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we demonstrate that prompt training is not the most efficient way to implement the CLIP guidance, and propose two novel methods, named CLIP-LIT-Latent and ResiduAl Vector Embedding (RAVE). In CLIP-LIT-Latent, we train vectors corresponding to well-lit and backlit images directly in the CLIP latent space rather than in text embedding space. We show that this approach has similar or superior performance when trained on paired or unpaired image data. Further, in RAVE, we form the guidance vector by subtracting the mean of CLIP latent vectors of backlit images from the mean of CLIP latent vectors of well-lit images. This residual vector points in a direction of moving backlit images to well-lit images in the CLIP embedding space. This vector is then used as guidance for the image enhancement model during training. We show that such guidance enables the enhancement model to produce high quality images, while requiring much less training time. While having all these benefits, both proposed CLIP-LIT-Latent and RAVE retain all the advantages of the original CLIP-LIT method, being lightweight for inference and suitable for training on paired or unpaired data. Moreover, we show that residual vector used for guidance in RAVE is interpretable, which opens up new possibilities for overcoming biases in the training data.", + "bbox": [ + 212, + 207, + 785, + 478 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The contributions of this paper can be summarized as follows:", + "bbox": [ + 238, + 486, + 686, + 501 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Inspired by CLIP-LIT, we present two novel approaches to CLIP guidance (CLIP-LIT-Latent and RAVE) working directly in the latent space for the task of backlit image enhancement.", + "2. We show that these approaches result in similar or better quality according to quantitative metrics in both settings of training on paired or unpaired data, and superior visual quality of resulting images, while requiring considerably less time to train.", + "3. We demonstrate that the embedding used by RAVE for guidance is interpretable, and its interpretation can reveal biases in the training data." + ], + "bbox": [ + 220, + 512, + 782, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 215, + 672, + 382, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image Enhancement. Image enhancement is the task of image processing with the goal of improving the visual appearance of an image. This is an essential task in numerous fields, such as photography, medical imaging, satellite imagery analysis, and other applications where image quality is crucial. As there are many types of image degradation such as noise, blurriness, low resolution, backlighting, low light, etc., there are a range of image enhancement methods, from traditional ones such as gray level transformation [8,12], histogram equalization [5], and methods based on Retinex theory [10], to machine learning and end-to-end trained deep learning methods [3,26]. A general overview of concepts", + "bbox": [ + 212, + 704, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "RAVE: Residual Vector Embedding", + "bbox": [ + 491, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and techniques commonly used for image enhancement is provided by Maini et al. [22].", + "bbox": [ + 215, + 146, + 784, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Backlit Image Enhancement. For the task of backlit image enhancement, many approaches have been proposed. Some techniques aim to identify backlit regions and correct them separately from the rest of the image [13]. Methods based on the Retinex [10] theory divide an image into a reflectance image and an illumination map, and both parts are then enhanced and then combined back together to form the resulting enhanced image. Another approach to enhancing backlit images is using end-to-end deep learning approaches, as in Restormer [26] or Retinexformer [3] methods. Combined techniques based on Retinex theory and deep learning were proposed for both backlit and low-light image enhancement, such as Diff-Retinex [25]. These methods decomposing an input image into a reflectance image and an illumination map and apply diffusion methods for enhancement before combining them to produce the output image. Finally, another set of methods are based on high dynamic range (HDR) imaging which combine multiple exposures to produce a composite well-lit image [4, 20, 28]. However, in this work we only focus on single-image methods where back-lit correction is applied as a post-process.", + "bbox": [ + 215, + 178, + 785, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fundamentally, backlit image correction is ill-posed, in that for a single backlit image, multiple well-lit variants with high perceptual quality exist. Relatedly, generation of paired training data requires hiring professional photographers, which can be expensive. To help address this problem, recently the BAID [15] dataset was proposed, which contains paired images generated by professional photographers. Nevertheless, development of methods for backlit image enhancement, which can be trained using non-paired data remains essential.", + "bbox": [ + 215, + 421, + 785, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CLIP guidance. CLIP [19] is a model that consists of two parts: image and text encoders, which learn to project images and text to a common latent embedding space. CLIP has been proven to provide strong image and text priors which are able to provide guidance in many tasks. CLIP guidance has been used in text-to-image generation [17], visual question answering [18], GAN inversion [2], 3D scene generation [7], domain adaptation of image generators [6] and more. Liang et al. [14] also used CLIP guidance for the task of backlit image enhancement. In our work, we aim to improve this technique by modifying the CLIP guidance.", + "bbox": [ + 215, + 527, + 785, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 217, + 679, + 377, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our work is built on top of the CLIP-LIT [14] approach, which uses CLIP guidance for training a backlit image enhancement model. We propose two modifications of CLIP guidance, which reduce training time, stabilize training and produce high quality enhanced images with fewer artifacts compared to CLIP-LIT.", + "bbox": [ + 215, + 717, + 784, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We start in Section 3.1 by reviewing the architecture of original CLIP-LIT method on which our work is based. Then, in Sections 3.2 and 3.3 we introduce our proposed methods CLIP-LIT-Latent and RAVE.", + "bbox": [ + 215, + 794, + 784, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "T. Gaintseva et al.", + "bbox": [ + 271, + 114, + 395, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d8392715a7afe95478d5da789a027ff9b1d0ff194e10a4f8a90415ea6d650bf5.jpg", + "image_caption": [ + "(a) Initial Prompts and Enhancement Training" + ], + "image_footnote": [], + "bbox": [ + 223, + 162, + 815, + 318 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1965875bbff3a3447b6ca8478c55d5c3d8e14101dfd5fc041ec04c8b294c1d67.jpg", + "image_caption": [ + "(b) Prompt Refinement and Enhancement Tuning", + "Fig. 2: Overview of the original CLIP-LIT approach and proposed CLIP-LIT-Latent. (a) depicts the first stage of training, which consists of prompt or latent vector initialization and the initial training of an enhancement network. (b) shows the second stage, where prompt/latent vector refinement and enhancement model fine-tuning are iteratively repeated. Blue and red boxes are related to CLIP-LIT and CLIP-LIT-Latent." + ], + "image_footnote": [], + "bbox": [ + 220, + 339, + 815, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 CLIP-LIT overview", + "text_level": 1, + "bbox": [ + 215, + 647, + 423, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The CLIP-LIT framework (Fig. 2) consists of two stages, described below.", + "bbox": [ + 215, + 698, + 754, + 714 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Initial Prompts and Enhancement Training", + "text_level": 1, + "bbox": [ + 212, + 734, + 575, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this stage, two prompts $T_{p}, T_{n} \\in \\mathbb{R}^{N \\times 512}$ representing positive and negative prompts are randomly initialized, where $N$ is a number of tokens in each prompt. Then they are tuned so that the CLIP embedding of $T_{p}$ is close in terms of the dot product to the CLIP embeddings of well-lit training images, and CLIP embedding of $T_{n}$ is close to the CLIP embeddings of backlit images from the training data. The binary cross-entropy loss of classifying the backlit and well", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "RAVE: Residual Vector Embedding", + "bbox": [ + 491, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "lit images based on $T_{p}$ and $T_{n}$ is used for this purpose:", + "bbox": [ + 215, + 146, + 612, + 162 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {i n i t i a l}} = - (y \\log (\\widehat {y}) + (1 - y) \\log (1 - \\widehat {y}))\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 172, + 648, + 191 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {y} = \\frac {e ^ {\\cos \\left(\\Phi_ {\\text {i m a g e}} (I) , \\Phi_ {\\text {t e x t}} \\left(T _ {p}\\right)\\right)}}{\\sum_ {i \\in \\{n , p \\}} e ^ {\\cos \\left(\\Phi_ {\\text {i m a g e}} (I) , \\Phi_ {\\text {t e x t}} \\left(T _ {i}\\right)\\right)}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 193, + 785, + 232 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here $I$ is a well-lit or backlit input image with label $y \\in \\{0,1\\}$ with 0 corresponding to backlit and 1 corresponding to well-lit; $\\varPhi_{\\mathrm{image}}$ is a CLIP [19] image encoder, $\\varPhi_{\\mathrm{text}}$ is a CLIP text encoder; cos is a cosine similarity function.", + "bbox": [ + 215, + 242, + 782, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Using these initialized prompts $T_{p}$ and $T_{n}$ it is now possible to train an image enhancement network with a CLIP-guided loss. The enhancement model is a Unet [21] which takes backlit image $I_{b}$ as input and outputs an estimated illumination map $I_{i} \\in \\mathbb{R}^{H \\times W \\times 1}$ . The final enhanced image $I_{t}$ is then obtained as $I_{t} = I_{b} / I_{i}$ . The image enhancement network is trained by combining two losses:", + "bbox": [ + 215, + 287, + 784, + 363 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {e n h a n c e}} = L _ {\\text {c l i p}} + \\omega L _ {\\text {i d e n t i t y}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 378, + 785, + 395 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here $L_{\\mathrm{clip}}$ measures the similarity between the enhanced image and the learned prompts $T_{p}$ and $T_{n}$ in the CLIP latent space. $L_{\\mathrm{identity}}$ is an identity loss designed to ensure that the enhanced image is similar to the original input image in terms of content and structure. The parameter $\\omega$ balances the two losses. These losses are defined as:", + "bbox": [ + 215, + 401, + 782, + 474 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {c l i p}} = \\frac {e ^ {\\cos \\left(\\Phi_ {\\mathrm {i m a g e}} \\left(I _ {t}\\right) , \\Phi_ {\\mathrm {t e x t}} \\left(T _ {n}\\right)\\right)}}{\\sum_ {i \\in \\{n , p \\}} e ^ {\\cos \\left(\\Phi_ {\\mathrm {i m a g e}} \\left(I _ {t}\\right) , \\Phi_ {\\mathrm {t e x t}} \\left(T _ {i}\\right)\\right)}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 473, + 785, + 512 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {i d e n t i t y}} = \\sum_ {l = 0} ^ {k} \\alpha_ {l} \\cdot \\left\\| \\Phi_ {\\text {i m a g e}} ^ {l} \\left(I _ {b}\\right) - \\Phi_ {\\text {i m a g e}} ^ {l} \\left(I _ {t}\\right) \\right\\| _ {2} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 518, + 785, + 560 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\Phi^l$ is $l^{th}$ layer of the CLIP image encoder, and $\\alpha_{l}$ is its corresponding weight. $k$ is number of layers used to calculate the metric.", + "bbox": [ + 215, + 566, + 782, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prompt Refinement and Enhancement Tuning", + "text_level": 1, + "bbox": [ + 215, + 613, + 599, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CLIP-LIT applies a second stage where several cycles of prompt refinement and enhancement network tuning are performed to improve the accuracy of the learned prompts.", + "bbox": [ + 215, + 628, + 782, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "First, learned prompts $T_{p}$ and $T_{n}$ are fine-tuned to provide better guidance for fine-tuning the image enhancement model. A margin ranking loss is used to update the prompts based on images given by the current enhancement model:", + "bbox": [ + 215, + 674, + 782, + 719 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {\\text {p r o m p r e t}} ^ {1} = \\max (0, S \\left(I _ {w}\\right) - S \\left(I _ {b}\\right) + m _ {0}) + \\max (0, S \\left(I _ {t - 1}\\right) - S \\left(I _ {b}\\right) + m _ {0}) \\\\ + \\max (0, S \\left(I _ {w}\\right) - S \\left(I _ {t}\\right) + m _ {1}) + \\max (0, S \\left(I _ {t}\\right) - S \\left(I _ {t - 1}\\right) + m _ {2}) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 729, + 743, + 767 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $S(I)$ is a negative similarity score between the prompt pair and an image defined as:", + "bbox": [ + 215, + 779, + 782, + 806 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nS (I) = \\frac {e ^ {\\cos \\left(\\Phi_ {\\text {i m a g e}} (I) , \\Phi_ {\\text {t e x t}} \\left(T _ {n}\\right)\\right)}}{\\sum_ {i \\in \\{n , p \\}} e ^ {\\cos \\left(\\Phi_ {\\text {i m a g e}} (I) , \\Phi_ {\\text {t e x t}} \\left(T _ {i}\\right)\\right)}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 361, + 805, + 785, + 844 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "T. Gaintseva et al.", + "bbox": [ + 271, + 114, + 397, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here $m_0 \\in [0,1]$ is the margin between the score of well-lit or enhanced result and the backlit image in the CLIP embedding space; $m_1 \\in [0,1]$ is the margin between the score of the enhanced result and the well-lit image in the CLIP embedding space; $m_2 \\in [0,1]$ is the margin between the newly enhanced result $I_t$ and previously enhanced result $I_{t-1}$ .", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After that, a stage of enhancement model fine-tuning is performed. This follows the same process as initial enhancement model training described above. The process of prompt refinement and enhancement tuning is repeated several times until satisfactory results are obtained.", + "bbox": [ + 212, + 222, + 782, + 281 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 CLIP-LIT-Latent", + "text_level": 1, + "bbox": [ + 215, + 304, + 405, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Like CLIP-LIT, CLIP-LIT-Latent also consists of two stages of initial prompts and enhancement training and prompt refinement and enhancement tuning. Fig. 2 illustrates the CLIP-LIT-Latent framework.", + "bbox": [ + 212, + 328, + 782, + 371 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Instead of learning prompts in the text embedding space, CLIP-LIT-Latent learns a pair of positive/negative vectors directly in the CLIP latent space. In the prompt initialization phase, we randomly initialize two vectors $E_{p}, E_{n} \\in \\mathbb{R}^{512}$ . These vectors are then tuned so that $E_{p}$ is close in terms of the dot product to the CLIP embedding of well-lit images from the training data, and $E_{n}$ is close to the CLIP embedding of backlit images. As in original CLIP-LIT approach, binary cross-entropy loss for classifying the backlit and well-lit images based on $E_{p}$ and $E_{n}$ is used for this purpose:", + "bbox": [ + 212, + 373, + 784, + 494 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {\\text {i n i t i a l}} = - (y \\log (\\widehat {y}) + (1 - y) \\log (1 - \\widehat {y})) \\\\ \\widehat {y} = \\frac {e ^ {\\cos \\left(\\Phi_ {\\mathrm {i m a g e}} (I) , E _ {p}\\right)}}{\\sum_ {i \\in \\{n , p \\}} e ^ {\\cos \\left(\\Phi_ {\\mathrm {i m a g e}} (I) , E _ {i}\\right)}} \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 506, + 782, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $I$ is a well-lit or backlit input image with label $y \\in 0, 1$ with 0 corresponding to backlit and 1 corresponding to well-lit; $\\varPhi_{\\mathrm{image}}$ is a CLIP image encoder.", + "bbox": [ + 212, + 570, + 782, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on these initialized vectors $E_{p}$ and $E_{n}$ , an enhancement network is trained with a CLIP-guided loss. As before, the enhancement model is a Unet which takes backlit image $I_{b}$ as input and outputs an estimated illumination map $I_{i} \\in \\mathbb{R}^{H \\times W \\times 1}$ . And, as before, it is trained using $L_{\\text{enhance}}$ loss, where the $L_{\\text{clip}}$ component is changed as follows:", + "bbox": [ + 212, + 601, + 784, + 676 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {c l i p}} = \\frac {e ^ {\\cos \\left(\\Phi_ {\\text {i m a g e}} \\left(I _ {r}\\right) , E _ {n}\\right)}}{\\sum_ {i \\in \\{n , p \\}} e ^ {\\cos \\left(\\Phi_ {\\text {i m a g e}} \\left(I _ {t}\\right) , E _ {i}\\right)}} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 382, + 688, + 782, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\varPhi^l$ is $l^{th}$ layer of CLIP image encoder, and $a_{l}$ is its corresponding weight.", + "bbox": [ + 212, + 734, + 782, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the second stage of prompt refinement and enhancement tuning, where learned vectors $E_{p}$ and $E_{n}$ are fine-tuned using margin ranking loss, the formula for $S(I)$ changes as follows:", + "bbox": [ + 212, + 751, + 782, + 795 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nS (I) = \\frac {e ^ {\\cos \\left(\\Phi_ {\\text {i m a g e}} (I) , E _ {n}\\right)}}{\\sum_ {i \\in \\{n , p \\}} e ^ {\\cos \\left(\\Phi_ {\\text {i m a g e}} (I) , E _ {i}\\right)}} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 806, + 782, + 843 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "RAVE: Residual Vector Embedding", + "bbox": [ + 491, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5e49f5cf0517de84b6d9d8538afbddbc9af2e198fc0c3162feefff15d98ddd33.jpg", + "image_caption": [ + "Fig. 3: Overview of the RAVE model. (a) First, we calculate residual vector $\\mathbf{v}_{\\mathrm{residual}}$ based on backlit and well-lit training data. (b) Then we switch to enhancement model training based on the identity loss and the loss based on the residual vector." + ], + "image_footnote": [ + "- CLIP vectors of backlit training images" + ], + "bbox": [ + 228, + 152, + 514, + 314 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cfa3f313b5396055e19418725ab6a4d126cbcafc9bef8110c21b540484222867.jpg", + "image_caption": [], + "image_footnote": [ + "- CLIP vectors of well-lit training images", + "(a) Calculation of residual vector $\\mathbf{v}_{\\text{residual}}$ based on (b) Training of enhancement model using paired or unpaired training data guidance from $\\mathbf{v}_{\\text{residual}}$" + ], + "bbox": [ + 532, + 247, + 785, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In all other aspects the training algorithm is similar to that of CLIP-LIT.", + "bbox": [ + 214, + 474, + 746, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CLIP-LIT-Latent has several advantages over CLIP-LIT. First, CLIP-LIT-Latent does not use the text encoder, which speeds up the training and inference processes, as it is no longer necessary to pass vectors and gradients through the text encoder. This also potentially enables the use of other vision models which do not have text encoder parts rather than CLIP for guidance. Second, images obtained from CLIP-LIT-Latent tend to have more contrast and have better visual quality (see Fig. 4 and Fig. 1). We speculate this results from tuning vectors $E_{p}$ and $E_{n}$ directly in the CLIP latent space. In contrast, in the original CLIP-LIT, vectors $T_{p}$ and $T_{n}$ belong to the text embedding space, and must be tuned so that their projection through the CLIP text encoder satisfies desired conditions. Such optimization is more complex and makes tuning of $T_{p}$ and $T_{n}$ harder and less accurate.", + "bbox": [ + 212, + 502, + 787, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 RAVE: Residual Vector Embedding", + "text_level": 1, + "bbox": [ + 214, + 714, + 558, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section we present our model RAVE (ResiduAl Vector Embedding). RAVE includes a further modification of the CLIP guidance. We show that RAVE does not require iterative stages of prompt and model updates. Instead it only requires one stage of enhancement model training.", + "bbox": [ + 212, + 747, + 787, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In RAVE we exploit arithmetic defined in the CLIP latent space. Using well-lit and backlit training data, we construct a residual vector $\\mathbf{v}_{\\mathrm{residual}}$ , which will", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "T. Gaintseva et al.", + "bbox": [ + 271, + 114, + 397, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "then be used for enhancement model guidance. $\\mathbf{v}_{\\mathrm{residual}}$ is defined as follows:", + "bbox": [ + 215, + 146, + 766, + 162 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {\\text {r e s i d u a l}} = f _ {\\text {n o r m}} \\left(\\mathbf {v} _ {\\text {w e l l - l i t}} - \\mathbf {v} _ {\\text {b a c k l i t}}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 169, + 593, + 184 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {\\text {w e l l - l i t}} = f _ {\\text {n o r m}} \\left(\\frac {\\sum_ {i = 1} ^ {N _ {p}} f _ {\\text {n o r m}} \\left(\\Phi_ {\\text {i m a g e}} \\left(I _ {i} ^ {p}\\right)\\right)}{N _ {p}}\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 188, + 784, + 228 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {\\text {b a c k l i t}} = f _ {n o r m} \\left(\\frac {\\sum_ {i = 1} ^ {N _ {n}} f _ {\\text {n o r m}} \\left(\\Phi_ {\\text {i m a g e}} \\left(I _ {i} ^ {n}\\right)\\right)}{N _ {n}}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 231, + 655, + 271 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $f_{\\mathrm{norm}}$ is a normalization function: $f_{\\mathrm{norm}}(\\mathbf{v}) = \\frac{\\mathbf{v}}{\\|\\mathbf{v}\\|_2}$ , $N_p$ is the number of training well-lit images, $N_n$ is the number of training backlit images, $I_i^p$ is the $i^{th}$ well-lit training image, $I_i^n$ is the $i^{th}$ backlit training image.", + "bbox": [ + 214, + 277, + 785, + 325 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "$\\mathbf{v}_{\\mathrm{residual}}$ is a vector that points in a direction moving from backlit images to well-lit images in the CLIP embedding space. We then use this vector as guidance for the image enhancement model during training. This will train the image enhancement model to produce images with CLIP latent vectors that are close to the CLIP latent vectors of well-lit training images. Overview of the RAVE approach is presented on Fig. 3.", + "bbox": [ + 214, + 325, + 782, + 415 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We use the same Unet [21] as enhancement model as in CLIP-LIT and CLIP-LIT-Latent. The enhancement model is trained using the following loss function:", + "bbox": [ + 214, + 416, + 782, + 445 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {r a v e}} = L _ {\\text {i d e n t i t y}} + \\omega L _ {\\text {r e s i d u a l}} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 462, + 784, + 477 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {r e s i d u a l}} = \\left\\| \\Phi_ {\\text {i m a g e}} (I) \\cdot \\mathbf {v} _ {\\text {r e s i d u a l}} - \\mathbf {v} _ {\\text {w e l l - l i t}} \\cdot \\mathbf {v} _ {\\text {r e s i d u a l}} \\right\\| ^ {2} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 494, + 784, + 512 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $I$ is an input image, and $v_{\\mathrm{well - lit}}$ is a normalized mean of CLIP embeddings of well-lit training images:", + "bbox": [ + 214, + 516, + 782, + 546 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {\\text {w e l l - l i t}} = f _ {\\text {n o r m}} \\left(\\frac {\\sum_ {i = 1} ^ {N _ {p}} f _ {\\text {n o r m}} \\left(\\Phi_ {\\text {i m a g e}} \\left(I _ {i} ^ {p}\\right)\\right)}{N _ {p}}\\right) \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 553, + 784, + 592 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Note that RAVE does not require several repeating stages during training. Before training, the residual vector $\\mathbf{v}_{\\mathrm{residual}}$ is calculated, and afterwards the training of the image enhancement model begins using $L_{rave}$ as the loss function. This makes RAVE training much more efficient than that of CLIP-LIT and CLIP-LIT-Latent. First, the training of the image enhancement model no longer depends on the quality of learned prompt embeddings, as a fixed residual vector is used for guidance. Also, the iterative manner of CLIP-LIT and CLIP-LIT-Latent training regimes makes training less stable, as the image enhancement model training depends on the quality of prompts obtained on the previous step and vice versa. In supplementary material we show that results obtained on intermediate steps of CLIP-LIT training can be of poor quality and can fluctuate from producing under-exposed to over-exposed resulting images from step to step. In contrast, RAVE has a well-defined fixed initial objective. Results show that RAVE converges up to 25 times faster than CLIP-LIT and CLIP-LIT-Latent. Furthermore, images obtained by RAVE have fewer over-exposed areas, producing high quality results.", + "bbox": [ + 214, + 598, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "RAVE: Residual Vector Embedding", + "bbox": [ + 491, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 143, + 375, + 162 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets Following CLIP-LIT, we use the BAID [15] and DIV2K [1] datasets for training. We train our models in two settings: using paired and unpaired training data. For the paired setting, we use 380 backlit and corresponding well-lit images from BAID dataset. For the unpaired setting, we use 380 backlit images from the BAID and 384 images from DIV2K dataset. For the unpaired setting, we use the same data as used in CLIP-LIT. For testing, we follow recent literature and use the BAID [15] test dataset, which has 368 backlit and corresponding well-lit images. The data is publicly available.", + "bbox": [ + 212, + 174, + 785, + 295 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training and inference We follow parameter settings of the original CLIP-LIT approach when setting parameters for CLIP-LIT-Latent and RAVE (where applicable), as well as for reproducing CLIP-LIT. Specific details of training and values of hyperparameters are presented in supplementary material.", + "bbox": [ + 212, + 296, + 785, + 356 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Metrics Following recent works in image enhancement, we report four metrics: PSNR [9], SSIM [23], and LPIPS [27] (AlexNet version) and FID [16].", + "bbox": [ + 212, + 357, + 785, + 387 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Methods for comparison The main model which we compare our approaches to is CLIP-LIT [14], as it showed superior performance on all mentioned metrics, except FID, to many prior image enhancement models. We also include recently published methods such as Restormer [26], Retinexformer [3], DiffIR [24] and Diff-Retinex [25] to the comparison.", + "bbox": [ + 212, + 387, + 785, + 463 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Results", + "text_level": 1, + "bbox": [ + 215, + 483, + 323, + 497 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Quantitative comparison The quantitative comparison on the BAID test dataset is presented in Tab. 1. We report test set metrics of the best model checkpoint obtained during training for all the models. Detailed metrics values for all the checkpoints after each training epoch can be found in supplementary material. Note that many approaches can only be trained using paired data, that is why their results are only present in the \"paired\" section. We see that our method achieves state-of-the-art performance in both settings. In supplementary material, we also present results on time needed to train all the models, and confirm that both CLIP-LIT-Latent and RAVE are much more efficient. Note that in the setting of unpaired data we have two versions of RAVE, referred as RAVE and RAVE shifted in Tab. 1. RAVE shifted is a variant of RAVE with a simple modification of residual vector, see \"Interpretation of the residual vector\" section below for more details.", + "bbox": [ + 212, + 507, + 785, + 702 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Visual comparison For visual comparison, we present representative examples from the BAID test dataset. First, in Fig. 4, we compare CLIP-LIT-Latent and RAVE to original CLIP-LIT. First, note that CLIP-LIT-Latent produces more contrast in general, giving more diverse illumination to different regions. Second, in Fig. 1, we present comparison of images produced by RAVE to images produced by CLIP-LIT. We show that images produced by RAVE have fewer artifacts. We hypothesize that this advantage of RAVE may result from its guiding mechanism: in RAVE we directly point the enhancement model to produce images which have similar CLIP embeddings to that of real well-lit images. This", + "bbox": [ + 212, + 704, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "T. Gaintseva et al.", + "bbox": [ + 271, + 114, + 395, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/6bccaa16bb1086cfa711d4b078891d3bd1337fe1410e77a0488d0e174c9b0e28.jpg", + "table_caption": [ + "Table 1: Quantitative comparison of different methods on the BAID test dataset. The best and second best performances in both settings are in bold and underlined." + ], + "table_footnote": [], + "table_body": "
MethodPSNR ↑SSIM ↑LPIPS ↓FID ↓
Original images16.910.760.2152.47
PairedRestormer21.070.8320.19241.17
Retinexformer22.030.8620.17345.27
DiffIR21.100.8350.17540.35
Diff-Retinex22.070.8610.16038.07
CLIP-LIT21.930.8750.15641.16
CLIP-LIT-Latent (ours)21.840.8770.15542.56
RAVE (ours)22.260.8800.13936.01
UnpairedEnlightenGAN17.550.8640.19643.50
CLIP-LIT21.590.8740.16046.49
CLIP-LIT-Latent (ours)21.450.8770.15646.20
RAVE (ours)20.390.8610.15540.11
RAVE shifted (ours)21.370.8770.14940.90
", + "bbox": [ + 223, + 184, + 779, + 390 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "does not allow model to produce regions with artifacts, as its CLIP embeddings won't match those of real well-lit images. In CLIP-LIT or CLIP-LIT-Latent, on the other hand, the objective of the enhancement model is to produce images for which the CLIP embeddings are simply closer to the positive prompt/vector than to the negative one. This allows the enhancement model to produce less accurate images, CLIP embeddings of which are just \"close enough\" to the CLIP embeddings of real well-lit images. More results of visual comparison of all the models can be found in supplementary material.", + "bbox": [ + 212, + 417, + 784, + 539 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Interpretation of the residual vector In RAVE, we build the residual vector $\\mathbf{v}_{\\text{residual}}$ following formula 16. We noted that $\\mathbf{v}_{\\text{residual}}$ is a vector that points to a direction of moving from backlit images to well-lit images in the CLIP embedding space. Here we prove this by interpreting vector $\\mathbf{v}_{\\text{residual}}$ using the fact that CLIP has a common vector space for embeddings of images and text.", + "bbox": [ + 212, + 539, + 784, + 614 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We go through the vocabulary of the CLIP text encoder and find those vocabulary tokens which have the closest and farthest CLIP embeddings to the $\\mathbf{v}_{\\text{residual}}$ in the CLIP latent space in terms of cosine similarity. More formally, we use the following metric:", + "bbox": [ + 212, + 614, + 784, + 675 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {s i m} (\\text {t o k e n}) = f _ {\\text {n o r m}} \\left(\\Phi_ {\\text {t e x t}} \\left(e _ {\\text {t o k e n}}\\right)\\right) \\cdot \\mathbf {v} _ {\\text {r e s i d u a l}} \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 689, + 784, + 705 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Here $e_{token}$ is embedding of vocabulary token, which is input to the CLIP text encoder $\\varPhi_{text}$ . We sort vocabulary tokens using sim. Table 2 shows 7 vocabulary tokens which have highest and lowest cosine similarity to the $v_{residual}$ vectors obtained from the paired dataset and from the unpaired dataset.", + "bbox": [ + 212, + 718, + 784, + 779 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In the case of paired data, many of the least similar tokens have a meaning around \"darkness\": \"silhouette\", \"dark\", \"shadow\", etc., and absolute value of their cosine similarity to $v_{\\mathrm{residual}}$ is around $0.12 - 0.15$ . At the same time, most similar tokens do not share any meaningful semantics, and the absolute value", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "RAVE: Residual Vector Embedding", + "bbox": [ + 491, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/04c78a3ca080e4259ef79c6fb999e0d808b813d5321854d749f9c4c7e01e525f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 147, + 421, + 222 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6fed4143d3a5ab7d022ff6e187c7f52189a5a9558400e76b513d97774b3c6c6d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 147, + 570, + 222 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/54805cd322501140ecc7b5532f9f0fed0865741ba26acfd77b5beea670b81a26.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 147, + 718, + 222 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7427c9ddc16216b987ae04de1d7222ecbe8906eaa1c07db025bd39eb07f93e61.jpg", + "image_caption": [ + "CLIP-LIT" + ], + "image_footnote": [], + "bbox": [ + 277, + 227, + 421, + 303 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/82348f4401a069e6f67850de01bbcf94810d68ccefe23c4e736cd39733d11bc3.jpg", + "image_caption": [ + "CLIP-LIT-Latent" + ], + "image_footnote": [], + "bbox": [ + 424, + 227, + 570, + 303 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/83f7b1f3420458904e917c1c9736ec137d30a2b24d35d76fe1ecea2bd0f817ff.jpg", + "image_caption": [ + "RAVE" + ], + "image_footnote": [], + "bbox": [ + 573, + 227, + 718, + 303 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ae3f81a5ea7377334f8366ff5e56242a551a3204f359a4edced299113d0d1319.jpg", + "image_caption": [], + "image_footnote": [ + "(a) Visual comparison of results by CLIP-LIT, CLIP-LIT-Latent and RAVE, all trained with paired data. RAVE produces results with more general contrast." + ], + "bbox": [ + 279, + 352, + 421, + 425 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7bbc3161de51bd7c8cc44a87dfdb5698b1d507e11a605e014e2e802bc262ccec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 352, + 570, + 425 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f8439baae704b93c0f573f02ca0f5cc3c183d24085937c8076806ca975e784b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 352, + 717, + 425 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d3faeb33581fbf8eb85efcc16a0ea4df9766c19a5bff3ced2689c54ef9182ec1.jpg", + "image_caption": [ + "CLIP-LIT" + ], + "image_footnote": [ + "(b) Visual comparison of results obtained by CLIP-LIT, CLIP-LIT-Latent and RAVE (shifted version), all trained with unpaired data. RAVE produces results with more general contrast." + ], + "bbox": [ + 279, + 431, + 421, + 505 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/480b1de885fcb754ab27679ecca57a84f6f6fabebc2ad223bb49a48239292337.jpg", + "image_caption": [ + "CLIP-LIT-Latent" + ], + "image_footnote": [], + "bbox": [ + 426, + 431, + 570, + 505 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/47cc75a10ec8f6532bed5d9d58e8f32a80ba0f60915653e40ed31cb8981ad4f2.jpg", + "image_caption": [ + "RAVE", + "Fig. 4: Visual comparison of results by CLIP-LIT, CLIP-LIT-Latent and RAVE." + ], + "image_footnote": [], + "bbox": [ + 573, + 431, + 715, + 505 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "of their cosine similarity to $v_{\\mathrm{residual}}$ is much smaller, being around $0.01 - 0.04$ . This supports the hypothesis that $v_{\\mathrm{residual}}$ has a meaning of shifting from \"dark\", \"shadowy\" images to \"normal\" images, having average level of lightness. $v_{\\mathrm{residual}}$ does not carry any other semantic meaning, as it was calculated using paired dataset, where for each backlit image there is its well-lit analogue, and all the semantics present in the BAID dataset is cancelled as a result of subtraction of the mean well-lit and backlit vectors when constructing $v_{\\mathrm{residual}}$ . On the other hand, in the case of unpaired data, we see that the least similar tokens have a meaning around \"asian\": \"busan\", \"beijing\", etc., and the most similar ones have a meaning around \"nature\": \"wildlife\", \"southafrica\", \"countrylife\", etc. Herein, an absolute value of cosine similarities for both sets of tokens if quite high, being around $0.08 - 0.12$ . This reflects the bias in the unpaired datasets. Indeed, in the BAID dataset there are mostly images of Chinese cities or people, while in DIV2K dataset there are many pictures of nature. Due to the bias, RAVE training with unpaired data becomes less effective, making the enhancement model", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "T. Gaintseva et al.", + "bbox": [ + 271, + 114, + 395, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/83ab36fb5e3305dc222244f2839ffcab08cd43ca30ccfe398fab6cd68d2cd7fc.jpg", + "table_caption": [ + "Table 2: Vocabulary tokens and their cosine similarities, which have the lowest and highest cosine similarity to the $v_{\\text{residual}}$ vectors calculated for paired and unpaired training datasets." + ], + "table_footnote": [], + "table_body": "
PairedUnpaired
Lowest similarityHighest similarityLowest similarityHighest similarity
silhouette-0.151healthtech0.035busan-0.108wildlifewednesdayday0.116
darkness-0.149whitepaper0.035beijing-0.103southafrica0.103
dark-0.139theeconomist0.030iu-0.100ecuador0.095
webcamtoy-0.139bona0.029guangzhou-0.095patrol0.095
blackand-0.132digitaltransformation0.027incheon-0.094countryfile0.093
vedere-0.131sua0.021shenzhen-0.094tasmanian0.093
shadow-0.130amarketing0.021eunhyuk-0.091womensart0.091
", + "bbox": [ + 215, + 198, + 802, + 335 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "shift images not in an ideal direction in the CLIP latent space. Nevertheless, RAVE still shows good performance when trained with unpaired data.", + "bbox": [ + 212, + 363, + 782, + 393 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We also propose a simple modification for correcting the residual vector in the unpaired setting. We simply take $n$ most and least similar tokens to form an additional residual vector as follows:", + "bbox": [ + 212, + 393, + 784, + 438 + ], + "page_idx": 12 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {v} _ {\\mathrm {a d d \\_ r e s i d u a l}} = f _ {\\mathrm {n o r m}} (\\mathbf {v} _ {\\mathrm {c l o s e s t}} - \\mathbf {v} _ {\\mathrm {f a r t h e s t}}) \\\\ \\mathbf {v} _ {\\text {c l o s e s t}} = f _ {\\text {n o r m}} \\left(\\frac {\\sum_ {i = 1} ^ {n} f _ {\\text {n o r m}} \\left(\\Phi_ {\\text {t e x t}} \\left(T _ {i} ^ {c}\\right)\\right)}{n}\\right) \\tag {15} \\\\ \\mathbf {v} _ {\\mathrm {f a r t h e s t}} = f _ {n o r m} \\left(\\frac {\\sum_ {i = 1} ^ {n} f _ {\\mathrm {n o r m}} (\\varPhi_ {\\mathrm {t e x t}} (T _ {i} ^ {f}))}{n}\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 464, + 784, + 560 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "$T_{i}^{c}$ and $T_{i}^{f}$ are tokens with $i^{th}$ most and least similar embeddings to the $\\mathbf{v}_{\\mathrm{residual}}$ . And finally we update the $\\mathbf{v}_{\\mathrm{residual}}$ as follows:", + "bbox": [ + 214, + 571, + 782, + 603 + ], + "page_idx": 12 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {\\text {r e s i d u a l}} = \\mathbf {v} _ {\\text {r e s i d u a l}} - c o s \\left(\\mathbf {v} _ {\\text {r e s i d u a l}}, \\mathbf {v} _ {\\text {a d d} \\_ \\text {r e s i d u a l}}\\right) \\cdot \\mathbf {v} _ {\\text {a d d} \\_ \\text {r e s i d u a l}} \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 284, + 616, + 784, + 633 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Resulting vector $\\mathbf{v}_{\\mathrm{residual}}$ will keep the meaning of shifting from \"dark\", images to \"normal\" images while losing the unwanted semantics. Experimental results show that RAVE trained with such updated $\\mathbf{v}_{\\mathrm{residual}}$ produces better results (see RAVE shifted in Tab. 1). In supplementary material we present visual comparison of results obtained by original and shifted versions of RAVE, where it can be clearly seen that RAVE shifted produces more well-lit results, while keeping all the advantages of RAVE. There we also present analysis of tokens with the most and least similar embeddings to the updated $\\mathbf{v}_{\\mathrm{residual}}$ . This analysis supports the claim about the meaning of updated $\\mathbf{v}_{\\mathrm{residual}}$ .", + "bbox": [ + 212, + 643, + 787, + 779 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Necessity of the residual vector Given the interpretation of $\\mathbf{v}_{\\mathrm{residual}}$ , we investigate whether it is necessary to compute $\\mathbf{v}_{\\mathrm{residual}}$ using the training data, or is it enough to use CLIP latent vectors of words such as \"dark\", \"darkness\" as a guidance vector for enhancement model. We trained RAVE using $v_{\\mathrm{residual}} =$", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "RAVE: Residual Vector Embedding", + "bbox": [ + 491, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/289014402d3f8e32cb70d26cd6ba198347b2ca0f94013970f37918f6cd011a6a.jpg", + "table_caption": [ + "Table 3: Mean similarity of CLIP latent vectors of well-lit and back-lit images from BAID and DIV2K datasets to the CLIP latent vectors of vocabulary tokens" + ], + "table_footnote": [], + "table_body": "
TokenBAID backlitBAID well-litDIV2K well-lit
silhouette0.2540.2270.223
dark0.2470.2220.232
darkness0.2420.2150.227
", + "bbox": [ + 290, + 184, + 710, + 252 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "$\\varPhi_{\\mathrm{text}}(token)$ as guidance vector, with token $\\in$ {\"silhouette\", \"dark\", \"darkness\"} but did not observe any meaningful enhancement in the resulting images. We suspect that this is due to the fact that latent vectors of backlit and well-lit images have similar cosine similarity to embeddings of these tokens, as presented in Tab 3. This suggests that residual vector $\\mathbf{v}_{\\mathrm{residual}}$ is not simply a combination of latent vectors of some tokens, but it contains more specific information with the meaning of shifting from backlit to well-lit data.", + "bbox": [ + 212, + 277, + 795, + 383 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.2 Limitations", + "text_level": 1, + "bbox": [ + 215, + 402, + 356, + 417 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Despite showing great performance in the task of backlit image enhancement, our models have limitation of non-optimal enhancement of the regions where the loss of information occurs, making restoration ambiguous. Our approach has no generative components and cannot regenerate highly-accurate textures when the information is lost. This leads to sub-optimal performance in such tasks as low-light image enhancement, where images are highly under-exposed (see supplementary material). Moreover, our models are not trained to exactly match pixels of resulting images with ground-truth data, so some sub-optimal lightening of pixels might occur, which results in non-ideal visual quality.", + "bbox": [ + 212, + 426, + 785, + 561 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "However, existing generative-based models also have limitations. Many of them cannot be trained with unpaired data setting, and require much time and memory to be trained and used. They may also hallucinate detail. As a consequence, the resolution of images it can work with is limited.", + "bbox": [ + 212, + 563, + 785, + 623 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 643, + 359, + 659 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Based on CLIP-LIT method, we have presented two improved CLIP guidance techniques for backlit image enhancement task, named CLIP-LIT-Latent and RAVE. Our methods show similar or superior performance to that of CLIP-LIT in both cases of training with paired or unpaired data, while having lower time and space requirements. Furthermore, latent vector constructed by RAVE is interpretable and can be used to find biases in the training data, as well as potentially correct them. Moreover, our proposed approach has several advantages over the current SOTA in backlit image enhancement, Diff-Retinex. First, unlike Diff-Retinex, our method can also be trained using unpaired data. Second, it is more lightweight, requiring less compute and space for training and inference, suitable for images with larger resolution.", + "bbox": [ + 212, + 672, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "T. Gaintseva et al.", + "bbox": [ + 271, + 114, + 395, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Funding for this research was provided by a Google-DeepMind PhD Studentship, and work utilised Queen Mary's Andrena HPC facility, supported by QMUL Research-IT. We also sincerely thank Laida Kushnareva for assisting with images and text composition in the paper.", + "bbox": [ + 212, + 180, + 787, + 242 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 268, + 321, + 285 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Agustsson, E., Timofte, R.: Ntire 2017 challenge on single image super-resolution: Dataset and study. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 1122-1131 (2017). https://doi.org/10.1109/CVPRW.2017.150 10", + "2. Baykal, A.C., Anees, A.B., Ceylan, D., Erdem, E., Erdem, A., Yuret, D.: Clip-guided stylegan inversion for text-driven real image editing (2023) 4", + "3. Cai, Y., Bian, H., Lin, J., Wang, H., Timofte, R., Zhang, Y.: Retinexformer: One-stage retinax-based transformer for low-light image enhancement. In: 2023 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 12470-12479 (2023). https://doi.org/10.1109/ICCV51070.2023.011492, 3, 4, 10", + "4. Catley-Chandar, S., Tanay, T., Vandroux, L., Leonardis, A., Slabaugh, G., Pérez-Pellitero, E.: Flexhdr: Modeling alignment and exposure uncertainties for flexible hdr imaging. IEEE Transactions on Image Processing 31, 5923-5935 (2022). https://doi.org/10.1109/TIP.2022.3203562 4", + "5. Cheng, H.D., Shi, X.: A simple and effective histogram equalization approach to image enhancement. Digital signal processing 14(2), 158-170 (2004) 3", + "6. Gal, R., Patashnik, O., Maron, H., Bermano, A.H., Chechik, G., Cohen-Or, D.: Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Trans. Graph. 41(4) (jul 2022). https://doi.org/10.1145/3528223.3530164, https://doi.org/10.1145/3528223.3530164 4", + "7. Goloujeh, A.M., Smith, J., Magerko, B.: Explainable clip-guided 3d-scene generation in an ai holodeck. In: Proceedings of the Eighteenth AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment. AIIDE'22, AAAI Press (2022). https://doi.org/10.1609/aide.v18i1.21973, https://doi.org/10.1609/aide.v18i1.21973 4", + "8. Guo, C., Li, C., Guo, J., Loy, C.C., Hou, J., Kwong, S., Cong, R.: Zero-reference deep curve estimation for low-light image enhancement. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1777-1786 (2020). https://doi.org/10.1109/CVPR42600.2020.001852, 3", + "9. Horé, A., Ziou, D.: Image quality metrics: Psnr vs. sSIM. In: 2010 20th International Conference on Pattern Recognition. pp. 2366-2369 (2010). https://doi.org/10.1109/ICPR.2010.579.10", + "0. Land, E.H.: The retina theory of color vision. Scientific american 237(6), 108-129 (1977) 2, 3, 4", + "1. Li, C., Guo, C., Han, L., Jiang, J., Cheng, M., Gu, J., Loy, C.: Low-light image and video enhancement using deep learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(12), 9396-9416 (dec 2022). https://doi.org/10.1109/TPAMI.2021.3126387 2" + ], + "bbox": [ + 225, + 304, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "RAVE: Residual Vector Embedding", + "bbox": [ + 491, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "12. Li, C., Guo, C., Loy, C.C.: Learning to enhance low-light image via zero-reference deep curve estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(8), 4225-4238 (2022). https://doi.org/10.1109/TPAMI.2021.3063604 2, 3", + "13. Li, Z., Wu, X.: Learning-based restoration of backlit images. IEEE Transactions on Image Processing 27(2), 976-986 (2018). https://doi.org/10.1109/TIP.2017.2771142 4", + "14. Liang, Z., Li, C., Zhou, S., Feng, R., Loy, C.C.: Iterative prompt learning for unsupervised backlit image enhancement. In: 2023 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 8060-8069 (2023). https://doi.org/10.1109/ICCV51070.2023.00743 2, 4, 10", + "15. Lv, X., Zhang, S., Liu, Q., Xie, H., Zhong, B., Zhou, H.: Backlitnet: A dataset and network for backlit image enhancement. Computer Vision and Image Understanding 218, 103403 (2022). https://doi.org/https://doi.org/10.1016/j.cviu.2022.103403, https://www.sciencedirect.com/science/article/pii/S1077314222000340 4, 10", + "16. Naeem, M.F., Oh, S.J., Uh, Y., Choi, Y., Yoo, J.: Reliable fidelity and diversity metrics for generative models. In: III, H.D., Singh, A. (eds.) Proceedings of the 37th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 119, pp. 7176-7185. PMLR (13-18 Jul 2020), https://proceedings.mlr.press/v119/naeem20a.html 10", + "17. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models (12 2021) 4", + "18. Parelli, M., Delitzas, A., Hars, N., Vlassis, G., Anagnostidis, S., Bachmann, G., Hofmann, T.: Clip-guided vision-language pre-training for question answering in 3d scenes. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 5607-5612 (2023). https://doi.org/10.1109/CVPRW59228.2023.005934", + "19. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (2021), https://apisemantic scholar.org/CorpusID: 231591445 2, 4, 6", + "20. Reinhard, E.: High Dynamic Range Imaging, pp. 1-6. Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-03243-2_843-1, https://doi.org/10.1007/978-3-030-03243-2_843-1 4", + "21. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Navab, N., Hornegger, J., Wells, W.M., Frangi, A.F. (eds.) Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015. pp. 234-241. Springer International Publishing, Cham (2015) 6, 9", + "22. Singh, K., Seth, A., Sandhu, H.S., Samdani, K.: A comprehensive review of convolutional neural network based image enhancement techniques. In: 2019 IEEE International Conference on System, Computation, Automation and Networking (ICSCAN). pp. 1-6 (2019). https://doi.org/10.1109/ICSCAN.2019.8878706 4", + "23. Wang, Z., Bovik, A., Sheikh, H., Simoncelli, E.: Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing 13(4), 600-612 (2004). https://doi.org/10.1109/TIP.2003.819861 10", + "24. Xia, B., Zhang, Y., Wang, S., Wang, Y., Wu, X., Tian, Y., Yang, W., Gool, L.V.: Diffir: Efficient diffusion model for image restoration. In: 2023 IEEE/CVF" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "T. Gaintseva et al.", + "bbox": [ + 271, + 114, + 395, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "International Conference on Computer Vision (ICCV). pp. 13049-13059. IEEE Computer Society, Los Alamitos, CA, USA (oct 2023). https://doi.org/10.1109/ICCV51070.2023.01204, https://doi.ieeecomputersociety.org/10.1109/ICCV51070.2023.01204 10", + "25. Yi, X., Xu, H., Zhang, H., Tang, L., Ma, J.: Diff-retinex: Rethinking low-light image enhancement with a generative diffusion model. In: 2023 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 12268-12277 (2023). https://doi.org/10.1109/ICCV51070.2023.011304, 10", + "26. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.: Restormer: Efficient transformer for high-resolution image restoration. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5718-5729 (2022). https://doi.org/10.1109/CVPR52688.2022.005642, 3, 4, 10", + "27. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 586-595 (2018). https://doi.org/10.1109/CVPR.2018.0006810", + "28. Zheng, B., Pan, X., Zhang, H., Zhou, X., Slabaugh, G., Yan, C., Yuan, S.: Domainplus: Cross transform domain learning towards high dynamic range imaging. In: Proceedings of the 30th ACM International Conference on Multimedia. p. 1954-1963. MM '22, Association for Computing Machinery, New York, NY, USA (2022). https://doi.org/10.1145/3503161.3547823, https://doi.org/10.1145/3503161.3547823 4" + ], + "bbox": [ + 215, + 146, + 785, + 450 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "RAVE: Residual Vector Embedding", + "bbox": [ + 491, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/RAVE_ Residual Vector Embedding for CLIP-Guided Backlit Image Enhancement/7f0c9e4e-7b51-4d42-aac8-de63b12e0ec8_model.json b/2024/RAVE_ Residual Vector Embedding for CLIP-Guided Backlit Image Enhancement/7f0c9e4e-7b51-4d42-aac8-de63b12e0ec8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..851879e99a3b61954d34b640e6f8342919e82ade --- /dev/null +++ b/2024/RAVE_ Residual Vector Embedding for CLIP-Guided Backlit Image Enhancement/7f0c9e4e-7b51-4d42-aac8-de63b12e0ec8_model.json @@ -0,0 +1,2434 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.26, + 0.141, + 0.744, + 0.187 + ], + "angle": 0, + "content": "RAVE: Residual Vector Embedding for CLIP-Guided Backlit Image Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.249, + 0.212, + 0.756, + 0.229 + ], + "angle": 0, + "content": "Tatiana Gaintseva1,2, Martin Benning3, and Gregory Slabaugh1,2" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.239, + 0.778, + 0.253 + ], + "angle": 0, + "content": "\\(^{1}\\) Digital Environment Research Institute, Queen Mary University of London, UK" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.253, + 0.785, + 0.281 + ], + "angle": 0, + "content": "\\(^{2}\\) School of Electronic Engineering and Computer Science, Queen Mary University of London, UK {t.gaintseva, g.slabaugh}@qmul.ac.uk" + }, + { + "type": "text", + "bbox": [ + 0.273, + 0.281, + 0.729, + 0.309 + ], + "angle": 0, + "content": "\\(^{3}\\) Department of Computer Science, University College London, UK \nmartin.benning@ucl.ac.uk" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.347, + 0.744, + 0.667 + ], + "angle": 0, + "content": "Abstract. In this paper we propose a novel modification of Contrastive Language-Image Pre-Training (CLIP) guidance for the task of backlit image enhancement. Our work builds on the state-of-the-art CLIP-LIT approach, which learns a prompt pair by constraining the text-image similarity between a prompt (negative/positive sample) and a corresponding image (backlit image/well-lit image) in the CLIP embedding space. Learned prompts then guide an image enhancement network. Based on the CLIP-LIT framework, we propose two novel methods for CLIP guidance. First, we show that instead of tuning prompts in the space of text embeddings, it is possible to directly tune their embeddings in the latent space without any loss in quality. This accelerates training and potentially enables the use of additional encoders that do not have a text encoder. Second, we propose a novel approach that does not require any prompt tuning. Instead, based on CLIP embeddings of backlit and well-lit images from training data, we compute the residual vector in the embedding space as a simple difference between the mean embeddings of the well-lit and backlit images. This vector then guides the enhancement network during training, pushing a backlit image towards the space of well-lit images. This approach further dramatically reduces training time, stabilizes training and produces high quality enhanced images without artifacts. Additionally, we show that residual vectors can be interpreted, revealing biases in training data, and thereby enabling potential bias correction. Code is available at https://github.com/Atmyre/RAVE" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.679, + 0.75, + 0.708 + ], + "angle": 0, + "content": "Keywords: Backlit image enhancement \\(\\cdot\\) Vision-language models \\(\\cdot\\) Residual vector embedding" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.733, + 0.377, + 0.75 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Backlit image enhancement aims to improve images that suffer from backlighting issues. Backlighting is a phenomenon when the light source is situated behind the photographed object, often resulting in a loss of detail and contrast of some areas due to underexposure, diminishing the overall visual quality of the image. The goal of backlit image enhancement is to correct these exposure discrepancies and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.398, + 0.128 + ], + "angle": 0, + "content": "T. Gaintseva et al." + }, + { + "type": "image", + "bbox": [ + 0.25, + 0.154, + 0.495, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.154, + 0.744, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.472, + 0.246, + 0.521, + 0.257 + ], + "angle": 0, + "content": "CLIP-LIT" + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.263, + 0.493, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.263, + 0.744, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.479, + 0.363, + 0.512, + 0.375 + ], + "angle": 0, + "content": "RAVE" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.39, + 0.789, + 0.447 + ], + "angle": 0, + "content": "Fig. 1: Visual comparison of results obtained by RAVE vs CLIP-LIT. RAVE produces well-lit images with fewer artifacts (note the dark green color of underexposed regions in CLIP-LIT), as shown in the zoomed-in sections. Further comparisons can be found in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.789, + 0.538 + ], + "angle": 0, + "content": "bring the details in the darkened areas out without overexposing the well-exposed parts of the image. Backlit image enhancement significantly improves photograph quality in uncontrolled lighting conditions. It is also vital for fields requiring clear image visibility, such as surveillance, security, and scientific imaging applications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.789, + 0.704 + ], + "angle": 0, + "content": "However, correcting backlit images is not an easy task. Manual correction requires skill using photo enhancement software, and often substantial time and effort. Automated solutions are preferred, but also face challenges. Global adjustment of the brightness level is not sufficient, as this typically overexposes regions of the image which were well lit prior to enhancement [11]. Spatially adaptive methods have appeared in the literature including early methods that rely on assumptions such as average luminance and a gray world model [8, 12] or Retinex theory [10]. More recent methods approximate well-lit image distributions via end-to-end training of deep networks [3, 26]. Despite progress in the field, considerable room for improvement exists for high quality automated backlit image enhancement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.79, + 0.842 + ], + "angle": 0, + "content": "Another challenge to high quality automatic backlit image enhancement is the lack of paired data, i.e. backlit images each with a corresponding well-lit image. It is complicated to collect such data, so approaches that work with unpaired data are imperative. Recently, CLIP-LIT [14] was proposed, which successfully utilizes CLIP [19] model guidance for training a backlit enhancement model with unpaired data. CLIP-LIT uses prompt learning techniques to provide CLIP guidance. More specifically, it constructs two learnable text prompts, which are trained to have CLIP embeddings close to the well-lit and backlit images, respectively. These learnt prompts then guide the image enhancement model" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.493, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RAVE: Residual Vector Embedding" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "during training. Next, several iterations of additional prompt correction followed by enhancement model fine-tuning using the updated prompts are required to produce the final image enhancement model. This approach is the current state-of-the-art in backlit image enhancement with training on unpaired data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.787, + 0.479 + ], + "angle": 0, + "content": "In this work, we demonstrate that prompt training is not the most efficient way to implement the CLIP guidance, and propose two novel methods, named CLIP-LIT-Latent and ResiduAl Vector Embedding (RAVE). In CLIP-LIT-Latent, we train vectors corresponding to well-lit and backlit images directly in the CLIP latent space rather than in text embedding space. We show that this approach has similar or superior performance when trained on paired or unpaired image data. Further, in RAVE, we form the guidance vector by subtracting the mean of CLIP latent vectors of backlit images from the mean of CLIP latent vectors of well-lit images. This residual vector points in a direction of moving backlit images to well-lit images in the CLIP embedding space. This vector is then used as guidance for the image enhancement model during training. We show that such guidance enables the enhancement model to produce high quality images, while requiring much less training time. While having all these benefits, both proposed CLIP-LIT-Latent and RAVE retain all the advantages of the original CLIP-LIT method, being lightweight for inference and suitable for training on paired or unpaired data. Moreover, we show that residual vector used for guidance in RAVE is interpretable, which opens up new possibilities for overcoming biases in the training data." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.487, + 0.687, + 0.502 + ], + "angle": 0, + "content": "The contributions of this paper can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.513, + 0.784, + 0.558 + ], + "angle": 0, + "content": "1. Inspired by CLIP-LIT, we present two novel approaches to CLIP guidance (CLIP-LIT-Latent and RAVE) working directly in the latent space for the task of backlit image enhancement." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.56, + 0.784, + 0.616 + ], + "angle": 0, + "content": "2. We show that these approaches result in similar or better quality according to quantitative metrics in both settings of training on paired or unpaired data, and superior visual quality of resulting images, while requiring considerably less time to train." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.619, + 0.783, + 0.648 + ], + "angle": 0, + "content": "3. We demonstrate that the embedding used by RAVE for guidance is interpretable, and its interpretation can reveal biases in the training data." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.513, + 0.784, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.673, + 0.383, + 0.689 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Image Enhancement. Image enhancement is the task of image processing with the goal of improving the visual appearance of an image. This is an essential task in numerous fields, such as photography, medical imaging, satellite imagery analysis, and other applications where image quality is crucial. As there are many types of image degradation such as noise, blurriness, low resolution, backlighting, low light, etc., there are a range of image enhancement methods, from traditional ones such as gray level transformation [8,12], histogram equalization [5], and methods based on Retinex theory [10], to machine learning and end-to-end trained deep learning methods [3,26]. A general overview of concepts" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.127 + ], + "angle": 0, + "content": "T. Gaintseva et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.177 + ], + "angle": 0, + "content": "and techniques commonly used for image enhancement is provided by Maini et al. [22]." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.179, + 0.786, + 0.42 + ], + "angle": 0, + "content": "Backlit Image Enhancement. For the task of backlit image enhancement, many approaches have been proposed. Some techniques aim to identify backlit regions and correct them separately from the rest of the image [13]. Methods based on the Retinex [10] theory divide an image into a reflectance image and an illumination map, and both parts are then enhanced and then combined back together to form the resulting enhanced image. Another approach to enhancing backlit images is using end-to-end deep learning approaches, as in Restormer [26] or Retinexformer [3] methods. Combined techniques based on Retinex theory and deep learning were proposed for both backlit and low-light image enhancement, such as Diff-Retinex [25]. These methods decomposing an input image into a reflectance image and an illumination map and apply diffusion methods for enhancement before combining them to produce the output image. Finally, another set of methods are based on high dynamic range (HDR) imaging which combine multiple exposures to produce a composite well-lit image [4, 20, 28]. However, in this work we only focus on single-image methods where back-lit correction is applied as a post-process." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.422, + 0.786, + 0.527 + ], + "angle": 0, + "content": "Fundamentally, backlit image correction is ill-posed, in that for a single backlit image, multiple well-lit variants with high perceptual quality exist. Relatedly, generation of paired training data requires hiring professional photographers, which can be expensive. To help address this problem, recently the BAID [15] dataset was proposed, which contains paired images generated by professional photographers. Nevertheless, development of methods for backlit image enhancement, which can be trained using non-paired data remains essential." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.529, + 0.786, + 0.65 + ], + "angle": 0, + "content": "CLIP guidance. CLIP [19] is a model that consists of two parts: image and text encoders, which learn to project images and text to a common latent embedding space. CLIP has been proven to provide strong image and text priors which are able to provide guidance in many tasks. CLIP guidance has been used in text-to-image generation [17], visual question answering [18], GAN inversion [2], 3D scene generation [7], domain adaptation of image generators [6] and more. Liang et al. [14] also used CLIP guidance for the task of backlit image enhancement. In our work, we aim to improve this technique by modifying the CLIP guidance." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.68, + 0.379, + 0.697 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.718, + 0.785, + 0.793 + ], + "angle": 0, + "content": "Our work is built on top of the CLIP-LIT [14] approach, which uses CLIP guidance for training a backlit image enhancement model. We propose two modifications of CLIP guidance, which reduce training time, stabilize training and produce high quality enhanced images with fewer artifacts compared to CLIP-LIT." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "We start in Section 3.1 by reviewing the architecture of original CLIP-LIT method on which our work is based. Then, in Sections 3.2 and 3.3 we introduce our proposed methods CLIP-LIT-Latent and RAVE." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.493, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "RAVE: Residual Vector Embedding" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.164, + 0.816, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.323, + 0.634, + 0.336 + ], + "angle": 0, + "content": "(a) Initial Prompts and Enhancement Training" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.34, + 0.816, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.513, + 0.643, + 0.527 + ], + "angle": 0, + "content": "(b) Prompt Refinement and Enhancement Tuning" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.537, + 0.788, + 0.607 + ], + "angle": 0, + "content": "Fig. 2: Overview of the original CLIP-LIT approach and proposed CLIP-LIT-Latent. (a) depicts the first stage of training, which consists of prompt or latent vector initialization and the initial training of an enhancement network. (b) shows the second stage, where prompt/latent vector refinement and enhancement model fine-tuning are iteratively repeated. Blue and red boxes are related to CLIP-LIT and CLIP-LIT-Latent." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.648, + 0.424, + 0.661 + ], + "angle": 0, + "content": "3.1 CLIP-LIT overview" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.699, + 0.755, + 0.715 + ], + "angle": 0, + "content": "The CLIP-LIT framework (Fig. 2) consists of two stages, described below." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.735, + 0.576, + 0.749 + ], + "angle": 0, + "content": "Initial Prompts and Enhancement Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In this stage, two prompts \\( T_{p}, T_{n} \\in \\mathbb{R}^{N \\times 512} \\) representing positive and negative prompts are randomly initialized, where \\( N \\) is a number of tokens in each prompt. Then they are tuned so that the CLIP embedding of \\( T_{p} \\) is close in terms of the dot product to the CLIP embeddings of well-lit training images, and CLIP embedding of \\( T_{n} \\) is close to the CLIP embeddings of backlit images from the training data. The binary cross-entropy loss of classifying the backlit and well" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.398, + 0.128 + ], + "angle": 0, + "content": "T. Gaintseva et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.614, + 0.164 + ], + "angle": 0, + "content": "lit images based on \\( T_{p} \\) and \\( T_{n} \\) is used for this purpose:" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.174, + 0.649, + 0.192 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {i n i t i a l}} = - (y \\log (\\widehat {y}) + (1 - y) \\log (1 - \\widehat {y}))\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.194, + 0.786, + 0.233 + ], + "angle": 0, + "content": "\\[\n\\widehat {y} = \\frac {e ^ {\\cos \\left(\\Phi_ {\\text {i m a g e}} (I) , \\Phi_ {\\text {t e x t}} \\left(T _ {p}\\right)\\right)}}{\\sum_ {i \\in \\{n , p \\}} e ^ {\\cos \\left(\\Phi_ {\\text {i m a g e}} (I) , \\Phi_ {\\text {t e x t}} \\left(T _ {i}\\right)\\right)}} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.243, + 0.784, + 0.288 + ], + "angle": 0, + "content": "Here \\( I \\) is a well-lit or backlit input image with label \\( y \\in \\{0,1\\} \\) with 0 corresponding to backlit and 1 corresponding to well-lit; \\( \\varPhi_{\\mathrm{image}} \\) is a CLIP [19] image encoder, \\( \\varPhi_{\\mathrm{text}} \\) is a CLIP text encoder; cos is a cosine similarity function." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.289, + 0.785, + 0.364 + ], + "angle": 0, + "content": "Using these initialized prompts \\( T_{p} \\) and \\( T_{n} \\) it is now possible to train an image enhancement network with a CLIP-guided loss. The enhancement model is a Unet [21] which takes backlit image \\( I_{b} \\) as input and outputs an estimated illumination map \\( I_{i} \\in \\mathbb{R}^{H \\times W \\times 1} \\). The final enhanced image \\( I_{t} \\) is then obtained as \\( I_{t} = I_{b} / I_{i} \\). The image enhancement network is trained by combining two losses:" + }, + { + "type": "equation", + "bbox": [ + 0.4, + 0.38, + 0.786, + 0.396 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {e n h a n c e}} = L _ {\\text {c l i p}} + \\omega L _ {\\text {i d e n t i t y}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.402, + 0.784, + 0.475 + ], + "angle": 0, + "content": "Here \\( L_{\\mathrm{clip}} \\) measures the similarity between the enhanced image and the learned prompts \\( T_{p} \\) and \\( T_{n} \\) in the CLIP latent space. \\( L_{\\mathrm{identity}} \\) is an identity loss designed to ensure that the enhanced image is similar to the original input image in terms of content and structure. The parameter \\( \\omega \\) balances the two losses. These losses are defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.361, + 0.474, + 0.786, + 0.513 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {c l i p}} = \\frac {e ^ {\\cos \\left(\\Phi_ {\\mathrm {i m a g e}} \\left(I _ {t}\\right) , \\Phi_ {\\mathrm {t e x t}} \\left(T _ {n}\\right)\\right)}}{\\sum_ {i \\in \\{n , p \\}} e ^ {\\cos \\left(\\Phi_ {\\mathrm {i m a g e}} \\left(I _ {t}\\right) , \\Phi_ {\\mathrm {t e x t}} \\left(T _ {i}\\right)\\right)}} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.34, + 0.519, + 0.786, + 0.561 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {i d e n t i t y}} = \\sum_ {l = 0} ^ {k} \\alpha_ {l} \\cdot \\left\\| \\Phi_ {\\text {i m a g e}} ^ {l} \\left(I _ {b}\\right) - \\Phi_ {\\text {i m a g e}} ^ {l} \\left(I _ {t}\\right) \\right\\| _ {2} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.568, + 0.784, + 0.599 + ], + "angle": 0, + "content": "where \\(\\Phi^l\\) is \\(l^{th}\\) layer of the CLIP image encoder, and \\(\\alpha_{l}\\) is its corresponding weight. \\(k\\) is number of layers used to calculate the metric." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.614, + 0.6, + 0.629 + ], + "angle": 0, + "content": "Prompt Refinement and Enhancement Tuning" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.629, + 0.784, + 0.674 + ], + "angle": 0, + "content": "CLIP-LIT applies a second stage where several cycles of prompt refinement and enhancement network tuning are performed to improve the accuracy of the learned prompts." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.675, + 0.784, + 0.72 + ], + "angle": 0, + "content": "First, learned prompts \\( T_{p} \\) and \\( T_{n} \\) are fine-tuned to provide better guidance for fine-tuning the image enhancement model. A margin ranking loss is used to update the prompts based on images given by the current enhancement model:" + }, + { + "type": "equation", + "bbox": [ + 0.235, + 0.73, + 0.744, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {\\text {p r o m p r e t}} ^ {1} = \\max (0, S \\left(I _ {w}\\right) - S \\left(I _ {b}\\right) + m _ {0}) + \\max (0, S \\left(I _ {t - 1}\\right) - S \\left(I _ {b}\\right) + m _ {0}) \\\\ + \\max (0, S \\left(I _ {w}\\right) - S \\left(I _ {t}\\right) + m _ {1}) + \\max (0, S \\left(I _ {t}\\right) - S \\left(I _ {t - 1}\\right) + m _ {2}) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.78, + 0.784, + 0.807 + ], + "angle": 0, + "content": "where \\( S(I) \\) is a negative similarity score between the prompt pair and an image defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.362, + 0.806, + 0.786, + 0.845 + ], + "angle": 0, + "content": "\\[\nS (I) = \\frac {e ^ {\\cos \\left(\\Phi_ {\\text {i m a g e}} (I) , \\Phi_ {\\text {t e x t}} \\left(T _ {n}\\right)\\right)}}{\\sum_ {i \\in \\{n , p \\}} e ^ {\\cos \\left(\\Phi_ {\\text {i m a g e}} (I) , \\Phi_ {\\text {t e x t}} \\left(T _ {i}\\right)\\right)}} \\tag {6}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.493, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "RAVE: Residual Vector Embedding" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.221 + ], + "angle": 0, + "content": "Here \\(m_0 \\in [0,1]\\) is the margin between the score of well-lit or enhanced result and the backlit image in the CLIP embedding space; \\(m_1 \\in [0,1]\\) is the margin between the score of the enhanced result and the well-lit image in the CLIP embedding space; \\(m_2 \\in [0,1]\\) is the margin between the newly enhanced result \\(I_t\\) and previously enhanced result \\(I_{t-1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.784, + 0.282 + ], + "angle": 0, + "content": "After that, a stage of enhancement model fine-tuning is performed. This follows the same process as initial enhancement model training described above. The process of prompt refinement and enhancement tuning is repeated several times until satisfactory results are obtained." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.305, + 0.406, + 0.318 + ], + "angle": 0, + "content": "3.2 CLIP-LIT-Latent" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.784, + 0.372 + ], + "angle": 0, + "content": "Like CLIP-LIT, CLIP-LIT-Latent also consists of two stages of initial prompts and enhancement training and prompt refinement and enhancement tuning. Fig. 2 illustrates the CLIP-LIT-Latent framework." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.785, + 0.495 + ], + "angle": 0, + "content": "Instead of learning prompts in the text embedding space, CLIP-LIT-Latent learns a pair of positive/negative vectors directly in the CLIP latent space. In the prompt initialization phase, we randomly initialize two vectors \\( E_{p}, E_{n} \\in \\mathbb{R}^{512} \\). These vectors are then tuned so that \\( E_{p} \\) is close in terms of the dot product to the CLIP embedding of well-lit images from the training data, and \\( E_{n} \\) is close to the CLIP embedding of backlit images. As in original CLIP-LIT approach, binary cross-entropy loss for classifying the backlit and well-lit images based on \\( E_{p} \\) and \\( E_{n} \\) is used for this purpose:" + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.507, + 0.784, + 0.563 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {\\text {i n i t i a l}} = - (y \\log (\\widehat {y}) + (1 - y) \\log (1 - \\widehat {y})) \\\\ \\widehat {y} = \\frac {e ^ {\\cos \\left(\\Phi_ {\\mathrm {i m a g e}} (I) , E _ {p}\\right)}}{\\sum_ {i \\in \\{n , p \\}} e ^ {\\cos \\left(\\Phi_ {\\mathrm {i m a g e}} (I) , E _ {i}\\right)}} \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.571, + 0.784, + 0.601 + ], + "angle": 0, + "content": "where \\(I\\) is a well-lit or backlit input image with label \\(y \\in 0, 1\\) with 0 corresponding to backlit and 1 corresponding to well-lit; \\(\\varPhi_{\\mathrm{image}}\\) is a CLIP image encoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.602, + 0.785, + 0.678 + ], + "angle": 0, + "content": "Based on these initialized vectors \\( E_{p} \\) and \\( E_{n} \\), an enhancement network is trained with a CLIP-guided loss. As before, the enhancement model is a Unet which takes backlit image \\( I_{b} \\) as input and outputs an estimated illumination map \\( I_{i} \\in \\mathbb{R}^{H \\times W \\times 1} \\). And, as before, it is trained using \\( L_{\\text{enhance}} \\) loss, where the \\( L_{\\text{clip}} \\) component is changed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.383, + 0.689, + 0.784, + 0.726 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {c l i p}} = \\frac {e ^ {\\cos \\left(\\Phi_ {\\text {i m a g e}} \\left(I _ {r}\\right) , E _ {n}\\right)}}{\\sum_ {i \\in \\{n , p \\}} e ^ {\\cos \\left(\\Phi_ {\\text {i m a g e}} \\left(I _ {t}\\right) , E _ {i}\\right)}} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.783, + 0.751 + ], + "angle": 0, + "content": "where \\(\\varPhi^l\\) is \\(l^{th}\\) layer of CLIP image encoder, and \\(a_{l}\\) is its corresponding weight." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.752, + 0.784, + 0.796 + ], + "angle": 0, + "content": "In the second stage of prompt refinement and enhancement tuning, where learned vectors \\( E_{p} \\) and \\( E_{n} \\) are fine-tuned using margin ranking loss, the formula for \\( S(I) \\) changes as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.385, + 0.807, + 0.784, + 0.844 + ], + "angle": 0, + "content": "\\[\nS (I) = \\frac {e ^ {\\cos \\left(\\Phi_ {\\text {i m a g e}} (I) , E _ {n}\\right)}}{\\sum_ {i \\in \\{n , p \\}} e ^ {\\cos \\left(\\Phi_ {\\text {i m a g e}} (I) , E _ {i}\\right)}} \\tag {9}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.398, + 0.128 + ], + "angle": 0, + "content": "T. Gaintseva et al." + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.154, + 0.516, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.238, + 0.324, + 0.452, + 0.338 + ], + "angle": 0, + "content": "- CLIP vectors of backlit training images" + }, + { + "type": "image_footnote", + "bbox": [ + 0.238, + 0.346, + 0.452, + 0.359 + ], + "angle": 0, + "content": "- CLIP vectors of well-lit training images" + }, + { + "type": "list", + "bbox": [ + 0.238, + 0.324, + 0.452, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.248, + 0.787, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.216, + 0.367, + 0.79, + 0.39 + ], + "angle": 0, + "content": "(a) Calculation of residual vector \\(\\mathbf{v}_{\\text{residual}}\\) based on (b) Training of enhancement model using paired or unpaired training data guidance from \\(\\mathbf{v}_{\\text{residual}}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.4, + 0.789, + 0.442 + ], + "angle": 0, + "content": "Fig. 3: Overview of the RAVE model. (a) First, we calculate residual vector \\(\\mathbf{v}_{\\mathrm{residual}}\\) based on backlit and well-lit training data. (b) Then we switch to enhancement model training based on the identity loss and the loss based on the residual vector." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.475, + 0.747, + 0.491 + ], + "angle": 0, + "content": "In all other aspects the training algorithm is similar to that of CLIP-LIT." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.789, + 0.685 + ], + "angle": 0, + "content": "CLIP-LIT-Latent has several advantages over CLIP-LIT. First, CLIP-LIT-Latent does not use the text encoder, which speeds up the training and inference processes, as it is no longer necessary to pass vectors and gradients through the text encoder. This also potentially enables the use of other vision models which do not have text encoder parts rather than CLIP for guidance. Second, images obtained from CLIP-LIT-Latent tend to have more contrast and have better visual quality (see Fig. 4 and Fig. 1). We speculate this results from tuning vectors \\( E_{p} \\) and \\( E_{n} \\) directly in the CLIP latent space. In contrast, in the original CLIP-LIT, vectors \\( T_{p} \\) and \\( T_{n} \\) belong to the text embedding space, and must be tuned so that their projection through the CLIP text encoder satisfies desired conditions. Such optimization is more complex and makes tuning of \\( T_{p} \\) and \\( T_{n} \\) harder and less accurate." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.715, + 0.559, + 0.731 + ], + "angle": 0, + "content": "3.3 RAVE: Residual Vector Embedding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.748, + 0.788, + 0.809 + ], + "angle": 0, + "content": "In this section we present our model RAVE (ResiduAl Vector Embedding). RAVE includes a further modification of the CLIP guidance. We show that RAVE does not require iterative stages of prompt and model updates. Instead it only requires one stage of enhancement model training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "In RAVE we exploit arithmetic defined in the CLIP latent space. Using well-lit and backlit training data, we construct a residual vector \\(\\mathbf{v}_{\\mathrm{residual}}\\), which will" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.493, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "RAVE: Residual Vector Embedding" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.767, + 0.163 + ], + "angle": 0, + "content": "then be used for enhancement model guidance. \\(\\mathbf{v}_{\\mathrm{residual}}\\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.346, + 0.17, + 0.594, + 0.185 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {\\text {r e s i d u a l}} = f _ {\\text {n o r m}} \\left(\\mathbf {v} _ {\\text {w e l l - l i t}} - \\mathbf {v} _ {\\text {b a c k l i t}}\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.346, + 0.189, + 0.785, + 0.229 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {\\text {w e l l - l i t}} = f _ {\\text {n o r m}} \\left(\\frac {\\sum_ {i = 1} ^ {N _ {p}} f _ {\\text {n o r m}} \\left(\\Phi_ {\\text {i m a g e}} \\left(I _ {i} ^ {p}\\right)\\right)}{N _ {p}}\\right) \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.346, + 0.232, + 0.656, + 0.272 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {\\text {b a c k l i t}} = f _ {n o r m} \\left(\\frac {\\sum_ {i = 1} ^ {N _ {n}} f _ {\\text {n o r m}} \\left(\\Phi_ {\\text {i m a g e}} \\left(I _ {i} ^ {n}\\right)\\right)}{N _ {n}}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.278, + 0.787, + 0.326 + ], + "angle": 0, + "content": "where \\( f_{\\mathrm{norm}} \\) is a normalization function: \\( f_{\\mathrm{norm}}(\\mathbf{v}) = \\frac{\\mathbf{v}}{\\|\\mathbf{v}\\|_2} \\), \\( N_p \\) is the number of training well-lit images, \\( N_n \\) is the number of training backlit images, \\( I_i^p \\) is the \\( i^{th} \\) well-lit training image, \\( I_i^n \\) is the \\( i^{th} \\) backlit training image." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.327, + 0.784, + 0.416 + ], + "angle": 0, + "content": "\\(\\mathbf{v}_{\\mathrm{residual}}\\) is a vector that points in a direction moving from backlit images to well-lit images in the CLIP embedding space. We then use this vector as guidance for the image enhancement model during training. This will train the image enhancement model to produce images with CLIP latent vectors that are close to the CLIP latent vectors of well-lit training images. Overview of the RAVE approach is presented on Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.417, + 0.784, + 0.446 + ], + "angle": 0, + "content": "We use the same Unet [21] as enhancement model as in CLIP-LIT and CLIP-LIT-Latent. The enhancement model is trained using the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.403, + 0.463, + 0.785, + 0.478 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {r a v e}} = L _ {\\text {i d e n t i t y}} + \\omega L _ {\\text {r e s i d u a l}} \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.496, + 0.785, + 0.513 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {r e s i d u a l}} = \\left\\| \\Phi_ {\\text {i m a g e}} (I) \\cdot \\mathbf {v} _ {\\text {r e s i d u a l}} - \\mathbf {v} _ {\\text {w e l l - l i t}} \\cdot \\mathbf {v} _ {\\text {r e s i d u a l}} \\right\\| ^ {2} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.517, + 0.784, + 0.547 + ], + "angle": 0, + "content": "where \\(I\\) is an input image, and \\(v_{\\mathrm{well - lit}}\\) is a normalized mean of CLIP embeddings of well-lit training images:" + }, + { + "type": "equation", + "bbox": [ + 0.348, + 0.554, + 0.785, + 0.593 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {\\text {w e l l - l i t}} = f _ {\\text {n o r m}} \\left(\\frac {\\sum_ {i = 1} ^ {N _ {p}} f _ {\\text {n o r m}} \\left(\\Phi_ {\\text {i m a g e}} \\left(I _ {i} ^ {p}\\right)\\right)}{N _ {p}}\\right) \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.599, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Note that RAVE does not require several repeating stages during training. Before training, the residual vector \\(\\mathbf{v}_{\\mathrm{residual}}\\) is calculated, and afterwards the training of the image enhancement model begins using \\(L_{rave}\\) as the loss function. This makes RAVE training much more efficient than that of CLIP-LIT and CLIP-LIT-Latent. First, the training of the image enhancement model no longer depends on the quality of learned prompt embeddings, as a fixed residual vector is used for guidance. Also, the iterative manner of CLIP-LIT and CLIP-LIT-Latent training regimes makes training less stable, as the image enhancement model training depends on the quality of prompts obtained on the previous step and vice versa. In supplementary material we show that results obtained on intermediate steps of CLIP-LIT training can be of poor quality and can fluctuate from producing under-exposed to over-exposed resulting images from step to step. In contrast, RAVE has a well-defined fixed initial objective. Results show that RAVE converges up to 25 times faster than CLIP-LIT and CLIP-LIT-Latent. Furthermore, images obtained by RAVE have fewer over-exposed areas, producing high quality results." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.127 + ], + "angle": 0, + "content": "T. Gaintseva et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.376, + 0.163 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.787, + 0.296 + ], + "angle": 0, + "content": "Datasets Following CLIP-LIT, we use the BAID [15] and DIV2K [1] datasets for training. We train our models in two settings: using paired and unpaired training data. For the paired setting, we use 380 backlit and corresponding well-lit images from BAID dataset. For the unpaired setting, we use 380 backlit images from the BAID and 384 images from DIV2K dataset. For the unpaired setting, we use the same data as used in CLIP-LIT. For testing, we follow recent literature and use the BAID [15] test dataset, which has 368 backlit and corresponding well-lit images. The data is publicly available." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.297, + 0.787, + 0.357 + ], + "angle": 0, + "content": "Training and inference We follow parameter settings of the original CLIP-LIT approach when setting parameters for CLIP-LIT-Latent and RAVE (where applicable), as well as for reproducing CLIP-LIT. Specific details of training and values of hyperparameters are presented in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.786, + 0.388 + ], + "angle": 0, + "content": "Metrics Following recent works in image enhancement, we report four metrics: PSNR [9], SSIM [23], and LPIPS [27] (AlexNet version) and FID [16]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.787, + 0.464 + ], + "angle": 0, + "content": "Methods for comparison The main model which we compare our approaches to is CLIP-LIT [14], as it showed superior performance on all mentioned metrics, except FID, to many prior image enhancement models. We also include recently published methods such as Restormer [26], Retinexformer [3], DiffIR [24] and Diff-Retinex [25] to the comparison." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.484, + 0.325, + 0.498 + ], + "angle": 0, + "content": "4.1 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.787, + 0.703 + ], + "angle": 0, + "content": "Quantitative comparison The quantitative comparison on the BAID test dataset is presented in Tab. 1. We report test set metrics of the best model checkpoint obtained during training for all the models. Detailed metrics values for all the checkpoints after each training epoch can be found in supplementary material. Note that many approaches can only be trained using paired data, that is why their results are only present in the \"paired\" section. We see that our method achieves state-of-the-art performance in both settings. In supplementary material, we also present results on time needed to train all the models, and confirm that both CLIP-LIT-Latent and RAVE are much more efficient. Note that in the setting of unpaired data we have two versions of RAVE, referred as RAVE and RAVE shifted in Tab. 1. RAVE shifted is a variant of RAVE with a simple modification of residual vector, see \"Interpretation of the residual vector\" section below for more details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Visual comparison For visual comparison, we present representative examples from the BAID test dataset. First, in Fig. 4, we compare CLIP-LIT-Latent and RAVE to original CLIP-LIT. First, note that CLIP-LIT-Latent produces more contrast in general, giving more diverse illumination to different regions. Second, in Fig. 1, we present comparison of images produced by RAVE to images produced by CLIP-LIT. We show that images produced by RAVE have fewer artifacts. We hypothesize that this advantage of RAVE may result from its guiding mechanism: in RAVE we directly point the enhancement model to produce images which have similar CLIP embeddings to that of real well-lit images. This" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.493, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RAVE: Residual Vector Embedding" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison of different methods on the BAID test dataset. The best and second best performances in both settings are in bold and underlined." + }, + { + "type": "table", + "bbox": [ + 0.224, + 0.185, + 0.78, + 0.391 + ], + "angle": 0, + "content": "
MethodPSNR ↑SSIM ↑LPIPS ↓FID ↓
Original images16.910.760.2152.47
PairedRestormer21.070.8320.19241.17
Retinexformer22.030.8620.17345.27
DiffIR21.100.8350.17540.35
Diff-Retinex22.070.8610.16038.07
CLIP-LIT21.930.8750.15641.16
CLIP-LIT-Latent (ours)21.840.8770.15542.56
RAVE (ours)22.260.8800.13936.01
UnpairedEnlightenGAN17.550.8640.19643.50
CLIP-LIT21.590.8740.16046.49
CLIP-LIT-Latent (ours)21.450.8770.15646.20
RAVE (ours)20.390.8610.15540.11
RAVE shifted (ours)21.370.8770.14940.90
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.785, + 0.54 + ], + "angle": 0, + "content": "does not allow model to produce regions with artifacts, as its CLIP embeddings won't match those of real well-lit images. In CLIP-LIT or CLIP-LIT-Latent, on the other hand, the objective of the enhancement model is to produce images for which the CLIP embeddings are simply closer to the positive prompt/vector than to the negative one. This allows the enhancement model to produce less accurate images, CLIP embeddings of which are just \"close enough\" to the CLIP embeddings of real well-lit images. More results of visual comparison of all the models can be found in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.54, + 0.785, + 0.616 + ], + "angle": 0, + "content": "Interpretation of the residual vector In RAVE, we build the residual vector \\(\\mathbf{v}_{\\text{residual}}\\) following formula 16. We noted that \\(\\mathbf{v}_{\\text{residual}}\\) is a vector that points to a direction of moving from backlit images to well-lit images in the CLIP embedding space. Here we prove this by interpreting vector \\(\\mathbf{v}_{\\text{residual}}\\) using the fact that CLIP has a common vector space for embeddings of images and text." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.616, + 0.785, + 0.676 + ], + "angle": 0, + "content": "We go through the vocabulary of the CLIP text encoder and find those vocabulary tokens which have the closest and farthest CLIP embeddings to the \\(\\mathbf{v}_{\\text{residual}}\\) in the CLIP latent space in terms of cosine similarity. More formally, we use the following metric:" + }, + { + "type": "equation", + "bbox": [ + 0.346, + 0.69, + 0.785, + 0.707 + ], + "angle": 0, + "content": "\\[\n\\operatorname {s i m} (\\text {t o k e n}) = f _ {\\text {n o r m}} \\left(\\Phi_ {\\text {t e x t}} \\left(e _ {\\text {t o k e n}}\\right)\\right) \\cdot \\mathbf {v} _ {\\text {r e s i d u a l}} \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.785, + 0.78 + ], + "angle": 0, + "content": "Here \\( e_{token} \\) is embedding of vocabulary token, which is input to the CLIP text encoder \\( \\varPhi_{text} \\). We sort vocabulary tokens using sim. Table 2 shows 7 vocabulary tokens which have highest and lowest cosine similarity to the \\( v_{residual} \\) vectors obtained from the paired dataset and from the unpaired dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In the case of paired data, many of the least similar tokens have a meaning around \"darkness\": \"silhouette\", \"dark\", \"shadow\", etc., and absolute value of their cosine similarity to \\( v_{\\mathrm{residual}} \\) is around \\( 0.12 - 0.15 \\). At the same time, most similar tokens do not share any meaningful semantics, and the absolute value" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.127 + ], + "angle": 0, + "content": "T. Gaintseva et al." + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.148, + 0.422, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.426, + 0.148, + 0.571, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.148, + 0.719, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.228, + 0.422, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.426, + 0.228, + 0.571, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.228, + 0.719, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.311, + 0.362, + 0.321 + ], + "angle": 0, + "content": "CLIP-LIT" + }, + { + "type": "image_caption", + "bbox": [ + 0.462, + 0.311, + 0.538, + 0.321 + ], + "angle": 0, + "content": "CLIP-LIT-Latent" + }, + { + "type": "image_caption", + "bbox": [ + 0.622, + 0.311, + 0.652, + 0.321 + ], + "angle": 0, + "content": "RAVE" + }, + { + "type": "image_footnote", + "bbox": [ + 0.216, + 0.325, + 0.777, + 0.347 + ], + "angle": 0, + "content": "(a) Visual comparison of results by CLIP-LIT, CLIP-LIT-Latent and RAVE, all trained with paired data. RAVE produces results with more general contrast." + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.353, + 0.423, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.426, + 0.353, + 0.571, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.353, + 0.718, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.281, + 0.433, + 0.423, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.428, + 0.433, + 0.571, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.433, + 0.717, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.514, + 0.362, + 0.524 + ], + "angle": 0, + "content": "CLIP-LIT" + }, + { + "type": "image_caption", + "bbox": [ + 0.463, + 0.514, + 0.538, + 0.524 + ], + "angle": 0, + "content": "CLIP-LIT-Latent" + }, + { + "type": "image_caption", + "bbox": [ + 0.622, + 0.514, + 0.652, + 0.524 + ], + "angle": 0, + "content": "RAVE" + }, + { + "type": "image_footnote", + "bbox": [ + 0.216, + 0.529, + 0.777, + 0.551 + ], + "angle": 0, + "content": "(b) Visual comparison of results obtained by CLIP-LIT, CLIP-LIT-Latent and RAVE (shifted version), all trained with unpaired data. RAVE produces results with more general contrast." + }, + { + "type": "image_caption", + "bbox": [ + 0.228, + 0.562, + 0.768, + 0.576 + ], + "angle": 0, + "content": "Fig. 4: Visual comparison of results by CLIP-LIT, CLIP-LIT-Latent and RAVE." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.841 + ], + "angle": 0, + "content": "of their cosine similarity to \\( v_{\\mathrm{residual}} \\) is much smaller, being around \\( 0.01 - 0.04 \\). This supports the hypothesis that \\( v_{\\mathrm{residual}} \\) has a meaning of shifting from \"dark\", \"shadowy\" images to \"normal\" images, having average level of lightness. \\( v_{\\mathrm{residual}} \\) does not carry any other semantic meaning, as it was calculated using paired dataset, where for each backlit image there is its well-lit analogue, and all the semantics present in the BAID dataset is cancelled as a result of subtraction of the mean well-lit and backlit vectors when constructing \\( v_{\\mathrm{residual}} \\). On the other hand, in the case of unpaired data, we see that the least similar tokens have a meaning around \"asian\": \"busan\", \"beijing\", etc., and the most similar ones have a meaning around \"nature\": \"wildlife\", \"southafrica\", \"countrylife\", etc. Herein, an absolute value of cosine similarities for both sets of tokens if quite high, being around \\( 0.08 - 0.12 \\). This reflects the bias in the unpaired datasets. Indeed, in the BAID dataset there are mostly images of Chinese cities or people, while in DIV2K dataset there are many pictures of nature. Due to the bias, RAVE training with unpaired data becomes less effective, making the enhancement model" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.493, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RAVE: Residual Vector Embedding" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 2: Vocabulary tokens and their cosine similarities, which have the lowest and highest cosine similarity to the \\( v_{\\text{residual}} \\) vectors calculated for paired and unpaired training datasets." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.199, + 0.803, + 0.336 + ], + "angle": 0, + "content": "
PairedUnpaired
Lowest similarityHighest similarityLowest similarityHighest similarity
silhouette-0.151healthtech0.035busan-0.108wildlifewednesdayday0.116
darkness-0.149whitepaper0.035beijing-0.103southafrica0.103
dark-0.139theeconomist0.030iu-0.100ecuador0.095
webcamtoy-0.139bona0.029guangzhou-0.095patrol0.095
blackand-0.132digitaltransformation0.027incheon-0.094countryfile0.093
vedere-0.131sua0.021shenzhen-0.094tasmanian0.093
shadow-0.130amarketing0.021eunhyuk-0.091womensart0.091
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.364, + 0.784, + 0.394 + ], + "angle": 0, + "content": "shift images not in an ideal direction in the CLIP latent space. Nevertheless, RAVE still shows good performance when trained with unpaired data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.395, + 0.785, + 0.439 + ], + "angle": 0, + "content": "We also propose a simple modification for correcting the residual vector in the unpaired setting. We simply take \\( n \\) most and least similar tokens to form an additional residual vector as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.465, + 0.785, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {v} _ {\\mathrm {a d d \\_ r e s i d u a l}} = f _ {\\mathrm {n o r m}} (\\mathbf {v} _ {\\mathrm {c l o s e s t}} - \\mathbf {v} _ {\\mathrm {f a r t h e s t}}) \\\\ \\mathbf {v} _ {\\text {c l o s e s t}} = f _ {\\text {n o r m}} \\left(\\frac {\\sum_ {i = 1} ^ {n} f _ {\\text {n o r m}} \\left(\\Phi_ {\\text {t e x t}} \\left(T _ {i} ^ {c}\\right)\\right)}{n}\\right) \\tag {15} \\\\ \\mathbf {v} _ {\\mathrm {f a r t h e s t}} = f _ {n o r m} \\left(\\frac {\\sum_ {i = 1} ^ {n} f _ {\\mathrm {n o r m}} (\\varPhi_ {\\mathrm {t e x t}} (T _ {i} ^ {f}))}{n}\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.572, + 0.784, + 0.604 + ], + "angle": 0, + "content": "\\(T_{i}^{c}\\) and \\(T_{i}^{f}\\) are tokens with \\(i^{th}\\) most and least similar embeddings to the \\(\\mathbf{v}_{\\mathrm{residual}}\\). And finally we update the \\(\\mathbf{v}_{\\mathrm{residual}}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.285, + 0.617, + 0.785, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {\\text {r e s i d u a l}} = \\mathbf {v} _ {\\text {r e s i d u a l}} - c o s \\left(\\mathbf {v} _ {\\text {r e s i d u a l}}, \\mathbf {v} _ {\\text {a d d} \\_ \\text {r e s i d u a l}}\\right) \\cdot \\mathbf {v} _ {\\text {a d d} \\_ \\text {r e s i d u a l}} \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Resulting vector \\(\\mathbf{v}_{\\mathrm{residual}}\\) will keep the meaning of shifting from \"dark\", images to \"normal\" images while losing the unwanted semantics. Experimental results show that RAVE trained with such updated \\(\\mathbf{v}_{\\mathrm{residual}}\\) produces better results (see RAVE shifted in Tab. 1). In supplementary material we present visual comparison of results obtained by original and shifted versions of RAVE, where it can be clearly seen that RAVE shifted produces more well-lit results, while keeping all the advantages of RAVE. There we also present analysis of tokens with the most and least similar embeddings to the updated \\(\\mathbf{v}_{\\mathrm{residual}}\\). This analysis supports the claim about the meaning of updated \\(\\mathbf{v}_{\\mathrm{residual}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Necessity of the residual vector Given the interpretation of \\(\\mathbf{v}_{\\mathrm{residual}}\\), we investigate whether it is necessary to compute \\(\\mathbf{v}_{\\mathrm{residual}}\\) using the training data, or is it enough to use CLIP latent vectors of words such as \"dark\", \"darkness\" as a guidance vector for enhancement model. We trained RAVE using \\(v_{\\mathrm{residual}} =\\)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.128 + ], + "angle": 0, + "content": "T. Gaintseva et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 3: Mean similarity of CLIP latent vectors of well-lit and back-lit images from BAID and DIV2K datasets to the CLIP latent vectors of vocabulary tokens" + }, + { + "type": "table", + "bbox": [ + 0.292, + 0.185, + 0.711, + 0.253 + ], + "angle": 0, + "content": "
TokenBAID backlitBAID well-litDIV2K well-lit
silhouette0.2540.2270.223
dark0.2470.2220.232
darkness0.2420.2150.227
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.279, + 0.796, + 0.384 + ], + "angle": 0, + "content": "\\(\\varPhi_{\\mathrm{text}}(token)\\) as guidance vector, with token \\(\\in\\) {\"silhouette\", \"dark\", \"darkness\"} but did not observe any meaningful enhancement in the resulting images. We suspect that this is due to the fact that latent vectors of backlit and well-lit images have similar cosine similarity to embeddings of these tokens, as presented in Tab 3. This suggests that residual vector \\(\\mathbf{v}_{\\mathrm{residual}}\\) is not simply a combination of latent vectors of some tokens, but it contains more specific information with the meaning of shifting from backlit to well-lit data." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.404, + 0.357, + 0.419 + ], + "angle": 0, + "content": "4.2 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.428, + 0.787, + 0.563 + ], + "angle": 0, + "content": "Despite showing great performance in the task of backlit image enhancement, our models have limitation of non-optimal enhancement of the regions where the loss of information occurs, making restoration ambiguous. Our approach has no generative components and cannot regenerate highly-accurate textures when the information is lost. This leads to sub-optimal performance in such tasks as low-light image enhancement, where images are highly under-exposed (see supplementary material). Moreover, our models are not trained to exactly match pixels of resulting images with ground-truth data, so some sub-optimal lightening of pixels might occur, which results in non-ideal visual quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.564, + 0.787, + 0.624 + ], + "angle": 0, + "content": "However, existing generative-based models also have limitations. Many of them cannot be trained with unpaired data setting, and require much time and memory to be trained and used. They may also hallucinate detail. As a consequence, the resolution of images it can work with is limited." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.645, + 0.36, + 0.66 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Based on CLIP-LIT method, we have presented two improved CLIP guidance techniques for backlit image enhancement task, named CLIP-LIT-Latent and RAVE. Our methods show similar or superior performance to that of CLIP-LIT in both cases of training with paired or unpaired data, while having lower time and space requirements. Furthermore, latent vector constructed by RAVE is interpretable and can be used to find biases in the training data, as well as potentially correct them. Moreover, our proposed approach has several advantages over the current SOTA in backlit image enhancement, Diff-Retinex. First, unlike Diff-Retinex, our method can also be trained using unpaired data. Second, it is more lightweight, requiring less compute and space for training and inference, suitable for images with larger resolution." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.493, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RAVE: Residual Vector Embedding" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.788, + 0.243 + ], + "angle": 0, + "content": "Funding for this research was provided by a Google-DeepMind PhD Studentship, and work utilised Queen Mary's Andrena HPC facility, supported by QMUL Research-IT. We also sincerely thank Laida Kushnareva for assisting with images and text composition in the paper." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.269, + 0.323, + 0.286 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.305, + 0.785, + 0.36 + ], + "angle": 0, + "content": "1. Agustsson, E., Timofte, R.: Ntire 2017 challenge on single image super-resolution: Dataset and study. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 1122-1131 (2017). https://doi.org/10.1109/CVPRW.2017.150 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.362, + 0.785, + 0.39 + ], + "angle": 0, + "content": "2. Baykal, A.C., Anees, A.B., Ceylan, D., Erdem, E., Erdem, A., Yuret, D.: Clip-guided stylegan inversion for text-driven real image editing (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.391, + 0.785, + 0.446 + ], + "angle": 0, + "content": "3. Cai, Y., Bian, H., Lin, J., Wang, H., Timofte, R., Zhang, Y.: Retinexformer: One-stage retinax-based transformer for low-light image enhancement. In: 2023 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 12470-12479 (2023). https://doi.org/10.1109/ICCV51070.2023.011492, 3, 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.447, + 0.785, + 0.502 + ], + "angle": 0, + "content": "4. Catley-Chandar, S., Tanay, T., Vandroux, L., Leonardis, A., Slabaugh, G., Pérez-Pellitero, E.: Flexhdr: Modeling alignment and exposure uncertainties for flexible hdr imaging. IEEE Transactions on Image Processing 31, 5923-5935 (2022). https://doi.org/10.1109/TIP.2022.3203562 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.504, + 0.785, + 0.531 + ], + "angle": 0, + "content": "5. Cheng, H.D., Shi, X.: A simple and effective histogram equalization approach to image enhancement. Digital signal processing 14(2), 158-170 (2004) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.532, + 0.785, + 0.587 + ], + "angle": 0, + "content": "6. Gal, R., Patashnik, O., Maron, H., Bermano, A.H., Chechik, G., Cohen-Or, D.: Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Trans. Graph. 41(4) (jul 2022). https://doi.org/10.1145/3528223.3530164, https://doi.org/10.1145/3528223.3530164 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.588, + 0.785, + 0.655 + ], + "angle": 0, + "content": "7. Goloujeh, A.M., Smith, J., Magerko, B.: Explainable clip-guided 3d-scene generation in an ai holodeck. In: Proceedings of the Eighteenth AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment. AIIDE'22, AAAI Press (2022). https://doi.org/10.1609/aide.v18i1.21973, https://doi.org/10.1609/aide.v18i1.21973 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.657, + 0.785, + 0.713 + ], + "angle": 0, + "content": "8. Guo, C., Li, C., Guo, J., Loy, C.C., Hou, J., Kwong, S., Cong, R.: Zero-reference deep curve estimation for low-light image enhancement. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1777-1786 (2020). https://doi.org/10.1109/CVPR42600.2020.001852, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.714, + 0.785, + 0.755 + ], + "angle": 0, + "content": "9. Horé, A., Ziou, D.: Image quality metrics: Psnr vs. sSIM. In: 2010 20th International Conference on Pattern Recognition. pp. 2366-2369 (2010). https://doi.org/10.1109/ICPR.2010.579.10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.785, + 0.784 + ], + "angle": 0, + "content": "0. Land, E.H.: The retina theory of color vision. Scientific american 237(6), 108-129 (1977) 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "1. Li, C., Guo, C., Han, L., Jiang, J., Cheng, M., Gu, J., Loy, C.: Low-light image and video enhancement using deep learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(12), 9396-9416 (dec 2022). https://doi.org/10.1109/TPAMI.2021.3126387 2" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.305, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.127 + ], + "angle": 0, + "content": "T. Gaintseva et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.202 + ], + "angle": 0, + "content": "12. Li, C., Guo, C., Loy, C.C.: Learning to enhance low-light image via zero-reference deep curve estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(8), 4225-4238 (2022). https://doi.org/10.1109/TPAMI.2021.3063604 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.787, + 0.243 + ], + "angle": 0, + "content": "13. Li, Z., Wu, X.: Learning-based restoration of backlit images. IEEE Transactions on Image Processing 27(2), 976-986 (2018). https://doi.org/10.1109/TIP.2017.2771142 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.787, + 0.299 + ], + "angle": 0, + "content": "14. Liang, Z., Li, C., Zhou, S., Feng, R., Loy, C.C.: Iterative prompt learning for unsupervised backlit image enhancement. In: 2023 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 8060-8069 (2023). https://doi.org/10.1109/ICCV51070.2023.00743 2, 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.369 + ], + "angle": 0, + "content": "15. Lv, X., Zhang, S., Liu, Q., Xie, H., Zhong, B., Zhou, H.: Backlitnet: A dataset and network for backlit image enhancement. Computer Vision and Image Understanding 218, 103403 (2022). https://doi.org/https://doi.org/10.1016/j.cviu.2022.103403, https://www.sciencedirect.com/science/article/pii/S1077314222000340 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.787, + 0.438 + ], + "angle": 0, + "content": "16. Naeem, M.F., Oh, S.J., Uh, Y., Choi, Y., Yoo, J.: Reliable fidelity and diversity metrics for generative models. In: III, H.D., Singh, A. (eds.) Proceedings of the 37th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 119, pp. 7176-7185. PMLR (13-18 Jul 2020), https://proceedings.mlr.press/v119/naeem20a.html 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "17. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models (12 2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.548 + ], + "angle": 0, + "content": "18. Parelli, M., Delitzas, A., Hars, N., Vlassis, G., Anagnostidis, S., Bachmann, G., Hofmann, T.: Clip-guided vision-language pre-training for question answering in 3d scenes. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 5607-5612 (2023). https://doi.org/10.1109/CVPRW59228.2023.005934" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.549, + 0.787, + 0.618 + ], + "angle": 0, + "content": "19. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (2021), https://apisemantic scholar.org/CorpusID: 231591445 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.787, + 0.66 + ], + "angle": 0, + "content": "20. Reinhard, E.: High Dynamic Range Imaging, pp. 1-6. Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-03243-2_843-1, https://doi.org/10.1007/978-3-030-03243-2_843-1 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.787, + 0.715 + ], + "angle": 0, + "content": "21. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Navab, N., Hornegger, J., Wells, W.M., Frangi, A.F. (eds.) Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015. pp. 234-241. Springer International Publishing, Cham (2015) 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.771 + ], + "angle": 0, + "content": "22. Singh, K., Seth, A., Sandhu, H.S., Samdani, K.: A comprehensive review of convolutional neural network based image enhancement techniques. In: 2019 IEEE International Conference on System, Computation, Automation and Networking (ICSCAN). pp. 1-6 (2019). https://doi.org/10.1109/ICSCAN.2019.8878706 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.787, + 0.812 + ], + "angle": 0, + "content": "23. Wang, Z., Bovik, A., Sheikh, H., Simoncelli, E.: Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing 13(4), 600-612 (2004). https://doi.org/10.1109/TIP.2003.819861 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.812, + 0.787, + 0.84 + ], + "angle": 0, + "content": "24. Xia, B., Zhang, Y., Wang, S., Wang, Y., Wu, X., Tian, Y., Yang, W., Gool, L.V.: Diffir: Efficient diffusion model for image restoration. In: 2023 IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.493, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RAVE: Residual Vector Embedding" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.147, + 0.785, + 0.202 + ], + "angle": 0, + "content": "International Conference on Computer Vision (ICCV). pp. 13049-13059. IEEE Computer Society, Los Alamitos, CA, USA (oct 2023). https://doi.org/10.1109/ICCV51070.2023.01204, https://doi.ieeecomputersociety.org/10.1109/ICCV51070.2023.01204 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.787, + 0.258 + ], + "angle": 0, + "content": "25. Yi, X., Xu, H., Zhang, H., Tang, L., Ma, J.: Diff-retinex: Rethinking low-light image enhancement with a generative diffusion model. In: 2023 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 12268-12277 (2023). https://doi.org/10.1109/ICCV51070.2023.011304, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.786, + 0.313 + ], + "angle": 0, + "content": "26. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.: Restormer: Efficient transformer for high-resolution image restoration. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5718-5729 (2022). https://doi.org/10.1109/CVPR52688.2022.005642, 3, 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.314, + 0.786, + 0.369 + ], + "angle": 0, + "content": "27. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 586-595 (2018). https://doi.org/10.1109/CVPR.2018.0006810" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.787, + 0.452 + ], + "angle": 0, + "content": "28. Zheng, B., Pan, X., Zhang, H., Zhou, X., Slabaugh, G., Yan, C., Yuan, S.: Domainplus: Cross transform domain learning towards high dynamic range imaging. In: Proceedings of the 30th ACM International Conference on Multimedia. p. 1954-1963. MM '22, Association for Computing Machinery, New York, NY, USA (2022). https://doi.org/10.1145/3503161.3547823, https://doi.org/10.1145/3503161.3547823 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.452 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/RAVE_ Residual Vector Embedding for CLIP-Guided Backlit Image Enhancement/7f0c9e4e-7b51-4d42-aac8-de63b12e0ec8_origin.pdf b/2024/RAVE_ Residual Vector Embedding for CLIP-Guided Backlit Image Enhancement/7f0c9e4e-7b51-4d42-aac8-de63b12e0ec8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..48d031735d8c2f755ec84dcc5984cbf6dca4dba7 --- /dev/null +++ b/2024/RAVE_ Residual Vector Embedding for CLIP-Guided Backlit Image Enhancement/7f0c9e4e-7b51-4d42-aac8-de63b12e0ec8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:584fba01bb27e1903bdd21a2d5dc31475dd2e449618a5e151f103977c125f12d +size 10013544 diff --git a/2024/RAVE_ Residual Vector Embedding for CLIP-Guided Backlit Image Enhancement/full.md b/2024/RAVE_ Residual Vector Embedding for CLIP-Guided Backlit Image Enhancement/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b73548903d0a63aebe098381659af090ed260b46 --- /dev/null +++ b/2024/RAVE_ Residual Vector Embedding for CLIP-Guided Backlit Image Enhancement/full.md @@ -0,0 +1,353 @@ +# RAVE: Residual Vector Embedding for CLIP-Guided Backlit Image Enhancement + +Tatiana Gaintseva1,2, Martin Benning3, and Gregory Slabaugh1,2 + +$^{1}$ Digital Environment Research Institute, Queen Mary University of London, UK + +$^{2}$ School of Electronic Engineering and Computer Science, Queen Mary University of London, UK {t.gaintseva, g.slabaugh}@qmul.ac.uk + +$^{3}$ Department of Computer Science, University College London, UK +martin.benning@ucl.ac.uk + +Abstract. In this paper we propose a novel modification of Contrastive Language-Image Pre-Training (CLIP) guidance for the task of backlit image enhancement. Our work builds on the state-of-the-art CLIP-LIT approach, which learns a prompt pair by constraining the text-image similarity between a prompt (negative/positive sample) and a corresponding image (backlit image/well-lit image) in the CLIP embedding space. Learned prompts then guide an image enhancement network. Based on the CLIP-LIT framework, we propose two novel methods for CLIP guidance. First, we show that instead of tuning prompts in the space of text embeddings, it is possible to directly tune their embeddings in the latent space without any loss in quality. This accelerates training and potentially enables the use of additional encoders that do not have a text encoder. Second, we propose a novel approach that does not require any prompt tuning. Instead, based on CLIP embeddings of backlit and well-lit images from training data, we compute the residual vector in the embedding space as a simple difference between the mean embeddings of the well-lit and backlit images. This vector then guides the enhancement network during training, pushing a backlit image towards the space of well-lit images. This approach further dramatically reduces training time, stabilizes training and produces high quality enhanced images without artifacts. Additionally, we show that residual vectors can be interpreted, revealing biases in training data, and thereby enabling potential bias correction. Code is available at https://github.com/Atmyre/RAVE + +Keywords: Backlit image enhancement $\cdot$ Vision-language models $\cdot$ Residual vector embedding + +# 1 Introduction + +Backlit image enhancement aims to improve images that suffer from backlighting issues. Backlighting is a phenomenon when the light source is situated behind the photographed object, often resulting in a loss of detail and contrast of some areas due to underexposure, diminishing the overall visual quality of the image. The goal of backlit image enhancement is to correct these exposure discrepancies and + +![](images/4c56d42defe9e54821bd923b385114375cbc9c9d19f5e4d519c205150b67e0df.jpg) +CLIP-LIT + +![](images/bc4de61b474a377884e6caada4a335757951b525ca3b7afd6dd4b7dc2690ab7d.jpg) + +![](images/2be1c07027c5f1c05064f6f0efc139829124c31266cce7e4ce4f25ded599013c.jpg) +RAVE + +![](images/2503d5a1fff65847e7ffbf956bb9c362f83d195ce80ae1c6c0e59ee4ca5558d3.jpg) +Fig. 1: Visual comparison of results obtained by RAVE vs CLIP-LIT. RAVE produces well-lit images with fewer artifacts (note the dark green color of underexposed regions in CLIP-LIT), as shown in the zoomed-in sections. Further comparisons can be found in supplementary material. + +bring the details in the darkened areas out without overexposing the well-exposed parts of the image. Backlit image enhancement significantly improves photograph quality in uncontrolled lighting conditions. It is also vital for fields requiring clear image visibility, such as surveillance, security, and scientific imaging applications. + +However, correcting backlit images is not an easy task. Manual correction requires skill using photo enhancement software, and often substantial time and effort. Automated solutions are preferred, but also face challenges. Global adjustment of the brightness level is not sufficient, as this typically overexposes regions of the image which were well lit prior to enhancement [11]. Spatially adaptive methods have appeared in the literature including early methods that rely on assumptions such as average luminance and a gray world model [8, 12] or Retinex theory [10]. More recent methods approximate well-lit image distributions via end-to-end training of deep networks [3, 26]. Despite progress in the field, considerable room for improvement exists for high quality automated backlit image enhancement. + +Another challenge to high quality automatic backlit image enhancement is the lack of paired data, i.e. backlit images each with a corresponding well-lit image. It is complicated to collect such data, so approaches that work with unpaired data are imperative. Recently, CLIP-LIT [14] was proposed, which successfully utilizes CLIP [19] model guidance for training a backlit enhancement model with unpaired data. CLIP-LIT uses prompt learning techniques to provide CLIP guidance. More specifically, it constructs two learnable text prompts, which are trained to have CLIP embeddings close to the well-lit and backlit images, respectively. These learnt prompts then guide the image enhancement model + +during training. Next, several iterations of additional prompt correction followed by enhancement model fine-tuning using the updated prompts are required to produce the final image enhancement model. This approach is the current state-of-the-art in backlit image enhancement with training on unpaired data. + +In this work, we demonstrate that prompt training is not the most efficient way to implement the CLIP guidance, and propose two novel methods, named CLIP-LIT-Latent and ResiduAl Vector Embedding (RAVE). In CLIP-LIT-Latent, we train vectors corresponding to well-lit and backlit images directly in the CLIP latent space rather than in text embedding space. We show that this approach has similar or superior performance when trained on paired or unpaired image data. Further, in RAVE, we form the guidance vector by subtracting the mean of CLIP latent vectors of backlit images from the mean of CLIP latent vectors of well-lit images. This residual vector points in a direction of moving backlit images to well-lit images in the CLIP embedding space. This vector is then used as guidance for the image enhancement model during training. We show that such guidance enables the enhancement model to produce high quality images, while requiring much less training time. While having all these benefits, both proposed CLIP-LIT-Latent and RAVE retain all the advantages of the original CLIP-LIT method, being lightweight for inference and suitable for training on paired or unpaired data. Moreover, we show that residual vector used for guidance in RAVE is interpretable, which opens up new possibilities for overcoming biases in the training data. + +The contributions of this paper can be summarized as follows: + +1. Inspired by CLIP-LIT, we present two novel approaches to CLIP guidance (CLIP-LIT-Latent and RAVE) working directly in the latent space for the task of backlit image enhancement. +2. We show that these approaches result in similar or better quality according to quantitative metrics in both settings of training on paired or unpaired data, and superior visual quality of resulting images, while requiring considerably less time to train. +3. We demonstrate that the embedding used by RAVE for guidance is interpretable, and its interpretation can reveal biases in the training data. + +# 2 Related work + +Image Enhancement. Image enhancement is the task of image processing with the goal of improving the visual appearance of an image. This is an essential task in numerous fields, such as photography, medical imaging, satellite imagery analysis, and other applications where image quality is crucial. As there are many types of image degradation such as noise, blurriness, low resolution, backlighting, low light, etc., there are a range of image enhancement methods, from traditional ones such as gray level transformation [8,12], histogram equalization [5], and methods based on Retinex theory [10], to machine learning and end-to-end trained deep learning methods [3,26]. A general overview of concepts + +and techniques commonly used for image enhancement is provided by Maini et al. [22]. + +Backlit Image Enhancement. For the task of backlit image enhancement, many approaches have been proposed. Some techniques aim to identify backlit regions and correct them separately from the rest of the image [13]. Methods based on the Retinex [10] theory divide an image into a reflectance image and an illumination map, and both parts are then enhanced and then combined back together to form the resulting enhanced image. Another approach to enhancing backlit images is using end-to-end deep learning approaches, as in Restormer [26] or Retinexformer [3] methods. Combined techniques based on Retinex theory and deep learning were proposed for both backlit and low-light image enhancement, such as Diff-Retinex [25]. These methods decomposing an input image into a reflectance image and an illumination map and apply diffusion methods for enhancement before combining them to produce the output image. Finally, another set of methods are based on high dynamic range (HDR) imaging which combine multiple exposures to produce a composite well-lit image [4, 20, 28]. However, in this work we only focus on single-image methods where back-lit correction is applied as a post-process. + +Fundamentally, backlit image correction is ill-posed, in that for a single backlit image, multiple well-lit variants with high perceptual quality exist. Relatedly, generation of paired training data requires hiring professional photographers, which can be expensive. To help address this problem, recently the BAID [15] dataset was proposed, which contains paired images generated by professional photographers. Nevertheless, development of methods for backlit image enhancement, which can be trained using non-paired data remains essential. + +CLIP guidance. CLIP [19] is a model that consists of two parts: image and text encoders, which learn to project images and text to a common latent embedding space. CLIP has been proven to provide strong image and text priors which are able to provide guidance in many tasks. CLIP guidance has been used in text-to-image generation [17], visual question answering [18], GAN inversion [2], 3D scene generation [7], domain adaptation of image generators [6] and more. Liang et al. [14] also used CLIP guidance for the task of backlit image enhancement. In our work, we aim to improve this technique by modifying the CLIP guidance. + +# 3 Methodology + +Our work is built on top of the CLIP-LIT [14] approach, which uses CLIP guidance for training a backlit image enhancement model. We propose two modifications of CLIP guidance, which reduce training time, stabilize training and produce high quality enhanced images with fewer artifacts compared to CLIP-LIT. + +We start in Section 3.1 by reviewing the architecture of original CLIP-LIT method on which our work is based. Then, in Sections 3.2 and 3.3 we introduce our proposed methods CLIP-LIT-Latent and RAVE. + +![](images/d8392715a7afe95478d5da789a027ff9b1d0ff194e10a4f8a90415ea6d650bf5.jpg) +(a) Initial Prompts and Enhancement Training + +![](images/1965875bbff3a3447b6ca8478c55d5c3d8e14101dfd5fc041ec04c8b294c1d67.jpg) +(b) Prompt Refinement and Enhancement Tuning +Fig. 2: Overview of the original CLIP-LIT approach and proposed CLIP-LIT-Latent. (a) depicts the first stage of training, which consists of prompt or latent vector initialization and the initial training of an enhancement network. (b) shows the second stage, where prompt/latent vector refinement and enhancement model fine-tuning are iteratively repeated. Blue and red boxes are related to CLIP-LIT and CLIP-LIT-Latent. + +# 3.1 CLIP-LIT overview + +The CLIP-LIT framework (Fig. 2) consists of two stages, described below. + +# Initial Prompts and Enhancement Training + +In this stage, two prompts $T_{p}, T_{n} \in \mathbb{R}^{N \times 512}$ representing positive and negative prompts are randomly initialized, where $N$ is a number of tokens in each prompt. Then they are tuned so that the CLIP embedding of $T_{p}$ is close in terms of the dot product to the CLIP embeddings of well-lit training images, and CLIP embedding of $T_{n}$ is close to the CLIP embeddings of backlit images from the training data. The binary cross-entropy loss of classifying the backlit and well + +lit images based on $T_{p}$ and $T_{n}$ is used for this purpose: + +$$ +L _ {\text {i n i t i a l}} = - (y \log (\widehat {y}) + (1 - y) \log (1 - \widehat {y})) +$$ + +$$ +\widehat {y} = \frac {e ^ {\cos \left(\Phi_ {\text {i m a g e}} (I) , \Phi_ {\text {t e x t}} \left(T _ {p}\right)\right)}}{\sum_ {i \in \{n , p \}} e ^ {\cos \left(\Phi_ {\text {i m a g e}} (I) , \Phi_ {\text {t e x t}} \left(T _ {i}\right)\right)}} \tag {1} +$$ + +Here $I$ is a well-lit or backlit input image with label $y \in \{0,1\}$ with 0 corresponding to backlit and 1 corresponding to well-lit; $\varPhi_{\mathrm{image}}$ is a CLIP [19] image encoder, $\varPhi_{\mathrm{text}}$ is a CLIP text encoder; cos is a cosine similarity function. + +Using these initialized prompts $T_{p}$ and $T_{n}$ it is now possible to train an image enhancement network with a CLIP-guided loss. The enhancement model is a Unet [21] which takes backlit image $I_{b}$ as input and outputs an estimated illumination map $I_{i} \in \mathbb{R}^{H \times W \times 1}$ . The final enhanced image $I_{t}$ is then obtained as $I_{t} = I_{b} / I_{i}$ . The image enhancement network is trained by combining two losses: + +$$ +L _ {\text {e n h a n c e}} = L _ {\text {c l i p}} + \omega L _ {\text {i d e n t i t y}} \tag {2} +$$ + +Here $L_{\mathrm{clip}}$ measures the similarity between the enhanced image and the learned prompts $T_{p}$ and $T_{n}$ in the CLIP latent space. $L_{\mathrm{identity}}$ is an identity loss designed to ensure that the enhanced image is similar to the original input image in terms of content and structure. The parameter $\omega$ balances the two losses. These losses are defined as: + +$$ +L _ {\mathrm {c l i p}} = \frac {e ^ {\cos \left(\Phi_ {\mathrm {i m a g e}} \left(I _ {t}\right) , \Phi_ {\mathrm {t e x t}} \left(T _ {n}\right)\right)}}{\sum_ {i \in \{n , p \}} e ^ {\cos \left(\Phi_ {\mathrm {i m a g e}} \left(I _ {t}\right) , \Phi_ {\mathrm {t e x t}} \left(T _ {i}\right)\right)}} \tag {3} +$$ + +$$ +L _ {\text {i d e n t i t y}} = \sum_ {l = 0} ^ {k} \alpha_ {l} \cdot \left\| \Phi_ {\text {i m a g e}} ^ {l} \left(I _ {b}\right) - \Phi_ {\text {i m a g e}} ^ {l} \left(I _ {t}\right) \right\| _ {2} \tag {4} +$$ + +where $\Phi^l$ is $l^{th}$ layer of the CLIP image encoder, and $\alpha_{l}$ is its corresponding weight. $k$ is number of layers used to calculate the metric. + +# Prompt Refinement and Enhancement Tuning + +CLIP-LIT applies a second stage where several cycles of prompt refinement and enhancement network tuning are performed to improve the accuracy of the learned prompts. + +First, learned prompts $T_{p}$ and $T_{n}$ are fine-tuned to provide better guidance for fine-tuning the image enhancement model. A margin ranking loss is used to update the prompts based on images given by the current enhancement model: + +$$ +\begin{array}{l} L _ {\text {p r o m p r e t}} ^ {1} = \max (0, S \left(I _ {w}\right) - S \left(I _ {b}\right) + m _ {0}) + \max (0, S \left(I _ {t - 1}\right) - S \left(I _ {b}\right) + m _ {0}) \\ + \max (0, S \left(I _ {w}\right) - S \left(I _ {t}\right) + m _ {1}) + \max (0, S \left(I _ {t}\right) - S \left(I _ {t - 1}\right) + m _ {2}) \\ \end{array} +$$ + +where $S(I)$ is a negative similarity score between the prompt pair and an image defined as: + +$$ +S (I) = \frac {e ^ {\cos \left(\Phi_ {\text {i m a g e}} (I) , \Phi_ {\text {t e x t}} \left(T _ {n}\right)\right)}}{\sum_ {i \in \{n , p \}} e ^ {\cos \left(\Phi_ {\text {i m a g e}} (I) , \Phi_ {\text {t e x t}} \left(T _ {i}\right)\right)}} \tag {6} +$$ + +Here $m_0 \in [0,1]$ is the margin between the score of well-lit or enhanced result and the backlit image in the CLIP embedding space; $m_1 \in [0,1]$ is the margin between the score of the enhanced result and the well-lit image in the CLIP embedding space; $m_2 \in [0,1]$ is the margin between the newly enhanced result $I_t$ and previously enhanced result $I_{t-1}$ . + +After that, a stage of enhancement model fine-tuning is performed. This follows the same process as initial enhancement model training described above. The process of prompt refinement and enhancement tuning is repeated several times until satisfactory results are obtained. + +# 3.2 CLIP-LIT-Latent + +Like CLIP-LIT, CLIP-LIT-Latent also consists of two stages of initial prompts and enhancement training and prompt refinement and enhancement tuning. Fig. 2 illustrates the CLIP-LIT-Latent framework. + +Instead of learning prompts in the text embedding space, CLIP-LIT-Latent learns a pair of positive/negative vectors directly in the CLIP latent space. In the prompt initialization phase, we randomly initialize two vectors $E_{p}, E_{n} \in \mathbb{R}^{512}$ . These vectors are then tuned so that $E_{p}$ is close in terms of the dot product to the CLIP embedding of well-lit images from the training data, and $E_{n}$ is close to the CLIP embedding of backlit images. As in original CLIP-LIT approach, binary cross-entropy loss for classifying the backlit and well-lit images based on $E_{p}$ and $E_{n}$ is used for this purpose: + +$$ +\begin{array}{l} L _ {\text {i n i t i a l}} = - (y \log (\widehat {y}) + (1 - y) \log (1 - \widehat {y})) \\ \widehat {y} = \frac {e ^ {\cos \left(\Phi_ {\mathrm {i m a g e}} (I) , E _ {p}\right)}}{\sum_ {i \in \{n , p \}} e ^ {\cos \left(\Phi_ {\mathrm {i m a g e}} (I) , E _ {i}\right)}} \tag {7} \\ \end{array} +$$ + +where $I$ is a well-lit or backlit input image with label $y \in 0, 1$ with 0 corresponding to backlit and 1 corresponding to well-lit; $\varPhi_{\mathrm{image}}$ is a CLIP image encoder. + +Based on these initialized vectors $E_{p}$ and $E_{n}$ , an enhancement network is trained with a CLIP-guided loss. As before, the enhancement model is a Unet which takes backlit image $I_{b}$ as input and outputs an estimated illumination map $I_{i} \in \mathbb{R}^{H \times W \times 1}$ . And, as before, it is trained using $L_{\text{enhance}}$ loss, where the $L_{\text{clip}}$ component is changed as follows: + +$$ +L _ {\text {c l i p}} = \frac {e ^ {\cos \left(\Phi_ {\text {i m a g e}} \left(I _ {r}\right) , E _ {n}\right)}}{\sum_ {i \in \{n , p \}} e ^ {\cos \left(\Phi_ {\text {i m a g e}} \left(I _ {t}\right) , E _ {i}\right)}} \tag {8} +$$ + +where $\varPhi^l$ is $l^{th}$ layer of CLIP image encoder, and $a_{l}$ is its corresponding weight. + +In the second stage of prompt refinement and enhancement tuning, where learned vectors $E_{p}$ and $E_{n}$ are fine-tuned using margin ranking loss, the formula for $S(I)$ changes as follows: + +$$ +S (I) = \frac {e ^ {\cos \left(\Phi_ {\text {i m a g e}} (I) , E _ {n}\right)}}{\sum_ {i \in \{n , p \}} e ^ {\cos \left(\Phi_ {\text {i m a g e}} (I) , E _ {i}\right)}} \tag {9} +$$ + +Fig. 3: Overview of the RAVE model. (a) First, we calculate residual vector $\mathbf{v}_{\mathrm{residual}}$ based on backlit and well-lit training data. (b) Then we switch to enhancement model training based on the identity loss and the loss based on the residual vector. +![](images/5e49f5cf0517de84b6d9d8538afbddbc9af2e198fc0c3162feefff15d98ddd33.jpg) +- CLIP vectors of backlit training images + +![](images/cfa3f313b5396055e19418725ab6a4d126cbcafc9bef8110c21b540484222867.jpg) +- CLIP vectors of well-lit training images +(a) Calculation of residual vector $\mathbf{v}_{\text{residual}}$ based on (b) Training of enhancement model using paired or unpaired training data guidance from $\mathbf{v}_{\text{residual}}$ + +In all other aspects the training algorithm is similar to that of CLIP-LIT. + +CLIP-LIT-Latent has several advantages over CLIP-LIT. First, CLIP-LIT-Latent does not use the text encoder, which speeds up the training and inference processes, as it is no longer necessary to pass vectors and gradients through the text encoder. This also potentially enables the use of other vision models which do not have text encoder parts rather than CLIP for guidance. Second, images obtained from CLIP-LIT-Latent tend to have more contrast and have better visual quality (see Fig. 4 and Fig. 1). We speculate this results from tuning vectors $E_{p}$ and $E_{n}$ directly in the CLIP latent space. In contrast, in the original CLIP-LIT, vectors $T_{p}$ and $T_{n}$ belong to the text embedding space, and must be tuned so that their projection through the CLIP text encoder satisfies desired conditions. Such optimization is more complex and makes tuning of $T_{p}$ and $T_{n}$ harder and less accurate. + +# 3.3 RAVE: Residual Vector Embedding + +In this section we present our model RAVE (ResiduAl Vector Embedding). RAVE includes a further modification of the CLIP guidance. We show that RAVE does not require iterative stages of prompt and model updates. Instead it only requires one stage of enhancement model training. + +In RAVE we exploit arithmetic defined in the CLIP latent space. Using well-lit and backlit training data, we construct a residual vector $\mathbf{v}_{\mathrm{residual}}$ , which will + +then be used for enhancement model guidance. $\mathbf{v}_{\mathrm{residual}}$ is defined as follows: + +$$ +\mathbf {v} _ {\text {r e s i d u a l}} = f _ {\text {n o r m}} \left(\mathbf {v} _ {\text {w e l l - l i t}} - \mathbf {v} _ {\text {b a c k l i t}}\right) +$$ + +$$ +\mathbf {v} _ {\text {w e l l - l i t}} = f _ {\text {n o r m}} \left(\frac {\sum_ {i = 1} ^ {N _ {p}} f _ {\text {n o r m}} \left(\Phi_ {\text {i m a g e}} \left(I _ {i} ^ {p}\right)\right)}{N _ {p}}\right) \tag {10} +$$ + +$$ +\mathbf {v} _ {\text {b a c k l i t}} = f _ {n o r m} \left(\frac {\sum_ {i = 1} ^ {N _ {n}} f _ {\text {n o r m}} \left(\Phi_ {\text {i m a g e}} \left(I _ {i} ^ {n}\right)\right)}{N _ {n}}\right) +$$ + +where $f_{\mathrm{norm}}$ is a normalization function: $f_{\mathrm{norm}}(\mathbf{v}) = \frac{\mathbf{v}}{\|\mathbf{v}\|_2}$ , $N_p$ is the number of training well-lit images, $N_n$ is the number of training backlit images, $I_i^p$ is the $i^{th}$ well-lit training image, $I_i^n$ is the $i^{th}$ backlit training image. + +$\mathbf{v}_{\mathrm{residual}}$ is a vector that points in a direction moving from backlit images to well-lit images in the CLIP embedding space. We then use this vector as guidance for the image enhancement model during training. This will train the image enhancement model to produce images with CLIP latent vectors that are close to the CLIP latent vectors of well-lit training images. Overview of the RAVE approach is presented on Fig. 3. + +We use the same Unet [21] as enhancement model as in CLIP-LIT and CLIP-LIT-Latent. The enhancement model is trained using the following loss function: + +$$ +L _ {\text {r a v e}} = L _ {\text {i d e n t i t y}} + \omega L _ {\text {r e s i d u a l}} \tag {11} +$$ + +$$ +L _ {\text {r e s i d u a l}} = \left\| \Phi_ {\text {i m a g e}} (I) \cdot \mathbf {v} _ {\text {r e s i d u a l}} - \mathbf {v} _ {\text {w e l l - l i t}} \cdot \mathbf {v} _ {\text {r e s i d u a l}} \right\| ^ {2} \tag {12} +$$ + +where $I$ is an input image, and $v_{\mathrm{well - lit}}$ is a normalized mean of CLIP embeddings of well-lit training images: + +$$ +\mathbf {v} _ {\text {w e l l - l i t}} = f _ {\text {n o r m}} \left(\frac {\sum_ {i = 1} ^ {N _ {p}} f _ {\text {n o r m}} \left(\Phi_ {\text {i m a g e}} \left(I _ {i} ^ {p}\right)\right)}{N _ {p}}\right) \tag {13} +$$ + +Note that RAVE does not require several repeating stages during training. Before training, the residual vector $\mathbf{v}_{\mathrm{residual}}$ is calculated, and afterwards the training of the image enhancement model begins using $L_{rave}$ as the loss function. This makes RAVE training much more efficient than that of CLIP-LIT and CLIP-LIT-Latent. First, the training of the image enhancement model no longer depends on the quality of learned prompt embeddings, as a fixed residual vector is used for guidance. Also, the iterative manner of CLIP-LIT and CLIP-LIT-Latent training regimes makes training less stable, as the image enhancement model training depends on the quality of prompts obtained on the previous step and vice versa. In supplementary material we show that results obtained on intermediate steps of CLIP-LIT training can be of poor quality and can fluctuate from producing under-exposed to over-exposed resulting images from step to step. In contrast, RAVE has a well-defined fixed initial objective. Results show that RAVE converges up to 25 times faster than CLIP-LIT and CLIP-LIT-Latent. Furthermore, images obtained by RAVE have fewer over-exposed areas, producing high quality results. + +# 4 Experiments + +Datasets Following CLIP-LIT, we use the BAID [15] and DIV2K [1] datasets for training. We train our models in two settings: using paired and unpaired training data. For the paired setting, we use 380 backlit and corresponding well-lit images from BAID dataset. For the unpaired setting, we use 380 backlit images from the BAID and 384 images from DIV2K dataset. For the unpaired setting, we use the same data as used in CLIP-LIT. For testing, we follow recent literature and use the BAID [15] test dataset, which has 368 backlit and corresponding well-lit images. The data is publicly available. + +Training and inference We follow parameter settings of the original CLIP-LIT approach when setting parameters for CLIP-LIT-Latent and RAVE (where applicable), as well as for reproducing CLIP-LIT. Specific details of training and values of hyperparameters are presented in supplementary material. + +Metrics Following recent works in image enhancement, we report four metrics: PSNR [9], SSIM [23], and LPIPS [27] (AlexNet version) and FID [16]. + +Methods for comparison The main model which we compare our approaches to is CLIP-LIT [14], as it showed superior performance on all mentioned metrics, except FID, to many prior image enhancement models. We also include recently published methods such as Restormer [26], Retinexformer [3], DiffIR [24] and Diff-Retinex [25] to the comparison. + +# 4.1 Results + +Quantitative comparison The quantitative comparison on the BAID test dataset is presented in Tab. 1. We report test set metrics of the best model checkpoint obtained during training for all the models. Detailed metrics values for all the checkpoints after each training epoch can be found in supplementary material. Note that many approaches can only be trained using paired data, that is why their results are only present in the "paired" section. We see that our method achieves state-of-the-art performance in both settings. In supplementary material, we also present results on time needed to train all the models, and confirm that both CLIP-LIT-Latent and RAVE are much more efficient. Note that in the setting of unpaired data we have two versions of RAVE, referred as RAVE and RAVE shifted in Tab. 1. RAVE shifted is a variant of RAVE with a simple modification of residual vector, see "Interpretation of the residual vector" section below for more details. + +Visual comparison For visual comparison, we present representative examples from the BAID test dataset. First, in Fig. 4, we compare CLIP-LIT-Latent and RAVE to original CLIP-LIT. First, note that CLIP-LIT-Latent produces more contrast in general, giving more diverse illumination to different regions. Second, in Fig. 1, we present comparison of images produced by RAVE to images produced by CLIP-LIT. We show that images produced by RAVE have fewer artifacts. We hypothesize that this advantage of RAVE may result from its guiding mechanism: in RAVE we directly point the enhancement model to produce images which have similar CLIP embeddings to that of real well-lit images. This + +Table 1: Quantitative comparison of different methods on the BAID test dataset. The best and second best performances in both settings are in bold and underlined. + +
MethodPSNR ↑SSIM ↑LPIPS ↓FID ↓
Original images16.910.760.2152.47
PairedRestormer21.070.8320.19241.17
Retinexformer22.030.8620.17345.27
DiffIR21.100.8350.17540.35
Diff-Retinex22.070.8610.16038.07
CLIP-LIT21.930.8750.15641.16
CLIP-LIT-Latent (ours)21.840.8770.15542.56
RAVE (ours)22.260.8800.13936.01
UnpairedEnlightenGAN17.550.8640.19643.50
CLIP-LIT21.590.8740.16046.49
CLIP-LIT-Latent (ours)21.450.8770.15646.20
RAVE (ours)20.390.8610.15540.11
RAVE shifted (ours)21.370.8770.14940.90
+ +does not allow model to produce regions with artifacts, as its CLIP embeddings won't match those of real well-lit images. In CLIP-LIT or CLIP-LIT-Latent, on the other hand, the objective of the enhancement model is to produce images for which the CLIP embeddings are simply closer to the positive prompt/vector than to the negative one. This allows the enhancement model to produce less accurate images, CLIP embeddings of which are just "close enough" to the CLIP embeddings of real well-lit images. More results of visual comparison of all the models can be found in supplementary material. + +Interpretation of the residual vector In RAVE, we build the residual vector $\mathbf{v}_{\text{residual}}$ following formula 16. We noted that $\mathbf{v}_{\text{residual}}$ is a vector that points to a direction of moving from backlit images to well-lit images in the CLIP embedding space. Here we prove this by interpreting vector $\mathbf{v}_{\text{residual}}$ using the fact that CLIP has a common vector space for embeddings of images and text. + +We go through the vocabulary of the CLIP text encoder and find those vocabulary tokens which have the closest and farthest CLIP embeddings to the $\mathbf{v}_{\text{residual}}$ in the CLIP latent space in terms of cosine similarity. More formally, we use the following metric: + +$$ +\operatorname {s i m} (\text {t o k e n}) = f _ {\text {n o r m}} \left(\Phi_ {\text {t e x t}} \left(e _ {\text {t o k e n}}\right)\right) \cdot \mathbf {v} _ {\text {r e s i d u a l}} \tag {14} +$$ + +Here $e_{token}$ is embedding of vocabulary token, which is input to the CLIP text encoder $\varPhi_{text}$ . We sort vocabulary tokens using sim. Table 2 shows 7 vocabulary tokens which have highest and lowest cosine similarity to the $v_{residual}$ vectors obtained from the paired dataset and from the unpaired dataset. + +In the case of paired data, many of the least similar tokens have a meaning around "darkness": "silhouette", "dark", "shadow", etc., and absolute value of their cosine similarity to $v_{\mathrm{residual}}$ is around $0.12 - 0.15$ . At the same time, most similar tokens do not share any meaningful semantics, and the absolute value + +![](images/04c78a3ca080e4259ef79c6fb999e0d808b813d5321854d749f9c4c7e01e525f.jpg) + +![](images/6fed4143d3a5ab7d022ff6e187c7f52189a5a9558400e76b513d97774b3c6c6d.jpg) + +![](images/54805cd322501140ecc7b5532f9f0fed0865741ba26acfd77b5beea670b81a26.jpg) + +![](images/7427c9ddc16216b987ae04de1d7222ecbe8906eaa1c07db025bd39eb07f93e61.jpg) +CLIP-LIT + +![](images/82348f4401a069e6f67850de01bbcf94810d68ccefe23c4e736cd39733d11bc3.jpg) +CLIP-LIT-Latent + +![](images/83f7b1f3420458904e917c1c9736ec137d30a2b24d35d76fe1ecea2bd0f817ff.jpg) +RAVE + +![](images/ae3f81a5ea7377334f8366ff5e56242a551a3204f359a4edced299113d0d1319.jpg) +(a) Visual comparison of results by CLIP-LIT, CLIP-LIT-Latent and RAVE, all trained with paired data. RAVE produces results with more general contrast. + +![](images/7bbc3161de51bd7c8cc44a87dfdb5698b1d507e11a605e014e2e802bc262ccec.jpg) + +![](images/f8439baae704b93c0f573f02ca0f5cc3c183d24085937c8076806ca975e784b2.jpg) + +CLIP-LIT +![](images/d3faeb33581fbf8eb85efcc16a0ea4df9766c19a5bff3ced2689c54ef9182ec1.jpg) +(b) Visual comparison of results obtained by CLIP-LIT, CLIP-LIT-Latent and RAVE (shifted version), all trained with unpaired data. RAVE produces results with more general contrast. + +![](images/480b1de885fcb754ab27679ecca57a84f6f6fabebc2ad223bb49a48239292337.jpg) +CLIP-LIT-Latent + +![](images/47cc75a10ec8f6532bed5d9d58e8f32a80ba0f60915653e40ed31cb8981ad4f2.jpg) +RAVE +Fig. 4: Visual comparison of results by CLIP-LIT, CLIP-LIT-Latent and RAVE. + +of their cosine similarity to $v_{\mathrm{residual}}$ is much smaller, being around $0.01 - 0.04$ . This supports the hypothesis that $v_{\mathrm{residual}}$ has a meaning of shifting from "dark", "shadowy" images to "normal" images, having average level of lightness. $v_{\mathrm{residual}}$ does not carry any other semantic meaning, as it was calculated using paired dataset, where for each backlit image there is its well-lit analogue, and all the semantics present in the BAID dataset is cancelled as a result of subtraction of the mean well-lit and backlit vectors when constructing $v_{\mathrm{residual}}$ . On the other hand, in the case of unpaired data, we see that the least similar tokens have a meaning around "asian": "busan", "beijing", etc., and the most similar ones have a meaning around "nature": "wildlife", "southafrica", "countrylife", etc. Herein, an absolute value of cosine similarities for both sets of tokens if quite high, being around $0.08 - 0.12$ . This reflects the bias in the unpaired datasets. Indeed, in the BAID dataset there are mostly images of Chinese cities or people, while in DIV2K dataset there are many pictures of nature. Due to the bias, RAVE training with unpaired data becomes less effective, making the enhancement model + +Table 2: Vocabulary tokens and their cosine similarities, which have the lowest and highest cosine similarity to the $v_{\text{residual}}$ vectors calculated for paired and unpaired training datasets. + +
PairedUnpaired
Lowest similarityHighest similarityLowest similarityHighest similarity
silhouette-0.151healthtech0.035busan-0.108wildlifewednesdayday0.116
darkness-0.149whitepaper0.035beijing-0.103southafrica0.103
dark-0.139theeconomist0.030iu-0.100ecuador0.095
webcamtoy-0.139bona0.029guangzhou-0.095patrol0.095
blackand-0.132digitaltransformation0.027incheon-0.094countryfile0.093
vedere-0.131sua0.021shenzhen-0.094tasmanian0.093
shadow-0.130amarketing0.021eunhyuk-0.091womensart0.091
+ +shift images not in an ideal direction in the CLIP latent space. Nevertheless, RAVE still shows good performance when trained with unpaired data. + +We also propose a simple modification for correcting the residual vector in the unpaired setting. We simply take $n$ most and least similar tokens to form an additional residual vector as follows: + +$$ +\begin{array}{l} \mathbf {v} _ {\mathrm {a d d \_ r e s i d u a l}} = f _ {\mathrm {n o r m}} (\mathbf {v} _ {\mathrm {c l o s e s t}} - \mathbf {v} _ {\mathrm {f a r t h e s t}}) \\ \mathbf {v} _ {\text {c l o s e s t}} = f _ {\text {n o r m}} \left(\frac {\sum_ {i = 1} ^ {n} f _ {\text {n o r m}} \left(\Phi_ {\text {t e x t}} \left(T _ {i} ^ {c}\right)\right)}{n}\right) \tag {15} \\ \mathbf {v} _ {\mathrm {f a r t h e s t}} = f _ {n o r m} \left(\frac {\sum_ {i = 1} ^ {n} f _ {\mathrm {n o r m}} (\varPhi_ {\mathrm {t e x t}} (T _ {i} ^ {f}))}{n}\right) \\ \end{array} +$$ + +$T_{i}^{c}$ and $T_{i}^{f}$ are tokens with $i^{th}$ most and least similar embeddings to the $\mathbf{v}_{\mathrm{residual}}$ . And finally we update the $\mathbf{v}_{\mathrm{residual}}$ as follows: + +$$ +\mathbf {v} _ {\text {r e s i d u a l}} = \mathbf {v} _ {\text {r e s i d u a l}} - c o s \left(\mathbf {v} _ {\text {r e s i d u a l}}, \mathbf {v} _ {\text {a d d} \_ \text {r e s i d u a l}}\right) \cdot \mathbf {v} _ {\text {a d d} \_ \text {r e s i d u a l}} \tag {16} +$$ + +Resulting vector $\mathbf{v}_{\mathrm{residual}}$ will keep the meaning of shifting from "dark", images to "normal" images while losing the unwanted semantics. Experimental results show that RAVE trained with such updated $\mathbf{v}_{\mathrm{residual}}$ produces better results (see RAVE shifted in Tab. 1). In supplementary material we present visual comparison of results obtained by original and shifted versions of RAVE, where it can be clearly seen that RAVE shifted produces more well-lit results, while keeping all the advantages of RAVE. There we also present analysis of tokens with the most and least similar embeddings to the updated $\mathbf{v}_{\mathrm{residual}}$ . This analysis supports the claim about the meaning of updated $\mathbf{v}_{\mathrm{residual}}$ . + +Necessity of the residual vector Given the interpretation of $\mathbf{v}_{\mathrm{residual}}$ , we investigate whether it is necessary to compute $\mathbf{v}_{\mathrm{residual}}$ using the training data, or is it enough to use CLIP latent vectors of words such as "dark", "darkness" as a guidance vector for enhancement model. We trained RAVE using $v_{\mathrm{residual}} =$ + +Table 3: Mean similarity of CLIP latent vectors of well-lit and back-lit images from BAID and DIV2K datasets to the CLIP latent vectors of vocabulary tokens + +
TokenBAID backlitBAID well-litDIV2K well-lit
silhouette0.2540.2270.223
dark0.2470.2220.232
darkness0.2420.2150.227
+ +$\varPhi_{\mathrm{text}}(token)$ as guidance vector, with token $\in$ {"silhouette", "dark", "darkness"} but did not observe any meaningful enhancement in the resulting images. We suspect that this is due to the fact that latent vectors of backlit and well-lit images have similar cosine similarity to embeddings of these tokens, as presented in Tab 3. This suggests that residual vector $\mathbf{v}_{\mathrm{residual}}$ is not simply a combination of latent vectors of some tokens, but it contains more specific information with the meaning of shifting from backlit to well-lit data. + +# 4.2 Limitations + +Despite showing great performance in the task of backlit image enhancement, our models have limitation of non-optimal enhancement of the regions where the loss of information occurs, making restoration ambiguous. Our approach has no generative components and cannot regenerate highly-accurate textures when the information is lost. This leads to sub-optimal performance in such tasks as low-light image enhancement, where images are highly under-exposed (see supplementary material). Moreover, our models are not trained to exactly match pixels of resulting images with ground-truth data, so some sub-optimal lightening of pixels might occur, which results in non-ideal visual quality. + +However, existing generative-based models also have limitations. Many of them cannot be trained with unpaired data setting, and require much time and memory to be trained and used. They may also hallucinate detail. As a consequence, the resolution of images it can work with is limited. + +# 5 Conclusion + +Based on CLIP-LIT method, we have presented two improved CLIP guidance techniques for backlit image enhancement task, named CLIP-LIT-Latent and RAVE. Our methods show similar or superior performance to that of CLIP-LIT in both cases of training with paired or unpaired data, while having lower time and space requirements. Furthermore, latent vector constructed by RAVE is interpretable and can be used to find biases in the training data, as well as potentially correct them. Moreover, our proposed approach has several advantages over the current SOTA in backlit image enhancement, Diff-Retinex. First, unlike Diff-Retinex, our method can also be trained using unpaired data. Second, it is more lightweight, requiring less compute and space for training and inference, suitable for images with larger resolution. + +# Acknowledgements + +Funding for this research was provided by a Google-DeepMind PhD Studentship, and work utilised Queen Mary's Andrena HPC facility, supported by QMUL Research-IT. We also sincerely thank Laida Kushnareva for assisting with images and text composition in the paper. + +# References + +1. Agustsson, E., Timofte, R.: Ntire 2017 challenge on single image super-resolution: Dataset and study. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 1122-1131 (2017). https://doi.org/10.1109/CVPRW.2017.150 10 +2. Baykal, A.C., Anees, A.B., Ceylan, D., Erdem, E., Erdem, A., Yuret, D.: Clip-guided stylegan inversion for text-driven real image editing (2023) 4 +3. Cai, Y., Bian, H., Lin, J., Wang, H., Timofte, R., Zhang, Y.: Retinexformer: One-stage retinax-based transformer for low-light image enhancement. In: 2023 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 12470-12479 (2023). https://doi.org/10.1109/ICCV51070.2023.011492, 3, 4, 10 +4. Catley-Chandar, S., Tanay, T., Vandroux, L., Leonardis, A., Slabaugh, G., Pérez-Pellitero, E.: Flexhdr: Modeling alignment and exposure uncertainties for flexible hdr imaging. IEEE Transactions on Image Processing 31, 5923-5935 (2022). https://doi.org/10.1109/TIP.2022.3203562 4 +5. Cheng, H.D., Shi, X.: A simple and effective histogram equalization approach to image enhancement. Digital signal processing 14(2), 158-170 (2004) 3 +6. Gal, R., Patashnik, O., Maron, H., Bermano, A.H., Chechik, G., Cohen-Or, D.: Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Trans. Graph. 41(4) (jul 2022). https://doi.org/10.1145/3528223.3530164, https://doi.org/10.1145/3528223.3530164 4 +7. Goloujeh, A.M., Smith, J., Magerko, B.: Explainable clip-guided 3d-scene generation in an ai holodeck. In: Proceedings of the Eighteenth AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment. AIIDE'22, AAAI Press (2022). https://doi.org/10.1609/aide.v18i1.21973, https://doi.org/10.1609/aide.v18i1.21973 4 +8. Guo, C., Li, C., Guo, J., Loy, C.C., Hou, J., Kwong, S., Cong, R.: Zero-reference deep curve estimation for low-light image enhancement. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1777-1786 (2020). https://doi.org/10.1109/CVPR42600.2020.001852, 3 +9. Horé, A., Ziou, D.: Image quality metrics: Psnr vs. sSIM. In: 2010 20th International Conference on Pattern Recognition. pp. 2366-2369 (2010). https://doi.org/10.1109/ICPR.2010.579.10 +0. Land, E.H.: The retina theory of color vision. Scientific american 237(6), 108-129 (1977) 2, 3, 4 +1. Li, C., Guo, C., Han, L., Jiang, J., Cheng, M., Gu, J., Loy, C.: Low-light image and video enhancement using deep learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(12), 9396-9416 (dec 2022). https://doi.org/10.1109/TPAMI.2021.3126387 2 + +12. Li, C., Guo, C., Loy, C.C.: Learning to enhance low-light image via zero-reference deep curve estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(8), 4225-4238 (2022). https://doi.org/10.1109/TPAMI.2021.3063604 2, 3 +13. Li, Z., Wu, X.: Learning-based restoration of backlit images. IEEE Transactions on Image Processing 27(2), 976-986 (2018). https://doi.org/10.1109/TIP.2017.2771142 4 +14. Liang, Z., Li, C., Zhou, S., Feng, R., Loy, C.C.: Iterative prompt learning for unsupervised backlit image enhancement. In: 2023 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 8060-8069 (2023). https://doi.org/10.1109/ICCV51070.2023.00743 2, 4, 10 +15. Lv, X., Zhang, S., Liu, Q., Xie, H., Zhong, B., Zhou, H.: Backlitnet: A dataset and network for backlit image enhancement. Computer Vision and Image Understanding 218, 103403 (2022). https://doi.org/https://doi.org/10.1016/j.cviu.2022.103403, https://www.sciencedirect.com/science/article/pii/S1077314222000340 4, 10 +16. Naeem, M.F., Oh, S.J., Uh, Y., Choi, Y., Yoo, J.: Reliable fidelity and diversity metrics for generative models. In: III, H.D., Singh, A. (eds.) Proceedings of the 37th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 119, pp. 7176-7185. PMLR (13-18 Jul 2020), https://proceedings.mlr.press/v119/naeem20a.html 10 +17. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models (12 2021) 4 +18. Parelli, M., Delitzas, A., Hars, N., Vlassis, G., Anagnostidis, S., Bachmann, G., Hofmann, T.: Clip-guided vision-language pre-training for question answering in 3d scenes. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 5607-5612 (2023). https://doi.org/10.1109/CVPRW59228.2023.005934 +19. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (2021), https://apisemantic scholar.org/CorpusID: 231591445 2, 4, 6 +20. Reinhard, E.: High Dynamic Range Imaging, pp. 1-6. Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-03243-2_843-1, https://doi.org/10.1007/978-3-030-03243-2_843-1 4 +21. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Navab, N., Hornegger, J., Wells, W.M., Frangi, A.F. (eds.) Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015. pp. 234-241. Springer International Publishing, Cham (2015) 6, 9 +22. Singh, K., Seth, A., Sandhu, H.S., Samdani, K.: A comprehensive review of convolutional neural network based image enhancement techniques. In: 2019 IEEE International Conference on System, Computation, Automation and Networking (ICSCAN). pp. 1-6 (2019). https://doi.org/10.1109/ICSCAN.2019.8878706 4 +23. Wang, Z., Bovik, A., Sheikh, H., Simoncelli, E.: Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing 13(4), 600-612 (2004). https://doi.org/10.1109/TIP.2003.819861 10 +24. Xia, B., Zhang, Y., Wang, S., Wang, Y., Wu, X., Tian, Y., Yang, W., Gool, L.V.: Diffir: Efficient diffusion model for image restoration. In: 2023 IEEE/CVF + +International Conference on Computer Vision (ICCV). pp. 13049-13059. IEEE Computer Society, Los Alamitos, CA, USA (oct 2023). https://doi.org/10.1109/ICCV51070.2023.01204, https://doi.ieeecomputersociety.org/10.1109/ICCV51070.2023.01204 10 +25. Yi, X., Xu, H., Zhang, H., Tang, L., Ma, J.: Diff-retinex: Rethinking low-light image enhancement with a generative diffusion model. In: 2023 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 12268-12277 (2023). https://doi.org/10.1109/ICCV51070.2023.011304, 10 +26. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.: Restormer: Efficient transformer for high-resolution image restoration. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5718-5729 (2022). https://doi.org/10.1109/CVPR52688.2022.005642, 3, 4, 10 +27. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 586-595 (2018). https://doi.org/10.1109/CVPR.2018.0006810 +28. Zheng, B., Pan, X., Zhang, H., Zhou, X., Slabaugh, G., Yan, C., Yuan, S.: Domainplus: Cross transform domain learning towards high dynamic range imaging. In: Proceedings of the 30th ACM International Conference on Multimedia. p. 1954-1963. MM '22, Association for Computing Machinery, New York, NY, USA (2022). https://doi.org/10.1145/3503161.3547823, https://doi.org/10.1145/3503161.3547823 4 \ No newline at end of file diff --git a/2024/RAVE_ Residual Vector Embedding for CLIP-Guided Backlit Image Enhancement/images.zip b/2024/RAVE_ Residual Vector Embedding for CLIP-Guided Backlit Image Enhancement/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..84e02f31b4067e7c50281481246cc05b74dcf34c --- /dev/null +++ b/2024/RAVE_ Residual Vector Embedding for CLIP-Guided Backlit Image Enhancement/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:655073f481b2079cd4671a7eea1362b780d295da1927b01b58d6129ef61870c1 +size 560889 diff --git a/2024/RAVE_ Residual Vector Embedding for CLIP-Guided Backlit Image Enhancement/layout.json b/2024/RAVE_ Residual Vector Embedding for CLIP-Guided Backlit Image Enhancement/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..99015659a70801bbb5881b197e17347f19652d99 --- /dev/null +++ b/2024/RAVE_ Residual Vector Embedding for CLIP-Guided Backlit Image Enhancement/layout.json @@ -0,0 +1,9775 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 159, + 111, + 455, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 111, + 455, + 148 + ], + "spans": [ + { + "bbox": [ + 159, + 111, + 455, + 148 + ], + "type": "text", + "content": "RAVE: Residual Vector Embedding for CLIP-Guided Backlit Image Enhancement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 152, + 167, + 462, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 167, + 462, + 181 + ], + "spans": [ + { + "bbox": [ + 152, + 167, + 462, + 181 + ], + "type": "text", + "content": "Tatiana Gaintseva1,2, Martin Benning3, and Gregory Slabaugh1,2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 189, + 476, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 189, + 476, + 200 + ], + "spans": [ + { + "bbox": [ + 132, + 189, + 476, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 189, + 476, + 200 + ], + "type": "text", + "content": " Digital Environment Research Institute, Queen Mary University of London, UK" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 200, + 480, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 200, + 480, + 222 + ], + "spans": [ + { + "bbox": [ + 132, + 200, + 480, + 222 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 200, + 480, + 222 + ], + "type": "text", + "content": " School of Electronic Engineering and Computer Science, Queen Mary University of London, UK {t.gaintseva, g.slabaugh}@qmul.ac.uk" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 167, + 222, + 446, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 222, + 446, + 244 + ], + "spans": [ + { + "bbox": [ + 167, + 222, + 446, + 244 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 167, + 222, + 446, + 244 + ], + "type": "text", + "content": " Department of Computer Science, University College London, UK \nmartin.benning@ucl.ac.uk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 274, + 455, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 274, + 455, + 528 + ], + "spans": [ + { + "bbox": [ + 160, + 274, + 455, + 528 + ], + "type": "text", + "content": "Abstract. In this paper we propose a novel modification of Contrastive Language-Image Pre-Training (CLIP) guidance for the task of backlit image enhancement. Our work builds on the state-of-the-art CLIP-LIT approach, which learns a prompt pair by constraining the text-image similarity between a prompt (negative/positive sample) and a corresponding image (backlit image/well-lit image) in the CLIP embedding space. Learned prompts then guide an image enhancement network. Based on the CLIP-LIT framework, we propose two novel methods for CLIP guidance. First, we show that instead of tuning prompts in the space of text embeddings, it is possible to directly tune their embeddings in the latent space without any loss in quality. This accelerates training and potentially enables the use of additional encoders that do not have a text encoder. Second, we propose a novel approach that does not require any prompt tuning. Instead, based on CLIP embeddings of backlit and well-lit images from training data, we compute the residual vector in the embedding space as a simple difference between the mean embeddings of the well-lit and backlit images. This vector then guides the enhancement network during training, pushing a backlit image towards the space of well-lit images. This approach further dramatically reduces training time, stabilizes training and produces high quality enhanced images without artifacts. Additionally, we show that residual vectors can be interpreted, revealing biases in training data, and thereby enabling potential bias correction. Code is available at https://github.com/Atmyre/RAVE" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 537, + 459, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 537, + 459, + 560 + ], + "spans": [ + { + "bbox": [ + 159, + 537, + 459, + 560 + ], + "type": "text", + "content": "Keywords: Backlit image enhancement " + }, + { + "bbox": [ + 159, + 537, + 459, + 560 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 537, + 459, + 560 + ], + "type": "text", + "content": " Vision-language models " + }, + { + "bbox": [ + 159, + 537, + 459, + 560 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 537, + 459, + 560 + ], + "type": "text", + "content": " Residual vector embedding" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 580, + 230, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 580, + 230, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 580, + 230, + 594 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Backlit image enhancement aims to improve images that suffer from backlighting issues. Backlighting is a phenomenon when the light source is situated behind the photographed object, often resulting in a loss of detail and contrast of some areas due to underexposure, diminishing the overall visual quality of the image. The goal of backlit image enhancement is to correct these exposure discrepancies and" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 121, + 302, + 190 + ], + "blocks": [ + { + "bbox": [ + 153, + 121, + 302, + 190 + ], + "lines": [ + { + "bbox": [ + 153, + 121, + 302, + 190 + ], + "spans": [ + { + "bbox": [ + 153, + 121, + 302, + 190 + ], + "type": "image", + "image_path": "4c56d42defe9e54821bd923b385114375cbc9c9d19f5e4d519c205150b67e0df.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 288, + 194, + 318, + 203 + ], + "lines": [ + { + "bbox": [ + 288, + 194, + 318, + 203 + ], + "spans": [ + { + "bbox": [ + 288, + 194, + 318, + 203 + ], + "type": "text", + "content": "CLIP-LIT" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 121, + 455, + 190 + ], + "blocks": [ + { + "bbox": [ + 309, + 121, + 455, + 190 + ], + "lines": [ + { + "bbox": [ + 309, + 121, + 455, + 190 + ], + "spans": [ + { + "bbox": [ + 309, + 121, + 455, + 190 + ], + "type": "image", + "image_path": "bc4de61b474a377884e6caada4a335757951b525ca3b7afd6dd4b7dc2690ab7d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 152, + 208, + 301, + 281 + ], + "blocks": [ + { + "bbox": [ + 152, + 208, + 301, + 281 + ], + "lines": [ + { + "bbox": [ + 152, + 208, + 301, + 281 + ], + "spans": [ + { + "bbox": [ + 152, + 208, + 301, + 281 + ], + "type": "image", + "image_path": "2be1c07027c5f1c05064f6f0efc139829124c31266cce7e4ce4f25ded599013c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 293, + 287, + 313, + 297 + ], + "lines": [ + { + "bbox": [ + 293, + 287, + 313, + 297 + ], + "spans": [ + { + "bbox": [ + 293, + 287, + 313, + 297 + ], + "type": "text", + "content": "RAVE" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 208, + 455, + 281 + ], + "blocks": [ + { + "bbox": [ + 309, + 208, + 455, + 281 + ], + "lines": [ + { + "bbox": [ + 309, + 208, + 455, + 281 + ], + "spans": [ + { + "bbox": [ + 309, + 208, + 455, + 281 + ], + "type": "image", + "image_path": "2503d5a1fff65847e7ffbf956bb9c362f83d195ce80ae1c6c0e59ee4ca5558d3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 308, + 482, + 354 + ], + "lines": [ + { + "bbox": [ + 131, + 308, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 131, + 308, + 482, + 354 + ], + "type": "text", + "content": "Fig. 1: Visual comparison of results obtained by RAVE vs CLIP-LIT. RAVE produces well-lit images with fewer artifacts (note the dark green color of underexposed regions in CLIP-LIT), as shown in the zoomed-in sections. Further comparisons can be found in supplementary material." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 377, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 482, + 426 + ], + "type": "text", + "content": "bring the details in the darkened areas out without overexposing the well-exposed parts of the image. Backlit image enhancement significantly improves photograph quality in uncontrolled lighting conditions. It is also vital for fields requiring clear image visibility, such as surveillance, security, and scientific imaging applications." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 426, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 557 + ], + "type": "text", + "content": "However, correcting backlit images is not an easy task. Manual correction requires skill using photo enhancement software, and often substantial time and effort. Automated solutions are preferred, but also face challenges. Global adjustment of the brightness level is not sufficient, as this typically overexposes regions of the image which were well lit prior to enhancement [11]. Spatially adaptive methods have appeared in the literature including early methods that rely on assumptions such as average luminance and a gray world model [8, 12] or Retinex theory [10]. More recent methods approximate well-lit image distributions via end-to-end training of deep networks [3, 26]. Despite progress in the field, considerable room for improvement exists for high quality automated backlit image enhancement." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 558, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 483, + 666 + ], + "type": "text", + "content": "Another challenge to high quality automatic backlit image enhancement is the lack of paired data, i.e. backlit images each with a corresponding well-lit image. It is complicated to collect such data, so approaches that work with unpaired data are imperative. Recently, CLIP-LIT [14] was proposed, which successfully utilizes CLIP [19] model guidance for training a backlit enhancement model with unpaired data. CLIP-LIT uses prompt learning techniques to provide CLIP guidance. More specifically, it constructs two learnable text prompts, which are trained to have CLIP embeddings close to the well-lit and backlit images, respectively. These learnt prompts then guide the image enhancement model" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 243, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 243, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 243, + 101 + ], + "type": "text", + "content": "T. Gaintseva et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "during training. Next, several iterations of additional prompt correction followed by enhancement model fine-tuning using the updated prompts are required to produce the final image enhancement model. This approach is the current state-of-the-art in backlit image enhancement with training on unpaired data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 481, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 481, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 481, + 379 + ], + "type": "text", + "content": "In this work, we demonstrate that prompt training is not the most efficient way to implement the CLIP guidance, and propose two novel methods, named CLIP-LIT-Latent and ResiduAl Vector Embedding (RAVE). In CLIP-LIT-Latent, we train vectors corresponding to well-lit and backlit images directly in the CLIP latent space rather than in text embedding space. We show that this approach has similar or superior performance when trained on paired or unpaired image data. Further, in RAVE, we form the guidance vector by subtracting the mean of CLIP latent vectors of backlit images from the mean of CLIP latent vectors of well-lit images. This residual vector points in a direction of moving backlit images to well-lit images in the CLIP embedding space. This vector is then used as guidance for the image enhancement model during training. We show that such guidance enables the enhancement model to produce high quality images, while requiring much less training time. While having all these benefits, both proposed CLIP-LIT-Latent and RAVE retain all the advantages of the original CLIP-LIT method, being lightweight for inference and suitable for training on paired or unpaired data. Moreover, we show that residual vector used for guidance in RAVE is interpretable, which opens up new possibilities for overcoming biases in the training data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 385, + 420, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 385, + 420, + 397 + ], + "spans": [ + { + "bbox": [ + 146, + 385, + 420, + 397 + ], + "type": "text", + "content": "The contributions of this paper can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 406, + 479, + 513 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 136, + 406, + 479, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 406, + 479, + 441 + ], + "spans": [ + { + "bbox": [ + 136, + 406, + 479, + 441 + ], + "type": "text", + "content": "1. Inspired by CLIP-LIT, we present two novel approaches to CLIP guidance (CLIP-LIT-Latent and RAVE) working directly in the latent space for the task of backlit image enhancement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 443, + 479, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 443, + 479, + 487 + ], + "spans": [ + { + "bbox": [ + 136, + 443, + 479, + 487 + ], + "type": "text", + "content": "2. We show that these approaches result in similar or better quality according to quantitative metrics in both settings of training on paired or unpaired data, and superior visual quality of resulting images, while requiring considerably less time to train." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 135, + 490, + 479, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 490, + 479, + 513 + ], + "spans": [ + { + "bbox": [ + 135, + 490, + 479, + 513 + ], + "type": "text", + "content": "3. We demonstrate that the embedding used by RAVE for guidance is interpretable, and its interpretation can reveal biases in the training data." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 533, + 234, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 234, + 545 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 234, + 545 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 558, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 666 + ], + "type": "text", + "content": "Image Enhancement. Image enhancement is the task of image processing with the goal of improving the visual appearance of an image. This is an essential task in numerous fields, such as photography, medical imaging, satellite imagery analysis, and other applications where image quality is crucial. As there are many types of image degradation such as noise, blurriness, low resolution, backlighting, low light, etc., there are a range of image enhancement methods, from traditional ones such as gray level transformation [8,12], histogram equalization [5], and methods based on Retinex theory [10], to machine learning and end-to-end trained deep learning methods [3,26]. A general overview of concepts" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "text", + "content": "RAVE: Residual Vector Embedding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": "and techniques commonly used for image enhancement is provided by Maini et al. [22]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 141, + 481, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 141, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 141, + 481, + 332 + ], + "type": "text", + "content": "Backlit Image Enhancement. For the task of backlit image enhancement, many approaches have been proposed. Some techniques aim to identify backlit regions and correct them separately from the rest of the image [13]. Methods based on the Retinex [10] theory divide an image into a reflectance image and an illumination map, and both parts are then enhanced and then combined back together to form the resulting enhanced image. Another approach to enhancing backlit images is using end-to-end deep learning approaches, as in Restormer [26] or Retinexformer [3] methods. Combined techniques based on Retinex theory and deep learning were proposed for both backlit and low-light image enhancement, such as Diff-Retinex [25]. These methods decomposing an input image into a reflectance image and an illumination map and apply diffusion methods for enhancement before combining them to produce the output image. Finally, another set of methods are based on high dynamic range (HDR) imaging which combine multiple exposures to produce a composite well-lit image [4, 20, 28]. However, in this work we only focus on single-image methods where back-lit correction is applied as a post-process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 334, + 481, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 334, + 481, + 417 + ], + "spans": [ + { + "bbox": [ + 132, + 334, + 481, + 417 + ], + "type": "text", + "content": "Fundamentally, backlit image correction is ill-posed, in that for a single backlit image, multiple well-lit variants with high perceptual quality exist. Relatedly, generation of paired training data requires hiring professional photographers, which can be expensive. To help address this problem, recently the BAID [15] dataset was proposed, which contains paired images generated by professional photographers. Nevertheless, development of methods for backlit image enhancement, which can be trained using non-paired data remains essential." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 418, + 481, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 514 + ], + "type": "text", + "content": "CLIP guidance. CLIP [19] is a model that consists of two parts: image and text encoders, which learn to project images and text to a common latent embedding space. CLIP has been proven to provide strong image and text priors which are able to provide guidance in many tasks. CLIP guidance has been used in text-to-image generation [17], visual question answering [18], GAN inversion [2], 3D scene generation [7], domain adaptation of image generators [6] and more. Liang et al. [14] also used CLIP guidance for the task of backlit image enhancement. In our work, we aim to improve this technique by modifying the CLIP guidance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 538, + 231, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 538, + 231, + 552 + ], + "spans": [ + { + "bbox": [ + 133, + 538, + 231, + 552 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 568, + 480, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 480, + 628 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 480, + 628 + ], + "type": "text", + "content": "Our work is built on top of the CLIP-LIT [14] approach, which uses CLIP guidance for training a backlit image enhancement model. We propose two modifications of CLIP guidance, which reduce training time, stabilize training and produce high quality enhanced images with fewer artifacts compared to CLIP-LIT." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 480, + 665 + ], + "type": "text", + "content": "We start in Section 3.1 by reviewing the architecture of original CLIP-LIT method on which our work is based. Then, in Sections 3.2 and 3.3 we introduce our proposed methods CLIP-LIT-Latent and RAVE." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 100 + ], + "type": "text", + "content": "T. Gaintseva et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 129, + 499, + 252 + ], + "blocks": [ + { + "bbox": [ + 137, + 129, + 499, + 252 + ], + "lines": [ + { + "bbox": [ + 137, + 129, + 499, + 252 + ], + "spans": [ + { + "bbox": [ + 137, + 129, + 499, + 252 + ], + "type": "image", + "image_path": "d8392715a7afe95478d5da789a027ff9b1d0ff194e10a4f8a90415ea6d650bf5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 255, + 388, + 266 + ], + "lines": [ + { + "bbox": [ + 220, + 255, + 388, + 266 + ], + "spans": [ + { + "bbox": [ + 220, + 255, + 388, + 266 + ], + "type": "text", + "content": "(a) Initial Prompts and Enhancement Training" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 135, + 269, + 499, + 403 + ], + "blocks": [ + { + "bbox": [ + 135, + 269, + 499, + 403 + ], + "lines": [ + { + "bbox": [ + 135, + 269, + 499, + 403 + ], + "spans": [ + { + "bbox": [ + 135, + 269, + 499, + 403 + ], + "type": "image", + "image_path": "1965875bbff3a3447b6ca8478c55d5c3d8e14101dfd5fc041ec04c8b294c1d67.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 215, + 406, + 393, + 417 + ], + "lines": [ + { + "bbox": [ + 215, + 406, + 393, + 417 + ], + "spans": [ + { + "bbox": [ + 215, + 406, + 393, + 417 + ], + "type": "text", + "content": "(b) Prompt Refinement and Enhancement Tuning" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 425, + 482, + 480 + ], + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 480 + ], + "type": "text", + "content": "Fig. 2: Overview of the original CLIP-LIT approach and proposed CLIP-LIT-Latent. (a) depicts the first stage of training, which consists of prompt or latent vector initialization and the initial training of an enhancement network. (b) shows the second stage, where prompt/latent vector refinement and enhancement model fine-tuning are iteratively repeated. Blue and red boxes are related to CLIP-LIT and CLIP-LIT-Latent." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 513, + 259, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 513, + 259, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 513, + 259, + 523 + ], + "type": "text", + "content": "3.1 CLIP-LIT overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 553, + 462, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 553, + 462, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 553, + 462, + 566 + ], + "type": "text", + "content": "The CLIP-LIT framework (Fig. 2) consists of two stages, described below." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 352, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 352, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 352, + 593 + ], + "type": "text", + "content": "Initial Prompts and Enhancement Training" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "In this stage, two prompts " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "T_{p}, T_{n} \\in \\mathbb{R}^{N \\times 512}" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " representing positive and negative prompts are randomly initialized, where " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " is a number of tokens in each prompt. Then they are tuned so that the CLIP embedding of " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "T_{p}" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " is close in terms of the dot product to the CLIP embeddings of well-lit training images, and CLIP embedding of " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "T_{n}" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " is close to the CLIP embeddings of backlit images from the training data. The binary cross-entropy loss of classifying the backlit and well" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "text", + "content": "RAVE: Residual Vector Embedding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 375, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 375, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 375, + 129 + ], + "type": "text", + "content": "lit images based on " + }, + { + "bbox": [ + 132, + 116, + 375, + 129 + ], + "type": "inline_equation", + "content": "T_{p}" + }, + { + "bbox": [ + 132, + 116, + 375, + 129 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 116, + 375, + 129 + ], + "type": "inline_equation", + "content": "T_{n}" + }, + { + "bbox": [ + 132, + 116, + 375, + 129 + ], + "type": "text", + "content": " is used for this purpose:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 216, + 137, + 397, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 137, + 397, + 152 + ], + "spans": [ + { + "bbox": [ + 216, + 137, + 397, + 152 + ], + "type": "interline_equation", + "content": "L _ {\\text {i n i t i a l}} = - (y \\log (\\widehat {y}) + (1 - y) \\log (1 - \\widehat {y}))", + "image_path": "18255d5027540b1f35f23fded60e1123aa6128db3a5ef766f461c90ee3d62c0b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 241, + 153, + 481, + 184 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 153, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 241, + 153, + 481, + 184 + ], + "type": "interline_equation", + "content": "\\widehat {y} = \\frac {e ^ {\\cos \\left(\\Phi_ {\\text {i m a g e}} (I) , \\Phi_ {\\text {t e x t}} \\left(T _ {p}\\right)\\right)}}{\\sum_ {i \\in \\{n , p \\}} e ^ {\\cos \\left(\\Phi_ {\\text {i m a g e}} (I) , \\Phi_ {\\text {t e x t}} \\left(T _ {i}\\right)\\right)}} \\tag {1}", + "image_path": "a970e6f06f6b46a1dbc2b0879850688814f3fb4e6c7236705efc3bae59801a64.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 479, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 479, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 479, + 228 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 132, + 192, + 479, + 228 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 132, + 192, + 479, + 228 + ], + "type": "text", + "content": " is a well-lit or backlit input image with label " + }, + { + "bbox": [ + 132, + 192, + 479, + 228 + ], + "type": "inline_equation", + "content": "y \\in \\{0,1\\}" + }, + { + "bbox": [ + 132, + 192, + 479, + 228 + ], + "type": "text", + "content": " with 0 corresponding to backlit and 1 corresponding to well-lit; " + }, + { + "bbox": [ + 132, + 192, + 479, + 228 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\mathrm{image}}" + }, + { + "bbox": [ + 132, + 192, + 479, + 228 + ], + "type": "text", + "content": " is a CLIP [19] image encoder, " + }, + { + "bbox": [ + 132, + 192, + 479, + 228 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\mathrm{text}}" + }, + { + "bbox": [ + 132, + 192, + 479, + 228 + ], + "type": "text", + "content": " is a CLIP text encoder; cos is a cosine similarity function." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 228, + 480, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 480, + 288 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 480, + 288 + ], + "type": "text", + "content": "Using these initialized prompts " + }, + { + "bbox": [ + 132, + 228, + 480, + 288 + ], + "type": "inline_equation", + "content": "T_{p}" + }, + { + "bbox": [ + 132, + 228, + 480, + 288 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 228, + 480, + 288 + ], + "type": "inline_equation", + "content": "T_{n}" + }, + { + "bbox": [ + 132, + 228, + 480, + 288 + ], + "type": "text", + "content": " it is now possible to train an image enhancement network with a CLIP-guided loss. The enhancement model is a Unet [21] which takes backlit image " + }, + { + "bbox": [ + 132, + 228, + 480, + 288 + ], + "type": "inline_equation", + "content": "I_{b}" + }, + { + "bbox": [ + 132, + 228, + 480, + 288 + ], + "type": "text", + "content": " as input and outputs an estimated illumination map " + }, + { + "bbox": [ + 132, + 228, + 480, + 288 + ], + "type": "inline_equation", + "content": "I_{i} \\in \\mathbb{R}^{H \\times W \\times 1}" + }, + { + "bbox": [ + 132, + 228, + 480, + 288 + ], + "type": "text", + "content": ". The final enhanced image " + }, + { + "bbox": [ + 132, + 228, + 480, + 288 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 132, + 228, + 480, + 288 + ], + "type": "text", + "content": " is then obtained as " + }, + { + "bbox": [ + 132, + 228, + 480, + 288 + ], + "type": "inline_equation", + "content": "I_{t} = I_{b} / I_{i}" + }, + { + "bbox": [ + 132, + 228, + 480, + 288 + ], + "type": "text", + "content": ". The image enhancement network is trained by combining two losses:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 244, + 300, + 481, + 313 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 300, + 481, + 313 + ], + "spans": [ + { + "bbox": [ + 244, + 300, + 481, + 313 + ], + "type": "interline_equation", + "content": "L _ {\\text {e n h a n c e}} = L _ {\\text {c l i p}} + \\omega L _ {\\text {i d e n t i t y}} \\tag {2}", + "image_path": "801c2cdd11159f8741e7be4af15ea293e1194ee0e2ea0d21d86dde27fa370eed.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 318, + 479, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 318, + 479, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 318, + 479, + 376 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 132, + 318, + 479, + 376 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{clip}}" + }, + { + "bbox": [ + 132, + 318, + 479, + 376 + ], + "type": "text", + "content": " measures the similarity between the enhanced image and the learned prompts " + }, + { + "bbox": [ + 132, + 318, + 479, + 376 + ], + "type": "inline_equation", + "content": "T_{p}" + }, + { + "bbox": [ + 132, + 318, + 479, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 318, + 479, + 376 + ], + "type": "inline_equation", + "content": "T_{n}" + }, + { + "bbox": [ + 132, + 318, + 479, + 376 + ], + "type": "text", + "content": " in the CLIP latent space. " + }, + { + "bbox": [ + 132, + 318, + 479, + 376 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{identity}}" + }, + { + "bbox": [ + 132, + 318, + 479, + 376 + ], + "type": "text", + "content": " is an identity loss designed to ensure that the enhanced image is similar to the original input image in terms of content and structure. The parameter " + }, + { + "bbox": [ + 132, + 318, + 479, + 376 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 132, + 318, + 479, + 376 + ], + "type": "text", + "content": " balances the two losses. These losses are defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 220, + 375, + 481, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 375, + 481, + 406 + ], + "spans": [ + { + "bbox": [ + 220, + 375, + 481, + 406 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {c l i p}} = \\frac {e ^ {\\cos \\left(\\Phi_ {\\mathrm {i m a g e}} \\left(I _ {t}\\right) , \\Phi_ {\\mathrm {t e x t}} \\left(T _ {n}\\right)\\right)}}{\\sum_ {i \\in \\{n , p \\}} e ^ {\\cos \\left(\\Phi_ {\\mathrm {i m a g e}} \\left(I _ {t}\\right) , \\Phi_ {\\mathrm {t e x t}} \\left(T _ {i}\\right)\\right)}} \\tag {3}", + "image_path": "12888955c3ad0641a0c7d2f785e0d951b4b14e7330c9b7ee1b02152efff54606.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 208, + 411, + 481, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 411, + 481, + 444 + ], + "spans": [ + { + "bbox": [ + 208, + 411, + 481, + 444 + ], + "type": "interline_equation", + "content": "L _ {\\text {i d e n t i t y}} = \\sum_ {l = 0} ^ {k} \\alpha_ {l} \\cdot \\left\\| \\Phi_ {\\text {i m a g e}} ^ {l} \\left(I _ {b}\\right) - \\Phi_ {\\text {i m a g e}} ^ {l} \\left(I _ {t}\\right) \\right\\| _ {2} \\tag {4}", + "image_path": "36c7a7681797fb3ac20b9eceab9bbdadacc43981c5451dffef6f2516c1398d38.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 449, + 479, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 479, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 479, + 474 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 449, + 479, + 474 + ], + "type": "inline_equation", + "content": "\\Phi^l" + }, + { + "bbox": [ + 132, + 449, + 479, + 474 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 132, + 449, + 479, + 474 + ], + "type": "inline_equation", + "content": "l^{th}" + }, + { + "bbox": [ + 132, + 449, + 479, + 474 + ], + "type": "text", + "content": " layer of the CLIP image encoder, and " + }, + { + "bbox": [ + 132, + 449, + 479, + 474 + ], + "type": "inline_equation", + "content": "\\alpha_{l}" + }, + { + "bbox": [ + 132, + 449, + 479, + 474 + ], + "type": "text", + "content": " is its corresponding weight. " + }, + { + "bbox": [ + 132, + 449, + 479, + 474 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 132, + 449, + 479, + 474 + ], + "type": "text", + "content": " is number of layers used to calculate the metric." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 486, + 367, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 367, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 367, + 498 + ], + "type": "text", + "content": "Prompt Refinement and Enhancement Tuning" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 498, + 479, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 479, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 479, + 533 + ], + "type": "text", + "content": "CLIP-LIT applies a second stage where several cycles of prompt refinement and enhancement network tuning are performed to improve the accuracy of the learned prompts." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 534, + 479, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 479, + 570 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 479, + 570 + ], + "type": "text", + "content": "First, learned prompts " + }, + { + "bbox": [ + 132, + 534, + 479, + 570 + ], + "type": "inline_equation", + "content": "T_{p}" + }, + { + "bbox": [ + 132, + 534, + 479, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 534, + 479, + 570 + ], + "type": "inline_equation", + "content": "T_{n}" + }, + { + "bbox": [ + 132, + 534, + 479, + 570 + ], + "type": "text", + "content": " are fine-tuned to provide better guidance for fine-tuning the image enhancement model. A margin ranking loss is used to update the prompts based on images given by the current enhancement model:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 143, + 578, + 455, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 578, + 455, + 608 + ], + "spans": [ + { + "bbox": [ + 143, + 578, + 455, + 608 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {\\text {p r o m p r e t}} ^ {1} = \\max (0, S \\left(I _ {w}\\right) - S \\left(I _ {b}\\right) + m _ {0}) + \\max (0, S \\left(I _ {t - 1}\\right) - S \\left(I _ {b}\\right) + m _ {0}) \\\\ + \\max (0, S \\left(I _ {w}\\right) - S \\left(I _ {t}\\right) + m _ {1}) + \\max (0, S \\left(I _ {t}\\right) - S \\left(I _ {t - 1}\\right) + m _ {2}) \\\\ \\end{array}", + "image_path": "385ff5010e462cce3cd13f21b758d62f34fb4161a1e7e7b9e01e73836d8b0d6b.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 617, + 479, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 617, + 479, + 639 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 479, + 639 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 617, + 479, + 639 + ], + "type": "inline_equation", + "content": "S(I)" + }, + { + "bbox": [ + 132, + 617, + 479, + 639 + ], + "type": "text", + "content": " is a negative similarity score between the prompt pair and an image defined as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 221, + 638, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 638, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 221, + 638, + 481, + 669 + ], + "type": "interline_equation", + "content": "S (I) = \\frac {e ^ {\\cos \\left(\\Phi_ {\\text {i m a g e}} (I) , \\Phi_ {\\text {t e x t}} \\left(T _ {n}\\right)\\right)}}{\\sum_ {i \\in \\{n , p \\}} e ^ {\\cos \\left(\\Phi_ {\\text {i m a g e}} (I) , \\Phi_ {\\text {t e x t}} \\left(T _ {i}\\right)\\right)}} \\tag {6}", + "image_path": "96e4dc391781eac50e7f657f149acc44af44963082189beec3d7cbee8fcaa0d9.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 243, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 243, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 243, + 101 + ], + "type": "text", + "content": "T. Gaintseva et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "m_0 \\in [0,1]" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " is the margin between the score of well-lit or enhanced result and the backlit image in the CLIP embedding space; " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "m_1 \\in [0,1]" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " is the margin between the score of the enhanced result and the well-lit image in the CLIP embedding space; " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "m_2 \\in [0,1]" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " is the margin between the newly enhanced result " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "I_t" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " and previously enhanced result " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "I_{t-1}" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 479, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 479, + 223 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 479, + 223 + ], + "type": "text", + "content": "After that, a stage of enhancement model fine-tuning is performed. This follows the same process as initial enhancement model training described above. The process of prompt refinement and enhancement tuning is repeated several times until satisfactory results are obtained." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 241, + 248, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 241, + 248, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 241, + 248, + 251 + ], + "type": "text", + "content": "3.2 CLIP-LIT-Latent" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 260, + 479, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 479, + 294 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 479, + 294 + ], + "type": "text", + "content": "Like CLIP-LIT, CLIP-LIT-Latent also consists of two stages of initial prompts and enhancement training and prompt refinement and enhancement tuning. Fig. 2 illustrates the CLIP-LIT-Latent framework." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 296, + 480, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 480, + 392 + ], + "type": "text", + "content": "Instead of learning prompts in the text embedding space, CLIP-LIT-Latent learns a pair of positive/negative vectors directly in the CLIP latent space. In the prompt initialization phase, we randomly initialize two vectors " + }, + { + "bbox": [ + 130, + 296, + 480, + 392 + ], + "type": "inline_equation", + "content": "E_{p}, E_{n} \\in \\mathbb{R}^{512}" + }, + { + "bbox": [ + 130, + 296, + 480, + 392 + ], + "type": "text", + "content": ". These vectors are then tuned so that " + }, + { + "bbox": [ + 130, + 296, + 480, + 392 + ], + "type": "inline_equation", + "content": "E_{p}" + }, + { + "bbox": [ + 130, + 296, + 480, + 392 + ], + "type": "text", + "content": " is close in terms of the dot product to the CLIP embedding of well-lit images from the training data, and " + }, + { + "bbox": [ + 130, + 296, + 480, + 392 + ], + "type": "inline_equation", + "content": "E_{n}" + }, + { + "bbox": [ + 130, + 296, + 480, + 392 + ], + "type": "text", + "content": " is close to the CLIP embedding of backlit images. As in original CLIP-LIT approach, binary cross-entropy loss for classifying the backlit and well-lit images based on " + }, + { + "bbox": [ + 130, + 296, + 480, + 392 + ], + "type": "inline_equation", + "content": "E_{p}" + }, + { + "bbox": [ + 130, + 296, + 480, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 296, + 480, + 392 + ], + "type": "inline_equation", + "content": "E_{n}" + }, + { + "bbox": [ + 130, + 296, + 480, + 392 + ], + "type": "text", + "content": " is used for this purpose:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 219, + 401, + 479, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 401, + 479, + 445 + ], + "spans": [ + { + "bbox": [ + 219, + 401, + 479, + 445 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {\\text {i n i t i a l}} = - (y \\log (\\widehat {y}) + (1 - y) \\log (1 - \\widehat {y})) \\\\ \\widehat {y} = \\frac {e ^ {\\cos \\left(\\Phi_ {\\mathrm {i m a g e}} (I) , E _ {p}\\right)}}{\\sum_ {i \\in \\{n , p \\}} e ^ {\\cos \\left(\\Phi_ {\\mathrm {i m a g e}} (I) , E _ {i}\\right)}} \\tag {7} \\\\ \\end{array}", + "image_path": "414cee85e86848803fbdc57d083ac06cba294dfe76635d4e5f958e7d65e4428c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 452, + 479, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 452, + 479, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 452, + 479, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 452, + 479, + 475 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 452, + 479, + 475 + ], + "type": "text", + "content": " is a well-lit or backlit input image with label " + }, + { + "bbox": [ + 130, + 452, + 479, + 475 + ], + "type": "inline_equation", + "content": "y \\in 0, 1" + }, + { + "bbox": [ + 130, + 452, + 479, + 475 + ], + "type": "text", + "content": " with 0 corresponding to backlit and 1 corresponding to well-lit; " + }, + { + "bbox": [ + 130, + 452, + 479, + 475 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\mathrm{image}}" + }, + { + "bbox": [ + 130, + 452, + 479, + 475 + ], + "type": "text", + "content": " is a CLIP image encoder." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 476, + 480, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 476, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 476, + 480, + 536 + ], + "type": "text", + "content": "Based on these initialized vectors " + }, + { + "bbox": [ + 130, + 476, + 480, + 536 + ], + "type": "inline_equation", + "content": "E_{p}" + }, + { + "bbox": [ + 130, + 476, + 480, + 536 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 476, + 480, + 536 + ], + "type": "inline_equation", + "content": "E_{n}" + }, + { + "bbox": [ + 130, + 476, + 480, + 536 + ], + "type": "text", + "content": ", an enhancement network is trained with a CLIP-guided loss. As before, the enhancement model is a Unet which takes backlit image " + }, + { + "bbox": [ + 130, + 476, + 480, + 536 + ], + "type": "inline_equation", + "content": "I_{b}" + }, + { + "bbox": [ + 130, + 476, + 480, + 536 + ], + "type": "text", + "content": " as input and outputs an estimated illumination map " + }, + { + "bbox": [ + 130, + 476, + 480, + 536 + ], + "type": "inline_equation", + "content": "I_{i} \\in \\mathbb{R}^{H \\times W \\times 1}" + }, + { + "bbox": [ + 130, + 476, + 480, + 536 + ], + "type": "text", + "content": ". And, as before, it is trained using " + }, + { + "bbox": [ + 130, + 476, + 480, + 536 + ], + "type": "inline_equation", + "content": "L_{\\text{enhance}}" + }, + { + "bbox": [ + 130, + 476, + 480, + 536 + ], + "type": "text", + "content": " loss, where the " + }, + { + "bbox": [ + 130, + 476, + 480, + 536 + ], + "type": "inline_equation", + "content": "L_{\\text{clip}}" + }, + { + "bbox": [ + 130, + 476, + 480, + 536 + ], + "type": "text", + "content": " component is changed as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 234, + 545, + 479, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 545, + 479, + 574 + ], + "spans": [ + { + "bbox": [ + 234, + 545, + 479, + 574 + ], + "type": "interline_equation", + "content": "L _ {\\text {c l i p}} = \\frac {e ^ {\\cos \\left(\\Phi_ {\\text {i m a g e}} \\left(I _ {r}\\right) , E _ {n}\\right)}}{\\sum_ {i \\in \\{n , p \\}} e ^ {\\cos \\left(\\Phi_ {\\text {i m a g e}} \\left(I _ {t}\\right) , E _ {i}\\right)}} \\tag {8}", + "image_path": "277658763377304f1ce7e80d5e82d6bf86a1033b8c3dcfc70942e8eee11608e7.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 582, + 479, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 479, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 479, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 582, + 479, + 594 + ], + "type": "inline_equation", + "content": "\\varPhi^l" + }, + { + "bbox": [ + 130, + 582, + 479, + 594 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 130, + 582, + 479, + 594 + ], + "type": "inline_equation", + "content": "l^{th}" + }, + { + "bbox": [ + 130, + 582, + 479, + 594 + ], + "type": "text", + "content": " layer of CLIP image encoder, and " + }, + { + "bbox": [ + 130, + 582, + 479, + 594 + ], + "type": "inline_equation", + "content": "a_{l}" + }, + { + "bbox": [ + 130, + 582, + 479, + 594 + ], + "type": "text", + "content": " is its corresponding weight." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 595, + 479, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 595, + 479, + 630 + ], + "spans": [ + { + "bbox": [ + 130, + 595, + 479, + 630 + ], + "type": "text", + "content": "In the second stage of prompt refinement and enhancement tuning, where learned vectors " + }, + { + "bbox": [ + 130, + 595, + 479, + 630 + ], + "type": "inline_equation", + "content": "E_{p}" + }, + { + "bbox": [ + 130, + 595, + 479, + 630 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 595, + 479, + 630 + ], + "type": "inline_equation", + "content": "E_{n}" + }, + { + "bbox": [ + 130, + 595, + 479, + 630 + ], + "type": "text", + "content": " are fine-tuned using margin ranking loss, the formula for " + }, + { + "bbox": [ + 130, + 595, + 479, + 630 + ], + "type": "inline_equation", + "content": "S(I)" + }, + { + "bbox": [ + 130, + 595, + 479, + 630 + ], + "type": "text", + "content": " changes as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 235, + 639, + 479, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 639, + 479, + 668 + ], + "spans": [ + { + "bbox": [ + 235, + 639, + 479, + 668 + ], + "type": "interline_equation", + "content": "S (I) = \\frac {e ^ {\\cos \\left(\\Phi_ {\\text {i m a g e}} (I) , E _ {n}\\right)}}{\\sum_ {i \\in \\{n , p \\}} e ^ {\\cos \\left(\\Phi_ {\\text {i m a g e}} (I) , E _ {i}\\right)}} \\tag {9}", + "image_path": "8dcb876f809ff1d3d8b75a2747e3953b534fa8f52c766542a9980ec436609fea.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 301, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 301, + 91, + 448, + 102 + ], + "type": "text", + "content": "RAVE: Residual Vector Embedding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 121, + 315, + 249 + ], + "blocks": [ + { + "bbox": [ + 140, + 121, + 315, + 249 + ], + "lines": [ + { + "bbox": [ + 140, + 121, + 315, + 249 + ], + "spans": [ + { + "bbox": [ + 140, + 121, + 315, + 249 + ], + "type": "image", + "image_path": "5e49f5cf0517de84b6d9d8538afbddbc9af2e198fc0c3162feefff15d98ddd33.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 256, + 276, + 267 + ], + "lines": [ + { + "bbox": [ + 145, + 256, + 276, + 267 + ], + "spans": [ + { + "bbox": [ + 145, + 256, + 276, + 267 + ], + "type": "text", + "content": "- CLIP vectors of backlit training images" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 131, + 316, + 482, + 350 + ], + "lines": [ + { + "bbox": [ + 131, + 316, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 131, + 316, + 482, + 350 + ], + "type": "text", + "content": "Fig. 3: Overview of the RAVE model. (a) First, we calculate residual vector " + }, + { + "bbox": [ + 131, + 316, + 482, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{residual}}" + }, + { + "bbox": [ + 131, + 316, + 482, + 350 + ], + "type": "text", + "content": " based on backlit and well-lit training data. (b) Then we switch to enhancement model training based on the identity loss and the loss based on the residual vector." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 326, + 196, + 481, + 281 + ], + "blocks": [ + { + "bbox": [ + 145, + 274, + 276, + 284 + ], + "lines": [ + { + "bbox": [ + 145, + 274, + 276, + 284 + ], + "spans": [ + { + "bbox": [ + 145, + 274, + 276, + 284 + ], + "type": "text", + "content": "- CLIP vectors of well-lit training images" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 326, + 196, + 481, + 281 + ], + "lines": [ + { + "bbox": [ + 326, + 196, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 326, + 196, + 481, + 281 + ], + "type": "image", + "image_path": "cfa3f313b5396055e19418725ab6a4d126cbcafc9bef8110c21b540484222867.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 290, + 483, + 308 + ], + "lines": [ + { + "bbox": [ + 132, + 290, + 483, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 290, + 483, + 308 + ], + "type": "text", + "content": "(a) Calculation of residual vector " + }, + { + "bbox": [ + 132, + 290, + 483, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\text{residual}}" + }, + { + "bbox": [ + 132, + 290, + 483, + 308 + ], + "type": "text", + "content": " based on (b) Training of enhancement model using paired or unpaired training data guidance from " + }, + { + "bbox": [ + 132, + 290, + 483, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\text{residual}}" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 376, + 457, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 376, + 457, + 388 + ], + "spans": [ + { + "bbox": [ + 131, + 376, + 457, + 388 + ], + "type": "text", + "content": "In all other aspects the training algorithm is similar to that of CLIP-LIT." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 398, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 482, + 542 + ], + "type": "text", + "content": "CLIP-LIT-Latent has several advantages over CLIP-LIT. First, CLIP-LIT-Latent does not use the text encoder, which speeds up the training and inference processes, as it is no longer necessary to pass vectors and gradients through the text encoder. This also potentially enables the use of other vision models which do not have text encoder parts rather than CLIP for guidance. Second, images obtained from CLIP-LIT-Latent tend to have more contrast and have better visual quality (see Fig. 4 and Fig. 1). We speculate this results from tuning vectors " + }, + { + "bbox": [ + 130, + 398, + 482, + 542 + ], + "type": "inline_equation", + "content": "E_{p}" + }, + { + "bbox": [ + 130, + 398, + 482, + 542 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 398, + 482, + 542 + ], + "type": "inline_equation", + "content": "E_{n}" + }, + { + "bbox": [ + 130, + 398, + 482, + 542 + ], + "type": "text", + "content": " directly in the CLIP latent space. In contrast, in the original CLIP-LIT, vectors " + }, + { + "bbox": [ + 130, + 398, + 482, + 542 + ], + "type": "inline_equation", + "content": "T_{p}" + }, + { + "bbox": [ + 130, + 398, + 482, + 542 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 398, + 482, + 542 + ], + "type": "inline_equation", + "content": "T_{n}" + }, + { + "bbox": [ + 130, + 398, + 482, + 542 + ], + "type": "text", + "content": " belong to the text embedding space, and must be tuned so that their projection through the CLIP text encoder satisfies desired conditions. Such optimization is more complex and makes tuning of " + }, + { + "bbox": [ + 130, + 398, + 482, + 542 + ], + "type": "inline_equation", + "content": "T_{p}" + }, + { + "bbox": [ + 130, + 398, + 482, + 542 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 398, + 482, + 542 + ], + "type": "inline_equation", + "content": "T_{n}" + }, + { + "bbox": [ + 130, + 398, + 482, + 542 + ], + "type": "text", + "content": " harder and less accurate." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 566, + 342, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 566, + 342, + 578 + ], + "spans": [ + { + "bbox": [ + 131, + 566, + 342, + 578 + ], + "type": "text", + "content": "3.3 RAVE: Residual Vector Embedding" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 592, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 592, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 592, + 482, + 640 + ], + "type": "text", + "content": "In this section we present our model RAVE (ResiduAl Vector Embedding). RAVE includes a further modification of the CLIP guidance. We show that RAVE does not require iterative stages of prompt and model updates. Instead it only requires one stage of enhancement model training." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "In RAVE we exploit arithmetic defined in the CLIP latent space. Using well-lit and backlit training data, we construct a residual vector " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{residual}}" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": ", which will" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 243, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 243, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 243, + 101 + ], + "type": "text", + "content": "T. Gaintseva et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 469, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 469, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 469, + 129 + ], + "type": "text", + "content": "then be used for enhancement model guidance. " + }, + { + "bbox": [ + 132, + 116, + 469, + 129 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{residual}}" + }, + { + "bbox": [ + 132, + 116, + 469, + 129 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 211, + 134, + 363, + 146 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 134, + 363, + 146 + ], + "spans": [ + { + "bbox": [ + 211, + 134, + 363, + 146 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {\\text {r e s i d u a l}} = f _ {\\text {n o r m}} \\left(\\mathbf {v} _ {\\text {w e l l - l i t}} - \\mathbf {v} _ {\\text {b a c k l i t}}\\right)", + "image_path": "36d854781b351b924ff995f0043fcd0c133379e71856f5ca20700dbc4256161c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 211, + 149, + 480, + 181 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 149, + 480, + 181 + ], + "spans": [ + { + "bbox": [ + 211, + 149, + 480, + 181 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {\\text {w e l l - l i t}} = f _ {\\text {n o r m}} \\left(\\frac {\\sum_ {i = 1} ^ {N _ {p}} f _ {\\text {n o r m}} \\left(\\Phi_ {\\text {i m a g e}} \\left(I _ {i} ^ {p}\\right)\\right)}{N _ {p}}\\right) \\tag {10}", + "image_path": "7aa5b14cae874df087e298f76ff0b653116d7736cccefc6e093d7024a0b66f37.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 211, + 183, + 401, + 215 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 183, + 401, + 215 + ], + "spans": [ + { + "bbox": [ + 211, + 183, + 401, + 215 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {\\text {b a c k l i t}} = f _ {n o r m} \\left(\\frac {\\sum_ {i = 1} ^ {N _ {n}} f _ {\\text {n o r m}} \\left(\\Phi_ {\\text {i m a g e}} \\left(I _ {i} ^ {n}\\right)\\right)}{N _ {n}}\\right)", + "image_path": "246fbabe9363d8d5be53475d5cc0a89c97f20da86375ebb2fc4eb586629fea9d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 220, + 481, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 220, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 131, + 220, + 481, + 258 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 220, + 481, + 258 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{norm}}" + }, + { + "bbox": [ + 131, + 220, + 481, + 258 + ], + "type": "text", + "content": " is a normalization function: " + }, + { + "bbox": [ + 131, + 220, + 481, + 258 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{norm}}(\\mathbf{v}) = \\frac{\\mathbf{v}}{\\|\\mathbf{v}\\|_2}" + }, + { + "bbox": [ + 131, + 220, + 481, + 258 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 220, + 481, + 258 + ], + "type": "inline_equation", + "content": "N_p" + }, + { + "bbox": [ + 131, + 220, + 481, + 258 + ], + "type": "text", + "content": " is the number of training well-lit images, " + }, + { + "bbox": [ + 131, + 220, + 481, + 258 + ], + "type": "inline_equation", + "content": "N_n" + }, + { + "bbox": [ + 131, + 220, + 481, + 258 + ], + "type": "text", + "content": " is the number of training backlit images, " + }, + { + "bbox": [ + 131, + 220, + 481, + 258 + ], + "type": "inline_equation", + "content": "I_i^p" + }, + { + "bbox": [ + 131, + 220, + 481, + 258 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 131, + 220, + 481, + 258 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 131, + 220, + 481, + 258 + ], + "type": "text", + "content": " well-lit training image, " + }, + { + "bbox": [ + 131, + 220, + 481, + 258 + ], + "type": "inline_equation", + "content": "I_i^n" + }, + { + "bbox": [ + 131, + 220, + 481, + 258 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 131, + 220, + 481, + 258 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 131, + 220, + 481, + 258 + ], + "type": "text", + "content": " backlit training image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 258, + 479, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 258, + 479, + 329 + ], + "spans": [ + { + "bbox": [ + 131, + 258, + 479, + 329 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{residual}}" + }, + { + "bbox": [ + 131, + 258, + 479, + 329 + ], + "type": "text", + "content": " is a vector that points in a direction moving from backlit images to well-lit images in the CLIP embedding space. We then use this vector as guidance for the image enhancement model during training. This will train the image enhancement model to produce images with CLIP latent vectors that are close to the CLIP latent vectors of well-lit training images. Overview of the RAVE approach is presented on Fig. 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 330, + 479, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 330, + 479, + 353 + ], + "spans": [ + { + "bbox": [ + 131, + 330, + 479, + 353 + ], + "type": "text", + "content": "We use the same Unet [21] as enhancement model as in CLIP-LIT and CLIP-LIT-Latent. The enhancement model is trained using the following loss function:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 246, + 366, + 480, + 378 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 366, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 246, + 366, + 480, + 378 + ], + "type": "interline_equation", + "content": "L _ {\\text {r a v e}} = L _ {\\text {i d e n t i t y}} + \\omega L _ {\\text {r e s i d u a l}} \\tag {11}", + "image_path": "a568296a7f52fd528991f798ff3eae74577ed76852cd37f739ca5b974a620d42.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 194, + 392, + 480, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 392, + 480, + 406 + ], + "spans": [ + { + "bbox": [ + 194, + 392, + 480, + 406 + ], + "type": "interline_equation", + "content": "L _ {\\text {r e s i d u a l}} = \\left\\| \\Phi_ {\\text {i m a g e}} (I) \\cdot \\mathbf {v} _ {\\text {r e s i d u a l}} - \\mathbf {v} _ {\\text {w e l l - l i t}} \\cdot \\mathbf {v} _ {\\text {r e s i d u a l}} \\right\\| ^ {2} \\tag {12}", + "image_path": "693f6ea20efa8e44299cd461210de447923a8b6a93432090b155fb49346bb764.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 409, + 479, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 409, + 479, + 433 + ], + "spans": [ + { + "bbox": [ + 131, + 409, + 479, + 433 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 409, + 479, + 433 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 131, + 409, + 479, + 433 + ], + "type": "text", + "content": " is an input image, and " + }, + { + "bbox": [ + 131, + 409, + 479, + 433 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{well - lit}}" + }, + { + "bbox": [ + 131, + 409, + 479, + 433 + ], + "type": "text", + "content": " is a normalized mean of CLIP embeddings of well-lit training images:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 212, + 438, + 480, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 438, + 480, + 469 + ], + "spans": [ + { + "bbox": [ + 212, + 438, + 480, + 469 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {\\text {w e l l - l i t}} = f _ {\\text {n o r m}} \\left(\\frac {\\sum_ {i = 1} ^ {N _ {p}} f _ {\\text {n o r m}} \\left(\\Phi_ {\\text {i m a g e}} \\left(I _ {i} ^ {p}\\right)\\right)}{N _ {p}}\\right) \\tag {13}", + "image_path": "dd12547c0ab026a67cb589c508e158fc1a6946a741df3dea4b9791b12131b552.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 474, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 474, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 474, + 481, + 666 + ], + "type": "text", + "content": "Note that RAVE does not require several repeating stages during training. Before training, the residual vector " + }, + { + "bbox": [ + 131, + 474, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{residual}}" + }, + { + "bbox": [ + 131, + 474, + 481, + 666 + ], + "type": "text", + "content": " is calculated, and afterwards the training of the image enhancement model begins using " + }, + { + "bbox": [ + 131, + 474, + 481, + 666 + ], + "type": "inline_equation", + "content": "L_{rave}" + }, + { + "bbox": [ + 131, + 474, + 481, + 666 + ], + "type": "text", + "content": " as the loss function. This makes RAVE training much more efficient than that of CLIP-LIT and CLIP-LIT-Latent. First, the training of the image enhancement model no longer depends on the quality of learned prompt embeddings, as a fixed residual vector is used for guidance. Also, the iterative manner of CLIP-LIT and CLIP-LIT-Latent training regimes makes training less stable, as the image enhancement model training depends on the quality of prompts obtained on the previous step and vice versa. In supplementary material we show that results obtained on intermediate steps of CLIP-LIT training can be of poor quality and can fluctuate from producing under-exposed to over-exposed resulting images from step to step. In contrast, RAVE has a well-defined fixed initial objective. Results show that RAVE converges up to 25 times faster than CLIP-LIT and CLIP-LIT-Latent. Furthermore, images obtained by RAVE have fewer over-exposed areas, producing high quality results." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 301, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 301, + 91, + 448, + 102 + ], + "type": "text", + "content": "RAVE: Residual Vector Embedding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 481, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 481, + 234 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 481, + 234 + ], + "type": "text", + "content": "Datasets Following CLIP-LIT, we use the BAID [15] and DIV2K [1] datasets for training. We train our models in two settings: using paired and unpaired training data. For the paired setting, we use 380 backlit and corresponding well-lit images from BAID dataset. For the unpaired setting, we use 380 backlit images from the BAID and 384 images from DIV2K dataset. For the unpaired setting, we use the same data as used in CLIP-LIT. For testing, we follow recent literature and use the BAID [15] test dataset, which has 368 backlit and corresponding well-lit images. The data is publicly available." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 235, + 481, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 235, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 235, + 481, + 282 + ], + "type": "text", + "content": "Training and inference We follow parameter settings of the original CLIP-LIT approach when setting parameters for CLIP-LIT-Latent and RAVE (where applicable), as well as for reproducing CLIP-LIT. Specific details of training and values of hyperparameters are presented in supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 283, + 481, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 481, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 481, + 307 + ], + "type": "text", + "content": "Metrics Following recent works in image enhancement, we report four metrics: PSNR [9], SSIM [23], and LPIPS [27] (AlexNet version) and FID [16]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 307, + 481, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 481, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 481, + 367 + ], + "type": "text", + "content": "Methods for comparison The main model which we compare our approaches to is CLIP-LIT [14], as it showed superior performance on all mentioned metrics, except FID, to many prior image enhancement models. We also include recently published methods such as Restormer [26], Retinexformer [3], DiffIR [24] and Diff-Retinex [25] to the comparison." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 383, + 198, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 383, + 198, + 394 + ], + "spans": [ + { + "bbox": [ + 132, + 383, + 198, + 394 + ], + "type": "text", + "content": "4.1 Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 402, + 481, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 481, + 556 + ], + "type": "text", + "content": "Quantitative comparison The quantitative comparison on the BAID test dataset is presented in Tab. 1. We report test set metrics of the best model checkpoint obtained during training for all the models. Detailed metrics values for all the checkpoints after each training epoch can be found in supplementary material. Note that many approaches can only be trained using paired data, that is why their results are only present in the \"paired\" section. We see that our method achieves state-of-the-art performance in both settings. In supplementary material, we also present results on time needed to train all the models, and confirm that both CLIP-LIT-Latent and RAVE are much more efficient. Note that in the setting of unpaired data we have two versions of RAVE, referred as RAVE and RAVE shifted in Tab. 1. RAVE shifted is a variant of RAVE with a simple modification of residual vector, see \"Interpretation of the residual vector\" section below for more details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "content": "Visual comparison For visual comparison, we present representative examples from the BAID test dataset. First, in Fig. 4, we compare CLIP-LIT-Latent and RAVE to original CLIP-LIT. First, note that CLIP-LIT-Latent produces more contrast in general, giving more diverse illumination to different regions. Second, in Fig. 1, we present comparison of images produced by RAVE to images produced by CLIP-LIT. We show that images produced by RAVE have fewer artifacts. We hypothesize that this advantage of RAVE may result from its guiding mechanism: in RAVE we directly point the enhancement model to produce images which have similar CLIP embeddings to that of real well-lit images. This" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 100 + ], + "type": "text", + "content": "T. Gaintseva et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 146, + 477, + 309 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 1: Quantitative comparison of different methods on the BAID test dataset. The best and second best performances in both settings are in bold and underlined." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 146, + 477, + 309 + ], + "lines": [ + { + "bbox": [ + 137, + 146, + 477, + 309 + ], + "spans": [ + { + "bbox": [ + 137, + 146, + 477, + 309 + ], + "type": "table", + "html": "
MethodPSNR ↑SSIM ↑LPIPS ↓FID ↓
Original images16.910.760.2152.47
PairedRestormer21.070.8320.19241.17
Retinexformer22.030.8620.17345.27
DiffIR21.100.8350.17540.35
Diff-Retinex22.070.8610.16038.07
CLIP-LIT21.930.8750.15641.16
CLIP-LIT-Latent (ours)21.840.8770.15542.56
RAVE (ours)22.260.8800.13936.01
UnpairedEnlightenGAN17.550.8640.19643.50
CLIP-LIT21.590.8740.16046.49
CLIP-LIT-Latent (ours)21.450.8770.15646.20
RAVE (ours)20.390.8610.15540.11
RAVE shifted (ours)21.370.8770.14940.90
", + "image_path": "6bccaa16bb1086cfa711d4b078891d3bd1337fe1410e77a0488d0e174c9b0e28.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 331, + 480, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 480, + 427 + ], + "type": "text", + "content": "does not allow model to produce regions with artifacts, as its CLIP embeddings won't match those of real well-lit images. In CLIP-LIT or CLIP-LIT-Latent, on the other hand, the objective of the enhancement model is to produce images for which the CLIP embeddings are simply closer to the positive prompt/vector than to the negative one. This allows the enhancement model to produce less accurate images, CLIP embeddings of which are just \"close enough\" to the CLIP embeddings of real well-lit images. More results of visual comparison of all the models can be found in supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 427, + 480, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 427, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 427, + 480, + 487 + ], + "type": "text", + "content": "Interpretation of the residual vector In RAVE, we build the residual vector " + }, + { + "bbox": [ + 130, + 427, + 480, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\text{residual}}" + }, + { + "bbox": [ + 130, + 427, + 480, + 487 + ], + "type": "text", + "content": " following formula 16. We noted that " + }, + { + "bbox": [ + 130, + 427, + 480, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\text{residual}}" + }, + { + "bbox": [ + 130, + 427, + 480, + 487 + ], + "type": "text", + "content": " is a vector that points to a direction of moving from backlit images to well-lit images in the CLIP embedding space. Here we prove this by interpreting vector " + }, + { + "bbox": [ + 130, + 427, + 480, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\text{residual}}" + }, + { + "bbox": [ + 130, + 427, + 480, + 487 + ], + "type": "text", + "content": " using the fact that CLIP has a common vector space for embeddings of images and text." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 487, + 480, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 480, + 535 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 480, + 535 + ], + "type": "text", + "content": "We go through the vocabulary of the CLIP text encoder and find those vocabulary tokens which have the closest and farthest CLIP embeddings to the " + }, + { + "bbox": [ + 130, + 487, + 480, + 535 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\text{residual}}" + }, + { + "bbox": [ + 130, + 487, + 480, + 535 + ], + "type": "text", + "content": " in the CLIP latent space in terms of cosine similarity. More formally, we use the following metric:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 211, + 546, + 480, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 546, + 480, + 559 + ], + "spans": [ + { + "bbox": [ + 211, + 546, + 480, + 559 + ], + "type": "interline_equation", + "content": "\\operatorname {s i m} (\\text {t o k e n}) = f _ {\\text {n o r m}} \\left(\\Phi_ {\\text {t e x t}} \\left(e _ {\\text {t o k e n}}\\right)\\right) \\cdot \\mathbf {v} _ {\\text {r e s i d u a l}} \\tag {14}", + "image_path": "fdb79e70cdd0dc6293ffc2652f4b96688dec6ed00f96e5a8fd3398e26565893f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 569, + 480, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 480, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 480, + 617 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 130, + 569, + 480, + 617 + ], + "type": "inline_equation", + "content": "e_{token}" + }, + { + "bbox": [ + 130, + 569, + 480, + 617 + ], + "type": "text", + "content": " is embedding of vocabulary token, which is input to the CLIP text encoder " + }, + { + "bbox": [ + 130, + 569, + 480, + 617 + ], + "type": "inline_equation", + "content": "\\varPhi_{text}" + }, + { + "bbox": [ + 130, + 569, + 480, + 617 + ], + "type": "text", + "content": ". We sort vocabulary tokens using sim. Table 2 shows 7 vocabulary tokens which have highest and lowest cosine similarity to the " + }, + { + "bbox": [ + 130, + 569, + 480, + 617 + ], + "type": "inline_equation", + "content": "v_{residual}" + }, + { + "bbox": [ + 130, + 569, + 480, + 617 + ], + "type": "text", + "content": " vectors obtained from the paired dataset and from the unpaired dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "In the case of paired data, many of the least similar tokens have a meaning around \"darkness\": \"silhouette\", \"dark\", \"shadow\", etc., and absolute value of their cosine similarity to " + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{residual}}" + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": " is around " + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "0.12 - 0.15" + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": ". At the same time, most similar tokens do not share any meaningful semantics, and the absolute value" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "text", + "content": "RAVE: Residual Vector Embedding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 170, + 117, + 258, + 176 + ], + "blocks": [ + { + "bbox": [ + 170, + 117, + 258, + 176 + ], + "lines": [ + { + "bbox": [ + 170, + 117, + 258, + 176 + ], + "spans": [ + { + "bbox": [ + 170, + 117, + 258, + 176 + ], + "type": "image", + "image_path": "04c78a3ca080e4259ef79c6fb999e0d808b813d5321854d749f9c4c7e01e525f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 260, + 117, + 349, + 176 + ], + "blocks": [ + { + "bbox": [ + 260, + 117, + 349, + 176 + ], + "lines": [ + { + "bbox": [ + 260, + 117, + 349, + 176 + ], + "spans": [ + { + "bbox": [ + 260, + 117, + 349, + 176 + ], + "type": "image", + "image_path": "6fed4143d3a5ab7d022ff6e187c7f52189a5a9558400e76b513d97774b3c6c6d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 351, + 117, + 440, + 176 + ], + "blocks": [ + { + "bbox": [ + 351, + 117, + 440, + 176 + ], + "lines": [ + { + "bbox": [ + 351, + 117, + 440, + 176 + ], + "spans": [ + { + "bbox": [ + 351, + 117, + 440, + 176 + ], + "type": "image", + "image_path": "54805cd322501140ecc7b5532f9f0fed0865741ba26acfd77b5beea670b81a26.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 170, + 180, + 258, + 240 + ], + "blocks": [ + { + "bbox": [ + 170, + 180, + 258, + 240 + ], + "lines": [ + { + "bbox": [ + 170, + 180, + 258, + 240 + ], + "spans": [ + { + "bbox": [ + 170, + 180, + 258, + 240 + ], + "type": "image", + "image_path": "7427c9ddc16216b987ae04de1d7222ecbe8906eaa1c07db025bd39eb07f93e61.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 246, + 221, + 254 + ], + "lines": [ + { + "bbox": [ + 194, + 246, + 221, + 254 + ], + "spans": [ + { + "bbox": [ + 194, + 246, + 221, + 254 + ], + "type": "text", + "content": "CLIP-LIT" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 260, + 180, + 349, + 240 + ], + "blocks": [ + { + "bbox": [ + 260, + 180, + 349, + 240 + ], + "lines": [ + { + "bbox": [ + 260, + 180, + 349, + 240 + ], + "spans": [ + { + "bbox": [ + 260, + 180, + 349, + 240 + ], + "type": "image", + "image_path": "82348f4401a069e6f67850de01bbcf94810d68ccefe23c4e736cd39733d11bc3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 282, + 246, + 329, + 254 + ], + "lines": [ + { + "bbox": [ + 282, + 246, + 329, + 254 + ], + "spans": [ + { + "bbox": [ + 282, + 246, + 329, + 254 + ], + "type": "text", + "content": "CLIP-LIT-Latent" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 351, + 180, + 440, + 240 + ], + "blocks": [ + { + "bbox": [ + 351, + 180, + 440, + 240 + ], + "lines": [ + { + "bbox": [ + 351, + 180, + 440, + 240 + ], + "spans": [ + { + "bbox": [ + 351, + 180, + 440, + 240 + ], + "type": "image", + "image_path": "83f7b1f3420458904e917c1c9736ec137d30a2b24d35d76fe1ecea2bd0f817ff.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 246, + 399, + 254 + ], + "lines": [ + { + "bbox": [ + 380, + 246, + 399, + 254 + ], + "spans": [ + { + "bbox": [ + 380, + 246, + 399, + 254 + ], + "type": "text", + "content": "RAVE" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 171, + 279, + 258, + 337 + ], + "blocks": [ + { + "bbox": [ + 132, + 257, + 475, + 274 + ], + "lines": [ + { + "bbox": [ + 132, + 257, + 475, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 475, + 274 + ], + "type": "text", + "content": "(a) Visual comparison of results by CLIP-LIT, CLIP-LIT-Latent and RAVE, all trained with paired data. RAVE produces results with more general contrast." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 171, + 279, + 258, + 337 + ], + "lines": [ + { + "bbox": [ + 171, + 279, + 258, + 337 + ], + "spans": [ + { + "bbox": [ + 171, + 279, + 258, + 337 + ], + "type": "image", + "image_path": "ae3f81a5ea7377334f8366ff5e56242a551a3204f359a4edced299113d0d1319.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 260, + 279, + 349, + 337 + ], + "blocks": [ + { + "bbox": [ + 260, + 279, + 349, + 337 + ], + "lines": [ + { + "bbox": [ + 260, + 279, + 349, + 337 + ], + "spans": [ + { + "bbox": [ + 260, + 279, + 349, + 337 + ], + "type": "image", + "image_path": "7bbc3161de51bd7c8cc44a87dfdb5698b1d507e11a605e014e2e802bc262ccec.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 351, + 279, + 439, + 337 + ], + "blocks": [ + { + "bbox": [ + 351, + 279, + 439, + 337 + ], + "lines": [ + { + "bbox": [ + 351, + 279, + 439, + 337 + ], + "spans": [ + { + "bbox": [ + 351, + 279, + 439, + 337 + ], + "type": "image", + "image_path": "f8439baae704b93c0f573f02ca0f5cc3c183d24085937c8076806ca975e784b2.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 171, + 342, + 258, + 400 + ], + "blocks": [ + { + "bbox": [ + 171, + 342, + 258, + 400 + ], + "lines": [ + { + "bbox": [ + 171, + 342, + 258, + 400 + ], + "spans": [ + { + "bbox": [ + 171, + 342, + 258, + 400 + ], + "type": "image", + "image_path": "d3faeb33581fbf8eb85efcc16a0ea4df9766c19a5bff3ced2689c54ef9182ec1.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 407, + 221, + 415 + ], + "lines": [ + { + "bbox": [ + 195, + 407, + 221, + 415 + ], + "spans": [ + { + "bbox": [ + 195, + 407, + 221, + 415 + ], + "type": "text", + "content": "CLIP-LIT" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 418, + 475, + 436 + ], + "lines": [ + { + "bbox": [ + 132, + 418, + 475, + 436 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 475, + 436 + ], + "type": "text", + "content": "(b) Visual comparison of results obtained by CLIP-LIT, CLIP-LIT-Latent and RAVE (shifted version), all trained with unpaired data. RAVE produces results with more general contrast." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 261, + 342, + 349, + 400 + ], + "blocks": [ + { + "bbox": [ + 261, + 342, + 349, + 400 + ], + "lines": [ + { + "bbox": [ + 261, + 342, + 349, + 400 + ], + "spans": [ + { + "bbox": [ + 261, + 342, + 349, + 400 + ], + "type": "image", + "image_path": "480b1de885fcb754ab27679ecca57a84f6f6fabebc2ad223bb49a48239292337.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 283, + 407, + 329, + 415 + ], + "lines": [ + { + "bbox": [ + 283, + 407, + 329, + 415 + ], + "spans": [ + { + "bbox": [ + 283, + 407, + 329, + 415 + ], + "type": "text", + "content": "CLIP-LIT-Latent" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 351, + 342, + 438, + 400 + ], + "blocks": [ + { + "bbox": [ + 351, + 342, + 438, + 400 + ], + "lines": [ + { + "bbox": [ + 351, + 342, + 438, + 400 + ], + "spans": [ + { + "bbox": [ + 351, + 342, + 438, + 400 + ], + "type": "image", + "image_path": "47cc75a10ec8f6532bed5d9d58e8f32a80ba0f60915653e40ed31cb8981ad4f2.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 407, + 399, + 415 + ], + "lines": [ + { + "bbox": [ + 380, + 407, + 399, + 415 + ], + "spans": [ + { + "bbox": [ + 380, + 407, + 399, + 415 + ], + "type": "text", + "content": "RAVE" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 139, + 445, + 470, + 456 + ], + "lines": [ + { + "bbox": [ + 139, + 445, + 470, + 456 + ], + "spans": [ + { + "bbox": [ + 139, + 445, + 470, + 456 + ], + "type": "text", + "content": "Fig. 4: Visual comparison of results by CLIP-LIT, CLIP-LIT-Latent and RAVE." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "of their cosine similarity to " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{residual}}" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " is much smaller, being around " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.01 - 0.04" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ". This supports the hypothesis that " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{residual}}" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " has a meaning of shifting from \"dark\", \"shadowy\" images to \"normal\" images, having average level of lightness. " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{residual}}" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " does not carry any other semantic meaning, as it was calculated using paired dataset, where for each backlit image there is its well-lit analogue, and all the semantics present in the BAID dataset is cancelled as a result of subtraction of the mean well-lit and backlit vectors when constructing " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{residual}}" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ". On the other hand, in the case of unpaired data, we see that the least similar tokens have a meaning around \"asian\": \"busan\", \"beijing\", etc., and the most similar ones have a meaning around \"nature\": \"wildlife\", \"southafrica\", \"countrylife\", etc. Herein, an absolute value of cosine similarities for both sets of tokens if quite high, being around " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.08 - 0.12" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ". This reflects the bias in the unpaired datasets. Indeed, in the BAID dataset there are mostly images of Chinese cities or people, while in DIV2K dataset there are many pictures of nature. Due to the bias, RAVE training with unpaired data becomes less effective, making the enhancement model" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 100 + ], + "type": "text", + "content": "T. Gaintseva et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 157, + 491, + 266 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 2: Vocabulary tokens and their cosine similarities, which have the lowest and highest cosine similarity to the " + }, + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "v_{\\text{residual}}" + }, + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "text", + "content": " vectors calculated for paired and unpaired training datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 157, + 491, + 266 + ], + "lines": [ + { + "bbox": [ + 132, + 157, + 491, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 157, + 491, + 266 + ], + "type": "table", + "html": "
PairedUnpaired
Lowest similarityHighest similarityLowest similarityHighest similarity
silhouette-0.151healthtech0.035busan-0.108wildlifewednesdayday0.116
darkness-0.149whitepaper0.035beijing-0.103southafrica0.103
dark-0.139theeconomist0.030iu-0.100ecuador0.095
webcamtoy-0.139bona0.029guangzhou-0.095patrol0.095
blackand-0.132digitaltransformation0.027incheon-0.094countryfile0.093
vedere-0.131sua0.021shenzhen-0.094tasmanian0.093
shadow-0.130amarketing0.021eunhyuk-0.091womensart0.091
", + "image_path": "83ab36fb5e3305dc222244f2839ffcab08cd43ca30ccfe398fab6cd68d2cd7fc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 288, + 479, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 288, + 479, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 288, + 479, + 312 + ], + "type": "text", + "content": "shift images not in an ideal direction in the CLIP latent space. Nevertheless, RAVE still shows good performance when trained with unpaired data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 312, + 480, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 480, + 347 + ], + "type": "text", + "content": "We also propose a simple modification for correcting the residual vector in the unpaired setting. We simply take " + }, + { + "bbox": [ + 130, + 312, + 480, + 347 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 312, + 480, + 347 + ], + "type": "text", + "content": " most and least similar tokens to form an additional residual vector as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 212, + 368, + 480, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 368, + 480, + 444 + ], + "spans": [ + { + "bbox": [ + 212, + 368, + 480, + 444 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {v} _ {\\mathrm {a d d \\_ r e s i d u a l}} = f _ {\\mathrm {n o r m}} (\\mathbf {v} _ {\\mathrm {c l o s e s t}} - \\mathbf {v} _ {\\mathrm {f a r t h e s t}}) \\\\ \\mathbf {v} _ {\\text {c l o s e s t}} = f _ {\\text {n o r m}} \\left(\\frac {\\sum_ {i = 1} ^ {n} f _ {\\text {n o r m}} \\left(\\Phi_ {\\text {t e x t}} \\left(T _ {i} ^ {c}\\right)\\right)}{n}\\right) \\tag {15} \\\\ \\mathbf {v} _ {\\mathrm {f a r t h e s t}} = f _ {n o r m} \\left(\\frac {\\sum_ {i = 1} ^ {n} f _ {\\mathrm {n o r m}} (\\varPhi_ {\\mathrm {t e x t}} (T _ {i} ^ {f}))}{n}\\right) \\\\ \\end{array}", + "image_path": "5e01c4c46240cf7ca2bfb299991063637bfd022ad3034b29b2b30eebcbf4ff98.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 453, + 479, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 453, + 479, + 478 + ], + "spans": [ + { + "bbox": [ + 131, + 453, + 479, + 478 + ], + "type": "inline_equation", + "content": "T_{i}^{c}" + }, + { + "bbox": [ + 131, + 453, + 479, + 478 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 453, + 479, + 478 + ], + "type": "inline_equation", + "content": "T_{i}^{f}" + }, + { + "bbox": [ + 131, + 453, + 479, + 478 + ], + "type": "text", + "content": " are tokens with " + }, + { + "bbox": [ + 131, + 453, + 479, + 478 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 131, + 453, + 479, + 478 + ], + "type": "text", + "content": " most and least similar embeddings to the " + }, + { + "bbox": [ + 131, + 453, + 479, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{residual}}" + }, + { + "bbox": [ + 131, + 453, + 479, + 478 + ], + "type": "text", + "content": ". And finally we update the " + }, + { + "bbox": [ + 131, + 453, + 479, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{residual}}" + }, + { + "bbox": [ + 131, + 453, + 479, + 478 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 174, + 488, + 480, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 488, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 174, + 488, + 480, + 502 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {\\text {r e s i d u a l}} = \\mathbf {v} _ {\\text {r e s i d u a l}} - c o s \\left(\\mathbf {v} _ {\\text {r e s i d u a l}}, \\mathbf {v} _ {\\text {a d d} \\_ \\text {r e s i d u a l}}\\right) \\cdot \\mathbf {v} _ {\\text {a d d} \\_ \\text {r e s i d u a l}} \\tag {16}", + "image_path": "61177353695079756f660b360ddfdb2c4ce154a2a5333e9df784a07bf724c5ac.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 617 + ], + "type": "text", + "content": "Resulting vector " + }, + { + "bbox": [ + 130, + 510, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{residual}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 617 + ], + "type": "text", + "content": " will keep the meaning of shifting from \"dark\", images to \"normal\" images while losing the unwanted semantics. Experimental results show that RAVE trained with such updated " + }, + { + "bbox": [ + 130, + 510, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{residual}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 617 + ], + "type": "text", + "content": " produces better results (see RAVE shifted in Tab. 1). In supplementary material we present visual comparison of results obtained by original and shifted versions of RAVE, where it can be clearly seen that RAVE shifted produces more well-lit results, while keeping all the advantages of RAVE. There we also present analysis of tokens with the most and least similar embeddings to the updated " + }, + { + "bbox": [ + 130, + 510, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{residual}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 617 + ], + "type": "text", + "content": ". This analysis supports the claim about the meaning of updated " + }, + { + "bbox": [ + 130, + 510, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{residual}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Necessity of the residual vector Given the interpretation of " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{residual}}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", we investigate whether it is necessary to compute " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{residual}}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " using the training data, or is it enough to use CLIP latent vectors of words such as \"dark\", \"darkness\" as a guidance vector for enhancement model. We trained RAVE using " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{residual}} =" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "text", + "content": "RAVE: Residual Vector Embedding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 178, + 146, + 435, + 200 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 3: Mean similarity of CLIP latent vectors of well-lit and back-lit images from BAID and DIV2K datasets to the CLIP latent vectors of vocabulary tokens" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 178, + 146, + 435, + 200 + ], + "lines": [ + { + "bbox": [ + 178, + 146, + 435, + 200 + ], + "spans": [ + { + "bbox": [ + 178, + 146, + 435, + 200 + ], + "type": "table", + "html": "
TokenBAID backlitBAID well-litDIV2K well-lit
silhouette0.2540.2270.223
dark0.2470.2220.232
darkness0.2420.2150.227
", + "image_path": "289014402d3f8e32cb70d26cd6ba198347b2ca0f94013970f37918f6cd011a6a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 220, + 487, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 220, + 487, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 487, + 304 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\mathrm{text}}(token)" + }, + { + "bbox": [ + 130, + 220, + 487, + 304 + ], + "type": "text", + "content": " as guidance vector, with token " + }, + { + "bbox": [ + 130, + 220, + 487, + 304 + ], + "type": "inline_equation", + "content": "\\in" + }, + { + "bbox": [ + 130, + 220, + 487, + 304 + ], + "type": "text", + "content": " {\"silhouette\", \"dark\", \"darkness\"} but did not observe any meaningful enhancement in the resulting images. We suspect that this is due to the fact that latent vectors of backlit and well-lit images have similar cosine similarity to embeddings of these tokens, as presented in Tab 3. This suggests that residual vector " + }, + { + "bbox": [ + 130, + 220, + 487, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{residual}}" + }, + { + "bbox": [ + 130, + 220, + 487, + 304 + ], + "type": "text", + "content": " is not simply a combination of latent vectors of some tokens, but it contains more specific information with the meaning of shifting from backlit to well-lit data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 319, + 218, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 218, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 218, + 331 + ], + "type": "text", + "content": "4.2 Limitations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 338, + 481, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 481, + 445 + ], + "type": "text", + "content": "Despite showing great performance in the task of backlit image enhancement, our models have limitation of non-optimal enhancement of the regions where the loss of information occurs, making restoration ambiguous. Our approach has no generative components and cannot regenerate highly-accurate textures when the information is lost. This leads to sub-optimal performance in such tasks as low-light image enhancement, where images are highly under-exposed (see supplementary material). Moreover, our models are not trained to exactly match pixels of resulting images with ground-truth data, so some sub-optimal lightening of pixels might occur, which results in non-ideal visual quality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 446, + 481, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 481, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 481, + 494 + ], + "type": "text", + "content": "However, existing generative-based models also have limitations. Many of them cannot be trained with unpaired data setting, and require much time and memory to be trained and used. They may also hallucinate detail. As a consequence, the resolution of images it can work with is limited." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 510, + 220, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 220, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 220, + 522 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 665 + ], + "type": "text", + "content": "Based on CLIP-LIT method, we have presented two improved CLIP guidance techniques for backlit image enhancement task, named CLIP-LIT-Latent and RAVE. Our methods show similar or superior performance to that of CLIP-LIT in both cases of training with paired or unpaired data, while having lower time and space requirements. Furthermore, latent vector constructed by RAVE is interpretable and can be used to find biases in the training data, as well as potentially correct them. Moreover, our proposed approach has several advantages over the current SOTA in backlit image enhancement, Diff-Retinex. First, unlike Diff-Retinex, our method can also be trained using unpaired data. Second, it is more lightweight, requiring less compute and space for training and inference, suitable for images with larger resolution." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "text", + "content": "T. Gaintseva et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 482, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 482, + 192 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 482, + 192 + ], + "type": "text", + "content": "Funding for this research was provided by a Google-DeepMind PhD Studentship, and work utilised Queen Mary's Andrena HPC facility, supported by QMUL Research-IT. We also sincerely thank Laida Kushnareva for assisting with images and text composition in the paper." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 213, + 197, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 213, + 197, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 213, + 197, + 226 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 241, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 241, + 480, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 241, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 138, + 241, + 480, + 285 + ], + "type": "text", + "content": "1. Agustsson, E., Timofte, R.: Ntire 2017 challenge on single image super-resolution: Dataset and study. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 1122-1131 (2017). https://doi.org/10.1109/CVPRW.2017.150 10" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 286, + 480, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 286, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 286, + 480, + 308 + ], + "type": "text", + "content": "2. Baykal, A.C., Anees, A.B., Ceylan, D., Erdem, E., Erdem, A., Yuret, D.: Clip-guided stylegan inversion for text-driven real image editing (2023) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 309, + 480, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 309, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 138, + 309, + 480, + 353 + ], + "type": "text", + "content": "3. Cai, Y., Bian, H., Lin, J., Wang, H., Timofte, R., Zhang, Y.: Retinexformer: One-stage retinax-based transformer for low-light image enhancement. In: 2023 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 12470-12479 (2023). https://doi.org/10.1109/ICCV51070.2023.011492, 3, 4, 10" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 354, + 480, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 480, + 397 + ], + "type": "text", + "content": "4. Catley-Chandar, S., Tanay, T., Vandroux, L., Leonardis, A., Slabaugh, G., Pérez-Pellitero, E.: Flexhdr: Modeling alignment and exposure uncertainties for flexible hdr imaging. IEEE Transactions on Image Processing 31, 5923-5935 (2022). https://doi.org/10.1109/TIP.2022.3203562 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 399, + 480, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 480, + 420 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 480, + 420 + ], + "type": "text", + "content": "5. Cheng, H.D., Shi, X.: A simple and effective histogram equalization approach to image enhancement. Digital signal processing 14(2), 158-170 (2004) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 421, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 421, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 421, + 480, + 464 + ], + "type": "text", + "content": "6. Gal, R., Patashnik, O., Maron, H., Bermano, A.H., Chechik, G., Cohen-Or, D.: Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Trans. Graph. 41(4) (jul 2022). https://doi.org/10.1145/3528223.3530164, https://doi.org/10.1145/3528223.3530164 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 465, + 480, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 465, + 480, + 518 + ], + "spans": [ + { + "bbox": [ + 138, + 465, + 480, + 518 + ], + "type": "text", + "content": "7. Goloujeh, A.M., Smith, J., Magerko, B.: Explainable clip-guided 3d-scene generation in an ai holodeck. In: Proceedings of the Eighteenth AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment. AIIDE'22, AAAI Press (2022). https://doi.org/10.1609/aide.v18i1.21973, https://doi.org/10.1609/aide.v18i1.21973 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 520, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 520, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 138, + 520, + 480, + 564 + ], + "type": "text", + "content": "8. Guo, C., Li, C., Guo, J., Loy, C.C., Hou, J., Kwong, S., Cong, R.: Zero-reference deep curve estimation for low-light image enhancement. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1777-1786 (2020). https://doi.org/10.1109/CVPR42600.2020.001852, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 565, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 565, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 565, + 480, + 597 + ], + "type": "text", + "content": "9. Horé, A., Ziou, D.: Image quality metrics: Psnr vs. sSIM. In: 2010 20th International Conference on Pattern Recognition. pp. 2366-2369 (2010). https://doi.org/10.1109/ICPR.2010.579.10" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "type": "text", + "content": "0. Land, E.H.: The retina theory of color vision. Scientific american 237(6), 108-129 (1977) 2, 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "type": "text", + "content": "1. Li, C., Guo, C., Han, L., Jiang, J., Cheng, M., Gu, J., Loy, C.: Low-light image and video enhancement using deep learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(12), 9396-9416 (dec 2022). https://doi.org/10.1109/TPAMI.2021.3126387 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "text", + "content": "RAVE: Residual Vector Embedding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 159 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 159 + ], + "type": "text", + "content": "12. Li, C., Guo, C., Loy, C.C.: Learning to enhance low-light image via zero-reference deep curve estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(8), 4225-4238 (2022). https://doi.org/10.1109/TPAMI.2021.3063604 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 481, + 192 + ], + "type": "text", + "content": "13. Li, Z., Wu, X.: Learning-based restoration of backlit images. IEEE Transactions on Image Processing 27(2), 976-986 (2018). https://doi.org/10.1109/TIP.2017.2771142 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 236 + ], + "type": "text", + "content": "14. Liang, Z., Li, C., Zhou, S., Feng, R., Loy, C.C.: Iterative prompt learning for unsupervised backlit image enhancement. In: 2023 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 8060-8069 (2023). https://doi.org/10.1109/ICCV51070.2023.00743 2, 4, 10" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 237, + 481, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 292 + ], + "type": "text", + "content": "15. Lv, X., Zhang, S., Liu, Q., Xie, H., Zhong, B., Zhou, H.: Backlitnet: A dataset and network for backlit image enhancement. Computer Vision and Image Understanding 218, 103403 (2022). https://doi.org/https://doi.org/10.1016/j.cviu.2022.103403, https://www.sciencedirect.com/science/article/pii/S1077314222000340 4, 10" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 293, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 346 + ], + "type": "text", + "content": "16. Naeem, M.F., Oh, S.J., Uh, Y., Choi, Y., Yoo, J.: Reliable fidelity and diversity metrics for generative models. In: III, H.D., Singh, A. (eds.) Proceedings of the 37th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 119, pp. 7176-7185. PMLR (13-18 Jul 2020), https://proceedings.mlr.press/v119/naeem20a.html 10" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "17. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models (12 2021) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 380, + 481, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 434 + ], + "type": "text", + "content": "18. Parelli, M., Delitzas, A., Hars, N., Vlassis, G., Anagnostidis, S., Bachmann, G., Hofmann, T.: Clip-guided vision-language pre-training for question answering in 3d scenes. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 5607-5612 (2023). https://doi.org/10.1109/CVPRW59228.2023.005934" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 434, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 481, + 489 + ], + "type": "text", + "content": "19. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (2021), https://apisemantic scholar.org/CorpusID: 231591445 2, 4, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "type": "text", + "content": "20. Reinhard, E.: High Dynamic Range Imaging, pp. 1-6. Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-03243-2_843-1, https://doi.org/10.1007/978-3-030-03243-2_843-1 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 522, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 481, + 566 + ], + "type": "text", + "content": "21. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Navab, N., Hornegger, J., Wells, W.M., Frangi, A.F. (eds.) Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015. pp. 234-241. Springer International Publishing, Cham (2015) 6, 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 567, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 610 + ], + "type": "text", + "content": "22. Singh, K., Seth, A., Sandhu, H.S., Samdani, K.: A comprehensive review of convolutional neural network based image enhancement techniques. In: 2019 IEEE International Conference on System, Computation, Automation and Networking (ICSCAN). pp. 1-6 (2019). https://doi.org/10.1109/ICSCAN.2019.8878706 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "text", + "content": "23. Wang, Z., Bovik, A., Sheikh, H., Simoncelli, E.: Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing 13(4), 600-612 (2004). https://doi.org/10.1109/TIP.2003.819861 10" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "24. Xia, B., Zhang, Y., Wang, S., Wang, Y., Wu, X., Tian, Y., Yang, W., Gool, L.V.: Diffir: Efficient diffusion model for image restoration. In: 2023 IEEE/CVF" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 100 + ], + "type": "text", + "content": "T. Gaintseva et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 357 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 149, + 116, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 116, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 149, + 116, + 480, + 159 + ], + "type": "text", + "content": "International Conference on Computer Vision (ICCV). pp. 13049-13059. IEEE Computer Society, Los Alamitos, CA, USA (oct 2023). https://doi.org/10.1109/ICCV51070.2023.01204, https://doi.ieeecomputersociety.org/10.1109/ICCV51070.2023.01204 10" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 481, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 481, + 204 + ], + "type": "text", + "content": "25. Yi, X., Xu, H., Zhang, H., Tang, L., Ma, J.: Diff-retinex: Rethinking low-light image enhancement with a generative diffusion model. In: 2023 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 12268-12277 (2023). https://doi.org/10.1109/ICCV51070.2023.011304, 10" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 247 + ], + "type": "text", + "content": "26. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.: Restormer: Efficient transformer for high-resolution image restoration. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5718-5729 (2022). https://doi.org/10.1109/CVPR52688.2022.005642, 3, 4, 10" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 248, + 481, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 248, + 481, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 248, + 481, + 292 + ], + "type": "text", + "content": "27. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 586-595 (2018). https://doi.org/10.1109/CVPR.2018.0006810" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 293, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 357 + ], + "type": "text", + "content": "28. Zheng, B., Pan, X., Zhang, H., Zhou, X., Slabaugh, G., Yan, C., Yuan, S.: Domainplus: Cross transform domain learning towards high dynamic range imaging. In: Proceedings of the 30th ACM International Conference on Multimedia. p. 1954-1963. MM '22, Association for Computing Machinery, New York, NY, USA (2022). https://doi.org/10.1145/3503161.3547823, https://doi.org/10.1145/3503161.3547823 4" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "text", + "content": "RAVE: Residual Vector Embedding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/RAW-Adapter_ Adapting Pretrained Visual Model to Camera RAW Images/a7bce0f0-b64e-4428-b3dc-7b7e2a217569_content_list.json b/2024/RAW-Adapter_ Adapting Pretrained Visual Model to Camera RAW Images/a7bce0f0-b64e-4428-b3dc-7b7e2a217569_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b96c2a1e1819be5393ba6bfbd9d6a67848a9714c --- /dev/null +++ b/2024/RAW-Adapter_ Adapting Pretrained Visual Model to Camera RAW Images/a7bce0f0-b64e-4428-b3dc-7b7e2a217569_content_list.json @@ -0,0 +1,2039 @@ +[ + { + "type": "text", + "text": "RAW-Adapter: Adapting Pre-trained Visual Model to Camera RAW Images", + "text_level": 1, + "bbox": [ + 248, + 140, + 754, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziteng Cui $^{1}$ and Tatsuya Harada $^{1,2}$", + "bbox": [ + 361, + 210, + 643, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 The University of Tokyo", + "bbox": [ + 411, + 239, + 589, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 RIKEN AIP", + "bbox": [ + 450, + 253, + 550, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{cui, harada}@mi.t.u-tokyo.ac.jp", + "bbox": [ + 377, + 268, + 625, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. sRGB images are now the predominant choice for pre-training visual models in computer vision research, owing to their ease of acquisition and efficient storage. Meanwhile, the advantage of RAW images lies in their rich physical information under variable real-world challenging lighting conditions. For computer vision tasks directly based on camera RAW data, most existing studies adopt methods of integrating image signal processor (ISP) with backend networks, yet often overlook the interaction capabilities between the ISP stages and subsequent networks. Drawing inspiration from ongoing adapter research in NLP and CV areas, we introduce RAW-Adapter, a novel approach aimed at adapting sRGB pre-trained models to camera RAW data. RAW-Adapter comprises input-level adapters that employ learnable ISP stages to adjust RAW inputs, as well as model-level adapters to build connections between ISP stages and subsequent high-level networks. Additionally, RAW-Adapter is a general framework that could be used in various computer vision frameworks. Abundant experiments under different lighting conditions have shown our algorithm's state-of-the-art (SOTA) performance, demonstrating its effectiveness and efficiency across a range of real-world and synthetic datasets. Code is available at this url", + "bbox": [ + 259, + 320, + 743, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 611, + 375, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, there has been a growing interest in revisiting vision tasks using unprocessed camera RAW images. How to leverage information-rich RAW image in computer vision tasks has become a current research hotspot in various sub-areas (i.e. denoising [27], view synthesis [46], object detection [27, 48]). Compared to commonly used sRGB image, RAW image directly acquired by the camera sensor, encompasses abundant information unaffected or compressed by image signal processor (ISP), also offers physically meaningful information like noise distributions [47, 68], owing to its linear correlation between image intensity and the radiant energy received by a camera. The acquisition of RAW data, facilitating enhanced detail capture and a higher dynamic range, imparts a unique advantage in addressing visual tasks under variable lighting conditions in the real world. For instance, sunlight irradiance can reach as high as $1.3 \\times 10^{3}W / m^{2}$ while bright planet irradiance can be as low as $2.0 \\times 10^{-6}W / m^{2}$ [32].", + "bbox": [ + 212, + 643, + 787, + 843 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d6f0e038f96a0dc83c9d3e0b6589ecb2e58610a9cbea038010c1e920d4bd3760.jpg", + "image_caption": [ + "Fig. 1: (a). An overview of basic image signal processor (ISP) pipeline. (b). ISP and current visual model have different objectives. (c) Previous methods optimize ISP with down-stream visual model. (d) Our proposed RAW-Adapter." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 787, + 321 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Meanwhile, sRGB has emerged as the primary choice for pre-training visual models in today's computer vision field, due to its scalability and ease of storage. Typically, sRGB images are derived from camera RAW data through the ISP pipeline. As shown in Fig. 1(a), the entire ISP pipeline consists of multiple modules to convert RAW images into vision-oriented sRGB images, each of these modules serves its own distinct purpose, with the majority being ill-posed and heavily dependent on prior information [5,16,26,35,42,56].", + "bbox": [ + 212, + 410, + 787, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "When adopt RAW data for computer vision tasks, the purpose of a manually designed ISP is to produce images that offer a superior visual experience [18,69], rather than optimizing for downstream visual tasks such as object detection or semantic segmentation (see Fig. ①b)). Additionally, most companies' ISPs are black boxes, making it difficult to obtain information about the specific steps inside. Consequently, utilizing human vision-oriented ISP in certain conditions is sometimes even less satisfactory than directly using RAW [22,27,40,81].", + "bbox": [ + 212, + 518, + 787, + 627 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To better take advantage of camera RAW data for various computer vision tasks. Researchers began to optimize the image signal processor (ISP) jointly with downstream networks. Since most ISP stages are non-differentiable and cannot be jointly backpropagation, there are two main lines of approaches to connect ISP stages with downstream networks (see Fig. 1(c)): ①. First-kind approaches maintain the modular design of the traditional ISP, involving the design of differentiable ISP modules through optimization algorithms 48,69,75, such as Hardware-in-the-loop 48 adopt covariance matrix adaptation evolution strategy (CMA-ES) 23. ②. Second-kind approaches replaced the ISP part entirely with a neural network 18,55,72,73, such as Dirty-Pixel 18 adopt a stack of residual UNets 58 as pre-encoder, however, the additional neural network introduces a significant computational burden, especially when dealing with high-resolution inputs. Beyond these, the above-mentioned methods still treat ISP and the backend neural network as two independent modules, lacking interac", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Cui, Harada.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4d3460696f9b178ee2425f48e832603f1c4e3961befcd0d770a0d98282388dcd.jpg", + "image_caption": [ + "Fig. 2: Performance of RAW-based visual tasks with and without sRGB pre-trained weights. We analyze two methods: Dirty-Pixel [18] and RAW-Adapter. Blue line represents trained with MS COCO [44] pre-train weights, the purple line indicates ImageNet [17] pre-train weights, and the yellow line signifies training from scratch." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 500, + 255 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d5317e4de476aa991ab27610f28f1e9956b0cfc91f38b155345715b96886099b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 143, + 787, + 255 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion capabilities between two separate models. Especially, current visual models predominantly pre-trained on large amounts of sRGB images, there exists a significant gap between pre-trained models and camera RAW images. As it shown in Fig. 2 training from scratch with RAW data can significantly impair the performance of RAW-based vision tasks. How to leveraging sRGB pre-trained weights proves to be highly beneficial for RAW image visual tasks. Hence, here we ask: Is there a more effective way to adapt the information-rich RAW data into the knowledge-rich sRGB pre-trained models?", + "bbox": [ + 212, + 358, + 787, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our solution, RAW-Adapter, differs from previous approaches that employed complex ISP or deeper neural networks at the input level. Instead, we prioritize simplifying the input stages and enhancing connectivity between ISP and subsequent networks at the model level. Inspired by recent advancements in prompt learning and adapter tuning [8,28,33,54], we developed two novel adapter approaches aimed at enhancing the integration between RAW input and sRGB pretrained models. Leveraging priors from ISP stages, our method includes input-level adapters and model-level adapters, as illustrated in Fig. [1(d)]. Input-level adapters are designed to adapt the input from RAW data to the backend network input, we follow traditional ISP design and predict ISP stages' key parameters to accommodate RAW image into the specific down-stream vision tasks, in this part, we employ strategies involving Query Adaptive Learning (QAL) and Implicit Neural Representation (INR) to ensure the differentiability of the core ISP processes while also upholding a lightweight design. Model-level adapters leverage prior knowledge from the input-level adapters by extracting intermediate stages as features, these features are then embedded into the downstream network, thus the network incorporates prior knowledge from the ISP stages, then jointly contributes to final machine-vision perception tasks. Our contributions could be summarized as follows:", + "bbox": [ + 212, + 481, + 787, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We introduce RAW-Adapter, a novel framework aimed at improving the integration of sRGB pre-trained models to camera RAW images. Through the implementation of two types of adapters, we strive to narrow the dispar", + "bbox": [ + 223, + 794, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "RAW-Adapter", + "bbox": [ + 633, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ity between RAW images and sRGB models, tackling discrepancies at both input and model levels.", + "bbox": [ + 240, + 146, + 784, + 174 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- By analyzing camera ISP steps, we've crafted input-level adapters that integrate query adaptive learning (QAL) and implicit neural representations (INR) to optimize ISP key parameters. Additionally, harnessing prior input stage information informs model-level adapters, enriching model understanding and improving downstream task performance.", + "bbox": [ + 225, + 176, + 782, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- We conducted detection and segmentation experiments involving different lighting scenarios, including normal-light, dark, and over-exposure scenes. Through comparisons with current mainstream ISPs and joint-training methods, sufficient experiments demonstrate that our algorithm achieved state-of-the-art (SOTA) performance.", + "bbox": [ + 223, + 253, + 782, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 352, + 395, + 368 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Image Signal Processor", + "text_level": 1, + "bbox": [ + 215, + 385, + 454, + 401 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Typically, a camera's Image Signal Processor (ISP) pipeline is required to reconstruct a high-quality sRGB image from camera RAW data, traditional ISP pipeline is formulated as a series of manually crafted modules executed sequentially [16,35,51,56], including some representative steps such as demosaicing, white balance, noise removal, tone mapping, color space transform and so on. There also exist alternative designs of ISP process such as Heide et al. [26] designed FlexISP which combine numerous ISP blocks to a joint optimization block, and Hasinoff et al. [24] modified some of the traditional ISP steps for burst photography under extreme low-light conditions. When it comes to deep learning era, Chen et al. [6] proposed SID to use a UNet [58] replace the traditional ISP steps, which translate low-light RAW data to normal-light sRGB images, and Hu et al. [29] proposed a white-box solution with eight differentiable filters. After that, many efforts began to substitute the traditional ISP process with a neural network, such as [6,10,30,34,37,42,60,62,79], however deep network models are constrained by the training dataset, leading to shortcomings in generalization performance. Meanwhile another line of work focuses on translate sRGB images back to RAW data [2,39,49,50,67,71,76]. Apart from the conversion between RAW and RGB, there is ongoing research specifically dedicated to harnessing RAW images for downstream computer vision tasks.", + "bbox": [ + 212, + 411, + 787, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Computer Vision based on RAW data", + "text_level": 1, + "bbox": [ + 215, + 722, + 578, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to effectively leverage information in RAW images for downstream visual tasks and save the time required for ISP, some early methods proposed to directly perform computer vision tasks on RAW images [3,81], which lacks consideration for the camera noise introduced during the conversion from photons to RAW images, especially in low-light conditions [41,47,68]. Meanwhile training from scratch on RAW data would abandon the current large-scale pre-trained", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Cui, Harada.", + "bbox": [ + 271, + 114, + 359, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "visual model on sRGB data (see Fig. 2), especially the quantity of RAW image datasets (i.e. 4259 images in PASCAL RAW 52 dataset and 2230 images in LOD 27 dataset) is incomparable to current RGB datasets (i.e. over 1M images in ImageNet 17 dataset and SAM 38 is trained with 11M images).", + "bbox": [ + 212, + 145, + 782, + 207 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Most subsequent research focuses on finding ways to integrate ISP and backend computer vision models [18,22,48,55,69,72,73,75]. For example Wu et al. [69] first propose VisionISP and emphasized the difference between human vision and machine vision, they introduced several trainable modules in the ISP to enhance backend object detection performance. Since then several methods attempt to replace the existing non-differentiable ISP with a differentiable ISP network [48,75] or directly using encoder networks to replace the ISP process [18,55,73], however training two consecutive networks simultaneously can result in a significant consumption of computational resources, previous work also rarely focused on how to fine-tune the current mainstream sRGB visual models more efficiently on RAW data. Additionally, research like Rawgment [74], focusing on achieving realistic data augmentation directly on RAW images, some studies address the decrease in computer vision performance caused by specific steps within the incamera process or ISP, such as white balance error [1], auto-exposure error [53], camera motion blur [59] and undesirable camera noise [15,21].", + "bbox": [ + 212, + 207, + 787, + 435 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.3 Adapters in Computer Vision", + "text_level": 1, + "bbox": [ + 214, + 455, + 504, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adapters have become prevalent NLP area, which introducing new modules in large language model (LLM) such as [28,63], these modules enable task-specific fine-tuning, allowing pre-trained models to swiftly adapt to downstream NLP tasks. In the computer vision area, adapters have been adopted in various areas such as incremental learning [31] and domain adaptation [57]. Recently, a series of adapters have been used to investigate how to better utilize pre-trained visual models [8,33,36] or pre-trained vision-language models [65,77]. These methods focus on utilizing prior knowledge to make pre-trained models quickly adapt to downstream tasks. Here, we introduce the RAW-Adapter to further improve the alignment between RAW inputs and sRGB pre-trained visual models.", + "bbox": [ + 212, + 479, + 787, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 RAW-Adapter", + "text_level": 1, + "bbox": [ + 214, + 654, + 392, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we introduce RAW-Adapter. Due to page limitation, an overview of the conventional image signal processor (ISP) in supplementary's Sec. A. For specific adapter designs, we selectively omit certain steps and focus on some key steps within the camera ISP. In Sec. 3.1 we introduce RAW-Adapter's input-level adapters, followed by an explanation of its model-level adapters in Sec. 3.2", + "bbox": [ + 212, + 686, + 784, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Input-level Adapters", + "text_level": 1, + "bbox": [ + 214, + 784, + 433, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 3(a) provides an overview of RAW-Adapter. The solid lines in Fig. 3(a) left represent input-level adapters. Input-level adapters are designed to convert", + "bbox": [ + 212, + 808, + 784, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "RAW-Adapter", + "bbox": [ + 633, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/faa998769cf87654cfecc86abd9f3911d3467d37db1a38152e4e333ddb53cca3.jpg", + "image_caption": [ + "Fig. 3: (a). Structure of RAW-Adapter. Solid line in left denotes input-level adapter's workflow and dotted line denotes model-level adapter's workflow, stage $1\\sim 4$ means different stage of visual model backbone. (b). Detailed structure of kernel & matrix predictors $\\mathbb{P}_{\\mathbb{K}}$ , $\\mathbb{P}_{\\mathbb{M}}$ . (c). Detailed structure of model-level adapter $\\mathbb{M}$ 's merge block." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 588, + 396 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6e4744b439528819ac969d3a50ece88406b7bd2cce496183f9e7b297ad424fee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 145, + 787, + 400 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the RAW image $\\mathbf{I}_1$ into the machine-vision oriented image $\\mathbf{I}_5$ . This process involves digital gain & denoise, demosacing, white balance adjustment, camera color matrix, and color manipulation.", + "bbox": [ + 212, + 498, + 787, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We maintain the ISP design while simultaneously using query adaptive learning (QAL) to estimate key parameters in ISP stages. The QAL strategy is motivated by previous transformer models [4,9,13] and detailed structure is shown in Fig. 3(b), input image $\\mathbf{I}_{i\\in (1,2)}$ would pass by 2 down-scale convolution blocks to generate feature, then feature pass by 2 linear layers to generate attention block's key $\\mathbb{k}$ and value $\\mathbb{V}$ , while query $\\mathbb{q}$ is a set of learnable dynamic parameters, the ISP parameters would be predicted by self-attention calculation [66]:", + "bbox": [ + 212, + 545, + 789, + 652 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np a r a m e t e r s = F F N \\left(\\operatorname {s o f t m a x} \\left(\\frac {\\mathbb {Q} \\cdot \\mathbb {k} ^ {T}}{\\sqrt {d _ {k}}}\\right) \\cdot \\mathbb {V}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 664, + 785, + 699 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $FFN$ denotes the feed-forward network, includes 2 linear layers and 1 activation layer, the predicted parameters would keep the same length as query q. We defined 2 QAL blocks $\\mathbb{P}_{\\mathbb{K}}$ and $\\mathbb{P}_{\\mathbb{M}}$ to predict different part parameters.", + "bbox": [ + 212, + 705, + 787, + 751 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The input RAW image $\\mathbf{I}_1$ would first go through the pre-process operations and a demosaing stage [16,35], followed by subsequent ISP processes:", + "bbox": [ + 212, + 752, + 787, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Gain & Denoise: Denoising algorithms always take various factors into account, such as input photon number, ISO gain level, exposure settings. Here we", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Cui, Harada.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "first use QAL block $\\mathbb{P}_{\\mathbb{K}}$ to predict a gain ratio $g$ [2,6] to adapt $\\mathbf{I}_1$ in different lighting scenarios, followed by an adaptive anisotropic Gaussian kernel to suppress noise under various noise conditions, $\\mathbb{P}_{\\mathbb{K}}$ will predict the appropriate Gaussian kernel $k$ for denoising to improve downstream visual tasks' performance, the predicted key parameters are the Gaussian kernel's major axis $r_1$ , minor axis $r_2$ (see Fig. 4 left), here we set the kernel angle $\\theta$ to 0 for simplification. Gaussian kernel $k$ at pixel $(x,y)$ would be:", + "bbox": [ + 214, + 145, + 787, + 252 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nk (x, y) = \\exp \\left(- \\left(b _ {0} x ^ {2} + 2 b _ {1} x y + b _ {2} y ^ {2}\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 266, + 784, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where:", + "bbox": [ + 215, + 287, + 266, + 299 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nb _ {0} = \\frac {\\cos (\\theta) ^ {2}}{2 r _ {1} ^ {2}} + \\frac {\\sin (\\theta) ^ {2}}{2 r _ {2} ^ {2}}, b _ {1} = \\frac {\\sin (2 \\theta)}{4 r _ {1} ^ {2}} ((\\frac {r _ {1}}{r _ {2}}) ^ {2} - 1), b _ {2} = \\frac {\\sin (\\theta) ^ {2}}{2 r _ {1} ^ {2}} + \\frac {\\cos (\\theta) ^ {2}}{2 r _ {2} ^ {2}}. (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 305, + 784, + 338 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After gain ratio $g$ and kernel $k$ process on image $\\mathbf{I}_1$ , $\\mathbb{P}_{\\mathbb{K}}$ also predict a filter parameter $\\sigma$ (initial at 0) to keep the sharpness and recover details of generated image $\\mathbf{I}_2$ . Eq. 4 shows the translation from $\\mathbf{I}_1$ to $\\mathbf{I}_2$ , where $*$ denotes kernel convolution and filter parameter $\\sigma$ is limited in a range of $(0, 1)$ by a Sigmoid activation. For more details please refer to our supplementary part Sec. C", + "bbox": [ + 214, + 342, + 782, + 419 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {P} _ {\\mathbb {K}} \\left(\\mathbf {I} _ {1}, \\mathrm {q}\\right)\\rightarrow k \\left\\{r _ {1}, r _ {2}, \\theta \\right\\}, g, \\sigma ,\n$$\n", + "text_format": "latex", + "bbox": [ + 397, + 430, + 604, + 445 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} _ {2} ^ {\\prime} = (g \\cdot \\mathbf {I} _ {1}) * k, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 397, + 449, + 784, + 464 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} _ {2} = \\mathbf {I} _ {2} ^ {\\prime} + (g \\cdot \\mathbf {I} _ {1} - \\mathbf {I} _ {2} ^ {\\prime}) \\cdot \\sigma .\n$$\n", + "text_format": "latex", + "bbox": [ + 397, + 467, + 573, + 483 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "White Balance & CCM Matrix: White balance (WB) mimics the color constancy of the human vision system (HVS) by aligning \"white\" color with the white object, resulting in the captured image reflecting the combination of light color and material reflectance [1,2,14]. In our work, we hope to find an adaptive white balance for different images under various lighting scenarios. Motivated by the design of Shades of Gray (SoG) [20] WB algorithm, where gray-world WB and Max-RGB WB can be regarded as a subcase, we've employed a learnable parameter $\\rho$ to replace the gray-world's L-1 average with an adaptive Minkowski distance average (see Eq. [5]), the QAL block $\\mathbb{P}_{\\mathbb{M}}$ predicts Minkowski distance's hyper-parameter $\\rho$ , a demonstration of various $\\rho$ is shown in Fig. [4]. After finding the suitable $\\rho$ , $\\mathbf{I}_2$ would multiply to white balance matrix to generate $\\mathbf{I}_3$ .", + "bbox": [ + 214, + 500, + 787, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Color Conversion Matrix (CCM) within the ISP is constrained by the specific camera model. Here we standardize the CCM as a single learnable $3 \\times 3$ matrix, initialized as a unit diagonal matrix $\\mathbf{E}_3$ . The QAL block $\\mathbb{P}_{\\mathbb{M}}$ predicts 9 parameters here, which are then added to $\\mathbf{E}_3$ to form the final $3 \\times 3$ matrix $\\mathbf{E}_{ccm}$ . Then $\\mathbf{I}_3$ would multiply to CCM $\\mathbf{E}_{ccm}$ to generate $\\mathbf{I}_4$ , equation as follows:", + "bbox": [ + 214, + 667, + 787, + 743 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {P} _ {\\mathbb {M}} \\left(\\mathbf {I} _ {2}, \\mathrm {q}\\right)\\rightarrow \\rho , \\mathbf {E} _ {c c m},\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 755, + 501, + 768 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nm _ {i \\in (r, g, b)} = \\sqrt [ r ]{a v g \\left(\\mathbf {I} _ {2} (i) ^ {\\rho}\\right)} / \\sqrt [ r ]{a v g \\left((\\mathbf {I} _ {2}) ^ {\\rho}\\right)},\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 775, + 640, + 792 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} _ {3} = \\mathbf {I} _ {2} * \\left[ \\begin{array}{c c c} m _ {r} & & \\\\ & m _ {g} & \\\\ & & m _ {b} \\end{array} \\right], \\quad \\mathbf {I} _ {4} = \\mathbf {I} _ {3} * \\mathbf {E} _ {c c m}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 792, + 784, + 843 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "RAW-Adapter", + "bbox": [ + 633, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/99aec0a1c43353c69e92d47427f2ed61936cfc1398fcfdab3e2c65b23660d24d.jpg", + "image_caption": [ + "Fig. 4: Left, we use query adaptive learning (QAL) to predict key parameters is ISP process. Right, we show RAW-Adapter different blocks' parameter." + ], + "image_footnote": [], + "bbox": [ + 233, + 143, + 772, + 262 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Color Manipulation Process: In ISP, color manipulation process is commonly achieved through lookup table (LUT), such as 1D and 3D LUTs [16]. Here, we consolidate color manipulation operations into a single 3D LUT, adjusting the color of image $\\mathbf{I}_4$ to produce output image $\\mathbf{I}_5$ . Leveraging advancements in LUT techniques, we choose the latest neural implicit 3D LUT (NILUT) [11], for its speed efficiency and ability to facilitate end-to-end learning in our pipeline [3]. We denote NILUT [11] as $\\mathbb{L}$ , $\\mathbb{L}$ maps the input pixel intensities $R,G,B$ to a continuous coordinate space, followed by the utilization of implicit neural representation (INR) [61], which involves using a multi-layer perceptron (MLP) network to map the output pixel intensities to $R^{\\prime},G^{\\prime},B^{\\prime}$ :", + "bbox": [ + 212, + 330, + 787, + 482 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} _ {5} \\left(R ^ {\\prime}, G ^ {\\prime}, B ^ {\\prime}\\right) = \\mathbb {L} \\left(\\mathbf {I} _ {4} (R, G, B)\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 497, + 785, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Image $\\mathbf{I}_5$ obtained through image-level adapters will be forwarded to the downstream network's backbone. Furthermore, $\\mathbf{I}_5$ 's features obtained in backbone will be fused with model-level adapters, which we will discuss in detail.", + "bbox": [ + 212, + 521, + 787, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.2 Model-level Adapters", + "text_level": 1, + "bbox": [ + 214, + 590, + 439, + 607 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Input-level adapters guarantee the production of machine vision-oriented image $\\mathbf{I}_5$ for high-level models. However, information at the ISP stages ( $\\mathbf{I}_1 \\sim \\mathbf{I}_4$ ) is almost overlooked. Inspired by adapter design in current NLP and CV area [8,28,33], we employ the prior information from the ISP stage as model-level adapters to guide subsequent models' perception. Additionally, model-level adapters promote a tighter integration between downstream tasks and the ISP stage. Dotted lines in Fig. 3(a) represent model-level adapters.", + "bbox": [ + 212, + 618, + 787, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Model-level adapters $\\mathbb{M}$ integrate the information from ISP stages to the network backbone. As shown in Fig. 3(a), we denote the stage $1\\sim 4$ as different stages in network backbone, image $\\mathbf{I}_5$ would pass by backbone stages $1\\sim 4$ and then followed by detection or segmentation head. We utilize convolution layer $c_{i}$ to extract features from $\\mathbf{I}_1$ through $\\mathbf{I}_4$ , these extracted features are concatenated as $\\mathbb{C}(\\mathbf{I}_{1\\sim 4}) = \\mathbb{C}(c_1(\\mathbf{I}_1),c_2(\\mathbf{I}_2),c_3(\\mathbf{I}_3),c_4(\\mathbf{I}_4))$ . Subsequently, $\\mathbb{C}(\\mathbf{I}_{1\\sim 4})$ goes through", + "bbox": [ + 212, + 724, + 787, + 816 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Cui, Harada.", + "bbox": [ + 271, + 114, + 359, + 128 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "3 To save memory occupy, we change the channel number 128 in [11] to 32.", + "bbox": [ + 217, + 823, + 723, + 842 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "two residual blocks [25] and generate adapter $f_{1}$ , $f_{1}$ would merge with stage 1's feature by a merge block, detail structure of merge block is shown in Fig. 3(c), which roughly includes concatenate process and a residual connect, finally merge block would output network feature for stage 2 and an additional adapter $f_{2}$ . Then process would repeat in stage 2 and stage 3 of network backbone. We collectively refer to all structures related to model-level adapters as $\\mathbb{M}$ .", + "bbox": [ + 212, + 145, + 782, + 236 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We show RAW-Adapter's different part parameter number in Fig. 4 right, including input-level adapters $\\{\\mathbb{P}_{\\mathbb{K}}$ (37.57K), $\\mathbb{P}_{\\mathbb{M}}$ (37.96K), $\\mathbb{L}$ (1.97K) and model level adapters $\\mathbb{M}$ (114.87K $\\sim$ 687.59K), model level adapters' parameter number depend on following network. Total parameter number of RAW-Adapter is around 0.2M to 0.8M, much smaller than following backbones ( $\\sim$ 25.6M in ResNet-50 [25], $\\sim$ 197M in Swin-L [45]), also smaller than previous SOTA methods like SID [6] (11.99M) and Dirty-Pixel [18] (4.28M). Additionally, subsequent experiments will demonstrate that the performance of RAW-Adapter on backbone networks with fewer parameters is even superior to that of previous algorithms on backbone networks with higher parameters.", + "bbox": [ + 212, + 237, + 787, + 388 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 415, + 375, + 431 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Dataset and Experimental Setting", + "text_level": 1, + "bbox": [ + 215, + 450, + 542, + 465 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Dataset. We conducted experiments on object detection and semantic segmentation tasks, utilizing a combination of various synthetic and real-world RAW image datasets, an overview of the datasets can be found in Table 1.", + "bbox": [ + 212, + 479, + 782, + 525 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For the object detection task, we take 2 open source real-world dataset PASCAL RAW [52] and LOD [27]. PASCAL RAW [52] is a normal-light condition dataset with 4259 RAW images, taken by a Nikon D3200 DSLR camera with 3 object classes. To verify the generalization capability of RAW-Adapter across various lighting conditions, we additionally synthesized low-light and overexposure datasets based on the PASCAL RAW dataset, named PASCAL RAW (dark) and PASCAL RAW (over-exp) respectively, the synthesized datasets are identical to the PASCAL RAW dataset in all aspects except for brightness levels. For PASCAL RAW (dark) and PASCAL RAW (over-exp) synthesis, the light intensity and environment irradiance exhibit a linear relationship on RAW images, thus we employed the synthesis method from the previous work [14,78]:", + "bbox": [ + 212, + 526, + 787, + 693 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} x _ {n} \\sim N (\\mu = l x, \\sigma^ {2} = \\delta_ {r} ^ {2} + \\delta_ {s} l x) \\\\ y = l x + x _ {n}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 705, + 614, + 739 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "(7)", + "bbox": [ + 761, + 715, + 782, + 729 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $x$ denotes the original normal-light RAW image and $y$ denotes degraded RAW image, $\\delta_s = \\sqrt{S}$ denotes shot noise while $\\sqrt{S}$ is signal of the sensor, $\\delta_r$ denotes read noise, $l$ denotes the light intensity parameter which randomly chosen from [0.05, 0.4] in PASCAL RAW (dark), and randomly chosen from [1.5, 2.5] in PASCAL RAW (over-exp). Additionally, we follow PASCAL RAW [52]'s dataset split to separate the training set and test set.", + "bbox": [ + 212, + 748, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "RAW-Adapter", + "bbox": [ + 633, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/9fa8b16e81d2e3fe71ee0d1b2830e12ebde4e47d1a627503640e79c84dac4e70.jpg", + "table_caption": [ + "Table 1: Dataset and framework setting in our experiments." + ], + "table_footnote": [], + "table_body": "
PASCAL RAWPASCAL RAW (dark/over-exp)LODADE20K RAWADE20K RAW (dark/over-exp)
TaskObject DetectionSemantic Segmentation
Number4259223027,574
Typereal-worldsynthesisreal-worldsynthesis
SensorNikon D3200 DSLRCanon EOS 5D Mark IV-
FrameworkRetinaNet43 & Sparse-RCNN64Segformer70
BackboneResNet25MIT70
pre-trainImageNet44 pre-train weights
", + "bbox": [ + 215, + 171, + 787, + 319 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Meanwhile, LOD [27] is a real-world dataset with 2230 low-light condition RAW images, taken by a Canon EOS 5D Mark IV camera with 8 object classes, we take 1800 images as training set and the other 430 images as test set.", + "bbox": [ + 212, + 345, + 782, + 391 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For the semantic segmentation task, we utilized the widely-used sRGB dataset ADE20K [80] to generate RAW dataset. Leveraging state-of-the-art unprocessing methods InvISP [71], we synthesized RAW images corresponding to the ADE20K sRGB dataset. Using InvISP, we projected input sRGB images into RAW format, effectively creating an ADE20K RAW dataset. To simulate various lighting conditions, we employed the same synthesis method from PASCAL RAW (dark/over-exp) to generate low-light and over-exposure RAW images, name as ADE20K RAW (dark/over-exp), training & test split is same as ADE20K.", + "bbox": [ + 212, + 391, + 787, + 513 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implement Details. We build our framework based on the open-source computer vision toolbox mmdetection 7 and mmsegmentation 12, both object detection tasks and semantic segmentation tasks are initialed with ImageNet pre-train weights (see Table. 1), and we apply the data augmentation pipeline in the default setting, mainly include random crop, random flip, and multi-scale test, etc. For the object detection task, we adopt the 2 mainstream object detectors: RetinaNet 43 and Sparse-RCNN 64 with ResNet 25 backbone. For the semantic segmentation task, we choose to use the mainstream segmentation framework Segformer 70 with MIT 70 backbone.", + "bbox": [ + 212, + 535, + 787, + 676 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Comparison Methods. We conducted comparative experiments with the current state-of-the-art (SOTA) algorithms, including various open-sourced ISP methods [6,34,35,71,79] and joint-training method DirtyPixels [18], among these Karaimer et al. [35] is a traditional ISP method with various human manipulate steps, and InvISP [71] & Lite-ISP [79] & SID [6] & DNF [34] are current SOTA network-based ISP methods, where SID [6] and DNF [34] is especially for the low-light condition RAW data. For fairness compared with the above ISP methods, both the training and test RAW data are rendered using the respective", + "bbox": [ + 212, + 694, + 787, + 816 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Cui, Harada.", + "bbox": [ + 271, + 114, + 359, + 128 + ], + "page_idx": 9 + }, + { + "type": "page_footnote", + "text": "4 For more experiments on other frameworks, please refer to supplementary materials.", + "bbox": [ + 217, + 824, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/42d81234c539124aac78a75e373137568ffe4efd7191f5e118450b771dbefc12.jpg", + "image_caption": [ + "Fig. 5: (a). Detection performance on PASCAL RAW [52] (normal/dark/over-exp). (b). Efficiency comparison (blue: vanilla, yellow: RAW-Adapter, gray: Dirty-Pixel [18])." + ], + "image_footnote": [], + "bbox": [ + 218, + 142, + 787, + 348 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "compared ISP. DirtyPixels [18] is the current SOTA joint-training method which uses a stack of residual UNet [58] as a low-level processor and joint optimization with the following task-specific networks. For fairness, all comparison methods adopt the same data augmentation process and the same training setting, we will introduce detailed experiment results in the following section.", + "bbox": [ + 212, + 414, + 787, + 492 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Object Detection Evaluation", + "text_level": 1, + "bbox": [ + 215, + 512, + 498, + 527 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For object detection task on PASCAL RAW [52] dataset, we adopt RetinaNet [43] with different size ResNet [25] backbones (ResNet-18, ResNet-50), all the models are trained on a single NVIDIA Tesla V100 GPU with SGD optimizer, the batch size is set to 4, training images are cropped into range of (400, 667) and training epochs are set to 50. Table. [2] shows the detection results with (a). ResNet-18 and (b). ResNet-50 backbone, with comparisons of demosaiced RAW data (\"Demosacing\"), camera default ISP in [52], various current SOTA ISP solutions [6,34,35,79] and Dirty-Pixel [18], we can observe that sometimes ISP algorithms in normal and over-exposure conditions may even degraded the detection performance, and ISP in dark scenarios sometimes could improves detection performance, additionally manual design ISP solution [35] even out-perform the deep learning solutions. Previous joint-training method Dirty-Pixel [18] can improve detection performance. Overall, our RAW-Adapter method achieves the best performance and even outperforms some ISP algorithms utilizing ResNet-50 backbone when employing ResNet-18 backbone, which can significantly reduce computational load and model parameters.", + "bbox": [ + 212, + 537, + 787, + 779 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Visualization of PASCAL RAW's detection results are shown in Fig. 5(a), we show the detection results under different lightness conditions, background are images generated from Dirty-Pixel [18] and RAW-Adapter $(\\mathbf{I}_5)$ . Our method could achieve satisfactory detection results across different lightness, while other", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "RAW-Adapter", + "bbox": [ + 633, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/358f237cb7c0bb96d33e1676e4cd9f34639804082ec47e540fe423a272bc318a.jpg", + "table_caption": [ + "Table 2: Comparison results on PASCAL RAW dataset [52], we take [43] with (a). ResNet-18 and (b). ResNet-50 backbone, bold denotes the best result.", + "(a) ResNet-18." + ], + "table_footnote": [], + "table_body": "
mAP Methodnormal over-expdark
Default ISP88.3--
Demosacing87.787.780.3
Karäimer et al. 3586.085.681.9
Lite-ISP 7985.284.271.9
InvISP 7184.186.670.9
SID 6--78.2
DNF 34--81.1
Dirty-Pixel88.688.080.8
RAW-Adapter88.788.782.5
", + "bbox": [ + 215, + 220, + 493, + 398 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/bd05fbdf835d93a0c3d0170cfa120b144c9b89399c44960d86b3d96829e20772.jpg", + "table_caption": [ + "(b) ResNet-50." + ], + "table_footnote": [], + "table_body": "
mAP Methodnormal over-expdark
Default ISP89.4--
Demosacing89.288.882.6
Karäimer et al. 3589.486.884.6
Lite-ISP 7989.385.173.5
InvISP 7189.187.374.7
SID 6--81.5
DNF 34--82.8
Dirty-Pixel89.089.083.6
RAW-Adapter89.689.486.6
", + "bbox": [ + 504, + 220, + 784, + 398 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/815c08731aa74c2fcd18e028fac35cee2213a2de136db054419fc44f3b564540.jpg", + "table_caption": [ + "Table 3: Comparison with ISP methods and Dirty-Pixel on LOD dataset [27]. We show the detection performance (mAP) ↑ of RetinaNet (R-Net) [43] and Sparse-RCNN (Sp-RCNN) [64], **bold** denotes the best result." + ], + "table_footnote": [], + "table_body": "
MethodsDemosaicing Karaimer et al. 35InvISP 71SID 6
mAP(R-Net)58.554.456.949.1
mAP(Sp-RCNN)57.752.249.443.1
MethodsDNF 34Dirty-Pixel 18RAW-Adapter (w/o M)RAW-Adapter
mAP(R-Net)58.861.661.662.2
mAP(Sp-RCNN)55.858.858.659.2
", + "bbox": [ + 215, + 465, + 787, + 580 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "methods face false alarm and miss detect. Efficiency comparison with Dirty-Pixel is shown in Fig. 5(b), our algorithm has achieved significant improvements in both training time acceleration and savings in Flops/Parameter Number.", + "bbox": [ + 212, + 608, + 784, + 652 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Detection performance on LOD [27] is shown in Table. [3], we adopt two detector RetinaNet [43] and Sparse-RCNN [64], both with ResNet-50 [25] backbone, training epochs are set to 35, Sparse-RCNN [64] are trained by 4 GPUs with Adam optimizer. With comparing of various ISP methods [6,34,35,79], Dirty-Pixel [18] and RAW-Adapter with only input-level adapters (w/o M). Table. [3] show that our algorithm can achieve the best performance on both two detectors.", + "bbox": [ + 212, + 654, + 787, + 746 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Semantic Segmentation Evaluation", + "text_level": 1, + "bbox": [ + 215, + 768, + 547, + 782 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "For semantic segmentation on ADE20K RAW dataset, we choose Segformer [70] as the segmentation framework with the different size MIT [70] backbones (MIT-B5, MIT-B3, MIT-B0), all the models are trained on 4 NVIDIA Tesla V100 GPU", + "bbox": [ + 212, + 792, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Cui, Harada.", + "bbox": [ + 271, + 114, + 359, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/43043d2a432760b276f0eaa25b628295141b3111f94a5b4507b9b95f2684bbe7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 143, + 330, + 204 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c8193287929b9033283ff25322727b8a24c30031cbe015f9b52a33d0f0ecc9c3.jpg", + "image_caption": [ + "Input / GT" + ], + "image_footnote": [], + "bbox": [ + 222, + 205, + 328, + 265 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e4f2a078d5327cf3870558799ae48cbe0487cc4586202ea0f209d8756e71de6e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 145, + 442, + 204 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/76af0ecf44d389d21af0d474151c8e9c61beb874a7b96e1f43881a5c46f24645.jpg", + "image_caption": [ + "LiteISP" + ], + "image_footnote": [], + "bbox": [ + 336, + 205, + 442, + 265 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/44c2d3340b05c06f0411900e178c0de0a616f8f8e2beb622cc0d24ccbb86a69a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 143, + 557, + 204 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8b21f471d384198de87064f75c2ed219886469de1c075a3d665cdbffcdcde120.jpg", + "image_caption": [ + "InvISP" + ], + "image_footnote": [], + "bbox": [ + 450, + 205, + 557, + 265 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/cf27c40c973a7170e58f2816d4c3e3d2d84f862a60eeba7b28d2aeaa320426bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 562, + 143, + 669, + 204 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/628ac741a03d8487f813f8227eb1e1eb9a93781a41656fa7e26475947a1f20c4.jpg", + "image_caption": [ + "Dirty-Pixel" + ], + "image_footnote": [], + "bbox": [ + 562, + 205, + 669, + 265 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c3d438cf47663ee167d54b4df4a571b1946916bde817ee7f245cb2f8b76c047c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 143, + 782, + 204 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d0c36cdfb0f926effb510b67d7ad725e9d8ae65ddd28d3a1dbbc69681d896acc.jpg", + "image_caption": [ + "RAW-Adapter" + ], + "image_footnote": [], + "bbox": [ + 674, + 205, + 782, + 265 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1514dfba31eff02384302b16778c9ac4ceceb7f59229d600c11bf0420f1dd62b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 279, + 330, + 340 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b44d4e10413562e366e88c7cc87cbde419ac36ebac93b11343f500c94b24eb4e.jpg", + "image_caption": [ + "Input / GT" + ], + "image_footnote": [], + "bbox": [ + 222, + 342, + 328, + 402 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/766e832983604b3b8e0a1bcbbb5d5f3c6b75844c98adc6d5eac514048e16e010.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 280, + 442, + 340 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/eaf10951855c1e7b00355437e9bf43731f5ff6348214d1a84434974dc6210f61.jpg", + "image_caption": [ + "SID" + ], + "image_footnote": [], + "bbox": [ + 336, + 342, + 442, + 402 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/546c157963da137b9df9751c2c0e23d803f8f81bf9f26209e8bebf0708c4459b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 279, + 555, + 340 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/dc170adc5338b02bf615689f40891bfc32bd0a0ff064c196e6f2aff99c71ae0e.jpg", + "image_caption": [ + "Karaimer et al." + ], + "image_footnote": [], + "bbox": [ + 450, + 342, + 555, + 402 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7a65f06628b24f3f75d006a461f46f87c194ffea49dddd0dfa54024eb7583f65.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 562, + 279, + 669, + 340 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/111aad8e97a5ac64533d6517df20d178561fb84fa8f690f86932affec39dd0d9.jpg", + "image_caption": [ + "Dirty-Pixel", + "Fig. 6: Semantic segmentation results on over-exposure and low-light RAW data, compare with various ISP methods [6], [35], [71], [79] and Dirty-Pixel [18]." + ], + "image_footnote": [], + "bbox": [ + 562, + 342, + 669, + 402 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/be7242f00efc4246400c470c9a2fec669fccc661513f726dd10c481fd309cdf2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 279, + 782, + 340 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d59502f2dfe0b97849a159e2a8fa294471a3852c3f2d4e4f5423d16306951ff7.jpg", + "image_caption": [ + "RAW-Adapter" + ], + "image_footnote": [], + "bbox": [ + 674, + 342, + 782, + 402 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "with Adam optimizer, the batch size is set to 4, training images are cropped into $512 \\times 512$ and training iter are set to 80000. We make comparison with single demosacing, various ISP methods [6,34,35,71,79] and Dirty-Pixel [18], and RAW-Adapter with only input-level adapters (w/o M).", + "bbox": [ + 212, + 489, + 784, + 549 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Comparison results are shown in Table. 4, where we compare both efficiency (parameters↓, inference time↓) and performance (mIOU↑), inference time is calculated on a single Tesla V100 GPU. We first show the results on vanilla demosaiced RAW data, then we show the results on RAW data processed by various ISP methods [6,34,35,71,79], followed by results of Dirty-Pixel [18], and our method without and with model-level adapter M. For Table. 4, we can find that directly employing ISP methods sometimes fails to enhance the performance of downstream segmentation tasks, particularly in low-light conditions. Joint-training method Dirty-Pixel [18] could improve down-stream performance and sometimes even degrade segmentation performance, but increase more network parameters. Overall, our proposed RAW-Adapter could effectively enhance performance while adding a slight parameter count and improving inference efficiency. RAW-Adapter on backbones with fewer parameters can even outperform previous methods on backbones with more parameters (i.e. The RAW-Adapter, employing the MIT-B3 backbone, achieves a dark condition mIOU of 40.52, surpassing other methods utilizing the MIT-B5 backbone despite having fewer than $\\sim$ 36M parameters.).", + "bbox": [ + 212, + 551, + 787, + 808 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The visualization of segmentation results are shown in Fig. 6, where row 1 and row 3 depict images rendered by different ISPs [6, 35, 71, 79] and images", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "RAW-Adapter", + "bbox": [ + 633, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/6221db2e5489ffb04015b9b29718336ab4e98da39e1da919e8970ba80c64effc.jpg", + "table_caption": [ + "Table 4: Comparison with previous methods on ADE 20K RAW (normal/over-exp/dark). Bold denotes the best result while underline denotes second best result." + ], + "table_footnote": [], + "table_body": "
backboneparams(M) ↓inference time(s) ↓mIOU ↑ (normal)mIOU ↑ (over-exp)mIOU ↑ (dark)
DemosacingMIT-B582.010.10545.2642.0830.97
Karaimer et al. [35]0.52546.4243.7139.79
InvISP [71]0.20347.5343.554.87
LiteISP [79]0.26143.2242.015.52
DNF [34]0.186--25.82
SID [6]0.312--26.44
Dirty-Pixel [18]86.290.15947.7646.8040.02
MIT-B348.920.09847.1945.3338.93
MIT-B08.000.04935.4332.1027.19
RAW-Adapter (w/o M)MIT-B582.090.14848.3346.4840.91
MIT-B344.720.08647.0145.0039.34
MIT-B03.800.03235.0932.8227.29
RAW-AdapterMIT-B582.530.16748.3847.0641.82
MIT-B345.160.10247.5945.6640.52
MIT-B04.260.05335.6733.4427.78
", + "bbox": [ + 215, + 185, + 787, + 476 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "generated from Dirty-Pixel [18] and RAW-Adapter $(\\mathbf{I}_5)$ , row 2 and row 4 denote our segmentation results appear superior compared to other methods. We were even pleased to discover that, although we do not add any human-vision-oriented loss function to constrain the model, RAW-Adapter still produces visually satisfactory image $\\mathbf{I}_5$ (see Fig. 6). This may be attributed to our preservation of traditional ISP stages in input-level adapters. Due to page limitations, we refer to more experimental results, visualization results, and ablation analysis of our method in the supplementary part.", + "bbox": [ + 212, + 502, + 789, + 625 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 643, + 359, + 660 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduce RAW-Adapter, an effective solution for adapting pre-trained sRGB models to camera RAW data. With input-level adapters and model-level adapters working in tandem, RAW-Adapter effectively forwards RAW images from various lighting conditions to downstream vision tasks, our method has achieved state-of-the-art (SOTA) results across multiple datasets.", + "bbox": [ + 212, + 672, + 790, + 748 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "For future research directions, we believe it is feasible to design a unified model capable of adapting to RAW visual tasks under different lighting conditions, without the need for retraining in each lighting scenario like RAW-Adapter. Additionally, designing multi-task decoders can enable the accommodation of various tasks, facilitating more effective integration of different visual tasks on RAW images with large-scale models.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Cui, Harada.", + "bbox": [ + 271, + 114, + 359, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This research is partially supported by JST Moonshot R&D Grant Number JPMJPS2011, CREST Grant Number JPMJCR2015 and Basic Research Grant (Super AI) of Institute for AI and Beyond of the University of Tokyo.", + "bbox": [ + 215, + 175, + 787, + 220 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 243, + 321, + 258 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Affifi, M., Brown, M.S.: What else can fool deep learning? addressing color constancy errors on deep neural network performance. 2019 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 243-252 (2019), https://apisemantic scholar.org/CorpusID:207995696", + "2. Brooks, T., Mildenhall, B., Xue, T., Chen, J., Sharlet, D., Barron, J.T.: Unprocessing images for learned raw denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11036-11045 (2019)", + "3. Buckler, M., Jayasuriya, S., Sampson, A.: Reconfiguring the imaging pipeline for computer vision. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 975-984 (2017)", + "4. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)", + "5. Chatterjee, P., Joshi, N., Kang, S.B., Matsushita, Y.: Noise suppression in low-light images through joint denoising and demosaicing. In: CVPR 2011. pp. 321-328. IEEE (2011)", + "6. Chen, C., Chen, Q., Xu, J., Koltun, V.: Learning to see in the dark. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3291-3300 (2018)", + "7. Chen, K., Wang, J., et al.: Mmdetection: Open mmlab detection toolbox and benchmark (2019)", + "8. Chen, Z., Duan, Y., Wang, W., He, J., Lu, T., Dai, J., Qiao, Y.: Vision transformer adapter for dense predictions. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=plKu2GByCNW", + "9. Cheng, B., Schwing, A.G., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. In: NeurIPS (2021)", + "0. Conde, M.V., McDonagh, S., Maggioni, M., Leonardis, A., Pérez-Pellitero, E.: Model-based image signal processors via learnable dictionaries. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 481-489 (2022)", + "1. Conde, M.V., Vazquez-Corral, J., Brown, M.S., Timofte, R.: Nilut: Conditional neural implicit 3d lookup tables for image enhancement. In: Proceedings of the AAAI Conference on Artificial Intelligence (2024)", + "2. Contributors, M.: MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation (2020)", + "3. Cui, Z., Li, K., Gu, L., Su, S., Gao, P., Jiang, Z., Qiao, Y., Harada, T.: You only need 90k parameters to adapt light: a light weight transformer for image enhancement and exposure correction. In: BMVC. p. 238 (2022)", + "4. Cui, Z., Qi, G.J., Gu, L., You, S., Zhang, Z., Harada, T.: Multitask aet with orthogonal tangent regularity for dark object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2553-2562 (October 2021)" + ], + "bbox": [ + 225, + 273, + 787, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "RAW-Adapter", + "bbox": [ + 633, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Cui, Z., Zhu, Y., Gu, L., Qi, G.J., Li, X., Zhang, R., Zhang, Z., Harada, T.: Exploring resolution and degradation clues as self-supervised signal for low quality object detection. In: European Conference on Computer Vision. pp. 473-491. Springer (2022)", + "16. Delbracio, M., Kelly, D., Brown, M.S., Milanfar, P.: Mobile computational photography: A tour. Annual Review of Vision Science 7(1), 571-604 (2021). https://doi.org/10.1146/annurev-vision-093019-115521, https://doi.org/10.1146/annurev-vision-093019-115521, pMID: 34524880", + "17. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A Large-Scale Hierarchical Image Database. In: CVPR09 (2009)", + "18. Diamond, S., Sitzmann, V., Julca-Aguilar, F., Boyd, S., Wetzstein, G., Heide, F.: Dirty pixels: Towards end-to-end image processing and perception. ACM Transactions on Graphics (SIGGRAPH) (2021)", + "19. Dong, X., Yokoya, N.: Understanding dark scenes by contrasting multi-modal observations. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 840-850 (January 2024)", + "20. Finlayson, G.D., Trezzi, E.: Shades of gray and colour constancy. In: Color and Imaging Conference. vol. 2004, pp. 37-41. Society for Imaging Science and Technology (2004)", + "21. Goyal, B., Lalonde, J.F., Li, Y., Gupta, M.: Robust scene inference under noise-blur dual corruptions. 2022 IEEE International Conference on Computational Photography (ICCP) pp. 1-12 (2022), https://api_semanticscholar.org/CorpusID:251040807", + "22. Guo, Y., Luo, F., Wu, X.: Learning degradation-independent representations for camera isp pipelines. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 25774-25783 (June 2024)", + "23. Hansen, N., Ostermeier, A.: Adapting arbitrary normal mutation distributions in evolution strategies: the covariance matrix adaptation. In: Proceedings of IEEE International Conference on Evolutionary Computation. pp. 312-317 (1996). https://doi.org/10.1109/ICEC.1996.542381", + "24. Hasinoff, S.W., Sharlet, D., Geiss, R., Adams, A., Barron, J.T., Kainz, F., Chen, J., Levoy, M.: Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM Transactions on Graphics (Proc. SIGGRAPH Asia) 35(6) (2016)", + "25. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "26. Heide, F., Steinberger, M., Tsai, Y.T., Rouf, M., Pajak, D., Reddy, D., Gallo, O., Liu, J., Heidrich, W., Egiazarian, K., et al.: Flexisp: A flexible camera image processing framework. ACM Transactions on Graphics (ToG) 33(6), 1-13 (2014)", + "27. Hong, Y., Wei, K., Chen, L., Fu, Y.: Crafting object detection in very low light. In: BMVC (2021)", + "28. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019)", + "29. Hu, Y., He, H., Xu, C., Wang, B., Lin, S.: Exposure: A white-box photo post-processing framework. ACM Transactions on Graphics (TOG) 37(2), 26 (2018)", + "30. Ignatov, A., Van Gool, L., Timofte, R.: Replacing mobile camera isp with a single deep learning model. arXiv preprint arXiv:2002.05509 (2020)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Cui, Harada.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Iscen, A., Zhang, J., Lazebnik, S., Schmid, C.: Memory-efficient incremental learning through feature adaptation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16. pp. 699-715. Springer (2020)", + "32. Jensen, H.W., Durand, F., Dorsey, J., Stark, M.M., Shirley, P., Premoze, S.: A physically-based night sky model. In: Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques. p. 399-408. SIGGRAPH '01, Association for Computing Machinery, New York, NY, USA (2001). https://doi.org/10.1145/383259.383306, https://doi.org/10.1145/383259.383306", + "33. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022)", + "34. Jin, X., Han, L., Li, Z., Chai, Z., Guo, C., Li, C.: Dnf: Decouple and feedback network for seeing in the dark. In: ICCV (2023)", + "35. Karaimer, H.C., Brown, M.S.: A software platform for manipulating the camera imaging pipeline. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) Computer Vision - ECCV 2016. pp. 429-444. Springer International Publishing, Cham (2016)", + "36. Ke, L., Ye, M., Danelljan, M., Liu, Y., Tai, Y.W., Tang, C.K., Yu, F.: Segment anything in high quality. In: NeurIPS (2023)", + "37. Kim, W., Kim, G., Lee, J., Lee, S., Baek, S.H., Cho, S.: Paramisp: Learned forward and inverse isps using camera parameters. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 26067-26076 (2024)", + "38. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4015-4026 (October 2023)", + "39. Koskinen, S., Yang, D., Kämäräinen, J.K.: Reverse imaging pipeline for raw rgb image augmentation. In: 2019 IEEE International Conference on Image Processing (ICIP). pp. 2896-2900 (2019). https://doi.org/10.1109/ICIP.2019.8804406", + "40. Li, Z., Lu, M., Zhang, X., Feng, X., Asif, M.S., Ma, Z.: Efficient visual computing with camera raw snapshots. TPAMI pp. 1-18 (2024). https://doi.org/10.1109/TPAMI.2024.3359326", + "41. Li, Z., Jiang, H., Cao, M., Zheng, Y.: Polarized color image denoising. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9873-9882 (2023). https://doi.org/10.1109/CVPR52729.2023.00952", + "42. Liang, Z., Cai, J., Cao, Z., Zhang, L.: Cameranet: A two-stage framework for effective camera isp learning. IEEE Transactions on Image Processing 30, 2248-2262 (2021). https://doi.org/10.1109/TIP.2021.3051486", + "43. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)", + "44. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Fleet, D., Pajdla, T., Schiele, B., Tuytelaars, T. (eds.) Computer Vision - ECCV 2014. pp. 740-755. Springer International Publishing, Cham (2014)", + "45. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021)", + "46. Mildenhall, B., Hedman, P., Martin-Brualla, R., Srinivasan, P.P., Barron, J.T.: Nerf in the dark: High dynamic range view synthesis from noisy raw images. In:" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "RAW-Adapter", + "bbox": [ + 633, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16190-16199 (2022)", + "47. Monakhova, K., Richter, S.R., Waller, L., Koltun, V.: Dancing under the stars: Video denoising in starlight. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16241-16251 (June 2022)", + "48. Mosleh, A., Sharma, A., Onzon, E., Mannan, F., Robidoux, N., Heide, F.: Hardware-in-the-loop end-to-end optimization of camera image processing pipelines. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)", + "49. Nam, S., Kim, S.J.: Modelling the scene dependent imaging in cameras with a deep neural network. 2017 IEEE International Conference on Computer Vision (ICCV) pp. 1726-1734 (2017), https://api.sementicscholar.org/CorpusID:23920472", + "50. Nguyen, R.M.H., Brown, M.S.: Raw image reconstruction using a self-contained srgb-jpeg image with only 64 kb overhead. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1655-1663 (2016). https://doi.org/10.1109/CVPR.2016.183", + "51. Nishimura, J., Gerasimow, T., Sushma, R., Sutic, A., Wu, C.T., Michael, G.: Automatic isp image quality tuning using nonlinear optimization. In: 2018 25th IEEE International Conference on Image Processing (ICIP). pp. 2471-2475. IEEE (2018)", + "52. Omid-Zohoor, A., Ta, D., Murmann, B.: Pascalraw: raw image database for object detection (2014)", + "53. Onzon, E., Mannan, F., Heide, F.: Neural auto-exposure for high-dynamic range object detection. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7706-7716 (2021). https://doi.org/10.1109/CVPR46437.2021.00762", + "54. Potlapalli, V., Zamir, S.W., Khan, S., Khan, F.: Prompt: Prompting for all-in-one image restoration. In: Thirty-seventh Conference on Neural Information Processing Systems (2023)", + "55. Qin, H., Han, L., Wang, J., Zhang, C., Li, Y., Li, B., Hu, W.: Attention-aware learning for hyperparameter prediction in image processing pipelines. In: European Conference on Computer Vision. pp. 271-287. Springer (2022)", + "56. Ramanath, R., Snyder, W.E., Yoo, Y., Drew, M.S.: Color image processing pipeline. IEEE Signal processing magazine 22(1), 34-43 (2005)", + "57. Rebuffi, S.A., Bilen, H., Vedaldi, A.: Learning multiple visual domains with residual adapters. Advances in neural information processing systems 30 (2017)", + "58. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Navab, N., Hornegger, J., Wells, W.M., Frangi, A.F. (eds.) Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015. pp. 234-241. Springer International Publishing, Cham (2015)", + "59. Sayed, M., Brostow, G.: Improved handling of motion blur in online object detection. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1706-1716 (2021). https://doi.org/10.1109/CVPR46437.2021.00175", + "60. Schwartz, E., Giryes, R., Bronstein, A.M.: Deepisp: Toward learning an end-to-end image processing pipeline. IEEE Transactions on Image Processing 28(2), 912-923 (2018)", + "61. Sitzmann, V., Martel, J.N., Bergman, A.W., Lindell, D.B., Wetzstein, G.: Implicit neural representations with periodic activation functions. In: Proc. NeurIPS (2020)", + "62. Souza, M., Heidrich, W.: MetaISP - Exploiting Global Scene Structure for Accurate Multi-Device Color Rendition. In: Guthe, M., Grosch, T. (eds.) Vision, Modeling," + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Cui, Harada.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Visualization. The Eurographics Association (2023). https://doi.org/10.2312/vmv.20231236", + "63. Stickland, A.C., Murray, I.: Bert and pals: Projected attention layers for efficient adaptation in multi-task learning. In: International Conference on Machine Learning. pp. 5986-5995. PMLR (2019)", + "64. Sun, P., Zhang, R., Jiang, Y., Kong, T., Xu, C., Zhan, W., Tomizuka, M., Li, L., Yuan, Z., Wang, C., Luo, P.: SparseR-CNN: End-to-end object detection with learnable proposals. arXiv preprint arXiv:2011.12450 (2021)", + "65. Sung, Y.L., Cho, J., Bansal, M.: Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5227-5237 (2022)", + "66. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)", + "67. Wang, Y., Yu, Y., Yang, W., Guo, L., Chau, L.P., Kot, A.C., Wen, B.: Raw image reconstruction with learned compact metadata. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18206-18215 (June 2023)", + "68. Wei, K., Fu, Y., Yang, J., Huang, H.: A physics-based noise formation model for extreme low-light raw denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)", + "69. Wu, C.T., Isikdogan, L.F., Rao, S., Nayak, B., Gerasimow, T., Sutic, A., Ainkedem, L., Michael, G.: Visionisp: Repurposing the image signal processor for computer vision applications. In: 2019 IEEE International Conference on Image Processing (ICIP). IEEE (Sep 2019). https://doi.org/10.1109/icip.2019.8803607, http://dx.doi.org/10.1109/ICIP.2019.8803607", + "70. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. arXiv preprint arXiv:2105.15203 (2021)", + "71. Xing, Y., Qian, Z., Chen, Q.: Invertible image signal processing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6287-6296 (2021)", + "72. Xu, R., Chen, C., Peng, J., Li, C., Huang, Y., Song, F., Yan, Y., Xiong, Z.: Toward raw object detection: A new benchmark and a new model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13384-13393 (2023)", + "73. Yoshimura, M., Otsuka, J., Irie, A., Ohashi, T.: Dynamically controlled image signal processor for image recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 12866-12876 (October 2023)", + "74. Yoshimura, M., Otsuka, J., Irie, A., Ohashi, T.: Rawgment: Noise-accounted raw augmentation enables recognition in a wide variety of environments. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14007-14017 (June 2023)", + "75. Yu, K., Li, Z., Peng, Y., Loy, C.C., Gu, J.: Reconfigisp: Reconfigurable camera image processing pipeline. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4248-4257 (2021)", + "76. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Cycleisp: Real image restoration via improved data synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2696-2705 (2020)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "RAW-Adapter", + "bbox": [ + 633, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "77. Zhang, R., Zhang, W., Fang, R., Gao, P., Li, K., Dai, J., Qiao, Y., Li, H.: Tip-adapter: Training-free adaption of clip for few-shot classification. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 493-510. Springer Nature Switzerland, Cham (2022)", + "78. Zhang, Y., Zhang, Y., Zhang, Z., Zhang, M., Tian, R., Ding, M.: Isp-teacher: Image signal process with disentanglement regularization for unsupervised domain adaptive dark object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 7387-7395 (2024)", + "79. Zhang, Z., Wang, H., Liu, M., Wang, R., Zuo, W., Zhang, J.: Learning raw-to-srgb mappings with inaccurately aligned supervision. In: ICCV (2021)", + "80. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing through ade20k dataset. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 633-641 (2017)", + "81. Zhou, W., Gao, S., Zhang, L., Lou, X.: Histogram of oriented gradients feature extraction from raw bayer pattern images. IEEE Transactions on Circuits and Systems II: Express Briefs 67(5), 946-950 (2020). https://doi.org/10.1109/TCSII.2020.2980557" + ], + "bbox": [ + 215, + 146, + 785, + 382 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "Cui, Harada.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/2024/RAW-Adapter_ Adapting Pretrained Visual Model to Camera RAW Images/a7bce0f0-b64e-4428-b3dc-7b7e2a217569_model.json b/2024/RAW-Adapter_ Adapting Pretrained Visual Model to Camera RAW Images/a7bce0f0-b64e-4428-b3dc-7b7e2a217569_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4a111a4fc9ef6c0eef0f7664987729899d2c66b8 --- /dev/null +++ b/2024/RAW-Adapter_ Adapting Pretrained Visual Model to Camera RAW Images/a7bce0f0-b64e-4428-b3dc-7b7e2a217569_model.json @@ -0,0 +1,3001 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.249, + 0.141, + 0.756, + 0.187 + ], + "angle": 0, + "content": "RAW-Adapter: Adapting Pre-trained Visual Model to Camera RAW Images" + }, + { + "type": "text", + "bbox": [ + 0.362, + 0.212, + 0.644, + 0.23 + ], + "angle": 0, + "content": "Ziteng Cui\\(^{1}\\) and Tatsuya Harada\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.413, + 0.24, + 0.591, + 0.254 + ], + "angle": 0, + "content": "1 The University of Tokyo" + }, + { + "type": "text", + "bbox": [ + 0.452, + 0.255, + 0.551, + 0.267 + ], + "angle": 0, + "content": "2 RIKEN AIP" + }, + { + "type": "text", + "bbox": [ + 0.378, + 0.269, + 0.626, + 0.283 + ], + "angle": 0, + "content": "{cui, harada}@mi.t.u-tokyo.ac.jp" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.321, + 0.744, + 0.587 + ], + "angle": 0, + "content": "Abstract. sRGB images are now the predominant choice for pre-training visual models in computer vision research, owing to their ease of acquisition and efficient storage. Meanwhile, the advantage of RAW images lies in their rich physical information under variable real-world challenging lighting conditions. For computer vision tasks directly based on camera RAW data, most existing studies adopt methods of integrating image signal processor (ISP) with backend networks, yet often overlook the interaction capabilities between the ISP stages and subsequent networks. Drawing inspiration from ongoing adapter research in NLP and CV areas, we introduce RAW-Adapter, a novel approach aimed at adapting sRGB pre-trained models to camera RAW data. RAW-Adapter comprises input-level adapters that employ learnable ISP stages to adjust RAW inputs, as well as model-level adapters to build connections between ISP stages and subsequent high-level networks. Additionally, RAW-Adapter is a general framework that could be used in various computer vision frameworks. Abundant experiments under different lighting conditions have shown our algorithm's state-of-the-art (SOTA) performance, demonstrating its effectiveness and efficiency across a range of real-world and synthetic datasets. Code is available at this url" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.612, + 0.377, + 0.628 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.844 + ], + "angle": 0, + "content": "In recent years, there has been a growing interest in revisiting vision tasks using unprocessed camera RAW images. How to leverage information-rich RAW image in computer vision tasks has become a current research hotspot in various sub-areas (i.e. denoising [27], view synthesis [46], object detection [27, 48]). Compared to commonly used sRGB image, RAW image directly acquired by the camera sensor, encompasses abundant information unaffected or compressed by image signal processor (ISP), also offers physically meaningful information like noise distributions [47, 68], owing to its linear correlation between image intensity and the radiant energy received by a camera. The acquisition of RAW data, facilitating enhanced detail capture and a higher dynamic range, imparts a unique advantage in addressing visual tasks under variable lighting conditions in the real world. For instance, sunlight irradiance can reach as high as \\(1.3 \\times 10^{3}W / m^{2}\\) while bright planet irradiance can be as low as \\(2.0 \\times 10^{-6}W / m^{2}\\) [32]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "Cui, Harada." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.788, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.332, + 0.788, + 0.377 + ], + "angle": 0, + "content": "Fig. 1: (a). An overview of basic image signal processor (ISP) pipeline. (b). ISP and current visual model have different objectives. (c) Previous methods optimize ISP with down-stream visual model. (d) Our proposed RAW-Adapter." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.411, + 0.788, + 0.519 + ], + "angle": 0, + "content": "Meanwhile, sRGB has emerged as the primary choice for pre-training visual models in today's computer vision field, due to its scalability and ease of storage. Typically, sRGB images are derived from camera RAW data through the ISP pipeline. As shown in Fig. 1(a), the entire ISP pipeline consists of multiple modules to convert RAW images into vision-oriented sRGB images, each of these modules serves its own distinct purpose, with the majority being ill-posed and heavily dependent on prior information [5,16,26,35,42,56]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.788, + 0.628 + ], + "angle": 0, + "content": "When adopt RAW data for computer vision tasks, the purpose of a manually designed ISP is to produce images that offer a superior visual experience [18,69], rather than optimizing for downstream visual tasks such as object detection or semantic segmentation (see Fig. ①b)). Additionally, most companies' ISPs are black boxes, making it difficult to obtain information about the specific steps inside. Consequently, utilizing human vision-oriented ISP in certain conditions is sometimes even less satisfactory than directly using RAW [22,27,40,81]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.789, + 0.842 + ], + "angle": 0, + "content": "To better take advantage of camera RAW data for various computer vision tasks. Researchers began to optimize the image signal processor (ISP) jointly with downstream networks. Since most ISP stages are non-differentiable and cannot be jointly backpropagation, there are two main lines of approaches to connect ISP stages with downstream networks (see Fig. 1(c)): ①. First-kind approaches maintain the modular design of the traditional ISP, involving the design of differentiable ISP modules through optimization algorithms 48,69,75, such as Hardware-in-the-loop 48 adopt covariance matrix adaptation evolution strategy (CMA-ES) 23. ②. Second-kind approaches replaced the ISP part entirely with a neural network 18,55,72,73, such as Dirty-Pixel 18 adopt a stack of residual UNets 58 as pre-encoder, however, the additional neural network introduces a significant computational burden, especially when dealing with high-resolution inputs. Beyond these, the above-mentioned methods still treat ISP and the backend neural network as two independent modules, lacking interac" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.634, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RAW-Adapter" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.145, + 0.501, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.145, + 0.788, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.267, + 0.788, + 0.327 + ], + "angle": 0, + "content": "Fig. 2: Performance of RAW-based visual tasks with and without sRGB pre-trained weights. We analyze two methods: Dirty-Pixel [18] and RAW-Adapter. Blue line represents trained with MS COCO [44] pre-train weights, the purple line indicates ImageNet [17] pre-train weights, and the yellow line signifies training from scratch." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.48 + ], + "angle": 0, + "content": "tion capabilities between two separate models. Especially, current visual models predominantly pre-trained on large amounts of sRGB images, there exists a significant gap between pre-trained models and camera RAW images. As it shown in Fig. 2 training from scratch with RAW data can significantly impair the performance of RAW-based vision tasks. How to leveraging sRGB pre-trained weights proves to be highly beneficial for RAW image visual tasks. Hence, here we ask: Is there a more effective way to adapt the information-rich RAW data into the knowledge-rich sRGB pre-trained models?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.482, + 0.789, + 0.769 + ], + "angle": 0, + "content": "Our solution, RAW-Adapter, differs from previous approaches that employed complex ISP or deeper neural networks at the input level. Instead, we prioritize simplifying the input stages and enhancing connectivity between ISP and subsequent networks at the model level. Inspired by recent advancements in prompt learning and adapter tuning [8,28,33,54], we developed two novel adapter approaches aimed at enhancing the integration between RAW input and sRGB pretrained models. Leveraging priors from ISP stages, our method includes input-level adapters and model-level adapters, as illustrated in Fig. [1(d)]. Input-level adapters are designed to adapt the input from RAW data to the backend network input, we follow traditional ISP design and predict ISP stages' key parameters to accommodate RAW image into the specific down-stream vision tasks, in this part, we employ strategies involving Query Adaptive Learning (QAL) and Implicit Neural Representation (INR) to ensure the differentiability of the core ISP processes while also upholding a lightweight design. Model-level adapters leverage prior knowledge from the input-level adapters by extracting intermediate stages as features, these features are then embedded into the downstream network, thus the network incorporates prior knowledge from the ISP stages, then jointly contributes to final machine-vision perception tasks. Our contributions could be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "- We introduce RAW-Adapter, a novel framework aimed at improving the integration of sRGB pre-trained models to camera RAW images. Through the implementation of two types of adapters, we strive to narrow the dispar" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.129 + ], + "angle": 0, + "content": "Cui, Harada." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "ity between RAW images and sRGB models, tackling discrepancies at both input and model levels." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.178, + 0.784, + 0.25 + ], + "angle": 0, + "content": "- By analyzing camera ISP steps, we've crafted input-level adapters that integrate query adaptive learning (QAL) and implicit neural representations (INR) to optimize ISP key parameters. Additionally, harnessing prior input stage information informs model-level adapters, enriching model understanding and improving downstream task performance." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.254, + 0.784, + 0.327 + ], + "angle": 0, + "content": "- We conducted detection and segmentation experiments involving different lighting scenarios, including normal-light, dark, and over-exposure scenes. Through comparisons with current mainstream ISPs and joint-training methods, sufficient experiments demonstrate that our algorithm achieved state-of-the-art (SOTA) performance." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.353, + 0.396, + 0.369 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.386, + 0.455, + 0.402 + ], + "angle": 0, + "content": "2.1 Image Signal Processor" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.789, + 0.701 + ], + "angle": 0, + "content": "Typically, a camera's Image Signal Processor (ISP) pipeline is required to reconstruct a high-quality sRGB image from camera RAW data, traditional ISP pipeline is formulated as a series of manually crafted modules executed sequentially [16,35,51,56], including some representative steps such as demosaicing, white balance, noise removal, tone mapping, color space transform and so on. There also exist alternative designs of ISP process such as Heide et al. [26] designed FlexISP which combine numerous ISP blocks to a joint optimization block, and Hasinoff et al. [24] modified some of the traditional ISP steps for burst photography under extreme low-light conditions. When it comes to deep learning era, Chen et al. [6] proposed SID to use a UNet [58] replace the traditional ISP steps, which translate low-light RAW data to normal-light sRGB images, and Hu et al. [29] proposed a white-box solution with eight differentiable filters. After that, many efforts began to substitute the traditional ISP process with a neural network, such as [6,10,30,34,37,42,60,62,79], however deep network models are constrained by the training dataset, leading to shortcomings in generalization performance. Meanwhile another line of work focuses on translate sRGB images back to RAW data [2,39,49,50,67,71,76]. Apart from the conversion between RAW and RGB, there is ongoing research specifically dedicated to harnessing RAW images for downstream computer vision tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.723, + 0.579, + 0.739 + ], + "angle": 0, + "content": "2.2 Computer Vision based on RAW data" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In order to effectively leverage information in RAW images for downstream visual tasks and save the time required for ISP, some early methods proposed to directly perform computer vision tasks on RAW images [3,81], which lacks consideration for the camera noise introduced during the conversion from photons to RAW images, especially in low-light conditions [41,47,68]. Meanwhile training from scratch on RAW data would abandon the current large-scale pre-trained" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.634, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RAW-Adapter" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.208 + ], + "angle": 0, + "content": "visual model on sRGB data (see Fig. 2), especially the quantity of RAW image datasets (i.e. 4259 images in PASCAL RAW 52 dataset and 2230 images in LOD 27 dataset) is incomparable to current RGB datasets (i.e. over 1M images in ImageNet 17 dataset and SAM 38 is trained with 11M images)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.788, + 0.436 + ], + "angle": 0, + "content": "Most subsequent research focuses on finding ways to integrate ISP and backend computer vision models [18,22,48,55,69,72,73,75]. For example Wu et al. [69] first propose VisionISP and emphasized the difference between human vision and machine vision, they introduced several trainable modules in the ISP to enhance backend object detection performance. Since then several methods attempt to replace the existing non-differentiable ISP with a differentiable ISP network [48,75] or directly using encoder networks to replace the ISP process [18,55,73], however training two consecutive networks simultaneously can result in a significant consumption of computational resources, previous work also rarely focused on how to fine-tune the current mainstream sRGB visual models more efficiently on RAW data. Additionally, research like Rawgment [74], focusing on achieving realistic data augmentation directly on RAW images, some studies address the decrease in computer vision performance caused by specific steps within the incamera process or ISP, such as white balance error [1], auto-exposure error [53], camera motion blur [59] and undesirable camera noise [15,21]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.456, + 0.506, + 0.472 + ], + "angle": 0, + "content": "2.3 Adapters in Computer Vision" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.481, + 0.788, + 0.633 + ], + "angle": 0, + "content": "Adapters have become prevalent NLP area, which introducing new modules in large language model (LLM) such as [28,63], these modules enable task-specific fine-tuning, allowing pre-trained models to swiftly adapt to downstream NLP tasks. In the computer vision area, adapters have been adopted in various areas such as incremental learning [31] and domain adaptation [57]. Recently, a series of adapters have been used to investigate how to better utilize pre-trained visual models [8,33,36] or pre-trained vision-language models [65,77]. These methods focus on utilizing prior knowledge to make pre-trained models quickly adapt to downstream tasks. Here, we introduce the RAW-Adapter to further improve the alignment between RAW inputs and sRGB pre-trained visual models." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.655, + 0.393, + 0.673 + ], + "angle": 0, + "content": "3 RAW-Adapter" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.785, + 0.765 + ], + "angle": 0, + "content": "In this section, we introduce RAW-Adapter. Due to page limitation, an overview of the conventional image signal processor (ISP) in supplementary's Sec. A. For specific adapter designs, we selectively omit certain steps and focus on some key steps within the camera ISP. In Sec. 3.1 we introduce RAW-Adapter's input-level adapters, followed by an explanation of its model-level adapters in Sec. 3.2" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.785, + 0.434, + 0.801 + ], + "angle": 0, + "content": "3.1 Input-level Adapters" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.809, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Fig. 3(a) provides an overview of RAW-Adapter. The solid lines in Fig. 3(a) left represent input-level adapters. Input-level adapters are designed to convert" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "Cui, Harada." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.589, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.146, + 0.788, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.411, + 0.789, + 0.468 + ], + "angle": 0, + "content": "Fig. 3: (a). Structure of RAW-Adapter. Solid line in left denotes input-level adapter's workflow and dotted line denotes model-level adapter's workflow, stage \\(1\\sim 4\\) means different stage of visual model backbone. (b). Detailed structure of kernel & matrix predictors \\(\\mathbb{P}_{\\mathbb{K}}\\), \\(\\mathbb{P}_{\\mathbb{M}}\\). (c). Detailed structure of model-level adapter \\(\\mathbb{M}\\)'s merge block." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.499, + 0.789, + 0.545 + ], + "angle": 0, + "content": "the RAW image \\(\\mathbf{I}_1\\) into the machine-vision oriented image \\(\\mathbf{I}_5\\). This process involves digital gain & denoise, demosacing, white balance adjustment, camera color matrix, and color manipulation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.546, + 0.79, + 0.653 + ], + "angle": 0, + "content": "We maintain the ISP design while simultaneously using query adaptive learning (QAL) to estimate key parameters in ISP stages. The QAL strategy is motivated by previous transformer models [4,9,13] and detailed structure is shown in Fig. 3(b), input image \\(\\mathbf{I}_{i\\in (1,2)}\\) would pass by 2 down-scale convolution blocks to generate feature, then feature pass by 2 linear layers to generate attention block's key \\(\\mathbb{k}\\) and value \\(\\mathbb{V}\\), while query \\(\\mathbb{q}\\) is a set of learnable dynamic parameters, the ISP parameters would be predicted by self-attention calculation [66]:" + }, + { + "type": "equation", + "bbox": [ + 0.343, + 0.665, + 0.787, + 0.7 + ], + "angle": 0, + "content": "\\[\np a r a m e t e r s = F F N \\left(\\operatorname {s o f t m a x} \\left(\\frac {\\mathbb {Q} \\cdot \\mathbb {k} ^ {T}}{\\sqrt {d _ {k}}}\\right) \\cdot \\mathbb {V}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.707, + 0.789, + 0.752 + ], + "angle": 0, + "content": "where \\( FFN \\) denotes the feed-forward network, includes 2 linear layers and 1 activation layer, the predicted parameters would keep the same length as query q. We defined 2 QAL blocks \\( \\mathbb{P}_{\\mathbb{K}} \\) and \\( \\mathbb{P}_{\\mathbb{M}} \\) to predict different part parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.753, + 0.789, + 0.786 + ], + "angle": 0, + "content": "The input RAW image \\(\\mathbf{I}_1\\) would first go through the pre-process operations and a demosaing stage [16,35], followed by subsequent ISP processes:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Gain & Denoise: Denoising algorithms always take various factors into account, such as input photon number, ISO gain level, exposure settings. Here we" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.634, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RAW-Adapter" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.146, + 0.788, + 0.253 + ], + "angle": 0, + "content": "first use QAL block \\(\\mathbb{P}_{\\mathbb{K}}\\) to predict a gain ratio \\(g\\) [2,6] to adapt \\(\\mathbf{I}_1\\) in different lighting scenarios, followed by an adaptive anisotropic Gaussian kernel to suppress noise under various noise conditions, \\(\\mathbb{P}_{\\mathbb{K}}\\) will predict the appropriate Gaussian kernel \\(k\\) for denoising to improve downstream visual tasks' performance, the predicted key parameters are the Gaussian kernel's major axis \\(r_1\\), minor axis \\(r_2\\) (see Fig. 4 left), here we set the kernel angle \\(\\theta\\) to 0 for simplification. Gaussian kernel \\(k\\) at pixel \\((x,y)\\) would be:" + }, + { + "type": "equation", + "bbox": [ + 0.363, + 0.267, + 0.785, + 0.284 + ], + "angle": 0, + "content": "\\[\nk (x, y) = \\exp \\left(- \\left(b _ {0} x ^ {2} + 2 b _ {1} x y + b _ {2} y ^ {2}\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.288, + 0.267, + 0.3 + ], + "angle": 0, + "content": "where:" + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.306, + 0.785, + 0.339 + ], + "angle": 0, + "content": "\\[\nb _ {0} = \\frac {\\cos (\\theta) ^ {2}}{2 r _ {1} ^ {2}} + \\frac {\\sin (\\theta) ^ {2}}{2 r _ {2} ^ {2}}, b _ {1} = \\frac {\\sin (2 \\theta)}{4 r _ {1} ^ {2}} ((\\frac {r _ {1}}{r _ {2}}) ^ {2} - 1), b _ {2} = \\frac {\\sin (\\theta) ^ {2}}{2 r _ {1} ^ {2}} + \\frac {\\cos (\\theta) ^ {2}}{2 r _ {2} ^ {2}}. (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.343, + 0.784, + 0.42 + ], + "angle": 0, + "content": "After gain ratio \\( g \\) and kernel \\( k \\) process on image \\( \\mathbf{I}_1 \\), \\( \\mathbb{P}_{\\mathbb{K}} \\) also predict a filter parameter \\( \\sigma \\) (initial at 0) to keep the sharpness and recover details of generated image \\( \\mathbf{I}_2 \\). Eq. 4 shows the translation from \\( \\mathbf{I}_1 \\) to \\( \\mathbf{I}_2 \\), where \\( * \\) denotes kernel convolution and filter parameter \\( \\sigma \\) is limited in a range of \\( (0, 1) \\) by a Sigmoid activation. For more details please refer to our supplementary part Sec. C" + }, + { + "type": "equation", + "bbox": [ + 0.398, + 0.431, + 0.606, + 0.446 + ], + "angle": 0, + "content": "\\[\n\\mathbb {P} _ {\\mathbb {K}} \\left(\\mathbf {I} _ {1}, \\mathrm {q}\\right)\\rightarrow k \\left\\{r _ {1}, r _ {2}, \\theta \\right\\}, g, \\sigma ,\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.398, + 0.45, + 0.785, + 0.465 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} _ {2} ^ {\\prime} = (g \\cdot \\mathbf {I} _ {1}) * k, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.398, + 0.468, + 0.575, + 0.484 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} _ {2} = \\mathbf {I} _ {2} ^ {\\prime} + (g \\cdot \\mathbf {I} _ {1} - \\mathbf {I} _ {2} ^ {\\prime}) \\cdot \\sigma .\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.501, + 0.788, + 0.667 + ], + "angle": 0, + "content": "White Balance & CCM Matrix: White balance (WB) mimics the color constancy of the human vision system (HVS) by aligning \"white\" color with the white object, resulting in the captured image reflecting the combination of light color and material reflectance [1,2,14]. In our work, we hope to find an adaptive white balance for different images under various lighting scenarios. Motivated by the design of Shades of Gray (SoG) [20] WB algorithm, where gray-world WB and Max-RGB WB can be regarded as a subcase, we've employed a learnable parameter \\(\\rho\\) to replace the gray-world's L-1 average with an adaptive Minkowski distance average (see Eq. [5]), the QAL block \\(\\mathbb{P}_{\\mathbb{M}}\\) predicts Minkowski distance's hyper-parameter \\(\\rho\\), a demonstration of various \\(\\rho\\) is shown in Fig. [4]. After finding the suitable \\(\\rho\\), \\(\\mathbf{I}_2\\) would multiply to white balance matrix to generate \\(\\mathbf{I}_3\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.668, + 0.788, + 0.744 + ], + "angle": 0, + "content": "Color Conversion Matrix (CCM) within the ISP is constrained by the specific camera model. Here we standardize the CCM as a single learnable \\(3 \\times 3\\) matrix, initialized as a unit diagonal matrix \\(\\mathbf{E}_3\\). The QAL block \\(\\mathbb{P}_{\\mathbb{M}}\\) predicts 9 parameters here, which are then added to \\(\\mathbf{E}_3\\) to form the final \\(3 \\times 3\\) matrix \\(\\mathbf{E}_{ccm}\\). Then \\(\\mathbf{I}_3\\) would multiply to CCM \\(\\mathbf{E}_{ccm}\\) to generate \\(\\mathbf{I}_4\\), equation as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.355, + 0.756, + 0.503, + 0.77 + ], + "angle": 0, + "content": "\\[\n\\mathbb {P} _ {\\mathbb {M}} \\left(\\mathbf {I} _ {2}, \\mathrm {q}\\right)\\rightarrow \\rho , \\mathbf {E} _ {c c m},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.355, + 0.776, + 0.641, + 0.793 + ], + "angle": 0, + "content": "\\[\nm _ {i \\in (r, g, b)} = \\sqrt [ r ]{a v g \\left(\\mathbf {I} _ {2} (i) ^ {\\rho}\\right)} / \\sqrt [ r ]{a v g \\left((\\mathbf {I} _ {2}) ^ {\\rho}\\right)},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.793, + 0.785, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} _ {3} = \\mathbf {I} _ {2} * \\left[ \\begin{array}{c c c} m _ {r} & & \\\\ & m _ {g} & \\\\ & & m _ {b} \\end{array} \\right], \\quad \\mathbf {I} _ {4} = \\mathbf {I} _ {3} * \\mathbf {E} _ {c c m}. \\tag {5}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.129 + ], + "angle": 0, + "content": "Cui, Harada." + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.145, + 0.773, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.272, + 0.788, + 0.302 + ], + "angle": 0, + "content": "Fig. 4: Left, we use query adaptive learning (QAL) to predict key parameters is ISP process. Right, we show RAW-Adapter different blocks' parameter." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.331, + 0.789, + 0.483 + ], + "angle": 0, + "content": "Color Manipulation Process: In ISP, color manipulation process is commonly achieved through lookup table (LUT), such as 1D and 3D LUTs [16]. Here, we consolidate color manipulation operations into a single 3D LUT, adjusting the color of image \\(\\mathbf{I}_4\\) to produce output image \\(\\mathbf{I}_5\\). Leveraging advancements in LUT techniques, we choose the latest neural implicit 3D LUT (NILUT) [11], for its speed efficiency and ability to facilitate end-to-end learning in our pipeline [3]. We denote NILUT [11] as \\(\\mathbb{L}\\), \\(\\mathbb{L}\\) maps the input pixel intensities \\(R,G,B\\) to a continuous coordinate space, followed by the utilization of implicit neural representation (INR) [61], which involves using a multi-layer perceptron (MLP) network to map the output pixel intensities to \\(R^{\\prime},G^{\\prime},B^{\\prime}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.385, + 0.498, + 0.787, + 0.515 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} _ {5} \\left(R ^ {\\prime}, G ^ {\\prime}, B ^ {\\prime}\\right) = \\mathbb {L} \\left(\\mathbf {I} _ {4} (R, G, B)\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.569 + ], + "angle": 0, + "content": "Image \\(\\mathbf{I}_5\\) obtained through image-level adapters will be forwarded to the downstream network's backbone. Furthermore, \\(\\mathbf{I}_5\\)'s features obtained in backbone will be fused with model-level adapters, which we will discuss in detail." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.592, + 0.441, + 0.608 + ], + "angle": 0, + "content": "3.2 Model-level Adapters" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.619, + 0.788, + 0.725 + ], + "angle": 0, + "content": "Input-level adapters guarantee the production of machine vision-oriented image \\(\\mathbf{I}_5\\) for high-level models. However, information at the ISP stages (\\(\\mathbf{I}_1 \\sim \\mathbf{I}_4\\)) is almost overlooked. Inspired by adapter design in current NLP and CV area [8,28,33], we employ the prior information from the ISP stage as model-level adapters to guide subsequent models' perception. Additionally, model-level adapters promote a tighter integration between downstream tasks and the ISP stage. Dotted lines in Fig. 3(a) represent model-level adapters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.725, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Model-level adapters \\(\\mathbb{M}\\) integrate the information from ISP stages to the network backbone. As shown in Fig. 3(a), we denote the stage \\(1\\sim 4\\) as different stages in network backbone, image \\(\\mathbf{I}_5\\) would pass by backbone stages \\(1\\sim 4\\) and then followed by detection or segmentation head. We utilize convolution layer \\(c_{i}\\) to extract features from \\(\\mathbf{I}_1\\) through \\(\\mathbf{I}_4\\), these extracted features are concatenated as \\(\\mathbb{C}(\\mathbf{I}_{1\\sim 4}) = \\mathbb{C}(c_1(\\mathbf{I}_1),c_2(\\mathbf{I}_2),c_3(\\mathbf{I}_3),c_4(\\mathbf{I}_4))\\). Subsequently, \\(\\mathbb{C}(\\mathbf{I}_{1\\sim 4})\\) goes through" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.824, + 0.724, + 0.843 + ], + "angle": 0, + "content": "3 To save memory occupy, we change the channel number 128 in [11] to 32." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.634, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RAW-Adapter" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.237 + ], + "angle": 0, + "content": "two residual blocks [25] and generate adapter \\( f_{1} \\), \\( f_{1} \\) would merge with stage 1's feature by a merge block, detail structure of merge block is shown in Fig. 3(c), which roughly includes concatenate process and a residual connect, finally merge block would output network feature for stage 2 and an additional adapter \\( f_{2} \\). Then process would repeat in stage 2 and stage 3 of network backbone. We collectively refer to all structures related to model-level adapters as \\( \\mathbb{M} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.788, + 0.389 + ], + "angle": 0, + "content": "We show RAW-Adapter's different part parameter number in Fig. 4 right, including input-level adapters \\(\\{\\mathbb{P}_{\\mathbb{K}}\\) (37.57K), \\(\\mathbb{P}_{\\mathbb{M}}\\) (37.96K), \\(\\mathbb{L}\\) (1.97K) and model level adapters \\(\\mathbb{M}\\) (114.87K \\(\\sim\\) 687.59K), model level adapters' parameter number depend on following network. Total parameter number of RAW-Adapter is around 0.2M to 0.8M, much smaller than following backbones (\\(\\sim\\)25.6M in ResNet-50 [25], \\(\\sim\\)197M in Swin-L [45]), also smaller than previous SOTA methods like SID [6] (11.99M) and Dirty-Pixel [18] (4.28M). Additionally, subsequent experiments will demonstrate that the performance of RAW-Adapter on backbone networks with fewer parameters is even superior to that of previous algorithms on backbone networks with higher parameters." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.416, + 0.376, + 0.433 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.452, + 0.544, + 0.467 + ], + "angle": 0, + "content": "4.1 Dataset and Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.784, + 0.526 + ], + "angle": 0, + "content": "Dataset. We conducted experiments on object detection and semantic segmentation tasks, utilizing a combination of various synthetic and real-world RAW image datasets, an overview of the datasets can be found in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.527, + 0.788, + 0.694 + ], + "angle": 0, + "content": "For the object detection task, we take 2 open source real-world dataset PASCAL RAW [52] and LOD [27]. PASCAL RAW [52] is a normal-light condition dataset with 4259 RAW images, taken by a Nikon D3200 DSLR camera with 3 object classes. To verify the generalization capability of RAW-Adapter across various lighting conditions, we additionally synthesized low-light and overexposure datasets based on the PASCAL RAW dataset, named PASCAL RAW (dark) and PASCAL RAW (over-exp) respectively, the synthesized datasets are identical to the PASCAL RAW dataset in all aspects except for brightness levels. For PASCAL RAW (dark) and PASCAL RAW (over-exp) synthesis, the light intensity and environment irradiance exhibit a linear relationship on RAW images, thus we employed the synthesis method from the previous work [14,78]:" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.706, + 0.616, + 0.741 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} x _ {n} \\sim N (\\mu = l x, \\sigma^ {2} = \\delta_ {r} ^ {2} + \\delta_ {s} l x) \\\\ y = l x + x _ {n}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.763, + 0.717, + 0.784, + 0.731 + ], + "angle": 0, + "content": "(7)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.84 + ], + "angle": 0, + "content": "where \\( x \\) denotes the original normal-light RAW image and \\( y \\) denotes degraded RAW image, \\( \\delta_s = \\sqrt{S} \\) denotes shot noise while \\( \\sqrt{S} \\) is signal of the sensor, \\( \\delta_r \\) denotes read noise, \\( l \\) denotes the light intensity parameter which randomly chosen from [0.05, 0.4] in PASCAL RAW (dark), and randomly chosen from [1.5, 2.5] in PASCAL RAW (over-exp). Additionally, we follow PASCAL RAW [52]'s dataset split to separate the training set and test set." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.129 + ], + "angle": 0, + "content": "Cui, Harada." + }, + { + "type": "table_caption", + "bbox": [ + 0.297, + 0.145, + 0.704, + 0.16 + ], + "angle": 0, + "content": "Table 1: Dataset and framework setting in our experiments." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.172, + 0.789, + 0.32 + ], + "angle": 0, + "content": "
PASCAL RAWPASCAL RAW (dark/over-exp)LODADE20K RAWADE20K RAW (dark/over-exp)
TaskObject DetectionSemantic Segmentation
Number4259223027,574
Typereal-worldsynthesisreal-worldsynthesis
SensorNikon D3200 DSLRCanon EOS 5D Mark IV-
FrameworkRetinaNet43 & Sparse-RCNN64Segformer70
BackboneResNet25MIT70
pre-trainImageNet44 pre-train weights
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.347, + 0.784, + 0.392 + ], + "angle": 0, + "content": "Meanwhile, LOD [27] is a real-world dataset with 2230 low-light condition RAW images, taken by a Canon EOS 5D Mark IV camera with 8 object classes, we take 1800 images as training set and the other 430 images as test set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.392, + 0.789, + 0.514 + ], + "angle": 0, + "content": "For the semantic segmentation task, we utilized the widely-used sRGB dataset ADE20K [80] to generate RAW dataset. Leveraging state-of-the-art unprocessing methods InvISP [71], we synthesized RAW images corresponding to the ADE20K sRGB dataset. Using InvISP, we projected input sRGB images into RAW format, effectively creating an ADE20K RAW dataset. To simulate various lighting conditions, we employed the same synthesis method from PASCAL RAW (dark/over-exp) to generate low-light and over-exposure RAW images, name as ADE20K RAW (dark/over-exp), training & test split is same as ADE20K." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.789, + 0.678 + ], + "angle": 0, + "content": "Implement Details. We build our framework based on the open-source computer vision toolbox mmdetection 7 and mmsegmentation 12, both object detection tasks and semantic segmentation tasks are initialed with ImageNet pre-train weights (see Table. 1), and we apply the data augmentation pipeline in the default setting, mainly include random crop, random flip, and multi-scale test, etc. For the object detection task, we adopt the 2 mainstream object detectors: RetinaNet 43 and Sparse-RCNN 64 with ResNet 25 backbone. For the semantic segmentation task, we choose to use the mainstream segmentation framework Segformer 70 with MIT 70 backbone." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.695, + 0.789, + 0.817 + ], + "angle": 0, + "content": "Comparison Methods. We conducted comparative experiments with the current state-of-the-art (SOTA) algorithms, including various open-sourced ISP methods [6,34,35,71,79] and joint-training method DirtyPixels [18], among these Karaimer et al. [35] is a traditional ISP method with various human manipulate steps, and InvISP [71] & Lite-ISP [79] & SID [6] & DNF [34] are current SOTA network-based ISP methods, where SID [6] and DNF [34] is especially for the low-light condition RAW data. For fairness compared with the above ISP methods, both the training and test RAW data are rendered using the respective" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.786, + 0.84 + ], + "angle": 0, + "content": "4 For more experiments on other frameworks, please refer to supplementary materials." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.634, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RAW-Adapter" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.143, + 0.788, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.358, + 0.785, + 0.389 + ], + "angle": 0, + "content": "Fig. 5: (a). Detection performance on PASCAL RAW [52] (normal/dark/over-exp). (b). Efficiency comparison (blue: vanilla, yellow: RAW-Adapter, gray: Dirty-Pixel [18])." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.788, + 0.493 + ], + "angle": 0, + "content": "compared ISP. DirtyPixels [18] is the current SOTA joint-training method which uses a stack of residual UNet [58] as a low-level processor and joint optimization with the following task-specific networks. For fairness, all comparison methods adopt the same data augmentation process and the same training setting, we will introduce detailed experiment results in the following section." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.513, + 0.499, + 0.529 + ], + "angle": 0, + "content": "4.2 Object Detection Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.789, + 0.78 + ], + "angle": 0, + "content": "For object detection task on PASCAL RAW [52] dataset, we adopt RetinaNet [43] with different size ResNet [25] backbones (ResNet-18, ResNet-50), all the models are trained on a single NVIDIA Tesla V100 GPU with SGD optimizer, the batch size is set to 4, training images are cropped into range of (400, 667) and training epochs are set to 50. Table. [2] shows the detection results with (a). ResNet-18 and (b). ResNet-50 backbone, with comparisons of demosaiced RAW data (\"Demosacing\"), camera default ISP in [52], various current SOTA ISP solutions [6,34,35,79] and Dirty-Pixel [18], we can observe that sometimes ISP algorithms in normal and over-exposure conditions may even degraded the detection performance, and ISP in dark scenarios sometimes could improves detection performance, additionally manual design ISP solution [35] even out-perform the deep learning solutions. Previous joint-training method Dirty-Pixel [18] can improve detection performance. Overall, our RAW-Adapter method achieves the best performance and even outperforms some ISP algorithms utilizing ResNet-50 backbone when employing ResNet-18 backbone, which can significantly reduce computational load and model parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Visualization of PASCAL RAW's detection results are shown in Fig. 5(a), we show the detection results under different lightness conditions, background are images generated from Dirty-Pixel [18] and RAW-Adapter \\((\\mathbf{I}_5)\\). Our method could achieve satisfactory detection results across different lightness, while other" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.129 + ], + "angle": 0, + "content": "Cui, Harada." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.144, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 2: Comparison results on PASCAL RAW dataset [52], we take [43] with (a). ResNet-18 and (b). ResNet-50 backbone, bold denotes the best result." + }, + { + "type": "table_caption", + "bbox": [ + 0.31, + 0.197, + 0.4, + 0.21 + ], + "angle": 0, + "content": "(a) ResNet-18." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.221, + 0.495, + 0.399 + ], + "angle": 0, + "content": "
mAP Methodnormal over-expdark
Default ISP88.3--
Demosacing87.787.780.3
Karäimer et al. 3586.085.681.9
Lite-ISP 7985.284.271.9
InvISP 7184.186.670.9
SID 6--78.2
DNF 34--81.1
Dirty-Pixel88.688.080.8
RAW-Adapter88.788.782.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.603, + 0.198, + 0.692, + 0.21 + ], + "angle": 0, + "content": "(b) ResNet-50." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.221, + 0.785, + 0.399 + ], + "angle": 0, + "content": "
mAP Methodnormal over-expdark
Default ISP89.4--
Demosacing89.288.882.6
Karäimer et al. 3589.486.884.6
Lite-ISP 7989.385.173.5
InvISP 7189.187.374.7
SID 6--81.5
DNF 34--82.8
Dirty-Pixel89.089.083.6
RAW-Adapter89.689.486.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.411, + 0.787, + 0.456 + ], + "angle": 0, + "content": "Table 3: Comparison with ISP methods and Dirty-Pixel on LOD dataset [27]. We show the detection performance (mAP) ↑ of RetinaNet (R-Net) [43] and Sparse-RCNN (Sp-RCNN) [64], **bold** denotes the best result." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.466, + 0.788, + 0.581 + ], + "angle": 0, + "content": "
MethodsDemosaicing Karaimer et al. 35InvISP 71SID 6
mAP(R-Net)58.554.456.949.1
mAP(Sp-RCNN)57.752.249.443.1
MethodsDNF 34Dirty-Pixel 18RAW-Adapter (w/o M)RAW-Adapter
mAP(R-Net)58.861.661.662.2
mAP(Sp-RCNN)55.858.858.659.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.609, + 0.785, + 0.654 + ], + "angle": 0, + "content": "methods face false alarm and miss detect. Efficiency comparison with Dirty-Pixel is shown in Fig. 5(b), our algorithm has achieved significant improvements in both training time acceleration and savings in Flops/Parameter Number." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.655, + 0.788, + 0.747 + ], + "angle": 0, + "content": "Detection performance on LOD [27] is shown in Table. [3], we adopt two detector RetinaNet [43] and Sparse-RCNN [64], both with ResNet-50 [25] backbone, training epochs are set to 35, Sparse-RCNN [64] are trained by 4 GPUs with Adam optimizer. With comparing of various ISP methods [6,34,35,79], Dirty-Pixel [18] and RAW-Adapter with only input-level adapters (w/o M). Table. [3] show that our algorithm can achieve the best performance on both two detectors." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.769, + 0.549, + 0.784 + ], + "angle": 0, + "content": "4.3 Semantic Segmentation Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.794, + 0.787, + 0.841 + ], + "angle": 0, + "content": "For semantic segmentation on ADE20K RAW dataset, we choose Segformer [70] as the segmentation framework with the different size MIT [70] backbones (MIT-B5, MIT-B3, MIT-B0), all the models are trained on 4 NVIDIA Tesla V100 GPU" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.634, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RAW-Adapter" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.331, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.206, + 0.33, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.253, + 0.267, + 0.31, + 0.277 + ], + "angle": 0, + "content": "Input / GT" + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.146, + 0.443, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.337, + 0.206, + 0.444, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.267, + 0.41, + 0.276 + ], + "angle": 0, + "content": "LiteISP" + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.145, + 0.558, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.206, + 0.558, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.486, + 0.267, + 0.521, + 0.276 + ], + "angle": 0, + "content": "InvISP" + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.145, + 0.67, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.206, + 0.67, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.267, + 0.645, + 0.277 + ], + "angle": 0, + "content": "Dirty-Pixel" + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.145, + 0.783, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.206, + 0.783, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.694, + 0.267, + 0.769, + 0.276 + ], + "angle": 0, + "content": "RAW-Adapter" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.28, + 0.331, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.343, + 0.33, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.406, + 0.311, + 0.417 + ], + "angle": 0, + "content": "Input / GT" + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.281, + 0.444, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.337, + 0.343, + 0.443, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.381, + 0.405, + 0.401, + 0.416 + ], + "angle": 0, + "content": "SID" + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.28, + 0.557, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.343, + 0.557, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.467, + 0.405, + 0.546, + 0.417 + ], + "angle": 0, + "content": "Karaimer et al." + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.28, + 0.67, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.343, + 0.67, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.586, + 0.405, + 0.642, + 0.417 + ], + "angle": 0, + "content": "Dirty-Pixel" + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.28, + 0.783, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.343, + 0.783, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.694, + 0.405, + 0.771, + 0.416 + ], + "angle": 0, + "content": "RAW-Adapter" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.429, + 0.785, + 0.46 + ], + "angle": 0, + "content": "Fig. 6: Semantic segmentation results on over-exposure and low-light RAW data, compare with various ISP methods [6], [35], [71], [79] and Dirty-Pixel [18]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.785, + 0.55 + ], + "angle": 0, + "content": "with Adam optimizer, the batch size is set to 4, training images are cropped into \\(512 \\times 512\\) and training iter are set to 80000. We make comparison with single demosacing, various ISP methods [6,34,35,71,79] and Dirty-Pixel [18], and RAW-Adapter with only input-level adapters (w/o M)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Comparison results are shown in Table. 4, where we compare both efficiency (parameters↓, inference time↓) and performance (mIOU↑), inference time is calculated on a single Tesla V100 GPU. We first show the results on vanilla demosaiced RAW data, then we show the results on RAW data processed by various ISP methods [6,34,35,71,79], followed by results of Dirty-Pixel [18], and our method without and with model-level adapter M. For Table. 4, we can find that directly employing ISP methods sometimes fails to enhance the performance of downstream segmentation tasks, particularly in low-light conditions. Joint-training method Dirty-Pixel [18] could improve down-stream performance and sometimes even degrade segmentation performance, but increase more network parameters. Overall, our proposed RAW-Adapter could effectively enhance performance while adding a slight parameter count and improving inference efficiency. RAW-Adapter on backbones with fewer parameters can even outperform previous methods on backbones with more parameters (i.e. The RAW-Adapter, employing the MIT-B3 backbone, achieves a dark condition mIOU of 40.52, surpassing other methods utilizing the MIT-B5 backbone despite having fewer than \\(\\sim\\) 36M parameters.)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "The visualization of segmentation results are shown in Fig. 6, where row 1 and row 3 depict images rendered by different ISPs [6, 35, 71, 79] and images" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.129 + ], + "angle": 0, + "content": "Cui, Harada." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 4: Comparison with previous methods on ADE 20K RAW (normal/over-exp/dark). Bold denotes the best result while underline denotes second best result." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.186, + 0.789, + 0.477 + ], + "angle": 0, + "content": "
backboneparams(M) ↓inference time(s) ↓mIOU ↑ (normal)mIOU ↑ (over-exp)mIOU ↑ (dark)
DemosacingMIT-B582.010.10545.2642.0830.97
Karaimer et al. [35]0.52546.4243.7139.79
InvISP [71]0.20347.5343.554.87
LiteISP [79]0.26143.2242.015.52
DNF [34]0.186--25.82
SID [6]0.312--26.44
Dirty-Pixel [18]86.290.15947.7646.8040.02
MIT-B348.920.09847.1945.3338.93
MIT-B08.000.04935.4332.1027.19
RAW-Adapter (w/o M)MIT-B582.090.14848.3346.4840.91
MIT-B344.720.08647.0145.0039.34
MIT-B03.800.03235.0932.8227.29
RAW-AdapterMIT-B582.530.16748.3847.0641.82
MIT-B345.160.10247.5945.6640.52
MIT-B04.260.05335.6733.4427.78
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.79, + 0.626 + ], + "angle": 0, + "content": "generated from Dirty-Pixel [18] and RAW-Adapter \\((\\mathbf{I}_5)\\), row 2 and row 4 denote our segmentation results appear superior compared to other methods. We were even pleased to discover that, although we do not add any human-vision-oriented loss function to constrain the model, RAW-Adapter still produces visually satisfactory image \\(\\mathbf{I}_5\\) (see Fig. 6). This may be attributed to our preservation of traditional ISP stages in input-level adapters. Due to page limitations, we refer to more experimental results, visualization results, and ablation analysis of our method in the supplementary part." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.645, + 0.36, + 0.661 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.791, + 0.749 + ], + "angle": 0, + "content": "In this paper, we introduce RAW-Adapter, an effective solution for adapting pre-trained sRGB models to camera RAW data. With input-level adapters and model-level adapters working in tandem, RAW-Adapter effectively forwards RAW images from various lighting conditions to downstream vision tasks, our method has achieved state-of-the-art (SOTA) results across multiple datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.789, + 0.84 + ], + "angle": 0, + "content": "For future research directions, we believe it is feasible to design a unified model capable of adapting to RAW visual tasks under different lighting conditions, without the need for retraining in each lighting scenario like RAW-Adapter. Additionally, designing multi-task decoders can enable the accommodation of various tasks, facilitating more effective integration of different visual tasks on RAW images with large-scale models." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.634, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RAW-Adapter" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.176, + 0.788, + 0.222 + ], + "angle": 0, + "content": "This research is partially supported by JST Moonshot R&D Grant Number JPMJPS2011, CREST Grant Number JPMJCR2015 and Basic Research Grant (Super AI) of Institute for AI and Beyond of the University of Tokyo." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.244, + 0.323, + 0.26 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.274, + 0.788, + 0.331 + ], + "angle": 0, + "content": "1. Affifi, M., Brown, M.S.: What else can fool deep learning? addressing color constancy errors on deep neural network performance. 2019 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 243-252 (2019), https://apisemantic scholar.org/CorpusID:207995696" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.331, + 0.788, + 0.372 + ], + "angle": 0, + "content": "2. Brooks, T., Mildenhall, B., Xue, T., Chen, J., Sharlet, D., Barron, J.T.: Unprocessing images for learned raw denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11036-11045 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.372, + 0.788, + 0.413 + ], + "angle": 0, + "content": "3. Buckler, M., Jayasuriya, S., Sampson, A.: Reconfiguring the imaging pipeline for computer vision. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 975-984 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.413, + 0.788, + 0.455 + ], + "angle": 0, + "content": "4. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.455, + 0.788, + 0.495 + ], + "angle": 0, + "content": "5. Chatterjee, P., Joshi, N., Kang, S.B., Matsushita, Y.: Noise suppression in low-light images through joint denoising and demosaicing. In: CVPR 2011. pp. 321-328. IEEE (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.496, + 0.788, + 0.537 + ], + "angle": 0, + "content": "6. Chen, C., Chen, Q., Xu, J., Koltun, V.: Learning to see in the dark. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3291-3300 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.537, + 0.788, + 0.565 + ], + "angle": 0, + "content": "7. Chen, K., Wang, J., et al.: Mmdetection: Open mmlab detection toolbox and benchmark (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.788, + 0.607 + ], + "angle": 0, + "content": "8. Chen, Z., Duan, Y., Wang, W., He, J., Lu, T., Dai, J., Qiao, Y.: Vision transformer adapter for dense predictions. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=plKu2GByCNW" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.607, + 0.788, + 0.634 + ], + "angle": 0, + "content": "9. Cheng, B., Schwing, A.G., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.788, + 0.675 + ], + "angle": 0, + "content": "0. Conde, M.V., McDonagh, S., Maggioni, M., Leonardis, A., Pérez-Pellitero, E.: Model-based image signal processors via learnable dictionaries. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 481-489 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.788, + 0.717 + ], + "angle": 0, + "content": "1. Conde, M.V., Vazquez-Corral, J., Brown, M.S., Timofte, R.: Nilut: Conditional neural implicit 3d lookup tables for image enhancement. In: Proceedings of the AAAI Conference on Artificial Intelligence (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.788, + 0.744 + ], + "angle": 0, + "content": "2. Contributors, M.: MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.788, + 0.785 + ], + "angle": 0, + "content": "3. Cui, Z., Li, K., Gu, L., Su, S., Gao, P., Jiang, Z., Qiao, Y., Harada, T.: You only need 90k parameters to adapt light: a light weight transformer for image enhancement and exposure correction. In: BMVC. p. 238 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.788, + 0.84 + ], + "angle": 0, + "content": "4. Cui, Z., Qi, G.J., Gu, L., You, S., Zhang, Z., Harada, T.: Multitask aet with orthogonal tangent regularity for dark object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2553-2562 (October 2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.274, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.128 + ], + "angle": 0, + "content": "Cui, Harada." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "15. Cui, Z., Zhu, Y., Gu, L., Qi, G.J., Li, X., Zhang, R., Zhang, Z., Harada, T.: Exploring resolution and degradation clues as self-supervised signal for low quality object detection. In: European Conference on Computer Vision. pp. 473-491. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.205, + 0.787, + 0.26 + ], + "angle": 0, + "content": "16. Delbracio, M., Kelly, D., Brown, M.S., Milanfar, P.: Mobile computational photography: A tour. Annual Review of Vision Science 7(1), 571-604 (2021). https://doi.org/10.1146/annurev-vision-093019-115521, https://doi.org/10.1146/annurev-vision-093019-115521, pMID: 34524880" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.261, + 0.787, + 0.288 + ], + "angle": 0, + "content": "17. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A Large-Scale Hierarchical Image Database. In: CVPR09 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.289, + 0.787, + 0.331 + ], + "angle": 0, + "content": "18. Diamond, S., Sitzmann, V., Julca-Aguilar, F., Boyd, S., Wetzstein, G., Heide, F.: Dirty pixels: Towards end-to-end image processing and perception. ACM Transactions on Graphics (SIGGRAPH) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.332, + 0.787, + 0.374 + ], + "angle": 0, + "content": "19. Dong, X., Yokoya, N.: Understanding dark scenes by contrasting multi-modal observations. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 840-850 (January 2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.787, + 0.416 + ], + "angle": 0, + "content": "20. Finlayson, G.D., Trezzi, E.: Shades of gray and colour constancy. In: Color and Imaging Conference. vol. 2004, pp. 37-41. Society for Imaging Science and Technology (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.417, + 0.787, + 0.471 + ], + "angle": 0, + "content": "21. Goyal, B., Lalonde, J.F., Li, Y., Gupta, M.: Robust scene inference under noise-blur dual corruptions. 2022 IEEE International Conference on Computational Photography (ICCP) pp. 1-12 (2022), https://api_semanticscholar.org/CorpusID:251040807" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.514 + ], + "angle": 0, + "content": "22. Guo, Y., Luo, F., Wu, X.: Learning degradation-independent representations for camera isp pipelines. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 25774-25783 (June 2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.787, + 0.571 + ], + "angle": 0, + "content": "23. Hansen, N., Ostermeier, A.: Adapting arbitrary normal mutation distributions in evolution strategies: the covariance matrix adaptation. In: Proceedings of IEEE International Conference on Evolutionary Computation. pp. 312-317 (1996). https://doi.org/10.1109/ICEC.1996.542381" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.572, + 0.787, + 0.627 + ], + "angle": 0, + "content": "24. Hasinoff, S.W., Sharlet, D., Geiss, R., Adams, A., Barron, J.T., Kainz, F., Chen, J., Levoy, M.: Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM Transactions on Graphics (Proc. SIGGRAPH Asia) 35(6) (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.628, + 0.787, + 0.669 + ], + "angle": 0, + "content": "25. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.67, + 0.787, + 0.712 + ], + "angle": 0, + "content": "26. Heide, F., Steinberger, M., Tsai, Y.T., Rouf, M., Pajak, D., Reddy, D., Gallo, O., Liu, J., Heidrich, W., Egiazarian, K., et al.: Flexisp: A flexible camera image processing framework. ACM Transactions on Graphics (ToG) 33(6), 1-13 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.713, + 0.787, + 0.74 + ], + "angle": 0, + "content": "27. Hong, Y., Wei, K., Chen, L., Fu, Y.: Crafting object detection in very low light. In: BMVC (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.741, + 0.787, + 0.783 + ], + "angle": 0, + "content": "28. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.784, + 0.787, + 0.812 + ], + "angle": 0, + "content": "29. Hu, Y., He, H., Xu, C., Wang, B., Lin, S.: Exposure: A white-box photo post-processing framework. ACM Transactions on Graphics (TOG) 37(2), 26 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "30. Ignatov, A., Van Gool, L., Timofte, R.: Replacing mobile camera isp with a single deep learning model. arXiv preprint arXiv:2002.05509 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.634, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RAW-Adapter" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "31. Iscen, A., Zhang, J., Lazebnik, S., Schmid, C.: Memory-efficient incremental learning through feature adaptation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16. pp. 699-715. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.204, + 0.788, + 0.272 + ], + "angle": 0, + "content": "32. Jensen, H.W., Durand, F., Dorsey, J., Stark, M.M., Shirley, P., Premoze, S.: A physically-based night sky model. In: Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques. p. 399-408. SIGGRAPH '01, Association for Computing Machinery, New York, NY, USA (2001). https://doi.org/10.1145/383259.383306, https://doi.org/10.1145/383259.383306" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "33. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.315, + 0.787, + 0.342 + ], + "angle": 0, + "content": "34. Jin, X., Han, L., Li, Z., Chai, Z., Guo, C., Li, C.: Dnf: Decouple and feedback network for seeing in the dark. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.787, + 0.384 + ], + "angle": 0, + "content": "35. Karaimer, H.C., Brown, M.S.: A software platform for manipulating the camera imaging pipeline. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) Computer Vision - ECCV 2016. pp. 429-444. Springer International Publishing, Cham (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.411 + ], + "angle": 0, + "content": "36. Ke, L., Ye, M., Danelljan, M., Liu, Y., Tai, Y.W., Tang, C.K., Yu, F.: Segment anything in high quality. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.787, + 0.452 + ], + "angle": 0, + "content": "37. Kim, W., Kim, G., Lee, J., Lee, S., Baek, S.H., Cho, S.: Paramisp: Learned forward and inverse isps using camera parameters. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 26067-26076 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.508 + ], + "angle": 0, + "content": "38. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4015-4026 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.787, + 0.55 + ], + "angle": 0, + "content": "39. Koskinen, S., Yang, D., Kämäräinen, J.K.: Reverse imaging pipeline for raw rgb image augmentation. In: 2019 IEEE International Conference on Image Processing (ICIP). pp. 2896-2900 (2019). https://doi.org/10.1109/ICIP.2019.8804406" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.59 + ], + "angle": 0, + "content": "40. Li, Z., Lu, M., Zhang, X., Feng, X., Asif, M.S., Ma, Z.: Efficient visual computing with camera raw snapshots. TPAMI pp. 1-18 (2024). https://doi.org/10.1109/TPAMI.2024.3359326" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.787, + 0.633 + ], + "angle": 0, + "content": "41. Li, Z., Jiang, H., Cao, M., Zheng, Y.: Polarized color image denoising. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9873-9882 (2023). https://doi.org/10.1109/CVPR52729.2023.00952" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.787, + 0.674 + ], + "angle": 0, + "content": "42. Liang, Z., Cai, J., Cao, Z., Zhang, L.: Cameranet: A two-stage framework for effective camera isp learning. IEEE Transactions on Image Processing 30, 2248-2262 (2021). https://doi.org/10.1109/TIP.2021.3051486" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.715 + ], + "angle": 0, + "content": "43. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.77 + ], + "angle": 0, + "content": "44. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Fleet, D., Pajdla, T., Schiele, B., Tuytelaars, T. (eds.) Computer Vision - ECCV 2014. pp. 740-755. Springer International Publishing, Cham (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.787, + 0.813 + ], + "angle": 0, + "content": "45. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "46. Mildenhall, B., Hedman, P., Martin-Brualla, R., Srinivasan, P.P., Barron, J.T.: Nerf in the dark: High dynamic range view synthesis from noisy raw images. In:" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.128 + ], + "angle": 0, + "content": "Cui, Harada." + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.147, + 0.786, + 0.176 + ], + "angle": 0, + "content": "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16190-16199 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.787, + 0.218 + ], + "angle": 0, + "content": "47. Monakhova, K., Richter, S.R., Waller, L., Koltun, V.: Dancing under the stars: Video denoising in starlight. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16241-16251 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.787, + 0.272 + ], + "angle": 0, + "content": "48. Mosleh, A., Sharma, A., Onzon, E., Mannan, F., Robidoux, N., Heide, F.: Hardware-in-the-loop end-to-end optimization of camera image processing pipelines. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.273, + 0.787, + 0.315 + ], + "angle": 0, + "content": "49. Nam, S., Kim, S.J.: Modelling the scene dependent imaging in cameras with a deep neural network. 2017 IEEE International Conference on Computer Vision (ICCV) pp. 1726-1734 (2017), https://api.sementicscholar.org/CorpusID:23920472" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.37 + ], + "angle": 0, + "content": "50. Nguyen, R.M.H., Brown, M.S.: Raw image reconstruction using a self-contained srgb-jpeg image with only 64 kb overhead. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1655-1663 (2016). https://doi.org/10.1109/CVPR.2016.183" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.412 + ], + "angle": 0, + "content": "51. Nishimura, J., Gerasimow, T., Sushma, R., Sutic, A., Wu, C.T., Michael, G.: Automatic isp image quality tuning using nonlinear optimization. In: 2018 25th IEEE International Conference on Image Processing (ICIP). pp. 2471-2475. IEEE (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.438 + ], + "angle": 0, + "content": "52. Omid-Zohoor, A., Ta, D., Murmann, B.: Pascalraw: raw image database for object detection (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.439, + 0.787, + 0.493 + ], + "angle": 0, + "content": "53. Onzon, E., Mannan, F., Heide, F.: Neural auto-exposure for high-dynamic range object detection. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7706-7716 (2021). https://doi.org/10.1109/CVPR46437.2021.00762" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.494, + 0.787, + 0.536 + ], + "angle": 0, + "content": "54. Potlapalli, V., Zamir, S.W., Khan, S., Khan, F.: Prompt: Prompting for all-in-one image restoration. In: Thirty-seventh Conference on Neural Information Processing Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.536, + 0.787, + 0.578 + ], + "angle": 0, + "content": "55. Qin, H., Han, L., Wang, J., Zhang, C., Li, Y., Li, B., Hu, W.: Attention-aware learning for hyperparameter prediction in image processing pipelines. In: European Conference on Computer Vision. pp. 271-287. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.578, + 0.787, + 0.605 + ], + "angle": 0, + "content": "56. Ramanath, R., Snyder, W.E., Yoo, Y., Drew, M.S.: Color image processing pipeline. IEEE Signal processing magazine 22(1), 34-43 (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.606, + 0.787, + 0.633 + ], + "angle": 0, + "content": "57. Rebuffi, S.A., Bilen, H., Vedaldi, A.: Learning multiple visual domains with residual adapters. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.633, + 0.787, + 0.689 + ], + "angle": 0, + "content": "58. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Navab, N., Hornegger, J., Wells, W.M., Frangi, A.F. (eds.) Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015. pp. 234-241. Springer International Publishing, Cham (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.689, + 0.787, + 0.743 + ], + "angle": 0, + "content": "59. Sayed, M., Brostow, G.: Improved handling of motion blur in online object detection. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1706-1716 (2021). https://doi.org/10.1109/CVPR46437.2021.00175" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.744, + 0.787, + 0.785 + ], + "angle": 0, + "content": "60. Schwartz, E., Giryes, R., Bronstein, A.M.: Deepisp: Toward learning an end-to-end image processing pipeline. IEEE Transactions on Image Processing 28(2), 912-923 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.785, + 0.787, + 0.813 + ], + "angle": 0, + "content": "61. Sitzmann, V., Martel, J.N., Bergman, A.W., Lindell, D.B., Wetzstein, G.: Implicit neural representations with periodic activation functions. In: Proc. NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.813, + 0.787, + 0.841 + ], + "angle": 0, + "content": "62. Souza, M., Heidrich, W.: MetaISP - Exploiting Global Scene Structure for Accurate Multi-Device Color Rendition. In: Guthe, M., Grosch, T. (eds.) Vision, Modeling," + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.634, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RAW-Adapter" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.147, + 0.786, + 0.175 + ], + "angle": 0, + "content": "and Visualization. The Eurographics Association (2023). https://doi.org/10.2312/vmv.20231236" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.175, + 0.787, + 0.215 + ], + "angle": 0, + "content": "63. Stickland, A.C., Murray, I.: Bert and pals: Projected attention layers for efficient adaptation in multi-task learning. In: International Conference on Machine Learning. pp. 5986-5995. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.215, + 0.787, + 0.257 + ], + "angle": 0, + "content": "64. Sun, P., Zhang, R., Jiang, Y., Kong, T., Xu, C., Zhan, W., Tomizuka, M., Li, L., Yuan, Z., Wang, C., Luo, P.: SparseR-CNN: End-to-end object detection with learnable proposals. arXiv preprint arXiv:2011.12450 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.787, + 0.297 + ], + "angle": 0, + "content": "65. Sung, Y.L., Cho, J., Bansal, M.: Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5227-5237 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.297, + 0.787, + 0.338 + ], + "angle": 0, + "content": "66. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.787, + 0.392 + ], + "angle": 0, + "content": "67. Wang, Y., Yu, Y., Yang, W., Guo, L., Chau, L.P., Kot, A.C., Wen, B.: Raw image reconstruction with learned compact metadata. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18206-18215 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.392, + 0.787, + 0.433 + ], + "angle": 0, + "content": "68. Wei, K., Fu, Y., Yang, J., Huang, H.: A physics-based noise formation model for extreme low-light raw denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.787, + 0.5 + ], + "angle": 0, + "content": "69. Wu, C.T., Isikdogan, L.F., Rao, S., Nayak, B., Gerasimow, T., Sutic, A., Ainkedem, L., Michael, G.: Visionisp: Repurposing the image signal processor for computer vision applications. In: 2019 IEEE International Conference on Image Processing (ICIP). IEEE (Sep 2019). https://doi.org/10.1109/icip.2019.8803607, http://dx.doi.org/10.1109/ICIP.2019.8803607" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.787, + 0.542 + ], + "angle": 0, + "content": "70. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. arXiv preprint arXiv:2105.15203 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.787, + 0.582 + ], + "angle": 0, + "content": "71. Xing, Y., Qian, Z., Chen, Q.: Invertible image signal processing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6287-6296 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.787, + 0.636 + ], + "angle": 0, + "content": "72. Xu, R., Chen, C., Peng, J., Li, C., Huang, Y., Song, F., Yan, Y., Xiong, Z.: Toward raw object detection: A new benchmark and a new model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13384-13393 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.636, + 0.787, + 0.69 + ], + "angle": 0, + "content": "73. Yoshimura, M., Otsuka, J., Irie, A., Ohashi, T.: Dynamically controlled image signal processor for image recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 12866-12876 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.69, + 0.787, + 0.746 + ], + "angle": 0, + "content": "74. Yoshimura, M., Otsuka, J., Irie, A., Ohashi, T.: Rawgment: Noise-accounted raw augmentation enables recognition in a wide variety of environments. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14007-14017 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.746, + 0.787, + 0.786 + ], + "angle": 0, + "content": "75. Yu, K., Li, Z., Peng, Y., Loy, C.C., Gu, J.: Reconfigisp: Reconfigurable camera image processing pipeline. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4248-4257 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.787, + 0.84 + ], + "angle": 0, + "content": "76. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Cycleisp: Real image restoration via improved data synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2696-2705 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.128 + ], + "angle": 0, + "content": "Cui, Harada." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "77. Zhang, R., Zhang, W., Fang, R., Gao, P., Li, K., Dai, J., Qiao, Y., Li, H.: Tip-adapter: Training-free adaption of clip for few-shot classification. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 493-510. Springer Nature Switzerland, Cham (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.259 + ], + "angle": 0, + "content": "78. Zhang, Y., Zhang, Y., Zhang, Z., Zhang, M., Tian, R., Ding, M.: Isp-teacher: Image signal process with disentanglement regularization for unsupervised domain adaptive dark object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 7387-7395 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.787, + 0.286 + ], + "angle": 0, + "content": "79. Zhang, Z., Wang, H., Liu, M., Wang, R., Zuo, W., Zhang, J.: Learning raw-to-srgb mappings with inaccurately aligned supervision. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.286, + 0.787, + 0.328 + ], + "angle": 0, + "content": "80. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing through ade20k dataset. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 633-641 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.383 + ], + "angle": 0, + "content": "81. Zhou, W., Gao, S., Zhang, L., Lou, X.: Histogram of oriented gradients feature extraction from raw bayer pattern images. IEEE Transactions on Circuits and Systems II: Express Briefs 67(5), 946-950 (2020). https://doi.org/10.1109/TCSII.2020.2980557" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.383 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/RAW-Adapter_ Adapting Pretrained Visual Model to Camera RAW Images/a7bce0f0-b64e-4428-b3dc-7b7e2a217569_origin.pdf b/2024/RAW-Adapter_ Adapting Pretrained Visual Model to Camera RAW Images/a7bce0f0-b64e-4428-b3dc-7b7e2a217569_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d094f758b53230f41ee9a189680e29331190af31 --- /dev/null +++ b/2024/RAW-Adapter_ Adapting Pretrained Visual Model to Camera RAW Images/a7bce0f0-b64e-4428-b3dc-7b7e2a217569_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:df3fab8531e899b3901ab7fc7286250e7174d560a92f783c93e32058a2b2b555 +size 4801923 diff --git a/2024/RAW-Adapter_ Adapting Pretrained Visual Model to Camera RAW Images/full.md b/2024/RAW-Adapter_ Adapting Pretrained Visual Model to Camera RAW Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..acaf3cdd10f1ec01e6347ef49a3e848504a27477 --- /dev/null +++ b/2024/RAW-Adapter_ Adapting Pretrained Visual Model to Camera RAW Images/full.md @@ -0,0 +1,373 @@ +# RAW-Adapter: Adapting Pre-trained Visual Model to Camera RAW Images + +Ziteng Cui $^{1}$ and Tatsuya Harada $^{1,2}$ + +1 The University of Tokyo + +2 RIKEN AIP + +{cui, harada}@mi.t.u-tokyo.ac.jp + +Abstract. sRGB images are now the predominant choice for pre-training visual models in computer vision research, owing to their ease of acquisition and efficient storage. Meanwhile, the advantage of RAW images lies in their rich physical information under variable real-world challenging lighting conditions. For computer vision tasks directly based on camera RAW data, most existing studies adopt methods of integrating image signal processor (ISP) with backend networks, yet often overlook the interaction capabilities between the ISP stages and subsequent networks. Drawing inspiration from ongoing adapter research in NLP and CV areas, we introduce RAW-Adapter, a novel approach aimed at adapting sRGB pre-trained models to camera RAW data. RAW-Adapter comprises input-level adapters that employ learnable ISP stages to adjust RAW inputs, as well as model-level adapters to build connections between ISP stages and subsequent high-level networks. Additionally, RAW-Adapter is a general framework that could be used in various computer vision frameworks. Abundant experiments under different lighting conditions have shown our algorithm's state-of-the-art (SOTA) performance, demonstrating its effectiveness and efficiency across a range of real-world and synthetic datasets. Code is available at this url + +# 1 Introduction + +In recent years, there has been a growing interest in revisiting vision tasks using unprocessed camera RAW images. How to leverage information-rich RAW image in computer vision tasks has become a current research hotspot in various sub-areas (i.e. denoising [27], view synthesis [46], object detection [27, 48]). Compared to commonly used sRGB image, RAW image directly acquired by the camera sensor, encompasses abundant information unaffected or compressed by image signal processor (ISP), also offers physically meaningful information like noise distributions [47, 68], owing to its linear correlation between image intensity and the radiant energy received by a camera. The acquisition of RAW data, facilitating enhanced detail capture and a higher dynamic range, imparts a unique advantage in addressing visual tasks under variable lighting conditions in the real world. For instance, sunlight irradiance can reach as high as $1.3 \times 10^{3}W / m^{2}$ while bright planet irradiance can be as low as $2.0 \times 10^{-6}W / m^{2}$ [32]. + +![](images/d6f0e038f96a0dc83c9d3e0b6589ecb2e58610a9cbea038010c1e920d4bd3760.jpg) +Fig. 1: (a). An overview of basic image signal processor (ISP) pipeline. (b). ISP and current visual model have different objectives. (c) Previous methods optimize ISP with down-stream visual model. (d) Our proposed RAW-Adapter. + +Meanwhile, sRGB has emerged as the primary choice for pre-training visual models in today's computer vision field, due to its scalability and ease of storage. Typically, sRGB images are derived from camera RAW data through the ISP pipeline. As shown in Fig. 1(a), the entire ISP pipeline consists of multiple modules to convert RAW images into vision-oriented sRGB images, each of these modules serves its own distinct purpose, with the majority being ill-posed and heavily dependent on prior information [5,16,26,35,42,56]. + +When adopt RAW data for computer vision tasks, the purpose of a manually designed ISP is to produce images that offer a superior visual experience [18,69], rather than optimizing for downstream visual tasks such as object detection or semantic segmentation (see Fig. ①b)). Additionally, most companies' ISPs are black boxes, making it difficult to obtain information about the specific steps inside. Consequently, utilizing human vision-oriented ISP in certain conditions is sometimes even less satisfactory than directly using RAW [22,27,40,81]. + +To better take advantage of camera RAW data for various computer vision tasks. Researchers began to optimize the image signal processor (ISP) jointly with downstream networks. Since most ISP stages are non-differentiable and cannot be jointly backpropagation, there are two main lines of approaches to connect ISP stages with downstream networks (see Fig. 1(c)): ①. First-kind approaches maintain the modular design of the traditional ISP, involving the design of differentiable ISP modules through optimization algorithms 48,69,75, such as Hardware-in-the-loop 48 adopt covariance matrix adaptation evolution strategy (CMA-ES) 23. ②. Second-kind approaches replaced the ISP part entirely with a neural network 18,55,72,73, such as Dirty-Pixel 18 adopt a stack of residual UNets 58 as pre-encoder, however, the additional neural network introduces a significant computational burden, especially when dealing with high-resolution inputs. Beyond these, the above-mentioned methods still treat ISP and the backend neural network as two independent modules, lacking interac + +![](images/4d3460696f9b178ee2425f48e832603f1c4e3961befcd0d770a0d98282388dcd.jpg) +Fig. 2: Performance of RAW-based visual tasks with and without sRGB pre-trained weights. We analyze two methods: Dirty-Pixel [18] and RAW-Adapter. Blue line represents trained with MS COCO [44] pre-train weights, the purple line indicates ImageNet [17] pre-train weights, and the yellow line signifies training from scratch. + +![](images/d5317e4de476aa991ab27610f28f1e9956b0cfc91f38b155345715b96886099b.jpg) + +tion capabilities between two separate models. Especially, current visual models predominantly pre-trained on large amounts of sRGB images, there exists a significant gap between pre-trained models and camera RAW images. As it shown in Fig. 2 training from scratch with RAW data can significantly impair the performance of RAW-based vision tasks. How to leveraging sRGB pre-trained weights proves to be highly beneficial for RAW image visual tasks. Hence, here we ask: Is there a more effective way to adapt the information-rich RAW data into the knowledge-rich sRGB pre-trained models? + +Our solution, RAW-Adapter, differs from previous approaches that employed complex ISP or deeper neural networks at the input level. Instead, we prioritize simplifying the input stages and enhancing connectivity between ISP and subsequent networks at the model level. Inspired by recent advancements in prompt learning and adapter tuning [8,28,33,54], we developed two novel adapter approaches aimed at enhancing the integration between RAW input and sRGB pretrained models. Leveraging priors from ISP stages, our method includes input-level adapters and model-level adapters, as illustrated in Fig. [1(d)]. Input-level adapters are designed to adapt the input from RAW data to the backend network input, we follow traditional ISP design and predict ISP stages' key parameters to accommodate RAW image into the specific down-stream vision tasks, in this part, we employ strategies involving Query Adaptive Learning (QAL) and Implicit Neural Representation (INR) to ensure the differentiability of the core ISP processes while also upholding a lightweight design. Model-level adapters leverage prior knowledge from the input-level adapters by extracting intermediate stages as features, these features are then embedded into the downstream network, thus the network incorporates prior knowledge from the ISP stages, then jointly contributes to final machine-vision perception tasks. Our contributions could be summarized as follows: + +- We introduce RAW-Adapter, a novel framework aimed at improving the integration of sRGB pre-trained models to camera RAW images. Through the implementation of two types of adapters, we strive to narrow the dispar + +ity between RAW images and sRGB models, tackling discrepancies at both input and model levels. + +- By analyzing camera ISP steps, we've crafted input-level adapters that integrate query adaptive learning (QAL) and implicit neural representations (INR) to optimize ISP key parameters. Additionally, harnessing prior input stage information informs model-level adapters, enriching model understanding and improving downstream task performance. + +- We conducted detection and segmentation experiments involving different lighting scenarios, including normal-light, dark, and over-exposure scenes. Through comparisons with current mainstream ISPs and joint-training methods, sufficient experiments demonstrate that our algorithm achieved state-of-the-art (SOTA) performance. + +# 2 Related Works + +# 2.1 Image Signal Processor + +Typically, a camera's Image Signal Processor (ISP) pipeline is required to reconstruct a high-quality sRGB image from camera RAW data, traditional ISP pipeline is formulated as a series of manually crafted modules executed sequentially [16,35,51,56], including some representative steps such as demosaicing, white balance, noise removal, tone mapping, color space transform and so on. There also exist alternative designs of ISP process such as Heide et al. [26] designed FlexISP which combine numerous ISP blocks to a joint optimization block, and Hasinoff et al. [24] modified some of the traditional ISP steps for burst photography under extreme low-light conditions. When it comes to deep learning era, Chen et al. [6] proposed SID to use a UNet [58] replace the traditional ISP steps, which translate low-light RAW data to normal-light sRGB images, and Hu et al. [29] proposed a white-box solution with eight differentiable filters. After that, many efforts began to substitute the traditional ISP process with a neural network, such as [6,10,30,34,37,42,60,62,79], however deep network models are constrained by the training dataset, leading to shortcomings in generalization performance. Meanwhile another line of work focuses on translate sRGB images back to RAW data [2,39,49,50,67,71,76]. Apart from the conversion between RAW and RGB, there is ongoing research specifically dedicated to harnessing RAW images for downstream computer vision tasks. + +# 2.2 Computer Vision based on RAW data + +In order to effectively leverage information in RAW images for downstream visual tasks and save the time required for ISP, some early methods proposed to directly perform computer vision tasks on RAW images [3,81], which lacks consideration for the camera noise introduced during the conversion from photons to RAW images, especially in low-light conditions [41,47,68]. Meanwhile training from scratch on RAW data would abandon the current large-scale pre-trained + +visual model on sRGB data (see Fig. 2), especially the quantity of RAW image datasets (i.e. 4259 images in PASCAL RAW 52 dataset and 2230 images in LOD 27 dataset) is incomparable to current RGB datasets (i.e. over 1M images in ImageNet 17 dataset and SAM 38 is trained with 11M images). + +Most subsequent research focuses on finding ways to integrate ISP and backend computer vision models [18,22,48,55,69,72,73,75]. For example Wu et al. [69] first propose VisionISP and emphasized the difference between human vision and machine vision, they introduced several trainable modules in the ISP to enhance backend object detection performance. Since then several methods attempt to replace the existing non-differentiable ISP with a differentiable ISP network [48,75] or directly using encoder networks to replace the ISP process [18,55,73], however training two consecutive networks simultaneously can result in a significant consumption of computational resources, previous work also rarely focused on how to fine-tune the current mainstream sRGB visual models more efficiently on RAW data. Additionally, research like Rawgment [74], focusing on achieving realistic data augmentation directly on RAW images, some studies address the decrease in computer vision performance caused by specific steps within the incamera process or ISP, such as white balance error [1], auto-exposure error [53], camera motion blur [59] and undesirable camera noise [15,21]. + +# 2.3 Adapters in Computer Vision + +Adapters have become prevalent NLP area, which introducing new modules in large language model (LLM) such as [28,63], these modules enable task-specific fine-tuning, allowing pre-trained models to swiftly adapt to downstream NLP tasks. In the computer vision area, adapters have been adopted in various areas such as incremental learning [31] and domain adaptation [57]. Recently, a series of adapters have been used to investigate how to better utilize pre-trained visual models [8,33,36] or pre-trained vision-language models [65,77]. These methods focus on utilizing prior knowledge to make pre-trained models quickly adapt to downstream tasks. Here, we introduce the RAW-Adapter to further improve the alignment between RAW inputs and sRGB pre-trained visual models. + +# 3 RAW-Adapter + +In this section, we introduce RAW-Adapter. Due to page limitation, an overview of the conventional image signal processor (ISP) in supplementary's Sec. A. For specific adapter designs, we selectively omit certain steps and focus on some key steps within the camera ISP. In Sec. 3.1 we introduce RAW-Adapter's input-level adapters, followed by an explanation of its model-level adapters in Sec. 3.2 + +# 3.1 Input-level Adapters + +Fig. 3(a) provides an overview of RAW-Adapter. The solid lines in Fig. 3(a) left represent input-level adapters. Input-level adapters are designed to convert + +![](images/faa998769cf87654cfecc86abd9f3911d3467d37db1a38152e4e333ddb53cca3.jpg) +Fig. 3: (a). Structure of RAW-Adapter. Solid line in left denotes input-level adapter's workflow and dotted line denotes model-level adapter's workflow, stage $1\sim 4$ means different stage of visual model backbone. (b). Detailed structure of kernel & matrix predictors $\mathbb{P}_{\mathbb{K}}$ , $\mathbb{P}_{\mathbb{M}}$ . (c). Detailed structure of model-level adapter $\mathbb{M}$ 's merge block. + +![](images/6e4744b439528819ac969d3a50ece88406b7bd2cce496183f9e7b297ad424fee.jpg) + +the RAW image $\mathbf{I}_1$ into the machine-vision oriented image $\mathbf{I}_5$ . This process involves digital gain & denoise, demosacing, white balance adjustment, camera color matrix, and color manipulation. + +We maintain the ISP design while simultaneously using query adaptive learning (QAL) to estimate key parameters in ISP stages. The QAL strategy is motivated by previous transformer models [4,9,13] and detailed structure is shown in Fig. 3(b), input image $\mathbf{I}_{i\in (1,2)}$ would pass by 2 down-scale convolution blocks to generate feature, then feature pass by 2 linear layers to generate attention block's key $\mathbb{k}$ and value $\mathbb{V}$ , while query $\mathbb{q}$ is a set of learnable dynamic parameters, the ISP parameters would be predicted by self-attention calculation [66]: + +$$ +p a r a m e t e r s = F F N \left(\operatorname {s o f t m a x} \left(\frac {\mathbb {Q} \cdot \mathbb {k} ^ {T}}{\sqrt {d _ {k}}}\right) \cdot \mathbb {V}\right), \tag {1} +$$ + +where $FFN$ denotes the feed-forward network, includes 2 linear layers and 1 activation layer, the predicted parameters would keep the same length as query q. We defined 2 QAL blocks $\mathbb{P}_{\mathbb{K}}$ and $\mathbb{P}_{\mathbb{M}}$ to predict different part parameters. + +The input RAW image $\mathbf{I}_1$ would first go through the pre-process operations and a demosaing stage [16,35], followed by subsequent ISP processes: + +Gain & Denoise: Denoising algorithms always take various factors into account, such as input photon number, ISO gain level, exposure settings. Here we + +first use QAL block $\mathbb{P}_{\mathbb{K}}$ to predict a gain ratio $g$ [2,6] to adapt $\mathbf{I}_1$ in different lighting scenarios, followed by an adaptive anisotropic Gaussian kernel to suppress noise under various noise conditions, $\mathbb{P}_{\mathbb{K}}$ will predict the appropriate Gaussian kernel $k$ for denoising to improve downstream visual tasks' performance, the predicted key parameters are the Gaussian kernel's major axis $r_1$ , minor axis $r_2$ (see Fig. 4 left), here we set the kernel angle $\theta$ to 0 for simplification. Gaussian kernel $k$ at pixel $(x,y)$ would be: + +$$ +k (x, y) = \exp \left(- \left(b _ {0} x ^ {2} + 2 b _ {1} x y + b _ {2} y ^ {2}\right)\right), \tag {2} +$$ + +where: + +$$ +b _ {0} = \frac {\cos (\theta) ^ {2}}{2 r _ {1} ^ {2}} + \frac {\sin (\theta) ^ {2}}{2 r _ {2} ^ {2}}, b _ {1} = \frac {\sin (2 \theta)}{4 r _ {1} ^ {2}} ((\frac {r _ {1}}{r _ {2}}) ^ {2} - 1), b _ {2} = \frac {\sin (\theta) ^ {2}}{2 r _ {1} ^ {2}} + \frac {\cos (\theta) ^ {2}}{2 r _ {2} ^ {2}}. (3) +$$ + +After gain ratio $g$ and kernel $k$ process on image $\mathbf{I}_1$ , $\mathbb{P}_{\mathbb{K}}$ also predict a filter parameter $\sigma$ (initial at 0) to keep the sharpness and recover details of generated image $\mathbf{I}_2$ . Eq. 4 shows the translation from $\mathbf{I}_1$ to $\mathbf{I}_2$ , where $*$ denotes kernel convolution and filter parameter $\sigma$ is limited in a range of $(0, 1)$ by a Sigmoid activation. For more details please refer to our supplementary part Sec. C + +$$ +\mathbb {P} _ {\mathbb {K}} \left(\mathbf {I} _ {1}, \mathrm {q}\right)\rightarrow k \left\{r _ {1}, r _ {2}, \theta \right\}, g, \sigma , +$$ + +$$ +\mathbf {I} _ {2} ^ {\prime} = (g \cdot \mathbf {I} _ {1}) * k, \tag {4} +$$ + +$$ +\mathbf {I} _ {2} = \mathbf {I} _ {2} ^ {\prime} + (g \cdot \mathbf {I} _ {1} - \mathbf {I} _ {2} ^ {\prime}) \cdot \sigma . +$$ + +White Balance & CCM Matrix: White balance (WB) mimics the color constancy of the human vision system (HVS) by aligning "white" color with the white object, resulting in the captured image reflecting the combination of light color and material reflectance [1,2,14]. In our work, we hope to find an adaptive white balance for different images under various lighting scenarios. Motivated by the design of Shades of Gray (SoG) [20] WB algorithm, where gray-world WB and Max-RGB WB can be regarded as a subcase, we've employed a learnable parameter $\rho$ to replace the gray-world's L-1 average with an adaptive Minkowski distance average (see Eq. [5]), the QAL block $\mathbb{P}_{\mathbb{M}}$ predicts Minkowski distance's hyper-parameter $\rho$ , a demonstration of various $\rho$ is shown in Fig. [4]. After finding the suitable $\rho$ , $\mathbf{I}_2$ would multiply to white balance matrix to generate $\mathbf{I}_3$ . + +Color Conversion Matrix (CCM) within the ISP is constrained by the specific camera model. Here we standardize the CCM as a single learnable $3 \times 3$ matrix, initialized as a unit diagonal matrix $\mathbf{E}_3$ . The QAL block $\mathbb{P}_{\mathbb{M}}$ predicts 9 parameters here, which are then added to $\mathbf{E}_3$ to form the final $3 \times 3$ matrix $\mathbf{E}_{ccm}$ . Then $\mathbf{I}_3$ would multiply to CCM $\mathbf{E}_{ccm}$ to generate $\mathbf{I}_4$ , equation as follows: + +$$ +\mathbb {P} _ {\mathbb {M}} \left(\mathbf {I} _ {2}, \mathrm {q}\right)\rightarrow \rho , \mathbf {E} _ {c c m}, +$$ + +$$ +m _ {i \in (r, g, b)} = \sqrt [ r ]{a v g \left(\mathbf {I} _ {2} (i) ^ {\rho}\right)} / \sqrt [ r ]{a v g \left((\mathbf {I} _ {2}) ^ {\rho}\right)}, +$$ + +$$ +\mathbf {I} _ {3} = \mathbf {I} _ {2} * \left[ \begin{array}{c c c} m _ {r} & & \\ & m _ {g} & \\ & & m _ {b} \end{array} \right], \quad \mathbf {I} _ {4} = \mathbf {I} _ {3} * \mathbf {E} _ {c c m}. \tag {5} +$$ + +![](images/99aec0a1c43353c69e92d47427f2ed61936cfc1398fcfdab3e2c65b23660d24d.jpg) +Fig. 4: Left, we use query adaptive learning (QAL) to predict key parameters is ISP process. Right, we show RAW-Adapter different blocks' parameter. + +Color Manipulation Process: In ISP, color manipulation process is commonly achieved through lookup table (LUT), such as 1D and 3D LUTs [16]. Here, we consolidate color manipulation operations into a single 3D LUT, adjusting the color of image $\mathbf{I}_4$ to produce output image $\mathbf{I}_5$ . Leveraging advancements in LUT techniques, we choose the latest neural implicit 3D LUT (NILUT) [11], for its speed efficiency and ability to facilitate end-to-end learning in our pipeline [3]. We denote NILUT [11] as $\mathbb{L}$ , $\mathbb{L}$ maps the input pixel intensities $R,G,B$ to a continuous coordinate space, followed by the utilization of implicit neural representation (INR) [61], which involves using a multi-layer perceptron (MLP) network to map the output pixel intensities to $R^{\prime},G^{\prime},B^{\prime}$ : + +$$ +\mathbf {I} _ {5} \left(R ^ {\prime}, G ^ {\prime}, B ^ {\prime}\right) = \mathbb {L} \left(\mathbf {I} _ {4} (R, G, B)\right). \tag {6} +$$ + +Image $\mathbf{I}_5$ obtained through image-level adapters will be forwarded to the downstream network's backbone. Furthermore, $\mathbf{I}_5$ 's features obtained in backbone will be fused with model-level adapters, which we will discuss in detail. + +# 3.2 Model-level Adapters + +Input-level adapters guarantee the production of machine vision-oriented image $\mathbf{I}_5$ for high-level models. However, information at the ISP stages ( $\mathbf{I}_1 \sim \mathbf{I}_4$ ) is almost overlooked. Inspired by adapter design in current NLP and CV area [8,28,33], we employ the prior information from the ISP stage as model-level adapters to guide subsequent models' perception. Additionally, model-level adapters promote a tighter integration between downstream tasks and the ISP stage. Dotted lines in Fig. 3(a) represent model-level adapters. + +Model-level adapters $\mathbb{M}$ integrate the information from ISP stages to the network backbone. As shown in Fig. 3(a), we denote the stage $1\sim 4$ as different stages in network backbone, image $\mathbf{I}_5$ would pass by backbone stages $1\sim 4$ and then followed by detection or segmentation head. We utilize convolution layer $c_{i}$ to extract features from $\mathbf{I}_1$ through $\mathbf{I}_4$ , these extracted features are concatenated as $\mathbb{C}(\mathbf{I}_{1\sim 4}) = \mathbb{C}(c_1(\mathbf{I}_1),c_2(\mathbf{I}_2),c_3(\mathbf{I}_3),c_4(\mathbf{I}_4))$ . Subsequently, $\mathbb{C}(\mathbf{I}_{1\sim 4})$ goes through + +two residual blocks [25] and generate adapter $f_{1}$ , $f_{1}$ would merge with stage 1's feature by a merge block, detail structure of merge block is shown in Fig. 3(c), which roughly includes concatenate process and a residual connect, finally merge block would output network feature for stage 2 and an additional adapter $f_{2}$ . Then process would repeat in stage 2 and stage 3 of network backbone. We collectively refer to all structures related to model-level adapters as $\mathbb{M}$ . + +We show RAW-Adapter's different part parameter number in Fig. 4 right, including input-level adapters $\{\mathbb{P}_{\mathbb{K}}$ (37.57K), $\mathbb{P}_{\mathbb{M}}$ (37.96K), $\mathbb{L}$ (1.97K) and model level adapters $\mathbb{M}$ (114.87K $\sim$ 687.59K), model level adapters' parameter number depend on following network. Total parameter number of RAW-Adapter is around 0.2M to 0.8M, much smaller than following backbones ( $\sim$ 25.6M in ResNet-50 [25], $\sim$ 197M in Swin-L [45]), also smaller than previous SOTA methods like SID [6] (11.99M) and Dirty-Pixel [18] (4.28M). Additionally, subsequent experiments will demonstrate that the performance of RAW-Adapter on backbone networks with fewer parameters is even superior to that of previous algorithms on backbone networks with higher parameters. + +# 4 Experiments + +# 4.1 Dataset and Experimental Setting + +Dataset. We conducted experiments on object detection and semantic segmentation tasks, utilizing a combination of various synthetic and real-world RAW image datasets, an overview of the datasets can be found in Table 1. + +For the object detection task, we take 2 open source real-world dataset PASCAL RAW [52] and LOD [27]. PASCAL RAW [52] is a normal-light condition dataset with 4259 RAW images, taken by a Nikon D3200 DSLR camera with 3 object classes. To verify the generalization capability of RAW-Adapter across various lighting conditions, we additionally synthesized low-light and overexposure datasets based on the PASCAL RAW dataset, named PASCAL RAW (dark) and PASCAL RAW (over-exp) respectively, the synthesized datasets are identical to the PASCAL RAW dataset in all aspects except for brightness levels. For PASCAL RAW (dark) and PASCAL RAW (over-exp) synthesis, the light intensity and environment irradiance exhibit a linear relationship on RAW images, thus we employed the synthesis method from the previous work [14,78]: + +$$ +\begin{array}{l} x _ {n} \sim N (\mu = l x, \sigma^ {2} = \delta_ {r} ^ {2} + \delta_ {s} l x) \\ y = l x + x _ {n}, \\ \end{array} +$$ + +(7) + +where $x$ denotes the original normal-light RAW image and $y$ denotes degraded RAW image, $\delta_s = \sqrt{S}$ denotes shot noise while $\sqrt{S}$ is signal of the sensor, $\delta_r$ denotes read noise, $l$ denotes the light intensity parameter which randomly chosen from [0.05, 0.4] in PASCAL RAW (dark), and randomly chosen from [1.5, 2.5] in PASCAL RAW (over-exp). Additionally, we follow PASCAL RAW [52]'s dataset split to separate the training set and test set. + +Table 1: Dataset and framework setting in our experiments. + +
PASCAL RAWPASCAL RAW (dark/over-exp)LODADE20K RAWADE20K RAW (dark/over-exp)
TaskObject DetectionSemantic Segmentation
Number4259223027,574
Typereal-worldsynthesisreal-worldsynthesis
SensorNikon D3200 DSLRCanon EOS 5D Mark IV-
FrameworkRetinaNet43 & Sparse-RCNN64Segformer70
BackboneResNet25MIT70
pre-trainImageNet44 pre-train weights
+ +Meanwhile, LOD [27] is a real-world dataset with 2230 low-light condition RAW images, taken by a Canon EOS 5D Mark IV camera with 8 object classes, we take 1800 images as training set and the other 430 images as test set. + +For the semantic segmentation task, we utilized the widely-used sRGB dataset ADE20K [80] to generate RAW dataset. Leveraging state-of-the-art unprocessing methods InvISP [71], we synthesized RAW images corresponding to the ADE20K sRGB dataset. Using InvISP, we projected input sRGB images into RAW format, effectively creating an ADE20K RAW dataset. To simulate various lighting conditions, we employed the same synthesis method from PASCAL RAW (dark/over-exp) to generate low-light and over-exposure RAW images, name as ADE20K RAW (dark/over-exp), training & test split is same as ADE20K. + +Implement Details. We build our framework based on the open-source computer vision toolbox mmdetection 7 and mmsegmentation 12, both object detection tasks and semantic segmentation tasks are initialed with ImageNet pre-train weights (see Table. 1), and we apply the data augmentation pipeline in the default setting, mainly include random crop, random flip, and multi-scale test, etc. For the object detection task, we adopt the 2 mainstream object detectors: RetinaNet 43 and Sparse-RCNN 64 with ResNet 25 backbone. For the semantic segmentation task, we choose to use the mainstream segmentation framework Segformer 70 with MIT 70 backbone. + +Comparison Methods. We conducted comparative experiments with the current state-of-the-art (SOTA) algorithms, including various open-sourced ISP methods [6,34,35,71,79] and joint-training method DirtyPixels [18], among these Karaimer et al. [35] is a traditional ISP method with various human manipulate steps, and InvISP [71] & Lite-ISP [79] & SID [6] & DNF [34] are current SOTA network-based ISP methods, where SID [6] and DNF [34] is especially for the low-light condition RAW data. For fairness compared with the above ISP methods, both the training and test RAW data are rendered using the respective + +![](images/42d81234c539124aac78a75e373137568ffe4efd7191f5e118450b771dbefc12.jpg) +Fig. 5: (a). Detection performance on PASCAL RAW [52] (normal/dark/over-exp). (b). Efficiency comparison (blue: vanilla, yellow: RAW-Adapter, gray: Dirty-Pixel [18]). + +compared ISP. DirtyPixels [18] is the current SOTA joint-training method which uses a stack of residual UNet [58] as a low-level processor and joint optimization with the following task-specific networks. For fairness, all comparison methods adopt the same data augmentation process and the same training setting, we will introduce detailed experiment results in the following section. + +# 4.2 Object Detection Evaluation + +For object detection task on PASCAL RAW [52] dataset, we adopt RetinaNet [43] with different size ResNet [25] backbones (ResNet-18, ResNet-50), all the models are trained on a single NVIDIA Tesla V100 GPU with SGD optimizer, the batch size is set to 4, training images are cropped into range of (400, 667) and training epochs are set to 50. Table. [2] shows the detection results with (a). ResNet-18 and (b). ResNet-50 backbone, with comparisons of demosaiced RAW data ("Demosacing"), camera default ISP in [52], various current SOTA ISP solutions [6,34,35,79] and Dirty-Pixel [18], we can observe that sometimes ISP algorithms in normal and over-exposure conditions may even degraded the detection performance, and ISP in dark scenarios sometimes could improves detection performance, additionally manual design ISP solution [35] even out-perform the deep learning solutions. Previous joint-training method Dirty-Pixel [18] can improve detection performance. Overall, our RAW-Adapter method achieves the best performance and even outperforms some ISP algorithms utilizing ResNet-50 backbone when employing ResNet-18 backbone, which can significantly reduce computational load and model parameters. + +Visualization of PASCAL RAW's detection results are shown in Fig. 5(a), we show the detection results under different lightness conditions, background are images generated from Dirty-Pixel [18] and RAW-Adapter $(\mathbf{I}_5)$ . Our method could achieve satisfactory detection results across different lightness, while other + +Table 2: Comparison results on PASCAL RAW dataset [52], we take [43] with (a). ResNet-18 and (b). ResNet-50 backbone, bold denotes the best result. +(a) ResNet-18. + +
mAP Methodnormal over-expdark
Default ISP88.3--
Demosacing87.787.780.3
Karäimer et al. 3586.085.681.9
Lite-ISP 7985.284.271.9
InvISP 7184.186.670.9
SID 6--78.2
DNF 34--81.1
Dirty-Pixel88.688.080.8
RAW-Adapter88.788.782.5
+ +(b) ResNet-50. + +
mAP Methodnormal over-expdark
Default ISP89.4--
Demosacing89.288.882.6
Karäimer et al. 3589.486.884.6
Lite-ISP 7989.385.173.5
InvISP 7189.187.374.7
SID 6--81.5
DNF 34--82.8
Dirty-Pixel89.089.083.6
RAW-Adapter89.689.486.6
+ +Table 3: Comparison with ISP methods and Dirty-Pixel on LOD dataset [27]. We show the detection performance (mAP) ↑ of RetinaNet (R-Net) [43] and Sparse-RCNN (Sp-RCNN) [64], **bold** denotes the best result. + +
MethodsDemosaicing Karaimer et al. 35InvISP 71SID 6
mAP(R-Net)58.554.456.949.1
mAP(Sp-RCNN)57.752.249.443.1
MethodsDNF 34Dirty-Pixel 18RAW-Adapter (w/o M)RAW-Adapter
mAP(R-Net)58.861.661.662.2
mAP(Sp-RCNN)55.858.858.659.2
+ +methods face false alarm and miss detect. Efficiency comparison with Dirty-Pixel is shown in Fig. 5(b), our algorithm has achieved significant improvements in both training time acceleration and savings in Flops/Parameter Number. + +Detection performance on LOD [27] is shown in Table. [3], we adopt two detector RetinaNet [43] and Sparse-RCNN [64], both with ResNet-50 [25] backbone, training epochs are set to 35, Sparse-RCNN [64] are trained by 4 GPUs with Adam optimizer. With comparing of various ISP methods [6,34,35,79], Dirty-Pixel [18] and RAW-Adapter with only input-level adapters (w/o M). Table. [3] show that our algorithm can achieve the best performance on both two detectors. + +# 4.3 Semantic Segmentation Evaluation + +For semantic segmentation on ADE20K RAW dataset, we choose Segformer [70] as the segmentation framework with the different size MIT [70] backbones (MIT-B5, MIT-B3, MIT-B0), all the models are trained on 4 NVIDIA Tesla V100 GPU + +![](images/43043d2a432760b276f0eaa25b628295141b3111f94a5b4507b9b95f2684bbe7.jpg) + +![](images/c8193287929b9033283ff25322727b8a24c30031cbe015f9b52a33d0f0ecc9c3.jpg) +Input / GT + +![](images/e4f2a078d5327cf3870558799ae48cbe0487cc4586202ea0f209d8756e71de6e.jpg) + +![](images/76af0ecf44d389d21af0d474151c8e9c61beb874a7b96e1f43881a5c46f24645.jpg) +LiteISP + +![](images/44c2d3340b05c06f0411900e178c0de0a616f8f8e2beb622cc0d24ccbb86a69a.jpg) + +![](images/8b21f471d384198de87064f75c2ed219886469de1c075a3d665cdbffcdcde120.jpg) +InvISP + +![](images/cf27c40c973a7170e58f2816d4c3e3d2d84f862a60eeba7b28d2aeaa320426bd.jpg) + +![](images/628ac741a03d8487f813f8227eb1e1eb9a93781a41656fa7e26475947a1f20c4.jpg) +Dirty-Pixel + +![](images/c3d438cf47663ee167d54b4df4a571b1946916bde817ee7f245cb2f8b76c047c.jpg) + +![](images/d0c36cdfb0f926effb510b67d7ad725e9d8ae65ddd28d3a1dbbc69681d896acc.jpg) +RAW-Adapter + +![](images/1514dfba31eff02384302b16778c9ac4ceceb7f59229d600c11bf0420f1dd62b.jpg) + +![](images/b44d4e10413562e366e88c7cc87cbde419ac36ebac93b11343f500c94b24eb4e.jpg) +Input / GT + +![](images/766e832983604b3b8e0a1bcbbb5d5f3c6b75844c98adc6d5eac514048e16e010.jpg) + +![](images/eaf10951855c1e7b00355437e9bf43731f5ff6348214d1a84434974dc6210f61.jpg) +SID + +![](images/546c157963da137b9df9751c2c0e23d803f8f81bf9f26209e8bebf0708c4459b.jpg) + +![](images/dc170adc5338b02bf615689f40891bfc32bd0a0ff064c196e6f2aff99c71ae0e.jpg) +Karaimer et al. + +![](images/7a65f06628b24f3f75d006a461f46f87c194ffea49dddd0dfa54024eb7583f65.jpg) + +![](images/111aad8e97a5ac64533d6517df20d178561fb84fa8f690f86932affec39dd0d9.jpg) +Dirty-Pixel +Fig. 6: Semantic segmentation results on over-exposure and low-light RAW data, compare with various ISP methods [6], [35], [71], [79] and Dirty-Pixel [18]. + +![](images/be7242f00efc4246400c470c9a2fec669fccc661513f726dd10c481fd309cdf2.jpg) + +![](images/d59502f2dfe0b97849a159e2a8fa294471a3852c3f2d4e4f5423d16306951ff7.jpg) +RAW-Adapter + +with Adam optimizer, the batch size is set to 4, training images are cropped into $512 \times 512$ and training iter are set to 80000. We make comparison with single demosacing, various ISP methods [6,34,35,71,79] and Dirty-Pixel [18], and RAW-Adapter with only input-level adapters (w/o M). + +Comparison results are shown in Table. 4, where we compare both efficiency (parameters↓, inference time↓) and performance (mIOU↑), inference time is calculated on a single Tesla V100 GPU. We first show the results on vanilla demosaiced RAW data, then we show the results on RAW data processed by various ISP methods [6,34,35,71,79], followed by results of Dirty-Pixel [18], and our method without and with model-level adapter M. For Table. 4, we can find that directly employing ISP methods sometimes fails to enhance the performance of downstream segmentation tasks, particularly in low-light conditions. Joint-training method Dirty-Pixel [18] could improve down-stream performance and sometimes even degrade segmentation performance, but increase more network parameters. Overall, our proposed RAW-Adapter could effectively enhance performance while adding a slight parameter count and improving inference efficiency. RAW-Adapter on backbones with fewer parameters can even outperform previous methods on backbones with more parameters (i.e. The RAW-Adapter, employing the MIT-B3 backbone, achieves a dark condition mIOU of 40.52, surpassing other methods utilizing the MIT-B5 backbone despite having fewer than $\sim$ 36M parameters.). + +The visualization of segmentation results are shown in Fig. 6, where row 1 and row 3 depict images rendered by different ISPs [6, 35, 71, 79] and images + +Table 4: Comparison with previous methods on ADE 20K RAW (normal/over-exp/dark). Bold denotes the best result while underline denotes second best result. + +
backboneparams(M) ↓inference time(s) ↓mIOU ↑ (normal)mIOU ↑ (over-exp)mIOU ↑ (dark)
DemosacingMIT-B582.010.10545.2642.0830.97
Karaimer et al. [35]0.52546.4243.7139.79
InvISP [71]0.20347.5343.554.87
LiteISP [79]0.26143.2242.015.52
DNF [34]0.186--25.82
SID [6]0.312--26.44
Dirty-Pixel [18]86.290.15947.7646.8040.02
MIT-B348.920.09847.1945.3338.93
MIT-B08.000.04935.4332.1027.19
RAW-Adapter (w/o M)MIT-B582.090.14848.3346.4840.91
MIT-B344.720.08647.0145.0039.34
MIT-B03.800.03235.0932.8227.29
RAW-AdapterMIT-B582.530.16748.3847.0641.82
MIT-B345.160.10247.5945.6640.52
MIT-B04.260.05335.6733.4427.78
+ +generated from Dirty-Pixel [18] and RAW-Adapter $(\mathbf{I}_5)$ , row 2 and row 4 denote our segmentation results appear superior compared to other methods. We were even pleased to discover that, although we do not add any human-vision-oriented loss function to constrain the model, RAW-Adapter still produces visually satisfactory image $\mathbf{I}_5$ (see Fig. 6). This may be attributed to our preservation of traditional ISP stages in input-level adapters. Due to page limitations, we refer to more experimental results, visualization results, and ablation analysis of our method in the supplementary part. + +# 5 Conclusion + +In this paper, we introduce RAW-Adapter, an effective solution for adapting pre-trained sRGB models to camera RAW data. With input-level adapters and model-level adapters working in tandem, RAW-Adapter effectively forwards RAW images from various lighting conditions to downstream vision tasks, our method has achieved state-of-the-art (SOTA) results across multiple datasets. + +For future research directions, we believe it is feasible to design a unified model capable of adapting to RAW visual tasks under different lighting conditions, without the need for retraining in each lighting scenario like RAW-Adapter. Additionally, designing multi-task decoders can enable the accommodation of various tasks, facilitating more effective integration of different visual tasks on RAW images with large-scale models. + +# Acknowledgements + +This research is partially supported by JST Moonshot R&D Grant Number JPMJPS2011, CREST Grant Number JPMJCR2015 and Basic Research Grant (Super AI) of Institute for AI and Beyond of the University of Tokyo. + +# References + +1. Affifi, M., Brown, M.S.: What else can fool deep learning? addressing color constancy errors on deep neural network performance. 2019 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 243-252 (2019), https://apisemantic scholar.org/CorpusID:207995696 +2. Brooks, T., Mildenhall, B., Xue, T., Chen, J., Sharlet, D., Barron, J.T.: Unprocessing images for learned raw denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11036-11045 (2019) +3. Buckler, M., Jayasuriya, S., Sampson, A.: Reconfiguring the imaging pipeline for computer vision. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 975-984 (2017) +4. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020) +5. Chatterjee, P., Joshi, N., Kang, S.B., Matsushita, Y.: Noise suppression in low-light images through joint denoising and demosaicing. In: CVPR 2011. pp. 321-328. IEEE (2011) +6. Chen, C., Chen, Q., Xu, J., Koltun, V.: Learning to see in the dark. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3291-3300 (2018) +7. Chen, K., Wang, J., et al.: Mmdetection: Open mmlab detection toolbox and benchmark (2019) +8. Chen, Z., Duan, Y., Wang, W., He, J., Lu, T., Dai, J., Qiao, Y.: Vision transformer adapter for dense predictions. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=plKu2GByCNW +9. Cheng, B., Schwing, A.G., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. In: NeurIPS (2021) +0. Conde, M.V., McDonagh, S., Maggioni, M., Leonardis, A., Pérez-Pellitero, E.: Model-based image signal processors via learnable dictionaries. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 481-489 (2022) +1. Conde, M.V., Vazquez-Corral, J., Brown, M.S., Timofte, R.: Nilut: Conditional neural implicit 3d lookup tables for image enhancement. In: Proceedings of the AAAI Conference on Artificial Intelligence (2024) +2. Contributors, M.: MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation (2020) +3. Cui, Z., Li, K., Gu, L., Su, S., Gao, P., Jiang, Z., Qiao, Y., Harada, T.: You only need 90k parameters to adapt light: a light weight transformer for image enhancement and exposure correction. In: BMVC. p. 238 (2022) +4. Cui, Z., Qi, G.J., Gu, L., You, S., Zhang, Z., Harada, T.: Multitask aet with orthogonal tangent regularity for dark object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2553-2562 (October 2021) + +15. Cui, Z., Zhu, Y., Gu, L., Qi, G.J., Li, X., Zhang, R., Zhang, Z., Harada, T.: Exploring resolution and degradation clues as self-supervised signal for low quality object detection. In: European Conference on Computer Vision. pp. 473-491. Springer (2022) +16. Delbracio, M., Kelly, D., Brown, M.S., Milanfar, P.: Mobile computational photography: A tour. Annual Review of Vision Science 7(1), 571-604 (2021). https://doi.org/10.1146/annurev-vision-093019-115521, https://doi.org/10.1146/annurev-vision-093019-115521, pMID: 34524880 +17. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A Large-Scale Hierarchical Image Database. In: CVPR09 (2009) +18. Diamond, S., Sitzmann, V., Julca-Aguilar, F., Boyd, S., Wetzstein, G., Heide, F.: Dirty pixels: Towards end-to-end image processing and perception. ACM Transactions on Graphics (SIGGRAPH) (2021) +19. Dong, X., Yokoya, N.: Understanding dark scenes by contrasting multi-modal observations. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 840-850 (January 2024) +20. Finlayson, G.D., Trezzi, E.: Shades of gray and colour constancy. In: Color and Imaging Conference. vol. 2004, pp. 37-41. Society for Imaging Science and Technology (2004) +21. Goyal, B., Lalonde, J.F., Li, Y., Gupta, M.: Robust scene inference under noise-blur dual corruptions. 2022 IEEE International Conference on Computational Photography (ICCP) pp. 1-12 (2022), https://api_semanticscholar.org/CorpusID:251040807 +22. Guo, Y., Luo, F., Wu, X.: Learning degradation-independent representations for camera isp pipelines. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 25774-25783 (June 2024) +23. Hansen, N., Ostermeier, A.: Adapting arbitrary normal mutation distributions in evolution strategies: the covariance matrix adaptation. In: Proceedings of IEEE International Conference on Evolutionary Computation. pp. 312-317 (1996). https://doi.org/10.1109/ICEC.1996.542381 +24. Hasinoff, S.W., Sharlet, D., Geiss, R., Adams, A., Barron, J.T., Kainz, F., Chen, J., Levoy, M.: Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM Transactions on Graphics (Proc. SIGGRAPH Asia) 35(6) (2016) +25. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +26. Heide, F., Steinberger, M., Tsai, Y.T., Rouf, M., Pajak, D., Reddy, D., Gallo, O., Liu, J., Heidrich, W., Egiazarian, K., et al.: Flexisp: A flexible camera image processing framework. ACM Transactions on Graphics (ToG) 33(6), 1-13 (2014) +27. Hong, Y., Wei, K., Chen, L., Fu, Y.: Crafting object detection in very low light. In: BMVC (2021) +28. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019) +29. Hu, Y., He, H., Xu, C., Wang, B., Lin, S.: Exposure: A white-box photo post-processing framework. ACM Transactions on Graphics (TOG) 37(2), 26 (2018) +30. Ignatov, A., Van Gool, L., Timofte, R.: Replacing mobile camera isp with a single deep learning model. arXiv preprint arXiv:2002.05509 (2020) + +31. Iscen, A., Zhang, J., Lazebnik, S., Schmid, C.: Memory-efficient incremental learning through feature adaptation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16. pp. 699-715. Springer (2020) +32. Jensen, H.W., Durand, F., Dorsey, J., Stark, M.M., Shirley, P., Premoze, S.: A physically-based night sky model. In: Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques. p. 399-408. SIGGRAPH '01, Association for Computing Machinery, New York, NY, USA (2001). https://doi.org/10.1145/383259.383306, https://doi.org/10.1145/383259.383306 +33. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022) +34. Jin, X., Han, L., Li, Z., Chai, Z., Guo, C., Li, C.: Dnf: Decouple and feedback network for seeing in the dark. In: ICCV (2023) +35. Karaimer, H.C., Brown, M.S.: A software platform for manipulating the camera imaging pipeline. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) Computer Vision - ECCV 2016. pp. 429-444. Springer International Publishing, Cham (2016) +36. Ke, L., Ye, M., Danelljan, M., Liu, Y., Tai, Y.W., Tang, C.K., Yu, F.: Segment anything in high quality. In: NeurIPS (2023) +37. Kim, W., Kim, G., Lee, J., Lee, S., Baek, S.H., Cho, S.: Paramisp: Learned forward and inverse isps using camera parameters. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 26067-26076 (2024) +38. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4015-4026 (October 2023) +39. Koskinen, S., Yang, D., Kämäräinen, J.K.: Reverse imaging pipeline for raw rgb image augmentation. In: 2019 IEEE International Conference on Image Processing (ICIP). pp. 2896-2900 (2019). https://doi.org/10.1109/ICIP.2019.8804406 +40. Li, Z., Lu, M., Zhang, X., Feng, X., Asif, M.S., Ma, Z.: Efficient visual computing with camera raw snapshots. TPAMI pp. 1-18 (2024). https://doi.org/10.1109/TPAMI.2024.3359326 +41. Li, Z., Jiang, H., Cao, M., Zheng, Y.: Polarized color image denoising. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9873-9882 (2023). https://doi.org/10.1109/CVPR52729.2023.00952 +42. Liang, Z., Cai, J., Cao, Z., Zhang, L.: Cameranet: A two-stage framework for effective camera isp learning. IEEE Transactions on Image Processing 30, 2248-2262 (2021). https://doi.org/10.1109/TIP.2021.3051486 +43. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017) +44. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Fleet, D., Pajdla, T., Schiele, B., Tuytelaars, T. (eds.) Computer Vision - ECCV 2014. pp. 740-755. Springer International Publishing, Cham (2014) +45. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021) +46. Mildenhall, B., Hedman, P., Martin-Brualla, R., Srinivasan, P.P., Barron, J.T.: Nerf in the dark: High dynamic range view synthesis from noisy raw images. In: + +Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16190-16199 (2022) +47. Monakhova, K., Richter, S.R., Waller, L., Koltun, V.: Dancing under the stars: Video denoising in starlight. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16241-16251 (June 2022) +48. Mosleh, A., Sharma, A., Onzon, E., Mannan, F., Robidoux, N., Heide, F.: Hardware-in-the-loop end-to-end optimization of camera image processing pipelines. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020) +49. Nam, S., Kim, S.J.: Modelling the scene dependent imaging in cameras with a deep neural network. 2017 IEEE International Conference on Computer Vision (ICCV) pp. 1726-1734 (2017), https://api.sementicscholar.org/CorpusID:23920472 +50. Nguyen, R.M.H., Brown, M.S.: Raw image reconstruction using a self-contained srgb-jpeg image with only 64 kb overhead. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1655-1663 (2016). https://doi.org/10.1109/CVPR.2016.183 +51. Nishimura, J., Gerasimow, T., Sushma, R., Sutic, A., Wu, C.T., Michael, G.: Automatic isp image quality tuning using nonlinear optimization. In: 2018 25th IEEE International Conference on Image Processing (ICIP). pp. 2471-2475. IEEE (2018) +52. Omid-Zohoor, A., Ta, D., Murmann, B.: Pascalraw: raw image database for object detection (2014) +53. Onzon, E., Mannan, F., Heide, F.: Neural auto-exposure for high-dynamic range object detection. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7706-7716 (2021). https://doi.org/10.1109/CVPR46437.2021.00762 +54. Potlapalli, V., Zamir, S.W., Khan, S., Khan, F.: Prompt: Prompting for all-in-one image restoration. In: Thirty-seventh Conference on Neural Information Processing Systems (2023) +55. Qin, H., Han, L., Wang, J., Zhang, C., Li, Y., Li, B., Hu, W.: Attention-aware learning for hyperparameter prediction in image processing pipelines. In: European Conference on Computer Vision. pp. 271-287. Springer (2022) +56. Ramanath, R., Snyder, W.E., Yoo, Y., Drew, M.S.: Color image processing pipeline. IEEE Signal processing magazine 22(1), 34-43 (2005) +57. Rebuffi, S.A., Bilen, H., Vedaldi, A.: Learning multiple visual domains with residual adapters. Advances in neural information processing systems 30 (2017) +58. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Navab, N., Hornegger, J., Wells, W.M., Frangi, A.F. (eds.) Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015. pp. 234-241. Springer International Publishing, Cham (2015) +59. Sayed, M., Brostow, G.: Improved handling of motion blur in online object detection. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1706-1716 (2021). https://doi.org/10.1109/CVPR46437.2021.00175 +60. Schwartz, E., Giryes, R., Bronstein, A.M.: Deepisp: Toward learning an end-to-end image processing pipeline. IEEE Transactions on Image Processing 28(2), 912-923 (2018) +61. Sitzmann, V., Martel, J.N., Bergman, A.W., Lindell, D.B., Wetzstein, G.: Implicit neural representations with periodic activation functions. In: Proc. NeurIPS (2020) +62. Souza, M., Heidrich, W.: MetaISP - Exploiting Global Scene Structure for Accurate Multi-Device Color Rendition. In: Guthe, M., Grosch, T. (eds.) Vision, Modeling, + +and Visualization. The Eurographics Association (2023). https://doi.org/10.2312/vmv.20231236 +63. Stickland, A.C., Murray, I.: Bert and pals: Projected attention layers for efficient adaptation in multi-task learning. In: International Conference on Machine Learning. pp. 5986-5995. PMLR (2019) +64. Sun, P., Zhang, R., Jiang, Y., Kong, T., Xu, C., Zhan, W., Tomizuka, M., Li, L., Yuan, Z., Wang, C., Luo, P.: SparseR-CNN: End-to-end object detection with learnable proposals. arXiv preprint arXiv:2011.12450 (2021) +65. Sung, Y.L., Cho, J., Bansal, M.: Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5227-5237 (2022) +66. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) +67. Wang, Y., Yu, Y., Yang, W., Guo, L., Chau, L.P., Kot, A.C., Wen, B.: Raw image reconstruction with learned compact metadata. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18206-18215 (June 2023) +68. Wei, K., Fu, Y., Yang, J., Huang, H.: A physics-based noise formation model for extreme low-light raw denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020) +69. Wu, C.T., Isikdogan, L.F., Rao, S., Nayak, B., Gerasimow, T., Sutic, A., Ainkedem, L., Michael, G.: Visionisp: Repurposing the image signal processor for computer vision applications. In: 2019 IEEE International Conference on Image Processing (ICIP). IEEE (Sep 2019). https://doi.org/10.1109/icip.2019.8803607, http://dx.doi.org/10.1109/ICIP.2019.8803607 +70. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. arXiv preprint arXiv:2105.15203 (2021) +71. Xing, Y., Qian, Z., Chen, Q.: Invertible image signal processing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6287-6296 (2021) +72. Xu, R., Chen, C., Peng, J., Li, C., Huang, Y., Song, F., Yan, Y., Xiong, Z.: Toward raw object detection: A new benchmark and a new model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13384-13393 (2023) +73. Yoshimura, M., Otsuka, J., Irie, A., Ohashi, T.: Dynamically controlled image signal processor for image recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 12866-12876 (October 2023) +74. Yoshimura, M., Otsuka, J., Irie, A., Ohashi, T.: Rawgment: Noise-accounted raw augmentation enables recognition in a wide variety of environments. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14007-14017 (June 2023) +75. Yu, K., Li, Z., Peng, Y., Loy, C.C., Gu, J.: Reconfigisp: Reconfigurable camera image processing pipeline. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4248-4257 (2021) +76. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Cycleisp: Real image restoration via improved data synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2696-2705 (2020) + +77. Zhang, R., Zhang, W., Fang, R., Gao, P., Li, K., Dai, J., Qiao, Y., Li, H.: Tip-adapter: Training-free adaption of clip for few-shot classification. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 493-510. Springer Nature Switzerland, Cham (2022) +78. Zhang, Y., Zhang, Y., Zhang, Z., Zhang, M., Tian, R., Ding, M.: Isp-teacher: Image signal process with disentanglement regularization for unsupervised domain adaptive dark object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 7387-7395 (2024) +79. Zhang, Z., Wang, H., Liu, M., Wang, R., Zuo, W., Zhang, J.: Learning raw-to-srgb mappings with inaccurately aligned supervision. In: ICCV (2021) +80. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing through ade20k dataset. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 633-641 (2017) +81. Zhou, W., Gao, S., Zhang, L., Lou, X.: Histogram of oriented gradients feature extraction from raw bayer pattern images. IEEE Transactions on Circuits and Systems II: Express Briefs 67(5), 946-950 (2020). https://doi.org/10.1109/TCSII.2020.2980557 \ No newline at end of file diff --git a/2024/RAW-Adapter_ Adapting Pretrained Visual Model to Camera RAW Images/images.zip b/2024/RAW-Adapter_ Adapting Pretrained Visual Model to Camera RAW Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8d9a7b8e6af8408b45fef18d2f3b14c4f8cf70af --- /dev/null +++ b/2024/RAW-Adapter_ Adapting Pretrained Visual Model to Camera RAW Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae289a124d1abf2577c8366c811591e0dd66c94ccb2e4c4da28b7a96ff0f95c4 +size 698646 diff --git a/2024/RAW-Adapter_ Adapting Pretrained Visual Model to Camera RAW Images/layout.json b/2024/RAW-Adapter_ Adapting Pretrained Visual Model to Camera RAW Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..072cb7a14b07d63c5373502a273c812873ffc18f --- /dev/null +++ b/2024/RAW-Adapter_ Adapting Pretrained Visual Model to Camera RAW Images/layout.json @@ -0,0 +1,11298 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 152, + 111, + 462, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 111, + 462, + 148 + ], + "spans": [ + { + "bbox": [ + 152, + 111, + 462, + 148 + ], + "type": "text", + "content": "RAW-Adapter: Adapting Pre-trained Visual Model to Camera RAW Images" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 221, + 167, + 394, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 167, + 394, + 182 + ], + "spans": [ + { + "bbox": [ + 221, + 167, + 394, + 182 + ], + "type": "text", + "content": "Ziteng Cui" + }, + { + "bbox": [ + 221, + 167, + 394, + 182 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 221, + 167, + 394, + 182 + ], + "type": "text", + "content": " and Tatsuya Harada" + }, + { + "bbox": [ + 221, + 167, + 394, + 182 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 252, + 190, + 361, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 190, + 361, + 201 + ], + "spans": [ + { + "bbox": [ + 252, + 190, + 361, + 201 + ], + "type": "text", + "content": "1 The University of Tokyo" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 276, + 201, + 337, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 201, + 337, + 211 + ], + "spans": [ + { + "bbox": [ + 276, + 201, + 337, + 211 + ], + "type": "text", + "content": "2 RIKEN AIP" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 231, + 213, + 383, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 213, + 383, + 224 + ], + "spans": [ + { + "bbox": [ + 231, + 213, + 383, + 224 + ], + "type": "text", + "content": "{cui, harada}@mi.t.u-tokyo.ac.jp" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 254, + 455, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 254, + 455, + 464 + ], + "spans": [ + { + "bbox": [ + 159, + 254, + 455, + 464 + ], + "type": "text", + "content": "Abstract. sRGB images are now the predominant choice for pre-training visual models in computer vision research, owing to their ease of acquisition and efficient storage. Meanwhile, the advantage of RAW images lies in their rich physical information under variable real-world challenging lighting conditions. For computer vision tasks directly based on camera RAW data, most existing studies adopt methods of integrating image signal processor (ISP) with backend networks, yet often overlook the interaction capabilities between the ISP stages and subsequent networks. Drawing inspiration from ongoing adapter research in NLP and CV areas, we introduce RAW-Adapter, a novel approach aimed at adapting sRGB pre-trained models to camera RAW data. RAW-Adapter comprises input-level adapters that employ learnable ISP stages to adjust RAW inputs, as well as model-level adapters to build connections between ISP stages and subsequent high-level networks. Additionally, RAW-Adapter is a general framework that could be used in various computer vision frameworks. Abundant experiments under different lighting conditions have shown our algorithm's state-of-the-art (SOTA) performance, demonstrating its effectiveness and efficiency across a range of real-world and synthetic datasets. Code is available at this url" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 484, + 230, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 484, + 230, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 484, + 230, + 497 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "type": "text", + "content": "In recent years, there has been a growing interest in revisiting vision tasks using unprocessed camera RAW images. How to leverage information-rich RAW image in computer vision tasks has become a current research hotspot in various sub-areas (i.e. denoising [27], view synthesis [46], object detection [27, 48]). Compared to commonly used sRGB image, RAW image directly acquired by the camera sensor, encompasses abundant information unaffected or compressed by image signal processor (ISP), also offers physically meaningful information like noise distributions [47, 68], owing to its linear correlation between image intensity and the radiant energy received by a camera. The acquisition of RAW data, facilitating enhanced detail capture and a higher dynamic range, imparts a unique advantage in addressing visual tasks under variable lighting conditions in the real world. For instance, sunlight irradiance can reach as high as " + }, + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "type": "inline_equation", + "content": "1.3 \\times 10^{3}W / m^{2}" + }, + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "type": "text", + "content": " while bright planet irradiance can be as low as " + }, + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "type": "inline_equation", + "content": "2.0 \\times 10^{-6}W / m^{2}" + }, + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "type": "text", + "content": " [32]." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 482, + 255 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 482, + 255 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 482, + 255 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 482, + 255 + ], + "type": "image", + "image_path": "d6f0e038f96a0dc83c9d3e0b6589ecb2e58610a9cbea038010c1e920d4bd3760.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 262, + 482, + 298 + ], + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 298 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 298 + ], + "type": "text", + "content": "Fig. 1: (a). An overview of basic image signal processor (ISP) pipeline. (b). ISP and current visual model have different objectives. (c) Previous methods optimize ISP with down-stream visual model. (d) Our proposed RAW-Adapter." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 325, + 482, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 482, + 411 + ], + "type": "text", + "content": "Meanwhile, sRGB has emerged as the primary choice for pre-training visual models in today's computer vision field, due to its scalability and ease of storage. Typically, sRGB images are derived from camera RAW data through the ISP pipeline. As shown in Fig. 1(a), the entire ISP pipeline consists of multiple modules to convert RAW images into vision-oriented sRGB images, each of these modules serves its own distinct purpose, with the majority being ill-posed and heavily dependent on prior information [5,16,26,35,42,56]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 411, + 482, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 497 + ], + "type": "text", + "content": "When adopt RAW data for computer vision tasks, the purpose of a manually designed ISP is to produce images that offer a superior visual experience [18,69], rather than optimizing for downstream visual tasks such as object detection or semantic segmentation (see Fig. ①b)). Additionally, most companies' ISPs are black boxes, making it difficult to obtain information about the specific steps inside. Consequently, utilizing human vision-oriented ISP in certain conditions is sometimes even less satisfactory than directly using RAW [22,27,40,81]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "To better take advantage of camera RAW data for various computer vision tasks. Researchers began to optimize the image signal processor (ISP) jointly with downstream networks. Since most ISP stages are non-differentiable and cannot be jointly backpropagation, there are two main lines of approaches to connect ISP stages with downstream networks (see Fig. 1(c)): ①. First-kind approaches maintain the modular design of the traditional ISP, involving the design of differentiable ISP modules through optimization algorithms 48,69,75, such as Hardware-in-the-loop 48 adopt covariance matrix adaptation evolution strategy (CMA-ES) 23. ②. Second-kind approaches replaced the ISP part entirely with a neural network 18,55,72,73, such as Dirty-Pixel 18 adopt a stack of residual UNets 58 as pre-encoder, however, the additional neural network introduces a significant computational burden, especially when dealing with high-resolution inputs. Beyond these, the above-mentioned methods still treat ISP and the backend neural network as two independent modules, lacking interac" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "Cui, Harada." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 306, + 202 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 306, + 202 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 306, + 202 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 306, + 202 + ], + "type": "image", + "image_path": "4d3460696f9b178ee2425f48e832603f1c4e3961befcd0d770a0d98282388dcd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 211, + 482, + 258 + ], + "lines": [ + { + "bbox": [ + 130, + 211, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 211, + 482, + 258 + ], + "type": "text", + "content": "Fig. 2: Performance of RAW-based visual tasks with and without sRGB pre-trained weights. We analyze two methods: Dirty-Pixel [18] and RAW-Adapter. Blue line represents trained with MS COCO [44] pre-train weights, the purple line indicates ImageNet [17] pre-train weights, and the yellow line signifies training from scratch." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 482, + 202 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 482, + 202 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 482, + 202 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 482, + 202 + ], + "type": "image", + "image_path": "d5317e4de476aa991ab27610f28f1e9956b0cfc91f38b155345715b96886099b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 284, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 380 + ], + "type": "text", + "content": "tion capabilities between two separate models. Especially, current visual models predominantly pre-trained on large amounts of sRGB images, there exists a significant gap between pre-trained models and camera RAW images. As it shown in Fig. 2 training from scratch with RAW data can significantly impair the performance of RAW-based vision tasks. How to leveraging sRGB pre-trained weights proves to be highly beneficial for RAW image visual tasks. Hence, here we ask: Is there a more effective way to adapt the information-rich RAW data into the knowledge-rich sRGB pre-trained models?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 381, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 381, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 482, + 609 + ], + "type": "text", + "content": "Our solution, RAW-Adapter, differs from previous approaches that employed complex ISP or deeper neural networks at the input level. Instead, we prioritize simplifying the input stages and enhancing connectivity between ISP and subsequent networks at the model level. Inspired by recent advancements in prompt learning and adapter tuning [8,28,33,54], we developed two novel adapter approaches aimed at enhancing the integration between RAW input and sRGB pretrained models. Leveraging priors from ISP stages, our method includes input-level adapters and model-level adapters, as illustrated in Fig. [1(d)]. Input-level adapters are designed to adapt the input from RAW data to the backend network input, we follow traditional ISP design and predict ISP stages' key parameters to accommodate RAW image into the specific down-stream vision tasks, in this part, we employ strategies involving Query Adaptive Learning (QAL) and Implicit Neural Representation (INR) to ensure the differentiability of the core ISP processes while also upholding a lightweight design. Model-level adapters leverage prior knowledge from the input-level adapters by extracting intermediate stages as features, these features are then embedded into the downstream network, thus the network incorporates prior knowledge from the ISP stages, then jointly contributes to final machine-vision perception tasks. Our contributions could be summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 137, + 629, + 482, + 666 + ], + "type": "text", + "content": "- We introduce RAW-Adapter, a novel framework aimed at improving the integration of sRGB pre-trained models to camera RAW images. Through the implementation of two types of adapters, we strive to narrow the dispar" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "type": "text", + "content": "RAW-Adapter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 147, + 116, + 480, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 480, + 138 + ], + "type": "text", + "content": "ity between RAW images and sRGB models, tackling discrepancies at both input and model levels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 479, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 479, + 198 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 479, + 198 + ], + "type": "text", + "content": "- By analyzing camera ISP steps, we've crafted input-level adapters that integrate query adaptive learning (QAL) and implicit neural representations (INR) to optimize ISP key parameters. Additionally, harnessing prior input stage information informs model-level adapters, enriching model understanding and improving downstream task performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 201, + 479, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 201, + 479, + 258 + ], + "spans": [ + { + "bbox": [ + 137, + 201, + 479, + 258 + ], + "type": "text", + "content": "- We conducted detection and segmentation experiments involving different lighting scenarios, including normal-light, dark, and over-exposure scenes. Through comparisons with current mainstream ISPs and joint-training methods, sufficient experiments demonstrate that our algorithm achieved state-of-the-art (SOTA) performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 279, + 242, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 279, + 242, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 279, + 242, + 292 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 305, + 278, + 318 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 305, + 278, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 305, + 278, + 318 + ], + "type": "text", + "content": "2.1 Image Signal Processor" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 326, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 482, + 555 + ], + "type": "text", + "content": "Typically, a camera's Image Signal Processor (ISP) pipeline is required to reconstruct a high-quality sRGB image from camera RAW data, traditional ISP pipeline is formulated as a series of manually crafted modules executed sequentially [16,35,51,56], including some representative steps such as demosaicing, white balance, noise removal, tone mapping, color space transform and so on. There also exist alternative designs of ISP process such as Heide et al. [26] designed FlexISP which combine numerous ISP blocks to a joint optimization block, and Hasinoff et al. [24] modified some of the traditional ISP steps for burst photography under extreme low-light conditions. When it comes to deep learning era, Chen et al. [6] proposed SID to use a UNet [58] replace the traditional ISP steps, which translate low-light RAW data to normal-light sRGB images, and Hu et al. [29] proposed a white-box solution with eight differentiable filters. After that, many efforts began to substitute the traditional ISP process with a neural network, such as [6,10,30,34,37,42,60,62,79], however deep network models are constrained by the training dataset, leading to shortcomings in generalization performance. Meanwhile another line of work focuses on translate sRGB images back to RAW data [2,39,49,50,67,71,76]. Apart from the conversion between RAW and RGB, there is ongoing research specifically dedicated to harnessing RAW images for downstream computer vision tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 572, + 354, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 572, + 354, + 585 + ], + "spans": [ + { + "bbox": [ + 132, + 572, + 354, + 585 + ], + "type": "text", + "content": "2.2 Computer Vision based on RAW data" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "In order to effectively leverage information in RAW images for downstream visual tasks and save the time required for ISP, some early methods proposed to directly perform computer vision tasks on RAW images [3,81], which lacks consideration for the camera noise introduced during the conversion from photons to RAW images, especially in low-light conditions [41,47,68]. Meanwhile training from scratch on RAW data would abandon the current large-scale pre-trained" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "text", + "content": "Cui, Harada." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 164 + ], + "type": "text", + "content": "visual model on sRGB data (see Fig. 2), especially the quantity of RAW image datasets (i.e. 4259 images in PASCAL RAW 52 dataset and 2230 images in LOD 27 dataset) is incomparable to current RGB datasets (i.e. over 1M images in ImageNet 17 dataset and SAM 38 is trained with 11M images)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 482, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 482, + 345 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 482, + 345 + ], + "type": "text", + "content": "Most subsequent research focuses on finding ways to integrate ISP and backend computer vision models [18,22,48,55,69,72,73,75]. For example Wu et al. [69] first propose VisionISP and emphasized the difference between human vision and machine vision, they introduced several trainable modules in the ISP to enhance backend object detection performance. Since then several methods attempt to replace the existing non-differentiable ISP with a differentiable ISP network [48,75] or directly using encoder networks to replace the ISP process [18,55,73], however training two consecutive networks simultaneously can result in a significant consumption of computational resources, previous work also rarely focused on how to fine-tune the current mainstream sRGB visual models more efficiently on RAW data. Additionally, research like Rawgment [74], focusing on achieving realistic data augmentation directly on RAW images, some studies address the decrease in computer vision performance caused by specific steps within the incamera process or ISP, such as white balance error [1], auto-exposure error [53], camera motion blur [59] and undesirable camera noise [15,21]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 361, + 309, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 361, + 309, + 373 + ], + "spans": [ + { + "bbox": [ + 131, + 361, + 309, + 373 + ], + "type": "text", + "content": "2.3 Adapters in Computer Vision" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 380, + 482, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 501 + ], + "type": "text", + "content": "Adapters have become prevalent NLP area, which introducing new modules in large language model (LLM) such as [28,63], these modules enable task-specific fine-tuning, allowing pre-trained models to swiftly adapt to downstream NLP tasks. In the computer vision area, adapters have been adopted in various areas such as incremental learning [31] and domain adaptation [57]. Recently, a series of adapters have been used to investigate how to better utilize pre-trained visual models [8,33,36] or pre-trained vision-language models [65,77]. These methods focus on utilizing prior knowledge to make pre-trained models quickly adapt to downstream tasks. Here, we introduce the RAW-Adapter to further improve the alignment between RAW inputs and sRGB pre-trained visual models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 518, + 240, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 518, + 240, + 533 + ], + "spans": [ + { + "bbox": [ + 131, + 518, + 240, + 533 + ], + "type": "text", + "content": "3 RAW-Adapter" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 544, + 480, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 480, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 480, + 605 + ], + "type": "text", + "content": "In this section, we introduce RAW-Adapter. Due to page limitation, an overview of the conventional image signal processor (ISP) in supplementary's Sec. A. For specific adapter designs, we selectively omit certain steps and focus on some key steps within the camera ISP. In Sec. 3.1 we introduce RAW-Adapter's input-level adapters, followed by an explanation of its model-level adapters in Sec. 3.2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 621, + 265, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 621, + 265, + 634 + ], + "spans": [ + { + "bbox": [ + 131, + 621, + 265, + 634 + ], + "type": "text", + "content": "3.1 Input-level Adapters" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 640, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 640, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 640, + 480, + 665 + ], + "type": "text", + "content": "Fig. 3(a) provides an overview of RAW-Adapter. The solid lines in Fig. 3(a) left represent input-level adapters. Input-level adapters are designed to convert" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "type": "text", + "content": "RAW-Adapter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 360, + 314 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 360, + 314 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 360, + 314 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 360, + 314 + ], + "type": "image", + "image_path": "faa998769cf87654cfecc86abd9f3911d3467d37db1a38152e4e333ddb53cca3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 325, + 482, + 370 + ], + "lines": [ + { + "bbox": [ + 130, + 325, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 482, + 370 + ], + "type": "text", + "content": "Fig. 3: (a). Structure of RAW-Adapter. Solid line in left denotes input-level adapter's workflow and dotted line denotes model-level adapter's workflow, stage " + }, + { + "bbox": [ + 130, + 325, + 482, + 370 + ], + "type": "inline_equation", + "content": "1\\sim 4" + }, + { + "bbox": [ + 130, + 325, + 482, + 370 + ], + "type": "text", + "content": " means different stage of visual model backbone. (b). Detailed structure of kernel & matrix predictors " + }, + { + "bbox": [ + 130, + 325, + 482, + 370 + ], + "type": "inline_equation", + "content": "\\mathbb{P}_{\\mathbb{K}}" + }, + { + "bbox": [ + 130, + 325, + 482, + 370 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 325, + 482, + 370 + ], + "type": "inline_equation", + "content": "\\mathbb{P}_{\\mathbb{M}}" + }, + { + "bbox": [ + 130, + 325, + 482, + 370 + ], + "type": "text", + "content": ". (c). Detailed structure of model-level adapter " + }, + { + "bbox": [ + 130, + 325, + 482, + 370 + ], + "type": "inline_equation", + "content": "\\mathbb{M}" + }, + { + "bbox": [ + 130, + 325, + 482, + 370 + ], + "type": "text", + "content": "'s merge block." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 361, + 115, + 482, + 317 + ], + "blocks": [ + { + "bbox": [ + 361, + 115, + 482, + 317 + ], + "lines": [ + { + "bbox": [ + 361, + 115, + 482, + 317 + ], + "spans": [ + { + "bbox": [ + 361, + 115, + 482, + 317 + ], + "type": "image", + "image_path": "6e4744b439528819ac969d3a50ece88406b7bd2cce496183f9e7b297ad424fee.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 395, + 482, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 395, + 482, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 395, + 482, + 431 + ], + "type": "text", + "content": "the RAW image " + }, + { + "bbox": [ + 130, + 395, + 482, + 431 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_1" + }, + { + "bbox": [ + 130, + 395, + 482, + 431 + ], + "type": "text", + "content": " into the machine-vision oriented image " + }, + { + "bbox": [ + 130, + 395, + 482, + 431 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_5" + }, + { + "bbox": [ + 130, + 395, + 482, + 431 + ], + "type": "text", + "content": ". This process involves digital gain & denoise, demosacing, white balance adjustment, camera color matrix, and color manipulation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 432, + 483, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 432, + 483, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 483, + 517 + ], + "type": "text", + "content": "We maintain the ISP design while simultaneously using query adaptive learning (QAL) to estimate key parameters in ISP stages. The QAL strategy is motivated by previous transformer models [4,9,13] and detailed structure is shown in Fig. 3(b), input image " + }, + { + "bbox": [ + 130, + 432, + 483, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{i\\in (1,2)}" + }, + { + "bbox": [ + 130, + 432, + 483, + 517 + ], + "type": "text", + "content": " would pass by 2 down-scale convolution blocks to generate feature, then feature pass by 2 linear layers to generate attention block's key " + }, + { + "bbox": [ + 130, + 432, + 483, + 517 + ], + "type": "inline_equation", + "content": "\\mathbb{k}" + }, + { + "bbox": [ + 130, + 432, + 483, + 517 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 130, + 432, + 483, + 517 + ], + "type": "inline_equation", + "content": "\\mathbb{V}" + }, + { + "bbox": [ + 130, + 432, + 483, + 517 + ], + "type": "text", + "content": ", while query " + }, + { + "bbox": [ + 130, + 432, + 483, + 517 + ], + "type": "inline_equation", + "content": "\\mathbb{q}" + }, + { + "bbox": [ + 130, + 432, + 483, + 517 + ], + "type": "text", + "content": " is a set of learnable dynamic parameters, the ISP parameters would be predicted by self-attention calculation [66]:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 209, + 526, + 481, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 526, + 481, + 554 + ], + "spans": [ + { + "bbox": [ + 209, + 526, + 481, + 554 + ], + "type": "interline_equation", + "content": "p a r a m e t e r s = F F N \\left(\\operatorname {s o f t m a x} \\left(\\frac {\\mathbb {Q} \\cdot \\mathbb {k} ^ {T}}{\\sqrt {d _ {k}}}\\right) \\cdot \\mathbb {V}\\right), \\tag {1}", + "image_path": "a82ed6aaa7e707f2686db719d7575b9f3001ed70baab0f3b1ab127690233bfdf.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 559, + 482, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 559, + 482, + 595 + ], + "spans": [ + { + "bbox": [ + 130, + 559, + 482, + 595 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 559, + 482, + 595 + ], + "type": "inline_equation", + "content": "FFN" + }, + { + "bbox": [ + 130, + 559, + 482, + 595 + ], + "type": "text", + "content": " denotes the feed-forward network, includes 2 linear layers and 1 activation layer, the predicted parameters would keep the same length as query q. We defined 2 QAL blocks " + }, + { + "bbox": [ + 130, + 559, + 482, + 595 + ], + "type": "inline_equation", + "content": "\\mathbb{P}_{\\mathbb{K}}" + }, + { + "bbox": [ + 130, + 559, + 482, + 595 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 559, + 482, + 595 + ], + "type": "inline_equation", + "content": "\\mathbb{P}_{\\mathbb{M}}" + }, + { + "bbox": [ + 130, + 559, + 482, + 595 + ], + "type": "text", + "content": " to predict different part parameters." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 596, + 482, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 596, + 482, + 622 + ], + "spans": [ + { + "bbox": [ + 130, + 596, + 482, + 622 + ], + "type": "text", + "content": "The input RAW image " + }, + { + "bbox": [ + 130, + 596, + 482, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_1" + }, + { + "bbox": [ + 130, + 596, + 482, + 622 + ], + "type": "text", + "content": " would first go through the pre-process operations and a demosaing stage [16,35], followed by subsequent ISP processes:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Gain & Denoise: Denoising algorithms always take various factors into account, such as input photon number, ISO gain level, exposure settings. Here we" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "Cui, Harada." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 115, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 115, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 131, + 115, + 482, + 200 + ], + "type": "text", + "content": "first use QAL block " + }, + { + "bbox": [ + 131, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\mathbb{P}_{\\mathbb{K}}" + }, + { + "bbox": [ + 131, + 115, + 482, + 200 + ], + "type": "text", + "content": " to predict a gain ratio " + }, + { + "bbox": [ + 131, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 131, + 115, + 482, + 200 + ], + "type": "text", + "content": " [2,6] to adapt " + }, + { + "bbox": [ + 131, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_1" + }, + { + "bbox": [ + 131, + 115, + 482, + 200 + ], + "type": "text", + "content": " in different lighting scenarios, followed by an adaptive anisotropic Gaussian kernel to suppress noise under various noise conditions, " + }, + { + "bbox": [ + 131, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\mathbb{P}_{\\mathbb{K}}" + }, + { + "bbox": [ + 131, + 115, + 482, + 200 + ], + "type": "text", + "content": " will predict the appropriate Gaussian kernel " + }, + { + "bbox": [ + 131, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 115, + 482, + 200 + ], + "type": "text", + "content": " for denoising to improve downstream visual tasks' performance, the predicted key parameters are the Gaussian kernel's major axis " + }, + { + "bbox": [ + 131, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "r_1" + }, + { + "bbox": [ + 131, + 115, + 482, + 200 + ], + "type": "text", + "content": ", minor axis " + }, + { + "bbox": [ + 131, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "r_2" + }, + { + "bbox": [ + 131, + 115, + 482, + 200 + ], + "type": "text", + "content": " (see Fig. 4 left), here we set the kernel angle " + }, + { + "bbox": [ + 131, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 131, + 115, + 482, + 200 + ], + "type": "text", + "content": " to 0 for simplification. Gaussian kernel " + }, + { + "bbox": [ + 131, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 115, + 482, + 200 + ], + "type": "text", + "content": " at pixel " + }, + { + "bbox": [ + 131, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 131, + 115, + 482, + 200 + ], + "type": "text", + "content": " would be:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 222, + 211, + 480, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 211, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 222, + 211, + 480, + 224 + ], + "type": "interline_equation", + "content": "k (x, y) = \\exp \\left(- \\left(b _ {0} x ^ {2} + 2 b _ {1} x y + b _ {2} y ^ {2}\\right)\\right), \\tag {2}", + "image_path": "76a8a8c47a32446d583e700c573192a775eef28d06c24736837208df23492c21.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 228, + 163, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 163, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 163, + 237 + ], + "type": "text", + "content": "where:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 140, + 242, + 480, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 242, + 480, + 268 + ], + "spans": [ + { + "bbox": [ + 140, + 242, + 480, + 268 + ], + "type": "interline_equation", + "content": "b _ {0} = \\frac {\\cos (\\theta) ^ {2}}{2 r _ {1} ^ {2}} + \\frac {\\sin (\\theta) ^ {2}}{2 r _ {2} ^ {2}}, b _ {1} = \\frac {\\sin (2 \\theta)}{4 r _ {1} ^ {2}} ((\\frac {r _ {1}}{r _ {2}}) ^ {2} - 1), b _ {2} = \\frac {\\sin (\\theta) ^ {2}}{2 r _ {1} ^ {2}} + \\frac {\\cos (\\theta) ^ {2}}{2 r _ {2} ^ {2}}. (3)", + "image_path": "c26870c1b205ef433adbaacf23707aa5f7c7d9eff538e1394544c04ce0312779.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 271, + 479, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 271, + 479, + 332 + ], + "spans": [ + { + "bbox": [ + 131, + 271, + 479, + 332 + ], + "type": "text", + "content": "After gain ratio " + }, + { + "bbox": [ + 131, + 271, + 479, + 332 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 131, + 271, + 479, + 332 + ], + "type": "text", + "content": " and kernel " + }, + { + "bbox": [ + 131, + 271, + 479, + 332 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 271, + 479, + 332 + ], + "type": "text", + "content": " process on image " + }, + { + "bbox": [ + 131, + 271, + 479, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_1" + }, + { + "bbox": [ + 131, + 271, + 479, + 332 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 271, + 479, + 332 + ], + "type": "inline_equation", + "content": "\\mathbb{P}_{\\mathbb{K}}" + }, + { + "bbox": [ + 131, + 271, + 479, + 332 + ], + "type": "text", + "content": " also predict a filter parameter " + }, + { + "bbox": [ + 131, + 271, + 479, + 332 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 131, + 271, + 479, + 332 + ], + "type": "text", + "content": " (initial at 0) to keep the sharpness and recover details of generated image " + }, + { + "bbox": [ + 131, + 271, + 479, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_2" + }, + { + "bbox": [ + 131, + 271, + 479, + 332 + ], + "type": "text", + "content": ". Eq. 4 shows the translation from " + }, + { + "bbox": [ + 131, + 271, + 479, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_1" + }, + { + "bbox": [ + 131, + 271, + 479, + 332 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 131, + 271, + 479, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_2" + }, + { + "bbox": [ + 131, + 271, + 479, + 332 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 271, + 479, + 332 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 131, + 271, + 479, + 332 + ], + "type": "text", + "content": " denotes kernel convolution and filter parameter " + }, + { + "bbox": [ + 131, + 271, + 479, + 332 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 131, + 271, + 479, + 332 + ], + "type": "text", + "content": " is limited in a range of " + }, + { + "bbox": [ + 131, + 271, + 479, + 332 + ], + "type": "inline_equation", + "content": "(0, 1)" + }, + { + "bbox": [ + 131, + 271, + 479, + 332 + ], + "type": "text", + "content": " by a Sigmoid activation. For more details please refer to our supplementary part Sec. C" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 243, + 341, + 370, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 341, + 370, + 353 + ], + "spans": [ + { + "bbox": [ + 243, + 341, + 370, + 353 + ], + "type": "interline_equation", + "content": "\\mathbb {P} _ {\\mathbb {K}} \\left(\\mathbf {I} _ {1}, \\mathrm {q}\\right)\\rightarrow k \\left\\{r _ {1}, r _ {2}, \\theta \\right\\}, g, \\sigma ,", + "image_path": "9500c046c1a1c4a37777c220692f0076290bd6c40996dddd29a11c152f720c89.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 243, + 356, + 480, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 356, + 480, + 368 + ], + "spans": [ + { + "bbox": [ + 243, + 356, + 480, + 368 + ], + "type": "interline_equation", + "content": "\\mathbf {I} _ {2} ^ {\\prime} = (g \\cdot \\mathbf {I} _ {1}) * k, \\tag {4}", + "image_path": "e3a718669f45298923bb8ea4f7de7f9e6e2cc36b37c6aa5074777a8eb0842fcd.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 243, + 370, + 351, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 370, + 351, + 383 + ], + "spans": [ + { + "bbox": [ + 243, + 370, + 351, + 383 + ], + "type": "interline_equation", + "content": "\\mathbf {I} _ {2} = \\mathbf {I} _ {2} ^ {\\prime} + (g \\cdot \\mathbf {I} _ {1} - \\mathbf {I} _ {2} ^ {\\prime}) \\cdot \\sigma .", + "image_path": "f8c87abfd4fa0c3925c437f854fdf0a30a89aa2cb5471dfd6250c279deb41ac7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 396, + 482, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 396, + 482, + 528 + ], + "spans": [ + { + "bbox": [ + 131, + 396, + 482, + 528 + ], + "type": "text", + "content": "White Balance & CCM Matrix: White balance (WB) mimics the color constancy of the human vision system (HVS) by aligning \"white\" color with the white object, resulting in the captured image reflecting the combination of light color and material reflectance [1,2,14]. In our work, we hope to find an adaptive white balance for different images under various lighting scenarios. Motivated by the design of Shades of Gray (SoG) [20] WB algorithm, where gray-world WB and Max-RGB WB can be regarded as a subcase, we've employed a learnable parameter " + }, + { + "bbox": [ + 131, + 396, + 482, + 528 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 131, + 396, + 482, + 528 + ], + "type": "text", + "content": " to replace the gray-world's L-1 average with an adaptive Minkowski distance average (see Eq. [5]), the QAL block " + }, + { + "bbox": [ + 131, + 396, + 482, + 528 + ], + "type": "inline_equation", + "content": "\\mathbb{P}_{\\mathbb{M}}" + }, + { + "bbox": [ + 131, + 396, + 482, + 528 + ], + "type": "text", + "content": " predicts Minkowski distance's hyper-parameter " + }, + { + "bbox": [ + 131, + 396, + 482, + 528 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 131, + 396, + 482, + 528 + ], + "type": "text", + "content": ", a demonstration of various " + }, + { + "bbox": [ + 131, + 396, + 482, + 528 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 131, + 396, + 482, + 528 + ], + "type": "text", + "content": " is shown in Fig. [4]. After finding the suitable " + }, + { + "bbox": [ + 131, + 396, + 482, + 528 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 131, + 396, + 482, + 528 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 396, + 482, + 528 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_2" + }, + { + "bbox": [ + 131, + 396, + 482, + 528 + ], + "type": "text", + "content": " would multiply to white balance matrix to generate " + }, + { + "bbox": [ + 131, + 396, + 482, + 528 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_3" + }, + { + "bbox": [ + 131, + 396, + 482, + 528 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 529, + 482, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 529, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 131, + 529, + 482, + 589 + ], + "type": "text", + "content": "Color Conversion Matrix (CCM) within the ISP is constrained by the specific camera model. Here we standardize the CCM as a single learnable " + }, + { + "bbox": [ + 131, + 529, + 482, + 589 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 131, + 529, + 482, + 589 + ], + "type": "text", + "content": " matrix, initialized as a unit diagonal matrix " + }, + { + "bbox": [ + 131, + 529, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_3" + }, + { + "bbox": [ + 131, + 529, + 482, + 589 + ], + "type": "text", + "content": ". The QAL block " + }, + { + "bbox": [ + 131, + 529, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\mathbb{P}_{\\mathbb{M}}" + }, + { + "bbox": [ + 131, + 529, + 482, + 589 + ], + "type": "text", + "content": " predicts 9 parameters here, which are then added to " + }, + { + "bbox": [ + 131, + 529, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_3" + }, + { + "bbox": [ + 131, + 529, + 482, + 589 + ], + "type": "text", + "content": " to form the final " + }, + { + "bbox": [ + 131, + 529, + 482, + 589 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 131, + 529, + 482, + 589 + ], + "type": "text", + "content": " matrix " + }, + { + "bbox": [ + 131, + 529, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{ccm}" + }, + { + "bbox": [ + 131, + 529, + 482, + 589 + ], + "type": "text", + "content": ". Then " + }, + { + "bbox": [ + 131, + 529, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_3" + }, + { + "bbox": [ + 131, + 529, + 482, + 589 + ], + "type": "text", + "content": " would multiply to CCM " + }, + { + "bbox": [ + 131, + 529, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{ccm}" + }, + { + "bbox": [ + 131, + 529, + 482, + 589 + ], + "type": "text", + "content": " to generate " + }, + { + "bbox": [ + 131, + 529, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_4" + }, + { + "bbox": [ + 131, + 529, + 482, + 589 + ], + "type": "text", + "content": ", equation as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 217, + 598, + 307, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 598, + 307, + 609 + ], + "spans": [ + { + "bbox": [ + 217, + 598, + 307, + 609 + ], + "type": "interline_equation", + "content": "\\mathbb {P} _ {\\mathbb {M}} \\left(\\mathbf {I} _ {2}, \\mathrm {q}\\right)\\rightarrow \\rho , \\mathbf {E} _ {c c m},", + "image_path": "51de6c0bc7b7186e276f1f790a8fb0f788cbc706b9059a474e921cf3b529e8a2.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 217, + 614, + 392, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 614, + 392, + 628 + ], + "spans": [ + { + "bbox": [ + 217, + 614, + 392, + 628 + ], + "type": "interline_equation", + "content": "m _ {i \\in (r, g, b)} = \\sqrt [ r ]{a v g \\left(\\mathbf {I} _ {2} (i) ^ {\\rho}\\right)} / \\sqrt [ r ]{a v g \\left((\\mathbf {I} _ {2}) ^ {\\rho}\\right)},", + "image_path": "06e7d0a1d330e853a2688053f659705e4cc0d316beb111e855e490a5d2b0bc64.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 216, + 628, + 480, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 628, + 480, + 668 + ], + "spans": [ + { + "bbox": [ + 216, + 628, + 480, + 668 + ], + "type": "interline_equation", + "content": "\\mathbf {I} _ {3} = \\mathbf {I} _ {2} * \\left[ \\begin{array}{c c c} m _ {r} & & \\\\ & m _ {g} & \\\\ & & m _ {b} \\end{array} \\right], \\quad \\mathbf {I} _ {4} = \\mathbf {I} _ {3} * \\mathbf {E} _ {c c m}. \\tag {5}", + "image_path": "5921339f7889acaafe528f4fb4408dc05c29c25d4e116eeea73c9d80f7f0dde1.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "type": "text", + "content": "RAW-Adapter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 473, + 208 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 473, + 208 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 473, + 208 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 473, + 208 + ], + "type": "image", + "image_path": "99aec0a1c43353c69e92d47427f2ed61936cfc1398fcfdab3e2c65b23660d24d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 215, + 482, + 239 + ], + "lines": [ + { + "bbox": [ + 131, + 215, + 482, + 239 + ], + "spans": [ + { + "bbox": [ + 131, + 215, + 482, + 239 + ], + "type": "text", + "content": "Fig. 4: Left, we use query adaptive learning (QAL) to predict key parameters is ISP process. Right, we show RAW-Adapter different blocks' parameter." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 262, + 482, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 382 + ], + "type": "text", + "content": "Color Manipulation Process: In ISP, color manipulation process is commonly achieved through lookup table (LUT), such as 1D and 3D LUTs [16]. Here, we consolidate color manipulation operations into a single 3D LUT, adjusting the color of image " + }, + { + "bbox": [ + 130, + 262, + 482, + 382 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_4" + }, + { + "bbox": [ + 130, + 262, + 482, + 382 + ], + "type": "text", + "content": " to produce output image " + }, + { + "bbox": [ + 130, + 262, + 482, + 382 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_5" + }, + { + "bbox": [ + 130, + 262, + 482, + 382 + ], + "type": "text", + "content": ". Leveraging advancements in LUT techniques, we choose the latest neural implicit 3D LUT (NILUT) [11], for its speed efficiency and ability to facilitate end-to-end learning in our pipeline [3]. We denote NILUT [11] as " + }, + { + "bbox": [ + 130, + 262, + 482, + 382 + ], + "type": "inline_equation", + "content": "\\mathbb{L}" + }, + { + "bbox": [ + 130, + 262, + 482, + 382 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 262, + 482, + 382 + ], + "type": "inline_equation", + "content": "\\mathbb{L}" + }, + { + "bbox": [ + 130, + 262, + 482, + 382 + ], + "type": "text", + "content": " maps the input pixel intensities " + }, + { + "bbox": [ + 130, + 262, + 482, + 382 + ], + "type": "inline_equation", + "content": "R,G,B" + }, + { + "bbox": [ + 130, + 262, + 482, + 382 + ], + "type": "text", + "content": " to a continuous coordinate space, followed by the utilization of implicit neural representation (INR) [61], which involves using a multi-layer perceptron (MLP) network to map the output pixel intensities to " + }, + { + "bbox": [ + 130, + 262, + 482, + 382 + ], + "type": "inline_equation", + "content": "R^{\\prime},G^{\\prime},B^{\\prime}" + }, + { + "bbox": [ + 130, + 262, + 482, + 382 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 235, + 394, + 481, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 394, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 235, + 394, + 481, + 407 + ], + "type": "interline_equation", + "content": "\\mathbf {I} _ {5} \\left(R ^ {\\prime}, G ^ {\\prime}, B ^ {\\prime}\\right) = \\mathbb {L} \\left(\\mathbf {I} _ {4} (R, G, B)\\right). \\tag {6}", + "image_path": "5ae2ab41496627455496752e4064b652a14f89b5639ef761f308679c3c6e794c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 413, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 450 + ], + "type": "text", + "content": "Image " + }, + { + "bbox": [ + 130, + 413, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_5" + }, + { + "bbox": [ + 130, + 413, + 482, + 450 + ], + "type": "text", + "content": " obtained through image-level adapters will be forwarded to the downstream network's backbone. Furthermore, " + }, + { + "bbox": [ + 130, + 413, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_5" + }, + { + "bbox": [ + 130, + 413, + 482, + 450 + ], + "type": "text", + "content": "'s features obtained in backbone will be fused with model-level adapters, which we will discuss in detail." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 468, + 269, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 468, + 269, + 481 + ], + "spans": [ + { + "bbox": [ + 131, + 468, + 269, + 481 + ], + "type": "text", + "content": "3.2 Model-level Adapters" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 490, + 482, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 482, + 574 + ], + "type": "text", + "content": "Input-level adapters guarantee the production of machine vision-oriented image " + }, + { + "bbox": [ + 130, + 490, + 482, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_5" + }, + { + "bbox": [ + 130, + 490, + 482, + 574 + ], + "type": "text", + "content": " for high-level models. However, information at the ISP stages (" + }, + { + "bbox": [ + 130, + 490, + 482, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_1 \\sim \\mathbf{I}_4" + }, + { + "bbox": [ + 130, + 490, + 482, + 574 + ], + "type": "text", + "content": ") is almost overlooked. Inspired by adapter design in current NLP and CV area [8,28,33], we employ the prior information from the ISP stage as model-level adapters to guide subsequent models' perception. Additionally, model-level adapters promote a tighter integration between downstream tasks and the ISP stage. Dotted lines in Fig. 3(a) represent model-level adapters." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "content": "Model-level adapters " + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathbb{M}" + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "content": " integrate the information from ISP stages to the network backbone. As shown in Fig. 3(a), we denote the stage " + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "inline_equation", + "content": "1\\sim 4" + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "content": " as different stages in network backbone, image " + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_5" + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "content": " would pass by backbone stages " + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "inline_equation", + "content": "1\\sim 4" + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "content": " and then followed by detection or segmentation head. We utilize convolution layer " + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "content": " to extract features from " + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_1" + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_4" + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "content": ", these extracted features are concatenated as " + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathbb{C}(\\mathbf{I}_{1\\sim 4}) = \\mathbb{C}(c_1(\\mathbf{I}_1),c_2(\\mathbf{I}_2),c_3(\\mathbf{I}_3),c_4(\\mathbf{I}_4))" + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "content": ". Subsequently, " + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathbb{C}(\\mathbf{I}_{1\\sim 4})" + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "content": " goes through" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "text", + "content": "Cui, Harada." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 652, + 443, + 667 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 652, + 443, + 667 + ], + "spans": [ + { + "bbox": [ + 133, + 652, + 443, + 667 + ], + "type": "text", + "content": "3 To save memory occupy, we change the channel number 128 in [11] to 32." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 187 + ], + "type": "text", + "content": "two residual blocks [25] and generate adapter " + }, + { + "bbox": [ + 130, + 115, + 479, + 187 + ], + "type": "inline_equation", + "content": "f_{1}" + }, + { + "bbox": [ + 130, + 115, + 479, + 187 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 115, + 479, + 187 + ], + "type": "inline_equation", + "content": "f_{1}" + }, + { + "bbox": [ + 130, + 115, + 479, + 187 + ], + "type": "text", + "content": " would merge with stage 1's feature by a merge block, detail structure of merge block is shown in Fig. 3(c), which roughly includes concatenate process and a residual connect, finally merge block would output network feature for stage 2 and an additional adapter " + }, + { + "bbox": [ + 130, + 115, + 479, + 187 + ], + "type": "inline_equation", + "content": "f_{2}" + }, + { + "bbox": [ + 130, + 115, + 479, + 187 + ], + "type": "text", + "content": ". Then process would repeat in stage 2 and stage 3 of network backbone. We collectively refer to all structures related to model-level adapters as " + }, + { + "bbox": [ + 130, + 115, + 479, + 187 + ], + "type": "inline_equation", + "content": "\\mathbb{M}" + }, + { + "bbox": [ + 130, + 115, + 479, + 187 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 308 + ], + "type": "text", + "content": "We show RAW-Adapter's different part parameter number in Fig. 4 right, including input-level adapters " + }, + { + "bbox": [ + 130, + 188, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\{\\mathbb{P}_{\\mathbb{K}}" + }, + { + "bbox": [ + 130, + 188, + 482, + 308 + ], + "type": "text", + "content": " (37.57K), " + }, + { + "bbox": [ + 130, + 188, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\mathbb{P}_{\\mathbb{M}}" + }, + { + "bbox": [ + 130, + 188, + 482, + 308 + ], + "type": "text", + "content": " (37.96K), " + }, + { + "bbox": [ + 130, + 188, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\mathbb{L}" + }, + { + "bbox": [ + 130, + 188, + 482, + 308 + ], + "type": "text", + "content": " (1.97K) and model level adapters " + }, + { + "bbox": [ + 130, + 188, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\mathbb{M}" + }, + { + "bbox": [ + 130, + 188, + 482, + 308 + ], + "type": "text", + "content": " (114.87K " + }, + { + "bbox": [ + 130, + 188, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 130, + 188, + 482, + 308 + ], + "type": "text", + "content": " 687.59K), model level adapters' parameter number depend on following network. Total parameter number of RAW-Adapter is around 0.2M to 0.8M, much smaller than following backbones (" + }, + { + "bbox": [ + 130, + 188, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 130, + 188, + 482, + 308 + ], + "type": "text", + "content": "25.6M in ResNet-50 [25], " + }, + { + "bbox": [ + 130, + 188, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 130, + 188, + 482, + 308 + ], + "type": "text", + "content": "197M in Swin-L [45]), also smaller than previous SOTA methods like SID [6] (11.99M) and Dirty-Pixel [18] (4.28M). Additionally, subsequent experiments will demonstrate that the performance of RAW-Adapter on backbone networks with fewer parameters is even superior to that of previous algorithms on backbone networks with higher parameters." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 329, + 230, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 329, + 230, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 329, + 230, + 342 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 357, + 332, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 357, + 332, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 332, + 369 + ], + "type": "text", + "content": "4.1 Dataset and Experimental Setting" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 380, + 479, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 479, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 479, + 416 + ], + "type": "text", + "content": "Dataset. We conducted experiments on object detection and semantic segmentation tasks, utilizing a combination of various synthetic and real-world RAW image datasets, an overview of the datasets can be found in Table 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 417, + 482, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 417, + 482, + 549 + ], + "spans": [ + { + "bbox": [ + 130, + 417, + 482, + 549 + ], + "type": "text", + "content": "For the object detection task, we take 2 open source real-world dataset PASCAL RAW [52] and LOD [27]. PASCAL RAW [52] is a normal-light condition dataset with 4259 RAW images, taken by a Nikon D3200 DSLR camera with 3 object classes. To verify the generalization capability of RAW-Adapter across various lighting conditions, we additionally synthesized low-light and overexposure datasets based on the PASCAL RAW dataset, named PASCAL RAW (dark) and PASCAL RAW (over-exp) respectively, the synthesized datasets are identical to the PASCAL RAW dataset in all aspects except for brightness levels. For PASCAL RAW (dark) and PASCAL RAW (over-exp) synthesis, the light intensity and environment irradiance exhibit a linear relationship on RAW images, thus we employed the synthesis method from the previous work [14,78]:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 238, + 559, + 376, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 559, + 376, + 586 + ], + "spans": [ + { + "bbox": [ + 238, + 559, + 376, + 586 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} x _ {n} \\sim N (\\mu = l x, \\sigma^ {2} = \\delta_ {r} ^ {2} + \\delta_ {s} l x) \\\\ y = l x + x _ {n}, \\\\ \\end{array}", + "image_path": "d7fadce4faa5b1e03e23306675fcac21b3e2d8921144676a64814a13554cbc66.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 466, + 567, + 479, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 466, + 567, + 479, + 578 + ], + "spans": [ + { + "bbox": [ + 466, + 567, + 479, + 578 + ], + "type": "text", + "content": "(7)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 593, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 593, + 482, + 665 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 593, + 482, + 665 + ], + "type": "text", + "content": " denotes the original normal-light RAW image and " + }, + { + "bbox": [ + 130, + 593, + 482, + 665 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 593, + 482, + 665 + ], + "type": "text", + "content": " denotes degraded RAW image, " + }, + { + "bbox": [ + 130, + 593, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\delta_s = \\sqrt{S}" + }, + { + "bbox": [ + 130, + 593, + 482, + 665 + ], + "type": "text", + "content": " denotes shot noise while " + }, + { + "bbox": [ + 130, + 593, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\sqrt{S}" + }, + { + "bbox": [ + 130, + 593, + 482, + 665 + ], + "type": "text", + "content": " is signal of the sensor, " + }, + { + "bbox": [ + 130, + 593, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\delta_r" + }, + { + "bbox": [ + 130, + 593, + 482, + 665 + ], + "type": "text", + "content": " denotes read noise, " + }, + { + "bbox": [ + 130, + 593, + 482, + 665 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 593, + 482, + 665 + ], + "type": "text", + "content": " denotes the light intensity parameter which randomly chosen from [0.05, 0.4] in PASCAL RAW (dark), and randomly chosen from [1.5, 2.5] in PASCAL RAW (over-exp). Additionally, we follow PASCAL RAW [52]'s dataset split to separate the training set and test set." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "type": "text", + "content": "RAW-Adapter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 136, + 482, + 253 + ], + "blocks": [ + { + "bbox": [ + 181, + 114, + 430, + 126 + ], + "lines": [ + { + "bbox": [ + 181, + 114, + 430, + 126 + ], + "spans": [ + { + "bbox": [ + 181, + 114, + 430, + 126 + ], + "type": "text", + "content": "Table 1: Dataset and framework setting in our experiments." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 136, + 482, + 253 + ], + "lines": [ + { + "bbox": [ + 132, + 136, + 482, + 253 + ], + "spans": [ + { + "bbox": [ + 132, + 136, + 482, + 253 + ], + "type": "table", + "html": "
PASCAL RAWPASCAL RAW (dark/over-exp)LODADE20K RAWADE20K RAW (dark/over-exp)
TaskObject DetectionSemantic Segmentation
Number4259223027,574
Typereal-worldsynthesisreal-worldsynthesis
SensorNikon D3200 DSLRCanon EOS 5D Mark IV-
FrameworkRetinaNet43 & Sparse-RCNN64Segformer70
BackboneResNet25MIT70
pre-trainImageNet44 pre-train weights
", + "image_path": "9fa8b16e81d2e3fe71ee0d1b2830e12ebde4e47d1a627503640e79c84dac4e70.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 274, + 479, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 479, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 479, + 310 + ], + "type": "text", + "content": "Meanwhile, LOD [27] is a real-world dataset with 2230 low-light condition RAW images, taken by a Canon EOS 5D Mark IV camera with 8 object classes, we take 1800 images as training set and the other 430 images as test set." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 310, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 482, + 407 + ], + "type": "text", + "content": "For the semantic segmentation task, we utilized the widely-used sRGB dataset ADE20K [80] to generate RAW dataset. Leveraging state-of-the-art unprocessing methods InvISP [71], we synthesized RAW images corresponding to the ADE20K sRGB dataset. Using InvISP, we projected input sRGB images into RAW format, effectively creating an ADE20K RAW dataset. To simulate various lighting conditions, we employed the same synthesis method from PASCAL RAW (dark/over-exp) to generate low-light and over-exposure RAW images, name as ADE20K RAW (dark/over-exp), training & test split is same as ADE20K." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 424, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 536 + ], + "type": "text", + "content": "Implement Details. We build our framework based on the open-source computer vision toolbox mmdetection 7 and mmsegmentation 12, both object detection tasks and semantic segmentation tasks are initialed with ImageNet pre-train weights (see Table. 1), and we apply the data augmentation pipeline in the default setting, mainly include random crop, random flip, and multi-scale test, etc. For the object detection task, we adopt the 2 mainstream object detectors: RetinaNet 43 and Sparse-RCNN 64 with ResNet 25 backbone. For the semantic segmentation task, we choose to use the mainstream segmentation framework Segformer 70 with MIT 70 backbone." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "type": "text", + "content": "Comparison Methods. We conducted comparative experiments with the current state-of-the-art (SOTA) algorithms, including various open-sourced ISP methods [6,34,35,71,79] and joint-training method DirtyPixels [18], among these Karaimer et al. [35] is a traditional ISP method with various human manipulate steps, and InvISP [71] & Lite-ISP [79] & SID [6] & DNF [34] are current SOTA network-based ISP methods, where SID [6] and DNF [34] is especially for the low-light condition RAW data. For fairness compared with the above ISP methods, both the training and test RAW data are rendered using the respective" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "text", + "content": "Cui, Harada." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 481, + 665 + ], + "type": "text", + "content": "4 For more experiments on other frameworks, please refer to supplementary materials." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 113, + 482, + 276 + ], + "blocks": [ + { + "bbox": [ + 134, + 113, + 482, + 276 + ], + "lines": [ + { + "bbox": [ + 134, + 113, + 482, + 276 + ], + "spans": [ + { + "bbox": [ + 134, + 113, + 482, + 276 + ], + "type": "image", + "image_path": "42d81234c539124aac78a75e373137568ffe4efd7191f5e118450b771dbefc12.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 283, + 480, + 308 + ], + "lines": [ + { + "bbox": [ + 132, + 283, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 283, + 480, + 308 + ], + "type": "text", + "content": "Fig. 5: (a). Detection performance on PASCAL RAW [52] (normal/dark/over-exp). (b). Efficiency comparison (blue: vanilla, yellow: RAW-Adapter, gray: Dirty-Pixel [18])." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 328, + 482, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 482, + 390 + ], + "type": "text", + "content": "compared ISP. DirtyPixels [18] is the current SOTA joint-training method which uses a stack of residual UNet [58] as a low-level processor and joint optimization with the following task-specific networks. For fairness, all comparison methods adopt the same data augmentation process and the same training setting, we will introduce detailed experiment results in the following section." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 406, + 305, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 406, + 305, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 406, + 305, + 418 + ], + "type": "text", + "content": "4.2 Object Detection Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 617 + ], + "type": "text", + "content": "For object detection task on PASCAL RAW [52] dataset, we adopt RetinaNet [43] with different size ResNet [25] backbones (ResNet-18, ResNet-50), all the models are trained on a single NVIDIA Tesla V100 GPU with SGD optimizer, the batch size is set to 4, training images are cropped into range of (400, 667) and training epochs are set to 50. Table. [2] shows the detection results with (a). ResNet-18 and (b). ResNet-50 backbone, with comparisons of demosaiced RAW data (\"Demosacing\"), camera default ISP in [52], various current SOTA ISP solutions [6,34,35,79] and Dirty-Pixel [18], we can observe that sometimes ISP algorithms in normal and over-exposure conditions may even degraded the detection performance, and ISP in dark scenarios sometimes could improves detection performance, additionally manual design ISP solution [35] even out-perform the deep learning solutions. Previous joint-training method Dirty-Pixel [18] can improve detection performance. Overall, our RAW-Adapter method achieves the best performance and even outperforms some ISP algorithms utilizing ResNet-50 backbone when employing ResNet-18 backbone, which can significantly reduce computational load and model parameters." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Visualization of PASCAL RAW's detection results are shown in Fig. 5(a), we show the detection results under different lightness conditions, background are images generated from Dirty-Pixel [18] and RAW-Adapter " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathbf{I}_5)" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ". Our method could achieve satisfactory detection results across different lightness, while other" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "type": "text", + "content": "RAW-Adapter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 175, + 302, + 316 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 2: Comparison results on PASCAL RAW dataset [52], we take [43] with (a). ResNet-18 and (b). ResNet-50 backbone, bold denotes the best result." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 189, + 156, + 244, + 166 + ], + "lines": [ + { + "bbox": [ + 189, + 156, + 244, + 166 + ], + "spans": [ + { + "bbox": [ + 189, + 156, + 244, + 166 + ], + "type": "text", + "content": "(a) ResNet-18." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 175, + 302, + 316 + ], + "lines": [ + { + "bbox": [ + 132, + 175, + 302, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 175, + 302, + 316 + ], + "type": "table", + "html": "
mAP Methodnormal over-expdark
Default ISP88.3--
Demosacing87.787.780.3
Karäimer et al. 3586.085.681.9
Lite-ISP 7985.284.271.9
InvISP 7184.186.670.9
SID 6--78.2
DNF 34--81.1
Dirty-Pixel88.688.080.8
RAW-Adapter88.788.782.5
", + "image_path": "358f237cb7c0bb96d33e1676e4cd9f34639804082ec47e540fe423a272bc318a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 309, + 175, + 480, + 316 + ], + "blocks": [ + { + "bbox": [ + 369, + 156, + 423, + 166 + ], + "lines": [ + { + "bbox": [ + 369, + 156, + 423, + 166 + ], + "spans": [ + { + "bbox": [ + 369, + 156, + 423, + 166 + ], + "type": "text", + "content": "(b) ResNet-50." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 175, + 480, + 316 + ], + "lines": [ + { + "bbox": [ + 309, + 175, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 309, + 175, + 480, + 316 + ], + "type": "table", + "html": "
mAP Methodnormal over-expdark
Default ISP89.4--
Demosacing89.288.882.6
Karäimer et al. 3589.486.884.6
Lite-ISP 7989.385.173.5
InvISP 7189.187.374.7
SID 6--81.5
DNF 34--82.8
Dirty-Pixel89.089.083.6
RAW-Adapter89.689.486.6
", + "image_path": "bd05fbdf835d93a0c3d0170cfa120b144c9b89399c44960d86b3d96829e20772.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 132, + 369, + 482, + 460 + ], + "blocks": [ + { + "bbox": [ + 132, + 325, + 481, + 361 + ], + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 361 + ], + "type": "text", + "content": "Table 3: Comparison with ISP methods and Dirty-Pixel on LOD dataset [27]. We show the detection performance (mAP) ↑ of RetinaNet (R-Net) [43] and Sparse-RCNN (Sp-RCNN) [64], **bold** denotes the best result." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 369, + 482, + 460 + ], + "lines": [ + { + "bbox": [ + 132, + 369, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 482, + 460 + ], + "type": "table", + "html": "
MethodsDemosaicing Karaimer et al. 35InvISP 71SID 6
mAP(R-Net)58.554.456.949.1
mAP(Sp-RCNN)57.752.249.443.1
MethodsDNF 34Dirty-Pixel 18RAW-Adapter (w/o M)RAW-Adapter
mAP(R-Net)58.861.661.662.2
mAP(Sp-RCNN)55.858.858.659.2
", + "image_path": "815c08731aa74c2fcd18e028fac35cee2213a2de136db054419fc44f3b564540.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 482, + 480, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 482, + 480, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 482, + 480, + 517 + ], + "type": "text", + "content": "methods face false alarm and miss detect. Efficiency comparison with Dirty-Pixel is shown in Fig. 5(b), our algorithm has achieved significant improvements in both training time acceleration and savings in Flops/Parameter Number." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 518, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 518, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 518, + 482, + 591 + ], + "type": "text", + "content": "Detection performance on LOD [27] is shown in Table. [3], we adopt two detector RetinaNet [43] and Sparse-RCNN [64], both with ResNet-50 [25] backbone, training epochs are set to 35, Sparse-RCNN [64] are trained by 4 GPUs with Adam optimizer. With comparing of various ISP methods [6,34,35,79], Dirty-Pixel [18] and RAW-Adapter with only input-level adapters (w/o M). Table. [3] show that our algorithm can achieve the best performance on both two detectors." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 609, + 335, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 335, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 335, + 620 + ], + "type": "text", + "content": "4.3 Semantic Segmentation Evaluation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 628, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 628, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 628, + 481, + 666 + ], + "type": "text", + "content": "For semantic segmentation on ADE20K RAW dataset, we choose Segformer [70] as the segmentation framework with the different size MIT [70] backbones (MIT-B5, MIT-B3, MIT-B0), all the models are trained on 4 NVIDIA Tesla V100 GPU" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "text", + "content": "Cui, Harada." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 202, + 162 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 202, + 162 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 202, + 162 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 202, + 162 + ], + "type": "image", + "image_path": "43043d2a432760b276f0eaa25b628295141b3111f94a5b4507b9b95f2684bbe7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 136, + 163, + 201, + 210 + ], + "blocks": [ + { + "bbox": [ + 136, + 163, + 201, + 210 + ], + "lines": [ + { + "bbox": [ + 136, + 163, + 201, + 210 + ], + "spans": [ + { + "bbox": [ + 136, + 163, + 201, + 210 + ], + "type": "image", + "image_path": "c8193287929b9033283ff25322727b8a24c30031cbe015f9b52a33d0f0ecc9c3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 211, + 189, + 219 + ], + "lines": [ + { + "bbox": [ + 154, + 211, + 189, + 219 + ], + "spans": [ + { + "bbox": [ + 154, + 211, + 189, + 219 + ], + "type": "text", + "content": "Input / GT" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 205, + 115, + 271, + 162 + ], + "blocks": [ + { + "bbox": [ + 205, + 115, + 271, + 162 + ], + "lines": [ + { + "bbox": [ + 205, + 115, + 271, + 162 + ], + "spans": [ + { + "bbox": [ + 205, + 115, + 271, + 162 + ], + "type": "image", + "image_path": "e4f2a078d5327cf3870558799ae48cbe0487cc4586202ea0f209d8756e71de6e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 206, + 163, + 271, + 210 + ], + "blocks": [ + { + "bbox": [ + 206, + 163, + 271, + 210 + ], + "lines": [ + { + "bbox": [ + 206, + 163, + 271, + 210 + ], + "spans": [ + { + "bbox": [ + 206, + 163, + 271, + 210 + ], + "type": "image", + "image_path": "76af0ecf44d389d21af0d474151c8e9c61beb874a7b96e1f43881a5c46f24645.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 211, + 250, + 218 + ], + "lines": [ + { + "bbox": [ + 227, + 211, + 250, + 218 + ], + "spans": [ + { + "bbox": [ + 227, + 211, + 250, + 218 + ], + "type": "text", + "content": "LiteISP" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 276, + 114, + 341, + 162 + ], + "blocks": [ + { + "bbox": [ + 276, + 114, + 341, + 162 + ], + "lines": [ + { + "bbox": [ + 276, + 114, + 341, + 162 + ], + "spans": [ + { + "bbox": [ + 276, + 114, + 341, + 162 + ], + "type": "image", + "image_path": "44c2d3340b05c06f0411900e178c0de0a616f8f8e2beb622cc0d24ccbb86a69a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 276, + 163, + 341, + 210 + ], + "blocks": [ + { + "bbox": [ + 276, + 163, + 341, + 210 + ], + "lines": [ + { + "bbox": [ + 276, + 163, + 341, + 210 + ], + "spans": [ + { + "bbox": [ + 276, + 163, + 341, + 210 + ], + "type": "image", + "image_path": "8b21f471d384198de87064f75c2ed219886469de1c075a3d665cdbffcdcde120.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 297, + 211, + 318, + 218 + ], + "lines": [ + { + "bbox": [ + 297, + 211, + 318, + 218 + ], + "spans": [ + { + "bbox": [ + 297, + 211, + 318, + 218 + ], + "type": "text", + "content": "InvISP" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 344, + 114, + 410, + 162 + ], + "blocks": [ + { + "bbox": [ + 344, + 114, + 410, + 162 + ], + "lines": [ + { + "bbox": [ + 344, + 114, + 410, + 162 + ], + "spans": [ + { + "bbox": [ + 344, + 114, + 410, + 162 + ], + "type": "image", + "image_path": "cf27c40c973a7170e58f2816d4c3e3d2d84f862a60eeba7b28d2aeaa320426bd.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 344, + 163, + 410, + 210 + ], + "blocks": [ + { + "bbox": [ + 344, + 163, + 410, + 210 + ], + "lines": [ + { + "bbox": [ + 344, + 163, + 410, + 210 + ], + "spans": [ + { + "bbox": [ + 344, + 163, + 410, + 210 + ], + "type": "image", + "image_path": "628ac741a03d8487f813f8227eb1e1eb9a93781a41656fa7e26475947a1f20c4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 360, + 211, + 394, + 219 + ], + "lines": [ + { + "bbox": [ + 360, + 211, + 394, + 219 + ], + "spans": [ + { + "bbox": [ + 360, + 211, + 394, + 219 + ], + "type": "text", + "content": "Dirty-Pixel" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 413, + 114, + 479, + 162 + ], + "blocks": [ + { + "bbox": [ + 413, + 114, + 479, + 162 + ], + "lines": [ + { + "bbox": [ + 413, + 114, + 479, + 162 + ], + "spans": [ + { + "bbox": [ + 413, + 114, + 479, + 162 + ], + "type": "image", + "image_path": "c3d438cf47663ee167d54b4df4a571b1946916bde817ee7f245cb2f8b76c047c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 413, + 163, + 479, + 210 + ], + "blocks": [ + { + "bbox": [ + 413, + 163, + 479, + 210 + ], + "lines": [ + { + "bbox": [ + 413, + 163, + 479, + 210 + ], + "spans": [ + { + "bbox": [ + 413, + 163, + 479, + 210 + ], + "type": "image", + "image_path": "d0c36cdfb0f926effb510b67d7ad725e9d8ae65ddd28d3a1dbbc69681d896acc.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 424, + 211, + 470, + 218 + ], + "lines": [ + { + "bbox": [ + 424, + 211, + 470, + 218 + ], + "spans": [ + { + "bbox": [ + 424, + 211, + 470, + 218 + ], + "type": "text", + "content": "RAW-Adapter" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 136, + 221, + 202, + 270 + ], + "blocks": [ + { + "bbox": [ + 136, + 221, + 202, + 270 + ], + "lines": [ + { + "bbox": [ + 136, + 221, + 202, + 270 + ], + "spans": [ + { + "bbox": [ + 136, + 221, + 202, + 270 + ], + "type": "image", + "image_path": "1514dfba31eff02384302b16778c9ac4ceceb7f59229d600c11bf0420f1dd62b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 136, + 271, + 201, + 319 + ], + "blocks": [ + { + "bbox": [ + 136, + 271, + 201, + 319 + ], + "lines": [ + { + "bbox": [ + 136, + 271, + 201, + 319 + ], + "spans": [ + { + "bbox": [ + 136, + 271, + 201, + 319 + ], + "type": "image", + "image_path": "b44d4e10413562e366e88c7cc87cbde419ac36ebac93b11343f500c94b24eb4e.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 321, + 190, + 330 + ], + "lines": [ + { + "bbox": [ + 156, + 321, + 190, + 330 + ], + "spans": [ + { + "bbox": [ + 156, + 321, + 190, + 330 + ], + "type": "text", + "content": "Input / GT" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 206, + 222, + 271, + 270 + ], + "blocks": [ + { + "bbox": [ + 206, + 222, + 271, + 270 + ], + "lines": [ + { + "bbox": [ + 206, + 222, + 271, + 270 + ], + "spans": [ + { + "bbox": [ + 206, + 222, + 271, + 270 + ], + "type": "image", + "image_path": "766e832983604b3b8e0a1bcbbb5d5f3c6b75844c98adc6d5eac514048e16e010.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 206, + 271, + 271, + 319 + ], + "blocks": [ + { + "bbox": [ + 206, + 271, + 271, + 319 + ], + "lines": [ + { + "bbox": [ + 206, + 271, + 271, + 319 + ], + "spans": [ + { + "bbox": [ + 206, + 271, + 271, + 319 + ], + "type": "image", + "image_path": "eaf10951855c1e7b00355437e9bf43731f5ff6348214d1a84434974dc6210f61.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 320, + 245, + 329 + ], + "lines": [ + { + "bbox": [ + 233, + 320, + 245, + 329 + ], + "spans": [ + { + "bbox": [ + 233, + 320, + 245, + 329 + ], + "type": "text", + "content": "SID" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 276, + 221, + 340, + 270 + ], + "blocks": [ + { + "bbox": [ + 276, + 221, + 340, + 270 + ], + "lines": [ + { + "bbox": [ + 276, + 221, + 340, + 270 + ], + "spans": [ + { + "bbox": [ + 276, + 221, + 340, + 270 + ], + "type": "image", + "image_path": "546c157963da137b9df9751c2c0e23d803f8f81bf9f26209e8bebf0708c4459b.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 276, + 271, + 340, + 319 + ], + "blocks": [ + { + "bbox": [ + 276, + 271, + 340, + 319 + ], + "lines": [ + { + "bbox": [ + 276, + 271, + 340, + 319 + ], + "spans": [ + { + "bbox": [ + 276, + 271, + 340, + 319 + ], + "type": "image", + "image_path": "dc170adc5338b02bf615689f40891bfc32bd0a0ff064c196e6f2aff99c71ae0e.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 285, + 320, + 334, + 330 + ], + "lines": [ + { + "bbox": [ + 285, + 320, + 334, + 330 + ], + "spans": [ + { + "bbox": [ + 285, + 320, + 334, + 330 + ], + "type": "text", + "content": "Karaimer et al." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 344, + 221, + 410, + 270 + ], + "blocks": [ + { + "bbox": [ + 344, + 221, + 410, + 270 + ], + "lines": [ + { + "bbox": [ + 344, + 221, + 410, + 270 + ], + "spans": [ + { + "bbox": [ + 344, + 221, + 410, + 270 + ], + "type": "image", + "image_path": "7a65f06628b24f3f75d006a461f46f87c194ffea49dddd0dfa54024eb7583f65.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 344, + 271, + 410, + 319 + ], + "blocks": [ + { + "bbox": [ + 344, + 271, + 410, + 319 + ], + "lines": [ + { + "bbox": [ + 344, + 271, + 410, + 319 + ], + "spans": [ + { + "bbox": [ + 344, + 271, + 410, + 319 + ], + "type": "image", + "image_path": "111aad8e97a5ac64533d6517df20d178561fb84fa8f690f86932affec39dd0d9.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 358, + 320, + 392, + 330 + ], + "lines": [ + { + "bbox": [ + 358, + 320, + 392, + 330 + ], + "spans": [ + { + "bbox": [ + 358, + 320, + 392, + 330 + ], + "type": "text", + "content": "Dirty-Pixel" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 339, + 480, + 364 + ], + "lines": [ + { + "bbox": [ + 132, + 339, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 339, + 480, + 364 + ], + "type": "text", + "content": "Fig. 6: Semantic segmentation results on over-exposure and low-light RAW data, compare with various ISP methods [6], [35], [71], [79] and Dirty-Pixel [18]." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 413, + 221, + 479, + 270 + ], + "blocks": [ + { + "bbox": [ + 413, + 221, + 479, + 270 + ], + "lines": [ + { + "bbox": [ + 413, + 221, + 479, + 270 + ], + "spans": [ + { + "bbox": [ + 413, + 221, + 479, + 270 + ], + "type": "image", + "image_path": "be7242f00efc4246400c470c9a2fec669fccc661513f726dd10c481fd309cdf2.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 413, + 271, + 479, + 319 + ], + "blocks": [ + { + "bbox": [ + 413, + 271, + 479, + 319 + ], + "lines": [ + { + "bbox": [ + 413, + 271, + 479, + 319 + ], + "spans": [ + { + "bbox": [ + 413, + 271, + 479, + 319 + ], + "type": "image", + "image_path": "d59502f2dfe0b97849a159e2a8fa294471a3852c3f2d4e4f5423d16306951ff7.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 424, + 320, + 471, + 329 + ], + "lines": [ + { + "bbox": [ + 424, + 320, + 471, + 329 + ], + "spans": [ + { + "bbox": [ + 424, + 320, + 471, + 329 + ], + "type": "text", + "content": "RAW-Adapter" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "bbox": [ + 130, + 388, + 480, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 480, + 435 + ], + "type": "text", + "content": "with Adam optimizer, the batch size is set to 4, training images are cropped into " + }, + { + "bbox": [ + 130, + 388, + 480, + 435 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 130, + 388, + 480, + 435 + ], + "type": "text", + "content": " and training iter are set to 80000. We make comparison with single demosacing, various ISP methods [6,34,35,71,79] and Dirty-Pixel [18], and RAW-Adapter with only input-level adapters (w/o M)." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 130, + 437, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 640 + ], + "type": "text", + "content": "Comparison results are shown in Table. 4, where we compare both efficiency (parameters↓, inference time↓) and performance (mIOU↑), inference time is calculated on a single Tesla V100 GPU. We first show the results on vanilla demosaiced RAW data, then we show the results on RAW data processed by various ISP methods [6,34,35,71,79], followed by results of Dirty-Pixel [18], and our method without and with model-level adapter M. For Table. 4, we can find that directly employing ISP methods sometimes fails to enhance the performance of downstream segmentation tasks, particularly in low-light conditions. Joint-training method Dirty-Pixel [18] could improve down-stream performance and sometimes even degrade segmentation performance, but increase more network parameters. Overall, our proposed RAW-Adapter could effectively enhance performance while adding a slight parameter count and improving inference efficiency. RAW-Adapter on backbones with fewer parameters can even outperform previous methods on backbones with more parameters (i.e. The RAW-Adapter, employing the MIT-B3 backbone, achieves a dark condition mIOU of 40.52, surpassing other methods utilizing the MIT-B5 backbone despite having fewer than " + }, + { + "bbox": [ + 130, + 437, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 130, + 437, + 482, + 640 + ], + "type": "text", + "content": " 36M parameters.)." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "The visualization of segmentation results are shown in Fig. 6, where row 1 and row 3 depict images rendered by different ISPs [6, 35, 71, 79] and images" + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "type": "text", + "content": "RAW-Adapter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 147, + 482, + 377 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 4: Comparison with previous methods on ADE 20K RAW (normal/over-exp/dark). Bold denotes the best result while underline denotes second best result." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 147, + 482, + 377 + ], + "lines": [ + { + "bbox": [ + 132, + 147, + 482, + 377 + ], + "spans": [ + { + "bbox": [ + 132, + 147, + 482, + 377 + ], + "type": "table", + "html": "
backboneparams(M) ↓inference time(s) ↓mIOU ↑ (normal)mIOU ↑ (over-exp)mIOU ↑ (dark)
DemosacingMIT-B582.010.10545.2642.0830.97
Karaimer et al. [35]0.52546.4243.7139.79
InvISP [71]0.20347.5343.554.87
LiteISP [79]0.26143.2242.015.52
DNF [34]0.186--25.82
SID [6]0.312--26.44
Dirty-Pixel [18]86.290.15947.7646.8040.02
MIT-B348.920.09847.1945.3338.93
MIT-B08.000.04935.4332.1027.19
RAW-Adapter (w/o M)MIT-B582.090.14848.3346.4840.91
MIT-B344.720.08647.0145.0039.34
MIT-B03.800.03235.0932.8227.29
RAW-AdapterMIT-B582.530.16748.3847.0641.82
MIT-B345.160.10247.5945.6640.52
MIT-B04.260.05335.6733.4427.78
", + "image_path": "6221db2e5489ffb04015b9b29718336ab4e98da39e1da919e8970ba80c64effc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 398, + 483, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 483, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 483, + 495 + ], + "type": "text", + "content": "generated from Dirty-Pixel [18] and RAW-Adapter " + }, + { + "bbox": [ + 130, + 398, + 483, + 495 + ], + "type": "inline_equation", + "content": "(\\mathbf{I}_5)" + }, + { + "bbox": [ + 130, + 398, + 483, + 495 + ], + "type": "text", + "content": ", row 2 and row 4 denote our segmentation results appear superior compared to other methods. We were even pleased to discover that, although we do not add any human-vision-oriented loss function to constrain the model, RAW-Adapter still produces visually satisfactory image " + }, + { + "bbox": [ + 130, + 398, + 483, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_5" + }, + { + "bbox": [ + 130, + 398, + 483, + 495 + ], + "type": "text", + "content": " (see Fig. 6). This may be attributed to our preservation of traditional ISP stages in input-level adapters. Due to page limitations, we refer to more experimental results, visualization results, and ablation analysis of our method in the supplementary part." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 510, + 220, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 220, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 220, + 523 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 484, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 484, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 484, + 593 + ], + "type": "text", + "content": "In this paper, we introduce RAW-Adapter, an effective solution for adapting pre-trained sRGB models to camera RAW data. With input-level adapters and model-level adapters working in tandem, RAW-Adapter effectively forwards RAW images from various lighting conditions to downstream vision tasks, our method has achieved state-of-the-art (SOTA) results across multiple datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "For future research directions, we believe it is feasible to design a unified model capable of adapting to RAW visual tasks under different lighting conditions, without the need for retraining in each lighting scenario like RAW-Adapter. Additionally, designing multi-task decoders can enable the accommodation of various tasks, facilitating more effective integration of different visual tasks on RAW images with large-scale models." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "text", + "content": "Cui, Harada." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 482, + 175 + ], + "type": "text", + "content": "This research is partially supported by JST Moonshot R&D Grant Number JPMJPS2011, CREST Grant Number JPMJCR2015 and Basic Research Grant (Super AI) of Institute for AI and Beyond of the University of Tokyo." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 217, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 217, + 482, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 482, + 262 + ], + "type": "text", + "content": "1. Affifi, M., Brown, M.S.: What else can fool deep learning? addressing color constancy errors on deep neural network performance. 2019 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 243-252 (2019), https://apisemantic scholar.org/CorpusID:207995696" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 262, + 482, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 482, + 294 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 482, + 294 + ], + "type": "text", + "content": "2. Brooks, T., Mildenhall, B., Xue, T., Chen, J., Sharlet, D., Barron, J.T.: Unprocessing images for learned raw denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11036-11045 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 294, + 482, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 294, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 138, + 294, + 482, + 327 + ], + "type": "text", + "content": "3. Buckler, M., Jayasuriya, S., Sampson, A.: Reconfiguring the imaging pipeline for computer vision. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 975-984 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 327, + 482, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 327, + 482, + 360 + ], + "spans": [ + { + "bbox": [ + 138, + 327, + 482, + 360 + ], + "type": "text", + "content": "4. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 360, + 482, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 360, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 360, + 482, + 392 + ], + "type": "text", + "content": "5. Chatterjee, P., Joshi, N., Kang, S.B., Matsushita, Y.: Noise suppression in low-light images through joint denoising and demosaicing. In: CVPR 2011. pp. 321-328. IEEE (2011)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 392, + 482, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 482, + 425 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 482, + 425 + ], + "type": "text", + "content": "6. Chen, C., Chen, Q., Xu, J., Koltun, V.: Learning to see in the dark. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3291-3300 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 425, + 482, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 482, + 447 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 482, + 447 + ], + "type": "text", + "content": "7. Chen, K., Wang, J., et al.: Mmdetection: Open mmlab detection toolbox and benchmark (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 447, + 482, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 482, + 480 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 482, + 480 + ], + "type": "text", + "content": "8. Chen, Z., Duan, Y., Wang, W., He, J., Lu, T., Dai, J., Qiao, Y.: Vision transformer adapter for dense predictions. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=plKu2GByCNW" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 480, + 482, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 480, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 480, + 482, + 502 + ], + "type": "text", + "content": "9. Cheng, B., Schwing, A.G., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. In: NeurIPS (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 502, + 482, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 482, + 534 + ], + "type": "text", + "content": "0. Conde, M.V., McDonagh, S., Maggioni, M., Leonardis, A., Pérez-Pellitero, E.: Model-based image signal processors via learnable dictionaries. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 481-489 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 534, + 482, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 482, + 567 + ], + "type": "text", + "content": "1. Conde, M.V., Vazquez-Corral, J., Brown, M.S., Timofte, R.: Nilut: Conditional neural implicit 3d lookup tables for image enhancement. In: Proceedings of the AAAI Conference on Artificial Intelligence (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 567, + 482, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 482, + 589 + ], + "type": "text", + "content": "2. Contributors, M.: MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 589, + 482, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 482, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 482, + 621 + ], + "type": "text", + "content": "3. Cui, Z., Li, K., Gu, L., Su, S., Gao, P., Jiang, Z., Qiao, Y., Harada, T.: You only need 90k parameters to adapt light: a light weight transformer for image enhancement and exposure correction. In: BMVC. p. 238 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 621, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 482, + 665 + ], + "type": "text", + "content": "4. Cui, Z., Qi, G.J., Gu, L., You, S., Zhang, Z., Harada, T.: Multitask aet with orthogonal tangent regularity for dark object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2553-2562 (October 2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "type": "text", + "content": "RAW-Adapter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "15. Cui, Z., Zhu, Y., Gu, L., Qi, G.J., Li, X., Zhang, R., Zhang, Z., Harada, T.: Exploring resolution and degradation clues as self-supervised signal for low quality object detection. In: European Conference on Computer Vision. pp. 473-491. Springer (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 162, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 162, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 481, + 205 + ], + "type": "text", + "content": "16. Delbracio, M., Kelly, D., Brown, M.S., Milanfar, P.: Mobile computational photography: A tour. Annual Review of Vision Science 7(1), 571-604 (2021). https://doi.org/10.1146/annurev-vision-093019-115521, https://doi.org/10.1146/annurev-vision-093019-115521, pMID: 34524880" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 206, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 481, + 228 + ], + "type": "text", + "content": "17. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A Large-Scale Hierarchical Image Database. In: CVPR09 (2009)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 228, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 481, + 262 + ], + "type": "text", + "content": "18. Diamond, S., Sitzmann, V., Julca-Aguilar, F., Boyd, S., Wetzstein, G., Heide, F.: Dirty pixels: Towards end-to-end image processing and perception. ACM Transactions on Graphics (SIGGRAPH) (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "type": "text", + "content": "19. Dong, X., Yokoya, N.: Understanding dark scenes by contrasting multi-modal observations. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 840-850 (January 2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "type": "text", + "content": "20. Finlayson, G.D., Trezzi, E.: Shades of gray and colour constancy. In: Color and Imaging Conference. vol. 2004, pp. 37-41. Society for Imaging Science and Technology (2004)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 330, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 481, + 373 + ], + "type": "text", + "content": "21. Goyal, B., Lalonde, J.F., Li, Y., Gupta, M.: Robust scene inference under noise-blur dual corruptions. 2022 IEEE International Conference on Computational Photography (ICCP) pp. 1-12 (2022), https://api_semanticscholar.org/CorpusID:251040807" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "text", + "content": "22. Guo, Y., Luo, F., Wu, X.: Learning degradation-independent representations for camera isp pipelines. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 25774-25783 (June 2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 407, + 481, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 452 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 452 + ], + "type": "text", + "content": "23. Hansen, N., Ostermeier, A.: Adapting arbitrary normal mutation distributions in evolution strategies: the covariance matrix adaptation. In: Proceedings of IEEE International Conference on Evolutionary Computation. pp. 312-317 (1996). https://doi.org/10.1109/ICEC.1996.542381" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 453, + 481, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 481, + 496 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 481, + 496 + ], + "type": "text", + "content": "24. Hasinoff, S.W., Sharlet, D., Geiss, R., Adams, A., Barron, J.T., Kainz, F., Chen, J., Levoy, M.: Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM Transactions on Graphics (Proc. SIGGRAPH Asia) 35(6) (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 497, + 481, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 497, + 481, + 529 + ], + "spans": [ + { + "bbox": [ + 132, + 497, + 481, + 529 + ], + "type": "text", + "content": "25. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 530, + 481, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 530, + 481, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 530, + 481, + 563 + ], + "type": "text", + "content": "26. Heide, F., Steinberger, M., Tsai, Y.T., Rouf, M., Pajak, D., Reddy, D., Gallo, O., Liu, J., Heidrich, W., Egiazarian, K., et al.: Flexisp: A flexible camera image processing framework. ACM Transactions on Graphics (ToG) 33(6), 1-13 (2014)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 564, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 564, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 564, + 481, + 586 + ], + "type": "text", + "content": "27. Hong, Y., Wei, K., Chen, L., Fu, Y.: Crafting object detection in very low light. In: BMVC (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 586, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 481, + 620 + ], + "type": "text", + "content": "28. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 620, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 481, + 643 + ], + "type": "text", + "content": "29. Hu, Y., He, H., Xu, C., Wang, B., Lin, S.: Exposure: A white-box photo post-processing framework. ACM Transactions on Graphics (TOG) 37(2), 26 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "30. Ignatov, A., Van Gool, L., Timofte, R.: Replacing mobile camera isp with a single deep learning model. arXiv preprint arXiv:2002.05509 (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "Cui, Harada." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "31. Iscen, A., Zhang, J., Lazebnik, S., Schmid, C.: Memory-efficient incremental learning through feature adaptation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16. pp. 699-715. Springer (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 161, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 482, + 215 + ], + "type": "text", + "content": "32. Jensen, H.W., Durand, F., Dorsey, J., Stark, M.M., Shirley, P., Premoze, S.: A physically-based night sky model. In: Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques. p. 399-408. SIGGRAPH '01, Association for Computing Machinery, New York, NY, USA (2001). https://doi.org/10.1145/383259.383306, https://doi.org/10.1145/383259.383306" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "33. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "text", + "content": "34. Jin, X., Han, L., Li, Z., Chai, Z., Guo, C., Li, C.: Dnf: Decouple and feedback network for seeing in the dark. In: ICCV (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 304 + ], + "type": "text", + "content": "35. Karaimer, H.C., Brown, M.S.: A software platform for manipulating the camera imaging pipeline. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) Computer Vision - ECCV 2016. pp. 429-444. Springer International Publishing, Cham (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "text", + "content": "36. Ke, L., Ye, M., Danelljan, M., Liu, Y., Tai, Y.W., Tang, C.K., Yu, F.: Segment anything in high quality. In: NeurIPS (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "type": "text", + "content": "37. Kim, W., Kim, G., Lee, J., Lee, S., Baek, S.H., Cho, S.: Paramisp: Learned forward and inverse isps using camera parameters. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 26067-26076 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "type": "text", + "content": "38. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4015-4026 (October 2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "type": "text", + "content": "39. Koskinen, S., Yang, D., Kämäräinen, J.K.: Reverse imaging pipeline for raw rgb image augmentation. In: 2019 IEEE International Conference on Image Processing (ICIP). pp. 2896-2900 (2019). https://doi.org/10.1109/ICIP.2019.8804406" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 435, + 481, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 467 + ], + "type": "text", + "content": "40. Li, Z., Lu, M., Zhang, X., Feng, X., Asif, M.S., Ma, Z.: Efficient visual computing with camera raw snapshots. TPAMI pp. 1-18 (2024). https://doi.org/10.1109/TPAMI.2024.3359326" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "text", + "content": "41. Li, Z., Jiang, H., Cao, M., Zheng, Y.: Polarized color image denoising. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9873-9882 (2023). https://doi.org/10.1109/CVPR52729.2023.00952" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "text", + "content": "42. Liang, Z., Cai, J., Cao, Z., Zhang, L.: Cameranet: A two-stage framework for effective camera isp learning. IEEE Transactions on Image Processing 30, 2248-2262 (2021). https://doi.org/10.1109/TIP.2021.3051486" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "text", + "content": "43. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 567, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 609 + ], + "type": "text", + "content": "44. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Fleet, D., Pajdla, T., Schiele, B., Tuytelaars, T. (eds.) Computer Vision - ECCV 2014. pp. 740-755. Springer International Publishing, Cham (2014)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "text", + "content": "45. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "46. Mildenhall, B., Hedman, P., Martin-Brualla, R., Srinivasan, P.P., Barron, J.T.: Nerf in the dark: High dynamic range view synthesis from noisy raw images. In:" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "type": "text", + "content": "RAW-Adapter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 147, + 116, + 481, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 481, + 139 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 481, + 139 + ], + "type": "text", + "content": "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16190-16199 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 481, + 172 + ], + "type": "text", + "content": "47. Monakhova, K., Richter, S.R., Waller, L., Koltun, V.: Dancing under the stars: Video denoising in starlight. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16241-16251 (June 2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 481, + 215 + ], + "type": "text", + "content": "48. Mosleh, A., Sharma, A., Onzon, E., Mannan, F., Robidoux, N., Heide, F.: Hardware-in-the-loop end-to-end optimization of camera image processing pipelines. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 216, + 481, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 481, + 249 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 481, + 249 + ], + "type": "text", + "content": "49. Nam, S., Kim, S.J.: Modelling the scene dependent imaging in cameras with a deep neural network. 2017 IEEE International Conference on Computer Vision (ICCV) pp. 1726-1734 (2017), https://api.sementicscholar.org/CorpusID:23920472" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 293 + ], + "type": "text", + "content": "50. Nguyen, R.M.H., Brown, M.S.: Raw image reconstruction using a self-contained srgb-jpeg image with only 64 kb overhead. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1655-1663 (2016). https://doi.org/10.1109/CVPR.2016.183" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "text", + "content": "51. Nishimura, J., Gerasimow, T., Sushma, R., Sutic, A., Wu, C.T., Michael, G.: Automatic isp image quality tuning using nonlinear optimization. In: 2018 25th IEEE International Conference on Image Processing (ICIP). pp. 2471-2475. IEEE (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "type": "text", + "content": "52. Omid-Zohoor, A., Ta, D., Murmann, B.: Pascalraw: raw image database for object detection (2014)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 347, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 347, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 133, + 347, + 481, + 390 + ], + "type": "text", + "content": "53. Onzon, E., Mannan, F., Heide, F.: Neural auto-exposure for high-dynamic range object detection. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7706-7716 (2021). https://doi.org/10.1109/CVPR46437.2021.00762" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 391, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 391, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 133, + 391, + 481, + 424 + ], + "type": "text", + "content": "54. Potlapalli, V., Zamir, S.W., Khan, S., Khan, F.: Prompt: Prompting for all-in-one image restoration. In: Thirty-seventh Conference on Neural Information Processing Systems (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 424, + 481, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 424, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 133, + 424, + 481, + 457 + ], + "type": "text", + "content": "55. Qin, H., Han, L., Wang, J., Zhang, C., Li, Y., Li, B., Hu, W.: Attention-aware learning for hyperparameter prediction in image processing pipelines. In: European Conference on Computer Vision. pp. 271-287. Springer (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 457, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 457, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 133, + 457, + 481, + 479 + ], + "type": "text", + "content": "56. Ramanath, R., Snyder, W.E., Yoo, Y., Drew, M.S.: Color image processing pipeline. IEEE Signal processing magazine 22(1), 34-43 (2005)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 479, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 479, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 133, + 479, + 481, + 501 + ], + "type": "text", + "content": "57. Rebuffi, S.A., Bilen, H., Vedaldi, A.: Learning multiple visual domains with residual adapters. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 501, + 481, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 501, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 133, + 501, + 481, + 545 + ], + "type": "text", + "content": "58. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Navab, N., Hornegger, J., Wells, W.M., Frangi, A.F. (eds.) Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015. pp. 234-241. Springer International Publishing, Cham (2015)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 545, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 545, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 133, + 545, + 481, + 588 + ], + "type": "text", + "content": "59. Sayed, M., Brostow, G.: Improved handling of motion blur in online object detection. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1706-1716 (2021). https://doi.org/10.1109/CVPR46437.2021.00175" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 589, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 589, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 133, + 589, + 481, + 621 + ], + "type": "text", + "content": "60. Schwartz, E., Giryes, R., Bronstein, A.M.: Deepisp: Toward learning an end-to-end image processing pipeline. IEEE Transactions on Image Processing 28(2), 912-923 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 621, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 621, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 621, + 481, + 643 + ], + "type": "text", + "content": "61. Sitzmann, V., Martel, J.N., Bergman, A.W., Lindell, D.B., Wetzstein, G.: Implicit neural representations with periodic activation functions. In: Proc. NeurIPS (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 481, + 666 + ], + "type": "text", + "content": "62. Souza, M., Heidrich, W.: MetaISP - Exploiting Global Scene Structure for Accurate Multi-Device Color Rendition. In: Guthe, M., Grosch, T. (eds.) Vision, Modeling," + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "Cui, Harada." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "type": "text", + "content": "and Visualization. The Eurographics Association (2023). https://doi.org/10.2312/vmv.20231236" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 138, + 481, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 138, + 481, + 170 + ], + "spans": [ + { + "bbox": [ + 133, + 138, + 481, + 170 + ], + "type": "text", + "content": "63. Stickland, A.C., Murray, I.: Bert and pals: Projected attention layers for efficient adaptation in multi-task learning. In: International Conference on Machine Learning. pp. 5986-5995. PMLR (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 170, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 170, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 170, + 481, + 203 + ], + "type": "text", + "content": "64. Sun, P., Zhang, R., Jiang, Y., Kong, T., Xu, C., Zhan, W., Tomizuka, M., Li, L., Yuan, Z., Wang, C., Luo, P.: SparseR-CNN: End-to-end object detection with learnable proposals. arXiv preprint arXiv:2011.12450 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 481, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 481, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 481, + 235 + ], + "type": "text", + "content": "65. Sung, Y.L., Cho, J., Bansal, M.: Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5227-5237 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 235, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 235, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 235, + 481, + 267 + ], + "type": "text", + "content": "66. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 267, + 481, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 481, + 310 + ], + "type": "text", + "content": "67. Wang, Y., Yu, Y., Yang, W., Guo, L., Chau, L.P., Kot, A.C., Wen, B.: Raw image reconstruction with learned compact metadata. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18206-18215 (June 2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 310, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 310, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 481, + 342 + ], + "type": "text", + "content": "68. Wei, K., Fu, Y., Yang, J., Huang, H.: A physics-based noise formation model for extreme low-light raw denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 342, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 481, + 396 + ], + "type": "text", + "content": "69. Wu, C.T., Isikdogan, L.F., Rao, S., Nayak, B., Gerasimow, T., Sutic, A., Ainkedem, L., Michael, G.: Visionisp: Repurposing the image signal processor for computer vision applications. In: 2019 IEEE International Conference on Image Processing (ICIP). IEEE (Sep 2019). https://doi.org/10.1109/icip.2019.8803607, http://dx.doi.org/10.1109/ICIP.2019.8803607" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 396, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 481, + 429 + ], + "type": "text", + "content": "70. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. arXiv preprint arXiv:2105.15203 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 429, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 481, + 460 + ], + "type": "text", + "content": "71. Xing, Y., Qian, Z., Chen, Q.: Invertible image signal processing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6287-6296 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 460, + 481, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 481, + 503 + ], + "type": "text", + "content": "72. Xu, R., Chen, C., Peng, J., Li, C., Huang, Y., Song, F., Yan, Y., Xiong, Z.: Toward raw object detection: A new benchmark and a new model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13384-13393 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 503, + 481, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 503, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 503, + 481, + 546 + ], + "type": "text", + "content": "73. Yoshimura, M., Otsuka, J., Irie, A., Ohashi, T.: Dynamically controlled image signal processor for image recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 12866-12876 (October 2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 546, + 481, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 546, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 481, + 590 + ], + "type": "text", + "content": "74. Yoshimura, M., Otsuka, J., Irie, A., Ohashi, T.: Rawgment: Noise-accounted raw augmentation enables recognition in a wide variety of environments. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14007-14017 (June 2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 590, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 481, + 622 + ], + "type": "text", + "content": "75. Yu, K., Li, Z., Peng, Y., Loy, C.C., Gu, J.: Reconfigisp: Reconfigurable camera image processing pipeline. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4248-4257 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "text", + "content": "76. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Cycleisp: Real image restoration via improved data synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2696-2705 (2020)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 388, + 91, + 447, + 102 + ], + "type": "text", + "content": "RAW-Adapter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 303 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "77. Zhang, R., Zhang, W., Fang, R., Gao, P., Li, K., Dai, J., Qiao, Y., Li, H.: Tip-adapter: Training-free adaption of clip for few-shot classification. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 493-510. Springer Nature Switzerland, Cham (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 205 + ], + "type": "text", + "content": "78. Zhang, Y., Zhang, Y., Zhang, Z., Zhang, M., Tian, R., Ding, M.: Isp-teacher: Image signal process with disentanglement regularization for unsupervised domain adaptive dark object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 7387-7395 (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "text", + "content": "79. Zhang, Z., Wang, H., Liu, M., Wang, R., Zuo, W., Zhang, J.: Learning raw-to-srgb mappings with inaccurately aligned supervision. In: ICCV (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 226, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 226, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 226, + 481, + 259 + ], + "type": "text", + "content": "80. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing through ade20k dataset. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 633-641 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 303 + ], + "type": "text", + "content": "81. Zhou, W., Gao, S., Zhang, L., Lou, X.: Histogram of oriented gradients feature extraction from raw bayer pattern images. IEEE Transactions on Circuits and Systems II: Express Briefs 67(5), 946-950 (2020). https://doi.org/10.1109/TCSII.2020.2980557" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "Cui, Harada." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/RCS-Prompt_ Learning Prompt to Rearrange Class Space for Prompt-based Continual Learning/6bf9dca4-8b37-4308-ad56-3bfced51f84f_content_list.json b/2024/RCS-Prompt_ Learning Prompt to Rearrange Class Space for Prompt-based Continual Learning/6bf9dca4-8b37-4308-ad56-3bfced51f84f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d1bfc00d12ada7697ddc168edc344eebd426ffba --- /dev/null +++ b/2024/RCS-Prompt_ Learning Prompt to Rearrange Class Space for Prompt-based Continual Learning/6bf9dca4-8b37-4308-ad56-3bfced51f84f_content_list.json @@ -0,0 +1,2012 @@ +[ + { + "type": "text", + "text": "RCS-Prompt: Learning Prompt to Rearrange Class Space for Prompt-based Continual Learning", + "text_level": 1, + "bbox": [ + 217, + 141, + 787, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Longrong Yang $^{1}$ , Hanbin Zhao $^{1}$ , Yunlong $\\mathrm{Yu}^{2\\star}$ , Xiaodong Zeng $^{3}$ , and Xi Li $^{1,4,5\\star}$", + "bbox": [ + 232, + 210, + 767, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1College of Computer Science & Technology, Zhejiang University", + "bbox": [ + 285, + 238, + 715, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ College of Information Science and Electronic Engineering, Zhejiang University $^{3}$ Ant Services Group $^{4}$ Shanghai Institute for Advanced Study of Zhejiang University", + "bbox": [ + 212, + 252, + 785, + 279 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{5}$ Zhejiang - Singapore Innovation and AI Joint Research Lab, Hangzhou", + "bbox": [ + 256, + 279, + 741, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{longrongyang, zhaohanbin, yuyunlong, xilizju}@zju.edu.cn", + "bbox": [ + 281, + 294, + 720, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{xiaodong.zxd}@antgroup.com", + "bbox": [ + 397, + 309, + 607, + 321 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Prompt-based Continual Learning is an emerging direction in leveraging pre-trained knowledge for downstream continual learning. While arriving at a new session, existing prompt-based continual learning methods usually adapt features from pre-trained models to new data by introducing prompts. However, these prompts lack an optimization objective explicitly modeling inter-session class relationships, thus failing to construct clear inter-session class margins. Moreover, some old samples use new prompts during inference, resulting in the prompt-ambiguity overlap space - a special situation where old and new class spaces overlap. To address these issues, we propose an innovative approach called RCS-Prompt to Rearrange Class Space by bidirectionally optimizing prompts. RCS-Prompt optimizes prompts to signify discriminative regions across different sessions in the class space. Additionally, it mitigates the prompt-ambiguity overlap space by altering the labels of a small subset of new samples to old classes and training them with a customized symmetric loss. The proposed method effectively reduces the overlap between old and new class spaces, thereby establishing clear inter-session class margins. We extensively evaluate RCS-Prompt on public datasets, demonstrating its effectiveness in prompt-based continual learning. Code is available at https://github.com/longrongyang/RCS-Prompt.", + "bbox": [ + 259, + 361, + 743, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Prompt-based Continual Learning $\\cdot$ Continual Learning $\\cdot$ Class Space Overlap $\\cdot$ Class Margin", + "bbox": [ + 259, + 651, + 740, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 707, + 375, + 722 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While deep models have achieved remarkable success across diverse fields [21, 53, 58], most research has focused on recognizing a fixed number of classes in static environments. However, in real-world scenarios, models often encounter streaming data where new classes may emerge over time [20]. To address this issue, Continual Learning has been proposed, which allows models to learn from", + "bbox": [ + 212, + 739, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 217, + 824, + 385, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/566a58d4abd0cbaa8f794e01c37e30db37dc8962cae970976550a16c97dd11fe.jpg", + "image_caption": [ + "(a) Without Prompts", + "Class Space Severe class overlap" + ], + "image_footnote": [], + "bbox": [ + 267, + 183, + 328, + 250 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/795bccccf132810738dbe6cfdad1249efb30aeb81890d0831e25a195fd06f133.jpg", + "image_caption": [ + "Class Space Similar prompts Still class overlap" + ], + "image_footnote": [], + "bbox": [ + 359, + 148, + 441, + 256 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ddc6742f07c5ad555a223a52095c41027896c60e6c027f220bc94747ea402801.jpg", + "image_caption": [ + "(c) Our Method", + "Class Space\nOptimize prompts to\nsolve severe class overlap" + ], + "image_footnote": [], + "bbox": [ + 467, + 172, + 622, + 257 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5671b9373790bd07156ffcc9f527dcd767839e52647ca3d23a8df7e956919494.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 143, + 658, + 152 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Samples in old session", + "bbox": [ + 663, + 143, + 753, + 152 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e9cd74167318525308d2c60e9a1d3c98c9b237678d157c06ac2c494d1fb7e6b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 157, + 658, + 166 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Samples in new session", + "bbox": [ + 666, + 157, + 756, + 166 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1060f61ba3283f745c874b6eb585d76116432be4b895f22435d6622f8390fba1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 172, + 658, + 186 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Margin", + "bbox": [ + 666, + 175, + 697, + 185 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/28b8010d5277cabb582d639aa3627396e46833e4e3cef6413e776ebb1d8bcb0a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 191, + 660, + 204 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Decision boundary", + "bbox": [ + 666, + 194, + 740, + 202 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1b6e490a7ae9ff9b11a2f25795973b0bb46ff6fcd9445b5dd0f222548b20f99b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 212, + 661, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Region signified by old prompts.", + "bbox": [ + 666, + 212, + 730, + 229 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1e2f65c2609d816b9537d019603a3b80bfb2082fadf845abaf11a39e9105ed58.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 234, + 663, + 247 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Region signified by new prompts", + "bbox": [ + 666, + 234, + 730, + 251 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/22eddcde71182843eaa648a75b9d172626620852d8a201de0dee9997981071cb.jpg", + "image_caption": [ + "Fig. 1: Models are incrementally learned across different sessions, where classes are accurately classified within each session. However, old and new class spaces overlap, so class margins lose abilities to distinguish different classes after learning a new session. Prompts signify distinct regions within the class space, yet without explicit inter-session prompt modeling, they fail to separate old and new class spaces. We aim to optimize prompts to create discriminative region and diminish prompt-ambiguity overlap space simultaneously, thereby establishing clear inter-session class margins." + ], + "image_footnote": [], + "bbox": [ + 630, + 258, + 653, + 268 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Old sample using new prompts Prompt-ambiguity overlap space", + "bbox": [ + 666, + 256, + 736, + 292 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "continuously changing data (i.e., different sessions) and dynamically build a unified classifier (i.e., model the relationships among different classes). However, Continual Learning encounters a significant challenge known as catastrophic forgetting [17], which arises when models, after incorporating new class data in a new learning session, tend to forget the previously acquired knowledge.", + "bbox": [ + 212, + 444, + 782, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To mitigate catastrophic forgetting, pre-training has been identified as a pivotal technique [41,43,48,65,81]. A recent innovation in this realm is the adoption of prompt-based methods. These methods freeze a pre-trained model and employ a few learnable prompt parameters to steer representation learning, referred to as Prompt-based Continual Learning methods [55,64,70,71], which exhibit distinct superiority in continual learning performance.", + "bbox": [ + 212, + 521, + 784, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In prompt-based continual learning methods, effective prompt optimization plays a pivotal role in solving the catastrophic forgetting problem, particularly when confronting the overlap between old and new class spaces. Concretely, across continual learning sessions, different classes are isolated, fostering an asynchronous learning process between sessions. Therefore, models tend to overfit the data and create boundaries specific to the classes within each session. This results in significant overlap between the old and new class spaces, as depicted in Fig. 1(a). Consequently, the margins between old classes lose abilities to distinguish between different classes, leading to the catastrophic forgetting problem. As shown in Fig. 1 (b), exiting prompt-based methods [55,64,70,71] use prompts to adjust features in each session. In essence, the learnable prompts create regions in the class space and relocate samples to the regions, thus being crucial for reducing the overlap between old and new class spaces.", + "bbox": [ + 212, + 613, + 785, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, in existing prompt-based continual learning methods [55,64,70,71], the optimization objective of prompts lacks explicit modeling for inter-session", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Longrong Yang et al.", + "bbox": [ + 271, + 114, + 411, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "class relationships. In a new session, prompts are optimized to adjust classification scores solely for classes within that session. Consequently, as shown in Fig. 1 (b), prompts lack the ability to effectively separate old and new class spaces. Additionally, some samples from old classes may use new prompts to adjust features during inference [64, 70]. In such instances, unilaterally optimizing new prompts to move samples away from old classes exacerbates the overlap. This phenomenon, termed the prompt-ambiguity overlap space, makes optimizing prompts to separate class spaces become more challenging. Another observed concern is the diminished efficacy of prompts when their magnitudes become notably smaller than features during optimization.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the above issues, we propose a novel method called Learning Prompt to Rearrange Class Space (RCS-Prompt) for prompt-based continual learning. As depicted in Fig. 1 (c), our main idea is to bidirectionally optimize prompts, making prompts signify discriminative regions in the class space while suppressing the prompt-ambiguity overlap space. Specifically, we firstly construct the Inter-session Bidirectional Prompt Loss. During training, if new prompts focused by new samples exhibit reduced influence on old samples, we explicitly reduce the classification scores of these new samples on old classes for optimizing new prompts. Consequently, these new prompts signify discriminative regions in the class space. Conversely, when new prompts focused by new samples may be attended by old samples, we select a subset of such new samples and relabel them as old classes for learning. Furthermore, the symmetric loss is introduced to prevent the negative impact brought by relabelling. Therefore, the prompt-ambiguity overlap space is reduced. Secondly, we propose the Prompt Magnitude Normalization to align prompt and feature magnitudes before using prompts, ensuring effective prompt tuning.", + "bbox": [ + 212, + 297, + 787, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method can serve as a plug-in to use, enlarging inter-session class margins by minimizing the overlap between old and new class spaces. The contributions are summarized as follows:", + "bbox": [ + 212, + 539, + 787, + 583 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel Inter-session Bidirectional Prompt Loss to optimize prompts, making session prompts signify discriminative regions and reducing the prompt-ambiguity overlap space. Thus, the overlap between old and new class spaces is minimized for constructing clear inter-session class margins.", + "- We additionally present the Prompt Magnitude Normalization, which aligns prompt and feature magnitudes before using prompts, ensuring prompts for effectively adjusting model outcomes.", + "- Without rehearsal, the proposed method achieves significant performance improvements compared to existing prompt-based continual learning methods, establishing state-of-the-art performance." + ], + "bbox": [ + 225, + 592, + 784, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 763, + 395, + 780 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Continual Learning: gradually integrates new concepts while preserving previous knowledge, roughly categorized into three groups. The first employs exemplars from old classes (i.e., rehearsal) to recover past knowledge [1,7,8,38,39,44,", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "RCS-Prompt", + "bbox": [ + 642, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "46, 51, 69, 72–74, 85, 90]. The second adjusts inductive bias through the regularization terms [3, 6, 10–12, 32, 36, 45, 50, 52, 56, 75, 80, 83, 88]. Some methods in this category use knowledge distillation [24] to align old and new model outputs, retaining old knowledge [14, 35, 49, 82]. Lastly, various approaches expand network capacity, including neuron-wise [77, 79], backbone-wise [37, 62, 63, 78, 81, 84, 87], token-wise [15, 67] and others [19, 26, 27, 66]. Some methods incorporate multiple techniques, e.g., [15] expands the network and uses rehearsal simultaneously. Furthermore, works on online continual learning, e.g., [54, 57], discuss the learning process without explicit task boundaries.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Some traditional continual learning methods [25, 27, 75, 86] ever discussed the old and new class space overlap. These methods mainly analyze the overlap problem based on the dynamically updating convolution neural networks, and they alleviate this problem by using old data for knowledge distillation or creating balanced training subsets between old and new sessions. In contrast, our method focuses on prompt-based continual learning, where prompts are special and vital, and it explicitly eliminates the overlap between old and new class spaces by solely using new data to optimizing new prompts.", + "bbox": [ + 212, + 282, + 787, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prompt-based continual learning: gains attraction with the surge of pre-trained models [13,47,60,67,89], using prompts to adapt pre-trained models for sequential data with new classes. L2P [71] applies visual prompt tuning [28] on pre-trained Transformer [13], learning a prompt pool for instance-specific selection. DualPrompt [70] extends L2P with general and expert prompts. S-Prompt [67] constructs only task-specific prompts for each task. CODA-Prompt [55] improves prompt selection via attention. SLCA [81] explores slow learners. DAP [29] generates instance-level prompts for use. LAE [18] reformulates different fin-tuning techniques into the Parameter-Efficient-Tuning module. PromptSRC [31] explores the use of prompts for continually updating CLIP. APG [59] develops a learnable adaptive prompt generator for weak pre-trained models. LGCL [30] introduces language guidance at the task level in the prompt pool. RanPAC [40] exploits training-free random projectors and class-prototype accumulation. Recently, HiDe-Prompt [64] decomposes the sub-optimality into hierarchical components and proposes different losses to solve them.", + "bbox": [ + 212, + 402, + 803, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While few approaches, e.g., HiDe-Prompt [64], suggest enhancing prompt disparity through regularization terms, their losses generally lack class-specificity characteristics, thereby failing to establish a connection between prompt disparity and inter-session class relationships. In contrast, our method optimizes prompts by explicitly adjusting classification logits to minimize overlap between old and new class spaces. Our approach effectively solves the prompt-ambiguity overlap space issue, a crucial concern overlooked by the existing methods.", + "bbox": [ + 212, + 628, + 789, + 737 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 756, + 380, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Overview", + "text_level": 1, + "bbox": [ + 215, + 786, + 341, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Problem setup: This study focuses on prompt-based continual learning. Existing prompt-based continual learning methods [55, 64, 70, 71] typically develop a", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Longrong Yang et al.", + "bbox": [ + 271, + 114, + 413, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7b5c24b05e944e7f8b621e91ad3b73556fcbce5cebd6fe952dc9fcbaae41465d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 143, + 754, + 202 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5028f922c50a1f9e2791ae4090e670ec0684240c6730a7c02329595327789649.jpg", + "image_caption": [ + "Fig. 2: (a) General Pipeline. Existing prompt-based continual learning methods usually need to generate and insert prompts, optimize prompts by the loss, and train a session indicator. During test, the session ID of each sample is inferred and used for select prompts. (b) New samples demand high logits for new classes and low logits for old classes, whereas old samples require low logits for new classes and high logits for old classes. (c) Our Inter-session Bidirectional Prompt Loss satisfies the two requirements simultaneously, reducing the overlap between old and new spaces." + ], + "image_footnote": [], + "bbox": [ + 246, + 203, + 467, + 378 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7f3a05069d3ec7446b37f55ed4732e822f8630d7069453c3f0115be6cb234a6f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 470, + 203, + 754, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": " unified classifier capable of learning from a dynamic data stream. The continual learning process involves a sequence of $B$ learning sessions, denoted as $\\{D^1, D^2, \\dots, D^B\\}$ . Each learning session $D^b = (\\pmb{x}_i^b, y_i^b)_{i=1}^{n_b}$ represents an incremental step with $n_b$ instances. The training instance $\\pmb{x}_i^b$ belongs to class $y_i^b \\in Y^b$ , where $Y^b$ represents the label space of session $b$ . Notably, the label sets $Y^b$ and $Y^{b'}$ are disjoint for $b \\neq b'$ . In this study, $D^b$ is only available at the session $b$ , where old data are not retrieved for rehearsal. Besides, session IDs are unavailable during inference. For simplicity, we refer to samples in the new session as \"new samples\" and those in the old session as \"old samples\". Likewise, we use the abbreviations \"new classes\", \"old classes\", \"new prompts\", and \"old prompts\" for convenience.", + "bbox": [ + 212, + 534, + 787, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Existing prompt-based continual learning methods [55,64,70,71] mainly contain a Pre-Trained Model (PTM) $\\phi(\\cdot)$ and some learnable prompts $\\pmb{pt}$ . As shown in Fig. 2 (a), there are three key components in these methods, which are prompt generation, prompt optimization, and session ID inference, respectively. First, in each session, a set of session-specific prompts $\\pmb{pt} \\in \\mathbb{R}^{L_p \\times D}$ is generated, where $L_p$ and $D$ denote the prompt number and prompt dimensionality, respectively. $\\pmb{pt} = \\{\\pmb{pt}_1, \\dots, \\pmb{pt}_{L_p}\\}$ . They adjust features through cross attention. Concatenating feature tokens $(N \\times D)$ with prompts $(L_p \\times D)$ generates new tokens $((N + L_p) \\times D)$ for attention, i.e., $\\phi(\\pmb{x}_i^b | \\pmb{pt})$ . Then, existing prompt-based continual learning methods use the below classification loss for optimizing prompts:", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "RCS-Prompt", + "bbox": [ + 642, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {z} ^ {b} (\\boldsymbol {x} _ {i} ^ {b}) = [ \\underbrace {- \\inf , \\cdots , - \\inf} _ {\\text {o l d s e s s i o n}}, \\underbrace {z _ {1} ^ {b} , \\cdots} _ {\\text {n e w s e s s i o n}} ],\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 157, + 630, + 191 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {p} ^ {b} \\left(\\boldsymbol {x} _ {i} ^ {b}\\right) = \\operatorname {s o f t m a x} \\left(\\boldsymbol {z} ^ {b} \\left(\\boldsymbol {x} _ {i} ^ {b}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 195, + 785, + 215 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c e} \\left(\\boldsymbol {x} _ {i} ^ {b}\\right) = - \\sum_ {k = 1} ^ {K} q \\left(k \\mid \\boldsymbol {x} _ {i} ^ {b}\\right) \\cdot \\log \\left(p ^ {b} \\left(k \\mid \\boldsymbol {x} _ {i} ^ {b}\\right)\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 218, + 643, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $b$ is the new session ID and $K$ is the class number, $z^b(\\boldsymbol{x}_i^b)$ indicates the logits of $\\boldsymbol{x}_i^b$ and $z_1^b$ indicates the logit of $\\boldsymbol{x}_i^b$ on a new class. In continual learning, unlike standard supervised learning using cross-entropy loss, model predictions for old classes are typically set to -inf to separate them from training. $q(k|\\boldsymbol{x}_i^b)$ represents the one-hot vector of the ground-truth label for sample $\\boldsymbol{x}_i^b$ . Lastly, during training, existing methods usually use features from the PTMs (i.e., PTM features, $\\phi(\\boldsymbol{x}_i^b)$ ) to train a session indicator. During inference, this session indicator is used to estimate the session ID of each sample. Based on the predicted session ID, samples select prompts from all old prompts $\\boldsymbol{p}\\boldsymbol{t}^{old}$ and new prompts $\\boldsymbol{p}\\boldsymbol{t}$ . Under this scenario, as shown in Fig. 2 (b), old samples may select new prompts, leading to the prompt-ambiguity overlap space.", + "bbox": [ + 212, + 263, + 787, + 430 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "RCS-Prompt: Our method focuses on prompt optimization. Specifically, asynchronous learning takes place between sessions in continual learning, resulting in significant overlap between old and new class spaces. In prompt-based continual learning methods, prompts are vital to signify specific regions within the class space, where samples are relocated. Different sets of prompts correspond to distinct regions, and within each set, diverse prompts can reposition samples to various positions within a region. However, existing methods tend to overlook the explicit modeling of inter-session class relationships. Consequently, prompt-defined regions still exhibit overlap, impeding the construction of clear class margins. To this end, we propose an innovative approach called RCS-Prompt to rearrange class space by learning prompts, aiming to minimize the overlap between old and new class spaces, as shown in Fig. 2 (c). In detail, RCS-Prompt comprises two modules: Inter-session Bidirectional Prompt Loss and Prompt Magnitude Normalization. The specifics of each module will be introduced in the following sections.", + "bbox": [ + 212, + 431, + 787, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Inter-session Bidirectional Prompt Loss", + "text_level": 1, + "bbox": [ + 215, + 679, + 588, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As mentioned above, our primary goal is to optimize new prompts during training on the new session, for minimizing the old and new class space overlap. However, achieving this goal is non-trivial due to a special challenge: When using new prompts to establish a region significantly distant from old class spaces and relocating samples to this region, old samples may inadvertently choose some new prompts to adjust features. Consequently, as new prompts shift old samples into the new region, overlap may emerge between old and new class spaces, a phenomenon referred to as the prompt-ambiguity overlap space. This challenge fundamentally stems from the divergence in indicator predictions for samples", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Longrong Yang et al.", + "bbox": [ + 271, + 114, + 411, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "within the same session, suggesting the potential to employ a session indicator for sample separation. Building on this basic idea, as shown in Fig. 2 (c), we propose the Inter-session Bidirectional Prompt Loss to bidirectionally optimize new prompts, for shifting new samples away from old class spaces (i.e., producing discriminative regions) while keeping old samples close to their original class spaces (i.e., depressing the prompt-ambiguity overlap space).", + "bbox": [ + 212, + 146, + 784, + 236 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use a session indicator to split samples, and design different losses to solve them respectively. Specifically, the first step is to generate a session indicator during training. Existing methods [64,70] use the frozen PTM features to train a session indicator, which is always changing during training. The changing session indicator may lead to fluctuating sample divisions, potentially increasing the convergence difficulty. Thus, before each training session, we first use the PTMs to extract the PTM features of all training samples within that session. Then, the PTM features of samples of each class are averaged, severing as the PTM feature prototype proto $_k$ . We store the PTM feature prototype of each class, resulting in a total storage of $K \\times D$ , which is merely the size of a classifier. During optimizing prompts, we compute the indicator score of each sample on all seen classes by:", + "bbox": [ + 212, + 237, + 787, + 417 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\ns _ {k} = \\cos \\left(\\phi \\left(\\boldsymbol {x} _ {i} ^ {b}\\right), p r o t o _ {k}\\right) = \\frac {\\phi \\left(\\boldsymbol {x} _ {i} ^ {b}\\right) ^ {T} \\cdot p r o t o _ {k}}{\\left\\| \\phi \\left(\\boldsymbol {x} _ {i} ^ {b}\\right) \\right\\| \\cdot \\left\\| p r o t o _ {k} \\right\\|}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 422, + 784, + 458 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\cos (\\cdot)$ denotes the cosine similarity. For each training sample, we can determine the class with the highest indicator score and its associated session, serving as its predicted session ID.", + "bbox": [ + 212, + 462, + 784, + 507 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The second step is categorizing new samples into two distinct groups. The first group is labeled as \"ID-clean\" samples, where new samples are predicted to belong to the new session during training. These ID-clean samples possess the PTM features that differ from old samples, indicating a reduced influence of new prompts on old samples during inference. Consequently, it becomes feasible to explicitly decrease the classification logits of these new samples for old classes for optimizing new prompts, as illustrated in Fig. 2 (c). This allows these new prompts to signify discriminative new regions, enabling the effective functioning of class margins. The second group is labeled as \"ID-noisy\" samples, where new samples are predicted to belong to old sessions during training. ID-noisy samples resemble old samples, suggesting that new prompts focused on them may be attended to by old samples during inference. Reducing the classification logits of these new samples for old classes would compel prompts to relocate old samples away from old class spaces, posing a risk of expanding the prompt-ambiguity overlap space. Therefore, we propose to relabel ID-noise samples to the old classes for learning, as depicted in Fig. 2 (c), to position old samples in the proximity of old class spaces during inference. In detail, for an ID-noisy sample $\\boldsymbol{x}_i^b$ , we manually assign its label to the old class with the highest indicator score, creating \"pseudo old labels\" $\\bar{y}$ : $\\bar{y} = \\mathrm{argmax}(s_k)$ , where $s_k$ denotes the fine-grained indicator score of a sample on the class $k$ ( $k \\in \\{1, \\dots, K\\}$ ).", + "bbox": [ + 212, + 507, + 787, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The third step is designing losses for distinct samples. Assigning the labels of new samples to old classes for learning may have a negative impact on new class", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "RCS-Prompt", + "bbox": [ + 642, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "learning. To mitigate the negative effect, we propose to apply a symmetric loss to prevent new samples from excessively drifting toward old class spaces. The symmetric loss, e.g., the reverse cross entropy loss [68], is a family of loss which adaptively weighs clean and noisy sample gradients, preventing noisy samples from dominating the learning. By conducting a Gradient analysis below, we establish a connection between the symmetric loss and the challenge we currently face, revealing the potential to leverage the symmetric loss to prevent the negative impact of using pseudo old labels for new classes.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In summary, in handling ID-clean samples, we maintain their original labels and explicitly decrease their classification logits on old classes. For ID-noisy samples, we relabel them as old classes and supervise them with the reverse cross-entropy loss. Thus, our Inter-session Bidirectional Prompt Loss is formulated as:", + "bbox": [ + 212, + 267, + 787, + 328 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {z} _ {d i s} ^ {b} (\\boldsymbol {x} _ {i} ^ {b}) = [ \\underbrace {z _ {1} ^ {1} , z _ {2} ^ {1} , \\cdots} _ {\\text {o l d s e s s i o n s}}, \\underbrace {z _ {1} ^ {b} , \\cdots} _ {\\text {n e w s e s s i o n}} ],\n$$\n", + "text_format": "latex", + "bbox": [ + 274, + 338, + 506, + 375 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {p} _ {d i s} ^ {b} \\left(\\boldsymbol {x} _ {i} ^ {b}\\right) = \\operatorname {s o f t m a x} \\left(\\boldsymbol {z} _ {d i s} ^ {b} \\left(\\boldsymbol {x} _ {i} ^ {b}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 272, + 377, + 784, + 396 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d i s} (\\boldsymbol {x} _ {i} ^ {b}) = \\left\\{ \\begin{array}{l l} - \\sum_ {k = 1} ^ {K} q (k | \\boldsymbol {x} _ {i} ^ {b}) \\cdot \\log (p _ {d i s} ^ {b} (k | \\boldsymbol {x} _ {i} ^ {b})) & \\text {i f} \\boldsymbol {x} _ {i} ^ {b} \\text {i s I D - c l e a n}, \\\\ - \\sum_ {k = 1} ^ {K} p _ {d i s} ^ {b} (k | \\boldsymbol {x} _ {i} ^ {b}) \\cdot \\log (\\bar {q} (k | \\boldsymbol {x} _ {i} ^ {b})) & \\text {i f} \\boldsymbol {x} _ {i} ^ {b} \\text {i s I D - n o i s y}, \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 272, + 398, + 728, + 438 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\bar{q}(k|\\pmb{x})$ represents the one-hot vector of the pseudo old labels, and ID-noisy samples use the reverse cross-entropy loss. The total loss is formulated as:", + "bbox": [ + 212, + 443, + 787, + 473 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} \\left(\\boldsymbol {x} _ {i} ^ {b}\\right) = \\mathcal {L} _ {\\text {c e}} \\left(\\boldsymbol {x} _ {i} ^ {b}\\right) + \\alpha \\cdot \\mathcal {L} _ {\\text {d i s}} \\left(\\boldsymbol {x} _ {i} ^ {b}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 484, + 784, + 501 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{ce}$ is the standard classification loss and $\\alpha$ is a hyper-parameter.", + "bbox": [ + 212, + 512, + 735, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Gradient analysis: By the gradient analysis, our objective is to clarify the adaptive weighting mechanism of the reverse cross-entropy loss for ID-noisy samples, for preventing the negative impact of pseudo old labels on new classes. The gradients of the reverse cross-entropy loss and the cross-entropy loss have been derived in [68]. For brevity, we denote $p_k$ and $q_k$ as abbreviations for $p(k|\\pmb{x})$ and $q(k|\\pmb{x})$ . The gradient of the reverse cross-entropy loss for ID-noisy samples with respect to the logit $z_k$ can be derived as:", + "bbox": [ + 212, + 527, + 787, + 632 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\mathcal {L} _ {d i s}}{\\partial z _ {k}} = \\left\\{ \\begin{array}{l l} A p _ {k} - A p _ {k} ^ {2} & \\text {i f} q _ {k} = q _ {y} = 1, \\\\ - A p _ {k} p _ {y} & \\text {i f} q _ {k} = 0, \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 643, + 784, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $A$ is a constant. The gradient of the cross-entropy loss $\\mathcal{L}_{ce}$ is indicated as:", + "bbox": [ + 212, + 691, + 787, + 708 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\mathcal {L} _ {c e}}{\\partial z _ {k}} = \\left\\{ \\begin{array}{l l} p _ {k} - 1 & \\text {i f} q _ {k} = q _ {y} = 1, \\\\ p _ {k} & \\text {i f} q _ {k} = 0. \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 382, + 718, + 784, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the case of $q_{k} = q_{y}$ , $Ap_{k} - Ap_{k}^{2}$ serves as an adaptive weighting term based on $p_{k}$ . In detail, when $p_{k}$ is close to 0, the cross-entropy loss generates large gradients to compel samples to increase scores on the labeled class. Conversely, for the reverse cross-entropy loss, \" $p_{k}$ is close to 0\" (an interval approaching 0, e.g., [0.0, 0.5)) indicates that the label is unreliable, resulting in the loss", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Longrong Yang et al.", + "bbox": [ + 271, + 114, + 413, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "decreasing its gradient to 0. When $p_k$ approaches 1 (an interval approaching 1), the reverse cross-entropy loss becomes similar to the cross-entropy loss.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Then, we establish a connection between the gradient and our current scenario. Throughout the training, low scores of samples on pseudo old labels signify their low resemblance to old samples. This scenario corresponds to \" $p_k$ is close to 0\" . In such cases, the cross-entropy loss generates large gradients to force samples to have high logits on pseudo old labels, while the reverse cross-entropy loss can decrease the gradients of these samples to prevent them from excessively drifting toward old class spaces. Conversely, high logits of samples on pseudo old labels signify their proximity to old samples. This scenario corresponds to \" $p_k$ is close to 1\", and the reverse cross-entropy loss becomes similar to the cross-entropy loss, treating these samples as if they were old samples.", + "bbox": [ + 212, + 178, + 787, + 328 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.3 Prompt Magnitude Normalization", + "text_level": 1, + "bbox": [ + 215, + 352, + 542, + 368 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Prompts operate by adjusting features to influence model outputs. We find that when the magnitude of prompts is significantly lower than that of features, their efficacy in altering outputs diminishes. When the magnitude of prompts is significantly higher than that of features, the feature distribution may be severely destroyed. Therefore, in this study, Prompt Magnitude Normalization is proposed to align the magnitude of prompts and features before using prompts, for enhancing the effectiveness of prompt tuning.", + "bbox": [ + 212, + 378, + 784, + 484 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To achieve our goal, we directly normalize prompts. Specifically, we compute the mean $\\mu_{\\pmb{x}}$ and variance $\\tau_{\\pmb{x}}$ of each sample feature, along with the mean $\\mu_{\\pmb{pt}}$ and variance $\\tau_{\\pmb{pt}}$ of its corresponding prompt. Then, we normalize the prompt:", + "bbox": [ + 212, + 486, + 785, + 532 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {p t} = \\frac {\\boldsymbol {p t} - \\mu_ {\\boldsymbol {p t}}}{\\tau_ {\\boldsymbol {p t}}} \\cdot \\tau_ {\\boldsymbol {x}} + \\mu_ {\\boldsymbol {x}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 545, + 785, + 575 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where this normalization ensures that the prompt and feature magnitudes are comparable before using prompts.", + "bbox": [ + 212, + 585, + 785, + 617 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 643, + 375, + 660 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 676, + 455, + 691 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Benchmark: This study uses the benchmark presented in the recent state-of-the-art method HiDe-Prompt [64], which mainly evaluates model performance on Split CIFAR-100 [33] and Split ImageNet-R [22]. Split CIFAR-100 features 100 classes evenly distributed across 10 incremental sessions. Split ImageNet-R includes 200 classes, encompassing challenging ImageNet examples. Similarly, these classes are evenly distributed across 10 incremental sessions. Then, the additional evaluation datasets in [64] include 5-Datasets and Split CUB-200. 5-Datasets [16] includes CIFAR-10 [33], MNIST [34], Fashion-MNIST [76], SVHN [42] and notMNIST [4] datasets, each treated as an incremental task to evaluate the impact of", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "RCS-Prompt", + "bbox": [ + 642, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/91b341e5bb5234d095ffb1db0694785feff74e80c9e20d54ba535bcd738ccf72.jpg", + "table_caption": [ + "Table 1: The comparison (\\%) between our method and five prompt-based continual methods. Following [64], the results of seeds 40, 42, and 44 are averaged to report. HiDe-Prompt uses Multi-Centroid statistical modeling [64]." + ], + "table_footnote": [], + "table_body": "
PTMMethodSplit CIFAR-100Split ImageNet-R
FAA (↑)CAA (↑)FFM (↓)FAA (↑)CAA (↑)FFM (↓)
Sup-21KL2P [71]83.0688.256.5863.6567.257.51
S-Prompt++ [67]88.8192.253.8769.6872.503.29
CODA-Prompt [55]86.9491.574.0470.0374.265.17
DualPrompt [70]86.6090.644.4568.7971.964.49
+RCS-Prompt (Ours)90.5394.214.0870.8074.384.69
HiDe-Prompt [64]92.9294.600.8673.5575.930.95
+RCS-Prompt (Ours)93.3495.340.8274.8577.520.91
DINO-1KL2P [71]70.6579.029.4657.4062.563.58
S-Prompt++ [67]74.9783.827.7857.6463.795.08
CODA-Prompt [55]77.5084.818.1063.1569.736.86
DualPrompt [70]74.9083.984.4558.5764.895.80
+RCS-Prompt (Ours)81.4489.324.2667.4171.735.10
HiDe-Prompt [64]90.0692.922.1369.3472.351.60
+RCS-Prompt (Ours)92.1194.091.9374.5478.022.41
MoCo-1KL2P [71]74.8583.146.5151.6458.872.37
S-Prompt++ [67]76.3083.8814.6753.1560.034.11
CODA-Prompt [55]76.8384.9712.6055.7565.4910.46
DualPrompt [70]77.7785.316.6152.5760.652.73
+RCS-Prompt (Ours)82.9189.374.6361.4165.572.61
HiDe-Prompt [64]90.7093.231.5663.0566.901.78
+RCS-Prompt (Ours)92.0093.651.5568.2471.041.74
", + "bbox": [ + 215, + 198, + 789, + 583 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "large inter-session differences. Split CUB-200 [61] includes 200-class fine-grained images of birds, randomly split into 10 incremental sessions of disjoint classes.", + "bbox": [ + 212, + 611, + 785, + 642 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baselines: On the evaluation benchmark, we extensively compare our method to recent prompt-based continual learning methods: L2P [71], S-Prompt++ [67], DualPrompt [70], CODA-Prompt [55] and HiDe-Prompt [64]. In experiments, we mainly add our method to HiDe-Prompt [64]. Extending upon this baseline, we add our proposed Inter-session Bidirectional Prompt Loss during training and apply the Prompt Magnitude Normalization before using prompts. Others keep unchanging. Besides, we provide results adding our method to DualPrompt [70], for showing the flexibility of our method as a plug-in.", + "bbox": [ + 212, + 643, + 787, + 763 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training details: Following HiDe-Prompt [64], we select ViT-B/16 Sup-21K, DINO-1K [5], and MoCo-1K [9] as the pre-trained models to conduct experiments. We set $\\alpha = 0.4$ . More details are provided in the Sec. 3 of supp.", + "bbox": [ + 212, + 763, + 785, + 809 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation protocol: We record the average accuracy of all seen classes after learning each session, presenting the last one as the final average accuracy", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Longrong Yang et al.", + "bbox": [ + 271, + 114, + 411, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/e4517a4f9089d6871cebfe597743d2d4c68600a5ba65d8d73648bd86ca8f461b.jpg", + "table_caption": [ + "Table 2: The comparison (\\%) between our method and five prompt-based continual learning methods on additional datasets. Following [64], the FAA metrics are report." + ], + "table_footnote": [], + "table_body": "
Method5-DatasetsSplitCUB-200
Sup-21KiBOT-21KSup-21KiBOT-21K
L2P [71]81.8482.2574.4844.29
DualPrompt [70]77.9168.0382.0541.31
S-Prompt++ [67]86.0677.2082.0842.73
CODA-Prompt [55]64.1851.6574.3447.79
HiDe-Prompt [64]93.8394.8886.5678.23
+RCS-Prompt (Ours)94.5196.1187.3279.42
", + "bbox": [ + 253, + 184, + 750, + 320 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "(FAA) and their historical average as the cumulative average accuracy (CAA). We also present the final forgetting measure (FFM) of all sessions. More details are provided in the Sec. 3 of supp.", + "bbox": [ + 212, + 348, + 787, + 395 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Main Results", + "text_level": 1, + "bbox": [ + 214, + 417, + 372, + 431 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Tab. 1 shows the results of our method and five prompt-based continual learning competitors on two datasets with three different pre-trained models. From the results, we observe that our RCS-Prompt performs very competitively.", + "bbox": [ + 212, + 445, + 785, + 491 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Specifically, when employing ViT-B/16 Sup-21K as the pre-trained model and augmenting it with RCS-Prompt, the performance of HiDe-Prompt increases by $0.42\\%$ and $1.30\\%$ FAA on the Split CIFAR-100 and Split ImageNet-R datasets, respectively. When using DINO-1K as the pre-trained model and augmenting it with RCS-Prompt, the performance of HiDe-Prompt increases by $2.05\\%$ and $5.20\\%$ FAA on the Split CIFAR-100 and Split ImageNet-R datasets, respectively. We find that our method exhibits higher performance improvements when the self-supervised pre-trained models are used. This phenomenon suggests that our approach excels in establishing clear class margins, particularly effective in scenarios where self-supervised pre-trained models struggle with defining clear class margins. Furthermore, equipping with RCS-Prompt leads to significant performance improvements in DualPrompt, which proves the flexibility our method as a plug-in.", + "bbox": [ + 212, + 491, + 787, + 686 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Additional datasets: Furthermore, we evaluate our method on 5-Datasets and Split CUB-200 in Tab. 2. On the 5-Datasets, when integrated with RCS-Prompt, the performance of HiDe-Prompt increases $0.68\\%$ and $1.23\\%$ with different PTMs, respectively. Then, on the Split CUB-200 dataset, employing RCS-Prompt leads to a notable improvement in the performance of HiDe-Prompt, with gains of $0.76\\%$ and $1.19\\%$ with different PTMs, respectively.", + "bbox": [ + 212, + 688, + 787, + 779 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Additional experimental outcomes involving more datasets (e.g., ImageNet-A [23] and ObjectNet [2]) and pre-trained models are detailed in the Sec. 4 of supp. The comparison between our method with DAP [29], LAE [18], APG [59] and RanPAC [40] on their benchmark is also provided.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "RCS-Prompt", + "bbox": [ + 640, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 785, + 127 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/d7728f4a9c2f7a585a46c0390521215e450c3660e4e16a3dbef79ebb27d39d9f.jpg", + "table_caption": [ + "Table 3: Component ablation study $(\\%)$ . Baseline indicate HiDe-Prompt [64] using DINO-1K and the seed 42. $\\mathcal{L}_{dis}$ means adding the Inter-session Bidirectional Prompt Loss. Normalization means adding the Prompt Magnitude Normalization." + ], + "table_footnote": [], + "table_body": "
BaselineL_disNormalizationFAACAA
✓71.6075.61
✓✓74.9278.20
✓✓71.9375.83
✓✓✓75.1278.35
", + "bbox": [ + 325, + 198, + 678, + 292 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Further Analysis", + "text_level": 1, + "bbox": [ + 215, + 328, + 401, + 343 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conduct all ablation studies on challenging Split ImageNet-R dataset [22] with DINO-1K using the seed as 42 (the widely used seed in [55,70,71]). More discussion about the continual learning performance trend, technical selections and visualization analyses are detailed in the Sec. 5 of supp.", + "bbox": [ + 212, + 375, + 782, + 434 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Component ablation: We conduct experiments to evaluate the impacts of the proposed Inter-session Bidirectional Prompt Loss and Prompt Magnitude Normalization. Our experiments involve: (i) Baseline: we establish HiDePrompt [64] as our baseline. (ii) Baseline $+\\mathcal{L}_{dis}$ : we add the Inter-session Bidirectional Prompt Loss during training. (iii) Baseline $+\\mathrm{Normalization}$ : we add the Prompt Magnitude Normalization before using prompts. (iv) Baseline $+\\mathcal{L}_{dis} + \\mathrm{Normalization}$ : two modules are used simultaneously. From the results in Tab. 3, we reach the following observations: (i) The primary contribution of this paper, the proposed Inter-session Bidirectional Prompt Loss $\\mathcal{L}_{dis}$ , yields a considerably $3.32\\%$ performance increase over the baseline. (ii) The Prompt Magnitude Normalization improves model performance by $0.33\\%$ . (iii) Leveraging both the Inter-session Bidirectional Prompt Loss and the Prompt Magnitude Normalization achieves a model performance of $75.12\\%$ .", + "bbox": [ + 212, + 438, + 787, + 635 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{dis}$ vs. contrastive regularization: This study aims to verify the superiority of $\\mathcal{L}_{dis}$ than Contrastive Regularization [64]. The Contrastive Regularization (CR) is designed to enhance the session-specific prompt differences, having the potential to reduce the overlap between old and new class spaces. In Tab. 4, we observe that our proposed loss significantly outperforms the contrastive regularization. This superiority stems from the fact that the contrastive loss is class-agnostic, not establishing a connection between prompt discrimination and class relationships. Besides, the contrastive regularization fails to alleviate the occurrence of the prompt-ambiguity overlap space.", + "bbox": [ + 212, + 638, + 787, + 776 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Loss design: This discussion aims to prove that bidirectional prompt optimization is vital for improving the model performance, and the symmetric loss works. For constructing $\\mathcal{L}_{dis}$ , we alter the labels of ID-noisy samples to pseudo old labels and design the symmetric loss to optimize prompts. In this discussion,", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Longrong Yang et al.", + "bbox": [ + 271, + 114, + 411, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/45e2a0afdfa8de62b93981559459224d88dfa41561674653415b89e8281d7dee.jpg", + "table_caption": [ + "Table 4: The comparison (\\%) between different losses for verifying the superiority of our proposed loss than simply enlarging the inter-session prompt difference. CR is a regularization term used in [64] to enlarge the inter-session prompt difference. In all experiments, the Prompt Magnitude Normalization is used." + ], + "table_footnote": [], + "table_body": "
Baseline w/o CR+CR+Ldis+CR & Ldis
FAA71.4771.9374.9075.12
CAA75.7375.8378.2978.35
", + "bbox": [ + 254, + 210, + 750, + 271 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/5228bbb3eef751765eefdf03f00b674da3b36b8b6efcca6e6572d020ddfedc30.jpg", + "table_caption": [ + "Table 5: The comparison (\\%) between different loss designs for verifying the necessity of the bidirectional prompt optimization. \"w/o old\" means not using pseudo old labels for ID-noisy samples and \"w old\" means using pseudo old labels for ID-noisy samples. In all experiments, the Prompt Magnitude Normalization is used." + ], + "table_footnote": [], + "table_body": "
Baseline+Lce dis+Ldis
w/o oldw oldw old
FAA71.9371.5073.0775.12
CAA75.8376.0576.8178.35
", + "bbox": [ + 297, + 349, + 705, + 425 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "we formulate the cross-entropy loss concerning both old and new classes as:", + "bbox": [ + 212, + 450, + 759, + 465 + ], + "page_idx": 12 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d i s} ^ {c e} \\left(\\boldsymbol {x} _ {i} ^ {b}\\right) = - \\sum_ {k = 1} ^ {K} q \\left(k \\mid \\boldsymbol {x} _ {i} ^ {b}\\right) \\cdot \\log \\left(p _ {d i s} ^ {b} \\left(k \\mid \\boldsymbol {x} _ {i} ^ {b}\\right)\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 473, + 785, + 516 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "where old class logits are not set to -inf. The analysis from Tab. 5 yields several key observations: (i) After adding $\\mathcal{L}_{dis}^{ce}$ but not using pseudo old labels, the model performance reduces from 71.93 to 71.50. This suggests that undirectionally enhancing inter-session prompt discrimination is ineffective, implying a potential enlargement of the prompt-ambiguity overlap space. (ii) Compared to $+ \\mathcal{L}_{dis}^{ce}$ without old\", $+ \\mathcal{L}_{dis}^{ce}$ with old\" results in a $1.57\\%$ higher model performance. Thus, bidirectionally optimizing prompts is vital for diminishing the class overlap. (iii) After using pseudo old labels, than adding $\\mathcal{L}_{dis}^{ce}$ , adding $\\mathcal{L}_{dis}$ improves the model performance from 73.07 to 75.12. This indicates that the symmetric loss can mitigate the negative impact brought by using pseudo old labels.", + "bbox": [ + 212, + 522, + 787, + 672 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Hyper-parameter analysis: We investigate the effects of the hyperparameter $\\alpha$ . More specifically, we explore a range of values for $\\alpha$ , namely $\\{0.0, 0.2, 0.4, 0.6, 1.0\\}$ . In the results presented in Tab. 6, it is evident that as the value of $\\alpha$ increases, there is a corresponding gradual improvement in performance. This improvement plateaus when $\\alpha$ attains a value of 0.4.", + "bbox": [ + 212, + 674, + 787, + 748 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Statistical analysis: This statistic aims to use inter-session misclassification to quantify the old and new class space overlap. In this analysis, without losing the generality, we focus on the predictions for test samples from session 0 (the first session) after learning all sessions. The total test sample count is 650 in session 0. For the predictions from the baseline (HiDe-Prompt), we compute the count of samples that are misclassified as classes from other sessions (session 0),", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "RCS-Prompt", + "bbox": [ + 640, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/39c5df419df7ab09e385d36d4edc57aed597ae513f36439d2f8b7677fbb1bb19.jpg", + "table_caption": [ + "Table 6: Impacts (%) of $\\alpha$ on Split ImageNet-R." + ], + "table_footnote": [], + "table_body": "
0.00.20.40.61.0
FAA71.9374.3375.1274.1273.58
CAA75.8377.8578.3577.6377.44
", + "bbox": [ + 336, + 170, + 666, + 229 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/1ffa2e08b20051eadf540f92a9dd523c8a07ef7d79c68a93ff9c8acebde2fbf6.jpg", + "image_caption": [ + "Fig. 3: The analysis about misclassification. (a) We analyze which type of misclassification is the main type for the baseline. Inter-session misclassification means that a sample is misclassified as a class from other sessions. (b) We quantify the solving of our method for the inter-session misclassification." + ], + "image_footnote": [], + "bbox": [ + 246, + 243, + 411, + 383 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/9aab9bb8f19d08373154688ce97f6b27ccf1e5b4ff25be814ba6994559ce6fa9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 243, + 754, + 383 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "termed as inter-session (intra-session) misclassification type, as shown in Fig. 3 (a). Then, as shown in Fig. 3 (b), focusing on inter-session misclassification, we compute the count of samples that are misclassified in each session to compare the baseline and our method. Fig. 3 reveals two key observations: $(i)$ For the baseline, 173 $(26.62\\%)$ test samples are misclassified as classes in another session, and only 31 $(4.77\\%)$ are misclassified as classes from session 0. This shows that the inter-session misclassification is the main problem for existing methods. $(ii)$ Than the baseline, our method significantly decreases the inter-session misclassification sample count from 173 to 67. This shows that our method can effectively solve the inter-session misclassification problem.", + "bbox": [ + 212, + 494, + 787, + 647 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 670, + 359, + 686 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our study reveals that the optimization objective of prompts lacks the explicit modeling of inter-session class relationships, not effectively separating new and old class spaces. To minimize the overlap between new and old class spaces for creating the clear inter-session class margins, we introduce an Inter-session Bidirectional Prompt Loss to place new and old samples to discriminative regions, while eliminating the prompt-ambiguity overlap space. Furthermore, we design the Prompt Magnitude Normalization to enhance the prompt tuning effectiveness. Extensive experiments demonstrate the superiority of our approach across different datasets and continual learning scenarios.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Longrong Yang et al.", + "bbox": [ + 271, + 114, + 411, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. This work is supported in part by National Natural Science Foundation of China under Grant No.62441602, National Science Foundation for Distinguished Young Scholars under Grant 62225605, Project 12326608 supported by NSFC, National Natural Science Foundation of China under Grant U20A20222, Zhejiang Key Research and Development Program under Grant 2023C03196, the Ningbo Science and Technology Innovation Project (No.2024Z294), the Fundamental Research Funds for the Central Universities, the Key R&D Program of Zhejiang Province, China 2023C01043, Ant Group through CCF-Ant Research Fund, and sponsored by CCF-AFSG Research Fund.", + "bbox": [ + 212, + 146, + 807, + 282 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 306, + 325, + 321 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aljundi, R., Lin, M., Goujaud, B., Bengio, Y.: Gradient based sample selection for online continual learning. Advances in neural information processing systems 32 (2019)", + "2. Barbu, A., Mayo, D., Alverio, J., Luo, W., Wang, C., Gutfreund, D., Tenenbaum, J., Katz, B.: Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems 32 (2019)", + "3. Belouadah, E., Popescu, A.: II2m: Class incremental learning with dual memory. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 583-592 (2019)", + "4. Bulatov, Y.: Notmnist dataset (2011)", + "5. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)", + "6. Cha, S., Cho, S., Hwang, D., Hong, S., Lee, M., Moon, T.: Rebalancing batch normalization for exemplar-based class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20127-20136 (2023)", + "7. Chaudhry, A., Dokania, P.K., Ajanthan, T., Torr, P.H.: Riemannian walk for incremental learning: Understanding forgetting and intransigence. In: Proceedings of the European conference on computer vision (ECCV). pp. 532-547 (2018)", + "8. Chaudhry, A., Ranzato, M., Rohrbach, M., Elhoseiny, M.: Efficient lifelong learning with a-gem. In: International Conference on Learning Representations (2018)", + "9. Chen, X., Xie, S., He, K.: An empirical study of training self-supervised vision transformers. in 2021 IEEE. In: CVF International Conference on Computer Vision (ICCV). pp. 9620-9629", + "10. Dong, J., Li, H., Cong, Y., Sun, G., Zhang, Y., Van Gool, L.: No one left behind: Real-world federated class-incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence 46(4), 2054-2070 (2024). https://doi.org/10.1109/TPAMI.2023.3334213", + "11. Dong, J., Liang, W., Cong, Y., Sun, G.: Heterogeneous forgetting compensation for class-incremental learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 11742-11751 (October 2023)", + "12. Dong, J., Wang, L., Fang, Z., Sun, G., Xu, S., Wang, X., Zhu, Q.: Federated class-incremental learning. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2022)" + ], + "bbox": [ + 217, + 338, + 785, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "RCS-Prompt", + "bbox": [ + 640, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020)", + "14. Douillard, A., Cord, M., Ollion, C., Robert, T., Valle, E.: Podnet: Pooled outputs distillation for small-tasks incremental learning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16. pp. 86-102. Springer (2020)", + "15. Douillard, A., Ramé, A., Couairon, G., Cord, M.: Dytox: Transformers for continual learning with dynamic token expansion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9285-9295 (2022)", + "16. Ebrahimi, S., Meier, F., Calandra, R., Darrell, T., Rohrbach, M.: Adversarial continual learning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16. pp. 386-402. Springer (2020)", + "17. French, R.M.: Catastrophic forgetting in connectionist networks. Trends in cognitive sciences 3(4), 128-135 (1999)", + "18. Gao, Q., Zhao, C., Sun, Y., Xi, T., Zhang, G., Ghanem, B., Zhang, J.: A unified continual learning framework with general parameter-efficient tuning. arXiv preprint arXiv:2303.10070 (2023)", + "19. Ge, Y., Li, Y., Ni, S., Zhao, J., Yang, M.H., Itti, L.: Clr: Channel-wise lightweight reprogramming for continual learning. arXiv preprint arXiv:2307.11386 (2023)", + "20. Gomes, H.M., Barddal, J.P., Enembreck, F., Bifet, A.: A survey on ensemble learning for data stream classification. ACM Computing Surveys (CSUR) 50(2), 1-36 (2017)", + "21. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "22. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8340-8349 (2021)", + "23. Hendrycks, D., Zhao, K., Basart, S., Steinhardt, J., Song, D.: Natural adversarial examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15262-15271 (2021)", + "24. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)", + "25. Hou, S., Pan, X., Loy, C.C., Wang, Z., Lin, D.: Learning a unified classifier incrementally via rebalancing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 831-839 (2019)", + "26. Hu, Z., Li, Y., Lyu, J., Gao, D., Vasconcelos, N.: Dense network expansion for class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11858-11867 (2023)", + "27. Huang, B., Chen, Z., Zhou, P., Chen, J., Wu, Z.: Resolving task confusion in dynamic expansion architectures for class incremental learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 908-916 (2023)", + "28. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII. pp. 709-727. Springer (2022)" + ], + "bbox": [ + 217, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Longrong Yang et al.", + "bbox": [ + 271, + 114, + 411, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "29. Jung, D., Han, D., Bang, J., Song, H.: Generating instance-level prompts for rehearsal-free continual learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11847-11857 (2023)", + "30. Khan, M.G.Z.A., Naeem, M.F., Van Gool, L., Stricker, D., Tombari, F., Afzal, M.Z.: Introducing language guidance in prompt-based continual learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11463-11473 (2023)", + "31. Khattak, M.U., Wasim, S.T., Naseer, M., Khan, S., Yang, M.H., Khan, F.S.: Self-regulating prompts: Foundational model adaptation without forgetting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15190-15200 (2023)", + "32. Kim, D., Han, B.: On the stability-plasticity dilemma of class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20196-20204 (2023)", + "33. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)", + "34. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (1998)", + "35. Li, Z., Hoiem, D.: Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence 40(12), 2935-2947 (2017)", + "36. Liu, Y., Li, Y., Schiele, B., Sun, Q.: Online hyperparameter optimization for class-incremental learning. Proceedings of the AAAI Conference on Artificial Intelligence (2023)", + "37. Liu, Y., Schiele, B., Sun, Q.: Adaptive aggregation networks for class-incremental learning. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 2544-2553 (2021)", + "38. Liu, Y., Su, Y., Liu, A.A., Schiele, B., Sun, Q.: Mnemonics training: Multi-class incremental learning without forgetting. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 12245-12254 (2020)", + "39. Luo, Z., Liu, Y., Schiele, B., Sun, Q.: Class-incremental exemplar compression for class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11371-11380 (2023)", + "40. McDonnell, M.D., Gong, D., Parvaneh, A., Abbasnejad, E., van den Hengel, A.: Ranpac: Random projections and pre-trained models for continual learning. Advances in Neural Information Processing Systems 36 (2024)", + "41. Mehta, S.V., Patil, D., Chandar, S., Strubell, E.: An empirical investigation of the role of pre-training in lifelong learning. Journal of Machine Learning Research 24(214), 1-50 (2023)", + "42. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011)", + "43. Panos, A., Kobe, Y., Reino, D.O., Aljundi, R., Turner, R.E.: First session adaptation: A strong replay-free baseline for class-incremental learning. arXiv preprint arXiv:2303.13199 (2023)", + "44. Petit, G., Popescu, A., Schindler, H., Picard, D., Delezoide, B.: Fetril: Feature translation for exemplar-free class-incremental learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3911-3920 (2023)", + "45. Pham, Q., Liu, C., Hoi, S.: Continual normalization: Rethinking batch normalization for online continual learning. In: International Conference on Learning Representations (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "RCS-Prompt", + "bbox": [ + 642, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "46. Qiang, S., Hou, J., Wan, J., Liang, Y., Lei, Z., Zhang, D.: Mixture uniform distribution modeling and asymmetric mix distillation for class incremental learning. In: Proceedings of the AAAI Conference on Artificial Intelligence (2023)", + "47. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "48. Ramasesh, V.V., Lewkowycz, A., Dyer, E.: Effect of scale on catastrophic forgetting in neural networks. In: International Conference on Learning Representations (2021)", + "49. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 2001-2010 (2017)", + "50. Rymarczyk, D., van de Weijer, J., Zielinski, B., Twardowski, B.: Icicle: Interpretable class incremental continual learning. arXiv preprint arXiv:2303.07811 (2023)", + "51. Sarfraz, F., Arani, E., Zonooz, B.: Sparse coding in a dual memory system for lifelong learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 9714-9722 (2023)", + "52. Shi, Y., Zhou, K., Liang, J., Jiang, Z., Feng, J., Torr, P.H., Bai, S., Tan, V.Y.: Mimicking the oracle: an initial phase decorrelation approach for class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16722-16731 (2022)", + "53. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)", + "54. Singh, P., Li, Y., Sikarwar, A., Lei, S.W., Gao, D., Talbot, M.B., Sun, Y., Shou, M.Z., Kreiman, G., Zhang, M.: Learning to learn: How to continuously teach humans and machines. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11708-11719 (2023)", + "55. Smith, J.S., Karlinsky, L., Gutta, V., Cascante-Bonilla, P., Kim, D., Arbelle, A., Panda, R., Feris, R., Kira, Z.: Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11909-11919 (2023)", + "56. Sun, Z., Mu, Y., Hua, G.: Regularizing second-order influences for continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20166-20175 (2023)", + "57. Talbot, M.B., Zawar, R., Badkundri, R., Zhang, M., Kreiman, G.: Tuned compositional feature replays for efficient stream learning. IEEE Transactions on Neural Networks and Learning Systems (2023)", + "58. Tan, M., Pang, R., Le, Q.V.: Efficientdet: Scalable and efficient object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10781-10790 (2020)", + "59. Tang, Y.M., Peng, Y.X., Zheng, W.S.: When prompt-based incremental learning does not meet strong pretraining. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1706-1716 (2023)", + "60. Villa, A., Alcázar, J.L., Alfarra, M., Alhamoud, K., Hurtado, J., Heilbron, F.C., Soto, A., Ghanem, B.: Pivot: Prompting for video continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24214-24223 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Longrong Yang et al.", + "bbox": [ + 271, + 114, + 411, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "61. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011)", + "62. Wang, F.Y., Zhou, D.W., Liu, L., Ye, H.J., Bian, Y., Zhan, D.C., Zhao, P.: Beef: Bi-compatible class-incremental learning via energy-based expansion and fusion. In: The Eleventh International Conference on Learning Representations (2023)", + "63. Wang, F.Y., Zhou, D.W., Ye, H.J., Zhan, D.C.: Foster: Feature boosting and compression for class-incremental learning. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXV. pp. 398-414. Springer (2022)", + "64. Wang, L., Xie, J., Zhang, X., Huang, M., Su, H., Zhu, J.: Hierarchical decomposition of prompt-based continual learning: Rethinking obscured sub-optimality. In: Advances in neural information processing systems (2023)", + "65. Wang, L., Zhang, X., Su, H., Zhu, J.: A comprehensive survey of continual learning: Theory, method and application. arXiv preprint arXiv:2302.00487 (2023)", + "66. Wang, W., Hu, Y., Chen, Q., Zhang, Y.: Task difficulty aware parameter allocation & regularization for lifelong learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7776-7785 (2023)", + "67. Wang, Y., Huang, Z., Hong, X.: S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. Annual Conference on Neural Information Processing Systems (2022)", + "68. Wang, Y., Ma, X., Chen, Z., Luo, Y., Yi, J., Bailey, J.: Symmetric cross entropy for robust learning with noisy labels. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 322-330 (2019)", + "69. Wang, Z., Zhao, Z., Wu, Y., Zhou, L., Xu, D.: Progressive target-styled feature augmentation for unsupervised domain adaptation on point clouds. In: ECCV (2024)", + "70. Wang, Z., Zhang, Z., Ebrahimi, S., Sun, R., Zhang, H., Lee, C.Y., Ren, X., Su, G., Perot, V., Dy, J., et al.: Dualprompt: Complementary prompting for rehearsal-free continual learning. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVI. pp. 631-648. Springer (2022)", + "71. Wang, Z., Zhang, Z., Lee, C.Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., Pfister, T.: Learning to prompt for continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 139-149 (2022)", + "72. Wu, X., Wang, H., Wu, Y., Li, X.: D3t-gan: Data-dependent domain transfer gans for image generation with limited data. ACM Trans. Multimedia Comput. Commun. Appl. 19(4) (mar 2023). https://doi.org/10.1145/3576858, https://doi.org/10.1145/3576858", + "73. Wu, Y., Li, H., Wang, F., Zhang, Y., Liang, R.: Self-distilled dynamic fusion network for language-based fashion retrieval. In: ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 3260-3264. IEEE (2024)", + "74. Wu, Y., Wu, X., Li, X., Tian, J.: Mgh: Metadata guided hypergraph modeling for unsupervised person re-identification. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 1571–1580 (2021)", + "75. Wu, Y., Chen, Y., Wang, L., Ye, Y., Liu, Z., Guo, Y., Fu, Y.: Large scale incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 374-382 (2019)", + "76. Xiao, H., Rasul, K., Vollgraf, R.: Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747 (2017)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "RCS-Prompt", + "bbox": [ + 642, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "77. Xu, J., Zhu, Z.: Reinforced continual learning. Advances in Neural Information Processing Systems 31 (2018)", + "78. Yan, S., Xie, J., He, X.: Der: Dynamically expandable representation for class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3014-3023 (2021)", + "79. Yoon, J., Yang, E., Lee, J., Hwang, S.J.: Lifelong learning with dynamically expandable networks. In: International Conference on Learning Representations (2018)", + "80. Yu, L., Twardowski, B., Liu, X., Herranz, L., Wang, K., Cheng, Y., Jui, S., Weijer, J.v.d.: Semantic drift compensation for class-incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6982-6991 (2020)", + "81. Zhang, G., Wang, L., Kang, G., Chen, L., Wei, Y.: Slow learner with classifier alignment for continual learning on a pre-trained model. arXiv preprint arXiv:2303.05118 (2023)", + "82. Zhang, J., Zhang, J., Ghosh, S., Li, D., Tasci, S., Heck, L., Zhang, H., Kuo, C.C.J.: Class-incremental learning via deep model consolidation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1131-1140 (2020)", + "83. Zhao, B., Xiao, X., Gan, G., Zhang, B., Xia, S.T.: Maintaining discrimination and fairness in class incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13208-13217 (2020)", + "84. Zhao, H., Fu, Y., Kang, M., Tian, Q., Wu, F., Li, X.: Mgmt: Multi-grained slow vs. fast framework for few-shot class-incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021)", + "85. Zhao, H., Wang, H., Fu, Y., Wu, F., Li, X.: Memory-efficient class-incremental learning for image classification. IEEE Transactions on Neural Networks and Learning Systems 33(10), 5966-5977 (2021)", + "86. Zhou, D.W., Wang, F.Y., Ye, H.J., Ma, L., Pu, S., Zhan, D.C.: Forward compatible few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9046-9056 (2022)", + "87. Zhou, D.W., Wang, Q.W., Ye, H.J., Zhan, D.C.: A model or 603 exemplars: Towards memory-efficient class-incremental learning. In: International Conference on Learning Representations (2023)", + "88. Zhou, D.W., Ye, H.J., Ma, L., Xie, D., Pu, S., Zhan, D.C.: Few-shot class incremental learning by sampling multi-phase tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)", + "89. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)", + "90. Zhuang, H., Weng, Z., Wei, H., Xie, R., Toh, K.A., Lin, Z.: Acil: Analytic class incremental learning with absolute memorization and privacy protection. Advances in Neural Information Processing Systems 35, 11602-11614 (2022)" + ], + "bbox": [ + 215, + 147, + 784, + 729 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "Longrong Yang et al.", + "bbox": [ + 271, + 114, + 411, + 128 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/2024/RCS-Prompt_ Learning Prompt to Rearrange Class Space for Prompt-based Continual Learning/6bf9dca4-8b37-4308-ad56-3bfced51f84f_model.json b/2024/RCS-Prompt_ Learning Prompt to Rearrange Class Space for Prompt-based Continual Learning/6bf9dca4-8b37-4308-ad56-3bfced51f84f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b362a721564e1249dd9b118cfb088448454dbf8f --- /dev/null +++ b/2024/RCS-Prompt_ Learning Prompt to Rearrange Class Space for Prompt-based Continual Learning/6bf9dca4-8b37-4308-ad56-3bfced51f84f_model.json @@ -0,0 +1,3023 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.218, + 0.142, + 0.788, + 0.187 + ], + "angle": 0, + "content": "RCS-Prompt: Learning Prompt to Rearrange Class Space for Prompt-based Continual Learning" + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.212, + 0.768, + 0.227 + ], + "angle": 0, + "content": "Longrong Yang\\(^{1}\\), Hanbin Zhao\\(^{1}\\), Yunlong \\(\\mathrm{Yu}^{2\\star}\\), Xiaodong Zeng\\(^{3}\\), and Xi Li\\(^{1,4,5\\star}\\)" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.239, + 0.717, + 0.253 + ], + "angle": 0, + "content": "1College of Computer Science & Technology, Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.786, + 0.28 + ], + "angle": 0, + "content": "\\(^{2}\\)College of Information Science and Electronic Engineering, Zhejiang University \\(^{3}\\)Ant Services Group \\(^{4}\\)Shanghai Institute for Advanced Study of Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.28, + 0.742, + 0.295 + ], + "angle": 0, + "content": "\\(^{5}\\)Zhejiang - Singapore Innovation and AI Joint Research Lab, Hangzhou" + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.295, + 0.722, + 0.309 + ], + "angle": 0, + "content": "{longrongyang, zhaohanbin, yuyunlong, xilizju}@zju.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.398, + 0.31, + 0.609, + 0.323 + ], + "angle": 0, + "content": "{xiaodong.zxd}@antgroup.com" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.362, + 0.744, + 0.64 + ], + "angle": 0, + "content": "Abstract. Prompt-based Continual Learning is an emerging direction in leveraging pre-trained knowledge for downstream continual learning. While arriving at a new session, existing prompt-based continual learning methods usually adapt features from pre-trained models to new data by introducing prompts. However, these prompts lack an optimization objective explicitly modeling inter-session class relationships, thus failing to construct clear inter-session class margins. Moreover, some old samples use new prompts during inference, resulting in the prompt-ambiguity overlap space - a special situation where old and new class spaces overlap. To address these issues, we propose an innovative approach called RCS-Prompt to Rearrange Class Space by bidirectionally optimizing prompts. RCS-Prompt optimizes prompts to signify discriminative regions across different sessions in the class space. Additionally, it mitigates the prompt-ambiguity overlap space by altering the labels of a small subset of new samples to old classes and training them with a customized symmetric loss. The proposed method effectively reduces the overlap between old and new class spaces, thereby establishing clear inter-session class margins. We extensively evaluate RCS-Prompt on public datasets, demonstrating its effectiveness in prompt-based continual learning. Code is available at https://github.com/longrongyang/RCS-Prompt." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.652, + 0.741, + 0.681 + ], + "angle": 0, + "content": "Keywords: Prompt-based Continual Learning \\(\\cdot\\) Continual Learning \\(\\cdot\\) Class Space Overlap \\(\\cdot\\) Class Margin" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.708, + 0.377, + 0.723 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.74, + 0.788, + 0.817 + ], + "angle": 0, + "content": "While deep models have achieved remarkable success across diverse fields [21, 53, 58], most research has focused on recognizing a fixed number of classes in static environments. However, in real-world scenarios, models often encounter streaming data where new classes may emerge over time [20]. To address this issue, Continual Learning has been proposed, which allows models to learn from" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.386, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.413, + 0.129 + ], + "angle": 0, + "content": "Longrong Yang et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.149, + 0.326, + 0.173 + ], + "angle": 0, + "content": "(a) Without Prompts" + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.184, + 0.329, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.256, + 0.258, + 0.337, + 0.282 + ], + "angle": 0, + "content": "Class Space Severe class overlap" + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.149, + 0.442, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.362, + 0.258, + 0.435, + 0.288 + ], + "angle": 0, + "content": "Class Space Similar prompts Still class overlap" + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.151, + 0.58, + 0.161 + ], + "angle": 0, + "content": "(c) Our Method" + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.173, + 0.624, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.492, + 0.259, + 0.592, + 0.289 + ], + "angle": 0, + "content": "Class Space\nOptimize prompts to\nsolve severe class overlap" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.145, + 0.66, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.664, + 0.145, + 0.754, + 0.153 + ], + "angle": 0, + "content": "Samples in old session" + }, + { + "type": "image", + "bbox": [ + 0.634, + 0.159, + 0.66, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.667, + 0.159, + 0.758, + 0.167 + ], + "angle": 0, + "content": "Samples in new session" + }, + { + "type": "image", + "bbox": [ + 0.634, + 0.174, + 0.66, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.668, + 0.176, + 0.698, + 0.186 + ], + "angle": 0, + "content": "Margin" + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.193, + 0.661, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.668, + 0.195, + 0.741, + 0.203 + ], + "angle": 0, + "content": "Decision boundary" + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.213, + 0.663, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.668, + 0.213, + 0.732, + 0.23 + ], + "angle": 0, + "content": "Region signified by old prompts." + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.236, + 0.664, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.668, + 0.235, + 0.732, + 0.252 + ], + "angle": 0, + "content": "Region signified by new prompts" + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.259, + 0.655, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.668, + 0.257, + 0.737, + 0.293 + ], + "angle": 0, + "content": "Old sample using new prompts Prompt-ambiguity overlap space" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.317, + 0.789, + 0.414 + ], + "angle": 0, + "content": "Fig. 1: Models are incrementally learned across different sessions, where classes are accurately classified within each session. However, old and new class spaces overlap, so class margins lose abilities to distinguish different classes after learning a new session. Prompts signify distinct regions within the class space, yet without explicit inter-session prompt modeling, they fail to separate old and new class spaces. We aim to optimize prompts to create discriminative region and diminish prompt-ambiguity overlap space simultaneously, thereby establishing clear inter-session class margins." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.784, + 0.52 + ], + "angle": 0, + "content": "continuously changing data (i.e., different sessions) and dynamically build a unified classifier (i.e., model the relationships among different classes). However, Continual Learning encounters a significant challenge known as catastrophic forgetting [17], which arises when models, after incorporating new class data in a new learning session, tend to forget the previously acquired knowledge." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.785, + 0.612 + ], + "angle": 0, + "content": "To mitigate catastrophic forgetting, pre-training has been identified as a pivotal technique [41,43,48,65,81]. A recent innovation in this realm is the adoption of prompt-based methods. These methods freeze a pre-trained model and employ a few learnable prompt parameters to steer representation learning, referred to as Prompt-based Continual Learning methods [55,64,70,71], which exhibit distinct superiority in continual learning performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.786, + 0.809 + ], + "angle": 0, + "content": "In prompt-based continual learning methods, effective prompt optimization plays a pivotal role in solving the catastrophic forgetting problem, particularly when confronting the overlap between old and new class spaces. Concretely, across continual learning sessions, different classes are isolated, fostering an asynchronous learning process between sessions. Therefore, models tend to overfit the data and create boundaries specific to the classes within each session. This results in significant overlap between the old and new class spaces, as depicted in Fig. 1(a). Consequently, the margins between old classes lose abilities to distinguish between different classes, leading to the catastrophic forgetting problem. As shown in Fig. 1 (b), exiting prompt-based methods [55,64,70,71] use prompts to adjust features in each session. In essence, the learnable prompts create regions in the class space and relocate samples to the regions, thus being crucial for reducing the overlap between old and new class spaces." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.786, + 0.84 + ], + "angle": 0, + "content": "However, in existing prompt-based continual learning methods [55,64,70,71], the optimization objective of prompts lacks explicit modeling for inter-session" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RCS-Prompt" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "class relationships. In a new session, prompts are optimized to adjust classification scores solely for classes within that session. Consequently, as shown in Fig. 1 (b), prompts lack the ability to effectively separate old and new class spaces. Additionally, some samples from old classes may use new prompts to adjust features during inference [64, 70]. In such instances, unilaterally optimizing new prompts to move samples away from old classes exacerbates the overlap. This phenomenon, termed the prompt-ambiguity overlap space, makes optimizing prompts to separate class spaces become more challenging. Another observed concern is the diminished efficacy of prompts when their magnitudes become notably smaller than features during optimization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.788, + 0.539 + ], + "angle": 0, + "content": "To address the above issues, we propose a novel method called Learning Prompt to Rearrange Class Space (RCS-Prompt) for prompt-based continual learning. As depicted in Fig. 1 (c), our main idea is to bidirectionally optimize prompts, making prompts signify discriminative regions in the class space while suppressing the prompt-ambiguity overlap space. Specifically, we firstly construct the Inter-session Bidirectional Prompt Loss. During training, if new prompts focused by new samples exhibit reduced influence on old samples, we explicitly reduce the classification scores of these new samples on old classes for optimizing new prompts. Consequently, these new prompts signify discriminative regions in the class space. Conversely, when new prompts focused by new samples may be attended by old samples, we select a subset of such new samples and relabel them as old classes for learning. Furthermore, the symmetric loss is introduced to prevent the negative impact brought by relabelling. Therefore, the prompt-ambiguity overlap space is reduced. Secondly, we propose the Prompt Magnitude Normalization to align prompt and feature magnitudes before using prompts, ensuring effective prompt tuning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.54, + 0.788, + 0.584 + ], + "angle": 0, + "content": "Our method can serve as a plug-in to use, enlarging inter-session class margins by minimizing the overlap between old and new class spaces. The contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.593, + 0.785, + 0.651 + ], + "angle": 0, + "content": "- We propose a novel Inter-session Bidirectional Prompt Loss to optimize prompts, making session prompts signify discriminative regions and reducing the prompt-ambiguity overlap space. Thus, the overlap between old and new class spaces is minimized for constructing clear inter-session class margins." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.653, + 0.785, + 0.696 + ], + "angle": 0, + "content": "- We additionally present the Prompt Magnitude Normalization, which aligns prompt and feature magnitudes before using prompts, ensuring prompts for effectively adjusting model outcomes." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.698, + 0.785, + 0.74 + ], + "angle": 0, + "content": "- Without rehearsal, the proposed method achieves significant performance improvements compared to existing prompt-based continual learning methods, establishing state-of-the-art performance." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.593, + 0.785, + 0.74 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.765, + 0.396, + 0.781 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Continual Learning: gradually integrates new concepts while preserving previous knowledge, roughly categorized into three groups. The first employs exemplars from old classes (i.e., rehearsal) to recover past knowledge [1,7,8,38,39,44," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.415, + 0.129 + ], + "angle": 0, + "content": "Longrong Yang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "46, 51, 69, 72–74, 85, 90]. The second adjusts inductive bias through the regularization terms [3, 6, 10–12, 32, 36, 45, 50, 52, 56, 75, 80, 83, 88]. Some methods in this category use knowledge distillation [24] to align old and new model outputs, retaining old knowledge [14, 35, 49, 82]. Lastly, various approaches expand network capacity, including neuron-wise [77, 79], backbone-wise [37, 62, 63, 78, 81, 84, 87], token-wise [15, 67] and others [19, 26, 27, 66]. Some methods incorporate multiple techniques, e.g., [15] expands the network and uses rehearsal simultaneously. Furthermore, works on online continual learning, e.g., [54, 57], discuss the learning process without explicit task boundaries." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.405 + ], + "angle": 0, + "content": "Some traditional continual learning methods [25, 27, 75, 86] ever discussed the old and new class space overlap. These methods mainly analyze the overlap problem based on the dynamically updating convolution neural networks, and they alleviate this problem by using old data for knowledge distillation or creating balanced training subsets between old and new sessions. In contrast, our method focuses on prompt-based continual learning, where prompts are special and vital, and it explicitly eliminates the overlap between old and new class spaces by solely using new data to optimizing new prompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.804, + 0.63 + ], + "angle": 0, + "content": "Prompt-based continual learning: gains attraction with the surge of pre-trained models [13,47,60,67,89], using prompts to adapt pre-trained models for sequential data with new classes. L2P [71] applies visual prompt tuning [28] on pre-trained Transformer [13], learning a prompt pool for instance-specific selection. DualPrompt [70] extends L2P with general and expert prompts. S-Prompt [67] constructs only task-specific prompts for each task. CODA-Prompt [55] improves prompt selection via attention. SLCA [81] explores slow learners. DAP [29] generates instance-level prompts for use. LAE [18] reformulates different fin-tuning techniques into the Parameter-Efficient-Tuning module. PromptSRC [31] explores the use of prompts for continually updating CLIP. APG [59] develops a learnable adaptive prompt generator for weak pre-trained models. LGCL [30] introduces language guidance at the task level in the prompt pool. RanPAC [40] exploits training-free random projectors and class-prototype accumulation. Recently, HiDe-Prompt [64] decomposes the sub-optimality into hierarchical components and proposes different losses to solve them." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.79, + 0.738 + ], + "angle": 0, + "content": "While few approaches, e.g., HiDe-Prompt [64], suggest enhancing prompt disparity through regularization terms, their losses generally lack class-specificity characteristics, thereby failing to establish a connection between prompt disparity and inter-session class relationships. In contrast, our method optimizes prompts by explicitly adjusting classification logits to minimize overlap between old and new class spaces. Our approach effectively solves the prompt-ambiguity overlap space issue, a crucial concern overlooked by the existing methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.757, + 0.381, + 0.775 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.787, + 0.342, + 0.802 + ], + "angle": 0, + "content": "3.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Problem setup: This study focuses on prompt-based continual learning. Existing prompt-based continual learning methods [55, 64, 70, 71] typically develop a" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RCS-Prompt" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.145, + 0.756, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.204, + 0.468, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.472, + 0.204, + 0.756, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.403, + 0.788, + 0.502 + ], + "angle": 0, + "content": "Fig. 2: (a) General Pipeline. Existing prompt-based continual learning methods usually need to generate and insert prompts, optimize prompts by the loss, and train a session indicator. During test, the session ID of each sample is inferred and used for select prompts. (b) New samples demand high logits for new classes and low logits for old classes, whereas old samples require low logits for new classes and high logits for old classes. (c) Our Inter-session Bidirectional Prompt Loss satisfies the two requirements simultaneously, reducing the overlap between old and new spaces." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.688 + ], + "angle": 0, + "content": " unified classifier capable of learning from a dynamic data stream. The continual learning process involves a sequence of \\(B\\) learning sessions, denoted as \\(\\{D^1, D^2, \\dots, D^B\\}\\). Each learning session \\(D^b = (\\pmb{x}_i^b, y_i^b)_{i=1}^{n_b}\\) represents an incremental step with \\(n_b\\) instances. The training instance \\(\\pmb{x}_i^b\\) belongs to class \\(y_i^b \\in Y^b\\), where \\(Y^b\\) represents the label space of session \\(b\\). Notably, the label sets \\(Y^b\\) and \\(Y^{b'}\\) are disjoint for \\(b \\neq b'\\). In this study, \\(D^b\\) is only available at the session \\(b\\), where old data are not retrieved for rehearsal. Besides, session IDs are unavailable during inference. For simplicity, we refer to samples in the new session as \"new samples\" and those in the old session as \"old samples\". Likewise, we use the abbreviations \"new classes\", \"old classes\", \"new prompts\", and \"old prompts\" for convenience." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Existing prompt-based continual learning methods [55,64,70,71] mainly contain a Pre-Trained Model (PTM) \\(\\phi(\\cdot)\\) and some learnable prompts \\(\\pmb{pt}\\). As shown in Fig. 2 (a), there are three key components in these methods, which are prompt generation, prompt optimization, and session ID inference, respectively. First, in each session, a set of session-specific prompts \\(\\pmb{pt} \\in \\mathbb{R}^{L_p \\times D}\\) is generated, where \\(L_p\\) and \\(D\\) denote the prompt number and prompt dimensionality, respectively. \\(\\pmb{pt} = \\{\\pmb{pt}_1, \\dots, \\pmb{pt}_{L_p}\\}\\). They adjust features through cross attention. Concatenating feature tokens \\((N \\times D)\\) with prompts \\((L_p \\times D)\\) generates new tokens \\(((N + L_p) \\times D)\\) for attention, i.e., \\(\\phi(\\pmb{x}_i^b | \\pmb{pt})\\). Then, existing prompt-based continual learning methods use the below classification loss for optimizing prompts:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.413, + 0.129 + ], + "angle": 0, + "content": "Longrong Yang et al." + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.159, + 0.632, + 0.193 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {z} ^ {b} (\\boldsymbol {x} _ {i} ^ {b}) = [ \\underbrace {- \\inf , \\cdots , - \\inf} _ {\\text {o l d s e s s i o n}}, \\underbrace {z _ {1} ^ {b} , \\cdots} _ {\\text {n e w s e s s i o n}} ],\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.196, + 0.786, + 0.217 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {p} ^ {b} \\left(\\boldsymbol {x} _ {i} ^ {b}\\right) = \\operatorname {s o f t m a x} \\left(\\boldsymbol {z} ^ {b} \\left(\\boldsymbol {x} _ {i} ^ {b}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.364, + 0.219, + 0.644, + 0.258 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c e} \\left(\\boldsymbol {x} _ {i} ^ {b}\\right) = - \\sum_ {k = 1} ^ {K} q \\left(k \\mid \\boldsymbol {x} _ {i} ^ {b}\\right) \\cdot \\log \\left(p ^ {b} \\left(k \\mid \\boldsymbol {x} _ {i} ^ {b}\\right)\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.265, + 0.788, + 0.431 + ], + "angle": 0, + "content": "where \\( b \\) is the new session ID and \\( K \\) is the class number, \\( z^b(\\boldsymbol{x}_i^b) \\) indicates the logits of \\( \\boldsymbol{x}_i^b \\) and \\( z_1^b \\) indicates the logit of \\( \\boldsymbol{x}_i^b \\) on a new class. In continual learning, unlike standard supervised learning using cross-entropy loss, model predictions for old classes are typically set to -inf to separate them from training. \\( q(k|\\boldsymbol{x}_i^b) \\) represents the one-hot vector of the ground-truth label for sample \\( \\boldsymbol{x}_i^b \\). Lastly, during training, existing methods usually use features from the PTMs (i.e., PTM features, \\( \\phi(\\boldsymbol{x}_i^b) \\)) to train a session indicator. During inference, this session indicator is used to estimate the session ID of each sample. Based on the predicted session ID, samples select prompts from all old prompts \\( \\boldsymbol{p}\\boldsymbol{t}^{old} \\) and new prompts \\( \\boldsymbol{p}\\boldsymbol{t} \\). Under this scenario, as shown in Fig. 2 (b), old samples may select new prompts, leading to the prompt-ambiguity overlap space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.658 + ], + "angle": 0, + "content": "RCS-Prompt: Our method focuses on prompt optimization. Specifically, asynchronous learning takes place between sessions in continual learning, resulting in significant overlap between old and new class spaces. In prompt-based continual learning methods, prompts are vital to signify specific regions within the class space, where samples are relocated. Different sets of prompts correspond to distinct regions, and within each set, diverse prompts can reposition samples to various positions within a region. However, existing methods tend to overlook the explicit modeling of inter-session class relationships. Consequently, prompt-defined regions still exhibit overlap, impeding the construction of clear class margins. To this end, we propose an innovative approach called RCS-Prompt to rearrange class space by learning prompts, aiming to minimize the overlap between old and new class spaces, as shown in Fig. 2 (c). In detail, RCS-Prompt comprises two modules: Inter-session Bidirectional Prompt Loss and Prompt Magnitude Normalization. The specifics of each module will be introduced in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.68, + 0.589, + 0.696 + ], + "angle": 0, + "content": "3.2 Inter-session Bidirectional Prompt Loss" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "As mentioned above, our primary goal is to optimize new prompts during training on the new session, for minimizing the old and new class space overlap. However, achieving this goal is non-trivial due to a special challenge: When using new prompts to establish a region significantly distant from old class spaces and relocating samples to this region, old samples may inadvertently choose some new prompts to adjust features. Consequently, as new prompts shift old samples into the new region, overlap may emerge between old and new class spaces, a phenomenon referred to as the prompt-ambiguity overlap space. This challenge fundamentally stems from the divergence in indicator predictions for samples" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "RCS-Prompt" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.237 + ], + "angle": 0, + "content": "within the same session, suggesting the potential to employ a session indicator for sample separation. Building on this basic idea, as shown in Fig. 2 (c), we propose the Inter-session Bidirectional Prompt Loss to bidirectionally optimize new prompts, for shifting new samples away from old class spaces (i.e., producing discriminative regions) while keeping old samples close to their original class spaces (i.e., depressing the prompt-ambiguity overlap space)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.788, + 0.419 + ], + "angle": 0, + "content": "We use a session indicator to split samples, and design different losses to solve them respectively. Specifically, the first step is to generate a session indicator during training. Existing methods [64,70] use the frozen PTM features to train a session indicator, which is always changing during training. The changing session indicator may lead to fluctuating sample divisions, potentially increasing the convergence difficulty. Thus, before each training session, we first use the PTMs to extract the PTM features of all training samples within that session. Then, the PTM features of samples of each class are averaged, severing as the PTM feature prototype proto\\(_k\\). We store the PTM feature prototype of each class, resulting in a total storage of \\( K \\times D \\), which is merely the size of a classifier. During optimizing prompts, we compute the indicator score of each sample on all seen classes by:" + }, + { + "type": "equation", + "bbox": [ + 0.34, + 0.424, + 0.785, + 0.459 + ], + "angle": 0, + "content": "\\[\ns _ {k} = \\cos \\left(\\phi \\left(\\boldsymbol {x} _ {i} ^ {b}\\right), p r o t o _ {k}\\right) = \\frac {\\phi \\left(\\boldsymbol {x} _ {i} ^ {b}\\right) ^ {T} \\cdot p r o t o _ {k}}{\\left\\| \\phi \\left(\\boldsymbol {x} _ {i} ^ {b}\\right) \\right\\| \\cdot \\left\\| p r o t o _ {k} \\right\\|}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.785, + 0.508 + ], + "angle": 0, + "content": "where \\(\\cos (\\cdot)\\) denotes the cosine similarity. For each training sample, we can determine the class with the highest indicator score and its associated session, serving as its predicted session ID." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.81 + ], + "angle": 0, + "content": "The second step is categorizing new samples into two distinct groups. The first group is labeled as \"ID-clean\" samples, where new samples are predicted to belong to the new session during training. These ID-clean samples possess the PTM features that differ from old samples, indicating a reduced influence of new prompts on old samples during inference. Consequently, it becomes feasible to explicitly decrease the classification logits of these new samples for old classes for optimizing new prompts, as illustrated in Fig. 2 (c). This allows these new prompts to signify discriminative new regions, enabling the effective functioning of class margins. The second group is labeled as \"ID-noisy\" samples, where new samples are predicted to belong to old sessions during training. ID-noisy samples resemble old samples, suggesting that new prompts focused on them may be attended to by old samples during inference. Reducing the classification logits of these new samples for old classes would compel prompts to relocate old samples away from old class spaces, posing a risk of expanding the prompt-ambiguity overlap space. Therefore, we propose to relabel ID-noise samples to the old classes for learning, as depicted in Fig. 2 (c), to position old samples in the proximity of old class spaces during inference. In detail, for an ID-noisy sample \\( \\boldsymbol{x}_i^b \\), we manually assign its label to the old class with the highest indicator score, creating \"pseudo old labels\" \\( \\bar{y} \\): \\( \\bar{y} = \\mathrm{argmax}(s_k) \\), where \\( s_k \\) denotes the fine-grained indicator score of a sample on the class \\( k \\) (\\( k \\in \\{1, \\dots, K\\} \\))." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "The third step is designing losses for distinct samples. Assigning the labels of new samples to old classes for learning may have a negative impact on new class" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.414, + 0.129 + ], + "angle": 0, + "content": "Longrong Yang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "learning. To mitigate the negative effect, we propose to apply a symmetric loss to prevent new samples from excessively drifting toward old class spaces. The symmetric loss, e.g., the reverse cross entropy loss [68], is a family of loss which adaptively weighs clean and noisy sample gradients, preventing noisy samples from dominating the learning. By conducting a Gradient analysis below, we establish a connection between the symmetric loss and the challenge we currently face, revealing the potential to leverage the symmetric loss to prevent the negative impact of using pseudo old labels for new classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.788, + 0.329 + ], + "angle": 0, + "content": "In summary, in handling ID-clean samples, we maintain their original labels and explicitly decrease their classification logits on old classes. For ID-noisy samples, we relabel them as old classes and supervise them with the reverse cross-entropy loss. Thus, our Inter-session Bidirectional Prompt Loss is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.276, + 0.339, + 0.508, + 0.375 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {z} _ {d i s} ^ {b} (\\boldsymbol {x} _ {i} ^ {b}) = [ \\underbrace {z _ {1} ^ {1} , z _ {2} ^ {1} , \\cdots} _ {\\text {o l d s e s s i o n s}}, \\underbrace {z _ {1} ^ {b} , \\cdots} _ {\\text {n e w s e s s i o n}} ],\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.274, + 0.378, + 0.785, + 0.397 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {p} _ {d i s} ^ {b} \\left(\\boldsymbol {x} _ {i} ^ {b}\\right) = \\operatorname {s o f t m a x} \\left(\\boldsymbol {z} _ {d i s} ^ {b} \\left(\\boldsymbol {x} _ {i} ^ {b}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.274, + 0.399, + 0.73, + 0.439 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d i s} (\\boldsymbol {x} _ {i} ^ {b}) = \\left\\{ \\begin{array}{l l} - \\sum_ {k = 1} ^ {K} q (k | \\boldsymbol {x} _ {i} ^ {b}) \\cdot \\log (p _ {d i s} ^ {b} (k | \\boldsymbol {x} _ {i} ^ {b})) & \\text {i f} \\boldsymbol {x} _ {i} ^ {b} \\text {i s I D - c l e a n}, \\\\ - \\sum_ {k = 1} ^ {K} p _ {d i s} ^ {b} (k | \\boldsymbol {x} _ {i} ^ {b}) \\cdot \\log (\\bar {q} (k | \\boldsymbol {x} _ {i} ^ {b})) & \\text {i f} \\boldsymbol {x} _ {i} ^ {b} \\text {i s I D - n o i s y}, \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.788, + 0.474 + ], + "angle": 0, + "content": "where \\(\\bar{q}(k|\\pmb{x})\\) represents the one-hot vector of the pseudo old labels, and ID-noisy samples use the reverse cross-entropy loss. The total loss is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.485, + 0.785, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} \\left(\\boldsymbol {x} _ {i} ^ {b}\\right) = \\mathcal {L} _ {\\text {c e}} \\left(\\boldsymbol {x} _ {i} ^ {b}\\right) + \\alpha \\cdot \\mathcal {L} _ {\\text {d i s}} \\left(\\boldsymbol {x} _ {i} ^ {b}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.513, + 0.736, + 0.527 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{ce}\\) is the standard classification loss and \\(\\alpha\\) is a hyper-parameter." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.528, + 0.788, + 0.633 + ], + "angle": 0, + "content": "Gradient analysis: By the gradient analysis, our objective is to clarify the adaptive weighting mechanism of the reverse cross-entropy loss for ID-noisy samples, for preventing the negative impact of pseudo old labels on new classes. The gradients of the reverse cross-entropy loss and the cross-entropy loss have been derived in [68]. For brevity, we denote \\( p_k \\) and \\( q_k \\) as abbreviations for \\( p(k|\\pmb{x}) \\) and \\( q(k|\\pmb{x}) \\). The gradient of the reverse cross-entropy loss for ID-noisy samples with respect to the logit \\( z_k \\) can be derived as:" + }, + { + "type": "equation", + "bbox": [ + 0.365, + 0.645, + 0.785, + 0.683 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\mathcal {L} _ {d i s}}{\\partial z _ {k}} = \\left\\{ \\begin{array}{l l} A p _ {k} - A p _ {k} ^ {2} & \\text {i f} q _ {k} = q _ {y} = 1, \\\\ - A p _ {k} p _ {y} & \\text {i f} q _ {k} = 0, \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.692, + 0.788, + 0.709 + ], + "angle": 0, + "content": "where \\( A \\) is a constant. The gradient of the cross-entropy loss \\( \\mathcal{L}_{ce} \\) is indicated as:" + }, + { + "type": "equation", + "bbox": [ + 0.383, + 0.719, + 0.785, + 0.759 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\mathcal {L} _ {c e}}{\\partial z _ {k}} = \\left\\{ \\begin{array}{l l} p _ {k} - 1 & \\text {i f} q _ {k} = q _ {y} = 1, \\\\ p _ {k} & \\text {i f} q _ {k} = 0. \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "In the case of \\( q_{k} = q_{y} \\), \\( Ap_{k} - Ap_{k}^{2} \\) serves as an adaptive weighting term based on \\( p_{k} \\). In detail, when \\( p_{k} \\) is close to 0, the cross-entropy loss generates large gradients to compel samples to increase scores on the labeled class. Conversely, for the reverse cross-entropy loss, \" \\( p_{k} \\) is close to 0\" (an interval approaching 0, e.g., [0.0, 0.5)) indicates that the label is unreliable, resulting in the loss" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "RCS-Prompt" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "decreasing its gradient to 0. When \\( p_k \\) approaches 1 (an interval approaching 1), the reverse cross-entropy loss becomes similar to the cross-entropy loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.329 + ], + "angle": 0, + "content": "Then, we establish a connection between the gradient and our current scenario. Throughout the training, low scores of samples on pseudo old labels signify their low resemblance to old samples. This scenario corresponds to \" \\( p_k \\) is close to 0\" . In such cases, the cross-entropy loss generates large gradients to force samples to have high logits on pseudo old labels, while the reverse cross-entropy loss can decrease the gradients of these samples to prevent them from excessively drifting toward old class spaces. Conversely, high logits of samples on pseudo old labels signify their proximity to old samples. This scenario corresponds to \" \\( p_k \\) is close to 1\", and the reverse cross-entropy loss becomes similar to the cross-entropy loss, treating these samples as if they were old samples." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.353, + 0.544, + 0.369 + ], + "angle": 0, + "content": "3.3 Prompt Magnitude Normalization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.785, + 0.486 + ], + "angle": 0, + "content": "Prompts operate by adjusting features to influence model outputs. We find that when the magnitude of prompts is significantly lower than that of features, their efficacy in altering outputs diminishes. When the magnitude of prompts is significantly higher than that of features, the feature distribution may be severely destroyed. Therefore, in this study, Prompt Magnitude Normalization is proposed to align the magnitude of prompts and features before using prompts, for enhancing the effectiveness of prompt tuning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.787, + 0.534 + ], + "angle": 0, + "content": "To achieve our goal, we directly normalize prompts. Specifically, we compute the mean \\(\\mu_{\\pmb{x}}\\) and variance \\(\\tau_{\\pmb{x}}\\) of each sample feature, along with the mean \\(\\mu_{\\pmb{pt}}\\) and variance \\(\\tau_{\\pmb{pt}}\\) of its corresponding prompt. Then, we normalize the prompt:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.546, + 0.786, + 0.577 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {p t} = \\frac {\\boldsymbol {p t} - \\mu_ {\\boldsymbol {p t}}}{\\tau_ {\\boldsymbol {p t}}} \\cdot \\tau_ {\\boldsymbol {x}} + \\mu_ {\\boldsymbol {x}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.787, + 0.618 + ], + "angle": 0, + "content": "where this normalization ensures that the prompt and feature magnitudes are comparable before using prompts." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.644, + 0.377, + 0.661 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.677, + 0.457, + 0.693 + ], + "angle": 0, + "content": "4.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Benchmark: This study uses the benchmark presented in the recent state-of-the-art method HiDe-Prompt [64], which mainly evaluates model performance on Split CIFAR-100 [33] and Split ImageNet-R [22]. Split CIFAR-100 features 100 classes evenly distributed across 10 incremental sessions. Split ImageNet-R includes 200 classes, encompassing challenging ImageNet examples. Similarly, these classes are evenly distributed across 10 incremental sessions. Then, the additional evaluation datasets in [64] include 5-Datasets and Split CUB-200. 5-Datasets [16] includes CIFAR-10 [33], MNIST [34], Fashion-MNIST [76], SVHN [42] and notMNIST [4] datasets, each treated as an incremental task to evaluate the impact of" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.413, + 0.129 + ], + "angle": 0, + "content": "Longrong Yang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.188 + ], + "angle": 0, + "content": "Table 1: The comparison (\\%) between our method and five prompt-based continual methods. Following [64], the results of seeds 40, 42, and 44 are averaged to report. HiDe-Prompt uses Multi-Centroid statistical modeling [64]." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.199, + 0.79, + 0.584 + ], + "angle": 0, + "content": "
PTMMethodSplit CIFAR-100Split ImageNet-R
FAA (↑)CAA (↑)FFM (↓)FAA (↑)CAA (↑)FFM (↓)
Sup-21KL2P [71]83.0688.256.5863.6567.257.51
S-Prompt++ [67]88.8192.253.8769.6872.503.29
CODA-Prompt [55]86.9491.574.0470.0374.265.17
DualPrompt [70]86.6090.644.4568.7971.964.49
+RCS-Prompt (Ours)90.5394.214.0870.8074.384.69
HiDe-Prompt [64]92.9294.600.8673.5575.930.95
+RCS-Prompt (Ours)93.3495.340.8274.8577.520.91
DINO-1KL2P [71]70.6579.029.4657.4062.563.58
S-Prompt++ [67]74.9783.827.7857.6463.795.08
CODA-Prompt [55]77.5084.818.1063.1569.736.86
DualPrompt [70]74.9083.984.4558.5764.895.80
+RCS-Prompt (Ours)81.4489.324.2667.4171.735.10
HiDe-Prompt [64]90.0692.922.1369.3472.351.60
+RCS-Prompt (Ours)92.1194.091.9374.5478.022.41
MoCo-1KL2P [71]74.8583.146.5151.6458.872.37
S-Prompt++ [67]76.3083.8814.6753.1560.034.11
CODA-Prompt [55]76.8384.9712.6055.7565.4910.46
DualPrompt [70]77.7785.316.6152.5760.652.73
+RCS-Prompt (Ours)82.9189.374.6361.4165.572.61
HiDe-Prompt [64]90.7093.231.5663.0566.901.78
+RCS-Prompt (Ours)92.0093.651.5568.2471.041.74
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.787, + 0.643 + ], + "angle": 0, + "content": "large inter-session differences. Split CUB-200 [61] includes 200-class fine-grained images of birds, randomly split into 10 incremental sessions of disjoint classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.764 + ], + "angle": 0, + "content": "Baselines: On the evaluation benchmark, we extensively compare our method to recent prompt-based continual learning methods: L2P [71], S-Prompt++ [67], DualPrompt [70], CODA-Prompt [55] and HiDe-Prompt [64]. In experiments, we mainly add our method to HiDe-Prompt [64]. Extending upon this baseline, we add our proposed Inter-session Bidirectional Prompt Loss during training and apply the Prompt Magnitude Normalization before using prompts. Others keep unchanging. Besides, we provide results adding our method to DualPrompt [70], for showing the flexibility of our method as a plug-in." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.81 + ], + "angle": 0, + "content": "Training details: Following HiDe-Prompt [64], we select ViT-B/16 Sup-21K, DINO-1K [5], and MoCo-1K [9] as the pre-trained models to conduct experiments. We set \\(\\alpha = 0.4\\). More details are provided in the Sec. 3 of supp." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Evaluation protocol: We record the average accuracy of all seen classes after learning each session, presenting the last one as the final average accuracy" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "RCS-Prompt" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.128 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.144, + 0.788, + 0.175 + ], + "angle": 0, + "content": "Table 2: The comparison (\\%) between our method and five prompt-based continual learning methods on additional datasets. Following [64], the FAA metrics are report." + }, + { + "type": "table", + "bbox": [ + 0.254, + 0.185, + 0.751, + 0.321 + ], + "angle": 0, + "content": "
Method5-DatasetsSplitCUB-200
Sup-21KiBOT-21KSup-21KiBOT-21K
L2P [71]81.8482.2574.4844.29
DualPrompt [70]77.9168.0382.0541.31
S-Prompt++ [67]86.0677.2082.0842.73
CODA-Prompt [55]64.1851.6574.3447.79
HiDe-Prompt [64]93.8394.8886.5678.23
+RCS-Prompt (Ours)94.5196.1187.3279.42
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.349, + 0.788, + 0.396 + ], + "angle": 0, + "content": "(FAA) and their historical average as the cumulative average accuracy (CAA). We also present the final forgetting measure (FFM) of all sessions. More details are provided in the Sec. 3 of supp." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.419, + 0.373, + 0.433 + ], + "angle": 0, + "content": "4.2 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.787, + 0.492 + ], + "angle": 0, + "content": "Tab. 1 shows the results of our method and five prompt-based continual learning competitors on two datasets with three different pre-trained models. From the results, we observe that our RCS-Prompt performs very competitively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.789, + 0.688 + ], + "angle": 0, + "content": "Specifically, when employing ViT-B/16 Sup-21K as the pre-trained model and augmenting it with RCS-Prompt, the performance of HiDe-Prompt increases by \\(0.42\\%\\) and \\(1.30\\%\\) FAA on the Split CIFAR-100 and Split ImageNet-R datasets, respectively. When using DINO-1K as the pre-trained model and augmenting it with RCS-Prompt, the performance of HiDe-Prompt increases by \\(2.05\\%\\) and \\(5.20\\%\\) FAA on the Split CIFAR-100 and Split ImageNet-R datasets, respectively. We find that our method exhibits higher performance improvements when the self-supervised pre-trained models are used. This phenomenon suggests that our approach excels in establishing clear class margins, particularly effective in scenarios where self-supervised pre-trained models struggle with defining clear class margins. Furthermore, equipping with RCS-Prompt leads to significant performance improvements in DualPrompt, which proves the flexibility our method as a plug-in." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.78 + ], + "angle": 0, + "content": "Additional datasets: Furthermore, we evaluate our method on 5-Datasets and Split CUB-200 in Tab. 2. On the 5-Datasets, when integrated with RCS-Prompt, the performance of HiDe-Prompt increases \\(0.68\\%\\) and \\(1.23\\%\\) with different PTMs, respectively. Then, on the Split CUB-200 dataset, employing RCS-Prompt leads to a notable improvement in the performance of HiDe-Prompt, with gains of \\(0.76\\%\\) and \\(1.19\\%\\) with different PTMs, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Additional experimental outcomes involving more datasets (e.g., ImageNet-A [23] and ObjectNet [2]) and pre-trained models are detailed in the Sec. 4 of supp. The comparison between our method with DAP [29], LAE [18], APG [59] and RanPAC [40] on their benchmark is also provided." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.412, + 0.129 + ], + "angle": 0, + "content": "Longrong Yang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 3: Component ablation study \\((\\%)\\). Baseline indicate HiDe-Prompt [64] using DINO-1K and the seed 42. \\(\\mathcal{L}_{dis}\\) means adding the Inter-session Bidirectional Prompt Loss. Normalization means adding the Prompt Magnitude Normalization." + }, + { + "type": "table", + "bbox": [ + 0.326, + 0.199, + 0.679, + 0.293 + ], + "angle": 0, + "content": "
BaselineL_disNormalizationFAACAA
✓71.6075.61
✓✓74.9278.20
✓✓71.9375.83
✓✓✓75.1278.35
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.329, + 0.402, + 0.344 + ], + "angle": 0, + "content": "4.3 Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.784, + 0.435 + ], + "angle": 0, + "content": "We conduct all ablation studies on challenging Split ImageNet-R dataset [22] with DINO-1K using the seed as 42 (the widely used seed in [55,70,71]). More discussion about the continual learning performance trend, technical selections and visualization analyses are detailed in the Sec. 5 of supp." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.636 + ], + "angle": 0, + "content": "Component ablation: We conduct experiments to evaluate the impacts of the proposed Inter-session Bidirectional Prompt Loss and Prompt Magnitude Normalization. Our experiments involve: (i) Baseline: we establish HiDePrompt [64] as our baseline. (ii) Baseline \\(+\\mathcal{L}_{dis}\\): we add the Inter-session Bidirectional Prompt Loss during training. (iii) Baseline \\(+\\mathrm{Normalization}\\): we add the Prompt Magnitude Normalization before using prompts. (iv) Baseline \\(+\\mathcal{L}_{dis} + \\mathrm{Normalization}\\): two modules are used simultaneously. From the results in Tab. 3, we reach the following observations: (i) The primary contribution of this paper, the proposed Inter-session Bidirectional Prompt Loss \\(\\mathcal{L}_{dis}\\), yields a considerably \\(3.32\\%\\) performance increase over the baseline. (ii) The Prompt Magnitude Normalization improves model performance by \\(0.33\\%\\). (iii) Leveraging both the Inter-session Bidirectional Prompt Loss and the Prompt Magnitude Normalization achieves a model performance of \\(75.12\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.788, + 0.777 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{dis}\\) vs. contrastive regularization: This study aims to verify the superiority of \\(\\mathcal{L}_{dis}\\) than Contrastive Regularization [64]. The Contrastive Regularization (CR) is designed to enhance the session-specific prompt differences, having the potential to reduce the overlap between old and new class spaces. In Tab. 4, we observe that our proposed loss significantly outperforms the contrastive regularization. This superiority stems from the fact that the contrastive loss is class-agnostic, not establishing a connection between prompt discrimination and class relationships. Besides, the contrastive regularization fails to alleviate the occurrence of the prompt-ambiguity overlap space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Loss design: This discussion aims to prove that bidirectional prompt optimization is vital for improving the model performance, and the symmetric loss works. For constructing \\(\\mathcal{L}_{dis}\\), we alter the labels of ID-noisy samples to pseudo old labels and design the symmetric loss to optimize prompts. In this discussion," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "RCS-Prompt" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 4: The comparison (\\%) between different losses for verifying the superiority of our proposed loss than simply enlarging the inter-session prompt difference. CR is a regularization term used in [64] to enlarge the inter-session prompt difference. In all experiments, the Prompt Magnitude Normalization is used." + }, + { + "type": "table", + "bbox": [ + 0.255, + 0.212, + 0.75, + 0.272 + ], + "angle": 0, + "content": "
Baseline w/o CR+CR+Ldis+CR & Ldis
FAA71.4771.9374.9075.12
CAA75.7375.8378.2978.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.34 + ], + "angle": 0, + "content": "Table 5: The comparison (\\%) between different loss designs for verifying the necessity of the bidirectional prompt optimization. \"w/o old\" means not using pseudo old labels for ID-noisy samples and \"w old\" means using pseudo old labels for ID-noisy samples. In all experiments, the Prompt Magnitude Normalization is used." + }, + { + "type": "table", + "bbox": [ + 0.299, + 0.351, + 0.706, + 0.426 + ], + "angle": 0, + "content": "
Baseline+Lce dis+Ldis
w/o oldw oldw old
FAA71.9371.5073.0775.12
CAA75.8376.0576.8178.35
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.451, + 0.76, + 0.466 + ], + "angle": 0, + "content": "we formulate the cross-entropy loss concerning both old and new classes as:" + }, + { + "type": "equation", + "bbox": [ + 0.342, + 0.474, + 0.787, + 0.517 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d i s} ^ {c e} \\left(\\boldsymbol {x} _ {i} ^ {b}\\right) = - \\sum_ {k = 1} ^ {K} q \\left(k \\mid \\boldsymbol {x} _ {i} ^ {b}\\right) \\cdot \\log \\left(p _ {d i s} ^ {b} \\left(k \\mid \\boldsymbol {x} _ {i} ^ {b}\\right)\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.788, + 0.674 + ], + "angle": 0, + "content": "where old class logits are not set to -inf. The analysis from Tab. 5 yields several key observations: (i) After adding \\(\\mathcal{L}_{dis}^{ce}\\) but not using pseudo old labels, the model performance reduces from 71.93 to 71.50. This suggests that undirectionally enhancing inter-session prompt discrimination is ineffective, implying a potential enlargement of the prompt-ambiguity overlap space. (ii) Compared to \\(+ \\mathcal{L}_{dis}^{ce}\\) without old\", \\(+ \\mathcal{L}_{dis}^{ce}\\) with old\" results in a \\(1.57\\%\\) higher model performance. Thus, bidirectionally optimizing prompts is vital for diminishing the class overlap. (iii) After using pseudo old labels, than adding \\(\\mathcal{L}_{dis}^{ce}\\), adding \\(\\mathcal{L}_{dis}\\) improves the model performance from 73.07 to 75.12. This indicates that the symmetric loss can mitigate the negative impact brought by using pseudo old labels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Hyper-parameter analysis: We investigate the effects of the hyperparameter \\(\\alpha\\). More specifically, we explore a range of values for \\(\\alpha\\), namely \\(\\{0.0, 0.2, 0.4, 0.6, 1.0\\}\\). In the results presented in Tab. 6, it is evident that as the value of \\(\\alpha\\) increases, there is a corresponding gradual improvement in performance. This improvement plateaus when \\(\\alpha\\) attains a value of 0.4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Statistical analysis: This statistic aims to use inter-session misclassification to quantify the old and new class space overlap. In this analysis, without losing the generality, we focus on the predictions for test samples from session 0 (the first session) after learning all sessions. The total test sample count is 650 in session 0. For the predictions from the baseline (HiDe-Prompt), we compute the count of samples that are misclassified as classes from other sessions (session 0)," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.413, + 0.129 + ], + "angle": 0, + "content": "Longrong Yang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.336, + 0.145, + 0.668, + 0.16 + ], + "angle": 0, + "content": "Table 6: Impacts (%) of \\( \\alpha \\) on Split ImageNet-R." + }, + { + "type": "table", + "bbox": [ + 0.338, + 0.171, + 0.668, + 0.23 + ], + "angle": 0, + "content": "
0.00.20.40.61.0
FAA71.9374.3375.1274.1273.58
CAA75.8377.8578.3577.6377.44
" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.244, + 0.413, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.244, + 0.756, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.409, + 0.789, + 0.465 + ], + "angle": 0, + "content": "Fig. 3: The analysis about misclassification. (a) We analyze which type of misclassification is the main type for the baseline. Inter-session misclassification means that a sample is misclassified as a class from other sessions. (b) We quantify the solving of our method for the inter-session misclassification." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.496, + 0.789, + 0.648 + ], + "angle": 0, + "content": "termed as inter-session (intra-session) misclassification type, as shown in Fig. 3 (a). Then, as shown in Fig. 3 (b), focusing on inter-session misclassification, we compute the count of samples that are misclassified in each session to compare the baseline and our method. Fig. 3 reveals two key observations: \\((i)\\) For the baseline, 173 \\((26.62\\%)\\) test samples are misclassified as classes in another session, and only 31 \\((4.77\\%)\\) are misclassified as classes from session 0. This shows that the inter-session misclassification is the main problem for existing methods. \\((ii)\\) Than the baseline, our method significantly decreases the inter-session misclassification sample count from 173 to 67. This shows that our method can effectively solve the inter-session misclassification problem." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.671, + 0.36, + 0.687 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Our study reveals that the optimization objective of prompts lacks the explicit modeling of inter-session class relationships, not effectively separating new and old class spaces. To minimize the overlap between new and old class spaces for creating the clear inter-session class margins, we introduce an Inter-session Bidirectional Prompt Loss to place new and old samples to discriminative regions, while eliminating the prompt-ambiguity overlap space. Furthermore, we design the Prompt Magnitude Normalization to enhance the prompt tuning effectiveness. Extensive experiments demonstrate the superiority of our approach across different datasets and continual learning scenarios." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RCS-Prompt" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.808, + 0.284 + ], + "angle": 0, + "content": "Acknowledgements. This work is supported in part by National Natural Science Foundation of China under Grant No.62441602, National Science Foundation for Distinguished Young Scholars under Grant 62225605, Project 12326608 supported by NSFC, National Natural Science Foundation of China under Grant U20A20222, Zhejiang Key Research and Development Program under Grant 2023C03196, the Ningbo Science and Technology Innovation Project (No.2024Z294), the Fundamental Research Funds for the Central Universities, the Key R&D Program of Zhejiang Province, China 2023C01043, Ant Group through CCF-Ant Research Fund, and sponsored by CCF-AFSG Research Fund." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.307, + 0.326, + 0.323 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.339, + 0.787, + 0.38 + ], + "angle": 0, + "content": "1. Aljundi, R., Lin, M., Goujaud, B., Bengio, Y.: Gradient based sample selection for online continual learning. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.381, + 0.787, + 0.436 + ], + "angle": 0, + "content": "2. Barbu, A., Mayo, D., Alverio, J., Luo, W., Wang, C., Gutfreund, D., Tenenbaum, J., Katz, B.: Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.437, + 0.787, + 0.478 + ], + "angle": 0, + "content": "3. Belouadah, E., Popescu, A.: II2m: Class incremental learning with dual memory. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 583-592 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.479, + 0.495, + 0.492 + ], + "angle": 0, + "content": "4. Bulatov, Y.: Notmnist dataset (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.493, + 0.787, + 0.535 + ], + "angle": 0, + "content": "5. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.535, + 0.787, + 0.589 + ], + "angle": 0, + "content": "6. Cha, S., Cho, S., Hwang, D., Hong, S., Lee, M., Moon, T.: Rebalancing batch normalization for exemplar-based class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20127-20136 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.59, + 0.787, + 0.632 + ], + "angle": 0, + "content": "7. Chaudhry, A., Dokania, P.K., Ajanthan, T., Torr, P.H.: Riemannian walk for incremental learning: Understanding forgetting and intransigence. In: Proceedings of the European conference on computer vision (ECCV). pp. 532-547 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.633, + 0.787, + 0.66 + ], + "angle": 0, + "content": "8. Chaudhry, A., Ranzato, M., Rohrbach, M., Elhoseiny, M.: Efficient lifelong learning with a-gem. In: International Conference on Learning Representations (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.66, + 0.787, + 0.701 + ], + "angle": 0, + "content": "9. Chen, X., Xie, S., He, K.: An empirical study of training self-supervised vision transformers. in 2021 IEEE. In: CVF International Conference on Computer Vision (ICCV). pp. 9620-9629" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.702, + 0.787, + 0.756 + ], + "angle": 0, + "content": "10. Dong, J., Li, H., Cong, Y., Sun, G., Zhang, Y., Van Gool, L.: No one left behind: Real-world federated class-incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence 46(4), 2054-2070 (2024). https://doi.org/10.1109/TPAMI.2023.3334213" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.757, + 0.787, + 0.799 + ], + "angle": 0, + "content": "11. Dong, J., Liang, W., Cong, Y., Sun, G.: Heterogeneous forgetting compensation for class-incremental learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 11742-11751 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "12. Dong, J., Wang, L., Fang, Z., Sun, G., Xu, S., Wang, X., Zhu, Q.: Federated class-incremental learning. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2022)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.339, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.412, + 0.129 + ], + "angle": 0, + "content": "Longrong Yang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "13. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.205, + 0.787, + 0.259 + ], + "angle": 0, + "content": "14. Douillard, A., Cord, M., Ollion, C., Robert, T., Valle, E.: Podnet: Pooled outputs distillation for small-tasks incremental learning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16. pp. 86-102. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.26, + 0.787, + 0.3 + ], + "angle": 0, + "content": "15. Douillard, A., Ramé, A., Couairon, G., Cord, M.: Dytox: Transformers for continual learning with dynamic token expansion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9285-9295 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.301, + 0.787, + 0.355 + ], + "angle": 0, + "content": "16. Ebrahimi, S., Meier, F., Calandra, R., Darrell, T., Rohrbach, M.: Adversarial continual learning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16. pp. 386-402. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.357, + 0.787, + 0.383 + ], + "angle": 0, + "content": "17. French, R.M.: Catastrophic forgetting in connectionist networks. Trends in cognitive sciences 3(4), 128-135 (1999)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.384, + 0.787, + 0.424 + ], + "angle": 0, + "content": "18. Gao, Q., Zhao, C., Sun, Y., Xi, T., Zhang, G., Ghanem, B., Zhang, J.: A unified continual learning framework with general parameter-efficient tuning. arXiv preprint arXiv:2303.10070 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.426, + 0.787, + 0.452 + ], + "angle": 0, + "content": "19. Ge, Y., Li, Y., Ni, S., Zhao, J., Yang, M.H., Itti, L.: Clr: Channel-wise lightweight reprogramming for continual learning. arXiv preprint arXiv:2307.11386 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.453, + 0.787, + 0.493 + ], + "angle": 0, + "content": "20. Gomes, H.M., Barddal, J.P., Enembreck, F., Bifet, A.: A survey on ensemble learning for data stream classification. ACM Computing Surveys (CSUR) 50(2), 1-36 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.495, + 0.787, + 0.535 + ], + "angle": 0, + "content": "21. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.537, + 0.787, + 0.591 + ], + "angle": 0, + "content": "22. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8340-8349 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.592, + 0.787, + 0.632 + ], + "angle": 0, + "content": "23. Hendrycks, D., Zhao, K., Basart, S., Steinhardt, J., Song, D.: Natural adversarial examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15262-15271 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.634, + 0.787, + 0.66 + ], + "angle": 0, + "content": "24. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.661, + 0.787, + 0.702 + ], + "angle": 0, + "content": "25. Hou, S., Pan, X., Loy, C.C., Wang, Z., Lin, D.: Learning a unified classifier incrementally via rebalancing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 831-839 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.703, + 0.787, + 0.742 + ], + "angle": 0, + "content": "26. Hu, Z., Li, Y., Lyu, J., Gao, D., Vasconcelos, N.: Dense network expansion for class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11858-11867 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.744, + 0.787, + 0.785 + ], + "angle": 0, + "content": "27. Huang, B., Chen, Z., Zhou, P., Chen, J., Wu, Z.: Resolving task confusion in dynamic expansion architectures for class incremental learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 908-916 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.786, + 0.787, + 0.84 + ], + "angle": 0, + "content": "28. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII. pp. 709-727. Springer (2022)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RCS-Prompt" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "29. Jung, D., Han, D., Bang, J., Song, H.: Generating instance-level prompts for rehearsal-free continual learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11847-11857 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.245 + ], + "angle": 0, + "content": "30. Khan, M.G.Z.A., Naeem, M.F., Van Gool, L., Stricker, D., Tombari, F., Afzal, M.Z.: Introducing language guidance in prompt-based continual learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11463-11473 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.246, + 0.788, + 0.3 + ], + "angle": 0, + "content": "31. Khattak, M.U., Wasim, S.T., Naseer, M., Khan, S., Yang, M.H., Khan, F.S.: Self-regulating prompts: Foundational model adaptation without forgetting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15190-15200 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.342 + ], + "angle": 0, + "content": "32. Kim, D., Han, B.: On the stability-plasticity dilemma of class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20196-20204 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.37 + ], + "angle": 0, + "content": "33. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.371, + 0.788, + 0.398 + ], + "angle": 0, + "content": "34. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.398, + 0.788, + 0.424 + ], + "angle": 0, + "content": "35. Li, Z., Hoiem, D.: Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence 40(12), 2935-2947 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.425, + 0.788, + 0.467 + ], + "angle": 0, + "content": "36. Liu, Y., Li, Y., Schiele, B., Sun, Q.: Online hyperparameter optimization for class-incremental learning. Proceedings of the AAAI Conference on Artificial Intelligence (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.467, + 0.788, + 0.509 + ], + "angle": 0, + "content": "37. Liu, Y., Schiele, B., Sun, Q.: Adaptive aggregation networks for class-incremental learning. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 2544-2553 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.509, + 0.788, + 0.55 + ], + "angle": 0, + "content": "38. Liu, Y., Su, Y., Liu, A.A., Schiele, B., Sun, Q.: Mnemonics training: Multi-class incremental learning without forgetting. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 12245-12254 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.55, + 0.788, + 0.592 + ], + "angle": 0, + "content": "39. Luo, Z., Liu, Y., Schiele, B., Sun, Q.: Class-incremental exemplar compression for class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11371-11380 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.592, + 0.788, + 0.633 + ], + "angle": 0, + "content": "40. McDonnell, M.D., Gong, D., Parvaneh, A., Abbasnejad, E., van den Hengel, A.: Ranpac: Random projections and pre-trained models for continual learning. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.633, + 0.788, + 0.675 + ], + "angle": 0, + "content": "41. Mehta, S.V., Patil, D., Chandar, S., Strubell, E.: An empirical investigation of the role of pre-training in lifelong learning. Journal of Machine Learning Research 24(214), 1-50 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.702 + ], + "angle": 0, + "content": "42. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.702, + 0.788, + 0.743 + ], + "angle": 0, + "content": "43. Panos, A., Kobe, Y., Reino, D.O., Aljundi, R., Turner, R.E.: First session adaptation: A strong replay-free baseline for class-incremental learning. arXiv preprint arXiv:2303.13199 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.743, + 0.788, + 0.799 + ], + "angle": 0, + "content": "44. Petit, G., Popescu, A., Schindler, H., Picard, D., Delezoide, B.: Fetril: Feature translation for exemplar-free class-incremental learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3911-3920 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.788, + 0.842 + ], + "angle": 0, + "content": "45. Pham, Q., Liu, C., Hoi, S.: Continual normalization: Rethinking batch normalization for online continual learning. In: International Conference on Learning Representations (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.413, + 0.129 + ], + "angle": 0, + "content": "Longrong Yang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "46. Qiang, S., Hou, J., Wan, J., Liang, Y., Lei, Z., Zhang, D.: Mixture uniform distribution modeling and asymmetric mix distillation for class incremental learning. In: Proceedings of the AAAI Conference on Artificial Intelligence (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "47. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.786, + 0.286 + ], + "angle": 0, + "content": "48. Ramasesh, V.V., Lewkowycz, A., Dyer, E.: Effect of scale on catastrophic forgetting in neural networks. In: International Conference on Learning Representations (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.786, + 0.328 + ], + "angle": 0, + "content": "49. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 2001-2010 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.786, + 0.37 + ], + "angle": 0, + "content": "50. Rymarczyk, D., van de Weijer, J., Zielinski, B., Twardowski, B.: Icicle: Interpretable class incremental continual learning. arXiv preprint arXiv:2303.07811 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.786, + 0.412 + ], + "angle": 0, + "content": "51. Sarfraz, F., Arani, E., Zonooz, B.: Sparse coding in a dual memory system for lifelong learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 9714-9722 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.786, + 0.467 + ], + "angle": 0, + "content": "52. Shi, Y., Zhou, K., Liang, J., Jiang, Z., Feng, J., Torr, P.H., Bai, S., Tan, V.Y.: Mimicking the oracle: an initial phase decorrelation approach for class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16722-16731 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.786, + 0.495 + ], + "angle": 0, + "content": "53. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.786, + 0.55 + ], + "angle": 0, + "content": "54. Singh, P., Li, Y., Sikarwar, A., Lei, S.W., Gao, D., Talbot, M.B., Sun, Y., Shou, M.Z., Kreiman, G., Zhang, M.: Learning to learn: How to continuously teach humans and machines. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11708-11719 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.786, + 0.619 + ], + "angle": 0, + "content": "55. Smith, J.S., Karlinsky, L., Gutta, V., Cascante-Bonilla, P., Kim, D., Arbelle, A., Panda, R., Feris, R., Kira, Z.: Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11909-11919 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.786, + 0.661 + ], + "angle": 0, + "content": "56. Sun, Z., Mu, Y., Hua, G.: Regularizing second-order influences for continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20166-20175 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.786, + 0.702 + ], + "angle": 0, + "content": "57. Talbot, M.B., Zawar, R., Badkundri, R., Zhang, M., Kreiman, G.: Tuned compositional feature replays for efficient stream learning. IEEE Transactions on Neural Networks and Learning Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.786, + 0.744 + ], + "angle": 0, + "content": "58. Tan, M., Pang, R., Le, Q.V.: Efficientdet: Scalable and efficient object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10781-10790 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.786, + 0.785 + ], + "angle": 0, + "content": "59. Tang, Y.M., Peng, Y.X., Zheng, W.S.: When prompt-based incremental learning does not meet strong pretraining. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1706-1716 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.786, + 0.841 + ], + "angle": 0, + "content": "60. Villa, A., Alcázar, J.L., Alfarra, M., Alhamoud, K., Hurtado, J., Heilbron, F.C., Soto, A., Ghanem, B.: Pivot: Prompting for video continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24214-24223 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RCS-Prompt" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "61. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.217 + ], + "angle": 0, + "content": "62. Wang, F.Y., Zhou, D.W., Liu, L., Ye, H.J., Bian, Y., Zhan, D.C., Zhao, P.: Beef: Bi-compatible class-incremental learning via energy-based expansion and fusion. In: The Eleventh International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.787, + 0.272 + ], + "angle": 0, + "content": "63. Wang, F.Y., Zhou, D.W., Ye, H.J., Zhan, D.C.: Foster: Feature boosting and compression for class-incremental learning. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXV. pp. 398-414. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "64. Wang, L., Xie, J., Zhang, X., Huang, M., Su, H., Zhu, J.: Hierarchical decomposition of prompt-based continual learning: Rethinking obscured sub-optimality. In: Advances in neural information processing systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.315, + 0.787, + 0.342 + ], + "angle": 0, + "content": "65. Wang, L., Zhang, X., Su, H., Zhu, J.: A comprehensive survey of continual learning: Theory, method and application. arXiv preprint arXiv:2302.00487 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.343, + 0.787, + 0.383 + ], + "angle": 0, + "content": "66. Wang, W., Hu, Y., Chen, Q., Zhang, Y.: Task difficulty aware parameter allocation & regularization for lifelong learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7776-7785 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.384, + 0.787, + 0.424 + ], + "angle": 0, + "content": "67. Wang, Y., Huang, Z., Hong, X.: S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. Annual Conference on Neural Information Processing Systems (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.425, + 0.787, + 0.466 + ], + "angle": 0, + "content": "68. Wang, Y., Ma, X., Chen, Z., Luo, Y., Yi, J., Bailey, J.: Symmetric cross entropy for robust learning with noisy labels. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 322-330 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.467, + 0.787, + 0.494 + ], + "angle": 0, + "content": "69. Wang, Z., Zhao, Z., Wu, Y., Zhou, L., Xu, D.: Progressive target-styled feature augmentation for unsupervised domain adaptation on point clouds. In: ECCV (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.495, + 0.787, + 0.563 + ], + "angle": 0, + "content": "70. Wang, Z., Zhang, Z., Ebrahimi, S., Sun, R., Zhang, H., Lee, C.Y., Ren, X., Su, G., Perot, V., Dy, J., et al.: Dualprompt: Complementary prompting for rehearsal-free continual learning. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVI. pp. 631-648. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.564, + 0.787, + 0.618 + ], + "angle": 0, + "content": "71. Wang, Z., Zhang, Z., Lee, C.Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., Pfister, T.: Learning to prompt for continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 139-149 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.619, + 0.787, + 0.673 + ], + "angle": 0, + "content": "72. Wu, X., Wang, H., Wu, Y., Li, X.: D3t-gan: Data-dependent domain transfer gans for image generation with limited data. ACM Trans. Multimedia Comput. Commun. Appl. 19(4) (mar 2023). https://doi.org/10.1145/3576858, https://doi.org/10.1145/3576858" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.674, + 0.787, + 0.729 + ], + "angle": 0, + "content": "73. Wu, Y., Li, H., Wang, F., Zhang, Y., Liang, R.: Self-distilled dynamic fusion network for language-based fashion retrieval. In: ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 3260-3264. IEEE (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.729, + 0.787, + 0.771 + ], + "angle": 0, + "content": "74. Wu, Y., Wu, X., Li, X., Tian, J.: Mgh: Metadata guided hypergraph modeling for unsupervised person re-identification. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 1571–1580 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.772, + 0.787, + 0.812 + ], + "angle": 0, + "content": "75. Wu, Y., Chen, Y., Wang, L., Ye, Y., Liu, Z., Guo, Y., Fu, Y.: Large scale incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 374-382 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "76. Xiao, H., Rasul, K., Vollgraf, R.: Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.412, + 0.129 + ], + "angle": 0, + "content": "Longrong Yang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "77. Xu, J., Zhu, Z.: Reinforced continual learning. Advances in Neural Information Processing Systems 31 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.217 + ], + "angle": 0, + "content": "78. Yan, S., Xie, J., He, X.: Der: Dynamically expandable representation for class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3014-3023 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.258 + ], + "angle": 0, + "content": "79. Yoon, J., Yang, E., Lee, J., Hwang, S.J.: Lifelong learning with dynamically expandable networks. In: International Conference on Learning Representations (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.259, + 0.785, + 0.313 + ], + "angle": 0, + "content": "80. Yu, L., Twardowski, B., Liu, X., Herranz, L., Wang, K., Cheng, Y., Jui, S., Weijer, J.v.d.: Semantic drift compensation for class-incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6982-6991 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "81. Zhang, G., Wang, L., Kang, G., Chen, L., Wei, Y.: Slow learner with classifier alignment for continual learning on a pre-trained model. arXiv preprint arXiv:2303.05118 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.357, + 0.785, + 0.41 + ], + "angle": 0, + "content": "82. Zhang, J., Zhang, J., Ghosh, S., Li, D., Tasci, S., Heck, L., Zhang, H., Kuo, C.C.J.: Class-incremental learning via deep model consolidation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1131-1140 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "83. Zhao, B., Xiao, X., Gan, G., Zhang, B., Xia, S.T.: Maintaining discrimination and fairness in class incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13208-13217 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.453, + 0.785, + 0.493 + ], + "angle": 0, + "content": "84. Zhao, H., Fu, Y., Kang, M., Tian, Q., Wu, F., Li, X.: Mgmt: Multi-grained slow vs. fast framework for few-shot class-incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.495, + 0.785, + 0.535 + ], + "angle": 0, + "content": "85. Zhao, H., Wang, H., Fu, Y., Wu, F., Li, X.: Memory-efficient class-incremental learning for image classification. IEEE Transactions on Neural Networks and Learning Systems 33(10), 5966-5977 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.537, + 0.785, + 0.577 + ], + "angle": 0, + "content": "86. Zhou, D.W., Wang, F.Y., Ye, H.J., Ma, L., Pu, S., Zhan, D.C.: Forward compatible few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9046-9056 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.579, + 0.785, + 0.618 + ], + "angle": 0, + "content": "87. Zhou, D.W., Wang, Q.W., Ye, H.J., Zhan, D.C.: A model or 603 exemplars: Towards memory-efficient class-incremental learning. In: International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.62, + 0.785, + 0.66 + ], + "angle": 0, + "content": "88. Zhou, D.W., Ye, H.J., Ma, L., Xie, D., Pu, S., Zhan, D.C.: Few-shot class incremental learning by sampling multi-phase tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.661, + 0.785, + 0.688 + ], + "angle": 0, + "content": "89. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.689, + 0.785, + 0.73 + ], + "angle": 0, + "content": "90. Zhuang, H., Weng, Z., Wei, H., Xie, R., Toh, K.A., Lin, Z.: Acil: Analytic class incremental learning with absolute memorization and privacy protection. Advances in Neural Information Processing Systems 35, 11602-11614 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.73 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/RCS-Prompt_ Learning Prompt to Rearrange Class Space for Prompt-based Continual Learning/6bf9dca4-8b37-4308-ad56-3bfced51f84f_origin.pdf b/2024/RCS-Prompt_ Learning Prompt to Rearrange Class Space for Prompt-based Continual Learning/6bf9dca4-8b37-4308-ad56-3bfced51f84f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e5eb6c97f2c44cdffe86a8122b4ec170fdfc4291 --- /dev/null +++ b/2024/RCS-Prompt_ Learning Prompt to Rearrange Class Space for Prompt-based Continual Learning/6bf9dca4-8b37-4308-ad56-3bfced51f84f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e0f6d453ab1a83751b08a78d3d4660aee206e6d4bdf75790175dc2feb71074e8 +size 640858 diff --git a/2024/RCS-Prompt_ Learning Prompt to Rearrange Class Space for Prompt-based Continual Learning/full.md b/2024/RCS-Prompt_ Learning Prompt to Rearrange Class Space for Prompt-based Continual Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..94837ea08d5910ed70c28f6e6e36d0ec2931d5ce --- /dev/null +++ b/2024/RCS-Prompt_ Learning Prompt to Rearrange Class Space for Prompt-based Continual Learning/full.md @@ -0,0 +1,381 @@ +# RCS-Prompt: Learning Prompt to Rearrange Class Space for Prompt-based Continual Learning + +Longrong Yang $^{1}$ , Hanbin Zhao $^{1}$ , Yunlong $\mathrm{Yu}^{2\star}$ , Xiaodong Zeng $^{3}$ , and Xi Li $^{1,4,5\star}$ + +1College of Computer Science & Technology, Zhejiang University + +$^{2}$ College of Information Science and Electronic Engineering, Zhejiang University $^{3}$ Ant Services Group $^{4}$ Shanghai Institute for Advanced Study of Zhejiang University + +$^{5}$ Zhejiang - Singapore Innovation and AI Joint Research Lab, Hangzhou + +{longrongyang, zhaohanbin, yuyunlong, xilizju}@zju.edu.cn + +{xiaodong.zxd}@antgroup.com + +Abstract. Prompt-based Continual Learning is an emerging direction in leveraging pre-trained knowledge for downstream continual learning. While arriving at a new session, existing prompt-based continual learning methods usually adapt features from pre-trained models to new data by introducing prompts. However, these prompts lack an optimization objective explicitly modeling inter-session class relationships, thus failing to construct clear inter-session class margins. Moreover, some old samples use new prompts during inference, resulting in the prompt-ambiguity overlap space - a special situation where old and new class spaces overlap. To address these issues, we propose an innovative approach called RCS-Prompt to Rearrange Class Space by bidirectionally optimizing prompts. RCS-Prompt optimizes prompts to signify discriminative regions across different sessions in the class space. Additionally, it mitigates the prompt-ambiguity overlap space by altering the labels of a small subset of new samples to old classes and training them with a customized symmetric loss. The proposed method effectively reduces the overlap between old and new class spaces, thereby establishing clear inter-session class margins. We extensively evaluate RCS-Prompt on public datasets, demonstrating its effectiveness in prompt-based continual learning. Code is available at https://github.com/longrongyang/RCS-Prompt. + +Keywords: Prompt-based Continual Learning $\cdot$ Continual Learning $\cdot$ Class Space Overlap $\cdot$ Class Margin + +# 1 Introduction + +While deep models have achieved remarkable success across diverse fields [21, 53, 58], most research has focused on recognizing a fixed number of classes in static environments. However, in real-world scenarios, models often encounter streaming data where new classes may emerge over time [20]. To address this issue, Continual Learning has been proposed, which allows models to learn from + +![](images/566a58d4abd0cbaa8f794e01c37e30db37dc8962cae970976550a16c97dd11fe.jpg) +(a) Without Prompts +Class Space Severe class overlap + +![](images/795bccccf132810738dbe6cfdad1249efb30aeb81890d0831e25a195fd06f133.jpg) +Class Space Similar prompts Still class overlap + +![](images/ddc6742f07c5ad555a223a52095c41027896c60e6c027f220bc94747ea402801.jpg) +(c) Our Method +Class Space +Optimize prompts to +solve severe class overlap + +![](images/5671b9373790bd07156ffcc9f527dcd767839e52647ca3d23a8df7e956919494.jpg) + +Samples in old session + +![](images/e9cd74167318525308d2c60e9a1d3c98c9b237678d157c06ac2c494d1fb7e6b7.jpg) + +Samples in new session + +![](images/1060f61ba3283f745c874b6eb585d76116432be4b895f22435d6622f8390fba1.jpg) + +Margin + +![](images/28b8010d5277cabb582d639aa3627396e46833e4e3cef6413e776ebb1d8bcb0a.jpg) + +Decision boundary + +![](images/1b6e490a7ae9ff9b11a2f25795973b0bb46ff6fcd9445b5dd0f222548b20f99b.jpg) + +Region signified by old prompts. + +![](images/1e2f65c2609d816b9537d019603a3b80bfb2082fadf845abaf11a39e9105ed58.jpg) + +Region signified by new prompts + +![](images/22eddcde71182843eaa648a75b9d172626620852d8a201de0dee9997981071cb.jpg) +Fig. 1: Models are incrementally learned across different sessions, where classes are accurately classified within each session. However, old and new class spaces overlap, so class margins lose abilities to distinguish different classes after learning a new session. Prompts signify distinct regions within the class space, yet without explicit inter-session prompt modeling, they fail to separate old and new class spaces. We aim to optimize prompts to create discriminative region and diminish prompt-ambiguity overlap space simultaneously, thereby establishing clear inter-session class margins. + +Old sample using new prompts Prompt-ambiguity overlap space + +continuously changing data (i.e., different sessions) and dynamically build a unified classifier (i.e., model the relationships among different classes). However, Continual Learning encounters a significant challenge known as catastrophic forgetting [17], which arises when models, after incorporating new class data in a new learning session, tend to forget the previously acquired knowledge. + +To mitigate catastrophic forgetting, pre-training has been identified as a pivotal technique [41,43,48,65,81]. A recent innovation in this realm is the adoption of prompt-based methods. These methods freeze a pre-trained model and employ a few learnable prompt parameters to steer representation learning, referred to as Prompt-based Continual Learning methods [55,64,70,71], which exhibit distinct superiority in continual learning performance. + +In prompt-based continual learning methods, effective prompt optimization plays a pivotal role in solving the catastrophic forgetting problem, particularly when confronting the overlap between old and new class spaces. Concretely, across continual learning sessions, different classes are isolated, fostering an asynchronous learning process between sessions. Therefore, models tend to overfit the data and create boundaries specific to the classes within each session. This results in significant overlap between the old and new class spaces, as depicted in Fig. 1(a). Consequently, the margins between old classes lose abilities to distinguish between different classes, leading to the catastrophic forgetting problem. As shown in Fig. 1 (b), exiting prompt-based methods [55,64,70,71] use prompts to adjust features in each session. In essence, the learnable prompts create regions in the class space and relocate samples to the regions, thus being crucial for reducing the overlap between old and new class spaces. + +However, in existing prompt-based continual learning methods [55,64,70,71], the optimization objective of prompts lacks explicit modeling for inter-session + +class relationships. In a new session, prompts are optimized to adjust classification scores solely for classes within that session. Consequently, as shown in Fig. 1 (b), prompts lack the ability to effectively separate old and new class spaces. Additionally, some samples from old classes may use new prompts to adjust features during inference [64, 70]. In such instances, unilaterally optimizing new prompts to move samples away from old classes exacerbates the overlap. This phenomenon, termed the prompt-ambiguity overlap space, makes optimizing prompts to separate class spaces become more challenging. Another observed concern is the diminished efficacy of prompts when their magnitudes become notably smaller than features during optimization. + +To address the above issues, we propose a novel method called Learning Prompt to Rearrange Class Space (RCS-Prompt) for prompt-based continual learning. As depicted in Fig. 1 (c), our main idea is to bidirectionally optimize prompts, making prompts signify discriminative regions in the class space while suppressing the prompt-ambiguity overlap space. Specifically, we firstly construct the Inter-session Bidirectional Prompt Loss. During training, if new prompts focused by new samples exhibit reduced influence on old samples, we explicitly reduce the classification scores of these new samples on old classes for optimizing new prompts. Consequently, these new prompts signify discriminative regions in the class space. Conversely, when new prompts focused by new samples may be attended by old samples, we select a subset of such new samples and relabel them as old classes for learning. Furthermore, the symmetric loss is introduced to prevent the negative impact brought by relabelling. Therefore, the prompt-ambiguity overlap space is reduced. Secondly, we propose the Prompt Magnitude Normalization to align prompt and feature magnitudes before using prompts, ensuring effective prompt tuning. + +Our method can serve as a plug-in to use, enlarging inter-session class margins by minimizing the overlap between old and new class spaces. The contributions are summarized as follows: + +- We propose a novel Inter-session Bidirectional Prompt Loss to optimize prompts, making session prompts signify discriminative regions and reducing the prompt-ambiguity overlap space. Thus, the overlap between old and new class spaces is minimized for constructing clear inter-session class margins. +- We additionally present the Prompt Magnitude Normalization, which aligns prompt and feature magnitudes before using prompts, ensuring prompts for effectively adjusting model outcomes. +- Without rehearsal, the proposed method achieves significant performance improvements compared to existing prompt-based continual learning methods, establishing state-of-the-art performance. + +# 2 Related Works + +Continual Learning: gradually integrates new concepts while preserving previous knowledge, roughly categorized into three groups. The first employs exemplars from old classes (i.e., rehearsal) to recover past knowledge [1,7,8,38,39,44, + +46, 51, 69, 72–74, 85, 90]. The second adjusts inductive bias through the regularization terms [3, 6, 10–12, 32, 36, 45, 50, 52, 56, 75, 80, 83, 88]. Some methods in this category use knowledge distillation [24] to align old and new model outputs, retaining old knowledge [14, 35, 49, 82]. Lastly, various approaches expand network capacity, including neuron-wise [77, 79], backbone-wise [37, 62, 63, 78, 81, 84, 87], token-wise [15, 67] and others [19, 26, 27, 66]. Some methods incorporate multiple techniques, e.g., [15] expands the network and uses rehearsal simultaneously. Furthermore, works on online continual learning, e.g., [54, 57], discuss the learning process without explicit task boundaries. + +Some traditional continual learning methods [25, 27, 75, 86] ever discussed the old and new class space overlap. These methods mainly analyze the overlap problem based on the dynamically updating convolution neural networks, and they alleviate this problem by using old data for knowledge distillation or creating balanced training subsets between old and new sessions. In contrast, our method focuses on prompt-based continual learning, where prompts are special and vital, and it explicitly eliminates the overlap between old and new class spaces by solely using new data to optimizing new prompts. + +Prompt-based continual learning: gains attraction with the surge of pre-trained models [13,47,60,67,89], using prompts to adapt pre-trained models for sequential data with new classes. L2P [71] applies visual prompt tuning [28] on pre-trained Transformer [13], learning a prompt pool for instance-specific selection. DualPrompt [70] extends L2P with general and expert prompts. S-Prompt [67] constructs only task-specific prompts for each task. CODA-Prompt [55] improves prompt selection via attention. SLCA [81] explores slow learners. DAP [29] generates instance-level prompts for use. LAE [18] reformulates different fin-tuning techniques into the Parameter-Efficient-Tuning module. PromptSRC [31] explores the use of prompts for continually updating CLIP. APG [59] develops a learnable adaptive prompt generator for weak pre-trained models. LGCL [30] introduces language guidance at the task level in the prompt pool. RanPAC [40] exploits training-free random projectors and class-prototype accumulation. Recently, HiDe-Prompt [64] decomposes the sub-optimality into hierarchical components and proposes different losses to solve them. + +While few approaches, e.g., HiDe-Prompt [64], suggest enhancing prompt disparity through regularization terms, their losses generally lack class-specificity characteristics, thereby failing to establish a connection between prompt disparity and inter-session class relationships. In contrast, our method optimizes prompts by explicitly adjusting classification logits to minimize overlap between old and new class spaces. Our approach effectively solves the prompt-ambiguity overlap space issue, a crucial concern overlooked by the existing methods. + +# 3 Methodology + +# 3.1 Overview + +Problem setup: This study focuses on prompt-based continual learning. Existing prompt-based continual learning methods [55, 64, 70, 71] typically develop a + +![](images/7b5c24b05e944e7f8b621e91ad3b73556fcbce5cebd6fe952dc9fcbaae41465d.jpg) + +![](images/5028f922c50a1f9e2791ae4090e670ec0684240c6730a7c02329595327789649.jpg) +Fig. 2: (a) General Pipeline. Existing prompt-based continual learning methods usually need to generate and insert prompts, optimize prompts by the loss, and train a session indicator. During test, the session ID of each sample is inferred and used for select prompts. (b) New samples demand high logits for new classes and low logits for old classes, whereas old samples require low logits for new classes and high logits for old classes. (c) Our Inter-session Bidirectional Prompt Loss satisfies the two requirements simultaneously, reducing the overlap between old and new spaces. + +![](images/7f3a05069d3ec7446b37f55ed4732e822f8630d7069453c3f0115be6cb234a6f.jpg) + +unified classifier capable of learning from a dynamic data stream. The continual learning process involves a sequence of $B$ learning sessions, denoted as $\{D^1, D^2, \dots, D^B\}$ . Each learning session $D^b = (\pmb{x}_i^b, y_i^b)_{i=1}^{n_b}$ represents an incremental step with $n_b$ instances. The training instance $\pmb{x}_i^b$ belongs to class $y_i^b \in Y^b$ , where $Y^b$ represents the label space of session $b$ . Notably, the label sets $Y^b$ and $Y^{b'}$ are disjoint for $b \neq b'$ . In this study, $D^b$ is only available at the session $b$ , where old data are not retrieved for rehearsal. Besides, session IDs are unavailable during inference. For simplicity, we refer to samples in the new session as "new samples" and those in the old session as "old samples". Likewise, we use the abbreviations "new classes", "old classes", "new prompts", and "old prompts" for convenience. + +Existing prompt-based continual learning methods [55,64,70,71] mainly contain a Pre-Trained Model (PTM) $\phi(\cdot)$ and some learnable prompts $\pmb{pt}$ . As shown in Fig. 2 (a), there are three key components in these methods, which are prompt generation, prompt optimization, and session ID inference, respectively. First, in each session, a set of session-specific prompts $\pmb{pt} \in \mathbb{R}^{L_p \times D}$ is generated, where $L_p$ and $D$ denote the prompt number and prompt dimensionality, respectively. $\pmb{pt} = \{\pmb{pt}_1, \dots, \pmb{pt}_{L_p}\}$ . They adjust features through cross attention. Concatenating feature tokens $(N \times D)$ with prompts $(L_p \times D)$ generates new tokens $((N + L_p) \times D)$ for attention, i.e., $\phi(\pmb{x}_i^b | \pmb{pt})$ . Then, existing prompt-based continual learning methods use the below classification loss for optimizing prompts: + +$$ +\boldsymbol {z} ^ {b} (\boldsymbol {x} _ {i} ^ {b}) = [ \underbrace {- \inf , \cdots , - \inf} _ {\text {o l d s e s s i o n}}, \underbrace {z _ {1} ^ {b} , \cdots} _ {\text {n e w s e s s i o n}} ], +$$ + +$$ +\boldsymbol {p} ^ {b} \left(\boldsymbol {x} _ {i} ^ {b}\right) = \operatorname {s o f t m a x} \left(\boldsymbol {z} ^ {b} \left(\boldsymbol {x} _ {i} ^ {b}\right)\right), \tag {1} +$$ + +$$ +\mathcal {L} _ {c e} \left(\boldsymbol {x} _ {i} ^ {b}\right) = - \sum_ {k = 1} ^ {K} q \left(k \mid \boldsymbol {x} _ {i} ^ {b}\right) \cdot \log \left(p ^ {b} \left(k \mid \boldsymbol {x} _ {i} ^ {b}\right)\right), +$$ + +where $b$ is the new session ID and $K$ is the class number, $z^b(\boldsymbol{x}_i^b)$ indicates the logits of $\boldsymbol{x}_i^b$ and $z_1^b$ indicates the logit of $\boldsymbol{x}_i^b$ on a new class. In continual learning, unlike standard supervised learning using cross-entropy loss, model predictions for old classes are typically set to -inf to separate them from training. $q(k|\boldsymbol{x}_i^b)$ represents the one-hot vector of the ground-truth label for sample $\boldsymbol{x}_i^b$ . Lastly, during training, existing methods usually use features from the PTMs (i.e., PTM features, $\phi(\boldsymbol{x}_i^b)$ ) to train a session indicator. During inference, this session indicator is used to estimate the session ID of each sample. Based on the predicted session ID, samples select prompts from all old prompts $\boldsymbol{p}\boldsymbol{t}^{old}$ and new prompts $\boldsymbol{p}\boldsymbol{t}$ . Under this scenario, as shown in Fig. 2 (b), old samples may select new prompts, leading to the prompt-ambiguity overlap space. + +RCS-Prompt: Our method focuses on prompt optimization. Specifically, asynchronous learning takes place between sessions in continual learning, resulting in significant overlap between old and new class spaces. In prompt-based continual learning methods, prompts are vital to signify specific regions within the class space, where samples are relocated. Different sets of prompts correspond to distinct regions, and within each set, diverse prompts can reposition samples to various positions within a region. However, existing methods tend to overlook the explicit modeling of inter-session class relationships. Consequently, prompt-defined regions still exhibit overlap, impeding the construction of clear class margins. To this end, we propose an innovative approach called RCS-Prompt to rearrange class space by learning prompts, aiming to minimize the overlap between old and new class spaces, as shown in Fig. 2 (c). In detail, RCS-Prompt comprises two modules: Inter-session Bidirectional Prompt Loss and Prompt Magnitude Normalization. The specifics of each module will be introduced in the following sections. + +# 3.2 Inter-session Bidirectional Prompt Loss + +As mentioned above, our primary goal is to optimize new prompts during training on the new session, for minimizing the old and new class space overlap. However, achieving this goal is non-trivial due to a special challenge: When using new prompts to establish a region significantly distant from old class spaces and relocating samples to this region, old samples may inadvertently choose some new prompts to adjust features. Consequently, as new prompts shift old samples into the new region, overlap may emerge between old and new class spaces, a phenomenon referred to as the prompt-ambiguity overlap space. This challenge fundamentally stems from the divergence in indicator predictions for samples + +within the same session, suggesting the potential to employ a session indicator for sample separation. Building on this basic idea, as shown in Fig. 2 (c), we propose the Inter-session Bidirectional Prompt Loss to bidirectionally optimize new prompts, for shifting new samples away from old class spaces (i.e., producing discriminative regions) while keeping old samples close to their original class spaces (i.e., depressing the prompt-ambiguity overlap space). + +We use a session indicator to split samples, and design different losses to solve them respectively. Specifically, the first step is to generate a session indicator during training. Existing methods [64,70] use the frozen PTM features to train a session indicator, which is always changing during training. The changing session indicator may lead to fluctuating sample divisions, potentially increasing the convergence difficulty. Thus, before each training session, we first use the PTMs to extract the PTM features of all training samples within that session. Then, the PTM features of samples of each class are averaged, severing as the PTM feature prototype proto $_k$ . We store the PTM feature prototype of each class, resulting in a total storage of $K \times D$ , which is merely the size of a classifier. During optimizing prompts, we compute the indicator score of each sample on all seen classes by: + +$$ +s _ {k} = \cos \left(\phi \left(\boldsymbol {x} _ {i} ^ {b}\right), p r o t o _ {k}\right) = \frac {\phi \left(\boldsymbol {x} _ {i} ^ {b}\right) ^ {T} \cdot p r o t o _ {k}}{\left\| \phi \left(\boldsymbol {x} _ {i} ^ {b}\right) \right\| \cdot \left\| p r o t o _ {k} \right\|}, \tag {2} +$$ + +where $\cos (\cdot)$ denotes the cosine similarity. For each training sample, we can determine the class with the highest indicator score and its associated session, serving as its predicted session ID. + +The second step is categorizing new samples into two distinct groups. The first group is labeled as "ID-clean" samples, where new samples are predicted to belong to the new session during training. These ID-clean samples possess the PTM features that differ from old samples, indicating a reduced influence of new prompts on old samples during inference. Consequently, it becomes feasible to explicitly decrease the classification logits of these new samples for old classes for optimizing new prompts, as illustrated in Fig. 2 (c). This allows these new prompts to signify discriminative new regions, enabling the effective functioning of class margins. The second group is labeled as "ID-noisy" samples, where new samples are predicted to belong to old sessions during training. ID-noisy samples resemble old samples, suggesting that new prompts focused on them may be attended to by old samples during inference. Reducing the classification logits of these new samples for old classes would compel prompts to relocate old samples away from old class spaces, posing a risk of expanding the prompt-ambiguity overlap space. Therefore, we propose to relabel ID-noise samples to the old classes for learning, as depicted in Fig. 2 (c), to position old samples in the proximity of old class spaces during inference. In detail, for an ID-noisy sample $\boldsymbol{x}_i^b$ , we manually assign its label to the old class with the highest indicator score, creating "pseudo old labels" $\bar{y}$ : $\bar{y} = \mathrm{argmax}(s_k)$ , where $s_k$ denotes the fine-grained indicator score of a sample on the class $k$ ( $k \in \{1, \dots, K\}$ ). + +The third step is designing losses for distinct samples. Assigning the labels of new samples to old classes for learning may have a negative impact on new class + +learning. To mitigate the negative effect, we propose to apply a symmetric loss to prevent new samples from excessively drifting toward old class spaces. The symmetric loss, e.g., the reverse cross entropy loss [68], is a family of loss which adaptively weighs clean and noisy sample gradients, preventing noisy samples from dominating the learning. By conducting a Gradient analysis below, we establish a connection between the symmetric loss and the challenge we currently face, revealing the potential to leverage the symmetric loss to prevent the negative impact of using pseudo old labels for new classes. + +In summary, in handling ID-clean samples, we maintain their original labels and explicitly decrease their classification logits on old classes. For ID-noisy samples, we relabel them as old classes and supervise them with the reverse cross-entropy loss. Thus, our Inter-session Bidirectional Prompt Loss is formulated as: + +$$ +\boldsymbol {z} _ {d i s} ^ {b} (\boldsymbol {x} _ {i} ^ {b}) = [ \underbrace {z _ {1} ^ {1} , z _ {2} ^ {1} , \cdots} _ {\text {o l d s e s s i o n s}}, \underbrace {z _ {1} ^ {b} , \cdots} _ {\text {n e w s e s s i o n}} ], +$$ + +$$ +\boldsymbol {p} _ {d i s} ^ {b} \left(\boldsymbol {x} _ {i} ^ {b}\right) = \operatorname {s o f t m a x} \left(\boldsymbol {z} _ {d i s} ^ {b} \left(\boldsymbol {x} _ {i} ^ {b}\right)\right), \tag {3} +$$ + +$$ +\mathcal {L} _ {d i s} (\boldsymbol {x} _ {i} ^ {b}) = \left\{ \begin{array}{l l} - \sum_ {k = 1} ^ {K} q (k | \boldsymbol {x} _ {i} ^ {b}) \cdot \log (p _ {d i s} ^ {b} (k | \boldsymbol {x} _ {i} ^ {b})) & \text {i f} \boldsymbol {x} _ {i} ^ {b} \text {i s I D - c l e a n}, \\ - \sum_ {k = 1} ^ {K} p _ {d i s} ^ {b} (k | \boldsymbol {x} _ {i} ^ {b}) \cdot \log (\bar {q} (k | \boldsymbol {x} _ {i} ^ {b})) & \text {i f} \boldsymbol {x} _ {i} ^ {b} \text {i s I D - n o i s y}, \end{array} \right. +$$ + +where $\bar{q}(k|\pmb{x})$ represents the one-hot vector of the pseudo old labels, and ID-noisy samples use the reverse cross-entropy loss. The total loss is formulated as: + +$$ +\mathcal {L} _ {\text {t o t a l}} \left(\boldsymbol {x} _ {i} ^ {b}\right) = \mathcal {L} _ {\text {c e}} \left(\boldsymbol {x} _ {i} ^ {b}\right) + \alpha \cdot \mathcal {L} _ {\text {d i s}} \left(\boldsymbol {x} _ {i} ^ {b}\right), \tag {4} +$$ + +where $\mathcal{L}_{ce}$ is the standard classification loss and $\alpha$ is a hyper-parameter. + +Gradient analysis: By the gradient analysis, our objective is to clarify the adaptive weighting mechanism of the reverse cross-entropy loss for ID-noisy samples, for preventing the negative impact of pseudo old labels on new classes. The gradients of the reverse cross-entropy loss and the cross-entropy loss have been derived in [68]. For brevity, we denote $p_k$ and $q_k$ as abbreviations for $p(k|\pmb{x})$ and $q(k|\pmb{x})$ . The gradient of the reverse cross-entropy loss for ID-noisy samples with respect to the logit $z_k$ can be derived as: + +$$ +\frac {\partial \mathcal {L} _ {d i s}}{\partial z _ {k}} = \left\{ \begin{array}{l l} A p _ {k} - A p _ {k} ^ {2} & \text {i f} q _ {k} = q _ {y} = 1, \\ - A p _ {k} p _ {y} & \text {i f} q _ {k} = 0, \end{array} \right. \tag {5} +$$ + +where $A$ is a constant. The gradient of the cross-entropy loss $\mathcal{L}_{ce}$ is indicated as: + +$$ +\frac {\partial \mathcal {L} _ {c e}}{\partial z _ {k}} = \left\{ \begin{array}{l l} p _ {k} - 1 & \text {i f} q _ {k} = q _ {y} = 1, \\ p _ {k} & \text {i f} q _ {k} = 0. \end{array} \right. \tag {6} +$$ + +In the case of $q_{k} = q_{y}$ , $Ap_{k} - Ap_{k}^{2}$ serves as an adaptive weighting term based on $p_{k}$ . In detail, when $p_{k}$ is close to 0, the cross-entropy loss generates large gradients to compel samples to increase scores on the labeled class. Conversely, for the reverse cross-entropy loss, " $p_{k}$ is close to 0" (an interval approaching 0, e.g., [0.0, 0.5)) indicates that the label is unreliable, resulting in the loss + +decreasing its gradient to 0. When $p_k$ approaches 1 (an interval approaching 1), the reverse cross-entropy loss becomes similar to the cross-entropy loss. + +Then, we establish a connection between the gradient and our current scenario. Throughout the training, low scores of samples on pseudo old labels signify their low resemblance to old samples. This scenario corresponds to " $p_k$ is close to 0" . In such cases, the cross-entropy loss generates large gradients to force samples to have high logits on pseudo old labels, while the reverse cross-entropy loss can decrease the gradients of these samples to prevent them from excessively drifting toward old class spaces. Conversely, high logits of samples on pseudo old labels signify their proximity to old samples. This scenario corresponds to " $p_k$ is close to 1", and the reverse cross-entropy loss becomes similar to the cross-entropy loss, treating these samples as if they were old samples. + +# 3.3 Prompt Magnitude Normalization + +Prompts operate by adjusting features to influence model outputs. We find that when the magnitude of prompts is significantly lower than that of features, their efficacy in altering outputs diminishes. When the magnitude of prompts is significantly higher than that of features, the feature distribution may be severely destroyed. Therefore, in this study, Prompt Magnitude Normalization is proposed to align the magnitude of prompts and features before using prompts, for enhancing the effectiveness of prompt tuning. + +To achieve our goal, we directly normalize prompts. Specifically, we compute the mean $\mu_{\pmb{x}}$ and variance $\tau_{\pmb{x}}$ of each sample feature, along with the mean $\mu_{\pmb{pt}}$ and variance $\tau_{\pmb{pt}}$ of its corresponding prompt. Then, we normalize the prompt: + +$$ +\boldsymbol {p t} = \frac {\boldsymbol {p t} - \mu_ {\boldsymbol {p t}}}{\tau_ {\boldsymbol {p t}}} \cdot \tau_ {\boldsymbol {x}} + \mu_ {\boldsymbol {x}}, \tag {7} +$$ + +where this normalization ensures that the prompt and feature magnitudes are comparable before using prompts. + +# 4 Experiments + +# 4.1 Implementation Details + +Benchmark: This study uses the benchmark presented in the recent state-of-the-art method HiDe-Prompt [64], which mainly evaluates model performance on Split CIFAR-100 [33] and Split ImageNet-R [22]. Split CIFAR-100 features 100 classes evenly distributed across 10 incremental sessions. Split ImageNet-R includes 200 classes, encompassing challenging ImageNet examples. Similarly, these classes are evenly distributed across 10 incremental sessions. Then, the additional evaluation datasets in [64] include 5-Datasets and Split CUB-200. 5-Datasets [16] includes CIFAR-10 [33], MNIST [34], Fashion-MNIST [76], SVHN [42] and notMNIST [4] datasets, each treated as an incremental task to evaluate the impact of + +Table 1: The comparison (\%) between our method and five prompt-based continual methods. Following [64], the results of seeds 40, 42, and 44 are averaged to report. HiDe-Prompt uses Multi-Centroid statistical modeling [64]. + +
PTMMethodSplit CIFAR-100Split ImageNet-R
FAA (↑)CAA (↑)FFM (↓)FAA (↑)CAA (↑)FFM (↓)
Sup-21KL2P [71]83.0688.256.5863.6567.257.51
S-Prompt++ [67]88.8192.253.8769.6872.503.29
CODA-Prompt [55]86.9491.574.0470.0374.265.17
DualPrompt [70]86.6090.644.4568.7971.964.49
+RCS-Prompt (Ours)90.5394.214.0870.8074.384.69
HiDe-Prompt [64]92.9294.600.8673.5575.930.95
+RCS-Prompt (Ours)93.3495.340.8274.8577.520.91
DINO-1KL2P [71]70.6579.029.4657.4062.563.58
S-Prompt++ [67]74.9783.827.7857.6463.795.08
CODA-Prompt [55]77.5084.818.1063.1569.736.86
DualPrompt [70]74.9083.984.4558.5764.895.80
+RCS-Prompt (Ours)81.4489.324.2667.4171.735.10
HiDe-Prompt [64]90.0692.922.1369.3472.351.60
+RCS-Prompt (Ours)92.1194.091.9374.5478.022.41
MoCo-1KL2P [71]74.8583.146.5151.6458.872.37
S-Prompt++ [67]76.3083.8814.6753.1560.034.11
CODA-Prompt [55]76.8384.9712.6055.7565.4910.46
DualPrompt [70]77.7785.316.6152.5760.652.73
+RCS-Prompt (Ours)82.9189.374.6361.4165.572.61
HiDe-Prompt [64]90.7093.231.5663.0566.901.78
+RCS-Prompt (Ours)92.0093.651.5568.2471.041.74
+ +large inter-session differences. Split CUB-200 [61] includes 200-class fine-grained images of birds, randomly split into 10 incremental sessions of disjoint classes. + +Baselines: On the evaluation benchmark, we extensively compare our method to recent prompt-based continual learning methods: L2P [71], S-Prompt++ [67], DualPrompt [70], CODA-Prompt [55] and HiDe-Prompt [64]. In experiments, we mainly add our method to HiDe-Prompt [64]. Extending upon this baseline, we add our proposed Inter-session Bidirectional Prompt Loss during training and apply the Prompt Magnitude Normalization before using prompts. Others keep unchanging. Besides, we provide results adding our method to DualPrompt [70], for showing the flexibility of our method as a plug-in. + +Training details: Following HiDe-Prompt [64], we select ViT-B/16 Sup-21K, DINO-1K [5], and MoCo-1K [9] as the pre-trained models to conduct experiments. We set $\alpha = 0.4$ . More details are provided in the Sec. 3 of supp. + +Evaluation protocol: We record the average accuracy of all seen classes after learning each session, presenting the last one as the final average accuracy + +Table 2: The comparison (\%) between our method and five prompt-based continual learning methods on additional datasets. Following [64], the FAA metrics are report. + +
Method5-DatasetsSplitCUB-200
Sup-21KiBOT-21KSup-21KiBOT-21K
L2P [71]81.8482.2574.4844.29
DualPrompt [70]77.9168.0382.0541.31
S-Prompt++ [67]86.0677.2082.0842.73
CODA-Prompt [55]64.1851.6574.3447.79
HiDe-Prompt [64]93.8394.8886.5678.23
+RCS-Prompt (Ours)94.5196.1187.3279.42
+ +(FAA) and their historical average as the cumulative average accuracy (CAA). We also present the final forgetting measure (FFM) of all sessions. More details are provided in the Sec. 3 of supp. + +# 4.2 Main Results + +Tab. 1 shows the results of our method and five prompt-based continual learning competitors on two datasets with three different pre-trained models. From the results, we observe that our RCS-Prompt performs very competitively. + +Specifically, when employing ViT-B/16 Sup-21K as the pre-trained model and augmenting it with RCS-Prompt, the performance of HiDe-Prompt increases by $0.42\%$ and $1.30\%$ FAA on the Split CIFAR-100 and Split ImageNet-R datasets, respectively. When using DINO-1K as the pre-trained model and augmenting it with RCS-Prompt, the performance of HiDe-Prompt increases by $2.05\%$ and $5.20\%$ FAA on the Split CIFAR-100 and Split ImageNet-R datasets, respectively. We find that our method exhibits higher performance improvements when the self-supervised pre-trained models are used. This phenomenon suggests that our approach excels in establishing clear class margins, particularly effective in scenarios where self-supervised pre-trained models struggle with defining clear class margins. Furthermore, equipping with RCS-Prompt leads to significant performance improvements in DualPrompt, which proves the flexibility our method as a plug-in. + +Additional datasets: Furthermore, we evaluate our method on 5-Datasets and Split CUB-200 in Tab. 2. On the 5-Datasets, when integrated with RCS-Prompt, the performance of HiDe-Prompt increases $0.68\%$ and $1.23\%$ with different PTMs, respectively. Then, on the Split CUB-200 dataset, employing RCS-Prompt leads to a notable improvement in the performance of HiDe-Prompt, with gains of $0.76\%$ and $1.19\%$ with different PTMs, respectively. + +Additional experimental outcomes involving more datasets (e.g., ImageNet-A [23] and ObjectNet [2]) and pre-trained models are detailed in the Sec. 4 of supp. The comparison between our method with DAP [29], LAE [18], APG [59] and RanPAC [40] on their benchmark is also provided. + +Table 3: Component ablation study $(\%)$ . Baseline indicate HiDe-Prompt [64] using DINO-1K and the seed 42. $\mathcal{L}_{dis}$ means adding the Inter-session Bidirectional Prompt Loss. Normalization means adding the Prompt Magnitude Normalization. + +
BaselineL_disNormalizationFAACAA
✓71.6075.61
✓✓74.9278.20
✓✓71.9375.83
✓✓✓75.1278.35
+ +# 4.3 Further Analysis + +We conduct all ablation studies on challenging Split ImageNet-R dataset [22] with DINO-1K using the seed as 42 (the widely used seed in [55,70,71]). More discussion about the continual learning performance trend, technical selections and visualization analyses are detailed in the Sec. 5 of supp. + +Component ablation: We conduct experiments to evaluate the impacts of the proposed Inter-session Bidirectional Prompt Loss and Prompt Magnitude Normalization. Our experiments involve: (i) Baseline: we establish HiDePrompt [64] as our baseline. (ii) Baseline $+\mathcal{L}_{dis}$ : we add the Inter-session Bidirectional Prompt Loss during training. (iii) Baseline $+\mathrm{Normalization}$ : we add the Prompt Magnitude Normalization before using prompts. (iv) Baseline $+\mathcal{L}_{dis} + \mathrm{Normalization}$ : two modules are used simultaneously. From the results in Tab. 3, we reach the following observations: (i) The primary contribution of this paper, the proposed Inter-session Bidirectional Prompt Loss $\mathcal{L}_{dis}$ , yields a considerably $3.32\%$ performance increase over the baseline. (ii) The Prompt Magnitude Normalization improves model performance by $0.33\%$ . (iii) Leveraging both the Inter-session Bidirectional Prompt Loss and the Prompt Magnitude Normalization achieves a model performance of $75.12\%$ . + +$\mathcal{L}_{dis}$ vs. contrastive regularization: This study aims to verify the superiority of $\mathcal{L}_{dis}$ than Contrastive Regularization [64]. The Contrastive Regularization (CR) is designed to enhance the session-specific prompt differences, having the potential to reduce the overlap between old and new class spaces. In Tab. 4, we observe that our proposed loss significantly outperforms the contrastive regularization. This superiority stems from the fact that the contrastive loss is class-agnostic, not establishing a connection between prompt discrimination and class relationships. Besides, the contrastive regularization fails to alleviate the occurrence of the prompt-ambiguity overlap space. + +Loss design: This discussion aims to prove that bidirectional prompt optimization is vital for improving the model performance, and the symmetric loss works. For constructing $\mathcal{L}_{dis}$ , we alter the labels of ID-noisy samples to pseudo old labels and design the symmetric loss to optimize prompts. In this discussion, + +Table 4: The comparison (\%) between different losses for verifying the superiority of our proposed loss than simply enlarging the inter-session prompt difference. CR is a regularization term used in [64] to enlarge the inter-session prompt difference. In all experiments, the Prompt Magnitude Normalization is used. + +
Baseline w/o CR+CR+Ldis+CR & Ldis
FAA71.4771.9374.9075.12
CAA75.7375.8378.2978.35
+ +Table 5: The comparison (\%) between different loss designs for verifying the necessity of the bidirectional prompt optimization. "w/o old" means not using pseudo old labels for ID-noisy samples and "w old" means using pseudo old labels for ID-noisy samples. In all experiments, the Prompt Magnitude Normalization is used. + +
Baseline+Lce dis+Ldis
w/o oldw oldw old
FAA71.9371.5073.0775.12
CAA75.8376.0576.8178.35
+ +we formulate the cross-entropy loss concerning both old and new classes as: + +$$ +\mathcal {L} _ {d i s} ^ {c e} \left(\boldsymbol {x} _ {i} ^ {b}\right) = - \sum_ {k = 1} ^ {K} q \left(k \mid \boldsymbol {x} _ {i} ^ {b}\right) \cdot \log \left(p _ {d i s} ^ {b} \left(k \mid \boldsymbol {x} _ {i} ^ {b}\right)\right), \tag {8} +$$ + +where old class logits are not set to -inf. The analysis from Tab. 5 yields several key observations: (i) After adding $\mathcal{L}_{dis}^{ce}$ but not using pseudo old labels, the model performance reduces from 71.93 to 71.50. This suggests that undirectionally enhancing inter-session prompt discrimination is ineffective, implying a potential enlargement of the prompt-ambiguity overlap space. (ii) Compared to $+ \mathcal{L}_{dis}^{ce}$ without old", $+ \mathcal{L}_{dis}^{ce}$ with old" results in a $1.57\%$ higher model performance. Thus, bidirectionally optimizing prompts is vital for diminishing the class overlap. (iii) After using pseudo old labels, than adding $\mathcal{L}_{dis}^{ce}$ , adding $\mathcal{L}_{dis}$ improves the model performance from 73.07 to 75.12. This indicates that the symmetric loss can mitigate the negative impact brought by using pseudo old labels. + +Hyper-parameter analysis: We investigate the effects of the hyperparameter $\alpha$ . More specifically, we explore a range of values for $\alpha$ , namely $\{0.0, 0.2, 0.4, 0.6, 1.0\}$ . In the results presented in Tab. 6, it is evident that as the value of $\alpha$ increases, there is a corresponding gradual improvement in performance. This improvement plateaus when $\alpha$ attains a value of 0.4. + +Statistical analysis: This statistic aims to use inter-session misclassification to quantify the old and new class space overlap. In this analysis, without losing the generality, we focus on the predictions for test samples from session 0 (the first session) after learning all sessions. The total test sample count is 650 in session 0. For the predictions from the baseline (HiDe-Prompt), we compute the count of samples that are misclassified as classes from other sessions (session 0), + +Table 6: Impacts (%) of $\alpha$ on Split ImageNet-R. + +
0.00.20.40.61.0
FAA71.9374.3375.1274.1273.58
CAA75.8377.8578.3577.6377.44
+ +![](images/1ffa2e08b20051eadf540f92a9dd523c8a07ef7d79c68a93ff9c8acebde2fbf6.jpg) +Fig. 3: The analysis about misclassification. (a) We analyze which type of misclassification is the main type for the baseline. Inter-session misclassification means that a sample is misclassified as a class from other sessions. (b) We quantify the solving of our method for the inter-session misclassification. + +![](images/9aab9bb8f19d08373154688ce97f6b27ccf1e5b4ff25be814ba6994559ce6fa9.jpg) + +termed as inter-session (intra-session) misclassification type, as shown in Fig. 3 (a). Then, as shown in Fig. 3 (b), focusing on inter-session misclassification, we compute the count of samples that are misclassified in each session to compare the baseline and our method. Fig. 3 reveals two key observations: $(i)$ For the baseline, 173 $(26.62\%)$ test samples are misclassified as classes in another session, and only 31 $(4.77\%)$ are misclassified as classes from session 0. This shows that the inter-session misclassification is the main problem for existing methods. $(ii)$ Than the baseline, our method significantly decreases the inter-session misclassification sample count from 173 to 67. This shows that our method can effectively solve the inter-session misclassification problem. + +# 5 Conclusion + +Our study reveals that the optimization objective of prompts lacks the explicit modeling of inter-session class relationships, not effectively separating new and old class spaces. To minimize the overlap between new and old class spaces for creating the clear inter-session class margins, we introduce an Inter-session Bidirectional Prompt Loss to place new and old samples to discriminative regions, while eliminating the prompt-ambiguity overlap space. Furthermore, we design the Prompt Magnitude Normalization to enhance the prompt tuning effectiveness. Extensive experiments demonstrate the superiority of our approach across different datasets and continual learning scenarios. + +Acknowledgements. This work is supported in part by National Natural Science Foundation of China under Grant No.62441602, National Science Foundation for Distinguished Young Scholars under Grant 62225605, Project 12326608 supported by NSFC, National Natural Science Foundation of China under Grant U20A20222, Zhejiang Key Research and Development Program under Grant 2023C03196, the Ningbo Science and Technology Innovation Project (No.2024Z294), the Fundamental Research Funds for the Central Universities, the Key R&D Program of Zhejiang Province, China 2023C01043, Ant Group through CCF-Ant Research Fund, and sponsored by CCF-AFSG Research Fund. + +# References + +1. Aljundi, R., Lin, M., Goujaud, B., Bengio, Y.: Gradient based sample selection for online continual learning. Advances in neural information processing systems 32 (2019) +2. Barbu, A., Mayo, D., Alverio, J., Luo, W., Wang, C., Gutfreund, D., Tenenbaum, J., Katz, B.: Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems 32 (2019) +3. Belouadah, E., Popescu, A.: II2m: Class incremental learning with dual memory. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 583-592 (2019) +4. Bulatov, Y.: Notmnist dataset (2011) +5. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021) +6. Cha, S., Cho, S., Hwang, D., Hong, S., Lee, M., Moon, T.: Rebalancing batch normalization for exemplar-based class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20127-20136 (2023) +7. Chaudhry, A., Dokania, P.K., Ajanthan, T., Torr, P.H.: Riemannian walk for incremental learning: Understanding forgetting and intransigence. In: Proceedings of the European conference on computer vision (ECCV). pp. 532-547 (2018) +8. Chaudhry, A., Ranzato, M., Rohrbach, M., Elhoseiny, M.: Efficient lifelong learning with a-gem. In: International Conference on Learning Representations (2018) +9. Chen, X., Xie, S., He, K.: An empirical study of training self-supervised vision transformers. in 2021 IEEE. In: CVF International Conference on Computer Vision (ICCV). pp. 9620-9629 +10. Dong, J., Li, H., Cong, Y., Sun, G., Zhang, Y., Van Gool, L.: No one left behind: Real-world federated class-incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence 46(4), 2054-2070 (2024). https://doi.org/10.1109/TPAMI.2023.3334213 +11. Dong, J., Liang, W., Cong, Y., Sun, G.: Heterogeneous forgetting compensation for class-incremental learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 11742-11751 (October 2023) +12. Dong, J., Wang, L., Fang, Z., Sun, G., Xu, S., Wang, X., Zhu, Q.: Federated class-incremental learning. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2022) + +13. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020) +14. Douillard, A., Cord, M., Ollion, C., Robert, T., Valle, E.: Podnet: Pooled outputs distillation for small-tasks incremental learning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16. pp. 86-102. Springer (2020) +15. Douillard, A., Ramé, A., Couairon, G., Cord, M.: Dytox: Transformers for continual learning with dynamic token expansion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9285-9295 (2022) +16. Ebrahimi, S., Meier, F., Calandra, R., Darrell, T., Rohrbach, M.: Adversarial continual learning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16. pp. 386-402. Springer (2020) +17. French, R.M.: Catastrophic forgetting in connectionist networks. Trends in cognitive sciences 3(4), 128-135 (1999) +18. Gao, Q., Zhao, C., Sun, Y., Xi, T., Zhang, G., Ghanem, B., Zhang, J.: A unified continual learning framework with general parameter-efficient tuning. arXiv preprint arXiv:2303.10070 (2023) +19. Ge, Y., Li, Y., Ni, S., Zhao, J., Yang, M.H., Itti, L.: Clr: Channel-wise lightweight reprogramming for continual learning. arXiv preprint arXiv:2307.11386 (2023) +20. Gomes, H.M., Barddal, J.P., Enembreck, F., Bifet, A.: A survey on ensemble learning for data stream classification. ACM Computing Surveys (CSUR) 50(2), 1-36 (2017) +21. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +22. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8340-8349 (2021) +23. Hendrycks, D., Zhao, K., Basart, S., Steinhardt, J., Song, D.: Natural adversarial examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15262-15271 (2021) +24. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015) +25. Hou, S., Pan, X., Loy, C.C., Wang, Z., Lin, D.: Learning a unified classifier incrementally via rebalancing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 831-839 (2019) +26. Hu, Z., Li, Y., Lyu, J., Gao, D., Vasconcelos, N.: Dense network expansion for class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11858-11867 (2023) +27. Huang, B., Chen, Z., Zhou, P., Chen, J., Wu, Z.: Resolving task confusion in dynamic expansion architectures for class incremental learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 908-916 (2023) +28. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII. pp. 709-727. Springer (2022) + +29. Jung, D., Han, D., Bang, J., Song, H.: Generating instance-level prompts for rehearsal-free continual learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11847-11857 (2023) +30. Khan, M.G.Z.A., Naeem, M.F., Van Gool, L., Stricker, D., Tombari, F., Afzal, M.Z.: Introducing language guidance in prompt-based continual learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11463-11473 (2023) +31. Khattak, M.U., Wasim, S.T., Naseer, M., Khan, S., Yang, M.H., Khan, F.S.: Self-regulating prompts: Foundational model adaptation without forgetting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15190-15200 (2023) +32. Kim, D., Han, B.: On the stability-plasticity dilemma of class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20196-20204 (2023) +33. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) +34. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (1998) +35. Li, Z., Hoiem, D.: Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence 40(12), 2935-2947 (2017) +36. Liu, Y., Li, Y., Schiele, B., Sun, Q.: Online hyperparameter optimization for class-incremental learning. Proceedings of the AAAI Conference on Artificial Intelligence (2023) +37. Liu, Y., Schiele, B., Sun, Q.: Adaptive aggregation networks for class-incremental learning. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 2544-2553 (2021) +38. Liu, Y., Su, Y., Liu, A.A., Schiele, B., Sun, Q.: Mnemonics training: Multi-class incremental learning without forgetting. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 12245-12254 (2020) +39. Luo, Z., Liu, Y., Schiele, B., Sun, Q.: Class-incremental exemplar compression for class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11371-11380 (2023) +40. McDonnell, M.D., Gong, D., Parvaneh, A., Abbasnejad, E., van den Hengel, A.: Ranpac: Random projections and pre-trained models for continual learning. Advances in Neural Information Processing Systems 36 (2024) +41. Mehta, S.V., Patil, D., Chandar, S., Strubell, E.: An empirical investigation of the role of pre-training in lifelong learning. Journal of Machine Learning Research 24(214), 1-50 (2023) +42. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011) +43. Panos, A., Kobe, Y., Reino, D.O., Aljundi, R., Turner, R.E.: First session adaptation: A strong replay-free baseline for class-incremental learning. arXiv preprint arXiv:2303.13199 (2023) +44. Petit, G., Popescu, A., Schindler, H., Picard, D., Delezoide, B.: Fetril: Feature translation for exemplar-free class-incremental learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3911-3920 (2023) +45. Pham, Q., Liu, C., Hoi, S.: Continual normalization: Rethinking batch normalization for online continual learning. In: International Conference on Learning Representations (2022) + +46. Qiang, S., Hou, J., Wan, J., Liang, Y., Lei, Z., Zhang, D.: Mixture uniform distribution modeling and asymmetric mix distillation for class incremental learning. In: Proceedings of the AAAI Conference on Artificial Intelligence (2023) +47. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +48. Ramasesh, V.V., Lewkowycz, A., Dyer, E.: Effect of scale on catastrophic forgetting in neural networks. In: International Conference on Learning Representations (2021) +49. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 2001-2010 (2017) +50. Rymarczyk, D., van de Weijer, J., Zielinski, B., Twardowski, B.: Icicle: Interpretable class incremental continual learning. arXiv preprint arXiv:2303.07811 (2023) +51. Sarfraz, F., Arani, E., Zonooz, B.: Sparse coding in a dual memory system for lifelong learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 9714-9722 (2023) +52. Shi, Y., Zhou, K., Liang, J., Jiang, Z., Feng, J., Torr, P.H., Bai, S., Tan, V.Y.: Mimicking the oracle: an initial phase decorrelation approach for class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16722-16731 (2022) +53. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014) +54. Singh, P., Li, Y., Sikarwar, A., Lei, S.W., Gao, D., Talbot, M.B., Sun, Y., Shou, M.Z., Kreiman, G., Zhang, M.: Learning to learn: How to continuously teach humans and machines. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11708-11719 (2023) +55. Smith, J.S., Karlinsky, L., Gutta, V., Cascante-Bonilla, P., Kim, D., Arbelle, A., Panda, R., Feris, R., Kira, Z.: Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11909-11919 (2023) +56. Sun, Z., Mu, Y., Hua, G.: Regularizing second-order influences for continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20166-20175 (2023) +57. Talbot, M.B., Zawar, R., Badkundri, R., Zhang, M., Kreiman, G.: Tuned compositional feature replays for efficient stream learning. IEEE Transactions on Neural Networks and Learning Systems (2023) +58. Tan, M., Pang, R., Le, Q.V.: Efficientdet: Scalable and efficient object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10781-10790 (2020) +59. Tang, Y.M., Peng, Y.X., Zheng, W.S.: When prompt-based incremental learning does not meet strong pretraining. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1706-1716 (2023) +60. Villa, A., Alcázar, J.L., Alfarra, M., Alhamoud, K., Hurtado, J., Heilbron, F.C., Soto, A., Ghanem, B.: Pivot: Prompting for video continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24214-24223 (2023) + +61. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011) +62. Wang, F.Y., Zhou, D.W., Liu, L., Ye, H.J., Bian, Y., Zhan, D.C., Zhao, P.: Beef: Bi-compatible class-incremental learning via energy-based expansion and fusion. In: The Eleventh International Conference on Learning Representations (2023) +63. Wang, F.Y., Zhou, D.W., Ye, H.J., Zhan, D.C.: Foster: Feature boosting and compression for class-incremental learning. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXV. pp. 398-414. Springer (2022) +64. Wang, L., Xie, J., Zhang, X., Huang, M., Su, H., Zhu, J.: Hierarchical decomposition of prompt-based continual learning: Rethinking obscured sub-optimality. In: Advances in neural information processing systems (2023) +65. Wang, L., Zhang, X., Su, H., Zhu, J.: A comprehensive survey of continual learning: Theory, method and application. arXiv preprint arXiv:2302.00487 (2023) +66. Wang, W., Hu, Y., Chen, Q., Zhang, Y.: Task difficulty aware parameter allocation & regularization for lifelong learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7776-7785 (2023) +67. Wang, Y., Huang, Z., Hong, X.: S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. Annual Conference on Neural Information Processing Systems (2022) +68. Wang, Y., Ma, X., Chen, Z., Luo, Y., Yi, J., Bailey, J.: Symmetric cross entropy for robust learning with noisy labels. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 322-330 (2019) +69. Wang, Z., Zhao, Z., Wu, Y., Zhou, L., Xu, D.: Progressive target-styled feature augmentation for unsupervised domain adaptation on point clouds. In: ECCV (2024) +70. Wang, Z., Zhang, Z., Ebrahimi, S., Sun, R., Zhang, H., Lee, C.Y., Ren, X., Su, G., Perot, V., Dy, J., et al.: Dualprompt: Complementary prompting for rehearsal-free continual learning. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVI. pp. 631-648. Springer (2022) +71. Wang, Z., Zhang, Z., Lee, C.Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., Pfister, T.: Learning to prompt for continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 139-149 (2022) +72. Wu, X., Wang, H., Wu, Y., Li, X.: D3t-gan: Data-dependent domain transfer gans for image generation with limited data. ACM Trans. Multimedia Comput. Commun. Appl. 19(4) (mar 2023). https://doi.org/10.1145/3576858, https://doi.org/10.1145/3576858 +73. Wu, Y., Li, H., Wang, F., Zhang, Y., Liang, R.: Self-distilled dynamic fusion network for language-based fashion retrieval. In: ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 3260-3264. IEEE (2024) +74. Wu, Y., Wu, X., Li, X., Tian, J.: Mgh: Metadata guided hypergraph modeling for unsupervised person re-identification. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 1571–1580 (2021) +75. Wu, Y., Chen, Y., Wang, L., Ye, Y., Liu, Z., Guo, Y., Fu, Y.: Large scale incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 374-382 (2019) +76. Xiao, H., Rasul, K., Vollgraf, R.: Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747 (2017) + +77. Xu, J., Zhu, Z.: Reinforced continual learning. Advances in Neural Information Processing Systems 31 (2018) +78. Yan, S., Xie, J., He, X.: Der: Dynamically expandable representation for class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3014-3023 (2021) +79. Yoon, J., Yang, E., Lee, J., Hwang, S.J.: Lifelong learning with dynamically expandable networks. In: International Conference on Learning Representations (2018) +80. Yu, L., Twardowski, B., Liu, X., Herranz, L., Wang, K., Cheng, Y., Jui, S., Weijer, J.v.d.: Semantic drift compensation for class-incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6982-6991 (2020) +81. Zhang, G., Wang, L., Kang, G., Chen, L., Wei, Y.: Slow learner with classifier alignment for continual learning on a pre-trained model. arXiv preprint arXiv:2303.05118 (2023) +82. Zhang, J., Zhang, J., Ghosh, S., Li, D., Tasci, S., Heck, L., Zhang, H., Kuo, C.C.J.: Class-incremental learning via deep model consolidation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1131-1140 (2020) +83. Zhao, B., Xiao, X., Gan, G., Zhang, B., Xia, S.T.: Maintaining discrimination and fairness in class incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13208-13217 (2020) +84. Zhao, H., Fu, Y., Kang, M., Tian, Q., Wu, F., Li, X.: Mgmt: Multi-grained slow vs. fast framework for few-shot class-incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021) +85. Zhao, H., Wang, H., Fu, Y., Wu, F., Li, X.: Memory-efficient class-incremental learning for image classification. IEEE Transactions on Neural Networks and Learning Systems 33(10), 5966-5977 (2021) +86. Zhou, D.W., Wang, F.Y., Ye, H.J., Ma, L., Pu, S., Zhan, D.C.: Forward compatible few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9046-9056 (2022) +87. Zhou, D.W., Wang, Q.W., Ye, H.J., Zhan, D.C.: A model or 603 exemplars: Towards memory-efficient class-incremental learning. In: International Conference on Learning Representations (2023) +88. Zhou, D.W., Ye, H.J., Ma, L., Xie, D., Pu, S., Zhan, D.C.: Few-shot class incremental learning by sampling multi-phase tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022) +89. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022) +90. Zhuang, H., Weng, Z., Wei, H., Xie, R., Toh, K.A., Lin, Z.: Acil: Analytic class incremental learning with absolute memorization and privacy protection. Advances in Neural Information Processing Systems 35, 11602-11614 (2022) \ No newline at end of file diff --git a/2024/RCS-Prompt_ Learning Prompt to Rearrange Class Space for Prompt-based Continual Learning/images.zip b/2024/RCS-Prompt_ Learning Prompt to Rearrange Class Space for Prompt-based Continual Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bc6b3c71e09d579e82e69e7346c41881cf1e1e07 --- /dev/null +++ b/2024/RCS-Prompt_ Learning Prompt to Rearrange Class Space for Prompt-based Continual Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69a8f57c1a5fc3b7fd69c8319c6b7a66006c21b815ebed3f55e3c5808929a2a9 +size 437412 diff --git a/2024/RCS-Prompt_ Learning Prompt to Rearrange Class Space for Prompt-based Continual Learning/layout.json b/2024/RCS-Prompt_ Learning Prompt to Rearrange Class Space for Prompt-based Continual Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8d135a9095c3b92ff2798022f67ad92974fb064e --- /dev/null +++ b/2024/RCS-Prompt_ Learning Prompt to Rearrange Class Space for Prompt-based Continual Learning/layout.json @@ -0,0 +1,11546 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 112, + 482, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 112, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 112, + 482, + 148 + ], + "type": "text", + "content": "RCS-Prompt: Learning Prompt to Rearrange Class Space for Prompt-based Continual Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 142, + 167, + 470, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 167, + 470, + 179 + ], + "spans": [ + { + "bbox": [ + 142, + 167, + 470, + 179 + ], + "type": "text", + "content": "Longrong Yang" + }, + { + "bbox": [ + 142, + 167, + 470, + 179 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 167, + 470, + 179 + ], + "type": "text", + "content": ", Hanbin Zhao" + }, + { + "bbox": [ + 142, + 167, + 470, + 179 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 167, + 470, + 179 + ], + "type": "text", + "content": ", Yunlong " + }, + { + "bbox": [ + 142, + 167, + 470, + 179 + ], + "type": "inline_equation", + "content": "\\mathrm{Yu}^{2\\star}" + }, + { + "bbox": [ + 142, + 167, + 470, + 179 + ], + "type": "text", + "content": ", Xiaodong Zeng" + }, + { + "bbox": [ + 142, + 167, + 470, + 179 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 142, + 167, + 470, + 179 + ], + "type": "text", + "content": ", and Xi Li" + }, + { + "bbox": [ + 142, + 167, + 470, + 179 + ], + "type": "inline_equation", + "content": "^{1,4,5\\star}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 175, + 189, + 438, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 189, + 438, + 200 + ], + "spans": [ + { + "bbox": [ + 175, + 189, + 438, + 200 + ], + "type": "text", + "content": "1College of Computer Science & Technology, Zhejiang University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 481, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 481, + 221 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 481, + 221 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 130, + 200, + 481, + 221 + ], + "type": "text", + "content": "College of Information Science and Electronic Engineering, Zhejiang University " + }, + { + "bbox": [ + 130, + 200, + 481, + 221 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 130, + 200, + 481, + 221 + ], + "type": "text", + "content": "Ant Services Group " + }, + { + "bbox": [ + 130, + 200, + 481, + 221 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 130, + 200, + 481, + 221 + ], + "type": "text", + "content": "Shanghai Institute for Advanced Study of Zhejiang University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 157, + 221, + 454, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 221, + 454, + 233 + ], + "spans": [ + { + "bbox": [ + 157, + 221, + 454, + 233 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 157, + 221, + 454, + 233 + ], + "type": "text", + "content": "Zhejiang - Singapore Innovation and AI Joint Research Lab, Hangzhou" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 172, + 233, + 441, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 233, + 441, + 244 + ], + "spans": [ + { + "bbox": [ + 172, + 233, + 441, + 244 + ], + "type": "text", + "content": "{longrongyang, zhaohanbin, yuyunlong, xilizju}@zju.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 243, + 245, + 372, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 245, + 372, + 255 + ], + "spans": [ + { + "bbox": [ + 243, + 245, + 372, + 255 + ], + "type": "text", + "content": "{xiaodong.zxd}@antgroup.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 286, + 455, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 286, + 455, + 506 + ], + "spans": [ + { + "bbox": [ + 159, + 286, + 455, + 506 + ], + "type": "text", + "content": "Abstract. Prompt-based Continual Learning is an emerging direction in leveraging pre-trained knowledge for downstream continual learning. While arriving at a new session, existing prompt-based continual learning methods usually adapt features from pre-trained models to new data by introducing prompts. However, these prompts lack an optimization objective explicitly modeling inter-session class relationships, thus failing to construct clear inter-session class margins. Moreover, some old samples use new prompts during inference, resulting in the prompt-ambiguity overlap space - a special situation where old and new class spaces overlap. To address these issues, we propose an innovative approach called RCS-Prompt to Rearrange Class Space by bidirectionally optimizing prompts. RCS-Prompt optimizes prompts to signify discriminative regions across different sessions in the class space. Additionally, it mitigates the prompt-ambiguity overlap space by altering the labels of a small subset of new samples to old classes and training them with a customized symmetric loss. The proposed method effectively reduces the overlap between old and new class spaces, thereby establishing clear inter-session class margins. We extensively evaluate RCS-Prompt on public datasets, demonstrating its effectiveness in prompt-based continual learning. Code is available at https://github.com/longrongyang/RCS-Prompt." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 516, + 453, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 516, + 453, + 539 + ], + "spans": [ + { + "bbox": [ + 159, + 516, + 453, + 539 + ], + "type": "text", + "content": "Keywords: Prompt-based Continual Learning " + }, + { + "bbox": [ + 159, + 516, + 453, + 539 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 516, + 453, + 539 + ], + "type": "text", + "content": " Continual Learning " + }, + { + "bbox": [ + 159, + 516, + 453, + 539 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 516, + 453, + 539 + ], + "type": "text", + "content": " Class Space Overlap " + }, + { + "bbox": [ + 159, + 516, + 453, + 539 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 516, + 453, + 539 + ], + "type": "text", + "content": " Class Margin" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 560, + 230, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 560, + 230, + 572 + ], + "spans": [ + { + "bbox": [ + 132, + 560, + 230, + 572 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": "While deep models have achieved remarkable success across diverse fields [21, 53, 58], most research has focused on recognizing a fixed number of classes in static environments. However, in real-world scenarios, models often encounter streaming data where new classes may emerge over time [20]. To address this issue, Continual Learning has been proposed, which allows models to learn from" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 164, + 145, + 201, + 198 + ], + "blocks": [ + { + "bbox": [ + 149, + 118, + 199, + 137 + ], + "lines": [ + { + "bbox": [ + 149, + 118, + 199, + 137 + ], + "spans": [ + { + "bbox": [ + 149, + 118, + 199, + 137 + ], + "type": "text", + "content": "(a) Without Prompts" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 164, + 145, + 201, + 198 + ], + "lines": [ + { + "bbox": [ + 164, + 145, + 201, + 198 + ], + "spans": [ + { + "bbox": [ + 164, + 145, + 201, + 198 + ], + "type": "image", + "image_path": "566a58d4abd0cbaa8f794e01c37e30db37dc8962cae970976550a16c97dd11fe.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 204, + 206, + 223 + ], + "lines": [ + { + "bbox": [ + 156, + 204, + 206, + 223 + ], + "spans": [ + { + "bbox": [ + 156, + 204, + 206, + 223 + ], + "type": "text", + "content": "Class Space Severe class overlap" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 220, + 118, + 270, + 203 + ], + "blocks": [ + { + "bbox": [ + 220, + 118, + 270, + 203 + ], + "lines": [ + { + "bbox": [ + 220, + 118, + 270, + 203 + ], + "spans": [ + { + "bbox": [ + 220, + 118, + 270, + 203 + ], + "type": "image", + "image_path": "795bccccf132810738dbe6cfdad1249efb30aeb81890d0831e25a195fd06f133.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 221, + 204, + 266, + 228 + ], + "lines": [ + { + "bbox": [ + 221, + 204, + 266, + 228 + ], + "spans": [ + { + "bbox": [ + 221, + 204, + 266, + 228 + ], + "type": "text", + "content": "Class Space Similar prompts Still class overlap" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 286, + 137, + 381, + 204 + ], + "blocks": [ + { + "bbox": [ + 308, + 119, + 354, + 127 + ], + "lines": [ + { + "bbox": [ + 308, + 119, + 354, + 127 + ], + "spans": [ + { + "bbox": [ + 308, + 119, + 354, + 127 + ], + "type": "text", + "content": "(c) Our Method" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 286, + 137, + 381, + 204 + ], + "lines": [ + { + "bbox": [ + 286, + 137, + 381, + 204 + ], + "spans": [ + { + "bbox": [ + 286, + 137, + 381, + 204 + ], + "type": "image", + "image_path": "ddc6742f07c5ad555a223a52095c41027896c60e6c027f220bc94747ea402801.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 301, + 205, + 362, + 228 + ], + "lines": [ + { + "bbox": [ + 301, + 205, + 362, + 228 + ], + "spans": [ + { + "bbox": [ + 301, + 205, + 362, + 228 + ], + "type": "text", + "content": "Class Space\nOptimize prompts to\nsolve severe class overlap" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 389, + 114, + 403, + 121 + ], + "blocks": [ + { + "bbox": [ + 389, + 114, + 403, + 121 + ], + "lines": [ + { + "bbox": [ + 389, + 114, + 403, + 121 + ], + "spans": [ + { + "bbox": [ + 389, + 114, + 403, + 121 + ], + "type": "image", + "image_path": "5671b9373790bd07156ffcc9f527dcd767839e52647ca3d23a8df7e956919494.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 406, + 114, + 461, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 114, + 461, + 121 + ], + "spans": [ + { + "bbox": [ + 406, + 114, + 461, + 121 + ], + "type": "text", + "content": "Samples in old session" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 388, + 125, + 403, + 132 + ], + "blocks": [ + { + "bbox": [ + 388, + 125, + 403, + 132 + ], + "lines": [ + { + "bbox": [ + 388, + 125, + 403, + 132 + ], + "spans": [ + { + "bbox": [ + 388, + 125, + 403, + 132 + ], + "type": "image", + "image_path": "e9cd74167318525308d2c60e9a1d3c98c9b237678d157c06ac2c494d1fb7e6b7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 408, + 125, + 463, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 125, + 463, + 132 + ], + "spans": [ + { + "bbox": [ + 408, + 125, + 463, + 132 + ], + "type": "text", + "content": "Samples in new session" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 388, + 137, + 403, + 148 + ], + "blocks": [ + { + "bbox": [ + 388, + 137, + 403, + 148 + ], + "lines": [ + { + "bbox": [ + 388, + 137, + 403, + 148 + ], + "spans": [ + { + "bbox": [ + 388, + 137, + 403, + 148 + ], + "type": "image", + "image_path": "1060f61ba3283f745c874b6eb585d76116432be4b895f22435d6622f8390fba1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 408, + 139, + 427, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 139, + 427, + 147 + ], + "spans": [ + { + "bbox": [ + 408, + 139, + 427, + 147 + ], + "type": "text", + "content": "Margin" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 386, + 152, + 404, + 162 + ], + "blocks": [ + { + "bbox": [ + 386, + 152, + 404, + 162 + ], + "lines": [ + { + "bbox": [ + 386, + 152, + 404, + 162 + ], + "spans": [ + { + "bbox": [ + 386, + 152, + 404, + 162 + ], + "type": "image", + "image_path": "28b8010d5277cabb582d639aa3627396e46833e4e3cef6413e776ebb1d8bcb0a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 408, + 154, + 453, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 154, + 453, + 160 + ], + "spans": [ + { + "bbox": [ + 408, + 154, + 453, + 160 + ], + "type": "text", + "content": "Decision boundary" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 386, + 168, + 405, + 178 + ], + "blocks": [ + { + "bbox": [ + 386, + 168, + 405, + 178 + ], + "lines": [ + { + "bbox": [ + 386, + 168, + 405, + 178 + ], + "spans": [ + { + "bbox": [ + 386, + 168, + 405, + 178 + ], + "type": "image", + "image_path": "1b6e490a7ae9ff9b11a2f25795973b0bb46ff6fcd9445b5dd0f222548b20f99b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 408, + 168, + 447, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 168, + 447, + 182 + ], + "spans": [ + { + "bbox": [ + 408, + 168, + 447, + 182 + ], + "type": "text", + "content": "Region signified by old prompts." + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 386, + 186, + 406, + 196 + ], + "blocks": [ + { + "bbox": [ + 386, + 186, + 406, + 196 + ], + "lines": [ + { + "bbox": [ + 386, + 186, + 406, + 196 + ], + "spans": [ + { + "bbox": [ + 386, + 186, + 406, + 196 + ], + "type": "image", + "image_path": "1e2f65c2609d816b9537d019603a3b80bfb2082fadf845abaf11a39e9105ed58.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 408, + 186, + 447, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 186, + 447, + 199 + ], + "spans": [ + { + "bbox": [ + 408, + 186, + 447, + 199 + ], + "type": "text", + "content": "Region signified by new prompts" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 386, + 205, + 400, + 213 + ], + "blocks": [ + { + "bbox": [ + 386, + 205, + 400, + 213 + ], + "lines": [ + { + "bbox": [ + 386, + 205, + 400, + 213 + ], + "spans": [ + { + "bbox": [ + 386, + 205, + 400, + 213 + ], + "type": "image", + "image_path": "22eddcde71182843eaa648a75b9d172626620852d8a201de0dee9997981071cb.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 251, + 482, + 327 + ], + "lines": [ + { + "bbox": [ + 131, + 251, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 131, + 251, + 482, + 327 + ], + "type": "text", + "content": "Fig. 1: Models are incrementally learned across different sessions, where classes are accurately classified within each session. However, old and new class spaces overlap, so class margins lose abilities to distinguish different classes after learning a new session. Prompts signify distinct regions within the class space, yet without explicit inter-session prompt modeling, they fail to separate old and new class spaces. We aim to optimize prompts to create discriminative region and diminish prompt-ambiguity overlap space simultaneously, thereby establishing clear inter-session class margins." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 408, + 203, + 451, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 203, + 451, + 232 + ], + "spans": [ + { + "bbox": [ + 408, + 203, + 451, + 232 + ], + "type": "text", + "content": "Old sample using new prompts Prompt-ambiguity overlap space" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 130, + 352, + 479, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 479, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 479, + 411 + ], + "type": "text", + "content": "continuously changing data (i.e., different sessions) and dynamically build a unified classifier (i.e., model the relationships among different classes). However, Continual Learning encounters a significant challenge known as catastrophic forgetting [17], which arises when models, after incorporating new class data in a new learning session, tend to forget the previously acquired knowledge." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 130, + 413, + 480, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 480, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 480, + 484 + ], + "type": "text", + "content": "To mitigate catastrophic forgetting, pre-training has been identified as a pivotal technique [41,43,48,65,81]. A recent innovation in this realm is the adoption of prompt-based methods. These methods freeze a pre-trained model and employ a few learnable prompt parameters to steer representation learning, referred to as Prompt-based Continual Learning methods [55,64,70,71], which exhibit distinct superiority in continual learning performance." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 130, + 486, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 481, + 640 + ], + "type": "text", + "content": "In prompt-based continual learning methods, effective prompt optimization plays a pivotal role in solving the catastrophic forgetting problem, particularly when confronting the overlap between old and new class spaces. Concretely, across continual learning sessions, different classes are isolated, fostering an asynchronous learning process between sessions. Therefore, models tend to overfit the data and create boundaries specific to the classes within each session. This results in significant overlap between the old and new class spaces, as depicted in Fig. 1(a). Consequently, the margins between old classes lose abilities to distinguish between different classes, leading to the catastrophic forgetting problem. As shown in Fig. 1 (b), exiting prompt-based methods [55,64,70,71] use prompts to adjust features in each session. In essence, the learnable prompts create regions in the class space and relocate samples to the regions, thus being crucial for reducing the overlap between old and new class spaces." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "However, in existing prompt-based continual learning methods [55,64,70,71], the optimization objective of prompts lacks explicit modeling for inter-session" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 252, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 252, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 252, + 102 + ], + "type": "text", + "content": "Longrong Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "class relationships. In a new session, prompts are optimized to adjust classification scores solely for classes within that session. Consequently, as shown in Fig. 1 (b), prompts lack the ability to effectively separate old and new class spaces. Additionally, some samples from old classes may use new prompts to adjust features during inference [64, 70]. In such instances, unilaterally optimizing new prompts to move samples away from old classes exacerbates the overlap. This phenomenon, termed the prompt-ambiguity overlap space, makes optimizing prompts to separate class spaces become more challenging. Another observed concern is the diminished efficacy of prompts when their magnitudes become notably smaller than features during optimization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 426 + ], + "type": "text", + "content": "To address the above issues, we propose a novel method called Learning Prompt to Rearrange Class Space (RCS-Prompt) for prompt-based continual learning. As depicted in Fig. 1 (c), our main idea is to bidirectionally optimize prompts, making prompts signify discriminative regions in the class space while suppressing the prompt-ambiguity overlap space. Specifically, we firstly construct the Inter-session Bidirectional Prompt Loss. During training, if new prompts focused by new samples exhibit reduced influence on old samples, we explicitly reduce the classification scores of these new samples on old classes for optimizing new prompts. Consequently, these new prompts signify discriminative regions in the class space. Conversely, when new prompts focused by new samples may be attended by old samples, we select a subset of such new samples and relabel them as old classes for learning. Furthermore, the symmetric loss is introduced to prevent the negative impact brought by relabelling. Therefore, the prompt-ambiguity overlap space is reduced. Secondly, we propose the Prompt Magnitude Normalization to align prompt and feature magnitudes before using prompts, ensuring effective prompt tuning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 427, + 482, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 427, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 427, + 482, + 462 + ], + "type": "text", + "content": "Our method can serve as a plug-in to use, enlarging inter-session class margins by minimizing the overlap between old and new class spaces. The contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 469, + 480, + 586 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 469, + 480, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 469, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 138, + 469, + 480, + 515 + ], + "type": "text", + "content": "- We propose a novel Inter-session Bidirectional Prompt Loss to optimize prompts, making session prompts signify discriminative regions and reducing the prompt-ambiguity overlap space. Thus, the overlap between old and new class spaces is minimized for constructing clear inter-session class margins." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 517, + 480, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 517, + 480, + 551 + ], + "spans": [ + { + "bbox": [ + 138, + 517, + 480, + 551 + ], + "type": "text", + "content": "- We additionally present the Prompt Magnitude Normalization, which aligns prompt and feature magnitudes before using prompts, ensuring prompts for effectively adjusting model outcomes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 552, + 480, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 552, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 138, + 552, + 480, + 586 + ], + "type": "text", + "content": "- Without rehearsal, the proposed method achieves significant performance improvements compared to existing prompt-based continual learning methods, establishing state-of-the-art performance." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 605, + 242, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 605, + 242, + 618 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 242, + 618 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Continual Learning: gradually integrates new concepts while preserving previous knowledge, roughly categorized into three groups. The first employs exemplars from old classes (i.e., rehearsal) to recover past knowledge [1,7,8,38,39,44," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 102 + ], + "type": "text", + "content": "RCS-Prompt" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "46, 51, 69, 72–74, 85, 90]. The second adjusts inductive bias through the regularization terms [3, 6, 10–12, 32, 36, 45, 50, 52, 56, 75, 80, 83, 88]. Some methods in this category use knowledge distillation [24] to align old and new model outputs, retaining old knowledge [14, 35, 49, 82]. Lastly, various approaches expand network capacity, including neuron-wise [77, 79], backbone-wise [37, 62, 63, 78, 81, 84, 87], token-wise [15, 67] and others [19, 26, 27, 66]. Some methods incorporate multiple techniques, e.g., [15] expands the network and uses rehearsal simultaneously. Furthermore, works on online continual learning, e.g., [54, 57], discuss the learning process without explicit task boundaries." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 482, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 320 + ], + "type": "text", + "content": "Some traditional continual learning methods [25, 27, 75, 86] ever discussed the old and new class space overlap. These methods mainly analyze the overlap problem based on the dynamically updating convolution neural networks, and they alleviate this problem by using old data for knowledge distillation or creating balanced training subsets between old and new sessions. In contrast, our method focuses on prompt-based continual learning, where prompts are special and vital, and it explicitly eliminates the overlap between old and new class spaces by solely using new data to optimizing new prompts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 319, + 492, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 492, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 492, + 498 + ], + "type": "text", + "content": "Prompt-based continual learning: gains attraction with the surge of pre-trained models [13,47,60,67,89], using prompts to adapt pre-trained models for sequential data with new classes. L2P [71] applies visual prompt tuning [28] on pre-trained Transformer [13], learning a prompt pool for instance-specific selection. DualPrompt [70] extends L2P with general and expert prompts. S-Prompt [67] constructs only task-specific prompts for each task. CODA-Prompt [55] improves prompt selection via attention. SLCA [81] explores slow learners. DAP [29] generates instance-level prompts for use. LAE [18] reformulates different fin-tuning techniques into the Parameter-Efficient-Tuning module. PromptSRC [31] explores the use of prompts for continually updating CLIP. APG [59] develops a learnable adaptive prompt generator for weak pre-trained models. LGCL [30] introduces language guidance at the task level in the prompt pool. RanPAC [40] exploits training-free random projectors and class-prototype accumulation. Recently, HiDe-Prompt [64] decomposes the sub-optimality into hierarchical components and proposes different losses to solve them." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 498, + 483, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 483, + 584 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 483, + 584 + ], + "type": "text", + "content": "While few approaches, e.g., HiDe-Prompt [64], suggest enhancing prompt disparity through regularization terms, their losses generally lack class-specificity characteristics, thereby failing to establish a connection between prompt disparity and inter-session class relationships. In contrast, our method optimizes prompts by explicitly adjusting classification logits to minimize overlap between old and new class spaces. Our approach effectively solves the prompt-ambiguity overlap space issue, a crucial concern overlooked by the existing methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 599, + 233, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 233, + 613 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 233, + 613 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 623, + 209, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 623, + 209, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 623, + 209, + 635 + ], + "type": "text", + "content": "3.1 Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Problem setup: This study focuses on prompt-based continual learning. Existing prompt-based continual learning methods [55, 64, 70, 71] typically develop a" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "type": "text", + "content": "Longrong Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 114, + 462, + 160 + ], + "blocks": [ + { + "bbox": [ + 151, + 114, + 462, + 160 + ], + "lines": [ + { + "bbox": [ + 151, + 114, + 462, + 160 + ], + "spans": [ + { + "bbox": [ + 151, + 114, + 462, + 160 + ], + "type": "image", + "image_path": "7b5c24b05e944e7f8b621e91ad3b73556fcbce5cebd6fe952dc9fcbaae41465d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 151, + 161, + 286, + 300 + ], + "blocks": [ + { + "bbox": [ + 151, + 161, + 286, + 300 + ], + "lines": [ + { + "bbox": [ + 151, + 161, + 286, + 300 + ], + "spans": [ + { + "bbox": [ + 151, + 161, + 286, + 300 + ], + "type": "image", + "image_path": "5028f922c50a1f9e2791ae4090e670ec0684240c6730a7c02329595327789649.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 319, + 482, + 397 + ], + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 397 + ], + "type": "text", + "content": "Fig. 2: (a) General Pipeline. Existing prompt-based continual learning methods usually need to generate and insert prompts, optimize prompts by the loss, and train a session indicator. During test, the session ID of each sample is inferred and used for select prompts. (b) New samples demand high logits for new classes and low logits for old classes, whereas old samples require low logits for new classes and high logits for old classes. (c) Our Inter-session Bidirectional Prompt Loss satisfies the two requirements simultaneously, reducing the overlap between old and new spaces." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 288, + 161, + 462, + 300 + ], + "blocks": [ + { + "bbox": [ + 288, + 161, + 462, + 300 + ], + "lines": [ + { + "bbox": [ + 288, + 161, + 462, + 300 + ], + "spans": [ + { + "bbox": [ + 288, + 161, + 462, + 300 + ], + "type": "image", + "image_path": "7f3a05069d3ec7446b37f55ed4732e822f8630d7069453c3f0115be6cb234a6f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "content": " unified classifier capable of learning from a dynamic data stream. The continual learning process involves a sequence of " + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "content": " learning sessions, denoted as " + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\{D^1, D^2, \\dots, D^B\\}" + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "content": ". Each learning session " + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "inline_equation", + "content": "D^b = (\\pmb{x}_i^b, y_i^b)_{i=1}^{n_b}" + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "content": " represents an incremental step with " + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "inline_equation", + "content": "n_b" + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "content": " instances. The training instance " + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^b" + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "content": " belongs to class " + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "inline_equation", + "content": "y_i^b \\in Y^b" + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "inline_equation", + "content": "Y^b" + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "content": " represents the label space of session " + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "content": ". Notably, the label sets " + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "inline_equation", + "content": "Y^b" + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "inline_equation", + "content": "Y^{b'}" + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "content": " are disjoint for " + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "inline_equation", + "content": "b \\neq b'" + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "content": ". In this study, " + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "inline_equation", + "content": "D^b" + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "content": " is only available at the session " + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "content": ", where old data are not retrieved for rehearsal. Besides, session IDs are unavailable during inference. For simplicity, we refer to samples in the new session as \"new samples\" and those in the old session as \"old samples\". Likewise, we use the abbreviations \"new classes\", \"old classes\", \"new prompts\", and \"old prompts\" for convenience." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Existing prompt-based continual learning methods [55,64,70,71] mainly contain a Pre-Trained Model (PTM) " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\phi(\\cdot)" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " and some learnable prompts " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{pt}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ". As shown in Fig. 2 (a), there are three key components in these methods, which are prompt generation, prompt optimization, and session ID inference, respectively. First, in each session, a set of session-specific prompts " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{pt} \\in \\mathbb{R}^{L_p \\times D}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " is generated, where " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "L_p" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " denote the prompt number and prompt dimensionality, respectively. " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{pt} = \\{\\pmb{pt}_1, \\dots, \\pmb{pt}_{L_p}\\}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ". They adjust features through cross attention. Concatenating feature tokens " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "(N \\times D)" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " with prompts " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "(L_p \\times D)" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " generates new tokens " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "((N + L_p) \\times D)" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " for attention, i.e., " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\phi(\\pmb{x}_i^b | \\pmb{pt})" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ". Then, existing prompt-based continual learning methods use the below classification loss for optimizing prompts:" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 102 + ], + "type": "text", + "content": "RCS-Prompt" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 227, + 125, + 386, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 125, + 386, + 152 + ], + "spans": [ + { + "bbox": [ + 227, + 125, + 386, + 152 + ], + "type": "interline_equation", + "content": "\\boldsymbol {z} ^ {b} (\\boldsymbol {x} _ {i} ^ {b}) = [ \\underbrace {- \\inf , \\cdots , - \\inf} _ {\\text {o l d s e s s i o n}}, \\underbrace {z _ {1} ^ {b} , \\cdots} _ {\\text {n e w s e s s i o n}} ],", + "image_path": "8010e5633110772de63c72065254908ef569cb750f7cee1918f11ccdfca7ac5d.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 227, + 155, + 481, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 155, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 227, + 155, + 481, + 171 + ], + "type": "interline_equation", + "content": "\\boldsymbol {p} ^ {b} \\left(\\boldsymbol {x} _ {i} ^ {b}\\right) = \\operatorname {s o f t m a x} \\left(\\boldsymbol {z} ^ {b} \\left(\\boldsymbol {x} _ {i} ^ {b}\\right)\\right), \\tag {1}", + "image_path": "77ff9e8b8c971b124b12e1d5b4aa5ab0074346707802f2e1c4d2d433c508afa8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 222, + 173, + 394, + 204 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 173, + 394, + 204 + ], + "spans": [ + { + "bbox": [ + 222, + 173, + 394, + 204 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c e} \\left(\\boldsymbol {x} _ {i} ^ {b}\\right) = - \\sum_ {k = 1} ^ {K} q \\left(k \\mid \\boldsymbol {x} _ {i} ^ {b}\\right) \\cdot \\log \\left(p ^ {b} \\left(k \\mid \\boldsymbol {x} _ {i} ^ {b}\\right)\\right),", + "image_path": "717396856ccbc7fa32f6a31697a2ff821dd5a11b865add407db6e0839f445129.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 209, + 482, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 209, + 482, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 209, + 482, + 341 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 209, + 482, + 341 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 209, + 482, + 341 + ], + "type": "text", + "content": " is the new session ID and " + }, + { + "bbox": [ + 130, + 209, + 482, + 341 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 209, + 482, + 341 + ], + "type": "text", + "content": " is the class number, " + }, + { + "bbox": [ + 130, + 209, + 482, + 341 + ], + "type": "inline_equation", + "content": "z^b(\\boldsymbol{x}_i^b)" + }, + { + "bbox": [ + 130, + 209, + 482, + 341 + ], + "type": "text", + "content": " indicates the logits of " + }, + { + "bbox": [ + 130, + 209, + 482, + 341 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_i^b" + }, + { + "bbox": [ + 130, + 209, + 482, + 341 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 209, + 482, + 341 + ], + "type": "inline_equation", + "content": "z_1^b" + }, + { + "bbox": [ + 130, + 209, + 482, + 341 + ], + "type": "text", + "content": " indicates the logit of " + }, + { + "bbox": [ + 130, + 209, + 482, + 341 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_i^b" + }, + { + "bbox": [ + 130, + 209, + 482, + 341 + ], + "type": "text", + "content": " on a new class. In continual learning, unlike standard supervised learning using cross-entropy loss, model predictions for old classes are typically set to -inf to separate them from training. " + }, + { + "bbox": [ + 130, + 209, + 482, + 341 + ], + "type": "inline_equation", + "content": "q(k|\\boldsymbol{x}_i^b)" + }, + { + "bbox": [ + 130, + 209, + 482, + 341 + ], + "type": "text", + "content": " represents the one-hot vector of the ground-truth label for sample " + }, + { + "bbox": [ + 130, + 209, + 482, + 341 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_i^b" + }, + { + "bbox": [ + 130, + 209, + 482, + 341 + ], + "type": "text", + "content": ". Lastly, during training, existing methods usually use features from the PTMs (i.e., PTM features, " + }, + { + "bbox": [ + 130, + 209, + 482, + 341 + ], + "type": "inline_equation", + "content": "\\phi(\\boldsymbol{x}_i^b)" + }, + { + "bbox": [ + 130, + 209, + 482, + 341 + ], + "type": "text", + "content": ") to train a session indicator. During inference, this session indicator is used to estimate the session ID of each sample. Based on the predicted session ID, samples select prompts from all old prompts " + }, + { + "bbox": [ + 130, + 209, + 482, + 341 + ], + "type": "inline_equation", + "content": "\\boldsymbol{p}\\boldsymbol{t}^{old}" + }, + { + "bbox": [ + 130, + 209, + 482, + 341 + ], + "type": "text", + "content": " and new prompts " + }, + { + "bbox": [ + 130, + 209, + 482, + 341 + ], + "type": "inline_equation", + "content": "\\boldsymbol{p}\\boldsymbol{t}" + }, + { + "bbox": [ + 130, + 209, + 482, + 341 + ], + "type": "text", + "content": ". Under this scenario, as shown in Fig. 2 (b), old samples may select new prompts, leading to the prompt-ambiguity overlap space." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 342, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 521 + ], + "type": "text", + "content": "RCS-Prompt: Our method focuses on prompt optimization. Specifically, asynchronous learning takes place between sessions in continual learning, resulting in significant overlap between old and new class spaces. In prompt-based continual learning methods, prompts are vital to signify specific regions within the class space, where samples are relocated. Different sets of prompts correspond to distinct regions, and within each set, diverse prompts can reposition samples to various positions within a region. However, existing methods tend to overlook the explicit modeling of inter-session class relationships. Consequently, prompt-defined regions still exhibit overlap, impeding the construction of clear class margins. To this end, we propose an innovative approach called RCS-Prompt to rearrange class space by learning prompts, aiming to minimize the overlap between old and new class spaces, as shown in Fig. 2 (c). In detail, RCS-Prompt comprises two modules: Inter-session Bidirectional Prompt Loss and Prompt Magnitude Normalization. The specifics of each module will be introduced in the following sections." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 538, + 360, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 538, + 360, + 551 + ], + "spans": [ + { + "bbox": [ + 132, + 538, + 360, + 551 + ], + "type": "text", + "content": "3.2 Inter-session Bidirectional Prompt Loss" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "As mentioned above, our primary goal is to optimize new prompts during training on the new session, for minimizing the old and new class space overlap. However, achieving this goal is non-trivial due to a special challenge: When using new prompts to establish a region significantly distant from old class spaces and relocating samples to this region, old samples may inadvertently choose some new prompts to adjust features. Consequently, as new prompts shift old samples into the new region, overlap may emerge between old and new class spaces, a phenomenon referred to as the prompt-ambiguity overlap space. This challenge fundamentally stems from the divergence in indicator predictions for samples" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 252, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 252, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 252, + 102 + ], + "type": "text", + "content": "Longrong Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 187 + ], + "type": "text", + "content": "within the same session, suggesting the potential to employ a session indicator for sample separation. Building on this basic idea, as shown in Fig. 2 (c), we propose the Inter-session Bidirectional Prompt Loss to bidirectionally optimize new prompts, for shifting new samples away from old class spaces (i.e., producing discriminative regions) while keeping old samples close to their original class spaces (i.e., depressing the prompt-ambiguity overlap space)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 331 + ], + "type": "text", + "content": "We use a session indicator to split samples, and design different losses to solve them respectively. Specifically, the first step is to generate a session indicator during training. Existing methods [64,70] use the frozen PTM features to train a session indicator, which is always changing during training. The changing session indicator may lead to fluctuating sample divisions, potentially increasing the convergence difficulty. Thus, before each training session, we first use the PTMs to extract the PTM features of all training samples within that session. Then, the PTM features of samples of each class are averaged, severing as the PTM feature prototype proto" + }, + { + "bbox": [ + 130, + 188, + 482, + 331 + ], + "type": "inline_equation", + "content": "_k" + }, + { + "bbox": [ + 130, + 188, + 482, + 331 + ], + "type": "text", + "content": ". We store the PTM feature prototype of each class, resulting in a total storage of " + }, + { + "bbox": [ + 130, + 188, + 482, + 331 + ], + "type": "inline_equation", + "content": "K \\times D" + }, + { + "bbox": [ + 130, + 188, + 482, + 331 + ], + "type": "text", + "content": ", which is merely the size of a classifier. During optimizing prompts, we compute the indicator score of each sample on all seen classes by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 208, + 335, + 480, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 335, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 208, + 335, + 480, + 363 + ], + "type": "interline_equation", + "content": "s _ {k} = \\cos \\left(\\phi \\left(\\boldsymbol {x} _ {i} ^ {b}\\right), p r o t o _ {k}\\right) = \\frac {\\phi \\left(\\boldsymbol {x} _ {i} ^ {b}\\right) ^ {T} \\cdot p r o t o _ {k}}{\\left\\| \\phi \\left(\\boldsymbol {x} _ {i} ^ {b}\\right) \\right\\| \\cdot \\left\\| p r o t o _ {k} \\right\\|}, \\tag {2}", + "image_path": "17c58b2e872a8cff26b3e4c6c355162b4f48bfe6244842eedb8889295f7747f5.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 366, + 480, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 480, + 402 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 366, + 480, + 402 + ], + "type": "inline_equation", + "content": "\\cos (\\cdot)" + }, + { + "bbox": [ + 130, + 366, + 480, + 402 + ], + "type": "text", + "content": " denotes the cosine similarity. For each training sample, we can determine the class with the highest indicator score and its associated session, serving as its predicted session ID." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 402, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 641 + ], + "type": "text", + "content": "The second step is categorizing new samples into two distinct groups. The first group is labeled as \"ID-clean\" samples, where new samples are predicted to belong to the new session during training. These ID-clean samples possess the PTM features that differ from old samples, indicating a reduced influence of new prompts on old samples during inference. Consequently, it becomes feasible to explicitly decrease the classification logits of these new samples for old classes for optimizing new prompts, as illustrated in Fig. 2 (c). This allows these new prompts to signify discriminative new regions, enabling the effective functioning of class margins. The second group is labeled as \"ID-noisy\" samples, where new samples are predicted to belong to old sessions during training. ID-noisy samples resemble old samples, suggesting that new prompts focused on them may be attended to by old samples during inference. Reducing the classification logits of these new samples for old classes would compel prompts to relocate old samples away from old class spaces, posing a risk of expanding the prompt-ambiguity overlap space. Therefore, we propose to relabel ID-noise samples to the old classes for learning, as depicted in Fig. 2 (c), to position old samples in the proximity of old class spaces during inference. In detail, for an ID-noisy sample " + }, + { + "bbox": [ + 130, + 402, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_i^b" + }, + { + "bbox": [ + 130, + 402, + 482, + 641 + ], + "type": "text", + "content": ", we manually assign its label to the old class with the highest indicator score, creating \"pseudo old labels\" " + }, + { + "bbox": [ + 130, + 402, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\bar{y}" + }, + { + "bbox": [ + 130, + 402, + 482, + 641 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 130, + 402, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\bar{y} = \\mathrm{argmax}(s_k)" + }, + { + "bbox": [ + 130, + 402, + 482, + 641 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 402, + 482, + 641 + ], + "type": "inline_equation", + "content": "s_k" + }, + { + "bbox": [ + 130, + 402, + 482, + 641 + ], + "type": "text", + "content": " denotes the fine-grained indicator score of a sample on the class " + }, + { + "bbox": [ + 130, + 402, + 482, + 641 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 402, + 482, + 641 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 402, + 482, + 641 + ], + "type": "inline_equation", + "content": "k \\in \\{1, \\dots, K\\}" + }, + { + "bbox": [ + 130, + 402, + 482, + 641 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "The third step is designing losses for distinct samples. Assigning the labels of new samples to old classes for learning may have a negative impact on new class" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 102 + ], + "type": "text", + "content": "RCS-Prompt" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "learning. To mitigate the negative effect, we propose to apply a symmetric loss to prevent new samples from excessively drifting toward old class spaces. The symmetric loss, e.g., the reverse cross entropy loss [68], is a family of loss which adaptively weighs clean and noisy sample gradients, preventing noisy samples from dominating the learning. By conducting a Gradient analysis below, we establish a connection between the symmetric loss and the challenge we currently face, revealing the potential to leverage the symmetric loss to prevent the negative impact of using pseudo old labels for new classes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 260 + ], + "type": "text", + "content": "In summary, in handling ID-clean samples, we maintain their original labels and explicitly decrease their classification logits on old classes. For ID-noisy samples, we relabel them as old classes and supervise them with the reverse cross-entropy loss. Thus, our Inter-session Bidirectional Prompt Loss is formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 168, + 268, + 310, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 268, + 310, + 297 + ], + "spans": [ + { + "bbox": [ + 168, + 268, + 310, + 297 + ], + "type": "interline_equation", + "content": "\\boldsymbol {z} _ {d i s} ^ {b} (\\boldsymbol {x} _ {i} ^ {b}) = [ \\underbrace {z _ {1} ^ {1} , z _ {2} ^ {1} , \\cdots} _ {\\text {o l d s e s s i o n s}}, \\underbrace {z _ {1} ^ {b} , \\cdots} _ {\\text {n e w s e s s i o n}} ],", + "image_path": "cc592720edba6c2ebfd1a6e429f2bebb07f986b46c9f8840e0b4942a9908d14a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 167, + 299, + 480, + 314 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 299, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 167, + 299, + 480, + 314 + ], + "type": "interline_equation", + "content": "\\boldsymbol {p} _ {d i s} ^ {b} \\left(\\boldsymbol {x} _ {i} ^ {b}\\right) = \\operatorname {s o f t m a x} \\left(\\boldsymbol {z} _ {d i s} ^ {b} \\left(\\boldsymbol {x} _ {i} ^ {b}\\right)\\right), \\tag {3}", + "image_path": "49731446f0f36b2723df6f4cb410f58a9ff9c258892f9d6fc07458741c0d98bd.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 167, + 316, + 446, + 347 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 316, + 446, + 347 + ], + "spans": [ + { + "bbox": [ + 167, + 316, + 446, + 347 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d i s} (\\boldsymbol {x} _ {i} ^ {b}) = \\left\\{ \\begin{array}{l l} - \\sum_ {k = 1} ^ {K} q (k | \\boldsymbol {x} _ {i} ^ {b}) \\cdot \\log (p _ {d i s} ^ {b} (k | \\boldsymbol {x} _ {i} ^ {b})) & \\text {i f} \\boldsymbol {x} _ {i} ^ {b} \\text {i s I D - c l e a n}, \\\\ - \\sum_ {k = 1} ^ {K} p _ {d i s} ^ {b} (k | \\boldsymbol {x} _ {i} ^ {b}) \\cdot \\log (\\bar {q} (k | \\boldsymbol {x} _ {i} ^ {b})) & \\text {i f} \\boldsymbol {x} _ {i} ^ {b} \\text {i s I D - n o i s y}, \\end{array} \\right.", + "image_path": "1019ce440c3c458b23c24659f2ca0ff7ac5e1d29cd923dbbe96f4b381fd880ec.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 351, + 482, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 482, + 375 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 482, + 375 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 351, + 482, + 375 + ], + "type": "inline_equation", + "content": "\\bar{q}(k|\\pmb{x})" + }, + { + "bbox": [ + 130, + 351, + 482, + 375 + ], + "type": "text", + "content": " represents the one-hot vector of the pseudo old labels, and ID-noisy samples use the reverse cross-entropy loss. The total loss is formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 230, + 384, + 480, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 384, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 230, + 384, + 480, + 397 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} \\left(\\boldsymbol {x} _ {i} ^ {b}\\right) = \\mathcal {L} _ {\\text {c e}} \\left(\\boldsymbol {x} _ {i} ^ {b}\\right) + \\alpha \\cdot \\mathcal {L} _ {\\text {d i s}} \\left(\\boldsymbol {x} _ {i} ^ {b}\\right), \\tag {4}", + "image_path": "ca17497a4a06bfd3c98a08b9c68160402998e1f27dd9c551e0f59660bc64afed.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 406, + 450, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 406, + 450, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 406, + 450, + 417 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 406, + 450, + 417 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 130, + 406, + 450, + 417 + ], + "type": "text", + "content": " is the standard classification loss and " + }, + { + "bbox": [ + 130, + 406, + 450, + 417 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 406, + 450, + 417 + ], + "type": "text", + "content": " is a hyper-parameter." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 418, + 482, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 501 + ], + "type": "text", + "content": "Gradient analysis: By the gradient analysis, our objective is to clarify the adaptive weighting mechanism of the reverse cross-entropy loss for ID-noisy samples, for preventing the negative impact of pseudo old labels on new classes. The gradients of the reverse cross-entropy loss and the cross-entropy loss have been derived in [68]. For brevity, we denote " + }, + { + "bbox": [ + 130, + 418, + 482, + 501 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 130, + 418, + 482, + 501 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 418, + 482, + 501 + ], + "type": "inline_equation", + "content": "q_k" + }, + { + "bbox": [ + 130, + 418, + 482, + 501 + ], + "type": "text", + "content": " as abbreviations for " + }, + { + "bbox": [ + 130, + 418, + 482, + 501 + ], + "type": "inline_equation", + "content": "p(k|\\pmb{x})" + }, + { + "bbox": [ + 130, + 418, + 482, + 501 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 418, + 482, + 501 + ], + "type": "inline_equation", + "content": "q(k|\\pmb{x})" + }, + { + "bbox": [ + 130, + 418, + 482, + 501 + ], + "type": "text", + "content": ". The gradient of the reverse cross-entropy loss for ID-noisy samples with respect to the logit " + }, + { + "bbox": [ + 130, + 418, + 482, + 501 + ], + "type": "inline_equation", + "content": "z_k" + }, + { + "bbox": [ + 130, + 418, + 482, + 501 + ], + "type": "text", + "content": " can be derived as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 223, + 510, + 480, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 510, + 480, + 540 + ], + "spans": [ + { + "bbox": [ + 223, + 510, + 480, + 540 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\mathcal {L} _ {d i s}}{\\partial z _ {k}} = \\left\\{ \\begin{array}{l l} A p _ {k} - A p _ {k} ^ {2} & \\text {i f} q _ {k} = q _ {y} = 1, \\\\ - A p _ {k} p _ {y} & \\text {i f} q _ {k} = 0, \\end{array} \\right. \\tag {5}", + "image_path": "c3bfc6c54417fbe1949d7972ca58a4f294dd2f91752b2f276368d748c1bf9334.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 548, + 482, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 482, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 482, + 561 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 548, + 482, + 561 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 548, + 482, + 561 + ], + "type": "text", + "content": " is a constant. The gradient of the cross-entropy loss " + }, + { + "bbox": [ + 130, + 548, + 482, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 130, + 548, + 482, + 561 + ], + "type": "text", + "content": " is indicated as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 234, + 569, + 480, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 569, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 234, + 569, + 480, + 601 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\mathcal {L} _ {c e}}{\\partial z _ {k}} = \\left\\{ \\begin{array}{l l} p _ {k} - 1 & \\text {i f} q _ {k} = q _ {y} = 1, \\\\ p _ {k} & \\text {i f} q _ {k} = 0. \\end{array} \\right. \\tag {6}", + "image_path": "a834f11825301e1dc10d177309738f8390596fc526d1364d20b19fccc0df90f1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "In the case of " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "q_{k} = q_{y}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "Ap_{k} - Ap_{k}^{2}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " serves as an adaptive weighting term based on " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "p_{k}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ". In detail, when " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "p_{k}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " is close to 0, the cross-entropy loss generates large gradients to compel samples to increase scores on the labeled class. Conversely, for the reverse cross-entropy loss, \" " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "p_{k}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " is close to 0\" (an interval approaching 0, e.g., [0.0, 0.5)) indicates that the label is unreliable, resulting in the loss" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "type": "text", + "content": "Longrong Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "decreasing its gradient to 0. When " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " approaches 1 (an interval approaching 1), the reverse cross-entropy loss becomes similar to the cross-entropy loss." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 260 + ], + "type": "text", + "content": "Then, we establish a connection between the gradient and our current scenario. Throughout the training, low scores of samples on pseudo old labels signify their low resemblance to old samples. This scenario corresponds to \" " + }, + { + "bbox": [ + 130, + 141, + 482, + 260 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 130, + 141, + 482, + 260 + ], + "type": "text", + "content": " is close to 0\" . In such cases, the cross-entropy loss generates large gradients to force samples to have high logits on pseudo old labels, while the reverse cross-entropy loss can decrease the gradients of these samples to prevent them from excessively drifting toward old class spaces. Conversely, high logits of samples on pseudo old labels signify their proximity to old samples. This scenario corresponds to \" " + }, + { + "bbox": [ + 130, + 141, + 482, + 260 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 130, + 141, + 482, + 260 + ], + "type": "text", + "content": " is close to 1\", and the reverse cross-entropy loss becomes similar to the cross-entropy loss, treating these samples as if they were old samples." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 279, + 332, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 279, + 332, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 279, + 332, + 292 + ], + "type": "text", + "content": "3.3 Prompt Magnitude Normalization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 300, + 480, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 480, + 384 + ], + "type": "text", + "content": "Prompts operate by adjusting features to influence model outputs. We find that when the magnitude of prompts is significantly lower than that of features, their efficacy in altering outputs diminishes. When the magnitude of prompts is significantly higher than that of features, the feature distribution may be severely destroyed. Therefore, in this study, Prompt Magnitude Normalization is proposed to align the magnitude of prompts and features before using prompts, for enhancing the effectiveness of prompt tuning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 385, + 481, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 481, + 422 + ], + "type": "text", + "content": "To achieve our goal, we directly normalize prompts. Specifically, we compute the mean " + }, + { + "bbox": [ + 130, + 385, + 481, + 422 + ], + "type": "inline_equation", + "content": "\\mu_{\\pmb{x}}" + }, + { + "bbox": [ + 130, + 385, + 481, + 422 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 130, + 385, + 481, + 422 + ], + "type": "inline_equation", + "content": "\\tau_{\\pmb{x}}" + }, + { + "bbox": [ + 130, + 385, + 481, + 422 + ], + "type": "text", + "content": " of each sample feature, along with the mean " + }, + { + "bbox": [ + 130, + 385, + 481, + 422 + ], + "type": "inline_equation", + "content": "\\mu_{\\pmb{pt}}" + }, + { + "bbox": [ + 130, + 385, + 481, + 422 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 130, + 385, + 481, + 422 + ], + "type": "inline_equation", + "content": "\\tau_{\\pmb{pt}}" + }, + { + "bbox": [ + 130, + 385, + 481, + 422 + ], + "type": "text", + "content": " of its corresponding prompt. Then, we normalize the prompt:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 252, + 432, + 481, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 432, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 252, + 432, + 481, + 456 + ], + "type": "interline_equation", + "content": "\\boldsymbol {p t} = \\frac {\\boldsymbol {p t} - \\mu_ {\\boldsymbol {p t}}}{\\tau_ {\\boldsymbol {p t}}} \\cdot \\tau_ {\\boldsymbol {x}} + \\mu_ {\\boldsymbol {x}}, \\tag {7}", + "image_path": "2d2e6277a7b3a781b061208fa1fc8335fae305f0dd078a7981d0a919e1a15276.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 464, + 481, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 481, + 489 + ], + "type": "text", + "content": "where this normalization ensures that the prompt and feature magnitudes are comparable before using prompts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 510, + 230, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 230, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 230, + 523 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 536, + 279, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 279, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 279, + 548 + ], + "type": "text", + "content": "4.1 Implementation Details" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Benchmark: This study uses the benchmark presented in the recent state-of-the-art method HiDe-Prompt [64], which mainly evaluates model performance on Split CIFAR-100 [33] and Split ImageNet-R [22]. Split CIFAR-100 features 100 classes evenly distributed across 10 incremental sessions. Split ImageNet-R includes 200 classes, encompassing challenging ImageNet examples. Similarly, these classes are evenly distributed across 10 incremental sessions. Then, the additional evaluation datasets in [64] include 5-Datasets and Split CUB-200. 5-Datasets [16] includes CIFAR-10 [33], MNIST [34], Fashion-MNIST [76], SVHN [42] and notMNIST [4] datasets, each treated as an incremental task to evaluate the impact of" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 102 + ], + "type": "text", + "content": "RCS-Prompt" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 157, + 483, + 462 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: The comparison (\\%) between our method and five prompt-based continual methods. Following [64], the results of seeds 40, 42, and 44 are averaged to report. HiDe-Prompt uses Multi-Centroid statistical modeling [64]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 157, + 483, + 462 + ], + "lines": [ + { + "bbox": [ + 132, + 157, + 483, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 157, + 483, + 462 + ], + "type": "table", + "html": "
PTMMethodSplit CIFAR-100Split ImageNet-R
FAA (↑)CAA (↑)FFM (↓)FAA (↑)CAA (↑)FFM (↓)
Sup-21KL2P [71]83.0688.256.5863.6567.257.51
S-Prompt++ [67]88.8192.253.8769.6872.503.29
CODA-Prompt [55]86.9491.574.0470.0374.265.17
DualPrompt [70]86.6090.644.4568.7971.964.49
+RCS-Prompt (Ours)90.5394.214.0870.8074.384.69
HiDe-Prompt [64]92.9294.600.8673.5575.930.95
+RCS-Prompt (Ours)93.3495.340.8274.8577.520.91
DINO-1KL2P [71]70.6579.029.4657.4062.563.58
S-Prompt++ [67]74.9783.827.7857.6463.795.08
CODA-Prompt [55]77.5084.818.1063.1569.736.86
DualPrompt [70]74.9083.984.4558.5764.895.80
+RCS-Prompt (Ours)81.4489.324.2667.4171.735.10
HiDe-Prompt [64]90.0692.922.1369.3472.351.60
+RCS-Prompt (Ours)92.1194.091.9374.5478.022.41
MoCo-1KL2P [71]74.8583.146.5151.6458.872.37
S-Prompt++ [67]76.3083.8814.6753.1560.034.11
CODA-Prompt [55]76.8384.9712.6055.7565.4910.46
DualPrompt [70]77.7785.316.6152.5760.652.73
+RCS-Prompt (Ours)82.9189.374.6361.4165.572.61
HiDe-Prompt [64]90.7093.231.5663.0566.901.78
+RCS-Prompt (Ours)92.0093.651.5568.2471.041.74
", + "image_path": "91b341e5bb5234d095ffb1db0694785feff74e80c9e20d54ba535bcd738ccf72.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 484, + 481, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 481, + 509 + ], + "type": "text", + "content": "large inter-session differences. Split CUB-200 [61] includes 200-class fine-grained images of birds, randomly split into 10 incremental sessions of disjoint classes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "type": "text", + "content": "Baselines: On the evaluation benchmark, we extensively compare our method to recent prompt-based continual learning methods: L2P [71], S-Prompt++ [67], DualPrompt [70], CODA-Prompt [55] and HiDe-Prompt [64]. In experiments, we mainly add our method to HiDe-Prompt [64]. Extending upon this baseline, we add our proposed Inter-session Bidirectional Prompt Loss during training and apply the Prompt Magnitude Normalization before using prompts. Others keep unchanging. Besides, we provide results adding our method to DualPrompt [70], for showing the flexibility of our method as a plug-in." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 641 + ], + "type": "text", + "content": "Training details: Following HiDe-Prompt [64], we select ViT-B/16 Sup-21K, DINO-1K [5], and MoCo-1K [9] as the pre-trained models to conduct experiments. We set " + }, + { + "bbox": [ + 130, + 605, + 481, + 641 + ], + "type": "inline_equation", + "content": "\\alpha = 0.4" + }, + { + "bbox": [ + 130, + 605, + 481, + 641 + ], + "type": "text", + "content": ". More details are provided in the Sec. 3 of supp." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Evaluation protocol: We record the average accuracy of all seen classes after learning each session, presenting the last one as the final average accuracy" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 252, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 252, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 252, + 102 + ], + "type": "text", + "content": "Longrong Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 155, + 146, + 459, + 254 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 138 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 138 + ], + "type": "text", + "content": "Table 2: The comparison (\\%) between our method and five prompt-based continual learning methods on additional datasets. Following [64], the FAA metrics are report." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 155, + 146, + 459, + 254 + ], + "lines": [ + { + "bbox": [ + 155, + 146, + 459, + 254 + ], + "spans": [ + { + "bbox": [ + 155, + 146, + 459, + 254 + ], + "type": "table", + "html": "
Method5-DatasetsSplitCUB-200
Sup-21KiBOT-21KSup-21KiBOT-21K
L2P [71]81.8482.2574.4844.29
DualPrompt [70]77.9168.0382.0541.31
S-Prompt++ [67]86.0677.2082.0842.73
CODA-Prompt [55]64.1851.6574.3447.79
HiDe-Prompt [64]93.8394.8886.5678.23
+RCS-Prompt (Ours)94.5196.1187.3279.42
", + "image_path": "e4517a4f9089d6871cebfe597743d2d4c68600a5ba65d8d73648bd86ca8f461b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 276, + 482, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 276, + 482, + 313 + ], + "spans": [ + { + "bbox": [ + 130, + 276, + 482, + 313 + ], + "type": "text", + "content": "(FAA) and their historical average as the cumulative average accuracy (CAA). We also present the final forgetting measure (FFM) of all sessions. More details are provided in the Sec. 3 of supp." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 331, + 228, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 331, + 228, + 342 + ], + "spans": [ + { + "bbox": [ + 131, + 331, + 228, + 342 + ], + "type": "text", + "content": "4.2 Main Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 353, + 481, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 481, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 481, + 389 + ], + "type": "text", + "content": "Tab. 1 shows the results of our method and five prompt-based continual learning competitors on two datasets with three different pre-trained models. From the results, we observe that our RCS-Prompt performs very competitively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 389, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 544 + ], + "type": "text", + "content": "Specifically, when employing ViT-B/16 Sup-21K as the pre-trained model and augmenting it with RCS-Prompt, the performance of HiDe-Prompt increases by " + }, + { + "bbox": [ + 130, + 389, + 482, + 544 + ], + "type": "inline_equation", + "content": "0.42\\%" + }, + { + "bbox": [ + 130, + 389, + 482, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 389, + 482, + 544 + ], + "type": "inline_equation", + "content": "1.30\\%" + }, + { + "bbox": [ + 130, + 389, + 482, + 544 + ], + "type": "text", + "content": " FAA on the Split CIFAR-100 and Split ImageNet-R datasets, respectively. When using DINO-1K as the pre-trained model and augmenting it with RCS-Prompt, the performance of HiDe-Prompt increases by " + }, + { + "bbox": [ + 130, + 389, + 482, + 544 + ], + "type": "inline_equation", + "content": "2.05\\%" + }, + { + "bbox": [ + 130, + 389, + 482, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 389, + 482, + 544 + ], + "type": "inline_equation", + "content": "5.20\\%" + }, + { + "bbox": [ + 130, + 389, + 482, + 544 + ], + "type": "text", + "content": " FAA on the Split CIFAR-100 and Split ImageNet-R datasets, respectively. We find that our method exhibits higher performance improvements when the self-supervised pre-trained models are used. This phenomenon suggests that our approach excels in establishing clear class margins, particularly effective in scenarios where self-supervised pre-trained models struggle with defining clear class margins. Furthermore, equipping with RCS-Prompt leads to significant performance improvements in DualPrompt, which proves the flexibility our method as a plug-in." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 617 + ], + "type": "text", + "content": "Additional datasets: Furthermore, we evaluate our method on 5-Datasets and Split CUB-200 in Tab. 2. On the 5-Datasets, when integrated with RCS-Prompt, the performance of HiDe-Prompt increases " + }, + { + "bbox": [ + 130, + 545, + 482, + 617 + ], + "type": "inline_equation", + "content": "0.68\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 545, + 482, + 617 + ], + "type": "inline_equation", + "content": "1.23\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 617 + ], + "type": "text", + "content": " with different PTMs, respectively. Then, on the Split CUB-200 dataset, employing RCS-Prompt leads to a notable improvement in the performance of HiDe-Prompt, with gains of " + }, + { + "bbox": [ + 130, + 545, + 482, + 617 + ], + "type": "inline_equation", + "content": "0.76\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 545, + 482, + 617 + ], + "type": "inline_equation", + "content": "1.19\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 617 + ], + "type": "text", + "content": " with different PTMs, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Additional experimental outcomes involving more datasets (e.g., ImageNet-A [23] and ObjectNet [2]) and pre-trained models are detailed in the Sec. 4 of supp. The comparison between our method with DAP [29], LAE [18], APG [59] and RanPAC [40] on their benchmark is also provided." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "type": "text", + "content": "RCS-Prompt" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 199, + 157, + 415, + 232 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 3: Component ablation study " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": ". Baseline indicate HiDe-Prompt [64] using DINO-1K and the seed 42. " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dis}" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " means adding the Inter-session Bidirectional Prompt Loss. Normalization means adding the Prompt Magnitude Normalization." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 199, + 157, + 415, + 232 + ], + "lines": [ + { + "bbox": [ + 199, + 157, + 415, + 232 + ], + "spans": [ + { + "bbox": [ + 199, + 157, + 415, + 232 + ], + "type": "table", + "html": "
BaselineL_disNormalizationFAACAA
✓71.6075.61
✓✓74.9278.20
✓✓71.9375.83
✓✓✓75.1278.35
", + "image_path": "d7728f4a9c2f7a585a46c0390521215e450c3660e4e16a3dbef79ebb27d39d9f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 260, + 246, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 246, + 272 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 246, + 272 + ], + "type": "text", + "content": "4.3 Further Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 297, + 479, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 479, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 479, + 344 + ], + "type": "text", + "content": "We conduct all ablation studies on challenging Split ImageNet-R dataset [22] with DINO-1K using the seed as 42 (the widely used seed in [55,70,71]). More discussion about the continual learning performance trend, technical selections and visualization analyses are detailed in the Sec. 5 of supp." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 347, + 482, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 503 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 503 + ], + "type": "text", + "content": "Component ablation: We conduct experiments to evaluate the impacts of the proposed Inter-session Bidirectional Prompt Loss and Prompt Magnitude Normalization. Our experiments involve: (i) Baseline: we establish HiDePrompt [64] as our baseline. (ii) Baseline " + }, + { + "bbox": [ + 130, + 347, + 482, + 503 + ], + "type": "inline_equation", + "content": "+\\mathcal{L}_{dis}" + }, + { + "bbox": [ + 130, + 347, + 482, + 503 + ], + "type": "text", + "content": ": we add the Inter-session Bidirectional Prompt Loss during training. (iii) Baseline " + }, + { + "bbox": [ + 130, + 347, + 482, + 503 + ], + "type": "inline_equation", + "content": "+\\mathrm{Normalization}" + }, + { + "bbox": [ + 130, + 347, + 482, + 503 + ], + "type": "text", + "content": ": we add the Prompt Magnitude Normalization before using prompts. (iv) Baseline " + }, + { + "bbox": [ + 130, + 347, + 482, + 503 + ], + "type": "inline_equation", + "content": "+\\mathcal{L}_{dis} + \\mathrm{Normalization}" + }, + { + "bbox": [ + 130, + 347, + 482, + 503 + ], + "type": "text", + "content": ": two modules are used simultaneously. From the results in Tab. 3, we reach the following observations: (i) The primary contribution of this paper, the proposed Inter-session Bidirectional Prompt Loss " + }, + { + "bbox": [ + 130, + 347, + 482, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dis}" + }, + { + "bbox": [ + 130, + 347, + 482, + 503 + ], + "type": "text", + "content": ", yields a considerably " + }, + { + "bbox": [ + 130, + 347, + 482, + 503 + ], + "type": "inline_equation", + "content": "3.32\\%" + }, + { + "bbox": [ + 130, + 347, + 482, + 503 + ], + "type": "text", + "content": " performance increase over the baseline. (ii) The Prompt Magnitude Normalization improves model performance by " + }, + { + "bbox": [ + 130, + 347, + 482, + 503 + ], + "type": "inline_equation", + "content": "0.33\\%" + }, + { + "bbox": [ + 130, + 347, + 482, + 503 + ], + "type": "text", + "content": ". (iii) Leveraging both the Inter-session Bidirectional Prompt Loss and the Prompt Magnitude Normalization achieves a model performance of " + }, + { + "bbox": [ + 130, + 347, + 482, + 503 + ], + "type": "inline_equation", + "content": "75.12\\%" + }, + { + "bbox": [ + 130, + 347, + 482, + 503 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 506, + 482, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dis}" + }, + { + "bbox": [ + 130, + 506, + 482, + 615 + ], + "type": "text", + "content": " vs. contrastive regularization: This study aims to verify the superiority of " + }, + { + "bbox": [ + 130, + 506, + 482, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dis}" + }, + { + "bbox": [ + 130, + 506, + 482, + 615 + ], + "type": "text", + "content": " than Contrastive Regularization [64]. The Contrastive Regularization (CR) is designed to enhance the session-specific prompt differences, having the potential to reduce the overlap between old and new class spaces. In Tab. 4, we observe that our proposed loss significantly outperforms the contrastive regularization. This superiority stems from the fact that the contrastive loss is class-agnostic, not establishing a connection between prompt discrimination and class relationships. Besides, the contrastive regularization fails to alleviate the occurrence of the prompt-ambiguity overlap space." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Loss design: This discussion aims to prove that bidirectional prompt optimization is vital for improving the model performance, and the symmetric loss works. For constructing " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dis}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", we alter the labels of ID-noisy samples to pseudo old labels and design the symmetric loss to optimize prompts. In this discussion," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 252, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 252, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 252, + 102 + ], + "type": "text", + "content": "Longrong Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 156, + 167, + 459, + 215 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": "Table 4: The comparison (\\%) between different losses for verifying the superiority of our proposed loss than simply enlarging the inter-session prompt difference. CR is a regularization term used in [64] to enlarge the inter-session prompt difference. In all experiments, the Prompt Magnitude Normalization is used." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 156, + 167, + 459, + 215 + ], + "lines": [ + { + "bbox": [ + 156, + 167, + 459, + 215 + ], + "spans": [ + { + "bbox": [ + 156, + 167, + 459, + 215 + ], + "type": "table", + "html": "
Baseline w/o CR+CR+Ldis+CR & Ldis
FAA71.4771.9374.9075.12
CAA75.7375.8378.2978.35
", + "image_path": "45e2a0afdfa8de62b93981559459224d88dfa41561674653415b89e8281d7dee.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 182, + 277, + 432, + 337 + ], + "blocks": [ + { + "bbox": [ + 130, + 224, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 269 + ], + "type": "text", + "content": "Table 5: The comparison (\\%) between different loss designs for verifying the necessity of the bidirectional prompt optimization. \"w/o old\" means not using pseudo old labels for ID-noisy samples and \"w old\" means using pseudo old labels for ID-noisy samples. In all experiments, the Prompt Magnitude Normalization is used." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 182, + 277, + 432, + 337 + ], + "lines": [ + { + "bbox": [ + 182, + 277, + 432, + 337 + ], + "spans": [ + { + "bbox": [ + 182, + 277, + 432, + 337 + ], + "type": "table", + "html": "
Baseline+Lce dis+Ldis
w/o oldw oldw old
FAA71.9371.5073.0775.12
CAA75.8376.0576.8178.35
", + "image_path": "5228bbb3eef751765eefdf03f00b674da3b36b8b6efcca6e6572d020ddfedc30.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 357, + 465, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 465, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 465, + 369 + ], + "type": "text", + "content": "we formulate the cross-entropy loss concerning both old and new classes as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 209, + 375, + 481, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 375, + 481, + 409 + ], + "spans": [ + { + "bbox": [ + 209, + 375, + 481, + 409 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d i s} ^ {c e} \\left(\\boldsymbol {x} _ {i} ^ {b}\\right) = - \\sum_ {k = 1} ^ {K} q \\left(k \\mid \\boldsymbol {x} _ {i} ^ {b}\\right) \\cdot \\log \\left(p _ {d i s} ^ {b} \\left(k \\mid \\boldsymbol {x} _ {i} ^ {b}\\right)\\right), \\tag {8}", + "image_path": "ba03abf16fb07a1d22d14bd34faa74dcecc957daf0e72d3aca28848b40d5ca43.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 414, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 533 + ], + "type": "text", + "content": "where old class logits are not set to -inf. The analysis from Tab. 5 yields several key observations: (i) After adding " + }, + { + "bbox": [ + 130, + 414, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dis}^{ce}" + }, + { + "bbox": [ + 130, + 414, + 482, + 533 + ], + "type": "text", + "content": " but not using pseudo old labels, the model performance reduces from 71.93 to 71.50. This suggests that undirectionally enhancing inter-session prompt discrimination is ineffective, implying a potential enlargement of the prompt-ambiguity overlap space. (ii) Compared to " + }, + { + "bbox": [ + 130, + 414, + 482, + 533 + ], + "type": "inline_equation", + "content": "+ \\mathcal{L}_{dis}^{ce}" + }, + { + "bbox": [ + 130, + 414, + 482, + 533 + ], + "type": "text", + "content": " without old\", " + }, + { + "bbox": [ + 130, + 414, + 482, + 533 + ], + "type": "inline_equation", + "content": "+ \\mathcal{L}_{dis}^{ce}" + }, + { + "bbox": [ + 130, + 414, + 482, + 533 + ], + "type": "text", + "content": " with old\" results in a " + }, + { + "bbox": [ + 130, + 414, + 482, + 533 + ], + "type": "inline_equation", + "content": "1.57\\%" + }, + { + "bbox": [ + 130, + 414, + 482, + 533 + ], + "type": "text", + "content": " higher model performance. Thus, bidirectionally optimizing prompts is vital for diminishing the class overlap. (iii) After using pseudo old labels, than adding " + }, + { + "bbox": [ + 130, + 414, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dis}^{ce}" + }, + { + "bbox": [ + 130, + 414, + 482, + 533 + ], + "type": "text", + "content": ", adding " + }, + { + "bbox": [ + 130, + 414, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dis}" + }, + { + "bbox": [ + 130, + 414, + 482, + 533 + ], + "type": "text", + "content": " improves the model performance from 73.07 to 75.12. This indicates that the symmetric loss can mitigate the negative impact brought by using pseudo old labels." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 534, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 593 + ], + "type": "text", + "content": "Hyper-parameter analysis: We investigate the effects of the hyperparameter " + }, + { + "bbox": [ + 130, + 534, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 534, + 482, + 593 + ], + "type": "text", + "content": ". More specifically, we explore a range of values for " + }, + { + "bbox": [ + 130, + 534, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 534, + 482, + 593 + ], + "type": "text", + "content": ", namely " + }, + { + "bbox": [ + 130, + 534, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\{0.0, 0.2, 0.4, 0.6, 1.0\\}" + }, + { + "bbox": [ + 130, + 534, + 482, + 593 + ], + "type": "text", + "content": ". In the results presented in Tab. 6, it is evident that as the value of " + }, + { + "bbox": [ + 130, + 534, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 534, + 482, + 593 + ], + "type": "text", + "content": " increases, there is a corresponding gradual improvement in performance. This improvement plateaus when " + }, + { + "bbox": [ + 130, + 534, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 534, + 482, + 593 + ], + "type": "text", + "content": " attains a value of 0.4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Statistical analysis: This statistic aims to use inter-session misclassification to quantify the old and new class space overlap. In this analysis, without losing the generality, we focus on the predictions for test samples from session 0 (the first session) after learning all sessions. The total test sample count is 650 in session 0. For the predictions from the baseline (HiDe-Prompt), we compute the count of samples that are misclassified as classes from other sessions (session 0)," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "type": "text", + "content": "RCS-Prompt" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 206, + 135, + 408, + 182 + ], + "blocks": [ + { + "bbox": [ + 205, + 114, + 408, + 126 + ], + "lines": [ + { + "bbox": [ + 205, + 114, + 408, + 126 + ], + "spans": [ + { + "bbox": [ + 205, + 114, + 408, + 126 + ], + "type": "text", + "content": "Table 6: Impacts (%) of " + }, + { + "bbox": [ + 205, + 114, + 408, + 126 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 205, + 114, + 408, + 126 + ], + "type": "text", + "content": " on Split ImageNet-R." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 206, + 135, + 408, + 182 + ], + "lines": [ + { + "bbox": [ + 206, + 135, + 408, + 182 + ], + "spans": [ + { + "bbox": [ + 206, + 135, + 408, + 182 + ], + "type": "table", + "html": "
0.00.20.40.61.0
FAA71.9374.3375.1274.1273.58
CAA75.8377.8578.3577.6377.44
", + "image_path": "39c5df419df7ab09e385d36d4edc57aed597ae513f36439d2f8b7677fbb1bb19.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 151, + 193, + 252, + 304 + ], + "blocks": [ + { + "bbox": [ + 151, + 193, + 252, + 304 + ], + "lines": [ + { + "bbox": [ + 151, + 193, + 252, + 304 + ], + "spans": [ + { + "bbox": [ + 151, + 193, + 252, + 304 + ], + "type": "image", + "image_path": "1ffa2e08b20051eadf540f92a9dd523c8a07ef7d79c68a93ff9c8acebde2fbf6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 323, + 482, + 368 + ], + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 368 + ], + "type": "text", + "content": "Fig. 3: The analysis about misclassification. (a) We analyze which type of misclassification is the main type for the baseline. Inter-session misclassification means that a sample is misclassified as a class from other sessions. (b) We quantify the solving of our method for the inter-session misclassification." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 256, + 193, + 462, + 304 + ], + "blocks": [ + { + "bbox": [ + 256, + 193, + 462, + 304 + ], + "lines": [ + { + "bbox": [ + 256, + 193, + 462, + 304 + ], + "spans": [ + { + "bbox": [ + 256, + 193, + 462, + 304 + ], + "type": "image", + "image_path": "9aab9bb8f19d08373154688ce97f6b27ccf1e5b4ff25be814ba6994559ce6fa9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 392, + 482, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 513 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 513 + ], + "type": "text", + "content": "termed as inter-session (intra-session) misclassification type, as shown in Fig. 3 (a). Then, as shown in Fig. 3 (b), focusing on inter-session misclassification, we compute the count of samples that are misclassified in each session to compare the baseline and our method. Fig. 3 reveals two key observations: " + }, + { + "bbox": [ + 130, + 392, + 482, + 513 + ], + "type": "inline_equation", + "content": "(i)" + }, + { + "bbox": [ + 130, + 392, + 482, + 513 + ], + "type": "text", + "content": " For the baseline, 173 " + }, + { + "bbox": [ + 130, + 392, + 482, + 513 + ], + "type": "inline_equation", + "content": "(26.62\\%)" + }, + { + "bbox": [ + 130, + 392, + 482, + 513 + ], + "type": "text", + "content": " test samples are misclassified as classes in another session, and only 31 " + }, + { + "bbox": [ + 130, + 392, + 482, + 513 + ], + "type": "inline_equation", + "content": "(4.77\\%)" + }, + { + "bbox": [ + 130, + 392, + 482, + 513 + ], + "type": "text", + "content": " are misclassified as classes from session 0. This shows that the inter-session misclassification is the main problem for existing methods. " + }, + { + "bbox": [ + 130, + 392, + 482, + 513 + ], + "type": "inline_equation", + "content": "(ii)" + }, + { + "bbox": [ + 130, + 392, + 482, + 513 + ], + "type": "text", + "content": " Than the baseline, our method significantly decreases the inter-session misclassification sample count from 173 to 67. This shows that our method can effectively solve the inter-session misclassification problem." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 531, + 220, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 220, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 220, + 544 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Our study reveals that the optimization objective of prompts lacks the explicit modeling of inter-session class relationships, not effectively separating new and old class spaces. To minimize the overlap between new and old class spaces for creating the clear inter-session class margins, we introduce an Inter-session Bidirectional Prompt Loss to place new and old samples to discriminative regions, while eliminating the prompt-ambiguity overlap space. Furthermore, we design the Prompt Magnitude Normalization to enhance the prompt tuning effectiveness. Extensive experiments demonstrate the superiority of our approach across different datasets and continual learning scenarios." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 252, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 252, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 252, + 102 + ], + "type": "text", + "content": "Longrong Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 494, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 494, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 494, + 224 + ], + "type": "text", + "content": "Acknowledgements. This work is supported in part by National Natural Science Foundation of China under Grant No.62441602, National Science Foundation for Distinguished Young Scholars under Grant 62225605, Project 12326608 supported by NSFC, National Natural Science Foundation of China under Grant U20A20222, Zhejiang Key Research and Development Program under Grant 2023C03196, the Ningbo Science and Technology Innovation Project (No.2024Z294), the Fundamental Research Funds for the Central Universities, the Key R&D Program of Zhejiang Province, China 2023C01043, Ant Group through CCF-Ant Research Fund, and sponsored by CCF-AFSG Research Fund." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 243, + 199, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 243, + 199, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 243, + 199, + 255 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 268, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 136, + 268, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 268, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 136, + 268, + 481, + 300 + ], + "type": "text", + "content": "1. Aljundi, R., Lin, M., Goujaud, B., Bengio, Y.: Gradient based sample selection for online continual learning. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 301, + 481, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 301, + 481, + 345 + ], + "spans": [ + { + "bbox": [ + 136, + 301, + 481, + 345 + ], + "type": "text", + "content": "2. Barbu, A., Mayo, D., Alverio, J., Luo, W., Wang, C., Gutfreund, D., Tenenbaum, J., Katz, B.: Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 346, + 481, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 346, + 481, + 378 + ], + "spans": [ + { + "bbox": [ + 136, + 346, + 481, + 378 + ], + "type": "text", + "content": "3. Belouadah, E., Popescu, A.: II2m: Class incremental learning with dual memory. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 583-592 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 379, + 302, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 379, + 302, + 389 + ], + "spans": [ + { + "bbox": [ + 137, + 379, + 302, + 389 + ], + "type": "text", + "content": "4. Bulatov, Y.: Notmnist dataset (2011)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 390, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 390, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 137, + 390, + 481, + 423 + ], + "type": "text", + "content": "5. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 423, + 481, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 423, + 481, + 466 + ], + "spans": [ + { + "bbox": [ + 137, + 423, + 481, + 466 + ], + "type": "text", + "content": "6. Cha, S., Cho, S., Hwang, D., Hong, S., Lee, M., Moon, T.: Rebalancing batch normalization for exemplar-based class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20127-20136 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 467, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 467, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 137, + 467, + 481, + 500 + ], + "type": "text", + "content": "7. Chaudhry, A., Dokania, P.K., Ajanthan, T., Torr, P.H.: Riemannian walk for incremental learning: Understanding forgetting and intransigence. In: Proceedings of the European conference on computer vision (ECCV). pp. 532-547 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 137, + 501, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 501, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 137, + 501, + 481, + 522 + ], + "type": "text", + "content": "8. Chaudhry, A., Ranzato, M., Rohrbach, M., Elhoseiny, M.: Efficient lifelong learning with a-gem. In: International Conference on Learning Representations (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 137, + 522, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 522, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 137, + 522, + 481, + 555 + ], + "type": "text", + "content": "9. Chen, X., Xie, S., He, K.: An empirical study of training self-supervised vision transformers. in 2021 IEEE. In: CVF International Conference on Computer Vision (ICCV). pp. 9620-9629" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 555, + 481, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 555, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 133, + 555, + 481, + 598 + ], + "type": "text", + "content": "10. Dong, J., Li, H., Cong, Y., Sun, G., Zhang, Y., Van Gool, L.: No one left behind: Real-world federated class-incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence 46(4), 2054-2070 (2024). https://doi.org/10.1109/TPAMI.2023.3334213" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 599, + 481, + 632 + ], + "type": "text", + "content": "11. Dong, J., Liang, W., Cong, Y., Sun, G.: Heterogeneous forgetting compensation for class-incremental learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 11742-11751 (October 2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 481, + 666 + ], + "type": "text", + "content": "12. Dong, J., Wang, L., Fang, Z., Sun, G., Xu, S., Wang, X., Zhu, Q.: Federated class-incremental learning. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2022)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "type": "text", + "content": "RCS-Prompt" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "13. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 162, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 162, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 481, + 205 + ], + "type": "text", + "content": "14. Douillard, A., Cord, M., Ollion, C., Robert, T., Valle, E.: Podnet: Pooled outputs distillation for small-tasks incremental learning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16. pp. 86-102. Springer (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 134, + 205, + 481, + 237 + ], + "type": "text", + "content": "15. Douillard, A., Ramé, A., Couairon, G., Cord, M.: Dytox: Transformers for continual learning with dynamic token expansion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9285-9295 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 238, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 238, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 134, + 238, + 481, + 281 + ], + "type": "text", + "content": "16. Ebrahimi, S., Meier, F., Calandra, R., Darrell, T., Rohrbach, M.: Adversarial continual learning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16. pp. 386-402. Springer (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 282, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 282, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 134, + 282, + 481, + 303 + ], + "type": "text", + "content": "17. French, R.M.: Catastrophic forgetting in connectionist networks. Trends in cognitive sciences 3(4), 128-135 (1999)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 304, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 304, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 134, + 304, + 481, + 335 + ], + "type": "text", + "content": "18. Gao, Q., Zhao, C., Sun, Y., Xi, T., Zhang, G., Ghanem, B., Zhang, J.: A unified continual learning framework with general parameter-efficient tuning. arXiv preprint arXiv:2303.10070 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 337, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 337, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 134, + 337, + 481, + 357 + ], + "type": "text", + "content": "19. Ge, Y., Li, Y., Ni, S., Zhao, J., Yang, M.H., Itti, L.: Clr: Channel-wise lightweight reprogramming for continual learning. arXiv preprint arXiv:2307.11386 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 358, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 358, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 134, + 358, + 481, + 390 + ], + "type": "text", + "content": "20. Gomes, H.M., Barddal, J.P., Enembreck, F., Bifet, A.: A survey on ensemble learning for data stream classification. ACM Computing Surveys (CSUR) 50(2), 1-36 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 392, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 392, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 134, + 392, + 481, + 423 + ], + "type": "text", + "content": "21. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 425, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 425, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 134, + 425, + 481, + 468 + ], + "type": "text", + "content": "22. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8340-8349 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 468, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 468, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 134, + 468, + 481, + 500 + ], + "type": "text", + "content": "23. Hendrycks, D., Zhao, K., Basart, S., Steinhardt, J., Song, D.: Natural adversarial examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15262-15271 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 502, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 502, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 134, + 502, + 481, + 522 + ], + "type": "text", + "content": "24. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 523, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 523, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 134, + 523, + 481, + 555 + ], + "type": "text", + "content": "25. Hou, S., Pan, X., Loy, C.C., Wang, Z., Lin, D.: Learning a unified classifier incrementally via rebalancing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 831-839 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 556, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 556, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 134, + 556, + 481, + 587 + ], + "type": "text", + "content": "26. Hu, Z., Li, Y., Lyu, J., Gao, D., Vasconcelos, N.: Dense network expansion for class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11858-11867 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 589, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 589, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 134, + 589, + 481, + 621 + ], + "type": "text", + "content": "27. Huang, B., Chen, Z., Zhou, P., Chen, J., Wu, Z.: Resolving task confusion in dynamic expansion architectures for class incremental learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 908-916 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 622, + 481, + 665 + ], + "type": "text", + "content": "28. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII. pp. 709-727. Springer (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 252, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 252, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 252, + 102 + ], + "type": "text", + "content": "Longrong Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "29. Jung, D., Han, D., Bang, J., Song, H.: Generating instance-level prompts for rehearsal-free continual learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11847-11857 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "type": "text", + "content": "30. Khan, M.G.Z.A., Naeem, M.F., Van Gool, L., Stricker, D., Tombari, F., Afzal, M.Z.: Introducing language guidance in prompt-based continual learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11463-11473 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 237 + ], + "type": "text", + "content": "31. Khattak, M.U., Wasim, S.T., Naseer, M., Khan, S., Yang, M.H., Khan, F.S.: Self-regulating prompts: Foundational model adaptation without forgetting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15190-15200 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "type": "text", + "content": "32. Kim, D., Han, B.: On the stability-plasticity dilemma of class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20196-20204 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 270, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 293 + ], + "type": "text", + "content": "33. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 293, + 482, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 315 + ], + "type": "text", + "content": "34. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (1998)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 315, + 482, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 335 + ], + "type": "text", + "content": "35. Li, Z., Hoiem, D.: Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence 40(12), 2935-2947 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 336, + 482, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 369 + ], + "type": "text", + "content": "36. Liu, Y., Li, Y., Schiele, B., Sun, Q.: Online hyperparameter optimization for class-incremental learning. Proceedings of the AAAI Conference on Artificial Intelligence (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 369, + 482, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 403 + ], + "type": "text", + "content": "37. Liu, Y., Schiele, B., Sun, Q.: Adaptive aggregation networks for class-incremental learning. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 2544-2553 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 403, + 482, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 435 + ], + "type": "text", + "content": "38. Liu, Y., Su, Y., Liu, A.A., Schiele, B., Sun, Q.: Mnemonics training: Multi-class incremental learning without forgetting. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 12245-12254 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 435, + 482, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 468 + ], + "type": "text", + "content": "39. Luo, Z., Liu, Y., Schiele, B., Sun, Q.: Class-incremental exemplar compression for class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11371-11380 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 468, + 482, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 501 + ], + "type": "text", + "content": "40. McDonnell, M.D., Gong, D., Parvaneh, A., Abbasnejad, E., van den Hengel, A.: Ranpac: Random projections and pre-trained models for continual learning. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 501, + 482, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 482, + 534 + ], + "type": "text", + "content": "41. Mehta, S.V., Patil, D., Chandar, S., Strubell, E.: An empirical investigation of the role of pre-training in lifelong learning. Journal of Machine Learning Research 24(214), 1-50 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 534, + 482, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 555 + ], + "type": "text", + "content": "42. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 555, + 482, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 482, + 588 + ], + "type": "text", + "content": "43. Panos, A., Kobe, Y., Reino, D.O., Aljundi, R., Turner, R.E.: First session adaptation: A strong replay-free baseline for class-incremental learning. arXiv preprint arXiv:2303.13199 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 588, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 588, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 588, + 482, + 632 + ], + "type": "text", + "content": "44. Petit, G., Popescu, A., Schindler, H., Picard, D., Delezoide, B.: Fetril: Feature translation for exemplar-free class-incremental learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3911-3920 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "text", + "content": "45. Pham, Q., Liu, C., Hoi, S.: Continual normalization: Rethinking batch normalization for online continual learning. In: International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 102 + ], + "type": "text", + "content": "RCS-Prompt" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "46. Qiang, S., Hou, J., Wan, J., Liang, Y., Lei, Z., Zhang, D.: Mixture uniform distribution modeling and asymmetric mix distillation for class incremental learning. In: Proceedings of the AAAI Conference on Artificial Intelligence (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "text", + "content": "47. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "48. Ramasesh, V.V., Lewkowycz, A., Dyer, E.: Effect of scale on catastrophic forgetting in neural networks. In: International Conference on Learning Representations (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "text", + "content": "49. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 2001-2010 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "50. Rymarczyk, D., van de Weijer, J., Zielinski, B., Twardowski, B.: Icicle: Interpretable class incremental continual learning. arXiv preprint arXiv:2303.07811 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "text", + "content": "51. Sarfraz, F., Arani, E., Zonooz, B.: Sparse coding in a dual memory system for lifelong learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 9714-9722 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "type": "text", + "content": "52. Shi, Y., Zhou, K., Liang, J., Jiang, Z., Feng, J., Torr, P.H., Bai, S., Tan, V.Y.: Mimicking the oracle: an initial phase decorrelation approach for class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16722-16731 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 369, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 392 + ], + "type": "text", + "content": "53. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 435 + ], + "type": "text", + "content": "54. Singh, P., Li, Y., Sikarwar, A., Lei, S.W., Gao, D., Talbot, M.B., Sun, Y., Shou, M.Z., Kreiman, G., Zhang, M.: Learning to learn: How to continuously teach humans and machines. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11708-11719 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 435, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 490 + ], + "type": "text", + "content": "55. Smith, J.S., Karlinsky, L., Gutta, V., Cascante-Bonilla, P., Kim, D., Arbelle, A., Panda, R., Feris, R., Kira, Z.: Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11909-11919 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 491, + 481, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 523 + ], + "type": "text", + "content": "56. Sun, Z., Mu, Y., Hua, G.: Regularizing second-order influences for continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20166-20175 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "type": "text", + "content": "57. Talbot, M.B., Zawar, R., Badkundri, R., Zhang, M., Kreiman, G.: Tuned compositional feature replays for efficient stream learning. IEEE Transactions on Neural Networks and Learning Systems (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 555, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 589 + ], + "type": "text", + "content": "58. Tan, M., Pang, R., Le, Q.V.: Efficientdet: Scalable and efficient object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10781-10790 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "type": "text", + "content": "59. Tang, Y.M., Peng, Y.X., Zheng, W.S.: When prompt-based incremental learning does not meet strong pretraining. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1706-1716 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 621, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 666 + ], + "type": "text", + "content": "60. Villa, A., Alcázar, J.L., Alfarra, M., Alhamoud, K., Hurtado, J., Heilbron, F.C., Soto, A., Ghanem, B.: Pivot: Prompting for video continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24214-24223 (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 252, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 252, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 252, + 102 + ], + "type": "text", + "content": "Longrong Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "61. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "type": "text", + "content": "62. Wang, F.Y., Zhou, D.W., Liu, L., Ye, H.J., Bian, Y., Zhan, D.C., Zhao, P.: Beef: Bi-compatible class-incremental learning via energy-based expansion and fusion. In: The Eleventh International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "type": "text", + "content": "63. Wang, F.Y., Zhou, D.W., Ye, H.J., Zhan, D.C.: Foster: Feature boosting and compression for class-incremental learning. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXV. pp. 398-414. Springer (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "64. Wang, L., Xie, J., Zhang, X., Huang, M., Su, H., Zhu, J.: Hierarchical decomposition of prompt-based continual learning: Rethinking obscured sub-optimality. In: Advances in neural information processing systems (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "text", + "content": "65. Wang, L., Zhang, X., Su, H., Zhu, J.: A comprehensive survey of continual learning: Theory, method and application. arXiv preprint arXiv:2302.00487 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "text", + "content": "66. Wang, W., Hu, Y., Chen, Q., Zhang, Y.: Task difficulty aware parameter allocation & regularization for lifelong learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7776-7785 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "text", + "content": "67. Wang, Y., Huang, Z., Hong, X.: S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. Annual Conference on Neural Information Processing Systems (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "68. Wang, Y., Ma, X., Chen, Z., Luo, Y., Yi, J., Bailey, J.: Symmetric cross entropy for robust learning with noisy labels. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 322-330 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "type": "text", + "content": "69. Wang, Z., Zhao, Z., Wu, Y., Zhou, L., Xu, D.: Progressive target-styled feature augmentation for unsupervised domain adaptation on point clouds. In: ECCV (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 392, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 445 + ], + "type": "text", + "content": "70. Wang, Z., Zhang, Z., Ebrahimi, S., Sun, R., Zhang, H., Lee, C.Y., Ren, X., Su, G., Perot, V., Dy, J., et al.: Dualprompt: Complementary prompting for rehearsal-free continual learning. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVI. pp. 631-648. Springer (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 446, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 489 + ], + "type": "text", + "content": "71. Wang, Z., Zhang, Z., Lee, C.Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., Pfister, T.: Learning to prompt for continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 139-149 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 490, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 481, + 533 + ], + "type": "text", + "content": "72. Wu, X., Wang, H., Wu, Y., Li, X.: D3t-gan: Data-dependent domain transfer gans for image generation with limited data. ACM Trans. Multimedia Comput. Commun. Appl. 19(4) (mar 2023). https://doi.org/10.1145/3576858, https://doi.org/10.1145/3576858" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 533, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 481, + 577 + ], + "type": "text", + "content": "73. Wu, Y., Li, H., Wang, F., Zhang, Y., Liang, R.: Self-distilled dynamic fusion network for language-based fashion retrieval. In: ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 3260-3264. IEEE (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 577, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 481, + 610 + ], + "type": "text", + "content": "74. Wu, Y., Wu, X., Li, X., Tian, J.: Mgh: Metadata guided hypergraph modeling for unsupervised person re-identification. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 1571–1580 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "text", + "content": "75. Wu, Y., Chen, Y., Wang, L., Ye, Y., Liu, Z., Guo, Y., Fu, Y.: Large scale incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 374-382 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "76. Xiao, H., Rasul, K., Vollgraf, R.: Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747 (2017)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 102 + ], + "type": "text", + "content": "RCS-Prompt" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 578 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "77. Xu, J., Zhu, Z.: Reinforced continual learning. Advances in Neural Information Processing Systems 31 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "type": "text", + "content": "78. Yan, S., Xie, J., He, X.: Der: Dynamically expandable representation for class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3014-3023 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 204 + ], + "type": "text", + "content": "79. Yoon, J., Yang, E., Lee, J., Hwang, S.J.: Lifelong learning with dynamically expandable networks. In: International Conference on Learning Representations (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 205, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 205, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 134, + 205, + 480, + 247 + ], + "type": "text", + "content": "80. Yu, L., Twardowski, B., Liu, X., Herranz, L., Wang, K., Cheng, Y., Jui, S., Weijer, J.v.d.: Semantic drift compensation for class-incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6982-6991 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 133, + 249, + 480, + 281 + ], + "type": "text", + "content": "81. Zhang, G., Wang, L., Kang, G., Chen, L., Wei, Y.: Slow learner with classifier alignment for continual learning on a pre-trained model. arXiv preprint arXiv:2303.05118 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 282, + 480, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 282, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 134, + 282, + 480, + 324 + ], + "type": "text", + "content": "82. Zhang, J., Zhang, J., Ghosh, S., Li, D., Tasci, S., Heck, L., Zhang, H., Kuo, C.C.J.: Class-incremental learning via deep model consolidation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1131-1140 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 133, + 326, + 480, + 357 + ], + "type": "text", + "content": "83. Zhao, B., Xiao, X., Gan, G., Zhang, B., Xia, S.T.: Maintaining discrimination and fairness in class incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13208-13217 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 358, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 358, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 133, + 358, + 480, + 390 + ], + "type": "text", + "content": "84. Zhao, H., Fu, Y., Kang, M., Tian, Q., Wu, F., Li, X.: Mgmt: Multi-grained slow vs. fast framework for few-shot class-incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 392, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 392, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 133, + 392, + 480, + 423 + ], + "type": "text", + "content": "85. Zhao, H., Wang, H., Fu, Y., Wu, F., Li, X.: Memory-efficient class-incremental learning for image classification. IEEE Transactions on Neural Networks and Learning Systems 33(10), 5966-5977 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 425, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 425, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 133, + 425, + 480, + 456 + ], + "type": "text", + "content": "86. Zhou, D.W., Wang, F.Y., Ye, H.J., Ma, L., Pu, S., Zhan, D.C.: Forward compatible few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9046-9056 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 458, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 458, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 133, + 458, + 480, + 489 + ], + "type": "text", + "content": "87. Zhou, D.W., Wang, Q.W., Ye, H.J., Zhan, D.C.: A model or 603 exemplars: Towards memory-efficient class-incremental learning. In: International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 491, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 491, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 133, + 491, + 480, + 522 + ], + "type": "text", + "content": "88. Zhou, D.W., Ye, H.J., Ma, L., Xie, D., Pu, S., Zhan, D.C.: Few-shot class incremental learning by sampling multi-phase tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 523, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 523, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 133, + 523, + 480, + 544 + ], + "type": "text", + "content": "89. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 545, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 545, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 133, + 545, + 480, + 578 + ], + "type": "text", + "content": "90. Zhuang, H., Weng, Z., Wei, H., Xie, R., Toh, K.A., Lin, Z.: Acil: Analytic class incremental learning with absolute memorization and privacy protection. Advances in Neural Information Processing Systems 35, 11602-11614 (2022)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 252, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 252, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 252, + 102 + ], + "type": "text", + "content": "Longrong Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/REDIR_ Refocus-free Event-based De-occlusion Image Reconstruction/59c82a8e-f2d6-49fa-a29f-6faeb9b1bbb2_content_list.json b/2024/REDIR_ Refocus-free Event-based De-occlusion Image Reconstruction/59c82a8e-f2d6-49fa-a29f-6faeb9b1bbb2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fee739396cf9464574bc4f5b4df15ffd5902bf86 --- /dev/null +++ b/2024/REDIR_ Refocus-free Event-based De-occlusion Image Reconstruction/59c82a8e-f2d6-49fa-a29f-6faeb9b1bbb2_content_list.json @@ -0,0 +1,1751 @@ +[ + { + "type": "text", + "text": "REDIR: Refocus-free Event-based De-occlusion Image Reconstruction", + "text_level": 1, + "bbox": [ + 225, + 140, + 779, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qi Guo $^{1,2}$ , Hailong Shi $^{1,2*}$ , Huan Li $^{1,2}$ , Jinsheng Xiao $^{3}$ , and Xingyu Gao $^{1,2*}$", + "bbox": [ + 218, + 210, + 782, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Institute of Microelectronics of the Chinese Academy of Sciences, Beijing, China", + "bbox": [ + 251, + 239, + 750, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ University of Chinese Academy of Sciences, Beijing, China", + "bbox": [ + 316, + 253, + 684, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Wuhan University, Hubei, China", + "bbox": [ + 395, + 267, + 604, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{shihailong, gaoxingyu}@ime.ac.cn", + "bbox": [ + 354, + 282, + 645, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/QiGuo-46/REDIR", + "bbox": [ + 354, + 296, + 648, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The employment of the event-based synthetic aperture imaging (E-SAI) technique, which has the capability to capture high-frequency light intensity variations, has facilitated its extensive application on scene de-occlusion reconstruction tasks. However, existing methods usually require prior information and have strict restriction of camera motion on SAI acquisition methods. This paper proposes a novel end-to-end refocus-free variable E-SAI de-occlusion image reconstruction approach REDIR, which can align the global and local features of the variable event data and effectively achieve high-resolution imaging of pure event streams. To further improve the reconstruction of the occluded target, we propose a perceptual mask-gated connection module to interlink information between modules, and incorporate a spatial-temporal attention mechanism into the SNN block to enhance target extraction ability of the model. Through extensive experiments, our model achieves state-of-the-art reconstruction quality on the traditional E-SAI dataset without prior information, while verifying the effectiveness of the variable event data feature registration method on our newly introduced V-ESAI dataset, which obviates the reliance on prior knowledge and extends the applicability of SAI acquisition methods by incorporating focus changes, lens rotations, and non-uniform motion.", + "bbox": [ + 259, + 345, + 743, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Event-based Synthetic Aperture Imaging $\\cdot$ De-occlusion Reconstruction $\\cdot$ Data Registration", + "bbox": [ + 259, + 609, + 743, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 662, + 356, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The event camera [4] is a biomimetic sensing device that can perceive changes in the intensity of reflected light from objects with high temporal resolution [16]. For application scenarios with dense occlusions, event cameras can be employed to collect target information from different perspectives [18] within the gaps between occlusions, such as between trees, dense fences, and sight obstruction caused by rain and snow, which holds significant value for search and rescue missions and target tracking operations [15].", + "bbox": [ + 212, + 695, + 787, + 785 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event-based synthetic aperture imaging (E-SAI) [31] is a data acquisition method, as displayed in Fig. 1, which utilizes a linear uniform motion of camera for events and APS", + "bbox": [ + 212, + 786, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 215, + 825, + 354, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "frames sampling. Through E-SAI acquisition method, the camera can capture events triggered by targets and occlusion in motion [33]. The deep learning models [17] are capable of removing the obstruction of target scenarios on linear uniform motion E-SAI method. However, it is susceptible to lens shakes and motion errors [28], leading to misalignment of event points and subsequent degradation of imaging performance [27].", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7a923e9eee67bf8fb23f858bb59f0241ab18a8f62ef1d180d25e044fcad41d6a.jpg", + "image_caption": [ + "Fig. 1: E-SAI De-occlusion Image Reconstruction (without prior information). The event streams are captured through the uniform linear E-SAI acquisition method, the pictures below illustrate the results of de-occlusion imaging by existing method Hybrid [31] and ours." + ], + "image_footnote": [], + "bbox": [ + 303, + 247, + 694, + 494 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Currently, the existing event point aligning methods [30, 31] mainly perform linear affine transformation on the event point by obtaining prior information such as the focal length, moving speed, and depth of field of the event camera, or obtain the affine transformation parameter matrix through model training and perform linear affine transformation. Therefore, the precise alignment of event points is unattainable due to the presence of errors caused by jitter during camera movement and alternative synthetic aperture imaging methods apart from uniform linear motion. Besides, obtaining event camera information in real-time and manually focusing will bring a huge workload. In real-world imaging scenarios, the aforementioned methods are facing two main challenges:", + "bbox": [ + 212, + 566, + 787, + 703 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Strict Restriction of SAI Acquisition Method: The reconstruction performance of imaging tasks relies on the acquisition method of event data. For imaging approaches beyond uniform linear motion, the reconstruction of the target presents substantial obstacles using existing techniques. The limited prediction correction parameters are employed for focusing tasks, while linear offset mapping is utilized to calibrate event data, resulting in inherent limitations of the imaging modality.", + "- Requires Prior Information: Focusing methods in previous work often require information that relies on the acquisition process and requires manual labeling of" + ], + "bbox": [ + 230, + 713, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Q. Guo et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "samples, which results in increased time costs for the production of training sets and is not generalizable to other acquisition parameters.", + "bbox": [ + 241, + 146, + 785, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the above problems, we propose an end-to-end event registration and reconstruction model REDIR. The synthetic aperture imaging de-occlusion task of event cameras mainly includes event alignment based on a combination of global and local features and reconstruction of preserving target objects by filtering out occlusions. The purpose of alignment registration [20] is primarily to address the non-uniform linear motion error in real synthetic aperture imaging. The adaptive independent registration [12] of multiple event frames aims to address the variable SAI imaging acquisition method without prior information, which lays the foundation for de-occlusion reconstruction [21] of variable SAI imaging acquisition methods without prior information.", + "bbox": [ + 212, + 191, + 787, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The event registration module extracts global and local features from the event data and performs fusion alignment, which utilizes spatial transformer network (STN) [9] to predict the affine matrix and perform spatial transformation for each timestamp event frame, while its feature map is downsampled by UNet. The Temporal-Spatial Attention spiking neural network (SNN) [29] module is used to filter noise and extract target features, remove discrete occlusion event points and retain target event points. Finally, the perceptual mask gate module is used to connect multi-scale local features, which can effectively interchange the information between modules to reconstruct the scene without occlusion.", + "bbox": [ + 212, + 328, + 787, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main contributions of the paper can be summarized as:", + "bbox": [ + 240, + 464, + 630, + 479 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel event data registration and alignment model, which can automatically map the target event point coordinates of multiple timestamps to the same position, effectively improving the event data reconstruction performance.", + "- We further propose an end-to-end reconstruction model REDIR that incorporates a perceptual mask-gated connection module and Temporal-Spatial Attention mechanism, which can reconstruct event data with occlusions into target scene images without visual obstruction.", + "- We propose a new multi-variable transformed data augmentation method to compile Variable E-SAI (V-ESAI) dataset, which simulates non-uniform motion, lens rotation, and focal length changes in actual acquisition process, proving our proposed method relevantly persists reconstruction ability under other SAI acquisition methods." + ], + "bbox": [ + 225, + 491, + 785, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 684, + 366, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Synthetic aperture imaging can capture the information of the target from different angles, capture more details in the imaging scene, and obtain a larger imaging range under the condition of limited depth of field and focal length. The difference between the event camera and the traditional sensor is that the light intensity change of the scene can be triggered by the moving DVS sensor, so the static image scene can be reconstructed. In addition, for the occluded imaging scene, the imaging method of synthetic aperture can obtain the information of the imaging plane of the target scene through the occlusion at different positions, and fuse and supplement the imaging data of multiple positions,", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "REDIR", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which can effectively remove the occlusion and restore the overall appearance of the target.", + "bbox": [ + 215, + 146, + 785, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, some scholars applied it to the field of optics. In order to obtain complete information on objects through occluding objects, Wang [25] et al. proposed a method of removing foreground occluding based on a deep encoder-decoder network, DeOccNet, for object surface reconstruction in occluding SAI scenarios. In order to obtain multidimensional features and improve the performance of detail reconstruction, Wang [24] et al. improved the method by integrating spatial features and angular features for SAI reconstruction, and the performance reached SOTA. People have gradually found that the event camera has better performance in object reconstruction of occluded scenes. Because the event sensor has the ability to perceive the change of light intensity, it can effectively capture the information of the object in the occluded gap. Zhang [31] et al. used an event camera to collect the results of linear uniform synthetic aperture imaging of the target under extremely dense occlusion and extreme lighting conditions. The event frame registration was realized through manual focusing, and the target scene was restored through filtering and reconstruction by SNN and CNN hybrid models. In order to solve the problem of automatic registration of event data, Yu [30] et al. proposed a method to predict the warping parameter and move the coordinates of each timestamp by linear offset according to the parameter, so as to achieve refocusing. Generally, in the actual acquisition process, the sensor device cannot carry out uniform linear motion perfectly. Therefore, this event registration method based on predicted linear transformation parameters is not suitable for acquisition methods with large motion errors, changes in rotation, focal length and depth of field, and non-uniform linear motion.", + "bbox": [ + 215, + 183, + 787, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Event-Based Image Registration is one of the necessary steps of E-SAI. The algorithms related to image registration and point cloud registration have been very mature, and are usually applied to the matching, stitching, or superposition of two images/point clouds. The goal of registration is to find the spatial mapping relationship between two images/point clouds, match the extracted features, and map and fuse them so that the two images/point clouds overlap to the same spatial position. For image fusion registration, Zhao [32] et al proposed a general recursive cascade network architecture that can realize deep cascade learning for deformable image registration. Through unsupervised training, the moving image is gradually deformed through each cascade and finally aligned with the fixed image. In order to combine global and local feature information for more refined registration, Balakrishna [1] et al. proposed a medical image registration method based on UNet architecture, which realized the alignment of MRI images of multiple coronal sections of the brain by predicting mapping transformation parameters. For the point cloud registration task, Qi [19] et al. proposed a PointNet, which extracts the global and local features of the point cloud through the T-Net transformation network, and realizes the registration of the point cloud by affine transformation through prediction transformation parameters.", + "bbox": [ + 215, + 522, + 787, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For event data registration, it is usually necessary to match the coordinates of multiple timestamps to align the event to the same position. Gallego [5] et al demonstrated the ability to handle several important visual problems in event cameras (3D reconstruction, optical flow, and motion estimation), all without additional information about", + "bbox": [ + 215, + 780, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 116, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Q. Guo et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the appearance of the scene: by maximizing the alignment of events on the trajectory of points on the image plane. Gu [6] et al. proposed a new probabilistic method for event alignment by modeling the events to be aligned as independent Poisson point processes or spatio-temporal Poisson point processes, so that every pixel is subject to this process. Based on this idea, a likelihood function is derived to estimate the camera motion that best explains the event. It is not limited to rotational motion, but can also be applied to other types of motion, such as translation. Huang [8] et al proposed a unified event-based motion estimation framework applicable to various motion models. By innovatively aligning events with the map, the rotation model and the 6-degree of freedom model are implemented on a standard CPU, and the motion estimation under a small number of sampled events is performed in an efficient manner. At the same time, a flexible batch size strategy is introduced to adapt to different scene textures, camera speed, and resolution. The comprehensive and detailed experimental results show that the framework reaches the current state of the art in terms of accuracy and efficiency. At present, there is no mature deep-learning method for event data registration without prior information.", + "bbox": [ + 212, + 146, + 787, + 387 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by the aforementioned approaches, we propose an event data registration method based on deep learning that enables fusion registration by extracting both global and local feature information from each time stamp. This method is not limited to uniform line SAI but can also be applied to other SAI acquisition methods.", + "bbox": [ + 212, + 388, + 787, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Problem Formulation", + "text_level": 1, + "bbox": [ + 214, + 476, + 431, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Event-Based SAI", + "text_level": 1, + "bbox": [ + 214, + 511, + 375, + 525 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When the event camera captures changes in light intensity within the event frame, the actual static light intensity at the last timestamp $n - 1$ and current timestamp $n$ can be expressed as $\\Gamma (x_{n},t_{n} - \\Delta t_{n})$ and $\\Gamma (x_{n},t_{n})$ [3].", + "bbox": [ + 212, + 540, + 787, + 585 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Gamma \\left(x _ {n}, t _ {n}\\right) - \\Gamma \\left(x _ {n}, t _ {n} - \\Delta t _ {n}\\right) \\geq p _ {n} \\cdot \\lambda , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 599, + 787, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where, $\\lambda$ represents the threshold, $p_n$ represents the polarity, and $\\Delta t_{n}$ represents the time interval between the adjacent timestamps. The collected event data includes events caused by occlusion and target $E_{OA}(n)$ , events caused by occlusion only $E_{OO}(n)$ , target only $E_{AA}(n)$ , as well as inherent noise of the camera $E_0$ . Event data can be represented as:", + "bbox": [ + 212, + 628, + 787, + 702 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE (n) = E _ {O A} (n) + E _ {A A} (n) + E _ {O O} (n) + E _ {0}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 720, + 785, + 738 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The static light intensity information corresponding to both target and occlusion objects in an actual occlusion imaging scene can be represented as $\\Gamma_A(n)$ and $\\Gamma_O(n)$ . Therefore, the target event information captured by an event camera in an occlusion scene $E_{OA}(n)$ can be described as:", + "bbox": [ + 212, + 746, + 787, + 806 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left| E _ {O A} (n) \\right| \\propto \\left| \\Gamma_ {A} (n) - \\Gamma_ {O} (n) \\right|. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 824, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "REDIR", + "bbox": [ + 683, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Event Registration", + "text_level": 1, + "bbox": [ + 215, + 146, + 387, + 161 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since the synthetic aperture imaging method necessitates camera movement to capture scene information from various viewing angles, accurate target registration is required. The mapping relationship between coordinates before and after the event can be formulated as follows:", + "bbox": [ + 212, + 171, + 787, + 231 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP _ {n} ^ {r e f} = K R _ {n} K ^ {- 1} P _ {n} + \\frac {K T _ {n}}{d}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 241, + 787, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $P_{n}^{ref}$ represents the pixel position of the target on the reference imaging plane, $K$ denotes the camera's internal matrix, $R_{n}$ and $T_{n}$ are rotation and translation matrices that describe the camera postures $\\theta_{n}$ and $\\theta_{ref}$ respectively, $d$ represents the distance between target A and the camera plane.", + "bbox": [ + 212, + 276, + 784, + 337 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The transformation formula [22] encompasses operations such as scaling $k$ , rotation $\\theta$ , and translation $\\Delta x, \\Delta y$ . Through registration, event points at different times can be aligned to a common position, enabling space-time dimension synthetic aperture imaging. The affine matrix $\\psi$ can be expressed as:", + "bbox": [ + 212, + 338, + 785, + 398 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\psi = \\left[ \\begin{array}{l l} \\cos \\theta / k & \\sin \\theta / k \\Delta x \\\\ - \\sin \\theta / k & \\cos \\theta / k \\Delta y \\end{array} \\right]. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 397, + 407, + 785, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Event Reconstruction", + "text_level": 1, + "bbox": [ + 215, + 460, + 408, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The reconstruction of events is primarily accomplished by accumulating event frame feature maps, with varying timestamps of event points after aligning and occlusion filtering, and performing normalization operations based on the number of timestamps N.", + "bbox": [ + 212, + 486, + 784, + 544 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF M _ {r e c o n} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} F M (n) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 405, + 542, + 785, + 584 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where, $FM_{recon}$ represents the cumulative reconstructed feature map, and $FM(n)$ represents the feature map of each timestamp $n$ .", + "bbox": [ + 212, + 594, + 785, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Methodology", + "text_level": 1, + "bbox": [ + 215, + 648, + 359, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Network Architecture", + "text_level": 1, + "bbox": [ + 215, + 679, + 410, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, the network architecture comprises an end-to-end model with a pure event stream as input data. Subsequently, E-SAI [31] de-occlusion imaging is conducted after the registration, filtering, and reconstruction modules. Specifically, the network consists of two UNet-like ANN models and a SNN [29] model, which are used to handle registration and reconstruction tasks respectively. These two models are connected in the perceptual mask module to effectively exchange information between the registration fusion feature and the reconstructed part.", + "bbox": [ + 212, + 703, + 785, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the registration module, the event stream data undergoes a series of processing, and the UNet with spatial transformer network (STN) [9] model is used for mapping", + "bbox": [ + 212, + 810, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Q. Guo et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/25630602509a326147da88bb1b26a175409d9fe0c4818e4822358625f0e4e59e.jpg", + "image_caption": [ + "Fig. 2: The Architecture of REDIR Network. The model incorporates the UNet registration module with STN, and the SNN filter module, which integrates a Temporal-Spatial Attention Mechanism. The registered target features are then transmitted to the CNN decoder module through the perceptual mask module." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 787, + 345 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and registration to ensure the spatial alignment of events. Then, the filtering module performs filtering operations on the registered data to enhance the expression of useful target information and reduce the impact of obstruction noise. Finally, through the reconstruction module, the SNN model is used to reconstruct the filtered data to achieve de-occlusion imaging of the event stream.", + "bbox": [ + 212, + 431, + 782, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To optimize the integration of registration fusion features and the reconstruction component, a perceptual mask module is introduced. This module facilitates effective information fusion of the outputs of the registration module and the reconstruction module to produce the final imaging result.", + "bbox": [ + 212, + 508, + 784, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Event Registration Module", + "text_level": 1, + "bbox": [ + 214, + 594, + 444, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As displayed in Fig. 3, this module adopts a UNet-like architecture with STN [9] affine transformation module. The STN module enables adaptive alignment of images and features, effectively fusing aligned original data features with subsequent fusion registration features to enhance the accuracy of feature registration.", + "bbox": [ + 212, + 623, + 784, + 685 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e72797d61cd2f0a0a818b19a2379f92ce3d0eb941ab3222f5a425c1e716dd713.jpg", + "image_caption": [ + "Fig. 3: Event Registration Module. The STN is utilized to predict matrices and perform affine transformations in each layer of UNet." + ], + "image_footnote": [], + "bbox": [ + 246, + 712, + 748, + 809 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "REDIR", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The registration network module adopts global and local mapping alignment to improve the registration accuracy of event data. First, the module predicts the transformation matrix $\\psi_t$ of event data for each timestamp $E_t$ , and performs affine registration based on the prediction results. In this way, event data with different timestamps can be effectively aligned into a unified coordinate system, as $E_t'$ . Subsequently, the registration network module utilizes multi-layer perceptron (MLP) to extract features. By performing the above operations on the feature map $F_{l,t}$ , more abstract and meaningful feature information can be further extracted. In order to fuse the feature information of different timestamps [7], the registration network module also uses an upsampling method. Upsampling can convert low-resolution feature maps into high-resolution feature maps to better capture detailed information. By fusing aligned features $F_{l,t}'$ and upsampling features, feature information at different timestamps can be made more consistent and accurate.", + "bbox": [ + 212, + 146, + 789, + 343 + ], + "page_idx": 7 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Event Feature Maps Registration" + ], + "code_body": "1 Input: current period $t$ , current downsampling layer $l$ , unaligned event data $E_{t}$ , unaligned feature maps $F_{l,t}$ \n2 Output: affine matrix $\\psi_{l,t}$ , aligned event data $E_{t}^{\\prime}$ , aligned feature maps $F_{l,t}^{\\prime}$ \n3 $\\psi_{0,t} \\gets STN(E_t)$ Predict the affine transformation matrix of the event frame. \n4 Perform the spatial transformation formula on the event frame: $E_{t}^{\\prime} \\gets TF(E_{t}, \\psi_{0,t})$ \n5 $l = 1$ ; \n6 while $l \\leq L$ do \n7 if $l = L$ then \n8 $\\begin{array}{c|c} F_{l,t} \\gets \\text{Downsampling}(E_t^{\\prime}); // \\text{Down sampling the Registered Event Frames}. \\\\ \\text{end} \\\\ \\text{else} \\\\ F_{l,t} \\gets \\text{Down sampling}(F_{l-1,t}^{\\prime}); // \\text{Down sampling the Registered Feature Frames on the } l - 1 \\text{th Layer.} \\\\ \\text{end} \\\\ t = 1; \\\\ \\text{while } t \\leq T \\text{ do} \\\\ \\text{Predict affine matrix: } \\psi_{l,t} = STN(F_{l,t}); \\\\ \\text{Spatial transform: } F_{l,t}^{\\prime} \\gets TF(F_{l,t}, \\psi_{l,t}); \\\\ t \\gets t + 1 \\\\ \\text{end} \\\\ l \\gets l + 1 \\\\ \\text{end} \\end{array}$ \n9 end", + "bbox": [ + 225, + 381, + 761, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Algorithm 1, the affine transformation parameters predicted by STN include specific values of rotation, scaling, and translation, so the algorithm can solve data registration for other acquisition methods besides uniform linear motion. The original event data of each input timestamp is subjected to parameter prediction and affine transformation to achieve preliminary overall registration; performing the same operation on the extracted feature map can further refine the alignment. The first two columns of the affine parameter matrix include information about the scaling and rotation, and the last column corresponds to the horizontal and vertical translation, achieving a", + "bbox": [ + 212, + 719, + 789, + 843 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Q. Guo et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "six-degree-of-freedom transformation [11]. By predicting this matrix and performing an affine transformation on the original event frame and its feature map, we can obtain the registered event frame, which will facilitate subsequent image reconstruction.", + "bbox": [ + 214, + 146, + 787, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Temporal-Spatial Attention SNN Filter Module", + "text_level": 1, + "bbox": [ + 214, + 214, + 584, + 229 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Fig. 4 illustrates the occlusion event filtering module employs a combination of Temporal-Spatial Attention (TSA) [29] mechanism and spiking neural network (SNN) model to filter out discrete noise events caused by obstructions. The resolving power of the model is improved by extracting the characteristics of targets and occlusions and assigning weights, and the Leaky Integrate-and-Fire mechanism (LIF) [2], is utilized to retain target events that have always existed in adjacent event frames and filter out discontinuous noise events.", + "bbox": [ + 212, + 239, + 787, + 345 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b380d2bfadde4f573533f844d8b77a2415183f0f44b3e120f1fb99ddfaa2d2dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 356, + 671, + 446 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/df2b12b2c21a9d00222369b56d93f66f1ea5b13e1b7440c39f34d7fc6dc50eae.jpg", + "image_caption": [ + "Fig. 4: TSA-SNN Occlusion Event Filtering Module. The SNN module incorporates a Temporal-Spatial Attention Mechanism for skip connection." + ], + "image_footnote": [], + "bbox": [ + 336, + 448, + 656, + 522 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The temporal-spatial attention (TSA) mechanism is incorporated into the image-filtering reconstruction model to enhance the model's perception of spatial and temporal correlations in the image. The SNN employs three neural layers with a leaky integrate-and-fire (LIF) mechanism, while introducing TSA skip connections between these layers.", + "bbox": [ + 212, + 583, + 787, + 643 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The TSA mechanism enables the model to dynamically allocate attention weights to regions and timestamps that contain crucial information in the input feature map $X$ . Subsequently, the output of the 1-st spiking neural layer is sent to the feature map $X_{C}$ obtained by the convolution layer to calculate the attention weight $g_{t}(X_{C})$ and $g_{s}(X_{C})$ in the temporal and spatial dimensions respectively. Weighted feature map $X_{A}$ is then integrated with the feature map $X$ , which is then forwarded to the activation layer and connected to the second spiking neural layer.", + "bbox": [ + 212, + 643, + 787, + 750 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The leakage mechanism of Spiking Neural Networks (SNNs) exclusively responds to persistent signals, effectively filtering out noise events induced by obstructions. The multi-dimensional attention mechanism can further extract the characteristic information of the target, calculate the attention weight, and multiply the feature map from the time, channel, and spatial dimensions, and the obtained result is jump-connected to the final output layer.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "REDIR", + "bbox": [ + 683, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.4 Perceptual Mask Connection Module", + "text_level": 1, + "bbox": [ + 215, + 146, + 516, + 161 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The perceptual mask connection module filters the occlusion events in registered feature maps through a mask operation to extract the target contour information, and then masks features through the convolution layer and sums them with the registered feature maps that have passed through the convolution layer. This module simplifies the information flow path in the network by creating a shortcut to directly transfer features and accumulating the filtered results of each timestamp into the encoder of the reconstruction module. This design not only enhances the network's robustness towards input data and bolsters the representation of salient features but also expedites information transmission. This enables the model to prioritize capturing crucial information, thereby effectively ameliorating the accuracy and efficiency of image reconstruction.", + "bbox": [ + 212, + 170, + 787, + 321 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.5 Loss Function", + "text_level": 1, + "bbox": [ + 215, + 342, + 356, + 356 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The perceptual loss emphasizes semantic similarity, the smoothing loss attends to the visual appearance of the image, and the pixel loss ensures accuracy at the finest level of detail.", + "bbox": [ + 212, + 367, + 787, + 411 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The Perception loss [10] can be expressed as:", + "bbox": [ + 238, + 412, + 540, + 426 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p e r} (\\Phi , \\hat {\\Phi}) = \\sum_ {k} \\frac {\\lambda_ {k}}{C _ {k} H _ {k} W _ {k}} \\| \\Phi_ {k} - \\hat {\\Phi} _ {k} \\| _ {2} ^ {2}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 436, + 785, + 470 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\varPhi_{k}$ and $\\hat{\\varPhi}_k$ with a shape of $C_k\\times H_k\\times W_k$ represent the output feature map of the k-th convolution layer of model output $\\varPhi$ and corresponding ground truth $\\hat{\\varPhi}$ . $\\lambda_{k}$ denotes the weight of the k-th convolution layer.", + "bbox": [ + 214, + 483, + 785, + 529 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The pixel loss [26] can be expressed as:", + "bbox": [ + 238, + 530, + 504, + 545 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p i x} (\\Phi , \\hat {\\Phi}) = \\frac {\\| \\Phi - \\hat {\\Phi} \\| _ {1}}{C H W}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 555, + 785, + 588 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $C\\times H\\times W$ represents the shape of $\\varPhi$ and $\\hat{\\varPhi}$", + "bbox": [ + 214, + 597, + 563, + 613 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The total variance loss [14] can be expressed as:", + "bbox": [ + 238, + 613, + 558, + 628 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t v} (\\Phi) = \\sum_ {i, j} \\left[ \\left(\\phi_ {i, j + 1} - \\phi_ {i, j}\\right) ^ {2} + \\left(\\phi_ {i + 1, j} - \\phi_ {i j}\\right) ^ {2} \\right] ^ {\\frac {1}{2}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 637, + 785, + 672 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\phi_{i,j}$ represents the pixel of $\\varPhi$ .", + "bbox": [ + 214, + 681, + 455, + 696 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The objective of employing these three losses in unison is to meticulously consider various levels and aspects of information, thereby guiding the training of the generative model more comprehensively. By amalgamating these losses, the generative model can be more thoroughly optimized to generate high-quality images across multiple dimensions.", + "bbox": [ + 212, + 696, + 785, + 771 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The weighted total loss function can be expressed as:", + "bbox": [ + 238, + 773, + 591, + 787 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\Phi , \\hat {\\Phi}) = \\beta_ {p e r} L _ {p e r} (\\Phi , \\hat {\\Phi}) + \\beta_ {p i x} L _ {p i x} (\\Phi , \\hat {\\Phi}) + \\beta_ {t v} L _ {t v} (\\Phi). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 302, + 797, + 785, + 815 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where, $\\beta_{per},\\beta_{pix},\\beta_{tv}$ are the weights of these functions.", + "bbox": [ + 214, + 824, + 591, + 842 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Q. Guo et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 143, + 356, + 162 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.1 Experimental Settings", + "text_level": 1, + "bbox": [ + 215, + 176, + 410, + 193 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "V-ESAI Dataset: V-ESAI dataset is formed by data enhancement based on the E-SAI dataset collected by the DAVIS346 camera through random affine transform operations such as rotation, scaling, and translation. It imitates the lens rotation, focal length change, and perspective change during the actual SAI acquisition process. This can improve the generalization ability of the model. The E-SAI dataset contains 488 groups of indoor scenes and 100 groups of outdoor scenes with occlusion E-SAI event data, each group of data is divided by 30 timestamps. V-ESAI dataset includes the original data and the above three random transformations for each E-SAI dataset sample group, expanded to 1,952 indoor groups and 400 outdoor groups, for a total of 2352 groups.", + "bbox": [ + 212, + 203, + 787, + 339 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation Details: The training experiment adopts a batch size of 12 and Adam optimizer with the SGDR schedule [13] reset the learning rate every 64 epochs through NVIDIA GeForce RTX3090 GPUs. The training experiment consists of three stages. First, the event stream is input into the filter reconstruction model for 200 epochs pre-training. The initial learning rate is $5 \\times 10^{-4}$ ; after the training is completed, the parameters of the second part of the model are fixed. The registration model was trained for 500 epochs, with an initial learning rate of $3 \\times 10^{-4}$ ; finally, based on the trained registration model and reconstruction model, the perceptual mask connection module was introduced to fine-tune the model parameters, and the overall training for 300 epochs was performed.", + "bbox": [ + 212, + 345, + 787, + 496 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Results of De-occlusion", + "text_level": 1, + "bbox": [ + 215, + 520, + 418, + 534 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Qualitative Analysis", + "text_level": 1, + "bbox": [ + 215, + 546, + 361, + 561 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As illustrated in Fig. 5, traditional image frame-based methods encounter challenges in scenarios with obstructions, as they struggle to extract target information and effectively mitigate obstructions. Consequently, leveraging the attributes of event cameras becomes imperative to capture event frames triggered during the mobile acquisition process, enabling alignment, filtering, and reconstruction of target and occlusion object information. The REDIR method we propose exhibits robust performance in occlusion removal and target image reconstruction.", + "bbox": [ + 212, + 568, + 787, + 672 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To demonstrate the superiority of our REDIR network, we present a comparative analysis of the visual outcomes generated by diverse reconstruction models on the E-SAI dataset in Fig. 5. The direct reconstruction method employed in the Hybrid [31] model solely possesses the capacity to filter occluders and accumulate misalignment in the target imaging scene, leading to artifacts, image blurring, and loss of detailed information. The reconstruction model incorporating the Refocus module experiences uneven local deformation and saturation distribution due to the discrepancy between the spatial transformation method and the actual imaging camera motion compensation technique. Nevertheless, the REDIR model effectively addresses these issues. By combining UNet with STN, our model can accurately predict the transformation parameters of the global image and local features at each timestamp, performing affine transformation to align", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "REDIR", + "bbox": [ + 683, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/04ebf8454c837d9704836b693da460f4432db3d4b1ed9be6699afbcd5d4d6650.jpg", + "image_caption": [ + "Fig. 5: Comparisons of E-SAI with different de-occlusion reconstruction methods. From left to right, we show four examples with the occluded scenarios, results of Hybrid [31], Refocus+Hybrid [30], and our approach REDIR respectively. Details are zoomed in for better view." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 779, + 369 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "them and thereby significantly reducing motion-induced disturbances. Consequently, the proposed REDIR network generates images with superior reconstruction performance.", + "bbox": [ + 212, + 446, + 787, + 479 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Quantitative Analysis", + "text_level": 1, + "bbox": [ + 215, + 491, + 372, + 507 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/646f83a20a7bb3e8c7f9ca6c57cdd0b74159c5dc03dbfe2f9a7fd6d31a1e86c3.jpg", + "table_caption": [ + "Table 1: De-occlusion reconstruction performance on F-SAI/E-SAI dataset. Discriminated based on the necessity of prior information." + ], + "table_footnote": [], + "table_body": "
MethodRequire Prior InformationIndoorOutdoorTotal
PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
FSAI+ACC [23]✓14.220.34840.295511.860.39250.282113.810.35610.2931
Refocus+Hybrid [30]✓29.550.80860.054620.390.69610.101327.940.78890.0628
FSAI+CNN [25]✗25.010.79330.124517.390.56160.134723.670.75270.1263
Hybrid [31]✗24.040.72410.139414.260.38670.301222.320.66490.1678
REDIR (Ours)✗27.280.76640.111018.530.61300.127525.740.73950.1139
", + "bbox": [ + 218, + 561, + 781, + 655 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In Tab. 1, we comparatively assess the reconstruction performance of the proposed models. Based on the E-SAI dataset for experiments, we conduct quantitative evaluations via PSNR, SSIM, and LPIPS metrics. The Hybrid model demonstrates inadequate handling of defocused data. For E-SAI imaging tasks with occlusion, the Hybrid model [31] directly reconstructs event data without focusing, which results in severe distortion due to the dislocation of event points across timestamps. The performance index of our model is slightly lower compared to that of Refocus+Hybrid [30], however, achieving such reconstruction accuracy requires pre-training on the focused E-SAI dataset with prior information, which necessitates pretraining on the manual-refocused dataset.", + "bbox": [ + 212, + 672, + 787, + 808 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To validate the state-of-the-art performance of our proposed method on a prior information-free dataset, we employ the Refocus+Hybrid [30] method for pre-training", + "bbox": [ + 212, + 810, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Q. Guo et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "and subsequent overall training. However, it is observed that this approach fails to effectively reconstruct the target image. During the training process, if there is no decrease in loss value, it indicates that the model lacks the processing capability for the E-SAI dataset without prior information. In contrast, our model successfully recovers the de-occluded target image without relying on any prior information. Our proposed REDIR method performs an affine transformation on each timestamp, effectively addressing the aforementioned issues and registering without being limited by motion paths and speeds. Consequently, it can effectively reconstruct the target under extremely dense occlusion conditions. Compared to the focused reconstruction model, the REDIR demonstrates superior performance with a PSNR increase of 3.42dB, SSIM increase of $7.46\\%$ , and LPIPS lower than $5.39\\%$ . This corroborates that the novel registration approach achieves data alignment, resulting in a significantly improved reconstruction quality compared to existing algorithms. The model can effectively restore the target scene without occlusion.", + "bbox": [ + 212, + 146, + 787, + 343 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Variable Motion Acquisition Reconstruction", + "text_level": 1, + "bbox": [ + 214, + 369, + 565, + 383 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The capability of our model to register data with scaling, rotation, and translation transformations was demonstrated through verification experiments conducted on the proposed multi-variable event V-ESAI dataset. As illustrated in Fig. 6, a comparative analysis of the outcomes obtained from the three transformation techniques demonstrates that our model exhibits comparable proficiency in handling diverse image transformation tasks, achieving satisfactory performance across all scenarios. Furthermore, the reconstructed image yields visually pleasing results, with.", + "bbox": [ + 212, + 398, + 787, + 503 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Currently, existing event data de-occlusion reconstruction models, such as ESAI and Refocus+Hybrid [30], lack the capability to process samples with transformations like rotation, scaling, and non-uniform translation. The prevailing automatic focusing method Refocus+Hybrid [30] significantly relies on prior information and requires manual focusing followed by pretraining, imposing strict restrictions on the acquisition method. In contrast, our proposed approach enables frame-by-frame alignment of transformed samples without the need for manual focusing step, thereby achieving truly adaptive registration. Experimental analysis demonstrates that our model exhibits an excellent de-occlusion reconstruction effect on samples subjected to non-uniform linear translation, scaling, and rotation transformations. Virtually no front occlusions are observed while effectively restoring the target scene.", + "bbox": [ + 212, + 505, + 787, + 671 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 696, + 362, + 713 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "As illustrated in Tab. 2, to assess the contribution of each module, we performed an ablation experiment. By progressively eliminating STN, PMCM, TSA, and other modules from the Hybrid or UNet-Hybrid model and comparing the reconstruction performance, it becomes evident that the model relies on each module. By incrementally incorporating modules, which contribute significantly to target alignment and feature fusion, the model has been gradually optimized. Finally, the accumulative effect of our contributions in comparison to the reconstruction-only baseline resulted in a substantial improvement of $9.3\\mathrm{dB}$ in PSNR, a $17.51\\%$ increase in SSIM, and a reduction of $12.02\\%$ in LPIPS, thereby establishing a new state of the art for event de-occlusion reconstruction.", + "bbox": [ + 212, + 726, + 787, + 863 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "REDIR", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 769, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9bb1207883e198f9d9ea6dc20308ebc9a804f3f8168f4914b4047d32a228d3fc.jpg", + "image_caption": [ + "Fig. 6: REDIR de-occlusion reconstruction results on V-ESAI dataset. The images from left to right correspond to the original image, the results of translation, scaling and rotation transformation respectively. Details are zoomed in for a better view." + ], + "image_footnote": [], + "bbox": [ + 323, + 143, + 678, + 428 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/7e3314e880eec0e9f2a0cfb05d13dceac7480d78d2652b03f60654277e6fb640.jpg", + "table_caption": [ + "Table 2: Ablation study of various components of our method on E-SAI dataset (part)." + ], + "table_footnote": [], + "table_body": "
ArchitectureSTNPMCMTSAPSNR↑SSIM↑LPIPS↓
Hybrid [31]XXX23.320.70910.1481
UNet+HybridXXX26.450.74380.1358
UNet+HybridX✓✓31.070.78020.0909
UNet+Hybrid✓X✓32.340.86090.0385
UNet+Hybrid✓✓X32.160.82320.0548
UNet+Hybrid✓✓✓32.610.88420.0279
", + "bbox": [ + 323, + 513, + 674, + 614 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 642, + 344, + 657 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we present an end-to-end method REDIR that has the capability of reconstructing events captured by moving cameras in occluded scenarios, while achieving generalized registration across events with variable transformations. As a result, our proposed method REDIR expands the applicability range for acquiring images using event cameras beyond uniform linear motion with fixed focal length or orbit. We introduce a new dataset V-ESAI, specifically designed to evaluate reconstruction capability under various challenging conditions including focus changes, lens rotation, as well as non-uniform motion. Our model achieves a higher level of imaging accuracy without prior information compared to state-of-the-art approaches on E-SAI dataset, while also demonstrating superior reconstruction performance on our V-ESAI dataset when compared to actual target scenarios without occlusion in terms of structural similarity.", + "bbox": [ + 212, + 672, + 789, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Q. Guo et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 145, + 382, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported in part by Science and Technology Innovation (STI) 2030—Major Projects under Grant 2022ZD0208700, and National Natural Science Foundation of China under Grant 62376264.", + "bbox": [ + 212, + 174, + 789, + 219 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 239, + 310, + 255 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Balakrishnan, G., Zhao, A., Sabuncu, M.R., Guttag, J., Dalca, A.V.: Voxelmorph: a learning framework for deformable medical image registration. IEEE transactions on medical imaging 38(8), 1788-1800 (2019) 4", + "2. Chowdhury, S.S., Lee, C., Roy, K.: Towards understanding the effect of leak in spiking neural networks. Neurocomputing 464, 83-94 (2021) 9", + "3. Delbrück, T., Linares-Barranco, B., Culurciello, E., Posch, C.: Activity-driven, event-based vision sensors. In: Proceedings of 2010 IEEE International Symposium on Circuits and Systems. pp. 2426-2429. IEEE (2010) 5", + "4. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., et al.: Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence 44(1), 154-180 (2020) 1", + "5. Gallego, G., Rebecq, H., Scaramuzza, D.: A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3867-3876 (2018) 4", + "6. Gu, C., Learned-Miller, E., Sheldon, D., Gallego, G., Bideau, P.: The spatio-temporaloisson point process: A simple model for the alignment of event camera data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13495-13504 (2021) 5", + "7. Gupta, A., Sharma, P., Ghosh, D., Honkote, V., Ghose, D.: Event-Based Time-To-Contact Estimation with Depth Image Fusion, p. 65-77 (Jan 2023). https://doi.org/10.1007/978-981-19-9379-4_6, http://dx.doi.org/10.1007/978-981-19-9379-4_6 8", + "8. Huang, X., Zhang, Y., Xiong, Z.: Progressive spatio-temporal alignment for efficient event-based motion estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1537-1546 (2023) 5", + "9. Jaderberg, M., Simonyan, K., Zisserman, A., et al.: Spatial transformer networks. Advances in neural information processing systems 28 (2015) 3, 6, 7", + "0. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and superresolution. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 694-711. Springer (2016) 10", + "1. Kim, H., Leutenegger, S., Davison, A.J.: Real-time 3d reconstruction and 6-dof tracking with an event camera. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VI 14. pp. 349-364. Springer (2016) 9", + "2. Liu, D., Parra, A., Chin, T.J.: Spatiotemporal registration for event-based visual odometry. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (Jun 2021). https://doi.org/10.1109/cvpr46437.2021.00490, http://dx.doi.org/10.1109/cvpr46437.2021.00490 3", + "3. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016) 11", + "4. Mahendran, A., Vedaldi, A.: Understanding deep image representations by inverting them. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5188-5196 (2015) 10" + ], + "bbox": [ + 225, + 266, + 787, + 838 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "REDIR", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 769, + 114, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Mitrokhin, A., Fermüller, C., Parameshwara, C., Aloimonos, Y.: Event-based moving object detection and tracking. In: 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1-9. IEEE (2018) 1", + "16. Munda, G., Reinbacher, C., Pock, T.: Real-time intensity-image reconstruction for event cameras using manifold regularisation. International Journal of Computer Vision 126, 1381-1393 (2018) 1", + "17. Neil, D., Pfeiffer, M., Liu, S.C.: Phased LSTM: Accelerating recurrent network training for long or event-based sequences. Cornell University - arXiv, Cornell University - arXiv (Oct 2016) 2", + "18. Pan, L., Hartley, R., Scheerlinck, C., Liu, M., Yu, X., Dai, Y.: High frame rate video reconstruction based on an event camera. IEEE Transactions on Pattern Analysis and Machine Intelligence p. 1-1 (Jan 2020). https://doi.org/10.1109/tpami.2020.3036667, http://dx.doi.org/10.1109/tpami.2020.3036667 1", + "19. Qi, C., Su, H., Mo, K., Guibas, L.: Pointnet: Deep learning on point sets for 3d classification and segmentation. arXiv 2016. arXiv preprint arXiv:1612.00593 4", + "20. Rebecq, H., Gallego, G., Mueggler, E., Scaramuzza, D.: Emvs: Event-based multi-view stereo—3d reconstruction with an event camera in real-time. International Journal of Computer Vision p. 1394–1414 (Dec 2018). https://doi.org/10.1007/s11263-017-1050-6, http://dx.doi.org/10.1007/s11263-017-1050-6 3", + "21. Scheerlinck, C., Rebecq, H., Gehrig, D., Barnes, N., Mahony, R.E., Scaramuzza, D.: Fast image reconstruction with an event camera. In: 2020 IEEE Winter Conference on Applications of Computer Vision (WACV) (Mar 2020). https://doi.org/10.1109/wacv45572.2020.9093366, http://dx.doi.org/10.1109/wacv45572.2020.9093366", + "22. Stearns, C.C., Kannappan, K.: Method for 2-d affine transformation of images (1995) 6", + "23. Vaish, V., Wilburn, B., Joshi, N., Levoy, M.: Using plane+ parallax for calibrating dense camera arrays. In: Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004. vol. 1, pp. I-I. IEEE (2004) 12", + "24. Wang, Y., Wang, L., Yang, J., An, W., Yu, J., Guo, Y.: Spatial-angular interaction for light field image super-resolution. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIII 16. pp. 290-308. Springer (2020) 4", + "25. Wang, Y., Wu, T., Yang, J., Wang, L., An, W., Guo, Y.: Deoccnet: Learning to see through foreground occlusions in light fields. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 118-127 (2020) 4, 12", + "26. Wang, Z., Chen, J., Hoi, S.C.: Deep learning for image super-resolution: A survey. IEEE transactions on pattern analysis and machine intelligence 43(10), 3365-3387 (2020) 10", + "27. Wang, Z., Ng, Y., Goor, P., Mahony, R.: Event camera calibration of per-pixel biased contrast threshold. arXiv: Computer Vision and Pattern Recognition, arXiv: Computer Vision and Pattern Recognition (Dec 2020) 2", + "28. Xu, J., Jiang, M., Yu, L., Yang, W., Wang, W.: Robust motion compensation for event cameras with smooth constraint. IEEE Transactions on Computational Imaging 6, 604-614 (2020) 2", + "29. Yao, M., Zhao, G., Zhang, H., Hu, Y., Deng, L., Tian, Y., Xu, B., Li, G.: Attention spiking neural networks. IEEE transactions on pattern analysis and machine intelligence (2023) 3, 6, 9", + "30. Yu, L., Zhang, X., Liao, W., Yang, W., Xia, G.S.: Learning to see through with events. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(7), 8660-8678 (2023). https://doi.org/10.1109/TPAMI.2022.3227448 2, 4, 12, 13", + "31. Zhang, X., Liao, W., Yu, L., Yang, W., Xia, G.S.: Event-based synthetic aperture imaging with a hybrid network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14235-14244 (2021) 1, 2, 4, 6, 11, 12, 14" + ], + "bbox": [ + 215, + 146, + 787, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Q. Guo et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Zhao, S., Dong, Y., Chang, E.I., Xu, Y., et al.: Recursive cascaded networks for unsupervised medical image registration. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10600-10610 (2019) 4", + "33. Zou, Y., Zheng, Y., Takatani, T., Fu, Y.: Learning to reconstruct high speed and high dynamic range videos from events. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2024-2033 (2021) 2" + ], + "bbox": [ + 215, + 146, + 787, + 231 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "REDIR", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/REDIR_ Refocus-free Event-based De-occlusion Image Reconstruction/59c82a8e-f2d6-49fa-a29f-6faeb9b1bbb2_model.json b/2024/REDIR_ Refocus-free Event-based De-occlusion Image Reconstruction/59c82a8e-f2d6-49fa-a29f-6faeb9b1bbb2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e608545ba49a08f4e1b9b693672e18bc960c0f1c --- /dev/null +++ b/2024/REDIR_ Refocus-free Event-based De-occlusion Image Reconstruction/59c82a8e-f2d6-49fa-a29f-6faeb9b1bbb2_model.json @@ -0,0 +1,2181 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.227, + 0.141, + 0.78, + 0.185 + ], + "angle": 0, + "content": "REDIR: Refocus-free Event-based De-occlusion Image Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.212, + 0.783, + 0.228 + ], + "angle": 0, + "content": "Qi Guo\\(^{1,2}\\), Hailong Shi\\(^{1,2*}\\), Huan Li\\(^{1,2}\\), Jinsheng Xiao\\(^{3}\\), and Xingyu Gao\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.24, + 0.751, + 0.254 + ], + "angle": 0, + "content": "\\(^{1}\\) Institute of Microelectronics of the Chinese Academy of Sciences, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.317, + 0.254, + 0.686, + 0.269 + ], + "angle": 0, + "content": "\\(^{2}\\) University of Chinese Academy of Sciences, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.397, + 0.268, + 0.606, + 0.282 + ], + "angle": 0, + "content": "3 Wuhan University, Hubei, China" + }, + { + "type": "text", + "bbox": [ + 0.355, + 0.283, + 0.646, + 0.296 + ], + "angle": 0, + "content": "{shihailong, gaoxingyu}@ime.ac.cn" + }, + { + "type": "text", + "bbox": [ + 0.355, + 0.297, + 0.65, + 0.309 + ], + "angle": 0, + "content": "https://github.com/QiGuo-46/REDIR" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.347, + 0.744, + 0.596 + ], + "angle": 0, + "content": "Abstract. The employment of the event-based synthetic aperture imaging (E-SAI) technique, which has the capability to capture high-frequency light intensity variations, has facilitated its extensive application on scene de-occlusion reconstruction tasks. However, existing methods usually require prior information and have strict restriction of camera motion on SAI acquisition methods. This paper proposes a novel end-to-end refocus-free variable E-SAI de-occlusion image reconstruction approach REDIR, which can align the global and local features of the variable event data and effectively achieve high-resolution imaging of pure event streams. To further improve the reconstruction of the occluded target, we propose a perceptual mask-gated connection module to interlink information between modules, and incorporate a spatial-temporal attention mechanism into the SNN block to enhance target extraction ability of the model. Through extensive experiments, our model achieves state-of-the-art reconstruction quality on the traditional E-SAI dataset without prior information, while verifying the effectiveness of the variable event data feature registration method on our newly introduced V-ESAI dataset, which obviates the reliance on prior knowledge and extends the applicability of SAI acquisition methods by incorporating focus changes, lens rotations, and non-uniform motion." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.61, + 0.744, + 0.639 + ], + "angle": 0, + "content": "Keywords: Event-based Synthetic Aperture Imaging \\(\\cdot\\) De-occlusion Reconstruction \\(\\cdot\\) Data Registration" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.664, + 0.357, + 0.68 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.696, + 0.788, + 0.786 + ], + "angle": 0, + "content": "The event camera [4] is a biomimetic sensing device that can perceive changes in the intensity of reflected light from objects with high temporal resolution [16]. For application scenarios with dense occlusions, event cameras can be employed to collect target information from different perspectives [18] within the gaps between occlusions, such as between trees, dense fences, and sight obstruction caused by rain and snow, which holds significant value for search and rescue missions and target tracking operations [15]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.787, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Event-based synthetic aperture imaging (E-SAI) [31] is a data acquisition method, as displayed in Fig. 1, which utilizes a linear uniform motion of camera for events and APS" + }, + { + "type": "page_footnote", + "bbox": [ + 0.216, + 0.826, + 0.356, + 0.84 + ], + "angle": 0, + "content": "*Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Q. Guo et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "frames sampling. Through E-SAI acquisition method, the camera can capture events triggered by targets and occlusion in motion [33]. The deep learning models [17] are capable of removing the obstruction of target scenarios on linear uniform motion E-SAI method. However, it is susceptible to lens shakes and motion errors [28], leading to misalignment of event points and subsequent degradation of imaging performance [27]." + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.248, + 0.695, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.551 + ], + "angle": 0, + "content": "Fig. 1: E-SAI De-occlusion Image Reconstruction (without prior information). The event streams are captured through the uniform linear E-SAI acquisition method, the pictures below illustrate the results of de-occlusion imaging by existing method Hybrid [31] and ours." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.789, + 0.704 + ], + "angle": 0, + "content": "Currently, the existing event point aligning methods [30, 31] mainly perform linear affine transformation on the event point by obtaining prior information such as the focal length, moving speed, and depth of field of the event camera, or obtain the affine transformation parameter matrix through model training and perform linear affine transformation. Therefore, the precise alignment of event points is unattainable due to the presence of errors caused by jitter during camera movement and alternative synthetic aperture imaging methods apart from uniform linear motion. Besides, obtaining event camera information in real-time and manually focusing will bring a huge workload. In real-world imaging scenarios, the aforementioned methods are facing two main challenges:" + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.714, + 0.788, + 0.806 + ], + "angle": 0, + "content": "- Strict Restriction of SAI Acquisition Method: The reconstruction performance of imaging tasks relies on the acquisition method of event data. For imaging approaches beyond uniform linear motion, the reconstruction of the target presents substantial obstacles using existing techniques. The limited prediction correction parameters are employed for focusing tasks, while linear offset mapping is utilized to calibrate event data, resulting in inherent limitations of the imaging modality." + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "- Requires Prior Information: Focusing methods in previous work often require information that relies on the acquisition process and requires manual labeling of" + }, + { + "type": "list", + "bbox": [ + 0.231, + 0.714, + 0.788, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.685, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "REDIR" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.147, + 0.786, + 0.178 + ], + "angle": 0, + "content": "samples, which results in increased time costs for the production of training sets and is not generalizable to other acquisition parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.788, + 0.328 + ], + "angle": 0, + "content": "To address the above problems, we propose an end-to-end event registration and reconstruction model REDIR. The synthetic aperture imaging de-occlusion task of event cameras mainly includes event alignment based on a combination of global and local features and reconstruction of preserving target objects by filtering out occlusions. The purpose of alignment registration [20] is primarily to address the non-uniform linear motion error in real synthetic aperture imaging. The adaptive independent registration [12] of multiple event frames aims to address the variable SAI imaging acquisition method without prior information, which lays the foundation for de-occlusion reconstruction [21] of variable SAI imaging acquisition methods without prior information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.788, + 0.463 + ], + "angle": 0, + "content": "The event registration module extracts global and local features from the event data and performs fusion alignment, which utilizes spatial transformer network (STN) [9] to predict the affine matrix and perform spatial transformation for each timestamp event frame, while its feature map is downsampled by UNet. The Temporal-Spatial Attention spiking neural network (SNN) [29] module is used to filter noise and extract target features, remove discrete occlusion event points and retain target event points. Finally, the perceptual mask gate module is used to connect multi-scale local features, which can effectively interchange the information between modules to reconstruct the scene without occlusion." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.465, + 0.631, + 0.48 + ], + "angle": 0, + "content": "The main contributions of the paper can be summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.492, + 0.786, + 0.535 + ], + "angle": 0, + "content": "- We propose a novel event data registration and alignment model, which can automatically map the target event point coordinates of multiple timestamps to the same position, effectively improving the event data reconstruction performance." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.538, + 0.786, + 0.595 + ], + "angle": 0, + "content": "- We further propose an end-to-end reconstruction model REDIR that incorporates a perceptual mask-gated connection module and Temporal-Spatial Attention mechanism, which can reconstruct event data with occlusions into target scene images without visual obstruction." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.599, + 0.785, + 0.657 + ], + "angle": 0, + "content": "- We propose a new multi-variable transformed data augmentation method to compile Variable E-SAI (V-ESAI) dataset, which simulates non-uniform motion, lens rotation, and focal length changes in actual acquisition process, proving our proposed method relevantly persists reconstruction ability under other SAI acquisition methods." + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.492, + 0.786, + 0.657 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.685, + 0.367, + 0.701 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Synthetic aperture imaging can capture the information of the target from different angles, capture more details in the imaging scene, and obtain a larger imaging range under the condition of limited depth of field and focal length. The difference between the event camera and the traditional sensor is that the light intensity change of the scene can be triggered by the moving DVS sensor, so the static image scene can be reconstructed. In addition, for the occluded imaging scene, the imaging method of synthetic aperture can obtain the information of the imaging plane of the target scene through the occlusion at different positions, and fuse and supplement the imaging data of multiple positions," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.117, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.35, + 0.128 + ], + "angle": 0, + "content": "Q. Guo et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.178 + ], + "angle": 0, + "content": "which can effectively remove the occlusion and restore the overall appearance of the target." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.184, + 0.788, + 0.515 + ], + "angle": 0, + "content": "Then, some scholars applied it to the field of optics. In order to obtain complete information on objects through occluding objects, Wang [25] et al. proposed a method of removing foreground occluding based on a deep encoder-decoder network, DeOccNet, for object surface reconstruction in occluding SAI scenarios. In order to obtain multidimensional features and improve the performance of detail reconstruction, Wang [24] et al. improved the method by integrating spatial features and angular features for SAI reconstruction, and the performance reached SOTA. People have gradually found that the event camera has better performance in object reconstruction of occluded scenes. Because the event sensor has the ability to perceive the change of light intensity, it can effectively capture the information of the object in the occluded gap. Zhang [31] et al. used an event camera to collect the results of linear uniform synthetic aperture imaging of the target under extremely dense occlusion and extreme lighting conditions. The event frame registration was realized through manual focusing, and the target scene was restored through filtering and reconstruction by SNN and CNN hybrid models. In order to solve the problem of automatic registration of event data, Yu [30] et al. proposed a method to predict the warping parameter and move the coordinates of each timestamp by linear offset according to the parameter, so as to achieve refocusing. Generally, in the actual acquisition process, the sensor device cannot carry out uniform linear motion perfectly. Therefore, this event registration method based on predicted linear transformation parameters is not suitable for acquisition methods with large motion errors, changes in rotation, focal length and depth of field, and non-uniform linear motion." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.523, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Event-Based Image Registration is one of the necessary steps of E-SAI. The algorithms related to image registration and point cloud registration have been very mature, and are usually applied to the matching, stitching, or superposition of two images/point clouds. The goal of registration is to find the spatial mapping relationship between two images/point clouds, match the extracted features, and map and fuse them so that the two images/point clouds overlap to the same spatial position. For image fusion registration, Zhao [32] et al proposed a general recursive cascade network architecture that can realize deep cascade learning for deformable image registration. Through unsupervised training, the moving image is gradually deformed through each cascade and finally aligned with the fixed image. In order to combine global and local feature information for more refined registration, Balakrishna [1] et al. proposed a medical image registration method based on UNet architecture, which realized the alignment of MRI images of multiple coronal sections of the brain by predicting mapping transformation parameters. For the point cloud registration task, Qi [19] et al. proposed a PointNet, which extracts the global and local features of the point cloud through the T-Net transformation network, and realizes the registration of the point cloud by affine transformation through prediction transformation parameters." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.781, + 0.788, + 0.84 + ], + "angle": 0, + "content": "For event data registration, it is usually necessary to match the coordinates of multiple timestamps to align the event to the same position. Gallego [5] et al demonstrated the ability to handle several important visual problems in event cameras (3D reconstruction, optical flow, and motion estimation), all without additional information about" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.684, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "REDIR" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.388 + ], + "angle": 0, + "content": "the appearance of the scene: by maximizing the alignment of events on the trajectory of points on the image plane. Gu [6] et al. proposed a new probabilistic method for event alignment by modeling the events to be aligned as independent Poisson point processes or spatio-temporal Poisson point processes, so that every pixel is subject to this process. Based on this idea, a likelihood function is derived to estimate the camera motion that best explains the event. It is not limited to rotational motion, but can also be applied to other types of motion, such as translation. Huang [8] et al proposed a unified event-based motion estimation framework applicable to various motion models. By innovatively aligning events with the map, the rotation model and the 6-degree of freedom model are implemented on a standard CPU, and the motion estimation under a small number of sampled events is performed in an efficient manner. At the same time, a flexible batch size strategy is introduced to adapt to different scene textures, camera speed, and resolution. The comprehensive and detailed experimental results show that the framework reaches the current state of the art in terms of accuracy and efficiency. At present, there is no mature deep-learning method for event data registration without prior information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.788, + 0.451 + ], + "angle": 0, + "content": "Inspired by the aforementioned approaches, we propose an event data registration method based on deep learning that enables fusion registration by extracting both global and local feature information from each time stamp. This method is not limited to uniform line SAI but can also be applied to other SAI acquisition methods." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.477, + 0.432, + 0.492 + ], + "angle": 0, + "content": "3 Problem Formulation" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.512, + 0.377, + 0.526 + ], + "angle": 0, + "content": "3.1 Event-Based SAI" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.541, + 0.788, + 0.587 + ], + "angle": 0, + "content": "When the event camera captures changes in light intensity within the event frame, the actual static light intensity at the last timestamp \\( n - 1 \\) and current timestamp \\( n \\) can be expressed as \\( \\Gamma (x_{n},t_{n} - \\Delta t_{n}) \\) and \\( \\Gamma (x_{n},t_{n}) \\) [3]." + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.6, + 0.788, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\Gamma \\left(x _ {n}, t _ {n}\\right) - \\Gamma \\left(x _ {n}, t _ {n} - \\Delta t _ {n}\\right) \\geq p _ {n} \\cdot \\lambda , \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.703 + ], + "angle": 0, + "content": "where, \\(\\lambda\\) represents the threshold, \\(p_n\\) represents the polarity, and \\(\\Delta t_{n}\\) represents the time interval between the adjacent timestamps. The collected event data includes events caused by occlusion and target \\(E_{OA}(n)\\), events caused by occlusion only \\(E_{OO}(n)\\), target only \\(E_{AA}(n)\\), as well as inherent noise of the camera \\(E_0\\). Event data can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.722, + 0.787, + 0.739 + ], + "angle": 0, + "content": "\\[\nE (n) = E _ {O A} (n) + E _ {A A} (n) + E _ {O O} (n) + E _ {0}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.747, + 0.788, + 0.807 + ], + "angle": 0, + "content": "The static light intensity information corresponding to both target and occlusion objects in an actual occlusion imaging scene can be represented as \\(\\Gamma_A(n)\\) and \\(\\Gamma_O(n)\\). Therefore, the target event information captured by an event camera in an occlusion scene \\(E_{OA}(n)\\) can be described as:" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.825, + 0.787, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\left| E _ {O A} (n) \\right| \\propto \\left| \\Gamma_ {A} (n) - \\Gamma_ {O} (n) \\right|. \\tag {3}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Q. Guo et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.388, + 0.162 + ], + "angle": 0, + "content": "3.2 Event Registration" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.172, + 0.788, + 0.232 + ], + "angle": 0, + "content": "Since the synthetic aperture imaging method necessitates camera movement to capture scene information from various viewing angles, accurate target registration is required. The mapping relationship between coordinates before and after the event can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.396, + 0.242, + 0.788, + 0.273 + ], + "angle": 0, + "content": "\\[\nP _ {n} ^ {r e f} = K R _ {n} K ^ {- 1} P _ {n} + \\frac {K T _ {n}}{d}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.277, + 0.785, + 0.338 + ], + "angle": 0, + "content": "where \\( P_{n}^{ref} \\) represents the pixel position of the target on the reference imaging plane, \\( K \\) denotes the camera's internal matrix, \\( R_{n} \\) and \\( T_{n} \\) are rotation and translation matrices that describe the camera postures \\( \\theta_{n} \\) and \\( \\theta_{ref} \\) respectively, \\( d \\) represents the distance between target A and the camera plane." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.787, + 0.399 + ], + "angle": 0, + "content": "The transformation formula [22] encompasses operations such as scaling \\( k \\), rotation \\( \\theta \\), and translation \\( \\Delta x, \\Delta y \\). Through registration, event points at different times can be aligned to a common position, enabling space-time dimension synthetic aperture imaging. The affine matrix \\( \\psi \\) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.398, + 0.409, + 0.787, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\psi = \\left[ \\begin{array}{l l} \\cos \\theta / k & \\sin \\theta / k \\Delta x \\\\ - \\sin \\theta / k & \\cos \\theta / k \\Delta y \\end{array} \\right]. \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.462, + 0.409, + 0.476 + ], + "angle": 0, + "content": "3.3 Event Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.785, + 0.545 + ], + "angle": 0, + "content": "The reconstruction of events is primarily accomplished by accumulating event frame feature maps, with varying timestamps of event points after aligning and occlusion filtering, and performing normalization operations based on the number of timestamps N." + }, + { + "type": "equation", + "bbox": [ + 0.406, + 0.544, + 0.787, + 0.585 + ], + "angle": 0, + "content": "\\[\nF M _ {r e c o n} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} F M (n) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.787, + 0.627 + ], + "angle": 0, + "content": "where, \\( FM_{recon} \\) represents the cumulative reconstructed feature map, and \\( FM(n) \\) represents the feature map of each timestamp \\( n \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.649, + 0.36, + 0.667 + ], + "angle": 0, + "content": "4 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.68, + 0.411, + 0.695 + ], + "angle": 0, + "content": "4.1 Network Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.787, + 0.81 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, the network architecture comprises an end-to-end model with a pure event stream as input data. Subsequently, E-SAI [31] de-occlusion imaging is conducted after the registration, filtering, and reconstruction modules. Specifically, the network consists of two UNet-like ANN models and a SNN [29] model, which are used to handle registration and reconstruction tasks respectively. These two models are connected in the perceptual mask module to effectively exchange information between the registration fusion feature and the reconstructed part." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.787, + 0.842 + ], + "angle": 0, + "content": "In the registration module, the event stream data undergoes a series of processing, and the UNet with spatial transformer network (STN) [9] model is used for mapping" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.685, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "REDIR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.144, + 0.788, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.357, + 0.788, + 0.412 + ], + "angle": 0, + "content": "Fig. 2: The Architecture of REDIR Network. The model incorporates the UNet registration module with STN, and the SNN filter module, which integrates a Temporal-Spatial Attention Mechanism. The registered target features are then transmitted to the CNN decoder module through the perceptual mask module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.784, + 0.507 + ], + "angle": 0, + "content": "and registration to ensure the spatial alignment of events. Then, the filtering module performs filtering operations on the registered data to enhance the expression of useful target information and reduce the impact of obstruction noise. Finally, through the reconstruction module, the SNN model is used to reconstruct the filtered data to achieve de-occlusion imaging of the event stream." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.785, + 0.569 + ], + "angle": 0, + "content": "To optimize the integration of registration fusion features and the reconstruction component, a perceptual mask module is introduced. This module facilitates effective information fusion of the outputs of the registration module and the reconstruction module to produce the final imaging result." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.595, + 0.446, + 0.61 + ], + "angle": 0, + "content": "4.2 Event Registration Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.785, + 0.686 + ], + "angle": 0, + "content": "As displayed in Fig. 3, this module adopts a UNet-like architecture with STN [9] affine transformation module. The STN module enables adaptive alignment of images and features, effectively fusing aligned original data features with subsequent fusion registration features to enhance the accuracy of feature registration." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.713, + 0.749, + 0.81 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.821, + 0.785, + 0.849 + ], + "angle": 0, + "content": "Fig. 3: Event Registration Module. The STN is utilized to predict matrices and perform affine transformations in each layer of UNet." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Q. Guo et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.79, + 0.344 + ], + "angle": 0, + "content": "The registration network module adopts global and local mapping alignment to improve the registration accuracy of event data. First, the module predicts the transformation matrix \\(\\psi_t\\) of event data for each timestamp \\(E_t\\), and performs affine registration based on the prediction results. In this way, event data with different timestamps can be effectively aligned into a unified coordinate system, as \\(E_t'\\). Subsequently, the registration network module utilizes multi-layer perceptron (MLP) to extract features. By performing the above operations on the feature map \\(F_{l,t}\\), more abstract and meaningful feature information can be further extracted. In order to fuse the feature information of different timestamps [7], the registration network module also uses an upsampling method. Upsampling can convert low-resolution feature maps into high-resolution feature maps to better capture detailed information. By fusing aligned features \\(F_{l,t}'\\) and upsampling features, feature information at different timestamps can be made more consistent and accurate." + }, + { + "type": "code_caption", + "bbox": [ + 0.228, + 0.364, + 0.542, + 0.379 + ], + "angle": 0, + "content": "Algorithm 1: Event Feature Maps Registration" + }, + { + "type": "algorithm", + "bbox": [ + 0.226, + 0.382, + 0.763, + 0.702 + ], + "angle": 0, + "content": "1 Input: current period \\(t\\), current downsampling layer \\(l\\), unaligned event data \\(E_{t}\\), unaligned feature maps \\(F_{l,t}\\) \n2 Output: affine matrix \\(\\psi_{l,t}\\), aligned event data \\(E_{t}^{\\prime}\\), aligned feature maps \\(F_{l,t}^{\\prime}\\) \n3 \\(\\psi_{0,t} \\gets STN(E_t)\\) Predict the affine transformation matrix of the event frame. \n4 Perform the spatial transformation formula on the event frame: \\(E_{t}^{\\prime} \\gets TF(E_{t}, \\psi_{0,t})\\) \n5 \\(l = 1\\); \n6 while \\(l \\leq L\\) do \n7 if \\(l = L\\) then \n8 \\(\\begin{array}{c|c} F_{l,t} \\gets \\text{Downsampling}(E_t^{\\prime}); // \\text{Down sampling the Registered Event Frames}. \\\\ \\text{end} \\\\ \\text{else} \\\\ F_{l,t} \\gets \\text{Down sampling}(F_{l-1,t}^{\\prime}); // \\text{Down sampling the Registered Feature Frames on the } l - 1 \\text{th Layer.} \\\\ \\text{end} \\\\ t = 1; \\\\ \\text{while } t \\leq T \\text{ do} \\\\ \\text{Predict affine matrix: } \\psi_{l,t} = STN(F_{l,t}); \\\\ \\text{Spatial transform: } F_{l,t}^{\\prime} \\gets TF(F_{l,t}, \\psi_{l,t}); \\\\ t \\gets t + 1 \\\\ \\text{end} \\\\ l \\gets l + 1 \\\\ \\text{end} \\end{array}\\) \n9 end" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.79, + 0.844 + ], + "angle": 0, + "content": "As shown in Algorithm 1, the affine transformation parameters predicted by STN include specific values of rotation, scaling, and translation, so the algorithm can solve data registration for other acquisition methods besides uniform linear motion. The original event data of each input timestamp is subjected to parameter prediction and affine transformation to achieve preliminary overall registration; performing the same operation on the extracted feature map can further refine the alignment. The first two columns of the affine parameter matrix include information about the scaling and rotation, and the last column corresponds to the horizontal and vertical translation, achieving a" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.684, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "REDIR" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "six-degree-of-freedom transformation [11]. By predicting this matrix and performing an affine transformation on the original event frame and its feature map, we can obtain the registered event frame, which will facilitate subsequent image reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.215, + 0.586, + 0.23 + ], + "angle": 0, + "content": "4.3 Temporal-Spatial Attention SNN Filter Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.24, + 0.789, + 0.346 + ], + "angle": 0, + "content": "Fig. 4 illustrates the occlusion event filtering module employs a combination of Temporal-Spatial Attention (TSA) [29] mechanism and spiking neural network (SNN) model to filter out discrete noise events caused by obstructions. The resolving power of the model is improved by extracting the characteristics of targets and occlusions and assigning weights, and the Leaky Integrate-and-Fire mechanism (LIF) [2], is utilized to retain target events that have always existed in adjacent event frames and filter out discontinuous noise events." + }, + { + "type": "image", + "bbox": [ + 0.33, + 0.357, + 0.672, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.449, + 0.657, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.536, + 0.79, + 0.565 + ], + "angle": 0, + "content": "Fig. 4: TSA-SNN Occlusion Event Filtering Module. The SNN module incorporates a Temporal-Spatial Attention Mechanism for skip connection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.789, + 0.644 + ], + "angle": 0, + "content": "The temporal-spatial attention (TSA) mechanism is incorporated into the image-filtering reconstruction model to enhance the model's perception of spatial and temporal correlations in the image. The SNN employs three neural layers with a leaky integrate-and-fire (LIF) mechanism, while introducing TSA skip connections between these layers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.789, + 0.75 + ], + "angle": 0, + "content": "The TSA mechanism enables the model to dynamically allocate attention weights to regions and timestamps that contain crucial information in the input feature map \\( X \\). Subsequently, the output of the 1-st spiking neural layer is sent to the feature map \\( X_{C} \\) obtained by the convolution layer to calculate the attention weight \\( g_{t}(X_{C}) \\) and \\( g_{s}(X_{C}) \\) in the temporal and spatial dimensions respectively. Weighted feature map \\( X_{A} \\) is then integrated with the feature map \\( X \\), which is then forwarded to the activation layer and connected to the second spiking neural layer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.789, + 0.84 + ], + "angle": 0, + "content": "The leakage mechanism of Spiking Neural Networks (SNNs) exclusively responds to persistent signals, effectively filtering out noise events induced by obstructions. The multi-dimensional attention mechanism can further extract the characteristic information of the target, calculate the attention weight, and multiply the feature map from the time, channel, and spatial dimensions, and the obtained result is jump-connected to the final output layer." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "Q. Guo et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.517, + 0.162 + ], + "angle": 0, + "content": "4.4 Perceptual Mask Connection Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.171, + 0.788, + 0.323 + ], + "angle": 0, + "content": "The perceptual mask connection module filters the occlusion events in registered feature maps through a mask operation to extract the target contour information, and then masks features through the convolution layer and sums them with the registered feature maps that have passed through the convolution layer. This module simplifies the information flow path in the network by creating a shortcut to directly transfer features and accumulating the filtered results of each timestamp into the encoder of the reconstruction module. This design not only enhances the network's robustness towards input data and bolsters the representation of salient features but also expedites information transmission. This enables the model to prioritize capturing crucial information, thereby effectively ameliorating the accuracy and efficiency of image reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.343, + 0.357, + 0.357 + ], + "angle": 0, + "content": "4.5 Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.368, + 0.788, + 0.412 + ], + "angle": 0, + "content": "The perceptual loss emphasizes semantic similarity, the smoothing loss attends to the visual appearance of the image, and the pixel loss ensures accuracy at the finest level of detail." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.413, + 0.542, + 0.428 + ], + "angle": 0, + "content": "The Perception loss [10] can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.36, + 0.437, + 0.787, + 0.472 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p e r} (\\Phi , \\hat {\\Phi}) = \\sum_ {k} \\frac {\\lambda_ {k}}{C _ {k} H _ {k} W _ {k}} \\| \\Phi_ {k} - \\hat {\\Phi} _ {k} \\| _ {2} ^ {2}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.484, + 0.787, + 0.53 + ], + "angle": 0, + "content": "where \\(\\varPhi_{k}\\) and \\(\\hat{\\varPhi}_k\\) with a shape of \\(C_k\\times H_k\\times W_k\\) represent the output feature map of the k-th convolution layer of model output \\(\\varPhi\\) and corresponding ground truth \\(\\hat{\\varPhi}\\). \\(\\lambda_{k}\\) denotes the weight of the k-th convolution layer." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.531, + 0.506, + 0.546 + ], + "angle": 0, + "content": "The pixel loss [26] can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.415, + 0.556, + 0.787, + 0.589 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p i x} (\\Phi , \\hat {\\Phi}) = \\frac {\\| \\Phi - \\hat {\\Phi} \\| _ {1}}{C H W}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.598, + 0.564, + 0.614 + ], + "angle": 0, + "content": "where \\(C\\times H\\times W\\) represents the shape of \\(\\varPhi\\) and \\(\\hat{\\varPhi}\\)" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.614, + 0.56, + 0.629 + ], + "angle": 0, + "content": "The total variance loss [14] can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.324, + 0.638, + 0.787, + 0.673 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t v} (\\Phi) = \\sum_ {i, j} \\left[ \\left(\\phi_ {i, j + 1} - \\phi_ {i, j}\\right) ^ {2} + \\left(\\phi_ {i + 1, j} - \\phi_ {i j}\\right) ^ {2} \\right] ^ {\\frac {1}{2}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.683, + 0.457, + 0.698 + ], + "angle": 0, + "content": "where \\(\\phi_{i,j}\\) represents the pixel of \\(\\varPhi\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.698, + 0.787, + 0.772 + ], + "angle": 0, + "content": "The objective of employing these three losses in unison is to meticulously consider various levels and aspects of information, thereby guiding the training of the generative model more comprehensively. By amalgamating these losses, the generative model can be more thoroughly optimized to generate high-quality images across multiple dimensions." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.774, + 0.592, + 0.789 + ], + "angle": 0, + "content": "The weighted total loss function can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.303, + 0.798, + 0.787, + 0.816 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\Phi , \\hat {\\Phi}) = \\beta_ {p e r} L _ {p e r} (\\Phi , \\hat {\\Phi}) + \\beta_ {p i x} L _ {p i x} (\\Phi , \\hat {\\Phi}) + \\beta_ {t v} L _ {t v} (\\Phi). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.825, + 0.592, + 0.843 + ], + "angle": 0, + "content": "where, \\(\\beta_{per},\\beta_{pix},\\beta_{tv}\\) are the weights of these functions." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.684, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "REDIR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.357, + 0.163 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.178, + 0.411, + 0.194 + ], + "angle": 0, + "content": "5.1 Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.204, + 0.788, + 0.34 + ], + "angle": 0, + "content": "V-ESAI Dataset: V-ESAI dataset is formed by data enhancement based on the E-SAI dataset collected by the DAVIS346 camera through random affine transform operations such as rotation, scaling, and translation. It imitates the lens rotation, focal length change, and perspective change during the actual SAI acquisition process. This can improve the generalization ability of the model. The E-SAI dataset contains 488 groups of indoor scenes and 100 groups of outdoor scenes with occlusion E-SAI event data, each group of data is divided by 30 timestamps. V-ESAI dataset includes the original data and the above three random transformations for each E-SAI dataset sample group, expanded to 1,952 indoor groups and 400 outdoor groups, for a total of 2352 groups." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.347, + 0.788, + 0.497 + ], + "angle": 0, + "content": "Implementation Details: The training experiment adopts a batch size of 12 and Adam optimizer with the SGDR schedule [13] reset the learning rate every 64 epochs through NVIDIA GeForce RTX3090 GPUs. The training experiment consists of three stages. First, the event stream is input into the filter reconstruction model for 200 epochs pre-training. The initial learning rate is \\(5 \\times 10^{-4}\\); after the training is completed, the parameters of the second part of the model are fixed. The registration model was trained for 500 epochs, with an initial learning rate of \\(3 \\times 10^{-4}\\); finally, based on the trained registration model and reconstruction model, the perceptual mask connection module was introduced to fine-tune the model parameters, and the overall training for 300 epochs was performed." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.521, + 0.419, + 0.535 + ], + "angle": 0, + "content": "5.2 Results of De-occlusion" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.547, + 0.362, + 0.562 + ], + "angle": 0, + "content": "Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.674 + ], + "angle": 0, + "content": "As illustrated in Fig. 5, traditional image frame-based methods encounter challenges in scenarios with obstructions, as they struggle to extract target information and effectively mitigate obstructions. Consequently, leveraging the attributes of event cameras becomes imperative to capture event frames triggered during the mobile acquisition process, enabling alignment, filtering, and reconstruction of target and occlusion object information. The REDIR method we propose exhibits robust performance in occlusion removal and target image reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To demonstrate the superiority of our REDIR network, we present a comparative analysis of the visual outcomes generated by diverse reconstruction models on the E-SAI dataset in Fig. 5. The direct reconstruction method employed in the Hybrid [31] model solely possesses the capacity to filter occluders and accumulate misalignment in the target imaging scene, leading to artifacts, image blurring, and loss of detailed information. The reconstruction model incorporating the Refocus module experiences uneven local deformation and saturation distribution due to the discrepancy between the spatial transformation method and the actual imaging camera motion compensation technique. Nevertheless, the REDIR model effectively addresses these issues. By combining UNet with STN, our model can accurately predict the transformation parameters of the global image and local features at each timestamp, performing affine transformation to align" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "Q. Guo et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.781, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.386, + 0.788, + 0.429 + ], + "angle": 0, + "content": "Fig. 5: Comparisons of E-SAI with different de-occlusion reconstruction methods. From left to right, we show four examples with the occluded scenarios, results of Hybrid [31], Refocus+Hybrid [30], and our approach REDIR respectively. Details are zoomed in for better view." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.48 + ], + "angle": 0, + "content": "them and thereby significantly reducing motion-induced disturbances. Consequently, the proposed REDIR network generates images with superior reconstruction performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.492, + 0.373, + 0.508 + ], + "angle": 0, + "content": "Quantitative Analysis" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.522, + 0.787, + 0.551 + ], + "angle": 0, + "content": "Table 1: De-occlusion reconstruction performance on F-SAI/E-SAI dataset. Discriminated based on the necessity of prior information." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.562, + 0.782, + 0.656 + ], + "angle": 0, + "content": "
MethodRequire Prior InformationIndoorOutdoorTotal
PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
FSAI+ACC [23]✓14.220.34840.295511.860.39250.282113.810.35610.2931
Refocus+Hybrid [30]✓29.550.80860.054620.390.69610.101327.940.78890.0628
FSAI+CNN [25]✗25.010.79330.124517.390.56160.134723.670.75270.1263
Hybrid [31]✗24.040.72410.139414.260.38670.301222.320.66490.1678
REDIR (Ours)✗27.280.76640.111018.530.61300.127525.740.73950.1139
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.788, + 0.809 + ], + "angle": 0, + "content": "In Tab. 1, we comparatively assess the reconstruction performance of the proposed models. Based on the E-SAI dataset for experiments, we conduct quantitative evaluations via PSNR, SSIM, and LPIPS metrics. The Hybrid model demonstrates inadequate handling of defocused data. For E-SAI imaging tasks with occlusion, the Hybrid model [31] directly reconstructs event data without focusing, which results in severe distortion due to the dislocation of event points across timestamps. The performance index of our model is slightly lower compared to that of Refocus+Hybrid [30], however, achieving such reconstruction accuracy requires pre-training on the focused E-SAI dataset with prior information, which necessitates pretraining on the manual-refocused dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To validate the state-of-the-art performance of our proposed method on a prior information-free dataset, we employ the Refocus+Hybrid [30] method for pre-training" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.685, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "REDIR" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.344 + ], + "angle": 0, + "content": "and subsequent overall training. However, it is observed that this approach fails to effectively reconstruct the target image. During the training process, if there is no decrease in loss value, it indicates that the model lacks the processing capability for the E-SAI dataset without prior information. In contrast, our model successfully recovers the de-occluded target image without relying on any prior information. Our proposed REDIR method performs an affine transformation on each timestamp, effectively addressing the aforementioned issues and registering without being limited by motion paths and speeds. Consequently, it can effectively reconstruct the target under extremely dense occlusion conditions. Compared to the focused reconstruction model, the REDIR demonstrates superior performance with a PSNR increase of 3.42dB, SSIM increase of \\(7.46\\%\\), and LPIPS lower than \\(5.39\\%\\). This corroborates that the novel registration approach achieves data alignment, resulting in a significantly improved reconstruction quality compared to existing algorithms. The model can effectively restore the target scene without occlusion." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.37, + 0.566, + 0.385 + ], + "angle": 0, + "content": "5.3 Variable Motion Acquisition Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.788, + 0.505 + ], + "angle": 0, + "content": "The capability of our model to register data with scaling, rotation, and translation transformations was demonstrated through verification experiments conducted on the proposed multi-variable event V-ESAI dataset. As illustrated in Fig. 6, a comparative analysis of the outcomes obtained from the three transformation techniques demonstrates that our model exhibits comparable proficiency in handling diverse image transformation tasks, achieving satisfactory performance across all scenarios. Furthermore, the reconstructed image yields visually pleasing results, with." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.788, + 0.672 + ], + "angle": 0, + "content": "Currently, existing event data de-occlusion reconstruction models, such as ESAI and Refocus+Hybrid [30], lack the capability to process samples with transformations like rotation, scaling, and non-uniform translation. The prevailing automatic focusing method Refocus+Hybrid [30] significantly relies on prior information and requires manual focusing followed by pretraining, imposing strict restrictions on the acquisition method. In contrast, our proposed approach enables frame-by-frame alignment of transformed samples without the need for manual focusing step, thereby achieving truly adaptive registration. Experimental analysis demonstrates that our model exhibits an excellent de-occlusion reconstruction effect on samples subjected to non-uniform linear translation, scaling, and rotation transformations. Virtually no front occlusions are observed while effectively restoring the target scene." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.698, + 0.364, + 0.714 + ], + "angle": 0, + "content": "5.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.727, + 0.788, + 0.864 + ], + "angle": 0, + "content": "As illustrated in Tab. 2, to assess the contribution of each module, we performed an ablation experiment. By progressively eliminating STN, PMCM, TSA, and other modules from the Hybrid or UNet-Hybrid model and comparing the reconstruction performance, it becomes evident that the model relies on each module. By incrementally incorporating modules, which contribute significantly to target alignment and feature fusion, the model has been gradually optimized. Finally, the accumulative effect of our contributions in comparison to the reconstruction-only baseline resulted in a substantial improvement of \\(9.3\\mathrm{dB}\\) in PSNR, a \\(17.51\\%\\) increase in SSIM, and a reduction of \\(12.02\\%\\) in LPIPS, thereby establishing a new state of the art for event de-occlusion reconstruction." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "Q. Guo et al." + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.144, + 0.679, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.443, + 0.788, + 0.485 + ], + "angle": 0, + "content": "Fig. 6: REDIR de-occlusion reconstruction results on V-ESAI dataset. The images from left to right correspond to the original image, the results of translation, scaling and rotation transformation respectively. Details are zoomed in for a better view." + }, + { + "type": "table_caption", + "bbox": [ + 0.244, + 0.488, + 0.757, + 0.502 + ], + "angle": 0, + "content": "Table 2: Ablation study of various components of our method on E-SAI dataset (part)." + }, + { + "type": "table", + "bbox": [ + 0.324, + 0.514, + 0.676, + 0.616 + ], + "angle": 0, + "content": "
ArchitectureSTNPMCMTSAPSNR↑SSIM↑LPIPS↓
Hybrid [31]XXX23.320.70910.1481
UNet+HybridXXX26.450.74380.1358
UNet+HybridX✓✓31.070.78020.0909
UNet+Hybrid✓X✓32.340.86090.0385
UNet+Hybrid✓✓X32.160.82320.0548
UNet+Hybrid✓✓✓32.610.88420.0279
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.643, + 0.345, + 0.658 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.79, + 0.841 + ], + "angle": 0, + "content": "In this paper, we present an end-to-end method REDIR that has the capability of reconstructing events captured by moving cameras in occluded scenarios, while achieving generalized registration across events with variable transformations. As a result, our proposed method REDIR expands the applicability range for acquiring images using event cameras beyond uniform linear motion with fixed focal length or orbit. We introduce a new dataset V-ESAI, specifically designed to evaluate reconstruction capability under various challenging conditions including focus changes, lens rotation, as well as non-uniform motion. Our model achieves a higher level of imaging accuracy without prior information compared to state-of-the-art approaches on E-SAI dataset, while also demonstrating superior reconstruction performance on our V-ESAI dataset when compared to actual target scenarios without occlusion in terms of structural similarity." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.685, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "REDIR" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.146, + 0.383, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.79, + 0.22 + ], + "angle": 0, + "content": "This work was supported in part by Science and Technology Innovation (STI) 2030—Major Projects under Grant 2022ZD0208700, and National Natural Science Foundation of China under Grant 62376264." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.24, + 0.312, + 0.256 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.267, + 0.787, + 0.308 + ], + "angle": 0, + "content": "1. Balakrishnan, G., Zhao, A., Sabuncu, M.R., Guttag, J., Dalca, A.V.: Voxelmorph: a learning framework for deformable medical image registration. IEEE transactions on medical imaging 38(8), 1788-1800 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.787, + 0.336 + ], + "angle": 0, + "content": "2. Chowdhury, S.S., Lee, C., Roy, K.: Towards understanding the effect of leak in spiking neural networks. Neurocomputing 464, 83-94 (2021) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.337, + 0.787, + 0.377 + ], + "angle": 0, + "content": "3. Delbrück, T., Linares-Barranco, B., Culurciello, E., Posch, C.: Activity-driven, event-based vision sensors. In: Proceedings of 2010 IEEE International Symposium on Circuits and Systems. pp. 2426-2429. IEEE (2010) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.379, + 0.787, + 0.418 + ], + "angle": 0, + "content": "4. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., et al.: Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence 44(1), 154-180 (2020) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.419, + 0.787, + 0.459 + ], + "angle": 0, + "content": "5. Gallego, G., Rebecq, H., Scaramuzza, D.: A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3867-3876 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.46, + 0.787, + 0.499 + ], + "angle": 0, + "content": "6. Gu, C., Learned-Miller, E., Sheldon, D., Gallego, G., Bideau, P.: The spatio-temporaloisson point process: A simple model for the alignment of event camera data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13495-13504 (2021) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.5, + 0.788, + 0.553 + ], + "angle": 0, + "content": "7. Gupta, A., Sharma, P., Ghosh, D., Honkote, V., Ghose, D.: Event-Based Time-To-Contact Estimation with Depth Image Fusion, p. 65-77 (Jan 2023). https://doi.org/10.1007/978-981-19-9379-4_6, http://dx.doi.org/10.1007/978-981-19-9379-4_6 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.554, + 0.787, + 0.595 + ], + "angle": 0, + "content": "8. Huang, X., Zhang, Y., Xiong, Z.: Progressive spatio-temporal alignment for efficient event-based motion estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1537-1546 (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.596, + 0.787, + 0.621 + ], + "angle": 0, + "content": "9. Jaderberg, M., Simonyan, K., Zisserman, A., et al.: Spatial transformer networks. Advances in neural information processing systems 28 (2015) 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.622, + 0.787, + 0.662 + ], + "angle": 0, + "content": "0. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and superresolution. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 694-711. Springer (2016) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.663, + 0.787, + 0.715 + ], + "angle": 0, + "content": "1. Kim, H., Leutenegger, S., Davison, A.J.: Real-time 3d reconstruction and 6-dof tracking with an event camera. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VI 14. pp. 349-364. Springer (2016) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.787, + 0.77 + ], + "angle": 0, + "content": "2. Liu, D., Parra, A., Chin, T.J.: Spatiotemporal registration for event-based visual odometry. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (Jun 2021). https://doi.org/10.1109/cvpr46437.2021.00490, http://dx.doi.org/10.1109/cvpr46437.2021.00490 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.787, + 0.799 + ], + "angle": 0, + "content": "3. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.787, + 0.839 + ], + "angle": 0, + "content": "4. Mahendran, A., Vedaldi, A.: Understanding deep image representations by inverting them. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5188-5196 (2015) 10" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.267, + 0.788, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.128 + ], + "angle": 0, + "content": "Q. Guo et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "15. Mitrokhin, A., Fermüller, C., Parameshwara, C., Aloimonos, Y.: Event-based moving object detection and tracking. In: 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1-9. IEEE (2018) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.788, + 0.231 + ], + "angle": 0, + "content": "16. Munda, G., Reinbacher, C., Pock, T.: Real-time intensity-image reconstruction for event cameras using manifold regularisation. International Journal of Computer Vision 126, 1381-1393 (2018) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.259 + ], + "angle": 0, + "content": "17. Neil, D., Pfeiffer, M., Liu, S.C.: Phased LSTM: Accelerating recurrent network training for long or event-based sequences. Cornell University - arXiv, Cornell University - arXiv (Oct 2016) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.26, + 0.788, + 0.313 + ], + "angle": 0, + "content": "18. Pan, L., Hartley, R., Scheerlinck, C., Liu, M., Yu, X., Dai, Y.: High frame rate video reconstruction based on an event camera. IEEE Transactions on Pattern Analysis and Machine Intelligence p. 1-1 (Jan 2020). https://doi.org/10.1109/tpami.2020.3036667, http://dx.doi.org/10.1109/tpami.2020.3036667 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.315, + 0.787, + 0.342 + ], + "angle": 0, + "content": "19. Qi, C., Su, H., Mo, K., Guibas, L.: Pointnet: Deep learning on point sets for 3d classification and segmentation. arXiv 2016. arXiv preprint arXiv:1612.00593 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.343, + 0.788, + 0.396 + ], + "angle": 0, + "content": "20. Rebecq, H., Gallego, G., Mueggler, E., Scaramuzza, D.: Emvs: Event-based multi-view stereo—3d reconstruction with an event camera in real-time. International Journal of Computer Vision p. 1394–1414 (Dec 2018). https://doi.org/10.1007/s11263-017-1050-6, http://dx.doi.org/10.1007/s11263-017-1050-6 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.398, + 0.788, + 0.464 + ], + "angle": 0, + "content": "21. Scheerlinck, C., Rebecq, H., Gehrig, D., Barnes, N., Mahony, R.E., Scaramuzza, D.: Fast image reconstruction with an event camera. In: 2020 IEEE Winter Conference on Applications of Computer Vision (WACV) (Mar 2020). https://doi.org/10.1109/wacv45572.2020.9093366, http://dx.doi.org/10.1109/wacv45572.2020.9093366" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.761, + 0.48 + ], + "angle": 0, + "content": "22. Stearns, C.C., Kannappan, K.: Method for 2-d affine transformation of images (1995) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "23. Vaish, V., Wilburn, B., Joshi, N., Levoy, M.: Using plane+ parallax for calibrating dense camera arrays. In: Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004. vol. 1, pp. I-I. IEEE (2004) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.523, + 0.787, + 0.576 + ], + "angle": 0, + "content": "24. Wang, Y., Wang, L., Yang, J., An, W., Yu, J., Guo, Y.: Spatial-angular interaction for light field image super-resolution. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIII 16. pp. 290-308. Springer (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.578, + 0.787, + 0.619 + ], + "angle": 0, + "content": "25. Wang, Y., Wu, T., Yang, J., Wang, L., An, W., Guo, Y.: Deoccnet: Learning to see through foreground occlusions in light fields. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 118-127 (2020) 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.62, + 0.787, + 0.646 + ], + "angle": 0, + "content": "26. Wang, Z., Chen, J., Hoi, S.C.: Deep learning for image super-resolution: A survey. IEEE transactions on pattern analysis and machine intelligence 43(10), 3365-3387 (2020) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.647, + 0.787, + 0.688 + ], + "angle": 0, + "content": "27. Wang, Z., Ng, Y., Goor, P., Mahony, R.: Event camera calibration of per-pixel biased contrast threshold. arXiv: Computer Vision and Pattern Recognition, arXiv: Computer Vision and Pattern Recognition (Dec 2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.716 + ], + "angle": 0, + "content": "28. Xu, J., Jiang, M., Yu, L., Yang, W., Wang, W.: Robust motion compensation for event cameras with smooth constraint. IEEE Transactions on Computational Imaging 6, 604-614 (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.717, + 0.787, + 0.755 + ], + "angle": 0, + "content": "29. Yao, M., Zhao, G., Zhang, H., Hu, Y., Deng, L., Tian, Y., Xu, B., Li, G.: Attention spiking neural networks. IEEE transactions on pattern analysis and machine intelligence (2023) 3, 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.757, + 0.787, + 0.798 + ], + "angle": 0, + "content": "30. Yu, L., Zhang, X., Liao, W., Yang, W., Xia, G.S.: Learning to see through with events. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(7), 8660-8678 (2023). https://doi.org/10.1109/TPAMI.2022.3227448 2, 4, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "31. Zhang, X., Liao, W., Yu, L., Yang, W., Xia, G.S.: Event-based synthetic aperture imaging with a hybrid network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14235-14244 (2021) 1, 2, 4, 6, 11, 12, 14" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.685, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "REDIR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "32. Zhao, S., Dong, Y., Chang, E.I., Xu, Y., et al.: Recursive cascaded networks for unsupervised medical image registration. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10600-10610 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "33. Zou, Y., Zheng, Y., Takatani, T., Fu, Y.: Learning to reconstruct high speed and high dynamic range videos from events. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2024-2033 (2021) 2" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.232 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/REDIR_ Refocus-free Event-based De-occlusion Image Reconstruction/59c82a8e-f2d6-49fa-a29f-6faeb9b1bbb2_origin.pdf b/2024/REDIR_ Refocus-free Event-based De-occlusion Image Reconstruction/59c82a8e-f2d6-49fa-a29f-6faeb9b1bbb2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..155dc6204a8118ab1c8b8292db7a21ada35f10a8 --- /dev/null +++ b/2024/REDIR_ Refocus-free Event-based De-occlusion Image Reconstruction/59c82a8e-f2d6-49fa-a29f-6faeb9b1bbb2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e4a868ea1e8e55be9c74fa937051c3f1b9e3412f97e3575523b8fab0de7a846 +size 3296265 diff --git a/2024/REDIR_ Refocus-free Event-based De-occlusion Image Reconstruction/full.md b/2024/REDIR_ Refocus-free Event-based De-occlusion Image Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..19f4b12cf8ed31cc04c74806e14c0ab04b07f6f7 --- /dev/null +++ b/2024/REDIR_ Refocus-free Event-based De-occlusion Image Reconstruction/full.md @@ -0,0 +1,301 @@ +# REDIR: Refocus-free Event-based De-occlusion Image Reconstruction + +Qi Guo $^{1,2}$ , Hailong Shi $^{1,2*}$ , Huan Li $^{1,2}$ , Jinsheng Xiao $^{3}$ , and Xingyu Gao $^{1,2*}$ + +$^{1}$ Institute of Microelectronics of the Chinese Academy of Sciences, Beijing, China + +$^{2}$ University of Chinese Academy of Sciences, Beijing, China + +3 Wuhan University, Hubei, China + +{shihailong, gaoxingyu}@ime.ac.cn + +https://github.com/QiGuo-46/REDIR + +Abstract. The employment of the event-based synthetic aperture imaging (E-SAI) technique, which has the capability to capture high-frequency light intensity variations, has facilitated its extensive application on scene de-occlusion reconstruction tasks. However, existing methods usually require prior information and have strict restriction of camera motion on SAI acquisition methods. This paper proposes a novel end-to-end refocus-free variable E-SAI de-occlusion image reconstruction approach REDIR, which can align the global and local features of the variable event data and effectively achieve high-resolution imaging of pure event streams. To further improve the reconstruction of the occluded target, we propose a perceptual mask-gated connection module to interlink information between modules, and incorporate a spatial-temporal attention mechanism into the SNN block to enhance target extraction ability of the model. Through extensive experiments, our model achieves state-of-the-art reconstruction quality on the traditional E-SAI dataset without prior information, while verifying the effectiveness of the variable event data feature registration method on our newly introduced V-ESAI dataset, which obviates the reliance on prior knowledge and extends the applicability of SAI acquisition methods by incorporating focus changes, lens rotations, and non-uniform motion. + +Keywords: Event-based Synthetic Aperture Imaging $\cdot$ De-occlusion Reconstruction $\cdot$ Data Registration + +# 1 Introduction + +The event camera [4] is a biomimetic sensing device that can perceive changes in the intensity of reflected light from objects with high temporal resolution [16]. For application scenarios with dense occlusions, event cameras can be employed to collect target information from different perspectives [18] within the gaps between occlusions, such as between trees, dense fences, and sight obstruction caused by rain and snow, which holds significant value for search and rescue missions and target tracking operations [15]. + +Event-based synthetic aperture imaging (E-SAI) [31] is a data acquisition method, as displayed in Fig. 1, which utilizes a linear uniform motion of camera for events and APS + +frames sampling. Through E-SAI acquisition method, the camera can capture events triggered by targets and occlusion in motion [33]. The deep learning models [17] are capable of removing the obstruction of target scenarios on linear uniform motion E-SAI method. However, it is susceptible to lens shakes and motion errors [28], leading to misalignment of event points and subsequent degradation of imaging performance [27]. + +![](images/7a923e9eee67bf8fb23f858bb59f0241ab18a8f62ef1d180d25e044fcad41d6a.jpg) +Fig. 1: E-SAI De-occlusion Image Reconstruction (without prior information). The event streams are captured through the uniform linear E-SAI acquisition method, the pictures below illustrate the results of de-occlusion imaging by existing method Hybrid [31] and ours. + +Currently, the existing event point aligning methods [30, 31] mainly perform linear affine transformation on the event point by obtaining prior information such as the focal length, moving speed, and depth of field of the event camera, or obtain the affine transformation parameter matrix through model training and perform linear affine transformation. Therefore, the precise alignment of event points is unattainable due to the presence of errors caused by jitter during camera movement and alternative synthetic aperture imaging methods apart from uniform linear motion. Besides, obtaining event camera information in real-time and manually focusing will bring a huge workload. In real-world imaging scenarios, the aforementioned methods are facing two main challenges: + +- Strict Restriction of SAI Acquisition Method: The reconstruction performance of imaging tasks relies on the acquisition method of event data. For imaging approaches beyond uniform linear motion, the reconstruction of the target presents substantial obstacles using existing techniques. The limited prediction correction parameters are employed for focusing tasks, while linear offset mapping is utilized to calibrate event data, resulting in inherent limitations of the imaging modality. +- Requires Prior Information: Focusing methods in previous work often require information that relies on the acquisition process and requires manual labeling of + +samples, which results in increased time costs for the production of training sets and is not generalizable to other acquisition parameters. + +To address the above problems, we propose an end-to-end event registration and reconstruction model REDIR. The synthetic aperture imaging de-occlusion task of event cameras mainly includes event alignment based on a combination of global and local features and reconstruction of preserving target objects by filtering out occlusions. The purpose of alignment registration [20] is primarily to address the non-uniform linear motion error in real synthetic aperture imaging. The adaptive independent registration [12] of multiple event frames aims to address the variable SAI imaging acquisition method without prior information, which lays the foundation for de-occlusion reconstruction [21] of variable SAI imaging acquisition methods without prior information. + +The event registration module extracts global and local features from the event data and performs fusion alignment, which utilizes spatial transformer network (STN) [9] to predict the affine matrix and perform spatial transformation for each timestamp event frame, while its feature map is downsampled by UNet. The Temporal-Spatial Attention spiking neural network (SNN) [29] module is used to filter noise and extract target features, remove discrete occlusion event points and retain target event points. Finally, the perceptual mask gate module is used to connect multi-scale local features, which can effectively interchange the information between modules to reconstruct the scene without occlusion. + +The main contributions of the paper can be summarized as: + +- We propose a novel event data registration and alignment model, which can automatically map the target event point coordinates of multiple timestamps to the same position, effectively improving the event data reconstruction performance. +- We further propose an end-to-end reconstruction model REDIR that incorporates a perceptual mask-gated connection module and Temporal-Spatial Attention mechanism, which can reconstruct event data with occlusions into target scene images without visual obstruction. +- We propose a new multi-variable transformed data augmentation method to compile Variable E-SAI (V-ESAI) dataset, which simulates non-uniform motion, lens rotation, and focal length changes in actual acquisition process, proving our proposed method relevantly persists reconstruction ability under other SAI acquisition methods. + +# 2 Related Work + +Synthetic aperture imaging can capture the information of the target from different angles, capture more details in the imaging scene, and obtain a larger imaging range under the condition of limited depth of field and focal length. The difference between the event camera and the traditional sensor is that the light intensity change of the scene can be triggered by the moving DVS sensor, so the static image scene can be reconstructed. In addition, for the occluded imaging scene, the imaging method of synthetic aperture can obtain the information of the imaging plane of the target scene through the occlusion at different positions, and fuse and supplement the imaging data of multiple positions, + +which can effectively remove the occlusion and restore the overall appearance of the target. + +Then, some scholars applied it to the field of optics. In order to obtain complete information on objects through occluding objects, Wang [25] et al. proposed a method of removing foreground occluding based on a deep encoder-decoder network, DeOccNet, for object surface reconstruction in occluding SAI scenarios. In order to obtain multidimensional features and improve the performance of detail reconstruction, Wang [24] et al. improved the method by integrating spatial features and angular features for SAI reconstruction, and the performance reached SOTA. People have gradually found that the event camera has better performance in object reconstruction of occluded scenes. Because the event sensor has the ability to perceive the change of light intensity, it can effectively capture the information of the object in the occluded gap. Zhang [31] et al. used an event camera to collect the results of linear uniform synthetic aperture imaging of the target under extremely dense occlusion and extreme lighting conditions. The event frame registration was realized through manual focusing, and the target scene was restored through filtering and reconstruction by SNN and CNN hybrid models. In order to solve the problem of automatic registration of event data, Yu [30] et al. proposed a method to predict the warping parameter and move the coordinates of each timestamp by linear offset according to the parameter, so as to achieve refocusing. Generally, in the actual acquisition process, the sensor device cannot carry out uniform linear motion perfectly. Therefore, this event registration method based on predicted linear transformation parameters is not suitable for acquisition methods with large motion errors, changes in rotation, focal length and depth of field, and non-uniform linear motion. + +Event-Based Image Registration is one of the necessary steps of E-SAI. The algorithms related to image registration and point cloud registration have been very mature, and are usually applied to the matching, stitching, or superposition of two images/point clouds. The goal of registration is to find the spatial mapping relationship between two images/point clouds, match the extracted features, and map and fuse them so that the two images/point clouds overlap to the same spatial position. For image fusion registration, Zhao [32] et al proposed a general recursive cascade network architecture that can realize deep cascade learning for deformable image registration. Through unsupervised training, the moving image is gradually deformed through each cascade and finally aligned with the fixed image. In order to combine global and local feature information for more refined registration, Balakrishna [1] et al. proposed a medical image registration method based on UNet architecture, which realized the alignment of MRI images of multiple coronal sections of the brain by predicting mapping transformation parameters. For the point cloud registration task, Qi [19] et al. proposed a PointNet, which extracts the global and local features of the point cloud through the T-Net transformation network, and realizes the registration of the point cloud by affine transformation through prediction transformation parameters. + +For event data registration, it is usually necessary to match the coordinates of multiple timestamps to align the event to the same position. Gallego [5] et al demonstrated the ability to handle several important visual problems in event cameras (3D reconstruction, optical flow, and motion estimation), all without additional information about + +the appearance of the scene: by maximizing the alignment of events on the trajectory of points on the image plane. Gu [6] et al. proposed a new probabilistic method for event alignment by modeling the events to be aligned as independent Poisson point processes or spatio-temporal Poisson point processes, so that every pixel is subject to this process. Based on this idea, a likelihood function is derived to estimate the camera motion that best explains the event. It is not limited to rotational motion, but can also be applied to other types of motion, such as translation. Huang [8] et al proposed a unified event-based motion estimation framework applicable to various motion models. By innovatively aligning events with the map, the rotation model and the 6-degree of freedom model are implemented on a standard CPU, and the motion estimation under a small number of sampled events is performed in an efficient manner. At the same time, a flexible batch size strategy is introduced to adapt to different scene textures, camera speed, and resolution. The comprehensive and detailed experimental results show that the framework reaches the current state of the art in terms of accuracy and efficiency. At present, there is no mature deep-learning method for event data registration without prior information. + +Inspired by the aforementioned approaches, we propose an event data registration method based on deep learning that enables fusion registration by extracting both global and local feature information from each time stamp. This method is not limited to uniform line SAI but can also be applied to other SAI acquisition methods. + +# 3 Problem Formulation + +# 3.1 Event-Based SAI + +When the event camera captures changes in light intensity within the event frame, the actual static light intensity at the last timestamp $n - 1$ and current timestamp $n$ can be expressed as $\Gamma (x_{n},t_{n} - \Delta t_{n})$ and $\Gamma (x_{n},t_{n})$ [3]. + +$$ +\Gamma \left(x _ {n}, t _ {n}\right) - \Gamma \left(x _ {n}, t _ {n} - \Delta t _ {n}\right) \geq p _ {n} \cdot \lambda , \tag {1} +$$ + +where, $\lambda$ represents the threshold, $p_n$ represents the polarity, and $\Delta t_{n}$ represents the time interval between the adjacent timestamps. The collected event data includes events caused by occlusion and target $E_{OA}(n)$ , events caused by occlusion only $E_{OO}(n)$ , target only $E_{AA}(n)$ , as well as inherent noise of the camera $E_0$ . Event data can be represented as: + +$$ +E (n) = E _ {O A} (n) + E _ {A A} (n) + E _ {O O} (n) + E _ {0}, \tag {2} +$$ + +The static light intensity information corresponding to both target and occlusion objects in an actual occlusion imaging scene can be represented as $\Gamma_A(n)$ and $\Gamma_O(n)$ . Therefore, the target event information captured by an event camera in an occlusion scene $E_{OA}(n)$ can be described as: + +$$ +\left| E _ {O A} (n) \right| \propto \left| \Gamma_ {A} (n) - \Gamma_ {O} (n) \right|. \tag {3} +$$ + +# 3.2 Event Registration + +Since the synthetic aperture imaging method necessitates camera movement to capture scene information from various viewing angles, accurate target registration is required. The mapping relationship between coordinates before and after the event can be formulated as follows: + +$$ +P _ {n} ^ {r e f} = K R _ {n} K ^ {- 1} P _ {n} + \frac {K T _ {n}}{d}, \tag {4} +$$ + +where $P_{n}^{ref}$ represents the pixel position of the target on the reference imaging plane, $K$ denotes the camera's internal matrix, $R_{n}$ and $T_{n}$ are rotation and translation matrices that describe the camera postures $\theta_{n}$ and $\theta_{ref}$ respectively, $d$ represents the distance between target A and the camera plane. + +The transformation formula [22] encompasses operations such as scaling $k$ , rotation $\theta$ , and translation $\Delta x, \Delta y$ . Through registration, event points at different times can be aligned to a common position, enabling space-time dimension synthetic aperture imaging. The affine matrix $\psi$ can be expressed as: + +$$ +\psi = \left[ \begin{array}{l l} \cos \theta / k & \sin \theta / k \Delta x \\ - \sin \theta / k & \cos \theta / k \Delta y \end{array} \right]. \tag {5} +$$ + +# 3.3 Event Reconstruction + +The reconstruction of events is primarily accomplished by accumulating event frame feature maps, with varying timestamps of event points after aligning and occlusion filtering, and performing normalization operations based on the number of timestamps N. + +$$ +F M _ {r e c o n} = \frac {1}{N} \sum_ {n = 1} ^ {N} F M (n) \tag {6} +$$ + +where, $FM_{recon}$ represents the cumulative reconstructed feature map, and $FM(n)$ represents the feature map of each timestamp $n$ . + +# 4 Methodology + +# 4.1 Network Architecture + +As illustrated in Fig. 2, the network architecture comprises an end-to-end model with a pure event stream as input data. Subsequently, E-SAI [31] de-occlusion imaging is conducted after the registration, filtering, and reconstruction modules. Specifically, the network consists of two UNet-like ANN models and a SNN [29] model, which are used to handle registration and reconstruction tasks respectively. These two models are connected in the perceptual mask module to effectively exchange information between the registration fusion feature and the reconstructed part. + +In the registration module, the event stream data undergoes a series of processing, and the UNet with spatial transformer network (STN) [9] model is used for mapping + +![](images/25630602509a326147da88bb1b26a175409d9fe0c4818e4822358625f0e4e59e.jpg) +Fig. 2: The Architecture of REDIR Network. The model incorporates the UNet registration module with STN, and the SNN filter module, which integrates a Temporal-Spatial Attention Mechanism. The registered target features are then transmitted to the CNN decoder module through the perceptual mask module. + +and registration to ensure the spatial alignment of events. Then, the filtering module performs filtering operations on the registered data to enhance the expression of useful target information and reduce the impact of obstruction noise. Finally, through the reconstruction module, the SNN model is used to reconstruct the filtered data to achieve de-occlusion imaging of the event stream. + +To optimize the integration of registration fusion features and the reconstruction component, a perceptual mask module is introduced. This module facilitates effective information fusion of the outputs of the registration module and the reconstruction module to produce the final imaging result. + +# 4.2 Event Registration Module + +As displayed in Fig. 3, this module adopts a UNet-like architecture with STN [9] affine transformation module. The STN module enables adaptive alignment of images and features, effectively fusing aligned original data features with subsequent fusion registration features to enhance the accuracy of feature registration. + +![](images/e72797d61cd2f0a0a818b19a2379f92ce3d0eb941ab3222f5a425c1e716dd713.jpg) +Fig. 3: Event Registration Module. The STN is utilized to predict matrices and perform affine transformations in each layer of UNet. + +The registration network module adopts global and local mapping alignment to improve the registration accuracy of event data. First, the module predicts the transformation matrix $\psi_t$ of event data for each timestamp $E_t$ , and performs affine registration based on the prediction results. In this way, event data with different timestamps can be effectively aligned into a unified coordinate system, as $E_t'$ . Subsequently, the registration network module utilizes multi-layer perceptron (MLP) to extract features. By performing the above operations on the feature map $F_{l,t}$ , more abstract and meaningful feature information can be further extracted. In order to fuse the feature information of different timestamps [7], the registration network module also uses an upsampling method. Upsampling can convert low-resolution feature maps into high-resolution feature maps to better capture detailed information. By fusing aligned features $F_{l,t}'$ and upsampling features, feature information at different timestamps can be made more consistent and accurate. + +Algorithm 1: Event Feature Maps Registration +1 Input: current period $t$ , current downsampling layer $l$ , unaligned event data $E_{t}$ , unaligned feature maps $F_{l,t}$ +2 Output: affine matrix $\psi_{l,t}$ , aligned event data $E_{t}^{\prime}$ , aligned feature maps $F_{l,t}^{\prime}$ +3 $\psi_{0,t} \gets STN(E_t)$ Predict the affine transformation matrix of the event frame. +4 Perform the spatial transformation formula on the event frame: $E_{t}^{\prime} \gets TF(E_{t}, \psi_{0,t})$ +5 $l = 1$ ; +6 while $l \leq L$ do +7 if $l = L$ then +8 $\begin{array}{c|c} F_{l,t} \gets \text{Downsampling}(E_t^{\prime}); // \text{Down sampling the Registered Event Frames}. \\ \text{end} \\ \text{else} \\ F_{l,t} \gets \text{Down sampling}(F_{l-1,t}^{\prime}); // \text{Down sampling the Registered Feature Frames on the } l - 1 \text{th Layer.} \\ \text{end} \\ t = 1; \\ \text{while } t \leq T \text{ do} \\ \text{Predict affine matrix: } \psi_{l,t} = STN(F_{l,t}); \\ \text{Spatial transform: } F_{l,t}^{\prime} \gets TF(F_{l,t}, \psi_{l,t}); \\ t \gets t + 1 \\ \text{end} \\ l \gets l + 1 \\ \text{end} \end{array}$ +9 end + +As shown in Algorithm 1, the affine transformation parameters predicted by STN include specific values of rotation, scaling, and translation, so the algorithm can solve data registration for other acquisition methods besides uniform linear motion. The original event data of each input timestamp is subjected to parameter prediction and affine transformation to achieve preliminary overall registration; performing the same operation on the extracted feature map can further refine the alignment. The first two columns of the affine parameter matrix include information about the scaling and rotation, and the last column corresponds to the horizontal and vertical translation, achieving a + +six-degree-of-freedom transformation [11]. By predicting this matrix and performing an affine transformation on the original event frame and its feature map, we can obtain the registered event frame, which will facilitate subsequent image reconstruction. + +# 4.3 Temporal-Spatial Attention SNN Filter Module + +Fig. 4 illustrates the occlusion event filtering module employs a combination of Temporal-Spatial Attention (TSA) [29] mechanism and spiking neural network (SNN) model to filter out discrete noise events caused by obstructions. The resolving power of the model is improved by extracting the characteristics of targets and occlusions and assigning weights, and the Leaky Integrate-and-Fire mechanism (LIF) [2], is utilized to retain target events that have always existed in adjacent event frames and filter out discontinuous noise events. + +![](images/b380d2bfadde4f573533f844d8b77a2415183f0f44b3e120f1fb99ddfaa2d2dd.jpg) + +![](images/df2b12b2c21a9d00222369b56d93f66f1ea5b13e1b7440c39f34d7fc6dc50eae.jpg) +Fig. 4: TSA-SNN Occlusion Event Filtering Module. The SNN module incorporates a Temporal-Spatial Attention Mechanism for skip connection. + +The temporal-spatial attention (TSA) mechanism is incorporated into the image-filtering reconstruction model to enhance the model's perception of spatial and temporal correlations in the image. The SNN employs three neural layers with a leaky integrate-and-fire (LIF) mechanism, while introducing TSA skip connections between these layers. + +The TSA mechanism enables the model to dynamically allocate attention weights to regions and timestamps that contain crucial information in the input feature map $X$ . Subsequently, the output of the 1-st spiking neural layer is sent to the feature map $X_{C}$ obtained by the convolution layer to calculate the attention weight $g_{t}(X_{C})$ and $g_{s}(X_{C})$ in the temporal and spatial dimensions respectively. Weighted feature map $X_{A}$ is then integrated with the feature map $X$ , which is then forwarded to the activation layer and connected to the second spiking neural layer. + +The leakage mechanism of Spiking Neural Networks (SNNs) exclusively responds to persistent signals, effectively filtering out noise events induced by obstructions. The multi-dimensional attention mechanism can further extract the characteristic information of the target, calculate the attention weight, and multiply the feature map from the time, channel, and spatial dimensions, and the obtained result is jump-connected to the final output layer. + +# 4.4 Perceptual Mask Connection Module + +The perceptual mask connection module filters the occlusion events in registered feature maps through a mask operation to extract the target contour information, and then masks features through the convolution layer and sums them with the registered feature maps that have passed through the convolution layer. This module simplifies the information flow path in the network by creating a shortcut to directly transfer features and accumulating the filtered results of each timestamp into the encoder of the reconstruction module. This design not only enhances the network's robustness towards input data and bolsters the representation of salient features but also expedites information transmission. This enables the model to prioritize capturing crucial information, thereby effectively ameliorating the accuracy and efficiency of image reconstruction. + +# 4.5 Loss Function + +The perceptual loss emphasizes semantic similarity, the smoothing loss attends to the visual appearance of the image, and the pixel loss ensures accuracy at the finest level of detail. + +The Perception loss [10] can be expressed as: + +$$ +\mathcal {L} _ {p e r} (\Phi , \hat {\Phi}) = \sum_ {k} \frac {\lambda_ {k}}{C _ {k} H _ {k} W _ {k}} \| \Phi_ {k} - \hat {\Phi} _ {k} \| _ {2} ^ {2}, \tag {7} +$$ + +where $\varPhi_{k}$ and $\hat{\varPhi}_k$ with a shape of $C_k\times H_k\times W_k$ represent the output feature map of the k-th convolution layer of model output $\varPhi$ and corresponding ground truth $\hat{\varPhi}$ . $\lambda_{k}$ denotes the weight of the k-th convolution layer. + +The pixel loss [26] can be expressed as: + +$$ +\mathcal {L} _ {p i x} (\Phi , \hat {\Phi}) = \frac {\| \Phi - \hat {\Phi} \| _ {1}}{C H W}, \tag {8} +$$ + +where $C\times H\times W$ represents the shape of $\varPhi$ and $\hat{\varPhi}$ + +The total variance loss [14] can be expressed as: + +$$ +\mathcal {L} _ {t v} (\Phi) = \sum_ {i, j} \left[ \left(\phi_ {i, j + 1} - \phi_ {i, j}\right) ^ {2} + \left(\phi_ {i + 1, j} - \phi_ {i j}\right) ^ {2} \right] ^ {\frac {1}{2}}, \tag {9} +$$ + +where $\phi_{i,j}$ represents the pixel of $\varPhi$ . + +The objective of employing these three losses in unison is to meticulously consider various levels and aspects of information, thereby guiding the training of the generative model more comprehensively. By amalgamating these losses, the generative model can be more thoroughly optimized to generate high-quality images across multiple dimensions. + +The weighted total loss function can be expressed as: + +$$ +\mathcal {L} (\Phi , \hat {\Phi}) = \beta_ {p e r} L _ {p e r} (\Phi , \hat {\Phi}) + \beta_ {p i x} L _ {p i x} (\Phi , \hat {\Phi}) + \beta_ {t v} L _ {t v} (\Phi). \tag {10} +$$ + +where, $\beta_{per},\beta_{pix},\beta_{tv}$ are the weights of these functions. + +# 5 Experiments + +# 5.1 Experimental Settings + +V-ESAI Dataset: V-ESAI dataset is formed by data enhancement based on the E-SAI dataset collected by the DAVIS346 camera through random affine transform operations such as rotation, scaling, and translation. It imitates the lens rotation, focal length change, and perspective change during the actual SAI acquisition process. This can improve the generalization ability of the model. The E-SAI dataset contains 488 groups of indoor scenes and 100 groups of outdoor scenes with occlusion E-SAI event data, each group of data is divided by 30 timestamps. V-ESAI dataset includes the original data and the above three random transformations for each E-SAI dataset sample group, expanded to 1,952 indoor groups and 400 outdoor groups, for a total of 2352 groups. + +Implementation Details: The training experiment adopts a batch size of 12 and Adam optimizer with the SGDR schedule [13] reset the learning rate every 64 epochs through NVIDIA GeForce RTX3090 GPUs. The training experiment consists of three stages. First, the event stream is input into the filter reconstruction model for 200 epochs pre-training. The initial learning rate is $5 \times 10^{-4}$ ; after the training is completed, the parameters of the second part of the model are fixed. The registration model was trained for 500 epochs, with an initial learning rate of $3 \times 10^{-4}$ ; finally, based on the trained registration model and reconstruction model, the perceptual mask connection module was introduced to fine-tune the model parameters, and the overall training for 300 epochs was performed. + +# 5.2 Results of De-occlusion + +# Qualitative Analysis + +As illustrated in Fig. 5, traditional image frame-based methods encounter challenges in scenarios with obstructions, as they struggle to extract target information and effectively mitigate obstructions. Consequently, leveraging the attributes of event cameras becomes imperative to capture event frames triggered during the mobile acquisition process, enabling alignment, filtering, and reconstruction of target and occlusion object information. The REDIR method we propose exhibits robust performance in occlusion removal and target image reconstruction. + +To demonstrate the superiority of our REDIR network, we present a comparative analysis of the visual outcomes generated by diverse reconstruction models on the E-SAI dataset in Fig. 5. The direct reconstruction method employed in the Hybrid [31] model solely possesses the capacity to filter occluders and accumulate misalignment in the target imaging scene, leading to artifacts, image blurring, and loss of detailed information. The reconstruction model incorporating the Refocus module experiences uneven local deformation and saturation distribution due to the discrepancy between the spatial transformation method and the actual imaging camera motion compensation technique. Nevertheless, the REDIR model effectively addresses these issues. By combining UNet with STN, our model can accurately predict the transformation parameters of the global image and local features at each timestamp, performing affine transformation to align + +![](images/04ebf8454c837d9704836b693da460f4432db3d4b1ed9be6699afbcd5d4d6650.jpg) +Fig. 5: Comparisons of E-SAI with different de-occlusion reconstruction methods. From left to right, we show four examples with the occluded scenarios, results of Hybrid [31], Refocus+Hybrid [30], and our approach REDIR respectively. Details are zoomed in for better view. + +them and thereby significantly reducing motion-induced disturbances. Consequently, the proposed REDIR network generates images with superior reconstruction performance. + +# Quantitative Analysis + +Table 1: De-occlusion reconstruction performance on F-SAI/E-SAI dataset. Discriminated based on the necessity of prior information. + +
MethodRequire Prior InformationIndoorOutdoorTotal
PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
FSAI+ACC [23]✓14.220.34840.295511.860.39250.282113.810.35610.2931
Refocus+Hybrid [30]✓29.550.80860.054620.390.69610.101327.940.78890.0628
FSAI+CNN [25]✗25.010.79330.124517.390.56160.134723.670.75270.1263
Hybrid [31]✗24.040.72410.139414.260.38670.301222.320.66490.1678
REDIR (Ours)✗27.280.76640.111018.530.61300.127525.740.73950.1139
+ +In Tab. 1, we comparatively assess the reconstruction performance of the proposed models. Based on the E-SAI dataset for experiments, we conduct quantitative evaluations via PSNR, SSIM, and LPIPS metrics. The Hybrid model demonstrates inadequate handling of defocused data. For E-SAI imaging tasks with occlusion, the Hybrid model [31] directly reconstructs event data without focusing, which results in severe distortion due to the dislocation of event points across timestamps. The performance index of our model is slightly lower compared to that of Refocus+Hybrid [30], however, achieving such reconstruction accuracy requires pre-training on the focused E-SAI dataset with prior information, which necessitates pretraining on the manual-refocused dataset. + +To validate the state-of-the-art performance of our proposed method on a prior information-free dataset, we employ the Refocus+Hybrid [30] method for pre-training + +and subsequent overall training. However, it is observed that this approach fails to effectively reconstruct the target image. During the training process, if there is no decrease in loss value, it indicates that the model lacks the processing capability for the E-SAI dataset without prior information. In contrast, our model successfully recovers the de-occluded target image without relying on any prior information. Our proposed REDIR method performs an affine transformation on each timestamp, effectively addressing the aforementioned issues and registering without being limited by motion paths and speeds. Consequently, it can effectively reconstruct the target under extremely dense occlusion conditions. Compared to the focused reconstruction model, the REDIR demonstrates superior performance with a PSNR increase of 3.42dB, SSIM increase of $7.46\%$ , and LPIPS lower than $5.39\%$ . This corroborates that the novel registration approach achieves data alignment, resulting in a significantly improved reconstruction quality compared to existing algorithms. The model can effectively restore the target scene without occlusion. + +# 5.3 Variable Motion Acquisition Reconstruction + +The capability of our model to register data with scaling, rotation, and translation transformations was demonstrated through verification experiments conducted on the proposed multi-variable event V-ESAI dataset. As illustrated in Fig. 6, a comparative analysis of the outcomes obtained from the three transformation techniques demonstrates that our model exhibits comparable proficiency in handling diverse image transformation tasks, achieving satisfactory performance across all scenarios. Furthermore, the reconstructed image yields visually pleasing results, with. + +Currently, existing event data de-occlusion reconstruction models, such as ESAI and Refocus+Hybrid [30], lack the capability to process samples with transformations like rotation, scaling, and non-uniform translation. The prevailing automatic focusing method Refocus+Hybrid [30] significantly relies on prior information and requires manual focusing followed by pretraining, imposing strict restrictions on the acquisition method. In contrast, our proposed approach enables frame-by-frame alignment of transformed samples without the need for manual focusing step, thereby achieving truly adaptive registration. Experimental analysis demonstrates that our model exhibits an excellent de-occlusion reconstruction effect on samples subjected to non-uniform linear translation, scaling, and rotation transformations. Virtually no front occlusions are observed while effectively restoring the target scene. + +# 5.4 Ablation Study + +As illustrated in Tab. 2, to assess the contribution of each module, we performed an ablation experiment. By progressively eliminating STN, PMCM, TSA, and other modules from the Hybrid or UNet-Hybrid model and comparing the reconstruction performance, it becomes evident that the model relies on each module. By incrementally incorporating modules, which contribute significantly to target alignment and feature fusion, the model has been gradually optimized. Finally, the accumulative effect of our contributions in comparison to the reconstruction-only baseline resulted in a substantial improvement of $9.3\mathrm{dB}$ in PSNR, a $17.51\%$ increase in SSIM, and a reduction of $12.02\%$ in LPIPS, thereby establishing a new state of the art for event de-occlusion reconstruction. + +![](images/9bb1207883e198f9d9ea6dc20308ebc9a804f3f8168f4914b4047d32a228d3fc.jpg) +Fig. 6: REDIR de-occlusion reconstruction results on V-ESAI dataset. The images from left to right correspond to the original image, the results of translation, scaling and rotation transformation respectively. Details are zoomed in for a better view. + +Table 2: Ablation study of various components of our method on E-SAI dataset (part). + +
ArchitectureSTNPMCMTSAPSNR↑SSIM↑LPIPS↓
Hybrid [31]XXX23.320.70910.1481
UNet+HybridXXX26.450.74380.1358
UNet+HybridX✓✓31.070.78020.0909
UNet+Hybrid✓X✓32.340.86090.0385
UNet+Hybrid✓✓X32.160.82320.0548
UNet+Hybrid✓✓✓32.610.88420.0279
+ +# 6 Conclusion + +In this paper, we present an end-to-end method REDIR that has the capability of reconstructing events captured by moving cameras in occluded scenarios, while achieving generalized registration across events with variable transformations. As a result, our proposed method REDIR expands the applicability range for acquiring images using event cameras beyond uniform linear motion with fixed focal length or orbit. We introduce a new dataset V-ESAI, specifically designed to evaluate reconstruction capability under various challenging conditions including focus changes, lens rotation, as well as non-uniform motion. Our model achieves a higher level of imaging accuracy without prior information compared to state-of-the-art approaches on E-SAI dataset, while also demonstrating superior reconstruction performance on our V-ESAI dataset when compared to actual target scenarios without occlusion in terms of structural similarity. + +# Acknowledgements + +This work was supported in part by Science and Technology Innovation (STI) 2030—Major Projects under Grant 2022ZD0208700, and National Natural Science Foundation of China under Grant 62376264. + +# References + +1. Balakrishnan, G., Zhao, A., Sabuncu, M.R., Guttag, J., Dalca, A.V.: Voxelmorph: a learning framework for deformable medical image registration. IEEE transactions on medical imaging 38(8), 1788-1800 (2019) 4 +2. Chowdhury, S.S., Lee, C., Roy, K.: Towards understanding the effect of leak in spiking neural networks. Neurocomputing 464, 83-94 (2021) 9 +3. Delbrück, T., Linares-Barranco, B., Culurciello, E., Posch, C.: Activity-driven, event-based vision sensors. In: Proceedings of 2010 IEEE International Symposium on Circuits and Systems. pp. 2426-2429. IEEE (2010) 5 +4. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., et al.: Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence 44(1), 154-180 (2020) 1 +5. Gallego, G., Rebecq, H., Scaramuzza, D.: A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3867-3876 (2018) 4 +6. Gu, C., Learned-Miller, E., Sheldon, D., Gallego, G., Bideau, P.: The spatio-temporaloisson point process: A simple model for the alignment of event camera data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13495-13504 (2021) 5 +7. Gupta, A., Sharma, P., Ghosh, D., Honkote, V., Ghose, D.: Event-Based Time-To-Contact Estimation with Depth Image Fusion, p. 65-77 (Jan 2023). https://doi.org/10.1007/978-981-19-9379-4_6, http://dx.doi.org/10.1007/978-981-19-9379-4_6 8 +8. Huang, X., Zhang, Y., Xiong, Z.: Progressive spatio-temporal alignment for efficient event-based motion estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1537-1546 (2023) 5 +9. Jaderberg, M., Simonyan, K., Zisserman, A., et al.: Spatial transformer networks. Advances in neural information processing systems 28 (2015) 3, 6, 7 +0. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and superresolution. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 694-711. Springer (2016) 10 +1. Kim, H., Leutenegger, S., Davison, A.J.: Real-time 3d reconstruction and 6-dof tracking with an event camera. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VI 14. pp. 349-364. Springer (2016) 9 +2. Liu, D., Parra, A., Chin, T.J.: Spatiotemporal registration for event-based visual odometry. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (Jun 2021). https://doi.org/10.1109/cvpr46437.2021.00490, http://dx.doi.org/10.1109/cvpr46437.2021.00490 3 +3. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016) 11 +4. Mahendran, A., Vedaldi, A.: Understanding deep image representations by inverting them. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5188-5196 (2015) 10 + +15. Mitrokhin, A., Fermüller, C., Parameshwara, C., Aloimonos, Y.: Event-based moving object detection and tracking. In: 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1-9. IEEE (2018) 1 +16. Munda, G., Reinbacher, C., Pock, T.: Real-time intensity-image reconstruction for event cameras using manifold regularisation. International Journal of Computer Vision 126, 1381-1393 (2018) 1 +17. Neil, D., Pfeiffer, M., Liu, S.C.: Phased LSTM: Accelerating recurrent network training for long or event-based sequences. Cornell University - arXiv, Cornell University - arXiv (Oct 2016) 2 +18. Pan, L., Hartley, R., Scheerlinck, C., Liu, M., Yu, X., Dai, Y.: High frame rate video reconstruction based on an event camera. IEEE Transactions on Pattern Analysis and Machine Intelligence p. 1-1 (Jan 2020). https://doi.org/10.1109/tpami.2020.3036667, http://dx.doi.org/10.1109/tpami.2020.3036667 1 +19. Qi, C., Su, H., Mo, K., Guibas, L.: Pointnet: Deep learning on point sets for 3d classification and segmentation. arXiv 2016. arXiv preprint arXiv:1612.00593 4 +20. Rebecq, H., Gallego, G., Mueggler, E., Scaramuzza, D.: Emvs: Event-based multi-view stereo—3d reconstruction with an event camera in real-time. International Journal of Computer Vision p. 1394–1414 (Dec 2018). https://doi.org/10.1007/s11263-017-1050-6, http://dx.doi.org/10.1007/s11263-017-1050-6 3 +21. Scheerlinck, C., Rebecq, H., Gehrig, D., Barnes, N., Mahony, R.E., Scaramuzza, D.: Fast image reconstruction with an event camera. In: 2020 IEEE Winter Conference on Applications of Computer Vision (WACV) (Mar 2020). https://doi.org/10.1109/wacv45572.2020.9093366, http://dx.doi.org/10.1109/wacv45572.2020.9093366 +22. Stearns, C.C., Kannappan, K.: Method for 2-d affine transformation of images (1995) 6 +23. Vaish, V., Wilburn, B., Joshi, N., Levoy, M.: Using plane+ parallax for calibrating dense camera arrays. In: Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004. vol. 1, pp. I-I. IEEE (2004) 12 +24. Wang, Y., Wang, L., Yang, J., An, W., Yu, J., Guo, Y.: Spatial-angular interaction for light field image super-resolution. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIII 16. pp. 290-308. Springer (2020) 4 +25. Wang, Y., Wu, T., Yang, J., Wang, L., An, W., Guo, Y.: Deoccnet: Learning to see through foreground occlusions in light fields. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 118-127 (2020) 4, 12 +26. Wang, Z., Chen, J., Hoi, S.C.: Deep learning for image super-resolution: A survey. IEEE transactions on pattern analysis and machine intelligence 43(10), 3365-3387 (2020) 10 +27. Wang, Z., Ng, Y., Goor, P., Mahony, R.: Event camera calibration of per-pixel biased contrast threshold. arXiv: Computer Vision and Pattern Recognition, arXiv: Computer Vision and Pattern Recognition (Dec 2020) 2 +28. Xu, J., Jiang, M., Yu, L., Yang, W., Wang, W.: Robust motion compensation for event cameras with smooth constraint. IEEE Transactions on Computational Imaging 6, 604-614 (2020) 2 +29. Yao, M., Zhao, G., Zhang, H., Hu, Y., Deng, L., Tian, Y., Xu, B., Li, G.: Attention spiking neural networks. IEEE transactions on pattern analysis and machine intelligence (2023) 3, 6, 9 +30. Yu, L., Zhang, X., Liao, W., Yang, W., Xia, G.S.: Learning to see through with events. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(7), 8660-8678 (2023). https://doi.org/10.1109/TPAMI.2022.3227448 2, 4, 12, 13 +31. Zhang, X., Liao, W., Yu, L., Yang, W., Xia, G.S.: Event-based synthetic aperture imaging with a hybrid network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14235-14244 (2021) 1, 2, 4, 6, 11, 12, 14 + +32. Zhao, S., Dong, Y., Chang, E.I., Xu, Y., et al.: Recursive cascaded networks for unsupervised medical image registration. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10600-10610 (2019) 4 +33. Zou, Y., Zheng, Y., Takatani, T., Fu, Y.: Learning to reconstruct high speed and high dynamic range videos from events. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2024-2033 (2021) 2 \ No newline at end of file diff --git a/2024/REDIR_ Refocus-free Event-based De-occlusion Image Reconstruction/images.zip b/2024/REDIR_ Refocus-free Event-based De-occlusion Image Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b2038a43ce53ecd551d354327c7be50bac2fe5e2 --- /dev/null +++ b/2024/REDIR_ Refocus-free Event-based De-occlusion Image Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e4c8ccbae7eb6805b74b295203be1af0746c8a6222c09a254e1a1b77046bdc7d +size 443072 diff --git a/2024/REDIR_ Refocus-free Event-based De-occlusion Image Reconstruction/layout.json b/2024/REDIR_ Refocus-free Event-based De-occlusion Image Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..371e976deb6d2c1f709a6c9882524916aca1c86d --- /dev/null +++ b/2024/REDIR_ Refocus-free Event-based De-occlusion Image Reconstruction/layout.json @@ -0,0 +1,8226 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 138, + 111, + 477, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 111, + 477, + 146 + ], + "spans": [ + { + "bbox": [ + 138, + 111, + 477, + 146 + ], + "type": "text", + "content": "REDIR: Refocus-free Event-based De-occlusion Image Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "spans": [ + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "type": "text", + "content": "Qi Guo" + }, + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "type": "text", + "content": ", Hailong Shi" + }, + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "type": "text", + "content": ", Huan Li" + }, + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "type": "text", + "content": ", Jinsheng Xiao" + }, + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "type": "text", + "content": ", and Xingyu Gao" + }, + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 154, + 190, + 459, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 190, + 459, + 201 + ], + "spans": [ + { + "bbox": [ + 154, + 190, + 459, + 201 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 154, + 190, + 459, + 201 + ], + "type": "text", + "content": " Institute of Microelectronics of the Chinese Academy of Sciences, Beijing, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 194, + 201, + 419, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 201, + 419, + 213 + ], + "spans": [ + { + "bbox": [ + 194, + 201, + 419, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 194, + 201, + 419, + 213 + ], + "type": "text", + "content": " University of Chinese Academy of Sciences, Beijing, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 242, + 212, + 370, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 212, + 370, + 223 + ], + "spans": [ + { + "bbox": [ + 242, + 212, + 370, + 223 + ], + "type": "text", + "content": "3 Wuhan University, Hubei, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 217, + 224, + 395, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 224, + 395, + 234 + ], + "spans": [ + { + "bbox": [ + 217, + 224, + 395, + 234 + ], + "type": "text", + "content": "{shihailong, gaoxingyu}@ime.ac.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 217, + 235, + 397, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 235, + 397, + 244 + ], + "spans": [ + { + "bbox": [ + 217, + 235, + 397, + 244 + ], + "type": "text", + "content": "https://github.com/QiGuo-46/REDIR" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 274, + 455, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 274, + 455, + 472 + ], + "spans": [ + { + "bbox": [ + 159, + 274, + 455, + 472 + ], + "type": "text", + "content": "Abstract. The employment of the event-based synthetic aperture imaging (E-SAI) technique, which has the capability to capture high-frequency light intensity variations, has facilitated its extensive application on scene de-occlusion reconstruction tasks. However, existing methods usually require prior information and have strict restriction of camera motion on SAI acquisition methods. This paper proposes a novel end-to-end refocus-free variable E-SAI de-occlusion image reconstruction approach REDIR, which can align the global and local features of the variable event data and effectively achieve high-resolution imaging of pure event streams. To further improve the reconstruction of the occluded target, we propose a perceptual mask-gated connection module to interlink information between modules, and incorporate a spatial-temporal attention mechanism into the SNN block to enhance target extraction ability of the model. Through extensive experiments, our model achieves state-of-the-art reconstruction quality on the traditional E-SAI dataset without prior information, while verifying the effectiveness of the variable event data feature registration method on our newly introduced V-ESAI dataset, which obviates the reliance on prior knowledge and extends the applicability of SAI acquisition methods by incorporating focus changes, lens rotations, and non-uniform motion." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 483, + 455, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 483, + 455, + 506 + ], + "spans": [ + { + "bbox": [ + 159, + 483, + 455, + 506 + ], + "type": "text", + "content": "Keywords: Event-based Synthetic Aperture Imaging " + }, + { + "bbox": [ + 159, + 483, + 455, + 506 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 483, + 455, + 506 + ], + "type": "text", + "content": " De-occlusion Reconstruction " + }, + { + "bbox": [ + 159, + 483, + 455, + 506 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 483, + 455, + 506 + ], + "type": "text", + "content": " Data Registration" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 525, + 218, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 218, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 218, + 538 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 551, + 482, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 551, + 482, + 622 + ], + "spans": [ + { + "bbox": [ + 130, + 551, + 482, + 622 + ], + "type": "text", + "content": "The event camera [4] is a biomimetic sensing device that can perceive changes in the intensity of reflected light from objects with high temporal resolution [16]. For application scenarios with dense occlusions, event cameras can be employed to collect target information from different perspectives [18] within the gaps between occlusions, such as between trees, dense fences, and sight obstruction caused by rain and snow, which holds significant value for search and rescue missions and target tracking operations [15]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 623, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 623, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 623, + 482, + 647 + ], + "type": "text", + "content": "Event-based synthetic aperture imaging (E-SAI) [31] is a data acquisition method, as displayed in Fig. 1, which utilizes a linear uniform motion of camera for events and APS" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 132, + 654, + 217, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 654, + 217, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 654, + 217, + 665 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "frames sampling. Through E-SAI acquisition method, the camera can capture events triggered by targets and occlusion in motion [33]. The deep learning models [17] are capable of removing the obstruction of target scenarios on linear uniform motion E-SAI method. However, it is susceptible to lens shakes and motion errors [28], leading to misalignment of event points and subsequent degradation of imaging performance [27]." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 186, + 196, + 425, + 392 + ], + "blocks": [ + { + "bbox": [ + 186, + 196, + 425, + 392 + ], + "lines": [ + { + "bbox": [ + 186, + 196, + 425, + 392 + ], + "spans": [ + { + "bbox": [ + 186, + 196, + 425, + 392 + ], + "type": "image", + "image_path": "7a923e9eee67bf8fb23f858bb59f0241ab18a8f62ef1d180d25e044fcad41d6a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 402, + 482, + 436 + ], + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 436 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 436 + ], + "type": "text", + "content": "Fig. 1: E-SAI De-occlusion Image Reconstruction (without prior information). The event streams are captured through the uniform linear E-SAI acquisition method, the pictures below illustrate the results of de-occlusion imaging by existing method Hybrid [31] and ours." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "content": "Currently, the existing event point aligning methods [30, 31] mainly perform linear affine transformation on the event point by obtaining prior information such as the focal length, moving speed, and depth of field of the event camera, or obtain the affine transformation parameter matrix through model training and perform linear affine transformation. Therefore, the precise alignment of event points is unattainable due to the presence of errors caused by jitter during camera movement and alternative synthetic aperture imaging methods apart from uniform linear motion. Besides, obtaining event camera information in real-time and manually focusing will bring a huge workload. In real-world imaging scenarios, the aforementioned methods are facing two main challenges:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 565, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 141, + 565, + 482, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 565, + 482, + 638 + ], + "spans": [ + { + "bbox": [ + 141, + 565, + 482, + 638 + ], + "type": "text", + "content": "- Strict Restriction of SAI Acquisition Method: The reconstruction performance of imaging tasks relies on the acquisition method of event data. For imaging approaches beyond uniform linear motion, the reconstruction of the target presents substantial obstacles using existing techniques. The limited prediction correction parameters are employed for focusing tasks, while linear offset mapping is utilized to calibrate event data, resulting in inherent limitations of the imaging modality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 141, + 641, + 482, + 666 + ], + "type": "text", + "content": "- Requires Prior Information: Focusing methods in previous work often require information that relies on the acquisition process and requires manual labeling of" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Q. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 148, + 116, + 481, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 116, + 481, + 140 + ], + "spans": [ + { + "bbox": [ + 148, + 116, + 481, + 140 + ], + "type": "text", + "content": "samples, which results in increased time costs for the production of training sets and is not generalizable to other acquisition parameters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 259 + ], + "type": "text", + "content": "To address the above problems, we propose an end-to-end event registration and reconstruction model REDIR. The synthetic aperture imaging de-occlusion task of event cameras mainly includes event alignment based on a combination of global and local features and reconstruction of preserving target objects by filtering out occlusions. The purpose of alignment registration [20] is primarily to address the non-uniform linear motion error in real synthetic aperture imaging. The adaptive independent registration [12] of multiple event frames aims to address the variable SAI imaging acquisition method without prior information, which lays the foundation for de-occlusion reconstruction [21] of variable SAI imaging acquisition methods without prior information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 260, + 482, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 482, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 482, + 366 + ], + "type": "text", + "content": "The event registration module extracts global and local features from the event data and performs fusion alignment, which utilizes spatial transformer network (STN) [9] to predict the affine matrix and perform spatial transformation for each timestamp event frame, while its feature map is downsampled by UNet. The Temporal-Spatial Attention spiking neural network (SNN) [29] module is used to filter noise and extract target features, remove discrete occlusion event points and retain target event points. Finally, the perceptual mask gate module is used to connect multi-scale local features, which can effectively interchange the information between modules to reconstruct the scene without occlusion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 147, + 368, + 386, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 368, + 386, + 380 + ], + "spans": [ + { + "bbox": [ + 147, + 368, + 386, + 380 + ], + "type": "text", + "content": "The main contributions of the paper can be summarized as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 389, + 481, + 520 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 389, + 481, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 389, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 138, + 389, + 481, + 423 + ], + "type": "text", + "content": "- We propose a novel event data registration and alignment model, which can automatically map the target event point coordinates of multiple timestamps to the same position, effectively improving the event data reconstruction performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 426, + 481, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 426, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 138, + 426, + 481, + 471 + ], + "type": "text", + "content": "- We further propose an end-to-end reconstruction model REDIR that incorporates a perceptual mask-gated connection module and Temporal-Spatial Attention mechanism, which can reconstruct event data with occlusions into target scene images without visual obstruction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 474, + 480, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 474, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 138, + 474, + 480, + 520 + ], + "type": "text", + "content": "- We propose a new multi-variable transformed data augmentation method to compile Variable E-SAI (V-ESAI) dataset, which simulates non-uniform motion, lens rotation, and focal length changes in actual acquisition process, proving our proposed method relevantly persists reconstruction ability under other SAI acquisition methods." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 542, + 224, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 542, + 224, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 542, + 224, + 555 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Synthetic aperture imaging can capture the information of the target from different angles, capture more details in the imaging scene, and obtain a larger imaging range under the condition of limited depth of field and focal length. The difference between the event camera and the traditional sensor is that the light intensity change of the scene can be triggered by the moving DVS sensor, so the static image scene can be reconstructed. In addition, for the occluded imaging scene, the imaging method of synthetic aperture can obtain the information of the imaging plane of the target scene through the occlusion at different positions, and fuse and supplement the imaging data of multiple positions," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "REDIR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 140 + ], + "type": "text", + "content": "which can effectively remove the occlusion and restore the overall appearance of the target." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 145, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 145, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 145, + 482, + 407 + ], + "type": "text", + "content": "Then, some scholars applied it to the field of optics. In order to obtain complete information on objects through occluding objects, Wang [25] et al. proposed a method of removing foreground occluding based on a deep encoder-decoder network, DeOccNet, for object surface reconstruction in occluding SAI scenarios. In order to obtain multidimensional features and improve the performance of detail reconstruction, Wang [24] et al. improved the method by integrating spatial features and angular features for SAI reconstruction, and the performance reached SOTA. People have gradually found that the event camera has better performance in object reconstruction of occluded scenes. Because the event sensor has the ability to perceive the change of light intensity, it can effectively capture the information of the object in the occluded gap. Zhang [31] et al. used an event camera to collect the results of linear uniform synthetic aperture imaging of the target under extremely dense occlusion and extreme lighting conditions. The event frame registration was realized through manual focusing, and the target scene was restored through filtering and reconstruction by SNN and CNN hybrid models. In order to solve the problem of automatic registration of event data, Yu [30] et al. proposed a method to predict the warping parameter and move the coordinates of each timestamp by linear offset according to the parameter, so as to achieve refocusing. Generally, in the actual acquisition process, the sensor device cannot carry out uniform linear motion perfectly. Therefore, this event registration method based on predicted linear transformation parameters is not suitable for acquisition methods with large motion errors, changes in rotation, focal length and depth of field, and non-uniform linear motion." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 414, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 482, + 617 + ], + "type": "text", + "content": "Event-Based Image Registration is one of the necessary steps of E-SAI. The algorithms related to image registration and point cloud registration have been very mature, and are usually applied to the matching, stitching, or superposition of two images/point clouds. The goal of registration is to find the spatial mapping relationship between two images/point clouds, match the extracted features, and map and fuse them so that the two images/point clouds overlap to the same spatial position. For image fusion registration, Zhao [32] et al proposed a general recursive cascade network architecture that can realize deep cascade learning for deformable image registration. Through unsupervised training, the moving image is gradually deformed through each cascade and finally aligned with the fixed image. In order to combine global and local feature information for more refined registration, Balakrishna [1] et al. proposed a medical image registration method based on UNet architecture, which realized the alignment of MRI images of multiple coronal sections of the brain by predicting mapping transformation parameters. For the point cloud registration task, Qi [19] et al. proposed a PointNet, which extracts the global and local features of the point cloud through the T-Net transformation network, and realizes the registration of the point cloud by affine transformation through prediction transformation parameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 618, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 618, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 618, + 482, + 665 + ], + "type": "text", + "content": "For event data registration, it is usually necessary to match the coordinates of multiple timestamps to align the event to the same position. Gallego [5] et al demonstrated the ability to handle several important visual problems in event cameras (3D reconstruction, optical flow, and motion estimation), all without additional information about" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "Q. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 307 + ], + "type": "text", + "content": "the appearance of the scene: by maximizing the alignment of events on the trajectory of points on the image plane. Gu [6] et al. proposed a new probabilistic method for event alignment by modeling the events to be aligned as independent Poisson point processes or spatio-temporal Poisson point processes, so that every pixel is subject to this process. Based on this idea, a likelihood function is derived to estimate the camera motion that best explains the event. It is not limited to rotational motion, but can also be applied to other types of motion, such as translation. Huang [8] et al proposed a unified event-based motion estimation framework applicable to various motion models. By innovatively aligning events with the map, the rotation model and the 6-degree of freedom model are implemented on a standard CPU, and the motion estimation under a small number of sampled events is performed in an efficient manner. At the same time, a flexible batch size strategy is introduced to adapt to different scene textures, camera speed, and resolution. The comprehensive and detailed experimental results show that the framework reaches the current state of the art in terms of accuracy and efficiency. At present, there is no mature deep-learning method for event data registration without prior information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 308, + 482, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 357 + ], + "type": "text", + "content": "Inspired by the aforementioned approaches, we propose an event data registration method based on deep learning that enables fusion registration by extracting both global and local feature information from each time stamp. This method is not limited to uniform line SAI but can also be applied to other SAI acquisition methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 377, + 264, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 377, + 264, + 389 + ], + "spans": [ + { + "bbox": [ + 131, + 377, + 264, + 389 + ], + "type": "text", + "content": "3 Problem Formulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 405, + 230, + 416 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 405, + 230, + 416 + ], + "spans": [ + { + "bbox": [ + 131, + 405, + 230, + 416 + ], + "type": "text", + "content": "3.1 Event-Based SAI" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 428, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 428, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 428, + 482, + 464 + ], + "type": "text", + "content": "When the event camera captures changes in light intensity within the event frame, the actual static light intensity at the last timestamp " + }, + { + "bbox": [ + 130, + 428, + 482, + 464 + ], + "type": "inline_equation", + "content": "n - 1" + }, + { + "bbox": [ + 130, + 428, + 482, + 464 + ], + "type": "text", + "content": " and current timestamp " + }, + { + "bbox": [ + 130, + 428, + 482, + 464 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 428, + 482, + 464 + ], + "type": "text", + "content": " can be expressed as " + }, + { + "bbox": [ + 130, + 428, + 482, + 464 + ], + "type": "inline_equation", + "content": "\\Gamma (x_{n},t_{n} - \\Delta t_{n})" + }, + { + "bbox": [ + 130, + 428, + 482, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 428, + 482, + 464 + ], + "type": "inline_equation", + "content": "\\Gamma (x_{n},t_{n})" + }, + { + "bbox": [ + 130, + 428, + 482, + 464 + ], + "type": "text", + "content": " [3]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 224, + 475, + 482, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 475, + 482, + 488 + ], + "spans": [ + { + "bbox": [ + 224, + 475, + 482, + 488 + ], + "type": "interline_equation", + "content": "\\Gamma \\left(x _ {n}, t _ {n}\\right) - \\Gamma \\left(x _ {n}, t _ {n} - \\Delta t _ {n}\\right) \\geq p _ {n} \\cdot \\lambda , \\tag {1}", + "image_path": "c26444d312ae4bf853f882cba11899404f6329af3fe2593d4e42b9220f5ace11.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 556 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 130, + 498, + 482, + 556 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 498, + 482, + 556 + ], + "type": "text", + "content": " represents the threshold, " + }, + { + "bbox": [ + 130, + 498, + 482, + 556 + ], + "type": "inline_equation", + "content": "p_n" + }, + { + "bbox": [ + 130, + 498, + 482, + 556 + ], + "type": "text", + "content": " represents the polarity, and " + }, + { + "bbox": [ + 130, + 498, + 482, + 556 + ], + "type": "inline_equation", + "content": "\\Delta t_{n}" + }, + { + "bbox": [ + 130, + 498, + 482, + 556 + ], + "type": "text", + "content": " represents the time interval between the adjacent timestamps. The collected event data includes events caused by occlusion and target " + }, + { + "bbox": [ + 130, + 498, + 482, + 556 + ], + "type": "inline_equation", + "content": "E_{OA}(n)" + }, + { + "bbox": [ + 130, + 498, + 482, + 556 + ], + "type": "text", + "content": ", events caused by occlusion only " + }, + { + "bbox": [ + 130, + 498, + 482, + 556 + ], + "type": "inline_equation", + "content": "E_{OO}(n)" + }, + { + "bbox": [ + 130, + 498, + 482, + 556 + ], + "type": "text", + "content": ", target only " + }, + { + "bbox": [ + 130, + 498, + 482, + 556 + ], + "type": "inline_equation", + "content": "E_{AA}(n)" + }, + { + "bbox": [ + 130, + 498, + 482, + 556 + ], + "type": "text", + "content": ", as well as inherent noise of the camera " + }, + { + "bbox": [ + 130, + 498, + 482, + 556 + ], + "type": "inline_equation", + "content": "E_0" + }, + { + "bbox": [ + 130, + 498, + 482, + 556 + ], + "type": "text", + "content": ". Event data can be represented as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 211, + 571, + 481, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 571, + 481, + 585 + ], + "spans": [ + { + "bbox": [ + 211, + 571, + 481, + 585 + ], + "type": "interline_equation", + "content": "E (n) = E _ {O A} (n) + E _ {A A} (n) + E _ {O O} (n) + E _ {0}, \\tag {2}", + "image_path": "65bbb2774a30bd2a7b6f957b7efcee3ec7d198cef3ddec63d5f75fd57628b916.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 591, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 591, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 591, + 482, + 639 + ], + "type": "text", + "content": "The static light intensity information corresponding to both target and occlusion objects in an actual occlusion imaging scene can be represented as " + }, + { + "bbox": [ + 130, + 591, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\Gamma_A(n)" + }, + { + "bbox": [ + 130, + 591, + 482, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 591, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\Gamma_O(n)" + }, + { + "bbox": [ + 130, + 591, + 482, + 639 + ], + "type": "text", + "content": ". Therefore, the target event information captured by an event camera in an occlusion scene " + }, + { + "bbox": [ + 130, + 591, + 482, + 639 + ], + "type": "inline_equation", + "content": "E_{OA}(n)" + }, + { + "bbox": [ + 130, + 591, + 482, + 639 + ], + "type": "text", + "content": " can be described as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 241, + 653, + 481, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 653, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 241, + 653, + 481, + 666 + ], + "type": "interline_equation", + "content": "\\left| E _ {O A} (n) \\right| \\propto \\left| \\Gamma_ {A} (n) - \\Gamma_ {O} (n) \\right|. \\tag {3}", + "image_path": "92e5d4b52af0e8f515cb7d4bc1309a0b6b2114e77795e0aec49f2f56da6ff5ad.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "text", + "content": "REDIR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 237, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 237, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 237, + 128 + ], + "type": "text", + "content": "3.2 Event Registration" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 136, + 482, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 136, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 482, + 183 + ], + "type": "text", + "content": "Since the synthetic aperture imaging method necessitates camera movement to capture scene information from various viewing angles, accurate target registration is required. The mapping relationship between coordinates before and after the event can be formulated as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 242, + 191, + 482, + 216 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 191, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 242, + 191, + 482, + 216 + ], + "type": "interline_equation", + "content": "P _ {n} ^ {r e f} = K R _ {n} K ^ {- 1} P _ {n} + \\frac {K T _ {n}}{d}, \\tag {4}", + "image_path": "1b9b451b1f66ec367e9686ed16f861b3e50d1bfca7a7fe9c8b9e1253699e43a9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 219, + 480, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 219, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 130, + 219, + 480, + 267 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 219, + 480, + 267 + ], + "type": "inline_equation", + "content": "P_{n}^{ref}" + }, + { + "bbox": [ + 130, + 219, + 480, + 267 + ], + "type": "text", + "content": " represents the pixel position of the target on the reference imaging plane, " + }, + { + "bbox": [ + 130, + 219, + 480, + 267 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 219, + 480, + 267 + ], + "type": "text", + "content": " denotes the camera's internal matrix, " + }, + { + "bbox": [ + 130, + 219, + 480, + 267 + ], + "type": "inline_equation", + "content": "R_{n}" + }, + { + "bbox": [ + 130, + 219, + 480, + 267 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 219, + 480, + 267 + ], + "type": "inline_equation", + "content": "T_{n}" + }, + { + "bbox": [ + 130, + 219, + 480, + 267 + ], + "type": "text", + "content": " are rotation and translation matrices that describe the camera postures " + }, + { + "bbox": [ + 130, + 219, + 480, + 267 + ], + "type": "inline_equation", + "content": "\\theta_{n}" + }, + { + "bbox": [ + 130, + 219, + 480, + 267 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 219, + 480, + 267 + ], + "type": "inline_equation", + "content": "\\theta_{ref}" + }, + { + "bbox": [ + 130, + 219, + 480, + 267 + ], + "type": "text", + "content": " respectively, " + }, + { + "bbox": [ + 130, + 219, + 480, + 267 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 219, + 480, + 267 + ], + "type": "text", + "content": " represents the distance between target A and the camera plane." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 268, + 481, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 481, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 481, + 316 + ], + "type": "text", + "content": "The transformation formula [22] encompasses operations such as scaling " + }, + { + "bbox": [ + 130, + 268, + 481, + 316 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 268, + 481, + 316 + ], + "type": "text", + "content": ", rotation " + }, + { + "bbox": [ + 130, + 268, + 481, + 316 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 268, + 481, + 316 + ], + "type": "text", + "content": ", and translation " + }, + { + "bbox": [ + 130, + 268, + 481, + 316 + ], + "type": "inline_equation", + "content": "\\Delta x, \\Delta y" + }, + { + "bbox": [ + 130, + 268, + 481, + 316 + ], + "type": "text", + "content": ". Through registration, event points at different times can be aligned to a common position, enabling space-time dimension synthetic aperture imaging. The affine matrix " + }, + { + "bbox": [ + 130, + 268, + 481, + 316 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 130, + 268, + 481, + 316 + ], + "type": "text", + "content": " can be expressed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 243, + 323, + 481, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 323, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 243, + 323, + 481, + 352 + ], + "type": "interline_equation", + "content": "\\psi = \\left[ \\begin{array}{l l} \\cos \\theta / k & \\sin \\theta / k \\Delta x \\\\ - \\sin \\theta / k & \\cos \\theta / k \\Delta y \\end{array} \\right]. \\tag {5}", + "image_path": "fcdfdf68d960c4ae66610839d28578805c3009a1f63e23b35036581252a7ee4d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 365, + 250, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 250, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 250, + 376 + ], + "type": "text", + "content": "3.3 Event Reconstruction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 385, + 480, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 480, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 480, + 431 + ], + "type": "text", + "content": "The reconstruction of events is primarily accomplished by accumulating event frame feature maps, with varying timestamps of event points after aligning and occlusion filtering, and performing normalization operations based on the number of timestamps N." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 248, + 430, + 481, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 430, + 481, + 463 + ], + "spans": [ + { + "bbox": [ + 248, + 430, + 481, + 463 + ], + "type": "interline_equation", + "content": "F M _ {r e c o n} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} F M (n) \\tag {6}", + "image_path": "7e636cfb85df945809c0de63eaa160d3127e8886a99195b2f8f207f0d9ea7ae5.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 471, + 481, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 481, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 481, + 496 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 130, + 471, + 481, + 496 + ], + "type": "inline_equation", + "content": "FM_{recon}" + }, + { + "bbox": [ + 130, + 471, + 481, + 496 + ], + "type": "text", + "content": " represents the cumulative reconstructed feature map, and " + }, + { + "bbox": [ + 130, + 471, + 481, + 496 + ], + "type": "inline_equation", + "content": "FM(n)" + }, + { + "bbox": [ + 130, + 471, + 481, + 496 + ], + "type": "text", + "content": " represents the feature map of each timestamp " + }, + { + "bbox": [ + 130, + 471, + 481, + 496 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 471, + 481, + 496 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 514, + 220, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 220, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 220, + 528 + ], + "type": "text", + "content": "4 Methodology" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 538, + 251, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 538, + 251, + 550 + ], + "spans": [ + { + "bbox": [ + 132, + 538, + 251, + 550 + ], + "type": "text", + "content": "4.1 Network Architecture" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 557, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 641 + ], + "type": "text", + "content": "As illustrated in Fig. 2, the network architecture comprises an end-to-end model with a pure event stream as input data. Subsequently, E-SAI [31] de-occlusion imaging is conducted after the registration, filtering, and reconstruction modules. Specifically, the network consists of two UNet-like ANN models and a SNN [29] model, which are used to handle registration and reconstruction tasks respectively. These two models are connected in the perceptual mask module to effectively exchange information between the registration fusion feature and the reconstructed part." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "text", + "content": "In the registration module, the event stream data undergoes a series of processing, and the UNet with spatial transformer network (STN) [9] model is used for mapping" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Q. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 482, + 274 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 274 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 274 + ], + "type": "image", + "image_path": "25630602509a326147da88bb1b26a175409d9fe0c4818e4822358625f0e4e59e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 282, + 482, + 326 + ], + "lines": [ + { + "bbox": [ + 130, + 282, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 482, + 326 + ], + "type": "text", + "content": "Fig. 2: The Architecture of REDIR Network. The model incorporates the UNet registration module with STN, and the SNN filter module, which integrates a Temporal-Spatial Attention Mechanism. The registered target features are then transmitted to the CNN decoder module through the perceptual mask module." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 342, + 479, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 479, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 479, + 401 + ], + "type": "text", + "content": "and registration to ensure the spatial alignment of events. Then, the filtering module performs filtering operations on the registered data to enhance the expression of useful target information and reduce the impact of obstruction noise. Finally, through the reconstruction module, the SNN model is used to reconstruct the filtered data to achieve de-occlusion imaging of the event stream." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 403, + 480, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 480, + 450 + ], + "type": "text", + "content": "To optimize the integration of registration fusion features and the reconstruction component, a perceptual mask module is introduced. This module facilitates effective information fusion of the outputs of the registration module and the reconstruction module to produce the final imaging result." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 471, + 272, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 471, + 272, + 483 + ], + "spans": [ + { + "bbox": [ + 131, + 471, + 272, + 483 + ], + "type": "text", + "content": "4.2 Event Registration Module" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 494, + 480, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 480, + 543 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 480, + 543 + ], + "type": "text", + "content": "As displayed in Fig. 3, this module adopts a UNet-like architecture with STN [9] affine transformation module. The STN module enables adaptive alignment of images and features, effectively fusing aligned original data features with subsequent fusion registration features to enhance the accuracy of feature registration." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 151, + 564, + 458, + 641 + ], + "blocks": [ + { + "bbox": [ + 151, + 564, + 458, + 641 + ], + "lines": [ + { + "bbox": [ + 151, + 564, + 458, + 641 + ], + "spans": [ + { + "bbox": [ + 151, + 564, + 458, + 641 + ], + "type": "image", + "image_path": "e72797d61cd2f0a0a818b19a2379f92ce3d0eb941ab3222f5a425c1e716dd713.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 650, + 480, + 672 + ], + "lines": [ + { + "bbox": [ + 130, + 650, + 480, + 672 + ], + "spans": [ + { + "bbox": [ + 130, + 650, + 480, + 672 + ], + "type": "text", + "content": "Fig. 3: Event Registration Module. The STN is utilized to predict matrices and perform affine transformations in each layer of UNet." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "REDIR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 483, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 483, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 483, + 272 + ], + "type": "text", + "content": "The registration network module adopts global and local mapping alignment to improve the registration accuracy of event data. First, the module predicts the transformation matrix " + }, + { + "bbox": [ + 130, + 116, + 483, + 272 + ], + "type": "inline_equation", + "content": "\\psi_t" + }, + { + "bbox": [ + 130, + 116, + 483, + 272 + ], + "type": "text", + "content": " of event data for each timestamp " + }, + { + "bbox": [ + 130, + 116, + 483, + 272 + ], + "type": "inline_equation", + "content": "E_t" + }, + { + "bbox": [ + 130, + 116, + 483, + 272 + ], + "type": "text", + "content": ", and performs affine registration based on the prediction results. In this way, event data with different timestamps can be effectively aligned into a unified coordinate system, as " + }, + { + "bbox": [ + 130, + 116, + 483, + 272 + ], + "type": "inline_equation", + "content": "E_t'" + }, + { + "bbox": [ + 130, + 116, + 483, + 272 + ], + "type": "text", + "content": ". Subsequently, the registration network module utilizes multi-layer perceptron (MLP) to extract features. By performing the above operations on the feature map " + }, + { + "bbox": [ + 130, + 116, + 483, + 272 + ], + "type": "inline_equation", + "content": "F_{l,t}" + }, + { + "bbox": [ + 130, + 116, + 483, + 272 + ], + "type": "text", + "content": ", more abstract and meaningful feature information can be further extracted. In order to fuse the feature information of different timestamps [7], the registration network module also uses an upsampling method. Upsampling can convert low-resolution feature maps into high-resolution feature maps to better capture detailed information. By fusing aligned features " + }, + { + "bbox": [ + 130, + 116, + 483, + 272 + ], + "type": "inline_equation", + "content": "F_{l,t}'" + }, + { + "bbox": [ + 130, + 116, + 483, + 272 + ], + "type": "text", + "content": " and upsampling features, feature information at different timestamps can be made more consistent and accurate." + } + ] + } + ], + "index": 2 + }, + { + "type": "code", + "bbox": [ + 138, + 302, + 466, + 555 + ], + "blocks": [ + { + "bbox": [ + 139, + 288, + 331, + 300 + ], + "lines": [ + { + "bbox": [ + 139, + 288, + 331, + 300 + ], + "spans": [ + { + "bbox": [ + 139, + 288, + 331, + 300 + ], + "type": "text", + "content": "Algorithm 1: Event Feature Maps Registration" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 138, + 302, + 466, + 555 + ], + "lines": [ + { + "bbox": [ + 138, + 302, + 466, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 302, + 466, + 555 + ], + "type": "text", + "content": "1 Input: current period " + }, + { + "bbox": [ + 138, + 302, + 466, + 555 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 138, + 302, + 466, + 555 + ], + "type": "text", + "content": ", current downsampling layer " + }, + { + "bbox": [ + 138, + 302, + 466, + 555 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 138, + 302, + 466, + 555 + ], + "type": "text", + "content": ", unaligned event data " + }, + { + "bbox": [ + 138, + 302, + 466, + 555 + ], + "type": "inline_equation", + "content": "E_{t}" + }, + { + "bbox": [ + 138, + 302, + 466, + 555 + ], + "type": "text", + "content": ", unaligned feature maps " + }, + { + "bbox": [ + 138, + 302, + 466, + 555 + ], + "type": "inline_equation", + "content": "F_{l,t}" + }, + { + "bbox": [ + 138, + 302, + 466, + 555 + ], + "type": "text", + "content": " \n2 Output: affine matrix " + }, + { + "bbox": [ + 138, + 302, + 466, + 555 + ], + "type": "inline_equation", + "content": "\\psi_{l,t}" + }, + { + "bbox": [ + 138, + 302, + 466, + 555 + ], + "type": "text", + "content": ", aligned event data " + }, + { + "bbox": [ + 138, + 302, + 466, + 555 + ], + "type": "inline_equation", + "content": "E_{t}^{\\prime}" + }, + { + "bbox": [ + 138, + 302, + 466, + 555 + ], + "type": "text", + "content": ", aligned feature maps " + }, + { + "bbox": [ + 138, + 302, + 466, + 555 + ], + "type": "inline_equation", + "content": "F_{l,t}^{\\prime}" + }, + { + "bbox": [ + 138, + 302, + 466, + 555 + ], + "type": "text", + "content": " \n3 " + }, + { + "bbox": [ + 138, + 302, + 466, + 555 + ], + "type": "inline_equation", + "content": "\\psi_{0,t} \\gets STN(E_t)" + }, + { + "bbox": [ + 138, + 302, + 466, + 555 + ], + "type": "text", + "content": " Predict the affine transformation matrix of the event frame. \n4 Perform the spatial transformation formula on the event frame: " + }, + { + "bbox": [ + 138, + 302, + 466, + 555 + ], + "type": "inline_equation", + "content": "E_{t}^{\\prime} \\gets TF(E_{t}, \\psi_{0,t})" + }, + { + "bbox": [ + 138, + 302, + 466, + 555 + ], + "type": "text", + "content": " \n5 " + }, + { + "bbox": [ + 138, + 302, + 466, + 555 + ], + "type": "inline_equation", + "content": "l = 1" + }, + { + "bbox": [ + 138, + 302, + 466, + 555 + ], + "type": "text", + "content": "; \n6 while " + }, + { + "bbox": [ + 138, + 302, + 466, + 555 + ], + "type": "inline_equation", + "content": "l \\leq L" + }, + { + "bbox": [ + 138, + 302, + 466, + 555 + ], + "type": "text", + "content": " do \n7 if " + }, + { + "bbox": [ + 138, + 302, + 466, + 555 + ], + "type": "inline_equation", + "content": "l = L" + }, + { + "bbox": [ + 138, + 302, + 466, + 555 + ], + "type": "text", + "content": " then \n8 " + }, + { + "bbox": [ + 138, + 302, + 466, + 555 + ], + "type": "inline_equation", + "content": "\\begin{array}{c|c} F_{l,t} \\gets \\text{Downsampling}(E_t^{\\prime}); // \\text{Down sampling the Registered Event Frames}. \\\\ \\text{end} \\\\ \\text{else} \\\\ F_{l,t} \\gets \\text{Down sampling}(F_{l-1,t}^{\\prime}); // \\text{Down sampling the Registered Feature Frames on the } l - 1 \\text{th Layer.} \\\\ \\text{end} \\\\ t = 1; \\\\ \\text{while } t \\leq T \\text{ do} \\\\ \\text{Predict affine matrix: } \\psi_{l,t} = STN(F_{l,t}); \\\\ \\text{Spatial transform: } F_{l,t}^{\\prime} \\gets TF(F_{l,t}, \\psi_{l,t}); \\\\ t \\gets t + 1 \\\\ \\text{end} \\\\ l \\gets l + 1 \\\\ \\text{end} \\end{array}" + }, + { + "bbox": [ + 138, + 302, + 466, + 555 + ], + "type": "text", + "content": " \n9 end" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "code_body" + } + ], + "index": 4, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 570, + 483, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 483, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 483, + 668 + ], + "type": "text", + "content": "As shown in Algorithm 1, the affine transformation parameters predicted by STN include specific values of rotation, scaling, and translation, so the algorithm can solve data registration for other acquisition methods besides uniform linear motion. The original event data of each input timestamp is subjected to parameter prediction and affine transformation to achieve preliminary overall registration; performing the same operation on the extracted feature map can further refine the alignment. The first two columns of the affine parameter matrix include information about the scaling and rotation, and the last column corresponds to the horizontal and vertical translation, achieving a" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Q. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "text", + "content": "six-degree-of-freedom transformation [11]. By predicting this matrix and performing an affine transformation on the original event frame and its feature map, we can obtain the registered event frame, which will facilitate subsequent image reconstruction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 170, + 358, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 170, + 358, + 182 + ], + "spans": [ + { + "bbox": [ + 131, + 170, + 358, + 182 + ], + "type": "text", + "content": "4.3 Temporal-Spatial Attention SNN Filter Module" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 190, + 482, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 482, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 482, + 274 + ], + "type": "text", + "content": "Fig. 4 illustrates the occlusion event filtering module employs a combination of Temporal-Spatial Attention (TSA) [29] mechanism and spiking neural network (SNN) model to filter out discrete noise events caused by obstructions. The resolving power of the model is improved by extracting the characteristics of targets and occlusions and assigning weights, and the Leaky Integrate-and-Fire mechanism (LIF) [2], is utilized to retain target events that have always existed in adjacent event frames and filter out discontinuous noise events." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 201, + 282, + 411, + 354 + ], + "blocks": [ + { + "bbox": [ + 201, + 282, + 411, + 354 + ], + "lines": [ + { + "bbox": [ + 201, + 282, + 411, + 354 + ], + "spans": [ + { + "bbox": [ + 201, + 282, + 411, + 354 + ], + "type": "image", + "image_path": "b380d2bfadde4f573533f844d8b77a2415183f0f44b3e120f1fb99ddfaa2d2dd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 206, + 355, + 402, + 414 + ], + "blocks": [ + { + "bbox": [ + 206, + 355, + 402, + 414 + ], + "lines": [ + { + "bbox": [ + 206, + 355, + 402, + 414 + ], + "spans": [ + { + "bbox": [ + 206, + 355, + 402, + 414 + ], + "type": "image", + "image_path": "df2b12b2c21a9d00222369b56d93f66f1ea5b13e1b7440c39f34d7fc6dc50eae.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 424, + 483, + 447 + ], + "lines": [ + { + "bbox": [ + 131, + 424, + 483, + 447 + ], + "spans": [ + { + "bbox": [ + 131, + 424, + 483, + 447 + ], + "type": "text", + "content": "Fig. 4: TSA-SNN Occlusion Event Filtering Module. The SNN module incorporates a Temporal-Spatial Attention Mechanism for skip connection." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 462, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 510 + ], + "type": "text", + "content": "The temporal-spatial attention (TSA) mechanism is incorporated into the image-filtering reconstruction model to enhance the model's perception of spatial and temporal correlations in the image. The SNN employs three neural layers with a leaky integrate-and-fire (LIF) mechanism, while introducing TSA skip connections between these layers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "text", + "content": "The TSA mechanism enables the model to dynamically allocate attention weights to regions and timestamps that contain crucial information in the input feature map " + }, + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "text", + "content": ". Subsequently, the output of the 1-st spiking neural layer is sent to the feature map " + }, + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "inline_equation", + "content": "X_{C}" + }, + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "text", + "content": " obtained by the convolution layer to calculate the attention weight " + }, + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "inline_equation", + "content": "g_{t}(X_{C})" + }, + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "inline_equation", + "content": "g_{s}(X_{C})" + }, + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "text", + "content": " in the temporal and spatial dimensions respectively. Weighted feature map " + }, + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "inline_equation", + "content": "X_{A}" + }, + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "text", + "content": " is then integrated with the feature map " + }, + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "text", + "content": ", which is then forwarded to the activation layer and connected to the second spiking neural layer." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "The leakage mechanism of Spiking Neural Networks (SNNs) exclusively responds to persistent signals, effectively filtering out noise events induced by obstructions. The multi-dimensional attention mechanism can further extract the characteristic information of the target, calculate the attention weight, and multiply the feature map from the time, channel, and spatial dimensions, and the obtained result is jump-connected to the final output layer." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "text", + "content": "REDIR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 316, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 316, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 316, + 128 + ], + "type": "text", + "content": "4.4 Perceptual Mask Connection Module" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 135, + 482, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 135, + 482, + 255 + ], + "spans": [ + { + "bbox": [ + 130, + 135, + 482, + 255 + ], + "type": "text", + "content": "The perceptual mask connection module filters the occlusion events in registered feature maps through a mask operation to extract the target contour information, and then masks features through the convolution layer and sums them with the registered feature maps that have passed through the convolution layer. This module simplifies the information flow path in the network by creating a shortcut to directly transfer features and accumulating the filtered results of each timestamp into the encoder of the reconstruction module. This design not only enhances the network's robustness towards input data and bolsters the representation of salient features but also expedites information transmission. This enables the model to prioritize capturing crucial information, thereby effectively ameliorating the accuracy and efficiency of image reconstruction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 271, + 218, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 218, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 218, + 282 + ], + "type": "text", + "content": "4.5 Loss Function" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 291, + 482, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 291, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 291, + 482, + 326 + ], + "type": "text", + "content": "The perceptual loss emphasizes semantic similarity, the smoothing loss attends to the visual appearance of the image, and the pixel loss ensures accuracy at the finest level of detail." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 146, + 327, + 331, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 327, + 331, + 338 + ], + "spans": [ + { + "bbox": [ + 146, + 327, + 331, + 338 + ], + "type": "text", + "content": "The Perception loss [10] can be expressed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 220, + 346, + 481, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 346, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 220, + 346, + 481, + 373 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p e r} (\\Phi , \\hat {\\Phi}) = \\sum_ {k} \\frac {\\lambda_ {k}}{C _ {k} H _ {k} W _ {k}} \\| \\Phi_ {k} - \\hat {\\Phi} _ {k} \\| _ {2} ^ {2}, \\tag {7}", + "image_path": "5c72a77dfec7d884b3d557a0cfa360f4f726d2df949266276bec47c6f49bc1cc.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 383, + 481, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 383, + 481, + 419 + ], + "spans": [ + { + "bbox": [ + 131, + 383, + 481, + 419 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 383, + 481, + 419 + ], + "type": "inline_equation", + "content": "\\varPhi_{k}" + }, + { + "bbox": [ + 131, + 383, + 481, + 419 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 383, + 481, + 419 + ], + "type": "inline_equation", + "content": "\\hat{\\varPhi}_k" + }, + { + "bbox": [ + 131, + 383, + 481, + 419 + ], + "type": "text", + "content": " with a shape of " + }, + { + "bbox": [ + 131, + 383, + 481, + 419 + ], + "type": "inline_equation", + "content": "C_k\\times H_k\\times W_k" + }, + { + "bbox": [ + 131, + 383, + 481, + 419 + ], + "type": "text", + "content": " represent the output feature map of the k-th convolution layer of model output " + }, + { + "bbox": [ + 131, + 383, + 481, + 419 + ], + "type": "inline_equation", + "content": "\\varPhi" + }, + { + "bbox": [ + 131, + 383, + 481, + 419 + ], + "type": "text", + "content": " and corresponding ground truth " + }, + { + "bbox": [ + 131, + 383, + 481, + 419 + ], + "type": "inline_equation", + "content": "\\hat{\\varPhi}" + }, + { + "bbox": [ + 131, + 383, + 481, + 419 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 383, + 481, + 419 + ], + "type": "inline_equation", + "content": "\\lambda_{k}" + }, + { + "bbox": [ + 131, + 383, + 481, + 419 + ], + "type": "text", + "content": " denotes the weight of the k-th convolution layer." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 146, + 420, + 309, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 420, + 309, + 432 + ], + "spans": [ + { + "bbox": [ + 146, + 420, + 309, + 432 + ], + "type": "text", + "content": "The pixel loss [26] can be expressed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 253, + 440, + 481, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 440, + 481, + 466 + ], + "spans": [ + { + "bbox": [ + 253, + 440, + 481, + 466 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p i x} (\\Phi , \\hat {\\Phi}) = \\frac {\\| \\Phi - \\hat {\\Phi} \\| _ {1}}{C H W}, \\tag {8}", + "image_path": "aa43b1c90ede26afd67d8fe8e5df2b50261307af117e9e583c6a11b1a86cd0c9.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 473, + 345, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 473, + 345, + 486 + ], + "spans": [ + { + "bbox": [ + 131, + 473, + 345, + 486 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 473, + 345, + 486 + ], + "type": "inline_equation", + "content": "C\\times H\\times W" + }, + { + "bbox": [ + 131, + 473, + 345, + 486 + ], + "type": "text", + "content": " represents the shape of " + }, + { + "bbox": [ + 131, + 473, + 345, + 486 + ], + "type": "inline_equation", + "content": "\\varPhi" + }, + { + "bbox": [ + 131, + 473, + 345, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 473, + 345, + 486 + ], + "type": "inline_equation", + "content": "\\hat{\\varPhi}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 146, + 486, + 342, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 486, + 342, + 498 + ], + "spans": [ + { + "bbox": [ + 146, + 486, + 342, + 498 + ], + "type": "text", + "content": "The total variance loss [14] can be expressed as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 198, + 505, + 481, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 505, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 198, + 505, + 481, + 533 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t v} (\\Phi) = \\sum_ {i, j} \\left[ \\left(\\phi_ {i, j + 1} - \\phi_ {i, j}\\right) ^ {2} + \\left(\\phi_ {i + 1, j} - \\phi_ {i j}\\right) ^ {2} \\right] ^ {\\frac {1}{2}}, \\tag {9}", + "image_path": "f7f51dec9fe47b8b4f684172b7a7bf7f3141593e0a6f12216fe4a58e9d03cd1d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 540, + 279, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 540, + 279, + 552 + ], + "spans": [ + { + "bbox": [ + 131, + 540, + 279, + 552 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 540, + 279, + 552 + ], + "type": "inline_equation", + "content": "\\phi_{i,j}" + }, + { + "bbox": [ + 131, + 540, + 279, + 552 + ], + "type": "text", + "content": " represents the pixel of " + }, + { + "bbox": [ + 131, + 540, + 279, + 552 + ], + "type": "inline_equation", + "content": "\\varPhi" + }, + { + "bbox": [ + 131, + 540, + 279, + 552 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 552, + 481, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 481, + 611 + ], + "type": "text", + "content": "The objective of employing these three losses in unison is to meticulously consider various levels and aspects of information, thereby guiding the training of the generative model more comprehensively. By amalgamating these losses, the generative model can be more thoroughly optimized to generate high-quality images across multiple dimensions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 146, + 613, + 362, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 613, + 362, + 624 + ], + "spans": [ + { + "bbox": [ + 146, + 613, + 362, + 624 + ], + "type": "text", + "content": "The weighted total loss function can be expressed as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 185, + 632, + 481, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 632, + 481, + 646 + ], + "spans": [ + { + "bbox": [ + 185, + 632, + 481, + 646 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\Phi , \\hat {\\Phi}) = \\beta_ {p e r} L _ {p e r} (\\Phi , \\hat {\\Phi}) + \\beta_ {p i x} L _ {p i x} (\\Phi , \\hat {\\Phi}) + \\beta_ {t v} L _ {t v} (\\Phi). \\tag {10}", + "image_path": "dd65a434814d34888047edcc099a8d17069a6ab4fc0f1ad39c0b4009c992e544.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 131, + 653, + 362, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 653, + 362, + 667 + ], + "spans": [ + { + "bbox": [ + 131, + 653, + 362, + 667 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 131, + 653, + 362, + 667 + ], + "type": "inline_equation", + "content": "\\beta_{per},\\beta_{pix},\\beta_{tv}" + }, + { + "bbox": [ + 131, + 653, + 362, + 667 + ], + "type": "text", + "content": " are the weights of these functions." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "Q. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 218, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 218, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 218, + 129 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 251, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 251, + 153 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 251, + 153 + ], + "type": "text", + "content": "5.1 Experimental Settings" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 161, + 482, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 482, + 269 + ], + "type": "text", + "content": "V-ESAI Dataset: V-ESAI dataset is formed by data enhancement based on the E-SAI dataset collected by the DAVIS346 camera through random affine transform operations such as rotation, scaling, and translation. It imitates the lens rotation, focal length change, and perspective change during the actual SAI acquisition process. This can improve the generalization ability of the model. The E-SAI dataset contains 488 groups of indoor scenes and 100 groups of outdoor scenes with occlusion E-SAI event data, each group of data is divided by 30 timestamps. V-ESAI dataset includes the original data and the above three random transformations for each E-SAI dataset sample group, expanded to 1,952 indoor groups and 400 outdoor groups, for a total of 2352 groups." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 274, + 482, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 393 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 393 + ], + "type": "text", + "content": "Implementation Details: The training experiment adopts a batch size of 12 and Adam optimizer with the SGDR schedule [13] reset the learning rate every 64 epochs through NVIDIA GeForce RTX3090 GPUs. The training experiment consists of three stages. First, the event stream is input into the filter reconstruction model for 200 epochs pre-training. The initial learning rate is " + }, + { + "bbox": [ + 130, + 274, + 482, + 393 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 274, + 482, + 393 + ], + "type": "text", + "content": "; after the training is completed, the parameters of the second part of the model are fixed. The registration model was trained for 500 epochs, with an initial learning rate of " + }, + { + "bbox": [ + 130, + 274, + 482, + 393 + ], + "type": "inline_equation", + "content": "3 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 274, + 482, + 393 + ], + "type": "text", + "content": "; finally, based on the trained registration model and reconstruction model, the perceptual mask connection module was introduced to fine-tune the model parameters, and the overall training for 300 epochs was performed." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 412, + 256, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 412, + 256, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 412, + 256, + 423 + ], + "type": "text", + "content": "5.2 Results of De-occlusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 433, + 221, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 433, + 221, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 433, + 221, + 445 + ], + "type": "text", + "content": "Qualitative Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 450, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 533 + ], + "type": "text", + "content": "As illustrated in Fig. 5, traditional image frame-based methods encounter challenges in scenarios with obstructions, as they struggle to extract target information and effectively mitigate obstructions. Consequently, leveraging the attributes of event cameras becomes imperative to capture event frames triggered during the mobile acquisition process, enabling alignment, filtering, and reconstruction of target and occlusion object information. The REDIR method we propose exhibits robust performance in occlusion removal and target image reconstruction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "To demonstrate the superiority of our REDIR network, we present a comparative analysis of the visual outcomes generated by diverse reconstruction models on the E-SAI dataset in Fig. 5. The direct reconstruction method employed in the Hybrid [31] model solely possesses the capacity to filter occluders and accumulate misalignment in the target imaging scene, leading to artifacts, image blurring, and loss of detailed information. The reconstruction model incorporating the Refocus module experiences uneven local deformation and saturation distribution due to the discrepancy between the spatial transformation method and the actual imaging camera motion compensation technique. Nevertheless, the REDIR model effectively addresses these issues. By combining UNet with STN, our model can accurately predict the transformation parameters of the global image and local features at each timestamp, performing affine transformation to align" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "text", + "content": "REDIR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 477, + 293 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 477, + 293 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 477, + 293 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 477, + 293 + ], + "type": "image", + "image_path": "04ebf8454c837d9704836b693da460f4432db3d4b1ed9be6699afbcd5d4d6650.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 305, + 482, + 339 + ], + "lines": [ + { + "bbox": [ + 130, + 305, + 482, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 305, + 482, + 339 + ], + "type": "text", + "content": "Fig. 5: Comparisons of E-SAI with different de-occlusion reconstruction methods. From left to right, we show four examples with the occluded scenarios, results of Hybrid [31], Refocus+Hybrid [30], and our approach REDIR respectively. Details are zoomed in for better view." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 354, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 380 + ], + "type": "text", + "content": "them and thereby significantly reducing motion-induced disturbances. Consequently, the proposed REDIR network generates images with superior reconstruction performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 389, + 228, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 389, + 228, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 389, + 228, + 402 + ], + "type": "text", + "content": "Quantitative Analysis" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 134, + 445, + 478, + 519 + ], + "blocks": [ + { + "bbox": [ + 130, + 413, + 481, + 436 + ], + "lines": [ + { + "bbox": [ + 130, + 413, + 481, + 436 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 481, + 436 + ], + "type": "text", + "content": "Table 1: De-occlusion reconstruction performance on F-SAI/E-SAI dataset. Discriminated based on the necessity of prior information." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 445, + 478, + 519 + ], + "lines": [ + { + "bbox": [ + 134, + 445, + 478, + 519 + ], + "spans": [ + { + "bbox": [ + 134, + 445, + 478, + 519 + ], + "type": "table", + "html": "
MethodRequire Prior InformationIndoorOutdoorTotal
PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
FSAI+ACC [23]✓14.220.34840.295511.860.39250.282113.810.35610.2931
Refocus+Hybrid [30]✓29.550.80860.054620.390.69610.101327.940.78890.0628
FSAI+CNN [25]✗25.010.79330.124517.390.56160.134723.670.75270.1263
Hybrid [31]✗24.040.72410.139414.260.38670.301222.320.66490.1678
REDIR (Ours)✗27.280.76640.111018.530.61300.127525.740.73950.1139
", + "image_path": "646f83a20a7bb3e8c7f9ca6c57cdd0b74159c5dc03dbfe2f9a7fd6d31a1e86c3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 640 + ], + "type": "text", + "content": "In Tab. 1, we comparatively assess the reconstruction performance of the proposed models. Based on the E-SAI dataset for experiments, we conduct quantitative evaluations via PSNR, SSIM, and LPIPS metrics. The Hybrid model demonstrates inadequate handling of defocused data. For E-SAI imaging tasks with occlusion, the Hybrid model [31] directly reconstructs event data without focusing, which results in severe distortion due to the dislocation of event points across timestamps. The performance index of our model is slightly lower compared to that of Refocus+Hybrid [30], however, achieving such reconstruction accuracy requires pre-training on the focused E-SAI dataset with prior information, which necessitates pretraining on the manual-refocused dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 642, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 482, + 666 + ], + "type": "text", + "content": "To validate the state-of-the-art performance of our proposed method on a prior information-free dataset, we employ the Refocus+Hybrid [30] method for pre-training" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "Q. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": "and subsequent overall training. However, it is observed that this approach fails to effectively reconstruct the target image. During the training process, if there is no decrease in loss value, it indicates that the model lacks the processing capability for the E-SAI dataset without prior information. In contrast, our model successfully recovers the de-occluded target image without relying on any prior information. Our proposed REDIR method performs an affine transformation on each timestamp, effectively addressing the aforementioned issues and registering without being limited by motion paths and speeds. Consequently, it can effectively reconstruct the target under extremely dense occlusion conditions. Compared to the focused reconstruction model, the REDIR demonstrates superior performance with a PSNR increase of 3.42dB, SSIM increase of " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "7.46\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": ", and LPIPS lower than " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "5.39\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": ". This corroborates that the novel registration approach achieves data alignment, resulting in a significantly improved reconstruction quality compared to existing algorithms. The model can effectively restore the target scene without occlusion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 293, + 346, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 293, + 346, + 304 + ], + "spans": [ + { + "bbox": [ + 131, + 293, + 346, + 304 + ], + "type": "text", + "content": "5.3 Variable Motion Acquisition Reconstruction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 316, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 399 + ], + "type": "text", + "content": "The capability of our model to register data with scaling, rotation, and translation transformations was demonstrated through verification experiments conducted on the proposed multi-variable event V-ESAI dataset. As illustrated in Fig. 6, a comparative analysis of the outcomes obtained from the three transformation techniques demonstrates that our model exhibits comparable proficiency in handling diverse image transformation tasks, achieving satisfactory performance across all scenarios. Furthermore, the reconstructed image yields visually pleasing results, with." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 400, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 532 + ], + "type": "text", + "content": "Currently, existing event data de-occlusion reconstruction models, such as ESAI and Refocus+Hybrid [30], lack the capability to process samples with transformations like rotation, scaling, and non-uniform translation. The prevailing automatic focusing method Refocus+Hybrid [30] significantly relies on prior information and requires manual focusing followed by pretraining, imposing strict restrictions on the acquisition method. In contrast, our proposed approach enables frame-by-frame alignment of transformed samples without the need for manual focusing step, thereby achieving truly adaptive registration. Experimental analysis demonstrates that our model exhibits an excellent de-occlusion reconstruction effect on samples subjected to non-uniform linear translation, scaling, and rotation transformations. Virtually no front occlusions are observed while effectively restoring the target scene." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 552, + 222, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 552, + 222, + 565 + ], + "spans": [ + { + "bbox": [ + 131, + 552, + 222, + 565 + ], + "type": "text", + "content": "5.4 Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 575, + 482, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 575, + 482, + 684 + ], + "spans": [ + { + "bbox": [ + 130, + 575, + 482, + 684 + ], + "type": "text", + "content": "As illustrated in Tab. 2, to assess the contribution of each module, we performed an ablation experiment. By progressively eliminating STN, PMCM, TSA, and other modules from the Hybrid or UNet-Hybrid model and comparing the reconstruction performance, it becomes evident that the model relies on each module. By incrementally incorporating modules, which contribute significantly to target alignment and feature fusion, the model has been gradually optimized. Finally, the accumulative effect of our contributions in comparison to the reconstruction-only baseline resulted in a substantial improvement of " + }, + { + "bbox": [ + 130, + 575, + 482, + 684 + ], + "type": "inline_equation", + "content": "9.3\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 575, + 482, + 684 + ], + "type": "text", + "content": " in PSNR, a " + }, + { + "bbox": [ + 130, + 575, + 482, + 684 + ], + "type": "inline_equation", + "content": "17.51\\%" + }, + { + "bbox": [ + 130, + 575, + 482, + 684 + ], + "type": "text", + "content": " increase in SSIM, and a reduction of " + }, + { + "bbox": [ + 130, + 575, + 482, + 684 + ], + "type": "inline_equation", + "content": "12.02\\%" + }, + { + "bbox": [ + 130, + 575, + 482, + 684 + ], + "type": "text", + "content": " in LPIPS, thereby establishing a new state of the art for event de-occlusion reconstruction." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "REDIR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 198, + 114, + 415, + 339 + ], + "blocks": [ + { + "bbox": [ + 198, + 114, + 415, + 339 + ], + "lines": [ + { + "bbox": [ + 198, + 114, + 415, + 339 + ], + "spans": [ + { + "bbox": [ + 198, + 114, + 415, + 339 + ], + "type": "image", + "image_path": "9bb1207883e198f9d9ea6dc20308ebc9a804f3f8168f4914b4047d32a228d3fc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 350, + 482, + 384 + ], + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 384 + ], + "type": "text", + "content": "Fig. 6: REDIR de-occlusion reconstruction results on V-ESAI dataset. The images from left to right correspond to the original image, the results of translation, scaling and rotation transformation respectively. Details are zoomed in for a better view." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 198, + 407, + 413, + 487 + ], + "blocks": [ + { + "bbox": [ + 149, + 386, + 463, + 397 + ], + "lines": [ + { + "bbox": [ + 149, + 386, + 463, + 397 + ], + "spans": [ + { + "bbox": [ + 149, + 386, + 463, + 397 + ], + "type": "text", + "content": "Table 2: Ablation study of various components of our method on E-SAI dataset (part)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 198, + 407, + 413, + 487 + ], + "lines": [ + { + "bbox": [ + 198, + 407, + 413, + 487 + ], + "spans": [ + { + "bbox": [ + 198, + 407, + 413, + 487 + ], + "type": "table", + "html": "
ArchitectureSTNPMCMTSAPSNR↑SSIM↑LPIPS↓
Hybrid [31]XXX23.320.70910.1481
UNet+HybridXXX26.450.74380.1358
UNet+HybridX✓✓31.070.78020.0909
UNet+Hybrid✓X✓32.340.86090.0385
UNet+Hybrid✓✓X32.160.82320.0548
UNet+Hybrid✓✓✓32.610.88420.0279
", + "image_path": "7e3314e880eec0e9f2a0cfb05d13dceac7480d78d2652b03f60654277e6fb640.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 509, + 211, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 211, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 211, + 521 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "content": "In this paper, we present an end-to-end method REDIR that has the capability of reconstructing events captured by moving cameras in occluded scenarios, while achieving generalized registration across events with variable transformations. As a result, our proposed method REDIR expands the applicability range for acquiring images using event cameras beyond uniform linear motion with fixed focal length or orbit. We introduce a new dataset V-ESAI, specifically designed to evaluate reconstruction capability under various challenging conditions including focus changes, lens rotation, as well as non-uniform motion. Our model achieves a higher level of imaging accuracy without prior information compared to state-of-the-art approaches on E-SAI dataset, while also demonstrating superior reconstruction performance on our V-ESAI dataset when compared to actual target scenarios without occlusion in terms of structural similarity." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "Q. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 234, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 234, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 234, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 483, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 483, + 174 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 483, + 174 + ], + "type": "text", + "content": "This work was supported in part by Science and Technology Innovation (STI) 2030—Major Projects under Grant 2022ZD0208700, and National Natural Science Foundation of China under Grant 62376264." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 190, + 190, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 190, + 190, + 202 + ], + "spans": [ + { + "bbox": [ + 133, + 190, + 190, + 202 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 211, + 482, + 664 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 211, + 481, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 211, + 481, + 243 + ], + "spans": [ + { + "bbox": [ + 138, + 211, + 481, + 243 + ], + "type": "text", + "content": "1. Balakrishnan, G., Zhao, A., Sabuncu, M.R., Guttag, J., Dalca, A.V.: Voxelmorph: a learning framework for deformable medical image registration. IEEE transactions on medical imaging 38(8), 1788-1800 (2019) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 245, + 481, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 481, + 266 + ], + "type": "text", + "content": "2. Chowdhury, S.S., Lee, C., Roy, K.: Towards understanding the effect of leak in spiking neural networks. Neurocomputing 464, 83-94 (2021) 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 266, + 481, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 481, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 481, + 298 + ], + "type": "text", + "content": "3. Delbrück, T., Linares-Barranco, B., Culurciello, E., Posch, C.: Activity-driven, event-based vision sensors. In: Proceedings of 2010 IEEE International Symposium on Circuits and Systems. pp. 2426-2429. IEEE (2010) 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 300, + 481, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 300, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 481, + 331 + ], + "type": "text", + "content": "4. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., et al.: Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence 44(1), 154-180 (2020) 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 331, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 481, + 363 + ], + "type": "text", + "content": "5. Gallego, G., Rebecq, H., Scaramuzza, D.: A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3867-3876 (2018) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 364, + 481, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 364, + 481, + 395 + ], + "spans": [ + { + "bbox": [ + 138, + 364, + 481, + 395 + ], + "type": "text", + "content": "6. Gu, C., Learned-Miller, E., Sheldon, D., Gallego, G., Bideau, P.: The spatio-temporaloisson point process: A simple model for the alignment of event camera data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13495-13504 (2021) 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 396, + 482, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 482, + 437 + ], + "type": "text", + "content": "7. Gupta, A., Sharma, P., Ghosh, D., Honkote, V., Ghose, D.: Event-Based Time-To-Contact Estimation with Depth Image Fusion, p. 65-77 (Jan 2023). https://doi.org/10.1007/978-981-19-9379-4_6, http://dx.doi.org/10.1007/978-981-19-9379-4_6 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 438, + 481, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 438, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 138, + 438, + 481, + 471 + ], + "type": "text", + "content": "8. Huang, X., Zhang, Y., Xiong, Z.: Progressive spatio-temporal alignment for efficient event-based motion estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1537-1546 (2023) 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 472, + 481, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 472, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 481, + 491 + ], + "type": "text", + "content": "9. Jaderberg, M., Simonyan, K., Zisserman, A., et al.: Spatial transformer networks. Advances in neural information processing systems 28 (2015) 3, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 492, + 481, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 492, + 481, + 524 + ], + "spans": [ + { + "bbox": [ + 138, + 492, + 481, + 524 + ], + "type": "text", + "content": "0. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and superresolution. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 694-711. Springer (2016) 10" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 525, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 481, + 566 + ], + "type": "text", + "content": "1. Kim, H., Leutenegger, S., Davison, A.J.: Real-time 3d reconstruction and 6-dof tracking with an event camera. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VI 14. pp. 349-364. Springer (2016) 9" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 567, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 481, + 609 + ], + "type": "text", + "content": "2. Liu, D., Parra, A., Chin, T.J.: Spatiotemporal registration for event-based visual odometry. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (Jun 2021). https://doi.org/10.1109/cvpr46437.2021.00490, http://dx.doi.org/10.1109/cvpr46437.2021.00490 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "type": "text", + "content": "3. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016) 11" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 633, + 481, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 481, + 664 + ], + "type": "text", + "content": "4. Mahendran, A., Vedaldi, A.: Understanding deep image representations by inverting them. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5188-5196 (2015) 10" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "REDIR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "15. Mitrokhin, A., Fermüller, C., Parameshwara, C., Aloimonos, Y.: Event-based moving object detection and tracking. In: 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1-9. IEEE (2018) 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 482, + 182 + ], + "type": "text", + "content": "16. Munda, G., Reinbacher, C., Pock, T.: Real-time intensity-image reconstruction for event cameras using manifold regularisation. International Journal of Computer Vision 126, 1381-1393 (2018) 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "text", + "content": "17. Neil, D., Pfeiffer, M., Liu, S.C.: Phased LSTM: Accelerating recurrent network training for long or event-based sequences. Cornell University - arXiv, Cornell University - arXiv (Oct 2016) 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 205, + 482, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 205, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 134, + 205, + 482, + 247 + ], + "type": "text", + "content": "18. Pan, L., Hartley, R., Scheerlinck, C., Liu, M., Yu, X., Dai, Y.: High frame rate video reconstruction based on an event camera. IEEE Transactions on Pattern Analysis and Machine Intelligence p. 1-1 (Jan 2020). https://doi.org/10.1109/tpami.2020.3036667, http://dx.doi.org/10.1109/tpami.2020.3036667 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 249, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 249, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 249, + 481, + 270 + ], + "type": "text", + "content": "19. Qi, C., Su, H., Mo, K., Guibas, L.: Pointnet: Deep learning on point sets for 3d classification and segmentation. arXiv 2016. arXiv preprint arXiv:1612.00593 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 271, + 482, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 271, + 482, + 313 + ], + "spans": [ + { + "bbox": [ + 134, + 271, + 482, + 313 + ], + "type": "text", + "content": "20. Rebecq, H., Gallego, G., Mueggler, E., Scaramuzza, D.: Emvs: Event-based multi-view stereo—3d reconstruction with an event camera in real-time. International Journal of Computer Vision p. 1394–1414 (Dec 2018). https://doi.org/10.1007/s11263-017-1050-6, http://dx.doi.org/10.1007/s11263-017-1050-6 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 315, + 482, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 315, + 482, + 367 + ], + "spans": [ + { + "bbox": [ + 133, + 315, + 482, + 367 + ], + "type": "text", + "content": "21. Scheerlinck, C., Rebecq, H., Gehrig, D., Barnes, N., Mahony, R.E., Scaramuzza, D.: Fast image reconstruction with an event camera. In: 2020 IEEE Winter Conference on Applications of Computer Vision (WACV) (Mar 2020). https://doi.org/10.1109/wacv45572.2020.9093366, http://dx.doi.org/10.1109/wacv45572.2020.9093366" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 369, + 465, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 465, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 465, + 380 + ], + "type": "text", + "content": "22. Stearns, C.C., Kannappan, K.: Method for 2-d affine transformation of images (1995) 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 133, + 380, + 481, + 413 + ], + "type": "text", + "content": "23. Vaish, V., Wilburn, B., Joshi, N., Levoy, M.: Using plane+ parallax for calibrating dense camera arrays. In: Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004. vol. 1, pp. I-I. IEEE (2004) 12" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 414, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 414, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 133, + 414, + 481, + 456 + ], + "type": "text", + "content": "24. Wang, Y., Wang, L., Yang, J., An, W., Yu, J., Guo, Y.: Spatial-angular interaction for light field image super-resolution. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIII 16. pp. 290-308. Springer (2020) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 457, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 457, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 133, + 457, + 481, + 490 + ], + "type": "text", + "content": "25. Wang, Y., Wu, T., Yang, J., Wang, L., An, W., Guo, Y.: Deoccnet: Learning to see through foreground occlusions in light fields. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 118-127 (2020) 4, 12" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 491, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 491, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 133, + 491, + 481, + 511 + ], + "type": "text", + "content": "26. Wang, Z., Chen, J., Hoi, S.C.: Deep learning for image super-resolution: A survey. IEEE transactions on pattern analysis and machine intelligence 43(10), 3365-3387 (2020) 10" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 133, + 512, + 481, + 544 + ], + "type": "text", + "content": "27. Wang, Z., Ng, Y., Goor, P., Mahony, R.: Event camera calibration of per-pixel biased contrast threshold. arXiv: Computer Vision and Pattern Recognition, arXiv: Computer Vision and Pattern Recognition (Dec 2020) 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "type": "text", + "content": "28. Xu, J., Jiang, M., Yu, L., Yang, W., Wang, W.: Robust motion compensation for event cameras with smooth constraint. IEEE Transactions on Computational Imaging 6, 604-614 (2020) 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 567, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 567, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 133, + 567, + 481, + 597 + ], + "type": "text", + "content": "29. Yao, M., Zhao, G., Zhang, H., Hu, Y., Deng, L., Tian, Y., Xu, B., Li, G.: Attention spiking neural networks. IEEE transactions on pattern analysis and machine intelligence (2023) 3, 6, 9" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 599, + 481, + 632 + ], + "type": "text", + "content": "30. Yu, L., Zhang, X., Liao, W., Yang, W., Xia, G.S.: Learning to see through with events. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(7), 8660-8678 (2023). https://doi.org/10.1109/TPAMI.2022.3227448 2, 4, 12, 13" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 633, + 481, + 665 + ], + "type": "text", + "content": "31. Zhang, X., Liao, W., Yu, L., Yang, W., Xia, G.S.: Event-based synthetic aperture imaging with a hybrid network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14235-14244 (2021) 1, 2, 4, 6, 11, 12, 14" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "Q. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 183 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "32. Zhao, S., Dong, Y., Chang, E.I., Xu, Y., et al.: Recursive cascaded networks for unsupervised medical image registration. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10600-10610 (2019) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "type": "text", + "content": "33. Zou, Y., Zheng, Y., Takatani, T., Fu, Y.: Learning to reconstruct high speed and high dynamic range videos from events. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2024-2033 (2021) 2" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "REDIR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/REFRAME_ Reflective Surface Real-Time Rendering for Mobile Devices/f5175c1c-adca-4484-8e6b-cf9e011d3e14_content_list.json b/2024/REFRAME_ Reflective Surface Real-Time Rendering for Mobile Devices/f5175c1c-adca-4484-8e6b-cf9e011d3e14_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0dc3a71db43e83babdb44bbfd18ec219e22379ef --- /dev/null +++ b/2024/REFRAME_ Reflective Surface Real-Time Rendering for Mobile Devices/f5175c1c-adca-4484-8e6b-cf9e011d3e14_content_list.json @@ -0,0 +1,2344 @@ +[ + { + "type": "text", + "text": "REFRAME: Reflective Surface Real-Time Rendering for Mobile Devices", + "text_level": 1, + "bbox": [ + 142, + 48, + 852, + 105 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chaojie Ji $^{1}$ , Yufeng Li $^{2}$ , and Yiyi Liao $^{1\\boxtimes}$", + "bbox": [ + 278, + 141, + 716, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Zhejiang University", + "bbox": [ + 389, + 177, + 607, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 University of Electronic Science and Technology of China", + "bbox": [ + 205, + 196, + 789, + 214 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/86e8b0d4d5bb40342a08760789063c0a8c8a19f5232a693dc46646ce57aef0a4.jpg", + "image_caption": [ + "Fig. 1: REFRAME enables real-time rendering on consumer GPUs and mobile devices, delivering superior subjective quality with a low face number compared to the baselines [9, 45]. Additionally, it effectively decouples the appearance properties and environmental information of the scene, which helps its capabilities for downstream scene editing tasks." + ], + "image_footnote": [], + "bbox": [ + 79, + 256, + 915, + 497 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. This work tackles the challenging task of achieving real-time novel view synthesis for reflective surfaces across various scenes. Existing real-time rendering methods, especially those based on meshes, often have subpar performance in modeling surfaces with rich view-dependent appearances. Our key idea lies in leveraging meshes for rendering acceleration while incorporating a novel approach to parameterize view-dependent information. We decompose the color into diffuse and specular, and model the specular color in the reflected direction based on a neural environment map. Our experiments demonstrate that our method achieves comparable reconstruction quality for highly reflective surfaces compared to state-of-the-art offline methods, while also efficiently enabling real-time rendering on edge devices such as smartphones. Our project page is at https://xdimlab.github.io/REFrame/.", + "bbox": [ + 142, + 642, + 849, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Reflective surface $\\cdot$ Real-time rendering $\\cdot$ Mobile device", + "bbox": [ + 142, + 894, + 820, + 913 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\mathbb{O}$ Corresponding author.", + "bbox": [ + 99, + 928, + 346, + 947 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 79, + 53, + 307, + 74 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Novel view synthesis (NVS) generates realistic images from novel viewpoints using multiple input views. While Neural Radiance Fields (NeRF) [33] excel at high-quality NVS through volume rendering, they struggle with modeling reflective appearance and lack real-time rendering capabilities.", + "bbox": [ + 79, + 92, + 917, + 170 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several methods [19, 52, 59, 63] extend NeRF to decouple the intrinsic scene properties, e.g., into materials and lighting, and obtain color by the rendering equation [20]. Decoupling the environmental lighting and the physical parameters of the object often helps in modeling reflective objects. In contrast to recovering the precise physical meaning which may harm the visual quality due to the approximated rendering equation, another line of works [14, 46, 56] avoids full decomposition but enables representing reflective objects by modeling the reflective radiance. However, none of the aforementioned methods is capable of real-time rendering, especially on mobile devices, due to the expensive query of the rendering equation, or the underlying volume rendering formulation.", + "bbox": [ + 79, + 170, + 917, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There are many existing methods focusing on accelerating the rendering of NeRF. One line of works sticks with the volume rendering pipeline where the color of a pixel is accumulated along the ray. Acceleration is achieved by improving the sampling strategy [25, 27, 31, 48], tabulating the intermediate output [12, 16, 18, 34, 40, 49, 55, 58], or leveraging super-resolution neural rendering [28]. However, deploying these methods on edge devices such as smartphones is often greatly hindered due to the demanding computational power they typically require. Besides, tabulation-based methods require large memory consumption and greatly increase communication costs when transmitting data between the cloud server and the client. Another line of works [9, 39, 45] distills radiance fields into a mesh for real-time rendering, combined with a small MLP to model view-dependent effects. Mesh-based methods can leverage traditional graphics pipelines for acceleration, enabling them to achieve real-time rendering even on edge devices. Nonetheless, real-time rendering methods [9, 15, 16, 45, 58] often struggle to model objects with highly reflective surfaces. Besides, these mesh-based real-time rendering methods [9, 45, 57] typically require a large number of vertices and faces to achieve high fidelity.", + "bbox": [ + 79, + 365, + 917, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose REFRAME, a mesh-based REFlective surface ReAl-time rendering method for MobiE devices (e.g., smartphones), see Fig. 1. We find the fact that existing NeRF distilled mesh rendering pipelines [9,45] do not perform well in rendering objects with highly reflective appearances mainly for two reasons. Firstly, these methods utilize the viewing direction to model the view-dependent appearance, approved to be less effective than using the reflection direction [46]. As such, our method adopts the reflection direction-based parameterization. Nevertheless, this parameterization requires surface normal, yet estimating precise vertex normal based on a mesh representation can be challenging. Therefore, we propose a geometry learner that learns vertex and normal offsets through two networks. This leads to good normal estimations to calculate the reflection direction while maintaining relatively low vertex and face numbers. Secondly and more importantly, to alleviate inference burden, real-time", + "bbox": [ + 79, + 695, + 917, + 947 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 79, + 17, + 94, + 29 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "C.Ji et al.", + "bbox": [ + 159, + 14, + 256, + 30 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "rendering methods often have limited capacity to model complex view-dependent information. In order to enhance the expressive power of the model without increasing the inference computational cost, we employ a four-layer MLP during training to learn the environmental features. This information is then baked into a two-dimensional environment feature map during inference. Remarkably, the distilled environment feature map incurs a memory overhead of less than 1MB and can be edited for relighting purposes. Finally, despite our method achieving real-time rendering with an advantage in mesh faces and vertices number compared to existing works, the reconstruction quality of our method is comparable to the current state-of-the-art (including non-real-time methods) work and even outperforms them in rendering highly reflective objects. We summarize our contributions as follows:", + "bbox": [ + 72, + 56, + 922, + 289 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel mesh geometry learner, allowing for robustly optimizing mesh vertex positions and normals. This leads to high rendering quality with relatively low mesh vertex and face numbers.", + "- We propose to use an environment feature map to model view-dependent appearances of highly reflective objects, which enhances the capacity to reconstruct complex reflective appearances without increasing the inference burden. This further enables relighting effects.", + "- Our rendering quality is on par with the current state-of-the-art methods while being able to achieve real-time rendering across various platform devices. Moreover, our method even surpasses the current non-real-time state-of-the-art approaches in rendering objects with highly reflective appearances." + ], + "bbox": [ + 92, + 299, + 917, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 75, + 539, + 331, + 562 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "NeRF-based Scene Representation: NeRF [33] and its derivative works [1-3,41,44,46,53] have employed ray-marching based volume rendering methods to achieve high-quality and realistic rendering of different types of objects in various environments. Many follow-up works have extended NeRF in different aspects, including dynamic scene modeling [36-38], 3D-aware generation [7,8,42], and semantic scene understanding [11,24,65]. In this work, we focus on addressing two limitations of NeRF, i.e., modeling of reflective objects and real-time rendering.", + "bbox": [ + 72, + 578, + 920, + 713 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reflectance Decomposition: A number of works [4, 5, 19, 43, 52, 59-64] investigate the task of inferring geometry and material properties based on neural fields representation, typically by formulating the image generation using the physically-based rendering. Despite achieving lighting and material control, these methods are typically inferior to state-of-the-art NeRF-based methods that directly model the radiance in terms of the rendering quality. This is due to the fact that these methods rely on simplified rendering equations of the real world. Another line of works achieves better rendering quality by avoiding full decomposition, yet allowing for modeling glossy objects by modeling the reflected radiance [14, 46, 56] or replacing the explicit rendering equation with a learned neural renderer [6, 30]. However, none of these methods are capable of real-time rendering due to the underlying volume rendering formulation. While", + "bbox": [ + 72, + 715, + 920, + 949 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "REFRAME", + "bbox": [ + 719, + 14, + 837, + 30 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 903, + 16, + 917, + 30 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/306ca9e23f659e462f884c3b3bca90db97570b01176230887ebf42ad705e7cf9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 79, + 53, + 917, + 258 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/80d0f593ad28999ffd556503fcaba6b9cea1c4a94ca30863167afa4c08c6314a.jpg", + "image_caption": [ + "Fig. 2: Pipeline for REFRAME. Components with a yellow background are either baked or omitted during the mobile rendering stage. Training Stage: The initial mesh is updated first before performing differentiable rasterization [26]. Next, we obtain the diffuse color $\\mathbf{c}_d$ and specular feature $\\mathbf{f}_s$ based on the position $\\mathbf{p}$ , and the environment feature $\\mathbf{f}_e$ based on reflective direction $\\omega_r$ . Then we obtain the specular color $\\mathbf{c}_s$ and combine it with $\\mathbf{c}_d$ to create the final full color $\\mathbf{c}$ . Mobile Rendering Stage: We bake the intermediate output for real-time rendering. This mesh-based rendering can be implemented using WebGL and easily deployed on various platforms (e.g., desktop and mobile devices). Here, we retrieve the $\\mathbf{c}_d$ and $\\mathbf{f}_s$ from baked texture images, and the $\\mathbf{f}_e$ from the environment feature map. $\\mathbf{c}_d$ and $\\mathbf{c}_s$ are processed the same as the training stage to obtain $\\mathbf{c}$ ." + ], + "image_footnote": [], + "bbox": [ + 79, + 261, + 917, + 362 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "NvDiffRec [35] enables real-time rendering based on the mesh representation, its quality is also limited by the full decomposition. In this work, we follow the partial decomposition pipeline [46] and propose to learn a neural environment map based on a NeRF-distilled mesh representation, with a focus on achieving real-time rendering.", + "bbox": [ + 72, + 614, + 917, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "NeRF Acceleration: Accelerating NeRF rendering is another significant research area, with approaches reducing sampling points along the ray [25,31,48], tabulating the intermediate output [12, 16, 18, 49, 55, 58], using thousands of small MLPs to represent the scene [10, 40], or utilizing super-resolution techniques [28, 51]. However, these methods typically require consumer GPUs for real-time rendering, making it challenging to deploy them on edge devices with limited computational resources.", + "bbox": [ + 72, + 713, + 920, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "More recently, a few approaches [9, 45, 57] propose to distill a NeRF representation into a mesh for real-time rendering, where MobileNeRF [9] and NeRF2Mesh [45] enable real-time rendering on mobile devices. However, MobileNeRF [9] does not model the accurate geometry of the scene, leading to expensive memory costs with a large number of vertices and faces. NeRF2Mesh [45]", + "bbox": [ + 72, + 850, + 920, + 949 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 79, + 17, + 94, + 29 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "C.Ji et al.", + "bbox": [ + 159, + 14, + 259, + 30 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is the most relevant work to ours, which distills grid-based representation into a mesh. The good geometry prior for NeRF2Mesh [45] leads to a relatively low number of vertices and faces. However, none of the aforementioned methods are capable of modeling highly reflective objects faithfully due to their way of view-dependent color modeling. We propose to use a neural environment map to tackle this challenge by modeling the reflected radiance. Furthermore, we propose to estimate a high-quality vertex normal, further reducing the number of faces and vertices while maintaining the quality.", + "bbox": [ + 77, + 53, + 922, + 211 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 79, + 245, + 239, + 266 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The pipeline of our method is illustrated in Fig. 2. Before training, we employ a volumetric rendering technique to obtain an initial coarse mesh similar to existing methods [9,39,45,57]. During the training stage, we leverage a geometry learner to update both the vertices and the vertex normals (Sec. 3.1). We further learn a shader that decomposes diffuse and specular color (Sec. 3.2), where the specular branch is designed to be able to model highly reflective surfaces by combining view-independent specular features and reflective direction-conditioned environment features. After training with loss functions introduced in Sec. 3.3, we perform UV unwrapping and bake both view-independent and view-dependent features to enable real-time rendering on various devices (Sec. 3.4).", + "bbox": [ + 77, + 291, + 917, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "More formally, let $\\mathcal{M} = (\\mathcal{V},\\mathcal{F})$ denote the initial coarse mesh where $\\mathcal{V} = \\{\\mathbf{v}\\in \\mathbb{R}^3\\}$ is the set of vertices and $\\mathcal{F}$ is the set of faces. We further compute initial vertex normals $\\mathcal{N} = \\{\\mathbf{n}_{\\mathbf{v}}\\in \\mathbb{R}^{3}\\}$ given the initial mesh. Let $\\mathbf{p}\\in \\mathbb{R}^3$ denote a rendered surface point on the mesh, $\\mathbf{n_p}\\in \\mathbb{R}^3$ the corresponding normal, $\\omega_{o}\\in \\mathbb{S}^{2}$ the viewing direction, and $\\omega_{r}\\in \\mathbb{S}^{2}$ the reflective direction. During training, our goal is to refine the mesh $\\mathcal{M}$ , as well as map $(\\mathbf{p},\\boldsymbol {\\omega}_r)$ to a color $\\mathbf{c}\\in \\mathbb{R}^3$ to enable mesh-based real-time rendering.", + "bbox": [ + 77, + 487, + 917, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Geometry Learner", + "text_level": 1, + "bbox": [ + 79, + 656, + 365, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Before training, we follow NeRF2Mesh [45] to extract the initial coarse mesh through grid-based representation [34]. This allows for better modeling of depth discontinuities in geometry [50] and prevents the optimization of mesh geometry from getting stuck in local optima. Then, we propose to leverage a geometry learner to refine the geometry of the initial coarse mesh, including refining the vertex positions and normals.", + "bbox": [ + 77, + 692, + 917, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Vertex Offset: As shown in Fig. 3, the initial mesh extracted from NeRF-based methods may be of low quality. Existing methods [45,54] directly update the vertex positions through gradient backpropagation, which means the vertex update step size is determined by the learning rate and the gradient backpropagated. Thus, costly and difficult manual learning rate tuning is needed to adapt update step sizes for different objects, with a lack of robustness as shown in our experiments. To address this, our method uses a hybrid representation to provide a", + "bbox": [ + 77, + 811, + 917, + 947 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "REFRAME", + "bbox": [ + 719, + 14, + 837, + 30 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 903, + 16, + 917, + 30 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b7361106c718e803c2afd8e8371c4e14a56504137a5862d6c26bfdb793b12a71.jpg", + "image_caption": [ + "(a) Initial normal." + ], + "image_footnote": [], + "bbox": [ + 89, + 56, + 297, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fa9c7a556c3c0802ffed3d4c5d6b0dbd2708fc116317f226a5f120869eb38d98.jpg", + "image_caption": [ + "(b) Optimized normal.", + "Fig. 3: Mesh Optimization. The geometry of the initial mesh is often poor. We are able to significantly improve the geometry of the mesh through geometry learner." + ], + "image_footnote": [], + "bbox": [ + 317, + 58, + 520, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c104953e624039d4ecddada899efba62b798bcae50c5268d9cb037dba0217a43.jpg", + "image_caption": [ + "(c) Initial mesh." + ], + "image_footnote": [], + "bbox": [ + 537, + 52, + 716, + 165 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a057b44de21f36a9b2c7d4d076ac7072ff68d7a040bbd87bad86d02c7a12fba4.jpg", + "image_caption": [ + "(d) Optimized mesh." + ], + "image_footnote": [], + "bbox": [ + 728, + 52, + 910, + 165 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "learned, adaptive step size for different vertices and different objects:", + "bbox": [ + 72, + 278, + 811, + 297 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} ^ {\\prime} = \\mathbf {v} + \\Delta \\mathbf {v}, \\quad \\Delta \\mathbf {v} = g _ {\\theta_ {\\mathbf {v}}} (\\mathbf {v}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 315, + 917, + 338 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{v}'$ is the updated vertex, $g_{\\theta_{\\mathbf{v}}}$ is a multi-resolution hash encoding in combined with a small MLP [34]. Though [47] also uses MLP for vertex offset learning, our coarse-to-fine hybrid representation allows us to leverage local information as well as global information, leading to better modeling of the geometry.", + "bbox": [ + 72, + 352, + 920, + 430 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Normal Offset: Our method relies on surface normals to estimate the reflection direction $\\omega_r$ to model view-dependent appearance. Following the classical idea of smooth shading, we leverage smoothly interpolated vertex normals to approximate our surface normal. Obtaining accurate vertex normals is yet challenging, especially when the number of faces is limited. Hence, we learn a per-vertex normal offset by taking $\\mathbf{v}$ and $\\mathbf{n_v}$ as input:", + "bbox": [ + 72, + 432, + 917, + 549 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {n} _ {\\mathbf {v}} ^ {\\prime} = \\mathbf {n} _ {\\mathbf {v}} + \\Delta \\mathbf {n} _ {\\mathbf {v}}, \\quad \\Delta \\mathbf {n} _ {\\mathbf {v}} = g _ {\\theta_ {\\mathbf {n}}} (\\mathbf {v}, \\mathbf {n} _ {\\mathbf {v}}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 297, + 572, + 917, + 593 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{n}_{\\mathbf{v}}^{\\prime}$ is the learned per-vertex normal and $g_{\\theta_{\\mathbf{n}}}$ is another multi-resolution hash encoding-based network. In the experimental section, we demonstrate that learning such a per-vertex normal allows for modeling more accurate surface normals without increasing the number of vertices and faces. Note that $g_{\\theta_{\\mathbf{v}}}$ and $g_{\\theta_{\\mathbf{n}}}$ are only used during training, as we can directly save the updated vertices $\\mathbf{v}^{\\prime}$ and normals $\\mathbf{n}_{\\mathbf{v}}^{\\prime}$ for real-time rendering.", + "bbox": [ + 72, + 604, + 920, + 723 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Color Formulation", + "text_level": 1, + "bbox": [ + 75, + 756, + 368, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We decompose the final color $\\mathbf{c}$ into a diffuse color $\\mathbf{c}_d$ and a specular one $\\mathbf{c}_s$ :", + "bbox": [ + 75, + 793, + 888, + 813 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} = \\min \\left(\\max \\left(\\mathbf {c} _ {d} + \\mathbf {c} _ {s}, 0\\right), 1\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 348, + 832, + 917, + 852 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{c}_d + \\mathbf{c}_s$ is clamped to [0, 1], both individual terms $\\mathbf{c}_d$ and $\\mathbf{c}_s$ from sigmoid are in [0, 1]. We now elaborate on the diffuse and specular color formulation.", + "bbox": [ + 72, + 869, + 917, + 908 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Diffuse Color Formulation: With the updated mesh $\\mathcal{M}' = (\\mathcal{V}',\\mathcal{N}',\\mathcal{F})$ , we perform differentiable rasterization [26] to obtain the position $\\mathbf{p}$ corresponding", + "bbox": [ + 72, + 908, + 920, + 949 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 77, + 16, + 94, + 29 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "C.Ji et al.", + "bbox": [ + 159, + 14, + 259, + 30 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to each pixel. Subsequently, $\\mathbf{p}$ is mapped to a diffuse color $\\mathbf{c}_d\\in \\mathbb{R}^3$ and a view-independent feature $\\mathbf{f}_s\\in \\mathbb{R}^3$ which will be used later for decoding specular color:", + "bbox": [ + 75, + 55, + 917, + 95 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {d}, \\mathbf {f} _ {s} = f _ {\\phi} (\\mathbf {p}) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 115, + 917, + 134 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While the number of channels of $\\mathbf{f}_s$ is flexible, we observe that such a compact three-channel representation is sufficient for our purpose.", + "bbox": [ + 72, + 154, + 917, + 193 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specular Color Formulation: Recent non-real-time methods [14, 46, 56] propose to model view-dependent color based on the reflective direction $\\omega_{r}$ , instead of the viewing direction used by NeRF. This approach has been shown to effectively model reflective surfaces. Motivated by these methods, we calculate the reflective direction as follows:", + "bbox": [ + 72, + 195, + 917, + 292 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\omega} _ {r} = 2 \\left(\\boldsymbol {\\omega} _ {o} \\cdot \\mathbf {n} _ {\\mathbf {p}}\\right) \\mathbf {n} _ {\\mathbf {p}} - \\boldsymbol {\\omega} _ {o} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 315, + 917, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{n}_{\\mathbf{p}}$ is smoothly interpolated surface normal at $\\mathbf{p}$ based on our learned vertex normal $\\mathbf{n}_{\\mathbf{v}}^{\\prime}$ . Next, the specular color can be obtained as follows:", + "bbox": [ + 75, + 352, + 917, + 391 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {s} = f _ {\\psi} \\left(\\mathbf {f} _ {s}, \\boldsymbol {\\omega} _ {r}, \\left(\\boldsymbol {\\omega} _ {o} \\cdot \\mathbf {n} _ {\\mathbf {p}}\\right)\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 365, + 414, + 917, + 432 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $(\\boldsymbol{\\omega}_o \\cdot \\mathbf{n_p})$ is considered as input to model the Fresnel effects [21] and $f_{\\psi}$ needs to be a tiny MLP to enable real-time rendering.", + "bbox": [ + 72, + 450, + 917, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Environment Learner: In practice, we observe that naively mapping the reflective viewing direction to specular color (see Eq. (6)) leads to unsatisfying performance due to the capacity limitation of the tiny MLP $f_{\\psi}$ . For a similar reason, Ref-NeRF adopts an 8-layer MLP to map $\\mathbf{f}_s$ and $\\omega_r$ to the view-dependent color. However, increasing the capacity of $f_{\\psi}$ hampers the real-time rendering capability. Therefore, instead of increasing the capacity of $f_{\\psi}$ , we propose to leverage a neural environment learner of large capacity that maps $\\omega_r$ to an environment feature $\\mathbf{f}_e \\in \\mathbb{R}^M$ ( $M = 3$ in our paper):", + "bbox": [ + 72, + 492, + 917, + 650 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {e} = f _ {e} \\left(\\gamma \\left(\\boldsymbol {\\omega} _ {r}\\right)\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 421, + 678, + 917, + 695 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\gamma (\\cdot)$ denotes positional encoding and $f_{e}$ is a 4-layer MLP. This leads to our final implementation of the specular color:", + "bbox": [ + 72, + 710, + 917, + 751 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {s} = f _ {\\psi} \\left(\\mathbf {f} _ {s}, \\mathbf {f} _ {e}, \\left(\\boldsymbol {\\omega} _ {o} \\cdot \\mathbf {n} _ {\\mathbf {p}}\\right)\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 772, + 917, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that $\\mathbf{f}_e$ can be baked into a two-dimensional neural environment feature map for real-time rendering as it solely depends on $\\omega_r$ .", + "bbox": [ + 72, + 809, + 917, + 850 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scene Editing: REFRAME can decouple geometry, diffuse and specular color, so we can perform simple scene editing tasks. For example, we can edit the geometry and appearance of objects by modifying the mesh or texture image. Additionally, we can perform relighting on the objects by modifying the environment feature map of the scene, as shown in Fig. 4.", + "bbox": [ + 72, + 850, + 917, + 949 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "REFRAME", + "bbox": [ + 719, + 14, + 837, + 30 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 903, + 16, + 917, + 30 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cfe3c5e600f39b784c1622cf7f94a1dd20ce647c1468c766b90186b4a8cf59fb.jpg", + "image_caption": [ + "(a) Original image." + ], + "image_footnote": [], + "bbox": [ + 108, + 50, + 334, + 173 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/500ef3ad23a8817ec90a1806390a84d599c4541f992dc732bb37e699148f6d5f.jpg", + "image_caption": [ + "(b) Flip lighting.", + "Fig. 4: Relighting. We can achieve relighting of objects by editing our environment feature map. The image illustrates the editing process of flipping or replacing the environmental feature map. The left side shows the rendering result, the top right corner presents the corresponding environment feature map, and the specular color is displayed in the bottom right corner." + ], + "image_footnote": [], + "bbox": [ + 380, + 50, + 607, + 173 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/943a0becffb385cdcf6bb0c86af562dba7afd1e8c12602e24405de41cb1af613.jpg", + "image_caption": [ + "(c) Change lighting." + ], + "image_footnote": [], + "bbox": [ + 656, + 50, + 881, + 173 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Loss Functions", + "text_level": 1, + "bbox": [ + 75, + 333, + 324, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We train our model based on image reconstruction loss and several regularization terms. At each iteration, we render a full image and train with the corresponding ground truth. Let $\\hat{\\mathbf{c}}$ denote the ground truth RGB value at one pixel. We optionally use ground truth binary mask for supervision at synthetic object-centric scenes. Our full loss function is shown in Eq. (9):", + "bbox": [ + 72, + 362, + 920, + 461 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\lambda_ {\\mathbf {c}} \\mathcal {L} _ {\\mathbf {c}} + \\lambda_ {\\mathbf {c} _ {d}} \\mathcal {L} _ {\\mathbf {c} _ {d}} + \\lambda_ {S S I M} \\mathcal {L} _ {S S I M} + \\tag {9} \\\\ \\lambda_ {m a s k} \\mathcal {L} _ {m a s k} + \\lambda_ {\\Delta \\mathbf {n} _ {\\mathbf {v}}} \\mathcal {L} _ {\\Delta \\mathbf {n} _ {\\mathbf {v}}} + \\lambda_ {m a x} \\mathcal {L} _ {m a x} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 268, + 474, + 917, + 520 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Image Reconstruction Loss: Our model is first supervised by the L2 loss between the final rendered color and the ground truth.", + "bbox": [ + 72, + 531, + 917, + 572 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathbf {c}} = \\left| \\left| \\mathbf {c} - \\hat {\\mathbf {c}} \\right| \\right| _ {2} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 585, + 917, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To encourage the correct decoupling of diffuse and specular colors, we add a diffuse loss to guide the diffuse color to be close to the ground truth color:", + "bbox": [ + 75, + 619, + 917, + 658 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathbf {c} _ {d}} = \\left\\| \\mathbf {c} _ {d} - \\hat {\\mathbf {c}} \\right\\| _ {2} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 671, + 917, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also introduce SSIM [17] loss for better perceptual quality. On object-centric datasets, we additionally utilize mask loss to ensure the rasterization mask closely aligns with the ground truth mask. Note that the mask loss is optional and is not required on unbounded datasets.", + "bbox": [ + 72, + 705, + 917, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Regularizations: We regularize the predicted vertex normal $\\mathbf{n}_{\\mathbf{v}}^{\\prime}$ to be close to the original vertex normal $\\mathbf{n}_{\\mathbf{v}}$ computed from the mesh. This allows for stabilizing the optimization of the vertex normals, e.g., avoid completely flipping the normal. This is achieved by regularizing the magnitude of the normal offset:", + "bbox": [ + 72, + 783, + 917, + 861 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\Delta \\mathbf {n} _ {\\mathbf {v}}} = \\left\\| \\Delta \\mathbf {n} _ {\\mathbf {v}} \\right\\| _ {1} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 404, + 876, + 917, + 897 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce another regularization to encourage the specular color $\\mathbf{c}_s$ to be reasonable when visualized alone. As shown in Eq. (3), the final color $\\mathbf{c}$ will be", + "bbox": [ + 75, + 910, + 920, + 949 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 77, + 16, + 94, + 29 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "C.Ji et al.", + "bbox": [ + 159, + 14, + 259, + 30 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "clamped when the sum of specular color $\\mathbf{c}_s$ and diffuse color $\\mathbf{c}_d$ exceeds 1. In this case, the gradient is truncated due to the clamp operation, preventing the backpropagation from updating the color values. This may leave an undesired large specular color $\\mathbf{c}_s$ with no penalization. Therefore, we encourage the sum of specular color and diffuse color to not exceed 1:", + "bbox": [ + 72, + 56, + 920, + 152 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\max } = \\left| \\left| \\operatorname {m a x} \\left(\\mathbf {c} _ {d} + \\mathbf {c} _ {s} - 1, 0\\right) \\right| \\right| \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 167, + 915, + 185 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Real-Time Rendering", + "text_level": 1, + "bbox": [ + 75, + 211, + 409, + 232 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Environment Feature Map: During the training process, we employ a four-layer MLP with a width of 256 as our environment learner. If we query $\\mathbf{f}_e$ through the environment learner module during the inference process, it would significantly slow down the inference speed. Hence, during the inference phase, we bake the learned environment feature $\\mathbf{f}_e$ into a 2D environment feature map by converting $\\omega_r \\in \\mathbb{S}^2$ to its polar coordinate, see supplementary for more details. This allows us to simply query the corresponding feature on the environment feature map based on $\\omega_r$ , thereby greatly accelerating our inference speed, reaching above 200 FPS on a single NVIDIA 3090 GPU.", + "bbox": [ + 72, + 240, + 920, + 414 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Texture Images: To further reduce the inference burden and better deploy our method across different platforms, we can also bake the $\\mathbf{c}_d$ and $\\mathbf{f}_s$ obtained from querying the $f_{\\phi}$ into two texture images. Specifically, we map the vertices of the mesh to UV coordinates and then bake the features into two texture images, respectively. In order to obtain the normal $\\mathbf{n_p}$ from the learned normal vector $\\mathbf{n_v}'$ during real-time rendering, we also bake the normal $\\mathbf{n_p}$ into a texture image.", + "bbox": [ + 72, + 417, + 917, + 534 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 75, + 560, + 300, + 583 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details. Training: Prior to the training stage, we employ the classic quadric error metrics algorithm [13] to simplify the initial mesh to 75,000 faces except for the outdoor scenes which typically have complex geometry. We train for 250 epochs on each scene, with a training time of approximately 2 hours. We utilize the Adam optimizer [23] and employ the cosine annealing learning rate adjustment strategy [32] during the training stage. Rendering: The resolution of our baked texture image is $4096 \\times 4096$ , and the resolution of the baked environment feature map is $720 \\times 360$ . As referenced in Sec. 3.4, once we have baked the environment learner into the environment feature map, our approach is already capable of achieving above 200 FPS on the GPU (referred to as Ours). This implementation ensures a fair comparison with the baselines since their results quoted are evaluated using GPU implementation that can not directly deploy on mobile devices. Given that we use texture images for rendering when implemented on mobile devices, we also employ texture images on the GPU for quality evaluation (referred to as Ours (Mobile)). Following NeRF2Mesh [45] and MobileNeRF [9], we employ anti-aliasing to enhance our rendering quality. For details on the implementation of anti-aliasing, please refer to the supplementary material. All experiments are performed on a single NVIDIA 3090 GPU.", + "bbox": [ + 72, + 598, + 917, + 947 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "REFRAME", + "bbox": [ + 719, + 14, + 837, + 30 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 900, + 16, + 917, + 30 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/9ee59315cb66968c5b8ffa488b6a67ab6d8c44dd4f9ea6bfde23802508bce61b.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1: Rendering Quality. Baseline comparisons of the rendering quality on three different datasets. Red represents the optimal, orange represents the second best, and yellow represents the third." + ], + "table_body": "
NeRF Synthetic DatasetShiny Blender DatasetReal Captured Dataset
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NeRF [33]31.010.9470.081------
Ref-NeRF [46]33.990.9660.03835.960.9670.05824.450.6650.142
3DGS [22]33.300.9690.03030.370.9470.08324.060.6610.259
NvDiffRec [35]29.050.9390.08129.050.9380.111---
MobileNeRF [9]30.900.9470.06226.620.8830.16322.960.4930.430
NeRF2Mesh [45]29.670.9400.07226.960.8960.17022.710.5230.419
Ours31.040.9610.04036.020.9830.04022.500.5980.325
Ours (Mobile)30.840.9590.04235.830.9810.04122.110.5240.423
", + "bbox": [ + 87, + 53, + 917, + 229 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Dataset. We validate the effectiveness and robustness of our method using three different datasets with varied scenes. 1) NeRF [33] Synthetic Dataset: This dataset consists of eight synthetic scenes. 2) Shiny Blender Dataset: This dataset is introduced by Ref-NeRF [46] and includes six glossy objects, serving as the primary dataset for assessing the reflective surface modeling capabilities of the methods. 3) Real Captured Dataset: We select three outdoor scenes with rich reflective appearances same as Ref-NeRF [46] to validate the effectiveness of our method in real-world outdoor capture scenarios. Due to the poor quality of the initial mesh extracted by Nerf2Mesh [45] when modeling glossy objects, we utilize the meshes extracted by Neuralangelo [29] and Ref-NeuS [14] as our initialization in certain scenes, see supplementary for more details.", + "bbox": [ + 72, + 339, + 920, + 552 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baseline. We compare with various types of baselines, including Ref-NeRF [46], 3DGS [22], NvDiffRec [35], MobileNeRF [9], and NeRF2Mesh [45] in our study. Ref-NeRF [46] serves as the state-of-the-art volume rendering method for handling the strong reflective appearance. 3DGS [22] represents a powerful new representation capable of real-time rendering on GPU. NvDiffRec [35] represents physically based rendering methods. MobileNeRF [9], and NeRF2Mesh [45] are representative works in the field of real-time rendering for mobile devices. We import the results of the baselines from their original paper and replicate the results that missing in their original paper.", + "bbox": [ + 72, + 554, + 920, + 731 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Reconstruction Quality", + "text_level": 1, + "bbox": [ + 75, + 760, + 428, + 780 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Rendering Quality. We compare the rendering quality of our method and baseline methods on three datasets, see Tab. 1. It is worth noting that Ref-NeRF [46] is the state-of-the-art method for modeling glossy objects but cannot achieve real-time rendering, even on high-end GPU. We introduce Ref-NeRF [46] as a benchmark for rendering quality and demonstrate that our method achieves comparable or even superior rendering quality to non-real-time state-of-the-art methods. While 3DGS [22] can render in real-time on GPU and has higher PSNR on both NeRF Synthetic and Real Captured dataset than ours, it's important", + "bbox": [ + 72, + 791, + 920, + 949 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "C.Ji et al.", + "bbox": [ + 157, + 14, + 259, + 30 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/e0cf1ca3563fe8025eda30447c9cdf9f16799b4892de70dfd0f30c6adb46474b.jpg", + "image_caption": [ + "C", + "NvDiffRec" + ], + "image_footnote": [], + "bbox": [ + 135, + 63, + 263, + 136 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/3ece761960d140f3cfa8d73a0bd4b325a80098ac5873a14900bb4dcdf8e282cc.jpg", + "image_caption": [ + "Helmert" + ], + "image_footnote": [], + "bbox": [ + 113, + 139, + 205, + 204 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0909144c648b932e13735cbc07ccaccd15639a4c851cbee9a295b69e6f30f129.jpg", + "image_caption": [ + "Toaster", + "Fig. 5: Rendering Quality on Shiny Blender Dataset. Our method achieves optimal rendering quality in most scenes and provides better modeling of reflective appearance compared to the comparison methods [9,35,45]." + ], + "image_footnote": [], + "bbox": [ + 123, + 214, + 193, + 287 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0673bde51653b9fdc6a5d1beac2d9586278bb6729d3131a96343e34cdd65484a.jpg", + "image_caption": [ + "MobileNeRF" + ], + "image_footnote": [], + "bbox": [ + 297, + 63, + 426, + 136 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d25fdc1caeeb7c21d3a652c1adf243a58530933dac34c4bcf0d7f483c8c3aefb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 278, + 139, + 370, + 204 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f552b8380a342be290b9a561c31bc861cab62f06830d8159b233d1ea8d277deb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 288, + 214, + 358, + 287 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/8e6819ca04d5ea4f4a2b01b1ed1d22a424cbf3f13a285108a1c67dd4fdfde17e.jpg", + "image_caption": [ + "NeRF2Mesh" + ], + "image_footnote": [], + "bbox": [ + 467, + 63, + 590, + 136 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/aad6df0036970970f779dfd8e995f7aa01b8ff9f429a7ca59f84a57790051d85.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 440, + 139, + 532, + 206 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/de997674c608a635f5e20925a181b8d930ffad02b559d08eac8dde5c0c614ddb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 139, + 426, + 211 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/4c8588d3154606b6943e65d590971c47f48feedfcaf03e8eaa5f7b9ea35ab660.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 213, + 426, + 286 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7506e90e827bae975ed0e2a2636a205723833c4487a6a2b65e7645beb7d027f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 139, + 590, + 208 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e12e0bc85bb5acefa4083d97ccbb56b04b60708963238e366021520feb597572.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 213, + 590, + 286 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/54a21d313b747c0c8016573d54fd77360827ddde6286cc47a38fa3fe7598d5ed.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 627, + 63, + 682, + 130 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1ba0963fb7223b5653a2893729329263342398f342de0d927a70a5744322d986.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 605, + 147, + 697, + 204 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/fe60be153fd9497be5a011fbf710fa3479823f4beb553f8431de0616c9e325fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 615, + 214, + 685, + 287 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/a3a3f28d0a4ccef1a37a731ad61cf8c106f8e177c0fcbdc714e45cf4f4fe7873.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 63, + 755, + 136 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/bc36c9703fadb2381fb1bb0d7a1e4266c978b00e1af3cad0823386b2d3fe1dc2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 139, + 753, + 211 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1be4200de4ea6c17e870d6863f7880617d427e16e2c652b7ce1e4f91d8ffcdc6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 213, + 753, + 286 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/4fc146c66927e528bc3e876fb7431f467163179cf4c51d7a5b608a95525f493b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 147, + 861, + 206 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/cc4c784cf2f9a14fa0c87b80e2b7651af25b75bfc9486290a6edda5d04056d98.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 214, + 849, + 287 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/72105cb9fdaf593c430f38330d6f26912ec829f2c9a837c161f5af898488df09.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 864, + 65, + 917, + 211 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/a1383c552b532b3c43356e2bdc786b6ce4539fec7c167e1fbf2dfbba400d0400.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 864, + 213, + 917, + 286 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "to note that 3DGS struggles to model reflective surfaces effectively (Fig. 6a). Moreover, they have a larger memory overhead than us. $E.g.$ , for gardenspheres, the memory cost for 3DGS is 1.4GB, whereas our method only requires 84.3MB.", + "bbox": [ + 72, + 421, + 917, + 481 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In comparison to mesh-based real-time rendering methods [9, 35, 45], we achieve optimal rendering quality in most object-centric datasets, as shown in Fig. 5. Despite having mask supervision on object-centric datasets as well, NvDiffRec [35] and NeRF2Mesh [45] struggle to achieve high rendering quality. This could be due to the fact that NvDiffRec relies on the simplified rendering equation, whereas NeRF2Mesh's simple color formulation struggles to model highly reflective objects. Note that NeRF2Mesh also requires an initial mesh in its second stage, and we provide NeRF2Mesh with the same initial mesh as ours for all experiments for a fair comparison.", + "bbox": [ + 72, + 492, + 917, + 669 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In unbounded outdoor scenes without masks, our method still achieves high-quality rendering, as demonstrated in Fig. 6a. Although our PSNR may be slightly lower compared to the baselines, we outperform them in rendering foreground glossy objects. However, this advantage is not accurately reflected in the PSNR metric since the foreground reflective objects occupy only a small region in the image. Further, our method mainly exhibits artifacts in the background, this is due to the fact that the initial mesh can only faithfully reconstruct the foreground but not the background, see supplementary for more details.", + "bbox": [ + 72, + 682, + 917, + 839 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Surface Quality. Furthermore, we observe that our method not only improves the rendering quality when modeling glossy objects but also generates more accurate meshes, as shown in Fig. 6b. It can be seen that the meshes obtained from [35,45] often have poor quality in the reflective region while our method is capable of accurately reconstructing the geometry.", + "bbox": [ + 72, + 850, + 917, + 949 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "REFRAME", + "bbox": [ + 719, + 14, + 837, + 30 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 893, + 16, + 915, + 29 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/36f40a5b6f218bf1d38b78034539f0f0c398d13ca834fa2291a5fd91a3214f60.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 79, + 52, + 181, + 134 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/511a37a008c361b3463a77a4d8274a538ffbb663e1449712d42357d100e19e52.jpg", + "image_caption": [ + "(a) Rendering Quality on Real Captured Dataset." + ], + "image_footnote": [], + "bbox": [ + 79, + 139, + 181, + 222 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d45d5777842ddfa6612b8b778c0eb5e2af6f2d2e14eb86bf9b75a6179fb31831.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 188, + 53, + 290, + 134 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/9c1804bdc99372cd289fe61aa7eb59e794efc80bc410df0ce3d1f33bb151e881.jpg", + "image_caption": [ + "Fig. 6: Qualitative Comparison. Compared to the baselines, our method reconstructs reflective regions of higher fidelity and yields more accurate surface geometry." + ], + "image_footnote": [], + "bbox": [ + 188, + 139, + 290, + 222 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6044dffec518b45ba37cafcae18efeb5ddd1affdad15f7433d851340ad457730.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 53, + 399, + 134 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2e870c184d25a65cabd834a6c84235a932f94a5cf52043fa99727a5e41f94aa5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 139, + 399, + 222 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/18d7cbe5d651a4cf241340fc313b9e74d0849b3f6c0bbe68350ac691fd91207f.jpg", + "image_caption": [ + "(b) Surface Quality on NeRF Synthetic Dataset." + ], + "image_footnote": [], + "bbox": [ + 411, + 53, + 590, + 235 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a5b217db35b9b2a8ca247afcc6e0c815c2109a9813c6a9bb30964b91b62d1255.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 53, + 753, + 235 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8642dd161fa89993d20b0b3e40f8424b44fa14f621eea6d57be3e387662a8df4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 762, + 53, + 917, + 235 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/c9173e073e3a041a4ec867f872cb526527029b801562924b3e25816fb8d905de.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MacBook Pro iPhone12 iPad Air3 Legion Y-7000 HUAWEI nova7
MobileNeRF [9]120.0060.0060.0055.3139.19
NeRF2Mesh [45]120.0060.0060.0048.6932.56
Ours (Mobile)120.0060.0060.0048.5031.71
", + "bbox": [ + 84, + 321, + 912, + 404 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 2: Computational Efficiency Comparison. We test the FPS of our method and the baseline methods [9,45] on the NeRF Synthetic dataset across multiple devices. On MacBook Pro, iPhone 12, and iPad Air3, both our method and the baseline methods [9,45] achieve the maximum screen refresh rate.", + "bbox": [ + 72, + 404, + 917, + 476 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 Rendering Efficiency", + "text_level": 1, + "bbox": [ + 75, + 530, + 394, + 549 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Computational efficiency. During the mobile rendering stage, the bottleneck of our rendering speed lies in querying the tiny MLP $f_{\\psi}$ to obtain the specular color. However, due to our MLP being only two layers with 64 widths, our rendering speed remains relatively fast. Therefore, our method achieves real-time rendering on both consumer GPUs and mobile devices. We compare the rendering speed with the baselines [9,45] that can render in real-time on mobile devices across multiple platform devices, as seen in Tab. 2.", + "bbox": [ + 72, + 560, + 917, + 695 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Memory efficiency. Our cache mainly consists of texture images, the environment feature map, and the mesh. Due to our high-quality estimation of vertex normals, our mesh can achieve high-quality rendering with a low number of vertices and faces, leading to a small cache size in most scenes, as shown in Tab. 3.", + "bbox": [ + 72, + 695, + 917, + 773 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 75, + 819, + 331, + 840 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Geometry Learner. [54] and [45] utilize direct gradient backpropagation to update the vertex positions of the mesh. We employ this update strategy in our method and verify that this update method lacks robustness. Directly using the backpropagated gradients to update the vertex positions can lead to a devastating blow to the mesh, especially when using a large learning rate. This", + "bbox": [ + 72, + 850, + 917, + 949 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "C.Ji et al.", + "bbox": [ + 159, + 14, + 259, + 30 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/cc6dc89390584d9bc0807e600edc51de9a1dcdb5ebcdc12836b6e43570b22b17.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
NeRF Synthetic DatasetShiny Blender Dataset
#V (103)#F (103)Cache#V (103)#F (103)Cache
MobileNeRF [9]494224125.751028343152.37
NeRF2Mesh [45]20019273.53459122.15
Ours (Mobile)377547.55387552.86
", + "bbox": [ + 77, + 53, + 927, + 152 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/1cdc3eb44d80f5f294296d3792749725a8dba6536d918db0712c5ff46a64a839.jpg", + "table_caption": [ + "Table 3: Memory Efficiency. While exhibiting higher quality, our method maintains a relatively low memory consumption compared to baselines. Cache reported in MB." + ], + "table_footnote": [], + "table_body": "
w/o gθvOurs
L=0.001L=0.0001L=0.001
Toaster17.0117.0124.95
Materials23.0129.0629.58
", + "bbox": [ + 106, + 227, + 464, + 304 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/ebcc2be1bcd99212021c070d066485ff89517aeda249f8b183abc0d520bb5d40.jpg", + "table_caption": [ + "Table 4: Ablation of Vertex Offset Learning. L represents the learning rate. Reported in PSNR." + ], + "table_footnote": [], + "table_body": "
w/o Δn_vOurs
Toaster#F=10k24.2424.62
#F=75k24.5524.95
Materials#F=10k27.6928.23
#F=75k29.4629.58
", + "bbox": [ + 552, + 227, + 886, + 321 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 5: Ablation of Normal Offset Learning. Reported in PSNR.", + "bbox": [ + 510, + 323, + 917, + 359 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "significantly compromises the rendering quality, as shown in Tab. 4. Additionally, different datasets require different learning rates, as indicated in Tab. 4. While a learning rate of 0.0001 works fine for the materials dataset, it performs poorly on the toaster dataset. In contrast, our vertex update strategy not only exhibits robustness across different datasets but also provides higher rendering quality.", + "bbox": [ + 72, + 417, + 917, + 515 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In order to verify the necessity of estimating an offset for the vertices normal, we conduct experiments with the w/o normal offset setting. In this setting, we do not estimate a normal offset for each vertex, but we still compute an updated normal based on the updated vertex positions. As shown in Tab. 5, introducing per-vertex normal offset provides a gain in reconstruction quality. This gain is particularly significant when the number of faces is low. Notably, in the case of the toaster dataset, a mesh with $10\\mathrm{k}$ faces with normal offset even achieves rendering quality surpassing that of a mesh with $75\\mathrm{k}$ faces without normal offset. Environment Learner. To validate the effectiveness of the environment learner $f_{e}$ , we directly input $\\omega_{r}$ values into the tiny MLP $f_{\\psi}$ as shown in Eq. (6), similar to [9] and [45], except they use viewing direction instead of reflective direction. Experimental results demonstrate that directly inputting $\\omega_{r}$ into the tiny MLP $f_{\\psi}$ cannot learn the correct reflection appearance as a result of the limited network's capacity, as depicted in Fig. 7. Therefore, introducing the environment learner module improves the reconstruction quality without compromising the inference speed, as demonstrated in Tab. 6.", + "bbox": [ + 72, + 517, + 920, + 830 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Environment Feature Map: We can directly optimize an environment feature map instead of learning it through the environment learner module and baking it. However, as shown in Tab. 7, the direct optimization of the environment feature map performs poorly when the map resolution is high. This is because, at higher resolutions, neighboring directions are mapped to distant grid points, leading to a non-smooth interpolation. As a result, the grid points can-", + "bbox": [ + 72, + 830, + 920, + 949 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "REFRAME", + "bbox": [ + 719, + 14, + 837, + 30 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 893, + 16, + 917, + 30 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9d790918db044f75c38db9020eb8d8e1b448e8a4729afbb55dadedb6283348da.jpg", + "image_caption": [ + "(a) $\\mathrm{W / o}f_{e}$" + ], + "image_footnote": [], + "bbox": [ + 111, + 58, + 213, + 131 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/04f9ad22e7b71099b85a84668bd4296fbb264a3853a342505b9988461b02dfe0.jpg", + "image_caption": [ + "(b) Ours", + "Fig. 7: Ablation of Environment Learner. Eliminating $f_{e}$ results in bad rendering quality." + ], + "image_footnote": [], + "bbox": [ + 292, + 58, + 394, + 131 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/e6bbdcc3e5b8faa4aafb69e096b2951c29ebd5478be512171be14075874aacb9.jpg", + "image_caption": [ + "(c) Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 464, + 58, + 566, + 131 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/9e89722c77c776fe26b4aca86dddeee80bc22fe5789e5ad8ff8dab9dc459506d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
w/o feOurs
Toaster19.3124.95
Materials27.3729.58
", + "bbox": [ + 622, + 65, + 898, + 130 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/3bde7766e825ead4ffee6125ca3844665b163501904078060159e55222299b45.jpg", + "table_caption": [ + "Table 6: Ablation of Environment Learner. Reported in PSNR." + ], + "table_footnote": [], + "table_body": "
Resolutions H*WOursDirectly optimization
PSNRCache(KB)PSNRCache(KB)
Toaster180*36024.864924.1691
900*180024.9650423.592460
Materials180*36029.556329.58114
900*180029.5870629.033000
", + "bbox": [ + 108, + 227, + 886, + 344 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 7: Ablation of Environment Feature Map. Our method outperforms in both cache size and rendering quality.", + "bbox": [ + 75, + 346, + 917, + 383 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "not share global information and only contain its own local information. On the other hand, when the resolution is too low, the rendering quality is limited by the capacity of the environment feature map. In contrast, our method exhibits good rendering quality across different resolutions by learning the environment feature map via an MLP. Additionally, our environment feature maps enjoy the smoothness bias of the MLP, resulting in relatively smooth tensor maps. Consequently, when saving images in PNG format, our method demonstrates good compression performance. This further reduces the overhead of our caching.", + "bbox": [ + 72, + 432, + 917, + 588 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 75, + 614, + 285, + 634 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Although our method successfully models the appearance of glossy objects, similar to Ref-NeRF [46], it still struggles with accurately modeling interreflections and non-distant illumination. Additionally, our method requires a certain level of quality in the initial mesh.", + "bbox": [ + 72, + 652, + 917, + 728 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In summary, our method achieves real-time rendering on edge devices with low hardware budgets. Besides, we propose a novel approach for modeling view-dependent appearance and can optimize the appearance and geometry of glossy objects with high computational efficiency and low memory footprint. Furthermore, by effectively decoupling the scene's geometry, appearance, and environmental information, our method can perform simple scene editing tasks.", + "bbox": [ + 72, + 730, + 917, + 847 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 77, + 873, + 351, + 895 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work is supported by NSFC under grant 62202418 and U21B2004. We thank Xiaoran Cao, Chenxi Tu, and Sicheng Li for their valuable discussions.", + "bbox": [ + 72, + 908, + 917, + 947 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "C.Ji et al.", + "bbox": [ + 159, + 14, + 259, + 30 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 77, + 53, + 237, + 74 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)", + "2. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mipnerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022)", + "3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-nerf: Anti-aliased grid-based neural radiance fields. arXiv preprint arXiv:2304.06706 (2023)", + "4. Bi, S., Xu, Z., Srinivasan, P., Mildenhall, B., Sunkavalli, K., Hasan, M., Hold-Geoffroy, Y., Kriegman, D., Ramamoorthi, R.: Neural reflectance fields for appearance acquisition. arXiv preprint arXiv:2008.03824 (2020)", + "5. Boss, M., Braun, R., Jampani, V., Barron, J.T., Liu, C., Lensch, H.: Nerd: Neural reflectance decomposition from image collections. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12684-12694 (2021)", + "6. Boss, M., Jampani, V., Braun, R., Liu, C., Barron, J., Lensch, H.: Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems 34, 10691-10704 (2021)", + "7. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022)", + "8. Chan, E.R., Monteiro, M., Kellnhofer, P., Wu, J., Wetzstein, G.: pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5799-5809 (2021)", + "9. Chen, Z., Funkhouser, T., Hedman, P., Tagliasacchi, A.: Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16569-16578 (2023)", + "10. Esposito, S., Baieri, D., Zellmann, S., Hinkenjann, A., Rodola, E.: Kiloneus: A versatile neural implicit surface representation for real-time rendering. arXiv preprint arXiv:2206.10885 (2022)", + "11. Fu, X., Zhang, S., Chen, T., Lu, Y., Zhu, L., Zhou, X., Geiger, A., Liao, Y.: Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation. In: 2022 International Conference on 3D Vision (3DV). pp. 1-11. IEEE (2022)", + "12. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: Fastnerf: High-fidelity neural rendering at 200fps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14346-14355 (2021)", + "13. Garland, M., Heckbert, P.S.: Surface simplification using quadric error metrics. In: Proceedings of the 24th annual conference on Computer graphics and interactive techniques. pp. 209-216 (1997)", + "14. Ge, W., Hu, T., Zhao, H., Liu, S., Chen, Y.C.: Ref-neus: Ambiguity-reduced neural implicit surface learning for multi-view reconstruction with reflection. arXiv preprint arXiv:2303.10840 (2023)" + ], + "bbox": [ + 79, + 97, + 920, + 949 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "REFRAME", + "bbox": [ + 719, + 14, + 837, + 30 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 893, + 16, + 917, + 30 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Guo, Y.C., Cao, Y.P., Wang, C., He, Y., Shan, Y., Zhang, S.H.: Vmesh: Hybrid volume-mesh representation for efficient view synthesis. In: SIGGRAPH Asia 2023 Conference Papers. pp. 1-11 (2023)", + "16. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking neural radiance fields for real-time view synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5875-5884 (2021)", + "17. Hore, A., Ziou, D.: Image quality metrics: Psnr vs. sSIM. In: 2010 20th international conference on pattern recognition. pp. 2366-2369. IEEE (2010)", + "18. Hu, T., Liu, S., Chen, Y., Shen, T., Jia, J.: Efficientnerf efficient neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12902-12911 (2022)", + "19. Jin, H., Liu, I., Xu, P., Zhang, X., Han, S., Bi, S., Zhou, X., Xu, Z., Su, H.: Tensoroir: Tensorial inverse rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 165-174 (2023)", + "20. Kajiya, J.T.: The rendering equation. In: Proceedings of the 13th annual conference on Computer graphics and interactive techniques. pp. 143-150 (1986)", + "21. Kautz, J., McCool, M.D.: Approximation of glossy reflection with prefiltered environment maps. In: Graphics Interface. vol. 2000, pp. 119-126 (2000)", + "22. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)", + "23. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)", + "24. Kundu, A., Genova, K., Yin, X., Fathi, A., Pantofaru, C., Guibas, L.J., Tagliasacchi, A., Dellaert, F., Funkhouser, T.: Panoptic neural fields: A semantic object-aware neural scene representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12871-12881 (2022)", + "25. Kurz, A., Neff, T., Lv, Z., Zollhöfer, M., Steinberger, M.: Adanerf: Adaptive sampling for real-time rendering of neural radiance fields. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII. pp. 254-270. Springer (2022)", + "26. Laine, S., Hellsten, J., Karras, T., Seol, Y., Lehtinen, J., Aila, T.: Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG) 39(6), 1-14 (2020)", + "27. Li, C., Li, S., Zhao, Y., Zhu, W., Lin, Y.: Rt-nerf: Real-time on-device neural radiance fields towards immersive ar/vr rendering. In: Proceedings of the 41st IEEE/ACM International Conference on Computer-Aided Design. pp. 1-9 (2022)", + "28. Li, S., Li, H., Wang, Y., Liao, Y., Yu, L.: Steernerf: Accelerating nef rendering via smooth viewpoint trajectory. arXiv preprint arXiv:2212.08476 (2022)", + "29. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8456-8465 (2023)", + "30. Liang, R., Chen, H., Li, C., Chen, F., Panneer, S., Vijaykumar, N.: Envidr: Implicit differentiable renderer with neural environment lighting. arXiv preprint arXiv:2303.13022 (2023)", + "31. Lin, H., Peng, S., Xu, Z., Yan, Y., Shuai, Q., Bao, H., Zhou, X.: Efficient neural radiance fields for interactive free-viewpoint video. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)", + "32. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016)" + ], + "bbox": [ + 79, + 56, + 917, + 947 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "C.Ji et al.", + "bbox": [ + 157, + 14, + 259, + 30 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)", + "34. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)", + "35. Munkberg, J., Hasselgren, J., Shen, T., Gao, J., Chen, W., Evans, A., Müller, T., Fidler, S.: Extracting triangular 3d models, materials, and lighting from images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8280-8290 (2022)", + "36. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5865-5874 (2021)", + "37. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., MartinBrualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228 (2021)", + "38. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10318-10327 (2021)", + "39. Rakotosaona, M.J., Manhardt, F., Arroyo, D.M., Niemeyer, M., Kundu, A., Tombari, F.: Nerfmeshing: Distilling neural radiance fields into geometrically-accurate 3d meshes. arXiv preprint arXiv:2303.09431 (2023)", + "40. Reiser, C., Peng, S., Liao, Y., Geiger, A.: Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14335-14345 (2021)", + "41. Rematas, K., Liu, A., Srinivasan, P.P., Barron, J.T., Tagliasacchi, A., Funkhouser, T., Ferrari, V.: Urban radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12932-12942 (2022)", + "42. Schwarz, K., Liao, Y., Niemeyer, M., Geiger, A.: Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems 33, 20154-20166 (2020)", + "43. Srinivasan, P.P., Deng, B., Zhang, X., Tancik, M., Mildenhall, B., Barron, J.T.: Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7495-7504 (June 2021)", + "44. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8248-8258 (2022)", + "45. Tang, J., Zhou, H., Chen, X., Hu, T., Ding, E., Wang, J., Zeng, G.: Delicate textured mesh recovery from nef via adaptive surface refinement. arXiv preprint arXiv:2303.02091 (2023)", + "46. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5481-5490. IEEE (2022)", + "47. Walker, T., Mariotti, O., Vaxman, A., Bilen, H.: Explicit neural surfaces: Learning continuous geometry with deformation fields. arXiv preprint arXiv:2306.02956 (2023)" + ], + "bbox": [ + 77, + 56, + 917, + 947 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "REFRAME", + "bbox": [ + 719, + 14, + 837, + 30 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 893, + 16, + 917, + 30 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "48. Wang, H., Ren, J., Huang, Z., Olszewski, K., Chai, M., Fu, Y., Tulyakov, S.: R2l: Distilling neural radiance field to neural light field for efficient novel view synthesis. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXI. pp. 612-629. Springer (2022)", + "49. Wang, L., Zhang, J., Liu, X., Zhao, F., Zhang, Y., Zhang, Y., Wu, M., Yu, J., Xu, L.: Fourier plenoctrees for dynamic radiance field rendering in real-time. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13524-13534 (2022)", + "50. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689 (2021)", + "51. Wang, Z., Li, L., Shen, Z., Shen, L., Bo, L.: 4k-nerf: High fidelity neural radiance fields at ultra high resolutions. arXiv preprint arXiv:2212.04701 (2022)", + "52. Wang, Z., Shen, T., Gao, J., Huang, S., Munkberg, J., Hasselgren, J., Gojcic, Z., Chen, W., Fidler, S.: Neural fields meet explicit geometric representations for inverse rendering of urban scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8370-8380 (2023)", + "53. Weng, C.Y., Curless, B., Srinivasan, P.P., Barron, J.T., Kemelmacher-Shlizerman, I.: Humannerf: Free-viewpoint rendering of moving people from monocular video. In: Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition. pp. 16210-16220 (2022)", + "54. Worchel, M., Diaz, R., Hu, W., Schreer, O., Feldmann, I., Eisert, P.: Multiview mesh reconstruction with neural deferred shading. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6187-6197 (2022)", + "55. Wu, L., Lee, J.Y., Bhattad, A., Wang, Y.X., Forsyth, D.: Diver: Real-time and accurate neural radiance fields with deterministic integration for volume rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16200-16209 (2022)", + "56. Wu, T., Sun, J.M., Lai, Y.K., Gao, L.: De-nerf: Decoupled neural radiance fields for view-consistent appearance editing and high-frequency environmental relighting. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-11 (2023)", + "57. Yariv, L., Hedman, P., Reiser, C., Verbin, D., Srinivasan, P.P., Szeliski, R., Barron, J.T., Mildenhall, B.: Bakedsdf: Meshing neural sdfs for real-time view synthesis. arXiv preprint arXiv:2302.14859 (2023)", + "58. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: Plenoctrees for real-time rendering of neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5752-5761 (2021)", + "59. Zhang, J., Yang, G., Tulsiani, S., Ramanan, D.: Ners: Neural reflectance surfaces for sparse-view 3d reconstruction in the wild. Advances in Neural Information Processing Systems 34, 29835-29847 (2021)", + "60. Zhang, J., Yao, Y., Li, S., Liu, J., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neilf++: Inter-reflectable light fields for geometry and material estimation. arXiv preprint arXiv:2303.17147 (2023)", + "61. Zhang, K., Luan, F., Li, Z., Snavely, N.: Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5565-5574 (2022)", + "62. Zhang, K., Luan, F., Wang, Q., Bala, K., Snavely, N.: Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5453-5462 (2021)" + ], + "bbox": [ + 75, + 56, + 917, + 947 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "C.Ji et al.", + "bbox": [ + 157, + 14, + 259, + 30 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "63. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG) 40(6), 1-18 (2021)", + "64. Zhang, Y., Sun, J., He, X., Fu, H., Jia, R., Zhou, X.: Modeling indirect illumination for inverse rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18643-18652 (June 2022)", + "65. Zhi, S., Laidlow, T., Leutenegger, S., Davison, A.J.: In-place scene labelling and understanding with implicit scene representation. In: ICCV (2021)" + ], + "bbox": [ + 75, + 56, + 917, + 200 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "REFRAME", + "bbox": [ + 719, + 14, + 837, + 30 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 893, + 16, + 917, + 30 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/REFRAME_ Reflective Surface Real-Time Rendering for Mobile Devices/f5175c1c-adca-4484-8e6b-cf9e011d3e14_model.json b/2024/REFRAME_ Reflective Surface Real-Time Rendering for Mobile Devices/f5175c1c-adca-4484-8e6b-cf9e011d3e14_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e3ca5ff02c1a3afd095f0a9be82d18103c81d093 --- /dev/null +++ b/2024/REFRAME_ Reflective Surface Real-Time Rendering for Mobile Devices/f5175c1c-adca-4484-8e6b-cf9e011d3e14_model.json @@ -0,0 +1,3197 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.145, + 0.049, + 0.853, + 0.107 + ], + "angle": 0, + "content": "REFRAME: Reflective Surface Real-Time Rendering for Mobile Devices" + }, + { + "type": "text", + "bbox": [ + 0.279, + 0.143, + 0.719, + 0.163 + ], + "angle": 0, + "content": "Chaojie Ji\\(^{1}\\), Yufeng Li\\(^{2}\\), and Yiyi Liao\\(^{1\\boxtimes}\\)" + }, + { + "type": "text", + "bbox": [ + 0.391, + 0.178, + 0.608, + 0.197 + ], + "angle": 0, + "content": "1 Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.197, + 0.791, + 0.215 + ], + "angle": 0, + "content": "2 University of Electronic Science and Technology of China" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.258, + 0.917, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.511, + 0.92, + 0.601 + ], + "angle": 0, + "content": "Fig. 1: REFRAME enables real-time rendering on consumer GPUs and mobile devices, delivering superior subjective quality with a low face number compared to the baselines [9, 45]. Additionally, it effectively decouples the appearance properties and environmental information of the scene, which helps its capabilities for downstream scene editing tasks." + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.643, + 0.852, + 0.877 + ], + "angle": 0, + "content": "Abstract. This work tackles the challenging task of achieving real-time novel view synthesis for reflective surfaces across various scenes. Existing real-time rendering methods, especially those based on meshes, often have subpar performance in modeling surfaces with rich view-dependent appearances. Our key idea lies in leveraging meshes for rendering acceleration while incorporating a novel approach to parameterize view-dependent information. We decompose the color into diffuse and specular, and model the specular color in the reflected direction based on a neural environment map. Our experiments demonstrate that our method achieves comparable reconstruction quality for highly reflective surfaces compared to state-of-the-art offline methods, while also efficiently enabling real-time rendering on edge devices such as smartphones. Our project page is at https://xdimlab.github.io/REFrame/." + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.895, + 0.821, + 0.914 + ], + "angle": 0, + "content": "Keywords: Reflective surface \\(\\cdot\\) Real-time rendering \\(\\cdot\\) Mobile device" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.93, + 0.348, + 0.948 + ], + "angle": 0, + "content": "\\(\\mathbb{O}\\) Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.018, + 0.095, + 0.03 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.16, + 0.016, + 0.258, + 0.032 + ], + "angle": 0, + "content": "C.Ji et al." + }, + { + "type": "title", + "bbox": [ + 0.082, + 0.054, + 0.309, + 0.075 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.093, + 0.918, + 0.171 + ], + "angle": 0, + "content": "Novel view synthesis (NVS) generates realistic images from novel viewpoints using multiple input views. While Neural Radiance Fields (NeRF) [33] excel at high-quality NVS through volume rendering, they struggle with modeling reflective appearance and lack real-time rendering capabilities." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.172, + 0.918, + 0.365 + ], + "angle": 0, + "content": "Several methods [19, 52, 59, 63] extend NeRF to decouple the intrinsic scene properties, e.g., into materials and lighting, and obtain color by the rendering equation [20]. Decoupling the environmental lighting and the physical parameters of the object often helps in modeling reflective objects. In contrast to recovering the precise physical meaning which may harm the visual quality due to the approximated rendering equation, another line of works [14, 46, 56] avoids full decomposition but enables representing reflective objects by modeling the reflective radiance. However, none of the aforementioned methods is capable of real-time rendering, especially on mobile devices, due to the expensive query of the rendering equation, or the underlying volume rendering formulation." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.366, + 0.918, + 0.695 + ], + "angle": 0, + "content": "There are many existing methods focusing on accelerating the rendering of NeRF. One line of works sticks with the volume rendering pipeline where the color of a pixel is accumulated along the ray. Acceleration is achieved by improving the sampling strategy [25, 27, 31, 48], tabulating the intermediate output [12, 16, 18, 34, 40, 49, 55, 58], or leveraging super-resolution neural rendering [28]. However, deploying these methods on edge devices such as smartphones is often greatly hindered due to the demanding computational power they typically require. Besides, tabulation-based methods require large memory consumption and greatly increase communication costs when transmitting data between the cloud server and the client. Another line of works [9, 39, 45] distills radiance fields into a mesh for real-time rendering, combined with a small MLP to model view-dependent effects. Mesh-based methods can leverage traditional graphics pipelines for acceleration, enabling them to achieve real-time rendering even on edge devices. Nonetheless, real-time rendering methods [9, 15, 16, 45, 58] often struggle to model objects with highly reflective surfaces. Besides, these mesh-based real-time rendering methods [9, 45, 57] typically require a large number of vertices and faces to achieve high fidelity." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.697, + 0.918, + 0.948 + ], + "angle": 0, + "content": "In this paper, we propose REFRAME, a mesh-based REFlective surface ReAl-time rendering method for MobiE devices (e.g., smartphones), see Fig. 1. We find the fact that existing NeRF distilled mesh rendering pipelines [9,45] do not perform well in rendering objects with highly reflective appearances mainly for two reasons. Firstly, these methods utilize the viewing direction to model the view-dependent appearance, approved to be less effective than using the reflection direction [46]. As such, our method adopts the reflection direction-based parameterization. Nevertheless, this parameterization requires surface normal, yet estimating precise vertex normal based on a mesh representation can be challenging. Therefore, we propose a geometry learner that learns vertex and normal offsets through two networks. This leads to good normal estimations to calculate the reflection direction while maintaining relatively low vertex and face numbers. Secondly and more importantly, to alleviate inference burden, real-time" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.72, + 0.016, + 0.839, + 0.032 + ], + "angle": 0, + "content": "REFRAME" + }, + { + "type": "page_number", + "bbox": [ + 0.904, + 0.017, + 0.919, + 0.031 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.057, + 0.923, + 0.29 + ], + "angle": 0, + "content": "rendering methods often have limited capacity to model complex view-dependent information. In order to enhance the expressive power of the model without increasing the inference computational cost, we employ a four-layer MLP during training to learn the environmental features. This information is then baked into a two-dimensional environment feature map during inference. Remarkably, the distilled environment feature map incurs a memory overhead of less than 1MB and can be edited for relighting purposes. Finally, despite our method achieving real-time rendering with an advantage in mesh faces and vertices number compared to existing works, the reconstruction quality of our method is comparable to the current state-of-the-art (including non-real-time methods) work and even outperforms them in rendering highly reflective objects. We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.3, + 0.918, + 0.358 + ], + "angle": 0, + "content": "- We propose a novel mesh geometry learner, allowing for robustly optimizing mesh vertex positions and normals. This leads to high rendering quality with relatively low mesh vertex and face numbers." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.358, + 0.918, + 0.435 + ], + "angle": 0, + "content": "- We propose to use an environment feature map to model view-dependent appearances of highly reflective objects, which enhances the capacity to reconstruct complex reflective appearances without increasing the inference burden. This further enables relighting effects." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.436, + 0.918, + 0.514 + ], + "angle": 0, + "content": "- Our rendering quality is on par with the current state-of-the-art methods while being able to achieve real-time rendering across various platform devices. Moreover, our method even surpasses the current non-real-time state-of-the-art approaches in rendering objects with highly reflective appearances." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.3, + 0.918, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.541, + 0.332, + 0.563 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.58, + 0.922, + 0.715 + ], + "angle": 0, + "content": "NeRF-based Scene Representation: NeRF [33] and its derivative works [1-3,41,44,46,53] have employed ray-marching based volume rendering methods to achieve high-quality and realistic rendering of different types of objects in various environments. Many follow-up works have extended NeRF in different aspects, including dynamic scene modeling [36-38], 3D-aware generation [7,8,42], and semantic scene understanding [11,24,65]. In this work, we focus on addressing two limitations of NeRF, i.e., modeling of reflective objects and real-time rendering." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.716, + 0.922, + 0.95 + ], + "angle": 0, + "content": "Reflectance Decomposition: A number of works [4, 5, 19, 43, 52, 59-64] investigate the task of inferring geometry and material properties based on neural fields representation, typically by formulating the image generation using the physically-based rendering. Despite achieving lighting and material control, these methods are typically inferior to state-of-the-art NeRF-based methods that directly model the radiance in terms of the rendering quality. This is due to the fact that these methods rely on simplified rendering equations of the real world. Another line of works achieves better rendering quality by avoiding full decomposition, yet allowing for modeling glossy objects by modeling the reflected radiance [14, 46, 56] or replacing the explicit rendering equation with a learned neural renderer [6, 30]. However, none of these methods are capable of real-time rendering due to the underlying volume rendering formulation. While" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.018, + 0.095, + 0.03 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.16, + 0.016, + 0.26, + 0.032 + ], + "angle": 0, + "content": "C.Ji et al." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.055, + 0.918, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.262, + 0.918, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.379, + 0.922, + 0.575 + ], + "angle": 0, + "content": "Fig. 2: Pipeline for REFRAME. Components with a yellow background are either baked or omitted during the mobile rendering stage. Training Stage: The initial mesh is updated first before performing differentiable rasterization [26]. Next, we obtain the diffuse color \\(\\mathbf{c}_d\\) and specular feature \\(\\mathbf{f}_s\\) based on the position \\(\\mathbf{p}\\), and the environment feature \\(\\mathbf{f}_e\\) based on reflective direction \\(\\omega_r\\). Then we obtain the specular color \\(\\mathbf{c}_s\\) and combine it with \\(\\mathbf{c}_d\\) to create the final full color \\(\\mathbf{c}\\). Mobile Rendering Stage: We bake the intermediate output for real-time rendering. This mesh-based rendering can be implemented using WebGL and easily deployed on various platforms (e.g., desktop and mobile devices). Here, we retrieve the \\(\\mathbf{c}_d\\) and \\(\\mathbf{f}_s\\) from baked texture images, and the \\(\\mathbf{f}_e\\) from the environment feature map. \\(\\mathbf{c}_d\\) and \\(\\mathbf{c}_s\\) are processed the same as the training stage to obtain \\(\\mathbf{c}\\)." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.616, + 0.92, + 0.713 + ], + "angle": 0, + "content": "NvDiffRec [35] enables real-time rendering based on the mesh representation, its quality is also limited by the full decomposition. In this work, we follow the partial decomposition pipeline [46] and propose to learn a neural environment map based on a NeRF-distilled mesh representation, with a focus on achieving real-time rendering." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.714, + 0.922, + 0.85 + ], + "angle": 0, + "content": "NeRF Acceleration: Accelerating NeRF rendering is another significant research area, with approaches reducing sampling points along the ray [25,31,48], tabulating the intermediate output [12, 16, 18, 49, 55, 58], using thousands of small MLPs to represent the scene [10, 40], or utilizing super-resolution techniques [28, 51]. However, these methods typically require consumer GPUs for real-time rendering, making it challenging to deploy them on edge devices with limited computational resources." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.852, + 0.922, + 0.951 + ], + "angle": 0, + "content": "More recently, a few approaches [9, 45, 57] propose to distill a NeRF representation into a mesh for real-time rendering, where MobileNeRF [9] and NeRF2Mesh [45] enable real-time rendering on mobile devices. However, MobileNeRF [9] does not model the accurate geometry of the scene, leading to expensive memory costs with a large number of vertices and faces. NeRF2Mesh [45]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.72, + 0.016, + 0.839, + 0.032 + ], + "angle": 0, + "content": "REFRAME" + }, + { + "type": "page_number", + "bbox": [ + 0.904, + 0.017, + 0.918, + 0.031 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.055, + 0.923, + 0.212 + ], + "angle": 0, + "content": "is the most relevant work to ours, which distills grid-based representation into a mesh. The good geometry prior for NeRF2Mesh [45] leads to a relatively low number of vertices and faces. However, none of the aforementioned methods are capable of modeling highly reflective objects faithfully due to their way of view-dependent color modeling. We propose to use a neural environment map to tackle this challenge by modeling the reflected radiance. Furthermore, we propose to estimate a high-quality vertex normal, further reducing the number of faces and vertices while maintaining the quality." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.246, + 0.241, + 0.267 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.292, + 0.92, + 0.486 + ], + "angle": 0, + "content": "The pipeline of our method is illustrated in Fig. 2. Before training, we employ a volumetric rendering technique to obtain an initial coarse mesh similar to existing methods [9,39,45,57]. During the training stage, we leverage a geometry learner to update both the vertices and the vertex normals (Sec. 3.1). We further learn a shader that decomposes diffuse and specular color (Sec. 3.2), where the specular branch is designed to be able to model highly reflective surfaces by combining view-independent specular features and reflective direction-conditioned environment features. After training with loss functions introduced in Sec. 3.3, we perform UV unwrapping and bake both view-independent and view-dependent features to enable real-time rendering on various devices (Sec. 3.4)." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.488, + 0.92, + 0.624 + ], + "angle": 0, + "content": "More formally, let \\(\\mathcal{M} = (\\mathcal{V},\\mathcal{F})\\) denote the initial coarse mesh where \\(\\mathcal{V} = \\{\\mathbf{v}\\in \\mathbb{R}^3\\}\\) is the set of vertices and \\(\\mathcal{F}\\) is the set of faces. We further compute initial vertex normals \\(\\mathcal{N} = \\{\\mathbf{n}_{\\mathbf{v}}\\in \\mathbb{R}^{3}\\}\\) given the initial mesh. Let \\(\\mathbf{p}\\in \\mathbb{R}^3\\) denote a rendered surface point on the mesh, \\(\\mathbf{n_p}\\in \\mathbb{R}^3\\) the corresponding normal, \\(\\omega_{o}\\in \\mathbb{S}^{2}\\) the viewing direction, and \\(\\omega_{r}\\in \\mathbb{S}^{2}\\) the reflective direction. During training, our goal is to refine the mesh \\(\\mathcal{M}\\), as well as map \\((\\mathbf{p},\\boldsymbol {\\omega}_r)\\) to a color \\(\\mathbf{c}\\in \\mathbb{R}^3\\) to enable mesh-based real-time rendering." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.658, + 0.368, + 0.677 + ], + "angle": 0, + "content": "3.1 Geometry Learner" + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.694, + 0.92, + 0.811 + ], + "angle": 0, + "content": "Before training, we follow NeRF2Mesh [45] to extract the initial coarse mesh through grid-based representation [34]. This allows for better modeling of depth discontinuities in geometry [50] and prevents the optimization of mesh geometry from getting stuck in local optima. Then, we propose to leverage a geometry learner to refine the geometry of the initial coarse mesh, including refining the vertex positions and normals." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.812, + 0.92, + 0.948 + ], + "angle": 0, + "content": "Vertex Offset: As shown in Fig. 3, the initial mesh extracted from NeRF-based methods may be of low quality. Existing methods [45,54] directly update the vertex positions through gradient backpropagation, which means the vertex update step size is determined by the learning rate and the gradient backpropagated. Thus, costly and difficult manual learning rate tuning is needed to adapt update step sizes for different objects, with a lack of robustness as shown in our experiments. To address this, our method uses a hybrid representation to provide a" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.079, + 0.017, + 0.096, + 0.03 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.16, + 0.016, + 0.26, + 0.032 + ], + "angle": 0, + "content": "C.Ji et al." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.058, + 0.298, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.172, + 0.273, + 0.187 + ], + "angle": 0, + "content": "(a) Initial normal." + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.059, + 0.523, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.172, + 0.518, + 0.187 + ], + "angle": 0, + "content": "(b) Optimized normal." + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.053, + 0.717, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.172, + 0.699, + 0.187 + ], + "angle": 0, + "content": "(c) Initial mesh." + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.053, + 0.911, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.731, + 0.172, + 0.911, + 0.187 + ], + "angle": 0, + "content": "(d) Optimized mesh." + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.202, + 0.921, + 0.237 + ], + "angle": 0, + "content": "Fig. 3: Mesh Optimization. The geometry of the initial mesh is often poor. We are able to significantly improve the geometry of the mesh through geometry learner." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.279, + 0.812, + 0.299 + ], + "angle": 0, + "content": "learned, adaptive step size for different vertices and different objects:" + }, + { + "type": "equation", + "bbox": [ + 0.342, + 0.316, + 0.919, + 0.339 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} ^ {\\prime} = \\mathbf {v} + \\Delta \\mathbf {v}, \\quad \\Delta \\mathbf {v} = g _ {\\theta_ {\\mathbf {v}}} (\\mathbf {v}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.354, + 0.921, + 0.432 + ], + "angle": 0, + "content": "where \\(\\mathbf{v}'\\) is the updated vertex, \\(g_{\\theta_{\\mathbf{v}}}\\) is a multi-resolution hash encoding in combined with a small MLP [34]. Though [47] also uses MLP for vertex offset learning, our coarse-to-fine hybrid representation allows us to leverage local information as well as global information, leading to better modeling of the geometry." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.434, + 0.92, + 0.55 + ], + "angle": 0, + "content": "Normal Offset: Our method relies on surface normals to estimate the reflection direction \\(\\omega_r\\) to model view-dependent appearance. Following the classical idea of smooth shading, we leverage smoothly interpolated vertex normals to approximate our surface normal. Obtaining accurate vertex normals is yet challenging, especially when the number of faces is limited. Hence, we learn a per-vertex normal offset by taking \\(\\mathbf{v}\\) and \\(\\mathbf{n_v}\\) as input:" + }, + { + "type": "equation", + "bbox": [ + 0.299, + 0.573, + 0.918, + 0.595 + ], + "angle": 0, + "content": "\\[\n\\mathbf {n} _ {\\mathbf {v}} ^ {\\prime} = \\mathbf {n} _ {\\mathbf {v}} + \\Delta \\mathbf {n} _ {\\mathbf {v}}, \\quad \\Delta \\mathbf {n} _ {\\mathbf {v}} = g _ {\\theta_ {\\mathbf {n}}} (\\mathbf {v}, \\mathbf {n} _ {\\mathbf {v}}) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.606, + 0.922, + 0.724 + ], + "angle": 0, + "content": "where \\(\\mathbf{n}_{\\mathbf{v}}^{\\prime}\\) is the learned per-vertex normal and \\(g_{\\theta_{\\mathbf{n}}}\\) is another multi-resolution hash encoding-based network. In the experimental section, we demonstrate that learning such a per-vertex normal allows for modeling more accurate surface normals without increasing the number of vertices and faces. Note that \\(g_{\\theta_{\\mathbf{v}}}\\) and \\(g_{\\theta_{\\mathbf{n}}}\\) are only used during training, as we can directly save the updated vertices \\(\\mathbf{v}^{\\prime}\\) and normals \\(\\mathbf{n}_{\\mathbf{v}}^{\\prime}\\) for real-time rendering." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.757, + 0.37, + 0.775 + ], + "angle": 0, + "content": "3.2 Color Formulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.891, + 0.814 + ], + "angle": 0, + "content": "We decompose the final color \\(\\mathbf{c}\\) into a diffuse color \\(\\mathbf{c}_d\\) and a specular one \\(\\mathbf{c}_s\\):" + }, + { + "type": "equation", + "bbox": [ + 0.349, + 0.833, + 0.918, + 0.853 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} = \\min \\left(\\max \\left(\\mathbf {c} _ {d} + \\mathbf {c} _ {s}, 0\\right), 1\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.87, + 0.92, + 0.909 + ], + "angle": 0, + "content": "where \\(\\mathbf{c}_d + \\mathbf{c}_s\\) is clamped to [0, 1], both individual terms \\(\\mathbf{c}_d\\) and \\(\\mathbf{c}_s\\) from sigmoid are in [0, 1]. We now elaborate on the diffuse and specular color formulation." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.91, + 0.922, + 0.951 + ], + "angle": 0, + "content": "Diffuse Color Formulation: With the updated mesh \\(\\mathcal{M}' = (\\mathcal{V}',\\mathcal{N}',\\mathcal{F})\\), we perform differentiable rasterization [26] to obtain the position \\(\\mathbf{p}\\) corresponding" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.72, + 0.016, + 0.84, + 0.032 + ], + "angle": 0, + "content": "REFRAME" + }, + { + "type": "page_number", + "bbox": [ + 0.904, + 0.017, + 0.92, + 0.031 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.056, + 0.92, + 0.096 + ], + "angle": 0, + "content": "to each pixel. Subsequently, \\(\\mathbf{p}\\) is mapped to a diffuse color \\(\\mathbf{c}_d\\in \\mathbb{R}^3\\) and a view-independent feature \\(\\mathbf{f}_s\\in \\mathbb{R}^3\\) which will be used later for decoding specular color:" + }, + { + "type": "equation", + "bbox": [ + 0.427, + 0.117, + 0.919, + 0.136 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {d}, \\mathbf {f} _ {s} = f _ {\\phi} (\\mathbf {p}) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.155, + 0.919, + 0.195 + ], + "angle": 0, + "content": "While the number of channels of \\(\\mathbf{f}_s\\) is flexible, we observe that such a compact three-channel representation is sufficient for our purpose." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.196, + 0.92, + 0.293 + ], + "angle": 0, + "content": "Specular Color Formulation: Recent non-real-time methods [14, 46, 56] propose to model view-dependent color based on the reflective direction \\(\\omega_{r}\\), instead of the viewing direction used by NeRF. This approach has been shown to effectively model reflective surfaces. Motivated by these methods, we calculate the reflective direction as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.372, + 0.316, + 0.918, + 0.334 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\omega} _ {r} = 2 \\left(\\boldsymbol {\\omega} _ {o} \\cdot \\mathbf {n} _ {\\mathbf {p}}\\right) \\mathbf {n} _ {\\mathbf {p}} - \\boldsymbol {\\omega} _ {o} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.919, + 0.393 + ], + "angle": 0, + "content": "where \\(\\mathbf{n}_{\\mathbf{p}}\\) is smoothly interpolated surface normal at \\(\\mathbf{p}\\) based on our learned vertex normal \\(\\mathbf{n}_{\\mathbf{v}}^{\\prime}\\). Next, the specular color can be obtained as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.415, + 0.918, + 0.433 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {s} = f _ {\\psi} \\left(\\mathbf {f} _ {s}, \\boldsymbol {\\omega} _ {r}, \\left(\\boldsymbol {\\omega} _ {o} \\cdot \\mathbf {n} _ {\\mathbf {p}}\\right)\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.452, + 0.919, + 0.492 + ], + "angle": 0, + "content": "where \\((\\boldsymbol{\\omega}_o \\cdot \\mathbf{n_p})\\) is considered as input to model the Fresnel effects [21] and \\(f_{\\psi}\\) needs to be a tiny MLP to enable real-time rendering." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.494, + 0.92, + 0.651 + ], + "angle": 0, + "content": "Environment Learner: In practice, we observe that naively mapping the reflective viewing direction to specular color (see Eq. (6)) leads to unsatisfying performance due to the capacity limitation of the tiny MLP \\( f_{\\psi} \\). For a similar reason, Ref-NeRF adopts an 8-layer MLP to map \\( \\mathbf{f}_s \\) and \\( \\omega_r \\) to the view-dependent color. However, increasing the capacity of \\( f_{\\psi} \\) hampers the real-time rendering capability. Therefore, instead of increasing the capacity of \\( f_{\\psi} \\), we propose to leverage a neural environment learner of large capacity that maps \\( \\omega_r \\) to an environment feature \\( \\mathbf{f}_e \\in \\mathbb{R}^M \\) (\\( M = 3 \\) in our paper):" + }, + { + "type": "equation", + "bbox": [ + 0.422, + 0.679, + 0.918, + 0.696 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {e} = f _ {e} \\left(\\gamma \\left(\\boldsymbol {\\omega} _ {r}\\right)\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.712, + 0.919, + 0.752 + ], + "angle": 0, + "content": "where \\(\\gamma (\\cdot)\\) denotes positional encoding and \\(f_{e}\\) is a 4-layer MLP. This leads to our final implementation of the specular color:" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.773, + 0.918, + 0.791 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {s} = f _ {\\psi} \\left(\\mathbf {f} _ {s}, \\mathbf {f} _ {e}, \\left(\\boldsymbol {\\omega} _ {o} \\cdot \\mathbf {n} _ {\\mathbf {p}}\\right)\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.811, + 0.919, + 0.851 + ], + "angle": 0, + "content": "Note that \\(\\mathbf{f}_e\\) can be baked into a two-dimensional neural environment feature map for real-time rendering as it solely depends on \\(\\omega_r\\)." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.852, + 0.92, + 0.95 + ], + "angle": 0, + "content": "Scene Editing: REFRAME can decouple geometry, diffuse and specular color, so we can perform simple scene editing tasks. For example, we can edit the geometry and appearance of objects by modifying the mesh or texture image. Additionally, we can perform relighting on the objects by modifying the environment feature map of the scene, as shown in Fig. 4." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.079, + 0.017, + 0.096, + 0.03 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.16, + 0.016, + 0.26, + 0.032 + ], + "angle": 0, + "content": "C.Ji et al." + }, + { + "type": "image", + "bbox": [ + 0.109, + 0.051, + 0.335, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.176, + 0.3, + 0.192 + ], + "angle": 0, + "content": "(a) Original image." + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.051, + 0.61, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.421, + 0.176, + 0.569, + 0.192 + ], + "angle": 0, + "content": "(b) Flip lighting." + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.051, + 0.883, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.176, + 0.858, + 0.192 + ], + "angle": 0, + "content": "(c) Change lighting." + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.206, + 0.921, + 0.298 + ], + "angle": 0, + "content": "Fig. 4: Relighting. We can achieve relighting of objects by editing our environment feature map. The image illustrates the editing process of flipping or replacing the environmental feature map. The left side shows the rendering result, the top right corner presents the corresponding environment feature map, and the specular color is displayed in the bottom right corner." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.334, + 0.325, + 0.351 + ], + "angle": 0, + "content": "3.3 Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.364, + 0.921, + 0.462 + ], + "angle": 0, + "content": "We train our model based on image reconstruction loss and several regularization terms. At each iteration, we render a full image and train with the corresponding ground truth. Let \\(\\hat{\\mathbf{c}}\\) denote the ground truth RGB value at one pixel. We optionally use ground truth binary mask for supervision at synthetic object-centric scenes. Our full loss function is shown in Eq. (9):" + }, + { + "type": "equation", + "bbox": [ + 0.269, + 0.476, + 0.918, + 0.521 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\lambda_ {\\mathbf {c}} \\mathcal {L} _ {\\mathbf {c}} + \\lambda_ {\\mathbf {c} _ {d}} \\mathcal {L} _ {\\mathbf {c} _ {d}} + \\lambda_ {S S I M} \\mathcal {L} _ {S S I M} + \\tag {9} \\\\ \\lambda_ {m a s k} \\mathcal {L} _ {m a s k} + \\lambda_ {\\Delta \\mathbf {n} _ {\\mathbf {v}}} \\mathcal {L} _ {\\Delta \\mathbf {n} _ {\\mathbf {v}}} + \\lambda_ {m a x} \\mathcal {L} _ {m a x} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.533, + 0.92, + 0.573 + ], + "angle": 0, + "content": "Image Reconstruction Loss: Our model is first supervised by the L2 loss between the final rendered color and the ground truth." + }, + { + "type": "equation", + "bbox": [ + 0.418, + 0.586, + 0.918, + 0.607 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathbf {c}} = \\left| \\left| \\mathbf {c} - \\hat {\\mathbf {c}} \\right| \\right| _ {2} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.62, + 0.918, + 0.659 + ], + "angle": 0, + "content": "To encourage the correct decoupling of diffuse and specular colors, we add a diffuse loss to guide the diffuse color to be close to the ground truth color:" + }, + { + "type": "equation", + "bbox": [ + 0.409, + 0.673, + 0.918, + 0.695 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathbf {c} _ {d}} = \\left\\| \\mathbf {c} _ {d} - \\hat {\\mathbf {c}} \\right\\| _ {2} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.706, + 0.918, + 0.783 + ], + "angle": 0, + "content": "We also introduce SSIM [17] loss for better perceptual quality. On object-centric datasets, we additionally utilize mask loss to ensure the rasterization mask closely aligns with the ground truth mask. Note that the mask loss is optional and is not required on unbounded datasets." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.785, + 0.92, + 0.862 + ], + "angle": 0, + "content": "Regularizations: We regularize the predicted vertex normal \\(\\mathbf{n}_{\\mathbf{v}}^{\\prime}\\) to be close to the original vertex normal \\(\\mathbf{n}_{\\mathbf{v}}\\) computed from the mesh. This allows for stabilizing the optimization of the vertex normals, e.g., avoid completely flipping the normal. This is achieved by regularizing the magnitude of the normal offset:" + }, + { + "type": "equation", + "bbox": [ + 0.406, + 0.877, + 0.918, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\Delta \\mathbf {n} _ {\\mathbf {v}}} = \\left\\| \\Delta \\mathbf {n} _ {\\mathbf {v}} \\right\\| _ {1} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.911, + 0.921, + 0.951 + ], + "angle": 0, + "content": "We introduce another regularization to encourage the specular color \\(\\mathbf{c}_s\\) to be reasonable when visualized alone. As shown in Eq. (3), the final color \\(\\mathbf{c}\\) will be" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.72, + 0.016, + 0.839, + 0.032 + ], + "angle": 0, + "content": "REFRAME" + }, + { + "type": "page_number", + "bbox": [ + 0.903, + 0.017, + 0.92, + 0.031 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.057, + 0.921, + 0.154 + ], + "angle": 0, + "content": "clamped when the sum of specular color \\(\\mathbf{c}_s\\) and diffuse color \\(\\mathbf{c}_d\\) exceeds 1. In this case, the gradient is truncated due to the clamp operation, preventing the backpropagation from updating the color values. This may leave an undesired large specular color \\(\\mathbf{c}_s\\) with no penalization. Therefore, we encourage the sum of specular color and diffuse color to not exceed 1:" + }, + { + "type": "equation", + "bbox": [ + 0.338, + 0.168, + 0.917, + 0.186 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\max } = \\left| \\left| \\operatorname {m a x} \\left(\\mathbf {c} _ {d} + \\mathbf {c} _ {s} - 1, 0\\right) \\right| \\right| \\tag {13}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.212, + 0.411, + 0.233 + ], + "angle": 0, + "content": "3.4 Real-Time Rendering" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.242, + 0.921, + 0.416 + ], + "angle": 0, + "content": "Environment Feature Map: During the training process, we employ a four-layer MLP with a width of 256 as our environment learner. If we query \\(\\mathbf{f}_e\\) through the environment learner module during the inference process, it would significantly slow down the inference speed. Hence, during the inference phase, we bake the learned environment feature \\(\\mathbf{f}_e\\) into a 2D environment feature map by converting \\(\\omega_r \\in \\mathbb{S}^2\\) to its polar coordinate, see supplementary for more details. This allows us to simply query the corresponding feature on the environment feature map based on \\(\\omega_r\\), thereby greatly accelerating our inference speed, reaching above 200 FPS on a single NVIDIA 3090 GPU." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.418, + 0.92, + 0.535 + ], + "angle": 0, + "content": "Texture Images: To further reduce the inference burden and better deploy our method across different platforms, we can also bake the \\(\\mathbf{c}_d\\) and \\(\\mathbf{f}_s\\) obtained from querying the \\(f_{\\phi}\\) into two texture images. Specifically, we map the vertices of the mesh to UV coordinates and then bake the features into two texture images, respectively. In order to obtain the normal \\(\\mathbf{n_p}\\) from the learned normal vector \\(\\mathbf{n_v}'\\) during real-time rendering, we also bake the normal \\(\\mathbf{n_p}\\) into a texture image." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.561, + 0.301, + 0.584 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.599, + 0.92, + 0.949 + ], + "angle": 0, + "content": "Implementation Details. Training: Prior to the training stage, we employ the classic quadric error metrics algorithm [13] to simplify the initial mesh to 75,000 faces except for the outdoor scenes which typically have complex geometry. We train for 250 epochs on each scene, with a training time of approximately 2 hours. We utilize the Adam optimizer [23] and employ the cosine annealing learning rate adjustment strategy [32] during the training stage. Rendering: The resolution of our baked texture image is \\(4096 \\times 4096\\), and the resolution of the baked environment feature map is \\(720 \\times 360\\). As referenced in Sec. 3.4, once we have baked the environment learner into the environment feature map, our approach is already capable of achieving above 200 FPS on the GPU (referred to as Ours). This implementation ensures a fair comparison with the baselines since their results quoted are evaluated using GPU implementation that can not directly deploy on mobile devices. Given that we use texture images for rendering when implemented on mobile devices, we also employ texture images on the GPU for quality evaluation (referred to as Ours (Mobile)). Following NeRF2Mesh [45] and MobileNeRF [9], we employ anti-aliasing to enhance our rendering quality. For details on the implementation of anti-aliasing, please refer to the supplementary material. All experiments are performed on a single NVIDIA 3090 GPU." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.105, + 0.032 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.261, + 0.032 + ], + "angle": 0, + "content": "C.Ji et al." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.055, + 0.919, + 0.23 + ], + "angle": 0, + "content": "
NeRF Synthetic DatasetShiny Blender DatasetReal Captured Dataset
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NeRF [33]31.010.9470.081------
Ref-NeRF [46]33.990.9660.03835.960.9670.05824.450.6650.142
3DGS [22]33.300.9690.03030.370.9470.08324.060.6610.259
NvDiffRec [35]29.050.9390.08129.050.9380.111---
MobileNeRF [9]30.900.9470.06226.620.8830.16322.960.4930.430
NeRF2Mesh [45]29.670.9400.07226.960.8960.17022.710.5230.419
Ours31.040.9610.04036.020.9830.04022.500.5980.325
Ours (Mobile)30.840.9590.04235.830.9810.04122.110.5240.423
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.232, + 0.921, + 0.285 + ], + "angle": 0, + "content": "Table 1: Rendering Quality. Baseline comparisons of the rendering quality on three different datasets. Red represents the optimal, orange represents the second best, and yellow represents the third." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.34, + 0.922, + 0.554 + ], + "angle": 0, + "content": "Dataset. We validate the effectiveness and robustness of our method using three different datasets with varied scenes. 1) NeRF [33] Synthetic Dataset: This dataset consists of eight synthetic scenes. 2) Shiny Blender Dataset: This dataset is introduced by Ref-NeRF [46] and includes six glossy objects, serving as the primary dataset for assessing the reflective surface modeling capabilities of the methods. 3) Real Captured Dataset: We select three outdoor scenes with rich reflective appearances same as Ref-NeRF [46] to validate the effectiveness of our method in real-world outdoor capture scenarios. Due to the poor quality of the initial mesh extracted by Nerf2Mesh [45] when modeling glossy objects, we utilize the meshes extracted by Neuralangelo [29] and Ref-NeuS [14] as our initialization in certain scenes, see supplementary for more details." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.555, + 0.922, + 0.732 + ], + "angle": 0, + "content": "Baseline. We compare with various types of baselines, including Ref-NeRF [46], 3DGS [22], NvDiffRec [35], MobileNeRF [9], and NeRF2Mesh [45] in our study. Ref-NeRF [46] serves as the state-of-the-art volume rendering method for handling the strong reflective appearance. 3DGS [22] represents a powerful new representation capable of real-time rendering on GPU. NvDiffRec [35] represents physically based rendering methods. MobileNeRF [9], and NeRF2Mesh [45] are representative works in the field of real-time rendering for mobile devices. We import the results of the baselines from their original paper and replicate the results that missing in their original paper." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.761, + 0.43, + 0.781 + ], + "angle": 0, + "content": "4.1 Reconstruction Quality" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.793, + 0.922, + 0.95 + ], + "angle": 0, + "content": "Rendering Quality. We compare the rendering quality of our method and baseline methods on three datasets, see Tab. 1. It is worth noting that Ref-NeRF [46] is the state-of-the-art method for modeling glossy objects but cannot achieve real-time rendering, even on high-end GPU. We introduce Ref-NeRF [46] as a benchmark for rendering quality and demonstrate that our method achieves comparable or even superior rendering quality to non-real-time state-of-the-art methods. While 3DGS [22] can render in real-time on GPU and has higher PSNR on both NeRF Synthetic and Real Captured dataset than ours, it's important" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.72, + 0.016, + 0.839, + 0.032 + ], + "angle": 0, + "content": "REFRAME" + }, + { + "type": "page_number", + "bbox": [ + 0.894, + 0.017, + 0.917, + 0.03 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.096, + 0.095, + 0.114 + ], + "angle": 0, + "content": "C" + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.159, + 0.095, + 0.194 + ], + "angle": 0, + "content": "Helmert" + }, + { + "type": "image_caption", + "bbox": [ + 0.079, + 0.241, + 0.096, + 0.274 + ], + "angle": 0, + "content": "Toaster" + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.052, + 0.21, + 0.064 + ], + "angle": 0, + "content": "NvDiffRec" + }, + { + "type": "image", + "bbox": [ + 0.138, + 0.065, + 0.264, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.14, + 0.207, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.215, + 0.196, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.053, + 0.379, + 0.063 + ], + "angle": 0, + "content": "MobileNeRF" + }, + { + "type": "image", + "bbox": [ + 0.299, + 0.065, + 0.428, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.14, + 0.371, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.215, + 0.359, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.459, + 0.053, + 0.541, + 0.063 + ], + "angle": 0, + "content": "NeRF2Mesh" + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.065, + 0.591, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.14, + 0.535, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.374, + 0.14, + 0.428, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.374, + 0.214, + 0.428, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.14, + 0.591, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.214, + 0.591, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.633, + 0.054, + 0.666, + 0.064 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.628, + 0.065, + 0.685, + 0.131 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.149, + 0.698, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.216, + 0.687, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.785, + 0.054, + 0.871, + 0.064 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.065, + 0.756, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.14, + 0.755, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.214, + 0.755, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.77, + 0.149, + 0.862, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.781, + 0.216, + 0.851, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.865, + 0.066, + 0.92, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.866, + 0.214, + 0.92, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.302, + 0.921, + 0.357 + ], + "angle": 0, + "content": "Fig. 5: Rendering Quality on Shiny Blender Dataset. Our method achieves optimal rendering quality in most scenes and provides better modeling of reflective appearance compared to the comparison methods [9,35,45]." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.422, + 0.918, + 0.482 + ], + "angle": 0, + "content": "to note that 3DGS struggles to model reflective surfaces effectively (Fig. 6a). Moreover, they have a larger memory overhead than us. \\( E.g. \\), for gardenspheres, the memory cost for 3DGS is 1.4GB, whereas our method only requires 84.3MB." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.494, + 0.92, + 0.671 + ], + "angle": 0, + "content": "In comparison to mesh-based real-time rendering methods [9, 35, 45], we achieve optimal rendering quality in most object-centric datasets, as shown in Fig. 5. Despite having mask supervision on object-centric datasets as well, NvDiffRec [35] and NeRF2Mesh [45] struggle to achieve high rendering quality. This could be due to the fact that NvDiffRec relies on the simplified rendering equation, whereas NeRF2Mesh's simple color formulation struggles to model highly reflective objects. Note that NeRF2Mesh also requires an initial mesh in its second stage, and we provide NeRF2Mesh with the same initial mesh as ours for all experiments for a fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.683, + 0.919, + 0.84 + ], + "angle": 0, + "content": "In unbounded outdoor scenes without masks, our method still achieves high-quality rendering, as demonstrated in Fig. 6a. Although our PSNR may be slightly lower compared to the baselines, we outperform them in rendering foreground glossy objects. However, this advantage is not accurately reflected in the PSNR metric since the foreground reflective objects occupy only a small region in the image. Further, our method mainly exhibits artifacts in the background, this is due to the fact that the initial mesh can only faithfully reconstruct the foreground but not the background, see supplementary for more details." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.852, + 0.92, + 0.95 + ], + "angle": 0, + "content": "Surface Quality. Furthermore, we observe that our method not only improves the rendering quality when modeling glossy objects but also generates more accurate meshes, as shown in Fig. 6b. It can be seen that the meshes obtained from [35,45] often have poor quality in the reflective region while our method is capable of accurately reconstructing the geometry." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.105, + 0.031 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.16, + 0.016, + 0.26, + 0.032 + ], + "angle": 0, + "content": "C.Ji et al." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.053, + 0.182, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.141, + 0.182, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.189, + 0.054, + 0.291, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.19, + 0.141, + 0.291, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.054, + 0.401, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.141, + 0.401, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.08, + 0.226, + 0.401, + 0.253 + ], + "angle": 0, + "content": "(a) Rendering Quality on Real Captured Dataset." + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.054, + 0.593, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.054, + 0.755, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.054, + 0.919, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.45, + 0.239, + 0.882, + 0.254 + ], + "angle": 0, + "content": "(b) Surface Quality on NeRF Synthetic Dataset." + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.269, + 0.919, + 0.306 + ], + "angle": 0, + "content": "Fig. 6: Qualitative Comparison. Compared to the baselines, our method reconstructs reflective regions of higher fidelity and yields more accurate surface geometry." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.322, + 0.913, + 0.406 + ], + "angle": 0, + "content": "
MacBook Pro iPhone12 iPad Air3 Legion Y-7000 HUAWEI nova7
MobileNeRF [9]120.0060.0060.0055.3139.19
NeRF2Mesh [45]120.0060.0060.0048.6932.56
Ours (Mobile)120.0060.0060.0048.5031.71
" + }, + { + "type": "table_caption", + "bbox": [ + 0.075, + 0.406, + 0.92, + 0.477 + ], + "angle": 0, + "content": "Table 2: Computational Efficiency Comparison. We test the FPS of our method and the baseline methods [9,45] on the NeRF Synthetic dataset across multiple devices. On MacBook Pro, iPhone 12, and iPad Air3, both our method and the baseline methods [9,45] achieve the maximum screen refresh rate." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.531, + 0.396, + 0.55 + ], + "angle": 0, + "content": "4.2 Rendering Efficiency" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.561, + 0.92, + 0.697 + ], + "angle": 0, + "content": "Computational efficiency. During the mobile rendering stage, the bottleneck of our rendering speed lies in querying the tiny MLP \\( f_{\\psi} \\) to obtain the specular color. However, due to our MLP being only two layers with 64 widths, our rendering speed remains relatively fast. Therefore, our method achieves real-time rendering on both consumer GPUs and mobile devices. We compare the rendering speed with the baselines [9,45] that can render in real-time on mobile devices across multiple platform devices, as seen in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.697, + 0.918, + 0.775 + ], + "angle": 0, + "content": "Memory efficiency. Our cache mainly consists of texture images, the environment feature map, and the mesh. Due to our high-quality estimation of vertex normals, our mesh can achieve high-quality rendering with a low number of vertices and faces, leading to a small cache size in most scenes, as shown in Tab. 3." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.821, + 0.334, + 0.841 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.852, + 0.92, + 0.95 + ], + "angle": 0, + "content": "Geometry Learner. [54] and [45] utilize direct gradient backpropagation to update the vertex positions of the mesh. We employ this update strategy in our method and verify that this update method lacks robustness. Directly using the backpropagated gradients to update the vertex positions can lead to a devastating blow to the mesh, especially when using a large learning rate. This" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.72, + 0.016, + 0.839, + 0.032 + ], + "angle": 0, + "content": "REFRAME" + }, + { + "type": "page_number", + "bbox": [ + 0.894, + 0.017, + 0.92, + 0.031 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.054, + 0.929, + 0.153 + ], + "angle": 0, + "content": "
NeRF Synthetic DatasetShiny Blender Dataset
#V (103)#F (103)Cache#V (103)#F (103)Cache
MobileNeRF [9]494224125.751028343152.37
NeRF2Mesh [45]20019273.53459122.15
Ours (Mobile)377547.55387552.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.154, + 0.92, + 0.191 + ], + "angle": 0, + "content": "Table 3: Memory Efficiency. While exhibiting higher quality, our method maintains a relatively low memory consumption compared to baselines. Cache reported in MB." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.228, + 0.465, + 0.305 + ], + "angle": 0, + "content": "
w/o gθvOurs
L=0.001L=0.0001L=0.001
Toaster17.0117.0124.95
Materials23.0129.0629.58
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.306, + 0.484, + 0.358 + ], + "angle": 0, + "content": "Table 4: Ablation of Vertex Offset Learning. L represents the learning rate. Reported in PSNR." + }, + { + "type": "table", + "bbox": [ + 0.554, + 0.228, + 0.887, + 0.323 + ], + "angle": 0, + "content": "
w/o Δn_vOurs
Toaster#F=10k24.2424.62
#F=75k24.5524.95
Materials#F=10k27.6928.23
#F=75k29.4629.58
" + }, + { + "type": "table_caption", + "bbox": [ + 0.511, + 0.324, + 0.918, + 0.36 + ], + "angle": 0, + "content": "Table 5: Ablation of Normal Offset Learning. Reported in PSNR." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.418, + 0.919, + 0.516 + ], + "angle": 0, + "content": "significantly compromises the rendering quality, as shown in Tab. 4. Additionally, different datasets require different learning rates, as indicated in Tab. 4. While a learning rate of 0.0001 works fine for the materials dataset, it performs poorly on the toaster dataset. In contrast, our vertex update strategy not only exhibits robustness across different datasets but also provides higher rendering quality." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.518, + 0.921, + 0.831 + ], + "angle": 0, + "content": "In order to verify the necessity of estimating an offset for the vertices normal, we conduct experiments with the w/o normal offset setting. In this setting, we do not estimate a normal offset for each vertex, but we still compute an updated normal based on the updated vertex positions. As shown in Tab. 5, introducing per-vertex normal offset provides a gain in reconstruction quality. This gain is particularly significant when the number of faces is low. Notably, in the case of the toaster dataset, a mesh with \\(10\\mathrm{k}\\) faces with normal offset even achieves rendering quality surpassing that of a mesh with \\(75\\mathrm{k}\\) faces without normal offset. Environment Learner. To validate the effectiveness of the environment learner \\(f_{e}\\), we directly input \\(\\omega_{r}\\) values into the tiny MLP \\(f_{\\psi}\\) as shown in Eq. (6), similar to [9] and [45], except they use viewing direction instead of reflective direction. Experimental results demonstrate that directly inputting \\(\\omega_{r}\\) into the tiny MLP \\(f_{\\psi}\\) cannot learn the correct reflection appearance as a result of the limited network's capacity, as depicted in Fig. 7. Therefore, introducing the environment learner module improves the reconstruction quality without compromising the inference speed, as demonstrated in Tab. 6." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.832, + 0.922, + 0.95 + ], + "angle": 0, + "content": "Environment Feature Map: We can directly optimize an environment feature map instead of learning it through the environment learner module and baking it. However, as shown in Tab. 7, the direct optimization of the environment feature map performs poorly when the map resolution is high. This is because, at higher resolutions, neighboring directions are mapped to distant grid points, leading to a non-smooth interpolation. As a result, the grid points can-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.105, + 0.031 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.16, + 0.016, + 0.26, + 0.032 + ], + "angle": 0, + "content": "C.Ji et al." + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.059, + 0.215, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.146, + 0.215, + 0.161 + ], + "angle": 0, + "content": "(a) \\(\\mathrm{W / o}f_{e}\\)" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.06, + 0.396, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.308, + 0.146, + 0.387, + 0.161 + ], + "angle": 0, + "content": "(b) Ours" + }, + { + "type": "image", + "bbox": [ + 0.466, + 0.06, + 0.569, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.444, + 0.146, + 0.598, + 0.16 + ], + "angle": 0, + "content": "(c) Ground Truth" + }, + { + "type": "table", + "bbox": [ + 0.624, + 0.066, + 0.899, + 0.131 + ], + "angle": 0, + "content": "
w/o feOurs
Toaster19.3124.95
Materials27.3729.58
" + }, + { + "type": "table_caption", + "bbox": [ + 0.612, + 0.146, + 0.908, + 0.198 + ], + "angle": 0, + "content": "Table 6: Ablation of Environment Learner. Reported in PSNR." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.176, + 0.609, + 0.212 + ], + "angle": 0, + "content": "Fig. 7: Ablation of Environment Learner. Eliminating \\( f_{e} \\) results in bad rendering quality." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.228, + 0.888, + 0.346 + ], + "angle": 0, + "content": "
Resolutions H*WOursDirectly optimization
PSNRCache(KB)PSNRCache(KB)
Toaster180*36024.864924.1691
900*180024.9650423.592460
Materials180*36029.556329.58114
900*180029.5870629.033000
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.347, + 0.918, + 0.385 + ], + "angle": 0, + "content": "Table 7: Ablation of Environment Feature Map. Our method outperforms in both cache size and rendering quality." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.434, + 0.92, + 0.59 + ], + "angle": 0, + "content": "not share global information and only contain its own local information. On the other hand, when the resolution is too low, the rendering quality is limited by the capacity of the environment feature map. In contrast, our method exhibits good rendering quality across different resolutions by learning the environment feature map via an MLP. Additionally, our environment feature maps enjoy the smoothness bias of the MLP, resulting in relatively smooth tensor maps. Consequently, when saving images in PNG format, our method demonstrates good compression performance. This further reduces the overhead of our caching." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.616, + 0.288, + 0.635 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.653, + 0.918, + 0.729 + ], + "angle": 0, + "content": "Although our method successfully models the appearance of glossy objects, similar to Ref-NeRF [46], it still struggles with accurately modeling interreflections and non-distant illumination. Additionally, our method requires a certain level of quality in the initial mesh." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.731, + 0.918, + 0.848 + ], + "angle": 0, + "content": "In summary, our method achieves real-time rendering on edge devices with low hardware budgets. Besides, we propose a novel approach for modeling view-dependent appearance and can optimize the appearance and geometry of glossy objects with high computational efficiency and low memory footprint. Furthermore, by effectively decoupling the scene's geometry, appearance, and environmental information, our method can perform simple scene editing tasks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.874, + 0.352, + 0.896 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.91, + 0.918, + 0.949 + ], + "angle": 0, + "content": "This work is supported by NSFC under grant 62202418 and U21B2004. We thank Xiaoran Cao, Chenxi Tu, and Sicheng Li for their valuable discussions." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.72, + 0.016, + 0.839, + 0.032 + ], + "angle": 0, + "content": "REFRAME" + }, + { + "type": "page_number", + "bbox": [ + 0.894, + 0.017, + 0.919, + 0.031 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.054, + 0.238, + 0.075 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.099, + 0.92, + 0.17 + ], + "angle": 0, + "content": "1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.172, + 0.92, + 0.243 + ], + "angle": 0, + "content": "2. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mipnerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.245, + 0.921, + 0.298 + ], + "angle": 0, + "content": "3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-nerf: Anti-aliased grid-based neural radiance fields. arXiv preprint arXiv:2304.06706 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.298, + 0.92, + 0.352 + ], + "angle": 0, + "content": "4. Bi, S., Xu, Z., Srinivasan, P., Mildenhall, B., Sunkavalli, K., Hasan, M., Hold-Geoffroy, Y., Kriegman, D., Ramamoorthi, R.: Neural reflectance fields for appearance acquisition. arXiv preprint arXiv:2008.03824 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.353, + 0.92, + 0.407 + ], + "angle": 0, + "content": "5. Boss, M., Braun, R., Jampani, V., Barron, J.T., Liu, C., Lensch, H.: Nerd: Neural reflectance decomposition from image collections. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12684-12694 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.407, + 0.918, + 0.462 + ], + "angle": 0, + "content": "6. Boss, M., Jampani, V., Braun, R., Liu, C., Barron, J., Lensch, H.: Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems 34, 10691-10704 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.462, + 0.92, + 0.534 + ], + "angle": 0, + "content": "7. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.534, + 0.92, + 0.605 + ], + "angle": 0, + "content": "8. Chan, E.R., Monteiro, M., Kellnhofer, P., Wu, J., Wetzstein, G.: pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5799-5809 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.606, + 0.92, + 0.678 + ], + "angle": 0, + "content": "9. Chen, Z., Funkhouser, T., Hedman, P., Tagliasacchi, A.: Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16569-16578 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.679, + 0.918, + 0.732 + ], + "angle": 0, + "content": "10. Esposito, S., Baieri, D., Zellmann, S., Hinkenjann, A., Rodola, E.: Kiloneus: A versatile neural implicit surface representation for real-time rendering. arXiv preprint arXiv:2206.10885 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.733, + 0.918, + 0.787 + ], + "angle": 0, + "content": "11. Fu, X., Zhang, S., Chen, T., Lu, Y., Zhu, L., Zhou, X., Geiger, A., Liao, Y.: Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation. In: 2022 International Conference on 3D Vision (3DV). pp. 1-11. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.788, + 0.918, + 0.841 + ], + "angle": 0, + "content": "12. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: Fastnerf: High-fidelity neural rendering at 200fps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14346-14355 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.842, + 0.918, + 0.895 + ], + "angle": 0, + "content": "13. Garland, M., Heckbert, P.S.: Surface simplification using quadric error metrics. In: Proceedings of the 24th annual conference on Computer graphics and interactive techniques. pp. 209-216 (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.896, + 0.918, + 0.95 + ], + "angle": 0, + "content": "14. Ge, W., Hu, T., Zhao, H., Liu, S., Chen, Y.C.: Ref-neus: Ambiguity-reduced neural implicit surface learning for multi-view reconstruction with reflection. arXiv preprint arXiv:2303.10840 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.081, + 0.099, + 0.921, + 0.95 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.106, + 0.032 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.261, + 0.032 + ], + "angle": 0, + "content": "C.Ji et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.057, + 0.918, + 0.111 + ], + "angle": 0, + "content": "15. Guo, Y.C., Cao, Y.P., Wang, C., He, Y., Shan, Y., Zhang, S.H.: Vmesh: Hybrid volume-mesh representation for efficient view synthesis. In: SIGGRAPH Asia 2023 Conference Papers. pp. 1-11 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.112, + 0.918, + 0.164 + ], + "angle": 0, + "content": "16. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking neural radiance fields for real-time view synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5875-5884 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.918, + 0.2 + ], + "angle": 0, + "content": "17. Hore, A., Ziou, D.: Image quality metrics: Psnr vs. sSIM. In: 2010 20th international conference on pattern recognition. pp. 2366-2369. IEEE (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.201, + 0.918, + 0.254 + ], + "angle": 0, + "content": "18. Hu, T., Liu, S., Chen, Y., Shen, T., Jia, J.: Efficientnerf efficient neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12902-12911 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.255, + 0.918, + 0.307 + ], + "angle": 0, + "content": "19. Jin, H., Liu, I., Xu, P., Zhang, X., Han, S., Bi, S., Zhou, X., Xu, Z., Su, H.: Tensoroir: Tensorial inverse rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 165-174 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.308, + 0.918, + 0.343 + ], + "angle": 0, + "content": "20. Kajiya, J.T.: The rendering equation. In: Proceedings of the 13th annual conference on Computer graphics and interactive techniques. pp. 143-150 (1986)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.918, + 0.379 + ], + "angle": 0, + "content": "21. Kautz, J., McCool, M.D.: Approximation of glossy reflection with prefiltered environment maps. In: Graphics Interface. vol. 2000, pp. 119-126 (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.379, + 0.918, + 0.414 + ], + "angle": 0, + "content": "22. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.415, + 0.918, + 0.45 + ], + "angle": 0, + "content": "23. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.451, + 0.918, + 0.522 + ], + "angle": 0, + "content": "24. Kundu, A., Genova, K., Yin, X., Fathi, A., Pantofaru, C., Guibas, L.J., Tagliasacchi, A., Dellaert, F., Funkhouser, T.: Panoptic neural fields: A semantic object-aware neural scene representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12871-12881 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.522, + 0.918, + 0.592 + ], + "angle": 0, + "content": "25. Kurz, A., Neff, T., Lv, Z., Zollhöfer, M., Steinberger, M.: Adanerf: Adaptive sampling for real-time rendering of neural radiance fields. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII. pp. 254-270. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.593, + 0.918, + 0.646 + ], + "angle": 0, + "content": "26. Laine, S., Hellsten, J., Karras, T., Seol, Y., Lehtinen, J., Aila, T.: Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG) 39(6), 1-14 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.918, + 0.699 + ], + "angle": 0, + "content": "27. Li, C., Li, S., Zhao, Y., Zhu, W., Lin, Y.: Rt-nerf: Real-time on-device neural radiance fields towards immersive ar/vr rendering. In: Proceedings of the 41st IEEE/ACM International Conference on Computer-Aided Design. pp. 1-9 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.7, + 0.918, + 0.735 + ], + "angle": 0, + "content": "28. Li, S., Li, H., Wang, Y., Liao, Y., Yu, L.: Steernerf: Accelerating nef rendering via smooth viewpoint trajectory. arXiv preprint arXiv:2212.08476 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.736, + 0.918, + 0.806 + ], + "angle": 0, + "content": "29. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8456-8465 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.807, + 0.918, + 0.859 + ], + "angle": 0, + "content": "30. Liang, R., Chen, H., Li, C., Chen, F., Panneer, S., Vijaykumar, N.: Envidr: Implicit differentiable renderer with neural environment lighting. arXiv preprint arXiv:2303.13022 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.918, + 0.913 + ], + "angle": 0, + "content": "31. Lin, H., Peng, S., Xu, Z., Yan, Y., Shuai, Q., Bao, H., Zhou, X.: Efficient neural radiance fields for interactive free-viewpoint video. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.914, + 0.918, + 0.949 + ], + "angle": 0, + "content": "32. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016)" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.057, + 0.918, + 0.949 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.72, + 0.016, + 0.839, + 0.032 + ], + "angle": 0, + "content": "REFRAME" + }, + { + "type": "page_number", + "bbox": [ + 0.894, + 0.017, + 0.92, + 0.031 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.057, + 0.918, + 0.111 + ], + "angle": 0, + "content": "33. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.113, + 0.92, + 0.166 + ], + "angle": 0, + "content": "34. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.167, + 0.918, + 0.239 + ], + "angle": 0, + "content": "35. Munkberg, J., Hasselgren, J., Shen, T., Gao, J., Chen, W., Evans, A., Müller, T., Fidler, S.: Extracting triangular 3d models, materials, and lighting from images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8280-8290 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.24, + 0.918, + 0.294 + ], + "angle": 0, + "content": "36. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5865-5874 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.295, + 0.918, + 0.349 + ], + "angle": 0, + "content": "37. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., MartinBrualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.35, + 0.918, + 0.403 + ], + "angle": 0, + "content": "38. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10318-10327 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.918, + 0.458 + ], + "angle": 0, + "content": "39. Rakotosaona, M.J., Manhardt, F., Arroyo, D.M., Niemeyer, M., Kundu, A., Tombari, F.: Nerfmeshing: Distilling neural radiance fields into geometrically-accurate 3d meshes. arXiv preprint arXiv:2303.09431 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.459, + 0.918, + 0.512 + ], + "angle": 0, + "content": "40. Reiser, C., Peng, S., Liao, Y., Geiger, A.: Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14335-14345 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.514, + 0.918, + 0.568 + ], + "angle": 0, + "content": "41. Rematas, K., Liu, A., Srinivasan, P.P., Barron, J.T., Tagliasacchi, A., Funkhouser, T., Ferrari, V.: Urban radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12932-12942 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.569, + 0.918, + 0.621 + ], + "angle": 0, + "content": "42. Schwarz, K., Liao, Y., Niemeyer, M., Geiger, A.: Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems 33, 20154-20166 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.623, + 0.918, + 0.695 + ], + "angle": 0, + "content": "43. Srinivasan, P.P., Deng, B., Zhang, X., Tancik, M., Mildenhall, B., Barron, J.T.: Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7495-7504 (June 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.696, + 0.918, + 0.767 + ], + "angle": 0, + "content": "44. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8248-8258 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.768, + 0.918, + 0.823 + ], + "angle": 0, + "content": "45. Tang, J., Zhou, H., Chen, X., Hu, T., Ding, E., Wang, J., Zeng, G.: Delicate textured mesh recovery from nef via adaptive surface refinement. arXiv preprint arXiv:2303.02091 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.824, + 0.918, + 0.894 + ], + "angle": 0, + "content": "46. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5481-5490. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.895, + 0.918, + 0.949 + ], + "angle": 0, + "content": "47. Walker, T., Mariotti, O., Vaxman, A., Bilen, H.: Explicit neural surfaces: Learning continuous geometry with deformation fields. arXiv preprint arXiv:2306.02956 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.057, + 0.92, + 0.949 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.105, + 0.031 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.261, + 0.032 + ], + "angle": 0, + "content": "C.Ji et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.057, + 0.918, + 0.129 + ], + "angle": 0, + "content": "48. Wang, H., Ren, J., Huang, Z., Olszewski, K., Chai, M., Fu, Y., Tulyakov, S.: R2l: Distilling neural radiance field to neural light field for efficient novel view synthesis. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXI. pp. 612-629. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.129, + 0.92, + 0.199 + ], + "angle": 0, + "content": "49. Wang, L., Zhang, J., Liu, X., Zhao, F., Zhang, Y., Zhang, Y., Wu, M., Yu, J., Xu, L.: Fourier plenoctrees for dynamic radiance field rendering in real-time. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13524-13534 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.199, + 0.918, + 0.251 + ], + "angle": 0, + "content": "50. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.251, + 0.918, + 0.285 + ], + "angle": 0, + "content": "51. Wang, Z., Li, L., Shen, Z., Shen, L., Bo, L.: 4k-nerf: High fidelity neural radiance fields at ultra high resolutions. arXiv preprint arXiv:2212.04701 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.285, + 0.918, + 0.355 + ], + "angle": 0, + "content": "52. Wang, Z., Shen, T., Gao, J., Huang, S., Munkberg, J., Hasselgren, J., Gojcic, Z., Chen, W., Fidler, S.: Neural fields meet explicit geometric representations for inverse rendering of urban scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8370-8380 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.354, + 0.918, + 0.425 + ], + "angle": 0, + "content": "53. Weng, C.Y., Curless, B., Srinivasan, P.P., Barron, J.T., Kemelmacher-Shlizerman, I.: Humannerf: Free-viewpoint rendering of moving people from monocular video. In: Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition. pp. 16210-16220 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.425, + 0.918, + 0.495 + ], + "angle": 0, + "content": "54. Worchel, M., Diaz, R., Hu, W., Schreer, O., Feldmann, I., Eisert, P.: Multiview mesh reconstruction with neural deferred shading. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6187-6197 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.495, + 0.918, + 0.566 + ], + "angle": 0, + "content": "55. Wu, L., Lee, J.Y., Bhattad, A., Wang, Y.X., Forsyth, D.: Diver: Real-time and accurate neural radiance fields with deterministic integration for volume rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16200-16209 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.566, + 0.918, + 0.618 + ], + "angle": 0, + "content": "56. Wu, T., Sun, J.M., Lai, Y.K., Gao, L.: De-nerf: Decoupled neural radiance fields for view-consistent appearance editing and high-frequency environmental relighting. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-11 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.618, + 0.918, + 0.669 + ], + "angle": 0, + "content": "57. Yariv, L., Hedman, P., Reiser, C., Verbin, D., Srinivasan, P.P., Szeliski, R., Barron, J.T., Mildenhall, B.: Bakedsdf: Meshing neural sdfs for real-time view synthesis. arXiv preprint arXiv:2302.14859 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.669, + 0.918, + 0.722 + ], + "angle": 0, + "content": "58. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: Plenoctrees for real-time rendering of neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5752-5761 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.722, + 0.918, + 0.774 + ], + "angle": 0, + "content": "59. Zhang, J., Yang, G., Tulsiani, S., Ramanan, D.: Ners: Neural reflectance surfaces for sparse-view 3d reconstruction in the wild. Advances in Neural Information Processing Systems 34, 29835-29847 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.774, + 0.918, + 0.826 + ], + "angle": 0, + "content": "60. Zhang, J., Yao, Y., Li, S., Liu, J., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neilf++: Inter-reflectable light fields for geometry and material estimation. arXiv preprint arXiv:2303.17147 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.826, + 0.918, + 0.879 + ], + "angle": 0, + "content": "61. Zhang, K., Luan, F., Li, Z., Snavely, N.: Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5565-5574 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.879, + 0.918, + 0.949 + ], + "angle": 0, + "content": "62. Zhang, K., Luan, F., Wang, Q., Bala, K., Snavely, N.: Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5453-5462 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.057, + 0.92, + 0.949 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.72, + 0.016, + 0.839, + 0.032 + ], + "angle": 0, + "content": "REFRAME" + }, + { + "type": "page_number", + "bbox": [ + 0.894, + 0.017, + 0.919, + 0.031 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.057, + 0.92, + 0.111 + ], + "angle": 0, + "content": "63. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG) 40(6), 1-18 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.112, + 0.92, + 0.164 + ], + "angle": 0, + "content": "64. Zhang, Y., Sun, J., He, X., Fu, H., Jia, R., Zhou, X.: Modeling indirect illumination for inverse rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18643-18652 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.165, + 0.918, + 0.2 + ], + "angle": 0, + "content": "65. Zhi, S., Laidlow, T., Leutenegger, S., Davison, A.J.: In-place scene labelling and understanding with implicit scene representation. In: ICCV (2021)" + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.057, + 0.92, + 0.2 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/REFRAME_ Reflective Surface Real-Time Rendering for Mobile Devices/f5175c1c-adca-4484-8e6b-cf9e011d3e14_origin.pdf b/2024/REFRAME_ Reflective Surface Real-Time Rendering for Mobile Devices/f5175c1c-adca-4484-8e6b-cf9e011d3e14_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d0c7f78dc58f401f6d7d78e05aa3f0b7703021f1 --- /dev/null +++ b/2024/REFRAME_ Reflective Surface Real-Time Rendering for Mobile Devices/f5175c1c-adca-4484-8e6b-cf9e011d3e14_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f8218cb86bb46b58148f3e5a1eb19968f7ce9d9b93f3a9297b31134ce9478da0 +size 9440954 diff --git a/2024/REFRAME_ Reflective Surface Real-Time Rendering for Mobile Devices/full.md b/2024/REFRAME_ Reflective Surface Real-Time Rendering for Mobile Devices/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2d8c6c34573a0b166c16f81ce9cd4adca9821c52 --- /dev/null +++ b/2024/REFRAME_ Reflective Surface Real-Time Rendering for Mobile Devices/full.md @@ -0,0 +1,424 @@ +# REFRAME: Reflective Surface Real-Time Rendering for Mobile Devices + +Chaojie Ji $^{1}$ , Yufeng Li $^{2}$ , and Yiyi Liao $^{1\boxtimes}$ + +1 Zhejiang University + +2 University of Electronic Science and Technology of China + +![](images/86e8b0d4d5bb40342a08760789063c0a8c8a19f5232a693dc46646ce57aef0a4.jpg) +Fig. 1: REFRAME enables real-time rendering on consumer GPUs and mobile devices, delivering superior subjective quality with a low face number compared to the baselines [9, 45]. Additionally, it effectively decouples the appearance properties and environmental information of the scene, which helps its capabilities for downstream scene editing tasks. + +Abstract. This work tackles the challenging task of achieving real-time novel view synthesis for reflective surfaces across various scenes. Existing real-time rendering methods, especially those based on meshes, often have subpar performance in modeling surfaces with rich view-dependent appearances. Our key idea lies in leveraging meshes for rendering acceleration while incorporating a novel approach to parameterize view-dependent information. We decompose the color into diffuse and specular, and model the specular color in the reflected direction based on a neural environment map. Our experiments demonstrate that our method achieves comparable reconstruction quality for highly reflective surfaces compared to state-of-the-art offline methods, while also efficiently enabling real-time rendering on edge devices such as smartphones. Our project page is at https://xdimlab.github.io/REFrame/. + +Keywords: Reflective surface $\cdot$ Real-time rendering $\cdot$ Mobile device + +# 1 Introduction + +Novel view synthesis (NVS) generates realistic images from novel viewpoints using multiple input views. While Neural Radiance Fields (NeRF) [33] excel at high-quality NVS through volume rendering, they struggle with modeling reflective appearance and lack real-time rendering capabilities. + +Several methods [19, 52, 59, 63] extend NeRF to decouple the intrinsic scene properties, e.g., into materials and lighting, and obtain color by the rendering equation [20]. Decoupling the environmental lighting and the physical parameters of the object often helps in modeling reflective objects. In contrast to recovering the precise physical meaning which may harm the visual quality due to the approximated rendering equation, another line of works [14, 46, 56] avoids full decomposition but enables representing reflective objects by modeling the reflective radiance. However, none of the aforementioned methods is capable of real-time rendering, especially on mobile devices, due to the expensive query of the rendering equation, or the underlying volume rendering formulation. + +There are many existing methods focusing on accelerating the rendering of NeRF. One line of works sticks with the volume rendering pipeline where the color of a pixel is accumulated along the ray. Acceleration is achieved by improving the sampling strategy [25, 27, 31, 48], tabulating the intermediate output [12, 16, 18, 34, 40, 49, 55, 58], or leveraging super-resolution neural rendering [28]. However, deploying these methods on edge devices such as smartphones is often greatly hindered due to the demanding computational power they typically require. Besides, tabulation-based methods require large memory consumption and greatly increase communication costs when transmitting data between the cloud server and the client. Another line of works [9, 39, 45] distills radiance fields into a mesh for real-time rendering, combined with a small MLP to model view-dependent effects. Mesh-based methods can leverage traditional graphics pipelines for acceleration, enabling them to achieve real-time rendering even on edge devices. Nonetheless, real-time rendering methods [9, 15, 16, 45, 58] often struggle to model objects with highly reflective surfaces. Besides, these mesh-based real-time rendering methods [9, 45, 57] typically require a large number of vertices and faces to achieve high fidelity. + +In this paper, we propose REFRAME, a mesh-based REFlective surface ReAl-time rendering method for MobiE devices (e.g., smartphones), see Fig. 1. We find the fact that existing NeRF distilled mesh rendering pipelines [9,45] do not perform well in rendering objects with highly reflective appearances mainly for two reasons. Firstly, these methods utilize the viewing direction to model the view-dependent appearance, approved to be less effective than using the reflection direction [46]. As such, our method adopts the reflection direction-based parameterization. Nevertheless, this parameterization requires surface normal, yet estimating precise vertex normal based on a mesh representation can be challenging. Therefore, we propose a geometry learner that learns vertex and normal offsets through two networks. This leads to good normal estimations to calculate the reflection direction while maintaining relatively low vertex and face numbers. Secondly and more importantly, to alleviate inference burden, real-time + +rendering methods often have limited capacity to model complex view-dependent information. In order to enhance the expressive power of the model without increasing the inference computational cost, we employ a four-layer MLP during training to learn the environmental features. This information is then baked into a two-dimensional environment feature map during inference. Remarkably, the distilled environment feature map incurs a memory overhead of less than 1MB and can be edited for relighting purposes. Finally, despite our method achieving real-time rendering with an advantage in mesh faces and vertices number compared to existing works, the reconstruction quality of our method is comparable to the current state-of-the-art (including non-real-time methods) work and even outperforms them in rendering highly reflective objects. We summarize our contributions as follows: + +- We propose a novel mesh geometry learner, allowing for robustly optimizing mesh vertex positions and normals. This leads to high rendering quality with relatively low mesh vertex and face numbers. +- We propose to use an environment feature map to model view-dependent appearances of highly reflective objects, which enhances the capacity to reconstruct complex reflective appearances without increasing the inference burden. This further enables relighting effects. +- Our rendering quality is on par with the current state-of-the-art methods while being able to achieve real-time rendering across various platform devices. Moreover, our method even surpasses the current non-real-time state-of-the-art approaches in rendering objects with highly reflective appearances. + +# 2 Related Work + +NeRF-based Scene Representation: NeRF [33] and its derivative works [1-3,41,44,46,53] have employed ray-marching based volume rendering methods to achieve high-quality and realistic rendering of different types of objects in various environments. Many follow-up works have extended NeRF in different aspects, including dynamic scene modeling [36-38], 3D-aware generation [7,8,42], and semantic scene understanding [11,24,65]. In this work, we focus on addressing two limitations of NeRF, i.e., modeling of reflective objects and real-time rendering. + +Reflectance Decomposition: A number of works [4, 5, 19, 43, 52, 59-64] investigate the task of inferring geometry and material properties based on neural fields representation, typically by formulating the image generation using the physically-based rendering. Despite achieving lighting and material control, these methods are typically inferior to state-of-the-art NeRF-based methods that directly model the radiance in terms of the rendering quality. This is due to the fact that these methods rely on simplified rendering equations of the real world. Another line of works achieves better rendering quality by avoiding full decomposition, yet allowing for modeling glossy objects by modeling the reflected radiance [14, 46, 56] or replacing the explicit rendering equation with a learned neural renderer [6, 30]. However, none of these methods are capable of real-time rendering due to the underlying volume rendering formulation. While + +![](images/306ca9e23f659e462f884c3b3bca90db97570b01176230887ebf42ad705e7cf9.jpg) + +![](images/80d0f593ad28999ffd556503fcaba6b9cea1c4a94ca30863167afa4c08c6314a.jpg) +Fig. 2: Pipeline for REFRAME. Components with a yellow background are either baked or omitted during the mobile rendering stage. Training Stage: The initial mesh is updated first before performing differentiable rasterization [26]. Next, we obtain the diffuse color $\mathbf{c}_d$ and specular feature $\mathbf{f}_s$ based on the position $\mathbf{p}$ , and the environment feature $\mathbf{f}_e$ based on reflective direction $\omega_r$ . Then we obtain the specular color $\mathbf{c}_s$ and combine it with $\mathbf{c}_d$ to create the final full color $\mathbf{c}$ . Mobile Rendering Stage: We bake the intermediate output for real-time rendering. This mesh-based rendering can be implemented using WebGL and easily deployed on various platforms (e.g., desktop and mobile devices). Here, we retrieve the $\mathbf{c}_d$ and $\mathbf{f}_s$ from baked texture images, and the $\mathbf{f}_e$ from the environment feature map. $\mathbf{c}_d$ and $\mathbf{c}_s$ are processed the same as the training stage to obtain $\mathbf{c}$ . + +NvDiffRec [35] enables real-time rendering based on the mesh representation, its quality is also limited by the full decomposition. In this work, we follow the partial decomposition pipeline [46] and propose to learn a neural environment map based on a NeRF-distilled mesh representation, with a focus on achieving real-time rendering. + +NeRF Acceleration: Accelerating NeRF rendering is another significant research area, with approaches reducing sampling points along the ray [25,31,48], tabulating the intermediate output [12, 16, 18, 49, 55, 58], using thousands of small MLPs to represent the scene [10, 40], or utilizing super-resolution techniques [28, 51]. However, these methods typically require consumer GPUs for real-time rendering, making it challenging to deploy them on edge devices with limited computational resources. + +More recently, a few approaches [9, 45, 57] propose to distill a NeRF representation into a mesh for real-time rendering, where MobileNeRF [9] and NeRF2Mesh [45] enable real-time rendering on mobile devices. However, MobileNeRF [9] does not model the accurate geometry of the scene, leading to expensive memory costs with a large number of vertices and faces. NeRF2Mesh [45] + +is the most relevant work to ours, which distills grid-based representation into a mesh. The good geometry prior for NeRF2Mesh [45] leads to a relatively low number of vertices and faces. However, none of the aforementioned methods are capable of modeling highly reflective objects faithfully due to their way of view-dependent color modeling. We propose to use a neural environment map to tackle this challenge by modeling the reflected radiance. Furthermore, we propose to estimate a high-quality vertex normal, further reducing the number of faces and vertices while maintaining the quality. + +# 3 Method + +The pipeline of our method is illustrated in Fig. 2. Before training, we employ a volumetric rendering technique to obtain an initial coarse mesh similar to existing methods [9,39,45,57]. During the training stage, we leverage a geometry learner to update both the vertices and the vertex normals (Sec. 3.1). We further learn a shader that decomposes diffuse and specular color (Sec. 3.2), where the specular branch is designed to be able to model highly reflective surfaces by combining view-independent specular features and reflective direction-conditioned environment features. After training with loss functions introduced in Sec. 3.3, we perform UV unwrapping and bake both view-independent and view-dependent features to enable real-time rendering on various devices (Sec. 3.4). + +More formally, let $\mathcal{M} = (\mathcal{V},\mathcal{F})$ denote the initial coarse mesh where $\mathcal{V} = \{\mathbf{v}\in \mathbb{R}^3\}$ is the set of vertices and $\mathcal{F}$ is the set of faces. We further compute initial vertex normals $\mathcal{N} = \{\mathbf{n}_{\mathbf{v}}\in \mathbb{R}^{3}\}$ given the initial mesh. Let $\mathbf{p}\in \mathbb{R}^3$ denote a rendered surface point on the mesh, $\mathbf{n_p}\in \mathbb{R}^3$ the corresponding normal, $\omega_{o}\in \mathbb{S}^{2}$ the viewing direction, and $\omega_{r}\in \mathbb{S}^{2}$ the reflective direction. During training, our goal is to refine the mesh $\mathcal{M}$ , as well as map $(\mathbf{p},\boldsymbol {\omega}_r)$ to a color $\mathbf{c}\in \mathbb{R}^3$ to enable mesh-based real-time rendering. + +# 3.1 Geometry Learner + +Before training, we follow NeRF2Mesh [45] to extract the initial coarse mesh through grid-based representation [34]. This allows for better modeling of depth discontinuities in geometry [50] and prevents the optimization of mesh geometry from getting stuck in local optima. Then, we propose to leverage a geometry learner to refine the geometry of the initial coarse mesh, including refining the vertex positions and normals. + +Vertex Offset: As shown in Fig. 3, the initial mesh extracted from NeRF-based methods may be of low quality. Existing methods [45,54] directly update the vertex positions through gradient backpropagation, which means the vertex update step size is determined by the learning rate and the gradient backpropagated. Thus, costly and difficult manual learning rate tuning is needed to adapt update step sizes for different objects, with a lack of robustness as shown in our experiments. To address this, our method uses a hybrid representation to provide a + +![](images/b7361106c718e803c2afd8e8371c4e14a56504137a5862d6c26bfdb793b12a71.jpg) +(a) Initial normal. + +![](images/fa9c7a556c3c0802ffed3d4c5d6b0dbd2708fc116317f226a5f120869eb38d98.jpg) +(b) Optimized normal. +Fig. 3: Mesh Optimization. The geometry of the initial mesh is often poor. We are able to significantly improve the geometry of the mesh through geometry learner. + +![](images/c104953e624039d4ecddada899efba62b798bcae50c5268d9cb037dba0217a43.jpg) +(c) Initial mesh. + +![](images/a057b44de21f36a9b2c7d4d076ac7072ff68d7a040bbd87bad86d02c7a12fba4.jpg) +(d) Optimized mesh. + +learned, adaptive step size for different vertices and different objects: + +$$ +\mathbf {v} ^ {\prime} = \mathbf {v} + \Delta \mathbf {v}, \quad \Delta \mathbf {v} = g _ {\theta_ {\mathbf {v}}} (\mathbf {v}) \tag {1} +$$ + +where $\mathbf{v}'$ is the updated vertex, $g_{\theta_{\mathbf{v}}}$ is a multi-resolution hash encoding in combined with a small MLP [34]. Though [47] also uses MLP for vertex offset learning, our coarse-to-fine hybrid representation allows us to leverage local information as well as global information, leading to better modeling of the geometry. + +Normal Offset: Our method relies on surface normals to estimate the reflection direction $\omega_r$ to model view-dependent appearance. Following the classical idea of smooth shading, we leverage smoothly interpolated vertex normals to approximate our surface normal. Obtaining accurate vertex normals is yet challenging, especially when the number of faces is limited. Hence, we learn a per-vertex normal offset by taking $\mathbf{v}$ and $\mathbf{n_v}$ as input: + +$$ +\mathbf {n} _ {\mathbf {v}} ^ {\prime} = \mathbf {n} _ {\mathbf {v}} + \Delta \mathbf {n} _ {\mathbf {v}}, \quad \Delta \mathbf {n} _ {\mathbf {v}} = g _ {\theta_ {\mathbf {n}}} (\mathbf {v}, \mathbf {n} _ {\mathbf {v}}) \tag {2} +$$ + +where $\mathbf{n}_{\mathbf{v}}^{\prime}$ is the learned per-vertex normal and $g_{\theta_{\mathbf{n}}}$ is another multi-resolution hash encoding-based network. In the experimental section, we demonstrate that learning such a per-vertex normal allows for modeling more accurate surface normals without increasing the number of vertices and faces. Note that $g_{\theta_{\mathbf{v}}}$ and $g_{\theta_{\mathbf{n}}}$ are only used during training, as we can directly save the updated vertices $\mathbf{v}^{\prime}$ and normals $\mathbf{n}_{\mathbf{v}}^{\prime}$ for real-time rendering. + +# 3.2 Color Formulation + +We decompose the final color $\mathbf{c}$ into a diffuse color $\mathbf{c}_d$ and a specular one $\mathbf{c}_s$ : + +$$ +\mathbf {c} = \min \left(\max \left(\mathbf {c} _ {d} + \mathbf {c} _ {s}, 0\right), 1\right) \tag {3} +$$ + +where $\mathbf{c}_d + \mathbf{c}_s$ is clamped to [0, 1], both individual terms $\mathbf{c}_d$ and $\mathbf{c}_s$ from sigmoid are in [0, 1]. We now elaborate on the diffuse and specular color formulation. + +Diffuse Color Formulation: With the updated mesh $\mathcal{M}' = (\mathcal{V}',\mathcal{N}',\mathcal{F})$ , we perform differentiable rasterization [26] to obtain the position $\mathbf{p}$ corresponding + +to each pixel. Subsequently, $\mathbf{p}$ is mapped to a diffuse color $\mathbf{c}_d\in \mathbb{R}^3$ and a view-independent feature $\mathbf{f}_s\in \mathbb{R}^3$ which will be used later for decoding specular color: + +$$ +\mathbf {c} _ {d}, \mathbf {f} _ {s} = f _ {\phi} (\mathbf {p}) \tag {4} +$$ + +While the number of channels of $\mathbf{f}_s$ is flexible, we observe that such a compact three-channel representation is sufficient for our purpose. + +Specular Color Formulation: Recent non-real-time methods [14, 46, 56] propose to model view-dependent color based on the reflective direction $\omega_{r}$ , instead of the viewing direction used by NeRF. This approach has been shown to effectively model reflective surfaces. Motivated by these methods, we calculate the reflective direction as follows: + +$$ +\boldsymbol {\omega} _ {r} = 2 \left(\boldsymbol {\omega} _ {o} \cdot \mathbf {n} _ {\mathbf {p}}\right) \mathbf {n} _ {\mathbf {p}} - \boldsymbol {\omega} _ {o} \tag {5} +$$ + +where $\mathbf{n}_{\mathbf{p}}$ is smoothly interpolated surface normal at $\mathbf{p}$ based on our learned vertex normal $\mathbf{n}_{\mathbf{v}}^{\prime}$ . Next, the specular color can be obtained as follows: + +$$ +\mathbf {c} _ {s} = f _ {\psi} \left(\mathbf {f} _ {s}, \boldsymbol {\omega} _ {r}, \left(\boldsymbol {\omega} _ {o} \cdot \mathbf {n} _ {\mathbf {p}}\right)\right) \tag {6} +$$ + +where $(\boldsymbol{\omega}_o \cdot \mathbf{n_p})$ is considered as input to model the Fresnel effects [21] and $f_{\psi}$ needs to be a tiny MLP to enable real-time rendering. + +Environment Learner: In practice, we observe that naively mapping the reflective viewing direction to specular color (see Eq. (6)) leads to unsatisfying performance due to the capacity limitation of the tiny MLP $f_{\psi}$ . For a similar reason, Ref-NeRF adopts an 8-layer MLP to map $\mathbf{f}_s$ and $\omega_r$ to the view-dependent color. However, increasing the capacity of $f_{\psi}$ hampers the real-time rendering capability. Therefore, instead of increasing the capacity of $f_{\psi}$ , we propose to leverage a neural environment learner of large capacity that maps $\omega_r$ to an environment feature $\mathbf{f}_e \in \mathbb{R}^M$ ( $M = 3$ in our paper): + +$$ +\mathbf {f} _ {e} = f _ {e} \left(\gamma \left(\boldsymbol {\omega} _ {r}\right)\right) \tag {7} +$$ + +where $\gamma (\cdot)$ denotes positional encoding and $f_{e}$ is a 4-layer MLP. This leads to our final implementation of the specular color: + +$$ +\mathbf {c} _ {s} = f _ {\psi} \left(\mathbf {f} _ {s}, \mathbf {f} _ {e}, \left(\boldsymbol {\omega} _ {o} \cdot \mathbf {n} _ {\mathbf {p}}\right)\right) \tag {8} +$$ + +Note that $\mathbf{f}_e$ can be baked into a two-dimensional neural environment feature map for real-time rendering as it solely depends on $\omega_r$ . + +Scene Editing: REFRAME can decouple geometry, diffuse and specular color, so we can perform simple scene editing tasks. For example, we can edit the geometry and appearance of objects by modifying the mesh or texture image. Additionally, we can perform relighting on the objects by modifying the environment feature map of the scene, as shown in Fig. 4. + +![](images/cfe3c5e600f39b784c1622cf7f94a1dd20ce647c1468c766b90186b4a8cf59fb.jpg) +(a) Original image. + +![](images/500ef3ad23a8817ec90a1806390a84d599c4541f992dc732bb37e699148f6d5f.jpg) +(b) Flip lighting. +Fig. 4: Relighting. We can achieve relighting of objects by editing our environment feature map. The image illustrates the editing process of flipping or replacing the environmental feature map. The left side shows the rendering result, the top right corner presents the corresponding environment feature map, and the specular color is displayed in the bottom right corner. + +![](images/943a0becffb385cdcf6bb0c86af562dba7afd1e8c12602e24405de41cb1af613.jpg) +(c) Change lighting. + +# 3.3 Loss Functions + +We train our model based on image reconstruction loss and several regularization terms. At each iteration, we render a full image and train with the corresponding ground truth. Let $\hat{\mathbf{c}}$ denote the ground truth RGB value at one pixel. We optionally use ground truth binary mask for supervision at synthetic object-centric scenes. Our full loss function is shown in Eq. (9): + +$$ +\begin{array}{l} \mathcal {L} = \lambda_ {\mathbf {c}} \mathcal {L} _ {\mathbf {c}} + \lambda_ {\mathbf {c} _ {d}} \mathcal {L} _ {\mathbf {c} _ {d}} + \lambda_ {S S I M} \mathcal {L} _ {S S I M} + \tag {9} \\ \lambda_ {m a s k} \mathcal {L} _ {m a s k} + \lambda_ {\Delta \mathbf {n} _ {\mathbf {v}}} \mathcal {L} _ {\Delta \mathbf {n} _ {\mathbf {v}}} + \lambda_ {m a x} \mathcal {L} _ {m a x} \\ \end{array} +$$ + +Image Reconstruction Loss: Our model is first supervised by the L2 loss between the final rendered color and the ground truth. + +$$ +\mathcal {L} _ {\mathbf {c}} = \left| \left| \mathbf {c} - \hat {\mathbf {c}} \right| \right| _ {2} \tag {10} +$$ + +To encourage the correct decoupling of diffuse and specular colors, we add a diffuse loss to guide the diffuse color to be close to the ground truth color: + +$$ +\mathcal {L} _ {\mathbf {c} _ {d}} = \left\| \mathbf {c} _ {d} - \hat {\mathbf {c}} \right\| _ {2} \tag {11} +$$ + +We also introduce SSIM [17] loss for better perceptual quality. On object-centric datasets, we additionally utilize mask loss to ensure the rasterization mask closely aligns with the ground truth mask. Note that the mask loss is optional and is not required on unbounded datasets. + +Regularizations: We regularize the predicted vertex normal $\mathbf{n}_{\mathbf{v}}^{\prime}$ to be close to the original vertex normal $\mathbf{n}_{\mathbf{v}}$ computed from the mesh. This allows for stabilizing the optimization of the vertex normals, e.g., avoid completely flipping the normal. This is achieved by regularizing the magnitude of the normal offset: + +$$ +\mathcal {L} _ {\Delta \mathbf {n} _ {\mathbf {v}}} = \left\| \Delta \mathbf {n} _ {\mathbf {v}} \right\| _ {1} \tag {12} +$$ + +We introduce another regularization to encourage the specular color $\mathbf{c}_s$ to be reasonable when visualized alone. As shown in Eq. (3), the final color $\mathbf{c}$ will be + +clamped when the sum of specular color $\mathbf{c}_s$ and diffuse color $\mathbf{c}_d$ exceeds 1. In this case, the gradient is truncated due to the clamp operation, preventing the backpropagation from updating the color values. This may leave an undesired large specular color $\mathbf{c}_s$ with no penalization. Therefore, we encourage the sum of specular color and diffuse color to not exceed 1: + +$$ +\mathcal {L} _ {\max } = \left| \left| \operatorname {m a x} \left(\mathbf {c} _ {d} + \mathbf {c} _ {s} - 1, 0\right) \right| \right| \tag {13} +$$ + +# 3.4 Real-Time Rendering + +Environment Feature Map: During the training process, we employ a four-layer MLP with a width of 256 as our environment learner. If we query $\mathbf{f}_e$ through the environment learner module during the inference process, it would significantly slow down the inference speed. Hence, during the inference phase, we bake the learned environment feature $\mathbf{f}_e$ into a 2D environment feature map by converting $\omega_r \in \mathbb{S}^2$ to its polar coordinate, see supplementary for more details. This allows us to simply query the corresponding feature on the environment feature map based on $\omega_r$ , thereby greatly accelerating our inference speed, reaching above 200 FPS on a single NVIDIA 3090 GPU. + +Texture Images: To further reduce the inference burden and better deploy our method across different platforms, we can also bake the $\mathbf{c}_d$ and $\mathbf{f}_s$ obtained from querying the $f_{\phi}$ into two texture images. Specifically, we map the vertices of the mesh to UV coordinates and then bake the features into two texture images, respectively. In order to obtain the normal $\mathbf{n_p}$ from the learned normal vector $\mathbf{n_v}'$ during real-time rendering, we also bake the normal $\mathbf{n_p}$ into a texture image. + +# 4 Experiment + +Implementation Details. Training: Prior to the training stage, we employ the classic quadric error metrics algorithm [13] to simplify the initial mesh to 75,000 faces except for the outdoor scenes which typically have complex geometry. We train for 250 epochs on each scene, with a training time of approximately 2 hours. We utilize the Adam optimizer [23] and employ the cosine annealing learning rate adjustment strategy [32] during the training stage. Rendering: The resolution of our baked texture image is $4096 \times 4096$ , and the resolution of the baked environment feature map is $720 \times 360$ . As referenced in Sec. 3.4, once we have baked the environment learner into the environment feature map, our approach is already capable of achieving above 200 FPS on the GPU (referred to as Ours). This implementation ensures a fair comparison with the baselines since their results quoted are evaluated using GPU implementation that can not directly deploy on mobile devices. Given that we use texture images for rendering when implemented on mobile devices, we also employ texture images on the GPU for quality evaluation (referred to as Ours (Mobile)). Following NeRF2Mesh [45] and MobileNeRF [9], we employ anti-aliasing to enhance our rendering quality. For details on the implementation of anti-aliasing, please refer to the supplementary material. All experiments are performed on a single NVIDIA 3090 GPU. + +
NeRF Synthetic DatasetShiny Blender DatasetReal Captured Dataset
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NeRF [33]31.010.9470.081------
Ref-NeRF [46]33.990.9660.03835.960.9670.05824.450.6650.142
3DGS [22]33.300.9690.03030.370.9470.08324.060.6610.259
NvDiffRec [35]29.050.9390.08129.050.9380.111---
MobileNeRF [9]30.900.9470.06226.620.8830.16322.960.4930.430
NeRF2Mesh [45]29.670.9400.07226.960.8960.17022.710.5230.419
Ours31.040.9610.04036.020.9830.04022.500.5980.325
Ours (Mobile)30.840.9590.04235.830.9810.04122.110.5240.423
+ +Table 1: Rendering Quality. Baseline comparisons of the rendering quality on three different datasets. Red represents the optimal, orange represents the second best, and yellow represents the third. + +Dataset. We validate the effectiveness and robustness of our method using three different datasets with varied scenes. 1) NeRF [33] Synthetic Dataset: This dataset consists of eight synthetic scenes. 2) Shiny Blender Dataset: This dataset is introduced by Ref-NeRF [46] and includes six glossy objects, serving as the primary dataset for assessing the reflective surface modeling capabilities of the methods. 3) Real Captured Dataset: We select three outdoor scenes with rich reflective appearances same as Ref-NeRF [46] to validate the effectiveness of our method in real-world outdoor capture scenarios. Due to the poor quality of the initial mesh extracted by Nerf2Mesh [45] when modeling glossy objects, we utilize the meshes extracted by Neuralangelo [29] and Ref-NeuS [14] as our initialization in certain scenes, see supplementary for more details. + +Baseline. We compare with various types of baselines, including Ref-NeRF [46], 3DGS [22], NvDiffRec [35], MobileNeRF [9], and NeRF2Mesh [45] in our study. Ref-NeRF [46] serves as the state-of-the-art volume rendering method for handling the strong reflective appearance. 3DGS [22] represents a powerful new representation capable of real-time rendering on GPU. NvDiffRec [35] represents physically based rendering methods. MobileNeRF [9], and NeRF2Mesh [45] are representative works in the field of real-time rendering for mobile devices. We import the results of the baselines from their original paper and replicate the results that missing in their original paper. + +# 4.1 Reconstruction Quality + +Rendering Quality. We compare the rendering quality of our method and baseline methods on three datasets, see Tab. 1. It is worth noting that Ref-NeRF [46] is the state-of-the-art method for modeling glossy objects but cannot achieve real-time rendering, even on high-end GPU. We introduce Ref-NeRF [46] as a benchmark for rendering quality and demonstrate that our method achieves comparable or even superior rendering quality to non-real-time state-of-the-art methods. While 3DGS [22] can render in real-time on GPU and has higher PSNR on both NeRF Synthetic and Real Captured dataset than ours, it's important + +![](images/e0cf1ca3563fe8025eda30447c9cdf9f16799b4892de70dfd0f30c6adb46474b.jpg) +C +NvDiffRec + +![](images/3ece761960d140f3cfa8d73a0bd4b325a80098ac5873a14900bb4dcdf8e282cc.jpg) +Helmert + +![](images/0909144c648b932e13735cbc07ccaccd15639a4c851cbee9a295b69e6f30f129.jpg) +Toaster +Fig. 5: Rendering Quality on Shiny Blender Dataset. Our method achieves optimal rendering quality in most scenes and provides better modeling of reflective appearance compared to the comparison methods [9,35,45]. + +![](images/0673bde51653b9fdc6a5d1beac2d9586278bb6729d3131a96343e34cdd65484a.jpg) +MobileNeRF + +![](images/d25fdc1caeeb7c21d3a652c1adf243a58530933dac34c4bcf0d7f483c8c3aefb.jpg) + +![](images/f552b8380a342be290b9a561c31bc861cab62f06830d8159b233d1ea8d277deb.jpg) + +![](images/8e6819ca04d5ea4f4a2b01b1ed1d22a424cbf3f13a285108a1c67dd4fdfde17e.jpg) +NeRF2Mesh + +![](images/aad6df0036970970f779dfd8e995f7aa01b8ff9f429a7ca59f84a57790051d85.jpg) + +![](images/de997674c608a635f5e20925a181b8d930ffad02b559d08eac8dde5c0c614ddb.jpg) + +![](images/4c8588d3154606b6943e65d590971c47f48feedfcaf03e8eaa5f7b9ea35ab660.jpg) + +![](images/7506e90e827bae975ed0e2a2636a205723833c4487a6a2b65e7645beb7d027f6.jpg) + +![](images/e12e0bc85bb5acefa4083d97ccbb56b04b60708963238e366021520feb597572.jpg) + +![](images/54a21d313b747c0c8016573d54fd77360827ddde6286cc47a38fa3fe7598d5ed.jpg) +Ours + +![](images/1ba0963fb7223b5653a2893729329263342398f342de0d927a70a5744322d986.jpg) + +![](images/fe60be153fd9497be5a011fbf710fa3479823f4beb553f8431de0616c9e325fb.jpg) + +![](images/a3a3f28d0a4ccef1a37a731ad61cf8c106f8e177c0fcbdc714e45cf4f4fe7873.jpg) + +![](images/bc36c9703fadb2381fb1bb0d7a1e4266c978b00e1af3cad0823386b2d3fe1dc2.jpg) + +![](images/1be4200de4ea6c17e870d6863f7880617d427e16e2c652b7ce1e4f91d8ffcdc6.jpg) + +![](images/4fc146c66927e528bc3e876fb7431f467163179cf4c51d7a5b608a95525f493b.jpg) + +![](images/cc4c784cf2f9a14fa0c87b80e2b7651af25b75bfc9486290a6edda5d04056d98.jpg) + +![](images/72105cb9fdaf593c430f38330d6f26912ec829f2c9a837c161f5af898488df09.jpg) +Ground Truth + +![](images/a1383c552b532b3c43356e2bdc786b6ce4539fec7c167e1fbf2dfbba400d0400.jpg) + +to note that 3DGS struggles to model reflective surfaces effectively (Fig. 6a). Moreover, they have a larger memory overhead than us. $E.g.$ , for gardenspheres, the memory cost for 3DGS is 1.4GB, whereas our method only requires 84.3MB. + +In comparison to mesh-based real-time rendering methods [9, 35, 45], we achieve optimal rendering quality in most object-centric datasets, as shown in Fig. 5. Despite having mask supervision on object-centric datasets as well, NvDiffRec [35] and NeRF2Mesh [45] struggle to achieve high rendering quality. This could be due to the fact that NvDiffRec relies on the simplified rendering equation, whereas NeRF2Mesh's simple color formulation struggles to model highly reflective objects. Note that NeRF2Mesh also requires an initial mesh in its second stage, and we provide NeRF2Mesh with the same initial mesh as ours for all experiments for a fair comparison. + +In unbounded outdoor scenes without masks, our method still achieves high-quality rendering, as demonstrated in Fig. 6a. Although our PSNR may be slightly lower compared to the baselines, we outperform them in rendering foreground glossy objects. However, this advantage is not accurately reflected in the PSNR metric since the foreground reflective objects occupy only a small region in the image. Further, our method mainly exhibits artifacts in the background, this is due to the fact that the initial mesh can only faithfully reconstruct the foreground but not the background, see supplementary for more details. + +Surface Quality. Furthermore, we observe that our method not only improves the rendering quality when modeling glossy objects but also generates more accurate meshes, as shown in Fig. 6b. It can be seen that the meshes obtained from [35,45] often have poor quality in the reflective region while our method is capable of accurately reconstructing the geometry. + +![](images/36f40a5b6f218bf1d38b78034539f0f0c398d13ca834fa2291a5fd91a3214f60.jpg) + +![](images/511a37a008c361b3463a77a4d8274a538ffbb663e1449712d42357d100e19e52.jpg) +(a) Rendering Quality on Real Captured Dataset. + +![](images/d45d5777842ddfa6612b8b778c0eb5e2af6f2d2e14eb86bf9b75a6179fb31831.jpg) + +![](images/9c1804bdc99372cd289fe61aa7eb59e794efc80bc410df0ce3d1f33bb151e881.jpg) +Fig. 6: Qualitative Comparison. Compared to the baselines, our method reconstructs reflective regions of higher fidelity and yields more accurate surface geometry. + +![](images/6044dffec518b45ba37cafcae18efeb5ddd1affdad15f7433d851340ad457730.jpg) + +![](images/2e870c184d25a65cabd834a6c84235a932f94a5cf52043fa99727a5e41f94aa5.jpg) + +![](images/18d7cbe5d651a4cf241340fc313b9e74d0849b3f6c0bbe68350ac691fd91207f.jpg) +(b) Surface Quality on NeRF Synthetic Dataset. + +![](images/a5b217db35b9b2a8ca247afcc6e0c815c2109a9813c6a9bb30964b91b62d1255.jpg) + +![](images/8642dd161fa89993d20b0b3e40f8424b44fa14f621eea6d57be3e387662a8df4.jpg) + +
MacBook Pro iPhone12 iPad Air3 Legion Y-7000 HUAWEI nova7
MobileNeRF [9]120.0060.0060.0055.3139.19
NeRF2Mesh [45]120.0060.0060.0048.6932.56
Ours (Mobile)120.0060.0060.0048.5031.71
+ +Table 2: Computational Efficiency Comparison. We test the FPS of our method and the baseline methods [9,45] on the NeRF Synthetic dataset across multiple devices. On MacBook Pro, iPhone 12, and iPad Air3, both our method and the baseline methods [9,45] achieve the maximum screen refresh rate. + +# 4.2 Rendering Efficiency + +Computational efficiency. During the mobile rendering stage, the bottleneck of our rendering speed lies in querying the tiny MLP $f_{\psi}$ to obtain the specular color. However, due to our MLP being only two layers with 64 widths, our rendering speed remains relatively fast. Therefore, our method achieves real-time rendering on both consumer GPUs and mobile devices. We compare the rendering speed with the baselines [9,45] that can render in real-time on mobile devices across multiple platform devices, as seen in Tab. 2. + +Memory efficiency. Our cache mainly consists of texture images, the environment feature map, and the mesh. Due to our high-quality estimation of vertex normals, our mesh can achieve high-quality rendering with a low number of vertices and faces, leading to a small cache size in most scenes, as shown in Tab. 3. + +# 4.3 Ablation Study + +Geometry Learner. [54] and [45] utilize direct gradient backpropagation to update the vertex positions of the mesh. We employ this update strategy in our method and verify that this update method lacks robustness. Directly using the backpropagated gradients to update the vertex positions can lead to a devastating blow to the mesh, especially when using a large learning rate. This + +
NeRF Synthetic DatasetShiny Blender Dataset
#V (103)#F (103)Cache#V (103)#F (103)Cache
MobileNeRF [9]494224125.751028343152.37
NeRF2Mesh [45]20019273.53459122.15
Ours (Mobile)377547.55387552.86
+ +Table 3: Memory Efficiency. While exhibiting higher quality, our method maintains a relatively low memory consumption compared to baselines. Cache reported in MB. + +
w/o gθvOurs
L=0.001L=0.0001L=0.001
Toaster17.0117.0124.95
Materials23.0129.0629.58
+ +Table 4: Ablation of Vertex Offset Learning. L represents the learning rate. Reported in PSNR. + +
w/o Δn_vOurs
Toaster#F=10k24.2424.62
#F=75k24.5524.95
Materials#F=10k27.6928.23
#F=75k29.4629.58
+ +Table 5: Ablation of Normal Offset Learning. Reported in PSNR. + +significantly compromises the rendering quality, as shown in Tab. 4. Additionally, different datasets require different learning rates, as indicated in Tab. 4. While a learning rate of 0.0001 works fine for the materials dataset, it performs poorly on the toaster dataset. In contrast, our vertex update strategy not only exhibits robustness across different datasets but also provides higher rendering quality. + +In order to verify the necessity of estimating an offset for the vertices normal, we conduct experiments with the w/o normal offset setting. In this setting, we do not estimate a normal offset for each vertex, but we still compute an updated normal based on the updated vertex positions. As shown in Tab. 5, introducing per-vertex normal offset provides a gain in reconstruction quality. This gain is particularly significant when the number of faces is low. Notably, in the case of the toaster dataset, a mesh with $10\mathrm{k}$ faces with normal offset even achieves rendering quality surpassing that of a mesh with $75\mathrm{k}$ faces without normal offset. Environment Learner. To validate the effectiveness of the environment learner $f_{e}$ , we directly input $\omega_{r}$ values into the tiny MLP $f_{\psi}$ as shown in Eq. (6), similar to [9] and [45], except they use viewing direction instead of reflective direction. Experimental results demonstrate that directly inputting $\omega_{r}$ into the tiny MLP $f_{\psi}$ cannot learn the correct reflection appearance as a result of the limited network's capacity, as depicted in Fig. 7. Therefore, introducing the environment learner module improves the reconstruction quality without compromising the inference speed, as demonstrated in Tab. 6. + +Environment Feature Map: We can directly optimize an environment feature map instead of learning it through the environment learner module and baking it. However, as shown in Tab. 7, the direct optimization of the environment feature map performs poorly when the map resolution is high. This is because, at higher resolutions, neighboring directions are mapped to distant grid points, leading to a non-smooth interpolation. As a result, the grid points can- + +![](images/9d790918db044f75c38db9020eb8d8e1b448e8a4729afbb55dadedb6283348da.jpg) +(a) $\mathrm{W / o}f_{e}$ + +![](images/04f9ad22e7b71099b85a84668bd4296fbb264a3853a342505b9988461b02dfe0.jpg) +(b) Ours +Fig. 7: Ablation of Environment Learner. Eliminating $f_{e}$ results in bad rendering quality. + +![](images/e6bbdcc3e5b8faa4aafb69e096b2951c29ebd5478be512171be14075874aacb9.jpg) +(c) Ground Truth + +
w/o feOurs
Toaster19.3124.95
Materials27.3729.58
+ +Table 6: Ablation of Environment Learner. Reported in PSNR. + +
Resolutions H*WOursDirectly optimization
PSNRCache(KB)PSNRCache(KB)
Toaster180*36024.864924.1691
900*180024.9650423.592460
Materials180*36029.556329.58114
900*180029.5870629.033000
+ +Table 7: Ablation of Environment Feature Map. Our method outperforms in both cache size and rendering quality. + +not share global information and only contain its own local information. On the other hand, when the resolution is too low, the rendering quality is limited by the capacity of the environment feature map. In contrast, our method exhibits good rendering quality across different resolutions by learning the environment feature map via an MLP. Additionally, our environment feature maps enjoy the smoothness bias of the MLP, resulting in relatively smooth tensor maps. Consequently, when saving images in PNG format, our method demonstrates good compression performance. This further reduces the overhead of our caching. + +# 5 Conclusion + +Although our method successfully models the appearance of glossy objects, similar to Ref-NeRF [46], it still struggles with accurately modeling interreflections and non-distant illumination. Additionally, our method requires a certain level of quality in the initial mesh. + +In summary, our method achieves real-time rendering on edge devices with low hardware budgets. Besides, we propose a novel approach for modeling view-dependent appearance and can optimize the appearance and geometry of glossy objects with high computational efficiency and low memory footprint. Furthermore, by effectively decoupling the scene's geometry, appearance, and environmental information, our method can perform simple scene editing tasks. + +# Acknowledgements + +This work is supported by NSFC under grant 62202418 and U21B2004. We thank Xiaoran Cao, Chenxi Tu, and Sicheng Li for their valuable discussions. + +# References + +1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021) +2. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mipnerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022) +3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-nerf: Anti-aliased grid-based neural radiance fields. arXiv preprint arXiv:2304.06706 (2023) +4. Bi, S., Xu, Z., Srinivasan, P., Mildenhall, B., Sunkavalli, K., Hasan, M., Hold-Geoffroy, Y., Kriegman, D., Ramamoorthi, R.: Neural reflectance fields for appearance acquisition. arXiv preprint arXiv:2008.03824 (2020) +5. Boss, M., Braun, R., Jampani, V., Barron, J.T., Liu, C., Lensch, H.: Nerd: Neural reflectance decomposition from image collections. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12684-12694 (2021) +6. Boss, M., Jampani, V., Braun, R., Liu, C., Barron, J., Lensch, H.: Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems 34, 10691-10704 (2021) +7. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022) +8. Chan, E.R., Monteiro, M., Kellnhofer, P., Wu, J., Wetzstein, G.: pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5799-5809 (2021) +9. Chen, Z., Funkhouser, T., Hedman, P., Tagliasacchi, A.: Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16569-16578 (2023) +10. Esposito, S., Baieri, D., Zellmann, S., Hinkenjann, A., Rodola, E.: Kiloneus: A versatile neural implicit surface representation for real-time rendering. arXiv preprint arXiv:2206.10885 (2022) +11. Fu, X., Zhang, S., Chen, T., Lu, Y., Zhu, L., Zhou, X., Geiger, A., Liao, Y.: Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation. In: 2022 International Conference on 3D Vision (3DV). pp. 1-11. IEEE (2022) +12. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: Fastnerf: High-fidelity neural rendering at 200fps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14346-14355 (2021) +13. Garland, M., Heckbert, P.S.: Surface simplification using quadric error metrics. In: Proceedings of the 24th annual conference on Computer graphics and interactive techniques. pp. 209-216 (1997) +14. Ge, W., Hu, T., Zhao, H., Liu, S., Chen, Y.C.: Ref-neus: Ambiguity-reduced neural implicit surface learning for multi-view reconstruction with reflection. arXiv preprint arXiv:2303.10840 (2023) + +15. Guo, Y.C., Cao, Y.P., Wang, C., He, Y., Shan, Y., Zhang, S.H.: Vmesh: Hybrid volume-mesh representation for efficient view synthesis. In: SIGGRAPH Asia 2023 Conference Papers. pp. 1-11 (2023) +16. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking neural radiance fields for real-time view synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5875-5884 (2021) +17. Hore, A., Ziou, D.: Image quality metrics: Psnr vs. sSIM. In: 2010 20th international conference on pattern recognition. pp. 2366-2369. IEEE (2010) +18. Hu, T., Liu, S., Chen, Y., Shen, T., Jia, J.: Efficientnerf efficient neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12902-12911 (2022) +19. Jin, H., Liu, I., Xu, P., Zhang, X., Han, S., Bi, S., Zhou, X., Xu, Z., Su, H.: Tensoroir: Tensorial inverse rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 165-174 (2023) +20. Kajiya, J.T.: The rendering equation. In: Proceedings of the 13th annual conference on Computer graphics and interactive techniques. pp. 143-150 (1986) +21. Kautz, J., McCool, M.D.: Approximation of glossy reflection with prefiltered environment maps. In: Graphics Interface. vol. 2000, pp. 119-126 (2000) +22. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023) +23. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) +24. Kundu, A., Genova, K., Yin, X., Fathi, A., Pantofaru, C., Guibas, L.J., Tagliasacchi, A., Dellaert, F., Funkhouser, T.: Panoptic neural fields: A semantic object-aware neural scene representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12871-12881 (2022) +25. Kurz, A., Neff, T., Lv, Z., Zollhöfer, M., Steinberger, M.: Adanerf: Adaptive sampling for real-time rendering of neural radiance fields. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII. pp. 254-270. Springer (2022) +26. Laine, S., Hellsten, J., Karras, T., Seol, Y., Lehtinen, J., Aila, T.: Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG) 39(6), 1-14 (2020) +27. Li, C., Li, S., Zhao, Y., Zhu, W., Lin, Y.: Rt-nerf: Real-time on-device neural radiance fields towards immersive ar/vr rendering. In: Proceedings of the 41st IEEE/ACM International Conference on Computer-Aided Design. pp. 1-9 (2022) +28. Li, S., Li, H., Wang, Y., Liao, Y., Yu, L.: Steernerf: Accelerating nef rendering via smooth viewpoint trajectory. arXiv preprint arXiv:2212.08476 (2022) +29. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8456-8465 (2023) +30. Liang, R., Chen, H., Li, C., Chen, F., Panneer, S., Vijaykumar, N.: Envidr: Implicit differentiable renderer with neural environment lighting. arXiv preprint arXiv:2303.13022 (2023) +31. Lin, H., Peng, S., Xu, Z., Yan, Y., Shuai, Q., Bao, H., Zhou, X.: Efficient neural radiance fields for interactive free-viewpoint video. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022) +32. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016) + +33. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) +34. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022) +35. Munkberg, J., Hasselgren, J., Shen, T., Gao, J., Chen, W., Evans, A., Müller, T., Fidler, S.: Extracting triangular 3d models, materials, and lighting from images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8280-8290 (2022) +36. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5865-5874 (2021) +37. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., MartinBrualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228 (2021) +38. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10318-10327 (2021) +39. Rakotosaona, M.J., Manhardt, F., Arroyo, D.M., Niemeyer, M., Kundu, A., Tombari, F.: Nerfmeshing: Distilling neural radiance fields into geometrically-accurate 3d meshes. arXiv preprint arXiv:2303.09431 (2023) +40. Reiser, C., Peng, S., Liao, Y., Geiger, A.: Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14335-14345 (2021) +41. Rematas, K., Liu, A., Srinivasan, P.P., Barron, J.T., Tagliasacchi, A., Funkhouser, T., Ferrari, V.: Urban radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12932-12942 (2022) +42. Schwarz, K., Liao, Y., Niemeyer, M., Geiger, A.: Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems 33, 20154-20166 (2020) +43. Srinivasan, P.P., Deng, B., Zhang, X., Tancik, M., Mildenhall, B., Barron, J.T.: Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7495-7504 (June 2021) +44. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8248-8258 (2022) +45. Tang, J., Zhou, H., Chen, X., Hu, T., Ding, E., Wang, J., Zeng, G.: Delicate textured mesh recovery from nef via adaptive surface refinement. arXiv preprint arXiv:2303.02091 (2023) +46. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5481-5490. IEEE (2022) +47. Walker, T., Mariotti, O., Vaxman, A., Bilen, H.: Explicit neural surfaces: Learning continuous geometry with deformation fields. arXiv preprint arXiv:2306.02956 (2023) + +48. Wang, H., Ren, J., Huang, Z., Olszewski, K., Chai, M., Fu, Y., Tulyakov, S.: R2l: Distilling neural radiance field to neural light field for efficient novel view synthesis. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXI. pp. 612-629. Springer (2022) +49. Wang, L., Zhang, J., Liu, X., Zhao, F., Zhang, Y., Zhang, Y., Wu, M., Yu, J., Xu, L.: Fourier plenoctrees for dynamic radiance field rendering in real-time. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13524-13534 (2022) +50. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689 (2021) +51. Wang, Z., Li, L., Shen, Z., Shen, L., Bo, L.: 4k-nerf: High fidelity neural radiance fields at ultra high resolutions. arXiv preprint arXiv:2212.04701 (2022) +52. Wang, Z., Shen, T., Gao, J., Huang, S., Munkberg, J., Hasselgren, J., Gojcic, Z., Chen, W., Fidler, S.: Neural fields meet explicit geometric representations for inverse rendering of urban scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8370-8380 (2023) +53. Weng, C.Y., Curless, B., Srinivasan, P.P., Barron, J.T., Kemelmacher-Shlizerman, I.: Humannerf: Free-viewpoint rendering of moving people from monocular video. In: Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition. pp. 16210-16220 (2022) +54. Worchel, M., Diaz, R., Hu, W., Schreer, O., Feldmann, I., Eisert, P.: Multiview mesh reconstruction with neural deferred shading. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6187-6197 (2022) +55. Wu, L., Lee, J.Y., Bhattad, A., Wang, Y.X., Forsyth, D.: Diver: Real-time and accurate neural radiance fields with deterministic integration for volume rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16200-16209 (2022) +56. Wu, T., Sun, J.M., Lai, Y.K., Gao, L.: De-nerf: Decoupled neural radiance fields for view-consistent appearance editing and high-frequency environmental relighting. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-11 (2023) +57. Yariv, L., Hedman, P., Reiser, C., Verbin, D., Srinivasan, P.P., Szeliski, R., Barron, J.T., Mildenhall, B.: Bakedsdf: Meshing neural sdfs for real-time view synthesis. arXiv preprint arXiv:2302.14859 (2023) +58. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: Plenoctrees for real-time rendering of neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5752-5761 (2021) +59. Zhang, J., Yang, G., Tulsiani, S., Ramanan, D.: Ners: Neural reflectance surfaces for sparse-view 3d reconstruction in the wild. Advances in Neural Information Processing Systems 34, 29835-29847 (2021) +60. Zhang, J., Yao, Y., Li, S., Liu, J., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neilf++: Inter-reflectable light fields for geometry and material estimation. arXiv preprint arXiv:2303.17147 (2023) +61. Zhang, K., Luan, F., Li, Z., Snavely, N.: Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5565-5574 (2022) +62. Zhang, K., Luan, F., Wang, Q., Bala, K., Snavely, N.: Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5453-5462 (2021) + +63. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG) 40(6), 1-18 (2021) +64. Zhang, Y., Sun, J., He, X., Fu, H., Jia, R., Zhou, X.: Modeling indirect illumination for inverse rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18643-18652 (June 2022) +65. Zhi, S., Laidlow, T., Leutenegger, S., Davison, A.J.: In-place scene labelling and understanding with implicit scene representation. In: ICCV (2021) \ No newline at end of file diff --git a/2024/REFRAME_ Reflective Surface Real-Time Rendering for Mobile Devices/images.zip b/2024/REFRAME_ Reflective Surface Real-Time Rendering for Mobile Devices/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4c3b7427919b06a509d5948dcd5609e8cc943c42 --- /dev/null +++ b/2024/REFRAME_ Reflective Surface Real-Time Rendering for Mobile Devices/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f88811b535e705df3b1db241f46941e79f1a6077783b015e8cfdf34d796f4821 +size 527334 diff --git a/2024/REFRAME_ Reflective Surface Real-Time Rendering for Mobile Devices/layout.json b/2024/REFRAME_ Reflective Surface Real-Time Rendering for Mobile Devices/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5c60b6dc2d32a3f2c12d8a09cba64b04cc876d54 --- /dev/null +++ b/2024/REFRAME_ Reflective Surface Real-Time Rendering for Mobile Devices/layout.json @@ -0,0 +1,12234 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 30, + 352, + 65 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 30, + 352, + 65 + ], + "spans": [ + { + "bbox": [ + 59, + 30, + 352, + 65 + ], + "type": "text", + "content": "REFRAME: Reflective Surface Real-Time Rendering for Mobile Devices" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 87, + 296, + 100 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 87, + 296, + 100 + ], + "spans": [ + { + "bbox": [ + 115, + 87, + 296, + 100 + ], + "type": "text", + "content": "Chaojie Ji" + }, + { + "bbox": [ + 115, + 87, + 296, + 100 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 115, + 87, + 296, + 100 + ], + "type": "text", + "content": ", Yufeng Li" + }, + { + "bbox": [ + 115, + 87, + 296, + 100 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 115, + 87, + 296, + 100 + ], + "type": "text", + "content": ", and Yiyi Liao" + }, + { + "bbox": [ + 115, + 87, + 296, + 100 + ], + "type": "inline_equation", + "content": "^{1\\boxtimes}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 161, + 109, + 251, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 109, + 251, + 121 + ], + "spans": [ + { + "bbox": [ + 161, + 109, + 251, + 121 + ], + "type": "text", + "content": "1 Zhejiang University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 121, + 326, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 121, + 326, + 132 + ], + "spans": [ + { + "bbox": [ + 85, + 121, + 326, + 132 + ], + "type": "text", + "content": "2 University of Electronic Science and Technology of China" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 33, + 158, + 378, + 306 + ], + "blocks": [ + { + "bbox": [ + 33, + 158, + 378, + 306 + ], + "lines": [ + { + "bbox": [ + 33, + 158, + 378, + 306 + ], + "spans": [ + { + "bbox": [ + 33, + 158, + 378, + 306 + ], + "type": "image", + "image_path": "86e8b0d4d5bb40342a08760789063c0a8c8a19f5232a693dc46646ce57aef0a4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 30, + 314, + 379, + 369 + ], + "lines": [ + { + "bbox": [ + 30, + 314, + 379, + 369 + ], + "spans": [ + { + "bbox": [ + 30, + 314, + 379, + 369 + ], + "type": "text", + "content": "Fig. 1: REFRAME enables real-time rendering on consumer GPUs and mobile devices, delivering superior subjective quality with a low face number compared to the baselines [9, 45]. Additionally, it effectively decouples the appearance properties and environmental information of the scene, which helps its capabilities for downstream scene editing tasks." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 395, + 351, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 395, + 351, + 539 + ], + "spans": [ + { + "bbox": [ + 59, + 395, + 351, + 539 + ], + "type": "text", + "content": "Abstract. This work tackles the challenging task of achieving real-time novel view synthesis for reflective surfaces across various scenes. Existing real-time rendering methods, especially those based on meshes, often have subpar performance in modeling surfaces with rich view-dependent appearances. Our key idea lies in leveraging meshes for rendering acceleration while incorporating a novel approach to parameterize view-dependent information. We decompose the color into diffuse and specular, and model the specular color in the reflected direction based on a neural environment map. Our experiments demonstrate that our method achieves comparable reconstruction quality for highly reflective surfaces compared to state-of-the-art offline methods, while also efficiently enabling real-time rendering on edge devices such as smartphones. Our project page is at https://xdimlab.github.io/REFrame/." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 550, + 339, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 550, + 339, + 562 + ], + "spans": [ + { + "bbox": [ + 59, + 550, + 339, + 562 + ], + "type": "text", + "content": "Keywords: Reflective surface " + }, + { + "bbox": [ + 59, + 550, + 339, + 562 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 59, + 550, + 339, + 562 + ], + "type": "text", + "content": " Real-time rendering " + }, + { + "bbox": [ + 59, + 550, + 339, + 562 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 59, + 550, + 339, + 562 + ], + "type": "text", + "content": " Mobile device" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 41, + 571, + 143, + 583 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 571, + 143, + 583 + ], + "spans": [ + { + "bbox": [ + 41, + 571, + 143, + 583 + ], + "type": "inline_equation", + "content": "\\mathbb{O}" + }, + { + "bbox": [ + 41, + 571, + 143, + 583 + ], + "type": "text", + "content": " Corresponding author." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 33, + 33, + 127, + 46 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 33, + 127, + 46 + ], + "spans": [ + { + "bbox": [ + 33, + 33, + 127, + 46 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 33, + 57, + 379, + 105 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 57, + 379, + 105 + ], + "spans": [ + { + "bbox": [ + 33, + 57, + 379, + 105 + ], + "type": "text", + "content": "Novel view synthesis (NVS) generates realistic images from novel viewpoints using multiple input views. While Neural Radiance Fields (NeRF) [33] excel at high-quality NVS through volume rendering, they struggle with modeling reflective appearance and lack real-time rendering capabilities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 105, + 379, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 105, + 379, + 224 + ], + "spans": [ + { + "bbox": [ + 33, + 105, + 379, + 224 + ], + "type": "text", + "content": "Several methods [19, 52, 59, 63] extend NeRF to decouple the intrinsic scene properties, e.g., into materials and lighting, and obtain color by the rendering equation [20]. Decoupling the environmental lighting and the physical parameters of the object often helps in modeling reflective objects. In contrast to recovering the precise physical meaning which may harm the visual quality due to the approximated rendering equation, another line of works [14, 46, 56] avoids full decomposition but enables representing reflective objects by modeling the reflective radiance. However, none of the aforementioned methods is capable of real-time rendering, especially on mobile devices, due to the expensive query of the rendering equation, or the underlying volume rendering formulation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 225, + 379, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 225, + 379, + 427 + ], + "spans": [ + { + "bbox": [ + 33, + 225, + 379, + 427 + ], + "type": "text", + "content": "There are many existing methods focusing on accelerating the rendering of NeRF. One line of works sticks with the volume rendering pipeline where the color of a pixel is accumulated along the ray. Acceleration is achieved by improving the sampling strategy [25, 27, 31, 48], tabulating the intermediate output [12, 16, 18, 34, 40, 49, 55, 58], or leveraging super-resolution neural rendering [28]. However, deploying these methods on edge devices such as smartphones is often greatly hindered due to the demanding computational power they typically require. Besides, tabulation-based methods require large memory consumption and greatly increase communication costs when transmitting data between the cloud server and the client. Another line of works [9, 39, 45] distills radiance fields into a mesh for real-time rendering, combined with a small MLP to model view-dependent effects. Mesh-based methods can leverage traditional graphics pipelines for acceleration, enabling them to achieve real-time rendering even on edge devices. Nonetheless, real-time rendering methods [9, 15, 16, 45, 58] often struggle to model objects with highly reflective surfaces. Besides, these mesh-based real-time rendering methods [9, 45, 57] typically require a large number of vertices and faces to achieve high fidelity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 428, + 379, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 428, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 428, + 379, + 583 + ], + "type": "text", + "content": "In this paper, we propose REFRAME, a mesh-based REFlective surface ReAl-time rendering method for MobiE devices (e.g., smartphones), see Fig. 1. We find the fact that existing NeRF distilled mesh rendering pipelines [9,45] do not perform well in rendering objects with highly reflective appearances mainly for two reasons. Firstly, these methods utilize the viewing direction to model the view-dependent appearance, approved to be less effective than using the reflection direction [46]. As such, our method adopts the reflection direction-based parameterization. Nevertheless, this parameterization requires surface normal, yet estimating precise vertex normal based on a mesh representation can be challenging. Therefore, we propose a geometry learner that learns vertex and normal offsets through two networks. This leads to good normal estimations to calculate the reflection direction while maintaining relatively low vertex and face numbers. Secondly and more importantly, to alleviate inference burden, real-time" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 11, + 39, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 11, + 39, + 18 + ], + "spans": [ + { + "bbox": [ + 33, + 11, + 39, + 18 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 9, + 106, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 9, + 106, + 19 + ], + "spans": [ + { + "bbox": [ + 66, + 9, + 106, + 19 + ], + "type": "text", + "content": "C.Ji et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 30, + 35, + 381, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 35, + 381, + 178 + ], + "spans": [ + { + "bbox": [ + 30, + 35, + 381, + 178 + ], + "type": "text", + "content": "rendering methods often have limited capacity to model complex view-dependent information. In order to enhance the expressive power of the model without increasing the inference computational cost, we employ a four-layer MLP during training to learn the environmental features. This information is then baked into a two-dimensional environment feature map during inference. Remarkably, the distilled environment feature map incurs a memory overhead of less than 1MB and can be edited for relighting purposes. Finally, despite our method achieving real-time rendering with an advantage in mesh faces and vertices number compared to existing works, the reconstruction quality of our method is comparable to the current state-of-the-art (including non-real-time methods) work and even outperforms them in rendering highly reflective objects. We summarize our contributions as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 38, + 184, + 379, + 316 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 38, + 184, + 379, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 184, + 379, + 220 + ], + "spans": [ + { + "bbox": [ + 38, + 184, + 379, + 220 + ], + "type": "text", + "content": "- We propose a novel mesh geometry learner, allowing for robustly optimizing mesh vertex positions and normals. This leads to high rendering quality with relatively low mesh vertex and face numbers." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 38, + 220, + 379, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 220, + 379, + 267 + ], + "spans": [ + { + "bbox": [ + 38, + 220, + 379, + 267 + ], + "type": "text", + "content": "- We propose to use an environment feature map to model view-dependent appearances of highly reflective objects, which enhances the capacity to reconstruct complex reflective appearances without increasing the inference burden. This further enables relighting effects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 38, + 268, + 379, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 268, + 379, + 316 + ], + "spans": [ + { + "bbox": [ + 38, + 268, + 379, + 316 + ], + "type": "text", + "content": "- Our rendering quality is on par with the current state-of-the-art methods while being able to achieve real-time rendering across various platform devices. Moreover, our method even surpasses the current non-real-time state-of-the-art approaches in rendering objects with highly reflective appearances." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 31, + 332, + 137, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 332, + 137, + 346 + ], + "spans": [ + { + "bbox": [ + 31, + 332, + 137, + 346 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 30, + 356, + 380, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 356, + 380, + 439 + ], + "spans": [ + { + "bbox": [ + 30, + 356, + 380, + 439 + ], + "type": "text", + "content": "NeRF-based Scene Representation: NeRF [33] and its derivative works [1-3,41,44,46,53] have employed ray-marching based volume rendering methods to achieve high-quality and realistic rendering of different types of objects in various environments. Many follow-up works have extended NeRF in different aspects, including dynamic scene modeling [36-38], 3D-aware generation [7,8,42], and semantic scene understanding [11,24,65]. In this work, we focus on addressing two limitations of NeRF, i.e., modeling of reflective objects and real-time rendering." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 30, + 440, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 440, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 440, + 380, + 584 + ], + "type": "text", + "content": "Reflectance Decomposition: A number of works [4, 5, 19, 43, 52, 59-64] investigate the task of inferring geometry and material properties based on neural fields representation, typically by formulating the image generation using the physically-based rendering. Despite achieving lighting and material control, these methods are typically inferior to state-of-the-art NeRF-based methods that directly model the radiance in terms of the rendering quality. This is due to the fact that these methods rely on simplified rendering equations of the real world. Another line of works achieves better rendering quality by avoiding full decomposition, yet allowing for modeling glossy objects by modeling the reflected radiance [14, 46, 56] or replacing the explicit rendering equation with a learned neural renderer [6, 30]. However, none of these methods are capable of real-time rendering due to the underlying volume rendering formulation. While" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 9, + 346, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 9, + 346, + 19 + ], + "spans": [ + { + "bbox": [ + 297, + 9, + 346, + 19 + ], + "type": "text", + "content": "REFRAME" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 33, + 33, + 379, + 159 + ], + "blocks": [ + { + "bbox": [ + 33, + 33, + 379, + 159 + ], + "lines": [ + { + "bbox": [ + 33, + 33, + 379, + 159 + ], + "spans": [ + { + "bbox": [ + 33, + 33, + 379, + 159 + ], + "type": "image", + "image_path": "306ca9e23f659e462f884c3b3bca90db97570b01176230887ebf42ad705e7cf9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 33, + 161, + 379, + 223 + ], + "blocks": [ + { + "bbox": [ + 33, + 161, + 379, + 223 + ], + "lines": [ + { + "bbox": [ + 33, + 161, + 379, + 223 + ], + "spans": [ + { + "bbox": [ + 33, + 161, + 379, + 223 + ], + "type": "image", + "image_path": "80d0f593ad28999ffd556503fcaba6b9cea1c4a94ca30863167afa4c08c6314a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 31, + 233, + 380, + 353 + ], + "lines": [ + { + "bbox": [ + 31, + 233, + 380, + 353 + ], + "spans": [ + { + "bbox": [ + 31, + 233, + 380, + 353 + ], + "type": "text", + "content": "Fig. 2: Pipeline for REFRAME. Components with a yellow background are either baked or omitted during the mobile rendering stage. Training Stage: The initial mesh is updated first before performing differentiable rasterization [26]. Next, we obtain the diffuse color " + }, + { + "bbox": [ + 31, + 233, + 380, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_d" + }, + { + "bbox": [ + 31, + 233, + 380, + 353 + ], + "type": "text", + "content": " and specular feature " + }, + { + "bbox": [ + 31, + 233, + 380, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_s" + }, + { + "bbox": [ + 31, + 233, + 380, + 353 + ], + "type": "text", + "content": " based on the position " + }, + { + "bbox": [ + 31, + 233, + 380, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 31, + 233, + 380, + 353 + ], + "type": "text", + "content": ", and the environment feature " + }, + { + "bbox": [ + 31, + 233, + 380, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_e" + }, + { + "bbox": [ + 31, + 233, + 380, + 353 + ], + "type": "text", + "content": " based on reflective direction " + }, + { + "bbox": [ + 31, + 233, + 380, + 353 + ], + "type": "inline_equation", + "content": "\\omega_r" + }, + { + "bbox": [ + 31, + 233, + 380, + 353 + ], + "type": "text", + "content": ". Then we obtain the specular color " + }, + { + "bbox": [ + 31, + 233, + 380, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_s" + }, + { + "bbox": [ + 31, + 233, + 380, + 353 + ], + "type": "text", + "content": " and combine it with " + }, + { + "bbox": [ + 31, + 233, + 380, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_d" + }, + { + "bbox": [ + 31, + 233, + 380, + 353 + ], + "type": "text", + "content": " to create the final full color " + }, + { + "bbox": [ + 31, + 233, + 380, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 31, + 233, + 380, + 353 + ], + "type": "text", + "content": ". Mobile Rendering Stage: We bake the intermediate output for real-time rendering. This mesh-based rendering can be implemented using WebGL and easily deployed on various platforms (e.g., desktop and mobile devices). Here, we retrieve the " + }, + { + "bbox": [ + 31, + 233, + 380, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_d" + }, + { + "bbox": [ + 31, + 233, + 380, + 353 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 31, + 233, + 380, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_s" + }, + { + "bbox": [ + 31, + 233, + 380, + 353 + ], + "type": "text", + "content": " from baked texture images, and the " + }, + { + "bbox": [ + 31, + 233, + 380, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_e" + }, + { + "bbox": [ + 31, + 233, + 380, + 353 + ], + "type": "text", + "content": " from the environment feature map. " + }, + { + "bbox": [ + 31, + 233, + 380, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_d" + }, + { + "bbox": [ + 31, + 233, + 380, + 353 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 31, + 233, + 380, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_s" + }, + { + "bbox": [ + 31, + 233, + 380, + 353 + ], + "type": "text", + "content": " are processed the same as the training stage to obtain " + }, + { + "bbox": [ + 31, + 233, + 380, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 31, + 233, + 380, + 353 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 30, + 378, + 379, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 378, + 379, + 438 + ], + "spans": [ + { + "bbox": [ + 30, + 378, + 379, + 438 + ], + "type": "text", + "content": "NvDiffRec [35] enables real-time rendering based on the mesh representation, its quality is also limited by the full decomposition. In this work, we follow the partial decomposition pipeline [46] and propose to learn a neural environment map based on a NeRF-distilled mesh representation, with a focus on achieving real-time rendering." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 439, + 380, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 439, + 380, + 522 + ], + "spans": [ + { + "bbox": [ + 30, + 439, + 380, + 522 + ], + "type": "text", + "content": "NeRF Acceleration: Accelerating NeRF rendering is another significant research area, with approaches reducing sampling points along the ray [25,31,48], tabulating the intermediate output [12, 16, 18, 49, 55, 58], using thousands of small MLPs to represent the scene [10, 40], or utilizing super-resolution techniques [28, 51]. However, these methods typically require consumer GPUs for real-time rendering, making it challenging to deploy them on edge devices with limited computational resources." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 30, + 523, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 523, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 523, + 380, + 584 + ], + "type": "text", + "content": "More recently, a few approaches [9, 45, 57] propose to distill a NeRF representation into a mesh for real-time rendering, where MobileNeRF [9] and NeRF2Mesh [45] enable real-time rendering on mobile devices. However, MobileNeRF [9] does not model the accurate geometry of the scene, leading to expensive memory costs with a large number of vertices and faces. NeRF2Mesh [45]" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 11, + 39, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 11, + 39, + 18 + ], + "spans": [ + { + "bbox": [ + 33, + 11, + 39, + 18 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 9, + 107, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 9, + 107, + 19 + ], + "spans": [ + { + "bbox": [ + 66, + 9, + 107, + 19 + ], + "type": "text", + "content": "C.Ji et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 33, + 381, + 130 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 33, + 381, + 130 + ], + "spans": [ + { + "bbox": [ + 32, + 33, + 381, + 130 + ], + "type": "text", + "content": "is the most relevant work to ours, which distills grid-based representation into a mesh. The good geometry prior for NeRF2Mesh [45] leads to a relatively low number of vertices and faces. However, none of the aforementioned methods are capable of modeling highly reflective objects faithfully due to their way of view-dependent color modeling. We propose to use a neural environment map to tackle this challenge by modeling the reflected radiance. Furthermore, we propose to estimate a high-quality vertex normal, further reducing the number of faces and vertices while maintaining the quality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 33, + 151, + 99, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 151, + 99, + 164 + ], + "spans": [ + { + "bbox": [ + 33, + 151, + 99, + 164 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 32, + 179, + 379, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 179, + 379, + 298 + ], + "spans": [ + { + "bbox": [ + 32, + 179, + 379, + 298 + ], + "type": "text", + "content": "The pipeline of our method is illustrated in Fig. 2. Before training, we employ a volumetric rendering technique to obtain an initial coarse mesh similar to existing methods [9,39,45,57]. During the training stage, we leverage a geometry learner to update both the vertices and the vertex normals (Sec. 3.1). We further learn a shader that decomposes diffuse and specular color (Sec. 3.2), where the specular branch is designed to be able to model highly reflective surfaces by combining view-independent specular features and reflective direction-conditioned environment features. After training with loss functions introduced in Sec. 3.3, we perform UV unwrapping and bake both view-independent and view-dependent features to enable real-time rendering on various devices (Sec. 3.4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 32, + 300, + 379, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 300, + 379, + 383 + ], + "spans": [ + { + "bbox": [ + 32, + 300, + 379, + 383 + ], + "type": "text", + "content": "More formally, let " + }, + { + "bbox": [ + 32, + 300, + 379, + 383 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = (\\mathcal{V},\\mathcal{F})" + }, + { + "bbox": [ + 32, + 300, + 379, + 383 + ], + "type": "text", + "content": " denote the initial coarse mesh where " + }, + { + "bbox": [ + 32, + 300, + 379, + 383 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{\\mathbf{v}\\in \\mathbb{R}^3\\}" + }, + { + "bbox": [ + 32, + 300, + 379, + 383 + ], + "type": "text", + "content": " is the set of vertices and " + }, + { + "bbox": [ + 32, + 300, + 379, + 383 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 32, + 300, + 379, + 383 + ], + "type": "text", + "content": " is the set of faces. We further compute initial vertex normals " + }, + { + "bbox": [ + 32, + 300, + 379, + 383 + ], + "type": "inline_equation", + "content": "\\mathcal{N} = \\{\\mathbf{n}_{\\mathbf{v}}\\in \\mathbb{R}^{3}\\}" + }, + { + "bbox": [ + 32, + 300, + 379, + 383 + ], + "type": "text", + "content": " given the initial mesh. Let " + }, + { + "bbox": [ + 32, + 300, + 379, + 383 + ], + "type": "inline_equation", + "content": "\\mathbf{p}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 32, + 300, + 379, + 383 + ], + "type": "text", + "content": " denote a rendered surface point on the mesh, " + }, + { + "bbox": [ + 32, + 300, + 379, + 383 + ], + "type": "inline_equation", + "content": "\\mathbf{n_p}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 32, + 300, + 379, + 383 + ], + "type": "text", + "content": " the corresponding normal, " + }, + { + "bbox": [ + 32, + 300, + 379, + 383 + ], + "type": "inline_equation", + "content": "\\omega_{o}\\in \\mathbb{S}^{2}" + }, + { + "bbox": [ + 32, + 300, + 379, + 383 + ], + "type": "text", + "content": " the viewing direction, and " + }, + { + "bbox": [ + 32, + 300, + 379, + 383 + ], + "type": "inline_equation", + "content": "\\omega_{r}\\in \\mathbb{S}^{2}" + }, + { + "bbox": [ + 32, + 300, + 379, + 383 + ], + "type": "text", + "content": " the reflective direction. During training, our goal is to refine the mesh " + }, + { + "bbox": [ + 32, + 300, + 379, + 383 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 32, + 300, + 379, + 383 + ], + "type": "text", + "content": ", as well as map " + }, + { + "bbox": [ + 32, + 300, + 379, + 383 + ], + "type": "inline_equation", + "content": "(\\mathbf{p},\\boldsymbol {\\omega}_r)" + }, + { + "bbox": [ + 32, + 300, + 379, + 383 + ], + "type": "text", + "content": " to a color " + }, + { + "bbox": [ + 32, + 300, + 379, + 383 + ], + "type": "inline_equation", + "content": "\\mathbf{c}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 32, + 300, + 379, + 383 + ], + "type": "text", + "content": " to enable mesh-based real-time rendering." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 404, + 151, + 416 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 404, + 151, + 416 + ], + "spans": [ + { + "bbox": [ + 33, + 404, + 151, + 416 + ], + "type": "text", + "content": "3.1 Geometry Learner" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 32, + 426, + 379, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 426, + 379, + 498 + ], + "spans": [ + { + "bbox": [ + 32, + 426, + 379, + 498 + ], + "type": "text", + "content": "Before training, we follow NeRF2Mesh [45] to extract the initial coarse mesh through grid-based representation [34]. This allows for better modeling of depth discontinuities in geometry [50] and prevents the optimization of mesh geometry from getting stuck in local optima. Then, we propose to leverage a geometry learner to refine the geometry of the initial coarse mesh, including refining the vertex positions and normals." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 32, + 499, + 379, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 499, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 32, + 499, + 379, + 583 + ], + "type": "text", + "content": "Vertex Offset: As shown in Fig. 3, the initial mesh extracted from NeRF-based methods may be of low quality. Existing methods [45,54] directly update the vertex positions through gradient backpropagation, which means the vertex update step size is determined by the learning rate and the gradient backpropagated. Thus, costly and difficult manual learning rate tuning is needed to adapt update step sizes for different objects, with a lack of robustness as shown in our experiments. To address this, our method uses a hybrid representation to provide a" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 9, + 346, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 9, + 346, + 19 + ], + "spans": [ + { + "bbox": [ + 297, + 9, + 346, + 19 + ], + "type": "text", + "content": "REFRAME" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 37, + 35, + 123, + 103 + ], + "blocks": [ + { + "bbox": [ + 37, + 35, + 123, + 103 + ], + "lines": [ + { + "bbox": [ + 37, + 35, + 123, + 103 + ], + "spans": [ + { + "bbox": [ + 37, + 35, + 123, + 103 + ], + "type": "image", + "image_path": "b7361106c718e803c2afd8e8371c4e14a56504137a5862d6c26bfdb793b12a71.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 105, + 112, + 115 + ], + "lines": [ + { + "bbox": [ + 47, + 105, + 112, + 115 + ], + "spans": [ + { + "bbox": [ + 47, + 105, + 112, + 115 + ], + "type": "text", + "content": "(a) Initial normal." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 131, + 36, + 215, + 103 + ], + "blocks": [ + { + "bbox": [ + 131, + 36, + 215, + 103 + ], + "lines": [ + { + "bbox": [ + 131, + 36, + 215, + 103 + ], + "spans": [ + { + "bbox": [ + 131, + 36, + 215, + 103 + ], + "type": "image", + "image_path": "fa9c7a556c3c0802ffed3d4c5d6b0dbd2708fc116317f226a5f120869eb38d98.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 105, + 213, + 115 + ], + "lines": [ + { + "bbox": [ + 132, + 105, + 213, + 115 + ], + "spans": [ + { + "bbox": [ + 132, + 105, + 213, + 115 + ], + "type": "text", + "content": "(b) Optimized normal." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 30, + 124, + 380, + 145 + ], + "lines": [ + { + "bbox": [ + 30, + 124, + 380, + 145 + ], + "spans": [ + { + "bbox": [ + 30, + 124, + 380, + 145 + ], + "type": "text", + "content": "Fig. 3: Mesh Optimization. The geometry of the initial mesh is often poor. We are able to significantly improve the geometry of the mesh through geometry learner." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 222, + 32, + 296, + 102 + ], + "blocks": [ + { + "bbox": [ + 222, + 32, + 296, + 102 + ], + "lines": [ + { + "bbox": [ + 222, + 32, + 296, + 102 + ], + "spans": [ + { + "bbox": [ + 222, + 32, + 296, + 102 + ], + "type": "image", + "image_path": "c104953e624039d4ecddada899efba62b798bcae50c5268d9cb037dba0217a43.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 105, + 288, + 115 + ], + "lines": [ + { + "bbox": [ + 230, + 105, + 288, + 115 + ], + "spans": [ + { + "bbox": [ + 230, + 105, + 288, + 115 + ], + "type": "text", + "content": "(c) Initial mesh." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 301, + 32, + 376, + 102 + ], + "blocks": [ + { + "bbox": [ + 301, + 32, + 376, + 102 + ], + "lines": [ + { + "bbox": [ + 301, + 32, + 376, + 102 + ], + "spans": [ + { + "bbox": [ + 301, + 32, + 376, + 102 + ], + "type": "image", + "image_path": "a057b44de21f36a9b2c7d4d076ac7072ff68d7a040bbd87bad86d02c7a12fba4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 301, + 105, + 376, + 115 + ], + "lines": [ + { + "bbox": [ + 301, + 105, + 376, + 115 + ], + "spans": [ + { + "bbox": [ + 301, + 105, + 376, + 115 + ], + "type": "text", + "content": "(d) Optimized mesh." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 30, + 171, + 335, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 171, + 335, + 183 + ], + "spans": [ + { + "bbox": [ + 30, + 171, + 335, + 183 + ], + "type": "text", + "content": "learned, adaptive step size for different vertices and different objects:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 194, + 379, + 208 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 194, + 379, + 208 + ], + "spans": [ + { + "bbox": [ + 141, + 194, + 379, + 208 + ], + "type": "interline_equation", + "content": "\\mathbf {v} ^ {\\prime} = \\mathbf {v} + \\Delta \\mathbf {v}, \\quad \\Delta \\mathbf {v} = g _ {\\theta_ {\\mathbf {v}}} (\\mathbf {v}) \\tag {1}", + "image_path": "e4072abee3a12b0061be179ed46facbc2e6d7262eb63034933d38f0e442a779d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 30, + 217, + 380, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 217, + 380, + 265 + ], + "spans": [ + { + "bbox": [ + 30, + 217, + 380, + 265 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 30, + 217, + 380, + 265 + ], + "type": "inline_equation", + "content": "\\mathbf{v}'" + }, + { + "bbox": [ + 30, + 217, + 380, + 265 + ], + "type": "text", + "content": " is the updated vertex, " + }, + { + "bbox": [ + 30, + 217, + 380, + 265 + ], + "type": "inline_equation", + "content": "g_{\\theta_{\\mathbf{v}}}" + }, + { + "bbox": [ + 30, + 217, + 380, + 265 + ], + "type": "text", + "content": " is a multi-resolution hash encoding in combined with a small MLP [34]. Though [47] also uses MLP for vertex offset learning, our coarse-to-fine hybrid representation allows us to leverage local information as well as global information, leading to better modeling of the geometry." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 30, + 266, + 379, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 266, + 379, + 338 + ], + "spans": [ + { + "bbox": [ + 30, + 266, + 379, + 338 + ], + "type": "text", + "content": "Normal Offset: Our method relies on surface normals to estimate the reflection direction " + }, + { + "bbox": [ + 30, + 266, + 379, + 338 + ], + "type": "inline_equation", + "content": "\\omega_r" + }, + { + "bbox": [ + 30, + 266, + 379, + 338 + ], + "type": "text", + "content": " to model view-dependent appearance. Following the classical idea of smooth shading, we leverage smoothly interpolated vertex normals to approximate our surface normal. Obtaining accurate vertex normals is yet challenging, especially when the number of faces is limited. Hence, we learn a per-vertex normal offset by taking " + }, + { + "bbox": [ + 30, + 266, + 379, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 30, + 266, + 379, + 338 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 266, + 379, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{n_v}" + }, + { + "bbox": [ + 30, + 266, + 379, + 338 + ], + "type": "text", + "content": " as input:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 123, + 352, + 379, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 352, + 379, + 365 + ], + "spans": [ + { + "bbox": [ + 123, + 352, + 379, + 365 + ], + "type": "interline_equation", + "content": "\\mathbf {n} _ {\\mathbf {v}} ^ {\\prime} = \\mathbf {n} _ {\\mathbf {v}} + \\Delta \\mathbf {n} _ {\\mathbf {v}}, \\quad \\Delta \\mathbf {n} _ {\\mathbf {v}} = g _ {\\theta_ {\\mathbf {n}}} (\\mathbf {v}, \\mathbf {n} _ {\\mathbf {v}}) \\tag {2}", + "image_path": "baf9ea714f10447e348b0caaf6a9d72915144afce3f813508c1e674088b44487.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 30, + 372, + 380, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 372, + 380, + 445 + ], + "spans": [ + { + "bbox": [ + 30, + 372, + 380, + 445 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 30, + 372, + 380, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_{\\mathbf{v}}^{\\prime}" + }, + { + "bbox": [ + 30, + 372, + 380, + 445 + ], + "type": "text", + "content": " is the learned per-vertex normal and " + }, + { + "bbox": [ + 30, + 372, + 380, + 445 + ], + "type": "inline_equation", + "content": "g_{\\theta_{\\mathbf{n}}}" + }, + { + "bbox": [ + 30, + 372, + 380, + 445 + ], + "type": "text", + "content": " is another multi-resolution hash encoding-based network. In the experimental section, we demonstrate that learning such a per-vertex normal allows for modeling more accurate surface normals without increasing the number of vertices and faces. Note that " + }, + { + "bbox": [ + 30, + 372, + 380, + 445 + ], + "type": "inline_equation", + "content": "g_{\\theta_{\\mathbf{v}}}" + }, + { + "bbox": [ + 30, + 372, + 380, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 372, + 380, + 445 + ], + "type": "inline_equation", + "content": "g_{\\theta_{\\mathbf{n}}}" + }, + { + "bbox": [ + 30, + 372, + 380, + 445 + ], + "type": "text", + "content": " are only used during training, as we can directly save the updated vertices " + }, + { + "bbox": [ + 30, + 372, + 380, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^{\\prime}" + }, + { + "bbox": [ + 30, + 372, + 380, + 445 + ], + "type": "text", + "content": " and normals " + }, + { + "bbox": [ + 30, + 372, + 380, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_{\\mathbf{v}}^{\\prime}" + }, + { + "bbox": [ + 30, + 372, + 380, + 445 + ], + "type": "text", + "content": " for real-time rendering." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 31, + 465, + 152, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 465, + 152, + 476 + ], + "spans": [ + { + "bbox": [ + 31, + 465, + 152, + 476 + ], + "type": "text", + "content": "3.2 Color Formulation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 31, + 488, + 367, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 488, + 367, + 500 + ], + "spans": [ + { + "bbox": [ + 31, + 488, + 367, + 500 + ], + "type": "text", + "content": "We decompose the final color " + }, + { + "bbox": [ + 31, + 488, + 367, + 500 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 31, + 488, + 367, + 500 + ], + "type": "text", + "content": " into a diffuse color " + }, + { + "bbox": [ + 31, + 488, + 367, + 500 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_d" + }, + { + "bbox": [ + 31, + 488, + 367, + 500 + ], + "type": "text", + "content": " and a specular one " + }, + { + "bbox": [ + 31, + 488, + 367, + 500 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_s" + }, + { + "bbox": [ + 31, + 488, + 367, + 500 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 144, + 512, + 379, + 524 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 512, + 379, + 524 + ], + "spans": [ + { + "bbox": [ + 144, + 512, + 379, + 524 + ], + "type": "interline_equation", + "content": "\\mathbf {c} = \\min \\left(\\max \\left(\\mathbf {c} _ {d} + \\mathbf {c} _ {s}, 0\\right), 1\\right) \\tag {3}", + "image_path": "84b50daaa2017c9f2ced9364450b25a12db2c62587e5d83f900719002b6c17ec.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 30, + 535, + 379, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 535, + 379, + 559 + ], + "spans": [ + { + "bbox": [ + 30, + 535, + 379, + 559 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 30, + 535, + 379, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_d + \\mathbf{c}_s" + }, + { + "bbox": [ + 30, + 535, + 379, + 559 + ], + "type": "text", + "content": " is clamped to [0, 1], both individual terms " + }, + { + "bbox": [ + 30, + 535, + 379, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_d" + }, + { + "bbox": [ + 30, + 535, + 379, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 535, + 379, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_s" + }, + { + "bbox": [ + 30, + 535, + 379, + 559 + ], + "type": "text", + "content": " from sigmoid are in [0, 1]. We now elaborate on the diffuse and specular color formulation." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 30, + 559, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 559, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 559, + 380, + 584 + ], + "type": "text", + "content": "Diffuse Color Formulation: With the updated mesh " + }, + { + "bbox": [ + 30, + 559, + 380, + 584 + ], + "type": "inline_equation", + "content": "\\mathcal{M}' = (\\mathcal{V}',\\mathcal{N}',\\mathcal{F})" + }, + { + "bbox": [ + 30, + 559, + 380, + 584 + ], + "type": "text", + "content": ", we perform differentiable rasterization [26] to obtain the position " + }, + { + "bbox": [ + 30, + 559, + 380, + 584 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 30, + 559, + 380, + 584 + ], + "type": "text", + "content": " corresponding" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 32, + 10, + 39, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 10, + 39, + 18 + ], + "spans": [ + { + "bbox": [ + 32, + 10, + 39, + 18 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 9, + 107, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 9, + 107, + 19 + ], + "spans": [ + { + "bbox": [ + 66, + 9, + 107, + 19 + ], + "type": "text", + "content": "C.Ji et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 31, + 34, + 379, + 59 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 34, + 379, + 59 + ], + "spans": [ + { + "bbox": [ + 31, + 34, + 379, + 59 + ], + "type": "text", + "content": "to each pixel. Subsequently, " + }, + { + "bbox": [ + 31, + 34, + 379, + 59 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 31, + 34, + 379, + 59 + ], + "type": "text", + "content": " is mapped to a diffuse color " + }, + { + "bbox": [ + 31, + 34, + 379, + 59 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_d\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 31, + 34, + 379, + 59 + ], + "type": "text", + "content": " and a view-independent feature " + }, + { + "bbox": [ + 31, + 34, + 379, + 59 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_s\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 31, + 34, + 379, + 59 + ], + "type": "text", + "content": " which will be used later for decoding specular color:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 176, + 71, + 379, + 83 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 71, + 379, + 83 + ], + "spans": [ + { + "bbox": [ + 176, + 71, + 379, + 83 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {d}, \\mathbf {f} _ {s} = f _ {\\phi} (\\mathbf {p}) \\tag {4}", + "image_path": "f5d111718ca8a3ba14466e4fdd1045472c9bf2add7b1f38a5afa7ebe9c0af0ec.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 30, + 95, + 379, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 95, + 379, + 119 + ], + "spans": [ + { + "bbox": [ + 30, + 95, + 379, + 119 + ], + "type": "text", + "content": "While the number of channels of " + }, + { + "bbox": [ + 30, + 95, + 379, + 119 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_s" + }, + { + "bbox": [ + 30, + 95, + 379, + 119 + ], + "type": "text", + "content": " is flexible, we observe that such a compact three-channel representation is sufficient for our purpose." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 30, + 120, + 379, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 120, + 379, + 180 + ], + "spans": [ + { + "bbox": [ + 30, + 120, + 379, + 180 + ], + "type": "text", + "content": "Specular Color Formulation: Recent non-real-time methods [14, 46, 56] propose to model view-dependent color based on the reflective direction " + }, + { + "bbox": [ + 30, + 120, + 379, + 180 + ], + "type": "inline_equation", + "content": "\\omega_{r}" + }, + { + "bbox": [ + 30, + 120, + 379, + 180 + ], + "type": "text", + "content": ", instead of the viewing direction used by NeRF. This approach has been shown to effectively model reflective surfaces. Motivated by these methods, we calculate the reflective direction as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 153, + 194, + 379, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 194, + 379, + 205 + ], + "spans": [ + { + "bbox": [ + 153, + 194, + 379, + 205 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\omega} _ {r} = 2 \\left(\\boldsymbol {\\omega} _ {o} \\cdot \\mathbf {n} _ {\\mathbf {p}}\\right) \\mathbf {n} _ {\\mathbf {p}} - \\boldsymbol {\\omega} _ {o} \\tag {5}", + "image_path": "a9963826c89e40c1162c3ed965f237e54ed8e9ea093592ba4ad93f5eca8cf3a9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 31, + 217, + 379, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 217, + 379, + 241 + ], + "spans": [ + { + "bbox": [ + 31, + 217, + 379, + 241 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 31, + 217, + 379, + 241 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_{\\mathbf{p}}" + }, + { + "bbox": [ + 31, + 217, + 379, + 241 + ], + "type": "text", + "content": " is smoothly interpolated surface normal at " + }, + { + "bbox": [ + 31, + 217, + 379, + 241 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 31, + 217, + 379, + 241 + ], + "type": "text", + "content": " based on our learned vertex normal " + }, + { + "bbox": [ + 31, + 217, + 379, + 241 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_{\\mathbf{v}}^{\\prime}" + }, + { + "bbox": [ + 31, + 217, + 379, + 241 + ], + "type": "text", + "content": ". Next, the specular color can be obtained as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 255, + 379, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 255, + 379, + 266 + ], + "spans": [ + { + "bbox": [ + 151, + 255, + 379, + 266 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {s} = f _ {\\psi} \\left(\\mathbf {f} _ {s}, \\boldsymbol {\\omega} _ {r}, \\left(\\boldsymbol {\\omega} _ {o} \\cdot \\mathbf {n} _ {\\mathbf {p}}\\right)\\right) \\tag {6}", + "image_path": "dd6ab364aa53c6577b0a17c9d03b7dc9f0e551689106d7fb697f22002271ccb1.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 30, + 277, + 379, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 277, + 379, + 302 + ], + "spans": [ + { + "bbox": [ + 30, + 277, + 379, + 302 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 30, + 277, + 379, + 302 + ], + "type": "inline_equation", + "content": "(\\boldsymbol{\\omega}_o \\cdot \\mathbf{n_p})" + }, + { + "bbox": [ + 30, + 277, + 379, + 302 + ], + "type": "text", + "content": " is considered as input to model the Fresnel effects [21] and " + }, + { + "bbox": [ + 30, + 277, + 379, + 302 + ], + "type": "inline_equation", + "content": "f_{\\psi}" + }, + { + "bbox": [ + 30, + 277, + 379, + 302 + ], + "type": "text", + "content": " needs to be a tiny MLP to enable real-time rendering." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 30, + 303, + 379, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 303, + 379, + 400 + ], + "spans": [ + { + "bbox": [ + 30, + 303, + 379, + 400 + ], + "type": "text", + "content": "Environment Learner: In practice, we observe that naively mapping the reflective viewing direction to specular color (see Eq. (6)) leads to unsatisfying performance due to the capacity limitation of the tiny MLP " + }, + { + "bbox": [ + 30, + 303, + 379, + 400 + ], + "type": "inline_equation", + "content": "f_{\\psi}" + }, + { + "bbox": [ + 30, + 303, + 379, + 400 + ], + "type": "text", + "content": ". For a similar reason, Ref-NeRF adopts an 8-layer MLP to map " + }, + { + "bbox": [ + 30, + 303, + 379, + 400 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_s" + }, + { + "bbox": [ + 30, + 303, + 379, + 400 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 303, + 379, + 400 + ], + "type": "inline_equation", + "content": "\\omega_r" + }, + { + "bbox": [ + 30, + 303, + 379, + 400 + ], + "type": "text", + "content": " to the view-dependent color. However, increasing the capacity of " + }, + { + "bbox": [ + 30, + 303, + 379, + 400 + ], + "type": "inline_equation", + "content": "f_{\\psi}" + }, + { + "bbox": [ + 30, + 303, + 379, + 400 + ], + "type": "text", + "content": " hampers the real-time rendering capability. Therefore, instead of increasing the capacity of " + }, + { + "bbox": [ + 30, + 303, + 379, + 400 + ], + "type": "inline_equation", + "content": "f_{\\psi}" + }, + { + "bbox": [ + 30, + 303, + 379, + 400 + ], + "type": "text", + "content": ", we propose to leverage a neural environment learner of large capacity that maps " + }, + { + "bbox": [ + 30, + 303, + 379, + 400 + ], + "type": "inline_equation", + "content": "\\omega_r" + }, + { + "bbox": [ + 30, + 303, + 379, + 400 + ], + "type": "text", + "content": " to an environment feature " + }, + { + "bbox": [ + 30, + 303, + 379, + 400 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_e \\in \\mathbb{R}^M" + }, + { + "bbox": [ + 30, + 303, + 379, + 400 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 30, + 303, + 379, + 400 + ], + "type": "inline_equation", + "content": "M = 3" + }, + { + "bbox": [ + 30, + 303, + 379, + 400 + ], + "type": "text", + "content": " in our paper):" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 174, + 417, + 379, + 428 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 417, + 379, + 428 + ], + "spans": [ + { + "bbox": [ + 174, + 417, + 379, + 428 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {e} = f _ {e} \\left(\\gamma \\left(\\boldsymbol {\\omega} _ {r}\\right)\\right) \\tag {7}", + "image_path": "0f8da53ebe82d9ec8db2a9c6aa85c55022586269eca03896756289cd2bc4b9fc.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 30, + 437, + 379, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 437, + 379, + 462 + ], + "spans": [ + { + "bbox": [ + 30, + 437, + 379, + 462 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 30, + 437, + 379, + 462 + ], + "type": "inline_equation", + "content": "\\gamma (\\cdot)" + }, + { + "bbox": [ + 30, + 437, + 379, + 462 + ], + "type": "text", + "content": " denotes positional encoding and " + }, + { + "bbox": [ + 30, + 437, + 379, + 462 + ], + "type": "inline_equation", + "content": "f_{e}" + }, + { + "bbox": [ + 30, + 437, + 379, + 462 + ], + "type": "text", + "content": " is a 4-layer MLP. This leads to our final implementation of the specular color:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 154, + 475, + 379, + 486 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 475, + 379, + 486 + ], + "spans": [ + { + "bbox": [ + 154, + 475, + 379, + 486 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {s} = f _ {\\psi} \\left(\\mathbf {f} _ {s}, \\mathbf {f} _ {e}, \\left(\\boldsymbol {\\omega} _ {o} \\cdot \\mathbf {n} _ {\\mathbf {p}}\\right)\\right) \\tag {8}", + "image_path": "be015839a8562349e58f03c26ea002950f709bf6f1bc5a56dcc9b454faab8c08.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 30, + 498, + 379, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 498, + 379, + 523 + ], + "spans": [ + { + "bbox": [ + 30, + 498, + 379, + 523 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 30, + 498, + 379, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_e" + }, + { + "bbox": [ + 30, + 498, + 379, + 523 + ], + "type": "text", + "content": " can be baked into a two-dimensional neural environment feature map for real-time rendering as it solely depends on " + }, + { + "bbox": [ + 30, + 498, + 379, + 523 + ], + "type": "inline_equation", + "content": "\\omega_r" + }, + { + "bbox": [ + 30, + 498, + 379, + 523 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 30, + 523, + 379, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 523, + 379, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 523, + 379, + 584 + ], + "type": "text", + "content": "Scene Editing: REFRAME can decouple geometry, diffuse and specular color, so we can perform simple scene editing tasks. For example, we can edit the geometry and appearance of objects by modifying the mesh or texture image. Additionally, we can perform relighting on the objects by modifying the environment feature map of the scene, as shown in Fig. 4." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 9, + 346, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 9, + 346, + 19 + ], + "spans": [ + { + "bbox": [ + 297, + 9, + 346, + 19 + ], + "type": "text", + "content": "REFRAME" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 45, + 31, + 138, + 107 + ], + "blocks": [ + { + "bbox": [ + 45, + 31, + 138, + 107 + ], + "lines": [ + { + "bbox": [ + 45, + 31, + 138, + 107 + ], + "spans": [ + { + "bbox": [ + 45, + 31, + 138, + 107 + ], + "type": "image", + "image_path": "cfe3c5e600f39b784c1622cf7f94a1dd20ce647c1468c766b90186b4a8cf59fb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 108, + 123, + 118 + ], + "lines": [ + { + "bbox": [ + 56, + 108, + 123, + 118 + ], + "spans": [ + { + "bbox": [ + 56, + 108, + 123, + 118 + ], + "type": "text", + "content": "(a) Original image." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 157, + 31, + 251, + 107 + ], + "blocks": [ + { + "bbox": [ + 157, + 31, + 251, + 107 + ], + "lines": [ + { + "bbox": [ + 157, + 31, + 251, + 107 + ], + "spans": [ + { + "bbox": [ + 157, + 31, + 251, + 107 + ], + "type": "image", + "image_path": "500ef3ad23a8817ec90a1806390a84d599c4541f992dc732bb37e699148f6d5f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 108, + 234, + 118 + ], + "lines": [ + { + "bbox": [ + 173, + 108, + 234, + 118 + ], + "spans": [ + { + "bbox": [ + 173, + 108, + 234, + 118 + ], + "type": "text", + "content": "(b) Flip lighting." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 30, + 126, + 380, + 183 + ], + "lines": [ + { + "bbox": [ + 30, + 126, + 380, + 183 + ], + "spans": [ + { + "bbox": [ + 30, + 126, + 380, + 183 + ], + "type": "text", + "content": "Fig. 4: Relighting. We can achieve relighting of objects by editing our environment feature map. The image illustrates the editing process of flipping or replacing the environmental feature map. The left side shows the rendering result, the top right corner presents the corresponding environment feature map, and the specular color is displayed in the bottom right corner." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 271, + 31, + 364, + 107 + ], + "blocks": [ + { + "bbox": [ + 271, + 31, + 364, + 107 + ], + "lines": [ + { + "bbox": [ + 271, + 31, + 364, + 107 + ], + "spans": [ + { + "bbox": [ + 271, + 31, + 364, + 107 + ], + "type": "image", + "image_path": "943a0becffb385cdcf6bb0c86af562dba7afd1e8c12602e24405de41cb1af613.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 282, + 108, + 354, + 118 + ], + "lines": [ + { + "bbox": [ + 282, + 108, + 354, + 118 + ], + "spans": [ + { + "bbox": [ + 282, + 108, + 354, + 118 + ], + "type": "text", + "content": "(c) Change lighting." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 31, + 205, + 134, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 205, + 134, + 215 + ], + "spans": [ + { + "bbox": [ + 31, + 205, + 134, + 215 + ], + "type": "text", + "content": "3.3 Loss Functions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 30, + 223, + 380, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 223, + 380, + 284 + ], + "spans": [ + { + "bbox": [ + 30, + 223, + 380, + 284 + ], + "type": "text", + "content": "We train our model based on image reconstruction loss and several regularization terms. At each iteration, we render a full image and train with the corresponding ground truth. Let " + }, + { + "bbox": [ + 30, + 223, + 380, + 284 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{c}}" + }, + { + "bbox": [ + 30, + 223, + 380, + 284 + ], + "type": "text", + "content": " denote the ground truth RGB value at one pixel. We optionally use ground truth binary mask for supervision at synthetic object-centric scenes. Our full loss function is shown in Eq. (9):" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 111, + 292, + 379, + 320 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 292, + 379, + 320 + ], + "spans": [ + { + "bbox": [ + 111, + 292, + 379, + 320 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\lambda_ {\\mathbf {c}} \\mathcal {L} _ {\\mathbf {c}} + \\lambda_ {\\mathbf {c} _ {d}} \\mathcal {L} _ {\\mathbf {c} _ {d}} + \\lambda_ {S S I M} \\mathcal {L} _ {S S I M} + \\tag {9} \\\\ \\lambda_ {m a s k} \\mathcal {L} _ {m a s k} + \\lambda_ {\\Delta \\mathbf {n} _ {\\mathbf {v}}} \\mathcal {L} _ {\\Delta \\mathbf {n} _ {\\mathbf {v}}} + \\lambda_ {m a x} \\mathcal {L} _ {m a x} \\\\ \\end{array}", + "image_path": "6254f342f41bde716708ee4e96d4e9dfefed9013ac42cae257fd4a0cb73d5b11.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 30, + 327, + 379, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 327, + 379, + 352 + ], + "spans": [ + { + "bbox": [ + 30, + 327, + 379, + 352 + ], + "type": "text", + "content": "Image Reconstruction Loss: Our model is first supervised by the L2 loss between the final rendered color and the ground truth." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 172, + 360, + 379, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 360, + 379, + 373 + ], + "spans": [ + { + "bbox": [ + 172, + 360, + 379, + 373 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathbf {c}} = \\left| \\left| \\mathbf {c} - \\hat {\\mathbf {c}} \\right| \\right| _ {2} \\tag {10}", + "image_path": "33647e5a21dd86726f884be456e6456737de5d0df64a02bebfae14e6c68b5f9e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 31, + 381, + 379, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 381, + 379, + 405 + ], + "spans": [ + { + "bbox": [ + 31, + 381, + 379, + 405 + ], + "type": "text", + "content": "To encourage the correct decoupling of diffuse and specular colors, we add a diffuse loss to guide the diffuse color to be close to the ground truth color:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 168, + 413, + 379, + 427 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 413, + 379, + 427 + ], + "spans": [ + { + "bbox": [ + 168, + 413, + 379, + 427 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathbf {c} _ {d}} = \\left\\| \\mathbf {c} _ {d} - \\hat {\\mathbf {c}} \\right\\| _ {2} \\tag {11}", + "image_path": "4a94ea2588abde1eacb04dfce0484427c3e4f99dd64cbc818d58e9de7a045e1f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 30, + 434, + 379, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 434, + 379, + 481 + ], + "spans": [ + { + "bbox": [ + 30, + 434, + 379, + 481 + ], + "type": "text", + "content": "We also introduce SSIM [17] loss for better perceptual quality. On object-centric datasets, we additionally utilize mask loss to ensure the rasterization mask closely aligns with the ground truth mask. Note that the mask loss is optional and is not required on unbounded datasets." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 30, + 482, + 379, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 482, + 379, + 530 + ], + "spans": [ + { + "bbox": [ + 30, + 482, + 379, + 530 + ], + "type": "text", + "content": "Regularizations: We regularize the predicted vertex normal " + }, + { + "bbox": [ + 30, + 482, + 379, + 530 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_{\\mathbf{v}}^{\\prime}" + }, + { + "bbox": [ + 30, + 482, + 379, + 530 + ], + "type": "text", + "content": " to be close to the original vertex normal " + }, + { + "bbox": [ + 30, + 482, + 379, + 530 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_{\\mathbf{v}}" + }, + { + "bbox": [ + 30, + 482, + 379, + 530 + ], + "type": "text", + "content": " computed from the mesh. This allows for stabilizing the optimization of the vertex normals, e.g., avoid completely flipping the normal. This is achieved by regularizing the magnitude of the normal offset:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 167, + 539, + 379, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 539, + 379, + 552 + ], + "spans": [ + { + "bbox": [ + 167, + 539, + 379, + 552 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\Delta \\mathbf {n} _ {\\mathbf {v}}} = \\left\\| \\Delta \\mathbf {n} _ {\\mathbf {v}} \\right\\| _ {1} \\tag {12}", + "image_path": "8a8346298039527784f72eec1cf7c77602ecad3a175da340900d8dfcc46989fa.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 31, + 560, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 560, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 31, + 560, + 380, + 584 + ], + "type": "text", + "content": "We introduce another regularization to encourage the specular color " + }, + { + "bbox": [ + 31, + 560, + 380, + 584 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_s" + }, + { + "bbox": [ + 31, + 560, + 380, + 584 + ], + "type": "text", + "content": " to be reasonable when visualized alone. As shown in Eq. (3), the final color " + }, + { + "bbox": [ + 31, + 560, + 380, + 584 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 31, + 560, + 380, + 584 + ], + "type": "text", + "content": " will be" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 32, + 10, + 39, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 10, + 39, + 18 + ], + "spans": [ + { + "bbox": [ + 32, + 10, + 39, + 18 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 9, + 107, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 9, + 107, + 19 + ], + "spans": [ + { + "bbox": [ + 66, + 9, + 107, + 19 + ], + "type": "text", + "content": "C.Ji et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 30, + 35, + 380, + 94 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 35, + 380, + 94 + ], + "spans": [ + { + "bbox": [ + 30, + 35, + 380, + 94 + ], + "type": "text", + "content": "clamped when the sum of specular color " + }, + { + "bbox": [ + 30, + 35, + 380, + 94 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_s" + }, + { + "bbox": [ + 30, + 35, + 380, + 94 + ], + "type": "text", + "content": " and diffuse color " + }, + { + "bbox": [ + 30, + 35, + 380, + 94 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_d" + }, + { + "bbox": [ + 30, + 35, + 380, + 94 + ], + "type": "text", + "content": " exceeds 1. In this case, the gradient is truncated due to the clamp operation, preventing the backpropagation from updating the color values. This may leave an undesired large specular color " + }, + { + "bbox": [ + 30, + 35, + 380, + 94 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_s" + }, + { + "bbox": [ + 30, + 35, + 380, + 94 + ], + "type": "text", + "content": " with no penalization. Therefore, we encourage the sum of specular color and diffuse color to not exceed 1:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 139, + 103, + 378, + 114 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 103, + 378, + 114 + ], + "spans": [ + { + "bbox": [ + 139, + 103, + 378, + 114 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\max } = \\left| \\left| \\operatorname {m a x} \\left(\\mathbf {c} _ {d} + \\mathbf {c} _ {s} - 1, 0\\right) \\right| \\right| \\tag {13}", + "image_path": "8a9e8a23a87601c57b1570f73916eebceecf78463858985660347a66ff03bd00.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 31, + 130, + 169, + 143 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 130, + 169, + 143 + ], + "spans": [ + { + "bbox": [ + 31, + 130, + 169, + 143 + ], + "type": "text", + "content": "3.4 Real-Time Rendering" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 30, + 148, + 380, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 148, + 380, + 255 + ], + "spans": [ + { + "bbox": [ + 30, + 148, + 380, + 255 + ], + "type": "text", + "content": "Environment Feature Map: During the training process, we employ a four-layer MLP with a width of 256 as our environment learner. If we query " + }, + { + "bbox": [ + 30, + 148, + 380, + 255 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_e" + }, + { + "bbox": [ + 30, + 148, + 380, + 255 + ], + "type": "text", + "content": " through the environment learner module during the inference process, it would significantly slow down the inference speed. Hence, during the inference phase, we bake the learned environment feature " + }, + { + "bbox": [ + 30, + 148, + 380, + 255 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_e" + }, + { + "bbox": [ + 30, + 148, + 380, + 255 + ], + "type": "text", + "content": " into a 2D environment feature map by converting " + }, + { + "bbox": [ + 30, + 148, + 380, + 255 + ], + "type": "inline_equation", + "content": "\\omega_r \\in \\mathbb{S}^2" + }, + { + "bbox": [ + 30, + 148, + 380, + 255 + ], + "type": "text", + "content": " to its polar coordinate, see supplementary for more details. This allows us to simply query the corresponding feature on the environment feature map based on " + }, + { + "bbox": [ + 30, + 148, + 380, + 255 + ], + "type": "inline_equation", + "content": "\\omega_r" + }, + { + "bbox": [ + 30, + 148, + 380, + 255 + ], + "type": "text", + "content": ", thereby greatly accelerating our inference speed, reaching above 200 FPS on a single NVIDIA 3090 GPU." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 257, + 379, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 257, + 379, + 329 + ], + "spans": [ + { + "bbox": [ + 30, + 257, + 379, + 329 + ], + "type": "text", + "content": "Texture Images: To further reduce the inference burden and better deploy our method across different platforms, we can also bake the " + }, + { + "bbox": [ + 30, + 257, + 379, + 329 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_d" + }, + { + "bbox": [ + 30, + 257, + 379, + 329 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 257, + 379, + 329 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_s" + }, + { + "bbox": [ + 30, + 257, + 379, + 329 + ], + "type": "text", + "content": " obtained from querying the " + }, + { + "bbox": [ + 30, + 257, + 379, + 329 + ], + "type": "inline_equation", + "content": "f_{\\phi}" + }, + { + "bbox": [ + 30, + 257, + 379, + 329 + ], + "type": "text", + "content": " into two texture images. Specifically, we map the vertices of the mesh to UV coordinates and then bake the features into two texture images, respectively. In order to obtain the normal " + }, + { + "bbox": [ + 30, + 257, + 379, + 329 + ], + "type": "inline_equation", + "content": "\\mathbf{n_p}" + }, + { + "bbox": [ + 30, + 257, + 379, + 329 + ], + "type": "text", + "content": " from the learned normal vector " + }, + { + "bbox": [ + 30, + 257, + 379, + 329 + ], + "type": "inline_equation", + "content": "\\mathbf{n_v}'" + }, + { + "bbox": [ + 30, + 257, + 379, + 329 + ], + "type": "text", + "content": " during real-time rendering, we also bake the normal " + }, + { + "bbox": [ + 30, + 257, + 379, + 329 + ], + "type": "inline_equation", + "content": "\\mathbf{n_p}" + }, + { + "bbox": [ + 30, + 257, + 379, + 329 + ], + "type": "text", + "content": " into a texture image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 31, + 345, + 124, + 359 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 345, + 124, + 359 + ], + "spans": [ + { + "bbox": [ + 31, + 345, + 124, + 359 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 30, + 368, + 379, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 368, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 30, + 368, + 379, + 583 + ], + "type": "text", + "content": "Implementation Details. Training: Prior to the training stage, we employ the classic quadric error metrics algorithm [13] to simplify the initial mesh to 75,000 faces except for the outdoor scenes which typically have complex geometry. We train for 250 epochs on each scene, with a training time of approximately 2 hours. We utilize the Adam optimizer [23] and employ the cosine annealing learning rate adjustment strategy [32] during the training stage. Rendering: The resolution of our baked texture image is " + }, + { + "bbox": [ + 30, + 368, + 379, + 583 + ], + "type": "inline_equation", + "content": "4096 \\times 4096" + }, + { + "bbox": [ + 30, + 368, + 379, + 583 + ], + "type": "text", + "content": ", and the resolution of the baked environment feature map is " + }, + { + "bbox": [ + 30, + 368, + 379, + 583 + ], + "type": "inline_equation", + "content": "720 \\times 360" + }, + { + "bbox": [ + 30, + 368, + 379, + 583 + ], + "type": "text", + "content": ". As referenced in Sec. 3.4, once we have baked the environment learner into the environment feature map, our approach is already capable of achieving above 200 FPS on the GPU (referred to as Ours). This implementation ensures a fair comparison with the baselines since their results quoted are evaluated using GPU implementation that can not directly deploy on mobile devices. Given that we use texture images for rendering when implemented on mobile devices, we also employ texture images on the GPU for quality evaluation (referred to as Ours (Mobile)). Following NeRF2Mesh [45] and MobileNeRF [9], we employ anti-aliasing to enhance our rendering quality. For details on the implementation of anti-aliasing, please refer to the supplementary material. All experiments are performed on a single NVIDIA 3090 GPU." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 9, + 346, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 9, + 346, + 19 + ], + "spans": [ + { + "bbox": [ + 297, + 9, + 346, + 19 + ], + "type": "text", + "content": "REFRAME" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 36, + 33, + 379, + 141 + ], + "blocks": [ + { + "bbox": [ + 36, + 33, + 379, + 141 + ], + "lines": [ + { + "bbox": [ + 36, + 33, + 379, + 141 + ], + "spans": [ + { + "bbox": [ + 36, + 33, + 379, + 141 + ], + "type": "table", + "html": "
NeRF Synthetic DatasetShiny Blender DatasetReal Captured Dataset
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NeRF [33]31.010.9470.081------
Ref-NeRF [46]33.990.9660.03835.960.9670.05824.450.6650.142
3DGS [22]33.300.9690.03030.370.9470.08324.060.6610.259
NvDiffRec [35]29.050.9390.08129.050.9380.111---
MobileNeRF [9]30.900.9470.06226.620.8830.16322.960.4930.430
NeRF2Mesh [45]29.670.9400.07226.960.8960.17022.710.5230.419
Ours31.040.9610.04036.020.9830.04022.500.5980.325
Ours (Mobile)30.840.9590.04235.830.9810.04122.110.5240.423
", + "image_path": "9ee59315cb66968c5b8ffa488b6a67ab6d8c44dd4f9ea6bfde23802508bce61b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 31, + 142, + 380, + 175 + ], + "lines": [ + { + "bbox": [ + 31, + 142, + 380, + 175 + ], + "spans": [ + { + "bbox": [ + 31, + 142, + 380, + 175 + ], + "type": "text", + "content": "Table 1: Rendering Quality. Baseline comparisons of the rendering quality on three different datasets. Red represents the optimal, orange represents the second best, and yellow represents the third." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 30, + 209, + 380, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 209, + 380, + 340 + ], + "spans": [ + { + "bbox": [ + 30, + 209, + 380, + 340 + ], + "type": "text", + "content": "Dataset. We validate the effectiveness and robustness of our method using three different datasets with varied scenes. 1) NeRF [33] Synthetic Dataset: This dataset consists of eight synthetic scenes. 2) Shiny Blender Dataset: This dataset is introduced by Ref-NeRF [46] and includes six glossy objects, serving as the primary dataset for assessing the reflective surface modeling capabilities of the methods. 3) Real Captured Dataset: We select three outdoor scenes with rich reflective appearances same as Ref-NeRF [46] to validate the effectiveness of our method in real-world outdoor capture scenarios. Due to the poor quality of the initial mesh extracted by Nerf2Mesh [45] when modeling glossy objects, we utilize the meshes extracted by Neuralangelo [29] and Ref-NeuS [14] as our initialization in certain scenes, see supplementary for more details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 30, + 341, + 380, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 341, + 380, + 450 + ], + "spans": [ + { + "bbox": [ + 30, + 341, + 380, + 450 + ], + "type": "text", + "content": "Baseline. We compare with various types of baselines, including Ref-NeRF [46], 3DGS [22], NvDiffRec [35], MobileNeRF [9], and NeRF2Mesh [45] in our study. Ref-NeRF [46] serves as the state-of-the-art volume rendering method for handling the strong reflective appearance. 3DGS [22] represents a powerful new representation capable of real-time rendering on GPU. NvDiffRec [35] represents physically based rendering methods. MobileNeRF [9], and NeRF2Mesh [45] are representative works in the field of real-time rendering for mobile devices. We import the results of the baselines from their original paper and replicate the results that missing in their original paper." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 31, + 468, + 177, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 468, + 177, + 480 + ], + "spans": [ + { + "bbox": [ + 31, + 468, + 177, + 480 + ], + "type": "text", + "content": "4.1 Reconstruction Quality" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 30, + 487, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 487, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 487, + 380, + 584 + ], + "type": "text", + "content": "Rendering Quality. We compare the rendering quality of our method and baseline methods on three datasets, see Tab. 1. It is worth noting that Ref-NeRF [46] is the state-of-the-art method for modeling glossy objects but cannot achieve real-time rendering, even on high-end GPU. We introduce Ref-NeRF [46] as a benchmark for rendering quality and demonstrate that our method achieves comparable or even superior rendering quality to non-real-time state-of-the-art methods. While 3DGS [22] can render in real-time on GPU and has higher PSNR on both NeRF Synthetic and Real Captured dataset than ours, it's important" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 107, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 107, + 19 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 107, + 19 + ], + "type": "text", + "content": "C.Ji et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 39, + 109, + 84 + ], + "blocks": [ + { + "bbox": [ + 32, + 59, + 39, + 70 + ], + "lines": [ + { + "bbox": [ + 32, + 59, + 39, + 70 + ], + "spans": [ + { + "bbox": [ + 32, + 59, + 39, + 70 + ], + "type": "text", + "content": "C" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 31, + 86, + 39 + ], + "lines": [ + { + "bbox": [ + 56, + 31, + 86, + 39 + ], + "spans": [ + { + "bbox": [ + 56, + 31, + 86, + 39 + ], + "type": "text", + "content": "NvDiffRec" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 39, + 109, + 84 + ], + "lines": [ + { + "bbox": [ + 56, + 39, + 109, + 84 + ], + "spans": [ + { + "bbox": [ + 56, + 39, + 109, + 84 + ], + "type": "image", + "image_path": "e0cf1ca3563fe8025eda30447c9cdf9f16799b4892de70dfd0f30c6adb46474b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 47, + 86, + 85, + 126 + ], + "blocks": [ + { + "bbox": [ + 32, + 97, + 39, + 119 + ], + "lines": [ + { + "bbox": [ + 32, + 97, + 39, + 119 + ], + "spans": [ + { + "bbox": [ + 32, + 97, + 39, + 119 + ], + "type": "text", + "content": "Helmert" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 86, + 85, + 126 + ], + "lines": [ + { + "bbox": [ + 47, + 86, + 85, + 126 + ], + "spans": [ + { + "bbox": [ + 47, + 86, + 85, + 126 + ], + "type": "image", + "image_path": "3ece761960d140f3cfa8d73a0bd4b325a80098ac5873a14900bb4dcdf8e282cc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 51, + 132, + 80, + 177 + ], + "blocks": [ + { + "bbox": [ + 32, + 148, + 39, + 168 + ], + "lines": [ + { + "bbox": [ + 32, + 148, + 39, + 168 + ], + "spans": [ + { + "bbox": [ + 32, + 148, + 39, + 168 + ], + "type": "text", + "content": "Toaster" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 132, + 80, + 177 + ], + "lines": [ + { + "bbox": [ + 51, + 132, + 80, + 177 + ], + "spans": [ + { + "bbox": [ + 51, + 132, + 80, + 177 + ], + "type": "image", + "image_path": "0909144c648b932e13735cbc07ccaccd15639a4c851cbee9a295b69e6f30f129.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 30, + 185, + 380, + 219 + ], + "lines": [ + { + "bbox": [ + 30, + 185, + 380, + 219 + ], + "spans": [ + { + "bbox": [ + 30, + 185, + 380, + 219 + ], + "type": "text", + "content": "Fig. 5: Rendering Quality on Shiny Blender Dataset. Our method achieves optimal rendering quality in most scenes and provides better modeling of reflective appearance compared to the comparison methods [9,35,45]." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 123, + 39, + 176, + 84 + ], + "blocks": [ + { + "bbox": [ + 121, + 32, + 156, + 38 + ], + "lines": [ + { + "bbox": [ + 121, + 32, + 156, + 38 + ], + "spans": [ + { + "bbox": [ + 121, + 32, + 156, + 38 + ], + "type": "text", + "content": "MobileNeRF" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 123, + 39, + 176, + 84 + ], + "lines": [ + { + "bbox": [ + 123, + 39, + 176, + 84 + ], + "spans": [ + { + "bbox": [ + 123, + 39, + 176, + 84 + ], + "type": "image", + "image_path": "0673bde51653b9fdc6a5d1beac2d9586278bb6729d3131a96343e34cdd65484a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 115, + 86, + 153, + 126 + ], + "blocks": [ + { + "bbox": [ + 115, + 86, + 153, + 126 + ], + "lines": [ + { + "bbox": [ + 115, + 86, + 153, + 126 + ], + "spans": [ + { + "bbox": [ + 115, + 86, + 153, + 126 + ], + "type": "image", + "image_path": "d25fdc1caeeb7c21d3a652c1adf243a58530933dac34c4bcf0d7f483c8c3aefb.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 119, + 132, + 148, + 177 + ], + "blocks": [ + { + "bbox": [ + 119, + 132, + 148, + 177 + ], + "lines": [ + { + "bbox": [ + 119, + 132, + 148, + 177 + ], + "spans": [ + { + "bbox": [ + 119, + 132, + 148, + 177 + ], + "type": "image", + "image_path": "f552b8380a342be290b9a561c31bc861cab62f06830d8159b233d1ea8d277deb.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 193, + 39, + 244, + 84 + ], + "blocks": [ + { + "bbox": [ + 189, + 32, + 223, + 38 + ], + "lines": [ + { + "bbox": [ + 189, + 32, + 223, + 38 + ], + "spans": [ + { + "bbox": [ + 189, + 32, + 223, + 38 + ], + "type": "text", + "content": "NeRF2Mesh" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 193, + 39, + 244, + 84 + ], + "lines": [ + { + "bbox": [ + 193, + 39, + 244, + 84 + ], + "spans": [ + { + "bbox": [ + 193, + 39, + 244, + 84 + ], + "type": "image", + "image_path": "8e6819ca04d5ea4f4a2b01b1ed1d22a424cbf3f13a285108a1c67dd4fdfde17e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 182, + 86, + 220, + 127 + ], + "blocks": [ + { + "bbox": [ + 182, + 86, + 220, + 127 + ], + "lines": [ + { + "bbox": [ + 182, + 86, + 220, + 127 + ], + "spans": [ + { + "bbox": [ + 182, + 86, + 220, + 127 + ], + "type": "image", + "image_path": "aad6df0036970970f779dfd8e995f7aa01b8ff9f429a7ca59f84a57790051d85.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 154, + 86, + 176, + 130 + ], + "blocks": [ + { + "bbox": [ + 154, + 86, + 176, + 130 + ], + "lines": [ + { + "bbox": [ + 154, + 86, + 176, + 130 + ], + "spans": [ + { + "bbox": [ + 154, + 86, + 176, + 130 + ], + "type": "image", + "image_path": "de997674c608a635f5e20925a181b8d930ffad02b559d08eac8dde5c0c614ddb.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 154, + 131, + 176, + 176 + ], + "blocks": [ + { + "bbox": [ + 154, + 131, + 176, + 176 + ], + "lines": [ + { + "bbox": [ + 154, + 131, + 176, + 176 + ], + "spans": [ + { + "bbox": [ + 154, + 131, + 176, + 176 + ], + "type": "image", + "image_path": "4c8588d3154606b6943e65d590971c47f48feedfcaf03e8eaa5f7b9ea35ab660.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 222, + 86, + 244, + 128 + ], + "blocks": [ + { + "bbox": [ + 222, + 86, + 244, + 128 + ], + "lines": [ + { + "bbox": [ + 222, + 86, + 244, + 128 + ], + "spans": [ + { + "bbox": [ + 222, + 86, + 244, + 128 + ], + "type": "image", + "image_path": "7506e90e827bae975ed0e2a2636a205723833c4487a6a2b65e7645beb7d027f6.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 222, + 131, + 244, + 176 + ], + "blocks": [ + { + "bbox": [ + 222, + 131, + 244, + 176 + ], + "lines": [ + { + "bbox": [ + 222, + 131, + 244, + 176 + ], + "spans": [ + { + "bbox": [ + 222, + 131, + 244, + 176 + ], + "type": "image", + "image_path": "e12e0bc85bb5acefa4083d97ccbb56b04b60708963238e366021520feb597572.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 259, + 39, + 282, + 80 + ], + "blocks": [ + { + "bbox": [ + 261, + 33, + 275, + 39 + ], + "lines": [ + { + "bbox": [ + 261, + 33, + 275, + 39 + ], + "spans": [ + { + "bbox": [ + 261, + 33, + 275, + 39 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 259, + 39, + 282, + 80 + ], + "lines": [ + { + "bbox": [ + 259, + 39, + 282, + 80 + ], + "spans": [ + { + "bbox": [ + 259, + 39, + 282, + 80 + ], + "type": "image", + "image_path": "54a21d313b747c0c8016573d54fd77360827ddde6286cc47a38fa3fe7598d5ed.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 250, + 91, + 288, + 126 + ], + "blocks": [ + { + "bbox": [ + 250, + 91, + 288, + 126 + ], + "lines": [ + { + "bbox": [ + 250, + 91, + 288, + 126 + ], + "spans": [ + { + "bbox": [ + 250, + 91, + 288, + 126 + ], + "type": "image", + "image_path": "1ba0963fb7223b5653a2893729329263342398f342de0d927a70a5744322d986.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 254, + 132, + 283, + 177 + ], + "blocks": [ + { + "bbox": [ + 254, + 132, + 283, + 177 + ], + "lines": [ + { + "bbox": [ + 254, + 132, + 283, + 177 + ], + "spans": [ + { + "bbox": [ + 254, + 132, + 283, + 177 + ], + "type": "image", + "image_path": "fe60be153fd9497be5a011fbf710fa3479823f4beb553f8431de0616c9e325fb.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 289, + 39, + 312, + 84 + ], + "blocks": [ + { + "bbox": [ + 289, + 39, + 312, + 84 + ], + "lines": [ + { + "bbox": [ + 289, + 39, + 312, + 84 + ], + "spans": [ + { + "bbox": [ + 289, + 39, + 312, + 84 + ], + "type": "image", + "image_path": "a3a3f28d0a4ccef1a37a731ad61cf8c106f8e177c0fcbdc714e45cf4f4fe7873.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 289, + 86, + 311, + 130 + ], + "blocks": [ + { + "bbox": [ + 289, + 86, + 311, + 130 + ], + "lines": [ + { + "bbox": [ + 289, + 86, + 311, + 130 + ], + "spans": [ + { + "bbox": [ + 289, + 86, + 311, + 130 + ], + "type": "image", + "image_path": "bc36c9703fadb2381fb1bb0d7a1e4266c978b00e1af3cad0823386b2d3fe1dc2.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 289, + 131, + 311, + 176 + ], + "blocks": [ + { + "bbox": [ + 289, + 131, + 311, + 176 + ], + "lines": [ + { + "bbox": [ + 289, + 131, + 311, + 176 + ], + "spans": [ + { + "bbox": [ + 289, + 131, + 311, + 176 + ], + "type": "image", + "image_path": "1be4200de4ea6c17e870d6863f7880617d427e16e2c652b7ce1e4f91d8ffcdc6.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 318, + 91, + 356, + 127 + ], + "blocks": [ + { + "bbox": [ + 318, + 91, + 356, + 127 + ], + "lines": [ + { + "bbox": [ + 318, + 91, + 356, + 127 + ], + "spans": [ + { + "bbox": [ + 318, + 91, + 356, + 127 + ], + "type": "image", + "image_path": "4fc146c66927e528bc3e876fb7431f467163179cf4c51d7a5b608a95525f493b.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 322, + 132, + 351, + 177 + ], + "blocks": [ + { + "bbox": [ + 322, + 132, + 351, + 177 + ], + "lines": [ + { + "bbox": [ + 322, + 132, + 351, + 177 + ], + "spans": [ + { + "bbox": [ + 322, + 132, + 351, + 177 + ], + "type": "image", + "image_path": "cc4c784cf2f9a14fa0c87b80e2b7651af25b75bfc9486290a6edda5d04056d98.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 357, + 40, + 379, + 130 + ], + "blocks": [ + { + "bbox": [ + 324, + 33, + 359, + 39 + ], + "lines": [ + { + "bbox": [ + 324, + 33, + 359, + 39 + ], + "spans": [ + { + "bbox": [ + 324, + 33, + 359, + 39 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 357, + 40, + 379, + 130 + ], + "lines": [ + { + "bbox": [ + 357, + 40, + 379, + 130 + ], + "spans": [ + { + "bbox": [ + 357, + 40, + 379, + 130 + ], + "type": "image", + "image_path": "72105cb9fdaf593c430f38330d6f26912ec829f2c9a837c161f5af898488df09.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 357, + 131, + 379, + 176 + ], + "blocks": [ + { + "bbox": [ + 357, + 131, + 379, + 176 + ], + "lines": [ + { + "bbox": [ + 357, + 131, + 379, + 176 + ], + "spans": [ + { + "bbox": [ + 357, + 131, + 379, + 176 + ], + "type": "image", + "image_path": "a1383c552b532b3c43356e2bdc786b6ce4539fec7c167e1fbf2dfbba400d0400.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "bbox": [ + 30, + 259, + 379, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 259, + 379, + 296 + ], + "spans": [ + { + "bbox": [ + 30, + 259, + 379, + 296 + ], + "type": "text", + "content": "to note that 3DGS struggles to model reflective surfaces effectively (Fig. 6a). Moreover, they have a larger memory overhead than us. " + }, + { + "bbox": [ + 30, + 259, + 379, + 296 + ], + "type": "inline_equation", + "content": "E.g." + }, + { + "bbox": [ + 30, + 259, + 379, + 296 + ], + "type": "text", + "content": ", for gardenspheres, the memory cost for 3DGS is 1.4GB, whereas our method only requires 84.3MB." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 30, + 303, + 379, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 303, + 379, + 412 + ], + "spans": [ + { + "bbox": [ + 30, + 303, + 379, + 412 + ], + "type": "text", + "content": "In comparison to mesh-based real-time rendering methods [9, 35, 45], we achieve optimal rendering quality in most object-centric datasets, as shown in Fig. 5. Despite having mask supervision on object-centric datasets as well, NvDiffRec [35] and NeRF2Mesh [45] struggle to achieve high rendering quality. This could be due to the fact that NvDiffRec relies on the simplified rendering equation, whereas NeRF2Mesh's simple color formulation struggles to model highly reflective objects. Note that NeRF2Mesh also requires an initial mesh in its second stage, and we provide NeRF2Mesh with the same initial mesh as ours for all experiments for a fair comparison." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 30, + 420, + 379, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 420, + 379, + 516 + ], + "spans": [ + { + "bbox": [ + 30, + 420, + 379, + 516 + ], + "type": "text", + "content": "In unbounded outdoor scenes without masks, our method still achieves high-quality rendering, as demonstrated in Fig. 6a. Although our PSNR may be slightly lower compared to the baselines, we outperform them in rendering foreground glossy objects. However, this advantage is not accurately reflected in the PSNR metric since the foreground reflective objects occupy only a small region in the image. Further, our method mainly exhibits artifacts in the background, this is due to the fact that the initial mesh can only faithfully reconstruct the foreground but not the background, see supplementary for more details." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 30, + 523, + 379, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 523, + 379, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 523, + 379, + 584 + ], + "type": "text", + "content": "Surface Quality. Furthermore, we observe that our method not only improves the rendering quality when modeling glossy objects but also generates more accurate meshes, as shown in Fig. 6b. It can be seen that the meshes obtained from [35,45] often have poor quality in the reflective region while our method is capable of accurately reconstructing the geometry." + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 9, + 346, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 9, + 346, + 19 + ], + "spans": [ + { + "bbox": [ + 297, + 9, + 346, + 19 + ], + "type": "text", + "content": "REFRAME" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 369, + 10, + 378, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 10, + 378, + 18 + ], + "spans": [ + { + "bbox": [ + 369, + 10, + 378, + 18 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 33, + 32, + 75, + 83 + ], + "blocks": [ + { + "bbox": [ + 33, + 32, + 75, + 83 + ], + "lines": [ + { + "bbox": [ + 33, + 32, + 75, + 83 + ], + "spans": [ + { + "bbox": [ + 33, + 32, + 75, + 83 + ], + "type": "image", + "image_path": "36f40a5b6f218bf1d38b78034539f0f0c398d13ca834fa2291a5fd91a3214f60.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 33, + 86, + 75, + 137 + ], + "blocks": [ + { + "bbox": [ + 33, + 86, + 75, + 137 + ], + "lines": [ + { + "bbox": [ + 33, + 86, + 75, + 137 + ], + "spans": [ + { + "bbox": [ + 33, + 86, + 75, + 137 + ], + "type": "image", + "image_path": "511a37a008c361b3463a77a4d8274a538ffbb663e1449712d42357d100e19e52.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 33, + 138, + 165, + 155 + ], + "lines": [ + { + "bbox": [ + 33, + 138, + 165, + 155 + ], + "spans": [ + { + "bbox": [ + 33, + 138, + 165, + 155 + ], + "type": "text", + "content": "(a) Rendering Quality on Real Captured Dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 78, + 33, + 120, + 83 + ], + "blocks": [ + { + "bbox": [ + 78, + 33, + 120, + 83 + ], + "lines": [ + { + "bbox": [ + 78, + 33, + 120, + 83 + ], + "spans": [ + { + "bbox": [ + 78, + 33, + 120, + 83 + ], + "type": "image", + "image_path": "d45d5777842ddfa6612b8b778c0eb5e2af6f2d2e14eb86bf9b75a6179fb31831.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 78, + 86, + 120, + 137 + ], + "blocks": [ + { + "bbox": [ + 78, + 86, + 120, + 137 + ], + "lines": [ + { + "bbox": [ + 78, + 86, + 120, + 137 + ], + "spans": [ + { + "bbox": [ + 78, + 86, + 120, + 137 + ], + "type": "image", + "image_path": "9c1804bdc99372cd289fe61aa7eb59e794efc80bc410df0ce3d1f33bb151e881.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 30, + 165, + 379, + 188 + ], + "lines": [ + { + "bbox": [ + 30, + 165, + 379, + 188 + ], + "spans": [ + { + "bbox": [ + 30, + 165, + 379, + 188 + ], + "type": "text", + "content": "Fig. 6: Qualitative Comparison. Compared to the baselines, our method reconstructs reflective regions of higher fidelity and yields more accurate surface geometry." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 122, + 33, + 165, + 83 + ], + "blocks": [ + { + "bbox": [ + 122, + 33, + 165, + 83 + ], + "lines": [ + { + "bbox": [ + 122, + 33, + 165, + 83 + ], + "spans": [ + { + "bbox": [ + 122, + 33, + 165, + 83 + ], + "type": "image", + "image_path": "6044dffec518b45ba37cafcae18efeb5ddd1affdad15f7433d851340ad457730.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 123, + 86, + 165, + 137 + ], + "blocks": [ + { + "bbox": [ + 123, + 86, + 165, + 137 + ], + "lines": [ + { + "bbox": [ + 123, + 86, + 165, + 137 + ], + "spans": [ + { + "bbox": [ + 123, + 86, + 165, + 137 + ], + "type": "image", + "image_path": "2e870c184d25a65cabd834a6c84235a932f94a5cf52043fa99727a5e41f94aa5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 170, + 33, + 244, + 145 + ], + "blocks": [ + { + "bbox": [ + 170, + 33, + 244, + 145 + ], + "lines": [ + { + "bbox": [ + 170, + 33, + 244, + 145 + ], + "spans": [ + { + "bbox": [ + 170, + 33, + 244, + 145 + ], + "type": "image", + "image_path": "18d7cbe5d651a4cf241340fc313b9e74d0849b3f6c0bbe68350ac691fd91207f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 185, + 146, + 364, + 156 + ], + "lines": [ + { + "bbox": [ + 185, + 146, + 364, + 156 + ], + "spans": [ + { + "bbox": [ + 185, + 146, + 364, + 156 + ], + "type": "text", + "content": "(b) Surface Quality on NeRF Synthetic Dataset." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 249, + 33, + 311, + 145 + ], + "blocks": [ + { + "bbox": [ + 249, + 33, + 311, + 145 + ], + "lines": [ + { + "bbox": [ + 249, + 33, + 311, + 145 + ], + "spans": [ + { + "bbox": [ + 249, + 33, + 311, + 145 + ], + "type": "image", + "image_path": "a5b217db35b9b2a8ca247afcc6e0c815c2109a9813c6a9bb30964b91b62d1255.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 315, + 33, + 379, + 145 + ], + "blocks": [ + { + "bbox": [ + 315, + 33, + 379, + 145 + ], + "lines": [ + { + "bbox": [ + 315, + 33, + 379, + 145 + ], + "spans": [ + { + "bbox": [ + 315, + 33, + 379, + 145 + ], + "type": "image", + "image_path": "8642dd161fa89993d20b0b3e40f8424b44fa14f621eea6d57be3e387662a8df4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 35, + 198, + 377, + 249 + ], + "blocks": [ + { + "bbox": [ + 35, + 198, + 377, + 249 + ], + "lines": [ + { + "bbox": [ + 35, + 198, + 377, + 249 + ], + "spans": [ + { + "bbox": [ + 35, + 198, + 377, + 249 + ], + "type": "table", + "html": "
MacBook Pro iPhone12 iPad Air3 Legion Y-7000 HUAWEI nova7
MobileNeRF [9]120.0060.0060.0055.3139.19
NeRF2Mesh [45]120.0060.0060.0048.6932.56
Ours (Mobile)120.0060.0060.0048.5031.71
", + "image_path": "c9173e073e3a041a4ec867f872cb526527029b801562924b3e25816fb8d905de.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 30, + 249, + 379, + 293 + ], + "lines": [ + { + "bbox": [ + 30, + 249, + 379, + 293 + ], + "spans": [ + { + "bbox": [ + 30, + 249, + 379, + 293 + ], + "type": "text", + "content": "Table 2: Computational Efficiency Comparison. We test the FPS of our method and the baseline methods [9,45] on the NeRF Synthetic dataset across multiple devices. On MacBook Pro, iPhone 12, and iPad Air3, both our method and the baseline methods [9,45] achieve the maximum screen refresh rate." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 31, + 326, + 163, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 326, + 163, + 338 + ], + "spans": [ + { + "bbox": [ + 31, + 326, + 163, + 338 + ], + "type": "text", + "content": "4.2 Rendering Efficiency" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 30, + 345, + 379, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 345, + 379, + 428 + ], + "spans": [ + { + "bbox": [ + 30, + 345, + 379, + 428 + ], + "type": "text", + "content": "Computational efficiency. During the mobile rendering stage, the bottleneck of our rendering speed lies in querying the tiny MLP " + }, + { + "bbox": [ + 30, + 345, + 379, + 428 + ], + "type": "inline_equation", + "content": "f_{\\psi}" + }, + { + "bbox": [ + 30, + 345, + 379, + 428 + ], + "type": "text", + "content": " to obtain the specular color. However, due to our MLP being only two layers with 64 widths, our rendering speed remains relatively fast. Therefore, our method achieves real-time rendering on both consumer GPUs and mobile devices. We compare the rendering speed with the baselines [9,45] that can render in real-time on mobile devices across multiple platform devices, as seen in Tab. 2." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 30, + 428, + 379, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 428, + 379, + 476 + ], + "spans": [ + { + "bbox": [ + 30, + 428, + 379, + 476 + ], + "type": "text", + "content": "Memory efficiency. Our cache mainly consists of texture images, the environment feature map, and the mesh. Due to our high-quality estimation of vertex normals, our mesh can achieve high-quality rendering with a low number of vertices and faces, leading to a small cache size in most scenes, as shown in Tab. 3." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 31, + 504, + 137, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 504, + 137, + 517 + ], + "spans": [ + { + "bbox": [ + 31, + 504, + 137, + 517 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 30, + 523, + 379, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 523, + 379, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 523, + 379, + 584 + ], + "type": "text", + "content": "Geometry Learner. [54] and [45] utilize direct gradient backpropagation to update the vertex positions of the mesh. We employ this update strategy in our method and verify that this update method lacks robustness. Directly using the backpropagated gradients to update the vertex positions can lead to a devastating blow to the mesh, especially when using a large learning rate. This" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 9, + 107, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 9, + 107, + 19 + ], + "spans": [ + { + "bbox": [ + 66, + 9, + 107, + 19 + ], + "type": "text", + "content": "C.Ji et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 32, + 33, + 383, + 94 + ], + "blocks": [ + { + "bbox": [ + 32, + 33, + 383, + 94 + ], + "lines": [ + { + "bbox": [ + 32, + 33, + 383, + 94 + ], + "spans": [ + { + "bbox": [ + 32, + 33, + 383, + 94 + ], + "type": "table", + "html": "
NeRF Synthetic DatasetShiny Blender Dataset
#V (103)#F (103)Cache#V (103)#F (103)Cache
MobileNeRF [9]494224125.751028343152.37
NeRF2Mesh [45]20019273.53459122.15
Ours (Mobile)377547.55387552.86
", + "image_path": "cc6dc89390584d9bc0807e600edc51de9a1dcdb5ebcdc12836b6e43570b22b17.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 44, + 140, + 192, + 187 + ], + "blocks": [ + { + "bbox": [ + 31, + 94, + 379, + 117 + ], + "lines": [ + { + "bbox": [ + 31, + 94, + 379, + 117 + ], + "spans": [ + { + "bbox": [ + 31, + 94, + 379, + 117 + ], + "type": "text", + "content": "Table 3: Memory Efficiency. While exhibiting higher quality, our method maintains a relatively low memory consumption compared to baselines. Cache reported in MB." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 44, + 140, + 192, + 187 + ], + "lines": [ + { + "bbox": [ + 44, + 140, + 192, + 187 + ], + "spans": [ + { + "bbox": [ + 44, + 140, + 192, + 187 + ], + "type": "table", + "html": "
w/o gθvOurs
L=0.001L=0.0001L=0.001
Toaster17.0117.0124.95
Materials23.0129.0629.58
", + "image_path": "1cdc3eb44d80f5f294296d3792749725a8dba6536d918db0712c5ff46a64a839.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 228, + 140, + 366, + 198 + ], + "blocks": [ + { + "bbox": [ + 31, + 188, + 199, + 220 + ], + "lines": [ + { + "bbox": [ + 31, + 188, + 199, + 220 + ], + "spans": [ + { + "bbox": [ + 31, + 188, + 199, + 220 + ], + "type": "text", + "content": "Table 4: Ablation of Vertex Offset Learning. L represents the learning rate. Reported in PSNR." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 228, + 140, + 366, + 198 + ], + "lines": [ + { + "bbox": [ + 228, + 140, + 366, + 198 + ], + "spans": [ + { + "bbox": [ + 228, + 140, + 366, + 198 + ], + "type": "table", + "html": "
w/o Δn_vOurs
Toaster#F=10k24.2424.62
#F=75k24.5524.95
Materials#F=10k27.6928.23
#F=75k29.4629.58
", + "image_path": "ebcc2be1bcd99212021c070d066485ff89517aeda249f8b183abc0d520bb5d40.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 211, + 199, + 379, + 221 + ], + "lines": [ + { + "bbox": [ + 211, + 199, + 379, + 221 + ], + "spans": [ + { + "bbox": [ + 211, + 199, + 379, + 221 + ], + "type": "text", + "content": "Table 5: Ablation of Normal Offset Learning. Reported in PSNR." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 30, + 257, + 379, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 257, + 379, + 317 + ], + "spans": [ + { + "bbox": [ + 30, + 257, + 379, + 317 + ], + "type": "text", + "content": "significantly compromises the rendering quality, as shown in Tab. 4. Additionally, different datasets require different learning rates, as indicated in Tab. 4. While a learning rate of 0.0001 works fine for the materials dataset, it performs poorly on the toaster dataset. In contrast, our vertex update strategy not only exhibits robustness across different datasets but also provides higher rendering quality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 30, + 318, + 380, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 318, + 380, + 511 + ], + "spans": [ + { + "bbox": [ + 30, + 318, + 380, + 511 + ], + "type": "text", + "content": "In order to verify the necessity of estimating an offset for the vertices normal, we conduct experiments with the w/o normal offset setting. In this setting, we do not estimate a normal offset for each vertex, but we still compute an updated normal based on the updated vertex positions. As shown in Tab. 5, introducing per-vertex normal offset provides a gain in reconstruction quality. This gain is particularly significant when the number of faces is low. Notably, in the case of the toaster dataset, a mesh with " + }, + { + "bbox": [ + 30, + 318, + 380, + 511 + ], + "type": "inline_equation", + "content": "10\\mathrm{k}" + }, + { + "bbox": [ + 30, + 318, + 380, + 511 + ], + "type": "text", + "content": " faces with normal offset even achieves rendering quality surpassing that of a mesh with " + }, + { + "bbox": [ + 30, + 318, + 380, + 511 + ], + "type": "inline_equation", + "content": "75\\mathrm{k}" + }, + { + "bbox": [ + 30, + 318, + 380, + 511 + ], + "type": "text", + "content": " faces without normal offset. Environment Learner. To validate the effectiveness of the environment learner " + }, + { + "bbox": [ + 30, + 318, + 380, + 511 + ], + "type": "inline_equation", + "content": "f_{e}" + }, + { + "bbox": [ + 30, + 318, + 380, + 511 + ], + "type": "text", + "content": ", we directly input " + }, + { + "bbox": [ + 30, + 318, + 380, + 511 + ], + "type": "inline_equation", + "content": "\\omega_{r}" + }, + { + "bbox": [ + 30, + 318, + 380, + 511 + ], + "type": "text", + "content": " values into the tiny MLP " + }, + { + "bbox": [ + 30, + 318, + 380, + 511 + ], + "type": "inline_equation", + "content": "f_{\\psi}" + }, + { + "bbox": [ + 30, + 318, + 380, + 511 + ], + "type": "text", + "content": " as shown in Eq. (6), similar to [9] and [45], except they use viewing direction instead of reflective direction. Experimental results demonstrate that directly inputting " + }, + { + "bbox": [ + 30, + 318, + 380, + 511 + ], + "type": "inline_equation", + "content": "\\omega_{r}" + }, + { + "bbox": [ + 30, + 318, + 380, + 511 + ], + "type": "text", + "content": " into the tiny MLP " + }, + { + "bbox": [ + 30, + 318, + 380, + 511 + ], + "type": "inline_equation", + "content": "f_{\\psi}" + }, + { + "bbox": [ + 30, + 318, + 380, + 511 + ], + "type": "text", + "content": " cannot learn the correct reflection appearance as a result of the limited network's capacity, as depicted in Fig. 7. Therefore, introducing the environment learner module improves the reconstruction quality without compromising the inference speed, as demonstrated in Tab. 6." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 30, + 511, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 511, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 511, + 380, + 584 + ], + "type": "text", + "content": "Environment Feature Map: We can directly optimize an environment feature map instead of learning it through the environment learner module and baking it. However, as shown in Tab. 7, the direct optimization of the environment feature map performs poorly when the map resolution is high. This is because, at higher resolutions, neighboring directions are mapped to distant grid points, leading to a non-smooth interpolation. As a result, the grid points can-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 9, + 346, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 9, + 346, + 19 + ], + "spans": [ + { + "bbox": [ + 297, + 9, + 346, + 19 + ], + "type": "text", + "content": "REFRAME" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 46, + 36, + 88, + 81 + ], + "blocks": [ + { + "bbox": [ + 46, + 36, + 88, + 81 + ], + "lines": [ + { + "bbox": [ + 46, + 36, + 88, + 81 + ], + "spans": [ + { + "bbox": [ + 46, + 36, + 88, + 81 + ], + "type": "image", + "image_path": "9d790918db044f75c38db9020eb8d8e1b448e8a4729afbb55dadedb6283348da.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 49, + 89, + 88, + 99 + ], + "lines": [ + { + "bbox": [ + 49, + 89, + 88, + 99 + ], + "spans": [ + { + "bbox": [ + 49, + 89, + 88, + 99 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 49, + 89, + 88, + 99 + ], + "type": "inline_equation", + "content": "\\mathrm{W / o}f_{e}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 121, + 36, + 163, + 81 + ], + "blocks": [ + { + "bbox": [ + 121, + 36, + 163, + 81 + ], + "lines": [ + { + "bbox": [ + 121, + 36, + 163, + 81 + ], + "spans": [ + { + "bbox": [ + 121, + 36, + 163, + 81 + ], + "type": "image", + "image_path": "04f9ad22e7b71099b85a84668bd4296fbb264a3853a342505b9988461b02dfe0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 89, + 159, + 99 + ], + "lines": [ + { + "bbox": [ + 127, + 89, + 159, + 99 + ], + "spans": [ + { + "bbox": [ + 127, + 89, + 159, + 99 + ], + "type": "text", + "content": "(b) Ours" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 31, + 108, + 251, + 130 + ], + "lines": [ + { + "bbox": [ + 31, + 108, + 251, + 130 + ], + "spans": [ + { + "bbox": [ + 31, + 108, + 251, + 130 + ], + "type": "text", + "content": "Fig. 7: Ablation of Environment Learner. Eliminating " + }, + { + "bbox": [ + 31, + 108, + 251, + 130 + ], + "type": "inline_equation", + "content": "f_{e}" + }, + { + "bbox": [ + 31, + 108, + 251, + 130 + ], + "type": "text", + "content": " results in bad rendering quality." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 192, + 36, + 234, + 81 + ], + "blocks": [ + { + "bbox": [ + 192, + 36, + 234, + 81 + ], + "lines": [ + { + "bbox": [ + 192, + 36, + 234, + 81 + ], + "spans": [ + { + "bbox": [ + 192, + 36, + 234, + 81 + ], + "type": "image", + "image_path": "e6bbdcc3e5b8faa4aafb69e096b2951c29ebd5478be512171be14075874aacb9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 183, + 89, + 246, + 98 + ], + "lines": [ + { + "bbox": [ + 183, + 89, + 246, + 98 + ], + "spans": [ + { + "bbox": [ + 183, + 89, + 246, + 98 + ], + "type": "text", + "content": "(c) Ground Truth" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 257, + 40, + 371, + 80 + ], + "blocks": [ + { + "bbox": [ + 257, + 40, + 371, + 80 + ], + "lines": [ + { + "bbox": [ + 257, + 40, + 371, + 80 + ], + "spans": [ + { + "bbox": [ + 257, + 40, + 371, + 80 + ], + "type": "table", + "html": "
w/o feOurs
Toaster19.3124.95
Materials27.3729.58
", + "image_path": "9e89722c77c776fe26b4aca86dddeee80bc22fe5789e5ad8ff8dab9dc459506d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 45, + 140, + 366, + 212 + ], + "blocks": [ + { + "bbox": [ + 252, + 89, + 375, + 121 + ], + "lines": [ + { + "bbox": [ + 252, + 89, + 375, + 121 + ], + "spans": [ + { + "bbox": [ + 252, + 89, + 375, + 121 + ], + "type": "text", + "content": "Table 6: Ablation of Environment Learner. Reported in PSNR." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 45, + 140, + 366, + 212 + ], + "lines": [ + { + "bbox": [ + 45, + 140, + 366, + 212 + ], + "spans": [ + { + "bbox": [ + 45, + 140, + 366, + 212 + ], + "type": "table", + "html": "
Resolutions H*WOursDirectly optimization
PSNRCache(KB)PSNRCache(KB)
Toaster180*36024.864924.1691
900*180024.9650423.592460
Materials180*36029.556329.58114
900*180029.5870629.033000
", + "image_path": "3bde7766e825ead4ffee6125ca3844665b163501904078060159e55222299b45.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 31, + 213, + 379, + 236 + ], + "lines": [ + { + "bbox": [ + 31, + 213, + 379, + 236 + ], + "spans": [ + { + "bbox": [ + 31, + 213, + 379, + 236 + ], + "type": "text", + "content": "Table 7: Ablation of Environment Feature Map. Our method outperforms in both cache size and rendering quality." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 30, + 266, + 379, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 266, + 379, + 362 + ], + "spans": [ + { + "bbox": [ + 30, + 266, + 379, + 362 + ], + "type": "text", + "content": "not share global information and only contain its own local information. On the other hand, when the resolution is too low, the rendering quality is limited by the capacity of the environment feature map. In contrast, our method exhibits good rendering quality across different resolutions by learning the environment feature map via an MLP. Additionally, our environment feature maps enjoy the smoothness bias of the MLP, resulting in relatively smooth tensor maps. Consequently, when saving images in PNG format, our method demonstrates good compression performance. This further reduces the overhead of our caching." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 31, + 378, + 118, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 378, + 118, + 390 + ], + "spans": [ + { + "bbox": [ + 31, + 378, + 118, + 390 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 30, + 401, + 379, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 401, + 379, + 448 + ], + "spans": [ + { + "bbox": [ + 30, + 401, + 379, + 448 + ], + "type": "text", + "content": "Although our method successfully models the appearance of glossy objects, similar to Ref-NeRF [46], it still struggles with accurately modeling interreflections and non-distant illumination. Additionally, our method requires a certain level of quality in the initial mesh." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 30, + 449, + 379, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 449, + 379, + 521 + ], + "spans": [ + { + "bbox": [ + 30, + 449, + 379, + 521 + ], + "type": "text", + "content": "In summary, our method achieves real-time rendering on edge devices with low hardware budgets. Besides, we propose a novel approach for modeling view-dependent appearance and can optimize the appearance and geometry of glossy objects with high computational efficiency and low memory footprint. Furthermore, by effectively decoupling the scene's geometry, appearance, and environmental information, our method can perform simple scene editing tasks." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 32, + 537, + 145, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 537, + 145, + 551 + ], + "spans": [ + { + "bbox": [ + 32, + 537, + 145, + 551 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 30, + 559, + 379, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 559, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 30, + 559, + 379, + 583 + ], + "type": "text", + "content": "This work is supported by NSFC under grant 62202418 and U21B2004. We thank Xiaoran Cao, Chenxi Tu, and Sicheng Li for their valuable discussions." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 9, + 107, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 9, + 107, + 19 + ], + "spans": [ + { + "bbox": [ + 66, + 9, + 107, + 19 + ], + "type": "text", + "content": "C.Ji et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 33, + 98, + 46 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 33, + 98, + 46 + ], + "spans": [ + { + "bbox": [ + 32, + 33, + 98, + 46 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 33, + 60, + 380, + 584 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 36, + 60, + 379, + 104 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 60, + 379, + 104 + ], + "spans": [ + { + "bbox": [ + 36, + 60, + 379, + 104 + ], + "type": "text", + "content": "1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 36, + 105, + 379, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 105, + 379, + 149 + ], + "spans": [ + { + "bbox": [ + 36, + 105, + 379, + 149 + ], + "type": "text", + "content": "2. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mipnerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 37, + 150, + 380, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 150, + 380, + 183 + ], + "spans": [ + { + "bbox": [ + 37, + 150, + 380, + 183 + ], + "type": "text", + "content": "3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-nerf: Anti-aliased grid-based neural radiance fields. arXiv preprint arXiv:2304.06706 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 37, + 183, + 379, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 183, + 379, + 216 + ], + "spans": [ + { + "bbox": [ + 37, + 183, + 379, + 216 + ], + "type": "text", + "content": "4. Bi, S., Xu, Z., Srinivasan, P., Mildenhall, B., Sunkavalli, K., Hasan, M., Hold-Geoffroy, Y., Kriegman, D., Ramamoorthi, R.: Neural reflectance fields for appearance acquisition. arXiv preprint arXiv:2008.03824 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 37, + 217, + 379, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 217, + 379, + 250 + ], + "spans": [ + { + "bbox": [ + 37, + 217, + 379, + 250 + ], + "type": "text", + "content": "5. Boss, M., Braun, R., Jampani, V., Barron, J.T., Liu, C., Lensch, H.: Nerd: Neural reflectance decomposition from image collections. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12684-12694 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 37, + 250, + 379, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 250, + 379, + 284 + ], + "spans": [ + { + "bbox": [ + 37, + 250, + 379, + 284 + ], + "type": "text", + "content": "6. Boss, M., Jampani, V., Braun, R., Liu, C., Barron, J., Lensch, H.: Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems 34, 10691-10704 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 37, + 284, + 379, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 284, + 379, + 328 + ], + "spans": [ + { + "bbox": [ + 37, + 284, + 379, + 328 + ], + "type": "text", + "content": "7. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 37, + 328, + 379, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 328, + 379, + 372 + ], + "spans": [ + { + "bbox": [ + 37, + 328, + 379, + 372 + ], + "type": "text", + "content": "8. Chan, E.R., Monteiro, M., Kellnhofer, P., Wu, J., Wetzstein, G.: pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5799-5809 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 37, + 372, + 379, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 372, + 379, + 416 + ], + "spans": [ + { + "bbox": [ + 37, + 372, + 379, + 416 + ], + "type": "text", + "content": "9. Chen, Z., Funkhouser, T., Hedman, P., Tagliasacchi, A.: Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16569-16578 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 33, + 417, + 379, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 417, + 379, + 450 + ], + "spans": [ + { + "bbox": [ + 33, + 417, + 379, + 450 + ], + "type": "text", + "content": "10. Esposito, S., Baieri, D., Zellmann, S., Hinkenjann, A., Rodola, E.: Kiloneus: A versatile neural implicit surface representation for real-time rendering. arXiv preprint arXiv:2206.10885 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 33, + 450, + 379, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 450, + 379, + 484 + ], + "spans": [ + { + "bbox": [ + 33, + 450, + 379, + 484 + ], + "type": "text", + "content": "11. Fu, X., Zhang, S., Chen, T., Lu, Y., Zhu, L., Zhou, X., Geiger, A., Liao, Y.: Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation. In: 2022 International Conference on 3D Vision (3DV). pp. 1-11. IEEE (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 33, + 484, + 379, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 484, + 379, + 517 + ], + "spans": [ + { + "bbox": [ + 33, + 484, + 379, + 517 + ], + "type": "text", + "content": "12. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: Fastnerf: High-fidelity neural rendering at 200fps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14346-14355 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 33, + 517, + 379, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 517, + 379, + 550 + ], + "spans": [ + { + "bbox": [ + 33, + 517, + 379, + 550 + ], + "type": "text", + "content": "13. Garland, M., Heckbert, P.S.: Surface simplification using quadric error metrics. In: Proceedings of the 24th annual conference on Computer graphics and interactive techniques. pp. 209-216 (1997)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 33, + 551, + 379, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 551, + 379, + 584 + ], + "spans": [ + { + "bbox": [ + 33, + 551, + 379, + 584 + ], + "type": "text", + "content": "14. Ge, W., Hu, T., Zhao, H., Liu, S., Chen, Y.C.: Ref-neus: Ambiguity-reduced neural implicit surface learning for multi-view reconstruction with reflection. arXiv preprint arXiv:2303.10840 (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 9, + 346, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 9, + 346, + 19 + ], + "spans": [ + { + "bbox": [ + 297, + 9, + 346, + 19 + ], + "type": "text", + "content": "REFRAME" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 33, + 35, + 379, + 583 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 33, + 35, + 379, + 68 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 35, + 379, + 68 + ], + "spans": [ + { + "bbox": [ + 33, + 35, + 379, + 68 + ], + "type": "text", + "content": "15. Guo, Y.C., Cao, Y.P., Wang, C., He, Y., Shan, Y., Zhang, S.H.: Vmesh: Hybrid volume-mesh representation for efficient view synthesis. In: SIGGRAPH Asia 2023 Conference Papers. pp. 1-11 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 33, + 68, + 379, + 100 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 68, + 379, + 100 + ], + "spans": [ + { + "bbox": [ + 33, + 68, + 379, + 100 + ], + "type": "text", + "content": "16. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking neural radiance fields for real-time view synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5875-5884 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 101, + 379, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 101, + 379, + 123 + ], + "spans": [ + { + "bbox": [ + 33, + 101, + 379, + 123 + ], + "type": "text", + "content": "17. Hore, A., Ziou, D.: Image quality metrics: Psnr vs. sSIM. In: 2010 20th international conference on pattern recognition. pp. 2366-2369. IEEE (2010)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 123, + 379, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 123, + 379, + 156 + ], + "spans": [ + { + "bbox": [ + 33, + 123, + 379, + 156 + ], + "type": "text", + "content": "18. Hu, T., Liu, S., Chen, Y., Shen, T., Jia, J.: Efficientnerf efficient neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12902-12911 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 156, + 379, + 188 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 156, + 379, + 188 + ], + "spans": [ + { + "bbox": [ + 33, + 156, + 379, + 188 + ], + "type": "text", + "content": "19. Jin, H., Liu, I., Xu, P., Zhang, X., Han, S., Bi, S., Zhou, X., Xu, Z., Su, H.: Tensoroir: Tensorial inverse rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 165-174 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 189, + 379, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 189, + 379, + 210 + ], + "spans": [ + { + "bbox": [ + 33, + 189, + 379, + 210 + ], + "type": "text", + "content": "20. Kajiya, J.T.: The rendering equation. In: Proceedings of the 13th annual conference on Computer graphics and interactive techniques. pp. 143-150 (1986)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 33, + 211, + 379, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 211, + 379, + 233 + ], + "spans": [ + { + "bbox": [ + 33, + 211, + 379, + 233 + ], + "type": "text", + "content": "21. Kautz, J., McCool, M.D.: Approximation of glossy reflection with prefiltered environment maps. In: Graphics Interface. vol. 2000, pp. 119-126 (2000)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 33, + 233, + 379, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 233, + 379, + 254 + ], + "spans": [ + { + "bbox": [ + 33, + 233, + 379, + 254 + ], + "type": "text", + "content": "22. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 33, + 255, + 379, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 255, + 379, + 276 + ], + "spans": [ + { + "bbox": [ + 33, + 255, + 379, + 276 + ], + "type": "text", + "content": "23. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 33, + 277, + 379, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 277, + 379, + 321 + ], + "spans": [ + { + "bbox": [ + 33, + 277, + 379, + 321 + ], + "type": "text", + "content": "24. Kundu, A., Genova, K., Yin, X., Fathi, A., Pantofaru, C., Guibas, L.J., Tagliasacchi, A., Dellaert, F., Funkhouser, T.: Panoptic neural fields: A semantic object-aware neural scene representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12871-12881 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 33, + 321, + 379, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 321, + 379, + 364 + ], + "spans": [ + { + "bbox": [ + 33, + 321, + 379, + 364 + ], + "type": "text", + "content": "25. Kurz, A., Neff, T., Lv, Z., Zollhöfer, M., Steinberger, M.: Adanerf: Adaptive sampling for real-time rendering of neural radiance fields. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII. pp. 254-270. Springer (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 33, + 364, + 379, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 364, + 379, + 397 + ], + "spans": [ + { + "bbox": [ + 33, + 364, + 379, + 397 + ], + "type": "text", + "content": "26. Laine, S., Hellsten, J., Karras, T., Seol, Y., Lehtinen, J., Aila, T.: Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG) 39(6), 1-14 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 33, + 397, + 379, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 397, + 379, + 429 + ], + "spans": [ + { + "bbox": [ + 33, + 397, + 379, + 429 + ], + "type": "text", + "content": "27. Li, C., Li, S., Zhao, Y., Zhu, W., Lin, Y.: Rt-nerf: Real-time on-device neural radiance fields towards immersive ar/vr rendering. In: Proceedings of the 41st IEEE/ACM International Conference on Computer-Aided Design. pp. 1-9 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 33, + 430, + 379, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 430, + 379, + 452 + ], + "spans": [ + { + "bbox": [ + 33, + 430, + 379, + 452 + ], + "type": "text", + "content": "28. Li, S., Li, H., Wang, Y., Liao, Y., Yu, L.: Steernerf: Accelerating nef rendering via smooth viewpoint trajectory. arXiv preprint arXiv:2212.08476 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 33, + 452, + 379, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 452, + 379, + 495 + ], + "spans": [ + { + "bbox": [ + 33, + 452, + 379, + 495 + ], + "type": "text", + "content": "29. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8456-8465 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 33, + 496, + 379, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 496, + 379, + 528 + ], + "spans": [ + { + "bbox": [ + 33, + 496, + 379, + 528 + ], + "type": "text", + "content": "30. Liang, R., Chen, H., Li, C., Chen, F., Panneer, S., Vijaykumar, N.: Envidr: Implicit differentiable renderer with neural environment lighting. arXiv preprint arXiv:2303.13022 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 33, + 528, + 379, + 561 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 528, + 379, + 561 + ], + "spans": [ + { + "bbox": [ + 33, + 528, + 379, + 561 + ], + "type": "text", + "content": "31. Lin, H., Peng, S., Xu, Z., Yan, Y., Shuai, Q., Bao, H., Zhou, X.: Efficient neural radiance fields for interactive free-viewpoint video. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 33, + 562, + 379, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 562, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 562, + 379, + 583 + ], + "type": "text", + "content": "32. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 107, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 107, + 19 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 107, + 19 + ], + "type": "text", + "content": "C.Ji et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 35, + 379, + 583 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 32, + 35, + 379, + 68 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 35, + 379, + 68 + ], + "spans": [ + { + "bbox": [ + 32, + 35, + 379, + 68 + ], + "type": "text", + "content": "33. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 32, + 69, + 379, + 102 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 69, + 379, + 102 + ], + "spans": [ + { + "bbox": [ + 32, + 69, + 379, + 102 + ], + "type": "text", + "content": "34. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 102, + 379, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 102, + 379, + 146 + ], + "spans": [ + { + "bbox": [ + 33, + 102, + 379, + 146 + ], + "type": "text", + "content": "35. Munkberg, J., Hasselgren, J., Shen, T., Gao, J., Chen, W., Evans, A., Müller, T., Fidler, S.: Extracting triangular 3d models, materials, and lighting from images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8280-8290 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 147, + 379, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 147, + 379, + 180 + ], + "spans": [ + { + "bbox": [ + 33, + 147, + 379, + 180 + ], + "type": "text", + "content": "36. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5865-5874 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 181, + 379, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 181, + 379, + 214 + ], + "spans": [ + { + "bbox": [ + 33, + 181, + 379, + 214 + ], + "type": "text", + "content": "37. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., MartinBrualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 215, + 379, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 215, + 379, + 247 + ], + "spans": [ + { + "bbox": [ + 33, + 215, + 379, + 247 + ], + "type": "text", + "content": "38. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10318-10327 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 33, + 249, + 379, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 249, + 379, + 281 + ], + "spans": [ + { + "bbox": [ + 33, + 249, + 379, + 281 + ], + "type": "text", + "content": "39. Rakotosaona, M.J., Manhardt, F., Arroyo, D.M., Niemeyer, M., Kundu, A., Tombari, F.: Nerfmeshing: Distilling neural radiance fields into geometrically-accurate 3d meshes. arXiv preprint arXiv:2303.09431 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 33, + 282, + 379, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 282, + 379, + 314 + ], + "spans": [ + { + "bbox": [ + 33, + 282, + 379, + 314 + ], + "type": "text", + "content": "40. Reiser, C., Peng, S., Liao, Y., Geiger, A.: Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14335-14345 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 33, + 316, + 379, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 316, + 379, + 349 + ], + "spans": [ + { + "bbox": [ + 33, + 316, + 379, + 349 + ], + "type": "text", + "content": "41. Rematas, K., Liu, A., Srinivasan, P.P., Barron, J.T., Tagliasacchi, A., Funkhouser, T., Ferrari, V.: Urban radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12932-12942 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 33, + 349, + 379, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 349, + 379, + 381 + ], + "spans": [ + { + "bbox": [ + 33, + 349, + 379, + 381 + ], + "type": "text", + "content": "42. Schwarz, K., Liao, Y., Niemeyer, M., Geiger, A.: Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems 33, 20154-20166 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 33, + 383, + 379, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 383, + 379, + 427 + ], + "spans": [ + { + "bbox": [ + 33, + 383, + 379, + 427 + ], + "type": "text", + "content": "43. Srinivasan, P.P., Deng, B., Zhang, X., Tancik, M., Mildenhall, B., Barron, J.T.: Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7495-7504 (June 2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 33, + 428, + 379, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 428, + 379, + 471 + ], + "spans": [ + { + "bbox": [ + 33, + 428, + 379, + 471 + ], + "type": "text", + "content": "44. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8248-8258 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 33, + 472, + 379, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 472, + 379, + 506 + ], + "spans": [ + { + "bbox": [ + 33, + 472, + 379, + 506 + ], + "type": "text", + "content": "45. Tang, J., Zhou, H., Chen, X., Hu, T., Ding, E., Wang, J., Zeng, G.: Delicate textured mesh recovery from nef via adaptive surface refinement. arXiv preprint arXiv:2303.02091 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 33, + 506, + 379, + 549 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 506, + 379, + 549 + ], + "spans": [ + { + "bbox": [ + 33, + 506, + 379, + 549 + ], + "type": "text", + "content": "46. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5481-5490. IEEE (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 33, + 550, + 379, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 550, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 550, + 379, + 583 + ], + "type": "text", + "content": "47. Walker, T., Mariotti, O., Vaxman, A., Bilen, H.: Explicit neural surfaces: Learning continuous geometry with deformation fields. arXiv preprint arXiv:2306.02956 (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 9, + 346, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 9, + 346, + 19 + ], + "spans": [ + { + "bbox": [ + 297, + 9, + 346, + 19 + ], + "type": "text", + "content": "REFRAME" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 31, + 35, + 379, + 583 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 33, + 35, + 379, + 79 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 35, + 379, + 79 + ], + "spans": [ + { + "bbox": [ + 33, + 35, + 379, + 79 + ], + "type": "text", + "content": "48. Wang, H., Ren, J., Huang, Z., Olszewski, K., Chai, M., Fu, Y., Tulyakov, S.: R2l: Distilling neural radiance field to neural light field for efficient novel view synthesis. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXI. pp. 612-629. Springer (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 31, + 79, + 379, + 122 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 79, + 379, + 122 + ], + "spans": [ + { + "bbox": [ + 31, + 79, + 379, + 122 + ], + "type": "text", + "content": "49. Wang, L., Zhang, J., Liu, X., Zhao, F., Zhang, Y., Zhang, Y., Wu, M., Yu, J., Xu, L.: Fourier plenoctrees for dynamic radiance field rendering in real-time. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13524-13534 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 31, + 122, + 379, + 154 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 122, + 379, + 154 + ], + "spans": [ + { + "bbox": [ + 31, + 122, + 379, + 154 + ], + "type": "text", + "content": "50. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 31, + 154, + 379, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 154, + 379, + 175 + ], + "spans": [ + { + "bbox": [ + 31, + 154, + 379, + 175 + ], + "type": "text", + "content": "51. Wang, Z., Li, L., Shen, Z., Shen, L., Bo, L.: 4k-nerf: High fidelity neural radiance fields at ultra high resolutions. arXiv preprint arXiv:2212.04701 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 31, + 175, + 379, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 175, + 379, + 218 + ], + "spans": [ + { + "bbox": [ + 31, + 175, + 379, + 218 + ], + "type": "text", + "content": "52. Wang, Z., Shen, T., Gao, J., Huang, S., Munkberg, J., Hasselgren, J., Gojcic, Z., Chen, W., Fidler, S.: Neural fields meet explicit geometric representations for inverse rendering of urban scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8370-8380 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 31, + 217, + 379, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 217, + 379, + 261 + ], + "spans": [ + { + "bbox": [ + 31, + 217, + 379, + 261 + ], + "type": "text", + "content": "53. Weng, C.Y., Curless, B., Srinivasan, P.P., Barron, J.T., Kemelmacher-Shlizerman, I.: Humannerf: Free-viewpoint rendering of moving people from monocular video. In: Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition. pp. 16210-16220 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 31, + 261, + 379, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 261, + 379, + 304 + ], + "spans": [ + { + "bbox": [ + 31, + 261, + 379, + 304 + ], + "type": "text", + "content": "54. Worchel, M., Diaz, R., Hu, W., Schreer, O., Feldmann, I., Eisert, P.: Multiview mesh reconstruction with neural deferred shading. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6187-6197 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 31, + 304, + 379, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 304, + 379, + 348 + ], + "spans": [ + { + "bbox": [ + 31, + 304, + 379, + 348 + ], + "type": "text", + "content": "55. Wu, L., Lee, J.Y., Bhattad, A., Wang, Y.X., Forsyth, D.: Diver: Real-time and accurate neural radiance fields with deterministic integration for volume rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16200-16209 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 31, + 348, + 379, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 348, + 379, + 380 + ], + "spans": [ + { + "bbox": [ + 31, + 348, + 379, + 380 + ], + "type": "text", + "content": "56. Wu, T., Sun, J.M., Lai, Y.K., Gao, L.: De-nerf: Decoupled neural radiance fields for view-consistent appearance editing and high-frequency environmental relighting. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-11 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 31, + 380, + 379, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 380, + 379, + 411 + ], + "spans": [ + { + "bbox": [ + 31, + 380, + 379, + 411 + ], + "type": "text", + "content": "57. Yariv, L., Hedman, P., Reiser, C., Verbin, D., Srinivasan, P.P., Szeliski, R., Barron, J.T., Mildenhall, B.: Bakedsdf: Meshing neural sdfs for real-time view synthesis. arXiv preprint arXiv:2302.14859 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 31, + 411, + 379, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 411, + 379, + 444 + ], + "spans": [ + { + "bbox": [ + 31, + 411, + 379, + 444 + ], + "type": "text", + "content": "58. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: Plenoctrees for real-time rendering of neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5752-5761 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 31, + 444, + 379, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 444, + 379, + 476 + ], + "spans": [ + { + "bbox": [ + 31, + 444, + 379, + 476 + ], + "type": "text", + "content": "59. Zhang, J., Yang, G., Tulsiani, S., Ramanan, D.: Ners: Neural reflectance surfaces for sparse-view 3d reconstruction in the wild. Advances in Neural Information Processing Systems 34, 29835-29847 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 31, + 476, + 379, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 476, + 379, + 507 + ], + "spans": [ + { + "bbox": [ + 31, + 476, + 379, + 507 + ], + "type": "text", + "content": "60. Zhang, J., Yao, Y., Li, S., Liu, J., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neilf++: Inter-reflectable light fields for geometry and material estimation. arXiv preprint arXiv:2303.17147 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 31, + 507, + 379, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 507, + 379, + 540 + ], + "spans": [ + { + "bbox": [ + 31, + 507, + 379, + 540 + ], + "type": "text", + "content": "61. Zhang, K., Luan, F., Li, Z., Snavely, N.: Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5565-5574 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 31, + 540, + 379, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 540, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 31, + 540, + 379, + 583 + ], + "type": "text", + "content": "62. Zhang, K., Luan, F., Wang, Q., Bala, K., Snavely, N.: Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5453-5462 (2021)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 107, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 107, + 19 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 107, + 19 + ], + "type": "text", + "content": "C.Ji et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 31, + 35, + 379, + 123 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 31, + 35, + 379, + 68 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 35, + 379, + 68 + ], + "spans": [ + { + "bbox": [ + 31, + 35, + 379, + 68 + ], + "type": "text", + "content": "63. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG) 40(6), 1-18 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 31, + 68, + 379, + 100 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 68, + 379, + 100 + ], + "spans": [ + { + "bbox": [ + 31, + 68, + 379, + 100 + ], + "type": "text", + "content": "64. Zhang, Y., Sun, J., He, X., Fu, H., Jia, R., Zhou, X.: Modeling indirect illumination for inverse rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18643-18652 (June 2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 31, + 101, + 379, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 101, + 379, + 123 + ], + "spans": [ + { + "bbox": [ + 31, + 101, + 379, + 123 + ], + "type": "text", + "content": "65. Zhi, S., Laidlow, T., Leutenegger, S., Davison, A.J.: In-place scene labelling and understanding with implicit scene representation. In: ICCV (2021)" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 9, + 346, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 9, + 346, + 19 + ], + "spans": [ + { + "bbox": [ + 297, + 9, + 346, + 19 + ], + "type": "text", + "content": "REFRAME" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/REVISION_ Rendering Tools Enable Spatial Fidelity in Vision-Language Models/f77a4b99-c39a-4693-bb9e-aff60ba50459_content_list.json b/2024/REVISION_ Rendering Tools Enable Spatial Fidelity in Vision-Language Models/f77a4b99-c39a-4693-bb9e-aff60ba50459_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a2e7c5a91a00e74e80667435bfa9374ef012b0fa --- /dev/null +++ b/2024/REVISION_ Rendering Tools Enable Spatial Fidelity in Vision-Language Models/f77a4b99-c39a-4693-bb9e-aff60ba50459_content_list.json @@ -0,0 +1,1895 @@ +[ + { + "type": "text", + "text": "REVISION: Rendering Tools Enable Spatial Fidelity in Vision-Language Models", + "text_level": 1, + "bbox": [ + 248, + 140, + 754, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Agneet Chatterjee $\\star 1$ Yiran Luo $\\star 1$ Tejas Gokhale $^{2}$ , Yezhou Yang $^{1}$ , and Chitta Baral $^{1}$", + "bbox": [ + 303, + 210, + 702, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Arizona State University \n $^{2}$ University of Maryland, Baltimore County", + "bbox": [ + 351, + 253, + 650, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Text-to-Image (T2I) and multimodal large language models (MLLMs) have been adopted in solutions for several computer vision and multimodal learning tasks. However, it has been found that such vision-language models lack the ability to correctly reason over spatial relationships. To tackle this shortcoming, we develop the REVISION framework which improves spatial fidelity in vision-language models. REVISION is a 3D rendering based pipeline that generates spatially accurate synthetic images, given a textual prompt. REVISION is an extendable framework, which currently supports $100+$ 3D assets, 11 spatial relationships, all with diverse camera perspectives and backgrounds. Leveraging images from REVISION as additional guidance in a training-free manner consistently improves the spatial consistency of T2I models across all spatial relationships, achieving competitive performance on the VISOR and T2I-CompBench benchmarks. We also design RevQA, a question-answering benchmark to evaluate the spatial reasoning abilities of MLLMs, and find that state-of-the-art models are not robust to complex spatial reasoning under adversarial settings. Our results and findings indicate that utilizing rendering-based frameworks is an effective approach for developing spatially-aware generative models. Code and data available at: https://github.com/agneet42/revision", + "bbox": [ + 259, + 325, + 743, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Text to Image $\\cdot$ Spatial Relationships $\\cdot$ Rendering Graphics", + "bbox": [ + 261, + 616, + 740, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 657, + 375, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generative vision-language models [36,44] represent a significant step towards developing multimodal systems that bridge the gap between computer vision and natural language processing. Text-to-image (T2I) models [5,38] convert text prompts to high-quality images, while multimodal large language models (MLLMs) [25,48] process images as inputs, and generate rich and coherent natural language outputs in response. As a result, these models have found diverse applications in robotics [45], image editing [17], image-to-image translation [31], and more. However, recent studies [20] and benchmarks such as", + "bbox": [ + 212, + 691, + 787, + 815 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. Correspondence to agneet@asu.edu", + "bbox": [ + 217, + 825, + 573, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7c26d986754b6ea01d566b4a44133234a2f36e3122ee36b599032b507cb86331.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 143, + 326, + 227 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9e04c3522ebd952e052c4b205571fe673306b234f82a0b5915a7a654558559e0.jpg", + "image_caption": [ + "A couch below an airplane" + ], + "image_footnote": [], + "bbox": [ + 328, + 143, + 434, + 227 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e35dd47badc8aabfbb8b7237ed66fef07c401b20f7eb70a643fdae6957fa70e3.jpg", + "image_caption": [ + "A suitcase above a bicycle" + ], + "image_footnote": [], + "bbox": [ + 437, + 143, + 545, + 227 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/90f4c76c193d88705f749e063f08281fab8b2ef784e66be09fa74c5b5a390a6b.jpg", + "image_caption": [ + "An umbrella above a car" + ], + "image_footnote": [], + "bbox": [ + 547, + 143, + 656, + 227 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/87fcd69edb6c4af5fef0e3ca554e07a68f07ddb7502e9d5411fef4a5fa853612.jpg", + "image_caption": [ + "Vertical Relationships" + ], + "image_footnote": [], + "bbox": [ + 658, + 143, + 769, + 227 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4ae198df6775e744f811e430e5e5bd8c428e40afc03b4a8c0d27f81eb5b53f75.jpg", + "image_caption": [ + "An apple on the top of a cake", + "A truck to the left of a train" + ], + "image_footnote": [], + "bbox": [ + 217, + 244, + 326, + 330 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e89d65537d44cae701d6ed8b9423f5ec0d04e7ffce5727d5c3c333c5ac1a0db6.jpg", + "image_caption": [ + "ATV to the right of a bird" + ], + "image_footnote": [], + "bbox": [ + 328, + 244, + 436, + 329 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dd9e0614269e975ec046219d275a4286e183545a69efb11318ff714eb243193e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 437, + 244, + 545, + 329 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7a1c1f888612761527d7ba055e32978e2a258dce5139a07d1c7564f1b7834ab0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 547, + 244, + 656, + 329 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/00556fc2135078c5ac2360f2c7ab278800bcfb52c396574dfc213f32a99a0db4.jpg", + "image_caption": [ + "A book on the bottom of a sports ball", + "Hrnnnne nnnnne" + ], + "image_footnote": [], + "bbox": [ + 658, + 244, + 767, + 329 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8e8bf2fee4fc48cc3c767a37d259699db593595abfdb306855a9f6b9241bc752.jpg", + "image_caption": [ + "A microwave behind an apple", + "Fig. 1: Text-to-Image models struggle to generate images that faithfully represent the spatial relationships mentioned in the input prompt. We develop REVISION, an efficient rendering pipeline that enables a training-free and guidance-based mechanism to address this shortcoming. Our method results in improvements in spatial reasoning for T2I models for three dimensional relationships demonstrated by consistently higher scores on VISOR and T2I-CompBench benchmarks." + ], + "image_footnote": [], + "bbox": [ + 217, + 345, + 326, + 430 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/76fca434fec087057a42d3462643fd2d53ac20f5be30c7c47ce731c9f18fee6b.jpg", + "image_caption": [ + "A cake in front of a cup" + ], + "image_footnote": [], + "bbox": [ + 328, + 345, + 436, + 430 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/765779b93df56a21d230d34ff9c36b2e6ea3428323415e795a67e6f8be813a29.jpg", + "image_caption": [ + "A bottle to the left of a boat", + "An oven behind a wine glass" + ], + "image_footnote": [], + "bbox": [ + 437, + 345, + 545, + 430 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ecdb6d160e935368543ee87bb0a3bd1689064e5f4f58931a6228eb103ecf6471.jpg", + "image_caption": [ + "A stop sign to the right of a dog", + "A cat in front of a boat" + ], + "image_footnote": [], + "bbox": [ + 549, + 345, + 656, + 430 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6359d10c9e112614aed27799b779bae880c535b1ac2d2b0391460d1261907017.jpg", + "image_caption": [ + "A wine glass to the left of a boat", + "A car behind a fire hydrant", + "Depth Relationships" + ], + "image_footnote": [], + "bbox": [ + 658, + 345, + 767, + 430 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DALL-Eval [8], VISOR [15], and T2I-CompBench [18] have found that generative vision-language models suffer from a common mode of failure - their inability to correctly reason over spatial relationships.", + "bbox": [ + 212, + 577, + 785, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We postulate that the lack of spatial understanding in generative vision-language models is a result of the lack of guidance from image-text datasets. Compared to T2I models, graphics rendering tools such as Blender allow deterministic and accurate object placement, but are limited by their lower visual detail and photorealism and do not have intuitive workflows such as T2I models where users can generate images by simply typing a sentence. To get the best of both worlds, in this work, we develop REVISION, a Blender-based image rendering pipeline which enables the synthesis of images with 101 3-dimensional object (assets), 11 spatial relationships, diverse backgrounds, camera perspectives, and lighting conditions. REVISION parses an input text prompt into assets and relationships and synthesizes the scene using Blender to exactly match the input prompt in terms of both objects and their spatial arrangement.", + "bbox": [ + 212, + 626, + 787, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In a training-free manner, we leverage images from REVISION as additional guidance for existing T2I methods to their ability to generate spatially accurate", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "A. Chatterjee et al.", + "bbox": [ + 271, + 114, + 401, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "images, and demonstrate improved performance on VISOR and T2I-CompBench benchmarks. We evaluate (i) the impact of utilizing diverse backgrounds from REVISION, (ii) the trade-off between controllability and photo-realism and (iii) the added generalization to complex prompts achieved by leveraging REVISION. For a holistic study, we introduce an extension to the VISOR benchmark, to include evaluation of depth relationships (in front of/behind).", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To assess the spatial and relational reasoning abilities of MLLMs, we also create the RevQA benchmark. We construct 16 diverse question types and their adversarial variations consisting of negations, conjunctions, and disjunctions. We perform holistic evaluations on 5 state-of-the-art MLLMs and discover significant shortcomings in their ability to accurately address complex spatial reasoning questions. These models also demonstrate a lack of robustness to adversarial perturbations, leading to a substantial decline in their performance.", + "bbox": [ + 212, + 237, + 787, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The key contributions and findings are summarized below:", + "bbox": [ + 238, + 342, + 663, + 357 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We develop the REVISION framework, a 3D rendering pipeline that is guaranteed to generate spatially accurate synthetic images, given an input text prompt. An extendable framework, REVISION currently accommodates $100+$ assets across 11 spatial relationships and 3 diverse backgrounds, and support for multiple lighting conditions, camera perspectives, and shadows.", + "- We present an approach that utilizes images from REVISION in an efficient training-free manner, which results in improved spatial reasoning across multiple benchmarks. Controlled experiments, ablations, and human studies reveal consistent improvements in generating images corresponding to the spatial relationships in the input prompt (as shown in Figure 1).", + "- We introduce the RevQA question-answering benchmark to evaluate spatial reasoning abilities of multimodal large language models. Our experiments reveal the shortcomings of state-of-the-art MLLMs in reasoning over complex spatial questions and their vulnerability to adversarial perturbations." + ], + "bbox": [ + 225, + 362, + 787, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 590, + 387, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generative Models for Image Synthesis. Image generation and synthesis methods have advanced rapidly, progressing from early approaches such as generative adversarial networks (GAN) [16], variational auto-encoders (VAE) [42], and auto-regressive models (ARM) [6], to contemporary text-to-image models including Stable Diffusion [38] and DALL-E [35]. GLIDE [30] adopts classifier-free guidance in T2I and explores the efficacy of CLIP [34] as a text encoder. Compared to GLIDE,Imagen [39] adopts a frozen language model as the text encoder, reducing computational overhead, allowing for usage of large text-only corpus. Multiple variants of T2I models have been developed by leveraging T5-based text encoder [5], T2I priors [32, 37], reward-based fine-tuning [18] and developing refiner models [33] for improved image-text alignment.", + "bbox": [ + 212, + 617, + 787, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Controllable Image Generation for Spatial Fidelity. To achieve better control over diffusion-based image synthesis, multiple methods have been proposed. ReCo [49], GLIGEN [21], Control-GPT [53], Composable Diffusion [26]", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "REVISION", + "bbox": [ + 651, + 113, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and ConPreDiff 47 all develop training-based methods to provide additional conditioning for T2I models. SPRIGHT 3 introduces a spatially-focused large-scale dataset, by re-captioning 6 million images from existing vision datasets and demonstrate performance gains through an efficient training methodology. Test-time adaptations have also been proposed - (i) Layout Guidance 7 restricts specific objects to their bounding box location through the modification of cross-attention maps; however it relies on bounding box annotations, (ii) LayoutGPT 12 and LLM-grounded Diffusion 23 leverage large language models (LLMs) to generate layouts and bounding box co-ordinates and, (iii) RealCompo 54 combines multiple generative models for better spatial control. By developing an annotation-free cost-efficient framework we overcome the shortcomings of existing methods through REVISION.", + "bbox": [ + 212, + 145, + 787, + 328 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Synthetic Images for Vision and Language. The flexibility and control provided during creation of synthetic images has led to various visuo-linguistic evaluation benchmarks using rendering tools. CLEVR [19] pioneered the utilization of synthetic objects in simulated scenes for visual compositionality reasoning. Many variants of CLEVR such as CLEVR-Hans [43], CLEVR-Hyp [41], Super-CLEVR [22], and CLEVRER [50] probe multiple facets of multimodal understanding with synthetic images and videos. PaintSkills introduced in DALL-EVAL [8] is an evaluation dataset that measures multiple aspects of a T2I model, which includes spatial reasoning, image-text alignment and social biases.", + "bbox": [ + 212, + 339, + 787, + 476 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Evaluation of Multimodal LLMs. Multiple benchmarks have been proposed that evaluate reasoning capabilities of MLLMs. MMBench [27] evaluates models across 20 different dimensions, for a total of 2974 evaluation instances. The distinctive abilities of MLLMS to differentiate between coarse and fine-grained vision tasks is explored by MME [13] with images sourced from COCO. A limitation across all these benchmarks is that they collect instances from common VL datasets, increasing risk of data leakage and do not evaluate spatial relationships at scale. RevQA fills this gap by developing a diverse set of synthetic and scalable image-question pairs for a holistic evaluation.", + "bbox": [ + 212, + 488, + 787, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 The REVISION Framework", + "text_level": 1, + "bbox": [ + 215, + 643, + 522, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "REVISION (Figure 2) is a rendering-based framework for generating spatially accurate images from an input prompt. Given a prompt, we generate an image in Blender, where the two object 3D models and the camera view are situated according to the spatial relationship derived from the prompt. The components of REVISION are described below.", + "bbox": [ + 212, + 672, + 787, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Asset Library includes a large human-inspected collection of 3D models of realistic objects with variations in texture and shape. Given an object name, the Asset Library randomly selects a matching asset rescaled to fit into a $1m$ cube to ensure that they are sufficiently visible in the final output. The Asset Library features 101 distinct classes of objects, 80 of which are from MS-COCO [24]. Each object class is associated with 3 to 5 royalty-free 3D model", + "bbox": [ + 212, + 750, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "A. Chatterjee et al.", + "bbox": [ + 271, + 114, + 401, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8743d3490a8fe3a6398e6472f6027a9774fd35c087901d18ff7f0213ce7cb3e4.jpg", + "image_caption": [ + "Fig. 2: REVISION parses a prompt into assets (objects) and the spatial relationship between them and synthesizes a symbolic image in Blender, placing the respective object assets at coordinates corresponding to the parsed spatial relationship." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 785, + 268 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/42bb63515541bdbe746827e18809470f818c15965c5de4d3b1193f0242d1bfc1.jpg", + "table_caption": [ + "Table 1: Spatial relationships in REVISION and their rules for the Coordinate Generator. The objects are positioned from the camera's perspective." + ], + "table_footnote": [], + "table_body": "
RelationSpatial PhrasesCoordinate Constraints Distance (m)
Horizontalto the left, to the rightX and Z are 0.[1, 1.5]
Verticalabove, below, top, bottomX and Y are 0.[0.75, 1]
Nearnear, next to, on the side ofZ is 0.[0.75, 1]
Depthin front of, behindZ is 0. Xobj1=-Xobj2.[1, 1.5]
", + "bbox": [ + 246, + 381, + 750, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "assets from sketchfab.com, with a total of 410 3D models. REVISION includes 3 background panoramicas (Indoor, Outdoor, and White) from polyhaven.com and a corresponding textured floor asset from Sketchfab.", + "bbox": [ + 212, + 489, + 787, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Coordinate Generator deterministically generates 3D coordinates for the objects and the camera, given the names of the objects and the spatial relation extracted from the prompt. As shown in Table [1], REVISION supports four categories of spatial relationships between objects. In our coordinate frame, the X-, Y-, and Z-axis represent depth, horizontal, and vertical relationships respectively. To ensure that the objects are visible and the spatial relationship is obvious from the camera's view, the coordinate values for the objects on all three axes are confined within the range of $[-1m, 1m]$ . The camera is placed at $x = 5m$ with its view always facing the origin point. The camera is at $z = 2.5$ for depth relationships and at $z = 1.5m$ otherwise.", + "bbox": [ + 212, + 536, + 787, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Scene Synthesizer assembles a 3D scene consisting of six main components: a camera, a light source, background, floor, and two objects. The two object assets and the camera are placed at their respective coordinates determined by the Coordinate Generator. Then the background asset, which is a 360-degree panorama image (modeled as a large sphere), is centered at the origin. The light source is added to a random position sufficiently higher than all objects in the scene. To prevent objects from appearing to float, the floor asset, a textured hyperplane orthogonal to the Z-axis, is positioned beneath the object asset with the lowest vertical coordinate. This floor placement also enables the object assets to cast shadows, enhancing the realism of the rendered image.", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "REVISION", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d1ed60142e665e16aeea66e96de7dbc1df7a9db6dd8cc1cafa3e5a4d40924b1c.jpg", + "image_caption": [ + "Fig. 3: Outputs from the REVISION rendering pipeline for 4 spatial relationships types for identical assets, with (bottom) and without a floor (top)." + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 756, + 343 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The Position Diversifier (Figure 3) ensures diversity in object orientations, background, and the camera angles every time REVISION is invoked. The background is rotated along the Z-axis, giving us a large number of static background options. In order to further diversify the perspective sizes and tilts of the object assets within the camera's view, we add random jitter to the position and orientation of the camera. We also add random small rotations to the objects along the Z-axis and vary the distance between the objects so that they are not always symmetric around the origin. See Supplementary Materials for more details.", + "bbox": [ + 212, + 410, + 789, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Improving Spatial Fidelity in T2I Generation", + "text_level": 1, + "bbox": [ + 214, + 556, + 689, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Training-Free Image Generation with REVISION", + "text_level": 1, + "bbox": [ + 214, + 587, + 669, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given an input prompt $(T)$ , we first generate a spatially accurate reference image $(x^{(g)})$ leveraging our REVISION pipeline. We then perform training-free image synthesis to generate an image $I$ , i.e. $\\phi(I|x^{(g)}, T)$ , where $\\phi$ is a T2I model. We reformulate the standard text-to-image pipeline into an image-to-image pipeline, conditioned by text, as shown in Figure 4.", + "bbox": [ + 212, + 612, + 787, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Standard diffusion methods such as Stable Diffusion (SD) generate an image by iteratively de-noising a Gaussian noise vector. Stochastic Differential Editing (SDEdit) [28], on the other hand, starts from a guide image $(x^{(g)}$ , in our case), adds Gaussian noise to it, and denoises it to produce the synthesized image $I$ . We use SDEdit within our Stable Diffusion pipeline and perform image generation guided by $x^{(g)}$ . We also explore the ControlNet [51] backbone, which allows fine-grained control over SD. Using ControlNet allows us to address two key points: a) our reference images provide enough spatial information even when low-level features are extracted from them and, b) we can mitigate any attribute-related biases present in the assets.", + "bbox": [ + 212, + 688, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "A. Chatterjee et al.", + "bbox": [ + 271, + 114, + 401, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/721ebdc641f63184247c42c6a9ff84b89a0befd9b1ca1f98e74ba3e1c2113c31.jpg", + "image_caption": [ + "Fig. 4: Given a user-provided input prompt $T$ , we generate a corresponding synthetic image $x^{(g)}$ using REVISION. With input prompt $T$ and guidance $x^{(g)}$ , we perform training-free image synthesis based on existing T2I pipelines such as Stable Diffusion or ControlNet to obtain a spatially accurate image." + ], + "image_footnote": [], + "bbox": [ + 232, + 148, + 787, + 282 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1d341b8b93ff4778660f453ee4647e163dfbd3bd9e5ea79d302ef71b0f6687f0.jpg", + "table_caption": [ + "Table 2: The incorporation of REVISION as a guiding framework significantly enhances the spatial reasoning performance of Stable Diffusion (SD) models. Results highlighted in green represent scores achieved with images from REVISION." + ], + "table_footnote": [], + "table_body": "
MethodOA (%)VISOR (%)
uncond cond1234
SD 1.429.8618.8162.9846.6020.116.891.63
+ REVISION53.9652.7197.6977.7961.0244.9027.15
SD 1.528.4317.5161.5944.2718.126.281.35
+ REVISION54.3353.0897.7278.0761.2745.4427.55
SD 2.147.8330.2563.2464.4235.7416.134.70
+ REVISION48.2647.1197.6176.0755.7537.1019.53
", + "bbox": [ + 299, + 421, + 699, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Experimental Setup", + "text_level": 1, + "bbox": [ + 214, + 579, + 426, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We study the efficacy of REVISION on two widely accepted benchmarks for spatial relationship, VISOR [15] and T2I-CompBench [18], which have 25,280 and 300 spatial prompts, respectively. For each evaluation prompt in the respective benchmarks, we generate a corresponding image from our REVISION pipeline and perform training-free image generation as described in Section 4.1.", + "bbox": [ + 212, + 599, + 784, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We leverage 3 variants of Stable Diffusion (SD), versions 1.4, 1.5, and 2.1 as our baseline models. For ControlNet, we use the canny edge-conditioned SD model. For holistic evaluations, we also report the Inception Score (IS) where applicable. For all subsequent tables, the bold values denote the best performance while underlined values indicate the second-best performance.", + "bbox": [ + 212, + 676, + 784, + 752 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 Results and Analysis", + "text_level": 1, + "bbox": [ + 214, + 771, + 434, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Improvements over Baseline Models - We summarize our representative improvements over the baseline and existing methods, on the VISOR and T2I-CompBench benchmarks in Table 2 and Figure 5 respectively. The results in", + "bbox": [ + 212, + 794, + 784, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "REVISION", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/29a8b81e2ecb320f2bbcdde45c0019047b5208bc41d660e545834c9d079ddce0.jpg", + "image_caption": [ + "Fig. 5: Comparing the T2I-CompBench spatial scores of REVISION-based guidance (green) with other leading T2I models and methods (blue)." + ], + "image_footnote": [], + "bbox": [ + 243, + 148, + 743, + 353 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8ed712fc26120e9cff7a6be685cbf64599bce2334730dc1232d80dc5e477bb08.jpg", + "table_caption": [ + "Table 3: Results on the VISOR Benchmark. With REVISION, we consistently outperform existing T2I methods on the VISOR benchmark." + ], + "table_footnote": [], + "table_body": "
MethodOA (%)VISOR (%)
uncondcond1234
GLIDE 303.361.9859.066.721.020.170.03
DALLE-mini 1027.1016.1759.6738.3117.506.891.96
DALLE-v2 3563.9337.8959.2773.5947.2323.267.49
Layout Guidance 740.0138.8095.95----
Control-GPT 5348.3344.1765.9769.8051.2035.6720.48
Structured Diffusion 1128.6517.8762.3644.7018.736.571.46
Attend-and-Excite 442.0725.7561.2149.2919.334.560.08
SD 1.4 + REVISION53.9652.7197.6977.7961.0244.9027.15
SD 1.5 + REVISION54.3353.0897.7278.0761.2745.4427.55
SD 2.1 + REVISION48.2647.1197.6176.0755.7537.1019.53
ControlNet + REVISION56.8855.4897.5478.8262.9348.5831.59
", + "bbox": [ + 256, + 460, + 741, + 646 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2 are shown with reference images on a white background and # of denoising steps = 30. As shown in Table 2 we improve on all aspects of spatial relationships compared to our baseline methods. On SD 1.5, we achieve a $91.1\\%$ improvement in Object Accuracy (OA) and a $58.6\\%$ improvement on the conditional score. Specifically, we generate objects more accurately and achieve a high $\\%$ of accuracy when spatially synthesizing them in the image. Interestingly, through REVISION, we increase the likelihood of consistently generating spatially correct images, as can be seen by the relatively high value of $\\mathrm{VISOR}_4$ . On VISOR (Table 3), REVISION enables baseline Stable Diffusion models to consistently outperform existing methods, across all aspects. Compared to the", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "A. Chatterjee et al.", + "bbox": [ + 271, + 114, + 403, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f909d1a2783da051162c7fde64c4d0e26df8ee412473a2bf53921cb0eef6d438.jpg", + "image_caption": [ + "Fig.6: REVISION improves T2I-CompBench Spatial Score (0 indicates missing objects, 1 denotes perfect object generation and spatial accuracy.)" + ], + "image_footnote": [], + "bbox": [ + 225, + 148, + 485, + 270 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/65024c20325d5f9998f015a6ac777ec08ffd30815e7db04fdea752e75cb325a4.jpg", + "image_caption": [ + "Fig.7: Benchmarking the trade-off between spatial accuracy (VISOR) and Inception Score, achieved with REVISION." + ], + "image_footnote": [], + "bbox": [ + 511, + 148, + 782, + 284 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/049ba8032c667c0749e422148543ac3975d373647eb2ef9e5f7459677900e7e8.jpg", + "table_caption": [ + "Table 4: VISORcond and Object Accuracy, split across relationship types. σVc and σ0A denote the respective metric's standard deviation w.r.t the relationships. Regardless of the spatial relation, REVISION enables T2I models to consistently produce spatially accurate images, a challenge faced by earlier approaches." + ], + "table_footnote": [], + "table_body": "
MethodVISORcond (%)Object Accuracy (%)
leftrightabovebelowσVcleftrightabovebelowσ0A
GLIDE57.7861.7160.3256.242.463.103.463.493.390.18
DALLE-mini57.8960.1663.7556.143.2922.2921.7433.6230.745.99
DALLE-v256.4756.5160.9963.243.3864.3064.3265.6661.451.77
Control-GPT72.5070.2867.8565.702.9549.8048.2747.9746.951.18
SD 1.4 + REVISION97.5397.4598.0997.660.2952.4252.1156.9354.382.22
SD 1.5 + REVISION97.5797.5398.0597.700.2452.9952.5956.8054.921.94
SD 2.1 + REVISION97.8197.4697.9197.280.3046.7047.9449.7048.711.27
ControlNet + REVISION97.5197.2597.6597.720.2155.1055.1458.9858.292.05
", + "bbox": [ + 222, + 438, + 777, + 588 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "best open-source model, Control-GPT, we achieve a $\\Delta$ improvement of $17.69\\%$ , $48.12\\%$ , and $25.6\\%$ on OA, VISORcond, and VISORuncond respectively.", + "bbox": [ + 212, + 628, + 782, + 660 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "On T2I-CompBench (Figure 5), we observe similar improvement trends across diverse backgrounds, with baseline models guided by REVISION achieving consistent performance gains on the benchmark. In addition to enhancing spatial accuracy, REVISION improves prompt fidelity by ensuring that images contain all objects mentioned in the input prompt (Figure 6).", + "bbox": [ + 212, + 666, + 787, + 743 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Consistent Performance Across Relationship Types - Across all spatial relationship types, REVISION achieve a consistently high performance score across the VISOR metrics as shown in Table 4; a shortcoming prevalent in other methods. For example, the largest deviation in $\\mathrm{VISOR}_{cond}$ performance for ControlNet + REVISION is $0.21\\%$ between left and below relationships; in comparison Control-GPT deviates as much as $6.8\\%$ for the same.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "REVISION", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/fd15fb30a3fd1352031689c16a05766b0f5d843d9624e9d8d6f922ba2d231e67.jpg", + "image_caption": [ + "An apple to the left of a book" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 405, + 218 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/2b9a9bd4fa3b80931dfc6fa651888bdd65b797e059bc8f00410dab0ef3129ff6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 224, + 403, + 297 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/c424fdc44c484ceb5164ef2b83d97b1335ef330f42a9aa4caace26541fd3b437.jpg", + "image_caption": [ + "A truck to the right of a fire hydrant", + "A potted plant above a dining table" + ], + "image_footnote": [], + "bbox": [ + 217, + 308, + 403, + 382 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/fde1ce4ac770e0449cb9a15a26bf1a14f487ffc730fd690b80e3c0b0d470607c.jpg", + "image_caption": [ + "A car below an elephant", + "Background Type = Indoor", + "Fig. 8: Illustrative examples depicting the variation of generated images across the three variants of backgrounds in REVISION. For each pair, the image on the left is from REVISION and the image on the right is generated from the T2I model." + ], + "image_footnote": [], + "bbox": [ + 217, + 390, + 403, + 465 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/eb49dec07cc0f367017790173f6813a6983e0dfdf00e1742decb61a41ebe1822.jpg", + "image_caption": [ + "A handbag to the left of a bird" + ], + "image_footnote": [], + "bbox": [ + 406, + 143, + 594, + 218 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/6f9fea7c8caefe11abfcbc55cb149408210efd0065bb88e3cd5424eaadb82135.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 224, + 594, + 297 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/220d1e3c8b880b9ad5d670f4f3bad2fb0d3556add840a58a257f7b2ff9c711b2.jpg", + "image_caption": [ + "A potted plant to the right of an oven", + "A vase above a teddy bear" + ], + "image_footnote": [], + "bbox": [ + 406, + 308, + 594, + 382 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/fa7a20d80e579ea3268a2016e24afc0f7308cf133f3f304295531026f6f82f82.jpg", + "image_caption": [ + "An apple below an airplane", + "Background Type $=$ Outdoor" + ], + "image_footnote": [], + "bbox": [ + 406, + 390, + 594, + 464 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/682b01f5b336b81caec330ec224a34751624a5b422129e437d2b6b57a93689a6.jpg", + "image_caption": [ + "An orange to the left of a sink" + ], + "image_footnote": [], + "bbox": [ + 598, + 143, + 784, + 217 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/54abcc2ae2487e02aaae0b9041085300b6979d4ff4938017a3ff6818cc8996d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 224, + 784, + 297 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/65e38c054bbcab72befa979e67b0dfbc583d78f9d1d235dd71fcf34232662300.jpg", + "image_caption": [ + "A cow to the right of a couch" + ], + "image_footnote": [], + "bbox": [ + 598, + 308, + 784, + 378 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/dc6a191b76edd8a97f826851aca78ddc8ad020e4c710b33f358e4f5c17bc2f25.jpg", + "image_caption": [ + "A bed above a sheep", + "A bottle below a wine glass", + "Background Type = White" + ], + "image_footnote": [], + "bbox": [ + 598, + 390, + 784, + 462 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.4 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 577, + 401, + 590 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Impact of Background - In Table 5 we enumerate the impact of the background types in the images from the REVISION pipeline and the downstream trade-off between VISOR performance and model diversity. Utilizing white backgrounds that exclusively feature the two objects in question minimizes potential distractions for the model. Conversely, when the model is presented with initial reference images incorporating indoor or outdoor backgrounds, it exhibits the capacity to identify and leverage distractor objects, resulting in the generation of diverse images. As shown in Figure 8 all generated images maintain spatial accuracy, but noisier reference images result in greater diversity.", + "bbox": [ + 212, + 607, + 787, + 744 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Controllability vs Photo-Realism - In this setup, we study the impact of the # of denoising steps and its trade-off with photo-realism. As shown in Figure 7 that while the performance on VISOR deteriorates with additional # of denoising steps, it improves the model's ability to be more diverse and photorealistic. In Figure 9, we demonstrate that by utilizing REVISION, baseline", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "A. Chatterjee et al.", + "bbox": [ + 271, + 114, + 401, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/291d3e429b17ca1673d2cee691d65ff10730c53efeeaaba367beddf452a8a007.jpg", + "table_caption": [ + "Table 5: The impact of the 3 background types in the REVISION pipeline on the VISOR benchmark. While best performance is achieved with a white background, diverse outputs are attained with the outdoor background type." + ], + "table_footnote": [], + "table_body": "
Model Background IS (↑) OA (%)VISOR (%)
uncondcond1234
SD 1.4White16.1653.9652.7197.6977.7961.0244.927.15
Indoor19.1148.5345.1292.9774.8253.7934.7817.09
Outdoor20.1644.3241.8094.3169.7949.3831.8616.17
SD 1.5White16.2754.3353.0897.7278.0761.2745.4427.55
Indoor19.1148.7745.2892.8574.9353.9634.7717.47
Outdoor19.6643.9941.5194.3669.4848.5831.4616.52
SD 2.1White12.7948.2647.1197.6176.0755.7537.1019.53
Indoor11.5231.0829.3794.5059.8033.9617.406.34
Outdoor10.5136.3734.6795.3465.0541.2323.059.36
", + "bbox": [ + 250, + 198, + 750, + 366 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b497e0b5395cedfff65b2586f1a911b345fdb59637f2c5e3b8a984328c3c93d1.jpg", + "image_caption": [ + "Fig. 9: Illustrative examples showing the trade-off between photo-realism and denoising steps, while maintaining generating spatially accurate images using REVISION." + ], + "image_footnote": [], + "bbox": [ + 217, + 382, + 781, + 563 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "models can preserve their spatial coherency while iteratively demonstrating a higher degree of photo-realism, through more # of denoising steps.", + "bbox": [ + 212, + 632, + 787, + 662 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.5 Extending VISOR for Depth Relationships", + "text_level": 1, + "bbox": [ + 214, + 689, + 616, + 705 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We further extend the VISOR benchmark for Depth relationships (in front of/behind). We utilize Depth Anything [46] for generation of depth maps and OWLv2 [29] for object detection. Given a T2I generated image $I$ and its prompt $T$ that contain two objects $o_1, o_2$ , we obtain its depth map $I_D$ using Depth Anything. We then retrieve the centroids detected for the two objects $c_{o_1}, c_{o_2}$ using OWLv2. At these centroid coordinates, we acquire the depth values for the two objects from the depth map $I_D(c_{o_1}), I_D(c_{o_2})$ . We check if the acquired depth values match the spatial relationship in the prompt, and evaluate similar", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "REVISION", + "bbox": [ + 651, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/427d14322bf539715816c1ca55ffd8c7fa2921ef22c1abf00240f014027606c9.jpg", + "table_caption": [ + "Table 6: Comparing baseline methods against REVISION-guided image synthesis on depth relationships. DS denotes the # of Denoising Steps." + ], + "table_footnote": [], + "table_body": "
SettingOA (%)VISOR (%)
uncond1234
SD 1.5 (baseline)41.5227.1560.1231.8213.193.47
+ REVISION (DS=30)58.3229.6264.7434.5314.994.22
+ REVISION (DS=35)52.0532.0868.1137.9217.514.78
+ REVISION (DS=40)47.4330.4665.5035.9915.744.64
", + "bbox": [ + 261, + 184, + 736, + 292 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "to VISOR. As shown in Table 6, REVISION improves VISOR scores across all metrics and across multiple denoising steps.", + "bbox": [ + 212, + 320, + 782, + 351 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.6 Human Evaluations", + "text_level": 1, + "bbox": [ + 215, + 378, + 426, + 391 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To verify the generalizability of REVISION-based guidance on T2I models, we perform 2 distinct experiments and conduct human evaluations for validation. For each experiment, we independently sample 200 generated images and take the average scores across 4 workers. We also report unanimous (100%) and majority (75%) agreements between the workers for each experiment.", + "bbox": [ + 212, + 406, + 784, + 482 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Prompts of Multiple Objects and Relationships - In this experiment, we generate reference images using prompts that include 2 spatial phrases and 3 objects, and use these images to guide T2I generation. Each generated image is evaluated for accuracy based on the input spatial prompts. We achieve an accuracy of $79.62\\%$ when at least 1 phrase is correctly represented in the image and $46.5\\%$ when both phrases are correctly represented. The unanimous and majority agreements among evaluators are $64.5\\%$ and $86.5\\%$ , respectively.", + "bbox": [ + 212, + 501, + 787, + 608 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Out-of-Distribution Objects - We consider prompts containing exactly one object not found in the REVISION Asset Library. Given a prompt that mentions an OOD object, we find the semantically closest object (list in Supplementary Material) in our library and use their corresponding image as guidance. For example, we generate an image of \"a helicopter above a bicycle\" by providing a reference image of \"an airplane above a bicycle\". An accuracy of $63.62\\%$ is found with an unanimous and majority agreement of $67\\%$ and $90.5\\%$ , respectively.", + "bbox": [ + 212, + 625, + 787, + 732 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5 RevQA: A Spatial Reasoning Benchmark for MLLMs", + "text_level": 1, + "bbox": [ + 215, + 758, + 769, + 776 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We leverage the determinism of the REVISION pipeline to construct a new visual question answering benchmark (RevQA) for evaluating the spatial reasoning abilities of multimodal large language models.", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "A. Chatterjee et al.", + "bbox": [ + 271, + 114, + 401, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c65468606980fc568f6fc02eda80285fdf7e395438d9b8b5474353829f684fa7.jpg", + "image_caption": [ + "Fig. 10: The RevQA Benchmark. Using the REVISION pipeline, we generate spatially accurate images and formulate 16 question types from a given caption. We leverage these generated questions and image, benchmarking Multimodal Large Language Models in their abilities to reason over spatial relationships." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 418, + 300 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/da2b7b10755c453e14d43409898aa44301a3519e98a9043bd7dde6ca01431564.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Question TypeQuestionAnswer
Simple SpatialIs there an airplane above a chair?Yes
Opposite SpatialIs there a chair above an airplane?No
ANDIs there a chair and an airplane?Yes
ORIs there a chair or an airplane?Yes
NOTIs there an airplane not above a chair?No
Double NegativeIs there an airplane not below a chair?Yes
Random ANDIs there an airplane and a hot dog?No
Random ORIs there an airplane or a hot dog?Yes
Random SpatialIs there an airplane above a hot dog?No
Random Combined ANDIs there an airplane above a chair and is there an airplane above a hot dog?No
Random Combined ORIs there an airplane above a hot dog or is there an airplane above a chair?Yes
Adversarial ANDIs there a helicopter and a chair?No
Adversarial ORIs there a helicopter or a chair?Yes
Adversarial SpatialIs there a helicopter above a chair?No
Adversarial Combined ANDIs there a helicopter above a chair and is there an airplane above a chair?No
Adversarial Combined ORIs there a helicopter above a chair or is there an airplane above a chair?Yes
", + "bbox": [ + 418, + 143, + 784, + 301 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Question Generation. The benchmark contains 16 types of yes-no questions for a REVISION-generated image, consisting of negations, conjunctions, and disjunctions, building on prior work on logic-based visual question answering [14]. Each question type evaluates a combination of spatial and logical reasoning abilities in multimodal large language models (MLLMs) (Figure 10).", + "bbox": [ + 212, + 401, + 787, + 477 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Among the 16 types, we incorporate Random and Adversarial types of questions to further evaluate the robustness and reliability of MLLMs using simple templated transformations. In Random types of questions, we replace an object (visible in the image) in the question with another randomly picked object from REVISION's Asset Library. For the Adversarial set of questions, we replace one of the objects with another that is semantically and visually close. In addition to benchmarking their robustness, these questions allow simultaneous evaluation of the fine-grained spatial perception and reasoning abilities of these models. To alleviate any order bias in instances which contain multiple questions (see Combined in Figure 10), we randomly switch the order between them.", + "bbox": [ + 212, + 477, + 789, + 630 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Evaluation Setup and Results We sample 50k image-question pairs and benchmark 5 open-source state-of-the-art MLLMs - LLaVA 1.5 [25], Fuyu-8B [2], InstructBLIP [9], LLaMA-Adapter 2.1 [52] and Qwen-VL-Chat [1]. We instruct all models to generate binary responses and set the temperature $= 0$ , to remove stochasticity in the generated responses.", + "bbox": [ + 212, + 643, + 787, + 718 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We present our evaluation results in Table 7 and find that all models have a large gap in performance in reasoning over spatial relationships. While most models reason well over simple spatial relationships, they have a large performance drop when presented with the opposite spatial relationships. For example, LLaVA-1.5, the best performing model, has a $58.17\\%$ decrease in performance when probed with simple vs opposite spatial questions. This can be attributed to: $(a)$ insufficient training data for rare object relationships, such as less instances of an \"elephant above a person\" than vice versa; $b)$ the inability of vision", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "REVISION", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/131c6bb1a60b9e98cc248132183dfbed43e29a840be60bc4289acf19b3dda005.jpg", + "table_caption": [ + "Table 7: Performances of 5 MLLMs across the 16 types of questions in RevQA. Most models perform worse than random (50%) when reasoning over Opposite Spatial relationships and Double Negative questions. All models have a significant drop in performance with Random/Adversarial questions, in comparison to their simpler versions." + ], + "table_footnote": [], + "table_body": "
Question TypeLLaVa 1.5Fuyu-8BInstructBLIPLLaMA-Adapter 2.1Qwen-VL Chat
Simple Spatial0.9420.7020.8340.5790.940
Opposite Spatial0.3940.2870.1840.4190.402
AND0.9350.8870.9570.8580.889
OR0.9950.3960.5980.7220.949
NOT0.9260.6190.3560.5040.583
Double Negative0.2670.3470.6650.4900.212
Random AND0.9340.3080.6750.6160.978
Random OR0.9250.1780.1940.1940.324
Random Spatial0.9250.3700.6860.7900.919
Random Combined AND0.1160.5020.6270.8000.567
Random Combined OR0.9680.5360.4140.0030.506
Adversarial AND0.6610.1840.5420.6410.789
Adversarial OR0.9210.1880.4430.1560.685
Adversarial Spatial0.5590.3350.7770.8930.615
Adversarial Combined AND0.1320.5390.6950.8050.695
Adversarial Combined OR0.9530.4560.3860.0030.254
Average0.7200.4460.5980.5780.642
", + "bbox": [ + 218, + 212, + 785, + 441 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "encoders like CLIP to capture subtle semantic differences. MLLMs also struggle with negation, possibly because image captions do not capture enough negations; e.g. COCO Captions only contain $0.97\\%$ occurrences of 'not'. All models significantly suffer when presented with questions that consist of double negatives, which evaluate the models' ability to reason of negations and spatial relationships in tandem. Furthermore, all models suffer under adversarial settings in comparison to their simpler counterparts; comparing LLaVA's performance for AND and Adversarial Combined AND questions, we find a $85.88\\%$ $(0.935\\rightarrow 0.132)$ drop in performance. We also observe a larger decline in performance for Adversarial questions than for the Random set of questions hinting that while models independently perform well at object recognition and simple spatial relationships, combining them adversarily significantly reduces performance.", + "bbox": [ + 212, + 469, + 787, + 652 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 672, + 359, + 690 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we introduce REVISION, a framework designed for training-free enhancement of spatial relationships in Text-to-Image models and RevQA, a benchmark to evaluate the spatial reasoning abilities of multimodal large language models. Our results demonstrate the effectiveness of leveraging 3D rendering pipelines as a cost-efficient approach for developing generative models with robust reasoning capabilities. REVISION is modular and can easily be extended to incorporate additional features, assets, and relationships. We hope our method inspires future research at the intersection of computer graphics and generative AI, enabling safe deployment of these systems in the real world.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "A. Chatterjee et al.", + "bbox": [ + 271, + 114, + 401, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "The authors acknowledge resources provided by Research Computing at Arizona State University. The authors also acknowledge technical access and support from ASU Enterprise Technology. This work was supported by NSF Robust Intelligence program grants #1750082 and #2132724. TG was supported by Microsoft's Accelerating Foundation Model Research (AFMR) program and UMBC's Strategic Award for Research Transitions (START). The views and opinions of the authors expressed herein do not necessarily state or reflect those of the funding agencies and employers.", + "bbox": [ + 212, + 174, + 787, + 295 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 316, + 321, + 332 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond (2023), https://arxiv.org/abs/2308.12966", + "2. Bavishi, R., Elsen, E., Hawthorne, C., Nye, M., Odena, A., Somani, A., Taşırlar, S.: Introducing our multimodal models (2023), https://www.adept.ai/blog/fuyu-8b", + "3. Chatterjee, A., Stan, G.B.M., Aflalo, E., Paul, S., Ghosh, D., Gokhale, T., Schmidt, L., Hajishirzi, H., Lal, V., Baral, C., Yang, Y.: Getting it right: Improving spatial consistency in text-to-image models (2024), https://arxiv.org/abs/2404.01197", + "4. Chefer, H., Alaluf, Y., Vinker, Y., Wolf, L., Cohen-Or, D.: Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG) 42(4), 1-10 (2023)", + "5. Chen, J., Jincheng, Y., Chongjian, G., Yao, L., Xie, E., Wang, Z., Kwok, J., Luo, P., Lu, H., Li, Z.: Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In: The Twelfth International Conference on Learning Representations (2023)", + "6. Chen, M., Radford, A., Child, R., Wu, J., Jun, H., Luan, D., Sutskever, I.: Generative pretraining from pixels. In: Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event. Proceedings of Machine Learning Research, vol. 119, pp. 1691-1703. PMLR (2020), http://proceedings.mlr.press/v119/chen20s.html", + "7. Chen, M., Laina, I., Vedaldi, A.: Training-free layout control with cross-attention guidance (2023)", + "8. Cho, J., Zala, A., Bansal, M.: Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3043-3054 (2023)", + "9. Dai, W., Li, J., Li, D., Tiong, A., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=vvoWPYqZJA", + "10. Dayma, B., Patil, S., Cuenca, P., Saifullah, K., Abraham, T., Le Khac, P., Melas, L., Ghosh, R.: Dall-e mini (2021). https://doi.org/10.5281/zenodo.5146400. https://github.com/borisdayma/dalle-mini", + "11. Feng, W., He, X., Fu, T.J., Jampani, V., Akula, A.R., Narayana, P., Basu, S., Wang, X.E., Wang, W.Y.: Training-free structured diffusion guidance for compositional text-to-image synthesis. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=PUIqjT4rzq7" + ], + "bbox": [ + 217, + 345, + 787, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "REVISION", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "12. Feng, W., Zhu, W., Fu, T.j., Jampani, V., Akula, A., He, X., Basu, S., Wang, X.E., Wang, W.Y.: Layoutgpt: Compositional visual planning and generation with large language models. Advances in Neural Information Processing Systems 36 (2024)", + "13. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Yang, J., Zheng, X., Li, K., Sun, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. ArXiv preprint abs/2306.13394 (2023), https://arxiv.org/abs/2306.13394", + "14. Gokhale, T., Banerjee, P., Baral, C., Yang, Y.: Vqa-lol: Visual question answering under the lens of logic. In: European conference on computer vision. pp. 379-396. Springer (2020)", + "15. Gokhale, T., Palangi, H., Nushi, B., Vineet, V., Horvitz, E., Kamar, E., Baral, C., Yang, Y.: Benchmarking spatial relationships in text-to-image generation. arXiv preprint arXiv:2212.10015 (2022)", + "16. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139–144 (2020)", + "17. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-or, D.: Prompt-to-prompt image editing with cross-attention control. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=CDixzkzeyb", + "18. Huang, K., Sun, K., Xie, E., Li, Z., Liu, X.: T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. Advances in Neural Information Processing Systems 36, 78723-78747 (2023)", + "19. Johnson, J., Hariharan, B., van der Maaten, L., Fei-Fei, L., Zitnick, C.L., Girshick, R.B.: CLEVR: A diagnostic dataset for compositional language and elementary visual reasoning. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017. pp. 1988-1997. IEEE Computer Society (2017). https://doi.org/10.1109/CVPR.2017.215. https://doi.org/10.1109/CVPR.2017.215", + "20. Kamath, A., Hessel, J., Chang, K.W.: What's \"up\" with vision-language models? investigating their struggle with spatial reasoning. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. pp. 9161-9175 (2023)", + "21. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22511-22521 (2023)", + "22. Li, Z., Wang, X., Stengel-Eskin, E., Kortylewski, A., Ma, W., Van Durme, B., Yuille, A.L.: Super-clevr: A virtual benchmark to diagnose domain robustness in visual reasoning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14963-14973 (2023)", + "23. Lian, L., Li, B., Yala, A., Darrell, T.: Llm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. ArXiv preprint abs/2305.13655 (2023), https://arxiv.org/abs/2305.13655", + "24. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "25. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "A. Chatterjee et al.", + "bbox": [ + 271, + 114, + 401, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "26. Liu, N., Li, S., Du, Y., Torralba, A., Tenenbaum, J.B.: Compositional visual generation with composable diffusion models. In: European Conference on Computer Vision. pp. 423-439. Springer (2022)", + "27. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281 (2023)", + "28. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. In: The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net (2022), https://openreview.net/forum?id=aBsCjcPu_tE", + "29. Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Dosovitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., et al.: Simple open-vocabulary object detection. In: European Conference on Computer Vision. pp. 728-755. Springer (2022)", + "30. Nichol, A.Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: GLIDE: towards photorealistic image generation and editing with text-guided diffusion models. In: Chaudhuri, K., Jegelka, S., Song, L., Szepesvári, C., Niu, G., Sabato, S. (eds.) International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA. Proceedings of Machine Learning Research, vol. 162, pp. 16784-16804. PMLR (2022), https://proceedings.mlr.press/v162/nichol22a.html", + "31. Parmar, G., Kumar Singh, K., Zhang, R., Li, Y., Lu, J., Zhu, J.Y.: Zero-shot image-to-image translation. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-11 (2023)", + "32. Patel, M., Kim, C., Cheng, S., Baral, C., Yang, Y.: Eclipse: A resource-efficient text-to-image prior for image generations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9069-9078 (2024)", + "33. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: SDXL: Improving latent diffusion models for high-resolution image synthesis. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=di52zR8xgf", + "34. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event. Proceedings of Machine Learning Research, vol. 139, pp. 8748-8763. PMLR (2021), http://proceedings.mlr.press/v139/radford21a.html", + "35. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022)", + "36. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event. Proceedings of Machine Learning Research, vol. 139, pp. 8821-8831. PMLR (2021), http://proceedings.mlr.press/v139/ramesh21a.html", + "37. Razzhigaev, A., Shakhmatov, A., Maltseva, A., Arkhipkin, V., Pavlov, I., Ryabov, I., Kuts, A., Panchenko, A., Kuznetsov, A., Dimitrov, D.: Kandinsky: An improved" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "REVISION", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "text-to-image synthesis with image prior and latent diffusion. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations. pp. 286-295 (2023)", + "38. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)", + "39. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)", + "40. Salimans, T., Goodfellow, I.J., Zaremba, W., Cheung, V., Radford, A., Chen, X.: Improved techniques for training gans. In: Lee, D.D., Sugiyama, M., von Luxburg, U., Guyon, I., Garnett, R. (eds.) Advances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, December 5-10, 2016, Barcelona, Spain. pp. 2226-2234 (2016), https://proceedings.neurips.cc/paper/2016/bitstream/8a3363abe792db2d8761d6403605aab7-Abstract.html", + "41. Sampat, S.K., Kumar, A., Yang, Y., Baral, C.: CLEVR_HYP: A challenge dataset and baselines for visual question answering with hypothetical actions over images. In: Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. pp. 3692-3709. Association for Computational Linguistics, Online (2021). https://doi.org/10.18653/v1/2021.naacl-main.289. https://aclanthology.org/2021.naacl-main.289", + "42. Sohn, K., Lee, H., Yan, X.: Learning structured output representation using deep conditional generative models. In: Cortes, C., Lawrence, N.D., Lee, D.D., Sugiyama, M., Garnett, R. (eds.) Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada. pp. 3483-3491 (2015), https://proceedings.neurips.cc/paper/2015/bitstream/8d55a249e6baa5c06772297520da2051-Abstract.html", + "43. Stammer, W., Schramowski, P., Kersting, K.: Right for the right concept: Revising neuro-symbolic concepts by interacting with their explanations. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 3619-3629. Computer Vision Foundation / IEEE (2021). https://doi.org/10.1109/CVPR46437.2021.00362, https://openaccess.thecvf.com/content/CVPR2021/html/Stammer_Right_for_the_Right_Concept_Revising_Neuro-Symbolic_Concepts_by_Interacting_CVPR_2021_paper.html", + "44. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. ArXiv preprint abs/2312.11805 (2023), https://arxiv.org/abs/2312.11805", + "45. Wake, N., Kanehira, A., Sasabuchi, K., Takamatsu, J., Ikeuchi, K.: Gpt-4v (ision) for robotics: Multimodal task planning from human demonstration. arXiv preprint arXiv:2311.12015 (2023)", + "46. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. ArXiv preprint abs/2401.10891 (2024), https://arxiv.org/abs/2401.10891" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "A. Chatterjee et al.", + "bbox": [ + 271, + 114, + 401, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "47. Yang, L., Liu, J., Hong, S., Zhang, Z., Huang, Z., Cai, Z., Zhang, W., Cui, B.: Improving diffusion-based image synthesis with context prediction. Advances in Neural Information Processing Systems 36 (2024)", + "48. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of lmms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421 9(1), 1 (2023)", + "49. Yang, Z., Wang, J., Gan, Z., Li, L., Lin, K., Wu, C., Duan, N., Liu, Z., Liu, C., Zeng, M., et al.: Reco: Region-controlled text-to-image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14246-14255 (2023)", + "50. Yi, K., Gan, C., Li, Y., Kohli, P., Wu, J., Torralba, A., Tenenbaum, J.B.: CLEVRER: collision events for video representation and reasoning. In: 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net (2020), https://openreview.net/forum?id=HkxYzANYDB", + "51. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023)", + "52. Zhang, R., Han, J., Liu, C., Gao, P., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199 (2023)", + "53. Zhang, T., Zhang, Y., Vineet, V., Joshi, N., Wang, X.: Controllable text-to-image generation with gpt-4. arXiv preprint arXiv:2305.18583 (2023)", + "54. Zhang, X., Yang, L., Cai, Y., Yu, Z., Xie, J., Tian, Y., Xu, M., Tang, Y., Yang, Y., Cui, B.: Realcompo: Dynamic equilibrium between realism and compositionality improves text-to-image diffusion models. arXiv preprint arXiv:2402.12908 (2024)" + ], + "bbox": [ + 212, + 146, + 787, + 508 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "REVISION", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/REVISION_ Rendering Tools Enable Spatial Fidelity in Vision-Language Models/f77a4b99-c39a-4693-bb9e-aff60ba50459_model.json b/2024/REVISION_ Rendering Tools Enable Spatial Fidelity in Vision-Language Models/f77a4b99-c39a-4693-bb9e-aff60ba50459_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c8532a7e6a7c5e2e1665403bb80edc17c079c8fb --- /dev/null +++ b/2024/REVISION_ Rendering Tools Enable Spatial Fidelity in Vision-Language Models/f77a4b99-c39a-4693-bb9e-aff60ba50459_model.json @@ -0,0 +1,2856 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.249, + 0.141, + 0.756, + 0.187 + ], + "angle": 0, + "content": "REVISION: Rendering Tools Enable Spatial Fidelity in Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.304, + 0.212, + 0.703, + 0.248 + ], + "angle": 0, + "content": "Agneet Chatterjee \\(\\star 1\\) Yiran Luo \\(\\star 1\\) Tejas Gokhale\\(^{2}\\), Yezhou Yang\\(^{1}\\), and Chitta Baral\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.352, + 0.255, + 0.651, + 0.284 + ], + "angle": 0, + "content": "\\(^{1}\\) Arizona State University \n\\(^{2}\\) University of Maryland, Baltimore County" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.326, + 0.744, + 0.605 + ], + "angle": 0, + "content": "Abstract. Text-to-Image (T2I) and multimodal large language models (MLLMs) have been adopted in solutions for several computer vision and multimodal learning tasks. However, it has been found that such vision-language models lack the ability to correctly reason over spatial relationships. To tackle this shortcoming, we develop the REVISION framework which improves spatial fidelity in vision-language models. REVISION is a 3D rendering based pipeline that generates spatially accurate synthetic images, given a textual prompt. REVISION is an extendable framework, which currently supports \\(100+\\) 3D assets, 11 spatial relationships, all with diverse camera perspectives and backgrounds. Leveraging images from REVISION as additional guidance in a training-free manner consistently improves the spatial consistency of T2I models across all spatial relationships, achieving competitive performance on the VISOR and T2I-CompBench benchmarks. We also design RevQA, a question-answering benchmark to evaluate the spatial reasoning abilities of MLLMs, and find that state-of-the-art models are not robust to complex spatial reasoning under adversarial settings. Our results and findings indicate that utilizing rendering-based frameworks is an effective approach for developing spatially-aware generative models. Code and data available at: https://github.com/agneet42/revision" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.617, + 0.741, + 0.631 + ], + "angle": 0, + "content": "Keywords: Text to Image \\(\\cdot\\) Spatial Relationships \\(\\cdot\\) Rendering Graphics" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.659, + 0.376, + 0.674 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.693, + 0.788, + 0.816 + ], + "angle": 0, + "content": "Generative vision-language models [36,44] represent a significant step towards developing multimodal systems that bridge the gap between computer vision and natural language processing. Text-to-image (T2I) models [5,38] convert text prompts to high-quality images, while multimodal large language models (MLLMs) [25,48] process images as inputs, and generate rich and coherent natural language outputs in response. As a result, these models have found diverse applications in robotics [45], image editing [17], image-to-image translation [31], and more. However, recent studies [20] and benchmarks such as" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.826, + 0.575, + 0.842 + ], + "angle": 0, + "content": "* Equal contribution. Correspondence to agneet@asu.edu" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.403, + 0.129 + ], + "angle": 0, + "content": "A. Chatterjee et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.327, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.235, + 0.229, + 0.307, + 0.245 + ], + "angle": 0, + "content": "An apple on the top of a cake" + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.145, + 0.436, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.333, + 0.232, + 0.432, + 0.242 + ], + "angle": 0, + "content": "A couch below an airplane" + }, + { + "type": "image", + "bbox": [ + 0.438, + 0.145, + 0.547, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.443, + 0.232, + 0.542, + 0.242 + ], + "angle": 0, + "content": "A suitcase above a bicycle" + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.145, + 0.657, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.232, + 0.648, + 0.242 + ], + "angle": 0, + "content": "An umbrella above a car" + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.145, + 0.77, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.229, + 0.762, + 0.245 + ], + "angle": 0, + "content": "A book on the bottom of a sports ball" + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.155, + 0.783, + 0.217 + ], + "angle": 0, + "content": "Vertical Relationships" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.245, + 0.327, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.334, + 0.321, + 0.343 + ], + "angle": 0, + "content": "A truck to the left of a train" + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.245, + 0.437, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.336, + 0.334, + 0.429, + 0.343 + ], + "angle": 0, + "content": "ATV to the right of a bird" + }, + { + "type": "image", + "bbox": [ + 0.439, + 0.245, + 0.547, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.443, + 0.334, + 0.542, + 0.344 + ], + "angle": 0, + "content": "A bottle to the left of a boat" + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.245, + 0.657, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.332, + 0.655, + 0.347 + ], + "angle": 0, + "content": "A stop sign to the right of a dog" + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.245, + 0.768, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.332, + 0.763, + 0.346 + ], + "angle": 0, + "content": "A wine glass to the left of a boat" + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.253, + 0.783, + 0.324 + ], + "angle": 0, + "content": "Hrnnnne nnnnne" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.347, + 0.327, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.228, + 0.434, + 0.316, + 0.449 + ], + "angle": 0, + "content": "A microwave behind an apple" + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.347, + 0.437, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.34, + 0.436, + 0.425, + 0.445 + ], + "angle": 0, + "content": "A cake in front of a cup" + }, + { + "type": "image", + "bbox": [ + 0.439, + 0.347, + 0.547, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.437, + 0.546, + 0.446 + ], + "angle": 0, + "content": "An oven behind a wine glass" + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.347, + 0.658, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.437, + 0.644, + 0.445 + ], + "angle": 0, + "content": "A cat in front of a boat" + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.347, + 0.768, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.437, + 0.762, + 0.445 + ], + "angle": 0, + "content": "A car behind a fire hydrant" + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.361, + 0.783, + 0.418 + ], + "angle": 0, + "content": "Depth Relationships" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.461, + 0.788, + 0.545 + ], + "angle": 0, + "content": "Fig. 1: Text-to-Image models struggle to generate images that faithfully represent the spatial relationships mentioned in the input prompt. We develop REVISION, an efficient rendering pipeline that enables a training-free and guidance-based mechanism to address this shortcoming. Our method results in improvements in spatial reasoning for T2I models for three dimensional relationships demonstrated by consistently higher scores on VISOR and T2I-CompBench benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.578, + 0.787, + 0.625 + ], + "angle": 0, + "content": "DALL-Eval [8], VISOR [15], and T2I-CompBench [18] have found that generative vision-language models suffer from a common mode of failure - their inability to correctly reason over spatial relationships." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.788, + 0.809 + ], + "angle": 0, + "content": "We postulate that the lack of spatial understanding in generative vision-language models is a result of the lack of guidance from image-text datasets. Compared to T2I models, graphics rendering tools such as Blender allow deterministic and accurate object placement, but are limited by their lower visual detail and photorealism and do not have intuitive workflows such as T2I models where users can generate images by simply typing a sentence. To get the best of both worlds, in this work, we develop REVISION, a Blender-based image rendering pipeline which enables the synthesis of images with 101 3-dimensional object (assets), 11 spatial relationships, diverse backgrounds, camera perspectives, and lighting conditions. REVISION parses an input text prompt into assets and relationships and synthesizes the scene using Blender to exactly match the input prompt in terms of both objects and their spatial arrangement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "In a training-free manner, we leverage images from REVISION as additional guidance for existing T2I methods to their ability to generate spatially accurate" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.653, + 0.114, + 0.732, + 0.127 + ], + "angle": 0, + "content": "REVISION" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "images, and demonstrate improved performance on VISOR and T2I-CompBench benchmarks. We evaluate (i) the impact of utilizing diverse backgrounds from REVISION, (ii) the trade-off between controllability and photo-realism and (iii) the added generalization to complex prompts achieved by leveraging REVISION. For a holistic study, we introduce an extension to the VISOR benchmark, to include evaluation of depth relationships (in front of/behind)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.788, + 0.342 + ], + "angle": 0, + "content": "To assess the spatial and relational reasoning abilities of MLLMs, we also create the RevQA benchmark. We construct 16 diverse question types and their adversarial variations consisting of negations, conjunctions, and disjunctions. We perform holistic evaluations on 5 state-of-the-art MLLMs and discover significant shortcomings in their ability to accurately address complex spatial reasoning questions. These models also demonstrate a lack of robustness to adversarial perturbations, leading to a substantial decline in their performance." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.343, + 0.665, + 0.358 + ], + "angle": 0, + "content": "The key contributions and findings are summarized below:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.363, + 0.788, + 0.437 + ], + "angle": 0, + "content": "- We develop the REVISION framework, a 3D rendering pipeline that is guaranteed to generate spatially accurate synthetic images, given an input text prompt. An extendable framework, REVISION currently accommodates \\(100+\\) assets across 11 spatial relationships and 3 diverse backgrounds, and support for multiple lighting conditions, camera perspectives, and shadows." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.438, + 0.788, + 0.513 + ], + "angle": 0, + "content": "- We present an approach that utilizes images from REVISION in an efficient training-free manner, which results in improved spatial reasoning across multiple benchmarks. Controlled experiments, ablations, and human studies reveal consistent improvements in generating images corresponding to the spatial relationships in the input prompt (as shown in Figure 1)." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.512, + 0.788, + 0.572 + ], + "angle": 0, + "content": "- We introduce the RevQA question-answering benchmark to evaluate spatial reasoning abilities of multimodal large language models. Our experiments reveal the shortcomings of state-of-the-art MLLMs in reasoning over complex spatial questions and their vulnerability to adversarial perturbations." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.363, + 0.788, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.591, + 0.388, + 0.607 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.618, + 0.788, + 0.786 + ], + "angle": 0, + "content": "Generative Models for Image Synthesis. Image generation and synthesis methods have advanced rapidly, progressing from early approaches such as generative adversarial networks (GAN) [16], variational auto-encoders (VAE) [42], and auto-regressive models (ARM) [6], to contemporary text-to-image models including Stable Diffusion [38] and DALL-E [35]. GLIDE [30] adopts classifier-free guidance in T2I and explores the efficacy of CLIP [34] as a text encoder. Compared to GLIDE,Imagen [39] adopts a frozen language model as the text encoder, reducing computational overhead, allowing for usage of large text-only corpus. Multiple variants of T2I models have been developed by leveraging T5-based text encoder [5], T2I priors [32, 37], reward-based fine-tuning [18] and developing refiner models [33] for improved image-text alignment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Controllable Image Generation for Spatial Fidelity. To achieve better control over diffusion-based image synthesis, multiple methods have been proposed. ReCo [49], GLIGEN [21], Control-GPT [53], Composable Diffusion [26]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.403, + 0.129 + ], + "angle": 0, + "content": "A. Chatterjee et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.329 + ], + "angle": 0, + "content": "and ConPreDiff 47 all develop training-based methods to provide additional conditioning for T2I models. SPRIGHT 3 introduces a spatially-focused large-scale dataset, by re-captioning 6 million images from existing vision datasets and demonstrate performance gains through an efficient training methodology. Test-time adaptations have also been proposed - (i) Layout Guidance 7 restricts specific objects to their bounding box location through the modification of cross-attention maps; however it relies on bounding box annotations, (ii) LayoutGPT 12 and LLM-grounded Diffusion 23 leverage large language models (LLMs) to generate layouts and bounding box co-ordinates and, (iii) RealCompo 54 combines multiple generative models for better spatial control. By developing an annotation-free cost-efficient framework we overcome the shortcomings of existing methods through REVISION." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.788, + 0.477 + ], + "angle": 0, + "content": "Synthetic Images for Vision and Language. The flexibility and control provided during creation of synthetic images has led to various visuo-linguistic evaluation benchmarks using rendering tools. CLEVR [19] pioneered the utilization of synthetic objects in simulated scenes for visual compositionality reasoning. Many variants of CLEVR such as CLEVR-Hans [43], CLEVR-Hyp [41], Super-CLEVR [22], and CLEVRER [50] probe multiple facets of multimodal understanding with synthetic images and videos. PaintSkills introduced in DALL-EVAL [8] is an evaluation dataset that measures multiple aspects of a T2I model, which includes spatial reasoning, image-text alignment and social biases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.788, + 0.625 + ], + "angle": 0, + "content": "Evaluation of Multimodal LLMs. Multiple benchmarks have been proposed that evaluate reasoning capabilities of MLLMs. MMBench [27] evaluates models across 20 different dimensions, for a total of 2974 evaluation instances. The distinctive abilities of MLLMS to differentiate between coarse and fine-grained vision tasks is explored by MME [13] with images sourced from COCO. A limitation across all these benchmarks is that they collect instances from common VL datasets, increasing risk of data leakage and do not evaluate spatial relationships at scale. RevQA fills this gap by developing a diverse set of synthetic and scalable image-question pairs for a holistic evaluation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.645, + 0.524, + 0.661 + ], + "angle": 0, + "content": "3 The REVISION Framework" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.788, + 0.748 + ], + "angle": 0, + "content": "REVISION (Figure 2) is a rendering-based framework for generating spatially accurate images from an input prompt. Given a prompt, we generate an image in Blender, where the two object 3D models and the camera view are situated according to the spatial relationship derived from the prompt. The components of REVISION are described below." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.843 + ], + "angle": 0, + "content": "The Asset Library includes a large human-inspected collection of 3D models of realistic objects with variations in texture and shape. Given an object name, the Asset Library randomly selects a matching asset rescaled to fit into a \\(1m\\) cube to ensure that they are sufficiently visible in the final output. The Asset Library features 101 distinct classes of objects, 80 of which are from MS-COCO [24]. Each object class is associated with 3 to 5 royalty-free 3D model" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.654, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "REVISION" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.144, + 0.787, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.281, + 0.788, + 0.325 + ], + "angle": 0, + "content": "Fig. 2: REVISION parses a prompt into assets (objects) and the spatial relationship between them and synthesizes a symbolic image in Blender, placing the respective object assets at coordinates corresponding to the parsed spatial relationship." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.341, + 0.788, + 0.37 + ], + "angle": 0, + "content": "Table 1: Spatial relationships in REVISION and their rules for the Coordinate Generator. The objects are positioned from the camera's perspective." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.382, + 0.751, + 0.459 + ], + "angle": 0, + "content": "
RelationSpatial PhrasesCoordinate Constraints Distance (m)
Horizontalto the left, to the rightX and Z are 0.[1, 1.5]
Verticalabove, below, top, bottomX and Y are 0.[0.75, 1]
Nearnear, next to, on the side ofZ is 0.[0.75, 1]
Depthin front of, behindZ is 0. Xobj1=-Xobj2.[1, 1.5]
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.788, + 0.535 + ], + "angle": 0, + "content": "assets from sketchfab.com, with a total of 410 3D models. REVISION includes 3 background panoramicas (Indoor, Outdoor, and White) from polyhaven.com and a corresponding textured floor asset from Sketchfab." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.788, + 0.688 + ], + "angle": 0, + "content": "The Coordinate Generator deterministically generates 3D coordinates for the objects and the camera, given the names of the objects and the spatial relation extracted from the prompt. As shown in Table [1], REVISION supports four categories of spatial relationships between objects. In our coordinate frame, the X-, Y-, and Z-axis represent depth, horizontal, and vertical relationships respectively. To ensure that the objects are visible and the spatial relationship is obvious from the camera's view, the coordinate values for the objects on all three axes are confined within the range of \\([-1m, 1m]\\). The camera is placed at \\(x = 5m\\) with its view always facing the origin point. The camera is at \\(z = 2.5\\) for depth relationships and at \\(z = 1.5m\\) otherwise." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The Scene Synthesizer assembles a 3D scene consisting of six main components: a camera, a light source, background, floor, and two objects. The two object assets and the camera are placed at their respective coordinates determined by the Coordinate Generator. Then the background asset, which is a 360-degree panorama image (modeled as a large sphere), is centered at the origin. The light source is added to a random position sufficiently higher than all objects in the scene. To prevent objects from appearing to float, the floor asset, a textured hyperplane orthogonal to the Z-axis, is positioned beneath the object asset with the lowest vertical coordinate. This floor placement also enables the object assets to cast shadows, enhancing the realism of the rendered image." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.403, + 0.129 + ], + "angle": 0, + "content": "A. Chatterjee et al." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.144, + 0.758, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.353, + 0.788, + 0.385 + ], + "angle": 0, + "content": "Fig. 3: Outputs from the REVISION rendering pipeline for 4 spatial relationships types for identical assets, with (bottom) and without a floor (top)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.411, + 0.79, + 0.534 + ], + "angle": 0, + "content": "The Position Diversifier (Figure 3) ensures diversity in object orientations, background, and the camera angles every time REVISION is invoked. The background is rotated along the Z-axis, giving us a large number of static background options. In order to further diversify the perspective sizes and tilts of the object assets within the camera's view, we add random jitter to the position and orientation of the camera. We also add random small rotations to the objects along the Z-axis and vary the distance between the objects so that they are not always symmetric around the origin. See Supplementary Materials for more details." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.557, + 0.691, + 0.575 + ], + "angle": 0, + "content": "4 Improving Spatial Fidelity in T2I Generation" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.588, + 0.671, + 0.604 + ], + "angle": 0, + "content": "4.1 Training-Free Image Generation with REVISION" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.689 + ], + "angle": 0, + "content": "Given an input prompt \\((T)\\), we first generate a spatially accurate reference image \\((x^{(g)})\\) leveraging our REVISION pipeline. We then perform training-free image synthesis to generate an image \\(I\\), i.e. \\(\\phi(I|x^{(g)}, T)\\), where \\(\\phi\\) is a T2I model. We reformulate the standard text-to-image pipeline into an image-to-image pipeline, conditioned by text, as shown in Figure 4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Standard diffusion methods such as Stable Diffusion (SD) generate an image by iteratively de-noising a Gaussian noise vector. Stochastic Differential Editing (SDEdit) [28], on the other hand, starts from a guide image \\((x^{(g)}\\), in our case), adds Gaussian noise to it, and denoises it to produce the synthesized image \\(I\\). We use SDEdit within our Stable Diffusion pipeline and perform image generation guided by \\(x^{(g)}\\). We also explore the ControlNet [51] backbone, which allows fine-grained control over SD. Using ControlNet allows us to address two key points: a) our reference images provide enough spatial information even when low-level features are extracted from them and, b) we can mitigate any attribute-related biases present in the assets." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.654, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "REVISION" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.149, + 0.788, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.299, + 0.785, + 0.355 + ], + "angle": 0, + "content": "Fig. 4: Given a user-provided input prompt \\( T \\), we generate a corresponding synthetic image \\( x^{(g)} \\) using REVISION. With input prompt \\( T \\) and guidance \\( x^{(g)} \\), we perform training-free image synthesis based on existing T2I pipelines such as Stable Diffusion or ControlNet to obtain a spatially accurate image." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.367, + 0.788, + 0.41 + ], + "angle": 0, + "content": "Table 2: The incorporation of REVISION as a guiding framework significantly enhances the spatial reasoning performance of Stable Diffusion (SD) models. Results highlighted in green represent scores achieved with images from REVISION." + }, + { + "type": "table", + "bbox": [ + 0.3, + 0.422, + 0.7, + 0.553 + ], + "angle": 0, + "content": "
MethodOA (%)VISOR (%)
uncond cond1234
SD 1.429.8618.8162.9846.6020.116.891.63
+ REVISION53.9652.7197.6977.7961.0244.9027.15
SD 1.528.4317.5161.5944.2718.126.281.35
+ REVISION54.3353.0897.7278.0761.2745.4427.55
SD 2.147.8330.2563.2464.4235.7416.134.70
+ REVISION48.2647.1197.6176.0755.7537.1019.53
" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.58, + 0.427, + 0.595 + ], + "angle": 0, + "content": "4.2 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.601, + 0.785, + 0.677 + ], + "angle": 0, + "content": "We study the efficacy of REVISION on two widely accepted benchmarks for spatial relationship, VISOR [15] and T2I-CompBench [18], which have 25,280 and 300 spatial prompts, respectively. For each evaluation prompt in the respective benchmarks, we generate a corresponding image from our REVISION pipeline and perform training-free image generation as described in Section 4.1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.678, + 0.785, + 0.753 + ], + "angle": 0, + "content": "We leverage 3 variants of Stable Diffusion (SD), versions 1.4, 1.5, and 2.1 as our baseline models. For ControlNet, we use the canny edge-conditioned SD model. For holistic evaluations, we also report the Inception Score (IS) where applicable. For all subsequent tables, the bold values denote the best performance while underlined values indicate the second-best performance." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.772, + 0.436, + 0.788 + ], + "angle": 0, + "content": "4.3 Results and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.785, + 0.843 + ], + "angle": 0, + "content": "Improvements over Baseline Models - We summarize our representative improvements over the baseline and existing methods, on the VISOR and T2I-CompBench benchmarks in Table 2 and Figure 5 respectively. The results in" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.404, + 0.129 + ], + "angle": 0, + "content": "A. Chatterjee et al." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.15, + 0.744, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.363, + 0.788, + 0.392 + ], + "angle": 0, + "content": "Fig. 5: Comparing the T2I-CompBench spatial scores of REVISION-based guidance (green) with other leading T2I models and methods (blue)." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.42, + 0.789, + 0.449 + ], + "angle": 0, + "content": "Table 3: Results on the VISOR Benchmark. With REVISION, we consistently outperform existing T2I methods on the VISOR benchmark." + }, + { + "type": "table", + "bbox": [ + 0.258, + 0.462, + 0.742, + 0.647 + ], + "angle": 0, + "content": "
MethodOA (%)VISOR (%)
uncondcond1234
GLIDE 303.361.9859.066.721.020.170.03
DALLE-mini 1027.1016.1759.6738.3117.506.891.96
DALLE-v2 3563.9337.8959.2773.5947.2323.267.49
Layout Guidance 740.0138.8095.95----
Control-GPT 5348.3344.1765.9769.8051.2035.6720.48
Structured Diffusion 1128.6517.8762.3644.7018.736.571.46
Attend-and-Excite 442.0725.7561.2149.2919.334.560.08
SD 1.4 + REVISION53.9652.7197.6977.7961.0244.9027.15
SD 1.5 + REVISION54.3353.0897.7278.0761.2745.4427.55
SD 2.1 + REVISION48.2647.1197.6176.0755.7537.1019.53
ControlNet + REVISION56.8855.4897.5478.8262.9348.5831.59
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Table 2 are shown with reference images on a white background and # of denoising steps = 30. As shown in Table 2 we improve on all aspects of spatial relationships compared to our baseline methods. On SD 1.5, we achieve a \\(91.1\\%\\) improvement in Object Accuracy (OA) and a \\(58.6\\%\\) improvement on the conditional score. Specifically, we generate objects more accurately and achieve a high \\(\\%\\) of accuracy when spatially synthesizing them in the image. Interestingly, through REVISION, we increase the likelihood of consistently generating spatially correct images, as can be seen by the relatively high value of \\(\\mathrm{VISOR}_4\\). On VISOR (Table 3), REVISION enables baseline Stable Diffusion models to consistently outperform existing methods, across all aspects. Compared to the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.654, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "REVISION" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.149, + 0.486, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.286, + 0.493, + 0.343 + ], + "angle": 0, + "content": "Fig.6: REVISION improves T2I-CompBench Spatial Score (0 indicates missing objects, 1 denotes perfect object generation and spatial accuracy.)" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.149, + 0.783, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.299, + 0.786, + 0.341 + ], + "angle": 0, + "content": "Fig.7: Benchmarking the trade-off between spatial accuracy (VISOR) and Inception Score, achieved with REVISION." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.371, + 0.788, + 0.428 + ], + "angle": 0, + "content": "Table 4: VISORcond and Object Accuracy, split across relationship types. σVc and σ0A denote the respective metric's standard deviation w.r.t the relationships. Regardless of the spatial relation, REVISION enables T2I models to consistently produce spatially accurate images, a challenge faced by earlier approaches." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.439, + 0.778, + 0.589 + ], + "angle": 0, + "content": "
MethodVISORcond (%)Object Accuracy (%)
leftrightabovebelowσVcleftrightabovebelowσ0A
GLIDE57.7861.7160.3256.242.463.103.463.493.390.18
DALLE-mini57.8960.1663.7556.143.2922.2921.7433.6230.745.99
DALLE-v256.4756.5160.9963.243.3864.3064.3265.6661.451.77
Control-GPT72.5070.2867.8565.702.9549.8048.2747.9746.951.18
SD 1.4 + REVISION97.5397.4598.0997.660.2952.4252.1156.9354.382.22
SD 1.5 + REVISION97.5797.5398.0597.700.2452.9952.5956.8054.921.94
SD 2.1 + REVISION97.8197.4697.9197.280.3046.7047.9449.7048.711.27
ControlNet + REVISION97.5197.2597.6597.720.2155.1055.1458.9858.292.05
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.784, + 0.661 + ], + "angle": 0, + "content": "best open-source model, Control-GPT, we achieve a \\(\\Delta\\) improvement of \\(17.69\\%\\), \\(48.12\\%\\), and \\(25.6\\%\\) on OA, VISORcond, and VISORuncond respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.667, + 0.788, + 0.744 + ], + "angle": 0, + "content": "On T2I-CompBench (Figure 5), we observe similar improvement trends across diverse backgrounds, with baseline models guided by REVISION achieving consistent performance gains on the benchmark. In addition to enhancing spatial accuracy, REVISION improves prompt fidelity by ensuring that images contain all objects mentioned in the input prompt (Figure 6)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Consistent Performance Across Relationship Types - Across all spatial relationship types, REVISION achieve a consistently high performance score across the VISOR metrics as shown in Table 4; a shortcoming prevalent in other methods. For example, the largest deviation in \\(\\mathrm{VISOR}_{cond}\\) performance for ControlNet + REVISION is \\(0.21\\%\\) between left and below relationships; in comparison Control-GPT deviates as much as \\(6.8\\%\\) for the same." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.403, + 0.129 + ], + "angle": 0, + "content": "A. Chatterjee et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.406, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.218, + 0.37, + 0.226 + ], + "angle": 0, + "content": "An apple to the left of a book" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.226, + 0.405, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.244, + 0.3, + 0.381, + 0.309 + ], + "angle": 0, + "content": "A truck to the right of a fire hydrant" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.309, + 0.405, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.244, + 0.383, + 0.381, + 0.391 + ], + "angle": 0, + "content": "A potted plant above a dining table" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.391, + 0.405, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.263, + 0.465, + 0.361, + 0.473 + ], + "angle": 0, + "content": "A car below an elephant" + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.479, + 0.365, + 0.488 + ], + "angle": 0, + "content": "Background Type = Indoor" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.144, + 0.596, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.443, + 0.218, + 0.561, + 0.226 + ], + "angle": 0, + "content": "A handbag to the left of a bird" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.226, + 0.596, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.429, + 0.3, + 0.574, + 0.309 + ], + "angle": 0, + "content": "A potted plant to the right of an oven" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.309, + 0.596, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.448, + 0.383, + 0.555, + 0.391 + ], + "angle": 0, + "content": "A vase above a teddy bear" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.391, + 0.596, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.448, + 0.465, + 0.556, + 0.473 + ], + "angle": 0, + "content": "An apple below an airplane" + }, + { + "type": "image_caption", + "bbox": [ + 0.446, + 0.479, + 0.558, + 0.488 + ], + "angle": 0, + "content": "Background Type \\(=\\) Outdoor" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.144, + 0.785, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.218, + 0.75, + 0.226 + ], + "angle": 0, + "content": "An orange to the left of a sink" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.226, + 0.785, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.3, + 0.749, + 0.309 + ], + "angle": 0, + "content": "A cow to the right of a couch" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.309, + 0.785, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.383, + 0.734, + 0.391 + ], + "angle": 0, + "content": "A bed above a sheep" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.391, + 0.785, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.637, + 0.464, + 0.746, + 0.473 + ], + "angle": 0, + "content": "A bottle below a wine glass" + }, + { + "type": "image_caption", + "bbox": [ + 0.64, + 0.479, + 0.743, + 0.488 + ], + "angle": 0, + "content": "Background Type = White" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.504, + 0.788, + 0.547 + ], + "angle": 0, + "content": "Fig. 8: Illustrative examples depicting the variation of generated images across the three variants of backgrounds in REVISION. For each pair, the image on the left is from REVISION and the image on the right is generated from the T2I model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.578, + 0.402, + 0.592 + ], + "angle": 0, + "content": "4.4 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.789, + 0.746 + ], + "angle": 0, + "content": "Impact of Background - In Table 5 we enumerate the impact of the background types in the images from the REVISION pipeline and the downstream trade-off between VISOR performance and model diversity. Utilizing white backgrounds that exclusively feature the two objects in question minimizes potential distractions for the model. Conversely, when the model is presented with initial reference images incorporating indoor or outdoor backgrounds, it exhibits the capacity to identify and leverage distractor objects, resulting in the generation of diverse images. As shown in Figure 8 all generated images maintain spatial accuracy, but noisier reference images result in greater diversity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Controllability vs Photo-Realism - In this setup, we study the impact of the # of denoising steps and its trade-off with photo-realism. As shown in Figure 7 that while the performance on VISOR deteriorates with additional # of denoising steps, it improves the model's ability to be more diverse and photorealistic. In Figure 9, we demonstrate that by utilizing REVISION, baseline" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.653, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "REVISION" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 5: The impact of the 3 background types in the REVISION pipeline on the VISOR benchmark. While best performance is achieved with a white background, diverse outputs are attained with the outdoor background type." + }, + { + "type": "table", + "bbox": [ + 0.25, + 0.199, + 0.75, + 0.367 + ], + "angle": 0, + "content": "
Model Background IS (↑) OA (%)VISOR (%)
uncondcond1234
SD 1.4White16.1653.9652.7197.6977.7961.0244.927.15
Indoor19.1148.5345.1292.9774.8253.7934.7817.09
Outdoor20.1644.3241.8094.3169.7949.3831.8616.17
SD 1.5White16.2754.3353.0897.7278.0761.2745.4427.55
Indoor19.1148.7745.2892.8574.9353.9634.7717.47
Outdoor19.6643.9941.5194.3669.4848.5831.4616.52
SD 2.1White12.7948.2647.1197.6176.0755.7537.1019.53
Indoor11.5231.0829.3794.5059.8033.9617.406.34
Outdoor10.5136.3734.6795.3465.0541.2323.059.36
" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.383, + 0.782, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.573, + 0.788, + 0.601 + ], + "angle": 0, + "content": "Fig. 9: Illustrative examples showing the trade-off between photo-realism and denoising steps, while maintaining generating spatially accurate images using REVISION." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.633, + 0.788, + 0.664 + ], + "angle": 0, + "content": "models can preserve their spatial coherency while iteratively demonstrating a higher degree of photo-realism, through more # of denoising steps." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.69, + 0.617, + 0.706 + ], + "angle": 0, + "content": "4.5 Extending VISOR for Depth Relationships" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We further extend the VISOR benchmark for Depth relationships (in front of/behind). We utilize Depth Anything [46] for generation of depth maps and OWLv2 [29] for object detection. Given a T2I generated image \\( I \\) and its prompt \\( T \\) that contain two objects \\( o_1, o_2 \\), we obtain its depth map \\( I_D \\) using Depth Anything. We then retrieve the centroids detected for the two objects \\( c_{o_1}, c_{o_2} \\) using OWLv2. At these centroid coordinates, we acquire the depth values for the two objects from the depth map \\( I_D(c_{o_1}), I_D(c_{o_2}) \\). We check if the acquired depth values match the spatial relationship in the prompt, and evaluate similar" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.403, + 0.129 + ], + "angle": 0, + "content": "A. Chatterjee et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 6: Comparing baseline methods against REVISION-guided image synthesis on depth relationships. DS denotes the # of Denoising Steps." + }, + { + "type": "table", + "bbox": [ + 0.263, + 0.185, + 0.737, + 0.293 + ], + "angle": 0, + "content": "
SettingOA (%)VISOR (%)
uncond1234
SD 1.5 (baseline)41.5227.1560.1231.8213.193.47
+ REVISION (DS=30)58.3229.6264.7434.5314.994.22
+ REVISION (DS=35)52.0532.0868.1137.9217.514.78
+ REVISION (DS=40)47.4330.4665.5035.9915.744.64
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.321, + 0.784, + 0.352 + ], + "angle": 0, + "content": "to VISOR. As shown in Table 6, REVISION improves VISOR scores across all metrics and across multiple denoising steps." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.379, + 0.427, + 0.392 + ], + "angle": 0, + "content": "4.6 Human Evaluations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.407, + 0.785, + 0.483 + ], + "angle": 0, + "content": "To verify the generalizability of REVISION-based guidance on T2I models, we perform 2 distinct experiments and conduct human evaluations for validation. For each experiment, we independently sample 200 generated images and take the average scores across 4 workers. We also report unanimous (100%) and majority (75%) agreements between the workers for each experiment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.502, + 0.788, + 0.609 + ], + "angle": 0, + "content": "Prompts of Multiple Objects and Relationships - In this experiment, we generate reference images using prompts that include 2 spatial phrases and 3 objects, and use these images to guide T2I generation. Each generated image is evaluated for accuracy based on the input spatial prompts. We achieve an accuracy of \\(79.62\\%\\) when at least 1 phrase is correctly represented in the image and \\(46.5\\%\\) when both phrases are correctly represented. The unanimous and majority agreements among evaluators are \\(64.5\\%\\) and \\(86.5\\%\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.788, + 0.733 + ], + "angle": 0, + "content": "Out-of-Distribution Objects - We consider prompts containing exactly one object not found in the REVISION Asset Library. Given a prompt that mentions an OOD object, we find the semantically closest object (list in Supplementary Material) in our library and use their corresponding image as guidance. For example, we generate an image of \"a helicopter above a bicycle\" by providing a reference image of \"an airplane above a bicycle\". An accuracy of \\(63.62\\%\\) is found with an unanimous and majority agreement of \\(67\\%\\) and \\(90.5\\%\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.759, + 0.77, + 0.777 + ], + "angle": 0, + "content": "5 RevQA: A Spatial Reasoning Benchmark for MLLMs" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We leverage the determinism of the REVISION pipeline to construct a new visual question answering benchmark (RevQA) for evaluating the spatial reasoning abilities of multimodal large language models." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.654, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "REVISION" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.419, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.419, + 0.145, + 0.785, + 0.303 + ], + "angle": 0, + "content": "
Question TypeQuestionAnswer
Simple SpatialIs there an airplane above a chair?Yes
Opposite SpatialIs there a chair above an airplane?No
ANDIs there a chair and an airplane?Yes
ORIs there a chair or an airplane?Yes
NOTIs there an airplane not above a chair?No
Double NegativeIs there an airplane not below a chair?Yes
Random ANDIs there an airplane and a hot dog?No
Random ORIs there an airplane or a hot dog?Yes
Random SpatialIs there an airplane above a hot dog?No
Random Combined ANDIs there an airplane above a chair and is there an airplane above a hot dog?No
Random Combined ORIs there an airplane above a hot dog or is there an airplane above a chair?Yes
Adversarial ANDIs there a helicopter and a chair?No
Adversarial ORIs there a helicopter or a chair?Yes
Adversarial SpatialIs there a helicopter above a chair?No
Adversarial Combined ANDIs there a helicopter above a chair and is there an airplane above a chair?No
Adversarial Combined ORIs there a helicopter above a chair or is there an airplane above a chair?Yes
" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.315, + 0.788, + 0.371 + ], + "angle": 0, + "content": "Fig. 10: The RevQA Benchmark. Using the REVISION pipeline, we generate spatially accurate images and formulate 16 question types from a given caption. We leverage these generated questions and image, benchmarking Multimodal Large Language Models in their abilities to reason over spatial relationships." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.789, + 0.478 + ], + "angle": 0, + "content": "Question Generation. The benchmark contains 16 types of yes-no questions for a REVISION-generated image, consisting of negations, conjunctions, and disjunctions, building on prior work on logic-based visual question answering [14]. Each question type evaluates a combination of spatial and logical reasoning abilities in multimodal large language models (MLLMs) (Figure 10)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.79, + 0.631 + ], + "angle": 0, + "content": "Among the 16 types, we incorporate Random and Adversarial types of questions to further evaluate the robustness and reliability of MLLMs using simple templated transformations. In Random types of questions, we replace an object (visible in the image) in the question with another randomly picked object from REVISION's Asset Library. For the Adversarial set of questions, we replace one of the objects with another that is semantically and visually close. In addition to benchmarking their robustness, these questions allow simultaneous evaluation of the fine-grained spatial perception and reasoning abilities of these models. To alleviate any order bias in instances which contain multiple questions (see Combined in Figure 10), we randomly switch the order between them." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.719 + ], + "angle": 0, + "content": "Evaluation Setup and Results We sample 50k image-question pairs and benchmark 5 open-source state-of-the-art MLLMs - LLaVA 1.5 [25], Fuyu-8B [2], InstructBLIP [9], LLaMA-Adapter 2.1 [52] and Qwen-VL-Chat [1]. We instruct all models to generate binary responses and set the temperature \\(= 0\\), to remove stochasticity in the generated responses." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We present our evaluation results in Table 7 and find that all models have a large gap in performance in reasoning over spatial relationships. While most models reason well over simple spatial relationships, they have a large performance drop when presented with the opposite spatial relationships. For example, LLaVA-1.5, the best performing model, has a \\(58.17\\%\\) decrease in performance when probed with simple vs opposite spatial questions. This can be attributed to: \\((a)\\) insufficient training data for rare object relationships, such as less instances of an \"elephant above a person\" than vice versa; \\(b)\\) the inability of vision" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.403, + 0.129 + ], + "angle": 0, + "content": "A. Chatterjee et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.787, + 0.201 + ], + "angle": 0, + "content": "Table 7: Performances of 5 MLLMs across the 16 types of questions in RevQA. Most models perform worse than random (50%) when reasoning over Opposite Spatial relationships and Double Negative questions. All models have a significant drop in performance with Random/Adversarial questions, in comparison to their simpler versions." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.213, + 0.787, + 0.442 + ], + "angle": 0, + "content": "
Question TypeLLaVa 1.5Fuyu-8BInstructBLIPLLaMA-Adapter 2.1Qwen-VL Chat
Simple Spatial0.9420.7020.8340.5790.940
Opposite Spatial0.3940.2870.1840.4190.402
AND0.9350.8870.9570.8580.889
OR0.9950.3960.5980.7220.949
NOT0.9260.6190.3560.5040.583
Double Negative0.2670.3470.6650.4900.212
Random AND0.9340.3080.6750.6160.978
Random OR0.9250.1780.1940.1940.324
Random Spatial0.9250.3700.6860.7900.919
Random Combined AND0.1160.5020.6270.8000.567
Random Combined OR0.9680.5360.4140.0030.506
Adversarial AND0.6610.1840.5420.6410.789
Adversarial OR0.9210.1880.4430.1560.685
Adversarial Spatial0.5590.3350.7770.8930.615
Adversarial Combined AND0.1320.5390.6950.8050.695
Adversarial Combined OR0.9530.4560.3860.0030.254
Average0.7200.4460.5980.5780.642
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.47, + 0.789, + 0.653 + ], + "angle": 0, + "content": "encoders like CLIP to capture subtle semantic differences. MLLMs also struggle with negation, possibly because image captions do not capture enough negations; e.g. COCO Captions only contain \\(0.97\\%\\) occurrences of 'not'. All models significantly suffer when presented with questions that consist of double negatives, which evaluate the models' ability to reason of negations and spatial relationships in tandem. Furthermore, all models suffer under adversarial settings in comparison to their simpler counterparts; comparing LLaVA's performance for AND and Adversarial Combined AND questions, we find a \\(85.88\\%\\) \\((0.935\\rightarrow 0.132)\\) drop in performance. We also observe a larger decline in performance for Adversarial questions than for the Random set of questions hinting that while models independently perform well at object recognition and simple spatial relationships, combining them adversarily significantly reduces performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.674, + 0.36, + 0.691 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In this work, we introduce REVISION, a framework designed for training-free enhancement of spatial relationships in Text-to-Image models and RevQA, a benchmark to evaluate the spatial reasoning abilities of multimodal large language models. Our results demonstrate the effectiveness of leveraging 3D rendering pipelines as a cost-efficient approach for developing generative models with robust reasoning capabilities. REVISION is modular and can easily be extended to incorporate additional features, assets, and relationships. We hope our method inspires future research at the intersection of computer graphics and generative AI, enabling safe deployment of these systems in the real world." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.654, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "REVISION" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.296 + ], + "angle": 0, + "content": "The authors acknowledge resources provided by Research Computing at Arizona State University. The authors also acknowledge technical access and support from ASU Enterprise Technology. This work was supported by NSF Robust Intelligence program grants #1750082 and #2132724. TG was supported by Microsoft's Accelerating Foundation Model Research (AFMR) program and UMBC's Strategic Award for Research Transitions (START). The views and opinions of the authors expressed herein do not necessarily state or reflect those of the funding agencies and employers." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.318, + 0.323, + 0.333 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.346, + 0.787, + 0.388 + ], + "angle": 0, + "content": "1. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond (2023), https://arxiv.org/abs/2308.12966" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.389, + 0.788, + 0.416 + ], + "angle": 0, + "content": "2. Bavishi, R., Elsen, E., Hawthorne, C., Nye, M., Odena, A., Somani, A., Taşırlar, S.: Introducing our multimodal models (2023), https://www.adept.ai/blog/fuyu-8b" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.416, + 0.787, + 0.457 + ], + "angle": 0, + "content": "3. Chatterjee, A., Stan, G.B.M., Aflalo, E., Paul, S., Ghosh, D., Gokhale, T., Schmidt, L., Hajishirzi, H., Lal, V., Baral, C., Yang, Y.: Getting it right: Improving spatial consistency in text-to-image models (2024), https://arxiv.org/abs/2404.01197" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.457, + 0.787, + 0.498 + ], + "angle": 0, + "content": "4. Chefer, H., Alaluf, Y., Vinker, Y., Wolf, L., Cohen-Or, D.: Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG) 42(4), 1-10 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.498, + 0.787, + 0.552 + ], + "angle": 0, + "content": "5. Chen, J., Jincheng, Y., Chongjian, G., Yao, L., Xie, E., Wang, Z., Kwok, J., Luo, P., Lu, H., Li, Z.: Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In: The Twelfth International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.553, + 0.787, + 0.621 + ], + "angle": 0, + "content": "6. Chen, M., Radford, A., Child, R., Wu, J., Jun, H., Luan, D., Sutskever, I.: Generative pretraining from pixels. In: Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event. Proceedings of Machine Learning Research, vol. 119, pp. 1691-1703. PMLR (2020), http://proceedings.mlr.press/v119/chen20s.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.622, + 0.787, + 0.648 + ], + "angle": 0, + "content": "7. Chen, M., Laina, I., Vedaldi, A.: Training-free layout control with cross-attention guidance (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.649, + 0.787, + 0.69 + ], + "angle": 0, + "content": "8. Cho, J., Zala, A., Bansal, M.: Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3043-3054 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.69, + 0.787, + 0.744 + ], + "angle": 0, + "content": "9. Dai, W., Li, J., Li, D., Tiong, A., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=vvoWPYqZJA" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.745, + 0.785, + 0.786 + ], + "angle": 0, + "content": "10. Dayma, B., Patil, S., Cuenca, P., Saifullah, K., Abraham, T., Le Khac, P., Melas, L., Ghosh, R.: Dall-e mini (2021). https://doi.org/10.5281/zenodo.5146400. https://github.com/borisdayma/dalle-mini" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.786, + 0.787, + 0.842 + ], + "angle": 0, + "content": "11. Feng, W., He, X., Fu, T.J., Jampani, V., Akula, A.R., Narayana, P., Basu, S., Wang, X.E., Wang, W.Y.: Training-free structured diffusion guidance for compositional text-to-image synthesis. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=PUIqjT4rzq7" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.346, + 0.788, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.403, + 0.128 + ], + "angle": 0, + "content": "A. Chatterjee et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.19 + ], + "angle": 0, + "content": "12. Feng, W., Zhu, W., Fu, T.j., Jampani, V., Akula, A., He, X., Basu, S., Wang, X.E., Wang, W.Y.: Layoutgpt: Compositional visual planning and generation with large language models. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.247 + ], + "angle": 0, + "content": "13. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Yang, J., Zheng, X., Li, K., Sun, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. ArXiv preprint abs/2306.13394 (2023), https://arxiv.org/abs/2306.13394" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.247, + 0.787, + 0.288 + ], + "angle": 0, + "content": "14. Gokhale, T., Banerjee, P., Baral, C., Yang, Y.: Vqa-lol: Visual question answering under the lens of logic. In: European conference on computer vision. pp. 379-396. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.289, + 0.787, + 0.331 + ], + "angle": 0, + "content": "15. Gokhale, T., Palangi, H., Nushi, B., Vineet, V., Horvitz, E., Kamar, E., Baral, C., Yang, Y.: Benchmarking spatial relationships in text-to-image generation. arXiv preprint arXiv:2212.10015 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.332, + 0.787, + 0.374 + ], + "angle": 0, + "content": "16. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139–144 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.374, + 0.787, + 0.431 + ], + "angle": 0, + "content": "17. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-or, D.: Prompt-to-prompt image editing with cross-attention control. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=CDixzkzeyb" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.431, + 0.787, + 0.472 + ], + "angle": 0, + "content": "18. Huang, K., Sun, K., Xie, E., Li, Z., Liu, X.: T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. Advances in Neural Information Processing Systems 36, 78723-78747 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.557 + ], + "angle": 0, + "content": "19. Johnson, J., Hariharan, B., van der Maaten, L., Fei-Fei, L., Zitnick, C.L., Girshick, R.B.: CLEVR: A diagnostic dataset for compositional language and elementary visual reasoning. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017. pp. 1988-1997. IEEE Computer Society (2017). https://doi.org/10.1109/CVPR.2017.215. https://doi.org/10.1109/CVPR.2017.215" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.558, + 0.787, + 0.613 + ], + "angle": 0, + "content": "20. Kamath, A., Hessel, J., Chang, K.W.: What's \"up\" with vision-language models? investigating their struggle with spatial reasoning. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. pp. 9161-9175 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.614, + 0.787, + 0.655 + ], + "angle": 0, + "content": "21. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22511-22521 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.656, + 0.787, + 0.712 + ], + "angle": 0, + "content": "22. Li, Z., Wang, X., Stengel-Eskin, E., Kortylewski, A., Ma, W., Van Durme, B., Yuille, A.L.: Super-clevr: A virtual benchmark to diagnose domain robustness in visual reasoning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14963-14973 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.713, + 0.787, + 0.755 + ], + "angle": 0, + "content": "23. Lian, L., Li, B., Yala, A., Darrell, T.: Llm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. ArXiv preprint abs/2305.13655 (2023), https://arxiv.org/abs/2305.13655" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.756, + 0.787, + 0.812 + ], + "angle": 0, + "content": "24. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.812, + 0.787, + 0.841 + ], + "angle": 0, + "content": "25. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.654, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "REVISION" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.189 + ], + "angle": 0, + "content": "26. Liu, N., Li, S., Du, Y., Torralba, A., Tenenbaum, J.B.: Compositional visual generation with composable diffusion models. In: European Conference on Computer Vision. pp. 423-439. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "27. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.3 + ], + "angle": 0, + "content": "28. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. In: The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net (2022), https://openreview.net/forum?id=aBsCjcPu_tE" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.787, + 0.357 + ], + "angle": 0, + "content": "29. Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Dosovitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., et al.: Simple open-vocabulary object detection. In: European Conference on Computer Vision. pp. 728-755. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.453 + ], + "angle": 0, + "content": "30. Nichol, A.Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: GLIDE: towards photorealistic image generation and editing with text-guided diffusion models. In: Chaudhuri, K., Jegelka, S., Song, L., Szepesvári, C., Niu, G., Sabato, S. (eds.) International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA. Proceedings of Machine Learning Research, vol. 162, pp. 16784-16804. PMLR (2022), https://proceedings.mlr.press/v162/nichol22a.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.495 + ], + "angle": 0, + "content": "31. Parmar, G., Kumar Singh, K., Zhang, R., Li, Y., Lu, J., Zhu, J.Y.: Zero-shot image-to-image translation. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-11 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.536 + ], + "angle": 0, + "content": "32. Patel, M., Kim, C., Cheng, S., Baral, C., Yang, Y.: Eclipse: A resource-efficient text-to-image prior for image generations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9069-9078 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.787, + 0.592 + ], + "angle": 0, + "content": "33. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: SDXL: Improving latent diffusion models for high-resolution image synthesis. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=di52zR8xgf" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.787, + 0.687 + ], + "angle": 0, + "content": "34. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event. Proceedings of Machine Learning Research, vol. 139, pp. 8748-8763. PMLR (2021), http://proceedings.mlr.press/v139/radford21a.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.688, + 0.787, + 0.73 + ], + "angle": 0, + "content": "35. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.787, + 0.811 + ], + "angle": 0, + "content": "36. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event. Proceedings of Machine Learning Research, vol. 139, pp. 8821-8831. PMLR (2021), http://proceedings.mlr.press/v139/ramesh21a.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.812, + 0.787, + 0.841 + ], + "angle": 0, + "content": "37. Razzhigaev, A., Shakhmatov, A., Maltseva, A., Arkhipkin, V., Pavlov, I., Ryabov, I., Kuts, A., Panchenko, A., Kuznetsov, A., Dimitrov, D.: Kandinsky: An improved" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.403, + 0.128 + ], + "angle": 0, + "content": "A. Chatterjee et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.147, + 0.786, + 0.19 + ], + "angle": 0, + "content": "text-to-image synthesis with image prior and latent diffusion. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations. pp. 286-295 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.192, + 0.787, + 0.234 + ], + "angle": 0, + "content": "38. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.236, + 0.787, + 0.293 + ], + "angle": 0, + "content": "39. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.295, + 0.787, + 0.392 + ], + "angle": 0, + "content": "40. Salimans, T., Goodfellow, I.J., Zaremba, W., Cheung, V., Radford, A., Chen, X.: Improved techniques for training gans. In: Lee, D.D., Sugiyama, M., von Luxburg, U., Guyon, I., Garnett, R. (eds.) Advances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, December 5-10, 2016, Barcelona, Spain. pp. 2226-2234 (2016), https://proceedings.neurips.cc/paper/2016/bitstream/8a3363abe792db2d8761d6403605aab7-Abstract.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.394, + 0.787, + 0.493 + ], + "angle": 0, + "content": "41. Sampat, S.K., Kumar, A., Yang, Y., Baral, C.: CLEVR_HYP: A challenge dataset and baselines for visual question answering with hypothetical actions over images. In: Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. pp. 3692-3709. Association for Computational Linguistics, Online (2021). https://doi.org/10.18653/v1/2021.naacl-main.289. https://aclanthology.org/2021.naacl-main.289" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.593 + ], + "angle": 0, + "content": "42. Sohn, K., Lee, H., Yan, X.: Learning structured output representation using deep conditional generative models. In: Cortes, C., Lawrence, N.D., Lee, D.D., Sugiyama, M., Garnett, R. (eds.) Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada. pp. 3483-3491 (2015), https://proceedings.neurips.cc/paper/2015/bitstream/8d55a249e6baa5c06772297520da2051-Abstract.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.595, + 0.787, + 0.693 + ], + "angle": 0, + "content": "43. Stammer, W., Schramowski, P., Kersting, K.: Right for the right concept: Revising neuro-symbolic concepts by interacting with their explanations. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 3619-3629. Computer Vision Foundation / IEEE (2021). https://doi.org/10.1109/CVPR46437.2021.00362, https://openaccess.thecvf.com/content/CVPR2021/html/Stammer_Right_for_the_Right_Concept_Revising_Neuro-Symbolic_Concepts_by_Interacting_CVPR_2021_paper.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.695, + 0.787, + 0.75 + ], + "angle": 0, + "content": "44. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. ArXiv preprint abs/2312.11805 (2023), https://arxiv.org/abs/2312.11805" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.753, + 0.787, + 0.796 + ], + "angle": 0, + "content": "45. Wake, N., Kanehira, A., Sasabuchi, K., Takamatsu, J., Ikeuchi, K.: Gpt-4v (ision) for robotics: Multimodal task planning from human demonstration. arXiv preprint arXiv:2311.12015 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.798, + 0.787, + 0.842 + ], + "angle": 0, + "content": "46. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. ArXiv preprint abs/2401.10891 (2024), https://arxiv.org/abs/2401.10891" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.654, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "REVISION" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "47. Yang, L., Liu, J., Hong, S., Zhang, Z., Huang, Z., Cai, Z., Zhang, W., Cui, B.: Improving diffusion-based image synthesis with context prediction. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.231 + ], + "angle": 0, + "content": "48. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of lmms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421 9(1), 1 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.286 + ], + "angle": 0, + "content": "49. Yang, Z., Wang, J., Gan, Z., Li, L., Lin, K., Wu, C., Duan, N., Liu, Z., Liu, C., Zeng, M., et al.: Reco: Region-controlled text-to-image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14246-14255 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.354 + ], + "angle": 0, + "content": "50. Yi, K., Gan, C., Li, Y., Kohli, P., Wu, J., Torralba, A., Tenenbaum, J.B.: CLEVRER: collision events for video representation and reasoning. In: 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net (2020), https://openreview.net/forum?id=HkxYzANYDB" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.354, + 0.788, + 0.397 + ], + "angle": 0, + "content": "51. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.397, + 0.788, + 0.438 + ], + "angle": 0, + "content": "52. Zhang, R., Han, J., Liu, C., Gao, P., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.438, + 0.788, + 0.466 + ], + "angle": 0, + "content": "53. Zhang, T., Zhang, Y., Vineet, V., Joshi, N., Wang, X.: Controllable text-to-image generation with gpt-4. arXiv preprint arXiv:2305.18583 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.466, + 0.788, + 0.509 + ], + "angle": 0, + "content": "54. Zhang, X., Yang, L., Cai, Y., Yu, Z., Xie, J., Tian, Y., Xu, M., Tang, Y., Yang, Y., Cui, B.: Realcompo: Dynamic equilibrium between realism and compositionality improves text-to-image diffusion models. arXiv preprint arXiv:2402.12908 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.509 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/REVISION_ Rendering Tools Enable Spatial Fidelity in Vision-Language Models/f77a4b99-c39a-4693-bb9e-aff60ba50459_origin.pdf b/2024/REVISION_ Rendering Tools Enable Spatial Fidelity in Vision-Language Models/f77a4b99-c39a-4693-bb9e-aff60ba50459_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0a12502f12f9b2d6f55d181bd85fadd51c9601e9 --- /dev/null +++ b/2024/REVISION_ Rendering Tools Enable Spatial Fidelity in Vision-Language Models/f77a4b99-c39a-4693-bb9e-aff60ba50459_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2487e28e252615015f1d576dee67bee61fc7eb0b34f3ecdb71243d2139989c66 +size 3817718 diff --git a/2024/REVISION_ Rendering Tools Enable Spatial Fidelity in Vision-Language Models/full.md b/2024/REVISION_ Rendering Tools Enable Spatial Fidelity in Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d17eb8145bcf1f718b1f40a438664213b6664f0c --- /dev/null +++ b/2024/REVISION_ Rendering Tools Enable Spatial Fidelity in Vision-Language Models/full.md @@ -0,0 +1,331 @@ +# REVISION: Rendering Tools Enable Spatial Fidelity in Vision-Language Models + +Agneet Chatterjee $\star 1$ Yiran Luo $\star 1$ Tejas Gokhale $^{2}$ , Yezhou Yang $^{1}$ , and Chitta Baral $^{1}$ + +$^{1}$ Arizona State University + $^{2}$ University of Maryland, Baltimore County + +Abstract. Text-to-Image (T2I) and multimodal large language models (MLLMs) have been adopted in solutions for several computer vision and multimodal learning tasks. However, it has been found that such vision-language models lack the ability to correctly reason over spatial relationships. To tackle this shortcoming, we develop the REVISION framework which improves spatial fidelity in vision-language models. REVISION is a 3D rendering based pipeline that generates spatially accurate synthetic images, given a textual prompt. REVISION is an extendable framework, which currently supports $100+$ 3D assets, 11 spatial relationships, all with diverse camera perspectives and backgrounds. Leveraging images from REVISION as additional guidance in a training-free manner consistently improves the spatial consistency of T2I models across all spatial relationships, achieving competitive performance on the VISOR and T2I-CompBench benchmarks. We also design RevQA, a question-answering benchmark to evaluate the spatial reasoning abilities of MLLMs, and find that state-of-the-art models are not robust to complex spatial reasoning under adversarial settings. Our results and findings indicate that utilizing rendering-based frameworks is an effective approach for developing spatially-aware generative models. Code and data available at: https://github.com/agneet42/revision + +Keywords: Text to Image $\cdot$ Spatial Relationships $\cdot$ Rendering Graphics + +# 1 Introduction + +Generative vision-language models [36,44] represent a significant step towards developing multimodal systems that bridge the gap between computer vision and natural language processing. Text-to-image (T2I) models [5,38] convert text prompts to high-quality images, while multimodal large language models (MLLMs) [25,48] process images as inputs, and generate rich and coherent natural language outputs in response. As a result, these models have found diverse applications in robotics [45], image editing [17], image-to-image translation [31], and more. However, recent studies [20] and benchmarks such as + +![](images/7c26d986754b6ea01d566b4a44133234a2f36e3122ee36b599032b507cb86331.jpg) + +![](images/9e04c3522ebd952e052c4b205571fe673306b234f82a0b5915a7a654558559e0.jpg) +A couch below an airplane + +![](images/e35dd47badc8aabfbb8b7237ed66fef07c401b20f7eb70a643fdae6957fa70e3.jpg) +A suitcase above a bicycle + +![](images/90f4c76c193d88705f749e063f08281fab8b2ef784e66be09fa74c5b5a390a6b.jpg) +An umbrella above a car + +![](images/87fcd69edb6c4af5fef0e3ca554e07a68f07ddb7502e9d5411fef4a5fa853612.jpg) +Vertical Relationships + +![](images/4ae198df6775e744f811e430e5e5bd8c428e40afc03b4a8c0d27f81eb5b53f75.jpg) +An apple on the top of a cake +A truck to the left of a train + +![](images/e89d65537d44cae701d6ed8b9423f5ec0d04e7ffce5727d5c3c333c5ac1a0db6.jpg) +ATV to the right of a bird + +![](images/dd9e0614269e975ec046219d275a4286e183545a69efb11318ff714eb243193e.jpg) + +![](images/7a1c1f888612761527d7ba055e32978e2a258dce5139a07d1c7564f1b7834ab0.jpg) + +![](images/00556fc2135078c5ac2360f2c7ab278800bcfb52c396574dfc213f32a99a0db4.jpg) +A book on the bottom of a sports ball +Hrnnnne nnnnne + +![](images/8e8bf2fee4fc48cc3c767a37d259699db593595abfdb306855a9f6b9241bc752.jpg) +A microwave behind an apple +Fig. 1: Text-to-Image models struggle to generate images that faithfully represent the spatial relationships mentioned in the input prompt. We develop REVISION, an efficient rendering pipeline that enables a training-free and guidance-based mechanism to address this shortcoming. Our method results in improvements in spatial reasoning for T2I models for three dimensional relationships demonstrated by consistently higher scores on VISOR and T2I-CompBench benchmarks. + +![](images/76fca434fec087057a42d3462643fd2d53ac20f5be30c7c47ce731c9f18fee6b.jpg) +A cake in front of a cup + +![](images/765779b93df56a21d230d34ff9c36b2e6ea3428323415e795a67e6f8be813a29.jpg) +A bottle to the left of a boat +An oven behind a wine glass + +![](images/ecdb6d160e935368543ee87bb0a3bd1689064e5f4f58931a6228eb103ecf6471.jpg) +A stop sign to the right of a dog +A cat in front of a boat + +![](images/6359d10c9e112614aed27799b779bae880c535b1ac2d2b0391460d1261907017.jpg) +A wine glass to the left of a boat +A car behind a fire hydrant +Depth Relationships + +DALL-Eval [8], VISOR [15], and T2I-CompBench [18] have found that generative vision-language models suffer from a common mode of failure - their inability to correctly reason over spatial relationships. + +We postulate that the lack of spatial understanding in generative vision-language models is a result of the lack of guidance from image-text datasets. Compared to T2I models, graphics rendering tools such as Blender allow deterministic and accurate object placement, but are limited by their lower visual detail and photorealism and do not have intuitive workflows such as T2I models where users can generate images by simply typing a sentence. To get the best of both worlds, in this work, we develop REVISION, a Blender-based image rendering pipeline which enables the synthesis of images with 101 3-dimensional object (assets), 11 spatial relationships, diverse backgrounds, camera perspectives, and lighting conditions. REVISION parses an input text prompt into assets and relationships and synthesizes the scene using Blender to exactly match the input prompt in terms of both objects and their spatial arrangement. + +In a training-free manner, we leverage images from REVISION as additional guidance for existing T2I methods to their ability to generate spatially accurate + +images, and demonstrate improved performance on VISOR and T2I-CompBench benchmarks. We evaluate (i) the impact of utilizing diverse backgrounds from REVISION, (ii) the trade-off between controllability and photo-realism and (iii) the added generalization to complex prompts achieved by leveraging REVISION. For a holistic study, we introduce an extension to the VISOR benchmark, to include evaluation of depth relationships (in front of/behind). + +To assess the spatial and relational reasoning abilities of MLLMs, we also create the RevQA benchmark. We construct 16 diverse question types and their adversarial variations consisting of negations, conjunctions, and disjunctions. We perform holistic evaluations on 5 state-of-the-art MLLMs and discover significant shortcomings in their ability to accurately address complex spatial reasoning questions. These models also demonstrate a lack of robustness to adversarial perturbations, leading to a substantial decline in their performance. + +The key contributions and findings are summarized below: + +- We develop the REVISION framework, a 3D rendering pipeline that is guaranteed to generate spatially accurate synthetic images, given an input text prompt. An extendable framework, REVISION currently accommodates $100+$ assets across 11 spatial relationships and 3 diverse backgrounds, and support for multiple lighting conditions, camera perspectives, and shadows. +- We present an approach that utilizes images from REVISION in an efficient training-free manner, which results in improved spatial reasoning across multiple benchmarks. Controlled experiments, ablations, and human studies reveal consistent improvements in generating images corresponding to the spatial relationships in the input prompt (as shown in Figure 1). +- We introduce the RevQA question-answering benchmark to evaluate spatial reasoning abilities of multimodal large language models. Our experiments reveal the shortcomings of state-of-the-art MLLMs in reasoning over complex spatial questions and their vulnerability to adversarial perturbations. + +# 2 Related Work + +Generative Models for Image Synthesis. Image generation and synthesis methods have advanced rapidly, progressing from early approaches such as generative adversarial networks (GAN) [16], variational auto-encoders (VAE) [42], and auto-regressive models (ARM) [6], to contemporary text-to-image models including Stable Diffusion [38] and DALL-E [35]. GLIDE [30] adopts classifier-free guidance in T2I and explores the efficacy of CLIP [34] as a text encoder. Compared to GLIDE,Imagen [39] adopts a frozen language model as the text encoder, reducing computational overhead, allowing for usage of large text-only corpus. Multiple variants of T2I models have been developed by leveraging T5-based text encoder [5], T2I priors [32, 37], reward-based fine-tuning [18] and developing refiner models [33] for improved image-text alignment. + +Controllable Image Generation for Spatial Fidelity. To achieve better control over diffusion-based image synthesis, multiple methods have been proposed. ReCo [49], GLIGEN [21], Control-GPT [53], Composable Diffusion [26] + +and ConPreDiff 47 all develop training-based methods to provide additional conditioning for T2I models. SPRIGHT 3 introduces a spatially-focused large-scale dataset, by re-captioning 6 million images from existing vision datasets and demonstrate performance gains through an efficient training methodology. Test-time adaptations have also been proposed - (i) Layout Guidance 7 restricts specific objects to their bounding box location through the modification of cross-attention maps; however it relies on bounding box annotations, (ii) LayoutGPT 12 and LLM-grounded Diffusion 23 leverage large language models (LLMs) to generate layouts and bounding box co-ordinates and, (iii) RealCompo 54 combines multiple generative models for better spatial control. By developing an annotation-free cost-efficient framework we overcome the shortcomings of existing methods through REVISION. + +Synthetic Images for Vision and Language. The flexibility and control provided during creation of synthetic images has led to various visuo-linguistic evaluation benchmarks using rendering tools. CLEVR [19] pioneered the utilization of synthetic objects in simulated scenes for visual compositionality reasoning. Many variants of CLEVR such as CLEVR-Hans [43], CLEVR-Hyp [41], Super-CLEVR [22], and CLEVRER [50] probe multiple facets of multimodal understanding with synthetic images and videos. PaintSkills introduced in DALL-EVAL [8] is an evaluation dataset that measures multiple aspects of a T2I model, which includes spatial reasoning, image-text alignment and social biases. + +Evaluation of Multimodal LLMs. Multiple benchmarks have been proposed that evaluate reasoning capabilities of MLLMs. MMBench [27] evaluates models across 20 different dimensions, for a total of 2974 evaluation instances. The distinctive abilities of MLLMS to differentiate between coarse and fine-grained vision tasks is explored by MME [13] with images sourced from COCO. A limitation across all these benchmarks is that they collect instances from common VL datasets, increasing risk of data leakage and do not evaluate spatial relationships at scale. RevQA fills this gap by developing a diverse set of synthetic and scalable image-question pairs for a holistic evaluation. + +# 3 The REVISION Framework + +REVISION (Figure 2) is a rendering-based framework for generating spatially accurate images from an input prompt. Given a prompt, we generate an image in Blender, where the two object 3D models and the camera view are situated according to the spatial relationship derived from the prompt. The components of REVISION are described below. + +The Asset Library includes a large human-inspected collection of 3D models of realistic objects with variations in texture and shape. Given an object name, the Asset Library randomly selects a matching asset rescaled to fit into a $1m$ cube to ensure that they are sufficiently visible in the final output. The Asset Library features 101 distinct classes of objects, 80 of which are from MS-COCO [24]. Each object class is associated with 3 to 5 royalty-free 3D model + +![](images/8743d3490a8fe3a6398e6472f6027a9774fd35c087901d18ff7f0213ce7cb3e4.jpg) +Fig. 2: REVISION parses a prompt into assets (objects) and the spatial relationship between them and synthesizes a symbolic image in Blender, placing the respective object assets at coordinates corresponding to the parsed spatial relationship. + +Table 1: Spatial relationships in REVISION and their rules for the Coordinate Generator. The objects are positioned from the camera's perspective. + +
RelationSpatial PhrasesCoordinate Constraints Distance (m)
Horizontalto the left, to the rightX and Z are 0.[1, 1.5]
Verticalabove, below, top, bottomX and Y are 0.[0.75, 1]
Nearnear, next to, on the side ofZ is 0.[0.75, 1]
Depthin front of, behindZ is 0. Xobj1=-Xobj2.[1, 1.5]
+ +assets from sketchfab.com, with a total of 410 3D models. REVISION includes 3 background panoramicas (Indoor, Outdoor, and White) from polyhaven.com and a corresponding textured floor asset from Sketchfab. + +The Coordinate Generator deterministically generates 3D coordinates for the objects and the camera, given the names of the objects and the spatial relation extracted from the prompt. As shown in Table [1], REVISION supports four categories of spatial relationships between objects. In our coordinate frame, the X-, Y-, and Z-axis represent depth, horizontal, and vertical relationships respectively. To ensure that the objects are visible and the spatial relationship is obvious from the camera's view, the coordinate values for the objects on all three axes are confined within the range of $[-1m, 1m]$ . The camera is placed at $x = 5m$ with its view always facing the origin point. The camera is at $z = 2.5$ for depth relationships and at $z = 1.5m$ otherwise. + +The Scene Synthesizer assembles a 3D scene consisting of six main components: a camera, a light source, background, floor, and two objects. The two object assets and the camera are placed at their respective coordinates determined by the Coordinate Generator. Then the background asset, which is a 360-degree panorama image (modeled as a large sphere), is centered at the origin. The light source is added to a random position sufficiently higher than all objects in the scene. To prevent objects from appearing to float, the floor asset, a textured hyperplane orthogonal to the Z-axis, is positioned beneath the object asset with the lowest vertical coordinate. This floor placement also enables the object assets to cast shadows, enhancing the realism of the rendered image. + +![](images/d1ed60142e665e16aeea66e96de7dbc1df7a9db6dd8cc1cafa3e5a4d40924b1c.jpg) +Fig. 3: Outputs from the REVISION rendering pipeline for 4 spatial relationships types for identical assets, with (bottom) and without a floor (top). + +The Position Diversifier (Figure 3) ensures diversity in object orientations, background, and the camera angles every time REVISION is invoked. The background is rotated along the Z-axis, giving us a large number of static background options. In order to further diversify the perspective sizes and tilts of the object assets within the camera's view, we add random jitter to the position and orientation of the camera. We also add random small rotations to the objects along the Z-axis and vary the distance between the objects so that they are not always symmetric around the origin. See Supplementary Materials for more details. + +# 4 Improving Spatial Fidelity in T2I Generation + +# 4.1 Training-Free Image Generation with REVISION + +Given an input prompt $(T)$ , we first generate a spatially accurate reference image $(x^{(g)})$ leveraging our REVISION pipeline. We then perform training-free image synthesis to generate an image $I$ , i.e. $\phi(I|x^{(g)}, T)$ , where $\phi$ is a T2I model. We reformulate the standard text-to-image pipeline into an image-to-image pipeline, conditioned by text, as shown in Figure 4. + +Standard diffusion methods such as Stable Diffusion (SD) generate an image by iteratively de-noising a Gaussian noise vector. Stochastic Differential Editing (SDEdit) [28], on the other hand, starts from a guide image $(x^{(g)}$ , in our case), adds Gaussian noise to it, and denoises it to produce the synthesized image $I$ . We use SDEdit within our Stable Diffusion pipeline and perform image generation guided by $x^{(g)}$ . We also explore the ControlNet [51] backbone, which allows fine-grained control over SD. Using ControlNet allows us to address two key points: a) our reference images provide enough spatial information even when low-level features are extracted from them and, b) we can mitigate any attribute-related biases present in the assets. + +![](images/721ebdc641f63184247c42c6a9ff84b89a0befd9b1ca1f98e74ba3e1c2113c31.jpg) +Fig. 4: Given a user-provided input prompt $T$ , we generate a corresponding synthetic image $x^{(g)}$ using REVISION. With input prompt $T$ and guidance $x^{(g)}$ , we perform training-free image synthesis based on existing T2I pipelines such as Stable Diffusion or ControlNet to obtain a spatially accurate image. + +Table 2: The incorporation of REVISION as a guiding framework significantly enhances the spatial reasoning performance of Stable Diffusion (SD) models. Results highlighted in green represent scores achieved with images from REVISION. + +
MethodOA (%)VISOR (%)
uncond cond1234
SD 1.429.8618.8162.9846.6020.116.891.63
+ REVISION53.9652.7197.6977.7961.0244.9027.15
SD 1.528.4317.5161.5944.2718.126.281.35
+ REVISION54.3353.0897.7278.0761.2745.4427.55
SD 2.147.8330.2563.2464.4235.7416.134.70
+ REVISION48.2647.1197.6176.0755.7537.1019.53
+ +# 4.2 Experimental Setup + +We study the efficacy of REVISION on two widely accepted benchmarks for spatial relationship, VISOR [15] and T2I-CompBench [18], which have 25,280 and 300 spatial prompts, respectively. For each evaluation prompt in the respective benchmarks, we generate a corresponding image from our REVISION pipeline and perform training-free image generation as described in Section 4.1. + +We leverage 3 variants of Stable Diffusion (SD), versions 1.4, 1.5, and 2.1 as our baseline models. For ControlNet, we use the canny edge-conditioned SD model. For holistic evaluations, we also report the Inception Score (IS) where applicable. For all subsequent tables, the bold values denote the best performance while underlined values indicate the second-best performance. + +# 4.3 Results and Analysis + +Improvements over Baseline Models - We summarize our representative improvements over the baseline and existing methods, on the VISOR and T2I-CompBench benchmarks in Table 2 and Figure 5 respectively. The results in + +![](images/29a8b81e2ecb320f2bbcdde45c0019047b5208bc41d660e545834c9d079ddce0.jpg) +Fig. 5: Comparing the T2I-CompBench spatial scores of REVISION-based guidance (green) with other leading T2I models and methods (blue). + +Table 3: Results on the VISOR Benchmark. With REVISION, we consistently outperform existing T2I methods on the VISOR benchmark. + +
MethodOA (%)VISOR (%)
uncondcond1234
GLIDE 303.361.9859.066.721.020.170.03
DALLE-mini 1027.1016.1759.6738.3117.506.891.96
DALLE-v2 3563.9337.8959.2773.5947.2323.267.49
Layout Guidance 740.0138.8095.95----
Control-GPT 5348.3344.1765.9769.8051.2035.6720.48
Structured Diffusion 1128.6517.8762.3644.7018.736.571.46
Attend-and-Excite 442.0725.7561.2149.2919.334.560.08
SD 1.4 + REVISION53.9652.7197.6977.7961.0244.9027.15
SD 1.5 + REVISION54.3353.0897.7278.0761.2745.4427.55
SD 2.1 + REVISION48.2647.1197.6176.0755.7537.1019.53
ControlNet + REVISION56.8855.4897.5478.8262.9348.5831.59
+ +Table 2 are shown with reference images on a white background and # of denoising steps = 30. As shown in Table 2 we improve on all aspects of spatial relationships compared to our baseline methods. On SD 1.5, we achieve a $91.1\%$ improvement in Object Accuracy (OA) and a $58.6\%$ improvement on the conditional score. Specifically, we generate objects more accurately and achieve a high $\%$ of accuracy when spatially synthesizing them in the image. Interestingly, through REVISION, we increase the likelihood of consistently generating spatially correct images, as can be seen by the relatively high value of $\mathrm{VISOR}_4$ . On VISOR (Table 3), REVISION enables baseline Stable Diffusion models to consistently outperform existing methods, across all aspects. Compared to the + +![](images/f909d1a2783da051162c7fde64c4d0e26df8ee412473a2bf53921cb0eef6d438.jpg) +Fig.6: REVISION improves T2I-CompBench Spatial Score (0 indicates missing objects, 1 denotes perfect object generation and spatial accuracy.) + +![](images/65024c20325d5f9998f015a6ac777ec08ffd30815e7db04fdea752e75cb325a4.jpg) +Fig.7: Benchmarking the trade-off between spatial accuracy (VISOR) and Inception Score, achieved with REVISION. + +Table 4: VISORcond and Object Accuracy, split across relationship types. σVc and σ0A denote the respective metric's standard deviation w.r.t the relationships. Regardless of the spatial relation, REVISION enables T2I models to consistently produce spatially accurate images, a challenge faced by earlier approaches. + +
MethodVISORcond (%)Object Accuracy (%)
leftrightabovebelowσVcleftrightabovebelowσ0A
GLIDE57.7861.7160.3256.242.463.103.463.493.390.18
DALLE-mini57.8960.1663.7556.143.2922.2921.7433.6230.745.99
DALLE-v256.4756.5160.9963.243.3864.3064.3265.6661.451.77
Control-GPT72.5070.2867.8565.702.9549.8048.2747.9746.951.18
SD 1.4 + REVISION97.5397.4598.0997.660.2952.4252.1156.9354.382.22
SD 1.5 + REVISION97.5797.5398.0597.700.2452.9952.5956.8054.921.94
SD 2.1 + REVISION97.8197.4697.9197.280.3046.7047.9449.7048.711.27
ControlNet + REVISION97.5197.2597.6597.720.2155.1055.1458.9858.292.05
+ +best open-source model, Control-GPT, we achieve a $\Delta$ improvement of $17.69\%$ , $48.12\%$ , and $25.6\%$ on OA, VISORcond, and VISORuncond respectively. + +On T2I-CompBench (Figure 5), we observe similar improvement trends across diverse backgrounds, with baseline models guided by REVISION achieving consistent performance gains on the benchmark. In addition to enhancing spatial accuracy, REVISION improves prompt fidelity by ensuring that images contain all objects mentioned in the input prompt (Figure 6). + +Consistent Performance Across Relationship Types - Across all spatial relationship types, REVISION achieve a consistently high performance score across the VISOR metrics as shown in Table 4; a shortcoming prevalent in other methods. For example, the largest deviation in $\mathrm{VISOR}_{cond}$ performance for ControlNet + REVISION is $0.21\%$ between left and below relationships; in comparison Control-GPT deviates as much as $6.8\%$ for the same. + +![](images/fd15fb30a3fd1352031689c16a05766b0f5d843d9624e9d8d6f922ba2d231e67.jpg) +An apple to the left of a book + +![](images/2b9a9bd4fa3b80931dfc6fa651888bdd65b797e059bc8f00410dab0ef3129ff6.jpg) + +![](images/c424fdc44c484ceb5164ef2b83d97b1335ef330f42a9aa4caace26541fd3b437.jpg) +A truck to the right of a fire hydrant +A potted plant above a dining table + +![](images/fde1ce4ac770e0449cb9a15a26bf1a14f487ffc730fd690b80e3c0b0d470607c.jpg) +A car below an elephant +Background Type = Indoor +Fig. 8: Illustrative examples depicting the variation of generated images across the three variants of backgrounds in REVISION. For each pair, the image on the left is from REVISION and the image on the right is generated from the T2I model. + +![](images/eb49dec07cc0f367017790173f6813a6983e0dfdf00e1742decb61a41ebe1822.jpg) +A handbag to the left of a bird + +![](images/6f9fea7c8caefe11abfcbc55cb149408210efd0065bb88e3cd5424eaadb82135.jpg) + +![](images/220d1e3c8b880b9ad5d670f4f3bad2fb0d3556add840a58a257f7b2ff9c711b2.jpg) +A potted plant to the right of an oven +A vase above a teddy bear + +![](images/fa7a20d80e579ea3268a2016e24afc0f7308cf133f3f304295531026f6f82f82.jpg) +An apple below an airplane +Background Type $=$ Outdoor + +![](images/682b01f5b336b81caec330ec224a34751624a5b422129e437d2b6b57a93689a6.jpg) +An orange to the left of a sink + +![](images/54abcc2ae2487e02aaae0b9041085300b6979d4ff4938017a3ff6818cc8996d3.jpg) + +![](images/65e38c054bbcab72befa979e67b0dfbc583d78f9d1d235dd71fcf34232662300.jpg) +A cow to the right of a couch + +![](images/dc6a191b76edd8a97f826851aca78ddc8ad020e4c710b33f358e4f5c17bc2f25.jpg) +A bed above a sheep +A bottle below a wine glass +Background Type = White + +# 4.4 Ablation Studies + +Impact of Background - In Table 5 we enumerate the impact of the background types in the images from the REVISION pipeline and the downstream trade-off between VISOR performance and model diversity. Utilizing white backgrounds that exclusively feature the two objects in question minimizes potential distractions for the model. Conversely, when the model is presented with initial reference images incorporating indoor or outdoor backgrounds, it exhibits the capacity to identify and leverage distractor objects, resulting in the generation of diverse images. As shown in Figure 8 all generated images maintain spatial accuracy, but noisier reference images result in greater diversity. + +Controllability vs Photo-Realism - In this setup, we study the impact of the # of denoising steps and its trade-off with photo-realism. As shown in Figure 7 that while the performance on VISOR deteriorates with additional # of denoising steps, it improves the model's ability to be more diverse and photorealistic. In Figure 9, we demonstrate that by utilizing REVISION, baseline + +Table 5: The impact of the 3 background types in the REVISION pipeline on the VISOR benchmark. While best performance is achieved with a white background, diverse outputs are attained with the outdoor background type. + +
Model Background IS (↑) OA (%)VISOR (%)
uncondcond1234
SD 1.4White16.1653.9652.7197.6977.7961.0244.927.15
Indoor19.1148.5345.1292.9774.8253.7934.7817.09
Outdoor20.1644.3241.8094.3169.7949.3831.8616.17
SD 1.5White16.2754.3353.0897.7278.0761.2745.4427.55
Indoor19.1148.7745.2892.8574.9353.9634.7717.47
Outdoor19.6643.9941.5194.3669.4848.5831.4616.52
SD 2.1White12.7948.2647.1197.6176.0755.7537.1019.53
Indoor11.5231.0829.3794.5059.8033.9617.406.34
Outdoor10.5136.3734.6795.3465.0541.2323.059.36
+ +![](images/b497e0b5395cedfff65b2586f1a911b345fdb59637f2c5e3b8a984328c3c93d1.jpg) +Fig. 9: Illustrative examples showing the trade-off between photo-realism and denoising steps, while maintaining generating spatially accurate images using REVISION. + +models can preserve their spatial coherency while iteratively demonstrating a higher degree of photo-realism, through more # of denoising steps. + +# 4.5 Extending VISOR for Depth Relationships + +We further extend the VISOR benchmark for Depth relationships (in front of/behind). We utilize Depth Anything [46] for generation of depth maps and OWLv2 [29] for object detection. Given a T2I generated image $I$ and its prompt $T$ that contain two objects $o_1, o_2$ , we obtain its depth map $I_D$ using Depth Anything. We then retrieve the centroids detected for the two objects $c_{o_1}, c_{o_2}$ using OWLv2. At these centroid coordinates, we acquire the depth values for the two objects from the depth map $I_D(c_{o_1}), I_D(c_{o_2})$ . We check if the acquired depth values match the spatial relationship in the prompt, and evaluate similar + +Table 6: Comparing baseline methods against REVISION-guided image synthesis on depth relationships. DS denotes the # of Denoising Steps. + +
SettingOA (%)VISOR (%)
uncond1234
SD 1.5 (baseline)41.5227.1560.1231.8213.193.47
+ REVISION (DS=30)58.3229.6264.7434.5314.994.22
+ REVISION (DS=35)52.0532.0868.1137.9217.514.78
+ REVISION (DS=40)47.4330.4665.5035.9915.744.64
+ +to VISOR. As shown in Table 6, REVISION improves VISOR scores across all metrics and across multiple denoising steps. + +# 4.6 Human Evaluations + +To verify the generalizability of REVISION-based guidance on T2I models, we perform 2 distinct experiments and conduct human evaluations for validation. For each experiment, we independently sample 200 generated images and take the average scores across 4 workers. We also report unanimous (100%) and majority (75%) agreements between the workers for each experiment. + +Prompts of Multiple Objects and Relationships - In this experiment, we generate reference images using prompts that include 2 spatial phrases and 3 objects, and use these images to guide T2I generation. Each generated image is evaluated for accuracy based on the input spatial prompts. We achieve an accuracy of $79.62\%$ when at least 1 phrase is correctly represented in the image and $46.5\%$ when both phrases are correctly represented. The unanimous and majority agreements among evaluators are $64.5\%$ and $86.5\%$ , respectively. + +Out-of-Distribution Objects - We consider prompts containing exactly one object not found in the REVISION Asset Library. Given a prompt that mentions an OOD object, we find the semantically closest object (list in Supplementary Material) in our library and use their corresponding image as guidance. For example, we generate an image of "a helicopter above a bicycle" by providing a reference image of "an airplane above a bicycle". An accuracy of $63.62\%$ is found with an unanimous and majority agreement of $67\%$ and $90.5\%$ , respectively. + +# 5 RevQA: A Spatial Reasoning Benchmark for MLLMs + +We leverage the determinism of the REVISION pipeline to construct a new visual question answering benchmark (RevQA) for evaluating the spatial reasoning abilities of multimodal large language models. + +![](images/c65468606980fc568f6fc02eda80285fdf7e395438d9b8b5474353829f684fa7.jpg) +Fig. 10: The RevQA Benchmark. Using the REVISION pipeline, we generate spatially accurate images and formulate 16 question types from a given caption. We leverage these generated questions and image, benchmarking Multimodal Large Language Models in their abilities to reason over spatial relationships. + +
Question TypeQuestionAnswer
Simple SpatialIs there an airplane above a chair?Yes
Opposite SpatialIs there a chair above an airplane?No
ANDIs there a chair and an airplane?Yes
ORIs there a chair or an airplane?Yes
NOTIs there an airplane not above a chair?No
Double NegativeIs there an airplane not below a chair?Yes
Random ANDIs there an airplane and a hot dog?No
Random ORIs there an airplane or a hot dog?Yes
Random SpatialIs there an airplane above a hot dog?No
Random Combined ANDIs there an airplane above a chair and is there an airplane above a hot dog?No
Random Combined ORIs there an airplane above a hot dog or is there an airplane above a chair?Yes
Adversarial ANDIs there a helicopter and a chair?No
Adversarial ORIs there a helicopter or a chair?Yes
Adversarial SpatialIs there a helicopter above a chair?No
Adversarial Combined ANDIs there a helicopter above a chair and is there an airplane above a chair?No
Adversarial Combined ORIs there a helicopter above a chair or is there an airplane above a chair?Yes
+ +Question Generation. The benchmark contains 16 types of yes-no questions for a REVISION-generated image, consisting of negations, conjunctions, and disjunctions, building on prior work on logic-based visual question answering [14]. Each question type evaluates a combination of spatial and logical reasoning abilities in multimodal large language models (MLLMs) (Figure 10). + +Among the 16 types, we incorporate Random and Adversarial types of questions to further evaluate the robustness and reliability of MLLMs using simple templated transformations. In Random types of questions, we replace an object (visible in the image) in the question with another randomly picked object from REVISION's Asset Library. For the Adversarial set of questions, we replace one of the objects with another that is semantically and visually close. In addition to benchmarking their robustness, these questions allow simultaneous evaluation of the fine-grained spatial perception and reasoning abilities of these models. To alleviate any order bias in instances which contain multiple questions (see Combined in Figure 10), we randomly switch the order between them. + +Evaluation Setup and Results We sample 50k image-question pairs and benchmark 5 open-source state-of-the-art MLLMs - LLaVA 1.5 [25], Fuyu-8B [2], InstructBLIP [9], LLaMA-Adapter 2.1 [52] and Qwen-VL-Chat [1]. We instruct all models to generate binary responses and set the temperature $= 0$ , to remove stochasticity in the generated responses. + +We present our evaluation results in Table 7 and find that all models have a large gap in performance in reasoning over spatial relationships. While most models reason well over simple spatial relationships, they have a large performance drop when presented with the opposite spatial relationships. For example, LLaVA-1.5, the best performing model, has a $58.17\%$ decrease in performance when probed with simple vs opposite spatial questions. This can be attributed to: $(a)$ insufficient training data for rare object relationships, such as less instances of an "elephant above a person" than vice versa; $b)$ the inability of vision + +Table 7: Performances of 5 MLLMs across the 16 types of questions in RevQA. Most models perform worse than random (50%) when reasoning over Opposite Spatial relationships and Double Negative questions. All models have a significant drop in performance with Random/Adversarial questions, in comparison to their simpler versions. + +
Question TypeLLaVa 1.5Fuyu-8BInstructBLIPLLaMA-Adapter 2.1Qwen-VL Chat
Simple Spatial0.9420.7020.8340.5790.940
Opposite Spatial0.3940.2870.1840.4190.402
AND0.9350.8870.9570.8580.889
OR0.9950.3960.5980.7220.949
NOT0.9260.6190.3560.5040.583
Double Negative0.2670.3470.6650.4900.212
Random AND0.9340.3080.6750.6160.978
Random OR0.9250.1780.1940.1940.324
Random Spatial0.9250.3700.6860.7900.919
Random Combined AND0.1160.5020.6270.8000.567
Random Combined OR0.9680.5360.4140.0030.506
Adversarial AND0.6610.1840.5420.6410.789
Adversarial OR0.9210.1880.4430.1560.685
Adversarial Spatial0.5590.3350.7770.8930.615
Adversarial Combined AND0.1320.5390.6950.8050.695
Adversarial Combined OR0.9530.4560.3860.0030.254
Average0.7200.4460.5980.5780.642
+ +encoders like CLIP to capture subtle semantic differences. MLLMs also struggle with negation, possibly because image captions do not capture enough negations; e.g. COCO Captions only contain $0.97\%$ occurrences of 'not'. All models significantly suffer when presented with questions that consist of double negatives, which evaluate the models' ability to reason of negations and spatial relationships in tandem. Furthermore, all models suffer under adversarial settings in comparison to their simpler counterparts; comparing LLaVA's performance for AND and Adversarial Combined AND questions, we find a $85.88\%$ $(0.935\rightarrow 0.132)$ drop in performance. We also observe a larger decline in performance for Adversarial questions than for the Random set of questions hinting that while models independently perform well at object recognition and simple spatial relationships, combining them adversarily significantly reduces performance. + +# 6 Conclusion + +In this work, we introduce REVISION, a framework designed for training-free enhancement of spatial relationships in Text-to-Image models and RevQA, a benchmark to evaluate the spatial reasoning abilities of multimodal large language models. Our results demonstrate the effectiveness of leveraging 3D rendering pipelines as a cost-efficient approach for developing generative models with robust reasoning capabilities. REVISION is modular and can easily be extended to incorporate additional features, assets, and relationships. We hope our method inspires future research at the intersection of computer graphics and generative AI, enabling safe deployment of these systems in the real world. + +# Acknowledgements + +The authors acknowledge resources provided by Research Computing at Arizona State University. The authors also acknowledge technical access and support from ASU Enterprise Technology. This work was supported by NSF Robust Intelligence program grants #1750082 and #2132724. TG was supported by Microsoft's Accelerating Foundation Model Research (AFMR) program and UMBC's Strategic Award for Research Transitions (START). The views and opinions of the authors expressed herein do not necessarily state or reflect those of the funding agencies and employers. + +# References + +1. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond (2023), https://arxiv.org/abs/2308.12966 +2. Bavishi, R., Elsen, E., Hawthorne, C., Nye, M., Odena, A., Somani, A., Taşırlar, S.: Introducing our multimodal models (2023), https://www.adept.ai/blog/fuyu-8b +3. Chatterjee, A., Stan, G.B.M., Aflalo, E., Paul, S., Ghosh, D., Gokhale, T., Schmidt, L., Hajishirzi, H., Lal, V., Baral, C., Yang, Y.: Getting it right: Improving spatial consistency in text-to-image models (2024), https://arxiv.org/abs/2404.01197 +4. Chefer, H., Alaluf, Y., Vinker, Y., Wolf, L., Cohen-Or, D.: Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG) 42(4), 1-10 (2023) +5. Chen, J., Jincheng, Y., Chongjian, G., Yao, L., Xie, E., Wang, Z., Kwok, J., Luo, P., Lu, H., Li, Z.: Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In: The Twelfth International Conference on Learning Representations (2023) +6. Chen, M., Radford, A., Child, R., Wu, J., Jun, H., Luan, D., Sutskever, I.: Generative pretraining from pixels. In: Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event. Proceedings of Machine Learning Research, vol. 119, pp. 1691-1703. PMLR (2020), http://proceedings.mlr.press/v119/chen20s.html +7. Chen, M., Laina, I., Vedaldi, A.: Training-free layout control with cross-attention guidance (2023) +8. Cho, J., Zala, A., Bansal, M.: Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3043-3054 (2023) +9. Dai, W., Li, J., Li, D., Tiong, A., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=vvoWPYqZJA +10. Dayma, B., Patil, S., Cuenca, P., Saifullah, K., Abraham, T., Le Khac, P., Melas, L., Ghosh, R.: Dall-e mini (2021). https://doi.org/10.5281/zenodo.5146400. https://github.com/borisdayma/dalle-mini +11. Feng, W., He, X., Fu, T.J., Jampani, V., Akula, A.R., Narayana, P., Basu, S., Wang, X.E., Wang, W.Y.: Training-free structured diffusion guidance for compositional text-to-image synthesis. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=PUIqjT4rzq7 + +12. Feng, W., Zhu, W., Fu, T.j., Jampani, V., Akula, A., He, X., Basu, S., Wang, X.E., Wang, W.Y.: Layoutgpt: Compositional visual planning and generation with large language models. Advances in Neural Information Processing Systems 36 (2024) +13. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Yang, J., Zheng, X., Li, K., Sun, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. ArXiv preprint abs/2306.13394 (2023), https://arxiv.org/abs/2306.13394 +14. Gokhale, T., Banerjee, P., Baral, C., Yang, Y.: Vqa-lol: Visual question answering under the lens of logic. In: European conference on computer vision. pp. 379-396. Springer (2020) +15. Gokhale, T., Palangi, H., Nushi, B., Vineet, V., Horvitz, E., Kamar, E., Baral, C., Yang, Y.: Benchmarking spatial relationships in text-to-image generation. arXiv preprint arXiv:2212.10015 (2022) +16. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139–144 (2020) +17. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-or, D.: Prompt-to-prompt image editing with cross-attention control. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=CDixzkzeyb +18. Huang, K., Sun, K., Xie, E., Li, Z., Liu, X.: T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. Advances in Neural Information Processing Systems 36, 78723-78747 (2023) +19. Johnson, J., Hariharan, B., van der Maaten, L., Fei-Fei, L., Zitnick, C.L., Girshick, R.B.: CLEVR: A diagnostic dataset for compositional language and elementary visual reasoning. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017. pp. 1988-1997. IEEE Computer Society (2017). https://doi.org/10.1109/CVPR.2017.215. https://doi.org/10.1109/CVPR.2017.215 +20. Kamath, A., Hessel, J., Chang, K.W.: What's "up" with vision-language models? investigating their struggle with spatial reasoning. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. pp. 9161-9175 (2023) +21. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22511-22521 (2023) +22. Li, Z., Wang, X., Stengel-Eskin, E., Kortylewski, A., Ma, W., Van Durme, B., Yuille, A.L.: Super-clevr: A virtual benchmark to diagnose domain robustness in visual reasoning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14963-14973 (2023) +23. Lian, L., Li, B., Yala, A., Darrell, T.: Llm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. ArXiv preprint abs/2305.13655 (2023), https://arxiv.org/abs/2305.13655 +24. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +25. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024) + +26. Liu, N., Li, S., Du, Y., Torralba, A., Tenenbaum, J.B.: Compositional visual generation with composable diffusion models. In: European Conference on Computer Vision. pp. 423-439. Springer (2022) +27. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281 (2023) +28. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. In: The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net (2022), https://openreview.net/forum?id=aBsCjcPu_tE +29. Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Dosovitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., et al.: Simple open-vocabulary object detection. In: European Conference on Computer Vision. pp. 728-755. Springer (2022) +30. Nichol, A.Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: GLIDE: towards photorealistic image generation and editing with text-guided diffusion models. In: Chaudhuri, K., Jegelka, S., Song, L., Szepesvári, C., Niu, G., Sabato, S. (eds.) International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA. Proceedings of Machine Learning Research, vol. 162, pp. 16784-16804. PMLR (2022), https://proceedings.mlr.press/v162/nichol22a.html +31. Parmar, G., Kumar Singh, K., Zhang, R., Li, Y., Lu, J., Zhu, J.Y.: Zero-shot image-to-image translation. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-11 (2023) +32. Patel, M., Kim, C., Cheng, S., Baral, C., Yang, Y.: Eclipse: A resource-efficient text-to-image prior for image generations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9069-9078 (2024) +33. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: SDXL: Improving latent diffusion models for high-resolution image synthesis. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=di52zR8xgf +34. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event. Proceedings of Machine Learning Research, vol. 139, pp. 8748-8763. PMLR (2021), http://proceedings.mlr.press/v139/radford21a.html +35. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022) +36. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event. Proceedings of Machine Learning Research, vol. 139, pp. 8821-8831. PMLR (2021), http://proceedings.mlr.press/v139/ramesh21a.html +37. Razzhigaev, A., Shakhmatov, A., Maltseva, A., Arkhipkin, V., Pavlov, I., Ryabov, I., Kuts, A., Panchenko, A., Kuznetsov, A., Dimitrov, D.: Kandinsky: An improved + +text-to-image synthesis with image prior and latent diffusion. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations. pp. 286-295 (2023) +38. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) +39. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022) +40. Salimans, T., Goodfellow, I.J., Zaremba, W., Cheung, V., Radford, A., Chen, X.: Improved techniques for training gans. In: Lee, D.D., Sugiyama, M., von Luxburg, U., Guyon, I., Garnett, R. (eds.) Advances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, December 5-10, 2016, Barcelona, Spain. pp. 2226-2234 (2016), https://proceedings.neurips.cc/paper/2016/bitstream/8a3363abe792db2d8761d6403605aab7-Abstract.html +41. Sampat, S.K., Kumar, A., Yang, Y., Baral, C.: CLEVR_HYP: A challenge dataset and baselines for visual question answering with hypothetical actions over images. In: Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. pp. 3692-3709. Association for Computational Linguistics, Online (2021). https://doi.org/10.18653/v1/2021.naacl-main.289. https://aclanthology.org/2021.naacl-main.289 +42. Sohn, K., Lee, H., Yan, X.: Learning structured output representation using deep conditional generative models. In: Cortes, C., Lawrence, N.D., Lee, D.D., Sugiyama, M., Garnett, R. (eds.) Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada. pp. 3483-3491 (2015), https://proceedings.neurips.cc/paper/2015/bitstream/8d55a249e6baa5c06772297520da2051-Abstract.html +43. Stammer, W., Schramowski, P., Kersting, K.: Right for the right concept: Revising neuro-symbolic concepts by interacting with their explanations. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 3619-3629. Computer Vision Foundation / IEEE (2021). https://doi.org/10.1109/CVPR46437.2021.00362, https://openaccess.thecvf.com/content/CVPR2021/html/Stammer_Right_for_the_Right_Concept_Revising_Neuro-Symbolic_Concepts_by_Interacting_CVPR_2021_paper.html +44. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. ArXiv preprint abs/2312.11805 (2023), https://arxiv.org/abs/2312.11805 +45. Wake, N., Kanehira, A., Sasabuchi, K., Takamatsu, J., Ikeuchi, K.: Gpt-4v (ision) for robotics: Multimodal task planning from human demonstration. arXiv preprint arXiv:2311.12015 (2023) +46. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. ArXiv preprint abs/2401.10891 (2024), https://arxiv.org/abs/2401.10891 + +47. Yang, L., Liu, J., Hong, S., Zhang, Z., Huang, Z., Cai, Z., Zhang, W., Cui, B.: Improving diffusion-based image synthesis with context prediction. Advances in Neural Information Processing Systems 36 (2024) +48. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of lmms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421 9(1), 1 (2023) +49. Yang, Z., Wang, J., Gan, Z., Li, L., Lin, K., Wu, C., Duan, N., Liu, Z., Liu, C., Zeng, M., et al.: Reco: Region-controlled text-to-image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14246-14255 (2023) +50. Yi, K., Gan, C., Li, Y., Kohli, P., Wu, J., Torralba, A., Tenenbaum, J.B.: CLEVRER: collision events for video representation and reasoning. In: 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net (2020), https://openreview.net/forum?id=HkxYzANYDB +51. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023) +52. Zhang, R., Han, J., Liu, C., Gao, P., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199 (2023) +53. Zhang, T., Zhang, Y., Vineet, V., Joshi, N., Wang, X.: Controllable text-to-image generation with gpt-4. arXiv preprint arXiv:2305.18583 (2023) +54. Zhang, X., Yang, L., Cai, Y., Yu, Z., Xie, J., Tian, Y., Xu, M., Tang, Y., Yang, Y., Cui, B.: Realcompo: Dynamic equilibrium between realism and compositionality improves text-to-image diffusion models. arXiv preprint arXiv:2402.12908 (2024) \ No newline at end of file diff --git a/2024/REVISION_ Rendering Tools Enable Spatial Fidelity in Vision-Language Models/images.zip b/2024/REVISION_ Rendering Tools Enable Spatial Fidelity in Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7d082c8d7c493b0c97e0f10a58c43ea79a6e581a --- /dev/null +++ b/2024/REVISION_ Rendering Tools Enable Spatial Fidelity in Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:18b3d673966bbfa497603b120e7d295a2f2642eacc25674fb1bc08386aad49cb +size 903784 diff --git a/2024/REVISION_ Rendering Tools Enable Spatial Fidelity in Vision-Language Models/layout.json b/2024/REVISION_ Rendering Tools Enable Spatial Fidelity in Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3e37f41808ce8efa093d5767331de8f13a5a487c --- /dev/null +++ b/2024/REVISION_ Rendering Tools Enable Spatial Fidelity in Vision-Language Models/layout.json @@ -0,0 +1,10243 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 152, + 111, + 462, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 111, + 462, + 148 + ], + "spans": [ + { + "bbox": [ + 152, + 111, + 462, + 148 + ], + "type": "text", + "content": "REVISION: Rendering Tools Enable Spatial Fidelity in Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 186, + 167, + 430, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 167, + 430, + 196 + ], + "spans": [ + { + "bbox": [ + 186, + 167, + 430, + 196 + ], + "type": "text", + "content": "Agneet Chatterjee " + }, + { + "bbox": [ + 186, + 167, + 430, + 196 + ], + "type": "inline_equation", + "content": "\\star 1" + }, + { + "bbox": [ + 186, + 167, + 430, + 196 + ], + "type": "text", + "content": " Yiran Luo " + }, + { + "bbox": [ + 186, + 167, + 430, + 196 + ], + "type": "inline_equation", + "content": "\\star 1" + }, + { + "bbox": [ + 186, + 167, + 430, + 196 + ], + "type": "text", + "content": " Tejas Gokhale" + }, + { + "bbox": [ + 186, + 167, + 430, + 196 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 186, + 167, + 430, + 196 + ], + "type": "text", + "content": ", Yezhou Yang" + }, + { + "bbox": [ + 186, + 167, + 430, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 186, + 167, + 430, + 196 + ], + "type": "text", + "content": ", and Chitta Baral" + }, + { + "bbox": [ + 186, + 167, + 430, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 215, + 201, + 398, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 201, + 398, + 224 + ], + "spans": [ + { + "bbox": [ + 215, + 201, + 398, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 215, + 201, + 398, + 224 + ], + "type": "text", + "content": " Arizona State University \n" + }, + { + "bbox": [ + 215, + 201, + 398, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 215, + 201, + 398, + 224 + ], + "type": "text", + "content": " University of Maryland, Baltimore County" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 258, + 455, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 258, + 455, + 479 + ], + "spans": [ + { + "bbox": [ + 159, + 258, + 455, + 479 + ], + "type": "text", + "content": "Abstract. Text-to-Image (T2I) and multimodal large language models (MLLMs) have been adopted in solutions for several computer vision and multimodal learning tasks. However, it has been found that such vision-language models lack the ability to correctly reason over spatial relationships. To tackle this shortcoming, we develop the REVISION framework which improves spatial fidelity in vision-language models. REVISION is a 3D rendering based pipeline that generates spatially accurate synthetic images, given a textual prompt. REVISION is an extendable framework, which currently supports " + }, + { + "bbox": [ + 159, + 258, + 455, + 479 + ], + "type": "inline_equation", + "content": "100+" + }, + { + "bbox": [ + 159, + 258, + 455, + 479 + ], + "type": "text", + "content": " 3D assets, 11 spatial relationships, all with diverse camera perspectives and backgrounds. Leveraging images from REVISION as additional guidance in a training-free manner consistently improves the spatial consistency of T2I models across all spatial relationships, achieving competitive performance on the VISOR and T2I-CompBench benchmarks. We also design RevQA, a question-answering benchmark to evaluate the spatial reasoning abilities of MLLMs, and find that state-of-the-art models are not robust to complex spatial reasoning under adversarial settings. Our results and findings indicate that utilizing rendering-based frameworks is an effective approach for developing spatially-aware generative models. Code and data available at: https://github.com/agneet42/revision" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 488, + 453, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 488, + 453, + 499 + ], + "spans": [ + { + "bbox": [ + 160, + 488, + 453, + 499 + ], + "type": "text", + "content": "Keywords: Text to Image " + }, + { + "bbox": [ + 160, + 488, + 453, + 499 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 488, + 453, + 499 + ], + "type": "text", + "content": " Spatial Relationships " + }, + { + "bbox": [ + 160, + 488, + 453, + 499 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 488, + 453, + 499 + ], + "type": "text", + "content": " Rendering Graphics" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 521, + 230, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 230, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 230, + 533 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 548, + 482, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 482, + 646 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 482, + 646 + ], + "type": "text", + "content": "Generative vision-language models [36,44] represent a significant step towards developing multimodal systems that bridge the gap between computer vision and natural language processing. Text-to-image (T2I) models [5,38] convert text prompts to high-quality images, while multimodal large language models (MLLMs) [25,48] process images as inputs, and generate rich and coherent natural language outputs in response. As a result, these models have found diverse applications in robotics [45], image editing [17], image-to-image translation [31], and more. However, recent studies [20] and benchmarks such as" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 654, + 351, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 654, + 351, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 654, + 351, + 666 + ], + "type": "text", + "content": "* Equal contribution. Correspondence to agneet@asu.edu" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 200, + 180 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 200, + 180 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 200, + 180 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 200, + 180 + ], + "type": "image", + "image_path": "7c26d986754b6ea01d566b4a44133234a2f36e3122ee36b599032b507cb86331.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 201, + 114, + 266, + 180 + ], + "blocks": [ + { + "bbox": [ + 201, + 114, + 266, + 180 + ], + "lines": [ + { + "bbox": [ + 201, + 114, + 266, + 180 + ], + "spans": [ + { + "bbox": [ + 201, + 114, + 266, + 180 + ], + "type": "image", + "image_path": "9e04c3522ebd952e052c4b205571fe673306b234f82a0b5915a7a654558559e0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 183, + 264, + 191 + ], + "lines": [ + { + "bbox": [ + 203, + 183, + 264, + 191 + ], + "spans": [ + { + "bbox": [ + 203, + 183, + 264, + 191 + ], + "type": "text", + "content": "A couch below an airplane" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 268, + 114, + 334, + 180 + ], + "blocks": [ + { + "bbox": [ + 268, + 114, + 334, + 180 + ], + "lines": [ + { + "bbox": [ + 268, + 114, + 334, + 180 + ], + "spans": [ + { + "bbox": [ + 268, + 114, + 334, + 180 + ], + "type": "image", + "image_path": "e35dd47badc8aabfbb8b7237ed66fef07c401b20f7eb70a643fdae6957fa70e3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 271, + 183, + 331, + 191 + ], + "lines": [ + { + "bbox": [ + 271, + 183, + 331, + 191 + ], + "spans": [ + { + "bbox": [ + 271, + 183, + 331, + 191 + ], + "type": "text", + "content": "A suitcase above a bicycle" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 335, + 114, + 402, + 180 + ], + "blocks": [ + { + "bbox": [ + 335, + 114, + 402, + 180 + ], + "lines": [ + { + "bbox": [ + 335, + 114, + 402, + 180 + ], + "spans": [ + { + "bbox": [ + 335, + 114, + 402, + 180 + ], + "type": "image", + "image_path": "90f4c76c193d88705f749e063f08281fab8b2ef784e66be09fa74c5b5a390a6b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 341, + 183, + 396, + 191 + ], + "lines": [ + { + "bbox": [ + 341, + 183, + 396, + 191 + ], + "spans": [ + { + "bbox": [ + 341, + 183, + 396, + 191 + ], + "type": "text", + "content": "An umbrella above a car" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 403, + 114, + 471, + 180 + ], + "blocks": [ + { + "bbox": [ + 403, + 114, + 471, + 180 + ], + "lines": [ + { + "bbox": [ + 403, + 114, + 471, + 180 + ], + "spans": [ + { + "bbox": [ + 403, + 114, + 471, + 180 + ], + "type": "image", + "image_path": "87fcd69edb6c4af5fef0e3ca554e07a68f07ddb7502e9d5411fef4a5fa853612.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 122, + 479, + 171 + ], + "lines": [ + { + "bbox": [ + 471, + 122, + 479, + 171 + ], + "spans": [ + { + "bbox": [ + 471, + 122, + 479, + 171 + ], + "type": "text", + "content": "Vertical Relationships" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 133, + 194, + 200, + 262 + ], + "blocks": [ + { + "bbox": [ + 143, + 181, + 187, + 194 + ], + "lines": [ + { + "bbox": [ + 143, + 181, + 187, + 194 + ], + "spans": [ + { + "bbox": [ + 143, + 181, + 187, + 194 + ], + "type": "text", + "content": "An apple on the top of a cake" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 194, + 200, + 262 + ], + "lines": [ + { + "bbox": [ + 133, + 194, + 200, + 262 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 200, + 262 + ], + "type": "image", + "image_path": "4ae198df6775e744f811e430e5e5bd8c428e40afc03b4a8c0d27f81eb5b53f75.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 264, + 196, + 271 + ], + "lines": [ + { + "bbox": [ + 136, + 264, + 196, + 271 + ], + "spans": [ + { + "bbox": [ + 136, + 264, + 196, + 271 + ], + "type": "text", + "content": "A truck to the left of a train" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 201, + 194, + 267, + 261 + ], + "blocks": [ + { + "bbox": [ + 201, + 194, + 267, + 261 + ], + "lines": [ + { + "bbox": [ + 201, + 194, + 267, + 261 + ], + "spans": [ + { + "bbox": [ + 201, + 194, + 267, + 261 + ], + "type": "image", + "image_path": "e89d65537d44cae701d6ed8b9423f5ec0d04e7ffce5727d5c3c333c5ac1a0db6.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 264, + 262, + 271 + ], + "lines": [ + { + "bbox": [ + 205, + 264, + 262, + 271 + ], + "spans": [ + { + "bbox": [ + 205, + 264, + 262, + 271 + ], + "type": "text", + "content": "ATV to the right of a bird" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 268, + 194, + 334, + 261 + ], + "blocks": [ + { + "bbox": [ + 268, + 194, + 334, + 261 + ], + "lines": [ + { + "bbox": [ + 268, + 194, + 334, + 261 + ], + "spans": [ + { + "bbox": [ + 268, + 194, + 334, + 261 + ], + "type": "image", + "image_path": "dd9e0614269e975ec046219d275a4286e183545a69efb11318ff714eb243193e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 335, + 194, + 402, + 261 + ], + "blocks": [ + { + "bbox": [ + 335, + 194, + 402, + 261 + ], + "lines": [ + { + "bbox": [ + 335, + 194, + 402, + 261 + ], + "spans": [ + { + "bbox": [ + 335, + 194, + 402, + 261 + ], + "type": "image", + "image_path": "7a1c1f888612761527d7ba055e32978e2a258dce5139a07d1c7564f1b7834ab0.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 403, + 194, + 470, + 261 + ], + "blocks": [ + { + "bbox": [ + 406, + 181, + 466, + 194 + ], + "lines": [ + { + "bbox": [ + 406, + 181, + 466, + 194 + ], + "spans": [ + { + "bbox": [ + 406, + 181, + 466, + 194 + ], + "type": "text", + "content": "A book on the bottom of a sports ball" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 403, + 194, + 470, + 261 + ], + "lines": [ + { + "bbox": [ + 403, + 194, + 470, + 261 + ], + "spans": [ + { + "bbox": [ + 403, + 194, + 470, + 261 + ], + "type": "image", + "image_path": "00556fc2135078c5ac2360f2c7ab278800bcfb52c396574dfc213f32a99a0db4.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 200, + 479, + 256 + ], + "lines": [ + { + "bbox": [ + 471, + 200, + 479, + 256 + ], + "spans": [ + { + "bbox": [ + 471, + 200, + 479, + 256 + ], + "type": "text", + "content": "Hrnnnne nnnnne" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 133, + 274, + 200, + 341 + ], + "blocks": [ + { + "bbox": [ + 133, + 274, + 200, + 341 + ], + "lines": [ + { + "bbox": [ + 133, + 274, + 200, + 341 + ], + "spans": [ + { + "bbox": [ + 133, + 274, + 200, + 341 + ], + "type": "image", + "image_path": "8e8bf2fee4fc48cc3c767a37d259699db593595abfdb306855a9f6b9241bc752.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 139, + 343, + 193, + 355 + ], + "lines": [ + { + "bbox": [ + 139, + 343, + 193, + 355 + ], + "spans": [ + { + "bbox": [ + 139, + 343, + 193, + 355 + ], + "type": "text", + "content": "A microwave behind an apple" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 365, + 482, + 431 + ], + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 431 + ], + "type": "text", + "content": "Fig. 1: Text-to-Image models struggle to generate images that faithfully represent the spatial relationships mentioned in the input prompt. We develop REVISION, an efficient rendering pipeline that enables a training-free and guidance-based mechanism to address this shortcoming. Our method results in improvements in spatial reasoning for T2I models for three dimensional relationships demonstrated by consistently higher scores on VISOR and T2I-CompBench benchmarks." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 201, + 274, + 267, + 341 + ], + "blocks": [ + { + "bbox": [ + 201, + 274, + 267, + 341 + ], + "lines": [ + { + "bbox": [ + 201, + 274, + 267, + 341 + ], + "spans": [ + { + "bbox": [ + 201, + 274, + 267, + 341 + ], + "type": "image", + "image_path": "76fca434fec087057a42d3462643fd2d53ac20f5be30c7c47ce731c9f18fee6b.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 345, + 260, + 352 + ], + "lines": [ + { + "bbox": [ + 208, + 345, + 260, + 352 + ], + "spans": [ + { + "bbox": [ + 208, + 345, + 260, + 352 + ], + "type": "text", + "content": "A cake in front of a cup" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 268, + 274, + 334, + 341 + ], + "blocks": [ + { + "bbox": [ + 271, + 264, + 331, + 272 + ], + "lines": [ + { + "bbox": [ + 271, + 264, + 331, + 272 + ], + "spans": [ + { + "bbox": [ + 271, + 264, + 331, + 272 + ], + "type": "text", + "content": "A bottle to the left of a boat" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 268, + 274, + 334, + 341 + ], + "lines": [ + { + "bbox": [ + 268, + 274, + 334, + 341 + ], + "spans": [ + { + "bbox": [ + 268, + 274, + 334, + 341 + ], + "type": "image", + "image_path": "765779b93df56a21d230d34ff9c36b2e6ea3428323415e795a67e6f8be813a29.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 270, + 346, + 334, + 353 + ], + "lines": [ + { + "bbox": [ + 270, + 346, + 334, + 353 + ], + "spans": [ + { + "bbox": [ + 270, + 346, + 334, + 353 + ], + "type": "text", + "content": "An oven behind a wine glass" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 336, + 274, + 402, + 341 + ], + "blocks": [ + { + "bbox": [ + 339, + 262, + 400, + 274 + ], + "lines": [ + { + "bbox": [ + 339, + 262, + 400, + 274 + ], + "spans": [ + { + "bbox": [ + 339, + 262, + 400, + 274 + ], + "type": "text", + "content": "A stop sign to the right of a dog" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 336, + 274, + 402, + 341 + ], + "lines": [ + { + "bbox": [ + 336, + 274, + 402, + 341 + ], + "spans": [ + { + "bbox": [ + 336, + 274, + 402, + 341 + ], + "type": "image", + "image_path": "ecdb6d160e935368543ee87bb0a3bd1689064e5f4f58931a6228eb103ecf6471.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 346, + 394, + 352 + ], + "lines": [ + { + "bbox": [ + 343, + 346, + 394, + 352 + ], + "spans": [ + { + "bbox": [ + 343, + 346, + 394, + 352 + ], + "type": "text", + "content": "A cat in front of a boat" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 403, + 274, + 470, + 341 + ], + "blocks": [ + { + "bbox": [ + 406, + 262, + 466, + 274 + ], + "lines": [ + { + "bbox": [ + 406, + 262, + 466, + 274 + ], + "spans": [ + { + "bbox": [ + 406, + 262, + 466, + 274 + ], + "type": "text", + "content": "A wine glass to the left of a boat" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 403, + 274, + 470, + 341 + ], + "lines": [ + { + "bbox": [ + 403, + 274, + 470, + 341 + ], + "spans": [ + { + "bbox": [ + 403, + 274, + 470, + 341 + ], + "type": "image", + "image_path": "6359d10c9e112614aed27799b779bae880c535b1ac2d2b0391460d1261907017.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 346, + 466, + 352 + ], + "lines": [ + { + "bbox": [ + 406, + 346, + 466, + 352 + ], + "spans": [ + { + "bbox": [ + 406, + 346, + 466, + 352 + ], + "type": "text", + "content": "A car behind a fire hydrant" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 471, + 285, + 479, + 331 + ], + "lines": [ + { + "bbox": [ + 471, + 285, + 479, + 331 + ], + "spans": [ + { + "bbox": [ + 471, + 285, + 479, + 331 + ], + "type": "text", + "content": "Depth Relationships" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "bbox": [ + 130, + 457, + 481, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 481, + 495 + ], + "type": "text", + "content": "DALL-Eval [8], VISOR [15], and T2I-CompBench [18] have found that generative vision-language models suffer from a common mode of failure - their inability to correctly reason over spatial relationships." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 130, + 496, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 482, + 640 + ], + "type": "text", + "content": "We postulate that the lack of spatial understanding in generative vision-language models is a result of the lack of guidance from image-text datasets. Compared to T2I models, graphics rendering tools such as Blender allow deterministic and accurate object placement, but are limited by their lower visual detail and photorealism and do not have intuitive workflows such as T2I models where users can generate images by simply typing a sentence. To get the best of both worlds, in this work, we develop REVISION, a Blender-based image rendering pipeline which enables the synthesis of images with 101 3-dimensional object (assets), 11 spatial relationships, diverse backgrounds, camera perspectives, and lighting conditions. REVISION parses an input text prompt into assets and relationships and synthesizes the scene using Blender to exactly match the input prompt in terms of both objects and their spatial arrangement." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "In a training-free manner, we leverage images from REVISION as additional guidance for existing T2I methods to their ability to generate spatially accurate" + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 246, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 246, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 246, + 102 + ], + "type": "text", + "content": "A. Chatterjee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "images, and demonstrate improved performance on VISOR and T2I-CompBench benchmarks. We evaluate (i) the impact of utilizing diverse backgrounds from REVISION, (ii) the trade-off between controllability and photo-realism and (iii) the added generalization to complex prompts achieved by leveraging REVISION. For a holistic study, we introduce an extension to the VISOR benchmark, to include evaluation of depth relationships (in front of/behind)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 482, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 270 + ], + "type": "text", + "content": "To assess the spatial and relational reasoning abilities of MLLMs, we also create the RevQA benchmark. We construct 16 diverse question types and their adversarial variations consisting of negations, conjunctions, and disjunctions. We perform holistic evaluations on 5 state-of-the-art MLLMs and discover significant shortcomings in their ability to accurately address complex spatial reasoning questions. These models also demonstrate a lack of robustness to adversarial perturbations, leading to a substantial decline in their performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 271, + 406, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 271, + 406, + 283 + ], + "spans": [ + { + "bbox": [ + 146, + 271, + 406, + 283 + ], + "type": "text", + "content": "The key contributions and findings are summarized below:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 287, + 482, + 453 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 287, + 482, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 287, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 138, + 287, + 482, + 346 + ], + "type": "text", + "content": "- We develop the REVISION framework, a 3D rendering pipeline that is guaranteed to generate spatially accurate synthetic images, given an input text prompt. An extendable framework, REVISION currently accommodates " + }, + { + "bbox": [ + 138, + 287, + 482, + 346 + ], + "type": "inline_equation", + "content": "100+" + }, + { + "bbox": [ + 138, + 287, + 482, + 346 + ], + "type": "text", + "content": " assets across 11 spatial relationships and 3 diverse backgrounds, and support for multiple lighting conditions, camera perspectives, and shadows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 346, + 482, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 482, + 406 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 482, + 406 + ], + "type": "text", + "content": "- We present an approach that utilizes images from REVISION in an efficient training-free manner, which results in improved spatial reasoning across multiple benchmarks. Controlled experiments, ablations, and human studies reveal consistent improvements in generating images corresponding to the spatial relationships in the input prompt (as shown in Figure 1)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 405, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 405, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 405, + 482, + 453 + ], + "type": "text", + "content": "- We introduce the RevQA question-answering benchmark to evaluate spatial reasoning abilities of multimodal large language models. Our experiments reveal the shortcomings of state-of-the-art MLLMs in reasoning over complex spatial questions and their vulnerability to adversarial perturbations." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 468, + 237, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 237, + 480 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 237, + 480 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 489, + 482, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 489, + 482, + 622 + ], + "spans": [ + { + "bbox": [ + 130, + 489, + 482, + 622 + ], + "type": "text", + "content": "Generative Models for Image Synthesis. Image generation and synthesis methods have advanced rapidly, progressing from early approaches such as generative adversarial networks (GAN) [16], variational auto-encoders (VAE) [42], and auto-regressive models (ARM) [6], to contemporary text-to-image models including Stable Diffusion [38] and DALL-E [35]. GLIDE [30] adopts classifier-free guidance in T2I and explores the efficacy of CLIP [34] as a text encoder. Compared to GLIDE,Imagen [39] adopts a frozen language model as the text encoder, reducing computational overhead, allowing for usage of large text-only corpus. Multiple variants of T2I models have been developed by leveraging T5-based text encoder [5], T2I priors [32, 37], reward-based fine-tuning [18] and developing refiner models [33] for improved image-text alignment." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "Controllable Image Generation for Spatial Fidelity. To achieve better control over diffusion-based image synthesis, multiple methods have been proposed. ReCo [49], GLIGEN [21], Control-GPT [53], Composable Diffusion [26]" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 90, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 90, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 399, + 90, + 447, + 100 + ], + "type": "text", + "content": "REVISION" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 260 + ], + "type": "text", + "content": "and ConPreDiff 47 all develop training-based methods to provide additional conditioning for T2I models. SPRIGHT 3 introduces a spatially-focused large-scale dataset, by re-captioning 6 million images from existing vision datasets and demonstrate performance gains through an efficient training methodology. Test-time adaptations have also been proposed - (i) Layout Guidance 7 restricts specific objects to their bounding box location through the modification of cross-attention maps; however it relies on bounding box annotations, (ii) LayoutGPT 12 and LLM-grounded Diffusion 23 leverage large language models (LLMs) to generate layouts and bounding box co-ordinates and, (iii) RealCompo 54 combines multiple generative models for better spatial control. By developing an annotation-free cost-efficient framework we overcome the shortcomings of existing methods through REVISION." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 269, + 482, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 482, + 377 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 482, + 377 + ], + "type": "text", + "content": "Synthetic Images for Vision and Language. The flexibility and control provided during creation of synthetic images has led to various visuo-linguistic evaluation benchmarks using rendering tools. CLEVR [19] pioneered the utilization of synthetic objects in simulated scenes for visual compositionality reasoning. Many variants of CLEVR such as CLEVR-Hans [43], CLEVR-Hyp [41], Super-CLEVR [22], and CLEVRER [50] probe multiple facets of multimodal understanding with synthetic images and videos. PaintSkills introduced in DALL-EVAL [8] is an evaluation dataset that measures multiple aspects of a T2I model, which includes spatial reasoning, image-text alignment and social biases." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 387, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 495 + ], + "type": "text", + "content": "Evaluation of Multimodal LLMs. Multiple benchmarks have been proposed that evaluate reasoning capabilities of MLLMs. MMBench [27] evaluates models across 20 different dimensions, for a total of 2974 evaluation instances. The distinctive abilities of MLLMS to differentiate between coarse and fine-grained vision tasks is explored by MME [13] with images sourced from COCO. A limitation across all these benchmarks is that they collect instances from common VL datasets, increasing risk of data leakage and do not evaluate spatial relationships at scale. RevQA fills this gap by developing a diverse set of synthetic and scalable image-question pairs for a holistic evaluation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 510, + 320, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 320, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 320, + 523 + ], + "type": "text", + "content": "3 The REVISION Framework" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 592 + ], + "type": "text", + "content": "REVISION (Figure 2) is a rendering-based framework for generating spatially accurate images from an input prompt. Given a prompt, we generate an image in Blender, where the two object 3D models and the camera view are situated according to the spatial relationship derived from the prompt. The components of REVISION are described below." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": "The Asset Library includes a large human-inspected collection of 3D models of realistic objects with variations in texture and shape. Given an object name, the Asset Library randomly selects a matching asset rescaled to fit into a " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "1m" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": " cube to ensure that they are sufficiently visible in the final output. The Asset Library features 101 distinct classes of objects, 80 of which are from MS-COCO [24]. Each object class is associated with 3 to 5 royalty-free 3D model" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 246, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 246, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 246, + 102 + ], + "type": "text", + "content": "A. Chatterjee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 481, + 213 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 213 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 213 + ], + "type": "image", + "image_path": "8743d3490a8fe3a6398e6472f6027a9774fd35c087901d18ff7f0213ce7cb3e4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 222, + 482, + 257 + ], + "lines": [ + { + "bbox": [ + 130, + 222, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 222, + 482, + 257 + ], + "type": "text", + "content": "Fig. 2: REVISION parses a prompt into assets (objects) and the spatial relationship between them and synthesizes a symbolic image in Blender, placing the respective object assets at coordinates corresponding to the parsed spatial relationship." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 151, + 302, + 459, + 363 + ], + "blocks": [ + { + "bbox": [ + 130, + 270, + 482, + 293 + ], + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 293 + ], + "type": "text", + "content": "Table 1: Spatial relationships in REVISION and their rules for the Coordinate Generator. The objects are positioned from the camera's perspective." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 302, + 459, + 363 + ], + "lines": [ + { + "bbox": [ + 151, + 302, + 459, + 363 + ], + "spans": [ + { + "bbox": [ + 151, + 302, + 459, + 363 + ], + "type": "table", + "html": "
RelationSpatial PhrasesCoordinate Constraints Distance (m)
Horizontalto the left, to the rightX and Z are 0.[1, 1.5]
Verticalabove, below, top, bottomX and Y are 0.[0.75, 1]
Nearnear, next to, on the side ofZ is 0.[0.75, 1]
Depthin front of, behindZ is 0. Xobj1=-Xobj2.[1, 1.5]
", + "image_path": "42bb63515541bdbe746827e18809470f818c15965c5de4d3b1193f0242d1bfc1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 388, + 482, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 423 + ], + "type": "text", + "content": "assets from sketchfab.com, with a total of 410 3D models. REVISION includes 3 background panoramicas (Indoor, Outdoor, and White) from polyhaven.com and a corresponding textured floor asset from Sketchfab." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "content": "The Coordinate Generator deterministically generates 3D coordinates for the objects and the camera, given the names of the objects and the spatial relation extracted from the prompt. As shown in Table [1], REVISION supports four categories of spatial relationships between objects. In our coordinate frame, the X-, Y-, and Z-axis represent depth, horizontal, and vertical relationships respectively. To ensure that the objects are visible and the spatial relationship is obvious from the camera's view, the coordinate values for the objects on all three axes are confined within the range of " + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "inline_equation", + "content": "[-1m, 1m]" + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "content": ". The camera is placed at " + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "inline_equation", + "content": "x = 5m" + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "content": " with its view always facing the origin point. The camera is at " + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "inline_equation", + "content": "z = 2.5" + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "content": " for depth relationships and at " + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "inline_equation", + "content": "z = 1.5m" + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "content": " otherwise." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "The Scene Synthesizer assembles a 3D scene consisting of six main components: a camera, a light source, background, floor, and two objects. The two object assets and the camera are placed at their respective coordinates determined by the Coordinate Generator. Then the background asset, which is a 360-degree panorama image (modeled as a large sphere), is centered at the origin. The light source is added to a random position sufficiently higher than all objects in the scene. To prevent objects from appearing to float, the floor asset, a textured hyperplane orthogonal to the Z-axis, is positioned beneath the object asset with the lowest vertical coordinate. This floor placement also enables the object assets to cast shadows, enhancing the realism of the rendered image." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "REVISION" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 463, + 272 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 463, + 272 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 463, + 272 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 463, + 272 + ], + "type": "image", + "image_path": "d1ed60142e665e16aeea66e96de7dbc1df7a9db6dd8cc1cafa3e5a4d40924b1c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 279, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 131, + 279, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 131, + 279, + 482, + 304 + ], + "type": "text", + "content": "Fig. 3: Outputs from the REVISION rendering pipeline for 4 spatial relationships types for identical assets, with (bottom) and without a floor (top)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 325, + 483, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 483, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 483, + 422 + ], + "type": "text", + "content": "The Position Diversifier (Figure 3) ensures diversity in object orientations, background, and the camera angles every time REVISION is invoked. The background is rotated along the Z-axis, giving us a large number of static background options. In order to further diversify the perspective sizes and tilts of the object assets within the camera's view, we add random jitter to the position and orientation of the camera. We also add random small rotations to the objects along the Z-axis and vary the distance between the objects so that they are not always symmetric around the origin. See Supplementary Materials for more details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 441, + 422, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 441, + 422, + 455 + ], + "spans": [ + { + "bbox": [ + 131, + 441, + 422, + 455 + ], + "type": "text", + "content": "4 Improving Spatial Fidelity in T2I Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 465, + 410, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 465, + 410, + 478 + ], + "spans": [ + { + "bbox": [ + 131, + 465, + 410, + 478 + ], + "type": "text", + "content": "4.1 Training-Free Image Generation with REVISION" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 485, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 545 + ], + "type": "text", + "content": "Given an input prompt " + }, + { + "bbox": [ + 130, + 485, + 482, + 545 + ], + "type": "inline_equation", + "content": "(T)" + }, + { + "bbox": [ + 130, + 485, + 482, + 545 + ], + "type": "text", + "content": ", we first generate a spatially accurate reference image " + }, + { + "bbox": [ + 130, + 485, + 482, + 545 + ], + "type": "inline_equation", + "content": "(x^{(g)})" + }, + { + "bbox": [ + 130, + 485, + 482, + 545 + ], + "type": "text", + "content": " leveraging our REVISION pipeline. We then perform training-free image synthesis to generate an image " + }, + { + "bbox": [ + 130, + 485, + 482, + 545 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 485, + 482, + 545 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 130, + 485, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\phi(I|x^{(g)}, T)" + }, + { + "bbox": [ + 130, + 485, + 482, + 545 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 485, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 485, + 482, + 545 + ], + "type": "text", + "content": " is a T2I model. We reformulate the standard text-to-image pipeline into an image-to-image pipeline, conditioned by text, as shown in Figure 4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": "Standard diffusion methods such as Stable Diffusion (SD) generate an image by iteratively de-noising a Gaussian noise vector. Stochastic Differential Editing (SDEdit) [28], on the other hand, starts from a guide image " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "(x^{(g)}" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": ", in our case), adds Gaussian noise to it, and denoises it to produce the synthesized image " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": ". We use SDEdit within our Stable Diffusion pipeline and perform image generation guided by " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "x^{(g)}" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": ". We also explore the ControlNet [51] backbone, which allows fine-grained control over SD. Using ControlNet allows us to address two key points: a) our reference images provide enough spatial information even when low-level features are extracted from them and, b) we can mitigate any attribute-related biases present in the assets." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 246, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 246, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 246, + 102 + ], + "type": "text", + "content": "A. Chatterjee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 118, + 482, + 224 + ], + "blocks": [ + { + "bbox": [ + 142, + 118, + 482, + 224 + ], + "lines": [ + { + "bbox": [ + 142, + 118, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 142, + 118, + 482, + 224 + ], + "type": "image", + "image_path": "721ebdc641f63184247c42c6a9ff84b89a0befd9b1ca1f98e74ba3e1c2113c31.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 236, + 480, + 281 + ], + "lines": [ + { + "bbox": [ + 130, + 236, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 480, + 281 + ], + "type": "text", + "content": "Fig. 4: Given a user-provided input prompt " + }, + { + "bbox": [ + 130, + 236, + 480, + 281 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 236, + 480, + 281 + ], + "type": "text", + "content": ", we generate a corresponding synthetic image " + }, + { + "bbox": [ + 130, + 236, + 480, + 281 + ], + "type": "inline_equation", + "content": "x^{(g)}" + }, + { + "bbox": [ + 130, + 236, + 480, + 281 + ], + "type": "text", + "content": " using REVISION. With input prompt " + }, + { + "bbox": [ + 130, + 236, + 480, + 281 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 236, + 480, + 281 + ], + "type": "text", + "content": " and guidance " + }, + { + "bbox": [ + 130, + 236, + 480, + 281 + ], + "type": "inline_equation", + "content": "x^{(g)}" + }, + { + "bbox": [ + 130, + 236, + 480, + 281 + ], + "type": "text", + "content": ", we perform training-free image synthesis based on existing T2I pipelines such as Stable Diffusion or ControlNet to obtain a spatially accurate image." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 183, + 334, + 428, + 437 + ], + "blocks": [ + { + "bbox": [ + 130, + 290, + 482, + 324 + ], + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 324 + ], + "type": "text", + "content": "Table 2: The incorporation of REVISION as a guiding framework significantly enhances the spatial reasoning performance of Stable Diffusion (SD) models. Results highlighted in green represent scores achieved with images from REVISION." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 183, + 334, + 428, + 437 + ], + "lines": [ + { + "bbox": [ + 183, + 334, + 428, + 437 + ], + "spans": [ + { + "bbox": [ + 183, + 334, + 428, + 437 + ], + "type": "table", + "html": "
MethodOA (%)VISOR (%)
uncond cond1234
SD 1.429.8618.8162.9846.6020.116.891.63
+ REVISION53.9652.7197.6977.7961.0244.9027.15
SD 1.528.4317.5161.5944.2718.126.281.35
+ REVISION54.3353.0897.7278.0761.2745.4427.55
SD 2.147.8330.2563.2464.4235.7416.134.70
+ REVISION48.2647.1197.6176.0755.7537.1019.53
", + "image_path": "1d341b8b93ff4778660f453ee4647e163dfbd3bd9e5ea79d302ef71b0f6687f0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 459, + 261, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 459, + 261, + 471 + ], + "spans": [ + { + "bbox": [ + 131, + 459, + 261, + 471 + ], + "type": "text", + "content": "4.2 Experimental Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 475, + 480, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 480, + 536 + ], + "type": "text", + "content": "We study the efficacy of REVISION on two widely accepted benchmarks for spatial relationship, VISOR [15] and T2I-CompBench [18], which have 25,280 and 300 spatial prompts, respectively. For each evaluation prompt in the respective benchmarks, we generate a corresponding image from our REVISION pipeline and perform training-free image generation as described in Section 4.1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 536, + 480, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 480, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 480, + 596 + ], + "type": "text", + "content": "We leverage 3 variants of Stable Diffusion (SD), versions 1.4, 1.5, and 2.1 as our baseline models. For ControlNet, we use the canny edge-conditioned SD model. For holistic evaluations, we also report the Inception Score (IS) where applicable. For all subsequent tables, the bold values denote the best performance while underlined values indicate the second-best performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 611, + 266, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 611, + 266, + 624 + ], + "spans": [ + { + "bbox": [ + 131, + 611, + 266, + 624 + ], + "type": "text", + "content": "4.3 Results and Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 480, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 480, + 667 + ], + "type": "text", + "content": "Improvements over Baseline Models - We summarize our representative improvements over the baseline and existing methods, on the VISOR and T2I-CompBench benchmarks in Table 2 and Figure 5 respectively. The results in" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "REVISION" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 118, + 455, + 280 + ], + "blocks": [ + { + "bbox": [ + 149, + 118, + 455, + 280 + ], + "lines": [ + { + "bbox": [ + 149, + 118, + 455, + 280 + ], + "spans": [ + { + "bbox": [ + 149, + 118, + 455, + 280 + ], + "type": "image", + "image_path": "29a8b81e2ecb320f2bbcdde45c0019047b5208bc41d660e545834c9d079ddce0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 287, + 482, + 310 + ], + "lines": [ + { + "bbox": [ + 132, + 287, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 287, + 482, + 310 + ], + "type": "text", + "content": "Fig. 5: Comparing the T2I-CompBench spatial scores of REVISION-based guidance (green) with other leading T2I models and methods (blue)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 157, + 365, + 454, + 512 + ], + "blocks": [ + { + "bbox": [ + 132, + 332, + 482, + 355 + ], + "lines": [ + { + "bbox": [ + 132, + 332, + 482, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 482, + 355 + ], + "type": "text", + "content": "Table 3: Results on the VISOR Benchmark. With REVISION, we consistently outperform existing T2I methods on the VISOR benchmark." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 157, + 365, + 454, + 512 + ], + "lines": [ + { + "bbox": [ + 157, + 365, + 454, + 512 + ], + "spans": [ + { + "bbox": [ + 157, + 365, + 454, + 512 + ], + "type": "table", + "html": "
MethodOA (%)VISOR (%)
uncondcond1234
GLIDE 303.361.9859.066.721.020.170.03
DALLE-mini 1027.1016.1759.6738.3117.506.891.96
DALLE-v2 3563.9337.8959.2773.5947.2323.267.49
Layout Guidance 740.0138.8095.95----
Control-GPT 5348.3344.1765.9769.8051.2035.6720.48
Structured Diffusion 1128.6517.8762.3644.7018.736.571.46
Attend-and-Excite 442.0725.7561.2149.2919.334.560.08
SD 1.4 + REVISION53.9652.7197.6977.7961.0244.9027.15
SD 1.5 + REVISION54.3353.0897.7278.0761.2745.4427.55
SD 2.1 + REVISION48.2647.1197.6176.0755.7537.1019.53
ControlNet + REVISION56.8855.4897.5478.8262.9348.5831.59
", + "image_path": "8ed712fc26120e9cff7a6be685cbf64599bce2334730dc1232d80dc5e477bb08.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Table 2 are shown with reference images on a white background and # of denoising steps = 30. As shown in Table 2 we improve on all aspects of spatial relationships compared to our baseline methods. On SD 1.5, we achieve a " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "91.1\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " improvement in Object Accuracy (OA) and a " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "58.6\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " improvement on the conditional score. Specifically, we generate objects more accurately and achieve a high " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " of accuracy when spatially synthesizing them in the image. Interestingly, through REVISION, we increase the likelihood of consistently generating spatially correct images, as can be seen by the relatively high value of " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{VISOR}_4" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ". On VISOR (Table 3), REVISION enables baseline Stable Diffusion models to consistently outperform existing methods, across all aspects. Compared to the" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 102 + ], + "type": "text", + "content": "A. Chatterjee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 118, + 297, + 214 + ], + "blocks": [ + { + "bbox": [ + 138, + 118, + 297, + 214 + ], + "lines": [ + { + "bbox": [ + 138, + 118, + 297, + 214 + ], + "spans": [ + { + "bbox": [ + 138, + 118, + 297, + 214 + ], + "type": "image", + "image_path": "f909d1a2783da051162c7fde64c4d0e26df8ee412473a2bf53921cb0eef6d438.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 226, + 301, + 271 + ], + "lines": [ + { + "bbox": [ + 132, + 226, + 301, + 271 + ], + "spans": [ + { + "bbox": [ + 132, + 226, + 301, + 271 + ], + "type": "text", + "content": "Fig.6: REVISION improves T2I-CompBench Spatial Score (0 indicates missing objects, 1 denotes perfect object generation and spatial accuracy.)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 313, + 118, + 479, + 225 + ], + "blocks": [ + { + "bbox": [ + 313, + 118, + 479, + 225 + ], + "lines": [ + { + "bbox": [ + 313, + 118, + 479, + 225 + ], + "spans": [ + { + "bbox": [ + 313, + 118, + 479, + 225 + ], + "type": "image", + "image_path": "65024c20325d5f9998f015a6ac777ec08ffd30815e7db04fdea752e75cb325a4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 236, + 481, + 270 + ], + "lines": [ + { + "bbox": [ + 311, + 236, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 311, + 236, + 481, + 270 + ], + "type": "text", + "content": "Fig.7: Benchmarking the trade-off between spatial accuracy (VISOR) and Inception Score, achieved with REVISION." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 136, + 347, + 476, + 466 + ], + "blocks": [ + { + "bbox": [ + 132, + 293, + 482, + 338 + ], + "lines": [ + { + "bbox": [ + 132, + 293, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 482, + 338 + ], + "type": "text", + "content": "Table 4: VISORcond and Object Accuracy, split across relationship types. σVc and σ0A denote the respective metric's standard deviation w.r.t the relationships. Regardless of the spatial relation, REVISION enables T2I models to consistently produce spatially accurate images, a challenge faced by earlier approaches." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 347, + 476, + 466 + ], + "lines": [ + { + "bbox": [ + 136, + 347, + 476, + 466 + ], + "spans": [ + { + "bbox": [ + 136, + 347, + 476, + 466 + ], + "type": "table", + "html": "
MethodVISORcond (%)Object Accuracy (%)
leftrightabovebelowσVcleftrightabovebelowσ0A
GLIDE57.7861.7160.3256.242.463.103.463.493.390.18
DALLE-mini57.8960.1663.7556.143.2922.2921.7433.6230.745.99
DALLE-v256.4756.5160.9963.243.3864.3064.3265.6661.451.77
Control-GPT72.5070.2867.8565.702.9549.8048.2747.9746.951.18
SD 1.4 + REVISION97.5397.4598.0997.660.2952.4252.1156.9354.382.22
SD 1.5 + REVISION97.5797.5398.0597.700.2452.9952.5956.8054.921.94
SD 2.1 + REVISION97.8197.4697.9197.280.3046.7047.9449.7048.711.27
ControlNet + REVISION97.5197.2597.6597.720.2155.1055.1458.9858.292.05
", + "image_path": "049ba8032c667c0749e422148543ac3975d373647eb2ef9e5f7459677900e7e8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 479, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 479, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 479, + 523 + ], + "type": "text", + "content": "best open-source model, Control-GPT, we achieve a " + }, + { + "bbox": [ + 130, + 498, + 479, + 523 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 130, + 498, + 479, + 523 + ], + "type": "text", + "content": " improvement of " + }, + { + "bbox": [ + 130, + 498, + 479, + 523 + ], + "type": "inline_equation", + "content": "17.69\\%" + }, + { + "bbox": [ + 130, + 498, + 479, + 523 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 498, + 479, + 523 + ], + "type": "inline_equation", + "content": "48.12\\%" + }, + { + "bbox": [ + 130, + 498, + 479, + 523 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 498, + 479, + 523 + ], + "type": "inline_equation", + "content": "25.6\\%" + }, + { + "bbox": [ + 130, + 498, + 479, + 523 + ], + "type": "text", + "content": " on OA, VISORcond, and VISORuncond respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 528, + 482, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 482, + 589 + ], + "type": "text", + "content": "On T2I-CompBench (Figure 5), we observe similar improvement trends across diverse backgrounds, with baseline models guided by REVISION achieving consistent performance gains on the benchmark. In addition to enhancing spatial accuracy, REVISION improves prompt fidelity by ensuring that images contain all objects mentioned in the input prompt (Figure 6)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "Consistent Performance Across Relationship Types - Across all spatial relationship types, REVISION achieve a consistently high performance score across the VISOR metrics as shown in Table 4; a shortcoming prevalent in other methods. For example, the largest deviation in " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{VISOR}_{cond}" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " performance for ControlNet + REVISION is " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "0.21\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " between left and below relationships; in comparison Control-GPT deviates as much as " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "6.8\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " for the same." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "REVISION" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 248, + 173 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 248, + 173 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 248, + 173 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 248, + 173 + ], + "type": "image", + "image_path": "fd15fb30a3fd1352031689c16a05766b0f5d843d9624e9d8d6f922ba2d231e67.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 172, + 226, + 178 + ], + "lines": [ + { + "bbox": [ + 155, + 172, + 226, + 178 + ], + "spans": [ + { + "bbox": [ + 155, + 172, + 226, + 178 + ], + "type": "text", + "content": "An apple to the left of a book" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 178, + 247, + 236 + ], + "blocks": [ + { + "bbox": [ + 133, + 178, + 247, + 236 + ], + "lines": [ + { + "bbox": [ + 133, + 178, + 247, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 178, + 247, + 236 + ], + "type": "image", + "image_path": "2b9a9bd4fa3b80931dfc6fa651888bdd65b797e059bc8f00410dab0ef3129ff6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 133, + 244, + 247, + 303 + ], + "blocks": [ + { + "bbox": [ + 149, + 237, + 233, + 244 + ], + "lines": [ + { + "bbox": [ + 149, + 237, + 233, + 244 + ], + "spans": [ + { + "bbox": [ + 149, + 237, + 233, + 244 + ], + "type": "text", + "content": "A truck to the right of a fire hydrant" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 244, + 247, + 303 + ], + "lines": [ + { + "bbox": [ + 133, + 244, + 247, + 303 + ], + "spans": [ + { + "bbox": [ + 133, + 244, + 247, + 303 + ], + "type": "image", + "image_path": "c424fdc44c484ceb5164ef2b83d97b1335ef330f42a9aa4caace26541fd3b437.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 303, + 233, + 309 + ], + "lines": [ + { + "bbox": [ + 149, + 303, + 233, + 309 + ], + "spans": [ + { + "bbox": [ + 149, + 303, + 233, + 309 + ], + "type": "text", + "content": "A potted plant above a dining table" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 133, + 309, + 247, + 369 + ], + "blocks": [ + { + "bbox": [ + 133, + 309, + 247, + 369 + ], + "lines": [ + { + "bbox": [ + 133, + 309, + 247, + 369 + ], + "spans": [ + { + "bbox": [ + 133, + 309, + 247, + 369 + ], + "type": "image", + "image_path": "fde1ce4ac770e0449cb9a15a26bf1a14f487ffc730fd690b80e3c0b0d470607c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 368, + 220, + 374 + ], + "lines": [ + { + "bbox": [ + 160, + 368, + 220, + 374 + ], + "spans": [ + { + "bbox": [ + 160, + 368, + 220, + 374 + ], + "type": "text", + "content": "A car below an elephant" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 158, + 379, + 223, + 386 + ], + "lines": [ + { + "bbox": [ + 158, + 379, + 223, + 386 + ], + "spans": [ + { + "bbox": [ + 158, + 379, + 223, + 386 + ], + "type": "text", + "content": "Background Type = Indoor" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 399, + 482, + 433 + ], + "lines": [ + { + "bbox": [ + 131, + 399, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 131, + 399, + 482, + 433 + ], + "type": "text", + "content": "Fig. 8: Illustrative examples depicting the variation of generated images across the three variants of backgrounds in REVISION. For each pair, the image on the left is from REVISION and the image on the right is generated from the T2I model." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 249, + 114, + 364, + 173 + ], + "blocks": [ + { + "bbox": [ + 249, + 114, + 364, + 173 + ], + "lines": [ + { + "bbox": [ + 249, + 114, + 364, + 173 + ], + "spans": [ + { + "bbox": [ + 249, + 114, + 364, + 173 + ], + "type": "image", + "image_path": "eb49dec07cc0f367017790173f6813a6983e0dfdf00e1742decb61a41ebe1822.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 271, + 172, + 343, + 178 + ], + "lines": [ + { + "bbox": [ + 271, + 172, + 343, + 178 + ], + "spans": [ + { + "bbox": [ + 271, + 172, + 343, + 178 + ], + "type": "text", + "content": "A handbag to the left of a bird" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 249, + 178, + 364, + 236 + ], + "blocks": [ + { + "bbox": [ + 249, + 178, + 364, + 236 + ], + "lines": [ + { + "bbox": [ + 249, + 178, + 364, + 236 + ], + "spans": [ + { + "bbox": [ + 249, + 178, + 364, + 236 + ], + "type": "image", + "image_path": "6f9fea7c8caefe11abfcbc55cb149408210efd0065bb88e3cd5424eaadb82135.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 249, + 244, + 364, + 303 + ], + "blocks": [ + { + "bbox": [ + 262, + 237, + 351, + 244 + ], + "lines": [ + { + "bbox": [ + 262, + 237, + 351, + 244 + ], + "spans": [ + { + "bbox": [ + 262, + 237, + 351, + 244 + ], + "type": "text", + "content": "A potted plant to the right of an oven" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 249, + 244, + 364, + 303 + ], + "lines": [ + { + "bbox": [ + 249, + 244, + 364, + 303 + ], + "spans": [ + { + "bbox": [ + 249, + 244, + 364, + 303 + ], + "type": "image", + "image_path": "220d1e3c8b880b9ad5d670f4f3bad2fb0d3556add840a58a257f7b2ff9c711b2.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 274, + 303, + 339, + 309 + ], + "lines": [ + { + "bbox": [ + 274, + 303, + 339, + 309 + ], + "spans": [ + { + "bbox": [ + 274, + 303, + 339, + 309 + ], + "type": "text", + "content": "A vase above a teddy bear" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 249, + 309, + 364, + 368 + ], + "blocks": [ + { + "bbox": [ + 249, + 309, + 364, + 368 + ], + "lines": [ + { + "bbox": [ + 249, + 309, + 364, + 368 + ], + "spans": [ + { + "bbox": [ + 249, + 309, + 364, + 368 + ], + "type": "image", + "image_path": "fa7a20d80e579ea3268a2016e24afc0f7308cf133f3f304295531026f6f82f82.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 274, + 368, + 340, + 374 + ], + "lines": [ + { + "bbox": [ + 274, + 368, + 340, + 374 + ], + "spans": [ + { + "bbox": [ + 274, + 368, + 340, + 374 + ], + "type": "text", + "content": "An apple below an airplane" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 272, + 379, + 341, + 386 + ], + "lines": [ + { + "bbox": [ + 272, + 379, + 341, + 386 + ], + "spans": [ + { + "bbox": [ + 272, + 379, + 341, + 386 + ], + "type": "text", + "content": "Background Type " + }, + { + "bbox": [ + 272, + 379, + 341, + 386 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 272, + 379, + 341, + 386 + ], + "type": "text", + "content": " Outdoor" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 366, + 114, + 480, + 172 + ], + "blocks": [ + { + "bbox": [ + 366, + 114, + 480, + 172 + ], + "lines": [ + { + "bbox": [ + 366, + 114, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 366, + 114, + 480, + 172 + ], + "type": "image", + "image_path": "682b01f5b336b81caec330ec224a34751624a5b422129e437d2b6b57a93689a6.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 172, + 459, + 178 + ], + "lines": [ + { + "bbox": [ + 388, + 172, + 459, + 178 + ], + "spans": [ + { + "bbox": [ + 388, + 172, + 459, + 178 + ], + "type": "text", + "content": "An orange to the left of a sink" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 366, + 178, + 480, + 236 + ], + "blocks": [ + { + "bbox": [ + 366, + 178, + 480, + 236 + ], + "lines": [ + { + "bbox": [ + 366, + 178, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 366, + 178, + 480, + 236 + ], + "type": "image", + "image_path": "54abcc2ae2487e02aaae0b9041085300b6979d4ff4938017a3ff6818cc8996d3.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 366, + 244, + 480, + 300 + ], + "blocks": [ + { + "bbox": [ + 388, + 237, + 458, + 244 + ], + "lines": [ + { + "bbox": [ + 388, + 237, + 458, + 244 + ], + "spans": [ + { + "bbox": [ + 388, + 237, + 458, + 244 + ], + "type": "text", + "content": "A cow to the right of a couch" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 366, + 244, + 480, + 300 + ], + "lines": [ + { + "bbox": [ + 366, + 244, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 366, + 244, + 480, + 300 + ], + "type": "image", + "image_path": "65e38c054bbcab72befa979e67b0dfbc583d78f9d1d235dd71fcf34232662300.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 366, + 309, + 480, + 366 + ], + "blocks": [ + { + "bbox": [ + 396, + 303, + 449, + 309 + ], + "lines": [ + { + "bbox": [ + 396, + 303, + 449, + 309 + ], + "spans": [ + { + "bbox": [ + 396, + 303, + 449, + 309 + ], + "type": "text", + "content": "A bed above a sheep" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 366, + 309, + 480, + 366 + ], + "lines": [ + { + "bbox": [ + 366, + 309, + 480, + 366 + ], + "spans": [ + { + "bbox": [ + 366, + 309, + 480, + 366 + ], + "type": "image", + "image_path": "dc6a191b76edd8a97f826851aca78ddc8ad020e4c710b33f358e4f5c17bc2f25.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 367, + 456, + 374 + ], + "lines": [ + { + "bbox": [ + 389, + 367, + 456, + 374 + ], + "spans": [ + { + "bbox": [ + 389, + 367, + 456, + 374 + ], + "type": "text", + "content": "A bottle below a wine glass" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 379, + 454, + 386 + ], + "lines": [ + { + "bbox": [ + 391, + 379, + 454, + 386 + ], + "spans": [ + { + "bbox": [ + 391, + 379, + 454, + 386 + ], + "type": "text", + "content": "Background Type = White" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "bbox": [ + 132, + 457, + 246, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 246, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 246, + 468 + ], + "type": "text", + "content": "4.4 Ablation Studies" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 130, + 481, + 482, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 482, + 590 + ], + "type": "text", + "content": "Impact of Background - In Table 5 we enumerate the impact of the background types in the images from the REVISION pipeline and the downstream trade-off between VISOR performance and model diversity. Utilizing white backgrounds that exclusively feature the two objects in question minimizes potential distractions for the model. Conversely, when the model is presented with initial reference images incorporating indoor or outdoor backgrounds, it exhibits the capacity to identify and leverage distractor objects, resulting in the generation of diverse images. As shown in Figure 8 all generated images maintain spatial accuracy, but noisier reference images result in greater diversity." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "Controllability vs Photo-Realism - In this setup, we study the impact of the # of denoising steps and its trade-off with photo-realism. As shown in Figure 7 that while the performance on VISOR deteriorates with additional # of denoising steps, it improves the model's ability to be more diverse and photorealistic. In Figure 9, we demonstrate that by utilizing REVISION, baseline" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 246, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 246, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 246, + 102 + ], + "type": "text", + "content": "A. Chatterjee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 153, + 157, + 459, + 290 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 5: The impact of the 3 background types in the REVISION pipeline on the VISOR benchmark. While best performance is achieved with a white background, diverse outputs are attained with the outdoor background type." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 153, + 157, + 459, + 290 + ], + "lines": [ + { + "bbox": [ + 153, + 157, + 459, + 290 + ], + "spans": [ + { + "bbox": [ + 153, + 157, + 459, + 290 + ], + "type": "table", + "html": "
Model Background IS (↑) OA (%)VISOR (%)
uncondcond1234
SD 1.4White16.1653.9652.7197.6977.7961.0244.927.15
Indoor19.1148.5345.1292.9774.8253.7934.7817.09
Outdoor20.1644.3241.8094.3169.7949.3831.8616.17
SD 1.5White16.2754.3353.0897.7278.0761.2745.4427.55
Indoor19.1148.7745.2892.8574.9353.9634.7717.47
Outdoor19.6643.9941.5194.3669.4848.5831.4616.52
SD 2.1White12.7948.2647.1197.6176.0755.7537.1019.53
Indoor11.5231.0829.3794.5059.8033.9617.406.34
Outdoor10.5136.3734.6795.3465.0541.2323.059.36
", + "image_path": "291d3e429b17ca1673d2cee691d65ff10730c53efeeaaba367beddf452a8a007.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 303, + 478, + 446 + ], + "blocks": [ + { + "bbox": [ + 133, + 303, + 478, + 446 + ], + "lines": [ + { + "bbox": [ + 133, + 303, + 478, + 446 + ], + "spans": [ + { + "bbox": [ + 133, + 303, + 478, + 446 + ], + "type": "image", + "image_path": "b497e0b5395cedfff65b2586f1a911b345fdb59637f2c5e3b8a984328c3c93d1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 453, + 482, + 475 + ], + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 475 + ], + "type": "text", + "content": "Fig. 9: Illustrative examples showing the trade-off between photo-realism and denoising steps, while maintaining generating spatially accurate images using REVISION." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 501, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 482, + 525 + ], + "type": "text", + "content": "models can preserve their spatial coherency while iteratively demonstrating a higher degree of photo-realism, through more # of denoising steps." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 546, + 377, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 546, + 377, + 559 + ], + "spans": [ + { + "bbox": [ + 131, + 546, + 377, + 559 + ], + "type": "text", + "content": "4.5 Extending VISOR for Depth Relationships" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "We further extend the VISOR benchmark for Depth relationships (in front of/behind). We utilize Depth Anything [46] for generation of depth maps and OWLv2 [29] for object detection. Given a T2I generated image " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " and its prompt " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " that contain two objects " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "o_1, o_2" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ", we obtain its depth map " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "I_D" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " using Depth Anything. We then retrieve the centroids detected for the two objects " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "c_{o_1}, c_{o_2}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " using OWLv2. At these centroid coordinates, we acquire the depth values for the two objects from the depth map " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "I_D(c_{o_1}), I_D(c_{o_2})" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ". We check if the acquired depth values match the spatial relationship in the prompt, and evaluate similar" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "text", + "content": "REVISION" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 160, + 146, + 451, + 232 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 6: Comparing baseline methods against REVISION-guided image synthesis on depth relationships. DS denotes the # of Denoising Steps." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 160, + 146, + 451, + 232 + ], + "lines": [ + { + "bbox": [ + 160, + 146, + 451, + 232 + ], + "spans": [ + { + "bbox": [ + 160, + 146, + 451, + 232 + ], + "type": "table", + "html": "
SettingOA (%)VISOR (%)
uncond1234
SD 1.5 (baseline)41.5227.1560.1231.8213.193.47
+ REVISION (DS=30)58.3229.6264.7434.5314.994.22
+ REVISION (DS=35)52.0532.0868.1137.9217.514.78
+ REVISION (DS=40)47.4330.4665.5035.9915.744.64
", + "image_path": "427d14322bf539715816c1ca55ffd8c7fa2921ef22c1abf00240f014027606c9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 254, + 479, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 254, + 479, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 479, + 278 + ], + "type": "text", + "content": "to VISOR. As shown in Table 6, REVISION improves VISOR scores across all metrics and across multiple denoising steps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 300, + 261, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 261, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 261, + 310 + ], + "type": "text", + "content": "4.6 Human Evaluations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 322, + 480, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 480, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 480, + 382 + ], + "type": "text", + "content": "To verify the generalizability of REVISION-based guidance on T2I models, we perform 2 distinct experiments and conduct human evaluations for validation. For each experiment, we independently sample 200 generated images and take the average scores across 4 workers. We also report unanimous (100%) and majority (75%) agreements between the workers for each experiment." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 397, + 482, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 482, + 482 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 482, + 482 + ], + "type": "text", + "content": "Prompts of Multiple Objects and Relationships - In this experiment, we generate reference images using prompts that include 2 spatial phrases and 3 objects, and use these images to guide T2I generation. Each generated image is evaluated for accuracy based on the input spatial prompts. We achieve an accuracy of " + }, + { + "bbox": [ + 130, + 397, + 482, + 482 + ], + "type": "inline_equation", + "content": "79.62\\%" + }, + { + "bbox": [ + 130, + 397, + 482, + 482 + ], + "type": "text", + "content": " when at least 1 phrase is correctly represented in the image and " + }, + { + "bbox": [ + 130, + 397, + 482, + 482 + ], + "type": "inline_equation", + "content": "46.5\\%" + }, + { + "bbox": [ + 130, + 397, + 482, + 482 + ], + "type": "text", + "content": " when both phrases are correctly represented. The unanimous and majority agreements among evaluators are " + }, + { + "bbox": [ + 130, + 397, + 482, + 482 + ], + "type": "inline_equation", + "content": "64.5\\%" + }, + { + "bbox": [ + 130, + 397, + 482, + 482 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 397, + 482, + 482 + ], + "type": "inline_equation", + "content": "86.5\\%" + }, + { + "bbox": [ + 130, + 397, + 482, + 482 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 495, + 482, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 580 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 580 + ], + "type": "text", + "content": "Out-of-Distribution Objects - We consider prompts containing exactly one object not found in the REVISION Asset Library. Given a prompt that mentions an OOD object, we find the semantically closest object (list in Supplementary Material) in our library and use their corresponding image as guidance. For example, we generate an image of \"a helicopter above a bicycle\" by providing a reference image of \"an airplane above a bicycle\". An accuracy of " + }, + { + "bbox": [ + 130, + 495, + 482, + 580 + ], + "type": "inline_equation", + "content": "63.62\\%" + }, + { + "bbox": [ + 130, + 495, + 482, + 580 + ], + "type": "text", + "content": " is found with an unanimous and majority agreement of " + }, + { + "bbox": [ + 130, + 495, + 482, + 580 + ], + "type": "inline_equation", + "content": "67\\%" + }, + { + "bbox": [ + 130, + 495, + 482, + 580 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 495, + 482, + 580 + ], + "type": "inline_equation", + "content": "90.5\\%" + }, + { + "bbox": [ + 130, + 495, + 482, + 580 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 601, + 471, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 471, + 615 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 471, + 615 + ], + "type": "text", + "content": "5 RevQA: A Spatial Reasoning Benchmark for MLLMs" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "We leverage the determinism of the REVISION pipeline to construct a new visual question answering benchmark (RevQA) for evaluating the spatial reasoning abilities of multimodal large language models." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 246, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 246, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 246, + 102 + ], + "type": "text", + "content": "A. Chatterjee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 256, + 238 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 256, + 238 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 256, + 238 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 256, + 238 + ], + "type": "image", + "image_path": "c65468606980fc568f6fc02eda80285fdf7e395438d9b8b5474353829f684fa7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 249, + 482, + 293 + ], + "lines": [ + { + "bbox": [ + 130, + 249, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 482, + 293 + ], + "type": "text", + "content": "Fig. 10: The RevQA Benchmark. Using the REVISION pipeline, we generate spatially accurate images and formulate 16 question types from a given caption. We leverage these generated questions and image, benchmarking Multimodal Large Language Models in their abilities to reason over spatial relationships." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 256, + 114, + 480, + 239 + ], + "blocks": [ + { + "bbox": [ + 256, + 114, + 480, + 239 + ], + "lines": [ + { + "bbox": [ + 256, + 114, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 256, + 114, + 480, + 239 + ], + "type": "table", + "html": "
Question TypeQuestionAnswer
Simple SpatialIs there an airplane above a chair?Yes
Opposite SpatialIs there a chair above an airplane?No
ANDIs there a chair and an airplane?Yes
ORIs there a chair or an airplane?Yes
NOTIs there an airplane not above a chair?No
Double NegativeIs there an airplane not below a chair?Yes
Random ANDIs there an airplane and a hot dog?No
Random ORIs there an airplane or a hot dog?Yes
Random SpatialIs there an airplane above a hot dog?No
Random Combined ANDIs there an airplane above a chair and is there an airplane above a hot dog?No
Random Combined ORIs there an airplane above a hot dog or is there an airplane above a chair?Yes
Adversarial ANDIs there a helicopter and a chair?No
Adversarial ORIs there a helicopter or a chair?Yes
Adversarial SpatialIs there a helicopter above a chair?No
Adversarial Combined ANDIs there a helicopter above a chair and is there an airplane above a chair?No
Adversarial Combined ORIs there a helicopter above a chair or is there an airplane above a chair?Yes
", + "image_path": "da2b7b10755c453e14d43409898aa44301a3519e98a9043bd7dde6ca01431564.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 318, + 482, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 482, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 482, + 378 + ], + "type": "text", + "content": "Question Generation. The benchmark contains 16 types of yes-no questions for a REVISION-generated image, consisting of negations, conjunctions, and disjunctions, building on prior work on logic-based visual question answering [14]. Each question type evaluates a combination of spatial and logical reasoning abilities in multimodal large language models (MLLMs) (Figure 10)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 378, + 483, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 483, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 483, + 499 + ], + "type": "text", + "content": "Among the 16 types, we incorporate Random and Adversarial types of questions to further evaluate the robustness and reliability of MLLMs using simple templated transformations. In Random types of questions, we replace an object (visible in the image) in the question with another randomly picked object from REVISION's Asset Library. For the Adversarial set of questions, we replace one of the objects with another that is semantically and visually close. In addition to benchmarking their robustness, these questions allow simultaneous evaluation of the fine-grained spatial perception and reasoning abilities of these models. To alleviate any order bias in instances which contain multiple questions (see Combined in Figure 10), we randomly switch the order between them." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "text", + "content": "Evaluation Setup and Results We sample 50k image-question pairs and benchmark 5 open-source state-of-the-art MLLMs - LLaVA 1.5 [25], Fuyu-8B [2], InstructBLIP [9], LLaMA-Adapter 2.1 [52] and Qwen-VL-Chat [1]. We instruct all models to generate binary responses and set the temperature " + }, + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "inline_equation", + "content": "= 0" + }, + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "text", + "content": ", to remove stochasticity in the generated responses." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "We present our evaluation results in Table 7 and find that all models have a large gap in performance in reasoning over spatial relationships. While most models reason well over simple spatial relationships, they have a large performance drop when presented with the opposite spatial relationships. For example, LLaVA-1.5, the best performing model, has a " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "58.17\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " decrease in performance when probed with simple vs opposite spatial questions. This can be attributed to: " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "(a)" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " insufficient training data for rare object relationships, such as less instances of an \"elephant above a person\" than vice versa; " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "b)" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " the inability of vision" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "REVISION" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 168, + 481, + 350 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 159 + ], + "type": "text", + "content": "Table 7: Performances of 5 MLLMs across the 16 types of questions in RevQA. Most models perform worse than random (50%) when reasoning over Opposite Spatial relationships and Double Negative questions. All models have a significant drop in performance with Random/Adversarial questions, in comparison to their simpler versions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 168, + 481, + 350 + ], + "lines": [ + { + "bbox": [ + 134, + 168, + 481, + 350 + ], + "spans": [ + { + "bbox": [ + 134, + 168, + 481, + 350 + ], + "type": "table", + "html": "
Question TypeLLaVa 1.5Fuyu-8BInstructBLIPLLaMA-Adapter 2.1Qwen-VL Chat
Simple Spatial0.9420.7020.8340.5790.940
Opposite Spatial0.3940.2870.1840.4190.402
AND0.9350.8870.9570.8580.889
OR0.9950.3960.5980.7220.949
NOT0.9260.6190.3560.5040.583
Double Negative0.2670.3470.6650.4900.212
Random AND0.9340.3080.6750.6160.978
Random OR0.9250.1780.1940.1940.324
Random Spatial0.9250.3700.6860.7900.919
Random Combined AND0.1160.5020.6270.8000.567
Random Combined OR0.9680.5360.4140.0030.506
Adversarial AND0.6610.1840.5420.6410.789
Adversarial OR0.9210.1880.4430.1560.685
Adversarial Spatial0.5590.3350.7770.8930.615
Adversarial Combined AND0.1320.5390.6950.8050.695
Adversarial Combined OR0.9530.4560.3860.0030.254
Average0.7200.4460.5980.5780.642
", + "image_path": "131c6bb1a60b9e98cc248132183dfbed43e29a840be60bc4289acf19b3dda005.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 372, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 372, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 372, + 482, + 517 + ], + "type": "text", + "content": "encoders like CLIP to capture subtle semantic differences. MLLMs also struggle with negation, possibly because image captions do not capture enough negations; e.g. COCO Captions only contain " + }, + { + "bbox": [ + 130, + 372, + 482, + 517 + ], + "type": "inline_equation", + "content": "0.97\\%" + }, + { + "bbox": [ + 130, + 372, + 482, + 517 + ], + "type": "text", + "content": " occurrences of 'not'. All models significantly suffer when presented with questions that consist of double negatives, which evaluate the models' ability to reason of negations and spatial relationships in tandem. Furthermore, all models suffer under adversarial settings in comparison to their simpler counterparts; comparing LLaVA's performance for AND and Adversarial Combined AND questions, we find a " + }, + { + "bbox": [ + 130, + 372, + 482, + 517 + ], + "type": "inline_equation", + "content": "85.88\\%" + }, + { + "bbox": [ + 130, + 372, + 482, + 517 + ], + "type": "inline_equation", + "content": "(0.935\\rightarrow 0.132)" + }, + { + "bbox": [ + 130, + 372, + 482, + 517 + ], + "type": "text", + "content": " drop in performance. We also observe a larger decline in performance for Adversarial questions than for the Random set of questions hinting that while models independently perform well at object recognition and simple spatial relationships, combining them adversarily significantly reduces performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 533, + 220, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 220, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 220, + 547 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "In this work, we introduce REVISION, a framework designed for training-free enhancement of spatial relationships in Text-to-Image models and RevQA, a benchmark to evaluate the spatial reasoning abilities of multimodal large language models. Our results demonstrate the effectiveness of leveraging 3D rendering pipelines as a cost-efficient approach for developing generative models with robust reasoning capabilities. REVISION is modular and can easily be extended to incorporate additional features, assets, and relationships. We hope our method inspires future research at the intersection of computer graphics and generative AI, enabling safe deployment of these systems in the real world." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 246, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 246, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 246, + 102 + ], + "type": "text", + "content": "A. Chatterjee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 234 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 234 + ], + "type": "text", + "content": "The authors acknowledge resources provided by Research Computing at Arizona State University. The authors also acknowledge technical access and support from ASU Enterprise Technology. This work was supported by NSF Robust Intelligence program grants #1750082 and #2132724. TG was supported by Microsoft's Accelerating Foundation Model Research (AFMR) program and UMBC's Strategic Award for Research Transitions (START). The views and opinions of the authors expressed herein do not necessarily state or reflect those of the funding agencies and employers." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 251, + 197, + 263 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 251, + 197, + 263 + ], + "spans": [ + { + "bbox": [ + 133, + 251, + 197, + 263 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 274, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 274, + 481, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 274, + 481, + 307 + ], + "spans": [ + { + "bbox": [ + 138, + 274, + 481, + 307 + ], + "type": "text", + "content": "1. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond (2023), https://arxiv.org/abs/2308.12966" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 308, + 482, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 308, + 482, + 329 + ], + "spans": [ + { + "bbox": [ + 138, + 308, + 482, + 329 + ], + "type": "text", + "content": "2. Bavishi, R., Elsen, E., Hawthorne, C., Nye, M., Odena, A., Somani, A., Taşırlar, S.: Introducing our multimodal models (2023), https://www.adept.ai/blog/fuyu-8b" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 329, + 481, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 329, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 138, + 329, + 481, + 361 + ], + "type": "text", + "content": "3. Chatterjee, A., Stan, G.B.M., Aflalo, E., Paul, S., Ghosh, D., Gokhale, T., Schmidt, L., Hajishirzi, H., Lal, V., Baral, C., Yang, Y.: Getting it right: Improving spatial consistency in text-to-image models (2024), https://arxiv.org/abs/2404.01197" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 361, + 481, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 361, + 481, + 394 + ], + "spans": [ + { + "bbox": [ + 138, + 361, + 481, + 394 + ], + "type": "text", + "content": "4. Chefer, H., Alaluf, Y., Vinker, Y., Wolf, L., Cohen-Or, D.: Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG) 42(4), 1-10 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 394, + 481, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 394, + 481, + 437 + ], + "spans": [ + { + "bbox": [ + 138, + 394, + 481, + 437 + ], + "type": "text", + "content": "5. Chen, J., Jincheng, Y., Chongjian, G., Yao, L., Xie, E., Wang, Z., Kwok, J., Luo, P., Lu, H., Li, Z.: Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In: The Twelfth International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 437, + 481, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 437, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 437, + 481, + 491 + ], + "type": "text", + "content": "6. Chen, M., Radford, A., Child, R., Wu, J., Jun, H., Luan, D., Sutskever, I.: Generative pretraining from pixels. In: Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event. Proceedings of Machine Learning Research, vol. 119, pp. 1691-1703. PMLR (2020), http://proceedings.mlr.press/v119/chen20s.html" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 492, + 481, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 492, + 481, + 513 + ], + "spans": [ + { + "bbox": [ + 138, + 492, + 481, + 513 + ], + "type": "text", + "content": "7. Chen, M., Laina, I., Vedaldi, A.: Training-free layout control with cross-attention guidance (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 514, + 481, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 481, + 546 + ], + "type": "text", + "content": "8. Cho, J., Zala, A., Bansal, M.: Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3043-3054 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 546, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 546, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 546, + 481, + 589 + ], + "type": "text", + "content": "9. Dai, W., Li, J., Li, D., Tiong, A., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=vvoWPYqZJA" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 590, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 590, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 134, + 590, + 480, + 622 + ], + "type": "text", + "content": "10. Dayma, B., Patil, S., Cuenca, P., Saifullah, K., Abraham, T., Le Khac, P., Melas, L., Ghosh, R.: Dall-e mini (2021). https://doi.org/10.5281/zenodo.5146400. https://github.com/borisdayma/dalle-mini" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 622, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 622, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 622, + 481, + 666 + ], + "type": "text", + "content": "11. Feng, W., He, X., Fu, T.J., Jampani, V., Akula, A.R., Narayana, P., Basu, S., Wang, X.E., Wang, W.Y.: Training-free structured diffusion guidance for compositional text-to-image synthesis. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=PUIqjT4rzq7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "REVISION" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 150 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 150 + ], + "type": "text", + "content": "12. Feng, W., Zhu, W., Fu, T.j., Jampani, V., Akula, A., He, X., Basu, S., Wang, X.E., Wang, W.Y.: Layoutgpt: Compositional visual planning and generation with large language models. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 195 + ], + "type": "text", + "content": "13. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Yang, J., Zheng, X., Li, K., Sun, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. ArXiv preprint abs/2306.13394 (2023), https://arxiv.org/abs/2306.13394" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 195, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 195, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 481, + 228 + ], + "type": "text", + "content": "14. Gokhale, T., Banerjee, P., Baral, C., Yang, Y.: Vqa-lol: Visual question answering under the lens of logic. In: European conference on computer vision. pp. 379-396. Springer (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 228, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 481, + 262 + ], + "type": "text", + "content": "15. Gokhale, T., Palangi, H., Nushi, B., Vineet, V., Horvitz, E., Kamar, E., Baral, C., Yang, Y.: Benchmarking spatial relationships in text-to-image generation. arXiv preprint arXiv:2212.10015 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "type": "text", + "content": "16. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139–144 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 296, + 481, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 296, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 296, + 481, + 341 + ], + "type": "text", + "content": "17. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-or, D.: Prompt-to-prompt image editing with cross-attention control. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=CDixzkzeyb" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 341, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 481, + 373 + ], + "type": "text", + "content": "18. Huang, K., Sun, K., Xie, E., Li, Z., Liu, X.: T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. Advances in Neural Information Processing Systems 36, 78723-78747 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 374, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 441 + ], + "type": "text", + "content": "19. Johnson, J., Hariharan, B., van der Maaten, L., Fei-Fei, L., Zitnick, C.L., Girshick, R.B.: CLEVR: A diagnostic dataset for compositional language and elementary visual reasoning. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017. pp. 1988-1997. IEEE Computer Society (2017). https://doi.org/10.1109/CVPR.2017.215. https://doi.org/10.1109/CVPR.2017.215" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 441, + 481, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 481, + 485 + ], + "type": "text", + "content": "20. Kamath, A., Hessel, J., Chang, K.W.: What's \"up\" with vision-language models? investigating their struggle with spatial reasoning. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. pp. 9161-9175 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 486, + 481, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 481, + 518 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 481, + 518 + ], + "type": "text", + "content": "21. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22511-22521 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 519, + 481, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 519, + 481, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 519, + 481, + 563 + ], + "type": "text", + "content": "22. Li, Z., Wang, X., Stengel-Eskin, E., Kortylewski, A., Ma, W., Van Durme, B., Yuille, A.L.: Super-clevr: A virtual benchmark to diagnose domain robustness in visual reasoning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14963-14973 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 564, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 564, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 564, + 481, + 597 + ], + "type": "text", + "content": "23. Lian, L., Li, B., Yala, A., Darrell, T.: Llm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. ArXiv preprint abs/2305.13655 (2023), https://arxiv.org/abs/2305.13655" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 598, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 481, + 643 + ], + "type": "text", + "content": "24. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "25. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "type": "text", + "content": "A. Chatterjee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "text", + "content": "26. Liu, N., Li, S., Du, Y., Torralba, A., Tenenbaum, J.B.: Compositional visual generation with composable diffusion models. In: European Conference on Computer Vision. pp. 423-439. Springer (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "text", + "content": "27. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 237 + ], + "type": "text", + "content": "28. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. In: The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net (2022), https://openreview.net/forum?id=aBsCjcPu_tE" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 238, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 282 + ], + "type": "text", + "content": "29. Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Dosovitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., et al.: Simple open-vocabulary object detection. In: European Conference on Computer Vision. pp. 728-755. Springer (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 282, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 358 + ], + "type": "text", + "content": "30. Nichol, A.Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: GLIDE: towards photorealistic image generation and editing with text-guided diffusion models. In: Chaudhuri, K., Jegelka, S., Song, L., Szepesvári, C., Niu, G., Sabato, S. (eds.) International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA. Proceedings of Machine Learning Research, vol. 162, pp. 16784-16804. PMLR (2022), https://proceedings.mlr.press/v162/nichol22a.html" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "text", + "content": "31. Parmar, G., Kumar Singh, K., Zhang, R., Li, Y., Lu, J., Zhu, J.Y.: Zero-shot image-to-image translation. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-11 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 392, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 424 + ], + "type": "text", + "content": "32. Patel, M., Kim, C., Cheng, S., Baral, C., Yang, Y.: Eclipse: A resource-efficient text-to-image prior for image generations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9069-9078 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 424, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 468 + ], + "type": "text", + "content": "33. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: SDXL: Improving latent diffusion models for high-resolution image synthesis. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=di52zR8xgf" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 468, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 544 + ], + "type": "text", + "content": "34. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event. Proceedings of Machine Learning Research, vol. 139, pp. 8748-8763. PMLR (2021), http://proceedings.mlr.press/v139/radford21a.html" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 544, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 481, + 578 + ], + "type": "text", + "content": "35. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 578, + 481, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 642 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 642 + ], + "type": "text", + "content": "36. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event. Proceedings of Machine Learning Research, vol. 139, pp. 8821-8831. PMLR (2021), http://proceedings.mlr.press/v139/ramesh21a.html" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "37. Razzhigaev, A., Shakhmatov, A., Maltseva, A., Arkhipkin, V., Pavlov, I., Ryabov, I., Kuts, A., Panchenko, A., Kuznetsov, A., Dimitrov, D.: Kandinsky: An improved" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "REVISION" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 147, + 116, + 481, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 481, + 150 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 481, + 150 + ], + "type": "text", + "content": "text-to-image synthesis with image prior and latent diffusion. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations. pp. 286-295 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 152, + 481, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 152, + 481, + 185 + ], + "spans": [ + { + "bbox": [ + 133, + 152, + 481, + 185 + ], + "type": "text", + "content": "38. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 186, + 481, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 186, + 481, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 186, + 481, + 232 + ], + "type": "text", + "content": "39. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 233, + 481, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 233, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 233, + 481, + 310 + ], + "type": "text", + "content": "40. Salimans, T., Goodfellow, I.J., Zaremba, W., Cheung, V., Radford, A., Chen, X.: Improved techniques for training gans. In: Lee, D.D., Sugiyama, M., von Luxburg, U., Guyon, I., Garnett, R. (eds.) Advances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, December 5-10, 2016, Barcelona, Spain. pp. 2226-2234 (2016), https://proceedings.neurips.cc/paper/2016/bitstream/8a3363abe792db2d8761d6403605aab7-Abstract.html" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 312, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 481, + 390 + ], + "type": "text", + "content": "41. Sampat, S.K., Kumar, A., Yang, Y., Baral, C.: CLEVR_HYP: A challenge dataset and baselines for visual question answering with hypothetical actions over images. In: Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. pp. 3692-3709. Association for Computational Linguistics, Online (2021). https://doi.org/10.18653/v1/2021.naacl-main.289. https://aclanthology.org/2021.naacl-main.289" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 392, + 481, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 469 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 469 + ], + "type": "text", + "content": "42. Sohn, K., Lee, H., Yan, X.: Learning structured output representation using deep conditional generative models. In: Cortes, C., Lawrence, N.D., Lee, D.D., Sugiyama, M., Garnett, R. (eds.) Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada. pp. 3483-3491 (2015), https://proceedings.neurips.cc/paper/2015/bitstream/8d55a249e6baa5c06772297520da2051-Abstract.html" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 471, + 481, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 481, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 481, + 548 + ], + "type": "text", + "content": "43. Stammer, W., Schramowski, P., Kersting, K.: Right for the right concept: Revising neuro-symbolic concepts by interacting with their explanations. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 3619-3629. Computer Vision Foundation / IEEE (2021). https://doi.org/10.1109/CVPR46437.2021.00362, https://openaccess.thecvf.com/content/CVPR2021/html/Stammer_Right_for_the_Right_Concept_Revising_Neuro-Symbolic_Concepts_by_Interacting_CVPR_2021_paper.html" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 550, + 481, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 550, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 550, + 481, + 594 + ], + "type": "text", + "content": "44. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. ArXiv preprint abs/2312.11805 (2023), https://arxiv.org/abs/2312.11805" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 596, + 481, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 596, + 481, + 630 + ], + "spans": [ + { + "bbox": [ + 132, + 596, + 481, + 630 + ], + "type": "text", + "content": "45. Wake, N., Kanehira, A., Sasabuchi, K., Takamatsu, J., Ikeuchi, K.: Gpt-4v (ision) for robotics: Multimodal task planning from human demonstration. arXiv preprint arXiv:2311.12015 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "46. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. ArXiv preprint abs/2401.10891 (2024), https://arxiv.org/abs/2401.10891" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "type": "text", + "content": "A. Chatterjee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "47. Yang, L., Liu, J., Hong, S., Zhang, Z., Huang, Z., Cai, Z., Zhang, W., Cui, B.: Improving diffusion-based image synthesis with context prediction. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "type": "text", + "content": "48. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of lmms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421 9(1), 1 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 226 + ], + "type": "text", + "content": "49. Yang, Z., Wang, J., Gan, Z., Li, L., Lin, K., Wu, C., Duan, N., Liu, Z., Liu, C., Zeng, M., et al.: Reco: Region-controlled text-to-image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14246-14255 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 226, + 482, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 280 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 280 + ], + "type": "text", + "content": "50. Yi, K., Gan, C., Li, Y., Kohli, P., Wu, J., Torralba, A., Tenenbaum, J.B.: CLEVRER: collision events for video representation and reasoning. In: 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net (2020), https://openreview.net/forum?id=HkxYzANYDB" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 280, + 482, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 280, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 482, + 314 + ], + "type": "text", + "content": "51. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 314, + 482, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 346 + ], + "type": "text", + "content": "52. Zhang, R., Han, J., Liu, C., Gao, P., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 346, + 482, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 369 + ], + "type": "text", + "content": "53. Zhang, T., Zhang, Y., Vineet, V., Joshi, N., Wang, X.: Controllable text-to-image generation with gpt-4. arXiv preprint arXiv:2305.18583 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 369, + 482, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 403 + ], + "type": "text", + "content": "54. Zhang, X., Yang, L., Cai, Y., Yu, Z., Xie, J., Tian, Y., Xu, M., Tang, Y., Yang, Y., Cui, B.: Realcompo: Dynamic equilibrium between realism and compositionality improves text-to-image diffusion models. arXiv preprint arXiv:2402.12908 (2024)" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "REVISION" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/RGBD GS-ICP SLAM/794618d9-8655-4def-b2e9-3b1608333e79_content_list.json b/2024/RGBD GS-ICP SLAM/794618d9-8655-4def-b2e9-3b1608333e79_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0197d9afb5ed2ce27de544bb928c8cfd83644028 --- /dev/null +++ b/2024/RGBD GS-ICP SLAM/794618d9-8655-4def-b2e9-3b1608333e79_content_list.json @@ -0,0 +1,1462 @@ +[ + { + "type": "text", + "text": "RGBD GS-ICP SLAM", + "text_level": 1, + "bbox": [ + 387, + 142, + 614, + 160 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Seongbo Ha, Jiung Yeon, and Hyeonwoo Yu", + "bbox": [ + 338, + 191, + 663, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sungkyunkwan University, Suwon, South Korea \n{sobo3607, wcr12st, hwyu}@skku.edu", + "bbox": [ + 357, + 218, + 643, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Simultaneous Localization and Mapping (SLAM) with dense representation plays a key role in robotics, Virtual Reality (VR), and Augmented Reality (AR) applications. Recent advancements in dense representation SLAM have highlighted the potential of leveraging neural scene representation and 3D Gaussian representation for high-fidelity spatial representation. In this paper, we propose a novel dense representation SLAM approach with a fusion of Generalized Iterative Closest Point (G-ICP) and 3D Gaussian Splatting (3DGS). In contrast to existing methods, we utilize a single Gaussian map for both tracking and mapping, resulting in mutual benefits. Through the exchange of covariances between tracking and mapping processes with scale alignment techniques, we minimize redundant computations and achieve an efficient system. Additionally, we enhance tracking accuracy and mapping quality through our keyframe selection methods. Experimental results demonstrate the effectiveness of our approach, showing incredibly fast speeds up to 107 FPS (for the entire system) and superior quality of the reconstructed map.", + "bbox": [ + 261, + 287, + 733, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The code is available at: https://github.com/Lab-of-AI-and-Robotics/GS-ICP-SLAM Video is: https://youtu.be/ebHh_uMMxE", + "bbox": [ + 261, + 494, + 756, + 522 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Coordinate-based 3D Representation $\\cdot$ G-ICP $\\cdot$ SLAM", + "bbox": [ + 261, + 537, + 658, + 550 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 579, + 356, + 593 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual Simultaneous Localization and Mapping (SLAM) is an algorithm that constructs maps of unknown environments while localizing poses of vision sensors simultaneously. As the influence of 3D visual SLAM in the fields of robotics, Virtual Reality (VR), and Augmented Reality (AR) has increased, higher rendering performance and more accurate trajectory are required. Thus, 3D reconstruction methods such as Signed Distance Field (SDF), and Truncated Signed Distance Field (TSDF) are utilized for traditional dense visual SLAM [3,5,10,27,30,42].", + "bbox": [ + 217, + 612, + 785, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, coordinate-based 3D Implicit Neural Representation (INR) [23] has been proposed for representing spatial information using neural radiance fields, showcasing high novel view synthesis capabilities and high-fidelity spatial representation prowess. Various approaches [8, 18, 32, 39, 45, 47] have been attempted to utilize INR for the real-time SLAM mapping process. However, INR requires computationally intensive raycasting to synthesize images, thus the rendering process in INR-based SLAM incurs significant time overhead, slowing down map optimization and rendering-loss based tracking. In contrast, coordinate-based 3D explicit representation such as 3D Gaussian", + "bbox": [ + 217, + 719, + 785, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9fd7eda37ed65e7f9c152ef05d387a65623c1897dcffc3e55e57a45c5241df0c.jpg", + "image_caption": [ + "Fig. 1: A comparison of PSNR with respect to FPS of entire system in recent research on SLAM algorithm utilizing dense representation such as neural scene representation and 3D Gaussian representation. Our method achieves state-of-the-art performance in rendering evaluation and FPS of entire system. Note that this FPS represents the overall system performance. Reported values are average of Replica 8 scenes." + ], + "image_footnote": [], + "bbox": [ + 316, + 152, + 671, + 309 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Splatting (3DGS) [15] represents the 3D space using 3D Gaussians as primitives, allowing for rendering speeds faster than NeRF using the rasterization method [43]. This rapid rendering capability of 3DGS results in fast optimization of 3D spatial information, making it suitable for dense representation in SLAM [12, 14, 22, 44].", + "bbox": [ + 215, + 433, + 785, + 494 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although 3DGS-based SLAM methods take advantages of the high-speed rendering, they fail to address the fundamental issue: the inability to directly utilize 3D explicit representations and the indirect tracking of 3D space through 2D image rendering. Even with the majority of current 3DGS-based SLAM [12, 14, 22, 32, 44] utilizing RGB-D data, the use of explicit representations is overlooked. For example, INR-based methods [32, 39, 47] and 3DGS-based methods [14, 22, 44] employ photometric-error based techniques, which estimate the optimal pose by iteratively minimizing the error between rendered and observed 2D images. However, due to limitations in tracking speed and performance, decoupled approaches [8, 12, 18] have been proposed incorporating well-crafted visual odometries [4, 26] into the tracking process. Although decoupled methods exhibit better performances in tracking by separating mapping and tracking processes, they require additional computational resources for the independent stages. Moreover, these methods need to store features in 3D space, unrelated to the 3DGS map.", + "bbox": [ + 215, + 501, + 787, + 698 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There exists a method that allows Gaussian, an explicit representation, to be directly used for tracking. The well-known Generalized Iterative Closest Point (G-ICP) [17, 33] from the 3D scan matching family is simple yet efficient for fast tracking of 3D point clouds. During preprocessing, it only requires computing Gaussians for the current frame and the map. Given that the map in 3DGS utilizes Gaussians as an explicit representation in 3D space, using G-ICP for tracking allows for the direct utilization of the 3DGS map without the need for post-processing. Furthermore, the Gaussians of the current frame computed during tracking with G-ICP can also be directly utilized as an explicit representation in the 3DGS map, without additional computations.", + "bbox": [ + 215, + 704, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Seongbo Ha et al.", + "bbox": [ + 271, + 114, + 379, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, we propose a dense representation SLAM framework that integrates G-ICP and 3DGS, allowing them to complement each other by sharing explicit representations. Unlike traditional methods, our approach actively leverages 3D information through G-ICP for tracking. The proposed method is a coupled approach that shares a single map during tracking and mapping processes, while maintaining fast tracking speed akin to decoupled approaches. Previous works using decoupled methods [8, 12, 18] require separate maps containing ORB feature information for tracking, along with additional resources and computations needed to obtain ORB features. Moreover, the values obtained during tracking are not utilized in mapping. Our approach exploits the covariance of each point computed during the G-ICP-based tracking process as the initial state of the 3DGS mapping. Meanwhile, the 3D Gaussians in the 3DGS map are also mutually used as 3D points and their covariances which are essential for G-ICP-based tracking, discarding the necessity to recalculate the covariances of the map. This is possible because the covariances of points computed during the G-ICP process and the 3D Gaussians representing the map commonly contain information about the surrounding space. In other words, G-ICP and 3DGS can share the same Gaussian world. Therefore, our system minimizes unnecessary computations and facilitates efficient system configuration by mutually utilizing the key elements, Gaussians, between tracking and mapping processes. To ensure optimal performance in tracking and mapping sharing information between G-ICP and 3DGS, we also introduce several techniques such as scale alignment. To summarize, our contributions are as follows:", + "bbox": [ + 217, + 146, + 787, + 463 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a real-time dense representation SLAM that combines G-ICP and 3DGS, achieving extremely high speed of the entire system (up to 107 FPS) and superior quality of the map.", + "- By incorporating G-ICP for tracking, our system utilizes 3D information actively and significantly reduces the time required for the tracking process.", + "- Reduction of computational cost and facilitating rapid convergence of primitives of 3DGS is achieved by sharing the covariances of G-ICP and 3DGS with scale aligning techniques." + ], + "bbox": [ + 225, + 470, + 785, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 217, + 613, + 366, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "G-ICP Scan-matching focuses on the registration of two point clouds that observe similar environments by selecting a transformation matrix that minimizes errors between two point clouds [2]. Iterative Closest Point (ICP) [1] algorithm is a widely used and influential method in the scan-matching area. ICP iteratively estimates point correspondences and finds a transformation that minimizes Euclidean distance between corresponding points, thereby optimizing the registration of point clouds. Because of simplicity and speed of ICP, there are many follow-up studies [7, 20, 33, 34, 41]. To improve robustness, Trimmed-ICP [7], proposed a method for correspondence selection. Point-to-plane ICP [20] augmented robustness and accuracy by taking point-to-plane distance as objective function. G-ICP [33] introduced probabilistic models to generalize ICP yielding notable improvements in robustness and accuracy. Scan matching is a common technique used in SLAM for pose tracking [6, 11, 21, 28, 29], and for enhanced robustness, feature-based scan matching is often employed [19, 35, 36, 46].", + "bbox": [ + 217, + 643, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "RGBD GS-ICP SLAM", + "bbox": [ + 593, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SLAM with Dense Representation SLAM with dense representation [9, 16] aims to construct maps in a dense form to enable interaction with the map in tasks such as AR, robotics, etc. To achieve this, classic approaches [3,5,10,27,30,42] creates dense maps by representing space in Signed Distance Field (SDF), and Truncated Signed Distance Field (TSDF), rather than in sparse forms such as point clouds or grids.", + "bbox": [ + 215, + 145, + 785, + 222 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recently, dense representation SLAM methods represent maps by utilizing NeRF [23], which demonstrates high spatial representation capabilities have been proposed. iMAP [39], NICE-SLAM [47], Point-SLAM [32], ESLAM [13] perform tracking by optimizing camera poses by reducing errors between synthesized and observed images. These methods have limitations in tracking speed and performance. Alternatively, Orbeez-SLAM [8], vMAP [18] incorporate well-crafted visual SLAM techniques into the tracking process, making mapping and tracking operate independently. Orbeez-SLAM utilizes ORB-SLAM2 [26] for tracking and instant-ngp [24] for learning spatial information to perform mapping. In a similar context, vMAP utilizes ORB-SLAM3 [4] for tracking while employing separate Multi-layer perceptrons [40] for each object to effectively represent the entire space.", + "bbox": [ + 215, + 223, + 787, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3DGS represents space in explicit form using 3D gaussians as primitives. It offers a level of high-fidelity spatial representation similar to NeRF but provides significantly faster rendering speeds. To leverage this advantage to dense representation SLAM, several approaches [12, 14, 22, 44] utilizing 3DGS for space representation have been proposed. Among them, GS-SLAM [44], SPLTAM [14], Gaussian Splating SLAM [22] perform tracking using dense photometric error. While 3DGS shows significantly improved rendering performance compared to NeRF, the tracking methods based on dense photometric error still suffer from tracking speed. To tackle this problem, Photo-SLAM [12] integrates ORB-SLAM3 for tracking, in a similar manner to Orbeez-SLAM and vMAP.", + "bbox": [ + 215, + 390, + 787, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 570, + 316, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To mutually benefit tracking and mapping, we introduce the fusion of G-ICP and GS, each representing the tracking and mapping processes, respectively. The key insight of our approach is that covariance can be considered as a fundamental common factor for this fusion. Suppose we have a point set (point cloud) $\\mathcal{X} = \\{\\pmb{x}_m\\}_{m=1,\\dots,M}$ and its corresponding covariance set $\\mathcal{C} = \\{C_m\\}_{m=1,\\dots,M}$ , where $\\pmb{x} = [x,y,z]^T$ . The covariance $C$ of one 3D point $\\pmb{x}$ is given by computing covariance matrix of $k$ -nearest neighbors of $\\pmb{x}$ . Let us define $\\pmb{G} = \\{\\pmb{\\mathcal{X}},\\pmb{\\mathcal{C}}\\}$ as a set of Gaussians. G-ICP aims to find a transformation $\\mathbf{T}$ that maximally aligns the source Gaussians (current frame) $\\pmb{G}^s = \\{\\pmb{\\mathcal{X}}^s,\\pmb{\\mathcal{C}}^s\\}$ and the target Gaussians (map) $\\pmb{G}^t = \\{\\pmb{\\mathcal{X}}^t,\\pmb{\\mathcal{C}}^t\\}$ . Assume we know the correspondences between $\\pmb{\\mathcal{X}}^s$ and $\\pmb{\\mathcal{X}}^t$ determined by nearest neighbor search. For example, we have $\\{\\pmb{x}_i^s\\}_{i=1,\\dots,N} \\subset \\pmb{\\mathcal{X}}^s$ , $\\{C_i^s\\}_{i=1,\\dots,N} \\subset \\pmb{\\mathcal{C}}^s$ and $\\{\\pmb{x}_i^t\\}_{i=1,\\dots,N} \\subset \\pmb{\\mathcal{X}}^t$ , $\\{C_i^t\\}_{i=1,\\dots,N} \\subset \\pmb{\\mathcal{C}}^t$ , where $\\pmb{x}_i^s$ is associated with $\\pmb{x}_i^t$ . To find the optimal transform $\\mathbf{T}^*$ , we exploit not the single point but the distribution of that point defined as a Gaussian distribution: $\\pmb{x}_i \\sim \\mathcal{N}(\\hat{\\pmb{x}}_i,C_i)$ . Let $d_i = \\pmb{x}_i^t - \\mathbf{T}\\pmb{x}_i^s$ be the error term, and if we assume that there is an optimal transformation $\\mathbf{T}^*$ , it is clear that $\\pmb{x}_i^t = \\mathbf{T}^*\\pmb{x}_i^s$ thus $\\hat{d}_i = 0$ . Since", + "bbox": [ + 215, + 609, + 787, + 843 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 116, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Seongbo Ha et al.", + "bbox": [ + 271, + 114, + 379, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we assume $\\pmb{x}$ is a Gaussian random variable, $d_{i}$ is also a Gaussian random variable as the following:", + "bbox": [ + 214, + 146, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} d _ {i} \\sim \\mathcal {N} \\left(\\hat {d} _ {i}, C _ {i} ^ {t} + \\mathbf {T} ^ {*} C _ {i} ^ {s} \\left(\\mathbf {T} ^ {*}\\right) ^ {T}\\right) \\\\ = \\mathcal {N} \\left(\\hat {x} _ {i} ^ {t} - \\mathbf {T} ^ {*} \\hat {x} _ {i} ^ {s}, C _ {i} ^ {t} + \\mathbf {T} ^ {*} C _ {i} ^ {s} \\left(\\mathbf {T} ^ {*}\\right) ^ {T}\\right) \\\\ = \\mathcal {N} (0, C _ {i} ^ {t} + \\mathbf {T} ^ {*} C _ {i} ^ {s} (\\mathbf {T} ^ {*}) ^ {T}). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 199, + 632, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To find the optimal transform $\\mathbf{T}^*$ for $\\boldsymbol{x}^s$ and $\\boldsymbol{x}^t$ , we use maximum likelihood estimation (MLE) as the following:", + "bbox": [ + 214, + 267, + 782, + 297 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {T} ^ {*} = \\underset {\\mathbf {T}} {\\operatorname {a r g m a x}} \\prod_ {i} ^ {N} p \\left(d _ {i}\\right) = \\underset {\\mathbf {T}} {\\operatorname {a r g m a x}} \\sum_ {i} ^ {N} \\log p \\left(d _ {i}\\right) \\\\ = \\underset {\\mathbf {T}} {\\operatorname {a r g m i n}} \\sum_ {i} ^ {N} d _ {i} ^ {T} \\left(C _ {i} ^ {B} + \\mathbf {T} C _ {i} ^ {A} \\mathbf {T} ^ {T}\\right) ^ {- 1} d _ {i}. \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 306, + 785, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, $\\mathbf{T}^*$ can be used as the relative pose between the current frame $G^{s} = \\{\\mathcal{X}^{s},\\mathcal{C}^{s}\\}$ and the map $G^{t} = \\{\\mathcal{X}^{t},\\mathcal{C}^{t}\\}$ .", + "bbox": [ + 214, + 398, + 792, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Meanwhile, for mapping purposes, GS also relies on Gaussians $G = \\{\\mathcal{X},\\mathcal{C}\\}$ of the 3D scene representation. Different from G-ICP, GS aims to find the optimal coordinates of the Gaussians $\\mathcal{X}^* = \\{x_m^*\\}_{m=1,\\dots,M}$ and the optimal covariances $\\mathcal{C}^* = \\{C_m^*\\}_{m=1,\\dots,M}$ as the following:", + "bbox": [ + 214, + 429, + 785, + 491 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol{\\mathcal{X}}^{*},\\boldsymbol{\\mathcal{C}}^{*},\\boldsymbol{H}^{*},\\boldsymbol{O}^{*} = \\operatorname *{argmin}_{\\boldsymbol {\\mathcal{X}},\\boldsymbol {\\mathcal{C}},\\boldsymbol {H},\\boldsymbol{O}}\\lambda_{I_{1}}\\boldsymbol{\\mathcal{L}}_{1}\\left(I,I_{gt}\\right) + \\lambda_{I_{2}}\\boldsymbol{\\mathcal{L}}_{D - SSIM}\\left(I,I_{gt}\\right) + \\lambda_{D}\\boldsymbol{\\mathcal{L}}_{1}\\left(D,D_{gt}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 497, + 792, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $H = \\{h_m\\}_{m=1,\\dots,M}$ and $O = \\{o_m\\}_{m=1,\\dots,M}$ are the color set and the opacity set of 3D points which are for the RGBD image rendering. $I$ and $D$ are the rendered RGB and depth images obtained by performing rasterization using $G$ , $H$ and $O$ .", + "bbox": [ + 214, + 531, + 782, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that in G-ICP and GS, the key common factor is Gaussians $G = \\{\\mathcal{X},\\mathcal{C}\\}$ , allowing these Gaussians to be shared mutually. During G-ICP tracking, the covariance of each frame is computed. Hence, when adding keyframes to expand the 3DGS map, there is no need to recalculate $\\mathcal{C}$ for every expansion. Also, G-ICP does not need to compute the covariances of the map because our GS map already contains Gaussians. Moreover, G-ICP, by aligning frames based on 3D geometric structure, inherently initializes a certain number of points and their coordinates that aptly represent the 3D structure. Therefore, it brings about an effect where an appropriate number of points and their poses are initialized, well-suited for depicting the 3D structure, significantly reducing the learning time to find the optimal pose $\\mathcal{X}^*$ and the optimal covariance $\\mathcal{C}^*$ of Gaussians $G$ in GS. Additionally, in the same vein, calculations like densifying or opacity reset to adjust the number of 3D points in GS become unnecessary. Consequently, by sharing a common source, each process becomes mutually beneficial, and the speed of execution accelerates due to reduced redundant computations. Our method can be simplified as follows:", + "bbox": [ + 214, + 577, + 785, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "1. Use G-ICP to align the current frame with the 3DGS map which contains covariance (solely need to compute the covariance for the current frame).", + "bbox": [ + 222, + 809, + 782, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "RGBD GS-ICP SLAM", + "bbox": [ + 591, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1ad7b8524217766017f170291f86933956bf6b83ec241d6c565bef0ae7db4a10.jpg", + "image_caption": [ + "Fig. 2: System Overview. The input of our system is RGBD frame. We generate a point cloud by downsampling and reprojecting the current depth image and utilize it in the GICP process. During the GICP process, we create source Gaussians from the point cloud and estimate the current camera pose by aligning them with target Gaussians, which are a subset of the 3DGS map. If the current frame is identified as a keyframe or a mapping-only keyframe, we add the source Gaussians to the 3DGS map as new primitives. Meanwhile, in the mapping process, we optimize the Gaussians along with the color and opacity set of the Gaussians concurrently with the tracking process." + ], + "image_footnote": [], + "bbox": [ + 228, + 145, + 769, + 262 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2. When adding keyframes to the 3DGS map, utilize the covariance computed in GICP during tracking (no need for densifying or opacity reset).", + "3. Repeat steps 1-2." + ], + "bbox": [ + 220, + 410, + 784, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The overview of our system is shown in Fig. 2. The detailed implementations of the proposed technique are introduced in the following Section.", + "bbox": [ + 212, + 460, + 787, + 492 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 G-ICP Tracking", + "text_level": 1, + "bbox": [ + 214, + 511, + 372, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Scale Regularization When using G-ICP for tracking, aligning the scale of the current frame with the map enables high-performance camera pose estimation [33]. Given the covariance $C$ , scale $\\boldsymbol{S} = [s_2, s_1, s_0]^T$ is given by the singular value decomposition (SVD) as the following:", + "bbox": [ + 212, + 532, + 787, + 594 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nC = \\boldsymbol {R} \\boldsymbol {\\Lambda} ^ {2} \\boldsymbol {R} ^ {T} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 450, + 601, + 785, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\Lambda = \\text{diag}(s_2, s_1, s_0) \\forall s_2 > s_1 > s_0$ is a scale matrix and $\\pmb{R}$ is the orientation of the Gaussian. To achieve the robust scan matching performance, point-to-plane ICP or voxelized point-to-plane ICP adopt a regularizing method that makes the scale as $\\pmb{S} = [1, 1, \\epsilon]^T$ in order to treat each Gaussian as a plane-like distribution.", + "bbox": [ + 212, + 627, + 787, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our framework, we utilize target Gaussians $\\pmb{G}^t = \\{\\pmb{x}^t,\\pmb{c}^t\\}$ from the 3DGS map for tracking. Here, $\\pmb{c}^t$ is optimized to represent the scene accurately, $\\pmb{G}^t$ involves not only planes but also lines, corners, and other features. Therefore, while tracking, instead of regularizing the scale into a plane-based form, preserving the original characteristics of the target Gaussians while performing regularization is more appropriate. Thus, in our framework, we propose the following ellipsoid regularization:", + "bbox": [ + 212, + 686, + 787, + 779 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Lambda} ^ {\\prime} = \\frac {1}{m e d i a n (\\boldsymbol {S})} \\operatorname {d i a g} \\left(s _ {2}, s _ {1}, s _ {0}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 382, + 785, + 785, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\pmb{\\Lambda}^{\\prime}$ is the ellipsoid-regularized scale matrix.", + "bbox": [ + 214, + 823, + 545, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 116, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Seongbo Ha et al.", + "bbox": [ + 271, + 114, + 379, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4fecd2b2fe34b0ccf75cc0977939a3930ab9e131e3d788583ce912a62a12304e.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 232, + 155, + 480, + 276 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ae889a95482c4e04d4a00af4e0c0799196234122024407ae5202eec361e2cdc3.jpg", + "image_caption": [ + "(b)", + "Fig. 3: Tracking Accuracy Comparison Based on Keyframe Selection Methods. The reported values represent the average results across eight scenes from the Replica dataset [37]. When selecting keyframes every n frame (depicted in blue), the tracking accuracy is notably low. Conversely, our keyframe selection method yielded the highest tracking accuracy." + ], + "image_footnote": [], + "bbox": [ + 516, + 155, + 767, + 275 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Keyframe Selection Similar to [25,31], we perform dynamic keyframe selection. Considering the geometric structure is critical for both G-ICP tracking and GS mapping, we exploit the geometric correspondence computed from G-ICP. As Eqn. (1) denotes, we can get the distance between $\\boldsymbol{x}_i^s$ and $\\boldsymbol{x}_i^t$ as an interim result of G-ICP procedure without additional calculations. So we can efficiently determine the correspondence by setting the distance threshold. By considering the proportion of correspondences between the current frame and map, we select keyframes while taking into account the characteristics of the scene. In more detail, if the proportion in the current frame goes below specific thresholds, the frame is selected as a keyframe. This allows us to achieve consistent tracking performance while simultaneously maintaining a consistent density of Gaussians $G = \\{\\mathcal{X},\\mathcal{C}\\}$ added to the map. The effect of this keyframe selection is shown in Fig. 3 (a). Note that, from selected keyframe, only Gaussians that do not overlap with the existing current map are considered as target Gaussians. Errors from discrepancies during tracking are directly incorporated into the map, leading to accumulated inaccuracies and reduced tracking precision. This strategy ensures more accurate tracking results and improves the overall robustness of the system.", + "bbox": [ + 212, + 386, + 787, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 GS Mapping", + "text_level": 1, + "bbox": [ + 215, + 648, + 346, + 665 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scale Aligning Suppose we have a fully trained 3DGS map. Since the 3DGS map is trained to represent the scene from any viewpoint, Gaussians $G$ should be uniformly distributed according to the structure of the 3D space. However, single frames obtained from radiance sensors such as RGBD cameras or LiDAR yield sparse representations of the 3D space as the distance from the sensor increases, due to the inherent characteristics of the sensor. In other words, as the distance from the sensor increases, the spacing between 3D points widens, and the scale of the corresponding covariance calculated based on the $k$ -nearest neighbors becomes excessively large. Adding such single frames as keyframes to the map during real-time training of GS mapping leads to an imbalance in the scale of Gaussians, ultimately resulting in a decrease in mapping performance. Therefore, to fundamentally alleviate the problem, our proposed method suggests the", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "RGBD GS-ICP SLAM", + "bbox": [ + 593, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/986f348d0f8449baaf0b34adfe9c8d720af5be7e608131bc957215d3cc8fb216.jpg", + "image_caption": [ + "Fig. 4: Separated Keyframe Selection on Replica office4. We demonstrate that a small number of tracking keyframes yield accurate trajectory estimation (case 1), while a large number of mapping keyframes result in high rendering performance (case 2). Thus, our method adopts case 3 that selects tracking keyframe and mapping keyframe separately at different intervals." + ], + "image_footnote": [], + "bbox": [ + 230, + 147, + 395, + 276 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d1280fceeaacaf642d05cfefae8894bcfa8c6c6e7c6bd82c421f693ed078b089.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 147, + 578, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c4a61d2aa57e457d2ad0ebbdc8857eaeb666befccf32c8f388bc4a2c12768665.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 146, + 774, + 273 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "following scale normalization for the current frame:", + "bbox": [ + 215, + 369, + 560, + 383 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\varLambda} ^ {\\prime \\prime} = \\frac {1}{z ^ {p}} \\boldsymbol {\\varLambda}.\n$$\n", + "text_format": "latex", + "bbox": [ + 457, + 391, + 545, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Where $\\Lambda''$ is the normalized scale matrix and $p$ is the parameter which is empirically determined. By adding such scale-normalized keyframes to the map to expand it, not only does the GS mapping performance increase, but also the performance of G-ICP tracking improves.", + "bbox": [ + 214, + 426, + 787, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additional Keyframe Selection for Mapping Mapping based on GS benefits from having a diverse set of frames from various viewpoints for training, as performance improves with a larger number of available frames. However, increasing the number of selected keyframes during tracking can lead to a degradation in tracking performance due to accumulated errors, as is common with G-ICP-based methods. To mitigate this issue, we propose adding mapping-only keyframe selection in addition to the existing keyframe selection process. In other words, tracking continues to proceed from selected keyframes in the existing tracking process, while mapping utilizes both the original keyframes and additional mapping-only keyframes.", + "bbox": [ + 212, + 507, + 787, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The overview of the additional keyframe selection is illustrated in Fig. 4. Cases 1 and 2 represent scenarios without considering mapping-only keyframes. In Case 1, to increase the number of keyframes, if no keyframe satisfying the criteria is found up to the $30^{th}$ frame from the previous keyframe, the $30^{th}$ frame is selected as a keyframe. However, since this may not yield sufficient mapping performance, as shown in Case 2, adding an additional keyframe at the $10^{th}$ frame can be considered. While this approach improves mapping performance as desired, it leads to a decline in tracking performance due to accumulated scan matching errors. To alleviate this, as shown in Case 3, we propose adding the $10^{th}$ frame as a mapping-only frame, allowing us to maximize tracking performance while simultaneously maximizing mapping performance. The effect of the proposed scheme, Case 3, is shown in Fig. 3 (b). Note that since G-ICP, which performs tracking, computes covariance for all frames anyway, there is no difference in the overall computational load.", + "bbox": [ + 214, + 643, + 789, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Seongbo Ha et al.", + "bbox": [ + 271, + 114, + 379, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Avoiding Local Minima while Mapping GS SLAM differs from traditional GS in that it operates in real-time and often lacks sufficient observations of the scene. Moreover, since it operates in real-time, focusing solely on images from the current viewpoint or nearby viewpoints during training can easily lead to local minima and degrade mapping quality [22]. For instance, GS simply considers image rendering from specific viewpoints, thus if we continuously train only on a specific viewpoint, the scale of Gaussians in the map tends to elongate in the direction of that viewpoint, leading to local minima.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To overcome this challenge, at each training iteration, we adopt a strategy of randomly choosing one keyframe for learning among the selected keyframes so far, ensuring that both the current observed scene and the entire map are uniformly learned. Additionally, we prune Gaussians that fall into local minima during training to improve mapping performance and ensure robust tracking through geometric preservation.", + "bbox": [ + 212, + 253, + 787, + 328 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 353, + 356, + 371 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 214, + 386, + 397, + 402 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We evaluate our method on Replica dataset [37] and TUM dataset [38]. Replica dataset contains synthetic scenes and high-quality RGB/depth images rendered from these scenes. TUM dataset includes images captured in the real world, with significant noise and blur, resulting in poor quality. Particularly, many parts of the depth images suffer from information loss. We validate the effectiveness of our approach by evaluating it on both synthetic and real-world datasets. All experiments are performed on a desktop with a Ryzen 7 7800x3d CPU, 32GB RAM, and an NVIDIA RTX 4090 24GB GPU. To evaluate camera tracking accuracy, we use the Root Mean Square Error (RMSE) of the Absolute Trajectory Error (ATE). For the quality of the reconstructed map, we report standard photometric rendering quality metrics (PSNR, SSIM, and LPIPS).", + "bbox": [ + 212, + 412, + 787, + 565 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/8dc1fdab1ca9e844d2e0399401a2e83ae1955bdbd948f443268cc6ed55db8e03.jpg", + "table_caption": [ + "Table 1: Tracking Performance on Replica and TUM-RGBD (ATE RMSE $\\downarrow$ [cm]). Our method archives state-of-the-art tracking accuracy on Replica. On TUM, Among coupled methods, we achieve competitive performance. The results of GS-SLAM/Gaussian Splitting SLAM are taken from [44]/ [22]. And the results of ORB-SLAM3 and Photo-SLAM are obtained from [12]." + ], + "table_footnote": [ + "* denotes the reproduced results by running official code." + ], + "table_body": "
MethodReplicaTUM
R0R1R2Of0Of1Of2Of3Of4Avg.fr1/deskfr2/xyzfr3/officeAvg.
DecoupledORB-SLAM3 [4]-------1.81.70.41.71.3
Photo-SLAM [12]-------0.602.60.31.01.3
CoupledNICE-SLAM* [47]1.611.481.610.950.811.461.761.691.422.82.17.24.0
Point-SLAM* [32]0.590.510.320.450.460.480.610.870.542.71.33.92.6
GS-SLAM [44]0.480.530.330.520.410.590.460.700.503.31.36.63.7
SplaTAM* [14]0.290.350.280.490.210.310.340.570.363.31.35.13.2
Gaussian Splitting SLAM* [22]0.330.220.290.360.190.250.120.810.321.51.41.51.5
Ours (limited to 30 FPS)0.150.160.110.180.120.170.160.210.162.71.82.72.4
", + "bbox": [ + 223, + 674, + 782, + 796 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "RGBD GS-ICP SLAM", + "bbox": [ + 593, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Camera Tracking Accuracy", + "text_level": 1, + "bbox": [ + 215, + 146, + 452, + 162 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Tab. 1 shows the tracking accuracy of our method compared to other approaches on both synthetic [37] and real-world [38] datasets. On Replica [37], the proposed method achieved state-of-the-art (SOTA) performance across all scenes, reducing the trajectory error by more than $50\\%$ compared to the previous SOTA. This result arises from the fact that our proposed method actively utilizes 3D information by employing G-ICP [33] for tracking, unlike other methods that perform tracking based on errors in 2D space. On TUM [38], our method shows competitive tracking accuracy among coupled methods [14,22,32,44,47] which conduct tracking and mapping based on a single map. Note that the results of Gaussian Splitting SLAM are evaluated only on keyframes. Decoupled method [12] and its baseline [4] show better performance, but these approaches require additional resources for storing separate map information and costly computations for extracting features solely for the tracking process.", + "bbox": [ + 212, + 180, + 789, + 363 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/3473d062f1bac9bf8fd26289dda744c5a8ac26f58dd808c49064673f9f9f1afd.jpg", + "table_caption": [ + "Table 2: Evaluation of System Speed and Map Quality on Replica. Our method outperforms all other frameworks in both system speed and quality of the reconstructed map." + ], + "table_footnote": [ + "* denotes the reproduced results by running official code. The result of GS-SLAM and Gaussian Splitting SLAM is taken from [44] and [22]." + ], + "table_body": "
MethodsMetricsR0R1R2Of0Of1Of2Of3Of4Avg.
Orbeez-SLAM* [8]PSNR[dB]↑12.1315.2815.8717.5919.2610.3011.5512.6514.33
SSIM ↑0.700.790.810.800.670.780.760.840.77
LPIPS ↓0.610.480.510.520.410.550.540.530.53
FPS ↑24.0524.7824.0019.2226.7423.8824.4526.1124.15
Point-SLAM* [32]PSNR[dB]↑33.3834.1036.3238.7239.3134.2234.1034.8235.62
SSIM ↑0.980.980.990.990.990.960.960.980.98
LPIPS ↓0.100.120.100.090.110.150.120.130.11
FPS ↑0.260.300.310.330.340.300.280.300.30
GS-SLAM [44]PSNR[dB]↑31.5632.8632.5938.7041.1732.3632.0332.9234.27
SSIM ↑0.970.970.970.990.990.980.970.970.98
LPIPS ↓0.090.080.090.050.030.090.110.110.08
FPS ↑8.34-------8.34
SplaTAM* [14]PSNR[dB]↑32.6033.5534.8338.0939.0231.9529.5331.5533.89
SSIM ↑0.980.970.980.980.980.970.950.950.97
LPIPS ↓0.070.100.070.090.090.100.120.150.10
FPS ↑0.240.190.190.200.220.270.260.240.23
Gaussian Splitting SLAM [22]PSNR[dB]↑34.8336.4337.4939.9542.0936.2436.7036.0737.50
SSIM ↑0.950.960.970.970.980.960.960.960.96
LPIPS ↓0.070.080.080.070.060.080.070.100.08
FPS ↑-----1.1---
Ours (no tracking speed limit)PSNR[dB]↑32.2035.3634.4240.3140.7533.8534.0836.4735.93
SSIM ↑0.940.960.960.980.980.960.950.960.96
LPIPS ↓0.080.070.080.050.050.070.070.070.07
FPS ↑100.9884.92103.3899.10107.0695.6097.2096.6698.11
Ours (limited to 30 FPS)PSNR[dB]↑35.3737.8038.5043.1343.2636.9336.9038.7538.83
SSIM ↑0.960.970.980.990.990.970.970.970.98
LPIPS ↓0.050.050.050.030.030.040.040.050.04
FPS ↑29.9729.9829.9829.9829.9929.9729.9729.9729.98
", + "bbox": [ + 235, + 433, + 772, + 811 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Seongbo Ha et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/bf4ff4d8d94b24eb3b818df2ba0b8278ecc3fe7b0125945f977753a3e3860ad4.jpg", + "table_caption": [ + "Table 3: Evaluation of System Speed and Map Quality on TUM-RGBD. Proposed method shows incredible system speed and competitive map quality." + ], + "table_footnote": [ + "* denotes the reproduced results by running official code. The result of Photo-SLAM is taken from [12]." + ], + "table_body": "
MethodsPSNR[dB] ↑SSIM ↑LPIPS ↓FPS ↑
NICE-SLAM* [47]14.100.5740.3950.08
Point-SLAM* [32]21.400.7380.4470.22
Photo-SLAM [12]21.900.7630.187-
SplaTAM* [14]23.460.9060.1560.32
Ours (unlimited tracking speed)19.620.7500.24073.92
Ours (limited to 30 FPS)20.720.7680.21829.99
", + "bbox": [ + 316, + 184, + 689, + 281 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 System Speed and Quality of Reconstructed Map", + "text_level": 1, + "bbox": [ + 215, + 363, + 598, + 378 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Tab. 2 and Tab. 3 demonstrate the FPS of the systems and quality of the reconstructed map on Replica [37] and TUM [38] dataset, respectively. FPS of the system is calculated by dividing the total number of frames by the total time. Note that this FPS represents the overall system performance including tracking and mapping processes, rather than just individual components. Photo-SLAM [12] reported the tracking FPS of their system, but did not include the system FPS, so we leave that field blank in the Tab. 3. Since our method implements mapping and tracking processes to operate in parallel, the number of map optimization iterations varies depending on the tracking speed, consequently affecting the quality of the reconstructed map. So we evaluate our system in two cases: one where the tracking speed is limited to the typical sensor input speed of 30 FPS, and the other where the tracking speed is not limited. Please refer to the supplementary material for results across various FPS and geometric quality evaluations.", + "bbox": [ + 212, + 393, + 787, + 575 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e969dc582149105ea922270bd1c198849e75b97ae6be9ed350ecc6ee3639711c.jpg", + "image_caption": [ + "Fig. 5: Comparison of Rendering Results. In the first scene, SplaTAM [14] failed to reconstruct the pillow and lamp and Point-SLAM [32] failed to represent the detailed pattern of the pillow. In the second case, SplaTAM and Point-SLAM failed to accurately reconstruct the details of the clock. However, our method exhibits rendering results that closely resemble the ground truth image, demonstrating a high level of accuracy." + ], + "image_footnote": [], + "bbox": [ + 222, + 614, + 782, + 756 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "RGBD GS-ICP SLAM", + "bbox": [ + 593, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 769, + 116, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In the limited case (30 FPS), our system achieves SOTA map quality across all scenes in the Replica dataset. When the tracking speed is not limited, our method shows extremely fast system speed up to 107 FPS. This speed is more than four times faster than that of the Orbeez-SLAM [8], based on the ORB-SLAM2 [26]. Remarkably, even under this condition, our approach maintains superior map quality and outperforms other methods. The proposed method effectively integrates Gaussians for map representation with appropriate initial states and has better detailed representation capability compared to NeRF-based methods. As shown in Fig. 5, SplaTAM suffers from incorporating appropriate Gaussians into the map, leading to regions of poor quality. Point-SLAM exhibits areas where detailed representation is lacking. In TUM-RGBD, our method demonstrates remarkably fast system speed with competitive map quality, showcasing its practicality in real-world scenarios. Compared to SplaTAM [14], our approach shows a slight decrease in map quality, with a PSNR reduction of approximately $11.7\\%$ . However, we achieve significant speed improvements, with a speedup of approximately 91.6 times under limited conditions and up to 227 times at maximum speed. The reason is that, unlike Replica, TUM dataset is captured with old-fashioned sensors, resulting in noisy and substantial information loss in depth images. While SplaTAM treats these factors by adding new Gaussians only in regions with significant depth loss, our method utilizes all structural information from the depth image, focusing on accurate tracking and rapid system rather than treating such factors.", + "bbox": [ + 212, + 146, + 787, + 450 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Ablation study", + "text_level": 1, + "bbox": [ + 215, + 468, + 359, + 484 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/2a5d0ffed9daf7f393f0f1eee15d218300941b5d02c7fd14bedf11e514372e92.jpg", + "table_caption": [ + "Table 4: Scale Regularization Ablation on TUM-RGBD. Reported results are the average ATE RMSE $\\downarrow$ [cm] of 3 scenes in TUM-RGBD dataset." + ], + "table_footnote": [], + "table_body": "
Scale Regularizationfr1-deskfr2-xyzfr3-officeAvg.
X136.32196.61376.40236.54
Plane65.635.5316.2129.12
ellipsoid2.661.772.672.37
", + "bbox": [ + 331, + 559, + 669, + 631 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Scale Regularization Tab. 4 shows the result of the ablation study on scale regularization while tracking. We perform scale regularization to enhance tracking accuracy by ensuring that the scales of Gaussians in the current frame and the map are within a similar range. Despite the application of plane regularization resulting in some improvement in tracking accuracy, it remains inadequate. This is because the plane regularization method considers all the Gaussians as planes, ignoring the characteristics of Gaussians existing in the map, which are optimized to represent the surrounding space through the mapping process. On the other hand, when applying the proposed ellipsoid regularization, the best tracking accuracy is achieved. This is attributed to the ability of the proposed ellipsoid regularization to consider the characteristics of Gaussians existing in the map while performing regularization.", + "bbox": [ + 212, + 657, + 787, + 824 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Scale Aligning We implement scale aligning to reduce the difference between scales", + "bbox": [ + 212, + 824, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Seongbo Ha et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/e249124b1b27d2afeaafaea8ae5ad79428d1c87c172951eb9eed0c303ec20a62.jpg", + "table_caption": [ + "Table 5: Ablation of Scale Aligning on Replica. The results are the average of 8 scenes in the Replica dataset." + ], + "table_footnote": [], + "table_body": "
Covs from G-ICPRe-initATE ↓PSNR ↑SSIM ↑LPIPS ↓
XX8.89324.810.8400.380
✓X0.25833.210.9220.177
✓constant0.15837.330.9640.074
✓z1.50.15738.830.9750.041
", + "bbox": [ + 346, + 184, + 658, + 258 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "of existing Gaussians in the map and newly added Gaussians. By doing this, we ensure that scales of newly added Gaussians closely resemble the optimal value they will reach through optimization. To demonstrate the effectiveness of this approach, we tested our method in various cases and Tab. 5 shows the results. When not utilizing the covariances computed during the G-ICP [33] process, we calculate the scales of Gaussians by using simple-knn module of vanilla 3DGS [15], and set rotations as identity. In this case, a significant drop in tracking performance appeared. This is because the newly added Gaussians lack sufficient prior information about the 3D structure, causing them to be utilized as target Gaussians without being optimally optimized in space. In the case of using the computed covariance from the G-ICP process, the tracking and mapping performance improves. Furthermore, the best performance is observed when dividing the scale by $z^{1.5}$ . These results indicate that utilizing covariances calculated during G-ICP procedure is effective for both tracking and mapping, and scale aligning facilitates smooth connection from G-ICP to 3DGS. Further analysis of the advantages of this connection is included in the supplementary material.", + "bbox": [ + 212, + 286, + 787, + 512 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/418d1cc2e0de27e8e81062fa263c5fa2400efe56b73e3e1f65b5a314a7aca559.jpg", + "table_caption": [ + "Table 6: Keyframe and Mapping-Only Keyframe Selection Ablation on Replica. Reported results are the average of 8 scenes in the Replica dataset." + ], + "table_footnote": [], + "table_body": "
KeyframeMapping-Only KeyframeATE [cm] ↓PSNR ↑SSIM ↑LPIPS ↓
tracking + every 30 frame-0.17437.520.9680.055
tracking + every 10 frame-0.17638.830.9740.041
trackingevery 10 frame0.15738.830.9750.041
", + "bbox": [ + 274, + 579, + 730, + 646 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Additional Keyframe Selection for Mapping For mapping quality, sufficient keyframes are required, but indiscriminate addition of keyframes causes a drop in tracking accuracy. To address this challenge and maximize both tracking accuracy and mapping quality, we pick mapping-only keyframes in addition to the original keyframes. Tab. 6 presents the result of the ablation study of this method. When using only keyframes without mapping-only keyframes, camera tracking accuracy suffers due to accumulated errors in scan-matching. However, employing our proposed method yields the best tracking accuracy and mapping quality. This shows that our method selectively adds keyframes essential for tracking, minimizing accumulated errors in scan-matching. While leveraging these benefits, we ensure sufficient primitives and images for training, which are crucial for map quality through the use of mapping-only keyframes.", + "bbox": [ + 212, + 672, + 789, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "RGBD GS-ICP SLAM", + "bbox": [ + 593, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 769, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/d71c09b959536ce1569f17ca350c1ff83168c436d9592c760b8cf9671f7e8acf.jpg", + "table_caption": [ + "Table 7: Ablation of Methods for Avoiding Local Minima While Mapping on Replica. The results are the average of 8 scenes in the Replica dataset." + ], + "table_footnote": [], + "table_body": "
(a)(b)
Keyframe choice for learningPSNRSSIMLPIPSdensifying pruningATE [cm]PSNRSSIMLPIPS
recent 1 keyframe26.850.8720.200✓✓0.18838.650.973
covisible keyframes31.340.9240.120✓✗0.17037.900.971
random keyframes38.830.9750.041✗✓0.15738.830.975
✗✗0.15938.720.974
", + "bbox": [ + 218, + 195, + 784, + 285 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Avoiding Local Minima while Mapping We employ two methods to prevent Gaussians from overfitting to the training image, which results in elongation to the viewpoint direction: (1) training with randomly selected keyframes and (2) Gaussian pruning. Tab. 7 presents the results of the ablation study on these methods. The worst rendering performance is observed when repeatedly training with the recently added keyframe, while using covisible keyframes for training yields better results. Covisible keyframes offer more diverse viewpoints for the Gaussians compared to a single keyframe, yet they still provide limited diversity in viewpoints. Random keyframes provide the most diverse viewpoints compared to other methods, leading to a significant improvement in map quality. In the vanilla 3DGS [15], Gaussian densifying/pruning techniques are utilized to manage Gaussians. Tab. 7 (b) shows the results of an ablation test of Gaussian densifying/pruning in our method. Since our method appropriately supplies Gaussians for scene representation, densifying is unnecessary. However, pruning improves tracking accuracy and map quality by removing Gaussians that have become elongated due to overfitting or are no longer essential for representing the map.", + "bbox": [ + 212, + 313, + 789, + 541 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 564, + 346, + 582 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we have proposed RGBD GS-ICP-SLAM, a dense representation SLAM system that leverages 3D Gaussian representation for high-fidelity spatial representation. We demonstrate that a fusion of G-ICP and 3DGS that utilizes a single 3D Gaussian map for both tracking and mapping yields mutual benefits. The exchange of Gaussians between tracking and mapping processes with scale alignments minimizes redundant computations and constructs an efficient system. Moreover, our dynamic keyframe selection method enhances both tracking and mapping performance. Through extensive experiments, the proposed approach presents state-of-the-art performance in spatial representation, camera pose estimation, and total system speed.", + "bbox": [ + 212, + 597, + 787, + 734 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. The proposed method achieved fast system speed by relying solely on depth for the 3D structure. In real-world, there are limitations in the quality of maps due to the depth noise in RGB-D sensors. Since the system speed is already exceptionally fast, it is expected that robust performance can be achieved by trading off a bit of speed to compensate for noisy depth images with relatively robust RGB information.", + "bbox": [ + 212, + 734, + 787, + 809 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgment. This work was supported by the National Research Foundation of Korea (NRF) grant (no. RS-2024-00359937).", + "bbox": [ + 214, + 809, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Seongbo Ha et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 310, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Besl, P.J., McKay, N.D.: Method for registration of 3-d shapes. In: Sensor fusion IV: control paradigms and data structures. vol. 1611, pp. 586-606. Spie (1992)", + "2. Biber, P., Straβer, W.: The normal distributions transform: A new approach to laser scan matching. In: Proceedings 2003 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2003)(Cat. No. 03CH37453). vol. 3, pp. 2743-2748. IEEE (2003)", + "3. Bylow, E., Sturm, J., Kerl, C., Kahl, F., Cremers, D.: Real-time camera tracking and 3d reconstruction using signed distance functions. In: Robotics: Science and Systems. vol. 2, p. 2 (2013)", + "4. Campos, C., Elvira, R., Rodríguez, J.J.G., Montiel, J.M., Tardós, J.D.: Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics 37(6), 1874-1890 (2021)", + "5. Canelhas, D.R., Stoyanov, T., Lilienthal, A.J.: Sdf tracker: A parallel algorithm for on-line pose estimation and scene reconstruction from depth images. In: 2013 IEEE/RSJ International Conference on Intelligent Robots and Systems. pp. 3671-3676. IEEE (2013)", + "6. Ceriani, S., Sánchez, C., Taddei, P., Wolfart, E., Sequeira, V.: Pose interpolation slam for large maps using moving 3d sensors. In: 2015 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 750-757. IEEE (2015)", + "7. Chetverikov, D., Svirko, D., Stepanov, D., Krsek, P.: The trimmed iterative closest point algorithm. In: 2002 International Conference on Pattern Recognition. vol. 3, pp. 545-548. IEEE (2002)", + "8. Chung, C.M., Tseng, Y.C., Hsu, Y.C., Shi, X.Q., Hua, Y.H., Yeh, J.F., Chen, W.C., Chen, Y.T., Hsu, W.H.: Orbeez-slam: A real-time monocular visual slam with orb features and nef-realized mapping. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 9400-9406. IEEE (2023)", + "9. Czarnowski, J., Laidlow, T., Clark, R., Davison, A.J.: Deepfactors: Real-time probabilistic dense monocular slam. IEEE Robotics and Automation Letters 5(2), 721-728 (2020)", + "0. Dai, A., Nießner, M., Zollhöfer, M., Izadi, S., Theobalt, C.: Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics (ToG) 36(4), 1 (2017)", + "1. Grisetti, G., Kümmerle, R., Stachniss, C., Burgard, W.: A tutorial on graph-based slam. IEEE Intelligent Transportation Systems Magazine 2(4), 31-43 (2010)", + "2. Huang, H., Li, L., Cheng, H., Yeung, S.K.: Photo-slam: Real-time simultaneous localization and photorealistic mapping for monocular, stereo, and rgb-d cameras. arXiv preprint arXiv:2311.16728 (2023)", + "3. Johari, M.M., Carta, C., Fleuret, F.: Eslam: Efficient dense slam system based on hybrid representation of signed distance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17408-17419 (2023)", + "4. Keetha, N., Karhade, J., Jatavallabhula, K.M., Yang, G., Scherer, S., Ramanan, D., Luiten, J.: Splatam: Splat, track & map 3d gaussians for dense rgb-d slam. arXiv preprint arXiv:2312.02126 (2023)", + "5. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)", + "6. Kerl, C., Sturm, J., Cremers, D.: Dense visual slam for rgb-d cameras. In: 2013 IEEE/RSJ International Conference on Intelligent Robots and Systems. pp. 2100-2106. IEEE (2013)", + "7. Koide, K., Yokozuka, M., Oishi, S., Banno, A.: Voxelized gicp for fast and accurate 3d point cloud registration. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 11054-11059. IEEE (2021)" + ], + "bbox": [ + 225, + 178, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "RGBD GS-ICP SLAM", + "bbox": [ + 593, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 769, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Kong, X., Liu, S., Taher, M., Davison, A.J.: vmap: Vectorised object mapping for neural field slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 952-961 (2023)", + "19. Lin, J., Zhang, F.: Loam livox: A fast, robust, high-precision lidar odometry and mapping package for lidars of small fov. In: 2020 IEEE International Conference on Robotics and Automation (ICRA). pp. 3126-3131. IEEE (2020)", + "20. Low, K.L.: Linear least-squares optimization for point-to-plane 1cp surface registration. Chapel Hill, University of North Carolina 4(10), 1-3 (2004)", + "21. Mallios, A., Ridao, P., Ribas, D., Hernández, E.: Scan matching slam in underwater environments. Autonomous Robots 36, 181-198 (2014)", + "22. Matsuki, H., Murai, R., Kelly, P.H., Davison, A.J.: Gaussian splatting slam. arXiv preprint arXiv:2312.06741 (2023)", + "23. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)", + "24. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)", + "25. Mur-Artal, R., Montiel, J.M.M., Tardos, J.D.: Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics 31(5), 1147-1163 (2015)", + "26. Mur-Artal, R., Tardós, J.D.: Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras. IEEE transactions on robotics 33(5), 1255-1262 (2017)", + "27. Newcombe, R.A., Izadi, S., Hilliges, O., Molyneaux, D., Kim, D., Davison, A.J., Kohi, P., Shotton, J., Hodges, S., Fitzgibbon, A.: Kinectfusion: Real-time dense surface mapping and tracking. In: 2011 10th IEEE international symposium on mixed and augmented reality. pp. 127-136. IEEE (2011)", + "28. Nieto, J., Bailey, T., Nebot, E.: Recursive scan-matching slam. Robotics and Autonomous systems 55(1), 39-49 (2007)", + "29. Palieri, M., Morrell, B., Thakur, A., Ebadi, K., Nash, J., Chatterjee, A., Kanellakis, C., Carlone, L., Guaragnella, C., Agha-Mohammadi, A.a.: Locus: A multi-sensor lidar-centric solution for high-precision odometry and 3d mapping in real-time. IEEE Robotics and Automation Letters 6(2), 421-428 (2020)", + "30. Prisacariu, V.A., Kähler, O., Golodetz, S., Sapienza, M., Cavallari, T., Torr, P.H., Murray, D.W.: Infinitam v3: A framework for large-scale 3d reconstruction with loop closure. arXiv preprint arXiv:1708.00783 (2017)", + "31. Qin, T., Li, P., Shen, S.: Vins-mono: A robust and versatile monocular visual-inertial state estimator. IEEE Transactions on Robotics 34(4), 1004-1020 (2018)", + "32. Sandström, E., Li, Y., Van Gool, L., Oswald, M.R.: Point-slam: Dense neural point cloud-based slam. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18433-18444 (2023)", + "33. Segal, A., Haehnel, D., Thrun, S.: Generalized-icp. In: Robotics: science and systems. vol. 2, p. 435. Seattle, WA (2009)", + "34. Serafin, J., Grisetti, G.: Nicp: Dense normal based point cloud registration. In: 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 742-749. IEEE (2015)", + "35. Shan, T., Englot, B.: Lego-loam: Lightweight and ground-optimized lidar odometry and mapping on variable terrain. In: 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 4758-4765. IEEE (2018)", + "36. Shan, T., Englot, B., Meyers, D., Wang, W., Ratti, C., Rus, D.: Lio-sam: Tightly-coupled lidar inertial odometry via smoothing and mapping. In: 2020 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 5135-5142. IEEE (2020)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Seongbo Ha et al.", + "bbox": [ + 271, + 114, + 379, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "37. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019)", + "38. Sturm, J., Engelhard, N., Endres, F., Burgard, W., Cremers, D.: A benchmark for the evaluation of rgb-d slam systems. In: 2012 IEEE/RSJ international conference on intelligent robots and systems. pp. 573-580. IEEE (2012)", + "39. Sucar, E., Liu, S., Ortiz, J., Davison, A.J.: imap: Implicit mapping and positioning in real-time. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6229-6238 (2021)", + "40. Taud, H., Mas, J.: Multilayer perceptron (mlp). Geomatic approaches for modeling land change scenarios pp. 451-455 (2018)", + "41. Vizzo, I., Guadagnino, T., Mersch, B., Wiesmann, L., Behley, J., Stachniss, C.: Kiss-icp: In defense of point-to-point icp-simple, accurate, and robust registration if done the right way. IEEE Robotics and Automation Letters 8(2), 1029–1036 (2023)", + "42. Whelan, T., Johannsson, H., Kaess, M., Leonard, J.J., McDonald, J.: Robust real-time visual odometry for dense rgb-d mapping. In: 2013 IEEE International Conference on Robotics and Automation. pp. 5724-5731. IEEE (2013)", + "43. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528 (2023)", + "44. Yan, C., Qu, D., Wang, D., Xu, D., Wang, Z., Zhao, B., Li, X.: Gs-slam: Dense visual slam with 3d gaussian splatting. arXiv preprint arXiv:2311.11700 (2023)", + "45. Yang, X., Li, H., Zhai, H., Ming, Y., Liu, Y., Zhang, G.: Vox-fusion: Dense tracking and mapping with voxel-based neural implicit representation. In: 2022 IEEE International Symposium on Mixed and Augmented Reality (ISMAR). pp. 499-507. IEEE (2022)", + "46. Zhang, J., Singh, S.: Loam: Lidar odometry and mapping in real-time. In: Robotics: Science and systems. vol. 2, pp. 1-9. Berkeley, CA (2014)", + "47. Zhu, Z., Peng, S., Larsson, V., Xu, W., Bao, H., Cui, Z., Oswald, M.R., Pollefeys, M.: Nice-slam: Neural implicit scalable encoding for slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12786-12796 (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 549 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "RGBD GS-ICP SLAM", + "bbox": [ + 593, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/RGBD GS-ICP SLAM/794618d9-8655-4def-b2e9-3b1608333e79_model.json b/2024/RGBD GS-ICP SLAM/794618d9-8655-4def-b2e9-3b1608333e79_model.json new file mode 100644 index 0000000000000000000000000000000000000000..75bfd7509416a1ea38533d15802bd222a12dce96 --- /dev/null +++ b/2024/RGBD GS-ICP SLAM/794618d9-8655-4def-b2e9-3b1608333e79_model.json @@ -0,0 +1,2104 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.388, + 0.143, + 0.616, + 0.161 + ], + "angle": 0, + "content": "RGBD GS-ICP SLAM" + }, + { + "type": "text", + "bbox": [ + 0.339, + 0.192, + 0.665, + 0.206 + ], + "angle": 0, + "content": "Seongbo Ha, Jiung Yeon, and Hyeonwoo Yu" + }, + { + "type": "text", + "bbox": [ + 0.358, + 0.219, + 0.644, + 0.246 + ], + "angle": 0, + "content": "Sungkyunkwan University, Suwon, South Korea \n{sobo3607, wcr12st, hwyu}@skku.edu" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.288, + 0.734, + 0.495 + ], + "angle": 0, + "content": "Abstract. Simultaneous Localization and Mapping (SLAM) with dense representation plays a key role in robotics, Virtual Reality (VR), and Augmented Reality (AR) applications. Recent advancements in dense representation SLAM have highlighted the potential of leveraging neural scene representation and 3D Gaussian representation for high-fidelity spatial representation. In this paper, we propose a novel dense representation SLAM approach with a fusion of Generalized Iterative Closest Point (G-ICP) and 3D Gaussian Splatting (3DGS). In contrast to existing methods, we utilize a single Gaussian map for both tracking and mapping, resulting in mutual benefits. Through the exchange of covariances between tracking and mapping processes with scale alignment techniques, we minimize redundant computations and achieve an efficient system. Additionally, we enhance tracking accuracy and mapping quality through our keyframe selection methods. Experimental results demonstrate the effectiveness of our approach, showing incredibly fast speeds up to 107 FPS (for the entire system) and superior quality of the reconstructed map." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.496, + 0.757, + 0.523 + ], + "angle": 0, + "content": "The code is available at: https://github.com/Lab-of-AI-and-Robotics/GS-ICP-SLAM Video is: https://youtu.be/ebHh_uMMxE" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.538, + 0.66, + 0.551 + ], + "angle": 0, + "content": "Keywords: Coordinate-based 3D Representation \\(\\cdot\\) G-ICP \\(\\cdot\\) SLAM" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.58, + 0.357, + 0.594 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.613, + 0.787, + 0.719 + ], + "angle": 0, + "content": "Visual Simultaneous Localization and Mapping (SLAM) is an algorithm that constructs maps of unknown environments while localizing poses of vision sensors simultaneously. As the influence of 3D visual SLAM in the fields of robotics, Virtual Reality (VR), and Augmented Reality (AR) has increased, higher rendering performance and more accurate trajectory are required. Thus, 3D reconstruction methods such as Signed Distance Field (SDF), and Truncated Signed Distance Field (TSDF) are utilized for traditional dense visual SLAM [3,5,10,27,30,42]." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.72, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Recently, coordinate-based 3D Implicit Neural Representation (INR) [23] has been proposed for representing spatial information using neural radiance fields, showcasing high novel view synthesis capabilities and high-fidelity spatial representation prowess. Various approaches [8, 18, 32, 39, 45, 47] have been attempted to utilize INR for the real-time SLAM mapping process. However, INR requires computationally intensive raycasting to synthesize images, thus the rendering process in INR-based SLAM incurs significant time overhead, slowing down map optimization and rendering-loss based tracking. In contrast, coordinate-based 3D explicit representation such as 3D Gaussian" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.38, + 0.129 + ], + "angle": 0, + "content": "Seongbo Ha et al." + }, + { + "type": "image", + "bbox": [ + 0.317, + 0.153, + 0.672, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.322, + 0.788, + 0.392 + ], + "angle": 0, + "content": "Fig. 1: A comparison of PSNR with respect to FPS of entire system in recent research on SLAM algorithm utilizing dense representation such as neural scene representation and 3D Gaussian representation. Our method achieves state-of-the-art performance in rendering evaluation and FPS of entire system. Note that this FPS represents the overall system performance. Reported values are average of Replica 8 scenes." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.434, + 0.787, + 0.496 + ], + "angle": 0, + "content": "Splatting (3DGS) [15] represents the 3D space using 3D Gaussians as primitives, allowing for rendering speeds faster than NeRF using the rasterization method [43]. This rapid rendering capability of 3DGS results in fast optimization of 3D spatial information, making it suitable for dense representation in SLAM [12, 14, 22, 44]." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.502, + 0.788, + 0.699 + ], + "angle": 0, + "content": "Although 3DGS-based SLAM methods take advantages of the high-speed rendering, they fail to address the fundamental issue: the inability to directly utilize 3D explicit representations and the indirect tracking of 3D space through 2D image rendering. Even with the majority of current 3DGS-based SLAM [12, 14, 22, 32, 44] utilizing RGB-D data, the use of explicit representations is overlooked. For example, INR-based methods [32, 39, 47] and 3DGS-based methods [14, 22, 44] employ photometric-error based techniques, which estimate the optimal pose by iteratively minimizing the error between rendered and observed 2D images. However, due to limitations in tracking speed and performance, decoupled approaches [8, 12, 18] have been proposed incorporating well-crafted visual odometries [4, 26] into the tracking process. Although decoupled methods exhibit better performances in tracking by separating mapping and tracking processes, they require additional computational resources for the independent stages. Moreover, these methods need to store features in 3D space, unrelated to the 3DGS map." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.705, + 0.788, + 0.84 + ], + "angle": 0, + "content": "There exists a method that allows Gaussian, an explicit representation, to be directly used for tracking. The well-known Generalized Iterative Closest Point (G-ICP) [17, 33] from the 3D scan matching family is simple yet efficient for fast tracking of 3D point clouds. During preprocessing, it only requires computing Gaussians for the current frame and the map. Given that the map in 3DGS utilizes Gaussians as an explicit representation in 3D space, using G-ICP for tracking allows for the direct utilization of the 3DGS map without the need for post-processing. Furthermore, the Gaussians of the current frame computed during tracking with G-ICP can also be directly utilized as an explicit representation in the 3DGS map, without additional computations." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.594, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RGBD GS-ICP SLAM" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.147, + 0.788, + 0.464 + ], + "angle": 0, + "content": "Therefore, we propose a dense representation SLAM framework that integrates G-ICP and 3DGS, allowing them to complement each other by sharing explicit representations. Unlike traditional methods, our approach actively leverages 3D information through G-ICP for tracking. The proposed method is a coupled approach that shares a single map during tracking and mapping processes, while maintaining fast tracking speed akin to decoupled approaches. Previous works using decoupled methods [8, 12, 18] require separate maps containing ORB feature information for tracking, along with additional resources and computations needed to obtain ORB features. Moreover, the values obtained during tracking are not utilized in mapping. Our approach exploits the covariance of each point computed during the G-ICP-based tracking process as the initial state of the 3DGS mapping. Meanwhile, the 3D Gaussians in the 3DGS map are also mutually used as 3D points and their covariances which are essential for G-ICP-based tracking, discarding the necessity to recalculate the covariances of the map. This is possible because the covariances of points computed during the G-ICP process and the 3D Gaussians representing the map commonly contain information about the surrounding space. In other words, G-ICP and 3DGS can share the same Gaussian world. Therefore, our system minimizes unnecessary computations and facilitates efficient system configuration by mutually utilizing the key elements, Gaussians, between tracking and mapping processes. To ensure optimal performance in tracking and mapping sharing information between G-ICP and 3DGS, we also introduce several techniques such as scale alignment. To summarize, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.472, + 0.786, + 0.514 + ], + "angle": 0, + "content": "- We present a real-time dense representation SLAM that combines G-ICP and 3DGS, achieving extremely high speed of the entire system (up to 107 FPS) and superior quality of the map." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.517, + 0.785, + 0.544 + ], + "angle": 0, + "content": "- By incorporating G-ICP for tracking, our system utilizes 3D information actively and significantly reduces the time required for the tracking process." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.547, + 0.785, + 0.589 + ], + "angle": 0, + "content": "- Reduction of computational cost and facilitating rapid convergence of primitives of 3DGS is achieved by sharing the covariances of G-ICP and 3DGS with scale aligning techniques." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.472, + 0.786, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.614, + 0.367, + 0.629 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.644, + 0.788, + 0.84 + ], + "angle": 0, + "content": "G-ICP Scan-matching focuses on the registration of two point clouds that observe similar environments by selecting a transformation matrix that minimizes errors between two point clouds [2]. Iterative Closest Point (ICP) [1] algorithm is a widely used and influential method in the scan-matching area. ICP iteratively estimates point correspondences and finds a transformation that minimizes Euclidean distance between corresponding points, thereby optimizing the registration of point clouds. Because of simplicity and speed of ICP, there are many follow-up studies [7, 20, 33, 34, 41]. To improve robustness, Trimmed-ICP [7], proposed a method for correspondence selection. Point-to-plane ICP [20] augmented robustness and accuracy by taking point-to-plane distance as objective function. G-ICP [33] introduced probabilistic models to generalize ICP yielding notable improvements in robustness and accuracy. Scan matching is a common technique used in SLAM for pose tracking [6, 11, 21, 28, 29], and for enhanced robustness, feature-based scan matching is often employed [19, 35, 36, 46]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.117, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.38, + 0.129 + ], + "angle": 0, + "content": "Seongbo Ha et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.146, + 0.787, + 0.223 + ], + "angle": 0, + "content": "SLAM with Dense Representation SLAM with dense representation [9, 16] aims to construct maps in a dense form to enable interaction with the map in tasks such as AR, robotics, etc. To achieve this, classic approaches [3,5,10,27,30,42] creates dense maps by representing space in Signed Distance Field (SDF), and Truncated Signed Distance Field (TSDF), rather than in sparse forms such as point clouds or grids." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.224, + 0.788, + 0.39 + ], + "angle": 0, + "content": "Recently, dense representation SLAM methods represent maps by utilizing NeRF [23], which demonstrates high spatial representation capabilities have been proposed. iMAP [39], NICE-SLAM [47], Point-SLAM [32], ESLAM [13] perform tracking by optimizing camera poses by reducing errors between synthesized and observed images. These methods have limitations in tracking speed and performance. Alternatively, Orbeez-SLAM [8], vMAP [18] incorporate well-crafted visual SLAM techniques into the tracking process, making mapping and tracking operate independently. Orbeez-SLAM utilizes ORB-SLAM2 [26] for tracking and instant-ngp [24] for learning spatial information to perform mapping. In a similar context, vMAP utilizes ORB-SLAM3 [4] for tracking while employing separate Multi-layer perceptrons [40] for each object to effectively represent the entire space." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.391, + 0.788, + 0.543 + ], + "angle": 0, + "content": "3DGS represents space in explicit form using 3D gaussians as primitives. It offers a level of high-fidelity spatial representation similar to NeRF but provides significantly faster rendering speeds. To leverage this advantage to dense representation SLAM, several approaches [12, 14, 22, 44] utilizing 3DGS for space representation have been proposed. Among them, GS-SLAM [44], SPLTAM [14], Gaussian Splating SLAM [22] perform tracking using dense photometric error. While 3DGS shows significantly improved rendering performance compared to NeRF, the tracking methods based on dense photometric error still suffer from tracking speed. To tackle this problem, Photo-SLAM [12] integrates ORB-SLAM3 for tracking, in a similar manner to Orbeez-SLAM and vMAP." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.571, + 0.317, + 0.589 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.61, + 0.788, + 0.844 + ], + "angle": 0, + "content": "To mutually benefit tracking and mapping, we introduce the fusion of G-ICP and GS, each representing the tracking and mapping processes, respectively. The key insight of our approach is that covariance can be considered as a fundamental common factor for this fusion. Suppose we have a point set (point cloud) \\(\\mathcal{X} = \\{\\pmb{x}_m\\}_{m=1,\\dots,M}\\) and its corresponding covariance set \\(\\mathcal{C} = \\{C_m\\}_{m=1,\\dots,M}\\), where \\(\\pmb{x} = [x,y,z]^T\\). The covariance \\(C\\) of one 3D point \\(\\pmb{x}\\) is given by computing covariance matrix of \\(k\\)-nearest neighbors of \\(\\pmb{x}\\). Let us define \\(\\pmb{G} = \\{\\pmb{\\mathcal{X}},\\pmb{\\mathcal{C}}\\}\\) as a set of Gaussians. G-ICP aims to find a transformation \\(\\mathbf{T}\\) that maximally aligns the source Gaussians (current frame) \\(\\pmb{G}^s = \\{\\pmb{\\mathcal{X}}^s,\\pmb{\\mathcal{C}}^s\\}\\) and the target Gaussians (map) \\(\\pmb{G}^t = \\{\\pmb{\\mathcal{X}}^t,\\pmb{\\mathcal{C}}^t\\}\\). Assume we know the correspondences between \\(\\pmb{\\mathcal{X}}^s\\) and \\(\\pmb{\\mathcal{X}}^t\\) determined by nearest neighbor search. For example, we have \\(\\{\\pmb{x}_i^s\\}_{i=1,\\dots,N} \\subset \\pmb{\\mathcal{X}}^s\\), \\(\\{C_i^s\\}_{i=1,\\dots,N} \\subset \\pmb{\\mathcal{C}}^s\\) and \\(\\{\\pmb{x}_i^t\\}_{i=1,\\dots,N} \\subset \\pmb{\\mathcal{X}}^t\\), \\(\\{C_i^t\\}_{i=1,\\dots,N} \\subset \\pmb{\\mathcal{C}}^t\\), where \\(\\pmb{x}_i^s\\) is associated with \\(\\pmb{x}_i^t\\). To find the optimal transform \\(\\mathbf{T}^*\\), we exploit not the single point but the distribution of that point defined as a Gaussian distribution: \\(\\pmb{x}_i \\sim \\mathcal{N}(\\hat{\\pmb{x}}_i,C_i)\\). Let \\(d_i = \\pmb{x}_i^t - \\mathbf{T}\\pmb{x}_i^s\\) be the error term, and if we assume that there is an optimal transformation \\(\\mathbf{T}^*\\), it is clear that \\(\\pmb{x}_i^t = \\mathbf{T}^*\\pmb{x}_i^s\\) thus \\(\\hat{d}_i = 0\\). Since" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.593, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RGBD GS-ICP SLAM" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "we assume \\( \\pmb{x} \\) is a Gaussian random variable, \\( d_{i} \\) is also a Gaussian random variable as the following:" + }, + { + "type": "equation", + "bbox": [ + 0.369, + 0.2, + 0.633, + 0.259 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} d _ {i} \\sim \\mathcal {N} \\left(\\hat {d} _ {i}, C _ {i} ^ {t} + \\mathbf {T} ^ {*} C _ {i} ^ {s} \\left(\\mathbf {T} ^ {*}\\right) ^ {T}\\right) \\\\ = \\mathcal {N} \\left(\\hat {x} _ {i} ^ {t} - \\mathbf {T} ^ {*} \\hat {x} _ {i} ^ {s}, C _ {i} ^ {t} + \\mathbf {T} ^ {*} C _ {i} ^ {s} \\left(\\mathbf {T} ^ {*}\\right) ^ {T}\\right) \\\\ = \\mathcal {N} (0, C _ {i} ^ {t} + \\mathbf {T} ^ {*} C _ {i} ^ {s} (\\mathbf {T} ^ {*}) ^ {T}). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.268, + 0.784, + 0.299 + ], + "angle": 0, + "content": "To find the optimal transform \\(\\mathbf{T}^*\\) for \\(\\boldsymbol{x}^s\\) and \\(\\boldsymbol{x}^t\\), we use maximum likelihood estimation (MLE) as the following:" + }, + { + "type": "equation", + "bbox": [ + 0.341, + 0.307, + 0.786, + 0.392 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {T} ^ {*} = \\underset {\\mathbf {T}} {\\operatorname {a r g m a x}} \\prod_ {i} ^ {N} p \\left(d _ {i}\\right) = \\underset {\\mathbf {T}} {\\operatorname {a r g m a x}} \\sum_ {i} ^ {N} \\log p \\left(d _ {i}\\right) \\\\ = \\underset {\\mathbf {T}} {\\operatorname {a r g m i n}} \\sum_ {i} ^ {N} d _ {i} ^ {T} \\left(C _ {i} ^ {B} + \\mathbf {T} C _ {i} ^ {A} \\mathbf {T} ^ {T}\\right) ^ {- 1} d _ {i}. \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.399, + 0.793, + 0.429 + ], + "angle": 0, + "content": "Therefore, \\(\\mathbf{T}^*\\) can be used as the relative pose between the current frame \\(G^{s} = \\{\\mathcal{X}^{s},\\mathcal{C}^{s}\\}\\) and the map \\(G^{t} = \\{\\mathcal{X}^{t},\\mathcal{C}^{t}\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.43, + 0.787, + 0.492 + ], + "angle": 0, + "content": "Meanwhile, for mapping purposes, GS also relies on Gaussians \\( G = \\{\\mathcal{X},\\mathcal{C}\\} \\) of the 3D scene representation. Different from G-ICP, GS aims to find the optimal coordinates of the Gaussians \\( \\mathcal{X}^* = \\{x_m^*\\}_{m=1,\\dots,M} \\) and the optimal covariances \\( \\mathcal{C}^* = \\{C_m^*\\}_{m=1,\\dots,M} \\) as the following:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.498, + 0.793, + 0.524 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol{\\mathcal{X}}^{*},\\boldsymbol{\\mathcal{C}}^{*},\\boldsymbol{H}^{*},\\boldsymbol{O}^{*} = \\operatorname *{argmin}_{\\boldsymbol {\\mathcal{X}},\\boldsymbol {\\mathcal{C}},\\boldsymbol {H},\\boldsymbol{O}}\\lambda_{I_{1}}\\boldsymbol{\\mathcal{L}}_{1}\\left(I,I_{gt}\\right) + \\lambda_{I_{2}}\\boldsymbol{\\mathcal{L}}_{D - SSIM}\\left(I,I_{gt}\\right) + \\lambda_{D}\\boldsymbol{\\mathcal{L}}_{1}\\left(D,D_{gt}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.532, + 0.784, + 0.577 + ], + "angle": 0, + "content": "where \\( H = \\{h_m\\}_{m=1,\\dots,M} \\) and \\( O = \\{o_m\\}_{m=1,\\dots,M} \\) are the color set and the opacity set of 3D points which are for the RGBD image rendering. \\( I \\) and \\( D \\) are the rendered RGB and depth images obtained by performing rasterization using \\( G \\), \\( H \\) and \\( O \\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.578, + 0.786, + 0.803 + ], + "angle": 0, + "content": "Note that in G-ICP and GS, the key common factor is Gaussians \\(G = \\{\\mathcal{X},\\mathcal{C}\\}\\), allowing these Gaussians to be shared mutually. During G-ICP tracking, the covariance of each frame is computed. Hence, when adding keyframes to expand the 3DGS map, there is no need to recalculate \\(\\mathcal{C}\\) for every expansion. Also, G-ICP does not need to compute the covariances of the map because our GS map already contains Gaussians. Moreover, G-ICP, by aligning frames based on 3D geometric structure, inherently initializes a certain number of points and their coordinates that aptly represent the 3D structure. Therefore, it brings about an effect where an appropriate number of points and their poses are initialized, well-suited for depicting the 3D structure, significantly reducing the learning time to find the optimal pose \\(\\mathcal{X}^*\\) and the optimal covariance \\(\\mathcal{C}^*\\) of Gaussians \\(G\\) in GS. Additionally, in the same vein, calculations like densifying or opacity reset to adjust the number of 3D points in GS become unnecessary. Consequently, by sharing a common source, each process becomes mutually beneficial, and the speed of execution accelerates due to reduced redundant computations. Our method can be simplified as follows:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "1. Use G-ICP to align the current frame with the 3DGS map which contains covariance (solely need to compute the covariance for the current frame)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.117, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.38, + 0.129 + ], + "angle": 0, + "content": "Seongbo Ha et al." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.146, + 0.77, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.384 + ], + "angle": 0, + "content": "Fig. 2: System Overview. The input of our system is RGBD frame. We generate a point cloud by downsampling and reprojecting the current depth image and utilize it in the GICP process. During the GICP process, we create source Gaussians from the point cloud and estimate the current camera pose by aligning them with target Gaussians, which are a subset of the 3DGS map. If the current frame is identified as a keyframe or a mapping-only keyframe, we add the source Gaussians to the 3DGS map as new primitives. Meanwhile, in the mapping process, we optimize the Gaussians along with the color and opacity set of the Gaussians concurrently with the tracking process." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.411, + 0.785, + 0.441 + ], + "angle": 0, + "content": "2. When adding keyframes to the 3DGS map, utilize the covariance computed in GICP during tracking (no need for densifying or opacity reset)." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.441, + 0.361, + 0.456 + ], + "angle": 0, + "content": "3. Repeat steps 1-2." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.411, + 0.785, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.493 + ], + "angle": 0, + "content": "The overview of our system is shown in Fig. 2. The detailed implementations of the proposed technique are introduced in the following Section." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.512, + 0.373, + 0.528 + ], + "angle": 0, + "content": "3.1 G-ICP Tracking" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.788, + 0.595 + ], + "angle": 0, + "content": "Scale Regularization When using G-ICP for tracking, aligning the scale of the current frame with the map enables high-performance camera pose estimation [33]. Given the covariance \\(C\\), scale \\(\\boldsymbol{S} = [s_2, s_1, s_0]^T\\) is given by the singular value decomposition (SVD) as the following:" + }, + { + "type": "equation", + "bbox": [ + 0.451, + 0.602, + 0.787, + 0.62 + ], + "angle": 0, + "content": "\\[\nC = \\boldsymbol {R} \\boldsymbol {\\Lambda} ^ {2} \\boldsymbol {R} ^ {T} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.687 + ], + "angle": 0, + "content": "where \\(\\Lambda = \\text{diag}(s_2, s_1, s_0) \\forall s_2 > s_1 > s_0\\) is a scale matrix and \\(\\pmb{R}\\) is the orientation of the Gaussian. To achieve the robust scan matching performance, point-to-plane ICP or voxelized point-to-plane ICP adopt a regularizing method that makes the scale as \\(\\pmb{S} = [1, 1, \\epsilon]^T\\) in order to treat each Gaussian as a plane-like distribution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.788, + 0.78 + ], + "angle": 0, + "content": "In our framework, we utilize target Gaussians \\(\\pmb{G}^t = \\{\\pmb{x}^t,\\pmb{c}^t\\}\\) from the 3DGS map for tracking. Here, \\(\\pmb{c}^t\\) is optimized to represent the scene accurately, \\(\\pmb{G}^t\\) involves not only planes but also lines, corners, and other features. Therefore, while tracking, instead of regularizing the scale into a plane-based form, preserving the original characteristics of the target Gaussians while performing regularization is more appropriate. Thus, in our framework, we propose the following ellipsoid regularization:" + }, + { + "type": "equation", + "bbox": [ + 0.383, + 0.786, + 0.787, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Lambda} ^ {\\prime} = \\frac {1}{m e d i a n (\\boldsymbol {S})} \\operatorname {d i a g} \\left(s _ {2}, s _ {1}, s _ {0}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.824, + 0.547, + 0.841 + ], + "angle": 0, + "content": "where \\(\\pmb{\\Lambda}^{\\prime}\\) is the ellipsoid-regularized scale matrix." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.594, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RGBD GS-ICP SLAM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.156, + 0.481, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.353, + 0.28, + 0.369, + 0.291 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.156, + 0.768, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.279, + 0.657, + 0.29 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.302, + 0.788, + 0.359 + ], + "angle": 0, + "content": "Fig. 3: Tracking Accuracy Comparison Based on Keyframe Selection Methods. The reported values represent the average results across eight scenes from the Replica dataset [37]. When selecting keyframes every n frame (depicted in blue), the tracking accuracy is notably low. Conversely, our keyframe selection method yielded the highest tracking accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.789, + 0.63 + ], + "angle": 0, + "content": "Keyframe Selection Similar to [25,31], we perform dynamic keyframe selection. Considering the geometric structure is critical for both G-ICP tracking and GS mapping, we exploit the geometric correspondence computed from G-ICP. As Eqn. (1) denotes, we can get the distance between \\( \\boldsymbol{x}_i^s \\) and \\( \\boldsymbol{x}_i^t \\) as an interim result of G-ICP procedure without additional calculations. So we can efficiently determine the correspondence by setting the distance threshold. By considering the proportion of correspondences between the current frame and map, we select keyframes while taking into account the characteristics of the scene. In more detail, if the proportion in the current frame goes below specific thresholds, the frame is selected as a keyframe. This allows us to achieve consistent tracking performance while simultaneously maintaining a consistent density of Gaussians \\( G = \\{\\mathcal{X},\\mathcal{C}\\} \\) added to the map. The effect of this keyframe selection is shown in Fig. 3 (a). Note that, from selected keyframe, only Gaussians that do not overlap with the existing current map are considered as target Gaussians. Errors from discrepancies during tracking are directly incorporated into the map, leading to accumulated inaccuracies and reduced tracking precision. This strategy ensures more accurate tracking results and improves the overall robustness of the system." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.65, + 0.348, + 0.666 + ], + "angle": 0, + "content": "3.2 GS Mapping" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Scale Aligning Suppose we have a fully trained 3DGS map. Since the 3DGS map is trained to represent the scene from any viewpoint, Gaussians \\( G \\) should be uniformly distributed according to the structure of the 3D space. However, single frames obtained from radiance sensors such as RGBD cameras or LiDAR yield sparse representations of the 3D space as the distance from the sensor increases, due to the inherent characteristics of the sensor. In other words, as the distance from the sensor increases, the spacing between 3D points widens, and the scale of the corresponding covariance calculated based on the \\( k \\)-nearest neighbors becomes excessively large. Adding such single frames as keyframes to the map during real-time training of GS mapping leads to an imbalance in the scale of Gaussians, ultimately resulting in a decrease in mapping performance. Therefore, to fundamentally alleviate the problem, our proposed method suggests the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.38, + 0.129 + ], + "angle": 0, + "content": "Seongbo Ha et al." + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.148, + 0.396, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.148, + 0.58, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.147, + 0.776, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.288, + 0.79, + 0.346 + ], + "angle": 0, + "content": "Fig. 4: Separated Keyframe Selection on Replica office4. We demonstrate that a small number of tracking keyframes yield accurate trajectory estimation (case 1), while a large number of mapping keyframes result in high rendering performance (case 2). Thus, our method adopts case 3 that selects tracking keyframe and mapping keyframe separately at different intervals." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.371, + 0.561, + 0.385 + ], + "angle": 0, + "content": "following scale normalization for the current frame:" + }, + { + "type": "equation", + "bbox": [ + 0.458, + 0.392, + 0.546, + 0.422 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\varLambda} ^ {\\prime \\prime} = \\frac {1}{z ^ {p}} \\boldsymbol {\\varLambda}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.428, + 0.789, + 0.492 + ], + "angle": 0, + "content": "Where \\(\\Lambda''\\) is the normalized scale matrix and \\(p\\) is the parameter which is empirically determined. By adding such scale-normalized keyframes to the map to expand it, not only does the GS mapping performance increase, but also the performance of G-ICP tracking improves." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.789, + 0.644 + ], + "angle": 0, + "content": "Additional Keyframe Selection for Mapping Mapping based on GS benefits from having a diverse set of frames from various viewpoints for training, as performance improves with a larger number of available frames. However, increasing the number of selected keyframes during tracking can lead to a degradation in tracking performance due to accumulated errors, as is common with G-ICP-based methods. To mitigate this issue, we propose adding mapping-only keyframe selection in addition to the existing keyframe selection process. In other words, tracking continues to proceed from selected keyframes in the existing tracking process, while mapping utilizes both the original keyframes and additional mapping-only keyframes." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.645, + 0.79, + 0.841 + ], + "angle": 0, + "content": "The overview of the additional keyframe selection is illustrated in Fig. 4. Cases 1 and 2 represent scenarios without considering mapping-only keyframes. In Case 1, to increase the number of keyframes, if no keyframe satisfying the criteria is found up to the \\(30^{th}\\) frame from the previous keyframe, the \\(30^{th}\\) frame is selected as a keyframe. However, since this may not yield sufficient mapping performance, as shown in Case 2, adding an additional keyframe at the \\(10^{th}\\) frame can be considered. While this approach improves mapping performance as desired, it leads to a decline in tracking performance due to accumulated scan matching errors. To alleviate this, as shown in Case 3, we propose adding the \\(10^{th}\\) frame as a mapping-only frame, allowing us to maximize tracking performance while simultaneously maximizing mapping performance. The effect of the proposed scheme, Case 3, is shown in Fig. 3 (b). Note that since G-ICP, which performs tracking, computes covariance for all frames anyway, there is no difference in the overall computational load." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.594, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RGBD GS-ICP SLAM" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "Avoiding Local Minima while Mapping GS SLAM differs from traditional GS in that it operates in real-time and often lacks sufficient observations of the scene. Moreover, since it operates in real-time, focusing solely on images from the current viewpoint or nearby viewpoints during training can easily lead to local minima and degrade mapping quality [22]. For instance, GS simply considers image rendering from specific viewpoints, thus if we continuously train only on a specific viewpoint, the scale of Gaussians in the map tends to elongate in the direction of that viewpoint, leading to local minima." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.254, + 0.788, + 0.329 + ], + "angle": 0, + "content": "To overcome this challenge, at each training iteration, we adopt a strategy of randomly choosing one keyframe for learning among the selected keyframes so far, ensuring that both the current observed scene and the entire map are uniformly learned. Additionally, we prune Gaussians that fall into local minima during training to improve mapping performance and ensure robust tracking through geometric preservation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.354, + 0.357, + 0.372 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.387, + 0.398, + 0.403 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.789, + 0.566 + ], + "angle": 0, + "content": "We evaluate our method on Replica dataset [37] and TUM dataset [38]. Replica dataset contains synthetic scenes and high-quality RGB/depth images rendered from these scenes. TUM dataset includes images captured in the real world, with significant noise and blur, resulting in poor quality. Particularly, many parts of the depth images suffer from information loss. We validate the effectiveness of our approach by evaluating it on both synthetic and real-world datasets. All experiments are performed on a desktop with a Ryzen 7 7800x3d CPU, 32GB RAM, and an NVIDIA RTX 4090 24GB GPU. To evaluate camera tracking accuracy, we use the Root Mean Square Error (RMSE) of the Absolute Trajectory Error (ATE). For the quality of the reconstructed map, we report standard photometric rendering quality metrics (PSNR, SSIM, and LPIPS)." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.592, + 0.789, + 0.661 + ], + "angle": 0, + "content": "Table 1: Tracking Performance on Replica and TUM-RGBD (ATE RMSE \\(\\downarrow\\) [cm]). Our method archives state-of-the-art tracking accuracy on Replica. On TUM, Among coupled methods, we achieve competitive performance. The results of GS-SLAM/Gaussian Splitting SLAM are taken from [44]/ [22]. And the results of ORB-SLAM3 and Photo-SLAM are obtained from [12]." + }, + { + "type": "table", + "bbox": [ + 0.224, + 0.675, + 0.784, + 0.797 + ], + "angle": 0, + "content": "
MethodReplicaTUM
R0R1R2Of0Of1Of2Of3Of4Avg.fr1/deskfr2/xyzfr3/officeAvg.
DecoupledORB-SLAM3 [4]-------1.81.70.41.71.3
Photo-SLAM [12]-------0.602.60.31.01.3
CoupledNICE-SLAM* [47]1.611.481.610.950.811.461.761.691.422.82.17.24.0
Point-SLAM* [32]0.590.510.320.450.460.480.610.870.542.71.33.92.6
GS-SLAM [44]0.480.530.330.520.410.590.460.700.503.31.36.63.7
SplaTAM* [14]0.290.350.280.490.210.310.340.570.363.31.35.13.2
Gaussian Splitting SLAM* [22]0.330.220.290.360.190.250.120.810.321.51.41.51.5
Ours (limited to 30 FPS)0.150.160.110.180.120.170.160.210.162.71.82.72.4
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.52, + 0.797, + 0.779, + 0.807 + ], + "angle": 0, + "content": "* denotes the reproduced results by running official code." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.129 + ], + "angle": 0, + "content": "Seongbo Ha et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.453, + 0.163 + ], + "angle": 0, + "content": "4.2 Camera Tracking Accuracy" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.79, + 0.364 + ], + "angle": 0, + "content": "Tab. 1 shows the tracking accuracy of our method compared to other approaches on both synthetic [37] and real-world [38] datasets. On Replica [37], the proposed method achieved state-of-the-art (SOTA) performance across all scenes, reducing the trajectory error by more than \\(50\\%\\) compared to the previous SOTA. This result arises from the fact that our proposed method actively utilizes 3D information by employing G-ICP [33] for tracking, unlike other methods that perform tracking based on errors in 2D space. On TUM [38], our method shows competitive tracking accuracy among coupled methods [14,22,32,44,47] which conduct tracking and mapping based on a single map. Note that the results of Gaussian Splitting SLAM are evaluated only on keyframes. Decoupled method [12] and its baseline [4] show better performance, but these approaches require additional resources for storing separate map information and costly computations for extracting features solely for the tracking process." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.392, + 0.79, + 0.422 + ], + "angle": 0, + "content": "Table 2: Evaluation of System Speed and Map Quality on Replica. Our method outperforms all other frameworks in both system speed and quality of the reconstructed map." + }, + { + "type": "table", + "bbox": [ + 0.236, + 0.434, + 0.773, + 0.813 + ], + "angle": 0, + "content": "
MethodsMetricsR0R1R2Of0Of1Of2Of3Of4Avg.
Orbeez-SLAM* [8]PSNR[dB]↑12.1315.2815.8717.5919.2610.3011.5512.6514.33
SSIM ↑0.700.790.810.800.670.780.760.840.77
LPIPS ↓0.610.480.510.520.410.550.540.530.53
FPS ↑24.0524.7824.0019.2226.7423.8824.4526.1124.15
Point-SLAM* [32]PSNR[dB]↑33.3834.1036.3238.7239.3134.2234.1034.8235.62
SSIM ↑0.980.980.990.990.990.960.960.980.98
LPIPS ↓0.100.120.100.090.110.150.120.130.11
FPS ↑0.260.300.310.330.340.300.280.300.30
GS-SLAM [44]PSNR[dB]↑31.5632.8632.5938.7041.1732.3632.0332.9234.27
SSIM ↑0.970.970.970.990.990.980.970.970.98
LPIPS ↓0.090.080.090.050.030.090.110.110.08
FPS ↑8.34-------8.34
SplaTAM* [14]PSNR[dB]↑32.6033.5534.8338.0939.0231.9529.5331.5533.89
SSIM ↑0.980.970.980.980.980.970.950.950.97
LPIPS ↓0.070.100.070.090.090.100.120.150.10
FPS ↑0.240.190.190.200.220.270.260.240.23
Gaussian Splitting SLAM [22]PSNR[dB]↑34.8336.4337.4939.9542.0936.2436.7036.0737.50
SSIM ↑0.950.960.970.970.980.960.960.960.96
LPIPS ↓0.070.080.080.070.060.080.070.100.08
FPS ↑-----1.1---
Ours (no tracking speed limit)PSNR[dB]↑32.2035.3634.4240.3140.7533.8534.0836.4735.93
SSIM ↑0.940.960.960.980.980.960.950.960.96
LPIPS ↓0.080.070.080.050.050.070.070.070.07
FPS ↑100.9884.92103.3899.10107.0695.6097.2096.6698.11
Ours (limited to 30 FPS)PSNR[dB]↑35.3737.8038.5043.1343.2636.9336.9038.7538.83
SSIM ↑0.960.970.980.990.990.970.970.970.98
LPIPS ↓0.050.050.050.030.030.040.040.050.04
FPS ↑29.9729.9829.9829.9829.9929.9729.9729.9729.98
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.352, + 0.814, + 0.77, + 0.838 + ], + "angle": 0, + "content": "* denotes the reproduced results by running official code. The result of GS-SLAM and Gaussian Splitting SLAM is taken from [44] and [22]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.594, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RGBD GS-ICP SLAM" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 3: Evaluation of System Speed and Map Quality on TUM-RGBD. Proposed method shows incredible system speed and competitive map quality." + }, + { + "type": "table", + "bbox": [ + 0.317, + 0.185, + 0.69, + 0.282 + ], + "angle": 0, + "content": "
MethodsPSNR[dB] ↑SSIM ↑LPIPS ↓FPS ↑
NICE-SLAM* [47]14.100.5740.3950.08
Point-SLAM* [32]21.400.7380.4470.22
Photo-SLAM [12]21.900.7630.187-
SplaTAM* [14]23.460.9060.1560.32
Ours (unlimited tracking speed)19.620.7500.24073.92
Ours (limited to 30 FPS)20.720.7680.21829.99
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.396, + 0.283, + 0.687, + 0.305 + ], + "angle": 0, + "content": "* denotes the reproduced results by running official code. The result of Photo-SLAM is taken from [12]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.364, + 0.599, + 0.38 + ], + "angle": 0, + "content": "4.3 System Speed and Quality of Reconstructed Map" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.788, + 0.576 + ], + "angle": 0, + "content": "Tab. 2 and Tab. 3 demonstrate the FPS of the systems and quality of the reconstructed map on Replica [37] and TUM [38] dataset, respectively. FPS of the system is calculated by dividing the total number of frames by the total time. Note that this FPS represents the overall system performance including tracking and mapping processes, rather than just individual components. Photo-SLAM [12] reported the tracking FPS of their system, but did not include the system FPS, so we leave that field blank in the Tab. 3. Since our method implements mapping and tracking processes to operate in parallel, the number of map optimization iterations varies depending on the tracking speed, consequently affecting the quality of the reconstructed map. So we evaluate our system in two cases: one where the tracking speed is limited to the typical sensor input speed of 30 FPS, and the other where the tracking speed is not limited. Please refer to the supplementary material for results across various FPS and geometric quality evaluations." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.616, + 0.784, + 0.757 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.767, + 0.788, + 0.837 + ], + "angle": 0, + "content": "Fig. 5: Comparison of Rendering Results. In the first scene, SplaTAM [14] failed to reconstruct the pillow and lamp and Point-SLAM [32] failed to represent the detailed pattern of the pillow. In the second case, SplaTAM and Point-SLAM failed to accurately reconstruct the details of the clock. However, our method exhibits rendering results that closely resemble the ground truth image, demonstrating a high level of accuracy." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.129 + ], + "angle": 0, + "content": "Seongbo Ha et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.451 + ], + "angle": 0, + "content": "In the limited case (30 FPS), our system achieves SOTA map quality across all scenes in the Replica dataset. When the tracking speed is not limited, our method shows extremely fast system speed up to 107 FPS. This speed is more than four times faster than that of the Orbeez-SLAM [8], based on the ORB-SLAM2 [26]. Remarkably, even under this condition, our approach maintains superior map quality and outperforms other methods. The proposed method effectively integrates Gaussians for map representation with appropriate initial states and has better detailed representation capability compared to NeRF-based methods. As shown in Fig. 5, SplaTAM suffers from incorporating appropriate Gaussians into the map, leading to regions of poor quality. Point-SLAM exhibits areas where detailed representation is lacking. In TUM-RGBD, our method demonstrates remarkably fast system speed with competitive map quality, showcasing its practicality in real-world scenarios. Compared to SplaTAM [14], our approach shows a slight decrease in map quality, with a PSNR reduction of approximately \\(11.7\\%\\). However, we achieve significant speed improvements, with a speedup of approximately 91.6 times under limited conditions and up to 227 times at maximum speed. The reason is that, unlike Replica, TUM dataset is captured with old-fashioned sensors, resulting in noisy and substantial information loss in depth images. While SplaTAM treats these factors by adding new Gaussians only in regions with significant depth loss, our method utilizes all structural information from the depth image, focusing on accurate tracking and rapid system rather than treating such factors." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.469, + 0.36, + 0.485 + ], + "angle": 0, + "content": "4.4 Ablation study" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.519, + 0.788, + 0.547 + ], + "angle": 0, + "content": "Table 4: Scale Regularization Ablation on TUM-RGBD. Reported results are the average ATE RMSE \\(\\downarrow\\) [cm] of 3 scenes in TUM-RGBD dataset." + }, + { + "type": "table", + "bbox": [ + 0.333, + 0.56, + 0.671, + 0.632 + ], + "angle": 0, + "content": "
Scale Regularizationfr1-deskfr2-xyzfr3-officeAvg.
X136.32196.61376.40236.54
Plane65.635.5316.2129.12
ellipsoid2.661.772.672.37
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.825 + ], + "angle": 0, + "content": "Scale Regularization Tab. 4 shows the result of the ablation study on scale regularization while tracking. We perform scale regularization to enhance tracking accuracy by ensuring that the scales of Gaussians in the current frame and the map are within a similar range. Despite the application of plane regularization resulting in some improvement in tracking accuracy, it remains inadequate. This is because the plane regularization method considers all the Gaussians as planes, ignoring the characteristics of Gaussians existing in the map, which are optimized to represent the surrounding space through the mapping process. On the other hand, when applying the proposed ellipsoid regularization, the best tracking accuracy is achieved. This is attributed to the ability of the proposed ellipsoid regularization to consider the characteristics of Gaussians existing in the map while performing regularization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.825, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Scale Aligning We implement scale aligning to reduce the difference between scales" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.594, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RGBD GS-ICP SLAM" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 5: Ablation of Scale Aligning on Replica. The results are the average of 8 scenes in the Replica dataset." + }, + { + "type": "table", + "bbox": [ + 0.348, + 0.185, + 0.66, + 0.259 + ], + "angle": 0, + "content": "
Covs from G-ICPRe-initATE ↓PSNR ↑SSIM ↑LPIPS ↓
XX8.89324.810.8400.380
✓X0.25833.210.9220.177
✓constant0.15837.330.9640.074
✓z1.50.15738.830.9750.041
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.287, + 0.788, + 0.513 + ], + "angle": 0, + "content": "of existing Gaussians in the map and newly added Gaussians. By doing this, we ensure that scales of newly added Gaussians closely resemble the optimal value they will reach through optimization. To demonstrate the effectiveness of this approach, we tested our method in various cases and Tab. 5 shows the results. When not utilizing the covariances computed during the G-ICP [33] process, we calculate the scales of Gaussians by using simple-knn module of vanilla 3DGS [15], and set rotations as identity. In this case, a significant drop in tracking performance appeared. This is because the newly added Gaussians lack sufficient prior information about the 3D structure, causing them to be utilized as target Gaussians without being optimally optimized in space. In the case of using the computed covariance from the G-ICP process, the tracking and mapping performance improves. Furthermore, the best performance is observed when dividing the scale by \\( z^{1.5} \\). These results indicate that utilizing covariances calculated during G-ICP procedure is effective for both tracking and mapping, and scale aligning facilitates smooth connection from G-ICP to 3DGS. Further analysis of the advantages of this connection is included in the supplementary material." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.54, + 0.785, + 0.568 + ], + "angle": 0, + "content": "Table 6: Keyframe and Mapping-Only Keyframe Selection Ablation on Replica. Reported results are the average of 8 scenes in the Replica dataset." + }, + { + "type": "table", + "bbox": [ + 0.276, + 0.58, + 0.731, + 0.647 + ], + "angle": 0, + "content": "
KeyframeMapping-Only KeyframeATE [cm] ↓PSNR ↑SSIM ↑LPIPS ↓
tracking + every 30 frame-0.17437.520.9680.055
tracking + every 10 frame-0.17638.830.9740.041
trackingevery 10 frame0.15738.830.9750.041
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.79, + 0.841 + ], + "angle": 0, + "content": "Additional Keyframe Selection for Mapping For mapping quality, sufficient keyframes are required, but indiscriminate addition of keyframes causes a drop in tracking accuracy. To address this challenge and maximize both tracking accuracy and mapping quality, we pick mapping-only keyframes in addition to the original keyframes. Tab. 6 presents the result of the ablation study of this method. When using only keyframes without mapping-only keyframes, camera tracking accuracy suffers due to accumulated errors in scan-matching. However, employing our proposed method yields the best tracking accuracy and mapping quality. This shows that our method selectively adds keyframes essential for tracking, minimizing accumulated errors in scan-matching. While leveraging these benefits, we ensure sufficient primitives and images for training, which are crucial for map quality through the use of mapping-only keyframes." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.129 + ], + "angle": 0, + "content": "Seongbo Ha et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.175 + ], + "angle": 0, + "content": "Table 7: Ablation of Methods for Avoiding Local Minima While Mapping on Replica. The results are the average of 8 scenes in the Replica dataset." + }, + { + "type": "table", + "bbox": [ + 0.219, + 0.196, + 0.785, + 0.286 + ], + "angle": 0, + "content": "
(a)(b)
Keyframe choice for learningPSNRSSIMLPIPSdensifying pruningATE [cm]PSNRSSIMLPIPS
recent 1 keyframe26.850.8720.200✓✓0.18838.650.973
covisible keyframes31.340.9240.120✓✗0.17037.900.971
random keyframes38.830.9750.041✗✓0.15738.830.975
✗✗0.15938.720.974
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.79, + 0.542 + ], + "angle": 0, + "content": "Avoiding Local Minima while Mapping We employ two methods to prevent Gaussians from overfitting to the training image, which results in elongation to the viewpoint direction: (1) training with randomly selected keyframes and (2) Gaussian pruning. Tab. 7 presents the results of the ablation study on these methods. The worst rendering performance is observed when repeatedly training with the recently added keyframe, while using covisible keyframes for training yields better results. Covisible keyframes offer more diverse viewpoints for the Gaussians compared to a single keyframe, yet they still provide limited diversity in viewpoints. Random keyframes provide the most diverse viewpoints compared to other methods, leading to a significant improvement in map quality. In the vanilla 3DGS [15], Gaussian densifying/pruning techniques are utilized to manage Gaussians. Tab. 7 (b) shows the results of an ablation test of Gaussian densifying/pruning in our method. Since our method appropriately supplies Gaussians for scene representation, densifying is unnecessary. However, pruning improves tracking accuracy and map quality by removing Gaussians that have become elongated due to overfitting or are no longer essential for representing the map." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.565, + 0.347, + 0.583 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.789, + 0.735 + ], + "angle": 0, + "content": "In this paper, we have proposed RGBD GS-ICP-SLAM, a dense representation SLAM system that leverages 3D Gaussian representation for high-fidelity spatial representation. We demonstrate that a fusion of G-ICP and 3DGS that utilizes a single 3D Gaussian map for both tracking and mapping yields mutual benefits. The exchange of Gaussians between tracking and mapping processes with scale alignments minimizes redundant computations and constructs an efficient system. Moreover, our dynamic keyframe selection method enhances both tracking and mapping performance. Through extensive experiments, the proposed approach presents state-of-the-art performance in spatial representation, camera pose estimation, and total system speed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.789, + 0.81 + ], + "angle": 0, + "content": "Limitations. The proposed method achieved fast system speed by relying solely on depth for the 3D structure. In real-world, there are limitations in the quality of maps due to the depth noise in RGB-D sensors. Since the system speed is already exceptionally fast, it is expected that robust performance can be achieved by trading off a bit of speed to compensate for noisy depth images with relatively robust RGB information." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Acknowledgment. This work was supported by the National Research Foundation of Korea (NRF) grant (no. RS-2024-00359937)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.594, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RGBD GS-ICP SLAM" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.312, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.208 + ], + "angle": 0, + "content": "1. Besl, P.J., McKay, N.D.: Method for registration of 3-d shapes. In: Sensor fusion IV: control paradigms and data structures. vol. 1611, pp. 586-606. Spie (1992)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.209, + 0.785, + 0.25 + ], + "angle": 0, + "content": "2. Biber, P., Straβer, W.: The normal distributions transform: A new approach to laser scan matching. In: Proceedings 2003 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2003)(Cat. No. 03CH37453). vol. 3, pp. 2743-2748. IEEE (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.251, + 0.785, + 0.292 + ], + "angle": 0, + "content": "3. Bylow, E., Sturm, J., Kerl, C., Kahl, F., Cremers, D.: Real-time camera tracking and 3d reconstruction using signed distance functions. In: Robotics: Science and Systems. vol. 2, p. 2 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.293, + 0.785, + 0.334 + ], + "angle": 0, + "content": "4. Campos, C., Elvira, R., Rodríguez, J.J.G., Montiel, J.M., Tardós, J.D.: Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics 37(6), 1874-1890 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.334, + 0.785, + 0.376 + ], + "angle": 0, + "content": "5. Canelhas, D.R., Stoyanov, T., Lilienthal, A.J.: Sdf tracker: A parallel algorithm for on-line pose estimation and scene reconstruction from depth images. In: 2013 IEEE/RSJ International Conference on Intelligent Robots and Systems. pp. 3671-3676. IEEE (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.377, + 0.785, + 0.418 + ], + "angle": 0, + "content": "6. Ceriani, S., Sánchez, C., Taddei, P., Wolfart, E., Sequeira, V.: Pose interpolation slam for large maps using moving 3d sensors. In: 2015 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 750-757. IEEE (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.42, + 0.785, + 0.46 + ], + "angle": 0, + "content": "7. Chetverikov, D., Svirko, D., Stepanov, D., Krsek, P.: The trimmed iterative closest point algorithm. In: 2002 International Conference on Pattern Recognition. vol. 3, pp. 545-548. IEEE (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.461, + 0.785, + 0.516 + ], + "angle": 0, + "content": "8. Chung, C.M., Tseng, Y.C., Hsu, Y.C., Shi, X.Q., Hua, Y.H., Yeh, J.F., Chen, W.C., Chen, Y.T., Hsu, W.H.: Orbeez-slam: A real-time monocular visual slam with orb features and nef-realized mapping. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 9400-9406. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.517, + 0.785, + 0.544 + ], + "angle": 0, + "content": "9. Czarnowski, J., Laidlow, T., Clark, R., Davison, A.J.: Deepfactors: Real-time probabilistic dense monocular slam. IEEE Robotics and Automation Letters 5(2), 721-728 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.546, + 0.785, + 0.587 + ], + "angle": 0, + "content": "0. Dai, A., Nießner, M., Zollhöfer, M., Izadi, S., Theobalt, C.: Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics (ToG) 36(4), 1 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.588, + 0.785, + 0.615 + ], + "angle": 0, + "content": "1. Grisetti, G., Kümmerle, R., Stachniss, C., Burgard, W.: A tutorial on graph-based slam. IEEE Intelligent Transportation Systems Magazine 2(4), 31-43 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.616, + 0.785, + 0.657 + ], + "angle": 0, + "content": "2. Huang, H., Li, L., Cheng, H., Yeung, S.K.: Photo-slam: Real-time simultaneous localization and photorealistic mapping for monocular, stereo, and rgb-d cameras. arXiv preprint arXiv:2311.16728 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.658, + 0.785, + 0.699 + ], + "angle": 0, + "content": "3. Johari, M.M., Carta, C., Fleuret, F.: Eslam: Efficient dense slam system based on hybrid representation of signed distance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17408-17419 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.7, + 0.785, + 0.741 + ], + "angle": 0, + "content": "4. Keetha, N., Karhade, J., Jatavallabhula, K.M., Yang, G., Scherer, S., Ramanan, D., Luiten, J.: Splatam: Splat, track & map 3d gaussians for dense rgb-d slam. arXiv preprint arXiv:2312.02126 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.742, + 0.785, + 0.77 + ], + "angle": 0, + "content": "5. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.785, + 0.798 + ], + "angle": 0, + "content": "6. Kerl, C., Sturm, J., Cremers, D.: Dense visual slam for rgb-d cameras. In: 2013 IEEE/RSJ International Conference on Intelligent Robots and Systems. pp. 2100-2106. IEEE (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Koide, K., Yokozuka, M., Oishi, S., Banno, A.: Voxelized gicp for fast and accurate 3d point cloud registration. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 11054-11059. IEEE (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.38, + 0.129 + ], + "angle": 0, + "content": "Seongbo Ha et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.189 + ], + "angle": 0, + "content": "18. Kong, X., Liu, S., Taher, M., Davison, A.J.: vmap: Vectorised object mapping for neural field slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 952-961 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "19. Lin, J., Zhang, F.: Loam livox: A fast, robust, high-precision lidar odometry and mapping package for lidars of small fov. In: 2020 IEEE International Conference on Robotics and Automation (ICRA). pp. 3126-3131. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.786, + 0.259 + ], + "angle": 0, + "content": "20. Low, K.L.: Linear least-squares optimization for point-to-plane 1cp surface registration. Chapel Hill, University of North Carolina 4(10), 1-3 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.786, + 0.285 + ], + "angle": 0, + "content": "21. Mallios, A., Ridao, P., Ribas, D., Hernández, E.: Scan matching slam in underwater environments. Autonomous Robots 36, 181-198 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.786, + 0.313 + ], + "angle": 0, + "content": "22. Matsuki, H., Murai, R., Kelly, P.H., Davison, A.J.: Gaussian splatting slam. arXiv preprint arXiv:2312.06741 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.786, + 0.355 + ], + "angle": 0, + "content": "23. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.786, + 0.383 + ], + "angle": 0, + "content": "24. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.786, + 0.411 + ], + "angle": 0, + "content": "25. Mur-Artal, R., Montiel, J.M.M., Tardos, J.D.: Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics 31(5), 1147-1163 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.786, + 0.438 + ], + "angle": 0, + "content": "26. Mur-Artal, R., Tardós, J.D.: Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras. IEEE transactions on robotics 33(5), 1255-1262 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.44, + 0.787, + 0.493 + ], + "angle": 0, + "content": "27. Newcombe, R.A., Izadi, S., Hilliges, O., Molyneaux, D., Kim, D., Davison, A.J., Kohi, P., Shotton, J., Hodges, S., Fitzgibbon, A.: Kinectfusion: Real-time dense surface mapping and tracking. In: 2011 10th IEEE international symposium on mixed and augmented reality. pp. 127-136. IEEE (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.786, + 0.522 + ], + "angle": 0, + "content": "28. Nieto, J., Bailey, T., Nebot, E.: Recursive scan-matching slam. Robotics and Autonomous systems 55(1), 39-49 (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.786, + 0.577 + ], + "angle": 0, + "content": "29. Palieri, M., Morrell, B., Thakur, A., Ebadi, K., Nash, J., Chatterjee, A., Kanellakis, C., Carlone, L., Guaragnella, C., Agha-Mohammadi, A.a.: Locus: A multi-sensor lidar-centric solution for high-precision odometry and 3d mapping in real-time. IEEE Robotics and Automation Letters 6(2), 421-428 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.786, + 0.619 + ], + "angle": 0, + "content": "30. Prisacariu, V.A., Kähler, O., Golodetz, S., Sapienza, M., Cavallari, T., Torr, P.H., Murray, D.W.: Infinitam v3: A framework for large-scale 3d reconstruction with loop closure. arXiv preprint arXiv:1708.00783 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.786, + 0.646 + ], + "angle": 0, + "content": "31. Qin, T., Li, P., Shen, S.: Vins-mono: A robust and versatile monocular visual-inertial state estimator. IEEE Transactions on Robotics 34(4), 1004-1020 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.786, + 0.688 + ], + "angle": 0, + "content": "32. Sandström, E., Li, Y., Van Gool, L., Oswald, M.R.: Point-slam: Dense neural point cloud-based slam. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18433-18444 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.786, + 0.716 + ], + "angle": 0, + "content": "33. Segal, A., Haehnel, D., Thrun, S.: Generalized-icp. In: Robotics: science and systems. vol. 2, p. 435. Seattle, WA (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.786, + 0.757 + ], + "angle": 0, + "content": "34. Serafin, J., Grisetti, G.: Nicp: Dense normal based point cloud registration. In: 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 742-749. IEEE (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.786, + 0.799 + ], + "angle": 0, + "content": "35. Shan, T., Englot, B.: Lego-loam: Lightweight and ground-optimized lidar odometry and mapping on variable terrain. In: 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 4758-4765. IEEE (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.786, + 0.84 + ], + "angle": 0, + "content": "36. Shan, T., Englot, B., Meyers, D., Wang, W., Ratti, C., Rus, D.: Lio-sam: Tightly-coupled lidar inertial odometry via smoothing and mapping. In: 2020 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 5135-5142. IEEE (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.594, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RGBD GS-ICP SLAM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "37. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "38. Sturm, J., Engelhard, N., Endres, F., Burgard, W., Cremers, D.: A benchmark for the evaluation of rgb-d slam systems. In: 2012 IEEE/RSJ international conference on intelligent robots and systems. pp. 573-580. IEEE (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.272 + ], + "angle": 0, + "content": "39. Sucar, E., Liu, S., Ortiz, J., Davison, A.J.: imap: Implicit mapping and positioning in real-time. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6229-6238 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.3 + ], + "angle": 0, + "content": "40. Taud, H., Mas, J.: Multilayer perceptron (mlp). Geomatic approaches for modeling land change scenarios pp. 451-455 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.342 + ], + "angle": 0, + "content": "41. Vizzo, I., Guadagnino, T., Mersch, B., Wiesmann, L., Behley, J., Stachniss, C.: Kiss-icp: In defense of point-to-point icp-simple, accurate, and robust registration if done the right way. IEEE Robotics and Automation Letters 8(2), 1029–1036 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.787, + 0.383 + ], + "angle": 0, + "content": "42. Whelan, T., Johannsson, H., Kaess, M., Leonard, J.J., McDonald, J.: Robust real-time visual odometry for dense rgb-d mapping. In: 2013 IEEE International Conference on Robotics and Automation. pp. 5724-5731. IEEE (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.383, + 0.787, + 0.411 + ], + "angle": 0, + "content": "43. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.787, + 0.438 + ], + "angle": 0, + "content": "44. Yan, C., Qu, D., Wang, D., Xu, D., Wang, Z., Zhao, B., Li, X.: Gs-slam: Dense visual slam with 3d gaussian splatting. arXiv preprint arXiv:2311.11700 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "45. Yang, X., Li, H., Zhai, H., Ming, Y., Liu, Y., Zhang, G.: Vox-fusion: Dense tracking and mapping with voxel-based neural implicit representation. In: 2022 IEEE International Symposium on Mixed and Augmented Reality (ISMAR). pp. 499-507. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.508 + ], + "angle": 0, + "content": "46. Zhang, J., Singh, S.: Loam: Lidar odometry and mapping in real-time. In: Robotics: Science and systems. vol. 2, pp. 1-9. Berkeley, CA (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.787, + 0.55 + ], + "angle": 0, + "content": "47. Zhu, Z., Peng, S., Larsson, V., Xu, W., Bao, H., Cui, Z., Oswald, M.R., Pollefeys, M.: Nice-slam: Neural implicit scalable encoding for slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12786-12796 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.55 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/RGBD GS-ICP SLAM/794618d9-8655-4def-b2e9-3b1608333e79_origin.pdf b/2024/RGBD GS-ICP SLAM/794618d9-8655-4def-b2e9-3b1608333e79_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dc15faeafb8f5eb3396eb7072fba41148439659d --- /dev/null +++ b/2024/RGBD GS-ICP SLAM/794618d9-8655-4def-b2e9-3b1608333e79_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:713cf41a9a1c2467e88d824375b109400b3ee2806fc877067fb5b41342062aed +size 8018861 diff --git a/2024/RGBD GS-ICP SLAM/full.md b/2024/RGBD GS-ICP SLAM/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7857723b88b403240e7d76926a5ed6fa0a1f98f2 --- /dev/null +++ b/2024/RGBD GS-ICP SLAM/full.md @@ -0,0 +1,260 @@ +# RGBD GS-ICP SLAM + +Seongbo Ha, Jiung Yeon, and Hyeonwoo Yu + +Sungkyunkwan University, Suwon, South Korea +{sobo3607, wcr12st, hwyu}@skku.edu + +Abstract. Simultaneous Localization and Mapping (SLAM) with dense representation plays a key role in robotics, Virtual Reality (VR), and Augmented Reality (AR) applications. Recent advancements in dense representation SLAM have highlighted the potential of leveraging neural scene representation and 3D Gaussian representation for high-fidelity spatial representation. In this paper, we propose a novel dense representation SLAM approach with a fusion of Generalized Iterative Closest Point (G-ICP) and 3D Gaussian Splatting (3DGS). In contrast to existing methods, we utilize a single Gaussian map for both tracking and mapping, resulting in mutual benefits. Through the exchange of covariances between tracking and mapping processes with scale alignment techniques, we minimize redundant computations and achieve an efficient system. Additionally, we enhance tracking accuracy and mapping quality through our keyframe selection methods. Experimental results demonstrate the effectiveness of our approach, showing incredibly fast speeds up to 107 FPS (for the entire system) and superior quality of the reconstructed map. + +The code is available at: https://github.com/Lab-of-AI-and-Robotics/GS-ICP-SLAM Video is: https://youtu.be/ebHh_uMMxE + +Keywords: Coordinate-based 3D Representation $\cdot$ G-ICP $\cdot$ SLAM + +# 1 Introduction + +Visual Simultaneous Localization and Mapping (SLAM) is an algorithm that constructs maps of unknown environments while localizing poses of vision sensors simultaneously. As the influence of 3D visual SLAM in the fields of robotics, Virtual Reality (VR), and Augmented Reality (AR) has increased, higher rendering performance and more accurate trajectory are required. Thus, 3D reconstruction methods such as Signed Distance Field (SDF), and Truncated Signed Distance Field (TSDF) are utilized for traditional dense visual SLAM [3,5,10,27,30,42]. + +Recently, coordinate-based 3D Implicit Neural Representation (INR) [23] has been proposed for representing spatial information using neural radiance fields, showcasing high novel view synthesis capabilities and high-fidelity spatial representation prowess. Various approaches [8, 18, 32, 39, 45, 47] have been attempted to utilize INR for the real-time SLAM mapping process. However, INR requires computationally intensive raycasting to synthesize images, thus the rendering process in INR-based SLAM incurs significant time overhead, slowing down map optimization and rendering-loss based tracking. In contrast, coordinate-based 3D explicit representation such as 3D Gaussian + +![](images/9fd7eda37ed65e7f9c152ef05d387a65623c1897dcffc3e55e57a45c5241df0c.jpg) +Fig. 1: A comparison of PSNR with respect to FPS of entire system in recent research on SLAM algorithm utilizing dense representation such as neural scene representation and 3D Gaussian representation. Our method achieves state-of-the-art performance in rendering evaluation and FPS of entire system. Note that this FPS represents the overall system performance. Reported values are average of Replica 8 scenes. + +Splatting (3DGS) [15] represents the 3D space using 3D Gaussians as primitives, allowing for rendering speeds faster than NeRF using the rasterization method [43]. This rapid rendering capability of 3DGS results in fast optimization of 3D spatial information, making it suitable for dense representation in SLAM [12, 14, 22, 44]. + +Although 3DGS-based SLAM methods take advantages of the high-speed rendering, they fail to address the fundamental issue: the inability to directly utilize 3D explicit representations and the indirect tracking of 3D space through 2D image rendering. Even with the majority of current 3DGS-based SLAM [12, 14, 22, 32, 44] utilizing RGB-D data, the use of explicit representations is overlooked. For example, INR-based methods [32, 39, 47] and 3DGS-based methods [14, 22, 44] employ photometric-error based techniques, which estimate the optimal pose by iteratively minimizing the error between rendered and observed 2D images. However, due to limitations in tracking speed and performance, decoupled approaches [8, 12, 18] have been proposed incorporating well-crafted visual odometries [4, 26] into the tracking process. Although decoupled methods exhibit better performances in tracking by separating mapping and tracking processes, they require additional computational resources for the independent stages. Moreover, these methods need to store features in 3D space, unrelated to the 3DGS map. + +There exists a method that allows Gaussian, an explicit representation, to be directly used for tracking. The well-known Generalized Iterative Closest Point (G-ICP) [17, 33] from the 3D scan matching family is simple yet efficient for fast tracking of 3D point clouds. During preprocessing, it only requires computing Gaussians for the current frame and the map. Given that the map in 3DGS utilizes Gaussians as an explicit representation in 3D space, using G-ICP for tracking allows for the direct utilization of the 3DGS map without the need for post-processing. Furthermore, the Gaussians of the current frame computed during tracking with G-ICP can also be directly utilized as an explicit representation in the 3DGS map, without additional computations. + +Therefore, we propose a dense representation SLAM framework that integrates G-ICP and 3DGS, allowing them to complement each other by sharing explicit representations. Unlike traditional methods, our approach actively leverages 3D information through G-ICP for tracking. The proposed method is a coupled approach that shares a single map during tracking and mapping processes, while maintaining fast tracking speed akin to decoupled approaches. Previous works using decoupled methods [8, 12, 18] require separate maps containing ORB feature information for tracking, along with additional resources and computations needed to obtain ORB features. Moreover, the values obtained during tracking are not utilized in mapping. Our approach exploits the covariance of each point computed during the G-ICP-based tracking process as the initial state of the 3DGS mapping. Meanwhile, the 3D Gaussians in the 3DGS map are also mutually used as 3D points and their covariances which are essential for G-ICP-based tracking, discarding the necessity to recalculate the covariances of the map. This is possible because the covariances of points computed during the G-ICP process and the 3D Gaussians representing the map commonly contain information about the surrounding space. In other words, G-ICP and 3DGS can share the same Gaussian world. Therefore, our system minimizes unnecessary computations and facilitates efficient system configuration by mutually utilizing the key elements, Gaussians, between tracking and mapping processes. To ensure optimal performance in tracking and mapping sharing information between G-ICP and 3DGS, we also introduce several techniques such as scale alignment. To summarize, our contributions are as follows: + +- We present a real-time dense representation SLAM that combines G-ICP and 3DGS, achieving extremely high speed of the entire system (up to 107 FPS) and superior quality of the map. +- By incorporating G-ICP for tracking, our system utilizes 3D information actively and significantly reduces the time required for the tracking process. +- Reduction of computational cost and facilitating rapid convergence of primitives of 3DGS is achieved by sharing the covariances of G-ICP and 3DGS with scale aligning techniques. + +# 2 Related Work + +G-ICP Scan-matching focuses on the registration of two point clouds that observe similar environments by selecting a transformation matrix that minimizes errors between two point clouds [2]. Iterative Closest Point (ICP) [1] algorithm is a widely used and influential method in the scan-matching area. ICP iteratively estimates point correspondences and finds a transformation that minimizes Euclidean distance between corresponding points, thereby optimizing the registration of point clouds. Because of simplicity and speed of ICP, there are many follow-up studies [7, 20, 33, 34, 41]. To improve robustness, Trimmed-ICP [7], proposed a method for correspondence selection. Point-to-plane ICP [20] augmented robustness and accuracy by taking point-to-plane distance as objective function. G-ICP [33] introduced probabilistic models to generalize ICP yielding notable improvements in robustness and accuracy. Scan matching is a common technique used in SLAM for pose tracking [6, 11, 21, 28, 29], and for enhanced robustness, feature-based scan matching is often employed [19, 35, 36, 46]. + +SLAM with Dense Representation SLAM with dense representation [9, 16] aims to construct maps in a dense form to enable interaction with the map in tasks such as AR, robotics, etc. To achieve this, classic approaches [3,5,10,27,30,42] creates dense maps by representing space in Signed Distance Field (SDF), and Truncated Signed Distance Field (TSDF), rather than in sparse forms such as point clouds or grids. + +Recently, dense representation SLAM methods represent maps by utilizing NeRF [23], which demonstrates high spatial representation capabilities have been proposed. iMAP [39], NICE-SLAM [47], Point-SLAM [32], ESLAM [13] perform tracking by optimizing camera poses by reducing errors between synthesized and observed images. These methods have limitations in tracking speed and performance. Alternatively, Orbeez-SLAM [8], vMAP [18] incorporate well-crafted visual SLAM techniques into the tracking process, making mapping and tracking operate independently. Orbeez-SLAM utilizes ORB-SLAM2 [26] for tracking and instant-ngp [24] for learning spatial information to perform mapping. In a similar context, vMAP utilizes ORB-SLAM3 [4] for tracking while employing separate Multi-layer perceptrons [40] for each object to effectively represent the entire space. + +3DGS represents space in explicit form using 3D gaussians as primitives. It offers a level of high-fidelity spatial representation similar to NeRF but provides significantly faster rendering speeds. To leverage this advantage to dense representation SLAM, several approaches [12, 14, 22, 44] utilizing 3DGS for space representation have been proposed. Among them, GS-SLAM [44], SPLTAM [14], Gaussian Splating SLAM [22] perform tracking using dense photometric error. While 3DGS shows significantly improved rendering performance compared to NeRF, the tracking methods based on dense photometric error still suffer from tracking speed. To tackle this problem, Photo-SLAM [12] integrates ORB-SLAM3 for tracking, in a similar manner to Orbeez-SLAM and vMAP. + +# 3 Method + +To mutually benefit tracking and mapping, we introduce the fusion of G-ICP and GS, each representing the tracking and mapping processes, respectively. The key insight of our approach is that covariance can be considered as a fundamental common factor for this fusion. Suppose we have a point set (point cloud) $\mathcal{X} = \{\pmb{x}_m\}_{m=1,\dots,M}$ and its corresponding covariance set $\mathcal{C} = \{C_m\}_{m=1,\dots,M}$ , where $\pmb{x} = [x,y,z]^T$ . The covariance $C$ of one 3D point $\pmb{x}$ is given by computing covariance matrix of $k$ -nearest neighbors of $\pmb{x}$ . Let us define $\pmb{G} = \{\pmb{\mathcal{X}},\pmb{\mathcal{C}}\}$ as a set of Gaussians. G-ICP aims to find a transformation $\mathbf{T}$ that maximally aligns the source Gaussians (current frame) $\pmb{G}^s = \{\pmb{\mathcal{X}}^s,\pmb{\mathcal{C}}^s\}$ and the target Gaussians (map) $\pmb{G}^t = \{\pmb{\mathcal{X}}^t,\pmb{\mathcal{C}}^t\}$ . Assume we know the correspondences between $\pmb{\mathcal{X}}^s$ and $\pmb{\mathcal{X}}^t$ determined by nearest neighbor search. For example, we have $\{\pmb{x}_i^s\}_{i=1,\dots,N} \subset \pmb{\mathcal{X}}^s$ , $\{C_i^s\}_{i=1,\dots,N} \subset \pmb{\mathcal{C}}^s$ and $\{\pmb{x}_i^t\}_{i=1,\dots,N} \subset \pmb{\mathcal{X}}^t$ , $\{C_i^t\}_{i=1,\dots,N} \subset \pmb{\mathcal{C}}^t$ , where $\pmb{x}_i^s$ is associated with $\pmb{x}_i^t$ . To find the optimal transform $\mathbf{T}^*$ , we exploit not the single point but the distribution of that point defined as a Gaussian distribution: $\pmb{x}_i \sim \mathcal{N}(\hat{\pmb{x}}_i,C_i)$ . Let $d_i = \pmb{x}_i^t - \mathbf{T}\pmb{x}_i^s$ be the error term, and if we assume that there is an optimal transformation $\mathbf{T}^*$ , it is clear that $\pmb{x}_i^t = \mathbf{T}^*\pmb{x}_i^s$ thus $\hat{d}_i = 0$ . Since + +we assume $\pmb{x}$ is a Gaussian random variable, $d_{i}$ is also a Gaussian random variable as the following: + +$$ +\begin{array}{l} d _ {i} \sim \mathcal {N} \left(\hat {d} _ {i}, C _ {i} ^ {t} + \mathbf {T} ^ {*} C _ {i} ^ {s} \left(\mathbf {T} ^ {*}\right) ^ {T}\right) \\ = \mathcal {N} \left(\hat {x} _ {i} ^ {t} - \mathbf {T} ^ {*} \hat {x} _ {i} ^ {s}, C _ {i} ^ {t} + \mathbf {T} ^ {*} C _ {i} ^ {s} \left(\mathbf {T} ^ {*}\right) ^ {T}\right) \\ = \mathcal {N} (0, C _ {i} ^ {t} + \mathbf {T} ^ {*} C _ {i} ^ {s} (\mathbf {T} ^ {*}) ^ {T}). \\ \end{array} +$$ + +To find the optimal transform $\mathbf{T}^*$ for $\boldsymbol{x}^s$ and $\boldsymbol{x}^t$ , we use maximum likelihood estimation (MLE) as the following: + +$$ +\begin{array}{l} \mathbf {T} ^ {*} = \underset {\mathbf {T}} {\operatorname {a r g m a x}} \prod_ {i} ^ {N} p \left(d _ {i}\right) = \underset {\mathbf {T}} {\operatorname {a r g m a x}} \sum_ {i} ^ {N} \log p \left(d _ {i}\right) \\ = \underset {\mathbf {T}} {\operatorname {a r g m i n}} \sum_ {i} ^ {N} d _ {i} ^ {T} \left(C _ {i} ^ {B} + \mathbf {T} C _ {i} ^ {A} \mathbf {T} ^ {T}\right) ^ {- 1} d _ {i}. \tag {1} \\ \end{array} +$$ + +Therefore, $\mathbf{T}^*$ can be used as the relative pose between the current frame $G^{s} = \{\mathcal{X}^{s},\mathcal{C}^{s}\}$ and the map $G^{t} = \{\mathcal{X}^{t},\mathcal{C}^{t}\}$ . + +Meanwhile, for mapping purposes, GS also relies on Gaussians $G = \{\mathcal{X},\mathcal{C}\}$ of the 3D scene representation. Different from G-ICP, GS aims to find the optimal coordinates of the Gaussians $\mathcal{X}^* = \{x_m^*\}_{m=1,\dots,M}$ and the optimal covariances $\mathcal{C}^* = \{C_m^*\}_{m=1,\dots,M}$ as the following: + +$$ +\boldsymbol{\mathcal{X}}^{*},\boldsymbol{\mathcal{C}}^{*},\boldsymbol{H}^{*},\boldsymbol{O}^{*} = \operatorname *{argmin}_{\boldsymbol {\mathcal{X}},\boldsymbol {\mathcal{C}},\boldsymbol {H},\boldsymbol{O}}\lambda_{I_{1}}\boldsymbol{\mathcal{L}}_{1}\left(I,I_{gt}\right) + \lambda_{I_{2}}\boldsymbol{\mathcal{L}}_{D - SSIM}\left(I,I_{gt}\right) + \lambda_{D}\boldsymbol{\mathcal{L}}_{1}\left(D,D_{gt}\right) +$$ + +where $H = \{h_m\}_{m=1,\dots,M}$ and $O = \{o_m\}_{m=1,\dots,M}$ are the color set and the opacity set of 3D points which are for the RGBD image rendering. $I$ and $D$ are the rendered RGB and depth images obtained by performing rasterization using $G$ , $H$ and $O$ . + +Note that in G-ICP and GS, the key common factor is Gaussians $G = \{\mathcal{X},\mathcal{C}\}$ , allowing these Gaussians to be shared mutually. During G-ICP tracking, the covariance of each frame is computed. Hence, when adding keyframes to expand the 3DGS map, there is no need to recalculate $\mathcal{C}$ for every expansion. Also, G-ICP does not need to compute the covariances of the map because our GS map already contains Gaussians. Moreover, G-ICP, by aligning frames based on 3D geometric structure, inherently initializes a certain number of points and their coordinates that aptly represent the 3D structure. Therefore, it brings about an effect where an appropriate number of points and their poses are initialized, well-suited for depicting the 3D structure, significantly reducing the learning time to find the optimal pose $\mathcal{X}^*$ and the optimal covariance $\mathcal{C}^*$ of Gaussians $G$ in GS. Additionally, in the same vein, calculations like densifying or opacity reset to adjust the number of 3D points in GS become unnecessary. Consequently, by sharing a common source, each process becomes mutually beneficial, and the speed of execution accelerates due to reduced redundant computations. Our method can be simplified as follows: + +1. Use G-ICP to align the current frame with the 3DGS map which contains covariance (solely need to compute the covariance for the current frame). + +![](images/1ad7b8524217766017f170291f86933956bf6b83ec241d6c565bef0ae7db4a10.jpg) +Fig. 2: System Overview. The input of our system is RGBD frame. We generate a point cloud by downsampling and reprojecting the current depth image and utilize it in the GICP process. During the GICP process, we create source Gaussians from the point cloud and estimate the current camera pose by aligning them with target Gaussians, which are a subset of the 3DGS map. If the current frame is identified as a keyframe or a mapping-only keyframe, we add the source Gaussians to the 3DGS map as new primitives. Meanwhile, in the mapping process, we optimize the Gaussians along with the color and opacity set of the Gaussians concurrently with the tracking process. + +2. When adding keyframes to the 3DGS map, utilize the covariance computed in GICP during tracking (no need for densifying or opacity reset). +3. Repeat steps 1-2. + +The overview of our system is shown in Fig. 2. The detailed implementations of the proposed technique are introduced in the following Section. + +# 3.1 G-ICP Tracking + +Scale Regularization When using G-ICP for tracking, aligning the scale of the current frame with the map enables high-performance camera pose estimation [33]. Given the covariance $C$ , scale $\boldsymbol{S} = [s_2, s_1, s_0]^T$ is given by the singular value decomposition (SVD) as the following: + +$$ +C = \boldsymbol {R} \boldsymbol {\Lambda} ^ {2} \boldsymbol {R} ^ {T} \tag {2} +$$ + +where $\Lambda = \text{diag}(s_2, s_1, s_0) \forall s_2 > s_1 > s_0$ is a scale matrix and $\pmb{R}$ is the orientation of the Gaussian. To achieve the robust scan matching performance, point-to-plane ICP or voxelized point-to-plane ICP adopt a regularizing method that makes the scale as $\pmb{S} = [1, 1, \epsilon]^T$ in order to treat each Gaussian as a plane-like distribution. + +In our framework, we utilize target Gaussians $\pmb{G}^t = \{\pmb{x}^t,\pmb{c}^t\}$ from the 3DGS map for tracking. Here, $\pmb{c}^t$ is optimized to represent the scene accurately, $\pmb{G}^t$ involves not only planes but also lines, corners, and other features. Therefore, while tracking, instead of regularizing the scale into a plane-based form, preserving the original characteristics of the target Gaussians while performing regularization is more appropriate. Thus, in our framework, we propose the following ellipsoid regularization: + +$$ +\boldsymbol {\Lambda} ^ {\prime} = \frac {1}{m e d i a n (\boldsymbol {S})} \operatorname {d i a g} \left(s _ {2}, s _ {1}, s _ {0}\right) \tag {3} +$$ + +where $\pmb{\Lambda}^{\prime}$ is the ellipsoid-regularized scale matrix. + +![](images/4fecd2b2fe34b0ccf75cc0977939a3930ab9e131e3d788583ce912a62a12304e.jpg) +(a) + +![](images/ae889a95482c4e04d4a00af4e0c0799196234122024407ae5202eec361e2cdc3.jpg) +(b) +Fig. 3: Tracking Accuracy Comparison Based on Keyframe Selection Methods. The reported values represent the average results across eight scenes from the Replica dataset [37]. When selecting keyframes every n frame (depicted in blue), the tracking accuracy is notably low. Conversely, our keyframe selection method yielded the highest tracking accuracy. + +Keyframe Selection Similar to [25,31], we perform dynamic keyframe selection. Considering the geometric structure is critical for both G-ICP tracking and GS mapping, we exploit the geometric correspondence computed from G-ICP. As Eqn. (1) denotes, we can get the distance between $\boldsymbol{x}_i^s$ and $\boldsymbol{x}_i^t$ as an interim result of G-ICP procedure without additional calculations. So we can efficiently determine the correspondence by setting the distance threshold. By considering the proportion of correspondences between the current frame and map, we select keyframes while taking into account the characteristics of the scene. In more detail, if the proportion in the current frame goes below specific thresholds, the frame is selected as a keyframe. This allows us to achieve consistent tracking performance while simultaneously maintaining a consistent density of Gaussians $G = \{\mathcal{X},\mathcal{C}\}$ added to the map. The effect of this keyframe selection is shown in Fig. 3 (a). Note that, from selected keyframe, only Gaussians that do not overlap with the existing current map are considered as target Gaussians. Errors from discrepancies during tracking are directly incorporated into the map, leading to accumulated inaccuracies and reduced tracking precision. This strategy ensures more accurate tracking results and improves the overall robustness of the system. + +# 3.2 GS Mapping + +Scale Aligning Suppose we have a fully trained 3DGS map. Since the 3DGS map is trained to represent the scene from any viewpoint, Gaussians $G$ should be uniformly distributed according to the structure of the 3D space. However, single frames obtained from radiance sensors such as RGBD cameras or LiDAR yield sparse representations of the 3D space as the distance from the sensor increases, due to the inherent characteristics of the sensor. In other words, as the distance from the sensor increases, the spacing between 3D points widens, and the scale of the corresponding covariance calculated based on the $k$ -nearest neighbors becomes excessively large. Adding such single frames as keyframes to the map during real-time training of GS mapping leads to an imbalance in the scale of Gaussians, ultimately resulting in a decrease in mapping performance. Therefore, to fundamentally alleviate the problem, our proposed method suggests the + +![](images/986f348d0f8449baaf0b34adfe9c8d720af5be7e608131bc957215d3cc8fb216.jpg) +Fig. 4: Separated Keyframe Selection on Replica office4. We demonstrate that a small number of tracking keyframes yield accurate trajectory estimation (case 1), while a large number of mapping keyframes result in high rendering performance (case 2). Thus, our method adopts case 3 that selects tracking keyframe and mapping keyframe separately at different intervals. + +![](images/d1280fceeaacaf642d05cfefae8894bcfa8c6c6e7c6bd82c421f693ed078b089.jpg) + +![](images/c4a61d2aa57e457d2ad0ebbdc8857eaeb666befccf32c8f388bc4a2c12768665.jpg) + +following scale normalization for the current frame: + +$$ +\boldsymbol {\varLambda} ^ {\prime \prime} = \frac {1}{z ^ {p}} \boldsymbol {\varLambda}. +$$ + +Where $\Lambda''$ is the normalized scale matrix and $p$ is the parameter which is empirically determined. By adding such scale-normalized keyframes to the map to expand it, not only does the GS mapping performance increase, but also the performance of G-ICP tracking improves. + +Additional Keyframe Selection for Mapping Mapping based on GS benefits from having a diverse set of frames from various viewpoints for training, as performance improves with a larger number of available frames. However, increasing the number of selected keyframes during tracking can lead to a degradation in tracking performance due to accumulated errors, as is common with G-ICP-based methods. To mitigate this issue, we propose adding mapping-only keyframe selection in addition to the existing keyframe selection process. In other words, tracking continues to proceed from selected keyframes in the existing tracking process, while mapping utilizes both the original keyframes and additional mapping-only keyframes. + +The overview of the additional keyframe selection is illustrated in Fig. 4. Cases 1 and 2 represent scenarios without considering mapping-only keyframes. In Case 1, to increase the number of keyframes, if no keyframe satisfying the criteria is found up to the $30^{th}$ frame from the previous keyframe, the $30^{th}$ frame is selected as a keyframe. However, since this may not yield sufficient mapping performance, as shown in Case 2, adding an additional keyframe at the $10^{th}$ frame can be considered. While this approach improves mapping performance as desired, it leads to a decline in tracking performance due to accumulated scan matching errors. To alleviate this, as shown in Case 3, we propose adding the $10^{th}$ frame as a mapping-only frame, allowing us to maximize tracking performance while simultaneously maximizing mapping performance. The effect of the proposed scheme, Case 3, is shown in Fig. 3 (b). Note that since G-ICP, which performs tracking, computes covariance for all frames anyway, there is no difference in the overall computational load. + +Avoiding Local Minima while Mapping GS SLAM differs from traditional GS in that it operates in real-time and often lacks sufficient observations of the scene. Moreover, since it operates in real-time, focusing solely on images from the current viewpoint or nearby viewpoints during training can easily lead to local minima and degrade mapping quality [22]. For instance, GS simply considers image rendering from specific viewpoints, thus if we continuously train only on a specific viewpoint, the scale of Gaussians in the map tends to elongate in the direction of that viewpoint, leading to local minima. + +To overcome this challenge, at each training iteration, we adopt a strategy of randomly choosing one keyframe for learning among the selected keyframes so far, ensuring that both the current observed scene and the entire map are uniformly learned. Additionally, we prune Gaussians that fall into local minima during training to improve mapping performance and ensure robust tracking through geometric preservation. + +# 4 Experiments + +# 4.1 Experimental Setup + +We evaluate our method on Replica dataset [37] and TUM dataset [38]. Replica dataset contains synthetic scenes and high-quality RGB/depth images rendered from these scenes. TUM dataset includes images captured in the real world, with significant noise and blur, resulting in poor quality. Particularly, many parts of the depth images suffer from information loss. We validate the effectiveness of our approach by evaluating it on both synthetic and real-world datasets. All experiments are performed on a desktop with a Ryzen 7 7800x3d CPU, 32GB RAM, and an NVIDIA RTX 4090 24GB GPU. To evaluate camera tracking accuracy, we use the Root Mean Square Error (RMSE) of the Absolute Trajectory Error (ATE). For the quality of the reconstructed map, we report standard photometric rendering quality metrics (PSNR, SSIM, and LPIPS). + +Table 1: Tracking Performance on Replica and TUM-RGBD (ATE RMSE $\downarrow$ [cm]). Our method archives state-of-the-art tracking accuracy on Replica. On TUM, Among coupled methods, we achieve competitive performance. The results of GS-SLAM/Gaussian Splitting SLAM are taken from [44]/ [22]. And the results of ORB-SLAM3 and Photo-SLAM are obtained from [12]. + +
MethodReplicaTUM
R0R1R2Of0Of1Of2Of3Of4Avg.fr1/deskfr2/xyzfr3/officeAvg.
DecoupledORB-SLAM3 [4]-------1.81.70.41.71.3
Photo-SLAM [12]-------0.602.60.31.01.3
CoupledNICE-SLAM* [47]1.611.481.610.950.811.461.761.691.422.82.17.24.0
Point-SLAM* [32]0.590.510.320.450.460.480.610.870.542.71.33.92.6
GS-SLAM [44]0.480.530.330.520.410.590.460.700.503.31.36.63.7
SplaTAM* [14]0.290.350.280.490.210.310.340.570.363.31.35.13.2
Gaussian Splitting SLAM* [22]0.330.220.290.360.190.250.120.810.321.51.41.51.5
Ours (limited to 30 FPS)0.150.160.110.180.120.170.160.210.162.71.82.72.4
+ +* denotes the reproduced results by running official code. + +# 4.2 Camera Tracking Accuracy + +Tab. 1 shows the tracking accuracy of our method compared to other approaches on both synthetic [37] and real-world [38] datasets. On Replica [37], the proposed method achieved state-of-the-art (SOTA) performance across all scenes, reducing the trajectory error by more than $50\%$ compared to the previous SOTA. This result arises from the fact that our proposed method actively utilizes 3D information by employing G-ICP [33] for tracking, unlike other methods that perform tracking based on errors in 2D space. On TUM [38], our method shows competitive tracking accuracy among coupled methods [14,22,32,44,47] which conduct tracking and mapping based on a single map. Note that the results of Gaussian Splitting SLAM are evaluated only on keyframes. Decoupled method [12] and its baseline [4] show better performance, but these approaches require additional resources for storing separate map information and costly computations for extracting features solely for the tracking process. + +Table 2: Evaluation of System Speed and Map Quality on Replica. Our method outperforms all other frameworks in both system speed and quality of the reconstructed map. + +
MethodsMetricsR0R1R2Of0Of1Of2Of3Of4Avg.
Orbeez-SLAM* [8]PSNR[dB]↑12.1315.2815.8717.5919.2610.3011.5512.6514.33
SSIM ↑0.700.790.810.800.670.780.760.840.77
LPIPS ↓0.610.480.510.520.410.550.540.530.53
FPS ↑24.0524.7824.0019.2226.7423.8824.4526.1124.15
Point-SLAM* [32]PSNR[dB]↑33.3834.1036.3238.7239.3134.2234.1034.8235.62
SSIM ↑0.980.980.990.990.990.960.960.980.98
LPIPS ↓0.100.120.100.090.110.150.120.130.11
FPS ↑0.260.300.310.330.340.300.280.300.30
GS-SLAM [44]PSNR[dB]↑31.5632.8632.5938.7041.1732.3632.0332.9234.27
SSIM ↑0.970.970.970.990.990.980.970.970.98
LPIPS ↓0.090.080.090.050.030.090.110.110.08
FPS ↑8.34-------8.34
SplaTAM* [14]PSNR[dB]↑32.6033.5534.8338.0939.0231.9529.5331.5533.89
SSIM ↑0.980.970.980.980.980.970.950.950.97
LPIPS ↓0.070.100.070.090.090.100.120.150.10
FPS ↑0.240.190.190.200.220.270.260.240.23
Gaussian Splitting SLAM [22]PSNR[dB]↑34.8336.4337.4939.9542.0936.2436.7036.0737.50
SSIM ↑0.950.960.970.970.980.960.960.960.96
LPIPS ↓0.070.080.080.070.060.080.070.100.08
FPS ↑-----1.1---
Ours (no tracking speed limit)PSNR[dB]↑32.2035.3634.4240.3140.7533.8534.0836.4735.93
SSIM ↑0.940.960.960.980.980.960.950.960.96
LPIPS ↓0.080.070.080.050.050.070.070.070.07
FPS ↑100.9884.92103.3899.10107.0695.6097.2096.6698.11
Ours (limited to 30 FPS)PSNR[dB]↑35.3737.8038.5043.1343.2636.9336.9038.7538.83
SSIM ↑0.960.970.980.990.990.970.970.970.98
LPIPS ↓0.050.050.050.030.030.040.040.050.04
FPS ↑29.9729.9829.9829.9829.9929.9729.9729.9729.98
+ +* denotes the reproduced results by running official code. The result of GS-SLAM and Gaussian Splitting SLAM is taken from [44] and [22]. + +Table 3: Evaluation of System Speed and Map Quality on TUM-RGBD. Proposed method shows incredible system speed and competitive map quality. + +
MethodsPSNR[dB] ↑SSIM ↑LPIPS ↓FPS ↑
NICE-SLAM* [47]14.100.5740.3950.08
Point-SLAM* [32]21.400.7380.4470.22
Photo-SLAM [12]21.900.7630.187-
SplaTAM* [14]23.460.9060.1560.32
Ours (unlimited tracking speed)19.620.7500.24073.92
Ours (limited to 30 FPS)20.720.7680.21829.99
+ +* denotes the reproduced results by running official code. The result of Photo-SLAM is taken from [12]. + +# 4.3 System Speed and Quality of Reconstructed Map + +Tab. 2 and Tab. 3 demonstrate the FPS of the systems and quality of the reconstructed map on Replica [37] and TUM [38] dataset, respectively. FPS of the system is calculated by dividing the total number of frames by the total time. Note that this FPS represents the overall system performance including tracking and mapping processes, rather than just individual components. Photo-SLAM [12] reported the tracking FPS of their system, but did not include the system FPS, so we leave that field blank in the Tab. 3. Since our method implements mapping and tracking processes to operate in parallel, the number of map optimization iterations varies depending on the tracking speed, consequently affecting the quality of the reconstructed map. So we evaluate our system in two cases: one where the tracking speed is limited to the typical sensor input speed of 30 FPS, and the other where the tracking speed is not limited. Please refer to the supplementary material for results across various FPS and geometric quality evaluations. + +![](images/e969dc582149105ea922270bd1c198849e75b97ae6be9ed350ecc6ee3639711c.jpg) +Fig. 5: Comparison of Rendering Results. In the first scene, SplaTAM [14] failed to reconstruct the pillow and lamp and Point-SLAM [32] failed to represent the detailed pattern of the pillow. In the second case, SplaTAM and Point-SLAM failed to accurately reconstruct the details of the clock. However, our method exhibits rendering results that closely resemble the ground truth image, demonstrating a high level of accuracy. + +In the limited case (30 FPS), our system achieves SOTA map quality across all scenes in the Replica dataset. When the tracking speed is not limited, our method shows extremely fast system speed up to 107 FPS. This speed is more than four times faster than that of the Orbeez-SLAM [8], based on the ORB-SLAM2 [26]. Remarkably, even under this condition, our approach maintains superior map quality and outperforms other methods. The proposed method effectively integrates Gaussians for map representation with appropriate initial states and has better detailed representation capability compared to NeRF-based methods. As shown in Fig. 5, SplaTAM suffers from incorporating appropriate Gaussians into the map, leading to regions of poor quality. Point-SLAM exhibits areas where detailed representation is lacking. In TUM-RGBD, our method demonstrates remarkably fast system speed with competitive map quality, showcasing its practicality in real-world scenarios. Compared to SplaTAM [14], our approach shows a slight decrease in map quality, with a PSNR reduction of approximately $11.7\%$ . However, we achieve significant speed improvements, with a speedup of approximately 91.6 times under limited conditions and up to 227 times at maximum speed. The reason is that, unlike Replica, TUM dataset is captured with old-fashioned sensors, resulting in noisy and substantial information loss in depth images. While SplaTAM treats these factors by adding new Gaussians only in regions with significant depth loss, our method utilizes all structural information from the depth image, focusing on accurate tracking and rapid system rather than treating such factors. + +# 4.4 Ablation study + +Table 4: Scale Regularization Ablation on TUM-RGBD. Reported results are the average ATE RMSE $\downarrow$ [cm] of 3 scenes in TUM-RGBD dataset. + +
Scale Regularizationfr1-deskfr2-xyzfr3-officeAvg.
X136.32196.61376.40236.54
Plane65.635.5316.2129.12
ellipsoid2.661.772.672.37
+ +Scale Regularization Tab. 4 shows the result of the ablation study on scale regularization while tracking. We perform scale regularization to enhance tracking accuracy by ensuring that the scales of Gaussians in the current frame and the map are within a similar range. Despite the application of plane regularization resulting in some improvement in tracking accuracy, it remains inadequate. This is because the plane regularization method considers all the Gaussians as planes, ignoring the characteristics of Gaussians existing in the map, which are optimized to represent the surrounding space through the mapping process. On the other hand, when applying the proposed ellipsoid regularization, the best tracking accuracy is achieved. This is attributed to the ability of the proposed ellipsoid regularization to consider the characteristics of Gaussians existing in the map while performing regularization. + +Scale Aligning We implement scale aligning to reduce the difference between scales + +Table 5: Ablation of Scale Aligning on Replica. The results are the average of 8 scenes in the Replica dataset. + +
Covs from G-ICPRe-initATE ↓PSNR ↑SSIM ↑LPIPS ↓
XX8.89324.810.8400.380
✓X0.25833.210.9220.177
✓constant0.15837.330.9640.074
✓z1.50.15738.830.9750.041
+ +of existing Gaussians in the map and newly added Gaussians. By doing this, we ensure that scales of newly added Gaussians closely resemble the optimal value they will reach through optimization. To demonstrate the effectiveness of this approach, we tested our method in various cases and Tab. 5 shows the results. When not utilizing the covariances computed during the G-ICP [33] process, we calculate the scales of Gaussians by using simple-knn module of vanilla 3DGS [15], and set rotations as identity. In this case, a significant drop in tracking performance appeared. This is because the newly added Gaussians lack sufficient prior information about the 3D structure, causing them to be utilized as target Gaussians without being optimally optimized in space. In the case of using the computed covariance from the G-ICP process, the tracking and mapping performance improves. Furthermore, the best performance is observed when dividing the scale by $z^{1.5}$ . These results indicate that utilizing covariances calculated during G-ICP procedure is effective for both tracking and mapping, and scale aligning facilitates smooth connection from G-ICP to 3DGS. Further analysis of the advantages of this connection is included in the supplementary material. + +Table 6: Keyframe and Mapping-Only Keyframe Selection Ablation on Replica. Reported results are the average of 8 scenes in the Replica dataset. + +
KeyframeMapping-Only KeyframeATE [cm] ↓PSNR ↑SSIM ↑LPIPS ↓
tracking + every 30 frame-0.17437.520.9680.055
tracking + every 10 frame-0.17638.830.9740.041
trackingevery 10 frame0.15738.830.9750.041
+ +Additional Keyframe Selection for Mapping For mapping quality, sufficient keyframes are required, but indiscriminate addition of keyframes causes a drop in tracking accuracy. To address this challenge and maximize both tracking accuracy and mapping quality, we pick mapping-only keyframes in addition to the original keyframes. Tab. 6 presents the result of the ablation study of this method. When using only keyframes without mapping-only keyframes, camera tracking accuracy suffers due to accumulated errors in scan-matching. However, employing our proposed method yields the best tracking accuracy and mapping quality. This shows that our method selectively adds keyframes essential for tracking, minimizing accumulated errors in scan-matching. While leveraging these benefits, we ensure sufficient primitives and images for training, which are crucial for map quality through the use of mapping-only keyframes. + +Table 7: Ablation of Methods for Avoiding Local Minima While Mapping on Replica. The results are the average of 8 scenes in the Replica dataset. + +
(a)(b)
Keyframe choice for learningPSNRSSIMLPIPSdensifying pruningATE [cm]PSNRSSIMLPIPS
recent 1 keyframe26.850.8720.200✓✓0.18838.650.973
covisible keyframes31.340.9240.120✓✗0.17037.900.971
random keyframes38.830.9750.041✗✓0.15738.830.975
✗✗0.15938.720.974
+ +Avoiding Local Minima while Mapping We employ two methods to prevent Gaussians from overfitting to the training image, which results in elongation to the viewpoint direction: (1) training with randomly selected keyframes and (2) Gaussian pruning. Tab. 7 presents the results of the ablation study on these methods. The worst rendering performance is observed when repeatedly training with the recently added keyframe, while using covisible keyframes for training yields better results. Covisible keyframes offer more diverse viewpoints for the Gaussians compared to a single keyframe, yet they still provide limited diversity in viewpoints. Random keyframes provide the most diverse viewpoints compared to other methods, leading to a significant improvement in map quality. In the vanilla 3DGS [15], Gaussian densifying/pruning techniques are utilized to manage Gaussians. Tab. 7 (b) shows the results of an ablation test of Gaussian densifying/pruning in our method. Since our method appropriately supplies Gaussians for scene representation, densifying is unnecessary. However, pruning improves tracking accuracy and map quality by removing Gaussians that have become elongated due to overfitting or are no longer essential for representing the map. + +# 5 Conclusion + +In this paper, we have proposed RGBD GS-ICP-SLAM, a dense representation SLAM system that leverages 3D Gaussian representation for high-fidelity spatial representation. We demonstrate that a fusion of G-ICP and 3DGS that utilizes a single 3D Gaussian map for both tracking and mapping yields mutual benefits. The exchange of Gaussians between tracking and mapping processes with scale alignments minimizes redundant computations and constructs an efficient system. Moreover, our dynamic keyframe selection method enhances both tracking and mapping performance. Through extensive experiments, the proposed approach presents state-of-the-art performance in spatial representation, camera pose estimation, and total system speed. + +Limitations. The proposed method achieved fast system speed by relying solely on depth for the 3D structure. In real-world, there are limitations in the quality of maps due to the depth noise in RGB-D sensors. Since the system speed is already exceptionally fast, it is expected that robust performance can be achieved by trading off a bit of speed to compensate for noisy depth images with relatively robust RGB information. + +Acknowledgment. This work was supported by the National Research Foundation of Korea (NRF) grant (no. RS-2024-00359937). + +# References + +1. Besl, P.J., McKay, N.D.: Method for registration of 3-d shapes. In: Sensor fusion IV: control paradigms and data structures. vol. 1611, pp. 586-606. Spie (1992) +2. Biber, P., Straβer, W.: The normal distributions transform: A new approach to laser scan matching. In: Proceedings 2003 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2003)(Cat. No. 03CH37453). vol. 3, pp. 2743-2748. IEEE (2003) +3. Bylow, E., Sturm, J., Kerl, C., Kahl, F., Cremers, D.: Real-time camera tracking and 3d reconstruction using signed distance functions. In: Robotics: Science and Systems. vol. 2, p. 2 (2013) +4. Campos, C., Elvira, R., Rodríguez, J.J.G., Montiel, J.M., Tardós, J.D.: Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics 37(6), 1874-1890 (2021) +5. Canelhas, D.R., Stoyanov, T., Lilienthal, A.J.: Sdf tracker: A parallel algorithm for on-line pose estimation and scene reconstruction from depth images. In: 2013 IEEE/RSJ International Conference on Intelligent Robots and Systems. pp. 3671-3676. IEEE (2013) +6. Ceriani, S., Sánchez, C., Taddei, P., Wolfart, E., Sequeira, V.: Pose interpolation slam for large maps using moving 3d sensors. In: 2015 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 750-757. IEEE (2015) +7. Chetverikov, D., Svirko, D., Stepanov, D., Krsek, P.: The trimmed iterative closest point algorithm. In: 2002 International Conference on Pattern Recognition. vol. 3, pp. 545-548. IEEE (2002) +8. Chung, C.M., Tseng, Y.C., Hsu, Y.C., Shi, X.Q., Hua, Y.H., Yeh, J.F., Chen, W.C., Chen, Y.T., Hsu, W.H.: Orbeez-slam: A real-time monocular visual slam with orb features and nef-realized mapping. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 9400-9406. IEEE (2023) +9. Czarnowski, J., Laidlow, T., Clark, R., Davison, A.J.: Deepfactors: Real-time probabilistic dense monocular slam. IEEE Robotics and Automation Letters 5(2), 721-728 (2020) +0. Dai, A., Nießner, M., Zollhöfer, M., Izadi, S., Theobalt, C.: Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics (ToG) 36(4), 1 (2017) +1. Grisetti, G., Kümmerle, R., Stachniss, C., Burgard, W.: A tutorial on graph-based slam. IEEE Intelligent Transportation Systems Magazine 2(4), 31-43 (2010) +2. Huang, H., Li, L., Cheng, H., Yeung, S.K.: Photo-slam: Real-time simultaneous localization and photorealistic mapping for monocular, stereo, and rgb-d cameras. arXiv preprint arXiv:2311.16728 (2023) +3. Johari, M.M., Carta, C., Fleuret, F.: Eslam: Efficient dense slam system based on hybrid representation of signed distance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17408-17419 (2023) +4. Keetha, N., Karhade, J., Jatavallabhula, K.M., Yang, G., Scherer, S., Ramanan, D., Luiten, J.: Splatam: Splat, track & map 3d gaussians for dense rgb-d slam. arXiv preprint arXiv:2312.02126 (2023) +5. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023) +6. Kerl, C., Sturm, J., Cremers, D.: Dense visual slam for rgb-d cameras. In: 2013 IEEE/RSJ International Conference on Intelligent Robots and Systems. pp. 2100-2106. IEEE (2013) +7. Koide, K., Yokozuka, M., Oishi, S., Banno, A.: Voxelized gicp for fast and accurate 3d point cloud registration. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 11054-11059. IEEE (2021) + +18. Kong, X., Liu, S., Taher, M., Davison, A.J.: vmap: Vectorised object mapping for neural field slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 952-961 (2023) +19. Lin, J., Zhang, F.: Loam livox: A fast, robust, high-precision lidar odometry and mapping package for lidars of small fov. In: 2020 IEEE International Conference on Robotics and Automation (ICRA). pp. 3126-3131. IEEE (2020) +20. Low, K.L.: Linear least-squares optimization for point-to-plane 1cp surface registration. Chapel Hill, University of North Carolina 4(10), 1-3 (2004) +21. Mallios, A., Ridao, P., Ribas, D., Hernández, E.: Scan matching slam in underwater environments. Autonomous Robots 36, 181-198 (2014) +22. Matsuki, H., Murai, R., Kelly, P.H., Davison, A.J.: Gaussian splatting slam. arXiv preprint arXiv:2312.06741 (2023) +23. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) +24. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022) +25. Mur-Artal, R., Montiel, J.M.M., Tardos, J.D.: Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics 31(5), 1147-1163 (2015) +26. Mur-Artal, R., Tardós, J.D.: Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras. IEEE transactions on robotics 33(5), 1255-1262 (2017) +27. Newcombe, R.A., Izadi, S., Hilliges, O., Molyneaux, D., Kim, D., Davison, A.J., Kohi, P., Shotton, J., Hodges, S., Fitzgibbon, A.: Kinectfusion: Real-time dense surface mapping and tracking. In: 2011 10th IEEE international symposium on mixed and augmented reality. pp. 127-136. IEEE (2011) +28. Nieto, J., Bailey, T., Nebot, E.: Recursive scan-matching slam. Robotics and Autonomous systems 55(1), 39-49 (2007) +29. Palieri, M., Morrell, B., Thakur, A., Ebadi, K., Nash, J., Chatterjee, A., Kanellakis, C., Carlone, L., Guaragnella, C., Agha-Mohammadi, A.a.: Locus: A multi-sensor lidar-centric solution for high-precision odometry and 3d mapping in real-time. IEEE Robotics and Automation Letters 6(2), 421-428 (2020) +30. Prisacariu, V.A., Kähler, O., Golodetz, S., Sapienza, M., Cavallari, T., Torr, P.H., Murray, D.W.: Infinitam v3: A framework for large-scale 3d reconstruction with loop closure. arXiv preprint arXiv:1708.00783 (2017) +31. Qin, T., Li, P., Shen, S.: Vins-mono: A robust and versatile monocular visual-inertial state estimator. IEEE Transactions on Robotics 34(4), 1004-1020 (2018) +32. Sandström, E., Li, Y., Van Gool, L., Oswald, M.R.: Point-slam: Dense neural point cloud-based slam. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18433-18444 (2023) +33. Segal, A., Haehnel, D., Thrun, S.: Generalized-icp. In: Robotics: science and systems. vol. 2, p. 435. Seattle, WA (2009) +34. Serafin, J., Grisetti, G.: Nicp: Dense normal based point cloud registration. In: 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 742-749. IEEE (2015) +35. Shan, T., Englot, B.: Lego-loam: Lightweight and ground-optimized lidar odometry and mapping on variable terrain. In: 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 4758-4765. IEEE (2018) +36. Shan, T., Englot, B., Meyers, D., Wang, W., Ratti, C., Rus, D.: Lio-sam: Tightly-coupled lidar inertial odometry via smoothing and mapping. In: 2020 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 5135-5142. IEEE (2020) + +37. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019) +38. Sturm, J., Engelhard, N., Endres, F., Burgard, W., Cremers, D.: A benchmark for the evaluation of rgb-d slam systems. In: 2012 IEEE/RSJ international conference on intelligent robots and systems. pp. 573-580. IEEE (2012) +39. Sucar, E., Liu, S., Ortiz, J., Davison, A.J.: imap: Implicit mapping and positioning in real-time. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6229-6238 (2021) +40. Taud, H., Mas, J.: Multilayer perceptron (mlp). Geomatic approaches for modeling land change scenarios pp. 451-455 (2018) +41. Vizzo, I., Guadagnino, T., Mersch, B., Wiesmann, L., Behley, J., Stachniss, C.: Kiss-icp: In defense of point-to-point icp-simple, accurate, and robust registration if done the right way. IEEE Robotics and Automation Letters 8(2), 1029–1036 (2023) +42. Whelan, T., Johannsson, H., Kaess, M., Leonard, J.J., McDonald, J.: Robust real-time visual odometry for dense rgb-d mapping. In: 2013 IEEE International Conference on Robotics and Automation. pp. 5724-5731. IEEE (2013) +43. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528 (2023) +44. Yan, C., Qu, D., Wang, D., Xu, D., Wang, Z., Zhao, B., Li, X.: Gs-slam: Dense visual slam with 3d gaussian splatting. arXiv preprint arXiv:2311.11700 (2023) +45. Yang, X., Li, H., Zhai, H., Ming, Y., Liu, Y., Zhang, G.: Vox-fusion: Dense tracking and mapping with voxel-based neural implicit representation. In: 2022 IEEE International Symposium on Mixed and Augmented Reality (ISMAR). pp. 499-507. IEEE (2022) +46. Zhang, J., Singh, S.: Loam: Lidar odometry and mapping in real-time. In: Robotics: Science and systems. vol. 2, pp. 1-9. Berkeley, CA (2014) +47. Zhu, Z., Peng, S., Larsson, V., Xu, W., Bao, H., Cui, Z., Oswald, M.R., Pollefeys, M.: Nice-slam: Neural implicit scalable encoding for slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12786-12796 (2022) \ No newline at end of file diff --git a/2024/RGBD GS-ICP SLAM/images.zip b/2024/RGBD GS-ICP SLAM/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..691d38f249eb5807c31c143043106d949e6ddab5 --- /dev/null +++ b/2024/RGBD GS-ICP SLAM/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f14940578cfab6d828a7244261e86c70c424f6b5bdf1dbbf634bc276de31877 +size 524875 diff --git a/2024/RGBD GS-ICP SLAM/layout.json b/2024/RGBD GS-ICP SLAM/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f8bc431bee9e497dcae07e101c6cdab90acca5da --- /dev/null +++ b/2024/RGBD GS-ICP SLAM/layout.json @@ -0,0 +1,7994 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 237, + 113, + 376, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 113, + 376, + 127 + ], + "spans": [ + { + "bbox": [ + 237, + 113, + 376, + 127 + ], + "type": "text", + "content": "RGBD GS-ICP SLAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 207, + 152, + 406, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 152, + 406, + 163 + ], + "spans": [ + { + "bbox": [ + 207, + 152, + 406, + 163 + ], + "type": "text", + "content": "Seongbo Ha, Jiung Yeon, and Hyeonwoo Yu" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 219, + 173, + 394, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 173, + 394, + 194 + ], + "spans": [ + { + "bbox": [ + 219, + 173, + 394, + 194 + ], + "type": "text", + "content": "Sungkyunkwan University, Suwon, South Korea \n{sobo3607, wcr12st, hwyu}@skku.edu" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 228, + 449, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 228, + 449, + 392 + ], + "spans": [ + { + "bbox": [ + 160, + 228, + 449, + 392 + ], + "type": "text", + "content": "Abstract. Simultaneous Localization and Mapping (SLAM) with dense representation plays a key role in robotics, Virtual Reality (VR), and Augmented Reality (AR) applications. Recent advancements in dense representation SLAM have highlighted the potential of leveraging neural scene representation and 3D Gaussian representation for high-fidelity spatial representation. In this paper, we propose a novel dense representation SLAM approach with a fusion of Generalized Iterative Closest Point (G-ICP) and 3D Gaussian Splatting (3DGS). In contrast to existing methods, we utilize a single Gaussian map for both tracking and mapping, resulting in mutual benefits. Through the exchange of covariances between tracking and mapping processes with scale alignment techniques, we minimize redundant computations and achieve an efficient system. Additionally, we enhance tracking accuracy and mapping quality through our keyframe selection methods. Experimental results demonstrate the effectiveness of our approach, showing incredibly fast speeds up to 107 FPS (for the entire system) and superior quality of the reconstructed map." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 392, + 463, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 392, + 463, + 414 + ], + "spans": [ + { + "bbox": [ + 160, + 392, + 463, + 414 + ], + "type": "text", + "content": "The code is available at: https://github.com/Lab-of-AI-and-Robotics/GS-ICP-SLAM Video is: https://youtu.be/ebHh_uMMxE" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 426, + 403, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 426, + 403, + 436 + ], + "spans": [ + { + "bbox": [ + 160, + 426, + 403, + 436 + ], + "type": "text", + "content": "Keywords: Coordinate-based 3D Representation " + }, + { + "bbox": [ + 160, + 426, + 403, + 436 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 426, + 403, + 436 + ], + "type": "text", + "content": " G-ICP " + }, + { + "bbox": [ + 160, + 426, + 403, + 436 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 426, + 403, + 436 + ], + "type": "text", + "content": " SLAM" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 459, + 218, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 459, + 218, + 470 + ], + "spans": [ + { + "bbox": [ + 133, + 459, + 218, + 470 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 485, + 481, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 485, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 133, + 485, + 481, + 569 + ], + "type": "text", + "content": "Visual Simultaneous Localization and Mapping (SLAM) is an algorithm that constructs maps of unknown environments while localizing poses of vision sensors simultaneously. As the influence of 3D visual SLAM in the fields of robotics, Virtual Reality (VR), and Augmented Reality (AR) has increased, higher rendering performance and more accurate trajectory are required. Thus, 3D reconstruction methods such as Signed Distance Field (SDF), and Truncated Signed Distance Field (TSDF) are utilized for traditional dense visual SLAM [3,5,10,27,30,42]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 570, + 481, + 665 + ], + "type": "text", + "content": "Recently, coordinate-based 3D Implicit Neural Representation (INR) [23] has been proposed for representing spatial information using neural radiance fields, showcasing high novel view synthesis capabilities and high-fidelity spatial representation prowess. Various approaches [8, 18, 32, 39, 45, 47] have been attempted to utilize INR for the real-time SLAM mapping process. However, INR requires computationally intensive raycasting to synthesize images, thus the rendering process in INR-based SLAM incurs significant time overhead, slowing down map optimization and rendering-loss based tracking. In contrast, coordinate-based 3D explicit representation such as 3D Gaussian" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 194, + 121, + 411, + 245 + ], + "blocks": [ + { + "bbox": [ + 194, + 121, + 411, + 245 + ], + "lines": [ + { + "bbox": [ + 194, + 121, + 411, + 245 + ], + "spans": [ + { + "bbox": [ + 194, + 121, + 411, + 245 + ], + "type": "image", + "image_path": "9fd7eda37ed65e7f9c152ef05d387a65623c1897dcffc3e55e57a45c5241df0c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 255, + 482, + 310 + ], + "lines": [ + { + "bbox": [ + 132, + 255, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 482, + 310 + ], + "type": "text", + "content": "Fig. 1: A comparison of PSNR with respect to FPS of entire system in recent research on SLAM algorithm utilizing dense representation such as neural scene representation and 3D Gaussian representation. Our method achieves state-of-the-art performance in rendering evaluation and FPS of entire system. Note that this FPS represents the overall system performance. Reported values are average of Replica 8 scenes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 343, + 481, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 481, + 392 + ], + "type": "text", + "content": "Splatting (3DGS) [15] represents the 3D space using 3D Gaussians as primitives, allowing for rendering speeds faster than NeRF using the rasterization method [43]. This rapid rendering capability of 3DGS results in fast optimization of 3D spatial information, making it suitable for dense representation in SLAM [12, 14, 22, 44]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 397, + 482, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 482, + 553 + ], + "type": "text", + "content": "Although 3DGS-based SLAM methods take advantages of the high-speed rendering, they fail to address the fundamental issue: the inability to directly utilize 3D explicit representations and the indirect tracking of 3D space through 2D image rendering. Even with the majority of current 3DGS-based SLAM [12, 14, 22, 32, 44] utilizing RGB-D data, the use of explicit representations is overlooked. For example, INR-based methods [32, 39, 47] and 3DGS-based methods [14, 22, 44] employ photometric-error based techniques, which estimate the optimal pose by iteratively minimizing the error between rendered and observed 2D images. However, due to limitations in tracking speed and performance, decoupled approaches [8, 12, 18] have been proposed incorporating well-crafted visual odometries [4, 26] into the tracking process. Although decoupled methods exhibit better performances in tracking by separating mapping and tracking processes, they require additional computational resources for the independent stages. Moreover, these methods need to store features in 3D space, unrelated to the 3DGS map." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 482, + 665 + ], + "type": "text", + "content": "There exists a method that allows Gaussian, an explicit representation, to be directly used for tracking. The well-known Generalized Iterative Closest Point (G-ICP) [17, 33] from the 3D scan matching family is simple yet efficient for fast tracking of 3D point clouds. During preprocessing, it only requires computing Gaussians for the current frame and the map. Given that the map in 3DGS utilizes Gaussians as an explicit representation in 3D space, using G-ICP for tracking allows for the direct utilization of the 3DGS map without the need for post-processing. Furthermore, the Gaussians of the current frame computed during tracking with G-ICP can also be directly utilized as an explicit representation in the 3DGS map, without additional computations." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "text", + "content": "Seongbo Ha et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 482, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 482, + 367 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 482, + 367 + ], + "type": "text", + "content": "Therefore, we propose a dense representation SLAM framework that integrates G-ICP and 3DGS, allowing them to complement each other by sharing explicit representations. Unlike traditional methods, our approach actively leverages 3D information through G-ICP for tracking. The proposed method is a coupled approach that shares a single map during tracking and mapping processes, while maintaining fast tracking speed akin to decoupled approaches. Previous works using decoupled methods [8, 12, 18] require separate maps containing ORB feature information for tracking, along with additional resources and computations needed to obtain ORB features. Moreover, the values obtained during tracking are not utilized in mapping. Our approach exploits the covariance of each point computed during the G-ICP-based tracking process as the initial state of the 3DGS mapping. Meanwhile, the 3D Gaussians in the 3DGS map are also mutually used as 3D points and their covariances which are essential for G-ICP-based tracking, discarding the necessity to recalculate the covariances of the map. This is possible because the covariances of points computed during the G-ICP process and the 3D Gaussians representing the map commonly contain information about the surrounding space. In other words, G-ICP and 3DGS can share the same Gaussian world. Therefore, our system minimizes unnecessary computations and facilitates efficient system configuration by mutually utilizing the key elements, Gaussians, between tracking and mapping processes. To ensure optimal performance in tracking and mapping sharing information between G-ICP and 3DGS, we also introduce several techniques such as scale alignment. To summarize, our contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 373, + 481, + 466 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 373, + 481, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 373, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 373, + 481, + 407 + ], + "type": "text", + "content": "- We present a real-time dense representation SLAM that combines G-ICP and 3DGS, achieving extremely high speed of the entire system (up to 107 FPS) and superior quality of the map." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 409, + 480, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 409, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 138, + 409, + 480, + 430 + ], + "type": "text", + "content": "- By incorporating G-ICP for tracking, our system utilizes 3D information actively and significantly reduces the time required for the tracking process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 433, + 480, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 433, + 480, + 466 + ], + "spans": [ + { + "bbox": [ + 138, + 433, + 480, + 466 + ], + "type": "text", + "content": "- Reduction of computational cost and facilitating rapid convergence of primitives of 3DGS is achieved by sharing the covariances of G-ICP and 3DGS with scale aligning techniques." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 133, + 486, + 224, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 486, + 224, + 498 + ], + "spans": [ + { + "bbox": [ + 133, + 486, + 224, + 498 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 510, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 510, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 510, + 482, + 665 + ], + "type": "text", + "content": "G-ICP Scan-matching focuses on the registration of two point clouds that observe similar environments by selecting a transformation matrix that minimizes errors between two point clouds [2]. Iterative Closest Point (ICP) [1] algorithm is a widely used and influential method in the scan-matching area. ICP iteratively estimates point correspondences and finds a transformation that minimizes Euclidean distance between corresponding points, thereby optimizing the registration of point clouds. Because of simplicity and speed of ICP, there are many follow-up studies [7, 20, 33, 34, 41]. To improve robustness, Trimmed-ICP [7], proposed a method for correspondence selection. Point-to-plane ICP [20] augmented robustness and accuracy by taking point-to-plane distance as objective function. G-ICP [33] introduced probabilistic models to generalize ICP yielding notable improvements in robustness and accuracy. Scan matching is a common technique used in SLAM for pose tracking [6, 11, 21, 28, 29], and for enhanced robustness, feature-based scan matching is often employed [19, 35, 36, 46]." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 363, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 363, + 91, + 447, + 100 + ], + "type": "text", + "content": "RGBD GS-ICP SLAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 481, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 481, + 176 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 481, + 176 + ], + "type": "text", + "content": "SLAM with Dense Representation SLAM with dense representation [9, 16] aims to construct maps in a dense form to enable interaction with the map in tasks such as AR, robotics, etc. To achieve this, classic approaches [3,5,10,27,30,42] creates dense maps by representing space in Signed Distance Field (SDF), and Truncated Signed Distance Field (TSDF), rather than in sparse forms such as point clouds or grids." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 177, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 177, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 177, + 482, + 308 + ], + "type": "text", + "content": "Recently, dense representation SLAM methods represent maps by utilizing NeRF [23], which demonstrates high spatial representation capabilities have been proposed. iMAP [39], NICE-SLAM [47], Point-SLAM [32], ESLAM [13] perform tracking by optimizing camera poses by reducing errors between synthesized and observed images. These methods have limitations in tracking speed and performance. Alternatively, Orbeez-SLAM [8], vMAP [18] incorporate well-crafted visual SLAM techniques into the tracking process, making mapping and tracking operate independently. Orbeez-SLAM utilizes ORB-SLAM2 [26] for tracking and instant-ngp [24] for learning spatial information to perform mapping. In a similar context, vMAP utilizes ORB-SLAM3 [4] for tracking while employing separate Multi-layer perceptrons [40] for each object to effectively represent the entire space." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 309, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 309, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 309, + 482, + 430 + ], + "type": "text", + "content": "3DGS represents space in explicit form using 3D gaussians as primitives. It offers a level of high-fidelity spatial representation similar to NeRF but provides significantly faster rendering speeds. To leverage this advantage to dense representation SLAM, several approaches [12, 14, 22, 44] utilizing 3DGS for space representation have been proposed. Among them, GS-SLAM [44], SPLTAM [14], Gaussian Splating SLAM [22] perform tracking using dense photometric error. While 3DGS shows significantly improved rendering performance compared to NeRF, the tracking methods based on dense photometric error still suffer from tracking speed. To tackle this problem, Photo-SLAM [12] integrates ORB-SLAM3 for tracking, in a similar manner to Orbeez-SLAM and vMAP." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 452, + 194, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 452, + 194, + 466 + ], + "spans": [ + { + "bbox": [ + 132, + 452, + 194, + 466 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "text", + "content": "To mutually benefit tracking and mapping, we introduce the fusion of G-ICP and GS, each representing the tracking and mapping processes, respectively. The key insight of our approach is that covariance can be considered as a fundamental common factor for this fusion. Suppose we have a point set (point cloud) " + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\{\\pmb{x}_m\\}_{m=1,\\dots,M}" + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "text", + "content": " and its corresponding covariance set " + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{C_m\\}_{m=1,\\dots,M}" + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\pmb{x} = [x,y,z]^T" + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "text", + "content": ". The covariance " + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "text", + "content": " of one 3D point " + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "text", + "content": " is given by computing covariance matrix of " + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "text", + "content": "-nearest neighbors of " + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "text", + "content": ". Let us define " + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\pmb{G} = \\{\\pmb{\\mathcal{X}},\\pmb{\\mathcal{C}}\\}" + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "text", + "content": " as a set of Gaussians. G-ICP aims to find a transformation " + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "text", + "content": " that maximally aligns the source Gaussians (current frame) " + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\pmb{G}^s = \\{\\pmb{\\mathcal{X}}^s,\\pmb{\\mathcal{C}}^s\\}" + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "text", + "content": " and the target Gaussians (map) " + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\pmb{G}^t = \\{\\pmb{\\mathcal{X}}^t,\\pmb{\\mathcal{C}}^t\\}" + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "text", + "content": ". Assume we know the correspondences between " + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\pmb{\\mathcal{X}}^s" + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\pmb{\\mathcal{X}}^t" + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "text", + "content": " determined by nearest neighbor search. For example, we have " + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\{\\pmb{x}_i^s\\}_{i=1,\\dots,N} \\subset \\pmb{\\mathcal{X}}^s" + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\{C_i^s\\}_{i=1,\\dots,N} \\subset \\pmb{\\mathcal{C}}^s" + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\{\\pmb{x}_i^t\\}_{i=1,\\dots,N} \\subset \\pmb{\\mathcal{X}}^t" + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\{C_i^t\\}_{i=1,\\dots,N} \\subset \\pmb{\\mathcal{C}}^t" + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^s" + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "text", + "content": " is associated with " + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^t" + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "text", + "content": ". To find the optimal transform " + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{T}^*" + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "text", + "content": ", we exploit not the single point but the distribution of that point defined as a Gaussian distribution: " + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i \\sim \\mathcal{N}(\\hat{\\pmb{x}}_i,C_i)" + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "inline_equation", + "content": "d_i = \\pmb{x}_i^t - \\mathbf{T}\\pmb{x}_i^s" + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "text", + "content": " be the error term, and if we assume that there is an optimal transformation " + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{T}^*" + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "text", + "content": ", it is clear that " + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^t = \\mathbf{T}^*\\pmb{x}_i^s" + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "text", + "content": " thus " + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\hat{d}_i = 0" + }, + { + "bbox": [ + 132, + 483, + 482, + 668 + ], + "type": "text", + "content": ". Since" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "text", + "content": "Seongbo Ha et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": "we assume " + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": " is a Gaussian random variable, " + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": " is also a Gaussian random variable as the following:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 225, + 158, + 387, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 158, + 387, + 205 + ], + "spans": [ + { + "bbox": [ + 225, + 158, + 387, + 205 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} d _ {i} \\sim \\mathcal {N} \\left(\\hat {d} _ {i}, C _ {i} ^ {t} + \\mathbf {T} ^ {*} C _ {i} ^ {s} \\left(\\mathbf {T} ^ {*}\\right) ^ {T}\\right) \\\\ = \\mathcal {N} \\left(\\hat {x} _ {i} ^ {t} - \\mathbf {T} ^ {*} \\hat {x} _ {i} ^ {s}, C _ {i} ^ {t} + \\mathbf {T} ^ {*} C _ {i} ^ {s} \\left(\\mathbf {T} ^ {*}\\right) ^ {T}\\right) \\\\ = \\mathcal {N} (0, C _ {i} ^ {t} + \\mathbf {T} ^ {*} C _ {i} ^ {s} (\\mathbf {T} ^ {*}) ^ {T}). \\\\ \\end{array}", + "image_path": "c5d8386cd1861588cbdadbcf75bdf34b707b5b93756303dddb35d22693852200.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 212, + 479, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 212, + 479, + 236 + ], + "spans": [ + { + "bbox": [ + 131, + 212, + 479, + 236 + ], + "type": "text", + "content": "To find the optimal transform " + }, + { + "bbox": [ + 131, + 212, + 479, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{T}^*" + }, + { + "bbox": [ + 131, + 212, + 479, + 236 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 131, + 212, + 479, + 236 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}^s" + }, + { + "bbox": [ + 131, + 212, + 479, + 236 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 212, + 479, + 236 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}^t" + }, + { + "bbox": [ + 131, + 212, + 479, + 236 + ], + "type": "text", + "content": ", we use maximum likelihood estimation (MLE) as the following:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 208, + 243, + 481, + 310 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 243, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 208, + 243, + 481, + 310 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {T} ^ {*} = \\underset {\\mathbf {T}} {\\operatorname {a r g m a x}} \\prod_ {i} ^ {N} p \\left(d _ {i}\\right) = \\underset {\\mathbf {T}} {\\operatorname {a r g m a x}} \\sum_ {i} ^ {N} \\log p \\left(d _ {i}\\right) \\\\ = \\underset {\\mathbf {T}} {\\operatorname {a r g m i n}} \\sum_ {i} ^ {N} d _ {i} ^ {T} \\left(C _ {i} ^ {B} + \\mathbf {T} C _ {i} ^ {A} \\mathbf {T} ^ {T}\\right) ^ {- 1} d _ {i}. \\tag {1} \\\\ \\end{array}", + "image_path": "28a5fe870b2708e27be7a64d78277e49d8274080b205f6156747d522127d616f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 316, + 485, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 316, + 485, + 339 + ], + "spans": [ + { + "bbox": [ + 131, + 316, + 485, + 339 + ], + "type": "text", + "content": "Therefore, " + }, + { + "bbox": [ + 131, + 316, + 485, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{T}^*" + }, + { + "bbox": [ + 131, + 316, + 485, + 339 + ], + "type": "text", + "content": " can be used as the relative pose between the current frame " + }, + { + "bbox": [ + 131, + 316, + 485, + 339 + ], + "type": "inline_equation", + "content": "G^{s} = \\{\\mathcal{X}^{s},\\mathcal{C}^{s}\\}" + }, + { + "bbox": [ + 131, + 316, + 485, + 339 + ], + "type": "text", + "content": " and the map " + }, + { + "bbox": [ + 131, + 316, + 485, + 339 + ], + "type": "inline_equation", + "content": "G^{t} = \\{\\mathcal{X}^{t},\\mathcal{C}^{t}\\}" + }, + { + "bbox": [ + 131, + 316, + 485, + 339 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 340, + 481, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 340, + 481, + 389 + ], + "spans": [ + { + "bbox": [ + 131, + 340, + 481, + 389 + ], + "type": "text", + "content": "Meanwhile, for mapping purposes, GS also relies on Gaussians " + }, + { + "bbox": [ + 131, + 340, + 481, + 389 + ], + "type": "inline_equation", + "content": "G = \\{\\mathcal{X},\\mathcal{C}\\}" + }, + { + "bbox": [ + 131, + 340, + 481, + 389 + ], + "type": "text", + "content": " of the 3D scene representation. Different from G-ICP, GS aims to find the optimal coordinates of the Gaussians " + }, + { + "bbox": [ + 131, + 340, + 481, + 389 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^* = \\{x_m^*\\}_{m=1,\\dots,M}" + }, + { + "bbox": [ + 131, + 340, + 481, + 389 + ], + "type": "text", + "content": " and the optimal covariances " + }, + { + "bbox": [ + 131, + 340, + 481, + 389 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^* = \\{C_m^*\\}_{m=1,\\dots,M}" + }, + { + "bbox": [ + 131, + 340, + 481, + 389 + ], + "type": "text", + "content": " as the following:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 394, + 485, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 394, + 485, + 415 + ], + "spans": [ + { + "bbox": [ + 132, + 394, + 485, + 415 + ], + "type": "interline_equation", + "content": "\\boldsymbol{\\mathcal{X}}^{*},\\boldsymbol{\\mathcal{C}}^{*},\\boldsymbol{H}^{*},\\boldsymbol{O}^{*} = \\operatorname *{argmin}_{\\boldsymbol {\\mathcal{X}},\\boldsymbol {\\mathcal{C}},\\boldsymbol {H},\\boldsymbol{O}}\\lambda_{I_{1}}\\boldsymbol{\\mathcal{L}}_{1}\\left(I,I_{gt}\\right) + \\lambda_{I_{2}}\\boldsymbol{\\mathcal{L}}_{D - SSIM}\\left(I,I_{gt}\\right) + \\lambda_{D}\\boldsymbol{\\mathcal{L}}_{1}\\left(D,D_{gt}\\right)", + "image_path": "098a43f124facf3bd7f5d8681b3569ccb9ed6673aca112a044c85ccae42b5003.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 421, + 479, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 421, + 479, + 456 + ], + "spans": [ + { + "bbox": [ + 131, + 421, + 479, + 456 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 421, + 479, + 456 + ], + "type": "inline_equation", + "content": "H = \\{h_m\\}_{m=1,\\dots,M}" + }, + { + "bbox": [ + 131, + 421, + 479, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 421, + 479, + 456 + ], + "type": "inline_equation", + "content": "O = \\{o_m\\}_{m=1,\\dots,M}" + }, + { + "bbox": [ + 131, + 421, + 479, + 456 + ], + "type": "text", + "content": " are the color set and the opacity set of 3D points which are for the RGBD image rendering. " + }, + { + "bbox": [ + 131, + 421, + 479, + 456 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 131, + 421, + 479, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 421, + 479, + 456 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 131, + 421, + 479, + 456 + ], + "type": "text", + "content": " are the rendered RGB and depth images obtained by performing rasterization using " + }, + { + "bbox": [ + 131, + 421, + 479, + 456 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 131, + 421, + 479, + 456 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 421, + 479, + 456 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 131, + 421, + 479, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 421, + 479, + 456 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 131, + 421, + 479, + 456 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 457, + 481, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 457, + 481, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 457, + 481, + 635 + ], + "type": "text", + "content": "Note that in G-ICP and GS, the key common factor is Gaussians " + }, + { + "bbox": [ + 131, + 457, + 481, + 635 + ], + "type": "inline_equation", + "content": "G = \\{\\mathcal{X},\\mathcal{C}\\}" + }, + { + "bbox": [ + 131, + 457, + 481, + 635 + ], + "type": "text", + "content": ", allowing these Gaussians to be shared mutually. During G-ICP tracking, the covariance of each frame is computed. Hence, when adding keyframes to expand the 3DGS map, there is no need to recalculate " + }, + { + "bbox": [ + 131, + 457, + 481, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 131, + 457, + 481, + 635 + ], + "type": "text", + "content": " for every expansion. Also, G-ICP does not need to compute the covariances of the map because our GS map already contains Gaussians. Moreover, G-ICP, by aligning frames based on 3D geometric structure, inherently initializes a certain number of points and their coordinates that aptly represent the 3D structure. Therefore, it brings about an effect where an appropriate number of points and their poses are initialized, well-suited for depicting the 3D structure, significantly reducing the learning time to find the optimal pose " + }, + { + "bbox": [ + 131, + 457, + 481, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^*" + }, + { + "bbox": [ + 131, + 457, + 481, + 635 + ], + "type": "text", + "content": " and the optimal covariance " + }, + { + "bbox": [ + 131, + 457, + 481, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^*" + }, + { + "bbox": [ + 131, + 457, + 481, + 635 + ], + "type": "text", + "content": " of Gaussians " + }, + { + "bbox": [ + 131, + 457, + 481, + 635 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 131, + 457, + 481, + 635 + ], + "type": "text", + "content": " in GS. Additionally, in the same vein, calculations like densifying or opacity reset to adjust the number of 3D points in GS become unnecessary. Consequently, by sharing a common source, each process becomes mutually beneficial, and the speed of execution accelerates due to reduced redundant computations. Our method can be simplified as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 136, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 136, + 641, + 479, + 665 + ], + "type": "text", + "content": "1. Use G-ICP to align the current frame with the 3DGS map which contains covariance (solely need to compute the covariance for the current frame)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 362, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 362, + 91, + 447, + 100 + ], + "type": "text", + "content": "RGBD GS-ICP SLAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 115, + 471, + 208 + ], + "blocks": [ + { + "bbox": [ + 140, + 115, + 471, + 208 + ], + "lines": [ + { + "bbox": [ + 140, + 115, + 471, + 208 + ], + "spans": [ + { + "bbox": [ + 140, + 115, + 471, + 208 + ], + "type": "image", + "image_path": "1ad7b8524217766017f170291f86933956bf6b83ec241d6c565bef0ae7db4a10.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 215, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 304 + ], + "type": "text", + "content": "Fig. 2: System Overview. The input of our system is RGBD frame. We generate a point cloud by downsampling and reprojecting the current depth image and utilize it in the GICP process. During the GICP process, we create source Gaussians from the point cloud and estimate the current camera pose by aligning them with target Gaussians, which are a subset of the 3DGS map. If the current frame is identified as a keyframe or a mapping-only keyframe, we add the source Gaussians to the 3DGS map as new primitives. Meanwhile, in the mapping process, we optimize the Gaussians along with the color and opacity set of the Gaussians concurrently with the tracking process." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 325, + 480, + 361 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 135, + 325, + 480, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 325, + 480, + 349 + ], + "spans": [ + { + "bbox": [ + 135, + 325, + 480, + 349 + ], + "type": "text", + "content": "2. When adding keyframes to the 3DGS map, utilize the covariance computed in GICP during tracking (no need for densifying or opacity reset)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 349, + 220, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 349, + 220, + 361 + ], + "spans": [ + { + "bbox": [ + 136, + 349, + 220, + 361 + ], + "type": "text", + "content": "3. Repeat steps 1-2." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 365, + 482, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 390 + ], + "type": "text", + "content": "The overview of our system is shown in Fig. 2. The detailed implementations of the proposed technique are introduced in the following Section." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 405, + 228, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 405, + 228, + 418 + ], + "spans": [ + { + "bbox": [ + 131, + 405, + 228, + 418 + ], + "type": "text", + "content": "3.1 G-ICP Tracking" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 422, + 482, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 471 + ], + "type": "text", + "content": "Scale Regularization When using G-ICP for tracking, aligning the scale of the current frame with the map enables high-performance camera pose estimation [33]. Given the covariance " + }, + { + "bbox": [ + 130, + 422, + 482, + 471 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 422, + 482, + 471 + ], + "type": "text", + "content": ", scale " + }, + { + "bbox": [ + 130, + 422, + 482, + 471 + ], + "type": "inline_equation", + "content": "\\boldsymbol{S} = [s_2, s_1, s_0]^T" + }, + { + "bbox": [ + 130, + 422, + 482, + 471 + ], + "type": "text", + "content": " is given by the singular value decomposition (SVD) as the following:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 276, + 476, + 481, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 476, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 276, + 476, + 481, + 491 + ], + "type": "interline_equation", + "content": "C = \\boldsymbol {R} \\boldsymbol {\\Lambda} ^ {2} \\boldsymbol {R} ^ {T} \\tag {2}", + "image_path": "62291fcf00d6ffba94ddf0d2964c32ae7540628b5fd1273265fac30acf557953.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 497, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 497, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\Lambda = \\text{diag}(s_2, s_1, s_0) \\forall s_2 > s_1 > s_0" + }, + { + "bbox": [ + 130, + 497, + 482, + 544 + ], + "type": "text", + "content": " is a scale matrix and " + }, + { + "bbox": [ + 130, + 497, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\pmb{R}" + }, + { + "bbox": [ + 130, + 497, + 482, + 544 + ], + "type": "text", + "content": " is the orientation of the Gaussian. To achieve the robust scan matching performance, point-to-plane ICP or voxelized point-to-plane ICP adopt a regularizing method that makes the scale as " + }, + { + "bbox": [ + 130, + 497, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\pmb{S} = [1, 1, \\epsilon]^T" + }, + { + "bbox": [ + 130, + 497, + 482, + 544 + ], + "type": "text", + "content": " in order to treat each Gaussian as a plane-like distribution." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 544, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 617 + ], + "type": "text", + "content": "In our framework, we utilize target Gaussians " + }, + { + "bbox": [ + 130, + 544, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\pmb{G}^t = \\{\\pmb{x}^t,\\pmb{c}^t\\}" + }, + { + "bbox": [ + 130, + 544, + 482, + 617 + ], + "type": "text", + "content": " from the 3DGS map for tracking. Here, " + }, + { + "bbox": [ + 130, + 544, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\pmb{c}^t" + }, + { + "bbox": [ + 130, + 544, + 482, + 617 + ], + "type": "text", + "content": " is optimized to represent the scene accurately, " + }, + { + "bbox": [ + 130, + 544, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\pmb{G}^t" + }, + { + "bbox": [ + 130, + 544, + 482, + 617 + ], + "type": "text", + "content": " involves not only planes but also lines, corners, and other features. Therefore, while tracking, instead of regularizing the scale into a plane-based form, preserving the original characteristics of the target Gaussians while performing regularization is more appropriate. Thus, in our framework, we propose the following ellipsoid regularization:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 234, + 622, + 481, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 622, + 481, + 647 + ], + "spans": [ + { + "bbox": [ + 234, + 622, + 481, + 647 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Lambda} ^ {\\prime} = \\frac {1}{m e d i a n (\\boldsymbol {S})} \\operatorname {d i a g} \\left(s _ {2}, s _ {1}, s _ {0}\\right) \\tag {3}", + "image_path": "7d3345be40c906081bb55d2577a2893c24c89b82489abdf73147477dd0c1d5c0.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 652, + 334, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 652, + 334, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 652, + 334, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 652, + 334, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{\\Lambda}^{\\prime}" + }, + { + "bbox": [ + 131, + 652, + 334, + 666 + ], + "type": "text", + "content": " is the ellipsoid-regularized scale matrix." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "text", + "content": "Seongbo Ha et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 123, + 294, + 219 + ], + "blocks": [ + { + "bbox": [ + 142, + 123, + 294, + 219 + ], + "lines": [ + { + "bbox": [ + 142, + 123, + 294, + 219 + ], + "spans": [ + { + "bbox": [ + 142, + 123, + 294, + 219 + ], + "type": "image", + "image_path": "4fecd2b2fe34b0ccf75cc0977939a3930ab9e131e3d788583ce912a62a12304e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 221, + 225, + 230 + ], + "lines": [ + { + "bbox": [ + 216, + 221, + 225, + 230 + ], + "spans": [ + { + "bbox": [ + 216, + 221, + 225, + 230 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 316, + 123, + 470, + 218 + ], + "blocks": [ + { + "bbox": [ + 316, + 123, + 470, + 218 + ], + "lines": [ + { + "bbox": [ + 316, + 123, + 470, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 123, + 470, + 218 + ], + "type": "image", + "image_path": "ae889a95482c4e04d4a00af4e0c0799196234122024407ae5202eec361e2cdc3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 220, + 402, + 229 + ], + "lines": [ + { + "bbox": [ + 392, + 220, + 402, + 229 + ], + "spans": [ + { + "bbox": [ + 392, + 220, + 402, + 229 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 239, + 482, + 284 + ], + "lines": [ + { + "bbox": [ + 131, + 239, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 131, + 239, + 482, + 284 + ], + "type": "text", + "content": "Fig. 3: Tracking Accuracy Comparison Based on Keyframe Selection Methods. The reported values represent the average results across eight scenes from the Replica dataset [37]. When selecting keyframes every n frame (depicted in blue), the tracking accuracy is notably low. Conversely, our keyframe selection method yielded the highest tracking accuracy." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 306, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 482, + 498 + ], + "type": "text", + "content": "Keyframe Selection Similar to [25,31], we perform dynamic keyframe selection. Considering the geometric structure is critical for both G-ICP tracking and GS mapping, we exploit the geometric correspondence computed from G-ICP. As Eqn. (1) denotes, we can get the distance between " + }, + { + "bbox": [ + 130, + 306, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_i^s" + }, + { + "bbox": [ + 130, + 306, + 482, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 306, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_i^t" + }, + { + "bbox": [ + 130, + 306, + 482, + 498 + ], + "type": "text", + "content": " as an interim result of G-ICP procedure without additional calculations. So we can efficiently determine the correspondence by setting the distance threshold. By considering the proportion of correspondences between the current frame and map, we select keyframes while taking into account the characteristics of the scene. In more detail, if the proportion in the current frame goes below specific thresholds, the frame is selected as a keyframe. This allows us to achieve consistent tracking performance while simultaneously maintaining a consistent density of Gaussians " + }, + { + "bbox": [ + 130, + 306, + 482, + 498 + ], + "type": "inline_equation", + "content": "G = \\{\\mathcal{X},\\mathcal{C}\\}" + }, + { + "bbox": [ + 130, + 306, + 482, + 498 + ], + "type": "text", + "content": " added to the map. The effect of this keyframe selection is shown in Fig. 3 (a). Note that, from selected keyframe, only Gaussians that do not overlap with the existing current map are considered as target Gaussians. Errors from discrepancies during tracking are directly incorporated into the map, leading to accumulated inaccuracies and reduced tracking precision. This strategy ensures more accurate tracking results and improves the overall robustness of the system." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 514, + 212, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 212, + 527 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 212, + 527 + ], + "type": "text", + "content": "3.2 GS Mapping" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Scale Aligning Suppose we have a fully trained 3DGS map. Since the 3DGS map is trained to represent the scene from any viewpoint, Gaussians " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " should be uniformly distributed according to the structure of the 3D space. However, single frames obtained from radiance sensors such as RGBD cameras or LiDAR yield sparse representations of the 3D space as the distance from the sensor increases, due to the inherent characteristics of the sensor. In other words, as the distance from the sensor increases, the spacing between 3D points widens, and the scale of the corresponding covariance calculated based on the " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "-nearest neighbors becomes excessively large. Adding such single frames as keyframes to the map during real-time training of GS mapping leads to an imbalance in the scale of Gaussians, ultimately resulting in a decrease in mapping performance. Therefore, to fundamentally alleviate the problem, our proposed method suggests the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 363, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 363, + 91, + 447, + 100 + ], + "type": "text", + "content": "RGBD GS-ICP SLAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 117, + 242, + 219 + ], + "blocks": [ + { + "bbox": [ + 141, + 117, + 242, + 219 + ], + "lines": [ + { + "bbox": [ + 141, + 117, + 242, + 219 + ], + "spans": [ + { + "bbox": [ + 141, + 117, + 242, + 219 + ], + "type": "image", + "image_path": "986f348d0f8449baaf0b34adfe9c8d720af5be7e608131bc957215d3cc8fb216.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 228, + 483, + 274 + ], + "lines": [ + { + "bbox": [ + 131, + 228, + 483, + 274 + ], + "spans": [ + { + "bbox": [ + 131, + 228, + 483, + 274 + ], + "type": "text", + "content": "Fig. 4: Separated Keyframe Selection on Replica office4. We demonstrate that a small number of tracking keyframes yield accurate trajectory estimation (case 1), while a large number of mapping keyframes result in high rendering performance (case 2). Thus, our method adopts case 3 that selects tracking keyframe and mapping keyframe separately at different intervals." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 256, + 117, + 354, + 217 + ], + "blocks": [ + { + "bbox": [ + 256, + 117, + 354, + 217 + ], + "lines": [ + { + "bbox": [ + 256, + 117, + 354, + 217 + ], + "spans": [ + { + "bbox": [ + 256, + 117, + 354, + 217 + ], + "type": "image", + "image_path": "d1280fceeaacaf642d05cfefae8894bcfa8c6c6e7c6bd82c421f693ed078b089.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 369, + 116, + 474, + 217 + ], + "blocks": [ + { + "bbox": [ + 369, + 116, + 474, + 217 + ], + "lines": [ + { + "bbox": [ + 369, + 116, + 474, + 217 + ], + "spans": [ + { + "bbox": [ + 369, + 116, + 474, + 217 + ], + "type": "image", + "image_path": "c4a61d2aa57e457d2ad0ebbdc8857eaeb666befccf32c8f388bc4a2c12768665.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 293, + 343, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 343, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 343, + 304 + ], + "type": "text", + "content": "following scale normalization for the current frame:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 280, + 310, + 334, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 280, + 310, + 334, + 334 + ], + "spans": [ + { + "bbox": [ + 280, + 310, + 334, + 334 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\varLambda} ^ {\\prime \\prime} = \\frac {1}{z ^ {p}} \\boldsymbol {\\varLambda}.", + "image_path": "7e6efa25587f134f39212c62dded49afb477a9de44798a9998de09597e79383c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 338, + 482, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 338, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 131, + 338, + 482, + 389 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 131, + 338, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\Lambda''" + }, + { + "bbox": [ + 131, + 338, + 482, + 389 + ], + "type": "text", + "content": " is the normalized scale matrix and " + }, + { + "bbox": [ + 131, + 338, + 482, + 389 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 131, + 338, + 482, + 389 + ], + "type": "text", + "content": " is the parameter which is empirically determined. By adding such scale-normalized keyframes to the map to expand it, not only does the GS mapping performance increase, but also the performance of G-ICP tracking improves." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 402, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 510 + ], + "type": "text", + "content": "Additional Keyframe Selection for Mapping Mapping based on GS benefits from having a diverse set of frames from various viewpoints for training, as performance improves with a larger number of available frames. However, increasing the number of selected keyframes during tracking can lead to a degradation in tracking performance due to accumulated errors, as is common with G-ICP-based methods. To mitigate this issue, we propose adding mapping-only keyframe selection in addition to the existing keyframe selection process. In other words, tracking continues to proceed from selected keyframes in the existing tracking process, while mapping utilizes both the original keyframes and additional mapping-only keyframes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 510, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 510, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 510, + 483, + 666 + ], + "type": "text", + "content": "The overview of the additional keyframe selection is illustrated in Fig. 4. Cases 1 and 2 represent scenarios without considering mapping-only keyframes. In Case 1, to increase the number of keyframes, if no keyframe satisfying the criteria is found up to the " + }, + { + "bbox": [ + 131, + 510, + 483, + 666 + ], + "type": "inline_equation", + "content": "30^{th}" + }, + { + "bbox": [ + 131, + 510, + 483, + 666 + ], + "type": "text", + "content": " frame from the previous keyframe, the " + }, + { + "bbox": [ + 131, + 510, + 483, + 666 + ], + "type": "inline_equation", + "content": "30^{th}" + }, + { + "bbox": [ + 131, + 510, + 483, + 666 + ], + "type": "text", + "content": " frame is selected as a keyframe. However, since this may not yield sufficient mapping performance, as shown in Case 2, adding an additional keyframe at the " + }, + { + "bbox": [ + 131, + 510, + 483, + 666 + ], + "type": "inline_equation", + "content": "10^{th}" + }, + { + "bbox": [ + 131, + 510, + 483, + 666 + ], + "type": "text", + "content": " frame can be considered. While this approach improves mapping performance as desired, it leads to a decline in tracking performance due to accumulated scan matching errors. To alleviate this, as shown in Case 3, we propose adding the " + }, + { + "bbox": [ + 131, + 510, + 483, + 666 + ], + "type": "inline_equation", + "content": "10^{th}" + }, + { + "bbox": [ + 131, + 510, + 483, + 666 + ], + "type": "text", + "content": " frame as a mapping-only frame, allowing us to maximize tracking performance while simultaneously maximizing mapping performance. The effect of the proposed scheme, Case 3, is shown in Fig. 3 (b). Note that since G-ICP, which performs tracking, computes covariance for all frames anyway, there is no difference in the overall computational load." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "text", + "content": "Seongbo Ha et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "Avoiding Local Minima while Mapping GS SLAM differs from traditional GS in that it operates in real-time and often lacks sufficient observations of the scene. Moreover, since it operates in real-time, focusing solely on images from the current viewpoint or nearby viewpoints during training can easily lead to local minima and degrade mapping quality [22]. For instance, GS simply considers image rendering from specific viewpoints, thus if we continuously train only on a specific viewpoint, the scale of Gaussians in the map tends to elongate in the direction of that viewpoint, leading to local minima." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 201, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 482, + 260 + ], + "type": "text", + "content": "To overcome this challenge, at each training iteration, we adopt a strategy of randomly choosing one keyframe for learning among the selected keyframes so far, ensuring that both the current observed scene and the entire map are uniformly learned. Additionally, we prune Gaussians that fall into local minima during training to improve mapping performance and ensure robust tracking through geometric preservation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 280, + 218, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 280, + 218, + 294 + ], + "spans": [ + { + "bbox": [ + 131, + 280, + 218, + 294 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 306, + 243, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 306, + 243, + 319 + ], + "spans": [ + { + "bbox": [ + 131, + 306, + 243, + 319 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 327, + 482, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 448 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 448 + ], + "type": "text", + "content": "We evaluate our method on Replica dataset [37] and TUM dataset [38]. Replica dataset contains synthetic scenes and high-quality RGB/depth images rendered from these scenes. TUM dataset includes images captured in the real world, with significant noise and blur, resulting in poor quality. Particularly, many parts of the depth images suffer from information loss. We validate the effectiveness of our approach by evaluating it on both synthetic and real-world datasets. All experiments are performed on a desktop with a Ryzen 7 7800x3d CPU, 32GB RAM, and an NVIDIA RTX 4090 24GB GPU. To evaluate camera tracking accuracy, we use the Root Mean Square Error (RMSE) of the Absolute Trajectory Error (ATE). For the quality of the reconstructed map, we report standard photometric rendering quality metrics (PSNR, SSIM, and LPIPS)." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 137, + 534, + 479, + 631 + ], + "blocks": [ + { + "bbox": [ + 130, + 468, + 482, + 523 + ], + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 523 + ], + "type": "text", + "content": "Table 1: Tracking Performance on Replica and TUM-RGBD (ATE RMSE " + }, + { + "bbox": [ + 130, + 468, + 482, + 523 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 130, + 468, + 482, + 523 + ], + "type": "text", + "content": " [cm]). Our method archives state-of-the-art tracking accuracy on Replica. On TUM, Among coupled methods, we achieve competitive performance. The results of GS-SLAM/Gaussian Splitting SLAM are taken from [44]/ [22]. And the results of ORB-SLAM3 and Photo-SLAM are obtained from [12]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 534, + 479, + 631 + ], + "lines": [ + { + "bbox": [ + 137, + 534, + 479, + 631 + ], + "spans": [ + { + "bbox": [ + 137, + 534, + 479, + 631 + ], + "type": "table", + "html": "
MethodReplicaTUM
R0R1R2Of0Of1Of2Of3Of4Avg.fr1/deskfr2/xyzfr3/officeAvg.
DecoupledORB-SLAM3 [4]-------1.81.70.41.71.3
Photo-SLAM [12]-------0.602.60.31.01.3
CoupledNICE-SLAM* [47]1.611.481.610.950.811.461.761.691.422.82.17.24.0
Point-SLAM* [32]0.590.510.320.450.460.480.610.870.542.71.33.92.6
GS-SLAM [44]0.480.530.330.520.410.590.460.700.503.31.36.63.7
SplaTAM* [14]0.290.350.280.490.210.310.340.570.363.31.35.13.2
Gaussian Splitting SLAM* [22]0.330.220.290.360.190.250.120.810.321.51.41.51.5
Ours (limited to 30 FPS)0.150.160.110.180.120.170.160.210.162.71.82.72.4
", + "image_path": "8dc1fdab1ca9e844d2e0399401a2e83ae1955bdbd948f443268cc6ed55db8e03.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 318, + 631, + 476, + 639 + ], + "lines": [ + { + "bbox": [ + 318, + 631, + 476, + 639 + ], + "spans": [ + { + "bbox": [ + 318, + 631, + 476, + 639 + ], + "type": "text", + "content": "* denotes the reproduced results by running official code." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 363, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 363, + 91, + 447, + 100 + ], + "type": "text", + "content": "RGBD GS-ICP SLAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 277, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 277, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 277, + 129 + ], + "type": "text", + "content": "4.2 Camera Tracking Accuracy" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 483, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 483, + 288 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 483, + 288 + ], + "type": "text", + "content": "Tab. 1 shows the tracking accuracy of our method compared to other approaches on both synthetic [37] and real-world [38] datasets. On Replica [37], the proposed method achieved state-of-the-art (SOTA) performance across all scenes, reducing the trajectory error by more than " + }, + { + "bbox": [ + 130, + 143, + 483, + 288 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 143, + 483, + 288 + ], + "type": "text", + "content": " compared to the previous SOTA. This result arises from the fact that our proposed method actively utilizes 3D information by employing G-ICP [33] for tracking, unlike other methods that perform tracking based on errors in 2D space. On TUM [38], our method shows competitive tracking accuracy among coupled methods [14,22,32,44,47] which conduct tracking and mapping based on a single map. Note that the results of Gaussian Splitting SLAM are evaluated only on keyframes. Decoupled method [12] and its baseline [4] show better performance, but these approaches require additional resources for storing separate map information and costly computations for extracting features solely for the tracking process." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 144, + 343, + 473, + 643 + ], + "blocks": [ + { + "bbox": [ + 132, + 310, + 483, + 334 + ], + "lines": [ + { + "bbox": [ + 132, + 310, + 483, + 334 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 483, + 334 + ], + "type": "text", + "content": "Table 2: Evaluation of System Speed and Map Quality on Replica. Our method outperforms all other frameworks in both system speed and quality of the reconstructed map." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 144, + 343, + 473, + 643 + ], + "lines": [ + { + "bbox": [ + 144, + 343, + 473, + 643 + ], + "spans": [ + { + "bbox": [ + 144, + 343, + 473, + 643 + ], + "type": "table", + "html": "
MethodsMetricsR0R1R2Of0Of1Of2Of3Of4Avg.
Orbeez-SLAM* [8]PSNR[dB]↑12.1315.2815.8717.5919.2610.3011.5512.6514.33
SSIM ↑0.700.790.810.800.670.780.760.840.77
LPIPS ↓0.610.480.510.520.410.550.540.530.53
FPS ↑24.0524.7824.0019.2226.7423.8824.4526.1124.15
Point-SLAM* [32]PSNR[dB]↑33.3834.1036.3238.7239.3134.2234.1034.8235.62
SSIM ↑0.980.980.990.990.990.960.960.980.98
LPIPS ↓0.100.120.100.090.110.150.120.130.11
FPS ↑0.260.300.310.330.340.300.280.300.30
GS-SLAM [44]PSNR[dB]↑31.5632.8632.5938.7041.1732.3632.0332.9234.27
SSIM ↑0.970.970.970.990.990.980.970.970.98
LPIPS ↓0.090.080.090.050.030.090.110.110.08
FPS ↑8.34-------8.34
SplaTAM* [14]PSNR[dB]↑32.6033.5534.8338.0939.0231.9529.5331.5533.89
SSIM ↑0.980.970.980.980.980.970.950.950.97
LPIPS ↓0.070.100.070.090.090.100.120.150.10
FPS ↑0.240.190.190.200.220.270.260.240.23
Gaussian Splitting SLAM [22]PSNR[dB]↑34.8336.4337.4939.9542.0936.2436.7036.0737.50
SSIM ↑0.950.960.970.970.980.960.960.960.96
LPIPS ↓0.070.080.080.070.060.080.070.100.08
FPS ↑-----1.1---
Ours (no tracking speed limit)PSNR[dB]↑32.2035.3634.4240.3140.7533.8534.0836.4735.93
SSIM ↑0.940.960.960.980.980.960.950.960.96
LPIPS ↓0.080.070.080.050.050.070.070.070.07
FPS ↑100.9884.92103.3899.10107.0695.6097.2096.6698.11
Ours (limited to 30 FPS)PSNR[dB]↑35.3737.8038.5043.1343.2636.9336.9038.7538.83
SSIM ↑0.960.970.980.990.990.970.970.970.98
LPIPS ↓0.050.050.050.030.030.040.040.050.04
FPS ↑29.9729.9829.9829.9829.9929.9729.9729.9729.98
", + "image_path": "3473d062f1bac9bf8fd26289dda744c5a8ac26f58dd808c49064673f9f9f1afd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 215, + 644, + 471, + 663 + ], + "lines": [ + { + "bbox": [ + 215, + 644, + 471, + 663 + ], + "spans": [ + { + "bbox": [ + 215, + 644, + 471, + 663 + ], + "type": "text", + "content": "* denotes the reproduced results by running official code. The result of GS-SLAM and Gaussian Splitting SLAM is taken from [44] and [22]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "Seongbo Ha et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 194, + 146, + 422, + 223 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 3: Evaluation of System Speed and Map Quality on TUM-RGBD. Proposed method shows incredible system speed and competitive map quality." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 194, + 146, + 422, + 223 + ], + "lines": [ + { + "bbox": [ + 194, + 146, + 422, + 223 + ], + "spans": [ + { + "bbox": [ + 194, + 146, + 422, + 223 + ], + "type": "table", + "html": "
MethodsPSNR[dB] ↑SSIM ↑LPIPS ↓FPS ↑
NICE-SLAM* [47]14.100.5740.3950.08
Point-SLAM* [32]21.400.7380.4470.22
Photo-SLAM [12]21.900.7630.187-
SplaTAM* [14]23.460.9060.1560.32
Ours (unlimited tracking speed)19.620.7500.24073.92
Ours (limited to 30 FPS)20.720.7680.21829.99
", + "image_path": "bf4ff4d8d94b24eb3b818df2ba0b8278ecc3fe7b0125945f977753a3e3860ad4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 242, + 224, + 420, + 241 + ], + "lines": [ + { + "bbox": [ + 242, + 224, + 420, + 241 + ], + "spans": [ + { + "bbox": [ + 242, + 224, + 420, + 241 + ], + "type": "text", + "content": "* denotes the reproduced results by running official code. The result of Photo-SLAM is taken from [12]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 288, + 366, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 288, + 366, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 288, + 366, + 300 + ], + "type": "text", + "content": "4.3 System Speed and Quality of Reconstructed Map" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 312, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 456 + ], + "type": "text", + "content": "Tab. 2 and Tab. 3 demonstrate the FPS of the systems and quality of the reconstructed map on Replica [37] and TUM [38] dataset, respectively. FPS of the system is calculated by dividing the total number of frames by the total time. Note that this FPS represents the overall system performance including tracking and mapping processes, rather than just individual components. Photo-SLAM [12] reported the tracking FPS of their system, but did not include the system FPS, so we leave that field blank in the Tab. 3. Since our method implements mapping and tracking processes to operate in parallel, the number of map optimization iterations varies depending on the tracking speed, consequently affecting the quality of the reconstructed map. So we evaluate our system in two cases: one where the tracking speed is limited to the typical sensor input speed of 30 FPS, and the other where the tracking speed is not limited. Please refer to the supplementary material for results across various FPS and geometric quality evaluations." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 136, + 487, + 479, + 599 + ], + "blocks": [ + { + "bbox": [ + 136, + 487, + 479, + 599 + ], + "lines": [ + { + "bbox": [ + 136, + 487, + 479, + 599 + ], + "spans": [ + { + "bbox": [ + 136, + 487, + 479, + 599 + ], + "type": "image", + "image_path": "e969dc582149105ea922270bd1c198849e75b97ae6be9ed350ecc6ee3639711c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 607, + 482, + 662 + ], + "lines": [ + { + "bbox": [ + 130, + 607, + 482, + 662 + ], + "spans": [ + { + "bbox": [ + 130, + 607, + 482, + 662 + ], + "type": "text", + "content": "Fig. 5: Comparison of Rendering Results. In the first scene, SplaTAM [14] failed to reconstruct the pillow and lamp and Point-SLAM [32] failed to represent the detailed pattern of the pillow. In the second case, SplaTAM and Point-SLAM failed to accurately reconstruct the details of the clock. However, our method exhibits rendering results that closely resemble the ground truth image, demonstrating a high level of accuracy." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 363, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 363, + 91, + 447, + 100 + ], + "type": "text", + "content": "RGBD GS-ICP SLAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 357 + ], + "type": "text", + "content": "In the limited case (30 FPS), our system achieves SOTA map quality across all scenes in the Replica dataset. When the tracking speed is not limited, our method shows extremely fast system speed up to 107 FPS. This speed is more than four times faster than that of the Orbeez-SLAM [8], based on the ORB-SLAM2 [26]. Remarkably, even under this condition, our approach maintains superior map quality and outperforms other methods. The proposed method effectively integrates Gaussians for map representation with appropriate initial states and has better detailed representation capability compared to NeRF-based methods. As shown in Fig. 5, SplaTAM suffers from incorporating appropriate Gaussians into the map, leading to regions of poor quality. Point-SLAM exhibits areas where detailed representation is lacking. In TUM-RGBD, our method demonstrates remarkably fast system speed with competitive map quality, showcasing its practicality in real-world scenarios. Compared to SplaTAM [14], our approach shows a slight decrease in map quality, with a PSNR reduction of approximately " + }, + { + "bbox": [ + 130, + 116, + 482, + 357 + ], + "type": "inline_equation", + "content": "11.7\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 357 + ], + "type": "text", + "content": ". However, we achieve significant speed improvements, with a speedup of approximately 91.6 times under limited conditions and up to 227 times at maximum speed. The reason is that, unlike Replica, TUM dataset is captured with old-fashioned sensors, resulting in noisy and substantial information loss in depth images. While SplaTAM treats these factors by adding new Gaussians only in regions with significant depth loss, our method utilizes all structural information from the depth image, focusing on accurate tracking and rapid system rather than treating such factors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 371, + 220, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 371, + 220, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 371, + 220, + 384 + ], + "type": "text", + "content": "4.4 Ablation study" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 203, + 443, + 410, + 500 + ], + "blocks": [ + { + "bbox": [ + 130, + 411, + 482, + 433 + ], + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 433 + ], + "type": "text", + "content": "Table 4: Scale Regularization Ablation on TUM-RGBD. Reported results are the average ATE RMSE " + }, + { + "bbox": [ + 130, + 411, + 482, + 433 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 130, + 411, + 482, + 433 + ], + "type": "text", + "content": " [cm] of 3 scenes in TUM-RGBD dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 203, + 443, + 410, + 500 + ], + "lines": [ + { + "bbox": [ + 203, + 443, + 410, + 500 + ], + "spans": [ + { + "bbox": [ + 203, + 443, + 410, + 500 + ], + "type": "table", + "html": "
Scale Regularizationfr1-deskfr2-xyzfr3-officeAvg.
X136.32196.61376.40236.54
Plane65.635.5316.2129.12
ellipsoid2.661.772.672.37
", + "image_path": "2a5d0ffed9daf7f393f0f1eee15d218300941b5d02c7fd14bedf11e514372e92.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 653 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 653 + ], + "type": "text", + "content": "Scale Regularization Tab. 4 shows the result of the ablation study on scale regularization while tracking. We perform scale regularization to enhance tracking accuracy by ensuring that the scales of Gaussians in the current frame and the map are within a similar range. Despite the application of plane regularization resulting in some improvement in tracking accuracy, it remains inadequate. This is because the plane regularization method considers all the Gaussians as planes, ignoring the characteristics of Gaussians existing in the map, which are optimized to represent the surrounding space through the mapping process. On the other hand, when applying the proposed ellipsoid regularization, the best tracking accuracy is achieved. This is attributed to the ability of the proposed ellipsoid regularization to consider the characteristics of Gaussians existing in the map while performing regularization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 653, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 653, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 653, + 481, + 665 + ], + "type": "text", + "content": "Scale Aligning We implement scale aligning to reduce the difference between scales" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "Seongbo Ha et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 212, + 146, + 403, + 205 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 5: Ablation of Scale Aligning on Replica. The results are the average of 8 scenes in the Replica dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 212, + 146, + 403, + 205 + ], + "lines": [ + { + "bbox": [ + 212, + 146, + 403, + 205 + ], + "spans": [ + { + "bbox": [ + 212, + 146, + 403, + 205 + ], + "type": "table", + "html": "
Covs from G-ICPRe-initATE ↓PSNR ↑SSIM ↑LPIPS ↓
XX8.89324.810.8400.380
✓X0.25833.210.9220.177
✓constant0.15837.330.9640.074
✓z1.50.15738.830.9750.041
", + "image_path": "e249124b1b27d2afeaafaea8ae5ad79428d1c87c172951eb9eed0c303ec20a62.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 227, + 482, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 227, + 482, + 406 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 482, + 406 + ], + "type": "text", + "content": "of existing Gaussians in the map and newly added Gaussians. By doing this, we ensure that scales of newly added Gaussians closely resemble the optimal value they will reach through optimization. To demonstrate the effectiveness of this approach, we tested our method in various cases and Tab. 5 shows the results. When not utilizing the covariances computed during the G-ICP [33] process, we calculate the scales of Gaussians by using simple-knn module of vanilla 3DGS [15], and set rotations as identity. In this case, a significant drop in tracking performance appeared. This is because the newly added Gaussians lack sufficient prior information about the 3D structure, causing them to be utilized as target Gaussians without being optimally optimized in space. In the case of using the computed covariance from the G-ICP process, the tracking and mapping performance improves. Furthermore, the best performance is observed when dividing the scale by " + }, + { + "bbox": [ + 130, + 227, + 482, + 406 + ], + "type": "inline_equation", + "content": "z^{1.5}" + }, + { + "bbox": [ + 130, + 227, + 482, + 406 + ], + "type": "text", + "content": ". These results indicate that utilizing covariances calculated during G-ICP procedure is effective for both tracking and mapping, and scale aligning facilitates smooth connection from G-ICP to 3DGS. Further analysis of the advantages of this connection is included in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 168, + 459, + 447, + 512 + ], + "blocks": [ + { + "bbox": [ + 132, + 427, + 480, + 449 + ], + "lines": [ + { + "bbox": [ + 132, + 427, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 427, + 480, + 449 + ], + "type": "text", + "content": "Table 6: Keyframe and Mapping-Only Keyframe Selection Ablation on Replica. Reported results are the average of 8 scenes in the Replica dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 168, + 459, + 447, + 512 + ], + "lines": [ + { + "bbox": [ + 168, + 459, + 447, + 512 + ], + "spans": [ + { + "bbox": [ + 168, + 459, + 447, + 512 + ], + "type": "table", + "html": "
KeyframeMapping-Only KeyframeATE [cm] ↓PSNR ↑SSIM ↑LPIPS ↓
tracking + every 30 frame-0.17437.520.9680.055
tracking + every 10 frame-0.17638.830.9740.041
trackingevery 10 frame0.15738.830.9750.041
", + "image_path": "418d1cc2e0de27e8e81062fa263c5fa2400efe56b73e3e1f65b5a314a7aca559.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "content": "Additional Keyframe Selection for Mapping For mapping quality, sufficient keyframes are required, but indiscriminate addition of keyframes causes a drop in tracking accuracy. To address this challenge and maximize both tracking accuracy and mapping quality, we pick mapping-only keyframes in addition to the original keyframes. Tab. 6 presents the result of the ablation study of this method. When using only keyframes without mapping-only keyframes, camera tracking accuracy suffers due to accumulated errors in scan-matching. However, employing our proposed method yields the best tracking accuracy and mapping quality. This shows that our method selectively adds keyframes essential for tracking, minimizing accumulated errors in scan-matching. While leveraging these benefits, we ensure sufficient primitives and images for training, which are crucial for map quality through the use of mapping-only keyframes." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 363, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 363, + 91, + 447, + 100 + ], + "type": "text", + "content": "RGBD GS-ICP SLAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 155, + 480, + 226 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 138 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 138 + ], + "type": "text", + "content": "Table 7: Ablation of Methods for Avoiding Local Minima While Mapping on Replica. The results are the average of 8 scenes in the Replica dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 155, + 480, + 226 + ], + "lines": [ + { + "bbox": [ + 134, + 155, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 134, + 155, + 480, + 226 + ], + "type": "table", + "html": "
(a)(b)
Keyframe choice for learningPSNRSSIMLPIPSdensifying pruningATE [cm]PSNRSSIMLPIPS
recent 1 keyframe26.850.8720.200✓✓0.18838.650.973
covisible keyframes31.340.9240.120✓✗0.17037.900.971
random keyframes38.830.9750.041✗✓0.15738.830.975
✗✗0.15938.720.974
", + "image_path": "d71c09b959536ce1569f17ca350c1ff83168c436d9592c760b8cf9671f7e8acf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 248, + 483, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 483, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 483, + 429 + ], + "type": "text", + "content": "Avoiding Local Minima while Mapping We employ two methods to prevent Gaussians from overfitting to the training image, which results in elongation to the viewpoint direction: (1) training with randomly selected keyframes and (2) Gaussian pruning. Tab. 7 presents the results of the ablation study on these methods. The worst rendering performance is observed when repeatedly training with the recently added keyframe, while using covisible keyframes for training yields better results. Covisible keyframes offer more diverse viewpoints for the Gaussians compared to a single keyframe, yet they still provide limited diversity in viewpoints. Random keyframes provide the most diverse viewpoints compared to other methods, leading to a significant improvement in map quality. In the vanilla 3DGS [15], Gaussian densifying/pruning techniques are utilized to manage Gaussians. Tab. 7 (b) shows the results of an ablation test of Gaussian densifying/pruning in our method. Since our method appropriately supplies Gaussians for scene representation, densifying is unnecessary. However, pruning improves tracking accuracy and map quality by removing Gaussians that have become elongated due to overfitting or are no longer essential for representing the map." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 447, + 212, + 461 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 212, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 212, + 461 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 473, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 582 + ], + "type": "text", + "content": "In this paper, we have proposed RGBD GS-ICP-SLAM, a dense representation SLAM system that leverages 3D Gaussian representation for high-fidelity spatial representation. We demonstrate that a fusion of G-ICP and 3DGS that utilizes a single 3D Gaussian map for both tracking and mapping yields mutual benefits. The exchange of Gaussians between tracking and mapping processes with scale alignments minimizes redundant computations and constructs an efficient system. Moreover, our dynamic keyframe selection method enhances both tracking and mapping performance. Through extensive experiments, the proposed approach presents state-of-the-art performance in spatial representation, camera pose estimation, and total system speed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 641 + ], + "type": "text", + "content": "Limitations. The proposed method achieved fast system speed by relying solely on depth for the 3D structure. In real-world, there are limitations in the quality of maps due to the depth noise in RGB-D sensors. Since the system speed is already exceptionally fast, it is expected that robust performance can be achieved by trading off a bit of speed to compensate for noisy depth images with relatively robust RGB information." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "type": "text", + "content": "Acknowledgment. This work was supported by the National Research Foundation of Korea (NRF) grant (no. RS-2024-00359937)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "Seongbo Ha et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 190, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 190, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 190, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 141, + 480, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 480, + 164 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 480, + 164 + ], + "type": "text", + "content": "1. Besl, P.J., McKay, N.D.: Method for registration of 3-d shapes. In: Sensor fusion IV: control paradigms and data structures. vol. 1611, pp. 586-606. Spie (1992)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 165, + 480, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 165, + 480, + 198 + ], + "spans": [ + { + "bbox": [ + 138, + 165, + 480, + 198 + ], + "type": "text", + "content": "2. Biber, P., Straβer, W.: The normal distributions transform: A new approach to laser scan matching. In: Proceedings 2003 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2003)(Cat. No. 03CH37453). vol. 3, pp. 2743-2748. IEEE (2003)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 198, + 480, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 198, + 480, + 231 + ], + "spans": [ + { + "bbox": [ + 138, + 198, + 480, + 231 + ], + "type": "text", + "content": "3. Bylow, E., Sturm, J., Kerl, C., Kahl, F., Cremers, D.: Real-time camera tracking and 3d reconstruction using signed distance functions. In: Robotics: Science and Systems. vol. 2, p. 2 (2013)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 232, + 480, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 232, + 480, + 264 + ], + "spans": [ + { + "bbox": [ + 138, + 232, + 480, + 264 + ], + "type": "text", + "content": "4. Campos, C., Elvira, R., Rodríguez, J.J.G., Montiel, J.M., Tardós, J.D.: Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics 37(6), 1874-1890 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 264, + 480, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 264, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 138, + 264, + 480, + 297 + ], + "type": "text", + "content": "5. Canelhas, D.R., Stoyanov, T., Lilienthal, A.J.: Sdf tracker: A parallel algorithm for on-line pose estimation and scene reconstruction from depth images. In: 2013 IEEE/RSJ International Conference on Intelligent Robots and Systems. pp. 3671-3676. IEEE (2013)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 298, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 298, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 298, + 480, + 331 + ], + "type": "text", + "content": "6. Ceriani, S., Sánchez, C., Taddei, P., Wolfart, E., Sequeira, V.: Pose interpolation slam for large maps using moving 3d sensors. In: 2015 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 750-757. IEEE (2015)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 332, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 480, + 364 + ], + "type": "text", + "content": "7. Chetverikov, D., Svirko, D., Stepanov, D., Krsek, P.: The trimmed iterative closest point algorithm. In: 2002 International Conference on Pattern Recognition. vol. 3, pp. 545-548. IEEE (2002)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 365, + 480, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 480, + 408 + ], + "type": "text", + "content": "8. Chung, C.M., Tseng, Y.C., Hsu, Y.C., Shi, X.Q., Hua, Y.H., Yeh, J.F., Chen, W.C., Chen, Y.T., Hsu, W.H.: Orbeez-slam: A real-time monocular visual slam with orb features and nef-realized mapping. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 9400-9406. IEEE (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 409, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 409, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 138, + 409, + 480, + 430 + ], + "type": "text", + "content": "9. Czarnowski, J., Laidlow, T., Clark, R., Davison, A.J.: Deepfactors: Real-time probabilistic dense monocular slam. IEEE Robotics and Automation Letters 5(2), 721-728 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 432, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 432, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 432, + 480, + 464 + ], + "type": "text", + "content": "0. Dai, A., Nießner, M., Zollhöfer, M., Izadi, S., Theobalt, C.: Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics (ToG) 36(4), 1 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 465, + 480, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 465, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 465, + 480, + 487 + ], + "type": "text", + "content": "1. Grisetti, G., Kümmerle, R., Stachniss, C., Burgard, W.: A tutorial on graph-based slam. IEEE Intelligent Transportation Systems Magazine 2(4), 31-43 (2010)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 487, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 480, + 520 + ], + "type": "text", + "content": "2. Huang, H., Li, L., Cheng, H., Yeung, S.K.: Photo-slam: Real-time simultaneous localization and photorealistic mapping for monocular, stereo, and rgb-d cameras. arXiv preprint arXiv:2311.16728 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 521, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 480, + 553 + ], + "type": "text", + "content": "3. Johari, M.M., Carta, C., Fleuret, F.: Eslam: Efficient dense slam system based on hybrid representation of signed distance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17408-17419 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 554, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 554, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 138, + 554, + 480, + 586 + ], + "type": "text", + "content": "4. Keetha, N., Karhade, J., Jatavallabhula, K.M., Yang, G., Scherer, S., Ramanan, D., Luiten, J.: Splatam: Splat, track & map 3d gaussians for dense rgb-d slam. arXiv preprint arXiv:2312.02126 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 587, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 587, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 587, + 480, + 609 + ], + "type": "text", + "content": "5. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 610, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 480, + 632 + ], + "type": "text", + "content": "6. Kerl, C., Sturm, J., Cremers, D.: Dense visual slam for rgb-d cameras. In: 2013 IEEE/RSJ International Conference on Intelligent Robots and Systems. pp. 2100-2106. IEEE (2013)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "7. Koide, K., Yokozuka, M., Oishi, S., Banno, A.: Voxelized gicp for fast and accurate 3d point cloud registration. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 11054-11059. IEEE (2021)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 363, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 363, + 91, + 447, + 100 + ], + "type": "text", + "content": "RGBD GS-ICP SLAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "type": "text", + "content": "18. Kong, X., Liu, S., Taher, M., Davison, A.J.: vmap: Vectorised object mapping for neural field slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 952-961 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "text", + "content": "19. Lin, J., Zhang, F.: Loam livox: A fast, robust, high-precision lidar odometry and mapping package for lidars of small fov. In: 2020 IEEE International Conference on Robotics and Automation (ICRA). pp. 3126-3131. IEEE (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "text", + "content": "20. Low, K.L.: Linear least-squares optimization for point-to-plane 1cp surface registration. Chapel Hill, University of North Carolina 4(10), 1-3 (2004)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 225 + ], + "type": "text", + "content": "21. Mallios, A., Ridao, P., Ribas, D., Hernández, E.: Scan matching slam in underwater environments. Autonomous Robots 36, 181-198 (2014)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 247 + ], + "type": "text", + "content": "22. Matsuki, H., Murai, R., Kelly, P.H., Davison, A.J.: Gaussian splatting slam. arXiv preprint arXiv:2312.06741 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "23. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "type": "text", + "content": "24. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "text", + "content": "25. Mur-Artal, R., Montiel, J.M.M., Tardos, J.D.: Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics 31(5), 1147-1163 (2015)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "type": "text", + "content": "26. Mur-Artal, R., Tardós, J.D.: Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras. IEEE transactions on robotics 33(5), 1255-1262 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 348, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 481, + 390 + ], + "type": "text", + "content": "27. Newcombe, R.A., Izadi, S., Hilliges, O., Molyneaux, D., Kim, D., Davison, A.J., Kohi, P., Shotton, J., Hodges, S., Fitzgibbon, A.: Kinectfusion: Real-time dense surface mapping and tracking. In: 2011 10th IEEE international symposium on mixed and augmented reality. pp. 127-136. IEEE (2011)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "text", + "content": "28. Nieto, J., Bailey, T., Nebot, E.: Recursive scan-matching slam. Robotics and Autonomous systems 55(1), 39-49 (2007)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 414, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 456 + ], + "type": "text", + "content": "29. Palieri, M., Morrell, B., Thakur, A., Ebadi, K., Nash, J., Chatterjee, A., Kanellakis, C., Carlone, L., Guaragnella, C., Agha-Mohammadi, A.a.: Locus: A multi-sensor lidar-centric solution for high-precision odometry and 3d mapping in real-time. IEEE Robotics and Automation Letters 6(2), 421-428 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "text", + "content": "30. Prisacariu, V.A., Kähler, O., Golodetz, S., Sapienza, M., Cavallari, T., Torr, P.H., Murray, D.W.: Infinitam v3: A framework for large-scale 3d reconstruction with loop closure. arXiv preprint arXiv:1708.00783 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 491, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 511 + ], + "type": "text", + "content": "31. Qin, T., Li, P., Shen, S.: Vins-mono: A robust and versatile monocular visual-inertial state estimator. IEEE Transactions on Robotics 34(4), 1004-1020 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "text", + "content": "32. Sandström, E., Li, Y., Van Gool, L., Oswald, M.R.: Point-slam: Dense neural point cloud-based slam. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18433-18444 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "type": "text", + "content": "33. Segal, A., Haehnel, D., Thrun, S.: Generalized-icp. In: Robotics: science and systems. vol. 2, p. 435. Seattle, WA (2009)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "text", + "content": "34. Serafin, J., Grisetti, G.: Nicp: Dense normal based point cloud registration. In: 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 742-749. IEEE (2015)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "text", + "content": "35. Shan, T., Englot, B.: Lego-loam: Lightweight and ground-optimized lidar odometry and mapping on variable terrain. In: 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 4758-4765. IEEE (2018)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "text", + "content": "36. Shan, T., Englot, B., Meyers, D., Wang, W., Ratti, C., Rus, D.: Lio-sam: Tightly-coupled lidar inertial odometry via smoothing and mapping. In: 2020 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 5135-5142. IEEE (2020)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "text", + "content": "Seongbo Ha et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 435 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "37. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "text", + "content": "38. Sturm, J., Engelhard, N., Endres, F., Burgard, W., Cremers, D.: A benchmark for the evaluation of rgb-d slam systems. In: 2012 IEEE/RSJ international conference on intelligent robots and systems. pp. 573-580. IEEE (2012)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "text", + "content": "39. Sucar, E., Liu, S., Ortiz, J., Davison, A.J.: imap: Implicit mapping and positioning in real-time. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6229-6238 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "text", + "content": "40. Taud, H., Mas, J.: Multilayer perceptron (mlp). Geomatic approaches for modeling land change scenarios pp. 451-455 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "text", + "content": "41. Vizzo, I., Guadagnino, T., Mersch, B., Wiesmann, L., Behley, J., Stachniss, C.: Kiss-icp: In defense of point-to-point icp-simple, accurate, and robust registration if done the right way. IEEE Robotics and Automation Letters 8(2), 1029–1036 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "type": "text", + "content": "42. Whelan, T., Johannsson, H., Kaess, M., Leonard, J.J., McDonald, J.: Robust real-time visual odometry for dense rgb-d mapping. In: 2013 IEEE International Conference on Robotics and Automation. pp. 5724-5731. IEEE (2013)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 303, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 481, + 325 + ], + "type": "text", + "content": "43. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 346 + ], + "type": "text", + "content": "44. Yan, C., Qu, D., Wang, D., Xu, D., Wang, Z., Zhao, B., Li, X.: Gs-slam: Dense visual slam with 3d gaussian splatting. arXiv preprint arXiv:2311.11700 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "45. Yang, X., Li, H., Zhai, H., Ming, Y., Liu, Y., Zhang, G.: Vox-fusion: Dense tracking and mapping with voxel-based neural implicit representation. In: 2022 IEEE International Symposium on Mixed and Augmented Reality (ISMAR). pp. 499-507. IEEE (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "text", + "content": "46. Zhang, J., Singh, S.: Loam: Lidar odometry and mapping in real-time. In: Robotics: Science and systems. vol. 2, pp. 1-9. Berkeley, CA (2014)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "type": "text", + "content": "47. Zhu, Z., Peng, S., Larsson, V., Xu, W., Bao, H., Cui, Z., Oswald, M.R., Pollefeys, M.: Nice-slam: Neural implicit scalable encoding for slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12786-12796 (2022)" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 363, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 363, + 91, + 447, + 100 + ], + "type": "text", + "content": "RGBD GS-ICP SLAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/RGNet_ A Unified Clip Retrieval and Grounding Network for Long Videos/f091b053-bc8d-4374-9687-69b5bf80ade6_content_list.json b/2024/RGNet_ A Unified Clip Retrieval and Grounding Network for Long Videos/f091b053-bc8d-4374-9687-69b5bf80ade6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a02240b2504386ac230ed6e149dc576a11c10008 --- /dev/null +++ b/2024/RGNet_ A Unified Clip Retrieval and Grounding Network for Long Videos/f091b053-bc8d-4374-9687-69b5bf80ade6_content_list.json @@ -0,0 +1,1726 @@ +[ + { + "type": "text", + "text": "RGNet: A Unified Clip Retrieval and Grounding Network for Long Videos", + "text_level": 1, + "bbox": [ + 223, + 140, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tanveer Hannan $^{1,2\\star}$ Md Mohaiminul Islam $^{3}$ Thomas Seidl $^{1,2}$ Gedas Bertasius $^{3}$", + "bbox": [ + 233, + 212, + 769, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 LMU Munich", + "bbox": [ + 446, + 253, + 553, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 MCML", + "bbox": [ + 467, + 268, + 534, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 UNC Chapel Hill", + "bbox": [ + 434, + 282, + 566, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Locating specific moments within long videos (20-120 minutes) presents a significant challenge, akin to finding a needle in a haystack. Adapting existing short video (5-30 seconds) grounding methods to this problem yields poor performance. Since most real-life videos, such as those on YouTube and AR/VR, are lengthy, addressing this issue is crucial. Existing methods typically operate in two stages: clip retrieval and grounding. However, this disjoint process limits the retrieval module's fine-grained event understanding, crucial for specific moment detection. We propose RGNet which deeply integrates clip retrieval and grounding into a single network capable of processing long videos into multiple granular levels, e.g., clips and frames. Its core component is a novel transformer encoder, RG-Encoder, that unifies the two stages through shared features and mutual optimization. The encoder incorporates a sparse attention mechanism and an attention loss to model both granularity jointly. Moreover, we introduce a contrastive clip sampling technique to mimic the long video paradigm closely during training. RGNet surpasses prior methods, showcasing state-of-the-art performance on long video temporal grounding (LVTG) datasets MAD and Ego4D. The code is released at https://github.com/Tanveer81/RGNet.", + "bbox": [ + 261, + 338, + 743, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Long Video Temporal Grounding $\\cdot$ Moment Localization", + "bbox": [ + 261, + 614, + 727, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 657, + 375, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The exponential rise in online videos has created a demand for effective content retrieval systems. The retrieval results of user-defined queries, particularly in long videos (20-120 minutes), often require locating specific moments. Most existing solutions [5,41,53,55], tailored for short videos (5-30 seconds), struggle when applied to hour-long videos, slowing down the retrieval and impacting the quality of the results [40, 54]. This challenge emerges because it is very difficult to locate short moments in a long video based on text queries, a task known as Long Video Temporal Grounding (LVTG) [14, 40].", + "bbox": [ + 212, + 691, + 787, + 815 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author: hannan@dbis.ifi.lmu.de", + "bbox": [ + 217, + 824, + 540, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a4aa5a3c60a7093dea15bbe7642f13a4f87b7a9a0be3ffc447826395f03b1919.jpg", + "image_caption": [ + "Fig. 1: Overview of RGNet. It predicts the moment boundary specified by textual queries from an hour-long video. First, our proposed RG-Encoder maps the video and text features to a joint space and retrieves the relevant clip feature. The subsequent grounding decoder processes the retrieved features to predict the beginning and end times of the moment. The encoder parallelly operates at multiple levels of granularity (e.g., clip and frame) to achieve an end-to-end-solution." + ], + "image_footnote": [], + "bbox": [ + 287, + 142, + 718, + 324 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A straightforward solution [16,37] for the LVTG task is to divide the video into shorter clips, retrieve the most relevant one, and apply a grounding network to predict the moment. However, the grounding network cannot rectify failure in the clip retrieval phase.", + "bbox": [ + 212, + 426, + 787, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Empirically, we evaluate the impact of the two stages in Sec. 5.3 and identify clip retrieval as the primary factor for poor performance.", + "bbox": [ + 212, + 488, + 785, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The existing methods have a separate module for clip retrieval, which is disjoint from their grounding network (see Fig. 2). They typically follow a text-video retrieval [13,23,29] technique to select the relevant clip. However, video retrieval only requires a high-level understanding of video topics. For example, a typical video retrieval query is \"A movie about a family surviving in a farmhouse\". In contrast, an example clip retrieval query in Fig. 1 is \"Find the moment when the mum hangs up laundry outside the farmhouse\". These specific moments from a long video require fine-grained event understanding. Thus, video retrieval models are suboptimal for these fine-grained moment localization tasks.", + "bbox": [ + 212, + 520, + 787, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We attribute this disjoint retrieval to the poor performance of these models. However, unifying clip selection and grounding is challenging due to their distinct setups. The former is a retrieval task, whereas the latter is a regression task. Moreover, it requires modeling two levels of video granularity: clip and frame. To address these technical challenges, we propose RGNet, a unified clip Retrieval and Grounding Network (see Fig. 1). The single network enables end-to-end training of both stages. This improves the retrieval module's fine-grained event understanding by directly optimizing it with the moment annotations. Parallely, the grounding network also benefits from the strong multimodal features produced by our encoder, further enhancing the localization capability.", + "bbox": [ + 212, + 657, + 787, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve this unification, we propose a novel transformer encoder, RGEncoder, which enables the grounding network to perform clip retrieval. This", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Hannan et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7ee1042d6ed0acb37be653a312ba6baf8af68724ee30a15c67925808283b8e4a.jpg", + "image_caption": [ + "Existing Disjoint Network" + ], + "image_footnote": [], + "bbox": [ + 285, + 143, + 553, + 296 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3d4b369cc3ea41360acbb01f69fc4e95f9ed760594bfee40902fac5a3eeaa769.jpg", + "image_caption": [ + "Proposed Unified RGNet", + "Fig. 2: Unified Solution. (left) Existing methods involve a separate retrieval and grounding network. The disjoint retrieval lacks fine-grained event understanding, which is crucial for moment localization. (right) Our unified network architecture overcomes it by deeply integrating the retrieval module with the grounding objective." + ], + "image_footnote": [], + "bbox": [ + 581, + 146, + 720, + 295 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "eliminates the suboptimal video retrieval network and effectively models long video clip retrieval. To enable the retrieval module to understand fine-grained events, we design a sparse-attention mechanism in the encoder. The sparse attention enables the retrieval module to focus on specific events in the video, which is more synergistic with the grounding task. Thus RG-Encoder is capable of operating both at clip and frame level granularity. We propose an Intra-Clip Attention Loss, which motivates the sparse attention to focus more on the video frames aligned with the specified event. Furthermore, we propose a negative clip mining technique to simulate clip retrieval from long videos during training. This negative mining enables our network to train on a large batch size with an inter-clip contrastive loss. The large number of negative clips in the batch closely mimics the long video paradigm and reduces the gap between the training and test phases.", + "bbox": [ + 212, + 388, + 784, + 584 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Together, these components of RGNet bridge the gap between the two stages of LVTG, enhancing the fine-grained event understanding in hour-long videos. In summary, our contributions are fourfold:", + "bbox": [ + 212, + 585, + 782, + 628 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We systematically deconstruct existing LVTG methods into clip retrieval and grounding stages. Through empirical evaluations, we discern that disjoint retrieval is the primary factor contributing to poor performance.", + "- Based on our observations, we introduce RGNet, which integrates clip retrieval with grounding through parallel clip and frame-level modeling. This obviates the necessity for a separate video retrieval network, replaced instead by an end-to-end clip retrieval module tailored specifically for long videos.", + "- We introduce sparse attention to the retriever and a corresponding loss to model fine-grained event understanding in long-range video. We propose a contrastive negative clip-mining strategy to simulate clip retrieval from a long video during training.", + "- RGNet achieves state-of-the-art performance across both LVTG datasets, Ego4D [14] and MAD [40]. For instance, RGNet outperforms the previous best Ego4D method [39] by a substantial margin (9.7%)." + ], + "bbox": [ + 222, + 630, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "RGNet", + "bbox": [ + 681, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Literature", + "text_level": 1, + "bbox": [ + 215, + 143, + 431, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Short Video Temporal Grounding. The recent grounding methods mainly focus on short videos. The best-performing ones utilize transformer variants [3,27,34,56]. For example, Moment-DETR [22] adopted transformers for combined video and text features. Subsequent works, such as UMT [28] and QD-DETR [35], split the cross-modal and unimodal modeling of video and text features for enhanced performance. EATR [18] improves its decoder with videospecific moment queries. These models were initially designed for shorter videos. However, when applied to hour-long videos, the moments become extremely tiny, posing a challenge akin to finding a needle in a haystack. Hence, these methods perform poorly on long videos, as noted by the authors of the MAD dataset [40]. In contrast, we effectively address the task by simultaneously processing long videos at two granular levels (video clips and frames) within a single network.", + "bbox": [ + 212, + 186, + 787, + 368 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Long Video Temporal Grounding. Recently, video temporal grounding has been adapted for long videos with MAD [40] and Ego4D [14] datasets. Typically, the methods designed for long videos involve two stages. Proposal-free methods [2, 26, 41, 53, 55] segment lengthy videos into smaller parts to predict candidate moments and then rank them to obtain the final predictions. Proposal-based methods [16, 37] generate proposal clips or anchors and subsequently apply their grounding model to the retrieved proposals. Some methods, like CONE [16], and M-Guidance [2], utilize the detection transformer [3] as the grounding network to improve the grounding phase. These proposal-based models are more efficient and perform better than their counterparts. However, they need a separate retrieval module to select the proposal clips. This disjoint two-stage architecture is what we find to be the main reason for the poor performance of these models. In contrast, we propose a novel transformer encoder that solves both clip retrieval and grounding in a unified way, enabling end-to-end optimization.", + "bbox": [ + 212, + 383, + 787, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video-text Retrieval. The early retrieval methods [4, 7, 21, 49, 50] designed multimodal fusion techniques to align pre-trained video and text models. To bridge the gap between pre-training and downstream tasks, large-scale pretrained video-language representations [1, 6, 9, 11, 31, 42, 45-47] have been introduced in various works. Several studies [8, 10, 19, 20, 32, 43, 44, 48, 51] have transferred the knowledge from CLIP [38] to text-video retrieval tasks. In recent studies, various sampling strategies [13, 23, 29] have been explored, enabling the model to selectively concentrate on pertinent video frames based on provided text inputs. For instance, Clip-BERT [23] introduces a sparse sampling strategy, X-Pool [13] utilizes a cross-modal attention model, and TS2-Net [29] adapts CLIP to capture temporal information and eliminate unimportant tokens. While existing methods aim to retrieve high-level video topics from a pool of video collections, moment localization demands a fine grained understanding to identify specific events within a video. Despite the similar setup to video retrieval, retrieving the corresponding clip from a long video necessitates deeper", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Hannan et al.", + "bbox": [ + 271, + 114, + 364, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "event comprehension. Therefore, we tailored our encoder to enhance event understanding and simulated precise clip retrieval conditions during training.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 RGNet", + "text_level": 1, + "bbox": [ + 215, + 200, + 321, + 217 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section presents a detailed description of our proposed unified LVTG method. As illustrated in Fig. 1, RGNet takes a long video and a text query as input and predicts the precise moment semantically correlated with the query in an end-to-end manner. The whole network is divided into the proposed RG-Encoder (Fig. 3) and a decoder (Sec. 3.3). The encoder retrieves the most relevant clip feature from the input long video. The decoder then processes the retrieved clip feature to predict the precise moment boundary. Additionally, an Intra-Clip Attention Loss and an Inter-Clip Contrastive Loss are incorporated to facilitate low-level event understanding in long-range videos.", + "bbox": [ + 212, + 234, + 789, + 369 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/172a7e7757a46c7be9bbd3ae7f46b1f5136b4f36025c4e25396febda8144b69c.jpg", + "image_caption": [ + "Fig. 3: Overview of RG-Encoder. It takes video clips and textual query as input and retrieves the relevant clip features. First, a cross-attention fuses the clips with text, and the sparsifier masks the out-of-moment frames. Based on the mask, the retrieval attention focuses on in-moment frames (colored red) and generates clip-level context and frame-level content features. We combine the context and content to generate the retrieved clip feature." + ], + "image_footnote": [], + "bbox": [ + 357, + 396, + 640, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Feature Extraction", + "text_level": 1, + "bbox": [ + 215, + 708, + 418, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let us assume we have an untrimmed long video $V$ consisting of $T$ frames and a query sentence $S$ with $N$ words corresponding to a target moment's center $\\tau_{c}$ and width $\\tau_{w}$ . Following previous methods [16,37], we use pre-trained frozen models to extract visual features $V = \\{f^{1}, f^{2}, \\dots, f^{T}\\} \\in R^{T \\times D_{f}}$ as well as textual features $S = \\{w^{1}, w^{2}, \\dots, w^{N}\\} \\in R^{N \\times D_{w}}$ , where $D_{f}$ and $D_{w}$ represent the feature dimensions of video frames and query words. We slice the long video $V$ into clips of length $L_{c}$ and feed them into our model. We employ a sliding window of length", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "RGNet", + "bbox": [ + 681, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$L_{c}$ and stride $L_{c} / 2$ to obtain a collection of $T_{c}$ clips $C = \\{C^{1}, C^{2}, \\dots, C^{T_{c}}\\} \\in R^{T_{c} \\times L_{c} \\times D_{f}}$ . Here, the $i^{th}$ clip, $C^{i} = \\{f^{s_{i} + 1}, f^{s_{i} + 2}, \\dots, f^{s_{i} + L_{c}}\\} \\in R^{L_{c} \\times D_{f}}$ , where $s_{i}$ is the start index of the clip.", + "bbox": [ + 215, + 145, + 787, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 RG-Encoder", + "text_level": 1, + "bbox": [ + 215, + 210, + 366, + 224 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All the clips $C$ from a single video comprise samples in a batch. The RG-Encoder (Fig. 3) processes them with the query $S$ to retrieve the relevant clip feature $P$ . First, the cross-attention generates query-conditioned frame features $F$ . Then, based on the frame-query correlation calculated from $F$ , the sparsifier determines the relevant frames inside the clip to produce a mask $M$ for retrieval attention. We incorporate a learnable token $R$ to capture the context of the clip. Additionally, the frame features $F$ undergo retrieval attention based on the predicted mask $M$ to generate the content feature $F_{c}$ . The combination of context $R$ and content $F_{c}$ forms the retrieved clip feature $P$ .", + "bbox": [ + 212, + 232, + 787, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cross Attention: To evaluate the relevance of each frame within the clip $C^i$ , we utilize cross-attention, where frame features act as queries and text features serve as keys and values. Cross-attention represents the frame features as a weighted average of the textual features, scaled by their mutual correlation. This process results in text-conditioned frame features $F^i$ , establishing a fine-grained correspondence between the modalities. The query, key and values are calculated as $Q^i = l_Q(C^i) \\in R^{L_c \\times D_f}$ , $K = l_K(S) \\in R^{N \\times D_q}$ and $V = l_V(S) \\in R^{N \\times D_q}$ respectively. Here, $l_Q(\\cdot), l_K(\\cdot)$ , and $l_V(\\cdot)$ denote the projection layers for the query, key, and value. Then, we apply the cross-attention layer as follows:", + "bbox": [ + 212, + 383, + 787, + 521 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF ^ {i} = \\operatorname {s o f t m a x} \\left(Q ^ {i} K ^ {T}\\right) V + Q ^ {i} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 400, + 527, + 784, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sparsifier: To retrieve the clip, we need to assess the relevance of each clip based on the text query. Clip relevance is determined by aggregating the relevancy of its frames. However, given the small moment duration, most frames are unrelated to the query. Hence, we classify frames into relevant and non-relevant categories using the sparsifier to assist retrieval attention in focusing on these fine-grained events. We calculate the relevancy $G^{j} \\in [0,1]$ of the $j^{th}$ frame with Eq. 2. We use a differentiable Gumbel-softmax function [15, 17] for enabling end-to-end training. The detailed calculation is in the supplementary materials.", + "bbox": [ + 212, + 566, + 787, + 686 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nG ^ {j} = \\operatorname {G u m b e l - s o f t m a x} \\left(\\operatorname {l i n e a r} \\left(f ^ {j}\\right)\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 695, + 784, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The intra-clip attention loss optimizes this classification using the ground truth frame relevancy information described in Sec. 3.4. To distinguish irrelevant frames from their counterparts, we constrain their updates in the retrieval attention. This restriction is implemented with an attention mask computed from the learned frame relevancy $G^{j}$ . The attention mask $M$ between $j^{th}$ and $k^{th}$ frame is calculated with Eq. 3. Here, the $j^{th}$ and $k^{th}$ frames are the attention query and key, respectively. This leads to a reduced set of relevant and informative frames for subsequent attention.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Hannan et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nM (j, k) = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} G ^ {j} > 0. 5 \\text {o r} j = k \\\\ - \\infty & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 157, + 784, + 189 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Retrieval Attention: We aim for a learnable clip retrieval function to train end-to-end with the grounding objective. Thus, we devise the retrieval attention for aggregating the text-conditioned frame features based on their relevancy. To condense the $i^{th}$ clip into a contextual feature embedding, we introduce a randomly initialized learnable vector $R^i$ . We denote this as retrieval token, which is trained with the inter-clip contrastive loss.", + "bbox": [ + 212, + 203, + 782, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "First, we concatenate the retrieval token with the text-conditioned frame features to create the queries $Q^i = [R^i,F^i]$ for retrieval attention. We also generate an attention mask containing all zeros for the retrieval token and concatenate it with $M$ . We set the key $K = Q^i$ , and the value $V = Q^i$ to calculate the retrieval attention according to Eq. 4. We get the clip-level context feature, $R^i = \\tilde{Q}^{i,1}$ and frame-level content features, $F_{c}^{i,j} = [\\tilde{Q}^{i,2},\\tilde{Q}^{i,3},\\dots,\\tilde{Q}^{i,L_{c} + 1}]$ . Here, $F_{c}^{i,j}$ is the $j^{th}$ frame feature of the $i^{th}$ clip.", + "bbox": [ + 212, + 294, + 784, + 398 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {Q} ^ {i} = \\operatorname {s o f t m a x} \\left(Q ^ {i} K ^ {T} + M\\right) V + Q ^ {i} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 411, + 784, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Further, the context feature $\\mathrm{R}$ is utilized to retrieve the clip. It undergoes a linear projection layer, $l_{s}(\\cdot)$ , to obtain context scores $S_{r} = \\operatorname{sigmoid}(l_{s}(R))$ . We retrieve the most relevant clip based on the context score $S_{r}$ . Importantly, the context $\\mathrm{R}$ represents the queried moment by attending to the reduced set of relevant frames. So, we fuse it with the frame features with Eq. 5 to produce stronger multimodal inputs for the decoder. It enhances the decoder's localization capability significantly. The decoder only processes features from the retrieved clip to predict the moment.", + "bbox": [ + 212, + 438, + 784, + 558 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nP ^ {i, j} = F _ {c} ^ {i, j} + R ^ {i} \\times G ^ {j}, i \\in P, \\forall j \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 387, + 571, + 784, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Grounding Decoder", + "text_level": 1, + "bbox": [ + 214, + 608, + 426, + 623 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We process the retrieved clip features with a transformer decoder to predict precise moment boundaries. Previous methods [16,22] feed both modalities into the decoder, introducing multiple positional information (e.g., time in the video and word order in the text) that complicates the detection task. In contrast, our encoder injects text information directly into the clip, eliminating the requirement to feed both modalities into the decoder. We incorporate learnable anchor queries [27] to represent the moment's center and width as $(\\tau_c,\\tau_w)$ . The decoder has 2 cross-attention and 2 self-attention layers.", + "bbox": [ + 212, + 626, + 782, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Loss Functions", + "text_level": 1, + "bbox": [ + 214, + 768, + 385, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Intra-Clip Attention Loss guides our sparsifier to distinguish frames within and outside of the ground truth moment. This enables the retrieval attention to focus on relevant video regions and achieve fine-grained event understanding.", + "bbox": [ + 212, + 794, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "RGNet", + "bbox": [ + 681, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, the in-moment frames are required to maintain higher relevancy scores than those outside the ground truth moment. The loss is defined as:", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {a t t n}} = \\max (0, \\Delta + S _ {c} (i, j _ {\\mathrm {o u t}}) - S _ {c} (i, j _ {\\mathrm {i n}})) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 184, + 785, + 199 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the equation, $S_{c}$ is the relevancy score, $\\Delta$ represents the margin, and anchor $i$ is an in-moment frame. We randomly chose in-moment and out-of-moment frames $j_{\\mathrm{in}}$ and $j_{\\mathrm{out}}$ from the same clip. To calculate the relevancy score, the context features $R_{i}$ , and encoded clip $P^{i,j}$ undergoes linear layers, $l_{r}(\\cdot)$ and $l_{c}(\\cdot)$ , to obtain projected context $R_{proj}^{i} = l_{r}(R^{i}) \\in R^{D_{l}}$ and content features, $P_{proj}^{i,j} = l_{c}(P^{i,j}) \\in R^{L_{c} \\times D_{l}}$ . Then, we calculate the frame-level relevancy score:", + "bbox": [ + 212, + 205, + 782, + 301 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nS _ {c} (i, j) = R _ {p r o j} ^ {i} \\cdot P _ {p r o j} ^ {i, j} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 421, + 306, + 785, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inter-Clip Contrastive loss. A long video contains many clips with similar environments and scenes. However, most of these clips do not contain the specific moment we are searching for. So, high-level video understanding is not enough to distinguish precise moments. During the test phase, we need to handle many such negative clips. To tackle this issue, unlike previous approaches [16], we sample a large number of negative clips and train our retriever on a large batch size. Hence, this negative clip sampling reduces the disparity between the train and test phases.", + "bbox": [ + 212, + 344, + 784, + 463 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For this section, we denote the context feature, $R^{i,j}$ for $i^{th}$ proposal and $j^{th}$ text query. We train our model contrastively with positive text-clip pairs $[R^{i,j}]_{i = j}$ and negative pairs $[R^{i,j}]_{i\\neq j}$ . We use the InfoNCE loss [36] to identify the positive text-clip pair amongst a set of unrelated negative samples. Here, $l_{cont}(\\cdot)$ is a linear projection layer that transforms the $D_{f}$ dimensional context feature into a one-dimensional logit.", + "bbox": [ + 212, + 465, + 784, + 556 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n t}} = - \\sum_ {i} \\log \\frac {\\exp \\left(l _ {\\text {c o n t}} \\left(R ^ {i , i}\\right)\\right)}{\\sum_ {j} \\exp \\left(l _ {\\text {c o n t}} \\left(R ^ {i , j}\\right)\\right)} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 561, + 785, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Grounding loss. The objective functions for grounding, which aim to locate desired moments, are adopted from the baseline approach [22]. The grounding loss $\\mathcal{L}_{\\mathrm{g}}$ measures the discrepancy between the ground truth (GT) and predicted moments. A one-to-one correspondence between GT and predicted moments is established using the Hungarian algorithm. This loss includes both an $\\mathcal{L}1$ loss and a generalized IoU loss ( $\\mathcal{L}\\mathrm{gIoU}$ ). Additionally, a cross-entropy loss $\\mathcal{L}_{\\mathrm{CE}}$ is employed to classify predicted moments as either foreground or background. Thus, $\\mathcal{L}_{\\mathrm{g}}$ is defined as follows:", + "bbox": [ + 212, + 608, + 785, + 731 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {g}} = \\lambda_ {\\mathcal {L} 1} | | \\tau - \\hat {\\tau} | | + \\lambda_ {\\mathrm {g I o U}} \\mathcal {L} _ {\\mathrm {g I o U}} (\\tau , \\hat {\\tau}) + \\lambda_ {\\mathrm {C E}} \\mathcal {L} _ {\\mathrm {C E}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 737, + 785, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\tau$ and $\\hat{\\tau}$ represent the ground-truth moment and its corresponding prediction, containing center coordinates $\\tau_{c}$ and width $\\tau_{w}$ . The hyper-parameters $\\lambda_{*}$ are used for balancing the losses. Finally, with the attention and contrastive loss, the total loss $\\mathcal{L}_{\\mathrm{total}}$ is defined as follows:", + "bbox": [ + 212, + 757, + 782, + 818 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {\\text {a t t n}} \\mathcal {L} _ {\\text {a t t n}} + \\lambda_ {\\text {c o n t}} \\mathcal {L} _ {\\text {c o n t}} + \\mathcal {L} _ {\\mathrm {g}} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 825, + 785, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Hannan et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 143, + 444, + 162 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets and Evaluation Metric", + "text_level": 1, + "bbox": [ + 215, + 176, + 524, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "MAD [40] is an extensive dataset comprising 1.2K hours of full-length movies and 384K natural language queries, each associated with specific moments in the videos. The videos are orders of magnitude longer than previous datasets, with an average duration of 110 minutes, while the specified text moments are on average 4.1 seconds. This small moment-to-video ratio poses a significant challenge for the grounding task.", + "bbox": [ + 215, + 200, + 785, + 292 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ego4D-NLQ [14] is a large-scale egocentric video data set with multiple challenges. Specifically, we use the episodic memory benchmark Ego4D-NLQ, which requires localizing where the answer to a natural language query can be seen. It contains around 13 template questions, amounting to around 74K queries. The train, val, and test sets contain 11.3K, 3.9K, and 4.0K queries. The video length ranges from 8 to 20 minutes, with an average of 8.25 minutes, and the average moment duration is 8.3 seconds. This means the moments constitute only $2\\%$ of the input video on average.", + "bbox": [ + 215, + 306, + 785, + 429 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Grounding Metric: For grounding, we follow previous methods [16, 40], and adopt the standard metric $\\mathrm{Recall}@\\mathbf{k}$ at IoU = $\\theta$ ( $\\mathrm{Rk}_{\\theta}$ ). This metric represents the percentage of testing samples with at least one grounding prediction whose intersection over union (IoU) with the ground truth (GT) is larger than $\\theta$ among the top- $k$ predictions.", + "bbox": [ + 215, + 443, + 785, + 518 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Retrieval Metric: To assess our proposal retrieval stage independently of the grounding, we utilize the standard retrieval metric [13], Recall at Rank $k$ ( $\\mathrm{R}@\\mathbf{k}$ ). This metric calculates the percentage of GT moments present in the top- $k$ retrieved proposals.", + "bbox": [ + 215, + 534, + 785, + 594 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Implementation Details:", + "text_level": 1, + "bbox": [ + 215, + 617, + 460, + 633 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We use pre-trained CLIP [38] and EgoVLP [24] models to extract video frames from MAD and Ego4D. Text features for both datasets are extracted using CLIP. The feature extractors are frozen, and their pre-trained weights remain unchanged during training. For Ego4D, we pre-train our model with NaQ annotations from previous work [39]. The proposal length, $W_{c}$ is set to 180s for MAD and 48s for Ego4D, with only the top-30 and top-5 proposals retrieved, respectively. Training on Ego4D and MAD is conducted on four Nvidia-RTX-A6000 GPUs while fine-tuning on Ego4D is performed on a single GPU. Considering the longer proposals in MAD, we utilize a batch size of 32 for Ego4D and 8 for MAD. The number of moment queries is set to 5. For loss computation, we set hyperparameters as $\\lambda_{L1} = 10$ , $\\lambda_{\\mathrm{gIoU}} = 1$ , $\\lambda_{\\mathrm{CE}} = 4$ , $\\lambda_{\\mathrm{samp}} = 1$ , $\\lambda_{\\mathrm{cont}} = 10$ , and $\\Delta = 0.2$ . We use Xavier initialization [12] and employ AdamW [30] with an initial learning rate of $1 \\times 10^{-4}$ . RGNet is trained for 35 and 200 epochs on", + "bbox": [ + 215, + 642, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "RGNet", + "bbox": [ + 681, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "MAD and Ego4D datasets. The initial learning rate is reduced by one order of magnitude at epochs 25 for MAD and 120 for Ego4D.", + "bbox": [ + 215, + 145, + 785, + 176 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Results and Analysis", + "text_level": 1, + "bbox": [ + 215, + 196, + 455, + 214 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We first compare our performance with previous state-of-the-art models. Then, we report detailed ablation studies of our proposed method and visualize the qualitative results compared to the disjoint baseline.", + "bbox": [ + 215, + 224, + 785, + 270 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/b930333388c1144e6b9e24c5938decc8a3fb51d0287f43559153e235342cfd1b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelNaQEgo4D-NLQ [14]MAD [40]Avg
R1.3R5.3R1.5R5.5AvgR1.1R5.1R1.3R5.3R1.5R5.5Avg
2D-TAN [55]X5.0412.892.025.886.463.2211.902.529.251.585.696.07
UniVTG [25]X11.747.543.257.887.60-------
VSLNet [53]✓10.2619.015.8112.6711.93-------
VLG-Net [41]X-----3.6411.662.769.311.655.995.84
M-DETR [22]X8.2323.235.0113.3712.463.6012.982.819.861.675.586.08
M-Guide [2]X-----9.3018.964.6513.062.167.409.26
SOONet [37]X8.0022.403.7611.0911.3111.2623.219.0019.645.3213.1413.59
H-Hands [52]X13.2023.307.9015.6015.00-------
CONE [16]X14.1530.338.1818.0217.678.9020.516.8716.114.109.5911.01
EgoVLP✓15.9026.389.4617.8017.38-------
ReLeR [39]✓19.3123.6211.5915.7517.57-------
Ours (Default)X18.2834.0212.0422.8921.8112.4325.129.4818.725.6110.8613.70
Ours✓20.6341.6712.4725.0824.9612.4325.129.4818.725.6110.8613.70
", + "bbox": [ + 215, + 292, + 785, + 474 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 1: Main results on Ego4D-NLQ and MAD. RGNet achieves state-of-the-art performance on both datasets. Our default network is trained without NaQ annotations [39] on Ego4D. Even without NaQ annotations, our default network shows a larger improvement, underscoring the effectiveness of our solution with limited data.", + "bbox": [ + 215, + 477, + 785, + 532 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Result on the Ego4D-NLQ Dataset", + "text_level": 1, + "bbox": [ + 215, + 550, + 549, + 566 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "RGNet demonstrates state-of-the-art performance on all metrics for the Ego4D-NLQ benchmark. As shown in Table 1, RGNet improves its primary evaluation metric, the R1.3 and R5.3 score average by $9.7\\%$ . Separately, we outperform the previous best methods on R1.3 and R5.3 by $1.32\\%$ and $18.1\\%$ , respectively. This improvement is consistent across other metrics as well. The unified clip and frame-level modeling in RGNet parallelly capture long- and short-range temporal dependencies, contributing to its superior performance. By modeling the entire video as a whole, RGNet learns better cross-modal alignment by contrasting visual scenarios from various events within the same video. Notably, compared to SooNet [37] and CONE [16], two prominent proposal-based methods, RGNet achieves an average score improvement of $13.7\\%$ and $7.3\\%$ . This substantial improvement is attributed to our unified modeling of clip retrieval, whereas they separately employ a retrieval similarity search between the video and the text.", + "bbox": [ + 215, + 571, + 785, + 768 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Result on the MAD Dataset", + "text_level": 1, + "bbox": [ + 215, + 787, + 495, + 801 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We compare our model's performance on the MAD dataset in Tab. 1. RGNet achieves state-of-the-art performance on both the R1.1 and R5.1 scores by out-", + "bbox": [ + 215, + 809, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Hannan et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "performing the previous best SooNet [37] by $1.2\\%$ and $1.9\\%$ . Together with its competitive performance on other metrics, RGNet demonstrates the importance of end-to-end modeling of long videos. Specifically, our performance of R1 is superior at all IoU thresholds, which testifies to the effectiveness of the proposed retrieval method. Notably, prior models rely on heuristic dot product similarity search applied to CLIP [38] features for the retrieval phase. In contrast, our shared retrieval feature and its unified optimization with the grounding objective drastically enhance the first stage retrieval, as seen in Tab. 2. It further helps the network capture more discriminative events crucial for detecting extremely tiny moments in hour-long videos.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.3 Disjoint vs. Unified Architecture", + "text_level": 1, + "bbox": [ + 214, + 316, + 532, + 333 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Since the long video temporal grounding (LVTG) performance relies on the first stage of clip retrieval, it does not perfectly reflect the actual moment localization capability of its grounding network. To assess the standalone grounding performance, we designed an oracle experiment operating exclusively on the clip where the ground truth moment is present. Compared to the oracle grounding, the LVTG $R1_{3}$ drops by $15.7\\%$ and $23.9\\%$ for Ego4D and MAD in Tab. 2. The poor grounding in LVTG emerges from the suboptimal retrieval accuracy of the disjoint network. Hence, retrieving clips where the ground truth moment exists is crucial, independent of the subsequent grounding accuracy. Regardless of the effectiveness of moment boundary detection, if the clip containing the moment boundary cannot be accurately retrieved. Our unified approach consis", + "bbox": [ + 212, + 338, + 787, + 505 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/fb0c45b8ba4c4ba58dbc050e7f22ed32784e3612bc281e8b0a165f1b1b919a3b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DataModelClipRetrievalOracleGroundingLVTG
R@1R@5R1.3R5.3R1.3R5.3
Ego4dBaseline31.7164.6329.8454.0114.1530.33
Ours42.0876.2836.5363.6420.6341.67
MADBaseline12.4124.5029.4953.026.8716.11
Ours25.0150.0233.4263.439.4818.72
", + "bbox": [ + 269, + 529, + 733, + 635 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 2: Empirical impact of two stages. To assess standalone grounding capability, we run an oracle experiment on the clip where the ground truth moment is present. The grounding capability degrades in LVTG evaluation because of incorrect selection by the disjoint clip retrieval network. Our unified model improves retrieval significantly, which leads to more effective temporal grounding in long videos.", + "bbox": [ + 212, + 637, + 787, + 708 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "tently improves all retrieval metrics compared to the disjoint baseline. Notably, we achieve a staggering $10.4\\%$ and $12.6\\%$ improved R@1 score for Ego4D and MAD. This improvement in clip retrieval leads to state-of-the-art LVTG performance. Moreover, the grounding also improves because of the stronger clip features generated by our encoder. Compared to the disjoint baseline, the oracle R1.3 scores improved by about $6.7\\%$ and $4.0\\%$ for Ego4D and MAD, respectively. Hence, the unified end-to-end architecture proves more effective at LVTG by mutually improving both stages.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "RGNet", + "bbox": [ + 681, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.4 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 146, + 401, + 159 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "This section reports detailed ablation studies on the proposed modules and loss functions. We also ablate with various clip lengths and retrieved clip numbers. Unless stated, the experiments are done with the Default network w/o NaQ annotations [39] for the Ego4D-NLQ dataset and CONE [16] as the baseline.", + "bbox": [ + 212, + 179, + 787, + 241 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/a7843a2c79407b07c6ebeb38f798b40a6016251a422474c3646cc3a08ba3d150.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelR1.3R5.3R1.5R5.5
RGNet (Default)18.2834.0212.0422.89
w/o Retrieval Token17.8033.9911.2522.32
w/o Sparsifier16.1231.579.9120.47
w/o RG-Encoder*14.1530.338.1818.02
w/o Contrastive Loss17.4132.1210.7922.80
w/o Attention Loss16.2131.599.9120.53
", + "bbox": [ + 302, + 268, + 702, + 393 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 3: Cumulative ablation study for the proposed modules and losses of RGNet. The experiment is conducted w/o NaQ augmentation on the Ego4D dataset. In each step, we remove one of our proposed modules or losses to evaluate their individual impact. *denotes the baseline model.", + "bbox": [ + 212, + 393, + 787, + 450 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Modules: In our initial experiments, we validate the effectiveness of each proposed module with a cumulative ablation in Tab. 3. The unified network demonstrates superior performance across all evaluation metrics, notably achieving an R1.3 score of $18.28\\%$ . Note that the disjoint baseline performs $4.2\\%$ worse than our unified network. This performance difference can be attributed to our models unified architecture.", + "bbox": [ + 212, + 472, + 787, + 561 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "With the removal of the retrieval token for generating the clip feature in Eq. 5, considerably drops the score by about $0.5\\%$ . This proves that modeling both clip level context and frame level content are necessary to produce strong multimodal features for the grounding decoder. Subsequently, eliminating the sparsifier, there is a decline in $\\mathrm{R1}_{.3}$ performance of approximately $1.7\\%$ . This verifies the impact of the sparsifier in modeling the fine-grained event boundaries. Finally, removing the RG-Encoder yields a notable degradation of approximately $2.0\\%$ in the $\\mathrm{R1}_{.3}$ score, which substantiates its capability to model clip and frame level granularity jointly.", + "bbox": [ + 212, + 565, + 787, + 702 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Loss Functions: Training RGNet without our negative clip sampling strategy for the contrastive loss (Tab. 3) results in an almost $0.9\\%$ decline in the R1.3 score. Our sampling strategy mimics long-video setup closely during training, which improves event discrimination capability inside the same scene and surroundings.", + "bbox": [ + 212, + 717, + 787, + 792 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "By eliminating the attention loss, we observe a further reduction of $1.2\\%$ in grounding performance. The attention loss enables our encoder to model fine-grained events in long videos which is crucial for specific moment detection.", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Hannan et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Number of Clips: The number of retrieved proposal clips significantly influences the Long Video Temporal Grounding (LVTG) performance. Our model consistently outperforms the baseline across various choices of top-k retrieved clips, as shown in Fig. 4. Even with only a single retrieved clip, our performance surpasses the baseline's optimal result for both datasets. Moreover, our model exhibits a less steep performance decline when fewer clips are retrieved. Importantly, with fewer clips, the grounding network runs fewer times, leading to an overall speedup. Therefore, our model strikes a favorable balance between efficiency and performance, showcasing superior performance than the baseline.", + "bbox": [ + 212, + 161, + 787, + 297 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7ee740ac8a82dad00098dd757937a2858c2e768cc2036277845a1f02225463e9.jpg", + "image_caption": [ + "Fig. 4: Impact of number of retrieved clips. Reducing the number of clips speeds up the network execution time. While the baseline model experiences a significant drop in performance with this reduction, RGNet shows a noticeably smaller decline in performance under the same conditions." + ], + "image_footnote": [], + "bbox": [ + 333, + 325, + 496, + 431 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2cea0e1ac86b544dfe9d3eb325bd2ab122c5ec943838025be54eaa3beb39a6b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 327, + 671, + 431 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "**Clip Length:** Varying clip length impacts both retrieval and grounding performance. The number of clip decreases with longer clips, making the retrieval task easier. Fig. 5 reports a monotonically improving retrieval with longer clips. However, after a certain length, moment localization quality starts to deteriorate. With longer clips, the task becomes extremely difficult. We see the decline after the 180s and 48s for the MAD and Ego4D datasets. Hence, we set the default clip length of our method to these optimal lengths. Importantly, the unified network helps outscore the disjoint baseline on all tested clip lengths (inversely proportional to the clip numbers in Fig. 4).", + "bbox": [ + 212, + 526, + 787, + 662 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/42237315c5214986a41bddd5bd1ff81f7a59e11cb15f98fb772901c7fcf5400a.jpg", + "image_caption": [ + "Fig. 5: Impact of retrieved clip length. Longer clips result in improved retrieval due to fewer candidates. However, grounding becomes exceedingly difficult with longer clips. For example, the grounding performance drops after 180 seconds and 48 seconds for the MAD and Ego4D datasets." + ], + "image_footnote": [], + "bbox": [ + 326, + 689, + 493, + 789 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/75b34f5a975d9f83f13448b05d8dfaee9e477ef4dffb730d8cf7bbdf8f7cd4e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 689, + 676, + 789 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "RGNet", + "bbox": [ + 681, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.5 Qualitative Analysis", + "text_level": 1, + "bbox": [ + 215, + 146, + 429, + 161 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We visualize the qualitative predictions of RGNet compared to the disjoint baseline in Fig. 6. For the first and second queries, the baseline struggles to retrieve the correct clip where the scissor appears. RGNet accurately retrieves the clip with the searched object and precisely localizes the moment it appeared in the video. Disjoint retrieval often leads to incorrect clip selection, which the final grounding network cannot recover from. For the third query, the baseline selects the correct clip but fails to localize the precise moment the man closes the car door, indicating a gap between the two stages. RGNet mitigates this gap by leveraging features from the retrieved clip to improve moment localization. More visual outputs are included the supplementary materials.", + "bbox": [ + 212, + 170, + 787, + 321 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/ea0a59213b6660ed4bc74e839ceb6b949aa4d53aa9c71487292fb574909eacaf.jpg", + "image_caption": [ + "Query 1: Where is the scissor?" + ], + "image_footnote": [], + "bbox": [ + 274, + 359, + 730, + 415 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/8f75fc2e4626d4706db20f6a1f70073f6b6ab6ecf466b45449ae5b0f07307483.jpg", + "image_caption": [ + "Query 2: How many stainless bowls did I pick?" + ], + "image_footnote": [], + "bbox": [ + 274, + 435, + 728, + 494 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/907b19be12350262c7b4fbc06206b8afdb4a9d8380bccd71e2f4c68195f155c5.jpg", + "image_caption": [ + "Query 3: Did I leave the car door open?", + "Fig. 6: Qualitative Results. The baseline fails to retrieve the correct clip in the first two queries. Since they primarily depict the same indoor room throughout the whole video, precise event discrimination is vital for accurate clip retrieval. In the third query, the baseline cannot identify the moment within the correctly retrieved clip, detecting it only after it has concluded. Precise event localization demands improved alignment between visual events and the queried text. RGNet correctly localizes all the moments." + ], + "image_footnote": [], + "bbox": [ + 276, + 513, + 684, + 566 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 672, + 359, + 689 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We introduce an end-to-end model for long video temporal grounding, unifying the two stages of prevailing methods with shared features and mutual optimization. We conduct independent analyses of the two stages, shaping our solution based on their distinct impacts. This leads to a better understanding of fine-grained events in long videos. Our approach demonstrates state-of-the-art performance on challenging long video grounding datasets, validating its effectiveness. Like all LVTG methods, we rely on a pre-trained image encoder to extract the visual features. A promising future direction includes eliminating the need for an image encoder and training directly on the raw video data.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Hannan et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: ICCV. pp. 1728-1738 (2021)", + "2. Barrios, W., Soldan, M., Ceballos-Arroyo, A.M., Heilbron, F.C., Ghanem, B.: Localizing moments in long video via multimodal guidance. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13667-13678 (2023)", + "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)", + "4. Chen, S., Zhao, Y., Jin, Q., Wu, Q.: Fine-grained video-text retrieval with hierarchical graph reasoning. In: CVPR. pp. 10638-10647 (2020)", + "5. Chen, Z., Jiang, X., Xu, X., Cao, Z., Mo, Y., Shen, H.T.: Joint searching and grounding: Multi-granularity video content retrieval. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 975-983 (2023)", + "6. Cheng, F., Wang, X., Lei, J., Crandall, D., Bansal, M., Bertasius, G.: Vindlu: A recipe for effective video-and-language pretraining. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023)", + "7. Dong, J., Li, X., Xu, C., Yang, X., Yang, G., Wang, X., Wang, M.: Dual encoding for video retrieval by text. IEEE TPAMI pp. 4065-4080 (2021)", + "8. Fang, B., Liu, C., Zhou, Y., Yang, M., Song, Y., Li, F., Wang, W., Ji, X., Ouyang, W., et al.: Uatvr: Uncertainty-adaptive text-video retrieval. arXiv preprint arXiv:2301.06309 (2023)", + "9. Gabeur, V., Sun, C., Alahari, K., Schmid, C.: Multi-modal transformer for video retrieval. In: ECCV. pp. 214-229 (2020)", + "0. Gao, Z., Liu, J., Sun, W., Chen, S., Chang, D., Zhao, L.: Clip2tv: Align, match and distill for video-text retrieval. arXiv preprint arXiv:2111.05610 (2021)", + "1. Ge, Y., Ge, Y., Liu, X., Li, D., Shan, Y., Qie, X., Luo, P.: Bridging video-text retrieval with multiple choice questions. In: CVPR. pp. 16167-16176 (2022)", + "2. Glorot, X., Bengio, Y.: Understanding the difficulty of training deep feedforward neural networks. In: Proceedings of the thirteenth international conference on artificial intelligence and statistics. pp. 249-256. JMLR Workshop and Conference Proceedings (2010)", + "3. Gorti, S.K., Vouitsis, N., Ma, J., Golestan, K., Volkovs, M., Garg, A., Yu, G.: X-pool: Cross-modal language-video attention for text-video retrieval. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5006-5015 (2022)", + "4. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: CVPR. pp. 18995-19012 (2022)", + "5. Hannan, T., Koner, R., Bernhard, M., Shit, S., Menze, B., Tresp, V., Schubert, M., Seidl, T.: Gratt-vis: Gated residual attention for auto rectifying video instance segmentation. arXiv preprint arXiv:2305.17096 (2023)", + "6. Hou, Z., Zhong, W., Ji, L., Gao, D., Yan, K., Chan, W.K., Ngo, C.W., Shou, Z., Duan, N.: Cone: An efficient coarse-to-fine alignment framework for long video temporal grounding. arXiv preprint arXiv:2209.10918 (2022)", + "7. Jang, E., Gu, S., Poole, B.: Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144 (2016)", + "8. Jang, J., Park, J., Kim, J., Kwon, H., Sohn, K.: Knowing where to focus: Event-aware transformer for video grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13846-13856 (2023)" + ], + "bbox": [ + 225, + 172, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "RGNet", + "bbox": [ + 681, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Jiang, H., Zhang, J., Huang, R., Ge, C., Ni, Z., Lu, J., Zhou, J., Song, S., Huang, G.: Cross-modal adapter for text-video retrieval. arXiv preprint arXiv:2211.09623 (2022)", + "20. Jin, P., Huang, J., Liu, F., Wu, X., Ge, S., Song, G., Clifton, D.A., Chen, J.: Expectation-maximization contrastive learning for compact video-and-language representations. NeurIPS (2022)", + "21. Kiros, R., Salakhutdinov, R., Zemel, R.S.: Unifying visual-semantic embeddings with multimodal neural language models. arXiv preprint arXiv:1411.2539 (2014)", + "22. Lei, J., Berg, T.L., Bansal, M.: Detecting moments and highlights in videos via natural language queries. Advances in Neural Information Processing Systems 34, 11846-11858 (2021)", + "23. Lei, J., Li, L., Zhou, L., Gan, Z., Berg, T.L., Bansal, M., Liu, J.: Less is more: Clipbert for video-and-language learning via sparse sampling. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7331-7341 (2021)", + "24. Lin, K.Q., Wang, A.J., Soldan, M., Wray, M., Yan, R., Xu, E.Z., Gao, D., Tu, R., Zhao, W., Kong, W., et al.: Egocentric video-language pretraining. arXiv preprint arXiv:2206.01670 (2022)", + "25. Lin, K.Q., Zhang, P., Chen, J., Pramanick, S., Gao, D., Wang, A.J., Yan, R., Shou, M.Z.: Univtg: Towards unified video-language temporal grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2794-2804 (2023)", + "26. Liu, N., Wang, X., Li, X., Yang, Y., Zhuang, Y.: Refer@ zju-alibaba submission to the ego4d natural language queries challenge 2022. arXiv preprint arXiv:2207.00383 (2022)", + "27. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329 (2022)", + "28. Liu, Y., Li, S., Wu, Y., Chen, C.W., Shan, Y., Qie, X.: Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3042-3051 (2022)", + "29. Liu, Y., Xiong, P., Xu, L., Cao, S., Jin, Q.: Ts2-net: Token shift and selection transformer for text-video retrieval. In: ECCV. pp. 319-335 (2022)", + "30. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)", + "31. Luo, H., Ji, L., Shi, B., Huang, H., Duan, N., Li, T., Li, J., Bharti, T., Zhou, M.: Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353 (2020)", + "32. Ma, Y., Xu, G., Sun, X., Yan, M., Zhang, J., Ji, R.: X-clip: End-to-end multigrained contrastive learning for video-text retrieval. In: ACM MM. pp. 638-647 (2022)", + "33. Maddison, C.J., Mnih, A., Teh, Y.W.: The concrete distribution: A continuous relaxation of discrete random variables. CoRR abs/1611.00712 (2016), http://arxiv.org/abs/1611.00712", + "34. Meng, D., Chen, X., Fan, Z., Zeng, G., Li, H., Yuan, Y., Sun, L., Wang, J.: Conditional detr for fast training convergence. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3651-3660 (2021)", + "35. Moon, W., Hyun, S., Park, S., Park, D., Heo, J.P.: Query-dependent video representation for moment retrieval and highlight detection. In: Proceedings of the" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Hannan et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23023-23033 (2023)", + "36. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)", + "37. Pan, Y., He, X., Gong, B., Lv, Y., Shen, Y., Peng, Y., Zhao, D.: Scanning only once: An end-to-end framework for fast temporal grounding in long videos. arXiv preprint arXiv:2303.08345 (2023)", + "38. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. pp. 8748-8763 (2021)", + "39. Ramakrishnan, S.K., Al-Halah, Z., Grauman, K.: Naq: Leveraging narrations as queries to supervise episodic memory. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6694-6703 (2023)", + "40. Soldan, M., Pardo, A., Alcázar, J.L., Caba, F., Zhao, C., Giancola, S., Ghanem, B.: Mad: A scalable dataset for language grounding in videos from movie audio descriptions. In: CVPR. pp. 5026-5035 (2022)", + "41. Soldan, M., Xu, M., Qu, S., Tegner, J., Ghanem, B.: Vlg-net: Video-language graph matching network for video grounding. In: ICCV. pp. 3224-3234 (2021)", + "42. Wang, A.J., Ge, Y., Yan, R., Ge, Y., Lin, X., Cai, G., Wu, J., Shan, Y., Qie, X., Shou, M.Z.: All in one: Exploring unified video-language pre-training. arXiv preprint arXiv:2203.07303 (2022)", + "43. Wang, Q., Zhang, Y., Zheng, Y., Pan, P., Hua, X.S.: Disentangled representation learning for text-video retrieval. arXiv preprint arXiv:2203.07111 (2022)", + "44. Wang, Z., Sung, Y.L., Cheng, F., Bertasius, G., Bansal, M.: Unified coarse-to-fine alignment for video-text retrieval. In: The IEEE International Conference on Computer Vision (ICCV) (October 2023)", + "45. Xu, H., Ghosh, G., Huang, P.Y., Arora, P., Aminzadeh, M., Feichtenhofer, C., Metze, F., Zettlemoyer, L.: Vlm: Task-agnostic video-language model pre-training for video understanding. ACLliu2021hit (2021)", + "46. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettlemoyer, L., Feichtenhofer, C.: Videoclip: Contrastive pre-training for zero-shot video-text understanding. In: EMNLP. pp. 6787-6800 (2021)", + "47. Xue, H., Hang, T., Zeng, Y., Sun, Y., Liu, B., Yang, H., Fu, J., Guo, B.: Advancing high-resolution video-language representation with large-scale video transcriptions. In: CVPR. pp. 5036-5045 (2022)", + "48. Xue, H., Sun, Y., Liu, B., Fu, J., Song, R., Li, H., Luo, J.: Clip-vip: Adapting pre-trained image-text model to video-language representation alignment. ICLR (2023)", + "49. Yu, Y., Kim, J., Kim, G.: A joint sequence fusion model for video question answering and retrieval. In: ECCV. pp. 471-487 (2018)", + "50. Yu, Y., Ko, H., Choi, J., Kim, G.: End-to-end concept word detection for video captioning, retrieval, and question answering. In: CVPR. pp. 3165-3173 (2017)", + "51. Zhang, B., Jin, X., Gong, W., Xu, K., Zhang, Z., Wang, P., Shen, X., Feng, J.: Multimodal video adapter for parameter efficient video text retrieval. arXiv preprint arXiv:2301.07868 (2023)", + "52. Zhang, C., Gupta, A., Zisserman, A.: Helping hands: An object-aware ego-centric video recognition model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13901-13912 (2023)", + "53. Zhang, H., Sun, A., Jing, W., Zhou, J.T.: Span-based localizing network for natural language video localization. arXiv preprint arXiv:2004.13931 (2020)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "RGNet", + "bbox": [ + 681, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "54. Zhang, H., Sun, A., Jing, W., Zhou, J.T.: The elements of temporal sentence grounding in videos: A survey and future directions. arXiv preprint arXiv:2201.08071 1(2) (2022)", + "55. Zhang, S., Peng, H., Fu, J., Luo, J.: Learning 2d temporal adjacent networks for moment localization with natural language. In: AAAI. pp. 12870-12877 (2020)", + "56. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)" + ], + "bbox": [ + 215, + 146, + 784, + 258 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Hannan et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/RGNet_ A Unified Clip Retrieval and Grounding Network for Long Videos/f091b053-bc8d-4374-9687-69b5bf80ade6_model.json b/2024/RGNet_ A Unified Clip Retrieval and Grounding Network for Long Videos/f091b053-bc8d-4374-9687-69b5bf80ade6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ab8e27a9aec05104ead49d712a9f11461b5563b3 --- /dev/null +++ b/2024/RGNet_ A Unified Clip Retrieval and Grounding Network for Long Videos/f091b053-bc8d-4374-9687-69b5bf80ade6_model.json @@ -0,0 +1,2403 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.224, + 0.141, + 0.78, + 0.187 + ], + "angle": 0, + "content": "RGNet: A Unified Clip Retrieval and Grounding Network for Long Videos" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.213, + 0.77, + 0.245 + ], + "angle": 0, + "content": "Tanveer Hannan\\(^{1,2\\star}\\) Md Mohaiminul Islam\\(^{3}\\) Thomas Seidl\\(^{1,2}\\) Gedas Bertasius\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.447, + 0.254, + 0.555, + 0.268 + ], + "angle": 0, + "content": "1 LMU Munich" + }, + { + "type": "text", + "bbox": [ + 0.468, + 0.269, + 0.535, + 0.282 + ], + "angle": 0, + "content": "2 MCML" + }, + { + "type": "text", + "bbox": [ + 0.435, + 0.283, + 0.567, + 0.297 + ], + "angle": 0, + "content": "3 UNC Chapel Hill" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.339, + 0.745, + 0.603 + ], + "angle": 0, + "content": "Abstract. Locating specific moments within long videos (20-120 minutes) presents a significant challenge, akin to finding a needle in a haystack. Adapting existing short video (5-30 seconds) grounding methods to this problem yields poor performance. Since most real-life videos, such as those on YouTube and AR/VR, are lengthy, addressing this issue is crucial. Existing methods typically operate in two stages: clip retrieval and grounding. However, this disjoint process limits the retrieval module's fine-grained event understanding, crucial for specific moment detection. We propose RGNet which deeply integrates clip retrieval and grounding into a single network capable of processing long videos into multiple granular levels, e.g., clips and frames. Its core component is a novel transformer encoder, RG-Encoder, that unifies the two stages through shared features and mutual optimization. The encoder incorporates a sparse attention mechanism and an attention loss to model both granularity jointly. Moreover, we introduce a contrastive clip sampling technique to mimic the long video paradigm closely during training. RGNet surpasses prior methods, showcasing state-of-the-art performance on long video temporal grounding (LVTG) datasets MAD and Ego4D. The code is released at https://github.com/Tanveer81/RGNet." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.616, + 0.728, + 0.631 + ], + "angle": 0, + "content": "Keywords: Long Video Temporal Grounding \\(\\cdot\\) Moment Localization" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.659, + 0.377, + 0.675 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.693, + 0.788, + 0.816 + ], + "angle": 0, + "content": "The exponential rise in online videos has created a demand for effective content retrieval systems. The retrieval results of user-defined queries, particularly in long videos (20-120 minutes), often require locating specific moments. Most existing solutions [5,41,53,55], tailored for short videos (5-30 seconds), struggle when applied to hour-long videos, slowing down the retrieval and impacting the quality of the results [40, 54]. This challenge emerges because it is very difficult to locate short moments in a long video based on text queries, a task known as Long Video Temporal Grounding (LVTG) [14, 40]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.541, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author: hannan@dbis.ifi.lmu.de" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "Hannan et al." + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.143, + 0.72, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.334, + 0.788, + 0.419 + ], + "angle": 0, + "content": "Fig. 1: Overview of RGNet. It predicts the moment boundary specified by textual queries from an hour-long video. First, our proposed RG-Encoder maps the video and text features to a joint space and retrieves the relevant clip feature. The subsequent grounding decoder processes the retrieved features to predict the beginning and end times of the moment. The encoder parallelly operates at multiple levels of granularity (e.g., clip and frame) to achieve an end-to-end-solution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.427, + 0.788, + 0.488 + ], + "angle": 0, + "content": "A straightforward solution [16,37] for the LVTG task is to divide the video into shorter clips, retrieve the most relevant one, and apply a grounding network to predict the moment. However, the grounding network cannot rectify failure in the clip retrieval phase." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.787, + 0.519 + ], + "angle": 0, + "content": "Empirically, we evaluate the impact of the two stages in Sec. 5.3 and identify clip retrieval as the primary factor for poor performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.788, + 0.656 + ], + "angle": 0, + "content": "The existing methods have a separate module for clip retrieval, which is disjoint from their grounding network (see Fig. 2). They typically follow a text-video retrieval [13,23,29] technique to select the relevant clip. However, video retrieval only requires a high-level understanding of video topics. For example, a typical video retrieval query is \"A movie about a family surviving in a farmhouse\". In contrast, an example clip retrieval query in Fig. 1 is \"Find the moment when the mum hangs up laundry outside the farmhouse\". These specific moments from a long video require fine-grained event understanding. Thus, video retrieval models are suboptimal for these fine-grained moment localization tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.788, + 0.809 + ], + "angle": 0, + "content": "We attribute this disjoint retrieval to the poor performance of these models. However, unifying clip selection and grounding is challenging due to their distinct setups. The former is a retrieval task, whereas the latter is a regression task. Moreover, it requires modeling two levels of video granularity: clip and frame. To address these technical challenges, we propose RGNet, a unified clip Retrieval and Grounding Network (see Fig. 1). The single network enables end-to-end training of both stages. This improves the retrieval module's fine-grained event understanding by directly optimizing it with the moment annotations. Parallely, the grounding network also benefits from the strong multimodal features produced by our encoder, further enhancing the localization capability." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To achieve this unification, we propose a novel transformer encoder, RGEncoder, which enables the grounding network to perform clip retrieval. This" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.682, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RGNet" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.145, + 0.554, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.376, + 0.303, + 0.49, + 0.313 + ], + "angle": 0, + "content": "Existing Disjoint Network" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.147, + 0.721, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.601, + 0.303, + 0.715, + 0.313 + ], + "angle": 0, + "content": "Proposed Unified RGNet" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.329, + 0.784, + 0.385 + ], + "angle": 0, + "content": "Fig. 2: Unified Solution. (left) Existing methods involve a separate retrieval and grounding network. The disjoint retrieval lacks fine-grained event understanding, which is crucial for moment localization. (right) Our unified network architecture overcomes it by deeply integrating the retrieval module with the grounding objective." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.785, + 0.585 + ], + "angle": 0, + "content": "eliminates the suboptimal video retrieval network and effectively models long video clip retrieval. To enable the retrieval module to understand fine-grained events, we design a sparse-attention mechanism in the encoder. The sparse attention enables the retrieval module to focus on specific events in the video, which is more synergistic with the grounding task. Thus RG-Encoder is capable of operating both at clip and frame level granularity. We propose an Intra-Clip Attention Loss, which motivates the sparse attention to focus more on the video frames aligned with the specified event. Furthermore, we propose a negative clip mining technique to simulate clip retrieval from long videos during training. This negative mining enables our network to train on a large batch size with an inter-clip contrastive loss. The large number of negative clips in the batch closely mimics the long video paradigm and reduces the gap between the training and test phases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.586, + 0.784, + 0.629 + ], + "angle": 0, + "content": "Together, these components of RGNet bridge the gap between the two stages of LVTG, enhancing the fine-grained event understanding in hour-long videos. In summary, our contributions are fourfold:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.631, + 0.784, + 0.675 + ], + "angle": 0, + "content": "- We systematically deconstruct existing LVTG methods into clip retrieval and grounding stages. Through empirical evaluations, we discern that disjoint retrieval is the primary factor contributing to poor performance." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.676, + 0.785, + 0.735 + ], + "angle": 0, + "content": "- Based on our observations, we introduce RGNet, which integrates clip retrieval with grounding through parallel clip and frame-level modeling. This obviates the necessity for a separate video retrieval network, replaced instead by an end-to-end clip retrieval module tailored specifically for long videos." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.736, + 0.785, + 0.795 + ], + "angle": 0, + "content": "- We introduce sparse attention to the retriever and a corresponding loss to model fine-grained event understanding in long-range video. We propose a contrastive negative clip-mining strategy to simulate clip retrieval from a long video during training." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.796, + 0.785, + 0.84 + ], + "angle": 0, + "content": "- RGNet achieves state-of-the-art performance across both LVTG datasets, Ego4D [14] and MAD [40]. For instance, RGNet outperforms the previous best Ego4D method [39] by a substantial margin (9.7%)." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.631, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.128 + ], + "angle": 0, + "content": "Hannan et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.433, + 0.161 + ], + "angle": 0, + "content": "2 Related Literature" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.187, + 0.788, + 0.369 + ], + "angle": 0, + "content": "Short Video Temporal Grounding. The recent grounding methods mainly focus on short videos. The best-performing ones utilize transformer variants [3,27,34,56]. For example, Moment-DETR [22] adopted transformers for combined video and text features. Subsequent works, such as UMT [28] and QD-DETR [35], split the cross-modal and unimodal modeling of video and text features for enhanced performance. EATR [18] improves its decoder with videospecific moment queries. These models were initially designed for shorter videos. However, when applied to hour-long videos, the moments become extremely tiny, posing a challenge akin to finding a needle in a haystack. Hence, these methods perform poorly on long videos, as noted by the authors of the MAD dataset [40]. In contrast, we effectively address the task by simultaneously processing long videos at two granular levels (video clips and frames) within a single network." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.788, + 0.598 + ], + "angle": 0, + "content": "Long Video Temporal Grounding. Recently, video temporal grounding has been adapted for long videos with MAD [40] and Ego4D [14] datasets. Typically, the methods designed for long videos involve two stages. Proposal-free methods [2, 26, 41, 53, 55] segment lengthy videos into smaller parts to predict candidate moments and then rank them to obtain the final predictions. Proposal-based methods [16, 37] generate proposal clips or anchors and subsequently apply their grounding model to the retrieved proposals. Some methods, like CONE [16], and M-Guidance [2], utilize the detection transformer [3] as the grounding network to improve the grounding phase. These proposal-based models are more efficient and perform better than their counterparts. However, they need a separate retrieval module to select the proposal clips. This disjoint two-stage architecture is what we find to be the main reason for the poor performance of these models. In contrast, we propose a novel transformer encoder that solves both clip retrieval and grounding in a unified way, enabling end-to-end optimization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Video-text Retrieval. The early retrieval methods [4, 7, 21, 49, 50] designed multimodal fusion techniques to align pre-trained video and text models. To bridge the gap between pre-training and downstream tasks, large-scale pretrained video-language representations [1, 6, 9, 11, 31, 42, 45-47] have been introduced in various works. Several studies [8, 10, 19, 20, 32, 43, 44, 48, 51] have transferred the knowledge from CLIP [38] to text-video retrieval tasks. In recent studies, various sampling strategies [13, 23, 29] have been explored, enabling the model to selectively concentrate on pertinent video frames based on provided text inputs. For instance, Clip-BERT [23] introduces a sparse sampling strategy, X-Pool [13] utilizes a cross-modal attention model, and TS2-Net [29] adapts CLIP to capture temporal information and eliminate unimportant tokens. While existing methods aim to retrieve high-level video topics from a pool of video collections, moment localization demands a fine grained understanding to identify specific events within a video. Despite the similar setup to video retrieval, retrieving the corresponding clip from a long video necessitates deeper" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.682, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RGNet" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "event comprehension. Therefore, we tailored our encoder to enhance event understanding and simulated precise clip retrieval conditions during training." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.202, + 0.322, + 0.218 + ], + "angle": 0, + "content": "3 RGNet" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.235, + 0.79, + 0.371 + ], + "angle": 0, + "content": "This section presents a detailed description of our proposed unified LVTG method. As illustrated in Fig. 1, RGNet takes a long video and a text query as input and predicts the precise moment semantically correlated with the query in an end-to-end manner. The whole network is divided into the proposed RG-Encoder (Fig. 3) and a decoder (Sec. 3.3). The encoder retrieves the most relevant clip feature from the input long video. The decoder then processes the retrieved clip feature to predict the precise moment boundary. Additionally, an Intra-Clip Attention Loss and an Inter-Clip Contrastive Loss are incorporated to facilitate low-level event understanding in long-range videos." + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.397, + 0.642, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.605, + 0.788, + 0.688 + ], + "angle": 0, + "content": "Fig. 3: Overview of RG-Encoder. It takes video clips and textual query as input and retrieves the relevant clip features. First, a cross-attention fuses the clips with text, and the sparsifier masks the out-of-moment frames. Based on the mask, the retrieval attention focuses on in-moment frames (colored red) and generates clip-level context and frame-level content features. We combine the context and content to generate the retrieved clip feature." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.709, + 0.419, + 0.722 + ], + "angle": 0, + "content": "3.1 Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Let us assume we have an untrimmed long video \\( V \\) consisting of \\( T \\) frames and a query sentence \\( S \\) with \\( N \\) words corresponding to a target moment's center \\( \\tau_{c} \\) and width \\( \\tau_{w} \\). Following previous methods [16,37], we use pre-trained frozen models to extract visual features \\( V = \\{f^{1}, f^{2}, \\dots, f^{T}\\} \\in R^{T \\times D_{f}} \\) as well as textual features \\( S = \\{w^{1}, w^{2}, \\dots, w^{N}\\} \\in R^{N \\times D_{w}} \\), where \\( D_{f} \\) and \\( D_{w} \\) represent the feature dimensions of video frames and query words. We slice the long video \\( V \\) into clips of length \\( L_{c} \\) and feed them into our model. We employ a sliding window of length" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "Hannan et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.146, + 0.788, + 0.193 + ], + "angle": 0, + "content": "\\(L_{c}\\) and stride \\(L_{c} / 2\\) to obtain a collection of \\(T_{c}\\) clips \\(C = \\{C^{1}, C^{2}, \\dots, C^{T_{c}}\\} \\in R^{T_{c} \\times L_{c} \\times D_{f}}\\). Here, the \\(i^{th}\\) clip, \\(C^{i} = \\{f^{s_{i} + 1}, f^{s_{i} + 2}, \\dots, f^{s_{i} + L_{c}}\\} \\in R^{L_{c} \\times D_{f}}\\), where \\(s_{i}\\) is the start index of the clip." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.211, + 0.367, + 0.226 + ], + "angle": 0, + "content": "3.2 RG-Encoder" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.233, + 0.789, + 0.37 + ], + "angle": 0, + "content": "All the clips \\( C \\) from a single video comprise samples in a batch. The RG-Encoder (Fig. 3) processes them with the query \\( S \\) to retrieve the relevant clip feature \\( P \\). First, the cross-attention generates query-conditioned frame features \\( F \\). Then, based on the frame-query correlation calculated from \\( F \\), the sparsifier determines the relevant frames inside the clip to produce a mask \\( M \\) for retrieval attention. We incorporate a learnable token \\( R \\) to capture the context of the clip. Additionally, the frame features \\( F \\) undergo retrieval attention based on the predicted mask \\( M \\) to generate the content feature \\( F_{c} \\). The combination of context \\( R \\) and content \\( F_{c} \\) forms the retrieved clip feature \\( P \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.384, + 0.789, + 0.522 + ], + "angle": 0, + "content": "Cross Attention: To evaluate the relevance of each frame within the clip \\( C^i \\), we utilize cross-attention, where frame features act as queries and text features serve as keys and values. Cross-attention represents the frame features as a weighted average of the textual features, scaled by their mutual correlation. This process results in text-conditioned frame features \\( F^i \\), establishing a fine-grained correspondence between the modalities. The query, key and values are calculated as \\( Q^i = l_Q(C^i) \\in R^{L_c \\times D_f} \\), \\( K = l_K(S) \\in R^{N \\times D_q} \\) and \\( V = l_V(S) \\in R^{N \\times D_q} \\) respectively. Here, \\( l_Q(\\cdot), l_K(\\cdot) \\), and \\( l_V(\\cdot) \\) denote the projection layers for the query, key, and value. Then, we apply the cross-attention layer as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.401, + 0.528, + 0.785, + 0.546 + ], + "angle": 0, + "content": "\\[\nF ^ {i} = \\operatorname {s o f t m a x} \\left(Q ^ {i} K ^ {T}\\right) V + Q ^ {i} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.789, + 0.688 + ], + "angle": 0, + "content": "Sparsifier: To retrieve the clip, we need to assess the relevance of each clip based on the text query. Clip relevance is determined by aggregating the relevancy of its frames. However, given the small moment duration, most frames are unrelated to the query. Hence, we classify frames into relevant and non-relevant categories using the sparsifier to assist retrieval attention in focusing on these fine-grained events. We calculate the relevancy \\( G^{j} \\in [0,1] \\) of the \\( j^{th} \\) frame with Eq. 2. We use a differentiable Gumbel-softmax function [15, 17] for enabling end-to-end training. The detailed calculation is in the supplementary materials." + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.696, + 0.785, + 0.712 + ], + "angle": 0, + "content": "\\[\nG ^ {j} = \\operatorname {G u m b e l - s o f t m a x} \\left(\\operatorname {l i n e a r} \\left(f ^ {j}\\right)\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.84 + ], + "angle": 0, + "content": "The intra-clip attention loss optimizes this classification using the ground truth frame relevancy information described in Sec. 3.4. To distinguish irrelevant frames from their counterparts, we constrain their updates in the retrieval attention. This restriction is implemented with an attention mask computed from the learned frame relevancy \\( G^{j} \\). The attention mask \\( M \\) between \\( j^{th} \\) and \\( k^{th} \\) frame is calculated with Eq. 3. Here, the \\( j^{th} \\) and \\( k^{th} \\) frames are the attention query and key, respectively. This leads to a reduced set of relevant and informative frames for subsequent attention." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.682, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RGNet" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "equation", + "bbox": [ + 0.363, + 0.159, + 0.785, + 0.19 + ], + "angle": 0, + "content": "\\[\nM (j, k) = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} G ^ {j} > 0. 5 \\text {o r} j = k \\\\ - \\infty & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.204, + 0.784, + 0.293 + ], + "angle": 0, + "content": "Retrieval Attention: We aim for a learnable clip retrieval function to train end-to-end with the grounding objective. Thus, we devise the retrieval attention for aggregating the text-conditioned frame features based on their relevancy. To condense the \\( i^{th} \\) clip into a contextual feature embedding, we introduce a randomly initialized learnable vector \\( R^i \\). We denote this as retrieval token, which is trained with the inter-clip contrastive loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.295, + 0.785, + 0.4 + ], + "angle": 0, + "content": "First, we concatenate the retrieval token with the text-conditioned frame features to create the queries \\( Q^i = [R^i,F^i] \\) for retrieval attention. We also generate an attention mask containing all zeros for the retrieval token and concatenate it with \\( M \\). We set the key \\( K = Q^i \\), and the value \\( V = Q^i \\) to calculate the retrieval attention according to Eq. 4. We get the clip-level context feature, \\( R^i = \\tilde{Q}^{i,1} \\) and frame-level content features, \\( F_{c}^{i,j} = [\\tilde{Q}^{i,2},\\tilde{Q}^{i,3},\\dots,\\tilde{Q}^{i,L_{c} + 1}] \\). Here, \\( F_{c}^{i,j} \\) is the \\( j^{th} \\) frame feature of the \\( i^{th} \\) clip." + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.412, + 0.785, + 0.428 + ], + "angle": 0, + "content": "\\[\n\\tilde {Q} ^ {i} = \\operatorname {s o f t m a x} \\left(Q ^ {i} K ^ {T} + M\\right) V + Q ^ {i} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.785, + 0.559 + ], + "angle": 0, + "content": "Further, the context feature \\( \\mathrm{R} \\) is utilized to retrieve the clip. It undergoes a linear projection layer, \\( l_{s}(\\cdot) \\), to obtain context scores \\( S_{r} = \\operatorname{sigmoid}(l_{s}(R)) \\). We retrieve the most relevant clip based on the context score \\( S_{r} \\). Importantly, the context \\( \\mathrm{R} \\) represents the queried moment by attending to the reduced set of relevant frames. So, we fuse it with the frame features with Eq. 5 to produce stronger multimodal inputs for the decoder. It enhances the decoder's localization capability significantly. The decoder only processes features from the retrieved clip to predict the moment." + }, + { + "type": "equation", + "bbox": [ + 0.388, + 0.572, + 0.785, + 0.587 + ], + "angle": 0, + "content": "\\[\nP ^ {i, j} = F _ {c} ^ {i, j} + R ^ {i} \\times G ^ {j}, i \\in P, \\forall j \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.609, + 0.427, + 0.624 + ], + "angle": 0, + "content": "3.3 Grounding Decoder" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.784, + 0.748 + ], + "angle": 0, + "content": "We process the retrieved clip features with a transformer decoder to predict precise moment boundaries. Previous methods [16,22] feed both modalities into the decoder, introducing multiple positional information (e.g., time in the video and word order in the text) that complicates the detection task. In contrast, our encoder injects text information directly into the clip, eliminating the requirement to feed both modalities into the decoder. We incorporate learnable anchor queries [27] to represent the moment's center and width as \\((\\tau_c,\\tau_w)\\). The decoder has 2 cross-attention and 2 self-attention layers." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.77, + 0.386, + 0.785 + ], + "angle": 0, + "content": "3.4 Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Intra-Clip Attention Loss guides our sparsifier to distinguish frames within and outside of the ground truth moment. This enables the retrieval attention to focus on relevant video regions and achieve fine-grained event understanding." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "Hannan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "Specifically, the in-moment frames are required to maintain higher relevancy scores than those outside the ground truth moment. The loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.185, + 0.786, + 0.2 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {a t t n}} = \\max (0, \\Delta + S _ {c} (i, j _ {\\mathrm {o u t}}) - S _ {c} (i, j _ {\\mathrm {i n}})) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.206, + 0.784, + 0.302 + ], + "angle": 0, + "content": "In the equation, \\( S_{c} \\) is the relevancy score, \\( \\Delta \\) represents the margin, and anchor \\( i \\) is an in-moment frame. We randomly chose in-moment and out-of-moment frames \\( j_{\\mathrm{in}} \\) and \\( j_{\\mathrm{out}} \\) from the same clip. To calculate the relevancy score, the context features \\( R_{i} \\), and encoded clip \\( P^{i,j} \\) undergoes linear layers, \\( l_{r}(\\cdot) \\) and \\( l_{c}(\\cdot) \\), to obtain projected context \\( R_{proj}^{i} = l_{r}(R^{i}) \\in R^{D_{l}} \\) and content features, \\( P_{proj}^{i,j} = l_{c}(P^{i,j}) \\in R^{L_{c} \\times D_{l}} \\). Then, we calculate the frame-level relevancy score:" + }, + { + "type": "equation", + "bbox": [ + 0.423, + 0.308, + 0.786, + 0.328 + ], + "angle": 0, + "content": "\\[\nS _ {c} (i, j) = R _ {p r o j} ^ {i} \\cdot P _ {p r o j} ^ {i, j} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.345, + 0.785, + 0.464 + ], + "angle": 0, + "content": "Inter-Clip Contrastive loss. A long video contains many clips with similar environments and scenes. However, most of these clips do not contain the specific moment we are searching for. So, high-level video understanding is not enough to distinguish precise moments. During the test phase, we need to handle many such negative clips. To tackle this issue, unlike previous approaches [16], we sample a large number of negative clips and train our retriever on a large batch size. Hence, this negative clip sampling reduces the disparity between the train and test phases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.466, + 0.785, + 0.557 + ], + "angle": 0, + "content": "For this section, we denote the context feature, \\( R^{i,j} \\) for \\( i^{th} \\) proposal and \\( j^{th} \\) text query. We train our model contrastively with positive text-clip pairs \\( [R^{i,j}]_{i = j} \\) and negative pairs \\( [R^{i,j}]_{i\\neq j} \\). We use the InfoNCE loss [36] to identify the positive text-clip pair amongst a set of unrelated negative samples. Here, \\( l_{cont}(\\cdot) \\) is a linear projection layer that transforms the \\( D_{f} \\) dimensional context feature into a one-dimensional logit." + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.563, + 0.786, + 0.599 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n t}} = - \\sum_ {i} \\log \\frac {\\exp \\left(l _ {\\text {c o n t}} \\left(R ^ {i , i}\\right)\\right)}{\\sum_ {j} \\exp \\left(l _ {\\text {c o n t}} \\left(R ^ {i , j}\\right)\\right)} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.609, + 0.786, + 0.732 + ], + "angle": 0, + "content": "Grounding loss. The objective functions for grounding, which aim to locate desired moments, are adopted from the baseline approach [22]. The grounding loss \\(\\mathcal{L}_{\\mathrm{g}}\\) measures the discrepancy between the ground truth (GT) and predicted moments. A one-to-one correspondence between GT and predicted moments is established using the Hungarian algorithm. This loss includes both an \\(\\mathcal{L}1\\) loss and a generalized IoU loss (\\(\\mathcal{L}\\mathrm{gIoU}\\)). Additionally, a cross-entropy loss \\(\\mathcal{L}_{\\mathrm{CE}}\\) is employed to classify predicted moments as either foreground or background. Thus, \\(\\mathcal{L}_{\\mathrm{g}}\\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.738, + 0.786, + 0.754 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {g}} = \\lambda_ {\\mathcal {L} 1} | | \\tau - \\hat {\\tau} | | + \\lambda_ {\\mathrm {g I o U}} \\mathcal {L} _ {\\mathrm {g I o U}} (\\tau , \\hat {\\tau}) + \\lambda_ {\\mathrm {C E}} \\mathcal {L} _ {\\mathrm {C E}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.758, + 0.784, + 0.819 + ], + "angle": 0, + "content": "where \\(\\tau\\) and \\(\\hat{\\tau}\\) represent the ground-truth moment and its corresponding prediction, containing center coordinates \\(\\tau_{c}\\) and width \\(\\tau_{w}\\). The hyper-parameters \\(\\lambda_{*}\\) are used for balancing the losses. Finally, with the attention and contrastive loss, the total loss \\(\\mathcal{L}_{\\mathrm{total}}\\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.827, + 0.786, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {\\text {a t t n}} \\mathcal {L} _ {\\text {a t t n}} + \\lambda_ {\\text {c o n t}} \\mathcal {L} _ {\\text {c o n t}} + \\mathcal {L} _ {\\mathrm {g}} \\tag {10}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.682, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RGNet" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.446, + 0.163 + ], + "angle": 0, + "content": "4 Experimental Setup" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.177, + 0.525, + 0.192 + ], + "angle": 0, + "content": "4.1 Datasets and Evaluation Metric" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.202, + 0.787, + 0.294 + ], + "angle": 0, + "content": "MAD [40] is an extensive dataset comprising 1.2K hours of full-length movies and 384K natural language queries, each associated with specific moments in the videos. The videos are orders of magnitude longer than previous datasets, with an average duration of 110 minutes, while the specified text moments are on average 4.1 seconds. This small moment-to-video ratio poses a significant challenge for the grounding task." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.308, + 0.787, + 0.43 + ], + "angle": 0, + "content": "Ego4D-NLQ [14] is a large-scale egocentric video data set with multiple challenges. Specifically, we use the episodic memory benchmark Ego4D-NLQ, which requires localizing where the answer to a natural language query can be seen. It contains around 13 template questions, amounting to around 74K queries. The train, val, and test sets contain 11.3K, 3.9K, and 4.0K queries. The video length ranges from 8 to 20 minutes, with an average of 8.25 minutes, and the average moment duration is 8.3 seconds. This means the moments constitute only \\(2\\%\\) of the input video on average." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.444, + 0.787, + 0.52 + ], + "angle": 0, + "content": "Grounding Metric: For grounding, we follow previous methods [16, 40], and adopt the standard metric \\(\\mathrm{Recall}@\\mathbf{k}\\) at IoU = \\(\\theta\\) (\\(\\mathrm{Rk}_{\\theta}\\)). This metric represents the percentage of testing samples with at least one grounding prediction whose intersection over union (IoU) with the ground truth (GT) is larger than \\(\\theta\\) among the top-\\(k\\) predictions." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.535, + 0.787, + 0.595 + ], + "angle": 0, + "content": "Retrieval Metric: To assess our proposal retrieval stage independently of the grounding, we utilize the standard retrieval metric [13], Recall at Rank \\( k \\) (\\( \\mathrm{R}@\\mathbf{k} \\)). This metric calculates the percentage of GT moments present in the top-\\( k \\) retrieved proposals." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.618, + 0.461, + 0.634 + ], + "angle": 0, + "content": "4.2 Implementation Details:" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.643, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We use pre-trained CLIP [38] and EgoVLP [24] models to extract video frames from MAD and Ego4D. Text features for both datasets are extracted using CLIP. The feature extractors are frozen, and their pre-trained weights remain unchanged during training. For Ego4D, we pre-train our model with NaQ annotations from previous work [39]. The proposal length, \\( W_{c} \\) is set to 180s for MAD and 48s for Ego4D, with only the top-30 and top-5 proposals retrieved, respectively. Training on Ego4D and MAD is conducted on four Nvidia-RTX-A6000 GPUs while fine-tuning on Ego4D is performed on a single GPU. Considering the longer proposals in MAD, we utilize a batch size of 32 for Ego4D and 8 for MAD. The number of moment queries is set to 5. For loss computation, we set hyperparameters as \\( \\lambda_{L1} = 10 \\), \\( \\lambda_{\\mathrm{gIoU}} = 1 \\), \\( \\lambda_{\\mathrm{CE}} = 4 \\), \\( \\lambda_{\\mathrm{samp}} = 1 \\), \\( \\lambda_{\\mathrm{cont}} = 10 \\), and \\( \\Delta = 0.2 \\). We use Xavier initialization [12] and employ AdamW [30] with an initial learning rate of \\( 1 \\times 10^{-4} \\). RGNet is trained for 35 and 200 epochs on" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "Hannan et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.146, + 0.787, + 0.178 + ], + "angle": 0, + "content": "MAD and Ego4D datasets. The initial learning rate is reduced by one order of magnitude at epochs 25 for MAD and 120 for Ego4D." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.197, + 0.456, + 0.215 + ], + "angle": 0, + "content": "5 Results and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.225, + 0.787, + 0.271 + ], + "angle": 0, + "content": "We first compare our performance with previous state-of-the-art models. Then, we report detailed ablation studies of our proposed method and visualize the qualitative results compared to the disjoint baseline." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.294, + 0.787, + 0.476 + ], + "angle": 0, + "content": "
ModelNaQEgo4D-NLQ [14]MAD [40]Avg
R1.3R5.3R1.5R5.5AvgR1.1R5.1R1.3R5.3R1.5R5.5Avg
2D-TAN [55]X5.0412.892.025.886.463.2211.902.529.251.585.696.07
UniVTG [25]X11.747.543.257.887.60-------
VSLNet [53]✓10.2619.015.8112.6711.93-------
VLG-Net [41]X-----3.6411.662.769.311.655.995.84
M-DETR [22]X8.2323.235.0113.3712.463.6012.982.819.861.675.586.08
M-Guide [2]X-----9.3018.964.6513.062.167.409.26
SOONet [37]X8.0022.403.7611.0911.3111.2623.219.0019.645.3213.1413.59
H-Hands [52]X13.2023.307.9015.6015.00-------
CONE [16]X14.1530.338.1818.0217.678.9020.516.8716.114.109.5911.01
EgoVLP✓15.9026.389.4617.8017.38-------
ReLeR [39]✓19.3123.6211.5915.7517.57-------
Ours (Default)X18.2834.0212.0422.8921.8112.4325.129.4818.725.6110.8613.70
Ours✓20.6341.6712.4725.0824.9612.4325.129.4818.725.6110.8613.70
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.478, + 0.787, + 0.533 + ], + "angle": 0, + "content": "Table 1: Main results on Ego4D-NLQ and MAD. RGNet achieves state-of-the-art performance on both datasets. Our default network is trained without NaQ annotations [39] on Ego4D. Even without NaQ annotations, our default network shows a larger improvement, underscoring the effectiveness of our solution with limited data." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.551, + 0.55, + 0.568 + ], + "angle": 0, + "content": "5.1 Result on the Ego4D-NLQ Dataset" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.573, + 0.787, + 0.77 + ], + "angle": 0, + "content": "RGNet demonstrates state-of-the-art performance on all metrics for the Ego4D-NLQ benchmark. As shown in Table 1, RGNet improves its primary evaluation metric, the R1.3 and R5.3 score average by \\(9.7\\%\\). Separately, we outperform the previous best methods on R1.3 and R5.3 by \\(1.32\\%\\) and \\(18.1\\%\\), respectively. This improvement is consistent across other metrics as well. The unified clip and frame-level modeling in RGNet parallelly capture long- and short-range temporal dependencies, contributing to its superior performance. By modeling the entire video as a whole, RGNet learns better cross-modal alignment by contrasting visual scenarios from various events within the same video. Notably, compared to SooNet [37] and CONE [16], two prominent proposal-based methods, RGNet achieves an average score improvement of \\(13.7\\%\\) and \\(7.3\\%\\). This substantial improvement is attributed to our unified modeling of clip retrieval, whereas they separately employ a retrieval similarity search between the video and the text." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.788, + 0.496, + 0.803 + ], + "angle": 0, + "content": "5.2 Result on the MAD Dataset" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We compare our model's performance on the MAD dataset in Tab. 1. RGNet achieves state-of-the-art performance on both the R1.1 and R5.1 scores by out-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.682, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RGNet" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "performing the previous best SooNet [37] by \\(1.2\\%\\) and \\(1.9\\%\\). Together with its competitive performance on other metrics, RGNet demonstrates the importance of end-to-end modeling of long videos. Specifically, our performance of R1 is superior at all IoU thresholds, which testifies to the effectiveness of the proposed retrieval method. Notably, prior models rely on heuristic dot product similarity search applied to CLIP [38] features for the retrieval phase. In contrast, our shared retrieval feature and its unified optimization with the grounding objective drastically enhance the first stage retrieval, as seen in Tab. 2. It further helps the network capture more discriminative events crucial for detecting extremely tiny moments in hour-long videos." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.317, + 0.533, + 0.334 + ], + "angle": 0, + "content": "5.3 Disjoint vs. Unified Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.789, + 0.506 + ], + "angle": 0, + "content": "Since the long video temporal grounding (LVTG) performance relies on the first stage of clip retrieval, it does not perfectly reflect the actual moment localization capability of its grounding network. To assess the standalone grounding performance, we designed an oracle experiment operating exclusively on the clip where the ground truth moment is present. Compared to the oracle grounding, the LVTG \\( R1_{3} \\) drops by \\( 15.7\\% \\) and \\( 23.9\\% \\) for Ego4D and MAD in Tab. 2. The poor grounding in LVTG emerges from the suboptimal retrieval accuracy of the disjoint network. Hence, retrieving clips where the ground truth moment exists is crucial, independent of the subsequent grounding accuracy. Regardless of the effectiveness of moment boundary detection, if the clip containing the moment boundary cannot be accurately retrieved. Our unified approach consis" + }, + { + "type": "table", + "bbox": [ + 0.27, + 0.53, + 0.735, + 0.636 + ], + "angle": 0, + "content": "
DataModelClipRetrievalOracleGroundingLVTG
R@1R@5R1.3R5.3R1.3R5.3
Ego4dBaseline31.7164.6329.8454.0114.1530.33
Ours42.0876.2836.5363.6420.6341.67
MADBaseline12.4124.5029.4953.026.8716.11
Ours25.0150.0233.4263.439.4818.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.638, + 0.788, + 0.709 + ], + "angle": 0, + "content": "Table 2: Empirical impact of two stages. To assess standalone grounding capability, we run an oracle experiment on the clip where the ground truth moment is present. The grounding capability degrades in LVTG evaluation because of incorrect selection by the disjoint clip retrieval network. Our unified model improves retrieval significantly, which leads to more effective temporal grounding in long videos." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "tently improves all retrieval metrics compared to the disjoint baseline. Notably, we achieve a staggering \\(10.4\\%\\) and \\(12.6\\%\\) improved R@1 score for Ego4D and MAD. This improvement in clip retrieval leads to state-of-the-art LVTG performance. Moreover, the grounding also improves because of the stronger clip features generated by our encoder. Compared to the disjoint baseline, the oracle R1.3 scores improved by about \\(6.7\\%\\) and \\(4.0\\%\\) for Ego4D and MAD, respectively. Hence, the unified end-to-end architecture proves more effective at LVTG by mutually improving both stages." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "Hannan et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.402, + 0.16 + ], + "angle": 0, + "content": "5.4 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.18, + 0.788, + 0.242 + ], + "angle": 0, + "content": "This section reports detailed ablation studies on the proposed modules and loss functions. We also ablate with various clip lengths and retrieved clip numbers. Unless stated, the experiments are done with the Default network w/o NaQ annotations [39] for the Ego4D-NLQ dataset and CONE [16] as the baseline." + }, + { + "type": "table", + "bbox": [ + 0.303, + 0.27, + 0.703, + 0.395 + ], + "angle": 0, + "content": "
ModelR1.3R5.3R1.5R5.5
RGNet (Default)18.2834.0212.0422.89
w/o Retrieval Token17.8033.9911.2522.32
w/o Sparsifier16.1231.579.9120.47
w/o RG-Encoder*14.1530.338.1818.02
w/o Contrastive Loss17.4132.1210.7922.80
w/o Attention Loss16.2131.599.9120.53
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.395, + 0.788, + 0.451 + ], + "angle": 0, + "content": "Table 3: Cumulative ablation study for the proposed modules and losses of RGNet. The experiment is conducted w/o NaQ augmentation on the Ego4D dataset. In each step, we remove one of our proposed modules or losses to evaluate their individual impact. *denotes the baseline model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.563 + ], + "angle": 0, + "content": "Modules: In our initial experiments, we validate the effectiveness of each proposed module with a cumulative ablation in Tab. 3. The unified network demonstrates superior performance across all evaluation metrics, notably achieving an R1.3 score of \\(18.28\\%\\). Note that the disjoint baseline performs \\(4.2\\%\\) worse than our unified network. This performance difference can be attributed to our models unified architecture." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.788, + 0.703 + ], + "angle": 0, + "content": "With the removal of the retrieval token for generating the clip feature in Eq. 5, considerably drops the score by about \\(0.5\\%\\). This proves that modeling both clip level context and frame level content are necessary to produce strong multimodal features for the grounding decoder. Subsequently, eliminating the sparsifier, there is a decline in \\(\\mathrm{R1}_{.3}\\) performance of approximately \\(1.7\\%\\). This verifies the impact of the sparsifier in modeling the fine-grained event boundaries. Finally, removing the RG-Encoder yields a notable degradation of approximately \\(2.0\\%\\) in the \\(\\mathrm{R1}_{.3}\\) score, which substantiates its capability to model clip and frame level granularity jointly." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.718, + 0.788, + 0.793 + ], + "angle": 0, + "content": "Loss Functions: Training RGNet without our negative clip sampling strategy for the contrastive loss (Tab. 3) results in an almost \\(0.9\\%\\) decline in the R1.3 score. Our sampling strategy mimics long-video setup closely during training, which improves event discrimination capability inside the same scene and surroundings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "By eliminating the attention loss, we observe a further reduction of \\(1.2\\%\\) in grounding performance. The attention loss enables our encoder to model fine-grained events in long videos which is crucial for specific moment detection." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.682, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RGNet" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.162, + 0.788, + 0.298 + ], + "angle": 0, + "content": "Number of Clips: The number of retrieved proposal clips significantly influences the Long Video Temporal Grounding (LVTG) performance. Our model consistently outperforms the baseline across various choices of top-k retrieved clips, as shown in Fig. 4. Even with only a single retrieved clip, our performance surpasses the baseline's optimal result for both datasets. Moreover, our model exhibits a less steep performance decline when fewer clips are retrieved. Importantly, with fewer clips, the grounding network runs fewer times, leading to an overall speedup. Therefore, our model strikes a favorable balance between efficiency and performance, showcasing superior performance than the baseline." + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.326, + 0.498, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.328, + 0.672, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.504 + ], + "angle": 0, + "content": "Fig. 4: Impact of number of retrieved clips. Reducing the number of clips speeds up the network execution time. While the baseline model experiences a significant drop in performance with this reduction, RGNet shows a noticeably smaller decline in performance under the same conditions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.527, + 0.788, + 0.664 + ], + "angle": 0, + "content": "**Clip Length:** Varying clip length impacts both retrieval and grounding performance. The number of clip decreases with longer clips, making the retrieval task easier. Fig. 5 reports a monotonically improving retrieval with longer clips. However, after a certain length, moment localization quality starts to deteriorate. With longer clips, the task becomes extremely difficult. We see the decline after the 180s and 48s for the MAD and Ego4D datasets. Hence, we set the default clip length of our method to these optimal lengths. Importantly, the unified network helps outscore the disjoint baseline on all tested clip lengths (inversely proportional to the clip numbers in Fig. 4)." + }, + { + "type": "image", + "bbox": [ + 0.328, + 0.69, + 0.495, + 0.79 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.69, + 0.678, + 0.79 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.804, + 0.788, + 0.86 + ], + "angle": 0, + "content": "Fig. 5: Impact of retrieved clip length. Longer clips result in improved retrieval due to fewer candidates. However, grounding becomes exceedingly difficult with longer clips. For example, the grounding performance drops after 180 seconds and 48 seconds for the MAD and Ego4D datasets." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "Hannan et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.43, + 0.162 + ], + "angle": 0, + "content": "5.5 Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.171, + 0.788, + 0.323 + ], + "angle": 0, + "content": "We visualize the qualitative predictions of RGNet compared to the disjoint baseline in Fig. 6. For the first and second queries, the baseline struggles to retrieve the correct clip where the scissor appears. RGNet accurately retrieves the clip with the searched object and precisely localizes the moment it appeared in the video. Disjoint retrieval often leads to incorrect clip selection, which the final grounding network cannot recover from. For the third query, the baseline selects the correct clip but fails to localize the precise moment the man closes the car door, indicating a gap between the two stages. RGNet mitigates this gap by leveraging features from the retrieved clip to improve moment localization. More visual outputs are included the supplementary materials." + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.347, + 0.555, + 0.359 + ], + "angle": 0, + "content": "Query 1: Where is the scissor?" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.36, + 0.731, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.378, + 0.42, + 0.596, + 0.433 + ], + "angle": 0, + "content": "Query 2: How many stainless bowls did I pick?" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.436, + 0.73, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.395, + 0.502, + 0.58, + 0.514 + ], + "angle": 0, + "content": "Query 3: Did I leave the car door open?" + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.515, + 0.685, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.579, + 0.788, + 0.664 + ], + "angle": 0, + "content": "Fig. 6: Qualitative Results. The baseline fails to retrieve the correct clip in the first two queries. Since they primarily depict the same indoor room throughout the whole video, precise event discrimination is vital for accurate clip retrieval. In the third query, the baseline cannot identify the moment within the correctly retrieved clip, detecting it only after it has concluded. Precise event localization demands improved alignment between visual events and the queried text. RGNet correctly localizes all the moments." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.674, + 0.36, + 0.69 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We introduce an end-to-end model for long video temporal grounding, unifying the two stages of prevailing methods with shared features and mutual optimization. We conduct independent analyses of the two stages, shaping our solution based on their distinct impacts. This leads to a better understanding of fine-grained events in long videos. Our approach demonstrates state-of-the-art performance on challenging long video grounding datasets, validating its effectiveness. Like all LVTG methods, we rely on a pre-trained image encoder to extract the visual features. A promising future direction includes eliminating the need for an image encoder and training directly on the raw video data." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.682, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RGNet" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.2 + ], + "angle": 0, + "content": "1. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: ICCV. pp. 1728-1738 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.201, + 0.785, + 0.242 + ], + "angle": 0, + "content": "2. Barrios, W., Soldan, M., Ceballos-Arroyo, A.M., Heilbron, F.C., Ghanem, B.: Localizing moments in long video via multimodal guidance. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13667-13678 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.242, + 0.785, + 0.283 + ], + "angle": 0, + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.283, + 0.785, + 0.31 + ], + "angle": 0, + "content": "4. Chen, S., Zhao, Y., Jin, Q., Wu, Q.: Fine-grained video-text retrieval with hierarchical graph reasoning. In: CVPR. pp. 10638-10647 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.785, + 0.351 + ], + "angle": 0, + "content": "5. Chen, Z., Jiang, X., Xu, X., Cao, Z., Mo, Y., Shen, H.T.: Joint searching and grounding: Multi-granularity video content retrieval. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 975-983 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.351, + 0.785, + 0.392 + ], + "angle": 0, + "content": "6. Cheng, F., Wang, X., Lei, J., Crandall, D., Bansal, M., Bertasius, G.: Vindlu: A recipe for effective video-and-language pretraining. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.392, + 0.785, + 0.418 + ], + "angle": 0, + "content": "7. Dong, J., Li, X., Xu, C., Yang, X., Yang, G., Wang, X., Wang, M.: Dual encoding for video retrieval by text. IEEE TPAMI pp. 4065-4080 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.419, + 0.785, + 0.459 + ], + "angle": 0, + "content": "8. Fang, B., Liu, C., Zhou, Y., Yang, M., Song, Y., Li, F., Wang, W., Ji, X., Ouyang, W., et al.: Uatvr: Uncertainty-adaptive text-video retrieval. arXiv preprint arXiv:2301.06309 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.459, + 0.785, + 0.486 + ], + "angle": 0, + "content": "9. Gabeur, V., Sun, C., Alahari, K., Schmid, C.: Multi-modal transformer for video retrieval. In: ECCV. pp. 214-229 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.486, + 0.785, + 0.514 + ], + "angle": 0, + "content": "0. Gao, Z., Liu, J., Sun, W., Chen, S., Chang, D., Zhao, L.: Clip2tv: Align, match and distill for video-text retrieval. arXiv preprint arXiv:2111.05610 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.514, + 0.785, + 0.54 + ], + "angle": 0, + "content": "1. Ge, Y., Ge, Y., Liu, X., Li, D., Shan, Y., Qie, X., Luo, P.: Bridging video-text retrieval with multiple choice questions. In: CVPR. pp. 16167-16176 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.54, + 0.785, + 0.595 + ], + "angle": 0, + "content": "2. Glorot, X., Bengio, Y.: Understanding the difficulty of training deep feedforward neural networks. In: Proceedings of the thirteenth international conference on artificial intelligence and statistics. pp. 249-256. JMLR Workshop and Conference Proceedings (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.595, + 0.785, + 0.65 + ], + "angle": 0, + "content": "3. Gorti, S.K., Vouitsis, N., Ma, J., Golestan, K., Volkovs, M., Garg, A., Yu, G.: X-pool: Cross-modal language-video attention for text-video retrieval. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5006-5015 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.65, + 0.785, + 0.691 + ], + "angle": 0, + "content": "4. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: CVPR. pp. 18995-19012 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.785, + 0.732 + ], + "angle": 0, + "content": "5. Hannan, T., Koner, R., Bernhard, M., Shit, S., Menze, B., Tresp, V., Schubert, M., Seidl, T.: Gratt-vis: Gated residual attention for auto rectifying video instance segmentation. arXiv preprint arXiv:2305.17096 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.732, + 0.785, + 0.772 + ], + "angle": 0, + "content": "6. Hou, Z., Zhong, W., Ji, L., Gao, D., Yan, K., Chan, W.K., Ngo, C.W., Shou, Z., Duan, N.: Cone: An efficient coarse-to-fine alignment framework for long video temporal grounding. arXiv preprint arXiv:2209.10918 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.8 + ], + "angle": 0, + "content": "7. Jang, E., Gu, S., Poole, B.: Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "8. Jang, J., Park, J., Kim, J., Kwon, H., Sohn, K.: Knowing where to focus: Event-aware transformer for video grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13846-13856 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "Hannan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "19. Jiang, H., Zhang, J., Huang, R., Ge, C., Ni, Z., Lu, J., Zhou, J., Song, S., Huang, G.: Cross-modal adapter for text-video retrieval. arXiv preprint arXiv:2211.09623 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.786, + 0.232 + ], + "angle": 0, + "content": "20. Jin, P., Huang, J., Liu, F., Wu, X., Ge, S., Song, G., Clifton, D.A., Chen, J.: Expectation-maximization contrastive learning for compact video-and-language representations. NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.786, + 0.259 + ], + "angle": 0, + "content": "21. Kiros, R., Salakhutdinov, R., Zemel, R.S.: Unifying visual-semantic embeddings with multimodal neural language models. arXiv preprint arXiv:1411.2539 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.786, + 0.3 + ], + "angle": 0, + "content": "22. Lei, J., Berg, T.L., Bansal, M.: Detecting moments and highlights in videos via natural language queries. Advances in Neural Information Processing Systems 34, 11846-11858 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.786, + 0.355 + ], + "angle": 0, + "content": "23. Lei, J., Li, L., Zhou, L., Gan, Z., Berg, T.L., Bansal, M., Liu, J.: Less is more: Clipbert for video-and-language learning via sparse sampling. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7331-7341 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.786, + 0.398 + ], + "angle": 0, + "content": "24. Lin, K.Q., Wang, A.J., Soldan, M., Wray, M., Yan, R., Xu, E.Z., Gao, D., Tu, R., Zhao, W., Kong, W., et al.: Egocentric video-language pretraining. arXiv preprint arXiv:2206.01670 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.786, + 0.453 + ], + "angle": 0, + "content": "25. Lin, K.Q., Zhang, P., Chen, J., Pramanick, S., Gao, D., Wang, A.J., Yan, R., Shou, M.Z.: Univtg: Towards unified video-language temporal grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2794-2804 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.786, + 0.495 + ], + "angle": 0, + "content": "26. Liu, N., Wang, X., Li, X., Yang, Y., Zhuang, Y.: Refer@ zju-alibaba submission to the ego4d natural language queries challenge 2022. arXiv preprint arXiv:2207.00383 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.786, + 0.536 + ], + "angle": 0, + "content": "27. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.786, + 0.592 + ], + "angle": 0, + "content": "28. Liu, Y., Li, S., Wu, Y., Chen, C.W., Shan, Y., Qie, X.: Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3042-3051 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.786, + 0.619 + ], + "angle": 0, + "content": "29. Liu, Y., Xiong, P., Xu, L., Cao, S., Jin, Q.: Ts2-net: Token shift and selection transformer for text-video retrieval. In: ECCV. pp. 319-335 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.786, + 0.647 + ], + "angle": 0, + "content": "30. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.786, + 0.688 + ], + "angle": 0, + "content": "31. Luo, H., Ji, L., Shi, B., Huang, H., Duan, N., Li, T., Li, J., Bharti, T., Zhou, M.: Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.688, + 0.786, + 0.73 + ], + "angle": 0, + "content": "32. Ma, Y., Xu, G., Sun, X., Yan, M., Zhang, J., Ji, R.: X-clip: End-to-end multigrained contrastive learning for video-text retrieval. In: ACM MM. pp. 638-647 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.786, + 0.771 + ], + "angle": 0, + "content": "33. Maddison, C.J., Mnih, A., Teh, Y.W.: The concrete distribution: A continuous relaxation of discrete random variables. CoRR abs/1611.00712 (2016), http://arxiv.org/abs/1611.00712" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.786, + 0.813 + ], + "angle": 0, + "content": "34. Meng, D., Chen, X., Fan, Z., Zeng, G., Li, H., Yuan, Y., Sun, L., Wang, J.: Conditional detr for fast training convergence. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3651-3660 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.786, + 0.84 + ], + "angle": 0, + "content": "35. Moon, W., Hyun, S., Park, S., Park, D., Heo, J.P.: Query-dependent video representation for moment retrieval and highlight detection. In: Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.682, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RGNet" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.147, + 0.786, + 0.175 + ], + "angle": 0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23023-23033 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.786, + 0.203 + ], + "angle": 0, + "content": "36. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "37. Pan, Y., He, X., Gong, B., Lv, Y., Shen, Y., Peng, Y., Zhao, D.: Scanning only once: An end-to-end framework for fast temporal grounding in long videos. arXiv preprint arXiv:2303.08345 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.786, + 0.286 + ], + "angle": 0, + "content": "38. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. pp. 8748-8763 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.786, + 0.327 + ], + "angle": 0, + "content": "39. Ramakrishnan, S.K., Al-Halah, Z., Grauman, K.: Naq: Leveraging narrations as queries to supervise episodic memory. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6694-6703 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.786, + 0.37 + ], + "angle": 0, + "content": "40. Soldan, M., Pardo, A., Alcázar, J.L., Caba, F., Zhao, C., Giancola, S., Ghanem, B.: Mad: A scalable dataset for language grounding in videos from movie audio descriptions. In: CVPR. pp. 5026-5035 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.786, + 0.397 + ], + "angle": 0, + "content": "41. Soldan, M., Xu, M., Qu, S., Tegner, J., Ghanem, B.: Vlg-net: Video-language graph matching network for video grounding. In: ICCV. pp. 3224-3234 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.786, + 0.438 + ], + "angle": 0, + "content": "42. Wang, A.J., Ge, Y., Yan, R., Ge, Y., Lin, X., Cai, G., Wu, J., Shan, Y., Qie, X., Shou, M.Z.: All in one: Exploring unified video-language pre-training. arXiv preprint arXiv:2203.07303 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.786, + 0.466 + ], + "angle": 0, + "content": "43. Wang, Q., Zhang, Y., Zheng, Y., Pan, P., Hua, X.S.: Disentangled representation learning for text-video retrieval. arXiv preprint arXiv:2203.07111 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.786, + 0.508 + ], + "angle": 0, + "content": "44. Wang, Z., Sung, Y.L., Cheng, F., Bertasius, G., Bansal, M.: Unified coarse-to-fine alignment for video-text retrieval. In: The IEEE International Conference on Computer Vision (ICCV) (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.786, + 0.55 + ], + "angle": 0, + "content": "45. Xu, H., Ghosh, G., Huang, P.Y., Arora, P., Aminzadeh, M., Feichtenhofer, C., Metze, F., Zettlemoyer, L.: Vlm: Task-agnostic video-language model pre-training for video understanding. ACLliu2021hit (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.786, + 0.591 + ], + "angle": 0, + "content": "46. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettlemoyer, L., Feichtenhofer, C.: Videoclip: Contrastive pre-training for zero-shot video-text understanding. In: EMNLP. pp. 6787-6800 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.786, + 0.632 + ], + "angle": 0, + "content": "47. Xue, H., Hang, T., Zeng, Y., Sun, Y., Liu, B., Yang, H., Fu, J., Guo, B.: Advancing high-resolution video-language representation with large-scale video transcriptions. In: CVPR. pp. 5036-5045 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.786, + 0.674 + ], + "angle": 0, + "content": "48. Xue, H., Sun, Y., Liu, B., Fu, J., Song, R., Li, H., Luo, J.: Clip-vip: Adapting pre-trained image-text model to video-language representation alignment. ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.786, + 0.702 + ], + "angle": 0, + "content": "49. Yu, Y., Kim, J., Kim, G.: A joint sequence fusion model for video question answering and retrieval. In: ECCV. pp. 471-487 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.786, + 0.73 + ], + "angle": 0, + "content": "50. Yu, Y., Ko, H., Choi, J., Kim, G.: End-to-end concept word detection for video captioning, retrieval, and question answering. In: CVPR. pp. 3165-3173 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.786, + 0.77 + ], + "angle": 0, + "content": "51. Zhang, B., Jin, X., Gong, W., Xu, K., Zhang, Z., Wang, P., Shen, X., Feng, J.: Multimodal video adapter for parameter efficient video text retrieval. arXiv preprint arXiv:2301.07868 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.786, + 0.812 + ], + "angle": 0, + "content": "52. Zhang, C., Gupta, A., Zisserman, A.: Helping hands: An object-aware ego-centric video recognition model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13901-13912 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.786, + 0.84 + ], + "angle": 0, + "content": "53. Zhang, H., Sun, A., Jing, W., Zhou, J.T.: Span-based localizing network for natural language video localization. arXiv preprint arXiv:2004.13931 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "Hannan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "54. Zhang, H., Sun, A., Jing, W., Zhou, J.T.: The elements of temporal sentence grounding in videos: A survey and future directions. arXiv preprint arXiv:2201.08071 1(2) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "55. Zhang, S., Peng, H., Fu, J., Luo, J.: Learning 2d temporal adjacent networks for moment localization with natural language. In: AAAI. pp. 12870-12877 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "56. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.259 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/RGNet_ A Unified Clip Retrieval and Grounding Network for Long Videos/f091b053-bc8d-4374-9687-69b5bf80ade6_origin.pdf b/2024/RGNet_ A Unified Clip Retrieval and Grounding Network for Long Videos/f091b053-bc8d-4374-9687-69b5bf80ade6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..90196c2d2b8035ab4c2273e94b7d0f1e4d34d027 --- /dev/null +++ b/2024/RGNet_ A Unified Clip Retrieval and Grounding Network for Long Videos/f091b053-bc8d-4374-9687-69b5bf80ade6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf76200c0dbc07fc2689cf73572a33f77ae32868017f67ae51ecfc75514f6015 +size 7583784 diff --git a/2024/RGNet_ A Unified Clip Retrieval and Grounding Network for Long Videos/full.md b/2024/RGNet_ A Unified Clip Retrieval and Grounding Network for Long Videos/full.md new file mode 100644 index 0000000000000000000000000000000000000000..933d0f49e47b870d8e596fd9d88681083b155da9 --- /dev/null +++ b/2024/RGNet_ A Unified Clip Retrieval and Grounding Network for Long Videos/full.md @@ -0,0 +1,302 @@ +# RGNet: A Unified Clip Retrieval and Grounding Network for Long Videos + +Tanveer Hannan $^{1,2\star}$ Md Mohaiminul Islam $^{3}$ Thomas Seidl $^{1,2}$ Gedas Bertasius $^{3}$ + +1 LMU Munich + +2 MCML + +3 UNC Chapel Hill + +Abstract. Locating specific moments within long videos (20-120 minutes) presents a significant challenge, akin to finding a needle in a haystack. Adapting existing short video (5-30 seconds) grounding methods to this problem yields poor performance. Since most real-life videos, such as those on YouTube and AR/VR, are lengthy, addressing this issue is crucial. Existing methods typically operate in two stages: clip retrieval and grounding. However, this disjoint process limits the retrieval module's fine-grained event understanding, crucial for specific moment detection. We propose RGNet which deeply integrates clip retrieval and grounding into a single network capable of processing long videos into multiple granular levels, e.g., clips and frames. Its core component is a novel transformer encoder, RG-Encoder, that unifies the two stages through shared features and mutual optimization. The encoder incorporates a sparse attention mechanism and an attention loss to model both granularity jointly. Moreover, we introduce a contrastive clip sampling technique to mimic the long video paradigm closely during training. RGNet surpasses prior methods, showcasing state-of-the-art performance on long video temporal grounding (LVTG) datasets MAD and Ego4D. The code is released at https://github.com/Tanveer81/RGNet. + +Keywords: Long Video Temporal Grounding $\cdot$ Moment Localization + +# 1 Introduction + +The exponential rise in online videos has created a demand for effective content retrieval systems. The retrieval results of user-defined queries, particularly in long videos (20-120 minutes), often require locating specific moments. Most existing solutions [5,41,53,55], tailored for short videos (5-30 seconds), struggle when applied to hour-long videos, slowing down the retrieval and impacting the quality of the results [40, 54]. This challenge emerges because it is very difficult to locate short moments in a long video based on text queries, a task known as Long Video Temporal Grounding (LVTG) [14, 40]. + +![](images/a4aa5a3c60a7093dea15bbe7642f13a4f87b7a9a0be3ffc447826395f03b1919.jpg) +Fig. 1: Overview of RGNet. It predicts the moment boundary specified by textual queries from an hour-long video. First, our proposed RG-Encoder maps the video and text features to a joint space and retrieves the relevant clip feature. The subsequent grounding decoder processes the retrieved features to predict the beginning and end times of the moment. The encoder parallelly operates at multiple levels of granularity (e.g., clip and frame) to achieve an end-to-end-solution. + +A straightforward solution [16,37] for the LVTG task is to divide the video into shorter clips, retrieve the most relevant one, and apply a grounding network to predict the moment. However, the grounding network cannot rectify failure in the clip retrieval phase. + +Empirically, we evaluate the impact of the two stages in Sec. 5.3 and identify clip retrieval as the primary factor for poor performance. + +The existing methods have a separate module for clip retrieval, which is disjoint from their grounding network (see Fig. 2). They typically follow a text-video retrieval [13,23,29] technique to select the relevant clip. However, video retrieval only requires a high-level understanding of video topics. For example, a typical video retrieval query is "A movie about a family surviving in a farmhouse". In contrast, an example clip retrieval query in Fig. 1 is "Find the moment when the mum hangs up laundry outside the farmhouse". These specific moments from a long video require fine-grained event understanding. Thus, video retrieval models are suboptimal for these fine-grained moment localization tasks. + +We attribute this disjoint retrieval to the poor performance of these models. However, unifying clip selection and grounding is challenging due to their distinct setups. The former is a retrieval task, whereas the latter is a regression task. Moreover, it requires modeling two levels of video granularity: clip and frame. To address these technical challenges, we propose RGNet, a unified clip Retrieval and Grounding Network (see Fig. 1). The single network enables end-to-end training of both stages. This improves the retrieval module's fine-grained event understanding by directly optimizing it with the moment annotations. Parallely, the grounding network also benefits from the strong multimodal features produced by our encoder, further enhancing the localization capability. + +To achieve this unification, we propose a novel transformer encoder, RGEncoder, which enables the grounding network to perform clip retrieval. This + +![](images/7ee1042d6ed0acb37be653a312ba6baf8af68724ee30a15c67925808283b8e4a.jpg) +Existing Disjoint Network + +![](images/3d4b369cc3ea41360acbb01f69fc4e95f9ed760594bfee40902fac5a3eeaa769.jpg) +Proposed Unified RGNet +Fig. 2: Unified Solution. (left) Existing methods involve a separate retrieval and grounding network. The disjoint retrieval lacks fine-grained event understanding, which is crucial for moment localization. (right) Our unified network architecture overcomes it by deeply integrating the retrieval module with the grounding objective. + +eliminates the suboptimal video retrieval network and effectively models long video clip retrieval. To enable the retrieval module to understand fine-grained events, we design a sparse-attention mechanism in the encoder. The sparse attention enables the retrieval module to focus on specific events in the video, which is more synergistic with the grounding task. Thus RG-Encoder is capable of operating both at clip and frame level granularity. We propose an Intra-Clip Attention Loss, which motivates the sparse attention to focus more on the video frames aligned with the specified event. Furthermore, we propose a negative clip mining technique to simulate clip retrieval from long videos during training. This negative mining enables our network to train on a large batch size with an inter-clip contrastive loss. The large number of negative clips in the batch closely mimics the long video paradigm and reduces the gap between the training and test phases. + +Together, these components of RGNet bridge the gap between the two stages of LVTG, enhancing the fine-grained event understanding in hour-long videos. In summary, our contributions are fourfold: + +- We systematically deconstruct existing LVTG methods into clip retrieval and grounding stages. Through empirical evaluations, we discern that disjoint retrieval is the primary factor contributing to poor performance. +- Based on our observations, we introduce RGNet, which integrates clip retrieval with grounding through parallel clip and frame-level modeling. This obviates the necessity for a separate video retrieval network, replaced instead by an end-to-end clip retrieval module tailored specifically for long videos. +- We introduce sparse attention to the retriever and a corresponding loss to model fine-grained event understanding in long-range video. We propose a contrastive negative clip-mining strategy to simulate clip retrieval from a long video during training. +- RGNet achieves state-of-the-art performance across both LVTG datasets, Ego4D [14] and MAD [40]. For instance, RGNet outperforms the previous best Ego4D method [39] by a substantial margin (9.7%). + +# 2 Related Literature + +Short Video Temporal Grounding. The recent grounding methods mainly focus on short videos. The best-performing ones utilize transformer variants [3,27,34,56]. For example, Moment-DETR [22] adopted transformers for combined video and text features. Subsequent works, such as UMT [28] and QD-DETR [35], split the cross-modal and unimodal modeling of video and text features for enhanced performance. EATR [18] improves its decoder with videospecific moment queries. These models were initially designed for shorter videos. However, when applied to hour-long videos, the moments become extremely tiny, posing a challenge akin to finding a needle in a haystack. Hence, these methods perform poorly on long videos, as noted by the authors of the MAD dataset [40]. In contrast, we effectively address the task by simultaneously processing long videos at two granular levels (video clips and frames) within a single network. + +Long Video Temporal Grounding. Recently, video temporal grounding has been adapted for long videos with MAD [40] and Ego4D [14] datasets. Typically, the methods designed for long videos involve two stages. Proposal-free methods [2, 26, 41, 53, 55] segment lengthy videos into smaller parts to predict candidate moments and then rank them to obtain the final predictions. Proposal-based methods [16, 37] generate proposal clips or anchors and subsequently apply their grounding model to the retrieved proposals. Some methods, like CONE [16], and M-Guidance [2], utilize the detection transformer [3] as the grounding network to improve the grounding phase. These proposal-based models are more efficient and perform better than their counterparts. However, they need a separate retrieval module to select the proposal clips. This disjoint two-stage architecture is what we find to be the main reason for the poor performance of these models. In contrast, we propose a novel transformer encoder that solves both clip retrieval and grounding in a unified way, enabling end-to-end optimization. + +Video-text Retrieval. The early retrieval methods [4, 7, 21, 49, 50] designed multimodal fusion techniques to align pre-trained video and text models. To bridge the gap between pre-training and downstream tasks, large-scale pretrained video-language representations [1, 6, 9, 11, 31, 42, 45-47] have been introduced in various works. Several studies [8, 10, 19, 20, 32, 43, 44, 48, 51] have transferred the knowledge from CLIP [38] to text-video retrieval tasks. In recent studies, various sampling strategies [13, 23, 29] have been explored, enabling the model to selectively concentrate on pertinent video frames based on provided text inputs. For instance, Clip-BERT [23] introduces a sparse sampling strategy, X-Pool [13] utilizes a cross-modal attention model, and TS2-Net [29] adapts CLIP to capture temporal information and eliminate unimportant tokens. While existing methods aim to retrieve high-level video topics from a pool of video collections, moment localization demands a fine grained understanding to identify specific events within a video. Despite the similar setup to video retrieval, retrieving the corresponding clip from a long video necessitates deeper + +event comprehension. Therefore, we tailored our encoder to enhance event understanding and simulated precise clip retrieval conditions during training. + +# 3 RGNet + +This section presents a detailed description of our proposed unified LVTG method. As illustrated in Fig. 1, RGNet takes a long video and a text query as input and predicts the precise moment semantically correlated with the query in an end-to-end manner. The whole network is divided into the proposed RG-Encoder (Fig. 3) and a decoder (Sec. 3.3). The encoder retrieves the most relevant clip feature from the input long video. The decoder then processes the retrieved clip feature to predict the precise moment boundary. Additionally, an Intra-Clip Attention Loss and an Inter-Clip Contrastive Loss are incorporated to facilitate low-level event understanding in long-range videos. + +![](images/172a7e7757a46c7be9bbd3ae7f46b1f5136b4f36025c4e25396febda8144b69c.jpg) +Fig. 3: Overview of RG-Encoder. It takes video clips and textual query as input and retrieves the relevant clip features. First, a cross-attention fuses the clips with text, and the sparsifier masks the out-of-moment frames. Based on the mask, the retrieval attention focuses on in-moment frames (colored red) and generates clip-level context and frame-level content features. We combine the context and content to generate the retrieved clip feature. + +# 3.1 Feature Extraction + +Let us assume we have an untrimmed long video $V$ consisting of $T$ frames and a query sentence $S$ with $N$ words corresponding to a target moment's center $\tau_{c}$ and width $\tau_{w}$ . Following previous methods [16,37], we use pre-trained frozen models to extract visual features $V = \{f^{1}, f^{2}, \dots, f^{T}\} \in R^{T \times D_{f}}$ as well as textual features $S = \{w^{1}, w^{2}, \dots, w^{N}\} \in R^{N \times D_{w}}$ , where $D_{f}$ and $D_{w}$ represent the feature dimensions of video frames and query words. We slice the long video $V$ into clips of length $L_{c}$ and feed them into our model. We employ a sliding window of length + +$L_{c}$ and stride $L_{c} / 2$ to obtain a collection of $T_{c}$ clips $C = \{C^{1}, C^{2}, \dots, C^{T_{c}}\} \in R^{T_{c} \times L_{c} \times D_{f}}$ . Here, the $i^{th}$ clip, $C^{i} = \{f^{s_{i} + 1}, f^{s_{i} + 2}, \dots, f^{s_{i} + L_{c}}\} \in R^{L_{c} \times D_{f}}$ , where $s_{i}$ is the start index of the clip. + +# 3.2 RG-Encoder + +All the clips $C$ from a single video comprise samples in a batch. The RG-Encoder (Fig. 3) processes them with the query $S$ to retrieve the relevant clip feature $P$ . First, the cross-attention generates query-conditioned frame features $F$ . Then, based on the frame-query correlation calculated from $F$ , the sparsifier determines the relevant frames inside the clip to produce a mask $M$ for retrieval attention. We incorporate a learnable token $R$ to capture the context of the clip. Additionally, the frame features $F$ undergo retrieval attention based on the predicted mask $M$ to generate the content feature $F_{c}$ . The combination of context $R$ and content $F_{c}$ forms the retrieved clip feature $P$ . + +Cross Attention: To evaluate the relevance of each frame within the clip $C^i$ , we utilize cross-attention, where frame features act as queries and text features serve as keys and values. Cross-attention represents the frame features as a weighted average of the textual features, scaled by their mutual correlation. This process results in text-conditioned frame features $F^i$ , establishing a fine-grained correspondence between the modalities. The query, key and values are calculated as $Q^i = l_Q(C^i) \in R^{L_c \times D_f}$ , $K = l_K(S) \in R^{N \times D_q}$ and $V = l_V(S) \in R^{N \times D_q}$ respectively. Here, $l_Q(\cdot), l_K(\cdot)$ , and $l_V(\cdot)$ denote the projection layers for the query, key, and value. Then, we apply the cross-attention layer as follows: + +$$ +F ^ {i} = \operatorname {s o f t m a x} \left(Q ^ {i} K ^ {T}\right) V + Q ^ {i} \tag {1} +$$ + +Sparsifier: To retrieve the clip, we need to assess the relevance of each clip based on the text query. Clip relevance is determined by aggregating the relevancy of its frames. However, given the small moment duration, most frames are unrelated to the query. Hence, we classify frames into relevant and non-relevant categories using the sparsifier to assist retrieval attention in focusing on these fine-grained events. We calculate the relevancy $G^{j} \in [0,1]$ of the $j^{th}$ frame with Eq. 2. We use a differentiable Gumbel-softmax function [15, 17] for enabling end-to-end training. The detailed calculation is in the supplementary materials. + +$$ +G ^ {j} = \operatorname {G u m b e l - s o f t m a x} \left(\operatorname {l i n e a r} \left(f ^ {j}\right)\right) \tag {2} +$$ + +The intra-clip attention loss optimizes this classification using the ground truth frame relevancy information described in Sec. 3.4. To distinguish irrelevant frames from their counterparts, we constrain their updates in the retrieval attention. This restriction is implemented with an attention mask computed from the learned frame relevancy $G^{j}$ . The attention mask $M$ between $j^{th}$ and $k^{th}$ frame is calculated with Eq. 3. Here, the $j^{th}$ and $k^{th}$ frames are the attention query and key, respectively. This leads to a reduced set of relevant and informative frames for subsequent attention. + +$$ +M (j, k) = \left\{ \begin{array}{l l} 0 & \text {i f} G ^ {j} > 0. 5 \text {o r} j = k \\ - \infty & \text {o t h e r w i s e} \end{array} \right. \tag {3} +$$ + +Retrieval Attention: We aim for a learnable clip retrieval function to train end-to-end with the grounding objective. Thus, we devise the retrieval attention for aggregating the text-conditioned frame features based on their relevancy. To condense the $i^{th}$ clip into a contextual feature embedding, we introduce a randomly initialized learnable vector $R^i$ . We denote this as retrieval token, which is trained with the inter-clip contrastive loss. + +First, we concatenate the retrieval token with the text-conditioned frame features to create the queries $Q^i = [R^i,F^i]$ for retrieval attention. We also generate an attention mask containing all zeros for the retrieval token and concatenate it with $M$ . We set the key $K = Q^i$ , and the value $V = Q^i$ to calculate the retrieval attention according to Eq. 4. We get the clip-level context feature, $R^i = \tilde{Q}^{i,1}$ and frame-level content features, $F_{c}^{i,j} = [\tilde{Q}^{i,2},\tilde{Q}^{i,3},\dots,\tilde{Q}^{i,L_{c} + 1}]$ . Here, $F_{c}^{i,j}$ is the $j^{th}$ frame feature of the $i^{th}$ clip. + +$$ +\tilde {Q} ^ {i} = \operatorname {s o f t m a x} \left(Q ^ {i} K ^ {T} + M\right) V + Q ^ {i} \tag {4} +$$ + +Further, the context feature $\mathrm{R}$ is utilized to retrieve the clip. It undergoes a linear projection layer, $l_{s}(\cdot)$ , to obtain context scores $S_{r} = \operatorname{sigmoid}(l_{s}(R))$ . We retrieve the most relevant clip based on the context score $S_{r}$ . Importantly, the context $\mathrm{R}$ represents the queried moment by attending to the reduced set of relevant frames. So, we fuse it with the frame features with Eq. 5 to produce stronger multimodal inputs for the decoder. It enhances the decoder's localization capability significantly. The decoder only processes features from the retrieved clip to predict the moment. + +$$ +P ^ {i, j} = F _ {c} ^ {i, j} + R ^ {i} \times G ^ {j}, i \in P, \forall j \tag {5} +$$ + +# 3.3 Grounding Decoder + +We process the retrieved clip features with a transformer decoder to predict precise moment boundaries. Previous methods [16,22] feed both modalities into the decoder, introducing multiple positional information (e.g., time in the video and word order in the text) that complicates the detection task. In contrast, our encoder injects text information directly into the clip, eliminating the requirement to feed both modalities into the decoder. We incorporate learnable anchor queries [27] to represent the moment's center and width as $(\tau_c,\tau_w)$ . The decoder has 2 cross-attention and 2 self-attention layers. + +# 3.4 Loss Functions + +Intra-Clip Attention Loss guides our sparsifier to distinguish frames within and outside of the ground truth moment. This enables the retrieval attention to focus on relevant video regions and achieve fine-grained event understanding. + +Specifically, the in-moment frames are required to maintain higher relevancy scores than those outside the ground truth moment. The loss is defined as: + +$$ +\mathcal {L} _ {\mathrm {a t t n}} = \max (0, \Delta + S _ {c} (i, j _ {\mathrm {o u t}}) - S _ {c} (i, j _ {\mathrm {i n}})) \tag {6} +$$ + +In the equation, $S_{c}$ is the relevancy score, $\Delta$ represents the margin, and anchor $i$ is an in-moment frame. We randomly chose in-moment and out-of-moment frames $j_{\mathrm{in}}$ and $j_{\mathrm{out}}$ from the same clip. To calculate the relevancy score, the context features $R_{i}$ , and encoded clip $P^{i,j}$ undergoes linear layers, $l_{r}(\cdot)$ and $l_{c}(\cdot)$ , to obtain projected context $R_{proj}^{i} = l_{r}(R^{i}) \in R^{D_{l}}$ and content features, $P_{proj}^{i,j} = l_{c}(P^{i,j}) \in R^{L_{c} \times D_{l}}$ . Then, we calculate the frame-level relevancy score: + +$$ +S _ {c} (i, j) = R _ {p r o j} ^ {i} \cdot P _ {p r o j} ^ {i, j} \tag {7} +$$ + +Inter-Clip Contrastive loss. A long video contains many clips with similar environments and scenes. However, most of these clips do not contain the specific moment we are searching for. So, high-level video understanding is not enough to distinguish precise moments. During the test phase, we need to handle many such negative clips. To tackle this issue, unlike previous approaches [16], we sample a large number of negative clips and train our retriever on a large batch size. Hence, this negative clip sampling reduces the disparity between the train and test phases. + +For this section, we denote the context feature, $R^{i,j}$ for $i^{th}$ proposal and $j^{th}$ text query. We train our model contrastively with positive text-clip pairs $[R^{i,j}]_{i = j}$ and negative pairs $[R^{i,j}]_{i\neq j}$ . We use the InfoNCE loss [36] to identify the positive text-clip pair amongst a set of unrelated negative samples. Here, $l_{cont}(\cdot)$ is a linear projection layer that transforms the $D_{f}$ dimensional context feature into a one-dimensional logit. + +$$ +\mathcal {L} _ {\text {c o n t}} = - \sum_ {i} \log \frac {\exp \left(l _ {\text {c o n t}} \left(R ^ {i , i}\right)\right)}{\sum_ {j} \exp \left(l _ {\text {c o n t}} \left(R ^ {i , j}\right)\right)} \tag {8} +$$ + +Grounding loss. The objective functions for grounding, which aim to locate desired moments, are adopted from the baseline approach [22]. The grounding loss $\mathcal{L}_{\mathrm{g}}$ measures the discrepancy between the ground truth (GT) and predicted moments. A one-to-one correspondence between GT and predicted moments is established using the Hungarian algorithm. This loss includes both an $\mathcal{L}1$ loss and a generalized IoU loss ( $\mathcal{L}\mathrm{gIoU}$ ). Additionally, a cross-entropy loss $\mathcal{L}_{\mathrm{CE}}$ is employed to classify predicted moments as either foreground or background. Thus, $\mathcal{L}_{\mathrm{g}}$ is defined as follows: + +$$ +\mathcal {L} _ {\mathrm {g}} = \lambda_ {\mathcal {L} 1} | | \tau - \hat {\tau} | | + \lambda_ {\mathrm {g I o U}} \mathcal {L} _ {\mathrm {g I o U}} (\tau , \hat {\tau}) + \lambda_ {\mathrm {C E}} \mathcal {L} _ {\mathrm {C E}}, \tag {9} +$$ + +where $\tau$ and $\hat{\tau}$ represent the ground-truth moment and its corresponding prediction, containing center coordinates $\tau_{c}$ and width $\tau_{w}$ . The hyper-parameters $\lambda_{*}$ are used for balancing the losses. Finally, with the attention and contrastive loss, the total loss $\mathcal{L}_{\mathrm{total}}$ is defined as follows: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \lambda_ {\text {a t t n}} \mathcal {L} _ {\text {a t t n}} + \lambda_ {\text {c o n t}} \mathcal {L} _ {\text {c o n t}} + \mathcal {L} _ {\mathrm {g}} \tag {10} +$$ + +# 4 Experimental Setup + +# 4.1 Datasets and Evaluation Metric + +MAD [40] is an extensive dataset comprising 1.2K hours of full-length movies and 384K natural language queries, each associated with specific moments in the videos. The videos are orders of magnitude longer than previous datasets, with an average duration of 110 minutes, while the specified text moments are on average 4.1 seconds. This small moment-to-video ratio poses a significant challenge for the grounding task. + +Ego4D-NLQ [14] is a large-scale egocentric video data set with multiple challenges. Specifically, we use the episodic memory benchmark Ego4D-NLQ, which requires localizing where the answer to a natural language query can be seen. It contains around 13 template questions, amounting to around 74K queries. The train, val, and test sets contain 11.3K, 3.9K, and 4.0K queries. The video length ranges from 8 to 20 minutes, with an average of 8.25 minutes, and the average moment duration is 8.3 seconds. This means the moments constitute only $2\%$ of the input video on average. + +Grounding Metric: For grounding, we follow previous methods [16, 40], and adopt the standard metric $\mathrm{Recall}@\mathbf{k}$ at IoU = $\theta$ ( $\mathrm{Rk}_{\theta}$ ). This metric represents the percentage of testing samples with at least one grounding prediction whose intersection over union (IoU) with the ground truth (GT) is larger than $\theta$ among the top- $k$ predictions. + +Retrieval Metric: To assess our proposal retrieval stage independently of the grounding, we utilize the standard retrieval metric [13], Recall at Rank $k$ ( $\mathrm{R}@\mathbf{k}$ ). This metric calculates the percentage of GT moments present in the top- $k$ retrieved proposals. + +# 4.2 Implementation Details: + +We use pre-trained CLIP [38] and EgoVLP [24] models to extract video frames from MAD and Ego4D. Text features for both datasets are extracted using CLIP. The feature extractors are frozen, and their pre-trained weights remain unchanged during training. For Ego4D, we pre-train our model with NaQ annotations from previous work [39]. The proposal length, $W_{c}$ is set to 180s for MAD and 48s for Ego4D, with only the top-30 and top-5 proposals retrieved, respectively. Training on Ego4D and MAD is conducted on four Nvidia-RTX-A6000 GPUs while fine-tuning on Ego4D is performed on a single GPU. Considering the longer proposals in MAD, we utilize a batch size of 32 for Ego4D and 8 for MAD. The number of moment queries is set to 5. For loss computation, we set hyperparameters as $\lambda_{L1} = 10$ , $\lambda_{\mathrm{gIoU}} = 1$ , $\lambda_{\mathrm{CE}} = 4$ , $\lambda_{\mathrm{samp}} = 1$ , $\lambda_{\mathrm{cont}} = 10$ , and $\Delta = 0.2$ . We use Xavier initialization [12] and employ AdamW [30] with an initial learning rate of $1 \times 10^{-4}$ . RGNet is trained for 35 and 200 epochs on + +MAD and Ego4D datasets. The initial learning rate is reduced by one order of magnitude at epochs 25 for MAD and 120 for Ego4D. + +# 5 Results and Analysis + +We first compare our performance with previous state-of-the-art models. Then, we report detailed ablation studies of our proposed method and visualize the qualitative results compared to the disjoint baseline. + +
ModelNaQEgo4D-NLQ [14]MAD [40]Avg
R1.3R5.3R1.5R5.5AvgR1.1R5.1R1.3R5.3R1.5R5.5Avg
2D-TAN [55]X5.0412.892.025.886.463.2211.902.529.251.585.696.07
UniVTG [25]X11.747.543.257.887.60-------
VSLNet [53]✓10.2619.015.8112.6711.93-------
VLG-Net [41]X-----3.6411.662.769.311.655.995.84
M-DETR [22]X8.2323.235.0113.3712.463.6012.982.819.861.675.586.08
M-Guide [2]X-----9.3018.964.6513.062.167.409.26
SOONet [37]X8.0022.403.7611.0911.3111.2623.219.0019.645.3213.1413.59
H-Hands [52]X13.2023.307.9015.6015.00-------
CONE [16]X14.1530.338.1818.0217.678.9020.516.8716.114.109.5911.01
EgoVLP✓15.9026.389.4617.8017.38-------
ReLeR [39]✓19.3123.6211.5915.7517.57-------
Ours (Default)X18.2834.0212.0422.8921.8112.4325.129.4818.725.6110.8613.70
Ours✓20.6341.6712.4725.0824.9612.4325.129.4818.725.6110.8613.70
+ +Table 1: Main results on Ego4D-NLQ and MAD. RGNet achieves state-of-the-art performance on both datasets. Our default network is trained without NaQ annotations [39] on Ego4D. Even without NaQ annotations, our default network shows a larger improvement, underscoring the effectiveness of our solution with limited data. + +# 5.1 Result on the Ego4D-NLQ Dataset + +RGNet demonstrates state-of-the-art performance on all metrics for the Ego4D-NLQ benchmark. As shown in Table 1, RGNet improves its primary evaluation metric, the R1.3 and R5.3 score average by $9.7\%$ . Separately, we outperform the previous best methods on R1.3 and R5.3 by $1.32\%$ and $18.1\%$ , respectively. This improvement is consistent across other metrics as well. The unified clip and frame-level modeling in RGNet parallelly capture long- and short-range temporal dependencies, contributing to its superior performance. By modeling the entire video as a whole, RGNet learns better cross-modal alignment by contrasting visual scenarios from various events within the same video. Notably, compared to SooNet [37] and CONE [16], two prominent proposal-based methods, RGNet achieves an average score improvement of $13.7\%$ and $7.3\%$ . This substantial improvement is attributed to our unified modeling of clip retrieval, whereas they separately employ a retrieval similarity search between the video and the text. + +# 5.2 Result on the MAD Dataset + +We compare our model's performance on the MAD dataset in Tab. 1. RGNet achieves state-of-the-art performance on both the R1.1 and R5.1 scores by out- + +performing the previous best SooNet [37] by $1.2\%$ and $1.9\%$ . Together with its competitive performance on other metrics, RGNet demonstrates the importance of end-to-end modeling of long videos. Specifically, our performance of R1 is superior at all IoU thresholds, which testifies to the effectiveness of the proposed retrieval method. Notably, prior models rely on heuristic dot product similarity search applied to CLIP [38] features for the retrieval phase. In contrast, our shared retrieval feature and its unified optimization with the grounding objective drastically enhance the first stage retrieval, as seen in Tab. 2. It further helps the network capture more discriminative events crucial for detecting extremely tiny moments in hour-long videos. + +# 5.3 Disjoint vs. Unified Architecture + +Since the long video temporal grounding (LVTG) performance relies on the first stage of clip retrieval, it does not perfectly reflect the actual moment localization capability of its grounding network. To assess the standalone grounding performance, we designed an oracle experiment operating exclusively on the clip where the ground truth moment is present. Compared to the oracle grounding, the LVTG $R1_{3}$ drops by $15.7\%$ and $23.9\%$ for Ego4D and MAD in Tab. 2. The poor grounding in LVTG emerges from the suboptimal retrieval accuracy of the disjoint network. Hence, retrieving clips where the ground truth moment exists is crucial, independent of the subsequent grounding accuracy. Regardless of the effectiveness of moment boundary detection, if the clip containing the moment boundary cannot be accurately retrieved. Our unified approach consis + +
DataModelClipRetrievalOracleGroundingLVTG
R@1R@5R1.3R5.3R1.3R5.3
Ego4dBaseline31.7164.6329.8454.0114.1530.33
Ours42.0876.2836.5363.6420.6341.67
MADBaseline12.4124.5029.4953.026.8716.11
Ours25.0150.0233.4263.439.4818.72
+ +Table 2: Empirical impact of two stages. To assess standalone grounding capability, we run an oracle experiment on the clip where the ground truth moment is present. The grounding capability degrades in LVTG evaluation because of incorrect selection by the disjoint clip retrieval network. Our unified model improves retrieval significantly, which leads to more effective temporal grounding in long videos. + +tently improves all retrieval metrics compared to the disjoint baseline. Notably, we achieve a staggering $10.4\%$ and $12.6\%$ improved R@1 score for Ego4D and MAD. This improvement in clip retrieval leads to state-of-the-art LVTG performance. Moreover, the grounding also improves because of the stronger clip features generated by our encoder. Compared to the disjoint baseline, the oracle R1.3 scores improved by about $6.7\%$ and $4.0\%$ for Ego4D and MAD, respectively. Hence, the unified end-to-end architecture proves more effective at LVTG by mutually improving both stages. + +# 5.4 Ablation Studies + +This section reports detailed ablation studies on the proposed modules and loss functions. We also ablate with various clip lengths and retrieved clip numbers. Unless stated, the experiments are done with the Default network w/o NaQ annotations [39] for the Ego4D-NLQ dataset and CONE [16] as the baseline. + +
ModelR1.3R5.3R1.5R5.5
RGNet (Default)18.2834.0212.0422.89
w/o Retrieval Token17.8033.9911.2522.32
w/o Sparsifier16.1231.579.9120.47
w/o RG-Encoder*14.1530.338.1818.02
w/o Contrastive Loss17.4132.1210.7922.80
w/o Attention Loss16.2131.599.9120.53
+ +Table 3: Cumulative ablation study for the proposed modules and losses of RGNet. The experiment is conducted w/o NaQ augmentation on the Ego4D dataset. In each step, we remove one of our proposed modules or losses to evaluate their individual impact. *denotes the baseline model. + +Modules: In our initial experiments, we validate the effectiveness of each proposed module with a cumulative ablation in Tab. 3. The unified network demonstrates superior performance across all evaluation metrics, notably achieving an R1.3 score of $18.28\%$ . Note that the disjoint baseline performs $4.2\%$ worse than our unified network. This performance difference can be attributed to our models unified architecture. + +With the removal of the retrieval token for generating the clip feature in Eq. 5, considerably drops the score by about $0.5\%$ . This proves that modeling both clip level context and frame level content are necessary to produce strong multimodal features for the grounding decoder. Subsequently, eliminating the sparsifier, there is a decline in $\mathrm{R1}_{.3}$ performance of approximately $1.7\%$ . This verifies the impact of the sparsifier in modeling the fine-grained event boundaries. Finally, removing the RG-Encoder yields a notable degradation of approximately $2.0\%$ in the $\mathrm{R1}_{.3}$ score, which substantiates its capability to model clip and frame level granularity jointly. + +Loss Functions: Training RGNet without our negative clip sampling strategy for the contrastive loss (Tab. 3) results in an almost $0.9\%$ decline in the R1.3 score. Our sampling strategy mimics long-video setup closely during training, which improves event discrimination capability inside the same scene and surroundings. + +By eliminating the attention loss, we observe a further reduction of $1.2\%$ in grounding performance. The attention loss enables our encoder to model fine-grained events in long videos which is crucial for specific moment detection. + +Number of Clips: The number of retrieved proposal clips significantly influences the Long Video Temporal Grounding (LVTG) performance. Our model consistently outperforms the baseline across various choices of top-k retrieved clips, as shown in Fig. 4. Even with only a single retrieved clip, our performance surpasses the baseline's optimal result for both datasets. Moreover, our model exhibits a less steep performance decline when fewer clips are retrieved. Importantly, with fewer clips, the grounding network runs fewer times, leading to an overall speedup. Therefore, our model strikes a favorable balance between efficiency and performance, showcasing superior performance than the baseline. + +![](images/7ee740ac8a82dad00098dd757937a2858c2e768cc2036277845a1f02225463e9.jpg) +Fig. 4: Impact of number of retrieved clips. Reducing the number of clips speeds up the network execution time. While the baseline model experiences a significant drop in performance with this reduction, RGNet shows a noticeably smaller decline in performance under the same conditions. + +![](images/2cea0e1ac86b544dfe9d3eb325bd2ab122c5ec943838025be54eaa3beb39a6b3.jpg) + +**Clip Length:** Varying clip length impacts both retrieval and grounding performance. The number of clip decreases with longer clips, making the retrieval task easier. Fig. 5 reports a monotonically improving retrieval with longer clips. However, after a certain length, moment localization quality starts to deteriorate. With longer clips, the task becomes extremely difficult. We see the decline after the 180s and 48s for the MAD and Ego4D datasets. Hence, we set the default clip length of our method to these optimal lengths. Importantly, the unified network helps outscore the disjoint baseline on all tested clip lengths (inversely proportional to the clip numbers in Fig. 4). + +![](images/42237315c5214986a41bddd5bd1ff81f7a59e11cb15f98fb772901c7fcf5400a.jpg) +Fig. 5: Impact of retrieved clip length. Longer clips result in improved retrieval due to fewer candidates. However, grounding becomes exceedingly difficult with longer clips. For example, the grounding performance drops after 180 seconds and 48 seconds for the MAD and Ego4D datasets. + +![](images/75b34f5a975d9f83f13448b05d8dfaee9e477ef4dffb730d8cf7bbdf8f7cd4e0.jpg) + +# 5.5 Qualitative Analysis + +We visualize the qualitative predictions of RGNet compared to the disjoint baseline in Fig. 6. For the first and second queries, the baseline struggles to retrieve the correct clip where the scissor appears. RGNet accurately retrieves the clip with the searched object and precisely localizes the moment it appeared in the video. Disjoint retrieval often leads to incorrect clip selection, which the final grounding network cannot recover from. For the third query, the baseline selects the correct clip but fails to localize the precise moment the man closes the car door, indicating a gap between the two stages. RGNet mitigates this gap by leveraging features from the retrieved clip to improve moment localization. More visual outputs are included the supplementary materials. + +![](images/ea0a59213b6660ed4bc74e839ceb6b949aa4d53aa9c71487292fb574909eacaf.jpg) +Query 1: Where is the scissor? + +![](images/8f75fc2e4626d4706db20f6a1f70073f6b6ab6ecf466b45449ae5b0f07307483.jpg) +Query 2: How many stainless bowls did I pick? + +![](images/907b19be12350262c7b4fbc06206b8afdb4a9d8380bccd71e2f4c68195f155c5.jpg) +Query 3: Did I leave the car door open? +Fig. 6: Qualitative Results. The baseline fails to retrieve the correct clip in the first two queries. Since they primarily depict the same indoor room throughout the whole video, precise event discrimination is vital for accurate clip retrieval. In the third query, the baseline cannot identify the moment within the correctly retrieved clip, detecting it only after it has concluded. Precise event localization demands improved alignment between visual events and the queried text. RGNet correctly localizes all the moments. + +# 6 Conclusion + +We introduce an end-to-end model for long video temporal grounding, unifying the two stages of prevailing methods with shared features and mutual optimization. We conduct independent analyses of the two stages, shaping our solution based on their distinct impacts. This leads to a better understanding of fine-grained events in long videos. Our approach demonstrates state-of-the-art performance on challenging long video grounding datasets, validating its effectiveness. Like all LVTG methods, we rely on a pre-trained image encoder to extract the visual features. A promising future direction includes eliminating the need for an image encoder and training directly on the raw video data. + +# References + +1. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: ICCV. pp. 1728-1738 (2021) +2. Barrios, W., Soldan, M., Ceballos-Arroyo, A.M., Heilbron, F.C., Ghanem, B.: Localizing moments in long video via multimodal guidance. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13667-13678 (2023) +3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020) +4. Chen, S., Zhao, Y., Jin, Q., Wu, Q.: Fine-grained video-text retrieval with hierarchical graph reasoning. In: CVPR. pp. 10638-10647 (2020) +5. Chen, Z., Jiang, X., Xu, X., Cao, Z., Mo, Y., Shen, H.T.: Joint searching and grounding: Multi-granularity video content retrieval. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 975-983 (2023) +6. Cheng, F., Wang, X., Lei, J., Crandall, D., Bansal, M., Bertasius, G.: Vindlu: A recipe for effective video-and-language pretraining. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023) +7. Dong, J., Li, X., Xu, C., Yang, X., Yang, G., Wang, X., Wang, M.: Dual encoding for video retrieval by text. IEEE TPAMI pp. 4065-4080 (2021) +8. Fang, B., Liu, C., Zhou, Y., Yang, M., Song, Y., Li, F., Wang, W., Ji, X., Ouyang, W., et al.: Uatvr: Uncertainty-adaptive text-video retrieval. arXiv preprint arXiv:2301.06309 (2023) +9. Gabeur, V., Sun, C., Alahari, K., Schmid, C.: Multi-modal transformer for video retrieval. In: ECCV. pp. 214-229 (2020) +0. Gao, Z., Liu, J., Sun, W., Chen, S., Chang, D., Zhao, L.: Clip2tv: Align, match and distill for video-text retrieval. arXiv preprint arXiv:2111.05610 (2021) +1. Ge, Y., Ge, Y., Liu, X., Li, D., Shan, Y., Qie, X., Luo, P.: Bridging video-text retrieval with multiple choice questions. In: CVPR. pp. 16167-16176 (2022) +2. Glorot, X., Bengio, Y.: Understanding the difficulty of training deep feedforward neural networks. In: Proceedings of the thirteenth international conference on artificial intelligence and statistics. pp. 249-256. JMLR Workshop and Conference Proceedings (2010) +3. Gorti, S.K., Vouitsis, N., Ma, J., Golestan, K., Volkovs, M., Garg, A., Yu, G.: X-pool: Cross-modal language-video attention for text-video retrieval. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5006-5015 (2022) +4. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: CVPR. pp. 18995-19012 (2022) +5. Hannan, T., Koner, R., Bernhard, M., Shit, S., Menze, B., Tresp, V., Schubert, M., Seidl, T.: Gratt-vis: Gated residual attention for auto rectifying video instance segmentation. arXiv preprint arXiv:2305.17096 (2023) +6. Hou, Z., Zhong, W., Ji, L., Gao, D., Yan, K., Chan, W.K., Ngo, C.W., Shou, Z., Duan, N.: Cone: An efficient coarse-to-fine alignment framework for long video temporal grounding. arXiv preprint arXiv:2209.10918 (2022) +7. Jang, E., Gu, S., Poole, B.: Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144 (2016) +8. Jang, J., Park, J., Kim, J., Kwon, H., Sohn, K.: Knowing where to focus: Event-aware transformer for video grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13846-13856 (2023) + +19. Jiang, H., Zhang, J., Huang, R., Ge, C., Ni, Z., Lu, J., Zhou, J., Song, S., Huang, G.: Cross-modal adapter for text-video retrieval. arXiv preprint arXiv:2211.09623 (2022) +20. Jin, P., Huang, J., Liu, F., Wu, X., Ge, S., Song, G., Clifton, D.A., Chen, J.: Expectation-maximization contrastive learning for compact video-and-language representations. NeurIPS (2022) +21. Kiros, R., Salakhutdinov, R., Zemel, R.S.: Unifying visual-semantic embeddings with multimodal neural language models. arXiv preprint arXiv:1411.2539 (2014) +22. Lei, J., Berg, T.L., Bansal, M.: Detecting moments and highlights in videos via natural language queries. Advances in Neural Information Processing Systems 34, 11846-11858 (2021) +23. Lei, J., Li, L., Zhou, L., Gan, Z., Berg, T.L., Bansal, M., Liu, J.: Less is more: Clipbert for video-and-language learning via sparse sampling. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7331-7341 (2021) +24. Lin, K.Q., Wang, A.J., Soldan, M., Wray, M., Yan, R., Xu, E.Z., Gao, D., Tu, R., Zhao, W., Kong, W., et al.: Egocentric video-language pretraining. arXiv preprint arXiv:2206.01670 (2022) +25. Lin, K.Q., Zhang, P., Chen, J., Pramanick, S., Gao, D., Wang, A.J., Yan, R., Shou, M.Z.: Univtg: Towards unified video-language temporal grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2794-2804 (2023) +26. Liu, N., Wang, X., Li, X., Yang, Y., Zhuang, Y.: Refer@ zju-alibaba submission to the ego4d natural language queries challenge 2022. arXiv preprint arXiv:2207.00383 (2022) +27. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329 (2022) +28. Liu, Y., Li, S., Wu, Y., Chen, C.W., Shan, Y., Qie, X.: Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3042-3051 (2022) +29. Liu, Y., Xiong, P., Xu, L., Cao, S., Jin, Q.: Ts2-net: Token shift and selection transformer for text-video retrieval. In: ECCV. pp. 319-335 (2022) +30. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017) +31. Luo, H., Ji, L., Shi, B., Huang, H., Duan, N., Li, T., Li, J., Bharti, T., Zhou, M.: Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353 (2020) +32. Ma, Y., Xu, G., Sun, X., Yan, M., Zhang, J., Ji, R.: X-clip: End-to-end multigrained contrastive learning for video-text retrieval. In: ACM MM. pp. 638-647 (2022) +33. Maddison, C.J., Mnih, A., Teh, Y.W.: The concrete distribution: A continuous relaxation of discrete random variables. CoRR abs/1611.00712 (2016), http://arxiv.org/abs/1611.00712 +34. Meng, D., Chen, X., Fan, Z., Zeng, G., Li, H., Yuan, Y., Sun, L., Wang, J.: Conditional detr for fast training convergence. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3651-3660 (2021) +35. Moon, W., Hyun, S., Park, S., Park, D., Heo, J.P.: Query-dependent video representation for moment retrieval and highlight detection. In: Proceedings of the + +IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23023-23033 (2023) +36. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018) +37. Pan, Y., He, X., Gong, B., Lv, Y., Shen, Y., Peng, Y., Zhao, D.: Scanning only once: An end-to-end framework for fast temporal grounding in long videos. arXiv preprint arXiv:2303.08345 (2023) +38. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. pp. 8748-8763 (2021) +39. Ramakrishnan, S.K., Al-Halah, Z., Grauman, K.: Naq: Leveraging narrations as queries to supervise episodic memory. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6694-6703 (2023) +40. Soldan, M., Pardo, A., Alcázar, J.L., Caba, F., Zhao, C., Giancola, S., Ghanem, B.: Mad: A scalable dataset for language grounding in videos from movie audio descriptions. In: CVPR. pp. 5026-5035 (2022) +41. Soldan, M., Xu, M., Qu, S., Tegner, J., Ghanem, B.: Vlg-net: Video-language graph matching network for video grounding. In: ICCV. pp. 3224-3234 (2021) +42. Wang, A.J., Ge, Y., Yan, R., Ge, Y., Lin, X., Cai, G., Wu, J., Shan, Y., Qie, X., Shou, M.Z.: All in one: Exploring unified video-language pre-training. arXiv preprint arXiv:2203.07303 (2022) +43. Wang, Q., Zhang, Y., Zheng, Y., Pan, P., Hua, X.S.: Disentangled representation learning for text-video retrieval. arXiv preprint arXiv:2203.07111 (2022) +44. Wang, Z., Sung, Y.L., Cheng, F., Bertasius, G., Bansal, M.: Unified coarse-to-fine alignment for video-text retrieval. In: The IEEE International Conference on Computer Vision (ICCV) (October 2023) +45. Xu, H., Ghosh, G., Huang, P.Y., Arora, P., Aminzadeh, M., Feichtenhofer, C., Metze, F., Zettlemoyer, L.: Vlm: Task-agnostic video-language model pre-training for video understanding. ACLliu2021hit (2021) +46. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettlemoyer, L., Feichtenhofer, C.: Videoclip: Contrastive pre-training for zero-shot video-text understanding. In: EMNLP. pp. 6787-6800 (2021) +47. Xue, H., Hang, T., Zeng, Y., Sun, Y., Liu, B., Yang, H., Fu, J., Guo, B.: Advancing high-resolution video-language representation with large-scale video transcriptions. In: CVPR. pp. 5036-5045 (2022) +48. Xue, H., Sun, Y., Liu, B., Fu, J., Song, R., Li, H., Luo, J.: Clip-vip: Adapting pre-trained image-text model to video-language representation alignment. ICLR (2023) +49. Yu, Y., Kim, J., Kim, G.: A joint sequence fusion model for video question answering and retrieval. In: ECCV. pp. 471-487 (2018) +50. Yu, Y., Ko, H., Choi, J., Kim, G.: End-to-end concept word detection for video captioning, retrieval, and question answering. In: CVPR. pp. 3165-3173 (2017) +51. Zhang, B., Jin, X., Gong, W., Xu, K., Zhang, Z., Wang, P., Shen, X., Feng, J.: Multimodal video adapter for parameter efficient video text retrieval. arXiv preprint arXiv:2301.07868 (2023) +52. Zhang, C., Gupta, A., Zisserman, A.: Helping hands: An object-aware ego-centric video recognition model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13901-13912 (2023) +53. Zhang, H., Sun, A., Jing, W., Zhou, J.T.: Span-based localizing network for natural language video localization. arXiv preprint arXiv:2004.13931 (2020) + +54. Zhang, H., Sun, A., Jing, W., Zhou, J.T.: The elements of temporal sentence grounding in videos: A survey and future directions. arXiv preprint arXiv:2201.08071 1(2) (2022) +55. Zhang, S., Peng, H., Fu, J., Luo, J.: Learning 2d temporal adjacent networks for moment localization with natural language. In: AAAI. pp. 12870-12877 (2020) +56. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020) \ No newline at end of file diff --git a/2024/RGNet_ A Unified Clip Retrieval and Grounding Network for Long Videos/images.zip b/2024/RGNet_ A Unified Clip Retrieval and Grounding Network for Long Videos/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..87e11f4d2bb2674ff7cbce176d7e022578f59e1d --- /dev/null +++ b/2024/RGNet_ A Unified Clip Retrieval and Grounding Network for Long Videos/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:52a6eeec49ceec5d649d731de9f3ccf1e5fab439047eee58477a29e1c6c05791 +size 373650 diff --git a/2024/RGNet_ A Unified Clip Retrieval and Grounding Network for Long Videos/layout.json b/2024/RGNet_ A Unified Clip Retrieval and Grounding Network for Long Videos/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..41c5e8155890d4487b78e0c5f3a78c53398bb8c8 --- /dev/null +++ b/2024/RGNet_ A Unified Clip Retrieval and Grounding Network for Long Videos/layout.json @@ -0,0 +1,10125 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "type": "text", + "content": "RGNet: A Unified Clip Retrieval and Grounding Network for Long Videos" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 143, + 168, + 471, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 168, + 471, + 194 + ], + "spans": [ + { + "bbox": [ + 143, + 168, + 471, + 194 + ], + "type": "text", + "content": "Tanveer Hannan" + }, + { + "bbox": [ + 143, + 168, + 471, + 194 + ], + "type": "inline_equation", + "content": "^{1,2\\star}" + }, + { + "bbox": [ + 143, + 168, + 471, + 194 + ], + "type": "text", + "content": " Md Mohaiminul Islam" + }, + { + "bbox": [ + 143, + 168, + 471, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 143, + 168, + 471, + 194 + ], + "type": "text", + "content": " Thomas Seidl" + }, + { + "bbox": [ + 143, + 168, + 471, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 143, + 168, + 471, + 194 + ], + "type": "text", + "content": " Gedas Bertasius" + }, + { + "bbox": [ + 143, + 168, + 471, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 273, + 201, + 339, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 273, + 201, + 339, + 212 + ], + "spans": [ + { + "bbox": [ + 273, + 201, + 339, + 212 + ], + "type": "text", + "content": "1 LMU Munich" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 286, + 213, + 327, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 213, + 327, + 223 + ], + "spans": [ + { + "bbox": [ + 286, + 213, + 327, + 223 + ], + "type": "text", + "content": "2 MCML" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 266, + 224, + 347, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 224, + 347, + 235 + ], + "spans": [ + { + "bbox": [ + 266, + 224, + 347, + 235 + ], + "type": "text", + "content": "3 UNC Chapel Hill" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 268, + 455, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 268, + 455, + 477 + ], + "spans": [ + { + "bbox": [ + 160, + 268, + 455, + 477 + ], + "type": "text", + "content": "Abstract. Locating specific moments within long videos (20-120 minutes) presents a significant challenge, akin to finding a needle in a haystack. Adapting existing short video (5-30 seconds) grounding methods to this problem yields poor performance. Since most real-life videos, such as those on YouTube and AR/VR, are lengthy, addressing this issue is crucial. Existing methods typically operate in two stages: clip retrieval and grounding. However, this disjoint process limits the retrieval module's fine-grained event understanding, crucial for specific moment detection. We propose RGNet which deeply integrates clip retrieval and grounding into a single network capable of processing long videos into multiple granular levels, e.g., clips and frames. Its core component is a novel transformer encoder, RG-Encoder, that unifies the two stages through shared features and mutual optimization. The encoder incorporates a sparse attention mechanism and an attention loss to model both granularity jointly. Moreover, we introduce a contrastive clip sampling technique to mimic the long video paradigm closely during training. RGNet surpasses prior methods, showcasing state-of-the-art performance on long video temporal grounding (LVTG) datasets MAD and Ego4D. The code is released at https://github.com/Tanveer81/RGNet." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 487, + 445, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 487, + 445, + 499 + ], + "spans": [ + { + "bbox": [ + 160, + 487, + 445, + 499 + ], + "type": "text", + "content": "Keywords: Long Video Temporal Grounding " + }, + { + "bbox": [ + 160, + 487, + 445, + 499 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 487, + 445, + 499 + ], + "type": "text", + "content": " Moment Localization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 521, + 230, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 230, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 230, + 534 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 548, + 482, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 482, + 646 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 482, + 646 + ], + "type": "text", + "content": "The exponential rise in online videos has created a demand for effective content retrieval systems. The retrieval results of user-defined queries, particularly in long videos (20-120 minutes), often require locating specific moments. Most existing solutions [5,41,53,55], tailored for short videos (5-30 seconds), struggle when applied to hour-long videos, slowing down the retrieval and impacting the quality of the results [40, 54]. This challenge emerges because it is very difficult to locate short moments in a long video based on text queries, a task known as Long Video Temporal Grounding (LVTG) [14, 40]." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 331, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 331, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 331, + 665 + ], + "type": "text", + "content": "* Corresponding author: hannan@dbis.ifi.lmu.de" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 176, + 113, + 440, + 257 + ], + "blocks": [ + { + "bbox": [ + 176, + 113, + 440, + 257 + ], + "lines": [ + { + "bbox": [ + 176, + 113, + 440, + 257 + ], + "spans": [ + { + "bbox": [ + 176, + 113, + 440, + 257 + ], + "type": "image", + "image_path": "a4aa5a3c60a7093dea15bbe7642f13a4f87b7a9a0be3ffc447826395f03b1919.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 264, + 482, + 331 + ], + "lines": [ + { + "bbox": [ + 130, + 264, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 482, + 331 + ], + "type": "text", + "content": "Fig. 1: Overview of RGNet. It predicts the moment boundary specified by textual queries from an hour-long video. First, our proposed RG-Encoder maps the video and text features to a joint space and retrieves the relevant clip feature. The subsequent grounding decoder processes the retrieved features to predict the beginning and end times of the moment. The encoder parallelly operates at multiple levels of granularity (e.g., clip and frame) to achieve an end-to-end-solution." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 338, + 482, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 386 + ], + "type": "text", + "content": "A straightforward solution [16,37] for the LVTG task is to divide the video into shorter clips, retrieve the most relevant one, and apply a grounding network to predict the moment. However, the grounding network cannot rectify failure in the clip retrieval phase." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 387, + 481, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 481, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 481, + 411 + ], + "type": "text", + "content": "Empirically, we evaluate the impact of the two stages in Sec. 5.3 and identify clip retrieval as the primary factor for poor performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 412, + 482, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 482, + 519 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 482, + 519 + ], + "type": "text", + "content": "The existing methods have a separate module for clip retrieval, which is disjoint from their grounding network (see Fig. 2). They typically follow a text-video retrieval [13,23,29] technique to select the relevant clip. However, video retrieval only requires a high-level understanding of video topics. For example, a typical video retrieval query is \"A movie about a family surviving in a farmhouse\". In contrast, an example clip retrieval query in Fig. 1 is \"Find the moment when the mum hangs up laundry outside the farmhouse\". These specific moments from a long video require fine-grained event understanding. Thus, video retrieval models are suboptimal for these fine-grained moment localization tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 640 + ], + "type": "text", + "content": "We attribute this disjoint retrieval to the poor performance of these models. However, unifying clip selection and grounding is challenging due to their distinct setups. The former is a retrieval task, whereas the latter is a regression task. Moreover, it requires modeling two levels of video granularity: clip and frame. To address these technical challenges, we propose RGNet, a unified clip Retrieval and Grounding Network (see Fig. 1). The single network enables end-to-end training of both stages. This improves the retrieval module's fine-grained event understanding by directly optimizing it with the moment annotations. Parallely, the grounding network also benefits from the strong multimodal features produced by our encoder, further enhancing the localization capability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "To achieve this unification, we propose a novel transformer encoder, RGEncoder, which enables the grounding network to perform clip retrieval. This" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "Hannan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 175, + 114, + 339, + 235 + ], + "blocks": [ + { + "bbox": [ + 175, + 114, + 339, + 235 + ], + "lines": [ + { + "bbox": [ + 175, + 114, + 339, + 235 + ], + "spans": [ + { + "bbox": [ + 175, + 114, + 339, + 235 + ], + "type": "image", + "image_path": "7ee1042d6ed0acb37be653a312ba6baf8af68724ee30a15c67925808283b8e4a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 239, + 299, + 247 + ], + "lines": [ + { + "bbox": [ + 230, + 239, + 299, + 247 + ], + "spans": [ + { + "bbox": [ + 230, + 239, + 299, + 247 + ], + "type": "text", + "content": "Existing Disjoint Network" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 356, + 116, + 441, + 234 + ], + "blocks": [ + { + "bbox": [ + 356, + 116, + 441, + 234 + ], + "lines": [ + { + "bbox": [ + 356, + 116, + 441, + 234 + ], + "spans": [ + { + "bbox": [ + 356, + 116, + 441, + 234 + ], + "type": "image", + "image_path": "3d4b369cc3ea41360acbb01f69fc4e95f9ed760594bfee40902fac5a3eeaa769.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 239, + 437, + 247 + ], + "lines": [ + { + "bbox": [ + 367, + 239, + 437, + 247 + ], + "spans": [ + { + "bbox": [ + 367, + 239, + 437, + 247 + ], + "type": "text", + "content": "Proposed Unified RGNet" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 260, + 479, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 260, + 479, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 479, + 304 + ], + "type": "text", + "content": "Fig. 2: Unified Solution. (left) Existing methods involve a separate retrieval and grounding network. The disjoint retrieval lacks fine-grained event understanding, which is crucial for moment localization. (right) Our unified network architecture overcomes it by deeply integrating the retrieval module with the grounding objective." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 308, + 480, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 480, + 463 + ], + "type": "text", + "content": "eliminates the suboptimal video retrieval network and effectively models long video clip retrieval. To enable the retrieval module to understand fine-grained events, we design a sparse-attention mechanism in the encoder. The sparse attention enables the retrieval module to focus on specific events in the video, which is more synergistic with the grounding task. Thus RG-Encoder is capable of operating both at clip and frame level granularity. We propose an Intra-Clip Attention Loss, which motivates the sparse attention to focus more on the video frames aligned with the specified event. Furthermore, we propose a negative clip mining technique to simulate clip retrieval from long videos during training. This negative mining enables our network to train on a large batch size with an inter-clip contrastive loss. The large number of negative clips in the batch closely mimics the long video paradigm and reduces the gap between the training and test phases." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 464, + 479, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 479, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 479, + 498 + ], + "type": "text", + "content": "Together, these components of RGNet bridge the gap between the two stages of LVTG, enhancing the fine-grained event understanding in hour-long videos. In summary, our contributions are fourfold:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 136, + 499, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 136, + 499, + 479, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 499, + 479, + 534 + ], + "spans": [ + { + "bbox": [ + 136, + 499, + 479, + 534 + ], + "type": "text", + "content": "- We systematically deconstruct existing LVTG methods into clip retrieval and grounding stages. Through empirical evaluations, we discern that disjoint retrieval is the primary factor contributing to poor performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 535, + 480, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 535, + 480, + 582 + ], + "spans": [ + { + "bbox": [ + 138, + 535, + 480, + 582 + ], + "type": "text", + "content": "- Based on our observations, we introduce RGNet, which integrates clip retrieval with grounding through parallel clip and frame-level modeling. This obviates the necessity for a separate video retrieval network, replaced instead by an end-to-end clip retrieval module tailored specifically for long videos." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 582, + 480, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 582, + 480, + 629 + ], + "spans": [ + { + "bbox": [ + 138, + 582, + 480, + 629 + ], + "type": "text", + "content": "- We introduce sparse attention to the retriever and a corresponding loss to model fine-grained event understanding in long-range video. We propose a contrastive negative clip-mining strategy to simulate clip retrieval from a long video during training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 630, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 630, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 630, + 480, + 665 + ], + "type": "text", + "content": "- RGNet achieves state-of-the-art performance across both LVTG datasets, Ego4D [14] and MAD [40]. For instance, RGNet outperforms the previous best Ego4D method [39] by a substantial margin (9.7%)." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "text", + "content": "RGNet" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 264, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 264, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 264, + 127 + ], + "type": "text", + "content": "2 Related Literature" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 148, + 482, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 148, + 482, + 292 + ], + "spans": [ + { + "bbox": [ + 130, + 148, + 482, + 292 + ], + "type": "text", + "content": "Short Video Temporal Grounding. The recent grounding methods mainly focus on short videos. The best-performing ones utilize transformer variants [3,27,34,56]. For example, Moment-DETR [22] adopted transformers for combined video and text features. Subsequent works, such as UMT [28] and QD-DETR [35], split the cross-modal and unimodal modeling of video and text features for enhanced performance. EATR [18] improves its decoder with videospecific moment queries. These models were initially designed for shorter videos. However, when applied to hour-long videos, the moments become extremely tiny, posing a challenge akin to finding a needle in a haystack. Hence, these methods perform poorly on long videos, as noted by the authors of the MAD dataset [40]. In contrast, we effectively address the task by simultaneously processing long videos at two granular levels (video clips and frames) within a single network." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 304, + 482, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 473 + ], + "type": "text", + "content": "Long Video Temporal Grounding. Recently, video temporal grounding has been adapted for long videos with MAD [40] and Ego4D [14] datasets. Typically, the methods designed for long videos involve two stages. Proposal-free methods [2, 26, 41, 53, 55] segment lengthy videos into smaller parts to predict candidate moments and then rank them to obtain the final predictions. Proposal-based methods [16, 37] generate proposal clips or anchors and subsequently apply their grounding model to the retrieved proposals. Some methods, like CONE [16], and M-Guidance [2], utilize the detection transformer [3] as the grounding network to improve the grounding phase. These proposal-based models are more efficient and perform better than their counterparts. However, they need a separate retrieval module to select the proposal clips. This disjoint two-stage architecture is what we find to be the main reason for the poor performance of these models. In contrast, we propose a novel transformer encoder that solves both clip retrieval and grounding in a unified way, enabling end-to-end optimization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "Video-text Retrieval. The early retrieval methods [4, 7, 21, 49, 50] designed multimodal fusion techniques to align pre-trained video and text models. To bridge the gap between pre-training and downstream tasks, large-scale pretrained video-language representations [1, 6, 9, 11, 31, 42, 45-47] have been introduced in various works. Several studies [8, 10, 19, 20, 32, 43, 44, 48, 51] have transferred the knowledge from CLIP [38] to text-video retrieval tasks. In recent studies, various sampling strategies [13, 23, 29] have been explored, enabling the model to selectively concentrate on pertinent video frames based on provided text inputs. For instance, Clip-BERT [23] introduces a sparse sampling strategy, X-Pool [13] utilizes a cross-modal attention model, and TS2-Net [29] adapts CLIP to capture temporal information and eliminate unimportant tokens. While existing methods aim to retrieve high-level video topics from a pool of video collections, moment localization demands a fine grained understanding to identify specific events within a video. Despite the similar setup to video retrieval, retrieving the corresponding clip from a long video necessitates deeper" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "text", + "content": "Hannan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "event comprehension. Therefore, we tailored our encoder to enhance event understanding and simulated precise clip retrieval conditions during training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 159, + 197, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 159, + 197, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 159, + 197, + 172 + ], + "type": "text", + "content": "3 RGNet" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 186, + 483, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 186, + 483, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 186, + 483, + 293 + ], + "type": "text", + "content": "This section presents a detailed description of our proposed unified LVTG method. As illustrated in Fig. 1, RGNet takes a long video and a text query as input and predicts the precise moment semantically correlated with the query in an end-to-end manner. The whole network is divided into the proposed RG-Encoder (Fig. 3) and a decoder (Sec. 3.3). The encoder retrieves the most relevant clip feature from the input long video. The decoder then processes the retrieved clip feature to predict the precise moment boundary. Additionally, an Intra-Clip Attention Loss and an Inter-Clip Contrastive Loss are incorporated to facilitate low-level event understanding in long-range videos." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 219, + 314, + 392, + 467 + ], + "blocks": [ + { + "bbox": [ + 219, + 314, + 392, + 467 + ], + "lines": [ + { + "bbox": [ + 219, + 314, + 392, + 467 + ], + "spans": [ + { + "bbox": [ + 219, + 314, + 392, + 467 + ], + "type": "image", + "image_path": "172a7e7757a46c7be9bbd3ae7f46b1f5136b4f36025c4e25396febda8144b69c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 479, + 482, + 544 + ], + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 544 + ], + "type": "text", + "content": "Fig. 3: Overview of RG-Encoder. It takes video clips and textual query as input and retrieves the relevant clip features. First, a cross-attention fuses the clips with text, and the sparsifier masks the out-of-moment frames. Based on the mask, the retrieval attention focuses on in-moment frames (colored red) and generates clip-level context and frame-level content features. We combine the context and content to generate the retrieved clip feature." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 561, + 256, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 561, + 256, + 571 + ], + "spans": [ + { + "bbox": [ + 132, + 561, + 256, + 571 + ], + "type": "text", + "content": "3.1 Feature Extraction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Let us assume we have an untrimmed long video " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " consisting of " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " frames and a query sentence " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " words corresponding to a target moment's center " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tau_{c}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tau_{w}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ". Following previous methods [16,37], we use pre-trained frozen models to extract visual features " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "V = \\{f^{1}, f^{2}, \\dots, f^{T}\\} \\in R^{T \\times D_{f}}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " as well as textual features " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "S = \\{w^{1}, w^{2}, \\dots, w^{N}\\} \\in R^{N \\times D_{w}}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "D_{f}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "D_{w}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " represent the feature dimensions of video frames and query words. We slice the long video " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " into clips of length " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "L_{c}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " and feed them into our model. We employ a sliding window of length" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "text", + "content": "RGNet" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 482, + 152 + ], + "type": "inline_equation", + "content": "L_{c}" + }, + { + "bbox": [ + 132, + 115, + 482, + 152 + ], + "type": "text", + "content": " and stride " + }, + { + "bbox": [ + 132, + 115, + 482, + 152 + ], + "type": "inline_equation", + "content": "L_{c} / 2" + }, + { + "bbox": [ + 132, + 115, + 482, + 152 + ], + "type": "text", + "content": " to obtain a collection of " + }, + { + "bbox": [ + 132, + 115, + 482, + 152 + ], + "type": "inline_equation", + "content": "T_{c}" + }, + { + "bbox": [ + 132, + 115, + 482, + 152 + ], + "type": "text", + "content": " clips " + }, + { + "bbox": [ + 132, + 115, + 482, + 152 + ], + "type": "inline_equation", + "content": "C = \\{C^{1}, C^{2}, \\dots, C^{T_{c}}\\} \\in R^{T_{c} \\times L_{c} \\times D_{f}}" + }, + { + "bbox": [ + 132, + 115, + 482, + 152 + ], + "type": "text", + "content": ". Here, the " + }, + { + "bbox": [ + 132, + 115, + 482, + 152 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 132, + 115, + 482, + 152 + ], + "type": "text", + "content": " clip, " + }, + { + "bbox": [ + 132, + 115, + 482, + 152 + ], + "type": "inline_equation", + "content": "C^{i} = \\{f^{s_{i} + 1}, f^{s_{i} + 2}, \\dots, f^{s_{i} + L_{c}}\\} \\in R^{L_{c} \\times D_{f}}" + }, + { + "bbox": [ + 132, + 115, + 482, + 152 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 132, + 115, + 482, + 152 + ], + "type": "inline_equation", + "content": "s_{i}" + }, + { + "bbox": [ + 132, + 115, + 482, + 152 + ], + "type": "text", + "content": " is the start index of the clip." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 167, + 224, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 167, + 224, + 178 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 224, + 178 + ], + "type": "text", + "content": "3.2 RG-Encoder" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 184, + 482, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 184, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 184, + 482, + 293 + ], + "type": "text", + "content": "All the clips " + }, + { + "bbox": [ + 130, + 184, + 482, + 293 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 184, + 482, + 293 + ], + "type": "text", + "content": " from a single video comprise samples in a batch. The RG-Encoder (Fig. 3) processes them with the query " + }, + { + "bbox": [ + 130, + 184, + 482, + 293 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 184, + 482, + 293 + ], + "type": "text", + "content": " to retrieve the relevant clip feature " + }, + { + "bbox": [ + 130, + 184, + 482, + 293 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 184, + 482, + 293 + ], + "type": "text", + "content": ". First, the cross-attention generates query-conditioned frame features " + }, + { + "bbox": [ + 130, + 184, + 482, + 293 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 184, + 482, + 293 + ], + "type": "text", + "content": ". Then, based on the frame-query correlation calculated from " + }, + { + "bbox": [ + 130, + 184, + 482, + 293 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 184, + 482, + 293 + ], + "type": "text", + "content": ", the sparsifier determines the relevant frames inside the clip to produce a mask " + }, + { + "bbox": [ + 130, + 184, + 482, + 293 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 184, + 482, + 293 + ], + "type": "text", + "content": " for retrieval attention. We incorporate a learnable token " + }, + { + "bbox": [ + 130, + 184, + 482, + 293 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 184, + 482, + 293 + ], + "type": "text", + "content": " to capture the context of the clip. Additionally, the frame features " + }, + { + "bbox": [ + 130, + 184, + 482, + 293 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 184, + 482, + 293 + ], + "type": "text", + "content": " undergo retrieval attention based on the predicted mask " + }, + { + "bbox": [ + 130, + 184, + 482, + 293 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 184, + 482, + 293 + ], + "type": "text", + "content": " to generate the content feature " + }, + { + "bbox": [ + 130, + 184, + 482, + 293 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 130, + 184, + 482, + 293 + ], + "type": "text", + "content": ". The combination of context " + }, + { + "bbox": [ + 130, + 184, + 482, + 293 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 184, + 482, + 293 + ], + "type": "text", + "content": " and content " + }, + { + "bbox": [ + 130, + 184, + 482, + 293 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 130, + 184, + 482, + 293 + ], + "type": "text", + "content": " forms the retrieved clip feature " + }, + { + "bbox": [ + 130, + 184, + 482, + 293 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 184, + 482, + 293 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 304, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 413 + ], + "type": "text", + "content": "Cross Attention: To evaluate the relevance of each frame within the clip " + }, + { + "bbox": [ + 130, + 304, + 482, + 413 + ], + "type": "inline_equation", + "content": "C^i" + }, + { + "bbox": [ + 130, + 304, + 482, + 413 + ], + "type": "text", + "content": ", we utilize cross-attention, where frame features act as queries and text features serve as keys and values. Cross-attention represents the frame features as a weighted average of the textual features, scaled by their mutual correlation. This process results in text-conditioned frame features " + }, + { + "bbox": [ + 130, + 304, + 482, + 413 + ], + "type": "inline_equation", + "content": "F^i" + }, + { + "bbox": [ + 130, + 304, + 482, + 413 + ], + "type": "text", + "content": ", establishing a fine-grained correspondence between the modalities. The query, key and values are calculated as " + }, + { + "bbox": [ + 130, + 304, + 482, + 413 + ], + "type": "inline_equation", + "content": "Q^i = l_Q(C^i) \\in R^{L_c \\times D_f}" + }, + { + "bbox": [ + 130, + 304, + 482, + 413 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 304, + 482, + 413 + ], + "type": "inline_equation", + "content": "K = l_K(S) \\in R^{N \\times D_q}" + }, + { + "bbox": [ + 130, + 304, + 482, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 304, + 482, + 413 + ], + "type": "inline_equation", + "content": "V = l_V(S) \\in R^{N \\times D_q}" + }, + { + "bbox": [ + 130, + 304, + 482, + 413 + ], + "type": "text", + "content": " respectively. Here, " + }, + { + "bbox": [ + 130, + 304, + 482, + 413 + ], + "type": "inline_equation", + "content": "l_Q(\\cdot), l_K(\\cdot)" + }, + { + "bbox": [ + 130, + 304, + 482, + 413 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 304, + 482, + 413 + ], + "type": "inline_equation", + "content": "l_V(\\cdot)" + }, + { + "bbox": [ + 130, + 304, + 482, + 413 + ], + "type": "text", + "content": " denote the projection layers for the query, key, and value. Then, we apply the cross-attention layer as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 245, + 418, + 480, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 418, + 480, + 432 + ], + "spans": [ + { + "bbox": [ + 245, + 418, + 480, + 432 + ], + "type": "interline_equation", + "content": "F ^ {i} = \\operatorname {s o f t m a x} \\left(Q ^ {i} K ^ {T}\\right) V + Q ^ {i} \\tag {1}", + "image_path": "c56dba39eddbc8ddd5260fae4f0dbcd59158f5efc78e7040a1d67f7a33cf3557.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "type": "text", + "content": "Sparsifier: To retrieve the clip, we need to assess the relevance of each clip based on the text query. Clip relevance is determined by aggregating the relevancy of its frames. However, given the small moment duration, most frames are unrelated to the query. Hence, we classify frames into relevant and non-relevant categories using the sparsifier to assist retrieval attention in focusing on these fine-grained events. We calculate the relevancy " + }, + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "type": "inline_equation", + "content": "G^{j} \\in [0,1]" + }, + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "type": "text", + "content": " frame with Eq. 2. We use a differentiable Gumbel-softmax function [15, 17] for enabling end-to-end training. The detailed calculation is in the supplementary materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 233, + 551, + 480, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 551, + 480, + 563 + ], + "spans": [ + { + "bbox": [ + 233, + 551, + 480, + 563 + ], + "type": "interline_equation", + "content": "G ^ {j} = \\operatorname {G u m b e l - s o f t m a x} \\left(\\operatorname {l i n e a r} \\left(f ^ {j}\\right)\\right) \\tag {2}", + "image_path": "acd62d6afcf7e4d1b2a3efa50c02c8198176f94e2c8b958025c12869b30c95bd.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "The intra-clip attention loss optimizes this classification using the ground truth frame relevancy information described in Sec. 3.4. To distinguish irrelevant frames from their counterparts, we constrain their updates in the retrieval attention. This restriction is implemented with an attention mask computed from the learned frame relevancy " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "G^{j}" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": ". The attention mask " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " frame is calculated with Eq. 3. Here, the " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " frames are the attention query and key, respectively. This leads to a reduced set of relevant and informative frames for subsequent attention." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "Hannan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 222, + 125, + 480, + 150 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 125, + 480, + 150 + ], + "spans": [ + { + "bbox": [ + 222, + 125, + 480, + 150 + ], + "type": "interline_equation", + "content": "M (j, k) = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} G ^ {j} > 0. 5 \\text {o r} j = k \\\\ - \\infty & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {3}", + "image_path": "0d53e2d3815af29af7cbe886977f9bdd5f0ddaf6b3071d8e2ae5a727b1105976.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 161, + 479, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 479, + 232 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 479, + 232 + ], + "type": "text", + "content": "Retrieval Attention: We aim for a learnable clip retrieval function to train end-to-end with the grounding objective. Thus, we devise the retrieval attention for aggregating the text-conditioned frame features based on their relevancy. To condense the " + }, + { + "bbox": [ + 130, + 161, + 479, + 232 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 130, + 161, + 479, + 232 + ], + "type": "text", + "content": " clip into a contextual feature embedding, we introduce a randomly initialized learnable vector " + }, + { + "bbox": [ + 130, + 161, + 479, + 232 + ], + "type": "inline_equation", + "content": "R^i" + }, + { + "bbox": [ + 130, + 161, + 479, + 232 + ], + "type": "text", + "content": ". We denote this as retrieval token, which is trained with the inter-clip contrastive loss." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 233, + 480, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 233, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 480, + 316 + ], + "type": "text", + "content": "First, we concatenate the retrieval token with the text-conditioned frame features to create the queries " + }, + { + "bbox": [ + 130, + 233, + 480, + 316 + ], + "type": "inline_equation", + "content": "Q^i = [R^i,F^i]" + }, + { + "bbox": [ + 130, + 233, + 480, + 316 + ], + "type": "text", + "content": " for retrieval attention. We also generate an attention mask containing all zeros for the retrieval token and concatenate it with " + }, + { + "bbox": [ + 130, + 233, + 480, + 316 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 233, + 480, + 316 + ], + "type": "text", + "content": ". We set the key " + }, + { + "bbox": [ + 130, + 233, + 480, + 316 + ], + "type": "inline_equation", + "content": "K = Q^i" + }, + { + "bbox": [ + 130, + 233, + 480, + 316 + ], + "type": "text", + "content": ", and the value " + }, + { + "bbox": [ + 130, + 233, + 480, + 316 + ], + "type": "inline_equation", + "content": "V = Q^i" + }, + { + "bbox": [ + 130, + 233, + 480, + 316 + ], + "type": "text", + "content": " to calculate the retrieval attention according to Eq. 4. We get the clip-level context feature, " + }, + { + "bbox": [ + 130, + 233, + 480, + 316 + ], + "type": "inline_equation", + "content": "R^i = \\tilde{Q}^{i,1}" + }, + { + "bbox": [ + 130, + 233, + 480, + 316 + ], + "type": "text", + "content": " and frame-level content features, " + }, + { + "bbox": [ + 130, + 233, + 480, + 316 + ], + "type": "inline_equation", + "content": "F_{c}^{i,j} = [\\tilde{Q}^{i,2},\\tilde{Q}^{i,3},\\dots,\\tilde{Q}^{i,L_{c} + 1}]" + }, + { + "bbox": [ + 130, + 233, + 480, + 316 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 233, + 480, + 316 + ], + "type": "inline_equation", + "content": "F_{c}^{i,j}" + }, + { + "bbox": [ + 130, + 233, + 480, + 316 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 130, + 233, + 480, + 316 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 130, + 233, + 480, + 316 + ], + "type": "text", + "content": " frame feature of the " + }, + { + "bbox": [ + 130, + 233, + 480, + 316 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 130, + 233, + 480, + 316 + ], + "type": "text", + "content": " clip." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 233, + 326, + 480, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 326, + 480, + 338 + ], + "spans": [ + { + "bbox": [ + 233, + 326, + 480, + 338 + ], + "type": "interline_equation", + "content": "\\tilde {Q} ^ {i} = \\operatorname {s o f t m a x} \\left(Q ^ {i} K ^ {T} + M\\right) V + Q ^ {i} \\tag {4}", + "image_path": "ac929a1e2f08b1dfd4341f109b88aa0680ea40d6e344227dc1645d544ea37500.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 347, + 480, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 480, + 442 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 480, + 442 + ], + "type": "text", + "content": "Further, the context feature " + }, + { + "bbox": [ + 130, + 347, + 480, + 442 + ], + "type": "inline_equation", + "content": "\\mathrm{R}" + }, + { + "bbox": [ + 130, + 347, + 480, + 442 + ], + "type": "text", + "content": " is utilized to retrieve the clip. It undergoes a linear projection layer, " + }, + { + "bbox": [ + 130, + 347, + 480, + 442 + ], + "type": "inline_equation", + "content": "l_{s}(\\cdot)" + }, + { + "bbox": [ + 130, + 347, + 480, + 442 + ], + "type": "text", + "content": ", to obtain context scores " + }, + { + "bbox": [ + 130, + 347, + 480, + 442 + ], + "type": "inline_equation", + "content": "S_{r} = \\operatorname{sigmoid}(l_{s}(R))" + }, + { + "bbox": [ + 130, + 347, + 480, + 442 + ], + "type": "text", + "content": ". We retrieve the most relevant clip based on the context score " + }, + { + "bbox": [ + 130, + 347, + 480, + 442 + ], + "type": "inline_equation", + "content": "S_{r}" + }, + { + "bbox": [ + 130, + 347, + 480, + 442 + ], + "type": "text", + "content": ". Importantly, the context " + }, + { + "bbox": [ + 130, + 347, + 480, + 442 + ], + "type": "inline_equation", + "content": "\\mathrm{R}" + }, + { + "bbox": [ + 130, + 347, + 480, + 442 + ], + "type": "text", + "content": " represents the queried moment by attending to the reduced set of relevant frames. So, we fuse it with the frame features with Eq. 5 to produce stronger multimodal inputs for the decoder. It enhances the decoder's localization capability significantly. The decoder only processes features from the retrieved clip to predict the moment." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 237, + 453, + 480, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 453, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 237, + 453, + 480, + 464 + ], + "type": "interline_equation", + "content": "P ^ {i, j} = F _ {c} ^ {i, j} + R ^ {i} \\times G ^ {j}, i \\in P, \\forall j \\tag {5}", + "image_path": "f72ec97bba4cdc4e8cb3c578db6b9bf5cc8a5b46e15444ac5ca749ea5e0c76f5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 482, + 261, + 494 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 482, + 261, + 494 + ], + "spans": [ + { + "bbox": [ + 131, + 482, + 261, + 494 + ], + "type": "text", + "content": "3.3 Grounding Decoder" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 496, + 479, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 479, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 479, + 592 + ], + "type": "text", + "content": "We process the retrieved clip features with a transformer decoder to predict precise moment boundaries. Previous methods [16,22] feed both modalities into the decoder, introducing multiple positional information (e.g., time in the video and word order in the text) that complicates the detection task. In contrast, our encoder injects text information directly into the clip, eliminating the requirement to feed both modalities into the decoder. We incorporate learnable anchor queries [27] to represent the moment's center and width as " + }, + { + "bbox": [ + 130, + 496, + 479, + 592 + ], + "type": "inline_equation", + "content": "(\\tau_c,\\tau_w)" + }, + { + "bbox": [ + 130, + 496, + 479, + 592 + ], + "type": "text", + "content": ". The decoder has 2 cross-attention and 2 self-attention layers." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 609, + 236, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 236, + 621 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 236, + 621 + ], + "type": "text", + "content": "3.4 Loss Functions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": "Intra-Clip Attention Loss guides our sparsifier to distinguish frames within and outside of the ground truth moment. This enables the retrieval attention to focus on relevant video regions and achieve fine-grained event understanding." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "text", + "content": "RGNet" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "Specifically, the in-moment frames are required to maintain higher relevancy scores than those outside the ground truth moment. The loss is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 216, + 146, + 481, + 158 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 146, + 481, + 158 + ], + "spans": [ + { + "bbox": [ + 216, + 146, + 481, + 158 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {a t t n}} = \\max (0, \\Delta + S _ {c} (i, j _ {\\mathrm {o u t}}) - S _ {c} (i, j _ {\\mathrm {i n}})) \\tag {6}", + "image_path": "4f99d9b440661d898ac05d7989a71c3ec083de652cdd33d6799d2331bad2f18f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 163, + 479, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 163, + 479, + 239 + ], + "spans": [ + { + "bbox": [ + 130, + 163, + 479, + 239 + ], + "type": "text", + "content": "In the equation, " + }, + { + "bbox": [ + 130, + 163, + 479, + 239 + ], + "type": "inline_equation", + "content": "S_{c}" + }, + { + "bbox": [ + 130, + 163, + 479, + 239 + ], + "type": "text", + "content": " is the relevancy score, " + }, + { + "bbox": [ + 130, + 163, + 479, + 239 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 130, + 163, + 479, + 239 + ], + "type": "text", + "content": " represents the margin, and anchor " + }, + { + "bbox": [ + 130, + 163, + 479, + 239 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 163, + 479, + 239 + ], + "type": "text", + "content": " is an in-moment frame. We randomly chose in-moment and out-of-moment frames " + }, + { + "bbox": [ + 130, + 163, + 479, + 239 + ], + "type": "inline_equation", + "content": "j_{\\mathrm{in}}" + }, + { + "bbox": [ + 130, + 163, + 479, + 239 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 163, + 479, + 239 + ], + "type": "inline_equation", + "content": "j_{\\mathrm{out}}" + }, + { + "bbox": [ + 130, + 163, + 479, + 239 + ], + "type": "text", + "content": " from the same clip. To calculate the relevancy score, the context features " + }, + { + "bbox": [ + 130, + 163, + 479, + 239 + ], + "type": "inline_equation", + "content": "R_{i}" + }, + { + "bbox": [ + 130, + 163, + 479, + 239 + ], + "type": "text", + "content": ", and encoded clip " + }, + { + "bbox": [ + 130, + 163, + 479, + 239 + ], + "type": "inline_equation", + "content": "P^{i,j}" + }, + { + "bbox": [ + 130, + 163, + 479, + 239 + ], + "type": "text", + "content": " undergoes linear layers, " + }, + { + "bbox": [ + 130, + 163, + 479, + 239 + ], + "type": "inline_equation", + "content": "l_{r}(\\cdot)" + }, + { + "bbox": [ + 130, + 163, + 479, + 239 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 163, + 479, + 239 + ], + "type": "inline_equation", + "content": "l_{c}(\\cdot)" + }, + { + "bbox": [ + 130, + 163, + 479, + 239 + ], + "type": "text", + "content": ", to obtain projected context " + }, + { + "bbox": [ + 130, + 163, + 479, + 239 + ], + "type": "inline_equation", + "content": "R_{proj}^{i} = l_{r}(R^{i}) \\in R^{D_{l}}" + }, + { + "bbox": [ + 130, + 163, + 479, + 239 + ], + "type": "text", + "content": " and content features, " + }, + { + "bbox": [ + 130, + 163, + 479, + 239 + ], + "type": "inline_equation", + "content": "P_{proj}^{i,j} = l_{c}(P^{i,j}) \\in R^{L_{c} \\times D_{l}}" + }, + { + "bbox": [ + 130, + 163, + 479, + 239 + ], + "type": "text", + "content": ". Then, we calculate the frame-level relevancy score:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 258, + 243, + 481, + 259 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 243, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 258, + 243, + 481, + 259 + ], + "type": "interline_equation", + "content": "S _ {c} (i, j) = R _ {p r o j} ^ {i} \\cdot P _ {p r o j} ^ {i, j} \\tag {7}", + "image_path": "17ccd3cd200d07d0c45752973b74fe7e6e03273c19ce9fe6b8378f4d4adc883c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 273, + 480, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 273, + 480, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 273, + 480, + 367 + ], + "type": "text", + "content": "Inter-Clip Contrastive loss. A long video contains many clips with similar environments and scenes. However, most of these clips do not contain the specific moment we are searching for. So, high-level video understanding is not enough to distinguish precise moments. During the test phase, we need to handle many such negative clips. To tackle this issue, unlike previous approaches [16], we sample a large number of negative clips and train our retriever on a large batch size. Hence, this negative clip sampling reduces the disparity between the train and test phases." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 369, + 480, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 480, + 441 + ], + "type": "text", + "content": "For this section, we denote the context feature, " + }, + { + "bbox": [ + 130, + 369, + 480, + 441 + ], + "type": "inline_equation", + "content": "R^{i,j}" + }, + { + "bbox": [ + 130, + 369, + 480, + 441 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 369, + 480, + 441 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 130, + 369, + 480, + 441 + ], + "type": "text", + "content": " proposal and " + }, + { + "bbox": [ + 130, + 369, + 480, + 441 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 130, + 369, + 480, + 441 + ], + "type": "text", + "content": " text query. We train our model contrastively with positive text-clip pairs " + }, + { + "bbox": [ + 130, + 369, + 480, + 441 + ], + "type": "inline_equation", + "content": "[R^{i,j}]_{i = j}" + }, + { + "bbox": [ + 130, + 369, + 480, + 441 + ], + "type": "text", + "content": " and negative pairs " + }, + { + "bbox": [ + 130, + 369, + 480, + 441 + ], + "type": "inline_equation", + "content": "[R^{i,j}]_{i\\neq j}" + }, + { + "bbox": [ + 130, + 369, + 480, + 441 + ], + "type": "text", + "content": ". We use the InfoNCE loss [36] to identify the positive text-clip pair amongst a set of unrelated negative samples. Here, " + }, + { + "bbox": [ + 130, + 369, + 480, + 441 + ], + "type": "inline_equation", + "content": "l_{cont}(\\cdot)" + }, + { + "bbox": [ + 130, + 369, + 480, + 441 + ], + "type": "text", + "content": " is a linear projection layer that transforms the " + }, + { + "bbox": [ + 130, + 369, + 480, + 441 + ], + "type": "inline_equation", + "content": "D_{f}" + }, + { + "bbox": [ + 130, + 369, + 480, + 441 + ], + "type": "text", + "content": " dimensional context feature into a one-dimensional logit." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 228, + 445, + 481, + 474 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 445, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 228, + 445, + 481, + 474 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n t}} = - \\sum_ {i} \\log \\frac {\\exp \\left(l _ {\\text {c o n t}} \\left(R ^ {i , i}\\right)\\right)}{\\sum_ {j} \\exp \\left(l _ {\\text {c o n t}} \\left(R ^ {i , j}\\right)\\right)} \\tag {8}", + "image_path": "95be59ebc8f3f41da1c459033fa9e10143afd4b2ac226e70fe682df77616442c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 482, + 481, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 482, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 482, + 481, + 579 + ], + "type": "text", + "content": "Grounding loss. The objective functions for grounding, which aim to locate desired moments, are adopted from the baseline approach [22]. The grounding loss " + }, + { + "bbox": [ + 130, + 482, + 481, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{g}}" + }, + { + "bbox": [ + 130, + 482, + 481, + 579 + ], + "type": "text", + "content": " measures the discrepancy between the ground truth (GT) and predicted moments. A one-to-one correspondence between GT and predicted moments is established using the Hungarian algorithm. This loss includes both an " + }, + { + "bbox": [ + 130, + 482, + 481, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}1" + }, + { + "bbox": [ + 130, + 482, + 481, + 579 + ], + "type": "text", + "content": " loss and a generalized IoU loss (" + }, + { + "bbox": [ + 130, + 482, + 481, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}\\mathrm{gIoU}" + }, + { + "bbox": [ + 130, + 482, + 481, + 579 + ], + "type": "text", + "content": "). Additionally, a cross-entropy loss " + }, + { + "bbox": [ + 130, + 482, + 481, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 130, + 482, + 481, + 579 + ], + "type": "text", + "content": " is employed to classify predicted moments as either foreground or background. Thus, " + }, + { + "bbox": [ + 130, + 482, + 481, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{g}}" + }, + { + "bbox": [ + 130, + 482, + 481, + 579 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 203, + 584, + 481, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 584, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 203, + 584, + 481, + 597 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {g}} = \\lambda_ {\\mathcal {L} 1} | | \\tau - \\hat {\\tau} | | + \\lambda_ {\\mathrm {g I o U}} \\mathcal {L} _ {\\mathrm {g I o U}} (\\tau , \\hat {\\tau}) + \\lambda_ {\\mathrm {C E}} \\mathcal {L} _ {\\mathrm {C E}}, \\tag {9}", + "image_path": "d6abe37413a4241d8510ddcadd56d4a25833cf7d23df96c4aecad460526ee2c0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 600, + 479, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 600, + 479, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 600, + 479, + 648 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 600, + 479, + 648 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 600, + 479, + 648 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 600, + 479, + 648 + ], + "type": "inline_equation", + "content": "\\hat{\\tau}" + }, + { + "bbox": [ + 130, + 600, + 479, + 648 + ], + "type": "text", + "content": " represent the ground-truth moment and its corresponding prediction, containing center coordinates " + }, + { + "bbox": [ + 130, + 600, + 479, + 648 + ], + "type": "inline_equation", + "content": "\\tau_{c}" + }, + { + "bbox": [ + 130, + 600, + 479, + 648 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 130, + 600, + 479, + 648 + ], + "type": "inline_equation", + "content": "\\tau_{w}" + }, + { + "bbox": [ + 130, + 600, + 479, + 648 + ], + "type": "text", + "content": ". The hyper-parameters " + }, + { + "bbox": [ + 130, + 600, + 479, + 648 + ], + "type": "inline_equation", + "content": "\\lambda_{*}" + }, + { + "bbox": [ + 130, + 600, + 479, + 648 + ], + "type": "text", + "content": " are used for balancing the losses. Finally, with the attention and contrastive loss, the total loss " + }, + { + "bbox": [ + 130, + 600, + 479, + 648 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{total}}" + }, + { + "bbox": [ + 130, + 600, + 479, + 648 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 228, + 654, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 654, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 228, + 654, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {\\text {a t t n}} \\mathcal {L} _ {\\text {a t t n}} + \\lambda_ {\\text {c o n t}} \\mathcal {L} _ {\\text {c o n t}} + \\mathcal {L} _ {\\mathrm {g}} \\tag {10}", + "image_path": "ea73b9d66dd465c711c958079b703257854cea676c51fa15929d15c9a723d527.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "Hannan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 272, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 272, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 272, + 129 + ], + "type": "text", + "content": "4 Experimental Setup" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 321, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 321, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 321, + 152 + ], + "type": "text", + "content": "4.1 Datasets and Evaluation Metric" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 159, + 481, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 159, + 481, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 159, + 481, + 232 + ], + "type": "text", + "content": "MAD [40] is an extensive dataset comprising 1.2K hours of full-length movies and 384K natural language queries, each associated with specific moments in the videos. The videos are orders of magnitude longer than previous datasets, with an average duration of 110 minutes, while the specified text moments are on average 4.1 seconds. This small moment-to-video ratio poses a significant challenge for the grounding task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 243, + 481, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 243, + 481, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 243, + 481, + 340 + ], + "type": "text", + "content": "Ego4D-NLQ [14] is a large-scale egocentric video data set with multiple challenges. Specifically, we use the episodic memory benchmark Ego4D-NLQ, which requires localizing where the answer to a natural language query can be seen. It contains around 13 template questions, amounting to around 74K queries. The train, val, and test sets contain 11.3K, 3.9K, and 4.0K queries. The video length ranges from 8 to 20 minutes, with an average of 8.25 minutes, and the average moment duration is 8.3 seconds. This means the moments constitute only " + }, + { + "bbox": [ + 132, + 243, + 481, + 340 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 132, + 243, + 481, + 340 + ], + "type": "text", + "content": " of the input video on average." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 351, + 481, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 351, + 481, + 411 + ], + "spans": [ + { + "bbox": [ + 132, + 351, + 481, + 411 + ], + "type": "text", + "content": "Grounding Metric: For grounding, we follow previous methods [16, 40], and adopt the standard metric " + }, + { + "bbox": [ + 132, + 351, + 481, + 411 + ], + "type": "inline_equation", + "content": "\\mathrm{Recall}@\\mathbf{k}" + }, + { + "bbox": [ + 132, + 351, + 481, + 411 + ], + "type": "text", + "content": " at IoU = " + }, + { + "bbox": [ + 132, + 351, + 481, + 411 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 132, + 351, + 481, + 411 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 132, + 351, + 481, + 411 + ], + "type": "inline_equation", + "content": "\\mathrm{Rk}_{\\theta}" + }, + { + "bbox": [ + 132, + 351, + 481, + 411 + ], + "type": "text", + "content": "). This metric represents the percentage of testing samples with at least one grounding prediction whose intersection over union (IoU) with the ground truth (GT) is larger than " + }, + { + "bbox": [ + 132, + 351, + 481, + 411 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 132, + 351, + 481, + 411 + ], + "type": "text", + "content": " among the top-" + }, + { + "bbox": [ + 132, + 351, + 481, + 411 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 132, + 351, + 481, + 411 + ], + "type": "text", + "content": " predictions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 423, + 481, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 423, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 423, + 481, + 471 + ], + "type": "text", + "content": "Retrieval Metric: To assess our proposal retrieval stage independently of the grounding, we utilize the standard retrieval metric [13], Recall at Rank " + }, + { + "bbox": [ + 132, + 423, + 481, + 471 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 132, + 423, + 481, + 471 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 132, + 423, + 481, + 471 + ], + "type": "inline_equation", + "content": "\\mathrm{R}@\\mathbf{k}" + }, + { + "bbox": [ + 132, + 423, + 481, + 471 + ], + "type": "text", + "content": "). This metric calculates the percentage of GT moments present in the top-" + }, + { + "bbox": [ + 132, + 423, + 481, + 471 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 132, + 423, + 481, + 471 + ], + "type": "text", + "content": " retrieved proposals." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 489, + 282, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 489, + 282, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 489, + 282, + 502 + ], + "type": "text", + "content": "4.2 Implementation Details:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 509, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 481, + 665 + ], + "type": "text", + "content": "We use pre-trained CLIP [38] and EgoVLP [24] models to extract video frames from MAD and Ego4D. Text features for both datasets are extracted using CLIP. The feature extractors are frozen, and their pre-trained weights remain unchanged during training. For Ego4D, we pre-train our model with NaQ annotations from previous work [39]. The proposal length, " + }, + { + "bbox": [ + 132, + 509, + 481, + 665 + ], + "type": "inline_equation", + "content": "W_{c}" + }, + { + "bbox": [ + 132, + 509, + 481, + 665 + ], + "type": "text", + "content": " is set to 180s for MAD and 48s for Ego4D, with only the top-30 and top-5 proposals retrieved, respectively. Training on Ego4D and MAD is conducted on four Nvidia-RTX-A6000 GPUs while fine-tuning on Ego4D is performed on a single GPU. Considering the longer proposals in MAD, we utilize a batch size of 32 for Ego4D and 8 for MAD. The number of moment queries is set to 5. For loss computation, we set hyperparameters as " + }, + { + "bbox": [ + 132, + 509, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{L1} = 10" + }, + { + "bbox": [ + 132, + 509, + 481, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 509, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{gIoU}} = 1" + }, + { + "bbox": [ + 132, + 509, + 481, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 509, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{CE}} = 4" + }, + { + "bbox": [ + 132, + 509, + 481, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 509, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{samp}} = 1" + }, + { + "bbox": [ + 132, + 509, + 481, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 509, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{cont}} = 10" + }, + { + "bbox": [ + 132, + 509, + 481, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 132, + 509, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\Delta = 0.2" + }, + { + "bbox": [ + 132, + 509, + 481, + 665 + ], + "type": "text", + "content": ". We use Xavier initialization [12] and employ AdamW [30] with an initial learning rate of " + }, + { + "bbox": [ + 132, + 509, + 481, + 665 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 132, + 509, + 481, + 665 + ], + "type": "text", + "content": ". RGNet is trained for 35 and 200 epochs on" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "text", + "content": "RGNet" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 481, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 481, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 481, + 140 + ], + "type": "text", + "content": "MAD and Ego4D datasets. The initial learning rate is reduced by one order of magnitude at epochs 25 for MAD and 120 for Ego4D." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 156, + 279, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 156, + 279, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 156, + 279, + 170 + ], + "type": "text", + "content": "5 Results and Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 178, + 481, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 178, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 178, + 481, + 214 + ], + "type": "text", + "content": "We first compare our performance with previous state-of-the-art models. Then, we report detailed ablation studies of our proposed method and visualize the qualitative results compared to the disjoint baseline." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 132, + 232, + 481, + 376 + ], + "blocks": [ + { + "bbox": [ + 132, + 232, + 481, + 376 + ], + "lines": [ + { + "bbox": [ + 132, + 232, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 232, + 481, + 376 + ], + "type": "table", + "html": "
ModelNaQEgo4D-NLQ [14]MAD [40]Avg
R1.3R5.3R1.5R5.5AvgR1.1R5.1R1.3R5.3R1.5R5.5Avg
2D-TAN [55]X5.0412.892.025.886.463.2211.902.529.251.585.696.07
UniVTG [25]X11.747.543.257.887.60-------
VSLNet [53]✓10.2619.015.8112.6711.93-------
VLG-Net [41]X-----3.6411.662.769.311.655.995.84
M-DETR [22]X8.2323.235.0113.3712.463.6012.982.819.861.675.586.08
M-Guide [2]X-----9.3018.964.6513.062.167.409.26
SOONet [37]X8.0022.403.7611.0911.3111.2623.219.0019.645.3213.1413.59
H-Hands [52]X13.2023.307.9015.6015.00-------
CONE [16]X14.1530.338.1818.0217.678.9020.516.8716.114.109.5911.01
EgoVLP✓15.9026.389.4617.8017.38-------
ReLeR [39]✓19.3123.6211.5915.7517.57-------
Ours (Default)X18.2834.0212.0422.8921.8112.4325.129.4818.725.6110.8613.70
Ours✓20.6341.6712.4725.0824.9612.4325.129.4818.725.6110.8613.70
", + "image_path": "b930333388c1144e6b9e24c5938decc8a3fb51d0287f43559153e235342cfd1b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 378, + 481, + 422 + ], + "lines": [ + { + "bbox": [ + 132, + 378, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 132, + 378, + 481, + 422 + ], + "type": "text", + "content": "Table 1: Main results on Ego4D-NLQ and MAD. RGNet achieves state-of-the-art performance on both datasets. Our default network is trained without NaQ annotations [39] on Ego4D. Even without NaQ annotations, our default network shows a larger improvement, underscoring the effectiveness of our solution with limited data." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 436, + 336, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 436, + 336, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 336, + 449 + ], + "type": "text", + "content": "5.1 Result on the Ego4D-NLQ Dataset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 453, + 481, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 481, + 609 + ], + "type": "text", + "content": "RGNet demonstrates state-of-the-art performance on all metrics for the Ego4D-NLQ benchmark. As shown in Table 1, RGNet improves its primary evaluation metric, the R1.3 and R5.3 score average by " + }, + { + "bbox": [ + 132, + 453, + 481, + 609 + ], + "type": "inline_equation", + "content": "9.7\\%" + }, + { + "bbox": [ + 132, + 453, + 481, + 609 + ], + "type": "text", + "content": ". Separately, we outperform the previous best methods on R1.3 and R5.3 by " + }, + { + "bbox": [ + 132, + 453, + 481, + 609 + ], + "type": "inline_equation", + "content": "1.32\\%" + }, + { + "bbox": [ + 132, + 453, + 481, + 609 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 453, + 481, + 609 + ], + "type": "inline_equation", + "content": "18.1\\%" + }, + { + "bbox": [ + 132, + 453, + 481, + 609 + ], + "type": "text", + "content": ", respectively. This improvement is consistent across other metrics as well. The unified clip and frame-level modeling in RGNet parallelly capture long- and short-range temporal dependencies, contributing to its superior performance. By modeling the entire video as a whole, RGNet learns better cross-modal alignment by contrasting visual scenarios from various events within the same video. Notably, compared to SooNet [37] and CONE [16], two prominent proposal-based methods, RGNet achieves an average score improvement of " + }, + { + "bbox": [ + 132, + 453, + 481, + 609 + ], + "type": "inline_equation", + "content": "13.7\\%" + }, + { + "bbox": [ + 132, + 453, + 481, + 609 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 453, + 481, + 609 + ], + "type": "inline_equation", + "content": "7.3\\%" + }, + { + "bbox": [ + 132, + 453, + 481, + 609 + ], + "type": "text", + "content": ". This substantial improvement is attributed to our unified modeling of clip retrieval, whereas they separately employ a retrieval similarity search between the video and the text." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 624, + 303, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 303, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 303, + 635 + ], + "type": "text", + "content": "5.2 Result on the MAD Dataset" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "content": "We compare our model's performance on the MAD dataset in Tab. 1. RGNet achieves state-of-the-art performance on both the R1.1 and R5.1 scores by out-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "Hannan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "performing the previous best SooNet [37] by " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": ". Together with its competitive performance on other metrics, RGNet demonstrates the importance of end-to-end modeling of long videos. Specifically, our performance of R1 is superior at all IoU thresholds, which testifies to the effectiveness of the proposed retrieval method. Notably, prior models rely on heuristic dot product similarity search applied to CLIP [38] features for the retrieval phase. In contrast, our shared retrieval feature and its unified optimization with the grounding objective drastically enhance the first stage retrieval, as seen in Tab. 2. It further helps the network capture more discriminative events crucial for detecting extremely tiny moments in hour-long videos." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 251, + 326, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 251, + 326, + 264 + ], + "spans": [ + { + "bbox": [ + 131, + 251, + 326, + 264 + ], + "type": "text", + "content": "5.3 Disjoint vs. Unified Architecture" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 268, + 482, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 482, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 482, + 400 + ], + "type": "text", + "content": "Since the long video temporal grounding (LVTG) performance relies on the first stage of clip retrieval, it does not perfectly reflect the actual moment localization capability of its grounding network. To assess the standalone grounding performance, we designed an oracle experiment operating exclusively on the clip where the ground truth moment is present. Compared to the oracle grounding, the LVTG " + }, + { + "bbox": [ + 130, + 268, + 482, + 400 + ], + "type": "inline_equation", + "content": "R1_{3}" + }, + { + "bbox": [ + 130, + 268, + 482, + 400 + ], + "type": "text", + "content": " drops by " + }, + { + "bbox": [ + 130, + 268, + 482, + 400 + ], + "type": "inline_equation", + "content": "15.7\\%" + }, + { + "bbox": [ + 130, + 268, + 482, + 400 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 268, + 482, + 400 + ], + "type": "inline_equation", + "content": "23.9\\%" + }, + { + "bbox": [ + 130, + 268, + 482, + 400 + ], + "type": "text", + "content": " for Ego4D and MAD in Tab. 2. The poor grounding in LVTG emerges from the suboptimal retrieval accuracy of the disjoint network. Hence, retrieving clips where the ground truth moment exists is crucial, independent of the subsequent grounding accuracy. Regardless of the effectiveness of moment boundary detection, if the clip containing the moment boundary cannot be accurately retrieved. Our unified approach consis" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 165, + 419, + 449, + 503 + ], + "blocks": [ + { + "bbox": [ + 165, + 419, + 449, + 503 + ], + "lines": [ + { + "bbox": [ + 165, + 419, + 449, + 503 + ], + "spans": [ + { + "bbox": [ + 165, + 419, + 449, + 503 + ], + "type": "table", + "html": "
DataModelClipRetrievalOracleGroundingLVTG
R@1R@5R1.3R5.3R1.3R5.3
Ego4dBaseline31.7164.6329.8454.0114.1530.33
Ours42.0876.2836.5363.6420.6341.67
MADBaseline12.4124.5029.4953.026.8716.11
Ours25.0150.0233.4263.439.4818.72
", + "image_path": "fb0c45b8ba4c4ba58dbc050e7f22ed32784e3612bc281e8b0a165f1b1b919a3b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 505, + 482, + 561 + ], + "lines": [ + { + "bbox": [ + 130, + 505, + 482, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 505, + 482, + 561 + ], + "type": "text", + "content": "Table 2: Empirical impact of two stages. To assess standalone grounding capability, we run an oracle experiment on the clip where the ground truth moment is present. The grounding capability degrades in LVTG evaluation because of incorrect selection by the disjoint clip retrieval network. Our unified model improves retrieval significantly, which leads to more effective temporal grounding in long videos." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "tently improves all retrieval metrics compared to the disjoint baseline. Notably, we achieve a staggering " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "10.4\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "12.6\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " improved R@1 score for Ego4D and MAD. This improvement in clip retrieval leads to state-of-the-art LVTG performance. Moreover, the grounding also improves because of the stronger clip features generated by our encoder. Compared to the disjoint baseline, the oracle R1.3 scores improved by about " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "6.7\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "4.0\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " for Ego4D and MAD, respectively. Hence, the unified end-to-end architecture proves more effective at LVTG by mutually improving both stages." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "text", + "content": "RGNet" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 246, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 246, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 246, + 126 + ], + "type": "text", + "content": "5.4 Ablation Studies" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 142, + 482, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 142, + 482, + 191 + ], + "spans": [ + { + "bbox": [ + 130, + 142, + 482, + 191 + ], + "type": "text", + "content": "This section reports detailed ablation studies on the proposed modules and loss functions. We also ablate with various clip lengths and retrieved clip numbers. Unless stated, the experiments are done with the Default network w/o NaQ annotations [39] for the Ego4D-NLQ dataset and CONE [16] as the baseline." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 185, + 213, + 430, + 312 + ], + "blocks": [ + { + "bbox": [ + 185, + 213, + 430, + 312 + ], + "lines": [ + { + "bbox": [ + 185, + 213, + 430, + 312 + ], + "spans": [ + { + "bbox": [ + 185, + 213, + 430, + 312 + ], + "type": "table", + "html": "
ModelR1.3R5.3R1.5R5.5
RGNet (Default)18.2834.0212.0422.89
w/o Retrieval Token17.8033.9911.2522.32
w/o Sparsifier16.1231.579.9120.47
w/o RG-Encoder*14.1530.338.1818.02
w/o Contrastive Loss17.4132.1210.7922.80
w/o Attention Loss16.2131.599.9120.53
", + "image_path": "a7843a2c79407b07c6ebeb38f798b40a6016251a422474c3646cc3a08ba3d150.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 312, + 482, + 357 + ], + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 357 + ], + "type": "text", + "content": "Table 3: Cumulative ablation study for the proposed modules and losses of RGNet. The experiment is conducted w/o NaQ augmentation on the Ego4D dataset. In each step, we remove one of our proposed modules or losses to evaluate their individual impact. *denotes the baseline model." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 374, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 445 + ], + "type": "text", + "content": "Modules: In our initial experiments, we validate the effectiveness of each proposed module with a cumulative ablation in Tab. 3. The unified network demonstrates superior performance across all evaluation metrics, notably achieving an R1.3 score of " + }, + { + "bbox": [ + 130, + 374, + 482, + 445 + ], + "type": "inline_equation", + "content": "18.28\\%" + }, + { + "bbox": [ + 130, + 374, + 482, + 445 + ], + "type": "text", + "content": ". Note that the disjoint baseline performs " + }, + { + "bbox": [ + 130, + 374, + 482, + 445 + ], + "type": "inline_equation", + "content": "4.2\\%" + }, + { + "bbox": [ + 130, + 374, + 482, + 445 + ], + "type": "text", + "content": " worse than our unified network. This performance difference can be attributed to our models unified architecture." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 448, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 556 + ], + "type": "text", + "content": "With the removal of the retrieval token for generating the clip feature in Eq. 5, considerably drops the score by about " + }, + { + "bbox": [ + 130, + 448, + 482, + 556 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 130, + 448, + 482, + 556 + ], + "type": "text", + "content": ". This proves that modeling both clip level context and frame level content are necessary to produce strong multimodal features for the grounding decoder. Subsequently, eliminating the sparsifier, there is a decline in " + }, + { + "bbox": [ + 130, + 448, + 482, + 556 + ], + "type": "inline_equation", + "content": "\\mathrm{R1}_{.3}" + }, + { + "bbox": [ + 130, + 448, + 482, + 556 + ], + "type": "text", + "content": " performance of approximately " + }, + { + "bbox": [ + 130, + 448, + 482, + 556 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 130, + 448, + 482, + 556 + ], + "type": "text", + "content": ". This verifies the impact of the sparsifier in modeling the fine-grained event boundaries. Finally, removing the RG-Encoder yields a notable degradation of approximately " + }, + { + "bbox": [ + 130, + 448, + 482, + 556 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 130, + 448, + 482, + 556 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 130, + 448, + 482, + 556 + ], + "type": "inline_equation", + "content": "\\mathrm{R1}_{.3}" + }, + { + "bbox": [ + 130, + 448, + 482, + 556 + ], + "type": "text", + "content": " score, which substantiates its capability to model clip and frame level granularity jointly." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 568, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 628 + ], + "type": "text", + "content": "Loss Functions: Training RGNet without our negative clip sampling strategy for the contrastive loss (Tab. 3) results in an almost " + }, + { + "bbox": [ + 130, + 568, + 482, + 628 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 130, + 568, + 482, + 628 + ], + "type": "text", + "content": " decline in the R1.3 score. Our sampling strategy mimics long-video setup closely during training, which improves event discrimination capability inside the same scene and surroundings." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "By eliminating the attention loss, we observe a further reduction of " + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": " in grounding performance. The attention loss enables our encoder to model fine-grained events in long videos which is crucial for specific moment detection." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "Hannan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 128, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 128, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 128, + 482, + 236 + ], + "type": "text", + "content": "Number of Clips: The number of retrieved proposal clips significantly influences the Long Video Temporal Grounding (LVTG) performance. Our model consistently outperforms the baseline across various choices of top-k retrieved clips, as shown in Fig. 4. Even with only a single retrieved clip, our performance surpasses the baseline's optimal result for both datasets. Moreover, our model exhibits a less steep performance decline when fewer clips are retrieved. Importantly, with fewer clips, the grounding network runs fewer times, leading to an overall speedup. Therefore, our model strikes a favorable balance between efficiency and performance, showcasing superior performance than the baseline." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 204, + 258, + 304, + 342 + ], + "blocks": [ + { + "bbox": [ + 204, + 258, + 304, + 342 + ], + "lines": [ + { + "bbox": [ + 204, + 258, + 304, + 342 + ], + "spans": [ + { + "bbox": [ + 204, + 258, + 304, + 342 + ], + "type": "image", + "image_path": "7ee740ac8a82dad00098dd757937a2858c2e768cc2036277845a1f02225463e9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 354, + 482, + 399 + ], + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 399 + ], + "type": "text", + "content": "Fig. 4: Impact of number of retrieved clips. Reducing the number of clips speeds up the network execution time. While the baseline model experiences a significant drop in performance with this reduction, RGNet shows a noticeably smaller decline in performance under the same conditions." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 309, + 259, + 411, + 342 + ], + "blocks": [ + { + "bbox": [ + 309, + 259, + 411, + 342 + ], + "lines": [ + { + "bbox": [ + 309, + 259, + 411, + 342 + ], + "spans": [ + { + "bbox": [ + 309, + 259, + 411, + 342 + ], + "type": "image", + "image_path": "2cea0e1ac86b544dfe9d3eb325bd2ab122c5ec943838025be54eaa3beb39a6b3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 417, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 417, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 417, + 482, + 525 + ], + "type": "text", + "content": "**Clip Length:** Varying clip length impacts both retrieval and grounding performance. The number of clip decreases with longer clips, making the retrieval task easier. Fig. 5 reports a monotonically improving retrieval with longer clips. However, after a certain length, moment localization quality starts to deteriorate. With longer clips, the task becomes extremely difficult. We see the decline after the 180s and 48s for the MAD and Ego4D datasets. Hence, we set the default clip length of our method to these optimal lengths. Importantly, the unified network helps outscore the disjoint baseline on all tested clip lengths (inversely proportional to the clip numbers in Fig. 4)." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 200, + 546, + 302, + 625 + ], + "blocks": [ + { + "bbox": [ + 200, + 546, + 302, + 625 + ], + "lines": [ + { + "bbox": [ + 200, + 546, + 302, + 625 + ], + "spans": [ + { + "bbox": [ + 200, + 546, + 302, + 625 + ], + "type": "image", + "image_path": "42237315c5214986a41bddd5bd1ff81f7a59e11cb15f98fb772901c7fcf5400a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 636, + 482, + 681 + ], + "lines": [ + { + "bbox": [ + 130, + 636, + 482, + 681 + ], + "spans": [ + { + "bbox": [ + 130, + 636, + 482, + 681 + ], + "type": "text", + "content": "Fig. 5: Impact of retrieved clip length. Longer clips result in improved retrieval due to fewer candidates. However, grounding becomes exceedingly difficult with longer clips. For example, the grounding performance drops after 180 seconds and 48 seconds for the MAD and Ego4D datasets." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 302, + 546, + 414, + 625 + ], + "blocks": [ + { + "bbox": [ + 302, + 546, + 414, + 625 + ], + "lines": [ + { + "bbox": [ + 302, + 546, + 414, + 625 + ], + "spans": [ + { + "bbox": [ + 302, + 546, + 414, + 625 + ], + "type": "image", + "image_path": "75b34f5a975d9f83f13448b05d8dfaee9e477ef4dffb730d8cf7bbdf8f7cd4e0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "text", + "content": "RGNet" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 263, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 263, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 263, + 128 + ], + "type": "text", + "content": "5.5 Qualitative Analysis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 135, + 482, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 135, + 482, + 255 + ], + "spans": [ + { + "bbox": [ + 130, + 135, + 482, + 255 + ], + "type": "text", + "content": "We visualize the qualitative predictions of RGNet compared to the disjoint baseline in Fig. 6. For the first and second queries, the baseline struggles to retrieve the correct clip where the scissor appears. RGNet accurately retrieves the clip with the searched object and precisely localizes the moment it appeared in the video. Disjoint retrieval often leads to incorrect clip selection, which the final grounding network cannot recover from. For the third query, the baseline selects the correct clip but fails to localize the precise moment the man closes the car door, indicating a gap between the two stages. RGNet mitigates this gap by leveraging features from the retrieved clip to improve moment localization. More visual outputs are included the supplementary materials." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 168, + 285, + 447, + 329 + ], + "blocks": [ + { + "bbox": [ + 247, + 274, + 339, + 284 + ], + "lines": [ + { + "bbox": [ + 247, + 274, + 339, + 284 + ], + "spans": [ + { + "bbox": [ + 247, + 274, + 339, + 284 + ], + "type": "text", + "content": "Query 1: Where is the scissor?" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 168, + 285, + 447, + 329 + ], + "lines": [ + { + "bbox": [ + 168, + 285, + 447, + 329 + ], + "spans": [ + { + "bbox": [ + 168, + 285, + 447, + 329 + ], + "type": "image", + "image_path": "ea0a59213b6660ed4bc74e839ceb6b949aa4d53aa9c71487292fb574909eacaf.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 168, + 345, + 446, + 392 + ], + "blocks": [ + { + "bbox": [ + 231, + 332, + 364, + 342 + ], + "lines": [ + { + "bbox": [ + 231, + 332, + 364, + 342 + ], + "spans": [ + { + "bbox": [ + 231, + 332, + 364, + 342 + ], + "type": "text", + "content": "Query 2: How many stainless bowls did I pick?" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 168, + 345, + 446, + 392 + ], + "lines": [ + { + "bbox": [ + 168, + 345, + 446, + 392 + ], + "spans": [ + { + "bbox": [ + 168, + 345, + 446, + 392 + ], + "type": "image", + "image_path": "8f75fc2e4626d4706db20f6a1f70073f6b6ab6ecf466b45449ae5b0f07307483.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 169, + 407, + 419, + 449 + ], + "blocks": [ + { + "bbox": [ + 241, + 397, + 354, + 407 + ], + "lines": [ + { + "bbox": [ + 241, + 397, + 354, + 407 + ], + "spans": [ + { + "bbox": [ + 241, + 397, + 354, + 407 + ], + "type": "text", + "content": "Query 3: Did I leave the car door open?" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 169, + 407, + 419, + 449 + ], + "lines": [ + { + "bbox": [ + 169, + 407, + 419, + 449 + ], + "spans": [ + { + "bbox": [ + 169, + 407, + 419, + 449 + ], + "type": "image", + "image_path": "907b19be12350262c7b4fbc06206b8afdb4a9d8380bccd71e2f4c68195f155c5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 458, + 482, + 525 + ], + "lines": [ + { + "bbox": [ + 130, + 458, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 482, + 525 + ], + "type": "text", + "content": "Fig. 6: Qualitative Results. The baseline fails to retrieve the correct clip in the first two queries. Since they primarily depict the same indoor room throughout the whole video, precise event discrimination is vital for accurate clip retrieval. In the third query, the baseline cannot identify the moment within the correctly retrieved clip, detecting it only after it has concluded. Precise event localization demands improved alignment between visual events and the queried text. RGNet correctly localizes all the moments." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 533, + 220, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 220, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 220, + 546 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "We introduce an end-to-end model for long video temporal grounding, unifying the two stages of prevailing methods with shared features and mutual optimization. We conduct independent analyses of the two stages, shaping our solution based on their distinct impacts. This leads to a better understanding of fine-grained events in long videos. Our approach demonstrates state-of-the-art performance on challenging long video grounding datasets, validating its effectiveness. Like all LVTG methods, we rely on a pre-trained image encoder to extract the visual features. A promising future direction includes eliminating the need for an image encoder and training directly on the raw video data." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "Hannan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "type": "text", + "content": "1. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: ICCV. pp. 1728-1738 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 159, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 159, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 138, + 159, + 480, + 191 + ], + "type": "text", + "content": "2. Barrios, W., Soldan, M., Ceballos-Arroyo, A.M., Heilbron, F.C., Ghanem, B.: Localizing moments in long video via multimodal guidance. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13667-13678 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 191, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 191, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 138, + 191, + 480, + 224 + ], + "type": "text", + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 224, + 480, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 480, + 245 + ], + "type": "text", + "content": "4. Chen, S., Zhao, Y., Jin, Q., Wu, Q.: Fine-grained video-text retrieval with hierarchical graph reasoning. In: CVPR. pp. 10638-10647 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 245, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 480, + 277 + ], + "type": "text", + "content": "5. Chen, Z., Jiang, X., Xu, X., Cao, Z., Mo, Y., Shen, H.T.: Joint searching and grounding: Multi-granularity video content retrieval. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 975-983 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 277, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 480, + 310 + ], + "type": "text", + "content": "6. Cheng, F., Wang, X., Lei, J., Crandall, D., Bansal, M., Bertasius, G.: Vindlu: A recipe for effective video-and-language pretraining. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 310, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 310, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 310, + 480, + 331 + ], + "type": "text", + "content": "7. Dong, J., Li, X., Xu, C., Yang, X., Yang, G., Wang, X., Wang, M.: Dual encoding for video retrieval by text. IEEE TPAMI pp. 4065-4080 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 331, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 480, + 363 + ], + "type": "text", + "content": "8. Fang, B., Liu, C., Zhou, Y., Yang, M., Song, Y., Li, F., Wang, W., Ji, X., Ouyang, W., et al.: Uatvr: Uncertainty-adaptive text-video retrieval. arXiv preprint arXiv:2301.06309 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 363, + 480, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 363, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 138, + 363, + 480, + 384 + ], + "type": "text", + "content": "9. Gabeur, V., Sun, C., Alahari, K., Schmid, C.: Multi-modal transformer for video retrieval. In: ECCV. pp. 214-229 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 384, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 384, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 384, + 480, + 407 + ], + "type": "text", + "content": "0. Gao, Z., Liu, J., Sun, W., Chen, S., Chang, D., Zhao, L.: Clip2tv: Align, match and distill for video-text retrieval. arXiv preprint arXiv:2111.05610 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 407, + 480, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 480, + 427 + ], + "type": "text", + "content": "1. Ge, Y., Ge, Y., Liu, X., Li, D., Shan, Y., Qie, X., Luo, P.: Bridging video-text retrieval with multiple choice questions. In: CVPR. pp. 16167-16176 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 427, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 427, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 138, + 427, + 480, + 471 + ], + "type": "text", + "content": "2. Glorot, X., Bengio, Y.: Understanding the difficulty of training deep feedforward neural networks. In: Proceedings of the thirteenth international conference on artificial intelligence and statistics. pp. 249-256. JMLR Workshop and Conference Proceedings (2010)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 471, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 471, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 138, + 471, + 480, + 514 + ], + "type": "text", + "content": "3. Gorti, S.K., Vouitsis, N., Ma, J., Golestan, K., Volkovs, M., Garg, A., Yu, G.: X-pool: Cross-modal language-video attention for text-video retrieval. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5006-5015 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 514, + 480, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 480, + 547 + ], + "type": "text", + "content": "4. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: CVPR. pp. 18995-19012 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 547, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 480, + 579 + ], + "type": "text", + "content": "5. Hannan, T., Koner, R., Bernhard, M., Shit, S., Menze, B., Tresp, V., Schubert, M., Seidl, T.: Gratt-vis: Gated residual attention for auto rectifying video instance segmentation. arXiv preprint arXiv:2305.17096 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 579, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 579, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 579, + 480, + 611 + ], + "type": "text", + "content": "6. Hou, Z., Zhong, W., Ji, L., Gao, D., Yan, K., Chan, W.K., Ngo, C.W., Shou, Z., Duan, N.: Cone: An efficient coarse-to-fine alignment framework for long video temporal grounding. arXiv preprint arXiv:2209.10918 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 611, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 633 + ], + "type": "text", + "content": "7. Jang, E., Gu, S., Poole, B.: Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144 (2016)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "8. Jang, J., Park, J., Kim, J., Kwon, H., Sohn, K.: Knowing where to focus: Event-aware transformer for video grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13846-13856 (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "text", + "content": "RGNet" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "19. Jiang, H., Zhang, J., Huang, R., Ge, C., Ni, Z., Lu, J., Zhou, J., Song, S., Huang, G.: Cross-modal adapter for text-video retrieval. arXiv preprint arXiv:2211.09623 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "text", + "content": "20. Jin, P., Huang, J., Liu, F., Wu, X., Ge, S., Song, G., Clifton, D.A., Chen, J.: Expectation-maximization contrastive learning for compact video-and-language representations. NeurIPS (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "text", + "content": "21. Kiros, R., Salakhutdinov, R., Zemel, R.S.: Unifying visual-semantic embeddings with multimodal neural language models. arXiv preprint arXiv:1411.2539 (2014)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "22. Lei, J., Berg, T.L., Bansal, M.: Detecting moments and highlights in videos via natural language queries. Advances in Neural Information Processing Systems 34, 11846-11858 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 281 + ], + "type": "text", + "content": "23. Lei, J., Li, L., Zhou, L., Gan, Z., Berg, T.L., Bansal, M., Liu, J.: Less is more: Clipbert for video-and-language learning via sparse sampling. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7331-7341 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 315 + ], + "type": "text", + "content": "24. Lin, K.Q., Wang, A.J., Soldan, M., Wray, M., Yan, R., Xu, E.Z., Gao, D., Tu, R., Zhao, W., Kong, W., et al.: Egocentric video-language pretraining. arXiv preprint arXiv:2206.01670 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 358 + ], + "type": "text", + "content": "25. Lin, K.Q., Zhang, P., Chen, J., Pramanick, S., Gao, D., Wang, A.J., Yan, R., Shou, M.Z.: Univtg: Towards unified video-language temporal grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2794-2804 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "text", + "content": "26. Liu, N., Wang, X., Li, X., Yang, Y., Zhuang, Y.: Refer@ zju-alibaba submission to the ego4d natural language queries challenge 2022. arXiv preprint arXiv:2207.00383 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 424 + ], + "type": "text", + "content": "27. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 424, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 468 + ], + "type": "text", + "content": "28. Liu, Y., Li, S., Wu, Y., Chen, C.W., Shan, Y., Qie, X.: Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3042-3051 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "text", + "content": "29. Liu, Y., Xiong, P., Xu, L., Cao, S., Jin, Q.: Ts2-net: Token shift and selection transformer for text-video retrieval. In: ECCV. pp. 319-335 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 490, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 481, + 512 + ], + "type": "text", + "content": "30. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "text", + "content": "31. Luo, H., Ji, L., Shi, B., Huang, H., Duan, N., Li, T., Li, J., Bharti, T., Zhou, M.: Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 544, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 481, + 578 + ], + "type": "text", + "content": "32. Ma, Y., Xu, G., Sun, X., Yan, M., Zhang, J., Ji, R.: X-clip: End-to-end multigrained contrastive learning for video-text retrieval. In: ACM MM. pp. 638-647 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "type": "text", + "content": "33. Maddison, C.J., Mnih, A., Teh, Y.W.: The concrete distribution: A continuous relaxation of discrete random variables. CoRR abs/1611.00712 (2016), http://arxiv.org/abs/1611.00712" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "text", + "content": "34. Meng, D., Chen, X., Fan, Z., Zeng, G., Li, H., Yuan, Y., Sun, L., Wang, J.: Conditional detr for fast training convergence. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3651-3660 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "35. Moon, W., Hyun, S., Park, S., Park, D., Heo, J.P.: Query-dependent video representation for moment retrieval and highlight detection. In: Proceedings of the" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "Hannan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "type": "text", + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23023-23033 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 160 + ], + "type": "text", + "content": "36. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "text", + "content": "37. Pan, Y., He, X., Gong, B., Lv, Y., Shen, Y., Peng, Y., Zhao, D.: Scanning only once: An end-to-end framework for fast temporal grounding in long videos. arXiv preprint arXiv:2303.08345 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "38. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. pp. 8748-8763 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "text", + "content": "39. Ramakrishnan, S.K., Al-Halah, Z., Grauman, K.: Naq: Leveraging narrations as queries to supervise episodic memory. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6694-6703 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "40. Soldan, M., Pardo, A., Alcázar, J.L., Caba, F., Zhao, C., Giancola, S., Ghanem, B.: Mad: A scalable dataset for language grounding in videos from movie audio descriptions. In: CVPR. pp. 5026-5035 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "text", + "content": "41. Soldan, M., Xu, M., Qu, S., Tegner, J., Ghanem, B.: Vlg-net: Video-language graph matching network for video grounding. In: ICCV. pp. 3224-3234 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "text", + "content": "42. Wang, A.J., Ge, Y., Yan, R., Ge, Y., Lin, X., Cai, G., Wu, J., Shan, Y., Qie, X., Shou, M.Z.: All in one: Exploring unified video-language pre-training. arXiv preprint arXiv:2203.07303 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "text", + "content": "43. Wang, Q., Zhang, Y., Zheng, Y., Pan, P., Hua, X.S.: Disentangled representation learning for text-video retrieval. arXiv preprint arXiv:2203.07111 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "text", + "content": "44. Wang, Z., Sung, Y.L., Cheng, F., Bertasius, G., Bansal, M.: Unified coarse-to-fine alignment for video-text retrieval. In: The IEEE International Conference on Computer Vision (ICCV) (October 2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "text", + "content": "45. Xu, H., Ghosh, G., Huang, P.Y., Arora, P., Aminzadeh, M., Feichtenhofer, C., Metze, F., Zettlemoyer, L.: Vlm: Task-agnostic video-language model pre-training for video understanding. ACLliu2021hit (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "text", + "content": "46. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettlemoyer, L., Feichtenhofer, C.: Videoclip: Contrastive pre-training for zero-shot video-text understanding. In: EMNLP. pp. 6787-6800 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "type": "text", + "content": "47. Xue, H., Hang, T., Zeng, Y., Sun, Y., Liu, B., Yang, H., Fu, J., Guo, B.: Advancing high-resolution video-language representation with large-scale video transcriptions. In: CVPR. pp. 5036-5045 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "text", + "content": "48. Xue, H., Sun, Y., Liu, B., Fu, J., Song, R., Li, H., Luo, J.: Clip-vip: Adapting pre-trained image-text model to video-language representation alignment. ICLR (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "type": "text", + "content": "49. Yu, Y., Kim, J., Kim, G.: A joint sequence fusion model for video question answering and retrieval. In: ECCV. pp. 471-487 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 555, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 578 + ], + "type": "text", + "content": "50. Yu, Y., Ko, H., Choi, J., Kim, G.: End-to-end concept word detection for video captioning, retrieval, and question answering. In: CVPR. pp. 3165-3173 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 578, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 609 + ], + "type": "text", + "content": "51. Zhang, B., Jin, X., Gong, W., Xu, K., Zhang, Z., Wang, P., Shen, X., Feng, J.: Multimodal video adapter for parameter efficient video text retrieval. arXiv preprint arXiv:2301.07868 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "text", + "content": "52. Zhang, C., Gupta, A., Zisserman, A.: Helping hands: An object-aware ego-centric video recognition model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13901-13912 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "53. Zhang, H., Sun, A., Jing, W., Zhou, J.T.: Span-based localizing network for natural language video localization. arXiv preprint arXiv:2004.13931 (2020)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "text", + "content": "RGNet" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 205 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "54. Zhang, H., Sun, A., Jing, W., Zhou, J.T.: The elements of temporal sentence grounding in videos: A survey and future directions. arXiv preprint arXiv:2201.08071 1(2) (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "text", + "content": "55. Zhang, S., Peng, H., Fu, J., Luo, J.: Learning 2d temporal adjacent networks for moment localization with natural language. In: AAAI. pp. 12870-12877 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "text", + "content": "56. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "Hannan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/RICA^2_ Rubric-Informed, Calibrated Assessment of Actions/805cf405-290f-425f-a265-d9aa43a99eb8_content_list.json b/2024/RICA^2_ Rubric-Informed, Calibrated Assessment of Actions/805cf405-290f-425f-a265-d9aa43a99eb8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f1bcc47f284688396d09843055de250969049998 --- /dev/null +++ b/2024/RICA^2_ Rubric-Informed, Calibrated Assessment of Actions/805cf405-290f-425f-a265-d9aa43a99eb8_content_list.json @@ -0,0 +1,1857 @@ +[ + { + "type": "text", + "text": "RICA $^2$ : Rubric-Informed, Calibrated Assessment of Actions", + "text_level": 1, + "bbox": [ + 290, + 142, + 709, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abrar Majeedi, Viswanatha Reddy Gajjala, Satya Sai Srinath Namburi GNVV, and Yin Li", + "bbox": [ + 233, + 214, + 767, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Wisconsin-Madison, Madison Wisconsin 53706 USA {majeedi,vgajjala,sgnamburi,yin.li}@wisc.edu", + "bbox": [ + 282, + 256, + 717, + 285 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The ability to quantify how well an action is carried out, also known as action quality assessment (AQA), has attracted recent interest in the vision community. Unfortunately, prior methods often ignore the score rubric used by human experts and fall short of quantifying the uncertainty of the model prediction. To bridge the gap, we present RICA² — a deep probabilistic model that integrates score rubric and accounts for prediction uncertainty for AQA. Central to our method lies in stochastic embeddings of action steps, defined on a graph structure that encodes the score rubric. The embeddings spread probabilistic density in the latent space and allow our method to represent model uncertainty. The graph encodes the scoring criteria, based on which the quality scores can be decoded. We demonstrate that our method establishes new state of the art on public benchmarks, including FineDiving, MTL-AQA, and JIGSAWS, with superior performance in score prediction and uncertainty calibration. Our code is available at https://abrarmajeedi.github.io/rica2_aqa/.", + "bbox": [ + 259, + 321, + 738, + 529 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Action Quality Assessment $\\cdot$ Video Understanding", + "bbox": [ + 261, + 542, + 684, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 582, + 374, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Action quality assessment (AQA), aiming at quantifying how well an action is carried out, has been widely studied across scientific disciplines due to its broad range of applications. AQA is key to sports science and analytics. The right way of performing actions maximizes an athlete's performance and minimizes injury risk. AQA is crucial to occupational safety and health. High-quality actions mitigate the physical stress and strain in the workspace. AQA is pivotal for physical therapies. The quality of actions reveals the progress in rehabilitation. AQA also plays a major role in surgical education. Proficient actions improve the outcome and reduce complications.", + "bbox": [ + 212, + 613, + 787, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Observational methods for AQA have been well established for various tasks, e.g., gymnastics [27], manual material handling [41], and surgery [19]. These methods involve a human expert observing an action and decomposing it into a series of key steps. Each of these steps, or a subset of them, can be grouped into a factor and then evaluated using a Likert scale [17] following a pre-defined criterion. Ratings for individual factors, sometimes complemented with impression-based", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/baf7824ac63e33344dcd192a11f8f7d2052e301734f4f126033754c0a29b7fcb.jpg", + "image_caption": [ + "Fig. 1: RICA2 integrates score rubric used by human experts and accounts for prediction uncertainty, resulting in accurate predictions and calibrated uncertainty estimates." + ], + "image_footnote": [], + "bbox": [ + 223, + 143, + 781, + 220 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "global ratings, are then summarized into a final quality score [19]. Multiple expert ratings are often considered to account for the variance in the scores. While these methods are commonly adopted, they require significant input from human experts and are thus costly and inefficient.", + "bbox": [ + 212, + 291, + 784, + 352 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There is a burgeoning interest in the vision community to develop video-based AQA [25, 26, 35, 43]. While current solutions have made steady progress across benchmarks [10, 24, 44], their decision-making processes differ largely from prior observational methods. Almost all prior solutions learn deep models to directly map input videos to scores. Many of them employ an exemplar-based approach, in which a model predicts relative scores by referencing exemplar videos with similar actions and known scores [2, 44, 47]. Few of them have considered the structure of the actions or their scoring criteria used by observational methods.", + "bbox": [ + 212, + 353, + 785, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Further, existing AQA methods face a key challenge in the accurate quantification of model uncertainty i.e., the uncertainty of model prediction that is calibrate to the expected error [12]. Knowing this uncertainty is particularly helpful for AQA, e.g., when assessing the quality of high-stakes competitions or surgical procedures. With proper calibration, videos that have uncertain predictions can be passed to human experts for a thorough evaluation. Several recent works have started to consider the variance among scores from multiple human experts [35,48,50,51]. Unfortunately, they still fall short of considering prediction uncertainty, leaving this challenge largely unaddressed.", + "bbox": [ + 212, + 474, + 787, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To bridge the gap, we develop a deep probabilistic model for AQA by integrating score rubrics and modeling the uncertainty of the prediction (see Fig. 1). Central to our method lies in the stochastic embedding of action steps, defined on a graph structure that encodes the score rubric. The embeddings spread probabilistic density in latent space and allow our method to represent model uncertainty. The graph encodes the scoring criteria, based on which the quality scores can be decoded. We also present a training scheme and describe an approach to estimate uncertainty. Putting things together, our method, dubbed RICA $^2$ (Rubric-informed, Calibrated Assessment of Actions), yields accurate action scores with additional uncertainty estimates.", + "bbox": [ + 212, + 612, + 787, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate RICA² on several public AQA datasets, covering sports and surgical videos. Particularly, RICA² establishes new state of the art on FineDiving [44], MTL-AQA [24] and JIGSAWS [10]. On FineDiving [44] - the largest and most challenging AQA benchmark, RICA² outperforms latest methods in prediction accuracy (a boost of $+0.94\\%$ in Spearman's Rank Correlation Coefficient", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "A. Majeedi et al.", + "bbox": [ + 271, + 114, + 387, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "$(SRCC))$ and demonstrates significantly improved uncertainty calibration (a gain of $+0.178$ in Kendall Tau [13]). Similarly, on MTL-AQA [24], the most commonly used dataset for AQA, $\\mathrm{RICA}^2$ attains state-of-the-art $SRCC$ , and again largely improved calibration (a gain of $+0.444$ in Kendall Tau). On JIGSAWS [10], $\\mathrm{RICA}^2$ beats the previous best results by a relative margin of $+3.37\\%$ in $SRCC$ . Further, we present extensive experiments to evaluate the key design of $\\mathrm{RICA}^2$ .", + "bbox": [ + 212, + 146, + 787, + 237 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our main contributions are summarized into three folds.", + "bbox": [ + 238, + 239, + 661, + 253 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose $RICA^2$ , a novel deep probabilistic method that incorporates scoring rubrics and uncertainty modeling for AQA, resulting in accurate scores and calibrated uncertainty estimates.", + "- Our technical innovations lie in (a) a graph neural network to model the scoring rubric in conjunction with stochastic embeddings on the graph to account for prediction uncertainty and (b) a training scheme under the variational information bottleneck framework.", + "- Our extensive set of experiments demonstrates that $\\mathrm{RICA}^2$ achieves state-of-the-art results in AQA, significantly outperforming prior methods in both prediction accuracy and calibration of uncertainty estimates." + ], + "bbox": [ + 223, + 256, + 784, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 431, + 387, + 448 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Action quality assessment (AQA). Early works in AQA [11,26] employed handcrafted features to estimate quality scores in videos. More recent methods developed various deep models, including convolutional [35,47], graph [23], recurrent [25, 43], and Transformer [2, 44] networks. AQA has also been widely considered in surgical education [18], rehabilitation [28], and ergonomics [5].", + "bbox": [ + 212, + 460, + 787, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, exemplar-based methods [2, 44, 47] have emerged as a promising solution for AQA due to their impressive performance across benchmarks. These methods predict the relative score of an input video by comparing it to selected exemplar videos with similar action steps and known scores. A limitation of this paradigm is the requirement of exemplar videos at inference time. This strategy largely deviates from existing observational methods used by human experts and leads to significantly higher computational costs. While RICA² also uses action steps in the input video, it further integrates the scoring rubric of these steps and offers a solution for no-reference AQA i.e. without using exemplars.", + "bbox": [ + 212, + 537, + 787, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Several recent works have started to consider the modeling of score uncertainty in AQA [35, 48, 50, 51]. For example, Tang et al. [35] proposed to model the final scores using a Gaussian distribution. They presented a model (MUSDL) trained to predict the score distribution. This distribution learning idea was further extended in [48, 50, 51]. However, modeling the score distribution does not warrant the quantification of model uncertainty, as the output distributions might not be calibrated with prediction errors. While RICA² also predicts a Gaussian distribution for the scores, our key design is to consider stochastic embeddings to quantify prediction uncertainty, resulting in calibrated uncertainty estimates.", + "bbox": [ + 212, + 674, + 787, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The most relevant work is IRIS [20]. IRIS incorporates score rubric into a convolutional network for AQA. This is done by segmenting key steps in the", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "$\\mathrm{RICA}^2$ : Rubric-Informed, Calibrated Assessment of Actions", + "bbox": [ + 333, + 113, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "video and predicting sub-scores for individual steps. Similar to IRIS, RICA $^2$ also considers rubric in a deep model. However, RICA $^2$ adapts a graph network, treats sub-scores as latent embeddings, predicts the final score, and further quantifies prediction uncertainty. These differences allow RICA $^2$ to be trained on major public datasets with only final scores, and to output calibrated uncertainty estimates, both of which cannot be achieved by IRIS.", + "bbox": [ + 212, + 145, + 787, + 236 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Modeling uncertainty with stochastic embedding. Stochastic embedding, initially introduced in NLP [21,38], treats each embedding as a distribution. This approach has gained recent attention for modeling uncertainty in deep models. Oh et al. [22] considered probabilistic embeddings for metric learning and proposed to model uncertainty based on the stochasticity of embeddings. This idea was further adopted in many vision tasks, including face verification [32], age estimation [16], pose estimation [34], and cross-modal retrieval [6]. Another related line of work is the conditional variational autoencoder [33], where a probabilistic representation of the input is used for a prediction task. Our approach shares a similar idea of using stochastic embeddings to model uncertainty yet is specifically designed for AQA. Our method significantly extends prior idea to embed action steps on a graph structure, and to propagate these stochastic embeddings on the graph.", + "bbox": [ + 212, + 239, + 789, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Graph neural networks (GNNs). GNNs [8, 15, 30] offer a powerful tool to leverage the relational inductive bias inherent in data [3, 45]. This inductive bias is beneficial to aggregate a global representation from a group of local ones [29]. Recently, Zhou et al. [51] proposed a hierarchical graph convolutional network for AQA, in which a GNN was used for video representation learning. In contrast, we adapt graph networks to model score rubrics used by observational methods.", + "bbox": [ + 212, + 424, + 789, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 AQA with Score Rubric and Uncertainty Modeling", + "text_level": 1, + "bbox": [ + 214, + 537, + 748, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our goal is to assess the quality of an action within an input video. Let $X$ be the video with the action and $Y$ as its quality score. Our method further considers the structure of the action and a scoring rubric based on the structure.", + "bbox": [ + 212, + 566, + 787, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Action steps. We assume that the action in $X$ comprises a known, ordered set of key steps, denoted as $\\mathbb{S} = (s_1, s_2, \\dots, s_k)$ . Each $s^1$ represents a necessary sub-action for successfully executing the action. Further, $s$ is associated with a text description that elucidates the specifics of the corresponding step, e.g., \"a front-facing takeoff\" for diving. This assumption is especially well suited for structured actions, such as diving or surgery, where the key steps are predetermined and follow a specific sequence. Note that the timing of the key steps is not presumed. Even if key steps are unavailable, they can be detected using action recognition methods [4,9] (see supplement Sec. C.4).", + "bbox": [ + 212, + 616, + 787, + 752 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scoring rubric. We further assume a pre-specified scoring rubric based on the key steps — a common strategy in technical skill assessment [19, 27, 41]. Specifically, each action step $s_k$ is independently scored, i.e., $s_k \\mapsto y_k$ . Subsequently, a rule-based rubric is employed to aggregate individual scores $\\{y_k\\}$ and calculate a final", + "bbox": [ + 212, + 755, + 787, + 816 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "A. Majeedi et al.", + "bbox": [ + 271, + 114, + 387, + 128 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1 For the sake of brevity, we omit the subscript as long as there is no confusion.", + "bbox": [ + 217, + 824, + 750, + 839 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d47ba5f9c11f940abc351e7ed272bf2c05b63595941f5eb066769fa5b21f5631.jpg", + "image_caption": [ + "(a) AQA Scoring Rubric" + ], + "image_footnote": [], + "bbox": [ + 230, + 157, + 401, + 300 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8ab449158ac2f05e7496eb8ec143cfd5d9fece82c172f23784d72d9df068b39f.jpg", + "image_caption": [ + "(b) GNN on DAG" + ], + "image_footnote": [], + "bbox": [ + 405, + 155, + 596, + 304 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0159e32da3c3e14823cbfe67fe915c37302b1c0a685906254e47a461e35648f5.jpg", + "image_caption": [ + "(c) Embedding Function $\\pmb{f}$", + "Fig. 2: Overview of RICA $^2$ . Leveraging scoring rubrics (a), RICA $^2$ integrates a graph representation of action step and rubric with uncertainty modeling (b). Specifically, RICA $^2$ takes an input of the video and its key action steps, encodes the input into embeddings (c), refines the embeddings through a deep probabilistic model, and outputs an action score in tandem with its uncertainty estimate." + ], + "image_footnote": [], + "bbox": [ + 596, + 155, + 771, + 304 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "quality score $Y$ , i.e., $\\{y_k\\} \\mapsto Y$ , in which steps might be grouped into intermediate stages (see an example in Fig. 2 (a-b)). This rubric follows a deterministic yet often non-injective mapping, e.g. a many-to-one mapping such as summation.", + "bbox": [ + 212, + 415, + 784, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Method overview. We now present RICA $^2$ — a deep probabilistic model for AQA that leverages known action steps and incorporates the scoring rubric for modeling. Importantly, RICA $^2$ accounts for prediction uncertainty, i.e., when the model prediction can and cannot be trusted. Fig. 2 presents an overview of RICA $^2$ . It consists of two main model components: (a) a graph neural network that integrates the key steps and scoring rubric (Sec. 3.1); and (b) stochastic embeddings defined on the graph to capture prediction uncertainty (Sec. 3.2), coupled with (c) a learning scheme under the variational information bottleneck framework (Sec. 3.3). In what follows, we delve into the details of RICA $^2$ .", + "bbox": [ + 212, + 463, + 787, + 601 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Integrating Actions Steps and Scoring Rubric with Graph", + "text_level": 1, + "bbox": [ + 214, + 619, + 740, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Steps and rubric as graph. We encode action steps and the corresponding scoring rubric with a directed acyclic graph (DAG). This DAG is denoted as $\\mathcal{G} = (\\mathbb{V},\\mathbb{E})$ with $\\mathbb{V}$ as the set of nodes and $\\mathbb{E}$ as the set of directed edges. $\\mathbb{V}$ consists of three types of nodes: (1) the leaf nodes, denoted as $V^s$ , correspond to individual action steps performed in the input video $X$ ; (2) the intermediate nodes capturing possible intermediate stages in the scoring criteria; and (3) a designated root node $V^r$ representing the final score of the action. Further, the edges $\\mathbb{E}$ indicate the scoring rubric, connecting steps (leaf nodes) to stages (intermediate nodes), and stages (intermediate nodes) to the final score (root node). We note that $\\mathcal{G}$ varies for every input video $X$ (assuming a single action), as different steps might be performed. Fig. 2 (a-b) show the example in diving where the key steps and scoring rubric are encoded using our DAG. Additional examples can be found in our supplement Fig. B.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "$\\mathrm{RICA}^2$ : Rubric-Informed, Calibrated Assessment of Actions", + "bbox": [ + 334, + 113, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Learning for quality assessment. Our approach involves a two-step process for quality assessment. First, we employ an embedding function $f$ , designed to map individual steps into a latent space representing action quality. Secondly, we leverage the key step embeddings $\\{Z^s\\}$ along with the score rubric encoded in $\\mathcal{G}$ to learn a scoring function $h$ . These functions $f$ and $h$ are defined as follows:", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nf: X, \\mathcal {G} \\mapsto \\{Z ^ {s} \\}; \\quad h: \\{Z ^ {s} \\}, \\mathcal {G} \\mapsto Y, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 382, + 228, + 787, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\{Z^s\\}$ are the embeddings for the set of steps $\\mathbb{S}$ in $X$ , corresponding to the leaf nodes $\\{V^s\\}$ on the DAG.", + "bbox": [ + 212, + 250, + 787, + 280 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Modeling Score Uncertainty with Stochastic Embeddings", + "text_level": 1, + "bbox": [ + 214, + 299, + 732, + 315 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To model the prediction uncertainty, we adopt stochastic step embeddings defined on the leaf nodes, such that $Z^{s} \\in \\mathbb{R}^{D} \\sim p(Z^{s}|X,\\{V^{s}\\})$ . Unlike deterministic embeddings, where $Z^{s}$ would be a fixed vector, stochastic embedding characterizes the distribution of $Z^{s}$ , allowing for uncertainty control. Specifically, we model $p(Z^{s}|X,V^{s})$ as a Gaussian distribution in $\\mathbb{R}^{D}$ with mean $\\mu^{s}$ and diagonal covariance $\\Sigma^{s}$ . The embedding function $f$ is thus tasked to predict the mean and covariance for the key steps $\\mathbb{S}$ , i.e., $\\{\\mu^s,\\Sigma^s\\} = f(X,\\{V^s\\})$ .", + "bbox": [ + 212, + 320, + 787, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Propagating stochastic embeddings on the graph. Our scoring function $h$ takes the stochastic embeddings $Z^s$ for leaf nodes in $\\mathcal{G}$ (provided by $f$ ), further computes the embeddings for all nodes in $\\mathcal{G}$ , and finally decodes a quality score $Y$ from the embedding $Z^r$ of the root node $V^r$ . To this end, we propose an extension of graph neural networks (GNNs), in which stochastic embeddings $Z^s$ are propagated from leaf nodes $V^s$ to the root node $V^r$ based on the graph structured informed by the scoring rubric of a particular task. Key to this GNN lies in a lightweight MLP that operates on each node, taking as input the embeddings of its direct predecessors, and generating a new embedding that is further propagated to its successors. This scoring function $h$ is thus given by", + "bbox": [ + 212, + 429, + 787, + 582 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\underbrace {Z ^ {s} \\sim \\mathcal {N} \\left(\\mu^ {s} , \\Sigma^ {s}\\right) , \\forall s \\in \\mathbb {S}} _ {\\text {S a m p l i n g f r o m l e a f n o d e s}}; \\quad \\underbrace {Z ^ {\\neg s} = G \\left(\\Sigma_ {V ^ {j} \\in \\mathcal {P} (V ^ {\\neg s})} Z ^ {j}\\right)} _ {\\text {P r o p a g a t i n g o n t h e D A G}}; \\quad \\underbrace {\\hat {Y} = \\mathcal {S} \\left(Z ^ {r}\\right)} _ {\\text {D e o c o d i n g t h e s c o r e}} \\quad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 248, + 585, + 787, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $V^{\\neg s}$ denotes a non-leaf node with its embedding $Z^{\\neg s}$ . $\\mathcal{P}(V^{\\neg s})$ is the set of predecessors of $V^{\\neg s}$ , $G(\\cdot)$ is the MLP aggregating features from predecessors, and $S(\\cdot)$ is another MLP decoding the final score $\\hat{Y}$ from the root node $V^{r}$ .", + "bbox": [ + 212, + 632, + 787, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It is important to note that each leaf embedding $Z^s$ is stochastic, characterized by a Gaussian distribution $(p(Z^s | X, \\mathcal{G}) = \\mathcal{N}(\\mu^s, \\Sigma^s))$ , with parameters predicted by $f$ . The non-leaf embeddings are however deterministic given samples from leaf distributions. This design is motivated by our assumption of the scoring rubric, where uncertainty lies only in assessing action steps and identical individual scores will yield the same final score.", + "bbox": [ + 212, + 679, + 787, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Learning with Variational Information Bottleneck", + "text_level": 1, + "bbox": [ + 214, + 787, + 671, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With stochastic embeddings, training of RICA $^2$ is a challenge. We design a training scheme under the variational information bottleneck framework.", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "A. Majeedi et al.", + "bbox": [ + 271, + 114, + 387, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Variational information bottleneck (VIB). To train our model $p(Y|X, \\mathcal{G})$ with stochastic step embeddings $\\{Z^s\\}$ , we adopt the information bottleneck principle [36], leading to the maximization of the following objective", + "bbox": [ + 212, + 145, + 787, + 191 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nI \\left(\\left\\{Z ^ {s} \\right\\}; Y | \\mathcal {G}\\right) - \\beta I \\left(\\left\\{Z ^ {s} \\right\\}; X | \\mathcal {G}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 200, + 787, + 215 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $I$ is the conditional mutual information, and $\\beta > 0$ controls the tradeoff between the sufficiency of using step embeddings $\\{Z^s\\}$ for predicting $Y$ given $\\mathcal{G}$ , and the size of the embeddings $\\{Z^s\\}$ derived from $X$ and $\\mathcal{G}$ .", + "bbox": [ + 212, + 224, + 785, + 268 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While mutual information is computationally intractable for high dimensional $\\{Z^s\\}$ , a common solution [1] is to assume Markov property ( $p(Z|X,Y,\\mathcal{G}) = p(Z|X,\\mathcal{G})$ ) and conditional independence ( $p(\\{Z^s\\}|X,\\mathcal{G}) = \\prod_s p(Z^s|X,\\mathcal{G})$ ), followed by the variational approximation for a tractable lower bound", + "bbox": [ + 212, + 270, + 787, + 330 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n- \\mathcal {L} _ {\\mathrm {V I B}} = \\mathbb {E} _ {Z ^ {s} \\sim p (Z ^ {s} | X, \\mathcal {G}), \\forall s \\in \\mathbb {S}} \\left[ \\log p (Y | \\{Z ^ {s} \\}, \\mathcal {G}) \\right] - \\beta \\Sigma_ {s \\in \\mathbb {S}} \\operatorname {K L} \\left(p (Z ^ {s} | X, \\mathcal {G}) | | p (Z ^ {s} | \\mathcal {G})\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 338, + 787, + 356 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $p(Y|\\{Z^s\\}, \\mathcal{G})$ is modeled by the scoring function $h$ , KL denotes the Kullback-Leibler divergence, and $p(Z^s|\\mathcal{G})$ is an approximate marginal prior.", + "bbox": [ + 212, + 363, + 787, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "VIB loss. The first term in Eq. (4) defines the log-likelihood of the score given the input. By assuming that output scores follow a Gaussian with a fixed variance $\\sigma$ , this term can be reduced to a mean squared error (MSE) loss", + "bbox": [ + 212, + 397, + 787, + 443 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {M S E} = \\frac {1}{N} \\Sigma_ {i} ^ {N} \\left(\\hat {Y} _ {i} - Y _ {i}\\right) ^ {2} / \\sigma^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 450, + 787, + 476 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $Y_{i}$ is the predicted score for a video indexed by $i$ , $\\hat{Y}_i$ is the corresponding ground-truth score, and $N$ is the total number of videos in the training set.", + "bbox": [ + 212, + 484, + 785, + 516 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The second term in Eq. (4) regularizes the latent space and encodes prediction uncertainty. By assuming a marginal prior of $\\mathcal{N}(0,I)$ for $p(Z^s|\\mathcal{G})$ , we have", + "bbox": [ + 212, + 517, + 785, + 547 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {K L} = \\Sigma_ {s \\in \\mathbb {S}} K L (\\mathcal {N} (\\mu^ {s} (x), \\Sigma^ {s} (x) | | \\mathcal {N} (0, I)) \\\\ = \\frac {1}{2} \\Sigma_ {s \\in \\mathbb {S}} \\Sigma_ {j = 1} ^ {D} \\left(\\left(\\mu_ {j} ^ {s}\\right) ^ {2} + \\left(\\sigma_ {j} ^ {s}\\right) ^ {2} - \\log \\left(\\sigma_ {j} ^ {s}\\right) ^ {2} - 1\\right), \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 556, + 785, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mu_j^s$ and $\\sigma_j^s$ , respectively, are the $j$ -th dimension of the mean ( $\\mu^s(x)$ ) and variance (square root of the diagonal of $\\Sigma^s(x)$ ), for the step $s$ .", + "bbox": [ + 212, + 606, + 785, + 637 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The VIB loss $(\\mathcal{L}_{VIB})$ is thus given by", + "bbox": [ + 238, + 637, + 516, + 652 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {V I B} = \\mathcal {L} _ {M S E} + \\beta \\mathcal {L} _ {K L}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 421, + 662, + 785, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{VIB}$ consists of (a) the MSE loss $\\mathcal{L}_{MSE}$ from the negative log-likelihood of the predicted scores, aiming at minimizing prediction errors; and (b) the KL divergence $\\mathcal{L}_{KL}$ between the predicted Gaussian and the prior, regularizing the stochastic embeddings. Further, the coefficient $\\beta$ balances between two loss terms.", + "bbox": [ + 212, + 685, + 787, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "During training, samples are drawn to compute the loss function. The output is matched to the Gaussian distribution with its mean equal to the average of the judge scores. The reparameterization trick [14] is used to allow the backpropagation of gradients through the sampling process.", + "bbox": [ + 212, + 746, + 787, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Estimating uncertainty. The diagonal covariance $\\Sigma^s (X)$ models the uncertainty of the predicted quality score of a step $s$ for an input video $X$ . A larger", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "$\\mathrm{RICA}^2$ : Rubric-Informed, Calibrated Assessment of Actions", + "bbox": [ + 333, + 113, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "value in its diagonal represents a wider distribution of scores and, hence, a lower confidence in the prediction. Following [16,22] we generate uncertainty scores by summing up the harmonic means of the predicted variances for individual steps", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {u n c e r t a i n t y} (Y) = \\Sigma_ {s \\in \\mathbb {S}} D / \\Sigma_ {j = 1} ^ {D} \\left(\\sigma_ {j} ^ {s}\\right) ^ {- 1}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 200, + 787, + 219 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $D$ is the dimensionality of the stochastic embeddings. Again, $\\sigma_{j}^{s}$ is the $j$ -th dimension of the predicted variance.", + "bbox": [ + 212, + 227, + 785, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Stochastic vs. deterministic modeling. An interesting variant of $\\mathrm{RICA}^2$ is to disable its stochastic component. Conceptually, this is equal to considering step embeddings $Z^r$ as vectors and removing the KL loss $\\mathcal{L}_{KL}$ . We refer to this deterministic version of our model as $\\mathrm{RICA}^{2\\dagger}$ . Without stochastic embeddings, $\\mathrm{RICA}^{2\\dagger}$ is unable to estimate prediction uncertainty, yet often yields slightly lower prediction errors. This trade-off is also observed in prior works [6, 16, 32]. We include this variant of our model in the experiments.", + "bbox": [ + 212, + 261, + 789, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Model Instantiation and Implementation", + "text_level": 1, + "bbox": [ + 214, + 390, + 599, + 406 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Video and step representation. For an input video $X$ , we adapt a pre-trained video backbone(e.g., I3D [4]) to extract its clip-level features, which are further pooled to produce video features $(x_{1}, x_{2}, \\ldots, x_{T})$ with fixed-length $T$ . To represent action steps, we make use of a pre-trained language model [7] (Flan-T5) to extract text features from their step descriptions, resulting in an ordered set of text embeddings $(s_{1}, s_{2}, \\ldots, s_{K})$ for $K$ steps. Note that the language model is not part of RICA². It is used solely to extract embeddings for text descriptions of the action steps (see supplement Tables I-L).", + "bbox": [ + 212, + 420, + 787, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "**Embedding function $f$ .** Our embedding function $f$ is realized using a Transformer model [37] (see Fig. 2(c)). $f$ first processes video features $(x_{1}, x_{2}, \\ldots, x_{T})$ with a self-attention block and text embeddings $(s_{1}, s_{2}, \\ldots, s_{K})$ using a MLP. It further makes use of cross-attention blocks $(2\\mathbf{x})$ to fuse video and text features, where video features are used to compute keys and values, and text embeddings of steps are projected into queries. Further, $f$ decodes stochastic embeddings of individual steps by predicting a mean vector $\\mu^{s} \\in \\mathbb{R}^{D}$ and a diagonal covariance vector $\\Sigma^{s} \\in \\mathbb{R}^{D}$ for each step $s$ .", + "bbox": [ + 212, + 544, + 789, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Scoring function $h$ . With Gaussian distributions for all steps specified by $\\{\\mu^s, \\Sigma^s\\}$ , we encode the steps and score rubric into a video-specific DAG $\\mathcal{G}$ , and realize $h$ as a GNN defined on $\\mathcal{G}$ following Eq. (2). $h$ is parameterized by its aggregation function $G$ , which is shared among nodes of the same type. $G$ is implemented using an averaging operation followed by a MLP (2 layers). Finally, $h$ decodes the final score at the root note $V^r$ of $\\mathcal{G}$ .", + "bbox": [ + 214, + 669, + 787, + 760 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training with auxiliary losses. While the VIB loss (Eq. (7)) is sufficient for training, it falls short of considering the temporal ordering of steps. This is because of the conditional independence assumption needed for the derivation of VIB, i.e., $p(\\{Z^s\\} | X, \\mathcal{G}) = \\prod_s p(Z^s | X, \\mathcal{G})$ , where the ordering of $\\{Z^s\\}$ is discarded. To bridge the gap, we incorporate an auxiliary loss term $\\mathcal{L}_{Aux}$ inspired by [2].", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "A. Majeedi et al.", + "bbox": [ + 271, + 114, + 387, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Specifically, we re-purpose the last cross-attention map $(\\mathbb{R}^{K\\times T})$ from $f$ as a step detector. This is done by computing a temporally-weighted center across the attention of each action step to every video time step (i.e., column-wise). We then enforce that (a) this center is co-located with the peak of the attention along video time steps using a sparsity loss [2]; and (b) all centers follow the temporal ordering of corresponding action steps using a ranking loss [2]. These two terms are summed up as the auxiliary loss, and further added to the VIB loss with a small weight (0.1). In our ablation, we empirically verify that adding the auxiliary loss leads to a minor performance boost.", + "bbox": [ + 212, + 145, + 787, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Inference with sampling. At the inference time, we enhance robustness by sampling 20 times and averaging their predictions to compute the final score.", + "bbox": [ + 212, + 286, + 787, + 316 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments and Results", + "text_level": 1, + "bbox": [ + 214, + 340, + 495, + 358 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. Our evaluations are primarily reported on three publicly available benchmark datasets, namely FineDiving [44], MTL-AQA [24], and JIGSAWS [10] in the main paper. In supplement Sec. B, we also include results on the Cataract-101 [31] with cataract surgery videos.", + "bbox": [ + 212, + 378, + 787, + 439 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation metrics. For all our experiments, we consider metrics on both the accuracy of the prediction and the calibration of the uncertainty estimates.", + "bbox": [ + 212, + 443, + 787, + 472 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "- For accuracy, we use two widely adopted metrics for AQA [2,35,44], namely Spearman's rank correlation (SRCC) and relative L2 distance $(R\\ell_{2})$ . SRCC measures how well the predicted scores are ranked w.r.t. the ground truth, while $R\\ell_{2}$ summarizes the prediction errors. A model with more accurate predictions will have higher SRCC and lower $R\\ell_{2}$ .", + "bbox": [ + 212, + 473, + 787, + 547 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "- For calibration, we report the uncertainty versus error curve following [16, 22]. To plot this curve, test samples are sorted by increasing uncertainty and divided into 10 equal-sized bins. The mean absolute error (MAE) is then computed for items in each bin. We also follow [16, 22] in employing Kendall's tau ( $\\tau$ ) [13], a numerical measure ranging from -1 to 1 to quantify the correlation between the uncertainties and the prediction errors. A higher $\\tau$ indicates better calibration, signifying that a model's uncertainty better aligns with prediction errors.", + "bbox": [ + 212, + 547, + 787, + 654 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Baselines. RICA $^2$ is benchmarked against a set of strong baselines, including exemplar-free methods such as DAE [48], UsDL and MUSDL [35], and exemplar-based ones such as CoRE [47], TPT [2] and TSA [44]. We further include the deterministic version of our model $\\mathrm{RICA}^{2\\dagger}$ , which trades the ability of uncertainty estimation for a minor boost in accuracy. Several baselines adopt a direct regression approach, without providing a confidence or uncertainty measure for predictions. UsDL [35] and TPT [2] implicitly offer a confidence value. In these works, the probability of the predicted score bin serves as a proxy for uncertainty, computed as (1.0-confidence). DAE [48] outputs the standard deviation of the score distribution, which represents uncertainty.", + "bbox": [ + 212, + 657, + 787, + 809 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We seek to ensure a fair comparison yet recognize that methods in our benchmark may consider different settings and/or various types of input. Most", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "$\\mathrm{RICA}^2$ : Rubric-Informed, Calibrated Assessment of Actions", + "bbox": [ + 333, + 113, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "prior exemplar-free methods only consider a video as input. While $\\mathrm{RICA}^2$ does not utilize exemplars, it takes additional input of step information, i.e., step presence and their temporal ordering. On the other hand, previous exemplar-based methods also require the step information as used by $\\mathrm{RICA}^2$ , in addition to an input video and an exemplar database. Notably, step information is used to select exemplars, leading to improved results. For example, for diving videos, CoRE [47], TPT [2] and TSA [44] use the diving number (DN) encoding steps and their ordering. Further, TSA [44] also requires the timing of individual steps during training. While it is infeasible to standardize the settings of all methods, we compare to the best reported results in our experiments.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Results on FineDiving", + "text_level": 1, + "bbox": [ + 215, + 320, + 447, + 335 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Dataset. FineDiving [44] is the largest public dataset for AQA, with 3000 video samples capturing various diving actions. The dataset covers 52 different action types, 29 sub-action types, and 23 difficulty degree types, providing a rich and diverse set of examples for AQA. While this dataset contains temporal annotations for the steps, which can be used to improve the performance of AQA as demonstrated in [44], we do not use these annotations for $\\mathrm{RICA}^2$ .", + "bbox": [ + 212, + 345, + 787, + 436 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Experiment setup. We adhere to the experimental setup of the most recent baseline [44] using their train-test split, with 2251 videos for training and 749 videos for testing. We follow the input video settings used in [44] for $\\mathrm{RICA}^2$ and the baselines. Specifically, for each video, we uniformly sample 96 frames, which are segmented into 9 overlapping clips, each containing 16 consecutive frames. We refer to supplement Sec. A.1 for further implementation details.", + "bbox": [ + 212, + 440, + 787, + 531 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results. Tab. 1a presents our results on FineDiving. Both our stochastic and deterministic versions (RICA $^2$ and RICA $^2\\dagger$ ) outperform the state-of-the-art TPT [2], an exemplar-based method. RICA $^2$ shows a relative margin of $0.7\\%$ / $1.4\\%$ on SRCC / $R\\ell_2$ , and RICA $^2\\dagger$ has a relative margin of $0.9\\%$ / $9.6\\%$ on SRCC / $R\\ell_2$ . This improvement is more pronounced when compared with the exemplar-free methods (MUSDL, DAE-MT) showcasing a significant relative gain of $4.9\\%$ , $1.5\\%$ on SRCC and $29.8\\%$ , $21.6\\%$ on $R\\ell_2$ . While the deterministic RICA $^2\\dagger$ has slightly higher accuracy, our stochastic RICA $^2$ demonstrates superior calibration of its uncertainty estimate ( $\\tau_{\\mathrm{RICA}^2} = 0.64$ vs. $\\tau_{\\mathrm{TPT}} = -0.56$ ). Fig. 3a further shows uncertainty calibration results. Uncertainty estimates from RICA $^2$ have a clear upward trend, indicating a higher calibration level. While MUSDL [35] also exhibits a reasonable level of calibration ( $\\tau_{\\mathrm{MUSDL}} = 0.47$ vs. $\\tau_{\\mathrm{RICA}^2} = 0.64$ ), the errors are significantly higher than RICA $^2$ across all uncertainty levels.", + "bbox": [ + 212, + 534, + 787, + 731 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Results on MTL-AQA", + "text_level": 1, + "bbox": [ + 215, + 753, + 447, + 768 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Dataset. MTL-AQA [24] is one of the most commonly used datasets for AQA. It consists of 1412 samples collected from 16 events with diverse views. The dataset has a rich set of annotations, including the steps performed during the dive, the difficulty score associated with the dive, and the individual judge scores.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "A. Majeedi et al.", + "bbox": [ + 271, + 114, + 387, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/2c7581abd2c1c7f19dfdba5c23dbc5e318ce76eeb0d9b778a151e0585f42050b.jpg", + "table_caption": [ + "Table 1: Main results on (a) FineDiving and (b) MTL-AQA. Prediction accuracy $(SRCC$ and $R\\ell_{2}$ ) and uncertainty calibration $(\\tau)$ metrics are reported. We compare our method with exemplar-based and exemplar-free baselines.", + "(a) Results on FineDiving" + ], + "table_footnote": [], + "table_body": "
Metrics
SRCC(↑)Rl2(↓)τ(↑)
Exemplar basedCoRe [47]0.90610.3615-
TSA [44]0.92030.3420-
TPT [2]0.93330.2877-0.5556
USDL [35,44]0.89130.38220.3778
MUSDL [35,44]0.89780.37040.4667
Exemplar freeDAE [48]0.88200.4919-0.1999
DAE-MT [48]0.92850.3320-0.4667
RICA2 (Ours)0.94020.28380.6444
RICA2† (Ours)0.94210.2600-
", + "bbox": [ + 220, + 214, + 500, + 338 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/42be03937e2de5dc35421a5362962d505335311c3877c55812c14d5a6ed0ced7.jpg", + "table_caption": [ + "(b) Results on MTL-AQA" + ], + "table_footnote": [], + "table_body": "
Metrics
SRCC(†)Rℓ2(↓)τ(↑)
Exemplar basedTSA-Net [40]0.9422--
CoRe [47]0.95120.2600-
DAE-CoRe [48]0.9589--
Exemplar freeTPT [2]0.96070.2378-0.1111
C3D-AVG-MTL [24]0.9044--
USDL [35]0.92310.46800.1556
MUSDL [35]0.92730.4510-0.0667
DAE [48]0.9231--
DAE-MT [48]0.94900.2738-0.4222
RICA2 (Ours)0.95940.25800.6000
RICA2† (Ours)0.96200.2280-
", + "bbox": [ + 517, + 214, + 784, + 340 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e61eb580869c267e53b222e1b379631aec9e94dfaf43a0046247600442b7a6b6.jpg", + "image_caption": [ + "(a) Uncertainty calibration on FineDiving" + ], + "image_footnote": [], + "bbox": [ + 256, + 359, + 478, + 496 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ed416ef7a3939acf35939cfd44ad642e6d94b6e40b82dfd61ed88849b2fc2a6a.jpg", + "image_caption": [ + "(b) Uncertainty calibration on MTL-AQA", + "Fig. 3: Uncertainty vs. prediction error (MAE) on (a) Finediving and (b) MTL-AQA. Results are reported on the test splits, with the X-axis as the uncertainty bin index (uncertainty increases from left to right) and the Y-axis as the MAE in the bin. In comparison to baselines, $\\mathrm{RICA}^2$ has improved calibration with lower prediction errors." + ], + "image_footnote": [], + "bbox": [ + 524, + 359, + 743, + 496 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Experiment setup. We follow the evaluation protocol of [2,43,47], dividing the dataset into the standard train set of 1059 videos and a test set of 353 videos. Further, we use the same input video settings as TPT [2] in our experiments to ensure a fair comparison. Specifically, for each video, we uniformly sample 103 frames segmented into 20 overlapping clips, each containing 8 continuous frames. Please refer to supplement Sec. A.2 for more details.", + "bbox": [ + 212, + 616, + 789, + 707 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results. Tab. 1b summarizes our results on MTL-AQA. Similar to FineDiving, $\\mathrm{RICA}^2$ shows state-of-the-art results on MTL-AQA across all evaluation metrics. Specifically, our $\\mathrm{RICA}^2\\dagger$ outperforms the best exemplar-free model (DAE-MT) by a relative margin of $1.4\\% / 16.7\\%$ on $SRCC / R\\ell_2$ . When compared with the competitive exemplar-based TPT, our has slightly better $SRCC$ ( $SRCC_{TPT} = 0.9607$ vs $SRCC_{\\mathrm{RICA}^2\\dagger} = 0.9620$ ) and $R\\ell_2$ ( $+4.1\\%$ relative margin). Again, compared to previous methods, our stochastic model shows improved calibration ( $\\tau_{\\mathrm{RICA}^2} = 0.60$ vs. $\\tau_{USDL} = 0.16$ vs. $\\tau_{TPT} = -0.11$ ) as shown in Fig. 3b.", + "bbox": [ + 212, + 719, + 787, + 842 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "$\\mathrm{RICA}^2$ : Rubric-Informed, Calibrated Assessment of Actions", + "bbox": [ + 333, + 113, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/b56cd2648cf2f389252b48a6928da16b656825ce9b9e448e26bf02a88aa71a7a.jpg", + "table_caption": [ + "Table 2: Results on JIGSAWS [10] dataset. Only prediction accuracy ( $SRCC$ ) is considered due to the limited sample size. RICA $^2$ outperforms all prior approaches." + ], + "table_footnote": [], + "table_body": "
Task
SNPKTAvg
Exemplar basedCoRe [47]0.840.860.860.85
TPT [2]0.880.880.910.89
ST-GCN [46]0.310.390.580.43
TSN [39]0.340.230.720.46
JRG [23]0.360.540.750.57
Exemplar freeUSDL [35]0.640.630.610.63
MUSDL [35]0.710.690.710.70
DAE [48]0.730.720.720.72
DAE-MT [48]0.780.740.740.76
RICA2† (Ours)0.880.930.880.90
RICA2 (Ours)0.920.940.900.92
", + "bbox": [ + 361, + 184, + 643, + 347 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Results on JIGSAWS", + "text_level": 1, + "bbox": [ + 215, + 372, + 439, + 386 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Dataset. In addition to diving videos, we also evaluate RICA² on JIGSAWS [10] — a robotic surgical video dataset. The dataset includes three tasks: \"Suturing (S),\" with 39 recordings, \"Needle Passing (NP),\" with 26 recordings and \"Knot Tying (KT)\" with 36 recordings. JIGSAWS is widely used for action quality assessment, despite its small scale.", + "bbox": [ + 210, + 393, + 787, + 470 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Experiment setup. Due to the limited number of samples in the dataset (as few as 7 videos in the test set), cross-validation is often considered for evaluation on JIGSAWS. To ensure a fair comparison, we follow the commonly adopted splits from [35], and the input video setting from [2]. Specifically, for each video, we uniformly sample 160 frames which are segmented into 20 non-overlapping clips. We opt to not include score calibration curves due to the limited sample size of the test sets. Additionally, the key steps in JIGSAWS are general motions (e.g. reaching for the needle, orienting the needle, etc.) and thus cannot be localized to any specific section of the video. Thus, we do not use the auxiliary losses for this experiment. More details are described in supplement Sec. A.3.", + "bbox": [ + 212, + 473, + 789, + 626 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Results. Tab. 2 summarizes our results on JIGSAWS. Similar to previous datasets, our models exhibit notable advancements over the previous state-of-the-art model TPT [35], showcasing substantial improvements of $1.1\\%$ (RICA²) to $3.4\\%$ (RICA²†) in terms of average SRCC relative to the exemplar-based state-of-the-art TPT [2]. When compared to the exemplar-free methods, our approach demonstrates an impressive $18.4\\%$ (RICA²) to $21.0\\%$ (RICA²†) relative gain in average SRCC compared to the latest method DAE-MT [48].", + "bbox": [ + 212, + 627, + 787, + 737 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 753, + 401, + 768 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To understand our model design choices, we conduct ablation studies on the MTL-AQA [24] dataset. Additional ablations are in supplement Sec. C.2.", + "bbox": [ + 212, + 776, + 785, + 806 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Experiment setup. To simplify our experiments, we opt for running our ablations using fixed I3D features. This allows us to precisely evaluate the", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "A. Majeedi et al.", + "bbox": [ + 271, + 114, + 387, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/d9f1317b57058c0258b0eb17011964c5b15e7317a8c85a75ed7bfe41ed2501ca.jpg", + "table_caption": [ + "Table 3: Ablation studies of model components on MTL-AQA dataset. * indicates that the text embeddings were frozen during training." + ], + "table_footnote": [], + "table_body": "
Step Rep.DAG (Rubric)KLAuxMetrics
SRCC(↑)Rl2(↓)τ(↑)Avg. Rank (↓)
Random×××0.94260.3882-5.50
Text×××0.94310.3509-4.50
Text✓××0.94300.3336-4.50
Text*✓××0.94370.3335-3.50
Text*✓✓×0.94480.33290.42221.83
Text*✓✓✓0.94600.33030.42221.17
", + "bbox": [ + 263, + 184, + 738, + 281 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "contribution of different components of our model. Specifically, we choose I3D weights from an intermediate checkpoint of our trained model and extract features for all videos with the frozen backbone.", + "bbox": [ + 212, + 308, + 782, + 354 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Base model. Our ablation constructs a base model using randomly initialized step embeddings, an averaging of these embeddings after cross attention with the video features, followed by an MLP for scoring. This base model is trained using only the MSE loss. We then gradually add modules from $\\mathrm{RICA}^2$ and study their effects. Tab. 3 presents our results using the same features and training epochs, with our base model in row 1.", + "bbox": [ + 212, + 357, + 787, + 446 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Text embeddings as step representations. We first replace randomly initialized step representations with the text embeddings of step descriptions. This leads to a major boost in $R\\ell_{2}$ (Tab. 3 row 1 vs. row 2), by leveraging knowledge encoded in the LLM [7]. Further, we find that freezing the text embeddings leads to comparable results and faster convergence (Tab. 3 row 3 vs. row 4).", + "bbox": [ + 212, + 450, + 787, + 527 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Does the scoring rubric help? We also investigate the effects of encoding steps and rubric as a DAG—a key design of our model. Adding the DAG results in a noteworthy boost in $R\\ell_{2}$ (Tab. 3 row 2 vs. row 3). This improvement can be ascribed to the DAG's proficiency in dissecting the action quality across steps.", + "bbox": [ + 212, + 530, + 787, + 593 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effects of loss functions. We now study the loss terms. Our loss function has three terms (a) the MSE loss $(\\mathcal{L}_{MSE})$ to minimize prediction error, (b) the KL loss $(\\mathcal{L}_{KL})$ to regularize the stochastic embeddings, and (c) the auxiliary loss $(\\mathcal{L}_{Aux})$ to ensure temporal ordering of steps. Adding the KL loss $\\mathcal{L}_{KL}$ yields similar results in $SRCC$ and $R\\ell_{2}$ , yet enables calibrated uncertainty estimation. Further attaching the auxiliary loss $\\mathcal{L}_{Aux}$ leads to improvement in both $SRCC$ and $R\\ell_{2}$ , while maintaining the calibration performance.", + "bbox": [ + 212, + 594, + 787, + 700 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Evaluating the cross-attention maps. To gain insight into $\\mathrm{RICA}^2$ , we now examine the cross-attention maps between the step representations and video features in our learned embedding function $f$ . Fig. 4 visualizes the attention map on two test videos on FineDiving. These maps reveal that a step representation is likely to attend to video features during which the step occurs, indicating that $\\mathrm{RICA}^2$ learns to encode the temporal location of individual steps.", + "bbox": [ + 212, + 703, + 787, + 795 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We further evaluate this localization ability following the Pointing game protocol [49], widely considered in weakly supervised / unsupervised localization tasks [42,52]. Pointing Game compares a generated heatmap with an annotated", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "$\\mathrm{RICA}^2$ : Rubric-Informed, Calibrated Assessment of Actions", + "bbox": [ + 333, + 113, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/12bae1831dc4a12c4483acf22d88943f6e48bbff84366f60663f5a0aeb9a87d9.jpg", + "image_caption": [ + "Fig. 4: Visualization of the cross-attention maps. Y-axis : attention value; Y-axis: clip indices (time). Each curve shows an attention map from a step representation to the temporal video features. The frames shown above are aligned with the timing of the corresponding attention plot. Curves and steps are colored accordingly." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 782, + 412 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "time interval and counts the chance of the heatmap's peak falling into the specified interval. Our evaluation focuses on the FineDiving dataset since it is the only dataset providing annotated time intervals for individual steps. When evaluated on the full test set, attention maps from our model attain an accuracy of $61.4\\%$ in the Pointing game protocol, significantly outperforming the chance level accuracy of $30.7\\%$ (given each video has 3.26 steps on average). Note that we did not use any annotated segmentation data for training.", + "bbox": [ + 212, + 513, + 787, + 619 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 215, + 650, + 509, + 666 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we present a deep probabilistic model for action quality assessment in videos. Our key innovation is to integrate score rubrics and to model prediction uncertainty. Specifically, we propose to adapt stochastic embeddings to quantify the uncertainty of individual steps, and to decode action scores using a variant of graph neural network operating on a DAG encoding the score rubric. Our method offers an exemplar-free approach for AQA, achieves new state-of-the-art results in terms of prediction accuracy on public benchmarks, and demonstrates superior calibration of the output uncertainty estimates. We believe that our work provides a solid step towards AQA. We hope that our method and findings can shed light on the challenging problem of trustworthy video recognition.", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "A. Majeedi et al.", + "bbox": [ + 271, + 114, + 387, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement: This work was supported by the UW Madison Office of the Vice Chancellor for Research with funding from the Wisconsin Alumni Research Foundation, by National Science Foundation under Grant No. CNS 2333491, and by the Army Research Lab under contract number W911NF-2020221.", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 229, + 321, + 244 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alemi, A.A., Fischer, I., Dillon, J.V., Murphy, K.: Deep variational information bottleneck. In: International Conference on Learning Representations (2016)", + "2. Bai, Y., Zhou, D., Zhang, S., Wang, J., Ding, E., Guan, Y., Long, Y., Wang, J.: Action quality assessment with temporal parsing transformer. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IV. pp. 422-438. Springer (2022)", + "3. Battaglia, P.W., Hamrick, J.B., Bapst, V., Sanchez-Gonzalez, A., Zambaldi, V., Malinowski, M., Tacchetti, A., Raposo, D., Santoro, A., Faulkner, R., et al.: Relational inductive biases, deep learning, and graph networks. arXiv preprint arXiv:1806.01261 (2018)", + "4. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017)", + "5. Chen, C.H., Hu, Y.H., Yen, T.Y., Radwin, R.G.: Automated video exposure assessment of repetitive hand activity level for a load transfer task. Human factors 55(2), 298-308 (2013)", + "6. Chun, S., Oh, S.J., De Rezende, R.S., Kalantidis, Y., Larlus, D.: Probabilistic embeddings for cross-modal retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8415-8424 (2021)", + "7. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., Webson, A., Gu, S.S., Dai, Z., Suzgun, M., Chen, X., Chowdhery, A., Castro-Ros, A., Pellat, M., Robinson, K., Valter, D., Narang, S., Mishra, G., Yu, A., Zhao, V., Huang, Y., Dai, A., Yu, H., Petrov, S., Chi, E.H., Dean, J., Devlin, J., Roberts, A., Zhou, D., Le, Q.V., Wei, J.: Scaling instructionfinetuned language models. Journal of Machine Learning Research 25(70), 1-53 (2024), http://jmlr.org/papers/v25/23-0870.html", + "8. Duvenaud, D.K., Maclaurin, D., Iparraguirre, J., Bombarell, R., Hirzel, T., Aspuru-Guzik, A., Adams, R.P.: Convolutional networks on graphs for learning molecular fingerprints. Advances in neural information processing systems 28 (2015)", + "9. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 6202-6211 (2019). https://doi.org/10.1109/ICCV.2019.00630", + "10. Gao, Y., Vedula, S.S., Reiley, C.E., Ahmidi, N., Varadarajan, B., Lin, H.C., Tao, L., Zappella, L., Bejar, B., Yuh, D.D., et al.: Jhu-isi gesture and skill assessment working set (jigsaws): A surgical activity dataset for human motion modeling. In: Modeling and Monitoring of Computer Assisted Interventions (M2CAI) - MICCAI Workshop (2014)", + "11. Gordon, A.S.: Automated video assessment of human performance. In: Proceedings of AI-ED. vol. 2 (1995)", + "12. Guo, C., Pleiss, G., Sun, Y., Weinberger, K.Q.: On calibration of modern neural networks. In: International conference on machine learning. pp. 1321-1330. PMLR (2017)" + ], + "bbox": [ + 218, + 258, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "RICA²: Rubric-Informed, Calibrated Assessment of Actions", + "bbox": [ + 334, + 113, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Kendall, M.G.: A new measure of rank correlation. Biometrika 30(1/2), 81-93 (1938)", + "14. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. In: International Conference on Learning Representations (2014)", + "15. Kipf, T.N., Welling, M.: Semi-supervised classification with graph convolutional networks. In: International Conference on Learning Representations (2017), https://openreview.net/forum?id=SJU4ayYgl", + "16. Li, W., Huang, X., Lu, J., Feng, J., Zhou, J.: Learning probabilistic ordinal embeddings for uncertainty-aware regression. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13896-13905 (2021)", + "17. Likert, R.: A technique for the measurement of attitudes. Archives of psychology (1932)", + "18. Liu, D., Li, Q., Jiang, T., Wang, Y., Miao, R., Shan, F., Li, Z.: Towards unified surgical skill assessment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9522-9531 (2021)", + "19. Martin, J., Regehr, G., Reznick, R., Macrae, H., Murnaghan, J., Hutchison, C., Brown, M.: Objective structured assessment of technical skill (osats) for surgical residents. British journal of surgery 84(2), 273-278 (1997)", + "20. Matsuyama, H., Kawaguchi, N., Lim, B.Y.: Iris: Interpretable rubric-informed segmentation for action quality assessment. In: Proceedings of the 28th International Conference on Intelligent User Interfaces. pp. 368-378 (2023)", + "21. Neelakantan, A., Shankar, J., Passos, A., McCallum, A.: Efficient non-parametric estimation of multiple embeddings per word in vector space. In: Moschitti, A., Pang, B., Daelemans, W. (eds.) Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing, EMNLP 2014, October 25-29, 2014, Doha, Qatar, A meeting of SIGDAT, a Special Interest Group of the ACL. pp. 1059-1069. ACL (2014)", + "22. Oh, S.J., Gallagher, A.C., Murphy, K.P., Schroff, F., Pan, J., Roth, J.: Modeling uncertainty with hedged instance embeddings. In: International Conference on Learning Representations (2019), https://openreview.net/forum?id=r1xQQhAqKX", + "23. Pan, J.H., Gao, J., Zheng, W.S.: Action assessment by joint relation graphs. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6331-6340 (2019)", + "24. Parmar, P., Morris, B.T.: What and how well you performed? a multitask learning approach to action quality assessment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 304-313 (2019)", + "25. Parmar, P., Tran Morris, B.: Learning to score olympic events. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops. pp 20-28 (2017)", + "26. Pirsiavash, H., Vondrick, C., Torralba, A.: Assessing the quality of actions. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13. pp. 556-571. Springer (2014)", + "27. Prassas, S., Kwon, Y.H., Sands, W.A.: Biomechanical research in artistic gymnastics: a review. Sports Biomechanics 5(2), 261-291 (2006)", + "28. Qiu, Y., Wang, J., Jin, Z., Chen, H., Zhang, M., Guo, L.: Pose-guided matching based on deep learning for assessing quality of action on rehabilitation training Biomedical Signal Processing and Control 72, 103323 (2022)", + "29. Santoro, A., Raposo, D., Barrett, D.G., Malinowski, M., Pascanu, R., Battaglia, P., Lillicrap, T.: A simple neural network module for relational reasoning. Advances in neural information processing systems 30 (2017)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "A. Majeedi et al.", + "bbox": [ + 271, + 114, + 385, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "30. Scarselli, F., Gori, M., Tsoi, A.C., Hagenbuchner, M., Monfardini, G.: The graph neural network model. IEEE transactions on neural networks 20(1), 61-80 (2008)", + "31. Schoeffmann, K., Taschwer, M., Sarny, S., Munzer, B., Primus, M.J., Putzgruber, D.: Cataract-101: video dataset of 101 cataract surgeries. In: Cesar, P., Zink, M., Murray, N. (eds.) Proceedings of the 9th ACM Multimedia Systems Conference, MMSys 2018, Amsterdam, The Netherlands, June 12-15, 2018. pp. 421-425. ACM (2018)", + "32. Shi, Y., Jain, A.K.: Probabilistic face embeddings. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6902-6911 (2019)", + "33. Sohn, K., Lee, H., Yan, X.: Learning structured output representation using deep conditional generative models. Advances in neural information processing systems 28 (2015)", + "34. Sun, J.J., Zhao, J., Chen, L.C., Schroff, F., Adam, H., Liu, T.: View-invariant probabilistic embedding for human pose. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16. pp. 53-70. Springer (2020)", + "35. Tang, Y., Ni, Z., Zhou, J., Zhang, D., Lu, J., Wu, Y., Zhou, J.: Uncertainty-aware score distribution learning for action quality assessment. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9839-9848 (2020)", + "36. TISHBY, N.: The information bottleneck method. In: Proc. of the 37th Allerton Conference on Communication and Computation, 1999 (1999)", + "37. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)", + "38. Vilnis, L., McCallum, A.: Word representations via gaussian embedding. In: International Conference on Learning Representations (2015)", + "39. Wang, L., Xiong, Y., Wang, Z., Qiao, Y., Lin, D., Tang, X., Van Gool, L.: Temporal segment networks: Towards good practices for deep action recognition. In: European conference on computer vision. pp. 20-36. Springer (2016)", + "40. Wang, S., Yang, D., Zhai, P., Chen, C., Zhang, L.: Tsa-net: Tube self-attention network for action quality assessment. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 4902-4910 (2021)", + "41. Waters, T.R., Putz-Anderson, V., Garg, A.: Applications manual for the revised niosh lifting equation (1994)", + "42. Xiao, F., Sigal, L., Jae Lee, Y.: Weakly-Supervised Visual Grounding of Phrases With Linguistic Structures. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5945-5954 (2017)", + "43. Xu, C., Fu, Y., Zhang, B., Chen, Z., Jiang, Y.G., Xue, X.: Learning to score figure skating sport videos. IEEE transactions on circuits and systems for video technology 30(12), 4578-4590 (2019)", + "44. Xu, J., Rao, Y., Yu, X., Chen, G., Zhou, J., Lu, J.: Finediving: A fine-grained dataset for procedure-aware action quality assessment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2949-2958 (2022)", + "45. Xu, K., Li, J., Zhang, M., Du, S.S., ichi Kawarabayashi, K., Jegelka, S.: What can neural networks reason about? In: International Conference on Learning Representations (2020), https://openreview.net/forum?id=rJxbJeHFPS", + "46. Yan, S., Xiong, Y., Lin, D.: Spatial temporal graph convolutional networks for skeleton-based action recognition. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "RICA²: Rubric-Informed, Calibrated Assessment of Actions", + "bbox": [ + 334, + 113, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "47. Yu, X., Rao, Y., Zhao, W., Lu, J., Zhou, J.: Group-aware contrastive regression for action quality assessment. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7899-7908. IEEE Computer Society, Los Alamitos, CA, USA (oct 2021)", + "48. Zhang, B., Chen, J., Xu, Y., Zhang, H., Yang, X., Geng, X.: Auto-encoding score distribution regression for action quality assessment. Neural Computing and Applications pp. 1-14 (2023)", + "49. Zhang, J., Bargal, S.A., Lin, Z., Brandt, J., Shen, X., Sclaroff, S.: Top-Down Neural Attention by Excitation Backprop. International Journal of Computer Vision 126(10), 1084-1102 (Oct 2018)", + "50. Zhou, C., Huang, Y.: Uncertainty-driven action quality assessment. arXiv preprint arXiv:2207.14513 (2022)", + "51. Zhou, K., Ma, Y., Shum, H.P.H., Liang, X.: Hierarchical graph convolutional networks for action quality assessment. IEEE Transactions on Circuits and Systems for Video Technology pp. 1-1 (2023)", + "52. Zhu, Y., Zhou, Y., Ye, Q., Qiu, Q., Jiao, J.: Soft Proposal Networks for Weakly Supervised Object Localization. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1841-1850 (2017)" + ], + "bbox": [ + 215, + 146, + 785, + 397 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "A. Majeedi et al.", + "bbox": [ + 271, + 114, + 385, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/RICA^2_ Rubric-Informed, Calibrated Assessment of Actions/805cf405-290f-425f-a265-d9aa43a99eb8_model.json b/2024/RICA^2_ Rubric-Informed, Calibrated Assessment of Actions/805cf405-290f-425f-a265-d9aa43a99eb8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..98bc0b6137ebfb4a407359cb1ded51c31f127fb7 --- /dev/null +++ b/2024/RICA^2_ Rubric-Informed, Calibrated Assessment of Actions/805cf405-290f-425f-a265-d9aa43a99eb8_model.json @@ -0,0 +1,2513 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.292, + 0.143, + 0.71, + 0.187 + ], + "angle": 0, + "content": "RICA\\(^2\\): Rubric-Informed, Calibrated Assessment of Actions" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.215, + 0.768, + 0.245 + ], + "angle": 0, + "content": "Abrar Majeedi, Viswanatha Reddy Gajjala, Satya Sai Srinath Namburi GNVV, and Yin Li" + }, + { + "type": "text", + "bbox": [ + 0.284, + 0.257, + 0.718, + 0.286 + ], + "angle": 0, + "content": "University of Wisconsin-Madison, Madison Wisconsin 53706 USA {majeedi,vgajjala,sgnamburi,yin.li}@wisc.edu" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.322, + 0.74, + 0.53 + ], + "angle": 0, + "content": "Abstract. The ability to quantify how well an action is carried out, also known as action quality assessment (AQA), has attracted recent interest in the vision community. Unfortunately, prior methods often ignore the score rubric used by human experts and fall short of quantifying the uncertainty of the model prediction. To bridge the gap, we present RICA² — a deep probabilistic model that integrates score rubric and accounts for prediction uncertainty for AQA. Central to our method lies in stochastic embeddings of action steps, defined on a graph structure that encodes the score rubric. The embeddings spread probabilistic density in the latent space and allow our method to represent model uncertainty. The graph encodes the scoring criteria, based on which the quality scores can be decoded. We demonstrate that our method establishes new state of the art on public benchmarks, including FineDiving, MTL-AQA, and JIGSAWS, with superior performance in score prediction and uncertainty calibration. Our code is available at https://abrarmajeedi.github.io/rica2_aqa/." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.543, + 0.686, + 0.558 + ], + "angle": 0, + "content": "Keywords: Action Quality Assessment \\(\\cdot\\) Video Understanding" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.583, + 0.375, + 0.598 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Action quality assessment (AQA), aiming at quantifying how well an action is carried out, has been widely studied across scientific disciplines due to its broad range of applications. AQA is key to sports science and analytics. The right way of performing actions maximizes an athlete's performance and minimizes injury risk. AQA is crucial to occupational safety and health. High-quality actions mitigate the physical stress and strain in the workspace. AQA is pivotal for physical therapies. The quality of actions reveals the progress in rehabilitation. AQA also plays a major role in surgical education. Proficient actions improve the outcome and reduce complications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Observational methods for AQA have been well established for various tasks, e.g., gymnastics [27], manual material handling [41], and surgery [19]. These methods involve a human expert observing an action and decomposing it into a series of key steps. Each of these steps, or a subset of them, can be grouped into a factor and then evaluated using a Likert scale [17] following a pre-defined criterion. Ratings for individual factors, sometimes complemented with impression-based" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.129 + ], + "angle": 0, + "content": "A. Majeedi et al." + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.144, + 0.782, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.232, + 0.788, + 0.261 + ], + "angle": 0, + "content": "Fig. 1: RICA2 integrates score rubric used by human experts and accounts for prediction uncertainty, resulting in accurate predictions and calibrated uncertainty estimates." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.292, + 0.785, + 0.353 + ], + "angle": 0, + "content": "global ratings, are then summarized into a final quality score [19]. Multiple expert ratings are often considered to account for the variance in the scores. While these methods are commonly adopted, they require significant input from human experts and are thus costly and inefficient." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.354, + 0.787, + 0.475 + ], + "angle": 0, + "content": "There is a burgeoning interest in the vision community to develop video-based AQA [25, 26, 35, 43]. While current solutions have made steady progress across benchmarks [10, 24, 44], their decision-making processes differ largely from prior observational methods. Almost all prior solutions learn deep models to directly map input videos to scores. Many of them employ an exemplar-based approach, in which a model predicts relative scores by referencing exemplar videos with similar actions and known scores [2, 44, 47]. Few of them have considered the structure of the actions or their scoring criteria used by observational methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.788, + 0.612 + ], + "angle": 0, + "content": "Further, existing AQA methods face a key challenge in the accurate quantification of model uncertainty i.e., the uncertainty of model prediction that is calibrate to the expected error [12]. Knowing this uncertainty is particularly helpful for AQA, e.g., when assessing the quality of high-stakes competitions or surgical procedures. With proper calibration, videos that have uncertain predictions can be passed to human experts for a thorough evaluation. Several recent works have started to consider the variance among scores from multiple human experts [35,48,50,51]. Unfortunately, they still fall short of considering prediction uncertainty, leaving this challenge largely unaddressed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.764 + ], + "angle": 0, + "content": "To bridge the gap, we develop a deep probabilistic model for AQA by integrating score rubrics and modeling the uncertainty of the prediction (see Fig. 1). Central to our method lies in the stochastic embedding of action steps, defined on a graph structure that encodes the score rubric. The embeddings spread probabilistic density in latent space and allow our method to represent model uncertainty. The graph encodes the scoring criteria, based on which the quality scores can be decoded. We also present a training scheme and describe an approach to estimate uncertainty. Putting things together, our method, dubbed RICA\\(^2\\) (Rubric-informed, Calibrated Assessment of Actions), yields accurate action scores with additional uncertainty estimates." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We evaluate RICA² on several public AQA datasets, covering sports and surgical videos. Particularly, RICA² establishes new state of the art on FineDiving [44], MTL-AQA [24] and JIGSAWS [10]. On FineDiving [44] - the largest and most challenging AQA benchmark, RICA² outperforms latest methods in prediction accuracy (a boost of \\(+0.94\\%\\) in Spearman's Rank Correlation Coefficient" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.334, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "\\(\\mathrm{RICA}^2\\) : Rubric-Informed, Calibrated Assessment of Actions" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.238 + ], + "angle": 0, + "content": "\\((SRCC))\\) and demonstrates significantly improved uncertainty calibration (a gain of \\(+0.178\\) in Kendall Tau [13]). Similarly, on MTL-AQA [24], the most commonly used dataset for AQA, \\(\\mathrm{RICA}^2\\) attains state-of-the-art \\(SRCC\\), and again largely improved calibration (a gain of \\(+0.444\\) in Kendall Tau). On JIGSAWS [10], \\(\\mathrm{RICA}^2\\) beats the previous best results by a relative margin of \\(+3.37\\%\\) in \\(SRCC\\). Further, we present extensive experiments to evaluate the key design of \\(\\mathrm{RICA}^2\\)." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.241, + 0.663, + 0.255 + ], + "angle": 0, + "content": "Our main contributions are summarized into three folds." + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.257, + 0.784, + 0.302 + ], + "angle": 0, + "content": "- We propose \\( RICA^2 \\), a novel deep probabilistic method that incorporates scoring rubrics and uncertainty modeling for AQA, resulting in accurate scores and calibrated uncertainty estimates." + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.304, + 0.785, + 0.363 + ], + "angle": 0, + "content": "- Our technical innovations lie in (a) a graph neural network to model the scoring rubric in conjunction with stochastic embeddings on the graph to account for prediction uncertainty and (b) a training scheme under the variational information bottleneck framework." + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.365, + 0.785, + 0.412 + ], + "angle": 0, + "content": "- Our extensive set of experiments demonstrates that \\(\\mathrm{RICA}^2\\) achieves state-of-the-art results in AQA, significantly outperforming prior methods in both prediction accuracy and calibration of uncertainty estimates." + }, + { + "type": "list", + "bbox": [ + 0.224, + 0.257, + 0.785, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.433, + 0.388, + 0.449 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.538 + ], + "angle": 0, + "content": "Action quality assessment (AQA). Early works in AQA [11,26] employed handcrafted features to estimate quality scores in videos. More recent methods developed various deep models, including convolutional [35,47], graph [23], recurrent [25, 43], and Transformer [2, 44] networks. AQA has also been widely considered in surgical education [18], rehabilitation [28], and ergonomics [5]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Recently, exemplar-based methods [2, 44, 47] have emerged as a promising solution for AQA due to their impressive performance across benchmarks. These methods predict the relative score of an input video by comparing it to selected exemplar videos with similar action steps and known scores. A limitation of this paradigm is the requirement of exemplar videos at inference time. This strategy largely deviates from existing observational methods used by human experts and leads to significantly higher computational costs. While RICA² also uses action steps in the input video, it further integrates the scoring rubric of these steps and offers a solution for no-reference AQA i.e. without using exemplars." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Several recent works have started to consider the modeling of score uncertainty in AQA [35, 48, 50, 51]. For example, Tang et al. [35] proposed to model the final scores using a Gaussian distribution. They presented a model (MUSDL) trained to predict the score distribution. This distribution learning idea was further extended in [48, 50, 51]. However, modeling the score distribution does not warrant the quantification of model uncertainty, as the output distributions might not be calibrated with prediction errors. While RICA² also predicts a Gaussian distribution for the scores, our key design is to consider stochastic embeddings to quantify prediction uncertainty, resulting in calibrated uncertainty estimates." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "The most relevant work is IRIS [20]. IRIS incorporates score rubric into a convolutional network for AQA. This is done by segmenting key steps in the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.129 + ], + "angle": 0, + "content": "A. Majeedi et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.237 + ], + "angle": 0, + "content": "video and predicting sub-scores for individual steps. Similar to IRIS, RICA\\(^2\\) also considers rubric in a deep model. However, RICA\\(^2\\) adapts a graph network, treats sub-scores as latent embeddings, predicts the final score, and further quantifies prediction uncertainty. These differences allow RICA\\(^2\\) to be trained on major public datasets with only final scores, and to output calibrated uncertainty estimates, both of which cannot be achieved by IRIS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.241, + 0.79, + 0.422 + ], + "angle": 0, + "content": "Modeling uncertainty with stochastic embedding. Stochastic embedding, initially introduced in NLP [21,38], treats each embedding as a distribution. This approach has gained recent attention for modeling uncertainty in deep models. Oh et al. [22] considered probabilistic embeddings for metric learning and proposed to model uncertainty based on the stochasticity of embeddings. This idea was further adopted in many vision tasks, including face verification [32], age estimation [16], pose estimation [34], and cross-modal retrieval [6]. Another related line of work is the conditional variational autoencoder [33], where a probabilistic representation of the input is used for a prediction task. Our approach shares a similar idea of using stochastic embeddings to model uncertainty yet is specifically designed for AQA. Our method significantly extends prior idea to embed action steps on a graph structure, and to propagate these stochastic embeddings on the graph." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.425, + 0.79, + 0.516 + ], + "angle": 0, + "content": "Graph neural networks (GNNs). GNNs [8, 15, 30] offer a powerful tool to leverage the relational inductive bias inherent in data [3, 45]. This inductive bias is beneficial to aggregate a global representation from a group of local ones [29]. Recently, Zhou et al. [51] proposed a hierarchical graph convolutional network for AQA, in which a GNN was used for video representation learning. In contrast, we adapt graph networks to model score rubrics used by observational methods." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.538, + 0.749, + 0.556 + ], + "angle": 0, + "content": "3 AQA with Score Rubric and Uncertainty Modeling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.613 + ], + "angle": 0, + "content": "Our goal is to assess the quality of an action within an input video. Let \\( X \\) be the video with the action and \\( Y \\) as its quality score. Our method further considers the structure of the action and a scoring rubric based on the structure." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.617, + 0.788, + 0.753 + ], + "angle": 0, + "content": "Action steps. We assume that the action in \\( X \\) comprises a known, ordered set of key steps, denoted as \\( \\mathbb{S} = (s_1, s_2, \\dots, s_k) \\). Each \\( s^1 \\) represents a necessary sub-action for successfully executing the action. Further, \\( s \\) is associated with a text description that elucidates the specifics of the corresponding step, e.g., \"a front-facing takeoff\" for diving. This assumption is especially well suited for structured actions, such as diving or surgery, where the key steps are predetermined and follow a specific sequence. Note that the timing of the key steps is not presumed. Even if key steps are unavailable, they can be detected using action recognition methods [4,9] (see supplement Sec. C.4)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.756, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Scoring rubric. We further assume a pre-specified scoring rubric based on the key steps — a common strategy in technical skill assessment [19, 27, 41]. Specifically, each action step \\( s_k \\) is independently scored, i.e., \\( s_k \\mapsto y_k \\). Subsequently, a rule-based rubric is employed to aggregate individual scores \\( \\{y_k\\} \\) and calculate a final" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.751, + 0.84 + ], + "angle": 0, + "content": "1 For the sake of brevity, we omit the subscript as long as there is no confusion." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.335, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "\\(\\mathrm{RICA}^2\\): Rubric-Informed, Calibrated Assessment of Actions" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.145, + 0.377, + 0.156 + ], + "angle": 0, + "content": "(a) AQA Scoring Rubric" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.159, + 0.403, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.46, + 0.145, + 0.543, + 0.155 + ], + "angle": 0, + "content": "(b) GNN on DAG" + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.156, + 0.597, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.622, + 0.145, + 0.744, + 0.156 + ], + "angle": 0, + "content": "(c) Embedding Function \\(\\pmb{f}\\)" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.156, + 0.772, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.317, + 0.788, + 0.388 + ], + "angle": 0, + "content": "Fig. 2: Overview of RICA\\(^2\\). Leveraging scoring rubrics (a), RICA\\(^2\\) integrates a graph representation of action step and rubric with uncertainty modeling (b). Specifically, RICA\\(^2\\) takes an input of the video and its key action steps, encodes the input into embeddings (c), refines the embeddings through a deep probabilistic model, and outputs an action score in tandem with its uncertainty estimate." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.785, + 0.461 + ], + "angle": 0, + "content": "quality score \\( Y \\), i.e., \\( \\{y_k\\} \\mapsto Y \\), in which steps might be grouped into intermediate stages (see an example in Fig. 2 (a-b)). This rubric follows a deterministic yet often non-injective mapping, e.g. a many-to-one mapping such as summation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.788, + 0.602 + ], + "angle": 0, + "content": "Method overview. We now present RICA\\(^2\\) — a deep probabilistic model for AQA that leverages known action steps and incorporates the scoring rubric for modeling. Importantly, RICA\\(^2\\) accounts for prediction uncertainty, i.e., when the model prediction can and cannot be trusted. Fig. 2 presents an overview of RICA\\(^2\\). It consists of two main model components: (a) a graph neural network that integrates the key steps and scoring rubric (Sec. 3.1); and (b) stochastic embeddings defined on the graph to capture prediction uncertainty (Sec. 3.2), coupled with (c) a learning scheme under the variational information bottleneck framework (Sec. 3.3). In what follows, we delve into the details of RICA\\(^2\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.621, + 0.741, + 0.637 + ], + "angle": 0, + "content": "3.1 Integrating Actions Steps and Scoring Rubric with Graph" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Steps and rubric as graph. We encode action steps and the corresponding scoring rubric with a directed acyclic graph (DAG). This DAG is denoted as \\(\\mathcal{G} = (\\mathbb{V},\\mathbb{E})\\) with \\(\\mathbb{V}\\) as the set of nodes and \\(\\mathbb{E}\\) as the set of directed edges. \\(\\mathbb{V}\\) consists of three types of nodes: (1) the leaf nodes, denoted as \\(V^s\\), correspond to individual action steps performed in the input video \\(X\\); (2) the intermediate nodes capturing possible intermediate stages in the scoring criteria; and (3) a designated root node \\(V^r\\) representing the final score of the action. Further, the edges \\(\\mathbb{E}\\) indicate the scoring rubric, connecting steps (leaf nodes) to stages (intermediate nodes), and stages (intermediate nodes) to the final score (root node). We note that \\(\\mathcal{G}\\) varies for every input video \\(X\\) (assuming a single action), as different steps might be performed. Fig. 2 (a-b) show the example in diving where the key steps and scoring rubric are encoded using our DAG. Additional examples can be found in our supplement Fig. B." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.129 + ], + "angle": 0, + "content": "A. Majeedi et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "Learning for quality assessment. Our approach involves a two-step process for quality assessment. First, we employ an embedding function \\( f \\), designed to map individual steps into a latent space representing action quality. Secondly, we leverage the key step embeddings \\( \\{Z^s\\} \\) along with the score rubric encoded in \\( \\mathcal{G} \\) to learn a scoring function \\( h \\). These functions \\( f \\) and \\( h \\) are defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.383, + 0.229, + 0.788, + 0.246 + ], + "angle": 0, + "content": "\\[\nf: X, \\mathcal {G} \\mapsto \\{Z ^ {s} \\}; \\quad h: \\{Z ^ {s} \\}, \\mathcal {G} \\mapsto Y, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.25, + 0.788, + 0.281 + ], + "angle": 0, + "content": "where \\(\\{Z^s\\}\\) are the embeddings for the set of steps \\(\\mathbb{S}\\) in \\(X\\), corresponding to the leaf nodes \\(\\{V^s\\}\\) on the DAG." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.3, + 0.733, + 0.316 + ], + "angle": 0, + "content": "3.2 Modeling Score Uncertainty with Stochastic Embeddings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.321, + 0.789, + 0.429 + ], + "angle": 0, + "content": "To model the prediction uncertainty, we adopt stochastic step embeddings defined on the leaf nodes, such that \\( Z^{s} \\in \\mathbb{R}^{D} \\sim p(Z^{s}|X,\\{V^{s}\\}) \\). Unlike deterministic embeddings, where \\( Z^{s} \\) would be a fixed vector, stochastic embedding characterizes the distribution of \\( Z^{s} \\), allowing for uncertainty control. Specifically, we model \\( p(Z^{s}|X,V^{s}) \\) as a Gaussian distribution in \\( \\mathbb{R}^{D} \\) with mean \\( \\mu^{s} \\) and diagonal covariance \\( \\Sigma^{s} \\). The embedding function \\( f \\) is thus tasked to predict the mean and covariance for the key steps \\( \\mathbb{S} \\), i.e., \\( \\{\\mu^s,\\Sigma^s\\} = f(X,\\{V^s\\}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.789, + 0.583 + ], + "angle": 0, + "content": "Propagating stochastic embeddings on the graph. Our scoring function \\( h \\) takes the stochastic embeddings \\( Z^s \\) for leaf nodes in \\( \\mathcal{G} \\) (provided by \\( f \\)), further computes the embeddings for all nodes in \\( \\mathcal{G} \\), and finally decodes a quality score \\( Y \\) from the embedding \\( Z^r \\) of the root node \\( V^r \\). To this end, we propose an extension of graph neural networks (GNNs), in which stochastic embeddings \\( Z^s \\) are propagated from leaf nodes \\( V^s \\) to the root node \\( V^r \\) based on the graph structured informed by the scoring rubric of a particular task. Key to this GNN lies in a lightweight MLP that operates on each node, taking as input the embeddings of its direct predecessors, and generating a new embedding that is further propagated to its successors. This scoring function \\( h \\) is thus given by" + }, + { + "type": "equation", + "bbox": [ + 0.249, + 0.587, + 0.788, + 0.628 + ], + "angle": 0, + "content": "\\[\n\\underbrace {Z ^ {s} \\sim \\mathcal {N} \\left(\\mu^ {s} , \\Sigma^ {s}\\right) , \\forall s \\in \\mathbb {S}} _ {\\text {S a m p l i n g f r o m l e a f n o d e s}}; \\quad \\underbrace {Z ^ {\\neg s} = G \\left(\\Sigma_ {V ^ {j} \\in \\mathcal {P} (V ^ {\\neg s})} Z ^ {j}\\right)} _ {\\text {P r o p a g a t i n g o n t h e D A G}}; \\quad \\underbrace {\\hat {Y} = \\mathcal {S} \\left(Z ^ {r}\\right)} _ {\\text {D e o c o d i n g t h e s c o r e}} \\quad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.633, + 0.789, + 0.679 + ], + "angle": 0, + "content": "where \\( V^{\\neg s} \\) denotes a non-leaf node with its embedding \\( Z^{\\neg s} \\). \\( \\mathcal{P}(V^{\\neg s}) \\) is the set of predecessors of \\( V^{\\neg s} \\), \\( G(\\cdot) \\) is the MLP aggregating features from predecessors, and \\( S(\\cdot) \\) is another MLP decoding the final score \\( \\hat{Y} \\) from the root node \\( V^{r} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.68, + 0.789, + 0.77 + ], + "angle": 0, + "content": "It is important to note that each leaf embedding \\( Z^s \\) is stochastic, characterized by a Gaussian distribution \\( (p(Z^s | X, \\mathcal{G}) = \\mathcal{N}(\\mu^s, \\Sigma^s)) \\), with parameters predicted by \\( f \\). The non-leaf embeddings are however deterministic given samples from leaf distributions. This design is motivated by our assumption of the scoring rubric, where uncertainty lies only in assessing action steps and identical individual scores will yield the same final score." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.789, + 0.673, + 0.804 + ], + "angle": 0, + "content": "3.3 Learning with Variational Information Bottleneck" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "With stochastic embeddings, training of RICA\\(^2\\) is a challenge. We design a training scheme under the variational information bottleneck framework." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.334, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "\\(\\mathrm{RICA}^2\\) : Rubric-Informed, Calibrated Assessment of Actions" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.193 + ], + "angle": 0, + "content": "Variational information bottleneck (VIB). To train our model \\( p(Y|X, \\mathcal{G}) \\) with stochastic step embeddings \\( \\{Z^s\\} \\), we adopt the information bottleneck principle [36], leading to the maximization of the following objective" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.201, + 0.788, + 0.217 + ], + "angle": 0, + "content": "\\[\nI \\left(\\left\\{Z ^ {s} \\right\\}; Y | \\mathcal {G}\\right) - \\beta I \\left(\\left\\{Z ^ {s} \\right\\}; X | \\mathcal {G}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.225, + 0.787, + 0.27 + ], + "angle": 0, + "content": "where \\(I\\) is the conditional mutual information, and \\(\\beta > 0\\) controls the tradeoff between the sufficiency of using step embeddings \\(\\{Z^s\\}\\) for predicting \\(Y\\) given \\(\\mathcal{G}\\), and the size of the embeddings \\(\\{Z^s\\}\\) derived from \\(X\\) and \\(\\mathcal{G}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.271, + 0.788, + 0.332 + ], + "angle": 0, + "content": "While mutual information is computationally intractable for high dimensional \\(\\{Z^s\\}\\), a common solution [1] is to assume Markov property (\\(p(Z|X,Y,\\mathcal{G}) = p(Z|X,\\mathcal{G})\\)) and conditional independence (\\(p(\\{Z^s\\}|X,\\mathcal{G}) = \\prod_s p(Z^s|X,\\mathcal{G})\\)), followed by the variational approximation for a tractable lower bound" + }, + { + "type": "equation", + "bbox": [ + 0.226, + 0.339, + 0.788, + 0.357 + ], + "angle": 0, + "content": "\\[\n- \\mathcal {L} _ {\\mathrm {V I B}} = \\mathbb {E} _ {Z ^ {s} \\sim p (Z ^ {s} | X, \\mathcal {G}), \\forall s \\in \\mathbb {S}} \\left[ \\log p (Y | \\{Z ^ {s} \\}, \\mathcal {G}) \\right] - \\beta \\Sigma_ {s \\in \\mathbb {S}} \\operatorname {K L} \\left(p (Z ^ {s} | X, \\mathcal {G}) | | p (Z ^ {s} | \\mathcal {G})\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.364, + 0.788, + 0.395 + ], + "angle": 0, + "content": "where \\(p(Y|\\{Z^s\\}, \\mathcal{G})\\) is modeled by the scoring function \\(h\\), KL denotes the Kullback-Leibler divergence, and \\(p(Z^s|\\mathcal{G})\\) is an approximate marginal prior." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.788, + 0.444 + ], + "angle": 0, + "content": "VIB loss. The first term in Eq. (4) defines the log-likelihood of the score given the input. By assuming that output scores follow a Gaussian with a fixed variance \\(\\sigma\\), this term can be reduced to a mean squared error (MSE) loss" + }, + { + "type": "equation", + "bbox": [ + 0.404, + 0.452, + 0.788, + 0.477 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {M S E} = \\frac {1}{N} \\Sigma_ {i} ^ {N} \\left(\\hat {Y} _ {i} - Y _ {i}\\right) ^ {2} / \\sigma^ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.787, + 0.517 + ], + "angle": 0, + "content": "where \\( Y_{i} \\) is the predicted score for a video indexed by \\( i \\), \\( \\hat{Y}_i \\) is the corresponding ground-truth score, and \\( N \\) is the total number of videos in the training set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.787, + 0.549 + ], + "angle": 0, + "content": "The second term in Eq. (4) regularizes the latent space and encodes prediction uncertainty. By assuming a marginal prior of \\(\\mathcal{N}(0,I)\\) for \\(p(Z^s|\\mathcal{G})\\), we have" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.557, + 0.787, + 0.6 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {K L} = \\Sigma_ {s \\in \\mathbb {S}} K L (\\mathcal {N} (\\mu^ {s} (x), \\Sigma^ {s} (x) | | \\mathcal {N} (0, I)) \\\\ = \\frac {1}{2} \\Sigma_ {s \\in \\mathbb {S}} \\Sigma_ {j = 1} ^ {D} \\left(\\left(\\mu_ {j} ^ {s}\\right) ^ {2} + \\left(\\sigma_ {j} ^ {s}\\right) ^ {2} - \\log \\left(\\sigma_ {j} ^ {s}\\right) ^ {2} - 1\\right), \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.607, + 0.787, + 0.638 + ], + "angle": 0, + "content": "where \\(\\mu_j^s\\) and \\(\\sigma_j^s\\), respectively, are the \\(j\\)-th dimension of the mean (\\(\\mu^s(x)\\)) and variance (square root of the diagonal of \\(\\Sigma^s(x)\\)), for the step \\(s\\)." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.638, + 0.517, + 0.653 + ], + "angle": 0, + "content": "The VIB loss \\((\\mathcal{L}_{VIB})\\) is thus given by" + }, + { + "type": "equation", + "bbox": [ + 0.422, + 0.663, + 0.787, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {V I B} = \\mathcal {L} _ {M S E} + \\beta \\mathcal {L} _ {K L}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.686, + 0.788, + 0.746 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{VIB}\\) consists of (a) the MSE loss \\(\\mathcal{L}_{MSE}\\) from the negative log-likelihood of the predicted scores, aiming at minimizing prediction errors; and (b) the KL divergence \\(\\mathcal{L}_{KL}\\) between the predicted Gaussian and the prior, regularizing the stochastic embeddings. Further, the coefficient \\(\\beta\\) balances between two loss terms." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.747, + 0.788, + 0.807 + ], + "angle": 0, + "content": "During training, samples are drawn to compute the loss function. The output is matched to the Gaussian distribution with its mean equal to the average of the judge scores. The reparameterization trick [14] is used to allow the backpropagation of gradients through the sampling process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Estimating uncertainty. The diagonal covariance \\(\\Sigma^s (X)\\) models the uncertainty of the predicted quality score of a step \\(s\\) for an input video \\(X\\). A larger" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.129 + ], + "angle": 0, + "content": "A. Majeedi et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "value in its diagonal represents a wider distribution of scores and, hence, a lower confidence in the prediction. Following [16,22] we generate uncertainty scores by summing up the harmonic means of the predicted variances for individual steps" + }, + { + "type": "equation", + "bbox": [ + 0.372, + 0.201, + 0.788, + 0.22 + ], + "angle": 0, + "content": "\\[\n\\operatorname {u n c e r t a i n t y} (Y) = \\Sigma_ {s \\in \\mathbb {S}} D / \\Sigma_ {j = 1} ^ {D} \\left(\\sigma_ {j} ^ {s}\\right) ^ {- 1}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.228, + 0.787, + 0.259 + ], + "angle": 0, + "content": "where \\(D\\) is the dimensionality of the stochastic embeddings. Again, \\(\\sigma_{j}^{s}\\) is the \\(j\\)-th dimension of the predicted variance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.262, + 0.79, + 0.368 + ], + "angle": 0, + "content": "Stochastic vs. deterministic modeling. An interesting variant of \\(\\mathrm{RICA}^2\\) is to disable its stochastic component. Conceptually, this is equal to considering step embeddings \\(Z^r\\) as vectors and removing the KL loss \\(\\mathcal{L}_{KL}\\). We refer to this deterministic version of our model as \\(\\mathrm{RICA}^{2\\dagger}\\). Without stochastic embeddings, \\(\\mathrm{RICA}^{2\\dagger}\\) is unable to estimate prediction uncertainty, yet often yields slightly lower prediction errors. This trade-off is also observed in prior works [6, 16, 32]. We include this variant of our model in the experiments." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.391, + 0.6, + 0.407 + ], + "angle": 0, + "content": "3.4 Model Instantiation and Implementation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.789, + 0.542 + ], + "angle": 0, + "content": "Video and step representation. For an input video \\(X\\), we adapt a pre-trained video backbone(e.g., I3D [4]) to extract its clip-level features, which are further pooled to produce video features \\((x_{1}, x_{2}, \\ldots, x_{T})\\) with fixed-length \\(T\\). To represent action steps, we make use of a pre-trained language model [7] (Flan-T5) to extract text features from their step descriptions, resulting in an ordered set of text embeddings \\((s_{1}, s_{2}, \\ldots, s_{K})\\) for \\(K\\) steps. Note that the language model is not part of RICA². It is used solely to extract embeddings for text descriptions of the action steps (see supplement Tables I-L)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.545, + 0.79, + 0.667 + ], + "angle": 0, + "content": "**Embedding function \\( f \\).** Our embedding function \\( f \\) is realized using a Transformer model [37] (see Fig. 2(c)). \\( f \\) first processes video features \\( (x_{1}, x_{2}, \\ldots, x_{T}) \\) with a self-attention block and text embeddings \\( (s_{1}, s_{2}, \\ldots, s_{K}) \\) using a MLP. It further makes use of cross-attention blocks \\( (2\\mathbf{x}) \\) to fuse video and text features, where video features are used to compute keys and values, and text embeddings of steps are projected into queries. Further, \\( f \\) decodes stochastic embeddings of individual steps by predicting a mean vector \\( \\mu^{s} \\in \\mathbb{R}^{D} \\) and a diagonal covariance vector \\( \\Sigma^{s} \\in \\mathbb{R}^{D} \\) for each step \\( s \\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.67, + 0.789, + 0.761 + ], + "angle": 0, + "content": "Scoring function \\( h \\). With Gaussian distributions for all steps specified by \\( \\{\\mu^s, \\Sigma^s\\} \\), we encode the steps and score rubric into a video-specific DAG \\( \\mathcal{G} \\), and realize \\( h \\) as a GNN defined on \\( \\mathcal{G} \\) following Eq. (2). \\( h \\) is parameterized by its aggregation function \\( G \\), which is shared among nodes of the same type. \\( G \\) is implemented using an averaging operation followed by a MLP (2 layers). Finally, \\( h \\) decodes the final score at the root note \\( V^r \\) of \\( \\mathcal{G} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Training with auxiliary losses. While the VIB loss (Eq. (7)) is sufficient for training, it falls short of considering the temporal ordering of steps. This is because of the conditional independence assumption needed for the derivation of VIB, i.e., \\( p(\\{Z^s\\} | X, \\mathcal{G}) = \\prod_s p(Z^s | X, \\mathcal{G}) \\), where the ordering of \\( \\{Z^s\\} \\) is discarded. To bridge the gap, we incorporate an auxiliary loss term \\( \\mathcal{L}_{Aux} \\) inspired by [2]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.334, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "\\(\\mathrm{RICA}^2\\) : Rubric-Informed, Calibrated Assessment of Actions" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.284 + ], + "angle": 0, + "content": "Specifically, we re-purpose the last cross-attention map \\((\\mathbb{R}^{K\\times T})\\) from \\(f\\) as a step detector. This is done by computing a temporally-weighted center across the attention of each action step to every video time step (i.e., column-wise). We then enforce that (a) this center is co-located with the peak of the attention along video time steps using a sparsity loss [2]; and (b) all centers follow the temporal ordering of corresponding action steps using a ranking loss [2]. These two terms are summed up as the auxiliary loss, and further added to the VIB loss with a small weight (0.1). In our ablation, we empirically verify that adding the auxiliary loss leads to a minor performance boost." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.287, + 0.788, + 0.318 + ], + "angle": 0, + "content": "Inference with sampling. At the inference time, we enhance robustness by sampling 20 times and averaging their predictions to compute the final score." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.342, + 0.496, + 0.359 + ], + "angle": 0, + "content": "4 Experiments and Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.379, + 0.788, + 0.44 + ], + "angle": 0, + "content": "Datasets. Our evaluations are primarily reported on three publicly available benchmark datasets, namely FineDiving [44], MTL-AQA [24], and JIGSAWS [10] in the main paper. In supplement Sec. B, we also include results on the Cataract-101 [31] with cataract surgery videos." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.788, + 0.473 + ], + "angle": 0, + "content": "Evaluation metrics. For all our experiments, we consider metrics on both the accuracy of the prediction and the calibration of the uncertainty estimates." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.788, + 0.548 + ], + "angle": 0, + "content": "- For accuracy, we use two widely adopted metrics for AQA [2,35,44], namely Spearman's rank correlation (SRCC) and relative L2 distance \\((R\\ell_{2})\\). SRCC measures how well the predicted scores are ranked w.r.t. the ground truth, while \\(R\\ell_{2}\\) summarizes the prediction errors. A model with more accurate predictions will have higher SRCC and lower \\(R\\ell_{2}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.788, + 0.655 + ], + "angle": 0, + "content": "- For calibration, we report the uncertainty versus error curve following [16, 22]. To plot this curve, test samples are sorted by increasing uncertainty and divided into 10 equal-sized bins. The mean absolute error (MAE) is then computed for items in each bin. We also follow [16, 22] in employing Kendall's tau (\\(\\tau\\)) [13], a numerical measure ranging from -1 to 1 to quantify the correlation between the uncertainties and the prediction errors. A higher \\(\\tau\\) indicates better calibration, signifying that a model's uncertainty better aligns with prediction errors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Baselines. RICA\\(^2\\) is benchmarked against a set of strong baselines, including exemplar-free methods such as DAE [48], UsDL and MUSDL [35], and exemplar-based ones such as CoRE [47], TPT [2] and TSA [44]. We further include the deterministic version of our model \\(\\mathrm{RICA}^{2\\dagger}\\), which trades the ability of uncertainty estimation for a minor boost in accuracy. Several baselines adopt a direct regression approach, without providing a confidence or uncertainty measure for predictions. UsDL [35] and TPT [2] implicitly offer a confidence value. In these works, the probability of the predicted score bin serves as a proxy for uncertainty, computed as (1.0-confidence). DAE [48] outputs the standard deviation of the score distribution, which represents uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We seek to ensure a fair comparison yet recognize that methods in our benchmark may consider different settings and/or various types of input. Most" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.129 + ], + "angle": 0, + "content": "A. Majeedi et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "prior exemplar-free methods only consider a video as input. While \\(\\mathrm{RICA}^2\\) does not utilize exemplars, it takes additional input of step information, i.e., step presence and their temporal ordering. On the other hand, previous exemplar-based methods also require the step information as used by \\(\\mathrm{RICA}^2\\), in addition to an input video and an exemplar database. Notably, step information is used to select exemplars, leading to improved results. For example, for diving videos, CoRE [47], TPT [2] and TSA [44] use the diving number (DN) encoding steps and their ordering. Further, TSA [44] also requires the timing of individual steps during training. While it is infeasible to standardize the settings of all methods, we compare to the best reported results in our experiments." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.321, + 0.449, + 0.337 + ], + "angle": 0, + "content": "4.1 Results on FineDiving" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.346, + 0.788, + 0.437 + ], + "angle": 0, + "content": "Dataset. FineDiving [44] is the largest public dataset for AQA, with 3000 video samples capturing various diving actions. The dataset covers 52 different action types, 29 sub-action types, and 23 difficulty degree types, providing a rich and diverse set of examples for AQA. While this dataset contains temporal annotations for the steps, which can be used to improve the performance of AQA as demonstrated in [44], we do not use these annotations for \\(\\mathrm{RICA}^2\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.441, + 0.788, + 0.532 + ], + "angle": 0, + "content": "Experiment setup. We adhere to the experimental setup of the most recent baseline [44] using their train-test split, with 2251 videos for training and 749 videos for testing. We follow the input video settings used in [44] for \\(\\mathrm{RICA}^2\\) and the baselines. Specifically, for each video, we uniformly sample 96 frames, which are segmented into 9 overlapping clips, each containing 16 consecutive frames. We refer to supplement Sec. A.1 for further implementation details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.789, + 0.732 + ], + "angle": 0, + "content": "Results. Tab. 1a presents our results on FineDiving. Both our stochastic and deterministic versions (RICA\\(^2\\) and RICA\\(^2\\dagger\\)) outperform the state-of-the-art TPT [2], an exemplar-based method. RICA\\(^2\\) shows a relative margin of \\(0.7\\%\\) / \\(1.4\\%\\) on SRCC / \\(R\\ell_2\\), and RICA\\(^2\\dagger\\) has a relative margin of \\(0.9\\%\\) / \\(9.6\\%\\) on SRCC / \\(R\\ell_2\\). This improvement is more pronounced when compared with the exemplar-free methods (MUSDL, DAE-MT) showcasing a significant relative gain of \\(4.9\\%\\), \\(1.5\\%\\) on SRCC and \\(29.8\\%\\), \\(21.6\\%\\) on \\(R\\ell_2\\). While the deterministic RICA\\(^2\\dagger\\) has slightly higher accuracy, our stochastic RICA\\(^2\\) demonstrates superior calibration of its uncertainty estimate (\\(\\tau_{\\mathrm{RICA}^2} = 0.64\\) vs. \\(\\tau_{\\mathrm{TPT}} = -0.56\\)). Fig. 3a further shows uncertainty calibration results. Uncertainty estimates from RICA\\(^2\\) have a clear upward trend, indicating a higher calibration level. While MUSDL [35] also exhibits a reasonable level of calibration (\\(\\tau_{\\mathrm{MUSDL}} = 0.47\\) vs. \\(\\tau_{\\mathrm{RICA}^2} = 0.64\\)), the errors are significantly higher than RICA\\(^2\\) across all uncertainty levels." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.755, + 0.449, + 0.77 + ], + "angle": 0, + "content": "4.2 Results on MTL-AQA" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Dataset. MTL-AQA [24] is one of the most commonly used datasets for AQA. It consists of 1412 samples collected from 16 events with diverse views. The dataset has a rich set of annotations, including the steps performed during the dive, the difficulty score associated with the dive, and the individual judge scores." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.334, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "\\(\\mathrm{RICA}^2\\) : Rubric-Informed, Calibrated Assessment of Actions" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.158, + 0.788, + 0.2 + ], + "angle": 0, + "content": "Table 1: Main results on (a) FineDiving and (b) MTL-AQA. Prediction accuracy \\((SRCC\\) and \\(R\\ell_{2}\\)) and uncertainty calibration \\((\\tau)\\) metrics are reported. We compare our method with exemplar-based and exemplar-free baselines." + }, + { + "type": "table_caption", + "bbox": [ + 0.28, + 0.203, + 0.437, + 0.215 + ], + "angle": 0, + "content": "(a) Results on FineDiving" + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.215, + 0.501, + 0.339 + ], + "angle": 0, + "content": "
Metrics
SRCC(↑)Rl2(↓)τ(↑)
Exemplar basedCoRe [47]0.90610.3615-
TSA [44]0.92030.3420-
TPT [2]0.93330.2877-0.5556
USDL [35,44]0.89130.38220.3778
MUSDL [35,44]0.89780.37040.4667
Exemplar freeDAE [48]0.88200.4919-0.1999
DAE-MT [48]0.92850.3320-0.4667
RICA2 (Ours)0.94020.28380.6444
RICA2† (Ours)0.94210.2600-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.574, + 0.203, + 0.728, + 0.215 + ], + "angle": 0, + "content": "(b) Results on MTL-AQA" + }, + { + "type": "table", + "bbox": [ + 0.519, + 0.215, + 0.785, + 0.342 + ], + "angle": 0, + "content": "
Metrics
SRCC(†)Rℓ2(↓)τ(↑)
Exemplar basedTSA-Net [40]0.9422--
CoRe [47]0.95120.2600-
DAE-CoRe [48]0.9589--
Exemplar freeTPT [2]0.96070.2378-0.1111
C3D-AVG-MTL [24]0.9044--
USDL [35]0.92310.46800.1556
MUSDL [35]0.92730.4510-0.0667
DAE [48]0.9231--
DAE-MT [48]0.94900.2738-0.4222
RICA2 (Ours)0.95940.25800.6000
RICA2† (Ours)0.96200.2280-
" + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.36, + 0.48, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.246, + 0.499, + 0.491, + 0.512 + ], + "angle": 0, + "content": "(a) Uncertainty calibration on FineDiving" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.36, + 0.744, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.499, + 0.756, + 0.512 + ], + "angle": 0, + "content": "(b) Uncertainty calibration on MTL-AQA" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.523, + 0.789, + 0.581 + ], + "angle": 0, + "content": "Fig. 3: Uncertainty vs. prediction error (MAE) on (a) Finediving and (b) MTL-AQA. Results are reported on the test splits, with the X-axis as the uncertainty bin index (uncertainty increases from left to right) and the Y-axis as the MAE in the bin. In comparison to baselines, \\(\\mathrm{RICA}^2\\) has improved calibration with lower prediction errors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.617, + 0.79, + 0.708 + ], + "angle": 0, + "content": "Experiment setup. We follow the evaluation protocol of [2,43,47], dividing the dataset into the standard train set of 1059 videos and a test set of 353 videos. Further, we use the same input video settings as TPT [2] in our experiments to ensure a fair comparison. Specifically, for each video, we uniformly sample 103 frames segmented into 20 overlapping clips, each containing 8 continuous frames. Please refer to supplement Sec. A.2 for more details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Results. Tab. 1b summarizes our results on MTL-AQA. Similar to FineDiving, \\(\\mathrm{RICA}^2\\) shows state-of-the-art results on MTL-AQA across all evaluation metrics. Specifically, our \\(\\mathrm{RICA}^2\\dagger\\) outperforms the best exemplar-free model (DAE-MT) by a relative margin of \\(1.4\\% / 16.7\\%\\) on \\(SRCC / R\\ell_2\\). When compared with the competitive exemplar-based TPT, our has slightly better \\(SRCC\\) (\\(SRCC_{TPT} = 0.9607\\) vs \\(SRCC_{\\mathrm{RICA}^2\\dagger} = 0.9620\\)) and \\(R\\ell_2\\) (\\(+4.1\\%\\) relative margin). Again, compared to previous methods, our stochastic model shows improved calibration (\\(\\tau_{\\mathrm{RICA}^2} = 0.60\\) vs. \\(\\tau_{USDL} = 0.16\\) vs. \\(\\tau_{TPT} = -0.11\\)) as shown in Fig. 3b." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.129 + ], + "angle": 0, + "content": "A. Majeedi et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.144, + 0.788, + 0.175 + ], + "angle": 0, + "content": "Table 2: Results on JIGSAWS [10] dataset. Only prediction accuracy (\\(SRCC\\)) is considered due to the limited sample size. RICA\\(^2\\) outperforms all prior approaches." + }, + { + "type": "table", + "bbox": [ + 0.362, + 0.185, + 0.644, + 0.348 + ], + "angle": 0, + "content": "
Task
SNPKTAvg
Exemplar basedCoRe [47]0.840.860.860.85
TPT [2]0.880.880.910.89
ST-GCN [46]0.310.390.580.43
TSN [39]0.340.230.720.46
JRG [23]0.360.540.750.57
Exemplar freeUSDL [35]0.640.630.610.63
MUSDL [35]0.710.690.710.70
DAE [48]0.730.720.720.72
DAE-MT [48]0.780.740.740.76
RICA2† (Ours)0.880.930.880.90
RICA2 (Ours)0.920.940.900.92
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.373, + 0.441, + 0.387 + ], + "angle": 0, + "content": "4.3 Results on JIGSAWS" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.395, + 0.789, + 0.472 + ], + "angle": 0, + "content": "Dataset. In addition to diving videos, we also evaluate RICA² on JIGSAWS [10] — a robotic surgical video dataset. The dataset includes three tasks: \"Suturing (S),\" with 39 recordings, \"Needle Passing (NP),\" with 26 recordings and \"Knot Tying (KT)\" with 36 recordings. JIGSAWS is widely used for action quality assessment, despite its small scale." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.79, + 0.627 + ], + "angle": 0, + "content": "Experiment setup. Due to the limited number of samples in the dataset (as few as 7 videos in the test set), cross-validation is often considered for evaluation on JIGSAWS. To ensure a fair comparison, we follow the commonly adopted splits from [35], and the input video setting from [2]. Specifically, for each video, we uniformly sample 160 frames which are segmented into 20 non-overlapping clips. We opt to not include score calibration curves due to the limited sample size of the test sets. Additionally, the key steps in JIGSAWS are general motions (e.g. reaching for the needle, orienting the needle, etc.) and thus cannot be localized to any specific section of the video. Thus, we do not use the auxiliary losses for this experiment. More details are described in supplement Sec. A.3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.789, + 0.738 + ], + "angle": 0, + "content": "Results. Tab. 2 summarizes our results on JIGSAWS. Similar to previous datasets, our models exhibit notable advancements over the previous state-of-the-art model TPT [35], showcasing substantial improvements of \\(1.1\\%\\) (RICA²) to \\(3.4\\%\\) (RICA²†) in terms of average SRCC relative to the exemplar-based state-of-the-art TPT [2]. When compared to the exemplar-free methods, our approach demonstrates an impressive \\(18.4\\%\\) (RICA²) to \\(21.0\\%\\) (RICA²†) relative gain in average SRCC compared to the latest method DAE-MT [48]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.754, + 0.403, + 0.769 + ], + "angle": 0, + "content": "4.4 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.777, + 0.787, + 0.808 + ], + "angle": 0, + "content": "To understand our model design choices, we conduct ablation studies on the MTL-AQA [24] dataset. Additional ablations are in supplement Sec. C.2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Experiment setup. To simplify our experiments, we opt for running our ablations using fixed I3D features. This allows us to precisely evaluate the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.334, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "\\(\\mathrm{RICA}^2\\) : Rubric-Informed, Calibrated Assessment of Actions" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 3: Ablation studies of model components on MTL-AQA dataset. * indicates that the text embeddings were frozen during training." + }, + { + "type": "table", + "bbox": [ + 0.264, + 0.185, + 0.74, + 0.282 + ], + "angle": 0, + "content": "
Step Rep.DAG (Rubric)KLAuxMetrics
SRCC(↑)Rl2(↓)τ(↑)Avg. Rank (↓)
Random×××0.94260.3882-5.50
Text×××0.94310.3509-4.50
Text✓××0.94300.3336-4.50
Text*✓××0.94370.3335-3.50
Text*✓✓×0.94480.33290.42221.83
Text*✓✓✓0.94600.33030.42221.17
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.309, + 0.784, + 0.355 + ], + "angle": 0, + "content": "contribution of different components of our model. Specifically, we choose I3D weights from an intermediate checkpoint of our trained model and extract features for all videos with the frozen backbone." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.788, + 0.448 + ], + "angle": 0, + "content": "Base model. Our ablation constructs a base model using randomly initialized step embeddings, an averaging of these embeddings after cross attention with the video features, followed by an MLP for scoring. This base model is trained using only the MSE loss. We then gradually add modules from \\(\\mathrm{RICA}^2\\) and study their effects. Tab. 3 presents our results using the same features and training epochs, with our base model in row 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.452, + 0.788, + 0.529 + ], + "angle": 0, + "content": "Text embeddings as step representations. We first replace randomly initialized step representations with the text embeddings of step descriptions. This leads to a major boost in \\( R\\ell_{2} \\) (Tab. 3 row 1 vs. row 2), by leveraging knowledge encoded in the LLM [7]. Further, we find that freezing the text embeddings leads to comparable results and faster convergence (Tab. 3 row 3 vs. row 4)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.531, + 0.788, + 0.594 + ], + "angle": 0, + "content": "Does the scoring rubric help? We also investigate the effects of encoding steps and rubric as a DAG—a key design of our model. Adding the DAG results in a noteworthy boost in \\( R\\ell_{2} \\) (Tab. 3 row 2 vs. row 3). This improvement can be ascribed to the DAG's proficiency in dissecting the action quality across steps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.788, + 0.702 + ], + "angle": 0, + "content": "Effects of loss functions. We now study the loss terms. Our loss function has three terms (a) the MSE loss \\((\\mathcal{L}_{MSE})\\) to minimize prediction error, (b) the KL loss \\((\\mathcal{L}_{KL})\\) to regularize the stochastic embeddings, and (c) the auxiliary loss \\((\\mathcal{L}_{Aux})\\) to ensure temporal ordering of steps. Adding the KL loss \\(\\mathcal{L}_{KL}\\) yields similar results in \\(SRCC\\) and \\(R\\ell_{2}\\), yet enables calibrated uncertainty estimation. Further attaching the auxiliary loss \\(\\mathcal{L}_{Aux}\\) leads to improvement in both \\(SRCC\\) and \\(R\\ell_{2}\\), while maintaining the calibration performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.796 + ], + "angle": 0, + "content": "Evaluating the cross-attention maps. To gain insight into \\(\\mathrm{RICA}^2\\), we now examine the cross-attention maps between the step representations and video features in our learned embedding function \\(f\\). Fig. 4 visualizes the attention map on two test videos on FineDiving. These maps reveal that a step representation is likely to attend to video features during which the step occurs, indicating that \\(\\mathrm{RICA}^2\\) learns to encode the temporal location of individual steps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "We further evaluate this localization ability following the Pointing game protocol [49], widely considered in weakly supervised / unsupervised localization tasks [42,52]. Pointing Game compares a generated heatmap with an annotated" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.129 + ], + "angle": 0, + "content": "A. Majeedi et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.784, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.425, + 0.79, + 0.482 + ], + "angle": 0, + "content": "Fig. 4: Visualization of the cross-attention maps. Y-axis : attention value; Y-axis: clip indices (time). Each curve shows an attention map from a step representation to the temporal video features. The frames shown above are aligned with the timing of the corresponding attention plot. Curves and steps are colored accordingly." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.514, + 0.788, + 0.621 + ], + "angle": 0, + "content": "time interval and counts the chance of the heatmap's peak falling into the specified interval. Our evaluation focuses on the FineDiving dataset since it is the only dataset providing annotated time intervals for individual steps. When evaluated on the full test set, attention maps from our model attain an accuracy of \\(61.4\\%\\) in the Pointing game protocol, significantly outperforming the chance level accuracy of \\(30.7\\%\\) (given each video has 3.26 steps on average). Note that we did not use any annotated segmentation data for training." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.651, + 0.51, + 0.667 + ], + "angle": 0, + "content": "5 Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we present a deep probabilistic model for action quality assessment in videos. Our key innovation is to integrate score rubrics and to model prediction uncertainty. Specifically, we propose to adapt stochastic embeddings to quantify the uncertainty of individual steps, and to decode action scores using a variant of graph neural network operating on a DAG encoding the score rubric. Our method offers an exemplar-free approach for AQA, achieves new state-of-the-art results in terms of prediction accuracy on public benchmarks, and demonstrates superior calibration of the output uncertainty estimates. We believe that our work provides a solid step towards AQA. We hope that our method and findings can shed light on the challenging problem of trustworthy video recognition." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.335, + 0.114, + 0.732, + 0.128 + ], + "angle": 0, + "content": "RICA²: Rubric-Informed, Calibrated Assessment of Actions" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "Acknowledgement: This work was supported by the UW Madison Office of the Vice Chancellor for Research with funding from the Wisconsin Alumni Research Foundation, by National Science Foundation under Grant No. CNS 2333491, and by the Army Research Lab under contract number W911NF-2020221." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.23, + 0.323, + 0.246 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.26, + 0.785, + 0.288 + ], + "angle": 0, + "content": "1. Alemi, A.A., Fischer, I., Dillon, J.V., Murphy, K.: Deep variational information bottleneck. In: International Conference on Learning Representations (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.787, + 0.344 + ], + "angle": 0, + "content": "2. Bai, Y., Zhou, D., Zhang, S., Wang, J., Ding, E., Guan, Y., Long, Y., Wang, J.: Action quality assessment with temporal parsing transformer. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IV. pp. 422-438. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.344, + 0.787, + 0.399 + ], + "angle": 0, + "content": "3. Battaglia, P.W., Hamrick, J.B., Bapst, V., Sanchez-Gonzalez, A., Zambaldi, V., Malinowski, M., Tacchetti, A., Raposo, D., Santoro, A., Faulkner, R., et al.: Relational inductive biases, deep learning, and graph networks. arXiv preprint arXiv:1806.01261 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.4, + 0.787, + 0.44 + ], + "angle": 0, + "content": "4. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.441, + 0.787, + 0.481 + ], + "angle": 0, + "content": "5. Chen, C.H., Hu, Y.H., Yen, T.Y., Radwin, R.G.: Automated video exposure assessment of repetitive hand activity level for a load transfer task. Human factors 55(2), 298-308 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.482, + 0.787, + 0.523 + ], + "angle": 0, + "content": "6. Chun, S., Oh, S.J., De Rezende, R.S., Kalantidis, Y., Larlus, D.: Probabilistic embeddings for cross-modal retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8415-8424 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.524, + 0.787, + 0.62 + ], + "angle": 0, + "content": "7. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., Webson, A., Gu, S.S., Dai, Z., Suzgun, M., Chen, X., Chowdhery, A., Castro-Ros, A., Pellat, M., Robinson, K., Valter, D., Narang, S., Mishra, G., Yu, A., Zhao, V., Huang, Y., Dai, A., Yu, H., Petrov, S., Chi, E.H., Dean, J., Devlin, J., Roberts, A., Zhou, D., Le, Q.V., Wei, J.: Scaling instructionfinetuned language models. Journal of Machine Learning Research 25(70), 1-53 (2024), http://jmlr.org/papers/v25/23-0870.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.787, + 0.661 + ], + "angle": 0, + "content": "8. Duvenaud, D.K., Maclaurin, D., Iparraguirre, J., Bombarell, R., Hirzel, T., Aspuru-Guzik, A., Adams, R.P.: Convolutional networks on graphs for learning molecular fingerprints. Advances in neural information processing systems 28 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.787, + 0.702 + ], + "angle": 0, + "content": "9. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 6202-6211 (2019). https://doi.org/10.1109/ICCV.2019.00630" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.702, + 0.787, + 0.771 + ], + "angle": 0, + "content": "10. Gao, Y., Vedula, S.S., Reiley, C.E., Ahmidi, N., Varadarajan, B., Lin, H.C., Tao, L., Zappella, L., Bejar, B., Yuh, D.D., et al.: Jhu-isi gesture and skill assessment working set (jigsaws): A surgical activity dataset for human motion modeling. In: Modeling and Monitoring of Computer Assisted Interventions (M2CAI) - MICCAI Workshop (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.771, + 0.787, + 0.799 + ], + "angle": 0, + "content": "11. Gordon, A.S.: Automated video assessment of human performance. In: Proceedings of AI-ED. vol. 2 (1995)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "12. Guo, C., Pleiss, G., Sun, Y., Weinberger, K.Q.: On calibration of modern neural networks. In: International conference on machine learning. pp. 1321-1330. PMLR (2017)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.26, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.386, + 0.129 + ], + "angle": 0, + "content": "A. Majeedi et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "13. Kendall, M.G.: A new measure of rank correlation. Biometrika 30(1/2), 81-93 (1938)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.203 + ], + "angle": 0, + "content": "14. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. In: International Conference on Learning Representations (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "15. Kipf, T.N., Welling, M.: Semi-supervised classification with graph convolutional networks. In: International Conference on Learning Representations (2017), https://openreview.net/forum?id=SJU4ayYgl" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "16. Li, W., Huang, X., Lu, J., Feng, J., Zhou, J.: Learning probabilistic ordinal embeddings for uncertainty-aware regression. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13896-13905 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.313 + ], + "angle": 0, + "content": "17. Likert, R.: A technique for the measurement of attitudes. Archives of psychology (1932)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "18. Liu, D., Li, Q., Jiang, T., Wang, Y., Miao, R., Shan, F., Li, Z.: Towards unified surgical skill assessment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9522-9531 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.397 + ], + "angle": 0, + "content": "19. Martin, J., Regehr, G., Reznick, R., Macrae, H., Murnaghan, J., Hutchison, C., Brown, M.: Objective structured assessment of technical skill (osats) for surgical residents. British journal of surgery 84(2), 273-278 (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "20. Matsuyama, H., Kawaguchi, N., Lim, B.Y.: Iris: Interpretable rubric-informed segmentation for action quality assessment. In: Proceedings of the 28th International Conference on Intelligent User Interfaces. pp. 368-378 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.44, + 0.785, + 0.522 + ], + "angle": 0, + "content": "21. Neelakantan, A., Shankar, J., Passos, A., McCallum, A.: Efficient non-parametric estimation of multiple embeddings per word in vector space. In: Moschitti, A., Pang, B., Daelemans, W. (eds.) Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing, EMNLP 2014, October 25-29, 2014, Doha, Qatar, A meeting of SIGDAT, a Special Interest Group of the ACL. pp. 1059-1069. ACL (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.563 + ], + "angle": 0, + "content": "22. Oh, S.J., Gallagher, A.C., Murphy, K.P., Schroff, F., Pan, J., Roth, J.: Modeling uncertainty with hedged instance embeddings. In: International Conference on Learning Representations (2019), https://openreview.net/forum?id=r1xQQhAqKX" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.605 + ], + "angle": 0, + "content": "23. Pan, J.H., Gao, J., Zheng, W.S.: Action assessment by joint relation graphs. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6331-6340 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.646 + ], + "angle": 0, + "content": "24. Parmar, P., Morris, B.T.: What and how well you performed? a multitask learning approach to action quality assessment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 304-313 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "25. Parmar, P., Tran Morris, B.: Learning to score olympic events. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops. pp 20-28 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.73 + ], + "angle": 0, + "content": "26. Pirsiavash, H., Vondrick, C., Torralba, A.: Assessing the quality of actions. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13. pp. 556-571. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.731, + 0.785, + 0.757 + ], + "angle": 0, + "content": "27. Prassas, S., Kwon, Y.H., Sands, W.A.: Biomechanical research in artistic gymnastics: a review. Sports Biomechanics 5(2), 261-291 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "28. Qiu, Y., Wang, J., Jin, Z., Chen, H., Zhang, M., Guo, L.: Pose-guided matching based on deep learning for assessing quality of action on rehabilitation training Biomedical Signal Processing and Control 72, 103323 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "29. Santoro, A., Raposo, D., Barrett, D.G., Malinowski, M., Pascanu, R., Battaglia, P., Lillicrap, T.: A simple neural network module for relational reasoning. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.335, + 0.114, + 0.732, + 0.128 + ], + "angle": 0, + "content": "RICA²: Rubric-Informed, Calibrated Assessment of Actions" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "30. Scarselli, F., Gori, M., Tsoi, A.C., Hagenbuchner, M., Monfardini, G.: The graph neural network model. IEEE transactions on neural networks 20(1), 61-80 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.244 + ], + "angle": 0, + "content": "31. Schoeffmann, K., Taschwer, M., Sarny, S., Munzer, B., Primus, M.J., Putzgruber, D.: Cataract-101: video dataset of 101 cataract surgeries. In: Cesar, P., Zink, M., Murray, N. (eds.) Proceedings of the 9th ACM Multimedia Systems Conference, MMSys 2018, Amsterdam, The Netherlands, June 12-15, 2018. pp. 421-425. ACM (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.246, + 0.786, + 0.273 + ], + "angle": 0, + "content": "32. Shi, Y., Jain, A.K.: Probabilistic face embeddings. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6902-6911 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.273, + 0.787, + 0.313 + ], + "angle": 0, + "content": "33. Sohn, K., Lee, H., Yan, X.: Learning structured output representation using deep conditional generative models. Advances in neural information processing systems 28 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.315, + 0.787, + 0.369 + ], + "angle": 0, + "content": "34. Sun, J.J., Zhao, J., Chen, L.C., Schroff, F., Adam, H., Liu, T.: View-invariant probabilistic embedding for human pose. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16. pp. 53-70. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.371, + 0.787, + 0.424 + ], + "angle": 0, + "content": "35. Tang, Y., Ni, Z., Zhou, J., Zhang, D., Lu, J., Wu, Y., Zhou, J.: Uncertainty-aware score distribution learning for action quality assessment. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9839-9848 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.425, + 0.787, + 0.452 + ], + "angle": 0, + "content": "36. TISHBY, N.: The information bottleneck method. In: Proc. of the 37th Allerton Conference on Communication and Computation, 1999 (1999)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.453, + 0.787, + 0.493 + ], + "angle": 0, + "content": "37. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.495, + 0.787, + 0.522 + ], + "angle": 0, + "content": "38. Vilnis, L., McCallum, A.: Word representations via gaussian embedding. In: International Conference on Learning Representations (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.523, + 0.787, + 0.563 + ], + "angle": 0, + "content": "39. Wang, L., Xiong, Y., Wang, Z., Qiao, Y., Lin, D., Tang, X., Van Gool, L.: Temporal segment networks: Towards good practices for deep action recognition. In: European conference on computer vision. pp. 20-36. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.565, + 0.787, + 0.605 + ], + "angle": 0, + "content": "40. Wang, S., Yang, D., Zhai, P., Chen, C., Zhang, L.: Tsa-net: Tube self-attention network for action quality assessment. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 4902-4910 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.606, + 0.787, + 0.633 + ], + "angle": 0, + "content": "41. Waters, T.R., Putz-Anderson, V., Garg, A.: Applications manual for the revised niosh lifting equation (1994)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.634, + 0.787, + 0.674 + ], + "angle": 0, + "content": "42. Xiao, F., Sigal, L., Jae Lee, Y.: Weakly-Supervised Visual Grounding of Phrases With Linguistic Structures. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5945-5954 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.675, + 0.787, + 0.715 + ], + "angle": 0, + "content": "43. Xu, C., Fu, Y., Zhang, B., Chen, Z., Jiang, Y.G., Xue, X.: Learning to score figure skating sport videos. IEEE transactions on circuits and systems for video technology 30(12), 4578-4590 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.717, + 0.787, + 0.757 + ], + "angle": 0, + "content": "44. Xu, J., Rao, Y., Yu, X., Chen, G., Zhou, J., Lu, J.: Finediving: A fine-grained dataset for procedure-aware action quality assessment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2949-2958 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.758, + 0.787, + 0.798 + ], + "angle": 0, + "content": "45. Xu, K., Li, J., Zhang, M., Du, S.S., ichi Kawarabayashi, K., Jegelka, S.: What can neural networks reason about? In: International Conference on Learning Representations (2020), https://openreview.net/forum?id=rJxbJeHFPS" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "46. Yan, S., Xiong, Y., Lin, D.: Spatial temporal graph convolutional networks for skeleton-based action recognition. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.387, + 0.129 + ], + "angle": 0, + "content": "A. Majeedi et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "47. Yu, X., Rao, Y., Zhao, W., Lu, J., Zhou, J.: Group-aware contrastive regression for action quality assessment. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7899-7908. IEEE Computer Society, Los Alamitos, CA, USA (oct 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "48. Zhang, B., Chen, J., Xu, Y., Zhang, H., Yang, X., Geng, X.: Auto-encoding score distribution regression for action quality assessment. Neural Computing and Applications pp. 1-14 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.787, + 0.286 + ], + "angle": 0, + "content": "49. Zhang, J., Bargal, S.A., Lin, Z., Brandt, J., Shen, X., Sclaroff, S.: Top-Down Neural Attention by Excitation Backprop. International Journal of Computer Vision 126(10), 1084-1102 (Oct 2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.314 + ], + "angle": 0, + "content": "50. Zhou, C., Huang, Y.: Uncertainty-driven action quality assessment. arXiv preprint arXiv:2207.14513 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.356 + ], + "angle": 0, + "content": "51. Zhou, K., Ma, Y., Shum, H.P.H., Liang, X.: Hierarchical graph convolutional networks for action quality assessment. IEEE Transactions on Circuits and Systems for Video Technology pp. 1-1 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.398 + ], + "angle": 0, + "content": "52. Zhu, Y., Zhou, Y., Ye, Q., Qiu, Q., Jiao, J.: Soft Proposal Networks for Weakly Supervised Object Localization. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1841-1850 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.398 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/RICA^2_ Rubric-Informed, Calibrated Assessment of Actions/805cf405-290f-425f-a265-d9aa43a99eb8_origin.pdf b/2024/RICA^2_ Rubric-Informed, Calibrated Assessment of Actions/805cf405-290f-425f-a265-d9aa43a99eb8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6fd9497ecb0fa560753e63451bbe060412304b2c --- /dev/null +++ b/2024/RICA^2_ Rubric-Informed, Calibrated Assessment of Actions/805cf405-290f-425f-a265-d9aa43a99eb8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3817dcd601672157f0e3969cda1736800f7373fd9323e211d8f58ac8ed0d9289 +size 2219681 diff --git a/2024/RICA^2_ Rubric-Informed, Calibrated Assessment of Actions/full.md b/2024/RICA^2_ Rubric-Informed, Calibrated Assessment of Actions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6e552059db5f7ad1bbd697eabeed639a3444d5a3 --- /dev/null +++ b/2024/RICA^2_ Rubric-Informed, Calibrated Assessment of Actions/full.md @@ -0,0 +1,325 @@ +# RICA $^2$ : Rubric-Informed, Calibrated Assessment of Actions + +Abrar Majeedi, Viswanatha Reddy Gajjala, Satya Sai Srinath Namburi GNVV, and Yin Li + +University of Wisconsin-Madison, Madison Wisconsin 53706 USA {majeedi,vgajjala,sgnamburi,yin.li}@wisc.edu + +Abstract. The ability to quantify how well an action is carried out, also known as action quality assessment (AQA), has attracted recent interest in the vision community. Unfortunately, prior methods often ignore the score rubric used by human experts and fall short of quantifying the uncertainty of the model prediction. To bridge the gap, we present RICA² — a deep probabilistic model that integrates score rubric and accounts for prediction uncertainty for AQA. Central to our method lies in stochastic embeddings of action steps, defined on a graph structure that encodes the score rubric. The embeddings spread probabilistic density in the latent space and allow our method to represent model uncertainty. The graph encodes the scoring criteria, based on which the quality scores can be decoded. We demonstrate that our method establishes new state of the art on public benchmarks, including FineDiving, MTL-AQA, and JIGSAWS, with superior performance in score prediction and uncertainty calibration. Our code is available at https://abrarmajeedi.github.io/rica2_aqa/. + +Keywords: Action Quality Assessment $\cdot$ Video Understanding + +# 1 Introduction + +Action quality assessment (AQA), aiming at quantifying how well an action is carried out, has been widely studied across scientific disciplines due to its broad range of applications. AQA is key to sports science and analytics. The right way of performing actions maximizes an athlete's performance and minimizes injury risk. AQA is crucial to occupational safety and health. High-quality actions mitigate the physical stress and strain in the workspace. AQA is pivotal for physical therapies. The quality of actions reveals the progress in rehabilitation. AQA also plays a major role in surgical education. Proficient actions improve the outcome and reduce complications. + +Observational methods for AQA have been well established for various tasks, e.g., gymnastics [27], manual material handling [41], and surgery [19]. These methods involve a human expert observing an action and decomposing it into a series of key steps. Each of these steps, or a subset of them, can be grouped into a factor and then evaluated using a Likert scale [17] following a pre-defined criterion. Ratings for individual factors, sometimes complemented with impression-based + +![](images/baf7824ac63e33344dcd192a11f8f7d2052e301734f4f126033754c0a29b7fcb.jpg) +Fig. 1: RICA2 integrates score rubric used by human experts and accounts for prediction uncertainty, resulting in accurate predictions and calibrated uncertainty estimates. + +global ratings, are then summarized into a final quality score [19]. Multiple expert ratings are often considered to account for the variance in the scores. While these methods are commonly adopted, they require significant input from human experts and are thus costly and inefficient. + +There is a burgeoning interest in the vision community to develop video-based AQA [25, 26, 35, 43]. While current solutions have made steady progress across benchmarks [10, 24, 44], their decision-making processes differ largely from prior observational methods. Almost all prior solutions learn deep models to directly map input videos to scores. Many of them employ an exemplar-based approach, in which a model predicts relative scores by referencing exemplar videos with similar actions and known scores [2, 44, 47]. Few of them have considered the structure of the actions or their scoring criteria used by observational methods. + +Further, existing AQA methods face a key challenge in the accurate quantification of model uncertainty i.e., the uncertainty of model prediction that is calibrate to the expected error [12]. Knowing this uncertainty is particularly helpful for AQA, e.g., when assessing the quality of high-stakes competitions or surgical procedures. With proper calibration, videos that have uncertain predictions can be passed to human experts for a thorough evaluation. Several recent works have started to consider the variance among scores from multiple human experts [35,48,50,51]. Unfortunately, they still fall short of considering prediction uncertainty, leaving this challenge largely unaddressed. + +To bridge the gap, we develop a deep probabilistic model for AQA by integrating score rubrics and modeling the uncertainty of the prediction (see Fig. 1). Central to our method lies in the stochastic embedding of action steps, defined on a graph structure that encodes the score rubric. The embeddings spread probabilistic density in latent space and allow our method to represent model uncertainty. The graph encodes the scoring criteria, based on which the quality scores can be decoded. We also present a training scheme and describe an approach to estimate uncertainty. Putting things together, our method, dubbed RICA $^2$ (Rubric-informed, Calibrated Assessment of Actions), yields accurate action scores with additional uncertainty estimates. + +We evaluate RICA² on several public AQA datasets, covering sports and surgical videos. Particularly, RICA² establishes new state of the art on FineDiving [44], MTL-AQA [24] and JIGSAWS [10]. On FineDiving [44] - the largest and most challenging AQA benchmark, RICA² outperforms latest methods in prediction accuracy (a boost of $+0.94\%$ in Spearman's Rank Correlation Coefficient + +$(SRCC))$ and demonstrates significantly improved uncertainty calibration (a gain of $+0.178$ in Kendall Tau [13]). Similarly, on MTL-AQA [24], the most commonly used dataset for AQA, $\mathrm{RICA}^2$ attains state-of-the-art $SRCC$ , and again largely improved calibration (a gain of $+0.444$ in Kendall Tau). On JIGSAWS [10], $\mathrm{RICA}^2$ beats the previous best results by a relative margin of $+3.37\%$ in $SRCC$ . Further, we present extensive experiments to evaluate the key design of $\mathrm{RICA}^2$ . + +Our main contributions are summarized into three folds. + +- We propose $RICA^2$ , a novel deep probabilistic method that incorporates scoring rubrics and uncertainty modeling for AQA, resulting in accurate scores and calibrated uncertainty estimates. +- Our technical innovations lie in (a) a graph neural network to model the scoring rubric in conjunction with stochastic embeddings on the graph to account for prediction uncertainty and (b) a training scheme under the variational information bottleneck framework. +- Our extensive set of experiments demonstrates that $\mathrm{RICA}^2$ achieves state-of-the-art results in AQA, significantly outperforming prior methods in both prediction accuracy and calibration of uncertainty estimates. + +# 2 Related Work + +Action quality assessment (AQA). Early works in AQA [11,26] employed handcrafted features to estimate quality scores in videos. More recent methods developed various deep models, including convolutional [35,47], graph [23], recurrent [25, 43], and Transformer [2, 44] networks. AQA has also been widely considered in surgical education [18], rehabilitation [28], and ergonomics [5]. + +Recently, exemplar-based methods [2, 44, 47] have emerged as a promising solution for AQA due to their impressive performance across benchmarks. These methods predict the relative score of an input video by comparing it to selected exemplar videos with similar action steps and known scores. A limitation of this paradigm is the requirement of exemplar videos at inference time. This strategy largely deviates from existing observational methods used by human experts and leads to significantly higher computational costs. While RICA² also uses action steps in the input video, it further integrates the scoring rubric of these steps and offers a solution for no-reference AQA i.e. without using exemplars. + +Several recent works have started to consider the modeling of score uncertainty in AQA [35, 48, 50, 51]. For example, Tang et al. [35] proposed to model the final scores using a Gaussian distribution. They presented a model (MUSDL) trained to predict the score distribution. This distribution learning idea was further extended in [48, 50, 51]. However, modeling the score distribution does not warrant the quantification of model uncertainty, as the output distributions might not be calibrated with prediction errors. While RICA² also predicts a Gaussian distribution for the scores, our key design is to consider stochastic embeddings to quantify prediction uncertainty, resulting in calibrated uncertainty estimates. + +The most relevant work is IRIS [20]. IRIS incorporates score rubric into a convolutional network for AQA. This is done by segmenting key steps in the + +video and predicting sub-scores for individual steps. Similar to IRIS, RICA $^2$ also considers rubric in a deep model. However, RICA $^2$ adapts a graph network, treats sub-scores as latent embeddings, predicts the final score, and further quantifies prediction uncertainty. These differences allow RICA $^2$ to be trained on major public datasets with only final scores, and to output calibrated uncertainty estimates, both of which cannot be achieved by IRIS. + +Modeling uncertainty with stochastic embedding. Stochastic embedding, initially introduced in NLP [21,38], treats each embedding as a distribution. This approach has gained recent attention for modeling uncertainty in deep models. Oh et al. [22] considered probabilistic embeddings for metric learning and proposed to model uncertainty based on the stochasticity of embeddings. This idea was further adopted in many vision tasks, including face verification [32], age estimation [16], pose estimation [34], and cross-modal retrieval [6]. Another related line of work is the conditional variational autoencoder [33], where a probabilistic representation of the input is used for a prediction task. Our approach shares a similar idea of using stochastic embeddings to model uncertainty yet is specifically designed for AQA. Our method significantly extends prior idea to embed action steps on a graph structure, and to propagate these stochastic embeddings on the graph. + +Graph neural networks (GNNs). GNNs [8, 15, 30] offer a powerful tool to leverage the relational inductive bias inherent in data [3, 45]. This inductive bias is beneficial to aggregate a global representation from a group of local ones [29]. Recently, Zhou et al. [51] proposed a hierarchical graph convolutional network for AQA, in which a GNN was used for video representation learning. In contrast, we adapt graph networks to model score rubrics used by observational methods. + +# 3 AQA with Score Rubric and Uncertainty Modeling + +Our goal is to assess the quality of an action within an input video. Let $X$ be the video with the action and $Y$ as its quality score. Our method further considers the structure of the action and a scoring rubric based on the structure. + +Action steps. We assume that the action in $X$ comprises a known, ordered set of key steps, denoted as $\mathbb{S} = (s_1, s_2, \dots, s_k)$ . Each $s^1$ represents a necessary sub-action for successfully executing the action. Further, $s$ is associated with a text description that elucidates the specifics of the corresponding step, e.g., "a front-facing takeoff" for diving. This assumption is especially well suited for structured actions, such as diving or surgery, where the key steps are predetermined and follow a specific sequence. Note that the timing of the key steps is not presumed. Even if key steps are unavailable, they can be detected using action recognition methods [4,9] (see supplement Sec. C.4). + +Scoring rubric. We further assume a pre-specified scoring rubric based on the key steps — a common strategy in technical skill assessment [19, 27, 41]. Specifically, each action step $s_k$ is independently scored, i.e., $s_k \mapsto y_k$ . Subsequently, a rule-based rubric is employed to aggregate individual scores $\{y_k\}$ and calculate a final + +![](images/d47ba5f9c11f940abc351e7ed272bf2c05b63595941f5eb066769fa5b21f5631.jpg) +(a) AQA Scoring Rubric + +![](images/8ab449158ac2f05e7496eb8ec143cfd5d9fece82c172f23784d72d9df068b39f.jpg) +(b) GNN on DAG + +![](images/0159e32da3c3e14823cbfe67fe915c37302b1c0a685906254e47a461e35648f5.jpg) +(c) Embedding Function $\pmb{f}$ +Fig. 2: Overview of RICA $^2$ . Leveraging scoring rubrics (a), RICA $^2$ integrates a graph representation of action step and rubric with uncertainty modeling (b). Specifically, RICA $^2$ takes an input of the video and its key action steps, encodes the input into embeddings (c), refines the embeddings through a deep probabilistic model, and outputs an action score in tandem with its uncertainty estimate. + +quality score $Y$ , i.e., $\{y_k\} \mapsto Y$ , in which steps might be grouped into intermediate stages (see an example in Fig. 2 (a-b)). This rubric follows a deterministic yet often non-injective mapping, e.g. a many-to-one mapping such as summation. + +Method overview. We now present RICA $^2$ — a deep probabilistic model for AQA that leverages known action steps and incorporates the scoring rubric for modeling. Importantly, RICA $^2$ accounts for prediction uncertainty, i.e., when the model prediction can and cannot be trusted. Fig. 2 presents an overview of RICA $^2$ . It consists of two main model components: (a) a graph neural network that integrates the key steps and scoring rubric (Sec. 3.1); and (b) stochastic embeddings defined on the graph to capture prediction uncertainty (Sec. 3.2), coupled with (c) a learning scheme under the variational information bottleneck framework (Sec. 3.3). In what follows, we delve into the details of RICA $^2$ . + +# 3.1 Integrating Actions Steps and Scoring Rubric with Graph + +Steps and rubric as graph. We encode action steps and the corresponding scoring rubric with a directed acyclic graph (DAG). This DAG is denoted as $\mathcal{G} = (\mathbb{V},\mathbb{E})$ with $\mathbb{V}$ as the set of nodes and $\mathbb{E}$ as the set of directed edges. $\mathbb{V}$ consists of three types of nodes: (1) the leaf nodes, denoted as $V^s$ , correspond to individual action steps performed in the input video $X$ ; (2) the intermediate nodes capturing possible intermediate stages in the scoring criteria; and (3) a designated root node $V^r$ representing the final score of the action. Further, the edges $\mathbb{E}$ indicate the scoring rubric, connecting steps (leaf nodes) to stages (intermediate nodes), and stages (intermediate nodes) to the final score (root node). We note that $\mathcal{G}$ varies for every input video $X$ (assuming a single action), as different steps might be performed. Fig. 2 (a-b) show the example in diving where the key steps and scoring rubric are encoded using our DAG. Additional examples can be found in our supplement Fig. B. + +Learning for quality assessment. Our approach involves a two-step process for quality assessment. First, we employ an embedding function $f$ , designed to map individual steps into a latent space representing action quality. Secondly, we leverage the key step embeddings $\{Z^s\}$ along with the score rubric encoded in $\mathcal{G}$ to learn a scoring function $h$ . These functions $f$ and $h$ are defined as follows: + +$$ +f: X, \mathcal {G} \mapsto \{Z ^ {s} \}; \quad h: \{Z ^ {s} \}, \mathcal {G} \mapsto Y, \tag {1} +$$ + +where $\{Z^s\}$ are the embeddings for the set of steps $\mathbb{S}$ in $X$ , corresponding to the leaf nodes $\{V^s\}$ on the DAG. + +# 3.2 Modeling Score Uncertainty with Stochastic Embeddings + +To model the prediction uncertainty, we adopt stochastic step embeddings defined on the leaf nodes, such that $Z^{s} \in \mathbb{R}^{D} \sim p(Z^{s}|X,\{V^{s}\})$ . Unlike deterministic embeddings, where $Z^{s}$ would be a fixed vector, stochastic embedding characterizes the distribution of $Z^{s}$ , allowing for uncertainty control. Specifically, we model $p(Z^{s}|X,V^{s})$ as a Gaussian distribution in $\mathbb{R}^{D}$ with mean $\mu^{s}$ and diagonal covariance $\Sigma^{s}$ . The embedding function $f$ is thus tasked to predict the mean and covariance for the key steps $\mathbb{S}$ , i.e., $\{\mu^s,\Sigma^s\} = f(X,\{V^s\})$ . + +Propagating stochastic embeddings on the graph. Our scoring function $h$ takes the stochastic embeddings $Z^s$ for leaf nodes in $\mathcal{G}$ (provided by $f$ ), further computes the embeddings for all nodes in $\mathcal{G}$ , and finally decodes a quality score $Y$ from the embedding $Z^r$ of the root node $V^r$ . To this end, we propose an extension of graph neural networks (GNNs), in which stochastic embeddings $Z^s$ are propagated from leaf nodes $V^s$ to the root node $V^r$ based on the graph structured informed by the scoring rubric of a particular task. Key to this GNN lies in a lightweight MLP that operates on each node, taking as input the embeddings of its direct predecessors, and generating a new embedding that is further propagated to its successors. This scoring function $h$ is thus given by + +$$ +\underbrace {Z ^ {s} \sim \mathcal {N} \left(\mu^ {s} , \Sigma^ {s}\right) , \forall s \in \mathbb {S}} _ {\text {S a m p l i n g f r o m l e a f n o d e s}}; \quad \underbrace {Z ^ {\neg s} = G \left(\Sigma_ {V ^ {j} \in \mathcal {P} (V ^ {\neg s})} Z ^ {j}\right)} _ {\text {P r o p a g a t i n g o n t h e D A G}}; \quad \underbrace {\hat {Y} = \mathcal {S} \left(Z ^ {r}\right)} _ {\text {D e o c o d i n g t h e s c o r e}} \quad (2) +$$ + +where $V^{\neg s}$ denotes a non-leaf node with its embedding $Z^{\neg s}$ . $\mathcal{P}(V^{\neg s})$ is the set of predecessors of $V^{\neg s}$ , $G(\cdot)$ is the MLP aggregating features from predecessors, and $S(\cdot)$ is another MLP decoding the final score $\hat{Y}$ from the root node $V^{r}$ . + +It is important to note that each leaf embedding $Z^s$ is stochastic, characterized by a Gaussian distribution $(p(Z^s | X, \mathcal{G}) = \mathcal{N}(\mu^s, \Sigma^s))$ , with parameters predicted by $f$ . The non-leaf embeddings are however deterministic given samples from leaf distributions. This design is motivated by our assumption of the scoring rubric, where uncertainty lies only in assessing action steps and identical individual scores will yield the same final score. + +# 3.3 Learning with Variational Information Bottleneck + +With stochastic embeddings, training of RICA $^2$ is a challenge. We design a training scheme under the variational information bottleneck framework. + +Variational information bottleneck (VIB). To train our model $p(Y|X, \mathcal{G})$ with stochastic step embeddings $\{Z^s\}$ , we adopt the information bottleneck principle [36], leading to the maximization of the following objective + +$$ +I \left(\left\{Z ^ {s} \right\}; Y | \mathcal {G}\right) - \beta I \left(\left\{Z ^ {s} \right\}; X | \mathcal {G}\right), \tag {3} +$$ + +where $I$ is the conditional mutual information, and $\beta > 0$ controls the tradeoff between the sufficiency of using step embeddings $\{Z^s\}$ for predicting $Y$ given $\mathcal{G}$ , and the size of the embeddings $\{Z^s\}$ derived from $X$ and $\mathcal{G}$ . + +While mutual information is computationally intractable for high dimensional $\{Z^s\}$ , a common solution [1] is to assume Markov property ( $p(Z|X,Y,\mathcal{G}) = p(Z|X,\mathcal{G})$ ) and conditional independence ( $p(\{Z^s\}|X,\mathcal{G}) = \prod_s p(Z^s|X,\mathcal{G})$ ), followed by the variational approximation for a tractable lower bound + +$$ +- \mathcal {L} _ {\mathrm {V I B}} = \mathbb {E} _ {Z ^ {s} \sim p (Z ^ {s} | X, \mathcal {G}), \forall s \in \mathbb {S}} \left[ \log p (Y | \{Z ^ {s} \}, \mathcal {G}) \right] - \beta \Sigma_ {s \in \mathbb {S}} \operatorname {K L} \left(p (Z ^ {s} | X, \mathcal {G}) | | p (Z ^ {s} | \mathcal {G})\right), \tag {4} +$$ + +where $p(Y|\{Z^s\}, \mathcal{G})$ is modeled by the scoring function $h$ , KL denotes the Kullback-Leibler divergence, and $p(Z^s|\mathcal{G})$ is an approximate marginal prior. + +VIB loss. The first term in Eq. (4) defines the log-likelihood of the score given the input. By assuming that output scores follow a Gaussian with a fixed variance $\sigma$ , this term can be reduced to a mean squared error (MSE) loss + +$$ +\mathcal {L} _ {M S E} = \frac {1}{N} \Sigma_ {i} ^ {N} \left(\hat {Y} _ {i} - Y _ {i}\right) ^ {2} / \sigma^ {2}, \tag {5} +$$ + +where $Y_{i}$ is the predicted score for a video indexed by $i$ , $\hat{Y}_i$ is the corresponding ground-truth score, and $N$ is the total number of videos in the training set. + +The second term in Eq. (4) regularizes the latent space and encodes prediction uncertainty. By assuming a marginal prior of $\mathcal{N}(0,I)$ for $p(Z^s|\mathcal{G})$ , we have + +$$ +\begin{array}{l} \mathcal {L} _ {K L} = \Sigma_ {s \in \mathbb {S}} K L (\mathcal {N} (\mu^ {s} (x), \Sigma^ {s} (x) | | \mathcal {N} (0, I)) \\ = \frac {1}{2} \Sigma_ {s \in \mathbb {S}} \Sigma_ {j = 1} ^ {D} \left(\left(\mu_ {j} ^ {s}\right) ^ {2} + \left(\sigma_ {j} ^ {s}\right) ^ {2} - \log \left(\sigma_ {j} ^ {s}\right) ^ {2} - 1\right), \tag {6} \\ \end{array} +$$ + +where $\mu_j^s$ and $\sigma_j^s$ , respectively, are the $j$ -th dimension of the mean ( $\mu^s(x)$ ) and variance (square root of the diagonal of $\Sigma^s(x)$ ), for the step $s$ . + +The VIB loss $(\mathcal{L}_{VIB})$ is thus given by + +$$ +\mathcal {L} _ {V I B} = \mathcal {L} _ {M S E} + \beta \mathcal {L} _ {K L}. \tag {7} +$$ + +$\mathcal{L}_{VIB}$ consists of (a) the MSE loss $\mathcal{L}_{MSE}$ from the negative log-likelihood of the predicted scores, aiming at minimizing prediction errors; and (b) the KL divergence $\mathcal{L}_{KL}$ between the predicted Gaussian and the prior, regularizing the stochastic embeddings. Further, the coefficient $\beta$ balances between two loss terms. + +During training, samples are drawn to compute the loss function. The output is matched to the Gaussian distribution with its mean equal to the average of the judge scores. The reparameterization trick [14] is used to allow the backpropagation of gradients through the sampling process. + +Estimating uncertainty. The diagonal covariance $\Sigma^s (X)$ models the uncertainty of the predicted quality score of a step $s$ for an input video $X$ . A larger + +value in its diagonal represents a wider distribution of scores and, hence, a lower confidence in the prediction. Following [16,22] we generate uncertainty scores by summing up the harmonic means of the predicted variances for individual steps + +$$ +\operatorname {u n c e r t a i n t y} (Y) = \Sigma_ {s \in \mathbb {S}} D / \Sigma_ {j = 1} ^ {D} \left(\sigma_ {j} ^ {s}\right) ^ {- 1}, \tag {8} +$$ + +where $D$ is the dimensionality of the stochastic embeddings. Again, $\sigma_{j}^{s}$ is the $j$ -th dimension of the predicted variance. + +Stochastic vs. deterministic modeling. An interesting variant of $\mathrm{RICA}^2$ is to disable its stochastic component. Conceptually, this is equal to considering step embeddings $Z^r$ as vectors and removing the KL loss $\mathcal{L}_{KL}$ . We refer to this deterministic version of our model as $\mathrm{RICA}^{2\dagger}$ . Without stochastic embeddings, $\mathrm{RICA}^{2\dagger}$ is unable to estimate prediction uncertainty, yet often yields slightly lower prediction errors. This trade-off is also observed in prior works [6, 16, 32]. We include this variant of our model in the experiments. + +# 3.4 Model Instantiation and Implementation + +Video and step representation. For an input video $X$ , we adapt a pre-trained video backbone(e.g., I3D [4]) to extract its clip-level features, which are further pooled to produce video features $(x_{1}, x_{2}, \ldots, x_{T})$ with fixed-length $T$ . To represent action steps, we make use of a pre-trained language model [7] (Flan-T5) to extract text features from their step descriptions, resulting in an ordered set of text embeddings $(s_{1}, s_{2}, \ldots, s_{K})$ for $K$ steps. Note that the language model is not part of RICA². It is used solely to extract embeddings for text descriptions of the action steps (see supplement Tables I-L). + +**Embedding function $f$ .** Our embedding function $f$ is realized using a Transformer model [37] (see Fig. 2(c)). $f$ first processes video features $(x_{1}, x_{2}, \ldots, x_{T})$ with a self-attention block and text embeddings $(s_{1}, s_{2}, \ldots, s_{K})$ using a MLP. It further makes use of cross-attention blocks $(2\mathbf{x})$ to fuse video and text features, where video features are used to compute keys and values, and text embeddings of steps are projected into queries. Further, $f$ decodes stochastic embeddings of individual steps by predicting a mean vector $\mu^{s} \in \mathbb{R}^{D}$ and a diagonal covariance vector $\Sigma^{s} \in \mathbb{R}^{D}$ for each step $s$ . + +Scoring function $h$ . With Gaussian distributions for all steps specified by $\{\mu^s, \Sigma^s\}$ , we encode the steps and score rubric into a video-specific DAG $\mathcal{G}$ , and realize $h$ as a GNN defined on $\mathcal{G}$ following Eq. (2). $h$ is parameterized by its aggregation function $G$ , which is shared among nodes of the same type. $G$ is implemented using an averaging operation followed by a MLP (2 layers). Finally, $h$ decodes the final score at the root note $V^r$ of $\mathcal{G}$ . + +Training with auxiliary losses. While the VIB loss (Eq. (7)) is sufficient for training, it falls short of considering the temporal ordering of steps. This is because of the conditional independence assumption needed for the derivation of VIB, i.e., $p(\{Z^s\} | X, \mathcal{G}) = \prod_s p(Z^s | X, \mathcal{G})$ , where the ordering of $\{Z^s\}$ is discarded. To bridge the gap, we incorporate an auxiliary loss term $\mathcal{L}_{Aux}$ inspired by [2]. + +Specifically, we re-purpose the last cross-attention map $(\mathbb{R}^{K\times T})$ from $f$ as a step detector. This is done by computing a temporally-weighted center across the attention of each action step to every video time step (i.e., column-wise). We then enforce that (a) this center is co-located with the peak of the attention along video time steps using a sparsity loss [2]; and (b) all centers follow the temporal ordering of corresponding action steps using a ranking loss [2]. These two terms are summed up as the auxiliary loss, and further added to the VIB loss with a small weight (0.1). In our ablation, we empirically verify that adding the auxiliary loss leads to a minor performance boost. + +Inference with sampling. At the inference time, we enhance robustness by sampling 20 times and averaging their predictions to compute the final score. + +# 4 Experiments and Results + +Datasets. Our evaluations are primarily reported on three publicly available benchmark datasets, namely FineDiving [44], MTL-AQA [24], and JIGSAWS [10] in the main paper. In supplement Sec. B, we also include results on the Cataract-101 [31] with cataract surgery videos. + +Evaluation metrics. For all our experiments, we consider metrics on both the accuracy of the prediction and the calibration of the uncertainty estimates. + +- For accuracy, we use two widely adopted metrics for AQA [2,35,44], namely Spearman's rank correlation (SRCC) and relative L2 distance $(R\ell_{2})$ . SRCC measures how well the predicted scores are ranked w.r.t. the ground truth, while $R\ell_{2}$ summarizes the prediction errors. A model with more accurate predictions will have higher SRCC and lower $R\ell_{2}$ . + +- For calibration, we report the uncertainty versus error curve following [16, 22]. To plot this curve, test samples are sorted by increasing uncertainty and divided into 10 equal-sized bins. The mean absolute error (MAE) is then computed for items in each bin. We also follow [16, 22] in employing Kendall's tau ( $\tau$ ) [13], a numerical measure ranging from -1 to 1 to quantify the correlation between the uncertainties and the prediction errors. A higher $\tau$ indicates better calibration, signifying that a model's uncertainty better aligns with prediction errors. + +Baselines. RICA $^2$ is benchmarked against a set of strong baselines, including exemplar-free methods such as DAE [48], UsDL and MUSDL [35], and exemplar-based ones such as CoRE [47], TPT [2] and TSA [44]. We further include the deterministic version of our model $\mathrm{RICA}^{2\dagger}$ , which trades the ability of uncertainty estimation for a minor boost in accuracy. Several baselines adopt a direct regression approach, without providing a confidence or uncertainty measure for predictions. UsDL [35] and TPT [2] implicitly offer a confidence value. In these works, the probability of the predicted score bin serves as a proxy for uncertainty, computed as (1.0-confidence). DAE [48] outputs the standard deviation of the score distribution, which represents uncertainty. + +We seek to ensure a fair comparison yet recognize that methods in our benchmark may consider different settings and/or various types of input. Most + +prior exemplar-free methods only consider a video as input. While $\mathrm{RICA}^2$ does not utilize exemplars, it takes additional input of step information, i.e., step presence and their temporal ordering. On the other hand, previous exemplar-based methods also require the step information as used by $\mathrm{RICA}^2$ , in addition to an input video and an exemplar database. Notably, step information is used to select exemplars, leading to improved results. For example, for diving videos, CoRE [47], TPT [2] and TSA [44] use the diving number (DN) encoding steps and their ordering. Further, TSA [44] also requires the timing of individual steps during training. While it is infeasible to standardize the settings of all methods, we compare to the best reported results in our experiments. + +# 4.1 Results on FineDiving + +Dataset. FineDiving [44] is the largest public dataset for AQA, with 3000 video samples capturing various diving actions. The dataset covers 52 different action types, 29 sub-action types, and 23 difficulty degree types, providing a rich and diverse set of examples for AQA. While this dataset contains temporal annotations for the steps, which can be used to improve the performance of AQA as demonstrated in [44], we do not use these annotations for $\mathrm{RICA}^2$ . + +Experiment setup. We adhere to the experimental setup of the most recent baseline [44] using their train-test split, with 2251 videos for training and 749 videos for testing. We follow the input video settings used in [44] for $\mathrm{RICA}^2$ and the baselines. Specifically, for each video, we uniformly sample 96 frames, which are segmented into 9 overlapping clips, each containing 16 consecutive frames. We refer to supplement Sec. A.1 for further implementation details. + +Results. Tab. 1a presents our results on FineDiving. Both our stochastic and deterministic versions (RICA $^2$ and RICA $^2\dagger$ ) outperform the state-of-the-art TPT [2], an exemplar-based method. RICA $^2$ shows a relative margin of $0.7\%$ / $1.4\%$ on SRCC / $R\ell_2$ , and RICA $^2\dagger$ has a relative margin of $0.9\%$ / $9.6\%$ on SRCC / $R\ell_2$ . This improvement is more pronounced when compared with the exemplar-free methods (MUSDL, DAE-MT) showcasing a significant relative gain of $4.9\%$ , $1.5\%$ on SRCC and $29.8\%$ , $21.6\%$ on $R\ell_2$ . While the deterministic RICA $^2\dagger$ has slightly higher accuracy, our stochastic RICA $^2$ demonstrates superior calibration of its uncertainty estimate ( $\tau_{\mathrm{RICA}^2} = 0.64$ vs. $\tau_{\mathrm{TPT}} = -0.56$ ). Fig. 3a further shows uncertainty calibration results. Uncertainty estimates from RICA $^2$ have a clear upward trend, indicating a higher calibration level. While MUSDL [35] also exhibits a reasonable level of calibration ( $\tau_{\mathrm{MUSDL}} = 0.47$ vs. $\tau_{\mathrm{RICA}^2} = 0.64$ ), the errors are significantly higher than RICA $^2$ across all uncertainty levels. + +# 4.2 Results on MTL-AQA + +Dataset. MTL-AQA [24] is one of the most commonly used datasets for AQA. It consists of 1412 samples collected from 16 events with diverse views. The dataset has a rich set of annotations, including the steps performed during the dive, the difficulty score associated with the dive, and the individual judge scores. + +Table 1: Main results on (a) FineDiving and (b) MTL-AQA. Prediction accuracy $(SRCC$ and $R\ell_{2}$ ) and uncertainty calibration $(\tau)$ metrics are reported. We compare our method with exemplar-based and exemplar-free baselines. +(a) Results on FineDiving + +
Metrics
SRCC(↑)Rl2(↓)τ(↑)
Exemplar basedCoRe [47]0.90610.3615-
TSA [44]0.92030.3420-
TPT [2]0.93330.2877-0.5556
USDL [35,44]0.89130.38220.3778
MUSDL [35,44]0.89780.37040.4667
Exemplar freeDAE [48]0.88200.4919-0.1999
DAE-MT [48]0.92850.3320-0.4667
RICA2 (Ours)0.94020.28380.6444
RICA2† (Ours)0.94210.2600-
+ +(b) Results on MTL-AQA + +
Metrics
SRCC(†)Rℓ2(↓)τ(↑)
Exemplar basedTSA-Net [40]0.9422--
CoRe [47]0.95120.2600-
DAE-CoRe [48]0.9589--
Exemplar freeTPT [2]0.96070.2378-0.1111
C3D-AVG-MTL [24]0.9044--
USDL [35]0.92310.46800.1556
MUSDL [35]0.92730.4510-0.0667
DAE [48]0.9231--
DAE-MT [48]0.94900.2738-0.4222
RICA2 (Ours)0.95940.25800.6000
RICA2† (Ours)0.96200.2280-
+ +![](images/e61eb580869c267e53b222e1b379631aec9e94dfaf43a0046247600442b7a6b6.jpg) +(a) Uncertainty calibration on FineDiving + +![](images/ed416ef7a3939acf35939cfd44ad642e6d94b6e40b82dfd61ed88849b2fc2a6a.jpg) +(b) Uncertainty calibration on MTL-AQA +Fig. 3: Uncertainty vs. prediction error (MAE) on (a) Finediving and (b) MTL-AQA. Results are reported on the test splits, with the X-axis as the uncertainty bin index (uncertainty increases from left to right) and the Y-axis as the MAE in the bin. In comparison to baselines, $\mathrm{RICA}^2$ has improved calibration with lower prediction errors. + +Experiment setup. We follow the evaluation protocol of [2,43,47], dividing the dataset into the standard train set of 1059 videos and a test set of 353 videos. Further, we use the same input video settings as TPT [2] in our experiments to ensure a fair comparison. Specifically, for each video, we uniformly sample 103 frames segmented into 20 overlapping clips, each containing 8 continuous frames. Please refer to supplement Sec. A.2 for more details. + +Results. Tab. 1b summarizes our results on MTL-AQA. Similar to FineDiving, $\mathrm{RICA}^2$ shows state-of-the-art results on MTL-AQA across all evaluation metrics. Specifically, our $\mathrm{RICA}^2\dagger$ outperforms the best exemplar-free model (DAE-MT) by a relative margin of $1.4\% / 16.7\%$ on $SRCC / R\ell_2$ . When compared with the competitive exemplar-based TPT, our has slightly better $SRCC$ ( $SRCC_{TPT} = 0.9607$ vs $SRCC_{\mathrm{RICA}^2\dagger} = 0.9620$ ) and $R\ell_2$ ( $+4.1\%$ relative margin). Again, compared to previous methods, our stochastic model shows improved calibration ( $\tau_{\mathrm{RICA}^2} = 0.60$ vs. $\tau_{USDL} = 0.16$ vs. $\tau_{TPT} = -0.11$ ) as shown in Fig. 3b. + +Table 2: Results on JIGSAWS [10] dataset. Only prediction accuracy ( $SRCC$ ) is considered due to the limited sample size. RICA $^2$ outperforms all prior approaches. + +
Task
SNPKTAvg
Exemplar basedCoRe [47]0.840.860.860.85
TPT [2]0.880.880.910.89
ST-GCN [46]0.310.390.580.43
TSN [39]0.340.230.720.46
JRG [23]0.360.540.750.57
Exemplar freeUSDL [35]0.640.630.610.63
MUSDL [35]0.710.690.710.70
DAE [48]0.730.720.720.72
DAE-MT [48]0.780.740.740.76
RICA2† (Ours)0.880.930.880.90
RICA2 (Ours)0.920.940.900.92
+ +# 4.3 Results on JIGSAWS + +Dataset. In addition to diving videos, we also evaluate RICA² on JIGSAWS [10] — a robotic surgical video dataset. The dataset includes three tasks: "Suturing (S)," with 39 recordings, "Needle Passing (NP)," with 26 recordings and "Knot Tying (KT)" with 36 recordings. JIGSAWS is widely used for action quality assessment, despite its small scale. + +Experiment setup. Due to the limited number of samples in the dataset (as few as 7 videos in the test set), cross-validation is often considered for evaluation on JIGSAWS. To ensure a fair comparison, we follow the commonly adopted splits from [35], and the input video setting from [2]. Specifically, for each video, we uniformly sample 160 frames which are segmented into 20 non-overlapping clips. We opt to not include score calibration curves due to the limited sample size of the test sets. Additionally, the key steps in JIGSAWS are general motions (e.g. reaching for the needle, orienting the needle, etc.) and thus cannot be localized to any specific section of the video. Thus, we do not use the auxiliary losses for this experiment. More details are described in supplement Sec. A.3. + +Results. Tab. 2 summarizes our results on JIGSAWS. Similar to previous datasets, our models exhibit notable advancements over the previous state-of-the-art model TPT [35], showcasing substantial improvements of $1.1\%$ (RICA²) to $3.4\%$ (RICA²†) in terms of average SRCC relative to the exemplar-based state-of-the-art TPT [2]. When compared to the exemplar-free methods, our approach demonstrates an impressive $18.4\%$ (RICA²) to $21.0\%$ (RICA²†) relative gain in average SRCC compared to the latest method DAE-MT [48]. + +# 4.4 Ablation Studies + +To understand our model design choices, we conduct ablation studies on the MTL-AQA [24] dataset. Additional ablations are in supplement Sec. C.2. + +Experiment setup. To simplify our experiments, we opt for running our ablations using fixed I3D features. This allows us to precisely evaluate the + +Table 3: Ablation studies of model components on MTL-AQA dataset. * indicates that the text embeddings were frozen during training. + +
Step Rep.DAG (Rubric)KLAuxMetrics
SRCC(↑)Rl2(↓)τ(↑)Avg. Rank (↓)
Random×××0.94260.3882-5.50
Text×××0.94310.3509-4.50
Text✓××0.94300.3336-4.50
Text*✓××0.94370.3335-3.50
Text*✓✓×0.94480.33290.42221.83
Text*✓✓✓0.94600.33030.42221.17
+ +contribution of different components of our model. Specifically, we choose I3D weights from an intermediate checkpoint of our trained model and extract features for all videos with the frozen backbone. + +Base model. Our ablation constructs a base model using randomly initialized step embeddings, an averaging of these embeddings after cross attention with the video features, followed by an MLP for scoring. This base model is trained using only the MSE loss. We then gradually add modules from $\mathrm{RICA}^2$ and study their effects. Tab. 3 presents our results using the same features and training epochs, with our base model in row 1. + +Text embeddings as step representations. We first replace randomly initialized step representations with the text embeddings of step descriptions. This leads to a major boost in $R\ell_{2}$ (Tab. 3 row 1 vs. row 2), by leveraging knowledge encoded in the LLM [7]. Further, we find that freezing the text embeddings leads to comparable results and faster convergence (Tab. 3 row 3 vs. row 4). + +Does the scoring rubric help? We also investigate the effects of encoding steps and rubric as a DAG—a key design of our model. Adding the DAG results in a noteworthy boost in $R\ell_{2}$ (Tab. 3 row 2 vs. row 3). This improvement can be ascribed to the DAG's proficiency in dissecting the action quality across steps. + +Effects of loss functions. We now study the loss terms. Our loss function has three terms (a) the MSE loss $(\mathcal{L}_{MSE})$ to minimize prediction error, (b) the KL loss $(\mathcal{L}_{KL})$ to regularize the stochastic embeddings, and (c) the auxiliary loss $(\mathcal{L}_{Aux})$ to ensure temporal ordering of steps. Adding the KL loss $\mathcal{L}_{KL}$ yields similar results in $SRCC$ and $R\ell_{2}$ , yet enables calibrated uncertainty estimation. Further attaching the auxiliary loss $\mathcal{L}_{Aux}$ leads to improvement in both $SRCC$ and $R\ell_{2}$ , while maintaining the calibration performance. + +Evaluating the cross-attention maps. To gain insight into $\mathrm{RICA}^2$ , we now examine the cross-attention maps between the step representations and video features in our learned embedding function $f$ . Fig. 4 visualizes the attention map on two test videos on FineDiving. These maps reveal that a step representation is likely to attend to video features during which the step occurs, indicating that $\mathrm{RICA}^2$ learns to encode the temporal location of individual steps. + +We further evaluate this localization ability following the Pointing game protocol [49], widely considered in weakly supervised / unsupervised localization tasks [42,52]. Pointing Game compares a generated heatmap with an annotated + +![](images/12bae1831dc4a12c4483acf22d88943f6e48bbff84366f60663f5a0aeb9a87d9.jpg) +Fig. 4: Visualization of the cross-attention maps. Y-axis : attention value; Y-axis: clip indices (time). Each curve shows an attention map from a step representation to the temporal video features. The frames shown above are aligned with the timing of the corresponding attention plot. Curves and steps are colored accordingly. + +time interval and counts the chance of the heatmap's peak falling into the specified interval. Our evaluation focuses on the FineDiving dataset since it is the only dataset providing annotated time intervals for individual steps. When evaluated on the full test set, attention maps from our model attain an accuracy of $61.4\%$ in the Pointing game protocol, significantly outperforming the chance level accuracy of $30.7\%$ (given each video has 3.26 steps on average). Note that we did not use any annotated segmentation data for training. + +# 5 Conclusion and Discussion + +In this paper, we present a deep probabilistic model for action quality assessment in videos. Our key innovation is to integrate score rubrics and to model prediction uncertainty. Specifically, we propose to adapt stochastic embeddings to quantify the uncertainty of individual steps, and to decode action scores using a variant of graph neural network operating on a DAG encoding the score rubric. Our method offers an exemplar-free approach for AQA, achieves new state-of-the-art results in terms of prediction accuracy on public benchmarks, and demonstrates superior calibration of the output uncertainty estimates. We believe that our work provides a solid step towards AQA. We hope that our method and findings can shed light on the challenging problem of trustworthy video recognition. + +Acknowledgement: This work was supported by the UW Madison Office of the Vice Chancellor for Research with funding from the Wisconsin Alumni Research Foundation, by National Science Foundation under Grant No. CNS 2333491, and by the Army Research Lab under contract number W911NF-2020221. + +# References + +1. Alemi, A.A., Fischer, I., Dillon, J.V., Murphy, K.: Deep variational information bottleneck. In: International Conference on Learning Representations (2016) +2. Bai, Y., Zhou, D., Zhang, S., Wang, J., Ding, E., Guan, Y., Long, Y., Wang, J.: Action quality assessment with temporal parsing transformer. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IV. pp. 422-438. Springer (2022) +3. Battaglia, P.W., Hamrick, J.B., Bapst, V., Sanchez-Gonzalez, A., Zambaldi, V., Malinowski, M., Tacchetti, A., Raposo, D., Santoro, A., Faulkner, R., et al.: Relational inductive biases, deep learning, and graph networks. arXiv preprint arXiv:1806.01261 (2018) +4. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017) +5. Chen, C.H., Hu, Y.H., Yen, T.Y., Radwin, R.G.: Automated video exposure assessment of repetitive hand activity level for a load transfer task. Human factors 55(2), 298-308 (2013) +6. Chun, S., Oh, S.J., De Rezende, R.S., Kalantidis, Y., Larlus, D.: Probabilistic embeddings for cross-modal retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8415-8424 (2021) +7. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., Webson, A., Gu, S.S., Dai, Z., Suzgun, M., Chen, X., Chowdhery, A., Castro-Ros, A., Pellat, M., Robinson, K., Valter, D., Narang, S., Mishra, G., Yu, A., Zhao, V., Huang, Y., Dai, A., Yu, H., Petrov, S., Chi, E.H., Dean, J., Devlin, J., Roberts, A., Zhou, D., Le, Q.V., Wei, J.: Scaling instructionfinetuned language models. Journal of Machine Learning Research 25(70), 1-53 (2024), http://jmlr.org/papers/v25/23-0870.html +8. Duvenaud, D.K., Maclaurin, D., Iparraguirre, J., Bombarell, R., Hirzel, T., Aspuru-Guzik, A., Adams, R.P.: Convolutional networks on graphs for learning molecular fingerprints. Advances in neural information processing systems 28 (2015) +9. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 6202-6211 (2019). https://doi.org/10.1109/ICCV.2019.00630 +10. Gao, Y., Vedula, S.S., Reiley, C.E., Ahmidi, N., Varadarajan, B., Lin, H.C., Tao, L., Zappella, L., Bejar, B., Yuh, D.D., et al.: Jhu-isi gesture and skill assessment working set (jigsaws): A surgical activity dataset for human motion modeling. In: Modeling and Monitoring of Computer Assisted Interventions (M2CAI) - MICCAI Workshop (2014) +11. Gordon, A.S.: Automated video assessment of human performance. In: Proceedings of AI-ED. vol. 2 (1995) +12. Guo, C., Pleiss, G., Sun, Y., Weinberger, K.Q.: On calibration of modern neural networks. In: International conference on machine learning. pp. 1321-1330. PMLR (2017) + +13. Kendall, M.G.: A new measure of rank correlation. Biometrika 30(1/2), 81-93 (1938) +14. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. In: International Conference on Learning Representations (2014) +15. Kipf, T.N., Welling, M.: Semi-supervised classification with graph convolutional networks. In: International Conference on Learning Representations (2017), https://openreview.net/forum?id=SJU4ayYgl +16. Li, W., Huang, X., Lu, J., Feng, J., Zhou, J.: Learning probabilistic ordinal embeddings for uncertainty-aware regression. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13896-13905 (2021) +17. Likert, R.: A technique for the measurement of attitudes. Archives of psychology (1932) +18. Liu, D., Li, Q., Jiang, T., Wang, Y., Miao, R., Shan, F., Li, Z.: Towards unified surgical skill assessment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9522-9531 (2021) +19. Martin, J., Regehr, G., Reznick, R., Macrae, H., Murnaghan, J., Hutchison, C., Brown, M.: Objective structured assessment of technical skill (osats) for surgical residents. British journal of surgery 84(2), 273-278 (1997) +20. Matsuyama, H., Kawaguchi, N., Lim, B.Y.: Iris: Interpretable rubric-informed segmentation for action quality assessment. In: Proceedings of the 28th International Conference on Intelligent User Interfaces. pp. 368-378 (2023) +21. Neelakantan, A., Shankar, J., Passos, A., McCallum, A.: Efficient non-parametric estimation of multiple embeddings per word in vector space. In: Moschitti, A., Pang, B., Daelemans, W. (eds.) Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing, EMNLP 2014, October 25-29, 2014, Doha, Qatar, A meeting of SIGDAT, a Special Interest Group of the ACL. pp. 1059-1069. ACL (2014) +22. Oh, S.J., Gallagher, A.C., Murphy, K.P., Schroff, F., Pan, J., Roth, J.: Modeling uncertainty with hedged instance embeddings. In: International Conference on Learning Representations (2019), https://openreview.net/forum?id=r1xQQhAqKX +23. Pan, J.H., Gao, J., Zheng, W.S.: Action assessment by joint relation graphs. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6331-6340 (2019) +24. Parmar, P., Morris, B.T.: What and how well you performed? a multitask learning approach to action quality assessment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 304-313 (2019) +25. Parmar, P., Tran Morris, B.: Learning to score olympic events. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops. pp 20-28 (2017) +26. Pirsiavash, H., Vondrick, C., Torralba, A.: Assessing the quality of actions. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13. pp. 556-571. Springer (2014) +27. Prassas, S., Kwon, Y.H., Sands, W.A.: Biomechanical research in artistic gymnastics: a review. Sports Biomechanics 5(2), 261-291 (2006) +28. Qiu, Y., Wang, J., Jin, Z., Chen, H., Zhang, M., Guo, L.: Pose-guided matching based on deep learning for assessing quality of action on rehabilitation training Biomedical Signal Processing and Control 72, 103323 (2022) +29. Santoro, A., Raposo, D., Barrett, D.G., Malinowski, M., Pascanu, R., Battaglia, P., Lillicrap, T.: A simple neural network module for relational reasoning. Advances in neural information processing systems 30 (2017) + +30. Scarselli, F., Gori, M., Tsoi, A.C., Hagenbuchner, M., Monfardini, G.: The graph neural network model. IEEE transactions on neural networks 20(1), 61-80 (2008) +31. Schoeffmann, K., Taschwer, M., Sarny, S., Munzer, B., Primus, M.J., Putzgruber, D.: Cataract-101: video dataset of 101 cataract surgeries. In: Cesar, P., Zink, M., Murray, N. (eds.) Proceedings of the 9th ACM Multimedia Systems Conference, MMSys 2018, Amsterdam, The Netherlands, June 12-15, 2018. pp. 421-425. ACM (2018) +32. Shi, Y., Jain, A.K.: Probabilistic face embeddings. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6902-6911 (2019) +33. Sohn, K., Lee, H., Yan, X.: Learning structured output representation using deep conditional generative models. Advances in neural information processing systems 28 (2015) +34. Sun, J.J., Zhao, J., Chen, L.C., Schroff, F., Adam, H., Liu, T.: View-invariant probabilistic embedding for human pose. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16. pp. 53-70. Springer (2020) +35. Tang, Y., Ni, Z., Zhou, J., Zhang, D., Lu, J., Wu, Y., Zhou, J.: Uncertainty-aware score distribution learning for action quality assessment. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9839-9848 (2020) +36. TISHBY, N.: The information bottleneck method. In: Proc. of the 37th Allerton Conference on Communication and Computation, 1999 (1999) +37. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) +38. Vilnis, L., McCallum, A.: Word representations via gaussian embedding. In: International Conference on Learning Representations (2015) +39. Wang, L., Xiong, Y., Wang, Z., Qiao, Y., Lin, D., Tang, X., Van Gool, L.: Temporal segment networks: Towards good practices for deep action recognition. In: European conference on computer vision. pp. 20-36. Springer (2016) +40. Wang, S., Yang, D., Zhai, P., Chen, C., Zhang, L.: Tsa-net: Tube self-attention network for action quality assessment. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 4902-4910 (2021) +41. Waters, T.R., Putz-Anderson, V., Garg, A.: Applications manual for the revised niosh lifting equation (1994) +42. Xiao, F., Sigal, L., Jae Lee, Y.: Weakly-Supervised Visual Grounding of Phrases With Linguistic Structures. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5945-5954 (2017) +43. Xu, C., Fu, Y., Zhang, B., Chen, Z., Jiang, Y.G., Xue, X.: Learning to score figure skating sport videos. IEEE transactions on circuits and systems for video technology 30(12), 4578-4590 (2019) +44. Xu, J., Rao, Y., Yu, X., Chen, G., Zhou, J., Lu, J.: Finediving: A fine-grained dataset for procedure-aware action quality assessment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2949-2958 (2022) +45. Xu, K., Li, J., Zhang, M., Du, S.S., ichi Kawarabayashi, K., Jegelka, S.: What can neural networks reason about? In: International Conference on Learning Representations (2020), https://openreview.net/forum?id=rJxbJeHFPS +46. Yan, S., Xiong, Y., Lin, D.: Spatial temporal graph convolutional networks for skeleton-based action recognition. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018) + +47. Yu, X., Rao, Y., Zhao, W., Lu, J., Zhou, J.: Group-aware contrastive regression for action quality assessment. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7899-7908. IEEE Computer Society, Los Alamitos, CA, USA (oct 2021) +48. Zhang, B., Chen, J., Xu, Y., Zhang, H., Yang, X., Geng, X.: Auto-encoding score distribution regression for action quality assessment. Neural Computing and Applications pp. 1-14 (2023) +49. Zhang, J., Bargal, S.A., Lin, Z., Brandt, J., Shen, X., Sclaroff, S.: Top-Down Neural Attention by Excitation Backprop. International Journal of Computer Vision 126(10), 1084-1102 (Oct 2018) +50. Zhou, C., Huang, Y.: Uncertainty-driven action quality assessment. arXiv preprint arXiv:2207.14513 (2022) +51. Zhou, K., Ma, Y., Shum, H.P.H., Liang, X.: Hierarchical graph convolutional networks for action quality assessment. IEEE Transactions on Circuits and Systems for Video Technology pp. 1-1 (2023) +52. Zhu, Y., Zhou, Y., Ye, Q., Qiu, Q., Jiao, J.: Soft Proposal Networks for Weakly Supervised Object Localization. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1841-1850 (2017) \ No newline at end of file diff --git a/2024/RICA^2_ Rubric-Informed, Calibrated Assessment of Actions/images.zip b/2024/RICA^2_ Rubric-Informed, Calibrated Assessment of Actions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..047ffa6868d3660e75e6d26e570e28ef8630ef5d --- /dev/null +++ b/2024/RICA^2_ Rubric-Informed, Calibrated Assessment of Actions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6eadc4a198b9f07dc2c98143da1b707eea389b4a4cf003ac386f6787ca678a5 +size 359246 diff --git a/2024/RICA^2_ Rubric-Informed, Calibrated Assessment of Actions/layout.json b/2024/RICA^2_ Rubric-Informed, Calibrated Assessment of Actions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..109ef360a8aa078916c93a5384d66f93ddb86c49 --- /dev/null +++ b/2024/RICA^2_ Rubric-Informed, Calibrated Assessment of Actions/layout.json @@ -0,0 +1,12507 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 178, + 113, + 434, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 113, + 434, + 148 + ], + "spans": [ + { + "bbox": [ + 178, + 113, + 434, + 148 + ], + "type": "text", + "content": "RICA" + }, + { + "bbox": [ + 178, + 113, + 434, + 148 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 178, + 113, + 434, + 148 + ], + "type": "text", + "content": ": Rubric-Informed, Calibrated Assessment of Actions" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 143, + 170, + 470, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 170, + 470, + 194 + ], + "spans": [ + { + "bbox": [ + 143, + 170, + 470, + 194 + ], + "type": "text", + "content": "Abrar Majeedi, Viswanatha Reddy Gajjala, Satya Sai Srinath Namburi GNVV, and Yin Li" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 173, + 203, + 439, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 203, + 439, + 226 + ], + "spans": [ + { + "bbox": [ + 173, + 203, + 439, + 226 + ], + "type": "text", + "content": "University of Wisconsin-Madison, Madison Wisconsin 53706 USA {majeedi,vgajjala,sgnamburi,yin.li}@wisc.edu" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 255, + 452, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 255, + 452, + 419 + ], + "spans": [ + { + "bbox": [ + 159, + 255, + 452, + 419 + ], + "type": "text", + "content": "Abstract. The ability to quantify how well an action is carried out, also known as action quality assessment (AQA), has attracted recent interest in the vision community. Unfortunately, prior methods often ignore the score rubric used by human experts and fall short of quantifying the uncertainty of the model prediction. To bridge the gap, we present RICA² — a deep probabilistic model that integrates score rubric and accounts for prediction uncertainty for AQA. Central to our method lies in stochastic embeddings of action steps, defined on a graph structure that encodes the score rubric. The embeddings spread probabilistic density in the latent space and allow our method to represent model uncertainty. The graph encodes the scoring criteria, based on which the quality scores can be decoded. We demonstrate that our method establishes new state of the art on public benchmarks, including FineDiving, MTL-AQA, and JIGSAWS, with superior performance in score prediction and uncertainty calibration. Our code is available at https://abrarmajeedi.github.io/rica2_aqa/." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 430, + 419, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 430, + 419, + 441 + ], + "spans": [ + { + "bbox": [ + 160, + 430, + 419, + 441 + ], + "type": "text", + "content": "Keywords: Action Quality Assessment " + }, + { + "bbox": [ + 160, + 430, + 419, + 441 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 430, + 419, + 441 + ], + "type": "text", + "content": " Video Understanding" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 461, + 229, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 461, + 229, + 473 + ], + "spans": [ + { + "bbox": [ + 133, + 461, + 229, + 473 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 593 + ], + "type": "text", + "content": "Action quality assessment (AQA), aiming at quantifying how well an action is carried out, has been widely studied across scientific disciplines due to its broad range of applications. AQA is key to sports science and analytics. The right way of performing actions maximizes an athlete's performance and minimizes injury risk. AQA is crucial to occupational safety and health. High-quality actions mitigate the physical stress and strain in the workspace. AQA is pivotal for physical therapies. The quality of actions reveals the progress in rehabilitation. AQA also plays a major role in surgical education. Proficient actions improve the outcome and reduce complications." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Observational methods for AQA have been well established for various tasks, e.g., gymnastics [27], manual material handling [41], and surgery [19]. These methods involve a human expert observing an action and decomposing it into a series of key steps. Each of these steps, or a subset of them, can be grouped into a factor and then evaluated using a Likert scale [17] following a pre-defined criterion. Ratings for individual factors, sometimes complemented with impression-based" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 114, + 478, + 175 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 478, + 175 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 478, + 175 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 478, + 175 + ], + "type": "image", + "image_path": "baf7824ac63e33344dcd192a11f8f7d2052e301734f4f126033754c0a29b7fcb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 183, + 482, + 206 + ], + "lines": [ + { + "bbox": [ + 132, + 183, + 482, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 482, + 206 + ], + "type": "text", + "content": "Fig. 1: RICA2 integrates score rubric used by human experts and accounts for prediction uncertainty, resulting in accurate predictions and calibrated uncertainty estimates." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 231, + 480, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 231, + 480, + 279 + ], + "spans": [ + { + "bbox": [ + 130, + 231, + 480, + 279 + ], + "type": "text", + "content": "global ratings, are then summarized into a final quality score [19]. Multiple expert ratings are often considered to account for the variance in the scores. While these methods are commonly adopted, they require significant input from human experts and are thus costly and inefficient." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 280, + 481, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 280, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 481, + 376 + ], + "type": "text", + "content": "There is a burgeoning interest in the vision community to develop video-based AQA [25, 26, 35, 43]. While current solutions have made steady progress across benchmarks [10, 24, 44], their decision-making processes differ largely from prior observational methods. Almost all prior solutions learn deep models to directly map input videos to scores. Many of them employ an exemplar-based approach, in which a model predicts relative scores by referencing exemplar videos with similar actions and known scores [2, 44, 47]. Few of them have considered the structure of the actions or their scoring criteria used by observational methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 376, + 482, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 484 + ], + "type": "text", + "content": "Further, existing AQA methods face a key challenge in the accurate quantification of model uncertainty i.e., the uncertainty of model prediction that is calibrate to the expected error [12]. Knowing this uncertainty is particularly helpful for AQA, e.g., when assessing the quality of high-stakes competitions or surgical procedures. With proper calibration, videos that have uncertain predictions can be passed to human experts for a thorough evaluation. Several recent works have started to consider the variance among scores from multiple human experts [35,48,50,51]. Unfortunately, they still fall short of considering prediction uncertainty, leaving this challenge largely unaddressed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 485, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 605 + ], + "type": "text", + "content": "To bridge the gap, we develop a deep probabilistic model for AQA by integrating score rubrics and modeling the uncertainty of the prediction (see Fig. 1). Central to our method lies in the stochastic embedding of action steps, defined on a graph structure that encodes the score rubric. The embeddings spread probabilistic density in latent space and allow our method to represent model uncertainty. The graph encodes the scoring criteria, based on which the quality scores can be decoded. We also present a training scheme and describe an approach to estimate uncertainty. Putting things together, our method, dubbed RICA" + }, + { + "bbox": [ + 130, + 485, + 482, + 605 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 130, + 485, + 482, + 605 + ], + "type": "text", + "content": " (Rubric-informed, Calibrated Assessment of Actions), yields accurate action scores with additional uncertainty estimates." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "We evaluate RICA² on several public AQA datasets, covering sports and surgical videos. Particularly, RICA² establishes new state of the art on FineDiving [44], MTL-AQA [24] and JIGSAWS [10]. On FineDiving [44] - the largest and most challenging AQA benchmark, RICA² outperforms latest methods in prediction accuracy (a boost of " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "+0.94\\%" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": " in Spearman's Rank Correlation Coefficient" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "text", + "content": "A. Majeedi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "inline_equation", + "content": "(SRCC))" + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "content": " and demonstrates significantly improved uncertainty calibration (a gain of " + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "inline_equation", + "content": "+0.178" + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "content": " in Kendall Tau [13]). Similarly, on MTL-AQA [24], the most commonly used dataset for AQA, " + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "inline_equation", + "content": "\\mathrm{RICA}^2" + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "content": " attains state-of-the-art " + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "inline_equation", + "content": "SRCC" + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "content": ", and again largely improved calibration (a gain of " + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "inline_equation", + "content": "+0.444" + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "content": " in Kendall Tau). On JIGSAWS [10], " + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "inline_equation", + "content": "\\mathrm{RICA}^2" + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "content": " beats the previous best results by a relative margin of " + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "inline_equation", + "content": "+3.37\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "inline_equation", + "content": "SRCC" + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "content": ". Further, we present extensive experiments to evaluate the key design of " + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "inline_equation", + "content": "\\mathrm{RICA}^2" + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 190, + 405, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 190, + 405, + 201 + ], + "spans": [ + { + "bbox": [ + 146, + 190, + 405, + 201 + ], + "type": "text", + "content": "Our main contributions are summarized into three folds." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 203, + 480, + 326 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 137, + 203, + 479, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 203, + 479, + 239 + ], + "spans": [ + { + "bbox": [ + 137, + 203, + 479, + 239 + ], + "type": "text", + "content": "- We propose " + }, + { + "bbox": [ + 137, + 203, + 479, + 239 + ], + "type": "inline_equation", + "content": "RICA^2" + }, + { + "bbox": [ + 137, + 203, + 479, + 239 + ], + "type": "text", + "content": ", a novel deep probabilistic method that incorporates scoring rubrics and uncertainty modeling for AQA, resulting in accurate scores and calibrated uncertainty estimates." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 240, + 480, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 240, + 480, + 287 + ], + "spans": [ + { + "bbox": [ + 137, + 240, + 480, + 287 + ], + "type": "text", + "content": "- Our technical innovations lie in (a) a graph neural network to model the scoring rubric in conjunction with stochastic embeddings on the graph to account for prediction uncertainty and (b) a training scheme under the variational information bottleneck framework." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 289, + 480, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 289, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 137, + 289, + 480, + 326 + ], + "type": "text", + "content": "- Our extensive set of experiments demonstrates that " + }, + { + "bbox": [ + 137, + 289, + 480, + 326 + ], + "type": "inline_equation", + "content": "\\mathrm{RICA}^2" + }, + { + "bbox": [ + 137, + 289, + 480, + 326 + ], + "type": "text", + "content": " achieves state-of-the-art results in AQA, significantly outperforming prior methods in both prediction accuracy and calibration of uncertainty estimates." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 342, + 237, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 237, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 237, + 355 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 365, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 426 + ], + "type": "text", + "content": "Action quality assessment (AQA). Early works in AQA [11,26] employed handcrafted features to estimate quality scores in videos. More recent methods developed various deep models, including convolutional [35,47], graph [23], recurrent [25, 43], and Transformer [2, 44] networks. AQA has also been widely considered in surgical education [18], rehabilitation [28], and ergonomics [5]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 426, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 533 + ], + "type": "text", + "content": "Recently, exemplar-based methods [2, 44, 47] have emerged as a promising solution for AQA due to their impressive performance across benchmarks. These methods predict the relative score of an input video by comparing it to selected exemplar videos with similar action steps and known scores. A limitation of this paradigm is the requirement of exemplar videos at inference time. This strategy largely deviates from existing observational methods used by human experts and leads to significantly higher computational costs. While RICA² also uses action steps in the input video, it further integrates the scoring rubric of these steps and offers a solution for no-reference AQA i.e. without using exemplars." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 534, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 641 + ], + "type": "text", + "content": "Several recent works have started to consider the modeling of score uncertainty in AQA [35, 48, 50, 51]. For example, Tang et al. [35] proposed to model the final scores using a Gaussian distribution. They presented a model (MUSDL) trained to predict the score distribution. This distribution learning idea was further extended in [48, 50, 51]. However, modeling the score distribution does not warrant the quantification of model uncertainty, as the output distributions might not be calibrated with prediction errors. While RICA² also predicts a Gaussian distribution for the scores, our key design is to consider stochastic embeddings to quantify prediction uncertainty, resulting in calibrated uncertainty estimates." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "The most relevant work is IRIS [20]. IRIS incorporates score rubric into a convolutional network for AQA. This is done by segmenting key steps in the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 204, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 204, + 90, + 448, + 102 + ], + "type": "inline_equation", + "content": "\\mathrm{RICA}^2" + }, + { + "bbox": [ + 204, + 90, + 448, + 102 + ], + "type": "text", + "content": " : Rubric-Informed, Calibrated Assessment of Actions" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 187 + ], + "type": "text", + "content": "video and predicting sub-scores for individual steps. Similar to IRIS, RICA" + }, + { + "bbox": [ + 130, + 115, + 482, + 187 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 130, + 115, + 482, + 187 + ], + "type": "text", + "content": " also considers rubric in a deep model. However, RICA" + }, + { + "bbox": [ + 130, + 115, + 482, + 187 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 130, + 115, + 482, + 187 + ], + "type": "text", + "content": " adapts a graph network, treats sub-scores as latent embeddings, predicts the final score, and further quantifies prediction uncertainty. These differences allow RICA" + }, + { + "bbox": [ + 130, + 115, + 482, + 187 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 130, + 115, + 482, + 187 + ], + "type": "text", + "content": " to be trained on major public datasets with only final scores, and to output calibrated uncertainty estimates, both of which cannot be achieved by IRIS." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 190, + 483, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 483, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 483, + 334 + ], + "type": "text", + "content": "Modeling uncertainty with stochastic embedding. Stochastic embedding, initially introduced in NLP [21,38], treats each embedding as a distribution. This approach has gained recent attention for modeling uncertainty in deep models. Oh et al. [22] considered probabilistic embeddings for metric learning and proposed to model uncertainty based on the stochasticity of embeddings. This idea was further adopted in many vision tasks, including face verification [32], age estimation [16], pose estimation [34], and cross-modal retrieval [6]. Another related line of work is the conditional variational autoencoder [33], where a probabilistic representation of the input is used for a prediction task. Our approach shares a similar idea of using stochastic embeddings to model uncertainty yet is specifically designed for AQA. Our method significantly extends prior idea to embed action steps on a graph structure, and to propagate these stochastic embeddings on the graph." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 336, + 483, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 483, + 408 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 483, + 408 + ], + "type": "text", + "content": "Graph neural networks (GNNs). GNNs [8, 15, 30] offer a powerful tool to leverage the relational inductive bias inherent in data [3, 45]. This inductive bias is beneficial to aggregate a global representation from a group of local ones [29]. Recently, Zhou et al. [51] proposed a hierarchical graph convolutional network for AQA, in which a GNN was used for video representation learning. In contrast, we adapt graph networks to model score rubrics used by observational methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 426, + 458, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 426, + 458, + 440 + ], + "spans": [ + { + "bbox": [ + 131, + 426, + 458, + 440 + ], + "type": "text", + "content": "3 AQA with Score Rubric and Uncertainty Modeling" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 449, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 485 + ], + "type": "text", + "content": "Our goal is to assess the quality of an action within an input video. Let " + }, + { + "bbox": [ + 130, + 449, + 482, + 485 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 449, + 482, + 485 + ], + "type": "text", + "content": " be the video with the action and " + }, + { + "bbox": [ + 130, + 449, + 482, + 485 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 449, + 482, + 485 + ], + "type": "text", + "content": " as its quality score. Our method further considers the structure of the action and a scoring rubric based on the structure." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 488, + 482, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 488, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 488, + 482, + 596 + ], + "type": "text", + "content": "Action steps. We assume that the action in " + }, + { + "bbox": [ + 130, + 488, + 482, + 596 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 488, + 482, + 596 + ], + "type": "text", + "content": " comprises a known, ordered set of key steps, denoted as " + }, + { + "bbox": [ + 130, + 488, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\mathbb{S} = (s_1, s_2, \\dots, s_k)" + }, + { + "bbox": [ + 130, + 488, + 482, + 596 + ], + "type": "text", + "content": ". Each " + }, + { + "bbox": [ + 130, + 488, + 482, + 596 + ], + "type": "inline_equation", + "content": "s^1" + }, + { + "bbox": [ + 130, + 488, + 482, + 596 + ], + "type": "text", + "content": " represents a necessary sub-action for successfully executing the action. Further, " + }, + { + "bbox": [ + 130, + 488, + 482, + 596 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 488, + 482, + 596 + ], + "type": "text", + "content": " is associated with a text description that elucidates the specifics of the corresponding step, e.g., \"a front-facing takeoff\" for diving. This assumption is especially well suited for structured actions, such as diving or surgery, where the key steps are predetermined and follow a specific sequence. Note that the timing of the key steps is not presumed. Even if key steps are unavailable, they can be detected using action recognition methods [4,9] (see supplement Sec. C.4)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": "Scoring rubric. We further assume a pre-specified scoring rubric based on the key steps — a common strategy in technical skill assessment [19, 27, 41]. Specifically, each action step " + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "inline_equation", + "content": "s_k" + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": " is independently scored, i.e., " + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "inline_equation", + "content": "s_k \\mapsto y_k" + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": ". Subsequently, a rule-based rubric is employed to aggregate individual scores " + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\{y_k\\}" + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": " and calculate a final" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "text", + "content": "A. Majeedi et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 459, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 459, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 459, + 665 + ], + "type": "text", + "content": "1 For the sake of brevity, we omit the subscript as long as there is no confusion." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 125, + 246, + 238 + ], + "blocks": [ + { + "bbox": [ + 160, + 114, + 230, + 123 + ], + "lines": [ + { + "bbox": [ + 160, + 114, + 230, + 123 + ], + "spans": [ + { + "bbox": [ + 160, + 114, + 230, + 123 + ], + "type": "text", + "content": "(a) AQA Scoring Rubric" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 141, + 125, + 246, + 238 + ], + "lines": [ + { + "bbox": [ + 141, + 125, + 246, + 238 + ], + "spans": [ + { + "bbox": [ + 141, + 125, + 246, + 238 + ], + "type": "image", + "image_path": "d47ba5f9c11f940abc351e7ed272bf2c05b63595941f5eb066769fa5b21f5631.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 248, + 123, + 365, + 241 + ], + "blocks": [ + { + "bbox": [ + 281, + 114, + 332, + 122 + ], + "lines": [ + { + "bbox": [ + 281, + 114, + 332, + 122 + ], + "spans": [ + { + "bbox": [ + 281, + 114, + 332, + 122 + ], + "type": "text", + "content": "(b) GNN on DAG" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 248, + 123, + 365, + 241 + ], + "lines": [ + { + "bbox": [ + 248, + 123, + 365, + 241 + ], + "spans": [ + { + "bbox": [ + 248, + 123, + 365, + 241 + ], + "type": "image", + "image_path": "8ab449158ac2f05e7496eb8ec143cfd5d9fece82c172f23784d72d9df068b39f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 365, + 123, + 472, + 241 + ], + "blocks": [ + { + "bbox": [ + 380, + 114, + 455, + 123 + ], + "lines": [ + { + "bbox": [ + 380, + 114, + 455, + 123 + ], + "spans": [ + { + "bbox": [ + 380, + 114, + 455, + 123 + ], + "type": "text", + "content": "(c) Embedding Function " + }, + { + "bbox": [ + 380, + 114, + 455, + 123 + ], + "type": "inline_equation", + "content": "\\pmb{f}" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 365, + 123, + 472, + 241 + ], + "lines": [ + { + "bbox": [ + 365, + 123, + 472, + 241 + ], + "spans": [ + { + "bbox": [ + 365, + 123, + 472, + 241 + ], + "type": "image", + "image_path": "0159e32da3c3e14823cbfe67fe915c37302b1c0a685906254e47a461e35648f5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 251, + 482, + 307 + ], + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 307 + ], + "type": "text", + "content": "Fig. 2: Overview of RICA" + }, + { + "bbox": [ + 130, + 251, + 482, + 307 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 130, + 251, + 482, + 307 + ], + "type": "text", + "content": ". Leveraging scoring rubrics (a), RICA" + }, + { + "bbox": [ + 130, + 251, + 482, + 307 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 130, + 251, + 482, + 307 + ], + "type": "text", + "content": " integrates a graph representation of action step and rubric with uncertainty modeling (b). Specifically, RICA" + }, + { + "bbox": [ + 130, + 251, + 482, + 307 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 130, + 251, + 482, + 307 + ], + "type": "text", + "content": " takes an input of the video and its key action steps, encodes the input into embeddings (c), refines the embeddings through a deep probabilistic model, and outputs an action score in tandem with its uncertainty estimate." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "type": "text", + "content": "quality score " + }, + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "type": "inline_equation", + "content": "\\{y_k\\} \\mapsto Y" + }, + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "type": "text", + "content": ", in which steps might be grouped into intermediate stages (see an example in Fig. 2 (a-b)). This rubric follows a deterministic yet often non-injective mapping, e.g. a many-to-one mapping such as summation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 367, + 482, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 482, + 476 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 482, + 476 + ], + "type": "text", + "content": "Method overview. We now present RICA" + }, + { + "bbox": [ + 130, + 367, + 482, + 476 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 130, + 367, + 482, + 476 + ], + "type": "text", + "content": " — a deep probabilistic model for AQA that leverages known action steps and incorporates the scoring rubric for modeling. Importantly, RICA" + }, + { + "bbox": [ + 130, + 367, + 482, + 476 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 130, + 367, + 482, + 476 + ], + "type": "text", + "content": " accounts for prediction uncertainty, i.e., when the model prediction can and cannot be trusted. Fig. 2 presents an overview of RICA" + }, + { + "bbox": [ + 130, + 367, + 482, + 476 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 130, + 367, + 482, + 476 + ], + "type": "text", + "content": ". It consists of two main model components: (a) a graph neural network that integrates the key steps and scoring rubric (Sec. 3.1); and (b) stochastic embeddings defined on the graph to capture prediction uncertainty (Sec. 3.2), coupled with (c) a learning scheme under the variational information bottleneck framework (Sec. 3.3). In what follows, we delve into the details of RICA" + }, + { + "bbox": [ + 130, + 367, + 482, + 476 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 130, + 367, + 482, + 476 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 491, + 453, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 491, + 453, + 504 + ], + "spans": [ + { + "bbox": [ + 131, + 491, + 453, + 504 + ], + "type": "text", + "content": "3.1 Integrating Actions Steps and Scoring Rubric with Graph" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Steps and rubric as graph. We encode action steps and the corresponding scoring rubric with a directed acyclic graph (DAG). This DAG is denoted as " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = (\\mathbb{V},\\mathbb{E})" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{V}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " as the set of nodes and " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{E}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " as the set of directed edges. " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{V}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " consists of three types of nodes: (1) the leaf nodes, denoted as " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "V^s" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ", correspond to individual action steps performed in the input video " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "; (2) the intermediate nodes capturing possible intermediate stages in the scoring criteria; and (3) a designated root node " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "V^r" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " representing the final score of the action. Further, the edges " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{E}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " indicate the scoring rubric, connecting steps (leaf nodes) to stages (intermediate nodes), and stages (intermediate nodes) to the final score (root node). We note that " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " varies for every input video " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " (assuming a single action), as different steps might be performed. Fig. 2 (a-b) show the example in diving where the key steps and scoring rubric are encoded using our DAG. Additional examples can be found in our supplement Fig. B." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 205, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 205, + 90, + 448, + 102 + ], + "type": "inline_equation", + "content": "\\mathrm{RICA}^2" + }, + { + "bbox": [ + 205, + 90, + 448, + 102 + ], + "type": "text", + "content": ": Rubric-Informed, Calibrated Assessment of Actions" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "Learning for quality assessment. Our approach involves a two-step process for quality assessment. First, we employ an embedding function " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": ", designed to map individual steps into a latent space representing action quality. Secondly, we leverage the key step embeddings " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\{Z^s\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " along with the score rubric encoded in " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " to learn a scoring function " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": ". These functions " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " are defined as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 234, + 181, + 482, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 181, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 234, + 181, + 482, + 194 + ], + "type": "interline_equation", + "content": "f: X, \\mathcal {G} \\mapsto \\{Z ^ {s} \\}; \\quad h: \\{Z ^ {s} \\}, \\mathcal {G} \\mapsto Y, \\tag {1}", + "image_path": "fc2aeb137923604341c4d3deaac01396db27086b328070ec83cea4de97580e3b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 198, + 482, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 198, + 482, + 222 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 482, + 222 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 198, + 482, + 222 + ], + "type": "inline_equation", + "content": "\\{Z^s\\}" + }, + { + "bbox": [ + 130, + 198, + 482, + 222 + ], + "type": "text", + "content": " are the embeddings for the set of steps " + }, + { + "bbox": [ + 130, + 198, + 482, + 222 + ], + "type": "inline_equation", + "content": "\\mathbb{S}" + }, + { + "bbox": [ + 130, + 198, + 482, + 222 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 198, + 482, + 222 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 198, + 482, + 222 + ], + "type": "text", + "content": ", corresponding to the leaf nodes " + }, + { + "bbox": [ + 130, + 198, + 482, + 222 + ], + "type": "inline_equation", + "content": "\\{V^s\\}" + }, + { + "bbox": [ + 130, + 198, + 482, + 222 + ], + "type": "text", + "content": " on the DAG." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 237, + 448, + 250 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 237, + 448, + 250 + ], + "spans": [ + { + "bbox": [ + 131, + 237, + 448, + 250 + ], + "type": "text", + "content": "3.2 Modeling Score Uncertainty with Stochastic Embeddings" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 254, + 482, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 254, + 482, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 482, + 339 + ], + "type": "text", + "content": "To model the prediction uncertainty, we adopt stochastic step embeddings defined on the leaf nodes, such that " + }, + { + "bbox": [ + 130, + 254, + 482, + 339 + ], + "type": "inline_equation", + "content": "Z^{s} \\in \\mathbb{R}^{D} \\sim p(Z^{s}|X,\\{V^{s}\\})" + }, + { + "bbox": [ + 130, + 254, + 482, + 339 + ], + "type": "text", + "content": ". Unlike deterministic embeddings, where " + }, + { + "bbox": [ + 130, + 254, + 482, + 339 + ], + "type": "inline_equation", + "content": "Z^{s}" + }, + { + "bbox": [ + 130, + 254, + 482, + 339 + ], + "type": "text", + "content": " would be a fixed vector, stochastic embedding characterizes the distribution of " + }, + { + "bbox": [ + 130, + 254, + 482, + 339 + ], + "type": "inline_equation", + "content": "Z^{s}" + }, + { + "bbox": [ + 130, + 254, + 482, + 339 + ], + "type": "text", + "content": ", allowing for uncertainty control. Specifically, we model " + }, + { + "bbox": [ + 130, + 254, + 482, + 339 + ], + "type": "inline_equation", + "content": "p(Z^{s}|X,V^{s})" + }, + { + "bbox": [ + 130, + 254, + 482, + 339 + ], + "type": "text", + "content": " as a Gaussian distribution in " + }, + { + "bbox": [ + 130, + 254, + 482, + 339 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{D}" + }, + { + "bbox": [ + 130, + 254, + 482, + 339 + ], + "type": "text", + "content": " with mean " + }, + { + "bbox": [ + 130, + 254, + 482, + 339 + ], + "type": "inline_equation", + "content": "\\mu^{s}" + }, + { + "bbox": [ + 130, + 254, + 482, + 339 + ], + "type": "text", + "content": " and diagonal covariance " + }, + { + "bbox": [ + 130, + 254, + 482, + 339 + ], + "type": "inline_equation", + "content": "\\Sigma^{s}" + }, + { + "bbox": [ + 130, + 254, + 482, + 339 + ], + "type": "text", + "content": ". The embedding function " + }, + { + "bbox": [ + 130, + 254, + 482, + 339 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 254, + 482, + 339 + ], + "type": "text", + "content": " is thus tasked to predict the mean and covariance for the key steps " + }, + { + "bbox": [ + 130, + 254, + 482, + 339 + ], + "type": "inline_equation", + "content": "\\mathbb{S}" + }, + { + "bbox": [ + 130, + 254, + 482, + 339 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 254, + 482, + 339 + ], + "type": "inline_equation", + "content": "\\{\\mu^s,\\Sigma^s\\} = f(X,\\{V^s\\})" + }, + { + "bbox": [ + 130, + 254, + 482, + 339 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 340, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 461 + ], + "type": "text", + "content": "Propagating stochastic embeddings on the graph. Our scoring function " + }, + { + "bbox": [ + 130, + 340, + 482, + 461 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 340, + 482, + 461 + ], + "type": "text", + "content": " takes the stochastic embeddings " + }, + { + "bbox": [ + 130, + 340, + 482, + 461 + ], + "type": "inline_equation", + "content": "Z^s" + }, + { + "bbox": [ + 130, + 340, + 482, + 461 + ], + "type": "text", + "content": " for leaf nodes in " + }, + { + "bbox": [ + 130, + 340, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 340, + 482, + 461 + ], + "type": "text", + "content": " (provided by " + }, + { + "bbox": [ + 130, + 340, + 482, + 461 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 340, + 482, + 461 + ], + "type": "text", + "content": "), further computes the embeddings for all nodes in " + }, + { + "bbox": [ + 130, + 340, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 340, + 482, + 461 + ], + "type": "text", + "content": ", and finally decodes a quality score " + }, + { + "bbox": [ + 130, + 340, + 482, + 461 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 340, + 482, + 461 + ], + "type": "text", + "content": " from the embedding " + }, + { + "bbox": [ + 130, + 340, + 482, + 461 + ], + "type": "inline_equation", + "content": "Z^r" + }, + { + "bbox": [ + 130, + 340, + 482, + 461 + ], + "type": "text", + "content": " of the root node " + }, + { + "bbox": [ + 130, + 340, + 482, + 461 + ], + "type": "inline_equation", + "content": "V^r" + }, + { + "bbox": [ + 130, + 340, + 482, + 461 + ], + "type": "text", + "content": ". To this end, we propose an extension of graph neural networks (GNNs), in which stochastic embeddings " + }, + { + "bbox": [ + 130, + 340, + 482, + 461 + ], + "type": "inline_equation", + "content": "Z^s" + }, + { + "bbox": [ + 130, + 340, + 482, + 461 + ], + "type": "text", + "content": " are propagated from leaf nodes " + }, + { + "bbox": [ + 130, + 340, + 482, + 461 + ], + "type": "inline_equation", + "content": "V^s" + }, + { + "bbox": [ + 130, + 340, + 482, + 461 + ], + "type": "text", + "content": " to the root node " + }, + { + "bbox": [ + 130, + 340, + 482, + 461 + ], + "type": "inline_equation", + "content": "V^r" + }, + { + "bbox": [ + 130, + 340, + 482, + 461 + ], + "type": "text", + "content": " based on the graph structured informed by the scoring rubric of a particular task. Key to this GNN lies in a lightweight MLP that operates on each node, taking as input the embeddings of its direct predecessors, and generating a new embedding that is further propagated to its successors. This scoring function " + }, + { + "bbox": [ + 130, + 340, + 482, + 461 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 340, + 482, + 461 + ], + "type": "text", + "content": " is thus given by" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 152, + 464, + 482, + 497 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 464, + 482, + 497 + ], + "spans": [ + { + "bbox": [ + 152, + 464, + 482, + 497 + ], + "type": "interline_equation", + "content": "\\underbrace {Z ^ {s} \\sim \\mathcal {N} \\left(\\mu^ {s} , \\Sigma^ {s}\\right) , \\forall s \\in \\mathbb {S}} _ {\\text {S a m p l i n g f r o m l e a f n o d e s}}; \\quad \\underbrace {Z ^ {\\neg s} = G \\left(\\Sigma_ {V ^ {j} \\in \\mathcal {P} (V ^ {\\neg s})} Z ^ {j}\\right)} _ {\\text {P r o p a g a t i n g o n t h e D A G}}; \\quad \\underbrace {\\hat {Y} = \\mathcal {S} \\left(Z ^ {r}\\right)} _ {\\text {D e o c o d i n g t h e s c o r e}} \\quad (2)", + "image_path": "725a773ee0ed6ba1257518e442466882f334917f490b37376bd095a9c9911296.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 501, + 482, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 482, + 537 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 482, + 537 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 501, + 482, + 537 + ], + "type": "inline_equation", + "content": "V^{\\neg s}" + }, + { + "bbox": [ + 130, + 501, + 482, + 537 + ], + "type": "text", + "content": " denotes a non-leaf node with its embedding " + }, + { + "bbox": [ + 130, + 501, + 482, + 537 + ], + "type": "inline_equation", + "content": "Z^{\\neg s}" + }, + { + "bbox": [ + 130, + 501, + 482, + 537 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 501, + 482, + 537 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(V^{\\neg s})" + }, + { + "bbox": [ + 130, + 501, + 482, + 537 + ], + "type": "text", + "content": " is the set of predecessors of " + }, + { + "bbox": [ + 130, + 501, + 482, + 537 + ], + "type": "inline_equation", + "content": "V^{\\neg s}" + }, + { + "bbox": [ + 130, + 501, + 482, + 537 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 501, + 482, + 537 + ], + "type": "inline_equation", + "content": "G(\\cdot)" + }, + { + "bbox": [ + 130, + 501, + 482, + 537 + ], + "type": "text", + "content": " is the MLP aggregating features from predecessors, and " + }, + { + "bbox": [ + 130, + 501, + 482, + 537 + ], + "type": "inline_equation", + "content": "S(\\cdot)" + }, + { + "bbox": [ + 130, + 501, + 482, + 537 + ], + "type": "text", + "content": " is another MLP decoding the final score " + }, + { + "bbox": [ + 130, + 501, + 482, + 537 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 130, + 501, + 482, + 537 + ], + "type": "text", + "content": " from the root node " + }, + { + "bbox": [ + 130, + 501, + 482, + 537 + ], + "type": "inline_equation", + "content": "V^{r}" + }, + { + "bbox": [ + 130, + 501, + 482, + 537 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 538, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 538, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 538, + 482, + 609 + ], + "type": "text", + "content": "It is important to note that each leaf embedding " + }, + { + "bbox": [ + 130, + 538, + 482, + 609 + ], + "type": "inline_equation", + "content": "Z^s" + }, + { + "bbox": [ + 130, + 538, + 482, + 609 + ], + "type": "text", + "content": " is stochastic, characterized by a Gaussian distribution " + }, + { + "bbox": [ + 130, + 538, + 482, + 609 + ], + "type": "inline_equation", + "content": "(p(Z^s | X, \\mathcal{G}) = \\mathcal{N}(\\mu^s, \\Sigma^s))" + }, + { + "bbox": [ + 130, + 538, + 482, + 609 + ], + "type": "text", + "content": ", with parameters predicted by " + }, + { + "bbox": [ + 130, + 538, + 482, + 609 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 538, + 482, + 609 + ], + "type": "text", + "content": ". The non-leaf embeddings are however deterministic given samples from leaf distributions. This design is motivated by our assumption of the scoring rubric, where uncertainty lies only in assessing action steps and identical individual scores will yield the same final score." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 624, + 411, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 411, + 636 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 411, + 636 + ], + "type": "text", + "content": "3.3 Learning with Variational Information Bottleneck" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "With stochastic embeddings, training of RICA" + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": " is a challenge. We design a training scheme under the variational information bottleneck framework." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "text", + "content": "A. Majeedi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "type": "text", + "content": "Variational information bottleneck (VIB). To train our model " + }, + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "type": "inline_equation", + "content": "p(Y|X, \\mathcal{G})" + }, + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "type": "text", + "content": " with stochastic step embeddings " + }, + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\{Z^s\\}" + }, + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "type": "text", + "content": ", we adopt the information bottleneck principle [36], leading to the maximization of the following objective" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 241, + 159, + 482, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 159, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 241, + 159, + 482, + 171 + ], + "type": "interline_equation", + "content": "I \\left(\\left\\{Z ^ {s} \\right\\}; Y | \\mathcal {G}\\right) - \\beta I \\left(\\left\\{Z ^ {s} \\right\\}; X | \\mathcal {G}\\right), \\tag {3}", + "image_path": "778da57ac4840d1318d6db2c44030bd72ac1ec78c37d78b79620e6cb2b3fa0ee.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 178, + 481, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 178, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 178, + 481, + 213 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 178, + 481, + 213 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 178, + 481, + 213 + ], + "type": "text", + "content": " is the conditional mutual information, and " + }, + { + "bbox": [ + 130, + 178, + 481, + 213 + ], + "type": "inline_equation", + "content": "\\beta > 0" + }, + { + "bbox": [ + 130, + 178, + 481, + 213 + ], + "type": "text", + "content": " controls the tradeoff between the sufficiency of using step embeddings " + }, + { + "bbox": [ + 130, + 178, + 481, + 213 + ], + "type": "inline_equation", + "content": "\\{Z^s\\}" + }, + { + "bbox": [ + 130, + 178, + 481, + 213 + ], + "type": "text", + "content": " for predicting " + }, + { + "bbox": [ + 130, + 178, + 481, + 213 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 178, + 481, + 213 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 130, + 178, + 481, + 213 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 178, + 481, + 213 + ], + "type": "text", + "content": ", and the size of the embeddings " + }, + { + "bbox": [ + 130, + 178, + 481, + 213 + ], + "type": "inline_equation", + "content": "\\{Z^s\\}" + }, + { + "bbox": [ + 130, + 178, + 481, + 213 + ], + "type": "text", + "content": " derived from " + }, + { + "bbox": [ + 130, + 178, + 481, + 213 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 178, + 481, + 213 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 178, + 481, + 213 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 178, + 481, + 213 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 214, + 482, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 482, + 262 + ], + "type": "text", + "content": "While mutual information is computationally intractable for high dimensional " + }, + { + "bbox": [ + 130, + 214, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\{Z^s\\}" + }, + { + "bbox": [ + 130, + 214, + 482, + 262 + ], + "type": "text", + "content": ", a common solution [1] is to assume Markov property (" + }, + { + "bbox": [ + 130, + 214, + 482, + 262 + ], + "type": "inline_equation", + "content": "p(Z|X,Y,\\mathcal{G}) = p(Z|X,\\mathcal{G})" + }, + { + "bbox": [ + 130, + 214, + 482, + 262 + ], + "type": "text", + "content": ") and conditional independence (" + }, + { + "bbox": [ + 130, + 214, + 482, + 262 + ], + "type": "inline_equation", + "content": "p(\\{Z^s\\}|X,\\mathcal{G}) = \\prod_s p(Z^s|X,\\mathcal{G})" + }, + { + "bbox": [ + 130, + 214, + 482, + 262 + ], + "type": "text", + "content": "), followed by the variational approximation for a tractable lower bound" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 268, + 482, + 282 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 268, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 138, + 268, + 482, + 282 + ], + "type": "interline_equation", + "content": "- \\mathcal {L} _ {\\mathrm {V I B}} = \\mathbb {E} _ {Z ^ {s} \\sim p (Z ^ {s} | X, \\mathcal {G}), \\forall s \\in \\mathbb {S}} \\left[ \\log p (Y | \\{Z ^ {s} \\}, \\mathcal {G}) \\right] - \\beta \\Sigma_ {s \\in \\mathbb {S}} \\operatorname {K L} \\left(p (Z ^ {s} | X, \\mathcal {G}) | | p (Z ^ {s} | \\mathcal {G})\\right), \\tag {4}", + "image_path": "27e7f23ad8a0d494a682a610021ffc7ee1ad370295c6ad6a1a3b9cd8930825bc.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 288, + 482, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 288, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 288, + 482, + 312 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 288, + 482, + 312 + ], + "type": "inline_equation", + "content": "p(Y|\\{Z^s\\}, \\mathcal{G})" + }, + { + "bbox": [ + 130, + 288, + 482, + 312 + ], + "type": "text", + "content": " is modeled by the scoring function " + }, + { + "bbox": [ + 130, + 288, + 482, + 312 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 288, + 482, + 312 + ], + "type": "text", + "content": ", KL denotes the Kullback-Leibler divergence, and " + }, + { + "bbox": [ + 130, + 288, + 482, + 312 + ], + "type": "inline_equation", + "content": "p(Z^s|\\mathcal{G})" + }, + { + "bbox": [ + 130, + 288, + 482, + 312 + ], + "type": "text", + "content": " is an approximate marginal prior." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 315, + 482, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 351 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 351 + ], + "type": "text", + "content": "VIB loss. The first term in Eq. (4) defines the log-likelihood of the score given the input. By assuming that output scores follow a Gaussian with a fixed variance " + }, + { + "bbox": [ + 130, + 315, + 482, + 351 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 315, + 482, + 351 + ], + "type": "text", + "content": ", this term can be reduced to a mean squared error (MSE) loss" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 247, + 357, + 482, + 377 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 357, + 482, + 377 + ], + "spans": [ + { + "bbox": [ + 247, + 357, + 482, + 377 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {M S E} = \\frac {1}{N} \\Sigma_ {i} ^ {N} \\left(\\hat {Y} _ {i} - Y _ {i}\\right) ^ {2} / \\sigma^ {2}, \\tag {5}", + "image_path": "0d9ad9b26ab287e516d246e4ebd4f8931b5a26f191d1c04f5f9b90c255a1a3be.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 384, + 481, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 481, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 481, + 409 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 384, + 481, + 409 + ], + "type": "inline_equation", + "content": "Y_{i}" + }, + { + "bbox": [ + 130, + 384, + 481, + 409 + ], + "type": "text", + "content": " is the predicted score for a video indexed by " + }, + { + "bbox": [ + 130, + 384, + 481, + 409 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 384, + 481, + 409 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 384, + 481, + 409 + ], + "type": "inline_equation", + "content": "\\hat{Y}_i" + }, + { + "bbox": [ + 130, + 384, + 481, + 409 + ], + "type": "text", + "content": " is the corresponding ground-truth score, and " + }, + { + "bbox": [ + 130, + 384, + 481, + 409 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 384, + 481, + 409 + ], + "type": "text", + "content": " is the total number of videos in the training set." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 410, + 481, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 481, + 434 + ], + "type": "text", + "content": "The second term in Eq. (4) regularizes the latent space and encodes prediction uncertainty. By assuming a marginal prior of " + }, + { + "bbox": [ + 130, + 410, + 481, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,I)" + }, + { + "bbox": [ + 130, + 410, + 481, + 434 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 410, + 481, + 434 + ], + "type": "inline_equation", + "content": "p(Z^s|\\mathcal{G})" + }, + { + "bbox": [ + 130, + 410, + 481, + 434 + ], + "type": "text", + "content": ", we have" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 203, + 441, + 481, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 441, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 203, + 441, + 481, + 475 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {K L} = \\Sigma_ {s \\in \\mathbb {S}} K L (\\mathcal {N} (\\mu^ {s} (x), \\Sigma^ {s} (x) | | \\mathcal {N} (0, I)) \\\\ = \\frac {1}{2} \\Sigma_ {s \\in \\mathbb {S}} \\Sigma_ {j = 1} ^ {D} \\left(\\left(\\mu_ {j} ^ {s}\\right) ^ {2} + \\left(\\sigma_ {j} ^ {s}\\right) ^ {2} - \\log \\left(\\sigma_ {j} ^ {s}\\right) ^ {2} - 1\\right), \\tag {6} \\\\ \\end{array}", + "image_path": "9de2b54a7eb4180e8af4ffbb76a0a64cfcfd310e4339f49e1cfc675181c635dc.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 480, + 481, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 480, + 481, + 505 + ], + "spans": [ + { + "bbox": [ + 130, + 480, + 481, + 505 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 480, + 481, + 505 + ], + "type": "inline_equation", + "content": "\\mu_j^s" + }, + { + "bbox": [ + 130, + 480, + 481, + 505 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 480, + 481, + 505 + ], + "type": "inline_equation", + "content": "\\sigma_j^s" + }, + { + "bbox": [ + 130, + 480, + 481, + 505 + ], + "type": "text", + "content": ", respectively, are the " + }, + { + "bbox": [ + 130, + 480, + 481, + 505 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 480, + 481, + 505 + ], + "type": "text", + "content": "-th dimension of the mean (" + }, + { + "bbox": [ + 130, + 480, + 481, + 505 + ], + "type": "inline_equation", + "content": "\\mu^s(x)" + }, + { + "bbox": [ + 130, + 480, + 481, + 505 + ], + "type": "text", + "content": ") and variance (square root of the diagonal of " + }, + { + "bbox": [ + 130, + 480, + 481, + 505 + ], + "type": "inline_equation", + "content": "\\Sigma^s(x)" + }, + { + "bbox": [ + 130, + 480, + 481, + 505 + ], + "type": "text", + "content": "), for the step " + }, + { + "bbox": [ + 130, + 480, + 481, + 505 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 480, + 481, + 505 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 146, + 505, + 316, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 505, + 316, + 517 + ], + "spans": [ + { + "bbox": [ + 146, + 505, + 316, + 517 + ], + "type": "text", + "content": "The VIB loss " + }, + { + "bbox": [ + 146, + 505, + 316, + 517 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{VIB})" + }, + { + "bbox": [ + 146, + 505, + 316, + 517 + ], + "type": "text", + "content": " is thus given by" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 258, + 525, + 481, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 525, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 258, + 525, + 481, + 536 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {V I B} = \\mathcal {L} _ {M S E} + \\beta \\mathcal {L} _ {K L}. \\tag {7}", + "image_path": "28e5d2ecaffa2d5f2b67af31de8eef8e9038f888aa1abcc52c54efaeeeb9343d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 543, + 482, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 543, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 543, + 482, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{VIB}" + }, + { + "bbox": [ + 130, + 543, + 482, + 590 + ], + "type": "text", + "content": " consists of (a) the MSE loss " + }, + { + "bbox": [ + 130, + 543, + 482, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MSE}" + }, + { + "bbox": [ + 130, + 543, + 482, + 590 + ], + "type": "text", + "content": " from the negative log-likelihood of the predicted scores, aiming at minimizing prediction errors; and (b) the KL divergence " + }, + { + "bbox": [ + 130, + 543, + 482, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{KL}" + }, + { + "bbox": [ + 130, + 543, + 482, + 590 + ], + "type": "text", + "content": " between the predicted Gaussian and the prior, regularizing the stochastic embeddings. Further, the coefficient " + }, + { + "bbox": [ + 130, + 543, + 482, + 590 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 543, + 482, + 590 + ], + "type": "text", + "content": " balances between two loss terms." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 591, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 591, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 591, + 482, + 639 + ], + "type": "text", + "content": "During training, samples are drawn to compute the loss function. The output is matched to the Gaussian distribution with its mean equal to the average of the judge scores. The reparameterization trick [14] is used to allow the backpropagation of gradients through the sampling process." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Estimating uncertainty. The diagonal covariance " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Sigma^s (X)" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " models the uncertainty of the predicted quality score of a step " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " for an input video " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": ". A larger" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 204, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 204, + 90, + 448, + 102 + ], + "type": "inline_equation", + "content": "\\mathrm{RICA}^2" + }, + { + "bbox": [ + 204, + 90, + 448, + 102 + ], + "type": "text", + "content": " : Rubric-Informed, Calibrated Assessment of Actions" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "value in its diagonal represents a wider distribution of scores and, hence, a lower confidence in the prediction. Following [16,22] we generate uncertainty scores by summing up the harmonic means of the predicted variances for individual steps" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 227, + 159, + 482, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 159, + 482, + 174 + ], + "spans": [ + { + "bbox": [ + 227, + 159, + 482, + 174 + ], + "type": "interline_equation", + "content": "\\operatorname {u n c e r t a i n t y} (Y) = \\Sigma_ {s \\in \\mathbb {S}} D / \\Sigma_ {j = 1} ^ {D} \\left(\\sigma_ {j} ^ {s}\\right) ^ {- 1}, \\tag {8}", + "image_path": "930e52e31786ea7926afcb22f01c286f363ba4b6b72f18460131597f171b07eb.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 180, + 481, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 180, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 180, + 481, + 205 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 180, + 481, + 205 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 180, + 481, + 205 + ], + "type": "text", + "content": " is the dimensionality of the stochastic embeddings. Again, " + }, + { + "bbox": [ + 130, + 180, + 481, + 205 + ], + "type": "inline_equation", + "content": "\\sigma_{j}^{s}" + }, + { + "bbox": [ + 130, + 180, + 481, + 205 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 130, + 180, + 481, + 205 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 180, + 481, + 205 + ], + "type": "text", + "content": "-th dimension of the predicted variance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 207, + 483, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 207, + 483, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 207, + 483, + 291 + ], + "type": "text", + "content": "Stochastic vs. deterministic modeling. An interesting variant of " + }, + { + "bbox": [ + 130, + 207, + 483, + 291 + ], + "type": "inline_equation", + "content": "\\mathrm{RICA}^2" + }, + { + "bbox": [ + 130, + 207, + 483, + 291 + ], + "type": "text", + "content": " is to disable its stochastic component. Conceptually, this is equal to considering step embeddings " + }, + { + "bbox": [ + 130, + 207, + 483, + 291 + ], + "type": "inline_equation", + "content": "Z^r" + }, + { + "bbox": [ + 130, + 207, + 483, + 291 + ], + "type": "text", + "content": " as vectors and removing the KL loss " + }, + { + "bbox": [ + 130, + 207, + 483, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{KL}" + }, + { + "bbox": [ + 130, + 207, + 483, + 291 + ], + "type": "text", + "content": ". We refer to this deterministic version of our model as " + }, + { + "bbox": [ + 130, + 207, + 483, + 291 + ], + "type": "inline_equation", + "content": "\\mathrm{RICA}^{2\\dagger}" + }, + { + "bbox": [ + 130, + 207, + 483, + 291 + ], + "type": "text", + "content": ". Without stochastic embeddings, " + }, + { + "bbox": [ + 130, + 207, + 483, + 291 + ], + "type": "inline_equation", + "content": "\\mathrm{RICA}^{2\\dagger}" + }, + { + "bbox": [ + 130, + 207, + 483, + 291 + ], + "type": "text", + "content": " is unable to estimate prediction uncertainty, yet often yields slightly lower prediction errors. This trade-off is also observed in prior works [6, 16, 32]. We include this variant of our model in the experiments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 309, + 367, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 309, + 367, + 322 + ], + "spans": [ + { + "bbox": [ + 131, + 309, + 367, + 322 + ], + "type": "text", + "content": "3.4 Model Instantiation and Implementation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 333, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 482, + 429 + ], + "type": "text", + "content": "Video and step representation. For an input video " + }, + { + "bbox": [ + 130, + 333, + 482, + 429 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 333, + 482, + 429 + ], + "type": "text", + "content": ", we adapt a pre-trained video backbone(e.g., I3D [4]) to extract its clip-level features, which are further pooled to produce video features " + }, + { + "bbox": [ + 130, + 333, + 482, + 429 + ], + "type": "inline_equation", + "content": "(x_{1}, x_{2}, \\ldots, x_{T})" + }, + { + "bbox": [ + 130, + 333, + 482, + 429 + ], + "type": "text", + "content": " with fixed-length " + }, + { + "bbox": [ + 130, + 333, + 482, + 429 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 333, + 482, + 429 + ], + "type": "text", + "content": ". To represent action steps, we make use of a pre-trained language model [7] (Flan-T5) to extract text features from their step descriptions, resulting in an ordered set of text embeddings " + }, + { + "bbox": [ + 130, + 333, + 482, + 429 + ], + "type": "inline_equation", + "content": "(s_{1}, s_{2}, \\ldots, s_{K})" + }, + { + "bbox": [ + 130, + 333, + 482, + 429 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 333, + 482, + 429 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 333, + 482, + 429 + ], + "type": "text", + "content": " steps. Note that the language model is not part of RICA². It is used solely to extract embeddings for text descriptions of the action steps (see supplement Tables I-L)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 431, + 483, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 431, + 483, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 431, + 483, + 528 + ], + "type": "text", + "content": "**Embedding function " + }, + { + "bbox": [ + 130, + 431, + 483, + 528 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 431, + 483, + 528 + ], + "type": "text", + "content": ".** Our embedding function " + }, + { + "bbox": [ + 130, + 431, + 483, + 528 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 431, + 483, + 528 + ], + "type": "text", + "content": " is realized using a Transformer model [37] (see Fig. 2(c)). " + }, + { + "bbox": [ + 130, + 431, + 483, + 528 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 431, + 483, + 528 + ], + "type": "text", + "content": " first processes video features " + }, + { + "bbox": [ + 130, + 431, + 483, + 528 + ], + "type": "inline_equation", + "content": "(x_{1}, x_{2}, \\ldots, x_{T})" + }, + { + "bbox": [ + 130, + 431, + 483, + 528 + ], + "type": "text", + "content": " with a self-attention block and text embeddings " + }, + { + "bbox": [ + 130, + 431, + 483, + 528 + ], + "type": "inline_equation", + "content": "(s_{1}, s_{2}, \\ldots, s_{K})" + }, + { + "bbox": [ + 130, + 431, + 483, + 528 + ], + "type": "text", + "content": " using a MLP. It further makes use of cross-attention blocks " + }, + { + "bbox": [ + 130, + 431, + 483, + 528 + ], + "type": "inline_equation", + "content": "(2\\mathbf{x})" + }, + { + "bbox": [ + 130, + 431, + 483, + 528 + ], + "type": "text", + "content": " to fuse video and text features, where video features are used to compute keys and values, and text embeddings of steps are projected into queries. Further, " + }, + { + "bbox": [ + 130, + 431, + 483, + 528 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 431, + 483, + 528 + ], + "type": "text", + "content": " decodes stochastic embeddings of individual steps by predicting a mean vector " + }, + { + "bbox": [ + 130, + 431, + 483, + 528 + ], + "type": "inline_equation", + "content": "\\mu^{s} \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 130, + 431, + 483, + 528 + ], + "type": "text", + "content": " and a diagonal covariance vector " + }, + { + "bbox": [ + 130, + 431, + 483, + 528 + ], + "type": "inline_equation", + "content": "\\Sigma^{s} \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 130, + 431, + 483, + 528 + ], + "type": "text", + "content": " for each step " + }, + { + "bbox": [ + 130, + 431, + 483, + 528 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 431, + 483, + 528 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 530, + 482, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 530, + 482, + 602 + ], + "spans": [ + { + "bbox": [ + 131, + 530, + 482, + 602 + ], + "type": "text", + "content": "Scoring function " + }, + { + "bbox": [ + 131, + 530, + 482, + 602 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 131, + 530, + 482, + 602 + ], + "type": "text", + "content": ". With Gaussian distributions for all steps specified by " + }, + { + "bbox": [ + 131, + 530, + 482, + 602 + ], + "type": "inline_equation", + "content": "\\{\\mu^s, \\Sigma^s\\}" + }, + { + "bbox": [ + 131, + 530, + 482, + 602 + ], + "type": "text", + "content": ", we encode the steps and score rubric into a video-specific DAG " + }, + { + "bbox": [ + 131, + 530, + 482, + 602 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 131, + 530, + 482, + 602 + ], + "type": "text", + "content": ", and realize " + }, + { + "bbox": [ + 131, + 530, + 482, + 602 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 131, + 530, + 482, + 602 + ], + "type": "text", + "content": " as a GNN defined on " + }, + { + "bbox": [ + 131, + 530, + 482, + 602 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 131, + 530, + 482, + 602 + ], + "type": "text", + "content": " following Eq. (2). " + }, + { + "bbox": [ + 131, + 530, + 482, + 602 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 131, + 530, + 482, + 602 + ], + "type": "text", + "content": " is parameterized by its aggregation function " + }, + { + "bbox": [ + 131, + 530, + 482, + 602 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 131, + 530, + 482, + 602 + ], + "type": "text", + "content": ", which is shared among nodes of the same type. " + }, + { + "bbox": [ + 131, + 530, + 482, + 602 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 131, + 530, + 482, + 602 + ], + "type": "text", + "content": " is implemented using an averaging operation followed by a MLP (2 layers). Finally, " + }, + { + "bbox": [ + 131, + 530, + 482, + 602 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 131, + 530, + 482, + 602 + ], + "type": "text", + "content": " decodes the final score at the root note " + }, + { + "bbox": [ + 131, + 530, + 482, + 602 + ], + "type": "inline_equation", + "content": "V^r" + }, + { + "bbox": [ + 131, + 530, + 482, + 602 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 131, + 530, + 482, + 602 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 131, + 530, + 482, + 602 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "Training with auxiliary losses. While the VIB loss (Eq. (7)) is sufficient for training, it falls short of considering the temporal ordering of steps. This is because of the conditional independence assumption needed for the derivation of VIB, i.e., " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "p(\\{Z^s\\} | X, \\mathcal{G}) = \\prod_s p(Z^s | X, \\mathcal{G})" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ", where the ordering of " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\{Z^s\\}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " is discarded. To bridge the gap, we incorporate an auxiliary loss term " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Aux}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " inspired by [2]." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "text", + "content": "A. Majeedi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": "Specifically, we re-purpose the last cross-attention map " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "(\\mathbb{R}^{K\\times T})" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": " as a step detector. This is done by computing a temporally-weighted center across the attention of each action step to every video time step (i.e., column-wise). We then enforce that (a) this center is co-located with the peak of the attention along video time steps using a sparsity loss [2]; and (b) all centers follow the temporal ordering of corresponding action steps using a ranking loss [2]. These two terms are summed up as the auxiliary loss, and further added to the VIB loss with a small weight (0.1). In our ablation, we empirically verify that adding the auxiliary loss leads to a minor performance boost." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 227, + 482, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 227, + 482, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 482, + 251 + ], + "type": "text", + "content": "Inference with sampling. At the inference time, we enhance robustness by sampling 20 times and averaging their predictions to compute the final score." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 270, + 303, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 270, + 303, + 284 + ], + "spans": [ + { + "bbox": [ + 131, + 270, + 303, + 284 + ], + "type": "text", + "content": "4 Experiments and Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 300, + 482, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 348 + ], + "type": "text", + "content": "Datasets. Our evaluations are primarily reported on three publicly available benchmark datasets, namely FineDiving [44], MTL-AQA [24], and JIGSAWS [10] in the main paper. In supplement Sec. B, we also include results on the Cataract-101 [31] with cataract surgery videos." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 351, + 482, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 482, + 374 + ], + "type": "text", + "content": "Evaluation metrics. For all our experiments, we consider metrics on both the accuracy of the prediction and the calibration of the uncertainty estimates." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 375, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 482, + 434 + ], + "type": "text", + "content": "- For accuracy, we use two widely adopted metrics for AQA [2,35,44], namely Spearman's rank correlation (SRCC) and relative L2 distance " + }, + { + "bbox": [ + 130, + 375, + 482, + 434 + ], + "type": "inline_equation", + "content": "(R\\ell_{2})" + }, + { + "bbox": [ + 130, + 375, + 482, + 434 + ], + "type": "text", + "content": ". SRCC measures how well the predicted scores are ranked w.r.t. the ground truth, while " + }, + { + "bbox": [ + 130, + 375, + 482, + 434 + ], + "type": "inline_equation", + "content": "R\\ell_{2}" + }, + { + "bbox": [ + 130, + 375, + 482, + 434 + ], + "type": "text", + "content": " summarizes the prediction errors. A model with more accurate predictions will have higher SRCC and lower " + }, + { + "bbox": [ + 130, + 375, + 482, + 434 + ], + "type": "inline_equation", + "content": "R\\ell_{2}" + }, + { + "bbox": [ + 130, + 375, + 482, + 434 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 434, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 518 + ], + "type": "text", + "content": "- For calibration, we report the uncertainty versus error curve following [16, 22]. To plot this curve, test samples are sorted by increasing uncertainty and divided into 10 equal-sized bins. The mean absolute error (MAE) is then computed for items in each bin. We also follow [16, 22] in employing Kendall's tau (" + }, + { + "bbox": [ + 130, + 434, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 434, + 482, + 518 + ], + "type": "text", + "content": ") [13], a numerical measure ranging from -1 to 1 to quantify the correlation between the uncertainties and the prediction errors. A higher " + }, + { + "bbox": [ + 130, + 434, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 434, + 482, + 518 + ], + "type": "text", + "content": " indicates better calibration, signifying that a model's uncertainty better aligns with prediction errors." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "type": "text", + "content": "Baselines. RICA" + }, + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "type": "text", + "content": " is benchmarked against a set of strong baselines, including exemplar-free methods such as DAE [48], UsDL and MUSDL [35], and exemplar-based ones such as CoRE [47], TPT [2] and TSA [44]. We further include the deterministic version of our model " + }, + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathrm{RICA}^{2\\dagger}" + }, + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "type": "text", + "content": ", which trades the ability of uncertainty estimation for a minor boost in accuracy. Several baselines adopt a direct regression approach, without providing a confidence or uncertainty measure for predictions. UsDL [35] and TPT [2] implicitly offer a confidence value. In these works, the probability of the predicted score bin serves as a proxy for uncertainty, computed as (1.0-confidence). DAE [48] outputs the standard deviation of the score distribution, which represents uncertainty." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "We seek to ensure a fair comparison yet recognize that methods in our benchmark may consider different settings and/or various types of input. Most" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 204, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 204, + 90, + 448, + 102 + ], + "type": "inline_equation", + "content": "\\mathrm{RICA}^2" + }, + { + "bbox": [ + 204, + 90, + 448, + 102 + ], + "type": "text", + "content": " : Rubric-Informed, Calibrated Assessment of Actions" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "prior exemplar-free methods only consider a video as input. While " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\mathrm{RICA}^2" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " does not utilize exemplars, it takes additional input of step information, i.e., step presence and their temporal ordering. On the other hand, previous exemplar-based methods also require the step information as used by " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\mathrm{RICA}^2" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": ", in addition to an input video and an exemplar database. Notably, step information is used to select exemplars, leading to improved results. For example, for diving videos, CoRE [47], TPT [2] and TSA [44] use the diving number (DN) encoding steps and their ordering. Further, TSA [44] also requires the timing of individual steps during training. While it is infeasible to standardize the settings of all methods, we compare to the best reported results in our experiments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 254, + 274, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 254, + 274, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 254, + 274, + 266 + ], + "type": "text", + "content": "4.1 Results on FineDiving" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 274, + 482, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 346 + ], + "type": "text", + "content": "Dataset. FineDiving [44] is the largest public dataset for AQA, with 3000 video samples capturing various diving actions. The dataset covers 52 different action types, 29 sub-action types, and 23 difficulty degree types, providing a rich and diverse set of examples for AQA. While this dataset contains temporal annotations for the steps, which can be used to improve the performance of AQA as demonstrated in [44], we do not use these annotations for " + }, + { + "bbox": [ + 130, + 274, + 482, + 346 + ], + "type": "inline_equation", + "content": "\\mathrm{RICA}^2" + }, + { + "bbox": [ + 130, + 274, + 482, + 346 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 349, + 482, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 349, + 482, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 349, + 482, + 421 + ], + "type": "text", + "content": "Experiment setup. We adhere to the experimental setup of the most recent baseline [44] using their train-test split, with 2251 videos for training and 749 videos for testing. We follow the input video settings used in [44] for " + }, + { + "bbox": [ + 130, + 349, + 482, + 421 + ], + "type": "inline_equation", + "content": "\\mathrm{RICA}^2" + }, + { + "bbox": [ + 130, + 349, + 482, + 421 + ], + "type": "text", + "content": " and the baselines. Specifically, for each video, we uniformly sample 96 frames, which are segmented into 9 overlapping clips, each containing 16 consecutive frames. We refer to supplement Sec. A.1 for further implementation details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "text", + "content": "Results. Tab. 1a presents our results on FineDiving. Both our stochastic and deterministic versions (RICA" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "text", + "content": " and RICA" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "inline_equation", + "content": "^2\\dagger" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "text", + "content": ") outperform the state-of-the-art TPT [2], an exemplar-based method. RICA" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "text", + "content": " shows a relative margin of " + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "text", + "content": " / " + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "text", + "content": " on SRCC / " + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "inline_equation", + "content": "R\\ell_2" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "text", + "content": ", and RICA" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "inline_equation", + "content": "^2\\dagger" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "text", + "content": " has a relative margin of " + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "text", + "content": " / " + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "inline_equation", + "content": "9.6\\%" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "text", + "content": " on SRCC / " + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "inline_equation", + "content": "R\\ell_2" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "text", + "content": ". This improvement is more pronounced when compared with the exemplar-free methods (MUSDL, DAE-MT) showcasing a significant relative gain of " + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "inline_equation", + "content": "4.9\\%" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "text", + "content": " on SRCC and " + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "inline_equation", + "content": "29.8\\%" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "inline_equation", + "content": "21.6\\%" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "inline_equation", + "content": "R\\ell_2" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "text", + "content": ". While the deterministic RICA" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "inline_equation", + "content": "^2\\dagger" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "text", + "content": " has slightly higher accuracy, our stochastic RICA" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "text", + "content": " demonstrates superior calibration of its uncertainty estimate (" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{RICA}^2} = 0.64" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{TPT}} = -0.56" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "text", + "content": "). Fig. 3a further shows uncertainty calibration results. Uncertainty estimates from RICA" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "text", + "content": " have a clear upward trend, indicating a higher calibration level. While MUSDL [35] also exhibits a reasonable level of calibration (" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{MUSDL}} = 0.47" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{RICA}^2} = 0.64" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "text", + "content": "), the errors are significantly higher than RICA" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 130, + 423, + 482, + 579 + ], + "type": "text", + "content": " across all uncertainty levels." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 597, + 274, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 274, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 274, + 609 + ], + "type": "text", + "content": "4.2 Results on MTL-AQA" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Dataset. MTL-AQA [24] is one of the most commonly used datasets for AQA. It consists of 1412 samples collected from 16 events with diverse views. The dataset has a rich set of annotations, including the steps performed during the dive, the difficulty score associated with the dive, and the individual judge scores." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "text", + "content": "A. Majeedi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 170, + 306, + 268 + ], + "blocks": [ + { + "bbox": [ + 130, + 125, + 482, + 158 + ], + "lines": [ + { + "bbox": [ + 130, + 125, + 482, + 158 + ], + "spans": [ + { + "bbox": [ + 130, + 125, + 482, + 158 + ], + "type": "text", + "content": "Table 1: Main results on (a) FineDiving and (b) MTL-AQA. Prediction accuracy " + }, + { + "bbox": [ + 130, + 125, + 482, + 158 + ], + "type": "inline_equation", + "content": "(SRCC" + }, + { + "bbox": [ + 130, + 125, + 482, + 158 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 125, + 482, + 158 + ], + "type": "inline_equation", + "content": "R\\ell_{2}" + }, + { + "bbox": [ + 130, + 125, + 482, + 158 + ], + "type": "text", + "content": ") and uncertainty calibration " + }, + { + "bbox": [ + 130, + 125, + 482, + 158 + ], + "type": "inline_equation", + "content": "(\\tau)" + }, + { + "bbox": [ + 130, + 125, + 482, + 158 + ], + "type": "text", + "content": " metrics are reported. We compare our method with exemplar-based and exemplar-free baselines." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 171, + 160, + 267, + 170 + ], + "lines": [ + { + "bbox": [ + 171, + 160, + 267, + 170 + ], + "spans": [ + { + "bbox": [ + 171, + 160, + 267, + 170 + ], + "type": "text", + "content": "(a) Results on FineDiving" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 170, + 306, + 268 + ], + "lines": [ + { + "bbox": [ + 135, + 170, + 306, + 268 + ], + "spans": [ + { + "bbox": [ + 135, + 170, + 306, + 268 + ], + "type": "table", + "html": "
Metrics
SRCC(↑)Rl2(↓)τ(↑)
Exemplar basedCoRe [47]0.90610.3615-
TSA [44]0.92030.3420-
TPT [2]0.93330.2877-0.5556
USDL [35,44]0.89130.38220.3778
MUSDL [35,44]0.89780.37040.4667
Exemplar freeDAE [48]0.88200.4919-0.1999
DAE-MT [48]0.92850.3320-0.4667
RICA2 (Ours)0.94020.28380.6444
RICA2† (Ours)0.94210.2600-
", + "image_path": "2c7581abd2c1c7f19dfdba5c23dbc5e318ce76eeb0d9b778a151e0585f42050b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 317, + 170, + 480, + 270 + ], + "blocks": [ + { + "bbox": [ + 351, + 160, + 445, + 170 + ], + "lines": [ + { + "bbox": [ + 351, + 160, + 445, + 170 + ], + "spans": [ + { + "bbox": [ + 351, + 160, + 445, + 170 + ], + "type": "text", + "content": "(b) Results on MTL-AQA" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 170, + 480, + 270 + ], + "lines": [ + { + "bbox": [ + 317, + 170, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 317, + 170, + 480, + 270 + ], + "type": "table", + "html": "
Metrics
SRCC(†)Rℓ2(↓)τ(↑)
Exemplar basedTSA-Net [40]0.9422--
CoRe [47]0.95120.2600-
DAE-CoRe [48]0.9589--
Exemplar freeTPT [2]0.96070.2378-0.1111
C3D-AVG-MTL [24]0.9044--
USDL [35]0.92310.46800.1556
MUSDL [35]0.92730.4510-0.0667
DAE [48]0.9231--
DAE-MT [48]0.94900.2738-0.4222
RICA2 (Ours)0.95940.25800.6000
RICA2† (Ours)0.96200.2280-
", + "image_path": "42be03937e2de5dc35421a5362962d505335311c3877c55812c14d5a6ed0ced7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 157, + 285, + 293, + 393 + ], + "blocks": [ + { + "bbox": [ + 157, + 285, + 293, + 393 + ], + "lines": [ + { + "bbox": [ + 157, + 285, + 293, + 393 + ], + "spans": [ + { + "bbox": [ + 157, + 285, + 293, + 393 + ], + "type": "image", + "image_path": "e61eb580869c267e53b222e1b379631aec9e94dfaf43a0046247600442b7a6b6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 150, + 395, + 300, + 405 + ], + "lines": [ + { + "bbox": [ + 150, + 395, + 300, + 405 + ], + "spans": [ + { + "bbox": [ + 150, + 395, + 300, + 405 + ], + "type": "text", + "content": "(a) Uncertainty calibration on FineDiving" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 321, + 285, + 455, + 393 + ], + "blocks": [ + { + "bbox": [ + 321, + 285, + 455, + 393 + ], + "lines": [ + { + "bbox": [ + 321, + 285, + 455, + 393 + ], + "spans": [ + { + "bbox": [ + 321, + 285, + 455, + 393 + ], + "type": "image", + "image_path": "ed416ef7a3939acf35939cfd44ad642e6d94b6e40b82dfd61ed88849b2fc2a6a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 395, + 462, + 405 + ], + "lines": [ + { + "bbox": [ + 313, + 395, + 462, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 462, + 405 + ], + "type": "text", + "content": "(b) Uncertainty calibration on MTL-AQA" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 414, + 482, + 460 + ], + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 460 + ], + "type": "text", + "content": "Fig. 3: Uncertainty vs. prediction error (MAE) on (a) Finediving and (b) MTL-AQA. Results are reported on the test splits, with the X-axis as the uncertainty bin index (uncertainty increases from left to right) and the Y-axis as the MAE in the bin. In comparison to baselines, " + }, + { + "bbox": [ + 130, + 414, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\mathrm{RICA}^2" + }, + { + "bbox": [ + 130, + 414, + 482, + 460 + ], + "type": "text", + "content": " has improved calibration with lower prediction errors." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 488, + 483, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 488, + 483, + 560 + ], + "spans": [ + { + "bbox": [ + 130, + 488, + 483, + 560 + ], + "type": "text", + "content": "Experiment setup. We follow the evaluation protocol of [2,43,47], dividing the dataset into the standard train set of 1059 videos and a test set of 353 videos. Further, we use the same input video settings as TPT [2] in our experiments to ensure a fair comparison. Specifically, for each video, we uniformly sample 103 frames segmented into 20 overlapping clips, each containing 8 continuous frames. Please refer to supplement Sec. A.2 for more details." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": "Results. Tab. 1b summarizes our results on MTL-AQA. Similar to FineDiving, " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathrm{RICA}^2" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " shows state-of-the-art results on MTL-AQA across all evaluation metrics. Specifically, our " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathrm{RICA}^2\\dagger" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " outperforms the best exemplar-free model (DAE-MT) by a relative margin of " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "1.4\\% / 16.7\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "SRCC / R\\ell_2" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": ". When compared with the competitive exemplar-based TPT, our has slightly better " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "SRCC" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "SRCC_{TPT} = 0.9607" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " vs " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "SRCC_{\\mathrm{RICA}^2\\dagger} = 0.9620" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "R\\ell_2" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "+4.1\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " relative margin). Again, compared to previous methods, our stochastic model shows improved calibration (" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{RICA}^2} = 0.60" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\tau_{USDL} = 0.16" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\tau_{TPT} = -0.11" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": ") as shown in Fig. 3b." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 204, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 204, + 90, + 448, + 102 + ], + "type": "inline_equation", + "content": "\\mathrm{RICA}^2" + }, + { + "bbox": [ + 204, + 90, + 448, + 102 + ], + "type": "text", + "content": " : Rubric-Informed, Calibrated Assessment of Actions" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 221, + 146, + 394, + 275 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 138 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 138 + ], + "type": "text", + "content": "Table 2: Results on JIGSAWS [10] dataset. Only prediction accuracy (" + }, + { + "bbox": [ + 130, + 114, + 482, + 138 + ], + "type": "inline_equation", + "content": "SRCC" + }, + { + "bbox": [ + 130, + 114, + 482, + 138 + ], + "type": "text", + "content": ") is considered due to the limited sample size. RICA" + }, + { + "bbox": [ + 130, + 114, + 482, + 138 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 130, + 114, + 482, + 138 + ], + "type": "text", + "content": " outperforms all prior approaches." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 221, + 146, + 394, + 275 + ], + "lines": [ + { + "bbox": [ + 221, + 146, + 394, + 275 + ], + "spans": [ + { + "bbox": [ + 221, + 146, + 394, + 275 + ], + "type": "table", + "html": "
Task
SNPKTAvg
Exemplar basedCoRe [47]0.840.860.860.85
TPT [2]0.880.880.910.89
ST-GCN [46]0.310.390.580.43
TSN [39]0.340.230.720.46
JRG [23]0.360.540.750.57
Exemplar freeUSDL [35]0.640.630.610.63
MUSDL [35]0.710.690.710.70
DAE [48]0.730.720.720.72
DAE-MT [48]0.780.740.740.76
RICA2† (Ours)0.880.930.880.90
RICA2 (Ours)0.920.940.900.92
", + "image_path": "b56cd2648cf2f389252b48a6928da16b656825ce9b9e448e26bf02a88aa71a7a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 295, + 269, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 295, + 269, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 295, + 269, + 306 + ], + "type": "text", + "content": "4.3 Results on JIGSAWS" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 129, + 312, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 312, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 129, + 312, + 482, + 373 + ], + "type": "text", + "content": "Dataset. In addition to diving videos, we also evaluate RICA² on JIGSAWS [10] — a robotic surgical video dataset. The dataset includes three tasks: \"Suturing (S),\" with 39 recordings, \"Needle Passing (NP),\" with 26 recordings and \"Knot Tying (KT)\" with 36 recordings. JIGSAWS is widely used for action quality assessment, despite its small scale." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 375, + 483, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 483, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 483, + 496 + ], + "type": "text", + "content": "Experiment setup. Due to the limited number of samples in the dataset (as few as 7 videos in the test set), cross-validation is often considered for evaluation on JIGSAWS. To ensure a fair comparison, we follow the commonly adopted splits from [35], and the input video setting from [2]. Specifically, for each video, we uniformly sample 160 frames which are segmented into 20 non-overlapping clips. We opt to not include score calibration curves due to the limited sample size of the test sets. Additionally, the key steps in JIGSAWS are general motions (e.g. reaching for the needle, orienting the needle, etc.) and thus cannot be localized to any specific section of the video. Thus, we do not use the auxiliary losses for this experiment. More details are described in supplement Sec. A.3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 497, + 482, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 584 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 584 + ], + "type": "text", + "content": "Results. Tab. 2 summarizes our results on JIGSAWS. Similar to previous datasets, our models exhibit notable advancements over the previous state-of-the-art model TPT [35], showcasing substantial improvements of " + }, + { + "bbox": [ + 130, + 497, + 482, + 584 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 130, + 497, + 482, + 584 + ], + "type": "text", + "content": " (RICA²) to " + }, + { + "bbox": [ + 130, + 497, + 482, + 584 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 130, + 497, + 482, + 584 + ], + "type": "text", + "content": " (RICA²†) in terms of average SRCC relative to the exemplar-based state-of-the-art TPT [2]. When compared to the exemplar-free methods, our approach demonstrates an impressive " + }, + { + "bbox": [ + 130, + 497, + 482, + 584 + ], + "type": "inline_equation", + "content": "18.4\\%" + }, + { + "bbox": [ + 130, + 497, + 482, + 584 + ], + "type": "text", + "content": " (RICA²) to " + }, + { + "bbox": [ + 130, + 497, + 482, + 584 + ], + "type": "inline_equation", + "content": "21.0\\%" + }, + { + "bbox": [ + 130, + 497, + 482, + 584 + ], + "type": "text", + "content": " (RICA²†) relative gain in average SRCC compared to the latest method DAE-MT [48]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 597, + 246, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 246, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 246, + 609 + ], + "type": "text", + "content": "4.4 Ablation Studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 615, + 481, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 615, + 481, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 615, + 481, + 639 + ], + "type": "text", + "content": "To understand our model design choices, we conduct ablation studies on the MTL-AQA [24] dataset. Additional ablations are in supplement Sec. C.2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Experiment setup. To simplify our experiments, we opt for running our ablations using fixed I3D features. This allows us to precisely evaluate the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "text", + "content": "A. Majeedi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 161, + 146, + 452, + 223 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 3: Ablation studies of model components on MTL-AQA dataset. * indicates that the text embeddings were frozen during training." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 161, + 146, + 452, + 223 + ], + "lines": [ + { + "bbox": [ + 161, + 146, + 452, + 223 + ], + "spans": [ + { + "bbox": [ + 161, + 146, + 452, + 223 + ], + "type": "table", + "html": "
Step Rep.DAG (Rubric)KLAuxMetrics
SRCC(↑)Rl2(↓)τ(↑)Avg. Rank (↓)
Random×××0.94260.3882-5.50
Text×××0.94310.3509-4.50
Text✓××0.94300.3336-4.50
Text*✓××0.94370.3335-3.50
Text*✓✓×0.94480.33290.42221.83
Text*✓✓✓0.94600.33030.42221.17
", + "image_path": "d9f1317b57058c0258b0eb17011964c5b15e7317a8c85a75ed7bfe41ed2501ca.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 244, + 479, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 244, + 479, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 244, + 479, + 281 + ], + "type": "text", + "content": "contribution of different components of our model. Specifically, we choose I3D weights from an intermediate checkpoint of our trained model and extract features for all videos with the frozen backbone." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 283, + 482, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 354 + ], + "type": "text", + "content": "Base model. Our ablation constructs a base model using randomly initialized step embeddings, an averaging of these embeddings after cross attention with the video features, followed by an MLP for scoring. This base model is trained using only the MSE loss. We then gradually add modules from " + }, + { + "bbox": [ + 130, + 283, + 482, + 354 + ], + "type": "inline_equation", + "content": "\\mathrm{RICA}^2" + }, + { + "bbox": [ + 130, + 283, + 482, + 354 + ], + "type": "text", + "content": " and study their effects. Tab. 3 presents our results using the same features and training epochs, with our base model in row 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 357, + 482, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 418 + ], + "type": "text", + "content": "Text embeddings as step representations. We first replace randomly initialized step representations with the text embeddings of step descriptions. This leads to a major boost in " + }, + { + "bbox": [ + 130, + 357, + 482, + 418 + ], + "type": "inline_equation", + "content": "R\\ell_{2}" + }, + { + "bbox": [ + 130, + 357, + 482, + 418 + ], + "type": "text", + "content": " (Tab. 3 row 1 vs. row 2), by leveraging knowledge encoded in the LLM [7]. Further, we find that freezing the text embeddings leads to comparable results and faster convergence (Tab. 3 row 3 vs. row 4)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 420, + 482, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 420, + 482, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 420, + 482, + 470 + ], + "type": "text", + "content": "Does the scoring rubric help? We also investigate the effects of encoding steps and rubric as a DAG—a key design of our model. Adding the DAG results in a noteworthy boost in " + }, + { + "bbox": [ + 130, + 420, + 482, + 470 + ], + "type": "inline_equation", + "content": "R\\ell_{2}" + }, + { + "bbox": [ + 130, + 420, + 482, + 470 + ], + "type": "text", + "content": " (Tab. 3 row 2 vs. row 3). This improvement can be ascribed to the DAG's proficiency in dissecting the action quality across steps." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 471, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 555 + ], + "type": "text", + "content": "Effects of loss functions. We now study the loss terms. Our loss function has three terms (a) the MSE loss " + }, + { + "bbox": [ + 130, + 471, + 482, + 555 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{MSE})" + }, + { + "bbox": [ + 130, + 471, + 482, + 555 + ], + "type": "text", + "content": " to minimize prediction error, (b) the KL loss " + }, + { + "bbox": [ + 130, + 471, + 482, + 555 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{KL})" + }, + { + "bbox": [ + 130, + 471, + 482, + 555 + ], + "type": "text", + "content": " to regularize the stochastic embeddings, and (c) the auxiliary loss " + }, + { + "bbox": [ + 130, + 471, + 482, + 555 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{Aux})" + }, + { + "bbox": [ + 130, + 471, + 482, + 555 + ], + "type": "text", + "content": " to ensure temporal ordering of steps. Adding the KL loss " + }, + { + "bbox": [ + 130, + 471, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{KL}" + }, + { + "bbox": [ + 130, + 471, + 482, + 555 + ], + "type": "text", + "content": " yields similar results in " + }, + { + "bbox": [ + 130, + 471, + 482, + 555 + ], + "type": "inline_equation", + "content": "SRCC" + }, + { + "bbox": [ + 130, + 471, + 482, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 471, + 482, + 555 + ], + "type": "inline_equation", + "content": "R\\ell_{2}" + }, + { + "bbox": [ + 130, + 471, + 482, + 555 + ], + "type": "text", + "content": ", yet enables calibrated uncertainty estimation. Further attaching the auxiliary loss " + }, + { + "bbox": [ + 130, + 471, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Aux}" + }, + { + "bbox": [ + 130, + 471, + 482, + 555 + ], + "type": "text", + "content": " leads to improvement in both " + }, + { + "bbox": [ + 130, + 471, + 482, + 555 + ], + "type": "inline_equation", + "content": "SRCC" + }, + { + "bbox": [ + 130, + 471, + 482, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 471, + 482, + 555 + ], + "type": "inline_equation", + "content": "R\\ell_{2}" + }, + { + "bbox": [ + 130, + 471, + 482, + 555 + ], + "type": "text", + "content": ", while maintaining the calibration performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 557, + 482, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 630 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 630 + ], + "type": "text", + "content": "Evaluating the cross-attention maps. To gain insight into " + }, + { + "bbox": [ + 130, + 557, + 482, + 630 + ], + "type": "inline_equation", + "content": "\\mathrm{RICA}^2" + }, + { + "bbox": [ + 130, + 557, + 482, + 630 + ], + "type": "text", + "content": ", we now examine the cross-attention maps between the step representations and video features in our learned embedding function " + }, + { + "bbox": [ + 130, + 557, + 482, + 630 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 557, + 482, + 630 + ], + "type": "text", + "content": ". Fig. 4 visualizes the attention map on two test videos on FineDiving. These maps reveal that a step representation is likely to attend to video features during which the step occurs, indicating that " + }, + { + "bbox": [ + 130, + 557, + 482, + 630 + ], + "type": "inline_equation", + "content": "\\mathrm{RICA}^2" + }, + { + "bbox": [ + 130, + 557, + 482, + 630 + ], + "type": "text", + "content": " learns to encode the temporal location of individual steps." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "We further evaluate this localization ability following the Pointing game protocol [49], widely considered in weakly supervised / unsupervised localization tasks [42,52]. Pointing Game compares a generated heatmap with an annotated" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 204, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 204, + 90, + 448, + 102 + ], + "type": "inline_equation", + "content": "\\mathrm{RICA}^2" + }, + { + "bbox": [ + 204, + 90, + 448, + 102 + ], + "type": "text", + "content": " : Rubric-Informed, Calibrated Assessment of Actions" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 479, + 327 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 479, + 327 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 479, + 327 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 479, + 327 + ], + "type": "image", + "image_path": "12bae1831dc4a12c4483acf22d88943f6e48bbff84366f60663f5a0aeb9a87d9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 336, + 483, + 381 + ], + "lines": [ + { + "bbox": [ + 130, + 336, + 483, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 483, + 381 + ], + "type": "text", + "content": "Fig. 4: Visualization of the cross-attention maps. Y-axis : attention value; Y-axis: clip indices (time). Each curve shows an attention map from a step representation to the temporal video features. The frames shown above are aligned with the timing of the corresponding attention plot. Curves and steps are colored accordingly." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 407, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 491 + ], + "type": "text", + "content": "time interval and counts the chance of the heatmap's peak falling into the specified interval. Our evaluation focuses on the FineDiving dataset since it is the only dataset providing annotated time intervals for individual steps. When evaluated on the full test set, attention maps from our model attain an accuracy of " + }, + { + "bbox": [ + 130, + 407, + 482, + 491 + ], + "type": "inline_equation", + "content": "61.4\\%" + }, + { + "bbox": [ + 130, + 407, + 482, + 491 + ], + "type": "text", + "content": " in the Pointing game protocol, significantly outperforming the chance level accuracy of " + }, + { + "bbox": [ + 130, + 407, + 482, + 491 + ], + "type": "inline_equation", + "content": "30.7\\%" + }, + { + "bbox": [ + 130, + 407, + 482, + 491 + ], + "type": "text", + "content": " (given each video has 3.26 steps on average). Note that we did not use any annotated segmentation data for training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 515, + 312, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 312, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 312, + 528 + ], + "type": "text", + "content": "5 Conclusion and Discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we present a deep probabilistic model for action quality assessment in videos. Our key innovation is to integrate score rubrics and to model prediction uncertainty. Specifically, we propose to adapt stochastic embeddings to quantify the uncertainty of individual steps, and to decode action scores using a variant of graph neural network operating on a DAG encoding the score rubric. Our method offers an exemplar-free approach for AQA, achieves new state-of-the-art results in terms of prediction accuracy on public benchmarks, and demonstrates superior calibration of the output uncertainty estimates. We believe that our work provides a solid step towards AQA. We hope that our method and findings can shed light on the challenging problem of trustworthy video recognition." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "text", + "content": "A. Majeedi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "Acknowledgement: This work was supported by the UW Madison Office of the Vice Chancellor for Research with funding from the Wisconsin Alumni Research Foundation, by National Science Foundation under Grant No. CNS 2333491, and by the Army Research Lab under contract number W911NF-2020221." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 182, + 197, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 197, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 197, + 194 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 205, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 205, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 480, + 228 + ], + "type": "text", + "content": "1. Alemi, A.A., Fischer, I., Dillon, J.V., Murphy, K.: Deep variational information bottleneck. In: International Conference on Learning Representations (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 228, + 481, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 481, + 272 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 481, + 272 + ], + "type": "text", + "content": "2. Bai, Y., Zhou, D., Zhang, S., Wang, J., Ding, E., Guan, Y., Long, Y., Wang, J.: Action quality assessment with temporal parsing transformer. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IV. pp. 422-438. Springer (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 272, + 481, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 272, + 481, + 316 + ], + "spans": [ + { + "bbox": [ + 138, + 272, + 481, + 316 + ], + "type": "text", + "content": "3. Battaglia, P.W., Hamrick, J.B., Bapst, V., Sanchez-Gonzalez, A., Zambaldi, V., Malinowski, M., Tacchetti, A., Raposo, D., Santoro, A., Faulkner, R., et al.: Relational inductive biases, deep learning, and graph networks. arXiv preprint arXiv:1806.01261 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 316, + 481, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 481, + 348 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 481, + 348 + ], + "type": "text", + "content": "4. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 349, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 349, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 349, + 481, + 380 + ], + "type": "text", + "content": "5. Chen, C.H., Hu, Y.H., Yen, T.Y., Radwin, R.G.: Automated video exposure assessment of repetitive hand activity level for a load transfer task. Human factors 55(2), 298-308 (2013)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 381, + 481, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 381, + 481, + 414 + ], + "spans": [ + { + "bbox": [ + 138, + 381, + 481, + 414 + ], + "type": "text", + "content": "6. Chun, S., Oh, S.J., De Rezende, R.S., Kalantidis, Y., Larlus, D.: Probabilistic embeddings for cross-modal retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8415-8424 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 415, + 481, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 415, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 415, + 481, + 491 + ], + "type": "text", + "content": "7. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., Webson, A., Gu, S.S., Dai, Z., Suzgun, M., Chen, X., Chowdhery, A., Castro-Ros, A., Pellat, M., Robinson, K., Valter, D., Narang, S., Mishra, G., Yu, A., Zhao, V., Huang, Y., Dai, A., Yu, H., Petrov, S., Chi, E.H., Dean, J., Devlin, J., Roberts, A., Zhou, D., Le, Q.V., Wei, J.: Scaling instructionfinetuned language models. Journal of Machine Learning Research 25(70), 1-53 (2024), http://jmlr.org/papers/v25/23-0870.html" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 491, + 481, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 481, + 523 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 481, + 523 + ], + "type": "text", + "content": "8. Duvenaud, D.K., Maclaurin, D., Iparraguirre, J., Bombarell, R., Hirzel, T., Aspuru-Guzik, A., Adams, R.P.: Convolutional networks on graphs for learning molecular fingerprints. Advances in neural information processing systems 28 (2015)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 523, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 481, + 555 + ], + "type": "text", + "content": "9. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 6202-6211 (2019). https://doi.org/10.1109/ICCV.2019.00630" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 555, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 555, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 134, + 555, + 481, + 610 + ], + "type": "text", + "content": "10. Gao, Y., Vedula, S.S., Reiley, C.E., Ahmidi, N., Varadarajan, B., Lin, H.C., Tao, L., Zappella, L., Bejar, B., Yuh, D.D., et al.: Jhu-isi gesture and skill assessment working set (jigsaws): A surgical activity dataset for human motion modeling. In: Modeling and Monitoring of Computer Assisted Interventions (M2CAI) - MICCAI Workshop (2014)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 610, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 610, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 610, + 481, + 632 + ], + "type": "text", + "content": "11. Gordon, A.S.: Automated video assessment of human performance. In: Proceedings of AI-ED. vol. 2 (1995)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "type": "text", + "content": "12. Guo, C., Pleiss, G., Sun, Y., Weinberger, K.Q.: On calibration of modern neural networks. In: International conference on machine learning. pp. 1321-1330. PMLR (2017)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 205, + 90, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 90, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 205, + 90, + 447, + 101 + ], + "type": "text", + "content": "RICA²: Rubric-Informed, Calibrated Assessment of Actions" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "13. Kendall, M.G.: A new measure of rank correlation. Biometrika 30(1/2), 81-93 (1938)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 160 + ], + "type": "text", + "content": "14. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. In: International Conference on Learning Representations (2014)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "type": "text", + "content": "15. Kipf, T.N., Welling, M.: Semi-supervised classification with graph convolutional networks. In: International Conference on Learning Representations (2017), https://openreview.net/forum?id=SJU4ayYgl" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "type": "text", + "content": "16. Li, W., Huang, X., Lu, J., Feng, J., Zhou, J.: Learning probabilistic ordinal embeddings for uncertainty-aware regression. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13896-13905 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 247 + ], + "type": "text", + "content": "17. Likert, R.: A technique for the measurement of attitudes. Archives of psychology (1932)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "text", + "content": "18. Liu, D., Li, Q., Jiang, T., Wang, Y., Miao, R., Shan, F., Li, Z.: Towards unified surgical skill assessment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9522-9531 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 282, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 314 + ], + "type": "text", + "content": "19. Martin, J., Regehr, G., Reznick, R., Macrae, H., Murnaghan, J., Hutchison, C., Brown, M.: Objective structured assessment of technical skill (osats) for surgical residents. British journal of surgery 84(2), 273-278 (1997)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "text", + "content": "20. Matsuyama, H., Kawaguchi, N., Lim, B.Y.: Iris: Interpretable rubric-informed segmentation for action quality assessment. In: Proceedings of the 28th International Conference on Intelligent User Interfaces. pp. 368-378 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 348, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 480, + 413 + ], + "type": "text", + "content": "21. Neelakantan, A., Shankar, J., Passos, A., McCallum, A.: Efficient non-parametric estimation of multiple embeddings per word in vector space. In: Moschitti, A., Pang, B., Daelemans, W. (eds.) Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing, EMNLP 2014, October 25-29, 2014, Doha, Qatar, A meeting of SIGDAT, a Special Interest Group of the ACL. pp. 1059-1069. ACL (2014)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "type": "text", + "content": "22. Oh, S.J., Gallagher, A.C., Murphy, K.P., Schroff, F., Pan, J., Roth, J.: Modeling uncertainty with hedged instance embeddings. In: International Conference on Learning Representations (2019), https://openreview.net/forum?id=r1xQQhAqKX" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "type": "text", + "content": "23. Pan, J.H., Gao, J., Zheng, W.S.: Action assessment by joint relation graphs. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6331-6340 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "text", + "content": "24. Parmar, P., Morris, B.T.: What and how well you performed? a multitask learning approach to action quality assessment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 304-313 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "25. Parmar, P., Tran Morris, B.: Learning to score olympic events. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops. pp 20-28 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 545, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 578 + ], + "type": "text", + "content": "26. Pirsiavash, H., Vondrick, C., Torralba, A.: Assessing the quality of actions. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13. pp. 556-571. Springer (2014)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "text", + "content": "27. Prassas, S., Kwon, Y.H., Sands, W.A.: Biomechanical research in artistic gymnastics: a review. Sports Biomechanics 5(2), 261-291 (2006)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "28. Qiu, Y., Wang, J., Jin, Z., Chen, H., Zhang, M., Guo, L.: Pose-guided matching based on deep learning for assessing quality of action on rehabilitation training Biomedical Signal Processing and Control 72, 103323 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "29. Santoro, A., Raposo, D., Barrett, D.G., Malinowski, M., Pascanu, R., Battaglia, P., Lillicrap, T.: A simple neural network module for relational reasoning. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 236, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 236, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 236, + 102 + ], + "type": "text", + "content": "A. Majeedi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "30. Scarselli, F., Gori, M., Tsoi, A.C., Hagenbuchner, M., Monfardini, G.: The graph neural network model. IEEE transactions on neural networks 20(1), 61-80 (2008)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 193 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 193 + ], + "type": "text", + "content": "31. Schoeffmann, K., Taschwer, M., Sarny, S., Munzer, B., Primus, M.J., Putzgruber, D.: Cataract-101: video dataset of 101 cataract surgeries. In: Cesar, P., Zink, M., Murray, N. (eds.) Proceedings of the 9th ACM Multimedia Systems Conference, MMSys 2018, Amsterdam, The Netherlands, June 12-15, 2018. pp. 421-425. ACM (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "text", + "content": "32. Shi, Y., Jain, A.K.: Probabilistic face embeddings. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6902-6911 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 247 + ], + "type": "text", + "content": "33. Sohn, K., Lee, H., Yan, X.: Learning structured output representation using deep conditional generative models. Advances in neural information processing systems 28 (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 292 + ], + "type": "text", + "content": "34. Sun, J.J., Zhao, J., Chen, L.C., Schroff, F., Adam, H., Liu, T.: View-invariant probabilistic embedding for human pose. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16. pp. 53-70. Springer (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "type": "text", + "content": "35. Tang, Y., Ni, Z., Zhou, J., Zhang, D., Lu, J., Wu, Y., Zhou, J.: Uncertainty-aware score distribution learning for action quality assessment. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9839-9848 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "type": "text", + "content": "36. TISHBY, N.: The information bottleneck method. In: Proc. of the 37th Allerton Conference on Communication and Computation, 1999 (1999)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "text", + "content": "37. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "text", + "content": "38. Vilnis, L., McCallum, A.: Word representations via gaussian embedding. In: International Conference on Learning Representations (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "text", + "content": "39. Wang, L., Xiong, Y., Wang, Z., Qiao, Y., Lin, D., Tang, X., Van Gool, L.: Temporal segment networks: Towards good practices for deep action recognition. In: European conference on computer vision. pp. 20-36. Springer (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 447, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 479 + ], + "type": "text", + "content": "40. Wang, S., Yang, D., Zhai, P., Chen, C., Zhang, L.: Tsa-net: Tube self-attention network for action quality assessment. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 4902-4910 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "type": "text", + "content": "41. Waters, T.R., Putz-Anderson, V., Garg, A.: Applications manual for the revised niosh lifting equation (1994)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 502, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 481, + 533 + ], + "type": "text", + "content": "42. Xiao, F., Sigal, L., Jae Lee, Y.: Weakly-Supervised Visual Grounding of Phrases With Linguistic Structures. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5945-5954 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "text", + "content": "43. Xu, C., Fu, Y., Zhang, B., Chen, Z., Jiang, Y.G., Xue, X.: Learning to score figure skating sport videos. IEEE transactions on circuits and systems for video technology 30(12), 4578-4590 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "text", + "content": "44. Xu, J., Rao, Y., Yu, X., Chen, G., Zhou, J., Lu, J.: Finediving: A fine-grained dataset for procedure-aware action quality assessment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2949-2958 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "text", + "content": "45. Xu, K., Li, J., Zhang, M., Du, S.S., ichi Kawarabayashi, K., Jegelka, S.: What can neural networks reason about? In: International Conference on Learning Representations (2020), https://openreview.net/forum?id=rJxbJeHFPS" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "text", + "content": "46. Yan, S., Xiong, Y., Lin, D.: Spatial temporal graph convolutional networks for skeleton-based action recognition. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 205, + 90, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 90, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 205, + 90, + 447, + 101 + ], + "type": "text", + "content": "RICA²: Rubric-Informed, Calibrated Assessment of Actions" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 315 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "47. Yu, X., Rao, Y., Zhao, W., Lu, J., Zhou, J.: Group-aware contrastive regression for action quality assessment. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7899-7908. IEEE Computer Society, Los Alamitos, CA, USA (oct 2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "text", + "content": "48. Zhang, B., Chen, J., Xu, Y., Zhang, H., Yang, X., Geng, X.: Auto-encoding score distribution regression for action quality assessment. Neural Computing and Applications pp. 1-14 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "49. Zhang, J., Bargal, S.A., Lin, Z., Brandt, J., Shen, X., Sclaroff, S.: Top-Down Neural Attention by Excitation Backprop. International Journal of Computer Vision 126(10), 1084-1102 (Oct 2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "text", + "content": "50. Zhou, C., Huang, Y.: Uncertainty-driven action quality assessment. arXiv preprint arXiv:2207.14513 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "51. Zhou, K., Ma, Y., Shum, H.P.H., Liang, X.: Hierarchical graph convolutional networks for action quality assessment. IEEE Transactions on Circuits and Systems for Video Technology pp. 1-1 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "type": "text", + "content": "52. Zhu, Y., Zhou, Y., Ye, Q., Qiu, Q., Jiao, J.: Soft Proposal Networks for Weakly Supervised Object Localization. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1841-1850 (2017)" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 236, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 236, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 236, + 102 + ], + "type": "text", + "content": "A. Majeedi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/RING-NeRF _ Rethinking Inductive Biases for Versatile and Efficient Neural Fields/5e093f34-06af-49a4-a5fe-2e79d5819847_content_list.json b/2024/RING-NeRF _ Rethinking Inductive Biases for Versatile and Efficient Neural Fields/5e093f34-06af-49a4-a5fe-2e79d5819847_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7f7f7fe339b9a4cd4b8addee3e3ed50dd1d28b70 --- /dev/null +++ b/2024/RING-NeRF _ Rethinking Inductive Biases for Versatile and Efficient Neural Fields/5e093f34-06af-49a4-a5fe-2e79d5819847_content_list.json @@ -0,0 +1,1748 @@ +[ + { + "type": "text", + "text": "RING-NeRF : Rethinking Inductive Biases for Versatile and Efficient Neural Fields", + "text_level": 1, + "bbox": [ + 235, + 140, + 767, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Doriand Petit $^{1,2}$ , Steve Bourgeois $^{1}$ , Dumitru Pavel $^{1}$ , Vincent Gay-Bellile $^{1}$ , Florian Chabot $^{1}$ , and Loic Barthe $^{2}$", + "bbox": [ + 233, + 210, + 767, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Université Paris-Saclay, CEA, List, F-91120, Palaiseau, France \n2 IRIT, Université Toulouse III, CNRS, France", + "bbox": [ + 284, + 253, + 717, + 282 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/72ab70ec195e7c177d49d3e764af15c3e127b41b5cb7f1c9512dfe05a18f63fb.jpg", + "image_caption": [ + "Fig. 1: RING-NeRF is a simple and versatile architecture which tackles many NeRF common issues such as robustness to distance of observation, few view supervision and lack of scene-specific initialization for SDF-based reconstruction. It provides on-par performances in terms of quality with SotA dedicated solutions [4,11,25] and and in terms of efficiency with fast methods [20,26]." + ], + "image_footnote": [], + "bbox": [ + 240, + 305, + 754, + 513 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Recent advances in Neural Fields mostly rely on developing task-specific supervision which often complicates the models. Rather than developing hard-to-combine and specific modules, another approach generally overlooked is to directly inject generic priors on the scene representation (also called inductive biases) into the NeRF architecture. Based on this idea, we propose the RING-NeRF architecture which includes two inductive biases: a continuous multi-scale representation of the scene and an invariance of the decoder's latent space over spatial and scale domains. We also design a single reconstruction process that takes advantage of those inductive biases and experimentally demonstrates on-par performances in terms of quality with dedicated architecture on multiple tasks (anti-aliasing, few view reconstruction, SDF reconstruction without scene-specific initialization) while being more efficient. Moreover, RING-NeRF has the distinctive ability to dynamically increase the resolution of the model, opening the way to adaptive reconstruction. Project page can be found at: https://cea-list.github.io/RING-NeRF", + "bbox": [ + 259, + 618, + 740, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 188, + 375, + 204 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Radiance Fields (NeRF) have emerged as a novel method for representing 3D scenes using neural networks. In its original design [13], a simple multi-layer perceptron (MLP) is trained to reproduce a continuous 5D function that outputs the density and radiance emitted in each direction $(\\theta, \\phi)$ at each point $(x, y, z)$ in space. This approach inspired many works due to the impressive quality of its novel view synthesis as well as its simplicity. However, the so-called \"vanilla\" NeRF architectures converge very slowly as they use large deep neural networks. Instant-NGP [14] introduced a new architecture based on a hierarchical hash-grid pyramid of learnable feature codes describing the 3D scene and combined with a much shallower MLP, called decoder, that transforms the concatenation of the codes, interpolated from the grids, into density and radiance. Resulting in local updates, this method reduced the training duration from hours to minutes.", + "bbox": [ + 212, + 220, + 787, + 400 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The great majority of current solutions are now based on these two standard architectures, and many of them are focused on overtaking their associated limitations in terms of:", + "bbox": [ + 212, + 402, + 785, + 445 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nature of scene - by transitioning from object-centric scenes to open unbounded scenes, using mostly space contraction [3];", + "bbox": [ + 212, + 446, + 785, + 477 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Robustness - by managing free motion trajectories with variations of the observation distance while avoiding aliasing artefacts, mostly through the integration of Level of Detail (LOD) in the model [2,4]; or by reducing drastically the number of supervised views through different kinds of regularization [8,15,25];", + "bbox": [ + 212, + 477, + 785, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensibility - by shifting from a holistic and fixed reconstruction process to an incremental (extensibility in the number of views) and adaptive (extensibility in resolution) process, mainly through the use of a frozen decoder [12, 29].", + "bbox": [ + 212, + 537, + 785, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, most solutions solely focus on one of these limitations and introduce specific and complex mechanisms that both increase the computational cost and lessen the possibilities of combination. The ability to solve jointly these main issues is however essential in real-world applications which often require a robust and extensible reconstruction in an unknown and unbounded environment.", + "bbox": [ + 212, + 583, + 785, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this article, rather than introducing yet another heavy and task-specific solution, we propose RING-NeRF, a versatile and simple NeRF architecture by rethinking usual grid-based models to introduce two inductive biases. We first represent a 3D scene as a continuous multi-scale representation and also make the decoder's latent space invariant in position and scale. Together, these two priors enable the production of intrinsic continuous LOD of the scene without explicit supervision. We demonstrate experimentally that, when combined with adapted cone casting and coarse-to-fine optimization, the resulting architecture is able to compete on several tasks with on-par quality performances with dedicated state-of-the-art solutions while improving speed, robustness and extensibility. The overall process is also simple, easy to implement and generic enough to be coupled with specific solutions. Our contributions can be summarized as follows:", + "bbox": [ + 212, + 657, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "D. Petit et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Keywords: NeRF $\\cdot$ 3D Reconstruction $\\cdot$ Implicit Neural Representation", + "bbox": [ + 261, + 147, + 738, + 161 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. an architecture that, by construction, represents the scene with a continuous and unbounded level of detail without the need for LOD-specific supervision and which permits resolution-adaptive reconstruction;", + "bbox": [ + 222, + 146, + 785, + 190 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. a distance-aware forward mapping compatible with scene contraction, that takes benefits of the continuous multi-scale representation with an adapted cone casting process to avoid aliasing artefacts when varying the observation distance.", + "bbox": [ + 222, + 191, + 785, + 251 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. a continuous coarse-to-fine reconstruction process that improves the convergence and stability (especially in challenging setups such as supervision with few viewpoints or no scene-specific initialization for SDF reconstruction).", + "bbox": [ + 220, + 253, + 785, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 217, + 323, + 387, + 339 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "From its original iteration [13], a majority of current research focuses on overtaking limitations of NeRF-based reconstruction in terms of adaptability to various natures of scene, robustness (to varying observation distances or limited amount of viewpoints) and reconstruction extensibility.", + "bbox": [ + 212, + 356, + 784, + 416 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adaptability to various natures of the scene. The ability of Neural Fields to reconstruct various natures of scene depends on three factors. The first one is related to its architecture itself. Tri-plane architectures [5, 7, 30] are mostly designed for object-centric reconstruction (as they provide a higher density of information in the center of the scene) while vanilla and 3D grid-based NeRF are able to cover a wider variety of scenes, though they initially were still constrained to a limited volume. A second factor is related to the representation of the 3D space of the scene, especially to represent distant elements in open scenes. Some approaches use two different NeRF models to reconstruct separately the foreground and the background [28], whereas others apply space contraction to the 3D scene coordinates [3, 4, 28] to map the infinite scene volume into a bounded one. The last factor is related to the initialization of the Neural Fields. While the random initialization of density-based NeRF can adjust to almost any nature of scenes, the convergence of SDF-based (Signed Distance Function) Neural Fields is extremely sensitive to their parameters' initialization, as stated in [1]. Current solutions rely on a scene-specific initialization (using an SDF field representing a sphere for outdoor scenes or an inverted sphere for indoor scenes), making them unable to adapt automatically to any scene.", + "bbox": [ + 212, + 417, + 787, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Robustness to observation distance variations. The initial NeRF model, as well as most of the subsequent works, relies implicitly on the hypothesis of a constant distance of observation to the scene. Indeed, the NeRF model provides a per-3D-point scene density representation and a rendering process which does not take into account the increasing volume covered by a pixel with respect to the distance to the cameras. As underlined in Mip-NeRF [2], this discrepancy induces artefacts such as over-contrasted images or aliasing phenomenon when the distance of observation differs from the ones used at reconstruction time. To avoid these artefacts, the rendering process needs to assess the density and color for a volume instead of a point. In the current state-of-the-art, two main", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "RING-NeRF", + "bbox": [ + 643, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "approaches are used (or their combination, as in Zip-NeRF [4]). The first solution consists of representing the scene with different levels of detail (LOD), in order to vary the precision of the reconstruction based on the observation distance. This is usually done by using a LOD-aware latent space [2,4,6,19,21,24], meaning that the LOD information is already encoded in the inputs of the MLP. This can be achieved by using a per-LOD decoder as in PyNeRF [21] or by incorporating LOD information in the latent feature as done in Zip-NeRF [4] and VR-NeRF [24]. One main flaw of these solutions is that they require the supervision of every used LOD which makes it impossible to vary the observation distances between the train views and novel synthesised views. A second approach consists of defining the latent representation of a volume as the mean of the latent features of the points included in the volume. It requires integrating the features over the volume, which can be achieved through convolutions for tri-plan representation [7,30], or through super-sampling of the latent space for 3D grids as also done in Zip-NeRF [4]. However, these approaches lengthen the training and rendering processes as they increase the number of computations.", + "bbox": [ + 212, + 145, + 787, + 387 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Robustness to limited amount of viewpoints. By construction, NeRF is subject to the shape-radiance ambiguity [28]. If not enough supervision viewpoints are available, the optimization might overfit them while not providing consistent 3D reconstruction nor generalization to non-supervised viewpoints. A first family of solutions to overcome this limitation consists of regularizing the reconstruction process through additional losses, whether via geometric [15,23,27] or semantic [8] regularization. The second family of solutions relies on a progressive reconstruction of the details of the scene, as introduced in FreeNeRF [25] and Nerfies [17]. Restraining the ability of the model to reconstruct a complex scene at early stages enforces the consistency of the reconstruction over the different supervision viewpoints. This presents the advantage of bringing stability while keeping a fairly simple training process. However, these latter solutions mostly rely on Vanilla models for stability purposes and require long training duration.", + "bbox": [ + 212, + 388, + 787, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reconstruction extensibility. Two kinds of extensibility should be distinguished: extensibility with respect to the number of views or the resolution. The first one is related to reconstructing new scene areas while keeping the previously reconstructed ones unchanged. The usual solution consists of using grid-based approaches with a pre-trained position-invariant decoder that is frozen during the reconstruction [29]. On the other side, the extensibility of the resolution consists in dynamically increasing the level of detail of a previously reconstructed scene. This problem is intractable for classic architectures such as vanilla NeRF and I-NGP since the number of layers (resp. grids) of those approaches cannot increase during the reconstruction. Some rare solutions, such as Neural Sparse Voxel Fields [12], combine a pre-trained decoder with a data structure allowing to dynamically allocate voxels to increase the reconstruction resolution. This is however a crucial stake, as being able to adapt the precision of the representation based on the complexity of the scene permits to optimize the computational efficiency (both in speed and memory requirements) with minimal loss in quality.", + "bbox": [ + 212, + 599, + 787, + 828 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "D. Petit et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d84823d8ac274faac4824caf00c8975b98a84de9c4e9946b64c03e02d8752d6a.jpg", + "image_caption": [ + "Fig. 2: Overview of RING-NeRF: to render a pixel, the casted cone is sampled with cubes. Depending on the cube volume, the corresponding LOD of the scene is selected and the latent feature is computed using a weighted sum of the grid hierarchy. The density (or SDF) and color of the cube are first decoded from the latent feature with a tiny MLP and then integrated with other samples through volume rendering." + ], + "image_footnote": [], + "bbox": [ + 230, + 143, + 777, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 RING-NeRF", + "text_level": 1, + "bbox": [ + 215, + 383, + 375, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Rather than focusing on solving one specific problem of NeRF using complex mechanisms, we propose a simple architecture called RING-NeRF constructed with novel inductive biases to tackle NeRF's common issues.", + "bbox": [ + 214, + 424, + 782, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Overview", + "text_level": 1, + "bbox": [ + 215, + 500, + 339, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "RING-NeRF relies on the classic NeRF [13] inverse rendering pipeline which is used to reconstruct a 3D scene from a set of localized frames. For a given image pixel (with its camera's pose), a 3D ray is cast and the 3D scene representation is sampled at $N$ various locations along the ray. The resulting density $\\sigma_{i}$ (or SDF converted to density [22]) and color $c_{i}$ of the samples are then combined with usual volume rendering techniques: $\\hat{C}(r) = \\sum_{i=0}^{N-1} T_{i} (1 - \\exp(-\\sigma_{i} \\delta_{i})) c_{i}$ with $T_{i} = \\exp(\\sum_{j=0}^{i-1} \\sigma_{j} \\delta_{j})$ and $\\delta_{i}$ is the distance between samples. The parameters of the scene representation are then optimized by minimizing the MSE loss $||\\hat{C}(r) - C(r)||^{2}$ between the rendered ray color and the ground truth pixel value, the rendering being differentiable3.", + "bbox": [ + 214, + 532, + 784, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The originality of RING-NeRF relies on its neural architecture, illustrated in figure 2, which is designed to represent the scene with a continuous and unbounded level of detail without the need for LOD-specific supervision (see section 3.2). We then use this LOD inductive bias to adjust the LOD of the samples with respect to the distance to the camera in contracted space for more accurate renderings (see section 3.3). Finally, combined with a continuous coarse-to-fine optimization, RING-NeRF results in a more robust reconstruction process with an intrinsic LOD extensibility property (see section 3.4).", + "bbox": [ + 214, + 688, + 784, + 809 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "RING-NeRF", + "bbox": [ + 643, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3 With SDF, an additional Eikonal loss is also being optimized for SDF regularization.", + "bbox": [ + 217, + 824, + 782, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 NeRF Representation with Inductive Biases", + "text_level": 1, + "bbox": [ + 215, + 146, + 627, + 161 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Classic grid-based representations, such as the one introduced by I-NGP [14], rely on two key elements: a unique MLP decoder which transforms a latent space into an output space (color, density or SDF), and a 3D (hierarchical) grid of latent features that implicitly defines a unique mapping function from the scene space onto the decoder latent space. We can distinguish two different approaches used to extend the mono-scale representation to a multi-scale representation: conditioning with scale information the decoding process [4,21] and/or defining separate per-LOD mapping functions [21]. In both cases, the training becomes more complex, leading to potential convergence issues as it relies solely on additional specific supervision. Instead, we propose to condition the mapping function itself with scale information and introduce inductive biases to guide the convergence.", + "bbox": [ + 212, + 170, + 787, + 351 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Scene representation as multi-scale mapping function. As illustrated in Figure 2, the mapping function at level $N$ is controlled by a grid $G_{N}$ that is implicitly defined through a recursive refinement process over scale:", + "bbox": [ + 212, + 352, + 787, + 398 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} G _ {0} \\leftarrow \\delta_ {0} \\\\ G _ {N} \\leftarrow S \\left(G _ {N - 1}\\right) + \\delta_ {N} \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 410, + 784, + 449 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\delta_{i}$ is the deviation defined by the i-th latent feature grid and S is the subdivision scheme consisting in increasing the resolution of $G_{N - 1}$ up to the $N$ -th grid resolution (with tri-linear interpolation). In practice, the latent vector associated with a point in the scene for a level of detail $N$ is simply obtained by interpolating linearly the point in the cell for each level inferior or equal to $N$ , and summing the results. Since linear interpolation is differentiable, our mapping function is also estimizable. The summing of the interpolation is comparable to a residual connection and gives the name of the architecture RING-NeRF for Residual Implicit Neural Grids.", + "bbox": [ + 212, + 453, + 787, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Such representation provides several advantages. First, since the mapping function is constructed in a top-down manner, refining recursively one level of detail from the previous one, its number of LOD is unbounded and adding a finer level of detail keeps the coarser ones valid. Secondly, because each $\\delta_{i}$ only represents a deviation of the coarser mapping value $G_{i - 1}$ , a continuous LOD representation can be easily obtained by incorporating a weighting factor $\\alpha \\in [0,1]$ in the recursive process:", + "bbox": [ + 212, + 589, + 787, + 696 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nG _ {L} \\leftarrow S \\left(G _ {N - 1}\\right) + \\alpha \\delta_ {N} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 419, + 708, + 784, + 723 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with $L\\in [N - 1,N]$ and $\\alpha = L - (N - 1)$", + "bbox": [ + 212, + 733, + 524, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Spatial and scale invariance of decoder. Since our architecture does not rely on a decoder conditioned on position (unlike [19]) nor scale (unlike [2-4, 21]), the decoder latent space is invariant to translations and scale changes in the scene coordinate. This property makes RING-NeRF more suited for incremental reconstruction in both spatial and scale space, since it ensures that local updates in the spatial and scale domain of the scene can be achieved through the", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "D. Petit et al.", + "bbox": [ + 271, + 114, + 364, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2db464ad350a9ef5df99d47be7363b59bd182042defe315f3196b4660b803091.jpg", + "image_caption": [ + "Fig. 3: We demonstrate the LOD inductive bias by training our model with a hierarchy of $N = 7$ levels where only the last level of the mapping function $G_{N}$ is supervised. We then compute renders at different levels $L \\leq N$ . Other examples (including of entire scenes) can be found in the supplementary materials." + ], + "image_footnote": [], + "bbox": [ + 220, + 147, + 330, + 212 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a526fb7ed0a150bac77bbb989deef339b0eb5fc05ecd7338a9a0489ffba4b002.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 148, + 442, + 213 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c96e543197cd47c549eaab334d8472b498fd341f24262108eb01c99c2ad1d362.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 148, + 553, + 212 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/361e48c86ec61bb06e172fe2e8bfa11ebc51ee8553609c8b4c0b67aa5cf744d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 148, + 666, + 212 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a309244325ed4cb5cfe637a343a4fbc7637c7c9bd1366844a7a819d47e257c16.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 148, + 781, + 212 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "hierarchical grid. The decoder architecture is illustrated in figure 2 and further details can be found in the supplementary materials.", + "bbox": [ + 212, + 305, + 782, + 334 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LOD inductive bias. During the reconstruction process, the gradients of the ray samples are backpropagated through the residual connections and aggregated for each grid level. Due to the pyramidal resolution of the hierarchical grid, a grid code at coarse levels influences a large scene volume and is thus supervised by more ray samples. Therefore, the gradient of a grid code increases as the level's resolution decreases, and as long as the associated samples' gradients are uniform. However, once the backpropagation through the residual connections reach a level whose associated samples have divergent enough gradients (meaning the error is finer than this level's resolution), the result of the aggregation will be mitigated. Hence, the level with the maximum correction is always the coarser level where the samples' gradients are still uniform. This property naturally induces corrections at the proper grid level and LOD. In figure 3, we illustrate this inductive bias by displaying different continuous LOD of a scene while the reconstruction is only supervised for its finest LOD. Not only this is a useful property when an unsupervised multi-scale representation is needed (as demonstrated in section 4.2), this also guarantees a more robust convergence (as illustrated in section 4.3 and section 4.4).", + "bbox": [ + 212, + 335, + 784, + 592 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Distance-Aware Forward Mapping", + "text_level": 1, + "bbox": [ + 214, + 616, + 542, + 632 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cone Casting in contracted space. In order to accommodate the scene rendering process to the variation of observation distances, we introduce a distance-aware forward mapping mechanism. Similarly to the cone casting of [7], it relies on assigning to each sample a latent feature whose LOD is inversely proportional to the sample-camera distance. However, unlike [7, 21], our model relies on the use of space contraction to allow the reconstruction of unbounded scenes.", + "bbox": [ + 212, + 643, + 784, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To define the LOD of a sample at distance $d$ of the camera, we first compute the associated volume of the cone cast pixel cube in the world space coordinates (see fig. 2). Assuming the pixel is a square of size $c$ at distance 1 of the camera ( $c$ depends on the image resolution and camera's FOV), the volume is thus $V = (dc)^3$ . To take into account the space contraction, we then proceed to contract the volume. Denoting $J$ the Jacobian of the contraction function at the sample's location $p$ , $V_{contract} = V\\det(J(p))$ . In practice, we compute the", + "bbox": [ + 212, + 734, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "RING-NeRF", + "bbox": [ + 643, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "analytical derivation of the contraction function depending on $p$ . More details can be found in the supplementary materials.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Assuming the $N$ -th feature grid's resolution in our hierarchy can be written as $f^{N}b$ with $b$ the resolution of the grid $\\delta_0$ and $f$ the growth factor, finding the appropriate LOD $L \\in \\mathbb{R}^+$ means finding a virtual grid4 of resolution $f^{L}b$ whose cell's volume is equal to the previously computed volume (in contracted space). Because we are working in the contracted space of size 1, the volume of a cell of the virtual grid of LOD $L$ is $(\\frac{1}{f^{L}b})^3$ . The LOD $L$ associated to one sample in the contracted space is thus given by:", + "bbox": [ + 212, + 176, + 787, + 282 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\left(d c\\right) ^ {3} \\det (J (p)) = \\left(\\frac {1}{f ^ {L} b}\\right) ^ {3} \\Longleftrightarrow L = - \\frac {\\log \\left(d c b \\sqrt [ 3 ]{\\det (J (p))}\\right)}{\\log (f)} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 284, + 294, + 785, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Note that this process is close to ZipNeRF [4] and VR-NeRF [24]. However, the first one rather derives a contracted scale factor from its Gaussian samples while the latter directly computes the LOD in the contracted space, which can be considered an approximation of our computation.", + "bbox": [ + 212, + 345, + 782, + 405 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Forward mapping. As illustrated in Figure 2, we use the determined LOD to compute a distance-dependent weighted sum of the features, which is fed to our decoder and transformed into density (or SDF) and radiance.", + "bbox": [ + 212, + 406, + 785, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Continuous Coarse-To-Fine and Resolution Extensibility", + "text_level": 1, + "bbox": [ + 214, + 473, + 725, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Recent works proposed to use coarse-to-fine optimization to improve the stability of NeRF models, especially when facing more challenging setups, including with few images [25] and surface-based models [11]. It consists of optimizing progressively the different LOD of the representation, from coarse levels to the most precise ones. The goal of this progressive optimization is to avoid the shape-radiance ambiguity [28] by introducing a strong regularization through LOD restriction, then relaxing progressively this regularization once the coarse geometry of the scene is reconstructed to recover the details of the scene.", + "bbox": [ + 212, + 498, + 787, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The coarse-to-fine reconstruction process of RING-NeRF consists of estimating progressively the LOD of the mapping function from the coarsest to the finest ones. In practice, it implies, during the cone casting, to clamp the samples' LOD up to a maximal LOD $l$ , the grids of level $l$ and above being set to zero and not optimized. Moreover, since our architecture provides continuous LOD, the coarse-to-fine optimization can be achieved continuously in the LOD space by using a linear scheduler $l = (l_0 + \\frac{n}{n_{ctf}}) \\in \\mathbb{R}^+$ with $n$ the current epoch, $n_{ctf}$ a hyperparameter describing the speed of the process and $l_0$ defining the number of used grids at initialization; up to a specified maximum resolution.", + "bbox": [ + 212, + 619, + 787, + 756 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Furthermore, the RING-NeRF architecture is more adapted than I-NGP-based architectures [11] for coarse-to-fine training. Indeed, for solutions based on the concatenation of features, keeping grid values to zero implies that some", + "bbox": [ + 212, + 757, + 787, + 801 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "D. Petit et al.", + "bbox": [ + 271, + 114, + 364, + 127 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "4 A virtual grid of LOD $L \\in \\mathbb{R}^{+}$ corresponds to a grid of resolution $f^{L}b$ that is not explicitly stored in memory but whose elements can be computed from other grids.", + "bbox": [ + 217, + 809, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dimensions of the decoder's latent space are not supervised. When a grid starts to be optimized, those unsupervised dimensions are suddenly used. The weights of the decoder thus need to be refined, with a global effect on the whole scene. On the opposite, our solution keeps supervising all the dimensions of the decoder latent space, and when a new grid gets optimized, it only implies more degrees of freedom to define the mapping function between the scene space and the decoder latent space. This also means that our scene reconstruction can be refined by adding dynamically new grid levels without modifying the decoder's weights or previously trained grids, as we demonstrate experimentally in section 4.5. This resolution extensibility property opens the path to adaptive resolution models, where the precision used to describe an area depends on the details needed, to optimize efficiency both in memory consumption and training duration.", + "bbox": [ + 212, + 146, + 787, + 328 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 353, + 375, + 372 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In these experiments, we intend to highlight the versatility of RING-NeRF by evaluating it on several tasks. After introducing implementation details in section 4.1, we evaluate our model on novel view synthesis with changes in observation distances (sec. 4.2). Then, we explore how robust is RING-NeRF first with few view reconstruction (sec. 4.3) and then without scene-specific initialization for SDF reconstruction (sec. 4.4). Finally, we demonstrate the capacity of our architecture to perform LOD extensibility, as a first step towards adaptive reconstruction (sec. 4.5).", + "bbox": [ + 212, + 387, + 787, + 508 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Implementation", + "text_level": 1, + "bbox": [ + 215, + 532, + 393, + 549 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our model is based on the PyTorch framework Nerfstudio [20]. We build upon its core method named Nerfacto, which combines ideas from several papers for fast and qualitative renders of unbounded complex scenes. This makes it an accessible baseline with a state-of-the-art quality/time ratio. Because NeRF pipelines contain a high number of small but decisive choices of implementation (eg. some frameworks choose to train their models image by image while Nerfstudio jointly and randomly samples across all images), we decided to use as much as possible Nerfstudio-based baselines for fairer comparisons. All of these models are trained on one Nvidia-A100 GPU. The reported times correspond to the approximated training duration of the models. Configuration details, further experiments and ablatives on our contributions can be found in the supplementary materials.", + "bbox": [ + 212, + 560, + 787, + 727 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Novel View Synthesis and Anti-Aliasing", + "text_level": 1, + "bbox": [ + 215, + 751, + 591, + 767 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This experiment aims to evaluate the reconstruction quality through the ability to synthesize viewpoints that are not supervised during the reconstruction. These new viewpoints differ from the angle of observation, but also from the distance of observation. The latter is particularly important since a reconstruction or a", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "RING-NeRF", + "bbox": [ + 643, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/7ca447ece0bbf99aef4a60643c13eb96a88b68e2850a5030e0555e3c0bd4fb90.jpg", + "table_caption": [ + "Table 1: Novel View Synthesis performances for the Mono-Scale setup (trained on the full resolution images only) on the 360 Dataset. The indicated resolutions refer to the resolution of the renders." + ], + "table_footnote": [], + "table_body": "
Full Res.1/2 Res.1/4 Res.1/8 Res.Time ↓
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Nerfacto [20]27.090.7790.18126.810.7820.16225.220.7110.23423.320.6360.2970.45h
PyNeRF [21]27.870.8020.16021.930.7130.21121.150.6620.26420.390.6190.3120.96h
Zip-NeRF [4]28.060.8080.15416.580.5960.31911.880.4240.4659.660.3230.5231.10h
RING-NeRF28.090.7990.15727.180.8040.13825.820.7860.14224.380.7370.1670.45h
", + "bbox": [ + 215, + 198, + 784, + 258 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "rendering process that does not take correctly into account the distance of observation leads to artefacts, from over-contrasted rendering to aliasing phenomenon. The challenge is to avoid these artefacts while keeping the reconstruction and rendering process as fast as possible.", + "bbox": [ + 212, + 276, + 782, + 335 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Dataset. The evaluation relies on the dataset introduced by Mip-NeRF-360 [3]. This dataset is composed of 9 scenes, each containing both a central area and complex background in both inside and outside setups. Since the trajectory keeps a constant distance to the central part, each viewpoint is represented with a pyramid of 4 different image resolutions to simulate a variation of the distance of observation combined with a change of the camera FOV, following the Mip-NeRF [2] original anti-aliasing evaluation pipeline.", + "bbox": [ + 212, + 335, + 784, + 441 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Algorithms. We compare our solution against several grid-based NeRF baselines, both with (PyNeRF [21] and Zip-NeRF [4]) and without anti-aliasing processing (Nerfacto), using their Nerfstudio implementations.", + "bbox": [ + 212, + 441, + 784, + 487 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Protocol. For each scene, we train the models with two different setups: the mono-scale setup that uses only the image with the highest resolution for each viewpoint, and the multi-scale setup which uses the whole pyramid of images for each viewpoint. Note that, for these two setups, we evaluate the performances on the whole resolution pyramid, with usual metrics (PSNR, SSIM, LPIPS).", + "bbox": [ + 212, + 488, + 784, + 561 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results. First of all, regarding novel view synthesis quality using the monoscale setup for both training and testing, as referred to in the \"Full Res.\" column of table 1, we notice that our architecture provides on par performances with Zip-NeRF, slightly better results than PyNeRF, and a more important gap with Nerfacto. While simple, RING-NeRF succeeds in performing state-of-the-art performances on a real single-scale dataset. Regarding quality for the multi-scale setup, as presented in table 2, we observe that all the algorithms that consider the distance of observation perform very similarly in terms of quality. On the", + "bbox": [ + 212, + 563, + 785, + 684 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/4e1ab55963528c40ffc68316a7943fac1db4d2db9816f2944fd16e7c8608c3f7.jpg", + "image_caption": [ + "Fig. 4: Examples of image renderings at 1/8th resolution from models solely trained with the full resolution images. RING-NeRF is the only method capable of producing coherent aliasing-free renderings thanks to its LOD inductive bias." + ], + "image_footnote": [], + "bbox": [ + 220, + 702, + 328, + 773 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/e10b9442e45182a7f15afc5fefbc310cee93cb4f1265391879c441bf6feed875.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 702, + 439, + 773 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/9b772adcc99b6dacb6a804e24fcf7d7855c02a82ccf420e8c917ed836757b4fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 702, + 555, + 773 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/9c58a02448d7a3bd308a112f08b9f77255d167ded6ad2008fc4ec33aa458832b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 557, + 702, + 666, + 773 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/08e2a35d1620d5f24568c3546003dbfd1345c07123da38c82a92fc70e56101fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 702, + 782, + 773 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "D. Petit et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/aaa74f02bea0e77ff08b7095c7dafc86f28b1bf9b737ea8c274863ad0f625bc3.jpg", + "table_caption": [ + "Table 2: Novel View Synthesis performances for the Multi-Scale setup (trained jointly on every resolution) on the 360 Dataset. The indicated resolutions refer to the resolution of the renders." + ], + "table_footnote": [], + "table_body": "
Full Res.1/2 Res.1/4 Res.1/8 Res.Time ↓
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Nerfacto [20]25.980.6680.36727.310.7920.17627.160.8050.13925.300.7430.1990.45h
PyNeRF [21]27.650.7810.19429.470.8570.09230.510.8870.06330.860.8930.0550.96h
Zip-NeRF [4]27.540.7810.19129.340.8580.09930.360.8890.06430.900.9000.0471.10h
RING-NeRF27.660.7820.19429.550.8560.09430.510.8880.06230.860.9010.0480.45h
", + "bbox": [ + 215, + 198, + 782, + 258 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "other side, Nerfacto performs quite poorly. Its performances are especially low at the coarsest resolution, with overflowing artefacts, as illustrated in figure 1.", + "bbox": [ + 212, + 275, + 782, + 303 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Finally, we evaluate the capacity of the models to generalize over new resolutions by training them on full-resolution images and then evaluate them on smaller resolutions. Qualitative results are shown in figure 4 while quantitative results can be found in the 1/2-th, 1/4-th and 1/8-th res. columns of table 1. Because most anti-aliasing methods need LOD-specific supervision to correctly function (including Zip-NeRF and PyNeRF), they cannot render coherent anti-aliased images in this setup. However, PyNeRF behaves better than ZipNeRF with coherent although very aliased renderings, as it decides to limit the LOD used for rendering inside the range of LOD seen during training. RING-NeRF, with his LOD inductive bias, is the only architecture capable of producing anti-aliased renderings from novel observation distances and thus outperforms every other method. While this experiment can seem somewhat esoteric as training on a multi-resolution images pyramid is rather easy, this increases GPU memory and total training time. Moreover, depending on the resolution and the trajectory, it is not trivial to choose the accurate number of scales in the image pyramid to supervise correctly every grid in the hierarchy and especially the coarsest grids.", + "bbox": [ + 212, + 305, + 784, + 546 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Regarding the reconstruction processing time reported in the \"Time\" column of both table 1 and table 2, since RING-NeRF does not rely on the multiplication of either sample or decoder, it processes as quickly as the fastest Ner-facto both on mono-scale and multi-scale setups. On the other hand, the other anti-aliasing methods, PyNeRF with its per-LOD MLP and Zip-NeRF with its super-sampling, are approximately 2.5 times slower.", + "bbox": [ + 212, + 547, + 784, + 637 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In conclusion, our solution provides the best quality-speed trade-off since it is both on par with the best quality method and the fastest method, in monoscale as well as in multi-scale setups. Furthermore, RING-NeRF is the only solution capable of creating coherent and anti-aliased renderings when facing novel observation distances unseen during training.", + "bbox": [ + 212, + 638, + 784, + 713 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Few Viewpoints Supervision", + "text_level": 1, + "bbox": [ + 214, + 737, + 493, + 752 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "This experiment aims to evaluate the influence of the RING-NERF architecture and pipeline on the reconstruction robustness to limited supervision viewpoints.", + "bbox": [ + 212, + 763, + 782, + 792 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Dataset. We evaluate our contribution on the object-centric real dataset DTU [9], often used in few-viewpoints evaluations.", + "bbox": [ + 212, + 795, + 782, + 823 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Algorithms. We compare our architecture against several baselines: Mip-NeRF", + "bbox": [ + 214, + 824, + 782, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "RING-NeRF", + "bbox": [ + 643, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "[2], and FreeNerf [25] (a state-of-the-art method for this task), for vanilla architectures and Nerfacto for grid-based architectures. To better demonstrate the intrinsic stability brought by RING-NeRF, we also developed the Nerfacto+ architecture, which corresponds to a Nerfacto architecture coupled with a coarse-to-fine training based on a progressive activation of the grids (the rest of the decoder's input being filled with zeros). For Nerfacto+ and RING-NeRF, we also add FreeNeRF's loss that penalizes the density of the first $M = 10$ samples of each ray to reduce as much as possible artefacts in front of the cameras. As an ablative experiment, we also evaluate RING-NeRF using discrete coarse-to-fine (fixed LOD increment of 1 rather than the proposed continuous increase).", + "bbox": [ + 212, + 145, + 787, + 297 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Protocol. We follow FreeNeRF's evaluation pipeline, including the number of supervision viewpoints (3 to 9), the choice of these views among the dataset and the evaluations using masks of the object. Since we are using the same protocol, the results of FreeNeRF and Mip-NeRF were taken out of FreeNeRF's article.", + "bbox": [ + 212, + 297, + 787, + 358 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Results. Evaluation results are reported in table 3. We first notice an important difference between vanilla and grid-based baselines. While Mip-NeRF faces troubles in reconstructing the scene with 3 views, the method seems to find coherency when adding more images. However, Nerfacto struggles much more to form a consistent 3D scene even when using 9 images (see figure 1). Even though the grid-based baseline is way faster to train, its design implies more instability when facing a small number of images. This does not mean however that few viewpoints are incompatible with grid-based methods. Using progressive training coupled with a very simple and generalizable regularization, both Nerfacto+ and our architecture succeed in creating coherent geometry. Nonetheless, RING-NeRF considerably outperforms Nerfacto+, with a PSNR difference varying from 3 to 4, demonstrating the stability increase brought by our architecture, and also outperforms Mip-NeRF for the configuration with 3 and 6 supervision images. The discrete coarse-to-fine version of RING-NeRF performs in-between Nerfacto+ and the complete RING-NeRF. This showcases both the intrinsic interest of the proposed architecture against the Nerfacto+ and the relevance of the continuous coarse-to-fine mechanism. FreeNerf remains the best performer of all in terms of quality, but with a reconstruction time that is extremely slower than RING-NeRF, the former achieving its reconstruction in 2.56 hours while the latter only requires less than 10 minutes. RING-NeRF thus offers a better quality-speed trade-off for the few-view reconstruction issue (see figure 1).", + "bbox": [ + 212, + 358, + 787, + 676 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/65d5f8bca8f22a6197d04fca006c945742a652685282a0fb7cc3a9072703fb6b.jpg", + "table_caption": [ + "Table 3: Performances of reconstruction from few viewpoints on the DTU dataset. The reported metrics are computed based on the mask of the object." + ], + "table_footnote": [], + "table_body": "
#ImagesPSNR ↑SSIM ↑LPIPS ↓Time ↓
369369369
Mip-NeRF [2]8.6816.5423.580.5710.7410.8790.3530.1980.0922.56h
FreeNeRF [25]19.9223.2525.380.7870.8440.8880.1350.0950.0672.56h
Nerfacto [20]9.359.759.780.5670.6040.6470.3850.3310.3260.15h
Nerfacto+13.6116.6119.330.6390.6990.7590.2760.2180.1510.15h
RING-NeRF16.1820.4723.190.7130.8080.8470.2000.1270.0850.15h
w/ discrete CtF15.7920.1622.930.7060.7850.8470.2010.1270.0850.15h
", + "bbox": [ + 243, + 731, + 761, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "D. Petit et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/7049b4c3facffb648f9dab42ba4e216fd0364798d6c1a1599c7d83699fa23156.jpg", + "table_caption": [ + "Table 4: SDF reconstruction performances when foregoing the scene-specific SDF Initialization on the Replica Dataset. The Chamfer distance is in centimeters." + ], + "table_footnote": [], + "table_body": "
PSNR ↑SSIM ↑LPIPS ↓Chamfer-L1 ↓Training Time ↓
NeuS-facto [26]24.620.7780.34717.611.4h
NeuralAngelo [11]30.790.9160.076113.693.40 h
RING-NeRF37.180.9690.01945.711.28 h
", + "bbox": [ + 300, + 184, + 702, + 232 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 SDF Reconstruction without Initialization", + "text_level": 1, + "bbox": [ + 215, + 250, + 609, + 265 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "SDF reconstruction is known to be a more unstable process than density-based NeRF [1], requiring a scene-specific initialization to converge. This initialization becomes an issue in complex environments and incremental setups, where several types of scenes can co-exist and are not necessarily known beforehand. Therefore, in this experiment, we evaluate the ability of RING-NeRF and other SotA architectures to achieve SDF reconstruction without scene-specific initialization.", + "bbox": [ + 212, + 271, + 784, + 361 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Dataset. The evaluation is achieved on a subset of 7 scenes of the Replica [18] synthetic indoor dataset. A Tanks & Temples [10] example is provided in supplementary materials with corresponding analysis.", + "bbox": [ + 212, + 362, + 782, + 407 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Algorithms. We compare our architecture to two SDF methods, all of them implemented in the same SDFStudio [26] branch of the Nerfstudio framework for fairer comparisons: NeuS-facto, an adaptation of NeuS for grid-based methods with Nerfacto modules, and an implementation of NeuralAngelo. For these experiments, our model is built upon the NeuS-facto baseline, using in particular the same NeuS-based SDF-to-density transformation [22].", + "bbox": [ + 212, + 407, + 784, + 497 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Protocol. To evaluate the impact of the architecture and pipeline over the convergence and stability, we suppress the inverted sphere SDF initialization scheme and use a random initialization for the model.", + "bbox": [ + 212, + 498, + 784, + 542 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Results. Evaluation results are shown in table 4. First of all, NeuS-facto faces low rendering and reconstruction metrics, due to catastrophic failure in most of the tested scenes (see figure 1) since the model tends to re-draw the 2D images in front of the camera. Regarding NeuralAngelo, its relatively high PSNR demonstrates its ability to synthesize satisfying RGB. However, as illustrated in figure 1 and highlighted by the reconstruction metrics, the underlying geometry of the scene is poorly reconstructed, without any fine details. Finally, our method RING-NeRF is by far the best performer, with much higher PSNR and a better geometry including fine details (see figure 1), although a bit noisy. The simple architecture of RING-NeRF also permits faster epochs, thus faster training.", + "bbox": [ + 212, + 542, + 785, + 694 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 LOD Extensibility", + "text_level": 1, + "bbox": [ + 215, + 712, + 413, + 728 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "This experiment aims to demonstrate RING-NeRF's unique ability to increase dynamically the level of detail of the scene representation.", + "bbox": [ + 212, + 734, + 782, + 763 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Dataset. The scan 114 of the DTU dataset is used for this experiment.", + "bbox": [ + 214, + 763, + 732, + 777 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Algorithms. Because I-NGP architectures [4, 11, 14] cannot perform LOD extensibility (due to the fixed decoder's input size), only RING-NeRF is evaluated.", + "bbox": [ + 212, + 779, + 782, + 809 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Protocol. We train our model using two different configurations : one low resolution with a 3 levels grid hierarchy and one high resolution with 5 levels (the", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "RING-NeRF", + "bbox": [ + 643, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/4eaf320a4870d993ff70708bcd7991e489b3dd4205e12abfd74d6e489faf691f.jpg", + "image_caption": [ + "Fig. 5: Learning curves and final renderings of RING-NeRF models with different grid configurations trained either jointly or incrementally." + ], + "image_footnote": [], + "bbox": [ + 225, + 146, + 524, + 287 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/f5bcb97c15a12b8e9aa96373f6fd3402f54633e4ca1063c0c5d15018049793c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 146, + 647, + 215 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/d0630a3dc21b8411ce3938d4ede0f764a4739c201d41375546cfe8f239a71d99.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 146, + 772, + 215 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/b638c4748cd31e44ad39d040f95fb7602f257637aead1d88f5d67aa4b382952d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 218, + 648, + 287 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/c6716be8e8b0c7f9a854e46f636030b993ff8f8586f0e5917d978598e0117d5f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 218, + 772, + 287 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "grid resolutions are 16, 32, 64, 128 and 256). For these two configurations named \"3 levels\" and \"5 levels\", every grid and the decoder are trained simultaneously. We proceed to showcase the extensibility of RING-NeRF by adding grids to the \"3 levels\" configuration that is previously trained. We first train one grid of resolutions 128 (\"3+1 levels\" in Figure 5) with the three initial grids and the decoder frozen and then train another grid of resolutions 256 (\"3+1+1 levels\" in Figures 5) with the four grids and the decoder frozen.", + "bbox": [ + 214, + 349, + 787, + 455 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Results. Figure 5 shows that the configuration \"3+1+1 levels\" results in the same rendering quality than the \"5 levels\" one. This demonstrates the ability of our model to dynamically change the resolution of the grid hierarchy. This is an important step towards the development of an adaptive architecture which locally chooses the resolution of the representation based on the scene's content. This allows to drastically reduce the number of parameters, helping in improving the memory footprint, the training duration and the model's robustness.", + "bbox": [ + 214, + 455, + 787, + 561 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion and Perspectives", + "text_level": 1, + "bbox": [ + 215, + 597, + 529, + 614 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we introduced RING-NeRF, a simple and versatile NeRF pipeline that provides two inductive biases by design: a continuous multi-scale representation of the scene, and an invariance of the decoder latent space over spatial and scale domains. Coupled with a distance-aware forward mapping and a continuous coarse-to-fine reconstruction process, our pipeline demonstrated experimentally its versatility with on-par performances with dedicated state-of-the-art solutions for anti-aliasing or reconstruction from few viewpoints. It even outperforms them in terms of robustness to scene-specific initialization for SDF reconstruction. Furthermore, it is highly efficient and is not limited to object-centric scenes.", + "bbox": [ + 212, + 641, + 787, + 777 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Future work will study the impact of RING-NeRF on other challenging use cases, such as facing inaccurate camera poses [16] and SLAM [29]. We will also use the extensibility property of our architecture to develop memory-efficient sparse Neural Fields, which is considered to be a limit of most grid-based models.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "D. Petit et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This publication was made possible by the use of the CEA List FactoryIA supercomputer, financially supported by the Ile-de-France Regional Council.", + "bbox": [ + 215, + 176, + 782, + 205 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 229, + 321, + 244 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Atzmon, M., Lipman, Y.: Sal: Sign agnostic learning of shapes from raw data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2565-2574 (2020)", + "2. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)", + "3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mipnerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022)", + "4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-nerf: Anti-aliased grid-based neural radiance fields. ICCV (2023)", + "5. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: European Conference on Computer Vision (ECCV) (2022)", + "6. Dou, Y., Zheng, Z., Jin, Q., Ni, B.: Multiplicative fourier level of detail. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1808-1817 (2023)", + "7. Hu, W., Wang, Y., Ma, L., Yang, B., Gao, L., Liu, X., Ma, Y.: Tri-miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19774-19783 (2023)", + "8. Jain, A., Tancik, M., Abbeel, P.: Putting nerf on a diet: Semantically consistent few-shot view synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5885-5894 (2021)", + "9. Jensen, R., Dahl, A., Vogiatzis, G., Tola, E., Aanaes, H.: Large scale multi-view stereopsis evaluation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 406-413 (2014)", + "0. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics 36(4) (2017)", + "11. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8456-8465 (2023)", + "12. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020)", + "13. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)", + "14. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + ], + "bbox": [ + 218, + 258, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "RING-NeRF", + "bbox": [ + 643, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5480-5490 (2022)", + "16. Park, K., Henzler, P., Mildenhall, B., Barron, J.T., Martin-Brualla, R.: Camp: Camera preconditioning for neural radiance fields. ACM Transactions on Graphics (TOG) 42(6), 1-11 (2023)", + "17. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. ICCV (2021)", + "18. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019)", + "19. Takikawa, T., Litalien, J., Yin, K., Kreis, K., Loop, C., Nowrouzezahrai, D., Jacobson, A., McGuire, M., Fidler, S.: Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11358-11367 (2021)", + "20. Tancik, M., Weber, E., Ng, E., Li, R., Yi, B., Wang, T., Kristoffersen, A., Austin, J., Salahi, K., Ahuja, A., et al.: Nerfstudio: A modular framework for neural radiance field development. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-12 (2023)", + "21. Turki, H., Zollhöfer, M., Richardt, C., Ramanan, D.: Pynerf: Pyramidal neural radiance fields. In: Thirty-Seventh Conference on Neural Information Processing Systems (2023)", + "22. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS (2021)", + "23. Wynn, J., Turmukhambetov, D.: DiffusioNeRF: Regularizing Neural Radiance Fields with Denoising Diffusion Models. In: CVPR (2023)", + "24. Xu, L., Agrawal, V., Laney, W., Garcia, T., Bansal, A., Kim, C., Rota Bulò, S., Porzi, L., Kontschieder, P., Božić, A., Lin, D., Zollhöfer, M., Richardt, C.: VR-NeRF: High-fidelity virtualized walkable spaces. In: SIGGRAPH Asia Conference Proceedings (2023). https://doi.org/10.1145/3610548.3618139", + "25. Yang, J., Pavone, M., Wang, Y.: Freenerf: Improving few-shot neural rendering with free frequency regularization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8254-8263 (2023)", + "26. Yu, Z., Chen, A., Antic, B., Peng, S., Bhattacharyya, A., Niemeyer, M., Tang, S., Sattler, T., Geiger, A.: Sdfstudio: A unified framework for surface reconstruction (2022), https://github.com/autonomousvision/sdfstudio", + "27. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems 35, 25018-25032 (2022)", + "28. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492 (2020)", + "29. Zhu, Z., Peng, S., Larsson, V., Xu, W., Bao, H., Cui, Z., Oswald, M.R., Pollefeys, M.: Nice-slam: Neural implicit scalable encoding for slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12786-12796 (2022)", + "30. Zhuang, Y., Zhang, Q., Feng, Y., Zhu, H., Yao, Y., Li, X., Cao, Y.P., Shan, Y., Cao, X.: Anti-aliased neural implicit surfaces with encoding level of detail. Siggraph Asia 2023 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 830 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "D. Petit et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 15 + } +] \ No newline at end of file diff --git a/2024/RING-NeRF _ Rethinking Inductive Biases for Versatile and Efficient Neural Fields/5e093f34-06af-49a4-a5fe-2e79d5819847_model.json b/2024/RING-NeRF _ Rethinking Inductive Biases for Versatile and Efficient Neural Fields/5e093f34-06af-49a4-a5fe-2e79d5819847_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ffa9fba94843353f7a728d9dc801f263eea686a8 --- /dev/null +++ b/2024/RING-NeRF _ Rethinking Inductive Biases for Versatile and Efficient Neural Fields/5e093f34-06af-49a4-a5fe-2e79d5819847_model.json @@ -0,0 +1,2091 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.236, + 0.141, + 0.768, + 0.185 + ], + "angle": 0, + "content": "RING-NeRF : Rethinking Inductive Biases for Versatile and Efficient Neural Fields" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.212, + 0.768, + 0.244 + ], + "angle": 0, + "content": "Doriand Petit\\(^{1,2}\\), Steve Bourgeois\\(^{1}\\), Dumitru Pavel\\(^{1}\\), Vincent Gay-Bellile\\(^{1}\\), Florian Chabot\\(^{1}\\), and Loic Barthe\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.285, + 0.255, + 0.718, + 0.283 + ], + "angle": 0, + "content": "1 Université Paris-Saclay, CEA, List, F-91120, Palaiseau, France \n2 IRIT, Université Toulouse III, CNRS, France" + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.306, + 0.756, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.526, + 0.788, + 0.596 + ], + "angle": 0, + "content": "Fig. 1: RING-NeRF is a simple and versatile architecture which tackles many NeRF common issues such as robustness to distance of observation, few view supervision and lack of scene-specific initialization for SDF-based reconstruction. It provides on-par performances in terms of quality with SotA dedicated solutions [4,11,25] and and in terms of efficiency with fast methods [20,26]." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.619, + 0.741, + 0.84 + ], + "angle": 0, + "content": "Abstract. Recent advances in Neural Fields mostly rely on developing task-specific supervision which often complicates the models. Rather than developing hard-to-combine and specific modules, another approach generally overlooked is to directly inject generic priors on the scene representation (also called inductive biases) into the NeRF architecture. Based on this idea, we propose the RING-NeRF architecture which includes two inductive biases: a continuous multi-scale representation of the scene and an invariance of the decoder's latent space over spatial and scale domains. We also design a single reconstruction process that takes advantage of those inductive biases and experimentally demonstrates on-par performances in terms of quality with dedicated architecture on multiple tasks (anti-aliasing, few view reconstruction, SDF reconstruction without scene-specific initialization) while being more efficient. Moreover, RING-NeRF has the distinctive ability to dynamically increase the resolution of the model, opening the way to adaptive reconstruction. Project page can be found at: https://cea-list.github.io/RING-NeRF" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "D. Petit et al." + }, + { + "type": "header", + "bbox": [ + 0.262, + 0.148, + 0.74, + 0.162 + ], + "angle": 0, + "content": "Keywords: NeRF \\(\\cdot\\) 3D Reconstruction \\(\\cdot\\) Implicit Neural Representation" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.189, + 0.376, + 0.205 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.221, + 0.788, + 0.401 + ], + "angle": 0, + "content": "Neural Radiance Fields (NeRF) have emerged as a novel method for representing 3D scenes using neural networks. In its original design [13], a simple multi-layer perceptron (MLP) is trained to reproduce a continuous 5D function that outputs the density and radiance emitted in each direction \\((\\theta, \\phi)\\) at each point \\((x, y, z)\\) in space. This approach inspired many works due to the impressive quality of its novel view synthesis as well as its simplicity. However, the so-called \"vanilla\" NeRF architectures converge very slowly as they use large deep neural networks. Instant-NGP [14] introduced a new architecture based on a hierarchical hash-grid pyramid of learnable feature codes describing the 3D scene and combined with a much shallower MLP, called decoder, that transforms the concatenation of the codes, interpolated from the grids, into density and radiance. Resulting in local updates, this method reduced the training duration from hours to minutes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.787, + 0.446 + ], + "angle": 0, + "content": "The great majority of current solutions are now based on these two standard architectures, and many of them are focused on overtaking their associated limitations in terms of:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.787, + 0.478 + ], + "angle": 0, + "content": "Nature of scene - by transitioning from object-centric scenes to open unbounded scenes, using mostly space contraction [3];" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.787, + 0.539 + ], + "angle": 0, + "content": "Robustness - by managing free motion trajectories with variations of the observation distance while avoiding aliasing artefacts, mostly through the integration of Level of Detail (LOD) in the model [2,4]; or by reducing drastically the number of supervised views through different kinds of regularization [8,15,25];" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.787, + 0.584 + ], + "angle": 0, + "content": "Extensibility - by shifting from a holistic and fixed reconstruction process to an incremental (extensibility in the number of views) and adaptive (extensibility in resolution) process, mainly through the use of a frozen decoder [12, 29]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.787, + 0.658 + ], + "angle": 0, + "content": "However, most solutions solely focus on one of these limitations and introduce specific and complex mechanisms that both increase the computational cost and lessen the possibilities of combination. The ability to solve jointly these main issues is however essential in real-world applications which often require a robust and extensible reconstruction in an unknown and unbounded environment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In this article, rather than introducing yet another heavy and task-specific solution, we propose RING-NeRF, a versatile and simple NeRF architecture by rethinking usual grid-based models to introduce two inductive biases. We first represent a 3D scene as a continuous multi-scale representation and also make the decoder's latent space invariant in position and scale. Together, these two priors enable the production of intrinsic continuous LOD of the scene without explicit supervision. We demonstrate experimentally that, when combined with adapted cone casting and coarse-to-fine optimization, the resulting architecture is able to compete on several tasks with on-par quality performances with dedicated state-of-the-art solutions while improving speed, robustness and extensibility. The overall process is also simple, easy to implement and generic enough to be coupled with specific solutions. Our contributions can be summarized as follows:" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.645, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RING-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.147, + 0.786, + 0.191 + ], + "angle": 0, + "content": "1. an architecture that, by construction, represents the scene with a continuous and unbounded level of detail without the need for LOD-specific supervision and which permits resolution-adaptive reconstruction;" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.193, + 0.787, + 0.252 + ], + "angle": 0, + "content": "2. a distance-aware forward mapping compatible with scene contraction, that takes benefits of the continuous multi-scale representation with an adapted cone casting process to avoid aliasing artefacts when varying the observation distance." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.254, + 0.786, + 0.299 + ], + "angle": 0, + "content": "3. a continuous coarse-to-fine reconstruction process that improves the convergence and stability (especially in challenging setups such as supervision with few viewpoints or no scene-specific initialization for SDF reconstruction)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.324, + 0.388, + 0.34 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.357, + 0.785, + 0.417 + ], + "angle": 0, + "content": "From its original iteration [13], a majority of current research focuses on overtaking limitations of NeRF-based reconstruction in terms of adaptability to various natures of scene, robustness (to varying observation distances or limited amount of viewpoints) and reconstruction extensibility." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.788, + 0.689 + ], + "angle": 0, + "content": "Adaptability to various natures of the scene. The ability of Neural Fields to reconstruct various natures of scene depends on three factors. The first one is related to its architecture itself. Tri-plane architectures [5, 7, 30] are mostly designed for object-centric reconstruction (as they provide a higher density of information in the center of the scene) while vanilla and 3D grid-based NeRF are able to cover a wider variety of scenes, though they initially were still constrained to a limited volume. A second factor is related to the representation of the 3D space of the scene, especially to represent distant elements in open scenes. Some approaches use two different NeRF models to reconstruct separately the foreground and the background [28], whereas others apply space contraction to the 3D scene coordinates [3, 4, 28] to map the infinite scene volume into a bounded one. The last factor is related to the initialization of the Neural Fields. While the random initialization of density-based NeRF can adjust to almost any nature of scenes, the convergence of SDF-based (Signed Distance Function) Neural Fields is extremely sensitive to their parameters' initialization, as stated in [1]. Current solutions rely on a scene-specific initialization (using an SDF field representing a sphere for outdoor scenes or an inverted sphere for indoor scenes), making them unable to adapt automatically to any scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Robustness to observation distance variations. The initial NeRF model, as well as most of the subsequent works, relies implicitly on the hypothesis of a constant distance of observation to the scene. Indeed, the NeRF model provides a per-3D-point scene density representation and a rendering process which does not take into account the increasing volume covered by a pixel with respect to the distance to the cameras. As underlined in Mip-NeRF [2], this discrepancy induces artefacts such as over-contrasted images or aliasing phenomenon when the distance of observation differs from the ones used at reconstruction time. To avoid these artefacts, the rendering process needs to assess the density and color for a volume instead of a point. In the current state-of-the-art, two main" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "D. Petit et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.388 + ], + "angle": 0, + "content": "approaches are used (or their combination, as in Zip-NeRF [4]). The first solution consists of representing the scene with different levels of detail (LOD), in order to vary the precision of the reconstruction based on the observation distance. This is usually done by using a LOD-aware latent space [2,4,6,19,21,24], meaning that the LOD information is already encoded in the inputs of the MLP. This can be achieved by using a per-LOD decoder as in PyNeRF [21] or by incorporating LOD information in the latent feature as done in Zip-NeRF [4] and VR-NeRF [24]. One main flaw of these solutions is that they require the supervision of every used LOD which makes it impossible to vary the observation distances between the train views and novel synthesised views. A second approach consists of defining the latent representation of a volume as the mean of the latent features of the points included in the volume. It requires integrating the features over the volume, which can be achieved through convolutions for tri-plan representation [7,30], or through super-sampling of the latent space for 3D grids as also done in Zip-NeRF [4]. However, these approaches lengthen the training and rendering processes as they increase the number of computations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.788, + 0.599 + ], + "angle": 0, + "content": "Robustness to limited amount of viewpoints. By construction, NeRF is subject to the shape-radiance ambiguity [28]. If not enough supervision viewpoints are available, the optimization might overfit them while not providing consistent 3D reconstruction nor generalization to non-supervised viewpoints. A first family of solutions to overcome this limitation consists of regularizing the reconstruction process through additional losses, whether via geometric [15,23,27] or semantic [8] regularization. The second family of solutions relies on a progressive reconstruction of the details of the scene, as introduced in FreeNeRF [25] and Nerfies [17]. Restraining the ability of the model to reconstruct a complex scene at early stages enforces the consistency of the reconstruction over the different supervision viewpoints. This presents the advantage of bringing stability while keeping a fairly simple training process. However, these latter solutions mostly rely on Vanilla models for stability purposes and require long training duration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.6, + 0.788, + 0.829 + ], + "angle": 0, + "content": "Reconstruction extensibility. Two kinds of extensibility should be distinguished: extensibility with respect to the number of views or the resolution. The first one is related to reconstructing new scene areas while keeping the previously reconstructed ones unchanged. The usual solution consists of using grid-based approaches with a pre-trained position-invariant decoder that is frozen during the reconstruction [29]. On the other side, the extensibility of the resolution consists in dynamically increasing the level of detail of a previously reconstructed scene. This problem is intractable for classic architectures such as vanilla NeRF and I-NGP since the number of layers (resp. grids) of those approaches cannot increase during the reconstruction. Some rare solutions, such as Neural Sparse Voxel Fields [12], combine a pre-trained decoder with a data structure allowing to dynamically allocate voxels to increase the reconstruction resolution. This is however a crucial stake, as being able to adapt the precision of the representation based on the complexity of the scene permits to optimize the computational efficiency (both in speed and memory requirements) with minimal loss in quality." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.645, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RING-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.144, + 0.779, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.298, + 0.784, + 0.368 + ], + "angle": 0, + "content": "Fig. 2: Overview of RING-NeRF: to render a pixel, the casted cone is sampled with cubes. Depending on the cube volume, the corresponding LOD of the scene is selected and the latent feature is computed using a weighted sum of the grid hierarchy. The density (or SDF) and color of the cube are first decoded from the latent feature with a tiny MLP and then integrated with other samples through volume rendering." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.385, + 0.376, + 0.4 + ], + "angle": 0, + "content": "3 RING-NeRF" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.425, + 0.784, + 0.469 + ], + "angle": 0, + "content": "Rather than focusing on solving one specific problem of NeRF using complex mechanisms, we propose a simple architecture called RING-NeRF constructed with novel inductive biases to tackle NeRF's common issues." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.501, + 0.341, + 0.514 + ], + "angle": 0, + "content": "3.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.533, + 0.785, + 0.687 + ], + "angle": 0, + "content": "RING-NeRF relies on the classic NeRF [13] inverse rendering pipeline which is used to reconstruct a 3D scene from a set of localized frames. For a given image pixel (with its camera's pose), a 3D ray is cast and the 3D scene representation is sampled at \\(N\\) various locations along the ray. The resulting density \\(\\sigma_{i}\\) (or SDF converted to density [22]) and color \\(c_{i}\\) of the samples are then combined with usual volume rendering techniques: \\(\\hat{C}(r) = \\sum_{i=0}^{N-1} T_{i} (1 - \\exp(-\\sigma_{i} \\delta_{i})) c_{i}\\) with \\(T_{i} = \\exp(\\sum_{j=0}^{i-1} \\sigma_{j} \\delta_{j})\\) and \\(\\delta_{i}\\) is the distance between samples. The parameters of the scene representation are then optimized by minimizing the MSE loss \\(||\\hat{C}(r) - C(r)||^{2}\\) between the rendered ray color and the ground truth pixel value, the rendering being differentiable3." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.689, + 0.785, + 0.81 + ], + "angle": 0, + "content": "The originality of RING-NeRF relies on its neural architecture, illustrated in figure 2, which is designed to represent the scene with a continuous and unbounded level of detail without the need for LOD-specific supervision (see section 3.2). We then use this LOD inductive bias to adjust the LOD of the samples with respect to the distance to the camera in contracted space for more accurate renderings (see section 3.3). Finally, combined with a continuous coarse-to-fine optimization, RING-NeRF results in a more robust reconstruction process with an intrinsic LOD extensibility property (see section 3.4)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.784, + 0.84 + ], + "angle": 0, + "content": "3 With SDF, an additional Eikonal loss is also being optimized for SDF regularization." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.128 + ], + "angle": 0, + "content": "D. Petit et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.628, + 0.162 + ], + "angle": 0, + "content": "3.2 NeRF Representation with Inductive Biases" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.171, + 0.788, + 0.352 + ], + "angle": 0, + "content": "Classic grid-based representations, such as the one introduced by I-NGP [14], rely on two key elements: a unique MLP decoder which transforms a latent space into an output space (color, density or SDF), and a 3D (hierarchical) grid of latent features that implicitly defines a unique mapping function from the scene space onto the decoder latent space. We can distinguish two different approaches used to extend the mono-scale representation to a multi-scale representation: conditioning with scale information the decoding process [4,21] and/or defining separate per-LOD mapping functions [21]. In both cases, the training becomes more complex, leading to potential convergence issues as it relies solely on additional specific supervision. Instead, we propose to condition the mapping function itself with scale information and introduce inductive biases to guide the convergence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.353, + 0.788, + 0.399 + ], + "angle": 0, + "content": "Scene representation as multi-scale mapping function. As illustrated in Figure 2, the mapping function at level \\( N \\) is controlled by a grid \\( G_{N} \\) that is implicitly defined through a recursive refinement process over scale:" + }, + { + "type": "equation", + "bbox": [ + 0.41, + 0.411, + 0.785, + 0.45 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} G _ {0} \\leftarrow \\delta_ {0} \\\\ G _ {N} \\leftarrow S \\left(G _ {N - 1}\\right) + \\delta_ {N} \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.454, + 0.788, + 0.589 + ], + "angle": 0, + "content": "where \\(\\delta_{i}\\) is the deviation defined by the i-th latent feature grid and S is the subdivision scheme consisting in increasing the resolution of \\(G_{N - 1}\\) up to the \\(N\\)-th grid resolution (with tri-linear interpolation). In practice, the latent vector associated with a point in the scene for a level of detail \\(N\\) is simply obtained by interpolating linearly the point in the cell for each level inferior or equal to \\(N\\), and summing the results. Since linear interpolation is differentiable, our mapping function is also estimizable. The summing of the interpolation is comparable to a residual connection and gives the name of the architecture RING-NeRF for Residual Implicit Neural Grids." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.59, + 0.788, + 0.697 + ], + "angle": 0, + "content": "Such representation provides several advantages. First, since the mapping function is constructed in a top-down manner, refining recursively one level of detail from the previous one, its number of LOD is unbounded and adding a finer level of detail keeps the coarser ones valid. Secondly, because each \\(\\delta_{i}\\) only represents a deviation of the coarser mapping value \\(G_{i - 1}\\), a continuous LOD representation can be easily obtained by incorporating a weighting factor \\(\\alpha \\in [0,1]\\) in the recursive process:" + }, + { + "type": "equation", + "bbox": [ + 0.42, + 0.709, + 0.785, + 0.724 + ], + "angle": 0, + "content": "\\[\nG _ {L} \\leftarrow S \\left(G _ {N - 1}\\right) + \\alpha \\delta_ {N} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.525, + 0.749 + ], + "angle": 0, + "content": "with \\(L\\in [N - 1,N]\\) and \\(\\alpha = L - (N - 1)\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Spatial and scale invariance of decoder. Since our architecture does not rely on a decoder conditioned on position (unlike [19]) nor scale (unlike [2-4, 21]), the decoder latent space is invariant to translations and scale changes in the scene coordinate. This property makes RING-NeRF more suited for incremental reconstruction in both spatial and scale space, since it ensures that local updates in the spatial and scale domain of the scene can be achieved through the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.645, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RING-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.148, + 0.331, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.149, + 0.443, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.149, + 0.555, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.149, + 0.668, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.149, + 0.782, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.23, + 0.785, + 0.285 + ], + "angle": 0, + "content": "Fig. 3: We demonstrate the LOD inductive bias by training our model with a hierarchy of \\( N = 7 \\) levels where only the last level of the mapping function \\( G_{N} \\) is supervised. We then compute renders at different levels \\( L \\leq N \\). Other examples (including of entire scenes) can be found in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.306, + 0.784, + 0.335 + ], + "angle": 0, + "content": "hierarchical grid. The decoder architecture is illustrated in figure 2 and further details can be found in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.336, + 0.785, + 0.593 + ], + "angle": 0, + "content": "LOD inductive bias. During the reconstruction process, the gradients of the ray samples are backpropagated through the residual connections and aggregated for each grid level. Due to the pyramidal resolution of the hierarchical grid, a grid code at coarse levels influences a large scene volume and is thus supervised by more ray samples. Therefore, the gradient of a grid code increases as the level's resolution decreases, and as long as the associated samples' gradients are uniform. However, once the backpropagation through the residual connections reach a level whose associated samples have divergent enough gradients (meaning the error is finer than this level's resolution), the result of the aggregation will be mitigated. Hence, the level with the maximum correction is always the coarser level where the samples' gradients are still uniform. This property naturally induces corrections at the proper grid level and LOD. In figure 3, we illustrate this inductive bias by displaying different continuous LOD of a scene while the reconstruction is only supervised for its finest LOD. Not only this is a useful property when an unsupervised multi-scale representation is needed (as demonstrated in section 4.2), this also guarantees a more robust convergence (as illustrated in section 4.3 and section 4.4)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.617, + 0.544, + 0.633 + ], + "angle": 0, + "content": "3.3 Distance-Aware Forward Mapping" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.785, + 0.734 + ], + "angle": 0, + "content": "Cone Casting in contracted space. In order to accommodate the scene rendering process to the variation of observation distances, we introduce a distance-aware forward mapping mechanism. Similarly to the cone casting of [7], it relies on assigning to each sample a latent feature whose LOD is inversely proportional to the sample-camera distance. However, unlike [7, 21], our model relies on the use of space contraction to allow the reconstruction of unbounded scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To define the LOD of a sample at distance \\( d \\) of the camera, we first compute the associated volume of the cone cast pixel cube in the world space coordinates (see fig. 2). Assuming the pixel is a square of size \\( c \\) at distance 1 of the camera (\\( c \\) depends on the image resolution and camera's FOV), the volume is thus \\( V = (dc)^3 \\). To take into account the space contraction, we then proceed to contract the volume. Denoting \\( J \\) the Jacobian of the contraction function at the sample's location \\( p \\), \\( V_{contract} = V\\det(J(p)) \\). In practice, we compute the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.128 + ], + "angle": 0, + "content": "D. Petit et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "analytical derivation of the contraction function depending on \\( p \\). More details can be found in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.284 + ], + "angle": 0, + "content": "Assuming the \\(N\\)-th feature grid's resolution in our hierarchy can be written as \\(f^{N}b\\) with \\(b\\) the resolution of the grid \\(\\delta_0\\) and \\(f\\) the growth factor, finding the appropriate LOD \\(L \\in \\mathbb{R}^+\\) means finding a virtual grid4 of resolution \\(f^{L}b\\) whose cell's volume is equal to the previously computed volume (in contracted space). Because we are working in the contracted space of size 1, the volume of a cell of the virtual grid of LOD \\(L\\) is \\((\\frac{1}{f^{L}b})^3\\). The LOD \\(L\\) associated to one sample in the contracted space is thus given by:" + }, + { + "type": "equation", + "bbox": [ + 0.285, + 0.295, + 0.786, + 0.337 + ], + "angle": 0, + "content": "\\[\n\\left(d c\\right) ^ {3} \\det (J (p)) = \\left(\\frac {1}{f ^ {L} b}\\right) ^ {3} \\Longleftrightarrow L = - \\frac {\\log \\left(d c b \\sqrt [ 3 ]{\\det (J (p))}\\right)}{\\log (f)} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.346, + 0.784, + 0.406 + ], + "angle": 0, + "content": "Note that this process is close to ZipNeRF [4] and VR-NeRF [24]. However, the first one rather derives a contracted scale factor from its Gaussian samples while the latter directly computes the LOD in the contracted space, which can be considered an approximation of our computation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.407, + 0.786, + 0.452 + ], + "angle": 0, + "content": "Forward mapping. As illustrated in Figure 2, we use the determined LOD to compute a distance-dependent weighted sum of the features, which is fed to our decoder and transformed into density (or SDF) and radiance." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.474, + 0.727, + 0.49 + ], + "angle": 0, + "content": "3.4 Continuous Coarse-To-Fine and Resolution Extensibility" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.499, + 0.788, + 0.62 + ], + "angle": 0, + "content": "Recent works proposed to use coarse-to-fine optimization to improve the stability of NeRF models, especially when facing more challenging setups, including with few images [25] and surface-based models [11]. It consists of optimizing progressively the different LOD of the representation, from coarse levels to the most precise ones. The goal of this progressive optimization is to avoid the shape-radiance ambiguity [28] by introducing a strong regularization through LOD restriction, then relaxing progressively this regularization once the coarse geometry of the scene is reconstructed to recover the details of the scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.62, + 0.788, + 0.757 + ], + "angle": 0, + "content": "The coarse-to-fine reconstruction process of RING-NeRF consists of estimating progressively the LOD of the mapping function from the coarsest to the finest ones. In practice, it implies, during the cone casting, to clamp the samples' LOD up to a maximal LOD \\( l \\), the grids of level \\( l \\) and above being set to zero and not optimized. Moreover, since our architecture provides continuous LOD, the coarse-to-fine optimization can be achieved continuously in the LOD space by using a linear scheduler \\( l = (l_0 + \\frac{n}{n_{ctf}}) \\in \\mathbb{R}^+ \\) with \\( n \\) the current epoch, \\( n_{ctf} \\) a hyperparameter describing the speed of the process and \\( l_0 \\) defining the number of used grids at initialization; up to a specified maximum resolution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.758, + 0.788, + 0.803 + ], + "angle": 0, + "content": "Furthermore, the RING-NeRF architecture is more adapted than I-NGP-based architectures [11] for coarse-to-fine training. Indeed, for solutions based on the concatenation of features, keeping grid values to zero implies that some" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "4 A virtual grid of LOD \\( L \\in \\mathbb{R}^{+} \\) corresponds to a grid of resolution \\( f^{L}b \\) that is not explicitly stored in memory but whose elements can be computed from other grids." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.645, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RING-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.329 + ], + "angle": 0, + "content": "dimensions of the decoder's latent space are not supervised. When a grid starts to be optimized, those unsupervised dimensions are suddenly used. The weights of the decoder thus need to be refined, with a global effect on the whole scene. On the opposite, our solution keeps supervising all the dimensions of the decoder latent space, and when a new grid gets optimized, it only implies more degrees of freedom to define the mapping function between the scene space and the decoder latent space. This also means that our scene reconstruction can be refined by adding dynamically new grid levels without modifying the decoder's weights or previously trained grids, as we demonstrate experimentally in section 4.5. This resolution extensibility property opens the path to adaptive resolution models, where the precision used to describe an area depends on the details needed, to optimize efficiency both in memory consumption and training duration." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.354, + 0.377, + 0.373 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.788, + 0.51 + ], + "angle": 0, + "content": "In these experiments, we intend to highlight the versatility of RING-NeRF by evaluating it on several tasks. After introducing implementation details in section 4.1, we evaluate our model on novel view synthesis with changes in observation distances (sec. 4.2). Then, we explore how robust is RING-NeRF first with few view reconstruction (sec. 4.3) and then without scene-specific initialization for SDF reconstruction (sec. 4.4). Finally, we demonstrate the capacity of our architecture to perform LOD extensibility, as a first step towards adaptive reconstruction (sec. 4.5)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.534, + 0.395, + 0.55 + ], + "angle": 0, + "content": "4.1 Implementation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.561, + 0.788, + 0.728 + ], + "angle": 0, + "content": "Our model is based on the PyTorch framework Nerfstudio [20]. We build upon its core method named Nerfacto, which combines ideas from several papers for fast and qualitative renders of unbounded complex scenes. This makes it an accessible baseline with a state-of-the-art quality/time ratio. Because NeRF pipelines contain a high number of small but decisive choices of implementation (eg. some frameworks choose to train their models image by image while Nerfstudio jointly and randomly samples across all images), we decided to use as much as possible Nerfstudio-based baselines for fairer comparisons. All of these models are trained on one Nvidia-A100 GPU. The reported times correspond to the approximated training duration of the models. Configuration details, further experiments and ablatives on our contributions can be found in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.752, + 0.593, + 0.768 + ], + "angle": 0, + "content": "4.2 Novel View Synthesis and Anti-Aliasing" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "This experiment aims to evaluate the reconstruction quality through the ability to synthesize viewpoints that are not supervised during the reconstruction. These new viewpoints differ from the angle of observation, but also from the distance of observation. The latter is particularly important since a reconstruction or a" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "D. Petit et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Novel View Synthesis performances for the Mono-Scale setup (trained on the full resolution images only) on the 360 Dataset. The indicated resolutions refer to the resolution of the renders." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.199, + 0.785, + 0.259 + ], + "angle": 0, + "content": "
Full Res.1/2 Res.1/4 Res.1/8 Res.Time ↓
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Nerfacto [20]27.090.7790.18126.810.7820.16225.220.7110.23423.320.6360.2970.45h
PyNeRF [21]27.870.8020.16021.930.7130.21121.150.6620.26420.390.6190.3120.96h
Zip-NeRF [4]28.060.8080.15416.580.5960.31911.880.4240.4659.660.3230.5231.10h
RING-NeRF28.090.7990.15727.180.8040.13825.820.7860.14224.380.7370.1670.45h
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.277, + 0.784, + 0.337 + ], + "angle": 0, + "content": "rendering process that does not take correctly into account the distance of observation leads to artefacts, from over-contrasted rendering to aliasing phenomenon. The challenge is to avoid these artefacts while keeping the reconstruction and rendering process as fast as possible." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.785, + 0.442 + ], + "angle": 0, + "content": "Dataset. The evaluation relies on the dataset introduced by Mip-NeRF-360 [3]. This dataset is composed of 9 scenes, each containing both a central area and complex background in both inside and outside setups. Since the trajectory keeps a constant distance to the central part, each viewpoint is represented with a pyramid of 4 different image resolutions to simulate a variation of the distance of observation combined with a change of the camera FOV, following the Mip-NeRF [2] original anti-aliasing evaluation pipeline." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.785, + 0.488 + ], + "angle": 0, + "content": "Algorithms. We compare our solution against several grid-based NeRF baselines, both with (PyNeRF [21] and Zip-NeRF [4]) and without anti-aliasing processing (Nerfacto), using their Nerfstudio implementations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.785, + 0.563 + ], + "angle": 0, + "content": "Protocol. For each scene, we train the models with two different setups: the mono-scale setup that uses only the image with the highest resolution for each viewpoint, and the multi-scale setup which uses the whole pyramid of images for each viewpoint. Note that, for these two setups, we evaluate the performances on the whole resolution pyramid, with usual metrics (PSNR, SSIM, LPIPS)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.564, + 0.787, + 0.685 + ], + "angle": 0, + "content": "Results. First of all, regarding novel view synthesis quality using the monoscale setup for both training and testing, as referred to in the \"Full Res.\" column of table 1, we notice that our architecture provides on par performances with Zip-NeRF, slightly better results than PyNeRF, and a more important gap with Nerfacto. While simple, RING-NeRF succeeds in performing state-of-the-art performances on a real single-scale dataset. Regarding quality for the multi-scale setup, as presented in table 2, we observe that all the algorithms that consider the distance of observation perform very similarly in terms of quality. On the" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.703, + 0.33, + 0.774 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.703, + 0.441, + 0.775 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.703, + 0.556, + 0.775 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.703, + 0.668, + 0.775 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.703, + 0.783, + 0.775 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.789, + 0.787, + 0.831 + ], + "angle": 0, + "content": "Fig. 4: Examples of image renderings at 1/8th resolution from models solely trained with the full resolution images. RING-NeRF is the only method capable of producing coherent aliasing-free renderings thanks to its LOD inductive bias." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.645, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RING-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.784, + 0.186 + ], + "angle": 0, + "content": "Table 2: Novel View Synthesis performances for the Multi-Scale setup (trained jointly on every resolution) on the 360 Dataset. The indicated resolutions refer to the resolution of the renders." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.199, + 0.784, + 0.259 + ], + "angle": 0, + "content": "
Full Res.1/2 Res.1/4 Res.1/8 Res.Time ↓
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Nerfacto [20]25.980.6680.36727.310.7920.17627.160.8050.13925.300.7430.1990.45h
PyNeRF [21]27.650.7810.19429.470.8570.09230.510.8870.06330.860.8930.0550.96h
Zip-NeRF [4]27.540.7810.19129.340.8580.09930.360.8890.06430.900.9000.0471.10h
RING-NeRF27.660.7820.19429.550.8560.09430.510.8880.06230.860.9010.0480.45h
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.276, + 0.784, + 0.304 + ], + "angle": 0, + "content": "other side, Nerfacto performs quite poorly. Its performances are especially low at the coarsest resolution, with overflowing artefacts, as illustrated in figure 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.306, + 0.785, + 0.547 + ], + "angle": 0, + "content": "Finally, we evaluate the capacity of the models to generalize over new resolutions by training them on full-resolution images and then evaluate them on smaller resolutions. Qualitative results are shown in figure 4 while quantitative results can be found in the 1/2-th, 1/4-th and 1/8-th res. columns of table 1. Because most anti-aliasing methods need LOD-specific supervision to correctly function (including Zip-NeRF and PyNeRF), they cannot render coherent anti-aliased images in this setup. However, PyNeRF behaves better than ZipNeRF with coherent although very aliased renderings, as it decides to limit the LOD used for rendering inside the range of LOD seen during training. RING-NeRF, with his LOD inductive bias, is the only architecture capable of producing anti-aliased renderings from novel observation distances and thus outperforms every other method. While this experiment can seem somewhat esoteric as training on a multi-resolution images pyramid is rather easy, this increases GPU memory and total training time. Moreover, depending on the resolution and the trajectory, it is not trivial to choose the accurate number of scales in the image pyramid to supervise correctly every grid in the hierarchy and especially the coarsest grids." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.785, + 0.638 + ], + "angle": 0, + "content": "Regarding the reconstruction processing time reported in the \"Time\" column of both table 1 and table 2, since RING-NeRF does not rely on the multiplication of either sample or decoder, it processes as quickly as the fastest Ner-facto both on mono-scale and multi-scale setups. On the other hand, the other anti-aliasing methods, PyNeRF with its per-LOD MLP and Zip-NeRF with its super-sampling, are approximately 2.5 times slower." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.785, + 0.714 + ], + "angle": 0, + "content": "In conclusion, our solution provides the best quality-speed trade-off since it is both on par with the best quality method and the fastest method, in monoscale as well as in multi-scale setups. Furthermore, RING-NeRF is the only solution capable of creating coherent and anti-aliased renderings when facing novel observation distances unseen during training." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.738, + 0.494, + 0.753 + ], + "angle": 0, + "content": "4.3 Few Viewpoints Supervision" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.784, + 0.794 + ], + "angle": 0, + "content": "This experiment aims to evaluate the influence of the RING-NERF architecture and pipeline on the reconstruction robustness to limited supervision viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.784, + 0.824 + ], + "angle": 0, + "content": "Dataset. We evaluate our contribution on the object-centric real dataset DTU [9], often used in few-viewpoints evaluations." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.825, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Algorithms. We compare our architecture against several baselines: Mip-NeRF" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "D. Petit et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.298 + ], + "angle": 0, + "content": "[2], and FreeNerf [25] (a state-of-the-art method for this task), for vanilla architectures and Nerfacto for grid-based architectures. To better demonstrate the intrinsic stability brought by RING-NeRF, we also developed the Nerfacto+ architecture, which corresponds to a Nerfacto architecture coupled with a coarse-to-fine training based on a progressive activation of the grids (the rest of the decoder's input being filled with zeros). For Nerfacto+ and RING-NeRF, we also add FreeNeRF's loss that penalizes the density of the first \\( M = 10 \\) samples of each ray to reduce as much as possible artefacts in front of the cameras. As an ablative experiment, we also evaluate RING-NeRF using discrete coarse-to-fine (fixed LOD increment of 1 rather than the proposed continuous increase)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.788, + 0.359 + ], + "angle": 0, + "content": "Protocol. We follow FreeNeRF's evaluation pipeline, including the number of supervision viewpoints (3 to 9), the choice of these views among the dataset and the evaluations using masks of the object. Since we are using the same protocol, the results of FreeNeRF and Mip-NeRF were taken out of FreeNeRF's article." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.678 + ], + "angle": 0, + "content": "Results. Evaluation results are reported in table 3. We first notice an important difference between vanilla and grid-based baselines. While Mip-NeRF faces troubles in reconstructing the scene with 3 views, the method seems to find coherency when adding more images. However, Nerfacto struggles much more to form a consistent 3D scene even when using 9 images (see figure 1). Even though the grid-based baseline is way faster to train, its design implies more instability when facing a small number of images. This does not mean however that few viewpoints are incompatible with grid-based methods. Using progressive training coupled with a very simple and generalizable regularization, both Nerfacto+ and our architecture succeed in creating coherent geometry. Nonetheless, RING-NeRF considerably outperforms Nerfacto+, with a PSNR difference varying from 3 to 4, demonstrating the stability increase brought by our architecture, and also outperforms Mip-NeRF for the configuration with 3 and 6 supervision images. The discrete coarse-to-fine version of RING-NeRF performs in-between Nerfacto+ and the complete RING-NeRF. This showcases both the intrinsic interest of the proposed architecture against the Nerfacto+ and the relevance of the continuous coarse-to-fine mechanism. FreeNerf remains the best performer of all in terms of quality, but with a reconstruction time that is extremely slower than RING-NeRF, the former achieving its reconstruction in 2.56 hours while the latter only requires less than 10 minutes. RING-NeRF thus offers a better quality-speed trade-off for the few-view reconstruction issue (see figure 1)." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.691, + 0.784, + 0.721 + ], + "angle": 0, + "content": "Table 3: Performances of reconstruction from few viewpoints on the DTU dataset. The reported metrics are computed based on the mask of the object." + }, + { + "type": "table", + "bbox": [ + 0.245, + 0.732, + 0.763, + 0.84 + ], + "angle": 0, + "content": "
#ImagesPSNR ↑SSIM ↑LPIPS ↓Time ↓
369369369
Mip-NeRF [2]8.6816.5423.580.5710.7410.8790.3530.1980.0922.56h
FreeNeRF [25]19.9223.2525.380.7870.8440.8880.1350.0950.0672.56h
Nerfacto [20]9.359.759.780.5670.6040.6470.3850.3310.3260.15h
Nerfacto+13.6116.6119.330.6390.6990.7590.2760.2180.1510.15h
RING-NeRF16.1820.4723.190.7130.8080.8470.2000.1270.0850.15h
w/ discrete CtF15.7920.1622.930.7060.7850.8470.2010.1270.0850.15h
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.645, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RING-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 4: SDF reconstruction performances when foregoing the scene-specific SDF Initialization on the Replica Dataset. The Chamfer distance is in centimeters." + }, + { + "type": "table", + "bbox": [ + 0.302, + 0.185, + 0.703, + 0.233 + ], + "angle": 0, + "content": "
PSNR ↑SSIM ↑LPIPS ↓Chamfer-L1 ↓Training Time ↓
NeuS-facto [26]24.620.7780.34717.611.4h
NeuralAngelo [11]30.790.9160.076113.693.40 h
RING-NeRF37.180.9690.01945.711.28 h
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.251, + 0.611, + 0.266 + ], + "angle": 0, + "content": "4.4 SDF Reconstruction without Initialization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.272, + 0.785, + 0.362 + ], + "angle": 0, + "content": "SDF reconstruction is known to be a more unstable process than density-based NeRF [1], requiring a scene-specific initialization to converge. This initialization becomes an issue in complex environments and incremental setups, where several types of scenes can co-exist and are not necessarily known beforehand. Therefore, in this experiment, we evaluate the ability of RING-NeRF and other SotA architectures to achieve SDF reconstruction without scene-specific initialization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.363, + 0.784, + 0.408 + ], + "angle": 0, + "content": "Dataset. The evaluation is achieved on a subset of 7 scenes of the Replica [18] synthetic indoor dataset. A Tanks & Temples [10] example is provided in supplementary materials with corresponding analysis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.408, + 0.785, + 0.498 + ], + "angle": 0, + "content": "Algorithms. We compare our architecture to two SDF methods, all of them implemented in the same SDFStudio [26] branch of the Nerfstudio framework for fairer comparisons: NeuS-facto, an adaptation of NeuS for grid-based methods with Nerfacto modules, and an implementation of NeuralAngelo. For these experiments, our model is built upon the NeuS-facto baseline, using in particular the same NeuS-based SDF-to-density transformation [22]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.499, + 0.785, + 0.543 + ], + "angle": 0, + "content": "Protocol. To evaluate the impact of the architecture and pipeline over the convergence and stability, we suppress the inverted sphere SDF initialization scheme and use a random initialization for the model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.544, + 0.786, + 0.695 + ], + "angle": 0, + "content": "Results. Evaluation results are shown in table 4. First of all, NeuS-facto faces low rendering and reconstruction metrics, due to catastrophic failure in most of the tested scenes (see figure 1) since the model tends to re-draw the 2D images in front of the camera. Regarding NeuralAngelo, its relatively high PSNR demonstrates its ability to synthesize satisfying RGB. However, as illustrated in figure 1 and highlighted by the reconstruction metrics, the underlying geometry of the scene is poorly reconstructed, without any fine details. Finally, our method RING-NeRF is by far the best performer, with much higher PSNR and a better geometry including fine details (see figure 1), although a bit noisy. The simple architecture of RING-NeRF also permits faster epochs, thus faster training." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.713, + 0.414, + 0.729 + ], + "angle": 0, + "content": "4.5 LOD Extensibility" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.784, + 0.764 + ], + "angle": 0, + "content": "This experiment aims to demonstrate RING-NeRF's unique ability to increase dynamically the level of detail of the scene representation." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.765, + 0.733, + 0.779 + ], + "angle": 0, + "content": "Dataset. The scan 114 of the DTU dataset is used for this experiment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.784, + 0.81 + ], + "angle": 0, + "content": "Algorithms. Because I-NGP architectures [4, 11, 14] cannot perform LOD extensibility (due to the fixed decoder's input size), only RING-NeRF is evaluated." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Protocol. We train our model using two different configurations : one low resolution with a 3 levels grid hierarchy and one high resolution with 5 levels (the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "D. Petit et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.147, + 0.526, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.147, + 0.648, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.147, + 0.774, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.219, + 0.649, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.219, + 0.774, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.303, + 0.787, + 0.334 + ], + "angle": 0, + "content": "Fig. 5: Learning curves and final renderings of RING-NeRF models with different grid configurations trained either jointly or incrementally." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.351, + 0.788, + 0.457 + ], + "angle": 0, + "content": "grid resolutions are 16, 32, 64, 128 and 256). For these two configurations named \"3 levels\" and \"5 levels\", every grid and the decoder are trained simultaneously. We proceed to showcase the extensibility of RING-NeRF by adding grids to the \"3 levels\" configuration that is previously trained. We first train one grid of resolutions 128 (\"3+1 levels\" in Figure 5) with the three initial grids and the decoder frozen and then train another grid of resolutions 256 (\"3+1+1 levels\" in Figures 5) with the four grids and the decoder frozen." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.457, + 0.788, + 0.563 + ], + "angle": 0, + "content": "Results. Figure 5 shows that the configuration \"3+1+1 levels\" results in the same rendering quality than the \"5 levels\" one. This demonstrates the ability of our model to dynamically change the resolution of the grid hierarchy. This is an important step towards the development of an adaptive architecture which locally chooses the resolution of the representation based on the scene's content. This allows to drastically reduce the number of parameters, helping in improving the memory footprint, the training duration and the model's robustness." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.598, + 0.53, + 0.615 + ], + "angle": 0, + "content": "5 Conclusion and Perspectives" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.788, + 0.779 + ], + "angle": 0, + "content": "In this work, we introduced RING-NeRF, a simple and versatile NeRF pipeline that provides two inductive biases by design: a continuous multi-scale representation of the scene, and an invariance of the decoder latent space over spatial and scale domains. Coupled with a distance-aware forward mapping and a continuous coarse-to-fine reconstruction process, our pipeline demonstrated experimentally its versatility with on-par performances with dedicated state-of-the-art solutions for anti-aliasing or reconstruction from few viewpoints. It even outperforms them in terms of robustness to scene-specific initialization for SDF reconstruction. Furthermore, it is highly efficient and is not limited to object-centric scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Future work will study the impact of RING-NeRF on other challenging use cases, such as facing inaccurate camera poses [16] and SLAM [29]. We will also use the extensibility property of our architecture to develop memory-efficient sparse Neural Fields, which is considered to be a limit of most grid-based models." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.645, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RING-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.177, + 0.784, + 0.207 + ], + "angle": 0, + "content": "This publication was made possible by the use of the CEA List FactoryIA supercomputer, financially supported by the Ile-de-France Regional Council." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.23, + 0.323, + 0.246 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.26, + 0.785, + 0.302 + ], + "angle": 0, + "content": "1. Atzmon, M., Lipman, Y.: Sal: Sign agnostic learning of shapes from raw data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2565-2574 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.302, + 0.785, + 0.357 + ], + "angle": 0, + "content": "2. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.359, + 0.785, + 0.412 + ], + "angle": 0, + "content": "3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mipnerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.413, + 0.785, + 0.44 + ], + "angle": 0, + "content": "4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-nerf: Anti-aliased grid-based neural radiance fields. ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.441, + 0.785, + 0.468 + ], + "angle": 0, + "content": "5. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: European Conference on Computer Vision (ECCV) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.469, + 0.785, + 0.509 + ], + "angle": 0, + "content": "6. Dou, Y., Zheng, Z., Jin, Q., Ni, B.: Multiplicative fourier level of detail. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1808-1817 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.51, + 0.785, + 0.564 + ], + "angle": 0, + "content": "7. Hu, W., Wang, Y., Ma, L., Yang, B., Gao, L., Liu, X., Ma, Y.: Tri-miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19774-19783 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.606 + ], + "angle": 0, + "content": "8. Jain, A., Tancik, M., Abbeel, P.: Putting nerf on a diet: Semantically consistent few-shot view synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5885-5894 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.607, + 0.785, + 0.647 + ], + "angle": 0, + "content": "9. Jensen, R., Dahl, A., Vogiatzis, G., Tola, E., Aanaes, H.: Large scale multi-view stereopsis evaluation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 406-413 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.648, + 0.785, + 0.674 + ], + "angle": 0, + "content": "0. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics 36(4) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.675, + 0.785, + 0.73 + ], + "angle": 0, + "content": "11. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8456-8465 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.731, + 0.785, + 0.757 + ], + "angle": 0, + "content": "12. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "13. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "14. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.26, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "D. Petit et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "15. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5480-5490 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.787, + 0.244 + ], + "angle": 0, + "content": "16. Park, K., Henzler, P., Mildenhall, B., Barron, J.T., Martin-Brualla, R.: Camp: Camera preconditioning for neural radiance fields. ACM Transactions on Graphics (TOG) 42(6), 1-11 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.244, + 0.787, + 0.271 + ], + "angle": 0, + "content": "17. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.787, + 0.312 + ], + "angle": 0, + "content": "18. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.312, + 0.787, + 0.367 + ], + "angle": 0, + "content": "19. Takikawa, T., Litalien, J., Yin, K., Kreis, K., Loop, C., Nowrouzezahrai, D., Jacobson, A., McGuire, M., Fidler, S.: Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11358-11367 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.367, + 0.787, + 0.422 + ], + "angle": 0, + "content": "20. Tancik, M., Weber, E., Ng, E., Li, R., Yi, B., Wang, T., Kristoffersen, A., Austin, J., Salahi, K., Ahuja, A., et al.: Nerfstudio: A modular framework for neural radiance field development. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-12 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.422, + 0.787, + 0.462 + ], + "angle": 0, + "content": "21. Turki, H., Zollhöfer, M., Richardt, C., Ramanan, D.: Pynerf: Pyramidal neural radiance fields. In: Thirty-Seventh Conference on Neural Information Processing Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.462, + 0.787, + 0.503 + ], + "angle": 0, + "content": "22. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.503, + 0.787, + 0.53 + ], + "angle": 0, + "content": "23. Wynn, J., Turmukhambetov, D.: DiffusioNeRF: Regularizing Neural Radiance Fields with Denoising Diffusion Models. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.787, + 0.585 + ], + "angle": 0, + "content": "24. Xu, L., Agrawal, V., Laney, W., Garcia, T., Bansal, A., Kim, C., Rota Bulò, S., Porzi, L., Kontschieder, P., Božić, A., Lin, D., Zollhöfer, M., Richardt, C.: VR-NeRF: High-fidelity virtualized walkable spaces. In: SIGGRAPH Asia Conference Proceedings (2023). https://doi.org/10.1145/3610548.3618139" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.585, + 0.787, + 0.626 + ], + "angle": 0, + "content": "25. Yang, J., Pavone, M., Wang, Y.: Freenerf: Improving few-shot neural rendering with free frequency regularization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8254-8263 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.626, + 0.787, + 0.667 + ], + "angle": 0, + "content": "26. Yu, Z., Chen, A., Antic, B., Peng, S., Bhattacharyya, A., Niemeyer, M., Tang, S., Sattler, T., Geiger, A.: Sdfstudio: A unified framework for surface reconstruction (2022), https://github.com/autonomousvision/sdfstudio" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.667, + 0.787, + 0.708 + ], + "angle": 0, + "content": "27. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems 35, 25018-25032 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.708, + 0.787, + 0.735 + ], + "angle": 0, + "content": "28. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.735, + 0.787, + 0.79 + ], + "angle": 0, + "content": "29. Zhu, Z., Peng, S., Larsson, V., Xu, W., Bao, H., Cui, Z., Oswald, M.R., Pollefeys, M.: Nice-slam: Neural implicit scalable encoding for slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12786-12796 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.79, + 0.787, + 0.831 + ], + "angle": 0, + "content": "30. Zhuang, Y., Zhang, Q., Feng, Y., Zhu, H., Yao, Y., Li, X., Cao, Y.P., Shan, Y., Cao, X.: Anti-aliased neural implicit surfaces with encoding level of detail. Siggraph Asia 2023 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.831 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/RING-NeRF _ Rethinking Inductive Biases for Versatile and Efficient Neural Fields/5e093f34-06af-49a4-a5fe-2e79d5819847_origin.pdf b/2024/RING-NeRF _ Rethinking Inductive Biases for Versatile and Efficient Neural Fields/5e093f34-06af-49a4-a5fe-2e79d5819847_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d2212b3880092c0cb84646c21b94b633d51d41db --- /dev/null +++ b/2024/RING-NeRF _ Rethinking Inductive Biases for Versatile and Efficient Neural Fields/5e093f34-06af-49a4-a5fe-2e79d5819847_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3499ba0f68322bb1a4a815d2cd2dd971adefa13ad7b25778af01067a99c5ea30 +size 2256595 diff --git a/2024/RING-NeRF _ Rethinking Inductive Biases for Versatile and Efficient Neural Fields/full.md b/2024/RING-NeRF _ Rethinking Inductive Biases for Versatile and Efficient Neural Fields/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a3bb78a59d1b76e56205cef716c2707a9b99ee6e --- /dev/null +++ b/2024/RING-NeRF _ Rethinking Inductive Biases for Versatile and Efficient Neural Fields/full.md @@ -0,0 +1,277 @@ +# RING-NeRF : Rethinking Inductive Biases for Versatile and Efficient Neural Fields + +Doriand Petit $^{1,2}$ , Steve Bourgeois $^{1}$ , Dumitru Pavel $^{1}$ , Vincent Gay-Bellile $^{1}$ , Florian Chabot $^{1}$ , and Loic Barthe $^{2}$ + +1 Université Paris-Saclay, CEA, List, F-91120, Palaiseau, France +2 IRIT, Université Toulouse III, CNRS, France + +![](images/72ab70ec195e7c177d49d3e764af15c3e127b41b5cb7f1c9512dfe05a18f63fb.jpg) +Fig. 1: RING-NeRF is a simple and versatile architecture which tackles many NeRF common issues such as robustness to distance of observation, few view supervision and lack of scene-specific initialization for SDF-based reconstruction. It provides on-par performances in terms of quality with SotA dedicated solutions [4,11,25] and and in terms of efficiency with fast methods [20,26]. + +Abstract. Recent advances in Neural Fields mostly rely on developing task-specific supervision which often complicates the models. Rather than developing hard-to-combine and specific modules, another approach generally overlooked is to directly inject generic priors on the scene representation (also called inductive biases) into the NeRF architecture. Based on this idea, we propose the RING-NeRF architecture which includes two inductive biases: a continuous multi-scale representation of the scene and an invariance of the decoder's latent space over spatial and scale domains. We also design a single reconstruction process that takes advantage of those inductive biases and experimentally demonstrates on-par performances in terms of quality with dedicated architecture on multiple tasks (anti-aliasing, few view reconstruction, SDF reconstruction without scene-specific initialization) while being more efficient. Moreover, RING-NeRF has the distinctive ability to dynamically increase the resolution of the model, opening the way to adaptive reconstruction. Project page can be found at: https://cea-list.github.io/RING-NeRF + +# 1 Introduction + +Neural Radiance Fields (NeRF) have emerged as a novel method for representing 3D scenes using neural networks. In its original design [13], a simple multi-layer perceptron (MLP) is trained to reproduce a continuous 5D function that outputs the density and radiance emitted in each direction $(\theta, \phi)$ at each point $(x, y, z)$ in space. This approach inspired many works due to the impressive quality of its novel view synthesis as well as its simplicity. However, the so-called "vanilla" NeRF architectures converge very slowly as they use large deep neural networks. Instant-NGP [14] introduced a new architecture based on a hierarchical hash-grid pyramid of learnable feature codes describing the 3D scene and combined with a much shallower MLP, called decoder, that transforms the concatenation of the codes, interpolated from the grids, into density and radiance. Resulting in local updates, this method reduced the training duration from hours to minutes. + +The great majority of current solutions are now based on these two standard architectures, and many of them are focused on overtaking their associated limitations in terms of: + +Nature of scene - by transitioning from object-centric scenes to open unbounded scenes, using mostly space contraction [3]; + +Robustness - by managing free motion trajectories with variations of the observation distance while avoiding aliasing artefacts, mostly through the integration of Level of Detail (LOD) in the model [2,4]; or by reducing drastically the number of supervised views through different kinds of regularization [8,15,25]; + +Extensibility - by shifting from a holistic and fixed reconstruction process to an incremental (extensibility in the number of views) and adaptive (extensibility in resolution) process, mainly through the use of a frozen decoder [12, 29]. + +However, most solutions solely focus on one of these limitations and introduce specific and complex mechanisms that both increase the computational cost and lessen the possibilities of combination. The ability to solve jointly these main issues is however essential in real-world applications which often require a robust and extensible reconstruction in an unknown and unbounded environment. + +In this article, rather than introducing yet another heavy and task-specific solution, we propose RING-NeRF, a versatile and simple NeRF architecture by rethinking usual grid-based models to introduce two inductive biases. We first represent a 3D scene as a continuous multi-scale representation and also make the decoder's latent space invariant in position and scale. Together, these two priors enable the production of intrinsic continuous LOD of the scene without explicit supervision. We demonstrate experimentally that, when combined with adapted cone casting and coarse-to-fine optimization, the resulting architecture is able to compete on several tasks with on-par quality performances with dedicated state-of-the-art solutions while improving speed, robustness and extensibility. The overall process is also simple, easy to implement and generic enough to be coupled with specific solutions. Our contributions can be summarized as follows: + +1. an architecture that, by construction, represents the scene with a continuous and unbounded level of detail without the need for LOD-specific supervision and which permits resolution-adaptive reconstruction; + +2. a distance-aware forward mapping compatible with scene contraction, that takes benefits of the continuous multi-scale representation with an adapted cone casting process to avoid aliasing artefacts when varying the observation distance. + +3. a continuous coarse-to-fine reconstruction process that improves the convergence and stability (especially in challenging setups such as supervision with few viewpoints or no scene-specific initialization for SDF reconstruction). + +# 2 Related Work + +From its original iteration [13], a majority of current research focuses on overtaking limitations of NeRF-based reconstruction in terms of adaptability to various natures of scene, robustness (to varying observation distances or limited amount of viewpoints) and reconstruction extensibility. + +Adaptability to various natures of the scene. The ability of Neural Fields to reconstruct various natures of scene depends on three factors. The first one is related to its architecture itself. Tri-plane architectures [5, 7, 30] are mostly designed for object-centric reconstruction (as they provide a higher density of information in the center of the scene) while vanilla and 3D grid-based NeRF are able to cover a wider variety of scenes, though they initially were still constrained to a limited volume. A second factor is related to the representation of the 3D space of the scene, especially to represent distant elements in open scenes. Some approaches use two different NeRF models to reconstruct separately the foreground and the background [28], whereas others apply space contraction to the 3D scene coordinates [3, 4, 28] to map the infinite scene volume into a bounded one. The last factor is related to the initialization of the Neural Fields. While the random initialization of density-based NeRF can adjust to almost any nature of scenes, the convergence of SDF-based (Signed Distance Function) Neural Fields is extremely sensitive to their parameters' initialization, as stated in [1]. Current solutions rely on a scene-specific initialization (using an SDF field representing a sphere for outdoor scenes or an inverted sphere for indoor scenes), making them unable to adapt automatically to any scene. + +Robustness to observation distance variations. The initial NeRF model, as well as most of the subsequent works, relies implicitly on the hypothesis of a constant distance of observation to the scene. Indeed, the NeRF model provides a per-3D-point scene density representation and a rendering process which does not take into account the increasing volume covered by a pixel with respect to the distance to the cameras. As underlined in Mip-NeRF [2], this discrepancy induces artefacts such as over-contrasted images or aliasing phenomenon when the distance of observation differs from the ones used at reconstruction time. To avoid these artefacts, the rendering process needs to assess the density and color for a volume instead of a point. In the current state-of-the-art, two main + +approaches are used (or their combination, as in Zip-NeRF [4]). The first solution consists of representing the scene with different levels of detail (LOD), in order to vary the precision of the reconstruction based on the observation distance. This is usually done by using a LOD-aware latent space [2,4,6,19,21,24], meaning that the LOD information is already encoded in the inputs of the MLP. This can be achieved by using a per-LOD decoder as in PyNeRF [21] or by incorporating LOD information in the latent feature as done in Zip-NeRF [4] and VR-NeRF [24]. One main flaw of these solutions is that they require the supervision of every used LOD which makes it impossible to vary the observation distances between the train views and novel synthesised views. A second approach consists of defining the latent representation of a volume as the mean of the latent features of the points included in the volume. It requires integrating the features over the volume, which can be achieved through convolutions for tri-plan representation [7,30], or through super-sampling of the latent space for 3D grids as also done in Zip-NeRF [4]. However, these approaches lengthen the training and rendering processes as they increase the number of computations. + +Robustness to limited amount of viewpoints. By construction, NeRF is subject to the shape-radiance ambiguity [28]. If not enough supervision viewpoints are available, the optimization might overfit them while not providing consistent 3D reconstruction nor generalization to non-supervised viewpoints. A first family of solutions to overcome this limitation consists of regularizing the reconstruction process through additional losses, whether via geometric [15,23,27] or semantic [8] regularization. The second family of solutions relies on a progressive reconstruction of the details of the scene, as introduced in FreeNeRF [25] and Nerfies [17]. Restraining the ability of the model to reconstruct a complex scene at early stages enforces the consistency of the reconstruction over the different supervision viewpoints. This presents the advantage of bringing stability while keeping a fairly simple training process. However, these latter solutions mostly rely on Vanilla models for stability purposes and require long training duration. + +Reconstruction extensibility. Two kinds of extensibility should be distinguished: extensibility with respect to the number of views or the resolution. The first one is related to reconstructing new scene areas while keeping the previously reconstructed ones unchanged. The usual solution consists of using grid-based approaches with a pre-trained position-invariant decoder that is frozen during the reconstruction [29]. On the other side, the extensibility of the resolution consists in dynamically increasing the level of detail of a previously reconstructed scene. This problem is intractable for classic architectures such as vanilla NeRF and I-NGP since the number of layers (resp. grids) of those approaches cannot increase during the reconstruction. Some rare solutions, such as Neural Sparse Voxel Fields [12], combine a pre-trained decoder with a data structure allowing to dynamically allocate voxels to increase the reconstruction resolution. This is however a crucial stake, as being able to adapt the precision of the representation based on the complexity of the scene permits to optimize the computational efficiency (both in speed and memory requirements) with minimal loss in quality. + +![](images/d84823d8ac274faac4824caf00c8975b98a84de9c4e9946b64c03e02d8752d6a.jpg) +Fig. 2: Overview of RING-NeRF: to render a pixel, the casted cone is sampled with cubes. Depending on the cube volume, the corresponding LOD of the scene is selected and the latent feature is computed using a weighted sum of the grid hierarchy. The density (or SDF) and color of the cube are first decoded from the latent feature with a tiny MLP and then integrated with other samples through volume rendering. + +# 3 RING-NeRF + +Rather than focusing on solving one specific problem of NeRF using complex mechanisms, we propose a simple architecture called RING-NeRF constructed with novel inductive biases to tackle NeRF's common issues. + +# 3.1 Overview + +RING-NeRF relies on the classic NeRF [13] inverse rendering pipeline which is used to reconstruct a 3D scene from a set of localized frames. For a given image pixel (with its camera's pose), a 3D ray is cast and the 3D scene representation is sampled at $N$ various locations along the ray. The resulting density $\sigma_{i}$ (or SDF converted to density [22]) and color $c_{i}$ of the samples are then combined with usual volume rendering techniques: $\hat{C}(r) = \sum_{i=0}^{N-1} T_{i} (1 - \exp(-\sigma_{i} \delta_{i})) c_{i}$ with $T_{i} = \exp(\sum_{j=0}^{i-1} \sigma_{j} \delta_{j})$ and $\delta_{i}$ is the distance between samples. The parameters of the scene representation are then optimized by minimizing the MSE loss $||\hat{C}(r) - C(r)||^{2}$ between the rendered ray color and the ground truth pixel value, the rendering being differentiable3. + +The originality of RING-NeRF relies on its neural architecture, illustrated in figure 2, which is designed to represent the scene with a continuous and unbounded level of detail without the need for LOD-specific supervision (see section 3.2). We then use this LOD inductive bias to adjust the LOD of the samples with respect to the distance to the camera in contracted space for more accurate renderings (see section 3.3). Finally, combined with a continuous coarse-to-fine optimization, RING-NeRF results in a more robust reconstruction process with an intrinsic LOD extensibility property (see section 3.4). + +# 3.2 NeRF Representation with Inductive Biases + +Classic grid-based representations, such as the one introduced by I-NGP [14], rely on two key elements: a unique MLP decoder which transforms a latent space into an output space (color, density or SDF), and a 3D (hierarchical) grid of latent features that implicitly defines a unique mapping function from the scene space onto the decoder latent space. We can distinguish two different approaches used to extend the mono-scale representation to a multi-scale representation: conditioning with scale information the decoding process [4,21] and/or defining separate per-LOD mapping functions [21]. In both cases, the training becomes more complex, leading to potential convergence issues as it relies solely on additional specific supervision. Instead, we propose to condition the mapping function itself with scale information and introduce inductive biases to guide the convergence. + +Scene representation as multi-scale mapping function. As illustrated in Figure 2, the mapping function at level $N$ is controlled by a grid $G_{N}$ that is implicitly defined through a recursive refinement process over scale: + +$$ +\left\{ \begin{array}{l} G _ {0} \leftarrow \delta_ {0} \\ G _ {N} \leftarrow S \left(G _ {N - 1}\right) + \delta_ {N} \end{array} \right. \tag {1} +$$ + +where $\delta_{i}$ is the deviation defined by the i-th latent feature grid and S is the subdivision scheme consisting in increasing the resolution of $G_{N - 1}$ up to the $N$ -th grid resolution (with tri-linear interpolation). In practice, the latent vector associated with a point in the scene for a level of detail $N$ is simply obtained by interpolating linearly the point in the cell for each level inferior or equal to $N$ , and summing the results. Since linear interpolation is differentiable, our mapping function is also estimizable. The summing of the interpolation is comparable to a residual connection and gives the name of the architecture RING-NeRF for Residual Implicit Neural Grids. + +Such representation provides several advantages. First, since the mapping function is constructed in a top-down manner, refining recursively one level of detail from the previous one, its number of LOD is unbounded and adding a finer level of detail keeps the coarser ones valid. Secondly, because each $\delta_{i}$ only represents a deviation of the coarser mapping value $G_{i - 1}$ , a continuous LOD representation can be easily obtained by incorporating a weighting factor $\alpha \in [0,1]$ in the recursive process: + +$$ +G _ {L} \leftarrow S \left(G _ {N - 1}\right) + \alpha \delta_ {N} \tag {2} +$$ + +with $L\in [N - 1,N]$ and $\alpha = L - (N - 1)$ + +Spatial and scale invariance of decoder. Since our architecture does not rely on a decoder conditioned on position (unlike [19]) nor scale (unlike [2-4, 21]), the decoder latent space is invariant to translations and scale changes in the scene coordinate. This property makes RING-NeRF more suited for incremental reconstruction in both spatial and scale space, since it ensures that local updates in the spatial and scale domain of the scene can be achieved through the + +![](images/2db464ad350a9ef5df99d47be7363b59bd182042defe315f3196b4660b803091.jpg) +Fig. 3: We demonstrate the LOD inductive bias by training our model with a hierarchy of $N = 7$ levels where only the last level of the mapping function $G_{N}$ is supervised. We then compute renders at different levels $L \leq N$ . Other examples (including of entire scenes) can be found in the supplementary materials. + +![](images/a526fb7ed0a150bac77bbb989deef339b0eb5fc05ecd7338a9a0489ffba4b002.jpg) + +![](images/c96e543197cd47c549eaab334d8472b498fd341f24262108eb01c99c2ad1d362.jpg) + +![](images/361e48c86ec61bb06e172fe2e8bfa11ebc51ee8553609c8b4c0b67aa5cf744d4.jpg) + +![](images/a309244325ed4cb5cfe637a343a4fbc7637c7c9bd1366844a7a819d47e257c16.jpg) + +hierarchical grid. The decoder architecture is illustrated in figure 2 and further details can be found in the supplementary materials. + +LOD inductive bias. During the reconstruction process, the gradients of the ray samples are backpropagated through the residual connections and aggregated for each grid level. Due to the pyramidal resolution of the hierarchical grid, a grid code at coarse levels influences a large scene volume and is thus supervised by more ray samples. Therefore, the gradient of a grid code increases as the level's resolution decreases, and as long as the associated samples' gradients are uniform. However, once the backpropagation through the residual connections reach a level whose associated samples have divergent enough gradients (meaning the error is finer than this level's resolution), the result of the aggregation will be mitigated. Hence, the level with the maximum correction is always the coarser level where the samples' gradients are still uniform. This property naturally induces corrections at the proper grid level and LOD. In figure 3, we illustrate this inductive bias by displaying different continuous LOD of a scene while the reconstruction is only supervised for its finest LOD. Not only this is a useful property when an unsupervised multi-scale representation is needed (as demonstrated in section 4.2), this also guarantees a more robust convergence (as illustrated in section 4.3 and section 4.4). + +# 3.3 Distance-Aware Forward Mapping + +Cone Casting in contracted space. In order to accommodate the scene rendering process to the variation of observation distances, we introduce a distance-aware forward mapping mechanism. Similarly to the cone casting of [7], it relies on assigning to each sample a latent feature whose LOD is inversely proportional to the sample-camera distance. However, unlike [7, 21], our model relies on the use of space contraction to allow the reconstruction of unbounded scenes. + +To define the LOD of a sample at distance $d$ of the camera, we first compute the associated volume of the cone cast pixel cube in the world space coordinates (see fig. 2). Assuming the pixel is a square of size $c$ at distance 1 of the camera ( $c$ depends on the image resolution and camera's FOV), the volume is thus $V = (dc)^3$ . To take into account the space contraction, we then proceed to contract the volume. Denoting $J$ the Jacobian of the contraction function at the sample's location $p$ , $V_{contract} = V\det(J(p))$ . In practice, we compute the + +analytical derivation of the contraction function depending on $p$ . More details can be found in the supplementary materials. + +Assuming the $N$ -th feature grid's resolution in our hierarchy can be written as $f^{N}b$ with $b$ the resolution of the grid $\delta_0$ and $f$ the growth factor, finding the appropriate LOD $L \in \mathbb{R}^+$ means finding a virtual grid4 of resolution $f^{L}b$ whose cell's volume is equal to the previously computed volume (in contracted space). Because we are working in the contracted space of size 1, the volume of a cell of the virtual grid of LOD $L$ is $(\frac{1}{f^{L}b})^3$ . The LOD $L$ associated to one sample in the contracted space is thus given by: + +$$ +\left(d c\right) ^ {3} \det (J (p)) = \left(\frac {1}{f ^ {L} b}\right) ^ {3} \Longleftrightarrow L = - \frac {\log \left(d c b \sqrt [ 3 ]{\det (J (p))}\right)}{\log (f)} \tag {3} +$$ + +Note that this process is close to ZipNeRF [4] and VR-NeRF [24]. However, the first one rather derives a contracted scale factor from its Gaussian samples while the latter directly computes the LOD in the contracted space, which can be considered an approximation of our computation. + +Forward mapping. As illustrated in Figure 2, we use the determined LOD to compute a distance-dependent weighted sum of the features, which is fed to our decoder and transformed into density (or SDF) and radiance. + +# 3.4 Continuous Coarse-To-Fine and Resolution Extensibility + +Recent works proposed to use coarse-to-fine optimization to improve the stability of NeRF models, especially when facing more challenging setups, including with few images [25] and surface-based models [11]. It consists of optimizing progressively the different LOD of the representation, from coarse levels to the most precise ones. The goal of this progressive optimization is to avoid the shape-radiance ambiguity [28] by introducing a strong regularization through LOD restriction, then relaxing progressively this regularization once the coarse geometry of the scene is reconstructed to recover the details of the scene. + +The coarse-to-fine reconstruction process of RING-NeRF consists of estimating progressively the LOD of the mapping function from the coarsest to the finest ones. In practice, it implies, during the cone casting, to clamp the samples' LOD up to a maximal LOD $l$ , the grids of level $l$ and above being set to zero and not optimized. Moreover, since our architecture provides continuous LOD, the coarse-to-fine optimization can be achieved continuously in the LOD space by using a linear scheduler $l = (l_0 + \frac{n}{n_{ctf}}) \in \mathbb{R}^+$ with $n$ the current epoch, $n_{ctf}$ a hyperparameter describing the speed of the process and $l_0$ defining the number of used grids at initialization; up to a specified maximum resolution. + +Furthermore, the RING-NeRF architecture is more adapted than I-NGP-based architectures [11] for coarse-to-fine training. Indeed, for solutions based on the concatenation of features, keeping grid values to zero implies that some + +dimensions of the decoder's latent space are not supervised. When a grid starts to be optimized, those unsupervised dimensions are suddenly used. The weights of the decoder thus need to be refined, with a global effect on the whole scene. On the opposite, our solution keeps supervising all the dimensions of the decoder latent space, and when a new grid gets optimized, it only implies more degrees of freedom to define the mapping function between the scene space and the decoder latent space. This also means that our scene reconstruction can be refined by adding dynamically new grid levels without modifying the decoder's weights or previously trained grids, as we demonstrate experimentally in section 4.5. This resolution extensibility property opens the path to adaptive resolution models, where the precision used to describe an area depends on the details needed, to optimize efficiency both in memory consumption and training duration. + +# 4 Experiments + +In these experiments, we intend to highlight the versatility of RING-NeRF by evaluating it on several tasks. After introducing implementation details in section 4.1, we evaluate our model on novel view synthesis with changes in observation distances (sec. 4.2). Then, we explore how robust is RING-NeRF first with few view reconstruction (sec. 4.3) and then without scene-specific initialization for SDF reconstruction (sec. 4.4). Finally, we demonstrate the capacity of our architecture to perform LOD extensibility, as a first step towards adaptive reconstruction (sec. 4.5). + +# 4.1 Implementation + +Our model is based on the PyTorch framework Nerfstudio [20]. We build upon its core method named Nerfacto, which combines ideas from several papers for fast and qualitative renders of unbounded complex scenes. This makes it an accessible baseline with a state-of-the-art quality/time ratio. Because NeRF pipelines contain a high number of small but decisive choices of implementation (eg. some frameworks choose to train their models image by image while Nerfstudio jointly and randomly samples across all images), we decided to use as much as possible Nerfstudio-based baselines for fairer comparisons. All of these models are trained on one Nvidia-A100 GPU. The reported times correspond to the approximated training duration of the models. Configuration details, further experiments and ablatives on our contributions can be found in the supplementary materials. + +# 4.2 Novel View Synthesis and Anti-Aliasing + +This experiment aims to evaluate the reconstruction quality through the ability to synthesize viewpoints that are not supervised during the reconstruction. These new viewpoints differ from the angle of observation, but also from the distance of observation. The latter is particularly important since a reconstruction or a + +Table 1: Novel View Synthesis performances for the Mono-Scale setup (trained on the full resolution images only) on the 360 Dataset. The indicated resolutions refer to the resolution of the renders. + +
Full Res.1/2 Res.1/4 Res.1/8 Res.Time ↓
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Nerfacto [20]27.090.7790.18126.810.7820.16225.220.7110.23423.320.6360.2970.45h
PyNeRF [21]27.870.8020.16021.930.7130.21121.150.6620.26420.390.6190.3120.96h
Zip-NeRF [4]28.060.8080.15416.580.5960.31911.880.4240.4659.660.3230.5231.10h
RING-NeRF28.090.7990.15727.180.8040.13825.820.7860.14224.380.7370.1670.45h
+ +rendering process that does not take correctly into account the distance of observation leads to artefacts, from over-contrasted rendering to aliasing phenomenon. The challenge is to avoid these artefacts while keeping the reconstruction and rendering process as fast as possible. + +Dataset. The evaluation relies on the dataset introduced by Mip-NeRF-360 [3]. This dataset is composed of 9 scenes, each containing both a central area and complex background in both inside and outside setups. Since the trajectory keeps a constant distance to the central part, each viewpoint is represented with a pyramid of 4 different image resolutions to simulate a variation of the distance of observation combined with a change of the camera FOV, following the Mip-NeRF [2] original anti-aliasing evaluation pipeline. + +Algorithms. We compare our solution against several grid-based NeRF baselines, both with (PyNeRF [21] and Zip-NeRF [4]) and without anti-aliasing processing (Nerfacto), using their Nerfstudio implementations. + +Protocol. For each scene, we train the models with two different setups: the mono-scale setup that uses only the image with the highest resolution for each viewpoint, and the multi-scale setup which uses the whole pyramid of images for each viewpoint. Note that, for these two setups, we evaluate the performances on the whole resolution pyramid, with usual metrics (PSNR, SSIM, LPIPS). + +Results. First of all, regarding novel view synthesis quality using the monoscale setup for both training and testing, as referred to in the "Full Res." column of table 1, we notice that our architecture provides on par performances with Zip-NeRF, slightly better results than PyNeRF, and a more important gap with Nerfacto. While simple, RING-NeRF succeeds in performing state-of-the-art performances on a real single-scale dataset. Regarding quality for the multi-scale setup, as presented in table 2, we observe that all the algorithms that consider the distance of observation perform very similarly in terms of quality. On the + +![](images/4e1ab55963528c40ffc68316a7943fac1db4d2db9816f2944fd16e7c8608c3f7.jpg) +Fig. 4: Examples of image renderings at 1/8th resolution from models solely trained with the full resolution images. RING-NeRF is the only method capable of producing coherent aliasing-free renderings thanks to its LOD inductive bias. + +![](images/e10b9442e45182a7f15afc5fefbc310cee93cb4f1265391879c441bf6feed875.jpg) + +![](images/9b772adcc99b6dacb6a804e24fcf7d7855c02a82ccf420e8c917ed836757b4fc.jpg) + +![](images/9c58a02448d7a3bd308a112f08b9f77255d167ded6ad2008fc4ec33aa458832b.jpg) + +![](images/08e2a35d1620d5f24568c3546003dbfd1345c07123da38c82a92fc70e56101fd.jpg) + +Table 2: Novel View Synthesis performances for the Multi-Scale setup (trained jointly on every resolution) on the 360 Dataset. The indicated resolutions refer to the resolution of the renders. + +
Full Res.1/2 Res.1/4 Res.1/8 Res.Time ↓
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Nerfacto [20]25.980.6680.36727.310.7920.17627.160.8050.13925.300.7430.1990.45h
PyNeRF [21]27.650.7810.19429.470.8570.09230.510.8870.06330.860.8930.0550.96h
Zip-NeRF [4]27.540.7810.19129.340.8580.09930.360.8890.06430.900.9000.0471.10h
RING-NeRF27.660.7820.19429.550.8560.09430.510.8880.06230.860.9010.0480.45h
+ +other side, Nerfacto performs quite poorly. Its performances are especially low at the coarsest resolution, with overflowing artefacts, as illustrated in figure 1. + +Finally, we evaluate the capacity of the models to generalize over new resolutions by training them on full-resolution images and then evaluate them on smaller resolutions. Qualitative results are shown in figure 4 while quantitative results can be found in the 1/2-th, 1/4-th and 1/8-th res. columns of table 1. Because most anti-aliasing methods need LOD-specific supervision to correctly function (including Zip-NeRF and PyNeRF), they cannot render coherent anti-aliased images in this setup. However, PyNeRF behaves better than ZipNeRF with coherent although very aliased renderings, as it decides to limit the LOD used for rendering inside the range of LOD seen during training. RING-NeRF, with his LOD inductive bias, is the only architecture capable of producing anti-aliased renderings from novel observation distances and thus outperforms every other method. While this experiment can seem somewhat esoteric as training on a multi-resolution images pyramid is rather easy, this increases GPU memory and total training time. Moreover, depending on the resolution and the trajectory, it is not trivial to choose the accurate number of scales in the image pyramid to supervise correctly every grid in the hierarchy and especially the coarsest grids. + +Regarding the reconstruction processing time reported in the "Time" column of both table 1 and table 2, since RING-NeRF does not rely on the multiplication of either sample or decoder, it processes as quickly as the fastest Ner-facto both on mono-scale and multi-scale setups. On the other hand, the other anti-aliasing methods, PyNeRF with its per-LOD MLP and Zip-NeRF with its super-sampling, are approximately 2.5 times slower. + +In conclusion, our solution provides the best quality-speed trade-off since it is both on par with the best quality method and the fastest method, in monoscale as well as in multi-scale setups. Furthermore, RING-NeRF is the only solution capable of creating coherent and anti-aliased renderings when facing novel observation distances unseen during training. + +# 4.3 Few Viewpoints Supervision + +This experiment aims to evaluate the influence of the RING-NERF architecture and pipeline on the reconstruction robustness to limited supervision viewpoints. + +Dataset. We evaluate our contribution on the object-centric real dataset DTU [9], often used in few-viewpoints evaluations. + +Algorithms. We compare our architecture against several baselines: Mip-NeRF + +[2], and FreeNerf [25] (a state-of-the-art method for this task), for vanilla architectures and Nerfacto for grid-based architectures. To better demonstrate the intrinsic stability brought by RING-NeRF, we also developed the Nerfacto+ architecture, which corresponds to a Nerfacto architecture coupled with a coarse-to-fine training based on a progressive activation of the grids (the rest of the decoder's input being filled with zeros). For Nerfacto+ and RING-NeRF, we also add FreeNeRF's loss that penalizes the density of the first $M = 10$ samples of each ray to reduce as much as possible artefacts in front of the cameras. As an ablative experiment, we also evaluate RING-NeRF using discrete coarse-to-fine (fixed LOD increment of 1 rather than the proposed continuous increase). + +Protocol. We follow FreeNeRF's evaluation pipeline, including the number of supervision viewpoints (3 to 9), the choice of these views among the dataset and the evaluations using masks of the object. Since we are using the same protocol, the results of FreeNeRF and Mip-NeRF were taken out of FreeNeRF's article. + +Results. Evaluation results are reported in table 3. We first notice an important difference between vanilla and grid-based baselines. While Mip-NeRF faces troubles in reconstructing the scene with 3 views, the method seems to find coherency when adding more images. However, Nerfacto struggles much more to form a consistent 3D scene even when using 9 images (see figure 1). Even though the grid-based baseline is way faster to train, its design implies more instability when facing a small number of images. This does not mean however that few viewpoints are incompatible with grid-based methods. Using progressive training coupled with a very simple and generalizable regularization, both Nerfacto+ and our architecture succeed in creating coherent geometry. Nonetheless, RING-NeRF considerably outperforms Nerfacto+, with a PSNR difference varying from 3 to 4, demonstrating the stability increase brought by our architecture, and also outperforms Mip-NeRF for the configuration with 3 and 6 supervision images. The discrete coarse-to-fine version of RING-NeRF performs in-between Nerfacto+ and the complete RING-NeRF. This showcases both the intrinsic interest of the proposed architecture against the Nerfacto+ and the relevance of the continuous coarse-to-fine mechanism. FreeNerf remains the best performer of all in terms of quality, but with a reconstruction time that is extremely slower than RING-NeRF, the former achieving its reconstruction in 2.56 hours while the latter only requires less than 10 minutes. RING-NeRF thus offers a better quality-speed trade-off for the few-view reconstruction issue (see figure 1). + +Table 3: Performances of reconstruction from few viewpoints on the DTU dataset. The reported metrics are computed based on the mask of the object. + +
#ImagesPSNR ↑SSIM ↑LPIPS ↓Time ↓
369369369
Mip-NeRF [2]8.6816.5423.580.5710.7410.8790.3530.1980.0922.56h
FreeNeRF [25]19.9223.2525.380.7870.8440.8880.1350.0950.0672.56h
Nerfacto [20]9.359.759.780.5670.6040.6470.3850.3310.3260.15h
Nerfacto+13.6116.6119.330.6390.6990.7590.2760.2180.1510.15h
RING-NeRF16.1820.4723.190.7130.8080.8470.2000.1270.0850.15h
w/ discrete CtF15.7920.1622.930.7060.7850.8470.2010.1270.0850.15h
+ +Table 4: SDF reconstruction performances when foregoing the scene-specific SDF Initialization on the Replica Dataset. The Chamfer distance is in centimeters. + +
PSNR ↑SSIM ↑LPIPS ↓Chamfer-L1 ↓Training Time ↓
NeuS-facto [26]24.620.7780.34717.611.4h
NeuralAngelo [11]30.790.9160.076113.693.40 h
RING-NeRF37.180.9690.01945.711.28 h
+ +# 4.4 SDF Reconstruction without Initialization + +SDF reconstruction is known to be a more unstable process than density-based NeRF [1], requiring a scene-specific initialization to converge. This initialization becomes an issue in complex environments and incremental setups, where several types of scenes can co-exist and are not necessarily known beforehand. Therefore, in this experiment, we evaluate the ability of RING-NeRF and other SotA architectures to achieve SDF reconstruction without scene-specific initialization. + +Dataset. The evaluation is achieved on a subset of 7 scenes of the Replica [18] synthetic indoor dataset. A Tanks & Temples [10] example is provided in supplementary materials with corresponding analysis. + +Algorithms. We compare our architecture to two SDF methods, all of them implemented in the same SDFStudio [26] branch of the Nerfstudio framework for fairer comparisons: NeuS-facto, an adaptation of NeuS for grid-based methods with Nerfacto modules, and an implementation of NeuralAngelo. For these experiments, our model is built upon the NeuS-facto baseline, using in particular the same NeuS-based SDF-to-density transformation [22]. + +Protocol. To evaluate the impact of the architecture and pipeline over the convergence and stability, we suppress the inverted sphere SDF initialization scheme and use a random initialization for the model. + +Results. Evaluation results are shown in table 4. First of all, NeuS-facto faces low rendering and reconstruction metrics, due to catastrophic failure in most of the tested scenes (see figure 1) since the model tends to re-draw the 2D images in front of the camera. Regarding NeuralAngelo, its relatively high PSNR demonstrates its ability to synthesize satisfying RGB. However, as illustrated in figure 1 and highlighted by the reconstruction metrics, the underlying geometry of the scene is poorly reconstructed, without any fine details. Finally, our method RING-NeRF is by far the best performer, with much higher PSNR and a better geometry including fine details (see figure 1), although a bit noisy. The simple architecture of RING-NeRF also permits faster epochs, thus faster training. + +# 4.5 LOD Extensibility + +This experiment aims to demonstrate RING-NeRF's unique ability to increase dynamically the level of detail of the scene representation. + +Dataset. The scan 114 of the DTU dataset is used for this experiment. + +Algorithms. Because I-NGP architectures [4, 11, 14] cannot perform LOD extensibility (due to the fixed decoder's input size), only RING-NeRF is evaluated. + +Protocol. We train our model using two different configurations : one low resolution with a 3 levels grid hierarchy and one high resolution with 5 levels (the + +![](images/4eaf320a4870d993ff70708bcd7991e489b3dd4205e12abfd74d6e489faf691f.jpg) +Fig. 5: Learning curves and final renderings of RING-NeRF models with different grid configurations trained either jointly or incrementally. + +![](images/f5bcb97c15a12b8e9aa96373f6fd3402f54633e4ca1063c0c5d15018049793c9.jpg) + +![](images/d0630a3dc21b8411ce3938d4ede0f764a4739c201d41375546cfe8f239a71d99.jpg) + +![](images/b638c4748cd31e44ad39d040f95fb7602f257637aead1d88f5d67aa4b382952d.jpg) + +![](images/c6716be8e8b0c7f9a854e46f636030b993ff8f8586f0e5917d978598e0117d5f.jpg) + +grid resolutions are 16, 32, 64, 128 and 256). For these two configurations named "3 levels" and "5 levels", every grid and the decoder are trained simultaneously. We proceed to showcase the extensibility of RING-NeRF by adding grids to the "3 levels" configuration that is previously trained. We first train one grid of resolutions 128 ("3+1 levels" in Figure 5) with the three initial grids and the decoder frozen and then train another grid of resolutions 256 ("3+1+1 levels" in Figures 5) with the four grids and the decoder frozen. + +Results. Figure 5 shows that the configuration "3+1+1 levels" results in the same rendering quality than the "5 levels" one. This demonstrates the ability of our model to dynamically change the resolution of the grid hierarchy. This is an important step towards the development of an adaptive architecture which locally chooses the resolution of the representation based on the scene's content. This allows to drastically reduce the number of parameters, helping in improving the memory footprint, the training duration and the model's robustness. + +# 5 Conclusion and Perspectives + +In this work, we introduced RING-NeRF, a simple and versatile NeRF pipeline that provides two inductive biases by design: a continuous multi-scale representation of the scene, and an invariance of the decoder latent space over spatial and scale domains. Coupled with a distance-aware forward mapping and a continuous coarse-to-fine reconstruction process, our pipeline demonstrated experimentally its versatility with on-par performances with dedicated state-of-the-art solutions for anti-aliasing or reconstruction from few viewpoints. It even outperforms them in terms of robustness to scene-specific initialization for SDF reconstruction. Furthermore, it is highly efficient and is not limited to object-centric scenes. + +Future work will study the impact of RING-NeRF on other challenging use cases, such as facing inaccurate camera poses [16] and SLAM [29]. We will also use the extensibility property of our architecture to develop memory-efficient sparse Neural Fields, which is considered to be a limit of most grid-based models. + +# Acknowledgements + +This publication was made possible by the use of the CEA List FactoryIA supercomputer, financially supported by the Ile-de-France Regional Council. + +# References + +1. Atzmon, M., Lipman, Y.: Sal: Sign agnostic learning of shapes from raw data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2565-2574 (2020) +2. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021) +3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mipnerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022) +4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-nerf: Anti-aliased grid-based neural radiance fields. ICCV (2023) +5. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: European Conference on Computer Vision (ECCV) (2022) +6. Dou, Y., Zheng, Z., Jin, Q., Ni, B.: Multiplicative fourier level of detail. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1808-1817 (2023) +7. Hu, W., Wang, Y., Ma, L., Yang, B., Gao, L., Liu, X., Ma, Y.: Tri-miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19774-19783 (2023) +8. Jain, A., Tancik, M., Abbeel, P.: Putting nerf on a diet: Semantically consistent few-shot view synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5885-5894 (2021) +9. Jensen, R., Dahl, A., Vogiatzis, G., Tola, E., Aanaes, H.: Large scale multi-view stereopsis evaluation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 406-413 (2014) +0. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics 36(4) (2017) +11. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8456-8465 (2023) +12. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020) +13. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) +14. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022) + +15. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5480-5490 (2022) +16. Park, K., Henzler, P., Mildenhall, B., Barron, J.T., Martin-Brualla, R.: Camp: Camera preconditioning for neural radiance fields. ACM Transactions on Graphics (TOG) 42(6), 1-11 (2023) +17. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. ICCV (2021) +18. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019) +19. Takikawa, T., Litalien, J., Yin, K., Kreis, K., Loop, C., Nowrouzezahrai, D., Jacobson, A., McGuire, M., Fidler, S.: Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11358-11367 (2021) +20. Tancik, M., Weber, E., Ng, E., Li, R., Yi, B., Wang, T., Kristoffersen, A., Austin, J., Salahi, K., Ahuja, A., et al.: Nerfstudio: A modular framework for neural radiance field development. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-12 (2023) +21. Turki, H., Zollhöfer, M., Richardt, C., Ramanan, D.: Pynerf: Pyramidal neural radiance fields. In: Thirty-Seventh Conference on Neural Information Processing Systems (2023) +22. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS (2021) +23. Wynn, J., Turmukhambetov, D.: DiffusioNeRF: Regularizing Neural Radiance Fields with Denoising Diffusion Models. In: CVPR (2023) +24. Xu, L., Agrawal, V., Laney, W., Garcia, T., Bansal, A., Kim, C., Rota Bulò, S., Porzi, L., Kontschieder, P., Božić, A., Lin, D., Zollhöfer, M., Richardt, C.: VR-NeRF: High-fidelity virtualized walkable spaces. In: SIGGRAPH Asia Conference Proceedings (2023). https://doi.org/10.1145/3610548.3618139 +25. Yang, J., Pavone, M., Wang, Y.: Freenerf: Improving few-shot neural rendering with free frequency regularization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8254-8263 (2023) +26. Yu, Z., Chen, A., Antic, B., Peng, S., Bhattacharyya, A., Niemeyer, M., Tang, S., Sattler, T., Geiger, A.: Sdfstudio: A unified framework for surface reconstruction (2022), https://github.com/autonomousvision/sdfstudio +27. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems 35, 25018-25032 (2022) +28. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492 (2020) +29. Zhu, Z., Peng, S., Larsson, V., Xu, W., Bao, H., Cui, Z., Oswald, M.R., Pollefeys, M.: Nice-slam: Neural implicit scalable encoding for slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12786-12796 (2022) +30. Zhuang, Y., Zhang, Q., Feng, Y., Zhu, H., Yao, Y., Li, X., Cao, Y.P., Shan, Y., Cao, X.: Anti-aliased neural implicit surfaces with encoding level of detail. Siggraph Asia 2023 (2023) \ No newline at end of file diff --git a/2024/RING-NeRF _ Rethinking Inductive Biases for Versatile and Efficient Neural Fields/images.zip b/2024/RING-NeRF _ Rethinking Inductive Biases for Versatile and Efficient Neural Fields/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a771802b356cc8ea4394bf2cde3d43a26758b6cb --- /dev/null +++ b/2024/RING-NeRF _ Rethinking Inductive Biases for Versatile and Efficient Neural Fields/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f7900b3da04495d8606a53ef792d6bb5c8862524ed76d1226eb050b8c2037a3 +size 392922 diff --git a/2024/RING-NeRF _ Rethinking Inductive Biases for Versatile and Efficient Neural Fields/layout.json b/2024/RING-NeRF _ Rethinking Inductive Biases for Versatile and Efficient Neural Fields/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6a284494e2e0ba3ca35e446cc5a3ccd7c1822285 --- /dev/null +++ b/2024/RING-NeRF _ Rethinking Inductive Biases for Versatile and Efficient Neural Fields/layout.json @@ -0,0 +1,7733 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 144, + 111, + 470, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 111, + 470, + 146 + ], + "spans": [ + { + "bbox": [ + 144, + 111, + 470, + 146 + ], + "type": "text", + "content": "RING-NeRF : Rethinking Inductive Biases for Versatile and Efficient Neural Fields" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "spans": [ + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "type": "text", + "content": "Doriand Petit" + }, + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "type": "text", + "content": ", Steve Bourgeois" + }, + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "type": "text", + "content": ", Dumitru Pavel" + }, + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "type": "text", + "content": ", Vincent Gay-Bellile" + }, + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "type": "text", + "content": ", Florian Chabot" + }, + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "type": "text", + "content": ", and Loic Barthe" + }, + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 174, + 201, + 439, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 201, + 439, + 224 + ], + "spans": [ + { + "bbox": [ + 174, + 201, + 439, + 224 + ], + "type": "text", + "content": "1 Université Paris-Saclay, CEA, List, F-91120, Palaiseau, France \n2 IRIT, Université Toulouse III, CNRS, France" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 147, + 242, + 462, + 407 + ], + "blocks": [ + { + "bbox": [ + 147, + 242, + 462, + 407 + ], + "lines": [ + { + "bbox": [ + 147, + 242, + 462, + 407 + ], + "spans": [ + { + "bbox": [ + 147, + 242, + 462, + 407 + ], + "type": "image", + "image_path": "72ab70ec195e7c177d49d3e764af15c3e127b41b5cb7f1c9512dfe05a18f63fb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 416, + 482, + 472 + ], + "lines": [ + { + "bbox": [ + 130, + 416, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 416, + 482, + 472 + ], + "type": "text", + "content": "Fig. 1: RING-NeRF is a simple and versatile architecture which tackles many NeRF common issues such as robustness to distance of observation, few view supervision and lack of scene-specific initialization for SDF-based reconstruction. It provides on-par performances in terms of quality with SotA dedicated solutions [4,11,25] and and in terms of efficiency with fast methods [20,26]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 490, + 453, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 490, + 453, + 665 + ], + "spans": [ + { + "bbox": [ + 159, + 490, + 453, + 665 + ], + "type": "text", + "content": "Abstract. Recent advances in Neural Fields mostly rely on developing task-specific supervision which often complicates the models. Rather than developing hard-to-combine and specific modules, another approach generally overlooked is to directly inject generic priors on the scene representation (also called inductive biases) into the NeRF architecture. Based on this idea, we propose the RING-NeRF architecture which includes two inductive biases: a continuous multi-scale representation of the scene and an invariance of the decoder's latent space over spatial and scale domains. We also design a single reconstruction process that takes advantage of those inductive biases and experimentally demonstrates on-par performances in terms of quality with dedicated architecture on multiple tasks (anti-aliasing, few view reconstruction, SDF reconstruction without scene-specific initialization) while being more efficient. Moreover, RING-NeRF has the distinctive ability to dynamically increase the resolution of the model, opening the way to adaptive reconstruction. Project page can be found at: https://cea-list.github.io/RING-NeRF" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 149, + 230, + 162 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 230, + 162 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 230, + 162 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 175, + 482, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 175, + 482, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 482, + 317 + ], + "type": "text", + "content": "Neural Radiance Fields (NeRF) have emerged as a novel method for representing 3D scenes using neural networks. In its original design [13], a simple multi-layer perceptron (MLP) is trained to reproduce a continuous 5D function that outputs the density and radiance emitted in each direction " + }, + { + "bbox": [ + 130, + 175, + 482, + 317 + ], + "type": "inline_equation", + "content": "(\\theta, \\phi)" + }, + { + "bbox": [ + 130, + 175, + 482, + 317 + ], + "type": "text", + "content": " at each point " + }, + { + "bbox": [ + 130, + 175, + 482, + 317 + ], + "type": "inline_equation", + "content": "(x, y, z)" + }, + { + "bbox": [ + 130, + 175, + 482, + 317 + ], + "type": "text", + "content": " in space. This approach inspired many works due to the impressive quality of its novel view synthesis as well as its simplicity. However, the so-called \"vanilla\" NeRF architectures converge very slowly as they use large deep neural networks. Instant-NGP [14] introduced a new architecture based on a hierarchical hash-grid pyramid of learnable feature codes describing the 3D scene and combined with a much shallower MLP, called decoder, that transforms the concatenation of the codes, interpolated from the grids, into density and radiance. Resulting in local updates, this method reduced the training duration from hours to minutes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 319, + 481, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 481, + 353 + ], + "type": "text", + "content": "The great majority of current solutions are now based on these two standard architectures, and many of them are focused on overtaking their associated limitations in terms of:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 354, + 481, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 481, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 481, + 378 + ], + "type": "text", + "content": "Nature of scene - by transitioning from object-centric scenes to open unbounded scenes, using mostly space contraction [3];" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 378, + 481, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 481, + 426 + ], + "type": "text", + "content": "Robustness - by managing free motion trajectories with variations of the observation distance while avoiding aliasing artefacts, mostly through the integration of Level of Detail (LOD) in the model [2,4]; or by reducing drastically the number of supervised views through different kinds of regularization [8,15,25];" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 426, + 481, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 481, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 481, + 462 + ], + "type": "text", + "content": "Extensibility - by shifting from a holistic and fixed reconstruction process to an incremental (extensibility in the number of views) and adaptive (extensibility in resolution) process, mainly through the use of a frozen decoder [12, 29]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 462, + 481, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 481, + 521 + ], + "type": "text", + "content": "However, most solutions solely focus on one of these limitations and introduce specific and complex mechanisms that both increase the computational cost and lessen the possibilities of combination. The ability to solve jointly these main issues is however essential in real-world applications which often require a robust and extensible reconstruction in an unknown and unbounded environment." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "content": "In this article, rather than introducing yet another heavy and task-specific solution, we propose RING-NeRF, a versatile and simple NeRF architecture by rethinking usual grid-based models to introduce two inductive biases. We first represent a 3D scene as a continuous multi-scale representation and also make the decoder's latent space invariant in position and scale. Together, these two priors enable the production of intrinsic continuous LOD of the scene without explicit supervision. We demonstrate experimentally that, when combined with adapted cone casting and coarse-to-fine optimization, the resulting architecture is able to compete on several tasks with on-par quality performances with dedicated state-of-the-art solutions while improving speed, robustness and extensibility. The overall process is also simple, easy to implement and generic enough to be coupled with specific solutions. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "D. Petit et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 160, + 117, + 452, + 128 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 117, + 452, + 128 + ], + "spans": [ + { + "bbox": [ + 160, + 117, + 452, + 128 + ], + "type": "text", + "content": "Keywords: NeRF " + }, + { + "bbox": [ + 160, + 117, + 452, + 128 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 117, + 452, + 128 + ], + "type": "text", + "content": " 3D Reconstruction " + }, + { + "bbox": [ + 160, + 117, + 452, + 128 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 117, + 452, + 128 + ], + "type": "text", + "content": " Implicit Neural Representation" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 136, + 116, + 481, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 116, + 481, + 151 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 481, + 151 + ], + "type": "text", + "content": "1. an architecture that, by construction, represents the scene with a continuous and unbounded level of detail without the need for LOD-specific supervision and which permits resolution-adaptive reconstruction;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 152, + 481, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 152, + 481, + 199 + ], + "spans": [ + { + "bbox": [ + 136, + 152, + 481, + 199 + ], + "type": "text", + "content": "2. a distance-aware forward mapping compatible with scene contraction, that takes benefits of the continuous multi-scale representation with an adapted cone casting process to avoid aliasing artefacts when varying the observation distance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 201, + 481, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 201, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 135, + 201, + 481, + 236 + ], + "type": "text", + "content": "3. a continuous coarse-to-fine reconstruction process that improves the convergence and stability (especially in challenging setups such as supervision with few viewpoints or no scene-specific initialization for SDF reconstruction)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 256, + 237, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 256, + 237, + 269 + ], + "spans": [ + { + "bbox": [ + 133, + 256, + 237, + 269 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 282, + 480, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 480, + 330 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 480, + 330 + ], + "type": "text", + "content": "From its original iteration [13], a majority of current research focuses on overtaking limitations of NeRF-based reconstruction in terms of adaptability to various natures of scene, robustness (to varying observation distances or limited amount of viewpoints) and reconstruction extensibility." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 331, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 545 + ], + "type": "text", + "content": "Adaptability to various natures of the scene. The ability of Neural Fields to reconstruct various natures of scene depends on three factors. The first one is related to its architecture itself. Tri-plane architectures [5, 7, 30] are mostly designed for object-centric reconstruction (as they provide a higher density of information in the center of the scene) while vanilla and 3D grid-based NeRF are able to cover a wider variety of scenes, though they initially were still constrained to a limited volume. A second factor is related to the representation of the 3D space of the scene, especially to represent distant elements in open scenes. Some approaches use two different NeRF models to reconstruct separately the foreground and the background [28], whereas others apply space contraction to the 3D scene coordinates [3, 4, 28] to map the infinite scene volume into a bounded one. The last factor is related to the initialization of the Neural Fields. While the random initialization of density-based NeRF can adjust to almost any nature of scenes, the convergence of SDF-based (Signed Distance Function) Neural Fields is extremely sensitive to their parameters' initialization, as stated in [1]. Current solutions rely on a scene-specific initialization (using an SDF field representing a sphere for outdoor scenes or an inverted sphere for indoor scenes), making them unable to adapt automatically to any scene." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Robustness to observation distance variations. The initial NeRF model, as well as most of the subsequent works, relies implicitly on the hypothesis of a constant distance of observation to the scene. Indeed, the NeRF model provides a per-3D-point scene density representation and a rendering process which does not take into account the increasing volume covered by a pixel with respect to the distance to the cameras. As underlined in Mip-NeRF [2], this discrepancy induces artefacts such as over-contrasted images or aliasing phenomenon when the distance of observation differs from the ones used at reconstruction time. To avoid these artefacts, the rendering process needs to assess the density and color for a volume instead of a point. In the current state-of-the-art, two main" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "text", + "content": "RING-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 307 + ], + "type": "text", + "content": "approaches are used (or their combination, as in Zip-NeRF [4]). The first solution consists of representing the scene with different levels of detail (LOD), in order to vary the precision of the reconstruction based on the observation distance. This is usually done by using a LOD-aware latent space [2,4,6,19,21,24], meaning that the LOD information is already encoded in the inputs of the MLP. This can be achieved by using a per-LOD decoder as in PyNeRF [21] or by incorporating LOD information in the latent feature as done in Zip-NeRF [4] and VR-NeRF [24]. One main flaw of these solutions is that they require the supervision of every used LOD which makes it impossible to vary the observation distances between the train views and novel synthesised views. A second approach consists of defining the latent representation of a volume as the mean of the latent features of the points included in the volume. It requires integrating the features over the volume, which can be achieved through convolutions for tri-plan representation [7,30], or through super-sampling of the latent space for 3D grids as also done in Zip-NeRF [4]. However, these approaches lengthen the training and rendering processes as they increase the number of computations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 308, + 482, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 474 + ], + "type": "text", + "content": "Robustness to limited amount of viewpoints. By construction, NeRF is subject to the shape-radiance ambiguity [28]. If not enough supervision viewpoints are available, the optimization might overfit them while not providing consistent 3D reconstruction nor generalization to non-supervised viewpoints. A first family of solutions to overcome this limitation consists of regularizing the reconstruction process through additional losses, whether via geometric [15,23,27] or semantic [8] regularization. The second family of solutions relies on a progressive reconstruction of the details of the scene, as introduced in FreeNeRF [25] and Nerfies [17]. Restraining the ability of the model to reconstruct a complex scene at early stages enforces the consistency of the reconstruction over the different supervision viewpoints. This presents the advantage of bringing stability while keeping a fairly simple training process. However, these latter solutions mostly rely on Vanilla models for stability purposes and require long training duration." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 475, + 482, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 656 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 656 + ], + "type": "text", + "content": "Reconstruction extensibility. Two kinds of extensibility should be distinguished: extensibility with respect to the number of views or the resolution. The first one is related to reconstructing new scene areas while keeping the previously reconstructed ones unchanged. The usual solution consists of using grid-based approaches with a pre-trained position-invariant decoder that is frozen during the reconstruction [29]. On the other side, the extensibility of the resolution consists in dynamically increasing the level of detail of a previously reconstructed scene. This problem is intractable for classic architectures such as vanilla NeRF and I-NGP since the number of layers (resp. grids) of those approaches cannot increase during the reconstruction. Some rare solutions, such as Neural Sparse Voxel Fields [12], combine a pre-trained decoder with a data structure allowing to dynamically allocate voxels to increase the reconstruction resolution. This is however a crucial stake, as being able to adapt the precision of the representation based on the complexity of the scene permits to optimize the computational efficiency (both in speed and memory requirements) with minimal loss in quality." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "D. Petit et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 114, + 476, + 228 + ], + "blocks": [ + { + "bbox": [ + 141, + 114, + 476, + 228 + ], + "lines": [ + { + "bbox": [ + 141, + 114, + 476, + 228 + ], + "spans": [ + { + "bbox": [ + 141, + 114, + 476, + 228 + ], + "type": "image", + "image_path": "d84823d8ac274faac4824caf00c8975b98a84de9c4e9946b64c03e02d8752d6a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 236, + 479, + 291 + ], + "lines": [ + { + "bbox": [ + 131, + 236, + 479, + 291 + ], + "spans": [ + { + "bbox": [ + 131, + 236, + 479, + 291 + ], + "type": "text", + "content": "Fig. 2: Overview of RING-NeRF: to render a pixel, the casted cone is sampled with cubes. Depending on the cube volume, the corresponding LOD of the scene is selected and the latent feature is computed using a weighted sum of the grid hierarchy. The density (or SDF) and color of the cube are first decoded from the latent feature with a tiny MLP and then integrated with other samples through volume rendering." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 304, + 230, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 230, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 230, + 316 + ], + "type": "text", + "content": "3 RING-NeRF" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 336, + 479, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 336, + 479, + 371 + ], + "spans": [ + { + "bbox": [ + 131, + 336, + 479, + 371 + ], + "type": "text", + "content": "Rather than focusing on solving one specific problem of NeRF using complex mechanisms, we propose a simple architecture called RING-NeRF constructed with novel inductive biases to tackle NeRF's common issues." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 396, + 208, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 208, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 208, + 407 + ], + "type": "text", + "content": "3.1 Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 422, + 480, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 422, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 131, + 422, + 480, + 544 + ], + "type": "text", + "content": "RING-NeRF relies on the classic NeRF [13] inverse rendering pipeline which is used to reconstruct a 3D scene from a set of localized frames. For a given image pixel (with its camera's pose), a 3D ray is cast and the 3D scene representation is sampled at " + }, + { + "bbox": [ + 131, + 422, + 480, + 544 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 422, + 480, + 544 + ], + "type": "text", + "content": " various locations along the ray. The resulting density " + }, + { + "bbox": [ + 131, + 422, + 480, + 544 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 131, + 422, + 480, + 544 + ], + "type": "text", + "content": " (or SDF converted to density [22]) and color " + }, + { + "bbox": [ + 131, + 422, + 480, + 544 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 131, + 422, + 480, + 544 + ], + "type": "text", + "content": " of the samples are then combined with usual volume rendering techniques: " + }, + { + "bbox": [ + 131, + 422, + 480, + 544 + ], + "type": "inline_equation", + "content": "\\hat{C}(r) = \\sum_{i=0}^{N-1} T_{i} (1 - \\exp(-\\sigma_{i} \\delta_{i})) c_{i}" + }, + { + "bbox": [ + 131, + 422, + 480, + 544 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 131, + 422, + 480, + 544 + ], + "type": "inline_equation", + "content": "T_{i} = \\exp(\\sum_{j=0}^{i-1} \\sigma_{j} \\delta_{j})" + }, + { + "bbox": [ + 131, + 422, + 480, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 422, + 480, + 544 + ], + "type": "inline_equation", + "content": "\\delta_{i}" + }, + { + "bbox": [ + 131, + 422, + 480, + 544 + ], + "type": "text", + "content": " is the distance between samples. The parameters of the scene representation are then optimized by minimizing the MSE loss " + }, + { + "bbox": [ + 131, + 422, + 480, + 544 + ], + "type": "inline_equation", + "content": "||\\hat{C}(r) - C(r)||^{2}" + }, + { + "bbox": [ + 131, + 422, + 480, + 544 + ], + "type": "text", + "content": " between the rendered ray color and the ground truth pixel value, the rendering being differentiable3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 545, + 480, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 545, + 480, + 641 + ], + "spans": [ + { + "bbox": [ + 131, + 545, + 480, + 641 + ], + "type": "text", + "content": "The originality of RING-NeRF relies on its neural architecture, illustrated in figure 2, which is designed to represent the scene with a continuous and unbounded level of detail without the need for LOD-specific supervision (see section 3.2). We then use this LOD inductive bias to adjust the LOD of the samples with respect to the distance to the camera in contracted space for more accurate renderings (see section 3.3). Finally, combined with a continuous coarse-to-fine optimization, RING-NeRF results in a more robust reconstruction process with an intrinsic LOD extensibility property (see section 3.4)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "text", + "content": "RING-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 479, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 479, + 665 + ], + "type": "text", + "content": "3 With SDF, an additional Eikonal loss is also being optimized for SDF regularization." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 384, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 384, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 384, + 128 + ], + "type": "text", + "content": "3.2 NeRF Representation with Inductive Biases" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 135, + 482, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 135, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 135, + 482, + 278 + ], + "type": "text", + "content": "Classic grid-based representations, such as the one introduced by I-NGP [14], rely on two key elements: a unique MLP decoder which transforms a latent space into an output space (color, density or SDF), and a 3D (hierarchical) grid of latent features that implicitly defines a unique mapping function from the scene space onto the decoder latent space. We can distinguish two different approaches used to extend the mono-scale representation to a multi-scale representation: conditioning with scale information the decoding process [4,21] and/or defining separate per-LOD mapping functions [21]. In both cases, the training becomes more complex, leading to potential convergence issues as it relies solely on additional specific supervision. Instead, we propose to condition the mapping function itself with scale information and introduce inductive biases to guide the convergence." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 279, + 482, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 482, + 316 + ], + "type": "text", + "content": "Scene representation as multi-scale mapping function. As illustrated in Figure 2, the mapping function at level " + }, + { + "bbox": [ + 130, + 279, + 482, + 316 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 279, + 482, + 316 + ], + "type": "text", + "content": " is controlled by a grid " + }, + { + "bbox": [ + 130, + 279, + 482, + 316 + ], + "type": "inline_equation", + "content": "G_{N}" + }, + { + "bbox": [ + 130, + 279, + 482, + 316 + ], + "type": "text", + "content": " that is implicitly defined through a recursive refinement process over scale:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 250, + 325, + 480, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 325, + 480, + 356 + ], + "spans": [ + { + "bbox": [ + 250, + 325, + 480, + 356 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} G _ {0} \\leftarrow \\delta_ {0} \\\\ G _ {N} \\leftarrow S \\left(G _ {N - 1}\\right) + \\delta_ {N} \\end{array} \\right. \\tag {1}", + "image_path": "86d83968b89a62cfe7cee389a9e45a590c18677374b1cc95f981e4189f4d6e5f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 359, + 482, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 359, + 482, + 466 + ], + "spans": [ + { + "bbox": [ + 130, + 359, + 482, + 466 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 359, + 482, + 466 + ], + "type": "inline_equation", + "content": "\\delta_{i}" + }, + { + "bbox": [ + 130, + 359, + 482, + 466 + ], + "type": "text", + "content": " is the deviation defined by the i-th latent feature grid and S is the subdivision scheme consisting in increasing the resolution of " + }, + { + "bbox": [ + 130, + 359, + 482, + 466 + ], + "type": "inline_equation", + "content": "G_{N - 1}" + }, + { + "bbox": [ + 130, + 359, + 482, + 466 + ], + "type": "text", + "content": " up to the " + }, + { + "bbox": [ + 130, + 359, + 482, + 466 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 359, + 482, + 466 + ], + "type": "text", + "content": "-th grid resolution (with tri-linear interpolation). In practice, the latent vector associated with a point in the scene for a level of detail " + }, + { + "bbox": [ + 130, + 359, + 482, + 466 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 359, + 482, + 466 + ], + "type": "text", + "content": " is simply obtained by interpolating linearly the point in the cell for each level inferior or equal to " + }, + { + "bbox": [ + 130, + 359, + 482, + 466 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 359, + 482, + 466 + ], + "type": "text", + "content": ", and summing the results. Since linear interpolation is differentiable, our mapping function is also estimizable. The summing of the interpolation is comparable to a residual connection and gives the name of the architecture RING-NeRF for Residual Implicit Neural Grids." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 467, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 467, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 467, + 482, + 552 + ], + "type": "text", + "content": "Such representation provides several advantages. First, since the mapping function is constructed in a top-down manner, refining recursively one level of detail from the previous one, its number of LOD is unbounded and adding a finer level of detail keeps the coarser ones valid. Secondly, because each " + }, + { + "bbox": [ + 130, + 467, + 482, + 552 + ], + "type": "inline_equation", + "content": "\\delta_{i}" + }, + { + "bbox": [ + 130, + 467, + 482, + 552 + ], + "type": "text", + "content": " only represents a deviation of the coarser mapping value " + }, + { + "bbox": [ + 130, + 467, + 482, + 552 + ], + "type": "inline_equation", + "content": "G_{i - 1}" + }, + { + "bbox": [ + 130, + 467, + 482, + 552 + ], + "type": "text", + "content": ", a continuous LOD representation can be easily obtained by incorporating a weighting factor " + }, + { + "bbox": [ + 130, + 467, + 482, + 552 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0,1]" + }, + { + "bbox": [ + 130, + 467, + 482, + 552 + ], + "type": "text", + "content": " in the recursive process:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 257, + 561, + 480, + 573 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 561, + 480, + 573 + ], + "spans": [ + { + "bbox": [ + 257, + 561, + 480, + 573 + ], + "type": "interline_equation", + "content": "G _ {L} \\leftarrow S \\left(G _ {N - 1}\\right) + \\alpha \\delta_ {N} \\tag {2}", + "image_path": "c258cd59f0d9467cafa4a7044dfc277831d34d8e3c095bcde21ed942125a7cf2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 581, + 321, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 321, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 321, + 593 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 130, + 581, + 321, + 593 + ], + "type": "inline_equation", + "content": "L\\in [N - 1,N]" + }, + { + "bbox": [ + 130, + 581, + 321, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 581, + 321, + 593 + ], + "type": "inline_equation", + "content": "\\alpha = L - (N - 1)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Spatial and scale invariance of decoder. Since our architecture does not rely on a decoder conditioned on position (unlike [19]) nor scale (unlike [2-4, 21]), the decoder latent space is invariant to translations and scale changes in the scene coordinate. This property makes RING-NeRF more suited for incremental reconstruction in both spatial and scale space, since it ensures that local updates in the spatial and scale domain of the scene can be achieved through the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "text", + "content": "D. Petit et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 117, + 202, + 168 + ], + "blocks": [ + { + "bbox": [ + 135, + 117, + 202, + 168 + ], + "lines": [ + { + "bbox": [ + 135, + 117, + 202, + 168 + ], + "spans": [ + { + "bbox": [ + 135, + 117, + 202, + 168 + ], + "type": "image", + "image_path": "2db464ad350a9ef5df99d47be7363b59bd182042defe315f3196b4660b803091.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 182, + 480, + 225 + ], + "lines": [ + { + "bbox": [ + 130, + 182, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 480, + 225 + ], + "type": "text", + "content": "Fig. 3: We demonstrate the LOD inductive bias by training our model with a hierarchy of " + }, + { + "bbox": [ + 130, + 182, + 480, + 225 + ], + "type": "inline_equation", + "content": "N = 7" + }, + { + "bbox": [ + 130, + 182, + 480, + 225 + ], + "type": "text", + "content": " levels where only the last level of the mapping function " + }, + { + "bbox": [ + 130, + 182, + 480, + 225 + ], + "type": "inline_equation", + "content": "G_{N}" + }, + { + "bbox": [ + 130, + 182, + 480, + 225 + ], + "type": "text", + "content": " is supervised. We then compute renders at different levels " + }, + { + "bbox": [ + 130, + 182, + 480, + 225 + ], + "type": "inline_equation", + "content": "L \\leq N" + }, + { + "bbox": [ + 130, + 182, + 480, + 225 + ], + "type": "text", + "content": ". Other examples (including of entire scenes) can be found in the supplementary materials." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 203, + 118, + 271, + 169 + ], + "blocks": [ + { + "bbox": [ + 203, + 118, + 271, + 169 + ], + "lines": [ + { + "bbox": [ + 203, + 118, + 271, + 169 + ], + "spans": [ + { + "bbox": [ + 203, + 118, + 271, + 169 + ], + "type": "image", + "image_path": "a526fb7ed0a150bac77bbb989deef339b0eb5fc05ecd7338a9a0489ffba4b002.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 272, + 118, + 339, + 168 + ], + "blocks": [ + { + "bbox": [ + 272, + 118, + 339, + 168 + ], + "lines": [ + { + "bbox": [ + 272, + 118, + 339, + 168 + ], + "spans": [ + { + "bbox": [ + 272, + 118, + 339, + 168 + ], + "type": "image", + "image_path": "c96e543197cd47c549eaab334d8472b498fd341f24262108eb01c99c2ad1d362.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 340, + 118, + 408, + 168 + ], + "blocks": [ + { + "bbox": [ + 340, + 118, + 408, + 168 + ], + "lines": [ + { + "bbox": [ + 340, + 118, + 408, + 168 + ], + "spans": [ + { + "bbox": [ + 340, + 118, + 408, + 168 + ], + "type": "image", + "image_path": "361e48c86ec61bb06e172fe2e8bfa11ebc51ee8553609c8b4c0b67aa5cf744d4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 410, + 118, + 478, + 168 + ], + "blocks": [ + { + "bbox": [ + 410, + 118, + 478, + 168 + ], + "lines": [ + { + "bbox": [ + 410, + 118, + 478, + 168 + ], + "spans": [ + { + "bbox": [ + 410, + 118, + 478, + 168 + ], + "type": "image", + "image_path": "a309244325ed4cb5cfe637a343a4fbc7637c7c9bd1366844a7a819d47e257c16.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 242, + 479, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 242, + 479, + 265 + ], + "spans": [ + { + "bbox": [ + 130, + 242, + 479, + 265 + ], + "type": "text", + "content": "hierarchical grid. The decoder architecture is illustrated in figure 2 and further details can be found in the supplementary materials." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 266, + 480, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 480, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 480, + 469 + ], + "type": "text", + "content": "LOD inductive bias. During the reconstruction process, the gradients of the ray samples are backpropagated through the residual connections and aggregated for each grid level. Due to the pyramidal resolution of the hierarchical grid, a grid code at coarse levels influences a large scene volume and is thus supervised by more ray samples. Therefore, the gradient of a grid code increases as the level's resolution decreases, and as long as the associated samples' gradients are uniform. However, once the backpropagation through the residual connections reach a level whose associated samples have divergent enough gradients (meaning the error is finer than this level's resolution), the result of the aggregation will be mitigated. Hence, the level with the maximum correction is always the coarser level where the samples' gradients are still uniform. This property naturally induces corrections at the proper grid level and LOD. In figure 3, we illustrate this inductive bias by displaying different continuous LOD of a scene while the reconstruction is only supervised for its finest LOD. Not only this is a useful property when an unsupervised multi-scale representation is needed (as demonstrated in section 4.2), this also guarantees a more robust convergence (as illustrated in section 4.3 and section 4.4)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 488, + 332, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 488, + 332, + 501 + ], + "spans": [ + { + "bbox": [ + 131, + 488, + 332, + 501 + ], + "type": "text", + "content": "3.3 Distance-Aware Forward Mapping" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 510, + 480, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 480, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 480, + 581 + ], + "type": "text", + "content": "Cone Casting in contracted space. In order to accommodate the scene rendering process to the variation of observation distances, we introduce a distance-aware forward mapping mechanism. Similarly to the cone casting of [7], it relies on assigning to each sample a latent feature whose LOD is inversely proportional to the sample-camera distance. However, unlike [7, 21], our model relies on the use of space contraction to allow the reconstruction of unbounded scenes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": "To define the LOD of a sample at distance " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": " of the camera, we first compute the associated volume of the cone cast pixel cube in the world space coordinates (see fig. 2). Assuming the pixel is a square of size " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": " at distance 1 of the camera (" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": " depends on the image resolution and camera's FOV), the volume is thus " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "V = (dc)^3" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": ". To take into account the space contraction, we then proceed to contract the volume. Denoting " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": " the Jacobian of the contraction function at the sample's location " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "V_{contract} = V\\det(J(p))" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": ". In practice, we compute the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "text", + "content": "RING-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "analytical derivation of the contraction function depending on " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": ". More details can be found in the supplementary materials." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": "Assuming the " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": "-th feature grid's resolution in our hierarchy can be written as " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "f^{N}b" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": " the resolution of the grid " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\delta_0" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": " the growth factor, finding the appropriate LOD " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "L \\in \\mathbb{R}^+" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": " means finding a virtual grid4 of resolution " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "f^{L}b" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": " whose cell's volume is equal to the previously computed volume (in contracted space). Because we are working in the contracted space of size 1, the volume of a cell of the virtual grid of LOD " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "(\\frac{1}{f^{L}b})^3" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": ". The LOD " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": " associated to one sample in the contracted space is thus given by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 174, + 233, + 481, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 233, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 174, + 233, + 481, + 266 + ], + "type": "interline_equation", + "content": "\\left(d c\\right) ^ {3} \\det (J (p)) = \\left(\\frac {1}{f ^ {L} b}\\right) ^ {3} \\Longleftrightarrow L = - \\frac {\\log \\left(d c b \\sqrt [ 3 ]{\\det (J (p))}\\right)}{\\log (f)} \\tag {3}", + "image_path": "016514c7f020bfa7273febbca6bbe3211155d15fc0db37b8cf9632277b0c8509.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 274, + 479, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 479, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 479, + 321 + ], + "type": "text", + "content": "Note that this process is close to ZipNeRF [4] and VR-NeRF [24]. However, the first one rather derives a contracted scale factor from its Gaussian samples while the latter directly computes the LOD in the contracted space, which can be considered an approximation of our computation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 322, + 481, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 481, + 357 + ], + "type": "text", + "content": "Forward mapping. As illustrated in Figure 2, we use the determined LOD to compute a distance-dependent weighted sum of the features, which is fed to our decoder and transformed into density (or SDF) and radiance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 375, + 444, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 375, + 444, + 388 + ], + "spans": [ + { + "bbox": [ + 131, + 375, + 444, + 388 + ], + "type": "text", + "content": "3.4 Continuous Coarse-To-Fine and Resolution Extensibility" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 395, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 395, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 395, + 482, + 491 + ], + "type": "text", + "content": "Recent works proposed to use coarse-to-fine optimization to improve the stability of NeRF models, especially when facing more challenging setups, including with few images [25] and surface-based models [11]. It consists of optimizing progressively the different LOD of the representation, from coarse levels to the most precise ones. The goal of this progressive optimization is to avoid the shape-radiance ambiguity [28] by introducing a strong regularization through LOD restriction, then relaxing progressively this regularization once the coarse geometry of the scene is reconstructed to recover the details of the scene." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 491, + 482, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 599 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 599 + ], + "type": "text", + "content": "The coarse-to-fine reconstruction process of RING-NeRF consists of estimating progressively the LOD of the mapping function from the coarsest to the finest ones. In practice, it implies, during the cone casting, to clamp the samples' LOD up to a maximal LOD " + }, + { + "bbox": [ + 130, + 491, + 482, + 599 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 491, + 482, + 599 + ], + "type": "text", + "content": ", the grids of level " + }, + { + "bbox": [ + 130, + 491, + 482, + 599 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 491, + 482, + 599 + ], + "type": "text", + "content": " and above being set to zero and not optimized. Moreover, since our architecture provides continuous LOD, the coarse-to-fine optimization can be achieved continuously in the LOD space by using a linear scheduler " + }, + { + "bbox": [ + 130, + 491, + 482, + 599 + ], + "type": "inline_equation", + "content": "l = (l_0 + \\frac{n}{n_{ctf}}) \\in \\mathbb{R}^+" + }, + { + "bbox": [ + 130, + 491, + 482, + 599 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 491, + 482, + 599 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 491, + 482, + 599 + ], + "type": "text", + "content": " the current epoch, " + }, + { + "bbox": [ + 130, + 491, + 482, + 599 + ], + "type": "inline_equation", + "content": "n_{ctf}" + }, + { + "bbox": [ + 130, + 491, + 482, + 599 + ], + "type": "text", + "content": " a hyperparameter describing the speed of the process and " + }, + { + "bbox": [ + 130, + 491, + 482, + 599 + ], + "type": "inline_equation", + "content": "l_0" + }, + { + "bbox": [ + 130, + 491, + 482, + 599 + ], + "type": "text", + "content": " defining the number of used grids at initialization; up to a specified maximum resolution." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 600, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 600, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 600, + 482, + 635 + ], + "type": "text", + "content": "Furthermore, the RING-NeRF architecture is more adapted than I-NGP-based architectures [11] for coarse-to-fine training. Indeed, for solutions based on the concatenation of features, keeping grid values to zero implies that some" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "text", + "content": "D. Petit et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 641, + 482, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 641, + 482, + 666 + ], + "type": "text", + "content": "4 A virtual grid of LOD " + }, + { + "bbox": [ + 133, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "L \\in \\mathbb{R}^{+}" + }, + { + "bbox": [ + 133, + 641, + 482, + 666 + ], + "type": "text", + "content": " corresponds to a grid of resolution " + }, + { + "bbox": [ + 133, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "f^{L}b" + }, + { + "bbox": [ + 133, + 641, + 482, + 666 + ], + "type": "text", + "content": " that is not explicitly stored in memory but whose elements can be computed from other grids." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": "dimensions of the decoder's latent space are not supervised. When a grid starts to be optimized, those unsupervised dimensions are suddenly used. The weights of the decoder thus need to be refined, with a global effect on the whole scene. On the opposite, our solution keeps supervising all the dimensions of the decoder latent space, and when a new grid gets optimized, it only implies more degrees of freedom to define the mapping function between the scene space and the decoder latent space. This also means that our scene reconstruction can be refined by adding dynamically new grid levels without modifying the decoder's weights or previously trained grids, as we demonstrate experimentally in section 4.5. This resolution extensibility property opens the path to adaptive resolution models, where the precision used to describe an area depends on the details needed, to optimize efficiency both in memory consumption and training duration." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 280, + 230, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 280, + 230, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 280, + 230, + 295 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 307, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 482, + 403 + ], + "type": "text", + "content": "In these experiments, we intend to highlight the versatility of RING-NeRF by evaluating it on several tasks. After introducing implementation details in section 4.1, we evaluate our model on novel view synthesis with changes in observation distances (sec. 4.2). Then, we explore how robust is RING-NeRF first with few view reconstruction (sec. 4.3) and then without scene-specific initialization for SDF reconstruction (sec. 4.4). Finally, we demonstrate the capacity of our architecture to perform LOD extensibility, as a first step towards adaptive reconstruction (sec. 4.5)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 422, + 241, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 422, + 241, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 422, + 241, + 435 + ], + "type": "text", + "content": "4.1 Implementation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 444, + 482, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 444, + 482, + 576 + ], + "spans": [ + { + "bbox": [ + 130, + 444, + 482, + 576 + ], + "type": "text", + "content": "Our model is based on the PyTorch framework Nerfstudio [20]. We build upon its core method named Nerfacto, which combines ideas from several papers for fast and qualitative renders of unbounded complex scenes. This makes it an accessible baseline with a state-of-the-art quality/time ratio. Because NeRF pipelines contain a high number of small but decisive choices of implementation (eg. some frameworks choose to train their models image by image while Nerfstudio jointly and randomly samples across all images), we decided to use as much as possible Nerfstudio-based baselines for fairer comparisons. All of these models are trained on one Nvidia-A100 GPU. The reported times correspond to the approximated training duration of the models. Configuration details, further experiments and ablatives on our contributions can be found in the supplementary materials." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 595, + 362, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 595, + 362, + 608 + ], + "spans": [ + { + "bbox": [ + 132, + 595, + 362, + 608 + ], + "type": "text", + "content": "4.2 Novel View Synthesis and Anti-Aliasing" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "This experiment aims to evaluate the reconstruction quality through the ability to synthesize viewpoints that are not supervised during the reconstruction. These new viewpoints differ from the angle of observation, but also from the distance of observation. The latter is particularly important since a reconstruction or a" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "text", + "content": "RING-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 157, + 480, + 205 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Novel View Synthesis performances for the Mono-Scale setup (trained on the full resolution images only) on the 360 Dataset. The indicated resolutions refer to the resolution of the renders." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 157, + 480, + 205 + ], + "lines": [ + { + "bbox": [ + 132, + 157, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 157, + 480, + 205 + ], + "type": "table", + "html": "
Full Res.1/2 Res.1/4 Res.1/8 Res.Time ↓
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Nerfacto [20]27.090.7790.18126.810.7820.16225.220.7110.23423.320.6360.2970.45h
PyNeRF [21]27.870.8020.16021.930.7130.21121.150.6620.26420.390.6190.3120.96h
Zip-NeRF [4]28.060.8080.15416.580.5960.31911.880.4240.4659.660.3230.5231.10h
RING-NeRF28.090.7990.15727.180.8040.13825.820.7860.14224.380.7370.1670.45h
", + "image_path": "7ca447ece0bbf99aef4a60643c13eb96a88b68e2850a5030e0555e3c0bd4fb90.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 219, + 479, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 219, + 479, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 219, + 479, + 266 + ], + "type": "text", + "content": "rendering process that does not take correctly into account the distance of observation leads to artefacts, from over-contrasted rendering to aliasing phenomenon. The challenge is to avoid these artefacts while keeping the reconstruction and rendering process as fast as possible." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 266, + 480, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 480, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 480, + 350 + ], + "type": "text", + "content": "Dataset. The evaluation relies on the dataset introduced by Mip-NeRF-360 [3]. This dataset is composed of 9 scenes, each containing both a central area and complex background in both inside and outside setups. Since the trajectory keeps a constant distance to the central part, each viewpoint is represented with a pyramid of 4 different image resolutions to simulate a variation of the distance of observation combined with a change of the camera FOV, following the Mip-NeRF [2] original anti-aliasing evaluation pipeline." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 350, + 480, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 480, + 386 + ], + "type": "text", + "content": "Algorithms. We compare our solution against several grid-based NeRF baselines, both with (PyNeRF [21] and Zip-NeRF [4]) and without anti-aliasing processing (Nerfacto), using their Nerfstudio implementations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 387, + 480, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 480, + 445 + ], + "type": "text", + "content": "Protocol. For each scene, we train the models with two different setups: the mono-scale setup that uses only the image with the highest resolution for each viewpoint, and the multi-scale setup which uses the whole pyramid of images for each viewpoint. Note that, for these two setups, we evaluate the performances on the whole resolution pyramid, with usual metrics (PSNR, SSIM, LPIPS)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 446, + 481, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 481, + 542 + ], + "type": "text", + "content": "Results. First of all, regarding novel view synthesis quality using the monoscale setup for both training and testing, as referred to in the \"Full Res.\" column of table 1, we notice that our architecture provides on par performances with Zip-NeRF, slightly better results than PyNeRF, and a more important gap with Nerfacto. While simple, RING-NeRF succeeds in performing state-of-the-art performances on a real single-scale dataset. Regarding quality for the multi-scale setup, as presented in table 2, we observe that all the algorithms that consider the distance of observation perform very similarly in terms of quality. On the" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 135, + 556, + 201, + 613 + ], + "blocks": [ + { + "bbox": [ + 135, + 556, + 201, + 613 + ], + "lines": [ + { + "bbox": [ + 135, + 556, + 201, + 613 + ], + "spans": [ + { + "bbox": [ + 135, + 556, + 201, + 613 + ], + "type": "image", + "image_path": "4e1ab55963528c40ffc68316a7943fac1db4d2db9816f2944fd16e7c8608c3f7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 624, + 481, + 658 + ], + "lines": [ + { + "bbox": [ + 130, + 624, + 481, + 658 + ], + "spans": [ + { + "bbox": [ + 130, + 624, + 481, + 658 + ], + "type": "text", + "content": "Fig. 4: Examples of image renderings at 1/8th resolution from models solely trained with the full resolution images. RING-NeRF is the only method capable of producing coherent aliasing-free renderings thanks to its LOD inductive bias." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 203, + 556, + 269, + 613 + ], + "blocks": [ + { + "bbox": [ + 203, + 556, + 269, + 613 + ], + "lines": [ + { + "bbox": [ + 203, + 556, + 269, + 613 + ], + "spans": [ + { + "bbox": [ + 203, + 556, + 269, + 613 + ], + "type": "image", + "image_path": "e10b9442e45182a7f15afc5fefbc310cee93cb4f1265391879c441bf6feed875.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 271, + 556, + 340, + 613 + ], + "blocks": [ + { + "bbox": [ + 271, + 556, + 340, + 613 + ], + "lines": [ + { + "bbox": [ + 271, + 556, + 340, + 613 + ], + "spans": [ + { + "bbox": [ + 271, + 556, + 340, + 613 + ], + "type": "image", + "image_path": "9b772adcc99b6dacb6a804e24fcf7d7855c02a82ccf420e8c917ed836757b4fc.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 341, + 556, + 408, + 613 + ], + "blocks": [ + { + "bbox": [ + 341, + 556, + 408, + 613 + ], + "lines": [ + { + "bbox": [ + 341, + 556, + 408, + 613 + ], + "spans": [ + { + "bbox": [ + 341, + 556, + 408, + 613 + ], + "type": "image", + "image_path": "9c58a02448d7a3bd308a112f08b9f77255d167ded6ad2008fc4ec33aa458832b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 410, + 556, + 479, + 613 + ], + "blocks": [ + { + "bbox": [ + 410, + 556, + 479, + 613 + ], + "lines": [ + { + "bbox": [ + 410, + 556, + 479, + 613 + ], + "spans": [ + { + "bbox": [ + 410, + 556, + 479, + 613 + ], + "type": "image", + "image_path": "08e2a35d1620d5f24568c3546003dbfd1345c07123da38c82a92fc70e56101fd.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "D. Petit et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 157, + 479, + 205 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 479, + 147 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 479, + 147 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 479, + 147 + ], + "type": "text", + "content": "Table 2: Novel View Synthesis performances for the Multi-Scale setup (trained jointly on every resolution) on the 360 Dataset. The indicated resolutions refer to the resolution of the renders." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 157, + 479, + 205 + ], + "lines": [ + { + "bbox": [ + 132, + 157, + 479, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 157, + 479, + 205 + ], + "type": "table", + "html": "
Full Res.1/2 Res.1/4 Res.1/8 Res.Time ↓
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Nerfacto [20]25.980.6680.36727.310.7920.17627.160.8050.13925.300.7430.1990.45h
PyNeRF [21]27.650.7810.19429.470.8570.09230.510.8870.06330.860.8930.0550.96h
Zip-NeRF [4]27.540.7810.19129.340.8580.09930.360.8890.06430.900.9000.0471.10h
RING-NeRF27.660.7820.19429.550.8560.09430.510.8880.06230.860.9010.0480.45h
", + "image_path": "aaa74f02bea0e77ff08b7095c7dafc86f28b1bf9b737ea8c274863ad0f625bc3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 218, + 479, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 218, + 479, + 240 + ], + "spans": [ + { + "bbox": [ + 130, + 218, + 479, + 240 + ], + "type": "text", + "content": "other side, Nerfacto performs quite poorly. Its performances are especially low at the coarsest resolution, with overflowing artefacts, as illustrated in figure 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 242, + 480, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 242, + 480, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 242, + 480, + 433 + ], + "type": "text", + "content": "Finally, we evaluate the capacity of the models to generalize over new resolutions by training them on full-resolution images and then evaluate them on smaller resolutions. Qualitative results are shown in figure 4 while quantitative results can be found in the 1/2-th, 1/4-th and 1/8-th res. columns of table 1. Because most anti-aliasing methods need LOD-specific supervision to correctly function (including Zip-NeRF and PyNeRF), they cannot render coherent anti-aliased images in this setup. However, PyNeRF behaves better than ZipNeRF with coherent although very aliased renderings, as it decides to limit the LOD used for rendering inside the range of LOD seen during training. RING-NeRF, with his LOD inductive bias, is the only architecture capable of producing anti-aliased renderings from novel observation distances and thus outperforms every other method. While this experiment can seem somewhat esoteric as training on a multi-resolution images pyramid is rather easy, this increases GPU memory and total training time. Moreover, depending on the resolution and the trajectory, it is not trivial to choose the accurate number of scales in the image pyramid to supervise correctly every grid in the hierarchy and especially the coarsest grids." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 434, + 480, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 480, + 505 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 480, + 505 + ], + "type": "text", + "content": "Regarding the reconstruction processing time reported in the \"Time\" column of both table 1 and table 2, since RING-NeRF does not rely on the multiplication of either sample or decoder, it processes as quickly as the fastest Ner-facto both on mono-scale and multi-scale setups. On the other hand, the other anti-aliasing methods, PyNeRF with its per-LOD MLP and Zip-NeRF with its super-sampling, are approximately 2.5 times slower." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 506, + 480, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 480, + 565 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 480, + 565 + ], + "type": "text", + "content": "In conclusion, our solution provides the best quality-speed trade-off since it is both on par with the best quality method and the fastest method, in monoscale as well as in multi-scale setups. Furthermore, RING-NeRF is the only solution capable of creating coherent and anti-aliased renderings when facing novel observation distances unseen during training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 584, + 302, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 584, + 302, + 596 + ], + "spans": [ + { + "bbox": [ + 131, + 584, + 302, + 596 + ], + "type": "text", + "content": "4.3 Few Viewpoints Supervision" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 479, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 479, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 479, + 628 + ], + "type": "text", + "content": "This experiment aims to evaluate the influence of the RING-NERF architecture and pipeline on the reconstruction robustness to limited supervision viewpoints." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 630, + 479, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 479, + 652 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 479, + 652 + ], + "type": "text", + "content": "Dataset. We evaluate our contribution on the object-centric real dataset DTU [9], often used in few-viewpoints evaluations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 653, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 653, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 653, + 479, + 665 + ], + "type": "text", + "content": "Algorithms. We compare our architecture against several baselines: Mip-NeRF" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "text", + "content": "RING-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "text", + "content": "[2], and FreeNerf [25] (a state-of-the-art method for this task), for vanilla architectures and Nerfacto for grid-based architectures. To better demonstrate the intrinsic stability brought by RING-NeRF, we also developed the Nerfacto+ architecture, which corresponds to a Nerfacto architecture coupled with a coarse-to-fine training based on a progressive activation of the grids (the rest of the decoder's input being filled with zeros). For Nerfacto+ and RING-NeRF, we also add FreeNeRF's loss that penalizes the density of the first " + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "inline_equation", + "content": "M = 10" + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "text", + "content": " samples of each ray to reduce as much as possible artefacts in front of the cameras. As an ablative experiment, we also evaluate RING-NeRF using discrete coarse-to-fine (fixed LOD increment of 1 rather than the proposed continuous increase)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 284 + ], + "type": "text", + "content": "Protocol. We follow FreeNeRF's evaluation pipeline, including the number of supervision viewpoints (3 to 9), the choice of these views among the dataset and the evaluations using masks of the object. Since we are using the same protocol, the results of FreeNeRF and Mip-NeRF were taken out of FreeNeRF's article." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 284, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 536 + ], + "type": "text", + "content": "Results. Evaluation results are reported in table 3. We first notice an important difference between vanilla and grid-based baselines. While Mip-NeRF faces troubles in reconstructing the scene with 3 views, the method seems to find coherency when adding more images. However, Nerfacto struggles much more to form a consistent 3D scene even when using 9 images (see figure 1). Even though the grid-based baseline is way faster to train, its design implies more instability when facing a small number of images. This does not mean however that few viewpoints are incompatible with grid-based methods. Using progressive training coupled with a very simple and generalizable regularization, both Nerfacto+ and our architecture succeed in creating coherent geometry. Nonetheless, RING-NeRF considerably outperforms Nerfacto+, with a PSNR difference varying from 3 to 4, demonstrating the stability increase brought by our architecture, and also outperforms Mip-NeRF for the configuration with 3 and 6 supervision images. The discrete coarse-to-fine version of RING-NeRF performs in-between Nerfacto+ and the complete RING-NeRF. This showcases both the intrinsic interest of the proposed architecture against the Nerfacto+ and the relevance of the continuous coarse-to-fine mechanism. FreeNerf remains the best performer of all in terms of quality, but with a reconstruction time that is extremely slower than RING-NeRF, the former achieving its reconstruction in 2.56 hours while the latter only requires less than 10 minutes. RING-NeRF thus offers a better quality-speed trade-off for the few-view reconstruction issue (see figure 1)." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 149, + 579, + 466, + 665 + ], + "blocks": [ + { + "bbox": [ + 130, + 547, + 479, + 571 + ], + "lines": [ + { + "bbox": [ + 130, + 547, + 479, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 547, + 479, + 571 + ], + "type": "text", + "content": "Table 3: Performances of reconstruction from few viewpoints on the DTU dataset. The reported metrics are computed based on the mask of the object." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 579, + 466, + 665 + ], + "lines": [ + { + "bbox": [ + 149, + 579, + 466, + 665 + ], + "spans": [ + { + "bbox": [ + 149, + 579, + 466, + 665 + ], + "type": "table", + "html": "
#ImagesPSNR ↑SSIM ↑LPIPS ↓Time ↓
369369369
Mip-NeRF [2]8.6816.5423.580.5710.7410.8790.3530.1980.0922.56h
FreeNeRF [25]19.9223.2525.380.7870.8440.8880.1350.0950.0672.56h
Nerfacto [20]9.359.759.780.5670.6040.6470.3850.3310.3260.15h
Nerfacto+13.6116.6119.330.6390.6990.7590.2760.2180.1510.15h
RING-NeRF16.1820.4723.190.7130.8080.8470.2000.1270.0850.15h
w/ discrete CtF15.7920.1622.930.7060.7850.8470.2010.1270.0850.15h
", + "image_path": "65d5f8bca8f22a6197d04fca006c945742a652685282a0fb7cc3a9072703fb6b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "D. Petit et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 184, + 146, + 430, + 184 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 4: SDF reconstruction performances when foregoing the scene-specific SDF Initialization on the Replica Dataset. The Chamfer distance is in centimeters." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 184, + 146, + 430, + 184 + ], + "lines": [ + { + "bbox": [ + 184, + 146, + 430, + 184 + ], + "spans": [ + { + "bbox": [ + 184, + 146, + 430, + 184 + ], + "type": "table", + "html": "
PSNR ↑SSIM ↑LPIPS ↓Chamfer-L1 ↓Training Time ↓
NeuS-facto [26]24.620.7780.34717.611.4h
NeuralAngelo [11]30.790.9160.076113.693.40 h
RING-NeRF37.180.9690.01945.711.28 h
", + "image_path": "7049b4c3facffb648f9dab42ba4e216fd0364798d6c1a1599c7d83699fa23156.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 198, + 373, + 210 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 198, + 373, + 210 + ], + "spans": [ + { + "bbox": [ + 132, + 198, + 373, + 210 + ], + "type": "text", + "content": "4.4 SDF Reconstruction without Initialization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 215, + 480, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 480, + 286 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 480, + 286 + ], + "type": "text", + "content": "SDF reconstruction is known to be a more unstable process than density-based NeRF [1], requiring a scene-specific initialization to converge. This initialization becomes an issue in complex environments and incremental setups, where several types of scenes can co-exist and are not necessarily known beforehand. Therefore, in this experiment, we evaluate the ability of RING-NeRF and other SotA architectures to achieve SDF reconstruction without scene-specific initialization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 287, + 479, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 287, + 479, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 287, + 479, + 323 + ], + "type": "text", + "content": "Dataset. The evaluation is achieved on a subset of 7 scenes of the Replica [18] synthetic indoor dataset. A Tanks & Temples [10] example is provided in supplementary materials with corresponding analysis." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 323, + 480, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 480, + 394 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 480, + 394 + ], + "type": "text", + "content": "Algorithms. We compare our architecture to two SDF methods, all of them implemented in the same SDFStudio [26] branch of the Nerfstudio framework for fairer comparisons: NeuS-facto, an adaptation of NeuS for grid-based methods with Nerfacto modules, and an implementation of NeuralAngelo. For these experiments, our model is built upon the NeuS-facto baseline, using in particular the same NeuS-based SDF-to-density transformation [22]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 395, + 480, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 395, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 395, + 480, + 430 + ], + "type": "text", + "content": "Protocol. To evaluate the impact of the architecture and pipeline over the convergence and stability, we suppress the inverted sphere SDF initialization scheme and use a random initialization for the model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 430, + 481, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 481, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 481, + 550 + ], + "type": "text", + "content": "Results. Evaluation results are shown in table 4. First of all, NeuS-facto faces low rendering and reconstruction metrics, due to catastrophic failure in most of the tested scenes (see figure 1) since the model tends to re-draw the 2D images in front of the camera. Regarding NeuralAngelo, its relatively high PSNR demonstrates its ability to synthesize satisfying RGB. However, as illustrated in figure 1 and highlighted by the reconstruction metrics, the underlying geometry of the scene is poorly reconstructed, without any fine details. Finally, our method RING-NeRF is by far the best performer, with much higher PSNR and a better geometry including fine details (see figure 1), although a bit noisy. The simple architecture of RING-NeRF also permits faster epochs, thus faster training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 564, + 253, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 564, + 253, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 564, + 253, + 577 + ], + "type": "text", + "content": "4.5 LOD Extensibility" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 582, + 479, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 479, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 479, + 605 + ], + "type": "text", + "content": "This experiment aims to demonstrate RING-NeRF's unique ability to increase dynamically the level of detail of the scene representation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 605, + 448, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 448, + 616 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 448, + 616 + ], + "type": "text", + "content": "Dataset. The scan 114 of the DTU dataset is used for this experiment." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 479, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 479, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 479, + 641 + ], + "type": "text", + "content": "Algorithms. Because I-NGP architectures [4, 11, 14] cannot perform LOD extensibility (due to the fixed decoder's input size), only RING-NeRF is evaluated." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Protocol. We train our model using two different configurations : one low resolution with a 3 levels grid hierarchy and one high resolution with 5 levels (the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "text", + "content": "RING-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 116, + 321, + 228 + ], + "blocks": [ + { + "bbox": [ + 138, + 116, + 321, + 228 + ], + "lines": [ + { + "bbox": [ + 138, + 116, + 321, + 228 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 321, + 228 + ], + "type": "image", + "image_path": "4eaf320a4870d993ff70708bcd7991e489b3dd4205e12abfd74d6e489faf691f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 239, + 481, + 264 + ], + "lines": [ + { + "bbox": [ + 131, + 239, + 481, + 264 + ], + "spans": [ + { + "bbox": [ + 131, + 239, + 481, + 264 + ], + "type": "text", + "content": "Fig. 5: Learning curves and final renderings of RING-NeRF models with different grid configurations trained either jointly or incrementally." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 323, + 116, + 396, + 171 + ], + "blocks": [ + { + "bbox": [ + 323, + 116, + 396, + 171 + ], + "lines": [ + { + "bbox": [ + 323, + 116, + 396, + 171 + ], + "spans": [ + { + "bbox": [ + 323, + 116, + 396, + 171 + ], + "type": "image", + "image_path": "f5bcb97c15a12b8e9aa96373f6fd3402f54633e4ca1063c0c5d15018049793c9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 399, + 116, + 473, + 171 + ], + "blocks": [ + { + "bbox": [ + 399, + 116, + 473, + 171 + ], + "lines": [ + { + "bbox": [ + 399, + 116, + 473, + 171 + ], + "spans": [ + { + "bbox": [ + 399, + 116, + 473, + 171 + ], + "type": "image", + "image_path": "d0630a3dc21b8411ce3938d4ede0f764a4739c201d41375546cfe8f239a71d99.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 323, + 173, + 397, + 228 + ], + "blocks": [ + { + "bbox": [ + 323, + 173, + 397, + 228 + ], + "lines": [ + { + "bbox": [ + 323, + 173, + 397, + 228 + ], + "spans": [ + { + "bbox": [ + 323, + 173, + 397, + 228 + ], + "type": "image", + "image_path": "b638c4748cd31e44ad39d040f95fb7602f257637aead1d88f5d67aa4b382952d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 399, + 173, + 473, + 228 + ], + "blocks": [ + { + "bbox": [ + 399, + 173, + 473, + 228 + ], + "lines": [ + { + "bbox": [ + 399, + 173, + 473, + 228 + ], + "spans": [ + { + "bbox": [ + 399, + 173, + 473, + 228 + ], + "type": "image", + "image_path": "c6716be8e8b0c7f9a854e46f636030b993ff8f8586f0e5917d978598e0117d5f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 277, + 482, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 277, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 131, + 277, + 482, + 361 + ], + "type": "text", + "content": "grid resolutions are 16, 32, 64, 128 and 256). For these two configurations named \"3 levels\" and \"5 levels\", every grid and the decoder are trained simultaneously. We proceed to showcase the extensibility of RING-NeRF by adding grids to the \"3 levels\" configuration that is previously trained. We first train one grid of resolutions 128 (\"3+1 levels\" in Figure 5) with the three initial grids and the decoder frozen and then train another grid of resolutions 256 (\"3+1+1 levels\" in Figures 5) with the four grids and the decoder frozen." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 361, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 361, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 131, + 361, + 482, + 445 + ], + "type": "text", + "content": "Results. Figure 5 shows that the configuration \"3+1+1 levels\" results in the same rendering quality than the \"5 levels\" one. This demonstrates the ability of our model to dynamically change the resolution of the grid hierarchy. This is an important step towards the development of an adaptive architecture which locally chooses the resolution of the representation based on the scene's content. This allows to drastically reduce the number of parameters, helping in improving the memory footprint, the training duration and the model's robustness." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 473, + 324, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 473, + 324, + 487 + ], + "spans": [ + { + "bbox": [ + 132, + 473, + 324, + 487 + ], + "type": "text", + "content": "5 Conclusion and Perspectives" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 508, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 482, + 616 + ], + "type": "text", + "content": "In this work, we introduced RING-NeRF, a simple and versatile NeRF pipeline that provides two inductive biases by design: a continuous multi-scale representation of the scene, and an invariance of the decoder latent space over spatial and scale domains. Coupled with a distance-aware forward mapping and a continuous coarse-to-fine reconstruction process, our pipeline demonstrated experimentally its versatility with on-par performances with dedicated state-of-the-art solutions for anti-aliasing or reconstruction from few viewpoints. It even outperforms them in terms of robustness to scene-specific initialization for SDF reconstruction. Furthermore, it is highly efficient and is not limited to object-centric scenes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Future work will study the impact of RING-NeRF on other challenging use cases, such as facing inaccurate camera poses [16] and SLAM [29]. We will also use the extensibility property of our architecture to develop memory-efficient sparse Neural Fields, which is considered to be a limit of most grid-based models." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "D. Petit et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 479, + 163 + ], + "type": "text", + "content": "This publication was made possible by the use of the CEA List FactoryIA supercomputer, financially supported by the Ile-de-France Regional Council." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 205, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 205, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 480, + 239 + ], + "type": "text", + "content": "1. Atzmon, M., Lipman, Y.: Sal: Sign agnostic learning of shapes from raw data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2565-2574 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 239, + 480, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 480, + 282 + ], + "type": "text", + "content": "2. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 284, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 284, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 138, + 284, + 480, + 326 + ], + "type": "text", + "content": "3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mipnerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 327, + 480, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 327, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 138, + 327, + 480, + 348 + ], + "type": "text", + "content": "4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-nerf: Anti-aliased grid-based neural radiance fields. ICCV (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 349, + 480, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 349, + 480, + 370 + ], + "spans": [ + { + "bbox": [ + 138, + 349, + 480, + 370 + ], + "type": "text", + "content": "5. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: European Conference on Computer Vision (ECCV) (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 371, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 371, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 371, + 480, + 403 + ], + "type": "text", + "content": "6. Dou, Y., Zheng, Z., Jin, Q., Ni, B.: Multiplicative fourier level of detail. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1808-1817 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 403, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 480, + 446 + ], + "type": "text", + "content": "7. Hu, W., Wang, Y., Ma, L., Yang, B., Gao, L., Liu, X., Ma, Y.: Tri-miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19774-19783 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "type": "text", + "content": "8. Jain, A., Tancik, M., Abbeel, P.: Putting nerf on a diet: Semantically consistent few-shot view synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5885-5894 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 480, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 480, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 480, + 480, + 512 + ], + "type": "text", + "content": "9. Jensen, R., Dahl, A., Vogiatzis, G., Tola, E., Aanaes, H.: Large scale multi-view stereopsis evaluation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 406-413 (2014)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 513, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 513, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 513, + 480, + 533 + ], + "type": "text", + "content": "0. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics 36(4) (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 534, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 534, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 134, + 534, + 480, + 578 + ], + "type": "text", + "content": "11. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8456-8465 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 578, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 578, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 134, + 578, + 480, + 599 + ], + "type": "text", + "content": "12. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 480, + 632 + ], + "type": "text", + "content": "13. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "text", + "content": "14. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "text", + "content": "RING-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 658 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "15. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5480-5490 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 481, + 193 + ], + "type": "text", + "content": "16. Park, K., Henzler, P., Mildenhall, B., Barron, J.T., Martin-Brualla, R.: Camp: Camera preconditioning for neural radiance fields. ACM Transactions on Graphics (TOG) 42(6), 1-11 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 193, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 193, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 193, + 481, + 214 + ], + "type": "text", + "content": "17. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. ICCV (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 481, + 247 + ], + "type": "text", + "content": "18. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 247, + 481, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 247, + 481, + 290 + ], + "spans": [ + { + "bbox": [ + 133, + 247, + 481, + 290 + ], + "type": "text", + "content": "19. Takikawa, T., Litalien, J., Yin, K., Kreis, K., Loop, C., Nowrouzezahrai, D., Jacobson, A., McGuire, M., Fidler, S.: Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11358-11367 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 290, + 481, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 290, + 481, + 334 + ], + "spans": [ + { + "bbox": [ + 132, + 290, + 481, + 334 + ], + "type": "text", + "content": "20. Tancik, M., Weber, E., Ng, E., Li, R., Yi, B., Wang, T., Kristoffersen, A., Austin, J., Salahi, K., Ahuja, A., et al.: Nerfstudio: A modular framework for neural radiance field development. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-12 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 334, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 334, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 334, + 481, + 365 + ], + "type": "text", + "content": "21. Turki, H., Zollhöfer, M., Richardt, C., Ramanan, D.: Pynerf: Pyramidal neural radiance fields. In: Thirty-Seventh Conference on Neural Information Processing Systems (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 365, + 481, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 481, + 398 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 481, + 398 + ], + "type": "text", + "content": "22. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 398, + 481, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 398, + 481, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 398, + 481, + 419 + ], + "type": "text", + "content": "23. Wynn, J., Turmukhambetov, D.: DiffusioNeRF: Regularizing Neural Radiance Fields with Denoising Diffusion Models. In: CVPR (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 419, + 481, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 481, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 481, + 463 + ], + "type": "text", + "content": "24. Xu, L., Agrawal, V., Laney, W., Garcia, T., Bansal, A., Kim, C., Rota Bulò, S., Porzi, L., Kontschieder, P., Božić, A., Lin, D., Zollhöfer, M., Richardt, C.: VR-NeRF: High-fidelity virtualized walkable spaces. In: SIGGRAPH Asia Conference Proceedings (2023). https://doi.org/10.1145/3610548.3618139" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 463, + 481, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 463, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 463, + 481, + 495 + ], + "type": "text", + "content": "25. Yang, J., Pavone, M., Wang, Y.: Freenerf: Improving few-shot neural rendering with free frequency regularization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8254-8263 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 495, + 481, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 495, + 481, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 495, + 481, + 528 + ], + "type": "text", + "content": "26. Yu, Z., Chen, A., Antic, B., Peng, S., Bhattacharyya, A., Niemeyer, M., Tang, S., Sattler, T., Geiger, A.: Sdfstudio: A unified framework for surface reconstruction (2022), https://github.com/autonomousvision/sdfstudio" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 528, + 481, + 560 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 528, + 481, + 560 + ], + "spans": [ + { + "bbox": [ + 132, + 528, + 481, + 560 + ], + "type": "text", + "content": "27. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems 35, 25018-25032 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 560, + 481, + 582 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 560, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 560, + 481, + 582 + ], + "type": "text", + "content": "28. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 582, + 481, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 582, + 481, + 625 + ], + "spans": [ + { + "bbox": [ + 132, + 582, + 481, + 625 + ], + "type": "text", + "content": "29. Zhu, Z., Peng, S., Larsson, V., Xu, W., Bao, H., Cui, Z., Oswald, M.R., Pollefeys, M.: Nice-slam: Neural implicit scalable encoding for slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12786-12796 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 625, + 481, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 625, + 481, + 658 + ], + "spans": [ + { + "bbox": [ + 132, + 625, + 481, + 658 + ], + "type": "text", + "content": "30. Zhuang, Y., Zhang, Q., Feng, Y., Zhu, H., Yao, Y., Li, X., Cao, Y.P., Shan, Y., Cao, X.: Anti-aliased neural implicit surfaces with encoding level of detail. Siggraph Asia 2023 (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "D. Petit et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/RISurConv_ Rotation Invariant Surface Attention-Augmented Convolutions for 3D Point Cloud Classification and Segmentation/85a1ebc6-6b7d-4e16-8cd8-c420684dc80d_content_list.json b/2024/RISurConv_ Rotation Invariant Surface Attention-Augmented Convolutions for 3D Point Cloud Classification and Segmentation/85a1ebc6-6b7d-4e16-8cd8-c420684dc80d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2a36e42ffd751fd895855f3799742397fb51ab5c --- /dev/null +++ b/2024/RISurConv_ Rotation Invariant Surface Attention-Augmented Convolutions for 3D Point Cloud Classification and Segmentation/85a1ebc6-6b7d-4e16-8cd8-c420684dc80d_content_list.json @@ -0,0 +1,1753 @@ +[ + { + "type": "text", + "text": "RISurConv: Rotation Invariant Surface Attention-Augmented Convolutions for 3D Point Cloud Classification and Segmentation", + "text_level": 1, + "bbox": [ + 223, + 140, + 781, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhiyuan Zhang $^{1}$ , Licheng Yang $^{2\\star}$ , and Zhiyu Xiang $^{2\\oplus}$", + "bbox": [ + 299, + 234, + 702, + 252 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{1}$ School of Computing and Information Systems, Singapore Management University https://github.com/cszyzhang/RISurConv", + "$^{2}$ College of Information Science and Electronic Engineering, Zhejiang University" + ], + "bbox": [ + 217, + 261, + 781, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Despite the progress on 3D point cloud deep learning, most prior works focus on learning features that are invariant to translation and point permutation, and very limited efforts have been devoted for rotation invariant property. Several recent studies achieve rotation invariance at the cost of lower accuracies. In this work, we close this gap by proposing a novel yet effective rotation invariant architecture for 3D point cloud classification and segmentation. Instead of traditional pointwise operations, we construct local triangle surfaces to capture more detailed surface structure, based on which we can extract highly expressive rotation invariant surface properties which are then integrated into an attention-augmented convolution operator named RISurConv to generate refined attention features via self-attention layers. Based on RISurConv we build an effective neural network for 3D point cloud analysis that is invariant to arbitrary rotations while maintaining high accuracy. We verify the performance on various benchmarks with supreme results obtained surpassing the previous state-of-the-art by a large margin. We achieve an overall accuracy of $96.0\\%$ $(+4.7\\%)$ on ModelNet40, $93.1\\%$ $(+12.8\\%)$ on ScanObjectNN, and class accuracies of $91.5\\%$ $(+3.6\\%)$ , $82.7\\%$ $(+5.1\\%)$ , and $78.5\\%$ $(+9.2\\%)$ on the three categories of the FG3D dataset for the fine-grained classification task. Additionally, we achieve $81.5\\%$ $(+1.0\\%)$ mIoU on ShapeNet for the segmentation task.", + "bbox": [ + 259, + 338, + 743, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Point cloud $\\cdot$ Rotation invariant $\\cdot$ Attention", + "bbox": [ + 261, + 642, + 643, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 681, + 375, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Point cloud has become the most promising 3D data representation for a wide range of immersive applications from robot navigation to autonomous driving. The increasing availability of 3D sensors and the emergence of large and mid-scale point cloud datasets [1,2,12,19,31,35] have spurred significant research in this area, leading to the development of numerous deep learning models for point cloud classification and segmentation, which are the fundamentals for various computer vision tasks.", + "bbox": [ + 212, + 710, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "\\* co-first author.", + "bbox": [ + 217, + 825, + 333, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, analyzing 3D point clouds remains challenging, mainly due to the irregular nature of point clouds and their inherent invariances such as translation, point permutation, and rotation. While significant progress has been made in learning translation and permutation-invariant features [17, 23, 25, 26, 44], achieving rotation invariance in point cloud convolution has been a relatively unexplored area.", + "bbox": [ + 212, + 146, + 787, + 234 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Rotation invariance is essential for 3D object classification and segmentation as objects can be viewed from different viewpoints, leading to variations in their orientation. Prior approaches [23,25,33] usually rely on rotation augmentation in training stage to relief the rotation affections for testing. However, such scheme is less effective as 3D data has more degrees of freedom. A rotation invariant model is, therefore, critical to accurately classify and segment objects in 3D point clouds.", + "bbox": [ + 212, + 237, + 787, + 343 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Various attempts have been made to address the problem of rotation invariance in 3D point clouds, as evidenced by several works [3,4,22,27,41]. These efforts have focused on designing rotation invariant properties to achieve consistent accuracies under arbitrary rotations without resorting to rotation augmentation. However, these methods have demonstrated inferior performance when compared to translation invariant approaches, primarily due to the loss of global information during the generation of rotation invariant properties. Recent research efforts [15,30,40] have attempted to overcome this limitation by employing local reference frame (LRF) or local reference axis (LRA) [42] to transform the data into a canonical coordinate system and encode global information. Despite the improvements, their accuracies are still lower than those of state-of-the-art non-rotation invariant methods because of the unstable LRF/LRA and the less descriptive rotation invariant properties that the useful surface information is not well preserved.", + "bbox": [ + 212, + 343, + 787, + 554 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above-mentioned issues, we propose a novel yet effective rotation invariant architecture. Specifically, we construct local triangle surfaces for each reference point of the input data to better capture the local surface structure. On the local surfaces, we design highly expressive rotation invariant surface properties which are then integrated into an attention-augmented convolution operator named RISurConv to extract refined rotation invariant features. Finally, we build up network based on RISurConv for rotation invariant object classification and segmentation. In summary, our main contributions include:", + "bbox": [ + 212, + 555, + 787, + 676 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Rotation Invariant Surface Properties. We construct local triangle surfaces from the reference point and its neighbors, based on which we are able to design highly expressive rotation invariant surface properties;", + "- RISurConv. We integrate the Rotation Invariant Surface Properties into an attention-augmented architecture named RISurConv comprising two self-attention layers to learn and generate refined rotation invariant features;", + "- Extensive experiments on a variety of classification and segmentation tasks. Our approach shows supreme performance, surpassing the state-of-the-art by a large margin under challenging rotation scenarios including an analysis of rotation-invariant features and an ablation study of our neural" + ], + "bbox": [ + 225, + 688, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z. Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "network to provide insights into the key factors contributing to the exceptional performance.", + "bbox": [ + 241, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 200, + 395, + 217 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we review the representative works in 3D point cloud deep learning and rotation invariant learning.", + "bbox": [ + 212, + 234, + 782, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Point Cloud Deep Learning. 3D point clouds is a more compact and intuitive representation for feature learning. PointNet [23] pioneered a point cloud convolution with global features by max-pooling per-point features from MLPs, and follow-up works [17, 25, 38] have focused on exploring convolution kernels that exploit geometric features [29], adding edges on top of points [33], parameterizing convolution using polynomials [38], and leveraging shape context [37]. Some methods are designed to combine with recurrent neural networks [13] and sequence models [18].", + "bbox": [ + 212, + 265, + 784, + 385 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In recent years, attention mechanism has become increasingly popular in various natural language processing (NLP) [32] and computer vision [6] tasks. With their ability to handle sequential and spatial information, attention mechanism has also shown promise in 3D point cloud processing tasks. Pioneer works that apply attention mechanism in point clouds include Point Transformer [43] and PCT [9], which use a self-attention mechanism to learn local and global features of point clouds. Other works have explored different types of transformers. Inspired by BERT [5], Point-BERT [39] was proposed to pre-train pure Transformer-based models with a Mask Point Modeling task for point cloud classification. Dual Transformer Network [11] aggregated the point-wise and channel-wise multi-head self-attention models simultaneously such that long-range context dependencies can be captured by investigating the point-wise and channel-wise relationships. Overall, these developments show promise in enhancing the efficiency and accuracy of point cloud analysis for a variety of applications. Readers can refer to the survey [10] for a comprehensive overview of 3D point cloud deep learning. However, most of these methods lack the property of rotation invariance which is critical for object classification and segmentation as the object can be viewed at any angle in reality. To address this issue, a common approach is to increase the training data by augmenting it with arbitrary rotations [17, 25, 33]. However, such approach has difficulty in generalizing predictions to unseen rotations, resulting in deteriorated performance. Therefore, it would be desirable to design a specific convolution possessing rotation-invariant features.", + "bbox": [ + 212, + 387, + 785, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Rotation Invariant Learning. Various methods have been proposed to address the issue of rotation invariance in feature learning of point clouds. For instance, Rao et al. [27] used a spherical domain to define a rotation-invariant convolution for point clouds. However, the discretized sphere is sensitive to global rotations, which can lead to a notable drop in performance for objects with arbitrary rotations. To overcome this limitation, Poulenard et al. [22] integrated spherical harmonics to their convolution. Similarly, Chen et al. [3] proposed a", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "RISurConv", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bbfdb259e870195d441844ef27250a76a3317bffbe790e18e1aa53036e6cd5e4.jpg", + "image_caption": [ + "Fig. 1: Rotation Invariant Surface Property (RISP) construction: Given a point $p$ as the reference point, $K$ ( $K = 8$ in this example) nearest points are selected (middle). For each neighbor $x_{i}$ , two adjacent neighbors $x_{i-1}$ and $x_{i+1}$ are used to form two triangular local surfaces (right), based on which rotation invariant properties are constructed." + ], + "image_footnote": [], + "bbox": [ + 259, + 142, + 743, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "hierarchical clustering scheme to encode the relative angles between two-point vectors and used vector norm to maintain rotation invariance. Zhang et al. [41] presented a simple convolution named RICnv that operates on handcrafted features built from Euclidean distances and angles that are rotation invariant by nature. However, this approach only considers local features resulting in accuracy degradation. GCACnv [40] addressed this limitation by building a global context-aware convolution based on anchors and Local Reference Frame (LRF) to achieve rotation invariance. RI-GCN [15] learned rotation-invariant local descriptors and applied graph convolutional neural networks to aggregate local features based on LRF. Thomas [30] also relied on LRF and used a multiple alignment scheme to achieve better results. RIF [16] presented a framework to construct both local and global features based on distances, angles, and reference points. To remove the uncertainty of LRF in x and y directions, RICnv++ [42] extracts informative features based on local reference axis (LRA). However, the LRF/LRA and global reference used in these methods may not be stable enough limiting their overall performance. We can clearly see the accuracy discrepancy of the object classification task on ModelNet40 dataset [36]. State-of-the-art translation-invariant convolutions such as PointNet++ [25] or Point Transformer v2 [34] achieve 89%-94% of accuracy while rotation-invariant convolutions only report up to 87%-91% of accuracy [14,42].", + "bbox": [ + 212, + 347, + 787, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our target in this paper is not only closing this performance gap but also surpassing the state-of-the-art methods while maintaining rotation invariance.", + "bbox": [ + 212, + 651, + 785, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Rotation Invariant Surface Property", + "text_level": 1, + "bbox": [ + 214, + 710, + 604, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we detail the Rotation Invariant Surface Property construction which is the first step of our method. Our goal for this step is to design highly expressive rotation invariant properties from underlying local surfaces. Different from previous works that rely on pointwise operations, we construct local surfaces around the reference point to better capture the local surface structure, based on which we then extract more expressive rotation invariant properties.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z. Zhang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The construction of rotation invariant surface property is shown in Fig. 1. Given a point $p$ as the reference point (red), we get $K$ nearest neighbors to form a local point set. In this case, $K$ is 8. For each neighbor $x_{i}$ , two adjacent neighbors, $x_{i-1}$ and $x_{i+1}$ , are identified based on the Euclidean distances, forming two triangular local surfaces, as shown in the figure. Then, we construct the rotation invariant surface properties (RISP) as follows:", + "bbox": [ + 214, + 146, + 787, + 237 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {R I S P} \\left(x _ {i}\\right) = \\left[ L _ {0}, \\phi_ {1}, \\phi_ {2}, \\phi_ {3}, \\phi_ {4}, \\phi_ {5}, \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 250, + 785, + 273 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left. \\alpha_ {1}, \\alpha_ {2}, \\beta_ {1}, \\beta_ {2}, \\theta_ {1}, \\theta_ {2}, \\gamma_ {1}, \\gamma_ {2} \\right]\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 268, + 617, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L_{0}$ measures the distance from reference $p$ to neighbor $x_{i}$ , and $\\phi_{1}$ to $\\phi_{5}$ measure the two triangles as well as the relationship between the two triangle surfaces with regard to the edge $\\overrightarrow{px_i}$ in the Euclidean space:", + "bbox": [ + 214, + 295, + 782, + 340 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {1} = \\angle (\\overrightarrow {x _ {i - 1} p}, \\overrightarrow {x _ {i} p}), \\phi_ {2} = \\angle (\\overrightarrow {x _ {i + 1} p}, \\overrightarrow {x _ {i} p}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 368, + 785, + 383 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {3} = \\angle (\\overrightarrow {x _ {i - 1} x _ {i}}, \\overrightarrow {x _ {i - 1} p}), \\phi_ {4} = \\angle (\\overrightarrow {x _ {i + 1} p}, \\overrightarrow {x _ {i + 1} x _ {i}}),\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 387, + 671, + 402 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {5} = \\angle (\\overrightarrow {x _ {i + 1} p} \\times \\overrightarrow {x _ {i} p}, \\overrightarrow {x _ {i - 1} p} \\times \\overrightarrow {x _ {i} p}),\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 405, + 583, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "while other properties describe the two surfaces in the tangent space, e.g., normal vectors can define the directions in which the surface is bending away from the tangent space:", + "bbox": [ + 214, + 431, + 782, + 477 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {1} = \\angle (\\overrightarrow {n _ {p}}, \\overrightarrow {x _ {i} p}), \\alpha_ {2} = \\angle (\\overrightarrow {n _ {p}}, \\overrightarrow {x _ {i - 1} p}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 491, + 784, + 506 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\beta_ {1} = \\angle (\\vec {n _ {i}}, \\vec {x _ {i}} \\vec {p}), \\beta_ {2} = \\angle (\\vec {n _ {i}}, \\vec {x _ {i - 1}} \\vec {x _ {i}}),\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 508, + 609, + 523 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {1} = \\angle (\\overrightarrow {n _ {i - 1}}, \\overrightarrow {x _ {i - 1} p}), \\theta_ {2} = \\angle (\\overrightarrow {n _ {i - 1}}, \\overrightarrow {x _ {i - 1} x _ {i} ^ {\\prime}}),\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 527, + 658, + 542 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma_ {1} = \\angle (\\overrightarrow {n _ {i + 1}}, \\overrightarrow {x _ {i + 1} x _ {i}}), \\gamma_ {2} = \\angle (\\overrightarrow {n _ {i + 1}}, \\overrightarrow {x _ {i + 1} p}).\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 547, + 658, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "RISP is able to fully describe the dual triangles as well as their relationship along different directions. Readers can refer to the supplementary for the mathematical proof of the completeness of RISP defined in Eq. (1).", + "bbox": [ + 214, + 574, + 782, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Number of local triangle surfaces. Throughout the paper, we define the number of local triangle surfaces as 2. Such setting is due to the fact that we wish to capture local surface as much as possible to extract more useful information to reflect the real shape of the object. So we construct 2 triangles rather than single triangle. However, this does not mean the more the better since more triangles can result in distorted surface and affect the computing efficiency. Please refer to the ablation study section for more details.", + "bbox": [ + 214, + 633, + 787, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 The RISurConv Operator", + "text_level": 1, + "bbox": [ + 214, + 762, + 501, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the Rotation Invariant Surface Properties (RISP), we are able to build up RISurConv. The main steps are detailed in Fig. 2. To start, we utilize a farthest point sampling strategy to generate uniformly distributed representative", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "RISurConv", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d1b087ed69e990380fbe05236df59dd01c39ae9cb602fc2950de414963679502.jpg", + "image_caption": [ + "Fig. 2: RISurConv operator. For a local point set with $p$ as the reference (red), K nearest neighbors are labelled as blue. Then, we compute the Rotation Invariant Surface Properties at each neighbor by constructing local dual triangle surfaces (Sec. 3), which is embedded to a high-dimensional space by a shared multi-layer perceptron (MLP) followed by a self-attention layer to produce refined features. Concatenated with previous layer features (if any), the features of these local points are further passed to MLPs, which are then summarized by maxpooling. To further refine the features, another self-attention layer follows." + ], + "image_footnote": [], + "bbox": [ + 230, + 141, + 776, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Algorithm 1 RISurConv operator.", + "text_level": 1, + "bbox": [ + 217, + 424, + 475, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Input: Reference point $p$ , point set $\\Omega$ , point features $\\mathbf{f}_{prev}$ from previous layer (if any) \nOutput: Convoluted features $\\mathbf{f}$", + "bbox": [ + 215, + 441, + 784, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "1: $\\mathbf{f} \\gets \\{\\mathrm{RISP}(x_i) : \\forall x_i \\in \\Omega\\}$ (Sec. 3)", + "bbox": [ + 222, + 470, + 416, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "* Construct Rotation Invariant Surface Properties", + "bbox": [ + 450, + 470, + 784, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "2: $\\mathbf{f}\\gets \\mathrm{MLP}(\\mathbf{f})$", + "bbox": [ + 222, + 500, + 331, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "* Embed each feature to a high-dimensional feature space", + "bbox": [ + 395, + 500, + 781, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3: $\\mathbf{f}\\gets \\mathrm{SA}(\\mathbf{f})$", + "bbox": [ + 222, + 513, + 318, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "* Refine features via self-attention layer", + "bbox": [ + 514, + 513, + 781, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4: $\\mathbf{f}_{in}\\gets [\\mathbf{f}_{prev},\\mathbf{f}]$ (if any)", + "bbox": [ + 222, + 527, + 339, + 554 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "* Concatenate the features from the local and the previous layer", + "bbox": [ + 357, + 527, + 784, + 554 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5: $\\mathbf{f}_{out} \\gets \\mathrm{MLP}(\\mathbf{f}_{in})$", + "bbox": [ + 222, + 554, + 354, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "* Feature embedding", + "bbox": [ + 638, + 554, + 779, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6: $\\mathbf{f}_{out}\\gets$ maxpool $(\\mathbf{f}_{out})$ return SA $(\\mathbf{f}_{out})$", + "bbox": [ + 222, + 568, + 383, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "* Maxpool features", + "bbox": [ + 656, + 569, + 785, + 582 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "* Self-attention and return", + "bbox": [ + 606, + 582, + 785, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "points. From this initial set, we perform K-nearest neighbor searches to obtain local point sets. Let us denote a local point set by $\\Omega = \\{x_{i}\\}$ , where $x_{i}$ represents the 3D coordinates of point $i$ . We define the attention-augmented convolution operation as follows:", + "bbox": [ + 212, + 630, + 787, + 691 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} (\\Omega) = S A \\left(\\sigma \\left(\\mathcal {A} \\left(\\left\\{\\mathcal {T} \\left(\\mathbf {f} _ {x _ {i}}\\right): \\forall i \\right\\}\\right)\\right)\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 703, + 785, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This formula indicates that features of each point in the point set are first transformed by $\\mathcal{T}$ before being aggregated by the aggregation function $\\mathcal{A}$ and passed to an activation function $\\sigma$ . SA is a channel-wise self-attention layer used to output refined features. We set the input features to our expressive rotation-invariant features $\\mathbf{f}_{x_i} = \\mathrm{RISP}(x_i)$ . We define the transformation function as", + "bbox": [ + 212, + 731, + 787, + 808 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} \\left(\\mathbf {f} _ {x _ {i}}\\right) = \\mathbf {w} _ {i} \\cdot \\mathbf {f} _ {x _ {i}} = \\mathbf {f} _ {x _ {i}} ^ {\\prime} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 418, + 819, + 785, + 837 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z. Zhang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\cdot$ indicates the element-wise product, and $\\mathbf{w}_i$ is the weight parameter to be learned by the network. Our transformation function is similar to PointNet++. A popular choice of the aggregation function $\\mathcal{A}$ is maxpooling, which supports permutation invariance of the input point features.", + "bbox": [ + 212, + 146, + 784, + 205 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To proceed with feature learning and be invariant to point permutation, we construct surfaces for each neighbor from which we construct RISPs. In Fig. 2, there are eight neighbors, so the resulted RISPs are in size of $8 \\times 14$ . Here, 14 is the length of the RISP as defined in Eq. (1) and 8 is the neighborhood size. Since RISPs are used as the input which are already rotation invariant, RISurConv is rotation invariant by nature. We then embed the RISPs into feature space using two layers of MLPs followed by a self-attention (SA) layer for feature refinement. We concatenate the refined features with previous layer features (if any), and embed the concatenated features into higher dimensional space via MLPs again. After the maxpooling, the aggregated features are passed into another SA layer again for further refinement. The detailed steps are shown in the Algorithm 1. Here, the two SA layers used in RISurConv are the standard SA module of the transformer [32].", + "bbox": [ + 212, + 207, + 785, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Self Attention Layers. In RISurConv there are two Self-Attention (SA) layers by which the extracted features are enhanced. Below, we provide a more detailed illustration of the structure of the SA layers. The SA structure is the standard SA module of the famous transformer [32] (Attention is all you need) which runs twice in RISurConv. The first SA is for the K neighbors, and the second SA works for the N representative points. The input tensor shape is $[B,N,K,C_0]$ for the first SA which goes into linear layers and outputs $Q,K$ and $V$ respectively in shape of $[B,N,K,C_0]$ . According to $\\text{att}(Q,K,V) = \\text{Softmax}(QK^{\\top} / \\text{sqrt}(d_k))V$ , we have the 1st attention score in $[B,N,K,K]$ . By multiplying with $V$ , we get the refined feature in $[B,N,K,C_0]$ . The same goes with the second SA where the input shape is $[B,N,C]$ and the attention score shape is $[B,N,N]$ .", + "bbox": [ + 212, + 402, + 787, + 587 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5 RISurConv Networks", + "text_level": 1, + "bbox": [ + 215, + 609, + 460, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We employed the RISurConv to develop neural networks for object classification and segmentation, as illustrated in Fig. 3. Our classification network follows a standard architecture similar to the single-scale grouping version of PointNet++, consisting of five consecutive layers of RISurConv followed by a transformer encoder [32] with 8 heads to enhance the extracted features. Finally, we connect the network to fully connected layers to produce the probability map. One of the major advantages of RISurConv is its ability to handle arbitrary rotation and point orders, which enables us to place each RISurConv layer consecutively without the need for complex preprocessing steps. Additionally, we apply batch normalization and ReLU activation to each convolution layer by default. The segmentation network utilizes an encoder-decoder structure similar to the U-Net [28]. We adopt the same definition of deconvolution as RISurConv. The key distinction lies in that the convolution produces a point subset with a greater", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "RISurConv", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/87105d49eaf443ae2b40917f81563b05648fb1b683c59903ffa1a82c5b00d65c.jpg", + "image_caption": [ + "Fig. 3: Our neural network architecture comprises five RISurConv layers to extract rotation invariant features followed by a Transformer Encoder to enhance the learnt features before fully connected layers for object classification. We add a decoder with skip connections for segmentation task." + ], + "image_footnote": [], + "bbox": [ + 220, + 142, + 782, + 280 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "number of feature channels and less points while the deconvolution outputs to a point set with more points compared to the input with fewer feature channels. Please refer to the supplementary for more details regarding the network architecture.", + "bbox": [ + 212, + 378, + 787, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6 Experiments", + "text_level": 1, + "bbox": [ + 215, + 476, + 375, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We report our evaluation results in this section. Our network is implemented in PyTorch using an Adam optimizer with initial learning rate of 0.001 for both classification and segmentation. Our training is executed on a computer with an Intel(R) Core(TM) i7-10700K CPU equipped with a NVIDIA GTX 3090 GPU.", + "bbox": [ + 212, + 520, + 785, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluated the performance of our method extensively on a series of classification and segmentation benchmarks. To train our model, we utilized 450 epochs for classification and 250 epochs for segmentation for most datasets except the Car and Chair categories of the fine-grained FG3D dataset due to small size and subtle variation where we applied 900 epochs. Our method can achieve convergence within 11 hours for classification and 24 hours for segmentation. To assess the robustness of our method, we followed the experimental design proposed by [8] and conducted experiments in three cases. The first case involved training and testing with data augmented with rotation about the gravity axis $(\\mathrm{z} / \\mathrm{z})$ , which is a common approach for evaluating translation-invariant point cloud learning methods. The second case involved training and testing with data augmented with arbitrary SO3 rotations (SO3/SO3), and the third case involved training with data by z-rotations and testing with data by SO3 rotations (z/SO3), which are used to evaluate rotation invariance. A rotation-invariant method should produce consistent results across all three cases. Our method demonstrated superior performance across all three cases, highlighting its robustness and effectiveness.", + "bbox": [ + 212, + 583, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z. Zhang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.1 Human-made Object Classification", + "text_level": 1, + "bbox": [ + 215, + 146, + 549, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ModelNet40 [36] that comprises 9843 training models and 2468 test models, divided into 40 categories is one of the most popular datasets for point cloud classification evaluation. The input data size is 1024 with each point equipped with (x; y; z; nx; ny; nz) which are 3D coordinates and 3D normals in the Euclidean space. Note that normals are optional to our method. When normals are not available (w/o n), we use weighted eigen vector corresponding to the smallest eigen value as the normal. This strategy has been used in SHOT and RICConv++ [42] which is highly robust to noise and efficient to compute.", + "bbox": [ + 212, + 180, + 787, + 301 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The experimental results are shown in Tab. 1. We use two criteria for evaluation: overall accuracy and accuracy standard deviation (Std.). True rotation invariant methods are expected to be unaffected by the rotations present in the training and testing data, resulting in low accuracy deviation.", + "bbox": [ + 212, + 303, + 787, + 364 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "From Tab. 1, we see that our method surpasses the state-of-the-art performance in all cases while maintaining zero accuracy deviation. It is worth noting that our method surpasses both rotation invariant and non-rotation invariant methods. To the best of our knowledge, this is the first work that achieves such high accuracy. It outperforms the state-of-the-art rotation invariant method RIConv++ by $4.7\\%$ , and outperforms the state-of-the-art non-rotation invariant method Point Transformer v2 by $1.8\\%$ under $\\mathrm{z / z}$ case.", + "bbox": [ + 212, + 366, + 787, + 470 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6.2 Real World Object Classification", + "text_level": 1, + "bbox": [ + 214, + 503, + 532, + 518 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We also evaluate the classification performance on a real-world 3D point cloud dataset ScanObjectNN [31]. It is composed of 2902 point clouds categorized into 15 categories sampled from real world indoor scenes. For our evaluation, we use the processed files and choose the hardest variant PB_T50_RS with $50\\%$ bounding box translation, rotation around the gravity axis, and random scaling that result in rotated and partial data. The results are shown", + "bbox": [ + 212, + 537, + 460, + 747 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/3b6c977530a0cfae5188ffb36b9a51b17c2506bdd8f72f71e9682d368ccaa3b7.jpg", + "table_caption": [ + "Table 2: Comparisons of real world 3D point cloud classification on hardest variant of ScanObjectNN dataset (Overall Accuracy)." + ], + "table_footnote": [], + "table_body": "
MethodPB_T50_RS
z/zSO3/SO3z/SO3
PointNet [23]68.242.217.1
PointNet++ [25]77.960.115.8
PointCNN [17]78.551.814.9
DGCNN [33]78.163.416.1
RICov [41]68.168.368.3
GCACov [40]69.870.069.8
RICov++ [42]80.380.380.3
Ours (w/o normal)93.193.193.1
", + "bbox": [ + 475, + 583, + 781, + 744 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "in Tab. 2. We see that RISurConv surpasses all the compared methods by a large margin. Particularly, our method significantly outperforms the state-of-the-art rotation invariant approach RICov++ [42] by $12.8\\%$ . This verifies that our method is effective for both synthetic and real world data. Note that we only test the 'w/o normal' case as the normal vectors are not provided in the processed files of this dataset.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "RISurConv", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/6ab3adbb3aa800e993c6bf96dfc181373938e4823ee1857537d58e260a41f566.jpg", + "table_caption": [ + "Table 1: Comparisons of the classification accuracy (\\%) on the ModelNet40 dataset. On average, our method has the best accuracy and lowest accuracy deviation in all cases (Overall Accuracy)." + ], + "table_footnote": [], + "table_body": "
MethodFormatInput Size Params.z/z↑SO3/ SO3↑z/ SO3↑Std.↓
TraditionalVoxNet [21]voxel3030.90M83.087.3-3.0
SubVolSup [24]voxel30317.00M88.582.736.628.4
PointNet [23]xyz1024 × 33.50M87.080.321.641.0
PointCNN [17]xyz1024 × 30.60M91.384.541.227.2
PointNet++ [25]xyz + nor1024 × 61.40M89.385.028.633.8
DGCNN [33]xyz1024 × 31.84M92.281.120.638.5
RS-CNN [20]xyz1024 × 31.41M90.382.648.722.1
Pt Transformer [43]xyz1024 × 3-93.785.950.119.1
Pt Transformer v2 [34]xyz1024 × 3-94.288.351.823.0
Rotation-invariantSpherical CNN [7]voxel2 × 6420.50M88.986.978.65.5
RICConv [41]xyz1024 × 30.70M86.586.486.40.1
SPHNet [22]xyz1024 × 32.90M87.087.686.60.5
SFCNN [27]xyz1024 × 3-91.490.184.83.5
ClusterNet [3]xyz1024 × 31.40M87.187.187.10.0
GCACov [40]xyz1024 × 30.41M89.089.289.10.0
RIF [16]xyz1024 × 3-89.489.389.40.0
RI-GCN [15]xyz + nor1024 × 64.38M91.091.091.00.0
RICConv++ [42]xyz1024 × 30.40M91.291.291.20.0
RICConv++ [42]xyz + nor1024 × 60.40M91.391.391.30.0
Ours (w/o normal)xyz1024 × 314.0M95.695.695.60.0
Ours (w/ normal)xyz + nor1024 × 614.0M96.096.096.00.0
", + "bbox": [ + 243, + 198, + 764, + 500 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6.3 Fine-Grained Object Classification", + "text_level": 1, + "bbox": [ + 212, + 531, + 547, + 546 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Fine-grained object classification is more challenging because the differences between subcategories are subtle and require a high level of precision in distinguishing them. We conduct experiments on the three categories from FD3D dataset [19]: Airplane, Chair and Car. To the best of our knowledge, this is the first work to test the performance of rotation invariant methods on fine-grained 3D point cloud dataset. The class accuracies are shown in Tab. 3. Again, our method outperforms the state-of-the-art approaches by large margins for all rotation cases and all the categories. Specifically, our method improves non-rotation invariant method FG3D-Net by $2.1\\%$ , $2.7\\%$ , and $4.5\\%$ for Airplane, Chair, and Car categories respectively under the $\\mathrm{z} / \\mathrm{z}$ rotation case, and the accuracy consistency is well preserved. Compared to the rotation invariant methods, our method outperforms the latest RICov++ by $3.6\\%$ , $5.1\\%$ , and $9.2\\%$ on the three categories. Such supreme performance shows that features extracted by RISurConv are not only rotation invariant but also highly expressive.", + "bbox": [ + 212, + 564, + 787, + 776 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6.4 Part Segmentation on ShapeNet", + "text_level": 1, + "bbox": [ + 212, + 805, + 529, + 821 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z. Zhang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/6574cfce0b23e77f1976bbcb33cd24597c8524d6b29ac2cb096b778e0660fb91.jpg", + "table_caption": [ + "Table 3: Comparisons of fine-grained 3D point cloud classification on FG3D dataset (Class Accuracy)." + ], + "table_footnote": [], + "table_body": "
MethodModalityAirplaneChairCar
z/zSO3/SO3z/SO3z/zSO3/SO3z/SO3z/zSO3/SO3z/SO3
VoxNetVoxel79.257.528.773.354.716.768.242.217.1
MVCNNView82.666.921.976.363.714.671.951.814.9
View-GCNView87.456.622.779.750.219.773.744.418.1
RotationNetView89.150.829.878.536.218.472.545.328.3
Part4FeatureView82.649.132.277.152.230.273.435.023.8
FG3D-NetView89.461.126.980.043.218.474.056.420.8
PointNetPoint82.755.529.772.133.918.268.132.116.4
PointNet++Point87.357.125.678.140.416.070.350.420.8
RS-CNNPoint82.845.932.275.150.718.671.231.813.9
DGCNNPoint88.460.622.771.752.319.765.353.418.1
RICConvPoint85.885.885.876.376.376.367.367.367.3
RICConv++Point87.987.987.977.677.677.669.369.369.3
OursPoint91.591.591.582.782.782.778.578.578.5
", + "bbox": [ + 225, + 184, + 774, + 404 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To evaluate the segmentation performance, we employed the ShapeNet dataset comprising 16880 CAD models distributed across 16 distinct categories. These models are annotated with 2 to 6 parts each, culminating in a dataset of 50 object parts. We follow the standard train/test split with 14006 models for training and 2874 models for testing, respectively.", + "bbox": [ + 212, + 463, + 446, + 643 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The evaluation results are shown in Tab. 4. Our method outperforms both traditional translation-invariant and latest rotation invariant methods significantly in the SO3/SO3 and z/SO3 scenario. Our method outperforms the RICov++ by", + "bbox": [ + 212, + 657, + 446, + 777 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/bdcbc11ea25b657b6a1cc3afd5b6ff43295c7486fe37f0a3f16728bfc4f63515.jpg", + "table_caption": [ + "Table 4: Comparisons of object part segmentation performed on ShapeNet dataset. The mean per-class IoU (mIoU, %) is used to measure the accuracy under two challenging rotation modes: SO3/SO3 and z/SO3." + ], + "table_footnote": [], + "table_body": "
MethodInputSO3/SO3 z/SO3
PointNet [23]xyz74.437.8
PointNet++ [25]xyz+nor76.748.2
PointCNN [17]xyz71.434.7
DGCNN [33]xyz73.337.4
RS-CNN [20]xyz72.536.5
SpiderCNNxyz+nor72.342.9
RICConv [41]xyz75.575.3
GCACConv [40]xyz77.377.2
RI-GCN [15]xyz77.077.0
RIF [16]xyz79.479.2
RICConv++ [42]xyz80.380.3
RICConv++ [42]xyz+nor80.580.5
Ours (w/o normal)xyz81.381.3
Ours (w/ normal)xyz+nor81.581.5
", + "bbox": [ + 473, + 536, + 776, + 765 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "1.0% mIoU. This result aligns well with the performance reported in the object classification task. We also show the qualitative results by error maps in Fig. 4. The wrong segmentation points are plotted as red. It clearly shows that our predictions are the closest to the ground truth.", + "bbox": [ + 212, + 779, + 784, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "RISurConv", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/04988c00896f6334af72dad5432b97b84261d7f7f701a72857e6be6179788f6d.jpg", + "image_caption": [ + "Fig. 4: Qualitative comparisons (Red indicates wrong)." + ], + "image_footnote": [], + "bbox": [ + 251, + 142, + 751, + 391 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "6.5 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 446, + 401, + 460 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We ablate some vital designs of our method on ModelNet40 [36] for an insightful exploration.", + "bbox": [ + 212, + 474, + 784, + 503 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "RISP Design. We first test the performance by turning on/off and adding different rotation invariant components used in the RISP construction (Algorithm 1). The results are shown in Tab. 5. Model A is our baseline setting with all rotation invariant features activated. Model B has only angle features but still achieves relatively high accuracy. Model C turns off both $L_{0}$ and $\\phi_{1\\dots 5}$ with accuracy decreases to $90.9\\%$ . This means that the distance $L_{0}$ is not as important as angles. In Model D, we turn off the angle features on the tangent direction with only $L_{0}$ and $\\phi_{1\\dots 5}$ are kept which is more like RICnv [41]. Compared to Model A, it shows that our proposed features is more effective than those by RICnv, and the improvement is explained by the additional consideration of the relations among the neighbor points. Model E adds more rotation invariant", + "bbox": [ + 212, + 505, + 787, + 672 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/13e4a140c656df0ee08152b0bcbea8c073da1fa088e1a0ad7100f728ddeea88a.jpg", + "table_caption": [ + "Table 5: Ablation study on the RISP design." + ], + "table_footnote": [], + "table_body": "
ModelL0φ1...5α1,2β1,2θ1,2γ1,2λμAcc.
A✓✓✓96.0
B✓✓95.5
C✓90.9
D✓✓88.2
E✓✓✓✓95.7
", + "bbox": [ + 222, + 739, + 480, + 835 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/f1155cc4c4c886c05df96ef5fcd8ac30bf78f29fec0266d378cee08e83cc1253.jpg", + "table_caption": [ + "Table 6: Ablation study on the Self-Attention module." + ], + "table_footnote": [], + "table_body": "
ModelSA1SA2TransformerEncoderAcc.
A✓✓✓96.0
B✓✓95.6
C✓✓95.2
D✓✓94.3
E92.8
", + "bbox": [ + 521, + 739, + 782, + 835 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z. Zhang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "features. $\\lambda$ and $\\mu$ represent the angles $\\angle (\\pmb{n}_{\\pmb{p}},\\pmb{n}_{i + 1})$ and $\\angle (\\pmb {n}_p,\\pmb {n}_{i - 1})$ respectively. The results show that the accuracy does not increase. This verifies the completeness of the proposed RISP.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Self-Attention Effects. We employ multiple self-attention (SA) modules to help produce refined rotation invariant features. So it is necessary to analyse the effects of different SA modules. In RISurConv, there are two SA modules and we name them as SA1 and SA2 respectively. Before fully connected layers, there is a Transformer Encoder module. We test these three modules by turning off one of them each time. From Tab. 6, we see that when removing the first SA module in RISurConv, the performance drops a bit, while removing the second SA module also decreases the accuracy. This indicates that both SA1 and SA2 are important for the success of our method. In addition, we also test the performance when Transformer Encoder is removed. The overall accuracy decreases to $94.3\\%$ but it is still higher than most existing methods. In model E, we test the performance when all SA modules are removed. The accuracy decreases to $92.8\\%$ which is still better than the state-of-the-art rotation invariant method [42] thanks to the well designed RISPs which capture sufficient surface structures.", + "bbox": [ + 212, + 196, + 784, + 407 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We also visualize the global feature vector before classification as shown in Fig. 5. We normalize the feature values to [0,1] and plot the histograms. We can see that by turning on the SA modules, the feature values become more evenly distributed and have less zeros because SA mechanism can reweight the features making the features more effective.", + "bbox": [ + 212, + 407, + 429, + 602 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/563e38b74af65103e4fbea3c41cfa12cb6705795b3829a6e5ec8dc24cfd593b3.jpg", + "image_caption": [ + "Fig. 5: Histogram comparison for normalized feature values without and with self-attention layers." + ], + "image_footnote": [], + "bbox": [ + 442, + 421, + 782, + 542 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Network Efficiency. In the experiments, we acknowledge that our method employs a higher number of parameters. Thus, it is imperative to conduct a comprehensive analysis of network efficiency during both the training and testing stages. We list the trainable parameters, FLOPs", + "bbox": [ + 212, + 608, + 413, + 773 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/9532ca7fd26e344857481956a1d8c87e3d83dfdc4b503057a13ddcfbbf6811a4.jpg", + "table_caption": [ + "Table 7: Trainable parameters, FLOPs and timing comparisons on ModelNet40. All are tested on the same platform. Time is measured per batch (batch size=16)." + ], + "table_footnote": [], + "table_body": "
MethodsParamsFLOPsTime (Train/Infer)
PointNet++ [25]1.41M0.86G0.145s / 0.129s
RICnv [41]0.70M0.92G0.121s / 0.092s
RI-GCN [15]4.19M1.24G0.148s / 0.112s
RICnv++ [42]0.42M0.72G0.074s / 0.029s
Ours13.96M1.12G0.114s / 0.049s
Ours w/o TR1.35M1.11G0.105s / 0.047s
", + "bbox": [ + 433, + 655, + 777, + 770 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "and running time comparisons in Tab. 7. We compare the performance with four recent works: PointNet++ [25], RICnv [41], RI-GCN [15], RICnv++ [42].", + "bbox": [ + 212, + 775, + 782, + 806 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "From the results, we can see that the number of parameters mainly comes from the transformer architectures used in RISurConv. When we turn off all", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "RISurConv", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "the transformer component, our method only takes up 1.35M parameters. More parameters do not mean longer inference time. It depends on lots of things including depth, parameter, number of operations etc. Our approach runs faster than most existing methods. It takes more time for training because more epochs are needed for self attention layers to converge. This is a common phenomenon of transformer based methods. The accuracy is also not due to the complexity model. As is shown in the ablation studies, by removing transformer encoder, it still achieves state-of-the-art performance due to the effectiveness of RISurConv.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Number of Local Surfaces. Local triangle surfaces used in Sec. 3 are the base for rotation invariant surface property (RISP) construction. Thus, it is important to analyse the effects of different number of surfaces. We conduct performance tests by setting the number of local surfaces to 1, 2, 3, and 4, resulting in overall accuracies of 94.7, 96.0, 91.9, and 89.8 respectively. It worth noting that for more surfaces, we can define more RISPs. From the results, we see the accuracy drops a bit when single", + "bbox": [ + 212, + 268, + 555, + 433 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/2e15e004c3454175b9ee2af6ae6f949fdf51f2d4757028bdacf6ddf6022fbc67.jpg", + "image_caption": [ + "Fig. 6: Distorted surface caused by 3 triangles." + ], + "image_footnote": [], + "bbox": [ + 571, + 280, + 782, + 391 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "surface is constructed. This is because smaller area cannot capture sufficient surface structures and less effective features can lower the performance. When we increase the number of surfaces to 3 and 4, the performance decreases. This shows that more surfaces can result in inferior performance because we construct triangle surfaces just to approximate the true surface. More surfaces may mess up the surface structure and affect the performance. Take Fig. 6 for example, suppose we are constructing the local surfaces $(\\triangle px_{i}x_{i + 1}\\triangle px_{i}x_{i - 1})$ with regard to the neighbor point $x_{i}$ , adding one more triangle $\\triangle px_{i}x_{i + 2}$ will result in distorted/non-manifold surfaces. Thus, we set number of surfaces as 2 throughout the paper.", + "bbox": [ + 212, + 433, + 787, + 584 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Discussion", + "text_level": 1, + "bbox": [ + 215, + 609, + 354, + 625 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "**Limitation.** Though effective, RISurConv may require longer training time (e.g. more epochs) to converge due to large number of training parameters (Tab. 1). This is a common issue of transformer. But this does not affect the inference speed as shown in the ablation study (Tab. 7).", + "bbox": [ + 212, + 642, + 785, + 703 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Conclusion. We have presented a new framework for achieving rotation invariance on 3D point cloud. We construct local triangle surfaces to better represent the local surface structures based on which we design highly expressive rotation invariant surface properties. We integrate the properties into an architecture named RISurConv to extract refined rotation invariant features. Based on RISurConv, we finally build up effective rotation invariant neural networks for 3D point cloud classification and segmentation with supreme performance achieved not only closing the performance gap with non-rotation invariant approaches but also surpassing the state-of-the-art methods by a large margin.", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z. Zhang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This research is supported by the Singapore Ministry of Education (MOE) Academic Research Fund (AcRF) Tier 1 grant (MSS23C010), Ningbo 2025 Science & Technology Innovation Major Project (No. 2022Z072, No.2023Z044), and Key Research & Development Plan of Zhejiang Province (2024C01017).", + "bbox": [ + 212, + 176, + 787, + 239 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 263, + 323, + 277 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3d semantic parsing of large-scale indoor spaces. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1534-1543 (2016)", + "2. Chang, A.X., Funkhouser, T.A., Guibas, L.J., Hanrahan, P., Huang, Q.X., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., Xiao, J., Yi, L., Yu, F.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015)", + "3. Chen, C., Li, G., Xu, R., Chen, T., Wang, M., Lin, L.: Clusternet: Deep hierarchical cluster network with rigorously rotation-invariant representation for point cloud analysis. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4994-5002 (2019)", + "4. Deng, H., Birdal, T., Ilic, S.: Ppf-foldnet: Unsupervised learning of rotation invariant 3d local descriptors. In: Eur. Conf. Comput. Vis. pp. 602-618 (2018)", + "5. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: NAACL-HLT (2018)", + "6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Int. Conf. Learn. Represent. (2020)", + "7. Esteves, C., Allen-Blanchette, C., Makadia, A., Daniilidis, K.: Learning so (3) equivariant representations with spherical cnns. In: Eur. Conf. Comput. Vis. pp. 52-68 (2018)", + "8. Esteves, C., Xu, Y., Allen-Blanchette, C., Daniilidis, K.: Equivariant multi-view networks. In: Int. Conf. Comput. Vis. pp. 1568-1577 (2019)", + "9. Guo, M.H., Cai, J.X., Liu, Z.N., Mu, T.J., Martin, R.R., Hu, S.M.: Pct: Point cloud transformer. Computational Visual Media 7, 187-199 (2021)", + "10. Guo, Y., Wang, H., Hu, Q., Liu, H., Liu, L., Bennamoun, M.: Deep learning for 3d point clouds: A survey. IEEE Trans. Pattern Anal. Mach. Intell. 43(12), 4338-4364 (2020)", + "1. Han, X.F., Jin, Y.F., Cheng, H.X., Xiao, G.Q.: Dual transformer for point cloud analysis. IEEE Trans. Multimedia (2022)", + "2. Hua, B.S., Pham, Q.H., Nguyen, D.T., Tran, M.K., Yu, L.F., Yeung, S.K.: Scenenn: A scene meshes dataset with annotations. In: International Conference on 3D Vision. pp. 92-101 (2016)", + "3. Huang, Q., Wang, W., Neumann, U.: Recurrent slice networks for 3d segmentation of point clouds. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2626-2635 (2018)", + "4. Kim, J., Jung, W., Kim, H., Lee, J.: Cynnn: a rotation invariant cnn using polar mapping and cylindrical convolution layers. arXiv preprint arXiv:2007.10588 (2020)", + "5. Kim, S., Park, J., Han, B.: Rotation-invariant local-to-global representation learning for 3d point cloud. Adv. Neural Inform. Process. Syst. 33, 8174-8185 (2020)" + ], + "bbox": [ + 225, + 295, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "RISurConv", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Li, X., Li, R., Chen, G., Fu, C.W., Cohen-Or, D., Heng, P.A.: A rotation-invariant framework for deep point cloud analysis. IEEE Trans. Vis. Comput. Graph. (2021)", + "17. Li, Y., Bu, R., Sun, M., Wu, W., Di, X., Chen, B.: Pointcnn: Convolution on x-transformed points. In: Adv. Neural Inform. Process. Syst. pp. 820-830 (2018)", + "18. Liu, X., Han, Z., Liu, Y.S., Zwicker, M.: Point2sequence: Learning the shape representation of 3d point clouds with an attention-based sequence to sequence network. In: AAAI. vol. 33, pp. 8778-8785 (2019)", + "19. Liu, X., Han, Z., Liu, Y.S., Zwicker, M.: Fine-grained 3d shape classification with hierarchical part-view attention. IEEE Trans. Image Process. 30, 1744-1758 (2021)", + "20. Liu, Y., Fan, B., Xiang, S., Pan, C.: Relation-shape convolutional neural network for point cloud analysis. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 8895-8904 (2019)", + "21. Maturana, D., Scherer, S.: Voxnet: A 3d convolutional neural network for real-time object recognition. In: 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 922-928 (2015)", + "22. Poulenard, A., Rakotosaona, M.J., Ponty, Y., Ovsjanikov, M.: Effective rotation-invariant point cnn with spherical harmonics kernels. International Conference on 3D Vision (2019)", + "23. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 652-660 (2017)", + "24. Qi, C.R., Su, H., Nießner, M., Dai, A., Yan, M., Guibas, L.J.: Volumetric and multi-view cnns for object classification on 3d data. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5648-5656 (2016)", + "25. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: Adv. Neural Inform. Process. Syst. pp. 5105-5114 (2017)", + "26. Qian, G., Li, Y., Peng, H., Mai, J., Hammoud, H., Elhoseiny, M., Ghanem, B.: Pointnext: Revisiting pointnet++ with improved training and scaling strategies. Adv. Neural Inform. Process. Syst. 35, 23192-23204 (2022)", + "27. Rao, Y., Lu, J., Zhou, J.: Spherical fractal convolutional neural networks for point cloud recognition. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019)", + "28. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: International Conference on Medical image computing and computer-assisted intervention. pp. 234-241 (2015)", + "29. Shen, Y., Feng, C., Yang, Y., Tian, D.: Mining point cloud local structures by kernel correlation and graph pooling. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4548-4557 (2018)", + "30. Thomas, H.: Rotation-invariant point convolution with multiple equivariant alignments. In: International Conference on 3D Vision (3DV). pp. 504-513 (2020)", + "31. Uy, M.A., Pham, Q.H., Hua, B.S., Nguyen, D.T., Yeung, S.K.: Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In: Int. Conf. Comput. Vis. (2019)", + "32. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Adv. Neural Inform. Process. Syst. 30 (2017)", + "33. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph cnn for learning on point clouds. ACM Trans. Graph. (2019)", + "34. Wu, X., Lao, Y., Jiang, L., Liu, X., Zhao, H.: Point transformer v2: Grouped vector attention and partition-based pooling. Adv. Neural Inform. Process. Syst. 35, 33330-33342 (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z. Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., Xiao, J.: 3d shapenets: A deep representation for volumetric shapes. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1912-1920 (2015)", + "36. Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., Xiao, J.: 3d shapenets: A deep representation for volumetric shapes. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1912-1920 (2015)", + "37. Xie, S., Liu, S., Chen, Z., Tu, Z.: Attentional shapecontextnet for point cloud recognition. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4606-4615 (2018)", + "38. Xu, Y., Fan, T., Xu, M., Zeng, L., Qiao, Y.: SpiderCNN: Deep learning on point sets with parameterized convolutional filters. In: Eur. Conf. Comput. Vis. pp. 87-102 (2018)", + "39. Yu, X., Tang, L., Rao, Y., Huang, T., Zhou, J., Lu, J.: Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 19313-19322 (2022)", + "40. Zhang, Z., Hua, B.S., Chen, W., Tian, Y., Yeung, S.K.: Global context aware convolutions for 3d point cloud understanding. In: International Conference on 3D Vision (2020)", + "41. Zhang, Z., Hua, B.S., Rosen, D.W., Yeung, S.K.: Rotation invariant convolutions for 3d point clouds deep learning. In: International Conference on 3D Vision. pp. 204-213 (2019)", + "42. Zhang, Z., Hua, B.S., Yeung, S.K.: Riconv++: Effective rotation invariant convolutions for 3d point clouds deep learning. Int. J. Comput. Vis. 130(5), 1228-1243 (2022)", + "43. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: Int. Conf. Comput. Vis. pp. 16259-16268 (2021)", + "44. Zhao, Y., Birdal, T., Lenssen, J.E., Menegatti, E., Guibas, L., Tombari, F.: Quaternion equivariant capsule networks for 3d point clouds. In: Eur. Conf. Comput. Vis. pp. 1-19. Springer (2020)" + ], + "bbox": [ + 212, + 146, + 787, + 535 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "RISurConv", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/RISurConv_ Rotation Invariant Surface Attention-Augmented Convolutions for 3D Point Cloud Classification and Segmentation/85a1ebc6-6b7d-4e16-8cd8-c420684dc80d_model.json b/2024/RISurConv_ Rotation Invariant Surface Attention-Augmented Convolutions for 3D Point Cloud Classification and Segmentation/85a1ebc6-6b7d-4e16-8cd8-c420684dc80d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b0037e2b367b26ca22367be7d4edcfc607e7cb5d --- /dev/null +++ b/2024/RISurConv_ Rotation Invariant Surface Attention-Augmented Convolutions for 3D Point Cloud Classification and Segmentation/85a1ebc6-6b7d-4e16-8cd8-c420684dc80d_model.json @@ -0,0 +1,2324 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.224, + 0.141, + 0.782, + 0.21 + ], + "angle": 0, + "content": "RISurConv: Rotation Invariant Surface Attention-Augmented Convolutions for 3D Point Cloud Classification and Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.3, + 0.235, + 0.703, + 0.253 + ], + "angle": 0, + "content": "Zhiyuan Zhang\\(^{1}\\), Licheng Yang\\(^{2\\star}\\), and Zhiyu Xiang\\(^{2\\oplus}\\)" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.262, + 0.782, + 0.291 + ], + "angle": 0, + "content": "\\(^{1}\\) School of Computing and Information Systems, Singapore Management University https://github.com/cszyzhang/RISurConv" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.291, + 0.771, + 0.305 + ], + "angle": 0, + "content": "\\(^{2}\\) College of Information Science and Electronic Engineering, Zhejiang University" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.262, + 0.782, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.339, + 0.744, + 0.631 + ], + "angle": 0, + "content": "Abstract. Despite the progress on 3D point cloud deep learning, most prior works focus on learning features that are invariant to translation and point permutation, and very limited efforts have been devoted for rotation invariant property. Several recent studies achieve rotation invariance at the cost of lower accuracies. In this work, we close this gap by proposing a novel yet effective rotation invariant architecture for 3D point cloud classification and segmentation. Instead of traditional pointwise operations, we construct local triangle surfaces to capture more detailed surface structure, based on which we can extract highly expressive rotation invariant surface properties which are then integrated into an attention-augmented convolution operator named RISurConv to generate refined attention features via self-attention layers. Based on RISurConv we build an effective neural network for 3D point cloud analysis that is invariant to arbitrary rotations while maintaining high accuracy. We verify the performance on various benchmarks with supreme results obtained surpassing the previous state-of-the-art by a large margin. We achieve an overall accuracy of \\(96.0\\%\\) \\((+4.7\\%)\\) on ModelNet40, \\(93.1\\%\\) \\((+12.8\\%)\\) on ScanObjectNN, and class accuracies of \\(91.5\\%\\) \\((+3.6\\%)\\), \\(82.7\\%\\) \\((+5.1\\%)\\), and \\(78.5\\%\\) \\((+9.2\\%)\\) on the three categories of the FG3D dataset for the fine-grained classification task. Additionally, we achieve \\(81.5\\%\\) \\((+1.0\\%)\\) mIoU on ShapeNet for the segmentation task." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.643, + 0.644, + 0.658 + ], + "angle": 0, + "content": "Keywords: Point cloud \\(\\cdot\\) Rotation invariant \\(\\cdot\\) Attention" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.682, + 0.377, + 0.698 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.712, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Point cloud has become the most promising 3D data representation for a wide range of immersive applications from robot navigation to autonomous driving. The increasing availability of 3D sensors and the emergence of large and mid-scale point cloud datasets [1,2,12,19,31,35] have spurred significant research in this area, leading to the development of numerous deep learning models for point cloud classification and segmentation, which are the fundamentals for various computer vision tasks." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.826, + 0.334, + 0.839 + ], + "angle": 0, + "content": "\\* co-first author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.129 + ], + "angle": 0, + "content": "Z. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.236 + ], + "angle": 0, + "content": "However, analyzing 3D point clouds remains challenging, mainly due to the irregular nature of point clouds and their inherent invariances such as translation, point permutation, and rotation. While significant progress has been made in learning translation and permutation-invariant features [17, 23, 25, 26, 44], achieving rotation invariance in point cloud convolution has been a relatively unexplored area." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.788, + 0.344 + ], + "angle": 0, + "content": "Rotation invariance is essential for 3D object classification and segmentation as objects can be viewed from different viewpoints, leading to variations in their orientation. Prior approaches [23,25,33] usually rely on rotation augmentation in training stage to relief the rotation affections for testing. However, such scheme is less effective as 3D data has more degrees of freedom. A rotation invariant model is, therefore, critical to accurately classify and segment objects in 3D point clouds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.555 + ], + "angle": 0, + "content": "Various attempts have been made to address the problem of rotation invariance in 3D point clouds, as evidenced by several works [3,4,22,27,41]. These efforts have focused on designing rotation invariant properties to achieve consistent accuracies under arbitrary rotations without resorting to rotation augmentation. However, these methods have demonstrated inferior performance when compared to translation invariant approaches, primarily due to the loss of global information during the generation of rotation invariant properties. Recent research efforts [15,30,40] have attempted to overcome this limitation by employing local reference frame (LRF) or local reference axis (LRA) [42] to transform the data into a canonical coordinate system and encode global information. Despite the improvements, their accuracies are still lower than those of state-of-the-art non-rotation invariant methods because of the unstable LRF/LRA and the less descriptive rotation invariant properties that the useful surface information is not well preserved." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.556, + 0.788, + 0.678 + ], + "angle": 0, + "content": "To address the above-mentioned issues, we propose a novel yet effective rotation invariant architecture. Specifically, we construct local triangle surfaces for each reference point of the input data to better capture the local surface structure. On the local surfaces, we design highly expressive rotation invariant surface properties which are then integrated into an attention-augmented convolution operator named RISurConv to extract refined rotation invariant features. Finally, we build up network based on RISurConv for rotation invariant object classification and segmentation. In summary, our main contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.733 + ], + "angle": 0, + "content": "- Rotation Invariant Surface Properties. We construct local triangle surfaces from the reference point and its neighbors, based on which we are able to design highly expressive rotation invariant surface properties;" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.735, + 0.785, + 0.779 + ], + "angle": 0, + "content": "- RISurConv. We integrate the Rotation Invariant Surface Properties into an attention-augmented architecture named RISurConv comprising two self-attention layers to learn and generate refined rotation invariant features;" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.781, + 0.787, + 0.84 + ], + "angle": 0, + "content": "- Extensive experiments on a variety of classification and segmentation tasks. Our approach shows supreme performance, surpassing the state-of-the-art by a large margin under challenging rotation scenarios including an analysis of rotation-invariant features and an ablation study of our neural" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.689, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RISurConv" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "network to provide insights into the key factors contributing to the exceptional performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.202, + 0.396, + 0.218 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.235, + 0.784, + 0.265 + ], + "angle": 0, + "content": "In this section, we review the representative works in 3D point cloud deep learning and rotation invariant learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.266, + 0.785, + 0.386 + ], + "angle": 0, + "content": "3D Point Cloud Deep Learning. 3D point clouds is a more compact and intuitive representation for feature learning. PointNet [23] pioneered a point cloud convolution with global features by max-pooling per-point features from MLPs, and follow-up works [17, 25, 38] have focused on exploring convolution kernels that exploit geometric features [29], adding edges on top of points [33], parameterizing convolution using polynomials [38], and leveraging shape context [37]. Some methods are designed to combine with recurrent neural networks [13] and sequence models [18]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.787, + 0.732 + ], + "angle": 0, + "content": "In recent years, attention mechanism has become increasingly popular in various natural language processing (NLP) [32] and computer vision [6] tasks. With their ability to handle sequential and spatial information, attention mechanism has also shown promise in 3D point cloud processing tasks. Pioneer works that apply attention mechanism in point clouds include Point Transformer [43] and PCT [9], which use a self-attention mechanism to learn local and global features of point clouds. Other works have explored different types of transformers. Inspired by BERT [5], Point-BERT [39] was proposed to pre-train pure Transformer-based models with a Mask Point Modeling task for point cloud classification. Dual Transformer Network [11] aggregated the point-wise and channel-wise multi-head self-attention models simultaneously such that long-range context dependencies can be captured by investigating the point-wise and channel-wise relationships. Overall, these developments show promise in enhancing the efficiency and accuracy of point cloud analysis for a variety of applications. Readers can refer to the survey [10] for a comprehensive overview of 3D point cloud deep learning. However, most of these methods lack the property of rotation invariance which is critical for object classification and segmentation as the object can be viewed at any angle in reality. To address this issue, a common approach is to increase the training data by augmenting it with arbitrary rotations [17, 25, 33]. However, such approach has difficulty in generalizing predictions to unseen rotations, resulting in deteriorated performance. Therefore, it would be desirable to design a specific convolution possessing rotation-invariant features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Rotation Invariant Learning. Various methods have been proposed to address the issue of rotation invariance in feature learning of point clouds. For instance, Rao et al. [27] used a spherical domain to define a rotation-invariant convolution for point clouds. However, the discretized sphere is sensitive to global rotations, which can lead to a notable drop in performance for objects with arbitrary rotations. To overcome this limitation, Poulenard et al. [22] integrated spherical harmonics to their convolution. Similarly, Chen et al. [3] proposed a" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "Z. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.143, + 0.745, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.26, + 0.788, + 0.318 + ], + "angle": 0, + "content": "Fig. 1: Rotation Invariant Surface Property (RISP) construction: Given a point \\( p \\) as the reference point, \\( K \\) (\\( K = 8 \\) in this example) nearest points are selected (middle). For each neighbor \\( x_{i} \\), two adjacent neighbors \\( x_{i-1} \\) and \\( x_{i+1} \\) are used to form two triangular local surfaces (right), based on which rotation invariant properties are constructed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.348, + 0.788, + 0.652 + ], + "angle": 0, + "content": "hierarchical clustering scheme to encode the relative angles between two-point vectors and used vector norm to maintain rotation invariance. Zhang et al. [41] presented a simple convolution named RICnv that operates on handcrafted features built from Euclidean distances and angles that are rotation invariant by nature. However, this approach only considers local features resulting in accuracy degradation. GCACnv [40] addressed this limitation by building a global context-aware convolution based on anchors and Local Reference Frame (LRF) to achieve rotation invariance. RI-GCN [15] learned rotation-invariant local descriptors and applied graph convolutional neural networks to aggregate local features based on LRF. Thomas [30] also relied on LRF and used a multiple alignment scheme to achieve better results. RIF [16] presented a framework to construct both local and global features based on distances, angles, and reference points. To remove the uncertainty of LRF in x and y directions, RICnv++ [42] extracts informative features based on local reference axis (LRA). However, the LRF/LRA and global reference used in these methods may not be stable enough limiting their overall performance. We can clearly see the accuracy discrepancy of the object classification task on ModelNet40 dataset [36]. State-of-the-art translation-invariant convolutions such as PointNet++ [25] or Point Transformer v2 [34] achieve 89%-94% of accuracy while rotation-invariant convolutions only report up to 87%-91% of accuracy [14,42]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.652, + 0.787, + 0.683 + ], + "angle": 0, + "content": "Our target in this paper is not only closing this performance gap but also surpassing the state-of-the-art methods while maintaining rotation invariance." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.711, + 0.605, + 0.731 + ], + "angle": 0, + "content": "3 Rotation Invariant Surface Property" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this section, we detail the Rotation Invariant Surface Property construction which is the first step of our method. Our goal for this step is to design highly expressive rotation invariant properties from underlying local surfaces. Different from previous works that rely on pointwise operations, we construct local surfaces around the reference point to better capture the local surface structure, based on which we then extract more expressive rotation invariant properties." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RISurConv" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.238 + ], + "angle": 0, + "content": "The construction of rotation invariant surface property is shown in Fig. 1. Given a point \\( p \\) as the reference point (red), we get \\( K \\) nearest neighbors to form a local point set. In this case, \\( K \\) is 8. For each neighbor \\( x_{i} \\), two adjacent neighbors, \\( x_{i-1} \\) and \\( x_{i+1} \\), are identified based on the Euclidean distances, forming two triangular local surfaces, as shown in the figure. Then, we construct the rotation invariant surface properties (RISP) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.251, + 0.786, + 0.274 + ], + "angle": 0, + "content": "\\[\n\\operatorname {R I S P} \\left(x _ {i}\\right) = \\left[ L _ {0}, \\phi_ {1}, \\phi_ {2}, \\phi_ {3}, \\phi_ {4}, \\phi_ {5}, \\right. \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.432, + 0.27, + 0.619, + 0.284 + ], + "angle": 0, + "content": "\\[\n\\left. \\alpha_ {1}, \\alpha_ {2}, \\beta_ {1}, \\beta_ {2}, \\theta_ {1}, \\theta_ {2}, \\gamma_ {1}, \\gamma_ {2} \\right]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.296, + 0.784, + 0.341 + ], + "angle": 0, + "content": "where \\( L_{0} \\) measures the distance from reference \\( p \\) to neighbor \\( x_{i} \\), and \\( \\phi_{1} \\) to \\( \\phi_{5} \\) measure the two triangles as well as the relationship between the two triangle surfaces with regard to the edge \\( \\overrightarrow{px_i} \\) in the Euclidean space:" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.369, + 0.786, + 0.384 + ], + "angle": 0, + "content": "\\[\n\\phi_ {1} = \\angle (\\overrightarrow {x _ {i - 1} p}, \\overrightarrow {x _ {i} p}), \\phi_ {2} = \\angle (\\overrightarrow {x _ {i + 1} p}, \\overrightarrow {x _ {i} p}), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.388, + 0.672, + 0.403 + ], + "angle": 0, + "content": "\\[\n\\phi_ {3} = \\angle (\\overrightarrow {x _ {i - 1} x _ {i}}, \\overrightarrow {x _ {i - 1} p}), \\phi_ {4} = \\angle (\\overrightarrow {x _ {i + 1} p}, \\overrightarrow {x _ {i + 1} x _ {i}}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.406, + 0.584, + 0.421 + ], + "angle": 0, + "content": "\\[\n\\phi_ {5} = \\angle (\\overrightarrow {x _ {i + 1} p} \\times \\overrightarrow {x _ {i} p}, \\overrightarrow {x _ {i - 1} p} \\times \\overrightarrow {x _ {i} p}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.433, + 0.784, + 0.478 + ], + "angle": 0, + "content": "while other properties describe the two surfaces in the tangent space, e.g., normal vectors can define the directions in which the surface is bending away from the tangent space:" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.492, + 0.785, + 0.507 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {1} = \\angle (\\overrightarrow {n _ {p}}, \\overrightarrow {x _ {i} p}), \\alpha_ {2} = \\angle (\\overrightarrow {n _ {p}}, \\overrightarrow {x _ {i - 1} p}), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.348, + 0.51, + 0.61, + 0.525 + ], + "angle": 0, + "content": "\\[\n\\beta_ {1} = \\angle (\\vec {n _ {i}}, \\vec {x _ {i}} \\vec {p}), \\beta_ {2} = \\angle (\\vec {n _ {i}}, \\vec {x _ {i - 1}} \\vec {x _ {i}}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.348, + 0.529, + 0.659, + 0.544 + ], + "angle": 0, + "content": "\\[\n\\theta_ {1} = \\angle (\\overrightarrow {n _ {i - 1}}, \\overrightarrow {x _ {i - 1} p}), \\theta_ {2} = \\angle (\\overrightarrow {n _ {i - 1}}, \\overrightarrow {x _ {i - 1} x _ {i} ^ {\\prime}}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.348, + 0.548, + 0.659, + 0.563 + ], + "angle": 0, + "content": "\\[\n\\gamma_ {1} = \\angle (\\overrightarrow {n _ {i + 1}}, \\overrightarrow {x _ {i + 1} x _ {i}}), \\gamma_ {2} = \\angle (\\overrightarrow {n _ {i + 1}}, \\overrightarrow {x _ {i + 1} p}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.575, + 0.784, + 0.62 + ], + "angle": 0, + "content": "RISP is able to fully describe the dual triangles as well as their relationship along different directions. Readers can refer to the supplementary for the mathematical proof of the completeness of RISP defined in Eq. (1)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.635, + 0.788, + 0.74 + ], + "angle": 0, + "content": "Number of local triangle surfaces. Throughout the paper, we define the number of local triangle surfaces as 2. Such setting is due to the fact that we wish to capture local surface as much as possible to extract more useful information to reflect the real shape of the object. So we construct 2 triangles rather than single triangle. However, this does not mean the more the better since more triangles can result in distorted surface and affect the computing efficiency. Please refer to the ablation study section for more details." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.763, + 0.502, + 0.781 + ], + "angle": 0, + "content": "4 The RISurConv Operator" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Based on the Rotation Invariant Surface Properties (RISP), we are able to build up RISurConv. The main steps are detailed in Fig. 2. To start, we utilize a farthest point sampling strategy to generate uniformly distributed representative" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "Z. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.142, + 0.777, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.297, + 0.789, + 0.41 + ], + "angle": 0, + "content": "Fig. 2: RISurConv operator. For a local point set with \\( p \\) as the reference (red), K nearest neighbors are labelled as blue. Then, we compute the Rotation Invariant Surface Properties at each neighbor by constructing local dual triangle surfaces (Sec. 3), which is embedded to a high-dimensional space by a shared multi-layer perceptron (MLP) followed by a self-attention layer to produce refined features. Concatenated with previous layer features (if any), the features of these local points are further passed to MLPs, which are then summarized by maxpooling. To further refine the features, another self-attention layer follows." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.425, + 0.476, + 0.44 + ], + "angle": 0, + "content": "Algorithm 1 RISurConv operator." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.442, + 0.785, + 0.469 + ], + "angle": 0, + "content": "Input: Reference point \\( p \\), point set \\( \\Omega \\), point features \\( \\mathbf{f}_{prev} \\) from previous layer (if any) \nOutput: Convoluted features \\( \\mathbf{f} \\)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.472, + 0.418, + 0.499 + ], + "angle": 0, + "content": "1: \\(\\mathbf{f} \\gets \\{\\mathrm{RISP}(x_i) : \\forall x_i \\in \\Omega\\}\\) (Sec. 3)" + }, + { + "type": "text", + "bbox": [ + 0.451, + 0.472, + 0.785, + 0.499 + ], + "angle": 0, + "content": "* Construct Rotation Invariant Surface Properties" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.5, + 0.333, + 0.514 + ], + "angle": 0, + "content": "2: \\(\\mathbf{f}\\gets \\mathrm{MLP}(\\mathbf{f})\\)" + }, + { + "type": "text", + "bbox": [ + 0.396, + 0.5, + 0.782, + 0.514 + ], + "angle": 0, + "content": "* Embed each feature to a high-dimensional feature space" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.514, + 0.319, + 0.527 + ], + "angle": 0, + "content": "3: \\(\\mathbf{f}\\gets \\mathrm{SA}(\\mathbf{f})\\)" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.514, + 0.782, + 0.528 + ], + "angle": 0, + "content": "* Refine features via self-attention layer" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.528, + 0.34, + 0.555 + ], + "angle": 0, + "content": "4: \\(\\mathbf{f}_{in}\\gets [\\mathbf{f}_{prev},\\mathbf{f}]\\) (if any)" + }, + { + "type": "text", + "bbox": [ + 0.358, + 0.528, + 0.785, + 0.555 + ], + "angle": 0, + "content": "* Concatenate the features from the local and the previous layer" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.555, + 0.356, + 0.569 + ], + "angle": 0, + "content": "5: \\(\\mathbf{f}_{out} \\gets \\mathrm{MLP}(\\mathbf{f}_{in})\\)" + }, + { + "type": "text", + "bbox": [ + 0.639, + 0.555, + 0.781, + 0.569 + ], + "angle": 0, + "content": "* Feature embedding" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.569, + 0.385, + 0.597 + ], + "angle": 0, + "content": "6: \\(\\mathbf{f}_{out}\\gets\\) maxpool \\((\\mathbf{f}_{out})\\) return SA \\((\\mathbf{f}_{out})\\)" + }, + { + "type": "text", + "bbox": [ + 0.657, + 0.57, + 0.786, + 0.583 + ], + "angle": 0, + "content": "* Maxpool features" + }, + { + "type": "text", + "bbox": [ + 0.607, + 0.583, + 0.786, + 0.596 + ], + "angle": 0, + "content": "* Self-attention and return" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.631, + 0.788, + 0.692 + ], + "angle": 0, + "content": "points. From this initial set, we perform K-nearest neighbor searches to obtain local point sets. Let us denote a local point set by \\(\\Omega = \\{x_{i}\\}\\), where \\(x_{i}\\) represents the 3D coordinates of point \\(i\\). We define the attention-augmented convolution operation as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.704, + 0.787, + 0.721 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} (\\Omega) = S A \\left(\\sigma \\left(\\mathcal {A} \\left(\\left\\{\\mathcal {T} \\left(\\mathbf {f} _ {x _ {i}}\\right): \\forall i \\right\\}\\right)\\right)\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.732, + 0.788, + 0.809 + ], + "angle": 0, + "content": "This formula indicates that features of each point in the point set are first transformed by \\(\\mathcal{T}\\) before being aggregated by the aggregation function \\(\\mathcal{A}\\) and passed to an activation function \\(\\sigma\\). SA is a channel-wise self-attention layer used to output refined features. We set the input features to our expressive rotation-invariant features \\(\\mathbf{f}_{x_i} = \\mathrm{RISP}(x_i)\\). We define the transformation function as" + }, + { + "type": "equation", + "bbox": [ + 0.419, + 0.82, + 0.787, + 0.838 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} \\left(\\mathbf {f} _ {x _ {i}}\\right) = \\mathbf {w} _ {i} \\cdot \\mathbf {f} _ {x _ {i}} = \\mathbf {f} _ {x _ {i}} ^ {\\prime} \\tag {5}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RISurConv" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.207 + ], + "angle": 0, + "content": "where \\(\\cdot\\) indicates the element-wise product, and \\(\\mathbf{w}_i\\) is the weight parameter to be learned by the network. Our transformation function is similar to PointNet++. A popular choice of the aggregation function \\(\\mathcal{A}\\) is maxpooling, which supports permutation invariance of the input point features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.787, + 0.403 + ], + "angle": 0, + "content": "To proceed with feature learning and be invariant to point permutation, we construct surfaces for each neighbor from which we construct RISPs. In Fig. 2, there are eight neighbors, so the resulted RISPs are in size of \\(8 \\times 14\\). Here, 14 is the length of the RISP as defined in Eq. (1) and 8 is the neighborhood size. Since RISPs are used as the input which are already rotation invariant, RISurConv is rotation invariant by nature. We then embed the RISPs into feature space using two layers of MLPs followed by a self-attention (SA) layer for feature refinement. We concatenate the refined features with previous layer features (if any), and embed the concatenated features into higher dimensional space via MLPs again. After the maxpooling, the aggregated features are passed into another SA layer again for further refinement. The detailed steps are shown in the Algorithm 1. Here, the two SA layers used in RISurConv are the standard SA module of the transformer [32]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.788, + 0.588 + ], + "angle": 0, + "content": "Self Attention Layers. In RISurConv there are two Self-Attention (SA) layers by which the extracted features are enhanced. Below, we provide a more detailed illustration of the structure of the SA layers. The SA structure is the standard SA module of the famous transformer [32] (Attention is all you need) which runs twice in RISurConv. The first SA is for the K neighbors, and the second SA works for the N representative points. The input tensor shape is \\([B,N,K,C_0]\\) for the first SA which goes into linear layers and outputs \\(Q,K\\) and \\(V\\) respectively in shape of \\([B,N,K,C_0]\\). According to \\(\\text{att}(Q,K,V) = \\text{Softmax}(QK^{\\top} / \\text{sqrt}(d_k))V\\), we have the 1st attention score in \\([B,N,K,K]\\). By multiplying with \\(V\\), we get the refined feature in \\([B,N,K,C_0]\\). The same goes with the second SA where the input shape is \\([B,N,C]\\) and the attention score shape is \\([B,N,N]\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.61, + 0.461, + 0.627 + ], + "angle": 0, + "content": "5 RISurConv Networks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We employed the RISurConv to develop neural networks for object classification and segmentation, as illustrated in Fig. 3. Our classification network follows a standard architecture similar to the single-scale grouping version of PointNet++, consisting of five consecutive layers of RISurConv followed by a transformer encoder [32] with 8 heads to enhance the extracted features. Finally, we connect the network to fully connected layers to produce the probability map. One of the major advantages of RISurConv is its ability to handle arbitrary rotation and point orders, which enables us to place each RISurConv layer consecutively without the need for complex preprocessing steps. Additionally, we apply batch normalization and ReLU activation to each convolution layer by default. The segmentation network utilizes an encoder-decoder structure similar to the U-Net [28]. We adopt the same definition of deconvolution as RISurConv. The key distinction lies in that the convolution produces a point subset with a greater" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "Z. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.143, + 0.784, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.289, + 0.788, + 0.348 + ], + "angle": 0, + "content": "Fig. 3: Our neural network architecture comprises five RISurConv layers to extract rotation invariant features followed by a Transformer Encoder to enhance the learnt features before fully connected layers for object classification. We add a decoder with skip connections for segmentation task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.788, + 0.442 + ], + "angle": 0, + "content": "number of feature channels and less points while the deconvolution outputs to a point set with more points compared to the input with fewer feature channels. Please refer to the supplementary for more details regarding the network architecture." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.477, + 0.377, + 0.495 + ], + "angle": 0, + "content": "6 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.787, + 0.581 + ], + "angle": 0, + "content": "We report our evaluation results in this section. Our network is implemented in PyTorch using an Adam optimizer with initial learning rate of 0.001 for both classification and segmentation. Our training is executed on a computer with an Intel(R) Core(TM) i7-10700K CPU equipped with a NVIDIA GTX 3090 GPU." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.789, + 0.84 + ], + "angle": 0, + "content": "We evaluated the performance of our method extensively on a series of classification and segmentation benchmarks. To train our model, we utilized 450 epochs for classification and 250 epochs for segmentation for most datasets except the Car and Chair categories of the fine-grained FG3D dataset due to small size and subtle variation where we applied 900 epochs. Our method can achieve convergence within 11 hours for classification and 24 hours for segmentation. To assess the robustness of our method, we followed the experimental design proposed by [8] and conducted experiments in three cases. The first case involved training and testing with data augmented with rotation about the gravity axis \\((\\mathrm{z} / \\mathrm{z})\\), which is a common approach for evaluating translation-invariant point cloud learning methods. The second case involved training and testing with data augmented with arbitrary SO3 rotations (SO3/SO3), and the third case involved training with data by z-rotations and testing with data by SO3 rotations (z/SO3), which are used to evaluate rotation invariance. A rotation-invariant method should produce consistent results across all three cases. Our method demonstrated superior performance across all three cases, highlighting its robustness and effectiveness." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RISurConv" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.55, + 0.162 + ], + "angle": 0, + "content": "6.1 Human-made Object Classification" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.788, + 0.303 + ], + "angle": 0, + "content": "ModelNet40 [36] that comprises 9843 training models and 2468 test models, divided into 40 categories is one of the most popular datasets for point cloud classification evaluation. The input data size is 1024 with each point equipped with (x; y; z; nx; ny; nz) which are 3D coordinates and 3D normals in the Euclidean space. Note that normals are optional to our method. When normals are not available (w/o n), we use weighted eigen vector corresponding to the smallest eigen value as the normal. This strategy has been used in SHOT and RICConv++ [42] which is highly robust to noise and efficient to compute." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.304, + 0.788, + 0.365 + ], + "angle": 0, + "content": "The experimental results are shown in Tab. 1. We use two criteria for evaluation: overall accuracy and accuracy standard deviation (Std.). True rotation invariant methods are expected to be unaffected by the rotations present in the training and testing data, resulting in low accuracy deviation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.788, + 0.472 + ], + "angle": 0, + "content": "From Tab. 1, we see that our method surpasses the state-of-the-art performance in all cases while maintaining zero accuracy deviation. It is worth noting that our method surpasses both rotation invariant and non-rotation invariant methods. To the best of our knowledge, this is the first work that achieves such high accuracy. It outperforms the state-of-the-art rotation invariant method RIConv++ by \\(4.7\\%\\), and outperforms the state-of-the-art non-rotation invariant method Point Transformer v2 by \\(1.8\\%\\) under \\(\\mathrm{z / z}\\) case." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.504, + 0.533, + 0.519 + ], + "angle": 0, + "content": "6.2 Real World Object Classification" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.461, + 0.748 + ], + "angle": 0, + "content": "We also evaluate the classification performance on a real-world 3D point cloud dataset ScanObjectNN [31]. It is composed of 2902 point clouds categorized into 15 categories sampled from real world indoor scenes. For our evaluation, we use the processed files and choose the hardest variant PB_T50_RS with \\(50\\%\\) bounding box translation, rotation around the gravity axis, and random scaling that result in rotated and partial data. The results are shown" + }, + { + "type": "table_caption", + "bbox": [ + 0.47, + 0.541, + 0.788, + 0.583 + ], + "angle": 0, + "content": "Table 2: Comparisons of real world 3D point cloud classification on hardest variant of ScanObjectNN dataset (Overall Accuracy)." + }, + { + "type": "table", + "bbox": [ + 0.477, + 0.584, + 0.782, + 0.746 + ], + "angle": 0, + "content": "
MethodPB_T50_RS
z/zSO3/SO3z/SO3
PointNet [23]68.242.217.1
PointNet++ [25]77.960.115.8
PointCNN [17]78.551.814.9
DGCNN [33]78.163.416.1
RICov [41]68.168.368.3
GCACov [40]69.870.069.8
RICov++ [42]80.380.380.3
Ours (w/o normal)93.193.193.1
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "in Tab. 2. We see that RISurConv surpasses all the compared methods by a large margin. Particularly, our method significantly outperforms the state-of-the-art rotation invariant approach RICov++ [42] by \\(12.8\\%\\). This verifies that our method is effective for both synthetic and real world data. Note that we only test the 'w/o normal' case as the normal vectors are not provided in the processed files of this dataset." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "Z. Zhang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 1: Comparisons of the classification accuracy (\\%) on the ModelNet40 dataset. On average, our method has the best accuracy and lowest accuracy deviation in all cases (Overall Accuracy)." + }, + { + "type": "table", + "bbox": [ + 0.244, + 0.199, + 0.766, + 0.501 + ], + "angle": 0, + "content": "
MethodFormatInput Size Params.z/z↑SO3/ SO3↑z/ SO3↑Std.↓
TraditionalVoxNet [21]voxel3030.90M83.087.3-3.0
SubVolSup [24]voxel30317.00M88.582.736.628.4
PointNet [23]xyz1024 × 33.50M87.080.321.641.0
PointCNN [17]xyz1024 × 30.60M91.384.541.227.2
PointNet++ [25]xyz + nor1024 × 61.40M89.385.028.633.8
DGCNN [33]xyz1024 × 31.84M92.281.120.638.5
RS-CNN [20]xyz1024 × 31.41M90.382.648.722.1
Pt Transformer [43]xyz1024 × 3-93.785.950.119.1
Pt Transformer v2 [34]xyz1024 × 3-94.288.351.823.0
Rotation-invariantSpherical CNN [7]voxel2 × 6420.50M88.986.978.65.5
RICConv [41]xyz1024 × 30.70M86.586.486.40.1
SPHNet [22]xyz1024 × 32.90M87.087.686.60.5
SFCNN [27]xyz1024 × 3-91.490.184.83.5
ClusterNet [3]xyz1024 × 31.40M87.187.187.10.0
GCACov [40]xyz1024 × 30.41M89.089.289.10.0
RIF [16]xyz1024 × 3-89.489.389.40.0
RI-GCN [15]xyz + nor1024 × 64.38M91.091.091.00.0
RICConv++ [42]xyz1024 × 30.40M91.291.291.20.0
RICConv++ [42]xyz + nor1024 × 60.40M91.391.391.30.0
Ours (w/o normal)xyz1024 × 314.0M95.695.695.60.0
Ours (w/ normal)xyz + nor1024 × 614.0M96.096.096.00.0
" + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.532, + 0.549, + 0.547 + ], + "angle": 0, + "content": "6.3 Fine-Grained Object Classification" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.788, + 0.777 + ], + "angle": 0, + "content": "Fine-grained object classification is more challenging because the differences between subcategories are subtle and require a high level of precision in distinguishing them. We conduct experiments on the three categories from FD3D dataset [19]: Airplane, Chair and Car. To the best of our knowledge, this is the first work to test the performance of rotation invariant methods on fine-grained 3D point cloud dataset. The class accuracies are shown in Tab. 3. Again, our method outperforms the state-of-the-art approaches by large margins for all rotation cases and all the categories. Specifically, our method improves non-rotation invariant method FG3D-Net by \\(2.1\\%\\), \\(2.7\\%\\), and \\(4.5\\%\\) for Airplane, Chair, and Car categories respectively under the \\(\\mathrm{z} / \\mathrm{z}\\) rotation case, and the accuracy consistency is well preserved. Compared to the rotation invariant methods, our method outperforms the latest RICov++ by \\(3.6\\%\\), \\(5.1\\%\\), and \\(9.2\\%\\) on the three categories. Such supreme performance shows that features extracted by RISurConv are not only rotation invariant but also highly expressive." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.806, + 0.53, + 0.822 + ], + "angle": 0, + "content": "6.4 Part Segmentation on ShapeNet" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RISurConv" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.784, + 0.174 + ], + "angle": 0, + "content": "Table 3: Comparisons of fine-grained 3D point cloud classification on FG3D dataset (Class Accuracy)." + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.185, + 0.776, + 0.405 + ], + "angle": 0, + "content": "
MethodModalityAirplaneChairCar
z/zSO3/SO3z/SO3z/zSO3/SO3z/SO3z/zSO3/SO3z/SO3
VoxNetVoxel79.257.528.773.354.716.768.242.217.1
MVCNNView82.666.921.976.363.714.671.951.814.9
View-GCNView87.456.622.779.750.219.773.744.418.1
RotationNetView89.150.829.878.536.218.472.545.328.3
Part4FeatureView82.649.132.277.152.230.273.435.023.8
FG3D-NetView89.461.126.980.043.218.474.056.420.8
PointNetPoint82.755.529.772.133.918.268.132.116.4
PointNet++Point87.357.125.678.140.416.070.350.420.8
RS-CNNPoint82.845.932.275.150.718.671.231.813.9
DGCNNPoint88.460.622.771.752.319.765.353.418.1
RICConvPoint85.885.885.876.376.376.367.367.367.3
RICConv++Point87.987.987.977.677.677.669.369.369.3
OursPoint91.591.591.582.782.782.778.578.578.5
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.447, + 0.645 + ], + "angle": 0, + "content": "To evaluate the segmentation performance, we employed the ShapeNet dataset comprising 16880 CAD models distributed across 16 distinct categories. These models are annotated with 2 to 6 parts each, culminating in a dataset of 50 object parts. We follow the standard train/test split with 14006 models for training and 2874 models for testing, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.447, + 0.779 + ], + "angle": 0, + "content": "The evaluation results are shown in Tab. 4. Our method outperforms both traditional translation-invariant and latest rotation invariant methods significantly in the SO3/SO3 and z/SO3 scenario. Our method outperforms the RICov++ by" + }, + { + "type": "table_caption", + "bbox": [ + 0.459, + 0.466, + 0.785, + 0.534 + ], + "angle": 0, + "content": "Table 4: Comparisons of object part segmentation performed on ShapeNet dataset. The mean per-class IoU (mIoU, %) is used to measure the accuracy under two challenging rotation modes: SO3/SO3 and z/SO3." + }, + { + "type": "table", + "bbox": [ + 0.475, + 0.537, + 0.777, + 0.766 + ], + "angle": 0, + "content": "
MethodInputSO3/SO3 z/SO3
PointNet [23]xyz74.437.8
PointNet++ [25]xyz+nor76.748.2
PointCNN [17]xyz71.434.7
DGCNN [33]xyz73.337.4
RS-CNN [20]xyz72.536.5
SpiderCNNxyz+nor72.342.9
RICConv [41]xyz75.575.3
GCACConv [40]xyz77.377.2
RI-GCN [15]xyz77.077.0
RIF [16]xyz79.479.2
RICConv++ [42]xyz80.380.3
RICConv++ [42]xyz+nor80.580.5
Ours (w/o normal)xyz81.381.3
Ours (w/ normal)xyz+nor81.581.5
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.84 + ], + "angle": 0, + "content": "1.0% mIoU. This result aligns well with the performance reported in the object classification task. We also show the qualitative results by error maps in Fig. 4. The wrong segmentation points are plotted as red. It clearly shows that our predictions are the closest to the ground truth." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "Z. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.143, + 0.753, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.402, + 0.685, + 0.417 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparisons (Red indicates wrong)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.447, + 0.402, + 0.461 + ], + "angle": 0, + "content": "6.5 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.785, + 0.505 + ], + "angle": 0, + "content": "We ablate some vital designs of our method on ModelNet40 [36] for an insightful exploration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.788, + 0.673 + ], + "angle": 0, + "content": "RISP Design. We first test the performance by turning on/off and adding different rotation invariant components used in the RISP construction (Algorithm 1). The results are shown in Tab. 5. Model A is our baseline setting with all rotation invariant features activated. Model B has only angle features but still achieves relatively high accuracy. Model C turns off both \\( L_{0} \\) and \\( \\phi_{1\\dots 5} \\) with accuracy decreases to \\( 90.9\\% \\). This means that the distance \\( L_{0} \\) is not as important as angles. In Model D, we turn off the angle features on the tangent direction with only \\( L_{0} \\) and \\( \\phi_{1\\dots 5} \\) are kept which is more like RICnv [41]. Compared to Model A, it shows that our proposed features is more effective than those by RICnv, and the improvement is explained by the additional consideration of the relations among the neighbor points. Model E adds more rotation invariant" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.699, + 0.486, + 0.728 + ], + "angle": 0, + "content": "Table 5: Ablation study on the RISP design." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.74, + 0.482, + 0.837 + ], + "angle": 0, + "content": "
ModelL0φ1...5α1,2β1,2θ1,2γ1,2λμAcc.
A✓✓✓96.0
B✓✓95.5
C✓90.9
D✓✓88.2
E✓✓✓✓95.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.516, + 0.699, + 0.785, + 0.727 + ], + "angle": 0, + "content": "Table 6: Ablation study on the Self-Attention module." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.74, + 0.784, + 0.837 + ], + "angle": 0, + "content": "
ModelSA1SA2TransformerEncoderAcc.
A✓✓✓96.0
B✓✓95.6
C✓✓95.2
D✓✓94.3
E92.8
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RISurConv" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "features. \\(\\lambda\\) and \\(\\mu\\) represent the angles \\(\\angle (\\pmb{n}_{\\pmb{p}},\\pmb{n}_{i + 1})\\) and \\(\\angle (\\pmb {n}_p,\\pmb {n}_{i - 1})\\) respectively. The results show that the accuracy does not increase. This verifies the completeness of the proposed RISP." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.197, + 0.785, + 0.408 + ], + "angle": 0, + "content": "Self-Attention Effects. We employ multiple self-attention (SA) modules to help produce refined rotation invariant features. So it is necessary to analyse the effects of different SA modules. In RISurConv, there are two SA modules and we name them as SA1 and SA2 respectively. Before fully connected layers, there is a Transformer Encoder module. We test these three modules by turning off one of them each time. From Tab. 6, we see that when removing the first SA module in RISurConv, the performance drops a bit, while removing the second SA module also decreases the accuracy. This indicates that both SA1 and SA2 are important for the success of our method. In addition, we also test the performance when Transformer Encoder is removed. The overall accuracy decreases to \\(94.3\\%\\) but it is still higher than most existing methods. In model E, we test the performance when all SA modules are removed. The accuracy decreases to \\(92.8\\%\\) which is still better than the state-of-the-art rotation invariant method [42] thanks to the well designed RISPs which capture sufficient surface structures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.409, + 0.431, + 0.603 + ], + "angle": 0, + "content": "We also visualize the global feature vector before classification as shown in Fig. 5. We normalize the feature values to [0,1] and plot the histograms. We can see that by turning on the SA modules, the feature values become more evenly distributed and have less zeros because SA mechanism can reweight the features making the features more effective." + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.422, + 0.784, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.554, + 0.785, + 0.583 + ], + "angle": 0, + "content": "Fig. 5: Histogram comparison for normalized feature values without and with self-attention layers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.609, + 0.415, + 0.774 + ], + "angle": 0, + "content": "Network Efficiency. In the experiments, we acknowledge that our method employs a higher number of parameters. Thus, it is imperative to conduct a comprehensive analysis of network efficiency during both the training and testing stages. We list the trainable parameters, FLOPs" + }, + { + "type": "table_caption", + "bbox": [ + 0.424, + 0.612, + 0.787, + 0.655 + ], + "angle": 0, + "content": "Table 7: Trainable parameters, FLOPs and timing comparisons on ModelNet40. All are tested on the same platform. Time is measured per batch (batch size=16)." + }, + { + "type": "table", + "bbox": [ + 0.434, + 0.656, + 0.779, + 0.771 + ], + "angle": 0, + "content": "
MethodsParamsFLOPsTime (Train/Infer)
PointNet++ [25]1.41M0.86G0.145s / 0.129s
RICnv [41]0.70M0.92G0.121s / 0.092s
RI-GCN [15]4.19M1.24G0.148s / 0.112s
RICnv++ [42]0.42M0.72G0.074s / 0.029s
Ours13.96M1.12G0.114s / 0.049s
Ours w/o TR1.35M1.11G0.105s / 0.047s
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.776, + 0.784, + 0.807 + ], + "angle": 0, + "content": "and running time comparisons in Tab. 7. We compare the performance with four recent works: PointNet++ [25], RICnv [41], RI-GCN [15], RICnv++ [42]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "From the results, we can see that the number of parameters mainly comes from the transformer architectures used in RISurConv. When we turn off all" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "Z. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "the transformer component, our method only takes up 1.35M parameters. More parameters do not mean longer inference time. It depends on lots of things including depth, parameter, number of operations etc. Our approach runs faster than most existing methods. It takes more time for training because more epochs are needed for self attention layers to converge. This is a common phenomenon of transformer based methods. The accuracy is also not due to the complexity model. As is shown in the ablation studies, by removing transformer encoder, it still achieves state-of-the-art performance due to the effectiveness of RISurConv." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.557, + 0.434 + ], + "angle": 0, + "content": "Number of Local Surfaces. Local triangle surfaces used in Sec. 3 are the base for rotation invariant surface property (RISP) construction. Thus, it is important to analyse the effects of different number of surfaces. We conduct performance tests by setting the number of local surfaces to 1, 2, 3, and 4, resulting in overall accuracies of 94.7, 96.0, 91.9, and 89.8 respectively. It worth noting that for more surfaces, we can define more RISPs. From the results, we see the accuracy drops a bit when single" + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.281, + 0.784, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.4, + 0.788, + 0.429 + ], + "angle": 0, + "content": "Fig. 6: Distorted surface caused by 3 triangles." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.585 + ], + "angle": 0, + "content": "surface is constructed. This is because smaller area cannot capture sufficient surface structures and less effective features can lower the performance. When we increase the number of surfaces to 3 and 4, the performance decreases. This shows that more surfaces can result in inferior performance because we construct triangle surfaces just to approximate the true surface. More surfaces may mess up the surface structure and affect the performance. Take Fig. 6 for example, suppose we are constructing the local surfaces \\((\\triangle px_{i}x_{i + 1}\\triangle px_{i}x_{i - 1})\\) with regard to the neighbor point \\(x_{i}\\), adding one more triangle \\(\\triangle px_{i}x_{i + 2}\\) will result in distorted/non-manifold surfaces. Thus, we set number of surfaces as 2 throughout the paper." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.61, + 0.355, + 0.626 + ], + "angle": 0, + "content": "7 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.787, + 0.704 + ], + "angle": 0, + "content": "**Limitation.** Though effective, RISurConv may require longer training time (e.g. more epochs) to converge due to large number of training parameters (Tab. 1). This is a common issue of transformer. But this does not affect the inference speed as shown in the ablation study (Tab. 7)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Conclusion. We have presented a new framework for achieving rotation invariance on 3D point cloud. We construct local triangle surfaces to better represent the local surface structures based on which we design highly expressive rotation invariant surface properties. We integrate the properties into an architecture named RISurConv to extract refined rotation invariant features. Based on RISurConv, we finally build up effective rotation invariant neural networks for 3D point cloud classification and segmentation with supreme performance achieved not only closing the performance gap with non-rotation invariant approaches but also surpassing the state-of-the-art methods by a large margin." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RISurConv" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.24 + ], + "angle": 0, + "content": "This research is supported by the Singapore Ministry of Education (MOE) Academic Research Fund (AcRF) Tier 1 grant (MSS23C010), Ningbo 2025 Science & Technology Innovation Major Project (No. 2022Z072, No.2023Z044), and Key Research & Development Plan of Zhejiang Province (2024C01017)." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.264, + 0.324, + 0.279 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.296, + 0.785, + 0.338 + ], + "angle": 0, + "content": "1. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3d semantic parsing of large-scale indoor spaces. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1534-1543 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.339, + 0.785, + 0.38 + ], + "angle": 0, + "content": "2. Chang, A.X., Funkhouser, T.A., Guibas, L.J., Hanrahan, P., Huang, Q.X., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., Xiao, J., Yi, L., Yu, F.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.381, + 0.785, + 0.422 + ], + "angle": 0, + "content": "3. Chen, C., Li, G., Xu, R., Chen, T., Wang, M., Lin, L.: Clusternet: Deep hierarchical cluster network with rigorously rotation-invariant representation for point cloud analysis. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4994-5002 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.423, + 0.785, + 0.45 + ], + "angle": 0, + "content": "4. Deng, H., Birdal, T., Ilic, S.: Ppf-foldnet: Unsupervised learning of rotation invariant 3d local descriptors. In: Eur. Conf. Comput. Vis. pp. 602-618 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.451, + 0.785, + 0.478 + ], + "angle": 0, + "content": "5. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: NAACL-HLT (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.479, + 0.785, + 0.533 + ], + "angle": 0, + "content": "6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Int. Conf. Learn. Represent. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.534, + 0.785, + 0.575 + ], + "angle": 0, + "content": "7. Esteves, C., Allen-Blanchette, C., Makadia, A., Daniilidis, K.: Learning so (3) equivariant representations with spherical cnns. In: Eur. Conf. Comput. Vis. pp. 52-68 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.576, + 0.785, + 0.603 + ], + "angle": 0, + "content": "8. Esteves, C., Xu, Y., Allen-Blanchette, C., Daniilidis, K.: Equivariant multi-view networks. In: Int. Conf. Comput. Vis. pp. 1568-1577 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.604, + 0.785, + 0.631 + ], + "angle": 0, + "content": "9. Guo, M.H., Cai, J.X., Liu, Z.N., Mu, T.J., Martin, R.R., Hu, S.M.: Pct: Point cloud transformer. Computational Visual Media 7, 187-199 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.632, + 0.785, + 0.672 + ], + "angle": 0, + "content": "10. Guo, Y., Wang, H., Hu, Q., Liu, H., Liu, L., Bennamoun, M.: Deep learning for 3d point clouds: A survey. IEEE Trans. Pattern Anal. Mach. Intell. 43(12), 4338-4364 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.673, + 0.785, + 0.701 + ], + "angle": 0, + "content": "1. Han, X.F., Jin, Y.F., Cheng, H.X., Xiao, G.Q.: Dual transformer for point cloud analysis. IEEE Trans. Multimedia (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.785, + 0.742 + ], + "angle": 0, + "content": "2. Hua, B.S., Pham, Q.H., Nguyen, D.T., Tran, M.K., Yu, L.F., Yeung, S.K.: Scenenn: A scene meshes dataset with annotations. In: International Conference on 3D Vision. pp. 92-101 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.743, + 0.785, + 0.771 + ], + "angle": 0, + "content": "3. Huang, Q., Wang, W., Neumann, U.: Recurrent slice networks for 3d segmentation of point clouds. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2626-2635 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "4. Kim, J., Jung, W., Kim, H., Lee, J.: Cynnn: a rotation invariant cnn using polar mapping and cylindrical convolution layers. arXiv preprint arXiv:2007.10588 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "5. Kim, S., Park, J., Han, B.: Rotation-invariant local-to-global representation learning for 3d point cloud. Adv. Neural Inform. Process. Syst. 33, 8174-8185 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.296, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.128 + ], + "angle": 0, + "content": "Z. Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "16. Li, X., Li, R., Chen, G., Fu, C.W., Cohen-Or, D., Heng, P.A.: A rotation-invariant framework for deep point cloud analysis. IEEE Trans. Vis. Comput. Graph. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.202 + ], + "angle": 0, + "content": "17. Li, Y., Bu, R., Sun, M., Wu, W., Di, X., Chen, B.: Pointcnn: Convolution on x-transformed points. In: Adv. Neural Inform. Process. Syst. pp. 820-830 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.785, + 0.242 + ], + "angle": 0, + "content": "18. Liu, X., Han, Z., Liu, Y.S., Zwicker, M.: Point2sequence: Learning the shape representation of 3d point clouds with an attention-based sequence to sequence network. In: AAAI. vol. 33, pp. 8778-8785 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.243, + 0.785, + 0.27 + ], + "angle": 0, + "content": "19. Liu, X., Han, Z., Liu, Y.S., Zwicker, M.: Fine-grained 3d shape classification with hierarchical part-view attention. IEEE Trans. Image Process. 30, 1744-1758 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.785, + 0.31 + ], + "angle": 0, + "content": "20. Liu, Y., Fan, B., Xiang, S., Pan, C.: Relation-shape convolutional neural network for point cloud analysis. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 8895-8904 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.785, + 0.351 + ], + "angle": 0, + "content": "21. Maturana, D., Scherer, S.: Voxnet: A 3d convolutional neural network for real-time object recognition. In: 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 922-928 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.785, + 0.392 + ], + "angle": 0, + "content": "22. Poulenard, A., Rakotosaona, M.J., Ponty, Y., Ovsjanikov, M.: Effective rotation-invariant point cnn with spherical harmonics kernels. International Conference on 3D Vision (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.785, + 0.433 + ], + "angle": 0, + "content": "23. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 652-660 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.785, + 0.473 + ], + "angle": 0, + "content": "24. Qi, C.R., Su, H., Nießner, M., Dai, A., Yan, M., Guibas, L.J.: Volumetric and multi-view cnns for object classification on 3d data. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5648-5656 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.785, + 0.514 + ], + "angle": 0, + "content": "25. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: Adv. Neural Inform. Process. Syst. pp. 5105-5114 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.785, + 0.555 + ], + "angle": 0, + "content": "26. Qian, G., Li, Y., Peng, H., Mai, J., Hammoud, H., Elhoseiny, M., Ghanem, B.: Pointnext: Revisiting pointnet++ with improved training and scaling strategies. Adv. Neural Inform. Process. Syst. 35, 23192-23204 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.556, + 0.785, + 0.582 + ], + "angle": 0, + "content": "27. Rao, Y., Lu, J., Zhou, J.: Spherical fractal convolutional neural networks for point cloud recognition. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.785, + 0.623 + ], + "angle": 0, + "content": "28. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: International Conference on Medical image computing and computer-assisted intervention. pp. 234-241 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.624, + 0.785, + 0.663 + ], + "angle": 0, + "content": "29. Shen, Y., Feng, C., Yang, Y., Tian, D.: Mining point cloud local structures by kernel correlation and graph pooling. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4548-4557 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.785, + 0.69 + ], + "angle": 0, + "content": "30. Thomas, H.: Rotation-invariant point convolution with multiple equivariant alignments. In: International Conference on 3D Vision (3DV). pp. 504-513 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.785, + 0.731 + ], + "angle": 0, + "content": "31. Uy, M.A., Pham, Q.H., Hua, B.S., Nguyen, D.T., Yeung, S.K.: Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In: Int. Conf. Comput. Vis. (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.785, + 0.771 + ], + "angle": 0, + "content": "32. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Adv. Neural Inform. Process. Syst. 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "33. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph cnn for learning on point clouds. ACM Trans. Graph. (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "34. Wu, X., Lao, Y., Jiang, L., Liu, X., Zhao, H.: Point transformer v2: Grouped vector attention and partition-based pooling. Adv. Neural Inform. Process. Syst. 35, 33330-33342 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RISurConv" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "35. Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., Xiao, J.: 3d shapenets: A deep representation for volumetric shapes. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1912-1920 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "36. Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., Xiao, J.: 3d shapenets: A deep representation for volumetric shapes. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1912-1920 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.259 + ], + "angle": 0, + "content": "37. Xie, S., Liu, S., Chen, Z., Tu, Z.: Attentional shapecontextnet for point cloud recognition. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4606-4615 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.3 + ], + "angle": 0, + "content": "38. Xu, Y., Fan, T., Xu, M., Zeng, L., Qiao, Y.: SpiderCNN: Deep learning on point sets with parameterized convolutional filters. In: Eur. Conf. Comput. Vis. pp. 87-102 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.342 + ], + "angle": 0, + "content": "39. Yu, X., Tang, L., Rao, Y., Huang, T., Zhou, J., Lu, J.: Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 19313-19322 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.383 + ], + "angle": 0, + "content": "40. Zhang, Z., Hua, B.S., Chen, W., Tian, Y., Yeung, S.K.: Global context aware convolutions for 3d point cloud understanding. In: International Conference on 3D Vision (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.383, + 0.788, + 0.425 + ], + "angle": 0, + "content": "41. Zhang, Z., Hua, B.S., Rosen, D.W., Yeung, S.K.: Rotation invariant convolutions for 3d point clouds deep learning. In: International Conference on 3D Vision. pp. 204-213 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.425, + 0.788, + 0.467 + ], + "angle": 0, + "content": "42. Zhang, Z., Hua, B.S., Yeung, S.K.: Riconv++: Effective rotation invariant convolutions for 3d point clouds deep learning. Int. J. Comput. Vis. 130(5), 1228-1243 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.467, + 0.788, + 0.494 + ], + "angle": 0, + "content": "43. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: Int. Conf. Comput. Vis. pp. 16259-16268 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.494, + 0.788, + 0.536 + ], + "angle": 0, + "content": "44. Zhao, Y., Birdal, T., Lenssen, J.E., Menegatti, E., Guibas, L., Tombari, F.: Quaternion equivariant capsule networks for 3d point clouds. In: Eur. Conf. Comput. Vis. pp. 1-19. Springer (2020)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.536 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/RISurConv_ Rotation Invariant Surface Attention-Augmented Convolutions for 3D Point Cloud Classification and Segmentation/85a1ebc6-6b7d-4e16-8cd8-c420684dc80d_origin.pdf b/2024/RISurConv_ Rotation Invariant Surface Attention-Augmented Convolutions for 3D Point Cloud Classification and Segmentation/85a1ebc6-6b7d-4e16-8cd8-c420684dc80d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..47e5280d49324ca94e923406662e13c28cec60c5 --- /dev/null +++ b/2024/RISurConv_ Rotation Invariant Surface Attention-Augmented Convolutions for 3D Point Cloud Classification and Segmentation/85a1ebc6-6b7d-4e16-8cd8-c420684dc80d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:904ce27ab52cd448695a33f2a1d6060290860baac95104c26282e09f1d76a8c8 +size 2445461 diff --git a/2024/RISurConv_ Rotation Invariant Surface Attention-Augmented Convolutions for 3D Point Cloud Classification and Segmentation/full.md b/2024/RISurConv_ Rotation Invariant Surface Attention-Augmented Convolutions for 3D Point Cloud Classification and Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1731349f29bfe49c2666e4cff7927198f609957b --- /dev/null +++ b/2024/RISurConv_ Rotation Invariant Surface Attention-Augmented Convolutions for 3D Point Cloud Classification and Segmentation/full.md @@ -0,0 +1,312 @@ +# RISurConv: Rotation Invariant Surface Attention-Augmented Convolutions for 3D Point Cloud Classification and Segmentation + +Zhiyuan Zhang $^{1}$ , Licheng Yang $^{2\star}$ , and Zhiyu Xiang $^{2\oplus}$ + +$^{1}$ School of Computing and Information Systems, Singapore Management University https://github.com/cszyzhang/RISurConv +$^{2}$ College of Information Science and Electronic Engineering, Zhejiang University + +Abstract. Despite the progress on 3D point cloud deep learning, most prior works focus on learning features that are invariant to translation and point permutation, and very limited efforts have been devoted for rotation invariant property. Several recent studies achieve rotation invariance at the cost of lower accuracies. In this work, we close this gap by proposing a novel yet effective rotation invariant architecture for 3D point cloud classification and segmentation. Instead of traditional pointwise operations, we construct local triangle surfaces to capture more detailed surface structure, based on which we can extract highly expressive rotation invariant surface properties which are then integrated into an attention-augmented convolution operator named RISurConv to generate refined attention features via self-attention layers. Based on RISurConv we build an effective neural network for 3D point cloud analysis that is invariant to arbitrary rotations while maintaining high accuracy. We verify the performance on various benchmarks with supreme results obtained surpassing the previous state-of-the-art by a large margin. We achieve an overall accuracy of $96.0\%$ $(+4.7\%)$ on ModelNet40, $93.1\%$ $(+12.8\%)$ on ScanObjectNN, and class accuracies of $91.5\%$ $(+3.6\%)$ , $82.7\%$ $(+5.1\%)$ , and $78.5\%$ $(+9.2\%)$ on the three categories of the FG3D dataset for the fine-grained classification task. Additionally, we achieve $81.5\%$ $(+1.0\%)$ mIoU on ShapeNet for the segmentation task. + +Keywords: Point cloud $\cdot$ Rotation invariant $\cdot$ Attention + +# 1 Introduction + +Point cloud has become the most promising 3D data representation for a wide range of immersive applications from robot navigation to autonomous driving. The increasing availability of 3D sensors and the emergence of large and mid-scale point cloud datasets [1,2,12,19,31,35] have spurred significant research in this area, leading to the development of numerous deep learning models for point cloud classification and segmentation, which are the fundamentals for various computer vision tasks. + +However, analyzing 3D point clouds remains challenging, mainly due to the irregular nature of point clouds and their inherent invariances such as translation, point permutation, and rotation. While significant progress has been made in learning translation and permutation-invariant features [17, 23, 25, 26, 44], achieving rotation invariance in point cloud convolution has been a relatively unexplored area. + +Rotation invariance is essential for 3D object classification and segmentation as objects can be viewed from different viewpoints, leading to variations in their orientation. Prior approaches [23,25,33] usually rely on rotation augmentation in training stage to relief the rotation affections for testing. However, such scheme is less effective as 3D data has more degrees of freedom. A rotation invariant model is, therefore, critical to accurately classify and segment objects in 3D point clouds. + +Various attempts have been made to address the problem of rotation invariance in 3D point clouds, as evidenced by several works [3,4,22,27,41]. These efforts have focused on designing rotation invariant properties to achieve consistent accuracies under arbitrary rotations without resorting to rotation augmentation. However, these methods have demonstrated inferior performance when compared to translation invariant approaches, primarily due to the loss of global information during the generation of rotation invariant properties. Recent research efforts [15,30,40] have attempted to overcome this limitation by employing local reference frame (LRF) or local reference axis (LRA) [42] to transform the data into a canonical coordinate system and encode global information. Despite the improvements, their accuracies are still lower than those of state-of-the-art non-rotation invariant methods because of the unstable LRF/LRA and the less descriptive rotation invariant properties that the useful surface information is not well preserved. + +To address the above-mentioned issues, we propose a novel yet effective rotation invariant architecture. Specifically, we construct local triangle surfaces for each reference point of the input data to better capture the local surface structure. On the local surfaces, we design highly expressive rotation invariant surface properties which are then integrated into an attention-augmented convolution operator named RISurConv to extract refined rotation invariant features. Finally, we build up network based on RISurConv for rotation invariant object classification and segmentation. In summary, our main contributions include: + +- Rotation Invariant Surface Properties. We construct local triangle surfaces from the reference point and its neighbors, based on which we are able to design highly expressive rotation invariant surface properties; +- RISurConv. We integrate the Rotation Invariant Surface Properties into an attention-augmented architecture named RISurConv comprising two self-attention layers to learn and generate refined rotation invariant features; +- Extensive experiments on a variety of classification and segmentation tasks. Our approach shows supreme performance, surpassing the state-of-the-art by a large margin under challenging rotation scenarios including an analysis of rotation-invariant features and an ablation study of our neural + +network to provide insights into the key factors contributing to the exceptional performance. + +# 2 Related Works + +In this section, we review the representative works in 3D point cloud deep learning and rotation invariant learning. + +3D Point Cloud Deep Learning. 3D point clouds is a more compact and intuitive representation for feature learning. PointNet [23] pioneered a point cloud convolution with global features by max-pooling per-point features from MLPs, and follow-up works [17, 25, 38] have focused on exploring convolution kernels that exploit geometric features [29], adding edges on top of points [33], parameterizing convolution using polynomials [38], and leveraging shape context [37]. Some methods are designed to combine with recurrent neural networks [13] and sequence models [18]. + +In recent years, attention mechanism has become increasingly popular in various natural language processing (NLP) [32] and computer vision [6] tasks. With their ability to handle sequential and spatial information, attention mechanism has also shown promise in 3D point cloud processing tasks. Pioneer works that apply attention mechanism in point clouds include Point Transformer [43] and PCT [9], which use a self-attention mechanism to learn local and global features of point clouds. Other works have explored different types of transformers. Inspired by BERT [5], Point-BERT [39] was proposed to pre-train pure Transformer-based models with a Mask Point Modeling task for point cloud classification. Dual Transformer Network [11] aggregated the point-wise and channel-wise multi-head self-attention models simultaneously such that long-range context dependencies can be captured by investigating the point-wise and channel-wise relationships. Overall, these developments show promise in enhancing the efficiency and accuracy of point cloud analysis for a variety of applications. Readers can refer to the survey [10] for a comprehensive overview of 3D point cloud deep learning. However, most of these methods lack the property of rotation invariance which is critical for object classification and segmentation as the object can be viewed at any angle in reality. To address this issue, a common approach is to increase the training data by augmenting it with arbitrary rotations [17, 25, 33]. However, such approach has difficulty in generalizing predictions to unseen rotations, resulting in deteriorated performance. Therefore, it would be desirable to design a specific convolution possessing rotation-invariant features. + +Rotation Invariant Learning. Various methods have been proposed to address the issue of rotation invariance in feature learning of point clouds. For instance, Rao et al. [27] used a spherical domain to define a rotation-invariant convolution for point clouds. However, the discretized sphere is sensitive to global rotations, which can lead to a notable drop in performance for objects with arbitrary rotations. To overcome this limitation, Poulenard et al. [22] integrated spherical harmonics to their convolution. Similarly, Chen et al. [3] proposed a + +![](images/bbfdb259e870195d441844ef27250a76a3317bffbe790e18e1aa53036e6cd5e4.jpg) +Fig. 1: Rotation Invariant Surface Property (RISP) construction: Given a point $p$ as the reference point, $K$ ( $K = 8$ in this example) nearest points are selected (middle). For each neighbor $x_{i}$ , two adjacent neighbors $x_{i-1}$ and $x_{i+1}$ are used to form two triangular local surfaces (right), based on which rotation invariant properties are constructed. + +hierarchical clustering scheme to encode the relative angles between two-point vectors and used vector norm to maintain rotation invariance. Zhang et al. [41] presented a simple convolution named RICnv that operates on handcrafted features built from Euclidean distances and angles that are rotation invariant by nature. However, this approach only considers local features resulting in accuracy degradation. GCACnv [40] addressed this limitation by building a global context-aware convolution based on anchors and Local Reference Frame (LRF) to achieve rotation invariance. RI-GCN [15] learned rotation-invariant local descriptors and applied graph convolutional neural networks to aggregate local features based on LRF. Thomas [30] also relied on LRF and used a multiple alignment scheme to achieve better results. RIF [16] presented a framework to construct both local and global features based on distances, angles, and reference points. To remove the uncertainty of LRF in x and y directions, RICnv++ [42] extracts informative features based on local reference axis (LRA). However, the LRF/LRA and global reference used in these methods may not be stable enough limiting their overall performance. We can clearly see the accuracy discrepancy of the object classification task on ModelNet40 dataset [36]. State-of-the-art translation-invariant convolutions such as PointNet++ [25] or Point Transformer v2 [34] achieve 89%-94% of accuracy while rotation-invariant convolutions only report up to 87%-91% of accuracy [14,42]. + +Our target in this paper is not only closing this performance gap but also surpassing the state-of-the-art methods while maintaining rotation invariance. + +# 3 Rotation Invariant Surface Property + +In this section, we detail the Rotation Invariant Surface Property construction which is the first step of our method. Our goal for this step is to design highly expressive rotation invariant properties from underlying local surfaces. Different from previous works that rely on pointwise operations, we construct local surfaces around the reference point to better capture the local surface structure, based on which we then extract more expressive rotation invariant properties. + +The construction of rotation invariant surface property is shown in Fig. 1. Given a point $p$ as the reference point (red), we get $K$ nearest neighbors to form a local point set. In this case, $K$ is 8. For each neighbor $x_{i}$ , two adjacent neighbors, $x_{i-1}$ and $x_{i+1}$ , are identified based on the Euclidean distances, forming two triangular local surfaces, as shown in the figure. Then, we construct the rotation invariant surface properties (RISP) as follows: + +$$ +\operatorname {R I S P} \left(x _ {i}\right) = \left[ L _ {0}, \phi_ {1}, \phi_ {2}, \phi_ {3}, \phi_ {4}, \phi_ {5}, \right. \tag {1} +$$ + +$$ +\left. \alpha_ {1}, \alpha_ {2}, \beta_ {1}, \beta_ {2}, \theta_ {1}, \theta_ {2}, \gamma_ {1}, \gamma_ {2} \right] +$$ + +where $L_{0}$ measures the distance from reference $p$ to neighbor $x_{i}$ , and $\phi_{1}$ to $\phi_{5}$ measure the two triangles as well as the relationship between the two triangle surfaces with regard to the edge $\overrightarrow{px_i}$ in the Euclidean space: + +$$ +\phi_ {1} = \angle (\overrightarrow {x _ {i - 1} p}, \overrightarrow {x _ {i} p}), \phi_ {2} = \angle (\overrightarrow {x _ {i + 1} p}, \overrightarrow {x _ {i} p}), \tag {2} +$$ + +$$ +\phi_ {3} = \angle (\overrightarrow {x _ {i - 1} x _ {i}}, \overrightarrow {x _ {i - 1} p}), \phi_ {4} = \angle (\overrightarrow {x _ {i + 1} p}, \overrightarrow {x _ {i + 1} x _ {i}}), +$$ + +$$ +\phi_ {5} = \angle (\overrightarrow {x _ {i + 1} p} \times \overrightarrow {x _ {i} p}, \overrightarrow {x _ {i - 1} p} \times \overrightarrow {x _ {i} p}), +$$ + +while other properties describe the two surfaces in the tangent space, e.g., normal vectors can define the directions in which the surface is bending away from the tangent space: + +$$ +\alpha_ {1} = \angle (\overrightarrow {n _ {p}}, \overrightarrow {x _ {i} p}), \alpha_ {2} = \angle (\overrightarrow {n _ {p}}, \overrightarrow {x _ {i - 1} p}), \tag {3} +$$ + +$$ +\beta_ {1} = \angle (\vec {n _ {i}}, \vec {x _ {i}} \vec {p}), \beta_ {2} = \angle (\vec {n _ {i}}, \vec {x _ {i - 1}} \vec {x _ {i}}), +$$ + +$$ +\theta_ {1} = \angle (\overrightarrow {n _ {i - 1}}, \overrightarrow {x _ {i - 1} p}), \theta_ {2} = \angle (\overrightarrow {n _ {i - 1}}, \overrightarrow {x _ {i - 1} x _ {i} ^ {\prime}}), +$$ + +$$ +\gamma_ {1} = \angle (\overrightarrow {n _ {i + 1}}, \overrightarrow {x _ {i + 1} x _ {i}}), \gamma_ {2} = \angle (\overrightarrow {n _ {i + 1}}, \overrightarrow {x _ {i + 1} p}). +$$ + +RISP is able to fully describe the dual triangles as well as their relationship along different directions. Readers can refer to the supplementary for the mathematical proof of the completeness of RISP defined in Eq. (1). + +Number of local triangle surfaces. Throughout the paper, we define the number of local triangle surfaces as 2. Such setting is due to the fact that we wish to capture local surface as much as possible to extract more useful information to reflect the real shape of the object. So we construct 2 triangles rather than single triangle. However, this does not mean the more the better since more triangles can result in distorted surface and affect the computing efficiency. Please refer to the ablation study section for more details. + +# 4 The RISurConv Operator + +Based on the Rotation Invariant Surface Properties (RISP), we are able to build up RISurConv. The main steps are detailed in Fig. 2. To start, we utilize a farthest point sampling strategy to generate uniformly distributed representative + +![](images/d1b087ed69e990380fbe05236df59dd01c39ae9cb602fc2950de414963679502.jpg) +Fig. 2: RISurConv operator. For a local point set with $p$ as the reference (red), K nearest neighbors are labelled as blue. Then, we compute the Rotation Invariant Surface Properties at each neighbor by constructing local dual triangle surfaces (Sec. 3), which is embedded to a high-dimensional space by a shared multi-layer perceptron (MLP) followed by a self-attention layer to produce refined features. Concatenated with previous layer features (if any), the features of these local points are further passed to MLPs, which are then summarized by maxpooling. To further refine the features, another self-attention layer follows. + +# Algorithm 1 RISurConv operator. + +Input: Reference point $p$ , point set $\Omega$ , point features $\mathbf{f}_{prev}$ from previous layer (if any) +Output: Convoluted features $\mathbf{f}$ + +1: $\mathbf{f} \gets \{\mathrm{RISP}(x_i) : \forall x_i \in \Omega\}$ (Sec. 3) + +* Construct Rotation Invariant Surface Properties + +2: $\mathbf{f}\gets \mathrm{MLP}(\mathbf{f})$ + +* Embed each feature to a high-dimensional feature space + +3: $\mathbf{f}\gets \mathrm{SA}(\mathbf{f})$ + +* Refine features via self-attention layer + +4: $\mathbf{f}_{in}\gets [\mathbf{f}_{prev},\mathbf{f}]$ (if any) + +* Concatenate the features from the local and the previous layer + +5: $\mathbf{f}_{out} \gets \mathrm{MLP}(\mathbf{f}_{in})$ + +* Feature embedding + +6: $\mathbf{f}_{out}\gets$ maxpool $(\mathbf{f}_{out})$ return SA $(\mathbf{f}_{out})$ + +* Maxpool features + +* Self-attention and return + +points. From this initial set, we perform K-nearest neighbor searches to obtain local point sets. Let us denote a local point set by $\Omega = \{x_{i}\}$ , where $x_{i}$ represents the 3D coordinates of point $i$ . We define the attention-augmented convolution operation as follows: + +$$ +\mathbf {f} (\Omega) = S A \left(\sigma \left(\mathcal {A} \left(\left\{\mathcal {T} \left(\mathbf {f} _ {x _ {i}}\right): \forall i \right\}\right)\right)\right). \tag {4} +$$ + +This formula indicates that features of each point in the point set are first transformed by $\mathcal{T}$ before being aggregated by the aggregation function $\mathcal{A}$ and passed to an activation function $\sigma$ . SA is a channel-wise self-attention layer used to output refined features. We set the input features to our expressive rotation-invariant features $\mathbf{f}_{x_i} = \mathrm{RISP}(x_i)$ . We define the transformation function as + +$$ +\mathcal {T} \left(\mathbf {f} _ {x _ {i}}\right) = \mathbf {w} _ {i} \cdot \mathbf {f} _ {x _ {i}} = \mathbf {f} _ {x _ {i}} ^ {\prime} \tag {5} +$$ + +where $\cdot$ indicates the element-wise product, and $\mathbf{w}_i$ is the weight parameter to be learned by the network. Our transformation function is similar to PointNet++. A popular choice of the aggregation function $\mathcal{A}$ is maxpooling, which supports permutation invariance of the input point features. + +To proceed with feature learning and be invariant to point permutation, we construct surfaces for each neighbor from which we construct RISPs. In Fig. 2, there are eight neighbors, so the resulted RISPs are in size of $8 \times 14$ . Here, 14 is the length of the RISP as defined in Eq. (1) and 8 is the neighborhood size. Since RISPs are used as the input which are already rotation invariant, RISurConv is rotation invariant by nature. We then embed the RISPs into feature space using two layers of MLPs followed by a self-attention (SA) layer for feature refinement. We concatenate the refined features with previous layer features (if any), and embed the concatenated features into higher dimensional space via MLPs again. After the maxpooling, the aggregated features are passed into another SA layer again for further refinement. The detailed steps are shown in the Algorithm 1. Here, the two SA layers used in RISurConv are the standard SA module of the transformer [32]. + +Self Attention Layers. In RISurConv there are two Self-Attention (SA) layers by which the extracted features are enhanced. Below, we provide a more detailed illustration of the structure of the SA layers. The SA structure is the standard SA module of the famous transformer [32] (Attention is all you need) which runs twice in RISurConv. The first SA is for the K neighbors, and the second SA works for the N representative points. The input tensor shape is $[B,N,K,C_0]$ for the first SA which goes into linear layers and outputs $Q,K$ and $V$ respectively in shape of $[B,N,K,C_0]$ . According to $\text{att}(Q,K,V) = \text{Softmax}(QK^{\top} / \text{sqrt}(d_k))V$ , we have the 1st attention score in $[B,N,K,K]$ . By multiplying with $V$ , we get the refined feature in $[B,N,K,C_0]$ . The same goes with the second SA where the input shape is $[B,N,C]$ and the attention score shape is $[B,N,N]$ . + +# 5 RISurConv Networks + +We employed the RISurConv to develop neural networks for object classification and segmentation, as illustrated in Fig. 3. Our classification network follows a standard architecture similar to the single-scale grouping version of PointNet++, consisting of five consecutive layers of RISurConv followed by a transformer encoder [32] with 8 heads to enhance the extracted features. Finally, we connect the network to fully connected layers to produce the probability map. One of the major advantages of RISurConv is its ability to handle arbitrary rotation and point orders, which enables us to place each RISurConv layer consecutively without the need for complex preprocessing steps. Additionally, we apply batch normalization and ReLU activation to each convolution layer by default. The segmentation network utilizes an encoder-decoder structure similar to the U-Net [28]. We adopt the same definition of deconvolution as RISurConv. The key distinction lies in that the convolution produces a point subset with a greater + +![](images/87105d49eaf443ae2b40917f81563b05648fb1b683c59903ffa1a82c5b00d65c.jpg) +Fig. 3: Our neural network architecture comprises five RISurConv layers to extract rotation invariant features followed by a Transformer Encoder to enhance the learnt features before fully connected layers for object classification. We add a decoder with skip connections for segmentation task. + +number of feature channels and less points while the deconvolution outputs to a point set with more points compared to the input with fewer feature channels. Please refer to the supplementary for more details regarding the network architecture. + +# 6 Experiments + +We report our evaluation results in this section. Our network is implemented in PyTorch using an Adam optimizer with initial learning rate of 0.001 for both classification and segmentation. Our training is executed on a computer with an Intel(R) Core(TM) i7-10700K CPU equipped with a NVIDIA GTX 3090 GPU. + +We evaluated the performance of our method extensively on a series of classification and segmentation benchmarks. To train our model, we utilized 450 epochs for classification and 250 epochs for segmentation for most datasets except the Car and Chair categories of the fine-grained FG3D dataset due to small size and subtle variation where we applied 900 epochs. Our method can achieve convergence within 11 hours for classification and 24 hours for segmentation. To assess the robustness of our method, we followed the experimental design proposed by [8] and conducted experiments in three cases. The first case involved training and testing with data augmented with rotation about the gravity axis $(\mathrm{z} / \mathrm{z})$ , which is a common approach for evaluating translation-invariant point cloud learning methods. The second case involved training and testing with data augmented with arbitrary SO3 rotations (SO3/SO3), and the third case involved training with data by z-rotations and testing with data by SO3 rotations (z/SO3), which are used to evaluate rotation invariance. A rotation-invariant method should produce consistent results across all three cases. Our method demonstrated superior performance across all three cases, highlighting its robustness and effectiveness. + +# 6.1 Human-made Object Classification + +ModelNet40 [36] that comprises 9843 training models and 2468 test models, divided into 40 categories is one of the most popular datasets for point cloud classification evaluation. The input data size is 1024 with each point equipped with (x; y; z; nx; ny; nz) which are 3D coordinates and 3D normals in the Euclidean space. Note that normals are optional to our method. When normals are not available (w/o n), we use weighted eigen vector corresponding to the smallest eigen value as the normal. This strategy has been used in SHOT and RICConv++ [42] which is highly robust to noise and efficient to compute. + +The experimental results are shown in Tab. 1. We use two criteria for evaluation: overall accuracy and accuracy standard deviation (Std.). True rotation invariant methods are expected to be unaffected by the rotations present in the training and testing data, resulting in low accuracy deviation. + +From Tab. 1, we see that our method surpasses the state-of-the-art performance in all cases while maintaining zero accuracy deviation. It is worth noting that our method surpasses both rotation invariant and non-rotation invariant methods. To the best of our knowledge, this is the first work that achieves such high accuracy. It outperforms the state-of-the-art rotation invariant method RIConv++ by $4.7\%$ , and outperforms the state-of-the-art non-rotation invariant method Point Transformer v2 by $1.8\%$ under $\mathrm{z / z}$ case. + +# 6.2 Real World Object Classification + +We also evaluate the classification performance on a real-world 3D point cloud dataset ScanObjectNN [31]. It is composed of 2902 point clouds categorized into 15 categories sampled from real world indoor scenes. For our evaluation, we use the processed files and choose the hardest variant PB_T50_RS with $50\%$ bounding box translation, rotation around the gravity axis, and random scaling that result in rotated and partial data. The results are shown + +Table 2: Comparisons of real world 3D point cloud classification on hardest variant of ScanObjectNN dataset (Overall Accuracy). + +
MethodPB_T50_RS
z/zSO3/SO3z/SO3
PointNet [23]68.242.217.1
PointNet++ [25]77.960.115.8
PointCNN [17]78.551.814.9
DGCNN [33]78.163.416.1
RICov [41]68.168.368.3
GCACov [40]69.870.069.8
RICov++ [42]80.380.380.3
Ours (w/o normal)93.193.193.1
+ +in Tab. 2. We see that RISurConv surpasses all the compared methods by a large margin. Particularly, our method significantly outperforms the state-of-the-art rotation invariant approach RICov++ [42] by $12.8\%$ . This verifies that our method is effective for both synthetic and real world data. Note that we only test the 'w/o normal' case as the normal vectors are not provided in the processed files of this dataset. + +Table 1: Comparisons of the classification accuracy (\%) on the ModelNet40 dataset. On average, our method has the best accuracy and lowest accuracy deviation in all cases (Overall Accuracy). + +
MethodFormatInput Size Params.z/z↑SO3/ SO3↑z/ SO3↑Std.↓
TraditionalVoxNet [21]voxel3030.90M83.087.3-3.0
SubVolSup [24]voxel30317.00M88.582.736.628.4
PointNet [23]xyz1024 × 33.50M87.080.321.641.0
PointCNN [17]xyz1024 × 30.60M91.384.541.227.2
PointNet++ [25]xyz + nor1024 × 61.40M89.385.028.633.8
DGCNN [33]xyz1024 × 31.84M92.281.120.638.5
RS-CNN [20]xyz1024 × 31.41M90.382.648.722.1
Pt Transformer [43]xyz1024 × 3-93.785.950.119.1
Pt Transformer v2 [34]xyz1024 × 3-94.288.351.823.0
Rotation-invariantSpherical CNN [7]voxel2 × 6420.50M88.986.978.65.5
RICConv [41]xyz1024 × 30.70M86.586.486.40.1
SPHNet [22]xyz1024 × 32.90M87.087.686.60.5
SFCNN [27]xyz1024 × 3-91.490.184.83.5
ClusterNet [3]xyz1024 × 31.40M87.187.187.10.0
GCACov [40]xyz1024 × 30.41M89.089.289.10.0
RIF [16]xyz1024 × 3-89.489.389.40.0
RI-GCN [15]xyz + nor1024 × 64.38M91.091.091.00.0
RICConv++ [42]xyz1024 × 30.40M91.291.291.20.0
RICConv++ [42]xyz + nor1024 × 60.40M91.391.391.30.0
Ours (w/o normal)xyz1024 × 314.0M95.695.695.60.0
Ours (w/ normal)xyz + nor1024 × 614.0M96.096.096.00.0
+ +# 6.3 Fine-Grained Object Classification + +Fine-grained object classification is more challenging because the differences between subcategories are subtle and require a high level of precision in distinguishing them. We conduct experiments on the three categories from FD3D dataset [19]: Airplane, Chair and Car. To the best of our knowledge, this is the first work to test the performance of rotation invariant methods on fine-grained 3D point cloud dataset. The class accuracies are shown in Tab. 3. Again, our method outperforms the state-of-the-art approaches by large margins for all rotation cases and all the categories. Specifically, our method improves non-rotation invariant method FG3D-Net by $2.1\%$ , $2.7\%$ , and $4.5\%$ for Airplane, Chair, and Car categories respectively under the $\mathrm{z} / \mathrm{z}$ rotation case, and the accuracy consistency is well preserved. Compared to the rotation invariant methods, our method outperforms the latest RICov++ by $3.6\%$ , $5.1\%$ , and $9.2\%$ on the three categories. Such supreme performance shows that features extracted by RISurConv are not only rotation invariant but also highly expressive. + +# 6.4 Part Segmentation on ShapeNet + +Table 3: Comparisons of fine-grained 3D point cloud classification on FG3D dataset (Class Accuracy). + +
MethodModalityAirplaneChairCar
z/zSO3/SO3z/SO3z/zSO3/SO3z/SO3z/zSO3/SO3z/SO3
VoxNetVoxel79.257.528.773.354.716.768.242.217.1
MVCNNView82.666.921.976.363.714.671.951.814.9
View-GCNView87.456.622.779.750.219.773.744.418.1
RotationNetView89.150.829.878.536.218.472.545.328.3
Part4FeatureView82.649.132.277.152.230.273.435.023.8
FG3D-NetView89.461.126.980.043.218.474.056.420.8
PointNetPoint82.755.529.772.133.918.268.132.116.4
PointNet++Point87.357.125.678.140.416.070.350.420.8
RS-CNNPoint82.845.932.275.150.718.671.231.813.9
DGCNNPoint88.460.622.771.752.319.765.353.418.1
RICConvPoint85.885.885.876.376.376.367.367.367.3
RICConv++Point87.987.987.977.677.677.669.369.369.3
OursPoint91.591.591.582.782.782.778.578.578.5
+ +To evaluate the segmentation performance, we employed the ShapeNet dataset comprising 16880 CAD models distributed across 16 distinct categories. These models are annotated with 2 to 6 parts each, culminating in a dataset of 50 object parts. We follow the standard train/test split with 14006 models for training and 2874 models for testing, respectively. + +The evaluation results are shown in Tab. 4. Our method outperforms both traditional translation-invariant and latest rotation invariant methods significantly in the SO3/SO3 and z/SO3 scenario. Our method outperforms the RICov++ by + +Table 4: Comparisons of object part segmentation performed on ShapeNet dataset. The mean per-class IoU (mIoU, %) is used to measure the accuracy under two challenging rotation modes: SO3/SO3 and z/SO3. + +
MethodInputSO3/SO3 z/SO3
PointNet [23]xyz74.437.8
PointNet++ [25]xyz+nor76.748.2
PointCNN [17]xyz71.434.7
DGCNN [33]xyz73.337.4
RS-CNN [20]xyz72.536.5
SpiderCNNxyz+nor72.342.9
RICConv [41]xyz75.575.3
GCACConv [40]xyz77.377.2
RI-GCN [15]xyz77.077.0
RIF [16]xyz79.479.2
RICConv++ [42]xyz80.380.3
RICConv++ [42]xyz+nor80.580.5
Ours (w/o normal)xyz81.381.3
Ours (w/ normal)xyz+nor81.581.5
+ +1.0% mIoU. This result aligns well with the performance reported in the object classification task. We also show the qualitative results by error maps in Fig. 4. The wrong segmentation points are plotted as red. It clearly shows that our predictions are the closest to the ground truth. + +![](images/04988c00896f6334af72dad5432b97b84261d7f7f701a72857e6be6179788f6d.jpg) +Fig. 4: Qualitative comparisons (Red indicates wrong). + +# 6.5 Ablation Studies + +We ablate some vital designs of our method on ModelNet40 [36] for an insightful exploration. + +RISP Design. We first test the performance by turning on/off and adding different rotation invariant components used in the RISP construction (Algorithm 1). The results are shown in Tab. 5. Model A is our baseline setting with all rotation invariant features activated. Model B has only angle features but still achieves relatively high accuracy. Model C turns off both $L_{0}$ and $\phi_{1\dots 5}$ with accuracy decreases to $90.9\%$ . This means that the distance $L_{0}$ is not as important as angles. In Model D, we turn off the angle features on the tangent direction with only $L_{0}$ and $\phi_{1\dots 5}$ are kept which is more like RICnv [41]. Compared to Model A, it shows that our proposed features is more effective than those by RICnv, and the improvement is explained by the additional consideration of the relations among the neighbor points. Model E adds more rotation invariant + +Table 5: Ablation study on the RISP design. + +
ModelL0φ1...5α1,2β1,2θ1,2γ1,2λμAcc.
A✓✓✓96.0
B✓✓95.5
C✓90.9
D✓✓88.2
E✓✓✓✓95.7
+ +Table 6: Ablation study on the Self-Attention module. + +
ModelSA1SA2TransformerEncoderAcc.
A✓✓✓96.0
B✓✓95.6
C✓✓95.2
D✓✓94.3
E92.8
+ +features. $\lambda$ and $\mu$ represent the angles $\angle (\pmb{n}_{\pmb{p}},\pmb{n}_{i + 1})$ and $\angle (\pmb {n}_p,\pmb {n}_{i - 1})$ respectively. The results show that the accuracy does not increase. This verifies the completeness of the proposed RISP. + +Self-Attention Effects. We employ multiple self-attention (SA) modules to help produce refined rotation invariant features. So it is necessary to analyse the effects of different SA modules. In RISurConv, there are two SA modules and we name them as SA1 and SA2 respectively. Before fully connected layers, there is a Transformer Encoder module. We test these three modules by turning off one of them each time. From Tab. 6, we see that when removing the first SA module in RISurConv, the performance drops a bit, while removing the second SA module also decreases the accuracy. This indicates that both SA1 and SA2 are important for the success of our method. In addition, we also test the performance when Transformer Encoder is removed. The overall accuracy decreases to $94.3\%$ but it is still higher than most existing methods. In model E, we test the performance when all SA modules are removed. The accuracy decreases to $92.8\%$ which is still better than the state-of-the-art rotation invariant method [42] thanks to the well designed RISPs which capture sufficient surface structures. + +We also visualize the global feature vector before classification as shown in Fig. 5. We normalize the feature values to [0,1] and plot the histograms. We can see that by turning on the SA modules, the feature values become more evenly distributed and have less zeros because SA mechanism can reweight the features making the features more effective. + +![](images/563e38b74af65103e4fbea3c41cfa12cb6705795b3829a6e5ec8dc24cfd593b3.jpg) +Fig. 5: Histogram comparison for normalized feature values without and with self-attention layers. + +Network Efficiency. In the experiments, we acknowledge that our method employs a higher number of parameters. Thus, it is imperative to conduct a comprehensive analysis of network efficiency during both the training and testing stages. We list the trainable parameters, FLOPs + +Table 7: Trainable parameters, FLOPs and timing comparisons on ModelNet40. All are tested on the same platform. Time is measured per batch (batch size=16). + +
MethodsParamsFLOPsTime (Train/Infer)
PointNet++ [25]1.41M0.86G0.145s / 0.129s
RICnv [41]0.70M0.92G0.121s / 0.092s
RI-GCN [15]4.19M1.24G0.148s / 0.112s
RICnv++ [42]0.42M0.72G0.074s / 0.029s
Ours13.96M1.12G0.114s / 0.049s
Ours w/o TR1.35M1.11G0.105s / 0.047s
+ +and running time comparisons in Tab. 7. We compare the performance with four recent works: PointNet++ [25], RICnv [41], RI-GCN [15], RICnv++ [42]. + +From the results, we can see that the number of parameters mainly comes from the transformer architectures used in RISurConv. When we turn off all + +the transformer component, our method only takes up 1.35M parameters. More parameters do not mean longer inference time. It depends on lots of things including depth, parameter, number of operations etc. Our approach runs faster than most existing methods. It takes more time for training because more epochs are needed for self attention layers to converge. This is a common phenomenon of transformer based methods. The accuracy is also not due to the complexity model. As is shown in the ablation studies, by removing transformer encoder, it still achieves state-of-the-art performance due to the effectiveness of RISurConv. + +Number of Local Surfaces. Local triangle surfaces used in Sec. 3 are the base for rotation invariant surface property (RISP) construction. Thus, it is important to analyse the effects of different number of surfaces. We conduct performance tests by setting the number of local surfaces to 1, 2, 3, and 4, resulting in overall accuracies of 94.7, 96.0, 91.9, and 89.8 respectively. It worth noting that for more surfaces, we can define more RISPs. From the results, we see the accuracy drops a bit when single + +![](images/2e15e004c3454175b9ee2af6ae6f949fdf51f2d4757028bdacf6ddf6022fbc67.jpg) +Fig. 6: Distorted surface caused by 3 triangles. + +surface is constructed. This is because smaller area cannot capture sufficient surface structures and less effective features can lower the performance. When we increase the number of surfaces to 3 and 4, the performance decreases. This shows that more surfaces can result in inferior performance because we construct triangle surfaces just to approximate the true surface. More surfaces may mess up the surface structure and affect the performance. Take Fig. 6 for example, suppose we are constructing the local surfaces $(\triangle px_{i}x_{i + 1}\triangle px_{i}x_{i - 1})$ with regard to the neighbor point $x_{i}$ , adding one more triangle $\triangle px_{i}x_{i + 2}$ will result in distorted/non-manifold surfaces. Thus, we set number of surfaces as 2 throughout the paper. + +# 7 Discussion + +**Limitation.** Though effective, RISurConv may require longer training time (e.g. more epochs) to converge due to large number of training parameters (Tab. 1). This is a common issue of transformer. But this does not affect the inference speed as shown in the ablation study (Tab. 7). + +Conclusion. We have presented a new framework for achieving rotation invariance on 3D point cloud. We construct local triangle surfaces to better represent the local surface structures based on which we design highly expressive rotation invariant surface properties. We integrate the properties into an architecture named RISurConv to extract refined rotation invariant features. Based on RISurConv, we finally build up effective rotation invariant neural networks for 3D point cloud classification and segmentation with supreme performance achieved not only closing the performance gap with non-rotation invariant approaches but also surpassing the state-of-the-art methods by a large margin. + +# Acknowledgements + +This research is supported by the Singapore Ministry of Education (MOE) Academic Research Fund (AcRF) Tier 1 grant (MSS23C010), Ningbo 2025 Science & Technology Innovation Major Project (No. 2022Z072, No.2023Z044), and Key Research & Development Plan of Zhejiang Province (2024C01017). + +# References + +1. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3d semantic parsing of large-scale indoor spaces. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1534-1543 (2016) +2. Chang, A.X., Funkhouser, T.A., Guibas, L.J., Hanrahan, P., Huang, Q.X., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., Xiao, J., Yi, L., Yu, F.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015) +3. Chen, C., Li, G., Xu, R., Chen, T., Wang, M., Lin, L.: Clusternet: Deep hierarchical cluster network with rigorously rotation-invariant representation for point cloud analysis. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4994-5002 (2019) +4. Deng, H., Birdal, T., Ilic, S.: Ppf-foldnet: Unsupervised learning of rotation invariant 3d local descriptors. In: Eur. Conf. Comput. Vis. pp. 602-618 (2018) +5. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: NAACL-HLT (2018) +6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Int. Conf. Learn. Represent. (2020) +7. Esteves, C., Allen-Blanchette, C., Makadia, A., Daniilidis, K.: Learning so (3) equivariant representations with spherical cnns. In: Eur. Conf. Comput. Vis. pp. 52-68 (2018) +8. Esteves, C., Xu, Y., Allen-Blanchette, C., Daniilidis, K.: Equivariant multi-view networks. In: Int. Conf. Comput. Vis. pp. 1568-1577 (2019) +9. Guo, M.H., Cai, J.X., Liu, Z.N., Mu, T.J., Martin, R.R., Hu, S.M.: Pct: Point cloud transformer. Computational Visual Media 7, 187-199 (2021) +10. Guo, Y., Wang, H., Hu, Q., Liu, H., Liu, L., Bennamoun, M.: Deep learning for 3d point clouds: A survey. IEEE Trans. Pattern Anal. Mach. Intell. 43(12), 4338-4364 (2020) +1. Han, X.F., Jin, Y.F., Cheng, H.X., Xiao, G.Q.: Dual transformer for point cloud analysis. IEEE Trans. Multimedia (2022) +2. Hua, B.S., Pham, Q.H., Nguyen, D.T., Tran, M.K., Yu, L.F., Yeung, S.K.: Scenenn: A scene meshes dataset with annotations. In: International Conference on 3D Vision. pp. 92-101 (2016) +3. Huang, Q., Wang, W., Neumann, U.: Recurrent slice networks for 3d segmentation of point clouds. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2626-2635 (2018) +4. Kim, J., Jung, W., Kim, H., Lee, J.: Cynnn: a rotation invariant cnn using polar mapping and cylindrical convolution layers. arXiv preprint arXiv:2007.10588 (2020) +5. Kim, S., Park, J., Han, B.: Rotation-invariant local-to-global representation learning for 3d point cloud. Adv. Neural Inform. Process. Syst. 33, 8174-8185 (2020) + +16. Li, X., Li, R., Chen, G., Fu, C.W., Cohen-Or, D., Heng, P.A.: A rotation-invariant framework for deep point cloud analysis. IEEE Trans. Vis. Comput. Graph. (2021) +17. Li, Y., Bu, R., Sun, M., Wu, W., Di, X., Chen, B.: Pointcnn: Convolution on x-transformed points. In: Adv. Neural Inform. Process. Syst. pp. 820-830 (2018) +18. Liu, X., Han, Z., Liu, Y.S., Zwicker, M.: Point2sequence: Learning the shape representation of 3d point clouds with an attention-based sequence to sequence network. In: AAAI. vol. 33, pp. 8778-8785 (2019) +19. Liu, X., Han, Z., Liu, Y.S., Zwicker, M.: Fine-grained 3d shape classification with hierarchical part-view attention. IEEE Trans. Image Process. 30, 1744-1758 (2021) +20. Liu, Y., Fan, B., Xiang, S., Pan, C.: Relation-shape convolutional neural network for point cloud analysis. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 8895-8904 (2019) +21. Maturana, D., Scherer, S.: Voxnet: A 3d convolutional neural network for real-time object recognition. In: 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 922-928 (2015) +22. Poulenard, A., Rakotosaona, M.J., Ponty, Y., Ovsjanikov, M.: Effective rotation-invariant point cnn with spherical harmonics kernels. International Conference on 3D Vision (2019) +23. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 652-660 (2017) +24. Qi, C.R., Su, H., Nießner, M., Dai, A., Yan, M., Guibas, L.J.: Volumetric and multi-view cnns for object classification on 3d data. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5648-5656 (2016) +25. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: Adv. Neural Inform. Process. Syst. pp. 5105-5114 (2017) +26. Qian, G., Li, Y., Peng, H., Mai, J., Hammoud, H., Elhoseiny, M., Ghanem, B.: Pointnext: Revisiting pointnet++ with improved training and scaling strategies. Adv. Neural Inform. Process. Syst. 35, 23192-23204 (2022) +27. Rao, Y., Lu, J., Zhou, J.: Spherical fractal convolutional neural networks for point cloud recognition. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019) +28. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: International Conference on Medical image computing and computer-assisted intervention. pp. 234-241 (2015) +29. Shen, Y., Feng, C., Yang, Y., Tian, D.: Mining point cloud local structures by kernel correlation and graph pooling. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4548-4557 (2018) +30. Thomas, H.: Rotation-invariant point convolution with multiple equivariant alignments. In: International Conference on 3D Vision (3DV). pp. 504-513 (2020) +31. Uy, M.A., Pham, Q.H., Hua, B.S., Nguyen, D.T., Yeung, S.K.: Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In: Int. Conf. Comput. Vis. (2019) +32. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Adv. Neural Inform. Process. Syst. 30 (2017) +33. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph cnn for learning on point clouds. ACM Trans. Graph. (2019) +34. Wu, X., Lao, Y., Jiang, L., Liu, X., Zhao, H.: Point transformer v2: Grouped vector attention and partition-based pooling. Adv. Neural Inform. Process. Syst. 35, 33330-33342 (2022) + +35. Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., Xiao, J.: 3d shapenets: A deep representation for volumetric shapes. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1912-1920 (2015) +36. Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., Xiao, J.: 3d shapenets: A deep representation for volumetric shapes. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1912-1920 (2015) +37. Xie, S., Liu, S., Chen, Z., Tu, Z.: Attentional shapecontextnet for point cloud recognition. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4606-4615 (2018) +38. Xu, Y., Fan, T., Xu, M., Zeng, L., Qiao, Y.: SpiderCNN: Deep learning on point sets with parameterized convolutional filters. In: Eur. Conf. Comput. Vis. pp. 87-102 (2018) +39. Yu, X., Tang, L., Rao, Y., Huang, T., Zhou, J., Lu, J.: Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 19313-19322 (2022) +40. Zhang, Z., Hua, B.S., Chen, W., Tian, Y., Yeung, S.K.: Global context aware convolutions for 3d point cloud understanding. In: International Conference on 3D Vision (2020) +41. Zhang, Z., Hua, B.S., Rosen, D.W., Yeung, S.K.: Rotation invariant convolutions for 3d point clouds deep learning. In: International Conference on 3D Vision. pp. 204-213 (2019) +42. Zhang, Z., Hua, B.S., Yeung, S.K.: Riconv++: Effective rotation invariant convolutions for 3d point clouds deep learning. Int. J. Comput. Vis. 130(5), 1228-1243 (2022) +43. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: Int. Conf. Comput. Vis. pp. 16259-16268 (2021) +44. Zhao, Y., Birdal, T., Lenssen, J.E., Menegatti, E., Guibas, L., Tombari, F.: Quaternion equivariant capsule networks for 3d point clouds. In: Eur. Conf. Comput. Vis. pp. 1-19. Springer (2020) \ No newline at end of file diff --git a/2024/RISurConv_ Rotation Invariant Surface Attention-Augmented Convolutions for 3D Point Cloud Classification and Segmentation/images.zip b/2024/RISurConv_ Rotation Invariant Surface Attention-Augmented Convolutions for 3D Point Cloud Classification and Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..517218e62ed1197d00ced0d219ca30a3d98ff82f --- /dev/null +++ b/2024/RISurConv_ Rotation Invariant Surface Attention-Augmented Convolutions for 3D Point Cloud Classification and Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:faeccaee158f16f2f87efdccfad6a98cc7803df794022daa81faf60acb2e8fa9 +size 615046 diff --git a/2024/RISurConv_ Rotation Invariant Surface Attention-Augmented Convolutions for 3D Point Cloud Classification and Segmentation/layout.json b/2024/RISurConv_ Rotation Invariant Surface Attention-Augmented Convolutions for 3D Point Cloud Classification and Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5bcb0d154362d51995f040a7f9828f1337d9906d --- /dev/null +++ b/2024/RISurConv_ Rotation Invariant Surface Attention-Augmented Convolutions for 3D Point Cloud Classification and Segmentation/layout.json @@ -0,0 +1,8976 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 137, + 111, + 478, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 111, + 478, + 166 + ], + "spans": [ + { + "bbox": [ + 137, + 111, + 478, + 166 + ], + "type": "text", + "content": "RISurConv: Rotation Invariant Surface Attention-Augmented Convolutions for 3D Point Cloud Classification and Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 183, + 186, + 430, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 186, + 430, + 200 + ], + "spans": [ + { + "bbox": [ + 183, + 186, + 430, + 200 + ], + "type": "text", + "content": "Zhiyuan Zhang" + }, + { + "bbox": [ + 183, + 186, + 430, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 183, + 186, + 430, + 200 + ], + "type": "text", + "content": ", Licheng Yang" + }, + { + "bbox": [ + 183, + 186, + 430, + 200 + ], + "type": "inline_equation", + "content": "^{2\\star}" + }, + { + "bbox": [ + 183, + 186, + 430, + 200 + ], + "type": "text", + "content": ", and Zhiyu Xiang" + }, + { + "bbox": [ + 183, + 186, + 430, + 200 + ], + "type": "inline_equation", + "content": "^{2\\oplus}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 207, + 478, + 241 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 133, + 207, + 478, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 207, + 478, + 230 + ], + "spans": [ + { + "bbox": [ + 133, + 207, + 478, + 230 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 207, + 478, + 230 + ], + "type": "text", + "content": " School of Computing and Information Systems, Singapore Management University https://github.com/cszyzhang/RISurConv" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 230, + 471, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 230, + 471, + 241 + ], + "spans": [ + { + "bbox": [ + 140, + 230, + 471, + 241 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 230, + 471, + 241 + ], + "type": "text", + "content": " College of Information Science and Electronic Engineering, Zhejiang University" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 159, + 268, + 455, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 268, + 455, + 499 + ], + "spans": [ + { + "bbox": [ + 159, + 268, + 455, + 499 + ], + "type": "text", + "content": "Abstract. Despite the progress on 3D point cloud deep learning, most prior works focus on learning features that are invariant to translation and point permutation, and very limited efforts have been devoted for rotation invariant property. Several recent studies achieve rotation invariance at the cost of lower accuracies. In this work, we close this gap by proposing a novel yet effective rotation invariant architecture for 3D point cloud classification and segmentation. Instead of traditional pointwise operations, we construct local triangle surfaces to capture more detailed surface structure, based on which we can extract highly expressive rotation invariant surface properties which are then integrated into an attention-augmented convolution operator named RISurConv to generate refined attention features via self-attention layers. Based on RISurConv we build an effective neural network for 3D point cloud analysis that is invariant to arbitrary rotations while maintaining high accuracy. We verify the performance on various benchmarks with supreme results obtained surpassing the previous state-of-the-art by a large margin. We achieve an overall accuracy of " + }, + { + "bbox": [ + 159, + 268, + 455, + 499 + ], + "type": "inline_equation", + "content": "96.0\\%" + }, + { + "bbox": [ + 159, + 268, + 455, + 499 + ], + "type": "inline_equation", + "content": "(+4.7\\%)" + }, + { + "bbox": [ + 159, + 268, + 455, + 499 + ], + "type": "text", + "content": " on ModelNet40, " + }, + { + "bbox": [ + 159, + 268, + 455, + 499 + ], + "type": "inline_equation", + "content": "93.1\\%" + }, + { + "bbox": [ + 159, + 268, + 455, + 499 + ], + "type": "inline_equation", + "content": "(+12.8\\%)" + }, + { + "bbox": [ + 159, + 268, + 455, + 499 + ], + "type": "text", + "content": " on ScanObjectNN, and class accuracies of " + }, + { + "bbox": [ + 159, + 268, + 455, + 499 + ], + "type": "inline_equation", + "content": "91.5\\%" + }, + { + "bbox": [ + 159, + 268, + 455, + 499 + ], + "type": "inline_equation", + "content": "(+3.6\\%)" + }, + { + "bbox": [ + 159, + 268, + 455, + 499 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 159, + 268, + 455, + 499 + ], + "type": "inline_equation", + "content": "82.7\\%" + }, + { + "bbox": [ + 159, + 268, + 455, + 499 + ], + "type": "inline_equation", + "content": "(+5.1\\%)" + }, + { + "bbox": [ + 159, + 268, + 455, + 499 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 159, + 268, + 455, + 499 + ], + "type": "inline_equation", + "content": "78.5\\%" + }, + { + "bbox": [ + 159, + 268, + 455, + 499 + ], + "type": "inline_equation", + "content": "(+9.2\\%)" + }, + { + "bbox": [ + 159, + 268, + 455, + 499 + ], + "type": "text", + "content": " on the three categories of the FG3D dataset for the fine-grained classification task. Additionally, we achieve " + }, + { + "bbox": [ + 159, + 268, + 455, + 499 + ], + "type": "inline_equation", + "content": "81.5\\%" + }, + { + "bbox": [ + 159, + 268, + 455, + 499 + ], + "type": "inline_equation", + "content": "(+1.0\\%)" + }, + { + "bbox": [ + 159, + 268, + 455, + 499 + ], + "type": "text", + "content": " mIoU on ShapeNet for the segmentation task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 509, + 394, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 509, + 394, + 521 + ], + "spans": [ + { + "bbox": [ + 160, + 509, + 394, + 521 + ], + "type": "text", + "content": "Keywords: Point cloud " + }, + { + "bbox": [ + 160, + 509, + 394, + 521 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 509, + 394, + 521 + ], + "type": "text", + "content": " Rotation invariant " + }, + { + "bbox": [ + 160, + 509, + 394, + 521 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 509, + 394, + 521 + ], + "type": "text", + "content": " Attention" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 540, + 230, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 540, + 230, + 552 + ], + "spans": [ + { + "bbox": [ + 132, + 540, + 230, + 552 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "text", + "content": "Point cloud has become the most promising 3D data representation for a wide range of immersive applications from robot navigation to autonomous driving. The increasing availability of 3D sensors and the emergence of large and mid-scale point cloud datasets [1,2,12,19,31,35] have spurred significant research in this area, leading to the development of numerous deep learning models for point cloud classification and segmentation, which are the fundamentals for various computer vision tasks." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 654, + 204, + 664 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 654, + 204, + 664 + ], + "spans": [ + { + "bbox": [ + 133, + 654, + 204, + 664 + ], + "type": "text", + "content": "\\* co-first author." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "type": "text", + "content": "However, analyzing 3D point clouds remains challenging, mainly due to the irregular nature of point clouds and their inherent invariances such as translation, point permutation, and rotation. While significant progress has been made in learning translation and permutation-invariant features [17, 23, 25, 26, 44], achieving rotation invariance in point cloud convolution has been a relatively unexplored area." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 272 + ], + "type": "text", + "content": "Rotation invariance is essential for 3D object classification and segmentation as objects can be viewed from different viewpoints, leading to variations in their orientation. Prior approaches [23,25,33] usually rely on rotation augmentation in training stage to relief the rotation affections for testing. However, such scheme is less effective as 3D data has more degrees of freedom. A rotation invariant model is, therefore, critical to accurately classify and segment objects in 3D point clouds." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 272, + 482, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 439 + ], + "type": "text", + "content": "Various attempts have been made to address the problem of rotation invariance in 3D point clouds, as evidenced by several works [3,4,22,27,41]. These efforts have focused on designing rotation invariant properties to achieve consistent accuracies under arbitrary rotations without resorting to rotation augmentation. However, these methods have demonstrated inferior performance when compared to translation invariant approaches, primarily due to the loss of global information during the generation of rotation invariant properties. Recent research efforts [15,30,40] have attempted to overcome this limitation by employing local reference frame (LRF) or local reference axis (LRA) [42] to transform the data into a canonical coordinate system and encode global information. Despite the improvements, their accuracies are still lower than those of state-of-the-art non-rotation invariant methods because of the unstable LRF/LRA and the less descriptive rotation invariant properties that the useful surface information is not well preserved." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 440, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 482, + 536 + ], + "type": "text", + "content": "To address the above-mentioned issues, we propose a novel yet effective rotation invariant architecture. Specifically, we construct local triangle surfaces for each reference point of the input data to better capture the local surface structure. On the local surfaces, we design highly expressive rotation invariant surface properties which are then integrated into an attention-augmented convolution operator named RISurConv to extract refined rotation invariant features. Finally, we build up network based on RISurConv for rotation invariant object classification and segmentation. In summary, our main contributions include:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 545, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 545, + 480, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 580 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 580 + ], + "type": "text", + "content": "- Rotation Invariant Surface Properties. We construct local triangle surfaces from the reference point and its neighbors, based on which we are able to design highly expressive rotation invariant surface properties;" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 582, + 480, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 582, + 480, + 616 + ], + "spans": [ + { + "bbox": [ + 138, + 582, + 480, + 616 + ], + "type": "text", + "content": "- RISurConv. We integrate the Rotation Invariant Surface Properties into an attention-augmented architecture named RISurConv comprising two self-attention layers to learn and generate refined rotation invariant features;" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 618, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 618, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 618, + 481, + 665 + ], + "type": "text", + "content": "- Extensive experiments on a variety of classification and segmentation tasks. Our approach shows supreme performance, surpassing the state-of-the-art by a large margin under challenging rotation scenarios including an analysis of rotation-invariant features and an ablation study of our neural" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Z. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 148, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 148, + 116, + 479, + 140 + ], + "type": "text", + "content": "network to provide insights into the key factors contributing to the exceptional performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 159, + 242, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 159, + 242, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 159, + 242, + 172 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 186, + 479, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 186, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 186, + 479, + 209 + ], + "type": "text", + "content": "In this section, we review the representative works in 3D point cloud deep learning and rotation invariant learning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 210, + 480, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 210, + 480, + 305 + ], + "spans": [ + { + "bbox": [ + 130, + 210, + 480, + 305 + ], + "type": "text", + "content": "3D Point Cloud Deep Learning. 3D point clouds is a more compact and intuitive representation for feature learning. PointNet [23] pioneered a point cloud convolution with global features by max-pooling per-point features from MLPs, and follow-up works [17, 25, 38] have focused on exploring convolution kernels that exploit geometric features [29], adding edges on top of points [33], parameterizing convolution using polynomials [38], and leveraging shape context [37]. Some methods are designed to combine with recurrent neural networks [13] and sequence models [18]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 307, + 481, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 481, + 579 + ], + "type": "text", + "content": "In recent years, attention mechanism has become increasingly popular in various natural language processing (NLP) [32] and computer vision [6] tasks. With their ability to handle sequential and spatial information, attention mechanism has also shown promise in 3D point cloud processing tasks. Pioneer works that apply attention mechanism in point clouds include Point Transformer [43] and PCT [9], which use a self-attention mechanism to learn local and global features of point clouds. Other works have explored different types of transformers. Inspired by BERT [5], Point-BERT [39] was proposed to pre-train pure Transformer-based models with a Mask Point Modeling task for point cloud classification. Dual Transformer Network [11] aggregated the point-wise and channel-wise multi-head self-attention models simultaneously such that long-range context dependencies can be captured by investigating the point-wise and channel-wise relationships. Overall, these developments show promise in enhancing the efficiency and accuracy of point cloud analysis for a variety of applications. Readers can refer to the survey [10] for a comprehensive overview of 3D point cloud deep learning. However, most of these methods lack the property of rotation invariance which is critical for object classification and segmentation as the object can be viewed at any angle in reality. To address this issue, a common approach is to increase the training data by augmenting it with arbitrary rotations [17, 25, 33]. However, such approach has difficulty in generalizing predictions to unseen rotations, resulting in deteriorated performance. Therefore, it would be desirable to design a specific convolution possessing rotation-invariant features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Rotation Invariant Learning. Various methods have been proposed to address the issue of rotation invariance in feature learning of point clouds. For instance, Rao et al. [27] used a spherical domain to define a rotation-invariant convolution for point clouds. However, the discretized sphere is sensitive to global rotations, which can lead to a notable drop in performance for objects with arbitrary rotations. To overcome this limitation, Poulenard et al. [22] integrated spherical harmonics to their convolution. Similarly, Chen et al. [3] proposed a" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "RISurConv" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 159, + 113, + 455, + 198 + ], + "blocks": [ + { + "bbox": [ + 159, + 113, + 455, + 198 + ], + "lines": [ + { + "bbox": [ + 159, + 113, + 455, + 198 + ], + "spans": [ + { + "bbox": [ + 159, + 113, + 455, + 198 + ], + "type": "image", + "image_path": "bbfdb259e870195d441844ef27250a76a3317bffbe790e18e1aa53036e6cd5e4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 205, + 482, + 251 + ], + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 251 + ], + "type": "text", + "content": "Fig. 1: Rotation Invariant Surface Property (RISP) construction: Given a point " + }, + { + "bbox": [ + 130, + 205, + 482, + 251 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 205, + 482, + 251 + ], + "type": "text", + "content": " as the reference point, " + }, + { + "bbox": [ + 130, + 205, + 482, + 251 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 205, + 482, + 251 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 205, + 482, + 251 + ], + "type": "inline_equation", + "content": "K = 8" + }, + { + "bbox": [ + 130, + 205, + 482, + 251 + ], + "type": "text", + "content": " in this example) nearest points are selected (middle). For each neighbor " + }, + { + "bbox": [ + 130, + 205, + 482, + 251 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 205, + 482, + 251 + ], + "type": "text", + "content": ", two adjacent neighbors " + }, + { + "bbox": [ + 130, + 205, + 482, + 251 + ], + "type": "inline_equation", + "content": "x_{i-1}" + }, + { + "bbox": [ + 130, + 205, + 482, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 205, + 482, + 251 + ], + "type": "inline_equation", + "content": "x_{i+1}" + }, + { + "bbox": [ + 130, + 205, + 482, + 251 + ], + "type": "text", + "content": " are used to form two triangular local surfaces (right), based on which rotation invariant properties are constructed." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 275, + 482, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 275, + 482, + 516 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 482, + 516 + ], + "type": "text", + "content": "hierarchical clustering scheme to encode the relative angles between two-point vectors and used vector norm to maintain rotation invariance. Zhang et al. [41] presented a simple convolution named RICnv that operates on handcrafted features built from Euclidean distances and angles that are rotation invariant by nature. However, this approach only considers local features resulting in accuracy degradation. GCACnv [40] addressed this limitation by building a global context-aware convolution based on anchors and Local Reference Frame (LRF) to achieve rotation invariance. RI-GCN [15] learned rotation-invariant local descriptors and applied graph convolutional neural networks to aggregate local features based on LRF. Thomas [30] also relied on LRF and used a multiple alignment scheme to achieve better results. RIF [16] presented a framework to construct both local and global features based on distances, angles, and reference points. To remove the uncertainty of LRF in x and y directions, RICnv++ [42] extracts informative features based on local reference axis (LRA). However, the LRF/LRA and global reference used in these methods may not be stable enough limiting their overall performance. We can clearly see the accuracy discrepancy of the object classification task on ModelNet40 dataset [36]. State-of-the-art translation-invariant convolutions such as PointNet++ [25] or Point Transformer v2 [34] achieve 89%-94% of accuracy while rotation-invariant convolutions only report up to 87%-91% of accuracy [14,42]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 516, + 481, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 516, + 481, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 516, + 481, + 540 + ], + "type": "text", + "content": "Our target in this paper is not only closing this performance gap but also surpassing the state-of-the-art methods while maintaining rotation invariance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 563, + 370, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 563, + 370, + 578 + ], + "spans": [ + { + "bbox": [ + 131, + 563, + 370, + 578 + ], + "type": "text", + "content": "3 Rotation Invariant Surface Property" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "In this section, we detail the Rotation Invariant Surface Property construction which is the first step of our method. Our goal for this step is to design highly expressive rotation invariant properties from underlying local surfaces. Different from previous works that rely on pointwise operations, we construct local surfaces around the reference point to better capture the local surface structure, based on which we then extract more expressive rotation invariant properties." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Z. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 188 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 188 + ], + "type": "text", + "content": "The construction of rotation invariant surface property is shown in Fig. 1. Given a point " + }, + { + "bbox": [ + 131, + 116, + 482, + 188 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 131, + 116, + 482, + 188 + ], + "type": "text", + "content": " as the reference point (red), we get " + }, + { + "bbox": [ + 131, + 116, + 482, + 188 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 116, + 482, + 188 + ], + "type": "text", + "content": " nearest neighbors to form a local point set. In this case, " + }, + { + "bbox": [ + 131, + 116, + 482, + 188 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 116, + 482, + 188 + ], + "type": "text", + "content": " is 8. For each neighbor " + }, + { + "bbox": [ + 131, + 116, + 482, + 188 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 131, + 116, + 482, + 188 + ], + "type": "text", + "content": ", two adjacent neighbors, " + }, + { + "bbox": [ + 131, + 116, + 482, + 188 + ], + "type": "inline_equation", + "content": "x_{i-1}" + }, + { + "bbox": [ + 131, + 116, + 482, + 188 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 116, + 482, + 188 + ], + "type": "inline_equation", + "content": "x_{i+1}" + }, + { + "bbox": [ + 131, + 116, + 482, + 188 + ], + "type": "text", + "content": ", are identified based on the Euclidean distances, forming two triangular local surfaces, as shown in the figure. Then, we construct the rotation invariant surface properties (RISP) as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 235, + 198, + 481, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 198, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 235, + 198, + 481, + 217 + ], + "type": "interline_equation", + "content": "\\operatorname {R I S P} \\left(x _ {i}\\right) = \\left[ L _ {0}, \\phi_ {1}, \\phi_ {2}, \\phi_ {3}, \\phi_ {4}, \\phi_ {5}, \\right. \\tag {1}", + "image_path": "4dbe80728e3ccd96f89641091e84a4cb7a31a730b954c0879a74fe323687b2e0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 264, + 213, + 378, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 213, + 378, + 224 + ], + "spans": [ + { + "bbox": [ + 264, + 213, + 378, + 224 + ], + "type": "interline_equation", + "content": "\\left. \\alpha_ {1}, \\alpha_ {2}, \\beta_ {1}, \\beta_ {2}, \\theta_ {1}, \\theta_ {2}, \\gamma_ {1}, \\gamma_ {2} \\right]", + "image_path": "f4db25b1cd3e8cb625941a153f47e007ff5472ff61c4a1589b11ff8d85b855a8.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 234, + 479, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 234, + 479, + 270 + ], + "spans": [ + { + "bbox": [ + 131, + 234, + 479, + 270 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 234, + 479, + 270 + ], + "type": "inline_equation", + "content": "L_{0}" + }, + { + "bbox": [ + 131, + 234, + 479, + 270 + ], + "type": "text", + "content": " measures the distance from reference " + }, + { + "bbox": [ + 131, + 234, + 479, + 270 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 131, + 234, + 479, + 270 + ], + "type": "text", + "content": " to neighbor " + }, + { + "bbox": [ + 131, + 234, + 479, + 270 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 131, + 234, + 479, + 270 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 234, + 479, + 270 + ], + "type": "inline_equation", + "content": "\\phi_{1}" + }, + { + "bbox": [ + 131, + 234, + 479, + 270 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 131, + 234, + 479, + 270 + ], + "type": "inline_equation", + "content": "\\phi_{5}" + }, + { + "bbox": [ + 131, + 234, + 479, + 270 + ], + "type": "text", + "content": " measure the two triangles as well as the relationship between the two triangle surfaces with regard to the edge " + }, + { + "bbox": [ + 131, + 234, + 479, + 270 + ], + "type": "inline_equation", + "content": "\\overrightarrow{px_i}" + }, + { + "bbox": [ + 131, + 234, + 479, + 270 + ], + "type": "text", + "content": " in the Euclidean space:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 205, + 292, + 481, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 292, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 205, + 292, + 481, + 304 + ], + "type": "interline_equation", + "content": "\\phi_ {1} = \\angle (\\overrightarrow {x _ {i - 1} p}, \\overrightarrow {x _ {i} p}), \\phi_ {2} = \\angle (\\overrightarrow {x _ {i + 1} p}, \\overrightarrow {x _ {i} p}), \\tag {2}", + "image_path": "20833637c9472ed7bec063ad064d5391d109bd259f66d8dceca366b0e9c2a6e8.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 205, + 307, + 411, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 307, + 411, + 319 + ], + "spans": [ + { + "bbox": [ + 205, + 307, + 411, + 319 + ], + "type": "interline_equation", + "content": "\\phi_ {3} = \\angle (\\overrightarrow {x _ {i - 1} x _ {i}}, \\overrightarrow {x _ {i - 1} p}), \\phi_ {4} = \\angle (\\overrightarrow {x _ {i + 1} p}, \\overrightarrow {x _ {i + 1} x _ {i}}),", + "image_path": "e8c8df65c731d29088578adefe6ab6cc153acf4ec76762cb69003a30c8f82e4c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 205, + 321, + 357, + 333 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 321, + 357, + 333 + ], + "spans": [ + { + "bbox": [ + 205, + 321, + 357, + 333 + ], + "type": "interline_equation", + "content": "\\phi_ {5} = \\angle (\\overrightarrow {x _ {i + 1} p} \\times \\overrightarrow {x _ {i} p}, \\overrightarrow {x _ {i - 1} p} \\times \\overrightarrow {x _ {i} p}),", + "image_path": "3b011f4a0f5c116ffe213d9cba97635314b98aad5526ea59dfd9b0d5eaaded45.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 342, + 479, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 342, + 479, + 378 + ], + "spans": [ + { + "bbox": [ + 131, + 342, + 479, + 378 + ], + "type": "text", + "content": "while other properties describe the two surfaces in the tangent space, e.g., normal vectors can define the directions in which the surface is bending away from the tangent space:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 212, + 389, + 480, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 389, + 480, + 401 + ], + "spans": [ + { + "bbox": [ + 212, + 389, + 480, + 401 + ], + "type": "interline_equation", + "content": "\\alpha_ {1} = \\angle (\\overrightarrow {n _ {p}}, \\overrightarrow {x _ {i} p}), \\alpha_ {2} = \\angle (\\overrightarrow {n _ {p}}, \\overrightarrow {x _ {i - 1} p}), \\tag {3}", + "image_path": "7f14cc903bee8c6df204b925600a9cd4a00feab52565c1da95b9e6677dbfbee4.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 212, + 403, + 373, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 403, + 373, + 415 + ], + "spans": [ + { + "bbox": [ + 212, + 403, + 373, + 415 + ], + "type": "interline_equation", + "content": "\\beta_ {1} = \\angle (\\vec {n _ {i}}, \\vec {x _ {i}} \\vec {p}), \\beta_ {2} = \\angle (\\vec {n _ {i}}, \\vec {x _ {i - 1}} \\vec {x _ {i}}),", + "image_path": "666ae03849d4e3cdce93d211557cc3f9b139ee9d0382fd299c549ced7718b4d3.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 212, + 418, + 403, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 418, + 403, + 430 + ], + "spans": [ + { + "bbox": [ + 212, + 418, + 403, + 430 + ], + "type": "interline_equation", + "content": "\\theta_ {1} = \\angle (\\overrightarrow {n _ {i - 1}}, \\overrightarrow {x _ {i - 1} p}), \\theta_ {2} = \\angle (\\overrightarrow {n _ {i - 1}}, \\overrightarrow {x _ {i - 1} x _ {i} ^ {\\prime}}),", + "image_path": "e193693ee879a208e2860883528efc64258bd0e8fe8945c8ee196aeb029e4fcd.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 212, + 434, + 403, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 434, + 403, + 445 + ], + "spans": [ + { + "bbox": [ + 212, + 434, + 403, + 445 + ], + "type": "interline_equation", + "content": "\\gamma_ {1} = \\angle (\\overrightarrow {n _ {i + 1}}, \\overrightarrow {x _ {i + 1} x _ {i}}), \\gamma_ {2} = \\angle (\\overrightarrow {n _ {i + 1}}, \\overrightarrow {x _ {i + 1} p}).", + "image_path": "abd6d4808f78a573871aa9052e9c39c7676d136cd45dc3600040cdde236d9142.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 455, + 479, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 455, + 479, + 491 + ], + "spans": [ + { + "bbox": [ + 131, + 455, + 479, + 491 + ], + "type": "text", + "content": "RISP is able to fully describe the dual triangles as well as their relationship along different directions. Readers can refer to the supplementary for the mathematical proof of the completeness of RISP defined in Eq. (1)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 502, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 502, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 502, + 482, + 586 + ], + "type": "text", + "content": "Number of local triangle surfaces. Throughout the paper, we define the number of local triangle surfaces as 2. Such setting is due to the fact that we wish to capture local surface as much as possible to extract more useful information to reflect the real shape of the object. So we construct 2 triangles rather than single triangle. However, this does not mean the more the better since more triangles can result in distorted surface and affect the computing efficiency. Please refer to the ablation study section for more details." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 604, + 307, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 604, + 307, + 618 + ], + "spans": [ + { + "bbox": [ + 131, + 604, + 307, + 618 + ], + "type": "text", + "content": "4 The RISurConv Operator" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Based on the Rotation Invariant Surface Properties (RISP), we are able to build up RISurConv. The main steps are detailed in Fig. 2. To start, we utilize a farthest point sampling strategy to generate uniformly distributed representative" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "RISurConv" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 112, + 475, + 228 + ], + "blocks": [ + { + "bbox": [ + 141, + 112, + 475, + 228 + ], + "lines": [ + { + "bbox": [ + 141, + 112, + 475, + 228 + ], + "spans": [ + { + "bbox": [ + 141, + 112, + 475, + 228 + ], + "type": "image", + "image_path": "d1b087ed69e990380fbe05236df59dd01c39ae9cb602fc2950de414963679502.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 235, + 482, + 324 + ], + "lines": [ + { + "bbox": [ + 130, + 235, + 482, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 235, + 482, + 324 + ], + "type": "text", + "content": "Fig. 2: RISurConv operator. For a local point set with " + }, + { + "bbox": [ + 130, + 235, + 482, + 324 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 235, + 482, + 324 + ], + "type": "text", + "content": " as the reference (red), K nearest neighbors are labelled as blue. Then, we compute the Rotation Invariant Surface Properties at each neighbor by constructing local dual triangle surfaces (Sec. 3), which is embedded to a high-dimensional space by a shared multi-layer perceptron (MLP) followed by a self-attention layer to produce refined features. Concatenated with previous layer features (if any), the features of these local points are further passed to MLPs, which are then summarized by maxpooling. To further refine the features, another self-attention layer follows." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 336, + 291, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 336, + 291, + 348 + ], + "spans": [ + { + "bbox": [ + 133, + 336, + 291, + 348 + ], + "type": "text", + "content": "Algorithm 1 RISurConv operator." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 350, + 480, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 350, + 480, + 371 + ], + "spans": [ + { + "bbox": [ + 132, + 350, + 480, + 371 + ], + "type": "text", + "content": "Input: Reference point " + }, + { + "bbox": [ + 132, + 350, + 480, + 371 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 132, + 350, + 480, + 371 + ], + "type": "text", + "content": ", point set " + }, + { + "bbox": [ + 132, + 350, + 480, + 371 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 132, + 350, + 480, + 371 + ], + "type": "text", + "content": ", point features " + }, + { + "bbox": [ + 132, + 350, + 480, + 371 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{prev}" + }, + { + "bbox": [ + 132, + 350, + 480, + 371 + ], + "type": "text", + "content": " from previous layer (if any) \nOutput: Convoluted features " + }, + { + "bbox": [ + 132, + 350, + 480, + 371 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 373, + 255, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 373, + 255, + 395 + ], + "spans": [ + { + "bbox": [ + 136, + 373, + 255, + 395 + ], + "type": "text", + "content": "1: " + }, + { + "bbox": [ + 136, + 373, + 255, + 395 + ], + "type": "inline_equation", + "content": "\\mathbf{f} \\gets \\{\\mathrm{RISP}(x_i) : \\forall x_i \\in \\Omega\\}" + }, + { + "bbox": [ + 136, + 373, + 255, + 395 + ], + "type": "text", + "content": " (Sec. 3)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 276, + 373, + 480, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 373, + 480, + 395 + ], + "spans": [ + { + "bbox": [ + 276, + 373, + 480, + 395 + ], + "type": "text", + "content": "* Construct Rotation Invariant Surface Properties" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 396, + 203, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 396, + 203, + 407 + ], + "spans": [ + { + "bbox": [ + 136, + 396, + 203, + 407 + ], + "type": "text", + "content": "2: " + }, + { + "bbox": [ + 136, + 396, + 203, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{f}\\gets \\mathrm{MLP}(\\mathbf{f})" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 242, + 396, + 478, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 396, + 478, + 407 + ], + "spans": [ + { + "bbox": [ + 242, + 396, + 478, + 407 + ], + "type": "text", + "content": "* Embed each feature to a high-dimensional feature space" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 136, + 407, + 195, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 407, + 195, + 417 + ], + "spans": [ + { + "bbox": [ + 136, + 407, + 195, + 417 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 136, + 407, + 195, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{f}\\gets \\mathrm{SA}(\\mathbf{f})" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 407, + 478, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 407, + 478, + 418 + ], + "spans": [ + { + "bbox": [ + 315, + 407, + 478, + 418 + ], + "type": "text", + "content": "* Refine features via self-attention layer" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 136, + 418, + 208, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 418, + 208, + 439 + ], + "spans": [ + { + "bbox": [ + 136, + 418, + 208, + 439 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 136, + 418, + 208, + 439 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{in}\\gets [\\mathbf{f}_{prev},\\mathbf{f}]" + }, + { + "bbox": [ + 136, + 418, + 208, + 439 + ], + "type": "text", + "content": " (if any)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 219, + 418, + 480, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 418, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 219, + 418, + 480, + 439 + ], + "type": "text", + "content": "* Concatenate the features from the local and the previous layer" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 136, + 439, + 217, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 439, + 217, + 450 + ], + "spans": [ + { + "bbox": [ + 136, + 439, + 217, + 450 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 136, + 439, + 217, + 450 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{out} \\gets \\mathrm{MLP}(\\mathbf{f}_{in})" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 391, + 439, + 477, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 439, + 477, + 450 + ], + "spans": [ + { + "bbox": [ + 391, + 439, + 477, + 450 + ], + "type": "text", + "content": "* Feature embedding" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 136, + 450, + 235, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 450, + 235, + 472 + ], + "spans": [ + { + "bbox": [ + 136, + 450, + 235, + 472 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 136, + 450, + 235, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{out}\\gets" + }, + { + "bbox": [ + 136, + 450, + 235, + 472 + ], + "type": "text", + "content": " maxpool " + }, + { + "bbox": [ + 136, + 450, + 235, + 472 + ], + "type": "inline_equation", + "content": "(\\mathbf{f}_{out})" + }, + { + "bbox": [ + 136, + 450, + 235, + 472 + ], + "type": "text", + "content": " return SA " + }, + { + "bbox": [ + 136, + 450, + 235, + 472 + ], + "type": "inline_equation", + "content": "(\\mathbf{f}_{out})" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 402, + 451, + 481, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 451, + 481, + 461 + ], + "spans": [ + { + "bbox": [ + 402, + 451, + 481, + 461 + ], + "type": "text", + "content": "* Maxpool features" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 371, + 461, + 481, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 461, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 371, + 461, + 481, + 472 + ], + "type": "text", + "content": "* Self-attention and return" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 499, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 499, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 499, + 482, + 548 + ], + "type": "text", + "content": "points. From this initial set, we perform K-nearest neighbor searches to obtain local point sets. Let us denote a local point set by " + }, + { + "bbox": [ + 130, + 499, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\Omega = \\{x_{i}\\}" + }, + { + "bbox": [ + 130, + 499, + 482, + 548 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 499, + 482, + 548 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 499, + 482, + 548 + ], + "type": "text", + "content": " represents the 3D coordinates of point " + }, + { + "bbox": [ + 130, + 499, + 482, + 548 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 499, + 482, + 548 + ], + "type": "text", + "content": ". We define the attention-augmented convolution operation as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 233, + 557, + 481, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 557, + 481, + 571 + ], + "spans": [ + { + "bbox": [ + 233, + 557, + 481, + 571 + ], + "type": "interline_equation", + "content": "\\mathbf {f} (\\Omega) = S A \\left(\\sigma \\left(\\mathcal {A} \\left(\\left\\{\\mathcal {T} \\left(\\mathbf {f} _ {x _ {i}}\\right): \\forall i \\right\\}\\right)\\right)\\right). \\tag {4}", + "image_path": "cdb455508edb5973c9f26ca55baeb11d912a8f19493e8135a393c8cddefbf93c.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 579, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 579, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 579, + 482, + 640 + ], + "type": "text", + "content": "This formula indicates that features of each point in the point set are first transformed by " + }, + { + "bbox": [ + 130, + 579, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 579, + 482, + 640 + ], + "type": "text", + "content": " before being aggregated by the aggregation function " + }, + { + "bbox": [ + 130, + 579, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 579, + 482, + 640 + ], + "type": "text", + "content": " and passed to an activation function " + }, + { + "bbox": [ + 130, + 579, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 579, + 482, + 640 + ], + "type": "text", + "content": ". SA is a channel-wise self-attention layer used to output refined features. We set the input features to our expressive rotation-invariant features " + }, + { + "bbox": [ + 130, + 579, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{x_i} = \\mathrm{RISP}(x_i)" + }, + { + "bbox": [ + 130, + 579, + 482, + 640 + ], + "type": "text", + "content": ". We define the transformation function as" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 256, + 649, + 481, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 649, + 481, + 663 + ], + "spans": [ + { + "bbox": [ + 256, + 649, + 481, + 663 + ], + "type": "interline_equation", + "content": "\\mathcal {T} \\left(\\mathbf {f} _ {x _ {i}}\\right) = \\mathbf {w} _ {i} \\cdot \\mathbf {f} _ {x _ {i}} = \\mathbf {f} _ {x _ {i}} ^ {\\prime} \\tag {5}", + "image_path": "e113b0b37d67d38e2f0950f86e2e5edd3f320e9ee8ed3ed42a03683fb22ae302.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Z. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "content": " indicates the element-wise product, and " + }, + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_i" + }, + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "content": " is the weight parameter to be learned by the network. Our transformation function is similar to PointNet++. A popular choice of the aggregation function " + }, + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "content": " is maxpooling, which supports permutation invariance of the input point features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 481, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 481, + 319 + ], + "type": "text", + "content": "To proceed with feature learning and be invariant to point permutation, we construct surfaces for each neighbor from which we construct RISPs. In Fig. 2, there are eight neighbors, so the resulted RISPs are in size of " + }, + { + "bbox": [ + 130, + 164, + 481, + 319 + ], + "type": "inline_equation", + "content": "8 \\times 14" + }, + { + "bbox": [ + 130, + 164, + 481, + 319 + ], + "type": "text", + "content": ". Here, 14 is the length of the RISP as defined in Eq. (1) and 8 is the neighborhood size. Since RISPs are used as the input which are already rotation invariant, RISurConv is rotation invariant by nature. We then embed the RISPs into feature space using two layers of MLPs followed by a self-attention (SA) layer for feature refinement. We concatenate the refined features with previous layer features (if any), and embed the concatenated features into higher dimensional space via MLPs again. After the maxpooling, the aggregated features are passed into another SA layer again for further refinement. The detailed steps are shown in the Algorithm 1. Here, the two SA layers used in RISurConv are the standard SA module of the transformer [32]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 319, + 482, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 465 + ], + "type": "text", + "content": "Self Attention Layers. In RISurConv there are two Self-Attention (SA) layers by which the extracted features are enhanced. Below, we provide a more detailed illustration of the structure of the SA layers. The SA structure is the standard SA module of the famous transformer [32] (Attention is all you need) which runs twice in RISurConv. The first SA is for the K neighbors, and the second SA works for the N representative points. The input tensor shape is " + }, + { + "bbox": [ + 130, + 319, + 482, + 465 + ], + "type": "inline_equation", + "content": "[B,N,K,C_0]" + }, + { + "bbox": [ + 130, + 319, + 482, + 465 + ], + "type": "text", + "content": " for the first SA which goes into linear layers and outputs " + }, + { + "bbox": [ + 130, + 319, + 482, + 465 + ], + "type": "inline_equation", + "content": "Q,K" + }, + { + "bbox": [ + 130, + 319, + 482, + 465 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 319, + 482, + 465 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 319, + 482, + 465 + ], + "type": "text", + "content": " respectively in shape of " + }, + { + "bbox": [ + 130, + 319, + 482, + 465 + ], + "type": "inline_equation", + "content": "[B,N,K,C_0]" + }, + { + "bbox": [ + 130, + 319, + 482, + 465 + ], + "type": "text", + "content": ". According to " + }, + { + "bbox": [ + 130, + 319, + 482, + 465 + ], + "type": "inline_equation", + "content": "\\text{att}(Q,K,V) = \\text{Softmax}(QK^{\\top} / \\text{sqrt}(d_k))V" + }, + { + "bbox": [ + 130, + 319, + 482, + 465 + ], + "type": "text", + "content": ", we have the 1st attention score in " + }, + { + "bbox": [ + 130, + 319, + 482, + 465 + ], + "type": "inline_equation", + "content": "[B,N,K,K]" + }, + { + "bbox": [ + 130, + 319, + 482, + 465 + ], + "type": "text", + "content": ". By multiplying with " + }, + { + "bbox": [ + 130, + 319, + 482, + 465 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 319, + 482, + 465 + ], + "type": "text", + "content": ", we get the refined feature in " + }, + { + "bbox": [ + 130, + 319, + 482, + 465 + ], + "type": "inline_equation", + "content": "[B,N,K,C_0]" + }, + { + "bbox": [ + 130, + 319, + 482, + 465 + ], + "type": "text", + "content": ". The same goes with the second SA where the input shape is " + }, + { + "bbox": [ + 130, + 319, + 482, + 465 + ], + "type": "inline_equation", + "content": "[B,N,C]" + }, + { + "bbox": [ + 130, + 319, + 482, + 465 + ], + "type": "text", + "content": " and the attention score shape is " + }, + { + "bbox": [ + 130, + 319, + 482, + 465 + ], + "type": "inline_equation", + "content": "[B,N,N]" + }, + { + "bbox": [ + 130, + 319, + 482, + 465 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 483, + 282, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 282, + 496 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 282, + 496 + ], + "type": "text", + "content": "5 RISurConv Networks" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "We employed the RISurConv to develop neural networks for object classification and segmentation, as illustrated in Fig. 3. Our classification network follows a standard architecture similar to the single-scale grouping version of PointNet++, consisting of five consecutive layers of RISurConv followed by a transformer encoder [32] with 8 heads to enhance the extracted features. Finally, we connect the network to fully connected layers to produce the probability map. One of the major advantages of RISurConv is its ability to handle arbitrary rotation and point orders, which enables us to place each RISurConv layer consecutively without the need for complex preprocessing steps. Additionally, we apply batch normalization and ReLU activation to each convolution layer by default. The segmentation network utilizes an encoder-decoder structure similar to the U-Net [28]. We adopt the same definition of deconvolution as RISurConv. The key distinction lies in that the convolution produces a point subset with a greater" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "RISurConv" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 113, + 479, + 222 + ], + "blocks": [ + { + "bbox": [ + 135, + 113, + 479, + 222 + ], + "lines": [ + { + "bbox": [ + 135, + 113, + 479, + 222 + ], + "spans": [ + { + "bbox": [ + 135, + 113, + 479, + 222 + ], + "type": "image", + "image_path": "87105d49eaf443ae2b40917f81563b05648fb1b683c59903ffa1a82c5b00d65c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 228, + 482, + 275 + ], + "lines": [ + { + "bbox": [ + 130, + 228, + 482, + 275 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 482, + 275 + ], + "type": "text", + "content": "Fig. 3: Our neural network architecture comprises five RISurConv layers to extract rotation invariant features followed by a Transformer Encoder to enhance the learnt features before fully connected layers for object classification. We add a decoder with skip connections for segmentation task." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 300, + 482, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 350 + ], + "type": "text", + "content": "number of feature channels and less points while the deconvolution outputs to a point set with more points compared to the input with fewer feature channels. Please refer to the supplementary for more details regarding the network architecture." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 377, + 230, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 377, + 230, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 377, + 230, + 392 + ], + "type": "text", + "content": "6 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 412, + 481, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 481, + 460 + ], + "type": "text", + "content": "We report our evaluation results in this section. Our network is implemented in PyTorch using an Adam optimizer with initial learning rate of 0.001 for both classification and segmentation. Our training is executed on a computer with an Intel(R) Core(TM) i7-10700K CPU equipped with a NVIDIA GTX 3090 GPU." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 462, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 665 + ], + "type": "text", + "content": "We evaluated the performance of our method extensively on a series of classification and segmentation benchmarks. To train our model, we utilized 450 epochs for classification and 250 epochs for segmentation for most datasets except the Car and Chair categories of the fine-grained FG3D dataset due to small size and subtle variation where we applied 900 epochs. Our method can achieve convergence within 11 hours for classification and 24 hours for segmentation. To assess the robustness of our method, we followed the experimental design proposed by [8] and conducted experiments in three cases. The first case involved training and testing with data augmented with rotation about the gravity axis " + }, + { + "bbox": [ + 130, + 462, + 482, + 665 + ], + "type": "inline_equation", + "content": "(\\mathrm{z} / \\mathrm{z})" + }, + { + "bbox": [ + 130, + 462, + 482, + 665 + ], + "type": "text", + "content": ", which is a common approach for evaluating translation-invariant point cloud learning methods. The second case involved training and testing with data augmented with arbitrary SO3 rotations (SO3/SO3), and the third case involved training with data by z-rotations and testing with data by SO3 rotations (z/SO3), which are used to evaluate rotation invariance. A rotation-invariant method should produce consistent results across all three cases. Our method demonstrated superior performance across all three cases, highlighting its robustness and effectiveness." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Z. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 336, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 336, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 336, + 128 + ], + "type": "text", + "content": "6.1 Human-made Object Classification" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 482, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 482, + 239 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 482, + 239 + ], + "type": "text", + "content": "ModelNet40 [36] that comprises 9843 training models and 2468 test models, divided into 40 categories is one of the most popular datasets for point cloud classification evaluation. The input data size is 1024 with each point equipped with (x; y; z; nx; ny; nz) which are 3D coordinates and 3D normals in the Euclidean space. Note that normals are optional to our method. When normals are not available (w/o n), we use weighted eigen vector corresponding to the smallest eigen value as the normal. This strategy has been used in SHOT and RICConv++ [42] which is highly robust to noise and efficient to compute." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 240, + 482, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 240, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 482, + 289 + ], + "type": "text", + "content": "The experimental results are shown in Tab. 1. We use two criteria for evaluation: overall accuracy and accuracy standard deviation (Std.). True rotation invariant methods are expected to be unaffected by the rotations present in the training and testing data, resulting in low accuracy deviation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "text", + "content": "From Tab. 1, we see that our method surpasses the state-of-the-art performance in all cases while maintaining zero accuracy deviation. It is worth noting that our method surpasses both rotation invariant and non-rotation invariant methods. To the best of our knowledge, this is the first work that achieves such high accuracy. It outperforms the state-of-the-art rotation invariant method RIConv++ by " + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "inline_equation", + "content": "4.7\\%" + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "text", + "content": ", and outperforms the state-of-the-art non-rotation invariant method Point Transformer v2 by " + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "text", + "content": " under " + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\mathrm{z / z}" + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "text", + "content": " case." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 399, + 326, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 399, + 326, + 411 + ], + "spans": [ + { + "bbox": [ + 131, + 399, + 326, + 411 + ], + "type": "text", + "content": "6.2 Real World Object Classification" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 426, + 282, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 282, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 282, + 592 + ], + "type": "text", + "content": "We also evaluate the classification performance on a real-world 3D point cloud dataset ScanObjectNN [31]. It is composed of 2902 point clouds categorized into 15 categories sampled from real world indoor scenes. For our evaluation, we use the processed files and choose the hardest variant PB_T50_RS with " + }, + { + "bbox": [ + 130, + 426, + 282, + 592 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 426, + 282, + 592 + ], + "type": "text", + "content": " bounding box translation, rotation around the gravity axis, and random scaling that result in rotated and partial data. The results are shown" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 291, + 462, + 478, + 590 + ], + "blocks": [ + { + "bbox": [ + 287, + 428, + 482, + 461 + ], + "lines": [ + { + "bbox": [ + 287, + 428, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 287, + 428, + 482, + 461 + ], + "type": "text", + "content": "Table 2: Comparisons of real world 3D point cloud classification on hardest variant of ScanObjectNN dataset (Overall Accuracy)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 291, + 462, + 478, + 590 + ], + "lines": [ + { + "bbox": [ + 291, + 462, + 478, + 590 + ], + "spans": [ + { + "bbox": [ + 291, + 462, + 478, + 590 + ], + "type": "table", + "html": "
MethodPB_T50_RS
z/zSO3/SO3z/SO3
PointNet [23]68.242.217.1
PointNet++ [25]77.960.115.8
PointCNN [17]78.551.814.9
DGCNN [33]78.163.416.1
RICov [41]68.168.368.3
GCACov [40]69.870.069.8
RICov++ [42]80.380.380.3
Ours (w/o normal)93.193.193.1
", + "image_path": "3b6c977530a0cfae5188ffb36b9a51b17c2506bdd8f72f71e9682d368ccaa3b7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "in Tab. 2. We see that RISurConv surpasses all the compared methods by a large margin. Particularly, our method significantly outperforms the state-of-the-art rotation invariant approach RICov++ [42] by " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "12.8\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": ". This verifies that our method is effective for both synthetic and real world data. Note that we only test the 'w/o normal' case as the normal vectors are not provided in the processed files of this dataset." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "RISurConv" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 149, + 157, + 468, + 396 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 1: Comparisons of the classification accuracy (\\%) on the ModelNet40 dataset. On average, our method has the best accuracy and lowest accuracy deviation in all cases (Overall Accuracy)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 157, + 468, + 396 + ], + "lines": [ + { + "bbox": [ + 149, + 157, + 468, + 396 + ], + "spans": [ + { + "bbox": [ + 149, + 157, + 468, + 396 + ], + "type": "table", + "html": "
MethodFormatInput Size Params.z/z↑SO3/ SO3↑z/ SO3↑Std.↓
TraditionalVoxNet [21]voxel3030.90M83.087.3-3.0
SubVolSup [24]voxel30317.00M88.582.736.628.4
PointNet [23]xyz1024 × 33.50M87.080.321.641.0
PointCNN [17]xyz1024 × 30.60M91.384.541.227.2
PointNet++ [25]xyz + nor1024 × 61.40M89.385.028.633.8
DGCNN [33]xyz1024 × 31.84M92.281.120.638.5
RS-CNN [20]xyz1024 × 31.41M90.382.648.722.1
Pt Transformer [43]xyz1024 × 3-93.785.950.119.1
Pt Transformer v2 [34]xyz1024 × 3-94.288.351.823.0
Rotation-invariantSpherical CNN [7]voxel2 × 6420.50M88.986.978.65.5
RICConv [41]xyz1024 × 30.70M86.586.486.40.1
SPHNet [22]xyz1024 × 32.90M87.087.686.60.5
SFCNN [27]xyz1024 × 3-91.490.184.83.5
ClusterNet [3]xyz1024 × 31.40M87.187.187.10.0
GCACov [40]xyz1024 × 30.41M89.089.289.10.0
RIF [16]xyz1024 × 3-89.489.389.40.0
RI-GCN [15]xyz + nor1024 × 64.38M91.091.091.00.0
RICConv++ [42]xyz1024 × 30.40M91.291.291.20.0
RICConv++ [42]xyz + nor1024 × 60.40M91.391.391.30.0
Ours (w/o normal)xyz1024 × 314.0M95.695.695.60.0
Ours (w/ normal)xyz + nor1024 × 614.0M96.096.096.00.0
", + "image_path": "6ab3adbb3aa800e993c6bf96dfc181373938e4823ee1857537d58e260a41f566.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 421, + 335, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 421, + 335, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 421, + 335, + 433 + ], + "type": "text", + "content": "6.3 Fine-Grained Object Classification" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 447, + 482, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 482, + 615 + ], + "type": "text", + "content": "Fine-grained object classification is more challenging because the differences between subcategories are subtle and require a high level of precision in distinguishing them. We conduct experiments on the three categories from FD3D dataset [19]: Airplane, Chair and Car. To the best of our knowledge, this is the first work to test the performance of rotation invariant methods on fine-grained 3D point cloud dataset. The class accuracies are shown in Tab. 3. Again, our method outperforms the state-of-the-art approaches by large margins for all rotation cases and all the categories. Specifically, our method improves non-rotation invariant method FG3D-Net by " + }, + { + "bbox": [ + 130, + 447, + 482, + 615 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 130, + 447, + 482, + 615 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 447, + 482, + 615 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 130, + 447, + 482, + 615 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 447, + 482, + 615 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 130, + 447, + 482, + 615 + ], + "type": "text", + "content": " for Airplane, Chair, and Car categories respectively under the " + }, + { + "bbox": [ + 130, + 447, + 482, + 615 + ], + "type": "inline_equation", + "content": "\\mathrm{z} / \\mathrm{z}" + }, + { + "bbox": [ + 130, + 447, + 482, + 615 + ], + "type": "text", + "content": " rotation case, and the accuracy consistency is well preserved. Compared to the rotation invariant methods, our method outperforms the latest RICov++ by " + }, + { + "bbox": [ + 130, + 447, + 482, + 615 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 130, + 447, + 482, + 615 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 447, + 482, + 615 + ], + "type": "inline_equation", + "content": "5.1\\%" + }, + { + "bbox": [ + 130, + 447, + 482, + 615 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 447, + 482, + 615 + ], + "type": "inline_equation", + "content": "9.2\\%" + }, + { + "bbox": [ + 130, + 447, + 482, + 615 + ], + "type": "text", + "content": " on the three categories. Such supreme performance shows that features extracted by RISurConv are not only rotation invariant but also highly expressive." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 638, + 324, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 638, + 324, + 651 + ], + "spans": [ + { + "bbox": [ + 130, + 638, + 324, + 651 + ], + "type": "text", + "content": "6.4 Part Segmentation on ShapeNet" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Z. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 146, + 474, + 320 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 3: Comparisons of fine-grained 3D point cloud classification on FG3D dataset (Class Accuracy)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 146, + 474, + 320 + ], + "lines": [ + { + "bbox": [ + 138, + 146, + 474, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 146, + 474, + 320 + ], + "type": "table", + "html": "
MethodModalityAirplaneChairCar
z/zSO3/SO3z/SO3z/zSO3/SO3z/SO3z/zSO3/SO3z/SO3
VoxNetVoxel79.257.528.773.354.716.768.242.217.1
MVCNNView82.666.921.976.363.714.671.951.814.9
View-GCNView87.456.622.779.750.219.773.744.418.1
RotationNetView89.150.829.878.536.218.472.545.328.3
Part4FeatureView82.649.132.277.152.230.273.435.023.8
FG3D-NetView89.461.126.980.043.218.474.056.420.8
PointNetPoint82.755.529.772.133.918.268.132.116.4
PointNet++Point87.357.125.678.140.416.070.350.420.8
RS-CNNPoint82.845.932.275.150.718.671.231.813.9
DGCNNPoint88.460.622.771.752.319.765.353.418.1
RICConvPoint85.885.885.876.376.376.367.367.367.3
RICConv++Point87.987.987.977.677.677.669.369.369.3
OursPoint91.591.591.582.782.782.778.578.578.5
", + "image_path": "6574cfce0b23e77f1976bbcb33cd24597c8524d6b29ac2cb096b778e0660fb91.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 367, + 273, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 273, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 273, + 510 + ], + "type": "text", + "content": "To evaluate the segmentation performance, we employed the ShapeNet dataset comprising 16880 CAD models distributed across 16 distinct categories. These models are annotated with 2 to 6 parts each, culminating in a dataset of 50 object parts. We follow the standard train/test split with 14006 models for training and 2874 models for testing, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 521, + 273, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 273, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 273, + 616 + ], + "type": "text", + "content": "The evaluation results are shown in Tab. 4. Our method outperforms both traditional translation-invariant and latest rotation invariant methods significantly in the SO3/SO3 and z/SO3 scenario. Our method outperforms the RICov++ by" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 290, + 425, + 475, + 606 + ], + "blocks": [ + { + "bbox": [ + 280, + 369, + 480, + 422 + ], + "lines": [ + { + "bbox": [ + 280, + 369, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 280, + 369, + 480, + 422 + ], + "type": "text", + "content": "Table 4: Comparisons of object part segmentation performed on ShapeNet dataset. The mean per-class IoU (mIoU, %) is used to measure the accuracy under two challenging rotation modes: SO3/SO3 and z/SO3." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 290, + 425, + 475, + 606 + ], + "lines": [ + { + "bbox": [ + 290, + 425, + 475, + 606 + ], + "spans": [ + { + "bbox": [ + 290, + 425, + 475, + 606 + ], + "type": "table", + "html": "
MethodInputSO3/SO3 z/SO3
PointNet [23]xyz74.437.8
PointNet++ [25]xyz+nor76.748.2
PointCNN [17]xyz71.434.7
DGCNN [33]xyz73.337.4
RS-CNN [20]xyz72.536.5
SpiderCNNxyz+nor72.342.9
RICConv [41]xyz75.575.3
GCACConv [40]xyz77.377.2
RI-GCN [15]xyz77.077.0
RIF [16]xyz79.479.2
RICConv++ [42]xyz80.380.3
RICConv++ [42]xyz+nor80.580.5
Ours (w/o normal)xyz81.381.3
Ours (w/ normal)xyz+nor81.581.5
", + "image_path": "bdcbc11ea25b657b6a1cc3afd5b6ff43295c7486fe37f0a3f16728bfc4f63515.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "content": "1.0% mIoU. This result aligns well with the performance reported in the object classification task. We also show the qualitative results by error maps in Fig. 4. The wrong segmentation points are plotted as red. It clearly shows that our predictions are the closest to the ground truth." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "RISurConv" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 154, + 113, + 460, + 310 + ], + "blocks": [ + { + "bbox": [ + 154, + 113, + 460, + 310 + ], + "lines": [ + { + "bbox": [ + 154, + 113, + 460, + 310 + ], + "spans": [ + { + "bbox": [ + 154, + 113, + 460, + 310 + ], + "type": "image", + "image_path": "04988c00896f6334af72dad5432b97b84261d7f7f701a72857e6be6179788f6d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 318, + 419, + 330 + ], + "lines": [ + { + "bbox": [ + 192, + 318, + 419, + 330 + ], + "spans": [ + { + "bbox": [ + 192, + 318, + 419, + 330 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparisons (Red indicates wrong)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 354, + 246, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 246, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 246, + 365 + ], + "type": "text", + "content": "6.5 Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 376, + 480, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 480, + 399 + ], + "type": "text", + "content": "We ablate some vital designs of our method on ModelNet40 [36] for an insightful exploration." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 400, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 533 + ], + "type": "text", + "content": "RISP Design. We first test the performance by turning on/off and adding different rotation invariant components used in the RISP construction (Algorithm 1). The results are shown in Tab. 5. Model A is our baseline setting with all rotation invariant features activated. Model B has only angle features but still achieves relatively high accuracy. Model C turns off both " + }, + { + "bbox": [ + 130, + 400, + 482, + 533 + ], + "type": "inline_equation", + "content": "L_{0}" + }, + { + "bbox": [ + 130, + 400, + 482, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 400, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\phi_{1\\dots 5}" + }, + { + "bbox": [ + 130, + 400, + 482, + 533 + ], + "type": "text", + "content": " with accuracy decreases to " + }, + { + "bbox": [ + 130, + 400, + 482, + 533 + ], + "type": "inline_equation", + "content": "90.9\\%" + }, + { + "bbox": [ + 130, + 400, + 482, + 533 + ], + "type": "text", + "content": ". This means that the distance " + }, + { + "bbox": [ + 130, + 400, + 482, + 533 + ], + "type": "inline_equation", + "content": "L_{0}" + }, + { + "bbox": [ + 130, + 400, + 482, + 533 + ], + "type": "text", + "content": " is not as important as angles. In Model D, we turn off the angle features on the tangent direction with only " + }, + { + "bbox": [ + 130, + 400, + 482, + 533 + ], + "type": "inline_equation", + "content": "L_{0}" + }, + { + "bbox": [ + 130, + 400, + 482, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 400, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\phi_{1\\dots 5}" + }, + { + "bbox": [ + 130, + 400, + 482, + 533 + ], + "type": "text", + "content": " are kept which is more like RICnv [41]. Compared to Model A, it shows that our proposed features is more effective than those by RICnv, and the improvement is explained by the additional consideration of the relations among the neighbor points. Model E adds more rotation invariant" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 136, + 586, + 294, + 662 + ], + "blocks": [ + { + "bbox": [ + 132, + 553, + 297, + 576 + ], + "lines": [ + { + "bbox": [ + 132, + 553, + 297, + 576 + ], + "spans": [ + { + "bbox": [ + 132, + 553, + 297, + 576 + ], + "type": "text", + "content": "Table 5: Ablation study on the RISP design." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 586, + 294, + 662 + ], + "lines": [ + { + "bbox": [ + 136, + 586, + 294, + 662 + ], + "spans": [ + { + "bbox": [ + 136, + 586, + 294, + 662 + ], + "type": "table", + "html": "
ModelL0φ1...5α1,2β1,2θ1,2γ1,2λμAcc.
A✓✓✓96.0
B✓✓95.5
C✓90.9
D✓✓88.2
E✓✓✓✓95.7
", + "image_path": "13e4a140c656df0ee08152b0bcbea8c073da1fa088e1a0ad7100f728ddeea88a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 319, + 586, + 479, + 662 + ], + "blocks": [ + { + "bbox": [ + 315, + 553, + 480, + 575 + ], + "lines": [ + { + "bbox": [ + 315, + 553, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 315, + 553, + 480, + 575 + ], + "type": "text", + "content": "Table 6: Ablation study on the Self-Attention module." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 586, + 479, + 662 + ], + "lines": [ + { + "bbox": [ + 319, + 586, + 479, + 662 + ], + "spans": [ + { + "bbox": [ + 319, + 586, + 479, + 662 + ], + "type": "table", + "html": "
ModelSA1SA2TransformerEncoderAcc.
A✓✓✓96.0
B✓✓95.6
C✓✓95.2
D✓✓94.3
E92.8
", + "image_path": "f1155cc4c4c886c05df96ef5fcd8ac30bf78f29fec0266d378cee08e83cc1253.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Z. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "features. " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " represent the angles " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\angle (\\pmb{n}_{\\pmb{p}},\\pmb{n}_{i + 1})" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\angle (\\pmb {n}_p,\\pmb {n}_{i - 1})" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " respectively. The results show that the accuracy does not increase. This verifies the completeness of the proposed RISP." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 156, + 480, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 156, + 480, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 156, + 480, + 323 + ], + "type": "text", + "content": "Self-Attention Effects. We employ multiple self-attention (SA) modules to help produce refined rotation invariant features. So it is necessary to analyse the effects of different SA modules. In RISurConv, there are two SA modules and we name them as SA1 and SA2 respectively. Before fully connected layers, there is a Transformer Encoder module. We test these three modules by turning off one of them each time. From Tab. 6, we see that when removing the first SA module in RISurConv, the performance drops a bit, while removing the second SA module also decreases the accuracy. This indicates that both SA1 and SA2 are important for the success of our method. In addition, we also test the performance when Transformer Encoder is removed. The overall accuracy decreases to " + }, + { + "bbox": [ + 130, + 156, + 480, + 323 + ], + "type": "inline_equation", + "content": "94.3\\%" + }, + { + "bbox": [ + 130, + 156, + 480, + 323 + ], + "type": "text", + "content": " but it is still higher than most existing methods. In model E, we test the performance when all SA modules are removed. The accuracy decreases to " + }, + { + "bbox": [ + 130, + 156, + 480, + 323 + ], + "type": "inline_equation", + "content": "92.8\\%" + }, + { + "bbox": [ + 130, + 156, + 480, + 323 + ], + "type": "text", + "content": " which is still better than the state-of-the-art rotation invariant method [42] thanks to the well designed RISPs which capture sufficient surface structures." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 323, + 263, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 263, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 263, + 477 + ], + "type": "text", + "content": "We also visualize the global feature vector before classification as shown in Fig. 5. We normalize the feature values to [0,1] and plot the histograms. We can see that by turning on the SA modules, the feature values become more evenly distributed and have less zeros because SA mechanism can reweight the features making the features more effective." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 271, + 334, + 479, + 430 + ], + "blocks": [ + { + "bbox": [ + 271, + 334, + 479, + 430 + ], + "lines": [ + { + "bbox": [ + 271, + 334, + 479, + 430 + ], + "spans": [ + { + "bbox": [ + 271, + 334, + 479, + 430 + ], + "type": "image", + "image_path": "563e38b74af65103e4fbea3c41cfa12cb6705795b3829a6e5ec8dc24cfd593b3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 270, + 438, + 480, + 461 + ], + "lines": [ + { + "bbox": [ + 270, + 438, + 480, + 461 + ], + "spans": [ + { + "bbox": [ + 270, + 438, + 480, + 461 + ], + "type": "text", + "content": "Fig. 5: Histogram comparison for normalized feature values without and with self-attention layers." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 482, + 253, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 482, + 253, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 482, + 253, + 613 + ], + "type": "text", + "content": "Network Efficiency. In the experiments, we acknowledge that our method employs a higher number of parameters. Thus, it is imperative to conduct a comprehensive analysis of network efficiency during both the training and testing stages. We list the trainable parameters, FLOPs" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 265, + 519, + 476, + 610 + ], + "blocks": [ + { + "bbox": [ + 259, + 484, + 481, + 518 + ], + "lines": [ + { + "bbox": [ + 259, + 484, + 481, + 518 + ], + "spans": [ + { + "bbox": [ + 259, + 484, + 481, + 518 + ], + "type": "text", + "content": "Table 7: Trainable parameters, FLOPs and timing comparisons on ModelNet40. All are tested on the same platform. Time is measured per batch (batch size=16)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 265, + 519, + 476, + 610 + ], + "lines": [ + { + "bbox": [ + 265, + 519, + 476, + 610 + ], + "spans": [ + { + "bbox": [ + 265, + 519, + 476, + 610 + ], + "type": "table", + "html": "
MethodsParamsFLOPsTime (Train/Infer)
PointNet++ [25]1.41M0.86G0.145s / 0.129s
RICnv [41]0.70M0.92G0.121s / 0.092s
RI-GCN [15]4.19M1.24G0.148s / 0.112s
RICnv++ [42]0.42M0.72G0.074s / 0.029s
Ours13.96M1.12G0.114s / 0.049s
Ours w/o TR1.35M1.11G0.105s / 0.047s
", + "image_path": "9532ca7fd26e344857481956a1d8c87e3d83dfdc4b503057a13ddcfbbf6811a4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 614, + 479, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 614, + 479, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 614, + 479, + 639 + ], + "type": "text", + "content": "and running time comparisons in Tab. 7. We compare the performance with four recent works: PointNet++ [25], RICnv [41], RI-GCN [15], RICnv++ [42]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "From the results, we can see that the number of parameters mainly comes from the transformer architectures used in RISurConv. When we turn off all" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "RISurConv" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "the transformer component, our method only takes up 1.35M parameters. More parameters do not mean longer inference time. It depends on lots of things including depth, parameter, number of operations etc. Our approach runs faster than most existing methods. It takes more time for training because more epochs are needed for self attention layers to converge. This is a common phenomenon of transformer based methods. The accuracy is also not due to the complexity model. As is shown in the ablation studies, by removing transformer encoder, it still achieves state-of-the-art performance due to the effectiveness of RISurConv." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 213, + 340, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 340, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 340, + 343 + ], + "type": "text", + "content": "Number of Local Surfaces. Local triangle surfaces used in Sec. 3 are the base for rotation invariant surface property (RISP) construction. Thus, it is important to analyse the effects of different number of surfaces. We conduct performance tests by setting the number of local surfaces to 1, 2, 3, and 4, resulting in overall accuracies of 94.7, 96.0, 91.9, and 89.8 respectively. It worth noting that for more surfaces, we can define more RISPs. From the results, we see the accuracy drops a bit when single" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 350, + 222, + 479, + 310 + ], + "blocks": [ + { + "bbox": [ + 350, + 222, + 479, + 310 + ], + "lines": [ + { + "bbox": [ + 350, + 222, + 479, + 310 + ], + "spans": [ + { + "bbox": [ + 350, + 222, + 479, + 310 + ], + "type": "image", + "image_path": "2e15e004c3454175b9ee2af6ae6f949fdf51f2d4757028bdacf6ddf6022fbc67.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 316, + 482, + 339 + ], + "lines": [ + { + "bbox": [ + 345, + 316, + 482, + 339 + ], + "spans": [ + { + "bbox": [ + 345, + 316, + 482, + 339 + ], + "type": "text", + "content": "Fig. 6: Distorted surface caused by 3 triangles." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 343, + 482, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 463 + ], + "type": "text", + "content": "surface is constructed. This is because smaller area cannot capture sufficient surface structures and less effective features can lower the performance. When we increase the number of surfaces to 3 and 4, the performance decreases. This shows that more surfaces can result in inferior performance because we construct triangle surfaces just to approximate the true surface. More surfaces may mess up the surface structure and affect the performance. Take Fig. 6 for example, suppose we are constructing the local surfaces " + }, + { + "bbox": [ + 130, + 343, + 482, + 463 + ], + "type": "inline_equation", + "content": "(\\triangle px_{i}x_{i + 1}\\triangle px_{i}x_{i - 1})" + }, + { + "bbox": [ + 130, + 343, + 482, + 463 + ], + "type": "text", + "content": " with regard to the neighbor point " + }, + { + "bbox": [ + 130, + 343, + 482, + 463 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 343, + 482, + 463 + ], + "type": "text", + "content": ", adding one more triangle " + }, + { + "bbox": [ + 130, + 343, + 482, + 463 + ], + "type": "inline_equation", + "content": "\\triangle px_{i}x_{i + 2}" + }, + { + "bbox": [ + 130, + 343, + 482, + 463 + ], + "type": "text", + "content": " will result in distorted/non-manifold surfaces. Thus, we set number of surfaces as 2 throughout the paper." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 483, + 217, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 217, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 217, + 495 + ], + "type": "text", + "content": "7 Discussion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 509, + 481, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 481, + 557 + ], + "type": "text", + "content": "**Limitation.** Though effective, RISurConv may require longer training time (e.g. more epochs) to converge due to large number of training parameters (Tab. 1). This is a common issue of transformer. But this does not affect the inference speed as shown in the ablation study (Tab. 7)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Conclusion. We have presented a new framework for achieving rotation invariance on 3D point cloud. We construct local triangle surfaces to better represent the local surface structures based on which we design highly expressive rotation invariant surface properties. We integrate the properties into an architecture named RISurConv to extract refined rotation invariant features. Based on RISurConv, we finally build up effective rotation invariant neural networks for 3D point cloud classification and segmentation with supreme performance achieved not only closing the performance gap with non-rotation invariant approaches but also surpassing the state-of-the-art methods by a large margin." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Z. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 190 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 190 + ], + "type": "text", + "content": "This research is supported by the Singapore Ministry of Education (MOE) Academic Research Fund (AcRF) Tier 1 grant (MSS23C010), Ningbo 2025 Science & Technology Innovation Major Project (No. 2022Z072, No.2023Z044), and Key Research & Development Plan of Zhejiang Province (2024C01017)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 209, + 198, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 209, + 198, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 209, + 198, + 220 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 234, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 234, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 234, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 234, + 480, + 267 + ], + "type": "text", + "content": "1. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3d semantic parsing of large-scale indoor spaces. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1534-1543 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 268, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 268, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 268, + 480, + 300 + ], + "type": "text", + "content": "2. Chang, A.X., Funkhouser, T.A., Guibas, L.J., Hanrahan, P., Huang, Q.X., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., Xiao, J., Yi, L., Yu, F.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 301, + 480, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 301, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 138, + 301, + 480, + 334 + ], + "type": "text", + "content": "3. Chen, C., Li, G., Xu, R., Chen, T., Wang, M., Lin, L.: Clusternet: Deep hierarchical cluster network with rigorously rotation-invariant representation for point cloud analysis. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4994-5002 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 335, + 480, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 335, + 480, + 356 + ], + "spans": [ + { + "bbox": [ + 138, + 335, + 480, + 356 + ], + "type": "text", + "content": "4. Deng, H., Birdal, T., Ilic, S.: Ppf-foldnet: Unsupervised learning of rotation invariant 3d local descriptors. In: Eur. Conf. Comput. Vis. pp. 602-618 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 357, + 480, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 480, + 378 + ], + "type": "text", + "content": "5. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: NAACL-HLT (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 379, + 480, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 379, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 379, + 480, + 422 + ], + "type": "text", + "content": "6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Int. Conf. Learn. Represent. (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 422, + 480, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 422, + 480, + 455 + ], + "spans": [ + { + "bbox": [ + 138, + 422, + 480, + 455 + ], + "type": "text", + "content": "7. Esteves, C., Allen-Blanchette, C., Makadia, A., Daniilidis, K.: Learning so (3) equivariant representations with spherical cnns. In: Eur. Conf. Comput. Vis. pp. 52-68 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 456, + 480, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 480, + 477 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 480, + 477 + ], + "type": "text", + "content": "8. Esteves, C., Xu, Y., Allen-Blanchette, C., Daniilidis, K.: Equivariant multi-view networks. In: Int. Conf. Comput. Vis. pp. 1568-1577 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 478, + 480, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 478, + 480, + 499 + ], + "spans": [ + { + "bbox": [ + 138, + 478, + 480, + 499 + ], + "type": "text", + "content": "9. Guo, M.H., Cai, J.X., Liu, Z.N., Mu, T.J., Martin, R.R., Hu, S.M.: Pct: Point cloud transformer. Computational Visual Media 7, 187-199 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 500, + 480, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 500, + 480, + 532 + ], + "spans": [ + { + "bbox": [ + 138, + 500, + 480, + 532 + ], + "type": "text", + "content": "10. Guo, Y., Wang, H., Hu, Q., Liu, H., Liu, L., Bennamoun, M.: Deep learning for 3d point clouds: A survey. IEEE Trans. Pattern Anal. Mach. Intell. 43(12), 4338-4364 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 533, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 480, + 555 + ], + "type": "text", + "content": "1. Han, X.F., Jin, Y.F., Cheng, H.X., Xiao, G.Q.: Dual transformer for point cloud analysis. IEEE Trans. Multimedia (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 555, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 587 + ], + "type": "text", + "content": "2. Hua, B.S., Pham, Q.H., Nguyen, D.T., Tran, M.K., Yu, L.F., Yeung, S.K.: Scenenn: A scene meshes dataset with annotations. In: International Conference on 3D Vision. pp. 92-101 (2016)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 588, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 588, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 588, + 480, + 610 + ], + "type": "text", + "content": "3. Huang, Q., Wang, W., Neumann, U.: Recurrent slice networks for 3d segmentation of point clouds. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2626-2635 (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "text", + "content": "4. Kim, J., Jung, W., Kim, H., Lee, J.: Cynnn: a rotation invariant cnn using polar mapping and cylindrical convolution layers. arXiv preprint arXiv:2007.10588 (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "5. Kim, S., Park, J., Han, B.: Rotation-invariant local-to-global representation learning for 3d point cloud. Adv. Neural Inform. Process. Syst. 33, 8174-8185 (2020)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "RISurConv" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "16. Li, X., Li, R., Chen, G., Fu, C.W., Cohen-Or, D., Heng, P.A.: A rotation-invariant framework for deep point cloud analysis. IEEE Trans. Vis. Comput. Graph. (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "type": "text", + "content": "17. Li, Y., Bu, R., Sun, M., Wu, W., Di, X., Chen, B.: Pointcnn: Convolution on x-transformed points. In: Adv. Neural Inform. Process. Syst. pp. 820-830 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 160, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 480, + 191 + ], + "type": "text", + "content": "18. Liu, X., Han, Z., Liu, Y.S., Zwicker, M.: Point2sequence: Learning the shape representation of 3d point clouds with an attention-based sequence to sequence network. In: AAAI. vol. 33, pp. 8778-8785 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 192, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 192, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 192, + 480, + 213 + ], + "type": "text", + "content": "19. Liu, X., Han, Z., Liu, Y.S., Zwicker, M.: Fine-grained 3d shape classification with hierarchical part-view attention. IEEE Trans. Image Process. 30, 1744-1758 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 214, + 480, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 480, + 245 + ], + "type": "text", + "content": "20. Liu, Y., Fan, B., Xiang, S., Pan, C.: Relation-shape convolutional neural network for point cloud analysis. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 8895-8904 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 246, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 480, + 277 + ], + "type": "text", + "content": "21. Maturana, D., Scherer, S.: Voxnet: A 3d convolutional neural network for real-time object recognition. In: 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 922-928 (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 278, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 480, + 310 + ], + "type": "text", + "content": "22. Poulenard, A., Rakotosaona, M.J., Ponty, Y., Ovsjanikov, M.: Effective rotation-invariant point cnn with spherical harmonics kernels. International Conference on 3D Vision (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 311, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 480, + 342 + ], + "type": "text", + "content": "23. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 652-660 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 343, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 480, + 374 + ], + "type": "text", + "content": "24. Qi, C.R., Su, H., Nießner, M., Dai, A., Yan, M., Guibas, L.J.: Volumetric and multi-view cnns for object classification on 3d data. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5648-5656 (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "type": "text", + "content": "25. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: Adv. Neural Inform. Process. Syst. pp. 5105-5114 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 407, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 439 + ], + "type": "text", + "content": "26. Qian, G., Li, Y., Peng, H., Mai, J., Hammoud, H., Elhoseiny, M., Ghanem, B.: Pointnext: Revisiting pointnet++ with improved training and scaling strategies. Adv. Neural Inform. Process. Syst. 35, 23192-23204 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 440, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 480, + 460 + ], + "type": "text", + "content": "27. Rao, Y., Lu, J., Zhou, J.: Spherical fractal convolutional neural networks for point cloud recognition. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 461, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 480, + 493 + ], + "type": "text", + "content": "28. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: International Conference on Medical image computing and computer-assisted intervention. pp. 234-241 (2015)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 494, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 480, + 525 + ], + "type": "text", + "content": "29. Shen, Y., Feng, C., Yang, Y., Tian, D.: Mining point cloud local structures by kernel correlation and graph pooling. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4548-4557 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 525, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 546 + ], + "type": "text", + "content": "30. Thomas, H.: Rotation-invariant point convolution with multiple equivariant alignments. In: International Conference on 3D Vision (3DV). pp. 504-513 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 547, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 480, + 578 + ], + "type": "text", + "content": "31. Uy, M.A., Pham, Q.H., Hua, B.S., Nguyen, D.T., Yeung, S.K.: Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In: Int. Conf. Comput. Vis. (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 579, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 480, + 610 + ], + "type": "text", + "content": "32. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Adv. Neural Inform. Process. Syst. 30 (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "text", + "content": "33. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph cnn for learning on point clouds. ACM Trans. Graph. (2019)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "34. Wu, X., Lao, Y., Jiang, L., Liu, X., Zhao, H.: Point transformer v2: Grouped vector attention and partition-based pooling. Adv. Neural Inform. Process. Syst. 35, 33330-33342 (2022)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "Z. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 424 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "35. Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., Xiao, J.: 3d shapenets: A deep representation for volumetric shapes. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1912-1920 (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "36. Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., Xiao, J.: 3d shapenets: A deep representation for volumetric shapes. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1912-1920 (2015)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "type": "text", + "content": "37. Xie, S., Liu, S., Chen, Z., Tu, Z.: Attentional shapecontextnet for point cloud recognition. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4606-4615 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "text", + "content": "38. Xu, Y., Fan, T., Xu, M., Zeng, L., Qiao, Y.: SpiderCNN: Deep learning on point sets with parameterized convolutional filters. In: Eur. Conf. Comput. Vis. pp. 87-102 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "type": "text", + "content": "39. Yu, X., Tang, L., Rao, Y., Huang, T., Zhou, J., Lu, J.: Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 19313-19322 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 270, + 482, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 303 + ], + "type": "text", + "content": "40. Zhang, Z., Hua, B.S., Chen, W., Tian, Y., Yeung, S.K.: Global context aware convolutions for 3d point cloud understanding. In: International Conference on 3D Vision (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 303, + 482, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 336 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 336 + ], + "type": "text", + "content": "41. Zhang, Z., Hua, B.S., Rosen, D.W., Yeung, S.K.: Rotation invariant convolutions for 3d point clouds deep learning. In: International Conference on 3D Vision. pp. 204-213 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 336, + 482, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 369 + ], + "type": "text", + "content": "42. Zhang, Z., Hua, B.S., Yeung, S.K.: Riconv++: Effective rotation invariant convolutions for 3d point clouds deep learning. Int. J. Comput. Vis. 130(5), 1228-1243 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 369, + 482, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 391 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 391 + ], + "type": "text", + "content": "43. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: Int. Conf. Comput. Vis. pp. 16259-16268 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 391, + 482, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 424 + ], + "type": "text", + "content": "44. Zhao, Y., Birdal, T., Lenssen, J.E., Menegatti, E., Guibas, L., Tombari, F.: Quaternion equivariant capsule networks for 3d point clouds. In: Eur. Conf. Comput. Vis. pp. 1-19. Springer (2020)" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "RISurConv" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/RPBG_ Towards Robust Neural Point-based Graphics in the Wild/ced6023f-74f5-424c-b6a1-715ff57d9512_content_list.json b/2024/RPBG_ Towards Robust Neural Point-based Graphics in the Wild/ced6023f-74f5-424c-b6a1-715ff57d9512_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7aa8a0514e9cef2008cbd1ad80ea217583e4b64c --- /dev/null +++ b/2024/RPBG_ Towards Robust Neural Point-based Graphics in the Wild/ced6023f-74f5-424c-b6a1-715ff57d9512_content_list.json @@ -0,0 +1,1818 @@ +[ + { + "type": "text", + "text": "RPBG: Towards Robust Neural Point-based Graphics in the Wild", + "text_level": 1, + "bbox": [ + 230, + 140, + 774, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qingtian Zhu $^{1\\star}$ , Zizhuang Wei $^{2,3}$ , Zhongtian Zheng $^{3}$ , Yifan Zhan $^{1}$ , Zhuyu Yao $^{4}$ , Jiawang Zhang $^{4}$ , Kejian Wu $^{4}$ , and Yinqiang Zheng $^{1\\boxtimes}$", + "bbox": [ + 235, + 210, + 769, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 The University of Tokyo", + "bbox": [ + 419, + 253, + 583, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Huawei Technologies", + "bbox": [ + 428, + 268, + 573, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Peking University", + "bbox": [ + 439, + 282, + 562, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 XREAL", + "bbox": [ + 468, + 296, + 535, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Point-based representations have recently gained popularity in novel view synthesis, for their unique advantages, e.g., intuitive geometric representation, simple manipulation, and faster convergence. However, based on our observation, these point-based neural re-rendering methods are only expected to perform well under ideal conditions and suffer from noisy, patchy points and unbounded scenes, which are challenging to handle but defacto common in real applications. To this end, we revisit one such influential method, known as Neural Point-based Graphics (NPBG), as our baseline, and propose Robust Point-based Graphics (RPBG). We in-depth analyze the factors that prevent NPBG from achieving satisfactory renderings on generic datasets, and accordingly reform the pipeline to make it more robust to varying datasets in-the-wild. Inspired by the practices in image restoration, we greatly enhance the neural renderer to enable the attention-based correction of point visibility and the inpainting of incomplete rasterization, with only acceptable overheads. We also seek for a simple and lightweight alternative for environment modeling and an iterative method to alleviate the problem of poor geometry. By thorough evaluation on a wide range of datasets with different shooting conditions and camera trajectories, RPBG stably outperforms the baseline by a large margin, and exhibits its great robustness over state-of-the-art NeRF-based variants. Code available at https://github.com/QT-Zhu/RPBG.", + "bbox": [ + 259, + 347, + 743, + 611 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Point-based graphics $\\cdot$ Novel view synthesis $\\cdot$ Neural rendering", + "bbox": [ + 261, + 625, + 712, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 662, + 357, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Novel view synthesis (NVS) aims to synthesize novel views under given camera poses by aid of a series of already posed images, which is a fundamental task in both computer vision and computer graphics and has been studied for decades.", + "bbox": [ + 212, + 695, + 787, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Among the feasible solutions, NeRF (Neural Radiance Field) [26] approximates an implicit scene representation, i.e., a radiance field (RF), encoded by a neural network mapping 3D coordinates and view directions to colors and densities. When rendering, the fitted RF is queried multiple times along a ray to volume-render the corresponding pixel for a novel view. Subsequent variants of NeRF follow a case-by-case design and", + "bbox": [ + 212, + 741, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Work done during Q. Zhu's internship at XREAL.", + "bbox": [ + 220, + 825, + 532, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b7ede94768135d06e5c41278c4cff3fbbf8927a1bd531dbd9bf20921066c4f63.jpg", + "image_caption": [ + "Fig. 1: Left: RPBG manages to achieve all-round good re-renderings (PSNR plotted) across generic datasets over the baseline [1] as well as state-of-the-art RF-based methods [42,44]. Right: We demonstrate the point clouds (with camera trajectories visualized) and corresponding re-rendered novel views of the representative scenes, revealing the great robustness and scalability of RPBG. Zoom in for best view." + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 385, + 273 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ec80b35a2520ba64f99530c13415086a3af529c84e449e0ffd1110e96a566457.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 147, + 781, + 275 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "adopt different parameterization techniques specifically for different scene types, e.g., object-centric [3, 54], free-trajectory [44], and large-scale scenes [38, 42]. Therefore, putting forward one robust method that can work stably across varying scene types with a unified parameterization technique is considered as challenging.", + "bbox": [ + 212, + 393, + 787, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, point-based alternatives [1, 18, 30, 33] have gained substantial attention for their unique advantages over implicit representations, e.g., ease of manipulation and faster training. The practice of employing point-based parameterization for rendering can be traced all the way back to [16, 20], and is intuitive especially under the context of NVs. Analogue to NeRF and its variants (collectively referred to as RF-based methods), whose optimization and convergence heavily rely on the inherent co-visibility among images, the 3D points triangulated from a series of 2D images according to epipolar geometry have already contain all the verified co-visibility information. In this way, by enforcing the triangulated points as strong prior knowledge for NVs, the expected freedom of optimization is greatly constrained, leading to faster training as well as better robustness. Recent attempts have achieved promising results in terms of fast optimization and rendering [18, 30] and accurate geometric representation [56]. Concretely, we focus on the specific category of methods that manage to integrate CNN-based neural renderers to yield re-renderings with featuremetric neural buffers [22, 30], for their concise pipeline and the potential capability of obtaining visually pleasing re-renderings.", + "bbox": [ + 212, + 457, + 787, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, by evaluating the representative baseline, NPBG (Neural Point-based Graphics) [1], with generic datasets, we realize that the original design only intends for ideal conditions, e.g., synthetic data [26] and well-captured human heads [31]. When synthesizing novel views under more general conditions, the performance of NPBG degrades severely. In this paper, we strive to boost the robustness of NPBG-like point-based neural re-rendering pipelines and reveal the true potential to achieve state-of-the-art performance across varying datasets in-the-wild, by analyzing the reasons for performance degradation and seeking for remedies.", + "bbox": [ + 212, + 686, + 787, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generally speaking, the major difficulties that the vanilla NPBG is faced up with include handling the background, handling patchy point clouds, and identifying correct", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Q. Zhu et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "point visibility. We reform the CNN-based neural renderer, with an inspiration from image restoration algorithms [6,8,52] that are able to identify downgraded patterns and restore the corresponding high-quality images. To make sure the neural renderer can capture sufficient and valid context information with extremely sparse rasterizations, we particularly leverages a Downgrade-aware Convolution (DAC) module to determine the correct point visibility with regard to a given camera pose and performs a pseudo point-wise back-face culling operation with visual self-attention. The background is also modeled in a lightweight manner. Instead of incorporating a massive environment map [33], a simple default trainable feature vector can reach similar quantitative results when working with the stronger neural renderer. We also discover the pseudo density calculated from the neural textures can roughly verify the existence of a given 3D position, which can be used to augment the poorly triangulated point clouds. In addition, we also simplify the phased training paradigm in [1,30] and optimize the parameters of both the neural textures and the renderer end-to-end collaboratively. We term our version of point-based re-rendering as RPBG (Robust Point-based Graphics), to emphasize its great robustness across different generic datasets.", + "bbox": [ + 217, + 145, + 787, + 387 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For thorough qualitative and quantitative evaluation, we collect 4 typical challenging scene types as the benchmark for robust NVs, i.e., $360^{\\circ}$ unbounded scenes (with free trajectories), inside-out scenes, large-scale scenes (at the scale of a block or campus), and sparse-view scenes. As a result, RPBG exhibits great robustness with perceptually satisfactory synthesis across the aforementioned typical scenes types as showcased in Fig. 1, where the high-quality re-renderings are obtained with an exactly identical parameterization strategy without any manual configuration, relieving the practitioners from the exhaustive per-scene search of hyper-parameters. Its stable superiority over state-of-the-art NVS methods, we believe, is of great significance especially for real applications.", + "bbox": [ + 217, + 388, + 787, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main contributions are three-fold as follows:", + "bbox": [ + 240, + 540, + 563, + 553 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We put forward RPBG, as a more robust and practical alternative for re-rendering high-quality images from triangulated 3D points.", + "- According to our in-depth analysis, we enhance the neural re-rendering pipeline regarding the neural renderer, environment modeling, point cloud augmentation, and the training paradigm.", + "- RPBG manages to greatly boost the performance of neural point re-renderer by a large margin, and exhibits stably greater robustness and generalizability over RF-based methods with even better perceptual rendering quality." + ], + "bbox": [ + 225, + 555, + 782, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 217, + 700, + 366, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Radiance Fields for NVs Along with the proposed NeRF [26], the scene representation of RF is becoming popular for its ease of optimization by the differentiable volume rendering. In a RF, each position is assigned with an anisotropic color and a density, and to render a pixel of a novel view, one needs to sample the trained field and conduct ray marching. The original NeRF [26], together with its variants [2,3,39,54], employs an MLP to represent the functional mapping from coordinates and view directions to the radiance values. Several attempts have been made to encode RFs with different", + "bbox": [ + 217, + 734, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "RPGB: Towards Robust Neural Point-based Graphics in the Wild", + "bbox": [ + 344, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "data structures for faster training and inference, e.g., grid voxels [12, 36], decomposed tensors [5], hash grids [27].", + "bbox": [ + 217, + 146, + 785, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In practice, the parameterization strategies have a great impact on the rendering quality when reconstructing different types of scenes, e.g., forward-facing ones, $360^{\\circ}$ unbounded ones, and large-scale ones. For unbounded scenes, $\\mathrm{NeRF}++$ [54] applies separate networks and different parameterization to model near and distant objects; mip-NeRF 360 [3] designs a smooth contraction operator to parameterize the whole unbounded scene into a ball; Mega-NeRF [42] and Block-NeRF [38] partition 3D scenes explicitly and use different networks to represent each scene partition; $\\mathrm{F}^2$ -NeRF [44] proposes to use perspective warping to handle sequential data with arbitrary trajectories. The varying case-by-case parameterization strategies of RF-based NVS greatly constraint the generalizability of methods.", + "bbox": [ + 217, + 178, + 787, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Point-based Graphics A point cloud is a collection of 3D coordinates that is usually used as a topology-agnostic coarse shape representation of the 3D geometry and favored for its flexibility and sparsity for storage. The development of techniques of employing points as modeling primitives for rendering (referred as point-based graphics [15]) can be traced back to [16, 20], the best practice of which is to replace each point with an oriented circular disk (a surfel) and reply on splatting to blend overlapping surfels [28].", + "bbox": [ + 217, + 343, + 787, + 433 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast to conventional physically based rendering (PBR) techniques, neural rendering [40] learns to render high-quality images in a data-driven manner. Particularly, we focus on the methods that conduct neural re-rendering with point clouds. Bui et al. [4] propose to enhance the coarse point-based rendering by a GAN for image super-resolution. NRW [25] and InvSFM [29] attempt to re-render the reconstructed point cloud with respective auxiliary buffers, e.g., latent appearance vectors, semantic masks and SIFT descriptors [23]. NPBG [1,30] adopts a U-Net-like [32] CNN to render neural point textures as RGB images, and exhibits better flexibility over mesh-based proxies [41]. The practical applications in the context of autonomous driving, e.g., scene editing and stitching, and large-scale training, are further explored in READ [21]. Recently, ADOP [33] and Gaussian Splatting [18] have demonstrated notable accomplishments. However, the differentiable rendering/splatting scheme employed requires an extensive amount of memory to maintain the computational graph and gradients during training. This limitation hinders their application on large-scale scenes.", + "bbox": [ + 217, + 434, + 787, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Neural Point-based Graphics Revisited", + "text_level": 1, + "bbox": [ + 217, + 667, + 571, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we revisit the pipeline of NPBG [1], analyze the existing drawbacks when attempting to re-render generic scenes, and attempt to explain why NPBG finds it difficult to handle the background and patchy points, and identify correct point visibility. We hope these discussions be of sufficient insights to support our modifications.", + "bbox": [ + 217, + 700, + 785, + 761 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 217, + 784, + 349, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inputs As the common practice in NVS, the inputs are a series of images with respective camera parameters (both intrinsics and extrinsics), and specifically for point-based", + "bbox": [ + 217, + 809, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 116, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Q. Zhu et al.", + "bbox": [ + 271, + 114, + 346, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b352c23801bc9ce10c0ac8b1ae1517c38d8b08f403fe65e1e53863e8924256e6.jpg", + "image_caption": [ + "(a) Rasterized Points" + ], + "image_footnote": [], + "bbox": [ + 223, + 147, + 362, + 273 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/43a2af0e2b26b0e374be61b1f1495cb66efdfeb3a427a7716d3bcfdf9846764a.jpg", + "image_caption": [ + "(b) Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 364, + 147, + 500, + 273 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c86057279b26509d122c492de35d5788f1fa7e38e97ecdd7482bf98ddb9e2282.jpg", + "image_caption": [ + "(c) NPBG", + "Fig. 2: Typically challenging scenes in T&T dataset [19] for NPBG. Top: Auditorium, where the walls and ceilings are extremely sparse. Bottom: Museum, where the point sparsity makes the occlusion and visibility complicated." + ], + "image_footnote": [], + "bbox": [ + 501, + 147, + 638, + 273 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/689cc3a0243e40edf745f0ea06e9b1930cebdee309b5611b5468e5a33add5c3c.jpg", + "image_caption": [ + "(d) Ours" + ], + "image_footnote": [], + "bbox": [ + 640, + 148, + 777, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "methods [1, 18, 30], the very first step is supposed to be the triangulation of 3D points from 2D observations. The original NPBG-like pipelines [1, 30] mainly consider this as a pre-processing without much attention.", + "bbox": [ + 214, + 367, + 785, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Point Rasterization NPBG [1] and NPBG++ [30] apply a non-differentiable hard point $z$ -buffering operation as an approximated back-face culling when rasterizing points as 2D fragments, where the fragment is updated when and only when the newly projected point has a smaller $z$ -depth than the current one. After a traversal of all the points, the fragment keeps the record of the indices of rasterized points. Then tensor scattering is performed to index the neural texture at the corresponding positions. Compared to differentiable rendering/splatting alternatives [18, 33], we consider the most significant advantage is its memory efficiency and thus better scalability, since the computation graph required to keep is much smaller for NPBG. We also want to keep this scalability without interfering the elegant rasterization paradigm.", + "bbox": [ + 214, + 425, + 787, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Neural Renderer The renderer $\\mathcal{R}_{\\Theta}:\\mathbb{R}^{H\\times W\\times C}\\to \\mathbb{R}^{H\\times W\\times 3}$ of NPBG [1] follows a vanilla U-Net-like architecture [32]. Since the target point clouds for rendering are supposed to be well-constructed, the vanilla U-Net is expressive enough to complete the assigned task of mapping higher dimensional features from the neural texture T to RGB values. Empirically, we find the expected properties of $\\mathcal{R}_{\\Theta}$ have much in common with a network for low-level vision tasks, e.g., image restoration.", + "bbox": [ + 214, + 585, + 787, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Problems", + "text_level": 1, + "bbox": [ + 215, + 696, + 323, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Patchy Triangulation Consider the inputs, for a NVS system aiming to synthesize images for generic scenes, the triangulation of points is not trivial. The point triangulation step to perform is similar to the step of multi-view stereo (MVS) reconstruction [14, 48], the key of which is to identify co-visible pixels across images and lift the 2D pixels to a 3D points according to epipolar geometry. Such algorithms suffer from non-Lambertian surfaces and textureless regions, both of which are very common and lead to poor triangulated and patchy points. We attach two typical scenes that NPBG fails to yield good renderings in Fig. 2.", + "bbox": [ + 214, + 719, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "RPBG: Towards Robust Neural Point-based Graphics in the Wild", + "bbox": [ + 344, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Wrong Point Visibility As can be imagined, such point visibility can be erroneous due to the poor quality of triangulated points, and also due to the inherent sparsity of points as a 3D representation. In this way, the points belonging to back faces, which should be considered as occluded, are rasterized as the fragment for further processing. Manually setting a depth threshold for bound check could not be helpful either, since being far from the camera does not necessarily indicate they should be occluded — they could be parts of the environment. While NHR [47] proposes to take the depth buffer as an additional rendering condition, we find its effectiveness not significant for generic scenes other than well-masked human captures.", + "bbox": [ + 217, + 146, + 787, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Lack of Context for Re-rendering Similarly, when re-rendering an incomplete point cloud, where the neural buffers are usually with significant flaws, the whole receptive field of a kernel at a given position may only capture the downgraded regions, leading to failure to yield reasonable restoration, especially for a high target resolution. It is worth noting that, though the relationship between the number of network layers and rendering quality has been discussed in [30, 33], there is still huge room for the improvement of the renderer.", + "bbox": [ + 217, + 297, + 787, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Failure to Effectively Model the Environment Unbounded scenes are a typical challenging scene type for NVS. RF-based methods that are designed on purpose to handle such scenes [3, 54] typically employ different parameterization to encode the areas outside a certain bounding sphere. For point-based methods, ADOP [33] proposes to apply an environment map of $H \\times W \\times C$ to model the environment, which is equivalent to wrapping the triangulated points with a sphere with $H \\times W$ points, resulting in more than $5 \\times 10^{5}$ points as overheads, according to the default configuration.", + "bbox": [ + 217, + 417, + 787, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Summary To summarize, we attribute the observed problems to two main causes: poor geometry and weak, local renderer. The two causes are to some extent coupled for a better geometry will relieve the difficulty of the renderer and vice versa.", + "bbox": [ + 217, + 539, + 787, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Robust Point-based Graphics", + "text_level": 1, + "bbox": [ + 217, + 608, + 496, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We are strongly convinced that the pipeline of point-based neural re-rendering has the great potential to outperform popular RF-based solutions, for its unified point-as-parameterization fashion and the incorporation of neural networks to yield visual details. Therefore in this section, we introduce Robust Point-based Graphics (RPBG), as an enhanced version of NPBG, with a particular focus on robustness under generic scenes. We will elaborate the modifications made and shed light on the underlying insights and motivation. The overall pipeline of RPBG as well as the training paradigm is briefly illustrated in Fig. 3.", + "bbox": [ + 217, + 640, + 787, + 761 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Downgrade-aware Neural Renderer", + "text_level": 1, + "bbox": [ + 217, + 782, + 506, + 799 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As is attributed as one key problem in NPBG, the U-Net-based neural renderer is considered as too naive to handle the challenging situations in generic scenes. The", + "bbox": [ + 217, + 809, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 116, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Q. Zhu et al.", + "bbox": [ + 271, + 114, + 346, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6b8cd81a3ed5ee1a60464a343d2dc2eda618897d578f7552b643a44052342850.jpg", + "image_caption": [ + "Fig. 3: The overall pipeline of RPBG. Point Triangulation: We first triangulate a 3D proxy for re-rendering with posed images, with its geometry-bounded neural texture initialized. Point Rasterization: The points are raterized to the given camera in a non-differentiable manner. By indexing the texture with the fragment, we obtain the neural buffer. A learnable point-size neural texture $\\mathbf{T}_{\\mathrm{env}}$ is also optimized. Neural Rendering: The restoration from downgraded neural buffer to photo-realistic images is performed by a CNN. The network and the neural texture are optimized end-to-end by image-level losses. An offline point cloud augmentation strategy is introduced to alleviate the problem of patchy triangulation under challenging conditions." + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 782, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "modifications made are deeply inspired by low-level vision tasks, where networks can adaptively determine downgraded parts, e.g., the deblurring/deraining networks are able to identify the blurring/raining pixels from the whole image, and correct them accordingly. We expect the neural renderer $\\mathcal{R}_{\\Theta}$ can benefit from relevant restoration-targeting techniques, and become able to decode high-quality visual information from patchy buffers.", + "bbox": [ + 212, + 518, + 787, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After evaluating three state-of-the-art fundamental architectures for image restoration, i.e., multi-scale fusion [8], multi-stage [52], and U-Net [6], following the taxonomy in [6]. According to the reported experiments in Sec. 5.4, we opt the paradigm in [8] for it achieves the best balance between performance and time/memory efficiency. However, the convolution layers in such paradigm is still with a fixed receptive field, lacking in robustness against point sparsity. We would like to further enlarge the receptive field to a global scale and attempt to explicitly model the visual attention to weigh the observed points.", + "bbox": [ + 212, + 611, + 787, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Transformer-based architectures for image restoration, e.g., Restormer [51], brings unacceptable memory overheads, so we rely on the frequency-domain alternative, Fast Fourier Convolution (FFC) [7], which can theoretically capture global contexts, to determine the correct point visibility adaptively. FFC performs channel-wise real 2D FFT (Fast Fourier Transform) and inverse real 2D FFT on 2D tensors. Real FFT uses only half of the spectrum and by convolving the transformed frequency-domain tensors. In this way, a receptive field covering the entire image is considered.", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "RPBG: Towards Robust Neural Point-based Graphics in the Wild", + "bbox": [ + 344, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/424413adaf0946be3980caff4e0c2453f8cdc98668babe46c712c681ab7db2dc.jpg", + "image_caption": [ + "Fig. 4: The architecture of the downgrade-aware neural renderer in RPBG, with some conventional modules omitted. From the visualized attention map, DAC manages to adaptively handle the severely erroneous point visibility." + ], + "image_footnote": [], + "bbox": [ + 289, + 148, + 723, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inspired by [37], we apply FFT (as the global branch) in parallel to conventional convolution layers (as the local branch) and rely on the fused features of both branches to determine the downgraded regions for the gated convolution [50] to filter. Based on the common practice in image inpainting, we leverage gated convolutions at the early stage of the renderer, to help locate the downgrade by wrong point visibility. We name such customized gated convolution module as the Downgrade-aware Convolution (DAC) module, and as can be inferred from the point attention maps before and after DAC in Fig. 4, DAC manages to determine complicated point visibility with patchy triangulation.", + "bbox": [ + 212, + 539, + 789, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Point Triangulation", + "text_level": 1, + "bbox": [ + 214, + 680, + 395, + 695 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Though there are advances in MVS empowered by deep learning [43, 45, 46], recovering a complete, accurate, and dense point cloud for certain regions remains a challenging. For RPBG, we adopt an off-the-shelf MVS method [46] to perform point triangulation for its memory efficiency to allow large-batch inference.", + "bbox": [ + 212, + 704, + 787, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As for problem the poor triangulation, we partially leave it to the neural renderer $\\mathcal{R}_{\\Theta}$ , relying on a stronger renderer to recover the re-rendering from downgraded buffers. Inspired by the point cloud extraction method of RF-based implementations [39], where the estimated radiance density $\\sigma$ can be a rough indicator for surface, we propose a point cloud augmentation technique to densify the initial triangulation.", + "bbox": [ + 212, + 765, + 789, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Q. Zhu et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The point cloud augmentation follows a trial-and-error paradigm by first assuming the existence of newly sampled points and then verify them by estimated pseudo densities $\\sigma_{i} = \\sum |\\mathbf{T}_{i}|$ . It is empirically observed that the absolute activation of the point-wise neural texture can roughly represent the reliability of the 3D position — it is reasonable since an outlier will be less distinguishable by the neural renderer and thus less visual attention will be given.", + "bbox": [ + 212, + 146, + 782, + 234 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Note that this strategy is optional, and we only apply such strategy to scenes with extremely poor triangulation. Please refer to the Supplementary Material for more discussions regarding the self-pruning.", + "bbox": [ + 212, + 237, + 784, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Environment Modeling", + "text_level": 1, + "bbox": [ + 214, + 305, + 419, + 320 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As a result of the incorporation of a stronger neural renderer, we find that given a neural texture with $C = 8$ , the dense environment map suggested in [33] is redundant. Instead, we shrink the overheads for environment modeling from $H \\times W \\times C$ to $1 \\times C$ , and relying on the stronger neural renderer to decode the background.", + "bbox": [ + 212, + 330, + 782, + 390 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "When rasterizing, we employ a tunable feature vector $\\mathbf{T}_{\\mathrm{env}}$ aside the neural point texture $\\mathbf{T}$ , as the default value for vacant pixels in the fragment, which is also involved in the end-to-end training. By experiments in Sec. 5.4, we demonstrate that the lightweight modeling strategy reaches the quantitative performance equivalent to applying an environment map, yet with a negligible overhead.", + "bbox": [ + 212, + 391, + 784, + 465 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.4 Collaborative Optimization", + "text_level": 1, + "bbox": [ + 214, + 488, + 447, + 503 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Recall that the point rasterization procedure is completely parameter-free. The overall collaborative optimization scheme can be thus formulated as", + "bbox": [ + 212, + 513, + 782, + 544 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} ^ {*}, \\Theta^ {*} = \\underset {\\mathbf {T}, \\Theta} {\\arg \\min } \\sum_ {k} \\mathcal {L} (\\mathbf {I} _ {k}, \\hat {\\mathbf {I}} _ {k}) = \\underset {\\mathbf {T}, \\Theta} {\\arg \\min } \\sum_ {k} \\mathcal {L} (\\mathbf {I} _ {k}, \\mathbf {K}, \\mathbf {R} _ {k}, \\mathbf {t} _ {k} | \\mathbf {T}, \\mathcal {R} _ {\\Theta}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 254, + 556, + 787, + 587 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathbf{T}$ and $\\Theta$ stand for the tunable parameters, i.e., the neural point texture (alongside the globally shared $\\mathbf{T}_{\\mathrm{env}}$ ) and the parameters of the renderer $\\mathcal{R}_{\\Theta}$ for neural re-rendering, while $\\mathbf{I}_k$ is the target image whose calibration is $\\mathbf{K}$ and $[\\mathbf{R}_k|\\mathbf{t}_k]$ and $\\hat{\\mathbf{I}}_k$ is the re-rendering", + "bbox": [ + 212, + 598, + 784, + 643 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Note that, we also discard all the typical tossing sampling [21] and optimization steps [30, 33] of point-based neural re-rendering, and manage to tune all the parameters involved in a simple but effective collaborative end-to-end way. The neural texture is initialized with all zeros while the rendering CNN is trained from scratch for each scene", + "bbox": [ + 212, + 643, + 784, + 703 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Loss Function Since the rendering scheme in both NPBG and RPBG is convolutional, where images are rendered in patches, we are able to apply patch-aware losses to enforce the involvement of neighboring pixels to ensure patch-to-patch consistency, offering perceptually good renderings. To this end, in addition to the pixel-wise Huber norm $\\mathcal{L}_{\\mathrm{huber}}$ providing the basic supervision and numerical stability, we apply two patch-aware losses to the collaborative optimization of RPBG, namely the perceptual VGG loss [10, 17] $\\mathcal{L}_{\\mathrm{vgg}}$ , and the FFT loss [8, 13] $\\mathcal{L}_{\\mathrm{fft}}$ . VGG loss compares the rendered image and the target ground-truth image in a high-dimensional feature space by a pre-trained VGG-19", + "bbox": [ + 212, + 719, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "RPBG: Towards Robust Neural Point-based Graphics in the Wild", + "bbox": [ + 344, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "network [35]. It reveals the perceptual similarity between images that pixel-wise metrics fail to measure. FFT loss measures the image-to-image distance in the frequency domain by carrying out 2D FFT towards images. The frequency components are considered to be crucial in terms of the perceptual quality [13].", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The final loss function applied in RPBG is composed as", + "bbox": [ + 238, + 207, + 609, + 220 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {\\text {h u b e r}} \\mathcal {L} _ {\\text {h u b e r}} + \\lambda_ {\\text {v g g}} \\mathcal {L} _ {\\text {v g g}} + \\lambda_ {\\text {f f t}} \\mathcal {L} _ {\\text {f f t}}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 227, + 785, + 244 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 214, + 262, + 356, + 277 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Datasets", + "text_level": 1, + "bbox": [ + 214, + 289, + 316, + 301 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We conduct experiments covering a diverse range of scene types to quantitatively examine the generalizability and robustness of NVs methods, including the Free dataset [44] (7 free-trajectory scans with long camera trajectories), Tanks and Temples (T&T) dataset [19] (4 unbounded scans and 5 inside-out scans, without salient dynamic objects), Mill19 dataset [42] (2 large-scale aerial scans with over 1600 images in each scan), and GigaMVS dataset [53] (8 sparse-view outdoor scans). Note that for T&T, we use the raw undistorted images provided by the benchmark without any masking [22]. We follow the common split protocol, that 1 frame out of every 8 frames is evaluated, for the Free dataset, T&T dataset, and GigaMVS dataset. For Mill19 dataset, we apply the officially recommended split to align with the experiments in [42].", + "bbox": [ + 212, + 309, + 785, + 459 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We also include some other challenging datasets for a thorough evaluation, e.g., ScanNet++ dataset [49], a challenging indoor benchmark, OMMO dataset [24], a multi-modal aerial NVS dataset and ETH-MS dataset [11], a super-large-scale dataset (around $5k$ images). Please refer to the Supplementary Material for more quantitative and qualitative results.", + "bbox": [ + 212, + 460, + 785, + 534 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 553, + 419, + 566 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Data Preparation To obtain the per-scene point cloud representation for re-rendering, we follow the mapping procedure of COLMAP [34] for sparse triangulation and the MVS reconstruction with a trained network [46] for dense triangulation. Note that the reconstruction only takes a small proportion of time over the whole training phase and we conduct ablation experiments to show that RPBG is robust against different triangulation configurations. Please refer to Sec. 5.4 as well as the Supplementary Material for more information.", + "bbox": [ + 212, + 573, + 785, + 676 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training Settings We randomly crop images to square patches of $256 \\times 256$ , with a batch size of 8. The learning rate for the neural point textures is $10^{-1}$ , and $10^{-4}$ for the rendering networks, which will decay by a factor of 0.5 if 5 consequent epochs witness no drop in the loss function. A dimension of 8 is applied for any neural texture regardless of the scene scale or complexity. For the weights of loss functions, we globally set $\\lambda_{\\mathrm{huber}} = 10^3$ , $\\lambda_{\\mathrm{vgg}} = 1$ , and $\\lambda_{\\mathrm{fft}} = 1$ . The training is performed on one NVIDIA GeForce RTX 3090, with a GPU memory consumption of up to 23 GB. It takes around 8 to 30 GPU hours for training, depending on the data scale. We would like to emphasize that, RPBG does not require case-by-case scene parameterization or grid search of training hyper-parameters for all the scenes covered in the experiments.", + "bbox": [ + 212, + 689, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Q. Zhu et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/382384e456bc329ff256bb9ad09ac248cec17f078676bb0f4c5e63dc006910b2.jpg", + "image_caption": [ + "Fig. 5: Visualized comparisons over varying scenes. From top to bottom: sky and hydrant of the Free dataset [44], Courtyard and Train of T&T dataset [19], and DayaTemple and MemorialHall of GigaMVS dataset [53]. We include the results of RPBG (Ours), Gaussian Splatting [18], NPBG [1], and $\\mathrm{F}^2$ -NeRF [44] for comparison. Zoom in for best view." + ], + "image_footnote": [], + "bbox": [ + 223, + 147, + 782, + 441 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation Metrics We adopt the metrics of PSNR, SSIM and LPIPS (VGG) [55] for the evaluation between the synthesized and the target images. According to [55], PSNR does not faithfully measure image sharpness and so cannot properly account for the nuances of human visual perception.", + "bbox": [ + 212, + 542, + 787, + 604 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.3 Results", + "text_level": 1, + "bbox": [ + 214, + 628, + 308, + 643 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In Fig. 5, we demonstrate several typical groups of visual comparisons of NVS results by RPBG, two representative point-based methods, i.e., Gaussian Splitting [18] and NPBG [1], and a state-of-the-art NeRF variant for unbounded scenes, $\\mathrm{F}^2$ -NeRF [44]. We group the datasets by category and report the corresponding scene-averaged quantitative scores in Tab. 1. Note that RPBG achieves the best SSIM and LPIPS, which are more relevant to the high-frequency components of images, across all datasets.", + "bbox": [ + 212, + 656, + 787, + 748 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Free-trajectory/Unbounded Scenes Compared with NVS methods designed in particular for unbounded scenes, $\\mathrm{NeRF}++$ [54] and $\\mathrm{F^2}$ -NeRF [44], RPBG achieves the best SSIM and LPIPS and comparable PSNR. For visual effects showcased in Fig. 5, since RPBG involves a neighborhood for rendering, its results appear more visually harmonious, especially than Gaussian Splatting [18] and $\\mathrm{F^2}$ -NeRF [44]. It is necessary to mention", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "RPBG: Towards Robust Neural Point-based Graphics in the Wild", + "bbox": [ + 344, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/4677386befd0997cacb294bda6e7416a15d87cf5a53804d3b5cea4157dbcbb85.jpg", + "table_caption": [ + "Table 1: Quantitative evaluation of state-of-the-art NVS methods [1,5,18,26,30,42,44,54] and RPBG across diverse scenes grouped by category, including free trajectory/unbounded scenes, inside-out scenes, large-scale scenes, and sparse-view scenes. The figure following the dataset name stands for the number of scenes the dataset contains." + ], + "table_footnote": [], + "table_body": "
MethodFree-trajectory/UnboundedInside-out T&T-5Large-scale Mill19-2Sparse-view GigaMVS-8
Free Dataset-7T&T-4
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NeRF [26]17.750.4050.59716.840.3960.73119.820.6150.52620.340.5240.52918.290.6320.499
TensoRF [5]21.740.5490.60020.730.6430.56619.430.6340.57020.070.4970.59718.700.6530.506
NeRF++ [54]23.470.6030.49921.660.6580.52919.250.6100.58520.190.5200.53118.380.6320.495
F2-NeRF [44]26.320.7790.27623.660.7640.30320.120.7060.394N/AN/AN/A17.440.5400.470
Mega-NeRF [42]22.600.5700.56218.730.5780.47819.160.6170.45122.490.5500.51018.250.5810.394
NPBG [1]21.400.6390.34019.850.6980.37620.570.6960.37116.210.3570.64418.340.6200.405
NPBG++ [30]20.060.5920.44517.230.6530.47418.300.6840.41117.040.4000.64819.300.6630.443
Gaussian Splatting [18]25.230.7400.29023.510.7820.29323.460.7830.277N/AN/AN/A16.840.5300.391
Ours26.330.8320.17722.500.7820.27623.290.8040.24222.620.5960.36820.540.6860.317
", + "bbox": [ + 217, + 212, + 785, + 329 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "that, for RF-based methods [5, 26, 42, 44], we have manually adjusted the scene-specific hyper-parameters to achieve better results.", + "bbox": [ + 212, + 359, + 784, + 390 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Inside-out Scenes The inside-out indoor scenes are strictly bounded, but lack high-quality ray intersections required for optimization, which will lead to under-fitting of the RF (Courtroom in Fig. 5). Similar to free-trajectory/unbounded scenes, RPBG outperforms state-of-the-art methods in SSIM and LPIPS.", + "bbox": [ + 212, + 402, + 785, + 463 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Large-scale Scenes We mainly evaluate RPBG against Mega-NeRF [42] on the massive imagery of Mill19 dataset [42]. Our method outperforms Mega-NeRF [42] in every metric. It is worth noting that Mega-NeRF takes 240 GPU hours for training while RPBG only takes 29 GPU hours. Notably, our system with 256 GB RAM and RTX 3090 fails to afford the training of $\\mathrm{F}^2$ -NeRF [44] and Gaussian Splitting [18] (marked as N/A in the table). In contrast, RPBG proves to be viable for large-scale scenes under the same hardware constraint.", + "bbox": [ + 212, + 477, + 787, + 583 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Sparse-view Scenes Sparse-view inputs are considered to be extremely challenging for NVS methods. Since RPBG, NPBG [1], and $\\mathrm{NPBG + + }$ [30] incorporate a point-based 3D proxy for re-rendering, they showcase better robustness over RF-based methods. As revealed in Fig. 5, Gaussian Splatting [18] fails to regularize the strong approximation power of point-wise Gaussians at training views, and yields obvious needle-like artifacts at novel views.", + "bbox": [ + 212, + 597, + 787, + 688 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 709, + 362, + 724 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Environment Modeling We study different strategies of environment modeling for NVS in-the-wild, namely leaving the blank pixels with zeros, filling the blanks with a default learnable feature (as is done in RPBG), and", + "bbox": [ + 212, + 734, + 433, + 839 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4921670c624329e449c60f71af44f725a85dfa9c3d7fef56b2841a5d3a75c0cc.jpg", + "image_caption": [ + "(a) $w / o$ Env. Modeling" + ], + "image_footnote": [], + "bbox": [ + 447, + 729, + 612, + 800 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/0b807bd6eaabce4251f38cd77661aa716bc13bbd0f006ac17f09b808cfcd4cca.jpg", + "image_caption": [ + "(b) $w / \\mathrm{Env}$ Modeling", + "Fig.6: The impact of leveraging the default feature vector for environment modeling in an unbounded scene." + ], + "image_footnote": [], + "bbox": [ + 616, + 729, + 779, + 800 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Q. Zhu et al.", + "bbox": [ + 271, + 114, + 346, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/3203be77f927afba80d6c5b91e101a0bb9549f19b6efca04192eb29a7dc1c391.jpg", + "table_caption": [ + "Table 2: Ablation on network designs on T&T dataset [19]. The reported inference time and memory consumption is tested with a target resolution of ${1920} \\times {1080}$ ." + ], + "table_footnote": [], + "table_body": "
MethodTime(s)Mem.(GB)PSNR↑SSIM↑LPIPS↓
Baseline1.035.4920.010.6660.390
+Multi-scale fusion1.6212.6721.880.7010.337
+DAC1.6713.0122.920.7690.320
+FFT loss--22.940.7940.257
+Multi-stage1.7717.1421.750.7130.338
+U-Net1.5810.3220.850.6920.364
", + "bbox": [ + 220, + 212, + 496, + 286 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/6c30e13af5ac60ecb50f3d99b650ff9d936ff37698e79b82362bc107874c1ff5.jpg", + "table_caption": [ + "Table 3: Ablation on environment modeling on two unbounded scenes, namely Caterpillar of T&T dataset [19] and sky of Free dataset [44]. #Points $(\\times 10^{6})$ represent the equivalent overheads when applying different strategies." + ], + "table_footnote": [], + "table_body": "
Env.Caterpillarsky
#PointsPSNR↑SSIM↑LPIPS↓#PointsPSNR↑SSIM↑LPIPS↓
Zeros7.9519.890.4560.40317.4921.770.5980.315
Learnable7.9521.780.6870.28917.4924.810.8660.199
Sphere8.9521.680.6860.28318.4925.150.8690.179
", + "bbox": [ + 508, + 226, + 784, + 273 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a7d1c1353fe2b69baabb0ea37f7774c619a9c75768268a61c948a0bc1b7b0d09.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 301, + 356, + 359 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/760a399218dc190a30b2d8acffc7b95906cdb04f56139e0d7e4adcc6b1ca2ba3.jpg", + "image_caption": [ + "(c) $w / \\theta$ DAC", + "Fig.7: A challenging case in Train, which is a zoomed-in view with complex occlusion and thin objects (e.g., the handrails)." + ], + "image_footnote": [], + "bbox": [ + 222, + 359, + 356, + 428 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/926f160a17c7c29ef5b47d8149db76a5d7294691b6471c62b216249e03995a58.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 301, + 491, + 361 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/4883f9a90237c97e80fcffc58f0920de95e37f1384ea51c16f0d3e0c7f17ab38.jpg", + "image_caption": [ + "(d) $w / \\mathrm{DAC}$", + "Fig. 8: Re-rendering quality (PSNR) with points of different quality on Church. RPBG manages to maintain its robustness against different levels of point sparsity with a $14\\%$ drop in PSNR (NPBG [1]: $25\\%$ ; NPBG++ [30]: $14\\%$ )." + ], + "image_footnote": [], + "bbox": [ + 357, + 361, + 491, + 428 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/cb069b7936944b5719e3c3ad478cc6f7c774a84f3db6cdd8334321284b607fc6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 303, + 782, + 412 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "wrapping the points with a sphere of $10^{6}$ points (equivalent to an environment map in [33]). The quantitative results are in Sec. 5.4, where the strategy incorporated by RPBG boosts the performance by a large margin, with almost no additional overhead.", + "bbox": [ + 212, + 525, + 787, + 571 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Network Architecture and Loss We study the key components in the rendering network in Sec. 5.4. The baseline refers to NPBG [1] trained per-scene from scratch with $\\mathcal{L}_1$ and VGG loss. We compare three fundamental architectures for image restoration (with necessary modifications on the first several layers), i.e., multi-scale fusion [8], multistage [52], and U-Net [6] following the taxonomy in [6]. Compared to the baseline, all modern restoration-oriented architectures bring remarkable improvement to the overall performance, and [8] is opted for a better balance between performance and time/memory efficiency. The gain by DAC is also considerable while the FFT loss mainly improves the perceptual quality. By the case in Fig. 7, we show the necessity of the DAC module when re-rendering challenging cases without in particular handling the erroneous point visibility.", + "bbox": [ + 212, + 592, + 787, + 758 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Triangulation Configurations We also evaluate RPBG with different triangulation configurations. Indicated by the results in Fig. 8, RPBG witnesses a total drop of $3.29\\mathrm{dB}$ in re-rendering PSNR when switching the triangulation from ground truth (GT) to randomly initialized points (Random). Note that even with sparse points (SfM),", + "bbox": [ + 212, + 779, + 789, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "RPBG: Towards Robust Neural Point-based Graphics in the Wild", + "bbox": [ + 344, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 769, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "which are usually yielded as a by-product when computing camera parameters, RPBG (21.41 dB) still outperforms $\\mathrm{F}^2$ -NeRF (20.57 dB).", + "bbox": [ + 215, + 146, + 784, + 176 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Discussion", + "text_level": 1, + "bbox": [ + 215, + 199, + 336, + 215 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Effectiveness of RPBG We would like to shed some light on the effectiveness of RPBG. The first factor should be the point representation. The triangulated 3D points have already contained all the verified co-visibility that is required for NVS. Compared to RF-based methods, which often struggle at local minima, such triangulation-as-parameterization paradigm greatly lower the complexity of searching, and is unified across different scene types. Besides, the patch-wise rendering scheme mentioned in Sec. 4.4 also improves the perceptual quality of re-rendering while RF-based methods apply pixel/ray-wise volume rendering, where the correlation between pixels are not modeled explicitly. For more discussions on the insights, we strongly refer the readers to the Supplementary Material.", + "bbox": [ + 215, + 231, + 785, + 382 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations An obvious problem with point-based re-rendering (RPBG and NPBG [1]) is that compared to lightweight RF-based variants [5, 27], they take more space to store the CNN parameters and the neural texture. Although it makes the scaling-up easier, the number of parameters grows linearly with the point cloud scale. Besides, due to the non-local neural renderer we employ, each point is encoded with visual context of a larger range, hindering the editability of RPBG. We have also noticed that, the CNN-based rendering scheme can lead to unsatisfactory temporal consistency especially when the triangulation is patchy so that we mainly count on the inpainting ability of the neural render for yielding re-renderings. Such flicker issue, as well as its potential solutions, has been discussed in [9].", + "bbox": [ + 215, + 397, + 785, + 547 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 571, + 343, + 588 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we present RPBG as a robust and practical alternative to NPBG, a baseline of point-based NVs methods, performing neural re-rendering on triangulated points. We analyze the key problems in NPBG, when attempting to generalize to more generic scenes other than the well-captured scans, and reform the pipeline to reveal the real potential of point-based graphics. Respectively motivated and inspired by RF-based methods and low-level image restoration methods, we reform the pipeline according to our analysis. By extensive experiments on diverse datasets, RPBG achieves stably superior results over state-of-the-art RF-based and point-based NVs methods, especially on the metrics with more attention paid to the nuances of human visual perception, without case-by-case parameterization across all scenes, indicating its robustness and generalizability.", + "bbox": [ + 215, + 604, + 785, + 770 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Q. Zhu et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 215, + 143, + 374, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This research was supported in part by JSPS KAKENHI Grant Numbers 24K22318, 22H00529, 20H05951, and JST-Mirai Program JPMJMI23G1.", + "bbox": [ + 212, + 174, + 787, + 204 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 224, + 310, + 242 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aliev, K.A., Sevastopolsky, A., Kolos, M., Ulyanov, D., Lempitsky, V.: Neural point-based graphics. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 696-712. Springer (2020)", + "2. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855–5864 (2021)", + "3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022)", + "4. Bui, G., Le, T., Morago, B., Duan, Y.: Point-based rendering enhancement via deep learning. The Visual Computer 34, 829-841 (2018)", + "5. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: European Conference on Computer Vision. pp. 333-350. Springer (2022)", + "6. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: European Conference on Computer Vision. pp. 17-33. Springer (2022)", + "7. Chi, L., Jiang, B., Mu, Y.: Fast fourier convolution. Advances in Neural Information Processing Systems 33, 4479-4488 (2020)", + "8. Cho, S.J., Ji, S.W., Hong, J.P., Jung, S.W., Ko, S.J.: Rethinking coarse-to-fine approach in single image deblurring. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4641-4650 (2021)", + "9. Dai, P., Zhang, Y., Li, Z., Liu, S., Zeng, B.: Neural point cloud rendering via multi-plane projection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7830-7839 (2020)", + "0. Dosovitskiy, A., Brox, T.: Generating images with perceptual similarity metrics based on deep networks. Advances in neural information processing systems 29 (2016)", + "11. ETH Zurich Computer Vision Group and Microsoft Mixed Reality & AI Lab Zurich: The ETH-Microsoft Localization Dataset. https://github.com/cvg/visloc-iccv2021 (2021)", + "12. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022)", + "13. Fuoli, D., Van Gool, L., Timofte, R.: Fourier space losses for efficient perceptual image super-resolution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2360-2369 (2021)", + "14. Furukawa, Y., Ponce, J.: Accurate, dense, and robust multiview stereopsis. IEEE transactions on pattern analysis and machine intelligence 32(8), 1362-1376 (2009)", + "15. Gross, M., Pfister, H.: Point-based graphics. Elsevier (2011)", + "16. Grossman, J.P., Dally, W.J.: Point sample rendering. In: Rendering Techniques' 98: Proceedings of the Eurographics Workshop in Vienna, Austria, June 29—July 1, 1998 9. pp. 181-192. Springer (1998)", + "17. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and superresolution. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 694-711. Springer (2016)" + ], + "bbox": [ + 218, + 253, + 787, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "RPBG: Towards Robust Neural Point-based Graphics in the Wild", + "bbox": [ + 344, + 114, + 732, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 769, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG) 42(4), 1-14 (2023)", + "19. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG) 36(4), 1-13 (2017)", + "20. Levoy, M., Whitted, T.: The use of points as a display primitive (1985)", + "21. Li, Z., Li, L., Zhu, J.: Read: Large-scale neural scene rendering for autonomous driving. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1522-1529 (2023)", + "22. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020)", + "23. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. International journal of computer vision 60, 91-110 (2004)", + "24. Lu, C., Yin, F., Chen, X., Chen, T., Yu, G., Fan, J.: A large-scale outdoor multi-modal dataset and benchmark for novel view synthesis and implicit scene reconstruction. arXiv preprint arXiv:2301.06782 (2023)", + "25. Meshry, M., Goldman, D.B., Khamis, S., Hoppe, H., Pandey, R., Snavely, N., Martin-Brualla, R.: Neural rerendering in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6878-6887 (2019)", + "26. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: European Conference on Computer Vision. pp. 405-421 (2020)", + "27. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)", + "28. Pfister, H., Zwicker, M., Van Baar, J., Gross, M.: Surfels: Surface elements as rendering primitives. In: Proceedings of the 27th annual conference on Computer graphics and interactive techniques. pp. 335-342 (2000)", + "29. Pittaluga, F., Koppal, S.J., Kang, S.B., Sinha, S.N.: Revealing scenes by inverting structure from motion reconstructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 145-154 (2019)", + "30. Rakhimov, R., Ardelean, A.T., Lempitsky, V., Burnaev, E.: Npbg++: Accelerating neural point-based graphics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15969-15979 (2022)", + "31. Ramon, E., Triginer, G., Escur, J., Pumarola, A., Garcia, J., Giro-i Nieto, X., Moreno-Noguer, F.: H3d-net: Few-shot high-fidelity 3d head reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5620-5629 (2021)", + "32. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)", + "33. Ruckert, D., Franke, L., Stamminger, M.: Adop: Approximate differentiable one-pixel point rendering. ACM Transactions on Graphics (ToG) 41(4), 1-14 (2022)", + "34. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4104-4113 (2016)", + "35. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)", + "36. Sun, C., Sun, M., Chen, H.T.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5459-5469 (2022)", + "37. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., Rabinovich, A.: Going deeper with convolutions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1-9 (2015)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Q. Zhu et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "38. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8248–8258 (2022)", + "39. Tancik, M., Weber, E., Ng, E., Li, R., Yi, B., Wang, T., Kristoffersen, A., Austin, J., Salahi, K., Ahuja, A., et al.: Nerfstudio: A modular framework for neural radiance field development. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-12 (2023)", + "40. Tewari, A., Fried, O., Thies, J., Sitzmann, V., Lombardi, S., Sunkavalli, K., Martin-Brualla, R., Simon, T., Saragih, J., Nießner, M., et al.: State of the art on neural rendering. In: Computer Graphics Forum. vol. 39, pp. 701-727. Wiley Online Library (2020)", + "41. Thies, J., Zollhöfer, M., Nießner, M.: Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG) 38(4), 1-12 (2019)", + "42. Turki, H., Ramanan, D., Satyanarayanan, M.: Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12922–12931 (2022)", + "43. Wang, F., Galliani, S., Vogel, C., Speciale, P., Pollefeys, M.: Patchmatchnet: Learned multiview patchmatch stereo. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14194-14203 (2021)", + "44. Wang, P., Liu, Y., Chen, Z., Liu, L., Liu, Z., Komura, T., Theobalt, C., Wang, W.: F2-nerf: Fast neural radiance field training with free camera trajectories. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4150–4159 (2023)", + "45. Wang, X., Zhu, Z., Huang, G., Qin, F., Ye, Y., He, Y., Chi, X., Wang, X.: Mvster: Epipolar transformer for efficient multi-view stereo. In: European Conference on Computer Vision. pp. 573-591. Springer (2022)", + "46. Wei, Z., Zhu, Q., Min, C., Chen, Y., Wang, G.: Aa-rmvvsnet: Adaptive aggregation recurrent multi-view stereo network. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6187-6196 (2021)", + "47. Wu, M., Wang, Y., Hu, Q., Yu, J.: Multi-view neural human rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1682-1691 (2020)", + "48. Yao, Y., Luo, Z., Li, S., Fang, T., Quan, L.: Mvsnet: Depth inference for unstructured multiview stereo. In: Proceedings of the European conference on computer vision (ECCV). pp. 767-783 (2018)", + "49. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12-22 (2023)", + "50. Yu, J., Lin, Z., Yang, J., Shen, X., Lu, X., Huang, T.S.: Free-form image inpainting with gated convolution. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4471-4480 (2019)", + "51. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5728-5739 (2022)", + "52. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14821-14831 (2021)", + "53. Zhang, J., Zhang, J., Mao, S., Ji, M., Wang, G., Chen, Z., Zhang, T., Yuan, X., Dai, Q., Fang, L.: Gigamvs: a benchmark for ultra-large-scale gigapixel-level 3d reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(11), 7534-7550 (2021)", + "54. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492 (2020)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "RPGB: Towards Robust Neural Point-based Graphics in the Wild", + "bbox": [ + 344, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 769, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "55. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)", + "56. Zhang, Y., Peng, S., Moazeni, A., Li, K.: Papr: Proximity attention point rendering. Advances in Neural Information Processing Systems 36 (2024)" + ], + "bbox": [ + 215, + 146, + 787, + 217 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Q. Zhu et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/RPBG_ Towards Robust Neural Point-based Graphics in the Wild/ced6023f-74f5-424c-b6a1-715ff57d9512_model.json b/2024/RPBG_ Towards Robust Neural Point-based Graphics in the Wild/ced6023f-74f5-424c-b6a1-715ff57d9512_model.json new file mode 100644 index 0000000000000000000000000000000000000000..aaaf778ef352732f418d035f8b7b4c9ca26e5a67 --- /dev/null +++ b/2024/RPBG_ Towards Robust Neural Point-based Graphics in the Wild/ced6023f-74f5-424c-b6a1-715ff57d9512_model.json @@ -0,0 +1,2546 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.231, + 0.141, + 0.776, + 0.185 + ], + "angle": 0, + "content": "RPBG: Towards Robust Neural Point-based Graphics in the Wild" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.212, + 0.77, + 0.246 + ], + "angle": 0, + "content": "Qingtian Zhu\\(^{1\\star}\\), Zizhuang Wei\\(^{2,3}\\), Zhongtian Zheng\\(^{3}\\), Yifan Zhan\\(^{1}\\), Zhuyu Yao\\(^{4}\\), Jiawang Zhang\\(^{4}\\), Kejian Wu\\(^{4}\\), and Yinqiang Zheng\\(^{1\\boxtimes}\\)" + }, + { + "type": "text", + "bbox": [ + 0.42, + 0.255, + 0.584, + 0.269 + ], + "angle": 0, + "content": "1 The University of Tokyo" + }, + { + "type": "text", + "bbox": [ + 0.429, + 0.269, + 0.574, + 0.283 + ], + "angle": 0, + "content": "\\(^{2}\\) Huawei Technologies" + }, + { + "type": "text", + "bbox": [ + 0.44, + 0.283, + 0.563, + 0.298 + ], + "angle": 0, + "content": "3 Peking University" + }, + { + "type": "text", + "bbox": [ + 0.469, + 0.297, + 0.536, + 0.31 + ], + "angle": 0, + "content": "4 XREAL" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.348, + 0.744, + 0.612 + ], + "angle": 0, + "content": "Abstract. Point-based representations have recently gained popularity in novel view synthesis, for their unique advantages, e.g., intuitive geometric representation, simple manipulation, and faster convergence. However, based on our observation, these point-based neural re-rendering methods are only expected to perform well under ideal conditions and suffer from noisy, patchy points and unbounded scenes, which are challenging to handle but defacto common in real applications. To this end, we revisit one such influential method, known as Neural Point-based Graphics (NPBG), as our baseline, and propose Robust Point-based Graphics (RPBG). We in-depth analyze the factors that prevent NPBG from achieving satisfactory renderings on generic datasets, and accordingly reform the pipeline to make it more robust to varying datasets in-the-wild. Inspired by the practices in image restoration, we greatly enhance the neural renderer to enable the attention-based correction of point visibility and the inpainting of incomplete rasterization, with only acceptable overheads. We also seek for a simple and lightweight alternative for environment modeling and an iterative method to alleviate the problem of poor geometry. By thorough evaluation on a wide range of datasets with different shooting conditions and camera trajectories, RPBG stably outperforms the baseline by a large margin, and exhibits its great robustness over state-of-the-art NeRF-based variants. Code available at https://github.com/QT-Zhu/RPBG." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.625, + 0.714, + 0.64 + ], + "angle": 0, + "content": "Keywords: Point-based graphics \\(\\cdot\\) Novel view synthesis \\(\\cdot\\) Neural rendering" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.664, + 0.358, + 0.681 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.696, + 0.788, + 0.741 + ], + "angle": 0, + "content": "Novel view synthesis (NVS) aims to synthesize novel views under given camera poses by aid of a series of already posed images, which is a fundamental task in both computer vision and computer graphics and has been studied for decades." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.742, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Among the feasible solutions, NeRF (Neural Radiance Field) [26] approximates an implicit scene representation, i.e., a radiance field (RF), encoded by a neural network mapping 3D coordinates and view directions to colors and densities. When rendering, the fitted RF is queried multiple times along a ray to volume-render the corresponding pixel for a novel view. Subsequent variants of NeRF follow a case-by-case design and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.221, + 0.826, + 0.533, + 0.841 + ], + "angle": 0, + "content": "* Work done during Q. Zhu's internship at XREAL." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.128 + ], + "angle": 0, + "content": "Q. Zhu et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.387, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.148, + 0.782, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.29, + 0.79, + 0.36 + ], + "angle": 0, + "content": "Fig. 1: Left: RPBG manages to achieve all-round good re-renderings (PSNR plotted) across generic datasets over the baseline [1] as well as state-of-the-art RF-based methods [42,44]. Right: We demonstrate the point clouds (with camera trajectories visualized) and corresponding re-rendered novel views of the representative scenes, revealing the great robustness and scalability of RPBG. Zoom in for best view." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.395, + 0.789, + 0.456 + ], + "angle": 0, + "content": "adopt different parameterization techniques specifically for different scene types, e.g., object-centric [3, 54], free-trajectory [44], and large-scale scenes [38, 42]. Therefore, putting forward one robust method that can work stably across varying scene types with a unified parameterization technique is considered as challenging." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.458, + 0.789, + 0.685 + ], + "angle": 0, + "content": "Recently, point-based alternatives [1, 18, 30, 33] have gained substantial attention for their unique advantages over implicit representations, e.g., ease of manipulation and faster training. The practice of employing point-based parameterization for rendering can be traced all the way back to [16, 20], and is intuitive especially under the context of NVs. Analogue to NeRF and its variants (collectively referred to as RF-based methods), whose optimization and convergence heavily rely on the inherent co-visibility among images, the 3D points triangulated from a series of 2D images according to epipolar geometry have already contain all the verified co-visibility information. In this way, by enforcing the triangulated points as strong prior knowledge for NVs, the expected freedom of optimization is greatly constrained, leading to faster training as well as better robustness. Recent attempts have achieved promising results in terms of fast optimization and rendering [18, 30] and accurate geometric representation [56]. Concretely, we focus on the specific category of methods that manage to integrate CNN-based neural renderers to yield re-renderings with featuremetric neural buffers [22, 30], for their concise pipeline and the potential capability of obtaining visually pleasing re-renderings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.789, + 0.808 + ], + "angle": 0, + "content": "However, by evaluating the representative baseline, NPBG (Neural Point-based Graphics) [1], with generic datasets, we realize that the original design only intends for ideal conditions, e.g., synthetic data [26] and well-captured human heads [31]. When synthesizing novel views under more general conditions, the performance of NPBG degrades severely. In this paper, we strive to boost the robustness of NPBG-like point-based neural re-rendering pipelines and reveal the true potential to achieve state-of-the-art performance across varying datasets in-the-wild, by analyzing the reasons for performance degradation and seeking for remedies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Generally speaking, the major difficulties that the vanilla NPBG is faced up with include handling the background, handling patchy point clouds, and identifying correct" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.346, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RPGB: Towards Robust Neural Point-based Graphics in the Wild" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.146, + 0.788, + 0.388 + ], + "angle": 0, + "content": "point visibility. We reform the CNN-based neural renderer, with an inspiration from image restoration algorithms [6,8,52] that are able to identify downgraded patterns and restore the corresponding high-quality images. To make sure the neural renderer can capture sufficient and valid context information with extremely sparse rasterizations, we particularly leverages a Downgrade-aware Convolution (DAC) module to determine the correct point visibility with regard to a given camera pose and performs a pseudo point-wise back-face culling operation with visual self-attention. The background is also modeled in a lightweight manner. Instead of incorporating a massive environment map [33], a simple default trainable feature vector can reach similar quantitative results when working with the stronger neural renderer. We also discover the pseudo density calculated from the neural textures can roughly verify the existence of a given 3D position, which can be used to augment the poorly triangulated point clouds. In addition, we also simplify the phased training paradigm in [1,30] and optimize the parameters of both the neural textures and the renderer end-to-end collaboratively. We term our version of point-based re-rendering as RPBG (Robust Point-based Graphics), to emphasize its great robustness across different generic datasets." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.389, + 0.788, + 0.539 + ], + "angle": 0, + "content": "For thorough qualitative and quantitative evaluation, we collect 4 typical challenging scene types as the benchmark for robust NVs, i.e., \\(360^{\\circ}\\) unbounded scenes (with free trajectories), inside-out scenes, large-scale scenes (at the scale of a block or campus), and sparse-view scenes. As a result, RPBG exhibits great robustness with perceptually satisfactory synthesis across the aforementioned typical scenes types as showcased in Fig. 1, where the high-quality re-renderings are obtained with an exactly identical parameterization strategy without any manual configuration, relieving the practitioners from the exhaustive per-scene search of hyper-parameters. Its stable superiority over state-of-the-art NVS methods, we believe, is of great significance especially for real applications." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.541, + 0.564, + 0.554 + ], + "angle": 0, + "content": "The main contributions are three-fold as follows:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.556, + 0.784, + 0.582 + ], + "angle": 0, + "content": "- We put forward RPBG, as a more robust and practical alternative for re-rendering high-quality images from triangulated 3D points." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.587, + 0.784, + 0.628 + ], + "angle": 0, + "content": "- According to our in-depth analysis, we enhance the neural re-rendering pipeline regarding the neural renderer, environment modeling, point cloud augmentation, and the training paradigm." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.631, + 0.784, + 0.674 + ], + "angle": 0, + "content": "- RPBG manages to greatly boost the performance of neural point re-renderer by a large margin, and exhibits stably greater robustness and generalizability over RF-based methods with even better perceptual rendering quality." + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.556, + 0.784, + 0.674 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.701, + 0.367, + 0.716 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Radiance Fields for NVs Along with the proposed NeRF [26], the scene representation of RF is becoming popular for its ease of optimization by the differentiable volume rendering. In a RF, each position is assigned with an anisotropic color and a density, and to render a pixel of a novel view, one needs to sample the trained field and conduct ray marching. The original NeRF [26], together with its variants [2,3,39,54], employs an MLP to represent the functional mapping from coordinates and view directions to the radiance values. Several attempts have been made to encode RFs with different" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.117, + 0.229, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.348, + 0.129 + ], + "angle": 0, + "content": "Q. Zhu et al." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.147, + 0.786, + 0.177 + ], + "angle": 0, + "content": "data structures for faster training and inference, e.g., grid voxels [12, 36], decomposed tensors [5], hash grids [27]." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.179, + 0.788, + 0.328 + ], + "angle": 0, + "content": "In practice, the parameterization strategies have a great impact on the rendering quality when reconstructing different types of scenes, e.g., forward-facing ones, \\(360^{\\circ}\\) unbounded ones, and large-scale ones. For unbounded scenes, \\(\\mathrm{NeRF}++\\) [54] applies separate networks and different parameterization to model near and distant objects; mip-NeRF 360 [3] designs a smooth contraction operator to parameterize the whole unbounded scene into a ball; Mega-NeRF [42] and Block-NeRF [38] partition 3D scenes explicitly and use different networks to represent each scene partition; \\(\\mathrm{F}^2\\)-NeRF [44] proposes to use perspective warping to handle sequential data with arbitrary trajectories. The varying case-by-case parameterization strategies of RF-based NVS greatly constraint the generalizability of methods." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.344, + 0.788, + 0.434 + ], + "angle": 0, + "content": "Point-based Graphics A point cloud is a collection of 3D coordinates that is usually used as a topology-agnostic coarse shape representation of the 3D geometry and favored for its flexibility and sparsity for storage. The development of techniques of employing points as modeling primitives for rendering (referred as point-based graphics [15]) can be traced back to [16, 20], the best practice of which is to replace each point with an oriented circular disk (a surfel) and reply on splatting to blend overlapping surfels [28]." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.435, + 0.788, + 0.646 + ], + "angle": 0, + "content": "In contrast to conventional physically based rendering (PBR) techniques, neural rendering [40] learns to render high-quality images in a data-driven manner. Particularly, we focus on the methods that conduct neural re-rendering with point clouds. Bui et al. [4] propose to enhance the coarse point-based rendering by a GAN for image super-resolution. NRW [25] and InvSFM [29] attempt to re-render the reconstructed point cloud with respective auxiliary buffers, e.g., latent appearance vectors, semantic masks and SIFT descriptors [23]. NPBG [1,30] adopts a U-Net-like [32] CNN to render neural point textures as RGB images, and exhibits better flexibility over mesh-based proxies [41]. The practical applications in the context of autonomous driving, e.g., scene editing and stitching, and large-scale training, are further explored in READ [21]. Recently, ADOP [33] and Gaussian Splatting [18] have demonstrated notable accomplishments. However, the differentiable rendering/splatting scheme employed requires an extensive amount of memory to maintain the computational graph and gradients during training. This limitation hinders their application on large-scale scenes." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.669, + 0.573, + 0.687 + ], + "angle": 0, + "content": "3 Neural Point-based Graphics Revisited" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.702, + 0.786, + 0.762 + ], + "angle": 0, + "content": "In this section, we revisit the pipeline of NPBG [1], analyze the existing drawbacks when attempting to re-render generic scenes, and attempt to explain why NPBG finds it difficult to handle the background and patchy points, and identify correct point visibility. We hope these discussions be of sufficient insights to support our modifications." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.785, + 0.35, + 0.799 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.81, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Inputs As the common practice in NVS, the inputs are a series of images with respective camera parameters (both intrinsics and extrinsics), and specifically for point-based" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.346, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "RPBG: Towards Robust Neural Point-based Graphics in the Wild" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.148, + 0.363, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.276, + 0.342, + 0.285 + ], + "angle": 0, + "content": "(a) Rasterized Points" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.148, + 0.501, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.276, + 0.466, + 0.285 + ], + "angle": 0, + "content": "(b) Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.148, + 0.639, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.276, + 0.591, + 0.285 + ], + "angle": 0, + "content": "(c) NPBG" + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.149, + 0.778, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.69, + 0.276, + 0.726, + 0.285 + ], + "angle": 0, + "content": "(d) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.3, + 0.787, + 0.342 + ], + "angle": 0, + "content": "Fig. 2: Typically challenging scenes in T&T dataset [19] for NPBG. Top: Auditorium, where the walls and ceilings are extremely sparse. Bottom: Museum, where the point sparsity makes the occlusion and visibility complicated." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.368, + 0.787, + 0.414 + ], + "angle": 0, + "content": "methods [1, 18, 30], the very first step is supposed to be the triangulation of 3D points from 2D observations. The original NPBG-like pipelines [1, 30] mainly consider this as a pre-processing without much attention." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.426, + 0.788, + 0.577 + ], + "angle": 0, + "content": "Point Rasterization NPBG [1] and NPBG++ [30] apply a non-differentiable hard point \\(z\\)-buffering operation as an approximated back-face culling when rasterizing points as 2D fragments, where the fragment is updated when and only when the newly projected point has a smaller \\(z\\)-depth than the current one. After a traversal of all the points, the fragment keeps the record of the indices of rasterized points. Then tensor scattering is performed to index the neural texture at the corresponding positions. Compared to differentiable rendering/splatting alternatives [18, 33], we consider the most significant advantage is its memory efficiency and thus better scalability, since the computation graph required to keep is much smaller for NPBG. We also want to keep this scalability without interfering the elegant rasterization paradigm." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.587, + 0.788, + 0.68 + ], + "angle": 0, + "content": "Neural Renderer The renderer \\(\\mathcal{R}_{\\Theta}:\\mathbb{R}^{H\\times W\\times C}\\to \\mathbb{R}^{H\\times W\\times 3}\\) of NPBG [1] follows a vanilla U-Net-like architecture [32]. Since the target point clouds for rendering are supposed to be well-constructed, the vanilla U-Net is expressive enough to complete the assigned task of mapping higher dimensional features from the neural texture T to RGB values. Empirically, we find the expected properties of \\(\\mathcal{R}_{\\Theta}\\) have much in common with a network for low-level vision tasks, e.g., image restoration." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.698, + 0.325, + 0.712 + ], + "angle": 0, + "content": "3.2 Problems" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Patchy Triangulation Consider the inputs, for a NVS system aiming to synthesize images for generic scenes, the triangulation of points is not trivial. The point triangulation step to perform is similar to the step of multi-view stereo (MVS) reconstruction [14, 48], the key of which is to identify co-visible pixels across images and lift the 2D pixels to a 3D points according to epipolar geometry. Such algorithms suffer from non-Lambertian surfaces and textureless regions, both of which are very common and lead to poor triangulated and patchy points. We attach two typical scenes that NPBG fails to yield good renderings in Fig. 2." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.117, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.348, + 0.129 + ], + "angle": 0, + "content": "Q. Zhu et al." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "Wrong Point Visibility As can be imagined, such point visibility can be erroneous due to the poor quality of triangulated points, and also due to the inherent sparsity of points as a 3D representation. In this way, the points belonging to back faces, which should be considered as occluded, are rasterized as the fragment for further processing. Manually setting a depth threshold for bound check could not be helpful either, since being far from the camera does not necessarily indicate they should be occluded — they could be parts of the environment. While NHR [47] proposes to take the depth buffer as an additional rendering condition, we find its effectiveness not significant for generic scenes other than well-masked human captures." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.298, + 0.788, + 0.403 + ], + "angle": 0, + "content": "Lack of Context for Re-rendering Similarly, when re-rendering an incomplete point cloud, where the neural buffers are usually with significant flaws, the whole receptive field of a kernel at a given position may only capture the downgraded regions, leading to failure to yield reasonable restoration, especially for a high target resolution. It is worth noting that, though the relationship between the number of network layers and rendering quality has been discussed in [30, 33], there is still huge room for the improvement of the renderer." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.419, + 0.788, + 0.525 + ], + "angle": 0, + "content": "Failure to Effectively Model the Environment Unbounded scenes are a typical challenging scene type for NVS. RF-based methods that are designed on purpose to handle such scenes [3, 54] typically employ different parameterization to encode the areas outside a certain bounding sphere. For point-based methods, ADOP [33] proposes to apply an environment map of \\( H \\times W \\times C \\) to model the environment, which is equivalent to wrapping the triangulated points with a sphere with \\( H \\times W \\) points, resulting in more than \\( 5 \\times 10^{5} \\) points as overheads, according to the default configuration." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.54, + 0.788, + 0.586 + ], + "angle": 0, + "content": "Summary To summarize, we attribute the observed problems to two main causes: poor geometry and weak, local renderer. The two causes are to some extent coupled for a better geometry will relieve the difficulty of the renderer and vice versa." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.609, + 0.497, + 0.627 + ], + "angle": 0, + "content": "4 Robust Point-based Graphics" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.641, + 0.788, + 0.762 + ], + "angle": 0, + "content": "We are strongly convinced that the pipeline of point-based neural re-rendering has the great potential to outperform popular RF-based solutions, for its unified point-as-parameterization fashion and the incorporation of neural networks to yield visual details. Therefore in this section, we introduce Robust Point-based Graphics (RPBG), as an enhanced version of NPBG, with a particular focus on robustness under generic scenes. We will elaborate the modifications made and shed light on the underlying insights and motivation. The overall pipeline of RPBG as well as the training paradigm is briefly illustrated in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.784, + 0.508, + 0.8 + ], + "angle": 0, + "content": "4.1 Downgrade-aware Neural Renderer" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "As is attributed as one key problem in NPBG, the U-Net-based neural renderer is considered as too naive to handle the challenging situations in generic scenes. The" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.346, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "RPBG: Towards Robust Neural Point-based Graphics in the Wild" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.784, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.485 + ], + "angle": 0, + "content": "Fig. 3: The overall pipeline of RPBG. Point Triangulation: We first triangulate a 3D proxy for re-rendering with posed images, with its geometry-bounded neural texture initialized. Point Rasterization: The points are raterized to the given camera in a non-differentiable manner. By indexing the texture with the fragment, we obtain the neural buffer. A learnable point-size neural texture \\(\\mathbf{T}_{\\mathrm{env}}\\) is also optimized. Neural Rendering: The restoration from downgraded neural buffer to photo-realistic images is performed by a CNN. The network and the neural texture are optimized end-to-end by image-level losses. An offline point cloud augmentation strategy is introduced to alleviate the problem of patchy triangulation under challenging conditions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.788, + 0.61 + ], + "angle": 0, + "content": "modifications made are deeply inspired by low-level vision tasks, where networks can adaptively determine downgraded parts, e.g., the deblurring/deraining networks are able to identify the blurring/raining pixels from the whole image, and correct them accordingly. We expect the neural renderer \\(\\mathcal{R}_{\\Theta}\\) can benefit from relevant restoration-targeting techniques, and become able to decode high-quality visual information from patchy buffers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.788, + 0.732 + ], + "angle": 0, + "content": "After evaluating three state-of-the-art fundamental architectures for image restoration, i.e., multi-scale fusion [8], multi-stage [52], and U-Net [6], following the taxonomy in [6]. According to the reported experiments in Sec. 5.4, we opt the paradigm in [8] for it achieves the best balance between performance and time/memory efficiency. However, the convolution layers in such paradigm is still with a fixed receptive field, lacking in robustness against point sparsity. We would like to further enlarge the receptive field to a global scale and attempt to explicitly model the visual attention to weigh the observed points." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Transformer-based architectures for image restoration, e.g., Restormer [51], brings unacceptable memory overheads, so we rely on the frequency-domain alternative, Fast Fourier Convolution (FFC) [7], which can theoretically capture global contexts, to determine the correct point visibility adaptively. FFC performs channel-wise real 2D FFT (Fast Fourier Transform) and inverse real 2D FFT on 2D tensors. Real FFT uses only half of the spectrum and by convolving the transformed frequency-domain tensors. In this way, a receptive field covering the entire image is considered." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.229, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.129 + ], + "angle": 0, + "content": "Q. Zhu et al." + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.15, + 0.725, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.469, + 0.79, + 0.514 + ], + "angle": 0, + "content": "Fig. 4: The architecture of the downgrade-aware neural renderer in RPBG, with some conventional modules omitted. From the visualized attention map, DAC manages to adaptively handle the severely erroneous point visibility." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.54, + 0.79, + 0.664 + ], + "angle": 0, + "content": "Inspired by [37], we apply FFT (as the global branch) in parallel to conventional convolution layers (as the local branch) and rely on the fused features of both branches to determine the downgraded regions for the gated convolution [50] to filter. Based on the common practice in image inpainting, we leverage gated convolutions at the early stage of the renderer, to help locate the downgrade by wrong point visibility. We name such customized gated convolution module as the Downgrade-aware Convolution (DAC) module, and as can be inferred from the point attention maps before and after DAC in Fig. 4, DAC manages to determine complicated point visibility with patchy triangulation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.681, + 0.396, + 0.696 + ], + "angle": 0, + "content": "4.2 Point Triangulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.765 + ], + "angle": 0, + "content": "Though there are advances in MVS empowered by deep learning [43, 45, 46], recovering a complete, accurate, and dense point cloud for certain regions remains a challenging. For RPBG, we adopt an off-the-shelf MVS method [46] to perform point triangulation for its memory efficiency to allow large-batch inference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.79, + 0.842 + ], + "angle": 0, + "content": "As for problem the poor triangulation, we partially leave it to the neural renderer \\(\\mathcal{R}_{\\Theta}\\), relying on a stronger renderer to recover the re-rendering from downgraded buffers. Inspired by the point cloud extraction method of RF-based implementations [39], where the estimated radiance density \\(\\sigma\\) can be a rough indicator for surface, we propose a point cloud augmentation technique to densify the initial triangulation." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.345, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "RPBG: Towards Robust Neural Point-based Graphics in the Wild" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.236 + ], + "angle": 0, + "content": "The point cloud augmentation follows a trial-and-error paradigm by first assuming the existence of newly sampled points and then verify them by estimated pseudo densities \\(\\sigma_{i} = \\sum |\\mathbf{T}_{i}|\\). It is empirically observed that the absolute activation of the point-wise neural texture can roughly represent the reliability of the 3D position — it is reasonable since an outlier will be less distinguishable by the neural renderer and thus less visual attention will be given." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.785, + 0.283 + ], + "angle": 0, + "content": "Note that this strategy is optional, and we only apply such strategy to scenes with extremely poor triangulation. Please refer to the Supplementary Material for more discussions regarding the self-pruning." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.306, + 0.421, + 0.321 + ], + "angle": 0, + "content": "4.3 Environment Modeling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.331, + 0.784, + 0.391 + ], + "angle": 0, + "content": "As a result of the incorporation of a stronger neural renderer, we find that given a neural texture with \\( C = 8 \\), the dense environment map suggested in [33] is redundant. Instead, we shrink the overheads for environment modeling from \\( H \\times W \\times C \\) to \\( 1 \\times C \\), and relying on the stronger neural renderer to decode the background." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.392, + 0.785, + 0.467 + ], + "angle": 0, + "content": "When rasterizing, we employ a tunable feature vector \\(\\mathbf{T}_{\\mathrm{env}}\\) aside the neural point texture \\(\\mathbf{T}\\), as the default value for vacant pixels in the fragment, which is also involved in the end-to-end training. By experiments in Sec. 5.4, we demonstrate that the lightweight modeling strategy reaches the quantitative performance equivalent to applying an environment map, yet with a negligible overhead." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.489, + 0.449, + 0.505 + ], + "angle": 0, + "content": "4.4 Collaborative Optimization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.784, + 0.545 + ], + "angle": 0, + "content": "Recall that the point rasterization procedure is completely parameter-free. The overall collaborative optimization scheme can be thus formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.256, + 0.557, + 0.788, + 0.588 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} ^ {*}, \\Theta^ {*} = \\underset {\\mathbf {T}, \\Theta} {\\arg \\min } \\sum_ {k} \\mathcal {L} (\\mathbf {I} _ {k}, \\hat {\\mathbf {I}} _ {k}) = \\underset {\\mathbf {T}, \\Theta} {\\arg \\min } \\sum_ {k} \\mathcal {L} (\\mathbf {I} _ {k}, \\mathbf {K}, \\mathbf {R} _ {k}, \\mathbf {t} _ {k} | \\mathbf {T}, \\mathcal {R} _ {\\Theta}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.785, + 0.644 + ], + "angle": 0, + "content": "where \\(\\mathbf{T}\\) and \\(\\Theta\\) stand for the tunable parameters, i.e., the neural point texture (alongside the globally shared \\(\\mathbf{T}_{\\mathrm{env}}\\)) and the parameters of the renderer \\(\\mathcal{R}_{\\Theta}\\) for neural re-rendering, while \\(\\mathbf{I}_k\\) is the target image whose calibration is \\(\\mathbf{K}\\) and \\([\\mathbf{R}_k|\\mathbf{t}_k]\\) and \\(\\hat{\\mathbf{I}}_k\\) is the re-rendering" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.785, + 0.704 + ], + "angle": 0, + "content": "Note that, we also discard all the typical tossing sampling [21] and optimization steps [30, 33] of point-based neural re-rendering, and manage to tune all the parameters involved in a simple but effective collaborative end-to-end way. The neural texture is initialized with all zeros while the rendering CNN is trained from scratch for each scene" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Loss Function Since the rendering scheme in both NPBG and RPBG is convolutional, where images are rendered in patches, we are able to apply patch-aware losses to enforce the involvement of neighboring pixels to ensure patch-to-patch consistency, offering perceptually good renderings. To this end, in addition to the pixel-wise Huber norm \\(\\mathcal{L}_{\\mathrm{huber}}\\) providing the basic supervision and numerical stability, we apply two patch-aware losses to the collaborative optimization of RPBG, namely the perceptual VGG loss [10, 17] \\(\\mathcal{L}_{\\mathrm{vgg}}\\), and the FFT loss [8, 13] \\(\\mathcal{L}_{\\mathrm{fft}}\\). VGG loss compares the rendered image and the target ground-truth image in a high-dimensional feature space by a pre-trained VGG-19" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.128 + ], + "angle": 0, + "content": "Q. Zhu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.207 + ], + "angle": 0, + "content": "network [35]. It reveals the perceptual similarity between images that pixel-wise metrics fail to measure. FFT loss measures the image-to-image distance in the frequency domain by carrying out 2D FFT towards images. The frequency components are considered to be crucial in terms of the perceptual quality [13]." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.208, + 0.61, + 0.222 + ], + "angle": 0, + "content": "The final loss function applied in RPBG is composed as" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.228, + 0.786, + 0.246 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {\\text {h u b e r}} \\mathcal {L} _ {\\text {h u b e r}} + \\lambda_ {\\text {v g g}} \\mathcal {L} _ {\\text {v g g}} + \\lambda_ {\\text {f f t}} \\mathcal {L} _ {\\text {f f t}}. \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.263, + 0.357, + 0.279 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.29, + 0.317, + 0.303 + ], + "angle": 0, + "content": "5.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.31, + 0.787, + 0.46 + ], + "angle": 0, + "content": "We conduct experiments covering a diverse range of scene types to quantitatively examine the generalizability and robustness of NVs methods, including the Free dataset [44] (7 free-trajectory scans with long camera trajectories), Tanks and Temples (T&T) dataset [19] (4 unbounded scans and 5 inside-out scans, without salient dynamic objects), Mill19 dataset [42] (2 large-scale aerial scans with over 1600 images in each scan), and GigaMVS dataset [53] (8 sparse-view outdoor scans). Note that for T&T, we use the raw undistorted images provided by the benchmark without any masking [22]. We follow the common split protocol, that 1 frame out of every 8 frames is evaluated, for the Free dataset, T&T dataset, and GigaMVS dataset. For Mill19 dataset, we apply the officially recommended split to align with the experiments in [42]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.787, + 0.535 + ], + "angle": 0, + "content": "We also include some other challenging datasets for a thorough evaluation, e.g., ScanNet++ dataset [49], a challenging indoor benchmark, OMMO dataset [24], a multi-modal aerial NVS dataset and ETH-MS dataset [11], a super-large-scale dataset (around \\(5k\\) images). Please refer to the Supplementary Material for more quantitative and qualitative results." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.554, + 0.421, + 0.568 + ], + "angle": 0, + "content": "5.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.574, + 0.787, + 0.678 + ], + "angle": 0, + "content": "Data Preparation To obtain the per-scene point cloud representation for re-rendering, we follow the mapping procedure of COLMAP [34] for sparse triangulation and the MVS reconstruction with a trained network [46] for dense triangulation. Note that the reconstruction only takes a small proportion of time over the whole training phase and we conduct ablation experiments to show that RPBG is robust against different triangulation configurations. Please refer to Sec. 5.4 as well as the Supplementary Material for more information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Training Settings We randomly crop images to square patches of \\(256 \\times 256\\), with a batch size of 8. The learning rate for the neural point textures is \\(10^{-1}\\), and \\(10^{-4}\\) for the rendering networks, which will decay by a factor of 0.5 if 5 consequent epochs witness no drop in the loss function. A dimension of 8 is applied for any neural texture regardless of the scene scale or complexity. For the weights of loss functions, we globally set \\(\\lambda_{\\mathrm{huber}} = 10^3\\), \\(\\lambda_{\\mathrm{vgg}} = 1\\), and \\(\\lambda_{\\mathrm{fft}} = 1\\). The training is performed on one NVIDIA GeForce RTX 3090, with a GPU memory consumption of up to 23 GB. It takes around 8 to 30 GPU hours for training, depending on the data scale. We would like to emphasize that, RPBG does not require case-by-case scene parameterization or grid search of training hyper-parameters for all the scenes covered in the experiments." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.346, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "RPBG: Towards Robust Neural Point-based Graphics in the Wild" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.148, + 0.784, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.458, + 0.79, + 0.514 + ], + "angle": 0, + "content": "Fig. 5: Visualized comparisons over varying scenes. From top to bottom: sky and hydrant of the Free dataset [44], Courtyard and Train of T&T dataset [19], and DayaTemple and MemorialHall of GigaMVS dataset [53]. We include the results of RPBG (Ours), Gaussian Splatting [18], NPBG [1], and \\(\\mathrm{F}^2\\)-NeRF [44] for comparison. Zoom in for best view." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.544, + 0.788, + 0.605 + ], + "angle": 0, + "content": "Evaluation Metrics We adopt the metrics of PSNR, SSIM and LPIPS (VGG) [55] for the evaluation between the synthesized and the target images. According to [55], PSNR does not faithfully measure image sharpness and so cannot properly account for the nuances of human visual perception." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.63, + 0.31, + 0.644 + ], + "angle": 0, + "content": "5.3 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.657, + 0.788, + 0.749 + ], + "angle": 0, + "content": "In Fig. 5, we demonstrate several typical groups of visual comparisons of NVS results by RPBG, two representative point-based methods, i.e., Gaussian Splitting [18] and NPBG [1], and a state-of-the-art NeRF variant for unbounded scenes, \\(\\mathrm{F}^2\\)-NeRF [44]. We group the datasets by category and report the corresponding scene-averaged quantitative scores in Tab. 1. Note that RPBG achieves the best SSIM and LPIPS, which are more relevant to the high-frequency components of images, across all datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Free-trajectory/Unbounded Scenes Compared with NVS methods designed in particular for unbounded scenes, \\(\\mathrm{NeRF}++\\) [54] and \\(\\mathrm{F^2}\\)-NeRF [44], RPBG achieves the best SSIM and LPIPS and comparable PSNR. For visual effects showcased in Fig. 5, since RPBG involves a neighborhood for rendering, its results appear more visually harmonious, especially than Gaussian Splatting [18] and \\(\\mathrm{F^2}\\)-NeRF [44]. It is necessary to mention" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.129 + ], + "angle": 0, + "content": "Q. Zhu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.2 + ], + "angle": 0, + "content": "Table 1: Quantitative evaluation of state-of-the-art NVS methods [1,5,18,26,30,42,44,54] and RPBG across diverse scenes grouped by category, including free trajectory/unbounded scenes, inside-out scenes, large-scale scenes, and sparse-view scenes. The figure following the dataset name stands for the number of scenes the dataset contains." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.213, + 0.787, + 0.33 + ], + "angle": 0, + "content": "
MethodFree-trajectory/UnboundedInside-out T&T-5Large-scale Mill19-2Sparse-view GigaMVS-8
Free Dataset-7T&T-4
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NeRF [26]17.750.4050.59716.840.3960.73119.820.6150.52620.340.5240.52918.290.6320.499
TensoRF [5]21.740.5490.60020.730.6430.56619.430.6340.57020.070.4970.59718.700.6530.506
NeRF++ [54]23.470.6030.49921.660.6580.52919.250.6100.58520.190.5200.53118.380.6320.495
F2-NeRF [44]26.320.7790.27623.660.7640.30320.120.7060.394N/AN/AN/A17.440.5400.470
Mega-NeRF [42]22.600.5700.56218.730.5780.47819.160.6170.45122.490.5500.51018.250.5810.394
NPBG [1]21.400.6390.34019.850.6980.37620.570.6960.37116.210.3570.64418.340.6200.405
NPBG++ [30]20.060.5920.44517.230.6530.47418.300.6840.41117.040.4000.64819.300.6630.443
Gaussian Splatting [18]25.230.7400.29023.510.7820.29323.460.7830.277N/AN/AN/A16.840.5300.391
Ours26.330.8320.17722.500.7820.27623.290.8040.24222.620.5960.36820.540.6860.317
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.785, + 0.391 + ], + "angle": 0, + "content": "that, for RF-based methods [5, 26, 42, 44], we have manually adjusted the scene-specific hyper-parameters to achieve better results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.787, + 0.464 + ], + "angle": 0, + "content": "Inside-out Scenes The inside-out indoor scenes are strictly bounded, but lack high-quality ray intersections required for optimization, which will lead to under-fitting of the RF (Courtroom in Fig. 5). Similar to free-trajectory/unbounded scenes, RPBG outperforms state-of-the-art methods in SSIM and LPIPS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.788, + 0.584 + ], + "angle": 0, + "content": "Large-scale Scenes We mainly evaluate RPBG against Mega-NeRF [42] on the massive imagery of Mill19 dataset [42]. Our method outperforms Mega-NeRF [42] in every metric. It is worth noting that Mega-NeRF takes 240 GPU hours for training while RPBG only takes 29 GPU hours. Notably, our system with 256 GB RAM and RTX 3090 fails to afford the training of \\(\\mathrm{F}^2\\)-NeRF [44] and Gaussian Splitting [18] (marked as N/A in the table). In contrast, RPBG proves to be viable for large-scale scenes under the same hardware constraint." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.788, + 0.689 + ], + "angle": 0, + "content": "Sparse-view Scenes Sparse-view inputs are considered to be extremely challenging for NVS methods. Since RPBG, NPBG [1], and \\(\\mathrm{NPBG + + }\\) [30] incorporate a point-based 3D proxy for re-rendering, they showcase better robustness over RF-based methods. As revealed in Fig. 5, Gaussian Splatting [18] fails to regularize the strong approximation power of point-wise Gaussians at training views, and yields obvious needle-like artifacts at novel views." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.71, + 0.363, + 0.726 + ], + "angle": 0, + "content": "5.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.434, + 0.84 + ], + "angle": 0, + "content": "Environment Modeling We study different strategies of environment modeling for NVS in-the-wild, namely leaving the blank pixels with zeros, filling the blanks with a default learnable feature (as is done in RPBG), and" + }, + { + "type": "image", + "bbox": [ + 0.449, + 0.73, + 0.614, + 0.801 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.802, + 0.578, + 0.812 + ], + "angle": 0, + "content": "(a) \\( w / o \\) Env. Modeling" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.73, + 0.78, + 0.801 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.801, + 0.741, + 0.811 + ], + "angle": 0, + "content": "(b) \\(w / \\mathrm{Env}\\) Modeling" + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.817, + 0.787, + 0.845 + ], + "angle": 0, + "content": "Fig.6: The impact of leveraging the default feature vector for environment modeling in an unbounded scene." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.345, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "RPBG: Towards Robust Neural Point-based Graphics in the Wild" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.5, + 0.2 + ], + "angle": 0, + "content": "Table 2: Ablation on network designs on T&T dataset [19]. The reported inference time and memory consumption is tested with a target resolution of \\( {1920} \\times {1080} \\) ." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.213, + 0.498, + 0.287 + ], + "angle": 0, + "content": "
MethodTime(s)Mem.(GB)PSNR↑SSIM↑LPIPS↓
Baseline1.035.4920.010.6660.390
+Multi-scale fusion1.6212.6721.880.7010.337
+DAC1.6713.0122.920.7690.320
+FFT loss--22.940.7940.257
+Multi-stage1.7717.1421.750.7130.338
+U-Net1.5810.3220.850.6920.364
" + }, + { + "type": "table_caption", + "bbox": [ + 0.503, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 3: Ablation on environment modeling on two unbounded scenes, namely Caterpillar of T&T dataset [19] and sky of Free dataset [44]. #Points \\((\\times 10^{6})\\) represent the equivalent overheads when applying different strategies." + }, + { + "type": "table", + "bbox": [ + 0.509, + 0.227, + 0.785, + 0.274 + ], + "angle": 0, + "content": "
Env.Caterpillarsky
#PointsPSNR↑SSIM↑LPIPS↓#PointsPSNR↑SSIM↑LPIPS↓
Zeros7.9519.890.4560.40317.4921.770.5980.315
Learnable7.9521.780.6870.28917.4924.810.8660.199
Sphere8.9521.680.6860.28318.4925.150.8690.179
" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.303, + 0.357, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.361, + 0.357, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.269, + 0.429, + 0.307, + 0.435 + ], + "angle": 0, + "content": "(c) \\( w / \\theta \\) DAC" + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.303, + 0.493, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.362, + 0.493, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.407, + 0.429, + 0.442, + 0.435 + ], + "angle": 0, + "content": "(d) \\(w / \\mathrm{DAC}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.448, + 0.499, + 0.49 + ], + "angle": 0, + "content": "Fig.7: A challenging case in Train, which is a zoomed-in view with complex occlusion and thin objects (e.g., the handrails)." + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.304, + 0.784, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.424, + 0.788, + 0.493 + ], + "angle": 0, + "content": "Fig. 8: Re-rendering quality (PSNR) with points of different quality on Church. RPBG manages to maintain its robustness against different levels of point sparsity with a \\(14\\%\\) drop in PSNR (NPBG [1]: \\(25\\%\\); NPBG++ [30]: \\(14\\%\\))." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.526, + 0.788, + 0.573 + ], + "angle": 0, + "content": "wrapping the points with a sphere of \\(10^{6}\\) points (equivalent to an environment map in [33]). The quantitative results are in Sec. 5.4, where the strategy incorporated by RPBG boosts the performance by a large margin, with almost no additional overhead." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.593, + 0.789, + 0.759 + ], + "angle": 0, + "content": "Network Architecture and Loss We study the key components in the rendering network in Sec. 5.4. The baseline refers to NPBG [1] trained per-scene from scratch with \\(\\mathcal{L}_1\\) and VGG loss. We compare three fundamental architectures for image restoration (with necessary modifications on the first several layers), i.e., multi-scale fusion [8], multistage [52], and U-Net [6] following the taxonomy in [6]. Compared to the baseline, all modern restoration-oriented architectures bring remarkable improvement to the overall performance, and [8] is opted for a better balance between performance and time/memory efficiency. The gain by DAC is also considerable while the FFT loss mainly improves the perceptual quality. By the case in Fig. 7, we show the necessity of the DAC module when re-rendering challenging cases without in particular handling the erroneous point visibility." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.79, + 0.842 + ], + "angle": 0, + "content": "Triangulation Configurations We also evaluate RPBG with different triangulation configurations. Indicated by the results in Fig. 8, RPBG witnesses a total drop of \\(3.29\\mathrm{dB}\\) in re-rendering PSNR when switching the triangulation from ground truth (GT) to randomly initialized points (Random). Note that even with sparse points (SfM)," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.348, + 0.128 + ], + "angle": 0, + "content": "Q. Zhu et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.177 + ], + "angle": 0, + "content": "which are usually yielded as a by-product when computing camera parameters, RPBG (21.41 dB) still outperforms \\(\\mathrm{F}^2\\)-NeRF (20.57 dB)." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.2, + 0.338, + 0.216 + ], + "angle": 0, + "content": "6 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.383 + ], + "angle": 0, + "content": "Effectiveness of RPBG We would like to shed some light on the effectiveness of RPBG. The first factor should be the point representation. The triangulated 3D points have already contained all the verified co-visibility that is required for NVS. Compared to RF-based methods, which often struggle at local minima, such triangulation-as-parameterization paradigm greatly lower the complexity of searching, and is unified across different scene types. Besides, the patch-wise rendering scheme mentioned in Sec. 4.4 also improves the perceptual quality of re-rendering while RF-based methods apply pixel/ray-wise volume rendering, where the correlation between pixels are not modeled explicitly. For more discussions on the insights, we strongly refer the readers to the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.549 + ], + "angle": 0, + "content": "Limitations An obvious problem with point-based re-rendering (RPBG and NPBG [1]) is that compared to lightweight RF-based variants [5, 27], they take more space to store the CNN parameters and the neural texture. Although it makes the scaling-up easier, the number of parameters grows linearly with the point cloud scale. Besides, due to the non-local neural renderer we employ, each point is encoded with visual context of a larger range, hindering the editability of RPBG. We have also noticed that, the CNN-based rendering scheme can lead to unsatisfactory temporal consistency especially when the triangulation is patchy so that we mainly count on the inpainting ability of the neural render for yielding re-renderings. Such flicker issue, as well as its potential solutions, has been discussed in [9]." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.572, + 0.344, + 0.589 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.605, + 0.787, + 0.771 + ], + "angle": 0, + "content": "In this paper, we present RPBG as a robust and practical alternative to NPBG, a baseline of point-based NVs methods, performing neural re-rendering on triangulated points. We analyze the key problems in NPBG, when attempting to generalize to more generic scenes other than the well-captured scans, and reform the pipeline to reveal the real potential of point-based graphics. Respectively motivated and inspired by RF-based methods and low-level image restoration methods, we reform the pipeline according to our analysis. By extensive experiments on diverse datasets, RPBG achieves stably superior results over state-of-the-art RF-based and point-based NVs methods, especially on the metrics with more attention paid to the nuances of human visual perception, without case-by-case parameterization across all scenes, indicating its robustness and generalizability." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.345, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "RPBG: Towards Robust Neural Point-based Graphics in the Wild" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.375, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.205 + ], + "angle": 0, + "content": "This research was supported in part by JSPS KAKENHI Grant Numbers 24K22318, 22H00529, 20H05951, and JST-Mirai Program JPMJMI23G1." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.226, + 0.312, + 0.243 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.254, + 0.788, + 0.296 + ], + "angle": 0, + "content": "1. Aliev, K.A., Sevastopolsky, A., Kolos, M., Ulyanov, D., Lempitsky, V.: Neural point-based graphics. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 696-712. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.297, + 0.788, + 0.337 + ], + "angle": 0, + "content": "2. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855–5864 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.338, + 0.788, + 0.378 + ], + "angle": 0, + "content": "3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.379, + 0.788, + 0.404 + ], + "angle": 0, + "content": "4. Bui, G., Le, T., Morago, B., Duan, Y.: Point-based rendering enhancement via deep learning. The Visual Computer 34, 829-841 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.788, + 0.432 + ], + "angle": 0, + "content": "5. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: European Conference on Computer Vision. pp. 333-350. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.433, + 0.788, + 0.459 + ], + "angle": 0, + "content": "6. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: European Conference on Computer Vision. pp. 17-33. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.46, + 0.788, + 0.485 + ], + "angle": 0, + "content": "7. Chi, L., Jiang, B., Mu, Y.: Fast fourier convolution. Advances in Neural Information Processing Systems 33, 4479-4488 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.486, + 0.788, + 0.527 + ], + "angle": 0, + "content": "8. Cho, S.J., Ji, S.W., Hong, J.P., Jung, S.W., Ko, S.J.: Rethinking coarse-to-fine approach in single image deblurring. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4641-4650 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.528, + 0.788, + 0.568 + ], + "angle": 0, + "content": "9. Dai, P., Zhang, Y., Li, Z., Liu, S., Zeng, B.: Neural point cloud rendering via multi-plane projection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7830-7839 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.569, + 0.788, + 0.595 + ], + "angle": 0, + "content": "0. Dosovitskiy, A., Brox, T.: Generating images with perceptual similarity metrics based on deep networks. Advances in neural information processing systems 29 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.596, + 0.788, + 0.635 + ], + "angle": 0, + "content": "11. ETH Zurich Computer Vision Group and Microsoft Mixed Reality & AI Lab Zurich: The ETH-Microsoft Localization Dataset. https://github.com/cvg/visloc-iccv2021 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.636, + 0.788, + 0.677 + ], + "angle": 0, + "content": "12. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.678, + 0.788, + 0.718 + ], + "angle": 0, + "content": "13. Fuoli, D., Van Gool, L., Timofte, R.: Fourier space losses for efficient perceptual image super-resolution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2360-2369 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.719, + 0.788, + 0.744 + ], + "angle": 0, + "content": "14. Furukawa, Y., Ponce, J.: Accurate, dense, and robust multiview stereopsis. IEEE transactions on pattern analysis and machine intelligence 32(8), 1362-1376 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.745, + 0.601, + 0.758 + ], + "angle": 0, + "content": "15. Gross, M., Pfister, H.: Point-based graphics. Elsevier (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.759, + 0.788, + 0.799 + ], + "angle": 0, + "content": "16. Grossman, J.P., Dally, W.J.: Point sample rendering. In: Rendering Techniques' 98: Proceedings of the Eurographics Workshop in Vienna, Austria, June 29—July 1, 1998 9. pp. 181-192. Springer (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.788, + 0.84 + ], + "angle": 0, + "content": "17. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and superresolution. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 694-711. Springer (2016)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.254, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.128 + ], + "angle": 0, + "content": "Q. Zhu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "18. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG) 42(4), 1-14 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "19. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG) 36(4), 1-13 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.661, + 0.218 + ], + "angle": 0, + "content": "20. Levoy, M., Whitted, T.: The use of points as a display primitive (1985)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.219, + 0.787, + 0.244 + ], + "angle": 0, + "content": "21. Li, Z., Li, L., Zhu, J.: Read: Large-scale neural scene rendering for autonomous driving. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1522-1529 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.786, + 0.272 + ], + "angle": 0, + "content": "22. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.786, + 0.299 + ], + "angle": 0, + "content": "23. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. International journal of computer vision 60, 91-110 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.786, + 0.34 + ], + "angle": 0, + "content": "24. Lu, C., Yin, F., Chen, X., Chen, T., Yu, G., Fan, J.: A large-scale outdoor multi-modal dataset and benchmark for novel view synthesis and implicit scene reconstruction. arXiv preprint arXiv:2301.06782 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.786, + 0.383 + ], + "angle": 0, + "content": "25. Meshry, M., Goldman, D.B., Khamis, S., Hoppe, H., Pandey, R., Snavely, N., Martin-Brualla, R.: Neural rerendering in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6878-6887 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.786, + 0.424 + ], + "angle": 0, + "content": "26. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: European Conference on Computer Vision. pp. 405-421 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.786, + 0.452 + ], + "angle": 0, + "content": "27. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.786, + 0.493 + ], + "angle": 0, + "content": "28. Pfister, H., Zwicker, M., Van Baar, J., Gross, M.: Surfels: Surface elements as rendering primitives. In: Proceedings of the 27th annual conference on Computer graphics and interactive techniques. pp. 335-342 (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.786, + 0.535 + ], + "angle": 0, + "content": "29. Pittaluga, F., Koppal, S.J., Kang, S.B., Sinha, S.N.: Revealing scenes by inverting structure from motion reconstructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 145-154 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.786, + 0.577 + ], + "angle": 0, + "content": "30. Rakhimov, R., Ardelean, A.T., Lempitsky, V., Burnaev, E.: Npbg++: Accelerating neural point-based graphics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15969-15979 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.786, + 0.618 + ], + "angle": 0, + "content": "31. Ramon, E., Triginer, G., Escur, J., Pumarola, A., Garcia, J., Giro-i Nieto, X., Moreno-Noguer, F.: H3d-net: Few-shot high-fidelity 3d head reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5620-5629 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.786, + 0.674 + ], + "angle": 0, + "content": "32. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.786, + 0.701 + ], + "angle": 0, + "content": "33. Ruckert, D., Franke, L., Stamminger, M.: Adop: Approximate differentiable one-pixel point rendering. ACM Transactions on Graphics (ToG) 41(4), 1-14 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.786, + 0.729 + ], + "angle": 0, + "content": "34. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4104-4113 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.787, + 0.756 + ], + "angle": 0, + "content": "35. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.786, + 0.798 + ], + "angle": 0, + "content": "36. Sun, C., Sun, M., Chen, H.T.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5459-5469 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "37. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., Rabinovich, A.: Going deeper with convolutions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1-9 (2015)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.346, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "RPGB: Towards Robust Neural Point-based Graphics in the Wild" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "38. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8248–8258 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.787, + 0.247 + ], + "angle": 0, + "content": "39. Tancik, M., Weber, E., Ng, E., Li, R., Yi, B., Wang, T., Kristoffersen, A., Austin, J., Salahi, K., Ahuja, A., et al.: Nerfstudio: A modular framework for neural radiance field development. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-12 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.248, + 0.787, + 0.288 + ], + "angle": 0, + "content": "40. Tewari, A., Fried, O., Thies, J., Sitzmann, V., Lombardi, S., Sunkavalli, K., Martin-Brualla, R., Simon, T., Saragih, J., Nießner, M., et al.: State of the art on neural rendering. In: Computer Graphics Forum. vol. 39, pp. 701-727. Wiley Online Library (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.29, + 0.787, + 0.317 + ], + "angle": 0, + "content": "41. Thies, J., Zollhöfer, M., Nießner, M.: Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG) 38(4), 1-12 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.787, + 0.359 + ], + "angle": 0, + "content": "42. Turki, H., Ramanan, D., Satyanarayanan, M.: Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12922–12931 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.361, + 0.787, + 0.402 + ], + "angle": 0, + "content": "43. Wang, F., Galliani, S., Vogel, C., Speciale, P., Pollefeys, M.: Patchmatchnet: Learned multiview patchmatch stereo. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14194-14203 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.403, + 0.787, + 0.445 + ], + "angle": 0, + "content": "44. Wang, P., Liu, Y., Chen, Z., Liu, L., Liu, Z., Komura, T., Theobalt, C., Wang, W.: F2-nerf: Fast neural radiance field training with free camera trajectories. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4150–4159 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.446, + 0.787, + 0.487 + ], + "angle": 0, + "content": "45. Wang, X., Zhu, Z., Huang, G., Qin, F., Ye, Y., He, Y., Chi, X., Wang, X.: Mvster: Epipolar transformer for efficient multi-view stereo. In: European Conference on Computer Vision. pp. 573-591. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.529 + ], + "angle": 0, + "content": "46. Wei, Z., Zhu, Q., Min, C., Chen, Y., Wang, G.: Aa-rmvvsnet: Adaptive aggregation recurrent multi-view stereo network. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6187-6196 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.531, + 0.787, + 0.558 + ], + "angle": 0, + "content": "47. Wu, M., Wang, Y., Hu, Q., Yu, J.: Multi-view neural human rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1682-1691 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.787, + 0.599 + ], + "angle": 0, + "content": "48. Yao, Y., Luo, Z., Li, S., Fang, T., Quan, L.: Mvsnet: Depth inference for unstructured multiview stereo. In: Proceedings of the European conference on computer vision (ECCV). pp. 767-783 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.787, + 0.642 + ], + "angle": 0, + "content": "49. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12-22 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.643, + 0.787, + 0.685 + ], + "angle": 0, + "content": "50. Yu, J., Lin, Z., Yang, J., Shen, X., Lu, X., Huang, T.S.: Free-form image inpainting with gated convolution. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4471-4480 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.787, + 0.728 + ], + "angle": 0, + "content": "51. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5728-5739 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.787, + 0.769 + ], + "angle": 0, + "content": "52. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14821-14831 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.77, + 0.787, + 0.812 + ], + "angle": 0, + "content": "53. Zhang, J., Zhang, J., Mao, S., Ji, M., Wang, G., Chen, Z., Zhang, T., Yuan, X., Dai, Q., Fang, L.: Gigamvs: a benchmark for ultra-large-scale gigapixel-level 3d reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(11), 7534-7550 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "54. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.115, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.129 + ], + "angle": 0, + "content": "Q. Zhu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "55. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.218 + ], + "angle": 0, + "content": "56. Zhang, Y., Peng, S., Moazeni, A., Li, K.: Papr: Proximity attention point rendering. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.218 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/RPBG_ Towards Robust Neural Point-based Graphics in the Wild/ced6023f-74f5-424c-b6a1-715ff57d9512_origin.pdf b/2024/RPBG_ Towards Robust Neural Point-based Graphics in the Wild/ced6023f-74f5-424c-b6a1-715ff57d9512_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..68d6a1cdee1c1746d568b8e2c0d5aad64b83f842 --- /dev/null +++ b/2024/RPBG_ Towards Robust Neural Point-based Graphics in the Wild/ced6023f-74f5-424c-b6a1-715ff57d9512_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a7fe7fe5de3ed632f1ac826f382fafdbcdc4a78dda940a1fd0b99d1d577647f2 +size 14383399 diff --git a/2024/RPBG_ Towards Robust Neural Point-based Graphics in the Wild/full.md b/2024/RPBG_ Towards Robust Neural Point-based Graphics in the Wild/full.md new file mode 100644 index 0000000000000000000000000000000000000000..08b8af955bb4a24ca3e839a32508e618cfb0e09f --- /dev/null +++ b/2024/RPBG_ Towards Robust Neural Point-based Graphics in the Wild/full.md @@ -0,0 +1,309 @@ +# RPBG: Towards Robust Neural Point-based Graphics in the Wild + +Qingtian Zhu $^{1\star}$ , Zizhuang Wei $^{2,3}$ , Zhongtian Zheng $^{3}$ , Yifan Zhan $^{1}$ , Zhuyu Yao $^{4}$ , Jiawang Zhang $^{4}$ , Kejian Wu $^{4}$ , and Yinqiang Zheng $^{1\boxtimes}$ + +1 The University of Tokyo + +$^{2}$ Huawei Technologies + +3 Peking University + +4 XREAL + +Abstract. Point-based representations have recently gained popularity in novel view synthesis, for their unique advantages, e.g., intuitive geometric representation, simple manipulation, and faster convergence. However, based on our observation, these point-based neural re-rendering methods are only expected to perform well under ideal conditions and suffer from noisy, patchy points and unbounded scenes, which are challenging to handle but defacto common in real applications. To this end, we revisit one such influential method, known as Neural Point-based Graphics (NPBG), as our baseline, and propose Robust Point-based Graphics (RPBG). We in-depth analyze the factors that prevent NPBG from achieving satisfactory renderings on generic datasets, and accordingly reform the pipeline to make it more robust to varying datasets in-the-wild. Inspired by the practices in image restoration, we greatly enhance the neural renderer to enable the attention-based correction of point visibility and the inpainting of incomplete rasterization, with only acceptable overheads. We also seek for a simple and lightweight alternative for environment modeling and an iterative method to alleviate the problem of poor geometry. By thorough evaluation on a wide range of datasets with different shooting conditions and camera trajectories, RPBG stably outperforms the baseline by a large margin, and exhibits its great robustness over state-of-the-art NeRF-based variants. Code available at https://github.com/QT-Zhu/RPBG. + +Keywords: Point-based graphics $\cdot$ Novel view synthesis $\cdot$ Neural rendering + +# 1 Introduction + +Novel view synthesis (NVS) aims to synthesize novel views under given camera poses by aid of a series of already posed images, which is a fundamental task in both computer vision and computer graphics and has been studied for decades. + +Among the feasible solutions, NeRF (Neural Radiance Field) [26] approximates an implicit scene representation, i.e., a radiance field (RF), encoded by a neural network mapping 3D coordinates and view directions to colors and densities. When rendering, the fitted RF is queried multiple times along a ray to volume-render the corresponding pixel for a novel view. Subsequent variants of NeRF follow a case-by-case design and + +![](images/b7ede94768135d06e5c41278c4cff3fbbf8927a1bd531dbd9bf20921066c4f63.jpg) +Fig. 1: Left: RPBG manages to achieve all-round good re-renderings (PSNR plotted) across generic datasets over the baseline [1] as well as state-of-the-art RF-based methods [42,44]. Right: We demonstrate the point clouds (with camera trajectories visualized) and corresponding re-rendered novel views of the representative scenes, revealing the great robustness and scalability of RPBG. Zoom in for best view. + +![](images/ec80b35a2520ba64f99530c13415086a3af529c84e449e0ffd1110e96a566457.jpg) + +adopt different parameterization techniques specifically for different scene types, e.g., object-centric [3, 54], free-trajectory [44], and large-scale scenes [38, 42]. Therefore, putting forward one robust method that can work stably across varying scene types with a unified parameterization technique is considered as challenging. + +Recently, point-based alternatives [1, 18, 30, 33] have gained substantial attention for their unique advantages over implicit representations, e.g., ease of manipulation and faster training. The practice of employing point-based parameterization for rendering can be traced all the way back to [16, 20], and is intuitive especially under the context of NVs. Analogue to NeRF and its variants (collectively referred to as RF-based methods), whose optimization and convergence heavily rely on the inherent co-visibility among images, the 3D points triangulated from a series of 2D images according to epipolar geometry have already contain all the verified co-visibility information. In this way, by enforcing the triangulated points as strong prior knowledge for NVs, the expected freedom of optimization is greatly constrained, leading to faster training as well as better robustness. Recent attempts have achieved promising results in terms of fast optimization and rendering [18, 30] and accurate geometric representation [56]. Concretely, we focus on the specific category of methods that manage to integrate CNN-based neural renderers to yield re-renderings with featuremetric neural buffers [22, 30], for their concise pipeline and the potential capability of obtaining visually pleasing re-renderings. + +However, by evaluating the representative baseline, NPBG (Neural Point-based Graphics) [1], with generic datasets, we realize that the original design only intends for ideal conditions, e.g., synthetic data [26] and well-captured human heads [31]. When synthesizing novel views under more general conditions, the performance of NPBG degrades severely. In this paper, we strive to boost the robustness of NPBG-like point-based neural re-rendering pipelines and reveal the true potential to achieve state-of-the-art performance across varying datasets in-the-wild, by analyzing the reasons for performance degradation and seeking for remedies. + +Generally speaking, the major difficulties that the vanilla NPBG is faced up with include handling the background, handling patchy point clouds, and identifying correct + +point visibility. We reform the CNN-based neural renderer, with an inspiration from image restoration algorithms [6,8,52] that are able to identify downgraded patterns and restore the corresponding high-quality images. To make sure the neural renderer can capture sufficient and valid context information with extremely sparse rasterizations, we particularly leverages a Downgrade-aware Convolution (DAC) module to determine the correct point visibility with regard to a given camera pose and performs a pseudo point-wise back-face culling operation with visual self-attention. The background is also modeled in a lightweight manner. Instead of incorporating a massive environment map [33], a simple default trainable feature vector can reach similar quantitative results when working with the stronger neural renderer. We also discover the pseudo density calculated from the neural textures can roughly verify the existence of a given 3D position, which can be used to augment the poorly triangulated point clouds. In addition, we also simplify the phased training paradigm in [1,30] and optimize the parameters of both the neural textures and the renderer end-to-end collaboratively. We term our version of point-based re-rendering as RPBG (Robust Point-based Graphics), to emphasize its great robustness across different generic datasets. + +For thorough qualitative and quantitative evaluation, we collect 4 typical challenging scene types as the benchmark for robust NVs, i.e., $360^{\circ}$ unbounded scenes (with free trajectories), inside-out scenes, large-scale scenes (at the scale of a block or campus), and sparse-view scenes. As a result, RPBG exhibits great robustness with perceptually satisfactory synthesis across the aforementioned typical scenes types as showcased in Fig. 1, where the high-quality re-renderings are obtained with an exactly identical parameterization strategy without any manual configuration, relieving the practitioners from the exhaustive per-scene search of hyper-parameters. Its stable superiority over state-of-the-art NVS methods, we believe, is of great significance especially for real applications. + +The main contributions are three-fold as follows: + +- We put forward RPBG, as a more robust and practical alternative for re-rendering high-quality images from triangulated 3D points. +- According to our in-depth analysis, we enhance the neural re-rendering pipeline regarding the neural renderer, environment modeling, point cloud augmentation, and the training paradigm. +- RPBG manages to greatly boost the performance of neural point re-renderer by a large margin, and exhibits stably greater robustness and generalizability over RF-based methods with even better perceptual rendering quality. + +# 2 Related Work + +Radiance Fields for NVs Along with the proposed NeRF [26], the scene representation of RF is becoming popular for its ease of optimization by the differentiable volume rendering. In a RF, each position is assigned with an anisotropic color and a density, and to render a pixel of a novel view, one needs to sample the trained field and conduct ray marching. The original NeRF [26], together with its variants [2,3,39,54], employs an MLP to represent the functional mapping from coordinates and view directions to the radiance values. Several attempts have been made to encode RFs with different + +data structures for faster training and inference, e.g., grid voxels [12, 36], decomposed tensors [5], hash grids [27]. + +In practice, the parameterization strategies have a great impact on the rendering quality when reconstructing different types of scenes, e.g., forward-facing ones, $360^{\circ}$ unbounded ones, and large-scale ones. For unbounded scenes, $\mathrm{NeRF}++$ [54] applies separate networks and different parameterization to model near and distant objects; mip-NeRF 360 [3] designs a smooth contraction operator to parameterize the whole unbounded scene into a ball; Mega-NeRF [42] and Block-NeRF [38] partition 3D scenes explicitly and use different networks to represent each scene partition; $\mathrm{F}^2$ -NeRF [44] proposes to use perspective warping to handle sequential data with arbitrary trajectories. The varying case-by-case parameterization strategies of RF-based NVS greatly constraint the generalizability of methods. + +Point-based Graphics A point cloud is a collection of 3D coordinates that is usually used as a topology-agnostic coarse shape representation of the 3D geometry and favored for its flexibility and sparsity for storage. The development of techniques of employing points as modeling primitives for rendering (referred as point-based graphics [15]) can be traced back to [16, 20], the best practice of which is to replace each point with an oriented circular disk (a surfel) and reply on splatting to blend overlapping surfels [28]. + +In contrast to conventional physically based rendering (PBR) techniques, neural rendering [40] learns to render high-quality images in a data-driven manner. Particularly, we focus on the methods that conduct neural re-rendering with point clouds. Bui et al. [4] propose to enhance the coarse point-based rendering by a GAN for image super-resolution. NRW [25] and InvSFM [29] attempt to re-render the reconstructed point cloud with respective auxiliary buffers, e.g., latent appearance vectors, semantic masks and SIFT descriptors [23]. NPBG [1,30] adopts a U-Net-like [32] CNN to render neural point textures as RGB images, and exhibits better flexibility over mesh-based proxies [41]. The practical applications in the context of autonomous driving, e.g., scene editing and stitching, and large-scale training, are further explored in READ [21]. Recently, ADOP [33] and Gaussian Splatting [18] have demonstrated notable accomplishments. However, the differentiable rendering/splatting scheme employed requires an extensive amount of memory to maintain the computational graph and gradients during training. This limitation hinders their application on large-scale scenes. + +# 3 Neural Point-based Graphics Revisited + +In this section, we revisit the pipeline of NPBG [1], analyze the existing drawbacks when attempting to re-render generic scenes, and attempt to explain why NPBG finds it difficult to handle the background and patchy points, and identify correct point visibility. We hope these discussions be of sufficient insights to support our modifications. + +# 3.1 Preliminaries + +Inputs As the common practice in NVS, the inputs are a series of images with respective camera parameters (both intrinsics and extrinsics), and specifically for point-based + +![](images/b352c23801bc9ce10c0ac8b1ae1517c38d8b08f403fe65e1e53863e8924256e6.jpg) +(a) Rasterized Points + +![](images/43a2af0e2b26b0e374be61b1f1495cb66efdfeb3a427a7716d3bcfdf9846764a.jpg) +(b) Ground Truth + +![](images/c86057279b26509d122c492de35d5788f1fa7e38e97ecdd7482bf98ddb9e2282.jpg) +(c) NPBG +Fig. 2: Typically challenging scenes in T&T dataset [19] for NPBG. Top: Auditorium, where the walls and ceilings are extremely sparse. Bottom: Museum, where the point sparsity makes the occlusion and visibility complicated. + +![](images/689cc3a0243e40edf745f0ea06e9b1930cebdee309b5611b5468e5a33add5c3c.jpg) +(d) Ours + +methods [1, 18, 30], the very first step is supposed to be the triangulation of 3D points from 2D observations. The original NPBG-like pipelines [1, 30] mainly consider this as a pre-processing without much attention. + +Point Rasterization NPBG [1] and NPBG++ [30] apply a non-differentiable hard point $z$ -buffering operation as an approximated back-face culling when rasterizing points as 2D fragments, where the fragment is updated when and only when the newly projected point has a smaller $z$ -depth than the current one. After a traversal of all the points, the fragment keeps the record of the indices of rasterized points. Then tensor scattering is performed to index the neural texture at the corresponding positions. Compared to differentiable rendering/splatting alternatives [18, 33], we consider the most significant advantage is its memory efficiency and thus better scalability, since the computation graph required to keep is much smaller for NPBG. We also want to keep this scalability without interfering the elegant rasterization paradigm. + +Neural Renderer The renderer $\mathcal{R}_{\Theta}:\mathbb{R}^{H\times W\times C}\to \mathbb{R}^{H\times W\times 3}$ of NPBG [1] follows a vanilla U-Net-like architecture [32]. Since the target point clouds for rendering are supposed to be well-constructed, the vanilla U-Net is expressive enough to complete the assigned task of mapping higher dimensional features from the neural texture T to RGB values. Empirically, we find the expected properties of $\mathcal{R}_{\Theta}$ have much in common with a network for low-level vision tasks, e.g., image restoration. + +# 3.2 Problems + +Patchy Triangulation Consider the inputs, for a NVS system aiming to synthesize images for generic scenes, the triangulation of points is not trivial. The point triangulation step to perform is similar to the step of multi-view stereo (MVS) reconstruction [14, 48], the key of which is to identify co-visible pixels across images and lift the 2D pixels to a 3D points according to epipolar geometry. Such algorithms suffer from non-Lambertian surfaces and textureless regions, both of which are very common and lead to poor triangulated and patchy points. We attach two typical scenes that NPBG fails to yield good renderings in Fig. 2. + +Wrong Point Visibility As can be imagined, such point visibility can be erroneous due to the poor quality of triangulated points, and also due to the inherent sparsity of points as a 3D representation. In this way, the points belonging to back faces, which should be considered as occluded, are rasterized as the fragment for further processing. Manually setting a depth threshold for bound check could not be helpful either, since being far from the camera does not necessarily indicate they should be occluded — they could be parts of the environment. While NHR [47] proposes to take the depth buffer as an additional rendering condition, we find its effectiveness not significant for generic scenes other than well-masked human captures. + +Lack of Context for Re-rendering Similarly, when re-rendering an incomplete point cloud, where the neural buffers are usually with significant flaws, the whole receptive field of a kernel at a given position may only capture the downgraded regions, leading to failure to yield reasonable restoration, especially for a high target resolution. It is worth noting that, though the relationship between the number of network layers and rendering quality has been discussed in [30, 33], there is still huge room for the improvement of the renderer. + +Failure to Effectively Model the Environment Unbounded scenes are a typical challenging scene type for NVS. RF-based methods that are designed on purpose to handle such scenes [3, 54] typically employ different parameterization to encode the areas outside a certain bounding sphere. For point-based methods, ADOP [33] proposes to apply an environment map of $H \times W \times C$ to model the environment, which is equivalent to wrapping the triangulated points with a sphere with $H \times W$ points, resulting in more than $5 \times 10^{5}$ points as overheads, according to the default configuration. + +Summary To summarize, we attribute the observed problems to two main causes: poor geometry and weak, local renderer. The two causes are to some extent coupled for a better geometry will relieve the difficulty of the renderer and vice versa. + +# 4 Robust Point-based Graphics + +We are strongly convinced that the pipeline of point-based neural re-rendering has the great potential to outperform popular RF-based solutions, for its unified point-as-parameterization fashion and the incorporation of neural networks to yield visual details. Therefore in this section, we introduce Robust Point-based Graphics (RPBG), as an enhanced version of NPBG, with a particular focus on robustness under generic scenes. We will elaborate the modifications made and shed light on the underlying insights and motivation. The overall pipeline of RPBG as well as the training paradigm is briefly illustrated in Fig. 3. + +# 4.1 Downgrade-aware Neural Renderer + +As is attributed as one key problem in NPBG, the U-Net-based neural renderer is considered as too naive to handle the challenging situations in generic scenes. The + +![](images/6b8cd81a3ed5ee1a60464a343d2dc2eda618897d578f7552b643a44052342850.jpg) +Fig. 3: The overall pipeline of RPBG. Point Triangulation: We first triangulate a 3D proxy for re-rendering with posed images, with its geometry-bounded neural texture initialized. Point Rasterization: The points are raterized to the given camera in a non-differentiable manner. By indexing the texture with the fragment, we obtain the neural buffer. A learnable point-size neural texture $\mathbf{T}_{\mathrm{env}}$ is also optimized. Neural Rendering: The restoration from downgraded neural buffer to photo-realistic images is performed by a CNN. The network and the neural texture are optimized end-to-end by image-level losses. An offline point cloud augmentation strategy is introduced to alleviate the problem of patchy triangulation under challenging conditions. + +modifications made are deeply inspired by low-level vision tasks, where networks can adaptively determine downgraded parts, e.g., the deblurring/deraining networks are able to identify the blurring/raining pixels from the whole image, and correct them accordingly. We expect the neural renderer $\mathcal{R}_{\Theta}$ can benefit from relevant restoration-targeting techniques, and become able to decode high-quality visual information from patchy buffers. + +After evaluating three state-of-the-art fundamental architectures for image restoration, i.e., multi-scale fusion [8], multi-stage [52], and U-Net [6], following the taxonomy in [6]. According to the reported experiments in Sec. 5.4, we opt the paradigm in [8] for it achieves the best balance between performance and time/memory efficiency. However, the convolution layers in such paradigm is still with a fixed receptive field, lacking in robustness against point sparsity. We would like to further enlarge the receptive field to a global scale and attempt to explicitly model the visual attention to weigh the observed points. + +Transformer-based architectures for image restoration, e.g., Restormer [51], brings unacceptable memory overheads, so we rely on the frequency-domain alternative, Fast Fourier Convolution (FFC) [7], which can theoretically capture global contexts, to determine the correct point visibility adaptively. FFC performs channel-wise real 2D FFT (Fast Fourier Transform) and inverse real 2D FFT on 2D tensors. Real FFT uses only half of the spectrum and by convolving the transformed frequency-domain tensors. In this way, a receptive field covering the entire image is considered. + +![](images/424413adaf0946be3980caff4e0c2453f8cdc98668babe46c712c681ab7db2dc.jpg) +Fig. 4: The architecture of the downgrade-aware neural renderer in RPBG, with some conventional modules omitted. From the visualized attention map, DAC manages to adaptively handle the severely erroneous point visibility. + +Inspired by [37], we apply FFT (as the global branch) in parallel to conventional convolution layers (as the local branch) and rely on the fused features of both branches to determine the downgraded regions for the gated convolution [50] to filter. Based on the common practice in image inpainting, we leverage gated convolutions at the early stage of the renderer, to help locate the downgrade by wrong point visibility. We name such customized gated convolution module as the Downgrade-aware Convolution (DAC) module, and as can be inferred from the point attention maps before and after DAC in Fig. 4, DAC manages to determine complicated point visibility with patchy triangulation. + +# 4.2 Point Triangulation + +Though there are advances in MVS empowered by deep learning [43, 45, 46], recovering a complete, accurate, and dense point cloud for certain regions remains a challenging. For RPBG, we adopt an off-the-shelf MVS method [46] to perform point triangulation for its memory efficiency to allow large-batch inference. + +As for problem the poor triangulation, we partially leave it to the neural renderer $\mathcal{R}_{\Theta}$ , relying on a stronger renderer to recover the re-rendering from downgraded buffers. Inspired by the point cloud extraction method of RF-based implementations [39], where the estimated radiance density $\sigma$ can be a rough indicator for surface, we propose a point cloud augmentation technique to densify the initial triangulation. + +The point cloud augmentation follows a trial-and-error paradigm by first assuming the existence of newly sampled points and then verify them by estimated pseudo densities $\sigma_{i} = \sum |\mathbf{T}_{i}|$ . It is empirically observed that the absolute activation of the point-wise neural texture can roughly represent the reliability of the 3D position — it is reasonable since an outlier will be less distinguishable by the neural renderer and thus less visual attention will be given. + +Note that this strategy is optional, and we only apply such strategy to scenes with extremely poor triangulation. Please refer to the Supplementary Material for more discussions regarding the self-pruning. + +# 4.3 Environment Modeling + +As a result of the incorporation of a stronger neural renderer, we find that given a neural texture with $C = 8$ , the dense environment map suggested in [33] is redundant. Instead, we shrink the overheads for environment modeling from $H \times W \times C$ to $1 \times C$ , and relying on the stronger neural renderer to decode the background. + +When rasterizing, we employ a tunable feature vector $\mathbf{T}_{\mathrm{env}}$ aside the neural point texture $\mathbf{T}$ , as the default value for vacant pixels in the fragment, which is also involved in the end-to-end training. By experiments in Sec. 5.4, we demonstrate that the lightweight modeling strategy reaches the quantitative performance equivalent to applying an environment map, yet with a negligible overhead. + +# 4.4 Collaborative Optimization + +Recall that the point rasterization procedure is completely parameter-free. The overall collaborative optimization scheme can be thus formulated as + +$$ +\mathbf {T} ^ {*}, \Theta^ {*} = \underset {\mathbf {T}, \Theta} {\arg \min } \sum_ {k} \mathcal {L} (\mathbf {I} _ {k}, \hat {\mathbf {I}} _ {k}) = \underset {\mathbf {T}, \Theta} {\arg \min } \sum_ {k} \mathcal {L} (\mathbf {I} _ {k}, \mathbf {K}, \mathbf {R} _ {k}, \mathbf {t} _ {k} | \mathbf {T}, \mathcal {R} _ {\Theta}), \tag {1} +$$ + +where $\mathbf{T}$ and $\Theta$ stand for the tunable parameters, i.e., the neural point texture (alongside the globally shared $\mathbf{T}_{\mathrm{env}}$ ) and the parameters of the renderer $\mathcal{R}_{\Theta}$ for neural re-rendering, while $\mathbf{I}_k$ is the target image whose calibration is $\mathbf{K}$ and $[\mathbf{R}_k|\mathbf{t}_k]$ and $\hat{\mathbf{I}}_k$ is the re-rendering + +Note that, we also discard all the typical tossing sampling [21] and optimization steps [30, 33] of point-based neural re-rendering, and manage to tune all the parameters involved in a simple but effective collaborative end-to-end way. The neural texture is initialized with all zeros while the rendering CNN is trained from scratch for each scene + +Loss Function Since the rendering scheme in both NPBG and RPBG is convolutional, where images are rendered in patches, we are able to apply patch-aware losses to enforce the involvement of neighboring pixels to ensure patch-to-patch consistency, offering perceptually good renderings. To this end, in addition to the pixel-wise Huber norm $\mathcal{L}_{\mathrm{huber}}$ providing the basic supervision and numerical stability, we apply two patch-aware losses to the collaborative optimization of RPBG, namely the perceptual VGG loss [10, 17] $\mathcal{L}_{\mathrm{vgg}}$ , and the FFT loss [8, 13] $\mathcal{L}_{\mathrm{fft}}$ . VGG loss compares the rendered image and the target ground-truth image in a high-dimensional feature space by a pre-trained VGG-19 + +network [35]. It reveals the perceptual similarity between images that pixel-wise metrics fail to measure. FFT loss measures the image-to-image distance in the frequency domain by carrying out 2D FFT towards images. The frequency components are considered to be crucial in terms of the perceptual quality [13]. + +The final loss function applied in RPBG is composed as + +$$ +\mathcal {L} = \lambda_ {\text {h u b e r}} \mathcal {L} _ {\text {h u b e r}} + \lambda_ {\text {v g g}} \mathcal {L} _ {\text {v g g}} + \lambda_ {\text {f f t}} \mathcal {L} _ {\text {f f t}}. \tag {2} +$$ + +# 5 Experiments + +# 5.1 Datasets + +We conduct experiments covering a diverse range of scene types to quantitatively examine the generalizability and robustness of NVs methods, including the Free dataset [44] (7 free-trajectory scans with long camera trajectories), Tanks and Temples (T&T) dataset [19] (4 unbounded scans and 5 inside-out scans, without salient dynamic objects), Mill19 dataset [42] (2 large-scale aerial scans with over 1600 images in each scan), and GigaMVS dataset [53] (8 sparse-view outdoor scans). Note that for T&T, we use the raw undistorted images provided by the benchmark without any masking [22]. We follow the common split protocol, that 1 frame out of every 8 frames is evaluated, for the Free dataset, T&T dataset, and GigaMVS dataset. For Mill19 dataset, we apply the officially recommended split to align with the experiments in [42]. + +We also include some other challenging datasets for a thorough evaluation, e.g., ScanNet++ dataset [49], a challenging indoor benchmark, OMMO dataset [24], a multi-modal aerial NVS dataset and ETH-MS dataset [11], a super-large-scale dataset (around $5k$ images). Please refer to the Supplementary Material for more quantitative and qualitative results. + +# 5.2 Implementation Details + +Data Preparation To obtain the per-scene point cloud representation for re-rendering, we follow the mapping procedure of COLMAP [34] for sparse triangulation and the MVS reconstruction with a trained network [46] for dense triangulation. Note that the reconstruction only takes a small proportion of time over the whole training phase and we conduct ablation experiments to show that RPBG is robust against different triangulation configurations. Please refer to Sec. 5.4 as well as the Supplementary Material for more information. + +Training Settings We randomly crop images to square patches of $256 \times 256$ , with a batch size of 8. The learning rate for the neural point textures is $10^{-1}$ , and $10^{-4}$ for the rendering networks, which will decay by a factor of 0.5 if 5 consequent epochs witness no drop in the loss function. A dimension of 8 is applied for any neural texture regardless of the scene scale or complexity. For the weights of loss functions, we globally set $\lambda_{\mathrm{huber}} = 10^3$ , $\lambda_{\mathrm{vgg}} = 1$ , and $\lambda_{\mathrm{fft}} = 1$ . The training is performed on one NVIDIA GeForce RTX 3090, with a GPU memory consumption of up to 23 GB. It takes around 8 to 30 GPU hours for training, depending on the data scale. We would like to emphasize that, RPBG does not require case-by-case scene parameterization or grid search of training hyper-parameters for all the scenes covered in the experiments. + +![](images/382384e456bc329ff256bb9ad09ac248cec17f078676bb0f4c5e63dc006910b2.jpg) +Fig. 5: Visualized comparisons over varying scenes. From top to bottom: sky and hydrant of the Free dataset [44], Courtyard and Train of T&T dataset [19], and DayaTemple and MemorialHall of GigaMVS dataset [53]. We include the results of RPBG (Ours), Gaussian Splatting [18], NPBG [1], and $\mathrm{F}^2$ -NeRF [44] for comparison. Zoom in for best view. + +Evaluation Metrics We adopt the metrics of PSNR, SSIM and LPIPS (VGG) [55] for the evaluation between the synthesized and the target images. According to [55], PSNR does not faithfully measure image sharpness and so cannot properly account for the nuances of human visual perception. + +# 5.3 Results + +In Fig. 5, we demonstrate several typical groups of visual comparisons of NVS results by RPBG, two representative point-based methods, i.e., Gaussian Splitting [18] and NPBG [1], and a state-of-the-art NeRF variant for unbounded scenes, $\mathrm{F}^2$ -NeRF [44]. We group the datasets by category and report the corresponding scene-averaged quantitative scores in Tab. 1. Note that RPBG achieves the best SSIM and LPIPS, which are more relevant to the high-frequency components of images, across all datasets. + +Free-trajectory/Unbounded Scenes Compared with NVS methods designed in particular for unbounded scenes, $\mathrm{NeRF}++$ [54] and $\mathrm{F^2}$ -NeRF [44], RPBG achieves the best SSIM and LPIPS and comparable PSNR. For visual effects showcased in Fig. 5, since RPBG involves a neighborhood for rendering, its results appear more visually harmonious, especially than Gaussian Splatting [18] and $\mathrm{F^2}$ -NeRF [44]. It is necessary to mention + +Table 1: Quantitative evaluation of state-of-the-art NVS methods [1,5,18,26,30,42,44,54] and RPBG across diverse scenes grouped by category, including free trajectory/unbounded scenes, inside-out scenes, large-scale scenes, and sparse-view scenes. The figure following the dataset name stands for the number of scenes the dataset contains. + +
MethodFree-trajectory/UnboundedInside-out T&T-5Large-scale Mill19-2Sparse-view GigaMVS-8
Free Dataset-7T&T-4
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NeRF [26]17.750.4050.59716.840.3960.73119.820.6150.52620.340.5240.52918.290.6320.499
TensoRF [5]21.740.5490.60020.730.6430.56619.430.6340.57020.070.4970.59718.700.6530.506
NeRF++ [54]23.470.6030.49921.660.6580.52919.250.6100.58520.190.5200.53118.380.6320.495
F2-NeRF [44]26.320.7790.27623.660.7640.30320.120.7060.394N/AN/AN/A17.440.5400.470
Mega-NeRF [42]22.600.5700.56218.730.5780.47819.160.6170.45122.490.5500.51018.250.5810.394
NPBG [1]21.400.6390.34019.850.6980.37620.570.6960.37116.210.3570.64418.340.6200.405
NPBG++ [30]20.060.5920.44517.230.6530.47418.300.6840.41117.040.4000.64819.300.6630.443
Gaussian Splatting [18]25.230.7400.29023.510.7820.29323.460.7830.277N/AN/AN/A16.840.5300.391
Ours26.330.8320.17722.500.7820.27623.290.8040.24222.620.5960.36820.540.6860.317
+ +that, for RF-based methods [5, 26, 42, 44], we have manually adjusted the scene-specific hyper-parameters to achieve better results. + +Inside-out Scenes The inside-out indoor scenes are strictly bounded, but lack high-quality ray intersections required for optimization, which will lead to under-fitting of the RF (Courtroom in Fig. 5). Similar to free-trajectory/unbounded scenes, RPBG outperforms state-of-the-art methods in SSIM and LPIPS. + +Large-scale Scenes We mainly evaluate RPBG against Mega-NeRF [42] on the massive imagery of Mill19 dataset [42]. Our method outperforms Mega-NeRF [42] in every metric. It is worth noting that Mega-NeRF takes 240 GPU hours for training while RPBG only takes 29 GPU hours. Notably, our system with 256 GB RAM and RTX 3090 fails to afford the training of $\mathrm{F}^2$ -NeRF [44] and Gaussian Splitting [18] (marked as N/A in the table). In contrast, RPBG proves to be viable for large-scale scenes under the same hardware constraint. + +Sparse-view Scenes Sparse-view inputs are considered to be extremely challenging for NVS methods. Since RPBG, NPBG [1], and $\mathrm{NPBG + + }$ [30] incorporate a point-based 3D proxy for re-rendering, they showcase better robustness over RF-based methods. As revealed in Fig. 5, Gaussian Splatting [18] fails to regularize the strong approximation power of point-wise Gaussians at training views, and yields obvious needle-like artifacts at novel views. + +# 5.4 Ablation Study + +Environment Modeling We study different strategies of environment modeling for NVS in-the-wild, namely leaving the blank pixels with zeros, filling the blanks with a default learnable feature (as is done in RPBG), and + +![](images/4921670c624329e449c60f71af44f725a85dfa9c3d7fef56b2841a5d3a75c0cc.jpg) +(a) $w / o$ Env. Modeling + +![](images/0b807bd6eaabce4251f38cd77661aa716bc13bbd0f006ac17f09b808cfcd4cca.jpg) +(b) $w / \mathrm{Env}$ Modeling +Fig.6: The impact of leveraging the default feature vector for environment modeling in an unbounded scene. + +Table 2: Ablation on network designs on T&T dataset [19]. The reported inference time and memory consumption is tested with a target resolution of ${1920} \times {1080}$ . + +
MethodTime(s)Mem.(GB)PSNR↑SSIM↑LPIPS↓
Baseline1.035.4920.010.6660.390
+Multi-scale fusion1.6212.6721.880.7010.337
+DAC1.6713.0122.920.7690.320
+FFT loss--22.940.7940.257
+Multi-stage1.7717.1421.750.7130.338
+U-Net1.5810.3220.850.6920.364
+ +Table 3: Ablation on environment modeling on two unbounded scenes, namely Caterpillar of T&T dataset [19] and sky of Free dataset [44]. #Points $(\times 10^{6})$ represent the equivalent overheads when applying different strategies. + +
Env.Caterpillarsky
#PointsPSNR↑SSIM↑LPIPS↓#PointsPSNR↑SSIM↑LPIPS↓
Zeros7.9519.890.4560.40317.4921.770.5980.315
Learnable7.9521.780.6870.28917.4924.810.8660.199
Sphere8.9521.680.6860.28318.4925.150.8690.179
+ +![](images/a7d1c1353fe2b69baabb0ea37f7774c619a9c75768268a61c948a0bc1b7b0d09.jpg) + +![](images/760a399218dc190a30b2d8acffc7b95906cdb04f56139e0d7e4adcc6b1ca2ba3.jpg) +(c) $w / \theta$ DAC +Fig.7: A challenging case in Train, which is a zoomed-in view with complex occlusion and thin objects (e.g., the handrails). + +![](images/926f160a17c7c29ef5b47d8149db76a5d7294691b6471c62b216249e03995a58.jpg) + +![](images/4883f9a90237c97e80fcffc58f0920de95e37f1384ea51c16f0d3e0c7f17ab38.jpg) +(d) $w / \mathrm{DAC}$ +Fig. 8: Re-rendering quality (PSNR) with points of different quality on Church. RPBG manages to maintain its robustness against different levels of point sparsity with a $14\%$ drop in PSNR (NPBG [1]: $25\%$ ; NPBG++ [30]: $14\%$ ). + +![](images/cb069b7936944b5719e3c3ad478cc6f7c774a84f3db6cdd8334321284b607fc6.jpg) + +wrapping the points with a sphere of $10^{6}$ points (equivalent to an environment map in [33]). The quantitative results are in Sec. 5.4, where the strategy incorporated by RPBG boosts the performance by a large margin, with almost no additional overhead. + +Network Architecture and Loss We study the key components in the rendering network in Sec. 5.4. The baseline refers to NPBG [1] trained per-scene from scratch with $\mathcal{L}_1$ and VGG loss. We compare three fundamental architectures for image restoration (with necessary modifications on the first several layers), i.e., multi-scale fusion [8], multistage [52], and U-Net [6] following the taxonomy in [6]. Compared to the baseline, all modern restoration-oriented architectures bring remarkable improvement to the overall performance, and [8] is opted for a better balance between performance and time/memory efficiency. The gain by DAC is also considerable while the FFT loss mainly improves the perceptual quality. By the case in Fig. 7, we show the necessity of the DAC module when re-rendering challenging cases without in particular handling the erroneous point visibility. + +Triangulation Configurations We also evaluate RPBG with different triangulation configurations. Indicated by the results in Fig. 8, RPBG witnesses a total drop of $3.29\mathrm{dB}$ in re-rendering PSNR when switching the triangulation from ground truth (GT) to randomly initialized points (Random). Note that even with sparse points (SfM), + +which are usually yielded as a by-product when computing camera parameters, RPBG (21.41 dB) still outperforms $\mathrm{F}^2$ -NeRF (20.57 dB). + +# 6 Discussion + +Effectiveness of RPBG We would like to shed some light on the effectiveness of RPBG. The first factor should be the point representation. The triangulated 3D points have already contained all the verified co-visibility that is required for NVS. Compared to RF-based methods, which often struggle at local minima, such triangulation-as-parameterization paradigm greatly lower the complexity of searching, and is unified across different scene types. Besides, the patch-wise rendering scheme mentioned in Sec. 4.4 also improves the perceptual quality of re-rendering while RF-based methods apply pixel/ray-wise volume rendering, where the correlation between pixels are not modeled explicitly. For more discussions on the insights, we strongly refer the readers to the Supplementary Material. + +Limitations An obvious problem with point-based re-rendering (RPBG and NPBG [1]) is that compared to lightweight RF-based variants [5, 27], they take more space to store the CNN parameters and the neural texture. Although it makes the scaling-up easier, the number of parameters grows linearly with the point cloud scale. Besides, due to the non-local neural renderer we employ, each point is encoded with visual context of a larger range, hindering the editability of RPBG. We have also noticed that, the CNN-based rendering scheme can lead to unsatisfactory temporal consistency especially when the triangulation is patchy so that we mainly count on the inpainting ability of the neural render for yielding re-renderings. Such flicker issue, as well as its potential solutions, has been discussed in [9]. + +# 7 Conclusion + +In this paper, we present RPBG as a robust and practical alternative to NPBG, a baseline of point-based NVs methods, performing neural re-rendering on triangulated points. We analyze the key problems in NPBG, when attempting to generalize to more generic scenes other than the well-captured scans, and reform the pipeline to reveal the real potential of point-based graphics. Respectively motivated and inspired by RF-based methods and low-level image restoration methods, we reform the pipeline according to our analysis. By extensive experiments on diverse datasets, RPBG achieves stably superior results over state-of-the-art RF-based and point-based NVs methods, especially on the metrics with more attention paid to the nuances of human visual perception, without case-by-case parameterization across all scenes, indicating its robustness and generalizability. + +# Acknowledgments + +This research was supported in part by JSPS KAKENHI Grant Numbers 24K22318, 22H00529, 20H05951, and JST-Mirai Program JPMJMI23G1. + +# References + +1. Aliev, K.A., Sevastopolsky, A., Kolos, M., Ulyanov, D., Lempitsky, V.: Neural point-based graphics. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 696-712. Springer (2020) +2. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855–5864 (2021) +3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022) +4. Bui, G., Le, T., Morago, B., Duan, Y.: Point-based rendering enhancement via deep learning. The Visual Computer 34, 829-841 (2018) +5. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: European Conference on Computer Vision. pp. 333-350. Springer (2022) +6. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: European Conference on Computer Vision. pp. 17-33. Springer (2022) +7. Chi, L., Jiang, B., Mu, Y.: Fast fourier convolution. Advances in Neural Information Processing Systems 33, 4479-4488 (2020) +8. Cho, S.J., Ji, S.W., Hong, J.P., Jung, S.W., Ko, S.J.: Rethinking coarse-to-fine approach in single image deblurring. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4641-4650 (2021) +9. Dai, P., Zhang, Y., Li, Z., Liu, S., Zeng, B.: Neural point cloud rendering via multi-plane projection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7830-7839 (2020) +0. Dosovitskiy, A., Brox, T.: Generating images with perceptual similarity metrics based on deep networks. Advances in neural information processing systems 29 (2016) +11. ETH Zurich Computer Vision Group and Microsoft Mixed Reality & AI Lab Zurich: The ETH-Microsoft Localization Dataset. https://github.com/cvg/visloc-iccv2021 (2021) +12. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022) +13. Fuoli, D., Van Gool, L., Timofte, R.: Fourier space losses for efficient perceptual image super-resolution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2360-2369 (2021) +14. Furukawa, Y., Ponce, J.: Accurate, dense, and robust multiview stereopsis. IEEE transactions on pattern analysis and machine intelligence 32(8), 1362-1376 (2009) +15. Gross, M., Pfister, H.: Point-based graphics. Elsevier (2011) +16. Grossman, J.P., Dally, W.J.: Point sample rendering. In: Rendering Techniques' 98: Proceedings of the Eurographics Workshop in Vienna, Austria, June 29—July 1, 1998 9. pp. 181-192. Springer (1998) +17. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and superresolution. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 694-711. Springer (2016) + +18. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG) 42(4), 1-14 (2023) +19. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG) 36(4), 1-13 (2017) +20. Levoy, M., Whitted, T.: The use of points as a display primitive (1985) +21. Li, Z., Li, L., Zhu, J.: Read: Large-scale neural scene rendering for autonomous driving. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1522-1529 (2023) +22. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020) +23. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. International journal of computer vision 60, 91-110 (2004) +24. Lu, C., Yin, F., Chen, X., Chen, T., Yu, G., Fan, J.: A large-scale outdoor multi-modal dataset and benchmark for novel view synthesis and implicit scene reconstruction. arXiv preprint arXiv:2301.06782 (2023) +25. Meshry, M., Goldman, D.B., Khamis, S., Hoppe, H., Pandey, R., Snavely, N., Martin-Brualla, R.: Neural rerendering in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6878-6887 (2019) +26. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: European Conference on Computer Vision. pp. 405-421 (2020) +27. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022) +28. Pfister, H., Zwicker, M., Van Baar, J., Gross, M.: Surfels: Surface elements as rendering primitives. In: Proceedings of the 27th annual conference on Computer graphics and interactive techniques. pp. 335-342 (2000) +29. Pittaluga, F., Koppal, S.J., Kang, S.B., Sinha, S.N.: Revealing scenes by inverting structure from motion reconstructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 145-154 (2019) +30. Rakhimov, R., Ardelean, A.T., Lempitsky, V., Burnaev, E.: Npbg++: Accelerating neural point-based graphics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15969-15979 (2022) +31. Ramon, E., Triginer, G., Escur, J., Pumarola, A., Garcia, J., Giro-i Nieto, X., Moreno-Noguer, F.: H3d-net: Few-shot high-fidelity 3d head reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5620-5629 (2021) +32. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015) +33. Ruckert, D., Franke, L., Stamminger, M.: Adop: Approximate differentiable one-pixel point rendering. ACM Transactions on Graphics (ToG) 41(4), 1-14 (2022) +34. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4104-4113 (2016) +35. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014) +36. Sun, C., Sun, M., Chen, H.T.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5459-5469 (2022) +37. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., Rabinovich, A.: Going deeper with convolutions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1-9 (2015) + +38. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8248–8258 (2022) +39. Tancik, M., Weber, E., Ng, E., Li, R., Yi, B., Wang, T., Kristoffersen, A., Austin, J., Salahi, K., Ahuja, A., et al.: Nerfstudio: A modular framework for neural radiance field development. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-12 (2023) +40. Tewari, A., Fried, O., Thies, J., Sitzmann, V., Lombardi, S., Sunkavalli, K., Martin-Brualla, R., Simon, T., Saragih, J., Nießner, M., et al.: State of the art on neural rendering. In: Computer Graphics Forum. vol. 39, pp. 701-727. Wiley Online Library (2020) +41. Thies, J., Zollhöfer, M., Nießner, M.: Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG) 38(4), 1-12 (2019) +42. Turki, H., Ramanan, D., Satyanarayanan, M.: Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12922–12931 (2022) +43. Wang, F., Galliani, S., Vogel, C., Speciale, P., Pollefeys, M.: Patchmatchnet: Learned multiview patchmatch stereo. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14194-14203 (2021) +44. Wang, P., Liu, Y., Chen, Z., Liu, L., Liu, Z., Komura, T., Theobalt, C., Wang, W.: F2-nerf: Fast neural radiance field training with free camera trajectories. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4150–4159 (2023) +45. Wang, X., Zhu, Z., Huang, G., Qin, F., Ye, Y., He, Y., Chi, X., Wang, X.: Mvster: Epipolar transformer for efficient multi-view stereo. In: European Conference on Computer Vision. pp. 573-591. Springer (2022) +46. Wei, Z., Zhu, Q., Min, C., Chen, Y., Wang, G.: Aa-rmvvsnet: Adaptive aggregation recurrent multi-view stereo network. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6187-6196 (2021) +47. Wu, M., Wang, Y., Hu, Q., Yu, J.: Multi-view neural human rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1682-1691 (2020) +48. Yao, Y., Luo, Z., Li, S., Fang, T., Quan, L.: Mvsnet: Depth inference for unstructured multiview stereo. In: Proceedings of the European conference on computer vision (ECCV). pp. 767-783 (2018) +49. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12-22 (2023) +50. Yu, J., Lin, Z., Yang, J., Shen, X., Lu, X., Huang, T.S.: Free-form image inpainting with gated convolution. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4471-4480 (2019) +51. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5728-5739 (2022) +52. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14821-14831 (2021) +53. Zhang, J., Zhang, J., Mao, S., Ji, M., Wang, G., Chen, Z., Zhang, T., Yuan, X., Dai, Q., Fang, L.: Gigamvs: a benchmark for ultra-large-scale gigapixel-level 3d reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(11), 7534-7550 (2021) +54. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492 (2020) + +55. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018) +56. Zhang, Y., Peng, S., Moazeni, A., Li, K.: Papr: Proximity attention point rendering. Advances in Neural Information Processing Systems 36 (2024) \ No newline at end of file diff --git a/2024/RPBG_ Towards Robust Neural Point-based Graphics in the Wild/images.zip b/2024/RPBG_ Towards Robust Neural Point-based Graphics in the Wild/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..676942aa8301ef37cb55d544c84b05e765ba449a --- /dev/null +++ b/2024/RPBG_ Towards Robust Neural Point-based Graphics in the Wild/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fc33e85ce371e3570ec92bbd5ab4e2de7cf88adcc90b042dec1ab89813579273 +size 570150 diff --git a/2024/RPBG_ Towards Robust Neural Point-based Graphics in the Wild/layout.json b/2024/RPBG_ Towards Robust Neural Point-based Graphics in the Wild/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a1bf43140c945ed4814f9f1e960cc253cbc0d3b8 --- /dev/null +++ b/2024/RPBG_ Towards Robust Neural Point-based Graphics in the Wild/layout.json @@ -0,0 +1,9236 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 141, + 111, + 474, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 111, + 474, + 146 + ], + "spans": [ + { + "bbox": [ + 141, + 111, + 474, + 146 + ], + "type": "text", + "content": "RPBG: Towards Robust Neural Point-based Graphics in the Wild" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 144, + 167, + 471, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 167, + 471, + 194 + ], + "spans": [ + { + "bbox": [ + 144, + 167, + 471, + 194 + ], + "type": "text", + "content": "Qingtian Zhu" + }, + { + "bbox": [ + 144, + 167, + 471, + 194 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 144, + 167, + 471, + 194 + ], + "type": "text", + "content": ", Zizhuang Wei" + }, + { + "bbox": [ + 144, + 167, + 471, + 194 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 144, + 167, + 471, + 194 + ], + "type": "text", + "content": ", Zhongtian Zheng" + }, + { + "bbox": [ + 144, + 167, + 471, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 144, + 167, + 471, + 194 + ], + "type": "text", + "content": ", Yifan Zhan" + }, + { + "bbox": [ + 144, + 167, + 471, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 144, + 167, + 471, + 194 + ], + "type": "text", + "content": ", Zhuyu Yao" + }, + { + "bbox": [ + 144, + 167, + 471, + 194 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 144, + 167, + 471, + 194 + ], + "type": "text", + "content": ", Jiawang Zhang" + }, + { + "bbox": [ + 144, + 167, + 471, + 194 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 144, + 167, + 471, + 194 + ], + "type": "text", + "content": ", Kejian Wu" + }, + { + "bbox": [ + 144, + 167, + 471, + 194 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 144, + 167, + 471, + 194 + ], + "type": "text", + "content": ", and Yinqiang Zheng" + }, + { + "bbox": [ + 144, + 167, + 471, + 194 + ], + "type": "inline_equation", + "content": "^{1\\boxtimes}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 257, + 201, + 357, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 201, + 357, + 213 + ], + "spans": [ + { + "bbox": [ + 257, + 201, + 357, + 213 + ], + "type": "text", + "content": "1 The University of Tokyo" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 262, + 213, + 351, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 213, + 351, + 224 + ], + "spans": [ + { + "bbox": [ + 262, + 213, + 351, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 262, + 213, + 351, + 224 + ], + "type": "text", + "content": " Huawei Technologies" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 269, + 224, + 344, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 224, + 344, + 236 + ], + "spans": [ + { + "bbox": [ + 269, + 224, + 344, + 236 + ], + "type": "text", + "content": "3 Peking University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 287, + 235, + 328, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 235, + 328, + 245 + ], + "spans": [ + { + "bbox": [ + 287, + 235, + 328, + 245 + ], + "type": "text", + "content": "4 XREAL" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 275, + 455, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 275, + 455, + 484 + ], + "spans": [ + { + "bbox": [ + 159, + 275, + 455, + 484 + ], + "type": "text", + "content": "Abstract. Point-based representations have recently gained popularity in novel view synthesis, for their unique advantages, e.g., intuitive geometric representation, simple manipulation, and faster convergence. However, based on our observation, these point-based neural re-rendering methods are only expected to perform well under ideal conditions and suffer from noisy, patchy points and unbounded scenes, which are challenging to handle but defacto common in real applications. To this end, we revisit one such influential method, known as Neural Point-based Graphics (NPBG), as our baseline, and propose Robust Point-based Graphics (RPBG). We in-depth analyze the factors that prevent NPBG from achieving satisfactory renderings on generic datasets, and accordingly reform the pipeline to make it more robust to varying datasets in-the-wild. Inspired by the practices in image restoration, we greatly enhance the neural renderer to enable the attention-based correction of point visibility and the inpainting of incomplete rasterization, with only acceptable overheads. We also seek for a simple and lightweight alternative for environment modeling and an iterative method to alleviate the problem of poor geometry. By thorough evaluation on a wide range of datasets with different shooting conditions and camera trajectories, RPBG stably outperforms the baseline by a large margin, and exhibits its great robustness over state-of-the-art NeRF-based variants. Code available at https://github.com/QT-Zhu/RPBG." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 495, + 436, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 495, + 436, + 506 + ], + "spans": [ + { + "bbox": [ + 160, + 495, + 436, + 506 + ], + "type": "text", + "content": "Keywords: Point-based graphics " + }, + { + "bbox": [ + 160, + 495, + 436, + 506 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 495, + 436, + 506 + ], + "type": "text", + "content": " Novel view synthesis " + }, + { + "bbox": [ + 160, + 495, + 436, + 506 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 495, + 436, + 506 + ], + "type": "text", + "content": " Neural rendering" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 525, + 219, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 219, + 539 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 219, + 539 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 551, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 551, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 551, + 482, + 586 + ], + "type": "text", + "content": "Novel view synthesis (NVS) aims to synthesize novel views under given camera poses by aid of a series of already posed images, which is a fundamental task in both computer vision and computer graphics and has been studied for decades." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 587, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 587, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 587, + 482, + 647 + ], + "type": "text", + "content": "Among the feasible solutions, NeRF (Neural Radiance Field) [26] approximates an implicit scene representation, i.e., a radiance field (RF), encoded by a neural network mapping 3D coordinates and view directions to colors and densities. When rendering, the fitted RF is queried multiple times along a ray to volume-render the corresponding pixel for a novel view. Subsequent variants of NeRF follow a case-by-case design and" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 654, + 326, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 654, + 326, + 666 + ], + "spans": [ + { + "bbox": [ + 135, + 654, + 326, + 666 + ], + "type": "text", + "content": "* Work done during Q. Zhu's internship at XREAL." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 236, + 217 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 236, + 217 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 236, + 217 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 236, + 217 + ], + "type": "image", + "image_path": "b7ede94768135d06e5c41278c4cff3fbbf8927a1bd531dbd9bf20921066c4f63.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 229, + 483, + 285 + ], + "lines": [ + { + "bbox": [ + 130, + 229, + 483, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 229, + 483, + 285 + ], + "type": "text", + "content": "Fig. 1: Left: RPBG manages to achieve all-round good re-renderings (PSNR plotted) across generic datasets over the baseline [1] as well as state-of-the-art RF-based methods [42,44]. Right: We demonstrate the point clouds (with camera trajectories visualized) and corresponding re-rendered novel views of the representative scenes, revealing the great robustness and scalability of RPBG. Zoom in for best view." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 236, + 117, + 478, + 218 + ], + "blocks": [ + { + "bbox": [ + 236, + 117, + 478, + 218 + ], + "lines": [ + { + "bbox": [ + 236, + 117, + 478, + 218 + ], + "spans": [ + { + "bbox": [ + 236, + 117, + 478, + 218 + ], + "type": "image", + "image_path": "ec80b35a2520ba64f99530c13415086a3af529c84e449e0ffd1110e96a566457.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 312, + 482, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 361 + ], + "type": "text", + "content": "adopt different parameterization techniques specifically for different scene types, e.g., object-centric [3, 54], free-trajectory [44], and large-scale scenes [38, 42]. Therefore, putting forward one robust method that can work stably across varying scene types with a unified parameterization technique is considered as challenging." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 362, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 362, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 362, + 482, + 542 + ], + "type": "text", + "content": "Recently, point-based alternatives [1, 18, 30, 33] have gained substantial attention for their unique advantages over implicit representations, e.g., ease of manipulation and faster training. The practice of employing point-based parameterization for rendering can be traced all the way back to [16, 20], and is intuitive especially under the context of NVs. Analogue to NeRF and its variants (collectively referred to as RF-based methods), whose optimization and convergence heavily rely on the inherent co-visibility among images, the 3D points triangulated from a series of 2D images according to epipolar geometry have already contain all the verified co-visibility information. In this way, by enforcing the triangulated points as strong prior knowledge for NVs, the expected freedom of optimization is greatly constrained, leading to faster training as well as better robustness. Recent attempts have achieved promising results in terms of fast optimization and rendering [18, 30] and accurate geometric representation [56]. Concretely, we focus on the specific category of methods that manage to integrate CNN-based neural renderers to yield re-renderings with featuremetric neural buffers [22, 30], for their concise pipeline and the potential capability of obtaining visually pleasing re-renderings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 544, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 639 + ], + "type": "text", + "content": "However, by evaluating the representative baseline, NPBG (Neural Point-based Graphics) [1], with generic datasets, we realize that the original design only intends for ideal conditions, e.g., synthetic data [26] and well-captured human heads [31]. When synthesizing novel views under more general conditions, the performance of NPBG degrades severely. In this paper, we strive to boost the robustness of NPBG-like point-based neural re-rendering pipelines and reveal the true potential to achieve state-of-the-art performance across varying datasets in-the-wild, by analyzing the reasons for performance degradation and seeking for remedies." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "Generally speaking, the major difficulties that the vanilla NPBG is faced up with include handling the background, handling patchy point clouds, and identifying correct" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "Q. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 115, + 482, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 115, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 482, + 307 + ], + "type": "text", + "content": "point visibility. We reform the CNN-based neural renderer, with an inspiration from image restoration algorithms [6,8,52] that are able to identify downgraded patterns and restore the corresponding high-quality images. To make sure the neural renderer can capture sufficient and valid context information with extremely sparse rasterizations, we particularly leverages a Downgrade-aware Convolution (DAC) module to determine the correct point visibility with regard to a given camera pose and performs a pseudo point-wise back-face culling operation with visual self-attention. The background is also modeled in a lightweight manner. Instead of incorporating a massive environment map [33], a simple default trainable feature vector can reach similar quantitative results when working with the stronger neural renderer. We also discover the pseudo density calculated from the neural textures can roughly verify the existence of a given 3D position, which can be used to augment the poorly triangulated point clouds. In addition, we also simplify the phased training paradigm in [1,30] and optimize the parameters of both the neural textures and the renderer end-to-end collaboratively. We term our version of point-based re-rendering as RPBG (Robust Point-based Graphics), to emphasize its great robustness across different generic datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 308, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 308, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 133, + 308, + 482, + 426 + ], + "type": "text", + "content": "For thorough qualitative and quantitative evaluation, we collect 4 typical challenging scene types as the benchmark for robust NVs, i.e., " + }, + { + "bbox": [ + 133, + 308, + 482, + 426 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 133, + 308, + 482, + 426 + ], + "type": "text", + "content": " unbounded scenes (with free trajectories), inside-out scenes, large-scale scenes (at the scale of a block or campus), and sparse-view scenes. As a result, RPBG exhibits great robustness with perceptually satisfactory synthesis across the aforementioned typical scenes types as showcased in Fig. 1, where the high-quality re-renderings are obtained with an exactly identical parameterization strategy without any manual configuration, relieving the practitioners from the exhaustive per-scene search of hyper-parameters. Its stable superiority over state-of-the-art NVS methods, we believe, is of great significance especially for real applications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 147, + 428, + 345, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 428, + 345, + 438 + ], + "spans": [ + { + "bbox": [ + 147, + 428, + 345, + 438 + ], + "type": "text", + "content": "The main contributions are three-fold as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 440, + 479, + 533 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 440, + 479, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 440, + 479, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 440, + 479, + 460 + ], + "type": "text", + "content": "- We put forward RPBG, as a more robust and practical alternative for re-rendering high-quality images from triangulated 3D points." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 464, + 479, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 464, + 479, + 497 + ], + "spans": [ + { + "bbox": [ + 138, + 464, + 479, + 497 + ], + "type": "text", + "content": "- According to our in-depth analysis, we enhance the neural re-rendering pipeline regarding the neural renderer, environment modeling, point cloud augmentation, and the training paradigm." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 499, + 479, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 499, + 479, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 499, + 479, + 533 + ], + "type": "text", + "content": "- RPBG manages to greatly boost the performance of neural point re-renderer by a large margin, and exhibits stably greater robustness and generalizability over RF-based methods with even better perceptual rendering quality." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 133, + 555, + 224, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 555, + 224, + 567 + ], + "spans": [ + { + "bbox": [ + 133, + 555, + 224, + 567 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 582, + 482, + 665 + ], + "type": "text", + "content": "Radiance Fields for NVs Along with the proposed NeRF [26], the scene representation of RF is becoming popular for its ease of optimization by the differentiable volume rendering. In a RF, each position is assigned with an anisotropic color and a density, and to render a pixel of a novel view, one needs to sample the trained field and conduct ray marching. The original NeRF [26], together with its variants [2,3,39,54], employs an MLP to represent the functional mapping from coordinates and view directions to the radiance values. Several attempts have been made to encode RFs with different" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 211, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 211, + 91, + 447, + 102 + ], + "type": "text", + "content": "RPGB: Towards Robust Neural Point-based Graphics in the Wild" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 140 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 140 + ], + "type": "text", + "content": "data structures for faster training and inference, e.g., grid voxels [12, 36], decomposed tensors [5], hash grids [27]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 141, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 141, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 133, + 141, + 482, + 259 + ], + "type": "text", + "content": "In practice, the parameterization strategies have a great impact on the rendering quality when reconstructing different types of scenes, e.g., forward-facing ones, " + }, + { + "bbox": [ + 133, + 141, + 482, + 259 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 133, + 141, + 482, + 259 + ], + "type": "text", + "content": " unbounded ones, and large-scale ones. For unbounded scenes, " + }, + { + "bbox": [ + 133, + 141, + 482, + 259 + ], + "type": "inline_equation", + "content": "\\mathrm{NeRF}++" + }, + { + "bbox": [ + 133, + 141, + 482, + 259 + ], + "type": "text", + "content": " [54] applies separate networks and different parameterization to model near and distant objects; mip-NeRF 360 [3] designs a smooth contraction operator to parameterize the whole unbounded scene into a ball; Mega-NeRF [42] and Block-NeRF [38] partition 3D scenes explicitly and use different networks to represent each scene partition; " + }, + { + "bbox": [ + 133, + 141, + 482, + 259 + ], + "type": "inline_equation", + "content": "\\mathrm{F}^2" + }, + { + "bbox": [ + 133, + 141, + 482, + 259 + ], + "type": "text", + "content": "-NeRF [44] proposes to use perspective warping to handle sequential data with arbitrary trajectories. The varying case-by-case parameterization strategies of RF-based NVS greatly constraint the generalizability of methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 272, + 482, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 272, + 482, + 343 + ], + "spans": [ + { + "bbox": [ + 133, + 272, + 482, + 343 + ], + "type": "text", + "content": "Point-based Graphics A point cloud is a collection of 3D coordinates that is usually used as a topology-agnostic coarse shape representation of the 3D geometry and favored for its flexibility and sparsity for storage. The development of techniques of employing points as modeling primitives for rendering (referred as point-based graphics [15]) can be traced back to [16, 20], the best practice of which is to replace each point with an oriented circular disk (a surfel) and reply on splatting to blend overlapping surfels [28]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 344, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 344, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 133, + 344, + 482, + 511 + ], + "type": "text", + "content": "In contrast to conventional physically based rendering (PBR) techniques, neural rendering [40] learns to render high-quality images in a data-driven manner. Particularly, we focus on the methods that conduct neural re-rendering with point clouds. Bui et al. [4] propose to enhance the coarse point-based rendering by a GAN for image super-resolution. NRW [25] and InvSFM [29] attempt to re-render the reconstructed point cloud with respective auxiliary buffers, e.g., latent appearance vectors, semantic masks and SIFT descriptors [23]. NPBG [1,30] adopts a U-Net-like [32] CNN to render neural point textures as RGB images, and exhibits better flexibility over mesh-based proxies [41]. The practical applications in the context of autonomous driving, e.g., scene editing and stitching, and large-scale training, are further explored in READ [21]. Recently, ADOP [33] and Gaussian Splatting [18] have demonstrated notable accomplishments. However, the differentiable rendering/splatting scheme employed requires an extensive amount of memory to maintain the computational graph and gradients during training. This limitation hinders their application on large-scale scenes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 529, + 350, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 529, + 350, + 544 + ], + "spans": [ + { + "bbox": [ + 133, + 529, + 350, + 544 + ], + "type": "text", + "content": "3 Neural Point-based Graphics Revisited" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 555, + 481, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 555, + 481, + 603 + ], + "spans": [ + { + "bbox": [ + 133, + 555, + 481, + 603 + ], + "type": "text", + "content": "In this section, we revisit the pipeline of NPBG [1], analyze the existing drawbacks when attempting to re-render generic scenes, and attempt to explain why NPBG finds it difficult to handle the background and patchy points, and identify correct point visibility. We hope these discussions be of sufficient insights to support our modifications." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 621, + 214, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 621, + 214, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 621, + 214, + 632 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 641, + 481, + 665 + ], + "type": "text", + "content": "Inputs As the common practice in NVS, the inputs are a series of images with respective camera parameters (both intrinsics and extrinsics), and specifically for point-based" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "text", + "content": "Q. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 117, + 222, + 217 + ], + "blocks": [ + { + "bbox": [ + 137, + 117, + 222, + 217 + ], + "lines": [ + { + "bbox": [ + 137, + 117, + 222, + 217 + ], + "spans": [ + { + "bbox": [ + 137, + 117, + 222, + 217 + ], + "type": "image", + "image_path": "b352c23801bc9ce10c0ac8b1ae1517c38d8b08f403fe65e1e53863e8924256e6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 218, + 209, + 225 + ], + "lines": [ + { + "bbox": [ + 153, + 218, + 209, + 225 + ], + "spans": [ + { + "bbox": [ + 153, + 218, + 209, + 225 + ], + "type": "text", + "content": "(a) Rasterized Points" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 223, + 117, + 306, + 217 + ], + "blocks": [ + { + "bbox": [ + 223, + 117, + 306, + 217 + ], + "lines": [ + { + "bbox": [ + 223, + 117, + 306, + 217 + ], + "spans": [ + { + "bbox": [ + 223, + 117, + 306, + 217 + ], + "type": "image", + "image_path": "43a2af0e2b26b0e374be61b1f1495cb66efdfeb3a427a7716d3bcfdf9846764a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 218, + 285, + 225 + ], + "lines": [ + { + "bbox": [ + 239, + 218, + 285, + 225 + ], + "spans": [ + { + "bbox": [ + 239, + 218, + 285, + 225 + ], + "type": "text", + "content": "(b) Ground Truth" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 117, + 391, + 217 + ], + "blocks": [ + { + "bbox": [ + 307, + 117, + 391, + 217 + ], + "lines": [ + { + "bbox": [ + 307, + 117, + 391, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 117, + 391, + 217 + ], + "type": "image", + "image_path": "c86057279b26509d122c492de35d5788f1fa7e38e97ecdd7482bf98ddb9e2282.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 218, + 361, + 225 + ], + "lines": [ + { + "bbox": [ + 335, + 218, + 361, + 225 + ], + "spans": [ + { + "bbox": [ + 335, + 218, + 361, + 225 + ], + "type": "text", + "content": "(c) NPBG" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 237, + 481, + 270 + ], + "lines": [ + { + "bbox": [ + 131, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 131, + 237, + 481, + 270 + ], + "type": "text", + "content": "Fig. 2: Typically challenging scenes in T&T dataset [19] for NPBG. Top: Auditorium, where the walls and ceilings are extremely sparse. Bottom: Museum, where the point sparsity makes the occlusion and visibility complicated." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 392, + 118, + 476, + 217 + ], + "blocks": [ + { + "bbox": [ + 392, + 118, + 476, + 217 + ], + "lines": [ + { + "bbox": [ + 392, + 118, + 476, + 217 + ], + "spans": [ + { + "bbox": [ + 392, + 118, + 476, + 217 + ], + "type": "image", + "image_path": "689cc3a0243e40edf745f0ea06e9b1930cebdee309b5611b5468e5a33add5c3c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 422, + 218, + 444, + 225 + ], + "lines": [ + { + "bbox": [ + 422, + 218, + 444, + 225 + ], + "spans": [ + { + "bbox": [ + 422, + 218, + 444, + 225 + ], + "type": "text", + "content": "(d) Ours" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 291, + 481, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 291, + 481, + 327 + ], + "spans": [ + { + "bbox": [ + 131, + 291, + 481, + 327 + ], + "type": "text", + "content": "methods [1, 18, 30], the very first step is supposed to be the triangulation of 3D points from 2D observations. The original NPBG-like pipelines [1, 30] mainly consider this as a pre-processing without much attention." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 337, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 337, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 131, + 337, + 482, + 456 + ], + "type": "text", + "content": "Point Rasterization NPBG [1] and NPBG++ [30] apply a non-differentiable hard point " + }, + { + "bbox": [ + 131, + 337, + 482, + 456 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 131, + 337, + 482, + 456 + ], + "type": "text", + "content": "-buffering operation as an approximated back-face culling when rasterizing points as 2D fragments, where the fragment is updated when and only when the newly projected point has a smaller " + }, + { + "bbox": [ + 131, + 337, + 482, + 456 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 131, + 337, + 482, + 456 + ], + "type": "text", + "content": "-depth than the current one. After a traversal of all the points, the fragment keeps the record of the indices of rasterized points. Then tensor scattering is performed to index the neural texture at the corresponding positions. Compared to differentiable rendering/splatting alternatives [18, 33], we consider the most significant advantage is its memory efficiency and thus better scalability, since the computation graph required to keep is much smaller for NPBG. We also want to keep this scalability without interfering the elegant rasterization paradigm." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 464, + 482, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 464, + 482, + 538 + ], + "spans": [ + { + "bbox": [ + 131, + 464, + 482, + 538 + ], + "type": "text", + "content": "Neural Renderer The renderer " + }, + { + "bbox": [ + 131, + 464, + 482, + 538 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\Theta}:\\mathbb{R}^{H\\times W\\times C}\\to \\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 131, + 464, + 482, + 538 + ], + "type": "text", + "content": " of NPBG [1] follows a vanilla U-Net-like architecture [32]. Since the target point clouds for rendering are supposed to be well-constructed, the vanilla U-Net is expressive enough to complete the assigned task of mapping higher dimensional features from the neural texture T to RGB values. Empirically, we find the expected properties of " + }, + { + "bbox": [ + 131, + 464, + 482, + 538 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\Theta}" + }, + { + "bbox": [ + 131, + 464, + 482, + 538 + ], + "type": "text", + "content": " have much in common with a network for low-level vision tasks, e.g., image restoration." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 552, + 198, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 552, + 198, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 552, + 198, + 563 + ], + "type": "text", + "content": "3.2 Problems" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 570, + 482, + 665 + ], + "type": "text", + "content": "Patchy Triangulation Consider the inputs, for a NVS system aiming to synthesize images for generic scenes, the triangulation of points is not trivial. The point triangulation step to perform is similar to the step of multi-view stereo (MVS) reconstruction [14, 48], the key of which is to identify co-visible pixels across images and lift the 2D pixels to a 3D points according to epipolar geometry. Such algorithms suffer from non-Lambertian surfaces and textureless regions, both of which are very common and lead to poor triangulated and patchy points. We attach two typical scenes that NPBG fails to yield good renderings in Fig. 2." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "type": "text", + "content": "RPBG: Towards Robust Neural Point-based Graphics in the Wild" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 482, + 224 + ], + "type": "text", + "content": "Wrong Point Visibility As can be imagined, such point visibility can be erroneous due to the poor quality of triangulated points, and also due to the inherent sparsity of points as a 3D representation. In this way, the points belonging to back faces, which should be considered as occluded, are rasterized as the fragment for further processing. Manually setting a depth threshold for bound check could not be helpful either, since being far from the camera does not necessarily indicate they should be occluded — they could be parts of the environment. While NHR [47] proposes to take the depth buffer as an additional rendering condition, we find its effectiveness not significant for generic scenes other than well-masked human captures." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 236, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 482, + 319 + ], + "type": "text", + "content": "Lack of Context for Re-rendering Similarly, when re-rendering an incomplete point cloud, where the neural buffers are usually with significant flaws, the whole receptive field of a kernel at a given position may only capture the downgraded regions, leading to failure to yield reasonable restoration, especially for a high target resolution. It is worth noting that, though the relationship between the number of network layers and rendering quality has been discussed in [30, 33], there is still huge room for the improvement of the renderer." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 331, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 331, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 133, + 331, + 482, + 415 + ], + "type": "text", + "content": "Failure to Effectively Model the Environment Unbounded scenes are a typical challenging scene type for NVS. RF-based methods that are designed on purpose to handle such scenes [3, 54] typically employ different parameterization to encode the areas outside a certain bounding sphere. For point-based methods, ADOP [33] proposes to apply an environment map of " + }, + { + "bbox": [ + 133, + 331, + 482, + 415 + ], + "type": "inline_equation", + "content": "H \\times W \\times C" + }, + { + "bbox": [ + 133, + 331, + 482, + 415 + ], + "type": "text", + "content": " to model the environment, which is equivalent to wrapping the triangulated points with a sphere with " + }, + { + "bbox": [ + 133, + 331, + 482, + 415 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 133, + 331, + 482, + 415 + ], + "type": "text", + "content": " points, resulting in more than " + }, + { + "bbox": [ + 133, + 331, + 482, + 415 + ], + "type": "inline_equation", + "content": "5 \\times 10^{5}" + }, + { + "bbox": [ + 133, + 331, + 482, + 415 + ], + "type": "text", + "content": " points as overheads, according to the default configuration." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 427, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 427, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 133, + 427, + 482, + 464 + ], + "type": "text", + "content": "Summary To summarize, we attribute the observed problems to two main causes: poor geometry and weak, local renderer. The two causes are to some extent coupled for a better geometry will relieve the difficulty of the renderer and vice versa." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 482, + 304, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 482, + 304, + 496 + ], + "spans": [ + { + "bbox": [ + 133, + 482, + 304, + 496 + ], + "type": "text", + "content": "4 Robust Point-based Graphics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 507, + 482, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 507, + 482, + 603 + ], + "spans": [ + { + "bbox": [ + 133, + 507, + 482, + 603 + ], + "type": "text", + "content": "We are strongly convinced that the pipeline of point-based neural re-rendering has the great potential to outperform popular RF-based solutions, for its unified point-as-parameterization fashion and the incorporation of neural networks to yield visual details. Therefore in this section, we introduce Robust Point-based Graphics (RPBG), as an enhanced version of NPBG, with a particular focus on robustness under generic scenes. We will elaborate the modifications made and shed light on the underlying insights and motivation. The overall pipeline of RPBG as well as the training paradigm is briefly illustrated in Fig. 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 620, + 310, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 620, + 310, + 633 + ], + "spans": [ + { + "bbox": [ + 133, + 620, + 310, + 633 + ], + "type": "text", + "content": "4.1 Downgrade-aware Neural Renderer" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 641, + 482, + 665 + ], + "type": "text", + "content": "As is attributed as one key problem in NPBG, the U-Net-based neural renderer is considered as too naive to handle the challenging situations in generic scenes. The" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "text", + "content": "Q. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 479, + 281 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 479, + 281 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 479, + 281 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 479, + 281 + ], + "type": "image", + "image_path": "6b8cd81a3ed5ee1a60464a343d2dc2eda618897d578f7552b643a44052342850.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 296, + 482, + 384 + ], + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 384 + ], + "type": "text", + "content": "Fig. 3: The overall pipeline of RPBG. Point Triangulation: We first triangulate a 3D proxy for re-rendering with posed images, with its geometry-bounded neural texture initialized. Point Rasterization: The points are raterized to the given camera in a non-differentiable manner. By indexing the texture with the fragment, we obtain the neural buffer. A learnable point-size neural texture " + }, + { + "bbox": [ + 130, + 296, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{env}}" + }, + { + "bbox": [ + 130, + 296, + 482, + 384 + ], + "type": "text", + "content": " is also optimized. Neural Rendering: The restoration from downgraded neural buffer to photo-realistic images is performed by a CNN. The network and the neural texture are optimized end-to-end by image-level losses. An offline point cloud augmentation strategy is introduced to alleviate the problem of patchy triangulation under challenging conditions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 411, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 483 + ], + "type": "text", + "content": "modifications made are deeply inspired by low-level vision tasks, where networks can adaptively determine downgraded parts, e.g., the deblurring/deraining networks are able to identify the blurring/raining pixels from the whole image, and correct them accordingly. We expect the neural renderer " + }, + { + "bbox": [ + 130, + 411, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\Theta}" + }, + { + "bbox": [ + 130, + 411, + 482, + 483 + ], + "type": "text", + "content": " can benefit from relevant restoration-targeting techniques, and become able to decode high-quality visual information from patchy buffers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 484, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 482, + 579 + ], + "type": "text", + "content": "After evaluating three state-of-the-art fundamental architectures for image restoration, i.e., multi-scale fusion [8], multi-stage [52], and U-Net [6], following the taxonomy in [6]. According to the reported experiments in Sec. 5.4, we opt the paradigm in [8] for it achieves the best balance between performance and time/memory efficiency. However, the convolution layers in such paradigm is still with a fixed receptive field, lacking in robustness against point sparsity. We would like to further enlarge the receptive field to a global scale and attempt to explicitly model the visual attention to weigh the observed points." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "Transformer-based architectures for image restoration, e.g., Restormer [51], brings unacceptable memory overheads, so we rely on the frequency-domain alternative, Fast Fourier Convolution (FFC) [7], which can theoretically capture global contexts, to determine the correct point visibility adaptively. FFC performs channel-wise real 2D FFT (Fast Fourier Transform) and inverse real 2D FFT on 2D tensors. Real FFT uses only half of the spectrum and by convolving the transformed frequency-domain tensors. In this way, a receptive field covering the entire image is considered." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "type": "text", + "content": "RPBG: Towards Robust Neural Point-based Graphics in the Wild" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 177, + 118, + 443, + 358 + ], + "blocks": [ + { + "bbox": [ + 177, + 118, + 443, + 358 + ], + "lines": [ + { + "bbox": [ + 177, + 118, + 443, + 358 + ], + "spans": [ + { + "bbox": [ + 177, + 118, + 443, + 358 + ], + "type": "image", + "image_path": "424413adaf0946be3980caff4e0c2453f8cdc98668babe46c712c681ab7db2dc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 371, + 483, + 407 + ], + "lines": [ + { + "bbox": [ + 130, + 371, + 483, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 483, + 407 + ], + "type": "text", + "content": "Fig. 4: The architecture of the downgrade-aware neural renderer in RPBG, with some conventional modules omitted. From the visualized attention map, DAC manages to adaptively handle the severely erroneous point visibility." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 427, + 483, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 427, + 483, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 427, + 483, + 525 + ], + "type": "text", + "content": "Inspired by [37], we apply FFT (as the global branch) in parallel to conventional convolution layers (as the local branch) and rely on the fused features of both branches to determine the downgraded regions for the gated convolution [50] to filter. Based on the common practice in image inpainting, we leverage gated convolutions at the early stage of the renderer, to help locate the downgrade by wrong point visibility. We name such customized gated convolution module as the Downgrade-aware Convolution (DAC) module, and as can be inferred from the point attention maps before and after DAC in Fig. 4, DAC manages to determine complicated point visibility with patchy triangulation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 539, + 242, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 539, + 242, + 551 + ], + "spans": [ + { + "bbox": [ + 131, + 539, + 242, + 551 + ], + "type": "text", + "content": "4.2 Point Triangulation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 558, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 605 + ], + "type": "text", + "content": "Though there are advances in MVS empowered by deep learning [43, 45, 46], recovering a complete, accurate, and dense point cloud for certain regions remains a challenging. For RPBG, we adopt an off-the-shelf MVS method [46] to perform point triangulation for its memory efficiency to allow large-batch inference." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 606, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 483, + 666 + ], + "type": "text", + "content": "As for problem the poor triangulation, we partially leave it to the neural renderer " + }, + { + "bbox": [ + 130, + 606, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\Theta}" + }, + { + "bbox": [ + 130, + 606, + 483, + 666 + ], + "type": "text", + "content": ", relying on a stronger renderer to recover the re-rendering from downgraded buffers. Inspired by the point cloud extraction method of RF-based implementations [39], where the estimated radiance density " + }, + { + "bbox": [ + 130, + 606, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 606, + 483, + 666 + ], + "type": "text", + "content": " can be a rough indicator for surface, we propose a point cloud augmentation technique to densify the initial triangulation." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Q. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": "The point cloud augmentation follows a trial-and-error paradigm by first assuming the existence of newly sampled points and then verify them by estimated pseudo densities " + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "inline_equation", + "content": "\\sigma_{i} = \\sum |\\mathbf{T}_{i}|" + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": ". It is empirically observed that the absolute activation of the point-wise neural texture can roughly represent the reliability of the 3D position — it is reasonable since an outlier will be less distinguishable by the neural renderer and thus less visual attention will be given." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 480, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 480, + 224 + ], + "type": "text", + "content": "Note that this strategy is optional, and we only apply such strategy to scenes with extremely poor triangulation. Please refer to the Supplementary Material for more discussions regarding the self-pruning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 242, + 257, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 242, + 257, + 254 + ], + "spans": [ + { + "bbox": [ + 131, + 242, + 257, + 254 + ], + "type": "text", + "content": "4.3 Environment Modeling" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 262, + 479, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 479, + 309 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 479, + 309 + ], + "type": "text", + "content": "As a result of the incorporation of a stronger neural renderer, we find that given a neural texture with " + }, + { + "bbox": [ + 130, + 262, + 479, + 309 + ], + "type": "inline_equation", + "content": "C = 8" + }, + { + "bbox": [ + 130, + 262, + 479, + 309 + ], + "type": "text", + "content": ", the dense environment map suggested in [33] is redundant. Instead, we shrink the overheads for environment modeling from " + }, + { + "bbox": [ + 130, + 262, + 479, + 309 + ], + "type": "inline_equation", + "content": "H \\times W \\times C" + }, + { + "bbox": [ + 130, + 262, + 479, + 309 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 262, + 479, + 309 + ], + "type": "inline_equation", + "content": "1 \\times C" + }, + { + "bbox": [ + 130, + 262, + 479, + 309 + ], + "type": "text", + "content": ", and relying on the stronger neural renderer to decode the background." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 310, + 480, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 480, + 369 + ], + "type": "text", + "content": "When rasterizing, we employ a tunable feature vector " + }, + { + "bbox": [ + 130, + 310, + 480, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{env}}" + }, + { + "bbox": [ + 130, + 310, + 480, + 369 + ], + "type": "text", + "content": " aside the neural point texture " + }, + { + "bbox": [ + 130, + 310, + 480, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 130, + 310, + 480, + 369 + ], + "type": "text", + "content": ", as the default value for vacant pixels in the fragment, which is also involved in the end-to-end training. By experiments in Sec. 5.4, we demonstrate that the lightweight modeling strategy reaches the quantitative performance equivalent to applying an environment map, yet with a negligible overhead." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 387, + 274, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 387, + 274, + 399 + ], + "spans": [ + { + "bbox": [ + 131, + 387, + 274, + 399 + ], + "type": "text", + "content": "4.4 Collaborative Optimization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 407, + 479, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 479, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 479, + 431 + ], + "type": "text", + "content": "Recall that the point rasterization procedure is completely parameter-free. The overall collaborative optimization scheme can be thus formulated as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 156, + 441, + 482, + 465 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 441, + 482, + 465 + ], + "spans": [ + { + "bbox": [ + 156, + 441, + 482, + 465 + ], + "type": "interline_equation", + "content": "\\mathbf {T} ^ {*}, \\Theta^ {*} = \\underset {\\mathbf {T}, \\Theta} {\\arg \\min } \\sum_ {k} \\mathcal {L} (\\mathbf {I} _ {k}, \\hat {\\mathbf {I}} _ {k}) = \\underset {\\mathbf {T}, \\Theta} {\\arg \\min } \\sum_ {k} \\mathcal {L} (\\mathbf {I} _ {k}, \\mathbf {K}, \\mathbf {R} _ {k}, \\mathbf {t} _ {k} | \\mathbf {T}, \\mathcal {R} _ {\\Theta}), \\tag {1}", + "image_path": "7ba3cccafd2b34ec70df742691c54d455562f8e4d5dc51f43bf340ec7b5e9cc5.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 474, + 480, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 480, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 474, + 480, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 130, + 474, + 480, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 474, + 480, + 510 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 474, + 480, + 510 + ], + "type": "text", + "content": " stand for the tunable parameters, i.e., the neural point texture (alongside the globally shared " + }, + { + "bbox": [ + 130, + 474, + 480, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{env}}" + }, + { + "bbox": [ + 130, + 474, + 480, + 510 + ], + "type": "text", + "content": ") and the parameters of the renderer " + }, + { + "bbox": [ + 130, + 474, + 480, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\Theta}" + }, + { + "bbox": [ + 130, + 474, + 480, + 510 + ], + "type": "text", + "content": " for neural re-rendering, while " + }, + { + "bbox": [ + 130, + 474, + 480, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_k" + }, + { + "bbox": [ + 130, + 474, + 480, + 510 + ], + "type": "text", + "content": " is the target image whose calibration is " + }, + { + "bbox": [ + 130, + 474, + 480, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 474, + 480, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 474, + 480, + 510 + ], + "type": "inline_equation", + "content": "[\\mathbf{R}_k|\\mathbf{t}_k]" + }, + { + "bbox": [ + 130, + 474, + 480, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 474, + 480, + 510 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_k" + }, + { + "bbox": [ + 130, + 474, + 480, + 510 + ], + "type": "text", + "content": " is the re-rendering" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 510, + 480, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 480, + 557 + ], + "type": "text", + "content": "Note that, we also discard all the typical tossing sampling [21] and optimization steps [30, 33] of point-based neural re-rendering, and manage to tune all the parameters involved in a simple but effective collaborative end-to-end way. The neural texture is initialized with all zeros while the rendering CNN is trained from scratch for each scene" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": "Loss Function Since the rendering scheme in both NPBG and RPBG is convolutional, where images are rendered in patches, we are able to apply patch-aware losses to enforce the involvement of neighboring pixels to ensure patch-to-patch consistency, offering perceptually good renderings. To this end, in addition to the pixel-wise Huber norm " + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{huber}}" + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": " providing the basic supervision and numerical stability, we apply two patch-aware losses to the collaborative optimization of RPBG, namely the perceptual VGG loss [10, 17] " + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{vgg}}" + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": ", and the FFT loss [8, 13] " + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{fft}}" + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": ". VGG loss compares the rendered image and the target ground-truth image in a high-dimensional feature space by a pre-trained VGG-19" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "type": "text", + "content": "RPBG: Towards Robust Neural Point-based Graphics in the Wild" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "network [35]. It reveals the perceptual similarity between images that pixel-wise metrics fail to measure. FFT loss measures the image-to-image distance in the frequency domain by carrying out 2D FFT towards images. The frequency components are considered to be crucial in terms of the perceptual quality [13]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 164, + 373, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 164, + 373, + 175 + ], + "spans": [ + { + "bbox": [ + 146, + 164, + 373, + 175 + ], + "type": "text", + "content": "The final loss function applied in RPBG is composed as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 228, + 180, + 481, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 180, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 228, + 180, + 481, + 194 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {\\text {h u b e r}} \\mathcal {L} _ {\\text {h u b e r}} + \\lambda_ {\\text {v g g}} \\mathcal {L} _ {\\text {v g g}} + \\lambda_ {\\text {f f t}} \\mathcal {L} _ {\\text {f f t}}. \\tag {2}", + "image_path": "86a00f75b94403cae56312c27a097ca047c07875edef0723ad7fcf0c56bbe13e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 208, + 218, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 208, + 218, + 220 + ], + "spans": [ + { + "bbox": [ + 131, + 208, + 218, + 220 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 229, + 194, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 229, + 194, + 239 + ], + "spans": [ + { + "bbox": [ + 131, + 229, + 194, + 239 + ], + "type": "text", + "content": "5.1 Datasets" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 245, + 481, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 245, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 481, + 364 + ], + "type": "text", + "content": "We conduct experiments covering a diverse range of scene types to quantitatively examine the generalizability and robustness of NVs methods, including the Free dataset [44] (7 free-trajectory scans with long camera trajectories), Tanks and Temples (T&T) dataset [19] (4 unbounded scans and 5 inside-out scans, without salient dynamic objects), Mill19 dataset [42] (2 large-scale aerial scans with over 1600 images in each scan), and GigaMVS dataset [53] (8 sparse-view outdoor scans). Note that for T&T, we use the raw undistorted images provided by the benchmark without any masking [22]. We follow the common split protocol, that 1 frame out of every 8 frames is evaluated, for the Free dataset, T&T dataset, and GigaMVS dataset. For Mill19 dataset, we apply the officially recommended split to align with the experiments in [42]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 365, + 481, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 481, + 423 + ], + "type": "text", + "content": "We also include some other challenging datasets for a thorough evaluation, e.g., ScanNet++ dataset [49], a challenging indoor benchmark, OMMO dataset [24], a multi-modal aerial NVS dataset and ETH-MS dataset [11], a super-large-scale dataset (around " + }, + { + "bbox": [ + 130, + 365, + 481, + 423 + ], + "type": "inline_equation", + "content": "5k" + }, + { + "bbox": [ + 130, + 365, + 481, + 423 + ], + "type": "text", + "content": " images). Please refer to the Supplementary Material for more quantitative and qualitative results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 438, + 257, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 438, + 257, + 449 + ], + "spans": [ + { + "bbox": [ + 131, + 438, + 257, + 449 + ], + "type": "text", + "content": "5.2 Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 454, + 481, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 454, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 454, + 481, + 536 + ], + "type": "text", + "content": "Data Preparation To obtain the per-scene point cloud representation for re-rendering, we follow the mapping procedure of COLMAP [34] for sparse triangulation and the MVS reconstruction with a trained network [46] for dense triangulation. Note that the reconstruction only takes a small proportion of time over the whole training phase and we conduct ablation experiments to show that RPBG is robust against different triangulation configurations. Please refer to Sec. 5.4 as well as the Supplementary Material for more information." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 546, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 481, + 665 + ], + "type": "text", + "content": "Training Settings We randomly crop images to square patches of " + }, + { + "bbox": [ + 130, + 546, + 481, + 665 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 546, + 481, + 665 + ], + "type": "text", + "content": ", with a batch size of 8. The learning rate for the neural point textures is " + }, + { + "bbox": [ + 130, + 546, + 481, + 665 + ], + "type": "inline_equation", + "content": "10^{-1}" + }, + { + "bbox": [ + 130, + 546, + 481, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 546, + 481, + 665 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 130, + 546, + 481, + 665 + ], + "type": "text", + "content": " for the rendering networks, which will decay by a factor of 0.5 if 5 consequent epochs witness no drop in the loss function. A dimension of 8 is applied for any neural texture regardless of the scene scale or complexity. For the weights of loss functions, we globally set " + }, + { + "bbox": [ + 130, + 546, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{huber}} = 10^3" + }, + { + "bbox": [ + 130, + 546, + 481, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 546, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{vgg}} = 1" + }, + { + "bbox": [ + 130, + 546, + 481, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 546, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{fft}} = 1" + }, + { + "bbox": [ + 130, + 546, + 481, + 665 + ], + "type": "text", + "content": ". The training is performed on one NVIDIA GeForce RTX 3090, with a GPU memory consumption of up to 23 GB. It takes around 8 to 30 GPU hours for training, depending on the data scale. We would like to emphasize that, RPBG does not require case-by-case scene parameterization or grid search of training hyper-parameters for all the scenes covered in the experiments." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "Q. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 117, + 479, + 350 + ], + "blocks": [ + { + "bbox": [ + 137, + 117, + 479, + 350 + ], + "lines": [ + { + "bbox": [ + 137, + 117, + 479, + 350 + ], + "spans": [ + { + "bbox": [ + 137, + 117, + 479, + 350 + ], + "type": "image", + "image_path": "382384e456bc329ff256bb9ad09ac248cec17f078676bb0f4c5e63dc006910b2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 362, + 483, + 407 + ], + "lines": [ + { + "bbox": [ + 130, + 362, + 483, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 362, + 483, + 407 + ], + "type": "text", + "content": "Fig. 5: Visualized comparisons over varying scenes. From top to bottom: sky and hydrant of the Free dataset [44], Courtyard and Train of T&T dataset [19], and DayaTemple and MemorialHall of GigaMVS dataset [53]. We include the results of RPBG (Ours), Gaussian Splatting [18], NPBG [1], and " + }, + { + "bbox": [ + 130, + 362, + 483, + 407 + ], + "type": "inline_equation", + "content": "\\mathrm{F}^2" + }, + { + "bbox": [ + 130, + 362, + 483, + 407 + ], + "type": "text", + "content": "-NeRF [44] for comparison. Zoom in for best view." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 430, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 479 + ], + "type": "text", + "content": "Evaluation Metrics We adopt the metrics of PSNR, SSIM and LPIPS (VGG) [55] for the evaluation between the synthesized and the target images. According to [55], PSNR does not faithfully measure image sharpness and so cannot properly account for the nuances of human visual perception." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 498, + 189, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 498, + 189, + 510 + ], + "spans": [ + { + "bbox": [ + 131, + 498, + 189, + 510 + ], + "type": "text", + "content": "5.3 Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 520, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 520, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 520, + 482, + 593 + ], + "type": "text", + "content": "In Fig. 5, we demonstrate several typical groups of visual comparisons of NVS results by RPBG, two representative point-based methods, i.e., Gaussian Splitting [18] and NPBG [1], and a state-of-the-art NeRF variant for unbounded scenes, " + }, + { + "bbox": [ + 130, + 520, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\mathrm{F}^2" + }, + { + "bbox": [ + 130, + 520, + 482, + 593 + ], + "type": "text", + "content": "-NeRF [44]. We group the datasets by category and report the corresponding scene-averaged quantitative scores in Tab. 1. Note that RPBG achieves the best SSIM and LPIPS, which are more relevant to the high-frequency components of images, across all datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "Free-trajectory/Unbounded Scenes Compared with NVS methods designed in particular for unbounded scenes, " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathrm{NeRF}++" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " [54] and " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathrm{F^2}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "-NeRF [44], RPBG achieves the best SSIM and LPIPS and comparable PSNR. For visual effects showcased in Fig. 5, since RPBG involves a neighborhood for rendering, its results appear more visually harmonious, especially than Gaussian Splatting [18] and " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathrm{F^2}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "-NeRF [44]. It is necessary to mention" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "type": "text", + "content": "RPBG: Towards Robust Neural Point-based Graphics in the Wild" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 168, + 481, + 261 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 158 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 158 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 158 + ], + "type": "text", + "content": "Table 1: Quantitative evaluation of state-of-the-art NVS methods [1,5,18,26,30,42,44,54] and RPBG across diverse scenes grouped by category, including free trajectory/unbounded scenes, inside-out scenes, large-scale scenes, and sparse-view scenes. The figure following the dataset name stands for the number of scenes the dataset contains." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 168, + 481, + 261 + ], + "lines": [ + { + "bbox": [ + 133, + 168, + 481, + 261 + ], + "spans": [ + { + "bbox": [ + 133, + 168, + 481, + 261 + ], + "type": "table", + "html": "
MethodFree-trajectory/UnboundedInside-out T&T-5Large-scale Mill19-2Sparse-view GigaMVS-8
Free Dataset-7T&T-4
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NeRF [26]17.750.4050.59716.840.3960.73119.820.6150.52620.340.5240.52918.290.6320.499
TensoRF [5]21.740.5490.60020.730.6430.56619.430.6340.57020.070.4970.59718.700.6530.506
NeRF++ [54]23.470.6030.49921.660.6580.52919.250.6100.58520.190.5200.53118.380.6320.495
F2-NeRF [44]26.320.7790.27623.660.7640.30320.120.7060.394N/AN/AN/A17.440.5400.470
Mega-NeRF [42]22.600.5700.56218.730.5780.47819.160.6170.45122.490.5500.51018.250.5810.394
NPBG [1]21.400.6390.34019.850.6980.37620.570.6960.37116.210.3570.64418.340.6200.405
NPBG++ [30]20.060.5920.44517.230.6530.47418.300.6840.41117.040.4000.64819.300.6630.443
Gaussian Splatting [18]25.230.7400.29023.510.7820.29323.460.7830.277N/AN/AN/A16.840.5300.391
Ours26.330.8320.17722.500.7820.27623.290.8040.24222.620.5960.36820.540.6860.317
", + "image_path": "4677386befd0997cacb294bda6e7416a15d87cf5a53804d3b5cea4157dbcbb85.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 285, + 480, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 480, + 309 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 480, + 309 + ], + "type": "text", + "content": "that, for RF-based methods [5, 26, 42, 44], we have manually adjusted the scene-specific hyper-parameters to achieve better results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 319, + 481, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 481, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 481, + 367 + ], + "type": "text", + "content": "Inside-out Scenes The inside-out indoor scenes are strictly bounded, but lack high-quality ray intersections required for optimization, which will lead to under-fitting of the RF (Courtroom in Fig. 5). Similar to free-trajectory/unbounded scenes, RPBG outperforms state-of-the-art methods in SSIM and LPIPS." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 378, + 482, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 462 + ], + "type": "text", + "content": "Large-scale Scenes We mainly evaluate RPBG against Mega-NeRF [42] on the massive imagery of Mill19 dataset [42]. Our method outperforms Mega-NeRF [42] in every metric. It is worth noting that Mega-NeRF takes 240 GPU hours for training while RPBG only takes 29 GPU hours. Notably, our system with 256 GB RAM and RTX 3090 fails to afford the training of " + }, + { + "bbox": [ + 130, + 378, + 482, + 462 + ], + "type": "inline_equation", + "content": "\\mathrm{F}^2" + }, + { + "bbox": [ + 130, + 378, + 482, + 462 + ], + "type": "text", + "content": "-NeRF [44] and Gaussian Splitting [18] (marked as N/A in the table). In contrast, RPBG proves to be viable for large-scale scenes under the same hardware constraint." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 473, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 545 + ], + "type": "text", + "content": "Sparse-view Scenes Sparse-view inputs are considered to be extremely challenging for NVS methods. Since RPBG, NPBG [1], and " + }, + { + "bbox": [ + 130, + 473, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\mathrm{NPBG + + }" + }, + { + "bbox": [ + 130, + 473, + 482, + 545 + ], + "type": "text", + "content": " [30] incorporate a point-based 3D proxy for re-rendering, they showcase better robustness over RF-based methods. As revealed in Fig. 5, Gaussian Splatting [18] fails to regularize the strong approximation power of point-wise Gaussians at training views, and yields obvious needle-like artifacts at novel views." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 562, + 222, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 562, + 222, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 562, + 222, + 574 + ], + "type": "text", + "content": "5.4 Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 265, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 265, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 265, + 665 + ], + "type": "text", + "content": "Environment Modeling We study different strategies of environment modeling for NVS in-the-wild, namely leaving the blank pixels with zeros, filling the blanks with a default learnable feature (as is done in RPBG), and" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 274, + 578, + 375, + 634 + ], + "blocks": [ + { + "bbox": [ + 274, + 578, + 375, + 634 + ], + "lines": [ + { + "bbox": [ + 274, + 578, + 375, + 634 + ], + "spans": [ + { + "bbox": [ + 274, + 578, + 375, + 634 + ], + "type": "image", + "image_path": "4921670c624329e449c60f71af44f725a85dfa9c3d7fef56b2841a5d3a75c0cc.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 635, + 353, + 643 + ], + "lines": [ + { + "bbox": [ + 302, + 635, + 353, + 643 + ], + "spans": [ + { + "bbox": [ + 302, + 635, + 353, + 643 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 302, + 635, + 353, + 643 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 302, + 635, + 353, + 643 + ], + "type": "text", + "content": " Env. Modeling" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 377, + 578, + 477, + 634 + ], + "blocks": [ + { + "bbox": [ + 377, + 578, + 477, + 634 + ], + "lines": [ + { + "bbox": [ + 377, + 578, + 477, + 634 + ], + "spans": [ + { + "bbox": [ + 377, + 578, + 477, + 634 + ], + "type": "image", + "image_path": "0b807bd6eaabce4251f38cd77661aa716bc13bbd0f006ac17f09b808cfcd4cca.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 634, + 453, + 642 + ], + "lines": [ + { + "bbox": [ + 403, + 634, + 453, + 642 + ], + "spans": [ + { + "bbox": [ + 403, + 634, + 453, + 642 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 403, + 634, + 453, + 642 + ], + "type": "inline_equation", + "content": "w / \\mathrm{Env}" + }, + { + "bbox": [ + 403, + 634, + 453, + 642 + ], + "type": "text", + "content": " Modeling" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 270, + 647, + 481, + 669 + ], + "lines": [ + { + "bbox": [ + 270, + 647, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 270, + 647, + 481, + 669 + ], + "type": "text", + "content": "Fig.6: The impact of leveraging the default feature vector for environment modeling in an unbounded scene." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "text", + "content": "Q. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 168, + 304, + 227 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 306, + 158 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 306, + 158 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 306, + 158 + ], + "type": "text", + "content": "Table 2: Ablation on network designs on T&T dataset [19]. The reported inference time and memory consumption is tested with a target resolution of " + }, + { + "bbox": [ + 131, + 114, + 306, + 158 + ], + "type": "inline_equation", + "content": "{1920} \\times {1080}" + }, + { + "bbox": [ + 131, + 114, + 306, + 158 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 168, + 304, + 227 + ], + "lines": [ + { + "bbox": [ + 135, + 168, + 304, + 227 + ], + "spans": [ + { + "bbox": [ + 135, + 168, + 304, + 227 + ], + "type": "table", + "html": "
MethodTime(s)Mem.(GB)PSNR↑SSIM↑LPIPS↓
Baseline1.035.4920.010.6660.390
+Multi-scale fusion1.6212.6721.880.7010.337
+DAC1.6713.0122.920.7690.320
+FFT loss--22.940.7940.257
+Multi-stage1.7717.1421.750.7130.338
+U-Net1.5810.3220.850.6920.364
", + "image_path": "3203be77f927afba80d6c5b91e101a0bb9549f19b6efca04192eb29a7dc1c391.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 311, + 179, + 480, + 217 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 3: Ablation on environment modeling on two unbounded scenes, namely Caterpillar of T&T dataset [19] and sky of Free dataset [44]. #Points " + }, + { + "bbox": [ + 307, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "(\\times 10^{6})" + }, + { + "bbox": [ + 307, + 114, + 482, + 170 + ], + "type": "text", + "content": " represent the equivalent overheads when applying different strategies." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 311, + 179, + 480, + 217 + ], + "lines": [ + { + "bbox": [ + 311, + 179, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 311, + 179, + 480, + 217 + ], + "type": "table", + "html": "
Env.Caterpillarsky
#PointsPSNR↑SSIM↑LPIPS↓#PointsPSNR↑SSIM↑LPIPS↓
Zeros7.9519.890.4560.40317.4921.770.5980.315
Learnable7.9521.780.6870.28917.4924.810.8660.199
Sphere8.9521.680.6860.28318.4925.150.8690.179
", + "image_path": "6c30e13af5ac60ecb50f3d99b650ff9d936ff37698e79b82362bc107874c1ff5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 136, + 239, + 218, + 285 + ], + "blocks": [ + { + "bbox": [ + 136, + 239, + 218, + 285 + ], + "lines": [ + { + "bbox": [ + 136, + 239, + 218, + 285 + ], + "spans": [ + { + "bbox": [ + 136, + 239, + 218, + 285 + ], + "type": "image", + "image_path": "a7d1c1353fe2b69baabb0ea37f7774c619a9c75768268a61c948a0bc1b7b0d09.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 136, + 285, + 218, + 339 + ], + "blocks": [ + { + "bbox": [ + 136, + 285, + 218, + 339 + ], + "lines": [ + { + "bbox": [ + 136, + 285, + 218, + 339 + ], + "spans": [ + { + "bbox": [ + 136, + 285, + 218, + 339 + ], + "type": "image", + "image_path": "760a399218dc190a30b2d8acffc7b95906cdb04f56139e0d7e4adcc6b1ca2ba3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 339, + 187, + 344 + ], + "lines": [ + { + "bbox": [ + 164, + 339, + 187, + 344 + ], + "spans": [ + { + "bbox": [ + 164, + 339, + 187, + 344 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 164, + 339, + 187, + 344 + ], + "type": "inline_equation", + "content": "w / \\theta" + }, + { + "bbox": [ + 164, + 339, + 187, + 344 + ], + "type": "text", + "content": " DAC" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 354, + 305, + 388 + ], + "lines": [ + { + "bbox": [ + 131, + 354, + 305, + 388 + ], + "spans": [ + { + "bbox": [ + 131, + 354, + 305, + 388 + ], + "type": "text", + "content": "Fig.7: A challenging case in Train, which is a zoomed-in view with complex occlusion and thin objects (e.g., the handrails)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 219, + 239, + 301, + 286 + ], + "blocks": [ + { + "bbox": [ + 219, + 239, + 301, + 286 + ], + "lines": [ + { + "bbox": [ + 219, + 239, + 301, + 286 + ], + "spans": [ + { + "bbox": [ + 219, + 239, + 301, + 286 + ], + "type": "image", + "image_path": "926f160a17c7c29ef5b47d8149db76a5d7294691b6471c62b216249e03995a58.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 219, + 286, + 301, + 339 + ], + "blocks": [ + { + "bbox": [ + 219, + 286, + 301, + 339 + ], + "lines": [ + { + "bbox": [ + 219, + 286, + 301, + 339 + ], + "spans": [ + { + "bbox": [ + 219, + 286, + 301, + 339 + ], + "type": "image", + "image_path": "4883f9a90237c97e80fcffc58f0920de95e37f1384ea51c16f0d3e0c7f17ab38.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 339, + 270, + 344 + ], + "lines": [ + { + "bbox": [ + 249, + 339, + 270, + 344 + ], + "spans": [ + { + "bbox": [ + 249, + 339, + 270, + 344 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 249, + 339, + 270, + 344 + ], + "type": "inline_equation", + "content": "w / \\mathrm{DAC}" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 335, + 482, + 390 + ], + "lines": [ + { + "bbox": [ + 307, + 335, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 307, + 335, + 482, + 390 + ], + "type": "text", + "content": "Fig. 8: Re-rendering quality (PSNR) with points of different quality on Church. RPBG manages to maintain its robustness against different levels of point sparsity with a " + }, + { + "bbox": [ + 307, + 335, + 482, + 390 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 307, + 335, + 482, + 390 + ], + "type": "text", + "content": " drop in PSNR (NPBG [1]: " + }, + { + "bbox": [ + 307, + 335, + 482, + 390 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 307, + 335, + 482, + 390 + ], + "type": "text", + "content": "; NPBG++ [30]: " + }, + { + "bbox": [ + 307, + 335, + 482, + 390 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 307, + 335, + 482, + 390 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 313, + 240, + 479, + 327 + ], + "blocks": [ + { + "bbox": [ + 313, + 240, + 479, + 327 + ], + "lines": [ + { + "bbox": [ + 313, + 240, + 479, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 479, + 327 + ], + "type": "image", + "image_path": "cb069b7936944b5719e3c3ad478cc6f7c774a84f3db6cdd8334321284b607fc6.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 416, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 416, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 416, + 482, + 453 + ], + "type": "text", + "content": "wrapping the points with a sphere of " + }, + { + "bbox": [ + 130, + 416, + 482, + 453 + ], + "type": "inline_equation", + "content": "10^{6}" + }, + { + "bbox": [ + 130, + 416, + 482, + 453 + ], + "type": "text", + "content": " points (equivalent to an environment map in [33]). The quantitative results are in Sec. 5.4, where the strategy incorporated by RPBG boosts the performance by a large margin, with almost no additional overhead." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 469, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 469, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 469, + 482, + 601 + ], + "type": "text", + "content": "Network Architecture and Loss We study the key components in the rendering network in Sec. 5.4. The baseline refers to NPBG [1] trained per-scene from scratch with " + }, + { + "bbox": [ + 130, + 469, + 482, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 130, + 469, + 482, + 601 + ], + "type": "text", + "content": " and VGG loss. We compare three fundamental architectures for image restoration (with necessary modifications on the first several layers), i.e., multi-scale fusion [8], multistage [52], and U-Net [6] following the taxonomy in [6]. Compared to the baseline, all modern restoration-oriented architectures bring remarkable improvement to the overall performance, and [8] is opted for a better balance between performance and time/memory efficiency. The gain by DAC is also considerable while the FFT loss mainly improves the perceptual quality. By the case in Fig. 7, we show the necessity of the DAC module when re-rendering challenging cases without in particular handling the erroneous point visibility." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "text", + "content": "Triangulation Configurations We also evaluate RPBG with different triangulation configurations. Indicated by the results in Fig. 8, RPBG witnesses a total drop of " + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "inline_equation", + "content": "3.29\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "text", + "content": " in re-rendering PSNR when switching the triangulation from ground truth (GT) to randomly initialized points (Random). Note that even with sparse points (SfM)," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "type": "text", + "content": "RPBG: Towards Robust Neural Point-based Graphics in the Wild" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": "which are usually yielded as a by-product when computing camera parameters, RPBG (21.41 dB) still outperforms " + }, + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "inline_equation", + "content": "\\mathrm{F}^2" + }, + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": "-NeRF (20.57 dB)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 158, + 206, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 158, + 206, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 158, + 206, + 171 + ], + "type": "text", + "content": "6 Discussion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 303 + ], + "type": "text", + "content": "Effectiveness of RPBG We would like to shed some light on the effectiveness of RPBG. The first factor should be the point representation. The triangulated 3D points have already contained all the verified co-visibility that is required for NVS. Compared to RF-based methods, which often struggle at local minima, such triangulation-as-parameterization paradigm greatly lower the complexity of searching, and is unified across different scene types. Besides, the patch-wise rendering scheme mentioned in Sec. 4.4 also improves the perceptual quality of re-rendering while RF-based methods apply pixel/ray-wise volume rendering, where the correlation between pixels are not modeled explicitly. For more discussions on the insights, we strongly refer the readers to the Supplementary Material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 315, + 481, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 434 + ], + "type": "text", + "content": "Limitations An obvious problem with point-based re-rendering (RPBG and NPBG [1]) is that compared to lightweight RF-based variants [5, 27], they take more space to store the CNN parameters and the neural texture. Although it makes the scaling-up easier, the number of parameters grows linearly with the point cloud scale. Besides, due to the non-local neural renderer we employ, each point is encoded with visual context of a larger range, hindering the editability of RPBG. We have also noticed that, the CNN-based rendering scheme can lead to unsatisfactory temporal consistency especially when the triangulation is patchy so that we mainly count on the inpainting ability of the neural render for yielding re-renderings. Such flicker issue, as well as its potential solutions, has been discussed in [9]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 453, + 210, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 210, + 466 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 210, + 466 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 479, + 481, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 610 + ], + "type": "text", + "content": "In this paper, we present RPBG as a robust and practical alternative to NPBG, a baseline of point-based NVs methods, performing neural re-rendering on triangulated points. We analyze the key problems in NPBG, when attempting to generalize to more generic scenes other than the well-captured scans, and reform the pipeline to reveal the real potential of point-based graphics. Respectively motivated and inspired by RF-based methods and low-level image restoration methods, we reform the pipeline according to our analysis. By extensive experiments on diverse datasets, RPBG achieves stably superior results over state-of-the-art RF-based and point-based NVs methods, especially on the metrics with more attention paid to the nuances of human visual perception, without case-by-case parameterization across all scenes, indicating its robustness and generalizability." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "Q. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 229, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 229, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 229, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 162 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 162 + ], + "type": "text", + "content": "This research was supported in part by JSPS KAKENHI Grant Numbers 24K22318, 22H00529, 20H05951, and JST-Mirai Program JPMJMI23G1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 178, + 190, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 178, + 190, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 178, + 190, + 192 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 201, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 201, + 482, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 201, + 482, + 234 + ], + "spans": [ + { + "bbox": [ + 138, + 201, + 482, + 234 + ], + "type": "text", + "content": "1. Aliev, K.A., Sevastopolsky, A., Kolos, M., Ulyanov, D., Lempitsky, V.: Neural point-based graphics. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 696-712. Springer (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 235, + 482, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 235, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 235, + 482, + 266 + ], + "type": "text", + "content": "2. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855–5864 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 267, + 482, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 482, + 299 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 482, + 299 + ], + "type": "text", + "content": "3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 300, + 482, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 300, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 482, + 319 + ], + "type": "text", + "content": "4. Bui, G., Le, T., Morago, B., Duan, Y.: Point-based rendering enhancement via deep learning. The Visual Computer 34, 829-841 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 321, + 482, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 482, + 342 + ], + "type": "text", + "content": "5. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: European Conference on Computer Vision. pp. 333-350. Springer (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 342, + 482, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 482, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 482, + 363 + ], + "type": "text", + "content": "6. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: European Conference on Computer Vision. pp. 17-33. Springer (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 364, + 482, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 364, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 138, + 364, + 482, + 384 + ], + "type": "text", + "content": "7. Chi, L., Jiang, B., Mu, Y.: Fast fourier convolution. Advances in Neural Information Processing Systems 33, 4479-4488 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 384, + 482, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 384, + 482, + 417 + ], + "spans": [ + { + "bbox": [ + 138, + 384, + 482, + 417 + ], + "type": "text", + "content": "8. Cho, S.J., Ji, S.W., Hong, J.P., Jung, S.W., Ko, S.J.: Rethinking coarse-to-fine approach in single image deblurring. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4641-4650 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 418, + 482, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 482, + 449 + ], + "type": "text", + "content": "9. Dai, P., Zhang, Y., Li, Z., Liu, S., Zeng, B.: Neural point cloud rendering via multi-plane projection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7830-7839 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 450, + 482, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 450, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 138, + 450, + 482, + 471 + ], + "type": "text", + "content": "0. Dosovitskiy, A., Brox, T.: Generating images with perceptual similarity metrics based on deep networks. Advances in neural information processing systems 29 (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 472, + 482, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 472, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 134, + 472, + 482, + 502 + ], + "type": "text", + "content": "11. ETH Zurich Computer Vision Group and Microsoft Mixed Reality & AI Lab Zurich: The ETH-Microsoft Localization Dataset. https://github.com/cvg/visloc-iccv2021 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 503, + 482, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 503, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 134, + 503, + 482, + 536 + ], + "type": "text", + "content": "12. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 536, + 482, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 536, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 134, + 536, + 482, + 568 + ], + "type": "text", + "content": "13. Fuoli, D., Van Gool, L., Timofte, R.: Fourier space losses for efficient perceptual image super-resolution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2360-2369 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 569, + 482, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 569, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 134, + 569, + 482, + 589 + ], + "type": "text", + "content": "14. Furukawa, Y., Ponce, J.: Accurate, dense, and robust multiview stereopsis. IEEE transactions on pattern analysis and machine intelligence 32(8), 1362-1376 (2009)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 590, + 367, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 590, + 367, + 600 + ], + "spans": [ + { + "bbox": [ + 134, + 590, + 367, + 600 + ], + "type": "text", + "content": "15. Gross, M., Pfister, H.: Point-based graphics. Elsevier (2011)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 134, + 601, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 601, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 601, + 482, + 632 + ], + "type": "text", + "content": "16. Grossman, J.P., Dally, W.J.: Point sample rendering. In: Rendering Techniques' 98: Proceedings of the Eurographics Workshop in Vienna, Austria, June 29—July 1, 1998 9. pp. 181-192. Springer (1998)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 134, + 633, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 482, + 665 + ], + "type": "text", + "content": "17. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and superresolution. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 694-711. Springer (2016)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "type": "text", + "content": "RPBG: Towards Robust Neural Point-based Graphics in the Wild" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "18. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG) 42(4), 1-14 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "text", + "content": "19. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG) 36(4), 1-13 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 404, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 404, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 404, + 172 + ], + "type": "text", + "content": "20. Levoy, M., Whitted, T.: The use of points as a display primitive (1985)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 173, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 173, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 173, + 481, + 193 + ], + "type": "text", + "content": "21. Li, Z., Li, L., Zhu, J.: Read: Large-scale neural scene rendering for autonomous driving. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1522-1529 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 194, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 215 + ], + "type": "text", + "content": "22. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 216, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 236 + ], + "type": "text", + "content": "23. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. International journal of computer vision 60, 91-110 (2004)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 237, + 481, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 269 + ], + "type": "text", + "content": "24. Lu, C., Yin, F., Chen, X., Chen, T., Yu, G., Fan, J.: A large-scale outdoor multi-modal dataset and benchmark for novel view synthesis and implicit scene reconstruction. arXiv preprint arXiv:2301.06782 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "type": "text", + "content": "25. Meshry, M., Goldman, D.B., Khamis, S., Hoppe, H., Pandey, R., Snavely, N., Martin-Brualla, R.: Neural rerendering in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6878-6887 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "text", + "content": "26. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: European Conference on Computer Vision. pp. 405-421 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "type": "text", + "content": "27. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "text", + "content": "28. Pfister, H., Zwicker, M., Van Baar, J., Gross, M.: Surfels: Surface elements as rendering primitives. In: Proceedings of the 27th annual conference on Computer graphics and interactive techniques. pp. 335-342 (2000)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 391, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 481, + 423 + ], + "type": "text", + "content": "29. Pittaluga, F., Koppal, S.J., Kang, S.B., Sinha, S.N.: Revealing scenes by inverting structure from motion reconstructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 145-154 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "text", + "content": "30. Rakhimov, R., Ardelean, A.T., Lempitsky, V., Burnaev, E.: Npbg++: Accelerating neural point-based graphics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15969-15979 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 457, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 489 + ], + "type": "text", + "content": "31. Ramon, E., Triginer, G., Escur, J., Pumarola, A., Garcia, J., Giro-i Nieto, X., Moreno-Noguer, F.: H3d-net: Few-shot high-fidelity 3d head reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5620-5629 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 490, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 481, + 533 + ], + "type": "text", + "content": "32. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "type": "text", + "content": "33. Ruckert, D., Franke, L., Stamminger, M.: Adop: Approximate differentiable one-pixel point rendering. ACM Transactions on Graphics (ToG) 41(4), 1-14 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 555, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 577 + ], + "type": "text", + "content": "34. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4104-4113 (2016)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 578, + 481, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 598 + ], + "type": "text", + "content": "35. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "text", + "content": "36. Sun, C., Sun, M., Chen, H.T.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5459-5469 (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "37. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., Rabinovich, A.: Going deeper with convolutions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1-9 (2015)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "Q. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "38. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8248–8258 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 162, + 481, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 481, + 195 + ], + "type": "text", + "content": "39. Tancik, M., Weber, E., Ng, E., Li, R., Yi, B., Wang, T., Kristoffersen, A., Austin, J., Salahi, K., Ahuja, A., et al.: Nerfstudio: A modular framework for neural radiance field development. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-12 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 196, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 196, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 481, + 228 + ], + "type": "text", + "content": "40. Tewari, A., Fried, O., Thies, J., Sitzmann, V., Lombardi, S., Sunkavalli, K., Martin-Brualla, R., Simon, T., Saragih, J., Nießner, M., et al.: State of the art on neural rendering. In: Computer Graphics Forum. vol. 39, pp. 701-727. Wiley Online Library (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 229, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 481, + 251 + ], + "type": "text", + "content": "41. Thies, J., Zollhöfer, M., Nießner, M.: Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG) 38(4), 1-12 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 251, + 481, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 481, + 284 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 481, + 284 + ], + "type": "text", + "content": "42. Turki, H., Ramanan, D., Satyanarayanan, M.: Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12922–12931 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 285, + 481, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 481, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 481, + 318 + ], + "type": "text", + "content": "43. Wang, F., Galliani, S., Vogel, C., Speciale, P., Pollefeys, M.: Patchmatchnet: Learned multiview patchmatch stereo. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14194-14203 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "type": "text", + "content": "44. Wang, P., Liu, Y., Chen, Z., Liu, L., Liu, Z., Komura, T., Theobalt, C., Wang, W.: F2-nerf: Fast neural radiance field training with free camera trajectories. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4150–4159 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 353, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 481, + 385 + ], + "type": "text", + "content": "45. Wang, X., Zhu, Z., Huang, G., Qin, F., Ye, Y., He, Y., Chi, X., Wang, X.: Mvster: Epipolar transformer for efficient multi-view stereo. In: European Conference on Computer Vision. pp. 573-591. Springer (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "text", + "content": "46. Wei, Z., Zhu, Q., Min, C., Chen, Y., Wang, G.: Aa-rmvvsnet: Adaptive aggregation recurrent multi-view stereo network. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6187-6196 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 420, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 481, + 441 + ], + "type": "text", + "content": "47. Wu, M., Wang, Y., Hu, Q., Yu, J.: Multi-view neural human rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1682-1691 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 442, + 481, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 481, + 474 + ], + "type": "text", + "content": "48. Yao, Y., Luo, Z., Li, S., Fang, T., Quan, L.: Mvsnet: Depth inference for unstructured multiview stereo. In: Proceedings of the European conference on computer vision (ECCV). pp. 767-783 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 475, + 481, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 481, + 508 + ], + "type": "text", + "content": "49. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12-22 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 509, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 481, + 542 + ], + "type": "text", + "content": "50. Yu, J., Lin, Z., Yang, J., Shen, X., Lu, X., Huang, T.S.: Free-form image inpainting with gated convolution. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4471-4480 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 543, + 481, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 481, + 576 + ], + "type": "text", + "content": "51. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5728-5739 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 577, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 481, + 609 + ], + "type": "text", + "content": "52. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14821-14831 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 609, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 481, + 643 + ], + "type": "text", + "content": "53. Zhang, J., Zhang, J., Mao, S., Ji, M., Wang, G., Chen, Z., Zhang, T., Yuan, X., Dai, Q., Fang, L.: Gigamvs: a benchmark for ultra-large-scale gigapixel-level 3d reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(11), 7534-7550 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "54. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492 (2020)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 211, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 211, + 91, + 447, + 101 + ], + "type": "text", + "content": "RPGB: Towards Robust Neural Point-based Graphics in the Wild" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 172 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "55. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 172 + ], + "type": "text", + "content": "56. Zhang, Y., Peng, S., Moazeni, A., Li, K.: Papr: Proximity attention point rendering. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Q. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/RS-NeRF_ Neural Radiance Fields from Rolling Shutter Images/cafbceee-b1fc-44b7-b4f6-4691cb686eec_content_list.json b/2024/RS-NeRF_ Neural Radiance Fields from Rolling Shutter Images/cafbceee-b1fc-44b7-b4f6-4691cb686eec_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ab6ebd0b85395738f175893b0077e839b54fa592 --- /dev/null +++ b/2024/RS-NeRF_ Neural Radiance Fields from Rolling Shutter Images/cafbceee-b1fc-44b7-b4f6-4691cb686eec_content_list.json @@ -0,0 +1,2014 @@ +[ + { + "type": "text", + "text": "RS-NeRF: Neural Radiance Fields from Rolling Shutter Images", + "text_level": 1, + "bbox": [ + 228, + 140, + 774, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Muyao Niu, Tong Chen, Yifan Zhan, Zhuoxiao Li, Xiang Ji, and Yinqiang Zheng*", + "bbox": [ + 320, + 213, + 679, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The University of Tokyo \nmuyao.niu@gmail.com, cocottt1023@gmail.com, \nzhan-yifan@g.ecc.u-tokyo.ac.jp, lizhuoxiao@g.ecc.u-tokyo.ac.jp, \nxiangji2016@gmail.com, yqzheng@ai.u-tokyo.ac.jp", + "bbox": [ + 261, + 256, + 738, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Neural Radiance Fields (NeRFs) have become increasingly popular because of their impressive ability for novel view synthesis. However, their effectiveness is hindered by the Rolling Shutter (RS) effects commonly found in most camera systems. To solve this, we present RS-NeRF, a method designed to synthesize normal images from novel views using input with RS distortions. This involves a physical model that replicates the image formation process under RS conditions and jointly optimizes NeRF parameters and camera extrinsic for each image row. We further address the inherent shortcomings of the basic RS-NeRF model by delving into the RS characteristics and developing algorithms to enhance its functionality. First, we impose a smoothness regularization to better estimate trajectories and improve the synthesis quality, in line with the camera movement prior. We also identify and address a fundamental flaw in the vanilla RS model by introducing a multi-sampling algorithm. This new approach improves the model's performance by comprehensively exploiting the RGB data across different rows for each intermediate camera pose. Through rigorous experimentation, we demonstrate that RS-NeRF surpasses previous methods in both synthetic and real-world scenarios, proving its ability to correct RS-related distortions effectively. Codes and data available: https://github.com/MyNiuuu/RS-NeRF", + "bbox": [ + 261, + 353, + 738, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 659, + 375, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "CMOS imaging sensors are commonly employed in a wide range of consumer and industrial products. Many of these sensors utilize a Rolling Shutter (RS) mechanism for image capture because of its cost-effectiveness and portability. In contrast to Global Shutter (GS) cameras, which capture all pixels simultaneously, rolling shutter cameras capture image pixels row by row in a sequential manner. Consequently, if the camera is in motion during image capture, visible distortions can occur. It is widely acknowledged that rolling shutter distortions pose challenges for various downstream computer vision tasks, as indicated by", + "bbox": [ + 212, + 691, + 787, + 815 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 218, + 824, + 382, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "previous research [1,14,15,18,29,30]. Consequently, the correction of rolling shutter effects has attracted significant attention in the past [11, 22, 28, 34, 39-42].", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the emerging field of novel view synthesis, technologies such as NeRF [25] have become increasingly proficient in generating highly realistic renderings. NeRF utilizes a continuous volumetric function that is parameterized by a multilayer perceptron (MLP) to associate 3D locations and 2D view directions with color and density information. It relies on volume rendering techniques to enable differentiable rendering.", + "bbox": [ + 212, + 179, + 785, + 270 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although traditional NeRF models deliver exceptional results when applied to well-calibrated and well-captured images, they encounter difficulties when dealing with RS effects. For example, in scenarios where the camera is in motion during image capture, these RS effects can significantly impact the performance of the model. Given the widespread use of RS cameras in various industrial applications, it is imperative to take into account this phenomenon when reconstructing 3D representations using NeRFs.", + "bbox": [ + 212, + 272, + 787, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One straightforward approach involves a two-stage baseline solution, where the first step corrects the rolling shutter effect in 2D image space, and then the NeRF is trained using these corrected images. While these baseline methods enhance the quality of novel view synthesis in NeRF to some extent by utilizing 2D rolling shutter correction techniques, they can not fully explore the 3D scene's geometry, leading to inaccuracies in correspondences, especially when there is significant camera motion.", + "bbox": [ + 212, + 381, + 787, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce RS-NeRF, a robust framework designed to explicitly incorporate the physics-based formation of rolling shutter effects into the rendering process. RS-NeRF has the ability to reconstruct a high-quality NeRF representation from only input data that contains rolling shutter artifacts. To achieve this, we explicitly model the camera's motion trajectory throughout the exposure duration for each frame and combine the RGB rendering outcomes from different camera poses to generate the rolling shutter image. Throughout this procedure, we optimize both the NeRF network and the pose estimation network simultaneously by minimizing the photometric loss.", + "bbox": [ + 212, + 489, + 787, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building upon the basic RS-NeRF framework, we delve into its shortcomings and enhance its functionality through various refinements. Initially, we impose a camera trajectory smoothness regularization to refine our model's trajectory prediction, by leveraging the sequential nature of camera movement to yield improved results. We further identify a critical shortcoming due to the vanilla RS camera motion model and introduce a multi-sampling technique that significantly enhances RS-NeRF's performance. Specifically, we reveal the issue of the standard camera pose formulation, which assigns distinct poses per row, potentially limiting RS-NeRF's efficiency due to limited training data for each pose. We then make a key observation that multiple poses may share the same 2D RGB observation for the same 3D point during the exposure time, which can drastically increase the available training data. To capitalize on this insight, we introduce and compute the PP-ratio—a measure of camera pose movement in 3D space relative to pixel movement in 2D space, and apply this to broaden our", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "M. Niu et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/29aee7f762130fd74118c5611f13df3e3cc1fcf90187e89e22bfd3109dce19e1.jpg", + "image_caption": [ + "Fig. 1: Pipeline of vanilla RS-NeRF. Given a series of images with RS distortion, RS-NeRF learns the underlying normal 3D representations by jointly estimating pose and RGB values for each row. For each RS image, we first sample the rows and obtain their camera poses via the pose estimation network. We then sample points for each row and feed them with the estimated camera pose to the NeRF network and optimize the photo-metric loss with ground truths." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 782, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "training data by sampling out-of-row pixel points. This approach significantly bolsters our model's performance by tapping into the inherent data augmentation possibilities presented by the RS formation process.", + "bbox": [ + 212, + 376, + 782, + 421 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We collect both synthetic and real data to evaluate our proposed RS-NeRF. The experimental results reveal that RS-NeRF outperforms the preceding models and demonstrate the effectiveness of our proposed refinements on the vanilla model. In summary, our contributions are as follows:", + "bbox": [ + 212, + 422, + 784, + 481 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a framework that can reconstruct NeRF from inputs with Rolling Shutter (RS) effects by modeling the camera trajectory and aggregating RGB rendering results of different poses during the exposure time.", + "- By analyzing the vanilla framework in detail, we proposed a series of improvements to it, including trajectory smoothness regularization and the multi-sampling algorithm.", + "- We contribute synthetic and real datasets to evaluate our framework and future works." + ], + "bbox": [ + 225, + 494, + 782, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 640, + 387, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2D Rolling Shutter Correction. Rolling Shutter Correction (RSC) aims to restore normal images given the input with RS distortions. In recent years, several appealing image-based RSC methods [22,28,34,39-42] have been developed. These methods take either single RS image or continuous RS frames as input and predict their global shutter (GS) counterparts. However, single-image RSC can be seriously ill-posed and struggle to correctly estimate the GS counterpart in complex scenarios. To reduce the ill-posedness of single-image RSC, more attention has been paid to multi-image RSC [6, 9-11, 22, 40]. However, these multi-image RS correction methods, which are built on image space operations, fail to exploit the 3D geometry of the scene. As a result, these methods produce multi-view inconsistency results which largely impact the reconstruction results", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "RS-NeRF", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of NeRF. In contrast, our proposed method can jointly reconstruct 3D geometry and correct RS images by effectively aggregating multi-view information of the scene.", + "bbox": [ + 212, + 146, + 782, + 190 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Neural Radiance Fields (NeRF). NeRF [25] is an end-to-end model that synthesizes novel views of complex 3D scenes by optimizing an implicit continuous volumetric scene function encoded by an MLP. Due to its simplicity in training and high-quality rendering results, NeRF-based methods [3,4,7,12,21,27,35,37] have attracted intensive attention from various computer vision fields, leading to significant improvements in terms of both accuracy and efficiency. NeRF-based models have also inspired many subsequent works that extend their continuous neural representation to different setups [2,8,13,17,24,36]. Deblur-NeRF [24] and BAD-NeRF [36] reconstruct sharp images given the blurry input from different view directions. One concurrent work [19] appears in ICLR'24, which formulates the RS process with B-cubic models. This work combines the basic modeling of camera motion with the NeRF network to consider the RS effect, without further technical improvements. In this paper, we first bake the RS model into NeRF formulation, then analyze several inherit limitations of the vanilla model, and propose a series of algorithms to further improve its performance.", + "bbox": [ + 212, + 191, + 787, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 RS-NeRF", + "text_level": 1, + "bbox": [ + 215, + 441, + 346, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we present the specifics of RS-NeRF. Initially, we explain the basic formulation of RS-NeRF in Sec. 3.1. We also note that the vanilla model possesses specific shortcomings that hinder its effectiveness. To address these, we suggest several improvements to enhance its capabilities, which are detailed in Sec. 3.2 and Sec. 3.3.", + "bbox": [ + 212, + 474, + 784, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Vanilla RS-NeRF Model", + "text_level": 1, + "bbox": [ + 215, + 574, + 465, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rolling shutter camera model. Unlike Global Shutter (GS) cameras, the Rolling Shutter (RS) camera captures each row at distinct timestamps. Without loss of generality, we consider the readout direction of the RS camera to be from top to bottom, assuming an infinitesimal exposure time for each row, following existing RS modeling techniques [10, 11, 22]. This mechanism can be described through a mathematical model as follows:", + "bbox": [ + 212, + 599, + 782, + 691 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n[ \\mathbf {I} ^ {R} (\\mathbf {x}) ] _ {i} = [ \\mathbf {I} _ {i} ^ {G} (\\mathbf {x}) ] _ {i}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 434, + 704, + 784, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $[\\mathbf{I}(\\mathbf{x})]_i$ is an operator to extract the $i$ -th row from an image $\\mathbf{I}(\\mathbf{x})$ , $\\mathbf{I}^R(\\mathbf{x}) \\in \\mathbb{R}^{H \\times W \\times 3}$ is the rolling shutter image, and $\\mathbf{I}_i^G(\\mathbf{x}) \\in \\mathbb{R}^{H \\times W \\times 3}$ is the virtual global shutter image captured at the same pose as the $i$ -th row of $\\mathbf{I}^R(\\mathbf{x})$ . As a result, for one RS image $\\mathbf{I}^R(\\mathbf{x})$ , we denote the pose of the $i$ -th row as $\\mathbf{T}_i$ .", + "bbox": [ + 212, + 733, + 782, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Camera motion trajectory modeling. In our rolling shutter camera model, it is necessary to establish the poses for each row in each image. Given the typically short exposure time of an RS image, we use a linear model to approximate the", + "bbox": [ + 212, + 795, + 784, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "M. Niu et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/eabc1255cf41091543613c6aa07177d3d28ab0faca1e30bcaecff64dd6293235.jpg", + "image_caption": [ + "Fig. 2: Smoothness regularization on camera trajectory. For each pair of adjacent trajectory vectors $\\vec{\\mathbf{d}}_k$ and $\\vec{\\mathbf{d}}_{k + 1}$ , we compute the mid-point trajectory vector $\\vec{\\mathbf{d}}_{mid}$ and their unit vectors $\\vec{\\mathbf{n}}_k$ , $\\vec{\\mathbf{n}}_{k + 1}$ , and $\\vec{\\mathbf{n}}_{mid}$ . We then apply the $L_{2}$ regularization between $\\vec{\\mathbf{n}}_{mid}$ and mean $(\\vec{\\mathbf{n}}_k,\\vec{\\mathbf{n}}_{k + 1})$ ." + ], + "image_footnote": [], + "bbox": [ + 223, + 147, + 493, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "motion of the camera. Specifically, we start by defining two camera poses: one for the first row, denoted as $\\mathbf{T}_0$ , and one for the last row, $\\mathbf{T}_{H - 1}$ . For the rows in between, we linearly interpolate their poses within the Lie algebra of $\\mathbf{SE}(3)$ as informed by prior research [20, 23, 36]:", + "bbox": [ + 212, + 290, + 782, + 351 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} _ {i} = \\mathbf {T} _ {0} \\cdot (1 - \\rho) + \\mathbf {T} _ {H - 1} \\cdot \\rho , \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 362, + 785, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the interpolation coefficient $\\rho = \\frac{i}{H - 1}$ is based on the row index $i$ . Thus, for each RS image, RS-NeRF receives one pose initialization, then optimizes the poses at both ends and then determines the poses for each intermediate row through linear interpolation. This interpolation method for calculating each row's pose is fully differentiable, enabling the optimization of pose parameters using a series of rolling shutter images.", + "bbox": [ + 212, + 390, + 782, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Optimizing NeRF from RS observation. For a set of $K$ rolling shutter images, we select $M$ rows per image and $N$ points per row, resulting in $K \\times M \\times N$ distinct camera rays for each iteration. These rays are used to concurrently optimize the NeRF network and the pose parameters. Specifically, for each image, we initially use the row indices to access the pose estimation network. This allows us to determine the camera ray $\\hat{\\mathbf{r}}(\\theta_p; u, v)$ for each sampled pixel point $(u, v)$ . Following this, these rays are employed to train the NeRF network by computing the ensuing photometric loss:", + "bbox": [ + 212, + 481, + 787, + 602 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\left| \\left| \\mathbf {C} (\\mathbf {r} (u, v)) - \\hat {\\mathbf {C}} \\left(\\theta_ {n}; \\hat {\\mathbf {r}} \\left(\\theta_ {p}; (u, v)\\right) \\right. \\right| \\right| _ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 613, + 784, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta_{n}$ represents the trainable parameters of the NeRF network, and $\\theta_{p}$ denotes the learnable parameters for camera poses. $\\hat{\\mathbf{C}} (\\theta_n;\\hat{\\mathbf{r}} (\\theta_p;(u,v))$ refers to the predicted RGB value for the pixel at coordinates $(u,v)$ , while $\\mathbf{C}(\\mathbf{r}(u,v))$ denotes the corresponding RGB ground truth value. The workflow of our vanilla RS-NeRF model is depicted in Fig. 1.", + "bbox": [ + 212, + 640, + 787, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Trajectory Smoothness Regularization.", + "text_level": 1, + "bbox": [ + 214, + 739, + 581, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the previous description of our model, it is evident that accurately estimating the camera pose for each row in RS frames is crucial for effectively reconstructing the normal NeRF representation. We observe that the basic RS-NeRF struggles with precise camera trajectory estimation, often resulting in high-frequency predictions, as further analyzed in our experiments. This issue", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "RS-NeRF", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a76f235469d71677fee3efc7000efff60cc89cdf6fe437688ed3e509e7c5a952.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
NposeON TRAJ. / OUT TRAJ.
PSNR ↑SSIM ↑LPIPS ↓
10024.36/25.16.7321/.7727.0703/.0578
40024.06/24.66.7245/.7536.0697/.0579
", + "bbox": [ + 217, + 151, + 457, + 213 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1: Quantitative results for different $N_{\\text{pose}}$ . $N_{\\text{pose}} = 100$ yields better results than $N_{\\text{pose}} = 400$ .", + "bbox": [ + 215, + 218, + 457, + 262 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a958c75ed494c3d7fceb1b81c4939f267a7da9f8a257120e2a3579d22b9f3639.jpg", + "image_caption": [ + "Fig. 3: Qualitative comparisons for different $N_{\\text{pose}}$ . The artifact decreases as $N_{\\text{pose}}$ grows." + ], + "image_footnote": [], + "bbox": [ + 460, + 143, + 784, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "arises because the network focuses only on the photometric loss for individual images, neglecting the consistent movement characteristics of the camera across different RS frames. To address this, we introduce a camera trajectory smoothness regularization, based on the expectation of a smooth and continuous camera trajectory in 3D space. Specifically, for every two adjacent RS frames $\\mathbf{I}_k^R$ and $\\mathbf{I}_{k + 1}^R$ , we aim to minimize the following loss during the optimization process:", + "bbox": [ + 212, + 301, + 784, + 393 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} = | | \\vec {\\mathbf {n}} _ {m i d} - \\operatorname {m e a n} (\\vec {\\mathbf {n}} _ {k} + \\vec {\\mathbf {n}} _ {k + 1}) | | _ {2}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 402, + 785, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, $\\vec{\\mathbf{n}}_k$ and $\\vec{\\mathbf{n}}_{k + 1}$ represent the unit direction vectors of the estimated camera trajectories for RS frames $\\mathbf{I}_k^R$ and $\\mathbf{I}_{k + 1}^R$ , respectively. Meanwhile, $\\vec{\\mathbf{n}}_{mid}$ denotes the unit direction vector for the camera trajectory during the exposure interval between these two RS frames.", + "bbox": [ + 212, + 430, + 787, + 492 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Multi-Sampling Algorithm", + "text_level": 1, + "bbox": [ + 215, + 513, + 480, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As discussed in Sec. 3.1, each row in an RS frame is assigned a unique camera pose. For example, in an RS frame with 400 rows, there are 400 distinct camera poses. This rigorous approach aligns with the real-world, physics-based photography process of RS cameras but also presents certain challenges. For example, under this framework, the data points for each pose are confined to a single row. This limitation can significantly affect our model's performance due to the sparse training data available for each pose.", + "bbox": [ + 212, + 537, + 787, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A straightforward solution to this problem is to reduce the number of poses per image. For example, assigning 200 poses to each RS frame means that all two adjacent rows would share the same pose. This approach can lead to blurriness since these rows are captured at different poses, but it also doubles the amount of data available for training each pose. To further explore this trade-off, we conducted a series of experiments. Fig. 3 shows qualitative results for different intermediate pose numbers $N_{\\text{pose}}$ , indicating that the artifacts decrease as $N_{\\text{pose}}$ increases. Interestingly, we observed that using 100 poses per frame often yields results comparable to or even better than using 400 poses. The quantitative results presented in Tab. 1 also support the superiority of having 100 poses per frame.", + "bbox": [ + 212, + 643, + 787, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This observation seems counter-intuitive at first, as one might assume that having fewer than 400 poses per frame would lead to poorer results due to", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "M. Niu et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3518bdd366c104f04ccdaff54f8406ce473f531de5996e9e5498d56150131e06.jpg", + "image_caption": [ + "Fig. 4: Motivation for the multi-sampling algorithm. Up: Three consecutive RS frames, labeled as (a), (b), and (c), along with the optical flow between each pair of frames, shown in (d) and (e). Down: For a single point in 3D space, we observe its 2D projections in the first (marked in yellow) and second (marked in red) RS frames. These projections exhibit a row displacement of 8 and a column displacement of 18, as depicted in (f). However, during this interval, the camera undergoes a shift of at least 392 poses, as indicated in (g)." + ], + "image_footnote": [], + "bbox": [ + 274, + 145, + 727, + 305 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "artifacts like blurriness. However, it can be rationally explained. As demonstrated in Fig. 4, for a given point in 3D space, the 2D projections on the first and second RS frames show relatively small pixel displacements (8 in row and 12 in column). Meanwhile, the camera shifts through at least 392 poses in 3D space. This is based on the assumption that there is no overlap in the trajectories of two adjacent frames (indicated as $\\geq 0$ poses' in the figure), which is naturally satisfied given the sparse observation setting and short exposure times typically involved.", + "bbox": [ + 212, + 452, + 787, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following our prior analysis, we introduced a multi-sampling algorithm aimed at explicitly increasing the sample points for each intermediate camera pose. The fundamental concept involves computing the ratio of camera pose displacement to pixel displacement (termed the PP-ratio) between two adjacent frames for the same 3D point. For example, a pixel displacement of 18 combined with a camera pose shift of 392 results in a PP-ratio of 21. This implies that at least 21 consecutive camera poses share the same pixel value, which means that all these 21 camera poses can utilize this RGB value for training purposes.", + "bbox": [ + 212, + 577, + 787, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For a pixel $\\mathbf{x} = (u,v)$ located in the $u$ -th column and $v$ -th row of the $k$ -th RS frame $\\mathbf{I}_k^R (\\mathbf{x})$ , we determine both the forward PP-ratio $\\rho_{k\\to k + 1}(\\mathbf{x})$ and the backward PP-ratio $\\rho_{k\\to k - 1}(\\mathbf{x})$ . This is because the point represented by $\\mathbf{x}$ can transition from $\\mathbf{I}_k^R (\\mathbf{x})$ to $\\mathbf{I}_{k + 1}^R (\\mathbf{x})$ or $\\mathbf{I}_{k - 1}^R (\\mathbf{x})$ . Consequently, all camera poses within the range of $[v - \\rho_{k\\to k - 1}(\\mathbf{x}), v + \\rho_{k\\to k + 1}(\\mathbf{x})]$ are capable of sampling the RGB value of $\\mathbf{x}$ for training purposes.", + "bbox": [ + 212, + 700, + 787, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To compute the forward PP-ratio $\\rho_{k\\rightarrow k + 1}(\\mathbf{x})$ for a pixel $\\mathbf{x}$ in the $k$ -th RS frame $\\mathbf{I}_k^R (\\mathbf{x})$ , the first step is to estimate the forward optical flow $F_{k\\rightarrow k + 1}(\\mathbf{x})$ , utilizing RAFT [33]. Consequently, the row displacement $\\varDelta_{k\\rightarrow k + 1}^{row}(\\mathbf{x})$ and column", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "RS-NeRF", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/edddefe08157b725531a95a409a992c1b2e71d0336d6ad8070f0ff4b3f55ad39.jpg", + "image_caption": [ + "Fig. 5: Comparing the accessible training points in one RS image under different strategies. The image size is $H \\times W$ , providing $H$ poses. Distinct circle hues represent different RGB values. In case $C$ , pixels transported from other poses are with dash contours, and the number signifying that it originates from the $i$ -th pose." + ], + "image_footnote": [], + "bbox": [ + 217, + 142, + 560, + 328 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "displacement $\\varDelta_{k\\rightarrow k+1}^{col}(\\mathbf{x})$ of $\\mathbf{x}$ from $\\mathbf{I}_k^R (\\mathbf{x})$ to $\\mathbf{I}_{k + 1}^R (\\mathbf{x})$ would be the second and the first element of $F_{k\\to k + 1}(\\mathbf{x})$ :", + "bbox": [ + 212, + 354, + 782, + 386 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nF _ {k \\rightarrow k + 1} (\\mathbf {x}) = \\operatorname {R A F T} \\left(\\mathbf {I} _ {k} ^ {R} (\\mathbf {x}), \\mathbf {I} _ {k + 1} ^ {R} (\\mathbf {x})\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 396, + 785, + 415 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\varDelta_ {k \\rightarrow k + 1} ^ {c o l} (\\mathbf {x}), \\varDelta_ {k \\rightarrow k + 1} ^ {r o w} (\\mathbf {x}) = F _ {k \\rightarrow k + 1} (\\mathbf {x}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 361, + 417, + 785, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and the calculation are similar for the backward situation.", + "bbox": [ + 215, + 446, + 635, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We then calculate the forward and backward camera pose shift as:", + "bbox": [ + 238, + 463, + 718, + 478 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\Delta_ {k \\rightarrow k + 1} ^ {\\text {p o s e}} (\\mathbf {x}) = v + \\Delta_ {k \\rightarrow k + 1} ^ {\\text {r o w}} (\\mathbf {x}) + (H - v), (7) \\\\ = H + \\Delta_ {k \\rightarrow k + 1} ^ {\\text {r o w}} (\\mathbf {x}), (8) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 489, + 785, + 527 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\Delta_ {k \\rightarrow k - 1} ^ {\\text {p o s e}} (\\mathbf {x}) = \\left(H - \\left(v + \\Delta_ {k \\rightarrow k - 1} ^ {\\text {r o w}} (\\mathbf {x})\\right)\\right) + v, (9) \\\\ = H - \\Delta_ {k \\rightarrow k - 1} ^ {\\text {r o w}} (\\mathbf {x}). (10) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 530, + 785, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, we calculate forward and backward PP-ratio as:", + "bbox": [ + 238, + 580, + 643, + 594 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\rho_ {k \\rightarrow k + 1} (\\mathbf {x}) = \\frac {\\Delta_ {k \\rightarrow k + 1} ^ {p o s e} (\\mathbf {x})}{\\max \\left(\\Delta_ {k \\rightarrow k + 1} ^ {c o l} (\\mathbf {x}) , \\Delta_ {k \\rightarrow k + 1} ^ {r o w} (\\mathbf {x})\\right)}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 604, + 785, + 642 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\rho_ {k \\rightarrow k - 1} (\\mathbf {x}) = \\frac {\\Delta_ {k \\rightarrow k - 1} ^ {p o s e} (\\mathbf {x})}{\\operatorname* {m a x} \\left(\\Delta_ {k \\rightarrow k - 1} ^ {c o l} (\\mathbf {x}) , \\Delta_ {k \\rightarrow k - 1} ^ {r o w} (\\mathbf {x})\\right)}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 643, + 785, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Therefore, when dealing with a sampled pixel $RGB(\\mathbf{x})$ in the RS frame $\\mathbf{I}_k^R (\\mathbf{x})$ , where $\\mathbf{x} = (u,v)$ , rather than directly using the $v$ -th pose $\\mathbf{T}_v$ of $\\mathbf{I}_k^R (\\mathbf{x})$ for training, we randomly sample $\\mathbf{T}_{\\hat{v}}$ where", + "bbox": [ + 214, + 691, + 787, + 739 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {v} \\in [ v - \\rho_ {k \\rightarrow k - 1} (\\mathbf {x}), v + \\rho_ {k \\rightarrow k + 1} (\\mathbf {x}) ]. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 369, + 751, + 785, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Why does Multi-Sampling work? Let us consider a basic task: optimizing NeRF with $H$ known poses and images (non-RS). Case $A$ : Only one row is available per pose. Case $B$ : all rows are accessible. Clearly, case $B$ would yield a better NeRF due to more data per pose. Moving to our RS setting as depicted", + "bbox": [ + 214, + 779, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "M. Niu et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/134dd0d4968a99f52c4f966ed8f47350b9727fc4f1c1ca4632e54224a0b3877b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LssMSTORIIWINEPOOLFACTORYTANABATAAVERAGE
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
ON TrBAI✓23.45.7535.044823.53.7019.068122.99.6111.078419.53.5018.107217.86.4921.103021.47.6121.0803
✓22.25.6729.056425.60.7749.071227.22.7705.058221.81.6128.094923.42.7916.067924.06.7245.0697
✓26.65.8572.034225.55.7637.068225.86.7303.055525.45.7754.070820.94.6802.083224.89.7614.0624
✓30.97.9058.039528.36.8482.067528.50.8137.056225.43.7673.086726.40.8614.065127.93.8393.0630
OUT TrBAI✓23.26.7561.042824.05.7169.060123.27.6218.073819.98.5426.089217.85.5065.094221.68.6288.0720
✓22.39.6875.050126.62.8113.057927.29.7783.048922.72.6624.081524.27.8286.051424.66.7536.0579
✓26.58.8611.030825.96.7750.061225.49.7080.054924.38.7446.074220.73.6850.075124.63.7547.0592
✓33.07.9499.024329.27.8703.056028.77.8313.045825.11.7738.080727.02.8881.049828.65.8627.0513
", + "bbox": [ + 217, + 143, + 785, + 229 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 2: Quantitative results on synthetic scenes for ablation study. $\\mathcal{L}_s$ denotes the camera trajectory smoothness regularization, and 'MS' denotes the multi-sampling algorithm. We color code each result as best and second best.", + "bbox": [ + 215, + 236, + 784, + 277 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "in Fig.5, case $A$ is the vanilla RS observation for one RS frame ( $H$ rows lead to $H$ poses), and case $B$ is the virtual GS observation, where all rows are available for each pose. Thus, case $B$ is the upper bound for case $A$ . Essentially, MS aims to transport all possible pixel values from nearby poses to fill the missing observation that can be potentially recovered (case $C$ in Fig.5), i.e., partially restore the virtual GS observation from the vanilla RS observation, by allocating additional poses for each pixel, which ensures better results than vanilla RS. Also, interpolating two poses does not harm MS. We optimize two poses for each frame, and each in-between pose becomes a fixed blend ratio of two poses. The best case is virtual GS (Fig.5(B)), where every pixel coordinate can be sampled in all ratios $\\left(\\frac{0}{H}, \\frac{1}{H-1}, \\dots, \\frac{H}{0}\\right)$ , resulting in a total of $H \\times W \\times H$ data points to optimize the two poses. In vanilla RS (Fig.5(A)), each coordinate only contributes to one ratio determined by row indices, producing only $1 \\times W \\times H$ data points. With MS (Fig.5(C)), some pixels are replicated to other poses, allowing multiple ratios for them. Thus, the number of data points is greater than $1 \\times W \\times H$ but less than $H \\times W \\times H$ . This ensures better pose estimation and view synthesis than vanilla RS.", + "bbox": [ + 212, + 320, + 787, + 575 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 601, + 375, + 618 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Settings", + "text_level": 1, + "bbox": [ + 215, + 633, + 330, + 648 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. Evaluating RS-aware view synthesis technologies requires synthetic datasets that have multi-view images with RS distortions and the corresponding GS ground truth. We synthesize 5 scenes via blender to evaluate our method. For each scene, we rendered images from various camera poses while moving the camera gradually around the scene in a forward-facing manner. We then combined rows from different poses to produce the final RS images. Each scene comprises 34 images with a resolution of $400 \\times 400$ . We employed the pose of the middle row of each RS image as the initial camera pose for our model. We also analyze our model's robustness to pose initialization in the experiment section. For evaluation, we used 34 on-trajectory views and 16 out-trajectory views to test RS correction and novel view synthesis results, respectively. To test our model in real-world scenarios, we also captured real scenes using the RS camera", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "RS-NeRF", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/4dc0615e6757afa002d38481f9f272938e24e0642c8abd588c3179c66348ebf2.jpg", + "image_caption": [ + "Vanilla", + "Fig.6: Qualitative results on synthetic scenes for ablation study. The corresponding error maps and zoomed regions are visualized in the bottom. Our full model has the smaller error to the GT, with less spatial misalignment and twist caused by the RS distortion." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 331, + 273 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/15da6701381db8447d3707d752b150f1bfe1b6ded98c90e2ef29e42158410eb8.jpg", + "image_caption": [ + "Vanilla $+\\mathcal{L}_s$" + ], + "image_footnote": [], + "bbox": [ + 333, + 143, + 444, + 272 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/a41cd36dfedf0ed62189e995797b3210080a2467e43af63fdf4715258a5ae6d8.jpg", + "image_caption": [ + "Vanilla $^+$ Mul.Sam." + ], + "image_footnote": [], + "bbox": [ + 446, + 143, + 557, + 273 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/04ffffeba08f72f7190ff0173a9d678dbdd1d821ba51718842f34fc909e0b661.jpg", + "image_caption": [ + "\\[\n\\begin{array}{r}\\mathsf{Vanilla} + \\mathcal{L}_s + \\mathsf{Mul.Sam}. \\end{array}\n\\]" + ], + "image_footnote": [], + "bbox": [ + 558, + 143, + 669, + 273 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/23e74af092b9a5b26580d441a02a2e37410b5d0337f05beb05b844878fa9d247.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 669, + 143, + 782, + 273 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/18bead5ad54558b077d423acfe4f24d9f3c56625d37b8cf47c66db1d895dab5f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
VanillaVanilla+LsVanilla+MSVanilla+Ls+MS
0.05090.03580.04620.0319
", + "bbox": [ + 217, + 366, + 531, + 401 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 3: Average MSTE (the lower the better) results for ablation study.", + "bbox": [ + 542, + 363, + 787, + 405 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "EO-1312C. These RS images were captured by manually moving the camera during exposure, in a forward-facing manner, with a resolution of $400 \\times 500$ . We utilize COLMAP [31,32] for the initial camera pose in our model.", + "bbox": [ + 212, + 434, + 784, + 479 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Metrics. The quality of the rendered image is evaluated with the commonly used metrics including PSNR, SSIM and LPIPS [38]. To evaluate the accuracy of the estimated camera trajectory, we calculate the Mean Square Trajectory Error (MSTE) with the ground-truth camera trajectory.", + "bbox": [ + 212, + 479, + 785, + 540 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 560, + 455, + 575 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In each training iteration, we randomly choose 64 rows from each of the images and 2 pixels from each of these rows. We employ the Adam optimizer [16] with $\\beta_{1} = 0.9$ and $\\beta_{2} = 0.999$ to optimize our model. The learning rate is initially set at $5 \\times 10^{-4}$ and undergoes exponential decay throughout the optimization process. To avoid potential errors from inaccurate optical flow estimation, we disable multi-sampling and use the vanilla sampling strategy in the later part of the training process. This makes the model use much fewer but absolutely accurate data points in the final training phase. Each scene is trained for 200,000 iterations. All this training is conducted on a single NVIDIA 4090 GPU.", + "bbox": [ + 212, + 584, + 787, + 720 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 739, + 390, + 756 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To comprehensively evaluate the effectiveness of each component in our RS-NeRF model, we carried out extensive ablation studies. The quantitative and qualitative results are presented in Tab. 2, Fig. 6 and Tab. 3, respectively.", + "bbox": [ + 212, + 763, + 784, + 810 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Trajectory Smoothness Regularization. The accurate estimation of camera trajectory is crucial for effectively restoring normal images from RS inputs, given", + "bbox": [ + 212, + 810, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "M. Niu et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/e0948a68ad70fd059f19418846722c10fdfb431eebbaa6229d395582448e9926.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsTORIIWINEPOOLFACTORYTANABATAAVERAGE
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
ON-TRIAL
\\(DU\\)21.236409.084519.87.4520.132923.08.5394.118919.95.5128.132518.17.5262.121320.46.5343.1180
\\(CVR\\)19.946148.145418.86.4090.205221.51.5173.188518.59.4851.188317.45.5277.184819.27.5108.1824
\\(JAM\\)20.666311.087219.50.4214.159522.43.5180.126319.34.4907.151017.87.5342.129219.96.5911.1307
RS-NeRF30.97.9058.039528.36.8482.067528.50.8137.056225.43.7673.086726.40.8614.065127.93.8393.0630
OUT-TRIAL
\\(NeRF\\)19.35.5976.253519.06.3845.355022.02.4740.319617.87.4075.357017.56.4958.301919.17.4719.3174
\\(DU+NeRF\\)20.33.6585.252820.85.4923.340423.25.5403.286019.16.5037.321918.14.5226.310220.35.5435.3022
\\(CVR+NeRF\\)23.67.7779.196722.27.6114.279124.56.6802.240819.81.5619.299618.58.6250.273621.78.6513.2580
\\(JAM+NeRF\\)24.13.7852.151222.14.5801.282225.51.6867.185520.46.5834.271718.63.6357.238822.18.6542.2259
RS-NeRF33.07.9499.024329.27.8703.056028.77.8313.045825.11.7738.080727.02.8881.049828.65.8627.0513
", + "bbox": [ + 217, + 143, + 785, + 232 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 4: Quantitative comparison on synthetic scenes of different methods. DU, CVR, and JAM represent DeepUnroll [22], CVR [10], and JAMNet [11] respectively. We color code each result as best and second best.", + "bbox": [ + 215, + 238, + 784, + 282 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "the inherent formation process of RS effects. As evidenced in Tab. 2, RS-NeRF enhanced with trajectory smoothness regularization $(\\mathcal{L}_s)$ shows notable improvements over the vanilla model. Specifically, it achieves a PSNR increase of 2.59 for on-trajectory views and 2.98 for out-trajectory views. Additionally, removing $\\mathcal{L}_s$ from our full model leads to a PSNR decrease of 3.04 for on-trajectory views and 4.02 for out-trajectory views. The qualitative results further underscore the effectiveness of $\\mathcal{L}_s$ . As illustrated in Fig. 6, the vanilla RS-NeRF, when complemented with $\\mathcal{L}_s$ , exhibits reduced spatial misalignment and a smaller synthesis error to the ground truth. Furthermore, Tab. 3 reveals that the vanilla RS-NeRF tends to provide an inaccurate trajectory estimation, primarily focusing on the photo-metric loss for each image and overlooking the pattern of camera movement across different RS frames. In contrast, the 'Vanilla + $\\mathcal{L}_s$ ' model, with the addition of smoothness regularization, yields a more accurate estimation of the camera trajectory, resulting in improved synthesis outcomes.", + "bbox": [ + 217, + 324, + 787, + 535 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Multi-Sampling Algorithm. The Multi-sampling algorithm in RS-NeRF significantly enhances the training data available for each intermediate pose by calculating the PP-ratio for each pixel, marking a notable advancement over the vanilla model. According to the results in Tab. 2, RS-NeRF equipped with the multi-sampling algorithm shows substantial improvements: a PSNR increase of 3.42 for on-trajectory views and 2.95 for out-trajectory views. Moreover, excluding the multi-sampling algorithm from our full model leads to a PSNR reduction of 3.87 for on-trajectory views and 3.99 for out-trajectory views. The qualitative results also affirm the effectiveness of it. As depicted in Fig. 6, our model, with the multi-sampling algorithm, displays fewer artifacts and a smaller synthesis error to the ground truth. Interestingly, the multi-sampling algorithm also seems to enhance trajectory estimation. As shown in Tab. 3, RS-NeRF with only the multi-sampling algorithm (Vanilla + MS) provides a more accurate trajectory estimation than the vanilla model, which likely contributes to its significantly improved synthesis results.", + "bbox": [ + 217, + 536, + 787, + 762 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Comparisons", + "text_level": 1, + "bbox": [ + 215, + 785, + 369, + 800 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of RS correction for on-trajectory images, our method is compared with three state-of-the-art 2D RS correction techniques:", + "bbox": [ + 215, + 809, + 784, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "RS-NeRF", + "bbox": [ + 665, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b2064caa354974159a7359d9d56da512a4b4b5a676905666150776e8ace0afda.jpg", + "image_caption": [ + "Fig. 7: Qualitative comparison results on synthetic scenes. The corresponding error maps and zoomed regions are visualized in the bottom. DU, CVR, and JAM represent DeepUnroll [22], CVR [10], and JAMNet [11] respectively." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 784, + 368 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/73b0b5e0cb4bc75864fcd46e39c4c021b9d467b02286186231c135b16a4b205f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsTORIIWINEPOOLFACTORYTANABATAAVERAGE
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
USB-NeRF [19]26.32.8283.082524.72.7105.167326.72.7130.159224.32.7213.136923.05.7359.133325.02.7418.1359
DRSC [26]23.04.8008.053921.84.6895.088022.85.6962.087420.85.6901.096721.19.7492.073121.95.7251.0798
RS-NeRF30.97.9058.039528.36.8482.067528.50.8137.056225.43.7673.086726.40.8614.065127.93.8393.0630
USB-NeRF [19]28.95.8932.073625.75.7446.161928.13.7765.157625.25.7676.142724.62.7853.120426.54.7935.1313
DRSC [26] + NeRF21.51.6790.283120.30.5145.300122.66.5614.002118.99.5295.290418.35.5643.296220.36.5697.2944
RS-NeRF33.07.9499.024329.27.8703.056028.77.8313.045825.11.7738.080727.02.8881.049828.65.8627.0513
", + "bbox": [ + 218, + 441, + 785, + 510 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 5: Quantitative comparisons on synthetic scenes with DRSC [26] + NeRF and USB-NeRF [19]. We color code each result as best and second best.", + "bbox": [ + 212, + 516, + 787, + 547 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "DeepUnroll [22], CVR [10], and JAMNet [11]. To evaluate out-trajectory views, we selected several baselines for comparison. The most basic approach is to train NeRF directly using RS images. We further compare our method with two-stage baselines that initially apply state-of-the-art 2D RS correction methods (DeepUnroll [22], CVR [10], and JAMNet [11]) to restore GS images, followed by training NeRF with these restored images. For 2D RS correction methods, we utilized the official checkpoints for evaluation. Furthermore, we compare RS-NeRF with DRSC [26] and USB-NeRF [19].", + "bbox": [ + 212, + 594, + 787, + 715 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Quantitative Results. Quantitative comparison results are shown in Tab. 4 and Tab. 5. 2D RS correction methods do not yield satisfactory outcomes and receive obvious lower metrics due to twist artifacts and spatial misalignment. In contrast, our approach significantly exceeds these methods in all metrics from both on-trajectory and out-trajectory views by incorporating the RS formation process into the modeling and utilizing the multi-view information from continuous RS frames. USB-NeRF [19] integrates basic camera motion modeling, i.e., assigning unique camera poses for different rows to account for the RS ef", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "M. Niu et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6657998b1800206c5019007af586d6b6f12243ea3fe9b38d97b62666fd9c90bc.jpg", + "image_caption": [ + "NeRF" + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 330, + 282 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/cfa306c44d6347bbb3f520c2194330c69089855cb9de3c26cc1e8642aab4a1ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 145, + 444, + 214 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1bdd2a722cd3e90373eb59c57669a2f2185cc660bdddbc980d5d2408c10e4097.jpg", + "image_caption": [ + "$\\mathrm{DU} + \\mathrm{NeRF}$" + ], + "image_footnote": [], + "bbox": [ + 331, + 215, + 444, + 282 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/581d6c56a6f8e06a1ea330c5266424fa921df0ce765f99106986eb1cd9baf294.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 145, + 557, + 214 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/eda2bdc76b23d396df6c13aaf3be996d55e54061ebac24968905db330f7448c2.jpg", + "image_caption": [ + "$\\mathrm{CVR} + \\mathrm{NeRF}$" + ], + "image_footnote": [], + "bbox": [ + 444, + 215, + 557, + 282 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2f50c8a6d9c15e61e37ba3248d4847d5dc46e3c6372b4587bba94a72522da7c2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 145, + 669, + 214 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3ba84ca3386737e101adba93a607f5587b02ebd79f2487527b4f44950c03c362.jpg", + "image_caption": [ + "JAM + NeRF" + ], + "image_footnote": [], + "bbox": [ + 558, + 215, + 669, + 282 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/40dd9cb377f2b751aff43979b3ac5efade0c1448402622ed2cbb8faa9dc0a85b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 145, + 782, + 215 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/12d2bcffde69cc2dbf2a2575efd95ab879bd09648d36d46b79229e76622564b2.jpg", + "image_caption": [ + "RS-NeRF" + ], + "image_footnote": [], + "bbox": [ + 669, + 215, + 782, + 282 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a966797d84ebcea10131cc5b272aa06e38607dd32a85c7e7679b1d25614d5084.jpg", + "image_caption": [ + "Fig. 8: Qualitative comparison results on real-world scenes. Our method achieves superior performance on real-world scenes." + ], + "image_footnote": [], + "bbox": [ + 218, + 348, + 359, + 420 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/cee2f2f030e6a46a5b34df4307c1fd2c8a9b7c36f6098ae17a6c339f198f3a96.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 349, + 477, + 420 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/43dabddcc826c9a19751822818e9e1ece21f5a01cfee20f9e16930d3e82a2c70.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 478, + 349, + 589, + 420 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e01f50d109a9aa8bacf9fddceb5a251544a8836f127f27474686ffb6338a4f99.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 349, + 640, + 420 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1b1d3c533dbfb143d2691a3ad57891c13634e753519c83df1d6b9200bd39f2d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 349, + 733, + 420 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/34a9457366ba9351f0e9ce7310caa70571d843988052656834c6accd1e89d893.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 349, + 808, + 420 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2b4fcce7d7e09f08c1876aea07e47000c65de683dd42c6fa67fba99f1b673f51.jpg", + "image_caption": [ + "Fig. 9: Qualitative comparison results on synthetic and real-world scenes with DRSC [26] + NeRF and USB-NeRF [19]." + ], + "image_footnote": [], + "bbox": [ + 218, + 422, + 349, + 508 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9699e66dd9bb65be2839711c92685d8455938ac7cf77f391f2f1a372c8adfc72.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 349, + 422, + 406, + 508 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0d6c6bc14c5273b51a44eb1048bce31fd3f4a449453b53b7d8a8fbbc8e84d992.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 422, + 593, + 508 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/cde5fd59937dfcff3c61b1ac3f17a8d1d821eb0bf23c9c7f8197dc546a4a4cc1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 422, + 727, + 508 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/001798100d7cb8cf2de661f89052da8f22a8ba7d07e0d2eeb3ece16e2202b48b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 727, + 422, + 782, + 508 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "fect, without additional technical improvements. The comparison shows that this approach fails to deliver satisfactory results compared to RS-NeRF.", + "bbox": [ + 212, + 582, + 784, + 612 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Qualitative Results. The qualitative performance of our method is compared to others through tests on synthetic and real-world datasets. In synthetic scenes, as shown in Fig. 7, 2D RS correction methods struggle with twisting artifacts and spatial misalignment for on-trajectory views. In contrast, our method takes advantage of multi-view information for RS correction, leading to significantly better results. For out-trajectory views, while 2D RS correction methods combined with NeRF moderately enhance the reconstruction results, they still exhibit noticeable visual flaws, such as blurriness. RS-NeRF excels by integrating the RS formation process into the model and accurately learning the underlying 3D representation for joint RS correction and novel view synthesis. Qualitative results in real world scenes are shown in Fig. 8. We can observe that baseline methods show obvious artifacts, particularly around object edges. Also, the performance of 2D RS correction combined with NeRF is poorer in real-world scenes than in synthetic ones, which further demonstrates the generalization issues of 2D RS correction methods with real-world data. RS-NeRF, on the other hand,", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "RS-NeRF", + "bbox": [ + 665, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/9eb7fcc1c51ffaa05a185d38fff9c6babfd8c5b86e157d6c078d3e7f41fa9c0d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PerturbationON TRAJ.OUT TRAJ.MSTE ↓
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
None27.93.8393.063028.65.8627.0513.0319
U[-0.01, 0.01]27.97.8403.061028.84.8680.0485.0297
U[-0.03, 0.03]28.31.8479.057829.00.8693.0470.0304
U[-0.05, 0.05]28.47.8530.058029.24.8769.0466.0308
U[-0.07, 0.07]28.43.8504.059729.25.8771.0471.0336
", + "bbox": [ + 220, + 143, + 566, + 227 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 6: Quantitative results on synthetic dataset for different noise perturbations on pose initialization. The results demonstrate that our model is robust to the accuracy of pose initialization.", + "bbox": [ + 576, + 143, + 787, + 243 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "significantly outperforms these methods with clear edges and detailed rendering, consistent with results in synthetic scenes. Fig. 9 also demonstrates the superiority of RS-NeRF over DRSC [26] + NeRF and USB-NeRF [19] on synthetic and real-world scenes.", + "bbox": [ + 212, + 268, + 784, + 328 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.5 Discussion and Limitations", + "text_level": 1, + "bbox": [ + 215, + 347, + 485, + 363 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Robustness to pose initialization. We evaluated the robustness of our method to the precision of pose initialization on synthetic dataset by introducing random noise of varying scales to the initial poses in the SE(3) space and using these altered poses for initialization. The results, as detailed in Tab. 6, indicate that our method maintains superior performance across different levels of noise perturbation. These results confirm the robustness of RS-NeRF in the accuracy of pose initialization. However, it is important to note that our method might fail when completely random poses are used for initialization. This challenge is also an ongoing issue in the field of NeRF-without-Poses, as discussed in works like [5, 20]. Addressing this limitation is a direction that we plan to pursue in future research.", + "bbox": [ + 212, + 368, + 787, + 532 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. Similar to previous works that address NeRF with abnormal inputs [24, 36], our model is designed to handle degradation (referred to the RS effects in our context) caused by camera movements, since we primarily model the camera trajectory. When encountering scenarios where RS effects are caused by object movements, i.e., dynamic scenes, our method may fail and produce unsatisfying results. We plan to involve the modeling of object movements based on dynamic NeRFs in future work.", + "bbox": [ + 212, + 535, + 787, + 642 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 660, + 359, + 676 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we present RS-NeRF, a method designed to synthesize normal images from novel views using input with Rolling Shutter (RS) distortions. We formulate a physical model that replicates the image formation process under RS conditions and jointly optimizes the NeRF parameters and the camera extrinsic for each image row. We then identify and address the fundamental limitations of the standard RS-NeRF model by introducing two innovative approaches: trajectory smoothness regularization and the multi-sampling algorithm. Through rigorous experimentation, we demonstrate that RS-NeRF surpasses previous methods in both synthetic and real-world scenarios, confirming its efficiency in rectifying distortions caused by RS cameras.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "M. Niu et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This research was supported in part by JSPS KAKENHI Grant Numbers 24K223-18, 22H00529, 20H05951, JST-Mirai Program JPMJMI23G1.", + "bbox": [ + 215, + 172, + 787, + 204 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 224, + 323, + 239 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Albl, C., Kukelova, Z., Larsson, V., Pajdla, T.: Rolling shutter camera absolute pose. IEEE transactions on pattern analysis and machine intelligence 42(6), 1439-1452 (2019)", + "2. Attal, B., Laidlaw, E., Gokaslan, A., Kim, C., Richardt, C., Tompkin, J., O'Toole, M.: Törf: Time-of-flight radiance fields for dynamic scene view synthesis. Advances in neural information processing systems 34, 26289-26301 (2021)", + "3. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)", + "4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mipnerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022)", + "5. Bian, W., Wang, Z., Li, K., Bian, J.W., Prisacariu, V.A.: Nope-nerf: Optimising neural radiance field with no pose prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4160-4169 (2023)", + "6. Cao, M., Zhong, Z., Wang, J., Zheng, Y., Yang, Y.: Learning adaptive warping for real-world rolling shutter correction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17785-17793 (2022)", + "7. Chen, T., Wang, P., Fan, Z., Wang, Z.: Aug-nerf: Training stronger neural radiance fields with triple-level physically-grounded augmentations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15191-15202 (2022)", + "8. Dave, A., Zhao, Y., Veeraraghavan, A.: Pandora: Polarization-aided neural decomposition of radiance. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part VII. pp. 538–556. Springer (2022)", + "9. Fan, B., Dai, Y., He, M.: Sunet: symmetric undistortion network for rolling shutter correction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4541-4550 (2021)", + "0. Fan, B., Dai, Y., Zhang, Z., Liu, Q., He, M.: Context-aware video reconstruction for rolling shutter cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17572-17582 (2022)", + "1. Fan, B., Mao, Y., Dai, Y., Wan, Z., Liu, Q.: Joint appearance and motion learning for efficient rolling shutter correction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5671-5681 (2023)", + "2. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: Fastnerf: High-fidelity neural rendering at 200fps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14346-14355 (2021)", + "3. Guo, Y.C., Kang, D., Bao, L., He, Y., Zhang, S.H.: Nerfren: Neural radiance fields with reflections. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18409-18418 (2022)" + ], + "bbox": [ + 225, + 252, + 785, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "RS-NeRF", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Hedborg, J., Forssén, P.E., Felsberg, M., Ringaby, E.: Rolling shutter bundle adjustment. In: 2012 IEEE Conference on Computer Vision and Pattern Recognition. pp. 1434-1441. IEEE (2012)", + "15. Kim, J.H., Latif, Y., Reid, I.: Rrd-slam: Radial-distorted rolling-shutter direct slam. In: 2017 IEEE International Conference on Robotics and Automation (ICRA). pp. 5148-5154. IEEE (2017)", + "16. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)", + "17. Klenk, S., Koestler, L., Scaramuzza, D., Cremers, D.: E-nerf: Neural radiance fields from a moving event camera. IEEE Robotics and Automation Letters (2023)", + "18. Klingner, B., Martin, D., Roseborough, J.: Street view motion-from-structure-from-motion. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 953–960 (2013)", + "19. Li, M., Wang, P., Zhao, L., Liao, B., Liu, P.: USB-neRF: Unrolling shutter bundle adjusted neural radiance fields. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=igfDXfMvm5", + "20. Lin, C.H., Ma, W.C., Torralba, A., Lucey, S.: Barf: Bundle-adjusting neural radiance fields. In: IEEE International Conference on Computer Vision (ICCV) (2021)", + "21. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020)", + "22. Liu, P., Cui, Z., Larsson, V., Pollefeys, M.: Deep shutter unrolling network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5941-5949 (2020)", + "23. Liu, P., Zuo, X., Larsson, V., Pollefeys, M.: Mba-vo: Motion blur aware visual odometry. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5550-5559 (2021)", + "24. Ma, L., Li, X., Liao, J., Zhang, Q., Wang, X., Wang, J., Sander, P.V.: Deblur-nerf: Neural radiance fields from blurry images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12861-12870 (2022)", + "25. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)", + "26. Qu, D., Liao, B., Zhang, H., Ait-Aider, O., Lao, Y.: Fast rolling shutter correction in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)", + "27. Reiser, C., Peng, S., Liao, Y., Geiger, A.: Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14335-14345 (2021)", + "28. Rengarajan, V., Balaji, Y., Rajagopalan, A.: Unrolling the shutter: Cnn to correct motion distortions. In: Proceedings of the IEEE Conference on computer Vision and Pattern Recognition. pp. 2291-2299 (2017)", + "29. Saurer, O., Koser, K., Bouguet, J.Y., Pollefeys, M.: Rolling shutter stereo. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 465-472 (2013)", + "30. Saurer, O., Pollefeys, M., Lee, G.H.: Sparse to dense 3d reconstruction from rolling shutter images. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3337-3345 (2016)", + "31. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4104-4113 (2016)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "M. Niu et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Schonberger, J.L., Zheng, E., Frahm, J.M., Pollefeys, M.: Pixelwise view selection for unstructured multi-view stereo. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14. pp. 501-518. Springer (2016)", + "33. Teed, Z., Deng, J.: Raft: Recurrent all-pairs field transforms for optical flow. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 402-419. Springer (2020)", + "34. Vasu, S., Rajagopalan, A., et al.: Occlusion-aware rolling shutter rectification of 3d scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 636-645 (2018)", + "35. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-NeRF: Structured view-dependent appearance for neural radiance fields. CVPR (2022)", + "36. Wang, P., Zhao, L., Ma, R., Liu, P.: Bad-nerf: Bundle adjusted deblur neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4170-4179 (2023)", + "37. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5438-5448 (2022)", + "38. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)", + "39. Zhong, Z., Cao, M., Sun, X., Wu, Z., Zhou, Z., Zheng, Y., Lin, S., Sato, I.: Bringing rolling shutter images alive with dual reversed distortion. In: European Conference on Computer Vision. pp. 233-249. Springer (2022)", + "40. Zhong, Z., Zheng, Y., Sato, I.: Towards rolling shutter correction and deblurring in dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9219-9228 (2021)", + "41. Zhou, X., Duan, P., Ma, Y., Shi, B.: Evunroll: Neuromorphic events based rolling shutter image correction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17775-17784 (2022)", + "42. Zhuang, B., Tran, Q.H., Ji, P., Cheong, L.F., Chandraker, M.: Learning structure-and-motion-aware rolling shutter correction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4551-4560 (2019)" + ], + "bbox": [ + 212, + 146, + 787, + 618 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "RS-NeRF", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/RS-NeRF_ Neural Radiance Fields from Rolling Shutter Images/cafbceee-b1fc-44b7-b4f6-4691cb686eec_model.json b/2024/RS-NeRF_ Neural Radiance Fields from Rolling Shutter Images/cafbceee-b1fc-44b7-b4f6-4691cb686eec_model.json new file mode 100644 index 0000000000000000000000000000000000000000..503c69fbb0b2a476d7821193878841f0d191c5a6 --- /dev/null +++ b/2024/RS-NeRF_ Neural Radiance Fields from Rolling Shutter Images/cafbceee-b1fc-44b7-b4f6-4691cb686eec_model.json @@ -0,0 +1,2555 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.23, + 0.141, + 0.776, + 0.187 + ], + "angle": 0, + "content": "RS-NeRF: Neural Radiance Fields from Rolling Shutter Images" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.214, + 0.681, + 0.245 + ], + "angle": 0, + "content": "Muyao Niu, Tong Chen, Yifan Zhan, Zhuoxiao Li, Xiang Ji, and Yinqiang Zheng*" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.257, + 0.74, + 0.313 + ], + "angle": 0, + "content": "The University of Tokyo \nmuyao.niu@gmail.com, cocottt1023@gmail.com, \nzhan-yifan@g.ecc.u-tokyo.ac.jp, lizhuoxiao@g.ecc.u-tokyo.ac.jp, \nxiangji2016@gmail.com, yqzheng@ai.u-tokyo.ac.jp" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.354, + 0.74, + 0.631 + ], + "angle": 0, + "content": "Abstract. Neural Radiance Fields (NeRFs) have become increasingly popular because of their impressive ability for novel view synthesis. However, their effectiveness is hindered by the Rolling Shutter (RS) effects commonly found in most camera systems. To solve this, we present RS-NeRF, a method designed to synthesize normal images from novel views using input with RS distortions. This involves a physical model that replicates the image formation process under RS conditions and jointly optimizes NeRF parameters and camera extrinsic for each image row. We further address the inherent shortcomings of the basic RS-NeRF model by delving into the RS characteristics and developing algorithms to enhance its functionality. First, we impose a smoothness regularization to better estimate trajectories and improve the synthesis quality, in line with the camera movement prior. We also identify and address a fundamental flaw in the vanilla RS model by introducing a multi-sampling algorithm. This new approach improves the model's performance by comprehensively exploiting the RGB data across different rows for each intermediate camera pose. Through rigorous experimentation, we demonstrate that RS-NeRF surpasses previous methods in both synthetic and real-world scenarios, proving its ability to correct RS-related distortions effectively. Codes and data available: https://github.com/MyNiuuu/RS-NeRF" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.66, + 0.377, + 0.675 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.693, + 0.788, + 0.816 + ], + "angle": 0, + "content": "CMOS imaging sensors are commonly employed in a wide range of consumer and industrial products. Many of these sensors utilize a Rolling Shutter (RS) mechanism for image capture because of its cost-effectiveness and portability. In contrast to Global Shutter (GS) cameras, which capture all pixels simultaneously, rolling shutter cameras capture image pixels row by row in a sequential manner. Consequently, if the camera is in motion during image capture, visible distortions can occur. It is widely acknowledged that rolling shutter distortions pose challenges for various downstream computer vision tasks, as indicated by" + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.825, + 0.383, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "M. Niu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "previous research [1,14,15,18,29,30]. Consequently, the correction of rolling shutter effects has attracted significant attention in the past [11, 22, 28, 34, 39-42]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.18, + 0.787, + 0.271 + ], + "angle": 0, + "content": "In the emerging field of novel view synthesis, technologies such as NeRF [25] have become increasingly proficient in generating highly realistic renderings. NeRF utilizes a continuous volumetric function that is parameterized by a multilayer perceptron (MLP) to associate 3D locations and 2D view directions with color and density information. It relies on volume rendering techniques to enable differentiable rendering." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.379 + ], + "angle": 0, + "content": "Although traditional NeRF models deliver exceptional results when applied to well-calibrated and well-captured images, they encounter difficulties when dealing with RS effects. For example, in scenarios where the camera is in motion during image capture, these RS effects can significantly impact the performance of the model. Given the widespread use of RS cameras in various industrial applications, it is imperative to take into account this phenomenon when reconstructing 3D representations using NeRFs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.788, + 0.487 + ], + "angle": 0, + "content": "One straightforward approach involves a two-stage baseline solution, where the first step corrects the rolling shutter effect in 2D image space, and then the NeRF is trained using these corrected images. While these baseline methods enhance the quality of novel view synthesis in NeRF to some extent by utilizing 2D rolling shutter correction techniques, they can not fully explore the 3D scene's geometry, leading to inaccuracies in correspondences, especially when there is significant camera motion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.788, + 0.627 + ], + "angle": 0, + "content": "In this paper, we introduce RS-NeRF, a robust framework designed to explicitly incorporate the physics-based formation of rolling shutter effects into the rendering process. RS-NeRF has the ability to reconstruct a high-quality NeRF representation from only input data that contains rolling shutter artifacts. To achieve this, we explicitly model the camera's motion trajectory throughout the exposure duration for each frame and combine the RGB rendering outcomes from different camera poses to generate the rolling shutter image. Throughout this procedure, we optimize both the NeRF network and the pose estimation network simultaneously by minimizing the photometric loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Building upon the basic RS-NeRF framework, we delve into its shortcomings and enhance its functionality through various refinements. Initially, we impose a camera trajectory smoothness regularization to refine our model's trajectory prediction, by leveraging the sequential nature of camera movement to yield improved results. We further identify a critical shortcoming due to the vanilla RS camera motion model and introduce a multi-sampling technique that significantly enhances RS-NeRF's performance. Specifically, we reveal the issue of the standard camera pose formulation, which assigns distinct poses per row, potentially limiting RS-NeRF's efficiency due to limited training data for each pose. We then make a key observation that multiple poses may share the same 2D RGB observation for the same 3D point during the exposure time, which can drastically increase the available training data. To capitalize on this insight, we introduce and compute the PP-ratio—a measure of camera pose movement in 3D space relative to pixel movement in 2D space, and apply this to broaden our" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.665, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RS-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.145, + 0.784, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.262, + 0.788, + 0.348 + ], + "angle": 0, + "content": "Fig. 1: Pipeline of vanilla RS-NeRF. Given a series of images with RS distortion, RS-NeRF learns the underlying normal 3D representations by jointly estimating pose and RGB values for each row. For each RS image, we first sample the rows and obtain their camera poses via the pose estimation network. We then sample points for each row and feed them with the estimated camera pose to the NeRF network and optimize the photo-metric loss with ground truths." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.784, + 0.422 + ], + "angle": 0, + "content": "training data by sampling out-of-row pixel points. This approach significantly bolsters our model's performance by tapping into the inherent data augmentation possibilities presented by the RS formation process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.423, + 0.785, + 0.482 + ], + "angle": 0, + "content": "We collect both synthetic and real data to evaluate our proposed RS-NeRF. The experimental results reveal that RS-NeRF outperforms the preceding models and demonstrate the effectiveness of our proposed refinements on the vanilla model. In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.496, + 0.784, + 0.54 + ], + "angle": 0, + "content": "- We propose a framework that can reconstruct NeRF from inputs with Rolling Shutter (RS) effects by modeling the camera trajectory and aggregating RGB rendering results of different poses during the exposure time." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.542, + 0.784, + 0.586 + ], + "angle": 0, + "content": "- By analyzing the vanilla framework in detail, we proposed a series of improvements to it, including trajectory smoothness regularization and the multi-sampling algorithm." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.587, + 0.784, + 0.615 + ], + "angle": 0, + "content": "- We contribute synthetic and real datasets to evaluate our framework and future works." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.496, + 0.784, + 0.615 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.641, + 0.388, + 0.657 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "2D Rolling Shutter Correction. Rolling Shutter Correction (RSC) aims to restore normal images given the input with RS distortions. In recent years, several appealing image-based RSC methods [22,28,34,39-42] have been developed. These methods take either single RS image or continuous RS frames as input and predict their global shutter (GS) counterparts. However, single-image RSC can be seriously ill-posed and struggle to correctly estimate the GS counterpart in complex scenarios. To reduce the ill-posedness of single-image RSC, more attention has been paid to multi-image RSC [6, 9-11, 22, 40]. However, these multi-image RS correction methods, which are built on image space operations, fail to exploit the 3D geometry of the scene. As a result, these methods produce multi-view inconsistency results which largely impact the reconstruction results" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "M. Niu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.191 + ], + "angle": 0, + "content": "of NeRF. In contrast, our proposed method can jointly reconstruct 3D geometry and correct RS images by effectively aggregating multi-view information of the scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.788, + 0.42 + ], + "angle": 0, + "content": "Neural Radiance Fields (NeRF). NeRF [25] is an end-to-end model that synthesizes novel views of complex 3D scenes by optimizing an implicit continuous volumetric scene function encoded by an MLP. Due to its simplicity in training and high-quality rendering results, NeRF-based methods [3,4,7,12,21,27,35,37] have attracted intensive attention from various computer vision fields, leading to significant improvements in terms of both accuracy and efficiency. NeRF-based models have also inspired many subsequent works that extend their continuous neural representation to different setups [2,8,13,17,24,36]. Deblur-NeRF [24] and BAD-NeRF [36] reconstruct sharp images given the blurry input from different view directions. One concurrent work [19] appears in ICLR'24, which formulates the RS process with B-cubic models. This work combines the basic modeling of camera motion with the NeRF network to consider the RS effect, without further technical improvements. In this paper, we first bake the RS model into NeRF formulation, then analyze several inherit limitations of the vanilla model, and propose a series of algorithms to further improve its performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.443, + 0.348, + 0.459 + ], + "angle": 0, + "content": "3 RS-NeRF" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.785, + 0.552 + ], + "angle": 0, + "content": "In this section, we present the specifics of RS-NeRF. Initially, we explain the basic formulation of RS-NeRF in Sec. 3.1. We also note that the vanilla model possesses specific shortcomings that hinder its effectiveness. To address these, we suggest several improvements to enhance its capabilities, which are detailed in Sec. 3.2 and Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.575, + 0.466, + 0.59 + ], + "angle": 0, + "content": "3.1 Vanilla RS-NeRF Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.601, + 0.784, + 0.692 + ], + "angle": 0, + "content": "Rolling shutter camera model. Unlike Global Shutter (GS) cameras, the Rolling Shutter (RS) camera captures each row at distinct timestamps. Without loss of generality, we consider the readout direction of the RS camera to be from top to bottom, assuming an infinitesimal exposure time for each row, following existing RS modeling techniques [10, 11, 22]. This mechanism can be described through a mathematical model as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.435, + 0.705, + 0.785, + 0.721 + ], + "angle": 0, + "content": "\\[\n[ \\mathbf {I} ^ {R} (\\mathbf {x}) ] _ {i} = [ \\mathbf {I} _ {i} ^ {G} (\\mathbf {x}) ] _ {i}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.784, + 0.795 + ], + "angle": 0, + "content": "where \\([\\mathbf{I}(\\mathbf{x})]_i\\) is an operator to extract the \\(i\\)-th row from an image \\(\\mathbf{I}(\\mathbf{x})\\), \\(\\mathbf{I}^R(\\mathbf{x}) \\in \\mathbb{R}^{H \\times W \\times 3}\\) is the rolling shutter image, and \\(\\mathbf{I}_i^G(\\mathbf{x}) \\in \\mathbb{R}^{H \\times W \\times 3}\\) is the virtual global shutter image captured at the same pose as the \\(i\\)-th row of \\(\\mathbf{I}^R(\\mathbf{x})\\). As a result, for one RS image \\(\\mathbf{I}^R(\\mathbf{x})\\), we denote the pose of the \\(i\\)-th row as \\(\\mathbf{T}_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Camera motion trajectory modeling. In our rolling shutter camera model, it is necessary to establish the poses for each row in each image. Given the typically short exposure time of an RS image, we use a linear model to approximate the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.665, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RS-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.148, + 0.494, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.145, + 0.788, + 0.262 + ], + "angle": 0, + "content": "Fig. 2: Smoothness regularization on camera trajectory. For each pair of adjacent trajectory vectors \\(\\vec{\\mathbf{d}}_k\\) and \\(\\vec{\\mathbf{d}}_{k + 1}\\), we compute the mid-point trajectory vector \\(\\vec{\\mathbf{d}}_{mid}\\) and their unit vectors \\(\\vec{\\mathbf{n}}_k\\), \\(\\vec{\\mathbf{n}}_{k + 1}\\), and \\(\\vec{\\mathbf{n}}_{mid}\\). We then apply the \\(L_{2}\\) regularization between \\(\\vec{\\mathbf{n}}_{mid}\\) and mean \\((\\vec{\\mathbf{n}}_k,\\vec{\\mathbf{n}}_{k + 1})\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.291, + 0.784, + 0.352 + ], + "angle": 0, + "content": "motion of the camera. Specifically, we start by defining two camera poses: one for the first row, denoted as \\(\\mathbf{T}_0\\), and one for the last row, \\(\\mathbf{T}_{H - 1}\\). For the rows in between, we linearly interpolate their poses within the Lie algebra of \\(\\mathbf{SE}(3)\\) as informed by prior research [20, 23, 36]:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.363, + 0.786, + 0.38 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} _ {i} = \\mathbf {T} _ {0} \\cdot (1 - \\rho) + \\mathbf {T} _ {H - 1} \\cdot \\rho , \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.784, + 0.481 + ], + "angle": 0, + "content": "where the interpolation coefficient \\(\\rho = \\frac{i}{H - 1}\\) is based on the row index \\(i\\). Thus, for each RS image, RS-NeRF receives one pose initialization, then optimizes the poses at both ends and then determines the poses for each intermediate row through linear interpolation. This interpolation method for calculating each row's pose is fully differentiable, enabling the optimization of pose parameters using a series of rolling shutter images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.482, + 0.788, + 0.603 + ], + "angle": 0, + "content": "Optimizing NeRF from RS observation. For a set of \\(K\\) rolling shutter images, we select \\(M\\) rows per image and \\(N\\) points per row, resulting in \\(K \\times M \\times N\\) distinct camera rays for each iteration. These rays are used to concurrently optimize the NeRF network and the pose parameters. Specifically, for each image, we initially use the row indices to access the pose estimation network. This allows us to determine the camera ray \\(\\hat{\\mathbf{r}}(\\theta_p; u, v)\\) for each sampled pixel point \\((u, v)\\). Following this, these rays are employed to train the NeRF network by computing the ensuing photometric loss:" + }, + { + "type": "equation", + "bbox": [ + 0.363, + 0.614, + 0.785, + 0.631 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\left| \\left| \\mathbf {C} (\\mathbf {r} (u, v)) - \\hat {\\mathbf {C}} \\left(\\theta_ {n}; \\hat {\\mathbf {r}} \\left(\\theta_ {p}; (u, v)\\right) \\right. \\right| \\right| _ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.788, + 0.718 + ], + "angle": 0, + "content": "where \\(\\theta_{n}\\) represents the trainable parameters of the NeRF network, and \\(\\theta_{p}\\) denotes the learnable parameters for camera poses. \\(\\hat{\\mathbf{C}} (\\theta_n;\\hat{\\mathbf{r}} (\\theta_p;(u,v))\\) refers to the predicted RGB value for the pixel at coordinates \\((u,v)\\), while \\(\\mathbf{C}(\\mathbf{r}(u,v))\\) denotes the corresponding RGB ground truth value. The workflow of our vanilla RS-NeRF model is depicted in Fig. 1." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.74, + 0.583, + 0.756 + ], + "angle": 0, + "content": "3.2 Trajectory Smoothness Regularization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Given the previous description of our model, it is evident that accurately estimating the camera pose for each row in RS frames is crucial for effectively reconstructing the normal NeRF representation. We observe that the basic RS-NeRF struggles with precise camera trajectory estimation, often resulting in high-frequency predictions, as further analyzed in our experiments. This issue" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "M. Niu et al." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.152, + 0.458, + 0.214 + ], + "angle": 0, + "content": "
NposeON TRAJ. / OUT TRAJ.
PSNR ↑SSIM ↑LPIPS ↓
10024.36/25.16.7321/.7727.0703/.0578
40024.06/24.66.7245/.7536.0697/.0579
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.219, + 0.458, + 0.263 + ], + "angle": 0, + "content": "Table 1: Quantitative results for different \\(N_{\\text{pose}}\\). \\(N_{\\text{pose}} = 100\\) yields better results than \\(N_{\\text{pose}} = 400\\)." + }, + { + "type": "image", + "bbox": [ + 0.462, + 0.145, + 0.785, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.459, + 0.233, + 0.788, + 0.263 + ], + "angle": 0, + "content": "Fig. 3: Qualitative comparisons for different \\( N_{\\text{pose}} \\). The artifact decreases as \\( N_{\\text{pose}} \\) grows." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.303, + 0.785, + 0.395 + ], + "angle": 0, + "content": "arises because the network focuses only on the photometric loss for individual images, neglecting the consistent movement characteristics of the camera across different RS frames. To address this, we introduce a camera trajectory smoothness regularization, based on the expectation of a smooth and continuous camera trajectory in 3D space. Specifically, for every two adjacent RS frames \\(\\mathbf{I}_k^R\\) and \\(\\mathbf{I}_{k + 1}^R\\), we aim to minimize the following loss during the optimization process:" + }, + { + "type": "equation", + "bbox": [ + 0.365, + 0.404, + 0.787, + 0.423 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} = | | \\vec {\\mathbf {n}} _ {m i d} - \\operatorname {m e a n} (\\vec {\\mathbf {n}} _ {k} + \\vec {\\mathbf {n}} _ {k + 1}) | | _ {2}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.788, + 0.493 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, \\(\\vec{\\mathbf{n}}_k\\) and \\(\\vec{\\mathbf{n}}_{k + 1}\\) represent the unit direction vectors of the estimated camera trajectories for RS frames \\(\\mathbf{I}_k^R\\) and \\(\\mathbf{I}_{k + 1}^R\\), respectively. Meanwhile, \\(\\vec{\\mathbf{n}}_{mid}\\) denotes the unit direction vector for the camera trajectory during the exposure interval between these two RS frames." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.514, + 0.482, + 0.53 + ], + "angle": 0, + "content": "3.3 Multi-Sampling Algorithm" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.644 + ], + "angle": 0, + "content": "As discussed in Sec. 3.1, each row in an RS frame is assigned a unique camera pose. For example, in an RS frame with 400 rows, there are 400 distinct camera poses. This rigorous approach aligns with the real-world, physics-based photography process of RS cameras but also presents certain challenges. For example, under this framework, the data points for each pose are confined to a single row. This limitation can significantly affect our model's performance due to the sparse training data available for each pose." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.788, + 0.809 + ], + "angle": 0, + "content": "A straightforward solution to this problem is to reduce the number of poses per image. For example, assigning 200 poses to each RS frame means that all two adjacent rows would share the same pose. This approach can lead to blurriness since these rows are captured at different poses, but it also doubles the amount of data available for training each pose. To further explore this trade-off, we conducted a series of experiments. Fig. 3 shows qualitative results for different intermediate pose numbers \\( N_{\\text{pose}} \\), indicating that the artifacts decrease as \\( N_{\\text{pose}} \\) increases. Interestingly, we observed that using 100 poses per frame often yields results comparable to or even better than using 400 poses. The quantitative results presented in Tab. 1 also support the superiority of having 100 poses per frame." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "This observation seems counter-intuitive at first, as one might assume that having fewer than 400 poses per frame would lead to poorer results due to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.665, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RS-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.146, + 0.728, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.32, + 0.788, + 0.417 + ], + "angle": 0, + "content": "Fig. 4: Motivation for the multi-sampling algorithm. Up: Three consecutive RS frames, labeled as (a), (b), and (c), along with the optical flow between each pair of frames, shown in (d) and (e). Down: For a single point in 3D space, we observe its 2D projections in the first (marked in yellow) and second (marked in red) RS frames. These projections exhibit a row displacement of 8 and a column displacement of 18, as depicted in (f). However, during this interval, the camera undergoes a shift of at least 392 poses, as indicated in (g)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.453, + 0.788, + 0.573 + ], + "angle": 0, + "content": "artifacts like blurriness. However, it can be rationally explained. As demonstrated in Fig. 4, for a given point in 3D space, the 2D projections on the first and second RS frames show relatively small pixel displacements (8 in row and 12 in column). Meanwhile, the camera shifts through at least 392 poses in 3D space. This is based on the assumption that there is no overlap in the trajectories of two adjacent frames (indicated as \\( \\geq 0 \\) poses' in the figure), which is naturally satisfied given the sparse observation setting and short exposure times typically involved." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.578, + 0.788, + 0.699 + ], + "angle": 0, + "content": "Following our prior analysis, we introduced a multi-sampling algorithm aimed at explicitly increasing the sample points for each intermediate camera pose. The fundamental concept involves computing the ratio of camera pose displacement to pixel displacement (termed the PP-ratio) between two adjacent frames for the same 3D point. For example, a pixel displacement of 18 combined with a camera pose shift of 392 results in a PP-ratio of 21. This implies that at least 21 consecutive camera poses share the same pixel value, which means that all these 21 camera poses can utilize this RGB value for training purposes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.701, + 0.788, + 0.793 + ], + "angle": 0, + "content": "For a pixel \\(\\mathbf{x} = (u,v)\\) located in the \\(u\\)-th column and \\(v\\)-th row of the \\(k\\)-th RS frame \\(\\mathbf{I}_k^R (\\mathbf{x})\\), we determine both the forward PP-ratio \\(\\rho_{k\\to k + 1}(\\mathbf{x})\\) and the backward PP-ratio \\(\\rho_{k\\to k - 1}(\\mathbf{x})\\). This is because the point represented by \\(\\mathbf{x}\\) can transition from \\(\\mathbf{I}_k^R (\\mathbf{x})\\) to \\(\\mathbf{I}_{k + 1}^R (\\mathbf{x})\\) or \\(\\mathbf{I}_{k - 1}^R (\\mathbf{x})\\). Consequently, all camera poses within the range of \\([v - \\rho_{k\\to k - 1}(\\mathbf{x}), v + \\rho_{k\\to k + 1}(\\mathbf{x})]\\) are capable of sampling the RGB value of \\(\\mathbf{x}\\) for training purposes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "To compute the forward PP-ratio \\(\\rho_{k\\rightarrow k + 1}(\\mathbf{x})\\) for a pixel \\(\\mathbf{x}\\) in the \\(k\\)-th RS frame \\(\\mathbf{I}_k^R (\\mathbf{x})\\), the first step is to estimate the forward optical flow \\(F_{k\\rightarrow k + 1}(\\mathbf{x})\\), utilizing RAFT [33]. Consequently, the row displacement \\(\\varDelta_{k\\rightarrow k + 1}^{row}(\\mathbf{x})\\) and column" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "M. Niu et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.143, + 0.561, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.145, + 0.788, + 0.312 + ], + "angle": 0, + "content": "Fig. 5: Comparing the accessible training points in one RS image under different strategies. The image size is \\( H \\times W \\), providing \\( H \\) poses. Distinct circle hues represent different RGB values. In case \\( C \\), pixels transported from other poses are with dash contours, and the number signifying that it originates from the \\( i \\)-th pose." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.355, + 0.784, + 0.387 + ], + "angle": 0, + "content": "displacement \\(\\varDelta_{k\\rightarrow k+1}^{col}(\\mathbf{x})\\) of \\(\\mathbf{x}\\) from \\(\\mathbf{I}_k^R (\\mathbf{x})\\) to \\(\\mathbf{I}_{k + 1}^R (\\mathbf{x})\\) would be the second and the first element of \\(F_{k\\to k + 1}(\\mathbf{x})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.361, + 0.397, + 0.786, + 0.416 + ], + "angle": 0, + "content": "\\[\nF _ {k \\rightarrow k + 1} (\\mathbf {x}) = \\operatorname {R A F T} \\left(\\mathbf {I} _ {k} ^ {R} (\\mathbf {x}), \\mathbf {I} _ {k + 1} ^ {R} (\\mathbf {x})\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.362, + 0.418, + 0.787, + 0.438 + ], + "angle": 0, + "content": "\\[\n\\varDelta_ {k \\rightarrow k + 1} ^ {c o l} (\\mathbf {x}), \\varDelta_ {k \\rightarrow k + 1} ^ {r o w} (\\mathbf {x}) = F _ {k \\rightarrow k + 1} (\\mathbf {x}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.448, + 0.637, + 0.461 + ], + "angle": 0, + "content": "and the calculation are similar for the backward situation." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.464, + 0.72, + 0.479 + ], + "angle": 0, + "content": "We then calculate the forward and backward camera pose shift as:" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.49, + 0.786, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\Delta_ {k \\rightarrow k + 1} ^ {\\text {p o s e}} (\\mathbf {x}) = v + \\Delta_ {k \\rightarrow k + 1} ^ {\\text {r o w}} (\\mathbf {x}) + (H - v), (7) \\\\ = H + \\Delta_ {k \\rightarrow k + 1} ^ {\\text {r o w}} (\\mathbf {x}), (8) \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.346, + 0.531, + 0.786, + 0.569 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\Delta_ {k \\rightarrow k - 1} ^ {\\text {p o s e}} (\\mathbf {x}) = \\left(H - \\left(v + \\Delta_ {k \\rightarrow k - 1} ^ {\\text {r o w}} (\\mathbf {x})\\right)\\right) + v, (9) \\\\ = H - \\Delta_ {k \\rightarrow k - 1} ^ {\\text {r o w}} (\\mathbf {x}). (10) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.581, + 0.645, + 0.595 + ], + "angle": 0, + "content": "Finally, we calculate forward and backward PP-ratio as:" + }, + { + "type": "equation", + "bbox": [ + 0.338, + 0.606, + 0.786, + 0.643 + ], + "angle": 0, + "content": "\\[\n\\rho_ {k \\rightarrow k + 1} (\\mathbf {x}) = \\frac {\\Delta_ {k \\rightarrow k + 1} ^ {p o s e} (\\mathbf {x})}{\\max \\left(\\Delta_ {k \\rightarrow k + 1} ^ {c o l} (\\mathbf {x}) , \\Delta_ {k \\rightarrow k + 1} ^ {r o w} (\\mathbf {x})\\right)}, \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.644, + 0.786, + 0.682 + ], + "angle": 0, + "content": "\\[\n\\rho_ {k \\rightarrow k - 1} (\\mathbf {x}) = \\frac {\\Delta_ {k \\rightarrow k - 1} ^ {p o s e} (\\mathbf {x})}{\\operatorname* {m a x} \\left(\\Delta_ {k \\rightarrow k - 1} ^ {c o l} (\\mathbf {x}) , \\Delta_ {k \\rightarrow k - 1} ^ {r o w} (\\mathbf {x})\\right)}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.693, + 0.788, + 0.74 + ], + "angle": 0, + "content": "Therefore, when dealing with a sampled pixel \\( RGB(\\mathbf{x}) \\) in the RS frame \\( \\mathbf{I}_k^R (\\mathbf{x}) \\), where \\( \\mathbf{x} = (u,v) \\), rather than directly using the \\( v \\)-th pose \\( \\mathbf{T}_v \\) of \\( \\mathbf{I}_k^R (\\mathbf{x}) \\) for training, we randomly sample \\( \\mathbf{T}_{\\hat{v}} \\) where" + }, + { + "type": "equation", + "bbox": [ + 0.37, + 0.752, + 0.786, + 0.769 + ], + "angle": 0, + "content": "\\[\n\\hat {v} \\in [ v - \\rho_ {k \\rightarrow k - 1} (\\mathbf {x}), v + \\rho_ {k \\rightarrow k + 1} (\\mathbf {x}) ]. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Why does Multi-Sampling work? Let us consider a basic task: optimizing NeRF with \\( H \\) known poses and images (non-RS). Case \\( A \\): Only one row is available per pose. Case \\( B \\): all rows are accessible. Clearly, case \\( B \\) would yield a better NeRF due to more data per pose. Moving to our RS setting as depicted" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.665, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RS-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.23 + ], + "angle": 0, + "content": "
LssMSTORIIWINEPOOLFACTORYTANABATAAVERAGE
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
ON TrBAI✓23.45.7535.044823.53.7019.068122.99.6111.078419.53.5018.107217.86.4921.103021.47.6121.0803
✓22.25.6729.056425.60.7749.071227.22.7705.058221.81.6128.094923.42.7916.067924.06.7245.0697
✓26.65.8572.034225.55.7637.068225.86.7303.055525.45.7754.070820.94.6802.083224.89.7614.0624
✓30.97.9058.039528.36.8482.067528.50.8137.056225.43.7673.086726.40.8614.065127.93.8393.0630
OUT TrBAI✓23.26.7561.042824.05.7169.060123.27.6218.073819.98.5426.089217.85.5065.094221.68.6288.0720
✓22.39.6875.050126.62.8113.057927.29.7783.048922.72.6624.081524.27.8286.051424.66.7536.0579
✓26.58.8611.030825.96.7750.061225.49.7080.054924.38.7446.074220.73.6850.075124.63.7547.0592
✓33.07.9499.024329.27.8703.056028.77.8313.045825.11.7738.080727.02.8881.049828.65.8627.0513
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.237, + 0.785, + 0.279 + ], + "angle": 0, + "content": "Table 2: Quantitative results on synthetic scenes for ablation study. \\( \\mathcal{L}_s \\) denotes the camera trajectory smoothness regularization, and 'MS' denotes the multi-sampling algorithm. We color code each result as best and second best." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.321, + 0.788, + 0.577 + ], + "angle": 0, + "content": "in Fig.5, case \\( A \\) is the vanilla RS observation for one RS frame (\\( H \\) rows lead to \\( H \\) poses), and case \\( B \\) is the virtual GS observation, where all rows are available for each pose. Thus, case \\( B \\) is the upper bound for case \\( A \\). Essentially, MS aims to transport all possible pixel values from nearby poses to fill the missing observation that can be potentially recovered (case \\( C \\) in Fig.5), i.e., partially restore the virtual GS observation from the vanilla RS observation, by allocating additional poses for each pixel, which ensures better results than vanilla RS. Also, interpolating two poses does not harm MS. We optimize two poses for each frame, and each in-between pose becomes a fixed blend ratio of two poses. The best case is virtual GS (Fig.5(B)), where every pixel coordinate can be sampled in all ratios \\( \\left(\\frac{0}{H}, \\frac{1}{H-1}, \\dots, \\frac{H}{0}\\right) \\), resulting in a total of \\( H \\times W \\times H \\) data points to optimize the two poses. In vanilla RS (Fig.5(A)), each coordinate only contributes to one ratio determined by row indices, producing only \\( 1 \\times W \\times H \\) data points. With MS (Fig.5(C)), some pixels are replicated to other poses, allowing multiple ratios for them. Thus, the number of data points is greater than \\( 1 \\times W \\times H \\) but less than \\( H \\times W \\times H \\). This ensures better pose estimation and view synthesis than vanilla RS." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.602, + 0.376, + 0.619 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.634, + 0.331, + 0.649 + ], + "angle": 0, + "content": "4.1 Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Datasets. Evaluating RS-aware view synthesis technologies requires synthetic datasets that have multi-view images with RS distortions and the corresponding GS ground truth. We synthesize 5 scenes via blender to evaluate our method. For each scene, we rendered images from various camera poses while moving the camera gradually around the scene in a forward-facing manner. We then combined rows from different poses to produce the final RS images. Each scene comprises 34 images with a resolution of \\(400 \\times 400\\). We employed the pose of the middle row of each RS image as the initial camera pose for our model. We also analyze our model's robustness to pose initialization in the experiment section. For evaluation, we used 34 on-trajectory views and 16 out-trajectory views to test RS correction and novel view synthesis results, respectively. To test our model in real-world scenarios, we also captured real scenes using the RS camera" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "M. Niu et al." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.145, + 0.333, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.274, + 0.292, + 0.282 + ], + "angle": 0, + "content": "Vanilla" + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.145, + 0.446, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.366, + 0.274, + 0.413, + 0.283 + ], + "angle": 0, + "content": "Vanilla \\(+\\mathcal{L}_s\\)" + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.145, + 0.558, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.466, + 0.274, + 0.538, + 0.283 + ], + "angle": 0, + "content": "Vanilla \\(^+\\) Mul.Sam." + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.145, + 0.671, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.57, + 0.274, + 0.661, + 0.283 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{r}\\mathsf{Vanilla} + \\mathcal{L}_s + \\mathsf{Mul.Sam}. \\end{array}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.145, + 0.784, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.274, + 0.753, + 0.282 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.295, + 0.785, + 0.351 + ], + "angle": 0, + "content": "Fig.6: Qualitative results on synthetic scenes for ablation study. The corresponding error maps and zoomed regions are visualized in the bottom. Our full model has the smaller error to the GT, with less spatial misalignment and twist caused by the RS distortion." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.367, + 0.532, + 0.402 + ], + "angle": 0, + "content": "
VanillaVanilla+LsVanilla+MSVanilla+Ls+MS
0.05090.03580.04620.0319
" + }, + { + "type": "table_caption", + "bbox": [ + 0.544, + 0.364, + 0.788, + 0.406 + ], + "angle": 0, + "content": "Table 3: Average MSTE (the lower the better) results for ablation study." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.785, + 0.48 + ], + "angle": 0, + "content": "EO-1312C. These RS images were captured by manually moving the camera during exposure, in a forward-facing manner, with a resolution of \\(400 \\times 500\\). We utilize COLMAP [31,32] for the initial camera pose in our model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.787, + 0.541 + ], + "angle": 0, + "content": "Evaluation Metrics. The quality of the rendered image is evaluated with the commonly used metrics including PSNR, SSIM and LPIPS [38]. To evaluate the accuracy of the estimated camera trajectory, we calculate the Mean Square Trajectory Error (MSTE) with the ground-truth camera trajectory." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.561, + 0.457, + 0.576 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.788, + 0.721 + ], + "angle": 0, + "content": "In each training iteration, we randomly choose 64 rows from each of the images and 2 pixels from each of these rows. We employ the Adam optimizer [16] with \\(\\beta_{1} = 0.9\\) and \\(\\beta_{2} = 0.999\\) to optimize our model. The learning rate is initially set at \\(5 \\times 10^{-4}\\) and undergoes exponential decay throughout the optimization process. To avoid potential errors from inaccurate optical flow estimation, we disable multi-sampling and use the vanilla sampling strategy in the later part of the training process. This makes the model use much fewer but absolutely accurate data points in the final training phase. Each scene is trained for 200,000 iterations. All this training is conducted on a single NVIDIA 4090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.741, + 0.391, + 0.757 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.785, + 0.811 + ], + "angle": 0, + "content": "To comprehensively evaluate the effectiveness of each component in our RS-NeRF model, we carried out extensive ablation studies. The quantitative and qualitative results are presented in Tab. 2, Fig. 6 and Tab. 3, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Trajectory Smoothness Regularization. The accurate estimation of camera trajectory is crucial for effectively restoring normal images from RS inputs, given" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.666, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RS-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.233 + ], + "angle": 0, + "content": "
MethodsTORIIWINEPOOLFACTORYTANABATAAVERAGE
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
ON-TRIAL
\\(DU\\)21.236409.084519.87.4520.132923.08.5394.118919.95.5128.132518.17.5262.121320.46.5343.1180
\\(CVR\\)19.946148.145418.86.4090.205221.51.5173.188518.59.4851.188317.45.5277.184819.27.5108.1824
\\(JAM\\)20.666311.087219.50.4214.159522.43.5180.126319.34.4907.151017.87.5342.129219.96.5911.1307
RS-NeRF30.97.9058.039528.36.8482.067528.50.8137.056225.43.7673.086726.40.8614.065127.93.8393.0630
OUT-TRIAL
\\(NeRF\\)19.35.5976.253519.06.3845.355022.02.4740.319617.87.4075.357017.56.4958.301919.17.4719.3174
\\(DU+NeRF\\)20.33.6585.252820.85.4923.340423.25.5403.286019.16.5037.321918.14.5226.310220.35.5435.3022
\\(CVR+NeRF\\)23.67.7779.196722.27.6114.279124.56.6802.240819.81.5619.299618.58.6250.273621.78.6513.2580
\\(JAM+NeRF\\)24.13.7852.151222.14.5801.282225.51.6867.185520.46.5834.271718.63.6357.238822.18.6542.2259
RS-NeRF33.07.9499.024329.27.8703.056028.77.8313.045825.11.7738.080727.02.8881.049828.65.8627.0513
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.239, + 0.785, + 0.283 + ], + "angle": 0, + "content": "Table 4: Quantitative comparison on synthetic scenes of different methods. DU, CVR, and JAM represent DeepUnroll [22], CVR [10], and JAMNet [11] respectively. We color code each result as best and second best." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.325, + 0.788, + 0.536 + ], + "angle": 0, + "content": "the inherent formation process of RS effects. As evidenced in Tab. 2, RS-NeRF enhanced with trajectory smoothness regularization \\((\\mathcal{L}_s)\\) shows notable improvements over the vanilla model. Specifically, it achieves a PSNR increase of 2.59 for on-trajectory views and 2.98 for out-trajectory views. Additionally, removing \\(\\mathcal{L}_s\\) from our full model leads to a PSNR decrease of 3.04 for on-trajectory views and 4.02 for out-trajectory views. The qualitative results further underscore the effectiveness of \\(\\mathcal{L}_s\\). As illustrated in Fig. 6, the vanilla RS-NeRF, when complemented with \\(\\mathcal{L}_s\\), exhibits reduced spatial misalignment and a smaller synthesis error to the ground truth. Furthermore, Tab. 3 reveals that the vanilla RS-NeRF tends to provide an inaccurate trajectory estimation, primarily focusing on the photo-metric loss for each image and overlooking the pattern of camera movement across different RS frames. In contrast, the 'Vanilla + \\(\\mathcal{L}_s\\)' model, with the addition of smoothness regularization, yields a more accurate estimation of the camera trajectory, resulting in improved synthesis outcomes." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.537, + 0.788, + 0.763 + ], + "angle": 0, + "content": "Multi-Sampling Algorithm. The Multi-sampling algorithm in RS-NeRF significantly enhances the training data available for each intermediate pose by calculating the PP-ratio for each pixel, marking a notable advancement over the vanilla model. According to the results in Tab. 2, RS-NeRF equipped with the multi-sampling algorithm shows substantial improvements: a PSNR increase of 3.42 for on-trajectory views and 2.95 for out-trajectory views. Moreover, excluding the multi-sampling algorithm from our full model leads to a PSNR reduction of 3.87 for on-trajectory views and 3.99 for out-trajectory views. The qualitative results also affirm the effectiveness of it. As depicted in Fig. 6, our model, with the multi-sampling algorithm, displays fewer artifacts and a smaller synthesis error to the ground truth. Interestingly, the multi-sampling algorithm also seems to enhance trajectory estimation. As shown in Tab. 3, RS-NeRF with only the multi-sampling algorithm (Vanilla + MS) provides a more accurate trajectory estimation than the vanilla model, which likely contributes to its significantly improved synthesis results." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.786, + 0.37, + 0.801 + ], + "angle": 0, + "content": "4.4 Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "To evaluate the effectiveness of RS correction for on-trajectory images, our method is compared with three state-of-the-art 2D RS correction techniques:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "M. Niu et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.785, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.38, + 0.788, + 0.424 + ], + "angle": 0, + "content": "Fig. 7: Qualitative comparison results on synthetic scenes. The corresponding error maps and zoomed regions are visualized in the bottom. DU, CVR, and JAM represent DeepUnroll [22], CVR [10], and JAMNet [11] respectively." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.442, + 0.787, + 0.511 + ], + "angle": 0, + "content": "
MethodsTORIIWINEPOOLFACTORYTANABATAAVERAGE
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
USB-NeRF [19]26.32.8283.082524.72.7105.167326.72.7130.159224.32.7213.136923.05.7359.133325.02.7418.1359
DRSC [26]23.04.8008.053921.84.6895.088022.85.6962.087420.85.6901.096721.19.7492.073121.95.7251.0798
RS-NeRF30.97.9058.039528.36.8482.067528.50.8137.056225.43.7673.086726.40.8614.065127.93.8393.0630
USB-NeRF [19]28.95.8932.073625.75.7446.161928.13.7765.157625.25.7676.142724.62.7853.120426.54.7935.1313
DRSC [26] + NeRF21.51.6790.283120.30.5145.300122.66.5614.002118.99.5295.290418.35.5643.296220.36.5697.2944
RS-NeRF33.07.9499.024329.27.8703.056028.77.8313.045825.11.7738.080727.02.8881.049828.65.8627.0513
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.517, + 0.788, + 0.548 + ], + "angle": 0, + "content": "Table 5: Quantitative comparisons on synthetic scenes with DRSC [26] + NeRF and USB-NeRF [19]. We color code each result as best and second best." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.788, + 0.717 + ], + "angle": 0, + "content": "DeepUnroll [22], CVR [10], and JAMNet [11]. To evaluate out-trajectory views, we selected several baselines for comparison. The most basic approach is to train NeRF directly using RS images. We further compare our method with two-stage baselines that initially apply state-of-the-art 2D RS correction methods (DeepUnroll [22], CVR [10], and JAMNet [11]) to restore GS images, followed by training NeRF with these restored images. For 2D RS correction methods, we utilized the official checkpoints for evaluation. Furthermore, we compare RS-NeRF with DRSC [26] and USB-NeRF [19]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Quantitative Results. Quantitative comparison results are shown in Tab. 4 and Tab. 5. 2D RS correction methods do not yield satisfactory outcomes and receive obvious lower metrics due to twist artifacts and spatial misalignment. In contrast, our approach significantly exceeds these methods in all metrics from both on-trajectory and out-trajectory views by incorporating the RS formation process into the modeling and utilizing the multi-view information from continuous RS frames. USB-NeRF [19] integrates basic camera motion modeling, i.e., assigning unique camera poses for different rows to account for the RS ef" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.666, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RS-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.331, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.285, + 0.285, + 0.293 + ], + "angle": 0, + "content": "NeRF" + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.146, + 0.445, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.216, + 0.445, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.285, + 0.402, + 0.293 + ], + "angle": 0, + "content": "\\(\\mathrm{DU} + \\mathrm{NeRF}\\)" + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.146, + 0.558, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.216, + 0.558, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.477, + 0.285, + 0.519, + 0.293 + ], + "angle": 0, + "content": "\\(\\mathrm{CVR} + \\mathrm{NeRF}\\)" + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.146, + 0.671, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.216, + 0.671, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.285, + 0.64, + 0.293 + ], + "angle": 0, + "content": "JAM + NeRF" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.146, + 0.784, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.216, + 0.784, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.285, + 0.743, + 0.293 + ], + "angle": 0, + "content": "RS-NeRF" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.305, + 0.785, + 0.333 + ], + "angle": 0, + "content": "Fig. 8: Qualitative comparison results on real-world scenes. Our method achieves superior performance on real-world scenes." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.349, + 0.361, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.35, + 0.478, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.479, + 0.35, + 0.591, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.35, + 0.642, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.35, + 0.735, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.35, + 0.809, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.424, + 0.351, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.35, + 0.424, + 0.408, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.424, + 0.594, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.424, + 0.728, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.424, + 0.784, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.523, + 0.785, + 0.552 + ], + "angle": 0, + "content": "Fig. 9: Qualitative comparison results on synthetic and real-world scenes with DRSC [26] + NeRF and USB-NeRF [19]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.785, + 0.613 + ], + "angle": 0, + "content": "fect, without additional technical improvements. The comparison shows that this approach fails to deliver satisfactory results compared to RS-NeRF." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Qualitative Results. The qualitative performance of our method is compared to others through tests on synthetic and real-world datasets. In synthetic scenes, as shown in Fig. 7, 2D RS correction methods struggle with twisting artifacts and spatial misalignment for on-trajectory views. In contrast, our method takes advantage of multi-view information for RS correction, leading to significantly better results. For out-trajectory views, while 2D RS correction methods combined with NeRF moderately enhance the reconstruction results, they still exhibit noticeable visual flaws, such as blurriness. RS-NeRF excels by integrating the RS formation process into the model and accurately learning the underlying 3D representation for joint RS correction and novel view synthesis. Qualitative results in real world scenes are shown in Fig. 8. We can observe that baseline methods show obvious artifacts, particularly around object edges. Also, the performance of 2D RS correction combined with NeRF is poorer in real-world scenes than in synthetic ones, which further demonstrates the generalization issues of 2D RS correction methods with real-world data. RS-NeRF, on the other hand," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "M. Niu et al." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.144, + 0.568, + 0.228 + ], + "angle": 0, + "content": "
PerturbationON TRAJ.OUT TRAJ.MSTE ↓
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
None27.93.8393.063028.65.8627.0513.0319
U[-0.01, 0.01]27.97.8403.061028.84.8680.0485.0297
U[-0.03, 0.03]28.31.8479.057829.00.8693.0470.0304
U[-0.05, 0.05]28.47.8530.058029.24.8769.0466.0308
U[-0.07, 0.07]28.43.8504.059729.25.8771.0471.0336
" + }, + { + "type": "table_caption", + "bbox": [ + 0.578, + 0.145, + 0.788, + 0.244 + ], + "angle": 0, + "content": "Table 6: Quantitative results on synthetic dataset for different noise perturbations on pose initialization. The results demonstrate that our model is robust to the accuracy of pose initialization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.785, + 0.329 + ], + "angle": 0, + "content": "significantly outperforms these methods with clear edges and detailed rendering, consistent with results in synthetic scenes. Fig. 9 also demonstrates the superiority of RS-NeRF over DRSC [26] + NeRF and USB-NeRF [19] on synthetic and real-world scenes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.348, + 0.486, + 0.364 + ], + "angle": 0, + "content": "4.5 Discussion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.369, + 0.788, + 0.534 + ], + "angle": 0, + "content": "Robustness to pose initialization. We evaluated the robustness of our method to the precision of pose initialization on synthetic dataset by introducing random noise of varying scales to the initial poses in the SE(3) space and using these altered poses for initialization. The results, as detailed in Tab. 6, indicate that our method maintains superior performance across different levels of noise perturbation. These results confirm the robustness of RS-NeRF in the accuracy of pose initialization. However, it is important to note that our method might fail when completely random poses are used for initialization. This challenge is also an ongoing issue in the field of NeRF-without-Poses, as discussed in works like [5, 20]. Addressing this limitation is a direction that we plan to pursue in future research." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.643 + ], + "angle": 0, + "content": "Limitations. Similar to previous works that address NeRF with abnormal inputs [24, 36], our model is designed to handle degradation (referred to the RS effects in our context) caused by camera movements, since we primarily model the camera trajectory. When encountering scenarios where RS effects are caused by object movements, i.e., dynamic scenes, our method may fail and produce unsatisfying results. We plan to involve the modeling of object movements based on dynamic NeRFs in future work." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.661, + 0.36, + 0.677 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we present RS-NeRF, a method designed to synthesize normal images from novel views using input with Rolling Shutter (RS) distortions. We formulate a physical model that replicates the image formation process under RS conditions and jointly optimizes the NeRF parameters and the camera extrinsic for each image row. We then identify and address the fundamental limitations of the standard RS-NeRF model by introducing two innovative approaches: trajectory smoothness regularization and the multi-sampling algorithm. Through rigorous experimentation, we demonstrate that RS-NeRF surpasses previous methods in both synthetic and real-world scenarios, confirming its efficiency in rectifying distortions caused by RS cameras." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.665, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RS-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.174, + 0.788, + 0.205 + ], + "angle": 0, + "content": "This research was supported in part by JSPS KAKENHI Grant Numbers 24K223-18, 22H00529, 20H05951, JST-Mirai Program JPMJMI23G1." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.225, + 0.324, + 0.241 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.253, + 0.787, + 0.295 + ], + "angle": 0, + "content": "1. Albl, C., Kukelova, Z., Larsson, V., Pajdla, T.: Rolling shutter camera absolute pose. IEEE transactions on pattern analysis and machine intelligence 42(6), 1439-1452 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.295, + 0.787, + 0.336 + ], + "angle": 0, + "content": "2. Attal, B., Laidlaw, E., Gokaslan, A., Kim, C., Richardt, C., Tompkin, J., O'Toole, M.: Törf: Time-of-flight radiance fields for dynamic scene view synthesis. Advances in neural information processing systems 34, 26289-26301 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.336, + 0.787, + 0.39 + ], + "angle": 0, + "content": "3. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.39, + 0.787, + 0.445 + ], + "angle": 0, + "content": "4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mipnerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.445, + 0.787, + 0.487 + ], + "angle": 0, + "content": "5. Bian, W., Wang, Z., Li, K., Bian, J.W., Prisacariu, V.A.: Nope-nerf: Optimising neural radiance field with no pose prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4160-4169 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.486, + 0.787, + 0.528 + ], + "angle": 0, + "content": "6. Cao, M., Zhong, Z., Wang, J., Zheng, Y., Yang, Y.: Learning adaptive warping for real-world rolling shutter correction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17785-17793 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.528, + 0.787, + 0.58 + ], + "angle": 0, + "content": "7. Chen, T., Wang, P., Fan, Z., Wang, Z.: Aug-nerf: Training stronger neural radiance fields with triple-level physically-grounded augmentations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15191-15202 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.58, + 0.787, + 0.636 + ], + "angle": 0, + "content": "8. Dave, A., Zhao, Y., Veeraraghavan, A.: Pandora: Polarization-aided neural decomposition of radiance. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part VII. pp. 538–556. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.636, + 0.787, + 0.678 + ], + "angle": 0, + "content": "9. Fan, B., Dai, Y., He, M.: Sunet: symmetric undistortion network for rolling shutter correction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4541-4550 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.678, + 0.787, + 0.718 + ], + "angle": 0, + "content": "0. Fan, B., Dai, Y., Zhang, Z., Liu, Q., He, M.: Context-aware video reconstruction for rolling shutter cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17572-17582 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.787, + 0.759 + ], + "angle": 0, + "content": "1. Fan, B., Mao, Y., Dai, Y., Wan, Z., Liu, Q.: Joint appearance and motion learning for efficient rolling shutter correction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5671-5681 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.787, + 0.8 + ], + "angle": 0, + "content": "2. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: Fastnerf: High-fidelity neural rendering at 200fps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14346-14355 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.787, + 0.841 + ], + "angle": 0, + "content": "3. Guo, Y.C., Kang, D., Bao, L., He, Y., Zhang, S.H.: Nerfren: Neural radiance fields with reflections. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18409-18418 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.253, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "M. Niu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "14. Hedborg, J., Forssén, P.E., Felsberg, M., Ringaby, E.: Rolling shutter bundle adjustment. In: 2012 IEEE Conference on Computer Vision and Pattern Recognition. pp. 1434-1441. IEEE (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.233 + ], + "angle": 0, + "content": "15. Kim, J.H., Latif, Y., Reid, I.: Rrd-slam: Radial-distorted rolling-shutter direct slam. In: 2017 IEEE International Conference on Robotics and Automation (ICRA). pp. 5148-5154. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.785, + 0.26 + ], + "angle": 0, + "content": "16. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.261, + 0.785, + 0.289 + ], + "angle": 0, + "content": "17. Klenk, S., Koestler, L., Scaramuzza, D., Cremers, D.: E-nerf: Neural radiance fields from a moving event camera. IEEE Robotics and Automation Letters (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.29, + 0.785, + 0.331 + ], + "angle": 0, + "content": "18. Klingner, B., Martin, D., Roseborough, J.: Street view motion-from-structure-from-motion. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 953–960 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.332, + 0.785, + 0.374 + ], + "angle": 0, + "content": "19. Li, M., Wang, P., Zhao, L., Liao, B., Liu, P.: USB-neRF: Unrolling shutter bundle adjusted neural radiance fields. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=igfDXfMvm5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.785, + 0.403 + ], + "angle": 0, + "content": "20. Lin, C.H., Ma, W.C., Torralba, A., Lucey, S.: Barf: Bundle-adjusting neural radiance fields. In: IEEE International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.403, + 0.785, + 0.431 + ], + "angle": 0, + "content": "21. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.432, + 0.785, + 0.473 + ], + "angle": 0, + "content": "22. Liu, P., Cui, Z., Larsson, V., Pollefeys, M.: Deep shutter unrolling network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5941-5949 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.785, + 0.515 + ], + "angle": 0, + "content": "23. Liu, P., Zuo, X., Larsson, V., Pollefeys, M.: Mba-vo: Motion blur aware visual odometry. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5550-5559 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.785, + 0.558 + ], + "angle": 0, + "content": "24. Ma, L., Li, X., Liao, J., Zhang, Q., Wang, X., Wang, J., Sander, P.V.: Deblur-nerf: Neural radiance fields from blurry images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12861-12870 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.785, + 0.6 + ], + "angle": 0, + "content": "25. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.785, + 0.629 + ], + "angle": 0, + "content": "26. Qu, D., Liao, B., Zhang, H., Ait-Aider, O., Lao, Y.: Fast rolling shutter correction in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.63, + 0.785, + 0.671 + ], + "angle": 0, + "content": "27. Reiser, C., Peng, S., Liao, Y., Geiger, A.: Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14335-14345 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.672, + 0.785, + 0.713 + ], + "angle": 0, + "content": "28. Rengarajan, V., Balaji, Y., Rajagopalan, A.: Unrolling the shutter: Cnn to correct motion distortions. In: Proceedings of the IEEE Conference on computer Vision and Pattern Recognition. pp. 2291-2299 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.785, + 0.755 + ], + "angle": 0, + "content": "29. Saurer, O., Koser, K., Bouguet, J.Y., Pollefeys, M.: Rolling shutter stereo. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 465-472 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.756, + 0.785, + 0.798 + ], + "angle": 0, + "content": "30. Saurer, O., Pollefeys, M., Lee, G.H.: Sparse to dense 3d reconstruction from rolling shutter images. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3337-3345 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "31. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4104-4113 (2016)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.665, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RS-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.204 + ], + "angle": 0, + "content": "32. Schonberger, J.L., Zheng, E., Frahm, J.M., Pollefeys, M.: Pixelwise view selection for unstructured multi-view stereo. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14. pp. 501-518. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.205, + 0.788, + 0.245 + ], + "angle": 0, + "content": "33. Teed, Z., Deng, J.: Raft: Recurrent all-pairs field transforms for optical flow. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 402-419. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.246, + 0.786, + 0.286 + ], + "angle": 0, + "content": "34. Vasu, S., Rajagopalan, A., et al.: Occlusion-aware rolling shutter rectification of 3d scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 636-645 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.287, + 0.787, + 0.327 + ], + "angle": 0, + "content": "35. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-NeRF: Structured view-dependent appearance for neural radiance fields. CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "36. Wang, P., Zhao, L., Ma, R., Liu, P.: Bad-nerf: Bundle adjusted deblur neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4170-4179 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.371, + 0.787, + 0.411 + ], + "angle": 0, + "content": "37. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5438-5448 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.412, + 0.787, + 0.452 + ], + "angle": 0, + "content": "38. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.453, + 0.787, + 0.493 + ], + "angle": 0, + "content": "39. Zhong, Z., Cao, M., Sun, X., Wu, Z., Zhou, Z., Zheng, Y., Lin, S., Sato, I.: Bringing rolling shutter images alive with dual reversed distortion. In: European Conference on Computer Vision. pp. 233-249. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.494, + 0.787, + 0.535 + ], + "angle": 0, + "content": "40. Zhong, Z., Zheng, Y., Sato, I.: Towards rolling shutter correction and deblurring in dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9219-9228 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.536, + 0.787, + 0.577 + ], + "angle": 0, + "content": "41. Zhou, X., Duan, P., Ma, Y., Shi, B.: Evunroll: Neuromorphic events based rolling shutter image correction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17775-17784 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.578, + 0.787, + 0.619 + ], + "angle": 0, + "content": "42. Zhuang, B., Tran, Q.H., Ji, P., Cheong, L.F., Chandraker, M.: Learning structure-and-motion-aware rolling shutter correction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4551-4560 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.619 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/RS-NeRF_ Neural Radiance Fields from Rolling Shutter Images/cafbceee-b1fc-44b7-b4f6-4691cb686eec_origin.pdf b/2024/RS-NeRF_ Neural Radiance Fields from Rolling Shutter Images/cafbceee-b1fc-44b7-b4f6-4691cb686eec_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..00caa9b60a6e17993feb6dcbc4542461d1a497c0 --- /dev/null +++ b/2024/RS-NeRF_ Neural Radiance Fields from Rolling Shutter Images/cafbceee-b1fc-44b7-b4f6-4691cb686eec_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c606602a8333a6dbef8de93a96b9ea8594eaf5b4814e36ee84691a574d0dc15 +size 3835250 diff --git a/2024/RS-NeRF_ Neural Radiance Fields from Rolling Shutter Images/full.md b/2024/RS-NeRF_ Neural Radiance Fields from Rolling Shutter Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7ad2f2290ef37b23fbe752df08a45140f43aa8f0 --- /dev/null +++ b/2024/RS-NeRF_ Neural Radiance Fields from Rolling Shutter Images/full.md @@ -0,0 +1,350 @@ +# RS-NeRF: Neural Radiance Fields from Rolling Shutter Images + +Muyao Niu, Tong Chen, Yifan Zhan, Zhuoxiao Li, Xiang Ji, and Yinqiang Zheng* + +The University of Tokyo +muyao.niu@gmail.com, cocottt1023@gmail.com, +zhan-yifan@g.ecc.u-tokyo.ac.jp, lizhuoxiao@g.ecc.u-tokyo.ac.jp, +xiangji2016@gmail.com, yqzheng@ai.u-tokyo.ac.jp + +Abstract. Neural Radiance Fields (NeRFs) have become increasingly popular because of their impressive ability for novel view synthesis. However, their effectiveness is hindered by the Rolling Shutter (RS) effects commonly found in most camera systems. To solve this, we present RS-NeRF, a method designed to synthesize normal images from novel views using input with RS distortions. This involves a physical model that replicates the image formation process under RS conditions and jointly optimizes NeRF parameters and camera extrinsic for each image row. We further address the inherent shortcomings of the basic RS-NeRF model by delving into the RS characteristics and developing algorithms to enhance its functionality. First, we impose a smoothness regularization to better estimate trajectories and improve the synthesis quality, in line with the camera movement prior. We also identify and address a fundamental flaw in the vanilla RS model by introducing a multi-sampling algorithm. This new approach improves the model's performance by comprehensively exploiting the RGB data across different rows for each intermediate camera pose. Through rigorous experimentation, we demonstrate that RS-NeRF surpasses previous methods in both synthetic and real-world scenarios, proving its ability to correct RS-related distortions effectively. Codes and data available: https://github.com/MyNiuuu/RS-NeRF + +# 1 Introduction + +CMOS imaging sensors are commonly employed in a wide range of consumer and industrial products. Many of these sensors utilize a Rolling Shutter (RS) mechanism for image capture because of its cost-effectiveness and portability. In contrast to Global Shutter (GS) cameras, which capture all pixels simultaneously, rolling shutter cameras capture image pixels row by row in a sequential manner. Consequently, if the camera is in motion during image capture, visible distortions can occur. It is widely acknowledged that rolling shutter distortions pose challenges for various downstream computer vision tasks, as indicated by + +previous research [1,14,15,18,29,30]. Consequently, the correction of rolling shutter effects has attracted significant attention in the past [11, 22, 28, 34, 39-42]. + +In the emerging field of novel view synthesis, technologies such as NeRF [25] have become increasingly proficient in generating highly realistic renderings. NeRF utilizes a continuous volumetric function that is parameterized by a multilayer perceptron (MLP) to associate 3D locations and 2D view directions with color and density information. It relies on volume rendering techniques to enable differentiable rendering. + +Although traditional NeRF models deliver exceptional results when applied to well-calibrated and well-captured images, they encounter difficulties when dealing with RS effects. For example, in scenarios where the camera is in motion during image capture, these RS effects can significantly impact the performance of the model. Given the widespread use of RS cameras in various industrial applications, it is imperative to take into account this phenomenon when reconstructing 3D representations using NeRFs. + +One straightforward approach involves a two-stage baseline solution, where the first step corrects the rolling shutter effect in 2D image space, and then the NeRF is trained using these corrected images. While these baseline methods enhance the quality of novel view synthesis in NeRF to some extent by utilizing 2D rolling shutter correction techniques, they can not fully explore the 3D scene's geometry, leading to inaccuracies in correspondences, especially when there is significant camera motion. + +In this paper, we introduce RS-NeRF, a robust framework designed to explicitly incorporate the physics-based formation of rolling shutter effects into the rendering process. RS-NeRF has the ability to reconstruct a high-quality NeRF representation from only input data that contains rolling shutter artifacts. To achieve this, we explicitly model the camera's motion trajectory throughout the exposure duration for each frame and combine the RGB rendering outcomes from different camera poses to generate the rolling shutter image. Throughout this procedure, we optimize both the NeRF network and the pose estimation network simultaneously by minimizing the photometric loss. + +Building upon the basic RS-NeRF framework, we delve into its shortcomings and enhance its functionality through various refinements. Initially, we impose a camera trajectory smoothness regularization to refine our model's trajectory prediction, by leveraging the sequential nature of camera movement to yield improved results. We further identify a critical shortcoming due to the vanilla RS camera motion model and introduce a multi-sampling technique that significantly enhances RS-NeRF's performance. Specifically, we reveal the issue of the standard camera pose formulation, which assigns distinct poses per row, potentially limiting RS-NeRF's efficiency due to limited training data for each pose. We then make a key observation that multiple poses may share the same 2D RGB observation for the same 3D point during the exposure time, which can drastically increase the available training data. To capitalize on this insight, we introduce and compute the PP-ratio—a measure of camera pose movement in 3D space relative to pixel movement in 2D space, and apply this to broaden our + +![](images/29aee7f762130fd74118c5611f13df3e3cc1fcf90187e89e22bfd3109dce19e1.jpg) +Fig. 1: Pipeline of vanilla RS-NeRF. Given a series of images with RS distortion, RS-NeRF learns the underlying normal 3D representations by jointly estimating pose and RGB values for each row. For each RS image, we first sample the rows and obtain their camera poses via the pose estimation network. We then sample points for each row and feed them with the estimated camera pose to the NeRF network and optimize the photo-metric loss with ground truths. + +training data by sampling out-of-row pixel points. This approach significantly bolsters our model's performance by tapping into the inherent data augmentation possibilities presented by the RS formation process. + +We collect both synthetic and real data to evaluate our proposed RS-NeRF. The experimental results reveal that RS-NeRF outperforms the preceding models and demonstrate the effectiveness of our proposed refinements on the vanilla model. In summary, our contributions are as follows: + +- We propose a framework that can reconstruct NeRF from inputs with Rolling Shutter (RS) effects by modeling the camera trajectory and aggregating RGB rendering results of different poses during the exposure time. +- By analyzing the vanilla framework in detail, we proposed a series of improvements to it, including trajectory smoothness regularization and the multi-sampling algorithm. +- We contribute synthetic and real datasets to evaluate our framework and future works. + +# 2 Related Work + +2D Rolling Shutter Correction. Rolling Shutter Correction (RSC) aims to restore normal images given the input with RS distortions. In recent years, several appealing image-based RSC methods [22,28,34,39-42] have been developed. These methods take either single RS image or continuous RS frames as input and predict their global shutter (GS) counterparts. However, single-image RSC can be seriously ill-posed and struggle to correctly estimate the GS counterpart in complex scenarios. To reduce the ill-posedness of single-image RSC, more attention has been paid to multi-image RSC [6, 9-11, 22, 40]. However, these multi-image RS correction methods, which are built on image space operations, fail to exploit the 3D geometry of the scene. As a result, these methods produce multi-view inconsistency results which largely impact the reconstruction results + +of NeRF. In contrast, our proposed method can jointly reconstruct 3D geometry and correct RS images by effectively aggregating multi-view information of the scene. + +Neural Radiance Fields (NeRF). NeRF [25] is an end-to-end model that synthesizes novel views of complex 3D scenes by optimizing an implicit continuous volumetric scene function encoded by an MLP. Due to its simplicity in training and high-quality rendering results, NeRF-based methods [3,4,7,12,21,27,35,37] have attracted intensive attention from various computer vision fields, leading to significant improvements in terms of both accuracy and efficiency. NeRF-based models have also inspired many subsequent works that extend their continuous neural representation to different setups [2,8,13,17,24,36]. Deblur-NeRF [24] and BAD-NeRF [36] reconstruct sharp images given the blurry input from different view directions. One concurrent work [19] appears in ICLR'24, which formulates the RS process with B-cubic models. This work combines the basic modeling of camera motion with the NeRF network to consider the RS effect, without further technical improvements. In this paper, we first bake the RS model into NeRF formulation, then analyze several inherit limitations of the vanilla model, and propose a series of algorithms to further improve its performance. + +# 3 RS-NeRF + +In this section, we present the specifics of RS-NeRF. Initially, we explain the basic formulation of RS-NeRF in Sec. 3.1. We also note that the vanilla model possesses specific shortcomings that hinder its effectiveness. To address these, we suggest several improvements to enhance its capabilities, which are detailed in Sec. 3.2 and Sec. 3.3. + +# 3.1 Vanilla RS-NeRF Model + +Rolling shutter camera model. Unlike Global Shutter (GS) cameras, the Rolling Shutter (RS) camera captures each row at distinct timestamps. Without loss of generality, we consider the readout direction of the RS camera to be from top to bottom, assuming an infinitesimal exposure time for each row, following existing RS modeling techniques [10, 11, 22]. This mechanism can be described through a mathematical model as follows: + +$$ +[ \mathbf {I} ^ {R} (\mathbf {x}) ] _ {i} = [ \mathbf {I} _ {i} ^ {G} (\mathbf {x}) ] _ {i}, \tag {1} +$$ + +where $[\mathbf{I}(\mathbf{x})]_i$ is an operator to extract the $i$ -th row from an image $\mathbf{I}(\mathbf{x})$ , $\mathbf{I}^R(\mathbf{x}) \in \mathbb{R}^{H \times W \times 3}$ is the rolling shutter image, and $\mathbf{I}_i^G(\mathbf{x}) \in \mathbb{R}^{H \times W \times 3}$ is the virtual global shutter image captured at the same pose as the $i$ -th row of $\mathbf{I}^R(\mathbf{x})$ . As a result, for one RS image $\mathbf{I}^R(\mathbf{x})$ , we denote the pose of the $i$ -th row as $\mathbf{T}_i$ . + +Camera motion trajectory modeling. In our rolling shutter camera model, it is necessary to establish the poses for each row in each image. Given the typically short exposure time of an RS image, we use a linear model to approximate the + +![](images/eabc1255cf41091543613c6aa07177d3d28ab0faca1e30bcaecff64dd6293235.jpg) +Fig. 2: Smoothness regularization on camera trajectory. For each pair of adjacent trajectory vectors $\vec{\mathbf{d}}_k$ and $\vec{\mathbf{d}}_{k + 1}$ , we compute the mid-point trajectory vector $\vec{\mathbf{d}}_{mid}$ and their unit vectors $\vec{\mathbf{n}}_k$ , $\vec{\mathbf{n}}_{k + 1}$ , and $\vec{\mathbf{n}}_{mid}$ . We then apply the $L_{2}$ regularization between $\vec{\mathbf{n}}_{mid}$ and mean $(\vec{\mathbf{n}}_k,\vec{\mathbf{n}}_{k + 1})$ . + +motion of the camera. Specifically, we start by defining two camera poses: one for the first row, denoted as $\mathbf{T}_0$ , and one for the last row, $\mathbf{T}_{H - 1}$ . For the rows in between, we linearly interpolate their poses within the Lie algebra of $\mathbf{SE}(3)$ as informed by prior research [20, 23, 36]: + +$$ +\mathbf {T} _ {i} = \mathbf {T} _ {0} \cdot (1 - \rho) + \mathbf {T} _ {H - 1} \cdot \rho , \tag {2} +$$ + +where the interpolation coefficient $\rho = \frac{i}{H - 1}$ is based on the row index $i$ . Thus, for each RS image, RS-NeRF receives one pose initialization, then optimizes the poses at both ends and then determines the poses for each intermediate row through linear interpolation. This interpolation method for calculating each row's pose is fully differentiable, enabling the optimization of pose parameters using a series of rolling shutter images. + +Optimizing NeRF from RS observation. For a set of $K$ rolling shutter images, we select $M$ rows per image and $N$ points per row, resulting in $K \times M \times N$ distinct camera rays for each iteration. These rays are used to concurrently optimize the NeRF network and the pose parameters. Specifically, for each image, we initially use the row indices to access the pose estimation network. This allows us to determine the camera ray $\hat{\mathbf{r}}(\theta_p; u, v)$ for each sampled pixel point $(u, v)$ . Following this, these rays are employed to train the NeRF network by computing the ensuing photometric loss: + +$$ +\mathcal {L} = \left| \left| \mathbf {C} (\mathbf {r} (u, v)) - \hat {\mathbf {C}} \left(\theta_ {n}; \hat {\mathbf {r}} \left(\theta_ {p}; (u, v)\right) \right. \right| \right| _ {2}, \tag {3} +$$ + +where $\theta_{n}$ represents the trainable parameters of the NeRF network, and $\theta_{p}$ denotes the learnable parameters for camera poses. $\hat{\mathbf{C}} (\theta_n;\hat{\mathbf{r}} (\theta_p;(u,v))$ refers to the predicted RGB value for the pixel at coordinates $(u,v)$ , while $\mathbf{C}(\mathbf{r}(u,v))$ denotes the corresponding RGB ground truth value. The workflow of our vanilla RS-NeRF model is depicted in Fig. 1. + +# 3.2 Trajectory Smoothness Regularization. + +Given the previous description of our model, it is evident that accurately estimating the camera pose for each row in RS frames is crucial for effectively reconstructing the normal NeRF representation. We observe that the basic RS-NeRF struggles with precise camera trajectory estimation, often resulting in high-frequency predictions, as further analyzed in our experiments. This issue + +
NposeON TRAJ. / OUT TRAJ.
PSNR ↑SSIM ↑LPIPS ↓
10024.36/25.16.7321/.7727.0703/.0578
40024.06/24.66.7245/.7536.0697/.0579
+ +Table 1: Quantitative results for different $N_{\text{pose}}$ . $N_{\text{pose}} = 100$ yields better results than $N_{\text{pose}} = 400$ . + +![](images/a958c75ed494c3d7fceb1b81c4939f267a7da9f8a257120e2a3579d22b9f3639.jpg) +Fig. 3: Qualitative comparisons for different $N_{\text{pose}}$ . The artifact decreases as $N_{\text{pose}}$ grows. + +arises because the network focuses only on the photometric loss for individual images, neglecting the consistent movement characteristics of the camera across different RS frames. To address this, we introduce a camera trajectory smoothness regularization, based on the expectation of a smooth and continuous camera trajectory in 3D space. Specifically, for every two adjacent RS frames $\mathbf{I}_k^R$ and $\mathbf{I}_{k + 1}^R$ , we aim to minimize the following loss during the optimization process: + +$$ +\mathcal {L} _ {s} = | | \vec {\mathbf {n}} _ {m i d} - \operatorname {m e a n} (\vec {\mathbf {n}} _ {k} + \vec {\mathbf {n}} _ {k + 1}) | | _ {2}. \tag {4} +$$ + +As illustrated in Fig. 2, $\vec{\mathbf{n}}_k$ and $\vec{\mathbf{n}}_{k + 1}$ represent the unit direction vectors of the estimated camera trajectories for RS frames $\mathbf{I}_k^R$ and $\mathbf{I}_{k + 1}^R$ , respectively. Meanwhile, $\vec{\mathbf{n}}_{mid}$ denotes the unit direction vector for the camera trajectory during the exposure interval between these two RS frames. + +# 3.3 Multi-Sampling Algorithm + +As discussed in Sec. 3.1, each row in an RS frame is assigned a unique camera pose. For example, in an RS frame with 400 rows, there are 400 distinct camera poses. This rigorous approach aligns with the real-world, physics-based photography process of RS cameras but also presents certain challenges. For example, under this framework, the data points for each pose are confined to a single row. This limitation can significantly affect our model's performance due to the sparse training data available for each pose. + +A straightforward solution to this problem is to reduce the number of poses per image. For example, assigning 200 poses to each RS frame means that all two adjacent rows would share the same pose. This approach can lead to blurriness since these rows are captured at different poses, but it also doubles the amount of data available for training each pose. To further explore this trade-off, we conducted a series of experiments. Fig. 3 shows qualitative results for different intermediate pose numbers $N_{\text{pose}}$ , indicating that the artifacts decrease as $N_{\text{pose}}$ increases. Interestingly, we observed that using 100 poses per frame often yields results comparable to or even better than using 400 poses. The quantitative results presented in Tab. 1 also support the superiority of having 100 poses per frame. + +This observation seems counter-intuitive at first, as one might assume that having fewer than 400 poses per frame would lead to poorer results due to + +![](images/3518bdd366c104f04ccdaff54f8406ce473f531de5996e9e5498d56150131e06.jpg) +Fig. 4: Motivation for the multi-sampling algorithm. Up: Three consecutive RS frames, labeled as (a), (b), and (c), along with the optical flow between each pair of frames, shown in (d) and (e). Down: For a single point in 3D space, we observe its 2D projections in the first (marked in yellow) and second (marked in red) RS frames. These projections exhibit a row displacement of 8 and a column displacement of 18, as depicted in (f). However, during this interval, the camera undergoes a shift of at least 392 poses, as indicated in (g). + +artifacts like blurriness. However, it can be rationally explained. As demonstrated in Fig. 4, for a given point in 3D space, the 2D projections on the first and second RS frames show relatively small pixel displacements (8 in row and 12 in column). Meanwhile, the camera shifts through at least 392 poses in 3D space. This is based on the assumption that there is no overlap in the trajectories of two adjacent frames (indicated as $\geq 0$ poses' in the figure), which is naturally satisfied given the sparse observation setting and short exposure times typically involved. + +Following our prior analysis, we introduced a multi-sampling algorithm aimed at explicitly increasing the sample points for each intermediate camera pose. The fundamental concept involves computing the ratio of camera pose displacement to pixel displacement (termed the PP-ratio) between two adjacent frames for the same 3D point. For example, a pixel displacement of 18 combined with a camera pose shift of 392 results in a PP-ratio of 21. This implies that at least 21 consecutive camera poses share the same pixel value, which means that all these 21 camera poses can utilize this RGB value for training purposes. + +For a pixel $\mathbf{x} = (u,v)$ located in the $u$ -th column and $v$ -th row of the $k$ -th RS frame $\mathbf{I}_k^R (\mathbf{x})$ , we determine both the forward PP-ratio $\rho_{k\to k + 1}(\mathbf{x})$ and the backward PP-ratio $\rho_{k\to k - 1}(\mathbf{x})$ . This is because the point represented by $\mathbf{x}$ can transition from $\mathbf{I}_k^R (\mathbf{x})$ to $\mathbf{I}_{k + 1}^R (\mathbf{x})$ or $\mathbf{I}_{k - 1}^R (\mathbf{x})$ . Consequently, all camera poses within the range of $[v - \rho_{k\to k - 1}(\mathbf{x}), v + \rho_{k\to k + 1}(\mathbf{x})]$ are capable of sampling the RGB value of $\mathbf{x}$ for training purposes. + +To compute the forward PP-ratio $\rho_{k\rightarrow k + 1}(\mathbf{x})$ for a pixel $\mathbf{x}$ in the $k$ -th RS frame $\mathbf{I}_k^R (\mathbf{x})$ , the first step is to estimate the forward optical flow $F_{k\rightarrow k + 1}(\mathbf{x})$ , utilizing RAFT [33]. Consequently, the row displacement $\varDelta_{k\rightarrow k + 1}^{row}(\mathbf{x})$ and column + +![](images/edddefe08157b725531a95a409a992c1b2e71d0336d6ad8070f0ff4b3f55ad39.jpg) +Fig. 5: Comparing the accessible training points in one RS image under different strategies. The image size is $H \times W$ , providing $H$ poses. Distinct circle hues represent different RGB values. In case $C$ , pixels transported from other poses are with dash contours, and the number signifying that it originates from the $i$ -th pose. + +displacement $\varDelta_{k\rightarrow k+1}^{col}(\mathbf{x})$ of $\mathbf{x}$ from $\mathbf{I}_k^R (\mathbf{x})$ to $\mathbf{I}_{k + 1}^R (\mathbf{x})$ would be the second and the first element of $F_{k\to k + 1}(\mathbf{x})$ : + +$$ +F _ {k \rightarrow k + 1} (\mathbf {x}) = \operatorname {R A F T} \left(\mathbf {I} _ {k} ^ {R} (\mathbf {x}), \mathbf {I} _ {k + 1} ^ {R} (\mathbf {x})\right), \tag {5} +$$ + +$$ +\varDelta_ {k \rightarrow k + 1} ^ {c o l} (\mathbf {x}), \varDelta_ {k \rightarrow k + 1} ^ {r o w} (\mathbf {x}) = F _ {k \rightarrow k + 1} (\mathbf {x}), \tag {6} +$$ + +and the calculation are similar for the backward situation. + +We then calculate the forward and backward camera pose shift as: + +$$ +\begin{array}{l} \Delta_ {k \rightarrow k + 1} ^ {\text {p o s e}} (\mathbf {x}) = v + \Delta_ {k \rightarrow k + 1} ^ {\text {r o w}} (\mathbf {x}) + (H - v), (7) \\ = H + \Delta_ {k \rightarrow k + 1} ^ {\text {r o w}} (\mathbf {x}), (8) \\ \end{array} +$$ + +$$ +\begin{array}{l} \Delta_ {k \rightarrow k - 1} ^ {\text {p o s e}} (\mathbf {x}) = \left(H - \left(v + \Delta_ {k \rightarrow k - 1} ^ {\text {r o w}} (\mathbf {x})\right)\right) + v, (9) \\ = H - \Delta_ {k \rightarrow k - 1} ^ {\text {r o w}} (\mathbf {x}). (10) \\ \end{array} +$$ + +Finally, we calculate forward and backward PP-ratio as: + +$$ +\rho_ {k \rightarrow k + 1} (\mathbf {x}) = \frac {\Delta_ {k \rightarrow k + 1} ^ {p o s e} (\mathbf {x})}{\max \left(\Delta_ {k \rightarrow k + 1} ^ {c o l} (\mathbf {x}) , \Delta_ {k \rightarrow k + 1} ^ {r o w} (\mathbf {x})\right)}, \tag {11} +$$ + +$$ +\rho_ {k \rightarrow k - 1} (\mathbf {x}) = \frac {\Delta_ {k \rightarrow k - 1} ^ {p o s e} (\mathbf {x})}{\operatorname* {m a x} \left(\Delta_ {k \rightarrow k - 1} ^ {c o l} (\mathbf {x}) , \Delta_ {k \rightarrow k - 1} ^ {r o w} (\mathbf {x})\right)}. \tag {12} +$$ + +Therefore, when dealing with a sampled pixel $RGB(\mathbf{x})$ in the RS frame $\mathbf{I}_k^R (\mathbf{x})$ , where $\mathbf{x} = (u,v)$ , rather than directly using the $v$ -th pose $\mathbf{T}_v$ of $\mathbf{I}_k^R (\mathbf{x})$ for training, we randomly sample $\mathbf{T}_{\hat{v}}$ where + +$$ +\hat {v} \in [ v - \rho_ {k \rightarrow k - 1} (\mathbf {x}), v + \rho_ {k \rightarrow k + 1} (\mathbf {x}) ]. \tag {13} +$$ + +Why does Multi-Sampling work? Let us consider a basic task: optimizing NeRF with $H$ known poses and images (non-RS). Case $A$ : Only one row is available per pose. Case $B$ : all rows are accessible. Clearly, case $B$ would yield a better NeRF due to more data per pose. Moving to our RS setting as depicted + +
LssMSTORIIWINEPOOLFACTORYTANABATAAVERAGE
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
ON TrBAI✓23.45.7535.044823.53.7019.068122.99.6111.078419.53.5018.107217.86.4921.103021.47.6121.0803
✓22.25.6729.056425.60.7749.071227.22.7705.058221.81.6128.094923.42.7916.067924.06.7245.0697
✓26.65.8572.034225.55.7637.068225.86.7303.055525.45.7754.070820.94.6802.083224.89.7614.0624
✓30.97.9058.039528.36.8482.067528.50.8137.056225.43.7673.086726.40.8614.065127.93.8393.0630
OUT TrBAI✓23.26.7561.042824.05.7169.060123.27.6218.073819.98.5426.089217.85.5065.094221.68.6288.0720
✓22.39.6875.050126.62.8113.057927.29.7783.048922.72.6624.081524.27.8286.051424.66.7536.0579
✓26.58.8611.030825.96.7750.061225.49.7080.054924.38.7446.074220.73.6850.075124.63.7547.0592
✓33.07.9499.024329.27.8703.056028.77.8313.045825.11.7738.080727.02.8881.049828.65.8627.0513
+ +Table 2: Quantitative results on synthetic scenes for ablation study. $\mathcal{L}_s$ denotes the camera trajectory smoothness regularization, and 'MS' denotes the multi-sampling algorithm. We color code each result as best and second best. + +in Fig.5, case $A$ is the vanilla RS observation for one RS frame ( $H$ rows lead to $H$ poses), and case $B$ is the virtual GS observation, where all rows are available for each pose. Thus, case $B$ is the upper bound for case $A$ . Essentially, MS aims to transport all possible pixel values from nearby poses to fill the missing observation that can be potentially recovered (case $C$ in Fig.5), i.e., partially restore the virtual GS observation from the vanilla RS observation, by allocating additional poses for each pixel, which ensures better results than vanilla RS. Also, interpolating two poses does not harm MS. We optimize two poses for each frame, and each in-between pose becomes a fixed blend ratio of two poses. The best case is virtual GS (Fig.5(B)), where every pixel coordinate can be sampled in all ratios $\left(\frac{0}{H}, \frac{1}{H-1}, \dots, \frac{H}{0}\right)$ , resulting in a total of $H \times W \times H$ data points to optimize the two poses. In vanilla RS (Fig.5(A)), each coordinate only contributes to one ratio determined by row indices, producing only $1 \times W \times H$ data points. With MS (Fig.5(C)), some pixels are replicated to other poses, allowing multiple ratios for them. Thus, the number of data points is greater than $1 \times W \times H$ but less than $H \times W \times H$ . This ensures better pose estimation and view synthesis than vanilla RS. + +# 4 Experiments + +# 4.1 Settings + +Datasets. Evaluating RS-aware view synthesis technologies requires synthetic datasets that have multi-view images with RS distortions and the corresponding GS ground truth. We synthesize 5 scenes via blender to evaluate our method. For each scene, we rendered images from various camera poses while moving the camera gradually around the scene in a forward-facing manner. We then combined rows from different poses to produce the final RS images. Each scene comprises 34 images with a resolution of $400 \times 400$ . We employed the pose of the middle row of each RS image as the initial camera pose for our model. We also analyze our model's robustness to pose initialization in the experiment section. For evaluation, we used 34 on-trajectory views and 16 out-trajectory views to test RS correction and novel view synthesis results, respectively. To test our model in real-world scenarios, we also captured real scenes using the RS camera + +![](images/4dc0615e6757afa002d38481f9f272938e24e0642c8abd588c3179c66348ebf2.jpg) +Vanilla +Fig.6: Qualitative results on synthetic scenes for ablation study. The corresponding error maps and zoomed regions are visualized in the bottom. Our full model has the smaller error to the GT, with less spatial misalignment and twist caused by the RS distortion. + +![](images/15da6701381db8447d3707d752b150f1bfe1b6ded98c90e2ef29e42158410eb8.jpg) +Vanilla $+\mathcal{L}_s$ + +![](images/a41cd36dfedf0ed62189e995797b3210080a2467e43af63fdf4715258a5ae6d8.jpg) +Vanilla $^+$ Mul.Sam. + +![](images/04ffffeba08f72f7190ff0173a9d678dbdd1d821ba51718842f34fc909e0b661.jpg) +\[ +\begin{array}{r}\mathsf{Vanilla} + \mathcal{L}_s + \mathsf{Mul.Sam}. \end{array} +\] + +![](images/23e74af092b9a5b26580d441a02a2e37410b5d0337f05beb05b844878fa9d247.jpg) +Ground Truth + +
VanillaVanilla+LsVanilla+MSVanilla+Ls+MS
0.05090.03580.04620.0319
+ +Table 3: Average MSTE (the lower the better) results for ablation study. + +EO-1312C. These RS images were captured by manually moving the camera during exposure, in a forward-facing manner, with a resolution of $400 \times 500$ . We utilize COLMAP [31,32] for the initial camera pose in our model. + +Evaluation Metrics. The quality of the rendered image is evaluated with the commonly used metrics including PSNR, SSIM and LPIPS [38]. To evaluate the accuracy of the estimated camera trajectory, we calculate the Mean Square Trajectory Error (MSTE) with the ground-truth camera trajectory. + +# 4.2 Implementation Details + +In each training iteration, we randomly choose 64 rows from each of the images and 2 pixels from each of these rows. We employ the Adam optimizer [16] with $\beta_{1} = 0.9$ and $\beta_{2} = 0.999$ to optimize our model. The learning rate is initially set at $5 \times 10^{-4}$ and undergoes exponential decay throughout the optimization process. To avoid potential errors from inaccurate optical flow estimation, we disable multi-sampling and use the vanilla sampling strategy in the later part of the training process. This makes the model use much fewer but absolutely accurate data points in the final training phase. Each scene is trained for 200,000 iterations. All this training is conducted on a single NVIDIA 4090 GPU. + +# 4.3 Ablation Study + +To comprehensively evaluate the effectiveness of each component in our RS-NeRF model, we carried out extensive ablation studies. The quantitative and qualitative results are presented in Tab. 2, Fig. 6 and Tab. 3, respectively. + +Trajectory Smoothness Regularization. The accurate estimation of camera trajectory is crucial for effectively restoring normal images from RS inputs, given + +
MethodsTORIIWINEPOOLFACTORYTANABATAAVERAGE
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
ON-TRIAL
\(DU\)21.236409.084519.87.4520.132923.08.5394.118919.95.5128.132518.17.5262.121320.46.5343.1180
\(CVR\)19.946148.145418.86.4090.205221.51.5173.188518.59.4851.188317.45.5277.184819.27.5108.1824
\(JAM\)20.666311.087219.50.4214.159522.43.5180.126319.34.4907.151017.87.5342.129219.96.5911.1307
RS-NeRF30.97.9058.039528.36.8482.067528.50.8137.056225.43.7673.086726.40.8614.065127.93.8393.0630
OUT-TRIAL
\(NeRF\)19.35.5976.253519.06.3845.355022.02.4740.319617.87.4075.357017.56.4958.301919.17.4719.3174
\(DU+NeRF\)20.33.6585.252820.85.4923.340423.25.5403.286019.16.5037.321918.14.5226.310220.35.5435.3022
\(CVR+NeRF\)23.67.7779.196722.27.6114.279124.56.6802.240819.81.5619.299618.58.6250.273621.78.6513.2580
\(JAM+NeRF\)24.13.7852.151222.14.5801.282225.51.6867.185520.46.5834.271718.63.6357.238822.18.6542.2259
RS-NeRF33.07.9499.024329.27.8703.056028.77.8313.045825.11.7738.080727.02.8881.049828.65.8627.0513
+ +Table 4: Quantitative comparison on synthetic scenes of different methods. DU, CVR, and JAM represent DeepUnroll [22], CVR [10], and JAMNet [11] respectively. We color code each result as best and second best. + +the inherent formation process of RS effects. As evidenced in Tab. 2, RS-NeRF enhanced with trajectory smoothness regularization $(\mathcal{L}_s)$ shows notable improvements over the vanilla model. Specifically, it achieves a PSNR increase of 2.59 for on-trajectory views and 2.98 for out-trajectory views. Additionally, removing $\mathcal{L}_s$ from our full model leads to a PSNR decrease of 3.04 for on-trajectory views and 4.02 for out-trajectory views. The qualitative results further underscore the effectiveness of $\mathcal{L}_s$ . As illustrated in Fig. 6, the vanilla RS-NeRF, when complemented with $\mathcal{L}_s$ , exhibits reduced spatial misalignment and a smaller synthesis error to the ground truth. Furthermore, Tab. 3 reveals that the vanilla RS-NeRF tends to provide an inaccurate trajectory estimation, primarily focusing on the photo-metric loss for each image and overlooking the pattern of camera movement across different RS frames. In contrast, the 'Vanilla + $\mathcal{L}_s$ ' model, with the addition of smoothness regularization, yields a more accurate estimation of the camera trajectory, resulting in improved synthesis outcomes. + +Multi-Sampling Algorithm. The Multi-sampling algorithm in RS-NeRF significantly enhances the training data available for each intermediate pose by calculating the PP-ratio for each pixel, marking a notable advancement over the vanilla model. According to the results in Tab. 2, RS-NeRF equipped with the multi-sampling algorithm shows substantial improvements: a PSNR increase of 3.42 for on-trajectory views and 2.95 for out-trajectory views. Moreover, excluding the multi-sampling algorithm from our full model leads to a PSNR reduction of 3.87 for on-trajectory views and 3.99 for out-trajectory views. The qualitative results also affirm the effectiveness of it. As depicted in Fig. 6, our model, with the multi-sampling algorithm, displays fewer artifacts and a smaller synthesis error to the ground truth. Interestingly, the multi-sampling algorithm also seems to enhance trajectory estimation. As shown in Tab. 3, RS-NeRF with only the multi-sampling algorithm (Vanilla + MS) provides a more accurate trajectory estimation than the vanilla model, which likely contributes to its significantly improved synthesis results. + +# 4.4 Comparisons + +To evaluate the effectiveness of RS correction for on-trajectory images, our method is compared with three state-of-the-art 2D RS correction techniques: + +![](images/b2064caa354974159a7359d9d56da512a4b4b5a676905666150776e8ace0afda.jpg) +Fig. 7: Qualitative comparison results on synthetic scenes. The corresponding error maps and zoomed regions are visualized in the bottom. DU, CVR, and JAM represent DeepUnroll [22], CVR [10], and JAMNet [11] respectively. + +
MethodsTORIIWINEPOOLFACTORYTANABATAAVERAGE
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
USB-NeRF [19]26.32.8283.082524.72.7105.167326.72.7130.159224.32.7213.136923.05.7359.133325.02.7418.1359
DRSC [26]23.04.8008.053921.84.6895.088022.85.6962.087420.85.6901.096721.19.7492.073121.95.7251.0798
RS-NeRF30.97.9058.039528.36.8482.067528.50.8137.056225.43.7673.086726.40.8614.065127.93.8393.0630
USB-NeRF [19]28.95.8932.073625.75.7446.161928.13.7765.157625.25.7676.142724.62.7853.120426.54.7935.1313
DRSC [26] + NeRF21.51.6790.283120.30.5145.300122.66.5614.002118.99.5295.290418.35.5643.296220.36.5697.2944
RS-NeRF33.07.9499.024329.27.8703.056028.77.8313.045825.11.7738.080727.02.8881.049828.65.8627.0513
+ +Table 5: Quantitative comparisons on synthetic scenes with DRSC [26] + NeRF and USB-NeRF [19]. We color code each result as best and second best. + +DeepUnroll [22], CVR [10], and JAMNet [11]. To evaluate out-trajectory views, we selected several baselines for comparison. The most basic approach is to train NeRF directly using RS images. We further compare our method with two-stage baselines that initially apply state-of-the-art 2D RS correction methods (DeepUnroll [22], CVR [10], and JAMNet [11]) to restore GS images, followed by training NeRF with these restored images. For 2D RS correction methods, we utilized the official checkpoints for evaluation. Furthermore, we compare RS-NeRF with DRSC [26] and USB-NeRF [19]. + +Quantitative Results. Quantitative comparison results are shown in Tab. 4 and Tab. 5. 2D RS correction methods do not yield satisfactory outcomes and receive obvious lower metrics due to twist artifacts and spatial misalignment. In contrast, our approach significantly exceeds these methods in all metrics from both on-trajectory and out-trajectory views by incorporating the RS formation process into the modeling and utilizing the multi-view information from continuous RS frames. USB-NeRF [19] integrates basic camera motion modeling, i.e., assigning unique camera poses for different rows to account for the RS ef + +![](images/6657998b1800206c5019007af586d6b6f12243ea3fe9b38d97b62666fd9c90bc.jpg) +NeRF + +![](images/cfa306c44d6347bbb3f520c2194330c69089855cb9de3c26cc1e8642aab4a1ce.jpg) + +![](images/1bdd2a722cd3e90373eb59c57669a2f2185cc660bdddbc980d5d2408c10e4097.jpg) +$\mathrm{DU} + \mathrm{NeRF}$ + +![](images/581d6c56a6f8e06a1ea330c5266424fa921df0ce765f99106986eb1cd9baf294.jpg) + +![](images/eda2bdc76b23d396df6c13aaf3be996d55e54061ebac24968905db330f7448c2.jpg) +$\mathrm{CVR} + \mathrm{NeRF}$ + +![](images/2f50c8a6d9c15e61e37ba3248d4847d5dc46e3c6372b4587bba94a72522da7c2.jpg) + +![](images/3ba84ca3386737e101adba93a607f5587b02ebd79f2487527b4f44950c03c362.jpg) +JAM + NeRF + +![](images/40dd9cb377f2b751aff43979b3ac5efade0c1448402622ed2cbb8faa9dc0a85b.jpg) + +![](images/12d2bcffde69cc2dbf2a2575efd95ab879bd09648d36d46b79229e76622564b2.jpg) +RS-NeRF + +![](images/a966797d84ebcea10131cc5b272aa06e38607dd32a85c7e7679b1d25614d5084.jpg) +Fig. 8: Qualitative comparison results on real-world scenes. Our method achieves superior performance on real-world scenes. + +![](images/cee2f2f030e6a46a5b34df4307c1fd2c8a9b7c36f6098ae17a6c339f198f3a96.jpg) + +![](images/43dabddcc826c9a19751822818e9e1ece21f5a01cfee20f9e16930d3e82a2c70.jpg) + +![](images/e01f50d109a9aa8bacf9fddceb5a251544a8836f127f27474686ffb6338a4f99.jpg) + +![](images/1b1d3c533dbfb143d2691a3ad57891c13634e753519c83df1d6b9200bd39f2d5.jpg) + +![](images/34a9457366ba9351f0e9ce7310caa70571d843988052656834c6accd1e89d893.jpg) + +![](images/2b4fcce7d7e09f08c1876aea07e47000c65de683dd42c6fa67fba99f1b673f51.jpg) +Fig. 9: Qualitative comparison results on synthetic and real-world scenes with DRSC [26] + NeRF and USB-NeRF [19]. + +![](images/9699e66dd9bb65be2839711c92685d8455938ac7cf77f391f2f1a372c8adfc72.jpg) + +![](images/0d6c6bc14c5273b51a44eb1048bce31fd3f4a449453b53b7d8a8fbbc8e84d992.jpg) + +![](images/cde5fd59937dfcff3c61b1ac3f17a8d1d821eb0bf23c9c7f8197dc546a4a4cc1.jpg) + +![](images/001798100d7cb8cf2de661f89052da8f22a8ba7d07e0d2eeb3ece16e2202b48b.jpg) + +fect, without additional technical improvements. The comparison shows that this approach fails to deliver satisfactory results compared to RS-NeRF. + +Qualitative Results. The qualitative performance of our method is compared to others through tests on synthetic and real-world datasets. In synthetic scenes, as shown in Fig. 7, 2D RS correction methods struggle with twisting artifacts and spatial misalignment for on-trajectory views. In contrast, our method takes advantage of multi-view information for RS correction, leading to significantly better results. For out-trajectory views, while 2D RS correction methods combined with NeRF moderately enhance the reconstruction results, they still exhibit noticeable visual flaws, such as blurriness. RS-NeRF excels by integrating the RS formation process into the model and accurately learning the underlying 3D representation for joint RS correction and novel view synthesis. Qualitative results in real world scenes are shown in Fig. 8. We can observe that baseline methods show obvious artifacts, particularly around object edges. Also, the performance of 2D RS correction combined with NeRF is poorer in real-world scenes than in synthetic ones, which further demonstrates the generalization issues of 2D RS correction methods with real-world data. RS-NeRF, on the other hand, + +
PerturbationON TRAJ.OUT TRAJ.MSTE ↓
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
None27.93.8393.063028.65.8627.0513.0319
U[-0.01, 0.01]27.97.8403.061028.84.8680.0485.0297
U[-0.03, 0.03]28.31.8479.057829.00.8693.0470.0304
U[-0.05, 0.05]28.47.8530.058029.24.8769.0466.0308
U[-0.07, 0.07]28.43.8504.059729.25.8771.0471.0336
+ +Table 6: Quantitative results on synthetic dataset for different noise perturbations on pose initialization. The results demonstrate that our model is robust to the accuracy of pose initialization. + +significantly outperforms these methods with clear edges and detailed rendering, consistent with results in synthetic scenes. Fig. 9 also demonstrates the superiority of RS-NeRF over DRSC [26] + NeRF and USB-NeRF [19] on synthetic and real-world scenes. + +# 4.5 Discussion and Limitations + +Robustness to pose initialization. We evaluated the robustness of our method to the precision of pose initialization on synthetic dataset by introducing random noise of varying scales to the initial poses in the SE(3) space and using these altered poses for initialization. The results, as detailed in Tab. 6, indicate that our method maintains superior performance across different levels of noise perturbation. These results confirm the robustness of RS-NeRF in the accuracy of pose initialization. However, it is important to note that our method might fail when completely random poses are used for initialization. This challenge is also an ongoing issue in the field of NeRF-without-Poses, as discussed in works like [5, 20]. Addressing this limitation is a direction that we plan to pursue in future research. + +Limitations. Similar to previous works that address NeRF with abnormal inputs [24, 36], our model is designed to handle degradation (referred to the RS effects in our context) caused by camera movements, since we primarily model the camera trajectory. When encountering scenarios where RS effects are caused by object movements, i.e., dynamic scenes, our method may fail and produce unsatisfying results. We plan to involve the modeling of object movements based on dynamic NeRFs in future work. + +# 5 Conclusion + +In this paper, we present RS-NeRF, a method designed to synthesize normal images from novel views using input with Rolling Shutter (RS) distortions. We formulate a physical model that replicates the image formation process under RS conditions and jointly optimizes the NeRF parameters and the camera extrinsic for each image row. We then identify and address the fundamental limitations of the standard RS-NeRF model by introducing two innovative approaches: trajectory smoothness regularization and the multi-sampling algorithm. Through rigorous experimentation, we demonstrate that RS-NeRF surpasses previous methods in both synthetic and real-world scenarios, confirming its efficiency in rectifying distortions caused by RS cameras. + +# Acknowledgements + +This research was supported in part by JSPS KAKENHI Grant Numbers 24K223-18, 22H00529, 20H05951, JST-Mirai Program JPMJMI23G1. + +# References + +1. Albl, C., Kukelova, Z., Larsson, V., Pajdla, T.: Rolling shutter camera absolute pose. IEEE transactions on pattern analysis and machine intelligence 42(6), 1439-1452 (2019) +2. Attal, B., Laidlaw, E., Gokaslan, A., Kim, C., Richardt, C., Tompkin, J., O'Toole, M.: Törf: Time-of-flight radiance fields for dynamic scene view synthesis. Advances in neural information processing systems 34, 26289-26301 (2021) +3. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021) +4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mipnerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022) +5. Bian, W., Wang, Z., Li, K., Bian, J.W., Prisacariu, V.A.: Nope-nerf: Optimising neural radiance field with no pose prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4160-4169 (2023) +6. Cao, M., Zhong, Z., Wang, J., Zheng, Y., Yang, Y.: Learning adaptive warping for real-world rolling shutter correction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17785-17793 (2022) +7. Chen, T., Wang, P., Fan, Z., Wang, Z.: Aug-nerf: Training stronger neural radiance fields with triple-level physically-grounded augmentations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15191-15202 (2022) +8. Dave, A., Zhao, Y., Veeraraghavan, A.: Pandora: Polarization-aided neural decomposition of radiance. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part VII. pp. 538–556. Springer (2022) +9. Fan, B., Dai, Y., He, M.: Sunet: symmetric undistortion network for rolling shutter correction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4541-4550 (2021) +0. Fan, B., Dai, Y., Zhang, Z., Liu, Q., He, M.: Context-aware video reconstruction for rolling shutter cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17572-17582 (2022) +1. Fan, B., Mao, Y., Dai, Y., Wan, Z., Liu, Q.: Joint appearance and motion learning for efficient rolling shutter correction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5671-5681 (2023) +2. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: Fastnerf: High-fidelity neural rendering at 200fps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14346-14355 (2021) +3. Guo, Y.C., Kang, D., Bao, L., He, Y., Zhang, S.H.: Nerfren: Neural radiance fields with reflections. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18409-18418 (2022) + +14. Hedborg, J., Forssén, P.E., Felsberg, M., Ringaby, E.: Rolling shutter bundle adjustment. In: 2012 IEEE Conference on Computer Vision and Pattern Recognition. pp. 1434-1441. IEEE (2012) +15. Kim, J.H., Latif, Y., Reid, I.: Rrd-slam: Radial-distorted rolling-shutter direct slam. In: 2017 IEEE International Conference on Robotics and Automation (ICRA). pp. 5148-5154. IEEE (2017) +16. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) +17. Klenk, S., Koestler, L., Scaramuzza, D., Cremers, D.: E-nerf: Neural radiance fields from a moving event camera. IEEE Robotics and Automation Letters (2023) +18. Klingner, B., Martin, D., Roseborough, J.: Street view motion-from-structure-from-motion. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 953–960 (2013) +19. Li, M., Wang, P., Zhao, L., Liao, B., Liu, P.: USB-neRF: Unrolling shutter bundle adjusted neural radiance fields. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=igfDXfMvm5 +20. Lin, C.H., Ma, W.C., Torralba, A., Lucey, S.: Barf: Bundle-adjusting neural radiance fields. In: IEEE International Conference on Computer Vision (ICCV) (2021) +21. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020) +22. Liu, P., Cui, Z., Larsson, V., Pollefeys, M.: Deep shutter unrolling network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5941-5949 (2020) +23. Liu, P., Zuo, X., Larsson, V., Pollefeys, M.: Mba-vo: Motion blur aware visual odometry. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5550-5559 (2021) +24. Ma, L., Li, X., Liao, J., Zhang, Q., Wang, X., Wang, J., Sander, P.V.: Deblur-nerf: Neural radiance fields from blurry images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12861-12870 (2022) +25. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) +26. Qu, D., Liao, B., Zhang, H., Ait-Aider, O., Lao, Y.: Fast rolling shutter correction in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) +27. Reiser, C., Peng, S., Liao, Y., Geiger, A.: Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14335-14345 (2021) +28. Rengarajan, V., Balaji, Y., Rajagopalan, A.: Unrolling the shutter: Cnn to correct motion distortions. In: Proceedings of the IEEE Conference on computer Vision and Pattern Recognition. pp. 2291-2299 (2017) +29. Saurer, O., Koser, K., Bouguet, J.Y., Pollefeys, M.: Rolling shutter stereo. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 465-472 (2013) +30. Saurer, O., Pollefeys, M., Lee, G.H.: Sparse to dense 3d reconstruction from rolling shutter images. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3337-3345 (2016) +31. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4104-4113 (2016) + +32. Schonberger, J.L., Zheng, E., Frahm, J.M., Pollefeys, M.: Pixelwise view selection for unstructured multi-view stereo. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14. pp. 501-518. Springer (2016) +33. Teed, Z., Deng, J.: Raft: Recurrent all-pairs field transforms for optical flow. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 402-419. Springer (2020) +34. Vasu, S., Rajagopalan, A., et al.: Occlusion-aware rolling shutter rectification of 3d scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 636-645 (2018) +35. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-NeRF: Structured view-dependent appearance for neural radiance fields. CVPR (2022) +36. Wang, P., Zhao, L., Ma, R., Liu, P.: Bad-nerf: Bundle adjusted deblur neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4170-4179 (2023) +37. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5438-5448 (2022) +38. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018) +39. Zhong, Z., Cao, M., Sun, X., Wu, Z., Zhou, Z., Zheng, Y., Lin, S., Sato, I.: Bringing rolling shutter images alive with dual reversed distortion. In: European Conference on Computer Vision. pp. 233-249. Springer (2022) +40. Zhong, Z., Zheng, Y., Sato, I.: Towards rolling shutter correction and deblurring in dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9219-9228 (2021) +41. Zhou, X., Duan, P., Ma, Y., Shi, B.: Evunroll: Neuromorphic events based rolling shutter image correction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17775-17784 (2022) +42. Zhuang, B., Tran, Q.H., Ji, P., Cheong, L.F., Chandraker, M.: Learning structure-and-motion-aware rolling shutter correction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4551-4560 (2019) \ No newline at end of file diff --git a/2024/RS-NeRF_ Neural Radiance Fields from Rolling Shutter Images/images.zip b/2024/RS-NeRF_ Neural Radiance Fields from Rolling Shutter Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..43664ecb8ad004d8ab1f0b05734cc9d18235c643 --- /dev/null +++ b/2024/RS-NeRF_ Neural Radiance Fields from Rolling Shutter Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8f42a1c767c19fe7d81c3d053df0be1772b4f2c6ae53e1f0468788630aff56d0 +size 746024 diff --git a/2024/RS-NeRF_ Neural Radiance Fields from Rolling Shutter Images/layout.json b/2024/RS-NeRF_ Neural Radiance Fields from Rolling Shutter Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d1b599920a110dba5e2b58a14caed106666c2ed5 --- /dev/null +++ b/2024/RS-NeRF_ Neural Radiance Fields from Rolling Shutter Images/layout.json @@ -0,0 +1,10405 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "spans": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "text", + "content": "RS-NeRF: Neural Radiance Fields from Rolling Shutter Images" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 196, + 169, + 416, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 169, + 416, + 194 + ], + "spans": [ + { + "bbox": [ + 196, + 169, + 416, + 194 + ], + "type": "text", + "content": "Muyao Niu, Tong Chen, Yifan Zhan, Zhuoxiao Li, Xiang Ji, and Yinqiang Zheng*" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 160, + 203, + 452, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 203, + 452, + 247 + ], + "spans": [ + { + "bbox": [ + 160, + 203, + 452, + 247 + ], + "type": "text", + "content": "The University of Tokyo \nmuyao.niu@gmail.com, cocottt1023@gmail.com, \nzhan-yifan@g.ecc.u-tokyo.ac.jp, lizhuoxiao@g.ecc.u-tokyo.ac.jp, \nxiangji2016@gmail.com, yqzheng@ai.u-tokyo.ac.jp" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 280, + 452, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 280, + 452, + 499 + ], + "spans": [ + { + "bbox": [ + 160, + 280, + 452, + 499 + ], + "type": "text", + "content": "Abstract. Neural Radiance Fields (NeRFs) have become increasingly popular because of their impressive ability for novel view synthesis. However, their effectiveness is hindered by the Rolling Shutter (RS) effects commonly found in most camera systems. To solve this, we present RS-NeRF, a method designed to synthesize normal images from novel views using input with RS distortions. This involves a physical model that replicates the image formation process under RS conditions and jointly optimizes NeRF parameters and camera extrinsic for each image row. We further address the inherent shortcomings of the basic RS-NeRF model by delving into the RS characteristics and developing algorithms to enhance its functionality. First, we impose a smoothness regularization to better estimate trajectories and improve the synthesis quality, in line with the camera movement prior. We also identify and address a fundamental flaw in the vanilla RS model by introducing a multi-sampling algorithm. This new approach improves the model's performance by comprehensively exploiting the RGB data across different rows for each intermediate camera pose. Through rigorous experimentation, we demonstrate that RS-NeRF surpasses previous methods in both synthetic and real-world scenarios, proving its ability to correct RS-related distortions effectively. Codes and data available: https://github.com/MyNiuuu/RS-NeRF" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 522, + 230, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 522, + 230, + 534 + ], + "spans": [ + { + "bbox": [ + 133, + 522, + 230, + 534 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 548, + 482, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 482, + 646 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 482, + 646 + ], + "type": "text", + "content": "CMOS imaging sensors are commonly employed in a wide range of consumer and industrial products. Many of these sensors utilize a Rolling Shutter (RS) mechanism for image capture because of its cost-effectiveness and portability. In contrast to Global Shutter (GS) cameras, which capture all pixels simultaneously, rolling shutter cameras capture image pixels row by row in a sequential manner. Consequently, if the camera is in motion during image capture, visible distortions can occur. It is widely acknowledged that rolling shutter distortions pose challenges for various downstream computer vision tasks, as indicated by" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 653, + 234, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 653, + 234, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 653, + 234, + 665 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "previous research [1,14,15,18,29,30]. Consequently, the correction of rolling shutter effects has attracted significant attention in the past [11, 22, 28, 34, 39-42]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 142, + 481, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 142, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 142, + 481, + 214 + ], + "type": "text", + "content": "In the emerging field of novel view synthesis, technologies such as NeRF [25] have become increasingly proficient in generating highly realistic renderings. NeRF utilizes a continuous volumetric function that is parameterized by a multilayer perceptron (MLP) to associate 3D locations and 2D view directions with color and density information. It relies on volume rendering techniques to enable differentiable rendering." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 216, + 482, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 300 + ], + "type": "text", + "content": "Although traditional NeRF models deliver exceptional results when applied to well-calibrated and well-captured images, they encounter difficulties when dealing with RS effects. For example, in scenarios where the camera is in motion during image capture, these RS effects can significantly impact the performance of the model. Given the widespread use of RS cameras in various industrial applications, it is imperative to take into account this phenomenon when reconstructing 3D representations using NeRFs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 302, + 482, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 482, + 385 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 482, + 385 + ], + "type": "text", + "content": "One straightforward approach involves a two-stage baseline solution, where the first step corrects the rolling shutter effect in 2D image space, and then the NeRF is trained using these corrected images. While these baseline methods enhance the quality of novel view synthesis in NeRF to some extent by utilizing 2D rolling shutter correction techniques, they can not fully explore the 3D scene's geometry, leading to inaccuracies in correspondences, especially when there is significant camera motion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 388, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 496 + ], + "type": "text", + "content": "In this paper, we introduce RS-NeRF, a robust framework designed to explicitly incorporate the physics-based formation of rolling shutter effects into the rendering process. RS-NeRF has the ability to reconstruct a high-quality NeRF representation from only input data that contains rolling shutter artifacts. To achieve this, we explicitly model the camera's motion trajectory throughout the exposure duration for each frame and combine the RGB rendering outcomes from different camera poses to generate the rolling shutter image. Throughout this procedure, we optimize both the NeRF network and the pose estimation network simultaneously by minimizing the photometric loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Building upon the basic RS-NeRF framework, we delve into its shortcomings and enhance its functionality through various refinements. Initially, we impose a camera trajectory smoothness regularization to refine our model's trajectory prediction, by leveraging the sequential nature of camera movement to yield improved results. We further identify a critical shortcoming due to the vanilla RS camera motion model and introduce a multi-sampling technique that significantly enhances RS-NeRF's performance. Specifically, we reveal the issue of the standard camera pose formulation, which assigns distinct poses per row, potentially limiting RS-NeRF's efficiency due to limited training data for each pose. We then make a key observation that multiple poses may share the same 2D RGB observation for the same 3D point during the exposure time, which can drastically increase the available training data. To capitalize on this insight, we introduce and compute the PP-ratio—a measure of camera pose movement in 3D space relative to pixel movement in 2D space, and apply this to broaden our" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "M. Niu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 479, + 198 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 479, + 198 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 479, + 198 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 479, + 198 + ], + "type": "image", + "image_path": "29aee7f762130fd74118c5611f13df3e3cc1fcf90187e89e22bfd3109dce19e1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 207, + 482, + 275 + ], + "lines": [ + { + "bbox": [ + 130, + 207, + 482, + 275 + ], + "spans": [ + { + "bbox": [ + 130, + 207, + 482, + 275 + ], + "type": "text", + "content": "Fig. 1: Pipeline of vanilla RS-NeRF. Given a series of images with RS distortion, RS-NeRF learns the underlying normal 3D representations by jointly estimating pose and RGB values for each row. For each RS image, we first sample the rows and obtain their camera poses via the pose estimation network. We then sample points for each row and feed them with the estimated camera pose to the NeRF network and optimize the photo-metric loss with ground truths." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 298, + 479, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 479, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 479, + 334 + ], + "type": "text", + "content": "training data by sampling out-of-row pixel points. This approach significantly bolsters our model's performance by tapping into the inherent data augmentation possibilities presented by the RS formation process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 335, + 480, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 480, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 480, + 381 + ], + "type": "text", + "content": "We collect both synthetic and real data to evaluate our proposed RS-NeRF. The experimental results reveal that RS-NeRF outperforms the preceding models and demonstrate the effectiveness of our proposed refinements on the vanilla model. In summary, our contributions are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 392, + 479, + 487 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 392, + 479, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 479, + 427 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 479, + 427 + ], + "type": "text", + "content": "- We propose a framework that can reconstruct NeRF from inputs with Rolling Shutter (RS) effects by modeling the camera trajectory and aggregating RGB rendering results of different poses during the exposure time." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 429, + 479, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 429, + 479, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 429, + 479, + 464 + ], + "type": "text", + "content": "- By analyzing the vanilla framework in detail, we proposed a series of improvements to it, including trajectory smoothness regularization and the multi-sampling algorithm." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 464, + 479, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 464, + 479, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 464, + 479, + 487 + ], + "type": "text", + "content": "- We contribute synthetic and real datasets to evaluate our framework and future works." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 507, + 237, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 507, + 237, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 507, + 237, + 520 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "2D Rolling Shutter Correction. Rolling Shutter Correction (RSC) aims to restore normal images given the input with RS distortions. In recent years, several appealing image-based RSC methods [22,28,34,39-42] have been developed. These methods take either single RS image or continuous RS frames as input and predict their global shutter (GS) counterparts. However, single-image RSC can be seriously ill-posed and struggle to correctly estimate the GS counterpart in complex scenarios. To reduce the ill-posedness of single-image RSC, more attention has been paid to multi-image RSC [6, 9-11, 22, 40]. However, these multi-image RS correction methods, which are built on image space operations, fail to exploit the 3D geometry of the scene. As a result, these methods produce multi-view inconsistency results which largely impact the reconstruction results" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "RS-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": "of NeRF. In contrast, our proposed method can jointly reconstruct 3D geometry and correct RS images by effectively aggregating multi-view information of the scene." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 332 + ], + "type": "text", + "content": "Neural Radiance Fields (NeRF). NeRF [25] is an end-to-end model that synthesizes novel views of complex 3D scenes by optimizing an implicit continuous volumetric scene function encoded by an MLP. Due to its simplicity in training and high-quality rendering results, NeRF-based methods [3,4,7,12,21,27,35,37] have attracted intensive attention from various computer vision fields, leading to significant improvements in terms of both accuracy and efficiency. NeRF-based models have also inspired many subsequent works that extend their continuous neural representation to different setups [2,8,13,17,24,36]. Deblur-NeRF [24] and BAD-NeRF [36] reconstruct sharp images given the blurry input from different view directions. One concurrent work [19] appears in ICLR'24, which formulates the RS process with B-cubic models. This work combines the basic modeling of camera motion with the NeRF network to consider the RS effect, without further technical improvements. In this paper, we first bake the RS model into NeRF formulation, then analyze several inherit limitations of the vanilla model, and propose a series of algorithms to further improve its performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 350, + 212, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 350, + 212, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 350, + 212, + 363 + ], + "type": "text", + "content": "3 RS-NeRF" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 376, + 480, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 480, + 437 + ], + "type": "text", + "content": "In this section, we present the specifics of RS-NeRF. Initially, we explain the basic formulation of RS-NeRF in Sec. 3.1. We also note that the vanilla model possesses specific shortcomings that hinder its effectiveness. To address these, we suggest several improvements to enhance its capabilities, which are detailed in Sec. 3.2 and Sec. 3.3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 455, + 285, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 455, + 285, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 455, + 285, + 467 + ], + "type": "text", + "content": "3.1 Vanilla RS-NeRF Model" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 475, + 479, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 479, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 479, + 548 + ], + "type": "text", + "content": "Rolling shutter camera model. Unlike Global Shutter (GS) cameras, the Rolling Shutter (RS) camera captures each row at distinct timestamps. Without loss of generality, we consider the readout direction of the RS camera to be from top to bottom, assuming an infinitesimal exposure time for each row, following existing RS modeling techniques [10, 11, 22]. This mechanism can be described through a mathematical model as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 266, + 558, + 480, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 558, + 480, + 571 + ], + "spans": [ + { + "bbox": [ + 266, + 558, + 480, + 571 + ], + "type": "interline_equation", + "content": "[ \\mathbf {I} ^ {R} (\\mathbf {x}) ] _ {i} = [ \\mathbf {I} _ {i} ^ {G} (\\mathbf {x}) ] _ {i}, \\tag {1}", + "image_path": "7a731011cdf9d60649b9528659394afbaebc6d10808289bf9d2d6f92495bb935.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 581, + 479, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 479, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 479, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 581, + 479, + 629 + ], + "type": "inline_equation", + "content": "[\\mathbf{I}(\\mathbf{x})]_i" + }, + { + "bbox": [ + 130, + 581, + 479, + 629 + ], + "type": "text", + "content": " is an operator to extract the " + }, + { + "bbox": [ + 130, + 581, + 479, + 629 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 581, + 479, + 629 + ], + "type": "text", + "content": "-th row from an image " + }, + { + "bbox": [ + 130, + 581, + 479, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{I}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 581, + 479, + 629 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 581, + 479, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^R(\\mathbf{x}) \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 581, + 479, + 629 + ], + "type": "text", + "content": " is the rolling shutter image, and " + }, + { + "bbox": [ + 130, + 581, + 479, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i^G(\\mathbf{x}) \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 581, + 479, + 629 + ], + "type": "text", + "content": " is the virtual global shutter image captured at the same pose as the " + }, + { + "bbox": [ + 130, + 581, + 479, + 629 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 581, + 479, + 629 + ], + "type": "text", + "content": "-th row of " + }, + { + "bbox": [ + 130, + 581, + 479, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^R(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 581, + 479, + 629 + ], + "type": "text", + "content": ". As a result, for one RS image " + }, + { + "bbox": [ + 130, + 581, + 479, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^R(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 581, + 479, + 629 + ], + "type": "text", + "content": ", we denote the pose of the " + }, + { + "bbox": [ + 130, + 581, + 479, + 629 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 581, + 479, + 629 + ], + "type": "text", + "content": "-th row as " + }, + { + "bbox": [ + 130, + 581, + 479, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_i" + }, + { + "bbox": [ + 130, + 581, + 479, + 629 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 630, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 480, + 665 + ], + "type": "text", + "content": "Camera motion trajectory modeling. In our rolling shutter camera model, it is necessary to establish the poses for each row in each image. Given the typically short exposure time of an RS image, we use a linear model to approximate the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "M. Niu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 117, + 302, + 190 + ], + "blocks": [ + { + "bbox": [ + 137, + 117, + 302, + 190 + ], + "lines": [ + { + "bbox": [ + 137, + 117, + 302, + 190 + ], + "spans": [ + { + "bbox": [ + 137, + 117, + 302, + 190 + ], + "type": "image", + "image_path": "eabc1255cf41091543613c6aa07177d3d28ab0faca1e30bcaecff64dd6293235.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 114, + 482, + 207 + ], + "lines": [ + { + "bbox": [ + 318, + 114, + 482, + 207 + ], + "spans": [ + { + "bbox": [ + 318, + 114, + 482, + 207 + ], + "type": "text", + "content": "Fig. 2: Smoothness regularization on camera trajectory. For each pair of adjacent trajectory vectors " + }, + { + "bbox": [ + 318, + 114, + 482, + 207 + ], + "type": "inline_equation", + "content": "\\vec{\\mathbf{d}}_k" + }, + { + "bbox": [ + 318, + 114, + 482, + 207 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 318, + 114, + 482, + 207 + ], + "type": "inline_equation", + "content": "\\vec{\\mathbf{d}}_{k + 1}" + }, + { + "bbox": [ + 318, + 114, + 482, + 207 + ], + "type": "text", + "content": ", we compute the mid-point trajectory vector " + }, + { + "bbox": [ + 318, + 114, + 482, + 207 + ], + "type": "inline_equation", + "content": "\\vec{\\mathbf{d}}_{mid}" + }, + { + "bbox": [ + 318, + 114, + 482, + 207 + ], + "type": "text", + "content": " and their unit vectors " + }, + { + "bbox": [ + 318, + 114, + 482, + 207 + ], + "type": "inline_equation", + "content": "\\vec{\\mathbf{n}}_k" + }, + { + "bbox": [ + 318, + 114, + 482, + 207 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 318, + 114, + 482, + 207 + ], + "type": "inline_equation", + "content": "\\vec{\\mathbf{n}}_{k + 1}" + }, + { + "bbox": [ + 318, + 114, + 482, + 207 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 318, + 114, + 482, + 207 + ], + "type": "inline_equation", + "content": "\\vec{\\mathbf{n}}_{mid}" + }, + { + "bbox": [ + 318, + 114, + 482, + 207 + ], + "type": "text", + "content": ". We then apply the " + }, + { + "bbox": [ + 318, + 114, + 482, + 207 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 318, + 114, + 482, + 207 + ], + "type": "text", + "content": " regularization between " + }, + { + "bbox": [ + 318, + 114, + 482, + 207 + ], + "type": "inline_equation", + "content": "\\vec{\\mathbf{n}}_{mid}" + }, + { + "bbox": [ + 318, + 114, + 482, + 207 + ], + "type": "text", + "content": " and mean " + }, + { + "bbox": [ + 318, + 114, + 482, + 207 + ], + "type": "inline_equation", + "content": "(\\vec{\\mathbf{n}}_k,\\vec{\\mathbf{n}}_{k + 1})" + }, + { + "bbox": [ + 318, + 114, + 482, + 207 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 230, + 479, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 230, + 479, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 230, + 479, + 278 + ], + "type": "text", + "content": "motion of the camera. Specifically, we start by defining two camera poses: one for the first row, denoted as " + }, + { + "bbox": [ + 130, + 230, + 479, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_0" + }, + { + "bbox": [ + 130, + 230, + 479, + 278 + ], + "type": "text", + "content": ", and one for the last row, " + }, + { + "bbox": [ + 130, + 230, + 479, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{H - 1}" + }, + { + "bbox": [ + 130, + 230, + 479, + 278 + ], + "type": "text", + "content": ". For the rows in between, we linearly interpolate their poses within the Lie algebra of " + }, + { + "bbox": [ + 130, + 230, + 479, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{SE}(3)" + }, + { + "bbox": [ + 130, + 230, + 479, + 278 + ], + "type": "text", + "content": " as informed by prior research [20, 23, 36]:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 242, + 287, + 481, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 287, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 242, + 287, + 481, + 300 + ], + "type": "interline_equation", + "content": "\\mathbf {T} _ {i} = \\mathbf {T} _ {0} \\cdot (1 - \\rho) + \\mathbf {T} _ {H - 1} \\cdot \\rho , \\tag {2}", + "image_path": "881fcbb73e31d915109a2b0511a5317e82e5fc8b3a5168aef502469d6bb90b46.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 309, + 479, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 479, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 479, + 380 + ], + "type": "text", + "content": "where the interpolation coefficient " + }, + { + "bbox": [ + 130, + 309, + 479, + 380 + ], + "type": "inline_equation", + "content": "\\rho = \\frac{i}{H - 1}" + }, + { + "bbox": [ + 130, + 309, + 479, + 380 + ], + "type": "text", + "content": " is based on the row index " + }, + { + "bbox": [ + 130, + 309, + 479, + 380 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 309, + 479, + 380 + ], + "type": "text", + "content": ". Thus, for each RS image, RS-NeRF receives one pose initialization, then optimizes the poses at both ends and then determines the poses for each intermediate row through linear interpolation. This interpolation method for calculating each row's pose is fully differentiable, enabling the optimization of pose parameters using a series of rolling shutter images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 381, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 381, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 482, + 477 + ], + "type": "text", + "content": "Optimizing NeRF from RS observation. For a set of " + }, + { + "bbox": [ + 130, + 381, + 482, + 477 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 381, + 482, + 477 + ], + "type": "text", + "content": " rolling shutter images, we select " + }, + { + "bbox": [ + 130, + 381, + 482, + 477 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 381, + 482, + 477 + ], + "type": "text", + "content": " rows per image and " + }, + { + "bbox": [ + 130, + 381, + 482, + 477 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 381, + 482, + 477 + ], + "type": "text", + "content": " points per row, resulting in " + }, + { + "bbox": [ + 130, + 381, + 482, + 477 + ], + "type": "inline_equation", + "content": "K \\times M \\times N" + }, + { + "bbox": [ + 130, + 381, + 482, + 477 + ], + "type": "text", + "content": " distinct camera rays for each iteration. These rays are used to concurrently optimize the NeRF network and the pose parameters. Specifically, for each image, we initially use the row indices to access the pose estimation network. This allows us to determine the camera ray " + }, + { + "bbox": [ + 130, + 381, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{r}}(\\theta_p; u, v)" + }, + { + "bbox": [ + 130, + 381, + 482, + 477 + ], + "type": "text", + "content": " for each sampled pixel point " + }, + { + "bbox": [ + 130, + 381, + 482, + 477 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 130, + 381, + 482, + 477 + ], + "type": "text", + "content": ". Following this, these rays are employed to train the NeRF network by computing the ensuing photometric loss:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 222, + 486, + 480, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 486, + 480, + 499 + ], + "spans": [ + { + "bbox": [ + 222, + 486, + 480, + 499 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\left| \\left| \\mathbf {C} (\\mathbf {r} (u, v)) - \\hat {\\mathbf {C}} \\left(\\theta_ {n}; \\hat {\\mathbf {r}} \\left(\\theta_ {p}; (u, v)\\right) \\right. \\right| \\right| _ {2}, \\tag {3}", + "image_path": "b3bd74991622e75c48ba7c394c41568b35dffee75158a0f669a7d892e0ebff03.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 507, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 482, + 568 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 507, + 482, + 568 + ], + "type": "inline_equation", + "content": "\\theta_{n}" + }, + { + "bbox": [ + 130, + 507, + 482, + 568 + ], + "type": "text", + "content": " represents the trainable parameters of the NeRF network, and " + }, + { + "bbox": [ + 130, + 507, + 482, + 568 + ], + "type": "inline_equation", + "content": "\\theta_{p}" + }, + { + "bbox": [ + 130, + 507, + 482, + 568 + ], + "type": "text", + "content": " denotes the learnable parameters for camera poses. " + }, + { + "bbox": [ + 130, + 507, + 482, + 568 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{C}} (\\theta_n;\\hat{\\mathbf{r}} (\\theta_p;(u,v))" + }, + { + "bbox": [ + 130, + 507, + 482, + 568 + ], + "type": "text", + "content": " refers to the predicted RGB value for the pixel at coordinates " + }, + { + "bbox": [ + 130, + 507, + 482, + 568 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 130, + 507, + 482, + 568 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 130, + 507, + 482, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{C}(\\mathbf{r}(u,v))" + }, + { + "bbox": [ + 130, + 507, + 482, + 568 + ], + "type": "text", + "content": " denotes the corresponding RGB ground truth value. The workflow of our vanilla RS-NeRF model is depicted in Fig. 1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 586, + 356, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 356, + 598 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 356, + 598 + ], + "type": "text", + "content": "3.2 Trajectory Smoothness Regularization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Given the previous description of our model, it is evident that accurately estimating the camera pose for each row in RS frames is crucial for effectively reconstructing the normal NeRF representation. We observe that the basic RS-NeRF struggles with precise camera trajectory estimation, often resulting in high-frequency predictions, as further analyzed in our experiments. This issue" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "RS-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 120, + 280, + 169 + ], + "blocks": [ + { + "bbox": [ + 133, + 120, + 280, + 169 + ], + "lines": [ + { + "bbox": [ + 133, + 120, + 280, + 169 + ], + "spans": [ + { + "bbox": [ + 133, + 120, + 280, + 169 + ], + "type": "table", + "html": "
NposeON TRAJ. / OUT TRAJ.
PSNR ↑SSIM ↑LPIPS ↓
10024.36/25.16.7321/.7727.0703/.0578
40024.06/24.66.7245/.7536.0697/.0579
", + "image_path": "a76f235469d71677fee3efc7000efff60cc89cdf6fe437688ed3e509e7c5a952.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 173, + 280, + 208 + ], + "lines": [ + { + "bbox": [ + 132, + 173, + 280, + 208 + ], + "spans": [ + { + "bbox": [ + 132, + 173, + 280, + 208 + ], + "type": "text", + "content": "Table 1: Quantitative results for different " + }, + { + "bbox": [ + 132, + 173, + 280, + 208 + ], + "type": "inline_equation", + "content": "N_{\\text{pose}}" + }, + { + "bbox": [ + 132, + 173, + 280, + 208 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 132, + 173, + 280, + 208 + ], + "type": "inline_equation", + "content": "N_{\\text{pose}} = 100" + }, + { + "bbox": [ + 132, + 173, + 280, + 208 + ], + "type": "text", + "content": " yields better results than " + }, + { + "bbox": [ + 132, + 173, + 280, + 208 + ], + "type": "inline_equation", + "content": "N_{\\text{pose}} = 400" + }, + { + "bbox": [ + 132, + 173, + 280, + 208 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 282, + 114, + 480, + 180 + ], + "blocks": [ + { + "bbox": [ + 282, + 114, + 480, + 180 + ], + "lines": [ + { + "bbox": [ + 282, + 114, + 480, + 180 + ], + "spans": [ + { + "bbox": [ + 282, + 114, + 480, + 180 + ], + "type": "image", + "image_path": "a958c75ed494c3d7fceb1b81c4939f267a7da9f8a257120e2a3579d22b9f3639.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 280, + 184, + 482, + 208 + ], + "lines": [ + { + "bbox": [ + 280, + 184, + 482, + 208 + ], + "spans": [ + { + "bbox": [ + 280, + 184, + 482, + 208 + ], + "type": "text", + "content": "Fig. 3: Qualitative comparisons for different " + }, + { + "bbox": [ + 280, + 184, + 482, + 208 + ], + "type": "inline_equation", + "content": "N_{\\text{pose}}" + }, + { + "bbox": [ + 280, + 184, + 482, + 208 + ], + "type": "text", + "content": ". The artifact decreases as " + }, + { + "bbox": [ + 280, + 184, + 482, + 208 + ], + "type": "inline_equation", + "content": "N_{\\text{pose}}" + }, + { + "bbox": [ + 280, + 184, + 482, + 208 + ], + "type": "text", + "content": " grows." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 239, + 480, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 480, + 312 + ], + "type": "text", + "content": "arises because the network focuses only on the photometric loss for individual images, neglecting the consistent movement characteristics of the camera across different RS frames. To address this, we introduce a camera trajectory smoothness regularization, based on the expectation of a smooth and continuous camera trajectory in 3D space. Specifically, for every two adjacent RS frames " + }, + { + "bbox": [ + 130, + 239, + 480, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_k^R" + }, + { + "bbox": [ + 130, + 239, + 480, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 239, + 480, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{k + 1}^R" + }, + { + "bbox": [ + 130, + 239, + 480, + 312 + ], + "type": "text", + "content": ", we aim to minimize the following loss during the optimization process:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 223, + 319, + 481, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 319, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 223, + 319, + 481, + 335 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} = | | \\vec {\\mathbf {n}} _ {m i d} - \\operatorname {m e a n} (\\vec {\\mathbf {n}} _ {k} + \\vec {\\mathbf {n}} _ {k + 1}) | | _ {2}. \\tag {4}", + "image_path": "a4d3319c493d744fff826b6438c69bc40c2b92e73bb08a94b881cba9141addd7.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 341, + 482, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 390 + ], + "type": "text", + "content": "As illustrated in Fig. 2, " + }, + { + "bbox": [ + 130, + 341, + 482, + 390 + ], + "type": "inline_equation", + "content": "\\vec{\\mathbf{n}}_k" + }, + { + "bbox": [ + 130, + 341, + 482, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 341, + 482, + 390 + ], + "type": "inline_equation", + "content": "\\vec{\\mathbf{n}}_{k + 1}" + }, + { + "bbox": [ + 130, + 341, + 482, + 390 + ], + "type": "text", + "content": " represent the unit direction vectors of the estimated camera trajectories for RS frames " + }, + { + "bbox": [ + 130, + 341, + 482, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_k^R" + }, + { + "bbox": [ + 130, + 341, + 482, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 341, + 482, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{k + 1}^R" + }, + { + "bbox": [ + 130, + 341, + 482, + 390 + ], + "type": "text", + "content": ", respectively. Meanwhile, " + }, + { + "bbox": [ + 130, + 341, + 482, + 390 + ], + "type": "inline_equation", + "content": "\\vec{\\mathbf{n}}_{mid}" + }, + { + "bbox": [ + 130, + 341, + 482, + 390 + ], + "type": "text", + "content": " denotes the unit direction vector for the camera trajectory during the exposure interval between these two RS frames." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 407, + 294, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 294, + 419 + ], + "type": "text", + "content": "3.3 Multi-Sampling Algorithm" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "text", + "content": "As discussed in Sec. 3.1, each row in an RS frame is assigned a unique camera pose. For example, in an RS frame with 400 rows, there are 400 distinct camera poses. This rigorous approach aligns with the real-world, physics-based photography process of RS cameras but also presents certain challenges. For example, under this framework, the data points for each pose are confined to a single row. This limitation can significantly affect our model's performance due to the sparse training data available for each pose." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 510, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 640 + ], + "type": "text", + "content": "A straightforward solution to this problem is to reduce the number of poses per image. For example, assigning 200 poses to each RS frame means that all two adjacent rows would share the same pose. This approach can lead to blurriness since these rows are captured at different poses, but it also doubles the amount of data available for training each pose. To further explore this trade-off, we conducted a series of experiments. Fig. 3 shows qualitative results for different intermediate pose numbers " + }, + { + "bbox": [ + 130, + 510, + 482, + 640 + ], + "type": "inline_equation", + "content": "N_{\\text{pose}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 640 + ], + "type": "text", + "content": ", indicating that the artifacts decrease as " + }, + { + "bbox": [ + 130, + 510, + 482, + 640 + ], + "type": "inline_equation", + "content": "N_{\\text{pose}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 640 + ], + "type": "text", + "content": " increases. Interestingly, we observed that using 100 poses per frame often yields results comparable to or even better than using 400 poses. The quantitative results presented in Tab. 1 also support the superiority of having 100 poses per frame." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "This observation seems counter-intuitive at first, as one might assume that having fewer than 400 poses per frame would lead to poorer results due to" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "M. Niu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 168, + 115, + 445, + 242 + ], + "blocks": [ + { + "bbox": [ + 168, + 115, + 445, + 242 + ], + "lines": [ + { + "bbox": [ + 168, + 115, + 445, + 242 + ], + "spans": [ + { + "bbox": [ + 168, + 115, + 445, + 242 + ], + "type": "image", + "image_path": "3518bdd366c104f04ccdaff54f8406ce473f531de5996e9e5498d56150131e06.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 253, + 482, + 330 + ], + "lines": [ + { + "bbox": [ + 130, + 253, + 482, + 330 + ], + "spans": [ + { + "bbox": [ + 130, + 253, + 482, + 330 + ], + "type": "text", + "content": "Fig. 4: Motivation for the multi-sampling algorithm. Up: Three consecutive RS frames, labeled as (a), (b), and (c), along with the optical flow between each pair of frames, shown in (d) and (e). Down: For a single point in 3D space, we observe its 2D projections in the first (marked in yellow) and second (marked in red) RS frames. These projections exhibit a row displacement of 8 and a column displacement of 18, as depicted in (f). However, during this interval, the camera undergoes a shift of at least 392 poses, as indicated in (g)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 358, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 453 + ], + "type": "text", + "content": "artifacts like blurriness. However, it can be rationally explained. As demonstrated in Fig. 4, for a given point in 3D space, the 2D projections on the first and second RS frames show relatively small pixel displacements (8 in row and 12 in column). Meanwhile, the camera shifts through at least 392 poses in 3D space. This is based on the assumption that there is no overlap in the trajectories of two adjacent frames (indicated as " + }, + { + "bbox": [ + 130, + 358, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\geq 0" + }, + { + "bbox": [ + 130, + 358, + 482, + 453 + ], + "type": "text", + "content": " poses' in the figure), which is naturally satisfied given the sparse observation setting and short exposure times typically involved." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 457, + 482, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 482, + 553 + ], + "type": "text", + "content": "Following our prior analysis, we introduced a multi-sampling algorithm aimed at explicitly increasing the sample points for each intermediate camera pose. The fundamental concept involves computing the ratio of camera pose displacement to pixel displacement (termed the PP-ratio) between two adjacent frames for the same 3D point. For example, a pixel displacement of 18 combined with a camera pose shift of 392 results in a PP-ratio of 21. This implies that at least 21 consecutive camera poses share the same pixel value, which means that all these 21 camera poses can utilize this RGB value for training purposes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "text", + "content": "For a pixel " + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = (u,v)" + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "text", + "content": " located in the " + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "text", + "content": "-th column and " + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "text", + "content": "-th row of the " + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "text", + "content": "-th RS frame " + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_k^R (\\mathbf{x})" + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "text", + "content": ", we determine both the forward PP-ratio " + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\rho_{k\\to k + 1}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "text", + "content": " and the backward PP-ratio " + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\rho_{k\\to k - 1}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "text", + "content": ". This is because the point represented by " + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "text", + "content": " can transition from " + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_k^R (\\mathbf{x})" + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{k + 1}^R (\\mathbf{x})" + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{k - 1}^R (\\mathbf{x})" + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "text", + "content": ". Consequently, all camera poses within the range of " + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "inline_equation", + "content": "[v - \\rho_{k\\to k - 1}(\\mathbf{x}), v + \\rho_{k\\to k + 1}(\\mathbf{x})]" + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "text", + "content": " are capable of sampling the RGB value of " + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "text", + "content": " for training purposes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "To compute the forward PP-ratio " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\rho_{k\\rightarrow k + 1}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " for a pixel " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "-th RS frame " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_k^R (\\mathbf{x})" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": ", the first step is to estimate the forward optical flow " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "F_{k\\rightarrow k + 1}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": ", utilizing RAFT [33]. Consequently, the row displacement " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\varDelta_{k\\rightarrow k + 1}^{row}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " and column" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "RS-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 113, + 343, + 260 + ], + "blocks": [ + { + "bbox": [ + 133, + 113, + 343, + 260 + ], + "lines": [ + { + "bbox": [ + 133, + 113, + 343, + 260 + ], + "spans": [ + { + "bbox": [ + 133, + 113, + 343, + 260 + ], + "type": "image", + "image_path": "edddefe08157b725531a95a409a992c1b2e71d0336d6ad8070f0ff4b3f55ad39.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 114, + 482, + 247 + ], + "lines": [ + { + "bbox": [ + 349, + 114, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 349, + 114, + 482, + 247 + ], + "type": "text", + "content": "Fig. 5: Comparing the accessible training points in one RS image under different strategies. The image size is " + }, + { + "bbox": [ + 349, + 114, + 482, + 247 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 349, + 114, + 482, + 247 + ], + "type": "text", + "content": ", providing " + }, + { + "bbox": [ + 349, + 114, + 482, + 247 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 349, + 114, + 482, + 247 + ], + "type": "text", + "content": " poses. Distinct circle hues represent different RGB values. In case " + }, + { + "bbox": [ + 349, + 114, + 482, + 247 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 349, + 114, + 482, + 247 + ], + "type": "text", + "content": ", pixels transported from other poses are with dash contours, and the number signifying that it originates from the " + }, + { + "bbox": [ + 349, + 114, + 482, + 247 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 349, + 114, + 482, + 247 + ], + "type": "text", + "content": "-th pose." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 281, + 479, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 479, + 306 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 479, + 306 + ], + "type": "text", + "content": "displacement " + }, + { + "bbox": [ + 130, + 281, + 479, + 306 + ], + "type": "inline_equation", + "content": "\\varDelta_{k\\rightarrow k+1}^{col}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 281, + 479, + 306 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 281, + 479, + 306 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 281, + 479, + 306 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 281, + 479, + 306 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_k^R (\\mathbf{x})" + }, + { + "bbox": [ + 130, + 281, + 479, + 306 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 281, + 479, + 306 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{k + 1}^R (\\mathbf{x})" + }, + { + "bbox": [ + 130, + 281, + 479, + 306 + ], + "type": "text", + "content": " would be the second and the first element of " + }, + { + "bbox": [ + 130, + 281, + 479, + 306 + ], + "type": "inline_equation", + "content": "F_{k\\to k + 1}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 281, + 479, + 306 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 220, + 314, + 481, + 329 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 314, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 220, + 314, + 481, + 329 + ], + "type": "interline_equation", + "content": "F _ {k \\rightarrow k + 1} (\\mathbf {x}) = \\operatorname {R A F T} \\left(\\mathbf {I} _ {k} ^ {R} (\\mathbf {x}), \\mathbf {I} _ {k + 1} ^ {R} (\\mathbf {x})\\right), \\tag {5}", + "image_path": "3709f450f05db7a8600e88494890aefd1dff10dd89e53c15dfcbc9c9a05e1a34.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 221, + 331, + 481, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 331, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 221, + 331, + 481, + 346 + ], + "type": "interline_equation", + "content": "\\varDelta_ {k \\rightarrow k + 1} ^ {c o l} (\\mathbf {x}), \\varDelta_ {k \\rightarrow k + 1} ^ {r o w} (\\mathbf {x}) = F _ {k \\rightarrow k + 1} (\\mathbf {x}), \\tag {6}", + "image_path": "9db1d4f078d7bdad300d0f7da7e9eb5cddc5756ccaedc9b7b657ae321f9c80ff.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 354, + 389, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 389, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 389, + 365 + ], + "type": "text", + "content": "and the calculation are similar for the backward situation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 146, + 367, + 440, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 367, + 440, + 379 + ], + "spans": [ + { + "bbox": [ + 146, + 367, + 440, + 379 + ], + "type": "text", + "content": "We then calculate the forward and backward camera pose shift as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 211, + 388, + 481, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 388, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 211, + 388, + 481, + 418 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\Delta_ {k \\rightarrow k + 1} ^ {\\text {p o s e}} (\\mathbf {x}) = v + \\Delta_ {k \\rightarrow k + 1} ^ {\\text {r o w}} (\\mathbf {x}) + (H - v), (7) \\\\ = H + \\Delta_ {k \\rightarrow k + 1} ^ {\\text {r o w}} (\\mathbf {x}), (8) \\\\ \\end{array}", + "image_path": "f6324a16789324345d208ddbed694144151fa28567333bbc549a1f9fb531a5a3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 211, + 420, + 481, + 450 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 420, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 211, + 420, + 481, + 450 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\Delta_ {k \\rightarrow k - 1} ^ {\\text {p o s e}} (\\mathbf {x}) = \\left(H - \\left(v + \\Delta_ {k \\rightarrow k - 1} ^ {\\text {r o w}} (\\mathbf {x})\\right)\\right) + v, (9) \\\\ = H - \\Delta_ {k \\rightarrow k - 1} ^ {\\text {r o w}} (\\mathbf {x}). (10) \\\\ \\end{array}", + "image_path": "5cf6dd6b3f64766203361b47849aeee3e0cc3280c8232c260d64115e77d68962.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 146, + 460, + 394, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 460, + 394, + 471 + ], + "spans": [ + { + "bbox": [ + 146, + 460, + 394, + 471 + ], + "type": "text", + "content": "Finally, we calculate forward and backward PP-ratio as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 206, + 479, + 481, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 479, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 206, + 479, + 481, + 509 + ], + "type": "interline_equation", + "content": "\\rho_ {k \\rightarrow k + 1} (\\mathbf {x}) = \\frac {\\Delta_ {k \\rightarrow k + 1} ^ {p o s e} (\\mathbf {x})}{\\max \\left(\\Delta_ {k \\rightarrow k + 1} ^ {c o l} (\\mathbf {x}) , \\Delta_ {k \\rightarrow k + 1} ^ {r o w} (\\mathbf {x})\\right)}, \\tag {11}", + "image_path": "ee29c4d7a4ffa604774b3d918ab59a76a30c4ebe774ab3eda118598a4ca0a204.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 207, + 510, + 481, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 510, + 481, + 540 + ], + "spans": [ + { + "bbox": [ + 207, + 510, + 481, + 540 + ], + "type": "interline_equation", + "content": "\\rho_ {k \\rightarrow k - 1} (\\mathbf {x}) = \\frac {\\Delta_ {k \\rightarrow k - 1} ^ {p o s e} (\\mathbf {x})}{\\operatorname* {m a x} \\left(\\Delta_ {k \\rightarrow k - 1} ^ {c o l} (\\mathbf {x}) , \\Delta_ {k \\rightarrow k - 1} ^ {r o w} (\\mathbf {x})\\right)}. \\tag {12}", + "image_path": "78bc5ec7d403633bc03dadf4b9266831109e199a4cff8aa38a6c4236ca90eb85.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 548, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 548, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 548, + 482, + 586 + ], + "type": "text", + "content": "Therefore, when dealing with a sampled pixel " + }, + { + "bbox": [ + 131, + 548, + 482, + 586 + ], + "type": "inline_equation", + "content": "RGB(\\mathbf{x})" + }, + { + "bbox": [ + 131, + 548, + 482, + 586 + ], + "type": "text", + "content": " in the RS frame " + }, + { + "bbox": [ + 131, + 548, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_k^R (\\mathbf{x})" + }, + { + "bbox": [ + 131, + 548, + 482, + 586 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 548, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = (u,v)" + }, + { + "bbox": [ + 131, + 548, + 482, + 586 + ], + "type": "text", + "content": ", rather than directly using the " + }, + { + "bbox": [ + 131, + 548, + 482, + 586 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 131, + 548, + 482, + 586 + ], + "type": "text", + "content": "-th pose " + }, + { + "bbox": [ + 131, + 548, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_v" + }, + { + "bbox": [ + 131, + 548, + 482, + 586 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 131, + 548, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_k^R (\\mathbf{x})" + }, + { + "bbox": [ + 131, + 548, + 482, + 586 + ], + "type": "text", + "content": " for training, we randomly sample " + }, + { + "bbox": [ + 131, + 548, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\hat{v}}" + }, + { + "bbox": [ + 131, + 548, + 482, + 586 + ], + "type": "text", + "content": " where" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 226, + 595, + 481, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 595, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 226, + 595, + 481, + 609 + ], + "type": "interline_equation", + "content": "\\hat {v} \\in [ v - \\rho_ {k \\rightarrow k - 1} (\\mathbf {x}), v + \\rho_ {k \\rightarrow k + 1} (\\mathbf {x}) ]. \\tag {13}", + "image_path": "b9d093c86bcb34cec89f294158c7d8e4204234f686ad4cebdce211f18ac1317f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "content": "Why does Multi-Sampling work? Let us consider a basic task: optimizing NeRF with " + }, + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "content": " known poses and images (non-RS). Case " + }, + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "content": ": Only one row is available per pose. Case " + }, + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "content": ": all rows are accessible. Clearly, case " + }, + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "content": " would yield a better NeRF due to more data per pose. Moving to our RS setting as depicted" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "M. Niu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 114, + 481, + 182 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 182 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 182 + ], + "type": "table", + "html": "
LssMSTORIIWINEPOOLFACTORYTANABATAAVERAGE
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
ON TrBAI✓23.45.7535.044823.53.7019.068122.99.6111.078419.53.5018.107217.86.4921.103021.47.6121.0803
✓22.25.6729.056425.60.7749.071227.22.7705.058221.81.6128.094923.42.7916.067924.06.7245.0697
✓26.65.8572.034225.55.7637.068225.86.7303.055525.45.7754.070820.94.6802.083224.89.7614.0624
✓30.97.9058.039528.36.8482.067528.50.8137.056225.43.7673.086726.40.8614.065127.93.8393.0630
OUT TrBAI✓23.26.7561.042824.05.7169.060123.27.6218.073819.98.5426.089217.85.5065.094221.68.6288.0720
✓22.39.6875.050126.62.8113.057927.29.7783.048922.72.6624.081524.27.8286.051424.66.7536.0579
✓26.58.8611.030825.96.7750.061225.49.7080.054924.38.7446.074220.73.6850.075124.63.7547.0592
✓33.07.9499.024329.27.8703.056028.77.8313.045825.11.7738.080727.02.8881.049828.65.8627.0513
", + "image_path": "134dd0d4968a99f52c4f966ed8f47350b9727fc4f1c1ca4632e54224a0b3877b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 187, + 480, + 220 + ], + "lines": [ + { + "bbox": [ + 132, + 187, + 480, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 187, + 480, + 220 + ], + "type": "text", + "content": "Table 2: Quantitative results on synthetic scenes for ablation study. " + }, + { + "bbox": [ + 132, + 187, + 480, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 132, + 187, + 480, + 220 + ], + "type": "text", + "content": " denotes the camera trajectory smoothness regularization, and 'MS' denotes the multi-sampling algorithm. We color code each result as best and second best." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 254, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 254, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 482, + 456 + ], + "type": "text", + "content": "in Fig.5, case " + }, + { + "bbox": [ + 130, + 254, + 482, + 456 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 254, + 482, + 456 + ], + "type": "text", + "content": " is the vanilla RS observation for one RS frame (" + }, + { + "bbox": [ + 130, + 254, + 482, + 456 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 254, + 482, + 456 + ], + "type": "text", + "content": " rows lead to " + }, + { + "bbox": [ + 130, + 254, + 482, + 456 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 254, + 482, + 456 + ], + "type": "text", + "content": " poses), and case " + }, + { + "bbox": [ + 130, + 254, + 482, + 456 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 254, + 482, + 456 + ], + "type": "text", + "content": " is the virtual GS observation, where all rows are available for each pose. Thus, case " + }, + { + "bbox": [ + 130, + 254, + 482, + 456 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 254, + 482, + 456 + ], + "type": "text", + "content": " is the upper bound for case " + }, + { + "bbox": [ + 130, + 254, + 482, + 456 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 254, + 482, + 456 + ], + "type": "text", + "content": ". Essentially, MS aims to transport all possible pixel values from nearby poses to fill the missing observation that can be potentially recovered (case " + }, + { + "bbox": [ + 130, + 254, + 482, + 456 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 254, + 482, + 456 + ], + "type": "text", + "content": " in Fig.5), i.e., partially restore the virtual GS observation from the vanilla RS observation, by allocating additional poses for each pixel, which ensures better results than vanilla RS. Also, interpolating two poses does not harm MS. We optimize two poses for each frame, and each in-between pose becomes a fixed blend ratio of two poses. The best case is virtual GS (Fig.5(B)), where every pixel coordinate can be sampled in all ratios " + }, + { + "bbox": [ + 130, + 254, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\left(\\frac{0}{H}, \\frac{1}{H-1}, \\dots, \\frac{H}{0}\\right)" + }, + { + "bbox": [ + 130, + 254, + 482, + 456 + ], + "type": "text", + "content": ", resulting in a total of " + }, + { + "bbox": [ + 130, + 254, + 482, + 456 + ], + "type": "inline_equation", + "content": "H \\times W \\times H" + }, + { + "bbox": [ + 130, + 254, + 482, + 456 + ], + "type": "text", + "content": " data points to optimize the two poses. In vanilla RS (Fig.5(A)), each coordinate only contributes to one ratio determined by row indices, producing only " + }, + { + "bbox": [ + 130, + 254, + 482, + 456 + ], + "type": "inline_equation", + "content": "1 \\times W \\times H" + }, + { + "bbox": [ + 130, + 254, + 482, + 456 + ], + "type": "text", + "content": " data points. With MS (Fig.5(C)), some pixels are replicated to other poses, allowing multiple ratios for them. Thus, the number of data points is greater than " + }, + { + "bbox": [ + 130, + 254, + 482, + 456 + ], + "type": "inline_equation", + "content": "1 \\times W \\times H" + }, + { + "bbox": [ + 130, + 254, + 482, + 456 + ], + "type": "text", + "content": " but less than " + }, + { + "bbox": [ + 130, + 254, + 482, + 456 + ], + "type": "inline_equation", + "content": "H \\times W \\times H" + }, + { + "bbox": [ + 130, + 254, + 482, + 456 + ], + "type": "text", + "content": ". This ensures better pose estimation and view synthesis than vanilla RS." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 476, + 230, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 476, + 230, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 230, + 490 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 502, + 202, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 202, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 202, + 514 + ], + "type": "text", + "content": "4.1 Settings" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Datasets. Evaluating RS-aware view synthesis technologies requires synthetic datasets that have multi-view images with RS distortions and the corresponding GS ground truth. We synthesize 5 scenes via blender to evaluate our method. For each scene, we rendered images from various camera poses while moving the camera gradually around the scene in a forward-facing manner. We then combined rows from different poses to produce the final RS images. Each scene comprises 34 images with a resolution of " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "400 \\times 400" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ". We employed the pose of the middle row of each RS image as the initial camera pose for our model. We also analyze our model's robustness to pose initialization in the experiment section. For evaluation, we used 34 on-trajectory views and 16 out-trajectory views to test RS correction and novel view synthesis results, respectively. To test our model in real-world scenarios, we also captured real scenes using the RS camera" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "RS-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 203, + 217 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 203, + 217 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 203, + 217 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 203, + 217 + ], + "type": "image", + "image_path": "4dc0615e6757afa002d38481f9f272938e24e0642c8abd588c3179c66348ebf2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 217, + 178, + 223 + ], + "lines": [ + { + "bbox": [ + 160, + 217, + 178, + 223 + ], + "spans": [ + { + "bbox": [ + 160, + 217, + 178, + 223 + ], + "type": "text", + "content": "Vanilla" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 233, + 480, + 277 + ], + "lines": [ + { + "bbox": [ + 130, + 233, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 480, + 277 + ], + "type": "text", + "content": "Fig.6: Qualitative results on synthetic scenes for ablation study. The corresponding error maps and zoomed regions are visualized in the bottom. Our full model has the smaller error to the GT, with less spatial misalignment and twist caused by the RS distortion." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 204, + 114, + 272, + 216 + ], + "blocks": [ + { + "bbox": [ + 204, + 114, + 272, + 216 + ], + "lines": [ + { + "bbox": [ + 204, + 114, + 272, + 216 + ], + "spans": [ + { + "bbox": [ + 204, + 114, + 272, + 216 + ], + "type": "image", + "image_path": "15da6701381db8447d3707d752b150f1bfe1b6ded98c90e2ef29e42158410eb8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 217, + 252, + 224 + ], + "lines": [ + { + "bbox": [ + 223, + 217, + 252, + 224 + ], + "spans": [ + { + "bbox": [ + 223, + 217, + 252, + 224 + ], + "type": "text", + "content": "Vanilla " + }, + { + "bbox": [ + 223, + 217, + 252, + 224 + ], + "type": "inline_equation", + "content": "+\\mathcal{L}_s" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 273, + 114, + 341, + 217 + ], + "blocks": [ + { + "bbox": [ + 273, + 114, + 341, + 217 + ], + "lines": [ + { + "bbox": [ + 273, + 114, + 341, + 217 + ], + "spans": [ + { + "bbox": [ + 273, + 114, + 341, + 217 + ], + "type": "image", + "image_path": "a41cd36dfedf0ed62189e995797b3210080a2467e43af63fdf4715258a5ae6d8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 285, + 217, + 329, + 224 + ], + "lines": [ + { + "bbox": [ + 285, + 217, + 329, + 224 + ], + "spans": [ + { + "bbox": [ + 285, + 217, + 329, + 224 + ], + "type": "text", + "content": "Vanilla " + }, + { + "bbox": [ + 285, + 217, + 329, + 224 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 285, + 217, + 329, + 224 + ], + "type": "text", + "content": " Mul.Sam." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 342, + 114, + 410, + 217 + ], + "blocks": [ + { + "bbox": [ + 342, + 114, + 410, + 217 + ], + "lines": [ + { + "bbox": [ + 342, + 114, + 410, + 217 + ], + "spans": [ + { + "bbox": [ + 342, + 114, + 410, + 217 + ], + "type": "image", + "image_path": "04ffffeba08f72f7190ff0173a9d678dbdd1d821ba51718842f34fc909e0b661.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 217, + 404, + 224 + ], + "lines": [ + { + "bbox": [ + 348, + 217, + 404, + 224 + ], + "spans": [ + { + "bbox": [ + 348, + 217, + 404, + 224 + ], + "type": "text", + "content": "\\[\n\\begin{array}{r}\\mathsf{Vanilla} + \\mathcal{L}_s + \\mathsf{Mul.Sam}. \\end{array}\n\\]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 410, + 114, + 479, + 217 + ], + "blocks": [ + { + "bbox": [ + 410, + 114, + 479, + 217 + ], + "lines": [ + { + "bbox": [ + 410, + 114, + 479, + 217 + ], + "spans": [ + { + "bbox": [ + 410, + 114, + 479, + 217 + ], + "type": "image", + "image_path": "23e74af092b9a5b26580d441a02a2e37410b5d0337f05beb05b844878fa9d247.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 217, + 460, + 223 + ], + "lines": [ + { + "bbox": [ + 429, + 217, + 460, + 223 + ], + "spans": [ + { + "bbox": [ + 429, + 217, + 460, + 223 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 133, + 290, + 325, + 318 + ], + "blocks": [ + { + "bbox": [ + 133, + 290, + 325, + 318 + ], + "lines": [ + { + "bbox": [ + 133, + 290, + 325, + 318 + ], + "spans": [ + { + "bbox": [ + 133, + 290, + 325, + 318 + ], + "type": "table", + "html": "
VanillaVanilla+LsVanilla+MSVanilla+Ls+MS
0.05090.03580.04620.0319
", + "image_path": "18bead5ad54558b077d423acfe4f24d9f3c56625d37b8cf47c66db1d895dab5f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 332, + 288, + 482, + 321 + ], + "lines": [ + { + "bbox": [ + 332, + 288, + 482, + 321 + ], + "spans": [ + { + "bbox": [ + 332, + 288, + 482, + 321 + ], + "type": "text", + "content": "Table 3: Average MSTE (the lower the better) results for ablation study." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 344, + 480, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 480, + 380 + ], + "type": "text", + "content": "EO-1312C. These RS images were captured by manually moving the camera during exposure, in a forward-facing manner, with a resolution of " + }, + { + "bbox": [ + 130, + 344, + 480, + 380 + ], + "type": "inline_equation", + "content": "400 \\times 500" + }, + { + "bbox": [ + 130, + 344, + 480, + 380 + ], + "type": "text", + "content": ". We utilize COLMAP [31,32] for the initial camera pose in our model." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 380, + 481, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 481, + 428 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 481, + 428 + ], + "type": "text", + "content": "Evaluation Metrics. The quality of the rendered image is evaluated with the commonly used metrics including PSNR, SSIM and LPIPS [38]. To evaluate the accuracy of the estimated camera trajectory, we calculate the Mean Square Trajectory Error (MSTE) with the ground-truth camera trajectory." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 444, + 279, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 444, + 279, + 456 + ], + "spans": [ + { + "bbox": [ + 131, + 444, + 279, + 456 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 463, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 482, + 571 + ], + "type": "text", + "content": "In each training iteration, we randomly choose 64 rows from each of the images and 2 pixels from each of these rows. We employ the Adam optimizer [16] with " + }, + { + "bbox": [ + 130, + 463, + 482, + 571 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 130, + 463, + 482, + 571 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 463, + 482, + 571 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 130, + 463, + 482, + 571 + ], + "type": "text", + "content": " to optimize our model. The learning rate is initially set at " + }, + { + "bbox": [ + 130, + 463, + 482, + 571 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 463, + 482, + 571 + ], + "type": "text", + "content": " and undergoes exponential decay throughout the optimization process. To avoid potential errors from inaccurate optical flow estimation, we disable multi-sampling and use the vanilla sampling strategy in the later part of the training process. This makes the model use much fewer but absolutely accurate data points in the final training phase. Each scene is trained for 200,000 iterations. All this training is conducted on a single NVIDIA 4090 GPU." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 131, + 586, + 239, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 239, + 599 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 239, + 599 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 605, + 480, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 480, + 642 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 480, + 642 + ], + "type": "text", + "content": "To comprehensively evaluate the effectiveness of each component in our RS-NeRF model, we carried out extensive ablation studies. The quantitative and qualitative results are presented in Tab. 2, Fig. 6 and Tab. 3, respectively." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "text", + "content": "Trajectory Smoothness Regularization. The accurate estimation of camera trajectory is crucial for effectively restoring normal images from RS inputs, given" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "M. Niu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 114, + 481, + 184 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 184 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 184 + ], + "type": "table", + "html": "
MethodsTORIIWINEPOOLFACTORYTANABATAAVERAGE
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
ON-TRIAL
\\(DU\\)21.236409.084519.87.4520.132923.08.5394.118919.95.5128.132518.17.5262.121320.46.5343.1180
\\(CVR\\)19.946148.145418.86.4090.205221.51.5173.188518.59.4851.188317.45.5277.184819.27.5108.1824
\\(JAM\\)20.666311.087219.50.4214.159522.43.5180.126319.34.4907.151017.87.5342.129219.96.5911.1307
RS-NeRF30.97.9058.039528.36.8482.067528.50.8137.056225.43.7673.086726.40.8614.065127.93.8393.0630
OUT-TRIAL
\\(NeRF\\)19.35.5976.253519.06.3845.355022.02.4740.319617.87.4075.357017.56.4958.301919.17.4719.3174
\\(DU+NeRF\\)20.33.6585.252820.85.4923.340423.25.5403.286019.16.5037.321918.14.5226.310220.35.5435.3022
\\(CVR+NeRF\\)23.67.7779.196722.27.6114.279124.56.6802.240819.81.5619.299618.58.6250.273621.78.6513.2580
\\(JAM+NeRF\\)24.13.7852.151222.14.5801.282225.51.6867.185520.46.5834.271718.63.6357.238822.18.6542.2259
RS-NeRF33.07.9499.024329.27.8703.056028.77.8313.045825.11.7738.080727.02.8881.049828.65.8627.0513
", + "image_path": "e0948a68ad70fd059f19418846722c10fdfb431eebbaa6229d395582448e9926.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 189, + 480, + 224 + ], + "lines": [ + { + "bbox": [ + 132, + 189, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 189, + 480, + 224 + ], + "type": "text", + "content": "Table 4: Quantitative comparison on synthetic scenes of different methods. DU, CVR, and JAM represent DeepUnroll [22], CVR [10], and JAMNet [11] respectively. We color code each result as best and second best." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 133, + 257, + 482, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 257, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 133, + 257, + 482, + 424 + ], + "type": "text", + "content": "the inherent formation process of RS effects. As evidenced in Tab. 2, RS-NeRF enhanced with trajectory smoothness regularization " + }, + { + "bbox": [ + 133, + 257, + 482, + 424 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_s)" + }, + { + "bbox": [ + 133, + 257, + 482, + 424 + ], + "type": "text", + "content": " shows notable improvements over the vanilla model. Specifically, it achieves a PSNR increase of 2.59 for on-trajectory views and 2.98 for out-trajectory views. Additionally, removing " + }, + { + "bbox": [ + 133, + 257, + 482, + 424 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 133, + 257, + 482, + 424 + ], + "type": "text", + "content": " from our full model leads to a PSNR decrease of 3.04 for on-trajectory views and 4.02 for out-trajectory views. The qualitative results further underscore the effectiveness of " + }, + { + "bbox": [ + 133, + 257, + 482, + 424 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 133, + 257, + 482, + 424 + ], + "type": "text", + "content": ". As illustrated in Fig. 6, the vanilla RS-NeRF, when complemented with " + }, + { + "bbox": [ + 133, + 257, + 482, + 424 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 133, + 257, + 482, + 424 + ], + "type": "text", + "content": ", exhibits reduced spatial misalignment and a smaller synthesis error to the ground truth. Furthermore, Tab. 3 reveals that the vanilla RS-NeRF tends to provide an inaccurate trajectory estimation, primarily focusing on the photo-metric loss for each image and overlooking the pattern of camera movement across different RS frames. In contrast, the 'Vanilla + " + }, + { + "bbox": [ + 133, + 257, + 482, + 424 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 133, + 257, + 482, + 424 + ], + "type": "text", + "content": "' model, with the addition of smoothness regularization, yields a more accurate estimation of the camera trajectory, resulting in improved synthesis outcomes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 425, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 425, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 133, + 425, + 482, + 604 + ], + "type": "text", + "content": "Multi-Sampling Algorithm. The Multi-sampling algorithm in RS-NeRF significantly enhances the training data available for each intermediate pose by calculating the PP-ratio for each pixel, marking a notable advancement over the vanilla model. According to the results in Tab. 2, RS-NeRF equipped with the multi-sampling algorithm shows substantial improvements: a PSNR increase of 3.42 for on-trajectory views and 2.95 for out-trajectory views. Moreover, excluding the multi-sampling algorithm from our full model leads to a PSNR reduction of 3.87 for on-trajectory views and 3.99 for out-trajectory views. The qualitative results also affirm the effectiveness of it. As depicted in Fig. 6, our model, with the multi-sampling algorithm, displays fewer artifacts and a smaller synthesis error to the ground truth. Interestingly, the multi-sampling algorithm also seems to enhance trajectory estimation. As shown in Tab. 3, RS-NeRF with only the multi-sampling algorithm (Vanilla + MS) provides a more accurate trajectory estimation than the vanilla model, which likely contributes to its significantly improved synthesis results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 622, + 226, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 226, + 634 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 226, + 634 + ], + "type": "text", + "content": "4.4 Comparisons" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 480, + 665 + ], + "type": "text", + "content": "To evaluate the effectiveness of RS correction for on-trajectory images, our method is compared with three state-of-the-art 2D RS correction techniques:" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 407, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 407, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 407, + 91, + 447, + 100 + ], + "type": "text", + "content": "RS-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 480, + 292 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 480, + 292 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 480, + 292 + ], + "type": "image", + "image_path": "b2064caa354974159a7359d9d56da512a4b4b5a676905666150776e8ace0afda.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 300, + 482, + 335 + ], + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 335 + ], + "type": "text", + "content": "Fig. 7: Qualitative comparison results on synthetic scenes. The corresponding error maps and zoomed regions are visualized in the bottom. DU, CVR, and JAM represent DeepUnroll [22], CVR [10], and JAMNet [11] respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 134, + 350, + 481, + 404 + ], + "blocks": [ + { + "bbox": [ + 134, + 350, + 481, + 404 + ], + "lines": [ + { + "bbox": [ + 134, + 350, + 481, + 404 + ], + "spans": [ + { + "bbox": [ + 134, + 350, + 481, + 404 + ], + "type": "table", + "html": "
MethodsTORIIWINEPOOLFACTORYTANABATAAVERAGE
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
USB-NeRF [19]26.32.8283.082524.72.7105.167326.72.7130.159224.32.7213.136923.05.7359.133325.02.7418.1359
DRSC [26]23.04.8008.053921.84.6895.088022.85.6962.087420.85.6901.096721.19.7492.073121.95.7251.0798
RS-NeRF30.97.9058.039528.36.8482.067528.50.8137.056225.43.7673.086726.40.8614.065127.93.8393.0630
USB-NeRF [19]28.95.8932.073625.75.7446.161928.13.7765.157625.25.7676.142724.62.7853.120426.54.7935.1313
DRSC [26] + NeRF21.51.6790.283120.30.5145.300122.66.5614.002118.99.5295.290418.35.5643.296220.36.5697.2944
RS-NeRF33.07.9499.024329.27.8703.056028.77.8313.045825.11.7738.080727.02.8881.049828.65.8627.0513
", + "image_path": "73b0b5e0cb4bc75864fcd46e39c4c021b9d467b02286186231c135b16a4b205f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 409, + 482, + 434 + ], + "lines": [ + { + "bbox": [ + 130, + 409, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 482, + 434 + ], + "type": "text", + "content": "Table 5: Quantitative comparisons on synthetic scenes with DRSC [26] + NeRF and USB-NeRF [19]. We color code each result as best and second best." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 471, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 567 + ], + "type": "text", + "content": "DeepUnroll [22], CVR [10], and JAMNet [11]. To evaluate out-trajectory views, we selected several baselines for comparison. The most basic approach is to train NeRF directly using RS images. We further compare our method with two-stage baselines that initially apply state-of-the-art 2D RS correction methods (DeepUnroll [22], CVR [10], and JAMNet [11]) to restore GS images, followed by training NeRF with these restored images. For 2D RS correction methods, we utilized the official checkpoints for evaluation. Furthermore, we compare RS-NeRF with DRSC [26] and USB-NeRF [19]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Quantitative Results. Quantitative comparison results are shown in Tab. 4 and Tab. 5. 2D RS correction methods do not yield satisfactory outcomes and receive obvious lower metrics due to twist artifacts and spatial misalignment. In contrast, our approach significantly exceeds these methods in all metrics from both on-trajectory and out-trajectory views by incorporating the RS formation process into the modeling and utilizing the multi-view information from continuous RS frames. USB-NeRF [19] integrates basic camera motion modeling, i.e., assigning unique camera poses for different rows to account for the RS ef" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "M. Niu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 202, + 224 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 202, + 224 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 202, + 224 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 202, + 224 + ], + "type": "image", + "image_path": "6657998b1800206c5019007af586d6b6f12243ea3fe9b38d97b62666fd9c90bc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 163, + 225, + 174, + 232 + ], + "lines": [ + { + "bbox": [ + 163, + 225, + 174, + 232 + ], + "spans": [ + { + "bbox": [ + 163, + 225, + 174, + 232 + ], + "type": "text", + "content": "NeRF" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 203, + 115, + 272, + 170 + ], + "blocks": [ + { + "bbox": [ + 203, + 115, + 272, + 170 + ], + "lines": [ + { + "bbox": [ + 203, + 115, + 272, + 170 + ], + "spans": [ + { + "bbox": [ + 203, + 115, + 272, + 170 + ], + "type": "image", + "image_path": "cfa306c44d6347bbb3f520c2194330c69089855cb9de3c26cc1e8642aab4a1ce.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 203, + 171, + 272, + 224 + ], + "blocks": [ + { + "bbox": [ + 203, + 171, + 272, + 224 + ], + "lines": [ + { + "bbox": [ + 203, + 171, + 272, + 224 + ], + "spans": [ + { + "bbox": [ + 203, + 171, + 272, + 224 + ], + "type": "image", + "image_path": "1bdd2a722cd3e90373eb59c57669a2f2185cc660bdddbc980d5d2408c10e4097.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 225, + 246, + 232 + ], + "lines": [ + { + "bbox": [ + 223, + 225, + 246, + 232 + ], + "spans": [ + { + "bbox": [ + 223, + 225, + 246, + 232 + ], + "type": "inline_equation", + "content": "\\mathrm{DU} + \\mathrm{NeRF}" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 272, + 115, + 341, + 170 + ], + "blocks": [ + { + "bbox": [ + 272, + 115, + 341, + 170 + ], + "lines": [ + { + "bbox": [ + 272, + 115, + 341, + 170 + ], + "spans": [ + { + "bbox": [ + 272, + 115, + 341, + 170 + ], + "type": "image", + "image_path": "581d6c56a6f8e06a1ea330c5266424fa921df0ce765f99106986eb1cd9baf294.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 272, + 171, + 341, + 224 + ], + "blocks": [ + { + "bbox": [ + 272, + 171, + 341, + 224 + ], + "lines": [ + { + "bbox": [ + 272, + 171, + 341, + 224 + ], + "spans": [ + { + "bbox": [ + 272, + 171, + 341, + 224 + ], + "type": "image", + "image_path": "eda2bdc76b23d396df6c13aaf3be996d55e54061ebac24968905db330f7448c2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 291, + 225, + 317, + 232 + ], + "lines": [ + { + "bbox": [ + 291, + 225, + 317, + 232 + ], + "spans": [ + { + "bbox": [ + 291, + 225, + 317, + 232 + ], + "type": "inline_equation", + "content": "\\mathrm{CVR} + \\mathrm{NeRF}" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 342, + 115, + 410, + 170 + ], + "blocks": [ + { + "bbox": [ + 342, + 115, + 410, + 170 + ], + "lines": [ + { + "bbox": [ + 342, + 115, + 410, + 170 + ], + "spans": [ + { + "bbox": [ + 342, + 115, + 410, + 170 + ], + "type": "image", + "image_path": "2f50c8a6d9c15e61e37ba3248d4847d5dc46e3c6372b4587bba94a72522da7c2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 342, + 171, + 410, + 224 + ], + "blocks": [ + { + "bbox": [ + 342, + 171, + 410, + 224 + ], + "lines": [ + { + "bbox": [ + 342, + 171, + 410, + 224 + ], + "spans": [ + { + "bbox": [ + 342, + 171, + 410, + 224 + ], + "type": "image", + "image_path": "3ba84ca3386737e101adba93a607f5587b02ebd79f2487527b4f44950c03c362.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 225, + 391, + 232 + ], + "lines": [ + { + "bbox": [ + 365, + 225, + 391, + 232 + ], + "spans": [ + { + "bbox": [ + 365, + 225, + 391, + 232 + ], + "type": "text", + "content": "JAM + NeRF" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 410, + 115, + 479, + 171 + ], + "blocks": [ + { + "bbox": [ + 410, + 115, + 479, + 171 + ], + "lines": [ + { + "bbox": [ + 410, + 115, + 479, + 171 + ], + "spans": [ + { + "bbox": [ + 410, + 115, + 479, + 171 + ], + "type": "image", + "image_path": "40dd9cb377f2b751aff43979b3ac5efade0c1448402622ed2cbb8faa9dc0a85b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 410, + 171, + 479, + 224 + ], + "blocks": [ + { + "bbox": [ + 410, + 171, + 479, + 224 + ], + "lines": [ + { + "bbox": [ + 410, + 171, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 410, + 171, + 479, + 224 + ], + "type": "image", + "image_path": "12d2bcffde69cc2dbf2a2575efd95ab879bd09648d36d46b79229e76622564b2.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 225, + 454, + 232 + ], + "lines": [ + { + "bbox": [ + 436, + 225, + 454, + 232 + ], + "spans": [ + { + "bbox": [ + 436, + 225, + 454, + 232 + ], + "type": "text", + "content": "RS-NeRF" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 134, + 276, + 220, + 333 + ], + "blocks": [ + { + "bbox": [ + 131, + 241, + 480, + 263 + ], + "lines": [ + { + "bbox": [ + 131, + 241, + 480, + 263 + ], + "spans": [ + { + "bbox": [ + 131, + 241, + 480, + 263 + ], + "type": "text", + "content": "Fig. 8: Qualitative comparison results on real-world scenes. Our method achieves superior performance on real-world scenes." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 134, + 276, + 220, + 333 + ], + "lines": [ + { + "bbox": [ + 134, + 276, + 220, + 333 + ], + "spans": [ + { + "bbox": [ + 134, + 276, + 220, + 333 + ], + "type": "image", + "image_path": "a966797d84ebcea10131cc5b272aa06e38607dd32a85c7e7679b1d25614d5084.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 221, + 277, + 292, + 333 + ], + "blocks": [ + { + "bbox": [ + 221, + 277, + 292, + 333 + ], + "lines": [ + { + "bbox": [ + 221, + 277, + 292, + 333 + ], + "spans": [ + { + "bbox": [ + 221, + 277, + 292, + 333 + ], + "type": "image", + "image_path": "cee2f2f030e6a46a5b34df4307c1fd2c8a9b7c36f6098ae17a6c339f198f3a96.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 293, + 277, + 361, + 333 + ], + "blocks": [ + { + "bbox": [ + 293, + 277, + 361, + 333 + ], + "lines": [ + { + "bbox": [ + 293, + 277, + 361, + 333 + ], + "spans": [ + { + "bbox": [ + 293, + 277, + 361, + 333 + ], + "type": "image", + "image_path": "43dabddcc826c9a19751822818e9e1ece21f5a01cfee20f9e16930d3e82a2c70.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 362, + 277, + 392, + 333 + ], + "blocks": [ + { + "bbox": [ + 362, + 277, + 392, + 333 + ], + "lines": [ + { + "bbox": [ + 362, + 277, + 392, + 333 + ], + "spans": [ + { + "bbox": [ + 362, + 277, + 392, + 333 + ], + "type": "image", + "image_path": "e01f50d109a9aa8bacf9fddceb5a251544a8836f127f27474686ffb6338a4f99.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 392, + 277, + 449, + 333 + ], + "blocks": [ + { + "bbox": [ + 392, + 277, + 449, + 333 + ], + "lines": [ + { + "bbox": [ + 392, + 277, + 449, + 333 + ], + "spans": [ + { + "bbox": [ + 392, + 277, + 449, + 333 + ], + "type": "image", + "image_path": "1b1d3c533dbfb143d2691a3ad57891c13634e753519c83df1d6b9200bd39f2d5.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 449, + 277, + 495, + 333 + ], + "blocks": [ + { + "bbox": [ + 449, + 277, + 495, + 333 + ], + "lines": [ + { + "bbox": [ + 449, + 277, + 495, + 333 + ], + "spans": [ + { + "bbox": [ + 449, + 277, + 495, + 333 + ], + "type": "image", + "image_path": "34a9457366ba9351f0e9ce7310caa70571d843988052656834c6accd1e89d893.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 134, + 335, + 214, + 403 + ], + "blocks": [ + { + "bbox": [ + 134, + 335, + 214, + 403 + ], + "lines": [ + { + "bbox": [ + 134, + 335, + 214, + 403 + ], + "spans": [ + { + "bbox": [ + 134, + 335, + 214, + 403 + ], + "type": "image", + "image_path": "2b4fcce7d7e09f08c1876aea07e47000c65de683dd42c6fa67fba99f1b673f51.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 414, + 480, + 437 + ], + "lines": [ + { + "bbox": [ + 131, + 414, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 131, + 414, + 480, + 437 + ], + "type": "text", + "content": "Fig. 9: Qualitative comparison results on synthetic and real-world scenes with DRSC [26] + NeRF and USB-NeRF [19]." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 214, + 335, + 249, + 403 + ], + "blocks": [ + { + "bbox": [ + 214, + 335, + 249, + 403 + ], + "lines": [ + { + "bbox": [ + 214, + 335, + 249, + 403 + ], + "spans": [ + { + "bbox": [ + 214, + 335, + 249, + 403 + ], + "type": "image", + "image_path": "9699e66dd9bb65be2839711c92685d8455938ac7cf77f391f2f1a372c8adfc72.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 250, + 335, + 363, + 403 + ], + "blocks": [ + { + "bbox": [ + 250, + 335, + 363, + 403 + ], + "lines": [ + { + "bbox": [ + 250, + 335, + 363, + 403 + ], + "spans": [ + { + "bbox": [ + 250, + 335, + 363, + 403 + ], + "type": "image", + "image_path": "0d6c6bc14c5273b51a44eb1048bce31fd3f4a449453b53b7d8a8fbbc8e84d992.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 364, + 335, + 445, + 403 + ], + "blocks": [ + { + "bbox": [ + 364, + 335, + 445, + 403 + ], + "lines": [ + { + "bbox": [ + 364, + 335, + 445, + 403 + ], + "spans": [ + { + "bbox": [ + 364, + 335, + 445, + 403 + ], + "type": "image", + "image_path": "cde5fd59937dfcff3c61b1ac3f17a8d1d821eb0bf23c9c7f8197dc546a4a4cc1.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 445, + 335, + 479, + 403 + ], + "blocks": [ + { + "bbox": [ + 445, + 335, + 479, + 403 + ], + "lines": [ + { + "bbox": [ + 445, + 335, + 479, + 403 + ], + "spans": [ + { + "bbox": [ + 445, + 335, + 479, + 403 + ], + "type": "image", + "image_path": "001798100d7cb8cf2de661f89052da8f22a8ba7d07e0d2eeb3ece16e2202b48b.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "bbox": [ + 130, + 461, + 480, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 480, + 485 + ], + "type": "text", + "content": "fect, without additional technical improvements. The comparison shows that this approach fails to deliver satisfactory results compared to RS-NeRF." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "Qualitative Results. The qualitative performance of our method is compared to others through tests on synthetic and real-world datasets. In synthetic scenes, as shown in Fig. 7, 2D RS correction methods struggle with twisting artifacts and spatial misalignment for on-trajectory views. In contrast, our method takes advantage of multi-view information for RS correction, leading to significantly better results. For out-trajectory views, while 2D RS correction methods combined with NeRF moderately enhance the reconstruction results, they still exhibit noticeable visual flaws, such as blurriness. RS-NeRF excels by integrating the RS formation process into the model and accurately learning the underlying 3D representation for joint RS correction and novel view synthesis. Qualitative results in real world scenes are shown in Fig. 8. We can observe that baseline methods show obvious artifacts, particularly around object edges. Also, the performance of 2D RS correction combined with NeRF is poorer in real-world scenes than in synthetic ones, which further demonstrates the generalization issues of 2D RS correction methods with real-world data. RS-NeRF, on the other hand," + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 407, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 407, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 407, + 91, + 447, + 100 + ], + "type": "text", + "content": "RS-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 114, + 347, + 180 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 347, + 180 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 347, + 180 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 347, + 180 + ], + "type": "table", + "html": "
PerturbationON TRAJ.OUT TRAJ.MSTE ↓
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
None27.93.8393.063028.65.8627.0513.0319
U[-0.01, 0.01]27.97.8403.061028.84.8680.0485.0297
U[-0.03, 0.03]28.31.8479.057829.00.8693.0470.0304
U[-0.05, 0.05]28.47.8530.058029.24.8769.0466.0308
U[-0.07, 0.07]28.43.8504.059729.25.8771.0471.0336
", + "image_path": "9eb7fcc1c51ffaa05a185d38fff9c6babfd8c5b86e157d6c078d3e7f41fa9c0d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 353, + 114, + 482, + 193 + ], + "lines": [ + { + "bbox": [ + 353, + 114, + 482, + 193 + ], + "spans": [ + { + "bbox": [ + 353, + 114, + 482, + 193 + ], + "type": "text", + "content": "Table 6: Quantitative results on synthetic dataset for different noise perturbations on pose initialization. The results demonstrate that our model is robust to the accuracy of pose initialization." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 213, + 480, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 480, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 480, + 260 + ], + "type": "text", + "content": "significantly outperforms these methods with clear edges and detailed rendering, consistent with results in synthetic scenes. Fig. 9 also demonstrates the superiority of RS-NeRF over DRSC [26] + NeRF and USB-NeRF [19] on synthetic and real-world scenes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 275, + 297, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 275, + 297, + 288 + ], + "spans": [ + { + "bbox": [ + 132, + 275, + 297, + 288 + ], + "type": "text", + "content": "4.5 Discussion and Limitations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 292, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 292, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 482, + 422 + ], + "type": "text", + "content": "Robustness to pose initialization. We evaluated the robustness of our method to the precision of pose initialization on synthetic dataset by introducing random noise of varying scales to the initial poses in the SE(3) space and using these altered poses for initialization. The results, as detailed in Tab. 6, indicate that our method maintains superior performance across different levels of noise perturbation. These results confirm the robustness of RS-NeRF in the accuracy of pose initialization. However, it is important to note that our method might fail when completely random poses are used for initialization. This challenge is also an ongoing issue in the field of NeRF-without-Poses, as discussed in works like [5, 20]. Addressing this limitation is a direction that we plan to pursue in future research." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 424, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 509 + ], + "type": "text", + "content": "Limitations. Similar to previous works that address NeRF with abnormal inputs [24, 36], our model is designed to handle degradation (referred to the RS effects in our context) caused by camera movements, since we primarily model the camera trajectory. When encountering scenarios where RS effects are caused by object movements, i.e., dynamic scenes, our method may fail and produce unsatisfying results. We plan to involve the modeling of object movements based on dynamic NeRFs in future work." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 523, + 220, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 220, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 220, + 536 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we present RS-NeRF, a method designed to synthesize normal images from novel views using input with Rolling Shutter (RS) distortions. We formulate a physical model that replicates the image formation process under RS conditions and jointly optimizes the NeRF parameters and the camera extrinsic for each image row. We then identify and address the fundamental limitations of the standard RS-NeRF model by introducing two innovative approaches: trajectory smoothness regularization and the multi-sampling algorithm. Through rigorous experimentation, we demonstrate that RS-NeRF surpasses previous methods in both synthetic and real-world scenarios, confirming its efficiency in rectifying distortions caused by RS cameras." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "M. Niu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 137, + 482, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 137, + 482, + 162 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 482, + 162 + ], + "type": "text", + "content": "This research was supported in part by JSPS KAKENHI Grant Numbers 24K223-18, 22H00529, 20H05951, JST-Mirai Program JPMJMI23G1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 178, + 198, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 178, + 198, + 190 + ], + "spans": [ + { + "bbox": [ + 133, + 178, + 198, + 190 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 200, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 200, + 481, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 200, + 481, + 233 + ], + "spans": [ + { + "bbox": [ + 138, + 200, + 481, + 233 + ], + "type": "text", + "content": "1. Albl, C., Kukelova, Z., Larsson, V., Pajdla, T.: Rolling shutter camera absolute pose. IEEE transactions on pattern analysis and machine intelligence 42(6), 1439-1452 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 233, + 481, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 233, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 233, + 481, + 266 + ], + "type": "text", + "content": "2. Attal, B., Laidlaw, E., Gokaslan, A., Kim, C., Richardt, C., Tompkin, J., O'Toole, M.: Törf: Time-of-flight radiance fields for dynamic scene view synthesis. Advances in neural information processing systems 34, 26289-26301 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 266, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 481, + 308 + ], + "type": "text", + "content": "3. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 308, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 308, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 308, + 481, + 352 + ], + "type": "text", + "content": "4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mipnerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 352, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 352, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 352, + 481, + 385 + ], + "type": "text", + "content": "5. Bian, W., Wang, Z., Li, K., Bian, J.W., Prisacariu, V.A.: Nope-nerf: Optimising neural radiance field with no pose prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4160-4169 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 384, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 384, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 384, + 481, + 418 + ], + "type": "text", + "content": "6. Cao, M., Zhong, Z., Wang, J., Zheng, Y., Yang, Y.: Learning adaptive warping for real-world rolling shutter correction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17785-17793 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 418, + 481, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 481, + 459 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 481, + 459 + ], + "type": "text", + "content": "7. Chen, T., Wang, P., Fan, Z., Wang, Z.: Aug-nerf: Training stronger neural radiance fields with triple-level physically-grounded augmentations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15191-15202 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 459, + 481, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 459, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 138, + 459, + 481, + 503 + ], + "type": "text", + "content": "8. Dave, A., Zhao, Y., Veeraraghavan, A.: Pandora: Polarization-aided neural decomposition of radiance. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part VII. pp. 538–556. Springer (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 503, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 503, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 138, + 503, + 481, + 536 + ], + "type": "text", + "content": "9. Fan, B., Dai, Y., He, M.: Sunet: symmetric undistortion network for rolling shutter correction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4541-4550 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 536, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 481, + 568 + ], + "type": "text", + "content": "0. Fan, B., Dai, Y., Zhang, Z., Liu, Q., He, M.: Context-aware video reconstruction for rolling shutter cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17572-17582 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 481, + 601 + ], + "type": "text", + "content": "1. Fan, B., Mao, Y., Dai, Y., Wan, Z., Liu, Q.: Joint appearance and motion learning for efficient rolling shutter correction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5671-5681 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "type": "text", + "content": "2. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: Fastnerf: High-fidelity neural rendering at 200fps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14346-14355 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 481, + 666 + ], + "type": "text", + "content": "3. Guo, Y.C., Kang, D., Bao, L., He, Y., Zhang, S.H.: Nerfren: Neural radiance fields with reflections. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18409-18418 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "RS-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "14. Hedborg, J., Forssén, P.E., Felsberg, M., Ringaby, E.: Rolling shutter bundle adjustment. In: 2012 IEEE Conference on Computer Vision and Pattern Recognition. pp. 1434-1441. IEEE (2012)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "type": "text", + "content": "15. Kim, J.H., Latif, Y., Reid, I.: Rrd-slam: Radial-distorted rolling-shutter direct slam. In: 2017 IEEE International Conference on Robotics and Automation (ICRA). pp. 5148-5154. IEEE (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 480, + 205 + ], + "type": "text", + "content": "16. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 206, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 480, + 228 + ], + "type": "text", + "content": "17. Klenk, S., Koestler, L., Scaramuzza, D., Cremers, D.: E-nerf: Neural radiance fields from a moving event camera. IEEE Robotics and Automation Letters (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 229, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 480, + 262 + ], + "type": "text", + "content": "18. Klingner, B., Martin, D., Roseborough, J.: Street view motion-from-structure-from-motion. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 953–960 (2013)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 262, + 480, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 480, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 480, + 296 + ], + "type": "text", + "content": "19. Li, M., Wang, P., Zhao, L., Liao, B., Liu, P.: USB-neRF: Unrolling shutter bundle adjusted neural radiance fields. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=igfDXfMvm5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 297, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 480, + 319 + ], + "type": "text", + "content": "20. Lin, C.H., Ma, W.C., Torralba, A., Lucey, S.: Barf: Bundle-adjusting neural radiance fields. In: IEEE International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 319, + 480, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 480, + 341 + ], + "type": "text", + "content": "21. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "type": "text", + "content": "22. Liu, P., Cui, Z., Larsson, V., Pollefeys, M.: Deep shutter unrolling network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5941-5949 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "type": "text", + "content": "23. Liu, P., Zuo, X., Larsson, V., Pollefeys, M.: Mba-vo: Motion blur aware visual odometry. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5550-5559 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 408, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 480, + 441 + ], + "type": "text", + "content": "24. Ma, L., Li, X., Liao, J., Zhang, Q., Wang, X., Wang, J., Sander, P.V.: Deblur-nerf: Neural radiance fields from blurry images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12861-12870 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 442, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 480, + 475 + ], + "type": "text", + "content": "25. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 475, + 480, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 480, + 498 + ], + "type": "text", + "content": "26. Qu, D., Liao, B., Zhang, H., Ait-Aider, O., Lao, Y.: Fast rolling shutter correction in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 498, + 480, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 480, + 531 + ], + "type": "text", + "content": "27. Reiser, C., Peng, S., Liao, Y., Geiger, A.: Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14335-14345 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 532, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 480, + 564 + ], + "type": "text", + "content": "28. Rengarajan, V., Balaji, Y., Rajagopalan, A.: Unrolling the shutter: Cnn to correct motion distortions. In: Proceedings of the IEEE Conference on computer Vision and Pattern Recognition. pp. 2291-2299 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 565, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 480, + 597 + ], + "type": "text", + "content": "29. Saurer, O., Koser, K., Bouguet, J.Y., Pollefeys, M.: Rolling shutter stereo. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 465-472 (2013)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 598, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 480, + 632 + ], + "type": "text", + "content": "30. Saurer, O., Pollefeys, M., Lee, G.H.: Sparse to dense 3d reconstruction from rolling shutter images. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3337-3345 (2016)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "31. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4104-4113 (2016)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "M. Niu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 490 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "type": "text", + "content": "32. Schonberger, J.L., Zheng, E., Frahm, J.M., Pollefeys, M.: Pixelwise view selection for unstructured multi-view stereo. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14. pp. 501-518. Springer (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 162, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 162, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 162, + 482, + 194 + ], + "type": "text", + "content": "33. Teed, Z., Deng, J.: Raft: Recurrent all-pairs field transforms for optical flow. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 402-419. Springer (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "34. Vasu, S., Rajagopalan, A., et al.: Occlusion-aware rolling shutter rectification of 3d scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 636-645 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "text", + "content": "35. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-NeRF: Structured view-dependent appearance for neural radiance fields. CVPR (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "36. Wang, P., Zhao, L., Ma, R., Liu, P.: Bad-nerf: Bundle adjusted deblur neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4170-4179 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "text", + "content": "37. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5438-5448 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "type": "text", + "content": "38. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "text", + "content": "39. Zhong, Z., Cao, M., Sun, X., Wu, Z., Zhou, Z., Zheng, Y., Lin, S., Sato, I.: Bringing rolling shutter images alive with dual reversed distortion. In: European Conference on Computer Vision. pp. 233-249. Springer (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 391, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 481, + 423 + ], + "type": "text", + "content": "40. Zhong, Z., Zheng, Y., Sato, I.: Towards rolling shutter correction and deblurring in dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9219-9228 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "text", + "content": "41. Zhou, X., Duan, P., Ma, Y., Shi, B.: Evunroll: Neuromorphic events based rolling shutter image correction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17775-17784 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "text", + "content": "42. Zhuang, B., Tran, Q.H., Ji, P., Cheong, L.F., Chandraker, M.: Learning structure-and-motion-aware rolling shutter correction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4551-4560 (2019)" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "RS-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/RSL-BA_ Rolling Shutter Line Bundle Adjustment/25dc3962-792b-46d2-8db3-b73a1efe60d9_content_list.json b/2024/RSL-BA_ Rolling Shutter Line Bundle Adjustment/25dc3962-792b-46d2-8db3-b73a1efe60d9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..34b7e94b0f69b72e0b7fa9b071b1d51bfac09383 --- /dev/null +++ b/2024/RSL-BA_ Rolling Shutter Line Bundle Adjustment/25dc3962-792b-46d2-8db3-b73a1efe60d9_content_list.json @@ -0,0 +1,2023 @@ +[ + { + "type": "text", + "text": "RSL-BA: Rolling Shutter Line Bundle Adjustment", + "text_level": 1, + "bbox": [ + 326, + 140, + 673, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yongcong Zhang $^{1,3,*}$ , Bangyan Liao $^{2,*}$ , Yifei Xue $^{1,3}$ , Chen Lu $^{4}$ , Peidong Liu $^{2}$ , and Yizhen Lao $^{1,\\dagger}$", + "bbox": [ + 251, + 210, + 750, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Hunan University $^{2}$ Westlake University $^{3}$ DaHe.AI $^{4}$ Dreame Technology Company", + "bbox": [ + 315, + 253, + 684, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The line is a prevalent element in man-made environments, inherently encoding spatial structural information, thus making it a more robust choice for feature representation in practical applications. Despite its apparent advantages, previous rolling shutter bundle adjustment (RSBA) methods have only supported sparse feature points, which lack robustness, particularly in degenerate environments. In this paper, we introduce the first rolling shutter line-based bundle adjustment solution, $RSL-BA$ . Specifically, we initially establish the rolling shutter camera line projection theory utilizing Plücker line parameterization. Subsequently, we derive a series of reprojection error formulations which are stable and efficient. Finally, we theoretically and experimentally demonstrate that our method can prevent three common degeneracies, one of which is first discovered in this paper. Extensive synthetic and real data experiments demonstrate that our method achieves efficiency and accuracy comparable to existing point-based rolling shutter bundle adjustment solutions.", + "bbox": [ + 261, + 323, + 738, + 531 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Rolling Shutter $\\cdot$ Bundle Adjustment", + "bbox": [ + 261, + 544, + 598, + 559 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 587, + 375, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bundle Adjustment (BA) is a crucial step in multi-view 3D reconstruction, as it jointly optimizes camera poses and scene structure. Although feature point-based BA [6,23,26] currently dominates the academic landscape, line-based BA has been gaining increasing attention from researchers [16,28]. Lines, being one of the most common features in man-made environments, effectively describe the structural information of 3D scenes, in contrast to points which rely solely on local image patches. This structural information provides robust constraints for visual reconstruction [27], making line a more suitable choice in challenging scenarios.", + "bbox": [ + 212, + 619, + 787, + 755 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In parallel with classical bundle adjustment [26], several methods [1,13] have incorporated the rolling shutter camera model to simultaneously estimate camera", + "bbox": [ + 214, + 756, + 785, + 787 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution", + "bbox": [ + 232, + 797, + 374, + 811 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author: (yizhenlao@hnu.edu.cn)", + "bbox": [ + 232, + 811, + 557, + 825 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Project page: https://github.com/zhangtaxue/RSL-BA", + "bbox": [ + 232, + 825, + 601, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "poses, structure, and instantaneous motion parameters. However, all existing rolling shutter bundle adjustment (RSBA) methods rely solely on point features, while line features have never been addressed in RSBA. We identify two challenges that hinder the direct combination of line-based BA and the rolling shutter camera model.", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The first challenge arises from the time-dependent exposure of RS cameras. Unlike CCD cameras and their global shutter (GS) counterparts, RS cameras capture images in a scanline-by-scanline manner. Consequently, as illustrated in Fig. 1, images taken by moving RS cameras exhibit distortions known as the RS effect. This effect curves the projection of each 3D straight line, making it non-trivial to directly transfer the line-based BA to the rolling shutter camera setting.", + "bbox": [ + 212, + 222, + 787, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The second challenge is associated with degeneracy. As shown in [1, 30], degeneracy is one of the main obstacles in RSBA since iterative optimization can easily be collapsed in some degenerate solutions which are far from the ground truth solutions. Although several strategies have been proposed to mitigate degeneracy in RSBA [1, 12, 14], a specific strategy to address degeneracy in line-based RSBA is still lacking.", + "bbox": [ + 212, + 327, + 787, + 417 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In conclusion, an accurate and robust solution to line-based RSBA is still missing. Such a method would be vital in the potential widespread deployment of 3D vision with RS imaging systems.", + "bbox": [ + 212, + 417, + 787, + 464 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1.1 Related Works", + "text_level": 1, + "bbox": [ + 215, + 484, + 385, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video-based RSBA. The assumption of smooth continuous trajectories is widely employed for RS video inputs to reduce the optimization parameter space and enhance algorithm robustness. In [8], Hedborg et al. present an RSBA algorithm that interpolates the motion between consecutive frames. Zhuang et al. [29] further propose an optical flow-based RSBA, which are developed to recover the relative pose of an RS camera that undergoes constant velocity and acceleration motion, respectively. Following this assumption, a spline-based camera trajectory motion model is proposed by [20].", + "bbox": [ + 212, + 507, + 792, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unordered RSBA. As the unordered image set is the standard input for SfM, Albl et al. first addresses this setting by explicitly velocity modeling and optimization. Besides, a planar degeneracy configuration has also been disclosed in [1]. While in [12], Lao et al. propose a camera-based RSBA to simulate the actual camera projection, exhibiting the degeneracy resilience ability. In [10], Ito et al. establish the equivalence between self-calibrated SfM and RSSFM based on the pure rotation instantaneous motion model and affine camera assumption, while the work of [13] draws the equivalence between RSSfM and non-rigid SfM. Recently, Liao et al. proposed two techniques to boost the accuracy and robustness of RSBA in [14].", + "bbox": [ + 212, + 628, + 787, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Line-based GSBA. In [2, 25], the authors propose a comprehensive line-based SFM pipeline, while [9, 18] introduce line-based incremental SFM frameworks. Wei et al. in [28] employs planes and points to guide matching, achieving efficient line reconstruction. Recently, Lui et al. introduced LIMAP in [16], where the", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Zhang, B. Liao et al.", + "bbox": [ + 271, + 114, + 431, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0f844bdff8bc036f12a958712c13ee7c6449121cf578752f82753cc210d6aa0d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 148, + 320, + 191 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c7a70a41c52a0b2a6bccf64dce3bcc1c6cf2f636c2948a8ad504aba5ed6fc62c.jpg", + "image_caption": [ + "input" + ], + "image_footnote": [], + "bbox": [ + 223, + 193, + 316, + 234 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2851336b6560c56be37b00028c47265d39d9850e08aaa71de39a2bd6db22a881.jpg", + "image_caption": [ + "Fig. 1: The pipeline of proposed RSL-BA with some example results. Starting with the input of some RS images (left), the optimization of camera poses and 3D line coordinates is performed using proposed RSL-BA (middle), yielding well-reconstructed 3D line segments and accurately estimated camera poses (right)." + ], + "image_footnote": [], + "bbox": [ + 330, + 148, + 625, + 252 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b0bbc13a2a053ad2dbff8f56de0c1f750023518efb77dea927cf1cea0f815aa4.jpg", + "image_caption": [ + "output" + ], + "image_footnote": [], + "bbox": [ + 629, + 150, + 767, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "authors devised various strategies and algorithms to achieve efficient and precise line reconstruction. In [15, 21, 31], hybrid strategies combining lines and feature points are utilized to enhance the accuracy and robustness of SLAM.", + "bbox": [ + 212, + 363, + 782, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To this end, we propose a novel $RSL-BA$ algorithm to solve the line-based RSBA problem. Specifically, we first establish the rolling shutter line projection theory using the Plücker line parameterization and subsequently derive a series of reprojection error formulations. We claim that the proposed reprojection error formulations are efficient and exhibit the degeneration-resistant ability. We also provide complete degeneration-resistant proof of our proposed error formulation. Our contributions are summarized as follows:", + "bbox": [ + 212, + 409, + 784, + 513 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To the best of our knowledge, this is the first line-based RSBA solution, which serves as the foundation for line-based or point-line-hybrid RS-SfM and RS-SLAM.", + "- we theoretically and experimentally demonstrate that the proposed $RSLBA$ can prevent three common degeneracies in RSBA, one of the degenerate cases is first discovered in this paper.", + "- The extensive evaluations in both synthetic and real datasets exhibit the comparable efficiency and accuracy of the proposed method over previous works." + ], + "bbox": [ + 223, + 522, + 784, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Background", + "text_level": 1, + "bbox": [ + 215, + 688, + 367, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we review the rolling shutter projection model in Sec. 2.1 and the parameterization of 3D lines in Sec. 2.2.", + "bbox": [ + 212, + 719, + 782, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Rolling Shutter Camera Projection Mode", + "text_level": 1, + "bbox": [ + 214, + 770, + 604, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathbf{R}(v) \\in \\mathbf{SO}(3)$ and $\\mathbf{t}(v) \\in \\mathbb{R}^3$ represent the camera rotation and translation, respectively, when the row index $v$ of measurement is acquired. Denote the intrinsic camera parameters as $\\mathbf{K}$ , and the initial rotation and translation of the", + "bbox": [ + 212, + 794, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "RSL-BA", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "camera at $v = 0$ as $\\mathbf{R}_0$ and $\\mathbf{t}_0$ , respectively. The projection matrix of the global shutter camera is defined as $\\mathbf{P}_0 = \\mathbf{K}[\\mathbf{R}_0, \\mathbf{t}_0]$ . Given the assumption of constant camera motion during frame capture, which is commonly made in RS 3D vision tasks [1,4,13], we can model the instantaneous motion as:", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R} (v) = (\\mathbf {I} + [ \\boldsymbol {\\omega} ] _ {\\times} v) \\mathbf {R} _ {0}, \\quad \\mathbf {t} (v) = \\mathbf {t} _ {0} + \\mathbf {d} v, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 348, + 223, + 785, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{d} = [d_x, d_y, d_z]^\\top$ is the translational velocity vector and $\\boldsymbol{\\omega} = [\\omega_x, \\omega_y, \\omega_z]^\\top$ is the rotational velocity vector, and $[\\boldsymbol{\\omega}]_\\times$ represents the skew-symmetric matrix of vector $\\boldsymbol{\\omega}$ . Then, the projection matrix of the camera under the RS model can be defined as:", + "bbox": [ + 212, + 253, + 784, + 313 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {v} = \\mathbf {K} \\left[ \\mathbf {R} _ {v}, \\mathbf {t} _ {v} \\right] = \\mathbf {P} _ {0} + v \\mathbf {Q} \\\\ \\mathbf {K} ^ {[ r ]} = \\left[ \\begin{array}{l l l l l} \\mathbf {P} _ {v} & \\mathbf {t} _ {v} & \\mathbf {Q} _ {v} & \\mathbf {I} ^ {[ r ]} & \\mathbf {D} _ {v} \\end{array} \\right] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 316, + 785, + 339 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {0} = \\mathbf {K} \\left[ \\mathbf {R} _ {0}, \\mathbf {t} _ {0} \\right] \\quad \\mathbf {Q} = \\mathbf {K} \\left[ \\left[ \\boldsymbol {\\omega} \\right] _ {\\times} \\mathbf {R} _ {0}, \\mathbf {d} \\right].\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 335, + 630, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 3D Line Segments Parameterization", + "text_level": 1, + "bbox": [ + 215, + 380, + 558, + 395 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We employ the Plücker matrix to parameterize the 3D line [7]. Specifically, given two homogeneous 3D points $P_{a} = [a_{1}, a_{2}, a_{3}, 1]^{\\top}$ and $P_{b} = [b_{1}, b_{2}, b_{3}, 1]^{\\top}$ . The Plücker matrix $\\mathbf{L}$ can be defined as:", + "bbox": [ + 214, + 414, + 785, + 458 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {L} = P _ {a} P _ {b} ^ {\\top} - P _ {b} P _ {a} ^ {\\top}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 470, + 751, + 488 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n= \\left[ \\begin{array}{c c c c} 0 & a _ {1} b _ {2} - a _ {2} b _ {1} & a _ {1} b _ {3} - a _ {3} b _ {1} & a _ {1} - b _ {1} \\\\ a _ {2} b _ {1} - a _ {1} b _ {2} & 0 & a _ {2} b _ {3} - a _ {3} b _ {2} & a _ {2} - b _ {2} \\\\ a _ {3} b _ {1} - a _ {1} b _ {3} & a _ {3} b _ {2} - a _ {2} b _ {3} & 0 & a _ {3} - b _ {3} \\\\ b _ {1} - a _ {1} & b _ {2} - a _ {2} & b _ {3} - a _ {3} & 0 \\end{array} \\right] = \\left[ \\begin{array}{c c c c} 0 & l _ {1 2} & l _ {1 3} & l _ {1 4} \\\\ - l _ {1 2} & 0 & l _ {2 3} & l _ {2 4} \\\\ - l _ {1 3} & - l _ {2 3} & 0 & l _ {3 4} \\\\ - l _ {1 4} & - l _ {2 4} & - l _ {3 4} & 0 \\end{array} \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 492, + 784, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the $[l_{14}, l_{24}, l_{34}]^{\\top}$ represents the direction of this straight line while the $[-l_{23}, l_{13}, -l_{12}]^{\\top}$ represents the normal direction of this line. The projection of 3D Plücker matrix is the skew-symmetric matrix of a 2D normalized line. It can be related to each other as:", + "bbox": [ + 214, + 569, + 787, + 630 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\mathbf {l} _ {g s} \\right] _ {\\times} = \\left[ \\mathbf {R}, \\mathbf {t} \\right] \\mathbf {L} \\left[ \\mathbf {R}, \\mathbf {t} \\right] ^ {\\top} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 421, + 646, + 784, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 693, + 380, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Sec. 3.1, we initially establish the rolling shutter projection equations using the Plücker line parameterization. These equations establish a connection between the 3D lines and the curve parameters on the image plane. Then in Sec. 3.2, we present a series of line-based reprojection error equations. On top of the error definition, we can formulate the entire rolling shutter line bundle adjustment problem in Sec. 3.3. Finally, in Sec. 3.4, we demonstrate the degeneracy resilience capability of the proposed error equations.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Zhang, B. Liao et al.", + "bbox": [ + 271, + 114, + 431, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Rolling Shutter Line Projection Formulation", + "text_level": 1, + "bbox": [ + 215, + 146, + 629, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As previously explained, the curved rolling shutter projection of 3D lines occurs because of the camera's nonlinear motion. Using the Plücker parameterization method, we can derive explicit expressions for the curve parameters.", + "bbox": [ + 212, + 171, + 782, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Firstly, we construct the skew-symmetric matrix $\\left[\\mathbf{l}_{rs}\\right]_{\\times}$ of the instantaneous projection line by combining Eq. (2) and Eq. (4):", + "bbox": [ + 212, + 217, + 784, + 248 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left[ \\mathbf {l} _ {r s} \\right] _ {\\times} = \\mathbf {P} _ {v} \\mathbf {L} \\mathbf {P} _ {v} ^ {\\top} \\\\ = \\mathbf {P} _ {0} \\mathbf {L} \\mathbf {P} _ {0} ^ {\\top} + v \\left(\\mathbf {P} _ {0} \\mathbf {L} \\mathbf {Q} ^ {\\top} + \\mathbf {Q} \\mathbf {L} \\mathbf {P} _ {0} ^ {\\top}\\right) + v ^ {2} \\mathbf {Q} \\mathbf {L} \\mathbf {Q} ^ {\\top} \\tag {5} \\\\ = \\mathbf {A} _ {1} + v \\mathbf {A} _ {2} + v ^ {2} \\mathbf {A} _ {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 258, + 784, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Secondly, as $\\mathbf{l}_{rs} = [l_1,l_2,l_3]^\\top$ and $[\\mathbf{l}_{rs}]_{\\times} = \\mathbf{A}_1 + v\\mathbf{A}_2 + v^2\\mathbf{A}_3$ , we can determine the instantaneous projected line parameter as follows:", + "bbox": [ + 212, + 329, + 782, + 359 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} l _ {1} = \\mathbf {A} _ {1} ^ {3 2} + v \\mathbf {A} _ {2} ^ {3 2} + v ^ {2} \\mathbf {A} _ {3} ^ {3 2} \\\\ l _ {2} = \\mathbf {A} _ {1} ^ {1 3} + v \\mathbf {A} _ {2} ^ {1 3} + v ^ {2} \\mathbf {A} _ {3} ^ {1 3} \\\\ l _ {3} = \\mathbf {A} _ {1} ^ {2 1} + v \\mathbf {A} _ {2} ^ {2 1} + v ^ {2} \\mathbf {A} _ {3} ^ {2 1} \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 372, + 784, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ultimately, we note that the instantaneous projected line parameter fulfills this $l_{1}u + l_{2}v + l_{3} = 0$ equation, which we can expand to obtain the final curve equation, denoted as $\\pmb{\\iota}$ :", + "bbox": [ + 212, + 429, + 782, + 474 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\iota : \\mathbf {A} _ {3} ^ {1 3} v ^ {3} + \\mathbf {A} _ {3} ^ {3 2} u v ^ {2} + \\left(\\mathbf {A} _ {2} ^ {1 3} + \\mathbf {A} _ {3} ^ {2 1}\\right) v ^ {2} + \\mathbf {A} _ {2} ^ {3 2} u v \\\\ + \\left(\\mathbf {A} _ {1} ^ {1 3} + \\mathbf {A} _ {2} ^ {2 1}\\right) v + \\mathbf {A} _ {1} ^ {3 2} u + \\mathbf {A} _ {1} ^ {2 1} = 0 \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 484, + 784, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our conclusion aligns with that of Lao et al. [11], yielding an identical polynomial curve degree. Nevertheless, our utilization of a 4-DoF orthogonal representation representation for 3D lines has distinct benefits when compared to Lao's 5-DoF over-parameterized representation. These advantages include the capability to conduct unconstrained optimization on spatial lines directly.", + "bbox": [ + 212, + 535, + 782, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark. The geometric interpretation underlying this derivation is quite straightforward. Given the parameters of a spatial line and the current camera pose, if point $(\\mathbf{u},\\mathbf{v})$ lies on the curve obtained by projecting this spatial line, then when the RS camera scans to the v-th row, the projection of the spatial line onto the image plane follows Eq. (6). Therefore, it is evident that point $(\\mathbf{u},\\mathbf{v})$ must be located on this line. Based on this reasoning, we can deduce Eq. (7).", + "bbox": [ + 212, + 611, + 784, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Line-based Reprojection Errors", + "text_level": 1, + "bbox": [ + 215, + 724, + 522, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As claimed in Sec. 3.1, a 3D straight line will become a polynomial curve on the captured image. In this section, we will establish a series of re-projection errors which are stable and robust. We first leverage the similar intuition in Sec. 3.1 to obtain the distance error representation. As shown in Fig. 2(a), Let there be a point $\\mathbf{q} = [u,v]^{\\top}$ on the observed curve $\\iota$ , then the pose of the camera when exposing that point is $\\mathbf{R}_v$ , $\\mathbf{t}_v$ , then projecting the 3D line $\\mathbf{L}$ onto the image", + "bbox": [ + 212, + 750, + 784, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "RSL-BA", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4f4d340e18b248320f6904fb07ad6a461bb1487e0d9e246341b03ccbf67c238a.jpg", + "image_caption": [ + "Fig. 2: The illustration of proposed line-based reprojection errors given the spatial line $L$ and observed curve $\\iota$ . Instead of directly measuring the distance between the observed curve and the projected curve, we split each curve into multiple points $q$ , and aggregate the distance errors. Specifically, we first utilize the virtual projected line $l$ when observing it, as shown in (a). To measure the distance between each point $q$ and the corresponding virtual lines, we can use the perpendicular distance (b), the horizontal distance (c,d) or the tangent distance (e) as our basic metrics." + ], + "image_footnote": [], + "bbox": [ + 230, + 143, + 776, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "plane, this results in the virtual 2D line $\\mathbf{1} = [l_1, l_2, l_3]^\\top$ . Under the condition of ground truth, this virtual line must theoretically intersect with that point $\\mathbf{q}$ . This constraint directly leads us to the two variants of distance error, which we have defined below.", + "bbox": [ + 214, + 429, + 787, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Perpendicular Distance Error $\\mathbf{e}_{d1}$ . The first distance error formulation, named perpendicular distance error, measures the shortest distance from point $\\mathbf{q}$ to the line $\\mathbf{l}$ , as shown in Fig. 2(b). The error is defined as:", + "bbox": [ + 214, + 515, + 787, + 561 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {e} _ {d 1} = \\frac {l _ {1} u + l _ {2} v + l _ {3}}{\\sqrt {l _ {1} ^ {2} + l _ {2} ^ {2}}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 423, + 571, + 785, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the line parameters $[l_1, l_2, l_3]$ can be calculated through Eq. (6).", + "bbox": [ + 214, + 617, + 723, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Horizontal Distance Error $\\mathbf{e}_{d2}$ . The second distance error formulation, named horizontal distance error, measures the u-axis distance from point $\\mathbf{q}$ to the line $\\mathbf{l}$ , as shown in Fig. 2(c). The error can be defined as:", + "bbox": [ + 214, + 657, + 787, + 704 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {e} _ {d 2} = u - u ^ {\\prime}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 450, + 715, + 785, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the corresponding $u'$ coordinate of virtual line 1 at the $v$ -th row can be calculated as:", + "bbox": [ + 214, + 744, + 787, + 772 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nu ^ {\\prime} = - \\frac {l _ {2} v + l _ {3}}{l _ {1}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 442, + 772, + 785, + 804 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Fig. 2(d), the $e_{d2}$ distance error is equivalent to the u-axis distance from the point to the projection curve.", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Zhang, B. Liao et al.", + "bbox": [ + 271, + 114, + 431, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tangent Error $\\mathbf{e}_t$ . While distance error provides an elegant constraint, it is insufficient to rely solely on distance error in certain scenarios. To further enhance the robustness, we introduce a second type of tangent error. As shown in Fig. 2(e), we constrain the tangent of the projected curve $\\iota'$ to match the observed tangent direction. The error can be defined as:", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {e} _ {t} = \\mathbf {s} ^ {\\top} \\mathbf {s} ^ {\\prime}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 465, + 233, + 785, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{s}$ represents the tangent direction observation. The $\\mathbf{s}'$ can be calculated analytically following the implicit function theorem as:", + "bbox": [ + 214, + 262, + 782, + 292 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} ^ {\\prime} = \\left[ \\frac {1}{\\sqrt {1 + s ^ {2}}}, \\frac {s}{\\sqrt {1 + s ^ {2}}} \\right], \\quad s = \\frac {\\partial u}{\\partial v} = - \\frac {\\partial \\iota / \\partial v}{\\partial \\iota / \\partial u} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 304, + 785, + 338 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Rolling Shutter Line Bundle Adjustment", + "text_level": 1, + "bbox": [ + 214, + 356, + 598, + 372 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The non-linear least squares solvers are used to find an optimal solution $\\pmb{\\theta}^{*}$ including camera poses $\\mathbf{R}^{*},\\mathbf{t}^{*}$ , instantaneous motion $\\omega^{*},\\mathbf{d}^{*}$ and 3D lines $\\pmb{\\tau}^{*}$ by minimizing the reprojection error $\\mathbf{e}_i^j$ from line $i$ to camera $j$ over all the camera index in set $\\mathcal{F}$ and corresponding observed 3D lines index in subset $\\mathcal{P}_j$ :", + "bbox": [ + 212, + 382, + 782, + 443 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\theta} ^ {*} = \\left\\{\\boldsymbol {\\tau} ^ {*}, \\mathbf {R} ^ {*}, \\mathbf {t} ^ {*}, \\boldsymbol {\\omega} ^ {*}, \\mathbf {d} ^ {*} \\right\\} = \\arg \\min _ {\\boldsymbol {\\theta}} \\sum_ {j \\in \\mathcal {F}} \\sum_ {i \\in \\mathcal {P} _ {j}} \\left\\| \\mathbf {e} _ {i} ^ {j} \\right\\| _ {2} ^ {2}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 320, + 455, + 785, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On top of errors proposed in Sec. 3.2, we provide two variants of reprojection error:", + "bbox": [ + 212, + 503, + 782, + 532 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {e} _ {1} = \\mathbf {e} _ {d 1} + \\lambda_ {1} \\mathbf {e} _ {t}, \\quad \\mathbf {e} _ {2} = \\mathbf {e} _ {d 2} + \\lambda_ {2} \\mathbf {e} _ {t}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 535, + 784, + 550 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\lambda_1, \\lambda_2$ are the relative weight. We have experimentally validated the optimality of these two error variants in Sec. 4.1. After definition, this non-linear least square problem can be iteratively solved through Levenberg-Marquardt [19] method.", + "bbox": [ + 212, + 556, + 782, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Besides, the full analytical Jacobian has been derived in the supplementary material.", + "bbox": [ + 214, + 617, + 782, + 646 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Degeneracy Analysis", + "text_level": 1, + "bbox": [ + 214, + 670, + 433, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conducted theoretical analysis on the resilience of $RSL-BA$ to three types of degeneracy scenarios. These three degeneracy scenarios include plane degeneracy case [1], 2-vIEWS pure translation case [30], and X-Y pure translation degeneracy case disclosed by us. We assume the positive direction of X Y Z axis is right down forward, respectively.", + "bbox": [ + 212, + 695, + 782, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(1) Resistance to Plane Degeneracy [1]. We first briefly describe the conditions under which the plane degeneration occurs: when the camera's y-axis is parallel, if a rotational velocity $\\omega = [-1,0,0]^{\\top}$ is given to the camera, then the", + "bbox": [ + 212, + 794, + 782, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "RSL-BA", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "reconstructed 3D points will be compressed onto a plane perpendicular to the $y$ -axis, and satisfy that the reprojection error is 0.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For line features, we also assume $\\mathbf{R}_0 = \\mathbf{I}$ , $\\mathbf{t}_0 = \\mathbf{0}$ , $\\boldsymbol{\\omega} = [-1,0,0]^\\top$ , $\\mathbf{d} = 0$ . Let there be two points, $\\mathbf{P}_a$ and $\\mathbf{P}_b$ , on the spatial line $\\mathbf{L}$ . $\\mathbf{P}_a = [a_1,a_2,a_3,1]^\\top$ and $\\mathbf{P}_b = [b_1,b_2,b_3,1]^\\top$ . When the space is compressed into a plane, the new coordinates of $\\mathbf{P}_a$ and $\\mathbf{P}_b$ are $\\mathbf{P}_a' = [a_1,0,a_3,1]^\\top$ and $\\mathbf{P}_b' = [b_1,0,b_3,1]^\\top$ , respectively. By referring to Sec. 3.1, we can solve for the projection parameters of the two-dimensional curve at this moment. The parameters of the curve at this time are:", + "bbox": [ + 212, + 176, + 784, + 280 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\mathbf {A} _ {3} ^ {1 3} = \\mathbf {A} _ {3} ^ {3 2} = \\mathbf {A} _ {2} ^ {1 3} = \\mathbf {A} _ {3} ^ {2 1} = \\mathbf {A} _ {2} ^ {3 2} = \\mathbf {A} _ {1} ^ {3 2} = \\mathbf {A} _ {1} ^ {2 1} = 0 \\\\ \\mathbf {A} _ {1} ^ {1 3} = a _ {3} b _ {1} - a _ {1} b _ {3} \\\\ \\mathbf {A} _ {2} ^ {2 1} = a _ {1} b _ {3} - a _ {3} b _ {1} \\end{array} \\right. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 282, + 784, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Substituting the parameters into the curve expression Eq. (7), we obtain $0 = 0$ , indicating that u and v are unrestricted. Under the influence of $\\omega = [-100]^{\\top}$ , the camera's imaging rows are always coplanar with the line, making the distance from any point on the image to the line zero and satisfying the point-to-line distance error function (Sec. 3.2). However, the slope constraint cannot be met. Using this projection equation to calculate the tangent direction of any point on the curve results in indeterminate forms $[0/0, 0/0]$ , which do not match the measured tangent direction. Therefore, using the tangent direction of points on the curve as the error function prevents this degeneration. As verified in Fig 5, our proposed RSL-BA can suppress plane degeneration, whereas conventional rolling shutter point-based bundle adjustment method NMRsBA [1] converges to a degenerate solution.", + "bbox": [ + 212, + 340, + 787, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(2) Resistance to 2-views Pure Translation Degeneracy [30]. We first briefly describe the conditions under which this degeneration occurs: when the camera moves in a straight line to take two pictures, assigning any translational velocity to the camera for both shots, it is possible to find new 3D points that satisfy the condition that the reprojection error of the points under this velocity is zero.", + "bbox": [ + 212, + 550, + 782, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Let's assume the real pose of the camera before the degeneration is: $\\mathbf{R}_0 = \\mathbf{I}$ , $\\mathbf{t}_0 = \\mathbf{0}$ , $\\boldsymbol{\\omega} = [0,0,0]^\\top$ , $\\mathbf{d} = [d_1,d_2,d_3]^\\top$ , $\\mathbf{L} = (l_{12},l_{13},l_{23},l_{14},l_{24},l_{34})$ In this case, substitute these parameters into Eq. (7) we can get curve expression:", + "bbox": [ + 212, + 642, + 782, + 688 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\left(l _ {1 4} d _ {3} - l _ {3 4} d _ {1}\\right) v ^ {2} + \\left(l _ {3 4} d _ {2} - l _ {2 4} d _ {3}\\right) u v + \\left(l _ {2 4} d _ {1} - l _ {1 4} d _ {2} + l _ {1 3}\\right) v - l _ {2 3} u - l _ {1 2} = 0 \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 232, + 702, + 784, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "After the degeneration occurs, $\\mathbf{R}_0^{\\prime} = \\mathbf{I}$ , $\\mathbf{t}_0^{\\prime} = \\mathbf{0}$ , $\\boldsymbol{\\omega}^{\\prime} = [0\\quad 0\\quad 0]^\\top$ , $\\mathbf{d}' = r[d_1,d_2,d_3]$ , $\\mathbf{L} = (l_{12}',l_{13}',l_{23}',l_{14}',l_{24}',l_{34}')$ . In this case, the curve expression is:", + "bbox": [ + 212, + 733, + 789, + 763 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nr \\left(l _ {1 4} ^ {\\prime} d _ {3} - l _ {3 4} ^ {\\prime} d _ {1}\\right) v ^ {2} + r \\left(l _ {3 4} ^ {\\prime} d _ {2} - l _ {2 4} ^ {\\prime} d _ {3}\\right) u v + \\left(r \\left(l _ {2 4} ^ {\\prime} d _ {1} - l _ {1 4} ^ {\\prime} d _ {2}\\right) + l _ {1 3} ^ {\\prime}\\right) v - l _ {2 3} ^ {\\prime} u - l _ {1 2} ^ {\\prime} = 0 \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 777, + 799, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To make the reprojection error zero, the curve expressions before and after degeneration must be the same, meaning the corresponding coefficients are in", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Zhang, B. Liao et al.", + "bbox": [ + 271, + 114, + 431, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d28cd589b623bfcde50423900425887091968c032fa33d960b982eecce6e6752.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 282, + 156, + 395, + 258 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0408ba474de9c4e0afae8cf88006c86e329f3306c9b0f8d2e32593c709f8dec1.jpg", + "image_caption": [ + "(b)", + "Fig. 3: The illustration of degeneracy resistance of our proposed method in 2-views pure translation degeneracy. (a) Degeneration in single-view scenarios. (b) Suppression of degeneration using line features in dual-view scenarios. (c) Inability of line features to suppress degeneration in special dual-view scenarios" + ], + "image_footnote": [], + "bbox": [ + 413, + 154, + 570, + 258 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b0fc43be8c524d8d4afeaa57a2ba0871d77002a78e9187ed1d9a53ccbad4c942.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 578, + 151, + 717, + 258 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "proportion.", + "bbox": [ + 215, + 383, + 299, + 397 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} l _ {1 4} d _ {3} - l _ {3 4} d _ {1} = s r \\left(l _ {1 4} ^ {\\prime} d _ {3} - l _ {3 4} ^ {\\prime} d _ {1}\\right) \\\\ l _ {3 4} d _ {2} - l _ {2 4} d _ {3} = s r \\left(l _ {3 4} ^ {\\prime} d _ {2} - l _ {2 4} ^ {\\prime} d _ {3}\\right) \\\\ l _ {2 4} d _ {1} - l _ {1 4} d _ {2} + l _ {1 3} = s r \\left(l _ {2 4} ^ {\\prime} d _ {1} - l _ {1 4} ^ {\\prime} d _ {2}\\right) + s l _ {1 3} ^ {\\prime} \\\\ l _ {2 3} = s l _ {2 3} ^ {\\prime} \\\\ l _ {1 2} = s l _ {1 2} ^ {\\prime} \\end{array} \\right. \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 407, + 784, + 486 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Substituting the first and second equations into the third equation yields $l_{13} = sl_{13}'$ . Combining this with:", + "bbox": [ + 214, + 496, + 784, + 526 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} l _ {1 2} l _ {3 4} - l _ {1 3} l _ {2 4} + l _ {1 4} l _ {2 3} = 0 \\\\ l _ {1 2} ^ {\\prime} l _ {3 4} ^ {\\prime} - l _ {1 3} ^ {\\prime} l _ {2 4} ^ {\\prime} + l _ {1 4} ^ {\\prime} l _ {2 3} ^ {\\prime} = 0 \\end{array} \\right. \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 537, + 784, + 570 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "we obtain:", + "bbox": [ + 215, + 580, + 294, + 593 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nl _ {1 2} ^ {\\prime} = \\frac {l _ {1 2} ^ {\\prime}}{s}, \\quad l _ {1 3} ^ {\\prime} = \\frac {l _ {1 3}}{s}, \\quad l _ {2 3} ^ {\\prime} = \\frac {l _ {2 3}}{s}, \\quad l _ {1 4} ^ {\\prime} = \\frac {l _ {1 4}}{s r}, \\quad l _ {2 4} ^ {\\prime} = \\frac {l _ {2 4}}{s r}, \\quad l _ {3 4} ^ {\\prime} = \\frac {l _ {3 4}}{s r} \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 243, + 604, + 784, + 626 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Note that the direction and normal direction of the line remain unchanged, indicating that the degenerated line stays in the original plane and parallel to the original line, as shown in Fig. 3(1), where $\\mathbf{L}^{gt}$ is the original line, and $\\mathbf{L}_1^{dg}$ is the degenerated line. In the dual-frame scenario (Fig.3(2)), when the plane $\\mathbf{O}_1\\mathbf{L}^{gt}$ is not coplanar with the plane $\\mathbf{O}_2\\mathbf{L}^{gt}$ , their degenerated states $\\mathbf{L}_1^{dg}$ and $\\mathbf{L}_2^{gt}$ cannot coincide (except at $\\mathbf{L}^{gt}$ ). In this case, line constraints can prevent degeneration. However, when $\\mathbf{L}^{gt}$ , $\\mathbf{O}_1$ , and $\\mathbf{O}_2$ are coplanar (Fig.3(3)), line constraints cannot suppress degeneration, though this is a very special condition requiring all lines to meet this criterion. This demonstrates that line constraints are more effective than point constraints in preventing degeneration.", + "bbox": [ + 212, + 635, + 782, + 787 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "(3) Resistance to X-Y Pure Translation Degeneracy (our new founding). We present a pure translation scenario without the z-direction, suitable for", + "bbox": [ + 214, + 809, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "RSL-BA", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/fc83d709053f01e4ee4124c0fa43a052c89b57a4c14e044d1c43d3d9fbbd21fc.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 285, + 151, + 395, + 241 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/24b87dcae962ef715b22a93d42baa8cc525ad05238c265e9c7b49408574d10fa.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 398, + 162, + 500, + 237 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/deef5e9ffb831fb95b02fb72e98086c7816d501055585862ea7cdfd5e54f60e7.jpg", + "image_caption": [ + "(c)", + "Fig. 4: In the case of X-Y pure translation, optimization compresses points into a straight line. In a single view, pixel points parallel to the motion trajectory backproject to the same 3D point ((a) and (b)). (c) Different intersection points in space are equidistant from the camera plane. (d) All intersection points formed by backprojection constitute a straight line." + ], + "image_footnote": [], + "bbox": [ + 500, + 155, + 602, + 242 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/768fe78bc1be96a55cf58fc83086a3e73383f5796a3bbe297b54c3ccf6aaa99a.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 609, + 152, + 720, + 241 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "multi-view BA, compressing space points into a straight line. Unlike the second type of degeneration, this does not depend on the number of images.", + "bbox": [ + 212, + 369, + 784, + 398 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Firstly, considering a single view, as shown in Fig.4(1), during exposure, the camera moves from position $\\mathbf{O}_1^1$ to $\\mathbf{O}_1^3$ , where pixel $\\mathbf{q}_1^1$ is located on the first row and $\\mathbf{q}_1^3$ is located on the $H$ -th row. Their corresponding spatial points are located on the rays $\\mathbf{O}_1^1\\mathbf{q}_1^1$ and $\\mathbf{O}_1^3\\mathbf{q}_1^3$ , intersecting at point $\\mathbf{P}_1$ . Actually, any point on the line $\\mathbf{q}_1^1\\mathbf{q}_1^3$ intersects with its corresponding 3D point at $\\mathbf{P}_1$ during exposure. Let $\\mathbf{O}_1^2$ be the camera position when exposed to the $h_1^2$ -th row, with corresponding feature point $\\mathbf{q}_1^2$ . We only consider the plane $\\mathbf{O}_1^1\\mathbf{P}_1\\mathbf{O}_1^3$ in Fig.4(2). The motion of $\\mathbf{O}$ is uniform rectilinear motion thus, there exists a length relationship:", + "bbox": [ + 212, + 398, + 787, + 521 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{c} l _ {\\mathbf {q} _ {1} ^ {1} \\mathbf {q} _ {1} ^ {2}} = \\frac {h _ {1} ^ {2}}{H} l _ {\\mathbf {q} _ {1} ^ {1} \\mathbf {q} _ {1} ^ {3}}, \\quad l _ {\\mathbf {q} _ {3} ^ {1} \\mathbf {q} _ {1} ^ {2}} = \\left(1 - \\frac {h _ {1} ^ {2}}{H}\\right) l _ {\\mathbf {q} _ {1} ^ {1} \\mathbf {q} _ {1} ^ {3}} \\\\ l _ {\\mathbf {O} _ {1} ^ {1} \\mathbf {O} _ {1} ^ {2}} = \\frac {h _ {1} ^ {2}}{H} l _ {\\mathbf {O} _ {1} ^ {1} \\mathbf {O} _ {1} ^ {3}}, \\quad l _ {\\mathbf {O} _ {3} ^ {1} \\mathbf {O} _ {1} ^ {2}} = \\left(1 - \\frac {h _ {1} ^ {2}}{H}\\right) l _ {\\mathbf {O} _ {1} ^ {1} \\mathbf {O} _ {1} ^ {3}} \\end{array} \\right. \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 527, + 784, + 571 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "so we have:", + "bbox": [ + 214, + 577, + 302, + 590 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {l _ {\\mathbf {q} _ {1} ^ {1} \\mathbf {q} _ {1} ^ {2}}}{l _ {\\mathbf {O} _ {1} ^ {1} \\mathbf {O} _ {1} ^ {2}}} = \\frac {l _ {\\mathbf {q} _ {1} ^ {3} \\mathbf {q} _ {1} ^ {2}}}{l _ {\\mathbf {O} _ {3} ^ {1} \\mathbf {O} _ {1} ^ {2}}} = \\frac {l _ {\\mathbf {q} _ {1} ^ {1} \\mathbf {q} _ {1} ^ {3}}}{l _ {\\mathbf {O} _ {1} ^ {1} \\mathbf {O} _ {1} ^ {3}}} \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 585, + 784, + 625 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Therefore, the ray $\\mathbf{O}_1^2\\mathbf{q}_1^2$ will also intersect at point $\\mathbf{P_1}$ . For every pixel on a line parallel to $\\mathbf{O}_1^1\\mathbf{O}_1^3$ , it can be traced back to a different spatial point. For pixel points on different lines parallel to the motion trajectory, as shown in Fig. 4(3), because the distance $l_{q^1 q^3}$ is the same and the camera focal length remains unchanged, according to similar triangles, these spatial points have equal distances to the image plane. Simultaneously, due to symmetry, these points also lie in the plane that is perpendicular to and bisects the line segment $\\mathbf{O}_1^1\\mathbf{O}_1^2$ . Therefore, the intersection line of these two planes represents the final degenerated configuration, which is a line parallel to the x-axis, as shown in Fig.4(4). In multiple views, with no z-direction movement, the imaging planes of different cameras coincide. We assign the same camera configuration to different views, including rotation, translation, angular velocity, and linear velocity. From the single-view case, corresponding points on the pixel plane parallel to the motion trajectory reconstruct to the same spatial point, with zero reprojection error.", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Zhang, B. Liao et al.", + "bbox": [ + 271, + 114, + 431, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/2a7dacdd8b01229a62e27383d3123c2cb57b2ca4b68e4fa1b139f1c753c76f63.jpg", + "image_caption": [ + "Fig.5: Comparison of degenerate resistance ability between NMRSBA [1] and proposed $RSL-BA$ . Ground truth camera poses and structure are colored with black. This example illustrates that our proposed $RSL-BA$ has the resistance ability against the plane and X-Y pure translation degeneracy." + ], + "image_footnote": [], + "bbox": [ + 308, + 146, + 697, + 291 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/9cfe949e393723ebaef0030e53b19349f3de2f40233bede2382dd9015eb1b8fa.jpg", + "image_caption": [ + "Fig. 6: Accuracy and computational time comparison of the proposed two combinations of reprojection errors." + ], + "image_footnote": [], + "bbox": [ + 259, + 376, + 750, + 445 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For the line constraints proposed in this paper, $RSL-BA$ can still effectively suppress degeneration. The proof process is similar to the previous section Sec. 3.4, with the only modification being setting $d_{3} = 0$ in the camera's motion direction. As verified in Fig 5, our proposed $RSL-BA$ can suppress X-Y pure translation degeneration, whereas conventional rolling shutter point-based bundle adjustment method NMRSBA [1] converges to a degenerate solution.", + "bbox": [ + 212, + 513, + 787, + 606 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 630, + 375, + 647 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We compare our method with two SOTA GS-based-method: 1) GSBA [17], 2) GLBA [25], and four SOTA RS-based-method: 1) MRSBA [5], 2) NMRSBA [1], 3) WMRSBA [14], 4) NWRSBA [14]. The experiments are conducted on a laptop with an Intel i7 CPU and all algorithms are implemented in MATLAB.", + "bbox": [ + 212, + 662, + 787, + 724 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation metrics. The accuracy measures employed encompass standard metrics such as rotation error, translation error, and algorithm efficiency. When conducting experiments related to line BA, we also incorporate error metrics that assess the accuracy of line reconstruction. Given the ground truth rotation $\\mathbf{R}_g$ , translation $\\mathbf{t}_g$ , and space line $\\mathbf{L}_g = (\\mathbf{n}_g^\\top, \\mathbf{a}_g^\\top)^\\top$ with an arbitrary point $\\mathbf{P}_g$ on it. Similarly, the optimized parameters are denoted as $\\mathbf{R}_d$ , $\\mathbf{t}_d$ , $\\mathbf{L}_d = (\\mathbf{n}_d^\\top, \\mathbf{a}_d^\\top)^\\top$ ,", + "bbox": [ + 212, + 747, + 787, + 842 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "RSL-BA", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/09e6b95d51c0b51709c85ef14adc60c63605de8df6f01b50a30b9d4cf41cd301.jpg", + "table_caption": [ + "Table 1: Robustness analysis against multiple levels of Gaussian noise (px)." + ], + "table_footnote": [], + "table_body": "
noise level (px)0.10.51.01.52.0
Rotation Error7.96e-69.24e-61.50e-51.71e-52.71e-5
Translation Error2.32e-45.93e-48.91e-49.54e-41.55e-3
L-R Error3.36e-33.83e-35.14e-35.22e-36.42e-3
L-D Error2.22e-42.43e-43.90e-44.43e-44.75e-4
", + "bbox": [ + 303, + 170, + 697, + 250 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/728eb8ed083f062042a63e82625d599e34fecd599cc267afe8ae7ef11a5918d0.jpg", + "image_caption": [ + "Fig. 7: Comparison of accuracy and time among different algorithms." + ], + "image_footnote": [], + "bbox": [ + 305, + 260, + 699, + 401 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "$\\mathbf{P}_d$ , respectively. The evaluation metrics are defined as:", + "bbox": [ + 215, + 454, + 612, + 469 + ], + "page_idx": 11 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{c} e _ {\\text {R o t a t i o n}} = \\operatorname {a r c c o s} \\left(\\frac {\\operatorname {T r a c e} \\left(\\mathbf {R} _ {g} ^ {\\top} \\mathbf {R} _ {d}\\right) - 1}{2}\\right) \\\\ e _ {\\text {T r a n s l a t i o n}} = \\operatorname {a r c c o s} \\left(\\mathbf {t} _ {g} ^ {\\top} \\mathbf {t} _ {d} / | | \\mathbf {t} _ {g} | | \\cdot | | \\mathbf {t} _ {d} |\\right) \\\\ e _ {L - R} = \\operatorname {a r c c o s} \\left(\\mathbf {a} _ {g} ^ {\\top} \\mathbf {a} _ {d} / | | \\mathbf {a} _ {g} | | \\cdot | | \\mathbf {a} _ {d} |\\right) \\\\ e _ {L - D} = \\frac {| (\\mathbf {a} _ {g} \\times \\mathbf {a} _ {d}) (\\mathbf {P} _ {d} - \\mathbf {P} _ {g}) |}{\\| \\mathbf {a} _ {g} \\times \\mathbf {a} _ {d} \\|} \\end{array} \\right. \\tag {23}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 477, + 785, + 553 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.1 Synthetic Experiments", + "text_level": 1, + "bbox": [ + 215, + 570, + 450, + 585 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Which error term in Sec. 3.2 is better? In the synthetic environment, We set up a cubic box with 12 edges in 3D space and simulated a series of cameras around it to perform RS imaging. The experimental results are shown in Fig. 6. It is evident that, the first type of reprojection error combination performs much better than the second combination in terms of accuracy and time. Therefore, in the forthcoming experiments, we shall exclusively employ the first type of reprojection error combination for comparison.", + "bbox": [ + 212, + 592, + 787, + 699 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "How significant is the impact of noise on RSL-BA? Following the same synthetic experimental setup as described above, we additionally introduce Gaussian noise to the endpoints of each line in this experiment. The noise levels are varied from 0.1 pixels to 2.0 pixels, and four different types of errors are recorded for the algorithm at each noise level. For each noise level, 50 experiments are conducted, and the median value is used as the final result. The results, presented in Table 1, demonstrate that RSL-BA provides stable estimations as the noise level increases from 0.1 pixels to 2 pixels.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Zhang, B. Liao et al.", + "bbox": [ + 271, + 114, + 431, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/f01781ca3e9ca6d900b40eede64a571453678cd30e22655036322b068ec05afd.jpg", + "table_caption": [ + "Table 2: The median absolute trajectory error (ATE) of different methods on WHU-RSVI [3] dataset. The best and second results are shown in green and blue, respectively." + ], + "table_footnote": [], + "table_body": "
WHU-RSVI1WHU-RSVI2WHU-RSVI3WHU-RSVI4
GSBA0.0809920.0613100.0304040.023698
GLBA0.0761730.0659850.0335540.024961
NMRSBA0.0509690.0416290.0423170.028183
NWRSBA0.0406400.0453130.0354510.022666
RLS-BA(ours)0.04435020.0393140.0233510.020675
", + "bbox": [ + 289, + 184, + 707, + 265 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "What are the advantages of $RSL-BA$ over point-based methods? Within this part, we will conduct a comparative analysis of our method alongside current SOTA methods, including point- and line-based methods. We set up a cubic box of points in the same positions as the 3D lines, totaling 56 points. To ensure fairness, we set up 8 lines with 7 points each to construct error functions, resulting in a total of precisely 56 points. Detailed experimental settings on the number of lines and the number of sampling points on each line are provided in the supplementary materials. We construct two scenarios: regular and classic scenarios with degeneration parallel to the y-axis. Due to significant differences in the experimental results of different methods, we uniformly take the logarithm (base 10) of the results for better observation. The experimental findings are shown in Fig. 7. It is evident that in both situations, our $RSL-BA$ and the current NWRSBA exhibit comparable accuracy, outperforming alternative approaches based on points and lines. Additionally, the computation time is equivalent to that of NWRSBA. This indicates that the application of lines on Rolling Shutter camera holds significant promise.", + "bbox": [ + 212, + 290, + 787, + 532 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.2 Synthetic Images", + "text_level": 1, + "bbox": [ + 215, + 550, + 405, + 566 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this section, we conduct experiments on input synthetic images. We use the WHU-RSVI [3] dataset, from which we select two sets of data from trajectory1-fast and trajectory2-fast for 3D reconstruction and pose estimation. We first employ [22] to detect RS curves by segmenting curves into multiple short-line segments and performing line fitting for initialization. The GS line-based SfM [16] is applied to initialize the $RSL-BA$ parameters. The comparative methods include GSBA, NMRSBA, NWRSBA, and GSLBA. Table 2 shows the median absolute trajectory error of different methods, it can be observed that the proposed $RSL-BA$ method is the most stable one, achieving optimal or near-optimal results in all cases. Qualitative comparison are also provided in Fig. 8. Unlike point-based methods, line-based methods often achieve good results with fewer feature lines. However, the GSL-BA method is not sufficiently stable when RS effects are prominent.", + "bbox": [ + 212, + 571, + 787, + 768 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Real Images", + "text_level": 1, + "bbox": [ + 215, + 787, + 364, + 803 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this section, we conduct experiments on the real image dataset TUM-RSVI [24]. The experimental setup and comparison methods are similar to those in Sec. 4.2.", + "bbox": [ + 212, + 809, + 790, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "RSL-BA", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/b461a3e0dd1311948f20d59dc9d60acc6d3f49d69bf934b4689915efd6bc231d.jpg", + "table_caption": [ + "Table 3: The absolute trajectory error (ATE) comparison of different methods in TUM-RSVI [24] dataset. The best and second results are shown in green and blue, respectively." + ], + "table_footnote": [], + "table_body": "
GSBAGLBANMRSBANWRSBARSL-BA(ours)
seq10.0694840.0864600.0523660.0450640.037816
seq20.0298210.0303790.0260280.0232270.025132
seq30.0651600.0637180.0579180.0550010.048187
seq40.0496130.0520260.0322140.0305340.031903
seq50.0318600.0358390.0194070.0160660.017659
seq60.0619660.0617920.0324340.0246580.026448
seq70.0516210.0565340.0391540.0396200.039701
seq80.0264030.0286900.0248070.0259260.024983
seq90.0983340.0982120.0824810.0735800.080523
seq100.811740.811800.593900.532960.57477
", + "bbox": [ + 313, + 198, + 679, + 340 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/e0d23ae6a4a6165cb387767c5b262fa2489d2272bb48cb228b4bb2e93ed384fa.jpg", + "image_caption": [ + "Fig. 8: Comparison of trajectories and 3D reconstruction on WHU-RSVI [3] and TUMRSVI [24] dataset. Each column represents a different bundle adjustment algorithm, and each row represents a different sequence." + ], + "image_footnote": [], + "bbox": [ + 310, + 356, + 692, + 545 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 3 and Fig 8 present some of the experimental results. As can be seen, in most cases, the $RSL-BA$ method outperforms other methods but is slightly weaker than NwRSBA. This is because the TUM-RSVI lacks line features and they are not visually prominent, making it less suitable for $RSL-BA$ .", + "bbox": [ + 212, + 633, + 787, + 696 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 717, + 359, + 734 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper presents the first solution of line-based RSL-BA. By utilizing points and tangent directions on curves, we have established a series of faster and more robust curve-to-line re-projection errors. The proposed RSL-BA method can prevent three common degeneracies in RSBA, one of which is newly introduced by us. Extensive experiments in real and synthetic data verify the effectiveness and efficiency of the proposed RSL-BA method.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Zhang, B. Liao et al.", + "bbox": [ + 271, + 114, + 431, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is supported by the National Key R&D Program of China (No. 2022ZD01190030), Nature Science Foundation of China (No. 62102145), Jiangxi Provincial 03 Special Foundation and 5G Program (Grant No. 20224ABC03A05), Lushan Lab Research Funding, and Changsha Science Fund for Distinguished Young Scholars (kq2306002).", + "bbox": [ + 212, + 176, + 787, + 253 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 277, + 321, + 292 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Albl, C., Sugimoto, A., Pajdla, T.: Degeneracies in rolling shutter sfm. In: ECCV (2016)", + "2. Bartoli, A., Sturm, P.: Structure-from-motion using lines: Representation, triangulation, and bundle adjustment. Computer vision and image understanding 100(3), 416-441 (2005)", + "3. Cao, L., Ling, J., Xiao, X.: The whu rolling shutter visual-inertial dataset. IEEE Access 8, 50771-50779 (2020)", + "4. Dai, Y., Li, H., Kneip, L.: Rolling shutter camera relative pose: generalized epipolar geometry. In: CVPR (2016)", + "5. Duchamp, G., Ait-Aider, O., Royer, E., Lavest, J.M.: A rolling shutter compliant method for localisation and reconstruction. In: VISAPP (2015)", + "6. Fisher, A., Cannizzaro, R., Cochrane, M., Nagahawatte, C., Palmer, J.L.: Colmap: A memory-efficient occupancy grid mapping framework. Robotics and Autonomous Systems 142, 103755 (2021)", + "7. Hartley, R., Zisserman, A.: Multiple view geometry in computer vision. Cambridge university press (2003)", + "8. Hedborg, J., Forssen, P.E., Felsberg, M., Ringaby, E.: Rolling shutter bundle adjustment. In: CVPR (2012)", + "9. Holynski, A., Geraghty, D., Frahm, J.M., Sweeney, C., Szeliski, R.: Reducing drift in structure from motion using extended features. In: 2020 International Conference on 3D Vision (3DV). pp. 51-60. IEEE (2020)", + "10. Ito, E., Okatani, T.: Self-calibration-based approach to critical motion sequences of rolling-shutter structure from motion. In: CVPR (2016)", + "1. Lao, Y., Ait-Aider, O.: A robust method for strong rolling shutter effects correction using lines with automatic feature selection. In: CVPR (2018)", + "2. Lao, Y., Ait-Aider, O., Araujo, H.: Robustified structure from motion with rolling-shutter camera using straightness constraint. Pattern Recognition Letters (2018)", + "3. Lao, Y., Ait-Aider, O., Bartoli, A.: Solving rolling shutter 3d vision problems using analogies with non-rigidity. IJCV (2021)", + "4. Liao, B., Qu, D., Xue, Y., Zhang, H., Lao, Y.: Revisiting rolling shutter bundle adjustment: Toward accurate and fast solution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4863-4871 (2023)", + "5. Lim, H., Jeon, J., Myung, H.: Unconstrained line-based slam using vanishing points for structural mapping. IEEE Robotics and Automation Letters 7(2), 1518-1525 (2022)", + "6. Liu, S., Yu, Y., Pautrat, R., Pollefeys, M., Larsson, V.: 3d line mapping revisited. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21445-21455 (2023)" + ], + "bbox": [ + 225, + 309, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "RSL-BA", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Lourakis, M.I., Argyros, A.A.: Sba: A software package for generic sparse bundle adjustment. ACM Transactions on Mathematical Software (TOMS) 36(1), 1-30 (2009)", + "18. Micusik, B., Wildenauer, H.: Structure from motion with line segments under relaxed endpoint constraints. International Journal of Computer Vision 124, 65-79 (2017)", + "19. Moré, J.J.: The levenberg-marquardt algorithm: implementation and theory. In: Numerical analysis: proceedings of the biennial Conference held at Dundee, June 28-July 1, 1977. pp. 105-116. Springer (2006)", + "20. Patron-Perez, A., Lovegrove, S., Sibley, G.: A spline-based trajectory representation for sensor fusion and rolling shutter cameras. IJCV (2015)", + "21. Pumarola, A., Vakhitov, A., Agudo, A., Sanfeliu, A., Moreno-Noguer, F.: Pl-slam: Real-time monocular visual slam with points and lines. In: 2017 IEEE international conference on robotics and automation (ICRA). pp. 4503-4508. IEEE (2017)", + "22. Purkait, P., Zach, C., Leonardis, A.: Rolling shutter correction in manhattan world. In: ICCV. pp. 882-890 (2017)", + "23. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4104-4113 (2016)", + "24. Schubert, D., Demmel, N., von Stumberg, L., Usenko, V., Cremers, D.: Rolling-shutter modelling for direct visual-inertial odometry. In: IROS (2019)", + "25. Taylor, C.J., Kriegman, D.J.: Structure and motion from line segments in multiple images. IEEE Transactions on Pattern Analysis and Machine Intelligence 17(11), 1021-1032 (1995)", + "26. Triggs, B., McLauchlan, P.F., Hartley, R.I., Fitzgibbon, A.W.: Bundle adjustment—a modern synthesis. In: Vision Algorithms: Theory and Practice: International Workshop on Vision Algorithms Corfu, Greece, September 21–22, 1999 Proceedings. pp. 298–372. Springer (2000)", + "27. Von Gioi, R.G., Jakubowicz, J., Morel, J.M., Randall, G.: Lsd: A fast line segment detector with a false detection control. IEEE transactions on pattern analysis and machine intelligence 32(4), 722-732 (2008)", + "28. Wei, D., Wan, Y., Zhang, Y., Liu, X., Zhang, B., Wang, X.: Elsr: Efficient line segment reconstruction with planes and points guidance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15807-15815 (2022)", + "29. Zhuang, B., Cheong, L.F., Lee, G.H.: Rolling-shutter-aware differential sfm and image rectification. In: ICCV (2017)", + "30. Zhuang, B., Tran, Q.H., Ji, P., Cheong, L.F., Chandraker, M.: Learning structure-and-motion-aware rolling shutter correction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4551-4560 (2019)", + "31. Zuo, X., Xie, X., Liu, Y., Huang, G.: Robust visual slam with point and line features. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1775-1782. IEEE (2017)" + ], + "bbox": [ + 215, + 147, + 785, + 743 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Zhang, B. Liao et al.", + "bbox": [ + 271, + 114, + 431, + 128 + ], + "page_idx": 15 + } +] \ No newline at end of file diff --git a/2024/RSL-BA_ Rolling Shutter Line Bundle Adjustment/25dc3962-792b-46d2-8db3-b73a1efe60d9_model.json b/2024/RSL-BA_ Rolling Shutter Line Bundle Adjustment/25dc3962-792b-46d2-8db3-b73a1efe60d9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..92adb7893231c229b5b83d45a6ee60d456e0aaef --- /dev/null +++ b/2024/RSL-BA_ Rolling Shutter Line Bundle Adjustment/25dc3962-792b-46d2-8db3-b73a1efe60d9_model.json @@ -0,0 +1,2498 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.328, + 0.141, + 0.674, + 0.187 + ], + "angle": 0, + "content": "RSL-BA: Rolling Shutter Line Bundle Adjustment" + }, + { + "type": "text", + "bbox": [ + 0.252, + 0.212, + 0.751, + 0.244 + ], + "angle": 0, + "content": "Yongcong Zhang\\(^{1,3,*}\\), Bangyan Liao\\(^{2,*}\\), Yifei Xue\\(^{1,3}\\), Chen Lu\\(^{4}\\), Peidong Liu\\(^{2}\\), and Yizhen Lao\\(^{1,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.316, + 0.254, + 0.686, + 0.285 + ], + "angle": 0, + "content": "\\(^{1}\\)Hunan University \\(^{2}\\)Westlake University \\(^{3}\\)DaHe.AI \\(^{4}\\)Dreame Technology Company" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.324, + 0.74, + 0.532 + ], + "angle": 0, + "content": "Abstract. The line is a prevalent element in man-made environments, inherently encoding spatial structural information, thus making it a more robust choice for feature representation in practical applications. Despite its apparent advantages, previous rolling shutter bundle adjustment (RSBA) methods have only supported sparse feature points, which lack robustness, particularly in degenerate environments. In this paper, we introduce the first rolling shutter line-based bundle adjustment solution, \\( RSL-BA \\). Specifically, we initially establish the rolling shutter camera line projection theory utilizing Plücker line parameterization. Subsequently, we derive a series of reprojection error formulations which are stable and efficient. Finally, we theoretically and experimentally demonstrate that our method can prevent three common degeneracies, one of which is first discovered in this paper. Extensive synthetic and real data experiments demonstrate that our method achieves efficiency and accuracy comparable to existing point-based rolling shutter bundle adjustment solutions." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.545, + 0.599, + 0.56 + ], + "angle": 0, + "content": "Keywords: Rolling Shutter \\(\\cdot\\) Bundle Adjustment" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.588, + 0.377, + 0.604 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.621, + 0.788, + 0.756 + ], + "angle": 0, + "content": "Bundle Adjustment (BA) is a crucial step in multi-view 3D reconstruction, as it jointly optimizes camera poses and scene structure. Although feature point-based BA [6,23,26] currently dominates the academic landscape, line-based BA has been gaining increasing attention from researchers [16,28]. Lines, being one of the most common features in man-made environments, effectively describe the structural information of 3D scenes, in contrast to points which rely solely on local image patches. This structural information provides robust constraints for visual reconstruction [27], making line a more suitable choice in challenging scenarios." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.757, + 0.787, + 0.789 + ], + "angle": 0, + "content": "In parallel with classical bundle adjustment [26], several methods [1,13] have incorporated the rolling shutter camera model to simultaneously estimate camera" + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.798, + 0.375, + 0.812 + ], + "angle": 0, + "content": "* Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.812, + 0.558, + 0.826 + ], + "angle": 0, + "content": "† Corresponding author: (yizhenlao@hnu.edu.cn)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.826, + 0.602, + 0.84 + ], + "angle": 0, + "content": "Project page: https://github.com/zhangtaxue/RSL-BA" + }, + { + "type": "list", + "bbox": [ + 0.233, + 0.798, + 0.602, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.432, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang, B. Liao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.221 + ], + "angle": 0, + "content": "poses, structure, and instantaneous motion parameters. However, all existing rolling shutter bundle adjustment (RSBA) methods rely solely on point features, while line features have never been addressed in RSBA. We identify two challenges that hinder the direct combination of line-based BA and the rolling shutter camera model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.788, + 0.327 + ], + "angle": 0, + "content": "The first challenge arises from the time-dependent exposure of RS cameras. Unlike CCD cameras and their global shutter (GS) counterparts, RS cameras capture images in a scanline-by-scanline manner. Consequently, as illustrated in Fig. 1, images taken by moving RS cameras exhibit distortions known as the RS effect. This effect curves the projection of each 3D straight line, making it non-trivial to directly transfer the line-based BA to the rolling shutter camera setting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.418 + ], + "angle": 0, + "content": "The second challenge is associated with degeneracy. As shown in [1, 30], degeneracy is one of the main obstacles in RSBA since iterative optimization can easily be collapsed in some degenerate solutions which are far from the ground truth solutions. Although several strategies have been proposed to mitigate degeneracy in RSBA [1, 12, 14], a specific strategy to address degeneracy in line-based RSBA is still lacking." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.788, + 0.465 + ], + "angle": 0, + "content": "In conclusion, an accurate and robust solution to line-based RSBA is still missing. Such a method would be vital in the potential widespread deployment of 3D vision with RS imaging systems." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.485, + 0.387, + 0.499 + ], + "angle": 0, + "content": "1.1 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.794, + 0.629 + ], + "angle": 0, + "content": "Video-based RSBA. The assumption of smooth continuous trajectories is widely employed for RS video inputs to reduce the optimization parameter space and enhance algorithm robustness. In [8], Hedborg et al. present an RSBA algorithm that interpolates the motion between consecutive frames. Zhuang et al. [29] further propose an optical flow-based RSBA, which are developed to recover the relative pose of an RS camera that undergoes constant velocity and acceleration motion, respectively. Following this assumption, a spline-based camera trajectory motion model is proposed by [20]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Unordered RSBA. As the unordered image set is the standard input for SfM, Albl et al. first addresses this setting by explicitly velocity modeling and optimization. Besides, a planar degeneracy configuration has also been disclosed in [1]. While in [12], Lao et al. propose a camera-based RSBA to simulate the actual camera projection, exhibiting the degeneracy resilience ability. In [10], Ito et al. establish the equivalence between self-calibrated SfM and RSSFM based on the pure rotation instantaneous motion model and affine camera assumption, while the work of [13] draws the equivalence between RSSfM and non-rigid SfM. Recently, Liao et al. proposed two techniques to boost the accuracy and robustness of RSBA in [14]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Line-based GSBA. In [2, 25], the authors propose a comprehensive line-based SFM pipeline, while [9, 18] introduce line-based incremental SFM frameworks. Wei et al. in [28] employs planes and points to guide matching, achieving efficient line reconstruction. Recently, Lui et al. introduced LIMAP in [16], where the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RSL-BA" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.149, + 0.321, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.194, + 0.318, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.24, + 0.274, + 0.248 + ], + "angle": 0, + "content": "input" + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.149, + 0.626, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.151, + 0.769, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.698, + 0.241, + 0.726, + 0.248 + ], + "angle": 0, + "content": "output" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.27, + 0.785, + 0.326 + ], + "angle": 0, + "content": "Fig. 1: The pipeline of proposed RSL-BA with some example results. Starting with the input of some RS images (left), the optimization of camera poses and 3D line coordinates is performed using proposed RSL-BA (middle), yielding well-reconstructed 3D line segments and accurately estimated camera poses (right)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.364, + 0.784, + 0.409 + ], + "angle": 0, + "content": "authors devised various strategies and algorithms to achieve efficient and precise line reconstruction. In [15, 21, 31], hybrid strategies combining lines and feature points are utilized to enhance the accuracy and robustness of SLAM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.785, + 0.514 + ], + "angle": 0, + "content": "To this end, we propose a novel \\( RSL-BA \\) algorithm to solve the line-based RSBA problem. Specifically, we first establish the rolling shutter line projection theory using the Plücker line parameterization and subsequently derive a series of reprojection error formulations. We claim that the proposed reprojection error formulations are efficient and exhibit the degeneration-resistant ability. We also provide complete degeneration-resistant proof of our proposed error formulation. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.523, + 0.784, + 0.565 + ], + "angle": 0, + "content": "- To the best of our knowledge, this is the first line-based RSBA solution, which serves as the foundation for line-based or point-line-hybrid RS-SfM and RS-SLAM." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.569, + 0.785, + 0.611 + ], + "angle": 0, + "content": "- we theoretically and experimentally demonstrate that the proposed \\( RSLBA \\) can prevent three common degeneracies in RSBA, one of the degenerate cases is first discovered in this paper." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.614, + 0.784, + 0.655 + ], + "angle": 0, + "content": "- The extensive evaluations in both synthetic and real datasets exhibit the comparable efficiency and accuracy of the proposed method over previous works." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.523, + 0.785, + 0.655 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.689, + 0.368, + 0.706 + ], + "angle": 0, + "content": "2 Background" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.784, + 0.75 + ], + "angle": 0, + "content": "In this section, we review the rolling shutter projection model in Sec. 2.1 and the parameterization of 3D lines in Sec. 2.2." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.771, + 0.605, + 0.787 + ], + "angle": 0, + "content": "2.1 Rolling Shutter Camera Projection Mode" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Let \\(\\mathbf{R}(v) \\in \\mathbf{SO}(3)\\) and \\(\\mathbf{t}(v) \\in \\mathbb{R}^3\\) represent the camera rotation and translation, respectively, when the row index \\(v\\) of measurement is acquired. Denote the intrinsic camera parameters as \\(\\mathbf{K}\\), and the initial rotation and translation of the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.432, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang, B. Liao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "camera at \\( v = 0 \\) as \\( \\mathbf{R}_0 \\) and \\( \\mathbf{t}_0 \\), respectively. The projection matrix of the global shutter camera is defined as \\( \\mathbf{P}_0 = \\mathbf{K}[\\mathbf{R}_0, \\mathbf{t}_0] \\). Given the assumption of constant camera motion during frame capture, which is commonly made in RS 3D vision tasks [1,4,13], we can model the instantaneous motion as:" + }, + { + "type": "equation", + "bbox": [ + 0.349, + 0.224, + 0.786, + 0.24 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R} (v) = (\\mathbf {I} + [ \\boldsymbol {\\omega} ] _ {\\times} v) \\mathbf {R} _ {0}, \\quad \\mathbf {t} (v) = \\mathbf {t} _ {0} + \\mathbf {d} v, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.254, + 0.785, + 0.314 + ], + "angle": 0, + "content": "where \\(\\mathbf{d} = [d_x, d_y, d_z]^\\top\\) is the translational velocity vector and \\(\\boldsymbol{\\omega} = [\\omega_x, \\omega_y, \\omega_z]^\\top\\) is the rotational velocity vector, and \\([\\boldsymbol{\\omega}]_\\times\\) represents the skew-symmetric matrix of vector \\(\\boldsymbol{\\omega}\\). Then, the projection matrix of the camera under the RS model can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.387, + 0.317, + 0.786, + 0.34 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {v} = \\mathbf {K} \\left[ \\mathbf {R} _ {v}, \\mathbf {t} _ {v} \\right] = \\mathbf {P} _ {0} + v \\mathbf {Q} \\\\ \\mathbf {K} ^ {[ r ]} = \\left[ \\begin{array}{l l l l l} \\mathbf {P} _ {v} & \\mathbf {t} _ {v} & \\mathbf {Q} _ {v} & \\mathbf {I} ^ {[ r ]} & \\mathbf {D} _ {v} \\end{array} \\right] \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.337, + 0.632, + 0.351 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {0} = \\mathbf {K} \\left[ \\mathbf {R} _ {0}, \\mathbf {t} _ {0} \\right] \\quad \\mathbf {Q} = \\mathbf {K} \\left[ \\left[ \\boldsymbol {\\omega} \\right] _ {\\times} \\mathbf {R} _ {0}, \\mathbf {d} \\right].\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.381, + 0.56, + 0.396 + ], + "angle": 0, + "content": "2.2 3D Line Segments Parameterization" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.415, + 0.787, + 0.459 + ], + "angle": 0, + "content": "We employ the Plücker matrix to parameterize the 3D line [7]. Specifically, given two homogeneous 3D points \\( P_{a} = [a_{1}, a_{2}, a_{3}, 1]^{\\top} \\) and \\( P_{b} = [b_{1}, b_{2}, b_{3}, 1]^{\\top} \\). The Plücker matrix \\( \\mathbf{L} \\) can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.472, + 0.752, + 0.489 + ], + "angle": 0, + "content": "\\[\n\\mathbf {L} = P _ {a} P _ {b} ^ {\\top} - P _ {b} P _ {a} ^ {\\top}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.231, + 0.493, + 0.785, + 0.556 + ], + "angle": 0, + "content": "\\[\n= \\left[ \\begin{array}{c c c c} 0 & a _ {1} b _ {2} - a _ {2} b _ {1} & a _ {1} b _ {3} - a _ {3} b _ {1} & a _ {1} - b _ {1} \\\\ a _ {2} b _ {1} - a _ {1} b _ {2} & 0 & a _ {2} b _ {3} - a _ {3} b _ {2} & a _ {2} - b _ {2} \\\\ a _ {3} b _ {1} - a _ {1} b _ {3} & a _ {3} b _ {2} - a _ {2} b _ {3} & 0 & a _ {3} - b _ {3} \\\\ b _ {1} - a _ {1} & b _ {2} - a _ {2} & b _ {3} - a _ {3} & 0 \\end{array} \\right] = \\left[ \\begin{array}{c c c c} 0 & l _ {1 2} & l _ {1 3} & l _ {1 4} \\\\ - l _ {1 2} & 0 & l _ {2 3} & l _ {2 4} \\\\ - l _ {1 3} & - l _ {2 3} & 0 & l _ {3 4} \\\\ - l _ {1 4} & - l _ {2 4} & - l _ {3 4} & 0 \\end{array} \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.57, + 0.788, + 0.631 + ], + "angle": 0, + "content": "where the \\([l_{14}, l_{24}, l_{34}]^{\\top}\\) represents the direction of this straight line while the \\([-l_{23}, l_{13}, -l_{12}]^{\\top}\\) represents the normal direction of this line. The projection of 3D Plücker matrix is the skew-symmetric matrix of a 2D normalized line. It can be related to each other as:" + }, + { + "type": "equation", + "bbox": [ + 0.423, + 0.647, + 0.785, + 0.664 + ], + "angle": 0, + "content": "\\[\n\\left[ \\mathbf {l} _ {g s} \\right] _ {\\times} = \\left[ \\mathbf {R}, \\mathbf {t} \\right] \\mathbf {L} \\left[ \\mathbf {R}, \\mathbf {t} \\right] ^ {\\top} \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.694, + 0.381, + 0.712 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In Sec. 3.1, we initially establish the rolling shutter projection equations using the Plücker line parameterization. These equations establish a connection between the 3D lines and the curve parameters on the image plane. Then in Sec. 3.2, we present a series of line-based reprojection error equations. On top of the error definition, we can formulate the entire rolling shutter line bundle adjustment problem in Sec. 3.3. Finally, in Sec. 3.4, we demonstrate the degeneracy resilience capability of the proposed error equations." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RSL-BA" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.63, + 0.163 + ], + "angle": 0, + "content": "3.1 Rolling Shutter Line Projection Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.172, + 0.784, + 0.217 + ], + "angle": 0, + "content": "As previously explained, the curved rolling shutter projection of 3D lines occurs because of the camera's nonlinear motion. Using the Plücker parameterization method, we can derive explicit expressions for the curve parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.218, + 0.785, + 0.249 + ], + "angle": 0, + "content": "Firstly, we construct the skew-symmetric matrix \\(\\left[\\mathbf{l}_{rs}\\right]_{\\times}\\) of the instantaneous projection line by combining Eq. (2) and Eq. (4):" + }, + { + "type": "equation", + "bbox": [ + 0.318, + 0.259, + 0.785, + 0.317 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left[ \\mathbf {l} _ {r s} \\right] _ {\\times} = \\mathbf {P} _ {v} \\mathbf {L} \\mathbf {P} _ {v} ^ {\\top} \\\\ = \\mathbf {P} _ {0} \\mathbf {L} \\mathbf {P} _ {0} ^ {\\top} + v \\left(\\mathbf {P} _ {0} \\mathbf {L} \\mathbf {Q} ^ {\\top} + \\mathbf {Q} \\mathbf {L} \\mathbf {P} _ {0} ^ {\\top}\\right) + v ^ {2} \\mathbf {Q} \\mathbf {L} \\mathbf {Q} ^ {\\top} \\tag {5} \\\\ = \\mathbf {A} _ {1} + v \\mathbf {A} _ {2} + v ^ {2} \\mathbf {A} _ {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.784, + 0.361 + ], + "angle": 0, + "content": "Secondly, as \\(\\mathbf{l}_{rs} = [l_1,l_2,l_3]^\\top\\) and \\([\\mathbf{l}_{rs}]_{\\times} = \\mathbf{A}_1 + v\\mathbf{A}_2 + v^2\\mathbf{A}_3\\), we can determine the instantaneous projected line parameter as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.403, + 0.373, + 0.785, + 0.42 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} l _ {1} = \\mathbf {A} _ {1} ^ {3 2} + v \\mathbf {A} _ {2} ^ {3 2} + v ^ {2} \\mathbf {A} _ {3} ^ {3 2} \\\\ l _ {2} = \\mathbf {A} _ {1} ^ {1 3} + v \\mathbf {A} _ {2} ^ {1 3} + v ^ {2} \\mathbf {A} _ {3} ^ {1 3} \\\\ l _ {3} = \\mathbf {A} _ {1} ^ {2 1} + v \\mathbf {A} _ {2} ^ {2 1} + v ^ {2} \\mathbf {A} _ {3} ^ {2 1} \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.784, + 0.475 + ], + "angle": 0, + "content": "Ultimately, we note that the instantaneous projected line parameter fulfills this \\( l_{1}u + l_{2}v + l_{3} = 0 \\) equation, which we can expand to obtain the final curve equation, denoted as \\( \\pmb{\\iota} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.486, + 0.785, + 0.524 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\iota : \\mathbf {A} _ {3} ^ {1 3} v ^ {3} + \\mathbf {A} _ {3} ^ {3 2} u v ^ {2} + \\left(\\mathbf {A} _ {2} ^ {1 3} + \\mathbf {A} _ {3} ^ {2 1}\\right) v ^ {2} + \\mathbf {A} _ {2} ^ {3 2} u v \\\\ + \\left(\\mathbf {A} _ {1} ^ {1 3} + \\mathbf {A} _ {2} ^ {2 1}\\right) v + \\mathbf {A} _ {1} ^ {3 2} u + \\mathbf {A} _ {1} ^ {2 1} = 0 \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.784, + 0.611 + ], + "angle": 0, + "content": "Our conclusion aligns with that of Lao et al. [11], yielding an identical polynomial curve degree. Nevertheless, our utilization of a 4-DoF orthogonal representation representation for 3D lines has distinct benefits when compared to Lao's 5-DoF over-parameterized representation. These advantages include the capability to conduct unconstrained optimization on spatial lines directly." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.785, + 0.703 + ], + "angle": 0, + "content": "Remark. The geometric interpretation underlying this derivation is quite straightforward. Given the parameters of a spatial line and the current camera pose, if point \\((\\mathbf{u},\\mathbf{v})\\) lies on the curve obtained by projecting this spatial line, then when the RS camera scans to the v-th row, the projection of the spatial line onto the image plane follows Eq. (6). Therefore, it is evident that point \\((\\mathbf{u},\\mathbf{v})\\) must be located on this line. Based on this reasoning, we can deduce Eq. (7)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.725, + 0.523, + 0.74 + ], + "angle": 0, + "content": "3.2 Line-based Reprojection Errors" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.841 + ], + "angle": 0, + "content": "As claimed in Sec. 3.1, a 3D straight line will become a polynomial curve on the captured image. In this section, we will establish a series of re-projection errors which are stable and robust. We first leverage the similar intuition in Sec. 3.1 to obtain the distance error representation. As shown in Fig. 2(a), Let there be a point \\(\\mathbf{q} = [u,v]^{\\top}\\) on the observed curve \\(\\iota\\), then the pose of the camera when exposing that point is \\(\\mathbf{R}_v\\), \\(\\mathbf{t}_v\\), then projecting the 3D line \\(\\mathbf{L}\\) onto the image" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.432, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang, B. Liao et al." + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.144, + 0.777, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.302, + 0.789, + 0.402 + ], + "angle": 0, + "content": "Fig. 2: The illustration of proposed line-based reprojection errors given the spatial line \\( L \\) and observed curve \\( \\iota \\). Instead of directly measuring the distance between the observed curve and the projected curve, we split each curve into multiple points \\( q \\), and aggregate the distance errors. Specifically, we first utilize the virtual projected line \\( l \\) when observing it, as shown in (a). To measure the distance between each point \\( q \\) and the corresponding virtual lines, we can use the perpendicular distance (b), the horizontal distance (c,d) or the tangent distance (e) as our basic metrics." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.43, + 0.789, + 0.492 + ], + "angle": 0, + "content": "plane, this results in the virtual 2D line \\(\\mathbf{1} = [l_1, l_2, l_3]^\\top\\). Under the condition of ground truth, this virtual line must theoretically intersect with that point \\(\\mathbf{q}\\). This constraint directly leads us to the two variants of distance error, which we have defined below." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.516, + 0.788, + 0.562 + ], + "angle": 0, + "content": "Perpendicular Distance Error \\(\\mathbf{e}_{d1}\\). The first distance error formulation, named perpendicular distance error, measures the shortest distance from point \\(\\mathbf{q}\\) to the line \\(\\mathbf{l}\\), as shown in Fig. 2(b). The error is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.424, + 0.572, + 0.787, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\mathbf {e} _ {d 1} = \\frac {l _ {1} u + l _ {2} v + l _ {3}}{\\sqrt {l _ {1} ^ {2} + l _ {2} ^ {2}}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.618, + 0.724, + 0.635 + ], + "angle": 0, + "content": "where the line parameters \\([l_1, l_2, l_3]\\) can be calculated through Eq. (6)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.659, + 0.789, + 0.705 + ], + "angle": 0, + "content": "Horizontal Distance Error \\(\\mathbf{e}_{d2}\\). The second distance error formulation, named horizontal distance error, measures the u-axis distance from point \\(\\mathbf{q}\\) to the line \\(\\mathbf{l}\\), as shown in Fig. 2(c). The error can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.452, + 0.716, + 0.787, + 0.734 + ], + "angle": 0, + "content": "\\[\n\\mathbf {e} _ {d 2} = u - u ^ {\\prime}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.746, + 0.788, + 0.773 + ], + "angle": 0, + "content": "where the corresponding \\( u' \\) coordinate of virtual line 1 at the \\( v \\)-th row can be calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.444, + 0.773, + 0.787, + 0.805 + ], + "angle": 0, + "content": "\\[\nu ^ {\\prime} = - \\frac {l _ {2} v + l _ {3}}{l _ {1}}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "As shown in Fig. 2(d), the \\( e_{d2} \\) distance error is equivalent to the u-axis distance from the point to the projection curve." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RSL-BA" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.222 + ], + "angle": 0, + "content": "Tangent Error \\(\\mathbf{e}_t\\). While distance error provides an elegant constraint, it is insufficient to rely solely on distance error in certain scenarios. To further enhance the robustness, we introduce a second type of tangent error. As shown in Fig. 2(e), we constrain the tangent of the projected curve \\(\\iota'\\) to match the observed tangent direction. The error can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.467, + 0.234, + 0.786, + 0.251 + ], + "angle": 0, + "content": "\\[\n\\mathbf {e} _ {t} = \\mathbf {s} ^ {\\top} \\mathbf {s} ^ {\\prime}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.263, + 0.784, + 0.294 + ], + "angle": 0, + "content": "where \\(\\mathbf{s}\\) represents the tangent direction observation. The \\(\\mathbf{s}'\\) can be calculated analytically following the implicit function theorem as:" + }, + { + "type": "equation", + "bbox": [ + 0.337, + 0.305, + 0.786, + 0.339 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} ^ {\\prime} = \\left[ \\frac {1}{\\sqrt {1 + s ^ {2}}}, \\frac {s}{\\sqrt {1 + s ^ {2}}} \\right], \\quad s = \\frac {\\partial u}{\\partial v} = - \\frac {\\partial \\iota / \\partial v}{\\partial \\iota / \\partial u} \\tag {12}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.357, + 0.599, + 0.373 + ], + "angle": 0, + "content": "3.3 Rolling Shutter Line Bundle Adjustment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.383, + 0.784, + 0.444 + ], + "angle": 0, + "content": "The non-linear least squares solvers are used to find an optimal solution \\(\\pmb{\\theta}^{*}\\) including camera poses \\(\\mathbf{R}^{*},\\mathbf{t}^{*}\\), instantaneous motion \\(\\omega^{*},\\mathbf{d}^{*}\\) and 3D lines \\(\\pmb{\\tau}^{*}\\) by minimizing the reprojection error \\(\\mathbf{e}_i^j\\) from line \\(i\\) to camera \\(j\\) over all the camera index in set \\(\\mathcal{F}\\) and corresponding observed 3D lines index in subset \\(\\mathcal{P}_j\\):" + }, + { + "type": "equation", + "bbox": [ + 0.321, + 0.457, + 0.786, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\theta} ^ {*} = \\left\\{\\boldsymbol {\\tau} ^ {*}, \\mathbf {R} ^ {*}, \\mathbf {t} ^ {*}, \\boldsymbol {\\omega} ^ {*}, \\mathbf {d} ^ {*} \\right\\} = \\arg \\min _ {\\boldsymbol {\\theta}} \\sum_ {j \\in \\mathcal {F}} \\sum_ {i \\in \\mathcal {P} _ {j}} \\left\\| \\mathbf {e} _ {i} ^ {j} \\right\\| _ {2} ^ {2}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.784, + 0.533 + ], + "angle": 0, + "content": "On top of errors proposed in Sec. 3.2, we provide two variants of reprojection error:" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.536, + 0.785, + 0.551 + ], + "angle": 0, + "content": "\\[\n\\mathbf {e} _ {1} = \\mathbf {e} _ {d 1} + \\lambda_ {1} \\mathbf {e} _ {t}, \\quad \\mathbf {e} _ {2} = \\mathbf {e} _ {d 2} + \\lambda_ {2} \\mathbf {e} _ {t}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.557, + 0.784, + 0.616 + ], + "angle": 0, + "content": "where \\(\\lambda_1, \\lambda_2\\) are the relative weight. We have experimentally validated the optimality of these two error variants in Sec. 4.1. After definition, this non-linear least square problem can be iteratively solved through Levenberg-Marquardt [19] method." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.618, + 0.784, + 0.647 + ], + "angle": 0, + "content": "Besides, the full analytical Jacobian has been derived in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.671, + 0.434, + 0.687 + ], + "angle": 0, + "content": "3.4 Degeneracy Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.696, + 0.784, + 0.773 + ], + "angle": 0, + "content": "In this section, we conducted theoretical analysis on the resilience of \\( RSL-BA \\) to three types of degeneracy scenarios. These three degeneracy scenarios include plane degeneracy case [1], 2-vIEWS pure translation case [30], and X-Y pure translation degeneracy case disclosed by us. We assume the positive direction of X Y Z axis is right down forward, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.843 + ], + "angle": 0, + "content": "(1) Resistance to Plane Degeneracy [1]. We first briefly describe the conditions under which the plane degeneration occurs: when the camera's y-axis is parallel, if a rotational velocity \\(\\omega = [-1,0,0]^{\\top}\\) is given to the camera, then the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.432, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang, B. Liao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "reconstructed 3D points will be compressed onto a plane perpendicular to the \\(y\\)-axis, and satisfy that the reprojection error is 0." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.281 + ], + "angle": 0, + "content": "For line features, we also assume \\(\\mathbf{R}_0 = \\mathbf{I}\\), \\(\\mathbf{t}_0 = \\mathbf{0}\\), \\(\\boldsymbol{\\omega} = [-1,0,0]^\\top\\), \\(\\mathbf{d} = 0\\). Let there be two points, \\(\\mathbf{P}_a\\) and \\(\\mathbf{P}_b\\), on the spatial line \\(\\mathbf{L}\\). \\(\\mathbf{P}_a = [a_1,a_2,a_3,1]^\\top\\) and \\(\\mathbf{P}_b = [b_1,b_2,b_3,1]^\\top\\). When the space is compressed into a plane, the new coordinates of \\(\\mathbf{P}_a\\) and \\(\\mathbf{P}_b\\) are \\(\\mathbf{P}_a' = [a_1,0,a_3,1]^\\top\\) and \\(\\mathbf{P}_b' = [b_1,0,b_3,1]^\\top\\), respectively. By referring to Sec. 3.1, we can solve for the projection parameters of the two-dimensional curve at this moment. The parameters of the curve at this time are:" + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.284, + 0.785, + 0.332 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\mathbf {A} _ {3} ^ {1 3} = \\mathbf {A} _ {3} ^ {3 2} = \\mathbf {A} _ {2} ^ {1 3} = \\mathbf {A} _ {3} ^ {2 1} = \\mathbf {A} _ {2} ^ {3 2} = \\mathbf {A} _ {1} ^ {3 2} = \\mathbf {A} _ {1} ^ {2 1} = 0 \\\\ \\mathbf {A} _ {1} ^ {1 3} = a _ {3} b _ {1} - a _ {1} b _ {3} \\\\ \\mathbf {A} _ {2} ^ {2 1} = a _ {1} b _ {3} - a _ {3} b _ {1} \\end{array} \\right. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.523 + ], + "angle": 0, + "content": "Substituting the parameters into the curve expression Eq. (7), we obtain \\(0 = 0\\), indicating that u and v are unrestricted. Under the influence of \\(\\omega = [-100]^{\\top}\\), the camera's imaging rows are always coplanar with the line, making the distance from any point on the image to the line zero and satisfying the point-to-line distance error function (Sec. 3.2). However, the slope constraint cannot be met. Using this projection equation to calculate the tangent direction of any point on the curve results in indeterminate forms \\([0/0, 0/0]\\), which do not match the measured tangent direction. Therefore, using the tangent direction of points on the curve as the error function prevents this degeneration. As verified in Fig 5, our proposed RSL-BA can suppress plane degeneration, whereas conventional rolling shutter point-based bundle adjustment method NMRsBA [1] converges to a degenerate solution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.784, + 0.64 + ], + "angle": 0, + "content": "(2) Resistance to 2-views Pure Translation Degeneracy [30]. We first briefly describe the conditions under which this degeneration occurs: when the camera moves in a straight line to take two pictures, assigning any translational velocity to the camera for both shots, it is possible to find new 3D points that satisfy the condition that the reprojection error of the points under this velocity is zero." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.784, + 0.689 + ], + "angle": 0, + "content": "Let's assume the real pose of the camera before the degeneration is: \\(\\mathbf{R}_0 = \\mathbf{I}\\), \\(\\mathbf{t}_0 = \\mathbf{0}\\), \\(\\boldsymbol{\\omega} = [0,0,0]^\\top\\), \\(\\mathbf{d} = [d_1,d_2,d_3]^\\top\\), \\(\\mathbf{L} = (l_{12},l_{13},l_{23},l_{14},l_{24},l_{34})\\) In this case, substitute these parameters into Eq. (7) we can get curve expression:" + }, + { + "type": "equation", + "bbox": [ + 0.233, + 0.703, + 0.785, + 0.734 + ], + "angle": 0, + "content": "\\[\n\\left(l _ {1 4} d _ {3} - l _ {3 4} d _ {1}\\right) v ^ {2} + \\left(l _ {3 4} d _ {2} - l _ {2 4} d _ {3}\\right) u v + \\left(l _ {2 4} d _ {1} - l _ {1 4} d _ {2} + l _ {1 3}\\right) v - l _ {2 3} u - l _ {1 2} = 0 \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.79, + 0.765 + ], + "angle": 0, + "content": "After the degeneration occurs, \\(\\mathbf{R}_0^{\\prime} = \\mathbf{I}\\), \\(\\mathbf{t}_0^{\\prime} = \\mathbf{0}\\), \\(\\boldsymbol{\\omega}^{\\prime} = [0\\quad 0\\quad 0]^\\top\\), \\(\\mathbf{d}' = r[d_1,d_2,d_3]\\), \\(\\mathbf{L} = (l_{12}',l_{13}',l_{23}',l_{14}',l_{24}',l_{34}')\\). In this case, the curve expression is:" + }, + { + "type": "equation", + "bbox": [ + 0.222, + 0.779, + 0.8, + 0.809 + ], + "angle": 0, + "content": "\\[\nr \\left(l _ {1 4} ^ {\\prime} d _ {3} - l _ {3 4} ^ {\\prime} d _ {1}\\right) v ^ {2} + r \\left(l _ {3 4} ^ {\\prime} d _ {2} - l _ {2 4} ^ {\\prime} d _ {3}\\right) u v + \\left(r \\left(l _ {2 4} ^ {\\prime} d _ {1} - l _ {1 4} ^ {\\prime} d _ {2}\\right) + l _ {1 3} ^ {\\prime}\\right) v - l _ {2 3} ^ {\\prime} u - l _ {1 2} ^ {\\prime} = 0 \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "To make the reprojection error zero, the curve expressions before and after degeneration must be the same, meaning the corresponding coefficients are in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RSL-BA" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.157, + 0.396, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.344, + 0.263, + 0.369, + 0.278 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.155, + 0.571, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.493, + 0.263, + 0.518, + 0.278 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.152, + 0.718, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.263, + 0.671, + 0.278 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.297, + 0.785, + 0.352 + ], + "angle": 0, + "content": "Fig. 3: The illustration of degeneracy resistance of our proposed method in 2-views pure translation degeneracy. (a) Degeneration in single-view scenarios. (b) Suppression of degeneration using line features in dual-view scenarios. (c) Inability of line features to suppress degeneration in special dual-view scenarios" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.384, + 0.3, + 0.398 + ], + "angle": 0, + "content": "proportion." + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.408, + 0.785, + 0.487 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} l _ {1 4} d _ {3} - l _ {3 4} d _ {1} = s r \\left(l _ {1 4} ^ {\\prime} d _ {3} - l _ {3 4} ^ {\\prime} d _ {1}\\right) \\\\ l _ {3 4} d _ {2} - l _ {2 4} d _ {3} = s r \\left(l _ {3 4} ^ {\\prime} d _ {2} - l _ {2 4} ^ {\\prime} d _ {3}\\right) \\\\ l _ {2 4} d _ {1} - l _ {1 4} d _ {2} + l _ {1 3} = s r \\left(l _ {2 4} ^ {\\prime} d _ {1} - l _ {1 4} ^ {\\prime} d _ {2}\\right) + s l _ {1 3} ^ {\\prime} \\\\ l _ {2 3} = s l _ {2 3} ^ {\\prime} \\\\ l _ {1 2} = s l _ {1 2} ^ {\\prime} \\end{array} \\right. \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.497, + 0.785, + 0.527 + ], + "angle": 0, + "content": "Substituting the first and second equations into the third equation yields \\( l_{13} = sl_{13}' \\). Combining this with:" + }, + { + "type": "equation", + "bbox": [ + 0.4, + 0.538, + 0.785, + 0.571 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} l _ {1 2} l _ {3 4} - l _ {1 3} l _ {2 4} + l _ {1 4} l _ {2 3} = 0 \\\\ l _ {1 2} ^ {\\prime} l _ {3 4} ^ {\\prime} - l _ {1 3} ^ {\\prime} l _ {2 4} ^ {\\prime} + l _ {1 4} ^ {\\prime} l _ {2 3} ^ {\\prime} = 0 \\end{array} \\right. \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.581, + 0.295, + 0.594 + ], + "angle": 0, + "content": "we obtain:" + }, + { + "type": "equation", + "bbox": [ + 0.244, + 0.606, + 0.785, + 0.627 + ], + "angle": 0, + "content": "\\[\nl _ {1 2} ^ {\\prime} = \\frac {l _ {1 2} ^ {\\prime}}{s}, \\quad l _ {1 3} ^ {\\prime} = \\frac {l _ {1 3}}{s}, \\quad l _ {2 3} ^ {\\prime} = \\frac {l _ {2 3}}{s}, \\quad l _ {1 4} ^ {\\prime} = \\frac {l _ {1 4}}{s r}, \\quad l _ {2 4} ^ {\\prime} = \\frac {l _ {2 4}}{s r}, \\quad l _ {3 4} ^ {\\prime} = \\frac {l _ {3 4}}{s r} \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.636, + 0.784, + 0.788 + ], + "angle": 0, + "content": "Note that the direction and normal direction of the line remain unchanged, indicating that the degenerated line stays in the original plane and parallel to the original line, as shown in Fig. 3(1), where \\(\\mathbf{L}^{gt}\\) is the original line, and \\(\\mathbf{L}_1^{dg}\\) is the degenerated line. In the dual-frame scenario (Fig.3(2)), when the plane \\(\\mathbf{O}_1\\mathbf{L}^{gt}\\) is not coplanar with the plane \\(\\mathbf{O}_2\\mathbf{L}^{gt}\\), their degenerated states \\(\\mathbf{L}_1^{dg}\\) and \\(\\mathbf{L}_2^{gt}\\) cannot coincide (except at \\(\\mathbf{L}^{gt}\\)). In this case, line constraints can prevent degeneration. However, when \\(\\mathbf{L}^{gt}\\), \\(\\mathbf{O}_1\\), and \\(\\mathbf{O}_2\\) are coplanar (Fig.3(3)), line constraints cannot suppress degeneration, though this is a very special condition requiring all lines to meet this criterion. This demonstrates that line constraints are more effective than point constraints in preventing degeneration." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "(3) Resistance to X-Y Pure Translation Degeneracy (our new founding). We present a pure translation scenario without the z-direction, suitable for" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.432, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang, B. Liao et al." + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.152, + 0.396, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.326, + 0.244, + 0.344, + 0.256 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.4, + 0.164, + 0.501, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.422, + 0.244, + 0.44, + 0.256 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.156, + 0.603, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.244, + 0.566, + 0.256 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.611, + 0.153, + 0.722, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.243, + 0.667, + 0.255 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.343 + ], + "angle": 0, + "content": "Fig. 4: In the case of X-Y pure translation, optimization compresses points into a straight line. In a single view, pixel points parallel to the motion trajectory backproject to the same 3D point ((a) and (b)). (c) Different intersection points in space are equidistant from the camera plane. (d) All intersection points formed by backprojection constitute a straight line." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.37, + 0.785, + 0.4 + ], + "angle": 0, + "content": "multi-view BA, compressing space points into a straight line. Unlike the second type of degeneration, this does not depend on the number of images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.4, + 0.789, + 0.522 + ], + "angle": 0, + "content": "Firstly, considering a single view, as shown in Fig.4(1), during exposure, the camera moves from position \\(\\mathbf{O}_1^1\\) to \\(\\mathbf{O}_1^3\\), where pixel \\(\\mathbf{q}_1^1\\) is located on the first row and \\(\\mathbf{q}_1^3\\) is located on the \\(H\\)-th row. Their corresponding spatial points are located on the rays \\(\\mathbf{O}_1^1\\mathbf{q}_1^1\\) and \\(\\mathbf{O}_1^3\\mathbf{q}_1^3\\), intersecting at point \\(\\mathbf{P}_1\\). Actually, any point on the line \\(\\mathbf{q}_1^1\\mathbf{q}_1^3\\) intersects with its corresponding 3D point at \\(\\mathbf{P}_1\\) during exposure. Let \\(\\mathbf{O}_1^2\\) be the camera position when exposed to the \\(h_1^2\\)-th row, with corresponding feature point \\(\\mathbf{q}_1^2\\). We only consider the plane \\(\\mathbf{O}_1^1\\mathbf{P}_1\\mathbf{O}_1^3\\) in Fig.4(2). The motion of \\(\\mathbf{O}\\) is uniform rectilinear motion thus, there exists a length relationship:" + }, + { + "type": "equation", + "bbox": [ + 0.336, + 0.528, + 0.785, + 0.573 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{c} l _ {\\mathbf {q} _ {1} ^ {1} \\mathbf {q} _ {1} ^ {2}} = \\frac {h _ {1} ^ {2}}{H} l _ {\\mathbf {q} _ {1} ^ {1} \\mathbf {q} _ {1} ^ {3}}, \\quad l _ {\\mathbf {q} _ {3} ^ {1} \\mathbf {q} _ {1} ^ {2}} = \\left(1 - \\frac {h _ {1} ^ {2}}{H}\\right) l _ {\\mathbf {q} _ {1} ^ {1} \\mathbf {q} _ {1} ^ {3}} \\\\ l _ {\\mathbf {O} _ {1} ^ {1} \\mathbf {O} _ {1} ^ {2}} = \\frac {h _ {1} ^ {2}}{H} l _ {\\mathbf {O} _ {1} ^ {1} \\mathbf {O} _ {1} ^ {3}}, \\quad l _ {\\mathbf {O} _ {3} ^ {1} \\mathbf {O} _ {1} ^ {2}} = \\left(1 - \\frac {h _ {1} ^ {2}}{H}\\right) l _ {\\mathbf {O} _ {1} ^ {1} \\mathbf {O} _ {1} ^ {3}} \\end{array} \\right. \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.578, + 0.303, + 0.591 + ], + "angle": 0, + "content": "so we have:" + }, + { + "type": "equation", + "bbox": [ + 0.41, + 0.587, + 0.785, + 0.626 + ], + "angle": 0, + "content": "\\[\n\\frac {l _ {\\mathbf {q} _ {1} ^ {1} \\mathbf {q} _ {1} ^ {2}}}{l _ {\\mathbf {O} _ {1} ^ {1} \\mathbf {O} _ {1} ^ {2}}} = \\frac {l _ {\\mathbf {q} _ {1} ^ {3} \\mathbf {q} _ {1} ^ {2}}}{l _ {\\mathbf {O} _ {3} ^ {1} \\mathbf {O} _ {1} ^ {2}}} = \\frac {l _ {\\mathbf {q} _ {1} ^ {1} \\mathbf {q} _ {1} ^ {3}}}{l _ {\\mathbf {O} _ {1} ^ {1} \\mathbf {O} _ {1} ^ {3}}} \\tag {22}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Therefore, the ray \\(\\mathbf{O}_1^2\\mathbf{q}_1^2\\) will also intersect at point \\(\\mathbf{P_1}\\). For every pixel on a line parallel to \\(\\mathbf{O}_1^1\\mathbf{O}_1^3\\), it can be traced back to a different spatial point. For pixel points on different lines parallel to the motion trajectory, as shown in Fig. 4(3), because the distance \\(l_{q^1 q^3}\\) is the same and the camera focal length remains unchanged, according to similar triangles, these spatial points have equal distances to the image plane. Simultaneously, due to symmetry, these points also lie in the plane that is perpendicular to and bisects the line segment \\(\\mathbf{O}_1^1\\mathbf{O}_1^2\\). Therefore, the intersection line of these two planes represents the final degenerated configuration, which is a line parallel to the x-axis, as shown in Fig.4(4). In multiple views, with no z-direction movement, the imaging planes of different cameras coincide. We assign the same camera configuration to different views, including rotation, translation, angular velocity, and linear velocity. From the single-view case, corresponding points on the pixel plane parallel to the motion trajectory reconstruct to the same spatial point, with zero reprojection error." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RSL-BA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.147, + 0.699, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.304, + 0.788, + 0.361 + ], + "angle": 0, + "content": "Fig.5: Comparison of degenerate resistance ability between NMRSBA [1] and proposed \\( RSL-BA \\). Ground truth camera poses and structure are colored with black. This example illustrates that our proposed \\( RSL-BA \\) has the resistance ability against the plane and X-Y pure translation degeneracy." + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.377, + 0.75, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.486 + ], + "angle": 0, + "content": "Fig. 6: Accuracy and computational time comparison of the proposed two combinations of reprojection errors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.788, + 0.607 + ], + "angle": 0, + "content": "For the line constraints proposed in this paper, \\( RSL-BA \\) can still effectively suppress degeneration. The proof process is similar to the previous section Sec. 3.4, with the only modification being setting \\( d_{3} = 0 \\) in the camera's motion direction. As verified in Fig 5, our proposed \\( RSL-BA \\) can suppress X-Y pure translation degeneration, whereas conventional rolling shutter point-based bundle adjustment method NMRSBA [1] converges to a degenerate solution." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.631, + 0.377, + 0.648 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.663, + 0.788, + 0.725 + ], + "angle": 0, + "content": "We compare our method with two SOTA GS-based-method: 1) GSBA [17], 2) GLBA [25], and four SOTA RS-based-method: 1) MRSBA [5], 2) NMRSBA [1], 3) WMRSBA [14], 4) NWRSBA [14]. The experiments are conducted on a laptop with an Intel i7 CPU and all algorithms are implemented in MATLAB." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.748, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Evaluation metrics. The accuracy measures employed encompass standard metrics such as rotation error, translation error, and algorithm efficiency. When conducting experiments related to line BA, we also incorporate error metrics that assess the accuracy of line reconstruction. Given the ground truth rotation \\(\\mathbf{R}_g\\), translation \\(\\mathbf{t}_g\\), and space line \\(\\mathbf{L}_g = (\\mathbf{n}_g^\\top, \\mathbf{a}_g^\\top)^\\top\\) with an arbitrary point \\(\\mathbf{P}_g\\) on it. Similarly, the optimized parameters are denoted as \\(\\mathbf{R}_d\\), \\(\\mathbf{t}_d\\), \\(\\mathbf{L}_d = (\\mathbf{n}_d^\\top, \\mathbf{a}_d^\\top)^\\top\\)," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.432, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang, B. Liao et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.246, + 0.145, + 0.756, + 0.16 + ], + "angle": 0, + "content": "Table 1: Robustness analysis against multiple levels of Gaussian noise (px)." + }, + { + "type": "table", + "bbox": [ + 0.304, + 0.171, + 0.699, + 0.251 + ], + "angle": 0, + "content": "
noise level (px)0.10.51.01.52.0
Rotation Error7.96e-69.24e-61.50e-51.71e-52.71e-5
Translation Error2.32e-45.93e-48.91e-49.54e-41.55e-3
L-R Error3.36e-33.83e-35.14e-35.22e-36.42e-3
L-D Error2.22e-42.43e-43.90e-44.43e-44.75e-4
" + }, + { + "type": "image", + "bbox": [ + 0.307, + 0.261, + 0.7, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.268, + 0.413, + 0.733, + 0.428 + ], + "angle": 0, + "content": "Fig. 7: Comparison of accuracy and time among different algorithms." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.455, + 0.614, + 0.47 + ], + "angle": 0, + "content": "\\(\\mathbf{P}_d\\), respectively. The evaluation metrics are defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.35, + 0.478, + 0.786, + 0.554 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{c} e _ {\\text {R o t a t i o n}} = \\operatorname {a r c c o s} \\left(\\frac {\\operatorname {T r a c e} \\left(\\mathbf {R} _ {g} ^ {\\top} \\mathbf {R} _ {d}\\right) - 1}{2}\\right) \\\\ e _ {\\text {T r a n s l a t i o n}} = \\operatorname {a r c c o s} \\left(\\mathbf {t} _ {g} ^ {\\top} \\mathbf {t} _ {d} / | | \\mathbf {t} _ {g} | | \\cdot | | \\mathbf {t} _ {d} |\\right) \\\\ e _ {L - R} = \\operatorname {a r c c o s} \\left(\\mathbf {a} _ {g} ^ {\\top} \\mathbf {a} _ {d} / | | \\mathbf {a} _ {g} | | \\cdot | | \\mathbf {a} _ {d} |\\right) \\\\ e _ {L - D} = \\frac {| (\\mathbf {a} _ {g} \\times \\mathbf {a} _ {d}) (\\mathbf {P} _ {d} - \\mathbf {P} _ {g}) |}{\\| \\mathbf {a} _ {g} \\times \\mathbf {a} _ {d} \\|} \\end{array} \\right. \\tag {23}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.571, + 0.452, + 0.586 + ], + "angle": 0, + "content": "4.1 Synthetic Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.593, + 0.788, + 0.7 + ], + "angle": 0, + "content": "Which error term in Sec. 3.2 is better? In the synthetic environment, We set up a cubic box with 12 edges in 3D space and simulated a series of cameras around it to perform RS imaging. The experimental results are shown in Fig. 6. It is evident that, the first type of reprojection error combination performs much better than the second combination in terms of accuracy and time. Therefore, in the forthcoming experiments, we shall exclusively employ the first type of reprojection error combination for comparison." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "How significant is the impact of noise on RSL-BA? Following the same synthetic experimental setup as described above, we additionally introduce Gaussian noise to the endpoints of each line in this experiment. The noise levels are varied from 0.1 pixels to 2.0 pixels, and four different types of errors are recorded for the algorithm at each noise level. For each noise level, 50 experiments are conducted, and the median value is used as the final result. The results, presented in Table 1, demonstrate that RSL-BA provides stable estimations as the noise level increases from 0.1 pixels to 2 pixels." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RSL-BA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 2: The median absolute trajectory error (ATE) of different methods on WHU-RSVI [3] dataset. The best and second results are shown in green and blue, respectively." + }, + { + "type": "table", + "bbox": [ + 0.29, + 0.185, + 0.709, + 0.266 + ], + "angle": 0, + "content": "
WHU-RSVI1WHU-RSVI2WHU-RSVI3WHU-RSVI4
GSBA0.0809920.0613100.0304040.023698
GLBA0.0761730.0659850.0335540.024961
NMRSBA0.0509690.0416290.0423170.028183
NWRSBA0.0406400.0453130.0354510.022666
RLS-BA(ours)0.04435020.0393140.0233510.020675
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.291, + 0.789, + 0.533 + ], + "angle": 0, + "content": "What are the advantages of \\( RSL-BA \\) over point-based methods? Within this part, we will conduct a comparative analysis of our method alongside current SOTA methods, including point- and line-based methods. We set up a cubic box of points in the same positions as the 3D lines, totaling 56 points. To ensure fairness, we set up 8 lines with 7 points each to construct error functions, resulting in a total of precisely 56 points. Detailed experimental settings on the number of lines and the number of sampling points on each line are provided in the supplementary materials. We construct two scenarios: regular and classic scenarios with degeneration parallel to the y-axis. Due to significant differences in the experimental results of different methods, we uniformly take the logarithm (base 10) of the results for better observation. The experimental findings are shown in Fig. 7. It is evident that in both situations, our \\( RSL-BA \\) and the current NWRSBA exhibit comparable accuracy, outperforming alternative approaches based on points and lines. Additionally, the computation time is equivalent to that of NWRSBA. This indicates that the application of lines on Rolling Shutter camera holds significant promise." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.551, + 0.406, + 0.567 + ], + "angle": 0, + "content": "4.2 Synthetic Images" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.573, + 0.789, + 0.77 + ], + "angle": 0, + "content": "In this section, we conduct experiments on input synthetic images. We use the WHU-RSVI [3] dataset, from which we select two sets of data from trajectory1-fast and trajectory2-fast for 3D reconstruction and pose estimation. We first employ [22] to detect RS curves by segmenting curves into multiple short-line segments and performing line fitting for initialization. The GS line-based SfM [16] is applied to initialize the \\( RSL-BA \\) parameters. The comparative methods include GSBA, NMRSBA, NWRSBA, and GSLBA. Table 2 shows the median absolute trajectory error of different methods, it can be observed that the proposed \\( RSL-BA \\) method is the most stable one, achieving optimal or near-optimal results in all cases. Qualitative comparison are also provided in Fig. 8. Unlike point-based methods, line-based methods often achieve good results with fewer feature lines. However, the GSL-BA method is not sufficiently stable when RS effects are prominent." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.789, + 0.365, + 0.804 + ], + "angle": 0, + "content": "4.3 Real Images" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.791, + 0.84 + ], + "angle": 0, + "content": "In this section, we conduct experiments on the real image dataset TUM-RSVI [24]. The experimental setup and comparison methods are similar to those in Sec. 4.2." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.432, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang, B. Liao et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 3: The absolute trajectory error (ATE) comparison of different methods in TUM-RSVI [24] dataset. The best and second results are shown in green and blue, respectively." + }, + { + "type": "table", + "bbox": [ + 0.315, + 0.199, + 0.681, + 0.342 + ], + "angle": 0, + "content": "
GSBAGLBANMRSBANWRSBARSL-BA(ours)
seq10.0694840.0864600.0523660.0450640.037816
seq20.0298210.0303790.0260280.0232270.025132
seq30.0651600.0637180.0579180.0550010.048187
seq40.0496130.0520260.0322140.0305340.031903
seq50.0318600.0358390.0194070.0160660.017659
seq60.0619660.0617920.0324340.0246580.026448
seq70.0516210.0565340.0391540.0396200.039701
seq80.0264030.0286900.0248070.0259260.024983
seq90.0983340.0982120.0824810.0735800.080523
seq100.811740.811800.593900.532960.57477
" + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.357, + 0.694, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.562, + 0.788, + 0.605 + ], + "angle": 0, + "content": "Fig. 8: Comparison of trajectories and 3D reconstruction on WHU-RSVI [3] and TUMRSVI [24] dataset. Each column represents a different bundle adjustment algorithm, and each row represents a different sequence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.634, + 0.788, + 0.697 + ], + "angle": 0, + "content": "Table 3 and Fig 8 present some of the experimental results. As can be seen, in most cases, the \\( RSL-BA \\) method outperforms other methods but is slightly weaker than NwRSBA. This is because the TUM-RSVI lacks line features and they are not visually prominent, making it less suitable for \\( RSL-BA \\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.718, + 0.36, + 0.735 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This paper presents the first solution of line-based RSL-BA. By utilizing points and tangent directions on curves, we have established a series of faster and more robust curve-to-line re-projection errors. The proposed RSL-BA method can prevent three common degeneracies in RSBA, one of which is newly introduced by us. Extensive experiments in real and synthetic data verify the effectiveness and efficiency of the proposed RSL-BA method." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RSL-BA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.255 + ], + "angle": 0, + "content": "This work is supported by the National Key R&D Program of China (No. 2022ZD01190030), Nature Science Foundation of China (No. 62102145), Jiangxi Provincial 03 Special Foundation and 5G Program (Grant No. 20224ABC03A05), Lushan Lab Research Funding, and Changsha Science Fund for Distinguished Young Scholars (kq2306002)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.278, + 0.323, + 0.294 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.785, + 0.338 + ], + "angle": 0, + "content": "1. Albl, C., Sugimoto, A., Pajdla, T.: Degeneracies in rolling shutter sfm. In: ECCV (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.34, + 0.785, + 0.38 + ], + "angle": 0, + "content": "2. Bartoli, A., Sturm, P.: Structure-from-motion using lines: Representation, triangulation, and bundle adjustment. Computer vision and image understanding 100(3), 416-441 (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.382, + 0.785, + 0.408 + ], + "angle": 0, + "content": "3. Cao, L., Ling, J., Xiao, X.: The whu rolling shutter visual-inertial dataset. IEEE Access 8, 50771-50779 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.41, + 0.785, + 0.436 + ], + "angle": 0, + "content": "4. Dai, Y., Li, H., Kneip, L.: Rolling shutter camera relative pose: generalized epipolar geometry. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.438, + 0.785, + 0.464 + ], + "angle": 0, + "content": "5. Duchamp, G., Ait-Aider, O., Royer, E., Lavest, J.M.: A rolling shutter compliant method for localisation and reconstruction. In: VISAPP (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.466, + 0.785, + 0.506 + ], + "angle": 0, + "content": "6. Fisher, A., Cannizzaro, R., Cochrane, M., Nagahawatte, C., Palmer, J.L.: Colmap: A memory-efficient occupancy grid mapping framework. Robotics and Autonomous Systems 142, 103755 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.507, + 0.785, + 0.534 + ], + "angle": 0, + "content": "7. Hartley, R., Zisserman, A.: Multiple view geometry in computer vision. Cambridge university press (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.535, + 0.785, + 0.562 + ], + "angle": 0, + "content": "8. Hedborg, J., Forssen, P.E., Felsberg, M., Ringaby, E.: Rolling shutter bundle adjustment. In: CVPR (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.563, + 0.785, + 0.604 + ], + "angle": 0, + "content": "9. Holynski, A., Geraghty, D., Frahm, J.M., Sweeney, C., Szeliski, R.: Reducing drift in structure from motion using extended features. In: 2020 International Conference on 3D Vision (3DV). pp. 51-60. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.605, + 0.785, + 0.632 + ], + "angle": 0, + "content": "10. Ito, E., Okatani, T.: Self-calibration-based approach to critical motion sequences of rolling-shutter structure from motion. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.785, + 0.659 + ], + "angle": 0, + "content": "1. Lao, Y., Ait-Aider, O.: A robust method for strong rolling shutter effects correction using lines with automatic feature selection. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.66, + 0.785, + 0.688 + ], + "angle": 0, + "content": "2. Lao, Y., Ait-Aider, O., Araujo, H.: Robustified structure from motion with rolling-shutter camera using straightness constraint. Pattern Recognition Letters (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.715 + ], + "angle": 0, + "content": "3. Lao, Y., Ait-Aider, O., Bartoli, A.: Solving rolling shutter 3d vision problems using analogies with non-rigidity. IJCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "4. Liao, B., Qu, D., Xue, Y., Zhang, H., Lao, Y.: Revisiting rolling shutter bundle adjustment: Toward accurate and fast solution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4863-4871 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.798 + ], + "angle": 0, + "content": "5. Lim, H., Jeon, J., Myung, H.: Unconstrained line-based slam using vanishing points for structural mapping. IEEE Robotics and Automation Letters 7(2), 1518-1525 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "6. Liu, S., Yu, Y., Pautrat, R., Pollefeys, M., Larsson, V.: 3d line mapping revisited. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21445-21455 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.31, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.432, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang, B. Liao et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.189 + ], + "angle": 0, + "content": "17. Lourakis, M.I., Argyros, A.A.: Sba: A software package for generic sparse bundle adjustment. ACM Transactions on Mathematical Software (TOMS) 36(1), 1-30 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "18. Micusik, B., Wildenauer, H.: Structure from motion with line segments under relaxed endpoint constraints. International Journal of Computer Vision 124, 65-79 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.273 + ], + "angle": 0, + "content": "19. Moré, J.J.: The levenberg-marquardt algorithm: implementation and theory. In: Numerical analysis: proceedings of the biennial Conference held at Dundee, June 28-July 1, 1977. pp. 105-116. Springer (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.3 + ], + "angle": 0, + "content": "20. Patron-Perez, A., Lovegrove, S., Sibley, G.: A spline-based trajectory representation for sensor fusion and rolling shutter cameras. IJCV (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.342 + ], + "angle": 0, + "content": "21. Pumarola, A., Vakhitov, A., Agudo, A., Sanfeliu, A., Moreno-Noguer, F.: Pl-slam: Real-time monocular visual slam with points and lines. In: 2017 IEEE international conference on robotics and automation (ICRA). pp. 4503-4508. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.787, + 0.37 + ], + "angle": 0, + "content": "22. Purkait, P., Zach, C., Leonardis, A.: Rolling shutter correction in manhattan world. In: ICCV. pp. 882-890 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.787, + 0.411 + ], + "angle": 0, + "content": "23. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4104-4113 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.787, + 0.439 + ], + "angle": 0, + "content": "24. Schubert, D., Demmel, N., von Stumberg, L., Usenko, V., Cremers, D.: Rolling-shutter modelling for direct visual-inertial odometry. In: IROS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "25. Taylor, C.J., Kriegman, D.J.: Structure and motion from line segments in multiple images. IEEE Transactions on Pattern Analysis and Machine Intelligence 17(11), 1021-1032 (1995)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.536 + ], + "angle": 0, + "content": "26. Triggs, B., McLauchlan, P.F., Hartley, R.I., Fitzgibbon, A.W.: Bundle adjustment—a modern synthesis. In: Vision Algorithms: Theory and Practice: International Workshop on Vision Algorithms Corfu, Greece, September 21–22, 1999 Proceedings. pp. 298–372. Springer (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.787, + 0.578 + ], + "angle": 0, + "content": "27. Von Gioi, R.G., Jakubowicz, J., Morel, J.M., Randall, G.: Lsd: A fast line segment detector with a false detection control. IEEE transactions on pattern analysis and machine intelligence 32(4), 722-732 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.632 + ], + "angle": 0, + "content": "28. Wei, D., Wan, Y., Zhang, Y., Liu, X., Zhang, B., Wang, X.: Elsr: Efficient line segment reconstruction with planes and points guidance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15807-15815 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.632, + 0.787, + 0.66 + ], + "angle": 0, + "content": "29. Zhuang, B., Cheong, L.F., Lee, G.H.: Rolling-shutter-aware differential sfm and image rectification. In: ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.787, + 0.702 + ], + "angle": 0, + "content": "30. Zhuang, B., Tran, Q.H., Ji, P., Cheong, L.F., Chandraker, M.: Learning structure-and-motion-aware rolling shutter correction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4551-4560 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.787, + 0.744 + ], + "angle": 0, + "content": "31. Zuo, X., Xie, X., Liu, Y., Huang, G.: Robust visual slam with point and line features. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1775-1782. IEEE (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.744 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/RSL-BA_ Rolling Shutter Line Bundle Adjustment/25dc3962-792b-46d2-8db3-b73a1efe60d9_origin.pdf b/2024/RSL-BA_ Rolling Shutter Line Bundle Adjustment/25dc3962-792b-46d2-8db3-b73a1efe60d9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..431dc65c2db32d97f7f659d0ea44bcfc5124580e --- /dev/null +++ b/2024/RSL-BA_ Rolling Shutter Line Bundle Adjustment/25dc3962-792b-46d2-8db3-b73a1efe60d9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e9f5a461dc0be14f23e986d1ac401f00613083f628e41db5ff6f1c2d4ed9e137 +size 8540121 diff --git a/2024/RSL-BA_ Rolling Shutter Line Bundle Adjustment/full.md b/2024/RSL-BA_ Rolling Shutter Line Bundle Adjustment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5142dc579fdf1083bf3caa715cd43de3cdb3e2e7 --- /dev/null +++ b/2024/RSL-BA_ Rolling Shutter Line Bundle Adjustment/full.md @@ -0,0 +1,375 @@ +# RSL-BA: Rolling Shutter Line Bundle Adjustment + +Yongcong Zhang $^{1,3,*}$ , Bangyan Liao $^{2,*}$ , Yifei Xue $^{1,3}$ , Chen Lu $^{4}$ , Peidong Liu $^{2}$ , and Yizhen Lao $^{1,\dagger}$ + +$^{1}$ Hunan University $^{2}$ Westlake University $^{3}$ DaHe.AI $^{4}$ Dreame Technology Company + +Abstract. The line is a prevalent element in man-made environments, inherently encoding spatial structural information, thus making it a more robust choice for feature representation in practical applications. Despite its apparent advantages, previous rolling shutter bundle adjustment (RSBA) methods have only supported sparse feature points, which lack robustness, particularly in degenerate environments. In this paper, we introduce the first rolling shutter line-based bundle adjustment solution, $RSL-BA$ . Specifically, we initially establish the rolling shutter camera line projection theory utilizing Plücker line parameterization. Subsequently, we derive a series of reprojection error formulations which are stable and efficient. Finally, we theoretically and experimentally demonstrate that our method can prevent three common degeneracies, one of which is first discovered in this paper. Extensive synthetic and real data experiments demonstrate that our method achieves efficiency and accuracy comparable to existing point-based rolling shutter bundle adjustment solutions. + +Keywords: Rolling Shutter $\cdot$ Bundle Adjustment + +# 1 Introduction + +Bundle Adjustment (BA) is a crucial step in multi-view 3D reconstruction, as it jointly optimizes camera poses and scene structure. Although feature point-based BA [6,23,26] currently dominates the academic landscape, line-based BA has been gaining increasing attention from researchers [16,28]. Lines, being one of the most common features in man-made environments, effectively describe the structural information of 3D scenes, in contrast to points which rely solely on local image patches. This structural information provides robust constraints for visual reconstruction [27], making line a more suitable choice in challenging scenarios. + +In parallel with classical bundle adjustment [26], several methods [1,13] have incorporated the rolling shutter camera model to simultaneously estimate camera + +poses, structure, and instantaneous motion parameters. However, all existing rolling shutter bundle adjustment (RSBA) methods rely solely on point features, while line features have never been addressed in RSBA. We identify two challenges that hinder the direct combination of line-based BA and the rolling shutter camera model. + +The first challenge arises from the time-dependent exposure of RS cameras. Unlike CCD cameras and their global shutter (GS) counterparts, RS cameras capture images in a scanline-by-scanline manner. Consequently, as illustrated in Fig. 1, images taken by moving RS cameras exhibit distortions known as the RS effect. This effect curves the projection of each 3D straight line, making it non-trivial to directly transfer the line-based BA to the rolling shutter camera setting. + +The second challenge is associated with degeneracy. As shown in [1, 30], degeneracy is one of the main obstacles in RSBA since iterative optimization can easily be collapsed in some degenerate solutions which are far from the ground truth solutions. Although several strategies have been proposed to mitigate degeneracy in RSBA [1, 12, 14], a specific strategy to address degeneracy in line-based RSBA is still lacking. + +In conclusion, an accurate and robust solution to line-based RSBA is still missing. Such a method would be vital in the potential widespread deployment of 3D vision with RS imaging systems. + +# 1.1 Related Works + +Video-based RSBA. The assumption of smooth continuous trajectories is widely employed for RS video inputs to reduce the optimization parameter space and enhance algorithm robustness. In [8], Hedborg et al. present an RSBA algorithm that interpolates the motion between consecutive frames. Zhuang et al. [29] further propose an optical flow-based RSBA, which are developed to recover the relative pose of an RS camera that undergoes constant velocity and acceleration motion, respectively. Following this assumption, a spline-based camera trajectory motion model is proposed by [20]. + +Unordered RSBA. As the unordered image set is the standard input for SfM, Albl et al. first addresses this setting by explicitly velocity modeling and optimization. Besides, a planar degeneracy configuration has also been disclosed in [1]. While in [12], Lao et al. propose a camera-based RSBA to simulate the actual camera projection, exhibiting the degeneracy resilience ability. In [10], Ito et al. establish the equivalence between self-calibrated SfM and RSSFM based on the pure rotation instantaneous motion model and affine camera assumption, while the work of [13] draws the equivalence between RSSfM and non-rigid SfM. Recently, Liao et al. proposed two techniques to boost the accuracy and robustness of RSBA in [14]. + +Line-based GSBA. In [2, 25], the authors propose a comprehensive line-based SFM pipeline, while [9, 18] introduce line-based incremental SFM frameworks. Wei et al. in [28] employs planes and points to guide matching, achieving efficient line reconstruction. Recently, Lui et al. introduced LIMAP in [16], where the + +![](images/0f844bdff8bc036f12a958712c13ee7c6449121cf578752f82753cc210d6aa0d.jpg) + +![](images/c7a70a41c52a0b2a6bccf64dce3bcc1c6cf2f636c2948a8ad504aba5ed6fc62c.jpg) +input + +![](images/2851336b6560c56be37b00028c47265d39d9850e08aaa71de39a2bd6db22a881.jpg) +Fig. 1: The pipeline of proposed RSL-BA with some example results. Starting with the input of some RS images (left), the optimization of camera poses and 3D line coordinates is performed using proposed RSL-BA (middle), yielding well-reconstructed 3D line segments and accurately estimated camera poses (right). + +![](images/b0bbc13a2a053ad2dbff8f56de0c1f750023518efb77dea927cf1cea0f815aa4.jpg) +output + +authors devised various strategies and algorithms to achieve efficient and precise line reconstruction. In [15, 21, 31], hybrid strategies combining lines and feature points are utilized to enhance the accuracy and robustness of SLAM. + +To this end, we propose a novel $RSL-BA$ algorithm to solve the line-based RSBA problem. Specifically, we first establish the rolling shutter line projection theory using the Plücker line parameterization and subsequently derive a series of reprojection error formulations. We claim that the proposed reprojection error formulations are efficient and exhibit the degeneration-resistant ability. We also provide complete degeneration-resistant proof of our proposed error formulation. Our contributions are summarized as follows: + +- To the best of our knowledge, this is the first line-based RSBA solution, which serves as the foundation for line-based or point-line-hybrid RS-SfM and RS-SLAM. +- we theoretically and experimentally demonstrate that the proposed $RSLBA$ can prevent three common degeneracies in RSBA, one of the degenerate cases is first discovered in this paper. +- The extensive evaluations in both synthetic and real datasets exhibit the comparable efficiency and accuracy of the proposed method over previous works. + +# 2 Background + +In this section, we review the rolling shutter projection model in Sec. 2.1 and the parameterization of 3D lines in Sec. 2.2. + +# 2.1 Rolling Shutter Camera Projection Mode + +Let $\mathbf{R}(v) \in \mathbf{SO}(3)$ and $\mathbf{t}(v) \in \mathbb{R}^3$ represent the camera rotation and translation, respectively, when the row index $v$ of measurement is acquired. Denote the intrinsic camera parameters as $\mathbf{K}$ , and the initial rotation and translation of the + +camera at $v = 0$ as $\mathbf{R}_0$ and $\mathbf{t}_0$ , respectively. The projection matrix of the global shutter camera is defined as $\mathbf{P}_0 = \mathbf{K}[\mathbf{R}_0, \mathbf{t}_0]$ . Given the assumption of constant camera motion during frame capture, which is commonly made in RS 3D vision tasks [1,4,13], we can model the instantaneous motion as: + +$$ +\mathbf {R} (v) = (\mathbf {I} + [ \boldsymbol {\omega} ] _ {\times} v) \mathbf {R} _ {0}, \quad \mathbf {t} (v) = \mathbf {t} _ {0} + \mathbf {d} v, \tag {1} +$$ + +where $\mathbf{d} = [d_x, d_y, d_z]^\top$ is the translational velocity vector and $\boldsymbol{\omega} = [\omega_x, \omega_y, \omega_z]^\top$ is the rotational velocity vector, and $[\boldsymbol{\omega}]_\times$ represents the skew-symmetric matrix of vector $\boldsymbol{\omega}$ . Then, the projection matrix of the camera under the RS model can be defined as: + +$$ +\mathbf {P} _ {v} = \mathbf {K} \left[ \mathbf {R} _ {v}, \mathbf {t} _ {v} \right] = \mathbf {P} _ {0} + v \mathbf {Q} \\ \mathbf {K} ^ {[ r ]} = \left[ \begin{array}{l l l l l} \mathbf {P} _ {v} & \mathbf {t} _ {v} & \mathbf {Q} _ {v} & \mathbf {I} ^ {[ r ]} & \mathbf {D} _ {v} \end{array} \right] \tag {2} +$$ + +$$ +\mathbf {P} _ {0} = \mathbf {K} \left[ \mathbf {R} _ {0}, \mathbf {t} _ {0} \right] \quad \mathbf {Q} = \mathbf {K} \left[ \left[ \boldsymbol {\omega} \right] _ {\times} \mathbf {R} _ {0}, \mathbf {d} \right]. +$$ + +# 2.2 3D Line Segments Parameterization + +We employ the Plücker matrix to parameterize the 3D line [7]. Specifically, given two homogeneous 3D points $P_{a} = [a_{1}, a_{2}, a_{3}, 1]^{\top}$ and $P_{b} = [b_{1}, b_{2}, b_{3}, 1]^{\top}$ . The Plücker matrix $\mathbf{L}$ can be defined as: + +$$ +\mathbf {L} = P _ {a} P _ {b} ^ {\top} - P _ {b} P _ {a} ^ {\top} +$$ + +$$ += \left[ \begin{array}{c c c c} 0 & a _ {1} b _ {2} - a _ {2} b _ {1} & a _ {1} b _ {3} - a _ {3} b _ {1} & a _ {1} - b _ {1} \\ a _ {2} b _ {1} - a _ {1} b _ {2} & 0 & a _ {2} b _ {3} - a _ {3} b _ {2} & a _ {2} - b _ {2} \\ a _ {3} b _ {1} - a _ {1} b _ {3} & a _ {3} b _ {2} - a _ {2} b _ {3} & 0 & a _ {3} - b _ {3} \\ b _ {1} - a _ {1} & b _ {2} - a _ {2} & b _ {3} - a _ {3} & 0 \end{array} \right] = \left[ \begin{array}{c c c c} 0 & l _ {1 2} & l _ {1 3} & l _ {1 4} \\ - l _ {1 2} & 0 & l _ {2 3} & l _ {2 4} \\ - l _ {1 3} & - l _ {2 3} & 0 & l _ {3 4} \\ - l _ {1 4} & - l _ {2 4} & - l _ {3 4} & 0 \end{array} \right], \tag {3} +$$ + +where the $[l_{14}, l_{24}, l_{34}]^{\top}$ represents the direction of this straight line while the $[-l_{23}, l_{13}, -l_{12}]^{\top}$ represents the normal direction of this line. The projection of 3D Plücker matrix is the skew-symmetric matrix of a 2D normalized line. It can be related to each other as: + +$$ +\left[ \mathbf {l} _ {g s} \right] _ {\times} = \left[ \mathbf {R}, \mathbf {t} \right] \mathbf {L} \left[ \mathbf {R}, \mathbf {t} \right] ^ {\top} \tag {4} +$$ + +# 3 Methodology + +In Sec. 3.1, we initially establish the rolling shutter projection equations using the Plücker line parameterization. These equations establish a connection between the 3D lines and the curve parameters on the image plane. Then in Sec. 3.2, we present a series of line-based reprojection error equations. On top of the error definition, we can formulate the entire rolling shutter line bundle adjustment problem in Sec. 3.3. Finally, in Sec. 3.4, we demonstrate the degeneracy resilience capability of the proposed error equations. + +# 3.1 Rolling Shutter Line Projection Formulation + +As previously explained, the curved rolling shutter projection of 3D lines occurs because of the camera's nonlinear motion. Using the Plücker parameterization method, we can derive explicit expressions for the curve parameters. + +Firstly, we construct the skew-symmetric matrix $\left[\mathbf{l}_{rs}\right]_{\times}$ of the instantaneous projection line by combining Eq. (2) and Eq. (4): + +$$ +\begin{array}{l} \left[ \mathbf {l} _ {r s} \right] _ {\times} = \mathbf {P} _ {v} \mathbf {L} \mathbf {P} _ {v} ^ {\top} \\ = \mathbf {P} _ {0} \mathbf {L} \mathbf {P} _ {0} ^ {\top} + v \left(\mathbf {P} _ {0} \mathbf {L} \mathbf {Q} ^ {\top} + \mathbf {Q} \mathbf {L} \mathbf {P} _ {0} ^ {\top}\right) + v ^ {2} \mathbf {Q} \mathbf {L} \mathbf {Q} ^ {\top} \tag {5} \\ = \mathbf {A} _ {1} + v \mathbf {A} _ {2} + v ^ {2} \mathbf {A} _ {3} \\ \end{array} +$$ + +Secondly, as $\mathbf{l}_{rs} = [l_1,l_2,l_3]^\top$ and $[\mathbf{l}_{rs}]_{\times} = \mathbf{A}_1 + v\mathbf{A}_2 + v^2\mathbf{A}_3$ , we can determine the instantaneous projected line parameter as follows: + +$$ +\left\{ \begin{array}{l} l _ {1} = \mathbf {A} _ {1} ^ {3 2} + v \mathbf {A} _ {2} ^ {3 2} + v ^ {2} \mathbf {A} _ {3} ^ {3 2} \\ l _ {2} = \mathbf {A} _ {1} ^ {1 3} + v \mathbf {A} _ {2} ^ {1 3} + v ^ {2} \mathbf {A} _ {3} ^ {1 3} \\ l _ {3} = \mathbf {A} _ {1} ^ {2 1} + v \mathbf {A} _ {2} ^ {2 1} + v ^ {2} \mathbf {A} _ {3} ^ {2 1} \end{array} \right. \tag {6} +$$ + +Ultimately, we note that the instantaneous projected line parameter fulfills this $l_{1}u + l_{2}v + l_{3} = 0$ equation, which we can expand to obtain the final curve equation, denoted as $\pmb{\iota}$ : + +$$ +\begin{array}{l} \iota : \mathbf {A} _ {3} ^ {1 3} v ^ {3} + \mathbf {A} _ {3} ^ {3 2} u v ^ {2} + \left(\mathbf {A} _ {2} ^ {1 3} + \mathbf {A} _ {3} ^ {2 1}\right) v ^ {2} + \mathbf {A} _ {2} ^ {3 2} u v \\ + \left(\mathbf {A} _ {1} ^ {1 3} + \mathbf {A} _ {2} ^ {2 1}\right) v + \mathbf {A} _ {1} ^ {3 2} u + \mathbf {A} _ {1} ^ {2 1} = 0 \\ \end{array} +$$ + +Our conclusion aligns with that of Lao et al. [11], yielding an identical polynomial curve degree. Nevertheless, our utilization of a 4-DoF orthogonal representation representation for 3D lines has distinct benefits when compared to Lao's 5-DoF over-parameterized representation. These advantages include the capability to conduct unconstrained optimization on spatial lines directly. + +Remark. The geometric interpretation underlying this derivation is quite straightforward. Given the parameters of a spatial line and the current camera pose, if point $(\mathbf{u},\mathbf{v})$ lies on the curve obtained by projecting this spatial line, then when the RS camera scans to the v-th row, the projection of the spatial line onto the image plane follows Eq. (6). Therefore, it is evident that point $(\mathbf{u},\mathbf{v})$ must be located on this line. Based on this reasoning, we can deduce Eq. (7). + +# 3.2 Line-based Reprojection Errors + +As claimed in Sec. 3.1, a 3D straight line will become a polynomial curve on the captured image. In this section, we will establish a series of re-projection errors which are stable and robust. We first leverage the similar intuition in Sec. 3.1 to obtain the distance error representation. As shown in Fig. 2(a), Let there be a point $\mathbf{q} = [u,v]^{\top}$ on the observed curve $\iota$ , then the pose of the camera when exposing that point is $\mathbf{R}_v$ , $\mathbf{t}_v$ , then projecting the 3D line $\mathbf{L}$ onto the image + +![](images/4f4d340e18b248320f6904fb07ad6a461bb1487e0d9e246341b03ccbf67c238a.jpg) +Fig. 2: The illustration of proposed line-based reprojection errors given the spatial line $L$ and observed curve $\iota$ . Instead of directly measuring the distance between the observed curve and the projected curve, we split each curve into multiple points $q$ , and aggregate the distance errors. Specifically, we first utilize the virtual projected line $l$ when observing it, as shown in (a). To measure the distance between each point $q$ and the corresponding virtual lines, we can use the perpendicular distance (b), the horizontal distance (c,d) or the tangent distance (e) as our basic metrics. + +plane, this results in the virtual 2D line $\mathbf{1} = [l_1, l_2, l_3]^\top$ . Under the condition of ground truth, this virtual line must theoretically intersect with that point $\mathbf{q}$ . This constraint directly leads us to the two variants of distance error, which we have defined below. + +Perpendicular Distance Error $\mathbf{e}_{d1}$ . The first distance error formulation, named perpendicular distance error, measures the shortest distance from point $\mathbf{q}$ to the line $\mathbf{l}$ , as shown in Fig. 2(b). The error is defined as: + +$$ +\mathbf {e} _ {d 1} = \frac {l _ {1} u + l _ {2} v + l _ {3}}{\sqrt {l _ {1} ^ {2} + l _ {2} ^ {2}}}, \tag {8} +$$ + +where the line parameters $[l_1, l_2, l_3]$ can be calculated through Eq. (6). + +Horizontal Distance Error $\mathbf{e}_{d2}$ . The second distance error formulation, named horizontal distance error, measures the u-axis distance from point $\mathbf{q}$ to the line $\mathbf{l}$ , as shown in Fig. 2(c). The error can be defined as: + +$$ +\mathbf {e} _ {d 2} = u - u ^ {\prime}, \tag {9} +$$ + +where the corresponding $u'$ coordinate of virtual line 1 at the $v$ -th row can be calculated as: + +$$ +u ^ {\prime} = - \frac {l _ {2} v + l _ {3}}{l _ {1}}, \tag {10} +$$ + +As shown in Fig. 2(d), the $e_{d2}$ distance error is equivalent to the u-axis distance from the point to the projection curve. + +Tangent Error $\mathbf{e}_t$ . While distance error provides an elegant constraint, it is insufficient to rely solely on distance error in certain scenarios. To further enhance the robustness, we introduce a second type of tangent error. As shown in Fig. 2(e), we constrain the tangent of the projected curve $\iota'$ to match the observed tangent direction. The error can be defined as: + +$$ +\mathbf {e} _ {t} = \mathbf {s} ^ {\top} \mathbf {s} ^ {\prime}, \tag {11} +$$ + +where $\mathbf{s}$ represents the tangent direction observation. The $\mathbf{s}'$ can be calculated analytically following the implicit function theorem as: + +$$ +\mathbf {s} ^ {\prime} = \left[ \frac {1}{\sqrt {1 + s ^ {2}}}, \frac {s}{\sqrt {1 + s ^ {2}}} \right], \quad s = \frac {\partial u}{\partial v} = - \frac {\partial \iota / \partial v}{\partial \iota / \partial u} \tag {12} +$$ + +# 3.3 Rolling Shutter Line Bundle Adjustment + +The non-linear least squares solvers are used to find an optimal solution $\pmb{\theta}^{*}$ including camera poses $\mathbf{R}^{*},\mathbf{t}^{*}$ , instantaneous motion $\omega^{*},\mathbf{d}^{*}$ and 3D lines $\pmb{\tau}^{*}$ by minimizing the reprojection error $\mathbf{e}_i^j$ from line $i$ to camera $j$ over all the camera index in set $\mathcal{F}$ and corresponding observed 3D lines index in subset $\mathcal{P}_j$ : + +$$ +\boldsymbol {\theta} ^ {*} = \left\{\boldsymbol {\tau} ^ {*}, \mathbf {R} ^ {*}, \mathbf {t} ^ {*}, \boldsymbol {\omega} ^ {*}, \mathbf {d} ^ {*} \right\} = \arg \min _ {\boldsymbol {\theta}} \sum_ {j \in \mathcal {F}} \sum_ {i \in \mathcal {P} _ {j}} \left\| \mathbf {e} _ {i} ^ {j} \right\| _ {2} ^ {2}. \tag {13} +$$ + +On top of errors proposed in Sec. 3.2, we provide two variants of reprojection error: + +$$ +\mathbf {e} _ {1} = \mathbf {e} _ {d 1} + \lambda_ {1} \mathbf {e} _ {t}, \quad \mathbf {e} _ {2} = \mathbf {e} _ {d 2} + \lambda_ {2} \mathbf {e} _ {t}, \tag {14} +$$ + +where $\lambda_1, \lambda_2$ are the relative weight. We have experimentally validated the optimality of these two error variants in Sec. 4.1. After definition, this non-linear least square problem can be iteratively solved through Levenberg-Marquardt [19] method. + +Besides, the full analytical Jacobian has been derived in the supplementary material. + +# 3.4 Degeneracy Analysis + +In this section, we conducted theoretical analysis on the resilience of $RSL-BA$ to three types of degeneracy scenarios. These three degeneracy scenarios include plane degeneracy case [1], 2-vIEWS pure translation case [30], and X-Y pure translation degeneracy case disclosed by us. We assume the positive direction of X Y Z axis is right down forward, respectively. + +(1) Resistance to Plane Degeneracy [1]. We first briefly describe the conditions under which the plane degeneration occurs: when the camera's y-axis is parallel, if a rotational velocity $\omega = [-1,0,0]^{\top}$ is given to the camera, then the + +reconstructed 3D points will be compressed onto a plane perpendicular to the $y$ -axis, and satisfy that the reprojection error is 0. + +For line features, we also assume $\mathbf{R}_0 = \mathbf{I}$ , $\mathbf{t}_0 = \mathbf{0}$ , $\boldsymbol{\omega} = [-1,0,0]^\top$ , $\mathbf{d} = 0$ . Let there be two points, $\mathbf{P}_a$ and $\mathbf{P}_b$ , on the spatial line $\mathbf{L}$ . $\mathbf{P}_a = [a_1,a_2,a_3,1]^\top$ and $\mathbf{P}_b = [b_1,b_2,b_3,1]^\top$ . When the space is compressed into a plane, the new coordinates of $\mathbf{P}_a$ and $\mathbf{P}_b$ are $\mathbf{P}_a' = [a_1,0,a_3,1]^\top$ and $\mathbf{P}_b' = [b_1,0,b_3,1]^\top$ , respectively. By referring to Sec. 3.1, we can solve for the projection parameters of the two-dimensional curve at this moment. The parameters of the curve at this time are: + +$$ +\left\{ \begin{array}{l} \mathbf {A} _ {3} ^ {1 3} = \mathbf {A} _ {3} ^ {3 2} = \mathbf {A} _ {2} ^ {1 3} = \mathbf {A} _ {3} ^ {2 1} = \mathbf {A} _ {2} ^ {3 2} = \mathbf {A} _ {1} ^ {3 2} = \mathbf {A} _ {1} ^ {2 1} = 0 \\ \mathbf {A} _ {1} ^ {1 3} = a _ {3} b _ {1} - a _ {1} b _ {3} \\ \mathbf {A} _ {2} ^ {2 1} = a _ {1} b _ {3} - a _ {3} b _ {1} \end{array} \right. \tag {15} +$$ + +Substituting the parameters into the curve expression Eq. (7), we obtain $0 = 0$ , indicating that u and v are unrestricted. Under the influence of $\omega = [-100]^{\top}$ , the camera's imaging rows are always coplanar with the line, making the distance from any point on the image to the line zero and satisfying the point-to-line distance error function (Sec. 3.2). However, the slope constraint cannot be met. Using this projection equation to calculate the tangent direction of any point on the curve results in indeterminate forms $[0/0, 0/0]$ , which do not match the measured tangent direction. Therefore, using the tangent direction of points on the curve as the error function prevents this degeneration. As verified in Fig 5, our proposed RSL-BA can suppress plane degeneration, whereas conventional rolling shutter point-based bundle adjustment method NMRsBA [1] converges to a degenerate solution. + +(2) Resistance to 2-views Pure Translation Degeneracy [30]. We first briefly describe the conditions under which this degeneration occurs: when the camera moves in a straight line to take two pictures, assigning any translational velocity to the camera for both shots, it is possible to find new 3D points that satisfy the condition that the reprojection error of the points under this velocity is zero. + +Let's assume the real pose of the camera before the degeneration is: $\mathbf{R}_0 = \mathbf{I}$ , $\mathbf{t}_0 = \mathbf{0}$ , $\boldsymbol{\omega} = [0,0,0]^\top$ , $\mathbf{d} = [d_1,d_2,d_3]^\top$ , $\mathbf{L} = (l_{12},l_{13},l_{23},l_{14},l_{24},l_{34})$ In this case, substitute these parameters into Eq. (7) we can get curve expression: + +$$ +\left(l _ {1 4} d _ {3} - l _ {3 4} d _ {1}\right) v ^ {2} + \left(l _ {3 4} d _ {2} - l _ {2 4} d _ {3}\right) u v + \left(l _ {2 4} d _ {1} - l _ {1 4} d _ {2} + l _ {1 3}\right) v - l _ {2 3} u - l _ {1 2} = 0 \tag {16} +$$ + +After the degeneration occurs, $\mathbf{R}_0^{\prime} = \mathbf{I}$ , $\mathbf{t}_0^{\prime} = \mathbf{0}$ , $\boldsymbol{\omega}^{\prime} = [0\quad 0\quad 0]^\top$ , $\mathbf{d}' = r[d_1,d_2,d_3]$ , $\mathbf{L} = (l_{12}',l_{13}',l_{23}',l_{14}',l_{24}',l_{34}')$ . In this case, the curve expression is: + +$$ +r \left(l _ {1 4} ^ {\prime} d _ {3} - l _ {3 4} ^ {\prime} d _ {1}\right) v ^ {2} + r \left(l _ {3 4} ^ {\prime} d _ {2} - l _ {2 4} ^ {\prime} d _ {3}\right) u v + \left(r \left(l _ {2 4} ^ {\prime} d _ {1} - l _ {1 4} ^ {\prime} d _ {2}\right) + l _ {1 3} ^ {\prime}\right) v - l _ {2 3} ^ {\prime} u - l _ {1 2} ^ {\prime} = 0 \tag {17} +$$ + +To make the reprojection error zero, the curve expressions before and after degeneration must be the same, meaning the corresponding coefficients are in + +![](images/d28cd589b623bfcde50423900425887091968c032fa33d960b982eecce6e6752.jpg) +(a) + +![](images/0408ba474de9c4e0afae8cf88006c86e329f3306c9b0f8d2e32593c709f8dec1.jpg) +(b) +Fig. 3: The illustration of degeneracy resistance of our proposed method in 2-views pure translation degeneracy. (a) Degeneration in single-view scenarios. (b) Suppression of degeneration using line features in dual-view scenarios. (c) Inability of line features to suppress degeneration in special dual-view scenarios + +![](images/b0fc43be8c524d8d4afeaa57a2ba0871d77002a78e9187ed1d9a53ccbad4c942.jpg) +(c) + +proportion. + +$$ +\left\{ \begin{array}{l} l _ {1 4} d _ {3} - l _ {3 4} d _ {1} = s r \left(l _ {1 4} ^ {\prime} d _ {3} - l _ {3 4} ^ {\prime} d _ {1}\right) \\ l _ {3 4} d _ {2} - l _ {2 4} d _ {3} = s r \left(l _ {3 4} ^ {\prime} d _ {2} - l _ {2 4} ^ {\prime} d _ {3}\right) \\ l _ {2 4} d _ {1} - l _ {1 4} d _ {2} + l _ {1 3} = s r \left(l _ {2 4} ^ {\prime} d _ {1} - l _ {1 4} ^ {\prime} d _ {2}\right) + s l _ {1 3} ^ {\prime} \\ l _ {2 3} = s l _ {2 3} ^ {\prime} \\ l _ {1 2} = s l _ {1 2} ^ {\prime} \end{array} \right. \tag {18} +$$ + +Substituting the first and second equations into the third equation yields $l_{13} = sl_{13}'$ . Combining this with: + +$$ +\left\{ \begin{array}{l} l _ {1 2} l _ {3 4} - l _ {1 3} l _ {2 4} + l _ {1 4} l _ {2 3} = 0 \\ l _ {1 2} ^ {\prime} l _ {3 4} ^ {\prime} - l _ {1 3} ^ {\prime} l _ {2 4} ^ {\prime} + l _ {1 4} ^ {\prime} l _ {2 3} ^ {\prime} = 0 \end{array} \right. \tag {19} +$$ + +we obtain: + +$$ +l _ {1 2} ^ {\prime} = \frac {l _ {1 2} ^ {\prime}}{s}, \quad l _ {1 3} ^ {\prime} = \frac {l _ {1 3}}{s}, \quad l _ {2 3} ^ {\prime} = \frac {l _ {2 3}}{s}, \quad l _ {1 4} ^ {\prime} = \frac {l _ {1 4}}{s r}, \quad l _ {2 4} ^ {\prime} = \frac {l _ {2 4}}{s r}, \quad l _ {3 4} ^ {\prime} = \frac {l _ {3 4}}{s r} \tag {20} +$$ + +Note that the direction and normal direction of the line remain unchanged, indicating that the degenerated line stays in the original plane and parallel to the original line, as shown in Fig. 3(1), where $\mathbf{L}^{gt}$ is the original line, and $\mathbf{L}_1^{dg}$ is the degenerated line. In the dual-frame scenario (Fig.3(2)), when the plane $\mathbf{O}_1\mathbf{L}^{gt}$ is not coplanar with the plane $\mathbf{O}_2\mathbf{L}^{gt}$ , their degenerated states $\mathbf{L}_1^{dg}$ and $\mathbf{L}_2^{gt}$ cannot coincide (except at $\mathbf{L}^{gt}$ ). In this case, line constraints can prevent degeneration. However, when $\mathbf{L}^{gt}$ , $\mathbf{O}_1$ , and $\mathbf{O}_2$ are coplanar (Fig.3(3)), line constraints cannot suppress degeneration, though this is a very special condition requiring all lines to meet this criterion. This demonstrates that line constraints are more effective than point constraints in preventing degeneration. + +(3) Resistance to X-Y Pure Translation Degeneracy (our new founding). We present a pure translation scenario without the z-direction, suitable for + +![](images/fc83d709053f01e4ee4124c0fa43a052c89b57a4c14e044d1c43d3d9fbbd21fc.jpg) +(a) + +![](images/24b87dcae962ef715b22a93d42baa8cc525ad05238c265e9c7b49408574d10fa.jpg) +(b) + +![](images/deef5e9ffb831fb95b02fb72e98086c7816d501055585862ea7cdfd5e54f60e7.jpg) +(c) +Fig. 4: In the case of X-Y pure translation, optimization compresses points into a straight line. In a single view, pixel points parallel to the motion trajectory backproject to the same 3D point ((a) and (b)). (c) Different intersection points in space are equidistant from the camera plane. (d) All intersection points formed by backprojection constitute a straight line. + +![](images/768fe78bc1be96a55cf58fc83086a3e73383f5796a3bbe297b54c3ccf6aaa99a.jpg) +(d) + +multi-view BA, compressing space points into a straight line. Unlike the second type of degeneration, this does not depend on the number of images. + +Firstly, considering a single view, as shown in Fig.4(1), during exposure, the camera moves from position $\mathbf{O}_1^1$ to $\mathbf{O}_1^3$ , where pixel $\mathbf{q}_1^1$ is located on the first row and $\mathbf{q}_1^3$ is located on the $H$ -th row. Their corresponding spatial points are located on the rays $\mathbf{O}_1^1\mathbf{q}_1^1$ and $\mathbf{O}_1^3\mathbf{q}_1^3$ , intersecting at point $\mathbf{P}_1$ . Actually, any point on the line $\mathbf{q}_1^1\mathbf{q}_1^3$ intersects with its corresponding 3D point at $\mathbf{P}_1$ during exposure. Let $\mathbf{O}_1^2$ be the camera position when exposed to the $h_1^2$ -th row, with corresponding feature point $\mathbf{q}_1^2$ . We only consider the plane $\mathbf{O}_1^1\mathbf{P}_1\mathbf{O}_1^3$ in Fig.4(2). The motion of $\mathbf{O}$ is uniform rectilinear motion thus, there exists a length relationship: + +$$ +\left\{ \begin{array}{c} l _ {\mathbf {q} _ {1} ^ {1} \mathbf {q} _ {1} ^ {2}} = \frac {h _ {1} ^ {2}}{H} l _ {\mathbf {q} _ {1} ^ {1} \mathbf {q} _ {1} ^ {3}}, \quad l _ {\mathbf {q} _ {3} ^ {1} \mathbf {q} _ {1} ^ {2}} = \left(1 - \frac {h _ {1} ^ {2}}{H}\right) l _ {\mathbf {q} _ {1} ^ {1} \mathbf {q} _ {1} ^ {3}} \\ l _ {\mathbf {O} _ {1} ^ {1} \mathbf {O} _ {1} ^ {2}} = \frac {h _ {1} ^ {2}}{H} l _ {\mathbf {O} _ {1} ^ {1} \mathbf {O} _ {1} ^ {3}}, \quad l _ {\mathbf {O} _ {3} ^ {1} \mathbf {O} _ {1} ^ {2}} = \left(1 - \frac {h _ {1} ^ {2}}{H}\right) l _ {\mathbf {O} _ {1} ^ {1} \mathbf {O} _ {1} ^ {3}} \end{array} \right. \tag {21} +$$ + +so we have: + +$$ +\frac {l _ {\mathbf {q} _ {1} ^ {1} \mathbf {q} _ {1} ^ {2}}}{l _ {\mathbf {O} _ {1} ^ {1} \mathbf {O} _ {1} ^ {2}}} = \frac {l _ {\mathbf {q} _ {1} ^ {3} \mathbf {q} _ {1} ^ {2}}}{l _ {\mathbf {O} _ {3} ^ {1} \mathbf {O} _ {1} ^ {2}}} = \frac {l _ {\mathbf {q} _ {1} ^ {1} \mathbf {q} _ {1} ^ {3}}}{l _ {\mathbf {O} _ {1} ^ {1} \mathbf {O} _ {1} ^ {3}}} \tag {22} +$$ + +Therefore, the ray $\mathbf{O}_1^2\mathbf{q}_1^2$ will also intersect at point $\mathbf{P_1}$ . For every pixel on a line parallel to $\mathbf{O}_1^1\mathbf{O}_1^3$ , it can be traced back to a different spatial point. For pixel points on different lines parallel to the motion trajectory, as shown in Fig. 4(3), because the distance $l_{q^1 q^3}$ is the same and the camera focal length remains unchanged, according to similar triangles, these spatial points have equal distances to the image plane. Simultaneously, due to symmetry, these points also lie in the plane that is perpendicular to and bisects the line segment $\mathbf{O}_1^1\mathbf{O}_1^2$ . Therefore, the intersection line of these two planes represents the final degenerated configuration, which is a line parallel to the x-axis, as shown in Fig.4(4). In multiple views, with no z-direction movement, the imaging planes of different cameras coincide. We assign the same camera configuration to different views, including rotation, translation, angular velocity, and linear velocity. From the single-view case, corresponding points on the pixel plane parallel to the motion trajectory reconstruct to the same spatial point, with zero reprojection error. + +![](images/2a7dacdd8b01229a62e27383d3123c2cb57b2ca4b68e4fa1b139f1c753c76f63.jpg) +Fig.5: Comparison of degenerate resistance ability between NMRSBA [1] and proposed $RSL-BA$ . Ground truth camera poses and structure are colored with black. This example illustrates that our proposed $RSL-BA$ has the resistance ability against the plane and X-Y pure translation degeneracy. + +![](images/9cfe949e393723ebaef0030e53b19349f3de2f40233bede2382dd9015eb1b8fa.jpg) +Fig. 6: Accuracy and computational time comparison of the proposed two combinations of reprojection errors. + +For the line constraints proposed in this paper, $RSL-BA$ can still effectively suppress degeneration. The proof process is similar to the previous section Sec. 3.4, with the only modification being setting $d_{3} = 0$ in the camera's motion direction. As verified in Fig 5, our proposed $RSL-BA$ can suppress X-Y pure translation degeneration, whereas conventional rolling shutter point-based bundle adjustment method NMRSBA [1] converges to a degenerate solution. + +# 4 Experiments + +We compare our method with two SOTA GS-based-method: 1) GSBA [17], 2) GLBA [25], and four SOTA RS-based-method: 1) MRSBA [5], 2) NMRSBA [1], 3) WMRSBA [14], 4) NWRSBA [14]. The experiments are conducted on a laptop with an Intel i7 CPU and all algorithms are implemented in MATLAB. + +Evaluation metrics. The accuracy measures employed encompass standard metrics such as rotation error, translation error, and algorithm efficiency. When conducting experiments related to line BA, we also incorporate error metrics that assess the accuracy of line reconstruction. Given the ground truth rotation $\mathbf{R}_g$ , translation $\mathbf{t}_g$ , and space line $\mathbf{L}_g = (\mathbf{n}_g^\top, \mathbf{a}_g^\top)^\top$ with an arbitrary point $\mathbf{P}_g$ on it. Similarly, the optimized parameters are denoted as $\mathbf{R}_d$ , $\mathbf{t}_d$ , $\mathbf{L}_d = (\mathbf{n}_d^\top, \mathbf{a}_d^\top)^\top$ , + +Table 1: Robustness analysis against multiple levels of Gaussian noise (px). + +
noise level (px)0.10.51.01.52.0
Rotation Error7.96e-69.24e-61.50e-51.71e-52.71e-5
Translation Error2.32e-45.93e-48.91e-49.54e-41.55e-3
L-R Error3.36e-33.83e-35.14e-35.22e-36.42e-3
L-D Error2.22e-42.43e-43.90e-44.43e-44.75e-4
+ +![](images/728eb8ed083f062042a63e82625d599e34fecd599cc267afe8ae7ef11a5918d0.jpg) +Fig. 7: Comparison of accuracy and time among different algorithms. + +$\mathbf{P}_d$ , respectively. The evaluation metrics are defined as: + +$$ +\left\{ \begin{array}{c} e _ {\text {R o t a t i o n}} = \operatorname {a r c c o s} \left(\frac {\operatorname {T r a c e} \left(\mathbf {R} _ {g} ^ {\top} \mathbf {R} _ {d}\right) - 1}{2}\right) \\ e _ {\text {T r a n s l a t i o n}} = \operatorname {a r c c o s} \left(\mathbf {t} _ {g} ^ {\top} \mathbf {t} _ {d} / | | \mathbf {t} _ {g} | | \cdot | | \mathbf {t} _ {d} |\right) \\ e _ {L - R} = \operatorname {a r c c o s} \left(\mathbf {a} _ {g} ^ {\top} \mathbf {a} _ {d} / | | \mathbf {a} _ {g} | | \cdot | | \mathbf {a} _ {d} |\right) \\ e _ {L - D} = \frac {| (\mathbf {a} _ {g} \times \mathbf {a} _ {d}) (\mathbf {P} _ {d} - \mathbf {P} _ {g}) |}{\| \mathbf {a} _ {g} \times \mathbf {a} _ {d} \|} \end{array} \right. \tag {23} +$$ + +# 4.1 Synthetic Experiments + +Which error term in Sec. 3.2 is better? In the synthetic environment, We set up a cubic box with 12 edges in 3D space and simulated a series of cameras around it to perform RS imaging. The experimental results are shown in Fig. 6. It is evident that, the first type of reprojection error combination performs much better than the second combination in terms of accuracy and time. Therefore, in the forthcoming experiments, we shall exclusively employ the first type of reprojection error combination for comparison. + +How significant is the impact of noise on RSL-BA? Following the same synthetic experimental setup as described above, we additionally introduce Gaussian noise to the endpoints of each line in this experiment. The noise levels are varied from 0.1 pixels to 2.0 pixels, and four different types of errors are recorded for the algorithm at each noise level. For each noise level, 50 experiments are conducted, and the median value is used as the final result. The results, presented in Table 1, demonstrate that RSL-BA provides stable estimations as the noise level increases from 0.1 pixels to 2 pixels. + +Table 2: The median absolute trajectory error (ATE) of different methods on WHU-RSVI [3] dataset. The best and second results are shown in green and blue, respectively. + +
WHU-RSVI1WHU-RSVI2WHU-RSVI3WHU-RSVI4
GSBA0.0809920.0613100.0304040.023698
GLBA0.0761730.0659850.0335540.024961
NMRSBA0.0509690.0416290.0423170.028183
NWRSBA0.0406400.0453130.0354510.022666
RLS-BA(ours)0.04435020.0393140.0233510.020675
+ +What are the advantages of $RSL-BA$ over point-based methods? Within this part, we will conduct a comparative analysis of our method alongside current SOTA methods, including point- and line-based methods. We set up a cubic box of points in the same positions as the 3D lines, totaling 56 points. To ensure fairness, we set up 8 lines with 7 points each to construct error functions, resulting in a total of precisely 56 points. Detailed experimental settings on the number of lines and the number of sampling points on each line are provided in the supplementary materials. We construct two scenarios: regular and classic scenarios with degeneration parallel to the y-axis. Due to significant differences in the experimental results of different methods, we uniformly take the logarithm (base 10) of the results for better observation. The experimental findings are shown in Fig. 7. It is evident that in both situations, our $RSL-BA$ and the current NWRSBA exhibit comparable accuracy, outperforming alternative approaches based on points and lines. Additionally, the computation time is equivalent to that of NWRSBA. This indicates that the application of lines on Rolling Shutter camera holds significant promise. + +# 4.2 Synthetic Images + +In this section, we conduct experiments on input synthetic images. We use the WHU-RSVI [3] dataset, from which we select two sets of data from trajectory1-fast and trajectory2-fast for 3D reconstruction and pose estimation. We first employ [22] to detect RS curves by segmenting curves into multiple short-line segments and performing line fitting for initialization. The GS line-based SfM [16] is applied to initialize the $RSL-BA$ parameters. The comparative methods include GSBA, NMRSBA, NWRSBA, and GSLBA. Table 2 shows the median absolute trajectory error of different methods, it can be observed that the proposed $RSL-BA$ method is the most stable one, achieving optimal or near-optimal results in all cases. Qualitative comparison are also provided in Fig. 8. Unlike point-based methods, line-based methods often achieve good results with fewer feature lines. However, the GSL-BA method is not sufficiently stable when RS effects are prominent. + +# 4.3 Real Images + +In this section, we conduct experiments on the real image dataset TUM-RSVI [24]. The experimental setup and comparison methods are similar to those in Sec. 4.2. + +Table 3: The absolute trajectory error (ATE) comparison of different methods in TUM-RSVI [24] dataset. The best and second results are shown in green and blue, respectively. + +
GSBAGLBANMRSBANWRSBARSL-BA(ours)
seq10.0694840.0864600.0523660.0450640.037816
seq20.0298210.0303790.0260280.0232270.025132
seq30.0651600.0637180.0579180.0550010.048187
seq40.0496130.0520260.0322140.0305340.031903
seq50.0318600.0358390.0194070.0160660.017659
seq60.0619660.0617920.0324340.0246580.026448
seq70.0516210.0565340.0391540.0396200.039701
seq80.0264030.0286900.0248070.0259260.024983
seq90.0983340.0982120.0824810.0735800.080523
seq100.811740.811800.593900.532960.57477
+ +![](images/e0d23ae6a4a6165cb387767c5b262fa2489d2272bb48cb228b4bb2e93ed384fa.jpg) +Fig. 8: Comparison of trajectories and 3D reconstruction on WHU-RSVI [3] and TUMRSVI [24] dataset. Each column represents a different bundle adjustment algorithm, and each row represents a different sequence. + +Table 3 and Fig 8 present some of the experimental results. As can be seen, in most cases, the $RSL-BA$ method outperforms other methods but is slightly weaker than NwRSBA. This is because the TUM-RSVI lacks line features and they are not visually prominent, making it less suitable for $RSL-BA$ . + +# 5 Conclusion + +This paper presents the first solution of line-based RSL-BA. By utilizing points and tangent directions on curves, we have established a series of faster and more robust curve-to-line re-projection errors. The proposed RSL-BA method can prevent three common degeneracies in RSBA, one of which is newly introduced by us. Extensive experiments in real and synthetic data verify the effectiveness and efficiency of the proposed RSL-BA method. + +# Acknowledgments + +This work is supported by the National Key R&D Program of China (No. 2022ZD01190030), Nature Science Foundation of China (No. 62102145), Jiangxi Provincial 03 Special Foundation and 5G Program (Grant No. 20224ABC03A05), Lushan Lab Research Funding, and Changsha Science Fund for Distinguished Young Scholars (kq2306002). + +# References + +1. Albl, C., Sugimoto, A., Pajdla, T.: Degeneracies in rolling shutter sfm. In: ECCV (2016) +2. Bartoli, A., Sturm, P.: Structure-from-motion using lines: Representation, triangulation, and bundle adjustment. Computer vision and image understanding 100(3), 416-441 (2005) +3. Cao, L., Ling, J., Xiao, X.: The whu rolling shutter visual-inertial dataset. IEEE Access 8, 50771-50779 (2020) +4. Dai, Y., Li, H., Kneip, L.: Rolling shutter camera relative pose: generalized epipolar geometry. In: CVPR (2016) +5. Duchamp, G., Ait-Aider, O., Royer, E., Lavest, J.M.: A rolling shutter compliant method for localisation and reconstruction. In: VISAPP (2015) +6. Fisher, A., Cannizzaro, R., Cochrane, M., Nagahawatte, C., Palmer, J.L.: Colmap: A memory-efficient occupancy grid mapping framework. Robotics and Autonomous Systems 142, 103755 (2021) +7. Hartley, R., Zisserman, A.: Multiple view geometry in computer vision. Cambridge university press (2003) +8. Hedborg, J., Forssen, P.E., Felsberg, M., Ringaby, E.: Rolling shutter bundle adjustment. In: CVPR (2012) +9. Holynski, A., Geraghty, D., Frahm, J.M., Sweeney, C., Szeliski, R.: Reducing drift in structure from motion using extended features. In: 2020 International Conference on 3D Vision (3DV). pp. 51-60. IEEE (2020) +10. Ito, E., Okatani, T.: Self-calibration-based approach to critical motion sequences of rolling-shutter structure from motion. In: CVPR (2016) +1. Lao, Y., Ait-Aider, O.: A robust method for strong rolling shutter effects correction using lines with automatic feature selection. In: CVPR (2018) +2. Lao, Y., Ait-Aider, O., Araujo, H.: Robustified structure from motion with rolling-shutter camera using straightness constraint. Pattern Recognition Letters (2018) +3. Lao, Y., Ait-Aider, O., Bartoli, A.: Solving rolling shutter 3d vision problems using analogies with non-rigidity. IJCV (2021) +4. Liao, B., Qu, D., Xue, Y., Zhang, H., Lao, Y.: Revisiting rolling shutter bundle adjustment: Toward accurate and fast solution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4863-4871 (2023) +5. Lim, H., Jeon, J., Myung, H.: Unconstrained line-based slam using vanishing points for structural mapping. IEEE Robotics and Automation Letters 7(2), 1518-1525 (2022) +6. Liu, S., Yu, Y., Pautrat, R., Pollefeys, M., Larsson, V.: 3d line mapping revisited. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21445-21455 (2023) + +17. Lourakis, M.I., Argyros, A.A.: Sba: A software package for generic sparse bundle adjustment. ACM Transactions on Mathematical Software (TOMS) 36(1), 1-30 (2009) +18. Micusik, B., Wildenauer, H.: Structure from motion with line segments under relaxed endpoint constraints. International Journal of Computer Vision 124, 65-79 (2017) +19. Moré, J.J.: The levenberg-marquardt algorithm: implementation and theory. In: Numerical analysis: proceedings of the biennial Conference held at Dundee, June 28-July 1, 1977. pp. 105-116. Springer (2006) +20. Patron-Perez, A., Lovegrove, S., Sibley, G.: A spline-based trajectory representation for sensor fusion and rolling shutter cameras. IJCV (2015) +21. Pumarola, A., Vakhitov, A., Agudo, A., Sanfeliu, A., Moreno-Noguer, F.: Pl-slam: Real-time monocular visual slam with points and lines. In: 2017 IEEE international conference on robotics and automation (ICRA). pp. 4503-4508. IEEE (2017) +22. Purkait, P., Zach, C., Leonardis, A.: Rolling shutter correction in manhattan world. In: ICCV. pp. 882-890 (2017) +23. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4104-4113 (2016) +24. Schubert, D., Demmel, N., von Stumberg, L., Usenko, V., Cremers, D.: Rolling-shutter modelling for direct visual-inertial odometry. In: IROS (2019) +25. Taylor, C.J., Kriegman, D.J.: Structure and motion from line segments in multiple images. IEEE Transactions on Pattern Analysis and Machine Intelligence 17(11), 1021-1032 (1995) +26. Triggs, B., McLauchlan, P.F., Hartley, R.I., Fitzgibbon, A.W.: Bundle adjustment—a modern synthesis. In: Vision Algorithms: Theory and Practice: International Workshop on Vision Algorithms Corfu, Greece, September 21–22, 1999 Proceedings. pp. 298–372. Springer (2000) +27. Von Gioi, R.G., Jakubowicz, J., Morel, J.M., Randall, G.: Lsd: A fast line segment detector with a false detection control. IEEE transactions on pattern analysis and machine intelligence 32(4), 722-732 (2008) +28. Wei, D., Wan, Y., Zhang, Y., Liu, X., Zhang, B., Wang, X.: Elsr: Efficient line segment reconstruction with planes and points guidance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15807-15815 (2022) +29. Zhuang, B., Cheong, L.F., Lee, G.H.: Rolling-shutter-aware differential sfm and image rectification. In: ICCV (2017) +30. Zhuang, B., Tran, Q.H., Ji, P., Cheong, L.F., Chandraker, M.: Learning structure-and-motion-aware rolling shutter correction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4551-4560 (2019) +31. Zuo, X., Xie, X., Liu, Y., Huang, G.: Robust visual slam with point and line features. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1775-1782. IEEE (2017) \ No newline at end of file diff --git a/2024/RSL-BA_ Rolling Shutter Line Bundle Adjustment/images.zip b/2024/RSL-BA_ Rolling Shutter Line Bundle Adjustment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6253f302fe09c6ff375cccef91b7ea99a5405a2d --- /dev/null +++ b/2024/RSL-BA_ Rolling Shutter Line Bundle Adjustment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d7250840c7e4cc223be03c0faedc7068b4f3858b7e8cb6f64c408a3bef7adddc +size 565834 diff --git a/2024/RSL-BA_ Rolling Shutter Line Bundle Adjustment/layout.json b/2024/RSL-BA_ Rolling Shutter Line Bundle Adjustment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..64c3d9eac41151ece576f437d4433af0b45dd8a0 --- /dev/null +++ b/2024/RSL-BA_ Rolling Shutter Line Bundle Adjustment/layout.json @@ -0,0 +1,10770 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 200, + 111, + 412, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 111, + 412, + 148 + ], + "spans": [ + { + "bbox": [ + 200, + 111, + 412, + 148 + ], + "type": "text", + "content": "RSL-BA: Rolling Shutter Line Bundle Adjustment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 154, + 167, + 459, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 167, + 459, + 193 + ], + "spans": [ + { + "bbox": [ + 154, + 167, + 459, + 193 + ], + "type": "text", + "content": "Yongcong Zhang" + }, + { + "bbox": [ + 154, + 167, + 459, + 193 + ], + "type": "inline_equation", + "content": "^{1,3,*}" + }, + { + "bbox": [ + 154, + 167, + 459, + 193 + ], + "type": "text", + "content": ", Bangyan Liao" + }, + { + "bbox": [ + 154, + 167, + 459, + 193 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 154, + 167, + 459, + 193 + ], + "type": "text", + "content": ", Yifei Xue" + }, + { + "bbox": [ + 154, + 167, + 459, + 193 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 154, + 167, + 459, + 193 + ], + "type": "text", + "content": ", Chen Lu" + }, + { + "bbox": [ + 154, + 167, + 459, + 193 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 154, + 167, + 459, + 193 + ], + "type": "text", + "content": ", Peidong Liu" + }, + { + "bbox": [ + 154, + 167, + 459, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 154, + 167, + 459, + 193 + ], + "type": "text", + "content": ", and Yizhen Lao" + }, + { + "bbox": [ + 154, + 167, + 459, + 193 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 193, + 201, + 419, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 201, + 419, + 225 + ], + "spans": [ + { + "bbox": [ + 193, + 201, + 419, + 225 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 193, + 201, + 419, + 225 + ], + "type": "text", + "content": "Hunan University " + }, + { + "bbox": [ + 193, + 201, + 419, + 225 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 193, + 201, + 419, + 225 + ], + "type": "text", + "content": "Westlake University " + }, + { + "bbox": [ + 193, + 201, + 419, + 225 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 193, + 201, + 419, + 225 + ], + "type": "text", + "content": "DaHe.AI " + }, + { + "bbox": [ + 193, + 201, + 419, + 225 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 193, + 201, + 419, + 225 + ], + "type": "text", + "content": "Dreame Technology Company" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 256, + 452, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 256, + 452, + 421 + ], + "spans": [ + { + "bbox": [ + 160, + 256, + 452, + 421 + ], + "type": "text", + "content": "Abstract. The line is a prevalent element in man-made environments, inherently encoding spatial structural information, thus making it a more robust choice for feature representation in practical applications. Despite its apparent advantages, previous rolling shutter bundle adjustment (RSBA) methods have only supported sparse feature points, which lack robustness, particularly in degenerate environments. In this paper, we introduce the first rolling shutter line-based bundle adjustment solution, " + }, + { + "bbox": [ + 160, + 256, + 452, + 421 + ], + "type": "inline_equation", + "content": "RSL-BA" + }, + { + "bbox": [ + 160, + 256, + 452, + 421 + ], + "type": "text", + "content": ". Specifically, we initially establish the rolling shutter camera line projection theory utilizing Plücker line parameterization. Subsequently, we derive a series of reprojection error formulations which are stable and efficient. Finally, we theoretically and experimentally demonstrate that our method can prevent three common degeneracies, one of which is first discovered in this paper. Extensive synthetic and real data experiments demonstrate that our method achieves efficiency and accuracy comparable to existing point-based rolling shutter bundle adjustment solutions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 431, + 366, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 431, + 366, + 443 + ], + "spans": [ + { + "bbox": [ + 160, + 431, + 366, + 443 + ], + "type": "text", + "content": "Keywords: Rolling Shutter " + }, + { + "bbox": [ + 160, + 431, + 366, + 443 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 431, + 366, + 443 + ], + "type": "text", + "content": " Bundle Adjustment" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 465, + 230, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 465, + 230, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 465, + 230, + 478 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 491, + 482, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 598 + ], + "type": "text", + "content": "Bundle Adjustment (BA) is a crucial step in multi-view 3D reconstruction, as it jointly optimizes camera poses and scene structure. Although feature point-based BA [6,23,26] currently dominates the academic landscape, line-based BA has been gaining increasing attention from researchers [16,28]. Lines, being one of the most common features in man-made environments, effectively describe the structural information of 3D scenes, in contrast to points which rely solely on local image patches. This structural information provides robust constraints for visual reconstruction [27], making line a more suitable choice in challenging scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 599, + 481, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 599, + 481, + 624 + ], + "spans": [ + { + "bbox": [ + 131, + 599, + 481, + 624 + ], + "type": "text", + "content": "In parallel with classical bundle adjustment [26], several methods [1,13] have incorporated the rolling shutter camera model to simultaneously estimate camera" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 142, + 632, + 229, + 643 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 632, + 229, + 643 + ], + "spans": [ + { + "bbox": [ + 142, + 632, + 229, + 643 + ], + "type": "text", + "content": "* Equal contribution" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 142, + 643, + 341, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 643, + 341, + 654 + ], + "spans": [ + { + "bbox": [ + 142, + 643, + 341, + 654 + ], + "type": "text", + "content": "† Corresponding author: (yizhenlao@hnu.edu.cn)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 654, + 368, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 654, + 368, + 665 + ], + "spans": [ + { + "bbox": [ + 142, + 654, + 368, + 665 + ], + "type": "text", + "content": "Project page: https://github.com/zhangtaxue/RSL-BA" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "poses, structure, and instantaneous motion parameters. However, all existing rolling shutter bundle adjustment (RSBA) methods rely solely on point features, while line features have never been addressed in RSBA. We identify two challenges that hinder the direct combination of line-based BA and the rolling shutter camera model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 482, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 482, + 258 + ], + "type": "text", + "content": "The first challenge arises from the time-dependent exposure of RS cameras. Unlike CCD cameras and their global shutter (GS) counterparts, RS cameras capture images in a scanline-by-scanline manner. Consequently, as illustrated in Fig. 1, images taken by moving RS cameras exhibit distortions known as the RS effect. This effect curves the projection of each 3D straight line, making it non-trivial to directly transfer the line-based BA to the rolling shutter camera setting." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 259, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 331 + ], + "type": "text", + "content": "The second challenge is associated with degeneracy. As shown in [1, 30], degeneracy is one of the main obstacles in RSBA since iterative optimization can easily be collapsed in some degenerate solutions which are far from the ground truth solutions. Although several strategies have been proposed to mitigate degeneracy in RSBA [1, 12, 14], a specific strategy to address degeneracy in line-based RSBA is still lacking." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 331, + 482, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 368 + ], + "type": "text", + "content": "In conclusion, an accurate and robust solution to line-based RSBA is still missing. Such a method would be vital in the potential widespread deployment of 3D vision with RS imaging systems." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 384, + 236, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 384, + 236, + 395 + ], + "spans": [ + { + "bbox": [ + 132, + 384, + 236, + 395 + ], + "type": "text", + "content": "1.1 Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 402, + 485, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 485, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 485, + 498 + ], + "type": "text", + "content": "Video-based RSBA. The assumption of smooth continuous trajectories is widely employed for RS video inputs to reduce the optimization parameter space and enhance algorithm robustness. In [8], Hedborg et al. present an RSBA algorithm that interpolates the motion between consecutive frames. Zhuang et al. [29] further propose an optical flow-based RSBA, which are developed to recover the relative pose of an RS camera that undergoes constant velocity and acceleration motion, respectively. Following this assumption, a spline-based camera trajectory motion model is proposed by [20]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 617 + ], + "type": "text", + "content": "Unordered RSBA. As the unordered image set is the standard input for SfM, Albl et al. first addresses this setting by explicitly velocity modeling and optimization. Besides, a planar degeneracy configuration has also been disclosed in [1]. While in [12], Lao et al. propose a camera-based RSBA to simulate the actual camera projection, exhibiting the degeneracy resilience ability. In [10], Ito et al. establish the equivalence between self-calibrated SfM and RSSFM based on the pure rotation instantaneous motion model and affine camera assumption, while the work of [13] draws the equivalence between RSSfM and non-rigid SfM. Recently, Liao et al. proposed two techniques to boost the accuracy and robustness of RSBA in [14]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Line-based GSBA. In [2, 25], the authors propose a comprehensive line-based SFM pipeline, while [9, 18] introduce line-based incremental SFM frameworks. Wei et al. in [28] employs planes and points to guide matching, achieving efficient line reconstruction. Recently, Lui et al. introduced LIMAP in [16], where the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "text", + "content": "Y. Zhang, B. Liao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 118, + 196, + 152 + ], + "blocks": [ + { + "bbox": [ + 137, + 118, + 196, + 152 + ], + "lines": [ + { + "bbox": [ + 137, + 118, + 196, + 152 + ], + "spans": [ + { + "bbox": [ + 137, + 118, + 196, + 152 + ], + "type": "image", + "image_path": "0f844bdff8bc036f12a958712c13ee7c6449121cf578752f82753cc210d6aa0d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 137, + 153, + 194, + 186 + ], + "blocks": [ + { + "bbox": [ + 137, + 153, + 194, + 186 + ], + "lines": [ + { + "bbox": [ + 137, + 153, + 194, + 186 + ], + "spans": [ + { + "bbox": [ + 137, + 153, + 194, + 186 + ], + "type": "image", + "image_path": "c7a70a41c52a0b2a6bccf64dce3bcc1c6cf2f636c2948a8ad504aba5ed6fc62c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 190, + 167, + 196 + ], + "lines": [ + { + "bbox": [ + 153, + 190, + 167, + 196 + ], + "spans": [ + { + "bbox": [ + 153, + 190, + 167, + 196 + ], + "type": "text", + "content": "input" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 202, + 118, + 383, + 200 + ], + "blocks": [ + { + "bbox": [ + 202, + 118, + 383, + 200 + ], + "lines": [ + { + "bbox": [ + 202, + 118, + 383, + 200 + ], + "spans": [ + { + "bbox": [ + 202, + 118, + 383, + 200 + ], + "type": "image", + "image_path": "2851336b6560c56be37b00028c47265d39d9850e08aaa71de39a2bd6db22a881.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 213, + 480, + 258 + ], + "lines": [ + { + "bbox": [ + 130, + 213, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 480, + 258 + ], + "type": "text", + "content": "Fig. 1: The pipeline of proposed RSL-BA with some example results. Starting with the input of some RS images (left), the optimization of camera poses and 3D line coordinates is performed using proposed RSL-BA (middle), yielding well-reconstructed 3D line segments and accurately estimated camera poses (right)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 385, + 119, + 470, + 182 + ], + "blocks": [ + { + "bbox": [ + 385, + 119, + 470, + 182 + ], + "lines": [ + { + "bbox": [ + 385, + 119, + 470, + 182 + ], + "spans": [ + { + "bbox": [ + 385, + 119, + 470, + 182 + ], + "type": "image", + "image_path": "b0bbc13a2a053ad2dbff8f56de0c1f750023518efb77dea927cf1cea0f815aa4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 190, + 444, + 196 + ], + "lines": [ + { + "bbox": [ + 427, + 190, + 444, + 196 + ], + "spans": [ + { + "bbox": [ + 427, + 190, + 444, + 196 + ], + "type": "text", + "content": "output" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 288, + 479, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 288, + 479, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 288, + 479, + 323 + ], + "type": "text", + "content": "authors devised various strategies and algorithms to achieve efficient and precise line reconstruction. In [15, 21, 31], hybrid strategies combining lines and feature points are utilized to enhance the accuracy and robustness of SLAM." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 324, + 480, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 480, + 407 + ], + "type": "text", + "content": "To this end, we propose a novel " + }, + { + "bbox": [ + 130, + 324, + 480, + 407 + ], + "type": "inline_equation", + "content": "RSL-BA" + }, + { + "bbox": [ + 130, + 324, + 480, + 407 + ], + "type": "text", + "content": " algorithm to solve the line-based RSBA problem. Specifically, we first establish the rolling shutter line projection theory using the Plücker line parameterization and subsequently derive a series of reprojection error formulations. We claim that the proposed reprojection error formulations are efficient and exhibit the degeneration-resistant ability. We also provide complete degeneration-resistant proof of our proposed error formulation. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 137, + 414, + 480, + 518 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 137, + 414, + 479, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 414, + 479, + 447 + ], + "spans": [ + { + "bbox": [ + 137, + 414, + 479, + 447 + ], + "type": "text", + "content": "- To the best of our knowledge, this is the first line-based RSBA solution, which serves as the foundation for line-based or point-line-hybrid RS-SfM and RS-SLAM." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 450, + 480, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 450, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 138, + 450, + 480, + 483 + ], + "type": "text", + "content": "- we theoretically and experimentally demonstrate that the proposed " + }, + { + "bbox": [ + 138, + 450, + 480, + 483 + ], + "type": "inline_equation", + "content": "RSLBA" + }, + { + "bbox": [ + 138, + 450, + 480, + 483 + ], + "type": "text", + "content": " can prevent three common degeneracies in RSBA, one of the degenerate cases is first discovered in this paper." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 137, + 486, + 479, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 486, + 479, + 518 + ], + "spans": [ + { + "bbox": [ + 137, + 486, + 479, + 518 + ], + "type": "text", + "content": "- The extensive evaluations in both synthetic and real datasets exhibit the comparable efficiency and accuracy of the proposed method over previous works." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 545, + 225, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 225, + 559 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 225, + 559 + ], + "type": "text", + "content": "2 Background" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 570, + 479, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 479, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 479, + 594 + ], + "type": "text", + "content": "In this section, we review the rolling shutter projection model in Sec. 2.1 and the parameterization of 3D lines in Sec. 2.2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 610, + 370, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 610, + 370, + 623 + ], + "spans": [ + { + "bbox": [ + 131, + 610, + 370, + 623 + ], + "type": "text", + "content": "2.1 Rolling Shutter Camera Projection Mode" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{R}(v) \\in \\mathbf{SO}(3)" + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{t}(v) \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": " represent the camera rotation and translation, respectively, when the row index " + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": " of measurement is acquired. Denote the intrinsic camera parameters as " + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": ", and the initial rotation and translation of the" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "RSL-BA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "camera at " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "v = 0" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_0" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_0" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": ", respectively. The projection matrix of the global shutter camera is defined as " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_0 = \\mathbf{K}[\\mathbf{R}_0, \\mathbf{t}_0]" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": ". Given the assumption of constant camera motion during frame capture, which is commonly made in RS 3D vision tasks [1,4,13], we can model the instantaneous motion as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 213, + 177, + 481, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 177, + 481, + 190 + ], + "spans": [ + { + "bbox": [ + 213, + 177, + 481, + 190 + ], + "type": "interline_equation", + "content": "\\mathbf {R} (v) = (\\mathbf {I} + [ \\boldsymbol {\\omega} ] _ {\\times} v) \\mathbf {R} _ {0}, \\quad \\mathbf {t} (v) = \\mathbf {t} _ {0} + \\mathbf {d} v, \\tag {1}", + "image_path": "567e34519038f7c609d465a53037e50989edaa74979736c31bfabf02bb46e887.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 201, + 480, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 480, + 248 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 201, + 480, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{d} = [d_x, d_y, d_z]^\\top" + }, + { + "bbox": [ + 130, + 201, + 480, + 248 + ], + "type": "text", + "content": " is the translational velocity vector and " + }, + { + "bbox": [ + 130, + 201, + 480, + 248 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\omega} = [\\omega_x, \\omega_y, \\omega_z]^\\top" + }, + { + "bbox": [ + 130, + 201, + 480, + 248 + ], + "type": "text", + "content": " is the rotational velocity vector, and " + }, + { + "bbox": [ + 130, + 201, + 480, + 248 + ], + "type": "inline_equation", + "content": "[\\boldsymbol{\\omega}]_\\times" + }, + { + "bbox": [ + 130, + 201, + 480, + 248 + ], + "type": "text", + "content": " represents the skew-symmetric matrix of vector " + }, + { + "bbox": [ + 130, + 201, + 480, + 248 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\omega}" + }, + { + "bbox": [ + 130, + 201, + 480, + 248 + ], + "type": "text", + "content": ". Then, the projection matrix of the camera under the RS model can be defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 236, + 251, + 481, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 251, + 481, + 269 + ], + "spans": [ + { + "bbox": [ + 236, + 251, + 481, + 269 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {v} = \\mathbf {K} \\left[ \\mathbf {R} _ {v}, \\mathbf {t} _ {v} \\right] = \\mathbf {P} _ {0} + v \\mathbf {Q} \\\\ \\mathbf {K} ^ {[ r ]} = \\left[ \\begin{array}{l l l l l} \\mathbf {P} _ {v} & \\mathbf {t} _ {v} & \\mathbf {Q} _ {v} & \\mathbf {I} ^ {[ r ]} & \\mathbf {D} _ {v} \\end{array} \\right] \\tag {2}", + "image_path": "5f0fe3d60ad0f79b8ce43057bdd2b430b8676709e4417650c5ebf72179c8103b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 227, + 266, + 386, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 266, + 386, + 277 + ], + "spans": [ + { + "bbox": [ + 227, + 266, + 386, + 277 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {0} = \\mathbf {K} \\left[ \\mathbf {R} _ {0}, \\mathbf {t} _ {0} \\right] \\quad \\mathbf {Q} = \\mathbf {K} \\left[ \\left[ \\boldsymbol {\\omega} \\right] _ {\\times} \\mathbf {R} _ {0}, \\mathbf {d} \\right].", + "image_path": "a26c76da9145a2b12e0b477fd75d93b5399876c511cc2b5a92497f8480091ee3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 301, + 342, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 301, + 342, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 301, + 342, + 313 + ], + "type": "text", + "content": "2.2 3D Line Segments Parameterization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 328, + 481, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 328, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 131, + 328, + 481, + 363 + ], + "type": "text", + "content": "We employ the Plücker matrix to parameterize the 3D line [7]. Specifically, given two homogeneous 3D points " + }, + { + "bbox": [ + 131, + 328, + 481, + 363 + ], + "type": "inline_equation", + "content": "P_{a} = [a_{1}, a_{2}, a_{3}, 1]^{\\top}" + }, + { + "bbox": [ + 131, + 328, + 481, + 363 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 328, + 481, + 363 + ], + "type": "inline_equation", + "content": "P_{b} = [b_{1}, b_{2}, b_{3}, 1]^{\\top}" + }, + { + "bbox": [ + 131, + 328, + 481, + 363 + ], + "type": "text", + "content": ". The Plücker matrix " + }, + { + "bbox": [ + 131, + 328, + 481, + 363 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 131, + 328, + 481, + 363 + ], + "type": "text", + "content": " can be defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 376, + 373, + 460, + 387 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 373, + 460, + 387 + ], + "spans": [ + { + "bbox": [ + 376, + 373, + 460, + 387 + ], + "type": "interline_equation", + "content": "\\mathbf {L} = P _ {a} P _ {b} ^ {\\top} - P _ {b} P _ {a} ^ {\\top}", + "image_path": "61cad15420dac8a52fb6a908495a4b9719c42f5d27f64de5ec43f00dc608dd10.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 390, + 480, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 390, + 480, + 440 + ], + "spans": [ + { + "bbox": [ + 141, + 390, + 480, + 440 + ], + "type": "interline_equation", + "content": "= \\left[ \\begin{array}{c c c c} 0 & a _ {1} b _ {2} - a _ {2} b _ {1} & a _ {1} b _ {3} - a _ {3} b _ {1} & a _ {1} - b _ {1} \\\\ a _ {2} b _ {1} - a _ {1} b _ {2} & 0 & a _ {2} b _ {3} - a _ {3} b _ {2} & a _ {2} - b _ {2} \\\\ a _ {3} b _ {1} - a _ {1} b _ {3} & a _ {3} b _ {2} - a _ {2} b _ {3} & 0 & a _ {3} - b _ {3} \\\\ b _ {1} - a _ {1} & b _ {2} - a _ {2} & b _ {3} - a _ {3} & 0 \\end{array} \\right] = \\left[ \\begin{array}{c c c c} 0 & l _ {1 2} & l _ {1 3} & l _ {1 4} \\\\ - l _ {1 2} & 0 & l _ {2 3} & l _ {2 4} \\\\ - l _ {1 3} & - l _ {2 3} & 0 & l _ {3 4} \\\\ - l _ {1 4} & - l _ {2 4} & - l _ {3 4} & 0 \\end{array} \\right], \\tag {3}", + "image_path": "e8b17d508dacecb6713b17c20afef1264e7b987293edf53da9268cbde7601eaa.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 451, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 451, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 131, + 451, + 482, + 499 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 131, + 451, + 482, + 499 + ], + "type": "inline_equation", + "content": "[l_{14}, l_{24}, l_{34}]^{\\top}" + }, + { + "bbox": [ + 131, + 451, + 482, + 499 + ], + "type": "text", + "content": " represents the direction of this straight line while the " + }, + { + "bbox": [ + 131, + 451, + 482, + 499 + ], + "type": "inline_equation", + "content": "[-l_{23}, l_{13}, -l_{12}]^{\\top}" + }, + { + "bbox": [ + 131, + 451, + 482, + 499 + ], + "type": "text", + "content": " represents the normal direction of this line. The projection of 3D Plücker matrix is the skew-symmetric matrix of a 2D normalized line. It can be related to each other as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 258, + 512, + 480, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 512, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 258, + 512, + 480, + 525 + ], + "type": "interline_equation", + "content": "\\left[ \\mathbf {l} _ {g s} \\right] _ {\\times} = \\left[ \\mathbf {R}, \\mathbf {t} \\right] \\mathbf {L} \\left[ \\mathbf {R}, \\mathbf {t} \\right] ^ {\\top} \\tag {4}", + "image_path": "95fd5cc538c3d2f82095cde22809d0b1dc0e9f21054575e31d99975b4f7a2698.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 549, + 233, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 549, + 233, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 549, + 233, + 563 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "In Sec. 3.1, we initially establish the rolling shutter projection equations using the Plücker line parameterization. These equations establish a connection between the 3D lines and the curve parameters on the image plane. Then in Sec. 3.2, we present a series of line-based reprojection error equations. On top of the error definition, we can formulate the entire rolling shutter line bundle adjustment problem in Sec. 3.3. Finally, in Sec. 3.4, we demonstrate the degeneracy resilience capability of the proposed error equations." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "text", + "content": "Y. Zhang, B. Liao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 385, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 385, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 385, + 129 + ], + "type": "text", + "content": "3.1 Rolling Shutter Line Projection Formulation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 136, + 479, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 136, + 479, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 479, + 171 + ], + "type": "text", + "content": "As previously explained, the curved rolling shutter projection of 3D lines occurs because of the camera's nonlinear motion. Using the Plücker parameterization method, we can derive explicit expressions for the curve parameters." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 172, + 480, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 172, + 480, + 197 + ], + "spans": [ + { + "bbox": [ + 130, + 172, + 480, + 197 + ], + "type": "text", + "content": "Firstly, we construct the skew-symmetric matrix " + }, + { + "bbox": [ + 130, + 172, + 480, + 197 + ], + "type": "inline_equation", + "content": "\\left[\\mathbf{l}_{rs}\\right]_{\\times}" + }, + { + "bbox": [ + 130, + 172, + 480, + 197 + ], + "type": "text", + "content": " of the instantaneous projection line by combining Eq. (2) and Eq. (4):" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 194, + 205, + 480, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 205, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 194, + 205, + 480, + 251 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left[ \\mathbf {l} _ {r s} \\right] _ {\\times} = \\mathbf {P} _ {v} \\mathbf {L} \\mathbf {P} _ {v} ^ {\\top} \\\\ = \\mathbf {P} _ {0} \\mathbf {L} \\mathbf {P} _ {0} ^ {\\top} + v \\left(\\mathbf {P} _ {0} \\mathbf {L} \\mathbf {Q} ^ {\\top} + \\mathbf {Q} \\mathbf {L} \\mathbf {P} _ {0} ^ {\\top}\\right) + v ^ {2} \\mathbf {Q} \\mathbf {L} \\mathbf {Q} ^ {\\top} \\tag {5} \\\\ = \\mathbf {A} _ {1} + v \\mathbf {A} _ {2} + v ^ {2} \\mathbf {A} _ {3} \\\\ \\end{array}", + "image_path": "39c3b7da05f5653491ae14498b2f60e72195415f75af97006e33cf7f496c82c5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 261, + 479, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 479, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 479, + 285 + ], + "type": "text", + "content": "Secondly, as " + }, + { + "bbox": [ + 130, + 261, + 479, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{l}_{rs} = [l_1,l_2,l_3]^\\top" + }, + { + "bbox": [ + 130, + 261, + 479, + 285 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 261, + 479, + 285 + ], + "type": "inline_equation", + "content": "[\\mathbf{l}_{rs}]_{\\times} = \\mathbf{A}_1 + v\\mathbf{A}_2 + v^2\\mathbf{A}_3" + }, + { + "bbox": [ + 130, + 261, + 479, + 285 + ], + "type": "text", + "content": ", we can determine the instantaneous projected line parameter as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 246, + 295, + 480, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 295, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 246, + 295, + 480, + 332 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} l _ {1} = \\mathbf {A} _ {1} ^ {3 2} + v \\mathbf {A} _ {2} ^ {3 2} + v ^ {2} \\mathbf {A} _ {3} ^ {3 2} \\\\ l _ {2} = \\mathbf {A} _ {1} ^ {1 3} + v \\mathbf {A} _ {2} ^ {1 3} + v ^ {2} \\mathbf {A} _ {3} ^ {1 3} \\\\ l _ {3} = \\mathbf {A} _ {1} ^ {2 1} + v \\mathbf {A} _ {2} ^ {2 1} + v ^ {2} \\mathbf {A} _ {3} ^ {2 1} \\end{array} \\right. \\tag {6}", + "image_path": "ba466316ef47d8f2e31e4105675c664e26ac84784477edecd24eaec3308c66dc.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 340, + 479, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 479, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 479, + 376 + ], + "type": "text", + "content": "Ultimately, we note that the instantaneous projected line parameter fulfills this " + }, + { + "bbox": [ + 130, + 340, + 479, + 376 + ], + "type": "inline_equation", + "content": "l_{1}u + l_{2}v + l_{3} = 0" + }, + { + "bbox": [ + 130, + 340, + 479, + 376 + ], + "type": "text", + "content": " equation, which we can expand to obtain the final curve equation, denoted as " + }, + { + "bbox": [ + 130, + 340, + 479, + 376 + ], + "type": "inline_equation", + "content": "\\pmb{\\iota}" + }, + { + "bbox": [ + 130, + 340, + 479, + 376 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 207, + 384, + 480, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 384, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 207, + 384, + 480, + 415 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\iota : \\mathbf {A} _ {3} ^ {1 3} v ^ {3} + \\mathbf {A} _ {3} ^ {3 2} u v ^ {2} + \\left(\\mathbf {A} _ {2} ^ {1 3} + \\mathbf {A} _ {3} ^ {2 1}\\right) v ^ {2} + \\mathbf {A} _ {2} ^ {3 2} u v \\\\ + \\left(\\mathbf {A} _ {1} ^ {1 3} + \\mathbf {A} _ {2} ^ {2 1}\\right) v + \\mathbf {A} _ {1} ^ {3 2} u + \\mathbf {A} _ {1} ^ {2 1} = 0 \\\\ \\end{array}", + "image_path": "2bb16df8ea43972b1c742976010f567ce2266a299fb2ed86c590da30f886371f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 424, + 479, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 479, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 479, + 483 + ], + "type": "text", + "content": "Our conclusion aligns with that of Lao et al. [11], yielding an identical polynomial curve degree. Nevertheless, our utilization of a 4-DoF orthogonal representation representation for 3D lines has distinct benefits when compared to Lao's 5-DoF over-parameterized representation. These advantages include the capability to conduct unconstrained optimization on spatial lines directly." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 484, + 480, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 480, + 556 + ], + "type": "text", + "content": "Remark. The geometric interpretation underlying this derivation is quite straightforward. Given the parameters of a spatial line and the current camera pose, if point " + }, + { + "bbox": [ + 130, + 484, + 480, + 556 + ], + "type": "inline_equation", + "content": "(\\mathbf{u},\\mathbf{v})" + }, + { + "bbox": [ + 130, + 484, + 480, + 556 + ], + "type": "text", + "content": " lies on the curve obtained by projecting this spatial line, then when the RS camera scans to the v-th row, the projection of the spatial line onto the image plane follows Eq. (6). Therefore, it is evident that point " + }, + { + "bbox": [ + 130, + 484, + 480, + 556 + ], + "type": "inline_equation", + "content": "(\\mathbf{u},\\mathbf{v})" + }, + { + "bbox": [ + 130, + 484, + 480, + 556 + ], + "type": "text", + "content": " must be located on this line. Based on this reasoning, we can deduce Eq. (7)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 574, + 320, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 320, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 320, + 586 + ], + "type": "text", + "content": "3.2 Line-based Reprojection Errors" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "text", + "content": "As claimed in Sec. 3.1, a 3D straight line will become a polynomial curve on the captured image. In this section, we will establish a series of re-projection errors which are stable and robust. We first leverage the similar intuition in Sec. 3.1 to obtain the distance error representation. As shown in Fig. 2(a), Let there be a point " + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{q} = [u,v]^{\\top}" + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "text", + "content": " on the observed curve " + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\iota" + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "text", + "content": ", then the pose of the camera when exposing that point is " + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_v" + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_v" + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "text", + "content": ", then projecting the 3D line " + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "text", + "content": " onto the image" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "RSL-BA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 114, + 475, + 228 + ], + "blocks": [ + { + "bbox": [ + 141, + 114, + 475, + 228 + ], + "lines": [ + { + "bbox": [ + 141, + 114, + 475, + 228 + ], + "spans": [ + { + "bbox": [ + 141, + 114, + 475, + 228 + ], + "type": "image", + "image_path": "4f4d340e18b248320f6904fb07ad6a461bb1487e0d9e246341b03ccbf67c238a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 239, + 482, + 318 + ], + "lines": [ + { + "bbox": [ + 131, + 239, + 482, + 318 + ], + "spans": [ + { + "bbox": [ + 131, + 239, + 482, + 318 + ], + "type": "text", + "content": "Fig. 2: The illustration of proposed line-based reprojection errors given the spatial line " + }, + { + "bbox": [ + 131, + 239, + 482, + 318 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 131, + 239, + 482, + 318 + ], + "type": "text", + "content": " and observed curve " + }, + { + "bbox": [ + 131, + 239, + 482, + 318 + ], + "type": "inline_equation", + "content": "\\iota" + }, + { + "bbox": [ + 131, + 239, + 482, + 318 + ], + "type": "text", + "content": ". Instead of directly measuring the distance between the observed curve and the projected curve, we split each curve into multiple points " + }, + { + "bbox": [ + 131, + 239, + 482, + 318 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 131, + 239, + 482, + 318 + ], + "type": "text", + "content": ", and aggregate the distance errors. Specifically, we first utilize the virtual projected line " + }, + { + "bbox": [ + 131, + 239, + 482, + 318 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 131, + 239, + 482, + 318 + ], + "type": "text", + "content": " when observing it, as shown in (a). To measure the distance between each point " + }, + { + "bbox": [ + 131, + 239, + 482, + 318 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 131, + 239, + 482, + 318 + ], + "type": "text", + "content": " and the corresponding virtual lines, we can use the perpendicular distance (b), the horizontal distance (c,d) or the tangent distance (e) as our basic metrics." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 340, + 482, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 340, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 131, + 340, + 482, + 389 + ], + "type": "text", + "content": "plane, this results in the virtual 2D line " + }, + { + "bbox": [ + 131, + 340, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{1} = [l_1, l_2, l_3]^\\top" + }, + { + "bbox": [ + 131, + 340, + 482, + 389 + ], + "type": "text", + "content": ". Under the condition of ground truth, this virtual line must theoretically intersect with that point " + }, + { + "bbox": [ + 131, + 340, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 131, + 340, + 482, + 389 + ], + "type": "text", + "content": ". This constraint directly leads us to the two variants of distance error, which we have defined below." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 408, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 408, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 131, + 408, + 482, + 445 + ], + "type": "text", + "content": "Perpendicular Distance Error " + }, + { + "bbox": [ + 131, + 408, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{d1}" + }, + { + "bbox": [ + 131, + 408, + 482, + 445 + ], + "type": "text", + "content": ". The first distance error formulation, named perpendicular distance error, measures the shortest distance from point " + }, + { + "bbox": [ + 131, + 408, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 131, + 408, + 482, + 445 + ], + "type": "text", + "content": " to the line " + }, + { + "bbox": [ + 131, + 408, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{l}" + }, + { + "bbox": [ + 131, + 408, + 482, + 445 + ], + "type": "text", + "content": ", as shown in Fig. 2(b). The error is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 259, + 453, + 481, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 453, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 259, + 453, + 481, + 482 + ], + "type": "interline_equation", + "content": "\\mathbf {e} _ {d 1} = \\frac {l _ {1} u + l _ {2} v + l _ {3}}{\\sqrt {l _ {1} ^ {2} + l _ {2} ^ {2}}}, \\tag {8}", + "image_path": "5d00dc0a421091fe0158b921b438175693d9ebc3f0cadc140cd98873ca385b44.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 489, + 443, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 489, + 443, + 502 + ], + "spans": [ + { + "bbox": [ + 131, + 489, + 443, + 502 + ], + "type": "text", + "content": "where the line parameters " + }, + { + "bbox": [ + 131, + 489, + 443, + 502 + ], + "type": "inline_equation", + "content": "[l_1, l_2, l_3]" + }, + { + "bbox": [ + 131, + 489, + 443, + 502 + ], + "type": "text", + "content": " can be calculated through Eq. (6)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 521, + 482, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 521, + 482, + 558 + ], + "spans": [ + { + "bbox": [ + 131, + 521, + 482, + 558 + ], + "type": "text", + "content": "Horizontal Distance Error " + }, + { + "bbox": [ + 131, + 521, + 482, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{d2}" + }, + { + "bbox": [ + 131, + 521, + 482, + 558 + ], + "type": "text", + "content": ". The second distance error formulation, named horizontal distance error, measures the u-axis distance from point " + }, + { + "bbox": [ + 131, + 521, + 482, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 131, + 521, + 482, + 558 + ], + "type": "text", + "content": " to the line " + }, + { + "bbox": [ + 131, + 521, + 482, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{l}" + }, + { + "bbox": [ + 131, + 521, + 482, + 558 + ], + "type": "text", + "content": ", as shown in Fig. 2(c). The error can be defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 276, + 567, + 481, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 567, + 481, + 581 + ], + "spans": [ + { + "bbox": [ + 276, + 567, + 481, + 581 + ], + "type": "interline_equation", + "content": "\\mathbf {e} _ {d 2} = u - u ^ {\\prime}, \\tag {9}", + "image_path": "8b7613b1a4c5a66abc15da383b6bffe149a1514c43b25fb8602185845a5647b2.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 590, + 482, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 590, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 131, + 590, + 482, + 612 + ], + "type": "text", + "content": "where the corresponding " + }, + { + "bbox": [ + 131, + 590, + 482, + 612 + ], + "type": "inline_equation", + "content": "u'" + }, + { + "bbox": [ + 131, + 590, + 482, + 612 + ], + "type": "text", + "content": " coordinate of virtual line 1 at the " + }, + { + "bbox": [ + 131, + 590, + 482, + 612 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 131, + 590, + 482, + 612 + ], + "type": "text", + "content": "-th row can be calculated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 271, + 612, + 481, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 612, + 481, + 637 + ], + "spans": [ + { + "bbox": [ + 271, + 612, + 481, + 637 + ], + "type": "interline_equation", + "content": "u ^ {\\prime} = - \\frac {l _ {2} v + l _ {3}}{l _ {1}}, \\tag {10}", + "image_path": "5777aa7551f872cdc5767c788c9b06bc7a675830e244bc4cf83299f7e08f5e66.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "As shown in Fig. 2(d), the " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "e_{d2}" + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": " distance error is equivalent to the u-axis distance from the point to the projection curve." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "text", + "content": "Y. Zhang, B. Liao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "Tangent Error " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_t" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": ". While distance error provides an elegant constraint, it is insufficient to rely solely on distance error in certain scenarios. To further enhance the robustness, we introduce a second type of tangent error. As shown in Fig. 2(e), we constrain the tangent of the projected curve " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\iota'" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " to match the observed tangent direction. The error can be defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 285, + 185, + 481, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 285, + 185, + 481, + 198 + ], + "spans": [ + { + "bbox": [ + 285, + 185, + 481, + 198 + ], + "type": "interline_equation", + "content": "\\mathbf {e} _ {t} = \\mathbf {s} ^ {\\top} \\mathbf {s} ^ {\\prime}, \\tag {11}", + "image_path": "514fe4032e18811309b690035ed118dd523ffd8f2fa14bc17208089f29b54b77.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 208, + 479, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 208, + 479, + 232 + ], + "spans": [ + { + "bbox": [ + 131, + 208, + 479, + 232 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 208, + 479, + 232 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 131, + 208, + 479, + 232 + ], + "type": "text", + "content": " represents the tangent direction observation. The " + }, + { + "bbox": [ + 131, + 208, + 479, + 232 + ], + "type": "inline_equation", + "content": "\\mathbf{s}'" + }, + { + "bbox": [ + 131, + 208, + 479, + 232 + ], + "type": "text", + "content": " can be calculated analytically following the implicit function theorem as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 206, + 241, + 481, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 241, + 481, + 268 + ], + "spans": [ + { + "bbox": [ + 206, + 241, + 481, + 268 + ], + "type": "interline_equation", + "content": "\\mathbf {s} ^ {\\prime} = \\left[ \\frac {1}{\\sqrt {1 + s ^ {2}}}, \\frac {s}{\\sqrt {1 + s ^ {2}}} \\right], \\quad s = \\frac {\\partial u}{\\partial v} = - \\frac {\\partial \\iota / \\partial v}{\\partial \\iota / \\partial u} \\tag {12}", + "image_path": "525480aae250b03f9c81ea7582aff92024868095dcfefdb4ef47bc13a6f201b8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 282, + 366, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 282, + 366, + 295 + ], + "spans": [ + { + "bbox": [ + 131, + 282, + 366, + 295 + ], + "type": "text", + "content": "3.3 Rolling Shutter Line Bundle Adjustment" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 303, + 479, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 479, + 351 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 479, + 351 + ], + "type": "text", + "content": "The non-linear least squares solvers are used to find an optimal solution " + }, + { + "bbox": [ + 130, + 303, + 479, + 351 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}^{*}" + }, + { + "bbox": [ + 130, + 303, + 479, + 351 + ], + "type": "text", + "content": " including camera poses " + }, + { + "bbox": [ + 130, + 303, + 479, + 351 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^{*},\\mathbf{t}^{*}" + }, + { + "bbox": [ + 130, + 303, + 479, + 351 + ], + "type": "text", + "content": ", instantaneous motion " + }, + { + "bbox": [ + 130, + 303, + 479, + 351 + ], + "type": "inline_equation", + "content": "\\omega^{*},\\mathbf{d}^{*}" + }, + { + "bbox": [ + 130, + 303, + 479, + 351 + ], + "type": "text", + "content": " and 3D lines " + }, + { + "bbox": [ + 130, + 303, + 479, + 351 + ], + "type": "inline_equation", + "content": "\\pmb{\\tau}^{*}" + }, + { + "bbox": [ + 130, + 303, + 479, + 351 + ], + "type": "text", + "content": " by minimizing the reprojection error " + }, + { + "bbox": [ + 130, + 303, + 479, + 351 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_i^j" + }, + { + "bbox": [ + 130, + 303, + 479, + 351 + ], + "type": "text", + "content": " from line " + }, + { + "bbox": [ + 130, + 303, + 479, + 351 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 303, + 479, + 351 + ], + "type": "text", + "content": " to camera " + }, + { + "bbox": [ + 130, + 303, + 479, + 351 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 303, + 479, + 351 + ], + "type": "text", + "content": " over all the camera index in set " + }, + { + "bbox": [ + 130, + 303, + 479, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 303, + 479, + 351 + ], + "type": "text", + "content": " and corresponding observed 3D lines index in subset " + }, + { + "bbox": [ + 130, + 303, + 479, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_j" + }, + { + "bbox": [ + 130, + 303, + 479, + 351 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 196, + 361, + 481, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 361, + 481, + 389 + ], + "spans": [ + { + "bbox": [ + 196, + 361, + 481, + 389 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\theta} ^ {*} = \\left\\{\\boldsymbol {\\tau} ^ {*}, \\mathbf {R} ^ {*}, \\mathbf {t} ^ {*}, \\boldsymbol {\\omega} ^ {*}, \\mathbf {d} ^ {*} \\right\\} = \\arg \\min _ {\\boldsymbol {\\theta}} \\sum_ {j \\in \\mathcal {F}} \\sum_ {i \\in \\mathcal {P} _ {j}} \\left\\| \\mathbf {e} _ {i} ^ {j} \\right\\| _ {2} ^ {2}. \\tag {13}", + "image_path": "b6df29075d8c25e6ab97bb3f775619fe42fe8dc344a96c0615b194263fd06428.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 399, + 479, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 479, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 479, + 422 + ], + "type": "text", + "content": "On top of errors proposed in Sec. 3.2, we provide two variants of reprojection error:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 230, + 424, + 480, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 424, + 480, + 436 + ], + "spans": [ + { + "bbox": [ + 230, + 424, + 480, + 436 + ], + "type": "interline_equation", + "content": "\\mathbf {e} _ {1} = \\mathbf {e} _ {d 1} + \\lambda_ {1} \\mathbf {e} _ {t}, \\quad \\mathbf {e} _ {2} = \\mathbf {e} _ {d 2} + \\lambda_ {2} \\mathbf {e} _ {t}, \\tag {14}", + "image_path": "1ddcea8f04dac4fbcfbe2652c19ddf4bdf4e4e7926527e04410a38af983567d5.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 441, + 479, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 479, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 479, + 487 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 441, + 479, + 487 + ], + "type": "inline_equation", + "content": "\\lambda_1, \\lambda_2" + }, + { + "bbox": [ + 130, + 441, + 479, + 487 + ], + "type": "text", + "content": " are the relative weight. We have experimentally validated the optimality of these two error variants in Sec. 4.1. After definition, this non-linear least square problem can be iteratively solved through Levenberg-Marquardt [19] method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 489, + 479, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 489, + 479, + 512 + ], + "spans": [ + { + "bbox": [ + 131, + 489, + 479, + 512 + ], + "type": "text", + "content": "Besides, the full analytical Jacobian has been derived in the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 531, + 265, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 531, + 265, + 544 + ], + "spans": [ + { + "bbox": [ + 131, + 531, + 265, + 544 + ], + "type": "text", + "content": "3.4 Degeneracy Analysis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 551, + 479, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 551, + 479, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 551, + 479, + 612 + ], + "type": "text", + "content": "In this section, we conducted theoretical analysis on the resilience of " + }, + { + "bbox": [ + 130, + 551, + 479, + 612 + ], + "type": "inline_equation", + "content": "RSL-BA" + }, + { + "bbox": [ + 130, + 551, + 479, + 612 + ], + "type": "text", + "content": " to three types of degeneracy scenarios. These three degeneracy scenarios include plane degeneracy case [1], 2-vIEWS pure translation case [30], and X-Y pure translation degeneracy case disclosed by us. We assume the positive direction of X Y Z axis is right down forward, respectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 629, + 479, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 667 + ], + "type": "text", + "content": "(1) Resistance to Plane Degeneracy [1]. We first briefly describe the conditions under which the plane degeneration occurs: when the camera's y-axis is parallel, if a rotational velocity " + }, + { + "bbox": [ + 130, + 629, + 479, + 667 + ], + "type": "inline_equation", + "content": "\\omega = [-1,0,0]^{\\top}" + }, + { + "bbox": [ + 130, + 629, + 479, + 667 + ], + "type": "text", + "content": " is given to the camera, then the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "RSL-BA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "reconstructed 3D points will be compressed onto a plane perpendicular to the " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "-axis, and satisfy that the reprojection error is 0." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 222 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 222 + ], + "type": "text", + "content": "For line features, we also assume " + }, + { + "bbox": [ + 130, + 140, + 480, + 222 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_0 = \\mathbf{I}" + }, + { + "bbox": [ + 130, + 140, + 480, + 222 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 140, + 480, + 222 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_0 = \\mathbf{0}" + }, + { + "bbox": [ + 130, + 140, + 480, + 222 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 140, + 480, + 222 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\omega} = [-1,0,0]^\\top" + }, + { + "bbox": [ + 130, + 140, + 480, + 222 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 140, + 480, + 222 + ], + "type": "inline_equation", + "content": "\\mathbf{d} = 0" + }, + { + "bbox": [ + 130, + 140, + 480, + 222 + ], + "type": "text", + "content": ". Let there be two points, " + }, + { + "bbox": [ + 130, + 140, + 480, + 222 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_a" + }, + { + "bbox": [ + 130, + 140, + 480, + 222 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 140, + 480, + 222 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_b" + }, + { + "bbox": [ + 130, + 140, + 480, + 222 + ], + "type": "text", + "content": ", on the spatial line " + }, + { + "bbox": [ + 130, + 140, + 480, + 222 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 130, + 140, + 480, + 222 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 140, + 480, + 222 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_a = [a_1,a_2,a_3,1]^\\top" + }, + { + "bbox": [ + 130, + 140, + 480, + 222 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 140, + 480, + 222 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_b = [b_1,b_2,b_3,1]^\\top" + }, + { + "bbox": [ + 130, + 140, + 480, + 222 + ], + "type": "text", + "content": ". When the space is compressed into a plane, the new coordinates of " + }, + { + "bbox": [ + 130, + 140, + 480, + 222 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_a" + }, + { + "bbox": [ + 130, + 140, + 480, + 222 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 140, + 480, + 222 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_b" + }, + { + "bbox": [ + 130, + 140, + 480, + 222 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 130, + 140, + 480, + 222 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_a' = [a_1,0,a_3,1]^\\top" + }, + { + "bbox": [ + 130, + 140, + 480, + 222 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 140, + 480, + 222 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_b' = [b_1,0,b_3,1]^\\top" + }, + { + "bbox": [ + 130, + 140, + 480, + 222 + ], + "type": "text", + "content": ", respectively. By referring to Sec. 3.1, we can solve for the projection parameters of the two-dimensional curve at this moment. The parameters of the curve at this time are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 194, + 224, + 480, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 224, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 194, + 224, + 480, + 262 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\mathbf {A} _ {3} ^ {1 3} = \\mathbf {A} _ {3} ^ {3 2} = \\mathbf {A} _ {2} ^ {1 3} = \\mathbf {A} _ {3} ^ {2 1} = \\mathbf {A} _ {2} ^ {3 2} = \\mathbf {A} _ {1} ^ {3 2} = \\mathbf {A} _ {1} ^ {2 1} = 0 \\\\ \\mathbf {A} _ {1} ^ {1 3} = a _ {3} b _ {1} - a _ {1} b _ {3} \\\\ \\mathbf {A} _ {2} ^ {2 1} = a _ {1} b _ {3} - a _ {3} b _ {1} \\end{array} \\right. \\tag {15}", + "image_path": "bfdbc6e0d955c87bd6296bf5ceb714d63d1d88fa8ccd32759f4f8e52f98be7e7.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 270, + 482, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 414 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 414 + ], + "type": "text", + "content": "Substituting the parameters into the curve expression Eq. (7), we obtain " + }, + { + "bbox": [ + 130, + 270, + 482, + 414 + ], + "type": "inline_equation", + "content": "0 = 0" + }, + { + "bbox": [ + 130, + 270, + 482, + 414 + ], + "type": "text", + "content": ", indicating that u and v are unrestricted. Under the influence of " + }, + { + "bbox": [ + 130, + 270, + 482, + 414 + ], + "type": "inline_equation", + "content": "\\omega = [-100]^{\\top}" + }, + { + "bbox": [ + 130, + 270, + 482, + 414 + ], + "type": "text", + "content": ", the camera's imaging rows are always coplanar with the line, making the distance from any point on the image to the line zero and satisfying the point-to-line distance error function (Sec. 3.2). However, the slope constraint cannot be met. Using this projection equation to calculate the tangent direction of any point on the curve results in indeterminate forms " + }, + { + "bbox": [ + 130, + 270, + 482, + 414 + ], + "type": "inline_equation", + "content": "[0/0, 0/0]" + }, + { + "bbox": [ + 130, + 270, + 482, + 414 + ], + "type": "text", + "content": ", which do not match the measured tangent direction. Therefore, using the tangent direction of points on the curve as the error function prevents this degeneration. As verified in Fig 5, our proposed RSL-BA can suppress plane degeneration, whereas conventional rolling shutter point-based bundle adjustment method NMRsBA [1] converges to a degenerate solution." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 436, + 479, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 479, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 479, + 506 + ], + "type": "text", + "content": "(2) Resistance to 2-views Pure Translation Degeneracy [30]. We first briefly describe the conditions under which this degeneration occurs: when the camera moves in a straight line to take two pictures, assigning any translational velocity to the camera for both shots, it is possible to find new 3D points that satisfy the condition that the reprojection error of the points under this velocity is zero." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 509, + 479, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 479, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 479, + 545 + ], + "type": "text", + "content": "Let's assume the real pose of the camera before the degeneration is: " + }, + { + "bbox": [ + 130, + 509, + 479, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_0 = \\mathbf{I}" + }, + { + "bbox": [ + 130, + 509, + 479, + 545 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 509, + 479, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_0 = \\mathbf{0}" + }, + { + "bbox": [ + 130, + 509, + 479, + 545 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 509, + 479, + 545 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\omega} = [0,0,0]^\\top" + }, + { + "bbox": [ + 130, + 509, + 479, + 545 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 509, + 479, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{d} = [d_1,d_2,d_3]^\\top" + }, + { + "bbox": [ + 130, + 509, + 479, + 545 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 509, + 479, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{L} = (l_{12},l_{13},l_{23},l_{14},l_{24},l_{34})" + }, + { + "bbox": [ + 130, + 509, + 479, + 545 + ], + "type": "text", + "content": " In this case, substitute these parameters into Eq. (7) we can get curve expression:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 142, + 556, + 480, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 556, + 480, + 581 + ], + "spans": [ + { + "bbox": [ + 142, + 556, + 480, + 581 + ], + "type": "interline_equation", + "content": "\\left(l _ {1 4} d _ {3} - l _ {3 4} d _ {1}\\right) v ^ {2} + \\left(l _ {3 4} d _ {2} - l _ {2 4} d _ {3}\\right) u v + \\left(l _ {2 4} d _ {1} - l _ {1 4} d _ {2} + l _ {1 3}\\right) v - l _ {2 3} u - l _ {1 2} = 0 \\tag {16}", + "image_path": "8e9289c432e7c68f0cccdd8006907ad5fc11bb1b7fcfff6f3b76913729ee7295.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 581, + 483, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 483, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 483, + 605 + ], + "type": "text", + "content": "After the degeneration occurs, " + }, + { + "bbox": [ + 130, + 581, + 483, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_0^{\\prime} = \\mathbf{I}" + }, + { + "bbox": [ + 130, + 581, + 483, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 581, + 483, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_0^{\\prime} = \\mathbf{0}" + }, + { + "bbox": [ + 130, + 581, + 483, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 581, + 483, + 605 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\omega}^{\\prime} = [0\\quad 0\\quad 0]^\\top" + }, + { + "bbox": [ + 130, + 581, + 483, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 581, + 483, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{d}' = r[d_1,d_2,d_3]" + }, + { + "bbox": [ + 130, + 581, + 483, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 581, + 483, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{L} = (l_{12}',l_{13}',l_{23}',l_{14}',l_{24}',l_{34}')" + }, + { + "bbox": [ + 130, + 581, + 483, + 605 + ], + "type": "text", + "content": ". In this case, the curve expression is:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 135, + 616, + 489, + 640 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 616, + 489, + 640 + ], + "spans": [ + { + "bbox": [ + 135, + 616, + 489, + 640 + ], + "type": "interline_equation", + "content": "r \\left(l _ {1 4} ^ {\\prime} d _ {3} - l _ {3 4} ^ {\\prime} d _ {1}\\right) v ^ {2} + r \\left(l _ {3 4} ^ {\\prime} d _ {2} - l _ {2 4} ^ {\\prime} d _ {3}\\right) u v + \\left(r \\left(l _ {2 4} ^ {\\prime} d _ {1} - l _ {1 4} ^ {\\prime} d _ {2}\\right) + l _ {1 3} ^ {\\prime}\\right) v - l _ {2 3} ^ {\\prime} u - l _ {1 2} ^ {\\prime} = 0 \\tag {17}", + "image_path": "499da1bad15cffaf69caed2988d7fd6b5f1d801b43c9c303ccc1d27510e2c4eb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "To make the reprojection error zero, the curve expressions before and after degeneration must be the same, meaning the corresponding coefficients are in" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "text", + "content": "Y. Zhang, B. Liao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 173, + 124, + 242, + 205 + ], + "blocks": [ + { + "bbox": [ + 173, + 124, + 242, + 205 + ], + "lines": [ + { + "bbox": [ + 173, + 124, + 242, + 205 + ], + "spans": [ + { + "bbox": [ + 173, + 124, + 242, + 205 + ], + "type": "image", + "image_path": "d28cd589b623bfcde50423900425887091968c032fa33d960b982eecce6e6752.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 210, + 208, + 225, + 220 + ], + "lines": [ + { + "bbox": [ + 210, + 208, + 225, + 220 + ], + "spans": [ + { + "bbox": [ + 210, + 208, + 225, + 220 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 253, + 122, + 349, + 205 + ], + "blocks": [ + { + "bbox": [ + 253, + 122, + 349, + 205 + ], + "lines": [ + { + "bbox": [ + 253, + 122, + 349, + 205 + ], + "spans": [ + { + "bbox": [ + 253, + 122, + 349, + 205 + ], + "type": "image", + "image_path": "0408ba474de9c4e0afae8cf88006c86e329f3306c9b0f8d2e32593c709f8dec1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 301, + 208, + 317, + 220 + ], + "lines": [ + { + "bbox": [ + 301, + 208, + 317, + 220 + ], + "spans": [ + { + "bbox": [ + 301, + 208, + 317, + 220 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 235, + 480, + 278 + ], + "lines": [ + { + "bbox": [ + 131, + 235, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 131, + 235, + 480, + 278 + ], + "type": "text", + "content": "Fig. 3: The illustration of degeneracy resistance of our proposed method in 2-views pure translation degeneracy. (a) Degeneration in single-view scenarios. (b) Suppression of degeneration using line features in dual-view scenarios. (c) Inability of line features to suppress degeneration in special dual-view scenarios" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 354, + 120, + 439, + 205 + ], + "blocks": [ + { + "bbox": [ + 354, + 120, + 439, + 205 + ], + "lines": [ + { + "bbox": [ + 354, + 120, + 439, + 205 + ], + "spans": [ + { + "bbox": [ + 354, + 120, + 439, + 205 + ], + "type": "image", + "image_path": "b0fc43be8c524d8d4afeaa57a2ba0871d77002a78e9187ed1d9a53ccbad4c942.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 396, + 208, + 410, + 220 + ], + "lines": [ + { + "bbox": [ + 396, + 208, + 410, + 220 + ], + "spans": [ + { + "bbox": [ + 396, + 208, + 410, + 220 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 183, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 183, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 183, + 315 + ], + "type": "text", + "content": "proportion." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 203, + 323, + 480, + 385 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 323, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 203, + 323, + 480, + 385 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} l _ {1 4} d _ {3} - l _ {3 4} d _ {1} = s r \\left(l _ {1 4} ^ {\\prime} d _ {3} - l _ {3 4} ^ {\\prime} d _ {1}\\right) \\\\ l _ {3 4} d _ {2} - l _ {2 4} d _ {3} = s r \\left(l _ {3 4} ^ {\\prime} d _ {2} - l _ {2 4} ^ {\\prime} d _ {3}\\right) \\\\ l _ {2 4} d _ {1} - l _ {1 4} d _ {2} + l _ {1 3} = s r \\left(l _ {2 4} ^ {\\prime} d _ {1} - l _ {1 4} ^ {\\prime} d _ {2}\\right) + s l _ {1 3} ^ {\\prime} \\\\ l _ {2 3} = s l _ {2 3} ^ {\\prime} \\\\ l _ {1 2} = s l _ {1 2} ^ {\\prime} \\end{array} \\right. \\tag {18}", + "image_path": "f0846ac594a8981426645e8db90e54bbcce94e0201c4d9209b1b8869e946dbdc.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 393, + 480, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 393, + 480, + 417 + ], + "spans": [ + { + "bbox": [ + 131, + 393, + 480, + 417 + ], + "type": "text", + "content": "Substituting the first and second equations into the third equation yields " + }, + { + "bbox": [ + 131, + 393, + 480, + 417 + ], + "type": "inline_equation", + "content": "l_{13} = sl_{13}'" + }, + { + "bbox": [ + 131, + 393, + 480, + 417 + ], + "type": "text", + "content": ". Combining this with:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 244, + 426, + 480, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 426, + 480, + 452 + ], + "spans": [ + { + "bbox": [ + 244, + 426, + 480, + 452 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} l _ {1 2} l _ {3 4} - l _ {1 3} l _ {2 4} + l _ {1 4} l _ {2 3} = 0 \\\\ l _ {1 2} ^ {\\prime} l _ {3 4} ^ {\\prime} - l _ {1 3} ^ {\\prime} l _ {2 4} ^ {\\prime} + l _ {1 4} ^ {\\prime} l _ {2 3} ^ {\\prime} = 0 \\end{array} \\right. \\tag {19}", + "image_path": "df371f123dd02e63218ee6947528055cdd1dee84be840f01d8fc07d1fd342173.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 460, + 180, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 180, + 470 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 180, + 470 + ], + "type": "text", + "content": "we obtain:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 149, + 479, + 480, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 479, + 480, + 496 + ], + "spans": [ + { + "bbox": [ + 149, + 479, + 480, + 496 + ], + "type": "interline_equation", + "content": "l _ {1 2} ^ {\\prime} = \\frac {l _ {1 2} ^ {\\prime}}{s}, \\quad l _ {1 3} ^ {\\prime} = \\frac {l _ {1 3}}{s}, \\quad l _ {2 3} ^ {\\prime} = \\frac {l _ {2 3}}{s}, \\quad l _ {1 4} ^ {\\prime} = \\frac {l _ {1 4}}{s r}, \\quad l _ {2 4} ^ {\\prime} = \\frac {l _ {2 4}}{s r}, \\quad l _ {3 4} ^ {\\prime} = \\frac {l _ {3 4}}{s r} \\tag {20}", + "image_path": "7468995e667d71e5edea462e0b636b08fa7455b8bb2ad2fc128ff05e98cfa12d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 503, + 479, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 479, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 479, + 624 + ], + "type": "text", + "content": "Note that the direction and normal direction of the line remain unchanged, indicating that the degenerated line stays in the original plane and parallel to the original line, as shown in Fig. 3(1), where " + }, + { + "bbox": [ + 130, + 503, + 479, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{L}^{gt}" + }, + { + "bbox": [ + 130, + 503, + 479, + 624 + ], + "type": "text", + "content": " is the original line, and " + }, + { + "bbox": [ + 130, + 503, + 479, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_1^{dg}" + }, + { + "bbox": [ + 130, + 503, + 479, + 624 + ], + "type": "text", + "content": " is the degenerated line. In the dual-frame scenario (Fig.3(2)), when the plane " + }, + { + "bbox": [ + 130, + 503, + 479, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_1\\mathbf{L}^{gt}" + }, + { + "bbox": [ + 130, + 503, + 479, + 624 + ], + "type": "text", + "content": " is not coplanar with the plane " + }, + { + "bbox": [ + 130, + 503, + 479, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_2\\mathbf{L}^{gt}" + }, + { + "bbox": [ + 130, + 503, + 479, + 624 + ], + "type": "text", + "content": ", their degenerated states " + }, + { + "bbox": [ + 130, + 503, + 479, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_1^{dg}" + }, + { + "bbox": [ + 130, + 503, + 479, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 503, + 479, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_2^{gt}" + }, + { + "bbox": [ + 130, + 503, + 479, + 624 + ], + "type": "text", + "content": " cannot coincide (except at " + }, + { + "bbox": [ + 130, + 503, + 479, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{L}^{gt}" + }, + { + "bbox": [ + 130, + 503, + 479, + 624 + ], + "type": "text", + "content": "). In this case, line constraints can prevent degeneration. However, when " + }, + { + "bbox": [ + 130, + 503, + 479, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{L}^{gt}" + }, + { + "bbox": [ + 130, + 503, + 479, + 624 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 503, + 479, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_1" + }, + { + "bbox": [ + 130, + 503, + 479, + 624 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 503, + 479, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_2" + }, + { + "bbox": [ + 130, + 503, + 479, + 624 + ], + "type": "text", + "content": " are coplanar (Fig.3(3)), line constraints cannot suppress degeneration, though this is a very special condition requiring all lines to meet this criterion. This demonstrates that line constraints are more effective than point constraints in preventing degeneration." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": "(3) Resistance to X-Y Pure Translation Degeneracy (our new founding). We present a pure translation scenario without the z-direction, suitable for" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "RSL-BA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 175, + 120, + 242, + 191 + ], + "blocks": [ + { + "bbox": [ + 175, + 120, + 242, + 191 + ], + "lines": [ + { + "bbox": [ + 175, + 120, + 242, + 191 + ], + "spans": [ + { + "bbox": [ + 175, + 120, + 242, + 191 + ], + "type": "image", + "image_path": "fc83d709053f01e4ee4124c0fa43a052c89b57a4c14e044d1c43d3d9fbbd21fc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 199, + 193, + 210, + 202 + ], + "lines": [ + { + "bbox": [ + 199, + 193, + 210, + 202 + ], + "spans": [ + { + "bbox": [ + 199, + 193, + 210, + 202 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 244, + 129, + 306, + 188 + ], + "blocks": [ + { + "bbox": [ + 244, + 129, + 306, + 188 + ], + "lines": [ + { + "bbox": [ + 244, + 129, + 306, + 188 + ], + "spans": [ + { + "bbox": [ + 244, + 129, + 306, + 188 + ], + "type": "image", + "image_path": "24b87dcae962ef715b22a93d42baa8cc525ad05238c265e9c7b49408574d10fa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 193, + 269, + 202 + ], + "lines": [ + { + "bbox": [ + 258, + 193, + 269, + 202 + ], + "spans": [ + { + "bbox": [ + 258, + 193, + 269, + 202 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 306, + 123, + 369, + 192 + ], + "blocks": [ + { + "bbox": [ + 306, + 123, + 369, + 192 + ], + "lines": [ + { + "bbox": [ + 306, + 123, + 369, + 192 + ], + "spans": [ + { + "bbox": [ + 306, + 123, + 369, + 192 + ], + "type": "image", + "image_path": "deef5e9ffb831fb95b02fb72e98086c7816d501055585862ea7cdfd5e54f60e7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 193, + 346, + 202 + ], + "lines": [ + { + "bbox": [ + 335, + 193, + 346, + 202 + ], + "spans": [ + { + "bbox": [ + 335, + 193, + 346, + 202 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 216, + 482, + 271 + ], + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 271 + ], + "type": "text", + "content": "Fig. 4: In the case of X-Y pure translation, optimization compresses points into a straight line. In a single view, pixel points parallel to the motion trajectory backproject to the same 3D point ((a) and (b)). (c) Different intersection points in space are equidistant from the camera plane. (d) All intersection points formed by backprojection constitute a straight line." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 373, + 121, + 441, + 191 + ], + "blocks": [ + { + "bbox": [ + 373, + 121, + 441, + 191 + ], + "lines": [ + { + "bbox": [ + 373, + 121, + 441, + 191 + ], + "spans": [ + { + "bbox": [ + 373, + 121, + 441, + 191 + ], + "type": "image", + "image_path": "768fe78bc1be96a55cf58fc83086a3e73383f5796a3bbe297b54c3ccf6aaa99a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 396, + 192, + 408, + 201 + ], + "lines": [ + { + "bbox": [ + 396, + 192, + 408, + 201 + ], + "spans": [ + { + "bbox": [ + 396, + 192, + 408, + 201 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 293, + 480, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 480, + 316 + ], + "type": "text", + "content": "multi-view BA, compressing space points into a straight line. Unlike the second type of degeneration, this does not depend on the number of images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "text", + "content": "Firstly, considering a single view, as shown in Fig.4(1), during exposure, the camera moves from position " + }, + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_1^1" + }, + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_1^3" + }, + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "text", + "content": ", where pixel " + }, + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_1^1" + }, + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "text", + "content": " is located on the first row and " + }, + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_1^3" + }, + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "text", + "content": " is located on the " + }, + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "text", + "content": "-th row. Their corresponding spatial points are located on the rays " + }, + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_1^1\\mathbf{q}_1^1" + }, + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_1^3\\mathbf{q}_1^3" + }, + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "text", + "content": ", intersecting at point " + }, + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_1" + }, + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "text", + "content": ". Actually, any point on the line " + }, + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_1^1\\mathbf{q}_1^3" + }, + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "text", + "content": " intersects with its corresponding 3D point at " + }, + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_1" + }, + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "text", + "content": " during exposure. Let " + }, + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_1^2" + }, + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "text", + "content": " be the camera position when exposed to the " + }, + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "inline_equation", + "content": "h_1^2" + }, + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "text", + "content": "-th row, with corresponding feature point " + }, + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_1^2" + }, + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "text", + "content": ". We only consider the plane " + }, + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_1^1\\mathbf{P}_1\\mathbf{O}_1^3" + }, + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "text", + "content": " in Fig.4(2). The motion of " + }, + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{O}" + }, + { + "bbox": [ + 130, + 316, + 482, + 413 + ], + "type": "text", + "content": " is uniform rectilinear motion thus, there exists a length relationship:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 205, + 418, + 480, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 418, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 205, + 418, + 480, + 453 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{c} l _ {\\mathbf {q} _ {1} ^ {1} \\mathbf {q} _ {1} ^ {2}} = \\frac {h _ {1} ^ {2}}{H} l _ {\\mathbf {q} _ {1} ^ {1} \\mathbf {q} _ {1} ^ {3}}, \\quad l _ {\\mathbf {q} _ {3} ^ {1} \\mathbf {q} _ {1} ^ {2}} = \\left(1 - \\frac {h _ {1} ^ {2}}{H}\\right) l _ {\\mathbf {q} _ {1} ^ {1} \\mathbf {q} _ {1} ^ {3}} \\\\ l _ {\\mathbf {O} _ {1} ^ {1} \\mathbf {O} _ {1} ^ {2}} = \\frac {h _ {1} ^ {2}}{H} l _ {\\mathbf {O} _ {1} ^ {1} \\mathbf {O} _ {1} ^ {3}}, \\quad l _ {\\mathbf {O} _ {3} ^ {1} \\mathbf {O} _ {1} ^ {2}} = \\left(1 - \\frac {h _ {1} ^ {2}}{H}\\right) l _ {\\mathbf {O} _ {1} ^ {1} \\mathbf {O} _ {1} ^ {3}} \\end{array} \\right. \\tag {21}", + "image_path": "32e4d72b105ab45c933f1c9c6e2a0dcd0c5ce52b16c95cbe5059ea538adb15a5.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 457, + 185, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 457, + 185, + 468 + ], + "spans": [ + { + "bbox": [ + 131, + 457, + 185, + 468 + ], + "type": "text", + "content": "so we have:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 250, + 464, + 480, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 464, + 480, + 495 + ], + "spans": [ + { + "bbox": [ + 250, + 464, + 480, + 495 + ], + "type": "interline_equation", + "content": "\\frac {l _ {\\mathbf {q} _ {1} ^ {1} \\mathbf {q} _ {1} ^ {2}}}{l _ {\\mathbf {O} _ {1} ^ {1} \\mathbf {O} _ {1} ^ {2}}} = \\frac {l _ {\\mathbf {q} _ {1} ^ {3} \\mathbf {q} _ {1} ^ {2}}}{l _ {\\mathbf {O} _ {3} ^ {1} \\mathbf {O} _ {1} ^ {2}}} = \\frac {l _ {\\mathbf {q} _ {1} ^ {1} \\mathbf {q} _ {1} ^ {3}}}{l _ {\\mathbf {O} _ {1} ^ {1} \\mathbf {O} _ {1} ^ {3}}} \\tag {22}", + "image_path": "221f527a3de1044449d640d2af468894ec34d7551377dd718da2b29176b90b9e.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Therefore, the ray " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_1^2\\mathbf{q}_1^2" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " will also intersect at point " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{P_1}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": ". For every pixel on a line parallel to " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_1^1\\mathbf{O}_1^3" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": ", it can be traced back to a different spatial point. For pixel points on different lines parallel to the motion trajectory, as shown in Fig. 4(3), because the distance " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "l_{q^1 q^3}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " is the same and the camera focal length remains unchanged, according to similar triangles, these spatial points have equal distances to the image plane. Simultaneously, due to symmetry, these points also lie in the plane that is perpendicular to and bisects the line segment " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_1^1\\mathbf{O}_1^2" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": ". Therefore, the intersection line of these two planes represents the final degenerated configuration, which is a line parallel to the x-axis, as shown in Fig.4(4). In multiple views, with no z-direction movement, the imaging planes of different cameras coincide. We assign the same camera configuration to different views, including rotation, translation, angular velocity, and linear velocity. From the single-view case, corresponding points on the pixel plane parallel to the motion trajectory reconstruct to the same spatial point, with zero reprojection error." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "text", + "content": "Y. Zhang, B. Liao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 189, + 116, + 427, + 231 + ], + "blocks": [ + { + "bbox": [ + 189, + 116, + 427, + 231 + ], + "lines": [ + { + "bbox": [ + 189, + 116, + 427, + 231 + ], + "spans": [ + { + "bbox": [ + 189, + 116, + 427, + 231 + ], + "type": "image", + "image_path": "2a7dacdd8b01229a62e27383d3123c2cb57b2ca4b68e4fa1b139f1c753c76f63.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 240, + 482, + 285 + ], + "lines": [ + { + "bbox": [ + 130, + 240, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 482, + 285 + ], + "type": "text", + "content": "Fig.5: Comparison of degenerate resistance ability between NMRSBA [1] and proposed " + }, + { + "bbox": [ + 130, + 240, + 482, + 285 + ], + "type": "inline_equation", + "content": "RSL-BA" + }, + { + "bbox": [ + 130, + 240, + 482, + 285 + ], + "type": "text", + "content": ". Ground truth camera poses and structure are colored with black. This example illustrates that our proposed " + }, + { + "bbox": [ + 130, + 240, + 482, + 285 + ], + "type": "inline_equation", + "content": "RSL-BA" + }, + { + "bbox": [ + 130, + 240, + 482, + 285 + ], + "type": "text", + "content": " has the resistance ability against the plane and X-Y pure translation degeneracy." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 159, + 298, + 459, + 353 + ], + "blocks": [ + { + "bbox": [ + 159, + 298, + 459, + 353 + ], + "lines": [ + { + "bbox": [ + 159, + 298, + 459, + 353 + ], + "spans": [ + { + "bbox": [ + 159, + 298, + 459, + 353 + ], + "type": "image", + "image_path": "9cfe949e393723ebaef0030e53b19349f3de2f40233bede2382dd9015eb1b8fa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 361, + 482, + 384 + ], + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 384 + ], + "type": "text", + "content": "Fig. 6: Accuracy and computational time comparison of the proposed two combinations of reprojection errors." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 407, + 482, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 480 + ], + "type": "text", + "content": "For the line constraints proposed in this paper, " + }, + { + "bbox": [ + 130, + 407, + 482, + 480 + ], + "type": "inline_equation", + "content": "RSL-BA" + }, + { + "bbox": [ + 130, + 407, + 482, + 480 + ], + "type": "text", + "content": " can still effectively suppress degeneration. The proof process is similar to the previous section Sec. 3.4, with the only modification being setting " + }, + { + "bbox": [ + 130, + 407, + 482, + 480 + ], + "type": "inline_equation", + "content": "d_{3} = 0" + }, + { + "bbox": [ + 130, + 407, + 482, + 480 + ], + "type": "text", + "content": " in the camera's motion direction. As verified in Fig 5, our proposed " + }, + { + "bbox": [ + 130, + 407, + 482, + 480 + ], + "type": "inline_equation", + "content": "RSL-BA" + }, + { + "bbox": [ + 130, + 407, + 482, + 480 + ], + "type": "text", + "content": " can suppress X-Y pure translation degeneration, whereas conventional rolling shutter point-based bundle adjustment method NMRSBA [1] converges to a degenerate solution." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 499, + 230, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 499, + 230, + 513 + ], + "spans": [ + { + "bbox": [ + 132, + 499, + 230, + 513 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 525, + 482, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 482, + 574 + ], + "type": "text", + "content": "We compare our method with two SOTA GS-based-method: 1) GSBA [17], 2) GLBA [25], and four SOTA RS-based-method: 1) MRSBA [5], 2) NMRSBA [1], 3) WMRSBA [14], 4) NWRSBA [14]. The experiments are conducted on a laptop with an Intel i7 CPU and all algorithms are implemented in MATLAB." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 592, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 592, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 592, + 482, + 667 + ], + "type": "text", + "content": "Evaluation metrics. The accuracy measures employed encompass standard metrics such as rotation error, translation error, and algorithm efficiency. When conducting experiments related to line BA, we also incorporate error metrics that assess the accuracy of line reconstruction. Given the ground truth rotation " + }, + { + "bbox": [ + 130, + 592, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_g" + }, + { + "bbox": [ + 130, + 592, + 482, + 667 + ], + "type": "text", + "content": ", translation " + }, + { + "bbox": [ + 130, + 592, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_g" + }, + { + "bbox": [ + 130, + 592, + 482, + 667 + ], + "type": "text", + "content": ", and space line " + }, + { + "bbox": [ + 130, + 592, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_g = (\\mathbf{n}_g^\\top, \\mathbf{a}_g^\\top)^\\top" + }, + { + "bbox": [ + 130, + 592, + 482, + 667 + ], + "type": "text", + "content": " with an arbitrary point " + }, + { + "bbox": [ + 130, + 592, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_g" + }, + { + "bbox": [ + 130, + 592, + 482, + 667 + ], + "type": "text", + "content": " on it. Similarly, the optimized parameters are denoted as " + }, + { + "bbox": [ + 130, + 592, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_d" + }, + { + "bbox": [ + 130, + 592, + 482, + 667 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 592, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_d" + }, + { + "bbox": [ + 130, + 592, + 482, + 667 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 592, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_d = (\\mathbf{n}_d^\\top, \\mathbf{a}_d^\\top)^\\top" + }, + { + "bbox": [ + 130, + 592, + 482, + 667 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "RSL-BA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 186, + 135, + 427, + 198 + ], + "blocks": [ + { + "bbox": [ + 150, + 114, + 462, + 126 + ], + "lines": [ + { + "bbox": [ + 150, + 114, + 462, + 126 + ], + "spans": [ + { + "bbox": [ + 150, + 114, + 462, + 126 + ], + "type": "text", + "content": "Table 1: Robustness analysis against multiple levels of Gaussian noise (px)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 186, + 135, + 427, + 198 + ], + "lines": [ + { + "bbox": [ + 186, + 135, + 427, + 198 + ], + "spans": [ + { + "bbox": [ + 186, + 135, + 427, + 198 + ], + "type": "table", + "html": "
noise level (px)0.10.51.01.52.0
Rotation Error7.96e-69.24e-61.50e-51.71e-52.71e-5
Translation Error2.32e-45.93e-48.91e-49.54e-41.55e-3
L-R Error3.36e-33.83e-35.14e-35.22e-36.42e-3
L-D Error2.22e-42.43e-43.90e-44.43e-44.75e-4
", + "image_path": "09e6b95d51c0b51709c85ef14adc60c63605de8df6f01b50a30b9d4cf41cd301.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 187, + 206, + 428, + 318 + ], + "blocks": [ + { + "bbox": [ + 187, + 206, + 428, + 318 + ], + "lines": [ + { + "bbox": [ + 187, + 206, + 428, + 318 + ], + "spans": [ + { + "bbox": [ + 187, + 206, + 428, + 318 + ], + "type": "image", + "image_path": "728eb8ed083f062042a63e82625d599e34fecd599cc267afe8ae7ef11a5918d0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 327, + 448, + 338 + ], + "lines": [ + { + "bbox": [ + 164, + 327, + 448, + 338 + ], + "spans": [ + { + "bbox": [ + 164, + 327, + 448, + 338 + ], + "type": "text", + "content": "Fig. 7: Comparison of accuracy and time among different algorithms." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 360, + 375, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 360, + 375, + 372 + ], + "spans": [ + { + "bbox": [ + 132, + 360, + 375, + 372 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_d" + }, + { + "bbox": [ + 132, + 360, + 375, + 372 + ], + "type": "text", + "content": ", respectively. The evaluation metrics are defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 214, + 378, + 481, + 438 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 378, + 481, + 438 + ], + "spans": [ + { + "bbox": [ + 214, + 378, + 481, + 438 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{c} e _ {\\text {R o t a t i o n}} = \\operatorname {a r c c o s} \\left(\\frac {\\operatorname {T r a c e} \\left(\\mathbf {R} _ {g} ^ {\\top} \\mathbf {R} _ {d}\\right) - 1}{2}\\right) \\\\ e _ {\\text {T r a n s l a t i o n}} = \\operatorname {a r c c o s} \\left(\\mathbf {t} _ {g} ^ {\\top} \\mathbf {t} _ {d} / | | \\mathbf {t} _ {g} | | \\cdot | | \\mathbf {t} _ {d} |\\right) \\\\ e _ {L - R} = \\operatorname {a r c c o s} \\left(\\mathbf {a} _ {g} ^ {\\top} \\mathbf {a} _ {d} / | | \\mathbf {a} _ {g} | | \\cdot | | \\mathbf {a} _ {d} |\\right) \\\\ e _ {L - D} = \\frac {| (\\mathbf {a} _ {g} \\times \\mathbf {a} _ {d}) (\\mathbf {P} _ {d} - \\mathbf {P} _ {g}) |}{\\| \\mathbf {a} _ {g} \\times \\mathbf {a} _ {d} \\|} \\end{array} \\right. \\tag {23}", + "image_path": "4466acce4b0623fb5862c88b7f968bf7f8894bd7e2300ff6cb31043ee029cd2f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 452, + 276, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 452, + 276, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 452, + 276, + 464 + ], + "type": "text", + "content": "4.1 Synthetic Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 469, + 482, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 469, + 482, + 554 + ], + "spans": [ + { + "bbox": [ + 130, + 469, + 482, + 554 + ], + "type": "text", + "content": "Which error term in Sec. 3.2 is better? In the synthetic environment, We set up a cubic box with 12 edges in 3D space and simulated a series of cameras around it to perform RS imaging. The experimental results are shown in Fig. 6. It is evident that, the first type of reprojection error combination performs much better than the second combination in terms of accuracy and time. Therefore, in the forthcoming experiments, we shall exclusively employ the first type of reprojection error combination for comparison." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "How significant is the impact of noise on RSL-BA? Following the same synthetic experimental setup as described above, we additionally introduce Gaussian noise to the endpoints of each line in this experiment. The noise levels are varied from 0.1 pixels to 2.0 pixels, and four different types of errors are recorded for the algorithm at each noise level. For each noise level, 50 experiments are conducted, and the median value is used as the final result. The results, presented in Table 1, demonstrate that RSL-BA provides stable estimations as the noise level increases from 0.1 pixels to 2 pixels." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "text", + "content": "Y. Zhang, B. Liao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 177, + 146, + 433, + 210 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 2: The median absolute trajectory error (ATE) of different methods on WHU-RSVI [3] dataset. The best and second results are shown in green and blue, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 177, + 146, + 433, + 210 + ], + "lines": [ + { + "bbox": [ + 177, + 146, + 433, + 210 + ], + "spans": [ + { + "bbox": [ + 177, + 146, + 433, + 210 + ], + "type": "table", + "html": "
WHU-RSVI1WHU-RSVI2WHU-RSVI3WHU-RSVI4
GSBA0.0809920.0613100.0304040.023698
GLBA0.0761730.0659850.0335540.024961
NMRSBA0.0509690.0416290.0423170.028183
NWRSBA0.0406400.0453130.0354510.022666
RLS-BA(ours)0.04435020.0393140.0233510.020675
", + "image_path": "f01781ca3e9ca6d900b40eede64a571453678cd30e22655036322b068ec05afd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 230, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 230, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 230, + 482, + 422 + ], + "type": "text", + "content": "What are the advantages of " + }, + { + "bbox": [ + 130, + 230, + 482, + 422 + ], + "type": "inline_equation", + "content": "RSL-BA" + }, + { + "bbox": [ + 130, + 230, + 482, + 422 + ], + "type": "text", + "content": " over point-based methods? Within this part, we will conduct a comparative analysis of our method alongside current SOTA methods, including point- and line-based methods. We set up a cubic box of points in the same positions as the 3D lines, totaling 56 points. To ensure fairness, we set up 8 lines with 7 points each to construct error functions, resulting in a total of precisely 56 points. Detailed experimental settings on the number of lines and the number of sampling points on each line are provided in the supplementary materials. We construct two scenarios: regular and classic scenarios with degeneration parallel to the y-axis. Due to significant differences in the experimental results of different methods, we uniformly take the logarithm (base 10) of the results for better observation. The experimental findings are shown in Fig. 7. It is evident that in both situations, our " + }, + { + "bbox": [ + 130, + 230, + 482, + 422 + ], + "type": "inline_equation", + "content": "RSL-BA" + }, + { + "bbox": [ + 130, + 230, + 482, + 422 + ], + "type": "text", + "content": " and the current NWRSBA exhibit comparable accuracy, outperforming alternative approaches based on points and lines. Additionally, the computation time is equivalent to that of NWRSBA. This indicates that the application of lines on Rolling Shutter camera holds significant promise." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 436, + 248, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 436, + 248, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 248, + 449 + ], + "type": "text", + "content": "4.2 Synthetic Images" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 453, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 609 + ], + "type": "text", + "content": "In this section, we conduct experiments on input synthetic images. We use the WHU-RSVI [3] dataset, from which we select two sets of data from trajectory1-fast and trajectory2-fast for 3D reconstruction and pose estimation. We first employ [22] to detect RS curves by segmenting curves into multiple short-line segments and performing line fitting for initialization. The GS line-based SfM [16] is applied to initialize the " + }, + { + "bbox": [ + 130, + 453, + 482, + 609 + ], + "type": "inline_equation", + "content": "RSL-BA" + }, + { + "bbox": [ + 130, + 453, + 482, + 609 + ], + "type": "text", + "content": " parameters. The comparative methods include GSBA, NMRSBA, NWRSBA, and GSLBA. Table 2 shows the median absolute trajectory error of different methods, it can be observed that the proposed " + }, + { + "bbox": [ + 130, + 453, + 482, + 609 + ], + "type": "inline_equation", + "content": "RSL-BA" + }, + { + "bbox": [ + 130, + 453, + 482, + 609 + ], + "type": "text", + "content": " method is the most stable one, achieving optimal or near-optimal results in all cases. Qualitative comparison are also provided in Fig. 8. Unlike point-based methods, line-based methods often achieve good results with fewer feature lines. However, the GSL-BA method is not sufficiently stable when RS effects are prominent." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 624, + 223, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 223, + 636 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 223, + 636 + ], + "type": "text", + "content": "4.3 Real Images" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 484, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 484, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 484, + 665 + ], + "type": "text", + "content": "In this section, we conduct experiments on the real image dataset TUM-RSVI [24]. The experimental setup and comparison methods are similar to those in Sec. 4.2." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "RSL-BA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 192, + 157, + 416, + 270 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 3: The absolute trajectory error (ATE) comparison of different methods in TUM-RSVI [24] dataset. The best and second results are shown in green and blue, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 192, + 157, + 416, + 270 + ], + "lines": [ + { + "bbox": [ + 192, + 157, + 416, + 270 + ], + "spans": [ + { + "bbox": [ + 192, + 157, + 416, + 270 + ], + "type": "table", + "html": "
GSBAGLBANMRSBANWRSBARSL-BA(ours)
seq10.0694840.0864600.0523660.0450640.037816
seq20.0298210.0303790.0260280.0232270.025132
seq30.0651600.0637180.0579180.0550010.048187
seq40.0496130.0520260.0322140.0305340.031903
seq50.0318600.0358390.0194070.0160660.017659
seq60.0619660.0617920.0324340.0246580.026448
seq70.0516210.0565340.0391540.0396200.039701
seq80.0264030.0286900.0248070.0259260.024983
seq90.0983340.0982120.0824810.0735800.080523
seq100.811740.811800.593900.532960.57477
", + "image_path": "b461a3e0dd1311948f20d59dc9d60acc6d3f49d69bf934b4689915efd6bc231d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 190, + 282, + 424, + 432 + ], + "blocks": [ + { + "bbox": [ + 190, + 282, + 424, + 432 + ], + "lines": [ + { + "bbox": [ + 190, + 282, + 424, + 432 + ], + "spans": [ + { + "bbox": [ + 190, + 282, + 424, + 432 + ], + "type": "image", + "image_path": "e0d23ae6a4a6165cb387767c5b262fa2489d2272bb48cb228b4bb2e93ed384fa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 445, + 482, + 479 + ], + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 479 + ], + "type": "text", + "content": "Fig. 8: Comparison of trajectories and 3D reconstruction on WHU-RSVI [3] and TUMRSVI [24] dataset. Each column represents a different bundle adjustment algorithm, and each row represents a different sequence." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 502, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 482, + 552 + ], + "type": "text", + "content": "Table 3 and Fig 8 present some of the experimental results. As can be seen, in most cases, the " + }, + { + "bbox": [ + 130, + 502, + 482, + 552 + ], + "type": "inline_equation", + "content": "RSL-BA" + }, + { + "bbox": [ + 130, + 502, + 482, + 552 + ], + "type": "text", + "content": " method outperforms other methods but is slightly weaker than NwRSBA. This is because the TUM-RSVI lacks line features and they are not visually prominent, making it less suitable for " + }, + { + "bbox": [ + 130, + 502, + 482, + 552 + ], + "type": "inline_equation", + "content": "RSL-BA" + }, + { + "bbox": [ + 130, + 502, + 482, + 552 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 568, + 220, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 568, + 220, + 582 + ], + "spans": [ + { + "bbox": [ + 131, + 568, + 220, + 582 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "This paper presents the first solution of line-based RSL-BA. By utilizing points and tangent directions on curves, we have established a series of faster and more robust curve-to-line re-projection errors. The proposed RSL-BA method can prevent three common degeneracies in RSBA, one of which is newly introduced by us. Extensive experiments in real and synthetic data verify the effectiveness and efficiency of the proposed RSL-BA method." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "text", + "content": "Y. Zhang, B. Liao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 201 + ], + "type": "text", + "content": "This work is supported by the National Key R&D Program of China (No. 2022ZD01190030), Nature Science Foundation of China (No. 62102145), Jiangxi Provincial 03 Special Foundation and 5G Program (Grant No. 20224ABC03A05), Lushan Lab Research Funding, and Changsha Science Fund for Distinguished Young Scholars (kq2306002)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 220, + 197, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 220, + 197, + 232 + ], + "spans": [ + { + "bbox": [ + 133, + 220, + 197, + 232 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 245, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 245, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 480, + 267 + ], + "type": "text", + "content": "1. Albl, C., Sugimoto, A., Pajdla, T.: Degeneracies in rolling shutter sfm. In: ECCV (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 269, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 269, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 269, + 480, + 300 + ], + "type": "text", + "content": "2. Bartoli, A., Sturm, P.: Structure-from-motion using lines: Representation, triangulation, and bundle adjustment. Computer vision and image understanding 100(3), 416-441 (2005)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 302, + 480, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 302, + 480, + 323 + ], + "spans": [ + { + "bbox": [ + 138, + 302, + 480, + 323 + ], + "type": "text", + "content": "3. Cao, L., Ling, J., Xiao, X.: The whu rolling shutter visual-inertial dataset. IEEE Access 8, 50771-50779 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 324, + 480, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 324, + 480, + 345 + ], + "spans": [ + { + "bbox": [ + 138, + 324, + 480, + 345 + ], + "type": "text", + "content": "4. Dai, Y., Li, H., Kneip, L.: Rolling shutter camera relative pose: generalized epipolar geometry. In: CVPR (2016)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 346, + 480, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 480, + 367 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 480, + 367 + ], + "type": "text", + "content": "5. Duchamp, G., Ait-Aider, O., Royer, E., Lavest, J.M.: A rolling shutter compliant method for localisation and reconstruction. In: VISAPP (2015)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 369, + 480, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 369, + 480, + 400 + ], + "spans": [ + { + "bbox": [ + 138, + 369, + 480, + 400 + ], + "type": "text", + "content": "6. Fisher, A., Cannizzaro, R., Cochrane, M., Nagahawatte, C., Palmer, J.L.: Colmap: A memory-efficient occupancy grid mapping framework. Robotics and Autonomous Systems 142, 103755 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 401, + 480, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 401, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 401, + 480, + 422 + ], + "type": "text", + "content": "7. Hartley, R., Zisserman, A.: Multiple view geometry in computer vision. Cambridge university press (2003)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 423, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 423, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 423, + 480, + 445 + ], + "type": "text", + "content": "8. Hedborg, J., Forssen, P.E., Felsberg, M., Ringaby, E.: Rolling shutter bundle adjustment. In: CVPR (2012)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 445, + 480, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 445, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 138, + 445, + 480, + 478 + ], + "type": "text", + "content": "9. Holynski, A., Geraghty, D., Frahm, J.M., Sweeney, C., Szeliski, R.: Reducing drift in structure from motion using extended features. In: 2020 International Conference on 3D Vision (3DV). pp. 51-60. IEEE (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 479, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 500 + ], + "type": "text", + "content": "10. Ito, E., Okatani, T.: Self-calibration-based approach to critical motion sequences of rolling-shutter structure from motion. In: CVPR (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 501, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 480, + 521 + ], + "type": "text", + "content": "1. Lao, Y., Ait-Aider, O.: A robust method for strong rolling shutter effects correction using lines with automatic feature selection. In: CVPR (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 522, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 480, + 544 + ], + "type": "text", + "content": "2. Lao, Y., Ait-Aider, O., Araujo, H.: Robustified structure from motion with rolling-shutter camera using straightness constraint. Pattern Recognition Letters (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 545, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 566 + ], + "type": "text", + "content": "3. Lao, Y., Ait-Aider, O., Bartoli, A.: Solving rolling shutter 3d vision problems using analogies with non-rigidity. IJCV (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "type": "text", + "content": "4. Liao, B., Qu, D., Xue, Y., Zhang, H., Lao, Y.: Revisiting rolling shutter bundle adjustment: Toward accurate and fast solution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4863-4871 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "text", + "content": "5. Lim, H., Jeon, J., Myung, H.: Unconstrained line-based slam using vanishing points for structural mapping. IEEE Robotics and Automation Letters 7(2), 1518-1525 (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "6. Liu, S., Yu, Y., Pautrat, R., Pollefeys, M., Larsson, V.: 3d line mapping revisited. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21445-21455 (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "RSL-BA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 589 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "type": "text", + "content": "17. Lourakis, M.I., Argyros, A.A.: Sba: A software package for generic sparse bundle adjustment. ACM Transactions on Mathematical Software (TOMS) 36(1), 1-30 (2009)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "text", + "content": "18. Micusik, B., Wildenauer, H.: Structure from motion with line segments under relaxed endpoint constraints. International Journal of Computer Vision 124, 65-79 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "text", + "content": "19. Moré, J.J.: The levenberg-marquardt algorithm: implementation and theory. In: Numerical analysis: proceedings of the biennial Conference held at Dundee, June 28-July 1, 1977. pp. 105-116. Springer (2006)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "text", + "content": "20. Patron-Perez, A., Lovegrove, S., Sibley, G.: A spline-based trajectory representation for sensor fusion and rolling shutter cameras. IJCV (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "text", + "content": "21. Pumarola, A., Vakhitov, A., Agudo, A., Sanfeliu, A., Moreno-Noguer, F.: Pl-slam: Real-time monocular visual slam with points and lines. In: 2017 IEEE international conference on robotics and automation (ICRA). pp. 4503-4508. IEEE (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "type": "text", + "content": "22. Purkait, P., Zach, C., Leonardis, A.: Rolling shutter correction in manhattan world. In: ICCV. pp. 882-890 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "text", + "content": "23. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4104-4113 (2016)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 347 + ], + "type": "text", + "content": "24. Schubert, D., Demmel, N., von Stumberg, L., Usenko, V., Cremers, D.: Rolling-shutter modelling for direct visual-inertial odometry. In: IROS (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "25. Taylor, C.J., Kriegman, D.J.: Structure and motion from line segments in multiple images. IEEE Transactions on Pattern Analysis and Machine Intelligence 17(11), 1021-1032 (1995)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 424 + ], + "type": "text", + "content": "26. Triggs, B., McLauchlan, P.F., Hartley, R.I., Fitzgibbon, A.W.: Bundle adjustment—a modern synthesis. In: Vision Algorithms: Theory and Practice: International Workshop on Vision Algorithms Corfu, Greece, September 21–22, 1999 Proceedings. pp. 298–372. Springer (2000)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 424, + 481, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 457 + ], + "type": "text", + "content": "27. Von Gioi, R.G., Jakubowicz, J., Morel, J.M., Randall, G.: Lsd: A fast line segment detector with a false detection control. IEEE transactions on pattern analysis and machine intelligence 32(4), 722-732 (2008)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 500 + ], + "type": "text", + "content": "28. Wei, D., Wan, Y., Zhang, Y., Liu, X., Zhang, B., Wang, X.: Elsr: Efficient line segment reconstruction with planes and points guidance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15807-15815 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 500, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 500, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 500, + 481, + 522 + ], + "type": "text", + "content": "29. Zhuang, B., Cheong, L.F., Lee, G.H.: Rolling-shutter-aware differential sfm and image rectification. In: ICCV (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 522, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 481, + 555 + ], + "type": "text", + "content": "30. Zhuang, B., Tran, Q.H., Ji, P., Cheong, L.F., Chandraker, M.: Learning structure-and-motion-aware rolling shutter correction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4551-4560 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 555, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 589 + ], + "type": "text", + "content": "31. Zuo, X., Xie, X., Liu, Y., Huang, G.: Robust visual slam with point and line features. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1775-1782. IEEE (2017)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "text", + "content": "Y. Zhang, B. Liao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/RT-Pose_ A 4D Radar-Tensor based 3D Human Pose Estimation and Localization Benchmark/e4fd786f-06ad-4f57-8db1-c3bab37f210b_content_list.json b/2024/RT-Pose_ A 4D Radar-Tensor based 3D Human Pose Estimation and Localization Benchmark/e4fd786f-06ad-4f57-8db1-c3bab37f210b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7e78430fd58a3161b7fd574549a7448b286b384e --- /dev/null +++ b/2024/RT-Pose_ A 4D Radar-Tensor based 3D Human Pose Estimation and Localization Benchmark/e4fd786f-06ad-4f57-8db1-c3bab37f210b_content_list.json @@ -0,0 +1,1931 @@ +[ + { + "type": "text", + "text": "RT-Pose: A 4D Radar Tensor-based 3D Human Pose Estimation and Localization Benchmark", + "text_level": 1, + "bbox": [ + 228, + 140, + 774, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuan-Hao $\\mathrm{Ho}^{1*}$ , Jen-Hao Cheng $^{2*}$ , Sheng Yao Kuan $^{2}$ , Zhongyu Jiang $^{2}$ ,", + "bbox": [ + 222, + 210, + 781, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenhao Chai² Hsiang-Wei Huang², Chih-Lung Lin¹, and", + "bbox": [ + 282, + 227, + 720, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jenq-Neng Hwang2", + "bbox": [ + 426, + 243, + 576, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "* Indicates equal contribution", + "bbox": [ + 393, + 258, + 609, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ National Cheng Kung University", + "bbox": [ + 383, + 284, + 617, + 299 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 University of Washington", + "bbox": [ + 406, + 299, + 594, + 313 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "n28081527@gs.ncku.edu.tw1,", + "bbox": [ + 398, + 313, + 601, + 327 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{andyhci, shengyao, zyjiang, wchai, hwhuang, hwang}@uw.edu²,", + "bbox": [ + 269, + 327, + 730, + 340 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cllin@ee.ncku.edu.tw", + "bbox": [ + 416, + 340, + 576, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Traditional methods for human localization and pose estimation (HPE), which mainly rely on RGB images as an input modality, confront substantial limitations in real-world applications due to privacy concerns. In contrast, radar-based HPE methods emerge as a promising alternative, characterized by distinctive attributes such as through-wall recognition and privacy-preserving, rendering the method more conducive to practical deployments. This paper presents a Radar Tensor-based human pose (RT-Pose) dataset and an open-source benchmarking framework. RT-Pose dataset comprises 4D radar tensors, Li-DAR point clouds, and RGB images, and is collected for a total of $72\\mathrm{k}$ frames across 240 sequences with six different complexity level actions. The 4D radar tensor provides raw spatio-temporal information, differentiating it from other radar point cloud-based datasets. We develop an annotation process, which uses RGB images and LiDAR point clouds to accurately label 3D human skeletons. In addition, we propose HRRadarPose, the first single-stage architecture that extracts the high-resolution representation of 4D radar tensors in 3D space to aid human keypoint estimation. HRRadarPose outperforms previous radar-based HPE work on the RT-Pose benchmark. The overall HRRadarPose performance on the RT-Pose dataset, as reflected in a mean per joint position error (MPJPE) of $9.91cm$ , indicates the persistent challenges in achieving accurate HPE in complex real-world scenarios. RT-Pose is available at https://huggingface.co/datasets/uwipl/RT-Pose.", + "bbox": [ + 259, + 391, + 743, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Human Pose Estimation $\\cdot$ 4D Radar $\\cdot$ Dataset $\\cdot$ Benchmark", + "bbox": [ + 261, + 723, + 740, + 737 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 762, + 375, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human localization and 3D pose estimation [6, 20-22, 31, 49, 53, 59] are indispensable in Augmented/Virtual Reality [5, 9, 15, 30], human-computer interaction [18, 51], healthcare [3, 8, 29]. However, capturing human poses across diverse", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "scenarios and generating comprehensive 3D representations for individuals in varying poses present significant challenges. Camera based motion capture systems are often used to collect human body posture [19,28,42]. However, these systems are sensitive to light and require careful multi-camera calibration, making them unreliable for outdoor scenarios and unable to provide overall pose estimation [7]. Furthermore, due to privacy concerns, deploying its application in home or long-term care center scenarios poses challenges.", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Radar-based method emerges as a promising alternative, characterized by distinctive attributes such as through-wall recognition [54,58]. In addition, radar is robust to lighting conditions and resilient to various weather conditions and occlusion [1,47]. These characteristics make it especially suitable for safety- and privacy-critical applications [2,39]. In smart automotive applications, radar, as a complementary sensing modality, can complement adverse scenarios such as lowlighting environments and adverse weather [25], challenging RGB-based sensing. In healthcare applications, RGB-based Human Pose Estimation (HPE)poses privacy risks and is hindered by occlusion, prompting a demand for radar in indoor care environments. Considering the advantages and prospects of radar technology, it is indispensable to provide a dataset that contains diverse scenarios for fostering radar-based HPE research.", + "bbox": [ + 212, + 257, + 787, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The previous radar-based HPE methods [14, 37, 48] often utilize Constant False Alarm Rate (CFAR) techniques [13, 45, 46] to extract radar point clouds. However, the effectiveness of CFAR-based point cloud extraction can be easily influenced by variations in radar module types and hardware parameters. Furthermore, different human body reflections in radar signals across diverse environments may require specific adjustments to CFAR parameters, making it challenging to generalize their applications. To address this issue, some works decompose the 4D radar tensor into vertical and horizontal directions as the input of the model to estimate human pose [26, 47, 50]. This approach computes the magnitude of radar signals for both directions, which effectively reduces data loss during the conversion to point cloud. However, a 4D radar tensor with Doppler values contains the velocity information, which is more informative than a decomposed radar tensor [10, 34]. Therefore, this work focuses on exploring the use of raw 4D radar tensors for HPE and establishing a benchmark based on this data format.", + "bbox": [ + 212, + 441, + 787, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The presented radar tensor-based human pose (RT-Pose) dataset is the first benchmark to integrate calibrated 4D radar tensors, RGB images, and LiDAR point clouds data in the field of HPE. Releasing a multi-modality dataset marks a significant advancement, particularly in the domain of human behavior analyses. There are a total of $72\\mathrm{k}$ frames in 240 sequences in the dataset. The actions are organized into sequences of increasing complexity, providing a realistic motion distribution. All experiments are conducted in 5 environmental conditions in 8 scenarios, which ensures the variety of scenarios for better comprehension of the model. Along with the release of this comprehensive RT-Pose dataset, we also build upon the High-Resolution Network (HRNet) [44] to propose HRRadarPose model, a robust baseline for 3D HPE utilizing 4D radar tensor data. Further", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Ho et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/0ec224de6211e114349e765dc0a44163207ce2a345e31c3335dceeeb40fd315a.jpg", + "table_caption": [ + "Table 1: Comparison of 3D human pose estimation datasets. The first group of rows shows RGB and LiDAR datasets. The second group of rows presents datasets with radar. The radar point cloud is denoted as RPC, and the 4D radar tensor is denoted as 4D-RT. * indicates a modified 4D radar tensor with a specific velocity range." + ], + "table_footnote": [], + "table_body": "
DatasetRGB DepthLiDARRPC4D-RT#Scenes#Actions#Seqs#FPSTime (min)
Human3.6M [19]✓-15839501200
HSC4D [12]✓✓320-2042
SLOPER4D [11]✓✓10-152083
MARS [4]✓110801080
mRI [3]✓✓✓11230010264
HuPR [26]✓✓*-323510240
RT-Pose (Ours)✓✓✓40624010120
", + "bbox": [ + 218, + 212, + 782, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "more, our study also includes a comparative evaluation of different radar data types within our proposed model, as well as benchmarking against previous radar-based methods. These results demonstrate the advantages of employing the 4D radar tensor method over traditional radar point cloud methods, especially in complex actions and diverse scenarios. The 4D radar tensor-based approach presents a challenging but promising direction for future research. In summary, our contributions include:", + "bbox": [ + 212, + 359, + 787, + 467 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The first dataset provides calibrated 4D radar tensors, LiDAR point clouds, and RGB images for human localization and 3D pose estimation of complex actions in diverse scenarios.", + "- We develop a reliable 3D human pose annotation workflow by joint optimization with LiDAR point cloud and RGB images for both outdoor and indoor scenarios.", + "- We propose HRRadarPose, the first single-stage human pose estimator specifically designed to learn 4D radar tensors' representation for the human pose estimation task." + ], + "bbox": [ + 223, + 474, + 784, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 214, + 633, + 397, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 3D Human Pose Estimation Datasets", + "text_level": 1, + "bbox": [ + 214, + 664, + 570, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As one of the fundamental tasks in the computer vision field, there are lots of works introducing datasets and benchmarks for 3D HPE, as shown in Table 1. Human3.6M [19] is the first large-scale dataset for 3D human sensing in lab environments, which contains 3.6-million frames of corresponding 2D and 3D human poses from mocap captured videos of 5 female and 6 male subjects. Compared to Human3.6M, MPI-INF-3DHP [32] is a more challenging 3D human pose dataset captured in the wild. There are 8 subjects with 8 actions captured by 14 cameras covering a greater diversity of poses. 3DPW [41], the first one to include video footage taken from a moving phone camera, includes 60 video sequences. Recently, several works have explored the possibility of HPE from", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "RT-Pose", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "other modalities like radar and LiDAR. mmBody [7] consists of synchronized and calibrated mmWave radar point clouds and RGB-D images in different scenes, as well as skeleton/mesh annotations for humans in the scenes.", + "bbox": [ + 212, + 146, + 782, + 190 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mRI [3], a multi-modal 3D HPE dataset using mmWave, RGB-D, and inertial sensors, generates a radar point cloud dataset based on the CFAR method. There are 12 rehabilitation exercises selected to evaluate radar application in home-based health monitoring. HuPR [26] collects 235 sequences of data by calibrated radar and camera for 2D HPE in an indoor environment.", + "bbox": [ + 212, + 191, + 784, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "All open datasets for radar-based human pose estimation are collected within a limited detection area, which contradicts the characteristics of radar applications for tracking. Although detecting within a small area may ensure system accuracy, it limits the system's practical application. Furthermore, to the best of our knowledge, our dataset is the first to include radar, LiDAR, and RGB sequence information with 3D pose annotations. This allows our dataset to provide pose annotations for both indoor and outdoor environments, distinguishing it from existing datasets.", + "bbox": [ + 212, + 267, + 785, + 387 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Radar-based Human Pose Estimation", + "text_level": 1, + "bbox": [ + 214, + 409, + 571, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In spite of the popularity of RGB-based HPE applications, privacy concerns have driven the exploration of radio or radar modalities for HPE. RF-Pose [54] is a pioneering application in 2D HPE utilizing radio frequency signals. RF-Pose demonstrates the feasibility of pose detection in through-wall scenarios and provide the possibility of predicting the 3D pose using radio signals.", + "bbox": [ + 212, + 431, + 782, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "On the other hand, frequency-modulated continuous wave (FMCW) radar, which operates in the millimeter-wave band (30-300GHz), has been shown to be suitable for detection and HPE tasks. The radar point cloud is widely used to represent the radar signal in HPE [3,4,7,36,37,48,52]. mmMesh [48] introduces a temporal model structure that concurrently captures global and local 3D point cloud structures in spatial dimensions, ensuring precise localization performance for pose estimation. Nonetheless, in cluttered environments, the probability and intensity of radar signal reflections from the human body decrease. Consequently, implementing point cloud-based methods in real-world applications becomes challenging.", + "bbox": [ + 212, + 508, + 784, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compared to the radar point cloud methods, utilizing 4D tensor radar signal proves to be more informative and reliable [26, 47, 55, 56]. RF-pose 3D [55] is the first work using the 4D RF tensor to predict 3D skeletons in 22 different locations. However, the used RF system is not commercially available, which is hard to reproduce. HuPR [26] utilizes two pieces of well-calibrated TI mmWave radar modules to capture vertical and horizontal radar signals, enhancing the resolution of radar elevation signals to improve pose estimation performance. RPM2.0 [47] adopts a similar hardware setup to HuPR but utilizes the HRNet [44] as the backbone and includes an attention model to reconstruct missing keypoints. However, their two-stage model is trained by only walking pose, restricting the actions' complexity for achieving better pose estimation performance. The methods proposed in HuPR and RPM2.0 fuse information from both", + "bbox": [ + 212, + 659, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Ho et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/59fcc3666594ad74f304d77ff1b378a4d0237a8eb9c73cb9433fe295b6b66cca.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 218, + 148, + 416, + 223 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0e1ac6a3978f2d7ec2db7f1789aca6815fdbbe986060c867eec575f47de5e33c.jpg", + "image_caption": [ + "(b)", + "Fig. 2: Data distribution for RT-Pose dataset: (a) Activities; (b) Environmental conditions; (c) Occlusion conditions" + ], + "image_footnote": [], + "bbox": [ + 426, + 148, + 614, + 223 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0ec4447a1856385525a9377a028ada88d698cdbf90aeeecb2c6d03b953fff01c.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 620, + 148, + 772, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "radar modules. While the dual-radar-module configuration potentially captures fine-grained information, its intricate setup poses reproducibility challenges.", + "bbox": [ + 212, + 310, + 782, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike previous works, our system employs a single radar module for vertical and horizontal sensing, simplifying the sensors' calibration and synchronization. We use 4D radar tensor as training data for 3D HPE, reducing the likelihood of data loss during radar point generation. Moreover, RT-Pose dataset includes complex actions and is recorded in diverse scenes, expanding the applicability to real-world scenarios.", + "bbox": [ + 212, + 340, + 784, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 RT-Pose Dataset", + "text_level": 1, + "bbox": [ + 215, + 458, + 415, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To construct the dataset, we recruit 10 participants to perform 6 types of actions. The experiments are divided into two categories. In the first category, participants are required to perform actions while standing, including waving, lifting legs, and random poses to increase the complexity of the actions. A video sequence of random poses involves a subject randomly performing one possible movement at a time, including stretching, bending, twisting, etc., each lasting 3 to 5 seconds. The second category focuses on walking, with additional actions added during the process, such as walking and waving or sitting down", + "bbox": [ + 212, + 492, + 486, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "after walking. Each type of action is inherited from previous radar-based datasets [4,35,48]. Each sequence contains one action for 30 seconds, providing a variety of different difficulty levels. As a result, people can thoroughly evaluate their 3D HPE methods on a broad distribution of actions with diverse difficulties. Furthermore, we provide the dataset development kit to facilitate the usage and evaluation of future methods.", + "bbox": [ + 212, + 750, + 782, + 839 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2f883ef3f042c30e0dc2f69f3e32a699d4c014be7d09a91f0a9648b9c2f41510.jpg", + "image_caption": [ + "Fig.1: Experimental hardware setup in an indoor environment for data collection." + ], + "image_footnote": [], + "bbox": [ + 498, + 523, + 785, + 676 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "RT-Pose", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Sensors", + "text_level": 1, + "bbox": [ + 215, + 146, + 326, + 161 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The data collection hardware system comprises two RGB cameras, a non-repetitive horizontal scanning LiDAR, and a cascade imaging radar module, as shown in Figure 1. To extract human pose information from radar signals, the radar module's parameters have been specifically configured. The radar operates at ten frames per second in this work. The radar module is equipped with 12 transmit (TX) and 16 receive (RX) antenna elements and operates in a multiple-input multiple-output (MIMO) mode [23]. This radar module provides the azimuth and elevation angle resolutions of 1.4 degrees and 18 degrees, respectively. Furthermore, this work implements a high-frequency slope with a long ramp time to efficiently utilize the hardware's sweep bandwidth, ranging from 77 GHz to almost 81 GHz. The longer sweep bandwidth of a chirp increases the radar range resolution [24,37], which is $4.5\\mathrm{cm}$ in this work. Each frame consists of 64 chirps used for calculating radar Doppler signals, providing a velocity resolution of 3.9 cm/s. These carefully chosen configurations ensure the system's accuracy and reliability in capturing human poses.", + "bbox": [ + 212, + 174, + 792, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Data Collection", + "text_level": 1, + "bbox": [ + 215, + 426, + 392, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We collect the dataset in 40 scenes with indoor and outdoor environments. Figure 3 shows some of the collected data. Indoor environments include clean and cluttered conditions, while outdoor ones include normal, rainy weather, and lowlighting conditions. Each experimental condition includes 8 scenarios, which are nonocclusive, and three occlusive conditions (cardboard, cloth, and plastic pad). Each nonocclusive and occlusive condition is recorded in single- and multi-person settings. To consider the different fields of view (FoV) of all sensors, the sensor suite is situated $109~\\mathrm{cm}$ above ground, and an appropriate detection area is planned based on different scenarios.", + "bbox": [ + 212, + 454, + 787, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Data Processing", + "text_level": 1, + "bbox": [ + 215, + 614, + 398, + 631 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The captured radar signal, referred to as 4D tensor data, is processed by a series of steps [16,33,45], as shown in Figure 4. On the transmitter side, the frequency of the transmitted signals, or chirps, periodically increases based on the set frequency slope. Consequently, the received signal exhibits a different frequency than the transmitted signal. The frequency difference between transmitted and received signals can be used to estimate the range of the object reflecting the signal to the radar sensor by the Fast Fourier Transform (FFT). At the beginning of the data processing, each radar frame consists of 64 chirps, resulting in 64 range estimation results. Utilizing this range information, the velocity of the responding object can be measured through FFT, a phenomenon known as the Doppler effect. Second, the radar signal data is re-modulated according to the antenna position. This step is essential for the calculation of the angle of arrival (AoA), enabling the generation of highangular resolution results. Third, the", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Ho et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/79ce2e9e719cb60a92a3492f1dbbf59853ad293e1a7bf2e224873558e2680697.jpg", + "image_caption": [ + "Fig. 3: Experimental instances across various indoor and outdoor conditions with diverse scenarios for data collection." + ], + "image_footnote": [], + "bbox": [ + 256, + 145, + 738, + 334 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b5b68f0127ea85644d9a4d36821f384d89bc2fa95a105c617582ad510eae9a31.jpg", + "image_caption": [ + "Fig. 4: Radar signal processing flow. The green arrow line is for radar point cloud generation and the blue line is for 4D radar tensor generation." + ], + "image_footnote": [], + "bbox": [ + 297, + 393, + 702, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "radar signal from the azimuth and elevation directions also undergoes FFT processing to represent velocity responses in different directions. Finally, the entire radar data is transformed from polar coordinates to Cartesian coordinates. This transformation enhances the intuitiveness for the subsequent pose estimation. In this study, the processed 4D radar tensor has the dimension of $64 \\times 32 \\times 128 \\times 256$ , which correspond to velocity, the $z$ -axis, the $y$ -axis, and the $x$ -axis, respectively.", + "bbox": [ + 212, + 599, + 782, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Annotation Workflow", + "text_level": 1, + "bbox": [ + 214, + 710, + 439, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To achieve accurate human subject detection in a larger area with high-quality ground truth, our system integrates LiDAR and RGB camera data, as illustrated in Figure 5. The images provided by the camera are processed using the HRNet [38] to extract 2D human poses. Then, a 3D HPE model, called ZeDO [22], utilizes the 2D poses to estimate the 3D poses as the initial 3D pseudo ground truth. However, estimating 3D poses using a monocular camera often lacks precision and stability in depth. Therefore, this work incorporates LiDAR for depth", + "bbox": [ + 212, + 733, + 784, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "RT-Pose", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/87050120d069cfb845b98cde6c52d02116fcbac86c8fe86a07a954642d1abdbe.jpg", + "image_caption": [ + "Fig. 5: Workflow of human localization and 3D pose ground truth annotations. Estimated 2D pose results, predicted by the pre-trained HRNet model, are denoted as $P_{2d}$ . The initial setting pose derived from LiDAR point clouds is denoted as $P_{init}$ . Both $P_{2d}$ and $P_{init}$ are inputs into ZeDO, an optimization-based pipeline for 3D pose estimation." + ], + "image_footnote": [], + "bbox": [ + 290, + 142, + 715, + 308 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "estimation of the human subject center. The LiDAR sensor captures environmental information and the subject's point cloud data, ensuring errors are under $2\\mathrm{cm}$ inside the detection range of $20\\mathrm{m}$ . The human subject is manually labeled with a 3D bounding box, which can provide an accurate 3D position.", + "bbox": [ + 212, + 406, + 784, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ZeDO, a state-of-the-art (SOTA) method for cross-dataset HPE, implements an optimization-based pipeline for 3D pose estimation using 2D poses. This method requires 2D poses as input, and ZeDO is able to estimate multiple-hypothesis 3D poses. In this research, the center of the 3D bounding box provided by LiDAR is utilized as the pelvis of the initial pose, which enhances the accuracy of depth estimation compared to relying solely on the monocular camera data, and improves the performance of ZeDO. With the imported 2D pose from HRNet [43], ZeDO iteratively refines the 3D poses. To further improve annotation accuracy, an annotation tool is developed for manual selection and optimization of each frame to retain correct 3D human poses. The 3D HPE results from ZeDO will be put in the LiDAR point cloud for visualization, which is beneficial for humans to filter and modify the error pose annotation results. About $30\\%$ of the data is removed or optimized after human correction, and this process ensures the quality of keypoints annotation. More details can be found in the supplementary material.", + "bbox": [ + 212, + 465, + 787, + 693 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 HRRadarPose", + "text_level": 1, + "bbox": [ + 215, + 715, + 392, + 732 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inspired by HRNet [43], renowned for its adeptness in extracting high-resolution representations, we build an architecture, HRRadarPose, with fully 3D convolutional layers. By treating the Doppler axis, $D$ , as the input channels, our model extracts volumetric spatial features along the $Z$ , $Y$ , and $X$ axes, representing vertical, horizontal, and depth directions in the Cartesian coordinate system. The HRRadarPose architecture is designed to preserve spatial resolution and assimi", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Ho et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "late semantic information from the 4D radar tensors, maintaining high-resolution representations and enabling the exchange of information between features of different resolutions during stage transitions. Figure 6 illustrates an instance of the HRRadarPose structure with three stages. The structure starts with using 3D convolutional layers to transform the 4D radar tensors, $T \\in \\mathbb{R}^{D \\times Z \\times Y \\times X}$ , into features, $f_{in} \\in \\mathbb{R}^{C_1 \\times Z_1 \\times Y_1 \\times X_1}$ . As the network proceeds from one stage to the next, it develops additional branches through strided convolution to widen the receptive field. Each branch $i$ employs strided convolution to elevate its input feature $f_i$ to a feature $f_{i+1}$ in the succeeding branch $i+1$ , simultaneously doubling the channels and halving the spatial dimensions, $f_{i+1} \\in \\mathbb{R}^{2C_i \\times \\frac{Z_i}{2} \\times \\frac{Y_i}{2} \\times \\frac{X_i}{2}}$ . Within each stage, a series of $P$ modules execute $M$ parallel convolutions. To align feature dimensions from different branches for information exchange, we use strided convolution followed by upsampling. At the end of the backbone, we take the feature with the highest resolution to serve as a unified representation for the pose estimation head. In the pose estimation head, two 3D convolutional branches are employed to process these features further. This bifurcation generates two outputs: a distribution map signifying the confidence of locating the human center and the location offset of joint keypoints. The right part of Figure 6 shows the decoding of the HRRadarPose network's output to multi-person poses. The process starts with picking top- $k$ human centers, $C_S$ , with confidence scores over a threshold value. We denote these picked centers' indices as $S$ . Then, we query the centers corresponding joint keypoints' offsets, $K_S$ , by looking up the output map of keypoint offsets head at $S$ . Lastly, we decode their joint keypoint locations, $J_S$ , by summing $K_S$ and $C_S$ . In scenarios with multiple people, we apply range non-maximum suppression (NMS) to reduce overlapping detections. Our method is favorable in two folds: The HRRadarPose's architecture gains efficiency through a single-stage workflow, avoiding complexity brought by region-proposal-based methods [55]. In addition, our pose estimation head inherently learns per person's joint keypoints refraining from ambiguity of association between people's identity and joints.", + "bbox": [ + 212, + 145, + 787, + 603 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To train the center probability head, we use the pixel-wise focal loss as the classification loss, $L_{\\text{class}}$ , defined as:", + "bbox": [ + 212, + 607, + 787, + 638 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {c l a s s}} = \\left\\{ \\begin{array}{l l} (1 - P _ {x y z}) ^ {\\alpha} \\log \\left(P _ {x y z}\\right), & \\text {i f} Y _ {x y z} = 1, \\\\ (1 - Y _ {x y z}) ^ {\\beta} \\left(P _ {x y z}\\right) ^ {\\alpha} \\log \\left(1 - P _ {x y z}\\right), & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 295, + 659, + 785, + 699 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $Y_{xyz}$ is the three-dimensional human center Gaussian distribution map generated by pelvis locations, $P_{xyz}$ is the probability of a person's presence at location, $(x,y,z)$ , $N$ is the number of foreground samples, $\\alpha$ and $\\beta$ are hyperparameters addressing class imbalance and focusing training on foreground samples. To train the keypoint offset head, we construct the regression loss $L_{reg}$ by taking the average of 15 keypoints' $\\mathcal{L}_1$ , comparing the distance between predicted offsets and ground-truth offsets to the center location of a voxel in $Y$ when $Y_{xyz}$ equals to 1. The final loss is a weighted sum of $L_{class}$ and $L_{reg}$ .", + "bbox": [ + 212, + 719, + 787, + 842 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "RT-Pose", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/241762dcba3e0b1f7538100561e8984227e37a1aa699bdb6d14b9d31a7694fbd.jpg", + "image_caption": [ + "Fig. 6: HRRadarPose's architecture consists of a 3D convolutional backbone and a pose estimation head, which generates a body center's confidence map and joint keypoint offsets map. To infer multi-person poses, we decode the prediction by adding the picked bodies' centers with their corresponding keypoint offsets." + ], + "image_footnote": [], + "bbox": [ + 228, + 152, + 774, + 304 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 401, + 375, + 419 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "15 obvious keypoints of a human body are selected, aligning with the body skeleton model of the Human3.6M dataset [19]. Each frame's 3D pose annotation in the sequence is manually validated to ensure the quality of the dataset. Figure 7 demonstrates the testing results across various levels of action complexity, presenting the feasibility of 3D pose estimation using the proposed baseline.", + "bbox": [ + 212, + 431, + 782, + 508 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation metrics We employ joint position error (JPE), which measures the mean of the Euclidean distance for every joint's predicted keypoints and the ground truth keypoints. Localization results are reported by the mean of the root position error (MRPE) [42], which is the JPE of the pelvis keypoint. For the evaluation of 3D HPE, we use Mean per joint position error (MPJPE) [57] and absolute-MPJPE (Abs-MPJPE) [40].", + "bbox": [ + 212, + 523, + 784, + 613 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baselines Comparison We evaluate the performance of the HRRadarPose in global localization and 3D HPE on the RT-Pose dataset. We compare our method with three baseline methods, as shown in Table 2. mm-Pose [37] and mmMesh [48], utilizing radar point clouds for 3D HPE, are re-implemented to test on the RT-Pose dataset. Notably, mmMesh introduces a global localization model, combined with an RNN HPE module, showing more accurate estimation results in both MRPE and MPJPE compared to mm-Pose, which employs a simpler CNN model. RF-Pose 3D [55] uses the 4D radar tensor with a region proposal network to locate the subjects and estimate the 3D pose. The experimental results show that our HRRadarPose outperforms other baseline methods. On the RT-Pose dataset, we obtain $9.93\\mathrm{cm}$ in MPJPE and $9.91\\mathrm{cm}$ in MRPE. We find that radar-tensor-based methods reach less error than the methods using radar point cloud. We attribute this performance gap to complicated scenarios since it is difficult for point cloud-based methods to discriminate different key", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Ho et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/63e42322a68c127528fe73d396d6a271b5ffbabd9b968ec1691aa24153cbd644.jpg", + "image_caption": [ + "R" + ], + "image_footnote": [], + "bbox": [ + 271, + 148, + 344, + 214 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7d52982fa3bf129fca7a2fa5286c7ff1af4f8280866da17da68d75ef34e15992.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 148, + 426, + 214 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/2df9a21432c8938e5aef8da4b312ee6440a1f79b157f2f32da8ba764b43577fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 148, + 504, + 214 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e96abddca39f5299a3c42fa75e5072282cbb909d8eefc2b0302c63bf15aada12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 148, + 586, + 214 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d96b9207b858c26dd2461d2efdc355e75335f2bb5662bff4bc4f630d3497c6f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 148, + 663, + 214 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/303c8131a5567b639ec9976f127c0c0116d037e7a56ca71cda437c23a808f572.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 148, + 743, + 214 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/478994c069a6abb065e22d4cca11396c331995ef319bc31b121c3136b15e6980.jpg", + "image_caption": [ + "BREV TraJ." + ], + "image_footnote": [], + "bbox": [ + 271, + 215, + 344, + 273 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/88098aad8f31ccec0135875708e0825fce8a50307c4210204a3d4332be7b5c08.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 215, + 426, + 273 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/5215f3e12d9cf52b1f1348b838588bdc51682ed34052fc434df41fae5f645614.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 215, + 504, + 273 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/62839eadb20b9da3e78537cdedf9cadd5696da9cc712da1ed23b2c536eb906d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 215, + 586, + 273 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/5d0809a9dac003984daf6643138040c398a899454bd15047d43cac3062305cb2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 215, + 665, + 273 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b74a7d4e620d40b328edd18013fc775ff7b08bd5ec974a30582134958a1e3aab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 215, + 743, + 273 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/a8605e18fe78ab26bfb98f21921c76915d3ccaab1bdbbc4ca6dacac130e8e22a.jpg", + "image_caption": [ + "3D pose", + "(a) Waving hand" + ], + "image_footnote": [], + "bbox": [ + 269, + 277, + 351, + 332 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f8651a21d82d7697fce8e273b7630003399a61ac706632f0cf61e8988e0737d9.jpg", + "image_caption": [ + "(b) Lifting leg" + ], + "image_footnote": [], + "bbox": [ + 352, + 277, + 429, + 332 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/68e55af94ad4b75044313768531c00e7f94ff9e04f9f62878bbc251bdd138fd3.jpg", + "image_caption": [ + "(c) Random pose" + ], + "image_footnote": [], + "bbox": [ + 429, + 277, + 513, + 332 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/8907c83bf211cf42fe61c003fb8fd1b2a42a11f85eebdcd34228c2c413b50112.jpg", + "image_caption": [ + "(d) Normal" + ], + "image_footnote": [], + "bbox": [ + 514, + 277, + 591, + 332 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/a76f2b5b3fd3dd1b7acfc221eb0f91d28f02ed4db7ad0f5044627352c12ba6a7.jpg", + "image_caption": [ + "(e) Sitting" + ], + "image_footnote": [], + "bbox": [ + 591, + 277, + 668, + 332 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/3ae9f7e212ad9f93384faa03c49d95912d49b683c2978e816088973c08b0d12b.jpg", + "image_caption": [ + "(f) Waving hand" + ], + "image_footnote": [], + "bbox": [ + 669, + 277, + 754, + 332 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/2c99d341c382c9eee4bf5a84506dff4d830d506f3db61d65e9124138a8557563.jpg", + "image_caption": [ + "Fig. 7: Visualization of pose estimation results in test frames for six activities. The second row is the trajectory of the test sequence in BEV. The third row is the instance of the global 3D pose estimated result for each action. In both rows, the blue line is the ground truth, and the red line represents estimated results." + ], + "image_footnote": [], + "bbox": [ + 271, + 347, + 751, + 364 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/9bb9b70ad09e3661ad372346bd32d7d5f05ae71ceed48f49162b0fc60b165b5e.jpg", + "table_caption": [ + "Table 2: Comparison of human localization and 3D pose estimation results of our HRRadarPose method and three baseline methods. † represents the radar point cloud-based methods." + ], + "table_footnote": [], + "table_body": "
MethodMRPE (cm)MPJPE (cm)Abs-MPJPE (cm)
mm-Pose † [37]102.2821.26102.7
mmMesh † [48]62.6913.7866.24
RF-Pose3D [55]29.1718.4338.05
HRRadarPose (Ours)9.919.9314.73
", + "bbox": [ + 303, + 505, + 694, + 585 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "points in various poses. The results highlight the importance of using 4D radar tensors, which preserve raw and rich spatial-temporal information for the HPE tasks. An end-to-end architecture directly takes in 4D radar tensors to better face different human action scenarios in the real world.", + "bbox": [ + 212, + 613, + 784, + 674 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Action complexity Analysis RT-posedataset provides six difficulty levels of actions to simulate real-world scenarios, enabling comprehensive evaluation of localization and 3D HPE results. We report the performance of the proposed HRRadarPose method in Table 3. Generally, the joints along the human longitudinal axis, such as the thorax and the head, are more stable than those in the limbs. This phenomenon may be attributed to the larger reflective surface of the body's trunk compared to the limbs, providing a more reliable HPE based on radar signals.", + "bbox": [ + 212, + 688, + 784, + 808 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The MRPE of standing posture is below $7.67~\\mathrm{cm}$ , indicating that the pose complexity slightly influences the localization accuracy. Due to the chair's strong", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "RT-Pose", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/30e002e62e7310ed3bf1d32b0e1f329226939114cee2691e97a0ceb743b6a19a.jpg", + "table_caption": [ + "Table 3: Comparison of the performance predicted by the HRRadar Pose in different complexity level activities, presenting different key points of the pose, the results of localization, and total 3D pose estimation results. The first group of rows lists the stand-related poses, while the other lists the walking-related poses. The complexity level increases from the top activity to the bottom within each group." + ], + "table_footnote": [], + "table_body": "
ActionsJPE(cm)MRPE (cm)MPJPE (cm)Abs-MPJPE (cm)
ThoraxHeadAnkleWrist
Stand: Lifting LegWaving Hand2.953.973.6125.653.356.917.97
6.127.9212.248.887.557.0610.37
Random Pose7.7811.2111.7532.127.6712.4715.12
Walk: SitingNormal4.316.6913.5215.9813.678.9316.92
9.3613.0313.9310.7714.119.4416.83
Waving Hand5.7410.3713.5343.1612.8215.2821.22
", + "bbox": [ + 220, + 226, + 782, + 348 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/0e19c1e2e75c5fecfb092a70f8b0919df69636e28e70b75392881e2db9134d30.jpg", + "image_caption": [ + "Fig. 8: HRRadarPose Results in different conditions. (a), (b), and (c) are occluded by cardboard, cloth, and plastic pad, respectively. (d) is in low-lighting conditions" + ], + "image_footnote": [], + "bbox": [ + 279, + 378, + 725, + 573 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "radar reflection, distinguishing between human subjects and the chair poses a challenge for localization. The MPJPE of walking and sitting is $9.44\\mathrm{cm}$ , showing our HRRadarPose model can reliably estimate the sitting pose. However, the MRPE of this action sequence achieves $14.11\\mathrm{cm}$ , demonstrating that sitting is a key factor for evaluating the performance of localization. The MPJPE of simple actions, such as standing and waving, standing and lifting legs, are around 7 cm, and the walking is $8.93\\mathrm{cm}$ , demonstrating that our HRRadarPose model is comparable to other recently published radar pose estimate methods [26, 47]. However, the current model still encounters difficulty in estimating HPE in complex actions such as walking and waving. Therefore, the RT-pose dataset is a challenging benchmark, which can enable HPE models to realistically learn diverse pose scenarios.", + "bbox": [ + 212, + 656, + 787, + 837 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Ho et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/218ba317d4367ec598643fb6c116508b3bcd4d5c6e878c9695af8cb7bfb0a1f6.jpg", + "table_caption": [ + "Table 4: HPE results from HRRadarPose in normal and occlusive scenarios." + ], + "table_footnote": [], + "table_body": "
OcclusionMPJPE (cm)Abs-MPJPE (cm)
✓9.8414.52
11.5718.55
", + "bbox": [ + 232, + 185, + 473, + 234 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/90c7a4f3275de3dba530c91b31e227e9ef90cc94a99440586eddbc520590f278.jpg", + "table_caption": [ + "Table 5: Ablation Study of Using Dopper information." + ], + "table_footnote": [], + "table_body": "
Doppler|MPJPE (cm) Abs-MPJPE (cm)
11.3122.34
✓9.9114.73
", + "bbox": [ + 524, + 184, + 764, + 234 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/e6f3dafe6a60196acc61c0939f7041d02469409a7da5cd375e57243c48ff9195.jpg", + "table_caption": [ + "Table 6: Our Radar-Doppler Convolutional Block in HR-Net." + ], + "table_footnote": [], + "table_body": "
HRNet-Block-DesignMPJPE (cm)Abs-MPJPE (cm)
Traditional Block (RGB-Vision)10.3517.59
Ours (Radar-Doppler)9.9314.73
", + "bbox": [ + 316, + 276, + 679, + 325 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Qualitative Result Figure 8 demonstrates the qualitative results of different occlusion conditions and a low-lighting instance. Due to the hardware setting of our system, supporting synchronized two camera data, the subject's pose can be visualized by another camera if one of them is occluded. Therefore, the 3D annotation results can be generated even on occlusion conditions. Our method provides reliable pose estimation results even when the radar module is blocked or with low-lighting conditions. Table 4 shows the HPE results of occlusion in normal outdoor conditions. Although the radar module is occluded, the mmWave signal is still robust enough as input for the HRRadarPose model to estimate HPE, demonstrating the characteristics of radar and the importance of using radar for HPE.", + "bbox": [ + 212, + 354, + 787, + 521 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation Study To illustrate the advantage of modeling motion features using Doppler information by our HRRadarPose, we remove Doppler information by averaging 4D radar tensors along the Doppler axis and train the models with and without Doppler information using the same setup. Table 5 shows that with the inclusion of Doppler information, our HRRadarPose enjoys a performance gain due to its capability to extract motion features, enhancing keypoint localization.", + "bbox": [ + 212, + 537, + 787, + 628 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Compared with RPM 2.0, our input data and backbone design are different. The proposed HRRadarPose network extracts better feature representation from 4D radar tensors through modified convolutional blocks. As shown in Figure 9, we replace the batch norm with the group norm of convolutional blocks in HRNet. Since we treat the channel-wise features as a Doppler axis, the group norm inherently handles the feature in different speed groups, which normalizes the data effectively. Table 6 shows the results, proving that the modified CNN block is more suitable for Doppler radar than the original HRNet. Compared with our method, most previous work did not leverage Doppler information [47,50]. We validate that motion features [17,27] are crucial for aiding deep learning models in capturing more human body characteristics over the environment.", + "bbox": [ + 212, + 643, + 787, + 808 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To ablate our design choice of HPE heads, we use the same backbone, replace our pose estimation head with RF-Pose3D's [55] head, and train both models for", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "RT-Pose", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e9d7b332004f99b083bd1c95c5fda7c8bdba0823064fa183233a3bc08fd23c6c.jpg", + "image_caption": [ + "Fig.9: Illustration of two variants of CNN blocks in HRNet. (a) Ours (Radar-Doppler). (b) Traditional Block (RGBVision)." + ], + "image_footnote": [], + "bbox": [ + 261, + 143, + 447, + 214 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/9d19895224c08b7ed7277770a0c80ca7ffe37ef0ea2721cbc716a5e35a09348c.jpg", + "table_caption": [ + "Table 7: Comparison of common 4D radar tensor-based HPE head and ours." + ], + "table_footnote": [], + "table_body": "
HPE HeadMPJPE (cm)Abs-MPJPE (cm)
RF-Pose3D [55]15.5218.08
Ours10.2417.63
", + "bbox": [ + 524, + 212, + 766, + 256 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "the same epochs. RF-Pose3D formulates the HPE as a multi-keypoints classification problem, where their model outputs confidence distributions of every human body joint. As shown in Table 7, our design is favorable in joint keypoints localization in the metric space, indicated by a lower Abs-MPJPE. Moreover, lower MPJPE from our HPE head's result validates stronger inter-joint relationships modeling capability.", + "bbox": [ + 212, + 311, + 784, + 402 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Limitation and Future Work", + "text_level": 1, + "bbox": [ + 215, + 424, + 527, + 439 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "First, to leverage the input data types for training the 3D HPE model, the computational resource is a crucial part. 4D radar tensor in Cartesian coordinate consumes up to $100\\mathrm{MB}$ per frame, which decreases the training speed. Second, we limit the experimental data collection scope from $2\\mathrm{m}$ to $8\\mathrm{m}$ because the camera is hard to clearly capture 2D pose skeletons at a large distance and LiDAR can not easily collect completed point clouds in near vision due to the constraints of FoV. Third, the proposed baseline HRRadarPose lacks robustness in accurately estimating poses during complex activities. Capturing human pose accurately on 4D radar tensor remains an unresolved challenge.", + "bbox": [ + 212, + 454, + 787, + 590 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 612, + 359, + 628 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we propose RT-Pose, the first dataset for human pose estimation (HPE) with synchronized and calibrated 4D radar tensors, LiDAR point clouds, and RGB images. RT-Pose provides cluttered scenarios and human activities of different complexity levels to enhance the difficulty of pose estimation for practical applications. In addition, having various modalities is beneficial for our semi-automatic 3D pose annotation optimization process. The proposed HRRadarPose is the first single-stage architecture designed for estimating human pose using 4D radar tensors as the input. The results indicate that the proposed HRRadarPose outperforms previous works, in which datasets are collected with simpler actions and cleaner scenarios. Our work contributes to offering a three-modality dataset with 4D radar tensors for HPE, which is promising to be applied for complex actions and different scenes. We hope this work encourages future development of 4D radar-based HPE methods.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Ho et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We thank Yizhou Wang $^{3}$ for helping us build the data collection system and the radar data processing pipeline. We thank Hou-I Liu $^{4}$ , Cheng-Yi Huang $^{5}$ , Yi Shiang Chen $^{6}$ , Chi-Yuan Chan $^{7}$ , and Yu-Hsien Lu $^{8}$ for helping us collect the RT-Pose dataset. We thank the National Science and Technology Counci of the Republic of China, Taiwan, under Projects of NSTC 112-2917-I-006-007 for its funding support.", + "bbox": [ + 212, + 180, + 787, + 272 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 297, + 321, + 314 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Adib, F., Hsu, C.Y., Mao, H., Katabi, D., Durand, F.: Rf-capture: Capturing the human figure through a wall. ACM SIGGRAPH Asia (2015)", + "2. Ahuja, K., Jiang, Y., Goel, M., Harrison, C.: Vid2doppler: Synthesizing doppler radar data from videos for training privacy-preserving activity recognition. In: Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. pp. 1-10 (2021)", + "3. An, S., Li, Y., Ogras, U.: mri: Multi-modal 3d human pose estimation dataset using mmwave, rgb-d, and inertial sensors. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems. vol. 35, pp. 27414-27426. Curran Associates, Inc. (2022)", + "4. An, S., Ogras, U.Y.: Mars: mmwave-based assistive rehabilitation system for smart healthcare. ACM Transactions on Embedded Computing Systems (TECS) 20(5s), 1-22 (2021)", + "5. Chai, W., Guo, X., Wang, G., Lu, Y.: Stablevideo: Text-driven consistency-aware diffusion video editing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 23040-23050 (2023)", + "6. Chai, W., Jiang, Z., Hwang, J.N., Wang, G.: Global adaptation meets local generalization: Unsupervised domain adaptation for 3d human pose estimation. arXiv preprint arXiv:2303.16456 (2023)", + "7. Chen, A., Wang, X., Zhu, S., Li, Y., Chen, J., Ye, Q.: mmbody benchmark: 3d body reconstruction dataset and analysis for millimeter wave radar. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 3501-3510 (2022)", + "8. Chen, K., Gabriel, P., Alasfour, A., Gong, C., Doyle, W.K., Devinsky, O., Friedman, D., Dugan, P., Melloni, L., Thesen, T., et al.: Patient-specific pose estimation in clinical environments. IEEE journal of translational engineering in health and medicine 6, 1-11 (2018)", + "9. Cheng, J.H., Chen, Y., Chang, T.Y., Lin, H.E., Wang, P.Y.C., Cheng, L.P.: Impossible staircase: Vertically real walking in an infinite virtual tower. In: 2021 IEEE Virtual Reality and 3D User Interfaces (VR). pp. 50-56 (2021)" + ], + "bbox": [ + 225, + 333, + 785, + 743 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "RT-Pose", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_footnote", + "text": "3 http://yizhouwang.net/", + "bbox": [ + 218, + 753, + 393, + 770 + ], + "page_idx": 14 + }, + { + "type": "page_footnote", + "text": "k39967.c@nycu.edu.tw", + "bbox": [ + 220, + 770, + 383, + 782 + ], + "page_idx": 14 + }, + { + "type": "page_footnote", + "text": "5 n28121521@gs.ncku.edu.tw", + "bbox": [ + 220, + 784, + 413, + 797 + ], + "page_idx": 14 + }, + { + "type": "page_footnote", + "text": "$^{6}$ m16121051@gs.ncku.edu.tw", + "bbox": [ + 220, + 797, + 413, + 811 + ], + "page_idx": 14 + }, + { + "type": "page_footnote", + "text": "7 n26120587@gs.ncku.edu.tw", + "bbox": [ + 220, + 811, + 413, + 825 + ], + "page_idx": 14 + }, + { + "type": "page_footnote", + "text": "n26120587@gs.ncku.edu.tw", + "bbox": [ + 220, + 825, + 413, + 839 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "10. Cheng, J.H., Kuan, S.Y., Liu, H.I., Latapie, H., Liu, G., Hwang, J.N.: Centerradar-net: Joint 3d object detection and tracking framework using 4d fmcw radar. In: 2024 IEEE International Conference on Image Processing (ICIP) (2024)", + "11. Dai, Y., Lin, Y., Lin, X., Wen, C., Xu, L., Yi, H., Shen, S., Ma, Y., Wang, C.: Sloper4d: A scene-aware dataset for global 4d human pose estimation in urban environments. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 682-692 (2023)", + "12. Dai, Y., Lin, Y., Wen, C., Shen, S., Xu, L., Yu, J., Ma, Y., Wang, C.: Hsc4d: Human-centered 4d scene capture in large-scale indoor-outdoor space using wearable imus and lidar. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6792-6802 (2022)", + "13. De Maio, A., Farina, A., Foglia, G.: Design and experimental validation of knowledge-based constant false alarm rate detectors. IET Radar, Sonar & Navigation 1(4), 308-316 (2007)", + "14. Endo, K., Ishikawa, T., Yamamoto, K., Ohtsuki, T.: Multi-person position estimation based on correlation between received signals using mimo fmcw radar. IEEE Access 11, 2610-2620 (2023)", + "15. Guzov, V., Mir, A., Sattler, T., Pons-Moll, G.: Human poseitioning system (hps): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4318-4329 (2021)", + "16. Heath, R.W., Gonzalez-Prelcic, N., Rangan, S., Roh, W., Sayeed, A.M.: An overview of signal processing techniques for millimeter wave mimo systems. IEEE journal of selected topics in signal processing 10(3), 436-453 (2016)", + "17. Huang, H.W., Yang, C.Y., Ramkumar, S., Huang, C.I., Hwang, J.N., Kim, P.K., Lee, K., Kim, K.: Observation centric and central distance recovery for athlete tracking. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 454-460 (2023)", + "18. Huang, H.Y., Ning, C.W., Wang, P.Y., Cheng, J.H., Cheng, L.P.: Haptic-go-round: A surrounding platform for encounter-type haptics in virtual reality experiences. In: Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems. CHI '20, Association for Computing Machinery (2020)", + "19. Ionescu, C., Papava, D., Olaru, V., Sminchisescu, C.: Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence 36(7), 1325-1339 (2013)", + "20. Jiang, Z., Chai, W., Li, L., Zhou, Z., Yang, C.Y., Hwang, J.N.: Unihpe: Towards unified human pose estimation via contrastive learning. arXiv preprint arXiv:2311.16477 (2023)", + "21. Jiang, Z., Ji, H., Yang, C.Y., Hwang, J.N.: 2d human pose estimation calibration and keypoint visibility classification. In: ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 6095-6099. IEEE (2024)", + "22. Jiang, Z., Zhou, Z., Li, L., Chai, W., Yang, C.Y., Hwang, J.N.: Back to optimization: Diffusion-based zero-shot 3d human pose estimation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6142-6152 (2024)", + "23. Kim, Y., Alnujaim, I., Oh, D.: Human activity classification based on point clouds measured by millimeter wave mimo radar with deep recurrent neural networks. IEEE Sensors Journal 21(12), 13522-13529 (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Ho et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "24. Klauder, J.R., Price, A., Darlington, S., Albersheim, W.J.: The theory and design of chirp radars. Bell System Technical Journal 39(4), 745-808 (1960)", + "25. Kuan, S.Y., Cheng, J.H., Huang, H.W., Chai, W., Yang, C.Y., Wu, B.F., Hwang, J.N.: Boosting online 3d multi-object tracking through camera-radar cross check. In: IEEE Intelligent Vehicles Symposium (IV) (2024)", + "26. Lee, S.P., Kini, N.P., Peng, W.H., Ma, C.W., Hwang, J.N.: Hupr: A benchmark for human pose estimation using millimeter wave radar. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5715-5724 (2023)", + "27. Lin, C.L., Chang, Y.M., Hung, C.C., Tu, C.D., Chuang, C.Y.: Position estimation and smooth tracking with a fuzzy-logic-based adaptive strong tracking kalman filter for capacitive touch panels. IEEE Transactions on Industrial Electronics 62(8), 5097-5108 (2015)", + "28. Lin, C.L., Chiu, W.C., Chu, T.C., Ho, Y.H., Chen, F.H., Hsu, C.C., Hsieh, P.H., Chen, C.H., Lin, C.C.K., Sung, P.S., et al.: Innovative head-mounted system based on inertial sensors and magnetometer for detecting falling movements. Sensors 20(20), 5774 (2020)", + "29. Lin, C.L., Ho, Y.H., Chiu, W.C., Chu, T.C., Liu, Y.H.: Innovative shoe-integrated system based on time-of-flight range sensors for fall detection on various terrains. IEEE Sensors Letters 5(10), 1-4 (2021)", + "30. Lin, H.Y., Chen, T.W.: Augmented reality with human body interaction based on monocular 3d pose estimation. In: International Conference on Advanced Concepts for Intelligent Vision Systems. pp. 321-331. Springer (2010)", + "31. Liu, H., He, J.Y., Cheng, Z.Q., Xiang, W., Yang, Q., Chai, W., Wang, G., Bao, X., Luo, B., Geng, Y., et al.: Posynda: Multi-hypothesis pose synthesis domain adaptation for robust 3d human pose estimation. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 5542-5551 (2023)", + "32. Mehta, D., Rhodin, H., Casas, D., Fua, P., Sotnychenko, O., Xu, W., Theobalt, C.: Monocular 3d human pose estimation in the wild using improved cnn supervision. In: 2017 international conference on 3D vision (3DV). pp. 506-516. IEEE (2017)", + "33. Neemat, S., Uysal, F., Krasnov, O., Yarovoy, A.: Reconfigurable range-doppler processing and range resolution improvement for fmcw radar. IEEE Sensors Journal 19(20), 9294-9303 (2019)", + "34. Paek, D.H., Kong, S.H., Wijaya, K.T.: K-radar: 4d radar object detection for autonomous driving in various weather conditions. Advances in Neural Information Processing Systems 35, 3819-3829 (2022)", + "35. Sengupta, A., Cao, S.: mmpose-nlp: A natural language processing approach to precise skeletal pose estimation using mmwave radars. IEEE Transactions on Neural Networks and Learning Systems (2022)", + "36. Sengupta, A., Jin, F., Cao, S.: Nlp based skeletal pose estimation using mmwave radar point-cloud: A simulation approach. In: 2020 IEEE Radar Conference (RadarConf20). pp. 1-6. IEEE (2020)", + "37. Sengupta, A., Jin, F., Zhang, R., Cao, S.: mm-pose: Real-time human skeletal posture estimation using mmwave radars and cnns. IEEE Sensors Journal 20(17), 10032-10044 (2020)", + "38. Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learning for human pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5693-5703 (2019)", + "39. Sun, Z., Ke, Q., Rahmani, H., Bennamoun, M., Wang, G., Liu, J.: Human action recognition from various data modalities: A review. IEEE transactions on pattern analysis and machine intelligence (2022)" + ], + "bbox": [ + 212, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "RT-Pose", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "40. Véges, M., Lorincz, A.: Absolute human pose estimation with depth prediction network. In: 2019 International Joint Conference on Neural Networks (IJCNN). pp. 1-7. IEEE (2019)", + "41. Von Marcard, T., Henschel, R., Black, M.J., Rosenhahn, B., Pons-Moll, G.: Recovering accurate 3d human pose in the wild using imus and a moving camera. In: Proceedings of the European conference on computer vision (ECCV). pp. 601-617 (2018)", + "42. Wang, J., Tan, S., Zhen, X., Xu, S., Zheng, F., He, Z., Shao, L.: Deep 3d human pose estimation: A review. Computer Vision and Image Understanding 210, 103225 (2021)", + "43. Wang, J., Sun, K., Cheng, T., Jiang, B., Deng, C., Zhao, Y., Liu, D., Mu, Y., Tan, M., Wang, X., Liu, W., Xiao, B.: Deep high-resolution representation learning for visual recognition. TPAMI (2019)", + "44. Wang, J., Sun, K., Cheng, T., Jiang, B., Deng, C., Zhao, Y., Liu, D., Mu, Y., Tan, M., Wang, X., et al.: Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence 43(10), 3349-3364 (2020)", + "45. Wang, Y., Cheng, J.H., Huang, J.T., Kuan, S.Y., Fu, Q., Ni, C., Hao, S., Wang, G., Xing, G., Liu, H., Hwang, J.N.: Vision meets mmwave radar: 3d object perception benchmark for autonomous driving. In: IEEE Intelligent Vehicles Symposium (IV) (2024)", + "46. Wang, Y., Jiang, Z., Li, Y., Hwang, J.N., Xing, G., Liu, H.: Rodnet: A real-time radar object detection network cross-supervised by camera-radar fused object 3d localization. IEEE Journal of Selected Topics in Signal Processing 15(4), 954-967 (2021)", + "47. Xie, C., Zhang, D., Wu, Z., Yu, C., Hu, Y., Chen, Y.: RPM 2.0: Rf-based pose machines for multi-person 3d pose estimation. IEEE Transactions on Circuits and Systems for Video Technology (2023)", + "48. Xue, H., Ju, Y., Miao, C., Wang, Y., Wang, S., Zhang, A., Su, L.: mmmesh: Towards 3d real-time dynamic human mesh construction using millimeter-wave. In: Proceedings of the 19th Annual International Conference on Mobile Systems, Applications, and Services. pp. 269-282 (2021)", + "49. Yang, C.Y., Luo, J., Xia, L., Sun, Y., Qiao, N., Zhang, K., Jiang, Z., Hwang, J.N., Kuo, C.H.: Camerapose: Weakly-supervised monocular 3d human pose estimation by leveraging in-the-wild 2d annotations. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2924-2933 (2023)", + "50. Yu, C., Zhang, D., Wu, Z., Xie, C., Lu, Z., Hu, Y., Chen, Y.: Mobirfpose: Portable rf-based 3d human pose camera. IEEE Transactions on Multimedia (2023)", + "51. Yuan, Y., Wei, S.E., Simon, T., Kitani, K., Saragih, J.: Simpoe: Simulated character control for 3d human pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7159-7169 (2021)", + "52. Zhang, G., Geng, X., Lin, Y.J.: Comprehensive mpoint: A method for 3d point cloud generation of human bodies utilizing fmcw mimo mm-wave radar. Sensors 21(19), 6455 (2021)", + "53. Zhang, Z., Chai, W., Jiang, Z., Ye, T., Song, M., Hwang, J.N., Wang, G.: Mpm: A unified 2d-3d human pose representation via masked pose modeling. arXiv preprint arXiv:2306.17201 (2023)", + "54. Zhao, M., Li, T., Abu Alsheikh, M., Tian, Y., Zhao, H., Torralba, A., Katabi, D.: Through-wall human pose estimation using radio signals. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7356-7365 (2018)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y. Ho et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "55. Zhao, M., Tian, Y., Zhao, H., Alsheikh, M.A., Li, T., Hristov, R., Kabelac, Z., Katabi, D., Torralba, A.: Rf-based 3d skeletons. In: Proceedings of the 2018 Conference of the ACM Special Interest Group on Data Communication. pp. 267-281 (2018)", + "56. Zhao, Y., Yarovoy, A., Fioranelli, F.: Angle-insensitive human motion and posture recognition based on 4d imaging radar and deep learning classifiers. IEEE Sensors Journal 22(12), 12173-12182 (2022)", + "57. Zheng, C., Wu, W., Chen, C., Yang, T., Zhu, S., Shen, J., Kehtarnavaz, N., Shah, M.: Deep learning-based human pose estimation: A survey. ACM Computing Surveys 56(1), 1-37 (2023)", + "58. Zheng, Z., Pan, J., Ni, Z., Shi, C., Ye, S., Fang, G.: Human posture reconstruction for through-the-wall radar imaging using convolutional neural networks. IEEE Geoscience and Remote Sensing Letters 19, 1-5 (2021)", + "59. Zhou, Z., Jiang, Z., Chai, W., Yang, C.Y., Li, L., Hwang, J.N.: Efficient domain adaptation via generative prior for 3d infant pose estimation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 41-49 (2024)" + ], + "bbox": [ + 212, + 146, + 787, + 383 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "RT-Pose", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/RT-Pose_ A 4D Radar-Tensor based 3D Human Pose Estimation and Localization Benchmark/e4fd786f-06ad-4f57-8db1-c3bab37f210b_model.json b/2024/RT-Pose_ A 4D Radar-Tensor based 3D Human Pose Estimation and Localization Benchmark/e4fd786f-06ad-4f57-8db1-c3bab37f210b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2c7dd5b8b0db0fcd6062448354cd815dcd642e7f --- /dev/null +++ b/2024/RT-Pose_ A 4D Radar-Tensor based 3D Human Pose Estimation and Localization Benchmark/e4fd786f-06ad-4f57-8db1-c3bab37f210b_model.json @@ -0,0 +1,2746 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.23, + 0.141, + 0.776, + 0.187 + ], + "angle": 0, + "content": "RT-Pose: A 4D Radar Tensor-based 3D Human Pose Estimation and Localization Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.212, + 0.782, + 0.228 + ], + "angle": 0, + "content": "Yuan-Hao \\(\\mathrm{Ho}^{1*}\\), Jen-Hao Cheng\\(^{2*}\\), Sheng Yao Kuan\\(^{2}\\), Zhongyu Jiang\\(^{2}\\)," + }, + { + "type": "text", + "bbox": [ + 0.284, + 0.228, + 0.721, + 0.244 + ], + "angle": 0, + "content": "Wenhao Chai² Hsiang-Wei Huang², Chih-Lung Lin¹, and" + }, + { + "type": "text", + "bbox": [ + 0.427, + 0.244, + 0.578, + 0.259 + ], + "angle": 0, + "content": "Jenq-Neng Hwang2" + }, + { + "type": "text", + "bbox": [ + 0.394, + 0.259, + 0.61, + 0.274 + ], + "angle": 0, + "content": "* Indicates equal contribution" + }, + { + "type": "text", + "bbox": [ + 0.384, + 0.285, + 0.619, + 0.3 + ], + "angle": 0, + "content": "\\(^{1}\\) National Cheng Kung University" + }, + { + "type": "text", + "bbox": [ + 0.408, + 0.3, + 0.595, + 0.314 + ], + "angle": 0, + "content": "2 University of Washington" + }, + { + "type": "text", + "bbox": [ + 0.4, + 0.314, + 0.602, + 0.328 + ], + "angle": 0, + "content": "n28081527@gs.ncku.edu.tw1," + }, + { + "type": "text", + "bbox": [ + 0.27, + 0.328, + 0.732, + 0.342 + ], + "angle": 0, + "content": "{andyhci, shengyao, zyjiang, wchai, hwhuang, hwang}@uw.edu²," + }, + { + "type": "text", + "bbox": [ + 0.417, + 0.342, + 0.578, + 0.354 + ], + "angle": 0, + "content": "cllin@ee.ncku.edu.tw" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.392, + 0.744, + 0.712 + ], + "angle": 0, + "content": "Abstract. Traditional methods for human localization and pose estimation (HPE), which mainly rely on RGB images as an input modality, confront substantial limitations in real-world applications due to privacy concerns. In contrast, radar-based HPE methods emerge as a promising alternative, characterized by distinctive attributes such as through-wall recognition and privacy-preserving, rendering the method more conducive to practical deployments. This paper presents a Radar Tensor-based human pose (RT-Pose) dataset and an open-source benchmarking framework. RT-Pose dataset comprises 4D radar tensors, Li-DAR point clouds, and RGB images, and is collected for a total of \\(72\\mathrm{k}\\) frames across 240 sequences with six different complexity level actions. The 4D radar tensor provides raw spatio-temporal information, differentiating it from other radar point cloud-based datasets. We develop an annotation process, which uses RGB images and LiDAR point clouds to accurately label 3D human skeletons. In addition, we propose HRRadarPose, the first single-stage architecture that extracts the high-resolution representation of 4D radar tensors in 3D space to aid human keypoint estimation. HRRadarPose outperforms previous radar-based HPE work on the RT-Pose benchmark. The overall HRRadarPose performance on the RT-Pose dataset, as reflected in a mean per joint position error (MPJPE) of \\(9.91cm\\), indicates the persistent challenges in achieving accurate HPE in complex real-world scenarios. RT-Pose is available at https://huggingface.co/datasets/uwipl/RT-Pose." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.724, + 0.741, + 0.738 + ], + "angle": 0, + "content": "Keywords: Human Pose Estimation \\(\\cdot\\) 4D Radar \\(\\cdot\\) Dataset \\(\\cdot\\) Benchmark" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.763, + 0.377, + 0.78 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Human localization and 3D pose estimation [6, 20-22, 31, 49, 53, 59] are indispensable in Augmented/Virtual Reality [5, 9, 15, 30], human-computer interaction [18, 51], healthcare [3, 8, 29]. However, capturing human poses across diverse" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "Y. Ho et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "scenarios and generating comprehensive 3D representations for individuals in varying poses present significant challenges. Camera based motion capture systems are often used to collect human body posture [19,28,42]. However, these systems are sensitive to light and require careful multi-camera calibration, making them unreliable for outdoor scenarios and unable to provide overall pose estimation [7]. Furthermore, due to privacy concerns, deploying its application in home or long-term care center scenarios poses challenges." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.258, + 0.788, + 0.438 + ], + "angle": 0, + "content": "Radar-based method emerges as a promising alternative, characterized by distinctive attributes such as through-wall recognition [54,58]. In addition, radar is robust to lighting conditions and resilient to various weather conditions and occlusion [1,47]. These characteristics make it especially suitable for safety- and privacy-critical applications [2,39]. In smart automotive applications, radar, as a complementary sensing modality, can complement adverse scenarios such as lowlighting environments and adverse weather [25], challenging RGB-based sensing. In healthcare applications, RGB-based Human Pose Estimation (HPE)poses privacy risks and is hindered by occlusion, prompting a demand for radar in indoor care environments. Considering the advantages and prospects of radar technology, it is indispensable to provide a dataset that contains diverse scenarios for fostering radar-based HPE research." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.788, + 0.669 + ], + "angle": 0, + "content": "The previous radar-based HPE methods [14, 37, 48] often utilize Constant False Alarm Rate (CFAR) techniques [13, 45, 46] to extract radar point clouds. However, the effectiveness of CFAR-based point cloud extraction can be easily influenced by variations in radar module types and hardware parameters. Furthermore, different human body reflections in radar signals across diverse environments may require specific adjustments to CFAR parameters, making it challenging to generalize their applications. To address this issue, some works decompose the 4D radar tensor into vertical and horizontal directions as the input of the model to estimate human pose [26, 47, 50]. This approach computes the magnitude of radar signals for both directions, which effectively reduces data loss during the conversion to point cloud. However, a 4D radar tensor with Doppler values contains the velocity information, which is more informative than a decomposed radar tensor [10, 34]. Therefore, this work focuses on exploring the use of raw 4D radar tensors for HPE and establishing a benchmark based on this data format." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The presented radar tensor-based human pose (RT-Pose) dataset is the first benchmark to integrate calibrated 4D radar tensors, RGB images, and LiDAR point clouds data in the field of HPE. Releasing a multi-modality dataset marks a significant advancement, particularly in the domain of human behavior analyses. There are a total of \\(72\\mathrm{k}\\) frames in 240 sequences in the dataset. The actions are organized into sequences of increasing complexity, providing a realistic motion distribution. All experiments are conducted in 5 environmental conditions in 8 scenarios, which ensures the variety of scenarios for better comprehension of the model. Along with the release of this comprehensive RT-Pose dataset, we also build upon the High-Resolution Network (HRNet) [44] to propose HRRadarPose model, a robust baseline for 3D HPE utilizing 4D radar tensor data. Further" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RT-Pose" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 1: Comparison of 3D human pose estimation datasets. The first group of rows shows RGB and LiDAR datasets. The second group of rows presents datasets with radar. The radar point cloud is denoted as RPC, and the 4D radar tensor is denoted as 4D-RT. * indicates a modified 4D radar tensor with a specific velocity range." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.213, + 0.784, + 0.334 + ], + "angle": 0, + "content": "
DatasetRGB DepthLiDARRPC4D-RT#Scenes#Actions#Seqs#FPSTime (min)
Human3.6M [19]✓-15839501200
HSC4D [12]✓✓320-2042
SLOPER4D [11]✓✓10-152083
MARS [4]✓110801080
mRI [3]✓✓✓11230010264
HuPR [26]✓✓*-323510240
RT-Pose (Ours)✓✓✓40624010120
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.361, + 0.788, + 0.468 + ], + "angle": 0, + "content": "more, our study also includes a comparative evaluation of different radar data types within our proposed model, as well as benchmarking against previous radar-based methods. These results demonstrate the advantages of employing the 4D radar tensor method over traditional radar point cloud methods, especially in complex actions and diverse scenarios. The 4D radar tensor-based approach presents a challenging but promising direction for future research. In summary, our contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.476, + 0.785, + 0.519 + ], + "angle": 0, + "content": "- The first dataset provides calibrated 4D radar tensors, LiDAR point clouds, and RGB images for human localization and 3D pose estimation of complex actions in diverse scenarios." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.521, + 0.785, + 0.564 + ], + "angle": 0, + "content": "- We develop a reliable 3D human pose annotation workflow by joint optimization with LiDAR point cloud and RGB images for both outdoor and indoor scenarios." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.566, + 0.785, + 0.61 + ], + "angle": 0, + "content": "- We propose HRRadarPose, the first single-stage human pose estimator specifically designed to learn 4D radar tensors' representation for the human pose estimation task." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.476, + 0.785, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.634, + 0.398, + 0.65 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.665, + 0.571, + 0.68 + ], + "angle": 0, + "content": "2.1 3D Human Pose Estimation Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "As one of the fundamental tasks in the computer vision field, there are lots of works introducing datasets and benchmarks for 3D HPE, as shown in Table 1. Human3.6M [19] is the first large-scale dataset for 3D human sensing in lab environments, which contains 3.6-million frames of corresponding 2D and 3D human poses from mocap captured videos of 5 female and 6 male subjects. Compared to Human3.6M, MPI-INF-3DHP [32] is a more challenging 3D human pose dataset captured in the wild. There are 8 subjects with 8 actions captured by 14 cameras covering a greater diversity of poses. 3DPW [41], the first one to include video footage taken from a moving phone camera, includes 60 video sequences. Recently, several works have explored the possibility of HPE from" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "Y. Ho et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.191 + ], + "angle": 0, + "content": "other modalities like radar and LiDAR. mmBody [7] consists of synchronized and calibrated mmWave radar point clouds and RGB-D images in different scenes, as well as skeleton/mesh annotations for humans in the scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.785, + 0.266 + ], + "angle": 0, + "content": "mRI [3], a multi-modal 3D HPE dataset using mmWave, RGB-D, and inertial sensors, generates a radar point cloud dataset based on the CFAR method. There are 12 rehabilitation exercises selected to evaluate radar application in home-based health monitoring. HuPR [26] collects 235 sequences of data by calibrated radar and camera for 2D HPE in an indoor environment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.787, + 0.388 + ], + "angle": 0, + "content": "All open datasets for radar-based human pose estimation are collected within a limited detection area, which contradicts the characteristics of radar applications for tracking. Although detecting within a small area may ensure system accuracy, it limits the system's practical application. Furthermore, to the best of our knowledge, our dataset is the first to include radar, LiDAR, and RGB sequence information with 3D pose annotations. This allows our dataset to provide pose annotations for both indoor and outdoor environments, distinguishing it from existing datasets." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.41, + 0.573, + 0.423 + ], + "angle": 0, + "content": "2.2 Radar-based Human Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.784, + 0.508 + ], + "angle": 0, + "content": "In spite of the popularity of RGB-based HPE applications, privacy concerns have driven the exploration of radio or radar modalities for HPE. RF-Pose [54] is a pioneering application in 2D HPE utilizing radio frequency signals. RF-Pose demonstrates the feasibility of pose detection in through-wall scenarios and provide the possibility of predicting the 3D pose using radio signals." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.785, + 0.659 + ], + "angle": 0, + "content": "On the other hand, frequency-modulated continuous wave (FMCW) radar, which operates in the millimeter-wave band (30-300GHz), has been shown to be suitable for detection and HPE tasks. The radar point cloud is widely used to represent the radar signal in HPE [3,4,7,36,37,48,52]. mmMesh [48] introduces a temporal model structure that concurrently captures global and local 3D point cloud structures in spatial dimensions, ensuring precise localization performance for pose estimation. Nonetheless, in cluttered environments, the probability and intensity of radar signal reflections from the human body decrease. Consequently, implementing point cloud-based methods in real-world applications becomes challenging." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Compared to the radar point cloud methods, utilizing 4D tensor radar signal proves to be more informative and reliable [26, 47, 55, 56]. RF-pose 3D [55] is the first work using the 4D RF tensor to predict 3D skeletons in 22 different locations. However, the used RF system is not commercially available, which is hard to reproduce. HuPR [26] utilizes two pieces of well-calibrated TI mmWave radar modules to capture vertical and horizontal radar signals, enhancing the resolution of radar elevation signals to improve pose estimation performance. RPM2.0 [47] adopts a similar hardware setup to HuPR but utilizes the HRNet [44] as the backbone and includes an attention model to reconstruct missing keypoints. However, their two-stage model is trained by only walking pose, restricting the actions' complexity for achieving better pose estimation performance. The methods proposed in HuPR and RPM2.0 fuse information from both" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RT-Pose" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.149, + 0.418, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.307, + 0.228, + 0.324, + 0.239 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.149, + 0.616, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.228, + 0.529, + 0.239 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.149, + 0.774, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.692, + 0.228, + 0.708, + 0.239 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.251, + 0.785, + 0.278 + ], + "angle": 0, + "content": "Fig. 2: Data distribution for RT-Pose dataset: (a) Activities; (b) Environmental conditions; (c) Occlusion conditions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.311, + 0.784, + 0.34 + ], + "angle": 0, + "content": "radar modules. While the dual-radar-module configuration potentially captures fine-grained information, its intricate setup poses reproducibility challenges." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.341, + 0.785, + 0.43 + ], + "angle": 0, + "content": "Unlike previous works, our system employs a single radar module for vertical and horizontal sensing, simplifying the sensors' calibration and synchronization. We use 4D radar tensor as training data for 3D HPE, reducing the likelihood of data loss during radar point generation. Moreover, RT-Pose dataset includes complex actions and is recorded in diverse scenes, expanding the applicability to real-world scenarios." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.459, + 0.416, + 0.473 + ], + "angle": 0, + "content": "3 RT-Pose Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.487, + 0.749 + ], + "angle": 0, + "content": "To construct the dataset, we recruit 10 participants to perform 6 types of actions. The experiments are divided into two categories. In the first category, participants are required to perform actions while standing, including waving, lifting legs, and random poses to increase the complexity of the actions. A video sequence of random poses involves a subject randomly performing one possible movement at a time, including stretching, bending, twisting, etc., each lasting 3 to 5 seconds. The second category focuses on walking, with additional actions added during the process, such as walking and waving or sitting down" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.784, + 0.84 + ], + "angle": 0, + "content": "after walking. Each type of action is inherited from previous radar-based datasets [4,35,48]. Each sequence contains one action for 30 seconds, providing a variety of different difficulty levels. As a result, people can thoroughly evaluate their 3D HPE methods on a broad distribution of actions with diverse difficulties. Furthermore, we provide the dataset development kit to facilitate the usage and evaluation of future methods." + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.524, + 0.787, + 0.677 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.686, + 0.785, + 0.713 + ], + "angle": 0, + "content": "Fig.1: Experimental hardware setup in an indoor environment for data collection." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "Y. Ho et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.327, + 0.162 + ], + "angle": 0, + "content": "3.1 Sensors" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.794, + 0.403 + ], + "angle": 0, + "content": "The data collection hardware system comprises two RGB cameras, a non-repetitive horizontal scanning LiDAR, and a cascade imaging radar module, as shown in Figure 1. To extract human pose information from radar signals, the radar module's parameters have been specifically configured. The radar operates at ten frames per second in this work. The radar module is equipped with 12 transmit (TX) and 16 receive (RX) antenna elements and operates in a multiple-input multiple-output (MIMO) mode [23]. This radar module provides the azimuth and elevation angle resolutions of 1.4 degrees and 18 degrees, respectively. Furthermore, this work implements a high-frequency slope with a long ramp time to efficiently utilize the hardware's sweep bandwidth, ranging from 77 GHz to almost 81 GHz. The longer sweep bandwidth of a chirp increases the radar range resolution [24,37], which is \\(4.5\\mathrm{cm}\\) in this work. Each frame consists of 64 chirps used for calculating radar Doppler signals, providing a velocity resolution of 3.9 cm/s. These carefully chosen configurations ensure the system's accuracy and reliability in capturing human poses." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.427, + 0.393, + 0.442 + ], + "angle": 0, + "content": "3.2 Data Collection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.788, + 0.592 + ], + "angle": 0, + "content": "We collect the dataset in 40 scenes with indoor and outdoor environments. Figure 3 shows some of the collected data. Indoor environments include clean and cluttered conditions, while outdoor ones include normal, rainy weather, and lowlighting conditions. Each experimental condition includes 8 scenarios, which are nonocclusive, and three occlusive conditions (cardboard, cloth, and plastic pad). Each nonocclusive and occlusive condition is recorded in single- and multi-person settings. To consider the different fields of view (FoV) of all sensors, the sensor suite is situated \\(109~\\mathrm{cm}\\) above ground, and an appropriate detection area is planned based on different scenarios." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.616, + 0.399, + 0.632 + ], + "angle": 0, + "content": "3.3 Data Processing" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The captured radar signal, referred to as 4D tensor data, is processed by a series of steps [16,33,45], as shown in Figure 4. On the transmitter side, the frequency of the transmitted signals, or chirps, periodically increases based on the set frequency slope. Consequently, the received signal exhibits a different frequency than the transmitted signal. The frequency difference between transmitted and received signals can be used to estimate the range of the object reflecting the signal to the radar sensor by the Fast Fourier Transform (FFT). At the beginning of the data processing, each radar frame consists of 64 chirps, resulting in 64 range estimation results. Utilizing this range information, the velocity of the responding object can be measured through FFT, a phenomenon known as the Doppler effect. Second, the radar signal data is re-modulated according to the antenna position. This step is essential for the calculation of the angle of arrival (AoA), enabling the generation of highangular resolution results. Third, the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RT-Pose" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.146, + 0.74, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.349, + 0.784, + 0.377 + ], + "angle": 0, + "content": "Fig. 3: Experimental instances across various indoor and outdoor conditions with diverse scenarios for data collection." + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.394, + 0.704, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.546, + 0.784, + 0.575 + ], + "angle": 0, + "content": "Fig. 4: Radar signal processing flow. The green arrow line is for radar point cloud generation and the blue line is for 4D radar tensor generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.601, + 0.784, + 0.693 + ], + "angle": 0, + "content": "radar signal from the azimuth and elevation directions also undergoes FFT processing to represent velocity responses in different directions. Finally, the entire radar data is transformed from polar coordinates to Cartesian coordinates. This transformation enhances the intuitiveness for the subsequent pose estimation. In this study, the processed 4D radar tensor has the dimension of \\(64 \\times 32 \\times 128 \\times 256\\), which correspond to velocity, the \\(z\\)-axis, the \\(y\\)-axis, and the \\(x\\)-axis, respectively." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.712, + 0.44, + 0.726 + ], + "angle": 0, + "content": "3.4 Annotation Workflow" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.785, + 0.84 + ], + "angle": 0, + "content": "To achieve accurate human subject detection in a larger area with high-quality ground truth, our system integrates LiDAR and RGB camera data, as illustrated in Figure 5. The images provided by the camera are processed using the HRNet [38] to extract 2D human poses. Then, a 3D HPE model, called ZeDO [22], utilizes the 2D poses to estimate the 3D poses as the initial 3D pseudo ground truth. However, estimating 3D poses using a monocular camera often lacks precision and stability in depth. Therefore, this work incorporates LiDAR for depth" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "Y. Ho et al." + }, + { + "type": "image", + "bbox": [ + 0.292, + 0.143, + 0.716, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.32, + 0.788, + 0.378 + ], + "angle": 0, + "content": "Fig. 5: Workflow of human localization and 3D pose ground truth annotations. Estimated 2D pose results, predicted by the pre-trained HRNet model, are denoted as \\( P_{2d} \\). The initial setting pose derived from LiDAR point clouds is denoted as \\( P_{init} \\). Both \\( P_{2d} \\) and \\( P_{init} \\) are inputs into ZeDO, an optimization-based pipeline for 3D pose estimation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.407, + 0.785, + 0.466 + ], + "angle": 0, + "content": "estimation of the human subject center. The LiDAR sensor captures environmental information and the subject's point cloud data, ensuring errors are under \\(2\\mathrm{cm}\\) inside the detection range of \\(20\\mathrm{m}\\). The human subject is manually labeled with a 3D bounding box, which can provide an accurate 3D position." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.467, + 0.788, + 0.694 + ], + "angle": 0, + "content": "ZeDO, a state-of-the-art (SOTA) method for cross-dataset HPE, implements an optimization-based pipeline for 3D pose estimation using 2D poses. This method requires 2D poses as input, and ZeDO is able to estimate multiple-hypothesis 3D poses. In this research, the center of the 3D bounding box provided by LiDAR is utilized as the pelvis of the initial pose, which enhances the accuracy of depth estimation compared to relying solely on the monocular camera data, and improves the performance of ZeDO. With the imported 2D pose from HRNet [43], ZeDO iteratively refines the 3D poses. To further improve annotation accuracy, an annotation tool is developed for manual selection and optimization of each frame to retain correct 3D human poses. The 3D HPE results from ZeDO will be put in the LiDAR point cloud for visualization, which is beneficial for humans to filter and modify the error pose annotation results. About \\(30\\%\\) of the data is removed or optimized after human correction, and this process ensures the quality of keypoints annotation. More details can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.717, + 0.393, + 0.733 + ], + "angle": 0, + "content": "4 HRRadarPose" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Inspired by HRNet [43], renowned for its adeptness in extracting high-resolution representations, we build an architecture, HRRadarPose, with fully 3D convolutional layers. By treating the Doppler axis, \\( D \\), as the input channels, our model extracts volumetric spatial features along the \\( Z \\), \\( Y \\), and \\( X \\) axes, representing vertical, horizontal, and depth directions in the Cartesian coordinate system. The HRRadarPose architecture is designed to preserve spatial resolution and assimi" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RT-Pose" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.604 + ], + "angle": 0, + "content": "late semantic information from the 4D radar tensors, maintaining high-resolution representations and enabling the exchange of information between features of different resolutions during stage transitions. Figure 6 illustrates an instance of the HRRadarPose structure with three stages. The structure starts with using 3D convolutional layers to transform the 4D radar tensors, \\( T \\in \\mathbb{R}^{D \\times Z \\times Y \\times X} \\), into features, \\( f_{in} \\in \\mathbb{R}^{C_1 \\times Z_1 \\times Y_1 \\times X_1} \\). As the network proceeds from one stage to the next, it develops additional branches through strided convolution to widen the receptive field. Each branch \\( i \\) employs strided convolution to elevate its input feature \\( f_i \\) to a feature \\( f_{i+1} \\) in the succeeding branch \\( i+1 \\), simultaneously doubling the channels and halving the spatial dimensions, \\( f_{i+1} \\in \\mathbb{R}^{2C_i \\times \\frac{Z_i}{2} \\times \\frac{Y_i}{2} \\times \\frac{X_i}{2}} \\). Within each stage, a series of \\( P \\) modules execute \\( M \\) parallel convolutions. To align feature dimensions from different branches for information exchange, we use strided convolution followed by upsampling. At the end of the backbone, we take the feature with the highest resolution to serve as a unified representation for the pose estimation head. In the pose estimation head, two 3D convolutional branches are employed to process these features further. This bifurcation generates two outputs: a distribution map signifying the confidence of locating the human center and the location offset of joint keypoints. The right part of Figure 6 shows the decoding of the HRRadarPose network's output to multi-person poses. The process starts with picking top- \\( k \\) human centers, \\( C_S \\), with confidence scores over a threshold value. We denote these picked centers' indices as \\( S \\). Then, we query the centers corresponding joint keypoints' offsets, \\( K_S \\), by looking up the output map of keypoint offsets head at \\( S \\). Lastly, we decode their joint keypoint locations, \\( J_S \\), by summing \\( K_S \\) and \\( C_S \\). In scenarios with multiple people, we apply range non-maximum suppression (NMS) to reduce overlapping detections. Our method is favorable in two folds: The HRRadarPose's architecture gains efficiency through a single-stage workflow, avoiding complexity brought by region-proposal-based methods [55]. In addition, our pose estimation head inherently learns per person's joint keypoints refraining from ambiguity of association between people's identity and joints." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.788, + 0.639 + ], + "angle": 0, + "content": "To train the center probability head, we use the pixel-wise focal loss as the classification loss, \\( L_{\\text{class}} \\), defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.297, + 0.66, + 0.786, + 0.7 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {c l a s s}} = \\left\\{ \\begin{array}{l l} (1 - P _ {x y z}) ^ {\\alpha} \\log \\left(P _ {x y z}\\right), & \\text {i f} Y _ {x y z} = 1, \\\\ (1 - Y _ {x y z}) ^ {\\beta} \\left(P _ {x y z}\\right) ^ {\\alpha} \\log \\left(1 - P _ {x y z}\\right), & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.843 + ], + "angle": 0, + "content": "where \\( Y_{xyz} \\) is the three-dimensional human center Gaussian distribution map generated by pelvis locations, \\( P_{xyz} \\) is the probability of a person's presence at location, \\( (x,y,z) \\), \\( N \\) is the number of foreground samples, \\( \\alpha \\) and \\( \\beta \\) are hyperparameters addressing class imbalance and focusing training on foreground samples. To train the keypoint offset head, we construct the regression loss \\( L_{reg} \\) by taking the average of 15 keypoints' \\( \\mathcal{L}_1 \\), comparing the distance between predicted offsets and ground-truth offsets to the center location of a voxel in \\( Y \\) when \\( Y_{xyz} \\) equals to 1. The final loss is a weighted sum of \\( L_{class} \\) and \\( L_{reg} \\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "Y. Ho et al." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.154, + 0.775, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.318, + 0.788, + 0.375 + ], + "angle": 0, + "content": "Fig. 6: HRRadarPose's architecture consists of a 3D convolutional backbone and a pose estimation head, which generates a body center's confidence map and joint keypoint offsets map. To infer multi-person poses, we decode the prediction by adding the picked bodies' centers with their corresponding keypoint offsets." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.402, + 0.377, + 0.42 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.784, + 0.51 + ], + "angle": 0, + "content": "15 obvious keypoints of a human body are selected, aligning with the body skeleton model of the Human3.6M dataset [19]. Each frame's 3D pose annotation in the sequence is manually validated to ensure the quality of the dataset. Figure 7 demonstrates the testing results across various levels of action complexity, presenting the feasibility of 3D pose estimation using the proposed baseline." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.785, + 0.614 + ], + "angle": 0, + "content": "Evaluation metrics We employ joint position error (JPE), which measures the mean of the Euclidean distance for every joint's predicted keypoints and the ground truth keypoints. Localization results are reported by the mean of the root position error (MRPE) [42], which is the JPE of the pelvis keypoint. For the evaluation of 3D HPE, we use Mean per joint position error (MPJPE) [57] and absolute-MPJPE (Abs-MPJPE) [40]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Baselines Comparison We evaluate the performance of the HRRadarPose in global localization and 3D HPE on the RT-Pose dataset. We compare our method with three baseline methods, as shown in Table 2. mm-Pose [37] and mmMesh [48], utilizing radar point clouds for 3D HPE, are re-implemented to test on the RT-Pose dataset. Notably, mmMesh introduces a global localization model, combined with an RNN HPE module, showing more accurate estimation results in both MRPE and MPJPE compared to mm-Pose, which employs a simpler CNN model. RF-Pose 3D [55] uses the 4D radar tensor with a region proposal network to locate the subjects and estimate the 3D pose. The experimental results show that our HRRadarPose outperforms other baseline methods. On the RT-Pose dataset, we obtain \\(9.93\\mathrm{cm}\\) in MPJPE and \\(9.91\\mathrm{cm}\\) in MRPE. We find that radar-tensor-based methods reach less error than the methods using radar point cloud. We attribute this performance gap to complicated scenarios since it is difficult for point cloud-based methods to discriminate different key" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RT-Pose" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image_caption", + "bbox": [ + 0.249, + 0.172, + 0.266, + 0.198 + ], + "angle": 0, + "content": "R" + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.15, + 0.346, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.15, + 0.427, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.431, + 0.15, + 0.506, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.15, + 0.587, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.15, + 0.664, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.15, + 0.744, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.219, + 0.267, + 0.269 + ], + "angle": 0, + "content": "BREV TraJ." + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.217, + 0.346, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.217, + 0.428, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.431, + 0.217, + 0.506, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.217, + 0.587, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.217, + 0.666, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.217, + 0.744, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.292, + 0.267, + 0.331 + ], + "angle": 0, + "content": "3D pose" + }, + { + "type": "image", + "bbox": [ + 0.271, + 0.279, + 0.352, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.272, + 0.334, + 0.344, + 0.344 + ], + "angle": 0, + "content": "(a) Waving hand" + }, + { + "type": "image", + "bbox": [ + 0.354, + 0.279, + 0.43, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.36, + 0.334, + 0.421, + 0.344 + ], + "angle": 0, + "content": "(b) Lifting leg" + }, + { + "type": "image", + "bbox": [ + 0.431, + 0.279, + 0.514, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.435, + 0.334, + 0.508, + 0.344 + ], + "angle": 0, + "content": "(c) Random pose" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.279, + 0.592, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.334, + 0.575, + 0.344 + ], + "angle": 0, + "content": "(d) Normal" + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.279, + 0.669, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.603, + 0.334, + 0.648, + 0.344 + ], + "angle": 0, + "content": "(e) Sitting" + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.279, + 0.755, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.673, + 0.334, + 0.742, + 0.344 + ], + "angle": 0, + "content": "(f) Waving hand" + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.348, + 0.753, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.381, + 0.785, + 0.436 + ], + "angle": 0, + "content": "Fig. 7: Visualization of pose estimation results in test frames for six activities. The second row is the trajectory of the test sequence in BEV. The third row is the instance of the global 3D pose estimated result for each action. In both rows, the blue line is the ground truth, and the red line represents estimated results." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.451, + 0.785, + 0.492 + ], + "angle": 0, + "content": "Table 2: Comparison of human localization and 3D pose estimation results of our HRRadarPose method and three baseline methods. † represents the radar point cloud-based methods." + }, + { + "type": "table", + "bbox": [ + 0.304, + 0.506, + 0.695, + 0.587 + ], + "angle": 0, + "content": "
MethodMRPE (cm)MPJPE (cm)Abs-MPJPE (cm)
mm-Pose † [37]102.2821.26102.7
mmMesh † [48]62.6913.7866.24
RF-Pose3D [55]29.1718.4338.05
HRRadarPose (Ours)9.919.9314.73
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.785, + 0.675 + ], + "angle": 0, + "content": "points in various poses. The results highlight the importance of using 4D radar tensors, which preserve raw and rich spatial-temporal information for the HPE tasks. An end-to-end architecture directly takes in 4D radar tensors to better face different human action scenarios in the real world." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.785, + 0.809 + ], + "angle": 0, + "content": "Action complexity Analysis RT-posedataset provides six difficulty levels of actions to simulate real-world scenarios, enabling comprehensive evaluation of localization and 3D HPE results. We report the performance of the proposed HRRadarPose method in Table 3. Generally, the joints along the human longitudinal axis, such as the thorax and the head, are more stable than those in the limbs. This phenomenon may be attributed to the larger reflective surface of the body's trunk compared to the limbs, providing a more reliable HPE based on radar signals." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "The MRPE of standing posture is below \\(7.67~\\mathrm{cm}\\), indicating that the pose complexity slightly influences the localization accuracy. Due to the chair's strong" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "Y. Ho et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 3: Comparison of the performance predicted by the HRRadar Pose in different complexity level activities, presenting different key points of the pose, the results of localization, and total 3D pose estimation results. The first group of rows lists the stand-related poses, while the other lists the walking-related poses. The complexity level increases from the top activity to the bottom within each group." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.227, + 0.784, + 0.349 + ], + "angle": 0, + "content": "
ActionsJPE(cm)MRPE (cm)MPJPE (cm)Abs-MPJPE (cm)
ThoraxHeadAnkleWrist
Stand: Lifting LegWaving Hand2.953.973.6125.653.356.917.97
6.127.9212.248.887.557.0610.37
Random Pose7.7811.2111.7532.127.6712.4715.12
Walk: SitingNormal4.316.6913.5215.9813.678.9316.92
9.3613.0313.9310.7714.119.4416.83
Waving Hand5.7410.3713.5343.1612.8215.2821.22
" + }, + { + "type": "image", + "bbox": [ + 0.281, + 0.379, + 0.727, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.613 + ], + "angle": 0, + "content": "Fig. 8: HRRadarPose Results in different conditions. (a), (b), and (c) are occluded by cardboard, cloth, and plastic pad, respectively. (d) is in low-lighting conditions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.657, + 0.789, + 0.838 + ], + "angle": 0, + "content": "radar reflection, distinguishing between human subjects and the chair poses a challenge for localization. The MPJPE of walking and sitting is \\(9.44\\mathrm{cm}\\), showing our HRRadarPose model can reliably estimate the sitting pose. However, the MRPE of this action sequence achieves \\(14.11\\mathrm{cm}\\), demonstrating that sitting is a key factor for evaluating the performance of localization. The MPJPE of simple actions, such as standing and waving, standing and lifting legs, are around 7 cm, and the walking is \\(8.93\\mathrm{cm}\\), demonstrating that our HRRadarPose model is comparable to other recently published radar pose estimate methods [26, 47]. However, the current model still encounters difficulty in estimating HPE in complex actions such as walking and waving. Therefore, the RT-pose dataset is a challenging benchmark, which can enable HPE models to realistically learn diverse pose scenarios." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RT-Pose" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.493, + 0.173 + ], + "angle": 0, + "content": "Table 4: HPE results from HRRadarPose in normal and occlusive scenarios." + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.186, + 0.475, + 0.236 + ], + "angle": 0, + "content": "
OcclusionMPJPE (cm)Abs-MPJPE (cm)
✓9.8414.52
11.5718.55
" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.145, + 0.785, + 0.172 + ], + "angle": 0, + "content": "Table 5: Ablation Study of Using Dopper information." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.185, + 0.766, + 0.236 + ], + "angle": 0, + "content": "
Doppler|MPJPE (cm) Abs-MPJPE (cm)
11.3122.34
✓9.9114.73
" + }, + { + "type": "table_caption", + "bbox": [ + 0.293, + 0.25, + 0.709, + 0.264 + ], + "angle": 0, + "content": "Table 6: Our Radar-Doppler Convolutional Block in HR-Net." + }, + { + "type": "table", + "bbox": [ + 0.317, + 0.277, + 0.681, + 0.326 + ], + "angle": 0, + "content": "
HRNet-Block-DesignMPJPE (cm)Abs-MPJPE (cm)
Traditional Block (RGB-Vision)10.3517.59
Ours (Radar-Doppler)9.9314.73
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.355, + 0.788, + 0.522 + ], + "angle": 0, + "content": "Qualitative Result Figure 8 demonstrates the qualitative results of different occlusion conditions and a low-lighting instance. Due to the hardware setting of our system, supporting synchronized two camera data, the subject's pose can be visualized by another camera if one of them is occluded. Therefore, the 3D annotation results can be generated even on occlusion conditions. Our method provides reliable pose estimation results even when the radar module is blocked or with low-lighting conditions. Table 4 shows the HPE results of occlusion in normal outdoor conditions. Although the radar module is occluded, the mmWave signal is still robust enough as input for the HRRadarPose model to estimate HPE, demonstrating the characteristics of radar and the importance of using radar for HPE." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.63 + ], + "angle": 0, + "content": "Ablation Study To illustrate the advantage of modeling motion features using Doppler information by our HRRadarPose, we remove Doppler information by averaging 4D radar tensors along the Doppler axis and train the models with and without Doppler information using the same setup. Table 5 shows that with the inclusion of Doppler information, our HRRadarPose enjoys a performance gain due to its capability to extract motion features, enhancing keypoint localization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Compared with RPM 2.0, our input data and backbone design are different. The proposed HRRadarPose network extracts better feature representation from 4D radar tensors through modified convolutional blocks. As shown in Figure 9, we replace the batch norm with the group norm of convolutional blocks in HRNet. Since we treat the channel-wise features as a Doppler axis, the group norm inherently handles the feature in different speed groups, which normalizes the data effectively. Table 6 shows the results, proving that the modified CNN block is more suitable for Doppler radar than the original HRNet. Compared with our method, most previous work did not leverage Doppler information [47,50]. We validate that motion features [17,27] are crucial for aiding deep learning models in capturing more human body characteristics over the environment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To ablate our design choice of HPE heads, we use the same backbone, replace our pose estimation head with RF-Pose3D's [55] head, and train both models for" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "Y. Ho et al." + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.145, + 0.449, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.215, + 0.495, + 0.271 + ], + "angle": 0, + "content": "Fig.9: Illustration of two variants of CNN blocks in HRNet. (a) Ours (Radar-Doppler). (b) Traditional Block (RGBVision)." + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.173, + 0.785, + 0.2 + ], + "angle": 0, + "content": "Table 7: Comparison of common 4D radar tensor-based HPE head and ours." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.213, + 0.767, + 0.257 + ], + "angle": 0, + "content": "
HPE HeadMPJPE (cm)Abs-MPJPE (cm)
RF-Pose3D [55]15.5218.08
Ours10.2417.63
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.785, + 0.404 + ], + "angle": 0, + "content": "the same epochs. RF-Pose3D formulates the HPE as a multi-keypoints classification problem, where their model outputs confidence distributions of every human body joint. As shown in Table 7, our design is favorable in joint keypoints localization in the metric space, indicated by a lower Abs-MPJPE. Moreover, lower MPJPE from our HPE head's result validates stronger inter-joint relationships modeling capability." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.425, + 0.528, + 0.44 + ], + "angle": 0, + "content": "6 Limitation and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.788, + 0.592 + ], + "angle": 0, + "content": "First, to leverage the input data types for training the 3D HPE model, the computational resource is a crucial part. 4D radar tensor in Cartesian coordinate consumes up to \\(100\\mathrm{MB}\\) per frame, which decreases the training speed. Second, we limit the experimental data collection scope from \\(2\\mathrm{m}\\) to \\(8\\mathrm{m}\\) because the camera is hard to clearly capture 2D pose skeletons at a large distance and LiDAR can not easily collect completed point clouds in near vision due to the constraints of FoV. Third, the proposed baseline HRRadarPose lacks robustness in accurately estimating poses during complex activities. Capturing human pose accurately on 4D radar tensor remains an unresolved challenge." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.613, + 0.36, + 0.629 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we propose RT-Pose, the first dataset for human pose estimation (HPE) with synchronized and calibrated 4D radar tensors, LiDAR point clouds, and RGB images. RT-Pose provides cluttered scenarios and human activities of different complexity levels to enhance the difficulty of pose estimation for practical applications. In addition, having various modalities is beneficial for our semi-automatic 3D pose annotation optimization process. The proposed HRRadarPose is the first single-stage architecture designed for estimating human pose using 4D radar tensors as the input. The results indicate that the proposed HRRadarPose outperforms previous works, in which datasets are collected with simpler actions and cleaner scenarios. Our work contributes to offering a three-modality dataset with 4D radar tensors for HPE, which is promising to be applied for complex actions and different scenes. We hope this work encourages future development of 4D radar-based HPE methods." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RT-Pose" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.788, + 0.273 + ], + "angle": 0, + "content": "We thank Yizhou Wang \\(^{3}\\) for helping us build the data collection system and the radar data processing pipeline. We thank Hou-I Liu \\(^{4}\\), Cheng-Yi Huang \\(^{5}\\), Yi Shiang Chen \\(^{6}\\), Chi-Yuan Chan \\(^{7}\\), and Yu-Hsien Lu \\(^{8}\\) for helping us collect the RT-Pose dataset. We thank the National Science and Technology Counci of the Republic of China, Taiwan, under Projects of NSTC 112-2917-I-006-007 for its funding support." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.299, + 0.323, + 0.315 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.334, + 0.787, + 0.363 + ], + "angle": 0, + "content": "1. Adib, F., Hsu, C.Y., Mao, H., Katabi, D., Durand, F.: Rf-capture: Capturing the human figure through a wall. ACM SIGGRAPH Asia (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.364, + 0.787, + 0.419 + ], + "angle": 0, + "content": "2. Ahuja, K., Jiang, Y., Goel, M., Harrison, C.: Vid2doppler: Synthesizing doppler radar data from videos for training privacy-preserving activity recognition. In: Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. pp. 1-10 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.421, + 0.787, + 0.476 + ], + "angle": 0, + "content": "3. An, S., Li, Y., Ogras, U.: mri: Multi-modal 3d human pose estimation dataset using mmwave, rgb-d, and inertial sensors. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems. vol. 35, pp. 27414-27426. Curran Associates, Inc. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.477, + 0.787, + 0.518 + ], + "angle": 0, + "content": "4. An, S., Ogras, U.Y.: Mars: mmwave-based assistive rehabilitation system for smart healthcare. ACM Transactions on Embedded Computing Systems (TECS) 20(5s), 1-22 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.519, + 0.787, + 0.56 + ], + "angle": 0, + "content": "5. Chai, W., Guo, X., Wang, G., Lu, Y.: Stablevideo: Text-driven consistency-aware diffusion video editing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 23040-23050 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.561, + 0.787, + 0.603 + ], + "angle": 0, + "content": "6. Chai, W., Jiang, Z., Hwang, J.N., Wang, G.: Global adaptation meets local generalization: Unsupervised domain adaptation for 3d human pose estimation. arXiv preprint arXiv:2303.16456 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.604, + 0.787, + 0.645 + ], + "angle": 0, + "content": "7. Chen, A., Wang, X., Zhu, S., Li, Y., Chen, J., Ye, Q.: mmbody benchmark: 3d body reconstruction dataset and analysis for millimeter wave radar. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 3501-3510 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.646, + 0.787, + 0.701 + ], + "angle": 0, + "content": "8. Chen, K., Gabriel, P., Alasfour, A., Gong, C., Doyle, W.K., Devinsky, O., Friedman, D., Dugan, P., Melloni, L., Thesen, T., et al.: Patient-specific pose estimation in clinical environments. IEEE journal of translational engineering in health and medicine 6, 1-11 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.787, + 0.744 + ], + "angle": 0, + "content": "9. Cheng, J.H., Chen, Y., Chang, T.Y., Lin, H.E., Wang, P.Y.C., Cheng, L.P.: Impossible staircase: Vertically real walking in an infinite virtual tower. In: 2021 IEEE Virtual Reality and 3D User Interfaces (VR). pp. 50-56 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.334, + 0.787, + 0.744 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.755, + 0.394, + 0.771 + ], + "angle": 0, + "content": "3 http://yizhouwang.net/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.222, + 0.771, + 0.384, + 0.784 + ], + "angle": 0, + "content": "k39967.c@nycu.edu.tw" + }, + { + "type": "page_footnote", + "bbox": [ + 0.222, + 0.785, + 0.414, + 0.798 + ], + "angle": 0, + "content": "5 n28121521@gs.ncku.edu.tw" + }, + { + "type": "page_footnote", + "bbox": [ + 0.222, + 0.799, + 0.415, + 0.812 + ], + "angle": 0, + "content": "\\(^{6}\\) m16121051@gs.ncku.edu.tw" + }, + { + "type": "page_footnote", + "bbox": [ + 0.222, + 0.812, + 0.414, + 0.826 + ], + "angle": 0, + "content": "7 n26120587@gs.ncku.edu.tw" + }, + { + "type": "page_footnote", + "bbox": [ + 0.222, + 0.826, + 0.414, + 0.84 + ], + "angle": 0, + "content": "n26120587@gs.ncku.edu.tw" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.755, + 0.415, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "Y. Ho et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "10. Cheng, J.H., Kuan, S.Y., Liu, H.I., Latapie, H., Liu, G., Hwang, J.N.: Centerradar-net: Joint 3d object detection and tracking framework using 4d fmcw radar. In: 2024 IEEE International Conference on Image Processing (ICIP) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "11. Dai, Y., Lin, Y., Lin, X., Wen, C., Xu, L., Yi, H., Shen, S., Ma, Y., Wang, C.: Sloper4d: A scene-aware dataset for global 4d human pose estimation in urban environments. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 682-692 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.785, + 0.3 + ], + "angle": 0, + "content": "12. Dai, Y., Lin, Y., Wen, C., Shen, S., Xu, L., Yu, J., Ma, Y., Wang, C.: Hsc4d: Human-centered 4d scene capture in large-scale indoor-outdoor space using wearable imus and lidar. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6792-6802 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.301, + 0.785, + 0.342 + ], + "angle": 0, + "content": "13. De Maio, A., Farina, A., Foglia, G.: Design and experimental validation of knowledge-based constant false alarm rate detectors. IET Radar, Sonar & Navigation 1(4), 308-316 (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.383 + ], + "angle": 0, + "content": "14. Endo, K., Ishikawa, T., Yamamoto, K., Ohtsuki, T.: Multi-person position estimation based on correlation between received signals using mimo fmcw radar. IEEE Access 11, 2610-2620 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.384, + 0.785, + 0.438 + ], + "angle": 0, + "content": "15. Guzov, V., Mir, A., Sattler, T., Pons-Moll, G.: Human poseitioning system (hps): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4318-4329 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "16. Heath, R.W., Gonzalez-Prelcic, N., Rangan, S., Roh, W., Sayeed, A.M.: An overview of signal processing techniques for millimeter wave mimo systems. IEEE journal of selected topics in signal processing 10(3), 436-453 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.481, + 0.785, + 0.536 + ], + "angle": 0, + "content": "17. Huang, H.W., Yang, C.Y., Ramkumar, S., Huang, C.I., Hwang, J.N., Kim, P.K., Lee, K., Kim, K.: Observation centric and central distance recovery for athlete tracking. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 454-460 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.537, + 0.785, + 0.591 + ], + "angle": 0, + "content": "18. Huang, H.Y., Ning, C.W., Wang, P.Y., Cheng, J.H., Cheng, L.P.: Haptic-go-round: A surrounding platform for encounter-type haptics in virtual reality experiences. In: Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems. CHI '20, Association for Computing Machinery (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.592, + 0.785, + 0.646 + ], + "angle": 0, + "content": "19. Ionescu, C., Papava, D., Olaru, V., Sminchisescu, C.: Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence 36(7), 1325-1339 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "20. Jiang, Z., Chai, W., Li, L., Zhou, Z., Yang, C.Y., Hwang, J.N.: Unihpe: Towards unified human pose estimation via contrastive learning. arXiv preprint arXiv:2311.16477 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.689, + 0.785, + 0.743 + ], + "angle": 0, + "content": "21. Jiang, Z., Ji, H., Yang, C.Y., Hwang, J.N.: 2d human pose estimation calibration and keypoint visibility classification. In: ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 6095-6099. IEEE (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.744, + 0.785, + 0.799 + ], + "angle": 0, + "content": "22. Jiang, Z., Zhou, Z., Li, L., Chai, W., Yang, C.Y., Hwang, J.N.: Back to optimization: Diffusion-based zero-shot 3d human pose estimation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6142-6152 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "23. Kim, Y., Alnujaim, I., Oh, D.: Human activity classification based on point clouds measured by millimeter wave mimo radar with deep recurrent neural networks. IEEE Sensors Journal 21(12), 13522-13529 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RT-Pose" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "24. Klauder, J.R., Price, A., Darlington, S., Albersheim, W.J.: The theory and design of chirp radars. Bell System Technical Journal 39(4), 745-808 (1960)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.787, + 0.217 + ], + "angle": 0, + "content": "25. Kuan, S.Y., Cheng, J.H., Huang, H.W., Chai, W., Yang, C.Y., Wu, B.F., Hwang, J.N.: Boosting online 3d multi-object tracking through camera-radar cross check. In: IEEE Intelligent Vehicles Symposium (IV) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.217, + 0.787, + 0.27 + ], + "angle": 0, + "content": "26. Lee, S.P., Kini, N.P., Peng, W.H., Ma, C.W., Hwang, J.N.: Hupr: A benchmark for human pose estimation using millimeter wave radar. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5715-5724 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.27, + 0.787, + 0.325 + ], + "angle": 0, + "content": "27. Lin, C.L., Chang, Y.M., Hung, C.C., Tu, C.D., Chuang, C.Y.: Position estimation and smooth tracking with a fuzzy-logic-based adaptive strong tracking kalman filter for capacitive touch panels. IEEE Transactions on Industrial Electronics 62(8), 5097-5108 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.325, + 0.787, + 0.379 + ], + "angle": 0, + "content": "28. Lin, C.L., Chiu, W.C., Chu, T.C., Ho, Y.H., Chen, F.H., Hsu, C.C., Hsieh, P.H., Chen, C.H., Lin, C.C.K., Sung, P.S., et al.: Innovative head-mounted system based on inertial sensors and magnetometer for detecting falling movements. Sensors 20(20), 5774 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.379, + 0.787, + 0.42 + ], + "angle": 0, + "content": "29. Lin, C.L., Ho, Y.H., Chiu, W.C., Chu, T.C., Liu, Y.H.: Innovative shoe-integrated system based on time-of-flight range sensors for fall detection on various terrains. IEEE Sensors Letters 5(10), 1-4 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.42, + 0.787, + 0.461 + ], + "angle": 0, + "content": "30. Lin, H.Y., Chen, T.W.: Augmented reality with human body interaction based on monocular 3d pose estimation. In: International Conference on Advanced Concepts for Intelligent Vision Systems. pp. 321-331. Springer (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.461, + 0.787, + 0.515 + ], + "angle": 0, + "content": "31. Liu, H., He, J.Y., Cheng, Z.Q., Xiang, W., Yang, Q., Chai, W., Wang, G., Bao, X., Luo, B., Geng, Y., et al.: Posynda: Multi-hypothesis pose synthesis domain adaptation for robust 3d human pose estimation. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 5542-5551 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.515, + 0.787, + 0.556 + ], + "angle": 0, + "content": "32. Mehta, D., Rhodin, H., Casas, D., Fua, P., Sotnychenko, O., Xu, W., Theobalt, C.: Monocular 3d human pose estimation in the wild using improved cnn supervision. In: 2017 international conference on 3D vision (3DV). pp. 506-516. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.556, + 0.787, + 0.597 + ], + "angle": 0, + "content": "33. Neemat, S., Uysal, F., Krasnov, O., Yarovoy, A.: Reconfigurable range-doppler processing and range resolution improvement for fmcw radar. IEEE Sensors Journal 19(20), 9294-9303 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.597, + 0.787, + 0.637 + ], + "angle": 0, + "content": "34. Paek, D.H., Kong, S.H., Wijaya, K.T.: K-radar: 4d radar object detection for autonomous driving in various weather conditions. Advances in Neural Information Processing Systems 35, 3819-3829 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.637, + 0.787, + 0.678 + ], + "angle": 0, + "content": "35. Sengupta, A., Cao, S.: mmpose-nlp: A natural language processing approach to precise skeletal pose estimation using mmwave radars. IEEE Transactions on Neural Networks and Learning Systems (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.678, + 0.787, + 0.718 + ], + "angle": 0, + "content": "36. Sengupta, A., Jin, F., Cao, S.: Nlp based skeletal pose estimation using mmwave radar point-cloud: A simulation approach. In: 2020 IEEE Radar Conference (RadarConf20). pp. 1-6. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.718, + 0.787, + 0.759 + ], + "angle": 0, + "content": "37. Sengupta, A., Jin, F., Zhang, R., Cao, S.: mm-pose: Real-time human skeletal posture estimation using mmwave radars and cnns. IEEE Sensors Journal 20(17), 10032-10044 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.759, + 0.787, + 0.8 + ], + "angle": 0, + "content": "38. Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learning for human pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5693-5703 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "39. Sun, Z., Ke, Q., Rahmani, H., Bennamoun, M., Wang, G., Liu, J.: Human action recognition from various data modalities: A review. IEEE transactions on pattern analysis and machine intelligence (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "Y. Ho et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "40. Véges, M., Lorincz, A.: Absolute human pose estimation with depth prediction network. In: 2019 International Joint Conference on Neural Networks (IJCNN). pp. 1-7. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.243 + ], + "angle": 0, + "content": "41. Von Marcard, T., Henschel, R., Black, M.J., Rosenhahn, B., Pons-Moll, G.: Recovering accurate 3d human pose in the wild using imus and a moving camera. In: Proceedings of the European conference on computer vision (ECCV). pp. 601-617 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.243, + 0.785, + 0.284 + ], + "angle": 0, + "content": "42. Wang, J., Tan, S., Zhen, X., Xu, S., Zheng, F., He, Z., Shao, L.: Deep 3d human pose estimation: A review. Computer Vision and Image Understanding 210, 103225 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.785, + 0.324 + ], + "angle": 0, + "content": "43. Wang, J., Sun, K., Cheng, T., Jiang, B., Deng, C., Zhao, Y., Liu, D., Mu, Y., Tan, M., Wang, X., Liu, W., Xiao, B.: Deep high-resolution representation learning for visual recognition. TPAMI (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.378 + ], + "angle": 0, + "content": "44. Wang, J., Sun, K., Cheng, T., Jiang, B., Deng, C., Zhao, Y., Liu, D., Mu, Y., Tan, M., Wang, X., et al.: Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence 43(10), 3349-3364 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.433 + ], + "angle": 0, + "content": "45. Wang, Y., Cheng, J.H., Huang, J.T., Kuan, S.Y., Fu, Q., Ni, C., Hao, S., Wang, G., Xing, G., Liu, H., Hwang, J.N.: Vision meets mmwave radar: 3d object perception benchmark for autonomous driving. In: IEEE Intelligent Vehicles Symposium (IV) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.785, + 0.488 + ], + "angle": 0, + "content": "46. Wang, Y., Jiang, Z., Li, Y., Hwang, J.N., Xing, G., Liu, H.: Rodnet: A real-time radar object detection network cross-supervised by camera-radar fused object 3d localization. IEEE Journal of Selected Topics in Signal Processing 15(4), 954-967 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.528 + ], + "angle": 0, + "content": "47. Xie, C., Zhang, D., Wu, Z., Yu, C., Hu, Y., Chen, Y.: RPM 2.0: Rf-based pose machines for multi-person 3d pose estimation. IEEE Transactions on Circuits and Systems for Video Technology (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.529, + 0.785, + 0.583 + ], + "angle": 0, + "content": "48. Xue, H., Ju, Y., Miao, C., Wang, Y., Wang, S., Zhang, A., Su, L.: mmmesh: Towards 3d real-time dynamic human mesh construction using millimeter-wave. In: Proceedings of the 19th Annual International Conference on Mobile Systems, Applications, and Services. pp. 269-282 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.785, + 0.637 + ], + "angle": 0, + "content": "49. Yang, C.Y., Luo, J., Xia, L., Sun, Y., Qiao, N., Zhang, K., Jiang, Z., Hwang, J.N., Kuo, C.H.: Camerapose: Weakly-supervised monocular 3d human pose estimation by leveraging in-the-wild 2d annotations. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2924-2933 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.664 + ], + "angle": 0, + "content": "50. Yu, C., Zhang, D., Wu, Z., Xie, C., Lu, Z., Hu, Y., Chen, Y.: Mobirfpose: Portable rf-based 3d human pose camera. IEEE Transactions on Multimedia (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.785, + 0.704 + ], + "angle": 0, + "content": "51. Yuan, Y., Wei, S.E., Simon, T., Kitani, K., Saragih, J.: Simpoe: Simulated character control for 3d human pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7159-7169 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.785, + 0.745 + ], + "angle": 0, + "content": "52. Zhang, G., Geng, X., Lin, Y.J.: Comprehensive mpoint: A method for 3d point cloud generation of human bodies utilizing fmcw mimo mm-wave radar. Sensors 21(19), 6455 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.785, + 0.785 + ], + "angle": 0, + "content": "53. Zhang, Z., Chai, W., Jiang, Z., Ye, T., Song, M., Hwang, J.N., Wang, G.: Mpm: A unified 2d-3d human pose representation via masked pose modeling. arXiv preprint arXiv:2306.17201 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "54. Zhao, M., Li, T., Abu Alsheikh, M., Tian, Y., Zhao, H., Torralba, A., Katabi, D.: Through-wall human pose estimation using radio signals. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7356-7365 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RT-Pose" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "55. Zhao, M., Tian, Y., Zhao, H., Alsheikh, M.A., Li, T., Hristov, R., Kabelac, Z., Katabi, D., Torralba, A.: Rf-based 3d skeletons. In: Proceedings of the 2018 Conference of the ACM Special Interest Group on Data Communication. pp. 267-281 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "56. Zhao, Y., Yarovoy, A., Fioranelli, F.: Angle-insensitive human motion and posture recognition based on 4d imaging radar and deep learning classifiers. IEEE Sensors Journal 22(12), 12173-12182 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.215, + 0.246, + 0.787, + 0.286 + ], + "angle": 0, + "content": "57. Zheng, C., Wu, W., Chen, C., Yang, T., Zhu, S., Shen, J., Kehtarnavaz, N., Shah, M.: Deep learning-based human pose estimation: A survey. ACM Computing Surveys 56(1), 1-37 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.215, + 0.287, + 0.787, + 0.328 + ], + "angle": 0, + "content": "58. Zheng, Z., Pan, J., Ni, Z., Shi, C., Ye, S., Fang, G.: Human posture reconstruction for through-the-wall radar imaging using convolutional neural networks. IEEE Geoscience and Remote Sensing Letters 19, 1-5 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.215, + 0.328, + 0.787, + 0.384 + ], + "angle": 0, + "content": "59. Zhou, Z., Jiang, Z., Chai, W., Yang, C.Y., Li, L., Hwang, J.N.: Efficient domain adaptation via generative prior for 3d infant pose estimation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 41-49 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.384 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/RT-Pose_ A 4D Radar-Tensor based 3D Human Pose Estimation and Localization Benchmark/e4fd786f-06ad-4f57-8db1-c3bab37f210b_origin.pdf b/2024/RT-Pose_ A 4D Radar-Tensor based 3D Human Pose Estimation and Localization Benchmark/e4fd786f-06ad-4f57-8db1-c3bab37f210b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..843146c99768c5e984d5f888dcf700a1c767f013 --- /dev/null +++ b/2024/RT-Pose_ A 4D Radar-Tensor based 3D Human Pose Estimation and Localization Benchmark/e4fd786f-06ad-4f57-8db1-c3bab37f210b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e52811e824aa9f6d72b4241cce0da7458c1fe166ca58f9746e7adf30db686f67 +size 3054718 diff --git a/2024/RT-Pose_ A 4D Radar-Tensor based 3D Human Pose Estimation and Localization Benchmark/full.md b/2024/RT-Pose_ A 4D Radar-Tensor based 3D Human Pose Estimation and Localization Benchmark/full.md new file mode 100644 index 0000000000000000000000000000000000000000..558d856db8e11b5ac868abed9a39a05bfee04929 --- /dev/null +++ b/2024/RT-Pose_ A 4D Radar-Tensor based 3D Human Pose Estimation and Localization Benchmark/full.md @@ -0,0 +1,318 @@ +# RT-Pose: A 4D Radar Tensor-based 3D Human Pose Estimation and Localization Benchmark + +Yuan-Hao $\mathrm{Ho}^{1*}$ , Jen-Hao Cheng $^{2*}$ , Sheng Yao Kuan $^{2}$ , Zhongyu Jiang $^{2}$ , + +Wenhao Chai² Hsiang-Wei Huang², Chih-Lung Lin¹, and + +Jenq-Neng Hwang2 + +* Indicates equal contribution + +$^{1}$ National Cheng Kung University + +2 University of Washington + +n28081527@gs.ncku.edu.tw1, + +{andyhci, shengyao, zyjiang, wchai, hwhuang, hwang}@uw.edu², + +cllin@ee.ncku.edu.tw + +Abstract. Traditional methods for human localization and pose estimation (HPE), which mainly rely on RGB images as an input modality, confront substantial limitations in real-world applications due to privacy concerns. In contrast, radar-based HPE methods emerge as a promising alternative, characterized by distinctive attributes such as through-wall recognition and privacy-preserving, rendering the method more conducive to practical deployments. This paper presents a Radar Tensor-based human pose (RT-Pose) dataset and an open-source benchmarking framework. RT-Pose dataset comprises 4D radar tensors, Li-DAR point clouds, and RGB images, and is collected for a total of $72\mathrm{k}$ frames across 240 sequences with six different complexity level actions. The 4D radar tensor provides raw spatio-temporal information, differentiating it from other radar point cloud-based datasets. We develop an annotation process, which uses RGB images and LiDAR point clouds to accurately label 3D human skeletons. In addition, we propose HRRadarPose, the first single-stage architecture that extracts the high-resolution representation of 4D radar tensors in 3D space to aid human keypoint estimation. HRRadarPose outperforms previous radar-based HPE work on the RT-Pose benchmark. The overall HRRadarPose performance on the RT-Pose dataset, as reflected in a mean per joint position error (MPJPE) of $9.91cm$ , indicates the persistent challenges in achieving accurate HPE in complex real-world scenarios. RT-Pose is available at https://huggingface.co/datasets/uwipl/RT-Pose. + +Keywords: Human Pose Estimation $\cdot$ 4D Radar $\cdot$ Dataset $\cdot$ Benchmark + +# 1 Introduction + +Human localization and 3D pose estimation [6, 20-22, 31, 49, 53, 59] are indispensable in Augmented/Virtual Reality [5, 9, 15, 30], human-computer interaction [18, 51], healthcare [3, 8, 29]. However, capturing human poses across diverse + +scenarios and generating comprehensive 3D representations for individuals in varying poses present significant challenges. Camera based motion capture systems are often used to collect human body posture [19,28,42]. However, these systems are sensitive to light and require careful multi-camera calibration, making them unreliable for outdoor scenarios and unable to provide overall pose estimation [7]. Furthermore, due to privacy concerns, deploying its application in home or long-term care center scenarios poses challenges. + +Radar-based method emerges as a promising alternative, characterized by distinctive attributes such as through-wall recognition [54,58]. In addition, radar is robust to lighting conditions and resilient to various weather conditions and occlusion [1,47]. These characteristics make it especially suitable for safety- and privacy-critical applications [2,39]. In smart automotive applications, radar, as a complementary sensing modality, can complement adverse scenarios such as lowlighting environments and adverse weather [25], challenging RGB-based sensing. In healthcare applications, RGB-based Human Pose Estimation (HPE)poses privacy risks and is hindered by occlusion, prompting a demand for radar in indoor care environments. Considering the advantages and prospects of radar technology, it is indispensable to provide a dataset that contains diverse scenarios for fostering radar-based HPE research. + +The previous radar-based HPE methods [14, 37, 48] often utilize Constant False Alarm Rate (CFAR) techniques [13, 45, 46] to extract radar point clouds. However, the effectiveness of CFAR-based point cloud extraction can be easily influenced by variations in radar module types and hardware parameters. Furthermore, different human body reflections in radar signals across diverse environments may require specific adjustments to CFAR parameters, making it challenging to generalize their applications. To address this issue, some works decompose the 4D radar tensor into vertical and horizontal directions as the input of the model to estimate human pose [26, 47, 50]. This approach computes the magnitude of radar signals for both directions, which effectively reduces data loss during the conversion to point cloud. However, a 4D radar tensor with Doppler values contains the velocity information, which is more informative than a decomposed radar tensor [10, 34]. Therefore, this work focuses on exploring the use of raw 4D radar tensors for HPE and establishing a benchmark based on this data format. + +The presented radar tensor-based human pose (RT-Pose) dataset is the first benchmark to integrate calibrated 4D radar tensors, RGB images, and LiDAR point clouds data in the field of HPE. Releasing a multi-modality dataset marks a significant advancement, particularly in the domain of human behavior analyses. There are a total of $72\mathrm{k}$ frames in 240 sequences in the dataset. The actions are organized into sequences of increasing complexity, providing a realistic motion distribution. All experiments are conducted in 5 environmental conditions in 8 scenarios, which ensures the variety of scenarios for better comprehension of the model. Along with the release of this comprehensive RT-Pose dataset, we also build upon the High-Resolution Network (HRNet) [44] to propose HRRadarPose model, a robust baseline for 3D HPE utilizing 4D radar tensor data. Further + +Table 1: Comparison of 3D human pose estimation datasets. The first group of rows shows RGB and LiDAR datasets. The second group of rows presents datasets with radar. The radar point cloud is denoted as RPC, and the 4D radar tensor is denoted as 4D-RT. * indicates a modified 4D radar tensor with a specific velocity range. + +
DatasetRGB DepthLiDARRPC4D-RT#Scenes#Actions#Seqs#FPSTime (min)
Human3.6M [19]✓-15839501200
HSC4D [12]✓✓320-2042
SLOPER4D [11]✓✓10-152083
MARS [4]✓110801080
mRI [3]✓✓✓11230010264
HuPR [26]✓✓*-323510240
RT-Pose (Ours)✓✓✓40624010120
+ +more, our study also includes a comparative evaluation of different radar data types within our proposed model, as well as benchmarking against previous radar-based methods. These results demonstrate the advantages of employing the 4D radar tensor method over traditional radar point cloud methods, especially in complex actions and diverse scenarios. The 4D radar tensor-based approach presents a challenging but promising direction for future research. In summary, our contributions include: + +- The first dataset provides calibrated 4D radar tensors, LiDAR point clouds, and RGB images for human localization and 3D pose estimation of complex actions in diverse scenarios. +- We develop a reliable 3D human pose annotation workflow by joint optimization with LiDAR point cloud and RGB images for both outdoor and indoor scenarios. +- We propose HRRadarPose, the first single-stage human pose estimator specifically designed to learn 4D radar tensors' representation for the human pose estimation task. + +# 2 Related Works + +# 2.1 3D Human Pose Estimation Datasets + +As one of the fundamental tasks in the computer vision field, there are lots of works introducing datasets and benchmarks for 3D HPE, as shown in Table 1. Human3.6M [19] is the first large-scale dataset for 3D human sensing in lab environments, which contains 3.6-million frames of corresponding 2D and 3D human poses from mocap captured videos of 5 female and 6 male subjects. Compared to Human3.6M, MPI-INF-3DHP [32] is a more challenging 3D human pose dataset captured in the wild. There are 8 subjects with 8 actions captured by 14 cameras covering a greater diversity of poses. 3DPW [41], the first one to include video footage taken from a moving phone camera, includes 60 video sequences. Recently, several works have explored the possibility of HPE from + +other modalities like radar and LiDAR. mmBody [7] consists of synchronized and calibrated mmWave radar point clouds and RGB-D images in different scenes, as well as skeleton/mesh annotations for humans in the scenes. + +mRI [3], a multi-modal 3D HPE dataset using mmWave, RGB-D, and inertial sensors, generates a radar point cloud dataset based on the CFAR method. There are 12 rehabilitation exercises selected to evaluate radar application in home-based health monitoring. HuPR [26] collects 235 sequences of data by calibrated radar and camera for 2D HPE in an indoor environment. + +All open datasets for radar-based human pose estimation are collected within a limited detection area, which contradicts the characteristics of radar applications for tracking. Although detecting within a small area may ensure system accuracy, it limits the system's practical application. Furthermore, to the best of our knowledge, our dataset is the first to include radar, LiDAR, and RGB sequence information with 3D pose annotations. This allows our dataset to provide pose annotations for both indoor and outdoor environments, distinguishing it from existing datasets. + +# 2.2 Radar-based Human Pose Estimation + +In spite of the popularity of RGB-based HPE applications, privacy concerns have driven the exploration of radio or radar modalities for HPE. RF-Pose [54] is a pioneering application in 2D HPE utilizing radio frequency signals. RF-Pose demonstrates the feasibility of pose detection in through-wall scenarios and provide the possibility of predicting the 3D pose using radio signals. + +On the other hand, frequency-modulated continuous wave (FMCW) radar, which operates in the millimeter-wave band (30-300GHz), has been shown to be suitable for detection and HPE tasks. The radar point cloud is widely used to represent the radar signal in HPE [3,4,7,36,37,48,52]. mmMesh [48] introduces a temporal model structure that concurrently captures global and local 3D point cloud structures in spatial dimensions, ensuring precise localization performance for pose estimation. Nonetheless, in cluttered environments, the probability and intensity of radar signal reflections from the human body decrease. Consequently, implementing point cloud-based methods in real-world applications becomes challenging. + +Compared to the radar point cloud methods, utilizing 4D tensor radar signal proves to be more informative and reliable [26, 47, 55, 56]. RF-pose 3D [55] is the first work using the 4D RF tensor to predict 3D skeletons in 22 different locations. However, the used RF system is not commercially available, which is hard to reproduce. HuPR [26] utilizes two pieces of well-calibrated TI mmWave radar modules to capture vertical and horizontal radar signals, enhancing the resolution of radar elevation signals to improve pose estimation performance. RPM2.0 [47] adopts a similar hardware setup to HuPR but utilizes the HRNet [44] as the backbone and includes an attention model to reconstruct missing keypoints. However, their two-stage model is trained by only walking pose, restricting the actions' complexity for achieving better pose estimation performance. The methods proposed in HuPR and RPM2.0 fuse information from both + +![](images/59fcc3666594ad74f304d77ff1b378a4d0237a8eb9c73cb9433fe295b6b66cca.jpg) +(a) + +![](images/0e1ac6a3978f2d7ec2db7f1789aca6815fdbbe986060c867eec575f47de5e33c.jpg) +(b) +Fig. 2: Data distribution for RT-Pose dataset: (a) Activities; (b) Environmental conditions; (c) Occlusion conditions + +![](images/0ec4447a1856385525a9377a028ada88d698cdbf90aeeecb2c6d03b953fff01c.jpg) +(c) + +radar modules. While the dual-radar-module configuration potentially captures fine-grained information, its intricate setup poses reproducibility challenges. + +Unlike previous works, our system employs a single radar module for vertical and horizontal sensing, simplifying the sensors' calibration and synchronization. We use 4D radar tensor as training data for 3D HPE, reducing the likelihood of data loss during radar point generation. Moreover, RT-Pose dataset includes complex actions and is recorded in diverse scenes, expanding the applicability to real-world scenarios. + +# 3 RT-Pose Dataset + +To construct the dataset, we recruit 10 participants to perform 6 types of actions. The experiments are divided into two categories. In the first category, participants are required to perform actions while standing, including waving, lifting legs, and random poses to increase the complexity of the actions. A video sequence of random poses involves a subject randomly performing one possible movement at a time, including stretching, bending, twisting, etc., each lasting 3 to 5 seconds. The second category focuses on walking, with additional actions added during the process, such as walking and waving or sitting down + +after walking. Each type of action is inherited from previous radar-based datasets [4,35,48]. Each sequence contains one action for 30 seconds, providing a variety of different difficulty levels. As a result, people can thoroughly evaluate their 3D HPE methods on a broad distribution of actions with diverse difficulties. Furthermore, we provide the dataset development kit to facilitate the usage and evaluation of future methods. + +![](images/2f883ef3f042c30e0dc2f69f3e32a699d4c014be7d09a91f0a9648b9c2f41510.jpg) +Fig.1: Experimental hardware setup in an indoor environment for data collection. + +# 3.1 Sensors + +The data collection hardware system comprises two RGB cameras, a non-repetitive horizontal scanning LiDAR, and a cascade imaging radar module, as shown in Figure 1. To extract human pose information from radar signals, the radar module's parameters have been specifically configured. The radar operates at ten frames per second in this work. The radar module is equipped with 12 transmit (TX) and 16 receive (RX) antenna elements and operates in a multiple-input multiple-output (MIMO) mode [23]. This radar module provides the azimuth and elevation angle resolutions of 1.4 degrees and 18 degrees, respectively. Furthermore, this work implements a high-frequency slope with a long ramp time to efficiently utilize the hardware's sweep bandwidth, ranging from 77 GHz to almost 81 GHz. The longer sweep bandwidth of a chirp increases the radar range resolution [24,37], which is $4.5\mathrm{cm}$ in this work. Each frame consists of 64 chirps used for calculating radar Doppler signals, providing a velocity resolution of 3.9 cm/s. These carefully chosen configurations ensure the system's accuracy and reliability in capturing human poses. + +# 3.2 Data Collection + +We collect the dataset in 40 scenes with indoor and outdoor environments. Figure 3 shows some of the collected data. Indoor environments include clean and cluttered conditions, while outdoor ones include normal, rainy weather, and lowlighting conditions. Each experimental condition includes 8 scenarios, which are nonocclusive, and three occlusive conditions (cardboard, cloth, and plastic pad). Each nonocclusive and occlusive condition is recorded in single- and multi-person settings. To consider the different fields of view (FoV) of all sensors, the sensor suite is situated $109~\mathrm{cm}$ above ground, and an appropriate detection area is planned based on different scenarios. + +# 3.3 Data Processing + +The captured radar signal, referred to as 4D tensor data, is processed by a series of steps [16,33,45], as shown in Figure 4. On the transmitter side, the frequency of the transmitted signals, or chirps, periodically increases based on the set frequency slope. Consequently, the received signal exhibits a different frequency than the transmitted signal. The frequency difference between transmitted and received signals can be used to estimate the range of the object reflecting the signal to the radar sensor by the Fast Fourier Transform (FFT). At the beginning of the data processing, each radar frame consists of 64 chirps, resulting in 64 range estimation results. Utilizing this range information, the velocity of the responding object can be measured through FFT, a phenomenon known as the Doppler effect. Second, the radar signal data is re-modulated according to the antenna position. This step is essential for the calculation of the angle of arrival (AoA), enabling the generation of highangular resolution results. Third, the + +![](images/79ce2e9e719cb60a92a3492f1dbbf59853ad293e1a7bf2e224873558e2680697.jpg) +Fig. 3: Experimental instances across various indoor and outdoor conditions with diverse scenarios for data collection. + +![](images/b5b68f0127ea85644d9a4d36821f384d89bc2fa95a105c617582ad510eae9a31.jpg) +Fig. 4: Radar signal processing flow. The green arrow line is for radar point cloud generation and the blue line is for 4D radar tensor generation. + +radar signal from the azimuth and elevation directions also undergoes FFT processing to represent velocity responses in different directions. Finally, the entire radar data is transformed from polar coordinates to Cartesian coordinates. This transformation enhances the intuitiveness for the subsequent pose estimation. In this study, the processed 4D radar tensor has the dimension of $64 \times 32 \times 128 \times 256$ , which correspond to velocity, the $z$ -axis, the $y$ -axis, and the $x$ -axis, respectively. + +# 3.4 Annotation Workflow + +To achieve accurate human subject detection in a larger area with high-quality ground truth, our system integrates LiDAR and RGB camera data, as illustrated in Figure 5. The images provided by the camera are processed using the HRNet [38] to extract 2D human poses. Then, a 3D HPE model, called ZeDO [22], utilizes the 2D poses to estimate the 3D poses as the initial 3D pseudo ground truth. However, estimating 3D poses using a monocular camera often lacks precision and stability in depth. Therefore, this work incorporates LiDAR for depth + +![](images/87050120d069cfb845b98cde6c52d02116fcbac86c8fe86a07a954642d1abdbe.jpg) +Fig. 5: Workflow of human localization and 3D pose ground truth annotations. Estimated 2D pose results, predicted by the pre-trained HRNet model, are denoted as $P_{2d}$ . The initial setting pose derived from LiDAR point clouds is denoted as $P_{init}$ . Both $P_{2d}$ and $P_{init}$ are inputs into ZeDO, an optimization-based pipeline for 3D pose estimation. + +estimation of the human subject center. The LiDAR sensor captures environmental information and the subject's point cloud data, ensuring errors are under $2\mathrm{cm}$ inside the detection range of $20\mathrm{m}$ . The human subject is manually labeled with a 3D bounding box, which can provide an accurate 3D position. + +ZeDO, a state-of-the-art (SOTA) method for cross-dataset HPE, implements an optimization-based pipeline for 3D pose estimation using 2D poses. This method requires 2D poses as input, and ZeDO is able to estimate multiple-hypothesis 3D poses. In this research, the center of the 3D bounding box provided by LiDAR is utilized as the pelvis of the initial pose, which enhances the accuracy of depth estimation compared to relying solely on the monocular camera data, and improves the performance of ZeDO. With the imported 2D pose from HRNet [43], ZeDO iteratively refines the 3D poses. To further improve annotation accuracy, an annotation tool is developed for manual selection and optimization of each frame to retain correct 3D human poses. The 3D HPE results from ZeDO will be put in the LiDAR point cloud for visualization, which is beneficial for humans to filter and modify the error pose annotation results. About $30\%$ of the data is removed or optimized after human correction, and this process ensures the quality of keypoints annotation. More details can be found in the supplementary material. + +# 4 HRRadarPose + +Inspired by HRNet [43], renowned for its adeptness in extracting high-resolution representations, we build an architecture, HRRadarPose, with fully 3D convolutional layers. By treating the Doppler axis, $D$ , as the input channels, our model extracts volumetric spatial features along the $Z$ , $Y$ , and $X$ axes, representing vertical, horizontal, and depth directions in the Cartesian coordinate system. The HRRadarPose architecture is designed to preserve spatial resolution and assimi + +late semantic information from the 4D radar tensors, maintaining high-resolution representations and enabling the exchange of information between features of different resolutions during stage transitions. Figure 6 illustrates an instance of the HRRadarPose structure with three stages. The structure starts with using 3D convolutional layers to transform the 4D radar tensors, $T \in \mathbb{R}^{D \times Z \times Y \times X}$ , into features, $f_{in} \in \mathbb{R}^{C_1 \times Z_1 \times Y_1 \times X_1}$ . As the network proceeds from one stage to the next, it develops additional branches through strided convolution to widen the receptive field. Each branch $i$ employs strided convolution to elevate its input feature $f_i$ to a feature $f_{i+1}$ in the succeeding branch $i+1$ , simultaneously doubling the channels and halving the spatial dimensions, $f_{i+1} \in \mathbb{R}^{2C_i \times \frac{Z_i}{2} \times \frac{Y_i}{2} \times \frac{X_i}{2}}$ . Within each stage, a series of $P$ modules execute $M$ parallel convolutions. To align feature dimensions from different branches for information exchange, we use strided convolution followed by upsampling. At the end of the backbone, we take the feature with the highest resolution to serve as a unified representation for the pose estimation head. In the pose estimation head, two 3D convolutional branches are employed to process these features further. This bifurcation generates two outputs: a distribution map signifying the confidence of locating the human center and the location offset of joint keypoints. The right part of Figure 6 shows the decoding of the HRRadarPose network's output to multi-person poses. The process starts with picking top- $k$ human centers, $C_S$ , with confidence scores over a threshold value. We denote these picked centers' indices as $S$ . Then, we query the centers corresponding joint keypoints' offsets, $K_S$ , by looking up the output map of keypoint offsets head at $S$ . Lastly, we decode their joint keypoint locations, $J_S$ , by summing $K_S$ and $C_S$ . In scenarios with multiple people, we apply range non-maximum suppression (NMS) to reduce overlapping detections. Our method is favorable in two folds: The HRRadarPose's architecture gains efficiency through a single-stage workflow, avoiding complexity brought by region-proposal-based methods [55]. In addition, our pose estimation head inherently learns per person's joint keypoints refraining from ambiguity of association between people's identity and joints. + +To train the center probability head, we use the pixel-wise focal loss as the classification loss, $L_{\text{class}}$ , defined as: + +$$ +L _ {\text {c l a s s}} = \left\{ \begin{array}{l l} (1 - P _ {x y z}) ^ {\alpha} \log \left(P _ {x y z}\right), & \text {i f} Y _ {x y z} = 1, \\ (1 - Y _ {x y z}) ^ {\beta} \left(P _ {x y z}\right) ^ {\alpha} \log \left(1 - P _ {x y z}\right), & \text {o t h e r w i s e}, \end{array} \right. \tag {1} +$$ + +where $Y_{xyz}$ is the three-dimensional human center Gaussian distribution map generated by pelvis locations, $P_{xyz}$ is the probability of a person's presence at location, $(x,y,z)$ , $N$ is the number of foreground samples, $\alpha$ and $\beta$ are hyperparameters addressing class imbalance and focusing training on foreground samples. To train the keypoint offset head, we construct the regression loss $L_{reg}$ by taking the average of 15 keypoints' $\mathcal{L}_1$ , comparing the distance between predicted offsets and ground-truth offsets to the center location of a voxel in $Y$ when $Y_{xyz}$ equals to 1. The final loss is a weighted sum of $L_{class}$ and $L_{reg}$ . + +![](images/241762dcba3e0b1f7538100561e8984227e37a1aa699bdb6d14b9d31a7694fbd.jpg) +Fig. 6: HRRadarPose's architecture consists of a 3D convolutional backbone and a pose estimation head, which generates a body center's confidence map and joint keypoint offsets map. To infer multi-person poses, we decode the prediction by adding the picked bodies' centers with their corresponding keypoint offsets. + +# 5 Experiments + +15 obvious keypoints of a human body are selected, aligning with the body skeleton model of the Human3.6M dataset [19]. Each frame's 3D pose annotation in the sequence is manually validated to ensure the quality of the dataset. Figure 7 demonstrates the testing results across various levels of action complexity, presenting the feasibility of 3D pose estimation using the proposed baseline. + +Evaluation metrics We employ joint position error (JPE), which measures the mean of the Euclidean distance for every joint's predicted keypoints and the ground truth keypoints. Localization results are reported by the mean of the root position error (MRPE) [42], which is the JPE of the pelvis keypoint. For the evaluation of 3D HPE, we use Mean per joint position error (MPJPE) [57] and absolute-MPJPE (Abs-MPJPE) [40]. + +Baselines Comparison We evaluate the performance of the HRRadarPose in global localization and 3D HPE on the RT-Pose dataset. We compare our method with three baseline methods, as shown in Table 2. mm-Pose [37] and mmMesh [48], utilizing radar point clouds for 3D HPE, are re-implemented to test on the RT-Pose dataset. Notably, mmMesh introduces a global localization model, combined with an RNN HPE module, showing more accurate estimation results in both MRPE and MPJPE compared to mm-Pose, which employs a simpler CNN model. RF-Pose 3D [55] uses the 4D radar tensor with a region proposal network to locate the subjects and estimate the 3D pose. The experimental results show that our HRRadarPose outperforms other baseline methods. On the RT-Pose dataset, we obtain $9.93\mathrm{cm}$ in MPJPE and $9.91\mathrm{cm}$ in MRPE. We find that radar-tensor-based methods reach less error than the methods using radar point cloud. We attribute this performance gap to complicated scenarios since it is difficult for point cloud-based methods to discriminate different key + +![](images/63e42322a68c127528fe73d396d6a271b5ffbabd9b968ec1691aa24153cbd644.jpg) +R + +![](images/7d52982fa3bf129fca7a2fa5286c7ff1af4f8280866da17da68d75ef34e15992.jpg) + +![](images/2df9a21432c8938e5aef8da4b312ee6440a1f79b157f2f32da8ba764b43577fc.jpg) + +![](images/e96abddca39f5299a3c42fa75e5072282cbb909d8eefc2b0302c63bf15aada12.jpg) + +![](images/d96b9207b858c26dd2461d2efdc355e75335f2bb5662bff4bc4f630d3497c6f4.jpg) + +![](images/303c8131a5567b639ec9976f127c0c0116d037e7a56ca71cda437c23a808f572.jpg) + +![](images/478994c069a6abb065e22d4cca11396c331995ef319bc31b121c3136b15e6980.jpg) +BREV TraJ. + +![](images/88098aad8f31ccec0135875708e0825fce8a50307c4210204a3d4332be7b5c08.jpg) + +![](images/5215f3e12d9cf52b1f1348b838588bdc51682ed34052fc434df41fae5f645614.jpg) + +![](images/62839eadb20b9da3e78537cdedf9cadd5696da9cc712da1ed23b2c536eb906d5.jpg) + +![](images/5d0809a9dac003984daf6643138040c398a899454bd15047d43cac3062305cb2.jpg) + +![](images/b74a7d4e620d40b328edd18013fc775ff7b08bd5ec974a30582134958a1e3aab.jpg) + +![](images/a8605e18fe78ab26bfb98f21921c76915d3ccaab1bdbbc4ca6dacac130e8e22a.jpg) +3D pose +(a) Waving hand + +![](images/f8651a21d82d7697fce8e273b7630003399a61ac706632f0cf61e8988e0737d9.jpg) +(b) Lifting leg + +![](images/68e55af94ad4b75044313768531c00e7f94ff9e04f9f62878bbc251bdd138fd3.jpg) +(c) Random pose + +![](images/8907c83bf211cf42fe61c003fb8fd1b2a42a11f85eebdcd34228c2c413b50112.jpg) +(d) Normal + +![](images/a76f2b5b3fd3dd1b7acfc221eb0f91d28f02ed4db7ad0f5044627352c12ba6a7.jpg) +(e) Sitting + +![](images/3ae9f7e212ad9f93384faa03c49d95912d49b683c2978e816088973c08b0d12b.jpg) +(f) Waving hand + +![](images/2c99d341c382c9eee4bf5a84506dff4d830d506f3db61d65e9124138a8557563.jpg) +Fig. 7: Visualization of pose estimation results in test frames for six activities. The second row is the trajectory of the test sequence in BEV. The third row is the instance of the global 3D pose estimated result for each action. In both rows, the blue line is the ground truth, and the red line represents estimated results. + +Table 2: Comparison of human localization and 3D pose estimation results of our HRRadarPose method and three baseline methods. † represents the radar point cloud-based methods. + +
MethodMRPE (cm)MPJPE (cm)Abs-MPJPE (cm)
mm-Pose † [37]102.2821.26102.7
mmMesh † [48]62.6913.7866.24
RF-Pose3D [55]29.1718.4338.05
HRRadarPose (Ours)9.919.9314.73
+ +points in various poses. The results highlight the importance of using 4D radar tensors, which preserve raw and rich spatial-temporal information for the HPE tasks. An end-to-end architecture directly takes in 4D radar tensors to better face different human action scenarios in the real world. + +Action complexity Analysis RT-posedataset provides six difficulty levels of actions to simulate real-world scenarios, enabling comprehensive evaluation of localization and 3D HPE results. We report the performance of the proposed HRRadarPose method in Table 3. Generally, the joints along the human longitudinal axis, such as the thorax and the head, are more stable than those in the limbs. This phenomenon may be attributed to the larger reflective surface of the body's trunk compared to the limbs, providing a more reliable HPE based on radar signals. + +The MRPE of standing posture is below $7.67~\mathrm{cm}$ , indicating that the pose complexity slightly influences the localization accuracy. Due to the chair's strong + +Table 3: Comparison of the performance predicted by the HRRadar Pose in different complexity level activities, presenting different key points of the pose, the results of localization, and total 3D pose estimation results. The first group of rows lists the stand-related poses, while the other lists the walking-related poses. The complexity level increases from the top activity to the bottom within each group. + +
ActionsJPE(cm)MRPE (cm)MPJPE (cm)Abs-MPJPE (cm)
ThoraxHeadAnkleWrist
Stand: Lifting LegWaving Hand2.953.973.6125.653.356.917.97
6.127.9212.248.887.557.0610.37
Random Pose7.7811.2111.7532.127.6712.4715.12
Walk: SitingNormal4.316.6913.5215.9813.678.9316.92
9.3613.0313.9310.7714.119.4416.83
Waving Hand5.7410.3713.5343.1612.8215.2821.22
+ +![](images/0e19c1e2e75c5fecfb092a70f8b0919df69636e28e70b75392881e2db9134d30.jpg) +Fig. 8: HRRadarPose Results in different conditions. (a), (b), and (c) are occluded by cardboard, cloth, and plastic pad, respectively. (d) is in low-lighting conditions + +radar reflection, distinguishing between human subjects and the chair poses a challenge for localization. The MPJPE of walking and sitting is $9.44\mathrm{cm}$ , showing our HRRadarPose model can reliably estimate the sitting pose. However, the MRPE of this action sequence achieves $14.11\mathrm{cm}$ , demonstrating that sitting is a key factor for evaluating the performance of localization. The MPJPE of simple actions, such as standing and waving, standing and lifting legs, are around 7 cm, and the walking is $8.93\mathrm{cm}$ , demonstrating that our HRRadarPose model is comparable to other recently published radar pose estimate methods [26, 47]. However, the current model still encounters difficulty in estimating HPE in complex actions such as walking and waving. Therefore, the RT-pose dataset is a challenging benchmark, which can enable HPE models to realistically learn diverse pose scenarios. + +Table 4: HPE results from HRRadarPose in normal and occlusive scenarios. + +
OcclusionMPJPE (cm)Abs-MPJPE (cm)
✓9.8414.52
11.5718.55
+ +Table 5: Ablation Study of Using Dopper information. + +
Doppler|MPJPE (cm) Abs-MPJPE (cm)
11.3122.34
✓9.9114.73
+ +Table 6: Our Radar-Doppler Convolutional Block in HR-Net. + +
HRNet-Block-DesignMPJPE (cm)Abs-MPJPE (cm)
Traditional Block (RGB-Vision)10.3517.59
Ours (Radar-Doppler)9.9314.73
+ +Qualitative Result Figure 8 demonstrates the qualitative results of different occlusion conditions and a low-lighting instance. Due to the hardware setting of our system, supporting synchronized two camera data, the subject's pose can be visualized by another camera if one of them is occluded. Therefore, the 3D annotation results can be generated even on occlusion conditions. Our method provides reliable pose estimation results even when the radar module is blocked or with low-lighting conditions. Table 4 shows the HPE results of occlusion in normal outdoor conditions. Although the radar module is occluded, the mmWave signal is still robust enough as input for the HRRadarPose model to estimate HPE, demonstrating the characteristics of radar and the importance of using radar for HPE. + +Ablation Study To illustrate the advantage of modeling motion features using Doppler information by our HRRadarPose, we remove Doppler information by averaging 4D radar tensors along the Doppler axis and train the models with and without Doppler information using the same setup. Table 5 shows that with the inclusion of Doppler information, our HRRadarPose enjoys a performance gain due to its capability to extract motion features, enhancing keypoint localization. + +Compared with RPM 2.0, our input data and backbone design are different. The proposed HRRadarPose network extracts better feature representation from 4D radar tensors through modified convolutional blocks. As shown in Figure 9, we replace the batch norm with the group norm of convolutional blocks in HRNet. Since we treat the channel-wise features as a Doppler axis, the group norm inherently handles the feature in different speed groups, which normalizes the data effectively. Table 6 shows the results, proving that the modified CNN block is more suitable for Doppler radar than the original HRNet. Compared with our method, most previous work did not leverage Doppler information [47,50]. We validate that motion features [17,27] are crucial for aiding deep learning models in capturing more human body characteristics over the environment. + +To ablate our design choice of HPE heads, we use the same backbone, replace our pose estimation head with RF-Pose3D's [55] head, and train both models for + +![](images/e9d7b332004f99b083bd1c95c5fda7c8bdba0823064fa183233a3bc08fd23c6c.jpg) +Fig.9: Illustration of two variants of CNN blocks in HRNet. (a) Ours (Radar-Doppler). (b) Traditional Block (RGBVision). + +Table 7: Comparison of common 4D radar tensor-based HPE head and ours. + +
HPE HeadMPJPE (cm)Abs-MPJPE (cm)
RF-Pose3D [55]15.5218.08
Ours10.2417.63
+ +the same epochs. RF-Pose3D formulates the HPE as a multi-keypoints classification problem, where their model outputs confidence distributions of every human body joint. As shown in Table 7, our design is favorable in joint keypoints localization in the metric space, indicated by a lower Abs-MPJPE. Moreover, lower MPJPE from our HPE head's result validates stronger inter-joint relationships modeling capability. + +# 6 Limitation and Future Work + +First, to leverage the input data types for training the 3D HPE model, the computational resource is a crucial part. 4D radar tensor in Cartesian coordinate consumes up to $100\mathrm{MB}$ per frame, which decreases the training speed. Second, we limit the experimental data collection scope from $2\mathrm{m}$ to $8\mathrm{m}$ because the camera is hard to clearly capture 2D pose skeletons at a large distance and LiDAR can not easily collect completed point clouds in near vision due to the constraints of FoV. Third, the proposed baseline HRRadarPose lacks robustness in accurately estimating poses during complex activities. Capturing human pose accurately on 4D radar tensor remains an unresolved challenge. + +# 7 Conclusion + +In this paper, we propose RT-Pose, the first dataset for human pose estimation (HPE) with synchronized and calibrated 4D radar tensors, LiDAR point clouds, and RGB images. RT-Pose provides cluttered scenarios and human activities of different complexity levels to enhance the difficulty of pose estimation for practical applications. In addition, having various modalities is beneficial for our semi-automatic 3D pose annotation optimization process. The proposed HRRadarPose is the first single-stage architecture designed for estimating human pose using 4D radar tensors as the input. The results indicate that the proposed HRRadarPose outperforms previous works, in which datasets are collected with simpler actions and cleaner scenarios. Our work contributes to offering a three-modality dataset with 4D radar tensors for HPE, which is promising to be applied for complex actions and different scenes. We hope this work encourages future development of 4D radar-based HPE methods. + +# Acknowledgements + +We thank Yizhou Wang $^{3}$ for helping us build the data collection system and the radar data processing pipeline. We thank Hou-I Liu $^{4}$ , Cheng-Yi Huang $^{5}$ , Yi Shiang Chen $^{6}$ , Chi-Yuan Chan $^{7}$ , and Yu-Hsien Lu $^{8}$ for helping us collect the RT-Pose dataset. We thank the National Science and Technology Counci of the Republic of China, Taiwan, under Projects of NSTC 112-2917-I-006-007 for its funding support. + +# References + +1. Adib, F., Hsu, C.Y., Mao, H., Katabi, D., Durand, F.: Rf-capture: Capturing the human figure through a wall. ACM SIGGRAPH Asia (2015) +2. Ahuja, K., Jiang, Y., Goel, M., Harrison, C.: Vid2doppler: Synthesizing doppler radar data from videos for training privacy-preserving activity recognition. In: Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. pp. 1-10 (2021) +3. An, S., Li, Y., Ogras, U.: mri: Multi-modal 3d human pose estimation dataset using mmwave, rgb-d, and inertial sensors. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems. vol. 35, pp. 27414-27426. Curran Associates, Inc. (2022) +4. An, S., Ogras, U.Y.: Mars: mmwave-based assistive rehabilitation system for smart healthcare. ACM Transactions on Embedded Computing Systems (TECS) 20(5s), 1-22 (2021) +5. Chai, W., Guo, X., Wang, G., Lu, Y.: Stablevideo: Text-driven consistency-aware diffusion video editing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 23040-23050 (2023) +6. Chai, W., Jiang, Z., Hwang, J.N., Wang, G.: Global adaptation meets local generalization: Unsupervised domain adaptation for 3d human pose estimation. arXiv preprint arXiv:2303.16456 (2023) +7. Chen, A., Wang, X., Zhu, S., Li, Y., Chen, J., Ye, Q.: mmbody benchmark: 3d body reconstruction dataset and analysis for millimeter wave radar. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 3501-3510 (2022) +8. Chen, K., Gabriel, P., Alasfour, A., Gong, C., Doyle, W.K., Devinsky, O., Friedman, D., Dugan, P., Melloni, L., Thesen, T., et al.: Patient-specific pose estimation in clinical environments. IEEE journal of translational engineering in health and medicine 6, 1-11 (2018) +9. Cheng, J.H., Chen, Y., Chang, T.Y., Lin, H.E., Wang, P.Y.C., Cheng, L.P.: Impossible staircase: Vertically real walking in an infinite virtual tower. In: 2021 IEEE Virtual Reality and 3D User Interfaces (VR). pp. 50-56 (2021) + +10. Cheng, J.H., Kuan, S.Y., Liu, H.I., Latapie, H., Liu, G., Hwang, J.N.: Centerradar-net: Joint 3d object detection and tracking framework using 4d fmcw radar. In: 2024 IEEE International Conference on Image Processing (ICIP) (2024) +11. Dai, Y., Lin, Y., Lin, X., Wen, C., Xu, L., Yi, H., Shen, S., Ma, Y., Wang, C.: Sloper4d: A scene-aware dataset for global 4d human pose estimation in urban environments. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 682-692 (2023) +12. Dai, Y., Lin, Y., Wen, C., Shen, S., Xu, L., Yu, J., Ma, Y., Wang, C.: Hsc4d: Human-centered 4d scene capture in large-scale indoor-outdoor space using wearable imus and lidar. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6792-6802 (2022) +13. De Maio, A., Farina, A., Foglia, G.: Design and experimental validation of knowledge-based constant false alarm rate detectors. IET Radar, Sonar & Navigation 1(4), 308-316 (2007) +14. Endo, K., Ishikawa, T., Yamamoto, K., Ohtsuki, T.: Multi-person position estimation based on correlation between received signals using mimo fmcw radar. IEEE Access 11, 2610-2620 (2023) +15. Guzov, V., Mir, A., Sattler, T., Pons-Moll, G.: Human poseitioning system (hps): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4318-4329 (2021) +16. Heath, R.W., Gonzalez-Prelcic, N., Rangan, S., Roh, W., Sayeed, A.M.: An overview of signal processing techniques for millimeter wave mimo systems. IEEE journal of selected topics in signal processing 10(3), 436-453 (2016) +17. Huang, H.W., Yang, C.Y., Ramkumar, S., Huang, C.I., Hwang, J.N., Kim, P.K., Lee, K., Kim, K.: Observation centric and central distance recovery for athlete tracking. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 454-460 (2023) +18. Huang, H.Y., Ning, C.W., Wang, P.Y., Cheng, J.H., Cheng, L.P.: Haptic-go-round: A surrounding platform for encounter-type haptics in virtual reality experiences. In: Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems. CHI '20, Association for Computing Machinery (2020) +19. Ionescu, C., Papava, D., Olaru, V., Sminchisescu, C.: Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence 36(7), 1325-1339 (2013) +20. Jiang, Z., Chai, W., Li, L., Zhou, Z., Yang, C.Y., Hwang, J.N.: Unihpe: Towards unified human pose estimation via contrastive learning. arXiv preprint arXiv:2311.16477 (2023) +21. Jiang, Z., Ji, H., Yang, C.Y., Hwang, J.N.: 2d human pose estimation calibration and keypoint visibility classification. In: ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 6095-6099. IEEE (2024) +22. Jiang, Z., Zhou, Z., Li, L., Chai, W., Yang, C.Y., Hwang, J.N.: Back to optimization: Diffusion-based zero-shot 3d human pose estimation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6142-6152 (2024) +23. Kim, Y., Alnujaim, I., Oh, D.: Human activity classification based on point clouds measured by millimeter wave mimo radar with deep recurrent neural networks. IEEE Sensors Journal 21(12), 13522-13529 (2021) + +24. Klauder, J.R., Price, A., Darlington, S., Albersheim, W.J.: The theory and design of chirp radars. Bell System Technical Journal 39(4), 745-808 (1960) +25. Kuan, S.Y., Cheng, J.H., Huang, H.W., Chai, W., Yang, C.Y., Wu, B.F., Hwang, J.N.: Boosting online 3d multi-object tracking through camera-radar cross check. In: IEEE Intelligent Vehicles Symposium (IV) (2024) +26. Lee, S.P., Kini, N.P., Peng, W.H., Ma, C.W., Hwang, J.N.: Hupr: A benchmark for human pose estimation using millimeter wave radar. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5715-5724 (2023) +27. Lin, C.L., Chang, Y.M., Hung, C.C., Tu, C.D., Chuang, C.Y.: Position estimation and smooth tracking with a fuzzy-logic-based adaptive strong tracking kalman filter for capacitive touch panels. IEEE Transactions on Industrial Electronics 62(8), 5097-5108 (2015) +28. Lin, C.L., Chiu, W.C., Chu, T.C., Ho, Y.H., Chen, F.H., Hsu, C.C., Hsieh, P.H., Chen, C.H., Lin, C.C.K., Sung, P.S., et al.: Innovative head-mounted system based on inertial sensors and magnetometer for detecting falling movements. Sensors 20(20), 5774 (2020) +29. Lin, C.L., Ho, Y.H., Chiu, W.C., Chu, T.C., Liu, Y.H.: Innovative shoe-integrated system based on time-of-flight range sensors for fall detection on various terrains. IEEE Sensors Letters 5(10), 1-4 (2021) +30. Lin, H.Y., Chen, T.W.: Augmented reality with human body interaction based on monocular 3d pose estimation. In: International Conference on Advanced Concepts for Intelligent Vision Systems. pp. 321-331. Springer (2010) +31. Liu, H., He, J.Y., Cheng, Z.Q., Xiang, W., Yang, Q., Chai, W., Wang, G., Bao, X., Luo, B., Geng, Y., et al.: Posynda: Multi-hypothesis pose synthesis domain adaptation for robust 3d human pose estimation. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 5542-5551 (2023) +32. Mehta, D., Rhodin, H., Casas, D., Fua, P., Sotnychenko, O., Xu, W., Theobalt, C.: Monocular 3d human pose estimation in the wild using improved cnn supervision. In: 2017 international conference on 3D vision (3DV). pp. 506-516. IEEE (2017) +33. Neemat, S., Uysal, F., Krasnov, O., Yarovoy, A.: Reconfigurable range-doppler processing and range resolution improvement for fmcw radar. IEEE Sensors Journal 19(20), 9294-9303 (2019) +34. Paek, D.H., Kong, S.H., Wijaya, K.T.: K-radar: 4d radar object detection for autonomous driving in various weather conditions. Advances in Neural Information Processing Systems 35, 3819-3829 (2022) +35. Sengupta, A., Cao, S.: mmpose-nlp: A natural language processing approach to precise skeletal pose estimation using mmwave radars. IEEE Transactions on Neural Networks and Learning Systems (2022) +36. Sengupta, A., Jin, F., Cao, S.: Nlp based skeletal pose estimation using mmwave radar point-cloud: A simulation approach. In: 2020 IEEE Radar Conference (RadarConf20). pp. 1-6. IEEE (2020) +37. Sengupta, A., Jin, F., Zhang, R., Cao, S.: mm-pose: Real-time human skeletal posture estimation using mmwave radars and cnns. IEEE Sensors Journal 20(17), 10032-10044 (2020) +38. Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learning for human pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5693-5703 (2019) +39. Sun, Z., Ke, Q., Rahmani, H., Bennamoun, M., Wang, G., Liu, J.: Human action recognition from various data modalities: A review. IEEE transactions on pattern analysis and machine intelligence (2022) + +40. Véges, M., Lorincz, A.: Absolute human pose estimation with depth prediction network. In: 2019 International Joint Conference on Neural Networks (IJCNN). pp. 1-7. IEEE (2019) +41. Von Marcard, T., Henschel, R., Black, M.J., Rosenhahn, B., Pons-Moll, G.: Recovering accurate 3d human pose in the wild using imus and a moving camera. In: Proceedings of the European conference on computer vision (ECCV). pp. 601-617 (2018) +42. Wang, J., Tan, S., Zhen, X., Xu, S., Zheng, F., He, Z., Shao, L.: Deep 3d human pose estimation: A review. Computer Vision and Image Understanding 210, 103225 (2021) +43. Wang, J., Sun, K., Cheng, T., Jiang, B., Deng, C., Zhao, Y., Liu, D., Mu, Y., Tan, M., Wang, X., Liu, W., Xiao, B.: Deep high-resolution representation learning for visual recognition. TPAMI (2019) +44. Wang, J., Sun, K., Cheng, T., Jiang, B., Deng, C., Zhao, Y., Liu, D., Mu, Y., Tan, M., Wang, X., et al.: Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence 43(10), 3349-3364 (2020) +45. Wang, Y., Cheng, J.H., Huang, J.T., Kuan, S.Y., Fu, Q., Ni, C., Hao, S., Wang, G., Xing, G., Liu, H., Hwang, J.N.: Vision meets mmwave radar: 3d object perception benchmark for autonomous driving. In: IEEE Intelligent Vehicles Symposium (IV) (2024) +46. Wang, Y., Jiang, Z., Li, Y., Hwang, J.N., Xing, G., Liu, H.: Rodnet: A real-time radar object detection network cross-supervised by camera-radar fused object 3d localization. IEEE Journal of Selected Topics in Signal Processing 15(4), 954-967 (2021) +47. Xie, C., Zhang, D., Wu, Z., Yu, C., Hu, Y., Chen, Y.: RPM 2.0: Rf-based pose machines for multi-person 3d pose estimation. IEEE Transactions on Circuits and Systems for Video Technology (2023) +48. Xue, H., Ju, Y., Miao, C., Wang, Y., Wang, S., Zhang, A., Su, L.: mmmesh: Towards 3d real-time dynamic human mesh construction using millimeter-wave. In: Proceedings of the 19th Annual International Conference on Mobile Systems, Applications, and Services. pp. 269-282 (2021) +49. Yang, C.Y., Luo, J., Xia, L., Sun, Y., Qiao, N., Zhang, K., Jiang, Z., Hwang, J.N., Kuo, C.H.: Camerapose: Weakly-supervised monocular 3d human pose estimation by leveraging in-the-wild 2d annotations. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2924-2933 (2023) +50. Yu, C., Zhang, D., Wu, Z., Xie, C., Lu, Z., Hu, Y., Chen, Y.: Mobirfpose: Portable rf-based 3d human pose camera. IEEE Transactions on Multimedia (2023) +51. Yuan, Y., Wei, S.E., Simon, T., Kitani, K., Saragih, J.: Simpoe: Simulated character control for 3d human pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7159-7169 (2021) +52. Zhang, G., Geng, X., Lin, Y.J.: Comprehensive mpoint: A method for 3d point cloud generation of human bodies utilizing fmcw mimo mm-wave radar. Sensors 21(19), 6455 (2021) +53. Zhang, Z., Chai, W., Jiang, Z., Ye, T., Song, M., Hwang, J.N., Wang, G.: Mpm: A unified 2d-3d human pose representation via masked pose modeling. arXiv preprint arXiv:2306.17201 (2023) +54. Zhao, M., Li, T., Abu Alsheikh, M., Tian, Y., Zhao, H., Torralba, A., Katabi, D.: Through-wall human pose estimation using radio signals. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7356-7365 (2018) + +55. Zhao, M., Tian, Y., Zhao, H., Alsheikh, M.A., Li, T., Hristov, R., Kabelac, Z., Katabi, D., Torralba, A.: Rf-based 3d skeletons. In: Proceedings of the 2018 Conference of the ACM Special Interest Group on Data Communication. pp. 267-281 (2018) +56. Zhao, Y., Yarovoy, A., Fioranelli, F.: Angle-insensitive human motion and posture recognition based on 4d imaging radar and deep learning classifiers. IEEE Sensors Journal 22(12), 12173-12182 (2022) +57. Zheng, C., Wu, W., Chen, C., Yang, T., Zhu, S., Shen, J., Kehtarnavaz, N., Shah, M.: Deep learning-based human pose estimation: A survey. ACM Computing Surveys 56(1), 1-37 (2023) +58. Zheng, Z., Pan, J., Ni, Z., Shi, C., Ye, S., Fang, G.: Human posture reconstruction for through-the-wall radar imaging using convolutional neural networks. IEEE Geoscience and Remote Sensing Letters 19, 1-5 (2021) +59. Zhou, Z., Jiang, Z., Chai, W., Yang, C.Y., Li, L., Hwang, J.N.: Efficient domain adaptation via generative prior for 3d infant pose estimation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 41-49 (2024) \ No newline at end of file diff --git a/2024/RT-Pose_ A 4D Radar-Tensor based 3D Human Pose Estimation and Localization Benchmark/images.zip b/2024/RT-Pose_ A 4D Radar-Tensor based 3D Human Pose Estimation and Localization Benchmark/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1a66dae04c4da57d6896f666c2bd925f1c9a1d13 --- /dev/null +++ b/2024/RT-Pose_ A 4D Radar-Tensor based 3D Human Pose Estimation and Localization Benchmark/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2c23572533e7a5d13876f9edc4b19a996e300736d9baadcab9e2531bb4524b1f +size 486638 diff --git a/2024/RT-Pose_ A 4D Radar-Tensor based 3D Human Pose Estimation and Localization Benchmark/layout.json b/2024/RT-Pose_ A 4D Radar-Tensor based 3D Human Pose Estimation and Localization Benchmark/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7dc7c79b7d7b325b640c231e6e7a9785672b2291 --- /dev/null +++ b/2024/RT-Pose_ A 4D Radar-Tensor based 3D Human Pose Estimation and Localization Benchmark/layout.json @@ -0,0 +1,10085 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "spans": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "text", + "content": "RT-Pose: A 4D Radar Tensor-based 3D Human Pose Estimation and Localization Benchmark" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 136, + 167, + 478, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 167, + 478, + 180 + ], + "spans": [ + { + "bbox": [ + 136, + 167, + 478, + 180 + ], + "type": "text", + "content": "Yuan-Hao " + }, + { + "bbox": [ + 136, + 167, + 478, + 180 + ], + "type": "inline_equation", + "content": "\\mathrm{Ho}^{1*}" + }, + { + "bbox": [ + 136, + 167, + 478, + 180 + ], + "type": "text", + "content": ", Jen-Hao Cheng" + }, + { + "bbox": [ + 136, + 167, + 478, + 180 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 136, + 167, + 478, + 180 + ], + "type": "text", + "content": ", Sheng Yao Kuan" + }, + { + "bbox": [ + 136, + 167, + 478, + 180 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 136, + 167, + 478, + 180 + ], + "type": "text", + "content": ", Zhongyu Jiang" + }, + { + "bbox": [ + 136, + 167, + 478, + 180 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 136, + 167, + 478, + 180 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 173, + 180, + 441, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 180, + 441, + 193 + ], + "spans": [ + { + "bbox": [ + 173, + 180, + 441, + 193 + ], + "type": "text", + "content": "Wenhao Chai² Hsiang-Wei Huang², Chih-Lung Lin¹, and" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 261, + 193, + 353, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 193, + 353, + 205 + ], + "spans": [ + { + "bbox": [ + 261, + 193, + 353, + 205 + ], + "type": "text", + "content": "Jenq-Neng Hwang2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 241, + 205, + 373, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 205, + 373, + 217 + ], + "spans": [ + { + "bbox": [ + 241, + 205, + 373, + 217 + ], + "type": "text", + "content": "* Indicates equal contribution" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 235, + 225, + 378, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 225, + 378, + 237 + ], + "spans": [ + { + "bbox": [ + 235, + 225, + 378, + 237 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 235, + 225, + 378, + 237 + ], + "type": "text", + "content": " National Cheng Kung University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 249, + 237, + 364, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 237, + 364, + 248 + ], + "spans": [ + { + "bbox": [ + 249, + 237, + 364, + 248 + ], + "type": "text", + "content": "2 University of Washington" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 244, + 248, + 368, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 248, + 368, + 259 + ], + "spans": [ + { + "bbox": [ + 244, + 248, + 368, + 259 + ], + "type": "text", + "content": "n28081527@gs.ncku.edu.tw1," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 165, + 259, + 447, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 259, + 447, + 270 + ], + "spans": [ + { + "bbox": [ + 165, + 259, + 447, + 270 + ], + "type": "text", + "content": "{andyhci, shengyao, zyjiang, wchai, hwhuang, hwang}@uw.edu²," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 255, + 270, + 353, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 270, + 353, + 280 + ], + "spans": [ + { + "bbox": [ + 255, + 270, + 353, + 280 + ], + "type": "text", + "content": "cllin@ee.ncku.edu.tw" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 159, + 310, + 455, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 310, + 455, + 563 + ], + "spans": [ + { + "bbox": [ + 159, + 310, + 455, + 563 + ], + "type": "text", + "content": "Abstract. Traditional methods for human localization and pose estimation (HPE), which mainly rely on RGB images as an input modality, confront substantial limitations in real-world applications due to privacy concerns. In contrast, radar-based HPE methods emerge as a promising alternative, characterized by distinctive attributes such as through-wall recognition and privacy-preserving, rendering the method more conducive to practical deployments. This paper presents a Radar Tensor-based human pose (RT-Pose) dataset and an open-source benchmarking framework. RT-Pose dataset comprises 4D radar tensors, Li-DAR point clouds, and RGB images, and is collected for a total of " + }, + { + "bbox": [ + 159, + 310, + 455, + 563 + ], + "type": "inline_equation", + "content": "72\\mathrm{k}" + }, + { + "bbox": [ + 159, + 310, + 455, + 563 + ], + "type": "text", + "content": " frames across 240 sequences with six different complexity level actions. The 4D radar tensor provides raw spatio-temporal information, differentiating it from other radar point cloud-based datasets. We develop an annotation process, which uses RGB images and LiDAR point clouds to accurately label 3D human skeletons. In addition, we propose HRRadarPose, the first single-stage architecture that extracts the high-resolution representation of 4D radar tensors in 3D space to aid human keypoint estimation. HRRadarPose outperforms previous radar-based HPE work on the RT-Pose benchmark. The overall HRRadarPose performance on the RT-Pose dataset, as reflected in a mean per joint position error (MPJPE) of " + }, + { + "bbox": [ + 159, + 310, + 455, + 563 + ], + "type": "inline_equation", + "content": "9.91cm" + }, + { + "bbox": [ + 159, + 310, + 455, + 563 + ], + "type": "text", + "content": ", indicates the persistent challenges in achieving accurate HPE in complex real-world scenarios. RT-Pose is available at https://huggingface.co/datasets/uwipl/RT-Pose." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 160, + 573, + 453, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 573, + 453, + 584 + ], + "spans": [ + { + "bbox": [ + 160, + 573, + 453, + 584 + ], + "type": "text", + "content": "Keywords: Human Pose Estimation " + }, + { + "bbox": [ + 160, + 573, + 453, + 584 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 573, + 453, + 584 + ], + "type": "text", + "content": " 4D Radar " + }, + { + "bbox": [ + 160, + 573, + 453, + 584 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 573, + 453, + 584 + ], + "type": "text", + "content": " Dataset " + }, + { + "bbox": [ + 160, + 573, + 453, + 584 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 573, + 453, + 584 + ], + "type": "text", + "content": " Benchmark" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 604, + 230, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 604, + 230, + 617 + ], + "spans": [ + { + "bbox": [ + 132, + 604, + 230, + 617 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "Human localization and 3D pose estimation [6, 20-22, 31, 49, 53, 59] are indispensable in Augmented/Virtual Reality [5, 9, 15, 30], human-computer interaction [18, 51], healthcare [3, 8, 29]. However, capturing human poses across diverse" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "scenarios and generating comprehensive 3D representations for individuals in varying poses present significant challenges. Camera based motion capture systems are often used to collect human body posture [19,28,42]. However, these systems are sensitive to light and require careful multi-camera calibration, making them unreliable for outdoor scenarios and unable to provide overall pose estimation [7]. Furthermore, due to privacy concerns, deploying its application in home or long-term care center scenarios poses challenges." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 204, + 482, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 204, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 204, + 482, + 346 + ], + "type": "text", + "content": "Radar-based method emerges as a promising alternative, characterized by distinctive attributes such as through-wall recognition [54,58]. In addition, radar is robust to lighting conditions and resilient to various weather conditions and occlusion [1,47]. These characteristics make it especially suitable for safety- and privacy-critical applications [2,39]. In smart automotive applications, radar, as a complementary sensing modality, can complement adverse scenarios such as lowlighting environments and adverse weather [25], challenging RGB-based sensing. In healthcare applications, RGB-based Human Pose Estimation (HPE)poses privacy risks and is hindered by occlusion, prompting a demand for radar in indoor care environments. Considering the advantages and prospects of radar technology, it is indispensable to provide a dataset that contains diverse scenarios for fostering radar-based HPE research." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 350, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 529 + ], + "type": "text", + "content": "The previous radar-based HPE methods [14, 37, 48] often utilize Constant False Alarm Rate (CFAR) techniques [13, 45, 46] to extract radar point clouds. However, the effectiveness of CFAR-based point cloud extraction can be easily influenced by variations in radar module types and hardware parameters. Furthermore, different human body reflections in radar signals across diverse environments may require specific adjustments to CFAR parameters, making it challenging to generalize their applications. To address this issue, some works decompose the 4D radar tensor into vertical and horizontal directions as the input of the model to estimate human pose [26, 47, 50]. This approach computes the magnitude of radar signals for both directions, which effectively reduces data loss during the conversion to point cloud. However, a 4D radar tensor with Doppler values contains the velocity information, which is more informative than a decomposed radar tensor [10, 34]. Therefore, this work focuses on exploring the use of raw 4D radar tensors for HPE and establishing a benchmark based on this data format." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "The presented radar tensor-based human pose (RT-Pose) dataset is the first benchmark to integrate calibrated 4D radar tensors, RGB images, and LiDAR point clouds data in the field of HPE. Releasing a multi-modality dataset marks a significant advancement, particularly in the domain of human behavior analyses. There are a total of " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "72\\mathrm{k}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " frames in 240 sequences in the dataset. The actions are organized into sequences of increasing complexity, providing a realistic motion distribution. All experiments are conducted in 5 environmental conditions in 8 scenarios, which ensures the variety of scenarios for better comprehension of the model. Along with the release of this comprehensive RT-Pose dataset, we also build upon the High-Resolution Network (HRNet) [44] to propose HRRadarPose model, a robust baseline for 3D HPE utilizing 4D radar tensor data. Further" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "Y. Ho et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 168, + 479, + 264 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": "Table 1: Comparison of 3D human pose estimation datasets. The first group of rows shows RGB and LiDAR datasets. The second group of rows presents datasets with radar. The radar point cloud is denoted as RPC, and the 4D radar tensor is denoted as 4D-RT. * indicates a modified 4D radar tensor with a specific velocity range." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 168, + 479, + 264 + ], + "lines": [ + { + "bbox": [ + 134, + 168, + 479, + 264 + ], + "spans": [ + { + "bbox": [ + 134, + 168, + 479, + 264 + ], + "type": "table", + "html": "
DatasetRGB DepthLiDARRPC4D-RT#Scenes#Actions#Seqs#FPSTime (min)
Human3.6M [19]✓-15839501200
HSC4D [12]✓✓320-2042
SLOPER4D [11]✓✓10-152083
MARS [4]✓110801080
mRI [3]✓✓✓11230010264
HuPR [26]✓✓*-323510240
RT-Pose (Ours)✓✓✓40624010120
", + "image_path": "0ec224de6211e114349e765dc0a44163207ce2a345e31c3335dceeeb40fd315a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 285, + 482, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 370 + ], + "type": "text", + "content": "more, our study also includes a comparative evaluation of different radar data types within our proposed model, as well as benchmarking against previous radar-based methods. These results demonstrate the advantages of employing the 4D radar tensor method over traditional radar point cloud methods, especially in complex actions and diverse scenarios. The 4D radar tensor-based approach presents a challenging but promising direction for future research. In summary, our contributions include:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 376, + 480, + 483 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 376, + 480, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 480, + 411 + ], + "type": "text", + "content": "- The first dataset provides calibrated 4D radar tensors, LiDAR point clouds, and RGB images for human localization and 3D pose estimation of complex actions in diverse scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 412, + 480, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 412, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 137, + 412, + 480, + 446 + ], + "type": "text", + "content": "- We develop a reliable 3D human pose annotation workflow by joint optimization with LiDAR point cloud and RGB images for both outdoor and indoor scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 448, + 480, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 448, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 137, + 448, + 480, + 483 + ], + "type": "text", + "content": "- We propose HRRadarPose, the first single-stage human pose estimator specifically designed to learn 4D radar tensors' representation for the human pose estimation task." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 502, + 243, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 502, + 243, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 502, + 243, + 514 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 526, + 349, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 526, + 349, + 538 + ], + "spans": [ + { + "bbox": [ + 131, + 526, + 349, + 538 + ], + "type": "text", + "content": "2.1 3D Human Pose Estimation Datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "As one of the fundamental tasks in the computer vision field, there are lots of works introducing datasets and benchmarks for 3D HPE, as shown in Table 1. Human3.6M [19] is the first large-scale dataset for 3D human sensing in lab environments, which contains 3.6-million frames of corresponding 2D and 3D human poses from mocap captured videos of 5 female and 6 male subjects. Compared to Human3.6M, MPI-INF-3DHP [32] is a more challenging 3D human pose dataset captured in the wild. There are 8 subjects with 8 actions captured by 14 cameras covering a greater diversity of poses. 3DPW [41], the first one to include video footage taken from a moving phone camera, includes 60 video sequences. Recently, several works have explored the possibility of HPE from" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "RT-Pose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": "other modalities like radar and LiDAR. mmBody [7] consists of synchronized and calibrated mmWave radar point clouds and RGB-D images in different scenes, as well as skeleton/mesh annotations for humans in the scenes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 210 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 210 + ], + "type": "text", + "content": "mRI [3], a multi-modal 3D HPE dataset using mmWave, RGB-D, and inertial sensors, generates a radar point cloud dataset based on the CFAR method. There are 12 rehabilitation exercises selected to evaluate radar application in home-based health monitoring. HuPR [26] collects 235 sequences of data by calibrated radar and camera for 2D HPE in an indoor environment." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 212, + 481, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 481, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 481, + 307 + ], + "type": "text", + "content": "All open datasets for radar-based human pose estimation are collected within a limited detection area, which contradicts the characteristics of radar applications for tracking. Although detecting within a small area may ensure system accuracy, it limits the system's practical application. Furthermore, to the best of our knowledge, our dataset is the first to include radar, LiDAR, and RGB sequence information with 3D pose annotations. This allows our dataset to provide pose annotations for both indoor and outdoor environments, distinguishing it from existing datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 324, + 350, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 324, + 350, + 335 + ], + "spans": [ + { + "bbox": [ + 131, + 324, + 350, + 335 + ], + "type": "text", + "content": "2.2 Radar-based Human Pose Estimation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 342, + 479, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 479, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 479, + 402 + ], + "type": "text", + "content": "In spite of the popularity of RGB-based HPE applications, privacy concerns have driven the exploration of radio or radar modalities for HPE. RF-Pose [54] is a pioneering application in 2D HPE utilizing radio frequency signals. RF-Pose demonstrates the feasibility of pose detection in through-wall scenarios and provide the possibility of predicting the 3D pose using radio signals." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 403, + 480, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 480, + 521 + ], + "type": "text", + "content": "On the other hand, frequency-modulated continuous wave (FMCW) radar, which operates in the millimeter-wave band (30-300GHz), has been shown to be suitable for detection and HPE tasks. The radar point cloud is widely used to represent the radar signal in HPE [3,4,7,36,37,48,52]. mmMesh [48] introduces a temporal model structure that concurrently captures global and local 3D point cloud structures in spatial dimensions, ensuring precise localization performance for pose estimation. Nonetheless, in cluttered environments, the probability and intensity of radar signal reflections from the human body decrease. Consequently, implementing point cloud-based methods in real-world applications becomes challenging." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 522, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 481, + 665 + ], + "type": "text", + "content": "Compared to the radar point cloud methods, utilizing 4D tensor radar signal proves to be more informative and reliable [26, 47, 55, 56]. RF-pose 3D [55] is the first work using the 4D RF tensor to predict 3D skeletons in 22 different locations. However, the used RF system is not commercially available, which is hard to reproduce. HuPR [26] utilizes two pieces of well-calibrated TI mmWave radar modules to capture vertical and horizontal radar signals, enhancing the resolution of radar elevation signals to improve pose estimation performance. RPM2.0 [47] adopts a similar hardware setup to HuPR but utilizes the HRNet [44] as the backbone and includes an attention model to reconstruct missing keypoints. However, their two-stage model is trained by only walking pose, restricting the actions' complexity for achieving better pose estimation performance. The methods proposed in HuPR and RPM2.0 fuse information from both" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "Y. Ho et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 118, + 255, + 177 + ], + "blocks": [ + { + "bbox": [ + 134, + 118, + 255, + 177 + ], + "lines": [ + { + "bbox": [ + 134, + 118, + 255, + 177 + ], + "spans": [ + { + "bbox": [ + 134, + 118, + 255, + 177 + ], + "type": "image", + "image_path": "59fcc3666594ad74f304d77ff1b378a4d0237a8eb9c73cb9433fe295b6b66cca.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 180, + 198, + 189 + ], + "lines": [ + { + "bbox": [ + 187, + 180, + 198, + 189 + ], + "spans": [ + { + "bbox": [ + 187, + 180, + 198, + 189 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 261, + 118, + 376, + 177 + ], + "blocks": [ + { + "bbox": [ + 261, + 118, + 376, + 177 + ], + "lines": [ + { + "bbox": [ + 261, + 118, + 376, + 177 + ], + "spans": [ + { + "bbox": [ + 261, + 118, + 376, + 177 + ], + "type": "image", + "image_path": "0e1ac6a3978f2d7ec2db7f1789aca6815fdbbe986060c867eec575f47de5e33c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 180, + 323, + 189 + ], + "lines": [ + { + "bbox": [ + 313, + 180, + 323, + 189 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 323, + 189 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 198, + 480, + 220 + ], + "lines": [ + { + "bbox": [ + 131, + 198, + 480, + 220 + ], + "spans": [ + { + "bbox": [ + 131, + 198, + 480, + 220 + ], + "type": "text", + "content": "Fig. 2: Data distribution for RT-Pose dataset: (a) Activities; (b) Environmental conditions; (c) Occlusion conditions" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 380, + 118, + 473, + 177 + ], + "blocks": [ + { + "bbox": [ + 380, + 118, + 473, + 177 + ], + "lines": [ + { + "bbox": [ + 380, + 118, + 473, + 177 + ], + "spans": [ + { + "bbox": [ + 380, + 118, + 473, + 177 + ], + "type": "image", + "image_path": "0ec4447a1856385525a9377a028ada88d698cdbf90aeeecb2c6d03b953fff01c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 423, + 180, + 433, + 189 + ], + "lines": [ + { + "bbox": [ + 423, + 180, + 433, + 189 + ], + "spans": [ + { + "bbox": [ + 423, + 180, + 433, + 189 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 246, + 479, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 479, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 479, + 269 + ], + "type": "text", + "content": "radar modules. While the dual-radar-module configuration potentially captures fine-grained information, its intricate setup poses reproducibility challenges." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 270, + 480, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 480, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 480, + 340 + ], + "type": "text", + "content": "Unlike previous works, our system employs a single radar module for vertical and horizontal sensing, simplifying the sensors' calibration and synchronization. We use 4D radar tensor as training data for 3D HPE, reducing the likelihood of data loss during radar point generation. Moreover, RT-Pose dataset includes complex actions and is recorded in diverse scenes, expanding the applicability to real-world scenarios." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 363, + 254, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 254, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 254, + 374 + ], + "type": "text", + "content": "3 RT-Pose Dataset" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 390, + 298, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 298, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 298, + 593 + ], + "type": "text", + "content": "To construct the dataset, we recruit 10 participants to perform 6 types of actions. The experiments are divided into two categories. In the first category, participants are required to perform actions while standing, including waving, lifting legs, and random poses to increase the complexity of the actions. A video sequence of random poses involves a subject randomly performing one possible movement at a time, including stretching, bending, twisting, etc., each lasting 3 to 5 seconds. The second category focuses on walking, with additional actions added during the process, such as walking and waving or sitting down" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "text", + "content": "after walking. Each type of action is inherited from previous radar-based datasets [4,35,48]. Each sequence contains one action for 30 seconds, providing a variety of different difficulty levels. As a result, people can thoroughly evaluate their 3D HPE methods on a broad distribution of actions with diverse difficulties. Furthermore, we provide the dataset development kit to facilitate the usage and evaluation of future methods." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 305, + 415, + 481, + 536 + ], + "blocks": [ + { + "bbox": [ + 305, + 415, + 481, + 536 + ], + "lines": [ + { + "bbox": [ + 305, + 415, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 305, + 415, + 481, + 536 + ], + "type": "image", + "image_path": "2f883ef3f042c30e0dc2f69f3e32a699d4c014be7d09a91f0a9648b9c2f41510.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 543, + 480, + 564 + ], + "lines": [ + { + "bbox": [ + 304, + 543, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 480, + 564 + ], + "type": "text", + "content": "Fig.1: Experimental hardware setup in an indoor environment for data collection." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "RT-Pose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 200, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 200, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 200, + 128 + ], + "type": "text", + "content": "3.1 Sensors" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 485, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 485, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 485, + 319 + ], + "type": "text", + "content": "The data collection hardware system comprises two RGB cameras, a non-repetitive horizontal scanning LiDAR, and a cascade imaging radar module, as shown in Figure 1. To extract human pose information from radar signals, the radar module's parameters have been specifically configured. The radar operates at ten frames per second in this work. The radar module is equipped with 12 transmit (TX) and 16 receive (RX) antenna elements and operates in a multiple-input multiple-output (MIMO) mode [23]. This radar module provides the azimuth and elevation angle resolutions of 1.4 degrees and 18 degrees, respectively. Furthermore, this work implements a high-frequency slope with a long ramp time to efficiently utilize the hardware's sweep bandwidth, ranging from 77 GHz to almost 81 GHz. The longer sweep bandwidth of a chirp increases the radar range resolution [24,37], which is " + }, + { + "bbox": [ + 130, + 138, + 485, + 319 + ], + "type": "inline_equation", + "content": "4.5\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 138, + 485, + 319 + ], + "type": "text", + "content": " in this work. Each frame consists of 64 chirps used for calculating radar Doppler signals, providing a velocity resolution of 3.9 cm/s. These carefully chosen configurations ensure the system's accuracy and reliability in capturing human poses." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 338, + 240, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 338, + 240, + 350 + ], + "spans": [ + { + "bbox": [ + 132, + 338, + 240, + 350 + ], + "type": "text", + "content": "3.2 Data Collection" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 360, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 482, + 468 + ], + "type": "text", + "content": "We collect the dataset in 40 scenes with indoor and outdoor environments. Figure 3 shows some of the collected data. Indoor environments include clean and cluttered conditions, while outdoor ones include normal, rainy weather, and lowlighting conditions. Each experimental condition includes 8 scenarios, which are nonocclusive, and three occlusive conditions (cardboard, cloth, and plastic pad). Each nonocclusive and occlusive condition is recorded in single- and multi-person settings. To consider the different fields of view (FoV) of all sensors, the sensor suite is situated " + }, + { + "bbox": [ + 130, + 360, + 482, + 468 + ], + "type": "inline_equation", + "content": "109~\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 360, + 482, + 468 + ], + "type": "text", + "content": " above ground, and an appropriate detection area is planned based on different scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 487, + 244, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 244, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 244, + 500 + ], + "type": "text", + "content": "3.3 Data Processing" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "The captured radar signal, referred to as 4D tensor data, is processed by a series of steps [16,33,45], as shown in Figure 4. On the transmitter side, the frequency of the transmitted signals, or chirps, periodically increases based on the set frequency slope. Consequently, the received signal exhibits a different frequency than the transmitted signal. The frequency difference between transmitted and received signals can be used to estimate the range of the object reflecting the signal to the radar sensor by the Fast Fourier Transform (FFT). At the beginning of the data processing, each radar frame consists of 64 chirps, resulting in 64 range estimation results. Utilizing this range information, the velocity of the responding object can be measured through FFT, a phenomenon known as the Doppler effect. Second, the radar signal data is re-modulated according to the antenna position. This step is essential for the calculation of the angle of arrival (AoA), enabling the generation of highangular resolution results. Third, the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "Y. Ho et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 157, + 115, + 452, + 265 + ], + "blocks": [ + { + "bbox": [ + 157, + 115, + 452, + 265 + ], + "lines": [ + { + "bbox": [ + 157, + 115, + 452, + 265 + ], + "spans": [ + { + "bbox": [ + 157, + 115, + 452, + 265 + ], + "type": "image", + "image_path": "79ce2e9e719cb60a92a3492f1dbbf59853ad293e1a7bf2e224873558e2680697.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 276, + 479, + 298 + ], + "lines": [ + { + "bbox": [ + 131, + 276, + 479, + 298 + ], + "spans": [ + { + "bbox": [ + 131, + 276, + 479, + 298 + ], + "type": "text", + "content": "Fig. 3: Experimental instances across various indoor and outdoor conditions with diverse scenarios for data collection." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 182, + 312, + 430, + 418 + ], + "blocks": [ + { + "bbox": [ + 182, + 312, + 430, + 418 + ], + "lines": [ + { + "bbox": [ + 182, + 312, + 430, + 418 + ], + "spans": [ + { + "bbox": [ + 182, + 312, + 430, + 418 + ], + "type": "image", + "image_path": "b5b68f0127ea85644d9a4d36821f384d89bc2fa95a105c617582ad510eae9a31.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 432, + 479, + 455 + ], + "lines": [ + { + "bbox": [ + 131, + 432, + 479, + 455 + ], + "spans": [ + { + "bbox": [ + 131, + 432, + 479, + 455 + ], + "type": "text", + "content": "Fig. 4: Radar signal processing flow. The green arrow line is for radar point cloud generation and the blue line is for 4D radar tensor generation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 475, + 479, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 479, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 479, + 548 + ], + "type": "text", + "content": "radar signal from the azimuth and elevation directions also undergoes FFT processing to represent velocity responses in different directions. Finally, the entire radar data is transformed from polar coordinates to Cartesian coordinates. This transformation enhances the intuitiveness for the subsequent pose estimation. In this study, the processed 4D radar tensor has the dimension of " + }, + { + "bbox": [ + 130, + 475, + 479, + 548 + ], + "type": "inline_equation", + "content": "64 \\times 32 \\times 128 \\times 256" + }, + { + "bbox": [ + 130, + 475, + 479, + 548 + ], + "type": "text", + "content": ", which correspond to velocity, the " + }, + { + "bbox": [ + 130, + 475, + 479, + 548 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 475, + 479, + 548 + ], + "type": "text", + "content": "-axis, the " + }, + { + "bbox": [ + 130, + 475, + 479, + 548 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 475, + 479, + 548 + ], + "type": "text", + "content": "-axis, and the " + }, + { + "bbox": [ + 130, + 475, + 479, + 548 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 475, + 479, + 548 + ], + "type": "text", + "content": "-axis, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 563, + 269, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 563, + 269, + 574 + ], + "spans": [ + { + "bbox": [ + 131, + 563, + 269, + 574 + ], + "type": "text", + "content": "3.4 Annotation Workflow" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 480, + 665 + ], + "type": "text", + "content": "To achieve accurate human subject detection in a larger area with high-quality ground truth, our system integrates LiDAR and RGB camera data, as illustrated in Figure 5. The images provided by the camera are processed using the HRNet [38] to extract 2D human poses. Then, a 3D HPE model, called ZeDO [22], utilizes the 2D poses to estimate the 3D poses as the initial 3D pseudo ground truth. However, estimating 3D poses using a monocular camera often lacks precision and stability in depth. Therefore, this work incorporates LiDAR for depth" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "RT-Pose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 178, + 113, + 438, + 244 + ], + "blocks": [ + { + "bbox": [ + 178, + 113, + 438, + 244 + ], + "lines": [ + { + "bbox": [ + 178, + 113, + 438, + 244 + ], + "spans": [ + { + "bbox": [ + 178, + 113, + 438, + 244 + ], + "type": "image", + "image_path": "87050120d069cfb845b98cde6c52d02116fcbac86c8fe86a07a954642d1abdbe.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 253, + 482, + 299 + ], + "lines": [ + { + "bbox": [ + 131, + 253, + 482, + 299 + ], + "spans": [ + { + "bbox": [ + 131, + 253, + 482, + 299 + ], + "type": "text", + "content": "Fig. 5: Workflow of human localization and 3D pose ground truth annotations. Estimated 2D pose results, predicted by the pre-trained HRNet model, are denoted as " + }, + { + "bbox": [ + 131, + 253, + 482, + 299 + ], + "type": "inline_equation", + "content": "P_{2d}" + }, + { + "bbox": [ + 131, + 253, + 482, + 299 + ], + "type": "text", + "content": ". The initial setting pose derived from LiDAR point clouds is denoted as " + }, + { + "bbox": [ + 131, + 253, + 482, + 299 + ], + "type": "inline_equation", + "content": "P_{init}" + }, + { + "bbox": [ + 131, + 253, + 482, + 299 + ], + "type": "text", + "content": ". Both " + }, + { + "bbox": [ + 131, + 253, + 482, + 299 + ], + "type": "inline_equation", + "content": "P_{2d}" + }, + { + "bbox": [ + 131, + 253, + 482, + 299 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 253, + 482, + 299 + ], + "type": "inline_equation", + "content": "P_{init}" + }, + { + "bbox": [ + 131, + 253, + 482, + 299 + ], + "type": "text", + "content": " are inputs into ZeDO, an optimization-based pipeline for 3D pose estimation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 322, + 480, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 480, + 369 + ], + "type": "text", + "content": "estimation of the human subject center. The LiDAR sensor captures environmental information and the subject's point cloud data, ensuring errors are under " + }, + { + "bbox": [ + 130, + 322, + 480, + 369 + ], + "type": "inline_equation", + "content": "2\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 322, + 480, + 369 + ], + "type": "text", + "content": " inside the detection range of " + }, + { + "bbox": [ + 130, + 322, + 480, + 369 + ], + "type": "inline_equation", + "content": "20\\mathrm{m}" + }, + { + "bbox": [ + 130, + 322, + 480, + 369 + ], + "type": "text", + "content": ". The human subject is manually labeled with a 3D bounding box, which can provide an accurate 3D position." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 369, + 482, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 549 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 549 + ], + "type": "text", + "content": "ZeDO, a state-of-the-art (SOTA) method for cross-dataset HPE, implements an optimization-based pipeline for 3D pose estimation using 2D poses. This method requires 2D poses as input, and ZeDO is able to estimate multiple-hypothesis 3D poses. In this research, the center of the 3D bounding box provided by LiDAR is utilized as the pelvis of the initial pose, which enhances the accuracy of depth estimation compared to relying solely on the monocular camera data, and improves the performance of ZeDO. With the imported 2D pose from HRNet [43], ZeDO iteratively refines the 3D poses. To further improve annotation accuracy, an annotation tool is developed for manual selection and optimization of each frame to retain correct 3D human poses. The 3D HPE results from ZeDO will be put in the LiDAR point cloud for visualization, which is beneficial for humans to filter and modify the error pose annotation results. About " + }, + { + "bbox": [ + 130, + 369, + 482, + 549 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 369, + 482, + 549 + ], + "type": "text", + "content": " of the data is removed or optimized after human correction, and this process ensures the quality of keypoints annotation. More details can be found in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 567, + 240, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 240, + 580 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 240, + 580 + ], + "type": "text", + "content": "4 HRRadarPose" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Inspired by HRNet [43], renowned for its adeptness in extracting high-resolution representations, we build an architecture, HRRadarPose, with fully 3D convolutional layers. By treating the Doppler axis, " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", as the input channels, our model extracts volumetric spatial features along the " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " axes, representing vertical, horizontal, and depth directions in the Cartesian coordinate system. The HRRadarPose architecture is designed to preserve spatial resolution and assimi" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "Y. Ho et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "text", + "content": "late semantic information from the 4D radar tensors, maintaining high-resolution representations and enabling the exchange of information between features of different resolutions during stage transitions. Figure 6 illustrates an instance of the HRRadarPose structure with three stages. The structure starts with using 3D convolutional layers to transform the 4D radar tensors, " + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "inline_equation", + "content": "T \\in \\mathbb{R}^{D \\times Z \\times Y \\times X}" + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "text", + "content": ", into features, " + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "inline_equation", + "content": "f_{in} \\in \\mathbb{R}^{C_1 \\times Z_1 \\times Y_1 \\times X_1}" + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "text", + "content": ". As the network proceeds from one stage to the next, it develops additional branches through strided convolution to widen the receptive field. Each branch " + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "text", + "content": " employs strided convolution to elevate its input feature " + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "text", + "content": " to a feature " + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "inline_equation", + "content": "f_{i+1}" + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "text", + "content": " in the succeeding branch " + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "inline_equation", + "content": "i+1" + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "text", + "content": ", simultaneously doubling the channels and halving the spatial dimensions, " + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "inline_equation", + "content": "f_{i+1} \\in \\mathbb{R}^{2C_i \\times \\frac{Z_i}{2} \\times \\frac{Y_i}{2} \\times \\frac{X_i}{2}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "text", + "content": ". Within each stage, a series of " + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "text", + "content": " modules execute " + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "text", + "content": " parallel convolutions. To align feature dimensions from different branches for information exchange, we use strided convolution followed by upsampling. At the end of the backbone, we take the feature with the highest resolution to serve as a unified representation for the pose estimation head. In the pose estimation head, two 3D convolutional branches are employed to process these features further. This bifurcation generates two outputs: a distribution map signifying the confidence of locating the human center and the location offset of joint keypoints. The right part of Figure 6 shows the decoding of the HRRadarPose network's output to multi-person poses. The process starts with picking top- " + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "text", + "content": " human centers, " + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "inline_equation", + "content": "C_S" + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "text", + "content": ", with confidence scores over a threshold value. We denote these picked centers' indices as " + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "text", + "content": ". Then, we query the centers corresponding joint keypoints' offsets, " + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "inline_equation", + "content": "K_S" + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "text", + "content": ", by looking up the output map of keypoint offsets head at " + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "text", + "content": ". Lastly, we decode their joint keypoint locations, " + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "inline_equation", + "content": "J_S" + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "text", + "content": ", by summing " + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "inline_equation", + "content": "K_S" + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "inline_equation", + "content": "C_S" + }, + { + "bbox": [ + 130, + 115, + 482, + 478 + ], + "type": "text", + "content": ". In scenarios with multiple people, we apply range non-maximum suppression (NMS) to reduce overlapping detections. Our method is favorable in two folds: The HRRadarPose's architecture gains efficiency through a single-stage workflow, avoiding complexity brought by region-proposal-based methods [55]. In addition, our pose estimation head inherently learns per person's joint keypoints refraining from ambiguity of association between people's identity and joints." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 481, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 482, + 506 + ], + "type": "text", + "content": "To train the center probability head, we use the pixel-wise focal loss as the classification loss, " + }, + { + "bbox": [ + 130, + 481, + 482, + 506 + ], + "type": "inline_equation", + "content": "L_{\\text{class}}" + }, + { + "bbox": [ + 130, + 481, + 482, + 506 + ], + "type": "text", + "content": ", defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 181, + 522, + 481, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 522, + 481, + 554 + ], + "spans": [ + { + "bbox": [ + 181, + 522, + 481, + 554 + ], + "type": "interline_equation", + "content": "L _ {\\text {c l a s s}} = \\left\\{ \\begin{array}{l l} (1 - P _ {x y z}) ^ {\\alpha} \\log \\left(P _ {x y z}\\right), & \\text {i f} Y _ {x y z} = 1, \\\\ (1 - Y _ {x y z}) ^ {\\beta} \\left(P _ {x y z}\\right) ^ {\\alpha} \\log \\left(1 - P _ {x y z}\\right), & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {1}", + "image_path": "a86b92e22e1dbe2166367a3113c02da98c19561a832373c59b78d81d50bae69b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "Y_{xyz}" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " is the three-dimensional human center Gaussian distribution map generated by pelvis locations, " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "P_{xyz}" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " is the probability of a person's presence at location, " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " is the number of foreground samples, " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " are hyperparameters addressing class imbalance and focusing training on foreground samples. To train the keypoint offset head, we construct the regression loss " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "L_{reg}" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " by taking the average of 15 keypoints' " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": ", comparing the distance between predicted offsets and ground-truth offsets to the center location of a voxel in " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "Y_{xyz}" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " equals to 1. The final loss is a weighted sum of " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "L_{class}" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "L_{reg}" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "RT-Pose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 121, + 474, + 241 + ], + "blocks": [ + { + "bbox": [ + 140, + 121, + 474, + 241 + ], + "lines": [ + { + "bbox": [ + 140, + 121, + 474, + 241 + ], + "spans": [ + { + "bbox": [ + 140, + 121, + 474, + 241 + ], + "type": "image", + "image_path": "241762dcba3e0b1f7538100561e8984227e37a1aa699bdb6d14b9d31a7694fbd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 251, + 482, + 297 + ], + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 297 + ], + "type": "text", + "content": "Fig. 6: HRRadarPose's architecture consists of a 3D convolutional backbone and a pose estimation head, which generates a body center's confidence map and joint keypoint offsets map. To infer multi-person poses, we decode the prediction by adding the picked bodies' centers with their corresponding keypoint offsets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 318, + 230, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 318, + 230, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 318, + 230, + 332 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 342, + 479, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 479, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 479, + 403 + ], + "type": "text", + "content": "15 obvious keypoints of a human body are selected, aligning with the body skeleton model of the Human3.6M dataset [19]. Each frame's 3D pose annotation in the sequence is manually validated to ensure the quality of the dataset. Figure 7 demonstrates the testing results across various levels of action complexity, presenting the feasibility of 3D pose estimation using the proposed baseline." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 415, + 480, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 480, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 480, + 486 + ], + "type": "text", + "content": "Evaluation metrics We employ joint position error (JPE), which measures the mean of the Euclidean distance for every joint's predicted keypoints and the ground truth keypoints. Localization results are reported by the mean of the root position error (MRPE) [42], which is the JPE of the pelvis keypoint. For the evaluation of 3D HPE, we use Mean per joint position error (MPJPE) [57] and absolute-MPJPE (Abs-MPJPE) [40]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Baselines Comparison We evaluate the performance of the HRRadarPose in global localization and 3D HPE on the RT-Pose dataset. We compare our method with three baseline methods, as shown in Table 2. mm-Pose [37] and mmMesh [48], utilizing radar point clouds for 3D HPE, are re-implemented to test on the RT-Pose dataset. Notably, mmMesh introduces a global localization model, combined with an RNN HPE module, showing more accurate estimation results in both MRPE and MPJPE compared to mm-Pose, which employs a simpler CNN model. RF-Pose 3D [55] uses the 4D radar tensor with a region proposal network to locate the subjects and estimate the 3D pose. The experimental results show that our HRRadarPose outperforms other baseline methods. On the RT-Pose dataset, we obtain " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "9.93\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " in MPJPE and " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "9.91\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " in MRPE. We find that radar-tensor-based methods reach less error than the methods using radar point cloud. We attribute this performance gap to complicated scenarios since it is difficult for point cloud-based methods to discriminate different key" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "Y. Ho et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 166, + 118, + 211, + 170 + ], + "blocks": [ + { + "bbox": [ + 152, + 136, + 162, + 156 + ], + "lines": [ + { + "bbox": [ + 152, + 136, + 162, + 156 + ], + "spans": [ + { + "bbox": [ + 152, + 136, + 162, + 156 + ], + "type": "text", + "content": "R" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 166, + 118, + 211, + 170 + ], + "lines": [ + { + "bbox": [ + 166, + 118, + 211, + 170 + ], + "spans": [ + { + "bbox": [ + 166, + 118, + 211, + 170 + ], + "type": "image", + "image_path": "63e42322a68c127528fe73d396d6a271b5ffbabd9b968ec1691aa24153cbd644.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 216, + 118, + 261, + 170 + ], + "blocks": [ + { + "bbox": [ + 216, + 118, + 261, + 170 + ], + "lines": [ + { + "bbox": [ + 216, + 118, + 261, + 170 + ], + "spans": [ + { + "bbox": [ + 216, + 118, + 261, + 170 + ], + "type": "image", + "image_path": "7d52982fa3bf129fca7a2fa5286c7ff1af4f8280866da17da68d75ef34e15992.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 263, + 118, + 309, + 170 + ], + "blocks": [ + { + "bbox": [ + 263, + 118, + 309, + 170 + ], + "lines": [ + { + "bbox": [ + 263, + 118, + 309, + 170 + ], + "spans": [ + { + "bbox": [ + 263, + 118, + 309, + 170 + ], + "type": "image", + "image_path": "2df9a21432c8938e5aef8da4b312ee6440a1f79b157f2f32da8ba764b43577fc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 313, + 118, + 359, + 170 + ], + "blocks": [ + { + "bbox": [ + 313, + 118, + 359, + 170 + ], + "lines": [ + { + "bbox": [ + 313, + 118, + 359, + 170 + ], + "spans": [ + { + "bbox": [ + 313, + 118, + 359, + 170 + ], + "type": "image", + "image_path": "e96abddca39f5299a3c42fa75e5072282cbb909d8eefc2b0302c63bf15aada12.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 361, + 118, + 406, + 170 + ], + "blocks": [ + { + "bbox": [ + 361, + 118, + 406, + 170 + ], + "lines": [ + { + "bbox": [ + 361, + 118, + 406, + 170 + ], + "spans": [ + { + "bbox": [ + 361, + 118, + 406, + 170 + ], + "type": "image", + "image_path": "d96b9207b858c26dd2461d2efdc355e75335f2bb5662bff4bc4f630d3497c6f4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 410, + 118, + 455, + 170 + ], + "blocks": [ + { + "bbox": [ + 410, + 118, + 455, + 170 + ], + "lines": [ + { + "bbox": [ + 410, + 118, + 455, + 170 + ], + "spans": [ + { + "bbox": [ + 410, + 118, + 455, + 170 + ], + "type": "image", + "image_path": "303c8131a5567b639ec9976f127c0c0116d037e7a56ca71cda437c23a808f572.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 166, + 171, + 211, + 217 + ], + "blocks": [ + { + "bbox": [ + 153, + 173, + 163, + 213 + ], + "lines": [ + { + "bbox": [ + 153, + 173, + 163, + 213 + ], + "spans": [ + { + "bbox": [ + 153, + 173, + 163, + 213 + ], + "type": "text", + "content": "BREV TraJ." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 166, + 171, + 211, + 217 + ], + "lines": [ + { + "bbox": [ + 166, + 171, + 211, + 217 + ], + "spans": [ + { + "bbox": [ + 166, + 171, + 211, + 217 + ], + "type": "image", + "image_path": "478994c069a6abb065e22d4cca11396c331995ef319bc31b121c3136b15e6980.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 216, + 171, + 261, + 217 + ], + "blocks": [ + { + "bbox": [ + 216, + 171, + 261, + 217 + ], + "lines": [ + { + "bbox": [ + 216, + 171, + 261, + 217 + ], + "spans": [ + { + "bbox": [ + 216, + 171, + 261, + 217 + ], + "type": "image", + "image_path": "88098aad8f31ccec0135875708e0825fce8a50307c4210204a3d4332be7b5c08.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 263, + 171, + 309, + 217 + ], + "blocks": [ + { + "bbox": [ + 263, + 171, + 309, + 217 + ], + "lines": [ + { + "bbox": [ + 263, + 171, + 309, + 217 + ], + "spans": [ + { + "bbox": [ + 263, + 171, + 309, + 217 + ], + "type": "image", + "image_path": "5215f3e12d9cf52b1f1348b838588bdc51682ed34052fc434df41fae5f645614.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 313, + 171, + 359, + 217 + ], + "blocks": [ + { + "bbox": [ + 313, + 171, + 359, + 217 + ], + "lines": [ + { + "bbox": [ + 313, + 171, + 359, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 359, + 217 + ], + "type": "image", + "image_path": "62839eadb20b9da3e78537cdedf9cadd5696da9cc712da1ed23b2c536eb906d5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 361, + 171, + 407, + 217 + ], + "blocks": [ + { + "bbox": [ + 361, + 171, + 407, + 217 + ], + "lines": [ + { + "bbox": [ + 361, + 171, + 407, + 217 + ], + "spans": [ + { + "bbox": [ + 361, + 171, + 407, + 217 + ], + "type": "image", + "image_path": "5d0809a9dac003984daf6643138040c398a899454bd15047d43cac3062305cb2.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 410, + 171, + 455, + 217 + ], + "blocks": [ + { + "bbox": [ + 410, + 171, + 455, + 217 + ], + "lines": [ + { + "bbox": [ + 410, + 171, + 455, + 217 + ], + "spans": [ + { + "bbox": [ + 410, + 171, + 455, + 217 + ], + "type": "image", + "image_path": "b74a7d4e620d40b328edd18013fc775ff7b08bd5ec974a30582134958a1e3aab.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 165, + 220, + 215, + 263 + ], + "blocks": [ + { + "bbox": [ + 153, + 231, + 163, + 262 + ], + "lines": [ + { + "bbox": [ + 153, + 231, + 163, + 262 + ], + "spans": [ + { + "bbox": [ + 153, + 231, + 163, + 262 + ], + "type": "text", + "content": "3D pose" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 165, + 220, + 215, + 263 + ], + "lines": [ + { + "bbox": [ + 165, + 220, + 215, + 263 + ], + "spans": [ + { + "bbox": [ + 165, + 220, + 215, + 263 + ], + "type": "image", + "image_path": "a8605e18fe78ab26bfb98f21921c76915d3ccaab1bdbbc4ca6dacac130e8e22a.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 166, + 264, + 210, + 272 + ], + "lines": [ + { + "bbox": [ + 166, + 264, + 210, + 272 + ], + "spans": [ + { + "bbox": [ + 166, + 264, + 210, + 272 + ], + "type": "text", + "content": "(a) Waving hand" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 216, + 220, + 263, + 263 + ], + "blocks": [ + { + "bbox": [ + 216, + 220, + 263, + 263 + ], + "lines": [ + { + "bbox": [ + 216, + 220, + 263, + 263 + ], + "spans": [ + { + "bbox": [ + 216, + 220, + 263, + 263 + ], + "type": "image", + "image_path": "f8651a21d82d7697fce8e273b7630003399a61ac706632f0cf61e8988e0737d9.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 264, + 257, + 272 + ], + "lines": [ + { + "bbox": [ + 220, + 264, + 257, + 272 + ], + "spans": [ + { + "bbox": [ + 220, + 264, + 257, + 272 + ], + "type": "text", + "content": "(b) Lifting leg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 263, + 220, + 314, + 263 + ], + "blocks": [ + { + "bbox": [ + 263, + 220, + 314, + 263 + ], + "lines": [ + { + "bbox": [ + 263, + 220, + 314, + 263 + ], + "spans": [ + { + "bbox": [ + 263, + 220, + 314, + 263 + ], + "type": "image", + "image_path": "68e55af94ad4b75044313768531c00e7f94ff9e04f9f62878bbc251bdd138fd3.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 266, + 264, + 310, + 272 + ], + "lines": [ + { + "bbox": [ + 266, + 264, + 310, + 272 + ], + "spans": [ + { + "bbox": [ + 266, + 264, + 310, + 272 + ], + "type": "text", + "content": "(c) Random pose" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 315, + 220, + 362, + 263 + ], + "blocks": [ + { + "bbox": [ + 315, + 220, + 362, + 263 + ], + "lines": [ + { + "bbox": [ + 315, + 220, + 362, + 263 + ], + "spans": [ + { + "bbox": [ + 315, + 220, + 362, + 263 + ], + "type": "image", + "image_path": "8907c83bf211cf42fe61c003fb8fd1b2a42a11f85eebdcd34228c2c413b50112.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 264, + 351, + 272 + ], + "lines": [ + { + "bbox": [ + 322, + 264, + 351, + 272 + ], + "spans": [ + { + "bbox": [ + 322, + 264, + 351, + 272 + ], + "type": "text", + "content": "(d) Normal" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 362, + 220, + 409, + 263 + ], + "blocks": [ + { + "bbox": [ + 362, + 220, + 409, + 263 + ], + "lines": [ + { + "bbox": [ + 362, + 220, + 409, + 263 + ], + "spans": [ + { + "bbox": [ + 362, + 220, + 409, + 263 + ], + "type": "image", + "image_path": "a76f2b5b3fd3dd1b7acfc221eb0f91d28f02ed4db7ad0f5044627352c12ba6a7.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 264, + 396, + 272 + ], + "lines": [ + { + "bbox": [ + 369, + 264, + 396, + 272 + ], + "spans": [ + { + "bbox": [ + 369, + 264, + 396, + 272 + ], + "type": "text", + "content": "(e) Sitting" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 410, + 220, + 462, + 263 + ], + "blocks": [ + { + "bbox": [ + 410, + 220, + 462, + 263 + ], + "lines": [ + { + "bbox": [ + 410, + 220, + 462, + 263 + ], + "spans": [ + { + "bbox": [ + 410, + 220, + 462, + 263 + ], + "type": "image", + "image_path": "3ae9f7e212ad9f93384faa03c49d95912d49b683c2978e816088973c08b0d12b.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 411, + 264, + 454, + 272 + ], + "lines": [ + { + "bbox": [ + 411, + 264, + 454, + 272 + ], + "spans": [ + { + "bbox": [ + 411, + 264, + 454, + 272 + ], + "type": "text", + "content": "(f) Waving hand" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 166, + 275, + 460, + 289 + ], + "blocks": [ + { + "bbox": [ + 166, + 275, + 460, + 289 + ], + "lines": [ + { + "bbox": [ + 166, + 275, + 460, + 289 + ], + "spans": [ + { + "bbox": [ + 166, + 275, + 460, + 289 + ], + "type": "image", + "image_path": "2c99d341c382c9eee4bf5a84506dff4d830d506f3db61d65e9124138a8557563.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 301, + 480, + 345 + ], + "lines": [ + { + "bbox": [ + 130, + 301, + 480, + 345 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 480, + 345 + ], + "type": "text", + "content": "Fig. 7: Visualization of pose estimation results in test frames for six activities. The second row is the trajectory of the test sequence in BEV. The third row is the instance of the global 3D pose estimated result for each action. In both rows, the blue line is the ground truth, and the red line represents estimated results." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "table", + "bbox": [ + 186, + 400, + 425, + 464 + ], + "blocks": [ + { + "bbox": [ + 130, + 357, + 480, + 389 + ], + "lines": [ + { + "bbox": [ + 130, + 357, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 480, + 389 + ], + "type": "text", + "content": "Table 2: Comparison of human localization and 3D pose estimation results of our HRRadarPose method and three baseline methods. † represents the radar point cloud-based methods." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 186, + 400, + 425, + 464 + ], + "lines": [ + { + "bbox": [ + 186, + 400, + 425, + 464 + ], + "spans": [ + { + "bbox": [ + 186, + 400, + 425, + 464 + ], + "type": "table", + "html": "
MethodMRPE (cm)MPJPE (cm)Abs-MPJPE (cm)
mm-Pose † [37]102.2821.26102.7
mmMesh † [48]62.6913.7866.24
RF-Pose3D [55]29.1718.4338.05
HRRadarPose (Ours)9.919.9314.73
", + "image_path": "9bb9b70ad09e3661ad372346bd32d7d5f05ae71ceed48f49162b0fc60b165b5e.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "table_body" + } + ], + "index": 32 + }, + { + "bbox": [ + 130, + 486, + 480, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 480, + 534 + ], + "type": "text", + "content": "points in various poses. The results highlight the importance of using 4D radar tensors, which preserve raw and rich spatial-temporal information for the HPE tasks. An end-to-end architecture directly takes in 4D radar tensors to better face different human action scenarios in the real world." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 130, + 545, + 480, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 480, + 640 + ], + "type": "text", + "content": "Action complexity Analysis RT-posedataset provides six difficulty levels of actions to simulate real-world scenarios, enabling comprehensive evaluation of localization and 3D HPE results. We report the performance of the proposed HRRadarPose method in Table 3. Generally, the joints along the human longitudinal axis, such as the thorax and the head, are more stable than those in the limbs. This phenomenon may be attributed to the larger reflective surface of the body's trunk compared to the limbs, providing a more reliable HPE based on radar signals." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "The MRPE of standing posture is below " + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "inline_equation", + "content": "7.67~\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": ", indicating that the pose complexity slightly influences the localization accuracy. Due to the chair's strong" + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "RT-Pose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 179, + 479, + 276 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 3: Comparison of the performance predicted by the HRRadar Pose in different complexity level activities, presenting different key points of the pose, the results of localization, and total 3D pose estimation results. The first group of rows lists the stand-related poses, while the other lists the walking-related poses. The complexity level increases from the top activity to the bottom within each group." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 179, + 479, + 276 + ], + "lines": [ + { + "bbox": [ + 135, + 179, + 479, + 276 + ], + "spans": [ + { + "bbox": [ + 135, + 179, + 479, + 276 + ], + "type": "table", + "html": "
ActionsJPE(cm)MRPE (cm)MPJPE (cm)Abs-MPJPE (cm)
ThoraxHeadAnkleWrist
Stand: Lifting LegWaving Hand2.953.973.6125.653.356.917.97
6.127.9212.248.887.557.0610.37
Random Pose7.7811.2111.7532.127.6712.4715.12
Walk: SitingNormal4.316.6913.5215.9813.678.9316.92
9.3613.0313.9310.7714.119.4416.83
Waving Hand5.7410.3713.5343.1612.8215.2821.22
", + "image_path": "30e002e62e7310ed3bf1d32b0e1f329226939114cee2691e97a0ceb743b6a19a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 171, + 300, + 444, + 454 + ], + "blocks": [ + { + "bbox": [ + 171, + 300, + 444, + 454 + ], + "lines": [ + { + "bbox": [ + 171, + 300, + 444, + 454 + ], + "spans": [ + { + "bbox": [ + 171, + 300, + 444, + 454 + ], + "type": "image", + "image_path": "0e19c1e2e75c5fecfb092a70f8b0919df69636e28e70b75392881e2db9134d30.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 462, + 482, + 485 + ], + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 485 + ], + "type": "text", + "content": "Fig. 8: HRRadarPose Results in different conditions. (a), (b), and (c) are occluded by cardboard, cloth, and plastic pad, respectively. (d) is in low-lighting conditions" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 520, + 482, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 520, + 482, + 663 + ], + "spans": [ + { + "bbox": [ + 130, + 520, + 482, + 663 + ], + "type": "text", + "content": "radar reflection, distinguishing between human subjects and the chair poses a challenge for localization. The MPJPE of walking and sitting is " + }, + { + "bbox": [ + 130, + 520, + 482, + 663 + ], + "type": "inline_equation", + "content": "9.44\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 520, + 482, + 663 + ], + "type": "text", + "content": ", showing our HRRadarPose model can reliably estimate the sitting pose. However, the MRPE of this action sequence achieves " + }, + { + "bbox": [ + 130, + 520, + 482, + 663 + ], + "type": "inline_equation", + "content": "14.11\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 520, + 482, + 663 + ], + "type": "text", + "content": ", demonstrating that sitting is a key factor for evaluating the performance of localization. The MPJPE of simple actions, such as standing and waving, standing and lifting legs, are around 7 cm, and the walking is " + }, + { + "bbox": [ + 130, + 520, + 482, + 663 + ], + "type": "inline_equation", + "content": "8.93\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 520, + 482, + 663 + ], + "type": "text", + "content": ", demonstrating that our HRRadarPose model is comparable to other recently published radar pose estimate methods [26, 47]. However, the current model still encounters difficulty in estimating HPE in complex actions such as walking and waving. Therefore, the RT-pose dataset is a challenging benchmark, which can enable HPE models to realistically learn diverse pose scenarios." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "Y. Ho et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 142, + 147, + 290, + 186 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 301, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 301, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 301, + 137 + ], + "type": "text", + "content": "Table 4: HPE results from HRRadarPose in normal and occlusive scenarios." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 142, + 147, + 290, + 186 + ], + "lines": [ + { + "bbox": [ + 142, + 147, + 290, + 186 + ], + "spans": [ + { + "bbox": [ + 142, + 147, + 290, + 186 + ], + "type": "table", + "html": "
OcclusionMPJPE (cm)Abs-MPJPE (cm)
✓9.8414.52
11.5718.55
", + "image_path": "218ba317d4367ec598643fb6c116508b3bcd4d5c6e878c9695af8cb7bfb0a1f6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 321, + 146, + 468, + 186 + ], + "blocks": [ + { + "bbox": [ + 312, + 114, + 480, + 136 + ], + "lines": [ + { + "bbox": [ + 312, + 114, + 480, + 136 + ], + "spans": [ + { + "bbox": [ + 312, + 114, + 480, + 136 + ], + "type": "text", + "content": "Table 5: Ablation Study of Using Dopper information." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 146, + 468, + 186 + ], + "lines": [ + { + "bbox": [ + 321, + 146, + 468, + 186 + ], + "spans": [ + { + "bbox": [ + 321, + 146, + 468, + 186 + ], + "type": "table", + "html": "
Doppler|MPJPE (cm) Abs-MPJPE (cm)
11.3122.34
✓9.9114.73
", + "image_path": "90c7a4f3275de3dba530c91b31e227e9ef90cc94a99440586eddbc520590f278.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 194, + 219, + 416, + 258 + ], + "blocks": [ + { + "bbox": [ + 179, + 198, + 433, + 209 + ], + "lines": [ + { + "bbox": [ + 179, + 198, + 433, + 209 + ], + "spans": [ + { + "bbox": [ + 179, + 198, + 433, + 209 + ], + "type": "text", + "content": "Table 6: Our Radar-Doppler Convolutional Block in HR-Net." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 194, + 219, + 416, + 258 + ], + "lines": [ + { + "bbox": [ + 194, + 219, + 416, + 258 + ], + "spans": [ + { + "bbox": [ + 194, + 219, + 416, + 258 + ], + "type": "table", + "html": "
HRNet-Block-DesignMPJPE (cm)Abs-MPJPE (cm)
Traditional Block (RGB-Vision)10.3517.59
Ours (Radar-Doppler)9.9314.73
", + "image_path": "e6f3dafe6a60196acc61c0939f7041d02469409a7da5cd375e57243c48ff9195.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 281, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 413 + ], + "type": "text", + "content": "Qualitative Result Figure 8 demonstrates the qualitative results of different occlusion conditions and a low-lighting instance. Due to the hardware setting of our system, supporting synchronized two camera data, the subject's pose can be visualized by another camera if one of them is occluded. Therefore, the 3D annotation results can be generated even on occlusion conditions. Our method provides reliable pose estimation results even when the radar module is blocked or with low-lighting conditions. Table 4 shows the HPE results of occlusion in normal outdoor conditions. Although the radar module is occluded, the mmWave signal is still robust enough as input for the HRRadarPose model to estimate HPE, demonstrating the characteristics of radar and the importance of using radar for HPE." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 426, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 498 + ], + "type": "text", + "content": "Ablation Study To illustrate the advantage of modeling motion features using Doppler information by our HRRadarPose, we remove Doppler information by averaging 4D radar tensors along the Doppler axis and train the models with and without Doppler information using the same setup. Table 5 shows that with the inclusion of Doppler information, our HRRadarPose enjoys a performance gain due to its capability to extract motion features, enhancing keypoint localization." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 640 + ], + "type": "text", + "content": "Compared with RPM 2.0, our input data and backbone design are different. The proposed HRRadarPose network extracts better feature representation from 4D radar tensors through modified convolutional blocks. As shown in Figure 9, we replace the batch norm with the group norm of convolutional blocks in HRNet. Since we treat the channel-wise features as a Doppler axis, the group norm inherently handles the feature in different speed groups, which normalizes the data effectively. Table 6 shows the results, proving that the modified CNN block is more suitable for Doppler radar than the original HRNet. Compared with our method, most previous work did not leverage Doppler information [47,50]. We validate that motion features [17,27] are crucial for aiding deep learning models in capturing more human body characteristics over the environment." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "To ablate our design choice of HPE heads, we use the same backbone, replace our pose estimation head with RF-Pose3D's [55] head, and train both models for" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "RT-Pose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 160, + 114, + 274, + 170 + ], + "blocks": [ + { + "bbox": [ + 160, + 114, + 274, + 170 + ], + "lines": [ + { + "bbox": [ + 160, + 114, + 274, + 170 + ], + "spans": [ + { + "bbox": [ + 160, + 114, + 274, + 170 + ], + "type": "image", + "image_path": "e9d7b332004f99b083bd1c95c5fda7c8bdba0823064fa183233a3bc08fd23c6c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 170, + 302, + 214 + ], + "lines": [ + { + "bbox": [ + 132, + 170, + 302, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 170, + 302, + 214 + ], + "type": "text", + "content": "Fig.9: Illustration of two variants of CNN blocks in HRNet. (a) Ours (Radar-Doppler). (b) Traditional Block (RGBVision)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 321, + 168, + 469, + 203 + ], + "blocks": [ + { + "bbox": [ + 311, + 137, + 480, + 158 + ], + "lines": [ + { + "bbox": [ + 311, + 137, + 480, + 158 + ], + "spans": [ + { + "bbox": [ + 311, + 137, + 480, + 158 + ], + "type": "text", + "content": "Table 7: Comparison of common 4D radar tensor-based HPE head and ours." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 168, + 469, + 203 + ], + "lines": [ + { + "bbox": [ + 321, + 168, + 469, + 203 + ], + "spans": [ + { + "bbox": [ + 321, + 168, + 469, + 203 + ], + "type": "table", + "html": "
HPE HeadMPJPE (cm)Abs-MPJPE (cm)
RF-Pose3D [55]15.5218.08
Ours10.2417.63
", + "image_path": "9d19895224c08b7ed7277770a0c80ca7ffe37ef0ea2721cbc716a5e35a09348c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 247, + 480, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 480, + 319 + ], + "type": "text", + "content": "the same epochs. RF-Pose3D formulates the HPE as a multi-keypoints classification problem, where their model outputs confidence distributions of every human body joint. As shown in Table 7, our design is favorable in joint keypoints localization in the metric space, indicated by a lower Abs-MPJPE. Moreover, lower MPJPE from our HPE head's result validates stronger inter-joint relationships modeling capability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 336, + 323, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 323, + 348 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 323, + 348 + ], + "type": "text", + "content": "6 Limitation and Future Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 360, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 482, + 468 + ], + "type": "text", + "content": "First, to leverage the input data types for training the 3D HPE model, the computational resource is a crucial part. 4D radar tensor in Cartesian coordinate consumes up to " + }, + { + "bbox": [ + 130, + 360, + 482, + 468 + ], + "type": "inline_equation", + "content": "100\\mathrm{MB}" + }, + { + "bbox": [ + 130, + 360, + 482, + 468 + ], + "type": "text", + "content": " per frame, which decreases the training speed. Second, we limit the experimental data collection scope from " + }, + { + "bbox": [ + 130, + 360, + 482, + 468 + ], + "type": "inline_equation", + "content": "2\\mathrm{m}" + }, + { + "bbox": [ + 130, + 360, + 482, + 468 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 360, + 482, + 468 + ], + "type": "inline_equation", + "content": "8\\mathrm{m}" + }, + { + "bbox": [ + 130, + 360, + 482, + 468 + ], + "type": "text", + "content": " because the camera is hard to clearly capture 2D pose skeletons at a large distance and LiDAR can not easily collect completed point clouds in near vision due to the constraints of FoV. Third, the proposed baseline HRRadarPose lacks robustness in accurately estimating poses during complex activities. Capturing human pose accurately on 4D radar tensor remains an unresolved challenge." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 485, + 220, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 485, + 220, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 485, + 220, + 498 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we propose RT-Pose, the first dataset for human pose estimation (HPE) with synchronized and calibrated 4D radar tensors, LiDAR point clouds, and RGB images. RT-Pose provides cluttered scenarios and human activities of different complexity levels to enhance the difficulty of pose estimation for practical applications. In addition, having various modalities is beneficial for our semi-automatic 3D pose annotation optimization process. The proposed HRRadarPose is the first single-stage architecture designed for estimating human pose using 4D radar tensors as the input. The results indicate that the proposed HRRadarPose outperforms previous works, in which datasets are collected with simpler actions and cleaner scenarios. Our work contributes to offering a three-modality dataset with 4D radar tensors for HPE, which is promising to be applied for complex actions and different scenes. We hope this work encourages future development of 4D radar-based HPE methods." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "Y. Ho et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 482, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 482, + 216 + ], + "type": "text", + "content": "We thank Yizhou Wang " + }, + { + "bbox": [ + 130, + 143, + 482, + 216 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 130, + 143, + 482, + 216 + ], + "type": "text", + "content": " for helping us build the data collection system and the radar data processing pipeline. We thank Hou-I Liu " + }, + { + "bbox": [ + 130, + 143, + 482, + 216 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 130, + 143, + 482, + 216 + ], + "type": "text", + "content": ", Cheng-Yi Huang " + }, + { + "bbox": [ + 130, + 143, + 482, + 216 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 130, + 143, + 482, + 216 + ], + "type": "text", + "content": ", Yi Shiang Chen " + }, + { + "bbox": [ + 130, + 143, + 482, + 216 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 130, + 143, + 482, + 216 + ], + "type": "text", + "content": ", Chi-Yuan Chan " + }, + { + "bbox": [ + 130, + 143, + 482, + 216 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 130, + 143, + 482, + 216 + ], + "type": "text", + "content": ", and Yu-Hsien Lu " + }, + { + "bbox": [ + 130, + 143, + 482, + 216 + ], + "type": "inline_equation", + "content": "^{8}" + }, + { + "bbox": [ + 130, + 143, + 482, + 216 + ], + "type": "text", + "content": " for helping us collect the RT-Pose dataset. We thank the National Science and Technology Counci of the Republic of China, Taiwan, under Projects of NSTC 112-2917-I-006-007 for its funding support." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 236, + 197, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 197, + 249 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 197, + 249 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 264, + 481, + 589 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 138, + 264, + 481, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 264, + 481, + 287 + ], + "spans": [ + { + "bbox": [ + 138, + 264, + 481, + 287 + ], + "type": "text", + "content": "1. Adib, F., Hsu, C.Y., Mao, H., Katabi, D., Durand, F.: Rf-capture: Capturing the human figure through a wall. ACM SIGGRAPH Asia (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 288, + 481, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 288, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 288, + 481, + 331 + ], + "type": "text", + "content": "2. Ahuja, K., Jiang, Y., Goel, M., Harrison, C.: Vid2doppler: Synthesizing doppler radar data from videos for training privacy-preserving activity recognition. In: Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. pp. 1-10 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 333, + 481, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 333, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 138, + 333, + 481, + 376 + ], + "type": "text", + "content": "3. An, S., Li, Y., Ogras, U.: mri: Multi-modal 3d human pose estimation dataset using mmwave, rgb-d, and inertial sensors. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems. vol. 35, pp. 27414-27426. Curran Associates, Inc. (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 377, + 481, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 377, + 481, + 410 + ], + "spans": [ + { + "bbox": [ + 138, + 377, + 481, + 410 + ], + "type": "text", + "content": "4. An, S., Ogras, U.Y.: Mars: mmwave-based assistive rehabilitation system for smart healthcare. ACM Transactions on Embedded Computing Systems (TECS) 20(5s), 1-22 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 411, + 481, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 481, + 443 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 481, + 443 + ], + "type": "text", + "content": "5. Chai, W., Guo, X., Wang, G., Lu, Y.: Stablevideo: Text-driven consistency-aware diffusion video editing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 23040-23050 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 444, + 481, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 444, + 481, + 477 + ], + "spans": [ + { + "bbox": [ + 138, + 444, + 481, + 477 + ], + "type": "text", + "content": "6. Chai, W., Jiang, Z., Hwang, J.N., Wang, G.: Global adaptation meets local generalization: Unsupervised domain adaptation for 3d human pose estimation. arXiv preprint arXiv:2303.16456 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 478, + 481, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 478, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 478, + 481, + 510 + ], + "type": "text", + "content": "7. Chen, A., Wang, X., Zhu, S., Li, Y., Chen, J., Ye, Q.: mmbody benchmark: 3d body reconstruction dataset and analysis for millimeter wave radar. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 3501-3510 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 511, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 511, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 511, + 481, + 555 + ], + "type": "text", + "content": "8. Chen, K., Gabriel, P., Alasfour, A., Gong, C., Doyle, W.K., Devinsky, O., Friedman, D., Dugan, P., Melloni, L., Thesen, T., et al.: Patient-specific pose estimation in clinical environments. IEEE journal of translational engineering in health and medicine 6, 1-11 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 555, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 481, + 589 + ], + "type": "text", + "content": "9. Cheng, J.H., Chen, Y., Chang, T.Y., Lin, H.E., Wang, P.Y.C., Cheng, L.P.: Impossible staircase: Vertically real walking in an infinite virtual tower. In: 2021 IEEE Virtual Reality and 3D User Interfaces (VR). pp. 50-56 (2021)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "RT-Pose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 134, + 597, + 241, + 610 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 597, + 241, + 610 + ], + "spans": [ + { + "bbox": [ + 134, + 597, + 241, + 610 + ], + "type": "text", + "content": "3 http://yizhouwang.net/" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 135, + 610, + 235, + 620 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 610, + 235, + 620 + ], + "spans": [ + { + "bbox": [ + 135, + 610, + 235, + 620 + ], + "type": "text", + "content": "k39967.c@nycu.edu.tw" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 135, + 621, + 253, + 632 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 621, + 253, + 632 + ], + "spans": [ + { + "bbox": [ + 135, + 621, + 253, + 632 + ], + "type": "text", + "content": "5 n28121521@gs.ncku.edu.tw" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 135, + 632, + 253, + 643 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 632, + 253, + 643 + ], + "spans": [ + { + "bbox": [ + 135, + 632, + 253, + 643 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 135, + 632, + 253, + 643 + ], + "type": "text", + "content": " m16121051@gs.ncku.edu.tw" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 135, + 643, + 253, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 643, + 253, + 654 + ], + "spans": [ + { + "bbox": [ + 135, + 643, + 253, + 654 + ], + "type": "text", + "content": "7 n26120587@gs.ncku.edu.tw" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 135, + 654, + 253, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 654, + 253, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 654, + 253, + 665 + ], + "type": "text", + "content": "n26120587@gs.ncku.edu.tw" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "10. Cheng, J.H., Kuan, S.Y., Liu, H.I., Latapie, H., Liu, G., Hwang, J.N.: Centerradar-net: Joint 3d object detection and tracking framework using 4d fmcw radar. In: 2024 IEEE International Conference on Image Processing (ICIP) (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "text", + "content": "11. Dai, Y., Lin, Y., Lin, X., Wen, C., Xu, L., Yi, H., Shen, S., Ma, Y., Wang, C.: Sloper4d: A scene-aware dataset for global 4d human pose estimation in urban environments. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 682-692 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 194, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 480, + 237 + ], + "type": "text", + "content": "12. Dai, Y., Lin, Y., Wen, C., Shen, S., Xu, L., Yu, J., Ma, Y., Wang, C.: Hsc4d: Human-centered 4d scene capture in large-scale indoor-outdoor space using wearable imus and lidar. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6792-6802 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 238, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 238, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 238, + 480, + 270 + ], + "type": "text", + "content": "13. De Maio, A., Farina, A., Foglia, G.: Design and experimental validation of knowledge-based constant false alarm rate detectors. IET Radar, Sonar & Navigation 1(4), 308-316 (2007)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "text", + "content": "14. Endo, K., Ishikawa, T., Yamamoto, K., Ohtsuki, T.: Multi-person position estimation based on correlation between received signals using mimo fmcw radar. IEEE Access 11, 2610-2620 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 304, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 304, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 133, + 304, + 480, + 346 + ], + "type": "text", + "content": "15. Guzov, V., Mir, A., Sattler, T., Pons-Moll, G.: Human poseitioning system (hps): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4318-4329 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 133, + 347, + 480, + 380 + ], + "type": "text", + "content": "16. Heath, R.W., Gonzalez-Prelcic, N., Rangan, S., Roh, W., Sayeed, A.M.: An overview of signal processing techniques for millimeter wave mimo systems. IEEE journal of selected topics in signal processing 10(3), 436-453 (2016)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 380, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 380, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 133, + 380, + 480, + 424 + ], + "type": "text", + "content": "17. Huang, H.W., Yang, C.Y., Ramkumar, S., Huang, C.I., Hwang, J.N., Kim, P.K., Lee, K., Kim, K.: Observation centric and central distance recovery for athlete tracking. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 454-460 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 425, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 425, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 133, + 425, + 480, + 468 + ], + "type": "text", + "content": "18. Huang, H.Y., Ning, C.W., Wang, P.Y., Cheng, J.H., Cheng, L.P.: Haptic-go-round: A surrounding platform for encounter-type haptics in virtual reality experiences. In: Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems. CHI '20, Association for Computing Machinery (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 468, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 468, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 133, + 468, + 480, + 511 + ], + "type": "text", + "content": "19. Ionescu, C., Papava, D., Olaru, V., Sminchisescu, C.: Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence 36(7), 1325-1339 (2013)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 133, + 512, + 480, + 544 + ], + "type": "text", + "content": "20. Jiang, Z., Chai, W., Li, L., Zhou, Z., Yang, C.Y., Hwang, J.N.: Unihpe: Towards unified human pose estimation via contrastive learning. arXiv preprint arXiv:2311.16477 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 545, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 545, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 133, + 545, + 480, + 588 + ], + "type": "text", + "content": "21. Jiang, Z., Ji, H., Yang, C.Y., Hwang, J.N.: 2d human pose estimation calibration and keypoint visibility classification. In: ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 6095-6099. IEEE (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 589, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 589, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 589, + 480, + 632 + ], + "type": "text", + "content": "22. Jiang, Z., Zhou, Z., Li, L., Chai, W., Yang, C.Y., Hwang, J.N.: Back to optimization: Diffusion-based zero-shot 3d human pose estimation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6142-6152 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 480, + 666 + ], + "type": "text", + "content": "23. Kim, Y., Alnujaim, I., Oh, D.: Human activity classification based on point clouds measured by millimeter wave mimo radar with deep recurrent neural networks. IEEE Sensors Journal 21(12), 13522-13529 (2021)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "Y. Ho et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "text", + "content": "24. Klauder, J.R., Price, A., Darlington, S., Albersheim, W.J.: The theory and design of chirp radars. Bell System Technical Journal 39(4), 745-808 (1960)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 481, + 171 + ], + "type": "text", + "content": "25. Kuan, S.Y., Cheng, J.H., Huang, H.W., Chai, W., Yang, C.Y., Wu, B.F., Hwang, J.N.: Boosting online 3d multi-object tracking through camera-radar cross check. In: IEEE Intelligent Vehicles Symposium (IV) (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 481, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 481, + 213 + ], + "type": "text", + "content": "26. Lee, S.P., Kini, N.P., Peng, W.H., Ma, C.W., Hwang, J.N.: Hupr: A benchmark for human pose estimation using millimeter wave radar. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5715-5724 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 213, + 481, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 213, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 213, + 481, + 257 + ], + "type": "text", + "content": "27. Lin, C.L., Chang, Y.M., Hung, C.C., Tu, C.D., Chuang, C.Y.: Position estimation and smooth tracking with a fuzzy-logic-based adaptive strong tracking kalman filter for capacitive touch panels. IEEE Transactions on Industrial Electronics 62(8), 5097-5108 (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 257, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 481, + 300 + ], + "type": "text", + "content": "28. Lin, C.L., Chiu, W.C., Chu, T.C., Ho, Y.H., Chen, F.H., Hsu, C.C., Hsieh, P.H., Chen, C.H., Lin, C.C.K., Sung, P.S., et al.: Innovative head-mounted system based on inertial sensors and magnetometer for detecting falling movements. Sensors 20(20), 5774 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "type": "text", + "content": "29. Lin, C.L., Ho, Y.H., Chiu, W.C., Chu, T.C., Liu, Y.H.: Innovative shoe-integrated system based on time-of-flight range sensors for fall detection on various terrains. IEEE Sensors Letters 5(10), 1-4 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "type": "text", + "content": "30. Lin, H.Y., Chen, T.W.: Augmented reality with human body interaction based on monocular 3d pose estimation. In: International Conference on Advanced Concepts for Intelligent Vision Systems. pp. 321-331. Springer (2010)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 365, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 481, + 407 + ], + "type": "text", + "content": "31. Liu, H., He, J.Y., Cheng, Z.Q., Xiang, W., Yang, Q., Chai, W., Wang, G., Bao, X., Luo, B., Geng, Y., et al.: Posynda: Multi-hypothesis pose synthesis domain adaptation for robust 3d human pose estimation. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 5542-5551 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 407, + 481, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 440 + ], + "type": "text", + "content": "32. Mehta, D., Rhodin, H., Casas, D., Fua, P., Sotnychenko, O., Xu, W., Theobalt, C.: Monocular 3d human pose estimation in the wild using improved cnn supervision. In: 2017 international conference on 3D vision (3DV). pp. 506-516. IEEE (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 440, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 481, + 472 + ], + "type": "text", + "content": "33. Neemat, S., Uysal, F., Krasnov, O., Yarovoy, A.: Reconfigurable range-doppler processing and range resolution improvement for fmcw radar. IEEE Sensors Journal 19(20), 9294-9303 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "type": "text", + "content": "34. Paek, D.H., Kong, S.H., Wijaya, K.T.: K-radar: 4d radar object detection for autonomous driving in various weather conditions. Advances in Neural Information Processing Systems 35, 3819-3829 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "type": "text", + "content": "35. Sengupta, A., Cao, S.: mmpose-nlp: A natural language processing approach to precise skeletal pose estimation using mmwave radars. IEEE Transactions on Neural Networks and Learning Systems (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "type": "text", + "content": "36. Sengupta, A., Jin, F., Cao, S.: Nlp based skeletal pose estimation using mmwave radar point-cloud: A simulation approach. In: 2020 IEEE Radar Conference (RadarConf20). pp. 1-6. IEEE (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "text", + "content": "37. Sengupta, A., Jin, F., Zhang, R., Cao, S.: mm-pose: Real-time human skeletal posture estimation using mmwave radars and cnns. IEEE Sensors Journal 20(17), 10032-10044 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "text", + "content": "38. Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learning for human pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5693-5703 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "text", + "content": "39. Sun, Z., Ke, Q., Rahmani, H., Bennamoun, M., Wang, G., Liu, J.: Human action recognition from various data modalities: A review. IEEE transactions on pattern analysis and machine intelligence (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "RT-Pose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "40. Véges, M., Lorincz, A.: Absolute human pose estimation with depth prediction network. In: 2019 International Joint Conference on Neural Networks (IJCNN). pp. 1-7. IEEE (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 192 + ], + "type": "text", + "content": "41. Von Marcard, T., Henschel, R., Black, M.J., Rosenhahn, B., Pons-Moll, G.: Recovering accurate 3d human pose in the wild using imus and a moving camera. In: Proceedings of the European conference on computer vision (ECCV). pp. 601-617 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 192, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 192, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 192, + 480, + 224 + ], + "type": "text", + "content": "42. Wang, J., Tan, S., Zhen, X., Xu, S., Zheng, F., He, Z., Shao, L.: Deep 3d human pose estimation: A review. Computer Vision and Image Understanding 210, 103225 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 224, + 480, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 256 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 256 + ], + "type": "text", + "content": "43. Wang, J., Sun, K., Cheng, T., Jiang, B., Deng, C., Zhao, Y., Liu, D., Mu, Y., Tan, M., Wang, X., Liu, W., Xiao, B.: Deep high-resolution representation learning for visual recognition. TPAMI (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 257, + 480, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 299 + ], + "type": "text", + "content": "44. Wang, J., Sun, K., Cheng, T., Jiang, B., Deng, C., Zhao, Y., Liu, D., Mu, Y., Tan, M., Wang, X., et al.: Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence 43(10), 3349-3364 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 300, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 342 + ], + "type": "text", + "content": "45. Wang, Y., Cheng, J.H., Huang, J.T., Kuan, S.Y., Fu, Q., Ni, C., Hao, S., Wang, G., Xing, G., Liu, H., Hwang, J.N.: Vision meets mmwave radar: 3d object perception benchmark for autonomous driving. In: IEEE Intelligent Vehicles Symposium (IV) (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 343, + 480, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 480, + 386 + ], + "type": "text", + "content": "46. Wang, Y., Jiang, Z., Li, Y., Hwang, J.N., Xing, G., Liu, H.: Rodnet: A real-time radar object detection network cross-supervised by camera-radar fused object 3d localization. IEEE Journal of Selected Topics in Signal Processing 15(4), 954-967 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "text", + "content": "47. Xie, C., Zhang, D., Wu, Z., Yu, C., Hu, Y., Chen, Y.: RPM 2.0: Rf-based pose machines for multi-person 3d pose estimation. IEEE Transactions on Circuits and Systems for Video Technology (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 418, + 480, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 461 + ], + "type": "text", + "content": "48. Xue, H., Ju, Y., Miao, C., Wang, Y., Wang, S., Zhang, A., Su, L.: mmmesh: Towards 3d real-time dynamic human mesh construction using millimeter-wave. In: Proceedings of the 19th Annual International Conference on Mobile Systems, Applications, and Services. pp. 269-282 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 461, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 480, + 504 + ], + "type": "text", + "content": "49. Yang, C.Y., Luo, J., Xia, L., Sun, Y., Qiao, N., Zhang, K., Jiang, Z., Hwang, J.N., Kuo, C.H.: Camerapose: Weakly-supervised monocular 3d human pose estimation by leveraging in-the-wild 2d annotations. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2924-2933 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "type": "text", + "content": "50. Yu, C., Zhang, D., Wu, Z., Xie, C., Lu, Z., Hu, Y., Chen, Y.: Mobirfpose: Portable rf-based 3d human pose camera. IEEE Transactions on Multimedia (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "type": "text", + "content": "51. Yuan, Y., Wei, S.E., Simon, T., Kitani, K., Saragih, J.: Simpoe: Simulated character control for 3d human pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7159-7169 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 558, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 480, + 590 + ], + "type": "text", + "content": "52. Zhang, G., Geng, X., Lin, Y.J.: Comprehensive mpoint: A method for 3d point cloud generation of human bodies utilizing fmcw mimo mm-wave radar. Sensors 21(19), 6455 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 590, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 621 + ], + "type": "text", + "content": "53. Zhang, Z., Chai, W., Jiang, Z., Ye, T., Song, M., Hwang, J.N., Wang, G.: Mpm: A unified 2d-3d human pose representation via masked pose modeling. arXiv preprint arXiv:2306.17201 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "54. Zhao, M., Li, T., Abu Alsheikh, M., Tian, Y., Zhao, H., Torralba, A., Katabi, D.: Through-wall human pose estimation using radio signals. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7356-7365 (2018)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "Y. Ho et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 304 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "55. Zhao, M., Tian, Y., Zhao, H., Alsheikh, M.A., Li, T., Hristov, R., Kabelac, Z., Katabi, D., Torralba, A.: Rf-based 3d skeletons. In: Proceedings of the 2018 Conference of the ACM Special Interest Group on Data Communication. pp. 267-281 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "type": "text", + "content": "56. Zhao, Y., Yarovoy, A., Fioranelli, F.: Angle-insensitive human motion and posture recognition based on 4d imaging radar and deep learning classifiers. IEEE Sensors Journal 22(12), 12173-12182 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 131, + 194, + 481, + 226 + ], + "type": "text", + "content": "57. Zheng, C., Wu, W., Chen, C., Yang, T., Zhu, S., Shen, J., Kehtarnavaz, N., Shah, M.: Deep learning-based human pose estimation: A survey. ACM Computing Surveys 56(1), 1-37 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 227, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 227, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 131, + 227, + 481, + 259 + ], + "type": "text", + "content": "58. Zheng, Z., Pan, J., Ni, Z., Shi, C., Ye, S., Fang, G.: Human posture reconstruction for through-the-wall radar imaging using convolutional neural networks. IEEE Geoscience and Remote Sensing Letters 19, 1-5 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 259, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 259, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 131, + 259, + 481, + 304 + ], + "type": "text", + "content": "59. Zhou, Z., Jiang, Z., Chai, W., Yang, C.Y., Li, L., Hwang, J.N.: Efficient domain adaptation via generative prior for 3d infant pose estimation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 41-49 (2024)" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "RT-Pose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/R^2-Bench_ Benchmarking the Robustness of Referring Perception Models under Perturbations/3cf440ac-b4a0-433a-943f-d4d176a64132_content_list.json b/2024/R^2-Bench_ Benchmarking the Robustness of Referring Perception Models under Perturbations/3cf440ac-b4a0-433a-943f-d4d176a64132_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..003a463fa99c4cec0fd0c39736a3382344e26926 --- /dev/null +++ b/2024/R^2-Bench_ Benchmarking the Robustness of Referring Perception Models under Perturbations/3cf440ac-b4a0-433a-943f-d4d176a64132_content_list.json @@ -0,0 +1,1817 @@ +[ + { + "type": "text", + "text": "$\\mathbf{R}^2$ -Bench: Benchmarking the Robustness of Referring Perception Models under Perturbations", + "text_level": 1, + "bbox": [ + 217, + 143, + 784, + 189 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiang Li $^{1}$ , Kai Qiu $^{1}$ , Jinglu Wang $^{2}$ , Xiaohao Xu $^{3}$ , Rita Singh $^{1}$ , Kashu Yamazaki $^{1}$ , Hao Chen $^{1}$ , Xiaonan Huang $^{3}$ , Bhiksha Raj $^{1,4}$", + "bbox": [ + 246, + 214, + 753, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 CMU, 2 Microsoft Research Asia, 3 University of Michigan, 4 MBZUAI", + "bbox": [ + 256, + 256, + 743, + 272 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/10aae151004360f04481c045134b269f97efaca8992350082a8d8f6834a038d2.jpg", + "image_caption": [ + "(a) Referring Perception in Practice" + ], + "image_footnote": [], + "bbox": [ + 217, + 282, + 493, + 404 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/27f810604090a0d5ebf36812144a519207af7753772bb7335da1cf77b9cf7049.jpg", + "image_caption": [ + "(b) Perturbation in Referring Guidance", + "Fig. 1: Motivation illustration. Referring perception models (RPMs) empower intelligent systems with their ability to perform object grounding within the environment based on referring guidance, such as textual descriptions, imagery exemplars, or auditory signals associated with the target object. However, RPMs' performance can be compromised by disturbances in real-world scenarios, such as environmental noise (e.g., extraneous sounds from a nearby radio), human-induced errors (e.g., typographical errors in textual input), and limitations in the sensor (e.g., motion blur in images). Conducting a rigorous analysis of RPMs' robustness to a wide array of perturbations is necessary for building reliable real-world applications." + ], + "image_footnote": [], + "bbox": [ + 522, + 282, + 785, + 405 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Referring perception, which aims at grounding visual objects with multimodal referring guidance, is essential for bridging the gap between humans, who provide instructions, and the environment where intelligent systems perceive. Despite progress in this field, the robustness of referring perception models (RPMs) against disruptive perturbations is not well explored. This work thoroughly assesses the resilience of RPMs against various perturbations in both general and specific contexts. Recognizing the complex nature of referring perception tasks, we present a comprehensive taxonomy of perturbations, and then develop a versatile toolbox for synthesizing and evaluating the effects of composite disturbances. Employing this toolbox, we construct $\\mathbf{R}^2$ -Bench, a benchmark for assessing the Robustness of Referring perception models under noisy conditions across five key tasks. Moreover, we propose the $\\mathbf{R}^2$ -Agent, an LLM-based agent that simplifies and automates model evaluation via natural language instructions. Our investigation uncovers the vulnerabilities of current RPMs to various perturbations and provides tools for assessing model robustness, potentially promoting the safe and resilient integration of intelligent systems into complex real-world scenarios.", + "bbox": [ + 259, + 563, + 743, + 813 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Referring perception $\\cdot$ Robustness & Perturbation $\\cdot$ Benchmark", + "bbox": [ + 259, + 818, + 743, + 845 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5a13cc43f38aaba38e8329e201d3c93c8a82d48baa9f88ae1a982df887bdbf27.jpg", + "image_caption": [ + "Fig. 2: Examples of $\\mathbf{R}^2$ -Bench. The \"Original\" row displays the original inputs alongside the outcomes from models of R-VOS [54], AVS [34] and Q3M [23], while the \"Perturbed\" row presents the inputs as synthesized by $\\mathbf{R}^2$ -Bench and the respective outcomes with the same models. RG: short for Referring Guidance." + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 774, + 279 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 369, + 375, + 386 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Perception systems [6-8,29-32,80,81] function as input channels for intelligent systems, analogous to human eyes. Referring perception, focuses on identification and contextualization of visual entities by referring multimodal guidance, illustrated in Fig. 1 (a). It effectively creates a communicative bridge between humans, who issue instructions, and the environment that is subject to perception. Multimodal referring cues used for such identification include textual descriptions, example imagery, or auditory signals corresponding to the target object (Fig. 1 (b)). Within the domain of referring perception, specific tasks such as referring expression segmentation [37,68], audiovisual source localization [17,34,79], and queryable 3D mapping [23], are essential modules of robotic control [2,22], autonomous navigation [23,40] and planning [21,48], and human-computer interaction [51].", + "bbox": [ + 212, + 401, + 787, + 568 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advancements in referring perception models [14, 33, 68] have been witnessed and deployed on resource-constrained platforms [54, 58]. However, existing datasets for model training often comprise delicate images with accurate annotations, a condition rarely met in real-world scenarios. In realistic settings, as illustrated in Fig. 1, perception systems may face various disturbances, such as environmental noise (buzzing sound from the radio), imprecise human instructions (misspelling), and sensor imperfections (image blur), which can significantly challenge the robustness of these models. Despite their critical importance for safer real-world applications, the impact of such perturbations on model performance is not thoroughly investigated in the literature. Evaluating model robustness is further complicated by the diversity of operational environments; the assessment process is typically labor-intensive and relies heavily on expert judgment. Addressing this gap, this paper introduces the first benchmark in the field to systematically assess the resilience of referring perception models, i.e., trying to obtain a reliable and comprehensive answer to the critical question: \"How robust are the current referring perception models under realistic perturbations?\"", + "bbox": [ + 212, + 568, + 789, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we introduce the Robust Referring Benchmark $(\\mathbf{R}^2$ -Bench), featuring: (1) a human-friendly taxonomy of perturbations for referring perception,", + "bbox": [ + 212, + 809, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Xiang Li et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "categorizing disruptions into source, environment, sensor, and transmission noise, and detailing perturbation types for each modality; (2) a customizable perturbation synthesis toolbox for creating reasonable noise-augmented datasets, enabling robustness evaluation of perception models; and (3) robustness evaluations across five tasks—referring image segmentation (RIS), video object segmentation (VOS), referring video object segmentation (R-VOS), audiovisual segmentation (AVS), and queryable 3D mapping (Q3M)—considering textual, visual, and acoustic reference modalities. We assess over twenty prevalent models in noisy conditions to benchmark their performance and analyze their vulnerabilities. (Examples are illustrated in Fig. 2.) We hope our benchmarking and analysis can benefit the community by shedding light on the vulnerability of existing models in the face of various perturbations.", + "bbox": [ + 212, + 146, + 787, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Notably, our work advances the evaluation of model robustness by incorporating tests against composite noise types, more closely mirroring real-world scenarios where different noises often coexist. The construction of complex noise for realistic robustness assessment is not trivial, as it must avoid non-contextual perturbations, such as the presence of snow in indoor imagery or excessive air absorption effects in indoor audio recordings, which are incongruent with the depicted environment. To address this complexity, we introduce the $\\mathbf{R}^2$ -Agent, a novel large language model-based (LLM-based) system that can comprehend natural language instructions provided by humans and autonomously generates test samples with contextual relevant perturbations that align with the deployment environment. For instance, given the human-provided instruction \"outdoor night scene\", the $\\mathbf{R}^2$ -Agent could probably generate perturbations of \"motion blur\" instead of \"glass blur\" and \"low background noise\" instead of \"room reverberation\". To improve the agent's abilities of instruction-following and commonsense-reasoning, we employ the multi-agent debating technique [16,36,77], wherein one agent proposes potential solutions while another verifies their validity. Consequently, the $\\mathbf{R}^2$ -Agent efficiently automates the identification and integration of mixed perturbations, thus optimizing the process of robustness evaluations tailored to specific domain contexts.", + "bbox": [ + 212, + 329, + 787, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our major contribution is three-fold:", + "bbox": [ + 238, + 619, + 598, + 633 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce the $\\mathbf{R}^2$ -Bench, a benchmarking framework that includes diverse perturbed data for five commonly encountered referring perception tasks. A comprehensive taxonomy and a corresponding customizable synthesis toolbox are developed to enable robustness assessment in noisy real-world settings.", + "- We propose the $\\mathbf{R}^2$ -Agent to streamline model evaluations tailored to particular use scenarios based on LLMs. This automated program executes the perturbation composition process in accordance with human-provided instructions, thereby allowing for context-specific robustness evaluations.", + "- Our systematic experimental investigations delve into the intrinsic characteristics of perturbations, such as their types, severity, dynamics, and correlations. These explorations yield valuable insights into the vulnerability of existing models to disturbances and elucidate the nature of these perturbations." + ], + "bbox": [ + 225, + 652, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Benchmarking the Robustness of Referring Perception Models under Perturbations", + "bbox": [ + 215, + 114, + 767, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 810, + 114, + 821, + 127 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 143, + 397, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Textual referring perception. Referring image segmentation (RIS) and referring video object segmentation (R-VOS) aim to segment objects in images and video sequences, respectively, based on a linguistic description. Recent RIS methods [15, 37,63,68,82-84] have achieved promising results using multimodal transformers. R-VOS [14,18,42,47,49,54] is more challenging as it requires leveraging both intra-frame and temporal cues. URVOS [47] is the first unified R-VOS framework with cross-modal attention and a memory attention module, significantly improving R-VOS performance. ReferFormer [55] employs a linguistic prior in the transformer decoder to focus on the referred object, while MTTR [5] uses a multimodal transformer encoder to fuse linguistic and visual features. $\\mathbb{R}^2$ -VOS introduces relational cyclic consistency to enhance the robustness of the R-VOS model. Unlike other vision-language tasks [65,66], R-VOS needs to construct object-level multimodal semantic consensus in dense visual representations. Relying on 2D models, 3D referring perception is also achievable [23, 64].", + "bbox": [ + 212, + 176, + 787, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Acoustic referring perception. Audiovisual segmentation (AVS) [35, 35, 38, 78] focuses on segmenting objects that produce sound in a given image frame. This pioneering work by Zhou et al. [78] introduced a method that uses cross-modal attention to identify the sound source. Building on this, Zhou et al. [78] proposed an extended task called audiovisual semantic segmentation (AVSS), which not only segments the sound-producing objects but also classifies them. AVSS is more challenging than AVS due to the complexity of audio semantics. To address this, Zhou et al. [78] utilized the TPAVI module from [79] for audiovisual interaction. Additionally, a recent development by Li et al. [27] introduced the CATR framework, which features a novel spatial-temporal audio-video fusion block for efficient audio-visual integration. Li et al. [34] propose a quantization-based semantic decomposition module to handle the complex acoustic environment and enhance the robustness. Sound source localization (SSL) is a related field that aims to identify the visual regions corresponding to sounds. Several SSL methods [3, 4, 12, 46] utilize the correspondence between audio and visual features to locate sounds, often represented as heatmaps. For example, Mo et al. [43] employed multi-level audiovisual contrastive learning to effectively pinpoint sound-producing objects. In addition, using speech as referring guidance is also explored in [35, 44]", + "bbox": [ + 212, + 401, + 787, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Visual referring perception. Visual referring perception typically relies on visual prompts such as example images, points, bounding boxes, or scribbles. Segment anything model (SAM) [26] is a powerful foundation model that supports multiple visual prompts. Several follow-up works [24,39,58] improve SAM in terms of inference cost, segmentation quality, and referring prompts. Beyond that, several works [67,73,75] explore the usage of SAM in downstream tasks. Li et al. [30] built a SAM-like model with stable diffusion. For video-level tasks, semi-video object segmentation (VOS) [9,28,60,70,71] aims to segment visual objects across frames given the first frame annotation. Recent offline models [9,11] focus more", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Xiang Li et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "on designing long-term information propagation modules to transfer previous image features and corresponding masks to the target frame to predict masks. This helps more precisely identify and track the movement trajectory of objects throughout the entire video sequence (more than 200 frames) but also makes the network architecture heavy [72]. Due to the slow speed of offline models, online models [56, 76] come back into focus, aiming to keep the right balance between speed and performance.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 $\\mathbf{R}^2$ -Bench: Customizable Perturbation Benchmark", + "text_level": 1, + "bbox": [ + 214, + 275, + 736, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we present the $\\mathbf{R}^2$ -Bench for evaluating robustness of referring perception tasks, including a comprehensive taxonomy that categorizes various perturbations and a customizable perturbation synthesis toolbox that can be tailored to generate specific disturbances.", + "bbox": [ + 212, + 306, + 785, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Taxonomic Perturbations in Referring Perception", + "text_level": 1, + "bbox": [ + 214, + 391, + 671, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Referring perception tasks typically employ multimodal guidance, spanning the visual, auditory, and textural domains, to ground objects in the visual contexts. Nonetheless, in real-world scenarios, each modality may undergo heterogeneous noises before input into the model. As depicted in Fig. 3, noises can be categorized into four classes based on their origins: source noises, environment noises, sensor noises, and transmission noises.", + "bbox": [ + 212, + 416, + 478, + 612 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/91b81622bee6eaa89487e40ef6b687a51d3a2312cfde49d3d3c2d7580915fa72.jpg", + "image_caption": [ + "Fig. 3: Noise categories based on their origins. Assuming airplane as the source of referring guidance, noise from it is categorized as source noise." + ], + "image_footnote": [], + "bbox": [ + 491, + 417, + 774, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Source noises. We define source noise as the inherent disturbances introduced spontaneously by the source of the referring guidance. For example, for text-referring tasks, the source noises can manifest as misspellings, grammar errors, or punctuation errors in the textual expression and instruction (human is usually the source of text reference).", + "bbox": [ + 212, + 628, + 787, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Environment noises. We define environment noise as the unintended disturbance introduced by the surrounding environment during the signal capture process. Background sound, room reverberation and air absorption for acoustic waveforms are common examples of this type.", + "bbox": [ + 212, + 719, + 787, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sensor noises. We define sensor noise as the unwanted disturbances in the output of sensors that arise from inherent limitations or imperfections. For example, defocus blur, motion blur and impulse noise for visual frames.", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Benchmarking the Robustness of Referring Perception Models under Perturbations", + "bbox": [ + 215, + 114, + 767, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 810, + 114, + 821, + 127 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2271e41a80c7137d63a06b809ae49c2bebdee9ab206335689ff40514be8380dd.jpg", + "table_caption": [ + "Table 1: Noise types considered in $\\mathbb{R}^2$ -Bench. Details are available in the Appendix." + ], + "table_footnote": [], + "table_body": "
TypeSourceEnvironmentSensorTransmission
Visual-snow (SN), fog (FG), frost (FT), spatter (SP), brightness (BR)defocus blur (DB), gaussian blur (GB), motion blur (MB), glass blur (GS), impulse noise (IN), shot noise (ST), speckle noise (SPN), contrast (CT), saturate (SA)JPEG compression (JPG), pixelated (PIX)
Acousticamplitude gain (GA)background noise (BN), air absorption (AA), room reverberation (RS)gaussian noise (GN), impulse noise (IN), peak filter (PF), time mask (TM), tanh distortion (TD)MP3 compression (MP3), lowpass filter (LP), highpass filter (HP)
Textualmisspelling (MS), mispunctuation (MP), grammar error (GE)-character missing (CM)-
", + "bbox": [ + 218, + 161, + 784, + 279 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Transmission noises. We define transmission noise as the signal distortion or loss during the transmission resulting from data compression or package loss. For example, JPEG and MP4 compression for visual and acoustic signals.", + "bbox": [ + 212, + 309, + 782, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We summarize the supported noises in visual, acoustic and textual modalities in Tab. 1. A total of 32 types of noises are considered in this paper. The implementation details of noise functions are available in the Appendix.", + "bbox": [ + 212, + 354, + 784, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Customizable Perturbation Synthesis", + "text_level": 1, + "bbox": [ + 215, + 422, + 568, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We approach the robustness evaluation through two conventional paradigms: (1) general evaluation under universal scenarios and (2) specific evaluation under designated instructions. To facilitate these evaluations, it is essential to generate perturbed datasets from the original clean data samples, denoted as $\\mathcal{X} = \\{x\\}$ , by introducing a spectrum of heterogeneous perturbations $\\varDelta=\\{\\delta_k\\}$ .", + "bbox": [ + 212, + 449, + 787, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Perturbation order. As discussed in Sec. 3.1, perturbations arise sequentially according to their origins in real-world contexts. To emulate this, perturbations are applied in a sequence that mirrors the natural order of disruption: source $\\rightarrow$ environment $\\rightarrow$ sensor $\\rightarrow$ transmission. Within a single category of origin, the order is randomly determined to account for the non-deterministic nature of real-world noise. To formalize the perturbation process, consider a clean data sample denoted by $x$ , the perturbed sample $x'$ undergoes a series of transformations that can be mathematically described as follows:", + "bbox": [ + 212, + 541, + 787, + 661 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {\\prime} = \\delta_ {t} \\circ \\delta_ {s e} \\circ \\delta_ {e} \\circ \\delta_ {s o} (x), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 675, + 785, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\circ$ denotes composition operation. $\\delta_{so},\\delta_e,\\delta_{se},\\delta_t$ represent generation functions for source, environment, sensor, and transmission noises, respectively. Given that the introduction of perturbations follows a sequential order, the operations that compose these noise functions are inherently order-variant.", + "bbox": [ + 212, + 703, + 787, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Perturbation severity and mode. In alignment with established methodologies [20, 61, 62], we quantitatively define perturbation severity across discrete levels that align with human perceptual categories, namely, {low, medium, high}. In addition, we investigate perturbation across two main modes based on dynamics:", + "bbox": [ + 215, + 779, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Xiang Li et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/96abbc7c9eeb9ae3d7a739ee1cea094ab1171bf9222f7b7e973cdaf97bc4a3c3.jpg", + "image_caption": [ + "Fig. 4: Overview of $\\mathbf{R}^2$ -Agent, the automatic evaluation assistant. Given a human instruction, clean datasets, perturbation functions, and evaluation functions, $\\mathbf{R}^2$ -Agent first proposes and verifies perturbed test samples that match the given instruction. After that, $\\mathbf{R}^2$ -Agent evaluates the model using the verified samples and provides a report that articulates the model's vulnerabilities and overall resilience." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 782, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "{static, dynamic}. Static perturbations remain a consistent level of severity throughout all sensor frames within a sequence, while dynamic perturbations exhibit varying severity and types from frame to frame, closely simulating the fluctuating nature of disturbances in real-world environments, taking the form:", + "bbox": [ + 212, + 398, + 787, + 460 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {\\prime} = \\circ_ {k = 1} ^ {K} \\delta_ {k} (x, \\lambda_ {k} (t)), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 419, + 472, + 787, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\lambda_{k}(t)$ denotes the severity of the $k$ -th perturbation at time step $t$ .", + "bbox": [ + 212, + 503, + 733, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the general evaluation under universal scenarios, we construct perturbed datasets for each task by leveraging the original clean benchmarks and all perturbations. Specifically, noisy datasets with low, medium, high and dynamic severity are considered. In each setting, a maximum of two perturbations are considered simultaneously. The detailed data-creating procedure is available in the Appendix. Comprehensive experiments and analyses discussing the impact of perturbations are provided to facilitate future research in Sec. 5.3.", + "bbox": [ + 212, + 518, + 787, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, we also consider specific evaluations under designated instructions in practice, e.g., virtual phone meetings without background noise. We detail $\\mathbf{R}^2$ -Agent in Sec. 4.", + "bbox": [ + 212, + 626, + 789, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 $\\mathbf{R}^2$ -Agent: Automatic Evaluation Assistant", + "text_level": 1, + "bbox": [ + 214, + 696, + 665, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With the versatile capabilities of LLMs [1, 50], LLM-based agents [52, 57] are employed to automatically perform various tasks in a human-like manner. In this section, we introduce $\\mathbf{R}^2$ -Agent, an LLM-based agent that facilitates automatic perturbation synthesis and robustness evaluation guided by human instructions. As depicted in Fig. 4, the $\\mathbf{R}^2$ -Agent consists of three primary steps: data proposal, data verification, and model analysis. We utilize Gemini-Vision-Pro [50] as the LLM component within the $\\mathbf{R}^2$ -Agent.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Benchmarking the Robustness of Referring Perception Models under Perturbations", + "bbox": [ + 215, + 114, + 767, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 810, + 114, + 821, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cdd590ad2708d573ea6e9a599e553b4cf4979a09c8bd5f0644212171300735af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 146, + 236, + 156 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Example", + "bbox": [ + 238, + 148, + 274, + 156 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "② Question", + "bbox": [ + 222, + 181, + 277, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Given a human instruction, you need to verify...Here is the list of all perturbations you need to consider: {ALL_PERTURBATIONS}. You will be given the results selected from another model as input. For example, if the input is {\"type\": \"snoweffect\", \"reason\": \"the snow effect is a noise type\", \"selected\": True} and the instruction is \"Indoor zoom meeting\", then this data sample is not correct since the instruction pointing to an indoor scenario while the snow effect is outdoor. You need to correct the \"selected\" and \"reason\" items in the dict... Now consider an of {HUMAN_INSTRUCTION}, what are the samples possibly happened within the following perturbations? Explain the reason. Reply me in a strict format of python list. For example, [\"type\": noise_type1, \"reason\": reason1, \"selected\": True/False], . Selected perturbations: {SELECTED_PERTURBATIONS}. \n``python \nperturbations = [\"type\": \"snoweffect\", \"reason\": \"...\", \"selected\": False], [\"type\": \"fog effect\", \"reason\": \"...\", \"selected\": False],", + "bbox": [ + 284, + 147, + 781, + 226 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 5: Chain-of-thought prompting template for data verification. $①\\&②$ Following the chain-of-thought spirit, we first give examples to LLM to boost the in-context-learning capability. After that, we ask the LLM to answer a question that is similar to the given example. Specifically, for the data verification task, we ask the LLM to verify the selected samples from the previous iteration, update the results, and explain the reason. The LLM response is instructed to be a Python-format list of dictionaries. $③$ : The response from LLM (which follows the desired Python-format list).", + "bbox": [ + 215, + 236, + 787, + 334 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Evaluation Pipeline", + "text_level": 1, + "bbox": [ + 215, + 353, + 424, + 368 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Given a set of clean data samples $\\mathcal{X} = \\{x_{i}\\}$ , perturbation functions $\\varDelta=\\{\\delta_k\\}$ , evaluation metric functions $\\mathcal{E} = \\{e_j\\}$ , and a human instruction $u$ , the objective of R²-Agent is to select an appropriate subset of clean data samples $\\mathcal{X}_u$ , potential perturbations $\\varDelta_u$ , and suitable evaluation functions $\\mathcal{E}_u$ to assess the model based on the given instruction $u$ . A desired human instruction for the R²-Agent should explicitly specify the scenario (e.g., indoor), the evaluation focus (e.g., segmentation quality), and, if applicable, any specific requirements (e.g., testing for motion blur) during the evaluation process. To enhance the robustness of the R²-Agent, we introduce a multi-agent debating strategy utilizing an operator agent $\\varPhi$ and a guardian agent $\\varPsi$ , each with their own independent memories.", + "bbox": [ + 212, + 378, + 787, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Data proposal. To propose data samples that align with the criteria delineated in the human instruction, we employ the operator agent $\\varPhi$ with the text prompts $<\\mathrm{SEL}>$ , $\\varPhi$ enables the traversal through the combined space of clean samples, pre-defined noise generation functions, and evaluation functions, collectively represented by the tuple $(\\mathcal{X},\\Delta,\\mathcal{E})$ , to select suitable samples. Additionally, we employ an LLM to generate captions $\\mathcal{C}$ for the clean samples $\\mathcal{X}$ . These captions are archived within an accessible memory to enhance the computational efficiency for subsequent stages. The data proposal stage can be formulated as:", + "bbox": [ + 212, + 544, + 787, + 667 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C}, \\mathcal {X} _ {u}, \\mathcal {E} _ {u}, \\Delta_ {u} = \\Phi (u, \\mathcal {X}, \\mathcal {E}, \\Delta , < \\mathrm {S E L} >). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 678, + 787, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Data verification. Although LLMs exhibit notable proficiency in reasoning tasks, their susceptibility to generating \"hallucinated\" content poses a significant challenge to the delivery of accurate responses. To mitigate the hallucination effects, we introduce a verification mechanism, designated as the safe guardian $\\varPsi$ , which monitors the data proposal made by the operator $\\varPhi$ through the utilization of chain-of-thought prompting [53]. We first create the proposed dataset $\\mathcal{X}_u^t$ and perturbations $\\Delta_u^t$ . $t$ denotes the iteration number. Then, we construct verification prompts $<\\mathrm{VER}>$ and employ $\\varPsi$ to validate the integrity of the data instances. Feedback from this verification is relayed to $\\varPhi$ , which then uses the information", + "bbox": [ + 212, + 703, + 789, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Xiang Li et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/be345852f4ffc4dd7e0bcc33a4bc461b27672e4ca52aba5653bf6f655ba2a5cf.jpg", + "table_caption": [ + "Table 2: Summary of considered tasks, metrics, referring modality, and datasets." + ], + "table_footnote": [], + "table_body": "
TaskRISVOSR-VOSAVSQ3M
Ref. ModalityTextImageTextAudioText
MetricsmIoUG, J, F [25]J, F [25]J, F [25]-
DatasetRefCOCO [74]/+ [74]/g [41]DAVIS [25], YTVOS [59]Ref-DAVIS [25], Ref-YTVOS [47]AVS-s4 [78], AVS-ms3 [78]ScanNet [13], ICL [19]
", + "bbox": [ + 218, + 160, + 787, + 215 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "to refine and enhance the data proposal. This iterative process is carried out until a state of convergence or until a maximum iteration number is reached, which is formulated as:", + "bbox": [ + 212, + 226, + 787, + 270 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nR ^ {t} = \\Psi \\left(\\mathcal {X} _ {u} ^ {t}, \\Delta_ {u} ^ {t}, < \\mathrm {V E R} > _ {\\Psi}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 277, + 589, + 294 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {X} _ {u} ^ {t + 1}, \\Delta_ {u} ^ {t + 1} = \\Phi (R ^ {t}, < \\mathrm {V E R} > _ {\\Phi}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 297, + 785, + 315 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $t$ is the iteration index, $R^t$ is the response of $\\varPsi$ at $t$ iteration. Note that we consider $\\varPhi$ and $\\varPsi$ have separate memories for previous inputs (similar to two separate ChatGPT sessions) and we omit the memorized inputs here for simplicity.", + "bbox": [ + 212, + 321, + 782, + 380 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Specifically, adversarial prompts are leveraged in the data verification step. In $\\langle \\mathrm{VER} \\rangle_{\\Phi}$ , we ask the LLM to select conservative choices with high confidence while, for $\\langle \\mathrm{VER} \\rangle_{\\Psi}$ , radical choices are acceptable. The adversarial prompts can proactively encourage a debate between two agents thus correcting potential mistakes in the original data proposal. To help understand the data verification step, we demonstrate an example of the data verification process in Fig. 5.", + "bbox": [ + 212, + 382, + 787, + 472 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Model analysis. With the selected samples and functions $\\mathcal{X}_u, \\mathcal{E}_u, \\Delta_u$ , we instantiate the noisy data $\\mathcal{X}_u'$ by applying perturbation functions to the clean data, $\\mathcal{X}_u' = \\Delta_u(\\mathcal{X}_u)$ . We then calculate a set of metrics $\\mathcal{M}_u = \\{m_i\\}$ that correspond to $\\mathcal{X}_u'$ by applying evaluating functions $\\mathcal{E}_u$ to noisy samples. Utilizing the chain-of-though prompt , we engage an LLM $\\Theta$ with empty memory to analyze the metrics and produce an output report $O$ for the model performance as:", + "bbox": [ + 214, + 483, + 787, + 574 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nO = \\Theta \\left(\\mathcal {M} _ {u}, \\mathcal {X} _ {u} ^ {\\prime}, < \\mathrm {A N A} >\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 580, + 785, + 595 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiment", + "text_level": 1, + "bbox": [ + 215, + 614, + 366, + 631 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Evaluation Setup and Metrics", + "text_level": 1, + "bbox": [ + 214, + 642, + 509, + 657 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In Tab. 2, we present a summary of the benchmark tasks, evaluated datasets, and corresponding metrics. Our benchmark encompasses five prevalent tasks, namely, RIS, VOS, R-VOS, AVS and Q3M.", + "bbox": [ + 212, + 662, + 787, + 709 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation setup. We conduct systematic evaluations to analyze the impact of perturbations on state-of-the-art mode five tasks. Both perturbations in the referring guidance and visual frames are considered in the benchmarking. We create noisy datasets with low, medium and high noise levels for each task. For video-level tasks, we additionally consider the scenario that noises change over time. We first benchmark the models' general performance with all types of noises and then investigate the impact of noises individually. More information about the dataset creation is available in the Appendix.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Benchmarking the Robustness of Referring Perception Models under Perturbations", + "bbox": [ + 215, + 114, + 767, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 810, + 114, + 821, + 127 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/9b7b17d60674959b454a72fc8fd02b132fb4d6f95809e0c74dded013e6a0c45a.jpg", + "table_caption": [ + "Table 3: Performance of referring image segmentation methods under low, medium, and high perturbation levels. Average performance change (APC) is averaged among all perturbation levels. We report mIoU following convention." + ], + "table_footnote": [], + "table_body": "
MethodVenueRefCOCORefCOCO+RefCOCOg
CleanLowMed.HighAPCCleanLowMed.HighAPCCleanLowMed.HighAPC
LAVT [69]CVPR 220.740.580.530.48-0.210.660.580.510.43-0.150.630.550.470.38-0.16
PolyFormer [37]CVPR 230.760.700.640.57-0.120.710.650.570.50-0.140.690.640.570.50-0.12
X-Decoder [82]CVPR 230.680.600.520.44-0.160.690.600.510.44-0.170.640.580.500.44-0.14
ETRIS [63]ICCV 230.710.640.570.49-0.140.600.560.480.40-0.120.600.530.450.38-0.15
SEEM [84]NeurIPS 230.650.590.540.47-0.120.640.580.510.44-0.130.620.560.490.43-0.12
", + "bbox": [ + 218, + 185, + 790, + 256 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/0212cd1f808b82c5e280cef7eea7f7cec5c9d3ea81c84af4e6d2c7c33346bff3.jpg", + "table_caption": [ + "Table 4: Performance of video object segmentation methods under low, medium, and high perturbation levels. APC is averaged among all perturbation levels." + ], + "table_footnote": [], + "table_body": "
MethodVenueDAVISYTVOS
CleanLowMediumHighAPCCleanLowMediumHighAPC
JFJFJFJJGsGuGsGuGsGu
AOT [70]NeurIPS 2176.083.071.779.269.776.965.371.1-7.1-7.386.484.284.980.982.078.176.170.4-5.4-7.7
DeAOT [71]NeurIPS 2276.984.573.281.072.480.468.275.5-5.6-6.486.784.586.683.784.581.880.876.8-2.7-3.7
XMem [11]ECCV 2277.484.572.280.071.879.768.474.7-6.6-6.486.584.584.081.281.879.078.674.9-5.1-6.1
DEVA [10]ICCV 2378.785.976.484.974.282.469.977.5-5.2-4.387.283.986.782.884.780.080.676.3-3.2-4.2
Cutie [9]CVPR 2480.687.775.883.575.082.769.375.4-7.2-7.287.884.586.582.884.280.580.375.1-4.1-5.0
", + "bbox": [ + 220, + 289, + 787, + 369 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/717cfb7530244a8033d07fc898fb4dd1cd074eef8b790fb1cfb2ab387712bdf6.jpg", + "table_caption": [ + "Table 5: Performance of referring video object segmentation methods under low, medium, and high perturbation levels. APC is averaged among all perturbation levels." + ], + "table_footnote": [], + "table_body": "
MethodVenueRef-DAVISRef-YTVOS
CleanLowMediumHighAPCCleanLowMediumHighAPC
J FJ FJ FJ FJ FJ FJ FJ FJ FJ FJ FJ F
Video-level Backbone
MTTR [5]CVPR 22------54.056.448.951.643.045.5
SgMg [42]ICCV 2359.064.857.362.651.957.7-6.8-7.757.760.052.454.5
Image-level Backbone
ReferFormer [55]CVPR 2255.861.346.550.143.148.038.541.9-13.1-14.654.856.5
OnlineRefer [54]ICCV 2355.762.951.657.948.054.342.246.3-8.4-10.055.658.9
R2-VOS [33]ICCV 2357.262.450.257.046.553.839.244.6-11.9-10.656.158.4
", + "bbox": [ + 220, + 402, + 790, + 494 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Metrics. Following the convention [41, 74], we leverage mIoU to evaluate the RIS task. For video-level tasks, the convention is to compute region similarity $\\mathcal{I}$ and contour accuracy $\\mathcal{F}$ as defined in [45]. Specifically, $\\mathcal{G} = \\frac{\\mathcal{I} + \\mathcal{F}}{2}$ is also used. In addition, to better demonstrate the performance degradation against perturbations, we define average performance change (APC) as $\\mathrm{APC} = \\sum_{i} \\frac{m_i^n - m_i^c}{N}$ where $m_i^n$ and $m_i^c$ denotes the performance of the $i$ -th sample of perturbed and clean data respectively. $N$ is the sample number. Additional results, evaluated using various metrics and settings, can be found in the Appendix.", + "bbox": [ + 212, + 513, + 787, + 638 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Performance Benchmarking", + "text_level": 1, + "bbox": [ + 212, + 661, + 493, + 678 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Referring image segmentation. As shown in Tab. 3, we evaluate the state-of-the-art RIS methods on perturbed RefCOCO $/ + / \\mathrm{g}$ [74] datasets. We notice that PolyFormer [37] achieves the best performance in terms of both performance and robustness. Different from other methods that generate pixel-level classification results to represent an object mask, PolyFormer utilizes multi-polygon vertices to represent the mask. We consider the robustness can result from the special sequential prediction and polygon representation which eases the reliance on the per-pixel representation. In addition, we notice that though SEEM [83] shows an inferior performance, its robustness to perturbations is promising. We consider that multi-task joint training can account for SEEM's robustness.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Xiang Li et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/096939eb02b38372271602d393300e1d457c327ee2faae0e79c3ee0fef8843dc.jpg", + "table_caption": [ + "Table 6: Performance of audiovisual segmentation methods under low, medium, and high perturbation levels. APC is averaged among all perturbation levels." + ], + "table_footnote": [], + "table_body": "
MethodVenueAVS-s4AVS-ms3
CleanLowMediumHighAPCCleanLowMediumHighAPC
J FJ FJ FJ FJ FJ FJ FJ FJ FJ FJ FJ F
AVS [79]ECCV 2272.8 84.868.4 79.760.8 71.255.6 66.8-11.2 -12.247.9 57.844.4 54.641.2 50.537.0 44.2-7.0 -8.0
CATR [27]MM 2374.9 87.171.1 84.265.7 79.058.4 72.4-9.8 -8.653.1 65.649.8 61.446.7 58.441.2 52.9-7.2 -8.0
AVSegFormer [17]AAAI 2376.4 86.770.9 82.362.4 74.857.3 69.8-12.9 -11.049.5 62.846.9 59.043.7 55.439.8 50.1-6.0 -8.0
QSD [35]CVPR 2477.6 85.674.8 83.668.3 78.459.4 71.2-10.1 -7.961.8 64.356.6 60.053.1 57.147.6 51.6-9.4 -8.0
", + "bbox": [ + 218, + 174, + 787, + 241 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/9cc248f997b73f32dd2a24b536b0b936ec18bc1c2026315e597823a55572b226.jpg", + "image_caption": [ + "Fig. 6: Queryable 3D mapping of ConceptFusion [23] with inaccurate referring guidance on ScanNet [13] and ICL [19] datasets. Due to the absence of publicly available evaluation codes, we did not provide quantitative evaluation for queryable 3D mapping." + ], + "image_footnote": [], + "bbox": [ + 223, + 244, + 784, + 401 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Video object segmentation. As depicted in Tab. 4, we demonstrate the performance of state-of-the-art methods on DAVIS [25] and YTVOS [59] datasets. We notice that, for simple scenarios in DAVIS datasets, Cutie [9] shows the best performance. While in the more complex YTVOS dataset, DeAOT [71] achieves the best. Cutie heavily relies on the pixel- and object-level correspondence across frames which can be disrupted by the perturbations. Differently, DeAOT leverages decoupled visual and ID embeddings which may be the reason for its robustness in noisy scenarios. In addition, we notice that the unseen categories generally have a larger performance drop compared to the seen categories which can impose more challenges in practical deployment in complex scenarios.", + "bbox": [ + 212, + 468, + 787, + 619 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Referring video object segmentation. We benchmark the referring video object segmentation task in Tab. 5. SgMg [42] and OnlineFormer [54] achieve the best performance and robustness among methods equipped with video- and image-level backbones respectively. Unlike other methods, OnlineRefer processes the visual features in a frame-by-frame manner which makes it rely less on the pixel-level temporal correspondence that is easily been disrupted by visual perturbations.", + "bbox": [ + 212, + 631, + 787, + 723 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Audio-visual segmentation. We show the performance of popular AVS methods in Tab. 6. We notice that QSD [34] and CATR [27] archive promising performance and robustness among all methods. CATR [27] and AVSegFormer [17] are similar methods that leverage acoustic query to query the visual frames with a transformer-based structure. QSD moves one step forward which additionally introduces a quantization operation and a local-global distillation to enhance the robustness of the acoustic representation to adapt to complex scenarios.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Benchmarking the Robustness of Referring Perception Models under Perturbations", + "bbox": [ + 215, + 114, + 767, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 803, + 116, + 820, + 127 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c237184825d1083df4354938ef2e717843c0572d78365cb3846347f120f9a9ba.jpg", + "image_caption": [ + "Fig. 7: Average performance changes (APC) for each perturbation type." + ], + "image_footnote": [], + "bbox": [ + 218, + 147, + 787, + 247 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Queryable 3D mapping. Recent queryable 3D mapping methods [23,64] typically leverage the results from 2D referring models and fuse them to a 3D map. Thereby, the failure of 2D referring perception can directly obstacle the success of 3D tasks. As shown in Fig. 6, we visualize the state-of-the-art queryable 3D mapping method, ConceptFusion [23]. We notice that with the accurate text query \"orange sofa\" ConceptFusion can successfully locate the object while when misspell occur in the query, an obvious performance degradation can be observed.", + "bbox": [ + 212, + 282, + 787, + 388 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Perturbation Analysis", + "text_level": 1, + "bbox": [ + 215, + 409, + 444, + 425 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Perturbation type. We conduct ablation studies on performance degradation against different perturbation types. Fig. 7 illustrates the aggregated average performance deviation among all benchmarked methods under visual, textual, and acoustic perturbations, respectively. For visual perturbations, we notice the frost effect (FT) induced the most substantial performance degradation, which is likely attributable to its extensive occlusion and distortion effects in the whole image. Conversely, color-based perturbations, such as brightness (BR) and saturate (SA), exhibit marginal impact on performance, possibly due to the preservation of shape information. For textual perturbations, our analysis indicated that all types of perturbations lead to comparable levels of degradation. The character missing (CA) has a slightly greater impact. For acoustic perturbations, we notice that most perturbation types only show a marginal impact on the performance. We consider this is because the target objects in AVS-s4/ms3 are salient objects, allowing models to rely predominantly on visual cues for object localization, even when the acoustic guidance lacks precision. Notably, the impulse response perturbation was associated with a significant performance drop, which could stem from the failure in feature extraction due to the impulse signal. Such abnormal acoustic features could potentially disrupt subsequent multimodal interactions.", + "bbox": [ + 212, + 433, + 789, + 705 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Perturbation correlation. To delve deeper into the nature of performance degradation, we computed correlation matrix for each modality, utilizing the Pearson product-moment correlation coefficients calculated from the performance under various perturbation types to populate the matrix. Fig. 8 displays the correlation matrix for textual perturbations.", + "bbox": [ + 212, + 719, + 540, + 839 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/60a3347f9f1ab8bf5ca7746429f2e861b828a20e2c72d2c53084317b5cb0387e.jpg", + "image_caption": [ + "Fig.8: Correlation of performance of textual perturbations." + ], + "image_footnote": [], + "bbox": [ + 558, + 728, + 772, + 805 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Xiang Li et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e1cd3cc6389dabacc940cfc734ab581c38a6bbb23004d68fd6876f9283982f06.jpg", + "image_caption": [ + "Fig. 9: Correlation matrix of degraded performance with visual and acoustic perturbations. We concatenate the sample-level performance of all methods on all testing datasets as a feature vector to calculate Pearson product-moment correlation." + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 498, + 258 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/cddd06cdc3252f8c855438348b3ba94245967bf0df35307f042789f907d77939.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 146, + 782, + 258 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Despite the average performance changes (in Fig. 7) being similar across different types, the low correlation between the types of performance degradation suggest that each perturbation affects the model in a distinct manner. The correlations for visual perturbations and the ones for acoustic perturbations are presented in Fig. 9 respectively. For visual perturbations, we notice several instances of high correlations. For instance, the group of blurring perturbations—defocus blur (DB), gaussian blur (GB), motion blur (MB), and glass blur (GS)—are highly correlated with one another. For the acoustic perturbations, impulse response (IR) and highpass filter (HF) exhibit different patterns compared to other perturbations, indicating unique effects on performance degradation. We systematically explored the correlations among perturbations and hope the analysis can provide insights for wider evaluation in the future.", + "bbox": [ + 212, + 321, + 787, + 503 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Static v.s. Dynamic. To analyze the dynamic changing of perturbation types in videos, we conduct an ablation study to explore the impact of perturbation changing across frames (for visual perturbation only). As shown in Fig. 10, we calculate the APC across all models on the Ref-DAVIS dataset. We notice the model performance", + "bbox": [ + 212, + 518, + 517, + 638 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b8aed8f2de4d606a0706778fc314252ef1b3bce727c5c9b9444d46df79dd6d7b.jpg", + "image_caption": [ + "Fig.10: Average performance change (APC) under static and dynamic perturbations." + ], + "image_footnote": [], + "bbox": [ + 531, + 521, + 774, + 595 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "decreases more with dynamic perturbations which can be due to the lack of temporal consistency resulting from various perturbation types.", + "bbox": [ + 212, + 640, + 787, + 671 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 $\\mathbf{R}^2$ -Agent: Evaluation Assistant", + "text_level": 1, + "bbox": [ + 214, + 691, + 521, + 709 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "User study of data selection. To evaluate the data selection performance of $\\mathbf{R}^2$ -Agent, we manually annotate an evaluation set. Since the instruction can be ambiguous in practice, we measure the alignment between $\\mathbf{R}^2$", + "bbox": [ + 214, + 719, + 486, + 808 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/3365ad42d51f4858f8e7fe9a93a019465a8c5cb1c26b5ba47ff3f0ddf9075b8f.jpg", + "table_caption": [ + "Table 7: Data selection performance." + ], + "table_footnote": [], + "table_body": "
# Ver. Iter.X123
ScoreACC RateACC RateACC RateACC Rate
Data Sample0.530.780.640.810.770.850.800.91
Perturbation0.650.870.770.860.860.910.870.93
Metric Func.1.001.001.001.001.001.001.001.00
", + "bbox": [ + 498, + 738, + 787, + 799 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Agent and human by (1) ACC: accuracy between the human-annotated samples and $\\mathbb{R}^2$ -Agent predictions and (2) Rate: rating the rationality of the prediction", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Benchmarking the Robustness of Referring Perception Models under Perturbations", + "bbox": [ + 215, + 114, + 767, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 803, + 114, + 821, + 127 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/471a469cc966a399caa11d51911ca264628e3f0602574063cd532553e5943044.jpg", + "image_caption": [ + "Fig. 11: Example of $\\mathbf{R}^2$ -Agent report. The evaluation setting and metrics are fed to $\\mathbf{R}^2$ -Agent with a set of hard-coded prompts to facilitate $\\mathbf{R}^2$ -Agent generate the human-like report. Beyond the hard-coded analysis, further QA is also supported." + ], + "image_footnote": [], + "bbox": [ + 225, + 145, + 354, + 250 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/bbced268dc80b1462669191b6abd5884fb9527b15c09dabad45b08796acb36e4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Scenario:Scenario: Car driving scenario with a fast motion\nPerturbation: snow (SN), fog (SG), frost (FT), spatter (SP), motion blur (MB),...
Evaluation Metric: mIoU (mean Intersection over Union)
Analysis:IoU Change: The spatter has a largest -12.1 change in mIoU suggesting that the high level of spatter effect has negatively impacted the model's ability. The spatter effect can lead to 1. Boundary Ambiguity...2. Fragmentation...3. Texture change...ForJPEG compression and animation have a marginal effect...
Case Analysis: The case with caption: “a red car is far away on the left” suffers severe performance drop under...The small size of the car may account for the...
Recommendation: Improve the resilience of models to compression artifacts through enhanced training, including a wider variety of...
", + "bbox": [ + 361, + 148, + 781, + 251 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "from $\\mathbf{R}^2$ -Agent (1 if reasonable else 0). As shown in Tab. 7, we ablate on the data verification iteration number. For the metrics function selection, since there are only several hard-coded metrics that can be selected, the accuracy remains perfect while, for data sample and perturbation, we notice a direct performance gain when increasing the iteration number. We notice the performance becomes saturated after 2 iterations, thus we leverage iteration number 2 as the final design choice. More details can be found in the Appendix.", + "bbox": [ + 212, + 325, + 787, + 434 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Automatic model analysis. With the evaluation setting and corresponding results, $\\mathbf{R}^2$ -Agent can generate a detailed evaluation report. We demonstrate an evaluation report example in Fig. 11. To make $\\mathbf{R}^2$ -Agent think in a human-like manner, we leverage a set of chain-of-thought prompts to instruct $\\mathbf{R}^2$ -Agent to give reasonable and detailed analysis based on the given evaluation results. We notice that $\\mathbf{R}^2$ -Agent can understand the evaluation metrics and give a human-like analysis which potentially helps to reduce the cost of model evaluation in practice. Furthermore, given the strong dialogue capability of LLMs, further question-answering about the results and report is also feasible. We demonstrate more qualitative results in the Appendix.", + "bbox": [ + 212, + 444, + 789, + 595 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 614, + 359, + 630 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we introduced $\\mathbb{R}^2$ -Bench, a comprehensive benchmark for evaluating the robustness of visual referring perception models against various perturbations. Our contributions include a detailed taxonomy of perturbations, a customizable perturbation synthesis toolbox, systematic perturbation analysis, and $\\mathbb{R}^2$ -Agent, a novel evaluation assistant based on large language models. Through extensive experiments, we observed and analyzed the intrinsic characteristics of different perturbations to current models and highlighted the importance of robustness in referring perception tasks. A benchmark with 5 popular referring tasks is also provided to facilitate future research. In addition, the data construction and analysis of robustness evaluation can be further simplified with our proposed $\\mathbb{R}^2$ -Agent. Our findings underscore the necessity for robustness in the deployment of intelligent systems in real-world scenarios, offering a foundation for future advancements in the field.", + "bbox": [ + 212, + 643, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Xiang Li et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023) 7", + "2. Ahn, H., Kwon, O., Kim, K., Jeong, J., Jun, H., Lee, H., Lee, D., Oh, S.: Visually grounding language instruction for history-dependent manipulation. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 675-682. IEEE (2022) 2", + "3. Arandjelovic, R., Zisserman, A.: Look, listen and learn. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 609-617 (2017) 4", + "4. Arandjelovic, R., Zisserman, A.: Objects that sound. In: Proceedings of the European conference on computer vision (ECCV). pp. 435-451 (2018) 4", + "5. Botach, A., Zheltonozhskii, E., Baskin, C.: End-to-end referring video object segmentation with multimodal transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4985-4995 (2022) 4, 10", + "6. Chen, F., Zhang, H., Hu, K., Huang, Y.k., Zhu, C., Savvides, M.: Enhanced training of query-based object detection via selective query recollection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23756-23765 (2023) 2", + "7. Chen, F., Zhang, H., Li, Z., Dou, J., Mo, S., Chen, H., Zhang, Y., Ahmed, U., Zhu, C., Savvides, M.: Unitail: Detecting, reading, and matching in retail scene (2022), https://arxiv.org/abs/2204.00298 2", + "8. Chen, F., Zhang, H., Yang, Z., Chen, H., Hu, K., Savvides, M.: Rtgen: Generating region-text pairs for open-vocabulary object detection (2024), https://arxiv.org/abs/2405.19854 2", + "9. Cheng, H.K., Oh, S.W., Price, B., Lee, J.Y., Schwing, A.: Putting the object back into video object segmentation. arXiv preprint arXiv:2310.12982 (2023) 4, 10, 11", + "0. Cheng, H.K., Oh, S.W., Price, B., Schwing, A., Lee, J.Y.: Tracking anything with decoupled video segmentation. In: ICCV (2023) 10", + "1. Cheng, H.K., Schwing, A.G.: Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In: European Conference on Computer Vision. pp. 640-658. Springer (2022) 4, 10", + "2. Cheng, Y., Wang, R., Pan, Z., Feng, R., Zhang, Y.: Look, listen, and attend: Co-attention network for self-supervised audio-visual representation learning. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 3884-3892 (2020) 4", + "3. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017) 9, 11", + "4. Ding, H., Liu, C., He, S., Jiang, X., Loy, C.C.: Mevis: A large-scale benchmark for video segmentation with motion expressions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2694-2703 (2023) 2, 4", + "5. Ding, H., Liu, C., Wang, S., Jiang, X.: Vision-language transformer and query generation for referring segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16321-16330 (2021) 4", + "6. Du, Y., Li, S., Torralba, A., Tenenbaum, J.B., Mordatch, I.: Improving factuality and reasoning in language models through multiagent debate. arXiv preprint arXiv:2305.14325 (2023) 3" + ], + "bbox": [ + 225, + 178, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Benchmarking the Robustness of Referring Perception Models under Perturbations", + "bbox": [ + 215, + 114, + 767, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 803, + 114, + 820, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Gao, S., Chen, Z., Chen, G., Wang, W., Lu, T.: Avsegformer: Audio-visual segmentation with transformer. arXiv preprint arXiv:2307.01146 (2023) 2, 11", + "18. Han, M., Wang, Y., Li, Z., Yao, L., Chang, X., Qiao, Y.: Hybrid temporal-scale multimodal learning framework for referring video object segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13414-13423 (2023) 4", + "19. Handa, A., Whelan, T., McDonald, J., Davison, A.: A benchmark for RGB-D visual odometry, 3D reconstruction and SLAM. In: IEEE Intl. Conf. on Robotics and Automation, ICRA. Hong Kong, China (May 2014) 9, 11", + "20. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261 (2019) 6", + "21. Hu, Y., Lin, F., Zhang, T., Yi, L., Gao, Y.: Look before you leap: Unveiling the power of gpt-4v in robotic vision-language planning. arXiv preprint arXiv:2311.17842 (2023) 2", + "22. Huang, W., Xia, F., Shah, D., Driess, D., Zeng, A., Lu, Y., Florence, P., Mordatch, I., Levine, S., Hausman, K., et al.: Grounded decoding: Guiding text generation with grounded models for robot control. arXiv preprint arXiv:2303.00855 (2023) 2", + "23. Jatavallabhula, K.M., Kuwajerwala, A., Gu, Q., Omama, M., Chen, T., Maalouf, A., Li, S., Iyer, G., Saryazdi, S., Keetha, N., et al.: Conceptfusion: Open-set multimodal 3d mapping. arXiv preprint arXiv:2302.07241 (2023) 2, 4, 11, 12", + "24. Ke, L., Ye, M., Danelljan, M., Liu, Y., Tai, Y.W., Tang, C.K., Yu, F.: Segment anything in high quality. In: NeurIPS (2023) 4", + "25. Khoreva, A., Rohrbach, A., Schiele, B.: Video object segmentation with language referring expressions. In: ACCV (2018) 9, 11", + "26. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) 4", + "27. Li, K., Yang, Z., Chen, L., Yang, Y., Xun, J.: Catr: Combinatorial-dependence audio-queried transformer for audio-visual video segmentation. arXiv preprint arXiv:2309.09709 (2023) 4, 11", + "28. Li, M., Li, S., Zhang, X., Zhang, L.: Univs: Unified and universal video segmentation with prompts as queries. arXiv preprint arXiv:2402.18115 (2024) 4", + "29. Li, X., Cao, H., Zhao, S., Li, J., Zhang, L., Raj, B.: Panoramic video salient object detection with ambisonic audio guidance. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1424-1432 (2023) 2", + "30. Li, X., Lin, C.C., Chen, Y., Liu, Z., Wang, J., Singh, R., Raj, B.: Paintseg: Painting pixels for training-free segmentation. In: Thirty-seventh Conference on Neural Information Processing Systems (2023) 2, 4", + "31. Li, X., Wang, J., Li, X., Lu, Y.: Hybrid instance-aware temporal fusion for online video instance segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 1429-1437 (2022) 2", + "32. Li, X., Wang, J., Li, X., Lu, Y.: Video instance segmentation by instance flow assembly. IEEE Transactions on Multimedia 25, 7469-7479 (2022) 2", + "33. Li, X., Wang, J., Xu, X., Li, X., Raj, B., Lu, Y.: Robust referring video object segmentation with cyclic structural consensus. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22236-22245 (2023) 2, 10", + "34. Li, X., Wang, J., Xu, X., Peng, X., Singh, R., Lu, Y., Raj, B.: Towards robust audiovisual segmentation in complex environments with quantization-based semantic decomposition. arXiv preprint arXiv:2310.00132 (2023) 2, 4, 11" + ], + "bbox": [ + 215, + 146, + 787, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Xiang Li et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Li, X., Wang, J., Xu, X., Yang, M., Yang, F., Zhao, Y., Singh, R., Raj, B.: Towards noise-tolerant speech-referring video object segmentation: Bridging speech and text. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. pp. 2283-2296 (2023) 4, 11", + "36. Liang, T., He, Z., Jiao, W., Wang, X., Wang, Y., Wang, R., Yang, Y., Tu, Z., Shi, S.: Encouraging divergent thinking in large language models through multi-agent debate. arXiv preprint arXiv:2305.19118 (2023) 3", + "37. Liu, J., Ding, H., Cai, Z., Zhang, Y., Satzoda, R.K., Mahadevan, V., Manmatha, R.: Polyformer: Referring image segmentation as sequential polygon generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18653-18663 (2023) 2, 4, 10", + "38. Liu, J., Wang, Y., Ju, C., Ma, C., Zhang, Y., Xie, W.: Annotation-free audio-visual segmentation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5604-5614 (2024) 4", + "39. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023) 4", + "40. Liu, S., Hasan, A., Hong, K., Wang, R., Chang, P., Mizrachi, Z., Lin, J., McPherson, D.L., Rogers, W.A., Driggs-Campbell, K.: Dragon: A dialogue-based robot for assistive navigation with visual language grounding. IEEE Robotics and Automation Letters (2024) 2", + "41. Mao, J., Huang, J., Toshev, A., Camburu, O., Yuille, A.L., Murphy, K.: Generation and comprehension of unambiguous object descriptions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 11-20 (2016) 9, 10", + "42. Miao, B., Bennamoun, M., Gao, Y., Mian, A.: Spectrum-guided multi-granularity referring video object segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 920-930 (2023) 4, 10, 11", + "43. Mo, S., Morgado, P.: A closer look at weakly-supervised audio-visual source localization. arXiv preprint arXiv:2209.09634 (2022) 4", + "44. Pan, W., Shi, H., Zhao, Z., Zhu, J., He, X., Pan, Z., Gao, L., Yu, J., Wu, F., Tian, Q.: Wnet: Audio-guided video object segmentation via wavelet-based cross-modal denoising networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1320-1331 (2022) 4", + "45. Pont-Tuset, J., Peruzzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675 (2017) 10", + "46. Senocak, A., Oh, T.H., Kim, J., Yang, M.H., Kweon, I.S.: Learning to localize sound source in visual scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4358-4366 (2018) 4", + "47. Seo, S., Lee, J.Y., Han, B.: Urvos: Unified referring video object segmentation network with a large-scale benchmark. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16. pp. 208-223. Springer (2020) 4, 9", + "48. Sun, J., Huang, D.A., Lu, B., Liu, Y.H., Zhou, B., Garg, A.: Visually-grounded planning with transformers in procedural tasks. IEEE Robotics and Automation Letters 7(2), 4924-4930 (2022) 2", + "49. Tang, J., Zheng, G., Yang, S.: Temporal collection and distribution for referring video object segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15466-15476 (2023) 4" + ], + "bbox": [ + 215, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Benchmarking the Robustness of Referring Perception Models under Perturbations", + "bbox": [ + 215, + 114, + 767, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 803, + 114, + 821, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "50. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023) 7", + "51. Tziafas, G., Kasaei, H.: Few-shot visual grounding for natural human-robot interaction. In: 2021 IEEE International Conference on Autonomous Robot Systems and Competitions (ICARSC). pp. 50-55. IEEE (2021) 2", + "52. Wang, Z., Cai, S., Chen, G., Liu, A., Ma, X.S., Liang, Y.: Describe, explain, plan and select: interactive planning with llms enables open-world multi-task agents. Advances in Neural Information Processing Systems 36 (2024) 7", + "53. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022) 8", + "54. Wu, D., Wang, T., Zhang, Y., Zhang, X., Shen, J.: Onlinerefer: A simple online baseline for referring video object segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2761-2770 (2023) 2, 4, 10, 11", + "55. Wu, J., Jiang, Y., Sun, P., Yuan, Z., Luo, P.: Language as queries for referring video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4974-4984 (2022) 4, 10", + "56. Wu, J., Jiang, Y., Yan, B., Lu, H., Yuan, Z., Luo, P.: Segment every reference object in spatial and temporal spaces. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2538-2550 (2023) 5", + "57. Xi, Z., Chen, W., Guo, X., He, W., Ding, Y., Hong, B., Zhang, M., Wang, J., Jin, S., Zhou, E., et al.: The rise and potential of large language model based agents: A survey. arXiv preprint arXiv:2309.07864 (2023) 7", + "58. Xiong, Y., Varadarajan, B., Wu, L., Xiang, X., Xiao, F., Zhu, C., Dai, X., Wang, D., Sun, F., Iandola, F., et al.: Efficientsam: Leveraged masked image pretraining for efficient segment anything. arXiv preprint arXiv:2312.00863 (2023) 2, 4", + "59. Xu, N., Yang, L., Fan, Y., Yue, D., Liang, Y., Yang, J., Huang, T.: Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327 (2018) 9, 11", + "60. Xu, X., Wang, J., Li, X., Lu, Y.: Reliable propagation-correction modulation for video object segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 2946-2954 (2022) 4", + "61. Xu, X., Wang, J., Ming, X., Lu, Y.: Towards robust video object segmentation with adaptive object calibration. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 2709-2718 (2022) 6", + "62. Xu, X., Zhang, T., Wang, S., Li, X., Chen, Y., Li, Y., Raj, B., Johnson-Roberson, M., Huang, X.: Customizable perturbation synthesis for robust slam benchmarking. arXiv preprint arXiv:2402.08125 (2024) 6", + "63. Xu, Z., Chen, Z., Zhang, Y., Song, Y., Wan, X., Li, G.: Bridging vision and language encoders: Parameter-efficient tuning for referring image segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17503-17512 (2023) 4, 10", + "64. Yamazaki, K., Hanyu, T., Vo, K., Pham, T., Tran, M., Doretto, G., Nguyen, A., Le, N.: Open-fusion: Real-time open-vocabulary 3d mapping and queryable scene representation. arXiv preprint arXiv:2310.03923 (2023) 4, 12", + "65. Yamazaki, K., Truong, S., Vo, K., Kidd, M., Rainwater, C., Luu, K., Le, N.: Vlcap: Vision-language with contrastive learning for coherent video paragraph captioning. In: 2022 IEEE International Conference on Image Processing (ICIP). pp. 3656-3661. IEEE (2022) 4" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Xiang Li et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "66. Yamazaki, K., Vo, K., Truong, Q.S., Raj, B., Le, N.: Vlnt: Visual-linguistic transformer-in-transformer for coherent video paragraph captioning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 3081-3090 (2023) 4", + "67. Yang, J., Gao, M., Li, Z., Gao, S., Wang, F., Zheng, F.: Track anything: Segment anything meets videos (2023) 4", + "68. Yang, Z., Wang, J., Tang, Y., Chen, K., Zhao, H., Torr, P.H.: Levt: Language-aware vision transformer for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18155-18165 (2022) 2, 4", + "69. Yang, Z., Wang, J., Tang, Y., Chen, K., Zhao, H., Torr, P.H.: Levt: Language-aware vision transformer for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18155-18165 (2022) 10", + "70. Yang, Z., Wei, Y., Yang, Y.: Associating objects with transformers for video object segmentation. Advances in Neural Information Processing Systems 34, 2491-2502 (2021) 4, 10", + "71. Yang, Z., Yang, Y.: Decoupling features in hierarchical propagation for video object segmentation. Advances in Neural Information Processing Systems 35, 36324-36336 (2022) 4, 10, 11", + "72. Yang, Z., Yang, Y.: Decoupling features in hierarchical propagation for video object segmentation. Advances in Neural Information Processing Systems 35, 36324-36336 (2022) 5", + "73. Yao, J., Wang, X., Ye, L., Liu, W.: Matte anything: Interactive natural image matting with segment anything models. arXiv preprint arXiv:2306.04121 (2023) 4", + "74. Yu, L., Poirson, P., Yang, S., Berg, A.C., Berg, T.L.: Modeling context in referring expressions. In: European Conference on Computer Vision. pp. 69-85. Springer (2016) 9, 10", + "75. Yu, T., Feng, R., Feng, R., Liu, J., Jin, X., Zeng, W., Chen, Z.: Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790 (2023) 4", + "76. Zhang, J., Cui, Y., Wu, G., Wang, L.: Joint modeling of feature, correspondence, and a compressed memory for video object segmentation. arXiv preprint arXiv:2308.13505 (2023) 5", + "77. Zhao, Q., Wang, J., Zhang, Y., Jin, Y., Zhu, K., Chen, H., Xie, X.: Competeai: Understanding the competition behaviors in large language model-based agents. arXiv preprint arXiv:2310.17512 (2023) 3", + "78. Zhou, J., Shen, X., Wang, J., Zhang, J., Sun, W., Zhang, J., Birchfield, S., Guo, D., Kong, L., Wang, M., et al.: Audio-visual segmentation with semantics. arXiv preprint arXiv:2301.13190 (2023) 4, 9", + "79. Zhou, J., Wang, J., Zhang, J., Sun, W., Zhang, J., Birchfield, S., Guo, D., Kong, L., Wang, M., Zhong, Y.: Audio-visual segmentation. In: European Conference on Computer Vision (2022) 2, 4, 11", + "80. Zhu, C., Chen, F., Ahmed, U., Shen, Z., Savvides, M.: Semantic relation reasoning for shot-stable few-shot object detection. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 8778-8787 (2021), https://api.sementicscholar.org/CorpusID:232093016 2", + "81. Zhu, C., Chen, F., Shen, Z., Savvides, M.: Soft anchor-point object detection. In: European Conference on Computer Vision (2019), https://api-semanticscholar.org/CorpusID:208512715 2" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Benchmarking the Robustness of Referring Perception Models under Perturbations", + "bbox": [ + 215, + 114, + 767, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 803, + 114, + 821, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "82. Zou, X., Dou, Z.Y., Yang, J., Gan, Z., Li, L., Li, C., Dai, X., Behl, H., Wang, J., Yuan, L., et al.: Generalized decoding for pixel, image, and language. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15116-15127 (2023) 4, 10", + "83. Zou, X., Yang, J., Zhang, H., Li, F., Li, L., Gao, J., Lee, Y.J.: Segment everything everywhere all at once. arXiv preprint arXiv:2304.06718 (2023) 4, 10", + "84. Zou, X., Yang, J., Zhang, H., Li, F., Li, L., Wang, J., Wang, L., Gao, J., Lee, Y.J.: Segment everything everywhere all at once. Advances in Neural Information Processing Systems 36 (2024) 4, 10" + ], + "bbox": [ + 215, + 146, + 787, + 272 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "Xiang Li et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/2024/R^2-Bench_ Benchmarking the Robustness of Referring Perception Models under Perturbations/3cf440ac-b4a0-433a-943f-d4d176a64132_model.json b/2024/R^2-Bench_ Benchmarking the Robustness of Referring Perception Models under Perturbations/3cf440ac-b4a0-433a-943f-d4d176a64132_model.json new file mode 100644 index 0000000000000000000000000000000000000000..93d2aa0e25154587d08bf247243afcea62293343 --- /dev/null +++ b/2024/R^2-Bench_ Benchmarking the Robustness of Referring Perception Models under Perturbations/3cf440ac-b4a0-433a-943f-d4d176a64132_model.json @@ -0,0 +1,2814 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.785, + 0.19 + ], + "angle": 0, + "content": "\\(\\mathbf{R}^2\\)-Bench: Benchmarking the Robustness of Referring Perception Models under Perturbations" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.215, + 0.754, + 0.248 + ], + "angle": 0, + "content": "Xiang Li\\(^{1}\\), Kai Qiu\\(^{1}\\), Jinglu Wang\\(^{2}\\), Xiaohao Xu\\(^{3}\\), Rita Singh\\(^{1}\\), Kashu Yamazaki\\(^{1}\\), Hao Chen\\(^{1}\\), Xiaonan Huang\\(^{3}\\), Bhiksha Raj\\(^{1,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.257, + 0.257, + 0.744, + 0.273 + ], + "angle": 0, + "content": "1 CMU, 2 Microsoft Research Asia, 3 University of Michigan, 4 MBZUAI" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.283, + 0.495, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.279, + 0.407, + 0.452, + 0.419 + ], + "angle": 0, + "content": "(a) Referring Perception in Practice" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.283, + 0.787, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.563, + 0.407, + 0.751, + 0.419 + ], + "angle": 0, + "content": "(b) Perturbation in Referring Guidance" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.423, + 0.789, + 0.549 + ], + "angle": 0, + "content": "Fig. 1: Motivation illustration. Referring perception models (RPMs) empower intelligent systems with their ability to perform object grounding within the environment based on referring guidance, such as textual descriptions, imagery exemplars, or auditory signals associated with the target object. However, RPMs' performance can be compromised by disturbances in real-world scenarios, such as environmental noise (e.g., extraneous sounds from a nearby radio), human-induced errors (e.g., typographical errors in textual input), and limitations in the sensor (e.g., motion blur in images). Conducting a rigorous analysis of RPMs' robustness to a wide array of perturbations is necessary for building reliable real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.564, + 0.744, + 0.814 + ], + "angle": 0, + "content": "Abstract. Referring perception, which aims at grounding visual objects with multimodal referring guidance, is essential for bridging the gap between humans, who provide instructions, and the environment where intelligent systems perceive. Despite progress in this field, the robustness of referring perception models (RPMs) against disruptive perturbations is not well explored. This work thoroughly assesses the resilience of RPMs against various perturbations in both general and specific contexts. Recognizing the complex nature of referring perception tasks, we present a comprehensive taxonomy of perturbations, and then develop a versatile toolbox for synthesizing and evaluating the effects of composite disturbances. Employing this toolbox, we construct \\(\\mathbf{R}^2\\)-Bench, a benchmark for assessing the Robustness of Referring perception models under noisy conditions across five key tasks. Moreover, we propose the \\(\\mathbf{R}^2\\)-Agent, an LLM-based agent that simplifies and automates model evaluation via natural language instructions. Our investigation uncovers the vulnerabilities of current RPMs to various perturbations and provides tools for assessing model robustness, potentially promoting the safe and resilient integration of intelligent systems into complex real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.819, + 0.744, + 0.846 + ], + "angle": 0, + "content": "Keywords: Referring perception \\(\\cdot\\) Robustness & Perturbation \\(\\cdot\\) Benchmark" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Xiang Li et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.775, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.285, + 0.789, + 0.342 + ], + "angle": 0, + "content": "Fig. 2: Examples of \\(\\mathbf{R}^2\\)-Bench. The \"Original\" row displays the original inputs alongside the outcomes from models of R-VOS [54], AVS [34] and Q3M [23], while the \"Perturbed\" row presents the inputs as synthesized by \\(\\mathbf{R}^2\\)-Bench and the respective outcomes with the same models. RG: short for Referring Guidance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.371, + 0.376, + 0.387 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.789, + 0.569 + ], + "angle": 0, + "content": "Perception systems [6-8,29-32,80,81] function as input channels for intelligent systems, analogous to human eyes. Referring perception, focuses on identification and contextualization of visual entities by referring multimodal guidance, illustrated in Fig. 1 (a). It effectively creates a communicative bridge between humans, who issue instructions, and the environment that is subject to perception. Multimodal referring cues used for such identification include textual descriptions, example imagery, or auditory signals corresponding to the target object (Fig. 1 (b)). Within the domain of referring perception, specific tasks such as referring expression segmentation [37,68], audiovisual source localization [17,34,79], and queryable 3D mapping [23], are essential modules of robotic control [2,22], autonomous navigation [23,40] and planning [21,48], and human-computer interaction [51]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.79, + 0.81 + ], + "angle": 0, + "content": "Recent advancements in referring perception models [14, 33, 68] have been witnessed and deployed on resource-constrained platforms [54, 58]. However, existing datasets for model training often comprise delicate images with accurate annotations, a condition rarely met in real-world scenarios. In realistic settings, as illustrated in Fig. 1, perception systems may face various disturbances, such as environmental noise (buzzing sound from the radio), imprecise human instructions (misspelling), and sensor imperfections (image blur), which can significantly challenge the robustness of these models. Despite their critical importance for safer real-world applications, the impact of such perturbations on model performance is not thoroughly investigated in the literature. Evaluating model robustness is further complicated by the diversity of operational environments; the assessment process is typically labor-intensive and relies heavily on expert judgment. Addressing this gap, this paper introduces the first benchmark in the field to systematically assess the resilience of referring perception models, i.e., trying to obtain a reliable and comprehensive answer to the critical question: \"How robust are the current referring perception models under realistic perturbations?\"" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.79, + 0.842 + ], + "angle": 0, + "content": "Specifically, we introduce the Robust Referring Benchmark \\((\\mathbf{R}^2\\)-Bench), featuring: (1) a human-friendly taxonomy of perturbations for referring perception," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.217, + 0.115, + 0.769, + 0.129 + ], + "angle": 0, + "content": "Benchmarking the Robustness of Referring Perception Models under Perturbations" + }, + { + "type": "page_number", + "bbox": [ + 0.811, + 0.116, + 0.823, + 0.128 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": "categorizing disruptions into source, environment, sensor, and transmission noise, and detailing perturbation types for each modality; (2) a customizable perturbation synthesis toolbox for creating reasonable noise-augmented datasets, enabling robustness evaluation of perception models; and (3) robustness evaluations across five tasks—referring image segmentation (RIS), video object segmentation (VOS), referring video object segmentation (R-VOS), audiovisual segmentation (AVS), and queryable 3D mapping (Q3M)—considering textual, visual, and acoustic reference modalities. We assess over twenty prevalent models in noisy conditions to benchmark their performance and analyze their vulnerabilities. (Examples are illustrated in Fig. 2.) We hope our benchmarking and analysis can benefit the community by shedding light on the vulnerability of existing models in the face of various perturbations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.789, + 0.618 + ], + "angle": 0, + "content": "Notably, our work advances the evaluation of model robustness by incorporating tests against composite noise types, more closely mirroring real-world scenarios where different noises often coexist. The construction of complex noise for realistic robustness assessment is not trivial, as it must avoid non-contextual perturbations, such as the presence of snow in indoor imagery or excessive air absorption effects in indoor audio recordings, which are incongruent with the depicted environment. To address this complexity, we introduce the \\(\\mathbf{R}^2\\)-Agent, a novel large language model-based (LLM-based) system that can comprehend natural language instructions provided by humans and autonomously generates test samples with contextual relevant perturbations that align with the deployment environment. For instance, given the human-provided instruction \"outdoor night scene\", the \\(\\mathbf{R}^2\\)-Agent could probably generate perturbations of \"motion blur\" instead of \"glass blur\" and \"low background noise\" instead of \"room reverberation\". To improve the agent's abilities of instruction-following and commonsense-reasoning, we employ the multi-agent debating technique [16,36,77], wherein one agent proposes potential solutions while another verifies their validity. Consequently, the \\(\\mathbf{R}^2\\)-Agent efficiently automates the identification and integration of mixed perturbations, thus optimizing the process of robustness evaluations tailored to specific domain contexts." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.62, + 0.599, + 0.635 + ], + "angle": 0, + "content": "In summary, our major contribution is three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.654, + 0.787, + 0.715 + ], + "angle": 0, + "content": "- We introduce the \\(\\mathbf{R}^2\\)-Bench, a benchmarking framework that includes diverse perturbed data for five commonly encountered referring perception tasks. A comprehensive taxonomy and a corresponding customizable synthesis toolbox are developed to enable robustness assessment in noisy real-world settings." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.717, + 0.788, + 0.778 + ], + "angle": 0, + "content": "- We propose the \\( \\mathbf{R}^2 \\)-Agent to streamline model evaluations tailored to particular use scenarios based on LLMs. This automated program executes the perturbation composition process in accordance with human-provided instructions, thereby allowing for context-specific robustness evaluations." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "- Our systematic experimental investigations delve into the intrinsic characteristics of perturbations, such as their types, severity, dynamics, and correlations. These explorations yield valuable insights into the vulnerability of existing models to disturbances and elucidate the nature of these perturbations." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.654, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Xiang Li et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.398, + 0.161 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.39 + ], + "angle": 0, + "content": "Textual referring perception. Referring image segmentation (RIS) and referring video object segmentation (R-VOS) aim to segment objects in images and video sequences, respectively, based on a linguistic description. Recent RIS methods [15, 37,63,68,82-84] have achieved promising results using multimodal transformers. R-VOS [14,18,42,47,49,54] is more challenging as it requires leveraging both intra-frame and temporal cues. URVOS [47] is the first unified R-VOS framework with cross-modal attention and a memory attention module, significantly improving R-VOS performance. ReferFormer [55] employs a linguistic prior in the transformer decoder to focus on the referred object, while MTTR [5] uses a multimodal transformer encoder to fuse linguistic and visual features. \\(\\mathbb{R}^2\\)-VOS introduces relational cyclic consistency to enhance the robustness of the R-VOS model. Unlike other vision-language tasks [65,66], R-VOS needs to construct object-level multimodal semantic consensus in dense visual representations. Relying on 2D models, 3D referring perception is also achievable [23, 64]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.789, + 0.692 + ], + "angle": 0, + "content": "Acoustic referring perception. Audiovisual segmentation (AVS) [35, 35, 38, 78] focuses on segmenting objects that produce sound in a given image frame. This pioneering work by Zhou et al. [78] introduced a method that uses cross-modal attention to identify the sound source. Building on this, Zhou et al. [78] proposed an extended task called audiovisual semantic segmentation (AVSS), which not only segments the sound-producing objects but also classifies them. AVSS is more challenging than AVS due to the complexity of audio semantics. To address this, Zhou et al. [78] utilized the TPAVI module from [79] for audiovisual interaction. Additionally, a recent development by Li et al. [27] introduced the CATR framework, which features a novel spatial-temporal audio-video fusion block for efficient audio-visual integration. Li et al. [34] propose a quantization-based semantic decomposition module to handle the complex acoustic environment and enhance the robustness. Sound source localization (SSL) is a related field that aims to identify the visual regions corresponding to sounds. Several SSL methods [3, 4, 12, 46] utilize the correspondence between audio and visual features to locate sounds, often represented as heatmaps. For example, Mo et al. [43] employed multi-level audiovisual contrastive learning to effectively pinpoint sound-producing objects. In addition, using speech as referring guidance is also explored in [35, 44]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Visual referring perception. Visual referring perception typically relies on visual prompts such as example images, points, bounding boxes, or scribbles. Segment anything model (SAM) [26] is a powerful foundation model that supports multiple visual prompts. Several follow-up works [24,39,58] improve SAM in terms of inference cost, segmentation quality, and referring prompts. Beyond that, several works [67,73,75] explore the usage of SAM in downstream tasks. Li et al. [30] built a SAM-like model with stable diffusion. For video-level tasks, semi-video object segmentation (VOS) [9,28,60,70,71] aims to segment visual objects across frames given the first frame annotation. Recent offline models [9,11] focus more" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.217, + 0.115, + 0.769, + 0.13 + ], + "angle": 0, + "content": "Benchmarking the Robustness of Referring Perception Models under Perturbations" + }, + { + "type": "page_number", + "bbox": [ + 0.811, + 0.116, + 0.823, + 0.128 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "on designing long-term information propagation modules to transfer previous image features and corresponding masks to the target frame to predict masks. This helps more precisely identify and track the movement trajectory of objects throughout the entire video sequence (more than 200 frames) but also makes the network architecture heavy [72]. Due to the slow speed of offline models, online models [56, 76] come back into focus, aiming to keep the right balance between speed and performance." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.276, + 0.737, + 0.294 + ], + "angle": 0, + "content": "3 \\(\\mathbf{R}^2\\)-Bench: Customizable Perturbation Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.308, + 0.787, + 0.37 + ], + "angle": 0, + "content": "In this section, we present the \\(\\mathbf{R}^2\\)-Bench for evaluating robustness of referring perception tasks, including a comprehensive taxonomy that categorizes various perturbations and a customizable perturbation synthesis toolbox that can be tailored to generate specific disturbances." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.392, + 0.673, + 0.409 + ], + "angle": 0, + "content": "3.1 Taxonomic Perturbations in Referring Perception" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.479, + 0.613 + ], + "angle": 0, + "content": "Referring perception tasks typically employ multimodal guidance, spanning the visual, auditory, and textural domains, to ground objects in the visual contexts. Nonetheless, in real-world scenarios, each modality may undergo heterogeneous noises before input into the model. As depicted in Fig. 3, noises can be categorized into four classes based on their origins: source noises, environment noises, sensor noises, and transmission noises." + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.419, + 0.776, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.484, + 0.557, + 0.788, + 0.612 + ], + "angle": 0, + "content": "Fig. 3: Noise categories based on their origins. Assuming airplane as the source of referring guidance, noise from it is categorized as source noise." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.705 + ], + "angle": 0, + "content": "Source noises. We define source noise as the inherent disturbances introduced spontaneously by the source of the referring guidance. For example, for text-referring tasks, the source noises can manifest as misspellings, grammar errors, or punctuation errors in the textual expression and instruction (human is usually the source of text reference)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.781 + ], + "angle": 0, + "content": "Environment noises. We define environment noise as the unintended disturbance introduced by the surrounding environment during the signal capture process. Background sound, room reverberation and air absorption for acoustic waveforms are common examples of this type." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Sensor noises. We define sensor noise as the unwanted disturbances in the output of sensors that arise from inherent limitations or imperfections. For example, defocus blur, motion blur and impulse noise for visual frames." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Xiang Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.145, + 0.782, + 0.159 + ], + "angle": 0, + "content": "Table 1: Noise types considered in \\( \\mathbb{R}^2 \\)-Bench. Details are available in the Appendix." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.162, + 0.785, + 0.28 + ], + "angle": 0, + "content": "
TypeSourceEnvironmentSensorTransmission
Visual-snow (SN), fog (FG), frost (FT), spatter (SP), brightness (BR)defocus blur (DB), gaussian blur (GB), motion blur (MB), glass blur (GS), impulse noise (IN), shot noise (ST), speckle noise (SPN), contrast (CT), saturate (SA)JPEG compression (JPG), pixelated (PIX)
Acousticamplitude gain (GA)background noise (BN), air absorption (AA), room reverberation (RS)gaussian noise (GN), impulse noise (IN), peak filter (PF), time mask (TM), tanh distortion (TD)MP3 compression (MP3), lowpass filter (LP), highpass filter (HP)
Textualmisspelling (MS), mispunctuation (MP), grammar error (GE)-character missing (CM)-
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.31, + 0.784, + 0.354 + ], + "angle": 0, + "content": "Transmission noises. We define transmission noise as the signal distortion or loss during the transmission resulting from data compression or package loss. For example, JPEG and MP4 compression for visual and acoustic signals." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.355, + 0.785, + 0.4 + ], + "angle": 0, + "content": "We summarize the supported noises in visual, acoustic and textual modalities in Tab. 1. A total of 32 types of noises are considered in this paper. The implementation details of noise functions are available in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.424, + 0.57, + 0.438 + ], + "angle": 0, + "content": "3.2 Customizable Perturbation Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.527 + ], + "angle": 0, + "content": "We approach the robustness evaluation through two conventional paradigms: (1) general evaluation under universal scenarios and (2) specific evaluation under designated instructions. To facilitate these evaluations, it is essential to generate perturbed datasets from the original clean data samples, denoted as \\(\\mathcal{X} = \\{x\\}\\), by introducing a spectrum of heterogeneous perturbations \\(\\varDelta=\\{\\delta_k\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.542, + 0.788, + 0.662 + ], + "angle": 0, + "content": "Perturbation order. As discussed in Sec. 3.1, perturbations arise sequentially according to their origins in real-world contexts. To emulate this, perturbations are applied in a sequence that mirrors the natural order of disruption: source \\(\\rightarrow\\) environment \\(\\rightarrow\\) sensor \\(\\rightarrow\\) transmission. Within a single category of origin, the order is randomly determined to account for the non-deterministic nature of real-world noise. To formalize the perturbation process, consider a clean data sample denoted by \\(x\\), the perturbed sample \\(x'\\) undergoes a series of transformations that can be mathematically described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.676, + 0.787, + 0.691 + ], + "angle": 0, + "content": "\\[\nx ^ {\\prime} = \\delta_ {t} \\circ \\delta_ {s e} \\circ \\delta_ {e} \\circ \\delta_ {s o} (x), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.764 + ], + "angle": 0, + "content": "where \\(\\circ\\) denotes composition operation. \\(\\delta_{so},\\delta_e,\\delta_{se},\\delta_t\\) represent generation functions for source, environment, sensor, and transmission noises, respectively. Given that the introduction of perturbations follows a sequential order, the operations that compose these noise functions are inherently order-variant." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Perturbation severity and mode. In alignment with established methodologies [20, 61, 62], we quantitatively define perturbation severity across discrete levels that align with human perceptual categories, namely, {low, medium, high}. In addition, we investigate perturbation across two main modes based on dynamics:" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.217, + 0.115, + 0.769, + 0.129 + ], + "angle": 0, + "content": "Benchmarking the Robustness of Referring Perception Models under Perturbations" + }, + { + "type": "page_number", + "bbox": [ + 0.811, + 0.116, + 0.823, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.784, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.298, + 0.788, + 0.37 + ], + "angle": 0, + "content": "Fig. 4: Overview of \\(\\mathbf{R}^2\\)-Agent, the automatic evaluation assistant. Given a human instruction, clean datasets, perturbation functions, and evaluation functions, \\(\\mathbf{R}^2\\)-Agent first proposes and verifies perturbed test samples that match the given instruction. After that, \\(\\mathbf{R}^2\\)-Agent evaluates the model using the verified samples and provides a report that articulates the model's vulnerabilities and overall resilience." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.4, + 0.789, + 0.462 + ], + "angle": 0, + "content": "{static, dynamic}. Static perturbations remain a consistent level of severity throughout all sensor frames within a sequence, while dynamic perturbations exhibit varying severity and types from frame to frame, closely simulating the fluctuating nature of disturbances in real-world environments, taking the form:" + }, + { + "type": "equation", + "bbox": [ + 0.421, + 0.473, + 0.788, + 0.492 + ], + "angle": 0, + "content": "\\[\nx ^ {\\prime} = \\circ_ {k = 1} ^ {K} \\delta_ {k} (x, \\lambda_ {k} (t)), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.735, + 0.519 + ], + "angle": 0, + "content": "where \\(\\lambda_{k}(t)\\) denotes the severity of the \\(k\\)-th perturbation at time step \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.788, + 0.625 + ], + "angle": 0, + "content": "For the general evaluation under universal scenarios, we construct perturbed datasets for each task by leveraging the original clean benchmarks and all perturbations. Specifically, noisy datasets with low, medium, high and dynamic severity are considered. In each setting, a maximum of two perturbations are considered simultaneously. The detailed data-creating procedure is available in the Appendix. Comprehensive experiments and analyses discussing the impact of perturbations are provided to facilitate future research in Sec. 5.3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.79, + 0.672 + ], + "angle": 0, + "content": "Furthermore, we also consider specific evaluations under designated instructions in practice, e.g., virtual phone meetings without background noise. We detail \\(\\mathbf{R}^2\\)-Agent in Sec. 4." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.698, + 0.666, + 0.716 + ], + "angle": 0, + "content": "4 \\(\\mathbf{R}^2\\)-Agent: Automatic Evaluation Assistant" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.789, + 0.841 + ], + "angle": 0, + "content": "With the versatile capabilities of LLMs [1, 50], LLM-based agents [52, 57] are employed to automatically perform various tasks in a human-like manner. In this section, we introduce \\(\\mathbf{R}^2\\)-Agent, an LLM-based agent that facilitates automatic perturbation synthesis and robustness evaluation guided by human instructions. As depicted in Fig. 4, the \\(\\mathbf{R}^2\\)-Agent consists of three primary steps: data proposal, data verification, and model analysis. We utilize Gemini-Vision-Pro [50] as the LLM component within the \\(\\mathbf{R}^2\\)-Agent." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Xiang Li et al." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.147, + 0.238, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.149, + 0.276, + 0.157 + ], + "angle": 0, + "content": "Example" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.182, + 0.278, + 0.192 + ], + "angle": 0, + "content": "② Question" + }, + { + "type": "text", + "bbox": [ + 0.285, + 0.148, + 0.782, + 0.227 + ], + "angle": 0, + "content": "Given a human instruction, you need to verify...Here is the list of all perturbations you need to consider: {ALL_PERTURBATIONS}. You will be given the results selected from another model as input. For example, if the input is {\"type\": \"snoweffect\", \"reason\": \"the snow effect is a noise type\", \"selected\": True} and the instruction is \"Indoor zoom meeting\", then this data sample is not correct since the instruction pointing to an indoor scenario while the snow effect is outdoor. You need to correct the \"selected\" and \"reason\" items in the dict... Now consider an of {HUMAN_INSTRUCTION}, what are the samples possibly happened within the following perturbations? Explain the reason. Reply me in a strict format of python list. For example, [\"type\": noise_type1, \"reason\": reason1, \"selected\": True/False], . Selected perturbations: {SELECTED_PERTURBATIONS}. \n``python \nperturbations = [\"type\": \"snoweffect\", \"reason\": \"...\", \"selected\": False], [\"type\": \"fog effect\", \"reason\": \"...\", \"selected\": False]," + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.237, + 0.789, + 0.335 + ], + "angle": 0, + "content": "Fig. 5: Chain-of-thought prompting template for data verification. \\(①\\&②\\) Following the chain-of-thought spirit, we first give examples to LLM to boost the in-context-learning capability. After that, we ask the LLM to answer a question that is similar to the given example. Specifically, for the data verification task, we ask the LLM to verify the selected samples from the previous iteration, update the results, and explain the reason. The LLM response is instructed to be a Python-format list of dictionaries. \\(③\\) : The response from LLM (which follows the desired Python-format list)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.354, + 0.425, + 0.369 + ], + "angle": 0, + "content": "4.1 Evaluation Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.379, + 0.789, + 0.533 + ], + "angle": 0, + "content": "Given a set of clean data samples \\(\\mathcal{X} = \\{x_{i}\\}\\), perturbation functions \\(\\varDelta=\\{\\delta_k\\}\\), evaluation metric functions \\(\\mathcal{E} = \\{e_j\\}\\), and a human instruction \\(u\\), the objective of R²-Agent is to select an appropriate subset of clean data samples \\(\\mathcal{X}_u\\), potential perturbations \\(\\varDelta_u\\), and suitable evaluation functions \\(\\mathcal{E}_u\\) to assess the model based on the given instruction \\(u\\). A desired human instruction for the R²-Agent should explicitly specify the scenario (e.g., indoor), the evaluation focus (e.g., segmentation quality), and, if applicable, any specific requirements (e.g., testing for motion blur) during the evaluation process. To enhance the robustness of the R²-Agent, we introduce a multi-agent debating strategy utilizing an operator agent \\(\\varPhi\\) and a guardian agent \\(\\varPsi\\), each with their own independent memories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.545, + 0.789, + 0.668 + ], + "angle": 0, + "content": "Data proposal. To propose data samples that align with the criteria delineated in the human instruction, we employ the operator agent \\(\\varPhi\\) with the text prompts \\(<\\mathrm{SEL}>\\), \\(\\varPhi\\) enables the traversal through the combined space of clean samples, pre-defined noise generation functions, and evaluation functions, collectively represented by the tuple \\((\\mathcal{X},\\Delta,\\mathcal{E})\\), to select suitable samples. Additionally, we employ an LLM to generate captions \\(\\mathcal{C}\\) for the clean samples \\(\\mathcal{X}\\). These captions are archived within an accessible memory to enhance the computational efficiency for subsequent stages. The data proposal stage can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.679, + 0.788, + 0.695 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C}, \\mathcal {X} _ {u}, \\mathcal {E} _ {u}, \\Delta_ {u} = \\Phi (u, \\mathcal {X}, \\mathcal {E}, \\Delta , < \\mathrm {S E L} >). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.79, + 0.841 + ], + "angle": 0, + "content": "Data verification. Although LLMs exhibit notable proficiency in reasoning tasks, their susceptibility to generating \"hallucinated\" content poses a significant challenge to the delivery of accurate responses. To mitigate the hallucination effects, we introduce a verification mechanism, designated as the safe guardian \\(\\varPsi\\), which monitors the data proposal made by the operator \\(\\varPhi\\) through the utilization of chain-of-thought prompting [53]. We first create the proposed dataset \\(\\mathcal{X}_u^t\\) and perturbations \\(\\Delta_u^t\\). \\(t\\) denotes the iteration number. Then, we construct verification prompts \\(<\\mathrm{VER}>\\) and employ \\(\\varPsi\\) to validate the integrity of the data instances. Feedback from this verification is relayed to \\(\\varPhi\\), which then uses the information" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.217, + 0.115, + 0.769, + 0.129 + ], + "angle": 0, + "content": "Benchmarking the Robustness of Referring Perception Models under Perturbations" + }, + { + "type": "page_number", + "bbox": [ + 0.811, + 0.116, + 0.823, + 0.128 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.228, + 0.145, + 0.772, + 0.159 + ], + "angle": 0, + "content": "Table 2: Summary of considered tasks, metrics, referring modality, and datasets." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.161, + 0.788, + 0.216 + ], + "angle": 0, + "content": "
TaskRISVOSR-VOSAVSQ3M
Ref. ModalityTextImageTextAudioText
MetricsmIoUG, J, F [25]J, F [25]J, F [25]-
DatasetRefCOCO [74]/+ [74]/g [41]DAVIS [25], YTVOS [59]Ref-DAVIS [25], Ref-YTVOS [47]AVS-s4 [78], AVS-ms3 [78]ScanNet [13], ICL [19]
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.227, + 0.789, + 0.271 + ], + "angle": 0, + "content": "to refine and enhance the data proposal. This iterative process is carried out until a state of convergence or until a maximum iteration number is reached, which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.279, + 0.59, + 0.295 + ], + "angle": 0, + "content": "\\[\nR ^ {t} = \\Psi \\left(\\mathcal {X} _ {u} ^ {t}, \\Delta_ {u} ^ {t}, < \\mathrm {V E R} > _ {\\Psi}\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.299, + 0.786, + 0.316 + ], + "angle": 0, + "content": "\\[\n\\mathcal {X} _ {u} ^ {t + 1}, \\Delta_ {u} ^ {t + 1} = \\Phi (R ^ {t}, < \\mathrm {V E R} > _ {\\Phi}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.322, + 0.784, + 0.381 + ], + "angle": 0, + "content": "where \\( t \\) is the iteration index, \\( R^t \\) is the response of \\( \\varPsi \\) at \\( t \\) iteration. Note that we consider \\( \\varPhi \\) and \\( \\varPsi \\) have separate memories for previous inputs (similar to two separate ChatGPT sessions) and we omit the memorized inputs here for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.383, + 0.788, + 0.473 + ], + "angle": 0, + "content": "Specifically, adversarial prompts are leveraged in the data verification step. In \\( \\langle \\mathrm{VER} \\rangle_{\\Phi} \\), we ask the LLM to select conservative choices with high confidence while, for \\( \\langle \\mathrm{VER} \\rangle_{\\Psi} \\), radical choices are acceptable. The adversarial prompts can proactively encourage a debate between two agents thus correcting potential mistakes in the original data proposal. To help understand the data verification step, we demonstrate an example of the data verification process in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.484, + 0.789, + 0.575 + ], + "angle": 0, + "content": "Model analysis. With the selected samples and functions \\(\\mathcal{X}_u, \\mathcal{E}_u, \\Delta_u\\), we instantiate the noisy data \\(\\mathcal{X}_u'\\) by applying perturbation functions to the clean data, \\(\\mathcal{X}_u' = \\Delta_u(\\mathcal{X}_u)\\). We then calculate a set of metrics \\(\\mathcal{M}_u = \\{m_i\\}\\) that correspond to \\(\\mathcal{X}_u'\\) by applying evaluating functions \\(\\mathcal{E}_u\\) to noisy samples. Utilizing the chain-of-though prompt , we engage an LLM \\(\\Theta\\) with empty memory to analyze the metrics and produce an output report \\(O\\) for the model performance as:" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.581, + 0.787, + 0.597 + ], + "angle": 0, + "content": "\\[\nO = \\Theta \\left(\\mathcal {M} _ {u}, \\mathcal {X} _ {u} ^ {\\prime}, < \\mathrm {A N A} >\\right). \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.615, + 0.367, + 0.632 + ], + "angle": 0, + "content": "5 Experiment" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.643, + 0.51, + 0.658 + ], + "angle": 0, + "content": "5.1 Evaluation Setup and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.664, + 0.788, + 0.71 + ], + "angle": 0, + "content": "In Tab. 2, we present a summary of the benchmark tasks, evaluated datasets, and corresponding metrics. Our benchmark encompasses five prevalent tasks, namely, RIS, VOS, R-VOS, AVS and Q3M." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Evaluation setup. We conduct systematic evaluations to analyze the impact of perturbations on state-of-the-art mode five tasks. Both perturbations in the referring guidance and visual frames are considered in the benchmarking. We create noisy datasets with low, medium and high noise levels for each task. For video-level tasks, we additionally consider the scenario that noises change over time. We first benchmark the models' general performance with all types of noises and then investigate the impact of noises individually. More information about the dataset creation is available in the Appendix." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Xiang Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.141, + 0.788, + 0.184 + ], + "angle": 0, + "content": "Table 3: Performance of referring image segmentation methods under low, medium, and high perturbation levels. Average performance change (APC) is averaged among all perturbation levels. We report mIoU following convention." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.186, + 0.792, + 0.257 + ], + "angle": 0, + "content": "
MethodVenueRefCOCORefCOCO+RefCOCOg
CleanLowMed.HighAPCCleanLowMed.HighAPCCleanLowMed.HighAPC
LAVT [69]CVPR 220.740.580.530.48-0.210.660.580.510.43-0.150.630.550.470.38-0.16
PolyFormer [37]CVPR 230.760.700.640.57-0.120.710.650.570.50-0.140.690.640.570.50-0.12
X-Decoder [82]CVPR 230.680.600.520.44-0.160.690.600.510.44-0.170.640.580.500.44-0.14
ETRIS [63]ICCV 230.710.640.570.49-0.140.600.560.480.40-0.120.600.530.450.38-0.15
SEEM [84]NeurIPS 230.650.590.540.47-0.120.640.580.510.44-0.130.620.560.490.43-0.12
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.261, + 0.786, + 0.289 + ], + "angle": 0, + "content": "Table 4: Performance of video object segmentation methods under low, medium, and high perturbation levels. APC is averaged among all perturbation levels." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.29, + 0.789, + 0.37 + ], + "angle": 0, + "content": "
MethodVenueDAVISYTVOS
CleanLowMediumHighAPCCleanLowMediumHighAPC
JFJFJFJJGsGuGsGuGsGu
AOT [70]NeurIPS 2176.083.071.779.269.776.965.371.1-7.1-7.386.484.284.980.982.078.176.170.4-5.4-7.7
DeAOT [71]NeurIPS 2276.984.573.281.072.480.468.275.5-5.6-6.486.784.586.683.784.581.880.876.8-2.7-3.7
XMem [11]ECCV 2277.484.572.280.071.879.768.474.7-6.6-6.486.584.584.081.281.879.078.674.9-5.1-6.1
DEVA [10]ICCV 2378.785.976.484.974.282.469.977.5-5.2-4.387.283.986.782.884.780.080.676.3-3.2-4.2
Cutie [9]CVPR 2480.687.775.883.575.082.769.375.4-7.2-7.287.884.586.582.884.280.580.375.1-4.1-5.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.402 + ], + "angle": 0, + "content": "Table 5: Performance of referring video object segmentation methods under low, medium, and high perturbation levels. APC is averaged among all perturbation levels." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.404, + 0.791, + 0.496 + ], + "angle": 0, + "content": "
MethodVenueRef-DAVISRef-YTVOS
CleanLowMediumHighAPCCleanLowMediumHighAPC
J FJ FJ FJ FJ FJ FJ FJ FJ FJ FJ FJ F
Video-level Backbone
MTTR [5]CVPR 22------54.056.448.951.643.045.5
SgMg [42]ICCV 2359.064.857.362.651.957.7-6.8-7.757.760.052.454.5
Image-level Backbone
ReferFormer [55]CVPR 2255.861.346.550.143.148.038.541.9-13.1-14.654.856.5
OnlineRefer [54]ICCV 2355.762.951.657.948.054.342.246.3-8.4-10.055.658.9
R2-VOS [33]ICCV 2357.262.450.257.046.553.839.244.6-11.9-10.656.158.4
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.788, + 0.639 + ], + "angle": 0, + "content": "Metrics. Following the convention [41, 74], we leverage mIoU to evaluate the RIS task. For video-level tasks, the convention is to compute region similarity \\(\\mathcal{I}\\) and contour accuracy \\(\\mathcal{F}\\) as defined in [45]. Specifically, \\(\\mathcal{G} = \\frac{\\mathcal{I} + \\mathcal{F}}{2}\\) is also used. In addition, to better demonstrate the performance degradation against perturbations, we define average performance change (APC) as \\(\\mathrm{APC} = \\sum_{i} \\frac{m_i^n - m_i^c}{N}\\) where \\(m_i^n\\) and \\(m_i^c\\) denotes the performance of the \\(i\\)-th sample of perturbed and clean data respectively. \\(N\\) is the sample number. Additional results, evaluated using various metrics and settings, can be found in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.662, + 0.494, + 0.679 + ], + "angle": 0, + "content": "5.2 Performance Benchmarking" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Referring image segmentation. As shown in Tab. 3, we evaluate the state-of-the-art RIS methods on perturbed RefCOCO \\( / + / \\mathrm{g} \\) [74] datasets. We notice that PolyFormer [37] achieves the best performance in terms of both performance and robustness. Different from other methods that generate pixel-level classification results to represent an object mask, PolyFormer utilizes multi-polygon vertices to represent the mask. We consider the robustness can result from the special sequential prediction and polygon representation which eases the reliance on the per-pixel representation. In addition, we notice that though SEEM [83] shows an inferior performance, its robustness to perturbations is promising. We consider that multi-task joint training can account for SEEM's robustness." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.217, + 0.115, + 0.769, + 0.129 + ], + "angle": 0, + "content": "Benchmarking the Robustness of Referring Perception Models under Perturbations" + }, + { + "type": "page_number", + "bbox": [ + 0.804, + 0.117, + 0.821, + 0.128 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.784, + 0.173 + ], + "angle": 0, + "content": "Table 6: Performance of audiovisual segmentation methods under low, medium, and high perturbation levels. APC is averaged among all perturbation levels." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.175, + 0.788, + 0.242 + ], + "angle": 0, + "content": "
MethodVenueAVS-s4AVS-ms3
CleanLowMediumHighAPCCleanLowMediumHighAPC
J FJ FJ FJ FJ FJ FJ FJ FJ FJ FJ FJ F
AVS [79]ECCV 2272.8 84.868.4 79.760.8 71.255.6 66.8-11.2 -12.247.9 57.844.4 54.641.2 50.537.0 44.2-7.0 -8.0
CATR [27]MM 2374.9 87.171.1 84.265.7 79.058.4 72.4-9.8 -8.653.1 65.649.8 61.446.7 58.441.2 52.9-7.2 -8.0
AVSegFormer [17]AAAI 2376.4 86.770.9 82.362.4 74.857.3 69.8-12.9 -11.049.5 62.846.9 59.043.7 55.439.8 50.1-6.0 -8.0
QSD [35]CVPR 2477.6 85.674.8 83.668.3 78.459.4 71.2-10.1 -7.961.8 64.356.6 60.053.1 57.147.6 51.6-9.4 -8.0
" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.246, + 0.785, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.41, + 0.785, + 0.453 + ], + "angle": 0, + "content": "Fig. 6: Queryable 3D mapping of ConceptFusion [23] with inaccurate referring guidance on ScanNet [13] and ICL [19] datasets. Due to the absence of publicly available evaluation codes, we did not provide quantitative evaluation for queryable 3D mapping." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.469, + 0.788, + 0.621 + ], + "angle": 0, + "content": "Video object segmentation. As depicted in Tab. 4, we demonstrate the performance of state-of-the-art methods on DAVIS [25] and YTVOS [59] datasets. We notice that, for simple scenarios in DAVIS datasets, Cutie [9] shows the best performance. While in the more complex YTVOS dataset, DeAOT [71] achieves the best. Cutie heavily relies on the pixel- and object-level correspondence across frames which can be disrupted by the perturbations. Differently, DeAOT leverages decoupled visual and ID embeddings which may be the reason for its robustness in noisy scenarios. In addition, we notice that the unseen categories generally have a larger performance drop compared to the seen categories which can impose more challenges in practical deployment in complex scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.632, + 0.788, + 0.724 + ], + "angle": 0, + "content": "Referring video object segmentation. We benchmark the referring video object segmentation task in Tab. 5. SgMg [42] and OnlineFormer [54] achieve the best performance and robustness among methods equipped with video- and image-level backbones respectively. Unlike other methods, OnlineRefer processes the visual features in a frame-by-frame manner which makes it rely less on the pixel-level temporal correspondence that is easily been disrupted by visual perturbations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Audio-visual segmentation. We show the performance of popular AVS methods in Tab. 6. We notice that QSD [34] and CATR [27] archive promising performance and robustness among all methods. CATR [27] and AVSegFormer [17] are similar methods that leverage acoustic query to query the visual frames with a transformer-based structure. QSD moves one step forward which additionally introduces a quantization operation and a local-global distillation to enhance the robustness of the acoustic representation to adapt to complex scenarios." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Xiang Li et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.148, + 0.788, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.25, + 0.773, + 0.266 + ], + "angle": 0, + "content": "Fig. 7: Average performance changes (APC) for each perturbation type." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.789, + 0.39 + ], + "angle": 0, + "content": "Queryable 3D mapping. Recent queryable 3D mapping methods [23,64] typically leverage the results from 2D referring models and fuse them to a 3D map. Thereby, the failure of 2D referring perception can directly obstacle the success of 3D tasks. As shown in Fig. 6, we visualize the state-of-the-art queryable 3D mapping method, ConceptFusion [23]. We notice that with the accurate text query \"orange sofa\" ConceptFusion can successfully locate the object while when misspell occur in the query, an obvious performance degradation can be observed." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.41, + 0.445, + 0.426 + ], + "angle": 0, + "content": "5.3 Perturbation Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.79, + 0.707 + ], + "angle": 0, + "content": "Perturbation type. We conduct ablation studies on performance degradation against different perturbation types. Fig. 7 illustrates the aggregated average performance deviation among all benchmarked methods under visual, textual, and acoustic perturbations, respectively. For visual perturbations, we notice the frost effect (FT) induced the most substantial performance degradation, which is likely attributable to its extensive occlusion and distortion effects in the whole image. Conversely, color-based perturbations, such as brightness (BR) and saturate (SA), exhibit marginal impact on performance, possibly due to the preservation of shape information. For textual perturbations, our analysis indicated that all types of perturbations lead to comparable levels of degradation. The character missing (CA) has a slightly greater impact. For acoustic perturbations, we notice that most perturbation types only show a marginal impact on the performance. We consider this is because the target objects in AVS-s4/ms3 are salient objects, allowing models to rely predominantly on visual cues for object localization, even when the acoustic guidance lacks precision. Notably, the impulse response perturbation was associated with a significant performance drop, which could stem from the failure in feature extraction due to the impulse signal. Such abnormal acoustic features could potentially disrupt subsequent multimodal interactions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.542, + 0.84 + ], + "angle": 0, + "content": "Perturbation correlation. To delve deeper into the nature of performance degradation, we computed correlation matrix for each modality, utilizing the Pearson product-moment correlation coefficients calculated from the performance under various perturbation types to populate the matrix. Fig. 8 displays the correlation matrix for textual perturbations." + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.729, + 0.773, + 0.806 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.809, + 0.788, + 0.837 + ], + "angle": 0, + "content": "Fig.8: Correlation of performance of textual perturbations." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.217, + 0.115, + 0.769, + 0.129 + ], + "angle": 0, + "content": "Benchmarking the Robustness of Referring Perception Models under Perturbations" + }, + { + "type": "page_number", + "bbox": [ + 0.804, + 0.116, + 0.822, + 0.128 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.499, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.147, + 0.784, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.262, + 0.788, + 0.304 + ], + "angle": 0, + "content": "Fig. 9: Correlation matrix of degraded performance with visual and acoustic perturbations. We concatenate the sample-level performance of all methods on all testing datasets as a feature vector to calculate Pearson product-moment correlation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.323, + 0.789, + 0.504 + ], + "angle": 0, + "content": "Despite the average performance changes (in Fig. 7) being similar across different types, the low correlation between the types of performance degradation suggest that each perturbation affects the model in a distinct manner. The correlations for visual perturbations and the ones for acoustic perturbations are presented in Fig. 9 respectively. For visual perturbations, we notice several instances of high correlations. For instance, the group of blurring perturbations—defocus blur (DB), gaussian blur (GB), motion blur (MB), and glass blur (GS)—are highly correlated with one another. For the acoustic perturbations, impulse response (IR) and highpass filter (HF) exhibit different patterns compared to other perturbations, indicating unique effects on performance degradation. We systematically explored the correlations among perturbations and hope the analysis can provide insights for wider evaluation in the future." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.518, + 0.64 + ], + "angle": 0, + "content": "Static v.s. Dynamic. To analyze the dynamic changing of perturbation types in videos, we conduct an ablation study to explore the impact of perturbation changing across frames (for visual perturbation only). As shown in Fig. 10, we calculate the APC across all models on the Ref-DAVIS dataset. We notice the model performance" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.522, + 0.776, + 0.596 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.598, + 0.788, + 0.639 + ], + "angle": 0, + "content": "Fig.10: Average performance change (APC) under static and dynamic perturbations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.788, + 0.672 + ], + "angle": 0, + "content": "decreases more with dynamic perturbations which can be due to the lack of temporal consistency resulting from various perturbation types." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.692, + 0.522, + 0.71 + ], + "angle": 0, + "content": "5.4 \\(\\mathbf{R}^2\\)-Agent: Evaluation Assistant" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.72, + 0.487, + 0.809 + ], + "angle": 0, + "content": "User study of data selection. To evaluate the data selection performance of \\(\\mathbf{R}^2\\)-Agent, we manually annotate an evaluation set. Since the instruction can be ambiguous in practice, we measure the alignment between \\(\\mathbf{R}^2\\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.511, + 0.725, + 0.766, + 0.738 + ], + "angle": 0, + "content": "Table 7: Data selection performance." + }, + { + "type": "table", + "bbox": [ + 0.499, + 0.739, + 0.788, + 0.8 + ], + "angle": 0, + "content": "
# Ver. Iter.X123
ScoreACC RateACC RateACC RateACC Rate
Data Sample0.530.780.640.810.770.850.800.91
Perturbation0.650.870.770.860.860.910.870.93
Metric Func.1.001.001.001.001.001.001.001.00
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Agent and human by (1) ACC: accuracy between the human-annotated samples and \\(\\mathbb{R}^2\\)-Agent predictions and (2) Rate: rating the rationality of the prediction" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Xiang Li et al." + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.146, + 0.356, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.362, + 0.15, + 0.782, + 0.252 + ], + "angle": 0, + "content": "
Scenario:Scenario: Car driving scenario with a fast motion\nPerturbation: snow (SN), fog (SG), frost (FT), spatter (SP), motion blur (MB),...
Evaluation Metric: mIoU (mean Intersection over Union)
Analysis:IoU Change: The spatter has a largest -12.1 change in mIoU suggesting that the high level of spatter effect has negatively impacted the model's ability. The spatter effect can lead to 1. Boundary Ambiguity...2. Fragmentation...3. Texture change...ForJPEG compression and animation have a marginal effect...
Case Analysis: The case with caption: “a red car is far away on the left” suffers severe performance drop under...The small size of the car may account for the...
Recommendation: Improve the resilience of models to compression artifacts through enhanced training, including a wider variety of...
" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.259, + 0.788, + 0.302 + ], + "angle": 0, + "content": "Fig. 11: Example of \\(\\mathbf{R}^2\\)-Agent report. The evaluation setting and metrics are fed to \\(\\mathbf{R}^2\\)-Agent with a set of hard-coded prompts to facilitate \\(\\mathbf{R}^2\\)-Agent generate the human-like report. Beyond the hard-coded analysis, further QA is also supported." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.327, + 0.788, + 0.435 + ], + "angle": 0, + "content": "from \\(\\mathbf{R}^2\\)-Agent (1 if reasonable else 0). As shown in Tab. 7, we ablate on the data verification iteration number. For the metrics function selection, since there are only several hard-coded metrics that can be selected, the accuracy remains perfect while, for data sample and perturbation, we notice a direct performance gain when increasing the iteration number. We notice the performance becomes saturated after 2 iterations, thus we leverage iteration number 2 as the final design choice. More details can be found in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.79, + 0.596 + ], + "angle": 0, + "content": "Automatic model analysis. With the evaluation setting and corresponding results, \\(\\mathbf{R}^2\\)-Agent can generate a detailed evaluation report. We demonstrate an evaluation report example in Fig. 11. To make \\(\\mathbf{R}^2\\)-Agent think in a human-like manner, we leverage a set of chain-of-thought prompts to instruct \\(\\mathbf{R}^2\\)-Agent to give reasonable and detailed analysis based on the given evaluation results. We notice that \\(\\mathbf{R}^2\\)-Agent can understand the evaluation metrics and give a human-like analysis which potentially helps to reduce the cost of model evaluation in practice. Furthermore, given the strong dialogue capability of LLMs, further question-answering about the results and report is also feasible. We demonstrate more qualitative results in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.616, + 0.36, + 0.631 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.84 + ], + "angle": 0, + "content": "In this work, we introduced \\(\\mathbb{R}^2\\)-Bench, a comprehensive benchmark for evaluating the robustness of visual referring perception models against various perturbations. Our contributions include a detailed taxonomy of perturbations, a customizable perturbation synthesis toolbox, systematic perturbation analysis, and \\(\\mathbb{R}^2\\)-Agent, a novel evaluation assistant based on large language models. Through extensive experiments, we observed and analyzed the intrinsic characteristics of different perturbations to current models and highlighted the importance of robustness in referring perception tasks. A benchmark with 5 popular referring tasks is also provided to facilitate future research. In addition, the data construction and analysis of robustness evaluation can be further simplified with our proposed \\(\\mathbb{R}^2\\)-Agent. Our findings underscore the necessity for robustness in the deployment of intelligent systems in real-world scenarios, offering a foundation for future advancements in the field." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.217, + 0.115, + 0.769, + 0.129 + ], + "angle": 0, + "content": "Benchmarking the Robustness of Referring Perception Models under Perturbations" + }, + { + "type": "page_number", + "bbox": [ + 0.804, + 0.116, + 0.821, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.787, + 0.223 + ], + "angle": 0, + "content": "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.223, + 0.787, + 0.278 + ], + "angle": 0, + "content": "2. Ahn, H., Kwon, O., Kim, K., Jeong, J., Jun, H., Lee, H., Lee, D., Oh, S.: Visually grounding language instruction for history-dependent manipulation. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 675-682. IEEE (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.279, + 0.787, + 0.306 + ], + "angle": 0, + "content": "3. Arandjelovic, R., Zisserman, A.: Look, listen and learn. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 609-617 (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.307, + 0.787, + 0.336 + ], + "angle": 0, + "content": "4. Arandjelovic, R., Zisserman, A.: Objects that sound. In: Proceedings of the European conference on computer vision (ECCV). pp. 435-451 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.336, + 0.787, + 0.377 + ], + "angle": 0, + "content": "5. Botach, A., Zheltonozhskii, E., Baskin, C.: End-to-end referring video object segmentation with multimodal transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4985-4995 (2022) 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.787, + 0.433 + ], + "angle": 0, + "content": "6. Chen, F., Zhang, H., Hu, K., Huang, Y.k., Zhu, C., Savvides, M.: Enhanced training of query-based object detection via selective query recollection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23756-23765 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.787, + 0.474 + ], + "angle": 0, + "content": "7. Chen, F., Zhang, H., Li, Z., Dou, J., Mo, S., Chen, H., Zhang, Y., Ahmed, U., Zhu, C., Savvides, M.: Unitail: Detecting, reading, and matching in retail scene (2022), https://arxiv.org/abs/2204.00298 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.476, + 0.787, + 0.516 + ], + "angle": 0, + "content": "8. Chen, F., Zhang, H., Yang, Z., Chen, H., Hu, K., Savvides, M.: Rtgen: Generating region-text pairs for open-vocabulary object detection (2024), https://arxiv.org/abs/2405.19854 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.517, + 0.787, + 0.546 + ], + "angle": 0, + "content": "9. Cheng, H.K., Oh, S.W., Price, B., Lee, J.Y., Schwing, A.: Putting the object back into video object segmentation. arXiv preprint arXiv:2310.12982 (2023) 4, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.546, + 0.787, + 0.574 + ], + "angle": 0, + "content": "0. Cheng, H.K., Oh, S.W., Price, B., Schwing, A., Lee, J.Y.: Tracking anything with decoupled video segmentation. In: ICCV (2023) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.575, + 0.787, + 0.616 + ], + "angle": 0, + "content": "1. Cheng, H.K., Schwing, A.G.: Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In: European Conference on Computer Vision. pp. 640-658. Springer (2022) 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.617, + 0.787, + 0.672 + ], + "angle": 0, + "content": "2. Cheng, Y., Wang, R., Pan, Z., Feng, R., Zhang, Y.: Look, listen, and attend: Co-attention network for self-supervised audio-visual representation learning. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 3884-3892 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.673, + 0.787, + 0.714 + ], + "angle": 0, + "content": "3. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017) 9, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.715, + 0.787, + 0.757 + ], + "angle": 0, + "content": "4. Ding, H., Liu, C., He, S., Jiang, X., Loy, C.C.: Mevis: A large-scale benchmark for video segmentation with motion expressions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2694-2703 (2023) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "5. Ding, H., Liu, C., Wang, S., Jiang, X.: Vision-language transformer and query generation for referring segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16321-16330 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "6. Du, Y., Li, S., Torralba, A., Tenenbaum, J.B., Mordatch, I.: Improving factuality and reasoning in language models through multiagent debate. arXiv preprint arXiv:2305.14325 (2023) 3" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Xiang Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.175 + ], + "angle": 0, + "content": "17. Gao, S., Chen, Z., Chen, G., Wang, W., Lu, T.: Avsegformer: Audio-visual segmentation with transformer. arXiv preprint arXiv:2307.01146 (2023) 2, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.788, + 0.232 + ], + "angle": 0, + "content": "18. Han, M., Wang, Y., Li, Z., Yao, L., Chang, X., Qiao, Y.: Hybrid temporal-scale multimodal learning framework for referring video object segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13414-13423 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.787, + 0.274 + ], + "angle": 0, + "content": "19. Handa, A., Whelan, T., McDonald, J., Davison, A.: A benchmark for RGB-D visual odometry, 3D reconstruction and SLAM. In: IEEE Intl. Conf. on Robotics and Automation, ICRA. Hong Kong, China (May 2014) 9, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.786, + 0.303 + ], + "angle": 0, + "content": "20. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261 (2019) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.786, + 0.345 + ], + "angle": 0, + "content": "21. Hu, Y., Lin, F., Zhang, T., Yi, L., Gao, Y.: Look before you leap: Unveiling the power of gpt-4v in robotic vision-language planning. arXiv preprint arXiv:2311.17842 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.787, + 0.388 + ], + "angle": 0, + "content": "22. Huang, W., Xia, F., Shah, D., Driess, D., Zeng, A., Lu, Y., Florence, P., Mordatch, I., Levine, S., Hausman, K., et al.: Grounded decoding: Guiding text generation with grounded models for robot control. arXiv preprint arXiv:2303.00855 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.389, + 0.787, + 0.43 + ], + "angle": 0, + "content": "23. Jatavallabhula, K.M., Kuwajerwala, A., Gu, Q., Omama, M., Chen, T., Maalouf, A., Li, S., Iyer, G., Saryazdi, S., Keetha, N., et al.: Conceptfusion: Open-set multimodal 3d mapping. arXiv preprint arXiv:2302.07241 (2023) 2, 4, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.431, + 0.787, + 0.458 + ], + "angle": 0, + "content": "24. Ke, L., Ye, M., Danelljan, M., Liu, Y., Tai, Y.W., Tang, C.K., Yu, F.: Segment anything in high quality. In: NeurIPS (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.459, + 0.787, + 0.487 + ], + "angle": 0, + "content": "25. Khoreva, A., Rohrbach, A., Schiele, B.: Video object segmentation with language referring expressions. In: ACCV (2018) 9, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.529 + ], + "angle": 0, + "content": "26. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.787, + 0.572 + ], + "angle": 0, + "content": "27. Li, K., Yang, Z., Chen, L., Yang, Y., Xun, J.: Catr: Combinatorial-dependence audio-queried transformer for audio-visual video segmentation. arXiv preprint arXiv:2309.09709 (2023) 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.787, + 0.6 + ], + "angle": 0, + "content": "28. Li, M., Li, S., Zhang, X., Zhang, L.: Univs: Unified and universal video segmentation with prompts as queries. arXiv preprint arXiv:2402.18115 (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.787, + 0.642 + ], + "angle": 0, + "content": "29. Li, X., Cao, H., Zhao, S., Li, J., Zhang, L., Raj, B.: Panoramic video salient object detection with ambisonic audio guidance. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1424-1432 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.643, + 0.787, + 0.685 + ], + "angle": 0, + "content": "30. Li, X., Lin, C.C., Chen, Y., Liu, Z., Wang, J., Singh, R., Raj, B.: Paintseg: Painting pixels for training-free segmentation. In: Thirty-seventh Conference on Neural Information Processing Systems (2023) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.787, + 0.727 + ], + "angle": 0, + "content": "31. Li, X., Wang, J., Li, X., Lu, Y.: Hybrid instance-aware temporal fusion for online video instance segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 1429-1437 (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.728, + 0.787, + 0.755 + ], + "angle": 0, + "content": "32. Li, X., Wang, J., Li, X., Lu, Y.: Video instance segmentation by instance flow assembly. IEEE Transactions on Multimedia 25, 7469-7479 (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.756, + 0.787, + 0.798 + ], + "angle": 0, + "content": "33. Li, X., Wang, J., Xu, X., Li, X., Raj, B., Lu, Y.: Robust referring video object segmentation with cyclic structural consensus. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22236-22245 (2023) 2, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "34. Li, X., Wang, J., Xu, X., Peng, X., Singh, R., Lu, Y., Raj, B.: Towards robust audiovisual segmentation in complex environments with quantization-based semantic decomposition. arXiv preprint arXiv:2310.00132 (2023) 2, 4, 11" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.217, + 0.115, + 0.769, + 0.129 + ], + "angle": 0, + "content": "Benchmarking the Robustness of Referring Perception Models under Perturbations" + }, + { + "type": "page_number", + "bbox": [ + 0.804, + 0.116, + 0.822, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "35. Li, X., Wang, J., Xu, X., Yang, M., Yang, F., Zhao, Y., Singh, R., Raj, B.: Towards noise-tolerant speech-referring video object segmentation: Bridging speech and text. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. pp. 2283-2296 (2023) 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.788, + 0.245 + ], + "angle": 0, + "content": "36. Liang, T., He, Z., Jiao, W., Wang, X., Wang, Y., Wang, R., Yang, Y., Tu, Z., Shi, S.: Encouraging divergent thinking in large language models through multi-agent debate. arXiv preprint arXiv:2305.19118 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.788, + 0.3 + ], + "angle": 0, + "content": "37. Liu, J., Ding, H., Cai, Z., Zhang, Y., Satzoda, R.K., Mahadevan, V., Manmatha, R.: Polyformer: Referring image segmentation as sequential polygon generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18653-18663 (2023) 2, 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.788, + 0.342 + ], + "angle": 0, + "content": "38. Liu, J., Wang, Y., Ju, C., Ma, C., Zhang, Y., Xie, W.: Annotation-free audio-visual segmentation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5604-5614 (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.788, + 0.384 + ], + "angle": 0, + "content": "39. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.788, + 0.438 + ], + "angle": 0, + "content": "40. Liu, S., Hasan, A., Hong, K., Wang, R., Chang, P., Mizrachi, Z., Lin, J., McPherson, D.L., Rogers, W.A., Driggs-Campbell, K.: Dragon: A dialogue-based robot for assistive navigation with visual language grounding. IEEE Robotics and Automation Letters (2024) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.788, + 0.492 + ], + "angle": 0, + "content": "41. Mao, J., Huang, J., Toshev, A., Camburu, O., Yuille, A.L., Murphy, K.: Generation and comprehension of unambiguous object descriptions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 11-20 (2016) 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.493, + 0.788, + 0.536 + ], + "angle": 0, + "content": "42. Miao, B., Bennamoun, M., Gao, Y., Mian, A.: Spectrum-guided multi-granularity referring video object segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 920-930 (2023) 4, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.788, + 0.563 + ], + "angle": 0, + "content": "43. Mo, S., Morgado, P.: A closer look at weakly-supervised audio-visual source localization. arXiv preprint arXiv:2209.09634 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.788, + 0.619 + ], + "angle": 0, + "content": "44. Pan, W., Shi, H., Zhao, Z., Zhu, J., He, X., Pan, Z., Gao, L., Yu, J., Wu, F., Tian, Q.: Wnet: Audio-guided video object segmentation via wavelet-based cross-modal denoising networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1320-1331 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.788, + 0.66 + ], + "angle": 0, + "content": "45. Pont-Tuset, J., Peruzzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675 (2017) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.788, + 0.702 + ], + "angle": 0, + "content": "46. Senocak, A., Oh, T.H., Kim, J., Yang, M.H., Kweon, I.S.: Learning to localize sound source in visual scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4358-4366 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.788, + 0.758 + ], + "angle": 0, + "content": "47. Seo, S., Lee, J.Y., Han, B.: Urvos: Unified referring video object segmentation network with a large-scale benchmark. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16. pp. 208-223. Springer (2020) 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.788, + 0.799 + ], + "angle": 0, + "content": "48. Sun, J., Huang, D.A., Lu, B., Liu, Y.H., Zhou, B., Garg, A.: Visually-grounded planning with transformers in procedural tasks. IEEE Robotics and Automation Letters 7(2), 4924-4930 (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.788, + 0.841 + ], + "angle": 0, + "content": "49. Tang, J., Zheng, G., Yang, S.: Temporal collection and distribution for referring video object segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15466-15476 (2023) 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Xiang Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "50. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "51. Tziafas, G., Kasaei, H.: Few-shot visual grounding for natural human-robot interaction. In: 2021 IEEE International Conference on Autonomous Robot Systems and Competitions (ICARSC). pp. 50-55. IEEE (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.272 + ], + "angle": 0, + "content": "52. Wang, Z., Cai, S., Chen, G., Liu, A., Ma, X.S., Liang, Y.: Describe, explain, plan and select: interactive planning with llms enables open-world multi-task agents. Advances in Neural Information Processing Systems 36 (2024) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "53. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "54. Wu, D., Wang, T., Zhang, Y., Zhang, X., Shen, J.: Onlinerefer: A simple online baseline for referring video object segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2761-2770 (2023) 2, 4, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.787, + 0.397 + ], + "angle": 0, + "content": "55. Wu, J., Jiang, Y., Sun, P., Yuan, Z., Luo, P.: Language as queries for referring video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4974-4984 (2022) 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.438 + ], + "angle": 0, + "content": "56. Wu, J., Jiang, Y., Yan, B., Lu, H., Yuan, Z., Luo, P.: Segment every reference object in spatial and temporal spaces. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2538-2550 (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "57. Xi, Z., Chen, W., Guo, X., He, W., Ding, Y., Hong, B., Zhang, M., Wang, J., Jin, S., Zhou, E., et al.: The rise and potential of large language model based agents: A survey. arXiv preprint arXiv:2309.07864 (2023) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "58. Xiong, Y., Varadarajan, B., Wu, L., Xiang, X., Xiao, F., Zhu, C., Dai, X., Wang, D., Sun, F., Iandola, F., et al.: Efficientsam: Leveraged masked image pretraining for efficient segment anything. arXiv preprint arXiv:2312.00863 (2023) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.787, + 0.563 + ], + "angle": 0, + "content": "59. Xu, N., Yang, L., Fan, Y., Yue, D., Liang, Y., Yang, J., Huang, T.: Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327 (2018) 9, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.787, + 0.605 + ], + "angle": 0, + "content": "60. Xu, X., Wang, J., Li, X., Lu, Y.: Reliable propagation-correction modulation for video object segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 2946-2954 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.646 + ], + "angle": 0, + "content": "61. Xu, X., Wang, J., Ming, X., Lu, Y.: Towards robust video object segmentation with adaptive object calibration. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 2709-2718 (2022) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.688 + ], + "angle": 0, + "content": "62. Xu, X., Zhang, T., Wang, S., Li, X., Chen, Y., Li, Y., Raj, B., Johnson-Roberson, M., Huang, X.: Customizable perturbation synthesis for robust slam benchmarking. arXiv preprint arXiv:2402.08125 (2024) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.742 + ], + "angle": 0, + "content": "63. Xu, Z., Chen, Z., Zhang, Y., Song, Y., Wan, X., Li, G.: Bridging vision and language encoders: Parameter-efficient tuning for referring image segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17503-17512 (2023) 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.787, + 0.785 + ], + "angle": 0, + "content": "64. Yamazaki, K., Hanyu, T., Vo, K., Pham, T., Tran, M., Doretto, G., Nguyen, A., Le, N.: Open-fusion: Real-time open-vocabulary 3d mapping and queryable scene representation. arXiv preprint arXiv:2310.03923 (2023) 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "65. Yamazaki, K., Truong, S., Vo, K., Kidd, M., Rainwater, C., Luu, K., Le, N.: Vlcap: Vision-language with contrastive learning for coherent video paragraph captioning. In: 2022 IEEE International Conference on Image Processing (ICIP). pp. 3656-3661. IEEE (2022) 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.217, + 0.115, + 0.769, + 0.129 + ], + "angle": 0, + "content": "Benchmarking the Robustness of Referring Perception Models under Perturbations" + }, + { + "type": "page_number", + "bbox": [ + 0.805, + 0.116, + 0.822, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "66. Yamazaki, K., Vo, K., Truong, Q.S., Raj, B., Le, N.: Vlnt: Visual-linguistic transformer-in-transformer for coherent video paragraph captioning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 3081-3090 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.219 + ], + "angle": 0, + "content": "67. Yang, J., Gao, M., Li, Z., Gao, S., Wang, F., Zheng, F.: Track anything: Segment anything meets videos (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.219, + 0.788, + 0.274 + ], + "angle": 0, + "content": "68. Yang, Z., Wang, J., Tang, Y., Chen, K., Zhao, H., Torr, P.H.: Levt: Language-aware vision transformer for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18155-18165 (2022) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.276, + 0.788, + 0.331 + ], + "angle": 0, + "content": "69. Yang, Z., Wang, J., Tang, Y., Chen, K., Zhao, H., Torr, P.H.: Levt: Language-aware vision transformer for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18155-18165 (2022) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.332, + 0.788, + 0.373 + ], + "angle": 0, + "content": "70. Yang, Z., Wei, Y., Yang, Y.: Associating objects with transformers for video object segmentation. Advances in Neural Information Processing Systems 34, 2491-2502 (2021) 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.415 + ], + "angle": 0, + "content": "71. Yang, Z., Yang, Y.: Decoupling features in hierarchical propagation for video object segmentation. Advances in Neural Information Processing Systems 35, 36324-36336 (2022) 4, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.417, + 0.788, + 0.458 + ], + "angle": 0, + "content": "72. Yang, Z., Yang, Y.: Decoupling features in hierarchical propagation for video object segmentation. Advances in Neural Information Processing Systems 35, 36324-36336 (2022) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.487 + ], + "angle": 0, + "content": "73. Yao, J., Wang, X., Ye, L., Liu, W.: Matte anything: Interactive natural image matting with segment anything models. arXiv preprint arXiv:2306.04121 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.488, + 0.788, + 0.529 + ], + "angle": 0, + "content": "74. Yu, L., Poirson, P., Yang, S., Berg, A.C., Berg, T.L.: Modeling context in referring expressions. In: European Conference on Computer Vision. pp. 69-85. Springer (2016) 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.53, + 0.788, + 0.571 + ], + "angle": 0, + "content": "75. Yu, T., Feng, R., Feng, R., Liu, J., Jin, X., Zeng, W., Chen, Z.: Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.572, + 0.788, + 0.613 + ], + "angle": 0, + "content": "76. Zhang, J., Cui, Y., Wu, G., Wang, L.: Joint modeling of feature, correspondence, and a compressed memory for video object segmentation. arXiv preprint arXiv:2308.13505 (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.615, + 0.788, + 0.655 + ], + "angle": 0, + "content": "77. Zhao, Q., Wang, J., Zhang, Y., Jin, Y., Zhu, K., Chen, H., Xie, X.: Competeai: Understanding the competition behaviors in large language model-based agents. arXiv preprint arXiv:2310.17512 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.657, + 0.788, + 0.699 + ], + "angle": 0, + "content": "78. Zhou, J., Shen, X., Wang, J., Zhang, J., Sun, W., Zhang, J., Birchfield, S., Guo, D., Kong, L., Wang, M., et al.: Audio-visual segmentation with semantics. arXiv preprint arXiv:2301.13190 (2023) 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.7, + 0.788, + 0.741 + ], + "angle": 0, + "content": "79. Zhou, J., Wang, J., Zhang, J., Sun, W., Zhang, J., Birchfield, S., Guo, D., Kong, L., Wang, M., Zhong, Y.: Audio-visual segmentation. In: European Conference on Computer Vision (2022) 2, 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.742, + 0.788, + 0.797 + ], + "angle": 0, + "content": "80. Zhu, C., Chen, F., Ahmed, U., Shen, Z., Savvides, M.: Semantic relation reasoning for shot-stable few-shot object detection. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 8778-8787 (2021), https://api.sementicscholar.org/CorpusID:232093016 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.788, + 0.84 + ], + "angle": 0, + "content": "81. Zhu, C., Chen, F., Shen, Z., Savvides, M.: Soft anchor-point object detection. In: European Conference on Computer Vision (2019), https://api-semanticscholar.org/CorpusID:208512715 2" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Xiang Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "82. Zou, X., Dou, Z.Y., Yang, J., Gan, Z., Li, L., Li, C., Dai, X., Behl, H., Wang, J., Yuan, L., et al.: Generalized decoding for pixel, image, and language. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15116-15127 (2023) 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.788, + 0.231 + ], + "angle": 0, + "content": "83. Zou, X., Yang, J., Zhang, H., Li, F., Li, L., Gao, J., Lee, Y.J.: Segment everything everywhere all at once. arXiv preprint arXiv:2304.06718 (2023) 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.788, + 0.273 + ], + "angle": 0, + "content": "84. Zou, X., Yang, J., Zhang, H., Li, F., Li, L., Wang, J., Wang, L., Gao, J., Lee, Y.J.: Segment everything everywhere all at once. Advances in Neural Information Processing Systems 36 (2024) 4, 10" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.273 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/R^2-Bench_ Benchmarking the Robustness of Referring Perception Models under Perturbations/3cf440ac-b4a0-433a-943f-d4d176a64132_origin.pdf b/2024/R^2-Bench_ Benchmarking the Robustness of Referring Perception Models under Perturbations/3cf440ac-b4a0-433a-943f-d4d176a64132_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4e92c7e46eb8be8bce7d0eb2c0ee279f2b682399 --- /dev/null +++ b/2024/R^2-Bench_ Benchmarking the Robustness of Referring Perception Models under Perturbations/3cf440ac-b4a0-433a-943f-d4d176a64132_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:40d6b10157e614ecaed5dc8c7c53cff3053d780b9fc15a6211e5acc59826ba26 +size 2885176 diff --git a/2024/R^2-Bench_ Benchmarking the Robustness of Referring Perception Models under Perturbations/full.md b/2024/R^2-Bench_ Benchmarking the Robustness of Referring Perception Models under Perturbations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5177024b7a738767e12409c1415ee6147e9c7acc --- /dev/null +++ b/2024/R^2-Bench_ Benchmarking the Robustness of Referring Perception Models under Perturbations/full.md @@ -0,0 +1,336 @@ +# $\mathbf{R}^2$ -Bench: Benchmarking the Robustness of Referring Perception Models under Perturbations + +Xiang Li $^{1}$ , Kai Qiu $^{1}$ , Jinglu Wang $^{2}$ , Xiaohao Xu $^{3}$ , Rita Singh $^{1}$ , Kashu Yamazaki $^{1}$ , Hao Chen $^{1}$ , Xiaonan Huang $^{3}$ , Bhiksha Raj $^{1,4}$ + +1 CMU, 2 Microsoft Research Asia, 3 University of Michigan, 4 MBZUAI + +![](images/10aae151004360f04481c045134b269f97efaca8992350082a8d8f6834a038d2.jpg) +(a) Referring Perception in Practice + +![](images/27f810604090a0d5ebf36812144a519207af7753772bb7335da1cf77b9cf7049.jpg) +(b) Perturbation in Referring Guidance +Fig. 1: Motivation illustration. Referring perception models (RPMs) empower intelligent systems with their ability to perform object grounding within the environment based on referring guidance, such as textual descriptions, imagery exemplars, or auditory signals associated with the target object. However, RPMs' performance can be compromised by disturbances in real-world scenarios, such as environmental noise (e.g., extraneous sounds from a nearby radio), human-induced errors (e.g., typographical errors in textual input), and limitations in the sensor (e.g., motion blur in images). Conducting a rigorous analysis of RPMs' robustness to a wide array of perturbations is necessary for building reliable real-world applications. + +Abstract. Referring perception, which aims at grounding visual objects with multimodal referring guidance, is essential for bridging the gap between humans, who provide instructions, and the environment where intelligent systems perceive. Despite progress in this field, the robustness of referring perception models (RPMs) against disruptive perturbations is not well explored. This work thoroughly assesses the resilience of RPMs against various perturbations in both general and specific contexts. Recognizing the complex nature of referring perception tasks, we present a comprehensive taxonomy of perturbations, and then develop a versatile toolbox for synthesizing and evaluating the effects of composite disturbances. Employing this toolbox, we construct $\mathbf{R}^2$ -Bench, a benchmark for assessing the Robustness of Referring perception models under noisy conditions across five key tasks. Moreover, we propose the $\mathbf{R}^2$ -Agent, an LLM-based agent that simplifies and automates model evaluation via natural language instructions. Our investigation uncovers the vulnerabilities of current RPMs to various perturbations and provides tools for assessing model robustness, potentially promoting the safe and resilient integration of intelligent systems into complex real-world scenarios. + +Keywords: Referring perception $\cdot$ Robustness & Perturbation $\cdot$ Benchmark + +![](images/5a13cc43f38aaba38e8329e201d3c93c8a82d48baa9f88ae1a982df887bdbf27.jpg) +Fig. 2: Examples of $\mathbf{R}^2$ -Bench. The "Original" row displays the original inputs alongside the outcomes from models of R-VOS [54], AVS [34] and Q3M [23], while the "Perturbed" row presents the inputs as synthesized by $\mathbf{R}^2$ -Bench and the respective outcomes with the same models. RG: short for Referring Guidance. + +# 1 Introduction + +Perception systems [6-8,29-32,80,81] function as input channels for intelligent systems, analogous to human eyes. Referring perception, focuses on identification and contextualization of visual entities by referring multimodal guidance, illustrated in Fig. 1 (a). It effectively creates a communicative bridge between humans, who issue instructions, and the environment that is subject to perception. Multimodal referring cues used for such identification include textual descriptions, example imagery, or auditory signals corresponding to the target object (Fig. 1 (b)). Within the domain of referring perception, specific tasks such as referring expression segmentation [37,68], audiovisual source localization [17,34,79], and queryable 3D mapping [23], are essential modules of robotic control [2,22], autonomous navigation [23,40] and planning [21,48], and human-computer interaction [51]. + +Recent advancements in referring perception models [14, 33, 68] have been witnessed and deployed on resource-constrained platforms [54, 58]. However, existing datasets for model training often comprise delicate images with accurate annotations, a condition rarely met in real-world scenarios. In realistic settings, as illustrated in Fig. 1, perception systems may face various disturbances, such as environmental noise (buzzing sound from the radio), imprecise human instructions (misspelling), and sensor imperfections (image blur), which can significantly challenge the robustness of these models. Despite their critical importance for safer real-world applications, the impact of such perturbations on model performance is not thoroughly investigated in the literature. Evaluating model robustness is further complicated by the diversity of operational environments; the assessment process is typically labor-intensive and relies heavily on expert judgment. Addressing this gap, this paper introduces the first benchmark in the field to systematically assess the resilience of referring perception models, i.e., trying to obtain a reliable and comprehensive answer to the critical question: "How robust are the current referring perception models under realistic perturbations?" + +Specifically, we introduce the Robust Referring Benchmark $(\mathbf{R}^2$ -Bench), featuring: (1) a human-friendly taxonomy of perturbations for referring perception, + +categorizing disruptions into source, environment, sensor, and transmission noise, and detailing perturbation types for each modality; (2) a customizable perturbation synthesis toolbox for creating reasonable noise-augmented datasets, enabling robustness evaluation of perception models; and (3) robustness evaluations across five tasks—referring image segmentation (RIS), video object segmentation (VOS), referring video object segmentation (R-VOS), audiovisual segmentation (AVS), and queryable 3D mapping (Q3M)—considering textual, visual, and acoustic reference modalities. We assess over twenty prevalent models in noisy conditions to benchmark their performance and analyze their vulnerabilities. (Examples are illustrated in Fig. 2.) We hope our benchmarking and analysis can benefit the community by shedding light on the vulnerability of existing models in the face of various perturbations. + +Notably, our work advances the evaluation of model robustness by incorporating tests against composite noise types, more closely mirroring real-world scenarios where different noises often coexist. The construction of complex noise for realistic robustness assessment is not trivial, as it must avoid non-contextual perturbations, such as the presence of snow in indoor imagery or excessive air absorption effects in indoor audio recordings, which are incongruent with the depicted environment. To address this complexity, we introduce the $\mathbf{R}^2$ -Agent, a novel large language model-based (LLM-based) system that can comprehend natural language instructions provided by humans and autonomously generates test samples with contextual relevant perturbations that align with the deployment environment. For instance, given the human-provided instruction "outdoor night scene", the $\mathbf{R}^2$ -Agent could probably generate perturbations of "motion blur" instead of "glass blur" and "low background noise" instead of "room reverberation". To improve the agent's abilities of instruction-following and commonsense-reasoning, we employ the multi-agent debating technique [16,36,77], wherein one agent proposes potential solutions while another verifies their validity. Consequently, the $\mathbf{R}^2$ -Agent efficiently automates the identification and integration of mixed perturbations, thus optimizing the process of robustness evaluations tailored to specific domain contexts. + +In summary, our major contribution is three-fold: + +- We introduce the $\mathbf{R}^2$ -Bench, a benchmarking framework that includes diverse perturbed data for five commonly encountered referring perception tasks. A comprehensive taxonomy and a corresponding customizable synthesis toolbox are developed to enable robustness assessment in noisy real-world settings. +- We propose the $\mathbf{R}^2$ -Agent to streamline model evaluations tailored to particular use scenarios based on LLMs. This automated program executes the perturbation composition process in accordance with human-provided instructions, thereby allowing for context-specific robustness evaluations. +- Our systematic experimental investigations delve into the intrinsic characteristics of perturbations, such as their types, severity, dynamics, and correlations. These explorations yield valuable insights into the vulnerability of existing models to disturbances and elucidate the nature of these perturbations. + +# 2 Related Works + +Textual referring perception. Referring image segmentation (RIS) and referring video object segmentation (R-VOS) aim to segment objects in images and video sequences, respectively, based on a linguistic description. Recent RIS methods [15, 37,63,68,82-84] have achieved promising results using multimodal transformers. R-VOS [14,18,42,47,49,54] is more challenging as it requires leveraging both intra-frame and temporal cues. URVOS [47] is the first unified R-VOS framework with cross-modal attention and a memory attention module, significantly improving R-VOS performance. ReferFormer [55] employs a linguistic prior in the transformer decoder to focus on the referred object, while MTTR [5] uses a multimodal transformer encoder to fuse linguistic and visual features. $\mathbb{R}^2$ -VOS introduces relational cyclic consistency to enhance the robustness of the R-VOS model. Unlike other vision-language tasks [65,66], R-VOS needs to construct object-level multimodal semantic consensus in dense visual representations. Relying on 2D models, 3D referring perception is also achievable [23, 64]. + +Acoustic referring perception. Audiovisual segmentation (AVS) [35, 35, 38, 78] focuses on segmenting objects that produce sound in a given image frame. This pioneering work by Zhou et al. [78] introduced a method that uses cross-modal attention to identify the sound source. Building on this, Zhou et al. [78] proposed an extended task called audiovisual semantic segmentation (AVSS), which not only segments the sound-producing objects but also classifies them. AVSS is more challenging than AVS due to the complexity of audio semantics. To address this, Zhou et al. [78] utilized the TPAVI module from [79] for audiovisual interaction. Additionally, a recent development by Li et al. [27] introduced the CATR framework, which features a novel spatial-temporal audio-video fusion block for efficient audio-visual integration. Li et al. [34] propose a quantization-based semantic decomposition module to handle the complex acoustic environment and enhance the robustness. Sound source localization (SSL) is a related field that aims to identify the visual regions corresponding to sounds. Several SSL methods [3, 4, 12, 46] utilize the correspondence between audio and visual features to locate sounds, often represented as heatmaps. For example, Mo et al. [43] employed multi-level audiovisual contrastive learning to effectively pinpoint sound-producing objects. In addition, using speech as referring guidance is also explored in [35, 44] + +Visual referring perception. Visual referring perception typically relies on visual prompts such as example images, points, bounding boxes, or scribbles. Segment anything model (SAM) [26] is a powerful foundation model that supports multiple visual prompts. Several follow-up works [24,39,58] improve SAM in terms of inference cost, segmentation quality, and referring prompts. Beyond that, several works [67,73,75] explore the usage of SAM in downstream tasks. Li et al. [30] built a SAM-like model with stable diffusion. For video-level tasks, semi-video object segmentation (VOS) [9,28,60,70,71] aims to segment visual objects across frames given the first frame annotation. Recent offline models [9,11] focus more + +on designing long-term information propagation modules to transfer previous image features and corresponding masks to the target frame to predict masks. This helps more precisely identify and track the movement trajectory of objects throughout the entire video sequence (more than 200 frames) but also makes the network architecture heavy [72]. Due to the slow speed of offline models, online models [56, 76] come back into focus, aiming to keep the right balance between speed and performance. + +# 3 $\mathbf{R}^2$ -Bench: Customizable Perturbation Benchmark + +In this section, we present the $\mathbf{R}^2$ -Bench for evaluating robustness of referring perception tasks, including a comprehensive taxonomy that categorizes various perturbations and a customizable perturbation synthesis toolbox that can be tailored to generate specific disturbances. + +# 3.1 Taxonomic Perturbations in Referring Perception + +Referring perception tasks typically employ multimodal guidance, spanning the visual, auditory, and textural domains, to ground objects in the visual contexts. Nonetheless, in real-world scenarios, each modality may undergo heterogeneous noises before input into the model. As depicted in Fig. 3, noises can be categorized into four classes based on their origins: source noises, environment noises, sensor noises, and transmission noises. + +![](images/91b81622bee6eaa89487e40ef6b687a51d3a2312cfde49d3d3c2d7580915fa72.jpg) +Fig. 3: Noise categories based on their origins. Assuming airplane as the source of referring guidance, noise from it is categorized as source noise. + +Source noises. We define source noise as the inherent disturbances introduced spontaneously by the source of the referring guidance. For example, for text-referring tasks, the source noises can manifest as misspellings, grammar errors, or punctuation errors in the textual expression and instruction (human is usually the source of text reference). + +Environment noises. We define environment noise as the unintended disturbance introduced by the surrounding environment during the signal capture process. Background sound, room reverberation and air absorption for acoustic waveforms are common examples of this type. + +Sensor noises. We define sensor noise as the unwanted disturbances in the output of sensors that arise from inherent limitations or imperfections. For example, defocus blur, motion blur and impulse noise for visual frames. + +Table 1: Noise types considered in $\mathbb{R}^2$ -Bench. Details are available in the Appendix. + +
TypeSourceEnvironmentSensorTransmission
Visual-snow (SN), fog (FG), frost (FT), spatter (SP), brightness (BR)defocus blur (DB), gaussian blur (GB), motion blur (MB), glass blur (GS), impulse noise (IN), shot noise (ST), speckle noise (SPN), contrast (CT), saturate (SA)JPEG compression (JPG), pixelated (PIX)
Acousticamplitude gain (GA)background noise (BN), air absorption (AA), room reverberation (RS)gaussian noise (GN), impulse noise (IN), peak filter (PF), time mask (TM), tanh distortion (TD)MP3 compression (MP3), lowpass filter (LP), highpass filter (HP)
Textualmisspelling (MS), mispunctuation (MP), grammar error (GE)-character missing (CM)-
+ +Transmission noises. We define transmission noise as the signal distortion or loss during the transmission resulting from data compression or package loss. For example, JPEG and MP4 compression for visual and acoustic signals. + +We summarize the supported noises in visual, acoustic and textual modalities in Tab. 1. A total of 32 types of noises are considered in this paper. The implementation details of noise functions are available in the Appendix. + +# 3.2 Customizable Perturbation Synthesis + +We approach the robustness evaluation through two conventional paradigms: (1) general evaluation under universal scenarios and (2) specific evaluation under designated instructions. To facilitate these evaluations, it is essential to generate perturbed datasets from the original clean data samples, denoted as $\mathcal{X} = \{x\}$ , by introducing a spectrum of heterogeneous perturbations $\varDelta=\{\delta_k\}$ . + +Perturbation order. As discussed in Sec. 3.1, perturbations arise sequentially according to their origins in real-world contexts. To emulate this, perturbations are applied in a sequence that mirrors the natural order of disruption: source $\rightarrow$ environment $\rightarrow$ sensor $\rightarrow$ transmission. Within a single category of origin, the order is randomly determined to account for the non-deterministic nature of real-world noise. To formalize the perturbation process, consider a clean data sample denoted by $x$ , the perturbed sample $x'$ undergoes a series of transformations that can be mathematically described as follows: + +$$ +x ^ {\prime} = \delta_ {t} \circ \delta_ {s e} \circ \delta_ {e} \circ \delta_ {s o} (x), \tag {1} +$$ + +where $\circ$ denotes composition operation. $\delta_{so},\delta_e,\delta_{se},\delta_t$ represent generation functions for source, environment, sensor, and transmission noises, respectively. Given that the introduction of perturbations follows a sequential order, the operations that compose these noise functions are inherently order-variant. + +Perturbation severity and mode. In alignment with established methodologies [20, 61, 62], we quantitatively define perturbation severity across discrete levels that align with human perceptual categories, namely, {low, medium, high}. In addition, we investigate perturbation across two main modes based on dynamics: + +![](images/96abbc7c9eeb9ae3d7a739ee1cea094ab1171bf9222f7b7e973cdaf97bc4a3c3.jpg) +Fig. 4: Overview of $\mathbf{R}^2$ -Agent, the automatic evaluation assistant. Given a human instruction, clean datasets, perturbation functions, and evaluation functions, $\mathbf{R}^2$ -Agent first proposes and verifies perturbed test samples that match the given instruction. After that, $\mathbf{R}^2$ -Agent evaluates the model using the verified samples and provides a report that articulates the model's vulnerabilities and overall resilience. + +{static, dynamic}. Static perturbations remain a consistent level of severity throughout all sensor frames within a sequence, while dynamic perturbations exhibit varying severity and types from frame to frame, closely simulating the fluctuating nature of disturbances in real-world environments, taking the form: + +$$ +x ^ {\prime} = \circ_ {k = 1} ^ {K} \delta_ {k} (x, \lambda_ {k} (t)), \tag {2} +$$ + +where $\lambda_{k}(t)$ denotes the severity of the $k$ -th perturbation at time step $t$ . + +For the general evaluation under universal scenarios, we construct perturbed datasets for each task by leveraging the original clean benchmarks and all perturbations. Specifically, noisy datasets with low, medium, high and dynamic severity are considered. In each setting, a maximum of two perturbations are considered simultaneously. The detailed data-creating procedure is available in the Appendix. Comprehensive experiments and analyses discussing the impact of perturbations are provided to facilitate future research in Sec. 5.3. + +Furthermore, we also consider specific evaluations under designated instructions in practice, e.g., virtual phone meetings without background noise. We detail $\mathbf{R}^2$ -Agent in Sec. 4. + +# 4 $\mathbf{R}^2$ -Agent: Automatic Evaluation Assistant + +With the versatile capabilities of LLMs [1, 50], LLM-based agents [52, 57] are employed to automatically perform various tasks in a human-like manner. In this section, we introduce $\mathbf{R}^2$ -Agent, an LLM-based agent that facilitates automatic perturbation synthesis and robustness evaluation guided by human instructions. As depicted in Fig. 4, the $\mathbf{R}^2$ -Agent consists of three primary steps: data proposal, data verification, and model analysis. We utilize Gemini-Vision-Pro [50] as the LLM component within the $\mathbf{R}^2$ -Agent. + +![](images/cdd590ad2708d573ea6e9a599e553b4cf4979a09c8bd5f0644212171300735af.jpg) + +Example + +② Question + +Given a human instruction, you need to verify...Here is the list of all perturbations you need to consider: {ALL_PERTURBATIONS}. You will be given the results selected from another model as input. For example, if the input is {"type": "snoweffect", "reason": "the snow effect is a noise type", "selected": True} and the instruction is "Indoor zoom meeting", then this data sample is not correct since the instruction pointing to an indoor scenario while the snow effect is outdoor. You need to correct the "selected" and "reason" items in the dict... Now consider an of {HUMAN_INSTRUCTION}, what are the samples possibly happened within the following perturbations? Explain the reason. Reply me in a strict format of python list. For example, ["type": noise_type1, "reason": reason1, "selected": True/False], . Selected perturbations: {SELECTED_PERTURBATIONS}. +``python +perturbations = ["type": "snoweffect", "reason": "...", "selected": False], ["type": "fog effect", "reason": "...", "selected": False], + +Fig. 5: Chain-of-thought prompting template for data verification. $①\&②$ Following the chain-of-thought spirit, we first give examples to LLM to boost the in-context-learning capability. After that, we ask the LLM to answer a question that is similar to the given example. Specifically, for the data verification task, we ask the LLM to verify the selected samples from the previous iteration, update the results, and explain the reason. The LLM response is instructed to be a Python-format list of dictionaries. $③$ : The response from LLM (which follows the desired Python-format list). + +# 4.1 Evaluation Pipeline + +Given a set of clean data samples $\mathcal{X} = \{x_{i}\}$ , perturbation functions $\varDelta=\{\delta_k\}$ , evaluation metric functions $\mathcal{E} = \{e_j\}$ , and a human instruction $u$ , the objective of R²-Agent is to select an appropriate subset of clean data samples $\mathcal{X}_u$ , potential perturbations $\varDelta_u$ , and suitable evaluation functions $\mathcal{E}_u$ to assess the model based on the given instruction $u$ . A desired human instruction for the R²-Agent should explicitly specify the scenario (e.g., indoor), the evaluation focus (e.g., segmentation quality), and, if applicable, any specific requirements (e.g., testing for motion blur) during the evaluation process. To enhance the robustness of the R²-Agent, we introduce a multi-agent debating strategy utilizing an operator agent $\varPhi$ and a guardian agent $\varPsi$ , each with their own independent memories. + +Data proposal. To propose data samples that align with the criteria delineated in the human instruction, we employ the operator agent $\varPhi$ with the text prompts $<\mathrm{SEL}>$ , $\varPhi$ enables the traversal through the combined space of clean samples, pre-defined noise generation functions, and evaluation functions, collectively represented by the tuple $(\mathcal{X},\Delta,\mathcal{E})$ , to select suitable samples. Additionally, we employ an LLM to generate captions $\mathcal{C}$ for the clean samples $\mathcal{X}$ . These captions are archived within an accessible memory to enhance the computational efficiency for subsequent stages. The data proposal stage can be formulated as: + +$$ +\mathcal {C}, \mathcal {X} _ {u}, \mathcal {E} _ {u}, \Delta_ {u} = \Phi (u, \mathcal {X}, \mathcal {E}, \Delta , < \mathrm {S E L} >). \tag {3} +$$ + +Data verification. Although LLMs exhibit notable proficiency in reasoning tasks, their susceptibility to generating "hallucinated" content poses a significant challenge to the delivery of accurate responses. To mitigate the hallucination effects, we introduce a verification mechanism, designated as the safe guardian $\varPsi$ , which monitors the data proposal made by the operator $\varPhi$ through the utilization of chain-of-thought prompting [53]. We first create the proposed dataset $\mathcal{X}_u^t$ and perturbations $\Delta_u^t$ . $t$ denotes the iteration number. Then, we construct verification prompts $<\mathrm{VER}>$ and employ $\varPsi$ to validate the integrity of the data instances. Feedback from this verification is relayed to $\varPhi$ , which then uses the information + +Table 2: Summary of considered tasks, metrics, referring modality, and datasets. + +
TaskRISVOSR-VOSAVSQ3M
Ref. ModalityTextImageTextAudioText
MetricsmIoUG, J, F [25]J, F [25]J, F [25]-
DatasetRefCOCO [74]/+ [74]/g [41]DAVIS [25], YTVOS [59]Ref-DAVIS [25], Ref-YTVOS [47]AVS-s4 [78], AVS-ms3 [78]ScanNet [13], ICL [19]
+ +to refine and enhance the data proposal. This iterative process is carried out until a state of convergence or until a maximum iteration number is reached, which is formulated as: + +$$ +R ^ {t} = \Psi \left(\mathcal {X} _ {u} ^ {t}, \Delta_ {u} ^ {t}, < \mathrm {V E R} > _ {\Psi}\right) +$$ + +$$ +\mathcal {X} _ {u} ^ {t + 1}, \Delta_ {u} ^ {t + 1} = \Phi (R ^ {t}, < \mathrm {V E R} > _ {\Phi}), \tag {4} +$$ + +where $t$ is the iteration index, $R^t$ is the response of $\varPsi$ at $t$ iteration. Note that we consider $\varPhi$ and $\varPsi$ have separate memories for previous inputs (similar to two separate ChatGPT sessions) and we omit the memorized inputs here for simplicity. + +Specifically, adversarial prompts are leveraged in the data verification step. In $\langle \mathrm{VER} \rangle_{\Phi}$ , we ask the LLM to select conservative choices with high confidence while, for $\langle \mathrm{VER} \rangle_{\Psi}$ , radical choices are acceptable. The adversarial prompts can proactively encourage a debate between two agents thus correcting potential mistakes in the original data proposal. To help understand the data verification step, we demonstrate an example of the data verification process in Fig. 5. + +Model analysis. With the selected samples and functions $\mathcal{X}_u, \mathcal{E}_u, \Delta_u$ , we instantiate the noisy data $\mathcal{X}_u'$ by applying perturbation functions to the clean data, $\mathcal{X}_u' = \Delta_u(\mathcal{X}_u)$ . We then calculate a set of metrics $\mathcal{M}_u = \{m_i\}$ that correspond to $\mathcal{X}_u'$ by applying evaluating functions $\mathcal{E}_u$ to noisy samples. Utilizing the chain-of-though prompt , we engage an LLM $\Theta$ with empty memory to analyze the metrics and produce an output report $O$ for the model performance as: + +$$ +O = \Theta \left(\mathcal {M} _ {u}, \mathcal {X} _ {u} ^ {\prime}, < \mathrm {A N A} >\right). \tag {5} +$$ + +# 5 Experiment + +# 5.1 Evaluation Setup and Metrics + +In Tab. 2, we present a summary of the benchmark tasks, evaluated datasets, and corresponding metrics. Our benchmark encompasses five prevalent tasks, namely, RIS, VOS, R-VOS, AVS and Q3M. + +Evaluation setup. We conduct systematic evaluations to analyze the impact of perturbations on state-of-the-art mode five tasks. Both perturbations in the referring guidance and visual frames are considered in the benchmarking. We create noisy datasets with low, medium and high noise levels for each task. For video-level tasks, we additionally consider the scenario that noises change over time. We first benchmark the models' general performance with all types of noises and then investigate the impact of noises individually. More information about the dataset creation is available in the Appendix. + +Table 3: Performance of referring image segmentation methods under low, medium, and high perturbation levels. Average performance change (APC) is averaged among all perturbation levels. We report mIoU following convention. + +
MethodVenueRefCOCORefCOCO+RefCOCOg
CleanLowMed.HighAPCCleanLowMed.HighAPCCleanLowMed.HighAPC
LAVT [69]CVPR 220.740.580.530.48-0.210.660.580.510.43-0.150.630.550.470.38-0.16
PolyFormer [37]CVPR 230.760.700.640.57-0.120.710.650.570.50-0.140.690.640.570.50-0.12
X-Decoder [82]CVPR 230.680.600.520.44-0.160.690.600.510.44-0.170.640.580.500.44-0.14
ETRIS [63]ICCV 230.710.640.570.49-0.140.600.560.480.40-0.120.600.530.450.38-0.15
SEEM [84]NeurIPS 230.650.590.540.47-0.120.640.580.510.44-0.130.620.560.490.43-0.12
+ +Table 4: Performance of video object segmentation methods under low, medium, and high perturbation levels. APC is averaged among all perturbation levels. + +
MethodVenueDAVISYTVOS
CleanLowMediumHighAPCCleanLowMediumHighAPC
JFJFJFJJGsGuGsGuGsGu
AOT [70]NeurIPS 2176.083.071.779.269.776.965.371.1-7.1-7.386.484.284.980.982.078.176.170.4-5.4-7.7
DeAOT [71]NeurIPS 2276.984.573.281.072.480.468.275.5-5.6-6.486.784.586.683.784.581.880.876.8-2.7-3.7
XMem [11]ECCV 2277.484.572.280.071.879.768.474.7-6.6-6.486.584.584.081.281.879.078.674.9-5.1-6.1
DEVA [10]ICCV 2378.785.976.484.974.282.469.977.5-5.2-4.387.283.986.782.884.780.080.676.3-3.2-4.2
Cutie [9]CVPR 2480.687.775.883.575.082.769.375.4-7.2-7.287.884.586.582.884.280.580.375.1-4.1-5.0
+ +Table 5: Performance of referring video object segmentation methods under low, medium, and high perturbation levels. APC is averaged among all perturbation levels. + +
MethodVenueRef-DAVISRef-YTVOS
CleanLowMediumHighAPCCleanLowMediumHighAPC
J FJ FJ FJ FJ FJ FJ FJ FJ FJ FJ FJ F
Video-level Backbone
MTTR [5]CVPR 22------54.056.448.951.643.045.5
SgMg [42]ICCV 2359.064.857.362.651.957.7-6.8-7.757.760.052.454.5
Image-level Backbone
ReferFormer [55]CVPR 2255.861.346.550.143.148.038.541.9-13.1-14.654.856.5
OnlineRefer [54]ICCV 2355.762.951.657.948.054.342.246.3-8.4-10.055.658.9
R2-VOS [33]ICCV 2357.262.450.257.046.553.839.244.6-11.9-10.656.158.4
+ +Metrics. Following the convention [41, 74], we leverage mIoU to evaluate the RIS task. For video-level tasks, the convention is to compute region similarity $\mathcal{I}$ and contour accuracy $\mathcal{F}$ as defined in [45]. Specifically, $\mathcal{G} = \frac{\mathcal{I} + \mathcal{F}}{2}$ is also used. In addition, to better demonstrate the performance degradation against perturbations, we define average performance change (APC) as $\mathrm{APC} = \sum_{i} \frac{m_i^n - m_i^c}{N}$ where $m_i^n$ and $m_i^c$ denotes the performance of the $i$ -th sample of perturbed and clean data respectively. $N$ is the sample number. Additional results, evaluated using various metrics and settings, can be found in the Appendix. + +# 5.2 Performance Benchmarking + +Referring image segmentation. As shown in Tab. 3, we evaluate the state-of-the-art RIS methods on perturbed RefCOCO $/ + / \mathrm{g}$ [74] datasets. We notice that PolyFormer [37] achieves the best performance in terms of both performance and robustness. Different from other methods that generate pixel-level classification results to represent an object mask, PolyFormer utilizes multi-polygon vertices to represent the mask. We consider the robustness can result from the special sequential prediction and polygon representation which eases the reliance on the per-pixel representation. In addition, we notice that though SEEM [83] shows an inferior performance, its robustness to perturbations is promising. We consider that multi-task joint training can account for SEEM's robustness. + +Table 6: Performance of audiovisual segmentation methods under low, medium, and high perturbation levels. APC is averaged among all perturbation levels. + +
MethodVenueAVS-s4AVS-ms3
CleanLowMediumHighAPCCleanLowMediumHighAPC
J FJ FJ FJ FJ FJ FJ FJ FJ FJ FJ FJ F
AVS [79]ECCV 2272.8 84.868.4 79.760.8 71.255.6 66.8-11.2 -12.247.9 57.844.4 54.641.2 50.537.0 44.2-7.0 -8.0
CATR [27]MM 2374.9 87.171.1 84.265.7 79.058.4 72.4-9.8 -8.653.1 65.649.8 61.446.7 58.441.2 52.9-7.2 -8.0
AVSegFormer [17]AAAI 2376.4 86.770.9 82.362.4 74.857.3 69.8-12.9 -11.049.5 62.846.9 59.043.7 55.439.8 50.1-6.0 -8.0
QSD [35]CVPR 2477.6 85.674.8 83.668.3 78.459.4 71.2-10.1 -7.961.8 64.356.6 60.053.1 57.147.6 51.6-9.4 -8.0
+ +![](images/9cc248f997b73f32dd2a24b536b0b936ec18bc1c2026315e597823a55572b226.jpg) +Fig. 6: Queryable 3D mapping of ConceptFusion [23] with inaccurate referring guidance on ScanNet [13] and ICL [19] datasets. Due to the absence of publicly available evaluation codes, we did not provide quantitative evaluation for queryable 3D mapping. + +Video object segmentation. As depicted in Tab. 4, we demonstrate the performance of state-of-the-art methods on DAVIS [25] and YTVOS [59] datasets. We notice that, for simple scenarios in DAVIS datasets, Cutie [9] shows the best performance. While in the more complex YTVOS dataset, DeAOT [71] achieves the best. Cutie heavily relies on the pixel- and object-level correspondence across frames which can be disrupted by the perturbations. Differently, DeAOT leverages decoupled visual and ID embeddings which may be the reason for its robustness in noisy scenarios. In addition, we notice that the unseen categories generally have a larger performance drop compared to the seen categories which can impose more challenges in practical deployment in complex scenarios. + +Referring video object segmentation. We benchmark the referring video object segmentation task in Tab. 5. SgMg [42] and OnlineFormer [54] achieve the best performance and robustness among methods equipped with video- and image-level backbones respectively. Unlike other methods, OnlineRefer processes the visual features in a frame-by-frame manner which makes it rely less on the pixel-level temporal correspondence that is easily been disrupted by visual perturbations. + +Audio-visual segmentation. We show the performance of popular AVS methods in Tab. 6. We notice that QSD [34] and CATR [27] archive promising performance and robustness among all methods. CATR [27] and AVSegFormer [17] are similar methods that leverage acoustic query to query the visual frames with a transformer-based structure. QSD moves one step forward which additionally introduces a quantization operation and a local-global distillation to enhance the robustness of the acoustic representation to adapt to complex scenarios. + +![](images/c237184825d1083df4354938ef2e717843c0572d78365cb3846347f120f9a9ba.jpg) +Fig. 7: Average performance changes (APC) for each perturbation type. + +Queryable 3D mapping. Recent queryable 3D mapping methods [23,64] typically leverage the results from 2D referring models and fuse them to a 3D map. Thereby, the failure of 2D referring perception can directly obstacle the success of 3D tasks. As shown in Fig. 6, we visualize the state-of-the-art queryable 3D mapping method, ConceptFusion [23]. We notice that with the accurate text query "orange sofa" ConceptFusion can successfully locate the object while when misspell occur in the query, an obvious performance degradation can be observed. + +# 5.3 Perturbation Analysis + +Perturbation type. We conduct ablation studies on performance degradation against different perturbation types. Fig. 7 illustrates the aggregated average performance deviation among all benchmarked methods under visual, textual, and acoustic perturbations, respectively. For visual perturbations, we notice the frost effect (FT) induced the most substantial performance degradation, which is likely attributable to its extensive occlusion and distortion effects in the whole image. Conversely, color-based perturbations, such as brightness (BR) and saturate (SA), exhibit marginal impact on performance, possibly due to the preservation of shape information. For textual perturbations, our analysis indicated that all types of perturbations lead to comparable levels of degradation. The character missing (CA) has a slightly greater impact. For acoustic perturbations, we notice that most perturbation types only show a marginal impact on the performance. We consider this is because the target objects in AVS-s4/ms3 are salient objects, allowing models to rely predominantly on visual cues for object localization, even when the acoustic guidance lacks precision. Notably, the impulse response perturbation was associated with a significant performance drop, which could stem from the failure in feature extraction due to the impulse signal. Such abnormal acoustic features could potentially disrupt subsequent multimodal interactions. + +Perturbation correlation. To delve deeper into the nature of performance degradation, we computed correlation matrix for each modality, utilizing the Pearson product-moment correlation coefficients calculated from the performance under various perturbation types to populate the matrix. Fig. 8 displays the correlation matrix for textual perturbations. + +![](images/60a3347f9f1ab8bf5ca7746429f2e861b828a20e2c72d2c53084317b5cb0387e.jpg) +Fig.8: Correlation of performance of textual perturbations. + +![](images/e1cd3cc6389dabacc940cfc734ab581c38a6bbb23004d68fd6876f9283982f06.jpg) +Fig. 9: Correlation matrix of degraded performance with visual and acoustic perturbations. We concatenate the sample-level performance of all methods on all testing datasets as a feature vector to calculate Pearson product-moment correlation. + +![](images/cddd06cdc3252f8c855438348b3ba94245967bf0df35307f042789f907d77939.jpg) + +Despite the average performance changes (in Fig. 7) being similar across different types, the low correlation between the types of performance degradation suggest that each perturbation affects the model in a distinct manner. The correlations for visual perturbations and the ones for acoustic perturbations are presented in Fig. 9 respectively. For visual perturbations, we notice several instances of high correlations. For instance, the group of blurring perturbations—defocus blur (DB), gaussian blur (GB), motion blur (MB), and glass blur (GS)—are highly correlated with one another. For the acoustic perturbations, impulse response (IR) and highpass filter (HF) exhibit different patterns compared to other perturbations, indicating unique effects on performance degradation. We systematically explored the correlations among perturbations and hope the analysis can provide insights for wider evaluation in the future. + +Static v.s. Dynamic. To analyze the dynamic changing of perturbation types in videos, we conduct an ablation study to explore the impact of perturbation changing across frames (for visual perturbation only). As shown in Fig. 10, we calculate the APC across all models on the Ref-DAVIS dataset. We notice the model performance + +![](images/b8aed8f2de4d606a0706778fc314252ef1b3bce727c5c9b9444d46df79dd6d7b.jpg) +Fig.10: Average performance change (APC) under static and dynamic perturbations. + +decreases more with dynamic perturbations which can be due to the lack of temporal consistency resulting from various perturbation types. + +# 5.4 $\mathbf{R}^2$ -Agent: Evaluation Assistant + +User study of data selection. To evaluate the data selection performance of $\mathbf{R}^2$ -Agent, we manually annotate an evaluation set. Since the instruction can be ambiguous in practice, we measure the alignment between $\mathbf{R}^2$ + +Table 7: Data selection performance. + +
# Ver. Iter.X123
ScoreACC RateACC RateACC RateACC Rate
Data Sample0.530.780.640.810.770.850.800.91
Perturbation0.650.870.770.860.860.910.870.93
Metric Func.1.001.001.001.001.001.001.001.00
+ +Agent and human by (1) ACC: accuracy between the human-annotated samples and $\mathbb{R}^2$ -Agent predictions and (2) Rate: rating the rationality of the prediction + +![](images/471a469cc966a399caa11d51911ca264628e3f0602574063cd532553e5943044.jpg) +Fig. 11: Example of $\mathbf{R}^2$ -Agent report. The evaluation setting and metrics are fed to $\mathbf{R}^2$ -Agent with a set of hard-coded prompts to facilitate $\mathbf{R}^2$ -Agent generate the human-like report. Beyond the hard-coded analysis, further QA is also supported. + +
Scenario:Scenario: Car driving scenario with a fast motion +Perturbation: snow (SN), fog (SG), frost (FT), spatter (SP), motion blur (MB),...
Evaluation Metric: mIoU (mean Intersection over Union)
Analysis:IoU Change: The spatter has a largest -12.1 change in mIoU suggesting that the high level of spatter effect has negatively impacted the model's ability. The spatter effect can lead to 1. Boundary Ambiguity...2. Fragmentation...3. Texture change...ForJPEG compression and animation have a marginal effect...
Case Analysis: The case with caption: “a red car is far away on the left” suffers severe performance drop under...The small size of the car may account for the...
Recommendation: Improve the resilience of models to compression artifacts through enhanced training, including a wider variety of...
+ +from $\mathbf{R}^2$ -Agent (1 if reasonable else 0). As shown in Tab. 7, we ablate on the data verification iteration number. For the metrics function selection, since there are only several hard-coded metrics that can be selected, the accuracy remains perfect while, for data sample and perturbation, we notice a direct performance gain when increasing the iteration number. We notice the performance becomes saturated after 2 iterations, thus we leverage iteration number 2 as the final design choice. More details can be found in the Appendix. + +Automatic model analysis. With the evaluation setting and corresponding results, $\mathbf{R}^2$ -Agent can generate a detailed evaluation report. We demonstrate an evaluation report example in Fig. 11. To make $\mathbf{R}^2$ -Agent think in a human-like manner, we leverage a set of chain-of-thought prompts to instruct $\mathbf{R}^2$ -Agent to give reasonable and detailed analysis based on the given evaluation results. We notice that $\mathbf{R}^2$ -Agent can understand the evaluation metrics and give a human-like analysis which potentially helps to reduce the cost of model evaluation in practice. Furthermore, given the strong dialogue capability of LLMs, further question-answering about the results and report is also feasible. We demonstrate more qualitative results in the Appendix. + +# 6 Conclusion + +In this work, we introduced $\mathbb{R}^2$ -Bench, a comprehensive benchmark for evaluating the robustness of visual referring perception models against various perturbations. Our contributions include a detailed taxonomy of perturbations, a customizable perturbation synthesis toolbox, systematic perturbation analysis, and $\mathbb{R}^2$ -Agent, a novel evaluation assistant based on large language models. Through extensive experiments, we observed and analyzed the intrinsic characteristics of different perturbations to current models and highlighted the importance of robustness in referring perception tasks. A benchmark with 5 popular referring tasks is also provided to facilitate future research. In addition, the data construction and analysis of robustness evaluation can be further simplified with our proposed $\mathbb{R}^2$ -Agent. Our findings underscore the necessity for robustness in the deployment of intelligent systems in real-world scenarios, offering a foundation for future advancements in the field. + +# References + +1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023) 7 +2. Ahn, H., Kwon, O., Kim, K., Jeong, J., Jun, H., Lee, H., Lee, D., Oh, S.: Visually grounding language instruction for history-dependent manipulation. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 675-682. IEEE (2022) 2 +3. Arandjelovic, R., Zisserman, A.: Look, listen and learn. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 609-617 (2017) 4 +4. Arandjelovic, R., Zisserman, A.: Objects that sound. In: Proceedings of the European conference on computer vision (ECCV). pp. 435-451 (2018) 4 +5. Botach, A., Zheltonozhskii, E., Baskin, C.: End-to-end referring video object segmentation with multimodal transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4985-4995 (2022) 4, 10 +6. Chen, F., Zhang, H., Hu, K., Huang, Y.k., Zhu, C., Savvides, M.: Enhanced training of query-based object detection via selective query recollection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23756-23765 (2023) 2 +7. Chen, F., Zhang, H., Li, Z., Dou, J., Mo, S., Chen, H., Zhang, Y., Ahmed, U., Zhu, C., Savvides, M.: Unitail: Detecting, reading, and matching in retail scene (2022), https://arxiv.org/abs/2204.00298 2 +8. Chen, F., Zhang, H., Yang, Z., Chen, H., Hu, K., Savvides, M.: Rtgen: Generating region-text pairs for open-vocabulary object detection (2024), https://arxiv.org/abs/2405.19854 2 +9. Cheng, H.K., Oh, S.W., Price, B., Lee, J.Y., Schwing, A.: Putting the object back into video object segmentation. arXiv preprint arXiv:2310.12982 (2023) 4, 10, 11 +0. Cheng, H.K., Oh, S.W., Price, B., Schwing, A., Lee, J.Y.: Tracking anything with decoupled video segmentation. In: ICCV (2023) 10 +1. Cheng, H.K., Schwing, A.G.: Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In: European Conference on Computer Vision. pp. 640-658. Springer (2022) 4, 10 +2. Cheng, Y., Wang, R., Pan, Z., Feng, R., Zhang, Y.: Look, listen, and attend: Co-attention network for self-supervised audio-visual representation learning. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 3884-3892 (2020) 4 +3. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017) 9, 11 +4. Ding, H., Liu, C., He, S., Jiang, X., Loy, C.C.: Mevis: A large-scale benchmark for video segmentation with motion expressions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2694-2703 (2023) 2, 4 +5. Ding, H., Liu, C., Wang, S., Jiang, X.: Vision-language transformer and query generation for referring segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16321-16330 (2021) 4 +6. Du, Y., Li, S., Torralba, A., Tenenbaum, J.B., Mordatch, I.: Improving factuality and reasoning in language models through multiagent debate. arXiv preprint arXiv:2305.14325 (2023) 3 + +17. Gao, S., Chen, Z., Chen, G., Wang, W., Lu, T.: Avsegformer: Audio-visual segmentation with transformer. arXiv preprint arXiv:2307.01146 (2023) 2, 11 +18. Han, M., Wang, Y., Li, Z., Yao, L., Chang, X., Qiao, Y.: Hybrid temporal-scale multimodal learning framework for referring video object segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13414-13423 (2023) 4 +19. Handa, A., Whelan, T., McDonald, J., Davison, A.: A benchmark for RGB-D visual odometry, 3D reconstruction and SLAM. In: IEEE Intl. Conf. on Robotics and Automation, ICRA. Hong Kong, China (May 2014) 9, 11 +20. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261 (2019) 6 +21. Hu, Y., Lin, F., Zhang, T., Yi, L., Gao, Y.: Look before you leap: Unveiling the power of gpt-4v in robotic vision-language planning. arXiv preprint arXiv:2311.17842 (2023) 2 +22. Huang, W., Xia, F., Shah, D., Driess, D., Zeng, A., Lu, Y., Florence, P., Mordatch, I., Levine, S., Hausman, K., et al.: Grounded decoding: Guiding text generation with grounded models for robot control. arXiv preprint arXiv:2303.00855 (2023) 2 +23. Jatavallabhula, K.M., Kuwajerwala, A., Gu, Q., Omama, M., Chen, T., Maalouf, A., Li, S., Iyer, G., Saryazdi, S., Keetha, N., et al.: Conceptfusion: Open-set multimodal 3d mapping. arXiv preprint arXiv:2302.07241 (2023) 2, 4, 11, 12 +24. Ke, L., Ye, M., Danelljan, M., Liu, Y., Tai, Y.W., Tang, C.K., Yu, F.: Segment anything in high quality. In: NeurIPS (2023) 4 +25. Khoreva, A., Rohrbach, A., Schiele, B.: Video object segmentation with language referring expressions. In: ACCV (2018) 9, 11 +26. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) 4 +27. Li, K., Yang, Z., Chen, L., Yang, Y., Xun, J.: Catr: Combinatorial-dependence audio-queried transformer for audio-visual video segmentation. arXiv preprint arXiv:2309.09709 (2023) 4, 11 +28. Li, M., Li, S., Zhang, X., Zhang, L.: Univs: Unified and universal video segmentation with prompts as queries. arXiv preprint arXiv:2402.18115 (2024) 4 +29. Li, X., Cao, H., Zhao, S., Li, J., Zhang, L., Raj, B.: Panoramic video salient object detection with ambisonic audio guidance. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1424-1432 (2023) 2 +30. Li, X., Lin, C.C., Chen, Y., Liu, Z., Wang, J., Singh, R., Raj, B.: Paintseg: Painting pixels for training-free segmentation. In: Thirty-seventh Conference on Neural Information Processing Systems (2023) 2, 4 +31. Li, X., Wang, J., Li, X., Lu, Y.: Hybrid instance-aware temporal fusion for online video instance segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 1429-1437 (2022) 2 +32. Li, X., Wang, J., Li, X., Lu, Y.: Video instance segmentation by instance flow assembly. IEEE Transactions on Multimedia 25, 7469-7479 (2022) 2 +33. Li, X., Wang, J., Xu, X., Li, X., Raj, B., Lu, Y.: Robust referring video object segmentation with cyclic structural consensus. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22236-22245 (2023) 2, 10 +34. Li, X., Wang, J., Xu, X., Peng, X., Singh, R., Lu, Y., Raj, B.: Towards robust audiovisual segmentation in complex environments with quantization-based semantic decomposition. arXiv preprint arXiv:2310.00132 (2023) 2, 4, 11 + +35. Li, X., Wang, J., Xu, X., Yang, M., Yang, F., Zhao, Y., Singh, R., Raj, B.: Towards noise-tolerant speech-referring video object segmentation: Bridging speech and text. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. pp. 2283-2296 (2023) 4, 11 +36. Liang, T., He, Z., Jiao, W., Wang, X., Wang, Y., Wang, R., Yang, Y., Tu, Z., Shi, S.: Encouraging divergent thinking in large language models through multi-agent debate. arXiv preprint arXiv:2305.19118 (2023) 3 +37. Liu, J., Ding, H., Cai, Z., Zhang, Y., Satzoda, R.K., Mahadevan, V., Manmatha, R.: Polyformer: Referring image segmentation as sequential polygon generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18653-18663 (2023) 2, 4, 10 +38. Liu, J., Wang, Y., Ju, C., Ma, C., Zhang, Y., Xie, W.: Annotation-free audio-visual segmentation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5604-5614 (2024) 4 +39. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023) 4 +40. Liu, S., Hasan, A., Hong, K., Wang, R., Chang, P., Mizrachi, Z., Lin, J., McPherson, D.L., Rogers, W.A., Driggs-Campbell, K.: Dragon: A dialogue-based robot for assistive navigation with visual language grounding. IEEE Robotics and Automation Letters (2024) 2 +41. Mao, J., Huang, J., Toshev, A., Camburu, O., Yuille, A.L., Murphy, K.: Generation and comprehension of unambiguous object descriptions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 11-20 (2016) 9, 10 +42. Miao, B., Bennamoun, M., Gao, Y., Mian, A.: Spectrum-guided multi-granularity referring video object segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 920-930 (2023) 4, 10, 11 +43. Mo, S., Morgado, P.: A closer look at weakly-supervised audio-visual source localization. arXiv preprint arXiv:2209.09634 (2022) 4 +44. Pan, W., Shi, H., Zhao, Z., Zhu, J., He, X., Pan, Z., Gao, L., Yu, J., Wu, F., Tian, Q.: Wnet: Audio-guided video object segmentation via wavelet-based cross-modal denoising networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1320-1331 (2022) 4 +45. Pont-Tuset, J., Peruzzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675 (2017) 10 +46. Senocak, A., Oh, T.H., Kim, J., Yang, M.H., Kweon, I.S.: Learning to localize sound source in visual scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4358-4366 (2018) 4 +47. Seo, S., Lee, J.Y., Han, B.: Urvos: Unified referring video object segmentation network with a large-scale benchmark. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16. pp. 208-223. Springer (2020) 4, 9 +48. Sun, J., Huang, D.A., Lu, B., Liu, Y.H., Zhou, B., Garg, A.: Visually-grounded planning with transformers in procedural tasks. IEEE Robotics and Automation Letters 7(2), 4924-4930 (2022) 2 +49. Tang, J., Zheng, G., Yang, S.: Temporal collection and distribution for referring video object segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15466-15476 (2023) 4 + +50. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023) 7 +51. Tziafas, G., Kasaei, H.: Few-shot visual grounding for natural human-robot interaction. In: 2021 IEEE International Conference on Autonomous Robot Systems and Competitions (ICARSC). pp. 50-55. IEEE (2021) 2 +52. Wang, Z., Cai, S., Chen, G., Liu, A., Ma, X.S., Liang, Y.: Describe, explain, plan and select: interactive planning with llms enables open-world multi-task agents. Advances in Neural Information Processing Systems 36 (2024) 7 +53. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022) 8 +54. Wu, D., Wang, T., Zhang, Y., Zhang, X., Shen, J.: Onlinerefer: A simple online baseline for referring video object segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2761-2770 (2023) 2, 4, 10, 11 +55. Wu, J., Jiang, Y., Sun, P., Yuan, Z., Luo, P.: Language as queries for referring video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4974-4984 (2022) 4, 10 +56. Wu, J., Jiang, Y., Yan, B., Lu, H., Yuan, Z., Luo, P.: Segment every reference object in spatial and temporal spaces. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2538-2550 (2023) 5 +57. Xi, Z., Chen, W., Guo, X., He, W., Ding, Y., Hong, B., Zhang, M., Wang, J., Jin, S., Zhou, E., et al.: The rise and potential of large language model based agents: A survey. arXiv preprint arXiv:2309.07864 (2023) 7 +58. Xiong, Y., Varadarajan, B., Wu, L., Xiang, X., Xiao, F., Zhu, C., Dai, X., Wang, D., Sun, F., Iandola, F., et al.: Efficientsam: Leveraged masked image pretraining for efficient segment anything. arXiv preprint arXiv:2312.00863 (2023) 2, 4 +59. Xu, N., Yang, L., Fan, Y., Yue, D., Liang, Y., Yang, J., Huang, T.: Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327 (2018) 9, 11 +60. Xu, X., Wang, J., Li, X., Lu, Y.: Reliable propagation-correction modulation for video object segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 2946-2954 (2022) 4 +61. Xu, X., Wang, J., Ming, X., Lu, Y.: Towards robust video object segmentation with adaptive object calibration. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 2709-2718 (2022) 6 +62. Xu, X., Zhang, T., Wang, S., Li, X., Chen, Y., Li, Y., Raj, B., Johnson-Roberson, M., Huang, X.: Customizable perturbation synthesis for robust slam benchmarking. arXiv preprint arXiv:2402.08125 (2024) 6 +63. Xu, Z., Chen, Z., Zhang, Y., Song, Y., Wan, X., Li, G.: Bridging vision and language encoders: Parameter-efficient tuning for referring image segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17503-17512 (2023) 4, 10 +64. Yamazaki, K., Hanyu, T., Vo, K., Pham, T., Tran, M., Doretto, G., Nguyen, A., Le, N.: Open-fusion: Real-time open-vocabulary 3d mapping and queryable scene representation. arXiv preprint arXiv:2310.03923 (2023) 4, 12 +65. Yamazaki, K., Truong, S., Vo, K., Kidd, M., Rainwater, C., Luu, K., Le, N.: Vlcap: Vision-language with contrastive learning for coherent video paragraph captioning. In: 2022 IEEE International Conference on Image Processing (ICIP). pp. 3656-3661. IEEE (2022) 4 + +66. Yamazaki, K., Vo, K., Truong, Q.S., Raj, B., Le, N.: Vlnt: Visual-linguistic transformer-in-transformer for coherent video paragraph captioning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 3081-3090 (2023) 4 +67. Yang, J., Gao, M., Li, Z., Gao, S., Wang, F., Zheng, F.: Track anything: Segment anything meets videos (2023) 4 +68. Yang, Z., Wang, J., Tang, Y., Chen, K., Zhao, H., Torr, P.H.: Levt: Language-aware vision transformer for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18155-18165 (2022) 2, 4 +69. Yang, Z., Wang, J., Tang, Y., Chen, K., Zhao, H., Torr, P.H.: Levt: Language-aware vision transformer for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18155-18165 (2022) 10 +70. Yang, Z., Wei, Y., Yang, Y.: Associating objects with transformers for video object segmentation. Advances in Neural Information Processing Systems 34, 2491-2502 (2021) 4, 10 +71. Yang, Z., Yang, Y.: Decoupling features in hierarchical propagation for video object segmentation. Advances in Neural Information Processing Systems 35, 36324-36336 (2022) 4, 10, 11 +72. Yang, Z., Yang, Y.: Decoupling features in hierarchical propagation for video object segmentation. Advances in Neural Information Processing Systems 35, 36324-36336 (2022) 5 +73. Yao, J., Wang, X., Ye, L., Liu, W.: Matte anything: Interactive natural image matting with segment anything models. arXiv preprint arXiv:2306.04121 (2023) 4 +74. Yu, L., Poirson, P., Yang, S., Berg, A.C., Berg, T.L.: Modeling context in referring expressions. In: European Conference on Computer Vision. pp. 69-85. Springer (2016) 9, 10 +75. Yu, T., Feng, R., Feng, R., Liu, J., Jin, X., Zeng, W., Chen, Z.: Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790 (2023) 4 +76. Zhang, J., Cui, Y., Wu, G., Wang, L.: Joint modeling of feature, correspondence, and a compressed memory for video object segmentation. arXiv preprint arXiv:2308.13505 (2023) 5 +77. Zhao, Q., Wang, J., Zhang, Y., Jin, Y., Zhu, K., Chen, H., Xie, X.: Competeai: Understanding the competition behaviors in large language model-based agents. arXiv preprint arXiv:2310.17512 (2023) 3 +78. Zhou, J., Shen, X., Wang, J., Zhang, J., Sun, W., Zhang, J., Birchfield, S., Guo, D., Kong, L., Wang, M., et al.: Audio-visual segmentation with semantics. arXiv preprint arXiv:2301.13190 (2023) 4, 9 +79. Zhou, J., Wang, J., Zhang, J., Sun, W., Zhang, J., Birchfield, S., Guo, D., Kong, L., Wang, M., Zhong, Y.: Audio-visual segmentation. In: European Conference on Computer Vision (2022) 2, 4, 11 +80. Zhu, C., Chen, F., Ahmed, U., Shen, Z., Savvides, M.: Semantic relation reasoning for shot-stable few-shot object detection. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 8778-8787 (2021), https://api.sementicscholar.org/CorpusID:232093016 2 +81. Zhu, C., Chen, F., Shen, Z., Savvides, M.: Soft anchor-point object detection. In: European Conference on Computer Vision (2019), https://api-semanticscholar.org/CorpusID:208512715 2 + +82. Zou, X., Dou, Z.Y., Yang, J., Gan, Z., Li, L., Li, C., Dai, X., Behl, H., Wang, J., Yuan, L., et al.: Generalized decoding for pixel, image, and language. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15116-15127 (2023) 4, 10 +83. Zou, X., Yang, J., Zhang, H., Li, F., Li, L., Gao, J., Lee, Y.J.: Segment everything everywhere all at once. arXiv preprint arXiv:2304.06718 (2023) 4, 10 +84. Zou, X., Yang, J., Zhang, H., Li, F., Li, L., Wang, J., Wang, L., Gao, J., Lee, Y.J.: Segment everything everywhere all at once. Advances in Neural Information Processing Systems 36 (2024) 4, 10 \ No newline at end of file diff --git a/2024/R^2-Bench_ Benchmarking the Robustness of Referring Perception Models under Perturbations/images.zip b/2024/R^2-Bench_ Benchmarking the Robustness of Referring Perception Models under Perturbations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8229c423194187260cb5cafcb52251d84313d2a4 --- /dev/null +++ b/2024/R^2-Bench_ Benchmarking the Robustness of Referring Perception Models under Perturbations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0afb2cd92283c00464290b076a6078e90f4054ca92b17044482591e9a80befe0 +size 701005 diff --git a/2024/R^2-Bench_ Benchmarking the Robustness of Referring Perception Models under Perturbations/layout.json b/2024/R^2-Bench_ Benchmarking the Robustness of Referring Perception Models under Perturbations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ba2613596326466170107230c69da855febb4f5c --- /dev/null +++ b/2024/R^2-Bench_ Benchmarking the Robustness of Referring Perception Models under Perturbations/layout.json @@ -0,0 +1,10884 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 480, + 150 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 480, + 150 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 480, + 150 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 133, + 114, + 480, + 150 + ], + "type": "text", + "content": "-Bench: Benchmarking the Robustness of Referring Perception Models under Perturbations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 151, + 170, + 461, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 170, + 461, + 196 + ], + "spans": [ + { + "bbox": [ + 151, + 170, + 461, + 196 + ], + "type": "text", + "content": "Xiang Li" + }, + { + "bbox": [ + 151, + 170, + 461, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 151, + 170, + 461, + 196 + ], + "type": "text", + "content": ", Kai Qiu" + }, + { + "bbox": [ + 151, + 170, + 461, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 151, + 170, + 461, + 196 + ], + "type": "text", + "content": ", Jinglu Wang" + }, + { + "bbox": [ + 151, + 170, + 461, + 196 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 151, + 170, + 461, + 196 + ], + "type": "text", + "content": ", Xiaohao Xu" + }, + { + "bbox": [ + 151, + 170, + 461, + 196 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 151, + 170, + 461, + 196 + ], + "type": "text", + "content": ", Rita Singh" + }, + { + "bbox": [ + 151, + 170, + 461, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 151, + 170, + 461, + 196 + ], + "type": "text", + "content": ", Kashu Yamazaki" + }, + { + "bbox": [ + 151, + 170, + 461, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 151, + 170, + 461, + 196 + ], + "type": "text", + "content": ", Hao Chen" + }, + { + "bbox": [ + 151, + 170, + 461, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 151, + 170, + 461, + 196 + ], + "type": "text", + "content": ", Xiaonan Huang" + }, + { + "bbox": [ + 151, + 170, + 461, + 196 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 151, + 170, + 461, + 196 + ], + "type": "text", + "content": ", Bhiksha Raj" + }, + { + "bbox": [ + 151, + 170, + 461, + 196 + ], + "type": "inline_equation", + "content": "^{1,4}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 157, + 203, + 455, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 203, + 455, + 216 + ], + "spans": [ + { + "bbox": [ + 157, + 203, + 455, + 216 + ], + "type": "text", + "content": "1 CMU, 2 Microsoft Research Asia, 3 University of Michigan, 4 MBZUAI" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 224, + 302, + 320 + ], + "blocks": [ + { + "bbox": [ + 133, + 224, + 302, + 320 + ], + "lines": [ + { + "bbox": [ + 133, + 224, + 302, + 320 + ], + "spans": [ + { + "bbox": [ + 133, + 224, + 302, + 320 + ], + "type": "image", + "image_path": "10aae151004360f04481c045134b269f97efaca8992350082a8d8f6834a038d2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 322, + 276, + 331 + ], + "lines": [ + { + "bbox": [ + 170, + 322, + 276, + 331 + ], + "spans": [ + { + "bbox": [ + 170, + 322, + 276, + 331 + ], + "type": "text", + "content": "(a) Referring Perception in Practice" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 320, + 224, + 481, + 321 + ], + "blocks": [ + { + "bbox": [ + 320, + 224, + 481, + 321 + ], + "lines": [ + { + "bbox": [ + 320, + 224, + 481, + 321 + ], + "spans": [ + { + "bbox": [ + 320, + 224, + 481, + 321 + ], + "type": "image", + "image_path": "27f810604090a0d5ebf36812144a519207af7753772bb7335da1cf77b9cf7049.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 344, + 322, + 459, + 331 + ], + "lines": [ + { + "bbox": [ + 344, + 322, + 459, + 331 + ], + "spans": [ + { + "bbox": [ + 344, + 322, + 459, + 331 + ], + "type": "text", + "content": "(b) Perturbation in Referring Guidance" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 335, + 482, + 434 + ], + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 434 + ], + "type": "text", + "content": "Fig. 1: Motivation illustration. Referring perception models (RPMs) empower intelligent systems with their ability to perform object grounding within the environment based on referring guidance, such as textual descriptions, imagery exemplars, or auditory signals associated with the target object. However, RPMs' performance can be compromised by disturbances in real-world scenarios, such as environmental noise (e.g., extraneous sounds from a nearby radio), human-induced errors (e.g., typographical errors in textual input), and limitations in the sensor (e.g., motion blur in images). Conducting a rigorous analysis of RPMs' robustness to a wide array of perturbations is necessary for building reliable real-world applications." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 446, + 455, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 446, + 455, + 644 + ], + "spans": [ + { + "bbox": [ + 159, + 446, + 455, + 644 + ], + "type": "text", + "content": "Abstract. Referring perception, which aims at grounding visual objects with multimodal referring guidance, is essential for bridging the gap between humans, who provide instructions, and the environment where intelligent systems perceive. Despite progress in this field, the robustness of referring perception models (RPMs) against disruptive perturbations is not well explored. This work thoroughly assesses the resilience of RPMs against various perturbations in both general and specific contexts. Recognizing the complex nature of referring perception tasks, we present a comprehensive taxonomy of perturbations, and then develop a versatile toolbox for synthesizing and evaluating the effects of composite disturbances. Employing this toolbox, we construct " + }, + { + "bbox": [ + 159, + 446, + 455, + 644 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 159, + 446, + 455, + 644 + ], + "type": "text", + "content": "-Bench, a benchmark for assessing the Robustness of Referring perception models under noisy conditions across five key tasks. Moreover, we propose the " + }, + { + "bbox": [ + 159, + 446, + 455, + 644 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 159, + 446, + 455, + 644 + ], + "type": "text", + "content": "-Agent, an LLM-based agent that simplifies and automates model evaluation via natural language instructions. Our investigation uncovers the vulnerabilities of current RPMs to various perturbations and provides tools for assessing model robustness, potentially promoting the safe and resilient integration of intelligent systems into complex real-world scenarios." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 159, + 648, + 455, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 648, + 455, + 670 + ], + "spans": [ + { + "bbox": [ + 159, + 648, + 455, + 670 + ], + "type": "text", + "content": "Keywords: Referring perception " + }, + { + "bbox": [ + 159, + 648, + 455, + 670 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 648, + 455, + 670 + ], + "type": "text", + "content": " Robustness & Perturbation " + }, + { + "bbox": [ + 159, + 648, + 455, + 670 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 648, + 455, + 670 + ], + "type": "text", + "content": " Benchmark" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 474, + 221 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 474, + 221 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 474, + 221 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 474, + 221 + ], + "type": "image", + "image_path": "5a13cc43f38aaba38e8329e201d3c93c8a82d48baa9f88ae1a982df887bdbf27.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 225, + 482, + 270 + ], + "lines": [ + { + "bbox": [ + 130, + 225, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 482, + 270 + ], + "type": "text", + "content": "Fig. 2: Examples of " + }, + { + "bbox": [ + 130, + 225, + 482, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 130, + 225, + 482, + 270 + ], + "type": "text", + "content": "-Bench. The \"Original\" row displays the original inputs alongside the outcomes from models of R-VOS [54], AVS [34] and Q3M [23], while the \"Perturbed\" row presents the inputs as synthesized by " + }, + { + "bbox": [ + 130, + 225, + 482, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 130, + 225, + 482, + 270 + ], + "type": "text", + "content": "-Bench and the respective outcomes with the same models. RG: short for Referring Guidance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 293, + 230, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 230, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 230, + 306 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 318, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 482, + 450 + ], + "type": "text", + "content": "Perception systems [6-8,29-32,80,81] function as input channels for intelligent systems, analogous to human eyes. Referring perception, focuses on identification and contextualization of visual entities by referring multimodal guidance, illustrated in Fig. 1 (a). It effectively creates a communicative bridge between humans, who issue instructions, and the environment that is subject to perception. Multimodal referring cues used for such identification include textual descriptions, example imagery, or auditory signals corresponding to the target object (Fig. 1 (b)). Within the domain of referring perception, specific tasks such as referring expression segmentation [37,68], audiovisual source localization [17,34,79], and queryable 3D mapping [23], are essential modules of robotic control [2,22], autonomous navigation [23,40] and planning [21,48], and human-computer interaction [51]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 450, + 483, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 483, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 483, + 641 + ], + "type": "text", + "content": "Recent advancements in referring perception models [14, 33, 68] have been witnessed and deployed on resource-constrained platforms [54, 58]. However, existing datasets for model training often comprise delicate images with accurate annotations, a condition rarely met in real-world scenarios. In realistic settings, as illustrated in Fig. 1, perception systems may face various disturbances, such as environmental noise (buzzing sound from the radio), imprecise human instructions (misspelling), and sensor imperfections (image blur), which can significantly challenge the robustness of these models. Despite their critical importance for safer real-world applications, the impact of such perturbations on model performance is not thoroughly investigated in the literature. Evaluating model robustness is further complicated by the diversity of operational environments; the assessment process is typically labor-intensive and relies heavily on expert judgment. Addressing this gap, this paper introduces the first benchmark in the field to systematically assess the resilience of referring perception models, i.e., trying to obtain a reliable and comprehensive answer to the critical question: \"How robust are the current referring perception models under realistic perturbations?\"" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 483, + 666 + ], + "type": "text", + "content": "Specifically, we introduce the Robust Referring Benchmark " + }, + { + "bbox": [ + 130, + 641, + 483, + 666 + ], + "type": "inline_equation", + "content": "(\\mathbf{R}^2" + }, + { + "bbox": [ + 130, + 641, + 483, + 666 + ], + "type": "text", + "content": "-Bench), featuring: (1) a human-friendly taxonomy of perturbations for referring perception," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Xiang Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": "categorizing disruptions into source, environment, sensor, and transmission noise, and detailing perturbation types for each modality; (2) a customizable perturbation synthesis toolbox for creating reasonable noise-augmented datasets, enabling robustness evaluation of perception models; and (3) robustness evaluations across five tasks—referring image segmentation (RIS), video object segmentation (VOS), referring video object segmentation (R-VOS), audiovisual segmentation (AVS), and queryable 3D mapping (Q3M)—considering textual, visual, and acoustic reference modalities. We assess over twenty prevalent models in noisy conditions to benchmark their performance and analyze their vulnerabilities. (Examples are illustrated in Fig. 2.) We hope our benchmarking and analysis can benefit the community by shedding light on the vulnerability of existing models in the face of various perturbations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 261, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 489 + ], + "type": "text", + "content": "Notably, our work advances the evaluation of model robustness by incorporating tests against composite noise types, more closely mirroring real-world scenarios where different noises often coexist. The construction of complex noise for realistic robustness assessment is not trivial, as it must avoid non-contextual perturbations, such as the presence of snow in indoor imagery or excessive air absorption effects in indoor audio recordings, which are incongruent with the depicted environment. To address this complexity, we introduce the " + }, + { + "bbox": [ + 130, + 261, + 482, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 130, + 261, + 482, + 489 + ], + "type": "text", + "content": "-Agent, a novel large language model-based (LLM-based) system that can comprehend natural language instructions provided by humans and autonomously generates test samples with contextual relevant perturbations that align with the deployment environment. For instance, given the human-provided instruction \"outdoor night scene\", the " + }, + { + "bbox": [ + 130, + 261, + 482, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 130, + 261, + 482, + 489 + ], + "type": "text", + "content": "-Agent could probably generate perturbations of \"motion blur\" instead of \"glass blur\" and \"low background noise\" instead of \"room reverberation\". To improve the agent's abilities of instruction-following and commonsense-reasoning, we employ the multi-agent debating technique [16,36,77], wherein one agent proposes potential solutions while another verifies their validity. Consequently, the " + }, + { + "bbox": [ + 130, + 261, + 482, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 130, + 261, + 482, + 489 + ], + "type": "text", + "content": "-Agent efficiently automates the identification and integration of mixed perturbations, thus optimizing the process of robustness evaluations tailored to specific domain contexts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 491, + 366, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 491, + 366, + 502 + ], + "spans": [ + { + "bbox": [ + 146, + 491, + 366, + 502 + ], + "type": "text", + "content": "In summary, our major contribution is three-fold:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 517, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 517, + 481, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 517, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 517, + 481, + 566 + ], + "type": "text", + "content": "- We introduce the " + }, + { + "bbox": [ + 138, + 517, + 481, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 138, + 517, + 481, + 566 + ], + "type": "text", + "content": "-Bench, a benchmarking framework that includes diverse perturbed data for five commonly encountered referring perception tasks. A comprehensive taxonomy and a corresponding customizable synthesis toolbox are developed to enable robustness assessment in noisy real-world settings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 567, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 482, + 616 + ], + "type": "text", + "content": "- We propose the " + }, + { + "bbox": [ + 138, + 567, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 138, + 567, + 482, + 616 + ], + "type": "text", + "content": "-Agent to streamline model evaluations tailored to particular use scenarios based on LLMs. This automated program executes the perturbation composition process in accordance with human-provided instructions, thereby allowing for context-specific robustness evaluations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 617, + 482, + 665 + ], + "type": "text", + "content": "- Our systematic experimental investigations delve into the intrinsic characteristics of perturbations, such as their types, severity, dynamics, and correlations. These explorations yield valuable insights into the vulnerability of existing models to disturbances and elucidate the nature of these perturbations." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 132, + 91, + 470, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 91, + 470, + 102 + ], + "spans": [ + { + "bbox": [ + 132, + 91, + 470, + 102 + ], + "type": "text", + "content": "Benchmarking the Robustness of Referring Perception Models under Perturbations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 496, + 91, + 503, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 496, + 91, + 503, + 101 + ], + "spans": [ + { + "bbox": [ + 496, + 91, + 503, + 101 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 243, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 243, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 243, + 127 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "text", + "content": "Textual referring perception. Referring image segmentation (RIS) and referring video object segmentation (R-VOS) aim to segment objects in images and video sequences, respectively, based on a linguistic description. Recent RIS methods [15, 37,63,68,82-84] have achieved promising results using multimodal transformers. R-VOS [14,18,42,47,49,54] is more challenging as it requires leveraging both intra-frame and temporal cues. URVOS [47] is the first unified R-VOS framework with cross-modal attention and a memory attention module, significantly improving R-VOS performance. ReferFormer [55] employs a linguistic prior in the transformer decoder to focus on the referred object, while MTTR [5] uses a multimodal transformer encoder to fuse linguistic and visual features. " + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "text", + "content": "-VOS introduces relational cyclic consistency to enhance the robustness of the R-VOS model. Unlike other vision-language tasks [65,66], R-VOS needs to construct object-level multimodal semantic consensus in dense visual representations. Relying on 2D models, 3D referring perception is also achievable [23, 64]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 318, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 482, + 548 + ], + "type": "text", + "content": "Acoustic referring perception. Audiovisual segmentation (AVS) [35, 35, 38, 78] focuses on segmenting objects that produce sound in a given image frame. This pioneering work by Zhou et al. [78] introduced a method that uses cross-modal attention to identify the sound source. Building on this, Zhou et al. [78] proposed an extended task called audiovisual semantic segmentation (AVSS), which not only segments the sound-producing objects but also classifies them. AVSS is more challenging than AVS due to the complexity of audio semantics. To address this, Zhou et al. [78] utilized the TPAVI module from [79] for audiovisual interaction. Additionally, a recent development by Li et al. [27] introduced the CATR framework, which features a novel spatial-temporal audio-video fusion block for efficient audio-visual integration. Li et al. [34] propose a quantization-based semantic decomposition module to handle the complex acoustic environment and enhance the robustness. Sound source localization (SSL) is a related field that aims to identify the visual regions corresponding to sounds. Several SSL methods [3, 4, 12, 46] utilize the correspondence between audio and visual features to locate sounds, often represented as heatmaps. For example, Mo et al. [43] employed multi-level audiovisual contrastive learning to effectively pinpoint sound-producing objects. In addition, using speech as referring guidance is also explored in [35, 44]" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Visual referring perception. Visual referring perception typically relies on visual prompts such as example images, points, bounding boxes, or scribbles. Segment anything model (SAM) [26] is a powerful foundation model that supports multiple visual prompts. Several follow-up works [24,39,58] improve SAM in terms of inference cost, segmentation quality, and referring prompts. Beyond that, several works [67,73,75] explore the usage of SAM in downstream tasks. Li et al. [30] built a SAM-like model with stable diffusion. For video-level tasks, semi-video object segmentation (VOS) [9,28,60,70,71] aims to segment visual objects across frames given the first frame annotation. Recent offline models [9,11] focus more" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Xiang Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "on designing long-term information propagation modules to transfer previous image features and corresponding masks to the target frame to predict masks. This helps more precisely identify and track the movement trajectory of objects throughout the entire video sequence (more than 200 frames) but also makes the network architecture heavy [72]. Due to the slow speed of offline models, online models [56, 76] come back into focus, aiming to keep the right balance between speed and performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 218, + 451, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 218, + 451, + 232 + ], + "spans": [ + { + "bbox": [ + 131, + 218, + 451, + 232 + ], + "type": "text", + "content": "3 " + }, + { + "bbox": [ + 131, + 218, + 451, + 232 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 131, + 218, + 451, + 232 + ], + "type": "text", + "content": "-Bench: Customizable Perturbation Benchmark" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 243, + 481, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 481, + 293 + ], + "type": "text", + "content": "In this section, we present the " + }, + { + "bbox": [ + 130, + 243, + 481, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 130, + 243, + 481, + 293 + ], + "type": "text", + "content": "-Bench for evaluating robustness of referring perception tasks, including a comprehensive taxonomy that categorizes various perturbations and a customizable perturbation synthesis toolbox that can be tailored to generate specific disturbances." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 310, + 411, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 310, + 411, + 323 + ], + "spans": [ + { + "bbox": [ + 131, + 310, + 411, + 323 + ], + "type": "text", + "content": "3.1 Taxonomic Perturbations in Referring Perception" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 330, + 293, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 293, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 293, + 485 + ], + "type": "text", + "content": "Referring perception tasks typically employ multimodal guidance, spanning the visual, auditory, and textural domains, to ground objects in the visual contexts. Nonetheless, in real-world scenarios, each modality may undergo heterogeneous noises before input into the model. As depicted in Fig. 3, noises can be categorized into four classes based on their origins: source noises, environment noises, sensor noises, and transmission noises." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 301, + 331, + 474, + 437 + ], + "blocks": [ + { + "bbox": [ + 301, + 331, + 474, + 437 + ], + "lines": [ + { + "bbox": [ + 301, + 331, + 474, + 437 + ], + "spans": [ + { + "bbox": [ + 301, + 331, + 474, + 437 + ], + "type": "image", + "image_path": "91b81622bee6eaa89487e40ef6b687a51d3a2312cfde49d3d3c2d7580915fa72.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 441, + 482, + 484 + ], + "lines": [ + { + "bbox": [ + 296, + 441, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 296, + 441, + 482, + 484 + ], + "type": "text", + "content": "Fig. 3: Noise categories based on their origins. Assuming airplane as the source of referring guidance, noise from it is categorized as source noise." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 482, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 558 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 558 + ], + "type": "text", + "content": "Source noises. We define source noise as the inherent disturbances introduced spontaneously by the source of the referring guidance. For example, for text-referring tasks, the source noises can manifest as misspellings, grammar errors, or punctuation errors in the textual expression and instruction (human is usually the source of text reference)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 482, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 618 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 618 + ], + "type": "text", + "content": "Environment noises. We define environment noise as the unintended disturbance introduced by the surrounding environment during the signal capture process. Background sound, room reverberation and air absorption for acoustic waveforms are common examples of this type." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "Sensor noises. We define sensor noise as the unwanted disturbances in the output of sensors that arise from inherent limitations or imperfections. For example, defocus blur, motion blur and impulse noise for visual frames." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 132, + 91, + 470, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 91, + 470, + 102 + ], + "spans": [ + { + "bbox": [ + 132, + 91, + 470, + 102 + ], + "type": "text", + "content": "Benchmarking the Robustness of Referring Perception Models under Perturbations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 496, + 91, + 503, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 496, + 91, + 503, + 101 + ], + "spans": [ + { + "bbox": [ + 496, + 91, + 503, + 101 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 128, + 480, + 221 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 478, + 125 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 478, + 125 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 478, + 125 + ], + "type": "text", + "content": "Table 1: Noise types considered in " + }, + { + "bbox": [ + 133, + 114, + 478, + 125 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 133, + 114, + 478, + 125 + ], + "type": "text", + "content": "-Bench. Details are available in the Appendix." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 128, + 480, + 221 + ], + "lines": [ + { + "bbox": [ + 134, + 128, + 480, + 221 + ], + "spans": [ + { + "bbox": [ + 134, + 128, + 480, + 221 + ], + "type": "table", + "html": "
TypeSourceEnvironmentSensorTransmission
Visual-snow (SN), fog (FG), frost (FT), spatter (SP), brightness (BR)defocus blur (DB), gaussian blur (GB), motion blur (MB), glass blur (GS), impulse noise (IN), shot noise (ST), speckle noise (SPN), contrast (CT), saturate (SA)JPEG compression (JPG), pixelated (PIX)
Acousticamplitude gain (GA)background noise (BN), air absorption (AA), room reverberation (RS)gaussian noise (GN), impulse noise (IN), peak filter (PF), time mask (TM), tanh distortion (TD)MP3 compression (MP3), lowpass filter (LP), highpass filter (HP)
Textualmisspelling (MS), mispunctuation (MP), grammar error (GE)-character missing (CM)-
", + "image_path": "2271e41a80c7137d63a06b809ae49c2bebdee9ab206335689ff40514be8380dd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 245, + 479, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 245, + 479, + 280 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 479, + 280 + ], + "type": "text", + "content": "Transmission noises. We define transmission noise as the signal distortion or loss during the transmission resulting from data compression or package loss. For example, JPEG and MP4 compression for visual and acoustic signals." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 281, + 480, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 480, + 316 + ], + "type": "text", + "content": "We summarize the supported noises in visual, acoustic and textual modalities in Tab. 1. A total of 32 types of noises are considered in this paper. The implementation details of noise functions are available in the Appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 335, + 348, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 335, + 348, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 335, + 348, + 346 + ], + "type": "text", + "content": "3.2 Customizable Perturbation Synthesis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 356, + 482, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 417 + ], + "type": "text", + "content": "We approach the robustness evaluation through two conventional paradigms: (1) general evaluation under universal scenarios and (2) specific evaluation under designated instructions. To facilitate these evaluations, it is essential to generate perturbed datasets from the original clean data samples, denoted as " + }, + { + "bbox": [ + 130, + 356, + 482, + 417 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\{x\\}" + }, + { + "bbox": [ + 130, + 356, + 482, + 417 + ], + "type": "text", + "content": ", by introducing a spectrum of heterogeneous perturbations " + }, + { + "bbox": [ + 130, + 356, + 482, + 417 + ], + "type": "inline_equation", + "content": "\\varDelta=\\{\\delta_k\\}" + }, + { + "bbox": [ + 130, + 356, + 482, + 417 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 429, + 482, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 482, + 524 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 482, + 524 + ], + "type": "text", + "content": "Perturbation order. As discussed in Sec. 3.1, perturbations arise sequentially according to their origins in real-world contexts. To emulate this, perturbations are applied in a sequence that mirrors the natural order of disruption: source " + }, + { + "bbox": [ + 130, + 429, + 482, + 524 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 429, + 482, + 524 + ], + "type": "text", + "content": " environment " + }, + { + "bbox": [ + 130, + 429, + 482, + 524 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 429, + 482, + 524 + ], + "type": "text", + "content": " sensor " + }, + { + "bbox": [ + 130, + 429, + 482, + 524 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 429, + 482, + 524 + ], + "type": "text", + "content": " transmission. Within a single category of origin, the order is randomly determined to account for the non-deterministic nature of real-world noise. To formalize the perturbation process, consider a clean data sample denoted by " + }, + { + "bbox": [ + 130, + 429, + 482, + 524 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 429, + 482, + 524 + ], + "type": "text", + "content": ", the perturbed sample " + }, + { + "bbox": [ + 130, + 429, + 482, + 524 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 130, + 429, + 482, + 524 + ], + "type": "text", + "content": " undergoes a series of transformations that can be mathematically described as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 252, + 535, + 481, + 547 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 535, + 481, + 547 + ], + "spans": [ + { + "bbox": [ + 252, + 535, + 481, + 547 + ], + "type": "interline_equation", + "content": "x ^ {\\prime} = \\delta_ {t} \\circ \\delta_ {s e} \\circ \\delta_ {e} \\circ \\delta_ {s o} (x), \\tag {1}", + "image_path": "e904201544ef94e72113f51e19438849f27b4ccd03ce232ba1dce8ffeacd403f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 557, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 557, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 130, + 557, + 482, + 605 + ], + "type": "text", + "content": " denotes composition operation. " + }, + { + "bbox": [ + 130, + 557, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\delta_{so},\\delta_e,\\delta_{se},\\delta_t" + }, + { + "bbox": [ + 130, + 557, + 482, + 605 + ], + "type": "text", + "content": " represent generation functions for source, environment, sensor, and transmission noises, respectively. Given that the introduction of perturbations follows a sequential order, the operations that compose these noise functions are inherently order-variant." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 482, + 666 + ], + "type": "text", + "content": "Perturbation severity and mode. In alignment with established methodologies [20, 61, 62], we quantitatively define perturbation severity across discrete levels that align with human perceptual categories, namely, {low, medium, high}. In addition, we investigate perturbation across two main modes based on dynamics:" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Xiang Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 479, + 232 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 479, + 232 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 479, + 232 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 479, + 232 + ], + "type": "image", + "image_path": "96abbc7c9eeb9ae3d7a739ee1cea094ab1171bf9222f7b7e973cdaf97bc4a3c3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 236, + 482, + 293 + ], + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 293 + ], + "type": "text", + "content": "Fig. 4: Overview of " + }, + { + "bbox": [ + 130, + 236, + 482, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 130, + 236, + 482, + 293 + ], + "type": "text", + "content": "-Agent, the automatic evaluation assistant. Given a human instruction, clean datasets, perturbation functions, and evaluation functions, " + }, + { + "bbox": [ + 130, + 236, + 482, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 130, + 236, + 482, + 293 + ], + "type": "text", + "content": "-Agent first proposes and verifies perturbed test samples that match the given instruction. After that, " + }, + { + "bbox": [ + 130, + 236, + 482, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 130, + 236, + 482, + 293 + ], + "type": "text", + "content": "-Agent evaluates the model using the verified samples and provides a report that articulates the model's vulnerabilities and overall resilience." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 316, + 482, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 365 + ], + "type": "text", + "content": "{static, dynamic}. Static perturbations remain a consistent level of severity throughout all sensor frames within a sequence, while dynamic perturbations exhibit varying severity and types from frame to frame, closely simulating the fluctuating nature of disturbances in real-world environments, taking the form:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 257, + 374, + 482, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 374, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 257, + 374, + 482, + 389 + ], + "type": "interline_equation", + "content": "x ^ {\\prime} = \\circ_ {k = 1} ^ {K} \\delta_ {k} (x, \\lambda_ {k} (t)), \\tag {2}", + "image_path": "ec8c4edb3288d39d3863aae185c9101573b4a91eabf7c57c5f89c5721e15d0f4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 399, + 449, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 449, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 449, + 411 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 399, + 449, + 411 + ], + "type": "inline_equation", + "content": "\\lambda_{k}(t)" + }, + { + "bbox": [ + 130, + 399, + 449, + 411 + ], + "type": "text", + "content": " denotes the severity of the " + }, + { + "bbox": [ + 130, + 399, + 449, + 411 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 399, + 449, + 411 + ], + "type": "text", + "content": "-th perturbation at time step " + }, + { + "bbox": [ + 130, + 399, + 449, + 411 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 399, + 449, + 411 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 411, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 495 + ], + "type": "text", + "content": "For the general evaluation under universal scenarios, we construct perturbed datasets for each task by leveraging the original clean benchmarks and all perturbations. Specifically, noisy datasets with low, medium, high and dynamic severity are considered. In each setting, a maximum of two perturbations are considered simultaneously. The detailed data-creating procedure is available in the Appendix. Comprehensive experiments and analyses discussing the impact of perturbations are provided to facilitate future research in Sec. 5.3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 496, + 483, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 483, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 483, + 532 + ], + "type": "text", + "content": "Furthermore, we also consider specific evaluations under designated instructions in practice, e.g., virtual phone meetings without background noise. We detail " + }, + { + "bbox": [ + 130, + 496, + 483, + 532 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 130, + 496, + 483, + 532 + ], + "type": "text", + "content": "-Agent in Sec. 4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 552, + 407, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 552, + 407, + 567 + ], + "spans": [ + { + "bbox": [ + 131, + 552, + 407, + 567 + ], + "type": "text", + "content": "4 " + }, + { + "bbox": [ + 131, + 552, + 407, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 131, + 552, + 407, + 567 + ], + "type": "text", + "content": "-Agent: Automatic Evaluation Assistant" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "With the versatile capabilities of LLMs [1, 50], LLM-based agents [52, 57] are employed to automatically perform various tasks in a human-like manner. In this section, we introduce " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "-Agent, an LLM-based agent that facilitates automatic perturbation synthesis and robustness evaluation guided by human instructions. As depicted in Fig. 4, the " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "-Agent consists of three primary steps: data proposal, data verification, and model analysis. We utilize Gemini-Vision-Pro [50] as the LLM component within the " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "-Agent." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 132, + 91, + 470, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 91, + 470, + 102 + ], + "spans": [ + { + "bbox": [ + 132, + 91, + 470, + 102 + ], + "type": "text", + "content": "Benchmarking the Robustness of Referring Perception Models under Perturbations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 496, + 91, + 503, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 496, + 91, + 503, + 100 + ], + "spans": [ + { + "bbox": [ + 496, + 91, + 503, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 116, + 145, + 124 + ], + "blocks": [ + { + "bbox": [ + 135, + 116, + 145, + 124 + ], + "lines": [ + { + "bbox": [ + 135, + 116, + 145, + 124 + ], + "spans": [ + { + "bbox": [ + 135, + 116, + 145, + 124 + ], + "type": "image", + "image_path": "cdd590ad2708d573ea6e9a599e553b4cf4979a09c8bd5f0644212171300735af.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 118, + 168, + 124 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 118, + 168, + 124 + ], + "spans": [ + { + "bbox": [ + 146, + 118, + 168, + 124 + ], + "type": "text", + "content": "Example" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 144, + 170, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 144, + 170, + 152 + ], + "spans": [ + { + "bbox": [ + 136, + 144, + 170, + 152 + ], + "type": "text", + "content": "② Question" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 174, + 117, + 478, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 117, + 478, + 179 + ], + "spans": [ + { + "bbox": [ + 174, + 117, + 478, + 179 + ], + "type": "text", + "content": "Given a human instruction, you need to verify...Here is the list of all perturbations you need to consider: {ALL_PERTURBATIONS}. You will be given the results selected from another model as input. For example, if the input is {\"type\": \"snoweffect\", \"reason\": \"the snow effect is a noise type\", \"selected\": True} and the instruction is \"Indoor zoom meeting\", then this data sample is not correct since the instruction pointing to an indoor scenario while the snow effect is outdoor. You need to correct the \"selected\" and \"reason\" items in the dict... Now consider an of {HUMAN_INSTRUCTION}, what are the samples possibly happened within the following perturbations? Explain the reason. Reply me in a strict format of python list. For example, [\"type\": noise_type1, \"reason\": reason1, \"selected\": True/False], . Selected perturbations: {SELECTED_PERTURBATIONS}. \n``python \nperturbations = [\"type\": \"snoweffect\", \"reason\": \"...\", \"selected\": False], [\"type\": \"fog effect\", \"reason\": \"...\", \"selected\": False]," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 187, + 482, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 187, + 482, + 265 + ], + "spans": [ + { + "bbox": [ + 132, + 187, + 482, + 265 + ], + "type": "text", + "content": "Fig. 5: Chain-of-thought prompting template for data verification. " + }, + { + "bbox": [ + 132, + 187, + 482, + 265 + ], + "type": "inline_equation", + "content": "①\\&②" + }, + { + "bbox": [ + 132, + 187, + 482, + 265 + ], + "type": "text", + "content": " Following the chain-of-thought spirit, we first give examples to LLM to boost the in-context-learning capability. After that, we ask the LLM to answer a question that is similar to the given example. Specifically, for the data verification task, we ask the LLM to verify the selected samples from the previous iteration, update the results, and explain the reason. The LLM response is instructed to be a Python-format list of dictionaries. " + }, + { + "bbox": [ + 132, + 187, + 482, + 265 + ], + "type": "inline_equation", + "content": "③" + }, + { + "bbox": [ + 132, + 187, + 482, + 265 + ], + "type": "text", + "content": " : The response from LLM (which follows the desired Python-format list)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 280, + 260, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 280, + 260, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 280, + 260, + 292 + ], + "type": "text", + "content": "4.1 Evaluation Pipeline" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 300, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 422 + ], + "type": "text", + "content": "Given a set of clean data samples " + }, + { + "bbox": [ + 130, + 300, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\{x_{i}\\}" + }, + { + "bbox": [ + 130, + 300, + 482, + 422 + ], + "type": "text", + "content": ", perturbation functions " + }, + { + "bbox": [ + 130, + 300, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\varDelta=\\{\\delta_k\\}" + }, + { + "bbox": [ + 130, + 300, + 482, + 422 + ], + "type": "text", + "content": ", evaluation metric functions " + }, + { + "bbox": [ + 130, + 300, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = \\{e_j\\}" + }, + { + "bbox": [ + 130, + 300, + 482, + 422 + ], + "type": "text", + "content": ", and a human instruction " + }, + { + "bbox": [ + 130, + 300, + 482, + 422 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 130, + 300, + 482, + 422 + ], + "type": "text", + "content": ", the objective of R²-Agent is to select an appropriate subset of clean data samples " + }, + { + "bbox": [ + 130, + 300, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_u" + }, + { + "bbox": [ + 130, + 300, + 482, + 422 + ], + "type": "text", + "content": ", potential perturbations " + }, + { + "bbox": [ + 130, + 300, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\varDelta_u" + }, + { + "bbox": [ + 130, + 300, + 482, + 422 + ], + "type": "text", + "content": ", and suitable evaluation functions " + }, + { + "bbox": [ + 130, + 300, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_u" + }, + { + "bbox": [ + 130, + 300, + 482, + 422 + ], + "type": "text", + "content": " to assess the model based on the given instruction " + }, + { + "bbox": [ + 130, + 300, + 482, + 422 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 130, + 300, + 482, + 422 + ], + "type": "text", + "content": ". A desired human instruction for the R²-Agent should explicitly specify the scenario (e.g., indoor), the evaluation focus (e.g., segmentation quality), and, if applicable, any specific requirements (e.g., testing for motion blur) during the evaluation process. To enhance the robustness of the R²-Agent, we introduce a multi-agent debating strategy utilizing an operator agent " + }, + { + "bbox": [ + 130, + 300, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\varPhi" + }, + { + "bbox": [ + 130, + 300, + 482, + 422 + ], + "type": "text", + "content": " and a guardian agent " + }, + { + "bbox": [ + 130, + 300, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\varPsi" + }, + { + "bbox": [ + 130, + 300, + 482, + 422 + ], + "type": "text", + "content": ", each with their own independent memories." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 431, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 431, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 431, + 482, + 529 + ], + "type": "text", + "content": "Data proposal. To propose data samples that align with the criteria delineated in the human instruction, we employ the operator agent " + }, + { + "bbox": [ + 130, + 431, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\varPhi" + }, + { + "bbox": [ + 130, + 431, + 482, + 529 + ], + "type": "text", + "content": " with the text prompts " + }, + { + "bbox": [ + 130, + 431, + 482, + 529 + ], + "type": "inline_equation", + "content": "<\\mathrm{SEL}>" + }, + { + "bbox": [ + 130, + 431, + 482, + 529 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 431, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\varPhi" + }, + { + "bbox": [ + 130, + 431, + 482, + 529 + ], + "type": "text", + "content": " enables the traversal through the combined space of clean samples, pre-defined noise generation functions, and evaluation functions, collectively represented by the tuple " + }, + { + "bbox": [ + 130, + 431, + 482, + 529 + ], + "type": "inline_equation", + "content": "(\\mathcal{X},\\Delta,\\mathcal{E})" + }, + { + "bbox": [ + 130, + 431, + 482, + 529 + ], + "type": "text", + "content": ", to select suitable samples. Additionally, we employ an LLM to generate captions " + }, + { + "bbox": [ + 130, + 431, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 431, + 482, + 529 + ], + "type": "text", + "content": " for the clean samples " + }, + { + "bbox": [ + 130, + 431, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 130, + 431, + 482, + 529 + ], + "type": "text", + "content": ". These captions are archived within an accessible memory to enhance the computational efficiency for subsequent stages. The data proposal stage can be formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 225, + 537, + 482, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 537, + 482, + 550 + ], + "spans": [ + { + "bbox": [ + 225, + 537, + 482, + 550 + ], + "type": "interline_equation", + "content": "\\mathcal {C}, \\mathcal {X} _ {u}, \\mathcal {E} _ {u}, \\Delta_ {u} = \\Phi (u, \\mathcal {X}, \\mathcal {E}, \\Delta , < \\mathrm {S E L} >). \\tag {3}", + "image_path": "c9e7558920adff753572e120141a5105eba669dc0ce3d1437ed095be17a09986.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 557, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 483, + 666 + ], + "type": "text", + "content": "Data verification. Although LLMs exhibit notable proficiency in reasoning tasks, their susceptibility to generating \"hallucinated\" content poses a significant challenge to the delivery of accurate responses. To mitigate the hallucination effects, we introduce a verification mechanism, designated as the safe guardian " + }, + { + "bbox": [ + 130, + 557, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\varPsi" + }, + { + "bbox": [ + 130, + 557, + 483, + 666 + ], + "type": "text", + "content": ", which monitors the data proposal made by the operator " + }, + { + "bbox": [ + 130, + 557, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\varPhi" + }, + { + "bbox": [ + 130, + 557, + 483, + 666 + ], + "type": "text", + "content": " through the utilization of chain-of-thought prompting [53]. We first create the proposed dataset " + }, + { + "bbox": [ + 130, + 557, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_u^t" + }, + { + "bbox": [ + 130, + 557, + 483, + 666 + ], + "type": "text", + "content": " and perturbations " + }, + { + "bbox": [ + 130, + 557, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\Delta_u^t" + }, + { + "bbox": [ + 130, + 557, + 483, + 666 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 557, + 483, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 557, + 483, + 666 + ], + "type": "text", + "content": " denotes the iteration number. Then, we construct verification prompts " + }, + { + "bbox": [ + 130, + 557, + 483, + 666 + ], + "type": "inline_equation", + "content": "<\\mathrm{VER}>" + }, + { + "bbox": [ + 130, + 557, + 483, + 666 + ], + "type": "text", + "content": " and employ " + }, + { + "bbox": [ + 130, + 557, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\varPsi" + }, + { + "bbox": [ + 130, + 557, + 483, + 666 + ], + "type": "text", + "content": " to validate the integrity of the data instances. Feedback from this verification is relayed to " + }, + { + "bbox": [ + 130, + 557, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\varPhi" + }, + { + "bbox": [ + 130, + 557, + 483, + 666 + ], + "type": "text", + "content": ", which then uses the information" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Xiang Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 127, + 482, + 171 + ], + "blocks": [ + { + "bbox": [ + 139, + 114, + 472, + 125 + ], + "lines": [ + { + "bbox": [ + 139, + 114, + 472, + 125 + ], + "spans": [ + { + "bbox": [ + 139, + 114, + 472, + 125 + ], + "type": "text", + "content": "Table 2: Summary of considered tasks, metrics, referring modality, and datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 127, + 482, + 171 + ], + "lines": [ + { + "bbox": [ + 134, + 127, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 134, + 127, + 482, + 171 + ], + "type": "table", + "html": "
TaskRISVOSR-VOSAVSQ3M
Ref. ModalityTextImageTextAudioText
MetricsmIoUG, J, F [25]J, F [25]J, F [25]-
DatasetRefCOCO [74]/+ [74]/g [41]DAVIS [25], YTVOS [59]Ref-DAVIS [25], Ref-YTVOS [47]AVS-s4 [78], AVS-ms3 [78]ScanNet [13], ICL [19]
", + "image_path": "be345852f4ffc4dd7e0bcc33a4bc461b27672e4ca52aba5653bf6f655ba2a5cf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 179, + 482, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 179, + 482, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 179, + 482, + 214 + ], + "type": "text", + "content": "to refine and enhance the data proposal. This iterative process is carried out until a state of convergence or until a maximum iteration number is reached, which is formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 255, + 220, + 361, + 233 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 220, + 361, + 233 + ], + "spans": [ + { + "bbox": [ + 255, + 220, + 361, + 233 + ], + "type": "interline_equation", + "content": "R ^ {t} = \\Psi \\left(\\mathcal {X} _ {u} ^ {t}, \\Delta_ {u} ^ {t}, < \\mathrm {V E R} > _ {\\Psi}\\right)", + "image_path": "27049195afe17d4174b91f6f82a85dd5f6acfdc965ccc462cc454a5b61bdf5a9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 244, + 236, + 481, + 250 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 236, + 481, + 250 + ], + "spans": [ + { + "bbox": [ + 244, + 236, + 481, + 250 + ], + "type": "interline_equation", + "content": "\\mathcal {X} _ {u} ^ {t + 1}, \\Delta_ {u} ^ {t + 1} = \\Phi (R ^ {t}, < \\mathrm {V E R} > _ {\\Phi}), \\tag {4}", + "image_path": "ab8b1184dbb1fdf85760e5596ec48af4beee9ffdbf392336fe00a9a85f1130fc.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 255, + 479, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 479, + 301 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 479, + 301 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 255, + 479, + 301 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 255, + 479, + 301 + ], + "type": "text", + "content": " is the iteration index, " + }, + { + "bbox": [ + 130, + 255, + 479, + 301 + ], + "type": "inline_equation", + "content": "R^t" + }, + { + "bbox": [ + 130, + 255, + 479, + 301 + ], + "type": "text", + "content": " is the response of " + }, + { + "bbox": [ + 130, + 255, + 479, + 301 + ], + "type": "inline_equation", + "content": "\\varPsi" + }, + { + "bbox": [ + 130, + 255, + 479, + 301 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 130, + 255, + 479, + 301 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 255, + 479, + 301 + ], + "type": "text", + "content": " iteration. Note that we consider " + }, + { + "bbox": [ + 130, + 255, + 479, + 301 + ], + "type": "inline_equation", + "content": "\\varPhi" + }, + { + "bbox": [ + 130, + 255, + 479, + 301 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 255, + 479, + 301 + ], + "type": "inline_equation", + "content": "\\varPsi" + }, + { + "bbox": [ + 130, + 255, + 479, + 301 + ], + "type": "text", + "content": " have separate memories for previous inputs (similar to two separate ChatGPT sessions) and we omit the memorized inputs here for simplicity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 303, + 482, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 374 + ], + "type": "text", + "content": "Specifically, adversarial prompts are leveraged in the data verification step. In " + }, + { + "bbox": [ + 130, + 303, + 482, + 374 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{VER} \\rangle_{\\Phi}" + }, + { + "bbox": [ + 130, + 303, + 482, + 374 + ], + "type": "text", + "content": ", we ask the LLM to select conservative choices with high confidence while, for " + }, + { + "bbox": [ + 130, + 303, + 482, + 374 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{VER} \\rangle_{\\Psi}" + }, + { + "bbox": [ + 130, + 303, + 482, + 374 + ], + "type": "text", + "content": ", radical choices are acceptable. The adversarial prompts can proactively encourage a debate between two agents thus correcting potential mistakes in the original data proposal. To help understand the data verification step, we demonstrate an example of the data verification process in Fig. 5." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 383, + 482, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 383, + 482, + 455 + ], + "spans": [ + { + "bbox": [ + 131, + 383, + 482, + 455 + ], + "type": "text", + "content": "Model analysis. With the selected samples and functions " + }, + { + "bbox": [ + 131, + 383, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_u, \\mathcal{E}_u, \\Delta_u" + }, + { + "bbox": [ + 131, + 383, + 482, + 455 + ], + "type": "text", + "content": ", we instantiate the noisy data " + }, + { + "bbox": [ + 131, + 383, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_u'" + }, + { + "bbox": [ + 131, + 383, + 482, + 455 + ], + "type": "text", + "content": " by applying perturbation functions to the clean data, " + }, + { + "bbox": [ + 131, + 383, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_u' = \\Delta_u(\\mathcal{X}_u)" + }, + { + "bbox": [ + 131, + 383, + 482, + 455 + ], + "type": "text", + "content": ". We then calculate a set of metrics " + }, + { + "bbox": [ + 131, + 383, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_u = \\{m_i\\}" + }, + { + "bbox": [ + 131, + 383, + 482, + 455 + ], + "type": "text", + "content": " that correspond to " + }, + { + "bbox": [ + 131, + 383, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_u'" + }, + { + "bbox": [ + 131, + 383, + 482, + 455 + ], + "type": "text", + "content": " by applying evaluating functions " + }, + { + "bbox": [ + 131, + 383, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_u" + }, + { + "bbox": [ + 131, + 383, + 482, + 455 + ], + "type": "text", + "content": " to noisy samples. Utilizing the chain-of-though prompt , we engage an LLM " + }, + { + "bbox": [ + 131, + 383, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 131, + 383, + 482, + 455 + ], + "type": "text", + "content": " with empty memory to analyze the metrics and produce an output report " + }, + { + "bbox": [ + 131, + 383, + 482, + 455 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 131, + 383, + 482, + 455 + ], + "type": "text", + "content": " for the model performance as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 255, + 460, + 481, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 460, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 255, + 460, + 481, + 472 + ], + "type": "interline_equation", + "content": "O = \\Theta \\left(\\mathcal {M} _ {u}, \\mathcal {X} _ {u} ^ {\\prime}, < \\mathrm {A N A} >\\right). \\tag {5}", + "image_path": "fb61a7c415eb1588ea27f6a55b40b3aa1434418effbea9430f2e3ac7edff1946.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 487, + 224, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 224, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 224, + 500 + ], + "type": "text", + "content": "5 Experiment" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 509, + 312, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 509, + 312, + 521 + ], + "spans": [ + { + "bbox": [ + 131, + 509, + 312, + 521 + ], + "type": "text", + "content": "5.1 Evaluation Setup and Metrics" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 525, + 482, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 482, + 562 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 482, + 562 + ], + "type": "text", + "content": "In Tab. 2, we present a summary of the benchmark tasks, evaluated datasets, and corresponding metrics. Our benchmark encompasses five prevalent tasks, namely, RIS, VOS, R-VOS, AVS and Q3M." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "Evaluation setup. We conduct systematic evaluations to analyze the impact of perturbations on state-of-the-art mode five tasks. Both perturbations in the referring guidance and visual frames are considered in the benchmarking. We create noisy datasets with low, medium and high noise levels for each task. For video-level tasks, we additionally consider the scenario that noises change over time. We first benchmark the models' general performance with all types of noises and then investigate the impact of noises individually. More information about the dataset creation is available in the Appendix." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 132, + 91, + 470, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 91, + 470, + 102 + ], + "spans": [ + { + "bbox": [ + 132, + 91, + 470, + 102 + ], + "type": "text", + "content": "Benchmarking the Robustness of Referring Perception Models under Perturbations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 496, + 91, + 503, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 496, + 91, + 503, + 101 + ], + "spans": [ + { + "bbox": [ + 496, + 91, + 503, + 101 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 147, + 484, + 203 + ], + "blocks": [ + { + "bbox": [ + 130, + 111, + 482, + 145 + ], + "lines": [ + { + "bbox": [ + 130, + 111, + 482, + 145 + ], + "spans": [ + { + "bbox": [ + 130, + 111, + 482, + 145 + ], + "type": "text", + "content": "Table 3: Performance of referring image segmentation methods under low, medium, and high perturbation levels. Average performance change (APC) is averaged among all perturbation levels. We report mIoU following convention." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 147, + 484, + 203 + ], + "lines": [ + { + "bbox": [ + 134, + 147, + 484, + 203 + ], + "spans": [ + { + "bbox": [ + 134, + 147, + 484, + 203 + ], + "type": "table", + "html": "
MethodVenueRefCOCORefCOCO+RefCOCOg
CleanLowMed.HighAPCCleanLowMed.HighAPCCleanLowMed.HighAPC
LAVT [69]CVPR 220.740.580.530.48-0.210.660.580.510.43-0.150.630.550.470.38-0.16
PolyFormer [37]CVPR 230.760.700.640.57-0.120.710.650.570.50-0.140.690.640.570.50-0.12
X-Decoder [82]CVPR 230.680.600.520.44-0.160.690.600.510.44-0.170.640.580.500.44-0.14
ETRIS [63]ICCV 230.710.640.570.49-0.140.600.560.480.40-0.120.600.530.450.38-0.15
SEEM [84]NeurIPS 230.650.590.540.47-0.120.640.580.510.44-0.130.620.560.490.43-0.12
", + "image_path": "9b7b17d60674959b454a72fc8fd02b132fb4d6f95809e0c74dded013e6a0c45a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 229, + 482, + 293 + ], + "blocks": [ + { + "bbox": [ + 130, + 206, + 481, + 228 + ], + "lines": [ + { + "bbox": [ + 130, + 206, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 130, + 206, + 481, + 228 + ], + "type": "text", + "content": "Table 4: Performance of video object segmentation methods under low, medium, and high perturbation levels. APC is averaged among all perturbation levels." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 229, + 482, + 293 + ], + "lines": [ + { + "bbox": [ + 135, + 229, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 135, + 229, + 482, + 293 + ], + "type": "table", + "html": "
MethodVenueDAVISYTVOS
CleanLowMediumHighAPCCleanLowMediumHighAPC
JFJFJFJJGsGuGsGuGsGu
AOT [70]NeurIPS 2176.083.071.779.269.776.965.371.1-7.1-7.386.484.284.980.982.078.176.170.4-5.4-7.7
DeAOT [71]NeurIPS 2276.984.573.281.072.480.468.275.5-5.6-6.486.784.586.683.784.581.880.876.8-2.7-3.7
XMem [11]ECCV 2277.484.572.280.071.879.768.474.7-6.6-6.486.584.584.081.281.879.078.674.9-5.1-6.1
DEVA [10]ICCV 2378.785.976.484.974.282.469.977.5-5.2-4.387.283.986.782.884.780.080.676.3-3.2-4.2
Cutie [9]CVPR 2480.687.775.883.575.082.769.375.4-7.2-7.287.884.586.582.884.280.580.375.1-4.1-5.0
", + "image_path": "0212cd1f808b82c5e280cef7eea7f7cec5c9d3ea81c84af4e6d2c7c33346bff3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 135, + 319, + 484, + 392 + ], + "blocks": [ + { + "bbox": [ + 130, + 296, + 482, + 318 + ], + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 318 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 318 + ], + "type": "text", + "content": "Table 5: Performance of referring video object segmentation methods under low, medium, and high perturbation levels. APC is averaged among all perturbation levels." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 319, + 484, + 392 + ], + "lines": [ + { + "bbox": [ + 135, + 319, + 484, + 392 + ], + "spans": [ + { + "bbox": [ + 135, + 319, + 484, + 392 + ], + "type": "table", + "html": "
MethodVenueRef-DAVISRef-YTVOS
CleanLowMediumHighAPCCleanLowMediumHighAPC
J FJ FJ FJ FJ FJ FJ FJ FJ FJ FJ FJ F
Video-level Backbone
MTTR [5]CVPR 22------54.056.448.951.643.045.5
SgMg [42]ICCV 2359.064.857.362.651.957.7-6.8-7.757.760.052.454.5
Image-level Backbone
ReferFormer [55]CVPR 2255.861.346.550.143.148.038.541.9-13.1-14.654.856.5
OnlineRefer [54]ICCV 2355.762.951.657.948.054.342.246.3-8.4-10.055.658.9
R2-VOS [33]ICCV 2357.262.450.257.046.553.839.244.6-11.9-10.656.158.4
", + "image_path": "717cfb7530244a8033d07fc898fb4dd1cd074eef8b790fb1cfb2ab387712bdf6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 407, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 506 + ], + "type": "text", + "content": "Metrics. Following the convention [41, 74], we leverage mIoU to evaluate the RIS task. For video-level tasks, the convention is to compute region similarity " + }, + { + "bbox": [ + 130, + 407, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 130, + 407, + 482, + 506 + ], + "type": "text", + "content": " and contour accuracy " + }, + { + "bbox": [ + 130, + 407, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 407, + 482, + 506 + ], + "type": "text", + "content": " as defined in [45]. Specifically, " + }, + { + "bbox": [ + 130, + 407, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = \\frac{\\mathcal{I} + \\mathcal{F}}{2}" + }, + { + "bbox": [ + 130, + 407, + 482, + 506 + ], + "type": "text", + "content": " is also used. In addition, to better demonstrate the performance degradation against perturbations, we define average performance change (APC) as " + }, + { + "bbox": [ + 130, + 407, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\mathrm{APC} = \\sum_{i} \\frac{m_i^n - m_i^c}{N}" + }, + { + "bbox": [ + 130, + 407, + 482, + 506 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 407, + 482, + 506 + ], + "type": "inline_equation", + "content": "m_i^n" + }, + { + "bbox": [ + 130, + 407, + 482, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 407, + 482, + 506 + ], + "type": "inline_equation", + "content": "m_i^c" + }, + { + "bbox": [ + 130, + 407, + 482, + 506 + ], + "type": "text", + "content": " denotes the performance of the " + }, + { + "bbox": [ + 130, + 407, + 482, + 506 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 407, + 482, + 506 + ], + "type": "text", + "content": "-th sample of perturbed and clean data respectively. " + }, + { + "bbox": [ + 130, + 407, + 482, + 506 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 407, + 482, + 506 + ], + "type": "text", + "content": " is the sample number. Additional results, evaluated using various metrics and settings, can be found in the Appendix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 524, + 302, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 524, + 302, + 537 + ], + "spans": [ + { + "bbox": [ + 130, + 524, + 302, + 537 + ], + "type": "text", + "content": "5.2 Performance Benchmarking" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Referring image segmentation. As shown in Tab. 3, we evaluate the state-of-the-art RIS methods on perturbed RefCOCO " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "/ + / \\mathrm{g}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " [74] datasets. We notice that PolyFormer [37] achieves the best performance in terms of both performance and robustness. Different from other methods that generate pixel-level classification results to represent an object mask, PolyFormer utilizes multi-polygon vertices to represent the mask. We consider the robustness can result from the special sequential prediction and polygon representation which eases the reliance on the per-pixel representation. In addition, we notice that though SEEM [83] shows an inferior performance, its robustness to perturbations is promising. We consider that multi-task joint training can account for SEEM's robustness." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Xiang Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 138, + 482, + 191 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 6: Performance of audiovisual segmentation methods under low, medium, and high perturbation levels. APC is averaged among all perturbation levels." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 138, + 482, + 191 + ], + "lines": [ + { + "bbox": [ + 134, + 138, + 482, + 191 + ], + "spans": [ + { + "bbox": [ + 134, + 138, + 482, + 191 + ], + "type": "table", + "html": "
MethodVenueAVS-s4AVS-ms3
CleanLowMediumHighAPCCleanLowMediumHighAPC
J FJ FJ FJ FJ FJ FJ FJ FJ FJ FJ FJ F
AVS [79]ECCV 2272.8 84.868.4 79.760.8 71.255.6 66.8-11.2 -12.247.9 57.844.4 54.641.2 50.537.0 44.2-7.0 -8.0
CATR [27]MM 2374.9 87.171.1 84.265.7 79.058.4 72.4-9.8 -8.653.1 65.649.8 61.446.7 58.441.2 52.9-7.2 -8.0
AVSegFormer [17]AAAI 2376.4 86.770.9 82.362.4 74.857.3 69.8-12.9 -11.049.5 62.846.9 59.043.7 55.439.8 50.1-6.0 -8.0
QSD [35]CVPR 2477.6 85.674.8 83.668.3 78.459.4 71.2-10.1 -7.961.8 64.356.6 60.053.1 57.147.6 51.6-9.4 -8.0
", + "image_path": "096939eb02b38372271602d393300e1d457c327ee2faae0e79c3ee0fef8843dc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 137, + 194, + 480, + 318 + ], + "blocks": [ + { + "bbox": [ + 137, + 194, + 480, + 318 + ], + "lines": [ + { + "bbox": [ + 137, + 194, + 480, + 318 + ], + "spans": [ + { + "bbox": [ + 137, + 194, + 480, + 318 + ], + "type": "image", + "image_path": "9cc248f997b73f32dd2a24b536b0b936ec18bc1c2026315e597823a55572b226.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 324, + 480, + 358 + ], + "lines": [ + { + "bbox": [ + 130, + 324, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 480, + 358 + ], + "type": "text", + "content": "Fig. 6: Queryable 3D mapping of ConceptFusion [23] with inaccurate referring guidance on ScanNet [13] and ICL [19] datasets. Due to the absence of publicly available evaluation codes, we did not provide quantitative evaluation for queryable 3D mapping." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 371, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 371, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 482, + 491 + ], + "type": "text", + "content": "Video object segmentation. As depicted in Tab. 4, we demonstrate the performance of state-of-the-art methods on DAVIS [25] and YTVOS [59] datasets. We notice that, for simple scenarios in DAVIS datasets, Cutie [9] shows the best performance. While in the more complex YTVOS dataset, DeAOT [71] achieves the best. Cutie heavily relies on the pixel- and object-level correspondence across frames which can be disrupted by the perturbations. Differently, DeAOT leverages decoupled visual and ID embeddings which may be the reason for its robustness in noisy scenarios. In addition, we notice that the unseen categories generally have a larger performance drop compared to the seen categories which can impose more challenges in practical deployment in complex scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 500, + 482, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 500, + 482, + 573 + ], + "spans": [ + { + "bbox": [ + 130, + 500, + 482, + 573 + ], + "type": "text", + "content": "Referring video object segmentation. We benchmark the referring video object segmentation task in Tab. 5. SgMg [42] and OnlineFormer [54] achieve the best performance and robustness among methods equipped with video- and image-level backbones respectively. Unlike other methods, OnlineRefer processes the visual features in a frame-by-frame manner which makes it rely less on the pixel-level temporal correspondence that is easily been disrupted by visual perturbations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Audio-visual segmentation. We show the performance of popular AVS methods in Tab. 6. We notice that QSD [34] and CATR [27] archive promising performance and robustness among all methods. CATR [27] and AVSegFormer [17] are similar methods that leverage acoustic query to query the visual frames with a transformer-based structure. QSD moves one step forward which additionally introduces a quantization operation and a local-global distillation to enhance the robustness of the acoustic representation to adapt to complex scenarios." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 132, + 91, + 470, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 91, + 470, + 102 + ], + "spans": [ + { + "bbox": [ + 132, + 91, + 470, + 102 + ], + "type": "text", + "content": "Benchmarking the Robustness of Referring Perception Models under Perturbations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 492, + 92, + 502, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 492, + 92, + 502, + 101 + ], + "spans": [ + { + "bbox": [ + 492, + 92, + 502, + 101 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 117, + 482, + 196 + ], + "blocks": [ + { + "bbox": [ + 134, + 117, + 482, + 196 + ], + "lines": [ + { + "bbox": [ + 134, + 117, + 482, + 196 + ], + "spans": [ + { + "bbox": [ + 134, + 117, + 482, + 196 + ], + "type": "image", + "image_path": "c237184825d1083df4354938ef2e717843c0572d78365cb3846347f120f9a9ba.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 198, + 473, + 210 + ], + "lines": [ + { + "bbox": [ + 136, + 198, + 473, + 210 + ], + "spans": [ + { + "bbox": [ + 136, + 198, + 473, + 210 + ], + "type": "text", + "content": "Fig. 7: Average performance changes (APC) for each perturbation type." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "text", + "content": "Queryable 3D mapping. Recent queryable 3D mapping methods [23,64] typically leverage the results from 2D referring models and fuse them to a 3D map. Thereby, the failure of 2D referring perception can directly obstacle the success of 3D tasks. As shown in Fig. 6, we visualize the state-of-the-art queryable 3D mapping method, ConceptFusion [23]. We notice that with the accurate text query \"orange sofa\" ConceptFusion can successfully locate the object while when misspell occur in the query, an obvious performance degradation can be observed." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 324, + 272, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 324, + 272, + 337 + ], + "spans": [ + { + "bbox": [ + 132, + 324, + 272, + 337 + ], + "type": "text", + "content": "5.3 Perturbation Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 343, + 483, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 483, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 483, + 559 + ], + "type": "text", + "content": "Perturbation type. We conduct ablation studies on performance degradation against different perturbation types. Fig. 7 illustrates the aggregated average performance deviation among all benchmarked methods under visual, textual, and acoustic perturbations, respectively. For visual perturbations, we notice the frost effect (FT) induced the most substantial performance degradation, which is likely attributable to its extensive occlusion and distortion effects in the whole image. Conversely, color-based perturbations, such as brightness (BR) and saturate (SA), exhibit marginal impact on performance, possibly due to the preservation of shape information. For textual perturbations, our analysis indicated that all types of perturbations lead to comparable levels of degradation. The character missing (CA) has a slightly greater impact. For acoustic perturbations, we notice that most perturbation types only show a marginal impact on the performance. We consider this is because the target objects in AVS-s4/ms3 are salient objects, allowing models to rely predominantly on visual cues for object localization, even when the acoustic guidance lacks precision. Notably, the impulse response perturbation was associated with a significant performance drop, which could stem from the failure in feature extraction due to the impulse signal. Such abnormal acoustic features could potentially disrupt subsequent multimodal interactions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 331, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 331, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 331, + 665 + ], + "type": "text", + "content": "Perturbation correlation. To delve deeper into the nature of performance degradation, we computed correlation matrix for each modality, utilizing the Pearson product-moment correlation coefficients calculated from the performance under various perturbation types to populate the matrix. Fig. 8 displays the correlation matrix for textual perturbations." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 342, + 577, + 473, + 638 + ], + "blocks": [ + { + "bbox": [ + 342, + 577, + 473, + 638 + ], + "lines": [ + { + "bbox": [ + 342, + 577, + 473, + 638 + ], + "spans": [ + { + "bbox": [ + 342, + 577, + 473, + 638 + ], + "type": "image", + "image_path": "60a3347f9f1ab8bf5ca7746429f2e861b828a20e2c72d2c53084317b5cb0387e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 640, + 482, + 662 + ], + "lines": [ + { + "bbox": [ + 335, + 640, + 482, + 662 + ], + "spans": [ + { + "bbox": [ + 335, + 640, + 482, + 662 + ], + "type": "text", + "content": "Fig.8: Correlation of performance of textual perturbations." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Xiang Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 305, + 205 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 305, + 205 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 305, + 205 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 305, + 205 + ], + "type": "image", + "image_path": "e1cd3cc6389dabacc940cfc734ab581c38a6bbb23004d68fd6876f9283982f06.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 207, + 482, + 240 + ], + "lines": [ + { + "bbox": [ + 130, + 207, + 482, + 240 + ], + "spans": [ + { + "bbox": [ + 130, + 207, + 482, + 240 + ], + "type": "text", + "content": "Fig. 9: Correlation matrix of degraded performance with visual and acoustic perturbations. We concatenate the sample-level performance of all methods on all testing datasets as a feature vector to calculate Pearson product-moment correlation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 116, + 479, + 205 + ], + "blocks": [ + { + "bbox": [ + 309, + 116, + 479, + 205 + ], + "lines": [ + { + "bbox": [ + 309, + 116, + 479, + 205 + ], + "spans": [ + { + "bbox": [ + 309, + 116, + 479, + 205 + ], + "type": "image", + "image_path": "cddd06cdc3252f8c855438348b3ba94245967bf0df35307f042789f907d77939.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 255, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 482, + 399 + ], + "type": "text", + "content": "Despite the average performance changes (in Fig. 7) being similar across different types, the low correlation between the types of performance degradation suggest that each perturbation affects the model in a distinct manner. The correlations for visual perturbations and the ones for acoustic perturbations are presented in Fig. 9 respectively. For visual perturbations, we notice several instances of high correlations. For instance, the group of blurring perturbations—defocus blur (DB), gaussian blur (GB), motion blur (MB), and glass blur (GS)—are highly correlated with one another. For the acoustic perturbations, impulse response (IR) and highpass filter (HF) exhibit different patterns compared to other perturbations, indicating unique effects on performance degradation. We systematically explored the correlations among perturbations and hope the analysis can provide insights for wider evaluation in the future." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 411, + 317, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 317, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 317, + 506 + ], + "type": "text", + "content": "Static v.s. Dynamic. To analyze the dynamic changing of perturbation types in videos, we conduct an ablation study to explore the impact of perturbation changing across frames (for visual perturbation only). As shown in Fig. 10, we calculate the APC across all models on the Ref-DAVIS dataset. We notice the model performance" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 325, + 413, + 474, + 472 + ], + "blocks": [ + { + "bbox": [ + 325, + 413, + 474, + 472 + ], + "lines": [ + { + "bbox": [ + 325, + 413, + 474, + 472 + ], + "spans": [ + { + "bbox": [ + 325, + 413, + 474, + 472 + ], + "type": "image", + "image_path": "b8aed8f2de4d606a0706778fc314252ef1b3bce727c5c9b9444d46df79dd6d7b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 473, + 482, + 506 + ], + "lines": [ + { + "bbox": [ + 321, + 473, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 321, + 473, + 482, + 506 + ], + "type": "text", + "content": "Fig.10: Average performance change (APC) under static and dynamic perturbations." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 507, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 482, + 532 + ], + "type": "text", + "content": "decreases more with dynamic perturbations which can be due to the lack of temporal consistency resulting from various perturbation types." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 548, + 319, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 548, + 319, + 562 + ], + "spans": [ + { + "bbox": [ + 131, + 548, + 319, + 562 + ], + "type": "text", + "content": "5.4 " + }, + { + "bbox": [ + 131, + 548, + 319, + 562 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 131, + 548, + 319, + 562 + ], + "type": "text", + "content": "-Agent: Evaluation Assistant" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 570, + 298, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 570, + 298, + 640 + ], + "spans": [ + { + "bbox": [ + 131, + 570, + 298, + 640 + ], + "type": "text", + "content": "User study of data selection. To evaluate the data selection performance of " + }, + { + "bbox": [ + 131, + 570, + 298, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 131, + 570, + 298, + 640 + ], + "type": "text", + "content": "-Agent, we manually annotate an evaluation set. Since the instruction can be ambiguous in practice, we measure the alignment between " + }, + { + "bbox": [ + 131, + 570, + 298, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 305, + 585, + 482, + 633 + ], + "blocks": [ + { + "bbox": [ + 312, + 574, + 468, + 584 + ], + "lines": [ + { + "bbox": [ + 312, + 574, + 468, + 584 + ], + "spans": [ + { + "bbox": [ + 312, + 574, + 468, + 584 + ], + "type": "text", + "content": "Table 7: Data selection performance." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 305, + 585, + 482, + 633 + ], + "lines": [ + { + "bbox": [ + 305, + 585, + 482, + 633 + ], + "spans": [ + { + "bbox": [ + 305, + 585, + 482, + 633 + ], + "type": "table", + "html": "
# Ver. Iter.X123
ScoreACC RateACC RateACC RateACC Rate
Data Sample0.530.780.640.810.770.850.800.91
Perturbation0.650.870.770.860.860.910.870.93
Metric Func.1.001.001.001.001.001.001.001.00
", + "image_path": "3365ad42d51f4858f8e7fe9a93a019465a8c5cb1c26b5ba47ff3f0ddf9075b8f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Agent and human by (1) ACC: accuracy between the human-annotated samples and " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "-Agent predictions and (2) Rate: rating the rationality of the prediction" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 132, + 91, + 470, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 91, + 470, + 102 + ], + "spans": [ + { + "bbox": [ + 132, + 91, + 470, + 102 + ], + "type": "text", + "content": "Benchmarking the Robustness of Referring Perception Models under Perturbations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 492, + 91, + 503, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 492, + 91, + 503, + 101 + ], + "spans": [ + { + "bbox": [ + 492, + 91, + 503, + 101 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 115, + 217, + 198 + ], + "blocks": [ + { + "bbox": [ + 138, + 115, + 217, + 198 + ], + "lines": [ + { + "bbox": [ + 138, + 115, + 217, + 198 + ], + "spans": [ + { + "bbox": [ + 138, + 115, + 217, + 198 + ], + "type": "image", + "image_path": "471a469cc966a399caa11d51911ca264628e3f0602574063cd532553e5943044.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 205, + 482, + 239 + ], + "lines": [ + { + "bbox": [ + 132, + 205, + 482, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 482, + 239 + ], + "type": "text", + "content": "Fig. 11: Example of " + }, + { + "bbox": [ + 132, + 205, + 482, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 132, + 205, + 482, + 239 + ], + "type": "text", + "content": "-Agent report. The evaluation setting and metrics are fed to " + }, + { + "bbox": [ + 132, + 205, + 482, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 132, + 205, + 482, + 239 + ], + "type": "text", + "content": "-Agent with a set of hard-coded prompts to facilitate " + }, + { + "bbox": [ + 132, + 205, + 482, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 132, + 205, + 482, + 239 + ], + "type": "text", + "content": "-Agent generate the human-like report. Beyond the hard-coded analysis, further QA is also supported." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 221, + 118, + 478, + 199 + ], + "blocks": [ + { + "bbox": [ + 221, + 118, + 478, + 199 + ], + "lines": [ + { + "bbox": [ + 221, + 118, + 478, + 199 + ], + "spans": [ + { + "bbox": [ + 221, + 118, + 478, + 199 + ], + "type": "table", + "html": "
Scenario:Scenario: Car driving scenario with a fast motion\nPerturbation: snow (SN), fog (SG), frost (FT), spatter (SP), motion blur (MB),...
Evaluation Metric: mIoU (mean Intersection over Union)
Analysis:IoU Change: The spatter has a largest -12.1 change in mIoU suggesting that the high level of spatter effect has negatively impacted the model's ability. The spatter effect can lead to 1. Boundary Ambiguity...2. Fragmentation...3. Texture change...ForJPEG compression and animation have a marginal effect...
Case Analysis: The case with caption: “a red car is far away on the left” suffers severe performance drop under...The small size of the car may account for the...
Recommendation: Improve the resilience of models to compression artifacts through enhanced training, including a wider variety of...
", + "image_path": "bbced268dc80b1462669191b6abd5884fb9527b15c09dabad45b08796acb36e4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 258, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 344 + ], + "type": "text", + "content": "from " + }, + { + "bbox": [ + 130, + 258, + 482, + 344 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 130, + 258, + 482, + 344 + ], + "type": "text", + "content": "-Agent (1 if reasonable else 0). As shown in Tab. 7, we ablate on the data verification iteration number. For the metrics function selection, since there are only several hard-coded metrics that can be selected, the accuracy remains perfect while, for data sample and perturbation, we notice a direct performance gain when increasing the iteration number. We notice the performance becomes saturated after 2 iterations, thus we leverage iteration number 2 as the final design choice. More details can be found in the Appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 352, + 483, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 483, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 483, + 472 + ], + "type": "text", + "content": "Automatic model analysis. With the evaluation setting and corresponding results, " + }, + { + "bbox": [ + 130, + 352, + 483, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 130, + 352, + 483, + 472 + ], + "type": "text", + "content": "-Agent can generate a detailed evaluation report. We demonstrate an evaluation report example in Fig. 11. To make " + }, + { + "bbox": [ + 130, + 352, + 483, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 130, + 352, + 483, + 472 + ], + "type": "text", + "content": "-Agent think in a human-like manner, we leverage a set of chain-of-thought prompts to instruct " + }, + { + "bbox": [ + 130, + 352, + 483, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 130, + 352, + 483, + 472 + ], + "type": "text", + "content": "-Agent to give reasonable and detailed analysis based on the given evaluation results. We notice that " + }, + { + "bbox": [ + 130, + 352, + 483, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 130, + 352, + 483, + 472 + ], + "type": "text", + "content": "-Agent can understand the evaluation metrics and give a human-like analysis which potentially helps to reduce the cost of model evaluation in practice. Furthermore, given the strong dialogue capability of LLMs, further question-answering about the results and report is also feasible. We demonstrate more qualitative results in the Appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 487, + 220, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 220, + 499 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 220, + 499 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": "In this work, we introduced " + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": "-Bench, a comprehensive benchmark for evaluating the robustness of visual referring perception models against various perturbations. Our contributions include a detailed taxonomy of perturbations, a customizable perturbation synthesis toolbox, systematic perturbation analysis, and " + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": "-Agent, a novel evaluation assistant based on large language models. Through extensive experiments, we observed and analyzed the intrinsic characteristics of different perturbations to current models and highlighted the importance of robustness in referring perception tasks. A benchmark with 5 popular referring tasks is also provided to facilitate future research. In addition, the data construction and analysis of robustness evaluation can be further simplified with our proposed " + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": "-Agent. Our findings underscore the necessity for robustness in the deployment of intelligent systems in real-world scenarios, offering a foundation for future advancements in the field." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Xiang Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 141, + 481, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 481, + 176 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 481, + 176 + ], + "type": "text", + "content": "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023) 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 176, + 481, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 176, + 481, + 220 + ], + "spans": [ + { + "bbox": [ + 138, + 176, + 481, + 220 + ], + "type": "text", + "content": "2. Ahn, H., Kwon, O., Kim, K., Jeong, J., Jun, H., Lee, H., Lee, D., Oh, S.: Visually grounding language instruction for history-dependent manipulation. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 675-682. IEEE (2022) 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 220, + 481, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 220, + 481, + 242 + ], + "spans": [ + { + "bbox": [ + 138, + 220, + 481, + 242 + ], + "type": "text", + "content": "3. Arandjelovic, R., Zisserman, A.: Look, listen and learn. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 609-617 (2017) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 243, + 481, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 243, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 243, + 481, + 266 + ], + "type": "text", + "content": "4. Arandjelovic, R., Zisserman, A.: Objects that sound. In: Proceedings of the European conference on computer vision (ECCV). pp. 435-451 (2018) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 266, + 481, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 481, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 481, + 298 + ], + "type": "text", + "content": "5. Botach, A., Zheltonozhskii, E., Baskin, C.: End-to-end referring video object segmentation with multimodal transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4985-4995 (2022) 4, 10" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 299, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 481, + 342 + ], + "type": "text", + "content": "6. Chen, F., Zhang, H., Hu, K., Huang, Y.k., Zhu, C., Savvides, M.: Enhanced training of query-based object detection via selective query recollection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23756-23765 (2023) 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 343, + 481, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 481, + 375 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 481, + 375 + ], + "type": "text", + "content": "7. Chen, F., Zhang, H., Li, Z., Dou, J., Mo, S., Chen, H., Zhang, Y., Ahmed, U., Zhu, C., Savvides, M.: Unitail: Detecting, reading, and matching in retail scene (2022), https://arxiv.org/abs/2204.00298 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 376, + 481, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 481, + 408 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 481, + 408 + ], + "type": "text", + "content": "8. Chen, F., Zhang, H., Yang, Z., Chen, H., Hu, K., Savvides, M.: Rtgen: Generating region-text pairs for open-vocabulary object detection (2024), https://arxiv.org/abs/2405.19854 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 409, + 481, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 409, + 481, + 432 + ], + "spans": [ + { + "bbox": [ + 138, + 409, + 481, + 432 + ], + "type": "text", + "content": "9. Cheng, H.K., Oh, S.W., Price, B., Lee, J.Y., Schwing, A.: Putting the object back into video object segmentation. arXiv preprint arXiv:2310.12982 (2023) 4, 10, 11" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 432, + 481, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 432, + 481, + 454 + ], + "spans": [ + { + "bbox": [ + 138, + 432, + 481, + 454 + ], + "type": "text", + "content": "0. Cheng, H.K., Oh, S.W., Price, B., Schwing, A., Lee, J.Y.: Tracking anything with decoupled video segmentation. In: ICCV (2023) 10" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 455, + 481, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 455, + 481, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 455, + 481, + 487 + ], + "type": "text", + "content": "1. Cheng, H.K., Schwing, A.G.: Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In: European Conference on Computer Vision. pp. 640-658. Springer (2022) 4, 10" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 488, + 481, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 488, + 481, + 532 + ], + "spans": [ + { + "bbox": [ + 138, + 488, + 481, + 532 + ], + "type": "text", + "content": "2. Cheng, Y., Wang, R., Pan, Z., Feng, R., Zhang, Y.: Look, listen, and attend: Co-attention network for self-supervised audio-visual representation learning. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 3884-3892 (2020) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 533, + 481, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 481, + 565 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 481, + 565 + ], + "type": "text", + "content": "3. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017) 9, 11" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 566, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 566, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 566, + 481, + 599 + ], + "type": "text", + "content": "4. Ding, H., Liu, C., He, S., Jiang, X., Loy, C.C.: Mevis: A large-scale benchmark for video segmentation with motion expressions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2694-2703 (2023) 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "type": "text", + "content": "5. Ding, H., Liu, C., Wang, S., Jiang, X.: Vision-language transformer and query generation for referring segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16321-16330 (2021) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "6. Du, Y., Li, S., Torralba, A., Tenenbaum, J.B., Mordatch, I.: Improving factuality and reasoning in language models through multiagent debate. arXiv preprint arXiv:2305.14325 (2023) 3" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 132, + 91, + 470, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 91, + 470, + 102 + ], + "spans": [ + { + "bbox": [ + 132, + 91, + 470, + 102 + ], + "type": "text", + "content": "Benchmarking the Robustness of Referring Perception Models under Perturbations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 492, + 91, + 502, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 492, + 91, + 502, + 100 + ], + "spans": [ + { + "bbox": [ + 492, + 91, + 502, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 138 + ], + "type": "text", + "content": "17. Gao, S., Chen, Z., Chen, G., Wang, W., Lu, T.: Avsegformer: Audio-visual segmentation with transformer. arXiv preprint arXiv:2307.01146 (2023) 2, 11" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 482, + 183 + ], + "type": "text", + "content": "18. Han, M., Wang, Y., Li, Z., Yao, L., Chang, X., Qiao, Y.: Hybrid temporal-scale multimodal learning framework for referring video object segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13414-13423 (2023) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "type": "text", + "content": "19. Handa, A., Whelan, T., McDonald, J., Davison, A.: A benchmark for RGB-D visual odometry, 3D reconstruction and SLAM. In: IEEE Intl. Conf. on Robotics and Automation, ICRA. Hong Kong, China (May 2014) 9, 11" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "type": "text", + "content": "20. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261 (2019) 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 481, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 481, + 273 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 481, + 273 + ], + "type": "text", + "content": "21. Hu, Y., Lin, F., Zhang, T., Yi, L., Gao, Y.: Look before you leap: Unveiling the power of gpt-4v in robotic vision-language planning. arXiv preprint arXiv:2311.17842 (2023) 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 481, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 307 + ], + "type": "text", + "content": "22. Huang, W., Xia, F., Shah, D., Driess, D., Zeng, A., Lu, Y., Florence, P., Mordatch, I., Levine, S., Hausman, K., et al.: Grounded decoding: Guiding text generation with grounded models for robot control. arXiv preprint arXiv:2303.00855 (2023) 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 308, + 481, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 340 + ], + "type": "text", + "content": "23. Jatavallabhula, K.M., Kuwajerwala, A., Gu, Q., Omama, M., Chen, T., Maalouf, A., Li, S., Iyer, G., Saryazdi, S., Keetha, N., et al.: Conceptfusion: Open-set multimodal 3d mapping. arXiv preprint arXiv:2302.07241 (2023) 2, 4, 11, 12" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 341, + 481, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 481, + 362 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 481, + 362 + ], + "type": "text", + "content": "24. Ke, L., Ye, M., Danelljan, M., Liu, Y., Tai, Y.W., Tang, C.K., Yu, F.: Segment anything in high quality. In: NeurIPS (2023) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 363, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 481, + 385 + ], + "type": "text", + "content": "25. Khoreva, A., Rohrbach, A., Schiele, B.: Video object segmentation with language referring expressions. In: ACCV (2018) 9, 11" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "text", + "content": "26. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 419, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 481, + 453 + ], + "type": "text", + "content": "27. Li, K., Yang, Z., Chen, L., Yang, Y., Xun, J.: Catr: Combinatorial-dependence audio-queried transformer for audio-visual video segmentation. arXiv preprint arXiv:2309.09709 (2023) 4, 11" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 453, + 481, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 481, + 475 + ], + "type": "text", + "content": "28. Li, M., Li, S., Zhang, X., Zhang, L.: Univs: Unified and universal video segmentation with prompts as queries. arXiv preprint arXiv:2402.18115 (2024) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 475, + 481, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 481, + 508 + ], + "type": "text", + "content": "29. Li, X., Cao, H., Zhao, S., Li, J., Zhang, L., Raj, B.: Panoramic video salient object detection with ambisonic audio guidance. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1424-1432 (2023) 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 509, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 481, + 542 + ], + "type": "text", + "content": "30. Li, X., Lin, C.C., Chen, Y., Liu, Z., Wang, J., Singh, R., Raj, B.: Paintseg: Painting pixels for training-free segmentation. In: Thirty-seventh Conference on Neural Information Processing Systems (2023) 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 543, + 481, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 481, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 481, + 575 + ], + "type": "text", + "content": "31. Li, X., Wang, J., Li, X., Lu, Y.: Hybrid instance-aware temporal fusion for online video instance segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 1429-1437 (2022) 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 576, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 481, + 597 + ], + "type": "text", + "content": "32. Li, X., Wang, J., Li, X., Lu, Y.: Video instance segmentation by instance flow assembly. IEEE Transactions on Multimedia 25, 7469-7479 (2022) 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 598, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 481, + 632 + ], + "type": "text", + "content": "33. Li, X., Wang, J., Xu, X., Li, X., Raj, B., Lu, Y.: Robust referring video object segmentation with cyclic structural consensus. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22236-22245 (2023) 2, 10" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "34. Li, X., Wang, J., Xu, X., Peng, X., Singh, R., Lu, Y., Raj, B.: Towards robust audiovisual segmentation in complex environments with quantization-based semantic decomposition. arXiv preprint arXiv:2310.00132 (2023) 2, 4, 11" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Xiang Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "type": "text", + "content": "35. Li, X., Wang, J., Xu, X., Yang, M., Yang, F., Zhao, Y., Singh, R., Raj, B.: Towards noise-tolerant speech-referring video object segmentation: Bridging speech and text. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. pp. 2283-2296 (2023) 4, 11" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 482, + 194 + ], + "type": "text", + "content": "36. Liang, T., He, Z., Jiao, W., Wang, X., Wang, Y., Wang, R., Yang, Y., Tu, Z., Shi, S.: Encouraging divergent thinking in large language models through multi-agent debate. arXiv preprint arXiv:2305.19118 (2023) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 482, + 237 + ], + "type": "text", + "content": "37. Liu, J., Ding, H., Cai, Z., Zhang, Y., Satzoda, R.K., Mahadevan, V., Manmatha, R.: Polyformer: Referring image segmentation as sequential polygon generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18653-18663 (2023) 2, 4, 10" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 237, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 482, + 270 + ], + "type": "text", + "content": "38. Liu, J., Wang, Y., Ju, C., Ma, C., Zhang, Y., Xie, W.: Annotation-free audio-visual segmentation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5604-5614 (2024) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 270, + 482, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 482, + 304 + ], + "type": "text", + "content": "39. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 482, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 482, + 346 + ], + "type": "text", + "content": "40. Liu, S., Hasan, A., Hong, K., Wang, R., Chang, P., Mizrachi, Z., Lin, J., McPherson, D.L., Rogers, W.A., Driggs-Campbell, K.: Dragon: A dialogue-based robot for assistive navigation with visual language grounding. IEEE Robotics and Automation Letters (2024) 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 347, + 482, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 482, + 389 + ], + "type": "text", + "content": "41. Mao, J., Huang, J., Toshev, A., Camburu, O., Yuille, A.L., Murphy, K.: Generation and comprehension of unambiguous object descriptions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 11-20 (2016) 9, 10" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 390, + 482, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 390, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 390, + 482, + 424 + ], + "type": "text", + "content": "42. Miao, B., Bennamoun, M., Gao, Y., Mian, A.: Spectrum-guided multi-granularity referring video object segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 920-930 (2023) 4, 10, 11" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 424, + 482, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 482, + 445 + ], + "type": "text", + "content": "43. Mo, S., Morgado, P.: A closer look at weakly-supervised audio-visual source localization. arXiv preprint arXiv:2209.09634 (2022) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 446, + 482, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 482, + 490 + ], + "type": "text", + "content": "44. Pan, W., Shi, H., Zhao, Z., Zhu, J., He, X., Pan, Z., Gao, L., Yu, J., Wu, F., Tian, Q.: Wnet: Audio-guided video object segmentation via wavelet-based cross-modal denoising networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1320-1331 (2022) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 490, + 482, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 482, + 522 + ], + "type": "text", + "content": "45. Pont-Tuset, J., Peruzzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675 (2017) 10" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 522, + 482, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 482, + 555 + ], + "type": "text", + "content": "46. Senocak, A., Oh, T.H., Kim, J., Yang, M.H., Kweon, I.S.: Learning to localize sound source in visual scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4358-4366 (2018) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 555, + 482, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 482, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 482, + 600 + ], + "type": "text", + "content": "47. Seo, S., Lee, J.Y., Han, B.: Urvos: Unified referring video object segmentation network with a large-scale benchmark. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16. pp. 208-223. Springer (2020) 4, 9" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 600, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 482, + 632 + ], + "type": "text", + "content": "48. Sun, J., Huang, D.A., Lu, B., Liu, Y.H., Zhou, B., Garg, A.: Visually-grounded planning with transformers in procedural tasks. IEEE Robotics and Automation Letters 7(2), 4924-4930 (2022) 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 482, + 666 + ], + "type": "text", + "content": "49. Tang, J., Zheng, G., Yang, S.: Temporal collection and distribution for referring video object segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15466-15476 (2023) 4" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 132, + 91, + 470, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 91, + 470, + 102 + ], + "spans": [ + { + "bbox": [ + 132, + 91, + 470, + 102 + ], + "type": "text", + "content": "Benchmarking the Robustness of Referring Perception Models under Perturbations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 492, + 91, + 503, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 492, + 91, + 503, + 100 + ], + "spans": [ + { + "bbox": [ + 492, + 91, + 503, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "50. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023) 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "text", + "content": "51. Tziafas, G., Kasaei, H.: Few-shot visual grounding for natural human-robot interaction. In: 2021 IEEE International Conference on Autonomous Robot Systems and Competitions (ICARSC). pp. 50-55. IEEE (2021) 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "text", + "content": "52. Wang, Z., Cai, S., Chen, G., Liu, A., Ma, X.S., Liang, Y.: Describe, explain, plan and select: interactive planning with llms enables open-world multi-task agents. Advances in Neural Information Processing Systems 36 (2024) 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "53. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022) 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "54. Wu, D., Wang, T., Zhang, Y., Zhang, X., Shen, J.: Onlinerefer: A simple online baseline for referring video object segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2761-2770 (2023) 2, 4, 10, 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "type": "text", + "content": "55. Wu, J., Jiang, Y., Sun, P., Yuan, Z., Luo, P.: Language as queries for referring video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4974-4984 (2022) 4, 10" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "text", + "content": "56. Wu, J., Jiang, Y., Yan, B., Lu, H., Yuan, Z., Luo, P.: Segment every reference object in spatial and temporal spaces. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2538-2550 (2023) 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "57. Xi, Z., Chen, W., Guo, X., He, W., Ding, Y., Hong, B., Zhang, M., Wang, J., Jin, S., Zhou, E., et al.: The rise and potential of large language model based agents: A survey. arXiv preprint arXiv:2309.07864 (2023) 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "58. Xiong, Y., Varadarajan, B., Wu, L., Xiang, X., Xiao, F., Zhu, C., Dai, X., Wang, D., Sun, F., Iandola, F., et al.: Efficientsam: Leveraged masked image pretraining for efficient segment anything. arXiv preprint arXiv:2312.00863 (2023) 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "type": "text", + "content": "59. Xu, N., Yang, L., Fan, Y., Yue, D., Liang, Y., Yang, J., Huang, T.: Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327 (2018) 9, 11" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "text", + "content": "60. Xu, X., Wang, J., Li, X., Lu, Y.: Reliable propagation-correction modulation for video object segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 2946-2954 (2022) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "61. Xu, X., Wang, J., Ming, X., Lu, Y.: Towards robust video object segmentation with adaptive object calibration. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 2709-2718 (2022) 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "text", + "content": "62. Xu, X., Zhang, T., Wang, S., Li, X., Chen, Y., Li, Y., Raj, B., Johnson-Roberson, M., Huang, X.: Customizable perturbation synthesis for robust slam benchmarking. arXiv preprint arXiv:2402.08125 (2024) 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 545, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 587 + ], + "type": "text", + "content": "63. Xu, Z., Chen, Z., Zhang, Y., Song, Y., Wan, X., Li, G.: Bridging vision and language encoders: Parameter-efficient tuning for referring image segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17503-17512 (2023) 4, 10" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 588, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 481, + 621 + ], + "type": "text", + "content": "64. Yamazaki, K., Hanyu, T., Vo, K., Pham, T., Tran, M., Doretto, G., Nguyen, A., Le, N.: Open-fusion: Real-time open-vocabulary 3d mapping and queryable scene representation. arXiv preprint arXiv:2310.03923 (2023) 4, 12" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "65. Yamazaki, K., Truong, S., Vo, K., Kidd, M., Rainwater, C., Luu, K., Le, N.: Vlcap: Vision-language with contrastive learning for coherent video paragraph captioning. In: 2022 IEEE International Conference on Image Processing (ICIP). pp. 3656-3661. IEEE (2022) 4" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Xiang Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "66. Yamazaki, K., Vo, K., Truong, Q.S., Raj, B., Le, N.: Vlnt: Visual-linguistic transformer-in-transformer for coherent video paragraph captioning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 3081-3090 (2023) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 173 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 173 + ], + "type": "text", + "content": "67. Yang, J., Gao, M., Li, Z., Gao, S., Wang, F., Zheng, F.: Track anything: Segment anything meets videos (2023) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 173, + 482, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 173, + 482, + 217 + ], + "spans": [ + { + "bbox": [ + 130, + 173, + 482, + 217 + ], + "type": "text", + "content": "68. Yang, Z., Wang, J., Tang, Y., Chen, K., Zhao, H., Torr, P.H.: Levt: Language-aware vision transformer for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18155-18165 (2022) 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 218, + 482, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 218, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 218, + 482, + 262 + ], + "type": "text", + "content": "69. Yang, Z., Wang, J., Tang, Y., Chen, K., Zhao, H., Torr, P.H.: Levt: Language-aware vision transformer for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18155-18165 (2022) 10" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 262, + 482, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 295 + ], + "type": "text", + "content": "70. Yang, Z., Wei, Y., Yang, Y.: Associating objects with transformers for video object segmentation. Advances in Neural Information Processing Systems 34, 2491-2502 (2021) 4, 10" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 296, + 482, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 328 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 328 + ], + "type": "text", + "content": "71. Yang, Z., Yang, Y.: Decoupling features in hierarchical propagation for video object segmentation. Advances in Neural Information Processing Systems 35, 36324-36336 (2022) 4, 10, 11" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 330, + 482, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 482, + 362 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 482, + 362 + ], + "type": "text", + "content": "72. Yang, Z., Yang, Y.: Decoupling features in hierarchical propagation for video object segmentation. Advances in Neural Information Processing Systems 35, 36324-36336 (2022) 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 363, + 482, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 385 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 385 + ], + "type": "text", + "content": "73. Yao, J., Wang, X., Ye, L., Liu, W.: Matte anything: Interactive natural image matting with segment anything models. arXiv preprint arXiv:2306.04121 (2023) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 386, + 482, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 418 + ], + "type": "text", + "content": "74. Yu, L., Poirson, P., Yang, S., Berg, A.C., Berg, T.L.: Modeling context in referring expressions. In: European Conference on Computer Vision. pp. 69-85. Springer (2016) 9, 10" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 419, + 482, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 482, + 452 + ], + "type": "text", + "content": "75. Yu, T., Feng, R., Feng, R., Liu, J., Jin, X., Zeng, W., Chen, Z.: Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790 (2023) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 453, + 482, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 485 + ], + "type": "text", + "content": "76. Zhang, J., Cui, Y., Wu, G., Wang, L.: Joint modeling of feature, correspondence, and a compressed memory for video object segmentation. arXiv preprint arXiv:2308.13505 (2023) 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 487, + 482, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 518 + ], + "type": "text", + "content": "77. Zhao, Q., Wang, J., Zhang, Y., Jin, Y., Zhu, K., Chen, H., Xie, X.: Competeai: Understanding the competition behaviors in large language model-based agents. arXiv preprint arXiv:2310.17512 (2023) 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 520, + 482, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 520, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 520, + 482, + 553 + ], + "type": "text", + "content": "78. Zhou, J., Shen, X., Wang, J., Zhang, J., Sun, W., Zhang, J., Birchfield, S., Guo, D., Kong, L., Wang, M., et al.: Audio-visual segmentation with semantics. arXiv preprint arXiv:2301.13190 (2023) 4, 9" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 554, + 482, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 554, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 554, + 482, + 586 + ], + "type": "text", + "content": "79. Zhou, J., Wang, J., Zhang, J., Sun, W., Zhang, J., Birchfield, S., Guo, D., Kong, L., Wang, M., Zhong, Y.: Audio-visual segmentation. In: European Conference on Computer Vision (2022) 2, 4, 11" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 587, + 482, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 587, + 482, + 631 + ], + "spans": [ + { + "bbox": [ + 130, + 587, + 482, + 631 + ], + "type": "text", + "content": "80. Zhu, C., Chen, F., Ahmed, U., Shen, Z., Savvides, M.: Semantic relation reasoning for shot-stable few-shot object detection. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 8778-8787 (2021), https://api.sementicscholar.org/CorpusID:232093016 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 632, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 665 + ], + "type": "text", + "content": "81. Zhu, C., Chen, F., Shen, Z., Savvides, M.: Soft anchor-point object detection. In: European Conference on Computer Vision (2019), https://api-semanticscholar.org/CorpusID:208512715 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 132, + 91, + 470, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 91, + 470, + 102 + ], + "spans": [ + { + "bbox": [ + 132, + 91, + 470, + 102 + ], + "type": "text", + "content": "Benchmarking the Robustness of Referring Perception Models under Perturbations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 492, + 91, + 503, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 492, + 91, + 503, + 100 + ], + "spans": [ + { + "bbox": [ + 492, + 91, + 503, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 216 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "type": "text", + "content": "82. Zou, X., Dou, Z.Y., Yang, J., Gan, Z., Li, L., Li, C., Dai, X., Behl, H., Wang, J., Yuan, L., et al.: Generalized decoding for pixel, image, and language. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15116-15127 (2023) 4, 10" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 482, + 182 + ], + "type": "text", + "content": "83. Zou, X., Yang, J., Zhang, H., Li, F., Li, L., Gao, J., Lee, Y.J.: Segment everything everywhere all at once. arXiv preprint arXiv:2304.06718 (2023) 4, 10" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 482, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 482, + 216 + ], + "type": "text", + "content": "84. Zou, X., Yang, J., Zhang, H., Li, F., Li, L., Wang, J., Wang, L., Gao, J., Lee, Y.J.: Segment everything everywhere all at once. Advances in Neural Information Processing Systems 36 (2024) 4, 10" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Xiang Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/R^2-Tuning_ Efficient Image-to-Video Transfer Learning for Video Temporal Grounding/148e844e-b528-47e5-9db4-f50b0b9b4b9c_content_list.json b/2024/R^2-Tuning_ Efficient Image-to-Video Transfer Learning for Video Temporal Grounding/148e844e-b528-47e5-9db4-f50b0b9b4b9c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..17d7a1358a68918a9e174394a7c9c5d0f0487283 --- /dev/null +++ b/2024/R^2-Tuning_ Efficient Image-to-Video Transfer Learning for Video Temporal Grounding/148e844e-b528-47e5-9db4-f50b0b9b4b9c_content_list.json @@ -0,0 +1,1853 @@ +[ + { + "type": "text", + "text": "$\\mathbf{R}^2$ -Tuning: Efficient Image-to-Video Transfer Learning for Video Temporal Grounding", + "text_level": 1, + "bbox": [ + 243, + 142, + 759, + 189 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ye Liu $^{1,3*}$ , Jixuan He $^{2,3}$ , Wanhua Li $^{3\\dagger}$ , Junsik Kim $^{3}$ , Donglai Wei $^{4}$ , Hanspeter Pfister $^{3}$ , and Chang Wen Chen $^{1\\dagger}$", + "bbox": [ + 228, + 214, + 774, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 The Hong Kong Polytechnic University 2 Tsinghua University 3 Harvard University 4 Boston College coco.ye.liu@connect.polyu.hk hejixuan2000@gmail.com wanhua@seas.harvard.edu mibastro@gmail.com donglai.wei@bc.edu pfister@seas.harvard.edu changwen.chen@polyu.edu.hk", + "bbox": [ + 261, + 256, + 740, + 330 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Video temporal grounding (VTG) is a fine-grained video understanding problem that aims to ground relevant clips in untrimmed videos given natural language queries. Most existing VTG models are built upon frame-wise final-layer CLIP features, aided by additional temporal backbones (e.g., SlowFast) with sophisticated temporal reasoning mechanisms. In this work, we claim that CLIP itself already shows great potential for fine-grained spatial-temporal modeling, as each layer offers distinct yet useful information under different granularity levels. Motivated by this, we propose Reversed Recurrent Tuning (R²-Tuning), a parameter- and memory-efficient transfer learning framework for video temporal grounding. Our method learns a lightweight R² Block containing only 1.5% of the total parameters to perform progressive spatial-temporal modeling. Starting from the last layer of CLIP, R² Block recurrently aggregates spatial features from earlier layers, then refines temporal correlation conditioning on the given query, resulting in a coarse-to-fine scheme. R²-Tuning achieves state-of-the-art performance across three VTG tasks (i.e., moment retrieval, highlight detection, and video summarization) on six public benchmarks (i.e., QVHighlights, CharadesSTA, Ego4D-NLQ, TACoS, YouTube Highlights, and TVSum) even without the additional backbone, demonstrating the significance and effectiveness of the proposed scheme. Our code is available at https://github.com/yeliudev/R2-Tuning.", + "bbox": [ + 261, + 369, + 743, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Video Temporal Grounding $\\cdot$ Transfer Learning $\\cdot$ CLIP", + "bbox": [ + 261, + 688, + 720, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 732, + 377, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video is becoming the major content media in our daily lives. The variety of video domains has extended beyond daily activities [5,22] but broader scenes such as egocentric [10] and surveillance [60]. They maintain large information", + "bbox": [ + 212, + 767, + 787, + 815 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Work done at Harvard University. † Corresponding authors.", + "bbox": [ + 230, + 824, + 640, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0545960a4f1b6cc8b4bd213c5958d2aaa6b0c29282b65178a9db990f60ef05fe.jpg", + "image_caption": [ + "Untrimmed Video" + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 277, + 234 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/44cc9f9c5c0dcc1f7cbc6f5eb8c68a404c5887900ee1dadc46f3088950515ee1.jpg", + "image_caption": [ + "Fig. 1: Video temporal grounding (VTG) contains three video-language understanding problems, i.e., moment retrieval (MR), highlight detection (HD), and video summarization (VS)." + ], + "image_footnote": [], + "bbox": [ + 279, + 145, + 524, + 244 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/909278c2716e565a2c52590c41482aea87c5dc7ae079358d20f589a7c12af3c1.jpg", + "image_caption": [ + "Fig. 2: Moment retrieval mAP with different backbones on QVHighlights val split. CLIP's potential of temporal modeling was not fully exploited." + ], + "image_footnote": [], + "bbox": [ + 540, + 143, + 784, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "capacities within multi-granularities, and thus can convey both high-level context and low-level details effectively.", + "bbox": [ + 212, + 340, + 782, + 369 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the content production and consumption industry, such a flourishing is accompanied by the increasing demand for browsing untrimmed videos efficiently with different user interests. This derives the topic of video temporal grounding (VTG) [26], which is a fine-grained video-language understanding problem aiming to ground video clips conditioning on natural language queries. As shown in Figure 1, VTG can be disentangled into three tasks based on different output formats, i.e., moment retrieval (MR) [9, 10, 22, 31] for regressing start-end timestamps, highlight detection (HD) [22, 47] for predicting frame-level saliency curves, and video summarization (VS) [2, 11, 45] for classifying disjoint shots.", + "bbox": [ + 212, + 371, + 784, + 507 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Encouraged by the recent success in adopting vision-language models (VLMs) for video understanding [4,15,19,28,32,37], most existing VTG methods [16,22, 31,34,35,46,56] are built upon frame-wise final-layer features from CLIP [41]. However, due to CLIP's misaligned pre-training objective (image-text contrast), these methods fail to capture temporal correlations well. As a feasible compromise, an additional backbone (e.g., SlowFast [8]) is incorporated to complement the temporal information, followed by carefully designed modules such as text-guided queries [31], dynamic anchors [35], and event reasoning [16].", + "bbox": [ + 212, + 507, + 785, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We refer to the paradigm above as a post-processing scheme (shown in Figure. 3 (a)), whereas two natural drawbacks exist due to the sub-optimal design. First, leveraging two backbones with similar capabilities is unintuitive and inefficient during inference. A single model with both vision-text alignment and spatial-temporal modeling abilities is more preferred. Second, queries for VTG could be of different granularities from coarse (e.g., the family is traveling) to fine (e.g., when did I take the golf club from the man with white hair). Leveraging only frame-wise & final-layer features is not granularity flexible, as it would focus more on high-level scene transitions while overlooking low-level details. Preliminary experiments in Figure 2 also demonstrate that the potential of spatial-temporal modeling for CLIP is not fully exploited by existing methods. Some recent works [15,19,57] tried to tackle the first problem by fine-tuning (part of) the CLIP encoders (Figure 3 (b) and (c)), but they are all inefficient in terms of data, learnable parameters, or memory.", + "bbox": [ + 212, + 628, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c47d92be81ded838e7ff94e79816dbc57933c7ede699e37d1c1477ee0e80e728.jpg", + "image_caption": [ + "Fig. 3: Different architectural designs for CLIP-based image-to-video transfer learning. The gray rectangle in (d) denotes the progressively refined spatial-temporal features." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 357, + 284 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ebd9a2205c8d88f5471fa446e5f8719ae9d78c0686598d11fc630beb91c9859b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 143, + 491, + 284 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6213e6af772ca6d4f3ef81f729e6fc1f7990cef419bc6273a4b2b3dbd2105c66.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 143, + 627, + 284 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4bd2d4ddef0e16fb72e7c86773e86ed259cba3f3517bfa6c02ddf98c7e2969e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 145, + 784, + 284 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This paper intends to answer the research question: how to efficiently transfer an image-language foundation model to video temporal grounding? We consider the above question by exploring two aspects: efficiency and granularity flexibility. To address these issues, a novel image-to-video transfer learning framework called Reversed Recurrent Tuning (R²-Tuning) is proposed for fine-grained understanding on untrimmed videos. Our insight is that multi-layer CLIP features offer distinct yet useful information, while their integration should be tailored to the downstream task. As illustrated in Figure 3 (d), based on a frozen CLIP [41], our method learns a side-block (R² Block) containing only $1.5\\%$ of the total parameters to perform spatial-temporal modeling. R² Block is recurrently attached to the last few layers of CLIP encoders from back to front, performing query-modulated spatial pooling and recurrent temporal refinement from coarse to fine. Aside from the new architecture, we also introduce video-level and layer-wise constraints to calibrate the granularities of visual and text encoders. During training, gradient flows do not pass through the CLIP encoders, thus our scheme is both parameter- and memory-efficient. It is also granularity-flexible as R² Block can adaptively control the spatial pooling strategy conditioning on queries.", + "bbox": [ + 212, + 356, + 787, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We conduct extensive experiments across three VTG tasks on six public benchmarks, including QVHighlights [22], Ego4D-NLQ [10], Charades-STA [9], TACoS [9], YouTube Highlights [47], and TVSum [45]. Without bells and whistles, $\\mathbb{R}^2$ -Tuning achieves more than 3 MR mAP gain on QVHighlights test split compared with $4\\times$ heavier counterparts [34] with additional temporal backbones and carefully designed post-processing modules. Further analysis also shows that our method can better handle multi-granularity information. Overall, our contributions are summarized as: (1) We introduced $\\mathbb{R}^2$ -Tuning, a novel image-to-video transfer learning framework tailored for video temporal grounding. (2) We designed two effective strategies, i.e., query-modulated spatial pooling and recurrent temporal refinement, to model spatial-temporal information from coarse to fine. (3) To calibrate the granularities of CLIP visual and text encoders, we further introduce video-level and layer-wise contrastive constraints to distill distinct information from each layer. (4) Extensive experiments across three tasks on six benchmarks demonstrate the significance and effectiveness of $\\mathbb{R}^2$ -Tuning.", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Reversed Recurrent Tuning", + "bbox": [ + 547, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 217, + 143, + 385, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CLIP for Video Understanding With powerful transfer learning abilities, CLIP [41] has been widely used for many image-language tasks [18, 29], and how to extend them to video understanding is an emerging and crucial research topic. Early attempts have been made to transfer CLIP to trimmed videos for action recognition [28,37,42] and video-text retrieval [4,15,19,32]. Specifically, Efficient-Prompt [19] introduces learnable prompt tokens to the text encoder, and ViFi-CLIP [42] further adopts learnable prompts at the visual encoder. Other works also tried to learn external modules designed for temporal modeling. EVL [28] proposes a temporal decoder in parallel with the main network. CLIP4Clip [32] investigated multiple temporal fusion strategies for video frames. Nevertheless, how to effectively transfer CLIP to untrimmed videos remains unexplored. To our best knowledge, $\\mathbb{R}^2$ -Tuning is the first solution for fine-grained video temporal grounding via memory-efficient transfer learning.", + "bbox": [ + 217, + 180, + 785, + 377 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video Temporal Grounding Video temporal grounding (VTG) [26] is a bundle of video understanding problems including moment retrieval (MR) [9,16,22, 31,35,36,52,56,57,61], highlight detection (HD) [3,14,47,55], and video summarization (VS) [11,45]. Although these tasks are closely related, they have not been jointly studied until recently. QVHighlights [22] is the first dataset that supports both MR and HD, breaking the moment distribution bias in previous MR datasets [9,43]. The authors also proposed Moment-DETR, a strong baseline method for MR/HD. Most following works were built upon the Moment-DETR framework with specially designed modules. Notably, UMT [31] tackles flexible MR and HD through a unified multi-modal architecture. QD-DETR [35] benefits from query-dependent video representations and dynamic anchors. EaTR [16] incorporates moment reasoning to provide reliable referential search areas for moment queries. Nevertheless, their design space is still within pre-extracted CLIP [41] and SlowFast [8] features. A recent work [57] revealed the possibility of fine-tuning CLIP for MR, while their solution (post pre-training) is extremely resource-consuming. Compared with existing works, $\\mathbb{R}^2$ -Tuning can effectively learn strong temporal modeling abilities without any pre-training on videos.", + "bbox": [ + 217, + 396, + 785, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Parameter-Efficient Transfer Learning Foundation models [41, 49] have achieved remarkable successes in vision and language understanding. Efforts have been made to efficiently transfer knowledge from these models to new scenarios. One line of research is to conduct context-based tuning (e.g., prompt-tuning [17, 65, 66]), which does not modify the model architectures but incorporates learnable embeddings in the inputs. Another paradigm is introducing an extra lightweight adapter to the original model, and keeping the rest of the parameters frozen [39]. Recent works also consider memory-efficient transfer learning (METL) that introduces parameters whose gradient flows do not go through the main model [28, 40, 48]. $\\mathbf{R}^2$ -Tuning is also a METL framework, where we adopt a novel recurrent tuning strategy to ensure parameter- and memory-efficient.", + "bbox": [ + 217, + 672, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/23773a00ff20ed68d2af5e26b0a9008d233cec11fe04d5e92343f9d9b6060da3.jpg", + "image_caption": [ + "Fig. 4: Overall architecture of our framework. The input video and query are first encoded by frozen CLIP [41] encoders. Their multi-layer outputs are then recurrently fused and refined by a learnable $\\mathbb{R}^2$ Block to construct spatial-temporal representations $h$ , which would be scaled up/down to construct a temporal feature pyramid, followed by three heads for MR, HD, and VS, respectively." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 386, + 380, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Problem Formulation", + "text_level": 1, + "bbox": [ + 215, + 417, + 437, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a video $V = \\{v_{i}\\}_{i=1}^{T}$ and a natural language query $Q = \\{q_{i}\\}_{i=1}^{L}$ , where $T$ and $L$ are the numbers of video frames and text tokens, VTG aims to densely predict a set of labels $(b_{i}, s_{i}, f_{i})$ for each frame, defined as follows:", + "bbox": [ + 212, + 443, + 785, + 491 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Moment Retrieval is to find the most relevant moments (i.e. sets of consecutive frames) in $V$ according to $Q$ , so that $b_{i} = [b_{i}^{s}, b_{i}^{e}] \\in \\mathbb{R}^{2}$ represents the temporal displacements from frame $v_{i}$ to the start and end timestamps of the nearest target moment.", + "- Highlight Detection requires estimating frame-level relevancies between $V$ and $Q$ , thus $s_i \\in [0,1]$ is a continuous saliency score denoting to what extent frame $v_i$ is semantically aligned with $Q$ .", + "- Video Summarization aims to select a subset of video frames according to $Q$ to form a concise summary, therefore, $f_{i} \\in \\{0,1\\}$ is a binary score indicating whether frame $v_{i}$ belongs to the summary." + ], + "bbox": [ + 225, + 503, + 785, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Overview", + "text_level": 1, + "bbox": [ + 215, + 676, + 339, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 4 shows an overview of the proposed framework. Our model derives from a pre-trained and frozen CLIP [41] with ViT [6] backbone, which has a two-stream architecture for spatial and textual encoding, respectively. A learnable $\\mathbb{R}^2$ block is iteratively attached to the last $K$ encoder layers to refine the temporal correlation. The resulting features are then scaled up/down to build a feature pyramid, followed by three heads to decode the task-specific predictions.", + "bbox": [ + 212, + 703, + 785, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, the input video $V$ and query $Q$ are first tokenized into frame patches and word tokens, then sent into the visual encoder $E_{v}$ and query encoder $E_{q}$ . The encoded visual and textual features can be denoted as $e_v \\in$", + "bbox": [ + 212, + 795, + 785, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Reversed Recurrent Tuning", + "bbox": [ + 545, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9a849fabc871996edfcd43ef75c076d8f644dcff3fb91447fbab73a7bbb56a50.jpg", + "image_caption": [ + "Fig. 5: Detailed architecture of the $\\mathbb{R}^2$ Block. It can be split into two parts: a) query-modulated spatial pooling, and b) recurrent temporal refinement. Note that the [CLS] tokens of visual features are omitted for clarity." + ], + "image_footnote": [], + "bbox": [ + 344, + 146, + 660, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\mathbb{R}^{B \\times N \\times T \\times (P + 1) \\times D_v}$ and $e_q \\in \\mathbb{R}^{B \\times N \\times L \\times D_q}$ , where $B, N, T, P, L, D_v$ and $D_q$ indicate batch size, number of encoder layers, number of video frames, number of patches per frame, number of query tokens, and the dimensions of visual/query features, respectively. These features are recurrently fused and refined by $\\mathbb{R}^2$ block to construct spatial-temporal features $h \\in \\mathbb{R}^{B \\times T \\times C}$ , in which each token preserves the $C$ -dimensional features for a frame. This process will be introduced in detail in Section 3.3.", + "bbox": [ + 212, + 387, + 787, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Reversed Recurrent Tuning", + "text_level": 1, + "bbox": [ + 215, + 517, + 491, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given pre-trained visual and query encoders, existing works [16,22,26,31,35] merely take features from the last hidden layer, which is sub-optimal due to the limited information. In this work, we propose $\\mathbb{R}^2$ -Tuning to exploit the potential of spatial-temporal modeling based on multi-layer CLIP features. Our scheme learns a lightweight $\\mathbb{R}^2$ Block which maintains a zero-initialized hidden state $h \\in \\mathbb{R}^{B \\times T \\times C}$ as frame-level spatial-temporal features. The $\\mathbb{R}^2$ Block is recurrently attached to the last $K$ layers of CLIP encoders from back to front to refine the hidden state $h$ for $K$ steps, with each step defined as follows:", + "bbox": [ + 212, + 542, + 787, + 664 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nh ^ {k} = \\mathcal {F} _ {\\theta} (e _ {v} ^ {n}, e _ {q} ^ {n}, h ^ {k - 1}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 421, + 674, + 785, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $\\mathcal{F}_{\\theta}$ is the refinement operation parameterized by $\\theta$ . $k \\in [1, K]$ is the index of current step. $n = N - k + 1$ denotes the CLIP layer index for step $k$ , thus $e_v^n$ and $e_q^n$ are visual and query features from the $n$ -th CLIP encoder layer. The refinement operation $\\mathcal{F}_{\\theta}$ can be decomposed into two parts: 1) query-modulated spatial pooling, and 2) recurrent temporal refinement. Details are discussed as follows.", + "bbox": [ + 214, + 703, + 787, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Query-Modulated Spatial Pooling Figure 5 (a) presents the data flow of query-modulated spatial pooling. The goal of this process is to adaptively pool", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "spatial features from patch-level representations $e_v^n$ to a single token $e_{pool}^n$ conditioning on the query $e_q^n$ . We first adopt two separate MLPs to map $e_v^n$ and $e_q^n$ into the same space:", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {e} _ {v} ^ {n} = \\operatorname {M L P} \\left(e _ {v} ^ {n}\\right) \\in \\mathbb {R} ^ {B \\times T \\times (P + 1) \\times C} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 200, + 785, + 217 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {e} _ {q} ^ {n} = \\operatorname {M L P} \\left(e _ {q} ^ {n}\\right) \\in \\mathbb {R} ^ {B \\times L \\times C} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 220, + 784, + 239 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, $C$ is the reduced hidden size. We further align the shape of $\\hat{e}_v^n$ and $\\hat{e}_q^n$ by merging dimension $T$ into $B$ and repeat $\\hat{e}_q^n$ for $T$ times in an interleaved style, resulting $\\hat{e}_v^n \\in \\mathbb{R}^{(B \\times T) \\times (P + 1) \\times C}$ and $\\hat{e}_q^n \\in \\mathbb{R}^{(B \\times T) \\times L \\times C}$ . Then, we compute the similarities for patch-token pairs using normalized Embedded Gaussian [50]:", + "bbox": [ + 214, + 247, + 782, + 311 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\na = \\operatorname {s o f t m a x} \\left(\\frac {\\left(w _ {q} \\hat {e} _ {q} ^ {n}\\right) ^ {\\top} w _ {v} \\hat {e} _ {v} ^ {n}}{\\sqrt {C}}\\right) \\in \\mathbb {R} ^ {(B \\times T) \\times L \\times P} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 320, + 784, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, $w_{v}$ and $w_{q}$ are learnable matrices for projecting features. This operation derives from cross-attention but discards the linear projection on value. We then multiply the similarities $a$ with visual features $\\hat{e}_{v}^{n}$ to pool them into each token, such that tokens can interact with patches independently. After that, a max pooling along the $L$ dimension is applied:", + "bbox": [ + 214, + 361, + 782, + 436 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\ne _ {t o k e n} ^ {n} = \\operatorname {M a x P o o l i n g} \\left(a \\hat {e} _ {v} ^ {n} + \\hat {e} _ {q} ^ {n}\\right) \\in \\mathbb {R} ^ {B \\times T \\times C} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 348, + 446, + 784, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, $\\hat{e}_q^n$ serves as residuals [13] to stabilize training. Our visualizations in Figure 8 show that this strategy can guide the model to focus more on query-related regions. Finally, we combine $e_{token}^n$ with the [CLS] token to generate query-modulated spatial features:", + "bbox": [ + 214, + 472, + 782, + 532 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\ne _ {p o o l} ^ {n} = e _ {v} ^ {n, 0} + g ^ {k} \\cdot e _ {t o k e n} ^ {n} \\in \\mathbb {R} ^ {B \\times T \\times C} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 542, + 784, + 560 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, $g^{k} \\in (-1,1)$ is a zero-initialized learnable gate for step $k$ constrained by Tanh. We allow negative values here to remove useless information from [CLS] tokens. $e_{pool}^{n}$ is then used to model temporal correlations and are refined in a recurrent manner.", + "bbox": [ + 214, + 569, + 782, + 630 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Recurrent Temporal Refinement Figure 5 (b) displays the flow of recurrent temporal refinement. Specifically, the pooled visual features $e_{pool}^{n}$ is first fused with the hidden state $h^{k-1}$ from the previous step:", + "bbox": [ + 214, + 643, + 782, + 691 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {h} ^ {k - 1} = \\varphi^ {k} \\cdot e _ {\\text {p o o l}} ^ {n} + (1 - \\varphi^ {k}) \\cdot h ^ {k - 1} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 699, + 784, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\varphi^k\\in (0,1)$ is the learnable gate for step $k$ . Then, we adopt a sequence of standard multi-head cross-attention $(\\hat{e}_q^n$ as $k,v)$ , multi-head self-attention, followed by a feed-forward network [50] to update the hidden state:", + "bbox": [ + 214, + 727, + 782, + 773 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nh ^ {k} = \\operatorname {F F N} (\\mathrm {M H S A} (\\mathrm {M H C A} (\\hat {h} ^ {k - 1}, \\hat {e} _ {q} ^ {n}))) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 782, + 784, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For each block, we adopt DropPath [21] with $p = 0.1$ to prevent overfitting. The query features $\\hat{e}_q^n$ are utilized both in spatial and temporal modeling as guidance.", + "bbox": [ + 214, + 809, + 782, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Reversed Recurrent Tuning", + "bbox": [ + 545, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Granularity Calibration The refinement process above is based on the assumption that visual and query features from the same layer of CLIP [41] are well aligned at the same granularity level. However, this cannot be guaranteed as the two encoders are learned in isolation during pre-training. Therefore, we need to add manual constraints to calibrate their granularities.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We apply a video-level constraint and a layer-wise constraint by designing two contrastive losses. We first denote $e_v^0 \\in \\mathbb{R}^{B \\times K \\times T \\times C}$ as the features of $K$ [CLS] tokens from Eq. 6. Then, we select all positive frames (i.e., frames within $[b_i^s, b_i^e]$ or with $s_i$ or $f_i$ higher than a threshold) from $e_v^0$ and perform average pooling to obtain video-level representations $\\widetilde{e}_v \\in \\mathbb{R}^{B \\times K \\times C}$ .", + "bbox": [ + 215, + 222, + 784, + 297 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {e} _ {v} = \\operatorname {A v g P o o l} \\left(\\left\\{e _ {v} ^ {0, i} \\right\\} _ {i \\in \\Omega}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 306, + 784, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here, $\\Omega$ is the set of positive frame indices. To obtain query-level representations $\\widetilde{e}_q\\in \\mathbb{R}^{B\\times K\\times C}$ , a token-wise adaptive pooling is applied following previous work [26]. We adopt InfoNCE loss [38] along two dimensions to calculate the video-level and layer-wise constraints:", + "bbox": [ + 214, + 330, + 782, + 390 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {v i d e o}} = \\lambda_ {\\text {v i d e o}} \\frac {1}{B} \\sum_ {b \\in B} \\operatorname {I n f o N C E} \\left(\\tilde {e} _ {v} ^ {b}, \\tilde {e} _ {q} ^ {b}\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 398, + 784, + 433 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {l a y e r}} = \\lambda_ {\\text {l a y e r}} \\frac {1}{K} \\sum_ {k \\in K} \\operatorname {I n f o N C E} \\left(\\widetilde {e} _ {v} ^ {k}, \\widetilde {e} _ {q} ^ {k}\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 436, + 784, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here, $\\mathcal{L}_{video}$ performs contrast among samples in the same batch (diversifying features are diverse among moment-query pairs), and averages the loss across $K$ layers. It also implicitly models temporal correlations between videos and queries. $\\mathcal{L}_{layer}$ performs contrast among layers and averages across the batch. $\\lambda_{video}$ and $\\lambda_{layer}$ serve as re-weighting terms for the losses. A fixed temperature parameter of 0.07 is used.", + "bbox": [ + 214, + 479, + 784, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Prediction Heads", + "text_level": 1, + "bbox": [ + 214, + 590, + 403, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "After refining spatial-temporal features $h$ , a temporal feature pyramid is constructed by applying 1D convolutions with stride $= 2$ on $h$ . The number of convolutions used for each level is subjected to $l - 1$ , where $l$ is the level index starting from 1. We concatenate features from all levels by the temporal dimension to form $\\hat{h}$ and predict once in all heads. Following a similar but simplified design from [26], we adopt three heads for video temporal grounding, described in detail as follows.", + "bbox": [ + 212, + 612, + 782, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Foreground-Background Classification Given $\\hat{h}$ , a two-layer 1D convolution module with $(1\\times 3)$ kernels is adopted to predict $\\hat{f}_i$ for each frame. It is optimized using a focal loss [27] with $\\alpha = 0.9$ and $\\gamma = 2.0$ :", + "bbox": [ + 214, + 729, + 782, + 775 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c l s} = - \\lambda_ {c l s} \\alpha (1 - \\hat {f} _ {i}) ^ {\\gamma} \\log (\\hat {f} _ {i}) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 784, + 784, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We find that focal loss perform better than binary cross-entropy when the numbers of foreground/background frames are severely imbalanced as noted in [26].", + "bbox": [ + 214, + 809, + 782, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/aaf408237a04185f85ea2a00aa2864eb9fb9b1e63b730526f125088ca475cf22.jpg", + "table_caption": [ + "Table 1: Video moment retrieval (MR) and highlight detection (HD) results on QVHighlights test split. Note that [16] and [57] only reported their results on val split. $\\bigcirc$ and $\\bigcirc$ are CLIP-B/32 and B/16 [41], $\\bigcirc$ denotes SlowFast R-50 [8], and $\\bigcirc$ is PANN [20] for audio features extraction. * means estimated parameters. The best and second-best metrics are marked with bold and underline, respectively." + ], + "table_footnote": [], + "table_body": "
MethodBackbonePost \nPre-trainMRHD
R1mAP≥ Very Good#Params
@0.5@0.7@0.5@0.75Avg.mAPHIT@1
BeautyThumb [44]○○X-----14.3620.88-
DVSE [30]○○X-----18.7521.79-
MCN [1]○○X11.412.7224.948.2210.67---
CAL [7]○○X25.4911.5423.407.659.89---
XML [23]○○X41.8330.3544.6331.7332.1434.4955.25-
XML+ [22]○○X46.6933.4647.8934.6734.9035.3855.06-
Moment-DETR [22]○○X52.8933.0254.8229.4030.7335.6955.604.8M
UMT [31]○○○X56.2341.1853.8337.0136.1238.1859.9914.9M
MomentDiff [24]○○○X58.2141.4854.5737.2136.84---
QD-DETR [35]○○X62.4044.9862.5239.8839.8638.9462.407.6M
MH-DETR [56]○○X60.0542.4860.7538.1338.3838.2260.518.2M
UniVTG [26]○○X58.8640.8657.6035.5935.4738.2060.9641.3M
TR-DETR [46]○○X64.6648.9663.9843.7342.6239.9163.427.9M
CG-DETR [34]○○X65.4348.3864.5142.7742.8640.3366.2112.0M
EaTR [16]○○X61.3645.7961.8641.9141.7437.1558.659.0M
Moment-DETR [22]○○236K ASR Cap.59.7840.3360.5135.3636.1437.4360.174.8M
UMT [31]○○○236K ASR Cap.60.8343.2657.3339.1238.0839.1262.3914.9M
UniVTG [26]○○4.2M Corpus65.4350.0664.0645.0243.6340.5466.2841.3M
UnLoc [57]○650K Videos64.5048.80-----87.9M*
R²-Tuning (Ours)○X68.0349.3569.0447.5646.1740.7564.202.7M
", + "bbox": [ + 215, + 224, + 787, + 486 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Boundary Regression Similar to foreground-background classification, boundary regression for moments is also realized by a two-layer 1D convolution module. The difference is that its output dimension is set to 2 instead of 1, representing the boundary displacements $[\\hat{b}_i^s,\\hat{b}_i^e ]$ for start-end timestamps. This head is optimized by an L1 loss:", + "bbox": [ + 212, + 510, + 787, + 585 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e g} = \\lambda_ {r e g} \\left(\\left| b _ {i} ^ {s} - \\hat {b} _ {i} ^ {s} \\right| + \\left| b _ {i} ^ {e} - \\hat {b} _ {i} ^ {e} \\right|\\right) \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 590, + 785, + 609 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $b_{i}^{s}$ and $b_{i}^{e}$ are the ground truths. This loss is only applied to frames inside the ground truth boundaries. We observe that using an L1 loss already works better than the combination of Smooth L1 Loss and GIoU Loss as in [26].", + "bbox": [ + 212, + 614, + 787, + 662 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Saliency Prediction To obtain the saliency scores $\\hat{s}_i$ for HD, we calculate the cosine similarities between the adaptive pooled query features at the last refinement step $\\widetilde{e}_q^K$ and each token in the spatial-temporal features $h$ :", + "bbox": [ + 212, + 670, + 787, + 719 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {s} _ {i} = \\frac {h _ {i} ^ {\\top} \\widetilde {e} _ {q} ^ {K}}{| | h _ {i} | | _ {2} | | \\widetilde {e} _ {q} ^ {K} | | _ {2}} \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 434, + 722, + 785, + 762 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The training objective is applied through a contrastive loss between sampled positive frames (with index $p \\in P$ ) and the adaptively pooled query $\\widetilde{e}_q^K$ :", + "bbox": [ + 212, + 765, + 787, + 797 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s a l} = - \\lambda_ {s a l} \\log \\frac {\\exp \\left(\\hat {s} _ {p} / \\tau\\right)}{\\exp \\left(\\hat {s} _ {p} / \\tau\\right) + \\sum_ {i \\in \\Theta} \\exp \\left(\\hat {s} _ {i} / \\tau\\right)} \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 803, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Reversed Recurrent Tuning", + "bbox": [ + 545, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/382be477779a22ef6bca88cd11c8cd55f1ecdf327c713374d1c7009af04aa285.jpg", + "table_caption": [ + "Table 2: Video moment retrieval results on Ego4D-NLQ, Charades-STA, and TACoS datasets. $\\bigcirc$ means using both CLIP-B/32 and SlowFast R-50 as feature extractors. $\\bigcirc$ indicates using CLIP-B/32 only. The best and second-best metrics are marked with bold and underline, respectively. Our method shows significant advantages in high-quality retrievals (R@0.7) even without a video backbone." + ], + "table_footnote": [], + "table_body": "
MethodEgo4D-NLQ [10]Charades-STA [9]TACoS [43]
R@0.3R@0.5R@0.7mIoUR@0.3R@0.5R@0.7mIoUR@0.3R@0.5R@0.7mIoU
○2D-TAN [64]4.331.830.603.3958.7646.0227.5041.2540.0127.9912.9227.22
○VSLNet [62]4.542.401.013.5460.3042.6924.1441.5835.5423.5413.1524.99
○Moment-DETR [22]4.341.810.653.5365.8352.0730.5945.5437.9724.6711.9725.49
○UniVTG [26]7.283.951.324.9170.8158.0135.6550.1051.4434.9717.3533.60
○R²-Tuning (Ours)7.204.492.124.9470.9159.7837.0250.8649.7138.7225.1235.92
", + "bbox": [ + 215, + 224, + 787, + 323 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Here, $\\Theta$ is the set of frame indices where $s_i < s_p$ , and $\\tau$ is a fixed temperature parameter set to 0.07.", + "bbox": [ + 215, + 352, + 785, + 383 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training & Inference The whole model is jointly optimized using a sum of the five losses mentioned above. In practice, the loss weights are set as $\\lambda_{video} = 0.1$ , $\\lambda_{layer} = 0.1$ , $\\lambda_{cls} = 1.0$ , $\\lambda_{reg} = 0.1$ , and $\\lambda_{sal} = 0.1$ . During inference, we form an moment retrieval prediction by combining $f_{i}$ and $b_{i}$ , i.e., calculating the start-end timestamps from frame index $i$ and boundary displacements $[b_i^s, b_i^e]$ , while regarding $f_{i}$ as the confidence. NMS with IoU threshold $\\theta_{IoU} = 0.7$ is applied to reduce duplicates. For highlight detection and video summarization, we directly use the frame-level output from saliency prediction $s_i$ .", + "bbox": [ + 215, + 401, + 787, + 521 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 547, + 375, + 565 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Datasets & Evaluation Metrics", + "text_level": 1, + "bbox": [ + 215, + 583, + 516, + 598 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We conduct experiments on six datasets covering various domains including daily vlogs & news (QVHighlights [22]), in-door scenes (TACoS [43] and Charades-STA [9]), egocentric videos (Ego4D-NLQ [10]), and sports (YouTube Highlights [47] and TVSum [45]). Details about the datasets are provided in the supplementary material.", + "bbox": [ + 215, + 612, + 785, + 686 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We adopt the same evaluation metrics with previous works [22, 26, 31]. To be specific, we compute Recall@1 with IoU threshold $\\theta_{IoU} = 0.5$ and 0.7, mean average precision (mAP) with $\\theta_{IoU} = 0.5$ and 0.7, and mAP with a series of thresholds [0.5:0.05:0.95] for MR on QVHighlights [22]. mAP and HIT@1 where positive samples are defined as with the saliency score of Very Good are adopted for HD. On Ego4D-NLQ [10], Charades-STA [9], and TACoS [43] datasets, we utilize Recall@1 with $\\theta_{IoU} = \\{0.3, 0.5, 0.7\\}$ and mIoU to measure the MR performance. On YouTube Highlights [47] for HD and TVSum [45] for VS, we follow [26, 31] and use the same train/val split with mAP and Top-5 mAP as metrics, respectively.", + "bbox": [ + 215, + 688, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/096e16fa5cd5b34df656829c9902a6b655d244d3917fca70281b627416c03760.jpg", + "table_caption": [ + "Table 3: Class-wise video highlight detection results (mAP) on YouTube Highlights dataset. $\\bigcirc$ , $\\bigcirc$ , and $\\bigcirc$ denote image, video, and audio backbones, respectively." + ], + "table_footnote": [], + "table_body": "
MethodDogGym.Par.Ska.Ski.Sur.Avg.
○ RRAE [58]49.035.050.025.022.049.038.3
○ GIFs [12]30.833.554.055.432.854.146.4
○ LSVM [47]60.041.061.062.036.061.053.6
○ LIM-S [54]57.941.767.057.848.665.156.4
○ SL-Module [55]70.853.277.272.566.176.269.3
○ PLD [53]74.970.277.957.570.779.073.0
○ MINI-Net [14]58.261.770.272.258.765.164.4
○ TCG [59]55.462.770.969.160.159.863.0
○ Joint-VA [3]64.571.980.862.073.278.371.8
○ CO-AV [25]60.966.089.074.169.081.174.7
○ UMT [31]65.975.281.671.872.382.774.9
○ UniVTG [26]71.876.573.973.373.282.275.2
○ R²-Tuning75.673.573.074.674.884.876.1
", + "bbox": [ + 300, + 184, + 702, + 347 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/4e17f519a0e140d7741faebe1e11ddee277471f2b2579cfb2037cad02af88ea9.jpg", + "table_caption": [ + "Table 4: Class-wise video summarization results (Top-5 mAP) on TVSum dataset. $\\odot$ , $\\odot$ , and $\\odot$ denote image, video, and audio backbones, respectively. Our method does not require any additional video/audio features." + ], + "table_footnote": [], + "table_body": "
MethodVTVUGAMSPKPRFMBKBTDSAvg.
○ sLSTM [63]41.146.246.347.744.846.145.240.647.145.545.1
○ SG [33]42.347.247.548.945.647.346.441.748.346.646.2
○ LIM-S [54]55.942.961.254.060.447.543.266.369.162.656.3
○ Trailer [51]61.354.665.760.859.170.158.264.765.668.162.8
○ SL-Module [55]86.568.774.986.279.063.258.972.678.964.073.3
○ PLD [53]84.580.970.372.576.487.271.974.074.479.177.1
○ MINI-Net [14]80.668.378.281.878.165.857.875.080.265.573.2
○ TCG [59]85.071.481.978.680.275.571.677.378.668.176.8
○ Joint-VA [3]83.757.378.586.180.169.270.073.097.467.576.3
○ CO-AV [25]90.872.884.685.078.378.072.877.189.572.380.1
○ UMT [31]87.581.588.278.881.487.076.086.984.479.683.1
○ UniVTG [26]83.985.189.080.184.681.470.991.773.569.381.0
○ R²-Tuning85.085.991.081.788.887.478.189.290.374.785.2
", + "bbox": [ + 264, + 414, + 738, + 577 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 603, + 455, + 619 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In all experiments, we adopt the visual and text encoders of CLIP ViT-B/32 [41] as our backbones. The whole CLIP model is frozen during training, i.e., only the parameters in $\\mathbb{R}^2$ Block, temporal feature pyramid, prediction heads are learnable. Without further specification, we set $K = 4$ , i.e., attaching $\\mathbb{R}^2$ Block to the last 4 layers of CLIP encoders. In $\\mathbb{R}^2$ Block, the hidden size is set to 256. Only one transformer layer with post-norm style and 8 attention heads is used for temporal modeling. The number of layers for the temporal feature pyramid is set to 4 for QVHighlights, Ego4D-NLQ, Charades-STA, and TACoS, while 1 for other datasets. Please refer to the supplementary material for details.", + "bbox": [ + 212, + 628, + 787, + 765 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Comparison with State-of-the-Arts", + "text_level": 1, + "bbox": [ + 214, + 785, + 550, + 800 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Joint Moment Retrieval and Highlight Detection We first evaluate our method on QVHighlights [22] test split, the only dataset that supports both", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Reversed Recurrent Tuning", + "bbox": [ + 545, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1965b07216fe806ed6b3dc139235ce5e0991176c050cdbcb7668eead2bfb9fa6.jpg", + "image_caption": [ + "QUERY: A man in white is browsing through clothes at a department store." + ], + "image_footnote": [], + "bbox": [ + 217, + 152, + 785, + 243 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/af00914a1c1c11765454382834f2ba07a7ad8ee70f499d2e86dfc101f4d57220.jpg", + "image_caption": [ + "QUERY: Man stands on playground balance bridge.", + "Fig.6: Visualization of joint moment retrieval and highlight detection results on QVHighlights val split. We compare our method with UniVTG [26], which is a strong baseline leveraging both additional temporal backbone and large-scale post pretraining. $\\mathbb{R}^2$ -Tuning can accurately regress the boundaries of moments and predict highlight saliency scores through its novel feature refinement design." + ], + "image_footnote": [], + "bbox": [ + 217, + 263, + 785, + 352 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "moment retrieval and highlight detection. The results are shown in Table 1. The first group of methods utilizes more than one feature extractor. The second group is that with large-scale post pre-training (defined as the extra pre-training stage between loading backbone weights and training on downstream datasets.). Without any additional backbone and extra pre-training, $\\mathbf{R}^2$ -Tuning achieves start-of-the-art performance compared with all previous methods with the fewest learnable parameters. Figure 6 visualizes the model predictions.", + "bbox": [ + 212, + 465, + 787, + 571 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Moment Retrieval We then evaluate our model on moment retrieval task in egocentric [10] and in-door [9, 43] domains. The results are shown in Table 2. We follow [26] and compare only with the methods using CLIP + SlowFast features. R²-Tuning still works better than all baselines with extra features. We also observe that R²-Tuning shows its significance on high-quality retrievals (R1@0.7), which requires accurate temporal modeling. This benefits from the fine-grained temporal modeling ability of reversed and recurrent designs.", + "bbox": [ + 212, + 590, + 787, + 699 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Highlight Detection & Video Summarization The performances of highlight detection on YouTube Highlights [47] and extractive video summarization on TVSum [45] are reported in Table 3 and Table 4, respectively. Following previous works [26, 31], we trained the model separately on each domain. The first group of methods is based on video backbones, which naturally have basic temporal modeling abilities. The second group of methods is enhanced by extra features such as image and audio. $\\mathbf{R}^2$ -Tuning can still reach the best performance when trained on small-scale data.", + "bbox": [ + 212, + 718, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b098f12db1afd9c6126afa7b3b43e782d7d214823744117017c0329b18d020d3.jpg", + "image_caption": [ + "(a) Naive Fusion" + ], + "image_footnote": [], + "bbox": [ + 215, + 154, + 352, + 253 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0e6de240e9b71af89bdea3aed47b4cc0fef6d4015cc7bd8e90467e9068ad9801.jpg", + "image_caption": [ + "Fig. 7: Different feature refinement strategies for MR on QVHighlights val split. Single means utilizing the $K$ -th layer only, while Mean denotes averaging the features from last $K$ layers. Forward and Reversed indicate different feature refinement directions." + ], + "image_footnote": [], + "bbox": [ + 361, + 155, + 496, + 253 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/666ac9bc80ecce0d6fce7c0a478eaba4331c40b70d2266f75bf76a9716c928da.jpg", + "image_caption": [ + "(b) Recurrent Refinement" + ], + "image_footnote": [], + "bbox": [ + 506, + 154, + 643, + 253 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/106051b73f23d014df24ae663665f4a8411385db7f49bd4c94d0aa60f2d780a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 154, + 787, + 253 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e40999bd62e249bae10d2ea3518aca5a5f9203fd71f7617dbdd3a0a57f4a0bc4.jpg", + "image_caption": [ + "Fig. 8: Visualization of the attention maps for query-modulated spatial pooling. Diverse regions are learned for different queries." + ], + "image_footnote": [], + "bbox": [ + 222, + 318, + 547, + 465 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/fc12baa482f5c650a0433afd55e3c1eeee525a401b4429e6ef262c09407d0a8b.jpg", + "image_caption": [ + "Fig. 9: Cosine similarities of video (x-axis) and text (y-axis) features among $7\\sim 12$ CLIP layers." + ], + "image_footnote": [], + "bbox": [ + 588, + 318, + 758, + 465 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Detailed Analysis", + "text_level": 1, + "bbox": [ + 215, + 546, + 408, + 561 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Multi-Granularity Features To investigate the significance of granularity flexibility, we set up a simple baseline that only attaches $\\mathbf{R}^2$ Block to one layer of CLIP encoders. Then, we make use of the multi-layer information by averaging the features from the last $K$ layers. These two variants are compared in Figure 7 (a). It can be observed that for the single-layer setting, higher-level features are more discriminative than lower-level ones. However, even a naive fusion strategy for the multi-layer information brings significant performance gains.", + "bbox": [ + 212, + 570, + 787, + 676 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Reversed Recurrent Feature Refinement We then further justify our design by comparing two refinement directions (i.e., forward and reversed) in Figure 7 (b). Note that when $K = 1$ , all four designs in Figure 7 are of the same architecture. When utilizing more layers, both forward and reversed refinement benefit from more information, while the reversed scheme statically performs better, as the multi-granularity features shall be refined from coarse to fine.", + "bbox": [ + 212, + 689, + 787, + 781 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Query-Modulated Spatial Pooling We visualize the attention maps learned by query-modulated spatial pooling in Figure 8. A query generally has multiple keywords that provide discriminative information to localize moments. The", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Reversed Recurrent Tuning", + "bbox": [ + 547, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/d057a0a670b4e97514c46f248cc81a7d1df26bee8ae95a162919bbc989856bb6.jpg", + "table_caption": [ + "Table 5: Effectiveness justification of granularity calibration on QVHighlights val split." + ], + "table_footnote": [], + "table_body": "
L VideoL layerMRHD
R1@0.5R1@0.7mAPmAPHIT@1
64.4848.6044.0137.9462.67
✓67.6851.6146.7439.8165.16
✓64.7148.8444.6038.9163.35
✓✓68.7152.0647.5940.5964.32
", + "bbox": [ + 215, + 184, + 511, + 265 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/5ad0783f38989c6f78b679e01d873e40e8c2b4962bf3846627d31d4f31387c9e.jpg", + "table_caption": [ + "Table 6: MR mAP with different query lengths on QVHighlights val split." + ], + "table_footnote": [], + "table_body": "
Method#Words
1-1011-2021-3031-40≥41
QD-DETR [35]40.3042.3229.010.1026.67
EaTR [16]38.9439.5424.5812.0341.43
UniVTG [26]32.3832.6520.880.0031.11
R2-Tuning47.0849.1328.7967.2472.38
", + "bbox": [ + 521, + 184, + 785, + 266 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "token-wise attention in Eq. 4 can guide the model to focus on multiple regions simultaneously, as can be seen in the diverse but meaningful patterns.", + "bbox": [ + 212, + 294, + 782, + 325 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Granularity Calibration To verify the effectiveness of granularity calibration, we also visualize the cosine distances of all the visual-query pairs from $7 \\sim 12$ layers in Fig. 9. Before calibration, the visual-text features are not well-aligned. Adding the video-level constraint makes higher-level features aligned with each other (b), while adding layer-wise constraint makes the model distill diverse information across different layers (c). The diagonal in (d) verifies that both calibration constraints can maximize the mutual similarities of paired visual and text features. Some quantitative results are presented in Table 5. Both video-level and layer-wise contrastive can help align the multi-granularity visual & query semantics, while their combination can further boost the performances. We also group the queries in QVHighlights into 5 bins with different lengths (acting as multi-granularities) and compute the means of MR mAP in Tab. 6. Although most training queries are coarse-grained ( $\\leqslant 30$ words), R²-Tuning can easily generalize to fine-grained queries ( $\\geqslant 31$ words), surpassing previous methods.", + "bbox": [ + 212, + 342, + 787, + 554 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 579, + 359, + 595 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper introduces R²-Tuning, a parameter- and memory-efficient transfer learning framework for video temporal grounding. It learns a lightweight side-adapter (R² Block) that is recurrently attached to the last few layers of a frozen CLIP to adaptively pool spatial details and refine temporal correlations. Two contrastive constraints (video-level and layer-wise) are utilized to calibrate the granularities of CLIP visual and text encoders. Experiments across three VTG tasks on six public benchmarks demonstrate the significance and effectiveness of the proposed scheme. We hope that the proposed framework sparks further research on efficient image-to-video transfer learning for untrimmed videos.", + "bbox": [ + 212, + 613, + 787, + 750 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 215, + 775, + 393, + 792 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work was supported in part by Hong Kong Research Grants Council GRF-15229423, US NIH grant R01HD104969, and NSF award IIS-2239688.", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Anne Hendricks, L., Wang, O., Shechtman, E., Sivic, J., Darrell, T., Russell, B.: Localizing moments in video with natural language. In: ICCV. pp. 5803-5812 (2017)", + "2. Apostolidis, E., Adamantidou, E., Metsai, A.I., Mezaris, V., Patras, I.: Video summarization using deep neural networks: A survey. Proceedings of the IEEE 109(11), 1838-1863 (2021)", + "3. Badamdorj, T., Rochan, M., Wang, Y., Cheng, L.: Joint visual and audio learning for video highlight detection. In: ICCV. pp. 8127-8137 (2021)", + "4. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: A clip-hitchhiker's guide to long video retrieval. Tech. Rep. arXiv:2205.08508 (2022)", + "5. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: CVPR. pp. 6299-6308 (2017)", + "6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2020)", + "7. Escorcia, V., Soldan, M., Sivic, J., Ghanem, B., Russell, B.: Temporal localization of moments in video collections with natural language. Tech. Rep. arXiv:1907.12763 (2019)", + "8. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: ICCV. pp. 6202-6211 (2019)", + "9. Gao, J., Sun, C., Yang, Z., Nevatia, R.: Tall: Temporal activity localization via language query. In: ICCV. pp. 5267-5275 (2017)", + "0. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: CVPR. pp. 18995-19012 (2022)", + "1. Gygli, M., Grabner, H., Riemenschneider, H., Van Gool, L.: Creating summaries from user videos. In: ECCV. pp. 505-520 (2014)", + "2. Gygli, M., Song, Y., Cao, L.: Video2gif: Automatic generation of animated gifs from video. In: CVPR. pp. 1001-1009 (2016)", + "3. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)", + "4. Hong, F.T., Huang, X., Li, W.H., Zheng, W.S.: Mini-net: Multiple instance ranking network for video highlight detection. In: ECCV. pp. 345-360 (2020)", + "5. Huang, S., Gong, B., Pan, Y., Jiang, J., Lv, Y., Li, Y., Wang, D.: Vop: Text-video co-operative prompt tuning for cross-modal retrieval. In: CVPR. pp. 6565-6574 (2023)", + "6. Jang, J., Park, J., Kim, J., Kwon, H., Sohn, K.: Knowing where to focus: Event-aware transformer for video grounding. In: ICCV. pp. 13846-13856 (2023)", + "7. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: ECCV. pp. 709-727 (2022)", + "8. Jiang, R., Liu, L., Chen, C.: Clip-count: Towards text-guided zero-shot object counting. In: ACM MM (2023)", + "9. Ju, C., Han, T., Zheng, K., Zhang, Y., Xie, W.: Prompting visual-language models for efficient video understanding. In: ECCV. pp. 105-124 (2022)", + "20. Kong, Q., Cao, Y., Iqbal, T., Wang, Y., Wang, W., Plumbley, M.D.: Panns: Large-scale pretrained audio neural networks for audio pattern recognition. IEEE/ACM Transactions on Audio, Speech, and Language Processing 28, 2880-2894 (2020)", + "21. Larsson, G., Maire, M., Shakhnarovich, G.: Fractalnet: Ultra-deep neural networks without residuals. Tech. Rep. arXiv:1605.07648 (2016)" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Reversed Recurrent Tuning", + "bbox": [ + 547, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "22. Lei, J., Berg, T.L., Bansal, M.: Qvhighlights: Detecting moments and highlights in videos via natural language queries. In: NeurIPS (2021)", + "23. Lei, J., Yu, L., Berg, T.L., Bansal, M.: Tv: A large-scale dataset for video-subtitle moment retrieval. In: ECCV. pp. 447-463 (2020)", + "24. Li, P., Xie, C.W., Xie, H., Zhao, L., Zhang, L., Zheng, Y., Zhao, D., Zhang, Y.: Momentdiff: Generative video moment retrieval from random to real. Tech. Rep. arXiv:2307.02869 (2023)", + "25. Li, S., Zhang, F., Yang, K., Liu, L., Liu, S., Hou, J., Yi, S.: Probing visual-audio representation for video highlight detection via hard-pairs guided contrastive learning. Tech. Rep. arXiv:2206.10157 (2022)", + "26. Lin, K.Q., Zhang, P., Chen, J., Pramanick, S., Gao, D., Wang, A.J., Yan, R., Shou, M.Z.: Univtg: Towards unified video-language temporal grounding. In: CVPR. pp. 2794-2804 (2023)", + "27. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: ICCV. pp. 2980-2988 (2017)", + "28. Lin, Z., Geng, S., Zhang, R., Gao, P., de Melo, G., Wang, X., Dai, J., Qiao, Y., Li, H.: Frozen clip models are efficient video learners. In: ECCV. pp. 388-404 (2022)", + "29. Liu, L., Yu, B.X., Chang, J., Tian, Q., Chen, C.W.: Prompt-matched semantic segmentation. Tech. Rep. arXiv:2208.10159 (2022)", + "30. Liu, W., Mei, T., Zhang, Y., Che, C., Luo, J.: Multi-task deep visual-semantic embedding for video thumbnail selection. In: CVPR. pp. 3707-3715 (2015)", + "31. Liu, Y., Li, S., Wu, Y., Chen, C.W., Shan, Y., Qie, X.: Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In: CVPR. pp. 3042-3051 (2022)", + "32. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: Clip4clip: An empirical study of clip for end to end video clip retrieval. Neurocomputing 508, 293-304 (2022)", + "33. Mahasseni, B., Lam, M., Todorovic, S.: Unsupervised video summarization with adversarial LSTM networks. In: CVPR. pp. 202-211 (2017)", + "34. Moon, W., Hyun, S., Lee, S., Heo, J.P.: Correlation-guided query-dependency calibration in video representation learning for temporal grounding. Tech. Rep. arXiv:2311.08835 (2023)", + "35. Moon, W., Hyun, S., Park, S., Park, D., Heo, J.P.: Query-dependent video representation for moment retrieval and highlight detection. In: CVPR. pp. 23023-23033 (2023)", + "36. Nan, G., Qiao, R., Xiao, Y., Liu, J., Leng, S., Zhang, H., Lu, W.: Interventional video grounding with dual contrastive learning. In: CVPR. pp. 2765-2775 (2021)", + "37. Ni, B., Peng, H., Chen, M., Zhang, S., Meng, G., Fu, J., Xiang, S., Ling, H.: Expanding language-image pretrained models for general video recognition. In: ECCV. pp. 1-18 (2022)", + "38. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. In: NeurIPS (2018)", + "39. Pan, J., Lin, Z., Zhu, X., Shao, J., Li, H.: St-adapter: Parameter-efficient image-to-video transfer learning. In: NeurIPS. pp. 26462-26477 (2022)", + "40. Qing, Z., Zhang, S., Huang, Z., Zhang, Y., Gao, C., Zhao, D., Sang, N.: Disentangling spatial and temporal learning for efficient image-to-video transfer learning. In: CVPR. pp. 13934-13944 (2023)", + "41. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "42. Rasheed, H., Khattak, M.U., Maaz, M., Khan, S., Khan, F.S.: Fine-tuned clip models are efficient video learners. In: CVPR. pp. 6545-6554 (2023)", + "43. Regneri, M., Rohrbach, M., Wetzel, D., Thater, S., Schiele, B., Pinkal, M.: Grounding action descriptions in videos. Transactions of the Association for Computational Linguistics 1, 25-36 (2013)", + "44. Song, Y., Redi, M., Vallmitjana, J., Jeimes, A.: To click or not to click: Automatic selection of beautiful thumbnails from videos. In: CIKM. pp. 659-668 (2016)", + "45. Song, Y., Vallmitjana, J., Stent, A., Jaimes, A.: Tvsum: Summarizing web videos using titles. In: CVPR. pp. 5179-5187 (2015)", + "46. Sun, H., Zhou, M., Chen, W., Xie, W.: Tr-detr: Task-reciprocal transformer for joint moment retrieval and highlight detection. In: AAAI (2024)", + "47. Sun, M., Farhadi, A., Seitz, S.: Ranking domain-specific highlights by analyzing edited videos. In: ECCV. pp. 787-802 (2014)", + "48. Sung, Y.L., Cho, J., Bansal, M.: Lst: Ladder side-tuning for parameter and memory efficient transfer learning. In: NeurIPS. pp. 12991-13005 (2022)", + "49. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. Tech. Rep. arXiv:2302.13971 (2023)", + "50. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS. pp. 5998-6008 (2017)", + "51. Wang, L., Liu, D., Puri, R., Metaxas, D.N.: Learning trailer moments in full-length movies. In: ECCV. pp. 300-316 (2020)", + "52. Wang, Z., Wang, L., Wu, T., Li, T., Wu, G.: Negative sample matters: A renaissance of metric learning for temporal grounding. In: AAAI. pp. 2613-2623 (2022)", + "53. Wei, F., Wang, B., Ge, T., Jiang, Y., Li, W., Duan, L.: Learning pixel-level distinctions for video highlight detection. In: CVPR. pp. 3073-3082 (2022)", + "54. Xiong, B., Kalantidis, Y., Ghadiyaram, D., Grauman, K.: Less is more: Learning highlight detection from video duration. In: CVPR. pp. 1258-1267 (2019)", + "55. Xu, M., Wang, H., Ni, B., Zhu, R., Sun, Z., Wang, C.: Cross-category video highlight detection via set-based learning. In: ICCV. pp. 7970-7979 (2021)", + "56. Xu, Y., Sun, Y., Li, Y., Shi, Y., Zhu, X., Du, S.: Mh-detr: Video moment and highlight detection with cross-modal transformer. Tech. Rep. arXiv:2305.00355 (2023)", + "57. Yan, S., Xiong, X., Nagrani, A., Arnab, A., Wang, Z., Ge, W., Ross, D., Schmid, C.: Unloc: A unified framework for video localization tasks. In: ICCV. pp. 13623-13633 (2023)", + "58. Yang, H., Wang, B., Lin, S., Wipf, D., Guo, M., Guo, B.: Unsupervised extraction of video highlights via robust recurrent auto-encoders. In: ICCV. pp. 4633-4641 (2015)", + "59. Ye, Q., Shen, X., Gao, Y., Wang, Z., Bi, Q., Li, P., Yang, G.: Temporal cue guided video highlight detection with low-rank audio-visual fusion. In: ICCV. pp. 7950-7959 (2021)", + "60. Yuan, T., Zhang, X., Liu, K., Liu, B., Jin, J., Jiao, Z.: Ucf-crime annotation: A benchmark for surveillance video-and-language understanding. Tech. Rep. arXiv:2309.13925 (2023)", + "61. Yuan, Y., Ma, L., Wang, J., Liu, W., Zhu, W.: Semantic conditioned dynamic modulation for temporal sentence grounding in videos. NeurIPS 32 (2019)", + "62. Zhang, H., Sun, A., Jing, W., Zhou, J.T.: Span-based localizing network for natural language video localization. Tech. Rep. arXiv:2004.13931 (2020)", + "63. Zhang, K., Chao, W.L., Sha, F., Grauman, K.: Video summarization with long short-term memory. In: ECCV. pp. 766-782. Springer (2016)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Reversed Recurrent Tuning", + "bbox": [ + 547, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "64. Zhang, S., Peng, H., Fu, J., Luo, J.: Learning 2d temporal adjacent networks for moment localization with natural language. In: AAAI. pp. 12870-12877 (2020)", + "65. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: CVPR. pp. 16816-16825 (2022)", + "66. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 229 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/R^2-Tuning_ Efficient Image-to-Video Transfer Learning for Video Temporal Grounding/148e844e-b528-47e5-9db4-f50b0b9b4b9c_model.json b/2024/R^2-Tuning_ Efficient Image-to-Video Transfer Learning for Video Temporal Grounding/148e844e-b528-47e5-9db4-f50b0b9b4b9c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d92287e13f71ce8425e0639c669c455d71aad5eb --- /dev/null +++ b/2024/R^2-Tuning_ Efficient Image-to-Video Transfer Learning for Video Temporal Grounding/148e844e-b528-47e5-9db4-f50b0b9b4b9c_model.json @@ -0,0 +1,2667 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.245, + 0.143, + 0.761, + 0.19 + ], + "angle": 0, + "content": "\\(\\mathbf{R}^2\\)-Tuning: Efficient Image-to-Video Transfer Learning for Video Temporal Grounding" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.215, + 0.776, + 0.248 + ], + "angle": 0, + "content": "Ye Liu\\(^{1,3*}\\), Jixuan He\\(^{2,3}\\), Wanhua Li\\(^{3\\dagger}\\), Junsik Kim\\(^{3}\\), Donglai Wei\\(^{4}\\), Hanspeter Pfister\\(^{3}\\), and Chang Wen Chen\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.257, + 0.741, + 0.331 + ], + "angle": 0, + "content": "1 The Hong Kong Polytechnic University 2 Tsinghua University 3 Harvard University 4 Boston College coco.ye.liu@connect.polyu.hk hejixuan2000@gmail.com wanhua@seas.harvard.edu mibastro@gmail.com donglai.wei@bc.edu pfister@seas.harvard.edu changwen.chen@polyu.edu.hk" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.371, + 0.744, + 0.678 + ], + "angle": 0, + "content": "Abstract. Video temporal grounding (VTG) is a fine-grained video understanding problem that aims to ground relevant clips in untrimmed videos given natural language queries. Most existing VTG models are built upon frame-wise final-layer CLIP features, aided by additional temporal backbones (e.g., SlowFast) with sophisticated temporal reasoning mechanisms. In this work, we claim that CLIP itself already shows great potential for fine-grained spatial-temporal modeling, as each layer offers distinct yet useful information under different granularity levels. Motivated by this, we propose Reversed Recurrent Tuning (R²-Tuning), a parameter- and memory-efficient transfer learning framework for video temporal grounding. Our method learns a lightweight R² Block containing only 1.5% of the total parameters to perform progressive spatial-temporal modeling. Starting from the last layer of CLIP, R² Block recurrently aggregates spatial features from earlier layers, then refines temporal correlation conditioning on the given query, resulting in a coarse-to-fine scheme. R²-Tuning achieves state-of-the-art performance across three VTG tasks (i.e., moment retrieval, highlight detection, and video summarization) on six public benchmarks (i.e., QVHighlights, CharadesSTA, Ego4D-NLQ, TACoS, YouTube Highlights, and TVSum) even without the additional backbone, demonstrating the significance and effectiveness of the proposed scheme. Our code is available at https://github.com/yeliudev/R2-Tuning." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.689, + 0.721, + 0.705 + ], + "angle": 0, + "content": "Keywords: Video Temporal Grounding \\(\\cdot\\) Transfer Learning \\(\\cdot\\) CLIP" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.733, + 0.378, + 0.749 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.768, + 0.788, + 0.816 + ], + "angle": 0, + "content": "Video is becoming the major content media in our daily lives. The variety of video domains has extended beyond daily activities [5,22] but broader scenes such as egocentric [10] and surveillance [60]. They maintain large information" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.825, + 0.642, + 0.842 + ], + "angle": 0, + "content": "* Work done at Harvard University. † Corresponding authors." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.278, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.236, + 0.278, + 0.243 + ], + "angle": 0, + "content": "Untrimmed Video" + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.146, + 0.526, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.254, + 0.529, + 0.311 + ], + "angle": 0, + "content": "Fig. 1: Video temporal grounding (VTG) contains three video-language understanding problems, i.e., moment retrieval (MR), highlight detection (HD), and video summarization (VS)." + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.144, + 0.785, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.54, + 0.254, + 0.788, + 0.311 + ], + "angle": 0, + "content": "Fig. 2: Moment retrieval mAP with different backbones on QVHighlights val split. CLIP's potential of temporal modeling was not fully exploited." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.341, + 0.784, + 0.371 + ], + "angle": 0, + "content": "capacities within multi-granularities, and thus can convey both high-level context and low-level details effectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.785, + 0.508 + ], + "angle": 0, + "content": "In the content production and consumption industry, such a flourishing is accompanied by the increasing demand for browsing untrimmed videos efficiently with different user interests. This derives the topic of video temporal grounding (VTG) [26], which is a fine-grained video-language understanding problem aiming to ground video clips conditioning on natural language queries. As shown in Figure 1, VTG can be disentangled into three tasks based on different output formats, i.e., moment retrieval (MR) [9, 10, 22, 31] for regressing start-end timestamps, highlight detection (HD) [22, 47] for predicting frame-level saliency curves, and video summarization (VS) [2, 11, 45] for classifying disjoint shots." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.786, + 0.629 + ], + "angle": 0, + "content": "Encouraged by the recent success in adopting vision-language models (VLMs) for video understanding [4,15,19,28,32,37], most existing VTG methods [16,22, 31,34,35,46,56] are built upon frame-wise final-layer features from CLIP [41]. However, due to CLIP's misaligned pre-training objective (image-text contrast), these methods fail to capture temporal correlations well. As a feasible compromise, an additional backbone (e.g., SlowFast [8]) is incorporated to complement the temporal information, followed by carefully designed modules such as text-guided queries [31], dynamic anchors [35], and event reasoning [16]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.787, + 0.841 + ], + "angle": 0, + "content": "We refer to the paradigm above as a post-processing scheme (shown in Figure. 3 (a)), whereas two natural drawbacks exist due to the sub-optimal design. First, leveraging two backbones with similar capabilities is unintuitive and inefficient during inference. A single model with both vision-text alignment and spatial-temporal modeling abilities is more preferred. Second, queries for VTG could be of different granularities from coarse (e.g., the family is traveling) to fine (e.g., when did I take the golf club from the man with white hair). Leveraging only frame-wise & final-layer features is not granularity flexible, as it would focus more on high-level scene transitions while overlooking low-level details. Preliminary experiments in Figure 2 also demonstrate that the potential of spatial-temporal modeling for CLIP is not fully exploited by existing methods. Some recent works [15,19,57] tried to tackle the first problem by fine-tuning (part of) the CLIP encoders (Figure 3 (b) and (c)), but they are all inefficient in terms of data, learnable parameters, or memory." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.548, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reversed Recurrent Tuning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.145, + 0.359, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.145, + 0.493, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.145, + 0.629, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.146, + 0.785, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.296, + 0.784, + 0.325 + ], + "angle": 0, + "content": "Fig. 3: Different architectural designs for CLIP-based image-to-video transfer learning. The gray rectangle in (d) denotes the progressively refined spatial-temporal features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.357, + 0.788, + 0.613 + ], + "angle": 0, + "content": "This paper intends to answer the research question: how to efficiently transfer an image-language foundation model to video temporal grounding? We consider the above question by exploring two aspects: efficiency and granularity flexibility. To address these issues, a novel image-to-video transfer learning framework called Reversed Recurrent Tuning (R²-Tuning) is proposed for fine-grained understanding on untrimmed videos. Our insight is that multi-layer CLIP features offer distinct yet useful information, while their integration should be tailored to the downstream task. As illustrated in Figure 3 (d), based on a frozen CLIP [41], our method learns a side-block (R² Block) containing only \\(1.5\\%\\) of the total parameters to perform spatial-temporal modeling. R² Block is recurrently attached to the last few layers of CLIP encoders from back to front, performing query-modulated spatial pooling and recurrent temporal refinement from coarse to fine. Aside from the new architecture, we also introduce video-level and layer-wise constraints to calibrate the granularities of visual and text encoders. During training, gradient flows do not pass through the CLIP encoders, thus our scheme is both parameter- and memory-efficient. It is also granularity-flexible as R² Block can adaptively control the spatial pooling strategy conditioning on queries." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We conduct extensive experiments across three VTG tasks on six public benchmarks, including QVHighlights [22], Ego4D-NLQ [10], Charades-STA [9], TACoS [9], YouTube Highlights [47], and TVSum [45]. Without bells and whistles, \\(\\mathbb{R}^2\\)-Tuning achieves more than 3 MR mAP gain on QVHighlights test split compared with \\(4\\times\\) heavier counterparts [34] with additional temporal backbones and carefully designed post-processing modules. Further analysis also shows that our method can better handle multi-granularity information. Overall, our contributions are summarized as: (1) We introduced \\(\\mathbb{R}^2\\)-Tuning, a novel image-to-video transfer learning framework tailored for video temporal grounding. (2) We designed two effective strategies, i.e., query-modulated spatial pooling and recurrent temporal refinement, to model spatial-temporal information from coarse to fine. (3) To calibrate the granularities of CLIP visual and text encoders, we further introduce video-level and layer-wise contrastive constraints to distill distinct information from each layer. (4) Extensive experiments across three tasks on six benchmarks demonstrate the significance and effectiveness of \\(\\mathbb{R}^2\\)-Tuning." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.144, + 0.387, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.181, + 0.786, + 0.378 + ], + "angle": 0, + "content": "CLIP for Video Understanding With powerful transfer learning abilities, CLIP [41] has been widely used for many image-language tasks [18, 29], and how to extend them to video understanding is an emerging and crucial research topic. Early attempts have been made to transfer CLIP to trimmed videos for action recognition [28,37,42] and video-text retrieval [4,15,19,32]. Specifically, Efficient-Prompt [19] introduces learnable prompt tokens to the text encoder, and ViFi-CLIP [42] further adopts learnable prompts at the visual encoder. Other works also tried to learn external modules designed for temporal modeling. EVL [28] proposes a temporal decoder in parallel with the main network. CLIP4Clip [32] investigated multiple temporal fusion strategies for video frames. Nevertheless, how to effectively transfer CLIP to untrimmed videos remains unexplored. To our best knowledge, \\(\\mathbb{R}^2\\)-Tuning is the first solution for fine-grained video temporal grounding via memory-efficient transfer learning." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.397, + 0.786, + 0.654 + ], + "angle": 0, + "content": "Video Temporal Grounding Video temporal grounding (VTG) [26] is a bundle of video understanding problems including moment retrieval (MR) [9,16,22, 31,35,36,52,56,57,61], highlight detection (HD) [3,14,47,55], and video summarization (VS) [11,45]. Although these tasks are closely related, they have not been jointly studied until recently. QVHighlights [22] is the first dataset that supports both MR and HD, breaking the moment distribution bias in previous MR datasets [9,43]. The authors also proposed Moment-DETR, a strong baseline method for MR/HD. Most following works were built upon the Moment-DETR framework with specially designed modules. Notably, UMT [31] tackles flexible MR and HD through a unified multi-modal architecture. QD-DETR [35] benefits from query-dependent video representations and dynamic anchors. EaTR [16] incorporates moment reasoning to provide reliable referential search areas for moment queries. Nevertheless, their design space is still within pre-extracted CLIP [41] and SlowFast [8] features. A recent work [57] revealed the possibility of fine-tuning CLIP for MR, while their solution (post pre-training) is extremely resource-consuming. Compared with existing works, \\(\\mathbb{R}^2\\)-Tuning can effectively learn strong temporal modeling abilities without any pre-training on videos." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.673, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Parameter-Efficient Transfer Learning Foundation models [41, 49] have achieved remarkable successes in vision and language understanding. Efforts have been made to efficiently transfer knowledge from these models to new scenarios. One line of research is to conduct context-based tuning (e.g., prompt-tuning [17, 65, 66]), which does not modify the model architectures but incorporates learnable embeddings in the inputs. Another paradigm is introducing an extra lightweight adapter to the original model, and keeping the rest of the parameters frozen [39]. Recent works also consider memory-efficient transfer learning (METL) that introduces parameters whose gradient flows do not go through the main model [28, 40, 48]. \\(\\mathbf{R}^2\\)-Tuning is also a METL framework, where we adopt a novel recurrent tuning strategy to ensure parameter- and memory-efficient." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.547, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Reversed Recurrent Tuning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.287, + 0.788, + 0.358 + ], + "angle": 0, + "content": "Fig. 4: Overall architecture of our framework. The input video and query are first encoded by frozen CLIP [41] encoders. Their multi-layer outputs are then recurrently fused and refined by a learnable \\(\\mathbb{R}^2\\) Block to construct spatial-temporal representations \\(h\\), which would be scaled up/down to construct a temporal feature pyramid, followed by three heads for MR, HD, and VS, respectively." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.387, + 0.381, + 0.404 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.419, + 0.439, + 0.433 + ], + "angle": 0, + "content": "3.1 Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.787, + 0.492 + ], + "angle": 0, + "content": "Given a video \\( V = \\{v_{i}\\}_{i=1}^{T} \\) and a natural language query \\( Q = \\{q_{i}\\}_{i=1}^{L} \\), where \\( T \\) and \\( L \\) are the numbers of video frames and text tokens, VTG aims to densely predict a set of labels \\( (b_{i}, s_{i}, f_{i}) \\) for each frame, defined as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.504, + 0.787, + 0.564 + ], + "angle": 0, + "content": "- Moment Retrieval is to find the most relevant moments (i.e. sets of consecutive frames) in \\( V \\) according to \\( Q \\), so that \\( b_{i} = [b_{i}^{s}, b_{i}^{e}] \\in \\mathbb{R}^{2} \\) represents the temporal displacements from frame \\( v_{i} \\) to the start and end timestamps of the nearest target moment." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.565, + 0.787, + 0.61 + ], + "angle": 0, + "content": "- Highlight Detection requires estimating frame-level relevancies between \\( V \\) and \\( Q \\), thus \\( s_i \\in [0,1] \\) is a continuous saliency score denoting to what extent frame \\( v_i \\) is semantically aligned with \\( Q \\)." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.61, + 0.787, + 0.656 + ], + "angle": 0, + "content": "- Video Summarization aims to select a subset of video frames according to \\( Q \\) to form a concise summary, therefore, \\( f_{i} \\in \\{0,1\\} \\) is a binary score indicating whether frame \\( v_{i} \\) belongs to the summary." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.504, + 0.787, + 0.656 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.678, + 0.341, + 0.692 + ], + "angle": 0, + "content": "3.2 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.787, + 0.795 + ], + "angle": 0, + "content": "Figure 4 shows an overview of the proposed framework. Our model derives from a pre-trained and frozen CLIP [41] with ViT [6] backbone, which has a two-stream architecture for spatial and textual encoding, respectively. A learnable \\(\\mathbb{R}^2\\) block is iteratively attached to the last \\(K\\) encoder layers to refine the temporal correlation. The resulting features are then scaled up/down to build a feature pyramid, followed by three heads to decode the task-specific predictions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.843 + ], + "angle": 0, + "content": "Specifically, the input video \\( V \\) and query \\( Q \\) are first tokenized into frame patches and word tokens, then sent into the visual encoder \\( E_{v} \\) and query encoder \\( E_{q} \\). The encoded visual and textual features can be denoted as \\( e_v \\in \\)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.345, + 0.147, + 0.661, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.317, + 0.788, + 0.36 + ], + "angle": 0, + "content": "Fig. 5: Detailed architecture of the \\(\\mathbb{R}^2\\) Block. It can be split into two parts: a) query-modulated spatial pooling, and b) recurrent temporal refinement. Note that the [CLS] tokens of visual features are omitted for clarity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.789, + 0.495 + ], + "angle": 0, + "content": "\\(\\mathbb{R}^{B \\times N \\times T \\times (P + 1) \\times D_v}\\) and \\(e_q \\in \\mathbb{R}^{B \\times N \\times L \\times D_q}\\), where \\(B, N, T, P, L, D_v\\) and \\(D_q\\) indicate batch size, number of encoder layers, number of video frames, number of patches per frame, number of query tokens, and the dimensions of visual/query features, respectively. These features are recurrently fused and refined by \\(\\mathbb{R}^2\\) block to construct spatial-temporal features \\(h \\in \\mathbb{R}^{B \\times T \\times C}\\), in which each token preserves the \\(C\\)-dimensional features for a frame. This process will be introduced in detail in Section 3.3." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.518, + 0.492, + 0.535 + ], + "angle": 0, + "content": "3.3 Reversed Recurrent Tuning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.543, + 0.788, + 0.665 + ], + "angle": 0, + "content": "Given pre-trained visual and query encoders, existing works [16,22,26,31,35] merely take features from the last hidden layer, which is sub-optimal due to the limited information. In this work, we propose \\(\\mathbb{R}^2\\)-Tuning to exploit the potential of spatial-temporal modeling based on multi-layer CLIP features. Our scheme learns a lightweight \\(\\mathbb{R}^2\\) Block which maintains a zero-initialized hidden state \\(h \\in \\mathbb{R}^{B \\times T \\times C}\\) as frame-level spatial-temporal features. The \\(\\mathbb{R}^2\\) Block is recurrently attached to the last \\(K\\) layers of CLIP encoders from back to front to refine the hidden state \\(h\\) for \\(K\\) steps, with each step defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.422, + 0.675, + 0.787, + 0.695 + ], + "angle": 0, + "content": "\\[\nh ^ {k} = \\mathcal {F} _ {\\theta} (e _ {v} ^ {n}, e _ {q} ^ {n}, h ^ {k - 1}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.704, + 0.789, + 0.794 + ], + "angle": 0, + "content": "Here, \\(\\mathcal{F}_{\\theta}\\) is the refinement operation parameterized by \\(\\theta\\). \\(k \\in [1, K]\\) is the index of current step. \\(n = N - k + 1\\) denotes the CLIP layer index for step \\(k\\), thus \\(e_v^n\\) and \\(e_q^n\\) are visual and query features from the \\(n\\)-th CLIP encoder layer. The refinement operation \\(\\mathcal{F}_{\\theta}\\) can be decomposed into two parts: 1) query-modulated spatial pooling, and 2) recurrent temporal refinement. Details are discussed as follows." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Query-Modulated Spatial Pooling Figure 5 (a) presents the data flow of query-modulated spatial pooling. The goal of this process is to adaptively pool" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.547, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Reversed Recurrent Tuning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "spatial features from patch-level representations \\( e_v^n \\) to a single token \\( e_{pool}^n \\) conditioning on the query \\( e_q^n \\). We first adopt two separate MLPs to map \\( e_v^n \\) and \\( e_q^n \\) into the same space:" + }, + { + "type": "equation", + "bbox": [ + 0.385, + 0.201, + 0.786, + 0.218 + ], + "angle": 0, + "content": "\\[\n\\hat {e} _ {v} ^ {n} = \\operatorname {M L P} \\left(e _ {v} ^ {n}\\right) \\in \\mathbb {R} ^ {B \\times T \\times (P + 1) \\times C} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.386, + 0.222, + 0.785, + 0.241 + ], + "angle": 0, + "content": "\\[\n\\hat {e} _ {q} ^ {n} = \\operatorname {M L P} \\left(e _ {q} ^ {n}\\right) \\in \\mathbb {R} ^ {B \\times L \\times C} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.248, + 0.784, + 0.312 + ], + "angle": 0, + "content": "Here, \\(C\\) is the reduced hidden size. We further align the shape of \\(\\hat{e}_v^n\\) and \\(\\hat{e}_q^n\\) by merging dimension \\(T\\) into \\(B\\) and repeat \\(\\hat{e}_q^n\\) for \\(T\\) times in an interleaved style, resulting \\(\\hat{e}_v^n \\in \\mathbb{R}^{(B \\times T) \\times (P + 1) \\times C}\\) and \\(\\hat{e}_q^n \\in \\mathbb{R}^{(B \\times T) \\times L \\times C}\\). Then, we compute the similarities for patch-token pairs using normalized Embedded Gaussian [50]:" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.321, + 0.785, + 0.356 + ], + "angle": 0, + "content": "\\[\na = \\operatorname {s o f t m a x} \\left(\\frac {\\left(w _ {q} \\hat {e} _ {q} ^ {n}\\right) ^ {\\top} w _ {v} \\hat {e} _ {v} ^ {n}}{\\sqrt {C}}\\right) \\in \\mathbb {R} ^ {(B \\times T) \\times L \\times P} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.362, + 0.784, + 0.438 + ], + "angle": 0, + "content": "Here, \\( w_{v} \\) and \\( w_{q} \\) are learnable matrices for projecting features. This operation derives from cross-attention but discards the linear projection on value. We then multiply the similarities \\( a \\) with visual features \\( \\hat{e}_{v}^{n} \\) to pool them into each token, such that tokens can interact with patches independently. After that, a max pooling along the \\( L \\) dimension is applied:" + }, + { + "type": "equation", + "bbox": [ + 0.349, + 0.447, + 0.785, + 0.466 + ], + "angle": 0, + "content": "\\[\ne _ {t o k e n} ^ {n} = \\operatorname {M a x P o o l i n g} \\left(a \\hat {e} _ {v} ^ {n} + \\hat {e} _ {q} ^ {n}\\right) \\in \\mathbb {R} ^ {B \\times T \\times C} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.473, + 0.784, + 0.533 + ], + "angle": 0, + "content": "Here, \\(\\hat{e}_q^n\\) serves as residuals [13] to stabilize training. Our visualizations in Figure 8 show that this strategy can guide the model to focus more on query-related regions. Finally, we combine \\(e_{token}^n\\) with the [CLS] token to generate query-modulated spatial features:" + }, + { + "type": "equation", + "bbox": [ + 0.376, + 0.543, + 0.785, + 0.561 + ], + "angle": 0, + "content": "\\[\ne _ {p o o l} ^ {n} = e _ {v} ^ {n, 0} + g ^ {k} \\cdot e _ {t o k e n} ^ {n} \\in \\mathbb {R} ^ {B \\times T \\times C} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.57, + 0.784, + 0.631 + ], + "angle": 0, + "content": "Here, \\( g^{k} \\in (-1,1) \\) is a zero-initialized learnable gate for step \\( k \\) constrained by Tanh. We allow negative values here to remove useless information from [CLS] tokens. \\( e_{pool}^{n} \\) is then used to model temporal correlations and are refined in a recurrent manner." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.644, + 0.784, + 0.692 + ], + "angle": 0, + "content": "Recurrent Temporal Refinement Figure 5 (b) displays the flow of recurrent temporal refinement. Specifically, the pooled visual features \\( e_{pool}^{n} \\) is first fused with the hidden state \\( h^{k-1} \\) from the previous step:" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.7, + 0.785, + 0.719 + ], + "angle": 0, + "content": "\\[\n\\hat {h} ^ {k - 1} = \\varphi^ {k} \\cdot e _ {\\text {p o o l}} ^ {n} + (1 - \\varphi^ {k}) \\cdot h ^ {k - 1} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.728, + 0.784, + 0.774 + ], + "angle": 0, + "content": "where \\(\\varphi^k\\in (0,1)\\) is the learnable gate for step \\(k\\). Then, we adopt a sequence of standard multi-head cross-attention \\((\\hat{e}_q^n\\) as \\(k,v)\\), multi-head self-attention, followed by a feed-forward network [50] to update the hidden state:" + }, + { + "type": "equation", + "bbox": [ + 0.369, + 0.784, + 0.785, + 0.803 + ], + "angle": 0, + "content": "\\[\nh ^ {k} = \\operatorname {F F N} (\\mathrm {M H S A} (\\mathrm {M H C A} (\\hat {h} ^ {k - 1}, \\hat {e} _ {q} ^ {n}))) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.784, + 0.842 + ], + "angle": 0, + "content": "For each block, we adopt DropPath [21] with \\( p = 0.1 \\) to prevent overfitting. The query features \\( \\hat{e}_q^n \\) are utilized both in spatial and temporal modeling as guidance." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.221 + ], + "angle": 0, + "content": "Granularity Calibration The refinement process above is based on the assumption that visual and query features from the same layer of CLIP [41] are well aligned at the same granularity level. However, this cannot be guaranteed as the two encoders are learned in isolation during pre-training. Therefore, we need to add manual constraints to calibrate their granularities." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.223, + 0.785, + 0.298 + ], + "angle": 0, + "content": "We apply a video-level constraint and a layer-wise constraint by designing two contrastive losses. We first denote \\( e_v^0 \\in \\mathbb{R}^{B \\times K \\times T \\times C} \\) as the features of \\( K \\) [CLS] tokens from Eq. 6. Then, we select all positive frames (i.e., frames within \\( [b_i^s, b_i^e] \\) or with \\( s_i \\) or \\( f_i \\) higher than a threshold) from \\( e_v^0 \\) and perform average pooling to obtain video-level representations \\( \\widetilde{e}_v \\in \\mathbb{R}^{B \\times K \\times C} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.414, + 0.307, + 0.785, + 0.323 + ], + "angle": 0, + "content": "\\[\n\\widetilde {e} _ {v} = \\operatorname {A v g P o o l} \\left(\\left\\{e _ {v} ^ {0, i} \\right\\} _ {i \\in \\Omega}\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.331, + 0.784, + 0.391 + ], + "angle": 0, + "content": "Here, \\(\\Omega\\) is the set of positive frame indices. To obtain query-level representations \\(\\widetilde{e}_q\\in \\mathbb{R}^{B\\times K\\times C}\\), a token-wise adaptive pooling is applied following previous work [26]. We adopt InfoNCE loss [38] along two dimensions to calculate the video-level and layer-wise constraints:" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.399, + 0.785, + 0.434 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {v i d e o}} = \\lambda_ {\\text {v i d e o}} \\frac {1}{B} \\sum_ {b \\in B} \\operatorname {I n f o N C E} \\left(\\tilde {e} _ {v} ^ {b}, \\tilde {e} _ {q} ^ {b}\\right) \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.438, + 0.785, + 0.472 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {l a y e r}} = \\lambda_ {\\text {l a y e r}} \\frac {1}{K} \\sum_ {k \\in K} \\operatorname {I n f o N C E} \\left(\\widetilde {e} _ {v} ^ {k}, \\widetilde {e} _ {q} ^ {k}\\right) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.481, + 0.785, + 0.571 + ], + "angle": 0, + "content": "Here, \\(\\mathcal{L}_{video}\\) performs contrast among samples in the same batch (diversifying features are diverse among moment-query pairs), and averages the loss across \\(K\\) layers. It also implicitly models temporal correlations between videos and queries. \\(\\mathcal{L}_{layer}\\) performs contrast among layers and averages across the batch. \\(\\lambda_{video}\\) and \\(\\lambda_{layer}\\) serve as re-weighting terms for the losses. A fixed temperature parameter of 0.07 is used." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.591, + 0.405, + 0.605 + ], + "angle": 0, + "content": "3.4 Prediction Heads" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.784, + 0.718 + ], + "angle": 0, + "content": "After refining spatial-temporal features \\( h \\), a temporal feature pyramid is constructed by applying 1D convolutions with stride \\( = 2 \\) on \\( h \\). The number of convolutions used for each level is subjected to \\( l - 1 \\), where \\( l \\) is the level index starting from 1. We concatenate features from all levels by the temporal dimension to form \\( \\hat{h} \\) and predict once in all heads. Following a similar but simplified design from [26], we adopt three heads for video temporal grounding, described in detail as follows." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.73, + 0.784, + 0.776 + ], + "angle": 0, + "content": "Foreground-Background Classification Given \\(\\hat{h}\\), a two-layer 1D convolution module with \\((1\\times 3)\\) kernels is adopted to predict \\(\\hat{f}_i\\) for each frame. It is optimized using a focal loss [27] with \\(\\alpha = 0.9\\) and \\(\\gamma = 2.0\\):" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.785, + 0.785, + 0.802 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c l s} = - \\lambda_ {c l s} \\alpha (1 - \\hat {f} _ {i}) ^ {\\gamma} \\log (\\hat {f} _ {i}) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.784, + 0.841 + ], + "angle": 0, + "content": "We find that focal loss perform better than binary cross-entropy when the numbers of foreground/background frames are severely imbalanced as noted in [26]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.547, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Reversed Recurrent Tuning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 1: Video moment retrieval (MR) and highlight detection (HD) results on QVHighlights test split. Note that [16] and [57] only reported their results on val split. \\(\\bigcirc\\) and \\(\\bigcirc\\) are CLIP-B/32 and B/16 [41], \\(\\bigcirc\\) denotes SlowFast R-50 [8], and \\(\\bigcirc\\) is PANN [20] for audio features extraction. * means estimated parameters. The best and second-best metrics are marked with bold and underline, respectively." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.226, + 0.789, + 0.487 + ], + "angle": 0, + "content": "
MethodBackbonePost \nPre-trainMRHD
R1mAP≥ Very Good#Params
@0.5@0.7@0.5@0.75Avg.mAPHIT@1
BeautyThumb [44]○○X-----14.3620.88-
DVSE [30]○○X-----18.7521.79-
MCN [1]○○X11.412.7224.948.2210.67---
CAL [7]○○X25.4911.5423.407.659.89---
XML [23]○○X41.8330.3544.6331.7332.1434.4955.25-
XML+ [22]○○X46.6933.4647.8934.6734.9035.3855.06-
Moment-DETR [22]○○X52.8933.0254.8229.4030.7335.6955.604.8M
UMT [31]○○○X56.2341.1853.8337.0136.1238.1859.9914.9M
MomentDiff [24]○○○X58.2141.4854.5737.2136.84---
QD-DETR [35]○○X62.4044.9862.5239.8839.8638.9462.407.6M
MH-DETR [56]○○X60.0542.4860.7538.1338.3838.2260.518.2M
UniVTG [26]○○X58.8640.8657.6035.5935.4738.2060.9641.3M
TR-DETR [46]○○X64.6648.9663.9843.7342.6239.9163.427.9M
CG-DETR [34]○○X65.4348.3864.5142.7742.8640.3366.2112.0M
EaTR [16]○○X61.3645.7961.8641.9141.7437.1558.659.0M
Moment-DETR [22]○○236K ASR Cap.59.7840.3360.5135.3636.1437.4360.174.8M
UMT [31]○○○236K ASR Cap.60.8343.2657.3339.1238.0839.1262.3914.9M
UniVTG [26]○○4.2M Corpus65.4350.0664.0645.0243.6340.5466.2841.3M
UnLoc [57]○650K Videos64.5048.80-----87.9M*
R²-Tuning (Ours)○X68.0349.3569.0447.5646.1740.7564.202.7M
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.511, + 0.788, + 0.586 + ], + "angle": 0, + "content": "Boundary Regression Similar to foreground-background classification, boundary regression for moments is also realized by a two-layer 1D convolution module. The difference is that its output dimension is set to 2 instead of 1, representing the boundary displacements \\([\\hat{b}_i^s,\\hat{b}_i^e ]\\) for start-end timestamps. This head is optimized by an L1 loss:" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.591, + 0.786, + 0.611 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e g} = \\lambda_ {r e g} \\left(\\left| b _ {i} ^ {s} - \\hat {b} _ {i} ^ {s} \\right| + \\left| b _ {i} ^ {e} - \\hat {b} _ {i} ^ {e} \\right|\\right) \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.788, + 0.663 + ], + "angle": 0, + "content": "where \\( b_{i}^{s} \\) and \\( b_{i}^{e} \\) are the ground truths. This loss is only applied to frames inside the ground truth boundaries. We observe that using an L1 loss already works better than the combination of Smooth L1 Loss and GIoU Loss as in [26]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.671, + 0.788, + 0.72 + ], + "angle": 0, + "content": "Saliency Prediction To obtain the saliency scores \\(\\hat{s}_i\\) for HD, we calculate the cosine similarities between the adaptive pooled query features at the last refinement step \\(\\widetilde{e}_q^K\\) and each token in the spatial-temporal features \\(h\\):" + }, + { + "type": "equation", + "bbox": [ + 0.435, + 0.723, + 0.786, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\hat {s} _ {i} = \\frac {h _ {i} ^ {\\top} \\widetilde {e} _ {q} ^ {K}}{| | h _ {i} | | _ {2} | | \\widetilde {e} _ {q} ^ {K} | | _ {2}} \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.799 + ], + "angle": 0, + "content": "The training objective is applied through a contrastive loss between sampled positive frames (with index \\( p \\in P \\)) and the adaptively pooled query \\( \\widetilde{e}_q^K \\):" + }, + { + "type": "equation", + "bbox": [ + 0.336, + 0.804, + 0.786, + 0.84 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s a l} = - \\lambda_ {s a l} \\log \\frac {\\exp \\left(\\hat {s} _ {p} / \\tau\\right)}{\\exp \\left(\\hat {s} _ {p} / \\tau\\right) + \\sum_ {i \\in \\Theta} \\exp \\left(\\hat {s} _ {i} / \\tau\\right)} \\tag {15}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 2: Video moment retrieval results on Ego4D-NLQ, Charades-STA, and TACoS datasets. \\(\\bigcirc\\) means using both CLIP-B/32 and SlowFast R-50 as feature extractors. \\(\\bigcirc\\) indicates using CLIP-B/32 only. The best and second-best metrics are marked with bold and underline, respectively. Our method shows significant advantages in high-quality retrievals (R@0.7) even without a video backbone." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.226, + 0.789, + 0.324 + ], + "angle": 0, + "content": "
MethodEgo4D-NLQ [10]Charades-STA [9]TACoS [43]
R@0.3R@0.5R@0.7mIoUR@0.3R@0.5R@0.7mIoUR@0.3R@0.5R@0.7mIoU
○2D-TAN [64]4.331.830.603.3958.7646.0227.5041.2540.0127.9912.9227.22
○VSLNet [62]4.542.401.013.5460.3042.6924.1441.5835.5423.5413.1524.99
○Moment-DETR [22]4.341.810.653.5365.8352.0730.5945.5437.9724.6711.9725.49
○UniVTG [26]7.283.951.324.9170.8158.0135.6550.1051.4434.9717.3533.60
○R²-Tuning (Ours)7.204.492.124.9470.9159.7837.0250.8649.7138.7225.1235.92
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.353, + 0.787, + 0.384 + ], + "angle": 0, + "content": "Here, \\(\\Theta\\) is the set of frame indices where \\(s_i < s_p\\), and \\(\\tau\\) is a fixed temperature parameter set to 0.07." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.402, + 0.788, + 0.522 + ], + "angle": 0, + "content": "Training & Inference The whole model is jointly optimized using a sum of the five losses mentioned above. In practice, the loss weights are set as \\(\\lambda_{video} = 0.1\\), \\(\\lambda_{layer} = 0.1\\), \\(\\lambda_{cls} = 1.0\\), \\(\\lambda_{reg} = 0.1\\), and \\(\\lambda_{sal} = 0.1\\). During inference, we form an moment retrieval prediction by combining \\(f_{i}\\) and \\(b_{i}\\), i.e., calculating the start-end timestamps from frame index \\(i\\) and boundary displacements \\([b_i^s, b_i^e]\\), while regarding \\(f_{i}\\) as the confidence. NMS with IoU threshold \\(\\theta_{IoU} = 0.7\\) is applied to reduce duplicates. For highlight detection and video summarization, we directly use the frame-level output from saliency prediction \\(s_i\\)." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.549, + 0.377, + 0.566 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.584, + 0.517, + 0.599 + ], + "angle": 0, + "content": "4.1 Datasets & Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.613, + 0.787, + 0.688 + ], + "angle": 0, + "content": "We conduct experiments on six datasets covering various domains including daily vlogs & news (QVHighlights [22]), in-door scenes (TACoS [43] and Charades-STA [9]), egocentric videos (Ego4D-NLQ [10]), and sports (YouTube Highlights [47] and TVSum [45]). Details about the datasets are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.689, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We adopt the same evaluation metrics with previous works [22, 26, 31]. To be specific, we compute Recall@1 with IoU threshold \\(\\theta_{IoU} = 0.5\\) and 0.7, mean average precision (mAP) with \\(\\theta_{IoU} = 0.5\\) and 0.7, and mAP with a series of thresholds [0.5:0.05:0.95] for MR on QVHighlights [22]. mAP and HIT@1 where positive samples are defined as with the saliency score of Very Good are adopted for HD. On Ego4D-NLQ [10], Charades-STA [9], and TACoS [43] datasets, we utilize Recall@1 with \\(\\theta_{IoU} = \\{0.3, 0.5, 0.7\\}\\) and mIoU to measure the MR performance. On YouTube Highlights [47] for HD and TVSum [45] for VS, we follow [26, 31] and use the same train/val split with mAP and Top-5 mAP as metrics, respectively." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.547, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reversed Recurrent Tuning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 3: Class-wise video highlight detection results (mAP) on YouTube Highlights dataset. \\(\\bigcirc\\), \\(\\bigcirc\\), and \\(\\bigcirc\\) denote image, video, and audio backbones, respectively." + }, + { + "type": "table", + "bbox": [ + 0.302, + 0.185, + 0.703, + 0.348 + ], + "angle": 0, + "content": "
MethodDogGym.Par.Ska.Ski.Sur.Avg.
○ RRAE [58]49.035.050.025.022.049.038.3
○ GIFs [12]30.833.554.055.432.854.146.4
○ LSVM [47]60.041.061.062.036.061.053.6
○ LIM-S [54]57.941.767.057.848.665.156.4
○ SL-Module [55]70.853.277.272.566.176.269.3
○ PLD [53]74.970.277.957.570.779.073.0
○ MINI-Net [14]58.261.770.272.258.765.164.4
○ TCG [59]55.462.770.969.160.159.863.0
○ Joint-VA [3]64.571.980.862.073.278.371.8
○ CO-AV [25]60.966.089.074.169.081.174.7
○ UMT [31]65.975.281.671.872.382.774.9
○ UniVTG [26]71.876.573.973.373.282.275.2
○ R²-Tuning75.673.573.074.674.884.876.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.36, + 0.788, + 0.403 + ], + "angle": 0, + "content": "Table 4: Class-wise video summarization results (Top-5 mAP) on TVSum dataset. \\(\\odot\\) , \\(\\odot\\) , and \\(\\odot\\) denote image, video, and audio backbones, respectively. Our method does not require any additional video/audio features." + }, + { + "type": "table", + "bbox": [ + 0.266, + 0.415, + 0.739, + 0.578 + ], + "angle": 0, + "content": "
MethodVTVUGAMSPKPRFMBKBTDSAvg.
○ sLSTM [63]41.146.246.347.744.846.145.240.647.145.545.1
○ SG [33]42.347.247.548.945.647.346.441.748.346.646.2
○ LIM-S [54]55.942.961.254.060.447.543.266.369.162.656.3
○ Trailer [51]61.354.665.760.859.170.158.264.765.668.162.8
○ SL-Module [55]86.568.774.986.279.063.258.972.678.964.073.3
○ PLD [53]84.580.970.372.576.487.271.974.074.479.177.1
○ MINI-Net [14]80.668.378.281.878.165.857.875.080.265.573.2
○ TCG [59]85.071.481.978.680.275.571.677.378.668.176.8
○ Joint-VA [3]83.757.378.586.180.169.270.073.097.467.576.3
○ CO-AV [25]90.872.884.685.078.378.072.877.189.572.380.1
○ UMT [31]87.581.588.278.881.487.076.086.984.479.683.1
○ UniVTG [26]83.985.189.080.184.681.470.991.773.569.381.0
○ R²-Tuning85.085.991.081.788.887.478.189.290.374.785.2
" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.604, + 0.457, + 0.62 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.766 + ], + "angle": 0, + "content": "In all experiments, we adopt the visual and text encoders of CLIP ViT-B/32 [41] as our backbones. The whole CLIP model is frozen during training, i.e., only the parameters in \\(\\mathbb{R}^2\\) Block, temporal feature pyramid, prediction heads are learnable. Without further specification, we set \\(K = 4\\), i.e., attaching \\(\\mathbb{R}^2\\) Block to the last 4 layers of CLIP encoders. In \\(\\mathbb{R}^2\\) Block, the hidden size is set to 256. Only one transformer layer with post-norm style and 8 attention heads is used for temporal modeling. The number of layers for the temporal feature pyramid is set to 4 for QVHighlights, Ego4D-NLQ, Charades-STA, and TACoS, while 1 for other datasets. Please refer to the supplementary material for details." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.786, + 0.551, + 0.801 + ], + "angle": 0, + "content": "4.3 Comparison with State-of-the-Arts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Joint Moment Retrieval and Highlight Detection We first evaluate our method on QVHighlights [22] test split, the only dataset that supports both" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.377, + 0.145, + 0.625, + 0.152 + ], + "angle": 0, + "content": "QUERY: A man in white is browsing through clothes at a department store." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.153, + 0.787, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.416, + 0.255, + 0.586, + 0.263 + ], + "angle": 0, + "content": "QUERY: Man stands on playground balance bridge." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.264, + 0.787, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.363, + 0.788, + 0.435 + ], + "angle": 0, + "content": "Fig.6: Visualization of joint moment retrieval and highlight detection results on QVHighlights val split. We compare our method with UniVTG [26], which is a strong baseline leveraging both additional temporal backbone and large-scale post pretraining. \\( \\mathbb{R}^2 \\)-Tuning can accurately regress the boundaries of moments and predict highlight saliency scores through its novel feature refinement design." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.466, + 0.788, + 0.572 + ], + "angle": 0, + "content": "moment retrieval and highlight detection. The results are shown in Table 1. The first group of methods utilizes more than one feature extractor. The second group is that with large-scale post pre-training (defined as the extra pre-training stage between loading backbone weights and training on downstream datasets.). Without any additional backbone and extra pre-training, \\(\\mathbf{R}^2\\)-Tuning achieves start-of-the-art performance compared with all previous methods with the fewest learnable parameters. Figure 6 visualizes the model predictions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.788, + 0.7 + ], + "angle": 0, + "content": "Moment Retrieval We then evaluate our model on moment retrieval task in egocentric [10] and in-door [9, 43] domains. The results are shown in Table 2. We follow [26] and compare only with the methods using CLIP + SlowFast features. R²-Tuning still works better than all baselines with extra features. We also observe that R²-Tuning shows its significance on high-quality retrievals (R1@0.7), which requires accurate temporal modeling. This benefits from the fine-grained temporal modeling ability of reversed and recurrent designs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Highlight Detection & Video Summarization The performances of highlight detection on YouTube Highlights [47] and extractive video summarization on TVSum [45] are reported in Table 3 and Table 4, respectively. Following previous works [26, 31], we trained the model separately on each domain. The first group of methods is based on video backbones, which naturally have basic temporal modeling abilities. The second group of methods is enhanced by extra features such as image and audio. \\(\\mathbf{R}^2\\)-Tuning can still reach the best performance when trained on small-scale data." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.548, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reversed Recurrent Tuning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.144, + 0.387, + 0.153 + ], + "angle": 0, + "content": "(a) Naive Fusion" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.155, + 0.353, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.156, + 0.497, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.144, + 0.704, + 0.153 + ], + "angle": 0, + "content": "(b) Recurrent Refinement" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.155, + 0.644, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.155, + 0.788, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.264, + 0.788, + 0.307 + ], + "angle": 0, + "content": "Fig. 7: Different feature refinement strategies for MR on QVHighlights val split. Single means utilizing the \\( K \\)-th layer only, while Mean denotes averaging the features from last \\( K \\) layers. Forward and Reversed indicate different feature refinement directions." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.319, + 0.548, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.476, + 0.549, + 0.518 + ], + "angle": 0, + "content": "Fig. 8: Visualization of the attention maps for query-modulated spatial pooling. Diverse regions are learned for different queries." + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.319, + 0.759, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.557, + 0.476, + 0.788, + 0.518 + ], + "angle": 0, + "content": "Fig. 9: Cosine similarities of video (x-axis) and text (y-axis) features among \\(7\\sim 12\\) CLIP layers." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.547, + 0.409, + 0.563 + ], + "angle": 0, + "content": "4.4 Detailed Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.571, + 0.788, + 0.678 + ], + "angle": 0, + "content": "Multi-Granularity Features To investigate the significance of granularity flexibility, we set up a simple baseline that only attaches \\(\\mathbf{R}^2\\) Block to one layer of CLIP encoders. Then, we make use of the multi-layer information by averaging the features from the last \\(K\\) layers. These two variants are compared in Figure 7 (a). It can be observed that for the single-layer setting, higher-level features are more discriminative than lower-level ones. However, even a naive fusion strategy for the multi-layer information brings significant performance gains." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.782 + ], + "angle": 0, + "content": "Reversed Recurrent Feature Refinement We then further justify our design by comparing two refinement directions (i.e., forward and reversed) in Figure 7 (b). Note that when \\( K = 1 \\), all four designs in Figure 7 are of the same architecture. When utilizing more layers, both forward and reversed refinement benefit from more information, while the reversed scheme statically performs better, as the multi-granularity features shall be refined from coarse to fine." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Query-Modulated Spatial Pooling We visualize the attention maps learned by query-modulated spatial pooling in Figure 8. A query generally has multiple keywords that provide discriminative information to localize moments. The" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.511, + 0.174 + ], + "angle": 0, + "content": "Table 5: Effectiveness justification of granularity calibration on QVHighlights val split." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.185, + 0.513, + 0.266 + ], + "angle": 0, + "content": "
L VideoL layerMRHD
R1@0.5R1@0.7mAPmAPHIT@1
64.4848.6044.0137.9462.67
✓67.6851.6146.7439.8165.16
✓64.7148.8444.6038.9163.35
✓✓68.7152.0647.5940.5964.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.521, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 6: MR mAP with different query lengths on QVHighlights val split." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.185, + 0.787, + 0.267 + ], + "angle": 0, + "content": "
Method#Words
1-1011-2021-3031-40≥41
QD-DETR [35]40.3042.3229.010.1026.67
EaTR [16]38.9439.5424.5812.0341.43
UniVTG [26]32.3832.6520.880.0031.11
R2-Tuning47.0849.1328.7967.2472.38
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.295, + 0.784, + 0.326 + ], + "angle": 0, + "content": "token-wise attention in Eq. 4 can guide the model to focus on multiple regions simultaneously, as can be seen in the diverse but meaningful patterns." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.788, + 0.555 + ], + "angle": 0, + "content": "Granularity Calibration To verify the effectiveness of granularity calibration, we also visualize the cosine distances of all the visual-query pairs from \\(7 \\sim 12\\) layers in Fig. 9. Before calibration, the visual-text features are not well-aligned. Adding the video-level constraint makes higher-level features aligned with each other (b), while adding layer-wise constraint makes the model distill diverse information across different layers (c). The diagonal in (d) verifies that both calibration constraints can maximize the mutual similarities of paired visual and text features. Some quantitative results are presented in Table 5. Both video-level and layer-wise contrastive can help align the multi-granularity visual & query semantics, while their combination can further boost the performances. We also group the queries in QVHighlights into 5 bins with different lengths (acting as multi-granularities) and compute the means of MR mAP in Tab. 6. Although most training queries are coarse-grained (\\(\\leqslant 30\\) words), R²-Tuning can easily generalize to fine-grained queries (\\(\\geqslant 31\\) words), surpassing previous methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.58, + 0.36, + 0.596 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.751 + ], + "angle": 0, + "content": "This paper introduces R²-Tuning, a parameter- and memory-efficient transfer learning framework for video temporal grounding. It learns a lightweight side-adapter (R² Block) that is recurrently attached to the last few layers of a frozen CLIP to adaptively pool spatial details and refine temporal correlations. Two contrastive constraints (video-level and layer-wise) are utilized to calibrate the granularities of CLIP visual and text encoders. Experiments across three VTG tasks on six public benchmarks demonstrate the significance and effectiveness of the proposed scheme. We hope that the proposed framework sparks further research on efficient image-to-video transfer learning for untrimmed videos." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.776, + 0.394, + 0.794 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "This work was supported in part by Hong Kong Research Grants Council GRF-15229423, US NIH grant R01HD104969, and NSF award IIS-2239688." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.548, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reversed Recurrent Tuning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "1. Anne Hendricks, L., Wang, O., Shechtman, E., Sivic, J., Darrell, T., Russell, B.: Localizing moments in video with natural language. In: ICCV. pp. 5803-5812 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.785, + 0.245 + ], + "angle": 0, + "content": "2. Apostolidis, E., Adamantidou, E., Metsai, A.I., Mezaris, V., Patras, I.: Video summarization using deep neural networks: A survey. Proceedings of the IEEE 109(11), 1838-1863 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.247, + 0.785, + 0.273 + ], + "angle": 0, + "content": "3. Badamdorj, T., Rochan, M., Wang, Y., Cheng, L.: Joint visual and audio learning for video highlight detection. In: ICCV. pp. 8127-8137 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.274, + 0.785, + 0.3 + ], + "angle": 0, + "content": "4. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: A clip-hitchhiker's guide to long video retrieval. Tech. Rep. arXiv:2205.08508 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.302, + 0.785, + 0.329 + ], + "angle": 0, + "content": "5. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: CVPR. pp. 6299-6308 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.33, + 0.785, + 0.37 + ], + "angle": 0, + "content": "6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.785, + 0.411 + ], + "angle": 0, + "content": "7. Escorcia, V., Soldan, M., Sivic, J., Ghanem, B., Russell, B.: Temporal localization of moments in video collections with natural language. Tech. Rep. arXiv:1907.12763 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.785, + 0.439 + ], + "angle": 0, + "content": "8. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: ICCV. pp. 6202-6211 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.44, + 0.785, + 0.467 + ], + "angle": 0, + "content": "9. Gao, J., Sun, C., Yang, Z., Nevatia, R.: Tall: Temporal activity localization via language query. In: ICCV. pp. 5267-5275 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.468, + 0.785, + 0.509 + ], + "angle": 0, + "content": "0. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: CVPR. pp. 18995-19012 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "1. Gygli, M., Grabner, H., Riemenschneider, H., Van Gool, L.: Creating summaries from user videos. In: ECCV. pp. 505-520 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.537, + 0.785, + 0.563 + ], + "angle": 0, + "content": "2. Gygli, M., Song, Y., Cao, L.: Video2gif: Automatic generation of animated gifs from video. In: CVPR. pp. 1001-1009 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.592 + ], + "angle": 0, + "content": "3. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.593, + 0.785, + 0.619 + ], + "angle": 0, + "content": "4. Hong, F.T., Huang, X., Li, W.H., Zheng, W.S.: Mini-net: Multiple instance ranking network for video highlight detection. In: ECCV. pp. 345-360 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.785, + 0.66 + ], + "angle": 0, + "content": "5. Huang, S., Gong, B., Pan, Y., Jiang, J., Lv, Y., Li, Y., Wang, D.: Vop: Text-video co-operative prompt tuning for cross-modal retrieval. In: CVPR. pp. 6565-6574 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.688 + ], + "angle": 0, + "content": "6. Jang, J., Park, J., Kim, J., Kwon, H., Sohn, K.: Knowing where to focus: Event-aware transformer for video grounding. In: ICCV. pp. 13846-13856 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.716 + ], + "angle": 0, + "content": "7. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: ECCV. pp. 709-727 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.785, + 0.744 + ], + "angle": 0, + "content": "8. Jiang, R., Liu, L., Chen, C.: Clip-count: Towards text-guided zero-shot object counting. In: ACM MM (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.785, + 0.771 + ], + "angle": 0, + "content": "9. Ju, C., Han, T., Zheng, K., Zhang, Y., Xie, W.: Prompting visual-language models for efficient video understanding. In: ECCV. pp. 105-124 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "20. Kong, Q., Cao, Y., Iqbal, T., Wang, Y., Wang, W., Plumbley, M.D.: Panns: Large-scale pretrained audio neural networks for audio pattern recognition. IEEE/ACM Transactions on Audio, Speech, and Language Processing 28, 2880-2894 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "21. Larsson, G., Maire, M., Shakhnarovich, G.: Fractalnet: Ultra-deep neural networks without residuals. Tech. Rep. arXiv:1605.07648 (2016)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "22. Lei, J., Berg, T.L., Bansal, M.: Qvhighlights: Detecting moments and highlights in videos via natural language queries. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "23. Lei, J., Yu, L., Berg, T.L., Bansal, M.: Tv: A large-scale dataset for video-subtitle moment retrieval. In: ECCV. pp. 447-463 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.244 + ], + "angle": 0, + "content": "24. Li, P., Xie, C.W., Xie, H., Zhao, L., Zhang, L., Zheng, Y., Zhao, D., Zhang, Y.: Momentdiff: Generative video moment retrieval from random to real. Tech. Rep. arXiv:2307.02869 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.285 + ], + "angle": 0, + "content": "25. Li, S., Zhang, F., Yang, K., Liu, L., Liu, S., Hou, J., Yi, S.: Probing visual-audio representation for video highlight detection via hard-pairs guided contrastive learning. Tech. Rep. arXiv:2206.10157 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "26. Lin, K.Q., Zhang, P., Chen, J., Pramanick, S., Gao, D., Wang, A.J., Yan, R., Shou, M.Z.: Univtg: Towards unified video-language temporal grounding. In: CVPR. pp. 2794-2804 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.355 + ], + "angle": 0, + "content": "27. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: ICCV. pp. 2980-2988 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.383 + ], + "angle": 0, + "content": "28. Lin, Z., Geng, S., Zhang, R., Gao, P., de Melo, G., Wang, X., Dai, J., Qiao, Y., Li, H.: Frozen clip models are efficient video learners. In: ECCV. pp. 388-404 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "29. Liu, L., Yu, B.X., Chang, J., Tian, Q., Chen, C.W.: Prompt-matched semantic segmentation. Tech. Rep. arXiv:2208.10159 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.438 + ], + "angle": 0, + "content": "30. Liu, W., Mei, T., Zhang, Y., Che, C., Luo, J.: Multi-task deep visual-semantic embedding for video thumbnail selection. In: CVPR. pp. 3707-3715 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.48 + ], + "angle": 0, + "content": "31. Liu, Y., Li, S., Wu, Y., Chen, C.W., Shan, Y., Qie, X.: Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In: CVPR. pp. 3042-3051 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.521 + ], + "angle": 0, + "content": "32. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: Clip4clip: An empirical study of clip for end to end video clip retrieval. Neurocomputing 508, 293-304 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.549 + ], + "angle": 0, + "content": "33. Mahasseni, B., Lam, M., Todorovic, S.: Unsupervised video summarization with adversarial LSTM networks. In: CVPR. pp. 202-211 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.59 + ], + "angle": 0, + "content": "34. Moon, W., Hyun, S., Lee, S., Heo, J.P.: Correlation-guided query-dependency calibration in video representation learning for temporal grounding. Tech. Rep. arXiv:2311.08835 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.785, + 0.632 + ], + "angle": 0, + "content": "35. Moon, W., Hyun, S., Park, S., Park, D., Heo, J.P.: Query-dependent video representation for moment retrieval and highlight detection. In: CVPR. pp. 23023-23033 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.66 + ], + "angle": 0, + "content": "36. Nan, G., Qiao, R., Xiao, Y., Liu, J., Leng, S., Zhang, H., Lu, W.: Interventional video grounding with dual contrastive learning. In: CVPR. pp. 2765-2775 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.785, + 0.702 + ], + "angle": 0, + "content": "37. Ni, B., Peng, H., Chen, M., Zhang, S., Meng, G., Fu, J., Xiang, S., Ling, H.: Expanding language-image pretrained models for general video recognition. In: ECCV. pp. 1-18 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.729 + ], + "angle": 0, + "content": "38. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. In: NeurIPS (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.785, + 0.757 + ], + "angle": 0, + "content": "39. Pan, J., Lin, Z., Zhu, X., Shao, J., Li, H.: St-adapter: Parameter-efficient image-to-video transfer learning. In: NeurIPS. pp. 26462-26477 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "40. Qing, Z., Zhang, S., Huang, Z., Zhang, Y., Gao, C., Zhao, D., Sang, N.: Disentangling spatial and temporal learning for efficient image-to-video transfer learning. In: CVPR. pp. 13934-13944 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "41. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.548, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reversed Recurrent Tuning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "42. Rasheed, H., Khattak, M.U., Maaz, M., Khan, S., Khan, F.S.: Fine-tuned clip models are efficient video learners. In: CVPR. pp. 6545-6554 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.217 + ], + "angle": 0, + "content": "43. Regneri, M., Rohrbach, M., Wetzel, D., Thater, S., Schiele, B., Pinkal, M.: Grounding action descriptions in videos. Transactions of the Association for Computational Linguistics 1, 25-36 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "44. Song, Y., Redi, M., Vallmitjana, J., Jeimes, A.: To click or not to click: Automatic selection of beautiful thumbnails from videos. In: CIKM. pp. 659-668 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.272 + ], + "angle": 0, + "content": "45. Song, Y., Vallmitjana, J., Stent, A., Jaimes, A.: Tvsum: Summarizing web videos using titles. In: CVPR. pp. 5179-5187 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "46. Sun, H., Zhou, M., Chen, W., Xie, W.: Tr-detr: Task-reciprocal transformer for joint moment retrieval and highlight detection. In: AAAI (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.327 + ], + "angle": 0, + "content": "47. Sun, M., Farhadi, A., Seitz, S.: Ranking domain-specific highlights by analyzing edited videos. In: ECCV. pp. 787-802 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.355 + ], + "angle": 0, + "content": "48. Sung, Y.L., Cho, J., Bansal, M.: Lst: Ladder side-tuning for parameter and memory efficient transfer learning. In: NeurIPS. pp. 12991-13005 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.397 + ], + "angle": 0, + "content": "49. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. Tech. Rep. arXiv:2302.13971 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "50. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS. pp. 5998-6008 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.452 + ], + "angle": 0, + "content": "51. Wang, L., Liu, D., Puri, R., Metaxas, D.N.: Learning trailer moments in full-length movies. In: ECCV. pp. 300-316 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.479 + ], + "angle": 0, + "content": "52. Wang, Z., Wang, L., Wu, T., Li, T., Wu, G.: Negative sample matters: A renaissance of metric learning for temporal grounding. In: AAAI. pp. 2613-2623 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.48, + 0.785, + 0.508 + ], + "angle": 0, + "content": "53. Wei, F., Wang, B., Ge, T., Jiang, Y., Li, W., Duan, L.: Learning pixel-level distinctions for video highlight detection. In: CVPR. pp. 3073-3082 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "54. Xiong, B., Kalantidis, Y., Ghadiyaram, D., Grauman, K.: Less is more: Learning highlight detection from video duration. In: CVPR. pp. 1258-1267 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.563 + ], + "angle": 0, + "content": "55. Xu, M., Wang, H., Ni, B., Zhu, R., Sun, Z., Wang, C.: Cross-category video highlight detection via set-based learning. In: ICCV. pp. 7970-7979 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.591 + ], + "angle": 0, + "content": "56. Xu, Y., Sun, Y., Li, Y., Shi, Y., Zhu, X., Du, S.: Mh-detr: Video moment and highlight detection with cross-modal transformer. Tech. Rep. arXiv:2305.00355 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.785, + 0.632 + ], + "angle": 0, + "content": "57. Yan, S., Xiong, X., Nagrani, A., Arnab, A., Wang, Z., Ge, W., Ross, D., Schmid, C.: Unloc: A unified framework for video localization tasks. In: ICCV. pp. 13623-13633 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.673 + ], + "angle": 0, + "content": "58. Yang, H., Wang, B., Lin, S., Wipf, D., Guo, M., Guo, B.: Unsupervised extraction of video highlights via robust recurrent auto-encoders. In: ICCV. pp. 4633-4641 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.674, + 0.785, + 0.715 + ], + "angle": 0, + "content": "59. Ye, Q., Shen, X., Gao, Y., Wang, Z., Bi, Q., Li, P., Yang, G.: Temporal cue guided video highlight detection with low-rank audio-visual fusion. In: ICCV. pp. 7950-7959 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "60. Yuan, T., Zhang, X., Liu, K., Liu, B., Jin, J., Jiao, Z.: Ucf-crime annotation: A benchmark for surveillance video-and-language understanding. Tech. Rep. arXiv:2309.13925 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.784 + ], + "angle": 0, + "content": "61. Yuan, Y., Ma, L., Wang, J., Liu, W., Zhu, W.: Semantic conditioned dynamic modulation for temporal sentence grounding in videos. NeurIPS 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.785, + 0.812 + ], + "angle": 0, + "content": "62. Zhang, H., Sun, A., Jing, W., Zhou, J.T.: Span-based localizing network for natural language video localization. Tech. Rep. arXiv:2004.13931 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "63. Zhang, K., Chao, W.L., Sha, F., Grauman, K.: Video summarization with long short-term memory. In: ECCV. pp. 766-782. Springer (2016)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "64. Zhang, S., Peng, H., Fu, J., Luo, J.: Learning 2d temporal adjacent networks for moment localization with natural language. In: AAAI. pp. 12870-12877 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "65. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: CVPR. pp. 16816-16825 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "66. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.231 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/R^2-Tuning_ Efficient Image-to-Video Transfer Learning for Video Temporal Grounding/148e844e-b528-47e5-9db4-f50b0b9b4b9c_origin.pdf b/2024/R^2-Tuning_ Efficient Image-to-Video Transfer Learning for Video Temporal Grounding/148e844e-b528-47e5-9db4-f50b0b9b4b9c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dc4c26ef1b7de365e63f5bc9a556d339ec70a39a --- /dev/null +++ b/2024/R^2-Tuning_ Efficient Image-to-Video Transfer Learning for Video Temporal Grounding/148e844e-b528-47e5-9db4-f50b0b9b4b9c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e0ca74e8d4907b906ba790cdeccd41953bc42d185d85843aa1455d4de5895f60 +size 5676505 diff --git a/2024/R^2-Tuning_ Efficient Image-to-Video Transfer Learning for Video Temporal Grounding/full.md b/2024/R^2-Tuning_ Efficient Image-to-Video Transfer Learning for Video Temporal Grounding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6cb2f2b60b4ef7339e90f20471ba01f78626b068 --- /dev/null +++ b/2024/R^2-Tuning_ Efficient Image-to-Video Transfer Learning for Video Temporal Grounding/full.md @@ -0,0 +1,350 @@ +# $\mathbf{R}^2$ -Tuning: Efficient Image-to-Video Transfer Learning for Video Temporal Grounding + +Ye Liu $^{1,3*}$ , Jixuan He $^{2,3}$ , Wanhua Li $^{3\dagger}$ , Junsik Kim $^{3}$ , Donglai Wei $^{4}$ , Hanspeter Pfister $^{3}$ , and Chang Wen Chen $^{1\dagger}$ + +1 The Hong Kong Polytechnic University 2 Tsinghua University 3 Harvard University 4 Boston College coco.ye.liu@connect.polyu.hk hejixuan2000@gmail.com wanhua@seas.harvard.edu mibastro@gmail.com donglai.wei@bc.edu pfister@seas.harvard.edu changwen.chen@polyu.edu.hk + +Abstract. Video temporal grounding (VTG) is a fine-grained video understanding problem that aims to ground relevant clips in untrimmed videos given natural language queries. Most existing VTG models are built upon frame-wise final-layer CLIP features, aided by additional temporal backbones (e.g., SlowFast) with sophisticated temporal reasoning mechanisms. In this work, we claim that CLIP itself already shows great potential for fine-grained spatial-temporal modeling, as each layer offers distinct yet useful information under different granularity levels. Motivated by this, we propose Reversed Recurrent Tuning (R²-Tuning), a parameter- and memory-efficient transfer learning framework for video temporal grounding. Our method learns a lightweight R² Block containing only 1.5% of the total parameters to perform progressive spatial-temporal modeling. Starting from the last layer of CLIP, R² Block recurrently aggregates spatial features from earlier layers, then refines temporal correlation conditioning on the given query, resulting in a coarse-to-fine scheme. R²-Tuning achieves state-of-the-art performance across three VTG tasks (i.e., moment retrieval, highlight detection, and video summarization) on six public benchmarks (i.e., QVHighlights, CharadesSTA, Ego4D-NLQ, TACoS, YouTube Highlights, and TVSum) even without the additional backbone, demonstrating the significance and effectiveness of the proposed scheme. Our code is available at https://github.com/yeliudev/R2-Tuning. + +Keywords: Video Temporal Grounding $\cdot$ Transfer Learning $\cdot$ CLIP + +# 1 Introduction + +Video is becoming the major content media in our daily lives. The variety of video domains has extended beyond daily activities [5,22] but broader scenes such as egocentric [10] and surveillance [60]. They maintain large information + +![](images/0545960a4f1b6cc8b4bd213c5958d2aaa6b0c29282b65178a9db990f60ef05fe.jpg) +Untrimmed Video + +![](images/44cc9f9c5c0dcc1f7cbc6f5eb8c68a404c5887900ee1dadc46f3088950515ee1.jpg) +Fig. 1: Video temporal grounding (VTG) contains three video-language understanding problems, i.e., moment retrieval (MR), highlight detection (HD), and video summarization (VS). + +![](images/909278c2716e565a2c52590c41482aea87c5dc7ae079358d20f589a7c12af3c1.jpg) +Fig. 2: Moment retrieval mAP with different backbones on QVHighlights val split. CLIP's potential of temporal modeling was not fully exploited. + +capacities within multi-granularities, and thus can convey both high-level context and low-level details effectively. + +In the content production and consumption industry, such a flourishing is accompanied by the increasing demand for browsing untrimmed videos efficiently with different user interests. This derives the topic of video temporal grounding (VTG) [26], which is a fine-grained video-language understanding problem aiming to ground video clips conditioning on natural language queries. As shown in Figure 1, VTG can be disentangled into three tasks based on different output formats, i.e., moment retrieval (MR) [9, 10, 22, 31] for regressing start-end timestamps, highlight detection (HD) [22, 47] for predicting frame-level saliency curves, and video summarization (VS) [2, 11, 45] for classifying disjoint shots. + +Encouraged by the recent success in adopting vision-language models (VLMs) for video understanding [4,15,19,28,32,37], most existing VTG methods [16,22, 31,34,35,46,56] are built upon frame-wise final-layer features from CLIP [41]. However, due to CLIP's misaligned pre-training objective (image-text contrast), these methods fail to capture temporal correlations well. As a feasible compromise, an additional backbone (e.g., SlowFast [8]) is incorporated to complement the temporal information, followed by carefully designed modules such as text-guided queries [31], dynamic anchors [35], and event reasoning [16]. + +We refer to the paradigm above as a post-processing scheme (shown in Figure. 3 (a)), whereas two natural drawbacks exist due to the sub-optimal design. First, leveraging two backbones with similar capabilities is unintuitive and inefficient during inference. A single model with both vision-text alignment and spatial-temporal modeling abilities is more preferred. Second, queries for VTG could be of different granularities from coarse (e.g., the family is traveling) to fine (e.g., when did I take the golf club from the man with white hair). Leveraging only frame-wise & final-layer features is not granularity flexible, as it would focus more on high-level scene transitions while overlooking low-level details. Preliminary experiments in Figure 2 also demonstrate that the potential of spatial-temporal modeling for CLIP is not fully exploited by existing methods. Some recent works [15,19,57] tried to tackle the first problem by fine-tuning (part of) the CLIP encoders (Figure 3 (b) and (c)), but they are all inefficient in terms of data, learnable parameters, or memory. + +![](images/c47d92be81ded838e7ff94e79816dbc57933c7ede699e37d1c1477ee0e80e728.jpg) +Fig. 3: Different architectural designs for CLIP-based image-to-video transfer learning. The gray rectangle in (d) denotes the progressively refined spatial-temporal features. + +![](images/ebd9a2205c8d88f5471fa446e5f8719ae9d78c0686598d11fc630beb91c9859b.jpg) + +![](images/6213e6af772ca6d4f3ef81f729e6fc1f7990cef419bc6273a4b2b3dbd2105c66.jpg) + +![](images/4bd2d4ddef0e16fb72e7c86773e86ed259cba3f3517bfa6c02ddf98c7e2969e3.jpg) + +This paper intends to answer the research question: how to efficiently transfer an image-language foundation model to video temporal grounding? We consider the above question by exploring two aspects: efficiency and granularity flexibility. To address these issues, a novel image-to-video transfer learning framework called Reversed Recurrent Tuning (R²-Tuning) is proposed for fine-grained understanding on untrimmed videos. Our insight is that multi-layer CLIP features offer distinct yet useful information, while their integration should be tailored to the downstream task. As illustrated in Figure 3 (d), based on a frozen CLIP [41], our method learns a side-block (R² Block) containing only $1.5\%$ of the total parameters to perform spatial-temporal modeling. R² Block is recurrently attached to the last few layers of CLIP encoders from back to front, performing query-modulated spatial pooling and recurrent temporal refinement from coarse to fine. Aside from the new architecture, we also introduce video-level and layer-wise constraints to calibrate the granularities of visual and text encoders. During training, gradient flows do not pass through the CLIP encoders, thus our scheme is both parameter- and memory-efficient. It is also granularity-flexible as R² Block can adaptively control the spatial pooling strategy conditioning on queries. + +We conduct extensive experiments across three VTG tasks on six public benchmarks, including QVHighlights [22], Ego4D-NLQ [10], Charades-STA [9], TACoS [9], YouTube Highlights [47], and TVSum [45]. Without bells and whistles, $\mathbb{R}^2$ -Tuning achieves more than 3 MR mAP gain on QVHighlights test split compared with $4\times$ heavier counterparts [34] with additional temporal backbones and carefully designed post-processing modules. Further analysis also shows that our method can better handle multi-granularity information. Overall, our contributions are summarized as: (1) We introduced $\mathbb{R}^2$ -Tuning, a novel image-to-video transfer learning framework tailored for video temporal grounding. (2) We designed two effective strategies, i.e., query-modulated spatial pooling and recurrent temporal refinement, to model spatial-temporal information from coarse to fine. (3) To calibrate the granularities of CLIP visual and text encoders, we further introduce video-level and layer-wise contrastive constraints to distill distinct information from each layer. (4) Extensive experiments across three tasks on six benchmarks demonstrate the significance and effectiveness of $\mathbb{R}^2$ -Tuning. + +# 2 Related Work + +CLIP for Video Understanding With powerful transfer learning abilities, CLIP [41] has been widely used for many image-language tasks [18, 29], and how to extend them to video understanding is an emerging and crucial research topic. Early attempts have been made to transfer CLIP to trimmed videos for action recognition [28,37,42] and video-text retrieval [4,15,19,32]. Specifically, Efficient-Prompt [19] introduces learnable prompt tokens to the text encoder, and ViFi-CLIP [42] further adopts learnable prompts at the visual encoder. Other works also tried to learn external modules designed for temporal modeling. EVL [28] proposes a temporal decoder in parallel with the main network. CLIP4Clip [32] investigated multiple temporal fusion strategies for video frames. Nevertheless, how to effectively transfer CLIP to untrimmed videos remains unexplored. To our best knowledge, $\mathbb{R}^2$ -Tuning is the first solution for fine-grained video temporal grounding via memory-efficient transfer learning. + +Video Temporal Grounding Video temporal grounding (VTG) [26] is a bundle of video understanding problems including moment retrieval (MR) [9,16,22, 31,35,36,52,56,57,61], highlight detection (HD) [3,14,47,55], and video summarization (VS) [11,45]. Although these tasks are closely related, they have not been jointly studied until recently. QVHighlights [22] is the first dataset that supports both MR and HD, breaking the moment distribution bias in previous MR datasets [9,43]. The authors also proposed Moment-DETR, a strong baseline method for MR/HD. Most following works were built upon the Moment-DETR framework with specially designed modules. Notably, UMT [31] tackles flexible MR and HD through a unified multi-modal architecture. QD-DETR [35] benefits from query-dependent video representations and dynamic anchors. EaTR [16] incorporates moment reasoning to provide reliable referential search areas for moment queries. Nevertheless, their design space is still within pre-extracted CLIP [41] and SlowFast [8] features. A recent work [57] revealed the possibility of fine-tuning CLIP for MR, while their solution (post pre-training) is extremely resource-consuming. Compared with existing works, $\mathbb{R}^2$ -Tuning can effectively learn strong temporal modeling abilities without any pre-training on videos. + +Parameter-Efficient Transfer Learning Foundation models [41, 49] have achieved remarkable successes in vision and language understanding. Efforts have been made to efficiently transfer knowledge from these models to new scenarios. One line of research is to conduct context-based tuning (e.g., prompt-tuning [17, 65, 66]), which does not modify the model architectures but incorporates learnable embeddings in the inputs. Another paradigm is introducing an extra lightweight adapter to the original model, and keeping the rest of the parameters frozen [39]. Recent works also consider memory-efficient transfer learning (METL) that introduces parameters whose gradient flows do not go through the main model [28, 40, 48]. $\mathbf{R}^2$ -Tuning is also a METL framework, where we adopt a novel recurrent tuning strategy to ensure parameter- and memory-efficient. + +![](images/23773a00ff20ed68d2af5e26b0a9008d233cec11fe04d5e92343f9d9b6060da3.jpg) +Fig. 4: Overall architecture of our framework. The input video and query are first encoded by frozen CLIP [41] encoders. Their multi-layer outputs are then recurrently fused and refined by a learnable $\mathbb{R}^2$ Block to construct spatial-temporal representations $h$ , which would be scaled up/down to construct a temporal feature pyramid, followed by three heads for MR, HD, and VS, respectively. + +# 3 Methodology + +# 3.1 Problem Formulation + +Given a video $V = \{v_{i}\}_{i=1}^{T}$ and a natural language query $Q = \{q_{i}\}_{i=1}^{L}$ , where $T$ and $L$ are the numbers of video frames and text tokens, VTG aims to densely predict a set of labels $(b_{i}, s_{i}, f_{i})$ for each frame, defined as follows: + +- Moment Retrieval is to find the most relevant moments (i.e. sets of consecutive frames) in $V$ according to $Q$ , so that $b_{i} = [b_{i}^{s}, b_{i}^{e}] \in \mathbb{R}^{2}$ represents the temporal displacements from frame $v_{i}$ to the start and end timestamps of the nearest target moment. +- Highlight Detection requires estimating frame-level relevancies between $V$ and $Q$ , thus $s_i \in [0,1]$ is a continuous saliency score denoting to what extent frame $v_i$ is semantically aligned with $Q$ . +- Video Summarization aims to select a subset of video frames according to $Q$ to form a concise summary, therefore, $f_{i} \in \{0,1\}$ is a binary score indicating whether frame $v_{i}$ belongs to the summary. + +# 3.2 Overview + +Figure 4 shows an overview of the proposed framework. Our model derives from a pre-trained and frozen CLIP [41] with ViT [6] backbone, which has a two-stream architecture for spatial and textual encoding, respectively. A learnable $\mathbb{R}^2$ block is iteratively attached to the last $K$ encoder layers to refine the temporal correlation. The resulting features are then scaled up/down to build a feature pyramid, followed by three heads to decode the task-specific predictions. + +Specifically, the input video $V$ and query $Q$ are first tokenized into frame patches and word tokens, then sent into the visual encoder $E_{v}$ and query encoder $E_{q}$ . The encoded visual and textual features can be denoted as $e_v \in$ + +![](images/9a849fabc871996edfcd43ef75c076d8f644dcff3fb91447fbab73a7bbb56a50.jpg) +Fig. 5: Detailed architecture of the $\mathbb{R}^2$ Block. It can be split into two parts: a) query-modulated spatial pooling, and b) recurrent temporal refinement. Note that the [CLS] tokens of visual features are omitted for clarity. + +$\mathbb{R}^{B \times N \times T \times (P + 1) \times D_v}$ and $e_q \in \mathbb{R}^{B \times N \times L \times D_q}$ , where $B, N, T, P, L, D_v$ and $D_q$ indicate batch size, number of encoder layers, number of video frames, number of patches per frame, number of query tokens, and the dimensions of visual/query features, respectively. These features are recurrently fused and refined by $\mathbb{R}^2$ block to construct spatial-temporal features $h \in \mathbb{R}^{B \times T \times C}$ , in which each token preserves the $C$ -dimensional features for a frame. This process will be introduced in detail in Section 3.3. + +# 3.3 Reversed Recurrent Tuning + +Given pre-trained visual and query encoders, existing works [16,22,26,31,35] merely take features from the last hidden layer, which is sub-optimal due to the limited information. In this work, we propose $\mathbb{R}^2$ -Tuning to exploit the potential of spatial-temporal modeling based on multi-layer CLIP features. Our scheme learns a lightweight $\mathbb{R}^2$ Block which maintains a zero-initialized hidden state $h \in \mathbb{R}^{B \times T \times C}$ as frame-level spatial-temporal features. The $\mathbb{R}^2$ Block is recurrently attached to the last $K$ layers of CLIP encoders from back to front to refine the hidden state $h$ for $K$ steps, with each step defined as follows: + +$$ +h ^ {k} = \mathcal {F} _ {\theta} (e _ {v} ^ {n}, e _ {q} ^ {n}, h ^ {k - 1}) \tag {1} +$$ + +Here, $\mathcal{F}_{\theta}$ is the refinement operation parameterized by $\theta$ . $k \in [1, K]$ is the index of current step. $n = N - k + 1$ denotes the CLIP layer index for step $k$ , thus $e_v^n$ and $e_q^n$ are visual and query features from the $n$ -th CLIP encoder layer. The refinement operation $\mathcal{F}_{\theta}$ can be decomposed into two parts: 1) query-modulated spatial pooling, and 2) recurrent temporal refinement. Details are discussed as follows. + +Query-Modulated Spatial Pooling Figure 5 (a) presents the data flow of query-modulated spatial pooling. The goal of this process is to adaptively pool + +spatial features from patch-level representations $e_v^n$ to a single token $e_{pool}^n$ conditioning on the query $e_q^n$ . We first adopt two separate MLPs to map $e_v^n$ and $e_q^n$ into the same space: + +$$ +\hat {e} _ {v} ^ {n} = \operatorname {M L P} \left(e _ {v} ^ {n}\right) \in \mathbb {R} ^ {B \times T \times (P + 1) \times C} \tag {2} +$$ + +$$ +\hat {e} _ {q} ^ {n} = \operatorname {M L P} \left(e _ {q} ^ {n}\right) \in \mathbb {R} ^ {B \times L \times C} \tag {3} +$$ + +Here, $C$ is the reduced hidden size. We further align the shape of $\hat{e}_v^n$ and $\hat{e}_q^n$ by merging dimension $T$ into $B$ and repeat $\hat{e}_q^n$ for $T$ times in an interleaved style, resulting $\hat{e}_v^n \in \mathbb{R}^{(B \times T) \times (P + 1) \times C}$ and $\hat{e}_q^n \in \mathbb{R}^{(B \times T) \times L \times C}$ . Then, we compute the similarities for patch-token pairs using normalized Embedded Gaussian [50]: + +$$ +a = \operatorname {s o f t m a x} \left(\frac {\left(w _ {q} \hat {e} _ {q} ^ {n}\right) ^ {\top} w _ {v} \hat {e} _ {v} ^ {n}}{\sqrt {C}}\right) \in \mathbb {R} ^ {(B \times T) \times L \times P} \tag {4} +$$ + +Here, $w_{v}$ and $w_{q}$ are learnable matrices for projecting features. This operation derives from cross-attention but discards the linear projection on value. We then multiply the similarities $a$ with visual features $\hat{e}_{v}^{n}$ to pool them into each token, such that tokens can interact with patches independently. After that, a max pooling along the $L$ dimension is applied: + +$$ +e _ {t o k e n} ^ {n} = \operatorname {M a x P o o l i n g} \left(a \hat {e} _ {v} ^ {n} + \hat {e} _ {q} ^ {n}\right) \in \mathbb {R} ^ {B \times T \times C} \tag {5} +$$ + +Here, $\hat{e}_q^n$ serves as residuals [13] to stabilize training. Our visualizations in Figure 8 show that this strategy can guide the model to focus more on query-related regions. Finally, we combine $e_{token}^n$ with the [CLS] token to generate query-modulated spatial features: + +$$ +e _ {p o o l} ^ {n} = e _ {v} ^ {n, 0} + g ^ {k} \cdot e _ {t o k e n} ^ {n} \in \mathbb {R} ^ {B \times T \times C} \tag {6} +$$ + +Here, $g^{k} \in (-1,1)$ is a zero-initialized learnable gate for step $k$ constrained by Tanh. We allow negative values here to remove useless information from [CLS] tokens. $e_{pool}^{n}$ is then used to model temporal correlations and are refined in a recurrent manner. + +Recurrent Temporal Refinement Figure 5 (b) displays the flow of recurrent temporal refinement. Specifically, the pooled visual features $e_{pool}^{n}$ is first fused with the hidden state $h^{k-1}$ from the previous step: + +$$ +\hat {h} ^ {k - 1} = \varphi^ {k} \cdot e _ {\text {p o o l}} ^ {n} + (1 - \varphi^ {k}) \cdot h ^ {k - 1} \tag {7} +$$ + +where $\varphi^k\in (0,1)$ is the learnable gate for step $k$ . Then, we adopt a sequence of standard multi-head cross-attention $(\hat{e}_q^n$ as $k,v)$ , multi-head self-attention, followed by a feed-forward network [50] to update the hidden state: + +$$ +h ^ {k} = \operatorname {F F N} (\mathrm {M H S A} (\mathrm {M H C A} (\hat {h} ^ {k - 1}, \hat {e} _ {q} ^ {n}))) \tag {8} +$$ + +For each block, we adopt DropPath [21] with $p = 0.1$ to prevent overfitting. The query features $\hat{e}_q^n$ are utilized both in spatial and temporal modeling as guidance. + +Granularity Calibration The refinement process above is based on the assumption that visual and query features from the same layer of CLIP [41] are well aligned at the same granularity level. However, this cannot be guaranteed as the two encoders are learned in isolation during pre-training. Therefore, we need to add manual constraints to calibrate their granularities. + +We apply a video-level constraint and a layer-wise constraint by designing two contrastive losses. We first denote $e_v^0 \in \mathbb{R}^{B \times K \times T \times C}$ as the features of $K$ [CLS] tokens from Eq. 6. Then, we select all positive frames (i.e., frames within $[b_i^s, b_i^e]$ or with $s_i$ or $f_i$ higher than a threshold) from $e_v^0$ and perform average pooling to obtain video-level representations $\widetilde{e}_v \in \mathbb{R}^{B \times K \times C}$ . + +$$ +\widetilde {e} _ {v} = \operatorname {A v g P o o l} \left(\left\{e _ {v} ^ {0, i} \right\} _ {i \in \Omega}\right) \tag {9} +$$ + +Here, $\Omega$ is the set of positive frame indices. To obtain query-level representations $\widetilde{e}_q\in \mathbb{R}^{B\times K\times C}$ , a token-wise adaptive pooling is applied following previous work [26]. We adopt InfoNCE loss [38] along two dimensions to calculate the video-level and layer-wise constraints: + +$$ +\mathcal {L} _ {\text {v i d e o}} = \lambda_ {\text {v i d e o}} \frac {1}{B} \sum_ {b \in B} \operatorname {I n f o N C E} \left(\tilde {e} _ {v} ^ {b}, \tilde {e} _ {q} ^ {b}\right) \tag {10} +$$ + +$$ +\mathcal {L} _ {\text {l a y e r}} = \lambda_ {\text {l a y e r}} \frac {1}{K} \sum_ {k \in K} \operatorname {I n f o N C E} \left(\widetilde {e} _ {v} ^ {k}, \widetilde {e} _ {q} ^ {k}\right) \tag {11} +$$ + +Here, $\mathcal{L}_{video}$ performs contrast among samples in the same batch (diversifying features are diverse among moment-query pairs), and averages the loss across $K$ layers. It also implicitly models temporal correlations between videos and queries. $\mathcal{L}_{layer}$ performs contrast among layers and averages across the batch. $\lambda_{video}$ and $\lambda_{layer}$ serve as re-weighting terms for the losses. A fixed temperature parameter of 0.07 is used. + +# 3.4 Prediction Heads + +After refining spatial-temporal features $h$ , a temporal feature pyramid is constructed by applying 1D convolutions with stride $= 2$ on $h$ . The number of convolutions used for each level is subjected to $l - 1$ , where $l$ is the level index starting from 1. We concatenate features from all levels by the temporal dimension to form $\hat{h}$ and predict once in all heads. Following a similar but simplified design from [26], we adopt three heads for video temporal grounding, described in detail as follows. + +Foreground-Background Classification Given $\hat{h}$ , a two-layer 1D convolution module with $(1\times 3)$ kernels is adopted to predict $\hat{f}_i$ for each frame. It is optimized using a focal loss [27] with $\alpha = 0.9$ and $\gamma = 2.0$ : + +$$ +\mathcal {L} _ {c l s} = - \lambda_ {c l s} \alpha (1 - \hat {f} _ {i}) ^ {\gamma} \log (\hat {f} _ {i}) \tag {12} +$$ + +We find that focal loss perform better than binary cross-entropy when the numbers of foreground/background frames are severely imbalanced as noted in [26]. + +Table 1: Video moment retrieval (MR) and highlight detection (HD) results on QVHighlights test split. Note that [16] and [57] only reported their results on val split. $\bigcirc$ and $\bigcirc$ are CLIP-B/32 and B/16 [41], $\bigcirc$ denotes SlowFast R-50 [8], and $\bigcirc$ is PANN [20] for audio features extraction. * means estimated parameters. The best and second-best metrics are marked with bold and underline, respectively. + +
MethodBackbonePost +Pre-trainMRHD
R1mAP≥ Very Good#Params
@0.5@0.7@0.5@0.75Avg.mAPHIT@1
BeautyThumb [44]○○X-----14.3620.88-
DVSE [30]○○X-----18.7521.79-
MCN [1]○○X11.412.7224.948.2210.67---
CAL [7]○○X25.4911.5423.407.659.89---
XML [23]○○X41.8330.3544.6331.7332.1434.4955.25-
XML+ [22]○○X46.6933.4647.8934.6734.9035.3855.06-
Moment-DETR [22]○○X52.8933.0254.8229.4030.7335.6955.604.8M
UMT [31]○○○X56.2341.1853.8337.0136.1238.1859.9914.9M
MomentDiff [24]○○○X58.2141.4854.5737.2136.84---
QD-DETR [35]○○X62.4044.9862.5239.8839.8638.9462.407.6M
MH-DETR [56]○○X60.0542.4860.7538.1338.3838.2260.518.2M
UniVTG [26]○○X58.8640.8657.6035.5935.4738.2060.9641.3M
TR-DETR [46]○○X64.6648.9663.9843.7342.6239.9163.427.9M
CG-DETR [34]○○X65.4348.3864.5142.7742.8640.3366.2112.0M
EaTR [16]○○X61.3645.7961.8641.9141.7437.1558.659.0M
Moment-DETR [22]○○236K ASR Cap.59.7840.3360.5135.3636.1437.4360.174.8M
UMT [31]○○○236K ASR Cap.60.8343.2657.3339.1238.0839.1262.3914.9M
UniVTG [26]○○4.2M Corpus65.4350.0664.0645.0243.6340.5466.2841.3M
UnLoc [57]○650K Videos64.5048.80-----87.9M*
R²-Tuning (Ours)○X68.0349.3569.0447.5646.1740.7564.202.7M
+ +Boundary Regression Similar to foreground-background classification, boundary regression for moments is also realized by a two-layer 1D convolution module. The difference is that its output dimension is set to 2 instead of 1, representing the boundary displacements $[\hat{b}_i^s,\hat{b}_i^e ]$ for start-end timestamps. This head is optimized by an L1 loss: + +$$ +\mathcal {L} _ {r e g} = \lambda_ {r e g} \left(\left| b _ {i} ^ {s} - \hat {b} _ {i} ^ {s} \right| + \left| b _ {i} ^ {e} - \hat {b} _ {i} ^ {e} \right|\right) \tag {13} +$$ + +where $b_{i}^{s}$ and $b_{i}^{e}$ are the ground truths. This loss is only applied to frames inside the ground truth boundaries. We observe that using an L1 loss already works better than the combination of Smooth L1 Loss and GIoU Loss as in [26]. + +Saliency Prediction To obtain the saliency scores $\hat{s}_i$ for HD, we calculate the cosine similarities between the adaptive pooled query features at the last refinement step $\widetilde{e}_q^K$ and each token in the spatial-temporal features $h$ : + +$$ +\hat {s} _ {i} = \frac {h _ {i} ^ {\top} \widetilde {e} _ {q} ^ {K}}{| | h _ {i} | | _ {2} | | \widetilde {e} _ {q} ^ {K} | | _ {2}} \tag {14} +$$ + +The training objective is applied through a contrastive loss between sampled positive frames (with index $p \in P$ ) and the adaptively pooled query $\widetilde{e}_q^K$ : + +$$ +\mathcal {L} _ {s a l} = - \lambda_ {s a l} \log \frac {\exp \left(\hat {s} _ {p} / \tau\right)}{\exp \left(\hat {s} _ {p} / \tau\right) + \sum_ {i \in \Theta} \exp \left(\hat {s} _ {i} / \tau\right)} \tag {15} +$$ + +Table 2: Video moment retrieval results on Ego4D-NLQ, Charades-STA, and TACoS datasets. $\bigcirc$ means using both CLIP-B/32 and SlowFast R-50 as feature extractors. $\bigcirc$ indicates using CLIP-B/32 only. The best and second-best metrics are marked with bold and underline, respectively. Our method shows significant advantages in high-quality retrievals (R@0.7) even without a video backbone. + +
MethodEgo4D-NLQ [10]Charades-STA [9]TACoS [43]
R@0.3R@0.5R@0.7mIoUR@0.3R@0.5R@0.7mIoUR@0.3R@0.5R@0.7mIoU
○2D-TAN [64]4.331.830.603.3958.7646.0227.5041.2540.0127.9912.9227.22
○VSLNet [62]4.542.401.013.5460.3042.6924.1441.5835.5423.5413.1524.99
○Moment-DETR [22]4.341.810.653.5365.8352.0730.5945.5437.9724.6711.9725.49
○UniVTG [26]7.283.951.324.9170.8158.0135.6550.1051.4434.9717.3533.60
○R²-Tuning (Ours)7.204.492.124.9470.9159.7837.0250.8649.7138.7225.1235.92
+ +Here, $\Theta$ is the set of frame indices where $s_i < s_p$ , and $\tau$ is a fixed temperature parameter set to 0.07. + +Training & Inference The whole model is jointly optimized using a sum of the five losses mentioned above. In practice, the loss weights are set as $\lambda_{video} = 0.1$ , $\lambda_{layer} = 0.1$ , $\lambda_{cls} = 1.0$ , $\lambda_{reg} = 0.1$ , and $\lambda_{sal} = 0.1$ . During inference, we form an moment retrieval prediction by combining $f_{i}$ and $b_{i}$ , i.e., calculating the start-end timestamps from frame index $i$ and boundary displacements $[b_i^s, b_i^e]$ , while regarding $f_{i}$ as the confidence. NMS with IoU threshold $\theta_{IoU} = 0.7$ is applied to reduce duplicates. For highlight detection and video summarization, we directly use the frame-level output from saliency prediction $s_i$ . + +# 4 Experiments + +# 4.1 Datasets & Evaluation Metrics + +We conduct experiments on six datasets covering various domains including daily vlogs & news (QVHighlights [22]), in-door scenes (TACoS [43] and Charades-STA [9]), egocentric videos (Ego4D-NLQ [10]), and sports (YouTube Highlights [47] and TVSum [45]). Details about the datasets are provided in the supplementary material. + +We adopt the same evaluation metrics with previous works [22, 26, 31]. To be specific, we compute Recall@1 with IoU threshold $\theta_{IoU} = 0.5$ and 0.7, mean average precision (mAP) with $\theta_{IoU} = 0.5$ and 0.7, and mAP with a series of thresholds [0.5:0.05:0.95] for MR on QVHighlights [22]. mAP and HIT@1 where positive samples are defined as with the saliency score of Very Good are adopted for HD. On Ego4D-NLQ [10], Charades-STA [9], and TACoS [43] datasets, we utilize Recall@1 with $\theta_{IoU} = \{0.3, 0.5, 0.7\}$ and mIoU to measure the MR performance. On YouTube Highlights [47] for HD and TVSum [45] for VS, we follow [26, 31] and use the same train/val split with mAP and Top-5 mAP as metrics, respectively. + +Table 3: Class-wise video highlight detection results (mAP) on YouTube Highlights dataset. $\bigcirc$ , $\bigcirc$ , and $\bigcirc$ denote image, video, and audio backbones, respectively. + +
MethodDogGym.Par.Ska.Ski.Sur.Avg.
○ RRAE [58]49.035.050.025.022.049.038.3
○ GIFs [12]30.833.554.055.432.854.146.4
○ LSVM [47]60.041.061.062.036.061.053.6
○ LIM-S [54]57.941.767.057.848.665.156.4
○ SL-Module [55]70.853.277.272.566.176.269.3
○ PLD [53]74.970.277.957.570.779.073.0
○ MINI-Net [14]58.261.770.272.258.765.164.4
○ TCG [59]55.462.770.969.160.159.863.0
○ Joint-VA [3]64.571.980.862.073.278.371.8
○ CO-AV [25]60.966.089.074.169.081.174.7
○ UMT [31]65.975.281.671.872.382.774.9
○ UniVTG [26]71.876.573.973.373.282.275.2
○ R²-Tuning75.673.573.074.674.884.876.1
+ +Table 4: Class-wise video summarization results (Top-5 mAP) on TVSum dataset. $\odot$ , $\odot$ , and $\odot$ denote image, video, and audio backbones, respectively. Our method does not require any additional video/audio features. + +
MethodVTVUGAMSPKPRFMBKBTDSAvg.
○ sLSTM [63]41.146.246.347.744.846.145.240.647.145.545.1
○ SG [33]42.347.247.548.945.647.346.441.748.346.646.2
○ LIM-S [54]55.942.961.254.060.447.543.266.369.162.656.3
○ Trailer [51]61.354.665.760.859.170.158.264.765.668.162.8
○ SL-Module [55]86.568.774.986.279.063.258.972.678.964.073.3
○ PLD [53]84.580.970.372.576.487.271.974.074.479.177.1
○ MINI-Net [14]80.668.378.281.878.165.857.875.080.265.573.2
○ TCG [59]85.071.481.978.680.275.571.677.378.668.176.8
○ Joint-VA [3]83.757.378.586.180.169.270.073.097.467.576.3
○ CO-AV [25]90.872.884.685.078.378.072.877.189.572.380.1
○ UMT [31]87.581.588.278.881.487.076.086.984.479.683.1
○ UniVTG [26]83.985.189.080.184.681.470.991.773.569.381.0
○ R²-Tuning85.085.991.081.788.887.478.189.290.374.785.2
+ +# 4.2 Implementation Details + +In all experiments, we adopt the visual and text encoders of CLIP ViT-B/32 [41] as our backbones. The whole CLIP model is frozen during training, i.e., only the parameters in $\mathbb{R}^2$ Block, temporal feature pyramid, prediction heads are learnable. Without further specification, we set $K = 4$ , i.e., attaching $\mathbb{R}^2$ Block to the last 4 layers of CLIP encoders. In $\mathbb{R}^2$ Block, the hidden size is set to 256. Only one transformer layer with post-norm style and 8 attention heads is used for temporal modeling. The number of layers for the temporal feature pyramid is set to 4 for QVHighlights, Ego4D-NLQ, Charades-STA, and TACoS, while 1 for other datasets. Please refer to the supplementary material for details. + +# 4.3 Comparison with State-of-the-Arts + +Joint Moment Retrieval and Highlight Detection We first evaluate our method on QVHighlights [22] test split, the only dataset that supports both + +![](images/1965b07216fe806ed6b3dc139235ce5e0991176c050cdbcb7668eead2bfb9fa6.jpg) +QUERY: A man in white is browsing through clothes at a department store. + +![](images/af00914a1c1c11765454382834f2ba07a7ad8ee70f499d2e86dfc101f4d57220.jpg) +QUERY: Man stands on playground balance bridge. +Fig.6: Visualization of joint moment retrieval and highlight detection results on QVHighlights val split. We compare our method with UniVTG [26], which is a strong baseline leveraging both additional temporal backbone and large-scale post pretraining. $\mathbb{R}^2$ -Tuning can accurately regress the boundaries of moments and predict highlight saliency scores through its novel feature refinement design. + +moment retrieval and highlight detection. The results are shown in Table 1. The first group of methods utilizes more than one feature extractor. The second group is that with large-scale post pre-training (defined as the extra pre-training stage between loading backbone weights and training on downstream datasets.). Without any additional backbone and extra pre-training, $\mathbf{R}^2$ -Tuning achieves start-of-the-art performance compared with all previous methods with the fewest learnable parameters. Figure 6 visualizes the model predictions. + +Moment Retrieval We then evaluate our model on moment retrieval task in egocentric [10] and in-door [9, 43] domains. The results are shown in Table 2. We follow [26] and compare only with the methods using CLIP + SlowFast features. R²-Tuning still works better than all baselines with extra features. We also observe that R²-Tuning shows its significance on high-quality retrievals (R1@0.7), which requires accurate temporal modeling. This benefits from the fine-grained temporal modeling ability of reversed and recurrent designs. + +Highlight Detection & Video Summarization The performances of highlight detection on YouTube Highlights [47] and extractive video summarization on TVSum [45] are reported in Table 3 and Table 4, respectively. Following previous works [26, 31], we trained the model separately on each domain. The first group of methods is based on video backbones, which naturally have basic temporal modeling abilities. The second group of methods is enhanced by extra features such as image and audio. $\mathbf{R}^2$ -Tuning can still reach the best performance when trained on small-scale data. + +![](images/b098f12db1afd9c6126afa7b3b43e782d7d214823744117017c0329b18d020d3.jpg) +(a) Naive Fusion + +![](images/0e6de240e9b71af89bdea3aed47b4cc0fef6d4015cc7bd8e90467e9068ad9801.jpg) +Fig. 7: Different feature refinement strategies for MR on QVHighlights val split. Single means utilizing the $K$ -th layer only, while Mean denotes averaging the features from last $K$ layers. Forward and Reversed indicate different feature refinement directions. + +![](images/666ac9bc80ecce0d6fce7c0a478eaba4331c40b70d2266f75bf76a9716c928da.jpg) +(b) Recurrent Refinement + +![](images/106051b73f23d014df24ae663665f4a8411385db7f49bd4c94d0aa60f2d780a0.jpg) + +![](images/e40999bd62e249bae10d2ea3518aca5a5f9203fd71f7617dbdd3a0a57f4a0bc4.jpg) +Fig. 8: Visualization of the attention maps for query-modulated spatial pooling. Diverse regions are learned for different queries. + +![](images/fc12baa482f5c650a0433afd55e3c1eeee525a401b4429e6ef262c09407d0a8b.jpg) +Fig. 9: Cosine similarities of video (x-axis) and text (y-axis) features among $7\sim 12$ CLIP layers. + +# 4.4 Detailed Analysis + +Multi-Granularity Features To investigate the significance of granularity flexibility, we set up a simple baseline that only attaches $\mathbf{R}^2$ Block to one layer of CLIP encoders. Then, we make use of the multi-layer information by averaging the features from the last $K$ layers. These two variants are compared in Figure 7 (a). It can be observed that for the single-layer setting, higher-level features are more discriminative than lower-level ones. However, even a naive fusion strategy for the multi-layer information brings significant performance gains. + +Reversed Recurrent Feature Refinement We then further justify our design by comparing two refinement directions (i.e., forward and reversed) in Figure 7 (b). Note that when $K = 1$ , all four designs in Figure 7 are of the same architecture. When utilizing more layers, both forward and reversed refinement benefit from more information, while the reversed scheme statically performs better, as the multi-granularity features shall be refined from coarse to fine. + +Query-Modulated Spatial Pooling We visualize the attention maps learned by query-modulated spatial pooling in Figure 8. A query generally has multiple keywords that provide discriminative information to localize moments. The + +Table 5: Effectiveness justification of granularity calibration on QVHighlights val split. + +
L VideoL layerMRHD
R1@0.5R1@0.7mAPmAPHIT@1
64.4848.6044.0137.9462.67
✓67.6851.6146.7439.8165.16
✓64.7148.8444.6038.9163.35
✓✓68.7152.0647.5940.5964.32
+ +Table 6: MR mAP with different query lengths on QVHighlights val split. + +
Method#Words
1-1011-2021-3031-40≥41
QD-DETR [35]40.3042.3229.010.1026.67
EaTR [16]38.9439.5424.5812.0341.43
UniVTG [26]32.3832.6520.880.0031.11
R2-Tuning47.0849.1328.7967.2472.38
+ +token-wise attention in Eq. 4 can guide the model to focus on multiple regions simultaneously, as can be seen in the diverse but meaningful patterns. + +Granularity Calibration To verify the effectiveness of granularity calibration, we also visualize the cosine distances of all the visual-query pairs from $7 \sim 12$ layers in Fig. 9. Before calibration, the visual-text features are not well-aligned. Adding the video-level constraint makes higher-level features aligned with each other (b), while adding layer-wise constraint makes the model distill diverse information across different layers (c). The diagonal in (d) verifies that both calibration constraints can maximize the mutual similarities of paired visual and text features. Some quantitative results are presented in Table 5. Both video-level and layer-wise contrastive can help align the multi-granularity visual & query semantics, while their combination can further boost the performances. We also group the queries in QVHighlights into 5 bins with different lengths (acting as multi-granularities) and compute the means of MR mAP in Tab. 6. Although most training queries are coarse-grained ( $\leqslant 30$ words), R²-Tuning can easily generalize to fine-grained queries ( $\geqslant 31$ words), surpassing previous methods. + +# 5 Conclusion + +This paper introduces R²-Tuning, a parameter- and memory-efficient transfer learning framework for video temporal grounding. It learns a lightweight side-adapter (R² Block) that is recurrently attached to the last few layers of a frozen CLIP to adaptively pool spatial details and refine temporal correlations. Two contrastive constraints (video-level and layer-wise) are utilized to calibrate the granularities of CLIP visual and text encoders. Experiments across three VTG tasks on six public benchmarks demonstrate the significance and effectiveness of the proposed scheme. We hope that the proposed framework sparks further research on efficient image-to-video transfer learning for untrimmed videos. + +# Acknowledgments + +This work was supported in part by Hong Kong Research Grants Council GRF-15229423, US NIH grant R01HD104969, and NSF award IIS-2239688. + +# References + +1. Anne Hendricks, L., Wang, O., Shechtman, E., Sivic, J., Darrell, T., Russell, B.: Localizing moments in video with natural language. In: ICCV. pp. 5803-5812 (2017) +2. Apostolidis, E., Adamantidou, E., Metsai, A.I., Mezaris, V., Patras, I.: Video summarization using deep neural networks: A survey. Proceedings of the IEEE 109(11), 1838-1863 (2021) +3. Badamdorj, T., Rochan, M., Wang, Y., Cheng, L.: Joint visual and audio learning for video highlight detection. In: ICCV. pp. 8127-8137 (2021) +4. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: A clip-hitchhiker's guide to long video retrieval. Tech. Rep. arXiv:2205.08508 (2022) +5. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: CVPR. pp. 6299-6308 (2017) +6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2020) +7. Escorcia, V., Soldan, M., Sivic, J., Ghanem, B., Russell, B.: Temporal localization of moments in video collections with natural language. Tech. Rep. arXiv:1907.12763 (2019) +8. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: ICCV. pp. 6202-6211 (2019) +9. Gao, J., Sun, C., Yang, Z., Nevatia, R.: Tall: Temporal activity localization via language query. In: ICCV. pp. 5267-5275 (2017) +0. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: CVPR. pp. 18995-19012 (2022) +1. Gygli, M., Grabner, H., Riemenschneider, H., Van Gool, L.: Creating summaries from user videos. In: ECCV. pp. 505-520 (2014) +2. Gygli, M., Song, Y., Cao, L.: Video2gif: Automatic generation of animated gifs from video. In: CVPR. pp. 1001-1009 (2016) +3. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016) +4. Hong, F.T., Huang, X., Li, W.H., Zheng, W.S.: Mini-net: Multiple instance ranking network for video highlight detection. In: ECCV. pp. 345-360 (2020) +5. Huang, S., Gong, B., Pan, Y., Jiang, J., Lv, Y., Li, Y., Wang, D.: Vop: Text-video co-operative prompt tuning for cross-modal retrieval. In: CVPR. pp. 6565-6574 (2023) +6. Jang, J., Park, J., Kim, J., Kwon, H., Sohn, K.: Knowing where to focus: Event-aware transformer for video grounding. In: ICCV. pp. 13846-13856 (2023) +7. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: ECCV. pp. 709-727 (2022) +8. Jiang, R., Liu, L., Chen, C.: Clip-count: Towards text-guided zero-shot object counting. In: ACM MM (2023) +9. Ju, C., Han, T., Zheng, K., Zhang, Y., Xie, W.: Prompting visual-language models for efficient video understanding. In: ECCV. pp. 105-124 (2022) +20. Kong, Q., Cao, Y., Iqbal, T., Wang, Y., Wang, W., Plumbley, M.D.: Panns: Large-scale pretrained audio neural networks for audio pattern recognition. IEEE/ACM Transactions on Audio, Speech, and Language Processing 28, 2880-2894 (2020) +21. Larsson, G., Maire, M., Shakhnarovich, G.: Fractalnet: Ultra-deep neural networks without residuals. Tech. Rep. arXiv:1605.07648 (2016) + +22. Lei, J., Berg, T.L., Bansal, M.: Qvhighlights: Detecting moments and highlights in videos via natural language queries. In: NeurIPS (2021) +23. Lei, J., Yu, L., Berg, T.L., Bansal, M.: Tv: A large-scale dataset for video-subtitle moment retrieval. In: ECCV. pp. 447-463 (2020) +24. Li, P., Xie, C.W., Xie, H., Zhao, L., Zhang, L., Zheng, Y., Zhao, D., Zhang, Y.: Momentdiff: Generative video moment retrieval from random to real. Tech. Rep. arXiv:2307.02869 (2023) +25. Li, S., Zhang, F., Yang, K., Liu, L., Liu, S., Hou, J., Yi, S.: Probing visual-audio representation for video highlight detection via hard-pairs guided contrastive learning. Tech. Rep. arXiv:2206.10157 (2022) +26. Lin, K.Q., Zhang, P., Chen, J., Pramanick, S., Gao, D., Wang, A.J., Yan, R., Shou, M.Z.: Univtg: Towards unified video-language temporal grounding. In: CVPR. pp. 2794-2804 (2023) +27. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: ICCV. pp. 2980-2988 (2017) +28. Lin, Z., Geng, S., Zhang, R., Gao, P., de Melo, G., Wang, X., Dai, J., Qiao, Y., Li, H.: Frozen clip models are efficient video learners. In: ECCV. pp. 388-404 (2022) +29. Liu, L., Yu, B.X., Chang, J., Tian, Q., Chen, C.W.: Prompt-matched semantic segmentation. Tech. Rep. arXiv:2208.10159 (2022) +30. Liu, W., Mei, T., Zhang, Y., Che, C., Luo, J.: Multi-task deep visual-semantic embedding for video thumbnail selection. In: CVPR. pp. 3707-3715 (2015) +31. Liu, Y., Li, S., Wu, Y., Chen, C.W., Shan, Y., Qie, X.: Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In: CVPR. pp. 3042-3051 (2022) +32. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: Clip4clip: An empirical study of clip for end to end video clip retrieval. Neurocomputing 508, 293-304 (2022) +33. Mahasseni, B., Lam, M., Todorovic, S.: Unsupervised video summarization with adversarial LSTM networks. In: CVPR. pp. 202-211 (2017) +34. Moon, W., Hyun, S., Lee, S., Heo, J.P.: Correlation-guided query-dependency calibration in video representation learning for temporal grounding. Tech. Rep. arXiv:2311.08835 (2023) +35. Moon, W., Hyun, S., Park, S., Park, D., Heo, J.P.: Query-dependent video representation for moment retrieval and highlight detection. In: CVPR. pp. 23023-23033 (2023) +36. Nan, G., Qiao, R., Xiao, Y., Liu, J., Leng, S., Zhang, H., Lu, W.: Interventional video grounding with dual contrastive learning. In: CVPR. pp. 2765-2775 (2021) +37. Ni, B., Peng, H., Chen, M., Zhang, S., Meng, G., Fu, J., Xiang, S., Ling, H.: Expanding language-image pretrained models for general video recognition. In: ECCV. pp. 1-18 (2022) +38. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. In: NeurIPS (2018) +39. Pan, J., Lin, Z., Zhu, X., Shao, J., Li, H.: St-adapter: Parameter-efficient image-to-video transfer learning. In: NeurIPS. pp. 26462-26477 (2022) +40. Qing, Z., Zhang, S., Huang, Z., Zhang, Y., Gao, C., Zhao, D., Sang, N.: Disentangling spatial and temporal learning for efficient image-to-video transfer learning. In: CVPR. pp. 13934-13944 (2023) +41. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021) + +42. Rasheed, H., Khattak, M.U., Maaz, M., Khan, S., Khan, F.S.: Fine-tuned clip models are efficient video learners. In: CVPR. pp. 6545-6554 (2023) +43. Regneri, M., Rohrbach, M., Wetzel, D., Thater, S., Schiele, B., Pinkal, M.: Grounding action descriptions in videos. Transactions of the Association for Computational Linguistics 1, 25-36 (2013) +44. Song, Y., Redi, M., Vallmitjana, J., Jeimes, A.: To click or not to click: Automatic selection of beautiful thumbnails from videos. In: CIKM. pp. 659-668 (2016) +45. Song, Y., Vallmitjana, J., Stent, A., Jaimes, A.: Tvsum: Summarizing web videos using titles. In: CVPR. pp. 5179-5187 (2015) +46. Sun, H., Zhou, M., Chen, W., Xie, W.: Tr-detr: Task-reciprocal transformer for joint moment retrieval and highlight detection. In: AAAI (2024) +47. Sun, M., Farhadi, A., Seitz, S.: Ranking domain-specific highlights by analyzing edited videos. In: ECCV. pp. 787-802 (2014) +48. Sung, Y.L., Cho, J., Bansal, M.: Lst: Ladder side-tuning for parameter and memory efficient transfer learning. In: NeurIPS. pp. 12991-13005 (2022) +49. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. Tech. Rep. arXiv:2302.13971 (2023) +50. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS. pp. 5998-6008 (2017) +51. Wang, L., Liu, D., Puri, R., Metaxas, D.N.: Learning trailer moments in full-length movies. In: ECCV. pp. 300-316 (2020) +52. Wang, Z., Wang, L., Wu, T., Li, T., Wu, G.: Negative sample matters: A renaissance of metric learning for temporal grounding. In: AAAI. pp. 2613-2623 (2022) +53. Wei, F., Wang, B., Ge, T., Jiang, Y., Li, W., Duan, L.: Learning pixel-level distinctions for video highlight detection. In: CVPR. pp. 3073-3082 (2022) +54. Xiong, B., Kalantidis, Y., Ghadiyaram, D., Grauman, K.: Less is more: Learning highlight detection from video duration. In: CVPR. pp. 1258-1267 (2019) +55. Xu, M., Wang, H., Ni, B., Zhu, R., Sun, Z., Wang, C.: Cross-category video highlight detection via set-based learning. In: ICCV. pp. 7970-7979 (2021) +56. Xu, Y., Sun, Y., Li, Y., Shi, Y., Zhu, X., Du, S.: Mh-detr: Video moment and highlight detection with cross-modal transformer. Tech. Rep. arXiv:2305.00355 (2023) +57. Yan, S., Xiong, X., Nagrani, A., Arnab, A., Wang, Z., Ge, W., Ross, D., Schmid, C.: Unloc: A unified framework for video localization tasks. In: ICCV. pp. 13623-13633 (2023) +58. Yang, H., Wang, B., Lin, S., Wipf, D., Guo, M., Guo, B.: Unsupervised extraction of video highlights via robust recurrent auto-encoders. In: ICCV. pp. 4633-4641 (2015) +59. Ye, Q., Shen, X., Gao, Y., Wang, Z., Bi, Q., Li, P., Yang, G.: Temporal cue guided video highlight detection with low-rank audio-visual fusion. In: ICCV. pp. 7950-7959 (2021) +60. Yuan, T., Zhang, X., Liu, K., Liu, B., Jin, J., Jiao, Z.: Ucf-crime annotation: A benchmark for surveillance video-and-language understanding. Tech. Rep. arXiv:2309.13925 (2023) +61. Yuan, Y., Ma, L., Wang, J., Liu, W., Zhu, W.: Semantic conditioned dynamic modulation for temporal sentence grounding in videos. NeurIPS 32 (2019) +62. Zhang, H., Sun, A., Jing, W., Zhou, J.T.: Span-based localizing network for natural language video localization. Tech. Rep. arXiv:2004.13931 (2020) +63. Zhang, K., Chao, W.L., Sha, F., Grauman, K.: Video summarization with long short-term memory. In: ECCV. pp. 766-782. Springer (2016) + +64. Zhang, S., Peng, H., Fu, J., Luo, J.: Learning 2d temporal adjacent networks for moment localization with natural language. In: AAAI. pp. 12870-12877 (2020) +65. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: CVPR. pp. 16816-16825 (2022) +66. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022) \ No newline at end of file diff --git a/2024/R^2-Tuning_ Efficient Image-to-Video Transfer Learning for Video Temporal Grounding/images.zip b/2024/R^2-Tuning_ Efficient Image-to-Video Transfer Learning for Video Temporal Grounding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0a45675ab08d669e0a39ce6255121500c82ee742 --- /dev/null +++ b/2024/R^2-Tuning_ Efficient Image-to-Video Transfer Learning for Video Temporal Grounding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:02dd1424d39690d69dd7f5556b2ae3587ad5324ece280572ca107e3bc42f792d +size 720366 diff --git a/2024/R^2-Tuning_ Efficient Image-to-Video Transfer Learning for Video Temporal Grounding/layout.json b/2024/R^2-Tuning_ Efficient Image-to-Video Transfer Learning for Video Temporal Grounding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fc358bb79ccc1fd6901d2f4920e40aecb199119c --- /dev/null +++ b/2024/R^2-Tuning_ Efficient Image-to-Video Transfer Learning for Video Temporal Grounding/layout.json @@ -0,0 +1,11833 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 113, + 465, + 150 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 113, + 465, + 150 + ], + "spans": [ + { + "bbox": [ + 149, + 113, + 465, + 150 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 149, + 113, + 465, + 150 + ], + "type": "text", + "content": "-Tuning: Efficient Image-to-Video Transfer Learning for Video Temporal Grounding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 140, + 170, + 474, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 170, + 474, + 196 + ], + "spans": [ + { + "bbox": [ + 140, + 170, + 474, + 196 + ], + "type": "text", + "content": "Ye Liu" + }, + { + "bbox": [ + 140, + 170, + 474, + 196 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 140, + 170, + 474, + 196 + ], + "type": "text", + "content": ", Jixuan He" + }, + { + "bbox": [ + 140, + 170, + 474, + 196 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 140, + 170, + 474, + 196 + ], + "type": "text", + "content": ", Wanhua Li" + }, + { + "bbox": [ + 140, + 170, + 474, + 196 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + }, + { + "bbox": [ + 140, + 170, + 474, + 196 + ], + "type": "text", + "content": ", Junsik Kim" + }, + { + "bbox": [ + 140, + 170, + 474, + 196 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 140, + 170, + 474, + 196 + ], + "type": "text", + "content": ", Donglai Wei" + }, + { + "bbox": [ + 140, + 170, + 474, + 196 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 140, + 170, + 474, + 196 + ], + "type": "text", + "content": ", Hanspeter Pfister" + }, + { + "bbox": [ + 140, + 170, + 474, + 196 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 140, + 170, + 474, + 196 + ], + "type": "text", + "content": ", and Chang Wen Chen" + }, + { + "bbox": [ + 140, + 170, + 474, + 196 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 160, + 203, + 453, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 203, + 453, + 262 + ], + "spans": [ + { + "bbox": [ + 160, + 203, + 453, + 262 + ], + "type": "text", + "content": "1 The Hong Kong Polytechnic University 2 Tsinghua University 3 Harvard University 4 Boston College coco.ye.liu@connect.polyu.hk hejixuan2000@gmail.com wanhua@seas.harvard.edu mibastro@gmail.com donglai.wei@bc.edu pfister@seas.harvard.edu changwen.chen@polyu.edu.hk" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 293, + 455, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 293, + 455, + 536 + ], + "spans": [ + { + "bbox": [ + 160, + 293, + 455, + 536 + ], + "type": "text", + "content": "Abstract. Video temporal grounding (VTG) is a fine-grained video understanding problem that aims to ground relevant clips in untrimmed videos given natural language queries. Most existing VTG models are built upon frame-wise final-layer CLIP features, aided by additional temporal backbones (e.g., SlowFast) with sophisticated temporal reasoning mechanisms. In this work, we claim that CLIP itself already shows great potential for fine-grained spatial-temporal modeling, as each layer offers distinct yet useful information under different granularity levels. Motivated by this, we propose Reversed Recurrent Tuning (R²-Tuning), a parameter- and memory-efficient transfer learning framework for video temporal grounding. Our method learns a lightweight R² Block containing only 1.5% of the total parameters to perform progressive spatial-temporal modeling. Starting from the last layer of CLIP, R² Block recurrently aggregates spatial features from earlier layers, then refines temporal correlation conditioning on the given query, resulting in a coarse-to-fine scheme. R²-Tuning achieves state-of-the-art performance across three VTG tasks (i.e., moment retrieval, highlight detection, and video summarization) on six public benchmarks (i.e., QVHighlights, CharadesSTA, Ego4D-NLQ, TACoS, YouTube Highlights, and TVSum) even without the additional backbone, demonstrating the significance and effectiveness of the proposed scheme. Our code is available at https://github.com/yeliudev/R2-Tuning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 545, + 441, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 545, + 441, + 558 + ], + "spans": [ + { + "bbox": [ + 160, + 545, + 441, + 558 + ], + "type": "text", + "content": "Keywords: Video Temporal Grounding " + }, + { + "bbox": [ + 160, + 545, + 441, + 558 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 545, + 441, + 558 + ], + "type": "text", + "content": " Transfer Learning " + }, + { + "bbox": [ + 160, + 545, + 441, + 558 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 545, + 441, + 558 + ], + "type": "text", + "content": " CLIP" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 580, + 231, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 580, + 231, + 593 + ], + "spans": [ + { + "bbox": [ + 132, + 580, + 231, + 593 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 608, + 482, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 608, + 482, + 646 + ], + "spans": [ + { + "bbox": [ + 130, + 608, + 482, + 646 + ], + "type": "text", + "content": "Video is becoming the major content media in our daily lives. The variety of video domains has extended beyond daily activities [5,22] but broader scenes such as egocentric [10] and surveillance [60]. They maintain large information" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 653, + 392, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 653, + 392, + 666 + ], + "spans": [ + { + "bbox": [ + 141, + 653, + 392, + 666 + ], + "type": "text", + "content": "* Work done at Harvard University. † Corresponding authors." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 170, + 186 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 170, + 186 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 170, + 186 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 170, + 186 + ], + "type": "image", + "image_path": "0545960a4f1b6cc8b4bd213c5958d2aaa6b0c29282b65178a9db990f60ef05fe.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 186, + 170, + 192 + ], + "lines": [ + { + "bbox": [ + 136, + 186, + 170, + 192 + ], + "spans": [ + { + "bbox": [ + 136, + 186, + 170, + 192 + ], + "type": "text", + "content": "Untrimmed Video" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 171, + 115, + 321, + 194 + ], + "blocks": [ + { + "bbox": [ + 171, + 115, + 321, + 194 + ], + "lines": [ + { + "bbox": [ + 171, + 115, + 321, + 194 + ], + "spans": [ + { + "bbox": [ + 171, + 115, + 321, + 194 + ], + "type": "image", + "image_path": "44cc9f9c5c0dcc1f7cbc6f5eb8c68a404c5887900ee1dadc46f3088950515ee1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 201, + 323, + 246 + ], + "lines": [ + { + "bbox": [ + 130, + 201, + 323, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 323, + 246 + ], + "type": "text", + "content": "Fig. 1: Video temporal grounding (VTG) contains three video-language understanding problems, i.e., moment retrieval (MR), highlight detection (HD), and video summarization (VS)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 331, + 114, + 480, + 192 + ], + "blocks": [ + { + "bbox": [ + 331, + 114, + 480, + 192 + ], + "lines": [ + { + "bbox": [ + 331, + 114, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 331, + 114, + 480, + 192 + ], + "type": "image", + "image_path": "909278c2716e565a2c52590c41482aea87c5dc7ae079358d20f589a7c12af3c1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 330, + 201, + 482, + 246 + ], + "lines": [ + { + "bbox": [ + 330, + 201, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 330, + 201, + 482, + 246 + ], + "type": "text", + "content": "Fig. 2: Moment retrieval mAP with different backbones on QVHighlights val split. CLIP's potential of temporal modeling was not fully exploited." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 270, + 479, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 479, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 479, + 293 + ], + "type": "text", + "content": "capacities within multi-granularities, and thus can convey both high-level context and low-level details effectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 294, + 480, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 480, + 402 + ], + "type": "text", + "content": "In the content production and consumption industry, such a flourishing is accompanied by the increasing demand for browsing untrimmed videos efficiently with different user interests. This derives the topic of video temporal grounding (VTG) [26], which is a fine-grained video-language understanding problem aiming to ground video clips conditioning on natural language queries. As shown in Figure 1, VTG can be disentangled into three tasks based on different output formats, i.e., moment retrieval (MR) [9, 10, 22, 31] for regressing start-end timestamps, highlight detection (HD) [22, 47] for predicting frame-level saliency curves, and video summarization (VS) [2, 11, 45] for classifying disjoint shots." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 402, + 481, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 481, + 498 + ], + "type": "text", + "content": "Encouraged by the recent success in adopting vision-language models (VLMs) for video understanding [4,15,19,28,32,37], most existing VTG methods [16,22, 31,34,35,46,56] are built upon frame-wise final-layer features from CLIP [41]. However, due to CLIP's misaligned pre-training objective (image-text contrast), these methods fail to capture temporal correlations well. As a feasible compromise, an additional backbone (e.g., SlowFast [8]) is incorporated to complement the temporal information, followed by carefully designed modules such as text-guided queries [31], dynamic anchors [35], and event reasoning [16]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 498, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 666 + ], + "type": "text", + "content": "We refer to the paradigm above as a post-processing scheme (shown in Figure. 3 (a)), whereas two natural drawbacks exist due to the sub-optimal design. First, leveraging two backbones with similar capabilities is unintuitive and inefficient during inference. A single model with both vision-text alignment and spatial-temporal modeling abilities is more preferred. Second, queries for VTG could be of different granularities from coarse (e.g., the family is traveling) to fine (e.g., when did I take the golf club from the man with white hair). Leveraging only frame-wise & final-layer features is not granularity flexible, as it would focus more on high-level scene transitions while overlooking low-level details. Preliminary experiments in Figure 2 also demonstrate that the potential of spatial-temporal modeling for CLIP is not fully exploited by existing methods. Some recent works [15,19,57] tried to tackle the first problem by fine-tuning (part of) the CLIP encoders (Figure 3 (b) and (c)), but they are all inefficient in terms of data, learnable parameters, or memory." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 219, + 225 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 219, + 225 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 219, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 219, + 225 + ], + "type": "image", + "image_path": "c47d92be81ded838e7ff94e79816dbc57933c7ede699e37d1c1477ee0e80e728.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 234, + 479, + 257 + ], + "lines": [ + { + "bbox": [ + 132, + 234, + 479, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 234, + 479, + 257 + ], + "type": "text", + "content": "Fig. 3: Different architectural designs for CLIP-based image-to-video transfer learning. The gray rectangle in (d) denotes the progressively refined spatial-temporal features." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 228, + 114, + 301, + 225 + ], + "blocks": [ + { + "bbox": [ + 228, + 114, + 301, + 225 + ], + "lines": [ + { + "bbox": [ + 228, + 114, + 301, + 225 + ], + "spans": [ + { + "bbox": [ + 228, + 114, + 301, + 225 + ], + "type": "image", + "image_path": "ebd9a2205c8d88f5471fa446e5f8719ae9d78c0686598d11fc630beb91c9859b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 310, + 114, + 384, + 225 + ], + "blocks": [ + { + "bbox": [ + 310, + 114, + 384, + 225 + ], + "lines": [ + { + "bbox": [ + 310, + 114, + 384, + 225 + ], + "spans": [ + { + "bbox": [ + 310, + 114, + 384, + 225 + ], + "type": "image", + "image_path": "6213e6af772ca6d4f3ef81f729e6fc1f7990cef419bc6273a4b2b3dbd2105c66.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 394, + 115, + 480, + 225 + ], + "blocks": [ + { + "bbox": [ + 394, + 115, + 480, + 225 + ], + "lines": [ + { + "bbox": [ + 394, + 115, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 394, + 115, + 480, + 225 + ], + "type": "image", + "image_path": "4bd2d4ddef0e16fb72e7c86773e86ed259cba3f3517bfa6c02ddf98c7e2969e3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 282, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 482, + 485 + ], + "type": "text", + "content": "This paper intends to answer the research question: how to efficiently transfer an image-language foundation model to video temporal grounding? We consider the above question by exploring two aspects: efficiency and granularity flexibility. To address these issues, a novel image-to-video transfer learning framework called Reversed Recurrent Tuning (R²-Tuning) is proposed for fine-grained understanding on untrimmed videos. Our insight is that multi-layer CLIP features offer distinct yet useful information, while their integration should be tailored to the downstream task. As illustrated in Figure 3 (d), based on a frozen CLIP [41], our method learns a side-block (R² Block) containing only " + }, + { + "bbox": [ + 130, + 282, + 482, + 485 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 130, + 282, + 482, + 485 + ], + "type": "text", + "content": " of the total parameters to perform spatial-temporal modeling. R² Block is recurrently attached to the last few layers of CLIP encoders from back to front, performing query-modulated spatial pooling and recurrent temporal refinement from coarse to fine. Aside from the new architecture, we also introduce video-level and layer-wise constraints to calibrate the granularities of visual and text encoders. During training, gradient flows do not pass through the CLIP encoders, thus our scheme is both parameter- and memory-efficient. It is also granularity-flexible as R² Block can adaptively control the spatial pooling strategy conditioning on queries." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "We conduct extensive experiments across three VTG tasks on six public benchmarks, including QVHighlights [22], Ego4D-NLQ [10], Charades-STA [9], TACoS [9], YouTube Highlights [47], and TVSum [45]. Without bells and whistles, " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "-Tuning achieves more than 3 MR mAP gain on QVHighlights test split compared with " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "4\\times" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " heavier counterparts [34] with additional temporal backbones and carefully designed post-processing modules. Further analysis also shows that our method can better handle multi-granularity information. Overall, our contributions are summarized as: (1) We introduced " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "-Tuning, a novel image-to-video transfer learning framework tailored for video temporal grounding. (2) We designed two effective strategies, i.e., query-modulated spatial pooling and recurrent temporal refinement, to model spatial-temporal information from coarse to fine. (3) To calibrate the granularities of CLIP visual and text encoders, we further introduce video-level and layer-wise contrastive constraints to distill distinct information from each layer. (4) Extensive experiments across three tasks on six benchmarks demonstrate the significance and effectiveness of " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "-Tuning." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reversed Recurrent Tuning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 236, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 236, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 236, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 143, + 481, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 143, + 481, + 299 + ], + "spans": [ + { + "bbox": [ + 133, + 143, + 481, + 299 + ], + "type": "text", + "content": "CLIP for Video Understanding With powerful transfer learning abilities, CLIP [41] has been widely used for many image-language tasks [18, 29], and how to extend them to video understanding is an emerging and crucial research topic. Early attempts have been made to transfer CLIP to trimmed videos for action recognition [28,37,42] and video-text retrieval [4,15,19,32]. Specifically, Efficient-Prompt [19] introduces learnable prompt tokens to the text encoder, and ViFi-CLIP [42] further adopts learnable prompts at the visual encoder. Other works also tried to learn external modules designed for temporal modeling. EVL [28] proposes a temporal decoder in parallel with the main network. CLIP4Clip [32] investigated multiple temporal fusion strategies for video frames. Nevertheless, how to effectively transfer CLIP to untrimmed videos remains unexplored. To our best knowledge, " + }, + { + "bbox": [ + 133, + 143, + 481, + 299 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 133, + 143, + 481, + 299 + ], + "type": "text", + "content": "-Tuning is the first solution for fine-grained video temporal grounding via memory-efficient transfer learning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 314, + 481, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 314, + 481, + 517 + ], + "spans": [ + { + "bbox": [ + 133, + 314, + 481, + 517 + ], + "type": "text", + "content": "Video Temporal Grounding Video temporal grounding (VTG) [26] is a bundle of video understanding problems including moment retrieval (MR) [9,16,22, 31,35,36,52,56,57,61], highlight detection (HD) [3,14,47,55], and video summarization (VS) [11,45]. Although these tasks are closely related, they have not been jointly studied until recently. QVHighlights [22] is the first dataset that supports both MR and HD, breaking the moment distribution bias in previous MR datasets [9,43]. The authors also proposed Moment-DETR, a strong baseline method for MR/HD. Most following works were built upon the Moment-DETR framework with specially designed modules. Notably, UMT [31] tackles flexible MR and HD through a unified multi-modal architecture. QD-DETR [35] benefits from query-dependent video representations and dynamic anchors. EaTR [16] incorporates moment reasoning to provide reliable referential search areas for moment queries. Nevertheless, their design space is still within pre-extracted CLIP [41] and SlowFast [8] features. A recent work [57] revealed the possibility of fine-tuning CLIP for MR, while their solution (post pre-training) is extremely resource-consuming. Compared with existing works, " + }, + { + "bbox": [ + 133, + 314, + 481, + 517 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 133, + 314, + 481, + 517 + ], + "type": "text", + "content": "-Tuning can effectively learn strong temporal modeling abilities without any pre-training on videos." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 533, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 533, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 533, + 481, + 665 + ], + "type": "text", + "content": "Parameter-Efficient Transfer Learning Foundation models [41, 49] have achieved remarkable successes in vision and language understanding. Efforts have been made to efficiently transfer knowledge from these models to new scenarios. One line of research is to conduct context-based tuning (e.g., prompt-tuning [17, 65, 66]), which does not modify the model architectures but incorporates learnable embeddings in the inputs. Another paradigm is introducing an extra lightweight adapter to the original model, and keeping the rest of the parameters frozen [39]. Recent works also consider memory-efficient transfer learning (METL) that introduces parameters whose gradient flows do not go through the main model [28, 40, 48]. " + }, + { + "bbox": [ + 133, + 533, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 133, + 533, + 481, + 665 + ], + "type": "text", + "content": "-Tuning is also a METL framework, where we adopt a novel recurrent tuning strategy to ensure parameter- and memory-efficient." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 219 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 219 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 219 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 219 + ], + "type": "image", + "image_path": "23773a00ff20ed68d2af5e26b0a9008d233cec11fe04d5e92343f9d9b6060da3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 227, + 482, + 283 + ], + "lines": [ + { + "bbox": [ + 130, + 227, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 482, + 283 + ], + "type": "text", + "content": "Fig. 4: Overall architecture of our framework. The input video and query are first encoded by frozen CLIP [41] encoders. Their multi-layer outputs are then recurrently fused and refined by a learnable " + }, + { + "bbox": [ + 130, + 227, + 482, + 283 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 130, + 227, + 482, + 283 + ], + "type": "text", + "content": " Block to construct spatial-temporal representations " + }, + { + "bbox": [ + 130, + 227, + 482, + 283 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 227, + 482, + 283 + ], + "type": "text", + "content": ", which would be scaled up/down to construct a temporal feature pyramid, followed by three heads for MR, HD, and VS, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 306, + 233, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 306, + 233, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 306, + 233, + 319 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 331, + 268, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 268, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 268, + 342 + ], + "type": "text", + "content": "3.1 Problem Formulation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 351, + 481, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 481, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 481, + 389 + ], + "type": "text", + "content": "Given a video " + }, + { + "bbox": [ + 130, + 351, + 481, + 389 + ], + "type": "inline_equation", + "content": "V = \\{v_{i}\\}_{i=1}^{T}" + }, + { + "bbox": [ + 130, + 351, + 481, + 389 + ], + "type": "text", + "content": " and a natural language query " + }, + { + "bbox": [ + 130, + 351, + 481, + 389 + ], + "type": "inline_equation", + "content": "Q = \\{q_{i}\\}_{i=1}^{L}" + }, + { + "bbox": [ + 130, + 351, + 481, + 389 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 351, + 481, + 389 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 351, + 481, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 351, + 481, + 389 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 351, + 481, + 389 + ], + "type": "text", + "content": " are the numbers of video frames and text tokens, VTG aims to densely predict a set of labels " + }, + { + "bbox": [ + 130, + 351, + 481, + 389 + ], + "type": "inline_equation", + "content": "(b_{i}, s_{i}, f_{i})" + }, + { + "bbox": [ + 130, + 351, + 481, + 389 + ], + "type": "text", + "content": " for each frame, defined as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 399, + 481, + 519 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 138, + 399, + 481, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 481, + 446 + ], + "type": "text", + "content": "- Moment Retrieval is to find the most relevant moments (i.e. sets of consecutive frames) in " + }, + { + "bbox": [ + 138, + 399, + 481, + 446 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 138, + 399, + 481, + 446 + ], + "type": "text", + "content": " according to " + }, + { + "bbox": [ + 138, + 399, + 481, + 446 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 138, + 399, + 481, + 446 + ], + "type": "text", + "content": ", so that " + }, + { + "bbox": [ + 138, + 399, + 481, + 446 + ], + "type": "inline_equation", + "content": "b_{i} = [b_{i}^{s}, b_{i}^{e}] \\in \\mathbb{R}^{2}" + }, + { + "bbox": [ + 138, + 399, + 481, + 446 + ], + "type": "text", + "content": " represents the temporal displacements from frame " + }, + { + "bbox": [ + 138, + 399, + 481, + 446 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 138, + 399, + 481, + 446 + ], + "type": "text", + "content": " to the start and end timestamps of the nearest target moment." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 447, + 481, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 481, + 483 + ], + "type": "text", + "content": "- Highlight Detection requires estimating frame-level relevancies between " + }, + { + "bbox": [ + 138, + 447, + 481, + 483 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 138, + 447, + 481, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 138, + 447, + 481, + 483 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 138, + 447, + 481, + 483 + ], + "type": "text", + "content": ", thus " + }, + { + "bbox": [ + 138, + 447, + 481, + 483 + ], + "type": "inline_equation", + "content": "s_i \\in [0,1]" + }, + { + "bbox": [ + 138, + 447, + 481, + 483 + ], + "type": "text", + "content": " is a continuous saliency score denoting to what extent frame " + }, + { + "bbox": [ + 138, + 447, + 481, + 483 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 138, + 447, + 481, + 483 + ], + "type": "text", + "content": " is semantically aligned with " + }, + { + "bbox": [ + 138, + 447, + 481, + 483 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 138, + 447, + 481, + 483 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 483, + 481, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 483, + 481, + 519 + ], + "spans": [ + { + "bbox": [ + 138, + 483, + 481, + 519 + ], + "type": "text", + "content": "- Video Summarization aims to select a subset of video frames according to " + }, + { + "bbox": [ + 138, + 483, + 481, + 519 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 138, + 483, + 481, + 519 + ], + "type": "text", + "content": " to form a concise summary, therefore, " + }, + { + "bbox": [ + 138, + 483, + 481, + 519 + ], + "type": "inline_equation", + "content": "f_{i} \\in \\{0,1\\}" + }, + { + "bbox": [ + 138, + 483, + 481, + 519 + ], + "type": "text", + "content": " is a binary score indicating whether frame " + }, + { + "bbox": [ + 138, + 483, + 481, + 519 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 138, + 483, + 481, + 519 + ], + "type": "text", + "content": " belongs to the summary." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 536, + 208, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 208, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 208, + 548 + ], + "type": "text", + "content": "3.2 Overview" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 557, + 481, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 629 + ], + "type": "text", + "content": "Figure 4 shows an overview of the proposed framework. Our model derives from a pre-trained and frozen CLIP [41] with ViT [6] backbone, which has a two-stream architecture for spatial and textual encoding, respectively. A learnable " + }, + { + "bbox": [ + 130, + 557, + 481, + 629 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 130, + 557, + 481, + 629 + ], + "type": "text", + "content": " block is iteratively attached to the last " + }, + { + "bbox": [ + 130, + 557, + 481, + 629 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 557, + 481, + 629 + ], + "type": "text", + "content": " encoder layers to refine the temporal correlation. The resulting features are then scaled up/down to build a feature pyramid, followed by three heads to decode the task-specific predictions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 630, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 667 + ], + "type": "text", + "content": "Specifically, the input video " + }, + { + "bbox": [ + 130, + 630, + 481, + 667 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 630, + 481, + 667 + ], + "type": "text", + "content": " and query " + }, + { + "bbox": [ + 130, + 630, + 481, + 667 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 630, + 481, + 667 + ], + "type": "text", + "content": " are first tokenized into frame patches and word tokens, then sent into the visual encoder " + }, + { + "bbox": [ + 130, + 630, + 481, + 667 + ], + "type": "inline_equation", + "content": "E_{v}" + }, + { + "bbox": [ + 130, + 630, + 481, + 667 + ], + "type": "text", + "content": " and query encoder " + }, + { + "bbox": [ + 130, + 630, + 481, + 667 + ], + "type": "inline_equation", + "content": "E_{q}" + }, + { + "bbox": [ + 130, + 630, + 481, + 667 + ], + "type": "text", + "content": ". The encoded visual and textual features can be denoted as " + }, + { + "bbox": [ + 130, + 630, + 481, + 667 + ], + "type": "inline_equation", + "content": "e_v \\in" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 334, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 334, + 91, + 448, + 102 + ], + "type": "text", + "content": "Reversed Recurrent Tuning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 211, + 116, + 404, + 240 + ], + "blocks": [ + { + "bbox": [ + 211, + 116, + 404, + 240 + ], + "lines": [ + { + "bbox": [ + 211, + 116, + 404, + 240 + ], + "spans": [ + { + "bbox": [ + 211, + 116, + 404, + 240 + ], + "type": "image", + "image_path": "9a849fabc871996edfcd43ef75c076d8f644dcff3fb91447fbab73a7bbb56a50.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 251, + 482, + 285 + ], + "lines": [ + { + "bbox": [ + 131, + 251, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 131, + 251, + 482, + 285 + ], + "type": "text", + "content": "Fig. 5: Detailed architecture of the " + }, + { + "bbox": [ + 131, + 251, + 482, + 285 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 131, + 251, + 482, + 285 + ], + "type": "text", + "content": " Block. It can be split into two parts: a) query-modulated spatial pooling, and b) recurrent temporal refinement. Note that the [CLS] tokens of visual features are omitted for clarity." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{B \\times N \\times T \\times (P + 1) \\times D_v}" + }, + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "inline_equation", + "content": "e_q \\in \\mathbb{R}^{B \\times N \\times L \\times D_q}" + }, + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "inline_equation", + "content": "B, N, T, P, L, D_v" + }, + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "inline_equation", + "content": "D_q" + }, + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "text", + "content": " indicate batch size, number of encoder layers, number of video frames, number of patches per frame, number of query tokens, and the dimensions of visual/query features, respectively. These features are recurrently fused and refined by " + }, + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "text", + "content": " block to construct spatial-temporal features " + }, + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "inline_equation", + "content": "h \\in \\mathbb{R}^{B \\times T \\times C}" + }, + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "text", + "content": ", in which each token preserves the " + }, + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 307, + 482, + 392 + ], + "type": "text", + "content": "-dimensional features for a frame. This process will be introduced in detail in Section 3.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 410, + 301, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 410, + 301, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 410, + 301, + 423 + ], + "type": "text", + "content": "3.3 Reversed Recurrent Tuning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "text", + "content": "Given pre-trained visual and query encoders, existing works [16,22,26,31,35] merely take features from the last hidden layer, which is sub-optimal due to the limited information. In this work, we propose " + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "text", + "content": "-Tuning to exploit the potential of spatial-temporal modeling based on multi-layer CLIP features. Our scheme learns a lightweight " + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "text", + "content": " Block which maintains a zero-initialized hidden state " + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "inline_equation", + "content": "h \\in \\mathbb{R}^{B \\times T \\times C}" + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "text", + "content": " as frame-level spatial-temporal features. The " + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "text", + "content": " Block is recurrently attached to the last " + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "text", + "content": " layers of CLIP encoders from back to front to refine the hidden state " + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "text", + "content": " steps, with each step defined as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 258, + 534, + 481, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 534, + 481, + 550 + ], + "spans": [ + { + "bbox": [ + 258, + 534, + 481, + 550 + ], + "type": "interline_equation", + "content": "h ^ {k} = \\mathcal {F} _ {\\theta} (e _ {v} ^ {n}, e _ {q} ^ {n}, h ^ {k - 1}) \\tag {1}", + "image_path": "b0903acf1d625de0a6a33e70061c254b0d8c4986646d3c6c9e1a6f3738035ae7.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 557, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 557, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 131, + 557, + 482, + 628 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 131, + 557, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}" + }, + { + "bbox": [ + 131, + 557, + 482, + 628 + ], + "type": "text", + "content": " is the refinement operation parameterized by " + }, + { + "bbox": [ + 131, + 557, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 131, + 557, + 482, + 628 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 557, + 482, + 628 + ], + "type": "inline_equation", + "content": "k \\in [1, K]" + }, + { + "bbox": [ + 131, + 557, + 482, + 628 + ], + "type": "text", + "content": " is the index of current step. " + }, + { + "bbox": [ + 131, + 557, + 482, + 628 + ], + "type": "inline_equation", + "content": "n = N - k + 1" + }, + { + "bbox": [ + 131, + 557, + 482, + 628 + ], + "type": "text", + "content": " denotes the CLIP layer index for step " + }, + { + "bbox": [ + 131, + 557, + 482, + 628 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 557, + 482, + 628 + ], + "type": "text", + "content": ", thus " + }, + { + "bbox": [ + 131, + 557, + 482, + 628 + ], + "type": "inline_equation", + "content": "e_v^n" + }, + { + "bbox": [ + 131, + 557, + 482, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 557, + 482, + 628 + ], + "type": "inline_equation", + "content": "e_q^n" + }, + { + "bbox": [ + 131, + 557, + 482, + 628 + ], + "type": "text", + "content": " are visual and query features from the " + }, + { + "bbox": [ + 131, + 557, + 482, + 628 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 131, + 557, + 482, + 628 + ], + "type": "text", + "content": "-th CLIP encoder layer. The refinement operation " + }, + { + "bbox": [ + 131, + 557, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}" + }, + { + "bbox": [ + 131, + 557, + 482, + 628 + ], + "type": "text", + "content": " can be decomposed into two parts: 1) query-modulated spatial pooling, and 2) recurrent temporal refinement. Details are discussed as follows." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "Query-Modulated Spatial Pooling Figure 5 (a) presents the data flow of query-modulated spatial pooling. The goal of this process is to adaptively pool" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "spatial features from patch-level representations " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "e_v^n" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " to a single token " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "e_{pool}^n" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " conditioning on the query " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "e_q^n" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": ". We first adopt two separate MLPs to map " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "e_v^n" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "e_q^n" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " into the same space:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 235, + 159, + 481, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 159, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 235, + 159, + 481, + 172 + ], + "type": "interline_equation", + "content": "\\hat {e} _ {v} ^ {n} = \\operatorname {M L P} \\left(e _ {v} ^ {n}\\right) \\in \\mathbb {R} ^ {B \\times T \\times (P + 1) \\times C} \\tag {2}", + "image_path": "6768fed2a3c9237515f6049c69a03b7cce0b9522c43bba4406122b93055d54df.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 236, + 175, + 480, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 175, + 480, + 190 + ], + "spans": [ + { + "bbox": [ + 236, + 175, + 480, + 190 + ], + "type": "interline_equation", + "content": "\\hat {e} _ {q} ^ {n} = \\operatorname {M L P} \\left(e _ {q} ^ {n}\\right) \\in \\mathbb {R} ^ {B \\times L \\times C} \\tag {3}", + "image_path": "4bc6e61244dc88614c2b9eeab794136a8886a0c58d2948e731212b938a471fc4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 196, + 479, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 196, + 479, + 247 + ], + "spans": [ + { + "bbox": [ + 131, + 196, + 479, + 247 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 131, + 196, + 479, + 247 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 131, + 196, + 479, + 247 + ], + "type": "text", + "content": " is the reduced hidden size. We further align the shape of " + }, + { + "bbox": [ + 131, + 196, + 479, + 247 + ], + "type": "inline_equation", + "content": "\\hat{e}_v^n" + }, + { + "bbox": [ + 131, + 196, + 479, + 247 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 196, + 479, + 247 + ], + "type": "inline_equation", + "content": "\\hat{e}_q^n" + }, + { + "bbox": [ + 131, + 196, + 479, + 247 + ], + "type": "text", + "content": " by merging dimension " + }, + { + "bbox": [ + 131, + 196, + 479, + 247 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 196, + 479, + 247 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 131, + 196, + 479, + 247 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 131, + 196, + 479, + 247 + ], + "type": "text", + "content": " and repeat " + }, + { + "bbox": [ + 131, + 196, + 479, + 247 + ], + "type": "inline_equation", + "content": "\\hat{e}_q^n" + }, + { + "bbox": [ + 131, + 196, + 479, + 247 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 131, + 196, + 479, + 247 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 196, + 479, + 247 + ], + "type": "text", + "content": " times in an interleaved style, resulting " + }, + { + "bbox": [ + 131, + 196, + 479, + 247 + ], + "type": "inline_equation", + "content": "\\hat{e}_v^n \\in \\mathbb{R}^{(B \\times T) \\times (P + 1) \\times C}" + }, + { + "bbox": [ + 131, + 196, + 479, + 247 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 196, + 479, + 247 + ], + "type": "inline_equation", + "content": "\\hat{e}_q^n \\in \\mathbb{R}^{(B \\times T) \\times L \\times C}" + }, + { + "bbox": [ + 131, + 196, + 479, + 247 + ], + "type": "text", + "content": ". Then, we compute the similarities for patch-token pairs using normalized Embedded Gaussian [50]:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 212, + 254, + 480, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 254, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 212, + 254, + 480, + 281 + ], + "type": "interline_equation", + "content": "a = \\operatorname {s o f t m a x} \\left(\\frac {\\left(w _ {q} \\hat {e} _ {q} ^ {n}\\right) ^ {\\top} w _ {v} \\hat {e} _ {v} ^ {n}}{\\sqrt {C}}\\right) \\in \\mathbb {R} ^ {(B \\times T) \\times L \\times P} \\tag {4}", + "image_path": "a1feaa4bff9d24d726823d26eccb2bfdf109934fe3998a79266378864352ab7d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 286, + 479, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 286, + 479, + 346 + ], + "spans": [ + { + "bbox": [ + 131, + 286, + 479, + 346 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 131, + 286, + 479, + 346 + ], + "type": "inline_equation", + "content": "w_{v}" + }, + { + "bbox": [ + 131, + 286, + 479, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 286, + 479, + 346 + ], + "type": "inline_equation", + "content": "w_{q}" + }, + { + "bbox": [ + 131, + 286, + 479, + 346 + ], + "type": "text", + "content": " are learnable matrices for projecting features. This operation derives from cross-attention but discards the linear projection on value. We then multiply the similarities " + }, + { + "bbox": [ + 131, + 286, + 479, + 346 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 131, + 286, + 479, + 346 + ], + "type": "text", + "content": " with visual features " + }, + { + "bbox": [ + 131, + 286, + 479, + 346 + ], + "type": "inline_equation", + "content": "\\hat{e}_{v}^{n}" + }, + { + "bbox": [ + 131, + 286, + 479, + 346 + ], + "type": "text", + "content": " to pool them into each token, such that tokens can interact with patches independently. After that, a max pooling along the " + }, + { + "bbox": [ + 131, + 286, + 479, + 346 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 131, + 286, + 479, + 346 + ], + "type": "text", + "content": " dimension is applied:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 213, + 354, + 480, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 354, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 213, + 354, + 480, + 369 + ], + "type": "interline_equation", + "content": "e _ {t o k e n} ^ {n} = \\operatorname {M a x P o o l i n g} \\left(a \\hat {e} _ {v} ^ {n} + \\hat {e} _ {q} ^ {n}\\right) \\in \\mathbb {R} ^ {B \\times T \\times C} \\tag {5}", + "image_path": "9873120283d251fac603a90bfe399c570a044ae7c0db7d3e8b626ace6576f8a6.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 374, + 479, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 374, + 479, + 422 + ], + "spans": [ + { + "bbox": [ + 131, + 374, + 479, + 422 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 131, + 374, + 479, + 422 + ], + "type": "inline_equation", + "content": "\\hat{e}_q^n" + }, + { + "bbox": [ + 131, + 374, + 479, + 422 + ], + "type": "text", + "content": " serves as residuals [13] to stabilize training. Our visualizations in Figure 8 show that this strategy can guide the model to focus more on query-related regions. Finally, we combine " + }, + { + "bbox": [ + 131, + 374, + 479, + 422 + ], + "type": "inline_equation", + "content": "e_{token}^n" + }, + { + "bbox": [ + 131, + 374, + 479, + 422 + ], + "type": "text", + "content": " with the [CLS] token to generate query-modulated spatial features:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 230, + 430, + 480, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 430, + 480, + 444 + ], + "spans": [ + { + "bbox": [ + 230, + 430, + 480, + 444 + ], + "type": "interline_equation", + "content": "e _ {p o o l} ^ {n} = e _ {v} ^ {n, 0} + g ^ {k} \\cdot e _ {t o k e n} ^ {n} \\in \\mathbb {R} ^ {B \\times T \\times C} \\tag {6}", + "image_path": "4aed7598c2ab2339a976b6ae2203cbdca53bff82d53e999b0a5729f103b5f991.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 451, + 479, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 451, + 479, + 499 + ], + "spans": [ + { + "bbox": [ + 131, + 451, + 479, + 499 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 131, + 451, + 479, + 499 + ], + "type": "inline_equation", + "content": "g^{k} \\in (-1,1)" + }, + { + "bbox": [ + 131, + 451, + 479, + 499 + ], + "type": "text", + "content": " is a zero-initialized learnable gate for step " + }, + { + "bbox": [ + 131, + 451, + 479, + 499 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 451, + 479, + 499 + ], + "type": "text", + "content": " constrained by Tanh. We allow negative values here to remove useless information from [CLS] tokens. " + }, + { + "bbox": [ + 131, + 451, + 479, + 499 + ], + "type": "inline_equation", + "content": "e_{pool}^{n}" + }, + { + "bbox": [ + 131, + 451, + 479, + 499 + ], + "type": "text", + "content": " is then used to model temporal correlations and are refined in a recurrent manner." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 510, + 479, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 510, + 479, + 548 + ], + "spans": [ + { + "bbox": [ + 131, + 510, + 479, + 548 + ], + "type": "text", + "content": "Recurrent Temporal Refinement Figure 5 (b) displays the flow of recurrent temporal refinement. Specifically, the pooled visual features " + }, + { + "bbox": [ + 131, + 510, + 479, + 548 + ], + "type": "inline_equation", + "content": "e_{pool}^{n}" + }, + { + "bbox": [ + 131, + 510, + 479, + 548 + ], + "type": "text", + "content": " is first fused with the hidden state " + }, + { + "bbox": [ + 131, + 510, + 479, + 548 + ], + "type": "inline_equation", + "content": "h^{k-1}" + }, + { + "bbox": [ + 131, + 510, + 479, + 548 + ], + "type": "text", + "content": " from the previous step:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 233, + 554, + 480, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 554, + 480, + 569 + ], + "spans": [ + { + "bbox": [ + 233, + 554, + 480, + 569 + ], + "type": "interline_equation", + "content": "\\hat {h} ^ {k - 1} = \\varphi^ {k} \\cdot e _ {\\text {p o o l}} ^ {n} + (1 - \\varphi^ {k}) \\cdot h ^ {k - 1} \\tag {7}", + "image_path": "ab68f05c2793c379d97e5177e582ad33313a2500a95e39d4fd9bd3e0e60eb506.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 576, + 479, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 576, + 479, + 613 + ], + "spans": [ + { + "bbox": [ + 131, + 576, + 479, + 613 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 576, + 479, + 613 + ], + "type": "inline_equation", + "content": "\\varphi^k\\in (0,1)" + }, + { + "bbox": [ + 131, + 576, + 479, + 613 + ], + "type": "text", + "content": " is the learnable gate for step " + }, + { + "bbox": [ + 131, + 576, + 479, + 613 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 576, + 479, + 613 + ], + "type": "text", + "content": ". Then, we adopt a sequence of standard multi-head cross-attention " + }, + { + "bbox": [ + 131, + 576, + 479, + 613 + ], + "type": "inline_equation", + "content": "(\\hat{e}_q^n" + }, + { + "bbox": [ + 131, + 576, + 479, + 613 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 131, + 576, + 479, + 613 + ], + "type": "inline_equation", + "content": "k,v)" + }, + { + "bbox": [ + 131, + 576, + 479, + 613 + ], + "type": "text", + "content": ", multi-head self-attention, followed by a feed-forward network [50] to update the hidden state:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 225, + 620, + 480, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 620, + 480, + 635 + ], + "spans": [ + { + "bbox": [ + 225, + 620, + 480, + 635 + ], + "type": "interline_equation", + "content": "h ^ {k} = \\operatorname {F F N} (\\mathrm {M H S A} (\\mathrm {M H C A} (\\hat {h} ^ {k - 1}, \\hat {e} _ {q} ^ {n}))) \\tag {8}", + "image_path": "2147c3286d703e9e4d56ca63d2ffbcd0abf729c6cc4d49e9ca1a0c05c752e6b8.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "type": "text", + "content": "For each block, we adopt DropPath [21] with " + }, + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "type": "inline_equation", + "content": "p = 0.1" + }, + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "type": "text", + "content": " to prevent overfitting. The query features " + }, + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\hat{e}_q^n" + }, + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "type": "text", + "content": " are utilized both in spatial and temporal modeling as guidance." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 334, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 334, + 91, + 448, + 102 + ], + "type": "text", + "content": "Reversed Recurrent Tuning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "Granularity Calibration The refinement process above is based on the assumption that visual and query features from the same layer of CLIP [41] are well aligned at the same granularity level. However, this cannot be guaranteed as the two encoders are learned in isolation during pre-training. Therefore, we need to add manual constraints to calibrate their granularities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 176, + 480, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 176, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 176, + 480, + 236 + ], + "type": "text", + "content": "We apply a video-level constraint and a layer-wise constraint by designing two contrastive losses. We first denote " + }, + { + "bbox": [ + 132, + 176, + 480, + 236 + ], + "type": "inline_equation", + "content": "e_v^0 \\in \\mathbb{R}^{B \\times K \\times T \\times C}" + }, + { + "bbox": [ + 132, + 176, + 480, + 236 + ], + "type": "text", + "content": " as the features of " + }, + { + "bbox": [ + 132, + 176, + 480, + 236 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 132, + 176, + 480, + 236 + ], + "type": "text", + "content": " [CLS] tokens from Eq. 6. Then, we select all positive frames (i.e., frames within " + }, + { + "bbox": [ + 132, + 176, + 480, + 236 + ], + "type": "inline_equation", + "content": "[b_i^s, b_i^e]" + }, + { + "bbox": [ + 132, + 176, + 480, + 236 + ], + "type": "text", + "content": " or with " + }, + { + "bbox": [ + 132, + 176, + 480, + 236 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 132, + 176, + 480, + 236 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 132, + 176, + 480, + 236 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 132, + 176, + 480, + 236 + ], + "type": "text", + "content": " higher than a threshold) from " + }, + { + "bbox": [ + 132, + 176, + 480, + 236 + ], + "type": "inline_equation", + "content": "e_v^0" + }, + { + "bbox": [ + 132, + 176, + 480, + 236 + ], + "type": "text", + "content": " and perform average pooling to obtain video-level representations " + }, + { + "bbox": [ + 132, + 176, + 480, + 236 + ], + "type": "inline_equation", + "content": "\\widetilde{e}_v \\in \\mathbb{R}^{B \\times K \\times C}" + }, + { + "bbox": [ + 132, + 176, + 480, + 236 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 253, + 243, + 480, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 243, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 253, + 243, + 480, + 255 + ], + "type": "interline_equation", + "content": "\\widetilde {e} _ {v} = \\operatorname {A v g P o o l} \\left(\\left\\{e _ {v} ^ {0, i} \\right\\} _ {i \\in \\Omega}\\right) \\tag {9}", + "image_path": "39eae0fb4b77e42a6cf035c4a96df0fcde8273483521437a4e9f13c7c376e177.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 262, + 479, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 262, + 479, + 309 + ], + "spans": [ + { + "bbox": [ + 131, + 262, + 479, + 309 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 131, + 262, + 479, + 309 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 131, + 262, + 479, + 309 + ], + "type": "text", + "content": " is the set of positive frame indices. To obtain query-level representations " + }, + { + "bbox": [ + 131, + 262, + 479, + 309 + ], + "type": "inline_equation", + "content": "\\widetilde{e}_q\\in \\mathbb{R}^{B\\times K\\times C}" + }, + { + "bbox": [ + 131, + 262, + 479, + 309 + ], + "type": "text", + "content": ", a token-wise adaptive pooling is applied following previous work [26]. We adopt InfoNCE loss [38] along two dimensions to calculate the video-level and layer-wise constraints:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 224, + 316, + 480, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 316, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 224, + 316, + 480, + 343 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {v i d e o}} = \\lambda_ {\\text {v i d e o}} \\frac {1}{B} \\sum_ {b \\in B} \\operatorname {I n f o N C E} \\left(\\tilde {e} _ {v} ^ {b}, \\tilde {e} _ {q} ^ {b}\\right) \\tag {10}", + "image_path": "00d19d7154ba4948706ba1a51722c429d5d08d47d104c112be36bc0269ba326b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 224, + 346, + 480, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 346, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 224, + 346, + 480, + 373 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {l a y e r}} = \\lambda_ {\\text {l a y e r}} \\frac {1}{K} \\sum_ {k \\in K} \\operatorname {I n f o N C E} \\left(\\widetilde {e} _ {v} ^ {k}, \\widetilde {e} _ {q} ^ {k}\\right) \\tag {11}", + "image_path": "7edc8e46fe988ac88cf0f18a00d64aadfe972ccf0e0346015230d7545e5a3784.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 380, + 480, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 380, + 480, + 452 + ], + "spans": [ + { + "bbox": [ + 131, + 380, + 480, + 452 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 131, + 380, + 480, + 452 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{video}" + }, + { + "bbox": [ + 131, + 380, + 480, + 452 + ], + "type": "text", + "content": " performs contrast among samples in the same batch (diversifying features are diverse among moment-query pairs), and averages the loss across " + }, + { + "bbox": [ + 131, + 380, + 480, + 452 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 380, + 480, + 452 + ], + "type": "text", + "content": " layers. It also implicitly models temporal correlations between videos and queries. " + }, + { + "bbox": [ + 131, + 380, + 480, + 452 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{layer}" + }, + { + "bbox": [ + 131, + 380, + 480, + 452 + ], + "type": "text", + "content": " performs contrast among layers and averages across the batch. " + }, + { + "bbox": [ + 131, + 380, + 480, + 452 + ], + "type": "inline_equation", + "content": "\\lambda_{video}" + }, + { + "bbox": [ + 131, + 380, + 480, + 452 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 380, + 480, + 452 + ], + "type": "inline_equation", + "content": "\\lambda_{layer}" + }, + { + "bbox": [ + 131, + 380, + 480, + 452 + ], + "type": "text", + "content": " serve as re-weighting terms for the losses. A fixed temperature parameter of 0.07 is used." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 468, + 247, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 468, + 247, + 479 + ], + "spans": [ + { + "bbox": [ + 131, + 468, + 247, + 479 + ], + "type": "text", + "content": "3.4 Prediction Heads" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 485, + 479, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 479, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 479, + 568 + ], + "type": "text", + "content": "After refining spatial-temporal features " + }, + { + "bbox": [ + 130, + 485, + 479, + 568 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 485, + 479, + 568 + ], + "type": "text", + "content": ", a temporal feature pyramid is constructed by applying 1D convolutions with stride " + }, + { + "bbox": [ + 130, + 485, + 479, + 568 + ], + "type": "inline_equation", + "content": "= 2" + }, + { + "bbox": [ + 130, + 485, + 479, + 568 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 130, + 485, + 479, + 568 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 485, + 479, + 568 + ], + "type": "text", + "content": ". The number of convolutions used for each level is subjected to " + }, + { + "bbox": [ + 130, + 485, + 479, + 568 + ], + "type": "inline_equation", + "content": "l - 1" + }, + { + "bbox": [ + 130, + 485, + 479, + 568 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 485, + 479, + 568 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 485, + 479, + 568 + ], + "type": "text", + "content": " is the level index starting from 1. We concatenate features from all levels by the temporal dimension to form " + }, + { + "bbox": [ + 130, + 485, + 479, + 568 + ], + "type": "inline_equation", + "content": "\\hat{h}" + }, + { + "bbox": [ + 130, + 485, + 479, + 568 + ], + "type": "text", + "content": " and predict once in all heads. Following a similar but simplified design from [26], we adopt three heads for video temporal grounding, described in detail as follows." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 578, + 479, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 578, + 479, + 614 + ], + "spans": [ + { + "bbox": [ + 131, + 578, + 479, + 614 + ], + "type": "text", + "content": "Foreground-Background Classification Given " + }, + { + "bbox": [ + 131, + 578, + 479, + 614 + ], + "type": "inline_equation", + "content": "\\hat{h}" + }, + { + "bbox": [ + 131, + 578, + 479, + 614 + ], + "type": "text", + "content": ", a two-layer 1D convolution module with " + }, + { + "bbox": [ + 131, + 578, + 479, + 614 + ], + "type": "inline_equation", + "content": "(1\\times 3)" + }, + { + "bbox": [ + 131, + 578, + 479, + 614 + ], + "type": "text", + "content": " kernels is adopted to predict " + }, + { + "bbox": [ + 131, + 578, + 479, + 614 + ], + "type": "inline_equation", + "content": "\\hat{f}_i" + }, + { + "bbox": [ + 131, + 578, + 479, + 614 + ], + "type": "text", + "content": " for each frame. It is optimized using a focal loss [27] with " + }, + { + "bbox": [ + 131, + 578, + 479, + 614 + ], + "type": "inline_equation", + "content": "\\alpha = 0.9" + }, + { + "bbox": [ + 131, + 578, + 479, + 614 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 578, + 479, + 614 + ], + "type": "inline_equation", + "content": "\\gamma = 2.0" + }, + { + "bbox": [ + 131, + 578, + 479, + 614 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 242, + 621, + 480, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 621, + 480, + 635 + ], + "spans": [ + { + "bbox": [ + 242, + 621, + 480, + 635 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c l s} = - \\lambda_ {c l s} \\alpha (1 - \\hat {f} _ {i}) ^ {\\gamma} \\log (\\hat {f} _ {i}) \\tag {12}", + "image_path": "4fdfe30c9c618f33566b9dd978f3a289a8343f5a525af3a941b5a5cc10498194.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "type": "text", + "content": "We find that focal loss perform better than binary cross-entropy when the numbers of foreground/background frames are severely imbalanced as noted in [26]." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 178, + 482, + 385 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 1: Video moment retrieval (MR) and highlight detection (HD) results on QVHighlights test split. Note that [16] and [57] only reported their results on val split. " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "\\bigcirc" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "\\bigcirc" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " are CLIP-B/32 and B/16 [41], " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "\\bigcirc" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " denotes SlowFast R-50 [8], and " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "\\bigcirc" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " is PANN [20] for audio features extraction. * means estimated parameters. The best and second-best metrics are marked with bold and underline, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 178, + 482, + 385 + ], + "lines": [ + { + "bbox": [ + 132, + 178, + 482, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 178, + 482, + 385 + ], + "type": "table", + "html": "
MethodBackbonePost \nPre-trainMRHD
R1mAP≥ Very Good#Params
@0.5@0.7@0.5@0.75Avg.mAPHIT@1
BeautyThumb [44]○○X-----14.3620.88-
DVSE [30]○○X-----18.7521.79-
MCN [1]○○X11.412.7224.948.2210.67---
CAL [7]○○X25.4911.5423.407.659.89---
XML [23]○○X41.8330.3544.6331.7332.1434.4955.25-
XML+ [22]○○X46.6933.4647.8934.6734.9035.3855.06-
Moment-DETR [22]○○X52.8933.0254.8229.4030.7335.6955.604.8M
UMT [31]○○○X56.2341.1853.8337.0136.1238.1859.9914.9M
MomentDiff [24]○○○X58.2141.4854.5737.2136.84---
QD-DETR [35]○○X62.4044.9862.5239.8839.8638.9462.407.6M
MH-DETR [56]○○X60.0542.4860.7538.1338.3838.2260.518.2M
UniVTG [26]○○X58.8640.8657.6035.5935.4738.2060.9641.3M
TR-DETR [46]○○X64.6648.9663.9843.7342.6239.9163.427.9M
CG-DETR [34]○○X65.4348.3864.5142.7742.8640.3366.2112.0M
EaTR [16]○○X61.3645.7961.8641.9141.7437.1558.659.0M
Moment-DETR [22]○○236K ASR Cap.59.7840.3360.5135.3636.1437.4360.174.8M
UMT [31]○○○236K ASR Cap.60.8343.2657.3339.1238.0839.1262.3914.9M
UniVTG [26]○○4.2M Corpus65.4350.0664.0645.0243.6340.5466.2841.3M
UnLoc [57]○650K Videos64.5048.80-----87.9M*
R²-Tuning (Ours)○X68.0349.3569.0447.5646.1740.7564.202.7M
", + "image_path": "aaf408237a04185f85ea2a00aa2864eb9fb9b1e63b730526f125088ca475cf22.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 404, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 404, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 482, + 464 + ], + "type": "text", + "content": "Boundary Regression Similar to foreground-background classification, boundary regression for moments is also realized by a two-layer 1D convolution module. The difference is that its output dimension is set to 2 instead of 1, representing the boundary displacements " + }, + { + "bbox": [ + 130, + 404, + 482, + 464 + ], + "type": "inline_equation", + "content": "[\\hat{b}_i^s,\\hat{b}_i^e ]" + }, + { + "bbox": [ + 130, + 404, + 482, + 464 + ], + "type": "text", + "content": " for start-end timestamps. This head is optimized by an L1 loss:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 235, + 468, + 481, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 468, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 235, + 468, + 481, + 483 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e g} = \\lambda_ {r e g} \\left(\\left| b _ {i} ^ {s} - \\hat {b} _ {i} ^ {s} \\right| + \\left| b _ {i} ^ {e} - \\hat {b} _ {i} ^ {e} \\right|\\right) \\tag {13}", + "image_path": "77eea03ab0aa1e1f841b4e644ff1c7a58671e20110992f961c7318ba206d75bb.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 487, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 525 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 487, + 482, + 525 + ], + "type": "inline_equation", + "content": "b_{i}^{s}" + }, + { + "bbox": [ + 130, + 487, + 482, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 487, + 482, + 525 + ], + "type": "inline_equation", + "content": "b_{i}^{e}" + }, + { + "bbox": [ + 130, + 487, + 482, + 525 + ], + "type": "text", + "content": " are the ground truths. This loss is only applied to frames inside the ground truth boundaries. We observe that using an L1 loss already works better than the combination of Smooth L1 Loss and GIoU Loss as in [26]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 531, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 482, + 570 + ], + "type": "text", + "content": "Saliency Prediction To obtain the saliency scores " + }, + { + "bbox": [ + 130, + 531, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\hat{s}_i" + }, + { + "bbox": [ + 130, + 531, + 482, + 570 + ], + "type": "text", + "content": " for HD, we calculate the cosine similarities between the adaptive pooled query features at the last refinement step " + }, + { + "bbox": [ + 130, + 531, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\widetilde{e}_q^K" + }, + { + "bbox": [ + 130, + 531, + 482, + 570 + ], + "type": "text", + "content": " and each token in the spatial-temporal features " + }, + { + "bbox": [ + 130, + 531, + 482, + 570 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 531, + 482, + 570 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 266, + 572, + 481, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 572, + 481, + 604 + ], + "spans": [ + { + "bbox": [ + 266, + 572, + 481, + 604 + ], + "type": "interline_equation", + "content": "\\hat {s} _ {i} = \\frac {h _ {i} ^ {\\top} \\widetilde {e} _ {q} ^ {K}}{| | h _ {i} | | _ {2} | | \\widetilde {e} _ {q} ^ {K} | | _ {2}} \\tag {14}", + "image_path": "f80b911398c0265c72fffd7a8c9c4db4a3800b323736069fde6471084948842a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 606, + 482, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 632 + ], + "type": "text", + "content": "The training objective is applied through a contrastive loss between sampled positive frames (with index " + }, + { + "bbox": [ + 130, + 606, + 482, + 632 + ], + "type": "inline_equation", + "content": "p \\in P" + }, + { + "bbox": [ + 130, + 606, + 482, + 632 + ], + "type": "text", + "content": ") and the adaptively pooled query " + }, + { + "bbox": [ + 130, + 606, + 482, + 632 + ], + "type": "inline_equation", + "content": "\\widetilde{e}_q^K" + }, + { + "bbox": [ + 130, + 606, + 482, + 632 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 205, + 636, + 481, + 665 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 636, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 205, + 636, + 481, + 665 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s a l} = - \\lambda_ {s a l} \\log \\frac {\\exp \\left(\\hat {s} _ {p} / \\tau\\right)}{\\exp \\left(\\hat {s} _ {p} / \\tau\\right) + \\sum_ {i \\in \\Theta} \\exp \\left(\\hat {s} _ {i} / \\tau\\right)} \\tag {15}", + "image_path": "a7c449e960b18146c945bd3526624449dd1c19c42ce64163d4da987ebbda8af7.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 334, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 334, + 91, + 448, + 102 + ], + "type": "text", + "content": "Reversed Recurrent Tuning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 178, + 482, + 256 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 2: Video moment retrieval results on Ego4D-NLQ, Charades-STA, and TACoS datasets. " + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "\\bigcirc" + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "text", + "content": " means using both CLIP-B/32 and SlowFast R-50 as feature extractors. " + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "\\bigcirc" + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "text", + "content": " indicates using CLIP-B/32 only. The best and second-best metrics are marked with bold and underline, respectively. Our method shows significant advantages in high-quality retrievals (R@0.7) even without a video backbone." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 178, + 482, + 256 + ], + "lines": [ + { + "bbox": [ + 132, + 178, + 482, + 256 + ], + "spans": [ + { + "bbox": [ + 132, + 178, + 482, + 256 + ], + "type": "table", + "html": "
MethodEgo4D-NLQ [10]Charades-STA [9]TACoS [43]
R@0.3R@0.5R@0.7mIoUR@0.3R@0.5R@0.7mIoUR@0.3R@0.5R@0.7mIoU
○2D-TAN [64]4.331.830.603.3958.7646.0227.5041.2540.0127.9912.9227.22
○VSLNet [62]4.542.401.013.5460.3042.6924.1441.5835.5423.5413.1524.99
○Moment-DETR [22]4.341.810.653.5365.8352.0730.5945.5437.9724.6711.9725.49
○UniVTG [26]7.283.951.324.9170.8158.0135.6550.1051.4434.9717.3533.60
○R²-Tuning (Ours)7.204.492.124.9470.9159.7837.0250.8649.7138.7225.1235.92
", + "image_path": "382be477779a22ef6bca88cd11c8cd55f1ecdf327c713374d1c7009af04aa285.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 279, + 481, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 279, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 279, + 481, + 304 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 132, + 279, + 481, + 304 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 132, + 279, + 481, + 304 + ], + "type": "text", + "content": " is the set of frame indices where " + }, + { + "bbox": [ + 132, + 279, + 481, + 304 + ], + "type": "inline_equation", + "content": "s_i < s_p" + }, + { + "bbox": [ + 132, + 279, + 481, + 304 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 132, + 279, + 481, + 304 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 132, + 279, + 481, + 304 + ], + "type": "text", + "content": " is a fixed temperature parameter set to 0.07." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 318, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 318, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 318, + 482, + 413 + ], + "type": "text", + "content": "Training & Inference The whole model is jointly optimized using a sum of the five losses mentioned above. In practice, the loss weights are set as " + }, + { + "bbox": [ + 132, + 318, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\lambda_{video} = 0.1" + }, + { + "bbox": [ + 132, + 318, + 482, + 413 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 318, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\lambda_{layer} = 0.1" + }, + { + "bbox": [ + 132, + 318, + 482, + 413 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 318, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\lambda_{cls} = 1.0" + }, + { + "bbox": [ + 132, + 318, + 482, + 413 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 318, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\lambda_{reg} = 0.1" + }, + { + "bbox": [ + 132, + 318, + 482, + 413 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 132, + 318, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\lambda_{sal} = 0.1" + }, + { + "bbox": [ + 132, + 318, + 482, + 413 + ], + "type": "text", + "content": ". During inference, we form an moment retrieval prediction by combining " + }, + { + "bbox": [ + 132, + 318, + 482, + 413 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 132, + 318, + 482, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 318, + 482, + 413 + ], + "type": "inline_equation", + "content": "b_{i}" + }, + { + "bbox": [ + 132, + 318, + 482, + 413 + ], + "type": "text", + "content": ", i.e., calculating the start-end timestamps from frame index " + }, + { + "bbox": [ + 132, + 318, + 482, + 413 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 132, + 318, + 482, + 413 + ], + "type": "text", + "content": " and boundary displacements " + }, + { + "bbox": [ + 132, + 318, + 482, + 413 + ], + "type": "inline_equation", + "content": "[b_i^s, b_i^e]" + }, + { + "bbox": [ + 132, + 318, + 482, + 413 + ], + "type": "text", + "content": ", while regarding " + }, + { + "bbox": [ + 132, + 318, + 482, + 413 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 132, + 318, + 482, + 413 + ], + "type": "text", + "content": " as the confidence. NMS with IoU threshold " + }, + { + "bbox": [ + 132, + 318, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\theta_{IoU} = 0.7" + }, + { + "bbox": [ + 132, + 318, + 482, + 413 + ], + "type": "text", + "content": " is applied to reduce duplicates. For highlight detection and video summarization, we directly use the frame-level output from saliency prediction " + }, + { + "bbox": [ + 132, + 318, + 482, + 413 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 132, + 318, + 482, + 413 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 434, + 230, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 230, + 448 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 230, + 448 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 462, + 316, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 462, + 316, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 462, + 316, + 474 + ], + "type": "text", + "content": "4.1 Datasets & Evaluation Metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 485, + 481, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 485, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 485, + 481, + 544 + ], + "type": "text", + "content": "We conduct experiments on six datasets covering various domains including daily vlogs & news (QVHighlights [22]), in-door scenes (TACoS [43] and Charades-STA [9]), egocentric videos (Ego4D-NLQ [10]), and sports (YouTube Highlights [47] and TVSum [45]). Details about the datasets are provided in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 545, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 482, + 665 + ], + "type": "text", + "content": "We adopt the same evaluation metrics with previous works [22, 26, 31]. To be specific, we compute Recall@1 with IoU threshold " + }, + { + "bbox": [ + 132, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\theta_{IoU} = 0.5" + }, + { + "bbox": [ + 132, + 545, + 482, + 665 + ], + "type": "text", + "content": " and 0.7, mean average precision (mAP) with " + }, + { + "bbox": [ + 132, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\theta_{IoU} = 0.5" + }, + { + "bbox": [ + 132, + 545, + 482, + 665 + ], + "type": "text", + "content": " and 0.7, and mAP with a series of thresholds [0.5:0.05:0.95] for MR on QVHighlights [22]. mAP and HIT@1 where positive samples are defined as with the saliency score of Very Good are adopted for HD. On Ego4D-NLQ [10], Charades-STA [9], and TACoS [43] datasets, we utilize Recall@1 with " + }, + { + "bbox": [ + 132, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\theta_{IoU} = \\{0.3, 0.5, 0.7\\}" + }, + { + "bbox": [ + 132, + 545, + 482, + 665 + ], + "type": "text", + "content": " and mIoU to measure the MR performance. On YouTube Highlights [47] for HD and TVSum [45] for VS, we follow [26, 31] and use the same train/val split with mAP and Top-5 mAP as metrics, respectively." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 184, + 146, + 430, + 275 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 3: Class-wise video highlight detection results (mAP) on YouTube Highlights dataset. " + }, + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "\\bigcirc" + }, + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "\\bigcirc" + }, + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "\\bigcirc" + }, + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": " denote image, video, and audio backbones, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 184, + 146, + 430, + 275 + ], + "lines": [ + { + "bbox": [ + 184, + 146, + 430, + 275 + ], + "spans": [ + { + "bbox": [ + 184, + 146, + 430, + 275 + ], + "type": "table", + "html": "
MethodDogGym.Par.Ska.Ski.Sur.Avg.
○ RRAE [58]49.035.050.025.022.049.038.3
○ GIFs [12]30.833.554.055.432.854.146.4
○ LSVM [47]60.041.061.062.036.061.053.6
○ LIM-S [54]57.941.767.057.848.665.156.4
○ SL-Module [55]70.853.277.272.566.176.269.3
○ PLD [53]74.970.277.957.570.779.073.0
○ MINI-Net [14]58.261.770.272.258.765.164.4
○ TCG [59]55.462.770.969.160.159.863.0
○ Joint-VA [3]64.571.980.862.073.278.371.8
○ CO-AV [25]60.966.089.074.169.081.174.7
○ UMT [31]65.975.281.671.872.382.774.9
○ UniVTG [26]71.876.573.973.373.282.275.2
○ R²-Tuning75.673.573.074.674.884.876.1
", + "image_path": "096e16fa5cd5b34df656829c9902a6b655d244d3917fca70281b627416c03760.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 162, + 328, + 452, + 457 + ], + "blocks": [ + { + "bbox": [ + 131, + 285, + 482, + 319 + ], + "lines": [ + { + "bbox": [ + 131, + 285, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 131, + 285, + 482, + 319 + ], + "type": "text", + "content": "Table 4: Class-wise video summarization results (Top-5 mAP) on TVSum dataset. " + }, + { + "bbox": [ + 131, + 285, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 131, + 285, + 482, + 319 + ], + "type": "text", + "content": " , " + }, + { + "bbox": [ + 131, + 285, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 131, + 285, + 482, + 319 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 131, + 285, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 131, + 285, + 482, + 319 + ], + "type": "text", + "content": " denote image, video, and audio backbones, respectively. Our method does not require any additional video/audio features." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 162, + 328, + 452, + 457 + ], + "lines": [ + { + "bbox": [ + 162, + 328, + 452, + 457 + ], + "spans": [ + { + "bbox": [ + 162, + 328, + 452, + 457 + ], + "type": "table", + "html": "
MethodVTVUGAMSPKPRFMBKBTDSAvg.
○ sLSTM [63]41.146.246.347.744.846.145.240.647.145.545.1
○ SG [33]42.347.247.548.945.647.346.441.748.346.646.2
○ LIM-S [54]55.942.961.254.060.447.543.266.369.162.656.3
○ Trailer [51]61.354.665.760.859.170.158.264.765.668.162.8
○ SL-Module [55]86.568.774.986.279.063.258.972.678.964.073.3
○ PLD [53]84.580.970.372.576.487.271.974.074.479.177.1
○ MINI-Net [14]80.668.378.281.878.165.857.875.080.265.573.2
○ TCG [59]85.071.481.978.680.275.571.677.378.668.176.8
○ Joint-VA [3]83.757.378.586.180.169.270.073.097.467.576.3
○ CO-AV [25]90.872.884.685.078.378.072.877.189.572.380.1
○ UMT [31]87.581.588.278.881.487.076.086.984.479.683.1
○ UniVTG [26]83.985.189.080.184.681.470.991.773.569.381.0
○ R²-Tuning85.085.991.081.788.887.478.189.290.374.785.2
", + "image_path": "4e17f519a0e140d7741faebe1e11ddee277471f2b2579cfb2037cad02af88ea9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 478, + 279, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 478, + 279, + 491 + ], + "spans": [ + { + "bbox": [ + 131, + 478, + 279, + 491 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 606 + ], + "type": "text", + "content": "In all experiments, we adopt the visual and text encoders of CLIP ViT-B/32 [41] as our backbones. The whole CLIP model is frozen during training, i.e., only the parameters in " + }, + { + "bbox": [ + 130, + 498, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 130, + 498, + 482, + 606 + ], + "type": "text", + "content": " Block, temporal feature pyramid, prediction heads are learnable. Without further specification, we set " + }, + { + "bbox": [ + 130, + 498, + 482, + 606 + ], + "type": "inline_equation", + "content": "K = 4" + }, + { + "bbox": [ + 130, + 498, + 482, + 606 + ], + "type": "text", + "content": ", i.e., attaching " + }, + { + "bbox": [ + 130, + 498, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 130, + 498, + 482, + 606 + ], + "type": "text", + "content": " Block to the last 4 layers of CLIP encoders. In " + }, + { + "bbox": [ + 130, + 498, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 130, + 498, + 482, + 606 + ], + "type": "text", + "content": " Block, the hidden size is set to 256. Only one transformer layer with post-norm style and 8 attention heads is used for temporal modeling. The number of layers for the temporal feature pyramid is set to 4 for QVHighlights, Ego4D-NLQ, Charades-STA, and TACoS, while 1 for other datasets. Please refer to the supplementary material for details." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 622, + 337, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 622, + 337, + 634 + ], + "spans": [ + { + "bbox": [ + 131, + 622, + 337, + 634 + ], + "type": "text", + "content": "4.3 Comparison with State-of-the-Arts" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Joint Moment Retrieval and Highlight Detection We first evaluate our method on QVHighlights [22] test split, the only dataset that supports both" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 334, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 334, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reversed Recurrent Tuning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 121, + 481, + 193 + ], + "blocks": [ + { + "bbox": [ + 230, + 114, + 382, + 120 + ], + "lines": [ + { + "bbox": [ + 230, + 114, + 382, + 120 + ], + "spans": [ + { + "bbox": [ + 230, + 114, + 382, + 120 + ], + "type": "text", + "content": "QUERY: A man in white is browsing through clothes at a department store." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 121, + 481, + 193 + ], + "lines": [ + { + "bbox": [ + 133, + 121, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 121, + 481, + 193 + ], + "type": "image", + "image_path": "1965b07216fe806ed6b3dc139235ce5e0991176c050cdbcb7668eead2bfb9fa6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 209, + 481, + 279 + ], + "blocks": [ + { + "bbox": [ + 254, + 201, + 358, + 208 + ], + "lines": [ + { + "bbox": [ + 254, + 201, + 358, + 208 + ], + "spans": [ + { + "bbox": [ + 254, + 201, + 358, + 208 + ], + "type": "text", + "content": "QUERY: Man stands on playground balance bridge." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 209, + 481, + 279 + ], + "lines": [ + { + "bbox": [ + 133, + 209, + 481, + 279 + ], + "spans": [ + { + "bbox": [ + 133, + 209, + 481, + 279 + ], + "type": "image", + "image_path": "af00914a1c1c11765454382834f2ba07a7ad8ee70f499d2e86dfc101f4d57220.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 287, + 482, + 344 + ], + "lines": [ + { + "bbox": [ + 130, + 287, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 287, + 482, + 344 + ], + "type": "text", + "content": "Fig.6: Visualization of joint moment retrieval and highlight detection results on QVHighlights val split. We compare our method with UniVTG [26], which is a strong baseline leveraging both additional temporal backbone and large-scale post pretraining. " + }, + { + "bbox": [ + 130, + 287, + 482, + 344 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 130, + 287, + 482, + 344 + ], + "type": "text", + "content": "-Tuning can accurately regress the boundaries of moments and predict highlight saliency scores through its novel feature refinement design." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 369, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 453 + ], + "type": "text", + "content": "moment retrieval and highlight detection. The results are shown in Table 1. The first group of methods utilizes more than one feature extractor. The second group is that with large-scale post pre-training (defined as the extra pre-training stage between loading backbone weights and training on downstream datasets.). Without any additional backbone and extra pre-training, " + }, + { + "bbox": [ + 130, + 369, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 130, + 369, + 482, + 453 + ], + "type": "text", + "content": "-Tuning achieves start-of-the-art performance compared with all previous methods with the fewest learnable parameters. Figure 6 visualizes the model predictions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 468, + 482, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 554 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 554 + ], + "type": "text", + "content": "Moment Retrieval We then evaluate our model on moment retrieval task in egocentric [10] and in-door [9, 43] domains. The results are shown in Table 2. We follow [26] and compare only with the methods using CLIP + SlowFast features. R²-Tuning still works better than all baselines with extra features. We also observe that R²-Tuning shows its significance on high-quality retrievals (R1@0.7), which requires accurate temporal modeling. This benefits from the fine-grained temporal modeling ability of reversed and recurrent designs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "text", + "content": "Highlight Detection & Video Summarization The performances of highlight detection on YouTube Highlights [47] and extractive video summarization on TVSum [45] are reported in Table 3 and Table 4, respectively. Following previous works [26, 31], we trained the model separately on each domain. The first group of methods is based on video backbones, which naturally have basic temporal modeling abilities. The second group of methods is enhanced by extra features such as image and audio. " + }, + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "text", + "content": "-Tuning can still reach the best performance when trained on small-scale data." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 122, + 216, + 201 + ], + "blocks": [ + { + "bbox": [ + 195, + 114, + 236, + 121 + ], + "lines": [ + { + "bbox": [ + 195, + 114, + 236, + 121 + ], + "spans": [ + { + "bbox": [ + 195, + 114, + 236, + 121 + ], + "type": "text", + "content": "(a) Naive Fusion" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 122, + 216, + 201 + ], + "lines": [ + { + "bbox": [ + 132, + 122, + 216, + 201 + ], + "spans": [ + { + "bbox": [ + 132, + 122, + 216, + 201 + ], + "type": "image", + "image_path": "b098f12db1afd9c6126afa7b3b43e782d7d214823744117017c0329b18d020d3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 221, + 123, + 304, + 201 + ], + "blocks": [ + { + "bbox": [ + 221, + 123, + 304, + 201 + ], + "lines": [ + { + "bbox": [ + 221, + 123, + 304, + 201 + ], + "spans": [ + { + "bbox": [ + 221, + 123, + 304, + 201 + ], + "type": "image", + "image_path": "0e6de240e9b71af89bdea3aed47b4cc0fef6d4015cc7bd8e90467e9068ad9801.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 209, + 482, + 243 + ], + "lines": [ + { + "bbox": [ + 132, + 209, + 482, + 243 + ], + "spans": [ + { + "bbox": [ + 132, + 209, + 482, + 243 + ], + "type": "text", + "content": "Fig. 7: Different feature refinement strategies for MR on QVHighlights val split. Single means utilizing the " + }, + { + "bbox": [ + 132, + 209, + 482, + 243 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 132, + 209, + 482, + 243 + ], + "type": "text", + "content": "-th layer only, while Mean denotes averaging the features from last " + }, + { + "bbox": [ + 132, + 209, + 482, + 243 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 132, + 209, + 482, + 243 + ], + "type": "text", + "content": " layers. Forward and Reversed indicate different feature refinement directions." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 122, + 394, + 201 + ], + "blocks": [ + { + "bbox": [ + 366, + 114, + 430, + 121 + ], + "lines": [ + { + "bbox": [ + 366, + 114, + 430, + 121 + ], + "spans": [ + { + "bbox": [ + 366, + 114, + 430, + 121 + ], + "type": "text", + "content": "(b) Recurrent Refinement" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 122, + 394, + 201 + ], + "lines": [ + { + "bbox": [ + 310, + 122, + 394, + 201 + ], + "spans": [ + { + "bbox": [ + 310, + 122, + 394, + 201 + ], + "type": "image", + "image_path": "666ac9bc80ecce0d6fce7c0a478eaba4331c40b70d2266f75bf76a9716c928da.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 399, + 122, + 482, + 201 + ], + "blocks": [ + { + "bbox": [ + 399, + 122, + 482, + 201 + ], + "lines": [ + { + "bbox": [ + 399, + 122, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 399, + 122, + 482, + 201 + ], + "type": "image", + "image_path": "106051b73f23d014df24ae663665f4a8411385db7f49bd4c94d0aa60f2d780a0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 136, + 252, + 335, + 369 + ], + "blocks": [ + { + "bbox": [ + 136, + 252, + 335, + 369 + ], + "lines": [ + { + "bbox": [ + 136, + 252, + 335, + 369 + ], + "spans": [ + { + "bbox": [ + 136, + 252, + 335, + 369 + ], + "type": "image", + "image_path": "e40999bd62e249bae10d2ea3518aca5a5f9203fd71f7617dbdd3a0a57f4a0bc4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 376, + 335, + 410 + ], + "lines": [ + { + "bbox": [ + 131, + 376, + 335, + 410 + ], + "spans": [ + { + "bbox": [ + 131, + 376, + 335, + 410 + ], + "type": "text", + "content": "Fig. 8: Visualization of the attention maps for query-modulated spatial pooling. Diverse regions are learned for different queries." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 360, + 252, + 464, + 369 + ], + "blocks": [ + { + "bbox": [ + 360, + 252, + 464, + 369 + ], + "lines": [ + { + "bbox": [ + 360, + 252, + 464, + 369 + ], + "spans": [ + { + "bbox": [ + 360, + 252, + 464, + 369 + ], + "type": "image", + "image_path": "fc12baa482f5c650a0433afd55e3c1eeee525a401b4429e6ef262c09407d0a8b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 376, + 482, + 410 + ], + "lines": [ + { + "bbox": [ + 340, + 376, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 340, + 376, + 482, + 410 + ], + "type": "text", + "content": "Fig. 9: Cosine similarities of video (x-axis) and text (y-axis) features among " + }, + { + "bbox": [ + 340, + 376, + 482, + 410 + ], + "type": "inline_equation", + "content": "7\\sim 12" + }, + { + "bbox": [ + 340, + 376, + 482, + 410 + ], + "type": "text", + "content": " CLIP layers." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 433, + 250, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 433, + 250, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 433, + 250, + 445 + ], + "type": "text", + "content": "4.4 Detailed Analysis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 452, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 452, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 452, + 482, + 536 + ], + "type": "text", + "content": "Multi-Granularity Features To investigate the significance of granularity flexibility, we set up a simple baseline that only attaches " + }, + { + "bbox": [ + 130, + 452, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2" + }, + { + "bbox": [ + 130, + 452, + 482, + 536 + ], + "type": "text", + "content": " Block to one layer of CLIP encoders. Then, we make use of the multi-layer information by averaging the features from the last " + }, + { + "bbox": [ + 130, + 452, + 482, + 536 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 452, + 482, + 536 + ], + "type": "text", + "content": " layers. These two variants are compared in Figure 7 (a). It can be observed that for the single-layer setting, higher-level features are more discriminative than lower-level ones. However, even a naive fusion strategy for the multi-layer information brings significant performance gains." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 546, + 482, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 619 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 619 + ], + "type": "text", + "content": "Reversed Recurrent Feature Refinement We then further justify our design by comparing two refinement directions (i.e., forward and reversed) in Figure 7 (b). Note that when " + }, + { + "bbox": [ + 130, + 546, + 482, + 619 + ], + "type": "inline_equation", + "content": "K = 1" + }, + { + "bbox": [ + 130, + 546, + 482, + 619 + ], + "type": "text", + "content": ", all four designs in Figure 7 are of the same architecture. When utilizing more layers, both forward and reversed refinement benefit from more information, while the reversed scheme statically performs better, as the multi-granularity features shall be refined from coarse to fine." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Query-Modulated Spatial Pooling We visualize the attention maps learned by query-modulated spatial pooling in Figure 8. A query generally has multiple keywords that provide discriminative information to localize moments. The" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reversed Recurrent Tuning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 146, + 313, + 210 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 312, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 312, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 312, + 137 + ], + "type": "text", + "content": "Table 5: Effectiveness justification of granularity calibration on QVHighlights val split." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 146, + 313, + 210 + ], + "lines": [ + { + "bbox": [ + 132, + 146, + 313, + 210 + ], + "spans": [ + { + "bbox": [ + 132, + 146, + 313, + 210 + ], + "type": "table", + "html": "
L VideoL layerMRHD
R1@0.5R1@0.7mAPmAPHIT@1
64.4848.6044.0137.9462.67
✓67.6851.6146.7439.8165.16
✓64.7148.8444.6038.9163.35
✓✓68.7152.0647.5940.5964.32
", + "image_path": "d057a0a670b4e97514c46f248cc81a7d1df26bee8ae95a162919bbc989856bb6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 319, + 146, + 481, + 211 + ], + "blocks": [ + { + "bbox": [ + 318, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 318, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 318, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 6: MR mAP with different query lengths on QVHighlights val split." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 146, + 481, + 211 + ], + "lines": [ + { + "bbox": [ + 319, + 146, + 481, + 211 + ], + "spans": [ + { + "bbox": [ + 319, + 146, + 481, + 211 + ], + "type": "table", + "html": "
Method#Words
1-1011-2021-3031-40≥41
QD-DETR [35]40.3042.3229.010.1026.67
EaTR [16]38.9439.5424.5812.0341.43
UniVTG [26]32.3832.6520.880.0031.11
R2-Tuning47.0849.1328.7967.2472.38
", + "image_path": "5ad0783f38989c6f78b679e01d873e40e8c2b4962bf3846627d31d4f31387c9e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 233, + 479, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 233, + 479, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 479, + 258 + ], + "type": "text", + "content": "token-wise attention in Eq. 4 can guide the model to focus on multiple regions simultaneously, as can be seen in the diverse but meaningful patterns." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 271, + 482, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 482, + 439 + ], + "type": "text", + "content": "Granularity Calibration To verify the effectiveness of granularity calibration, we also visualize the cosine distances of all the visual-query pairs from " + }, + { + "bbox": [ + 130, + 271, + 482, + 439 + ], + "type": "inline_equation", + "content": "7 \\sim 12" + }, + { + "bbox": [ + 130, + 271, + 482, + 439 + ], + "type": "text", + "content": " layers in Fig. 9. Before calibration, the visual-text features are not well-aligned. Adding the video-level constraint makes higher-level features aligned with each other (b), while adding layer-wise constraint makes the model distill diverse information across different layers (c). The diagonal in (d) verifies that both calibration constraints can maximize the mutual similarities of paired visual and text features. Some quantitative results are presented in Table 5. Both video-level and layer-wise contrastive can help align the multi-granularity visual & query semantics, while their combination can further boost the performances. We also group the queries in QVHighlights into 5 bins with different lengths (acting as multi-granularities) and compute the means of MR mAP in Tab. 6. Although most training queries are coarse-grained (" + }, + { + "bbox": [ + 130, + 271, + 482, + 439 + ], + "type": "inline_equation", + "content": "\\leqslant 30" + }, + { + "bbox": [ + 130, + 271, + 482, + 439 + ], + "type": "text", + "content": " words), R²-Tuning can easily generalize to fine-grained queries (" + }, + { + "bbox": [ + 130, + 271, + 482, + 439 + ], + "type": "inline_equation", + "content": "\\geqslant 31" + }, + { + "bbox": [ + 130, + 271, + 482, + 439 + ], + "type": "text", + "content": " words), surpassing previous methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 459, + 220, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 459, + 220, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 459, + 220, + 472 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 486, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 594 + ], + "type": "text", + "content": "This paper introduces R²-Tuning, a parameter- and memory-efficient transfer learning framework for video temporal grounding. It learns a lightweight side-adapter (R² Block) that is recurrently attached to the last few layers of a frozen CLIP to adaptively pool spatial details and refine temporal correlations. Two contrastive constraints (video-level and layer-wise) are utilized to calibrate the granularities of CLIP visual and text encoders. Experiments across three VTG tasks on six public benchmarks demonstrate the significance and effectiveness of the proposed scheme. We hope that the proposed framework sparks further research on efficient image-to-video transfer learning for untrimmed videos." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 614, + 241, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 614, + 241, + 628 + ], + "spans": [ + { + "bbox": [ + 132, + 614, + 241, + 628 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "This work was supported in part by Hong Kong Research Grants Council GRF-15229423, US NIH grant R01HD104969, and NSF award IIS-2239688." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "text", + "content": "1. Anne Hendricks, L., Wang, O., Shechtman, E., Sivic, J., Darrell, T., Russell, B.: Localizing moments in video with natural language. In: ICCV. pp. 5803-5812 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 162, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 480, + 194 + ], + "type": "text", + "content": "2. Apostolidis, E., Adamantidou, E., Metsai, A.I., Mezaris, V., Patras, I.: Video summarization using deep neural networks: A survey. Proceedings of the IEEE 109(11), 1838-1863 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 195, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 195, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 138, + 195, + 480, + 216 + ], + "type": "text", + "content": "3. Badamdorj, T., Rochan, M., Wang, Y., Cheng, L.: Joint visual and audio learning for video highlight detection. In: ICCV. pp. 8127-8137 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 217, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 480, + 237 + ], + "type": "text", + "content": "4. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: A clip-hitchhiker's guide to long video retrieval. Tech. Rep. arXiv:2205.08508 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 239, + 480, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 480, + 260 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 480, + 260 + ], + "type": "text", + "content": "5. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: CVPR. pp. 6299-6308 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 261, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 261, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 261, + 480, + 293 + ], + "type": "text", + "content": "6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 480, + 325 + ], + "type": "text", + "content": "7. Escorcia, V., Soldan, M., Sivic, J., Ghanem, B., Russell, B.: Temporal localization of moments in video collections with natural language. Tech. Rep. arXiv:1907.12763 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 326, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 480, + 347 + ], + "type": "text", + "content": "8. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: ICCV. pp. 6202-6211 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 348, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 480, + 369 + ], + "type": "text", + "content": "9. Gao, J., Sun, C., Yang, Z., Nevatia, R.: Tall: Temporal activity localization via language query. In: ICCV. pp. 5267-5275 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 370, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 480, + 403 + ], + "type": "text", + "content": "0. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: CVPR. pp. 18995-19012 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 480, + 424 + ], + "type": "text", + "content": "1. Gygli, M., Grabner, H., Riemenschneider, H., Van Gool, L.: Creating summaries from user videos. In: ECCV. pp. 505-520 (2014)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 425, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 480, + 445 + ], + "type": "text", + "content": "2. Gygli, M., Song, Y., Cao, L.: Video2gif: Automatic generation of animated gifs from video. In: CVPR. pp. 1001-1009 (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "text", + "content": "3. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 469, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 469, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 138, + 469, + 480, + 490 + ], + "type": "text", + "content": "4. Hong, F.T., Huang, X., Li, W.H., Zheng, W.S.: Mini-net: Multiple instance ranking network for video highlight detection. In: ECCV. pp. 345-360 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "type": "text", + "content": "5. Huang, S., Gong, B., Pan, Y., Jiang, J., Lv, Y., Li, Y., Wang, D.: Vop: Text-video co-operative prompt tuning for cross-modal retrieval. In: CVPR. pp. 6565-6574 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "type": "text", + "content": "6. Jang, J., Park, J., Kim, J., Kwon, H., Sohn, K.: Knowing where to focus: Event-aware transformer for video grounding. In: ICCV. pp. 13846-13856 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "type": "text", + "content": "7. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: ECCV. pp. 709-727 (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "type": "text", + "content": "8. Jiang, R., Liu, L., Chen, C.: Clip-count: Towards text-guided zero-shot object counting. In: ACM MM (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 590, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 610 + ], + "type": "text", + "content": "9. Ju, C., Han, T., Zheng, K., Zhang, Y., Xie, W.: Prompting visual-language models for efficient video understanding. In: ECCV. pp. 105-124 (2022)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "text", + "content": "20. Kong, Q., Cao, Y., Iqbal, T., Wang, Y., Wang, W., Plumbley, M.D.: Panns: Large-scale pretrained audio neural networks for audio pattern recognition. IEEE/ACM Transactions on Audio, Speech, and Language Processing 28, 2880-2894 (2020)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "21. Larsson, G., Maire, M., Shakhnarovich, G.: Fractalnet: Ultra-deep neural networks without residuals. Tech. Rep. arXiv:1605.07648 (2016)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reversed Recurrent Tuning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "22. Lei, J., Berg, T.L., Bansal, M.: Qvhighlights: Detecting moments and highlights in videos via natural language queries. In: NeurIPS (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "text", + "content": "23. Lei, J., Yu, L., Berg, T.L., Bansal, M.: Tv: A large-scale dataset for video-subtitle moment retrieval. In: ECCV. pp. 447-463 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 193 + ], + "type": "text", + "content": "24. Li, P., Xie, C.W., Xie, H., Zhao, L., Zhang, L., Zheng, Y., Zhao, D., Zhang, Y.: Momentdiff: Generative video moment retrieval from random to real. Tech. Rep. arXiv:2307.02869 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "text", + "content": "25. Li, S., Zhang, F., Yang, K., Liu, L., Liu, S., Hou, J., Yi, S.: Probing visual-audio representation for video highlight detection via hard-pairs guided contrastive learning. Tech. Rep. arXiv:2206.10157 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "26. Lin, K.Q., Zhang, P., Chen, J., Pramanick, S., Gao, D., Wang, A.J., Yan, R., Shou, M.Z.: Univtg: Towards unified video-language temporal grounding. In: CVPR. pp. 2794-2804 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "27. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: ICCV. pp. 2980-2988 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "text", + "content": "28. Lin, Z., Geng, S., Zhang, R., Gao, P., de Melo, G., Wang, X., Dai, J., Qiao, Y., Li, H.: Frozen clip models are efficient video learners. In: ECCV. pp. 388-404 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "29. Liu, L., Yu, B.X., Chang, J., Tian, Q., Chen, C.W.: Prompt-matched semantic segmentation. Tech. Rep. arXiv:2208.10159 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "text", + "content": "30. Liu, W., Mei, T., Zhang, Y., Che, C., Luo, J.: Multi-task deep visual-semantic embedding for video thumbnail selection. In: CVPR. pp. 3707-3715 (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "31. Liu, Y., Li, S., Wu, Y., Chen, C.W., Shan, Y., Qie, X.: Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In: CVPR. pp. 3042-3051 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 380, + 480, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 412 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 412 + ], + "type": "text", + "content": "32. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: Clip4clip: An empirical study of clip for end to end video clip retrieval. Neurocomputing 508, 293-304 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 413, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 434 + ], + "type": "text", + "content": "33. Mahasseni, B., Lam, M., Todorovic, S.: Unsupervised video summarization with adversarial LSTM networks. In: CVPR. pp. 202-211 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 435, + 480, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 467 + ], + "type": "text", + "content": "34. Moon, W., Hyun, S., Lee, S., Heo, J.P.: Correlation-guided query-dependency calibration in video representation learning for temporal grounding. Tech. Rep. arXiv:2311.08835 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "text", + "content": "35. Moon, W., Hyun, S., Park, S., Park, D., Heo, J.P.: Query-dependent video representation for moment retrieval and highlight detection. In: CVPR. pp. 23023-23033 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "text", + "content": "36. Nan, G., Qiao, R., Xiao, Y., Liu, J., Leng, S., Zhang, H., Lu, W.: Interventional video grounding with dual contrastive learning. In: CVPR. pp. 2765-2775 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "type": "text", + "content": "37. Ni, B., Peng, H., Chen, M., Zhang, S., Meng, G., Fu, J., Xiang, S., Ling, H.: Expanding language-image pretrained models for general video recognition. In: ECCV. pp. 1-18 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 555, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 577 + ], + "type": "text", + "content": "38. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. In: NeurIPS (2018)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "text", + "content": "39. Pan, J., Lin, Z., Zhu, X., Shao, J., Li, H.: St-adapter: Parameter-efficient image-to-video transfer learning. In: NeurIPS. pp. 26462-26477 (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "40. Qing, Z., Zhang, S., Huang, Z., Zhang, Y., Gao, C., Zhao, D., Sang, N.: Disentangling spatial and temporal learning for efficient image-to-video transfer learning. In: CVPR. pp. 13934-13944 (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "41. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "42. Rasheed, H., Khattak, M.U., Maaz, M., Khan, S., Khan, F.S.: Fine-tuned clip models are efficient video learners. In: CVPR. pp. 6545-6554 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "type": "text", + "content": "43. Regneri, M., Rohrbach, M., Wetzel, D., Thater, S., Schiele, B., Pinkal, M.: Grounding action descriptions in videos. Transactions of the Association for Computational Linguistics 1, 25-36 (2013)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "text", + "content": "44. Song, Y., Redi, M., Vallmitjana, J., Jeimes, A.: To click or not to click: Automatic selection of beautiful thumbnails from videos. In: CIKM. pp. 659-668 (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "type": "text", + "content": "45. Song, Y., Vallmitjana, J., Stent, A., Jaimes, A.: Tvsum: Summarizing web videos using titles. In: CVPR. pp. 5179-5187 (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "text", + "content": "46. Sun, H., Zhou, M., Chen, W., Xie, W.: Tr-detr: Task-reciprocal transformer for joint moment retrieval and highlight detection. In: AAAI (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 237, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 258 + ], + "type": "text", + "content": "47. Sun, M., Farhadi, A., Seitz, S.: Ranking domain-specific highlights by analyzing edited videos. In: ECCV. pp. 787-802 (2014)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "48. Sung, Y.L., Cho, J., Bansal, M.: Lst: Ladder side-tuning for parameter and memory efficient transfer learning. In: NeurIPS. pp. 12991-13005 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "type": "text", + "content": "49. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. Tech. Rep. arXiv:2302.13971 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "text", + "content": "50. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS. pp. 5998-6008 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "text", + "content": "51. Wang, L., Liu, D., Puri, R., Metaxas, D.N.: Learning trailer moments in full-length movies. In: ECCV. pp. 300-316 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "text", + "content": "52. Wang, Z., Wang, L., Wu, T., Li, T., Wu, G.: Negative sample matters: A renaissance of metric learning for temporal grounding. In: AAAI. pp. 2613-2623 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "text", + "content": "53. Wei, F., Wang, B., Ge, T., Jiang, Y., Li, W., Duan, L.: Learning pixel-level distinctions for video highlight detection. In: CVPR. pp. 3073-3082 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "text", + "content": "54. Xiong, B., Kalantidis, Y., Ghadiyaram, D., Grauman, K.: Less is more: Learning highlight detection from video duration. In: CVPR. pp. 1258-1267 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "text", + "content": "55. Xu, M., Wang, H., Ni, B., Zhu, R., Sun, Z., Wang, C.: Cross-category video highlight detection via set-based learning. In: ICCV. pp. 7970-7979 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "text", + "content": "56. Xu, Y., Sun, Y., Li, Y., Shi, Y., Zhu, X., Du, S.: Mh-detr: Video moment and highlight detection with cross-modal transformer. Tech. Rep. arXiv:2305.00355 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "text", + "content": "57. Yan, S., Xiong, X., Nagrani, A., Arnab, A., Wang, Z., Ge, W., Ross, D., Schmid, C.: Unloc: A unified framework for video localization tasks. In: ICCV. pp. 13623-13633 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "text", + "content": "58. Yang, H., Wang, B., Lin, S., Wipf, D., Guo, M., Guo, B.: Unsupervised extraction of video highlights via robust recurrent auto-encoders. In: ICCV. pp. 4633-4641 (2015)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 533, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 480, + 566 + ], + "type": "text", + "content": "59. Ye, Q., Shen, X., Gao, Y., Wang, Z., Bi, Q., Li, P., Yang, G.: Temporal cue guided video highlight detection with low-rank audio-visual fusion. In: ICCV. pp. 7950-7959 (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "text", + "content": "60. Yuan, T., Zhang, X., Liu, K., Liu, B., Jin, J., Jiao, Z.: Ucf-crime annotation: A benchmark for surveillance video-and-language understanding. Tech. Rep. arXiv:2309.13925 (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 600, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 620 + ], + "type": "text", + "content": "61. Yuan, Y., Ma, L., Wang, J., Liu, W., Zhu, W.: Semantic conditioned dynamic modulation for temporal sentence grounding in videos. NeurIPS 32 (2019)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "text", + "content": "62. Zhang, H., Sun, A., Jing, W., Zhou, J.T.: Span-based localizing network for natural language video localization. Tech. Rep. arXiv:2004.13931 (2020)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "63. Zhang, K., Chao, W.L., Sha, F., Grauman, K.: Video summarization with long short-term memory. In: ECCV. pp. 766-782. Springer (2016)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reversed Recurrent Tuning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 182 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "64. Zhang, S., Peng, H., Fu, J., Luo, J.: Learning 2d temporal adjacent networks for moment localization with natural language. In: AAAI. pp. 12870-12877 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "text", + "content": "65. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: CVPR. pp. 16816-16825 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "text", + "content": "66. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/RaFE_ Generative Radiance Fields Restoration/8e187213-9e85-483c-8762-09f958b4b9d0_content_list.json b/2024/RaFE_ Generative Radiance Fields Restoration/8e187213-9e85-483c-8762-09f958b4b9d0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b43a6c7d00247267f963483fbe4849d05d0ea7a9 --- /dev/null +++ b/2024/RaFE_ Generative Radiance Fields Restoration/8e187213-9e85-483c-8762-09f958b4b9d0_content_list.json @@ -0,0 +1,2386 @@ +[ + { + "type": "text", + "text": "RaFE: Generative Radiance Fields Restoration", + "text_level": 1, + "bbox": [ + 232, + 141, + 772, + 161 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhongkai Wu $^{1}$ , Ziyu Wan $^{2}$ , Jing Zhang $^{1\\star}$ , Jing Liao $^{2}$ , and Dong Xu $^{3}$", + "bbox": [ + 251, + 189, + 750, + 207 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{1}$ School of Software, Beihang University, China", + "2 City University of Hong Kong, China", + "3 The University of Hong Kong, China" + ], + "bbox": [ + 339, + 215, + 661, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ZhongkaiWu@buaa.edu.cn", + "bbox": [ + 415, + 260, + 586, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zkaiwu.github.io/RaFE", + "bbox": [ + 423, + 273, + 580, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. NeRF (Neural Radiance Fields) has demonstrated tremendous potential in novel view synthesis and 3D reconstruction, but its performance is sensitive to input image quality, which struggles to achieve high-fidelity rendering when provided with low-quality sparse input viewpoints. Previous methods for NeRF restoration are tailored for specific degradation type, ignoring the generality of restoration. To overcome this limitation, we propose a generic radiance fields restoration pipeline, named RaFE, which applies to various types of degradations, such as low resolution, blurriness, noise, compression artifacts, or their combinations. Our approach leverages the success of off-the-shelf 2D restoration methods to recover the multi-view images individually. Instead of reconstructing a blurred NeRF by averaging inconsistencies, we introduce a novel approach using Generative Adversarial Networks (GANs) for NeRF generation to better accommodate the geometric and appearance inconsistencies present in the multi-view images. Specifically, we adopt a two-level tri-plane architecture, where the coarse level remains fixed to represent the low-quality NeRF, and a fine-level residual tri-plane to be added to the coarse level is modeled as a distribution with GAN to capture potential variations in restoration. We validate RaFE on both synthetic and real cases for various restoration tasks, demonstrating superior performance in both quantitative and qualitative evaluations, surpassing other 3D restoration methods specific to single task. Please see our project website zkaiwu.github.io/RaFE.", + "bbox": [ + 259, + 329, + 743, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Neural Rendering $\\cdot$ Generative Model $\\cdot$ 3D Restoration $\\cdot$ Neural Radiance Fields", + "bbox": [ + 259, + 661, + 740, + 689 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 718, + 375, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, Neural Radiance Fields (NeRFs) [3,4,7,12,31,32,38,41] have achieved great success in novel view synthesis and 3D reconstruction. However, most NeRF methods are designed based on well-captured images from multiple views with calibrated camera parameters. In real-world applications of NeRF, the data", + "bbox": [ + 212, + 753, + 787, + 815 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 217, + 824, + 385, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4a5460e32f517f42f4190399c34c360676c2727dcc26e60c00bcfe5bdbd7926e.jpg", + "image_caption": [ + "Low Resolution Image Set" + ], + "image_footnote": [], + "bbox": [ + 223, + 148, + 295, + 204 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a275bf643662d43a087ca5cfdc10d6354ffda6dd7a1085956e515c6780a6de88.jpg", + "image_caption": [ + "NeRF" + ], + "image_footnote": [], + "bbox": [ + 303, + 148, + 454, + 210 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4217439998a23fc74817423ba11c85a3f2485ee38f6677a2760e70423673dc47.jpg", + "image_caption": [ + "Ours (x4)" + ], + "image_footnote": [], + "bbox": [ + 460, + 154, + 614, + 210 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4d648344512e15ab63f3009c91098c6e6fb7ae6b909086f0c8923a5ffcf89049.jpg", + "image_caption": [ + "Ours (x8)" + ], + "image_footnote": [], + "bbox": [ + 627, + 155, + 772, + 210 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f809c55d9dcd61fa009dcb98dd895cb661a4705e86d5c4ebfe4080b6608c8aa1.jpg", + "image_caption": [ + "Mixed Degradation Image Set" + ], + "image_footnote": [], + "bbox": [ + 218, + 227, + 310, + 282 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/be5d652e9257010efd4d055a512bb1619d37f2289a69293bebef0a8c66d7c3aa.jpg", + "image_caption": [ + "NeRF" + ], + "image_footnote": [], + "bbox": [ + 312, + 229, + 398, + 282 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b2bcd2da6c9795d089ddf88e7079cb751d3f5ea82b2b73d62fc049d3c32f8f24.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 400, + 229, + 485, + 282 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fee409245c09265cace66219303880641c1612dbe3b38933d06274815b766f38.jpg", + "image_caption": [ + "Real Blur Image Set" + ], + "image_footnote": [], + "bbox": [ + 486, + 229, + 586, + 282 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f2f6aaf43d3f0f50b8d335e1bc345bd4cd44cac233cf8f0d0c15b3090c70feca.jpg", + "image_caption": [ + "NeRF" + ], + "image_footnote": [], + "bbox": [ + 589, + 229, + 718, + 282 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fd235358287efbd47fb0d4cc1bae7ca0c154c40b971b02d1e82a2ecd3cfaf8fa.jpg", + "image_caption": [ + "Ours", + "Fig. 1: High-quality restoration of radiance fields from various types of degradations. Given only degraded images, our method can restore high-quality NeRF. It is a generic approach that can be applied to various types of degradations, resulting in refinement of both geometry and appearance." + ], + "image_footnote": [], + "bbox": [ + 720, + 229, + 782, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "capture or transmission process often introduces various forms of image degradations, such as noise generated during photography in low-light conditions [29,33] and blur caused by camera motion [27,44], or JPEG compression and down-sampling during transmission [2,43]. Simply restoring degraded images frame-by-frame can result in inconsistencies of geometry and appearance across different viewpoints. Directly reconstructing 3D models over these per-frame restoration results can easily induce inferior quality since current NeRF methods heavily rely on pixel-wise independent ray optimization with local computations, which are highly susceptible to noise and degradation.", + "bbox": [ + 212, + 402, + 784, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several NeRF variants have attempted to reconstruct 3D scenes with degraded multi-view images by introducing specific strategies or additional constraints for the optimization of radiance fields. For example, [19,20,27] deal with image blur artifacts by modeling the degradation kernel with NeRF, while [2,43] super-sampling on rays or tri-planes to obtain high-resolution 3D from low-resolution observations. Additionally, [31] modifies NeRF to reconstruct the scene in linear HDR space by supervising directly on noisy raw images to address the noise generated in low-light conditions. [53] tries to improve the view synthesis quality by removing NeRF-specific rendering artifacts. Even with great success, all these approaches are only designed to handle specific types of degradation. To the best of our knowledge, currently there is no generic pipeline which supports the restoration of radiance fields under various types of degradation.", + "bbox": [ + 212, + 537, + 785, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose RaFE, a generic NeRF restoration framework that enables high-quality radiance field reconstruction from captured images containing various types of degradations in a generative manner. Firstly, we leverage the success of off-the-shelf image restoration methods to restore multi-view images with different forms of degradations, such as super-resolution, deblurring, denoising, removing compression artifacts, or a combination thereof. It is important to note that since the images from different views are independently restored, there inevitably exist geometric and appearance inconsistencies between them.", + "bbox": [ + 212, + 719, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Wu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Naively optimizing a radiance field with these refined images would average out the inconsistencies and result in blurry outputs. To overcome this challenge, our insight here is, instead of reconstructing a single 3D using inconsistent frames, we could consider these restored multi-view images as the renderings from multiple distinct high-quality NeRF models with varied geometry and appearance. In this case, we abandon the commonly-used pixel-wise reconstruction objective and propose to leverage the generative adversarial networks (GANs) to model the distribution of these different high-quality NeRF models, which could effectively capture the inherent variability in ill-posed inverse problem, allowing for a better accommodation of the inconsistencies present in different views.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, our pipeline consists of two main stages. In the first stage, based on the degradation type, we employ the corresponding off-the-shelf image restoration methods [17, 25, 34, 35, 46, 49] to obtain a set of high-quality multiview images. In practice, we prefer choosing restoration methods which have strong capabilities to recover high-quality and realistic texture details. In the second stage, we train a 3D generative model based on these restored multi-view images. Drawing inspirations from recent 3D generation works [6, 9, 37, 40], we construct a convolutional neural network (CNN) to generate tri-plane features, which are subsequently sampled and decoded into density and colors using MLP networks for NeRF rendering. Here, instead of generating single-level tri-plane features as previous works did, we decompose the tri-planes into two levels. The coarse-level tri-planes are constructed directly from low-quality images and remain fixed during training, representing the coarse structure of the modeled 3D distribution. Simultaneously, we train a generator to output the diverse fine-level tri-plane features, which act as residuals to be added to the coarse-level features for NeRF rendering. By focusing on learning the residual representations instead of the entire tri-planes for NeRF, we simplify the modeling and learning of restoration variations since we only need to learn the details while the coarse structure is provided by coarse-level tri-planes, which makes great improvement in rendering quality for more complex regions. To train the generator, we adopt an adversarial loss defined on rendered 2D images to encourage them to be indistinguishable from the high-quality restored images. We also incorporate a perceptual loss between the rendered images and the restored images to calculate structure constraints. Additionally, we propose patch sampling strategies to stabilize the generator training procedure. Once the generator has been trained, we can generate restored radiance fields with high quality renderings and a certain level of diversity by sampling different code in the latent space.", + "bbox": [ + 212, + 311, + 787, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We conducted extensive experiments to validate the effectiveness of our method, both qualitatively and quantitatively. The experimental results showcase the superiority of our approach in various restoration tasks, such as super-resolution (upper row of Figure 1), camera motion blur (a real-world case at the right part of the lower row of Figure 1) and the restoration of mixed degradation consisting of noise, blur, and compression (left part of the lower row of Figure 1). Our method not only generates images with richer and enhanced tex", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "RaFE: Generative Radiance Fields Restoration", + "bbox": [ + 416, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ture details but also achieves significant improvements in geometric refinement, as demonstrated by the mesh visualization in Figure 1.", + "bbox": [ + 215, + 146, + 782, + 175 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To summarize, our contributions are three-fold:", + "bbox": [ + 240, + 176, + 581, + 191 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a generic radiance fields restoration pipeline that is applicable to various types of degradations.", + "- We introduce a generative method for NeRF restoration that enables better accommodation of geometric and appearance inconsistencies present in the multi-view images, thus allowing us to incorporate the success of image restoration into 3D restoration.", + "- We show the restoration method performs well on various degradation scenarios with both enhanced appearance and refined geometry." + ], + "bbox": [ + 225, + 203, + 782, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 345, + 395, + 361 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 2D image restoration", + "text_level": 1, + "bbox": [ + 215, + 376, + 434, + 391 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image restoration is a long-standing problem in low-level vision domain and significant progress has been achieved in different specific tasks including image super-resolution, deblur, denoise and blind restoration. Previously reconstruction-based methods [11,21,23,52,54] show their success in these tasks. However, those reconstruction-based methods are struggling to generate abundant high-quality details. Subsequently, generative restoration methods [10,17,25,34,35,39,42,46, 49], particularly those based on diffusion model, have shown the great capability to render high-quality details. Deepfloyd [35] proposes a super-resolution model, which concatenates the low-resolution input with random noise at pixel level as the condition image. DiffBIR [25] designs a degraded pipeline to simulate real-world degradation and utilizes the pre-trained diffusion model to generate photorealistic images. For camera motion blur, HiDiff [10] recovers images by using diffusion to generate features with abundant detailed information.", + "bbox": [ + 215, + 401, + 787, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Radiance Fields Restoration", + "text_level": 1, + "bbox": [ + 215, + 619, + 493, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "NeRF restoration aims to reconstruct high-quality NeRF given only degraded images with various artifacts such as blur, noise, or low resolution. Up to now, several works [2,8,14,19,20,26,27,31,33,43,44,53] have explored this task under specific types of degradation. [20,27,44] deal with blurred input images by designing blur kernel or optimizing camera paths for NeRF rendering process. For NeRF super-resolution, [43] increased the ray sampling density, forcing multiple rays to render pixels equal to the same pixel, and applied a 2D refinement model to get final output images. [14] introduces a CEM [1] refinement model to adjust the output of the super-resolution model for better multi-view consistency. However, the CEM refinement model ruins the structure details of the images and inconsistencies still exist, leading to a smooth reconstruction result. [31,33] mainly focus on the noise degradation of the input image. [31] modifies NeRF to reconstruct the scene in linear HDR color space to address the noise generated", + "bbox": [ + 215, + 643, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Wuet al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "during image post-processing, while [33] achieve NeRF restoration on noisy images by using a noise-awarded encoder to aggregate features across views. [53] considers solving the degradation that occurs in NeRF reconstruction by training a 2D refinement model using simulated degraded images for typical NeRF-style artifacts, but they can not achieve refinement on geometry since the restoration only happens on rendered views. None of the existing approaches can restore NeRF in 3D space directly with flexible forms of degradation. NVSR [2] archives 3D geometry refinement by upsampling tri-plane representation, but their training processing requires tremendous amounts of 3D data, which are extremely hard to obtain in practice. [26] achieves NeRF translation by designing a VAE-like NeRF to model the distribution of the underlying 3D edits. By contrast, our method has the ability to handle more flexible forms of degradation and restore 3D geometry and appearance with the only need of an image set for an object or scene, making the 3D restoration more practical in real-world applications.", + "bbox": [ + 212, + 146, + 787, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 385, + 330, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we elaborate the details of the proposed RaFE method. We introduce how to refine the degraded views using pretrained 2D restoration model, to capture the high-quality appearance distribution in Sec. 3.1. Then, we describe our generative restoration framework including the neural representation, generator architecture and optimization in Sec. 3.2. The training strategy is introduced in Sec. 3.3. The overall pipeline could be found in Fig. 2.", + "bbox": [ + 212, + 420, + 787, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 High-quality Image Restoration", + "text_level": 1, + "bbox": [ + 215, + 537, + 522, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The recent success of the 2D image restoration task is dominated by the denoising diffusion probabilistic model, benefiting from its powerful appearance generation capability and extensive prior knowledge. Hence in this paper, we mainly borrow the recent diffusion-based restoration methods [25,35] to obtain the high-quality multi-view images from their low-quality counterparts. It should be noted that our 3D restoration framework could also work well based on non-diffusion restoration approaches, which has been verified by our experiments.", + "bbox": [ + 212, + 566, + 787, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To assist in the restoration of image details, we employ powerful image caption models [5,22] to get an accurate textual prompt for the scene, denoted as $P$ . Given a set of multi-view low-quality images $I_{l}$ , the textual prompt is produced by selecting a view that contains as much information of the scene as possible, e.g. a side view of a synthetic Lego model. The prompt $P$ and low-quality images $I_{l}$ are then fed into the restoration model [25,35] to achieve high-quality per-frame refinement. We denote the restored high-quality images as $I_{h}$ .", + "bbox": [ + 212, + 672, + 787, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The diffusion model-based image restoration methods can effectively reconstruct high-quality images from various real-world degradation types. Moreover, thanks to the powerful generalization ability of diffusion models preserved in these restoration methods, we can handle the scene restoration in open domain", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "RaFE: Generative Radiance Fields Restoration", + "bbox": [ + 416, + 114, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/21c1cae40575dd9c4b3ead499c76e68862b7c764397c2b5e76a7f6993bd6ddae.jpg", + "image_caption": [ + "Fig. 2: Overview of our pipeline. Given multi-view degraded images, we utilize the off-the-shelf methods to restore high-quality multi-view images. Then, we train our Generative Restoration NeRF to generate a high-quality scene." + ], + "image_footnote": [], + "bbox": [ + 225, + 145, + 782, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "without the need of re-training. However, though the restored images have better quality, more variations among different views occur due to the generative nature of these restoration models, leading to serious multi-view inconsistencies. Next, we focus on dealing with this issue with the proposed generative pipeline.", + "bbox": [ + 214, + 377, + 787, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Generative NeRF Restoration", + "text_level": 1, + "bbox": [ + 215, + 463, + 511, + 478 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To deal with geometric and appearance inconsistencies across views, we treat these restored multi-view images distributed similarly to the rendered images from multiple slightly different high-quality NeRF models. Consequently, instead of directly fitting a NeRF model using the refined high-quality views which usually leads to blurry reconstruction, we are trying to learn the distribution of these diverse NeRF models by leveraging a generative method, allowing us to sample distinct restored 3D under the same degraded inputs.", + "bbox": [ + 212, + 491, + 787, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D Representation. Following the recent 3D generative model [6,9,37], we adopt the hybrid explicit-implicit tri-plane representation for the feature field. This representation combines both explicit and implicit components to effectively model the density and RGB values. More specifically, to obtain the descriptor for any query location $\\pmb{x} \\in \\mathbb{R}^3$ , we project the point onto three planes $(P_{xy}, P_{yz}, P_{zx})$ to retrieve the corresponding features $(f_{xy}, f_{yz}, f_{zx})$ using interpolation. Then we calculate the mean value of these three sampled features to obtain the final feature representation.", + "bbox": [ + 214, + 597, + 787, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Once we obtain the features for a point along the ray, we use two MLPs to decode the density and RGB values. The first MLP, denoted as $\\mathcal{M}_{\\text{dens}}$ , maps the features to the density value $\\sigma \\in \\mathbb{R}$ and a color feature $\\pmb{f}_{color}$ . The second MLP, denoted as $\\mathcal{M}_{color}$ , takes the color feature and the view direction as inputs and maps them to the RGB value $\\pmb{c} \\in \\mathbb{R}^3$ . We empirically show that incorporating the view direction allows us to effectively model view-dependent effects, particularly when there are non-Lambertian surfaces in the scene. After obtaining the densities and RGB values for each sampled point along the emitted ray, we can", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Wu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "apply the classic volumetric rendering to obtain the color value for each pixel:", + "bbox": [ + 215, + 146, + 774, + 161 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nC (\\boldsymbol {r}) = \\sum_ {i = 1} ^ {N} T _ {i} (1 - \\exp (- \\sigma_ {i} \\delta_ {i})) \\boldsymbol {c} _ {i}, \\quad T _ {i} = \\exp (- \\sum_ {j = 1} ^ {i - 1} \\sigma_ {i} \\delta_ {i}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 169, + 784, + 205 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\pmb{c}_i, \\sigma_i$ represents RGB and density value of the $i_{th}$ point alone the ray, respectively. $N$ is the sampled number and $\\delta_i$ is the distance between samples.", + "bbox": [ + 212, + 212, + 782, + 241 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "NeRF Generator. Given restored multi-view images, we regard them as renderings from several diverse high-quality NeRF models with varied geometry and appearance. The distribution formed by these distinct NeRF models, formulated as $p_r(\\phi | I_h)$ , can be modeled by a generator. With the hybrid explicit-implicit tri-plane representation, we introduce a StyleGAN2 [15]-like CNN-based generator, which receives a latent code $\\mathbf{w}$ mapped from random code $\\mathbf{z}$ to generate fine tri-planes $P_f$ for high-quality NeRF rendering. To fully leverage the degraded images and make the generator only focus on generating the necessary refinements, we also pre-train a coarse NeRF model with tri-plane features denoted as $P_c$ using the input low-quality images. Specifically, the coarse NeRF model consists of a coarse tri-plane $P_c$ features and a decoder (where the decoder of coarse tri-planes is shared with the one of fine tri-planes). To obtain the final refined tri-planes $P$ , we combine the reconstructed tri-plane representations with the residual tri-planes generated by the generator via: $\\mathbf{f} = \\mathbf{f}_c + \\mathbf{f}_f$ . By merging these components, RaFE effectively captures both the global geometric guidance provided by the coarse NeRF and the local refinements learned from 3D generator, enabling us to obtain the restored tri-planes that exhibit enhanced geometric accuracy and appearance fidelity.", + "bbox": [ + 212, + 241, + 784, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Optimization. To supervise the generator and NeRF parameters, we propose to minimize the distribution discrepancy between the rendered images and the restored high-quality images. We adopt a saturate GAN [13] loss with an image level discriminator. Specifically, we treat the high-quality images restored by 2D model as the real samples while the rendered images as the fake samples, and utilize adversarial loss with R1 regularization between the real and fake images:", + "bbox": [ + 212, + 513, + 784, + 604 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {a d v} = \\mathbb {E} _ {\\boldsymbol {z} \\sim p _ {z}, \\boldsymbol {\\theta} \\sim p _ {\\theta}} [ f (D (G (\\boldsymbol {z}, \\boldsymbol {\\theta})) ] \\\\ + \\mathbb {E} _ {I _ {h} \\sim p _ {h}} [ f (- D (I _ {h})) + \\lambda \\| \\nabla D (I _ {h}) \\| ^ {2} ], \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 351, + 627, + 784, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $z, \\theta$ represent random code and view point, respectively, and $I_h$ is the restored high quality image. However, we observed relying solely on a GAN loss for training can lead to significant geometric mismatches between the restored images and rendered views. We argue that although the GAN loss helps align the distribution of 2D renderings, it still lacks geometry-level constraints. Therefore, we also incorporate a perceptual loss that encourages the rendered images to resemble the geometry of pre-frame restoration.", + "bbox": [ + 212, + 670, + 784, + 775 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {g e o m e t r y}} = L P I P S \\left(I _ {h}, G (\\boldsymbol {z}, \\boldsymbol {\\theta})\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 785, + 784, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $LPIPS(\\cdot, \\cdot)$ refers to the learned perceptual image patch similarity proposed in [50]. $I_h$ is the restored high quality image paired with view point $\\theta$ , and", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "RaFE: Generative Radiance Fields Restoration", + "bbox": [ + 416, + 114, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$G(\\pmb{z}, \\pmb{\\theta})$ are the rendered image using the same training view $\\pmb{\\theta}$ and a random sampled latent code $\\pmb{z}$ . We also supervise the coarse NeRF branch:", + "bbox": [ + 215, + 146, + 784, + 176 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e c} = \\mathbb {E} _ {\\boldsymbol {\\theta} \\sim p _ {\\theta}} \\left[ \\| G _ {c} (\\boldsymbol {\\theta}) - I _ {l} ^ {\\theta} \\| ^ {2} \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 188, + 784, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $p_{\\theta}$ indicate the view point distribution, and $I_l^\\theta$ is the low quality image corresponding to view point $\\pmb{\\theta}$ . Overall, the complete training objective is:", + "bbox": [ + 215, + 215, + 782, + 246 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {\\text {g e o m e t r y}} \\mathcal {L} _ {\\text {g e o m e t r y}} + \\lambda_ {\\text {a d v}} \\mathcal {L} _ {\\text {a d v}} + \\lambda_ {\\text {r e c}} \\mathcal {L} _ {\\text {r e c}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 256, + 784, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\lambda_{\\text{geometry}}, \\lambda_{\\text{adv}}, \\lambda_{\\text{rec}}$ are trade-off parameters.", + "bbox": [ + 215, + 280, + 588, + 296 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Patch-based Training Strategy", + "text_level": 1, + "bbox": [ + 215, + 315, + 511, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During the training process, fully rendering the entire image, such as $256^2$ or $512^2$ pixels, and then feeding it into the discriminator can be computationally intensive and resource-consuming. This is because volume rendering requires the computation of density and color values for sampling points along the ray for each pixel, which can become prohibitively expensive for a whole image. Therefore, we only render a patch of the view at a time (i.e. $64^2$ ). The rendered patches and high-quality image patches are randomly selected, and the discriminator in Eq. 2 receives patches of rendered images and patches of high-quality images as fake images and real images respectively. For the perceptual loss in Eq. 3, the rendered patches and the high-quality patches have the same spacial coordinate on their original images to guarantee that the cropped patches have the same semantic and local structure.", + "bbox": [ + 212, + 338, + 784, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "One limitation of using patches instead of entire images for training is that it is challenging to sample the local scenes at the boundary regions of images, particularly for forward-facing data. The imbalanced training distribution will result in mode collapse during the training process. To mitigate this issue, we employ a beta sampling strategy to determine the positions of the sampled patches. This strategy ensures that patches in the boundary regions of the image are adequately sampled. More specifically, the beta sampling can be formulated as:", + "bbox": [ + 212, + 520, + 784, + 626 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\delta_ {x}, \\delta_ {y} \\sim \\operatorname {B e t a} (\\beta (t), \\beta (t)), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 636, + 784, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\delta_x$ and $\\delta_y$ denote the position offset in $x$ and $y$ directions respectively, while $\\beta(t)$ are linearly annealed from $\\beta(0) = 1$ to some final value $\\beta(T)$ smaller than 1. By using the beta sampling strategy, we can maintain a more balanced distribution of training data that focuses more on the boundary patches, alleviating mode collapse issue and improving the overall training stability.", + "bbox": [ + 212, + 660, + 784, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 757, + 375, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Setup", + "text_level": 1, + "bbox": [ + 215, + 787, + 310, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets. We evaluate our model on the NeRF-Synthetic benchmark dataset [31], which contains 8 synthetic objects with images taken from different viewpoints", + "bbox": [ + 215, + 809, + 792, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Wu et al.", + "bbox": [ + 271, + 114, + 334, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "uniformly distributed in the hemisphere. Following the original setting, We hold 200 viewpoints for generating high-quality training data and 200 viewpoints for testing. Further, to demonstrate the generalization ability of our method, we also evaluate our method on complex real-world LLFF scenes [30] which consists of 8 scenes captured with roughly forward-facing images. We also demonstrate the superior performance of RaFE on real-world blur [27] and noise [29] data.", + "bbox": [ + 212, + 146, + 782, + 234 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following the common practice of 3D reconstruction, we try to evaluate each method with two standard image quality metrics: peak signal-to-noise ratio (PSNR), structural similarity index (SSIM) [47], which however could not reflect the real 3D restoration performance according to our observations. Due to the generative characteristic of RaFE, the recovered radiance field from RaFE is very high-quality, but may not faithfully follow the \"ground-truth\" 3D, since inversing a degraded signal is a highly ill-posed problem. Moreover, we also found the baselines with better scores over PSNR and SSIM still pose a degraded appearance with smooth texture details, as shown in Figure 3b. Hence, a better way is to leverage perceptual metric: learned perceptual image patch similarity (LPIPS) [50] which computes the mean squared error (MSE) between normalized features from all layers of a pre-trained VGG [36] encoder and is deemed to better correlate with human perception. We also leverage the latest non-reference-based image quality assessment metrics including LIQE [51] and MANIQA [48] to demonstrate the superior rendering quality of our method.", + "bbox": [ + 212, + 239, + 787, + 467 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details. We implement all the experiments by PyTorch. For the 2D generator and discriminator, we use a convolutional-based generator and discriminator used in StyleGAN2 [16]. In all experiments, we choose Adam optimizer for all the modules in our pipeline, with hyperparameters $\\beta_{1} = 0, \\beta_{2} = 0.99$ . We use learning rate $2 \\times 10^{-3}$ for both generator and discriminator. For loss weights, we use $\\lambda_{mimic} = 0.5$ , $\\lambda_{adv} = 1.0$ , and $\\lambda_{rec} = 1.0$ for almost all experiments. We evaluate RaFE framework on 4 different 3D restoration tasks:", + "bbox": [ + 212, + 469, + 787, + 575 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- $4 \\times$ Super-Resolution: On the blender dataset, we resize the image to $64 \\times 64$ resolutions to get low-resolution images. As for LLFF data, we first center crop the training image to $188 \\times 252$ to adapt to our $4 \\times$ super-resolution task and then resize to $47 \\times 63$ to get low-resolution images. And we use Deepfloyd [35] for 2D super-resolution.", + "- Deblur: On the LLFF dataset, we construct camera motion blur by applying the blur kernel following equation $I_{blur} = A * I_{clear}$ , where $A$ is the blur kernel and $*$ stands for the convolution operator. Different from the blur dataset proposed in Deblur-NeRF [27], which contains blurred images with varying degrees of blur, and even includes a certain amount of high-resolution images, we apply a large blur kernel size (e.g. 13) and a more complex camera motion path to all the images in the dataset. We additionally construct consistent blur datasets by using the same blur kernel to the training image in a scene, which means the camera motion trajectory is the same for the training image set. The resolution of images is the same as the super-resolution task. We use HiDiff [10] to recover high-quality images." + ], + "bbox": [ + 225, + 595, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "RaFE: Generative Radiance Fields Restoration", + "bbox": [ + 416, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "- Denoise: On the LLFF dataset, we follow the noise model used in [28, 33] and we get the noisy version of a clean image $I$ according to equation $I_{\\text{noisy}}(x) = \\mathcal{N}(I(x), \\delta_r^2 + \\delta_s^2 I^2(x))$ , where $\\sigma_r$ is the signal-independent read-noise parameter, $\\sigma_s$ is the signal-dependent shot-noise. Following [28, 33], we use the Gain level to represent the noise strength. We use gain levels $= 8$ to get our noisy image. We use DiffBIR [25] to get high-quality images.", + "bbox": [ + 225, + 145, + 787, + 236 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "- Mixed degradation: The degradation pipeline consists of three stages: blur, noise, and JPEG compression. First, we utilize Gaussian blur with a radius of 7 for the blender dataset and 3 for the LLFF dataset. Second, we add noise with std 25. And then, we apply JPEG compression. The quality of JPEG compression is 50. The resolution of images is the same as the super-resolution task. We use DiffBIR [25] to get high-quality images.", + "bbox": [ + 225, + 237, + 787, + 328 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Results", + "text_level": 1, + "bbox": [ + 217, + 348, + 323, + 361 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baseline Methods. For general tasks, we try a baseline that firstly restores the degraded images and uses the restored high-quality images to reconstruct a NeRF directly, denoted as NeRF-Perframe. We also use the 2D-based restoration model SwinIR [24] to do the per-view refinement for the renderings of NeRF trained by degraded image, denoted as NeRF-SwinIR. Note that we do not evaluate NeRF-SwinIR for the deblur task since there are no corresponding checkpoints.", + "bbox": [ + 212, + 371, + 785, + 462 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To more thoroughly test the effectiveness of our method, we also select some task-specific competitors. For super-resolution task, we choose NeRF-SR [43], Neural Volume Super-Resolution (NVSR) [1] as baselines. For mixed degradation, since there is no existing method tailored for mixed degradation, we choose NeRFLiX, which tries to solve the NeRF-like degradation by training a 2D refinement model using degradation images constructed by a degradation simulator for typical NeRF-style artifacts. We consider this to be the most relevant method. For the deblur task, we compare with two state-of-the-art methods Deblur-NeRF [27] and BAD-NeRF [45]. Deblur-NeRF designs a learnable blur kernel and applies it to rays to simulate the degrading process and BAD-NeRF directly models the camera trajectories to solve motion blur. For the denoising task, we compare with NAN [33], which uses a noise-aware encoder to aggregate the feature of multi-view images for restoration.", + "bbox": [ + 212, + 462, + 785, + 657 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Quantitative Results. We conduct extensive quantitative comparisons with various baselines across different restoration tasks in Tab. 1a for super-resolution, Tab. 2b and Tab. 2a for deblurring, Tab. 2c for denoising and Tab. 1b for mixed degradation. As analyzed before, although most of the time our method falls slightly behind on the reconstruction metrics like PSNR and SSIM compared with other baselines, which only measure local pixel-aligned similarity between the rendered novel views and the ground truth images, are less indicative since uncertainties naturally exist in generative procedure. Taking the super-resolution results on Blender data as an example (Tab. 1a), the simplest baseline NeRF-Perframe has already achieved the best reconstruction metrics, but as shown in Figure. 3a, its visual quality is vastly inferior compared with our results. Through the error map, we found the misalignment between the generated 3D and input", + "bbox": [ + 212, + 659, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Wuet al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/c9a0709fa184a3568877df92199e5fb4bfc74c60b33df35fd922378921f6f1c4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BleederLLFFLLFF
PSNR†SSIM†LPiPS‡LIPE†MANIQA†PSNR†SSIM†LPiPS‡LIPE†MANIQA†MethodPSNR†SSIM†LPiPS‡LIPE†MANIQA†PSNR†SSIM†LPiPS‡LIPE†MANIQA†
NevRF-SR26.850.9120.1352.3720.36622.650.7020.3271.3200.220
NVSR26.130.8790.1322.3010.34321.290.6060.4422.1080.177
NeRF-SwnIR24.070.8780.1193.060.38121.730.6550.3651.8080.267
NeRF-Perframe27.390.9220.0833.2760.38623.680.7450.2611.4170.257
Ours24.990.9010.0624.6210.54323.850.7520.1972.3970.322
", + "bbox": [ + 218, + 143, + 769, + 191 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/7625f6936d77b80ea53cf1b5296c30edc72cd1744ed36e8e37df8d304aa9f5ce.jpg", + "table_caption": [ + "(a) Super-resolution", + "(b) Mixed degradation", + "Table 1: Quantitative comparisons on super-resolution and mixed degradation tasks. The best result without using reference is highlighted. Our method achieves great superiorities on perceptual metrics and image quality when compared with others." + ], + "table_footnote": [], + "table_body": "
MethodLLFF
PSNR†SSIM†LPIPS‡LIPE†MANIQA†
Deblur-NeRF23.750.7990.3071.3650.157
BAD-NeRF24.0270.7880.3131.0940.151
NeRF-Perframe21.020.6950.3620.3620.150
Ours23.230.8110.2941.1440.177
", + "bbox": [ + 217, + 285, + 401, + 335 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/f2f4f5f8ddd082ab92bf220695e8dcec77dcfbe55213a6cc1c2643d4fab81459.jpg", + "table_caption": [ + "(a) Deblurring" + ], + "table_footnote": [], + "table_body": "
MethodLLFF
PSNR↑SSIM↑LPIPS↓LIPE↑MANIQA↑
Debhr-NeRF21.710.7490.3331.1040.122
BAD-NeRF25.400.8360.2781.1400.173
NeRF-Perframe21.570.7710.3101.1030.199
Ours22.930.7890.2521.1430.224
", + "bbox": [ + 403, + 286, + 586, + 335 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/245218e2c283968f4b4640daf71d44ac73b357f06d7eda544143e697eede65e9.jpg", + "table_caption": [ + "(b) Deblurring(consistent blur)" + ], + "table_footnote": [], + "table_body": "
MethodLLFF
PSNR↑SSIM↑LPIPS↓LIPE↑MANIQA↑
NAN25.990.8220.32081.30320.241
NeRF-SwinIR24.370.7780.3461.5790.210
NeRF-Perframe23.660.7860.2811.0950.214
Ours23.780.7910.25611.6070.257
", + "bbox": [ + 589, + 286, + 772, + 335 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "(c) Denoising with Gain 8", + "bbox": [ + 604, + 338, + 756, + 349 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 2: Quantitative comparisons for deblurring and denoising. The best result without using reference is highlighted. Our method performs the best on perceptual metrics.", + "bbox": [ + 215, + 359, + 784, + 388 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "3D causes the drop of PSNR and SSIM. By contrast, on the perceptual metrics like LPIPS and non-reference-based metrics including LIPE and MANIQA, which could more effectively reflect the restoration performance, our method consistently achieves better results when compared with other baselines.", + "bbox": [ + 212, + 431, + 784, + 491 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For mixed degradation tasks, the best results for LPIPS metrics are achieved by NeRFLix w. ref [53]. This is because NeRFLiX can see two high-quality ground-truth images from the nearest two viewpoints when inference, which leads to information leakage. However, high-quality ground-truth information is not accessed in our setting or in any real-world cases. After eliminating the impact of ground truth (NeRFLiX w/o. ref) by replacing the reference ground truth images with degraded images, our method performs better than NeRFLiX. Qualitative Results. To further verify the restoration capability, we present visual results for different degradations in Fig. 4. Our method is able to generate realistic details while other methods tend to generate smooth results which lack high-quality details. For example, for the super-resolution task, we show the drum restoration for each method, as can be seen in Fig. 4, our method generates high-fidelity drum surfaces with sharp edges, while other methods suffer from severe blur on the drum surfaces and edges. For mixed degradation tasks, our method successfully restores the intricate golden textures on the chair surface while other methods only rendered extremely blurred texture.", + "bbox": [ + 212, + 492, + 785, + 733 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Geometry Refinement. Compared with previous NeRF restoration methods, most of which focus on using 2D refinement to resolve frame-wise defects, our method firstly leverages the general priors of large foundation models and GAN, achieving the open-domain 3D-based restoration. Thus, our restoration algorithm could not only recover the high-quality and view-consistent images, but also refine the 3D geometry. We demonstrate this advantage using the Ficus data of Blender, as shown in Fig. 5a, through comparing the extracted mesh from the", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "RaFE: Generative Radiance Fields Restoration", + "bbox": [ + 416, + 114, + 730, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/125f23c222781059bc595eedaea13010d055178e7ebcac57c84819d2ae1a1405.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 233, + 148, + 287, + 223 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4fb27322685fb0d02ed831c9e62001a4fae8672e70366101e354d56f18f0ff05.jpg", + "image_caption": [ + "Perframe", + "(a)" + ], + "image_footnote": [], + "bbox": [ + 302, + 148, + 357, + 228 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ae7b5738179baad23c5d8a99885c6f85f119e5aa35df0cd0b3570b743304d515.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 359, + 150, + 413, + 228 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a64e405d9bcb6565f98ca273acb8570683d6b5615f5d1573d821fa7e94289239.jpg", + "image_caption": [ + "GT", + "Fig. 3: (a) Error map between ground truth and our method/NeRF-Perframe. (b) we showcase that the visual quality can be much better even with lower PSNR scores." + ], + "image_footnote": [], + "bbox": [ + 524, + 181, + 571, + 224 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d159837477fa8b4e28a84abad3508150a2487fc4e1db10d9a2fbfa28d147f1a6.jpg", + "image_caption": [ + "Ours", + "(b)" + ], + "image_footnote": [], + "bbox": [ + 591, + 147, + 647, + 224 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8db7cda11d49d7b836014492629f658286453844e06275227b4437fedc0ea210.jpg", + "image_caption": [ + "NeRF-SR" + ], + "image_footnote": [], + "bbox": [ + 651, + 147, + 754, + 224 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "restored NeRF models, our method effectively learns the restoration in the 3D tri-plane space and recovers better geometry.", + "bbox": [ + 212, + 324, + 782, + 354 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 378, + 387, + 393 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effects of different restoration models. To investigate the influence of using different 2D restoration models, we tested two additional off-the-shelf restoration models for the super-resolution task, including diffusion-based DiffBIR [25] and non-diffusion-based SRFormer [54]. As shown in Fig. 5b, the diffusion-based model DiffBIR shows larger restoration diversity over SRFormer by measuring the diversity score. When the repaired images exhibit diversity, direct reconstruction inevitably leads to blurriness to varying degrees due to the existing multi-view inconsistency. Through modeling the distribution of the potential high-quality NeRFs, our method accommodates these inconsistencies and consistently achieves better performance over NeRF-Perframe, demonstrating the great generalization capability of RaFE to different 2D restoration models.", + "bbox": [ + 212, + 404, + 785, + 570 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effects of the generator. In this ablation study, we examine the influence of the generator by comparing it with the baseline that directly optimizes the NeRF parameters using GAN loss and LPIPS mentioned above on the Blender dataset. As we can observe in Fig. 6, the image rendered by generative NeRF exhibits varied fine-textured details under different random code $\\mathbf{z}$ . Once the generator is removed, the rendered images will contain a blurry and smooth appearance, showing the importance of using the generator to model the distribution, which can also be demonstrated by the numerical metrics in the right part of Fig. 6.", + "bbox": [ + 212, + 571, + 787, + 691 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 715, + 367, + 732 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "This paper proposes a novel generic NeRF restoration method that applies to various types of degradations, such as low resolution, blurriness, noise, and mixed degradation. The proposed method leverages the off-the-shelf image restoration methods to restore the multi-view input images individually. To tackle the geometric and appearance inconsistencies presented in multi-view images due to individual restoration, we propose to train a GAN for NeRF generation, where", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Wu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a0f86ebb671cbff4cf90754998e0ade07ffe51711e143e64f23ef30bcd37d54b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 159, + 318, + 231 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d8e1ded7c54cb6468732d67f90c87b48af9310694f04558c95be0524de092366.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 320, + 159, + 410, + 231 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c17ac47c480de50c8dfded832d4b7848515dce9b0170084527e7336e876ea0cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 159, + 503, + 232 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/13bf006035cd16f100b8c3b89289d50fe112a5703727cce4cc9f7d2a764b3c70.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 159, + 594, + 231 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/911a6932edaaf60e960934bd95848eadce2f014775a105e683d74f4f31ccc3e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 159, + 687, + 231 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/35ef8f558319326a2bfb1f2246ebefe2d16ecc02176799cd64c83aa6d31cbfe5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 159, + 779, + 231 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3cc06a44f37365d2e012ead3fdc9ebf408275baf97c017666a5bcf9af1d8292b.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 223, + 243, + 316, + 297 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/28acb84c0328be37bc399141a94bd93f3c37c784e9cc537d1ada7eabd305954b.jpg", + "image_caption": [ + "(b) NeRF-SR" + ], + "image_footnote": [], + "bbox": [ + 318, + 243, + 410, + 297 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f7d479743418b09c805deeb021d3c03f90a4021a2583ba28da44482b81b6d29a.jpg", + "image_caption": [ + "(c) NVSR" + ], + "image_footnote": [], + "bbox": [ + 411, + 243, + 501, + 297 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/98e21da0fcb3a5f87de597e418ea26c35ecfc93c6f045d52bcf2df70625498ff.jpg", + "image_caption": [ + "(d) NeRF-SwinIR" + ], + "image_footnote": [], + "bbox": [ + 503, + 243, + 594, + 297 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8b2bb7880364077387a24a719eb58bccffc16b84e1ab7f25824f3f3845a012dd.jpg", + "image_caption": [ + "(e) NeRF-Perframe" + ], + "image_footnote": [], + "bbox": [ + 594, + 243, + 686, + 297 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/abb39be8fe5ecc748691484efff0dac8454ccbd0b0b548855a175198eaf7479f.jpg", + "image_caption": [ + "(f) Ours" + ], + "image_footnote": [], + "bbox": [ + 689, + 243, + 779, + 297 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1594ac270108d3833ebac6fb052496e01f03be459c5932d6c1d2b81b64011382.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 223, + 325, + 331, + 391 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7e4dfcc2a8e6244ad57aead5ed03ac60bf85a6f27ff5adc11f5a4a608c76c353.jpg", + "image_caption": [ + "(b) Deblur-NeRF" + ], + "image_footnote": [], + "bbox": [ + 334, + 325, + 442, + 391 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1e5017fc02880138eb92352ff014f0f0a9b8e64be4af6ad42dc060bd244ac411.jpg", + "image_caption": [ + "(c) BAD-NeRF" + ], + "image_footnote": [], + "bbox": [ + 444, + 325, + 553, + 391 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7bd334d5a22bb26792f48ec7dc3fce40df7cb4d8738b8cf98d7b2ea9bb74ee84.jpg", + "image_caption": [ + "Super-resolution", + "(b)NeRF-Perframe" + ], + "image_footnote": [], + "bbox": [ + 553, + 325, + 663, + 391 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9c0e71139d2b51146cc179012e9f061ed74fc7288eb411b3db401d21a4e075f0.jpg", + "image_caption": [ + "(e) ours" + ], + "image_footnote": [], + "bbox": [ + 665, + 325, + 774, + 391 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3f1d6a3eb2977f569a43f14b0b359d2e2bb1221aa20155a4b8ce3a6ad379b397.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 225, + 419, + 333, + 483 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d017ae851f9be427de08ae8cbd18290c7549858df7b18f8785c2faf06a984efe.jpg", + "image_caption": [ + "(b) Deblur-NeRF" + ], + "image_footnote": [], + "bbox": [ + 334, + 419, + 442, + 483 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e88d6f5bfd51e822affb7c15a1456a49f80fb2dff9329fd9d6f1c25b1eb8b0f6.jpg", + "image_caption": [ + "Inconsistent Motion Deblur", + "(c) BAD-NeRF" + ], + "image_footnote": [], + "bbox": [ + 444, + 419, + 553, + 483 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f8a167f4bb04e0f54a25cbef71d6a2555fff3b68b4a88f5f136233a70cb4b398.jpg", + "image_caption": [ + "(b)NeRF-Perframe" + ], + "image_footnote": [], + "bbox": [ + 553, + 419, + 663, + 483 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e803680c6d14ef2f6955c239b74e836f16191fd3551d1caed6d366355cf34ead.jpg", + "image_caption": [ + "(e) ours" + ], + "image_footnote": [], + "bbox": [ + 666, + 419, + 776, + 483 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/6d4017c6f82e9a244871b873e5b5f55f81ced7b302bd7b3fa1809623bc829193.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 223, + 508, + 333, + 571 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/576d9f561108186776eed629b11dd3f7d7738d47d1f74bf063dd158067df36ff.jpg", + "image_caption": [ + "(b) NAN" + ], + "image_footnote": [], + "bbox": [ + 334, + 508, + 442, + 571 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b820cf0acfe28c5c28a0590109a68a998b76746e4180566cf5d30823ac74cf84.jpg", + "image_caption": [ + "Consistent Motion Deblur", + "(b) NeRF-SwinIR" + ], + "image_footnote": [], + "bbox": [ + 444, + 508, + 553, + 571 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d3e723c5b981640f8a68b34f66616ed1f9e23c7d53bdf2e57da994057eecc1b4.jpg", + "image_caption": [ + "(c) NeRF-Perframe" + ], + "image_footnote": [], + "bbox": [ + 553, + 508, + 661, + 571 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/5731ea98f68847a46743ed17894e068f7a14c66126025f60eec937a8b5199feb.jpg", + "image_caption": [ + "(d) ours" + ], + "image_footnote": [], + "bbox": [ + 665, + 508, + 774, + 571 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/72ee40d01719d85eaaf89f3d1605e7bf4269853ea204384584fcad68a9c5a810.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 594, + 316, + 669 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8880675a671b5a78cfa68d1a31b542c496e0f0965c7351444692a041ed0c9e76.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 595, + 408, + 669 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/976fc9f1aa14575ecb4aff242d6f6d79d5d65dc1df58e2f114656a26736dacd2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 595, + 500, + 669 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/4064ad80d76071bb260ebe49be1520946845fab17a0119d79677bb6733b0586f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 595, + 589, + 669 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/68dee7be4e7bfda9e21d06c4cbf40717ecef00e944b2ffce5fea1ab6c470e10b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 595, + 679, + 669 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9ac75037c80568de0fff6dc282dc547ab99f3bbcf4d79005038e14984d9fe1ac.jpg", + "image_caption": [ + "(f) Ours" + ], + "image_footnote": [], + "bbox": [ + 679, + 595, + 772, + 734 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/688c9e2f34235af8266913dae2f012b88a637f642ff5b760ebe95ad520cb24b6.jpg", + "image_caption": [ + "(a) Input", + "Fig. 4: Visual results for super-resolution and mixed-degradation tasks. We show that our method is capable of generating detailed geometry and texture while other methods tend to be smooth in both geometry and texture. We recommend zooming in for better visualization." + ], + "image_footnote": [], + "bbox": [ + 223, + 678, + 316, + 734 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ba662c96531ef2f998329ee09d48e8c516c4ba42c4d159bfee8dca2623aff237.jpg", + "image_caption": [ + "(b) NeRF-SwinIR" + ], + "image_footnote": [], + "bbox": [ + 316, + 680, + 408, + 734 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/77d42aa4df6c959a9feac61b2a638254271513d592115dae84e0c4f4a048f28a.jpg", + "image_caption": [ + "(c) NeRFLiX w.ref", + "Mixed Degradation" + ], + "image_footnote": [], + "bbox": [ + 408, + 680, + 500, + 734 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1b461d05594929ab4749a6a2293f58f629e3c0d2414ccc6460d9eb0da61516c7.jpg", + "image_caption": [ + "(d) NeRFLiX w/o.ref" + ], + "image_footnote": [], + "bbox": [ + 500, + 680, + 589, + 734 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d76df3b73bbc7363bd0e5c3b02cb019a29cea3d25905ca4bf7d7820e34e562e5.jpg", + "image_caption": [ + "(e) NeRF-Perframe" + ], + "image_footnote": [], + "bbox": [ + 589, + 680, + 679, + 734 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "", + "bbox": [ + 679, + 680, + 772, + 734 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "RaFE: Generative Radiance Fields Restoration", + "bbox": [ + 416, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f1e4d071783c70441558ce6999370b1d1069e7d4d9e515f853a2b861a54fc25f.jpg", + "image_caption": [ + "Fig. 5: (a) Geometry comparisons between NeRF-SWINIR/NeRFLIX and our method. (b) Comparisons between using different 2D restoration models." + ], + "image_footnote": [], + "bbox": [ + 225, + 141, + 459, + 246 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/59d8628f343e2e2e2ac333b3d29fa6606dc8b9bb600d020c0e923886b9809268.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 147, + 784, + 247 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/84a5b24e4bfdbf7aa0e1424316ee20b5007be85010842144e65634f8c83bcf6d.jpg", + "image_caption": [ + "Fig. 6: Effectiveness of the tri-plane generator.Left: image rendered by NeRF without generator and images rendered by generative NeRF under different random codes $z(z_{0},z_{1},z_{2})$ . Right: numerical metrics to evaluate the efficacy of the generator. The results show the effectiveness of using the generator to model the distribution." + ], + "image_footnote": [], + "bbox": [ + 233, + 316, + 599, + 391 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/b137d44b8bd60e3194d18c56dc810d2f17b67dcd67e3669abec6d539e74d0d13.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
GeneratorX✓
LPIPS↓0.0710.062
LIPE↑3.3704.621
MANIQA↑0.3830.543
", + "bbox": [ + 656, + 323, + 785, + 383 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "a two-level tri-plane structure is adopted. The coarse-level tri-plane pre-trained by low-quality images remains fixed, while the fine-level residual tri-plane to be added to the coarse level is modeled by a GAN-based generator to capture variations in restoration. Extensive experiments on various restoration tasks with both synthetic and real cases demonstrate the superior performance of our method. Limitations and Future Work: One limitation of our method is instability when performing the restoration at extremely high resolutions, such as 4k, due to the patch-rendering strategy. Moreover, due to the inherent slow efficiency of NeRF rendering, currently the long training time also needs to be optimized. To overcome these limitations, a potential solution would be to integrate more efficient rendering techniques like Gaussian splatting [18] into RaFE, enabling the rendering of entire images. We plan to resolve these issues in the future work.", + "bbox": [ + 212, + 503, + 787, + 686 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 215, + 712, + 392, + 729 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work was supported in part by the Hong Kong Research Grants Council General Research Fund (17203023), in part by The Hong Kong Jockey Club Charities Trust under Grant 2022-0174, in part by the Startup Fund and the Seed Fund for Basic Research for New Staff from The University of Hong Kong, in part by the funding from UBTECH Robotics, in part by a GRF grant from the Research Grants Council (RGC) of the Hong Kong Special Administrative", + "bbox": [ + 212, + 748, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Wu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Region, China [Project No. CityU 11208123], and in part by the National Natural Science Foundation of China (62132001)", + "bbox": [ + 215, + 146, + 784, + 176 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 202, + 321, + 218 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bahat, Y., Michaeli, T.: Explorable super resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2716-2725 (2020)", + "2. Bahat, Y., Zhang, Y., Sommerhoff, H., Kolb, A., Heide, F.: Neural volume superresolution. arXiv preprint arXiv:2212.04666 (2022)", + "3. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields (2021)", + "4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR (2022)", + "5. Beyer, L., Zhai, X., Kolesnikov, A.: Big vision. https://github.com/google-research/big_vision (2022)", + "6. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022)", + "7. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: European Conference on Computer Vision (ECCV) (2022)", + "8. Chen, W.T., Yifan, W., Kuo, S.Y., Wetzstein, G.: Dehazenerf: Multiple image haze removal and 3d shape reconstruction using neural radiance fields. arXiv preprint arXiv:2303.11364 (2023)", + "9. Chen, X., Deng, Y., Wang, B.: Mimic3d: Thriving 3d-aware gans via 3d-to-2d imitation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)", + "0. Chen, Z., Zhang, Y., Ding, L., Bin, X., Gu, J., Kong, L., Yuan, X.: Hierarchical integration diffusion model for realistic image deblurring. In: NeurIPS (2023)", + "1. Chen, Z., Zhang, Y., Gu, J., Kong, L., Yang, X., Yu, F.: Dual aggregation transformer for image super-resolution. In: ICCV (2023)", + "2. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit radiance fields in space, time, and appearance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12479-12488 (2023)", + "3. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014)", + "4. Han, Y., Yu, T., Yu, X., Wang, Y., Dai, Q.: Super-nerf: View-consistent detail generation for nef super-resolution. arXiv preprint arXiv:2304.13518 (2023)", + "5. Karras, T., Aittala, M., Laine, S., Härkönen, E., Hellsten, J., Lehtinen, J., Aila, T.: Alias-free generative adversarial networks. In: Proc. NeurIPS (2021)", + "6. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of StyleGAN. In: Proc. CVPR (2020)", + "7. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. In: Advances in Neural Information Processing Systems (2022)" + ], + "bbox": [ + 225, + 236, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "RaFE: Generative Radiance Fields Restoration", + "bbox": [ + 416, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/", + "19. Lee, D., Lee, M., Shin, C., Lee, S.: Dp-nerf: Deblurred neural radiance field with physical scene priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12386-12396 (2023)", + "20. Lee, D., Oh, J., Rim, J., Cho, S., Lee, K.M.: Exblurf: Efficient radiance fields for extreme motion blurred images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17639-17648 (2023)", + "21. Li, H., Zhang, Z., Jiang, T., Luo, P., Feng, H., Xu, Z.: Real-world deep local motion deblurring. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1314-1322 (2023)", + "22. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: ICML (2022)", + "23. Li, J., Zhang, Z., Liu, X., Feng, C., Wang, X., Lei, L., Zuo, W.: Spatially adaptive self-supervised learning for real-world image denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9914-9924 (June 2023)", + "24. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. arXiv preprint arXiv:2108.10257 (2021)", + "25. Lin, X., He, J., Chen, Z., Lyu, Z., Fei, B., Dai, B., Ouyang, W., Qiao, Y., Dong, C.: Diffbir: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070 (2023)", + "26. Liu, X., Xue, H., Luo, K., Tan, P., Yi, L.: Genn2n: Generative nerf2nerf translation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5105-5114 (2024)", + "27. Ma, L., Li, X., Liao, J., Zhang, Q., Wang, X., Wang, J., Sander, P.V.: Deblur-nerf: Neural radiance fields from blurry images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12861-12870 (2022)", + "28. Mildenhall, B., Barron, J.T., Chen, J., Sharlet, D., Ng, R., Carroll, R.: Burst denoising with kernel prediction networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2502-2510 (2018)", + "29. Mildenhall, B., Hedman, P., Martin-Brualla, R., Srinivasan, P.P., Barron, J.T.: NeRF in the dark: High dynamic range view synthesis from noisy raw images. CVPR (2022)", + "30. Mildenhall, B., Srinivasan, P.P., Ortiz-Cayon, R., Kalantari, N.K., Ramamoorthi, R., Ng, R., Kar, A.: Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG) (2019)", + "31. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)", + "32. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (Jul 2022). https://doi.org/10.1145/3528223.3530127, https://doi.org/10.1145/3528223.3530127", + "33. Pearl, N., Treibitz, T., Korman, S.: Nan: Noise-aware nerfs for burst-denoising. In: CVPR (2022)", + "34. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Wu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)", + "36. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)", + "37. Skorokhodov, I., Tulyakov, S., Wang, Y., Wonka, P.: Epigraph: Rethinking training of 3d gans. Advances in Neural Information Processing Systems 35, 24487-24501 (2022)", + "38. Sun, C., Sun, M., Chen, H.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: CVPR (2022)", + "39. Tian, K., Jiang, Y., Yuan, Z., Bingyue, P., Wang, L.: Visual autoregressive modeling: Scalable image generation via next-scale prediction. arXiv preprint arXiv:2404.02905 (2024)", + "40. Wan, Z., Paschalidou, D., Huang, I., Liu, H., Shen, B., Xiang, X., Liao, J., Guibas, L.: Cad: Photorealistic 3d generation via adversarial distillation. arXiv preprint arXiv:2312.06663 (2023)", + "41. Wan, Z., Richardt, C., Božić, A., Li, C., Rengarajan, V., Nam, S., Xiang, X., Li, T., Zhu, B., Ranjan, R., Liao, J.: Learning neural duplex radiance fields for real-time view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8307-8316 (June 2023)", + "42. Wan, Z., Zhang, B., Chen, D., Zhang, P., Chen, D., Liao, J., Wen, F.: Bringing old photos back to life. In: proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2747-2757 (2020)", + "43. Wang, C., Wu, X., Guo, Y.C., Zhang, S.H., Tai, Y.W., Hu, S.M.: Nerf-sr: High quality neural radiance fields using supersampling. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 6445-6454 (2022)", + "44. Wang, P., Zhao, L., Ma, R., Liu, P.: BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4170-4179 (June 2023)", + "45. Wang, P., Zhao, L., Ma, R., Liu, P.: Bad-nerf: Bundle adjusted deblur neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4170-4179 (2023)", + "46. Wang, Y., Yu, J., Zhang, J.: Zero-shot image restoration using denoising diffusion null-space model. The Eleventh International Conference on Learning Representations (2023)", + "47. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)", + "48. Yang, S., Wu, T., Shi, S., Lao, S., Gong, Y., Cao, M., Wang, J., Yang, Y.: Maniaq: Multi-dimension attention network for no-reference image quality assessment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1191-1200 (2022)", + "49. Zhang, K., Liang, J., Van Gool, L., Timofte, R.: Designing a practical degradation model for deep blind image super-resolution. In: IEEE International Conference on Computer Vision. pp. 4791-4800 (2021)", + "50. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)", + "51. Zhang, W., Zhai, G., Wei, Y., Yang, X., Ma, K.: Blind image quality assessment via vision-language correspondence: A multitask learning perspective. In: IEEE Conference on Computer Vision and Pattern Recognition. pp. 14071-14081 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "RaFE: Generative Radiance Fields Restoration", + "bbox": [ + 416, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "52. Zhang, W., Li, X., Chen, X., Qiao, Y., Wu, X.M., Dong, C.: Seal: A framework for systematic evaluation of real-world super-resolution. arXiv preprint arXiv:2309.03020 (2023)", + "53. Zhou, K., Li, W., Wang, Y., Hu, T., Jiang, N., Han, X., Lu, J.: Nerflix: High-quality neural view synthesis by learning a degradation-driven inter-viewpoint mixer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12363-12374 (2023)", + "54. Zhou, Y., Li, Z., Guo, C.L., Bai, S., Cheng, M.M., Hou, Q.: Srformer: Permuted self-attention for single image super-resolution. arXiv preprint arXiv:2303.09735 (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 285 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Wu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/RaFE_ Generative Radiance Fields Restoration/8e187213-9e85-483c-8762-09f958b4b9d0_model.json b/2024/RaFE_ Generative Radiance Fields Restoration/8e187213-9e85-483c-8762-09f958b4b9d0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c2703e32051909d1ddd5ee260b6995dc34789c45 --- /dev/null +++ b/2024/RaFE_ Generative Radiance Fields Restoration/8e187213-9e85-483c-8762-09f958b4b9d0_model.json @@ -0,0 +1,3426 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.233, + 0.142, + 0.773, + 0.162 + ], + "angle": 0, + "content": "RaFE: Generative Radiance Fields Restoration" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.19, + 0.751, + 0.208 + ], + "angle": 0, + "content": "Zhongkai Wu\\(^{1}\\), Ziyu Wan\\(^{2}\\), Jing Zhang\\(^{1\\star}\\), Jing Liao\\(^{2}\\), and Dong Xu\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.34, + 0.217, + 0.663, + 0.232 + ], + "angle": 0, + "content": "\\(^{1}\\) School of Software, Beihang University, China" + }, + { + "type": "text", + "bbox": [ + 0.37, + 0.232, + 0.633, + 0.246 + ], + "angle": 0, + "content": "2 City University of Hong Kong, China" + }, + { + "type": "text", + "bbox": [ + 0.37, + 0.246, + 0.632, + 0.26 + ], + "angle": 0, + "content": "3 The University of Hong Kong, China" + }, + { + "type": "list", + "bbox": [ + 0.34, + 0.217, + 0.663, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.416, + 0.261, + 0.588, + 0.274 + ], + "angle": 0, + "content": "ZhongkaiWu@buaa.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.424, + 0.275, + 0.581, + 0.288 + ], + "angle": 0, + "content": "zkaiwu.github.io/RaFE" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.33, + 0.744, + 0.65 + ], + "angle": 0, + "content": "Abstract. NeRF (Neural Radiance Fields) has demonstrated tremendous potential in novel view synthesis and 3D reconstruction, but its performance is sensitive to input image quality, which struggles to achieve high-fidelity rendering when provided with low-quality sparse input viewpoints. Previous methods for NeRF restoration are tailored for specific degradation type, ignoring the generality of restoration. To overcome this limitation, we propose a generic radiance fields restoration pipeline, named RaFE, which applies to various types of degradations, such as low resolution, blurriness, noise, compression artifacts, or their combinations. Our approach leverages the success of off-the-shelf 2D restoration methods to recover the multi-view images individually. Instead of reconstructing a blurred NeRF by averaging inconsistencies, we introduce a novel approach using Generative Adversarial Networks (GANs) for NeRF generation to better accommodate the geometric and appearance inconsistencies present in the multi-view images. Specifically, we adopt a two-level tri-plane architecture, where the coarse level remains fixed to represent the low-quality NeRF, and a fine-level residual tri-plane to be added to the coarse level is modeled as a distribution with GAN to capture potential variations in restoration. We validate RaFE on both synthetic and real cases for various restoration tasks, demonstrating superior performance in both quantitative and qualitative evaluations, surpassing other 3D restoration methods specific to single task. Please see our project website zkaiwu.github.io/RaFE." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.662, + 0.741, + 0.69 + ], + "angle": 0, + "content": "Keywords: Neural Rendering \\(\\cdot\\) Generative Model \\(\\cdot\\) 3D Restoration \\(\\cdot\\) Neural Radiance Fields" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.719, + 0.377, + 0.736 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.754, + 0.788, + 0.816 + ], + "angle": 0, + "content": "Recently, Neural Radiance Fields (NeRFs) [3,4,7,12,31,32,38,41] have achieved great success in novel view synthesis and 3D reconstruction. However, most NeRF methods are designed based on well-captured images from multiple views with calibrated camera parameters. In real-world applications of NeRF, the data" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.387, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Wu et al." + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.15, + 0.296, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.206, + 0.296, + 0.223 + ], + "angle": 0, + "content": "Low Resolution Image Set" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.149, + 0.455, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.363, + 0.212, + 0.391, + 0.221 + ], + "angle": 0, + "content": "NeRF" + }, + { + "type": "image", + "bbox": [ + 0.462, + 0.155, + 0.616, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.212, + 0.558, + 0.221 + ], + "angle": 0, + "content": "Ours (x4)" + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.156, + 0.774, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.656, + 0.212, + 0.698, + 0.221 + ], + "angle": 0, + "content": "Ours (x8)" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.228, + 0.311, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.222, + 0.286, + 0.309, + 0.304 + ], + "angle": 0, + "content": "Mixed Degradation Image Set" + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.23, + 0.4, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.344, + 0.29, + 0.371, + 0.299 + ], + "angle": 0, + "content": "NeRF" + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.23, + 0.486, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.43, + 0.29, + 0.454, + 0.299 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.23, + 0.588, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.29, + 0.582, + 0.3 + ], + "angle": 0, + "content": "Real Blur Image Set" + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.23, + 0.72, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.625, + 0.29, + 0.652, + 0.299 + ], + "angle": 0, + "content": "NeRF" + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.23, + 0.784, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.29, + 0.747, + 0.298 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.316, + 0.788, + 0.372 + ], + "angle": 0, + "content": "Fig. 1: High-quality restoration of radiance fields from various types of degradations. Given only degraded images, our method can restore high-quality NeRF. It is a generic approach that can be applied to various types of degradations, resulting in refinement of both geometry and appearance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.785, + 0.538 + ], + "angle": 0, + "content": "capture or transmission process often introduces various forms of image degradations, such as noise generated during photography in low-light conditions [29,33] and blur caused by camera motion [27,44], or JPEG compression and down-sampling during transmission [2,43]. Simply restoring degraded images frame-by-frame can result in inconsistencies of geometry and appearance across different viewpoints. Directly reconstructing 3D models over these per-frame restoration results can easily induce inferior quality since current NeRF methods heavily rely on pixel-wise independent ray optimization with local computations, which are highly susceptible to noise and degradation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.786, + 0.719 + ], + "angle": 0, + "content": "Several NeRF variants have attempted to reconstruct 3D scenes with degraded multi-view images by introducing specific strategies or additional constraints for the optimization of radiance fields. For example, [19,20,27] deal with image blur artifacts by modeling the degradation kernel with NeRF, while [2,43] super-sampling on rays or tri-planes to obtain high-resolution 3D from low-resolution observations. Additionally, [31] modifies NeRF to reconstruct the scene in linear HDR space by supervising directly on noisy raw images to address the noise generated in low-light conditions. [53] tries to improve the view synthesis quality by removing NeRF-specific rendering artifacts. Even with great success, all these approaches are only designed to handle specific types of degradation. To the best of our knowledge, currently there is no generic pipeline which supports the restoration of radiance fields under various types of degradation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this paper, we propose RaFE, a generic NeRF restoration framework that enables high-quality radiance field reconstruction from captured images containing various types of degradations in a generative manner. Firstly, we leverage the success of off-the-shelf image restoration methods to restore multi-view images with different forms of degradations, such as super-resolution, deblurring, denoising, removing compression artifacts, or a combination thereof. It is important to note that since the images from different views are independently restored, there inevitably exist geometric and appearance inconsistencies between them." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "RaFE: Generative Radiance Fields Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "Naively optimizing a radiance field with these refined images would average out the inconsistencies and result in blurry outputs. To overcome this challenge, our insight here is, instead of reconstructing a single 3D using inconsistent frames, we could consider these restored multi-view images as the renderings from multiple distinct high-quality NeRF models with varied geometry and appearance. In this case, we abandon the commonly-used pixel-wise reconstruction objective and propose to leverage the generative adversarial networks (GANs) to model the distribution of these different high-quality NeRF models, which could effectively capture the inherent variability in ill-posed inverse problem, allowing for a better accommodation of the inconsistencies present in different views." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.788, + 0.721 + ], + "angle": 0, + "content": "Specifically, our pipeline consists of two main stages. In the first stage, based on the degradation type, we employ the corresponding off-the-shelf image restoration methods [17, 25, 34, 35, 46, 49] to obtain a set of high-quality multiview images. In practice, we prefer choosing restoration methods which have strong capabilities to recover high-quality and realistic texture details. In the second stage, we train a 3D generative model based on these restored multi-view images. Drawing inspirations from recent 3D generation works [6, 9, 37, 40], we construct a convolutional neural network (CNN) to generate tri-plane features, which are subsequently sampled and decoded into density and colors using MLP networks for NeRF rendering. Here, instead of generating single-level tri-plane features as previous works did, we decompose the tri-planes into two levels. The coarse-level tri-planes are constructed directly from low-quality images and remain fixed during training, representing the coarse structure of the modeled 3D distribution. Simultaneously, we train a generator to output the diverse fine-level tri-plane features, which act as residuals to be added to the coarse-level features for NeRF rendering. By focusing on learning the residual representations instead of the entire tri-planes for NeRF, we simplify the modeling and learning of restoration variations since we only need to learn the details while the coarse structure is provided by coarse-level tri-planes, which makes great improvement in rendering quality for more complex regions. To train the generator, we adopt an adversarial loss defined on rendered 2D images to encourage them to be indistinguishable from the high-quality restored images. We also incorporate a perceptual loss between the rendered images and the restored images to calculate structure constraints. Additionally, we propose patch sampling strategies to stabilize the generator training procedure. Once the generator has been trained, we can generate restored radiance fields with high quality renderings and a certain level of diversity by sampling different code in the latent space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "We conducted extensive experiments to validate the effectiveness of our method, both qualitatively and quantitatively. The experimental results showcase the superiority of our approach in various restoration tasks, such as super-resolution (upper row of Figure 1), camera motion blur (a real-world case at the right part of the lower row of Figure 1) and the restoration of mixed degradation consisting of noise, blur, and compression (left part of the lower row of Figure 1). Our method not only generates images with richer and enhanced tex" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Wuet al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "ture details but also achieves significant improvements in geometric refinement, as demonstrated by the mesh visualization in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.178, + 0.582, + 0.192 + ], + "angle": 0, + "content": "To summarize, our contributions are three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.204, + 0.784, + 0.233 + ], + "angle": 0, + "content": "- We propose a generic radiance fields restoration pipeline that is applicable to various types of degradations." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.234, + 0.784, + 0.292 + ], + "angle": 0, + "content": "- We introduce a generative method for NeRF restoration that enables better accommodation of geometric and appearance inconsistencies present in the multi-view images, thus allowing us to incorporate the success of image restoration into 3D restoration." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.294, + 0.784, + 0.323 + ], + "angle": 0, + "content": "- We show the restoration method performs well on various degradation scenarios with both enhanced appearance and refined geometry." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.204, + 0.784, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.346, + 0.396, + 0.362 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.377, + 0.436, + 0.392 + ], + "angle": 0, + "content": "2.1 2D image restoration" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.402, + 0.788, + 0.598 + ], + "angle": 0, + "content": "Image restoration is a long-standing problem in low-level vision domain and significant progress has been achieved in different specific tasks including image super-resolution, deblur, denoise and blind restoration. Previously reconstruction-based methods [11,21,23,52,54] show their success in these tasks. However, those reconstruction-based methods are struggling to generate abundant high-quality details. Subsequently, generative restoration methods [10,17,25,34,35,39,42,46, 49], particularly those based on diffusion model, have shown the great capability to render high-quality details. Deepfloyd [35] proposes a super-resolution model, which concatenates the low-resolution input with random noise at pixel level as the condition image. DiffBIR [25] designs a degraded pipeline to simulate real-world degradation and utilizes the pre-trained diffusion model to generate photorealistic images. For camera motion blur, HiDiff [10] recovers images by using diffusion to generate features with abundant detailed information." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.62, + 0.495, + 0.634 + ], + "angle": 0, + "content": "2.2 Radiance Fields Restoration" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "NeRF restoration aims to reconstruct high-quality NeRF given only degraded images with various artifacts such as blur, noise, or low resolution. Up to now, several works [2,8,14,19,20,26,27,31,33,43,44,53] have explored this task under specific types of degradation. [20,27,44] deal with blurred input images by designing blur kernel or optimizing camera paths for NeRF rendering process. For NeRF super-resolution, [43] increased the ray sampling density, forcing multiple rays to render pixels equal to the same pixel, and applied a 2D refinement model to get final output images. [14] introduces a CEM [1] refinement model to adjust the output of the super-resolution model for better multi-view consistency. However, the CEM refinement model ruins the structure details of the images and inconsistencies still exist, leading to a smooth reconstruction result. [31,33] mainly focus on the noise degradation of the input image. [31] modifies NeRF to reconstruct the scene in linear HDR color space to address the noise generated" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "RaFE: Generative Radiance Fields Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.36 + ], + "angle": 0, + "content": "during image post-processing, while [33] achieve NeRF restoration on noisy images by using a noise-awarded encoder to aggregate features across views. [53] considers solving the degradation that occurs in NeRF reconstruction by training a 2D refinement model using simulated degraded images for typical NeRF-style artifacts, but they can not achieve refinement on geometry since the restoration only happens on rendered views. None of the existing approaches can restore NeRF in 3D space directly with flexible forms of degradation. NVSR [2] archives 3D geometry refinement by upsampling tri-plane representation, but their training processing requires tremendous amounts of 3D data, which are extremely hard to obtain in practice. [26] achieves NeRF translation by designing a VAE-like NeRF to model the distribution of the underlying 3D edits. By contrast, our method has the ability to handle more flexible forms of degradation and restore 3D geometry and appearance with the only need of an image set for an object or scene, making the 3D restoration more practical in real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.386, + 0.331, + 0.401 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.788, + 0.513 + ], + "angle": 0, + "content": "In this section, we elaborate the details of the proposed RaFE method. We introduce how to refine the degraded views using pretrained 2D restoration model, to capture the high-quality appearance distribution in Sec. 3.1. Then, we describe our generative restoration framework including the neural representation, generator architecture and optimization in Sec. 3.2. The training strategy is introduced in Sec. 3.3. The overall pipeline could be found in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.538, + 0.524, + 0.554 + ], + "angle": 0, + "content": "3.1 High-quality Image Restoration" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.788, + 0.673 + ], + "angle": 0, + "content": "The recent success of the 2D image restoration task is dominated by the denoising diffusion probabilistic model, benefiting from its powerful appearance generation capability and extensive prior knowledge. Hence in this paper, we mainly borrow the recent diffusion-based restoration methods [25,35] to obtain the high-quality multi-view images from their low-quality counterparts. It should be noted that our 3D restoration framework could also work well based on non-diffusion restoration approaches, which has been verified by our experiments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.779 + ], + "angle": 0, + "content": "To assist in the restoration of image details, we employ powerful image caption models [5,22] to get an accurate textual prompt for the scene, denoted as \\( P \\). Given a set of multi-view low-quality images \\( I_{l} \\), the textual prompt is produced by selecting a view that contains as much information of the scene as possible, e.g. a side view of a synthetic Lego model. The prompt \\( P \\) and low-quality images \\( I_{l} \\) are then fed into the restoration model [25,35] to achieve high-quality per-frame refinement. We denote the restored high-quality images as \\( I_{h} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The diffusion model-based image restoration methods can effectively reconstruct high-quality images from various real-world degradation types. Moreover, thanks to the powerful generalization ability of diffusion models preserved in these restoration methods, we can handle the scene restoration in open domain" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Wu et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.146, + 0.784, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.305, + 0.788, + 0.348 + ], + "angle": 0, + "content": "Fig. 2: Overview of our pipeline. Given multi-view degraded images, we utilize the off-the-shelf methods to restore high-quality multi-view images. Then, we train our Generative Restoration NeRF to generate a high-quality scene." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.378, + 0.788, + 0.44 + ], + "angle": 0, + "content": "without the need of re-training. However, though the restored images have better quality, more variations among different views occur due to the generative nature of these restoration models, leading to serious multi-view inconsistencies. Next, we focus on dealing with this issue with the proposed generative pipeline." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.464, + 0.513, + 0.479 + ], + "angle": 0, + "content": "3.2 Generative NeRF Restoration" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.788, + 0.598 + ], + "angle": 0, + "content": "To deal with geometric and appearance inconsistencies across views, we treat these restored multi-view images distributed similarly to the rendered images from multiple slightly different high-quality NeRF models. Consequently, instead of directly fitting a NeRF model using the refined high-quality views which usually leads to blurry reconstruction, we are trying to learn the distribution of these diverse NeRF models by leveraging a generative method, allowing us to sample distinct restored 3D under the same degraded inputs." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.598, + 0.788, + 0.719 + ], + "angle": 0, + "content": "3D Representation. Following the recent 3D generative model [6,9,37], we adopt the hybrid explicit-implicit tri-plane representation for the feature field. This representation combines both explicit and implicit components to effectively model the density and RGB values. More specifically, to obtain the descriptor for any query location \\( \\pmb{x} \\in \\mathbb{R}^3 \\), we project the point onto three planes \\( (P_{xy}, P_{yz}, P_{zx}) \\) to retrieve the corresponding features \\( (f_{xy}, f_{yz}, f_{zx}) \\) using interpolation. Then we calculate the mean value of these three sampled features to obtain the final feature representation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Once we obtain the features for a point along the ray, we use two MLPs to decode the density and RGB values. The first MLP, denoted as \\(\\mathcal{M}_{\\text{dens}}\\), maps the features to the density value \\(\\sigma \\in \\mathbb{R}\\) and a color feature \\(\\pmb{f}_{color}\\). The second MLP, denoted as \\(\\mathcal{M}_{color}\\), takes the color feature and the view direction as inputs and maps them to the RGB value \\(\\pmb{c} \\in \\mathbb{R}^3\\). We empirically show that incorporating the view direction allows us to effectively model view-dependent effects, particularly when there are non-Lambertian surfaces in the scene. After obtaining the densities and RGB values for each sampled point along the emitted ray, we can" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.731, + 0.128 + ], + "angle": 0, + "content": "RaFE: Generative Radiance Fields Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.776, + 0.162 + ], + "angle": 0, + "content": "apply the classic volumetric rendering to obtain the color value for each pixel:" + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.17, + 0.785, + 0.207 + ], + "angle": 0, + "content": "\\[\nC (\\boldsymbol {r}) = \\sum_ {i = 1} ^ {N} T _ {i} (1 - \\exp (- \\sigma_ {i} \\delta_ {i})) \\boldsymbol {c} _ {i}, \\quad T _ {i} = \\exp (- \\sum_ {j = 1} ^ {i - 1} \\sigma_ {i} \\delta_ {i}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.213, + 0.784, + 0.242 + ], + "angle": 0, + "content": "where \\( \\pmb{c}_i, \\sigma_i \\) represents RGB and density value of the \\( i_{th} \\) point alone the ray, respectively. \\( N \\) is the sampled number and \\( \\delta_i \\) is the distance between samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.242, + 0.785, + 0.515 + ], + "angle": 0, + "content": "NeRF Generator. Given restored multi-view images, we regard them as renderings from several diverse high-quality NeRF models with varied geometry and appearance. The distribution formed by these distinct NeRF models, formulated as \\( p_r(\\phi | I_h) \\), can be modeled by a generator. With the hybrid explicit-implicit tri-plane representation, we introduce a StyleGAN2 [15]-like CNN-based generator, which receives a latent code \\( \\mathbf{w} \\) mapped from random code \\( \\mathbf{z} \\) to generate fine tri-planes \\( P_f \\) for high-quality NeRF rendering. To fully leverage the degraded images and make the generator only focus on generating the necessary refinements, we also pre-train a coarse NeRF model with tri-plane features denoted as \\( P_c \\) using the input low-quality images. Specifically, the coarse NeRF model consists of a coarse tri-plane \\( P_c \\) features and a decoder (where the decoder of coarse tri-planes is shared with the one of fine tri-planes). To obtain the final refined tri-planes \\( P \\), we combine the reconstructed tri-plane representations with the residual tri-planes generated by the generator via: \\( \\mathbf{f} = \\mathbf{f}_c + \\mathbf{f}_f \\). By merging these components, RaFE effectively captures both the global geometric guidance provided by the coarse NeRF and the local refinements learned from 3D generator, enabling us to obtain the restored tri-planes that exhibit enhanced geometric accuracy and appearance fidelity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.785, + 0.605 + ], + "angle": 0, + "content": "Optimization. To supervise the generator and NeRF parameters, we propose to minimize the distribution discrepancy between the rendered images and the restored high-quality images. We adopt a saturate GAN [13] loss with an image level discriminator. Specifically, we treat the high-quality images restored by 2D model as the real samples while the rendered images as the fake samples, and utilize adversarial loss with R1 regularization between the real and fake images:" + }, + { + "type": "equation", + "bbox": [ + 0.352, + 0.628, + 0.785, + 0.664 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {a d v} = \\mathbb {E} _ {\\boldsymbol {z} \\sim p _ {z}, \\boldsymbol {\\theta} \\sim p _ {\\theta}} [ f (D (G (\\boldsymbol {z}, \\boldsymbol {\\theta})) ] \\\\ + \\mathbb {E} _ {I _ {h} \\sim p _ {h}} [ f (- D (I _ {h})) + \\lambda \\| \\nabla D (I _ {h}) \\| ^ {2} ], \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.671, + 0.785, + 0.776 + ], + "angle": 0, + "content": "where \\( z, \\theta \\) represent random code and view point, respectively, and \\( I_h \\) is the restored high quality image. However, we observed relying solely on a GAN loss for training can lead to significant geometric mismatches between the restored images and rendered views. We argue that although the GAN loss helps align the distribution of 2D renderings, it still lacks geometry-level constraints. Therefore, we also incorporate a perceptual loss that encourages the rendered images to resemble the geometry of pre-frame restoration." + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.786, + 0.785, + 0.802 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {g e o m e t r y}} = L P I P S \\left(I _ {h}, G (\\boldsymbol {z}, \\boldsymbol {\\theta})\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "where \\(LPIPS(\\cdot, \\cdot)\\) refers to the learned perceptual image patch similarity proposed in [50]. \\(I_h\\) is the restored high quality image paired with view point \\(\\theta\\), and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.128 + ], + "angle": 0, + "content": "Wu et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.785, + 0.178 + ], + "angle": 0, + "content": "\\(G(\\pmb{z}, \\pmb{\\theta})\\) are the rendered image using the same training view \\(\\pmb{\\theta}\\) and a random sampled latent code \\(\\pmb{z}\\). We also supervise the coarse NeRF branch:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.189, + 0.785, + 0.206 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e c} = \\mathbb {E} _ {\\boldsymbol {\\theta} \\sim p _ {\\theta}} \\left[ \\| G _ {c} (\\boldsymbol {\\theta}) - I _ {l} ^ {\\theta} \\| ^ {2} \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.216, + 0.784, + 0.247 + ], + "angle": 0, + "content": "where \\( p_{\\theta} \\) indicate the view point distribution, and \\( I_l^\\theta \\) is the low quality image corresponding to view point \\( \\pmb{\\theta} \\). Overall, the complete training objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.257, + 0.785, + 0.273 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {\\text {g e o m e t r y}} \\mathcal {L} _ {\\text {g e o m e t r y}} + \\lambda_ {\\text {a d v}} \\mathcal {L} _ {\\text {a d v}} + \\lambda_ {\\text {r e c}} \\mathcal {L} _ {\\text {r e c}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.281, + 0.589, + 0.297 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\text{geometry}}, \\lambda_{\\text{adv}}, \\lambda_{\\text{rec}}\\) are trade-off parameters." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.316, + 0.513, + 0.332 + ], + "angle": 0, + "content": "3.3 Patch-based Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.785, + 0.519 + ], + "angle": 0, + "content": "During the training process, fully rendering the entire image, such as \\(256^2\\) or \\(512^2\\) pixels, and then feeding it into the discriminator can be computationally intensive and resource-consuming. This is because volume rendering requires the computation of density and color values for sampling points along the ray for each pixel, which can become prohibitively expensive for a whole image. Therefore, we only render a patch of the view at a time (i.e. \\(64^2\\)). The rendered patches and high-quality image patches are randomly selected, and the discriminator in Eq. 2 receives patches of rendered images and patches of high-quality images as fake images and real images respectively. For the perceptual loss in Eq. 3, the rendered patches and the high-quality patches have the same spacial coordinate on their original images to guarantee that the cropped patches have the same semantic and local structure." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.785, + 0.627 + ], + "angle": 0, + "content": "One limitation of using patches instead of entire images for training is that it is challenging to sample the local scenes at the boundary regions of images, particularly for forward-facing data. The imbalanced training distribution will result in mode collapse during the training process. To mitigate this issue, we employ a beta sampling strategy to determine the positions of the sampled patches. This strategy ensures that patches in the boundary regions of the image are adequately sampled. More specifically, the beta sampling can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.637, + 0.785, + 0.653 + ], + "angle": 0, + "content": "\\[\n\\delta_ {x}, \\delta_ {y} \\sim \\operatorname {B e t a} (\\beta (t), \\beta (t)), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.661, + 0.785, + 0.737 + ], + "angle": 0, + "content": "where \\(\\delta_x\\) and \\(\\delta_y\\) denote the position offset in \\(x\\) and \\(y\\) directions respectively, while \\(\\beta(t)\\) are linearly annealed from \\(\\beta(0) = 1\\) to some final value \\(\\beta(T)\\) smaller than 1. By using the beta sampling strategy, we can maintain a more balanced distribution of training data that focuses more on the boundary patches, alleviating mode collapse issue and improving the overall training stability." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.758, + 0.376, + 0.775 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.312, + 0.803 + ], + "angle": 0, + "content": "4.1 Setup" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.794, + 0.841 + ], + "angle": 0, + "content": "Datasets. We evaluate our model on the NeRF-Synthetic benchmark dataset [31], which contains 8 synthetic objects with images taken from different viewpoints" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "RaFE: Generative Radiance Fields Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.236 + ], + "angle": 0, + "content": "uniformly distributed in the hemisphere. Following the original setting, We hold 200 viewpoints for generating high-quality training data and 200 viewpoints for testing. Further, to demonstrate the generalization ability of our method, we also evaluate our method on complex real-world LLFF scenes [30] which consists of 8 scenes captured with roughly forward-facing images. We also demonstrate the superior performance of RaFE on real-world blur [27] and noise [29] data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.24, + 0.788, + 0.468 + ], + "angle": 0, + "content": "Evaluation Metrics. Following the common practice of 3D reconstruction, we try to evaluate each method with two standard image quality metrics: peak signal-to-noise ratio (PSNR), structural similarity index (SSIM) [47], which however could not reflect the real 3D restoration performance according to our observations. Due to the generative characteristic of RaFE, the recovered radiance field from RaFE is very high-quality, but may not faithfully follow the \"ground-truth\" 3D, since inversing a degraded signal is a highly ill-posed problem. Moreover, we also found the baselines with better scores over PSNR and SSIM still pose a degraded appearance with smooth texture details, as shown in Figure 3b. Hence, a better way is to leverage perceptual metric: learned perceptual image patch similarity (LPIPS) [50] which computes the mean squared error (MSE) between normalized features from all layers of a pre-trained VGG [36] encoder and is deemed to better correlate with human perception. We also leverage the latest non-reference-based image quality assessment metrics including LIQE [51] and MANIQA [48] to demonstrate the superior rendering quality of our method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.47, + 0.788, + 0.577 + ], + "angle": 0, + "content": "Implementation Details. We implement all the experiments by PyTorch. For the 2D generator and discriminator, we use a convolutional-based generator and discriminator used in StyleGAN2 [16]. In all experiments, we choose Adam optimizer for all the modules in our pipeline, with hyperparameters \\(\\beta_{1} = 0, \\beta_{2} = 0.99\\). We use learning rate \\(2 \\times 10^{-3}\\) for both generator and discriminator. For loss weights, we use \\(\\lambda_{mimic} = 0.5\\), \\(\\lambda_{adv} = 1.0\\), and \\(\\lambda_{rec} = 1.0\\) for almost all experiments. We evaluate RaFE framework on 4 different 3D restoration tasks:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.596, + 0.784, + 0.671 + ], + "angle": 0, + "content": "- \\(4 \\times\\) Super-Resolution: On the blender dataset, we resize the image to \\(64 \\times 64\\) resolutions to get low-resolution images. As for LLFF data, we first center crop the training image to \\(188 \\times 252\\) to adapt to our \\(4 \\times\\) super-resolution task and then resize to \\(47 \\times 63\\) to get low-resolution images. And we use Deepfloyd [35] for 2D super-resolution." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "- Deblur: On the LLFF dataset, we construct camera motion blur by applying the blur kernel following equation \\( I_{blur} = A * I_{clear} \\), where \\( A \\) is the blur kernel and \\( * \\) stands for the convolution operator. Different from the blur dataset proposed in Deblur-NeRF [27], which contains blurred images with varying degrees of blur, and even includes a certain amount of high-resolution images, we apply a large blur kernel size (e.g. 13) and a more complex camera motion path to all the images in the dataset. We additionally construct consistent blur datasets by using the same blur kernel to the training image in a scene, which means the camera motion trajectory is the same for the training image set. The resolution of images is the same as the super-resolution task. We use HiDiff [10] to recover high-quality images." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.596, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Wuet al." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.146, + 0.788, + 0.237 + ], + "angle": 0, + "content": "- Denoise: On the LLFF dataset, we follow the noise model used in [28, 33] and we get the noisy version of a clean image \\(I\\) according to equation \\(I_{\\text{noisy}}(x) = \\mathcal{N}(I(x), \\delta_r^2 + \\delta_s^2 I^2(x))\\), where \\(\\sigma_r\\) is the signal-independent read-noise parameter, \\(\\sigma_s\\) is the signal-dependent shot-noise. Following [28, 33], we use the Gain level to represent the noise strength. We use gain levels \\(= 8\\) to get our noisy image. We use DiffBIR [25] to get high-quality images." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.238, + 0.788, + 0.329 + ], + "angle": 0, + "content": "- Mixed degradation: The degradation pipeline consists of three stages: blur, noise, and JPEG compression. First, we utilize Gaussian blur with a radius of 7 for the blender dataset and 3 for the LLFF dataset. Second, we add noise with std 25. And then, we apply JPEG compression. The quality of JPEG compression is 50. The resolution of images is the same as the super-resolution task. We use DiffBIR [25] to get high-quality images." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.349, + 0.324, + 0.362 + ], + "angle": 0, + "content": "4.2 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.787, + 0.463 + ], + "angle": 0, + "content": "Baseline Methods. For general tasks, we try a baseline that firstly restores the degraded images and uses the restored high-quality images to reconstruct a NeRF directly, denoted as NeRF-Perframe. We also use the 2D-based restoration model SwinIR [24] to do the per-view refinement for the renderings of NeRF trained by degraded image, denoted as NeRF-SwinIR. Note that we do not evaluate NeRF-SwinIR for the deblur task since there are no corresponding checkpoints." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.787, + 0.658 + ], + "angle": 0, + "content": "To more thoroughly test the effectiveness of our method, we also select some task-specific competitors. For super-resolution task, we choose NeRF-SR [43], Neural Volume Super-Resolution (NVSR) [1] as baselines. For mixed degradation, since there is no existing method tailored for mixed degradation, we choose NeRFLiX, which tries to solve the NeRF-like degradation by training a 2D refinement model using degradation images constructed by a degradation simulator for typical NeRF-style artifacts. We consider this to be the most relevant method. For the deblur task, we compare with two state-of-the-art methods Deblur-NeRF [27] and BAD-NeRF [45]. Deblur-NeRF designs a learnable blur kernel and applies it to rays to simulate the degrading process and BAD-NeRF directly models the camera trajectories to solve motion blur. For the denoising task, we compare with NAN [33], which uses a noise-aware encoder to aggregate the feature of multi-view images for restoration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Quantitative Results. We conduct extensive quantitative comparisons with various baselines across different restoration tasks in Tab. 1a for super-resolution, Tab. 2b and Tab. 2a for deblurring, Tab. 2c for denoising and Tab. 1b for mixed degradation. As analyzed before, although most of the time our method falls slightly behind on the reconstruction metrics like PSNR and SSIM compared with other baselines, which only measure local pixel-aligned similarity between the rendered novel views and the ground truth images, are less indicative since uncertainties naturally exist in generative procedure. Taking the super-resolution results on Blender data as an example (Tab. 1a), the simplest baseline NeRF-Perframe has already achieved the best reconstruction metrics, but as shown in Figure. 3a, its visual quality is vastly inferior compared with our results. Through the error map, we found the misalignment between the generated 3D and input" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.731, + 0.128 + ], + "angle": 0, + "content": "RaFE: Generative Radiance Fields Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.219, + 0.144, + 0.77, + 0.193 + ], + "angle": 0, + "content": "
BleederLLFFLLFF
PSNR†SSIM†LPiPS‡LIPE†MANIQA†PSNR†SSIM†LPiPS‡LIPE†MANIQA†MethodPSNR†SSIM†LPiPS‡LIPE†MANIQA†PSNR†SSIM†LPiPS‡LIPE†MANIQA†
NevRF-SR26.850.9120.1352.3720.36622.650.7020.3271.3200.220
NVSR26.130.8790.1322.3010.34321.290.6060.4422.1080.177
NeRF-SwnIR24.070.8780.1193.060.38121.730.6550.3651.8080.267
NeRF-Perframe27.390.9220.0833.2760.38623.680.7450.2611.4170.257
Ours24.990.9010.0624.6210.54323.850.7520.1972.3970.322
" + }, + { + "type": "table_caption", + "bbox": [ + 0.296, + 0.195, + 0.414, + 0.207 + ], + "angle": 0, + "content": "(a) Super-resolution" + }, + { + "type": "table_caption", + "bbox": [ + 0.565, + 0.194, + 0.698, + 0.206 + ], + "angle": 0, + "content": "(b) Mixed degradation" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.218, + 0.785, + 0.261 + ], + "angle": 0, + "content": "Table 1: Quantitative comparisons on super-resolution and mixed degradation tasks. The best result without using reference is highlighted. Our method achieves great superiorities on perceptual metrics and image quality when compared with others." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.286, + 0.402, + 0.337 + ], + "angle": 0, + "content": "
MethodLLFF
PSNR†SSIM†LPIPS‡LIPE†MANIQA†
Deblur-NeRF23.750.7990.3071.3650.157
BAD-NeRF24.0270.7880.3131.0940.151
NeRF-Perframe21.020.6950.3620.3620.150
Ours23.230.8110.2941.1440.177
" + }, + { + "type": "table_caption", + "bbox": [ + 0.267, + 0.339, + 0.354, + 0.35 + ], + "angle": 0, + "content": "(a) Deblurring" + }, + { + "type": "table", + "bbox": [ + 0.405, + 0.287, + 0.588, + 0.337 + ], + "angle": 0, + "content": "
MethodLLFF
PSNR↑SSIM↑LPIPS↓LIPE↑MANIQA↑
Debhr-NeRF21.710.7490.3331.1040.122
BAD-NeRF25.400.8360.2781.1400.173
NeRF-Perframe21.570.7710.3101.1030.199
Ours22.930.7890.2521.1430.224
" + }, + { + "type": "table_caption", + "bbox": [ + 0.406, + 0.339, + 0.585, + 0.35 + ], + "angle": 0, + "content": "(b) Deblurring(consistent blur)" + }, + { + "type": "table", + "bbox": [ + 0.591, + 0.287, + 0.774, + 0.337 + ], + "angle": 0, + "content": "
MethodLLFF
PSNR↑SSIM↑LPIPS↓LIPE↑MANIQA↑
NAN25.990.8220.32081.30320.241
NeRF-SwinIR24.370.7780.3461.5790.210
NeRF-Perframe23.660.7860.2811.0950.214
Ours23.780.7910.25611.6070.257
" + }, + { + "type": "table_caption", + "bbox": [ + 0.606, + 0.339, + 0.758, + 0.35 + ], + "angle": 0, + "content": "(c) Denoising with Gain 8" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.361, + 0.785, + 0.39 + ], + "angle": 0, + "content": "Table 2: Quantitative comparisons for deblurring and denoising. The best result without using reference is highlighted. Our method performs the best on perceptual metrics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.785, + 0.492 + ], + "angle": 0, + "content": "3D causes the drop of PSNR and SSIM. By contrast, on the perceptual metrics like LPIPS and non-reference-based metrics including LIPE and MANIQA, which could more effectively reflect the restoration performance, our method consistently achieves better results when compared with other baselines." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.787, + 0.734 + ], + "angle": 0, + "content": "For mixed degradation tasks, the best results for LPIPS metrics are achieved by NeRFLix w. ref [53]. This is because NeRFLiX can see two high-quality ground-truth images from the nearest two viewpoints when inference, which leads to information leakage. However, high-quality ground-truth information is not accessed in our setting or in any real-world cases. After eliminating the impact of ground truth (NeRFLiX w/o. ref) by replacing the reference ground truth images with degraded images, our method performs better than NeRFLiX. Qualitative Results. To further verify the restoration capability, we present visual results for different degradations in Fig. 4. Our method is able to generate realistic details while other methods tend to generate smooth results which lack high-quality details. For example, for the super-resolution task, we show the drum restoration for each method, as can be seen in Fig. 4, our method generates high-fidelity drum surfaces with sharp edges, while other methods suffer from severe blur on the drum surfaces and edges. For mixed degradation tasks, our method successfully restores the intricate golden textures on the chair surface while other methods only rendered extremely blurred texture." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Geometry Refinement. Compared with previous NeRF restoration methods, most of which focus on using 2D refinement to resolve frame-wise defects, our method firstly leverages the general priors of large foundation models and GAN, achieving the open-domain 3D-based restoration. Thus, our restoration algorithm could not only recover the high-quality and view-consistent images, but also refine the 3D geometry. We demonstrate this advantage using the Ficus data of Blender, as shown in Fig. 5a, through comparing the extracted mesh from the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Wu et al." + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.149, + 0.288, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.23, + 0.27, + 0.24 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.15, + 0.359, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.307, + 0.231, + 0.352, + 0.241 + ], + "angle": 0, + "content": "Perframe" + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.151, + 0.415, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.377, + 0.231, + 0.401, + 0.24 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.244, + 0.348, + 0.255 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.182, + 0.573, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.543, + 0.23, + 0.56, + 0.238 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.148, + 0.648, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.62, + 0.229, + 0.643, + 0.238 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.148, + 0.755, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.691, + 0.229, + 0.737, + 0.238 + ], + "angle": 0, + "content": "NeRF-SR" + }, + { + "type": "image_caption", + "bbox": [ + 0.627, + 0.243, + 0.648, + 0.255 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.266, + 0.785, + 0.294 + ], + "angle": 0, + "content": "Fig. 3: (a) Error map between ground truth and our method/NeRF-Perframe. (b) we showcase that the visual quality can be much better even with lower PSNR scores." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.325, + 0.784, + 0.356 + ], + "angle": 0, + "content": "restored NeRF models, our method effectively learns the restoration in the 3D tri-plane space and recovers better geometry." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.379, + 0.388, + 0.394 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.787, + 0.571 + ], + "angle": 0, + "content": "Effects of different restoration models. To investigate the influence of using different 2D restoration models, we tested two additional off-the-shelf restoration models for the super-resolution task, including diffusion-based DiffBIR [25] and non-diffusion-based SRFormer [54]. As shown in Fig. 5b, the diffusion-based model DiffBIR shows larger restoration diversity over SRFormer by measuring the diversity score. When the repaired images exhibit diversity, direct reconstruction inevitably leads to blurriness to varying degrees due to the existing multi-view inconsistency. Through modeling the distribution of the potential high-quality NeRFs, our method accommodates these inconsistencies and consistently achieves better performance over NeRF-Perframe, demonstrating the great generalization capability of RaFE to different 2D restoration models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.572, + 0.788, + 0.693 + ], + "angle": 0, + "content": "Effects of the generator. In this ablation study, we examine the influence of the generator by comparing it with the baseline that directly optimizes the NeRF parameters using GAN loss and LPIPS mentioned above on the Blender dataset. As we can observe in Fig. 6, the image rendered by generative NeRF exhibits varied fine-textured details under different random code \\( \\mathbf{z} \\). Once the generator is removed, the rendered images will contain a blurry and smooth appearance, showing the importance of using the generator to model the distribution, which can also be demonstrated by the numerical metrics in the right part of Fig. 6." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.717, + 0.368, + 0.733 + ], + "angle": 0, + "content": "5 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This paper proposes a novel generic NeRF restoration method that applies to various types of degradations, such as low resolution, blurriness, noise, and mixed degradation. The proposed method leverages the off-the-shelf image restoration methods to restore the multi-view input images individually. To tackle the geometric and appearance inconsistencies presented in multi-view images due to individual restoration, we propose to train a GAN for NeRF generation, where" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "RaFE: Generative Radiance Fields Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.16, + 0.319, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.16, + 0.411, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.16, + 0.504, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.16, + 0.596, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.16, + 0.688, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.16, + 0.78, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.244, + 0.318, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.256, + 0.301, + 0.288, + 0.309 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.244, + 0.411, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.341, + 0.301, + 0.387, + 0.308 + ], + "angle": 0, + "content": "(b) NeRF-SR" + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.244, + 0.503, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.439, + 0.301, + 0.475, + 0.308 + ], + "angle": 0, + "content": "(c) NVSR" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.244, + 0.595, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.301, + 0.584, + 0.308 + ], + "angle": 0, + "content": "(d) NeRF-SwinIR" + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.244, + 0.687, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.61, + 0.301, + 0.677, + 0.308 + ], + "angle": 0, + "content": "(e) NeRF-Perframe" + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.244, + 0.78, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.721, + 0.301, + 0.75, + 0.308 + ], + "angle": 0, + "content": "(f) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.471, + 0.311, + 0.536, + 0.319 + ], + "angle": 0, + "content": "Super-resolution" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.327, + 0.333, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.261, + 0.396, + 0.292, + 0.405 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.327, + 0.444, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.395, + 0.421, + 0.403 + ], + "angle": 0, + "content": "(b) Deblur-NeRF" + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.327, + 0.554, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.469, + 0.396, + 0.522, + 0.403 + ], + "angle": 0, + "content": "(c) BAD-NeRF" + }, + { + "type": "image", + "bbox": [ + 0.555, + 0.327, + 0.664, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.396, + 0.647, + 0.404 + ], + "angle": 0, + "content": "(b)NeRF-Perframe" + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.327, + 0.775, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.396, + 0.743, + 0.404 + ], + "angle": 0, + "content": "(e) ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.455, + 0.408, + 0.561, + 0.415 + ], + "angle": 0, + "content": "Inconsistent Motion Deblur" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.42, + 0.334, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.257, + 0.487, + 0.288, + 0.495 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.42, + 0.444, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.36, + 0.486, + 0.419, + 0.494 + ], + "angle": 0, + "content": "(b) Deblur-NeRF" + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.42, + 0.554, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.464, + 0.486, + 0.517, + 0.493 + ], + "angle": 0, + "content": "(c) BAD-NeRF" + }, + { + "type": "image", + "bbox": [ + 0.555, + 0.42, + 0.665, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.58, + 0.486, + 0.645, + 0.494 + ], + "angle": 0, + "content": "(b)NeRF-Perframe" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.42, + 0.777, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.486, + 0.74, + 0.494 + ], + "angle": 0, + "content": "(e) ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.455, + 0.497, + 0.551, + 0.504 + ], + "angle": 0, + "content": "Consistent Motion Deblur" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.509, + 0.334, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.576, + 0.29, + 0.584 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.51, + 0.444, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.576, + 0.402, + 0.583 + ], + "angle": 0, + "content": "(b) NAN" + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.509, + 0.554, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.473, + 0.576, + 0.535, + 0.583 + ], + "angle": 0, + "content": "(b) NeRF-SwinIR" + }, + { + "type": "image", + "bbox": [ + 0.555, + 0.509, + 0.663, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.58, + 0.576, + 0.646, + 0.583 + ], + "angle": 0, + "content": "(c) NeRF-Perframe" + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.509, + 0.775, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.576, + 0.73, + 0.584 + ], + "angle": 0, + "content": "(d) ours" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.595, + 0.318, + 0.67 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.597, + 0.409, + 0.67 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.597, + 0.5, + 0.67 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.597, + 0.59, + 0.67 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.597, + 0.681, + 0.67 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.681, + 0.597, + 0.774, + 0.736 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.679, + 0.318, + 0.735 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.738, + 0.289, + 0.746 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.681, + 0.409, + 0.735 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.329, + 0.738, + 0.392, + 0.745 + ], + "angle": 0, + "content": "(b) NeRF-SwinIR" + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.681, + 0.5, + 0.735 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.421, + 0.738, + 0.489, + 0.745 + ], + "angle": 0, + "content": "(c) NeRFLiX w.ref" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.681, + 0.59, + 0.735 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.738, + 0.584, + 0.745 + ], + "angle": 0, + "content": "(d) NeRFLiX w/o.ref" + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.681, + 0.68, + 0.735 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.738, + 0.671, + 0.745 + ], + "angle": 0, + "content": "(e) NeRF-Perframe" + }, + { + "type": "image", + "bbox": [ + 0.681, + 0.681, + 0.774, + 0.735 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.738, + 0.744, + 0.745 + ], + "angle": 0, + "content": "(f) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.465, + 0.748, + 0.539, + 0.756 + ], + "angle": 0, + "content": "Mixed Degradation" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.774, + 0.788, + 0.83 + ], + "angle": 0, + "content": "Fig. 4: Visual results for super-resolution and mixed-degradation tasks. We show that our method is capable of generating detailed geometry and texture while other methods tend to be smooth in both geometry and texture. We recommend zooming in for better visualization." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Wu et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.142, + 0.46, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.148, + 0.785, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.258, + 0.784, + 0.288 + ], + "angle": 0, + "content": "Fig. 5: (a) Geometry comparisons between NeRF-SWINIR/NeRFLIX and our method. (b) Comparisons between using different 2D restoration models." + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.317, + 0.601, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.657, + 0.324, + 0.787, + 0.384 + ], + "angle": 0, + "content": "
GeneratorX✓
LPIPS↓0.0710.062
LIPE↑3.3704.621
MANIQA↑0.3830.543
" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.416, + 0.785, + 0.473 + ], + "angle": 0, + "content": "Fig. 6: Effectiveness of the tri-plane generator.Left: image rendered by NeRF without generator and images rendered by generative NeRF under different random codes \\( z(z_{0},z_{1},z_{2}) \\). Right: numerical metrics to evaluate the efficacy of the generator. The results show the effectiveness of using the generator to model the distribution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.788, + 0.687 + ], + "angle": 0, + "content": "a two-level tri-plane structure is adopted. The coarse-level tri-plane pre-trained by low-quality images remains fixed, while the fine-level residual tri-plane to be added to the coarse level is modeled by a GAN-based generator to capture variations in restoration. Extensive experiments on various restoration tasks with both synthetic and real cases demonstrate the superior performance of our method. Limitations and Future Work: One limitation of our method is instability when performing the restoration at extremely high resolutions, such as 4k, due to the patch-rendering strategy. Moreover, due to the inherent slow efficiency of NeRF rendering, currently the long training time also needs to be optimized. To overcome these limitations, a potential solution would be to integrate more efficient rendering techniques like Gaussian splatting [18] into RaFE, enabling the rendering of entire images. We plan to resolve these issues in the future work." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.713, + 0.393, + 0.731 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This work was supported in part by the Hong Kong Research Grants Council General Research Fund (17203023), in part by The Hong Kong Jockey Club Charities Trust under Grant 2022-0174, in part by the Startup Fund and the Seed Fund for Basic Research for New Staff from The University of Hong Kong, in part by the funding from UBTECH Robotics, in part by a GRF grant from the Research Grants Council (RGC) of the Hong Kong Special Administrative" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "RaFE: Generative Radiance Fields Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.178 + ], + "angle": 0, + "content": "Region, China [Project No. CityU 11208123], and in part by the National Natural Science Foundation of China (62132001)" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.203, + 0.323, + 0.219 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.237, + 0.785, + 0.279 + ], + "angle": 0, + "content": "1. Bahat, Y., Michaeli, T.: Explorable super resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2716-2725 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.28, + 0.785, + 0.307 + ], + "angle": 0, + "content": "2. Bahat, Y., Zhang, Y., Sommerhoff, H., Kolb, A., Heide, F.: Neural volume superresolution. arXiv preprint arXiv:2212.04666 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.308, + 0.785, + 0.349 + ], + "angle": 0, + "content": "3. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.35, + 0.785, + 0.377 + ], + "angle": 0, + "content": "4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.379, + 0.785, + 0.405 + ], + "angle": 0, + "content": "5. Beyer, L., Zhai, X., Kolesnikov, A.: Big vision. https://github.com/google-research/big_vision (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.407, + 0.785, + 0.462 + ], + "angle": 0, + "content": "6. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.463, + 0.785, + 0.49 + ], + "angle": 0, + "content": "7. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: European Conference on Computer Vision (ECCV) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.491, + 0.785, + 0.532 + ], + "angle": 0, + "content": "8. Chen, W.T., Yifan, W., Kuo, S.Y., Wetzstein, G.: Dehazenerf: Multiple image haze removal and 3d shape reconstruction using neural radiance fields. arXiv preprint arXiv:2303.11364 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.533, + 0.785, + 0.574 + ], + "angle": 0, + "content": "9. Chen, X., Deng, Y., Wang, B.: Mimic3d: Thriving 3d-aware gans via 3d-to-2d imitation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.575, + 0.785, + 0.602 + ], + "angle": 0, + "content": "0. Chen, Z., Zhang, Y., Ding, L., Bin, X., Gu, J., Kong, L., Yuan, X.: Hierarchical integration diffusion model for realistic image deblurring. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.603, + 0.785, + 0.629 + ], + "angle": 0, + "content": "1. Chen, Z., Zhang, Y., Gu, J., Kong, L., Yang, X., Yu, F.: Dual aggregation transformer for image super-resolution. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.631, + 0.785, + 0.685 + ], + "angle": 0, + "content": "2. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit radiance fields in space, time, and appearance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12479-12488 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.687, + 0.785, + 0.728 + ], + "angle": 0, + "content": "3. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.785, + 0.756 + ], + "angle": 0, + "content": "4. Han, Y., Yu, T., Yu, X., Wang, Y., Dai, Q.: Super-nerf: View-consistent detail generation for nef super-resolution. arXiv preprint arXiv:2304.13518 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.785, + 0.784 + ], + "angle": 0, + "content": "5. Karras, T., Aittala, M., Laine, S., Härkönen, E., Hellsten, J., Lehtinen, J., Aila, T.: Alias-free generative adversarial networks. In: Proc. NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.812 + ], + "angle": 0, + "content": "6. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of StyleGAN. In: Proc. CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. In: Advances in Neural Information Processing Systems (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.237, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Wu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "18. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.231 + ], + "angle": 0, + "content": "19. Lee, D., Lee, M., Shin, C., Lee, S.: Dp-nerf: Deblurred neural radiance field with physical scene priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12386-12396 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "20. Lee, D., Oh, J., Rim, J., Cho, S., Lee, K.M.: Exblurf: Efficient radiance fields for extreme motion blurred images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17639-17648 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.314 + ], + "angle": 0, + "content": "21. Li, H., Zhang, Z., Jiang, T., Luo, P., Feng, H., Xu, Z.: Real-world deep local motion deblurring. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1314-1322 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "22. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: ICML (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.397 + ], + "angle": 0, + "content": "23. Li, J., Zhang, Z., Liu, X., Feng, C., Wang, X., Lei, L., Zuo, W.: Spatially adaptive self-supervised learning for real-world image denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9914-9924 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "24. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. arXiv preprint arXiv:2108.10257 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.466 + ], + "angle": 0, + "content": "25. Lin, X., He, J., Chen, Z., Lyu, Z., Fei, B., Dai, B., Ouyang, W., Qiao, Y., Dong, C.: Diffbir: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "26. Liu, X., Xue, H., Luo, K., Tan, P., Yi, L.: Genn2n: Generative nerf2nerf translation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5105-5114 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "27. Ma, L., Li, X., Liao, J., Zhang, Q., Wang, X., Wang, J., Sander, P.V.: Deblur-nerf: Neural radiance fields from blurry images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12861-12870 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "28. Mildenhall, B., Barron, J.T., Chen, J., Sharlet, D., Ng, R., Carroll, R.: Burst denoising with kernel prediction networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2502-2510 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.632 + ], + "angle": 0, + "content": "29. Mildenhall, B., Hedman, P., Martin-Brualla, R., Srinivasan, P.P., Barron, J.T.: NeRF in the dark: High dynamic range view synthesis from noisy raw images. CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.674 + ], + "angle": 0, + "content": "30. Mildenhall, B., Srinivasan, P.P., Ortiz-Cayon, R., Kalantari, N.K., Ramamoorthi, R., Ng, R., Kar, A.: Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "31. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.769 + ], + "angle": 0, + "content": "32. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (Jul 2022). https://doi.org/10.1145/3528223.3530127, https://doi.org/10.1145/3528223.3530127" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.785, + 0.798 + ], + "angle": 0, + "content": "33. Pearl, N., Treibitz, T., Korman, S.: Nan: Noise-aware nerfs for burst-denoising. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "34. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "RaFE: Generative Radiance Fields Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "35. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.787, + 0.23 + ], + "angle": 0, + "content": "36. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.23, + 0.787, + 0.271 + ], + "angle": 0, + "content": "37. Skorokhodov, I., Tulyakov, S., Wang, Y., Wonka, P.: Epigraph: Rethinking training of 3d gans. Advances in Neural Information Processing Systems 35, 24487-24501 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.271, + 0.787, + 0.298 + ], + "angle": 0, + "content": "38. Sun, C., Sun, M., Chen, H.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.298, + 0.787, + 0.338 + ], + "angle": 0, + "content": "39. Tian, K., Jiang, Y., Yuan, Z., Bingyue, P., Wang, L.: Visual autoregressive modeling: Scalable image generation via next-scale prediction. arXiv preprint arXiv:2404.02905 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.338, + 0.787, + 0.379 + ], + "angle": 0, + "content": "40. Wan, Z., Paschalidou, D., Huang, I., Liu, H., Shen, B., Xiang, X., Liao, J., Guibas, L.: Cad: Photorealistic 3d generation via adversarial distillation. arXiv preprint arXiv:2312.06663 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.379, + 0.787, + 0.434 + ], + "angle": 0, + "content": "41. Wan, Z., Richardt, C., Božić, A., Li, C., Rengarajan, V., Nam, S., Xiang, X., Li, T., Zhu, B., Ranjan, R., Liao, J.: Learning neural duplex radiance fields for real-time view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8307-8316 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.434, + 0.787, + 0.473 + ], + "angle": 0, + "content": "42. Wan, Z., Zhang, B., Chen, D., Zhang, P., Chen, D., Liao, J., Wen, F.: Bringing old photos back to life. In: proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2747-2757 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.473, + 0.787, + 0.514 + ], + "angle": 0, + "content": "43. Wang, C., Wu, X., Guo, Y.C., Zhang, S.H., Tai, Y.W., Hu, S.M.: Nerf-sr: High quality neural radiance fields using supersampling. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 6445-6454 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.514, + 0.787, + 0.556 + ], + "angle": 0, + "content": "44. Wang, P., Zhao, L., Ma, R., Liu, P.: BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4170-4179 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.556, + 0.787, + 0.597 + ], + "angle": 0, + "content": "45. Wang, P., Zhao, L., Ma, R., Liu, P.: Bad-nerf: Bundle adjusted deblur neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4170-4179 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.597, + 0.787, + 0.637 + ], + "angle": 0, + "content": "46. Wang, Y., Yu, J., Zhang, J.: Zero-shot image restoration using denoising diffusion null-space model. The Eleventh International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.637, + 0.787, + 0.678 + ], + "angle": 0, + "content": "47. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.678, + 0.787, + 0.733 + ], + "angle": 0, + "content": "48. Yang, S., Wu, T., Shi, S., Lao, S., Gong, Y., Cao, M., Wang, J., Yang, Y.: Maniaq: Multi-dimension attention network for no-reference image quality assessment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1191-1200 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.733, + 0.787, + 0.772 + ], + "angle": 0, + "content": "49. Zhang, K., Liang, J., Van Gool, L., Timofte, R.: Designing a practical degradation model for deep blind image super-resolution. In: IEEE International Conference on Computer Vision. pp. 4791-4800 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.772, + 0.787, + 0.8 + ], + "angle": 0, + "content": "50. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.8, + 0.787, + 0.841 + ], + "angle": 0, + "content": "51. Zhang, W., Zhai, G., Wei, Y., Yang, X., Ma, K.: Blind image quality assessment via vision-language correspondence: A multitask learning perspective. In: IEEE Conference on Computer Vision and Pattern Recognition. pp. 14071-14081 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Wu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.188 + ], + "angle": 0, + "content": "52. Zhang, W., Li, X., Chen, X., Qiao, Y., Wu, X.M., Dong, C.: Seal: A framework for systematic evaluation of real-world super-resolution. arXiv preprint arXiv:2309.03020 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.245 + ], + "angle": 0, + "content": "53. Zhou, K., Li, W., Wang, Y., Hu, T., Jiang, N., Han, X., Lu, J.: Nerflix: High-quality neural view synthesis by learning a degradation-driven inter-viewpoint mixer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12363-12374 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.286 + ], + "angle": 0, + "content": "54. Zhou, Y., Li, Z., Guo, C.L., Bai, S., Cheng, M.M., Hou, Q.: Srformer: Permuted self-attention for single image super-resolution. arXiv preprint arXiv:2303.09735 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.286 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/RaFE_ Generative Radiance Fields Restoration/8e187213-9e85-483c-8762-09f958b4b9d0_origin.pdf b/2024/RaFE_ Generative Radiance Fields Restoration/8e187213-9e85-483c-8762-09f958b4b9d0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a09404175402199b4e08dd67f21a46a912c018d6 --- /dev/null +++ b/2024/RaFE_ Generative Radiance Fields Restoration/8e187213-9e85-483c-8762-09f958b4b9d0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8ee4b176b1c017faa8277695202a1d9133764f355989baa3807c5c4d5338f674 +size 4934082 diff --git a/2024/RaFE_ Generative Radiance Fields Restoration/full.md b/2024/RaFE_ Generative Radiance Fields Restoration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a9cc123c4a2475cdb7a06dee423f49fc74077da2 --- /dev/null +++ b/2024/RaFE_ Generative Radiance Fields Restoration/full.md @@ -0,0 +1,429 @@ +# RaFE: Generative Radiance Fields Restoration + +Zhongkai Wu $^{1}$ , Ziyu Wan $^{2}$ , Jing Zhang $^{1\star}$ , Jing Liao $^{2}$ , and Dong Xu $^{3}$ + +$^{1}$ School of Software, Beihang University, China +2 City University of Hong Kong, China +3 The University of Hong Kong, China + +ZhongkaiWu@buaa.edu.cn + +zkaiwu.github.io/RaFE + +Abstract. NeRF (Neural Radiance Fields) has demonstrated tremendous potential in novel view synthesis and 3D reconstruction, but its performance is sensitive to input image quality, which struggles to achieve high-fidelity rendering when provided with low-quality sparse input viewpoints. Previous methods for NeRF restoration are tailored for specific degradation type, ignoring the generality of restoration. To overcome this limitation, we propose a generic radiance fields restoration pipeline, named RaFE, which applies to various types of degradations, such as low resolution, blurriness, noise, compression artifacts, or their combinations. Our approach leverages the success of off-the-shelf 2D restoration methods to recover the multi-view images individually. Instead of reconstructing a blurred NeRF by averaging inconsistencies, we introduce a novel approach using Generative Adversarial Networks (GANs) for NeRF generation to better accommodate the geometric and appearance inconsistencies present in the multi-view images. Specifically, we adopt a two-level tri-plane architecture, where the coarse level remains fixed to represent the low-quality NeRF, and a fine-level residual tri-plane to be added to the coarse level is modeled as a distribution with GAN to capture potential variations in restoration. We validate RaFE on both synthetic and real cases for various restoration tasks, demonstrating superior performance in both quantitative and qualitative evaluations, surpassing other 3D restoration methods specific to single task. Please see our project website zkaiwu.github.io/RaFE. + +Keywords: Neural Rendering $\cdot$ Generative Model $\cdot$ 3D Restoration $\cdot$ Neural Radiance Fields + +# 1 Introduction + +Recently, Neural Radiance Fields (NeRFs) [3,4,7,12,31,32,38,41] have achieved great success in novel view synthesis and 3D reconstruction. However, most NeRF methods are designed based on well-captured images from multiple views with calibrated camera parameters. In real-world applications of NeRF, the data + +![](images/4a5460e32f517f42f4190399c34c360676c2727dcc26e60c00bcfe5bdbd7926e.jpg) +Low Resolution Image Set + +![](images/a275bf643662d43a087ca5cfdc10d6354ffda6dd7a1085956e515c6780a6de88.jpg) +NeRF + +![](images/4217439998a23fc74817423ba11c85a3f2485ee38f6677a2760e70423673dc47.jpg) +Ours (x4) + +![](images/4d648344512e15ab63f3009c91098c6e6fb7ae6b909086f0c8923a5ffcf89049.jpg) +Ours (x8) + +![](images/f809c55d9dcd61fa009dcb98dd895cb661a4705e86d5c4ebfe4080b6608c8aa1.jpg) +Mixed Degradation Image Set + +![](images/be5d652e9257010efd4d055a512bb1619d37f2289a69293bebef0a8c66d7c3aa.jpg) +NeRF + +![](images/b2bcd2da6c9795d089ddf88e7079cb751d3f5ea82b2b73d62fc049d3c32f8f24.jpg) +Ours + +![](images/fee409245c09265cace66219303880641c1612dbe3b38933d06274815b766f38.jpg) +Real Blur Image Set + +![](images/f2f6aaf43d3f0f50b8d335e1bc345bd4cd44cac233cf8f0d0c15b3090c70feca.jpg) +NeRF + +![](images/fd235358287efbd47fb0d4cc1bae7ca0c154c40b971b02d1e82a2ecd3cfaf8fa.jpg) +Ours +Fig. 1: High-quality restoration of radiance fields from various types of degradations. Given only degraded images, our method can restore high-quality NeRF. It is a generic approach that can be applied to various types of degradations, resulting in refinement of both geometry and appearance. + +capture or transmission process often introduces various forms of image degradations, such as noise generated during photography in low-light conditions [29,33] and blur caused by camera motion [27,44], or JPEG compression and down-sampling during transmission [2,43]. Simply restoring degraded images frame-by-frame can result in inconsistencies of geometry and appearance across different viewpoints. Directly reconstructing 3D models over these per-frame restoration results can easily induce inferior quality since current NeRF methods heavily rely on pixel-wise independent ray optimization with local computations, which are highly susceptible to noise and degradation. + +Several NeRF variants have attempted to reconstruct 3D scenes with degraded multi-view images by introducing specific strategies or additional constraints for the optimization of radiance fields. For example, [19,20,27] deal with image blur artifacts by modeling the degradation kernel with NeRF, while [2,43] super-sampling on rays or tri-planes to obtain high-resolution 3D from low-resolution observations. Additionally, [31] modifies NeRF to reconstruct the scene in linear HDR space by supervising directly on noisy raw images to address the noise generated in low-light conditions. [53] tries to improve the view synthesis quality by removing NeRF-specific rendering artifacts. Even with great success, all these approaches are only designed to handle specific types of degradation. To the best of our knowledge, currently there is no generic pipeline which supports the restoration of radiance fields under various types of degradation. + +In this paper, we propose RaFE, a generic NeRF restoration framework that enables high-quality radiance field reconstruction from captured images containing various types of degradations in a generative manner. Firstly, we leverage the success of off-the-shelf image restoration methods to restore multi-view images with different forms of degradations, such as super-resolution, deblurring, denoising, removing compression artifacts, or a combination thereof. It is important to note that since the images from different views are independently restored, there inevitably exist geometric and appearance inconsistencies between them. + +Naively optimizing a radiance field with these refined images would average out the inconsistencies and result in blurry outputs. To overcome this challenge, our insight here is, instead of reconstructing a single 3D using inconsistent frames, we could consider these restored multi-view images as the renderings from multiple distinct high-quality NeRF models with varied geometry and appearance. In this case, we abandon the commonly-used pixel-wise reconstruction objective and propose to leverage the generative adversarial networks (GANs) to model the distribution of these different high-quality NeRF models, which could effectively capture the inherent variability in ill-posed inverse problem, allowing for a better accommodation of the inconsistencies present in different views. + +Specifically, our pipeline consists of two main stages. In the first stage, based on the degradation type, we employ the corresponding off-the-shelf image restoration methods [17, 25, 34, 35, 46, 49] to obtain a set of high-quality multiview images. In practice, we prefer choosing restoration methods which have strong capabilities to recover high-quality and realistic texture details. In the second stage, we train a 3D generative model based on these restored multi-view images. Drawing inspirations from recent 3D generation works [6, 9, 37, 40], we construct a convolutional neural network (CNN) to generate tri-plane features, which are subsequently sampled and decoded into density and colors using MLP networks for NeRF rendering. Here, instead of generating single-level tri-plane features as previous works did, we decompose the tri-planes into two levels. The coarse-level tri-planes are constructed directly from low-quality images and remain fixed during training, representing the coarse structure of the modeled 3D distribution. Simultaneously, we train a generator to output the diverse fine-level tri-plane features, which act as residuals to be added to the coarse-level features for NeRF rendering. By focusing on learning the residual representations instead of the entire tri-planes for NeRF, we simplify the modeling and learning of restoration variations since we only need to learn the details while the coarse structure is provided by coarse-level tri-planes, which makes great improvement in rendering quality for more complex regions. To train the generator, we adopt an adversarial loss defined on rendered 2D images to encourage them to be indistinguishable from the high-quality restored images. We also incorporate a perceptual loss between the rendered images and the restored images to calculate structure constraints. Additionally, we propose patch sampling strategies to stabilize the generator training procedure. Once the generator has been trained, we can generate restored radiance fields with high quality renderings and a certain level of diversity by sampling different code in the latent space. + +We conducted extensive experiments to validate the effectiveness of our method, both qualitatively and quantitatively. The experimental results showcase the superiority of our approach in various restoration tasks, such as super-resolution (upper row of Figure 1), camera motion blur (a real-world case at the right part of the lower row of Figure 1) and the restoration of mixed degradation consisting of noise, blur, and compression (left part of the lower row of Figure 1). Our method not only generates images with richer and enhanced tex + +ture details but also achieves significant improvements in geometric refinement, as demonstrated by the mesh visualization in Figure 1. + +To summarize, our contributions are three-fold: + +- We propose a generic radiance fields restoration pipeline that is applicable to various types of degradations. +- We introduce a generative method for NeRF restoration that enables better accommodation of geometric and appearance inconsistencies present in the multi-view images, thus allowing us to incorporate the success of image restoration into 3D restoration. +- We show the restoration method performs well on various degradation scenarios with both enhanced appearance and refined geometry. + +# 2 Related Works + +# 2.1 2D image restoration + +Image restoration is a long-standing problem in low-level vision domain and significant progress has been achieved in different specific tasks including image super-resolution, deblur, denoise and blind restoration. Previously reconstruction-based methods [11,21,23,52,54] show their success in these tasks. However, those reconstruction-based methods are struggling to generate abundant high-quality details. Subsequently, generative restoration methods [10,17,25,34,35,39,42,46, 49], particularly those based on diffusion model, have shown the great capability to render high-quality details. Deepfloyd [35] proposes a super-resolution model, which concatenates the low-resolution input with random noise at pixel level as the condition image. DiffBIR [25] designs a degraded pipeline to simulate real-world degradation and utilizes the pre-trained diffusion model to generate photorealistic images. For camera motion blur, HiDiff [10] recovers images by using diffusion to generate features with abundant detailed information. + +# 2.2 Radiance Fields Restoration + +NeRF restoration aims to reconstruct high-quality NeRF given only degraded images with various artifacts such as blur, noise, or low resolution. Up to now, several works [2,8,14,19,20,26,27,31,33,43,44,53] have explored this task under specific types of degradation. [20,27,44] deal with blurred input images by designing blur kernel or optimizing camera paths for NeRF rendering process. For NeRF super-resolution, [43] increased the ray sampling density, forcing multiple rays to render pixels equal to the same pixel, and applied a 2D refinement model to get final output images. [14] introduces a CEM [1] refinement model to adjust the output of the super-resolution model for better multi-view consistency. However, the CEM refinement model ruins the structure details of the images and inconsistencies still exist, leading to a smooth reconstruction result. [31,33] mainly focus on the noise degradation of the input image. [31] modifies NeRF to reconstruct the scene in linear HDR color space to address the noise generated + +during image post-processing, while [33] achieve NeRF restoration on noisy images by using a noise-awarded encoder to aggregate features across views. [53] considers solving the degradation that occurs in NeRF reconstruction by training a 2D refinement model using simulated degraded images for typical NeRF-style artifacts, but they can not achieve refinement on geometry since the restoration only happens on rendered views. None of the existing approaches can restore NeRF in 3D space directly with flexible forms of degradation. NVSR [2] archives 3D geometry refinement by upsampling tri-plane representation, but their training processing requires tremendous amounts of 3D data, which are extremely hard to obtain in practice. [26] achieves NeRF translation by designing a VAE-like NeRF to model the distribution of the underlying 3D edits. By contrast, our method has the ability to handle more flexible forms of degradation and restore 3D geometry and appearance with the only need of an image set for an object or scene, making the 3D restoration more practical in real-world applications. + +# 3 Method + +In this section, we elaborate the details of the proposed RaFE method. We introduce how to refine the degraded views using pretrained 2D restoration model, to capture the high-quality appearance distribution in Sec. 3.1. Then, we describe our generative restoration framework including the neural representation, generator architecture and optimization in Sec. 3.2. The training strategy is introduced in Sec. 3.3. The overall pipeline could be found in Fig. 2. + +# 3.1 High-quality Image Restoration + +The recent success of the 2D image restoration task is dominated by the denoising diffusion probabilistic model, benefiting from its powerful appearance generation capability and extensive prior knowledge. Hence in this paper, we mainly borrow the recent diffusion-based restoration methods [25,35] to obtain the high-quality multi-view images from their low-quality counterparts. It should be noted that our 3D restoration framework could also work well based on non-diffusion restoration approaches, which has been verified by our experiments. + +To assist in the restoration of image details, we employ powerful image caption models [5,22] to get an accurate textual prompt for the scene, denoted as $P$ . Given a set of multi-view low-quality images $I_{l}$ , the textual prompt is produced by selecting a view that contains as much information of the scene as possible, e.g. a side view of a synthetic Lego model. The prompt $P$ and low-quality images $I_{l}$ are then fed into the restoration model [25,35] to achieve high-quality per-frame refinement. We denote the restored high-quality images as $I_{h}$ . + +The diffusion model-based image restoration methods can effectively reconstruct high-quality images from various real-world degradation types. Moreover, thanks to the powerful generalization ability of diffusion models preserved in these restoration methods, we can handle the scene restoration in open domain + +![](images/21c1cae40575dd9c4b3ead499c76e68862b7c764397c2b5e76a7f6993bd6ddae.jpg) +Fig. 2: Overview of our pipeline. Given multi-view degraded images, we utilize the off-the-shelf methods to restore high-quality multi-view images. Then, we train our Generative Restoration NeRF to generate a high-quality scene. + +without the need of re-training. However, though the restored images have better quality, more variations among different views occur due to the generative nature of these restoration models, leading to serious multi-view inconsistencies. Next, we focus on dealing with this issue with the proposed generative pipeline. + +# 3.2 Generative NeRF Restoration + +To deal with geometric and appearance inconsistencies across views, we treat these restored multi-view images distributed similarly to the rendered images from multiple slightly different high-quality NeRF models. Consequently, instead of directly fitting a NeRF model using the refined high-quality views which usually leads to blurry reconstruction, we are trying to learn the distribution of these diverse NeRF models by leveraging a generative method, allowing us to sample distinct restored 3D under the same degraded inputs. + +3D Representation. Following the recent 3D generative model [6,9,37], we adopt the hybrid explicit-implicit tri-plane representation for the feature field. This representation combines both explicit and implicit components to effectively model the density and RGB values. More specifically, to obtain the descriptor for any query location $\pmb{x} \in \mathbb{R}^3$ , we project the point onto three planes $(P_{xy}, P_{yz}, P_{zx})$ to retrieve the corresponding features $(f_{xy}, f_{yz}, f_{zx})$ using interpolation. Then we calculate the mean value of these three sampled features to obtain the final feature representation. + +Once we obtain the features for a point along the ray, we use two MLPs to decode the density and RGB values. The first MLP, denoted as $\mathcal{M}_{\text{dens}}$ , maps the features to the density value $\sigma \in \mathbb{R}$ and a color feature $\pmb{f}_{color}$ . The second MLP, denoted as $\mathcal{M}_{color}$ , takes the color feature and the view direction as inputs and maps them to the RGB value $\pmb{c} \in \mathbb{R}^3$ . We empirically show that incorporating the view direction allows us to effectively model view-dependent effects, particularly when there are non-Lambertian surfaces in the scene. After obtaining the densities and RGB values for each sampled point along the emitted ray, we can + +apply the classic volumetric rendering to obtain the color value for each pixel: + +$$ +C (\boldsymbol {r}) = \sum_ {i = 1} ^ {N} T _ {i} (1 - \exp (- \sigma_ {i} \delta_ {i})) \boldsymbol {c} _ {i}, \quad T _ {i} = \exp (- \sum_ {j = 1} ^ {i - 1} \sigma_ {i} \delta_ {i}), \tag {1} +$$ + +where $\pmb{c}_i, \sigma_i$ represents RGB and density value of the $i_{th}$ point alone the ray, respectively. $N$ is the sampled number and $\delta_i$ is the distance between samples. + +NeRF Generator. Given restored multi-view images, we regard them as renderings from several diverse high-quality NeRF models with varied geometry and appearance. The distribution formed by these distinct NeRF models, formulated as $p_r(\phi | I_h)$ , can be modeled by a generator. With the hybrid explicit-implicit tri-plane representation, we introduce a StyleGAN2 [15]-like CNN-based generator, which receives a latent code $\mathbf{w}$ mapped from random code $\mathbf{z}$ to generate fine tri-planes $P_f$ for high-quality NeRF rendering. To fully leverage the degraded images and make the generator only focus on generating the necessary refinements, we also pre-train a coarse NeRF model with tri-plane features denoted as $P_c$ using the input low-quality images. Specifically, the coarse NeRF model consists of a coarse tri-plane $P_c$ features and a decoder (where the decoder of coarse tri-planes is shared with the one of fine tri-planes). To obtain the final refined tri-planes $P$ , we combine the reconstructed tri-plane representations with the residual tri-planes generated by the generator via: $\mathbf{f} = \mathbf{f}_c + \mathbf{f}_f$ . By merging these components, RaFE effectively captures both the global geometric guidance provided by the coarse NeRF and the local refinements learned from 3D generator, enabling us to obtain the restored tri-planes that exhibit enhanced geometric accuracy and appearance fidelity. + +Optimization. To supervise the generator and NeRF parameters, we propose to minimize the distribution discrepancy between the rendered images and the restored high-quality images. We adopt a saturate GAN [13] loss with an image level discriminator. Specifically, we treat the high-quality images restored by 2D model as the real samples while the rendered images as the fake samples, and utilize adversarial loss with R1 regularization between the real and fake images: + +$$ +\begin{array}{l} \mathcal {L} _ {a d v} = \mathbb {E} _ {\boldsymbol {z} \sim p _ {z}, \boldsymbol {\theta} \sim p _ {\theta}} [ f (D (G (\boldsymbol {z}, \boldsymbol {\theta})) ] \\ + \mathbb {E} _ {I _ {h} \sim p _ {h}} [ f (- D (I _ {h})) + \lambda \| \nabla D (I _ {h}) \| ^ {2} ], \tag {2} \\ \end{array} +$$ + +where $z, \theta$ represent random code and view point, respectively, and $I_h$ is the restored high quality image. However, we observed relying solely on a GAN loss for training can lead to significant geometric mismatches between the restored images and rendered views. We argue that although the GAN loss helps align the distribution of 2D renderings, it still lacks geometry-level constraints. Therefore, we also incorporate a perceptual loss that encourages the rendered images to resemble the geometry of pre-frame restoration. + +$$ +\mathcal {L} _ {\text {g e o m e t r y}} = L P I P S \left(I _ {h}, G (\boldsymbol {z}, \boldsymbol {\theta})\right), \tag {3} +$$ + +where $LPIPS(\cdot, \cdot)$ refers to the learned perceptual image patch similarity proposed in [50]. $I_h$ is the restored high quality image paired with view point $\theta$ , and + +$G(\pmb{z}, \pmb{\theta})$ are the rendered image using the same training view $\pmb{\theta}$ and a random sampled latent code $\pmb{z}$ . We also supervise the coarse NeRF branch: + +$$ +\mathcal {L} _ {r e c} = \mathbb {E} _ {\boldsymbol {\theta} \sim p _ {\theta}} \left[ \| G _ {c} (\boldsymbol {\theta}) - I _ {l} ^ {\theta} \| ^ {2} \right], \tag {4} +$$ + +where $p_{\theta}$ indicate the view point distribution, and $I_l^\theta$ is the low quality image corresponding to view point $\pmb{\theta}$ . Overall, the complete training objective is: + +$$ +\mathcal {L} = \lambda_ {\text {g e o m e t r y}} \mathcal {L} _ {\text {g e o m e t r y}} + \lambda_ {\text {a d v}} \mathcal {L} _ {\text {a d v}} + \lambda_ {\text {r e c}} \mathcal {L} _ {\text {r e c}}, \tag {5} +$$ + +where $\lambda_{\text{geometry}}, \lambda_{\text{adv}}, \lambda_{\text{rec}}$ are trade-off parameters. + +# 3.3 Patch-based Training Strategy + +During the training process, fully rendering the entire image, such as $256^2$ or $512^2$ pixels, and then feeding it into the discriminator can be computationally intensive and resource-consuming. This is because volume rendering requires the computation of density and color values for sampling points along the ray for each pixel, which can become prohibitively expensive for a whole image. Therefore, we only render a patch of the view at a time (i.e. $64^2$ ). The rendered patches and high-quality image patches are randomly selected, and the discriminator in Eq. 2 receives patches of rendered images and patches of high-quality images as fake images and real images respectively. For the perceptual loss in Eq. 3, the rendered patches and the high-quality patches have the same spacial coordinate on their original images to guarantee that the cropped patches have the same semantic and local structure. + +One limitation of using patches instead of entire images for training is that it is challenging to sample the local scenes at the boundary regions of images, particularly for forward-facing data. The imbalanced training distribution will result in mode collapse during the training process. To mitigate this issue, we employ a beta sampling strategy to determine the positions of the sampled patches. This strategy ensures that patches in the boundary regions of the image are adequately sampled. More specifically, the beta sampling can be formulated as: + +$$ +\delta_ {x}, \delta_ {y} \sim \operatorname {B e t a} (\beta (t), \beta (t)), \tag {6} +$$ + +where $\delta_x$ and $\delta_y$ denote the position offset in $x$ and $y$ directions respectively, while $\beta(t)$ are linearly annealed from $\beta(0) = 1$ to some final value $\beta(T)$ smaller than 1. By using the beta sampling strategy, we can maintain a more balanced distribution of training data that focuses more on the boundary patches, alleviating mode collapse issue and improving the overall training stability. + +# 4 Experiments + +# 4.1 Setup + +Datasets. We evaluate our model on the NeRF-Synthetic benchmark dataset [31], which contains 8 synthetic objects with images taken from different viewpoints + +uniformly distributed in the hemisphere. Following the original setting, We hold 200 viewpoints for generating high-quality training data and 200 viewpoints for testing. Further, to demonstrate the generalization ability of our method, we also evaluate our method on complex real-world LLFF scenes [30] which consists of 8 scenes captured with roughly forward-facing images. We also demonstrate the superior performance of RaFE on real-world blur [27] and noise [29] data. + +Evaluation Metrics. Following the common practice of 3D reconstruction, we try to evaluate each method with two standard image quality metrics: peak signal-to-noise ratio (PSNR), structural similarity index (SSIM) [47], which however could not reflect the real 3D restoration performance according to our observations. Due to the generative characteristic of RaFE, the recovered radiance field from RaFE is very high-quality, but may not faithfully follow the "ground-truth" 3D, since inversing a degraded signal is a highly ill-posed problem. Moreover, we also found the baselines with better scores over PSNR and SSIM still pose a degraded appearance with smooth texture details, as shown in Figure 3b. Hence, a better way is to leverage perceptual metric: learned perceptual image patch similarity (LPIPS) [50] which computes the mean squared error (MSE) between normalized features from all layers of a pre-trained VGG [36] encoder and is deemed to better correlate with human perception. We also leverage the latest non-reference-based image quality assessment metrics including LIQE [51] and MANIQA [48] to demonstrate the superior rendering quality of our method. + +Implementation Details. We implement all the experiments by PyTorch. For the 2D generator and discriminator, we use a convolutional-based generator and discriminator used in StyleGAN2 [16]. In all experiments, we choose Adam optimizer for all the modules in our pipeline, with hyperparameters $\beta_{1} = 0, \beta_{2} = 0.99$ . We use learning rate $2 \times 10^{-3}$ for both generator and discriminator. For loss weights, we use $\lambda_{mimic} = 0.5$ , $\lambda_{adv} = 1.0$ , and $\lambda_{rec} = 1.0$ for almost all experiments. We evaluate RaFE framework on 4 different 3D restoration tasks: + +- $4 \times$ Super-Resolution: On the blender dataset, we resize the image to $64 \times 64$ resolutions to get low-resolution images. As for LLFF data, we first center crop the training image to $188 \times 252$ to adapt to our $4 \times$ super-resolution task and then resize to $47 \times 63$ to get low-resolution images. And we use Deepfloyd [35] for 2D super-resolution. +- Deblur: On the LLFF dataset, we construct camera motion blur by applying the blur kernel following equation $I_{blur} = A * I_{clear}$ , where $A$ is the blur kernel and $*$ stands for the convolution operator. Different from the blur dataset proposed in Deblur-NeRF [27], which contains blurred images with varying degrees of blur, and even includes a certain amount of high-resolution images, we apply a large blur kernel size (e.g. 13) and a more complex camera motion path to all the images in the dataset. We additionally construct consistent blur datasets by using the same blur kernel to the training image in a scene, which means the camera motion trajectory is the same for the training image set. The resolution of images is the same as the super-resolution task. We use HiDiff [10] to recover high-quality images. + +- Denoise: On the LLFF dataset, we follow the noise model used in [28, 33] and we get the noisy version of a clean image $I$ according to equation $I_{\text{noisy}}(x) = \mathcal{N}(I(x), \delta_r^2 + \delta_s^2 I^2(x))$ , where $\sigma_r$ is the signal-independent read-noise parameter, $\sigma_s$ is the signal-dependent shot-noise. Following [28, 33], we use the Gain level to represent the noise strength. We use gain levels $= 8$ to get our noisy image. We use DiffBIR [25] to get high-quality images. + +- Mixed degradation: The degradation pipeline consists of three stages: blur, noise, and JPEG compression. First, we utilize Gaussian blur with a radius of 7 for the blender dataset and 3 for the LLFF dataset. Second, we add noise with std 25. And then, we apply JPEG compression. The quality of JPEG compression is 50. The resolution of images is the same as the super-resolution task. We use DiffBIR [25] to get high-quality images. + +# 4.2 Results + +Baseline Methods. For general tasks, we try a baseline that firstly restores the degraded images and uses the restored high-quality images to reconstruct a NeRF directly, denoted as NeRF-Perframe. We also use the 2D-based restoration model SwinIR [24] to do the per-view refinement for the renderings of NeRF trained by degraded image, denoted as NeRF-SwinIR. Note that we do not evaluate NeRF-SwinIR for the deblur task since there are no corresponding checkpoints. + +To more thoroughly test the effectiveness of our method, we also select some task-specific competitors. For super-resolution task, we choose NeRF-SR [43], Neural Volume Super-Resolution (NVSR) [1] as baselines. For mixed degradation, since there is no existing method tailored for mixed degradation, we choose NeRFLiX, which tries to solve the NeRF-like degradation by training a 2D refinement model using degradation images constructed by a degradation simulator for typical NeRF-style artifacts. We consider this to be the most relevant method. For the deblur task, we compare with two state-of-the-art methods Deblur-NeRF [27] and BAD-NeRF [45]. Deblur-NeRF designs a learnable blur kernel and applies it to rays to simulate the degrading process and BAD-NeRF directly models the camera trajectories to solve motion blur. For the denoising task, we compare with NAN [33], which uses a noise-aware encoder to aggregate the feature of multi-view images for restoration. + +Quantitative Results. We conduct extensive quantitative comparisons with various baselines across different restoration tasks in Tab. 1a for super-resolution, Tab. 2b and Tab. 2a for deblurring, Tab. 2c for denoising and Tab. 1b for mixed degradation. As analyzed before, although most of the time our method falls slightly behind on the reconstruction metrics like PSNR and SSIM compared with other baselines, which only measure local pixel-aligned similarity between the rendered novel views and the ground truth images, are less indicative since uncertainties naturally exist in generative procedure. Taking the super-resolution results on Blender data as an example (Tab. 1a), the simplest baseline NeRF-Perframe has already achieved the best reconstruction metrics, but as shown in Figure. 3a, its visual quality is vastly inferior compared with our results. Through the error map, we found the misalignment between the generated 3D and input + +
BleederLLFFLLFF
PSNR†SSIM†LPiPS‡LIPE†MANIQA†PSNR†SSIM†LPiPS‡LIPE†MANIQA†MethodPSNR†SSIM†LPiPS‡LIPE†MANIQA†PSNR†SSIM†LPiPS‡LIPE†MANIQA†
NevRF-SR26.850.9120.1352.3720.36622.650.7020.3271.3200.220
NVSR26.130.8790.1322.3010.34321.290.6060.4422.1080.177
NeRF-SwnIR24.070.8780.1193.060.38121.730.6550.3651.8080.267
NeRF-Perframe27.390.9220.0833.2760.38623.680.7450.2611.4170.257
Ours24.990.9010.0624.6210.54323.850.7520.1972.3970.322
+ +(a) Super-resolution +(b) Mixed degradation +Table 1: Quantitative comparisons on super-resolution and mixed degradation tasks. The best result without using reference is highlighted. Our method achieves great superiorities on perceptual metrics and image quality when compared with others. + +
MethodLLFF
PSNR†SSIM†LPIPS‡LIPE†MANIQA†
Deblur-NeRF23.750.7990.3071.3650.157
BAD-NeRF24.0270.7880.3131.0940.151
NeRF-Perframe21.020.6950.3620.3620.150
Ours23.230.8110.2941.1440.177
+ +(a) Deblurring + +
MethodLLFF
PSNR↑SSIM↑LPIPS↓LIPE↑MANIQA↑
Debhr-NeRF21.710.7490.3331.1040.122
BAD-NeRF25.400.8360.2781.1400.173
NeRF-Perframe21.570.7710.3101.1030.199
Ours22.930.7890.2521.1430.224
+ +(b) Deblurring(consistent blur) + +
MethodLLFF
PSNR↑SSIM↑LPIPS↓LIPE↑MANIQA↑
NAN25.990.8220.32081.30320.241
NeRF-SwinIR24.370.7780.3461.5790.210
NeRF-Perframe23.660.7860.2811.0950.214
Ours23.780.7910.25611.6070.257
+ +(c) Denoising with Gain 8 + +Table 2: Quantitative comparisons for deblurring and denoising. The best result without using reference is highlighted. Our method performs the best on perceptual metrics. + +3D causes the drop of PSNR and SSIM. By contrast, on the perceptual metrics like LPIPS and non-reference-based metrics including LIPE and MANIQA, which could more effectively reflect the restoration performance, our method consistently achieves better results when compared with other baselines. + +For mixed degradation tasks, the best results for LPIPS metrics are achieved by NeRFLix w. ref [53]. This is because NeRFLiX can see two high-quality ground-truth images from the nearest two viewpoints when inference, which leads to information leakage. However, high-quality ground-truth information is not accessed in our setting or in any real-world cases. After eliminating the impact of ground truth (NeRFLiX w/o. ref) by replacing the reference ground truth images with degraded images, our method performs better than NeRFLiX. Qualitative Results. To further verify the restoration capability, we present visual results for different degradations in Fig. 4. Our method is able to generate realistic details while other methods tend to generate smooth results which lack high-quality details. For example, for the super-resolution task, we show the drum restoration for each method, as can be seen in Fig. 4, our method generates high-fidelity drum surfaces with sharp edges, while other methods suffer from severe blur on the drum surfaces and edges. For mixed degradation tasks, our method successfully restores the intricate golden textures on the chair surface while other methods only rendered extremely blurred texture. + +Geometry Refinement. Compared with previous NeRF restoration methods, most of which focus on using 2D refinement to resolve frame-wise defects, our method firstly leverages the general priors of large foundation models and GAN, achieving the open-domain 3D-based restoration. Thus, our restoration algorithm could not only recover the high-quality and view-consistent images, but also refine the 3D geometry. We demonstrate this advantage using the Ficus data of Blender, as shown in Fig. 5a, through comparing the extracted mesh from the + +![](images/125f23c222781059bc595eedaea13010d055178e7ebcac57c84819d2ae1a1405.jpg) +GT + +![](images/4fb27322685fb0d02ed831c9e62001a4fae8672e70366101e354d56f18f0ff05.jpg) +Perframe +(a) + +![](images/ae7b5738179baad23c5d8a99885c6f85f119e5aa35df0cd0b3570b743304d515.jpg) +Ours + +![](images/a64e405d9bcb6565f98ca273acb8570683d6b5615f5d1573d821fa7e94289239.jpg) +GT +Fig. 3: (a) Error map between ground truth and our method/NeRF-Perframe. (b) we showcase that the visual quality can be much better even with lower PSNR scores. + +![](images/d159837477fa8b4e28a84abad3508150a2487fc4e1db10d9a2fbfa28d147f1a6.jpg) +Ours +(b) + +![](images/8db7cda11d49d7b836014492629f658286453844e06275227b4437fedc0ea210.jpg) +NeRF-SR + +restored NeRF models, our method effectively learns the restoration in the 3D tri-plane space and recovers better geometry. + +# 4.3 Ablation Study + +Effects of different restoration models. To investigate the influence of using different 2D restoration models, we tested two additional off-the-shelf restoration models for the super-resolution task, including diffusion-based DiffBIR [25] and non-diffusion-based SRFormer [54]. As shown in Fig. 5b, the diffusion-based model DiffBIR shows larger restoration diversity over SRFormer by measuring the diversity score. When the repaired images exhibit diversity, direct reconstruction inevitably leads to blurriness to varying degrees due to the existing multi-view inconsistency. Through modeling the distribution of the potential high-quality NeRFs, our method accommodates these inconsistencies and consistently achieves better performance over NeRF-Perframe, demonstrating the great generalization capability of RaFE to different 2D restoration models. + +Effects of the generator. In this ablation study, we examine the influence of the generator by comparing it with the baseline that directly optimizes the NeRF parameters using GAN loss and LPIPS mentioned above on the Blender dataset. As we can observe in Fig. 6, the image rendered by generative NeRF exhibits varied fine-textured details under different random code $\mathbf{z}$ . Once the generator is removed, the rendered images will contain a blurry and smooth appearance, showing the importance of using the generator to model the distribution, which can also be demonstrated by the numerical metrics in the right part of Fig. 6. + +# 5 Conclusions + +This paper proposes a novel generic NeRF restoration method that applies to various types of degradations, such as low resolution, blurriness, noise, and mixed degradation. The proposed method leverages the off-the-shelf image restoration methods to restore the multi-view input images individually. To tackle the geometric and appearance inconsistencies presented in multi-view images due to individual restoration, we propose to train a GAN for NeRF generation, where + +![](images/a0f86ebb671cbff4cf90754998e0ade07ffe51711e143e64f23ef30bcd37d54b.jpg) + +![](images/d8e1ded7c54cb6468732d67f90c87b48af9310694f04558c95be0524de092366.jpg) + +![](images/c17ac47c480de50c8dfded832d4b7848515dce9b0170084527e7336e876ea0cf.jpg) + +![](images/13bf006035cd16f100b8c3b89289d50fe112a5703727cce4cc9f7d2a764b3c70.jpg) + +![](images/911a6932edaaf60e960934bd95848eadce2f014775a105e683d74f4f31ccc3e9.jpg) + +![](images/35ef8f558319326a2bfb1f2246ebefe2d16ecc02176799cd64c83aa6d31cbfe5.jpg) + +![](images/3cc06a44f37365d2e012ead3fdc9ebf408275baf97c017666a5bcf9af1d8292b.jpg) +(a) Input + +![](images/28acb84c0328be37bc399141a94bd93f3c37c784e9cc537d1ada7eabd305954b.jpg) +(b) NeRF-SR + +![](images/f7d479743418b09c805deeb021d3c03f90a4021a2583ba28da44482b81b6d29a.jpg) +(c) NVSR + +![](images/98e21da0fcb3a5f87de597e418ea26c35ecfc93c6f045d52bcf2df70625498ff.jpg) +(d) NeRF-SwinIR + +![](images/8b2bb7880364077387a24a719eb58bccffc16b84e1ab7f25824f3f3845a012dd.jpg) +(e) NeRF-Perframe + +![](images/abb39be8fe5ecc748691484efff0dac8454ccbd0b0b548855a175198eaf7479f.jpg) +(f) Ours + +![](images/1594ac270108d3833ebac6fb052496e01f03be459c5932d6c1d2b81b64011382.jpg) +(a) Input + +![](images/7e4dfcc2a8e6244ad57aead5ed03ac60bf85a6f27ff5adc11f5a4a608c76c353.jpg) +(b) Deblur-NeRF + +![](images/1e5017fc02880138eb92352ff014f0f0a9b8e64be4af6ad42dc060bd244ac411.jpg) +(c) BAD-NeRF + +![](images/7bd334d5a22bb26792f48ec7dc3fce40df7cb4d8738b8cf98d7b2ea9bb74ee84.jpg) +Super-resolution +(b)NeRF-Perframe + +![](images/9c0e71139d2b51146cc179012e9f061ed74fc7288eb411b3db401d21a4e075f0.jpg) +(e) ours + +![](images/3f1d6a3eb2977f569a43f14b0b359d2e2bb1221aa20155a4b8ce3a6ad379b397.jpg) +(a) Input + +![](images/d017ae851f9be427de08ae8cbd18290c7549858df7b18f8785c2faf06a984efe.jpg) +(b) Deblur-NeRF + +![](images/e88d6f5bfd51e822affb7c15a1456a49f80fb2dff9329fd9d6f1c25b1eb8b0f6.jpg) +Inconsistent Motion Deblur +(c) BAD-NeRF + +![](images/f8a167f4bb04e0f54a25cbef71d6a2555fff3b68b4a88f5f136233a70cb4b398.jpg) +(b)NeRF-Perframe + +![](images/e803680c6d14ef2f6955c239b74e836f16191fd3551d1caed6d366355cf34ead.jpg) +(e) ours + +![](images/6d4017c6f82e9a244871b873e5b5f55f81ced7b302bd7b3fa1809623bc829193.jpg) +(a) Input + +![](images/576d9f561108186776eed629b11dd3f7d7738d47d1f74bf063dd158067df36ff.jpg) +(b) NAN + +![](images/b820cf0acfe28c5c28a0590109a68a998b76746e4180566cf5d30823ac74cf84.jpg) +Consistent Motion Deblur +(b) NeRF-SwinIR + +![](images/d3e723c5b981640f8a68b34f66616ed1f9e23c7d53bdf2e57da994057eecc1b4.jpg) +(c) NeRF-Perframe + +![](images/5731ea98f68847a46743ed17894e068f7a14c66126025f60eec937a8b5199feb.jpg) +(d) ours + +![](images/72ee40d01719d85eaaf89f3d1605e7bf4269853ea204384584fcad68a9c5a810.jpg) + +![](images/8880675a671b5a78cfa68d1a31b542c496e0f0965c7351444692a041ed0c9e76.jpg) + +![](images/976fc9f1aa14575ecb4aff242d6f6d79d5d65dc1df58e2f114656a26736dacd2.jpg) + +![](images/4064ad80d76071bb260ebe49be1520946845fab17a0119d79677bb6733b0586f.jpg) + +![](images/68dee7be4e7bfda9e21d06c4cbf40717ecef00e944b2ffce5fea1ab6c470e10b.jpg) + +![](images/9ac75037c80568de0fff6dc282dc547ab99f3bbcf4d79005038e14984d9fe1ac.jpg) +(f) Ours + +![](images/688c9e2f34235af8266913dae2f012b88a637f642ff5b760ebe95ad520cb24b6.jpg) +(a) Input +Fig. 4: Visual results for super-resolution and mixed-degradation tasks. We show that our method is capable of generating detailed geometry and texture while other methods tend to be smooth in both geometry and texture. We recommend zooming in for better visualization. + +![](images/ba662c96531ef2f998329ee09d48e8c516c4ba42c4d159bfee8dca2623aff237.jpg) +(b) NeRF-SwinIR + +![](images/77d42aa4df6c959a9feac61b2a638254271513d592115dae84e0c4f4a048f28a.jpg) +(c) NeRFLiX w.ref +Mixed Degradation + +![](images/1b461d05594929ab4749a6a2293f58f629e3c0d2414ccc6460d9eb0da61516c7.jpg) +(d) NeRFLiX w/o.ref + +![](images/d76df3b73bbc7363bd0e5c3b02cb019a29cea3d25905ca4bf7d7820e34e562e5.jpg) +(e) NeRF-Perframe + +![](images/f1e4d071783c70441558ce6999370b1d1069e7d4d9e515f853a2b861a54fc25f.jpg) +Fig. 5: (a) Geometry comparisons between NeRF-SWINIR/NeRFLIX and our method. (b) Comparisons between using different 2D restoration models. + +![](images/59d8628f343e2e2e2ac333b3d29fa6606dc8b9bb600d020c0e923886b9809268.jpg) + +![](images/84a5b24e4bfdbf7aa0e1424316ee20b5007be85010842144e65634f8c83bcf6d.jpg) +Fig. 6: Effectiveness of the tri-plane generator.Left: image rendered by NeRF without generator and images rendered by generative NeRF under different random codes $z(z_{0},z_{1},z_{2})$ . Right: numerical metrics to evaluate the efficacy of the generator. The results show the effectiveness of using the generator to model the distribution. + +
GeneratorX✓
LPIPS↓0.0710.062
LIPE↑3.3704.621
MANIQA↑0.3830.543
+ +a two-level tri-plane structure is adopted. The coarse-level tri-plane pre-trained by low-quality images remains fixed, while the fine-level residual tri-plane to be added to the coarse level is modeled by a GAN-based generator to capture variations in restoration. Extensive experiments on various restoration tasks with both synthetic and real cases demonstrate the superior performance of our method. Limitations and Future Work: One limitation of our method is instability when performing the restoration at extremely high resolutions, such as 4k, due to the patch-rendering strategy. Moreover, due to the inherent slow efficiency of NeRF rendering, currently the long training time also needs to be optimized. To overcome these limitations, a potential solution would be to integrate more efficient rendering techniques like Gaussian splatting [18] into RaFE, enabling the rendering of entire images. We plan to resolve these issues in the future work. + +# Acknowledgments + +This work was supported in part by the Hong Kong Research Grants Council General Research Fund (17203023), in part by The Hong Kong Jockey Club Charities Trust under Grant 2022-0174, in part by the Startup Fund and the Seed Fund for Basic Research for New Staff from The University of Hong Kong, in part by the funding from UBTECH Robotics, in part by a GRF grant from the Research Grants Council (RGC) of the Hong Kong Special Administrative + +Region, China [Project No. CityU 11208123], and in part by the National Natural Science Foundation of China (62132001) + +# References + +1. Bahat, Y., Michaeli, T.: Explorable super resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2716-2725 (2020) +2. Bahat, Y., Zhang, Y., Sommerhoff, H., Kolb, A., Heide, F.: Neural volume superresolution. arXiv preprint arXiv:2212.04666 (2022) +3. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields (2021) +4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR (2022) +5. Beyer, L., Zhai, X., Kolesnikov, A.: Big vision. https://github.com/google-research/big_vision (2022) +6. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022) +7. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: European Conference on Computer Vision (ECCV) (2022) +8. Chen, W.T., Yifan, W., Kuo, S.Y., Wetzstein, G.: Dehazenerf: Multiple image haze removal and 3d shape reconstruction using neural radiance fields. arXiv preprint arXiv:2303.11364 (2023) +9. Chen, X., Deng, Y., Wang, B.: Mimic3d: Thriving 3d-aware gans via 3d-to-2d imitation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023) +0. Chen, Z., Zhang, Y., Ding, L., Bin, X., Gu, J., Kong, L., Yuan, X.: Hierarchical integration diffusion model for realistic image deblurring. In: NeurIPS (2023) +1. Chen, Z., Zhang, Y., Gu, J., Kong, L., Yang, X., Yu, F.: Dual aggregation transformer for image super-resolution. In: ICCV (2023) +2. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit radiance fields in space, time, and appearance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12479-12488 (2023) +3. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014) +4. Han, Y., Yu, T., Yu, X., Wang, Y., Dai, Q.: Super-nerf: View-consistent detail generation for nef super-resolution. arXiv preprint arXiv:2304.13518 (2023) +5. Karras, T., Aittala, M., Laine, S., Härkönen, E., Hellsten, J., Lehtinen, J., Aila, T.: Alias-free generative adversarial networks. In: Proc. NeurIPS (2021) +6. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of StyleGAN. In: Proc. CVPR (2020) +7. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. In: Advances in Neural Information Processing Systems (2022) + +18. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/ +19. Lee, D., Lee, M., Shin, C., Lee, S.: Dp-nerf: Deblurred neural radiance field with physical scene priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12386-12396 (2023) +20. Lee, D., Oh, J., Rim, J., Cho, S., Lee, K.M.: Exblurf: Efficient radiance fields for extreme motion blurred images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17639-17648 (2023) +21. Li, H., Zhang, Z., Jiang, T., Luo, P., Feng, H., Xu, Z.: Real-world deep local motion deblurring. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1314-1322 (2023) +22. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: ICML (2022) +23. Li, J., Zhang, Z., Liu, X., Feng, C., Wang, X., Lei, L., Zuo, W.: Spatially adaptive self-supervised learning for real-world image denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9914-9924 (June 2023) +24. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. arXiv preprint arXiv:2108.10257 (2021) +25. Lin, X., He, J., Chen, Z., Lyu, Z., Fei, B., Dai, B., Ouyang, W., Qiao, Y., Dong, C.: Diffbir: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070 (2023) +26. Liu, X., Xue, H., Luo, K., Tan, P., Yi, L.: Genn2n: Generative nerf2nerf translation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5105-5114 (2024) +27. Ma, L., Li, X., Liao, J., Zhang, Q., Wang, X., Wang, J., Sander, P.V.: Deblur-nerf: Neural radiance fields from blurry images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12861-12870 (2022) +28. Mildenhall, B., Barron, J.T., Chen, J., Sharlet, D., Ng, R., Carroll, R.: Burst denoising with kernel prediction networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2502-2510 (2018) +29. Mildenhall, B., Hedman, P., Martin-Brualla, R., Srinivasan, P.P., Barron, J.T.: NeRF in the dark: High dynamic range view synthesis from noisy raw images. CVPR (2022) +30. Mildenhall, B., Srinivasan, P.P., Ortiz-Cayon, R., Kalantari, N.K., Ramamoorthi, R., Ng, R., Kar, A.: Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG) (2019) +31. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) +32. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (Jul 2022). https://doi.org/10.1145/3528223.3530127, https://doi.org/10.1145/3528223.3530127 +33. Pearl, N., Treibitz, T., Korman, S.: Nan: Noise-aware nerfs for burst-denoising. In: CVPR (2022) +34. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022) + +35. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022) +36. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014) +37. Skorokhodov, I., Tulyakov, S., Wang, Y., Wonka, P.: Epigraph: Rethinking training of 3d gans. Advances in Neural Information Processing Systems 35, 24487-24501 (2022) +38. Sun, C., Sun, M., Chen, H.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: CVPR (2022) +39. Tian, K., Jiang, Y., Yuan, Z., Bingyue, P., Wang, L.: Visual autoregressive modeling: Scalable image generation via next-scale prediction. arXiv preprint arXiv:2404.02905 (2024) +40. Wan, Z., Paschalidou, D., Huang, I., Liu, H., Shen, B., Xiang, X., Liao, J., Guibas, L.: Cad: Photorealistic 3d generation via adversarial distillation. arXiv preprint arXiv:2312.06663 (2023) +41. Wan, Z., Richardt, C., Božić, A., Li, C., Rengarajan, V., Nam, S., Xiang, X., Li, T., Zhu, B., Ranjan, R., Liao, J.: Learning neural duplex radiance fields for real-time view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8307-8316 (June 2023) +42. Wan, Z., Zhang, B., Chen, D., Zhang, P., Chen, D., Liao, J., Wen, F.: Bringing old photos back to life. In: proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2747-2757 (2020) +43. Wang, C., Wu, X., Guo, Y.C., Zhang, S.H., Tai, Y.W., Hu, S.M.: Nerf-sr: High quality neural radiance fields using supersampling. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 6445-6454 (2022) +44. Wang, P., Zhao, L., Ma, R., Liu, P.: BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4170-4179 (June 2023) +45. Wang, P., Zhao, L., Ma, R., Liu, P.: Bad-nerf: Bundle adjusted deblur neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4170-4179 (2023) +46. Wang, Y., Yu, J., Zhang, J.: Zero-shot image restoration using denoising diffusion null-space model. The Eleventh International Conference on Learning Representations (2023) +47. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004) +48. Yang, S., Wu, T., Shi, S., Lao, S., Gong, Y., Cao, M., Wang, J., Yang, Y.: Maniaq: Multi-dimension attention network for no-reference image quality assessment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1191-1200 (2022) +49. Zhang, K., Liang, J., Van Gool, L., Timofte, R.: Designing a practical degradation model for deep blind image super-resolution. In: IEEE International Conference on Computer Vision. pp. 4791-4800 (2021) +50. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018) +51. Zhang, W., Zhai, G., Wei, Y., Yang, X., Ma, K.: Blind image quality assessment via vision-language correspondence: A multitask learning perspective. In: IEEE Conference on Computer Vision and Pattern Recognition. pp. 14071-14081 (2023) + +52. Zhang, W., Li, X., Chen, X., Qiao, Y., Wu, X.M., Dong, C.: Seal: A framework for systematic evaluation of real-world super-resolution. arXiv preprint arXiv:2309.03020 (2023) +53. Zhou, K., Li, W., Wang, Y., Hu, T., Jiang, N., Han, X., Lu, J.: Nerflix: High-quality neural view synthesis by learning a degradation-driven inter-viewpoint mixer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12363-12374 (2023) +54. Zhou, Y., Li, Z., Guo, C.L., Bai, S., Cheng, M.M., Hou, Q.: Srformer: Permuted self-attention for single image super-resolution. arXiv preprint arXiv:2303.09735 (2023) \ No newline at end of file diff --git a/2024/RaFE_ Generative Radiance Fields Restoration/images.zip b/2024/RaFE_ Generative Radiance Fields Restoration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..40aa380caff72981144d66a50519f41c761a5e91 --- /dev/null +++ b/2024/RaFE_ Generative Radiance Fields Restoration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f2d7d006dcc6343c46f875b790ed7cf8f4101481d052e241e3099e8dd95bd920 +size 512436 diff --git a/2024/RaFE_ Generative Radiance Fields Restoration/layout.json b/2024/RaFE_ Generative Radiance Fields Restoration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3701afab61fa9670f564d6c9d8379de53528061d --- /dev/null +++ b/2024/RaFE_ Generative Radiance Fields Restoration/layout.json @@ -0,0 +1,12492 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 142, + 112, + 473, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 112, + 473, + 128 + ], + "spans": [ + { + "bbox": [ + 142, + 112, + 473, + 128 + ], + "type": "text", + "content": "RaFE: Generative Radiance Fields Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 154, + 150, + 459, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 150, + 459, + 164 + ], + "spans": [ + { + "bbox": [ + 154, + 150, + 459, + 164 + ], + "type": "text", + "content": "Zhongkai Wu" + }, + { + "bbox": [ + 154, + 150, + 459, + 164 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 154, + 150, + 459, + 164 + ], + "type": "text", + "content": ", Ziyu Wan" + }, + { + "bbox": [ + 154, + 150, + 459, + 164 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 154, + 150, + 459, + 164 + ], + "type": "text", + "content": ", Jing Zhang" + }, + { + "bbox": [ + 154, + 150, + 459, + 164 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 154, + 150, + 459, + 164 + ], + "type": "text", + "content": ", Jing Liao" + }, + { + "bbox": [ + 154, + 150, + 459, + 164 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 154, + 150, + 459, + 164 + ], + "type": "text", + "content": ", and Dong Xu" + }, + { + "bbox": [ + 154, + 150, + 459, + 164 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 208, + 171, + 405, + 205 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 208, + 171, + 405, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 171, + 405, + 183 + ], + "spans": [ + { + "bbox": [ + 208, + 171, + 405, + 183 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 208, + 171, + 405, + 183 + ], + "type": "text", + "content": " School of Software, Beihang University, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 226, + 183, + 387, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 183, + 387, + 194 + ], + "spans": [ + { + "bbox": [ + 226, + 183, + 387, + 194 + ], + "type": "text", + "content": "2 City University of Hong Kong, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 226, + 194, + 386, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 194, + 386, + 205 + ], + "spans": [ + { + "bbox": [ + 226, + 194, + 386, + 205 + ], + "type": "text", + "content": "3 The University of Hong Kong, China" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 254, + 206, + 359, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 206, + 359, + 217 + ], + "spans": [ + { + "bbox": [ + 254, + 206, + 359, + 217 + ], + "type": "text", + "content": "ZhongkaiWu@buaa.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 259, + 217, + 355, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 217, + 355, + 228 + ], + "spans": [ + { + "bbox": [ + 259, + 217, + 355, + 228 + ], + "type": "text", + "content": "zkaiwu.github.io/RaFE" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 261, + 455, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 261, + 455, + 514 + ], + "spans": [ + { + "bbox": [ + 159, + 261, + 455, + 514 + ], + "type": "text", + "content": "Abstract. NeRF (Neural Radiance Fields) has demonstrated tremendous potential in novel view synthesis and 3D reconstruction, but its performance is sensitive to input image quality, which struggles to achieve high-fidelity rendering when provided with low-quality sparse input viewpoints. Previous methods for NeRF restoration are tailored for specific degradation type, ignoring the generality of restoration. To overcome this limitation, we propose a generic radiance fields restoration pipeline, named RaFE, which applies to various types of degradations, such as low resolution, blurriness, noise, compression artifacts, or their combinations. Our approach leverages the success of off-the-shelf 2D restoration methods to recover the multi-view images individually. Instead of reconstructing a blurred NeRF by averaging inconsistencies, we introduce a novel approach using Generative Adversarial Networks (GANs) for NeRF generation to better accommodate the geometric and appearance inconsistencies present in the multi-view images. Specifically, we adopt a two-level tri-plane architecture, where the coarse level remains fixed to represent the low-quality NeRF, and a fine-level residual tri-plane to be added to the coarse level is modeled as a distribution with GAN to capture potential variations in restoration. We validate RaFE on both synthetic and real cases for various restoration tasks, demonstrating superior performance in both quantitative and qualitative evaluations, surpassing other 3D restoration methods specific to single task. Please see our project website zkaiwu.github.io/RaFE." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 159, + 524, + 453, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 524, + 453, + 546 + ], + "spans": [ + { + "bbox": [ + 159, + 524, + 453, + 546 + ], + "type": "text", + "content": "Keywords: Neural Rendering " + }, + { + "bbox": [ + 159, + 524, + 453, + 546 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 524, + 453, + 546 + ], + "type": "text", + "content": " Generative Model " + }, + { + "bbox": [ + 159, + 524, + 453, + 546 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 524, + 453, + 546 + ], + "type": "text", + "content": " 3D Restoration " + }, + { + "bbox": [ + 159, + 524, + 453, + 546 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 524, + 453, + 546 + ], + "type": "text", + "content": " Neural Radiance Fields" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 569, + 230, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 230, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 230, + 582 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 597, + 482, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 597, + 482, + 646 + ], + "spans": [ + { + "bbox": [ + 130, + 597, + 482, + 646 + ], + "type": "text", + "content": "Recently, Neural Radiance Fields (NeRFs) [3,4,7,12,31,32,38,41] have achieved great success in novel view synthesis and 3D reconstruction. However, most NeRF methods are designed based on well-captured images from multiple views with calibrated camera parameters. In real-world applications of NeRF, the data" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 118, + 181, + 162 + ], + "blocks": [ + { + "bbox": [ + 137, + 118, + 181, + 162 + ], + "lines": [ + { + "bbox": [ + 137, + 118, + 181, + 162 + ], + "spans": [ + { + "bbox": [ + 137, + 118, + 181, + 162 + ], + "type": "image", + "image_path": "4a5460e32f517f42f4190399c34c360676c2727dcc26e60c00bcfe5bdbd7926e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 163, + 181, + 176 + ], + "lines": [ + { + "bbox": [ + 138, + 163, + 181, + 176 + ], + "spans": [ + { + "bbox": [ + 138, + 163, + 181, + 176 + ], + "type": "text", + "content": "Low Resolution Image Set" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 186, + 118, + 278, + 167 + ], + "blocks": [ + { + "bbox": [ + 186, + 118, + 278, + 167 + ], + "lines": [ + { + "bbox": [ + 186, + 118, + 278, + 167 + ], + "spans": [ + { + "bbox": [ + 186, + 118, + 278, + 167 + ], + "type": "image", + "image_path": "a275bf643662d43a087ca5cfdc10d6354ffda6dd7a1085956e515c6780a6de88.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 222, + 167, + 239, + 175 + ], + "lines": [ + { + "bbox": [ + 222, + 167, + 239, + 175 + ], + "spans": [ + { + "bbox": [ + 222, + 167, + 239, + 175 + ], + "type": "text", + "content": "NeRF" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 282, + 122, + 376, + 167 + ], + "blocks": [ + { + "bbox": [ + 282, + 122, + 376, + 167 + ], + "lines": [ + { + "bbox": [ + 282, + 122, + 376, + 167 + ], + "spans": [ + { + "bbox": [ + 282, + 122, + 376, + 167 + ], + "type": "image", + "image_path": "4217439998a23fc74817423ba11c85a3f2485ee38f6677a2760e70423673dc47.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 167, + 341, + 175 + ], + "lines": [ + { + "bbox": [ + 315, + 167, + 341, + 175 + ], + "spans": [ + { + "bbox": [ + 315, + 167, + 341, + 175 + ], + "type": "text", + "content": "Ours (x4)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 384, + 123, + 473, + 167 + ], + "blocks": [ + { + "bbox": [ + 384, + 123, + 473, + 167 + ], + "lines": [ + { + "bbox": [ + 384, + 123, + 473, + 167 + ], + "spans": [ + { + "bbox": [ + 384, + 123, + 473, + 167 + ], + "type": "image", + "image_path": "4d648344512e15ab63f3009c91098c6e6fb7ae6b909086f0c8923a5ffcf89049.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 401, + 167, + 427, + 175 + ], + "lines": [ + { + "bbox": [ + 401, + 167, + 427, + 175 + ], + "spans": [ + { + "bbox": [ + 401, + 167, + 427, + 175 + ], + "type": "text", + "content": "Ours (x8)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 134, + 180, + 190, + 224 + ], + "blocks": [ + { + "bbox": [ + 134, + 180, + 190, + 224 + ], + "lines": [ + { + "bbox": [ + 134, + 180, + 190, + 224 + ], + "spans": [ + { + "bbox": [ + 134, + 180, + 190, + 224 + ], + "type": "image", + "image_path": "f809c55d9dcd61fa009dcb98dd895cb661a4705e86d5c4ebfe4080b6608c8aa1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 226, + 189, + 240 + ], + "lines": [ + { + "bbox": [ + 135, + 226, + 189, + 240 + ], + "spans": [ + { + "bbox": [ + 135, + 226, + 189, + 240 + ], + "type": "text", + "content": "Mixed Degradation Image Set" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 191, + 182, + 244, + 224 + ], + "blocks": [ + { + "bbox": [ + 191, + 182, + 244, + 224 + ], + "lines": [ + { + "bbox": [ + 191, + 182, + 244, + 224 + ], + "spans": [ + { + "bbox": [ + 191, + 182, + 244, + 224 + ], + "type": "image", + "image_path": "be5d652e9257010efd4d055a512bb1619d37f2289a69293bebef0a8c66d7c3aa.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 210, + 229, + 227, + 236 + ], + "lines": [ + { + "bbox": [ + 210, + 229, + 227, + 236 + ], + "spans": [ + { + "bbox": [ + 210, + 229, + 227, + 236 + ], + "type": "text", + "content": "NeRF" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 245, + 182, + 297, + 224 + ], + "blocks": [ + { + "bbox": [ + 245, + 182, + 297, + 224 + ], + "lines": [ + { + "bbox": [ + 245, + 182, + 297, + 224 + ], + "spans": [ + { + "bbox": [ + 245, + 182, + 297, + 224 + ], + "type": "image", + "image_path": "b2bcd2da6c9795d089ddf88e7079cb751d3f5ea82b2b73d62fc049d3c32f8f24.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 263, + 229, + 277, + 236 + ], + "lines": [ + { + "bbox": [ + 263, + 229, + 277, + 236 + ], + "spans": [ + { + "bbox": [ + 263, + 229, + 277, + 236 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 298, + 182, + 359, + 224 + ], + "blocks": [ + { + "bbox": [ + 298, + 182, + 359, + 224 + ], + "lines": [ + { + "bbox": [ + 298, + 182, + 359, + 224 + ], + "spans": [ + { + "bbox": [ + 298, + 182, + 359, + 224 + ], + "type": "image", + "image_path": "fee409245c09265cace66219303880641c1612dbe3b38933d06274815b766f38.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 229, + 356, + 237 + ], + "lines": [ + { + "bbox": [ + 302, + 229, + 356, + 237 + ], + "spans": [ + { + "bbox": [ + 302, + 229, + 356, + 237 + ], + "type": "text", + "content": "Real Blur Image Set" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 361, + 182, + 440, + 224 + ], + "blocks": [ + { + "bbox": [ + 361, + 182, + 440, + 224 + ], + "lines": [ + { + "bbox": [ + 361, + 182, + 440, + 224 + ], + "spans": [ + { + "bbox": [ + 361, + 182, + 440, + 224 + ], + "type": "image", + "image_path": "f2f6aaf43d3f0f50b8d335e1bc345bd4cd44cac233cf8f0d0c15b3090c70feca.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 382, + 229, + 399, + 236 + ], + "lines": [ + { + "bbox": [ + 382, + 229, + 399, + 236 + ], + "spans": [ + { + "bbox": [ + 382, + 229, + 399, + 236 + ], + "type": "text", + "content": "NeRF" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 441, + 182, + 479, + 224 + ], + "blocks": [ + { + "bbox": [ + 441, + 182, + 479, + 224 + ], + "lines": [ + { + "bbox": [ + 441, + 182, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 441, + 182, + 479, + 224 + ], + "type": "image", + "image_path": "fd235358287efbd47fb0d4cc1bae7ca0c154c40b971b02d1e82a2ecd3cfaf8fa.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 229, + 457, + 236 + ], + "lines": [ + { + "bbox": [ + 443, + 229, + 457, + 236 + ], + "spans": [ + { + "bbox": [ + 443, + 229, + 457, + 236 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 250, + 482, + 294 + ], + "lines": [ + { + "bbox": [ + 131, + 250, + 482, + 294 + ], + "spans": [ + { + "bbox": [ + 131, + 250, + 482, + 294 + ], + "type": "text", + "content": "Fig. 1: High-quality restoration of radiance fields from various types of degradations. Given only degraded images, our method can restore high-quality NeRF. It is a generic approach that can be applied to various types of degradations, resulting in refinement of both geometry and appearance." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 319, + 480, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 480, + 426 + ], + "type": "text", + "content": "capture or transmission process often introduces various forms of image degradations, such as noise generated during photography in low-light conditions [29,33] and blur caused by camera motion [27,44], or JPEG compression and down-sampling during transmission [2,43]. Simply restoring degraded images frame-by-frame can result in inconsistencies of geometry and appearance across different viewpoints. Directly reconstructing 3D models over these per-frame restoration results can easily induce inferior quality since current NeRF methods heavily rely on pixel-wise independent ray optimization with local computations, which are highly susceptible to noise and degradation." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 130, + 426, + 481, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 481, + 569 + ], + "type": "text", + "content": "Several NeRF variants have attempted to reconstruct 3D scenes with degraded multi-view images by introducing specific strategies or additional constraints for the optimization of radiance fields. For example, [19,20,27] deal with image blur artifacts by modeling the degradation kernel with NeRF, while [2,43] super-sampling on rays or tri-planes to obtain high-resolution 3D from low-resolution observations. Additionally, [31] modifies NeRF to reconstruct the scene in linear HDR space by supervising directly on noisy raw images to address the noise generated in low-light conditions. [53] tries to improve the view synthesis quality by removing NeRF-specific rendering artifacts. Even with great success, all these approaches are only designed to handle specific types of degradation. To the best of our knowledge, currently there is no generic pipeline which supports the restoration of radiance fields under various types of degradation." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": "In this paper, we propose RaFE, a generic NeRF restoration framework that enables high-quality radiance field reconstruction from captured images containing various types of degradations in a generative manner. Firstly, we leverage the success of off-the-shelf image restoration methods to restore multi-view images with different forms of degradations, such as super-resolution, deblurring, denoising, removing compression artifacts, or a combination thereof. It is important to note that since the images from different views are independently restored, there inevitably exist geometric and appearance inconsistencies between them." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "Naively optimizing a radiance field with these refined images would average out the inconsistencies and result in blurry outputs. To overcome this challenge, our insight here is, instead of reconstructing a single 3D using inconsistent frames, we could consider these restored multi-view images as the renderings from multiple distinct high-quality NeRF models with varied geometry and appearance. In this case, we abandon the commonly-used pixel-wise reconstruction objective and propose to leverage the generative adversarial networks (GANs) to model the distribution of these different high-quality NeRF models, which could effectively capture the inherent variability in ill-posed inverse problem, allowing for a better accommodation of the inconsistencies present in different views." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 247, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 571 + ], + "type": "text", + "content": "Specifically, our pipeline consists of two main stages. In the first stage, based on the degradation type, we employ the corresponding off-the-shelf image restoration methods [17, 25, 34, 35, 46, 49] to obtain a set of high-quality multiview images. In practice, we prefer choosing restoration methods which have strong capabilities to recover high-quality and realistic texture details. In the second stage, we train a 3D generative model based on these restored multi-view images. Drawing inspirations from recent 3D generation works [6, 9, 37, 40], we construct a convolutional neural network (CNN) to generate tri-plane features, which are subsequently sampled and decoded into density and colors using MLP networks for NeRF rendering. Here, instead of generating single-level tri-plane features as previous works did, we decompose the tri-planes into two levels. The coarse-level tri-planes are constructed directly from low-quality images and remain fixed during training, representing the coarse structure of the modeled 3D distribution. Simultaneously, we train a generator to output the diverse fine-level tri-plane features, which act as residuals to be added to the coarse-level features for NeRF rendering. By focusing on learning the residual representations instead of the entire tri-planes for NeRF, we simplify the modeling and learning of restoration variations since we only need to learn the details while the coarse structure is provided by coarse-level tri-planes, which makes great improvement in rendering quality for more complex regions. To train the generator, we adopt an adversarial loss defined on rendered 2D images to encourage them to be indistinguishable from the high-quality restored images. We also incorporate a perceptual loss between the rendered images and the restored images to calculate structure constraints. Additionally, we propose patch sampling strategies to stabilize the generator training procedure. Once the generator has been trained, we can generate restored radiance fields with high quality renderings and a certain level of diversity by sampling different code in the latent space." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "We conducted extensive experiments to validate the effectiveness of our method, both qualitatively and quantitatively. The experimental results showcase the superiority of our approach in various restoration tasks, such as super-resolution (upper row of Figure 1), camera motion blur (a real-world case at the right part of the lower row of Figure 1) and the restoration of mixed degradation consisting of noise, blur, and compression (left part of the lower row of Figure 1). Our method not only generates images with richer and enhanced tex" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 447, + 101 + ], + "type": "text", + "content": "RaFE: Generative Radiance Fields Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 479, + 139 + ], + "type": "text", + "content": "ture details but also achieves significant improvements in geometric refinement, as demonstrated by the mesh visualization in Figure 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 147, + 140, + 356, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 140, + 356, + 152 + ], + "spans": [ + { + "bbox": [ + 147, + 140, + 356, + 152 + ], + "type": "text", + "content": "To summarize, our contributions are three-fold:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 161, + 479, + 255 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 161, + 479, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 161, + 479, + 184 + ], + "spans": [ + { + "bbox": [ + 138, + 161, + 479, + 184 + ], + "type": "text", + "content": "- We propose a generic radiance fields restoration pipeline that is applicable to various types of degradations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 185, + 479, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 185, + 479, + 231 + ], + "spans": [ + { + "bbox": [ + 138, + 185, + 479, + 231 + ], + "type": "text", + "content": "- We introduce a generative method for NeRF restoration that enables better accommodation of geometric and appearance inconsistencies present in the multi-view images, thus allowing us to incorporate the success of image restoration into 3D restoration." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 232, + 479, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 232, + 479, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 232, + 479, + 255 + ], + "type": "text", + "content": "- We show the restoration method performs well on various degradation scenarios with both enhanced appearance and refined geometry." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 274, + 242, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 242, + 286 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 242, + 286 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 298, + 266, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 298, + 266, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 298, + 266, + 310 + ], + "type": "text", + "content": "2.1 2D image restoration" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 318, + 482, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 318, + 482, + 473 + ], + "spans": [ + { + "bbox": [ + 132, + 318, + 482, + 473 + ], + "type": "text", + "content": "Image restoration is a long-standing problem in low-level vision domain and significant progress has been achieved in different specific tasks including image super-resolution, deblur, denoise and blind restoration. Previously reconstruction-based methods [11,21,23,52,54] show their success in these tasks. However, those reconstruction-based methods are struggling to generate abundant high-quality details. Subsequently, generative restoration methods [10,17,25,34,35,39,42,46, 49], particularly those based on diffusion model, have shown the great capability to render high-quality details. Deepfloyd [35] proposes a super-resolution model, which concatenates the low-resolution input with random noise at pixel level as the condition image. DiffBIR [25] designs a degraded pipeline to simulate real-world degradation and utilizes the pre-trained diffusion model to generate photorealistic images. For camera motion blur, HiDiff [10] recovers images by using diffusion to generate features with abundant detailed information." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 491, + 302, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 302, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 302, + 502 + ], + "type": "text", + "content": "2.2 Radiance Fields Restoration" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "content": "NeRF restoration aims to reconstruct high-quality NeRF given only degraded images with various artifacts such as blur, noise, or low resolution. Up to now, several works [2,8,14,19,20,26,27,31,33,43,44,53] have explored this task under specific types of degradation. [20,27,44] deal with blurred input images by designing blur kernel or optimizing camera paths for NeRF rendering process. For NeRF super-resolution, [43] increased the ray sampling density, forcing multiple rays to render pixels equal to the same pixel, and applied a 2D refinement model to get final output images. [14] introduces a CEM [1] refinement model to adjust the output of the super-resolution model for better multi-view consistency. However, the CEM refinement model ruins the structure details of the images and inconsistencies still exist, leading to a smooth reconstruction result. [31,33] mainly focus on the noise degradation of the input image. [31] modifies NeRF to reconstruct the scene in linear HDR color space to address the noise generated" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Wuet al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": "during image post-processing, while [33] achieve NeRF restoration on noisy images by using a noise-awarded encoder to aggregate features across views. [53] considers solving the degradation that occurs in NeRF reconstruction by training a 2D refinement model using simulated degraded images for typical NeRF-style artifacts, but they can not achieve refinement on geometry since the restoration only happens on rendered views. None of the existing approaches can restore NeRF in 3D space directly with flexible forms of degradation. NVSR [2] archives 3D geometry refinement by upsampling tri-plane representation, but their training processing requires tremendous amounts of 3D data, which are extremely hard to obtain in practice. [26] achieves NeRF translation by designing a VAE-like NeRF to model the distribution of the underlying 3D edits. By contrast, our method has the ability to handle more flexible forms of degradation and restore 3D geometry and appearance with the only need of an image set for an object or scene, making the 3D restoration more practical in real-world applications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 305, + 202, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 305, + 202, + 317 + ], + "spans": [ + { + "bbox": [ + 132, + 305, + 202, + 317 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 333, + 482, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 482, + 406 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 482, + 406 + ], + "type": "text", + "content": "In this section, we elaborate the details of the proposed RaFE method. We introduce how to refine the degraded views using pretrained 2D restoration model, to capture the high-quality appearance distribution in Sec. 3.1. Then, we describe our generative restoration framework including the neural representation, generator architecture and optimization in Sec. 3.2. The training strategy is introduced in Sec. 3.3. The overall pipeline could be found in Fig. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 426, + 320, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 426, + 320, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 426, + 320, + 438 + ], + "type": "text", + "content": "3.1 High-quality Image Restoration" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "text", + "content": "The recent success of the 2D image restoration task is dominated by the denoising diffusion probabilistic model, benefiting from its powerful appearance generation capability and extensive prior knowledge. Hence in this paper, we mainly borrow the recent diffusion-based restoration methods [25,35] to obtain the high-quality multi-view images from their low-quality counterparts. It should be noted that our 3D restoration framework could also work well based on non-diffusion restoration approaches, which has been verified by our experiments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "text", + "content": "To assist in the restoration of image details, we employ powerful image caption models [5,22] to get an accurate textual prompt for the scene, denoted as " + }, + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "text", + "content": ". Given a set of multi-view low-quality images " + }, + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "inline_equation", + "content": "I_{l}" + }, + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "text", + "content": ", the textual prompt is produced by selecting a view that contains as much information of the scene as possible, e.g. a side view of a synthetic Lego model. The prompt " + }, + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "text", + "content": " and low-quality images " + }, + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "inline_equation", + "content": "I_{l}" + }, + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "text", + "content": " are then fed into the restoration model [25,35] to achieve high-quality per-frame refinement. We denote the restored high-quality images as " + }, + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "inline_equation", + "content": "I_{h}" + }, + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "The diffusion model-based image restoration methods can effectively reconstruct high-quality images from various real-world degradation types. Moreover, thanks to the powerful generalization ability of diffusion models preserved in these restoration methods, we can handle the scene restoration in open domain" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 447, + 101 + ], + "type": "text", + "content": "RaFE: Generative Radiance Fields Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 115, + 479, + 232 + ], + "blocks": [ + { + "bbox": [ + 138, + 115, + 479, + 232 + ], + "lines": [ + { + "bbox": [ + 138, + 115, + 479, + 232 + ], + "spans": [ + { + "bbox": [ + 138, + 115, + 479, + 232 + ], + "type": "image", + "image_path": "21c1cae40575dd9c4b3ead499c76e68862b7c764397c2b5e76a7f6993bd6ddae.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 241, + 482, + 275 + ], + "lines": [ + { + "bbox": [ + 131, + 241, + 482, + 275 + ], + "spans": [ + { + "bbox": [ + 131, + 241, + 482, + 275 + ], + "type": "text", + "content": "Fig. 2: Overview of our pipeline. Given multi-view degraded images, we utilize the off-the-shelf methods to restore high-quality multi-view images. Then, we train our Generative Restoration NeRF to generate a high-quality scene." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 299, + 482, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 299, + 482, + 348 + ], + "spans": [ + { + "bbox": [ + 131, + 299, + 482, + 348 + ], + "type": "text", + "content": "without the need of re-training. However, though the restored images have better quality, more variations among different views occur due to the generative nature of these restoration models, leading to serious multi-view inconsistencies. Next, we focus on dealing with this issue with the proposed generative pipeline." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 367, + 313, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 367, + 313, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 367, + 313, + 379 + ], + "type": "text", + "content": "3.2 Generative NeRF Restoration" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 389, + 482, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 473 + ], + "type": "text", + "content": "To deal with geometric and appearance inconsistencies across views, we treat these restored multi-view images distributed similarly to the rendered images from multiple slightly different high-quality NeRF models. Consequently, instead of directly fitting a NeRF model using the refined high-quality views which usually leads to blurry reconstruction, we are trying to learn the distribution of these diverse NeRF models by leveraging a generative method, allowing us to sample distinct restored 3D under the same degraded inputs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 473, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 473, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 131, + 473, + 482, + 569 + ], + "type": "text", + "content": "3D Representation. Following the recent 3D generative model [6,9,37], we adopt the hybrid explicit-implicit tri-plane representation for the feature field. This representation combines both explicit and implicit components to effectively model the density and RGB values. More specifically, to obtain the descriptor for any query location " + }, + { + "bbox": [ + 131, + 473, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 131, + 473, + 482, + 569 + ], + "type": "text", + "content": ", we project the point onto three planes " + }, + { + "bbox": [ + 131, + 473, + 482, + 569 + ], + "type": "inline_equation", + "content": "(P_{xy}, P_{yz}, P_{zx})" + }, + { + "bbox": [ + 131, + 473, + 482, + 569 + ], + "type": "text", + "content": " to retrieve the corresponding features " + }, + { + "bbox": [ + 131, + 473, + 482, + 569 + ], + "type": "inline_equation", + "content": "(f_{xy}, f_{yz}, f_{zx})" + }, + { + "bbox": [ + 131, + 473, + 482, + 569 + ], + "type": "text", + "content": " using interpolation. Then we calculate the mean value of these three sampled features to obtain the final feature representation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Once we obtain the features for a point along the ray, we use two MLPs to decode the density and RGB values. The first MLP, denoted as " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\text{dens}}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ", maps the features to the density value " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sigma \\in \\mathbb{R}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " and a color feature " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{color}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ". The second MLP, denoted as " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{color}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ", takes the color feature and the view direction as inputs and maps them to the RGB value " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{c} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ". We empirically show that incorporating the view direction allows us to effectively model view-dependent effects, particularly when there are non-Lambertian surfaces in the scene. After obtaining the densities and RGB values for each sampled point along the emitted ray, we can" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 474, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 474, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 474, + 128 + ], + "type": "text", + "content": "apply the classic volumetric rendering to obtain the color value for each pixel:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 194, + 134, + 480, + 163 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 134, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 194, + 134, + 480, + 163 + ], + "type": "interline_equation", + "content": "C (\\boldsymbol {r}) = \\sum_ {i = 1} ^ {N} T _ {i} (1 - \\exp (- \\sigma_ {i} \\delta_ {i})) \\boldsymbol {c} _ {i}, \\quad T _ {i} = \\exp (- \\sum_ {j = 1} ^ {i - 1} \\sigma_ {i} \\delta_ {i}), \\tag {1}", + "image_path": "95a9fe93c2b1f076ff8a382eb3102bec1e09c89014bd4949c103b629d1ab88e1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 168, + 479, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 168, + 479, + 191 + ], + "spans": [ + { + "bbox": [ + 130, + 168, + 479, + 191 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 168, + 479, + 191 + ], + "type": "inline_equation", + "content": "\\pmb{c}_i, \\sigma_i" + }, + { + "bbox": [ + 130, + 168, + 479, + 191 + ], + "type": "text", + "content": " represents RGB and density value of the " + }, + { + "bbox": [ + 130, + 168, + 479, + 191 + ], + "type": "inline_equation", + "content": "i_{th}" + }, + { + "bbox": [ + 130, + 168, + 479, + 191 + ], + "type": "text", + "content": " point alone the ray, respectively. " + }, + { + "bbox": [ + 130, + 168, + 479, + 191 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 168, + 479, + 191 + ], + "type": "text", + "content": " is the sampled number and " + }, + { + "bbox": [ + 130, + 168, + 479, + 191 + ], + "type": "inline_equation", + "content": "\\delta_i" + }, + { + "bbox": [ + 130, + 168, + 479, + 191 + ], + "type": "text", + "content": " is the distance between samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 191, + 480, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 191, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 191, + 480, + 407 + ], + "type": "text", + "content": "NeRF Generator. Given restored multi-view images, we regard them as renderings from several diverse high-quality NeRF models with varied geometry and appearance. The distribution formed by these distinct NeRF models, formulated as " + }, + { + "bbox": [ + 130, + 191, + 480, + 407 + ], + "type": "inline_equation", + "content": "p_r(\\phi | I_h)" + }, + { + "bbox": [ + 130, + 191, + 480, + 407 + ], + "type": "text", + "content": ", can be modeled by a generator. With the hybrid explicit-implicit tri-plane representation, we introduce a StyleGAN2 [15]-like CNN-based generator, which receives a latent code " + }, + { + "bbox": [ + 130, + 191, + 480, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 130, + 191, + 480, + 407 + ], + "type": "text", + "content": " mapped from random code " + }, + { + "bbox": [ + 130, + 191, + 480, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 191, + 480, + 407 + ], + "type": "text", + "content": " to generate fine tri-planes " + }, + { + "bbox": [ + 130, + 191, + 480, + 407 + ], + "type": "inline_equation", + "content": "P_f" + }, + { + "bbox": [ + 130, + 191, + 480, + 407 + ], + "type": "text", + "content": " for high-quality NeRF rendering. To fully leverage the degraded images and make the generator only focus on generating the necessary refinements, we also pre-train a coarse NeRF model with tri-plane features denoted as " + }, + { + "bbox": [ + 130, + 191, + 480, + 407 + ], + "type": "inline_equation", + "content": "P_c" + }, + { + "bbox": [ + 130, + 191, + 480, + 407 + ], + "type": "text", + "content": " using the input low-quality images. Specifically, the coarse NeRF model consists of a coarse tri-plane " + }, + { + "bbox": [ + 130, + 191, + 480, + 407 + ], + "type": "inline_equation", + "content": "P_c" + }, + { + "bbox": [ + 130, + 191, + 480, + 407 + ], + "type": "text", + "content": " features and a decoder (where the decoder of coarse tri-planes is shared with the one of fine tri-planes). To obtain the final refined tri-planes " + }, + { + "bbox": [ + 130, + 191, + 480, + 407 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 191, + 480, + 407 + ], + "type": "text", + "content": ", we combine the reconstructed tri-plane representations with the residual tri-planes generated by the generator via: " + }, + { + "bbox": [ + 130, + 191, + 480, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{f} = \\mathbf{f}_c + \\mathbf{f}_f" + }, + { + "bbox": [ + 130, + 191, + 480, + 407 + ], + "type": "text", + "content": ". By merging these components, RaFE effectively captures both the global geometric guidance provided by the coarse NeRF and the local refinements learned from 3D generator, enabling us to obtain the restored tri-planes that exhibit enhanced geometric accuracy and appearance fidelity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 407, + 480, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 480, + 479 + ], + "type": "text", + "content": "Optimization. To supervise the generator and NeRF parameters, we propose to minimize the distribution discrepancy between the rendered images and the restored high-quality images. We adopt a saturate GAN [13] loss with an image level discriminator. Specifically, we treat the high-quality images restored by 2D model as the real samples while the rendered images as the fake samples, and utilize adversarial loss with R1 regularization between the real and fake images:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 215, + 497, + 480, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 497, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 215, + 497, + 480, + 525 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {a d v} = \\mathbb {E} _ {\\boldsymbol {z} \\sim p _ {z}, \\boldsymbol {\\theta} \\sim p _ {\\theta}} [ f (D (G (\\boldsymbol {z}, \\boldsymbol {\\theta})) ] \\\\ + \\mathbb {E} _ {I _ {h} \\sim p _ {h}} [ f (- D (I _ {h})) + \\lambda \\| \\nabla D (I _ {h}) \\| ^ {2} ], \\tag {2} \\\\ \\end{array}", + "image_path": "d80a98facd82fc0268038abe68dd22f1f67dd30dfb33ac61b69b51d879b6623e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 531, + 480, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 480, + 614 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 480, + 614 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 531, + 480, + 614 + ], + "type": "inline_equation", + "content": "z, \\theta" + }, + { + "bbox": [ + 130, + 531, + 480, + 614 + ], + "type": "text", + "content": " represent random code and view point, respectively, and " + }, + { + "bbox": [ + 130, + 531, + 480, + 614 + ], + "type": "inline_equation", + "content": "I_h" + }, + { + "bbox": [ + 130, + 531, + 480, + 614 + ], + "type": "text", + "content": " is the restored high quality image. However, we observed relying solely on a GAN loss for training can lead to significant geometric mismatches between the restored images and rendered views. We argue that although the GAN loss helps align the distribution of 2D renderings, it still lacks geometry-level constraints. Therefore, we also incorporate a perceptual loss that encourages the rendered images to resemble the geometry of pre-frame restoration." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 235, + 622, + 480, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 622, + 480, + 635 + ], + "spans": [ + { + "bbox": [ + 235, + 622, + 480, + 635 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {g e o m e t r y}} = L P I P S \\left(I _ {h}, G (\\boldsymbol {z}, \\boldsymbol {\\theta})\\right), \\tag {3}", + "image_path": "de5786711ccbe931c63ff1e25fa0ea467046d7f08b9a2e464c9b42d99005eb77.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "LPIPS(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": " refers to the learned perceptual image patch similarity proposed in [50]. " + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "I_h" + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": " is the restored high quality image paired with view point " + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": ", and" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 447, + 101 + ], + "type": "text", + "content": "RaFE: Generative Radiance Fields Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "inline_equation", + "content": "G(\\pmb{z}, \\pmb{\\theta})" + }, + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": " are the rendered image using the same training view " + }, + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": " and a random sampled latent code " + }, + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "inline_equation", + "content": "\\pmb{z}" + }, + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": ". We also supervise the coarse NeRF branch:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 242, + 149, + 480, + 163 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 149, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 242, + 149, + 480, + 163 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e c} = \\mathbb {E} _ {\\boldsymbol {\\theta} \\sim p _ {\\theta}} \\left[ \\| G _ {c} (\\boldsymbol {\\theta}) - I _ {l} ^ {\\theta} \\| ^ {2} \\right], \\tag {4}", + "image_path": "6b6052ab8f60bd4668aca46b3d0110205a3bda0bc67a6e5a0cd90ad6bbd7e968.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 479, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 479, + 195 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 479, + 195 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 171, + 479, + 195 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 132, + 171, + 479, + 195 + ], + "type": "text", + "content": " indicate the view point distribution, and " + }, + { + "bbox": [ + 132, + 171, + 479, + 195 + ], + "type": "inline_equation", + "content": "I_l^\\theta" + }, + { + "bbox": [ + 132, + 171, + 479, + 195 + ], + "type": "text", + "content": " is the low quality image corresponding to view point " + }, + { + "bbox": [ + 132, + 171, + 479, + 195 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 132, + 171, + 479, + 195 + ], + "type": "text", + "content": ". Overall, the complete training objective is:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 205, + 203, + 480, + 216 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 203, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 205, + 203, + 480, + 216 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {\\text {g e o m e t r y}} \\mathcal {L} _ {\\text {g e o m e t r y}} + \\lambda_ {\\text {a d v}} \\mathcal {L} _ {\\text {a d v}} + \\lambda_ {\\text {r e c}} \\mathcal {L} _ {\\text {r e c}}, \\tag {5}", + "image_path": "042e9a96d2194e742b47bbaa3958274cb5e010fdccf8549fe36d7f587a7fa86c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 222, + 360, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 222, + 360, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 222, + 360, + 235 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 222, + 360, + 235 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{geometry}}, \\lambda_{\\text{adv}}, \\lambda_{\\text{rec}}" + }, + { + "bbox": [ + 132, + 222, + 360, + 235 + ], + "type": "text", + "content": " are trade-off parameters." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 250, + 313, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 250, + 313, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 250, + 313, + 262 + ], + "type": "text", + "content": "3.3 Patch-based Training Strategy" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 268, + 480, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 480, + 411 + ], + "type": "text", + "content": "During the training process, fully rendering the entire image, such as " + }, + { + "bbox": [ + 130, + 268, + 480, + 411 + ], + "type": "inline_equation", + "content": "256^2" + }, + { + "bbox": [ + 130, + 268, + 480, + 411 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 268, + 480, + 411 + ], + "type": "inline_equation", + "content": "512^2" + }, + { + "bbox": [ + 130, + 268, + 480, + 411 + ], + "type": "text", + "content": " pixels, and then feeding it into the discriminator can be computationally intensive and resource-consuming. This is because volume rendering requires the computation of density and color values for sampling points along the ray for each pixel, which can become prohibitively expensive for a whole image. Therefore, we only render a patch of the view at a time (i.e. " + }, + { + "bbox": [ + 130, + 268, + 480, + 411 + ], + "type": "inline_equation", + "content": "64^2" + }, + { + "bbox": [ + 130, + 268, + 480, + 411 + ], + "type": "text", + "content": "). The rendered patches and high-quality image patches are randomly selected, and the discriminator in Eq. 2 receives patches of rendered images and patches of high-quality images as fake images and real images respectively. For the perceptual loss in Eq. 3, the rendered patches and the high-quality patches have the same spacial coordinate on their original images to guarantee that the cropped patches have the same semantic and local structure." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 412, + 480, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 480, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 480, + 496 + ], + "type": "text", + "content": "One limitation of using patches instead of entire images for training is that it is challenging to sample the local scenes at the boundary regions of images, particularly for forward-facing data. The imbalanced training distribution will result in mode collapse during the training process. To mitigate this issue, we employ a beta sampling strategy to determine the positions of the sampled patches. This strategy ensures that patches in the boundary regions of the image are adequately sampled. More specifically, the beta sampling can be formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 252, + 504, + 480, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 504, + 480, + 517 + ], + "spans": [ + { + "bbox": [ + 252, + 504, + 480, + 517 + ], + "type": "interline_equation", + "content": "\\delta_ {x}, \\delta_ {y} \\sim \\operatorname {B e t a} (\\beta (t), \\beta (t)), \\tag {6}", + "image_path": "eee4e30f79e07be7be07d35bc8f0c1b57742a326b595a837985079559a00b5bc.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 523, + 480, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 523, + 480, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 523, + 480, + 583 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 523, + 480, + 583 + ], + "type": "inline_equation", + "content": "\\delta_x" + }, + { + "bbox": [ + 130, + 523, + 480, + 583 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 523, + 480, + 583 + ], + "type": "inline_equation", + "content": "\\delta_y" + }, + { + "bbox": [ + 130, + 523, + 480, + 583 + ], + "type": "text", + "content": " denote the position offset in " + }, + { + "bbox": [ + 130, + 523, + 480, + 583 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 523, + 480, + 583 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 523, + 480, + 583 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 523, + 480, + 583 + ], + "type": "text", + "content": " directions respectively, while " + }, + { + "bbox": [ + 130, + 523, + 480, + 583 + ], + "type": "inline_equation", + "content": "\\beta(t)" + }, + { + "bbox": [ + 130, + 523, + 480, + 583 + ], + "type": "text", + "content": " are linearly annealed from " + }, + { + "bbox": [ + 130, + 523, + 480, + 583 + ], + "type": "inline_equation", + "content": "\\beta(0) = 1" + }, + { + "bbox": [ + 130, + 523, + 480, + 583 + ], + "type": "text", + "content": " to some final value " + }, + { + "bbox": [ + 130, + 523, + 480, + 583 + ], + "type": "inline_equation", + "content": "\\beta(T)" + }, + { + "bbox": [ + 130, + 523, + 480, + 583 + ], + "type": "text", + "content": " smaller than 1. By using the beta sampling strategy, we can maintain a more balanced distribution of training data that focuses more on the boundary patches, alleviating mode collapse issue and improving the overall training stability." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 600, + 230, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 230, + 613 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 230, + 613 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 624, + 190, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 190, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 190, + 635 + ], + "type": "text", + "content": "4.1 Setup" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 641, + 485, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 485, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 485, + 666 + ], + "type": "text", + "content": "Datasets. We evaluate our model on the NeRF-Synthetic benchmark dataset [31], which contains 8 synthetic objects with images taken from different viewpoints" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 101 + ], + "type": "text", + "content": "Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": "uniformly distributed in the hemisphere. Following the original setting, We hold 200 viewpoints for generating high-quality training data and 200 viewpoints for testing. Further, to demonstrate the generalization ability of our method, we also evaluate our method on complex real-world LLFF scenes [30] which consists of 8 scenes captured with roughly forward-facing images. We also demonstrate the superior performance of RaFE on real-world blur [27] and noise [29] data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 190, + 482, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 482, + 370 + ], + "type": "text", + "content": "Evaluation Metrics. Following the common practice of 3D reconstruction, we try to evaluate each method with two standard image quality metrics: peak signal-to-noise ratio (PSNR), structural similarity index (SSIM) [47], which however could not reflect the real 3D restoration performance according to our observations. Due to the generative characteristic of RaFE, the recovered radiance field from RaFE is very high-quality, but may not faithfully follow the \"ground-truth\" 3D, since inversing a degraded signal is a highly ill-posed problem. Moreover, we also found the baselines with better scores over PSNR and SSIM still pose a degraded appearance with smooth texture details, as shown in Figure 3b. Hence, a better way is to leverage perceptual metric: learned perceptual image patch similarity (LPIPS) [50] which computes the mean squared error (MSE) between normalized features from all layers of a pre-trained VGG [36] encoder and is deemed to better correlate with human perception. We also leverage the latest non-reference-based image quality assessment metrics including LIQE [51] and MANIQA [48] to demonstrate the superior rendering quality of our method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 372, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 372, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 372, + 482, + 456 + ], + "type": "text", + "content": "Implementation Details. We implement all the experiments by PyTorch. For the 2D generator and discriminator, we use a convolutional-based generator and discriminator used in StyleGAN2 [16]. In all experiments, we choose Adam optimizer for all the modules in our pipeline, with hyperparameters " + }, + { + "bbox": [ + 130, + 372, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0, \\beta_{2} = 0.99" + }, + { + "bbox": [ + 130, + 372, + 482, + 456 + ], + "type": "text", + "content": ". We use learning rate " + }, + { + "bbox": [ + 130, + 372, + 482, + 456 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-3}" + }, + { + "bbox": [ + 130, + 372, + 482, + 456 + ], + "type": "text", + "content": " for both generator and discriminator. For loss weights, we use " + }, + { + "bbox": [ + 130, + 372, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\lambda_{mimic} = 0.5" + }, + { + "bbox": [ + 130, + 372, + 482, + 456 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 372, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\lambda_{adv} = 1.0" + }, + { + "bbox": [ + 130, + 372, + 482, + 456 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 372, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\lambda_{rec} = 1.0" + }, + { + "bbox": [ + 130, + 372, + 482, + 456 + ], + "type": "text", + "content": " for almost all experiments. We evaluate RaFE framework on 4 different 3D restoration tasks:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 472, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 472, + 479, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 472, + 479, + 531 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 479, + 531 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 138, + 472, + 479, + 531 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 138, + 472, + 479, + 531 + ], + "type": "text", + "content": " Super-Resolution: On the blender dataset, we resize the image to " + }, + { + "bbox": [ + 138, + 472, + 479, + 531 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 138, + 472, + 479, + 531 + ], + "type": "text", + "content": " resolutions to get low-resolution images. As for LLFF data, we first center crop the training image to " + }, + { + "bbox": [ + 138, + 472, + 479, + 531 + ], + "type": "inline_equation", + "content": "188 \\times 252" + }, + { + "bbox": [ + 138, + 472, + 479, + 531 + ], + "type": "text", + "content": " to adapt to our " + }, + { + "bbox": [ + 138, + 472, + 479, + 531 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 138, + 472, + 479, + 531 + ], + "type": "text", + "content": " super-resolution task and then resize to " + }, + { + "bbox": [ + 138, + 472, + 479, + 531 + ], + "type": "inline_equation", + "content": "47 \\times 63" + }, + { + "bbox": [ + 138, + 472, + 479, + 531 + ], + "type": "text", + "content": " to get low-resolution images. And we use Deepfloyd [35] for 2D super-resolution." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 482, + 666 + ], + "type": "text", + "content": "- Deblur: On the LLFF dataset, we construct camera motion blur by applying the blur kernel following equation " + }, + { + "bbox": [ + 138, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "I_{blur} = A * I_{clear}" + }, + { + "bbox": [ + 138, + 533, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 138, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 138, + 533, + 482, + 666 + ], + "type": "text", + "content": " is the blur kernel and " + }, + { + "bbox": [ + 138, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 138, + 533, + 482, + 666 + ], + "type": "text", + "content": " stands for the convolution operator. Different from the blur dataset proposed in Deblur-NeRF [27], which contains blurred images with varying degrees of blur, and even includes a certain amount of high-resolution images, we apply a large blur kernel size (e.g. 13) and a more complex camera motion path to all the images in the dataset. We additionally construct consistent blur datasets by using the same blur kernel to the training image in a scene, which means the camera motion trajectory is the same for the training image set. The resolution of images is the same as the super-resolution task. We use HiDiff [10] to recover high-quality images." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 447, + 101 + ], + "type": "text", + "content": "RaFE: Generative Radiance Fields Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 115, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 115, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 138, + 115, + 482, + 187 + ], + "type": "text", + "content": "- Denoise: On the LLFF dataset, we follow the noise model used in [28, 33] and we get the noisy version of a clean image " + }, + { + "bbox": [ + 138, + 115, + 482, + 187 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 138, + 115, + 482, + 187 + ], + "type": "text", + "content": " according to equation " + }, + { + "bbox": [ + 138, + 115, + 482, + 187 + ], + "type": "inline_equation", + "content": "I_{\\text{noisy}}(x) = \\mathcal{N}(I(x), \\delta_r^2 + \\delta_s^2 I^2(x))" + }, + { + "bbox": [ + 138, + 115, + 482, + 187 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 138, + 115, + 482, + 187 + ], + "type": "inline_equation", + "content": "\\sigma_r" + }, + { + "bbox": [ + 138, + 115, + 482, + 187 + ], + "type": "text", + "content": " is the signal-independent read-noise parameter, " + }, + { + "bbox": [ + 138, + 115, + 482, + 187 + ], + "type": "inline_equation", + "content": "\\sigma_s" + }, + { + "bbox": [ + 138, + 115, + 482, + 187 + ], + "type": "text", + "content": " is the signal-dependent shot-noise. Following [28, 33], we use the Gain level to represent the noise strength. We use gain levels " + }, + { + "bbox": [ + 138, + 115, + 482, + 187 + ], + "type": "inline_equation", + "content": "= 8" + }, + { + "bbox": [ + 138, + 115, + 482, + 187 + ], + "type": "text", + "content": " to get our noisy image. We use DiffBIR [25] to get high-quality images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 188, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 188, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 138, + 188, + 482, + 260 + ], + "type": "text", + "content": "- Mixed degradation: The degradation pipeline consists of three stages: blur, noise, and JPEG compression. First, we utilize Gaussian blur with a radius of 7 for the blender dataset and 3 for the LLFF dataset. Second, we add noise with std 25. And then, we apply JPEG compression. The quality of JPEG compression is 50. The resolution of images is the same as the super-resolution task. We use DiffBIR [25] to get high-quality images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 276, + 198, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 276, + 198, + 286 + ], + "spans": [ + { + "bbox": [ + 133, + 276, + 198, + 286 + ], + "type": "text", + "content": "4.2 Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 294, + 481, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 481, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 481, + 366 + ], + "type": "text", + "content": "Baseline Methods. For general tasks, we try a baseline that firstly restores the degraded images and uses the restored high-quality images to reconstruct a NeRF directly, denoted as NeRF-Perframe. We also use the 2D-based restoration model SwinIR [24] to do the per-view refinement for the renderings of NeRF trained by degraded image, denoted as NeRF-SwinIR. Note that we do not evaluate NeRF-SwinIR for the deblur task since there are no corresponding checkpoints." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 366, + 481, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 481, + 521 + ], + "type": "text", + "content": "To more thoroughly test the effectiveness of our method, we also select some task-specific competitors. For super-resolution task, we choose NeRF-SR [43], Neural Volume Super-Resolution (NVSR) [1] as baselines. For mixed degradation, since there is no existing method tailored for mixed degradation, we choose NeRFLiX, which tries to solve the NeRF-like degradation by training a 2D refinement model using degradation images constructed by a degradation simulator for typical NeRF-style artifacts. We consider this to be the most relevant method. For the deblur task, we compare with two state-of-the-art methods Deblur-NeRF [27] and BAD-NeRF [45]. Deblur-NeRF designs a learnable blur kernel and applies it to rays to simulate the degrading process and BAD-NeRF directly models the camera trajectories to solve motion blur. For the denoising task, we compare with NAN [33], which uses a noise-aware encoder to aggregate the feature of multi-view images for restoration." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 522, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 481, + 666 + ], + "type": "text", + "content": "Quantitative Results. We conduct extensive quantitative comparisons with various baselines across different restoration tasks in Tab. 1a for super-resolution, Tab. 2b and Tab. 2a for deblurring, Tab. 2c for denoising and Tab. 1b for mixed degradation. As analyzed before, although most of the time our method falls slightly behind on the reconstruction metrics like PSNR and SSIM compared with other baselines, which only measure local pixel-aligned similarity between the rendered novel views and the ground truth images, are less indicative since uncertainties naturally exist in generative procedure. Taking the super-resolution results on Blender data as an example (Tab. 1a), the simplest baseline NeRF-Perframe has already achieved the best reconstruction metrics, but as shown in Figure. 3a, its visual quality is vastly inferior compared with our results. Through the error map, we found the misalignment between the generated 3D and input" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Wuet al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 114, + 471, + 152 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 471, + 152 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 471, + 152 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 471, + 152 + ], + "type": "table", + "html": "
BleederLLFFLLFF
PSNR†SSIM†LPiPS‡LIPE†MANIQA†PSNR†SSIM†LPiPS‡LIPE†MANIQA†MethodPSNR†SSIM†LPiPS‡LIPE†MANIQA†PSNR†SSIM†LPiPS‡LIPE†MANIQA†
NevRF-SR26.850.9120.1352.3720.36622.650.7020.3271.3200.220
NVSR26.130.8790.1322.3010.34321.290.6060.4422.1080.177
NeRF-SwnIR24.070.8780.1193.060.38121.730.6550.3651.8080.267
NeRF-Perframe27.390.9220.0833.2760.38623.680.7450.2611.4170.257
Ours24.990.9010.0624.6210.54323.850.7520.1972.3970.322
", + "image_path": "c9a0709fa184a3568877df92199e5fb4bfc74c60b33df35fd922378921f6f1c4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 133, + 226, + 246, + 266 + ], + "blocks": [ + { + "bbox": [ + 181, + 154, + 253, + 163 + ], + "lines": [ + { + "bbox": [ + 181, + 154, + 253, + 163 + ], + "spans": [ + { + "bbox": [ + 181, + 154, + 253, + 163 + ], + "type": "text", + "content": "(a) Super-resolution" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 345, + 153, + 427, + 163 + ], + "lines": [ + { + "bbox": [ + 345, + 153, + 427, + 163 + ], + "spans": [ + { + "bbox": [ + 345, + 153, + 427, + 163 + ], + "type": "text", + "content": "(b) Mixed degradation" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 172, + 480, + 206 + ], + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 206 + ], + "type": "text", + "content": "Table 1: Quantitative comparisons on super-resolution and mixed degradation tasks. The best result without using reference is highlighted. Our method achieves great superiorities on perceptual metrics and image quality when compared with others." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 226, + 246, + 266 + ], + "lines": [ + { + "bbox": [ + 133, + 226, + 246, + 266 + ], + "spans": [ + { + "bbox": [ + 133, + 226, + 246, + 266 + ], + "type": "table", + "html": "
MethodLLFF
PSNR†SSIM†LPIPS‡LIPE†MANIQA†
Deblur-NeRF23.750.7990.3071.3650.157
BAD-NeRF24.0270.7880.3131.0940.151
NeRF-Perframe21.020.6950.3620.3620.150
Ours23.230.8110.2941.1440.177
", + "image_path": "7625f6936d77b80ea53cf1b5296c30edc72cd1744ed36e8e37df8d304aa9f5ce.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 247, + 227, + 359, + 266 + ], + "blocks": [ + { + "bbox": [ + 163, + 268, + 216, + 277 + ], + "lines": [ + { + "bbox": [ + 163, + 268, + 216, + 277 + ], + "spans": [ + { + "bbox": [ + 163, + 268, + 216, + 277 + ], + "type": "text", + "content": "(a) Deblurring" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 247, + 227, + 359, + 266 + ], + "lines": [ + { + "bbox": [ + 247, + 227, + 359, + 266 + ], + "spans": [ + { + "bbox": [ + 247, + 227, + 359, + 266 + ], + "type": "table", + "html": "
MethodLLFF
PSNR↑SSIM↑LPIPS↓LIPE↑MANIQA↑
Debhr-NeRF21.710.7490.3331.1040.122
BAD-NeRF25.400.8360.2781.1400.173
NeRF-Perframe21.570.7710.3101.1030.199
Ours22.930.7890.2521.1430.224
", + "image_path": "f2f4f5f8ddd082ab92bf220695e8dcec77dcfbe55213a6cc1c2643d4fab81459.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 361, + 227, + 473, + 266 + ], + "blocks": [ + { + "bbox": [ + 248, + 268, + 358, + 277 + ], + "lines": [ + { + "bbox": [ + 248, + 268, + 358, + 277 + ], + "spans": [ + { + "bbox": [ + 248, + 268, + 358, + 277 + ], + "type": "text", + "content": "(b) Deblurring(consistent blur)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 361, + 227, + 473, + 266 + ], + "lines": [ + { + "bbox": [ + 361, + 227, + 473, + 266 + ], + "spans": [ + { + "bbox": [ + 361, + 227, + 473, + 266 + ], + "type": "table", + "html": "
MethodLLFF
PSNR↑SSIM↑LPIPS↓LIPE↑MANIQA↑
NAN25.990.8220.32081.30320.241
NeRF-SwinIR24.370.7780.3461.5790.210
NeRF-Perframe23.660.7860.2811.0950.214
Ours23.780.7910.25611.6070.257
", + "image_path": "245218e2c283968f4b4640daf71d44ac73b357f06d7eda544143e697eede65e9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 370, + 268, + 463, + 277 + ], + "lines": [ + { + "bbox": [ + 370, + 268, + 463, + 277 + ], + "spans": [ + { + "bbox": [ + 370, + 268, + 463, + 277 + ], + "type": "text", + "content": "(c) Denoising with Gain 8" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 285, + 480, + 308 + ], + "lines": [ + { + "bbox": [ + 132, + 285, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 480, + 308 + ], + "type": "text", + "content": "Table 2: Quantitative comparisons for deblurring and denoising. The best result without using reference is highlighted. Our method performs the best on perceptual metrics." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 342, + 480, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 480, + 389 + ], + "type": "text", + "content": "3D causes the drop of PSNR and SSIM. By contrast, on the perceptual metrics like LPIPS and non-reference-based metrics including LIPE and MANIQA, which could more effectively reflect the restoration performance, our method consistently achieves better results when compared with other baselines." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 390, + 481, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 481, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 481, + 581 + ], + "type": "text", + "content": "For mixed degradation tasks, the best results for LPIPS metrics are achieved by NeRFLix w. ref [53]. This is because NeRFLiX can see two high-quality ground-truth images from the nearest two viewpoints when inference, which leads to information leakage. However, high-quality ground-truth information is not accessed in our setting or in any real-world cases. After eliminating the impact of ground truth (NeRFLiX w/o. ref) by replacing the reference ground truth images with degraded images, our method performs better than NeRFLiX. Qualitative Results. To further verify the restoration capability, we present visual results for different degradations in Fig. 4. Our method is able to generate realistic details while other methods tend to generate smooth results which lack high-quality details. For example, for the super-resolution task, we show the drum restoration for each method, as can be seen in Fig. 4, our method generates high-fidelity drum surfaces with sharp edges, while other methods suffer from severe blur on the drum surfaces and edges. For mixed degradation tasks, our method successfully restores the intricate golden textures on the chair surface while other methods only rendered extremely blurred texture." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Geometry Refinement. Compared with previous NeRF restoration methods, most of which focus on using 2D refinement to resolve frame-wise defects, our method firstly leverages the general priors of large foundation models and GAN, achieving the open-domain 3D-based restoration. Thus, our restoration algorithm could not only recover the high-quality and view-consistent images, but also refine the 3D geometry. We demonstrate this advantage using the Ficus data of Blender, as shown in Fig. 5a, through comparing the extracted mesh from the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 447, + 101 + ], + "type": "text", + "content": "RaFE: Generative Radiance Fields Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 118, + 176, + 177 + ], + "blocks": [ + { + "bbox": [ + 143, + 118, + 176, + 177 + ], + "lines": [ + { + "bbox": [ + 143, + 118, + 176, + 177 + ], + "spans": [ + { + "bbox": [ + 143, + 118, + 176, + 177 + ], + "type": "image", + "image_path": "125f23c222781059bc595eedaea13010d055178e7ebcac57c84819d2ae1a1405.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 182, + 165, + 190 + ], + "lines": [ + { + "bbox": [ + 154, + 182, + 165, + 190 + ], + "spans": [ + { + "bbox": [ + 154, + 182, + 165, + 190 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 185, + 118, + 219, + 181 + ], + "blocks": [ + { + "bbox": [ + 185, + 118, + 219, + 181 + ], + "lines": [ + { + "bbox": [ + 185, + 118, + 219, + 181 + ], + "spans": [ + { + "bbox": [ + 185, + 118, + 219, + 181 + ], + "type": "image", + "image_path": "4fb27322685fb0d02ed831c9e62001a4fae8672e70366101e354d56f18f0ff05.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 182, + 215, + 190 + ], + "lines": [ + { + "bbox": [ + 187, + 182, + 215, + 190 + ], + "spans": [ + { + "bbox": [ + 187, + 182, + 215, + 190 + ], + "type": "text", + "content": "Perframe" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 200, + 193, + 212, + 201 + ], + "lines": [ + { + "bbox": [ + 200, + 193, + 212, + 201 + ], + "spans": [ + { + "bbox": [ + 200, + 193, + 212, + 201 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 220, + 119, + 253, + 181 + ], + "blocks": [ + { + "bbox": [ + 220, + 119, + 253, + 181 + ], + "lines": [ + { + "bbox": [ + 220, + 119, + 253, + 181 + ], + "spans": [ + { + "bbox": [ + 220, + 119, + 253, + 181 + ], + "type": "image", + "image_path": "ae7b5738179baad23c5d8a99885c6f85f119e5aa35df0cd0b3570b743304d515.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 182, + 245, + 190 + ], + "lines": [ + { + "bbox": [ + 230, + 182, + 245, + 190 + ], + "spans": [ + { + "bbox": [ + 230, + 182, + 245, + 190 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 321, + 144, + 350, + 178 + ], + "blocks": [ + { + "bbox": [ + 321, + 144, + 350, + 178 + ], + "lines": [ + { + "bbox": [ + 321, + 144, + 350, + 178 + ], + "spans": [ + { + "bbox": [ + 321, + 144, + 350, + 178 + ], + "type": "image", + "image_path": "a64e405d9bcb6565f98ca273acb8570683d6b5615f5d1573d821fa7e94289239.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 182, + 342, + 188 + ], + "lines": [ + { + "bbox": [ + 332, + 182, + 342, + 188 + ], + "spans": [ + { + "bbox": [ + 332, + 182, + 342, + 188 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 210, + 480, + 232 + ], + "lines": [ + { + "bbox": [ + 131, + 210, + 480, + 232 + ], + "spans": [ + { + "bbox": [ + 131, + 210, + 480, + 232 + ], + "type": "text", + "content": "Fig. 3: (a) Error map between ground truth and our method/NeRF-Perframe. (b) we showcase that the visual quality can be much better even with lower PSNR scores." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 362, + 117, + 396, + 178 + ], + "blocks": [ + { + "bbox": [ + 362, + 117, + 396, + 178 + ], + "lines": [ + { + "bbox": [ + 362, + 117, + 396, + 178 + ], + "spans": [ + { + "bbox": [ + 362, + 117, + 396, + 178 + ], + "type": "image", + "image_path": "d159837477fa8b4e28a84abad3508150a2487fc4e1db10d9a2fbfa28d147f1a6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 379, + 181, + 393, + 188 + ], + "lines": [ + { + "bbox": [ + 379, + 181, + 393, + 188 + ], + "spans": [ + { + "bbox": [ + 379, + 181, + 393, + 188 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 383, + 192, + 396, + 201 + ], + "lines": [ + { + "bbox": [ + 383, + 192, + 396, + 201 + ], + "spans": [ + { + "bbox": [ + 383, + 192, + 396, + 201 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 399, + 117, + 462, + 178 + ], + "blocks": [ + { + "bbox": [ + 399, + 117, + 462, + 178 + ], + "lines": [ + { + "bbox": [ + 399, + 117, + 462, + 178 + ], + "spans": [ + { + "bbox": [ + 399, + 117, + 462, + 178 + ], + "type": "image", + "image_path": "8db7cda11d49d7b836014492629f658286453844e06275227b4437fedc0ea210.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 422, + 181, + 451, + 188 + ], + "lines": [ + { + "bbox": [ + 422, + 181, + 451, + 188 + ], + "spans": [ + { + "bbox": [ + 422, + 181, + 451, + 188 + ], + "type": "text", + "content": "NeRF-SR" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 257, + 479, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 479, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 479, + 281 + ], + "type": "text", + "content": "restored NeRF models, our method effectively learns the restoration in the 3D tri-plane space and recovers better geometry." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 300, + 237, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 237, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 237, + 312 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 320, + 481, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 481, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 481, + 452 + ], + "type": "text", + "content": "Effects of different restoration models. To investigate the influence of using different 2D restoration models, we tested two additional off-the-shelf restoration models for the super-resolution task, including diffusion-based DiffBIR [25] and non-diffusion-based SRFormer [54]. As shown in Fig. 5b, the diffusion-based model DiffBIR shows larger restoration diversity over SRFormer by measuring the diversity score. When the repaired images exhibit diversity, direct reconstruction inevitably leads to blurriness to varying degrees due to the existing multi-view inconsistency. Through modeling the distribution of the potential high-quality NeRFs, our method accommodates these inconsistencies and consistently achieves better performance over NeRF-Perframe, demonstrating the great generalization capability of RaFE to different 2D restoration models." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 453, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 548 + ], + "type": "text", + "content": "Effects of the generator. In this ablation study, we examine the influence of the generator by comparing it with the baseline that directly optimizes the NeRF parameters using GAN loss and LPIPS mentioned above on the Blender dataset. As we can observe in Fig. 6, the image rendered by generative NeRF exhibits varied fine-textured details under different random code " + }, + { + "bbox": [ + 130, + 453, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 453, + 482, + 548 + ], + "type": "text", + "content": ". Once the generator is removed, the rendered images will contain a blurry and smooth appearance, showing the importance of using the generator to model the distribution, which can also be demonstrated by the numerical metrics in the right part of Fig. 6." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 567, + 225, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 225, + 580 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 225, + 580 + ], + "type": "text", + "content": "5 Conclusions" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "This paper proposes a novel generic NeRF restoration method that applies to various types of degradations, such as low resolution, blurriness, noise, and mixed degradation. The proposed method leverages the off-the-shelf image restoration methods to restore the multi-view input images individually. To tackle the geometric and appearance inconsistencies presented in multi-view images due to individual restoration, we propose to train a GAN for NeRF generation, where" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 126, + 195, + 183 + ], + "blocks": [ + { + "bbox": [ + 137, + 126, + 195, + 183 + ], + "lines": [ + { + "bbox": [ + 137, + 126, + 195, + 183 + ], + "spans": [ + { + "bbox": [ + 137, + 126, + 195, + 183 + ], + "type": "image", + "image_path": "a0f86ebb671cbff4cf90754998e0ade07ffe51711e143e64f23ef30bcd37d54b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 196, + 126, + 251, + 183 + ], + "blocks": [ + { + "bbox": [ + 196, + 126, + 251, + 183 + ], + "lines": [ + { + "bbox": [ + 196, + 126, + 251, + 183 + ], + "spans": [ + { + "bbox": [ + 196, + 126, + 251, + 183 + ], + "type": "image", + "image_path": "d8e1ded7c54cb6468732d67f90c87b48af9310694f04558c95be0524de092366.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 252, + 126, + 308, + 184 + ], + "blocks": [ + { + "bbox": [ + 252, + 126, + 308, + 184 + ], + "lines": [ + { + "bbox": [ + 252, + 126, + 308, + 184 + ], + "spans": [ + { + "bbox": [ + 252, + 126, + 308, + 184 + ], + "type": "image", + "image_path": "c17ac47c480de50c8dfded832d4b7848515dce9b0170084527e7336e876ea0cf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 126, + 364, + 183 + ], + "blocks": [ + { + "bbox": [ + 310, + 126, + 364, + 183 + ], + "lines": [ + { + "bbox": [ + 310, + 126, + 364, + 183 + ], + "spans": [ + { + "bbox": [ + 310, + 126, + 364, + 183 + ], + "type": "image", + "image_path": "13bf006035cd16f100b8c3b89289d50fe112a5703727cce4cc9f7d2a764b3c70.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 365, + 126, + 421, + 183 + ], + "blocks": [ + { + "bbox": [ + 365, + 126, + 421, + 183 + ], + "lines": [ + { + "bbox": [ + 365, + 126, + 421, + 183 + ], + "spans": [ + { + "bbox": [ + 365, + 126, + 421, + 183 + ], + "type": "image", + "image_path": "911a6932edaaf60e960934bd95848eadce2f014775a105e683d74f4f31ccc3e9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 422, + 126, + 477, + 183 + ], + "blocks": [ + { + "bbox": [ + 422, + 126, + 477, + 183 + ], + "lines": [ + { + "bbox": [ + 422, + 126, + 477, + 183 + ], + "spans": [ + { + "bbox": [ + 422, + 126, + 477, + 183 + ], + "type": "image", + "image_path": "35ef8f558319326a2bfb1f2246ebefe2d16ecc02176799cd64c83aa6d31cbfe5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 137, + 193, + 194, + 236 + ], + "blocks": [ + { + "bbox": [ + 137, + 193, + 194, + 236 + ], + "lines": [ + { + "bbox": [ + 137, + 193, + 194, + 236 + ], + "spans": [ + { + "bbox": [ + 137, + 193, + 194, + 236 + ], + "type": "image", + "image_path": "3cc06a44f37365d2e012ead3fdc9ebf408275baf97c017666a5bcf9af1d8292b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 238, + 176, + 244 + ], + "lines": [ + { + "bbox": [ + 156, + 238, + 176, + 244 + ], + "spans": [ + { + "bbox": [ + 156, + 238, + 176, + 244 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 195, + 193, + 251, + 236 + ], + "blocks": [ + { + "bbox": [ + 195, + 193, + 251, + 236 + ], + "lines": [ + { + "bbox": [ + 195, + 193, + 251, + 236 + ], + "spans": [ + { + "bbox": [ + 195, + 193, + 251, + 236 + ], + "type": "image", + "image_path": "28acb84c0328be37bc399141a94bd93f3c37c784e9cc537d1ada7eabd305954b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 238, + 236, + 243 + ], + "lines": [ + { + "bbox": [ + 208, + 238, + 236, + 243 + ], + "spans": [ + { + "bbox": [ + 208, + 238, + 236, + 243 + ], + "type": "text", + "content": "(b) NeRF-SR" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 252, + 193, + 307, + 236 + ], + "blocks": [ + { + "bbox": [ + 252, + 193, + 307, + 236 + ], + "lines": [ + { + "bbox": [ + 252, + 193, + 307, + 236 + ], + "spans": [ + { + "bbox": [ + 252, + 193, + 307, + 236 + ], + "type": "image", + "image_path": "f7d479743418b09c805deeb021d3c03f90a4021a2583ba28da44482b81b6d29a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 268, + 238, + 290, + 243 + ], + "lines": [ + { + "bbox": [ + 268, + 238, + 290, + 243 + ], + "spans": [ + { + "bbox": [ + 268, + 238, + 290, + 243 + ], + "type": "text", + "content": "(c) NVSR" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 308, + 193, + 364, + 236 + ], + "blocks": [ + { + "bbox": [ + 308, + 193, + 364, + 236 + ], + "lines": [ + { + "bbox": [ + 308, + 193, + 364, + 236 + ], + "spans": [ + { + "bbox": [ + 308, + 193, + 364, + 236 + ], + "type": "image", + "image_path": "98e21da0fcb3a5f87de597e418ea26c35ecfc93c6f045d52bcf2df70625498ff.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 319, + 238, + 357, + 243 + ], + "lines": [ + { + "bbox": [ + 319, + 238, + 357, + 243 + ], + "spans": [ + { + "bbox": [ + 319, + 238, + 357, + 243 + ], + "type": "text", + "content": "(d) NeRF-SwinIR" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 364, + 193, + 420, + 236 + ], + "blocks": [ + { + "bbox": [ + 364, + 193, + 420, + 236 + ], + "lines": [ + { + "bbox": [ + 364, + 193, + 420, + 236 + ], + "spans": [ + { + "bbox": [ + 364, + 193, + 420, + 236 + ], + "type": "image", + "image_path": "8b2bb7880364077387a24a719eb58bccffc16b84e1ab7f25824f3f3845a012dd.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 238, + 414, + 243 + ], + "lines": [ + { + "bbox": [ + 373, + 238, + 414, + 243 + ], + "spans": [ + { + "bbox": [ + 373, + 238, + 414, + 243 + ], + "type": "text", + "content": "(e) NeRF-Perframe" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 422, + 193, + 477, + 236 + ], + "blocks": [ + { + "bbox": [ + 422, + 193, + 477, + 236 + ], + "lines": [ + { + "bbox": [ + 422, + 193, + 477, + 236 + ], + "spans": [ + { + "bbox": [ + 422, + 193, + 477, + 236 + ], + "type": "image", + "image_path": "abb39be8fe5ecc748691484efff0dac8454ccbd0b0b548855a175198eaf7479f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 441, + 238, + 459, + 243 + ], + "lines": [ + { + "bbox": [ + 441, + 238, + 459, + 243 + ], + "spans": [ + { + "bbox": [ + 441, + 238, + 459, + 243 + ], + "type": "text", + "content": "(f) Ours" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 137, + 258, + 203, + 310 + ], + "blocks": [ + { + "bbox": [ + 137, + 258, + 203, + 310 + ], + "lines": [ + { + "bbox": [ + 137, + 258, + 203, + 310 + ], + "spans": [ + { + "bbox": [ + 137, + 258, + 203, + 310 + ], + "type": "image", + "image_path": "1594ac270108d3833ebac6fb052496e01f03be459c5932d6c1d2b81b64011382.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 313, + 178, + 320 + ], + "lines": [ + { + "bbox": [ + 159, + 313, + 178, + 320 + ], + "spans": [ + { + "bbox": [ + 159, + 313, + 178, + 320 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 205, + 258, + 271, + 310 + ], + "blocks": [ + { + "bbox": [ + 205, + 258, + 271, + 310 + ], + "lines": [ + { + "bbox": [ + 205, + 258, + 271, + 310 + ], + "spans": [ + { + "bbox": [ + 205, + 258, + 271, + 310 + ], + "type": "image", + "image_path": "7e4dfcc2a8e6244ad57aead5ed03ac60bf85a6f27ff5adc11f5a4a608c76c353.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 312, + 257, + 319 + ], + "lines": [ + { + "bbox": [ + 220, + 312, + 257, + 319 + ], + "spans": [ + { + "bbox": [ + 220, + 312, + 257, + 319 + ], + "type": "text", + "content": "(b) Deblur-NeRF" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 272, + 258, + 339, + 310 + ], + "blocks": [ + { + "bbox": [ + 272, + 258, + 339, + 310 + ], + "lines": [ + { + "bbox": [ + 272, + 258, + 339, + 310 + ], + "spans": [ + { + "bbox": [ + 272, + 258, + 339, + 310 + ], + "type": "image", + "image_path": "1e5017fc02880138eb92352ff014f0f0a9b8e64be4af6ad42dc060bd244ac411.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 287, + 313, + 319, + 319 + ], + "lines": [ + { + "bbox": [ + 287, + 313, + 319, + 319 + ], + "spans": [ + { + "bbox": [ + 287, + 313, + 319, + 319 + ], + "type": "text", + "content": "(c) BAD-NeRF" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 339, + 258, + 406, + 310 + ], + "blocks": [ + { + "bbox": [ + 288, + 246, + 328, + 252 + ], + "lines": [ + { + "bbox": [ + 288, + 246, + 328, + 252 + ], + "spans": [ + { + "bbox": [ + 288, + 246, + 328, + 252 + ], + "type": "text", + "content": "Super-resolution" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 339, + 258, + 406, + 310 + ], + "lines": [ + { + "bbox": [ + 339, + 258, + 406, + 310 + ], + "spans": [ + { + "bbox": [ + 339, + 258, + 406, + 310 + ], + "type": "image", + "image_path": "7bd334d5a22bb26792f48ec7dc3fce40df7cb4d8738b8cf98d7b2ea9bb74ee84.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 313, + 395, + 319 + ], + "lines": [ + { + "bbox": [ + 356, + 313, + 395, + 319 + ], + "spans": [ + { + "bbox": [ + 356, + 313, + 395, + 319 + ], + "type": "text", + "content": "(b)NeRF-Perframe" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 407, + 258, + 474, + 310 + ], + "blocks": [ + { + "bbox": [ + 407, + 258, + 474, + 310 + ], + "lines": [ + { + "bbox": [ + 407, + 258, + 474, + 310 + ], + "spans": [ + { + "bbox": [ + 407, + 258, + 474, + 310 + ], + "type": "image", + "image_path": "9c0e71139d2b51146cc179012e9f061ed74fc7288eb411b3db401d21a4e075f0.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 437, + 313, + 454, + 319 + ], + "lines": [ + { + "bbox": [ + 437, + 313, + 454, + 319 + ], + "spans": [ + { + "bbox": [ + 437, + 313, + 454, + 319 + ], + "type": "text", + "content": "(e) ours" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 138, + 332, + 204, + 383 + ], + "blocks": [ + { + "bbox": [ + 138, + 332, + 204, + 383 + ], + "lines": [ + { + "bbox": [ + 138, + 332, + 204, + 383 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 204, + 383 + ], + "type": "image", + "image_path": "3f1d6a3eb2977f569a43f14b0b359d2e2bb1221aa20155a4b8ce3a6ad379b397.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 385, + 176, + 392 + ], + "lines": [ + { + "bbox": [ + 157, + 385, + 176, + 392 + ], + "spans": [ + { + "bbox": [ + 157, + 385, + 176, + 392 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 205, + 332, + 271, + 383 + ], + "blocks": [ + { + "bbox": [ + 205, + 332, + 271, + 383 + ], + "lines": [ + { + "bbox": [ + 205, + 332, + 271, + 383 + ], + "spans": [ + { + "bbox": [ + 205, + 332, + 271, + 383 + ], + "type": "image", + "image_path": "d017ae851f9be427de08ae8cbd18290c7549858df7b18f8785c2faf06a984efe.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 384, + 256, + 391 + ], + "lines": [ + { + "bbox": [ + 220, + 384, + 256, + 391 + ], + "spans": [ + { + "bbox": [ + 220, + 384, + 256, + 391 + ], + "type": "text", + "content": "(b) Deblur-NeRF" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 272, + 332, + 339, + 383 + ], + "blocks": [ + { + "bbox": [ + 278, + 323, + 343, + 328 + ], + "lines": [ + { + "bbox": [ + 278, + 323, + 343, + 328 + ], + "spans": [ + { + "bbox": [ + 278, + 323, + 343, + 328 + ], + "type": "text", + "content": "Inconsistent Motion Deblur" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 272, + 332, + 339, + 383 + ], + "lines": [ + { + "bbox": [ + 272, + 332, + 339, + 383 + ], + "spans": [ + { + "bbox": [ + 272, + 332, + 339, + 383 + ], + "type": "image", + "image_path": "e88d6f5bfd51e822affb7c15a1456a49f80fb2dff9329fd9d6f1c25b1eb8b0f6.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 283, + 384, + 316, + 390 + ], + "lines": [ + { + "bbox": [ + 283, + 384, + 316, + 390 + ], + "spans": [ + { + "bbox": [ + 283, + 384, + 316, + 390 + ], + "type": "text", + "content": "(c) BAD-NeRF" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 339, + 332, + 406, + 383 + ], + "blocks": [ + { + "bbox": [ + 339, + 332, + 406, + 383 + ], + "lines": [ + { + "bbox": [ + 339, + 332, + 406, + 383 + ], + "spans": [ + { + "bbox": [ + 339, + 332, + 406, + 383 + ], + "type": "image", + "image_path": "f8a167f4bb04e0f54a25cbef71d6a2555fff3b68b4a88f5f136233a70cb4b398.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 384, + 394, + 391 + ], + "lines": [ + { + "bbox": [ + 354, + 384, + 394, + 391 + ], + "spans": [ + { + "bbox": [ + 354, + 384, + 394, + 391 + ], + "type": "text", + "content": "(b)NeRF-Perframe" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 408, + 332, + 475, + 383 + ], + "blocks": [ + { + "bbox": [ + 408, + 332, + 475, + 383 + ], + "lines": [ + { + "bbox": [ + 408, + 332, + 475, + 383 + ], + "spans": [ + { + "bbox": [ + 408, + 332, + 475, + 383 + ], + "type": "image", + "image_path": "e803680c6d14ef2f6955c239b74e836f16191fd3551d1caed6d366355cf34ead.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 384, + 452, + 391 + ], + "lines": [ + { + "bbox": [ + 436, + 384, + 452, + 391 + ], + "spans": [ + { + "bbox": [ + 436, + 384, + 452, + 391 + ], + "type": "text", + "content": "(e) ours" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 137, + 403, + 204, + 453 + ], + "blocks": [ + { + "bbox": [ + 137, + 403, + 204, + 453 + ], + "lines": [ + { + "bbox": [ + 137, + 403, + 204, + 453 + ], + "spans": [ + { + "bbox": [ + 137, + 403, + 204, + 453 + ], + "type": "image", + "image_path": "6d4017c6f82e9a244871b873e5b5f55f81ced7b302bd7b3fa1809623bc829193.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 456, + 177, + 462 + ], + "lines": [ + { + "bbox": [ + 159, + 456, + 177, + 462 + ], + "spans": [ + { + "bbox": [ + 159, + 456, + 177, + 462 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 205, + 403, + 271, + 453 + ], + "blocks": [ + { + "bbox": [ + 205, + 403, + 271, + 453 + ], + "lines": [ + { + "bbox": [ + 205, + 403, + 271, + 453 + ], + "spans": [ + { + "bbox": [ + 205, + 403, + 271, + 453 + ], + "type": "image", + "image_path": "576d9f561108186776eed629b11dd3f7d7738d47d1f74bf063dd158067df36ff.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 456, + 246, + 461 + ], + "lines": [ + { + "bbox": [ + 227, + 456, + 246, + 461 + ], + "spans": [ + { + "bbox": [ + 227, + 456, + 246, + 461 + ], + "type": "text", + "content": "(b) NAN" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 272, + 403, + 339, + 453 + ], + "blocks": [ + { + "bbox": [ + 278, + 393, + 337, + 399 + ], + "lines": [ + { + "bbox": [ + 278, + 393, + 337, + 399 + ], + "spans": [ + { + "bbox": [ + 278, + 393, + 337, + 399 + ], + "type": "text", + "content": "Consistent Motion Deblur" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 272, + 403, + 339, + 453 + ], + "lines": [ + { + "bbox": [ + 272, + 403, + 339, + 453 + ], + "spans": [ + { + "bbox": [ + 272, + 403, + 339, + 453 + ], + "type": "image", + "image_path": "b820cf0acfe28c5c28a0590109a68a998b76746e4180566cf5d30823ac74cf84.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 289, + 456, + 327, + 461 + ], + "lines": [ + { + "bbox": [ + 289, + 456, + 327, + 461 + ], + "spans": [ + { + "bbox": [ + 289, + 456, + 327, + 461 + ], + "type": "text", + "content": "(b) NeRF-SwinIR" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 339, + 403, + 405, + 453 + ], + "blocks": [ + { + "bbox": [ + 339, + 403, + 405, + 453 + ], + "lines": [ + { + "bbox": [ + 339, + 403, + 405, + 453 + ], + "spans": [ + { + "bbox": [ + 339, + 403, + 405, + 453 + ], + "type": "image", + "image_path": "d3e723c5b981640f8a68b34f66616ed1f9e23c7d53bdf2e57da994057eecc1b4.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 456, + 395, + 461 + ], + "lines": [ + { + "bbox": [ + 354, + 456, + 395, + 461 + ], + "spans": [ + { + "bbox": [ + 354, + 456, + 395, + 461 + ], + "type": "text", + "content": "(c) NeRF-Perframe" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 407, + 403, + 474, + 453 + ], + "blocks": [ + { + "bbox": [ + 407, + 403, + 474, + 453 + ], + "lines": [ + { + "bbox": [ + 407, + 403, + 474, + 453 + ], + "spans": [ + { + "bbox": [ + 407, + 403, + 474, + 453 + ], + "type": "image", + "image_path": "5731ea98f68847a46743ed17894e068f7a14c66126025f60eec937a8b5199feb.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 456, + 446, + 462 + ], + "lines": [ + { + "bbox": [ + 429, + 456, + 446, + 462 + ], + "spans": [ + { + "bbox": [ + 429, + 456, + 446, + 462 + ], + "type": "text", + "content": "(d) ours" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_caption" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 137, + 471, + 194, + 530 + ], + "blocks": [ + { + "bbox": [ + 137, + 471, + 194, + 530 + ], + "lines": [ + { + "bbox": [ + 137, + 471, + 194, + 530 + ], + "spans": [ + { + "bbox": [ + 137, + 471, + 194, + 530 + ], + "type": "image", + "image_path": "72ee40d01719d85eaaf89f3d1605e7bf4269853ea204384584fcad68a9c5a810.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 194, + 472, + 250, + 530 + ], + "blocks": [ + { + "bbox": [ + 194, + 472, + 250, + 530 + ], + "lines": [ + { + "bbox": [ + 194, + 472, + 250, + 530 + ], + "spans": [ + { + "bbox": [ + 194, + 472, + 250, + 530 + ], + "type": "image", + "image_path": "8880675a671b5a78cfa68d1a31b542c496e0f0965c7351444692a041ed0c9e76.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 250, + 472, + 306, + 530 + ], + "blocks": [ + { + "bbox": [ + 250, + 472, + 306, + 530 + ], + "lines": [ + { + "bbox": [ + 250, + 472, + 306, + 530 + ], + "spans": [ + { + "bbox": [ + 250, + 472, + 306, + 530 + ], + "type": "image", + "image_path": "976fc9f1aa14575ecb4aff242d6f6d79d5d65dc1df58e2f114656a26736dacd2.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 306, + 472, + 361, + 530 + ], + "blocks": [ + { + "bbox": [ + 306, + 472, + 361, + 530 + ], + "lines": [ + { + "bbox": [ + 306, + 472, + 361, + 530 + ], + "spans": [ + { + "bbox": [ + 306, + 472, + 361, + 530 + ], + "type": "image", + "image_path": "4064ad80d76071bb260ebe49be1520946845fab17a0119d79677bb6733b0586f.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 361, + 472, + 416, + 530 + ], + "blocks": [ + { + "bbox": [ + 361, + 472, + 416, + 530 + ], + "lines": [ + { + "bbox": [ + 361, + 472, + 416, + 530 + ], + "spans": [ + { + "bbox": [ + 361, + 472, + 416, + 530 + ], + "type": "image", + "image_path": "68dee7be4e7bfda9e21d06c4cbf40717ecef00e944b2ffce5fea1ab6c470e10b.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 416, + 472, + 473, + 582 + ], + "blocks": [ + { + "bbox": [ + 416, + 472, + 473, + 582 + ], + "lines": [ + { + "bbox": [ + 416, + 472, + 473, + 582 + ], + "spans": [ + { + "bbox": [ + 416, + 472, + 473, + 582 + ], + "type": "image", + "image_path": "9ac75037c80568de0fff6dc282dc547ab99f3bbcf4d79005038e14984d9fe1ac.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 437, + 584, + 455, + 590 + ], + "lines": [ + { + "bbox": [ + 437, + 584, + 455, + 590 + ], + "spans": [ + { + "bbox": [ + 437, + 584, + 455, + 590 + ], + "type": "text", + "content": "(f) Ours" + } + ] + } + ], + "index": 70, + "angle": 0, + "type": "image_caption" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 137, + 537, + 194, + 582 + ], + "blocks": [ + { + "bbox": [ + 137, + 537, + 194, + 582 + ], + "lines": [ + { + "bbox": [ + 137, + 537, + 194, + 582 + ], + "spans": [ + { + "bbox": [ + 137, + 537, + 194, + 582 + ], + "type": "image", + "image_path": "688c9e2f34235af8266913dae2f012b88a637f642ff5b760ebe95ad520cb24b6.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 584, + 176, + 590 + ], + "lines": [ + { + "bbox": [ + 157, + 584, + 176, + 590 + ], + "spans": [ + { + "bbox": [ + 157, + 584, + 176, + 590 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 613, + 482, + 657 + ], + "lines": [ + { + "bbox": [ + 132, + 613, + 482, + 657 + ], + "spans": [ + { + "bbox": [ + 132, + 613, + 482, + 657 + ], + "type": "text", + "content": "Fig. 4: Visual results for super-resolution and mixed-degradation tasks. We show that our method is capable of generating detailed geometry and texture while other methods tend to be smooth in both geometry and texture. We recommend zooming in for better visualization." + } + ] + } + ], + "index": 72, + "angle": 0, + "type": "image_caption" + } + ], + "index": 59 + }, + { + "type": "image", + "bbox": [ + 194, + 539, + 250, + 582 + ], + "blocks": [ + { + "bbox": [ + 194, + 539, + 250, + 582 + ], + "lines": [ + { + "bbox": [ + 194, + 539, + 250, + 582 + ], + "spans": [ + { + "bbox": [ + 194, + 539, + 250, + 582 + ], + "type": "image", + "image_path": "ba662c96531ef2f998329ee09d48e8c516c4ba42c4d159bfee8dca2623aff237.jpg" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 201, + 584, + 239, + 590 + ], + "lines": [ + { + "bbox": [ + 201, + 584, + 239, + 590 + ], + "spans": [ + { + "bbox": [ + 201, + 584, + 239, + 590 + ], + "type": "text", + "content": "(b) NeRF-SwinIR" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_caption" + } + ], + "index": 61 + }, + { + "type": "image", + "bbox": [ + 250, + 539, + 306, + 582 + ], + "blocks": [ + { + "bbox": [ + 250, + 539, + 306, + 582 + ], + "lines": [ + { + "bbox": [ + 250, + 539, + 306, + 582 + ], + "spans": [ + { + "bbox": [ + 250, + 539, + 306, + 582 + ], + "type": "image", + "image_path": "77d42aa4df6c959a9feac61b2a638254271513d592115dae84e0c4f4a048f28a.jpg" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 584, + 299, + 590 + ], + "lines": [ + { + "bbox": [ + 257, + 584, + 299, + 590 + ], + "spans": [ + { + "bbox": [ + 257, + 584, + 299, + 590 + ], + "type": "text", + "content": "(c) NeRFLiX w.ref" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 284, + 592, + 329, + 598 + ], + "lines": [ + { + "bbox": [ + 284, + 592, + 329, + 598 + ], + "spans": [ + { + "bbox": [ + 284, + 592, + 329, + 598 + ], + "type": "text", + "content": "Mixed Degradation" + } + ] + } + ], + "index": 71, + "angle": 0, + "type": "image_caption" + } + ], + "index": 63 + }, + { + "type": "image", + "bbox": [ + 306, + 539, + 361, + 582 + ], + "blocks": [ + { + "bbox": [ + 306, + 539, + 361, + 582 + ], + "lines": [ + { + "bbox": [ + 306, + 539, + 361, + 582 + ], + "spans": [ + { + "bbox": [ + 306, + 539, + 361, + 582 + ], + "type": "image", + "image_path": "1b461d05594929ab4749a6a2293f58f629e3c0d2414ccc6460d9eb0da61516c7.jpg" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 584, + 357, + 590 + ], + "lines": [ + { + "bbox": [ + 312, + 584, + 357, + 590 + ], + "spans": [ + { + "bbox": [ + 312, + 584, + 357, + 590 + ], + "type": "text", + "content": "(d) NeRFLiX w/o.ref" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_caption" + } + ], + "index": 65 + }, + { + "type": "image", + "bbox": [ + 361, + 539, + 416, + 582 + ], + "blocks": [ + { + "bbox": [ + 361, + 539, + 416, + 582 + ], + "lines": [ + { + "bbox": [ + 361, + 539, + 416, + 582 + ], + "spans": [ + { + "bbox": [ + 361, + 539, + 416, + 582 + ], + "type": "image", + "image_path": "d76df3b73bbc7363bd0e5c3b02cb019a29cea3d25905ca4bf7d7820e34e562e5.jpg" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 584, + 410, + 590 + ], + "lines": [ + { + "bbox": [ + 369, + 584, + 410, + 590 + ], + "spans": [ + { + "bbox": [ + 369, + 584, + 410, + 590 + ], + "type": "text", + "content": "(e) NeRF-Perframe" + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_caption" + } + ], + "index": 67 + }, + { + "bbox": [ + 416, + 539, + 473, + 582 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 539, + 473, + 582 + ], + "spans": [ + { + "bbox": [ + 416, + 539, + 473, + 582 + ], + "type": "image", + "image_path": "08f76fb3be7011e4448e3f31c2730f66df14f9c62facd7e7b4a395578f785de6.jpg" + } + ] + } + ], + "index": 69, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 447, + 101 + ], + "type": "text", + "content": "RaFE: Generative Radiance Fields Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 112, + 281, + 195 + ], + "blocks": [ + { + "bbox": [ + 138, + 112, + 281, + 195 + ], + "lines": [ + { + "bbox": [ + 138, + 112, + 281, + 195 + ], + "spans": [ + { + "bbox": [ + 138, + 112, + 281, + 195 + ], + "type": "image", + "image_path": "f1e4d071783c70441558ce6999370b1d1069e7d4d9e515f853a2b861a54fc25f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 204, + 479, + 228 + ], + "lines": [ + { + "bbox": [ + 132, + 204, + 479, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 204, + 479, + 228 + ], + "type": "text", + "content": "Fig. 5: (a) Geometry comparisons between NeRF-SWINIR/NeRFLIX and our method. (b) Comparisons between using different 2D restoration models." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 117, + 480, + 196 + ], + "blocks": [ + { + "bbox": [ + 307, + 117, + 480, + 196 + ], + "lines": [ + { + "bbox": [ + 307, + 117, + 480, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 117, + 480, + 196 + ], + "type": "image", + "image_path": "59d8628f343e2e2e2ac333b3d29fa6606dc8b9bb600d020c0e923886b9809268.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 143, + 251, + 367, + 310 + ], + "blocks": [ + { + "bbox": [ + 143, + 251, + 367, + 310 + ], + "lines": [ + { + "bbox": [ + 143, + 251, + 367, + 310 + ], + "spans": [ + { + "bbox": [ + 143, + 251, + 367, + 310 + ], + "type": "image", + "image_path": "84a5b24e4bfdbf7aa0e1424316ee20b5007be85010842144e65634f8c83bcf6d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 329, + 480, + 374 + ], + "lines": [ + { + "bbox": [ + 130, + 329, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 480, + 374 + ], + "type": "text", + "content": "Fig. 6: Effectiveness of the tri-plane generator.Left: image rendered by NeRF without generator and images rendered by generative NeRF under different random codes " + }, + { + "bbox": [ + 130, + 329, + 480, + 374 + ], + "type": "inline_equation", + "content": "z(z_{0},z_{1},z_{2})" + }, + { + "bbox": [ + 130, + 329, + 480, + 374 + ], + "type": "text", + "content": ". Right: numerical metrics to evaluate the efficacy of the generator. The results show the effectiveness of using the generator to model the distribution." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 402, + 256, + 481, + 304 + ], + "blocks": [ + { + "bbox": [ + 402, + 256, + 481, + 304 + ], + "lines": [ + { + "bbox": [ + 402, + 256, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 402, + 256, + 481, + 304 + ], + "type": "table", + "html": "
GeneratorX✓
LPIPS↓0.0710.062
LIPE↑3.3704.621
MANIQA↑0.3830.543
", + "image_path": "b137d44b8bd60e3194d18c56dc810d2f17b67dcd67e3669abec6d539e74d0d13.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 399, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 544 + ], + "type": "text", + "content": "a two-level tri-plane structure is adopted. The coarse-level tri-plane pre-trained by low-quality images remains fixed, while the fine-level residual tri-plane to be added to the coarse level is modeled by a GAN-based generator to capture variations in restoration. Extensive experiments on various restoration tasks with both synthetic and real cases demonstrate the superior performance of our method. Limitations and Future Work: One limitation of our method is instability when performing the restoration at extremely high resolutions, such as 4k, due to the patch-rendering strategy. Moreover, due to the inherent slow efficiency of NeRF rendering, currently the long training time also needs to be optimized. To overcome these limitations, a potential solution would be to integrate more efficient rendering techniques like Gaussian splatting [18] into RaFE, enabling the rendering of entire images. We plan to resolve these issues in the future work." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 564, + 240, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 564, + 240, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 564, + 240, + 578 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": "This work was supported in part by the Hong Kong Research Grants Council General Research Fund (17203023), in part by The Hong Kong Jockey Club Charities Trust under Grant 2022-0174, in part by the Startup Fund and the Seed Fund for Basic Research for New Staff from The University of Hong Kong, in part by the funding from UBTECH Robotics, in part by a GRF grant from the Research Grants Council (RGC) of the Hong Kong Special Administrative" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": "Region, China [Project No. CityU 11208123], and in part by the National Natural Science Foundation of China (62132001)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 197, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 197, + 173 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 197, + 173 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 187, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 187, + 480, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 187, + 480, + 220 + ], + "spans": [ + { + "bbox": [ + 138, + 187, + 480, + 220 + ], + "type": "text", + "content": "1. Bahat, Y., Michaeli, T.: Explorable super resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2716-2725 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 221, + 480, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 221, + 480, + 243 + ], + "spans": [ + { + "bbox": [ + 138, + 221, + 480, + 243 + ], + "type": "text", + "content": "2. Bahat, Y., Zhang, Y., Sommerhoff, H., Kolb, A., Heide, F.: Neural volume superresolution. arXiv preprint arXiv:2212.04666 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 243, + 480, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 243, + 480, + 276 + ], + "spans": [ + { + "bbox": [ + 138, + 243, + 480, + 276 + ], + "type": "text", + "content": "3. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 277, + 480, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 480, + 298 + ], + "type": "text", + "content": "4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 300, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 300, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 480, + 320 + ], + "type": "text", + "content": "5. Beyer, L., Zhai, X., Kolesnikov, A.: Big vision. https://github.com/google-research/big_vision (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 322, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 322, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 138, + 322, + 480, + 365 + ], + "type": "text", + "content": "6. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 366, + 480, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 366, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 366, + 480, + 388 + ], + "type": "text", + "content": "7. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: European Conference on Computer Vision (ECCV) (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 388, + 480, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 480, + 421 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 480, + 421 + ], + "type": "text", + "content": "8. Chen, W.T., Yifan, W., Kuo, S.Y., Wetzstein, G.: Dehazenerf: Multiple image haze removal and 3d shape reconstruction using neural radiance fields. arXiv preprint arXiv:2303.11364 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 422, + 480, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 422, + 480, + 454 + ], + "spans": [ + { + "bbox": [ + 138, + 422, + 480, + 454 + ], + "type": "text", + "content": "9. Chen, X., Deng, Y., Wang, B.: Mimic3d: Thriving 3d-aware gans via 3d-to-2d imitation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 455, + 480, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 455, + 480, + 476 + ], + "spans": [ + { + "bbox": [ + 138, + 455, + 480, + 476 + ], + "type": "text", + "content": "0. Chen, Z., Zhang, Y., Ding, L., Bin, X., Gu, J., Kong, L., Yuan, X.: Hierarchical integration diffusion model for realistic image deblurring. In: NeurIPS (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 477, + 480, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 477, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 477, + 480, + 498 + ], + "type": "text", + "content": "1. Chen, Z., Zhang, Y., Gu, J., Kong, L., Yang, X., Yu, F.: Dual aggregation transformer for image super-resolution. In: ICCV (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 499, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 499, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 138, + 499, + 480, + 542 + ], + "type": "text", + "content": "2. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit radiance fields in space, time, and appearance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12479-12488 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 544, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 480, + 576 + ], + "type": "text", + "content": "3. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 577, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 480, + 598 + ], + "type": "text", + "content": "4. Han, Y., Yu, T., Yu, X., Wang, Y., Dai, Q.: Super-nerf: View-consistent detail generation for nef super-resolution. arXiv preprint arXiv:2304.13518 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "type": "text", + "content": "5. Karras, T., Aittala, M., Laine, S., Härkönen, E., Hellsten, J., Lehtinen, J., Aila, T.: Alias-free generative adversarial networks. In: Proc. NeurIPS (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "text", + "content": "6. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of StyleGAN. In: Proc. CVPR (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "7. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. In: Advances in Neural Information Processing Systems (2022)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 447, + 101 + ], + "type": "text", + "content": "RaFE: Generative Radiance Fields Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "18. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "type": "text", + "content": "19. Lee, D., Lee, M., Shin, C., Lee, S.: Dp-nerf: Deblurred neural radiance field with physical scene priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12386-12396 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "text", + "content": "20. Lee, D., Oh, J., Rim, J., Cho, S., Lee, K.M.: Exblurf: Efficient radiance fields for extreme motion blurred images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17639-17648 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "text", + "content": "21. Li, H., Zhang, Z., Jiang, T., Luo, P., Feng, H., Xu, Z.: Real-world deep local motion deblurring. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1314-1322 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "22. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: ICML (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 314 + ], + "type": "text", + "content": "23. Li, J., Zhang, Z., Liu, X., Feng, C., Wang, X., Lei, L., Zuo, W.: Spatially adaptive self-supervised learning for real-world image denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9914-9924 (June 2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "text", + "content": "24. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. arXiv preprint arXiv:2108.10257 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "text", + "content": "25. Lin, X., He, J., Chen, Z., Lyu, Z., Fei, B., Dai, B., Ouyang, W., Qiao, Y., Dong, C.: Diffbir: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "26. Liu, X., Xue, H., Luo, K., Tan, P., Yi, L.: Genn2n: Generative nerf2nerf translation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5105-5114 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "text", + "content": "27. Ma, L., Li, X., Liao, J., Zhang, Q., Wang, X., Wang, J., Sander, P.V.: Deblur-nerf: Neural radiance fields from blurry images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12861-12870 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "text", + "content": "28. Mildenhall, B., Barron, J.T., Chen, J., Sharlet, D., Ng, R., Carroll, R.: Burst denoising with kernel prediction networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2502-2510 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "text", + "content": "29. Mildenhall, B., Hedman, P., Martin-Brualla, R., Srinivasan, P.P., Barron, J.T.: NeRF in the dark: High dynamic range view synthesis from noisy raw images. CVPR (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "text", + "content": "30. Mildenhall, B., Srinivasan, P.P., Ortiz-Cayon, R., Kalantari, N.K., Ramamoorthi, R., Ng, R., Kar, A.: Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG) (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "text", + "content": "31. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 609 + ], + "type": "text", + "content": "32. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (Jul 2022). https://doi.org/10.1145/3528223.3530127, https://doi.org/10.1145/3528223.3530127" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "type": "text", + "content": "33. Pearl, N., Treibitz, T., Korman, S.: Nan: Noise-aware nerfs for burst-denoising. In: CVPR (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "34. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "35. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 160, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 481, + 182 + ], + "type": "text", + "content": "36. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 481, + 214 + ], + "type": "text", + "content": "37. Skorokhodov, I., Tulyakov, S., Wang, Y., Wonka, P.: Epigraph: Rethinking training of 3d gans. Advances in Neural Information Processing Systems 35, 24487-24501 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 214, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 481, + 236 + ], + "type": "text", + "content": "38. Sun, C., Sun, M., Chen, H.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: CVPR (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 236, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 481, + 267 + ], + "type": "text", + "content": "39. Tian, K., Jiang, Y., Yuan, Z., Bingyue, P., Wang, L.: Visual autoregressive modeling: Scalable image generation via next-scale prediction. arXiv preprint arXiv:2404.02905 (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 267, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 481, + 300 + ], + "type": "text", + "content": "40. Wan, Z., Paschalidou, D., Huang, I., Liu, H., Shen, B., Xiang, X., Liao, J., Guibas, L.: Cad: Photorealistic 3d generation via adversarial distillation. arXiv preprint arXiv:2312.06663 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 300, + 481, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 481, + 343 + ], + "type": "text", + "content": "41. Wan, Z., Richardt, C., Božić, A., Li, C., Rengarajan, V., Nam, S., Xiang, X., Li, T., Zhu, B., Ranjan, R., Liao, J.: Learning neural duplex radiance fields for real-time view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8307-8316 (June 2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 343, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 481, + 374 + ], + "type": "text", + "content": "42. Wan, Z., Zhang, B., Chen, D., Zhang, P., Chen, D., Liao, J., Wen, F.: Bringing old photos back to life. In: proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2747-2757 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 481, + 407 + ], + "type": "text", + "content": "43. Wang, C., Wu, X., Guo, Y.C., Zhang, S.H., Tai, Y.W., Hu, S.M.: Nerf-sr: High quality neural radiance fields using supersampling. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 6445-6454 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 407, + 481, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 481, + 440 + ], + "type": "text", + "content": "44. Wang, P., Zhao, L., Ma, R., Liu, P.: BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4170-4179 (June 2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 440, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 481, + 472 + ], + "type": "text", + "content": "45. Wang, P., Zhao, L., Ma, R., Liu, P.: Bad-nerf: Bundle adjusted deblur neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4170-4179 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 472, + 481, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 481, + 504 + ], + "type": "text", + "content": "46. Wang, Y., Yu, J., Zhang, J.: Zero-shot image restoration using denoising diffusion null-space model. The Eleventh International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 504, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 504, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 504, + 481, + 536 + ], + "type": "text", + "content": "47. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 536, + 481, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 481, + 580 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 481, + 580 + ], + "type": "text", + "content": "48. Yang, S., Wu, T., Shi, S., Lao, S., Gong, Y., Cao, M., Wang, J., Yang, Y.: Maniaq: Multi-dimension attention network for no-reference image quality assessment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1191-1200 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 580, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 580, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 580, + 481, + 611 + ], + "type": "text", + "content": "49. Zhang, K., Liang, J., Van Gool, L., Timofte, R.: Designing a practical degradation model for deep blind image super-resolution. In: IEEE International Conference on Computer Vision. pp. 4791-4800 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 611, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 481, + 633 + ], + "type": "text", + "content": "50. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 633, + 481, + 666 + ], + "type": "text", + "content": "51. Zhang, W., Zhai, G., Wei, Y., Yang, X., Ma, K.: Blind image quality assessment via vision-language correspondence: A multitask learning perspective. In: IEEE Conference on Computer Vision and Pattern Recognition. pp. 14071-14081 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 447, + 101 + ], + "type": "text", + "content": "RaFE: Generative Radiance Fields Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 226 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "type": "text", + "content": "52. Zhang, W., Li, X., Chen, X., Qiao, Y., Wu, X.M., Dong, C.: Seal: A framework for systematic evaluation of real-world super-resolution. arXiv preprint arXiv:2309.03020 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 194 + ], + "type": "text", + "content": "53. Zhou, K., Li, W., Wang, Y., Hu, T., Jiang, N., Han, X., Lu, J.: Nerflix: High-quality neural view synthesis by learning a degradation-driven inter-viewpoint mixer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12363-12374 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "text", + "content": "54. Zhou, Y., Li, Z., Guo, C.L., Bai, S., Cheng, M.M., Hou, Q.: Srformer: Permuted self-attention for single image super-resolution. arXiv preprint arXiv:2303.09735 (2023)" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/RadEdit_ stress-testing biomedical vision models via diffusion image editing/730452c5-1675-40d8-8c46-f903c2fa2703_content_list.json b/2024/RadEdit_ stress-testing biomedical vision models via diffusion image editing/730452c5-1675-40d8-8c46-f903c2fa2703_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b64a5003b2d3055012e5ef6950541d4c10ffc15f --- /dev/null +++ b/2024/RadEdit_ stress-testing biomedical vision models via diffusion image editing/730452c5-1675-40d8-8c46-f903c2fa2703_content_list.json @@ -0,0 +1,2108 @@ +[ + { + "type": "text", + "text": "RadEdit: stress-testing biomedical vision models via diffusion image editing", + "text_level": 1, + "bbox": [ + 238, + 141, + 763, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fernando Pérez-García\\*,1, Sam Bond-Taylor\\*,1, Pedro P. Sanchez+,2, Boris van Breugel+,3, Daniel C. Castro1, Harshita Sharma1, Valentina Salvatelli1, Maria T.A. Wetscherek1, Hannah Richardson1, Matthew P. Lungren1,4,5, Aditya Nori1, Javier Alvarez-Valle1, Ozan Oktay†,1, and Maximilian Ilse†,1", + "bbox": [ + 227, + 212, + 774, + 273 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1 Microsoft Health Futures", + "2 University of Edinburgh", + "3 University of Cambridge", + "4 University of California", + "5 Stanford University", + "* Shared first author" + ], + "bbox": [ + 416, + 284, + 583, + 367 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "+ Work done at Microsoft Health Futures", + "bbox": [ + 377, + 368, + 624, + 381 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "† Shared last author", + "bbox": [ + 444, + 382, + 560, + 395 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Biomedical imaging datasets are often small and biased, meaning that real-world performance of predictive models can be substantially lower than expected from internal testing. This work proposes using generative image editing to simulate dataset shifts and diagnose failure modes of biomedical vision models; this can be used in advance of deployment to assess readiness, potentially reducing cost and patient harm. Existing editing methods can produce undesirable changes, with spurious correlations learned due to the co-occurrence of disease and treatment interventions, limiting practical applicability. To address this, we train a text-to-image diffusion model on multiple chest X-ray datasets and introduce a new editing method, RadEdit, that uses multiple image masks, if present, to constrain changes and ensure consistency in the edited images, minimising bias. We consider three types of dataset shifts: acquisition shift, manifestation shift, and population shift, and demonstrate that our approach can diagnose failures and quantify model robustness without additional data collection, complementing more qualitative tools for explainable AI.", + "bbox": [ + 259, + 436, + 740, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Image editing $\\cdot$ diffusion models $\\cdot$ biomedical imaging", + "bbox": [ + 261, + 659, + 660, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 700, + 356, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Developing accurate and robust models for biomedical image analysis requires large and diverse datasets that are often difficult to obtain due to ethical, legal, geographical, and financial constraints [41]. This leads to biased training datasets that affect the performance of trained models and generalisation to real-world scenarios [59, 40]. Such data mismatch may arise from genuine differences in upstream data acquisition as well as from the selection criteria for dataset creation, which materialise as various forms of dataset shifts (population, acquisition, annotation, prevalence, manifestation) [7].", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Top: editing out COVID-19 features results in false positives since the classifier relies on acquisition differences, e.g., radiographic markers (white arrow).", + "bbox": [ + 215, + 172, + 457, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Middle: editing out a pneumothorax (PTX) results in false positives since the classifier instead detects chest drains.", + "bbox": [ + 215, + 227, + 457, + 268 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Bottom: editing abnormalities into lungs causes a lung segmentation model to mislabel (blue: ground-truth segmentation; red: model prediction).", + "bbox": [ + 215, + 268, + 457, + 324 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d2a8b5c9a55a2a7c69814c666b6e44508a804669c7e9df2845998298be920b81.jpg", + "image_caption": [ + "Fig.1: Stress-testing models by simulating dataset shifts via image editing." + ], + "image_footnote": [], + "bbox": [ + 472, + 142, + 787, + 328 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Biomedical vision models, when put into real-world use, can be unhelpful or potentially even harmful to patients if they are affected by dataset shifts, leading to missed diagnoses [22, 74, 76, 56]. For example, the COVID-19 pandemic led to hundreds of detection tools being developed, with some put into use in hospitals; yet Roberts et al. [56] found that \"none of the models identified are of potential clinical use due to methodological flaws and/or underlying biases.\" It is therefore crucial to properly assess models for bias, prior to real-world use.", + "bbox": [ + 212, + 367, + 787, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent deep generative models have made remarkable improvements in terms of sample quality, diversity, and steerability [57, 48, 35, 28]. These models have been shown to generalise to out-of-distribution domains [42, 6, 32, 19], opening up avenues for new applications. One such application is generating synthetic data for stress-testing models [51, 42, 72]. This involves creating data that is realistic, yet can represent settings, domains, or populations that do not appear (enough) in the real training/test data.", + "bbox": [ + 212, + 473, + 787, + 564 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we investigate how deep generative models can be used for stress-testing biomedical imaging models. We consider three dataset shift scenarios:", + "bbox": [ + 212, + 564, + 785, + 594 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Acquisition shift: classifying COVID-19 cases when the positive and negative cases were acquired at different hospitals (Sec. 5.2).", + "2. Manifestation shift: detecting if pneumothorax1 was resolved when chest drains (inserted to treat pneumothorax) are present (Sec. 5.3).", + "3. Population shift: segmenting lungs in the presence of abnormalities rarely or never seen in the training dataset (Sec. 5.4)." + ], + "bbox": [ + 222, + 598, + 784, + 689 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For each of these scenarios, we simulate dataset shifts, producing stress-test sets which can occur in the real world but do not appear or are underrepresented in the original training/test sets. Following prior work, these test sets are synthesised using generative image editing, which unlike generating images from scratch, only minimally modifies the images, hence, better retains fidelity and diversity [51, 42]. For the above scenarios, we use generative editing to 1. remove only COVID-19 features while keeping visual indicators of the different hospitals; 2. remove only pneumothorax while keeping the chest drain; and 3. add abnormalities that occlude lung structures in the image.", + "bbox": [ + 212, + 694, + 787, + 816 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "F. Pérez-García, S. Bond-Taylor et al.", + "bbox": [ + 271, + 114, + 495, + 128 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1 We provide descriptions of the medical terms used throughout the paper in Appendix A.", + "bbox": [ + 238, + 824, + 767, + 840 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We train a diffusion model [27, 57] on a large collection of chest X-rays from a variety of biomedical imaging datasets (Sec. 5.1), enabling us to add and remove a wide variety of pathologies and support devices when editing. Despite the diversity within these datasets, substantial biases are still present, some of which are learned by the generative model; as a result, correlated features may also be modified. For example, in Scenario 2, removing the pneumothorax might also remove the chest drains as both features typically co-occur in datasets [58], since chest drains are used to treat pneumothorax. Furthermore, when editing only within editing masks, artefacts often appear at the border of the masks. Lastly, artefacts occur when editing images outside of the training dataset domain of the diffusion model used for editing. To overcome these challenges, we propose using multiple masks to break existing correlations. This involves defining which regions must change, and explicitly forcing correlated regions to remain unchanged. In addition, we allow the area outside of the masks to be modified by the diffusion model to ensure image consistency. Since our proposed editing method, which we call RadEdit, leads to only minimal overall changes of chest X-rays, we are able to generate synthetic datasets that can be used to stress-test segmentation models (Scenario 3), which, to the best of our knowledge, we are the first to demonstrate.", + "bbox": [ + 217, + 146, + 785, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 238, + 404, + 539, + 417 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel editing approach that reduces the presence of artefacts in edited images and simplifies prompt construction compared to prior work [10, 51].", + "- Our editing approach allows us to construct synthetic datasets with specific data shifts by performing zero-shot edits on datasets/abnormalities not seen in training.", + "- We conduct a broad set of experiments using these synthetic datasets to stress-test and expose biases in biomedical classification and, for the first time, segmentation models, introducing a new use case of synthetic data into the medical setting." + ], + "bbox": [ + 225, + 419, + 782, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Preliminaries", + "text_level": 1, + "bbox": [ + 217, + 544, + 361, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce background context for stress-testing biomedical imaging models: failure modes of biomedical imaging models caused by different dataset shifts; diffusion models as versatile generative models; and diffusion-based image editing.", + "bbox": [ + 217, + 573, + 785, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Dataset shifts", + "text_level": 1, + "bbox": [ + 217, + 638, + 349, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dataset shift refers to a discrepancy between the training and test data distributions due to external factors [7, 34]. Such shifts are regularly observed in machine learning for biomedical imaging, often due to data scarcity. For example, collected training datasets might consist primarily of healthy patients. However, when the model is used in practice after training, there could be a shift towards unhealthy patients. A taxonomy of different types of dataset shifts in the context of biomedical imaging was developed by Castro et al. [7]. In this paper, we consider three dataset shifts of particular interest.", + "bbox": [ + 217, + 661, + 785, + 766 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Acquisition shift results from the use of different scanners (manufacturer, hardware, and software) or imaging protocols as often encountered when using data from multiple cohorts. These changes affect factors such as image resolution, contrast, patient positioning, and image markings.", + "bbox": [ + 217, + 779, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "RadEdit: stress-testing biomedical vision models via diffusion image editing", + "bbox": [ + 279, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Manifestation shift results from the way the prediction targets physically manifest in anatomy changes between domains. For example, training datasets could consist of more severe pathological cases than observed in practice, or a pathology may co-occur with different visual features, e.g., support devices.", + "bbox": [ + 215, + 146, + 787, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Population shift results from differences in intrinsic characteristics of the populations under study, changing the anatomical appearance distribution. This definition encompasses examples such as age, sex, ethnicity, and comorbidities, but also abnormalities such as pleural effusion and support devices. In contrast to manifestation shift, the shift in anatomical appearance is not affected by prediction targets.", + "bbox": [ + 215, + 217, + 787, + 294 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Diffusion models", + "text_level": 1, + "bbox": [ + 215, + 310, + 375, + 324 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Denoising diffusion probabilistic models (DDPMs) [27, 67] are a versatile and effective class of generative models that enable sampling from the data distribution by learning to denoise samples corrupted with Gaussian noise. DDPMs are formed by defining a forward time process that gradually adds noise to data points $x_0$ through the recursion", + "bbox": [ + 215, + 330, + 785, + 391 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = \\sqrt {1 - \\beta_ {t}} x _ {t - 1} + \\sqrt {\\beta_ {t}} \\epsilon_ {t}, \\quad t = 1, \\dots , T, \\quad \\text {s . t .} x _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\bar {\\epsilon} _ {t}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 396, + 785, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon_{1:T},\\bar{\\epsilon}_{1:T}\\sim \\mathcal{N}(0,I),\\beta_{1:T}$ is a predefined noise schedule that determines how quickly to corrupt the data and ensures that $x_{T}$ contains little to no information about $x_0$ , and $\\bar{\\alpha}_t = \\prod_{s = 1}^t (1 - \\beta_s)$ . To form a generative model, the process is reversed in time, gradually transforming Gaussian noise into samples from the learned distribution. While the exact reversal is intractable, a variational approximation is defined by [68]:", + "bbox": [ + 215, + 419, + 787, + 496 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t - 1} = \\hat {\\mu} _ {t} \\left(x _ {t}, f _ {\\theta} \\left(x _ {t}, t, c\\right)\\right) + \\sigma_ {t} z _ {t}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 503, + 785, + 518 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mu} _ {t} \\left(x _ {t}, \\epsilon_ {t}\\right) = \\sqrt {\\bar {\\alpha} _ {t - 1}} \\frac {x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t}}{\\sqrt {\\bar {\\alpha} _ {t}}} + \\sqrt {1 - \\bar {\\alpha} _ {t - 1} - \\sigma_ {t} ^ {2}} \\epsilon_ {t}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 523, + 785, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c$ is a conditioning signal such as a text description, $f_{\\theta}(x_t,t,c)$ is a learned approximation of the noise $\\bar{\\epsilon}_t$ that corrupted the image $x_0$ to obtain $x_{t}$ , $z_{1:T} \\sim \\mathcal{N}(0,I)$ , and $\\sigma_{1:T}$ controls how much noise is introduced. The process is Markovian and known as a DDPM [27] when $\\sigma_t = \\sqrt{(1 - \\bar{\\alpha}_{t - 1}) / (1 - \\bar{\\alpha}_t)}\\sqrt{1 - \\bar{\\alpha}_t / \\bar{\\alpha}_{t - 1}}$ , while for $\\sigma_t = 0$ the process is deterministic and is called a denoising diffusion implicit model (DDIM) [68].", + "bbox": [ + 215, + 560, + 787, + 637 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Image editing", + "text_level": 1, + "bbox": [ + 215, + 652, + 354, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since DDIMs have a one-to-one correspondence with latent vectors $x_{T}$ , we can deterministically map data points to latents by running the DDIM generative process in reverse [68], called DDIM inversion. Several approaches [10, 46] have shown that images can be edited by running the reverse diffusion process augmented by the latent vectors and a modified prompt $c$ . However, this method can lead to undesired artefacts in the edited images. For example, structures unrelated to the desired edit may also change shape, size, or location. To address this, Huberman-Spiegelglas et al. [29] propose DDPM inversion. Here, the original forward process defined in Eq. (1) is adapted, replacing the correlated vectors $\\bar{\\epsilon}_{1:T}$ with statistically independent vectors $\\tilde{\\epsilon}_{1:T}$ (more details in Appendix B). These noise vectors are then used in the generative process, retaining the structure of the original image better than DDIM inversion.", + "bbox": [ + 215, + 672, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "F. Pérez-García, S. Bond-Taylor et al.", + "bbox": [ + 269, + 114, + 495, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Related work", + "text_level": 1, + "bbox": [ + 217, + 143, + 361, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Generative image editing", + "text_level": 1, + "bbox": [ + 217, + 175, + 429, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With advances in deep generative modelling, several approaches to image editing have emerged. Many of these early approaches use compressed latent manipulation [12, 52, 64, 71] where fine-grained edits are difficult to achieve and can result in unwanted changes. More recently, the unparalleled flexibility of diffusion models, together with advances in plain text conditioning, have opened up new avenues for editing techniques.", + "bbox": [ + 217, + 200, + 785, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, we describe some notable diffusion editing methods. SDEdit [46] shows that diffusion models trained solely on real images can be used to generate images from sketches by perturbing sketches with noise, then running the reverse diffusion process. Palette [60] is a diffusion model trained for inpainting by filling regions with noise and denoising. Blended diffusion [2, 3] uses masks with CLIP [53] conditioning to guide local edits. Multiple works show that injecting U-Net activations, obtained by encoding the original image into the generation process, better retains image structure [24, 70]. DiffEdit [10] uses text prompts to determine the appropriate region to edit. Mokady et al. [47] improve diffusion inversion quality by optimising the diffusion trajectory.", + "bbox": [ + 217, + 277, + 785, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Crucially, in the works which use masks for editing, a single type of mask is always used to define the region of interest. In this work, we argue that a second type of mask is required to avoid the loss of features caused by spurious correlations. As better editing approaches are developed, this requirement should be kept in mind.", + "bbox": [ + 217, + 412, + 785, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Biomedical imaging counterfactuals", + "text_level": 1, + "bbox": [ + 217, + 494, + 504, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generative models have also been applied to biomedical counterfactual generation. Reinhold et al. [55] manipulate causes of multiple sclerosis in brain MRI with deep structural causal models [49]. Sanchez et al. [62] and Fontanella et al. [15] use editing to remove pathologies for abnormality detection. Ktena et al. [39] generate out-of-distribution samples to improve classifier performance. Gu et al. [21] train a diffusion model to model disease progression by conditioning on a prior X-ray and text progression description. Unlike our approach, these methods do not use masks to enforce which regions may or may not be edited, meaning that spurious correlations might affect edits. Additionally, these methods use synthetic data to augment and improve model performance whereas we focus on using synthetic medical data for stress-testing.", + "bbox": [ + 217, + 520, + 785, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 Stress-testing", + "text_level": 1, + "bbox": [ + 217, + 691, + 349, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Several approaches have used non-deep-generative-model methods to stress-test networks. Hendrycks and Dietterich [23] evaluate classification models' robustness to corruptions such as blurring, Gaussian noise, and JPEG artefacts. Sakaridis et al. [61] stress-test a segmentation model for roads by using an optical model to add synthetic fog to scenes. Koh et al. [38] collate a dataset presenting various distribution shifts.", + "bbox": [ + 217, + 719, + 785, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "More recent models have made use of conditional generative models to simulate shifts. Prabhu et al. [51] propose LANCE, which stress-tests classification models by using diffusion-based image editing to modify image subjects via caption editing with", + "bbox": [ + 217, + 795, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "RadEdit: stress-testing biomedical vision models via diffusion image editing", + "bbox": [ + 279, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "a large language model (LLM); Kattakinda et al. [36] do similar, but instead modify the background. Li et al. [42] use diffusion models with a single subject mask to separately edit backgrounds and subjects. Van Breugel et al. [72] use generative adversarial networks to simulate distribution shifts on tabular data. This line of research is partially related to adversarial attacks [20], where the focus is on minimally modifying images such that they are visually indistinguishable to a human, but the attacked model fails.", + "bbox": [ + 217, + 146, + 785, + 237 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4 Limitations of existing editing-based stress-testing methods", + "text_level": 1, + "bbox": [ + 217, + 258, + 666, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Recent advancements in diffusion modelling have drastically improved image editing. However, two prevalent approaches, LANCE [51] and DiffEdit [10], produce artefacts in medical images, making them unsuitable for stress-testing biomedical vision models.", + "bbox": [ + 217, + 284, + 785, + 329 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LANCE only uses a global prompt (no mask) for image editing. While effective in the natural image domain, it leads to artefacts in the biomedical domain. For example in Sec. 5.4, we add pathologies and support devices to images of healthy lungs to stress-test lung segmentation models; in such cases we must ensure that the position and shape of the lung borders are not altered during editing. However, we find that LANCE changes the position and shape of the lung border thus making edited images unsuitable for stress-testing such models (Sec. 5.5). In addition, we find that LANCE potentially removes support devices when prompted to remove pathologies, a direct effect of the correlations in the diffusion model's training datasets (Sec. 5.1), making LANCE unsuited for testing the robustness of models to manifestation shift.", + "bbox": [ + 217, + 330, + 785, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DiffEdit addresses these issues by editing only inside an automatically predicted mask $m_{\\mathrm{edit}}$ . However, these predicted masks often mismatch the manually annotated ground-truth, especially for small and complex abnormalities like pneumothorax1 (see Sec. 5.5). Moreover, spurious correlations learned by the diffusion model can lead to the inclusion of support devices in the automatically predicted masks. Furthermore, even when relying on manually annotated masks, DiffEdit can introduce sharp discrepancies at mask boundaries, leading to unrealistic artefacts, such as when adding consolidation that should partially occlude the lung border (Fig. 10b in the Appendix).", + "bbox": [ + 217, + 481, + 785, + 601 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Method", + "text_level": 1, + "bbox": [ + 217, + 625, + 313, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our objective is to create synthetic test data through image editing that simulates specific data shifts, to rigorously evaluate biomedical imaging models. This synthetic data is used to predict model robustness, eliminating need for additional real-world test data.", + "bbox": [ + 217, + 656, + 785, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Improved editing with RadEdit", + "text_level": 1, + "bbox": [ + 217, + 723, + 473, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To address the issues outlined in Sec. 3.4, we propose RadEdit: by introducing 'keep' and 'edit' masks into the editing process, RadEdit explicitly specifies which areas must remain unchanged (keep) and which should be actively modified based on the conditioning signal (edit). Crucially, these masks need not be mutually exclusive, allowing changes in the unmasked regions to ensure global consistency. Using masks, we assume that spurious correlations are mostly non-overlapping [44].", + "bbox": [ + 217, + 750, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 116, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "F. Pérez-García, S. Bond-Taylor et al.", + "bbox": [ + 271, + 114, + 495, + 128 + ], + "page_idx": 5 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 RadEdit (ours) uses multiple masks to decouple spurious correlations" + ], + "code_body": "Require: original image $x_0$ , inversion prompt $c_{\\mathrm{inv}}$ , editing prompt c, edit mask medit, keep mask $m_{\\mathrm{keep}}$ CFG weight w, diffusion model $f_{\\theta}$ $(\\hat{x}_{1:T},z_{1:T})\\leftarrow$ DDPMINVERSION(xo,cinv) $\\triangleright$ Encode image. Procedure in Appendix B \n $x_{T}\\gets \\hat{x}_{T}$ \nfor $t\\gets T$ to 1 do \n $\\epsilon_{\\mathrm{cond},t}\\gets f_{\\theta}(xt,t,c)$ $\\triangleright$ Predict conditional noise \n $\\epsilon_{\\mathrm{uncond},t}\\gets f_{\\theta}(xt,t,c = \\emptyset)$ $\\triangleright$ Predict unconditional noise \n $\\epsilon_t\\gets \\epsilon_{\\mathrm{uncond},t} + w(\\epsilon_{\\mathrm{cond},t} - \\epsilon_{\\mathrm{uncond},t})$ $\\triangleright$ Combine noise predictions with CFG \n $\\epsilon_t\\gets m_{\\mathrm{edit}}\\odot \\epsilon_t + (1 - m_{\\mathrm{edit}})\\odot \\epsilon_{\\mathrm{uncond},t}$ $\\triangleright$ Use CFG only within medit \n $x_{t - 1}\\gets \\hat{\\mu}_t(x_t,\\epsilon_t) + \\sigma_t z_t$ $\\triangleright$ Move to next time step \n $x_{t - 1}\\gets m_{\\mathrm{keep}}\\odot \\hat{x}_{t - 1} + (1 - m_{\\mathrm{keep}})\\odot x_{t - 1}$ $\\triangleright$ Undo edits within mkeep \nreturn edited version of $x_0$", + "bbox": [ + 215, + 164, + 785, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "RadEdit is detailed in Algorithm 1, where a number of key properties make RadEdit more suitable for biomedical image editing than prior editing methods. Firstly, since we aim to edit only within the edit mask $m_{\\mathrm{edit}}$ , classifier-free guidance (CFG) [26] is used only within this region, with high guidance values (following [29], we use a value of 15) ensuring that pathologies are completely removed without drastically changing the rest of the image. This approach also simplifies choosing a prompt for editing since we do not have to take into account the effect of the prompt on the rest of the image. Secondly, we allow the area outside $m_{\\mathrm{edit}}$ to be modified via unconditional generation to ensure image consistency. Lastly, the region of the keep mask $m_{\\mathrm{keep}}$ is reverted to the encoding, ensuring that this region remains the same. Instead of initiating our generating process from pure noise we set $x_{T} = \\hat{x}_{T}$ , where $\\hat{x}_T$ is the last output of the DDPM inversion.", + "bbox": [ + 212, + 359, + 787, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig. 3c, 10c, we show that RadEdit enables artefact-free editing while preserving structures of interest. Because the anatomical layout remains intact after editing, masks still correspond to the same structures, therefore the same masks can be reused to stress-test segmentation models (Sec. 5.4). In practice, we use a latent diffusion model [57], therefore all operations in Algorithm 1 are performed in the latent space of a variational autoencoder (VAE) [57]; this does not limit the generality of the approach.", + "bbox": [ + 212, + 527, + 787, + 617 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Using synthetic images to uncover bias", + "text_level": 1, + "bbox": [ + 214, + 636, + 526, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Despite advancements in biomedical computer vision, recent studies have shown that bias in training and test data can lead to unrealistically high performance of machine learning models on the test set [59, 11]. In our experiments, we use RadEdit to create high quality synthetic test datasets that realistically capture specific dataset shifts, allowing us to quantify the robustness of models to these dataset shifts. By using masks, we can precisely edit the original training data to represent either acquisition shift, manifestation shift, or population shift [7] (Secs. 5.2 to 5.4). These synthetic test sets are used to stress-test (potentially biased) biomedical vision models by comparing performance to the real (biased) test set; a significant drop in performance indicates that the vision model is not robust to the dataset shift that can occur in clinical settings. This serves as a complementary tool to visual explainable AI tools like Grad-CAM [63] and saliency maps [66, 1], which offer qualitative insight into the robustness of models.", + "bbox": [ + 212, + 659, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "RadEdit: stress-testing biomedical vision models via diffusion image editing", + "bbox": [ + 279, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 BioViL-T editing score", + "text_level": 1, + "bbox": [ + 215, + 146, + 416, + 161 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since generative models result in samples of varying quality, poor-quality samples can be filtered out using image-text alignment scores, which quantitatively assess how closely related image-text pairs are via a pre-trained model that embeds similar images and text to nearby vectors [4, 54, 53, 14]. For image editing, we instead assess how similar the change in text and image embeddings are after editing: for a real image-text pair $(I_{\\mathrm{real}}, T_{\\mathrm{real}})$ , edited image-text pair $(I_{\\mathrm{edit}}, T_{\\mathrm{edit}})$ , image encoder $E_I$ , and text encoder $E_T$ , the editing score is defined based on directional similarity [17]:", + "bbox": [ + 212, + 172, + 787, + 277 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nS _ {\\mathrm {B i o V i L - T}} = \\frac {\\Delta I \\cdot \\Delta T}{\\| \\Delta I \\| \\| \\Delta T \\|}, \\quad \\text {w h e r e} \\quad \\begin{array}{l} \\Delta I = E _ {I} \\left(I _ {\\mathrm {e d i t}}\\right) - E _ {I} \\left(I _ {\\mathrm {r e a l}}\\right), \\text {a n d} \\\\ \\Delta T = E _ {T} \\left(T _ {\\mathrm {e d i t}}\\right) - E _ {T} \\left(T _ {\\mathrm {r e a l}}\\right). \\end{array} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 267, + 287, + 785, + 324 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Given the focus on biomedical data, we use the BioViL-T [5] image and text encoders: domain-specific vision-language models trained to analyse chest X-rays and radiology reports, therefore well suited to measure changes in the edited image, such as removed pathologies. Following Prabhu et al. [51], we discard images with $S_{\\mathrm{BioViL - T}} < 0.2$ . This is not only effective for filtering out poor quality edits but is also able to detect whether the original image $I_{\\mathrm{real}}$ does not match the original text description $T_{\\mathrm{real}}$ well.", + "bbox": [ + 212, + 334, + 787, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 449, + 356, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1 Diffusion model", + "text_level": 1, + "bbox": [ + 215, + 479, + 369, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our editing method is heavily dependent on a latent diffusion model [57] that can generate realistic chest X-rays. We use the VAE [37, 25] of SDXL [50] which can adequately reconstruct chest X-rays [8]. The VAE is frozen, and the denoising U-Net is trained on three datasets downsampled and centre-cropped to $512 \\times 512$ pixels: MIMIC-CXR [33], ChestX-ray8 [75], and CheXpert [30], totalling 487680 training images. This data diversity allows us to perform zero-shot edits on datasets not seen during training.", + "bbox": [ + 212, + 505, + 787, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For MIMIC-CXR, we only include frontal view chest X-rays, and condition the denoising U-Net on the corresponding impression section in the radiology report (a short clinically actionable outline of the main findings). We employ the tokeniser and frozen text encoder from BioViL-T [5]. For ChestX-ray8 and CheXpert, we condition on a list of all abnormalities present in an image as indicated by the labels, e.g., 'Cardiomegaly. Pneumothorax'. If the list of abnormalities is empty, we use the string 'No findings'. More details on the datasets and diffusion model training can be found in Appendix E.", + "bbox": [ + 212, + 595, + 787, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2 Acquisition shift", + "text_level": 1, + "bbox": [ + 215, + 724, + 372, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Background To show how RadEdit can be used to quantify the robustness of models to acquisition shift, we closely follow the experimental setup of DeGrave et al. [11], who show that deep learning systems built to detect COVID-19 from chest X-rays rely on confounding factors rather than pathology features. This problem arises when COVID-19-positive and -negative images come from disparate sources. In our setup, all COVID-19-positive cases come from the BIMCV dataset [73] (denoted BIMCV+),", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "F. Pérez-García, S. Bond-Taylor et al.", + "bbox": [ + 271, + 114, + 495, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9111de6a974fbc3b2c5917a8c2007f88b2d20157edd2638845627203f073c19f.jpg", + "image_caption": [ + "Fig.2: Removing COVID-19 features with LANCE $^2$ (b) also changes the laterality markers and reduces contrast. In contrast, RadEdit (c; ours) preserves anatomical structures and laterality markers, and retains the original contrast.", + "(a) Original Image" + ], + "image_footnote": [], + "bbox": [ + 406, + 148, + 535, + 248 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/fc36787c107450cc0af8b9e69ae3edd6634b3472011cf568d7520bb5ec21c319.jpg", + "image_caption": [ + "(b) LANCE [51]" + ], + "image_footnote": [], + "bbox": [ + 535, + 148, + 656, + 248 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/354372db538e0480126b35eabce74802facc1997cc588ad2fe8ec5b211b46288.jpg", + "image_caption": [ + "(c) RadEdit (ours)" + ], + "image_footnote": [], + "bbox": [ + 656, + 148, + 784, + 248 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "and all COVID-19-negative cases from MIMIC-CXR [33] (see Fig. 11). A classifier trained on these datasets will rely on spurious features indicative of the data's origin, e.g., laterality markers or the field of view, instead of features caused by the pathology.", + "bbox": [ + 214, + 297, + 782, + 345 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Setup A synthetic test set is created by applying RadEdit to remove COVID-19 features1 from BIMCV+ images using the prompt 'No acute cardiopulmonary process'3 (Fig. 2); the included bounding boxes of COVID-19 features are used as the edit mask $m_{\\mathrm{edit}}$ . Since this is the only mask available, we set the keep mask as $m_{\\mathrm{keep}} = 1 - m_{\\mathrm{edit}}$ . After filtering using the BioViL-T editing (Sec. 4.3), this results in a synthetic dataset of 2774 COVID-19-negative images containing the same spurious features as BIMCV+. Both the weak and strong predictor models are classifiers trained using a ResNet50 architecture (see Appendix F for more implementation details).", + "bbox": [ + 214, + 359, + 787, + 479 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Findings Tab. 1, shows the performance of a COVID-19 classifier (weak predictor) trained on BIMCV+ and MIMIC-CXR. In accordance with DeGrave et al. [11], we find that the weak predictor performs exceptionally well on the real test set (i.e. test splits of both datasets) since the model learned to distinguish the two data sources instead of learning visual features related to COVID-19. However, in the second row of Tab. 1, we see a drop of $95\\%$ in accuracy meaning that the model fails to classify the synthetic images as COVID-19-negative. The weak predictor is not robust to a shift in acquisition.", + "bbox": [ + 214, + 494, + 787, + 599 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To show that the decreased performance of the weak predictor is not caused by artefacts in the edited images, we train a more robust COVID-19 classifier (strong predictor), using the BIMCV+ and BIMCV- datasets, as in [11], where the BIMCV- dataset consists of only COVID-19-negative cases from BIMCV, and test on the same two test", + "bbox": [ + 214, + 601, + 787, + 660 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/a1be34d50db1492e81c3452b6ecfd8e322d799b6a444cbd7f0b0f481c12a6f00.jpg", + "table_caption": [ + "Table 1: Quantifying robustness of COVID-19 detectors to acquisition shift. We train a weak predictor on the 'Biased' dataset—a combination of BIMCV+ [73] and MIMIC-CXR [33]; and a strong predictor on an unbiased dataset—a combination of BIMCV+ and BIMCV; the 'Synthetic' test set consists of 2774 COVID-19-negative images with the same spurious features as the BIMCV+ datasets, e.g. laterality markers. We report mean accuracy and standard deviation across 5 runs." + ], + "table_footnote": [], + "table_body": "
PredictorTest dataAccuracy
WeakBiased99.1 ± 0.2
WeakSynthetic5.5 ± 2.1
StrongBiased74.4 ± 3.0
StrongSynthetic76.0 ± 7.7
", + "bbox": [ + 593, + 700, + 784, + 791 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "RadEdit: stress-testing biomedical vision models via diffusion image editing", + "bbox": [ + 279, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "2For LANCE, we perform the text perturbation manually.", + "bbox": [ + 238, + 810, + 580, + 824 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "This is a common radiological description of a 'normal' chest X-ray.", + "bbox": [ + 240, + 824, + 650, + 839 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/d8b08b200a2c06a0e26d7200596a7eaa7c8afb6de45548faa4a8189e7973fe82.jpg", + "image_caption": [ + "Fig.3: Removing pneumothorax (red) with LANCE $^2$ (b) also removes the spuriously correlated chest drain (blue) and reduces contrast. In contrast, RadEdit (c; ours) preserves the chest drain and better preserves anatomical structures.", + "(a) Original Image" + ], + "image_footnote": [], + "bbox": [ + 415, + 145, + 535, + 242 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/fc158dd38debd7237738a6c816ded24d5d19c7a6cf421d03d19ba7f44547e67f.jpg", + "image_caption": [ + "(b) LANCE [51]" + ], + "image_footnote": [], + "bbox": [ + 535, + 146, + 656, + 242 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/bb6ce92710ee1361641cde70c894f33ba86a2916f4f05e6cf11783410a8f5ebf.jpg", + "image_caption": [ + "(c) RadEdit (ours)" + ], + "image_footnote": [], + "bbox": [ + 656, + 146, + 784, + 242 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "datasets. Comparing rows one and three of Tab. 1, we find that the strong predictor performs worse on the test set containing samples from BIMCV+ and MIMIC-CXR than the weak predictor (row one). This is expected as the strong predictor relies on actual pathology features. Lastly, rows three and four of Tab. 1 show that the strong predictor performs similarly on the real and synthetic test sets, attesting to the quality of our edits.", + "bbox": [ + 214, + 290, + 782, + 364 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.3 Manifestation shift", + "text_level": 1, + "bbox": [ + 215, + 388, + 388, + 402 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Background In this section, we show how RadEdit can be used to quantify the robustness of biomedical vision models to manifestation shift. We closely follow the experimental setup of Rueckel et al. [59], who demonstrate that pneumothorax classification models are strongly biased by the presence of chest drains: while the average performance of pneumothorax classifiers is high, performance on the subset of images with a chest drain but no pneumothorax is significantly lower. This is due to chest drains being a common treatment for pneumothorax, resulting in the majority of images in datasets like CANDID-PTX [13] containing a chest drain only if there is a pneumothorax. As a result, only $1\\%$ of images in CANDID-PTX contain a chest drain but no pneumothorax.", + "bbox": [ + 214, + 414, + 784, + 551 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Setup We use RadEdit to create a synthetic dataset containing images with chest drains but no pneumothorax, by editing out the pneumothorax from CANDID-PTX images using the prompt 'No acute cardiopulmonary process' (Fig. 3). The edit mask $m_{\\mathrm{edit}}$ is set as a mask of the pneumothorax, and the keep mask $m_{\\mathrm{keep}}$ set as the chest drain mask. This ensures that the chest drain is not removed, while preventing border artefacts. After filtering using the BioViL-T editing score (Sec. 4.3), 628 images are left; in contrast, the real test set contains only 16 of cases with drains but no pneumothorax. The weak predictor is a ResNet50 trained to classify pneumothorax. Following Rueckel et al. [59], the strong predictor is an EfficientNet U-Net [69] trained on SIIM-ACR [77] to both segment pneumothorax and classify pneumothorax (more details in Appendix G).", + "bbox": [ + 214, + 566, + 784, + 719 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Findings In accordance with [59], we show in Tab. 2 that a pneumothorax classifier (weak predictor) trained on CANDID-PTX performs exceptionally well on the test split of CANDID-PTX, since very few images contain a chest drain and no pneumothorax. However, in row two of Tab. 2, we show a drastic drop in performance on the synthetic test set, i.e., the weak predictor is not robust to manifestation shift. To show that the drop in performance on the synthetic dataset does not come from editing artefacts, we evaluate the more robust segmentation model (strong predictor) on the same test datasets", + "bbox": [ + 214, + 734, + 782, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "F. Pérez-García, S. Bond-Taylor et al.", + "bbox": [ + 271, + 114, + 493, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/ea7c63fdedf15994d67632ed6840aa312c2aa8288c0ff54f200e75f16485ec7e.jpg", + "table_caption": [ + "Table 2: Quantifying robustness of pneumothorax detectors to manifestation shift. The weak predictor is trained on the biased CANDID-PTX [13] dataset to classify pneumothorax; the strong predictor is trained on SIIM-ACR [77] to classify and segment the pneumothorax. Real 'Biased' test data comes from CANDID-PTX which exhibits strong confounding between the pneumothorax and chest tubes; 'Synthetic' test data is 629 solely edited images containing chest drains but no pneumothorax. We report mean accuracy and standard deviation across 5 runs." + ], + "table_footnote": [], + "table_body": "
PredictorTest dataAccuracy
WeakBiased93.3 ± 0.6
WeakSynthetic17.9 ± 3.7
StrongBiased93.7 ± 1.3
StrongSynthetic81.7 ± 7.1
", + "bbox": [ + 594, + 174, + 789, + 268 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": " Rows three and four of Tab. 2), we find that the strong predictor performs on par with the weak predictor in row one; however, the strong predictor closes the majority of the gap between the real test set and the synthetic one, attesting to the quality of our edits. In agreement with Rueckel et al. [59], there is still a performance gap, indicating that the strong predictor still suffers from mild manifestation shift.", + "bbox": [ + 214, + 324, + 787, + 401 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.4 Population shift", + "text_level": 1, + "bbox": [ + 215, + 421, + 370, + 436 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Background In this section, we show how RadEdit can be used to quantify the robustness of lung segmentation models to population shifts. Manually segmenting X-ray images is labour intensive and requires high expertise, leading to small datasets often limited to single pathologies or healthy patients [65, 31], e.g., MIMIC-Seg [9]. These models are thus sensitive to occlusions such as medical devices or pathologies, which typically appear as white regions on X-rays [43]. Evaluating model robustness requires further image collection for each occlusion type, which is time-consuming and costly.", + "bbox": [ + 214, + 445, + 787, + 551 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Setup Here, abnormalities are added to the lung region in healthy X-rays from MIMIC-Seg (Fig. 4). The edit mask $m_{\\mathrm{edit}}$ is set as a mask of the lung(s). When editing a single", + "bbox": [ + 214, + 564, + 787, + 595 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d1509fd8ad578531c7ac80d7555bd446cc49b7e0b76c72db5475b17b150e4d51.jpg", + "image_caption": [ + "Fig.4: Adding pulmonary edema (top), pacemakers (middle), and consolidation (bottom) with RadEdit. The 'strong predictor' (d), a segmentation model trained on CheXmask [16] (a large dataset containing various abnormalities) is more robust to these abnormalities than the 'weak predictor' (c), a segmentation model trained on MIMIC-Seg [9] (a small set of mostly healthy patients): the weak predictor traces around the pacemaker and poorly annotates the consolidated lung. Blue: ground-truth annotation; red: predicted segmentation." + ], + "image_footnote": [], + "bbox": [ + 465, + 621, + 787, + 833 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "RadEdit: stress-testing biomedical vision models via diffusion image editing", + "bbox": [ + 279, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/7dde52658bee57b516c34c59ca481e039836d98bfda9979ce6734f3668ef1ba4.jpg", + "table_caption": [ + "Table 3: Quantifying robustness of lung segmentation models to population shift. The 'weak predictor' is trained on MIMIC-Seg (a small set of predominantly healthy patients); the 'strong predictor' is trained on CheXmask (a large mixed set of patients with various abnormalities). Synthetic test data is created by using RadEdit to add edema, pacemakers, and consolidation. We report the change $(\\Delta)$ in Dice score and AHD with respect to the segmentation models evaluated on the ground-truth test set." + ], + "table_footnote": [], + "table_body": "
Test dataWeak PredictorStrong Predictor
Dice↑Δ↓AHD↓Δ↓Dice↑Δ↓AHD↓Δ↓
Real data97.4—6.1—95.5—11.6—
Healthy → edema93.83.621.815.793.91.622.811.2
Healthy → pacemaker85.012.449.843.787.38.229.517.9
Healthy → consolidation85.911.544.138.188.17.429.417.8
", + "bbox": [ + 251, + 239, + 750, + 354 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "lung, the keep mask $m_{\\mathrm{keep}}$ corresponds to the lung which must not change, while when editing both lungs we set $m_{\\mathrm{keep}} = 0$ to allow opacity adjustments, or for elements to be added outside of the lungs. Stress-test sets are generated for three abnormalities: pulmonary edema, pacemakers, and consolidation1. Prompts are phrased to match similar impressions in the training data (see Appendix H). Both predictors are EfficientNet U-Net [69] models; the weak predictor is trained on MIMIC-Seg, while the strong predictor is trained on CheXmask [16], a larger dataset with various lung abnormalities (more details in Appendix H). We evaluate segmentation quality using Dice similarity coefficient, which is the harmonic mean of the precision and recall, and 95th percentile average Hausdorff distance (AHD), a measure of the distance between two sets [45].", + "bbox": [ + 212, + 382, + 787, + 532 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Findings Tab. 3 shows that the weak predictor model performs well on the real biased test data, since it is mostly composed of healthy subjects. However, testing on the synthetic lung abnormality datasets (rows two to four), causes performance to drop substantially, i.e. the weak predictor is not robust to population shift. To show that this drop in performance does not come from editing artefacts, we evaluate the strong predictor on the synthetic test sets and see considerably smaller changes in performance. This can be seen in Fig. 4: for pulmonary edema, both models can accurately segment, despite the abnormality; for pacemakers, the weak predictor incorrectly segments around the pacemakers, while the strong predictor more accurately segments the lungs; and for consolidation, both models are less able to segment the lungs accurately, however, the strong predictor gets closer to the ground-truth. See Appendix H for more examples.", + "bbox": [ + 212, + 547, + 787, + 715 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.5 Quantifying the limitations of existing editing methods", + "text_level": 1, + "bbox": [ + 214, + 738, + 635, + 753 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "LANCE As seen in the second row of Tab. 3, adding edema leads only to a small drop in performance of the strong predictor. We hypothesise that further drops in performances stem from a mismatch of the original mask and the edited images. We therefore use this setup to quantify how well LANCE and RadEdit preserve the shape and position of the lung borders. Additionally, we study the difference between results using", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "F. Pérez-García, S. Bond-Taylor et al.", + "bbox": [ + 271, + 114, + 493, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/375fc1d5fb456937dfea497ad62d2705686a2b443a6a0a34aa4834acdb4f4e09.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 326, + 233 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d4bd5f26608b23783c54424f27891562ec463a3ab4e5603f08ea6e297c24f377.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 330, + 145, + 444, + 232 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9f465dfb4e44cb0f02d1cd45eb24d43345bf71b88185e79f866293b897d8461f.jpg", + "image_caption": [ + "(c)", + "Fig. 5: Comparison of LANCE $^2$ and RadEdit. We measure how well the strong predictor from Tab. 3's outputs matches the ground-truth lung masks (blue) for four synthetic datasets created by adding edema using LANCE and RadEdit with DDIM or DDPM inversion. High Dice / low AHD indicates that the editing method well preserves the lung border's location and shape." + ], + "image_footnote": [], + "bbox": [ + 444, + 145, + 557, + 232 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1b3c5ac8ffcec0dc13caeb3a545a3d2c7b44d250909b74a47e72414330f29ad5.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 557, + 145, + 668, + 232 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/af6befdacdf2b5fb7813977631443e3169b0f3008a4cd3cd061e61b24aa5cff7.jpg", + "image_caption": [ + "(e)" + ], + "image_footnote": [], + "bbox": [ + 669, + 145, + 785, + 233 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/4d5a72cfccddfa19ddfe3a6600b2b90edaa7b9a21ed2602d9341520e8093e60b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Editing MethodDice↑AHD↓
(a) Original data95.511.6
(b) LANCE w/ DDIM78.965.1
(c) LANCE w/ DDPM80.169.5
(d) RadEdit w/ DDIM86.239.8
(e) RadEdit w/ DDPM93.922.8
", + "bbox": [ + 550, + 251, + 776, + 351 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "DDIM or DDPM inversion. For all four methods in Fig. 5, we use the same setup as in Sec. 5.4: we first edit the original image with the prompt 'Moderate pulmonary edema. The heart size is normal', and then compare the outputs of the strong predictor with the original ground-truth lung masks. Here, we find that using masks and DDPM inversion is necessary for RadEdit to preserve the shape and position of the lung border.", + "bbox": [ + 214, + 373, + 787, + 450 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "DiffEdit We quantify how well DiffEdit's automatically predicted masks match the manual ground-truth using the same setup as in Sec. 5.3: we take an image containing pneumothorax and a chest drain, and try to remove only the pneumothorax. The editing prompt is created by keeping only the parts of the impressions related to pneumothorax and chest drains, and replacing the description of the pneumothorax with 'No pneumothorax'. DiffEdit should therefore predict a mask containing only the pneumothorax. We perform a grid search on the MIMIC-Seg [9] validation set over DiffEdit's hyperparameters (noise strength and binarising threshold) to optimise pneumothorax segmentation metrics, then evaluate on the training set. In Fig. 6 we see that DiffEdit's predicted masks obtain poor quantitative metrics where parts of the pneumothorax are often missing, and the spuriously correlated chest drain is often included in the predicted mask. As a result, DiffEdit's predicted masks are unsuitable for stress-testing.", + "bbox": [ + 214, + 463, + 787, + 645 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7f9e7b069302971b2be8f0e75a6ff32d5734c9ba94f8c8563f98d4282702d961.jpg", + "image_caption": [ + "(a) Examples of pneumothorax masks predicted using DiffEdit [10]. Blue: ground-truth annotation; red: predicted editing mask." + ], + "image_footnote": [], + "bbox": [ + 217, + 672, + 328, + 762 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9a69c7837b7f1cfa312e9da9f6d9a67e5d85557e42ec66badfe8ccd74cad0ccb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 341, + 672, + 452, + 762 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b3f352f725da320efcb78cb78f5944d6971e44d5b73a91c410a4c774baa79c52.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 457, + 672, + 570, + 762 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Hyperparameters Dice AHD", + "text_level": 1, + "bbox": [ + 586, + 676, + 777, + 689 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Tuned per image 33.8 97.7", + "bbox": [ + 586, + 696, + 777, + 708 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Tuned on validation 18.4 256.8", + "bbox": [ + 586, + 710, + 776, + 722 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "(b) Segmentation metrics for the pneumothorax mask predicted by DiffEdit [10], for hyperparameters tuned on the validation set (bottom) and tuned per image (top; which requires ground-truth masks).", + "bbox": [ + 576, + 729, + 785, + 781 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Fig. 6: Evaluating pneumothorax masks predicted using DiffEdit [10]. (a) Predicted masks (red) are noisy, with chest drains often incorrectly segmented as well as or instead of the pneumothorax (blue); (b) this is demonstrated quantitatively with low Dice score and high AHD.", + "bbox": [ + 214, + 792, + 787, + 835 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "RadEdit: stress-testing biomedical vision models via diffusion image editing", + "bbox": [ + 279, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 769, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6 Limitations and future work", + "text_level": 1, + "bbox": [ + 217, + 143, + 486, + 161 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Despite the encouraging results presented in the paper, RadEdit is not without limitations and more work is needed to extend it to more applications. Currently, training datasets and models must be manually analysed to predict potential failure cases, simulate these failures to test the hypothesis, and finally quantitatively evaluate the model; future work could automate such failure mode discovery. Another limitation is that current editing techniques do not enable all types of stress-testing; for example, with current approaches, we are unable to test segmentation models' behaviour to cardiomegaly (enlarged heart) since this would require segmentation maps to be adjusted after editing. However, this could potentially be enabled by enlarging heart segmentations to simulate cardiomegaly and adjusting the ground-truth lung segmentation accordingly.", + "bbox": [ + 217, + 172, + 785, + 323 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "When using generative editing, it is not possible to guarantee that unwanted changes will not occur. With RadEdit, we minimise this by forcing spuriously correlated regions to remain the same, only using classifier-free guidance within the editing mask, and filtering via image-text alignment. Future work to better maintain structure when editing will help with this issue, but masks will still be necessary to bypass spurious correlations. Furthermore, due to potential overlap of 3D structures, editing 2D X-rays can be limited. In such cases, editing single structures using RadEdit may not be successful. To address this, 3D CT images could be edited, then projected to synthetic X-rays [18].", + "bbox": [ + 217, + 324, + 785, + 444 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "When producing simulated stress test sets, several factors affect edit quality including classifier-free guidance weight, number of inference steps, and time step to encode to. Additionally, the text encoder must well understand specified pathologies to provide informative features to condition the generative model on; similarly, the diffusion model must be able to capture fine details and well cover the data distribution.", + "bbox": [ + 217, + 444, + 785, + 518 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Finally, more research is required to develop better approaches for quantifying edit quality for downstream tasks. In particular, observing a change in downstream performance is not necessarily indicative of real-world performance as edit quality may be poor. While the introduced BioViL-T editing score can be used to quantify edit quality, this introduces reliance on a potentially biased model. Additionally, the BioViL-T editing score is not suited to detect the artefacts introduced by LANCE and DiffEdit.", + "bbox": [ + 217, + 518, + 785, + 609 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 217, + 628, + 341, + 646 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this study, we illustrate the efficacy of generative image editing as a robust tool for stress-testing biomedical vision models. Our focus is on assessing their robustness against three types of dataset shifts commonly encountered in biomedical imaging: acquisition shift, manifestation shift, and population shift. We highlight that one of the significant challenges in biomedical image editing is the correlations learned by the generative model, which can result in artefacts during the editing process. To mitigate these artefacts, RadEdit relies on various types of masks to restrict the effects of the editing to certain areas while ensuring the consistency of the edited images. This approach enables us to generate high fidelity synthetic test sets that exhibit common dataset shifts. These synthetic test sets are used to identify and quantify the failure modes of biomedical classification and segmentation models. This provides a valuable supplement to explainable AI approaches such as Grad-CAM [63] and saliency maps [66, 1].", + "bbox": [ + 217, + 657, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "F. Pérez-García, S. Bond-Taylor et al.", + "bbox": [ + 271, + 114, + 493, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Bibliography", + "text_level": 1, + "bbox": [ + 433, + 142, + 568, + 162 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. Advances in neural information processing systems, 31, 2018.", + "[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022.", + "[3] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM Transactions on Graphics (TOG), 42(4):1-11, 2023.", + "[4] Samaneh Azadi, Catherine Olsson, Trevor Darrell, Ian Goodfellow, and Augustus Odena. Discriminator rejection sampling. arXiv preprint arXiv:1810.06758, 2018.", + "[5] Shruthi Bannur, Stephanie Hyland, Qianchu Liu, Fernando Perez-Garcia, Maximilian Ilse, Daniel C. Castro, Benedikt Boecking, Harshita Sharma, Kenya Bouzid, Anja Thieme, Anton Schwaighofer, Maria Wetscherek, Matthew P. Lungren, Aditya Nori, Javier Alvarez-Valle, and Ozan Oktay. Learning to Exploit Temporal Structure for Biomedical Vision-Language Processing. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 1 2023. https://doi.org/10.48550/arxiv.2301.04558.", + "[6] Riccardo Barbano, Alexander Denker, Hyungjin Chung, Tae Hoon Roh, Simon Arridge, Peter Maass, Bangti Jin, and Jong Chul Ye. Steerable conditional diffusion for out-of-distribution adaptation in imaging inverse problems. arXiv preprint arXiv:2308.14409, 2023.", + "[7] Daniel C Castro, Ian Walker, and Ben Glocker. Causality matters in medical imaging. Nature Communications, 11(1):3673, 2020.", + "[8] Pierre Chambon, Christian Bluethgen, Jean-Benoit Delbrouck, Rogier Van der Sluijs, Malgorzata Polacin, Juan Manuel Zambrano Chaves, Tanishq Mathew Abraham, Shivanshu Purohit, Curtis P Langlotz, and Akshay Chaudhari. Roentgen: vision-language foundation model for chest x-ray generation. arXiv preprint arXiv:2211.12737, 2022.", + "[9] Li-Ching Chen, Po-Chih Kuo, Ryan Wang, Judy Gichoya, and Leo Anthony Celi. Chest X-ray segmentation images based on MIMIC-CXR (version 1.0.0). PhysioNet, 2022.", + "[10] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. DiffEdit: Diffusion-based semantic image editing with mask guidance, 2022.", + "[11] Alex J. DeGrave, Joseph D. Janizek, and Su-In Lee. AI for radiographic COVID-19 detection selects shortcuts over signal. Nature Machine Intelligence, 3(7):610-619, 2021. ISSN 2522-5839. https://doi.org/10.1038/s42256-021-00338-7.", + "[12] Alexey Dosovitskiy, Jost Tobias Springenberg, and Thomas Brox. Learning to generate chairs with convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1538-1546, 2015.", + "[13] Sijing Feng, Damian Azzollini, Ji Soo Kim, Cheng-Kai Jin, Simon P. Gordon, Jason Yeoh, Eve Kim, Mina Han, Andrew Lee, Aakash Patel, Joy Wu, Martin" + ], + "bbox": [ + 217, + 210, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Urschler, Amy Fong, Cameron Simmers, Gregory P. Tarr, Stuart Barnard, and Ben Wilson. Curation of the CANDID-PTX dataset with free-text reports. *Radiology: Artificial Intelligence*, 3(6):e210136, 2021. ISSN 2638-6100. https://doi.org/10.1148/ryai.2021210136.", + "[14] Virginia Fernandez, Pedro Sanchez, Walter Hugo Lopez Pinaya, Grzegorz Jacenków, Sotirios A Tsaftaris, and Jorge Cardoso. Privacy distillation: Reducing re-identification risk of multimodal diffusion models. arXiv preprint arXiv:2306.01322, 2023.", + "[15] Alessandro Fontanella, Grant Mair, Joanna Wardlaw, Emanuele Trucco, and Amos Storkey. Diffusion models for counterfactual generation and anomaly detection in brain images. URL http://arxiv.org/abs/2308.02062.", + "[16] Nicolas Gaggion, Candelaria Mosquera, Lucas Mansilla, Martina Aineseder, Diego H Milone, and Enzo Ferrante. CheXmask: a large-scale dataset of anatomical segmentation masks for multi-center chest X-ray images. arXiv preprint arXiv:2307.03293, 2023.", + "[17] Rinon Gal, Or Patashnik, Haggai Maron, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. StyleGAN-NADA: CLIP-guided domain adaptation of image generators. ACM Transactions on Graphics, 41(4), 2022. https://doi.org/10.1145/3528223.3530164.", + "[18] Cong Gao, Benjamin D Killeen, Yicheng Hu, Robert B Grupp, Russell H Taylor, Mehran Armand, and Mathias Unberath. Synthetic data accelerates the development of generalizable learning-based algorithms for x-ray image analysis. Nature Machine Intelligence, 5(3):294-308, 2023.", + "[19] Timur Garipov, Sebastiaan De Peuter, Ge Yang, Vikas Garg, Samuel Kaski, and Tommi Jaakkola. Compositional sculpting of iterative generative processes. arXiv preprint arXiv:2309.16115, 2023.", + "[20] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples, 2015.", + "[21] Yu Gu, Jianwei Yang, Naoto Usuyama, Chunyuan Li, Sheng Zhang, Matthew P Lungren, Jianfeng Gao, and Hoifung Poon. Biomedjourney: Counterfactual biomedical image generation by instruction-learning from multimodal patient journeys. arXiv preprint arXiv:2310.10765, 2023.", + "[22] Will Douglas Heaven. Hundreds of ai tools have been built to catch covid. none of them helped. MIT Technology Review. Retrieved December 2023, 2021.", + "[23] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2018.", + "[24] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations, 2022.", + "[25] Irina Higgins, Loic Matthew, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. $\\beta$ -VAE: Learning basic visual concepts with a constrained variational framework. In International Conference on Learning Representations, 2016.", + "[26] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022." + ], + "bbox": [ + 215, + 145, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "F. Pérez-García, S. Bond-Taylor et al.", + "bbox": [ + 271, + 114, + 493, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020.", + "[28] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research, 23(1):2249-2281, 2022.", + "[29] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly DDPM noise space: Inversion and manipulations, 2023.", + "[30] Jeremy Irvin, Pranav Rajpurkar, Michael Ko, Yifan Yu, Silviana Ciurea-Ilicus, Chris Chute, Henrik Marklund, Behzad Haghgoo, Robyn Ball, Katie Shpanskaya, Jayne Seekins, David A. Mong, Safwan S. Halabi, Jesse K. Sandberg, Ricky Jones, David B. Larson, Curtis P. Langlotz, Bhavik N. Patel, Matthew P. Lungren, and Andrew Y. Ng. CheXpert: A large chest radiograph dataset with uncertainty labels and expert comparison, 2019.", + "[31] Stefan Jaeger, Sema Candemir, Sameer Antani, Yi-Xiang J Wang, Pu-Xuan Lu, and George Thoma. Two public chest x-ray datasets for computer-aided screening of pulmonary diseases. Quantitative imaging in medicine and surgery, 4(6):475, 2014.", + "[32] Priyank Jaini, Kevin Clark, and Robert Geirhos. Intriguing properties of generative classifiers. arXiv preprint arXiv:2309.16779, 2023.", + "[33] Alistair E. W. Johnson, Tom J. Pollard, Seth J. Berkowitz, Nathaniel R. Greenbaum, Matthew P. Lungren, Chih-ying Deng, Roger G. Mark, and Steven Horng. MIMIC-CXR, a de-identified publicly available database of chest radiographs with free-text reports. Scientific Data, 6(1):317, 2019. ISSN 2052-4463. https://doi.org/10.1038/s41597-019-0322-0.", + "[34] Charles Jones, Daniel C Castro, Fabio De Sousa Ribeiro, Ozan Oktay, Melissa McCradden, and Ben Glocker. No fair lunch: A causal perspective on dataset bias in machine learning for medical imaging. arXiv preprint arXiv:2307.16526, 2023.", + "[35] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10124–10134, 2023.", + "[36] Priyatham Kattakinda, Alexander Levine, and Soheil Feizi. Invariant learning via diffusion dreamed distribution shifts. arXiv preprint arXiv:2211.10370, 2022.", + "[37] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes, 2022.", + "[38] Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas Phillips, Irena Gao, et al. Wilds: A benchmark of in-the-wild distribution shifts. In International Conference on Machine Learning, pages 5637-5664. PMLR, 2021.", + "[39] Ira Ktena, Olivia Wiles, Isabela Albuquerque, Sylvestre-Alvise Rebuffi, Ryutaro Tanno, Abhijit Guha Roy, Shekoofeh Azizi, Danielle Belgrave, Pushmeet Kohli, Alan Karthikesalingam, et al. Generative models improve fairness of medical classifiers under distribution shifts. arXiv preprint arXiv:2304.09218, 2023.", + "[40] Agostina J Larrazabal, Nicolás Nieto, Victoria Peterson, Diego H Milone, and Enzo Ferrante. Gender imbalance in medical imaging datasets produces biased classifiers for computer-aided diagnosis. Proceedings of the National Academy of Sciences, 117(23):12592–12594, 2020." + ], + "bbox": [ + 215, + 145, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "RadEdit: stress-testing biomedical vision models via diffusion image editing", + "bbox": [ + 279, + 114, + 732, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Choong Ho Lee and Hyung-Jin Yoon. Medical big data: promise and challenges. *Kidney research and clinical practice*, 36(1):3, 2017.", + "[42] Xiaodan Li, Yuefeng Chen, Yao Zhu, Shuhui Wang, Rong Zhang, and Hui Xue. ImageNet-e: Benchmarking neural network robustness via attribute editing, 2023.", + "[43] Wufeng Liu, Jiaxin Luo, Yan Yang, Wenlian Wang, Junkui Deng, and Liang Yu. Automatic lung segmentation in chest x-ray images using improved u-net. Scientific Reports, 12(1):8649, 2022.", + "[44] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. Advances in Neural Information Processing Systems, 33:11525-11538, 2020.", + "[45] Lena Maier-Hein, Annika Reinke, Patrick Godau, Minu D. Tizabi, Florian Buettner, Evangelia Christodoulou, Ben Glocker, Fabian Isensee, Jens Kleesiek, Michal Kozubek, Mauricio Reyes, Michael A. Riegler, Manuel Wiesenfarth, A. Emre Kavur, Carole H. Sudre, Michael Baumgartner, Matthias Eisenmann, Doreen Heckmann-Nötzel, A. Tim Rädsch, Laura Acion, Michela Antonelli, Tal Arbel, Spyridon Bakas, Arrielenis, Matthew Blaschko, M. Jorge Cardoso, Veronika Cheplygina, Beth A. Cimini, Gary S. Collins, Keyvan Farahani, Luciana Ferrer, Adrian Galdran, Bram van Ginneken, Robert Haase, Daniel A. Hashimoto, Michael M. Hoffman, Merel Huisman, Pierre Jannin, Charles E. Kahn, Dagmar Kainmueller, Bernhard Kainz, Alexandros Karargyris, Alan Karthikesalingam, Hannes Kenngott, Florian Kofler, Annette Kopp-Schneider, Anna Kreshuk, Tahsin Kurc, Bennett A. Landman, Geert Litjens, Amin Madani, Klaus Maier-Hein, Anne L. Martel, Peter Mattson, Erik Meijering, Bjoern Menze, Karel G. M. Moons, Henning Müller, Brennan Nichyporuk, Felix Nickel, Jens Petersen, Nasir Rajpoot, Nicola Rieke, Julio Saez-Rodriguez, Clara I. Sánchez, Shravya Shetty, Maarten van Smeden, Ronald M. Summers, Abdel A. Taha, Aleksei Tiulpin, Sotirios A. Tsaftaris, Ben Van Calster, Gaël Varoquaux, and Paul F. Jäger. Metrics reloaded: Recommendations for image analysis validation, 2023.", + "[46] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations, 2022.", + "[47] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Nulltext inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023.", + "[48] Gustav Müller-Franzes, Jan Moritz Niehues, Firas Khader, Soroosh Tayebi Arasteh, Christoph Haarburger, Christiane Kuhl, Tianci Wang, Tianyu Han, Sven Nebelung, Jakob Nikolas Kather, et al. Diffusion probabilistic models beat gans on medical images. arXiv preprint arXiv:2212.07501, 2022.", + "[49] Nick Pawlowski, Daniel C. Castro, and Ben Glocker. Deep structural causal models for tractable counterfactual inference. In Advances in Neural Information Processing Systems, volume 33, pages 857-869, 2020.", + "[50] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis, 2023." + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "F. Pérez-García, S. Bond-Taylor et al.", + "bbox": [ + 271, + 114, + 493, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Viraj Prabhu, Sriram Yenamandra, Prithvijit Chattopadhyay, and Judy Hoffman. LANCE: Stress-testing visual models by generating language-guided counterfactual images, 2023.", + "[52] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015.", + "[53] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021.", + "[54] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019.", + "[55] Jacob C. Reinhold, Aaron Carass, and Jerry L. Prince. A structural causal model for MR images of multiple sclerosis. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2021, volume 12905 of LNCS, pages 782–792, 2021. https://doi.org/10.1007/978-3-030-87240-3_75.", + "[56] Michael Roberts, Derek Driggs, Matthew Thorpe, Julian Gilbey, Michael Yeung, Stephan Ursprung, Angelica I Aviles-Rivero, Christian Etmann, Cathal McCague, Lucian Beer, et al. Common pitfalls and recommendations for using machine learning to detect and prognosticate for Covid-19 using chest radiographs and ct scans. Nature Machine Intelligence, 3(3):199-217, 2021.", + "[57] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022.", + "[58] Johannes Rueckel, Lena Trappmann, Balthasar Schachtner, Philipp Wesp, Boj Friedrich Hoppe, Nicola Fink, Jens Ricke, Julien Dinkel, Michael Ingrisch, and Bastian Oliver Sabel. Impact of confounding thoracic tubes and pleural dehiscence extent on artificial intelligence pneumothorax detection in chest radiographs. Investigative Radiology, 55(12):792-798, July 2020. ISSN 0020-9996. https://doi.org/10.1097/rli.0000000000000707.", + "[59] Johannes Rueckel, Christian Huemmer, Andreas Fieselmann, Florin-Cristian Gheu, Awais Mansoor, Balthasar Schachtner, Philipp Wesp, Lena Trappmann, Basel Munawwar, Jens Ricke, Michael Ingrisch, and Bastian O. Sabel. Pneumothorax detection in chest radiographs: optimizing artificial intelligence system for accuracy and confounding bias reduction using in-image annotations in algorithm training. European Radiology, 31(10):7888-7900, 2021. https://doi.org/10.1007/s00330-021-07833-w.", + "[60] Chitwan Sahara, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models, 2022.", + "[61] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Semantic foggy scene understanding with synthetic data. International Journal of Computer Vision, 126: 973-992, 2018.", + "[62] Pedro Sanchez, Antanas Kascenas, Xiao Liu, Alison Q O'Neil, and Sotirios A Tsaftaris. What is healthy? generative counterfactual diffusion for lesion localization" + ], + "bbox": [ + 215, + 145, + 787, + 840 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "RadEdit: stress-testing biomedical vision models via diffusion image editing", + "bbox": [ + 279, + 114, + 732, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 769, + 114, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tion. In MICCAI Workshop on Deep Generative Models, pages 34-44. Springer, 2022.", + "[63] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017.", + "[64] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9243–9252, 2020.", + "[65] Junji Shiraishi, Shigehiko Katsuragawa, Junpei Ikezoe, Tsuneo Matsumoto, Takeshi Kobayashi, Ken-ichi Komatsu, Mitate Matsui, Hiroshi Fujita, Yoshie Kodera, and Kunio Doi. Development of a digital image database for chest radiographs with and without a lung nodule: receiver operating characteristic analysis of radiologists' detection of pulmonary nodules. American Journal of Roentgenology, 174(1):71-74, 2000.", + "[66] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034, 2013.", + "[67] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015.", + "[68] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models, 2022.", + "[69] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019.", + "[70] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023.", + "[71] Paul Upchurch, Jacob Gardner, Geoff Pleiss, Robert Pless, Noah Snavely, Kavita Bala, and Kilian Weinberger. Deep feature interpolation for image content changes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7064-7073, 2017.", + "[72] Boris Van Breugel, Nabeel Seedat, Fergus Imrie, and Mihaela van der Schaar. Can you rely on your model evaluation? improving model evaluation with synthetic test data, 2023.", + "[73] Maria de la Iglesia Vayá, Jose Manuel Saborit, Joaquim Angel Montell, Antonio Pertusa, Aurelia Bustos, Miguel Cazorla, Joaquin Galant, Xavier Barber, Domingo Orozco-Beltrán, Francisco García-García, Marisa Caparrós, Germán González, and Jose María Salinas. BIMCV COVID-19+: a large annotated dataset of RX and CT images from COVID-19 patients, 2020. version: 3.", + "[74] I von Borzyskowski, A Mazumder, B Mateen, and M Wooldridge. Data science and ai in the age of covid-19, 2021.", + "[75] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M. Summers. ChestX-ray8: Hospital-scale chest x-ray database and" + ], + "bbox": [ + 215, + 145, + 785, + 839 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "F. Pérez-García, S. Bond-Taylor et al.", + "bbox": [ + 271, + 114, + 493, + 128 + ], + "page_idx": 19 + }, + { + "type": "ref_text", + "text": "benchmarks on weakly-supervised classification and localization of common thorax diseases. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3462-3471, 2017. https://doi.org/10.1109/CVPR.2017.369.", + "bbox": [ + 251, + 146, + 785, + 205 + ], + "page_idx": 20 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[76] Laure Wynants, Ben Van Calster, Gary S Collins, Richard D Riley, Georg Heinze, Ewoud Schuit, Elena Albu, Banafsheh Arshi, Vanessa Bellou, Marc MJ Bonten, et al. Prediction models for diagnosis and prognosis of Covid-19: systematic review and critical appraisal. bmj, 369, 2020.", + "[77] Anna Zawacki, Carol Wu, George Shih, Julia Elliott, Mikhail Fomitchev, Mohannad ParasLakhani Hussain, Phil Culliton, and Shunxing Bao. Siim-acr pneumothorax segmentation, 2019." + ], + "bbox": [ + 217, + 207, + 785, + 313 + ], + "page_idx": 20 + }, + { + "type": "header", + "text": "RadEdit: stress-testing biomedical vision models via diffusion image editing", + "bbox": [ + 279, + 114, + 730, + 128 + ], + "page_idx": 20 + }, + { + "type": "page_number", + "text": "21", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 20 + } +] \ No newline at end of file diff --git a/2024/RadEdit_ stress-testing biomedical vision models via diffusion image editing/730452c5-1675-40d8-8c46-f903c2fa2703_model.json b/2024/RadEdit_ stress-testing biomedical vision models via diffusion image editing/730452c5-1675-40d8-8c46-f903c2fa2703_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2e241e4fe95af7378c3910f27ff23155029fe00c --- /dev/null +++ b/2024/RadEdit_ stress-testing biomedical vision models via diffusion image editing/730452c5-1675-40d8-8c46-f903c2fa2703_model.json @@ -0,0 +1,3168 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.24, + 0.142, + 0.764, + 0.187 + ], + "angle": 0, + "content": "RadEdit: stress-testing biomedical vision models via diffusion image editing" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.213, + 0.776, + 0.275 + ], + "angle": 0, + "content": "Fernando Pérez-García\\*,1, Sam Bond-Taylor\\*,1, Pedro P. Sanchez+,2, Boris van Breugel+,3, Daniel C. Castro1, Harshita Sharma1, Valentina Salvatelli1, Maria T.A. Wetscherek1, Hannah Richardson1, Matthew P. Lungren1,4,5, Aditya Nori1, Javier Alvarez-Valle1, Ozan Oktay†,1, and Maximilian Ilse†,1" + }, + { + "type": "text", + "bbox": [ + 0.418, + 0.285, + 0.584, + 0.299 + ], + "angle": 0, + "content": "1 Microsoft Health Futures" + }, + { + "type": "text", + "bbox": [ + 0.422, + 0.3, + 0.582, + 0.313 + ], + "angle": 0, + "content": "2 University of Edinburgh" + }, + { + "type": "text", + "bbox": [ + 0.422, + 0.313, + 0.583, + 0.327 + ], + "angle": 0, + "content": "3 University of Cambridge" + }, + { + "type": "text", + "bbox": [ + 0.422, + 0.328, + 0.581, + 0.341 + ], + "angle": 0, + "content": "4 University of California" + }, + { + "type": "text", + "bbox": [ + 0.437, + 0.341, + 0.568, + 0.355 + ], + "angle": 0, + "content": "5 Stanford University" + }, + { + "type": "text", + "bbox": [ + 0.442, + 0.356, + 0.564, + 0.368 + ], + "angle": 0, + "content": "* Shared first author" + }, + { + "type": "list", + "bbox": [ + 0.418, + 0.285, + 0.584, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.379, + 0.369, + 0.625, + 0.382 + ], + "angle": 0, + "content": "+ Work done at Microsoft Health Futures" + }, + { + "type": "text", + "bbox": [ + 0.446, + 0.383, + 0.562, + 0.396 + ], + "angle": 0, + "content": "† Shared last author" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.438, + 0.741, + 0.646 + ], + "angle": 0, + "content": "Abstract. Biomedical imaging datasets are often small and biased, meaning that real-world performance of predictive models can be substantially lower than expected from internal testing. This work proposes using generative image editing to simulate dataset shifts and diagnose failure modes of biomedical vision models; this can be used in advance of deployment to assess readiness, potentially reducing cost and patient harm. Existing editing methods can produce undesirable changes, with spurious correlations learned due to the co-occurrence of disease and treatment interventions, limiting practical applicability. To address this, we train a text-to-image diffusion model on multiple chest X-ray datasets and introduce a new editing method, RadEdit, that uses multiple image masks, if present, to constrain changes and ensure consistency in the edited images, minimising bias. We consider three types of dataset shifts: acquisition shift, manifestation shift, and population shift, and demonstrate that our approach can diagnose failures and quantify model robustness without additional data collection, complementing more qualitative tools for explainable AI." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.66, + 0.661, + 0.674 + ], + "angle": 0, + "content": "Keywords: Image editing \\(\\cdot\\) diffusion models \\(\\cdot\\) biomedical imaging" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.701, + 0.357, + 0.716 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Developing accurate and robust models for biomedical image analysis requires large and diverse datasets that are often difficult to obtain due to ethical, legal, geographical, and financial constraints [41]. This leads to biased training datasets that affect the performance of trained models and generalisation to real-world scenarios [59, 40]. Such data mismatch may arise from genuine differences in upstream data acquisition as well as from the selection criteria for dataset creation, which materialise as various forms of dataset shifts (population, acquisition, annotation, prevalence, manifestation) [7]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.496, + 0.129 + ], + "angle": 0, + "content": "F. Pérez-García, S. Bond-Taylor et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.145, + 0.459, + 0.173 + ], + "angle": 0, + "content": "Fig.1: Stress-testing models by simulating dataset shifts via image editing." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.174, + 0.459, + 0.227 + ], + "angle": 0, + "content": "Top: editing out COVID-19 features results in false positives since the classifier relies on acquisition differences, e.g., radiographic markers (white arrow)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.228, + 0.459, + 0.269 + ], + "angle": 0, + "content": "Middle: editing out a pneumothorax (PTX) results in false positives since the classifier instead detects chest drains." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.27, + 0.459, + 0.325 + ], + "angle": 0, + "content": "Bottom: editing abnormalities into lungs causes a lung segmentation model to mislabel (blue: ground-truth segmentation; red: model prediction)." + }, + { + "type": "image", + "bbox": [ + 0.473, + 0.143, + 0.788, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.368, + 0.788, + 0.473 + ], + "angle": 0, + "content": "Biomedical vision models, when put into real-world use, can be unhelpful or potentially even harmful to patients if they are affected by dataset shifts, leading to missed diagnoses [22, 74, 76, 56]. For example, the COVID-19 pandemic led to hundreds of detection tools being developed, with some put into use in hospitals; yet Roberts et al. [56] found that \"none of the models identified are of potential clinical use due to methodological flaws and/or underlying biases.\" It is therefore crucial to properly assess models for bias, prior to real-world use." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.788, + 0.565 + ], + "angle": 0, + "content": "Recent deep generative models have made remarkable improvements in terms of sample quality, diversity, and steerability [57, 48, 35, 28]. These models have been shown to generalise to out-of-distribution domains [42, 6, 32, 19], opening up avenues for new applications. One such application is generating synthetic data for stress-testing models [51, 42, 72]. This involves creating data that is realistic, yet can represent settings, domains, or populations that do not appear (enough) in the real training/test data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.787, + 0.595 + ], + "angle": 0, + "content": "In this work, we investigate how deep generative models can be used for stress-testing biomedical imaging models. We consider three dataset shift scenarios:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.599, + 0.785, + 0.629 + ], + "angle": 0, + "content": "1. Acquisition shift: classifying COVID-19 cases when the positive and negative cases were acquired at different hospitals (Sec. 5.2)." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.63, + 0.785, + 0.659 + ], + "angle": 0, + "content": "2. Manifestation shift: detecting if pneumothorax1 was resolved when chest drains (inserted to treat pneumothorax) are present (Sec. 5.3)." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.66, + 0.785, + 0.69 + ], + "angle": 0, + "content": "3. Population shift: segmenting lungs in the presence of abnormalities rarely or never seen in the training dataset (Sec. 5.4)." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.599, + 0.785, + 0.69 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.695, + 0.788, + 0.817 + ], + "angle": 0, + "content": "For each of these scenarios, we simulate dataset shifts, producing stress-test sets which can occur in the real world but do not appear or are underrepresented in the original training/test sets. Following prior work, these test sets are synthesised using generative image editing, which unlike generating images from scratch, only minimally modifies the images, hence, better retains fidelity and diversity [51, 42]. For the above scenarios, we use generative editing to 1. remove only COVID-19 features while keeping visual indicators of the different hospitals; 2. remove only pneumothorax while keeping the chest drain; and 3. add abnormalities that occlude lung structures in the image." + }, + { + "type": "page_footnote", + "bbox": [ + 0.239, + 0.825, + 0.768, + 0.841 + ], + "angle": 0, + "content": "1 We provide descriptions of the medical terms used throughout the paper in Appendix A." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.28, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "RadEdit: stress-testing biomedical vision models via diffusion image editing" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.147, + 0.786, + 0.403 + ], + "angle": 0, + "content": "We train a diffusion model [27, 57] on a large collection of chest X-rays from a variety of biomedical imaging datasets (Sec. 5.1), enabling us to add and remove a wide variety of pathologies and support devices when editing. Despite the diversity within these datasets, substantial biases are still present, some of which are learned by the generative model; as a result, correlated features may also be modified. For example, in Scenario 2, removing the pneumothorax might also remove the chest drains as both features typically co-occur in datasets [58], since chest drains are used to treat pneumothorax. Furthermore, when editing only within editing masks, artefacts often appear at the border of the masks. Lastly, artefacts occur when editing images outside of the training dataset domain of the diffusion model used for editing. To overcome these challenges, we propose using multiple masks to break existing correlations. This involves defining which regions must change, and explicitly forcing correlated regions to remain unchanged. In addition, we allow the area outside of the masks to be modified by the diffusion model to ensure image consistency. Since our proposed editing method, which we call RadEdit, leads to only minimal overall changes of chest X-rays, we are able to generate synthetic datasets that can be used to stress-test segmentation models (Scenario 3), which, to the best of our knowledge, we are the first to demonstrate." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.405, + 0.54, + 0.418 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.42, + 0.784, + 0.448 + ], + "angle": 0, + "content": "- We introduce a novel editing approach that reduces the presence of artefacts in edited images and simplifies prompt construction compared to prior work [10, 51]." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.45, + 0.784, + 0.477 + ], + "angle": 0, + "content": "- Our editing approach allows us to construct synthetic datasets with specific data shifts by performing zero-shot edits on datasets/abnormalities not seen in training." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.48, + 0.784, + 0.522 + ], + "angle": 0, + "content": "- We conduct a broad set of experiments using these synthetic datasets to stress-test and expose biases in biomedical classification and, for the first time, segmentation models, introducing a new use case of synthetic data into the medical setting." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.42, + 0.784, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.545, + 0.362, + 0.56 + ], + "angle": 0, + "content": "2 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.574, + 0.786, + 0.619 + ], + "angle": 0, + "content": "In this section, we introduce background context for stress-testing biomedical imaging models: failure modes of biomedical imaging models caused by different dataset shifts; diffusion models as versatile generative models; and diffusion-based image editing." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.639, + 0.351, + 0.653 + ], + "angle": 0, + "content": "2.1 Dataset shifts" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.662, + 0.786, + 0.767 + ], + "angle": 0, + "content": "Dataset shift refers to a discrepancy between the training and test data distributions due to external factors [7, 34]. Such shifts are regularly observed in machine learning for biomedical imaging, often due to data scarcity. For example, collected training datasets might consist primarily of healthy patients. However, when the model is used in practice after training, there could be a shift towards unhealthy patients. A taxonomy of different types of dataset shifts in the context of biomedical imaging was developed by Castro et al. [7]. In this paper, we consider three dataset shifts of particular interest." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.78, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Acquisition shift results from the use of different scanners (manufacturer, hardware, and software) or imaging protocols as often encountered when using data from multiple cohorts. These changes affect factors such as image resolution, contrast, patient positioning, and image markings." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.496, + 0.129 + ], + "angle": 0, + "content": "F. Pérez-García, S. Bond-Taylor et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "Manifestation shift results from the way the prediction targets physically manifest in anatomy changes between domains. For example, training datasets could consist of more severe pathological cases than observed in practice, or a pathology may co-occur with different visual features, e.g., support devices." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.218, + 0.788, + 0.295 + ], + "angle": 0, + "content": "Population shift results from differences in intrinsic characteristics of the populations under study, changing the anatomical appearance distribution. This definition encompasses examples such as age, sex, ethnicity, and comorbidities, but also abnormalities such as pleural effusion and support devices. In contrast to manifestation shift, the shift in anatomical appearance is not affected by prediction targets." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.311, + 0.376, + 0.325 + ], + "angle": 0, + "content": "2.2 Diffusion models" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.331, + 0.787, + 0.392 + ], + "angle": 0, + "content": "Denoising diffusion probabilistic models (DDPMs) [27, 67] are a versatile and effective class of generative models that enable sampling from the data distribution by learning to denoise samples corrupted with Gaussian noise. DDPMs are formed by defining a forward time process that gradually adds noise to data points \\( x_0 \\) through the recursion" + }, + { + "type": "equation", + "bbox": [ + 0.226, + 0.397, + 0.786, + 0.416 + ], + "angle": 0, + "content": "\\[\nx _ {t} = \\sqrt {1 - \\beta_ {t}} x _ {t - 1} + \\sqrt {\\beta_ {t}} \\epsilon_ {t}, \\quad t = 1, \\dots , T, \\quad \\text {s . t .} x _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\bar {\\epsilon} _ {t}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.42, + 0.788, + 0.497 + ], + "angle": 0, + "content": "where \\(\\epsilon_{1:T},\\bar{\\epsilon}_{1:T}\\sim \\mathcal{N}(0,I),\\beta_{1:T}\\) is a predefined noise schedule that determines how quickly to corrupt the data and ensures that \\(x_{T}\\) contains little to no information about \\(x_0\\), and \\(\\bar{\\alpha}_t = \\prod_{s = 1}^t (1 - \\beta_s)\\). To form a generative model, the process is reversed in time, gradually transforming Gaussian noise into samples from the learned distribution. While the exact reversal is intractable, a variational approximation is defined by [68]:" + }, + { + "type": "equation", + "bbox": [ + 0.387, + 0.504, + 0.786, + 0.52 + ], + "angle": 0, + "content": "\\[\nx _ {t - 1} = \\hat {\\mu} _ {t} \\left(x _ {t}, f _ {\\theta} \\left(x _ {t}, t, c\\right)\\right) + \\sigma_ {t} z _ {t}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.524, + 0.786, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mu} _ {t} \\left(x _ {t}, \\epsilon_ {t}\\right) = \\sqrt {\\bar {\\alpha} _ {t - 1}} \\frac {x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t}}{\\sqrt {\\bar {\\alpha} _ {t}}} + \\sqrt {1 - \\bar {\\alpha} _ {t - 1} - \\sigma_ {t} ^ {2}} \\epsilon_ {t}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.561, + 0.788, + 0.638 + ], + "angle": 0, + "content": "where \\(c\\) is a conditioning signal such as a text description, \\(f_{\\theta}(x_t,t,c)\\) is a learned approximation of the noise \\(\\bar{\\epsilon}_t\\) that corrupted the image \\(x_0\\) to obtain \\(x_{t}\\), \\(z_{1:T} \\sim \\mathcal{N}(0,I)\\), and \\(\\sigma_{1:T}\\) controls how much noise is introduced. The process is Markovian and known as a DDPM [27] when \\(\\sigma_t = \\sqrt{(1 - \\bar{\\alpha}_{t - 1}) / (1 - \\bar{\\alpha}_t)}\\sqrt{1 - \\bar{\\alpha}_t / \\bar{\\alpha}_{t - 1}}\\), while for \\(\\sigma_t = 0\\) the process is deterministic and is called a denoising diffusion implicit model (DDIM) [68]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.654, + 0.356, + 0.669 + ], + "angle": 0, + "content": "2.3 Image editing" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Since DDIMs have a one-to-one correspondence with latent vectors \\( x_{T} \\), we can deterministically map data points to latents by running the DDIM generative process in reverse [68], called DDIM inversion. Several approaches [10, 46] have shown that images can be edited by running the reverse diffusion process augmented by the latent vectors and a modified prompt \\( c \\). However, this method can lead to undesired artefacts in the edited images. For example, structures unrelated to the desired edit may also change shape, size, or location. To address this, Huberman-Spiegelglas et al. [29] propose DDPM inversion. Here, the original forward process defined in Eq. (1) is adapted, replacing the correlated vectors \\( \\bar{\\epsilon}_{1:T} \\) with statistically independent vectors \\( \\tilde{\\epsilon}_{1:T} \\) (more details in Appendix B). These noise vectors are then used in the generative process, retaining the structure of the original image better than DDIM inversion." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.281, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RadEdit: stress-testing biomedical vision models via diffusion image editing" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.362, + 0.161 + ], + "angle": 0, + "content": "3 Related work" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.176, + 0.431, + 0.192 + ], + "angle": 0, + "content": "3.1 Generative image editing" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.201, + 0.786, + 0.277 + ], + "angle": 0, + "content": "With advances in deep generative modelling, several approaches to image editing have emerged. Many of these early approaches use compressed latent manipulation [12, 52, 64, 71] where fine-grained edits are difficult to achieve and can result in unwanted changes. More recently, the unparalleled flexibility of diffusion models, together with advances in plain text conditioning, have opened up new avenues for editing techniques." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.278, + 0.786, + 0.413 + ], + "angle": 0, + "content": "Here, we describe some notable diffusion editing methods. SDEdit [46] shows that diffusion models trained solely on real images can be used to generate images from sketches by perturbing sketches with noise, then running the reverse diffusion process. Palette [60] is a diffusion model trained for inpainting by filling regions with noise and denoising. Blended diffusion [2, 3] uses masks with CLIP [53] conditioning to guide local edits. Multiple works show that injecting U-Net activations, obtained by encoding the original image into the generation process, better retains image structure [24, 70]. DiffEdit [10] uses text prompts to determine the appropriate region to edit. Mokady et al. [47] improve diffusion inversion quality by optimising the diffusion trajectory." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.414, + 0.786, + 0.473 + ], + "angle": 0, + "content": "Crucially, in the works which use masks for editing, a single type of mask is always used to define the region of interest. In this work, we argue that a second type of mask is required to avoid the loss of features caused by spurious correlations. As better editing approaches are developed, this requirement should be kept in mind." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.495, + 0.506, + 0.512 + ], + "angle": 0, + "content": "3.2 Biomedical imaging counterfactuals" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.521, + 0.786, + 0.671 + ], + "angle": 0, + "content": "Generative models have also been applied to biomedical counterfactual generation. Reinhold et al. [55] manipulate causes of multiple sclerosis in brain MRI with deep structural causal models [49]. Sanchez et al. [62] and Fontanella et al. [15] use editing to remove pathologies for abnormality detection. Ktena et al. [39] generate out-of-distribution samples to improve classifier performance. Gu et al. [21] train a diffusion model to model disease progression by conditioning on a prior X-ray and text progression description. Unlike our approach, these methods do not use masks to enforce which regions may or may not be edited, meaning that spurious correlations might affect edits. Additionally, these methods use synthetic data to augment and improve model performance whereas we focus on using synthetic medical data for stress-testing." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.693, + 0.35, + 0.71 + ], + "angle": 0, + "content": "3.3 Stress-testing" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.72, + 0.786, + 0.794 + ], + "angle": 0, + "content": "Several approaches have used non-deep-generative-model methods to stress-test networks. Hendrycks and Dietterich [23] evaluate classification models' robustness to corruptions such as blurring, Gaussian noise, and JPEG artefacts. Sakaridis et al. [61] stress-test a segmentation model for roads by using an optical model to add synthetic fog to scenes. Koh et al. [38] collate a dataset presenting various distribution shifts." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.796, + 0.786, + 0.84 + ], + "angle": 0, + "content": "More recent models have made use of conditional generative models to simulate shifts. Prabhu et al. [51] propose LANCE, which stress-tests classification models by using diffusion-based image editing to modify image subjects via caption editing with" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.117, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.496, + 0.129 + ], + "angle": 0, + "content": "F. Pérez-García, S. Bond-Taylor et al." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.238 + ], + "angle": 0, + "content": "a large language model (LLM); Kattakinda et al. [36] do similar, but instead modify the background. Li et al. [42] use diffusion models with a single subject mask to separately edit backgrounds and subjects. Van Breugel et al. [72] use generative adversarial networks to simulate distribution shifts on tabular data. This line of research is partially related to adversarial attacks [20], where the focus is on minimally modifying images such that they are visually indistinguishable to a human, but the attacked model fails." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.26, + 0.667, + 0.275 + ], + "angle": 0, + "content": "3.4 Limitations of existing editing-based stress-testing methods" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.285, + 0.787, + 0.33 + ], + "angle": 0, + "content": "Recent advancements in diffusion modelling have drastically improved image editing. However, two prevalent approaches, LANCE [51] and DiffEdit [10], produce artefacts in medical images, making them unsuitable for stress-testing biomedical vision models." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.331, + 0.787, + 0.481 + ], + "angle": 0, + "content": "LANCE only uses a global prompt (no mask) for image editing. While effective in the natural image domain, it leads to artefacts in the biomedical domain. For example in Sec. 5.4, we add pathologies and support devices to images of healthy lungs to stress-test lung segmentation models; in such cases we must ensure that the position and shape of the lung borders are not altered during editing. However, we find that LANCE changes the position and shape of the lung border thus making edited images unsuitable for stress-testing such models (Sec. 5.5). In addition, we find that LANCE potentially removes support devices when prompted to remove pathologies, a direct effect of the correlations in the diffusion model's training datasets (Sec. 5.1), making LANCE unsuited for testing the robustness of models to manifestation shift." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.482, + 0.787, + 0.602 + ], + "angle": 0, + "content": "DiffEdit addresses these issues by editing only inside an automatically predicted mask \\( m_{\\mathrm{edit}} \\). However, these predicted masks often mismatch the manually annotated ground-truth, especially for small and complex abnormalities like pneumothorax1 (see Sec. 5.5). Moreover, spurious correlations learned by the diffusion model can lead to the inclusion of support devices in the automatically predicted masks. Furthermore, even when relying on manually annotated masks, DiffEdit can introduce sharp discrepancies at mask boundaries, leading to unrealistic artefacts, such as when adding consolidation that should partially occlude the lung border (Fig. 10b in the Appendix)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.625, + 0.315, + 0.64 + ], + "angle": 0, + "content": "4 Method" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.657, + 0.787, + 0.703 + ], + "angle": 0, + "content": "Our objective is to create synthetic test data through image editing that simulates specific data shifts, to rigorously evaluate biomedical imaging models. This synthetic data is used to predict model robustness, eliminating need for additional real-world test data." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.724, + 0.475, + 0.74 + ], + "angle": 0, + "content": "4.1 Improved editing with RadEdit" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To address the issues outlined in Sec. 3.4, we propose RadEdit: by introducing 'keep' and 'edit' masks into the editing process, RadEdit explicitly specifies which areas must remain unchanged (keep) and which should be actively modified based on the conditioning signal (edit). Crucially, these masks need not be mutually exclusive, allowing changes in the unmasked regions to ensure global consistency. Using masks, we assume that spurious correlations are mostly non-overlapping [44]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.28, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "RadEdit: stress-testing biomedical vision models via diffusion image editing" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "code_caption", + "bbox": [ + 0.217, + 0.146, + 0.707, + 0.161 + ], + "angle": 0, + "content": "Algorithm 1 RadEdit (ours) uses multiple masks to decouple spurious correlations" + }, + { + "type": "algorithm", + "bbox": [ + 0.217, + 0.165, + 0.787, + 0.334 + ], + "angle": 0, + "content": "Require: original image \\(x_0\\) , inversion prompt \\(c_{\\mathrm{inv}}\\) , editing prompt c, edit mask medit, keep mask \\(m_{\\mathrm{keep}}\\) CFG weight w, diffusion model \\(f_{\\theta}\\) \\((\\hat{x}_{1:T},z_{1:T})\\leftarrow\\) DDPMINVERSION(xo,cinv) \\(\\triangleright\\) Encode image. Procedure in Appendix B \n\\(x_{T}\\gets \\hat{x}_{T}\\) \nfor \\(t\\gets T\\) to 1 do \n\\(\\epsilon_{\\mathrm{cond},t}\\gets f_{\\theta}(xt,t,c)\\) \\(\\triangleright\\) Predict conditional noise \n\\(\\epsilon_{\\mathrm{uncond},t}\\gets f_{\\theta}(xt,t,c = \\emptyset)\\) \\(\\triangleright\\) Predict unconditional noise \n\\(\\epsilon_t\\gets \\epsilon_{\\mathrm{uncond},t} + w(\\epsilon_{\\mathrm{cond},t} - \\epsilon_{\\mathrm{uncond},t})\\) \\(\\triangleright\\) Combine noise predictions with CFG \n\\(\\epsilon_t\\gets m_{\\mathrm{edit}}\\odot \\epsilon_t + (1 - m_{\\mathrm{edit}})\\odot \\epsilon_{\\mathrm{uncond},t}\\) \\(\\triangleright\\) Use CFG only within medit \n\\(x_{t - 1}\\gets \\hat{\\mu}_t(x_t,\\epsilon_t) + \\sigma_t z_t\\) \\(\\triangleright\\) Move to next time step \n\\(x_{t - 1}\\gets m_{\\mathrm{keep}}\\odot \\hat{x}_{t - 1} + (1 - m_{\\mathrm{keep}})\\odot x_{t - 1}\\) \\(\\triangleright\\) Undo edits within mkeep \nreturn edited version of \\(x_0\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.361, + 0.788, + 0.527 + ], + "angle": 0, + "content": "RadEdit is detailed in Algorithm 1, where a number of key properties make RadEdit more suitable for biomedical image editing than prior editing methods. Firstly, since we aim to edit only within the edit mask \\( m_{\\mathrm{edit}} \\), classifier-free guidance (CFG) [26] is used only within this region, with high guidance values (following [29], we use a value of 15) ensuring that pathologies are completely removed without drastically changing the rest of the image. This approach also simplifies choosing a prompt for editing since we do not have to take into account the effect of the prompt on the rest of the image. Secondly, we allow the area outside \\( m_{\\mathrm{edit}} \\) to be modified via unconditional generation to ensure image consistency. Lastly, the region of the keep mask \\( m_{\\mathrm{keep}} \\) is reverted to the encoding, ensuring that this region remains the same. Instead of initiating our generating process from pure noise we set \\( x_{T} = \\hat{x}_{T} \\), where \\( \\hat{x}_T \\) is the last output of the DDPM inversion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.528, + 0.788, + 0.618 + ], + "angle": 0, + "content": "In Fig. 3c, 10c, we show that RadEdit enables artefact-free editing while preserving structures of interest. Because the anatomical layout remains intact after editing, masks still correspond to the same structures, therefore the same masks can be reused to stress-test segmentation models (Sec. 5.4). In practice, we use a latent diffusion model [57], therefore all operations in Algorithm 1 are performed in the latent space of a variational autoencoder (VAE) [57]; this does not limit the generality of the approach." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.637, + 0.527, + 0.652 + ], + "angle": 0, + "content": "4.2 Using synthetic images to uncover bias" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Despite advancements in biomedical computer vision, recent studies have shown that bias in training and test data can lead to unrealistically high performance of machine learning models on the test set [59, 11]. In our experiments, we use RadEdit to create high quality synthetic test datasets that realistically capture specific dataset shifts, allowing us to quantify the robustness of models to these dataset shifts. By using masks, we can precisely edit the original training data to represent either acquisition shift, manifestation shift, or population shift [7] (Secs. 5.2 to 5.4). These synthetic test sets are used to stress-test (potentially biased) biomedical vision models by comparing performance to the real (biased) test set; a significant drop in performance indicates that the vision model is not robust to the dataset shift that can occur in clinical settings. This serves as a complementary tool to visual explainable AI tools like Grad-CAM [63] and saliency maps [66, 1], which offer qualitative insight into the robustness of models." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.496, + 0.129 + ], + "angle": 0, + "content": "F. Pérez-García, S. Bond-Taylor et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.418, + 0.162 + ], + "angle": 0, + "content": "4.3 BioViL-T editing score" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.788, + 0.279 + ], + "angle": 0, + "content": "Since generative models result in samples of varying quality, poor-quality samples can be filtered out using image-text alignment scores, which quantitatively assess how closely related image-text pairs are via a pre-trained model that embeds similar images and text to nearby vectors [4, 54, 53, 14]. For image editing, we instead assess how similar the change in text and image embeddings are after editing: for a real image-text pair \\((I_{\\mathrm{real}}, T_{\\mathrm{real}})\\), edited image-text pair \\((I_{\\mathrm{edit}}, T_{\\mathrm{edit}})\\), image encoder \\(E_I\\), and text encoder \\(E_T\\), the editing score is defined based on directional similarity [17]:" + }, + { + "type": "equation", + "bbox": [ + 0.268, + 0.289, + 0.786, + 0.325 + ], + "angle": 0, + "content": "\\[\nS _ {\\mathrm {B i o V i L - T}} = \\frac {\\Delta I \\cdot \\Delta T}{\\| \\Delta I \\| \\| \\Delta T \\|}, \\quad \\text {w h e r e} \\quad \\begin{array}{l} \\Delta I = E _ {I} \\left(I _ {\\mathrm {e d i t}}\\right) - E _ {I} \\left(I _ {\\mathrm {r e a l}}\\right), \\text {a n d} \\\\ \\Delta T = E _ {T} \\left(T _ {\\mathrm {e d i t}}\\right) - E _ {T} \\left(T _ {\\mathrm {r e a l}}\\right). \\end{array} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.335, + 0.788, + 0.428 + ], + "angle": 0, + "content": "Given the focus on biomedical data, we use the BioViL-T [5] image and text encoders: domain-specific vision-language models trained to analyse chest X-rays and radiology reports, therefore well suited to measure changes in the edited image, such as removed pathologies. Following Prabhu et al. [51], we discard images with \\( S_{\\mathrm{BioViL - T}} < 0.2 \\). This is not only effective for filtering out poor quality edits but is also able to detect whether the original image \\( I_{\\mathrm{real}} \\) does not match the original text description \\( T_{\\mathrm{real}} \\) well." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.45, + 0.357, + 0.466 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.481, + 0.37, + 0.495 + ], + "angle": 0, + "content": "5.1 Diffusion model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.788, + 0.596 + ], + "angle": 0, + "content": "Our editing method is heavily dependent on a latent diffusion model [57] that can generate realistic chest X-rays. We use the VAE [37, 25] of SDXL [50] which can adequately reconstruct chest X-rays [8]. The VAE is frozen, and the denoising U-Net is trained on three datasets downsampled and centre-cropped to \\(512 \\times 512\\) pixels: MIMIC-CXR [33], ChestX-ray8 [75], and CheXpert [30], totalling 487680 training images. This data diversity allows us to perform zero-shot edits on datasets not seen during training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.788, + 0.703 + ], + "angle": 0, + "content": "For MIMIC-CXR, we only include frontal view chest X-rays, and condition the denoising U-Net on the corresponding impression section in the radiology report (a short clinically actionable outline of the main findings). We employ the tokeniser and frozen text encoder from BioViL-T [5]. For ChestX-ray8 and CheXpert, we condition on a list of all abnormalities present in an image as indicated by the labels, e.g., 'Cardiomegaly. Pneumothorax'. If the list of abnormalities is empty, we use the string 'No findings'. More details on the datasets and diffusion model training can be found in Appendix E." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.725, + 0.373, + 0.74 + ], + "angle": 0, + "content": "5.2 Acquisition shift" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Background To show how RadEdit can be used to quantify the robustness of models to acquisition shift, we closely follow the experimental setup of DeGrave et al. [11], who show that deep learning systems built to detect COVID-19 from chest X-rays rely on confounding factors rather than pathology features. This problem arises when COVID-19-positive and -negative images come from disparate sources. In our setup, all COVID-19-positive cases come from the BIMCV dataset [73] (denoted BIMCV+)," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.28, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "RadEdit: stress-testing biomedical vision models via diffusion image editing" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.145, + 0.396, + 0.268 + ], + "angle": 0, + "content": "Fig.2: Removing COVID-19 features with LANCE\\(^2\\) (b) also changes the laterality markers and reduces contrast. In contrast, RadEdit (c; ours) preserves anatomical structures and laterality markers, and retains the original contrast." + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.15, + 0.536, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.253, + 0.514, + 0.264 + ], + "angle": 0, + "content": "(a) Original Image" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.15, + 0.657, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.555, + 0.253, + 0.634, + 0.264 + ], + "angle": 0, + "content": "(b) LANCE [51]" + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.15, + 0.785, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.253, + 0.762, + 0.264 + ], + "angle": 0, + "content": "(c) RadEdit (ours)" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.299, + 0.784, + 0.346 + ], + "angle": 0, + "content": "and all COVID-19-negative cases from MIMIC-CXR [33] (see Fig. 11). A classifier trained on these datasets will rely on spurious features indicative of the data's origin, e.g., laterality markers or the field of view, instead of features caused by the pathology." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.36, + 0.788, + 0.48 + ], + "angle": 0, + "content": "Setup A synthetic test set is created by applying RadEdit to remove COVID-19 features1 from BIMCV+ images using the prompt 'No acute cardiopulmonary process'3 (Fig. 2); the included bounding boxes of COVID-19 features are used as the edit mask \\( m_{\\mathrm{edit}} \\). Since this is the only mask available, we set the keep mask as \\( m_{\\mathrm{keep}} = 1 - m_{\\mathrm{edit}} \\). After filtering using the BioViL-T editing (Sec. 4.3), this results in a synthetic dataset of 2774 COVID-19-negative images containing the same spurious features as BIMCV+. Both the weak and strong predictor models are classifiers trained using a ResNet50 architecture (see Appendix F for more implementation details)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.495, + 0.788, + 0.601 + ], + "angle": 0, + "content": "Findings Tab. 1, shows the performance of a COVID-19 classifier (weak predictor) trained on BIMCV+ and MIMIC-CXR. In accordance with DeGrave et al. [11], we find that the weak predictor performs exceptionally well on the real test set (i.e. test splits of both datasets) since the model learned to distinguish the two data sources instead of learning visual features related to COVID-19. However, in the second row of Tab. 1, we see a drop of \\(95\\%\\) in accuracy meaning that the model fails to classify the synthetic images as COVID-19-negative. The weak predictor is not robust to a shift in acquisition." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.602, + 0.788, + 0.661 + ], + "angle": 0, + "content": "To show that the decreased performance of the weak predictor is not caused by artefacts in the edited images, we train a more robust COVID-19 classifier (strong predictor), using the BIMCV+ and BIMCV- datasets, as in [11], where the BIMCV- dataset consists of only COVID-19-negative cases from BIMCV, and test on the same two test" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.688, + 0.584, + 0.799 + ], + "angle": 0, + "content": "Table 1: Quantifying robustness of COVID-19 detectors to acquisition shift. We train a weak predictor on the 'Biased' dataset—a combination of BIMCV+ [73] and MIMIC-CXR [33]; and a strong predictor on an unbiased dataset—a combination of BIMCV+ and BIMCV; the 'Synthetic' test set consists of 2774 COVID-19-negative images with the same spurious features as the BIMCV+ datasets, e.g. laterality markers. We report mean accuracy and standard deviation across 5 runs." + }, + { + "type": "table", + "bbox": [ + 0.594, + 0.702, + 0.785, + 0.792 + ], + "angle": 0, + "content": "
PredictorTest dataAccuracy
WeakBiased99.1 ± 0.2
WeakSynthetic5.5 ± 2.1
StrongBiased74.4 ± 3.0
StrongSynthetic76.0 ± 7.7
" + }, + { + "type": "page_footnote", + "bbox": [ + 0.239, + 0.811, + 0.581, + 0.825 + ], + "angle": 0, + "content": "2For LANCE, we perform the text perturbation manually." + }, + { + "type": "page_footnote", + "bbox": [ + 0.241, + 0.825, + 0.651, + 0.84 + ], + "angle": 0, + "content": "This is a common radiological description of a 'normal' chest X-ray." + }, + { + "type": "list", + "bbox": [ + 0.239, + 0.811, + 0.651, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.495, + 0.129 + ], + "angle": 0, + "content": "F. Pérez-García, S. Bond-Taylor et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.145, + 0.409, + 0.255 + ], + "angle": 0, + "content": "Fig.3: Removing pneumothorax (red) with LANCE\\(^2\\) (b) also removes the spuriously correlated chest drain (blue) and reduces contrast. In contrast, RadEdit (c; ours) preserves the chest drain and better preserves anatomical structures." + }, + { + "type": "image", + "bbox": [ + 0.416, + 0.146, + 0.536, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.43, + 0.247, + 0.519, + 0.259 + ], + "angle": 0, + "content": "(a) Original Image" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.147, + 0.657, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.557, + 0.247, + 0.637, + 0.259 + ], + "angle": 0, + "content": "(b) LANCE [51]" + }, + { + "type": "image", + "bbox": [ + 0.657, + 0.147, + 0.785, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.247, + 0.762, + 0.259 + ], + "angle": 0, + "content": "(c) RadEdit (ours)" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.291, + 0.784, + 0.366 + ], + "angle": 0, + "content": "datasets. Comparing rows one and three of Tab. 1, we find that the strong predictor performs worse on the test set containing samples from BIMCV+ and MIMIC-CXR than the weak predictor (row one). This is expected as the strong predictor relies on actual pathology features. Lastly, rows three and four of Tab. 1 show that the strong predictor performs similarly on the real and synthetic test sets, attesting to the quality of our edits." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.389, + 0.389, + 0.403 + ], + "angle": 0, + "content": "5.3 Manifestation shift" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.415, + 0.785, + 0.552 + ], + "angle": 0, + "content": "Background In this section, we show how RadEdit can be used to quantify the robustness of biomedical vision models to manifestation shift. We closely follow the experimental setup of Rueckel et al. [59], who demonstrate that pneumothorax classification models are strongly biased by the presence of chest drains: while the average performance of pneumothorax classifiers is high, performance on the subset of images with a chest drain but no pneumothorax is significantly lower. This is due to chest drains being a common treatment for pneumothorax, resulting in the majority of images in datasets like CANDID-PTX [13] containing a chest drain only if there is a pneumothorax. As a result, only \\(1\\%\\) of images in CANDID-PTX contain a chest drain but no pneumothorax." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.568, + 0.785, + 0.72 + ], + "angle": 0, + "content": "Setup We use RadEdit to create a synthetic dataset containing images with chest drains but no pneumothorax, by editing out the pneumothorax from CANDID-PTX images using the prompt 'No acute cardiopulmonary process' (Fig. 3). The edit mask \\( m_{\\mathrm{edit}} \\) is set as a mask of the pneumothorax, and the keep mask \\( m_{\\mathrm{keep}} \\) set as the chest drain mask. This ensures that the chest drain is not removed, while preventing border artefacts. After filtering using the BioViL-T editing score (Sec. 4.3), 628 images are left; in contrast, the real test set contains only 16 of cases with drains but no pneumothorax. The weak predictor is a ResNet50 trained to classify pneumothorax. Following Rueckel et al. [59], the strong predictor is an EfficientNet U-Net [69] trained on SIIM-ACR [77] to both segment pneumothorax and classify pneumothorax (more details in Appendix G)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.735, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Findings In accordance with [59], we show in Tab. 2 that a pneumothorax classifier (weak predictor) trained on CANDID-PTX performs exceptionally well on the test split of CANDID-PTX, since very few images contain a chest drain and no pneumothorax. However, in row two of Tab. 2, we show a drastic drop in performance on the synthetic test set, i.e., the weak predictor is not robust to manifestation shift. To show that the drop in performance on the synthetic dataset does not come from editing artefacts, we evaluate the more robust segmentation model (strong predictor) on the same test datasets" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.28, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "RadEdit: stress-testing biomedical vision models via diffusion image editing" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.158, + 0.579, + 0.296 + ], + "angle": 0, + "content": "Table 2: Quantifying robustness of pneumothorax detectors to manifestation shift. The weak predictor is trained on the biased CANDID-PTX [13] dataset to classify pneumothorax; the strong predictor is trained on SIIM-ACR [77] to classify and segment the pneumothorax. Real 'Biased' test data comes from CANDID-PTX which exhibits strong confounding between the pneumothorax and chest tubes; 'Synthetic' test data is 629 solely edited images containing chest drains but no pneumothorax. We report mean accuracy and standard deviation across 5 runs." + }, + { + "type": "table", + "bbox": [ + 0.596, + 0.175, + 0.79, + 0.269 + ], + "angle": 0, + "content": "
PredictorTest dataAccuracy
WeakBiased93.3 ± 0.6
WeakSynthetic17.9 ± 3.7
StrongBiased93.7 ± 1.3
StrongSynthetic81.7 ± 7.1
" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.325, + 0.788, + 0.402 + ], + "angle": 0, + "content": " Rows three and four of Tab. 2), we find that the strong predictor performs on par with the weak predictor in row one; however, the strong predictor closes the majority of the gap between the real test set and the synthetic one, attesting to the quality of our edits. In agreement with Rueckel et al. [59], there is still a performance gap, indicating that the strong predictor still suffers from mild manifestation shift." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.422, + 0.371, + 0.437 + ], + "angle": 0, + "content": "5.4 Population shift" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.446, + 0.788, + 0.553 + ], + "angle": 0, + "content": "Background In this section, we show how RadEdit can be used to quantify the robustness of lung segmentation models to population shifts. Manually segmenting X-ray images is labour intensive and requires high expertise, leading to small datasets often limited to single pathologies or healthy patients [65, 31], e.g., MIMIC-Seg [9]. These models are thus sensitive to occlusions such as medical devices or pathologies, which typically appear as white regions on X-rays [43]. Evaluating model robustness requires further image collection for each occlusion type, which is time-consuming and costly." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.565, + 0.788, + 0.597 + ], + "angle": 0, + "content": "Setup Here, abnormalities are added to the lung region in healthy X-rays from MIMIC-Seg (Fig. 4). The edit mask \\( m_{\\mathrm{edit}} \\) is set as a mask of the lung(s). When editing a single" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.622, + 0.449, + 0.83 + ], + "angle": 0, + "content": "Fig.4: Adding pulmonary edema (top), pacemakers (middle), and consolidation (bottom) with RadEdit. The 'strong predictor' (d), a segmentation model trained on CheXmask [16] (a large dataset containing various abnormalities) is more robust to these abnormalities than the 'weak predictor' (c), a segmentation model trained on MIMIC-Seg [9] (a small set of mostly healthy patients): the weak predictor traces around the pacemaker and poorly annotates the consolidated lung. Blue: ground-truth annotation; red: predicted segmentation." + }, + { + "type": "image", + "bbox": [ + 0.466, + 0.622, + 0.788, + 0.834 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.495, + 0.129 + ], + "angle": 0, + "content": "F. Pérez-García, S. Bond-Taylor et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Table 3: Quantifying robustness of lung segmentation models to population shift. The 'weak predictor' is trained on MIMIC-Seg (a small set of predominantly healthy patients); the 'strong predictor' is trained on CheXmask (a large mixed set of patients with various abnormalities). Synthetic test data is created by using RadEdit to add edema, pacemakers, and consolidation. We report the change \\((\\Delta)\\) in Dice score and AHD with respect to the segmentation models evaluated on the ground-truth test set." + }, + { + "type": "table", + "bbox": [ + 0.252, + 0.24, + 0.75, + 0.356 + ], + "angle": 0, + "content": "
Test dataWeak PredictorStrong Predictor
Dice↑Δ↓AHD↓Δ↓Dice↑Δ↓AHD↓Δ↓
Real data97.4—6.1—95.5—11.6—
Healthy → edema93.83.621.815.793.91.622.811.2
Healthy → pacemaker85.012.449.843.787.38.229.517.9
Healthy → consolidation85.911.544.138.188.17.429.417.8
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.383, + 0.788, + 0.534 + ], + "angle": 0, + "content": "lung, the keep mask \\( m_{\\mathrm{keep}} \\) corresponds to the lung which must not change, while when editing both lungs we set \\( m_{\\mathrm{keep}} = 0 \\) to allow opacity adjustments, or for elements to be added outside of the lungs. Stress-test sets are generated for three abnormalities: pulmonary edema, pacemakers, and consolidation1. Prompts are phrased to match similar impressions in the training data (see Appendix H). Both predictors are EfficientNet U-Net [69] models; the weak predictor is trained on MIMIC-Seg, while the strong predictor is trained on CheXmask [16], a larger dataset with various lung abnormalities (more details in Appendix H). We evaluate segmentation quality using Dice similarity coefficient, which is the harmonic mean of the precision and recall, and 95th percentile average Hausdorff distance (AHD), a measure of the distance between two sets [45]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.788, + 0.716 + ], + "angle": 0, + "content": "Findings Tab. 3 shows that the weak predictor model performs well on the real biased test data, since it is mostly composed of healthy subjects. However, testing on the synthetic lung abnormality datasets (rows two to four), causes performance to drop substantially, i.e. the weak predictor is not robust to population shift. To show that this drop in performance does not come from editing artefacts, we evaluate the strong predictor on the synthetic test sets and see considerably smaller changes in performance. This can be seen in Fig. 4: for pulmonary edema, both models can accurately segment, despite the abnormality; for pacemakers, the weak predictor incorrectly segments around the pacemakers, while the strong predictor more accurately segments the lungs; and for consolidation, both models are less able to segment the lungs accurately, however, the strong predictor gets closer to the ground-truth. See Appendix H for more examples." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.739, + 0.637, + 0.755 + ], + "angle": 0, + "content": "5.5 Quantifying the limitations of existing editing methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "LANCE As seen in the second row of Tab. 3, adding edema leads only to a small drop in performance of the strong predictor. We hypothesise that further drops in performances stem from a mismatch of the original mask and the edited images. We therefore use this setup to quantify how well LANCE and RadEdit preserve the shape and position of the lung borders. Additionally, we study the difference between results using" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.28, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "RadEdit: stress-testing biomedical vision models via diffusion image editing" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.328, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.265, + 0.238, + 0.281, + 0.248 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.146, + 0.445, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.378, + 0.238, + 0.395, + 0.248 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.146, + 0.558, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.492, + 0.238, + 0.507, + 0.248 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.146, + 0.669, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.238, + 0.621, + 0.248 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.146, + 0.787, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.718, + 0.238, + 0.733, + 0.248 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.256, + 0.544, + 0.354 + ], + "angle": 0, + "content": "Fig. 5: Comparison of LANCE\\(^2\\) and RadEdit. We measure how well the strong predictor from Tab. 3's outputs matches the ground-truth lung masks (blue) for four synthetic datasets created by adding edema using LANCE and RadEdit with DDIM or DDPM inversion. High Dice / low AHD indicates that the editing method well preserves the lung border's location and shape." + }, + { + "type": "table", + "bbox": [ + 0.551, + 0.252, + 0.777, + 0.352 + ], + "angle": 0, + "content": "
Editing MethodDice↑AHD↓
(a) Original data95.511.6
(b) LANCE w/ DDIM78.965.1
(c) LANCE w/ DDPM80.169.5
(d) RadEdit w/ DDIM86.239.8
(e) RadEdit w/ DDPM93.922.8
" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.374, + 0.788, + 0.451 + ], + "angle": 0, + "content": "DDIM or DDPM inversion. For all four methods in Fig. 5, we use the same setup as in Sec. 5.4: we first edit the original image with the prompt 'Moderate pulmonary edema. The heart size is normal', and then compare the outputs of the strong predictor with the original ground-truth lung masks. Here, we find that using masks and DDPM inversion is necessary for RadEdit to preserve the shape and position of the lung border." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.464, + 0.789, + 0.646 + ], + "angle": 0, + "content": "DiffEdit We quantify how well DiffEdit's automatically predicted masks match the manual ground-truth using the same setup as in Sec. 5.3: we take an image containing pneumothorax and a chest drain, and try to remove only the pneumothorax. The editing prompt is created by keeping only the parts of the impressions related to pneumothorax and chest drains, and replacing the description of the pneumothorax with 'No pneumothorax'. DiffEdit should therefore predict a mask containing only the pneumothorax. We perform a grid search on the MIMIC-Seg [9] validation set over DiffEdit's hyperparameters (noise strength and binarising threshold) to optimise pneumothorax segmentation metrics, then evaluate on the training set. In Fig. 6 we see that DiffEdit's predicted masks obtain poor quantitative metrics where parts of the pneumothorax are often missing, and the spuriously correlated chest drain is often included in the predicted mask. As a result, DiffEdit's predicted masks are unsuitable for stress-testing." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.673, + 0.329, + 0.763 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.765, + 0.571, + 0.787 + ], + "angle": 0, + "content": "(a) Examples of pneumothorax masks predicted using DiffEdit [10]. Blue: ground-truth annotation; red: predicted editing mask." + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.674, + 0.453, + 0.763 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.673, + 0.571, + 0.763 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.587, + 0.677, + 0.778, + 0.69 + ], + "angle": 0, + "content": "Hyperparameters Dice AHD" + }, + { + "type": "text", + "bbox": [ + 0.587, + 0.697, + 0.778, + 0.709 + ], + "angle": 0, + "content": "Tuned per image 33.8 97.7" + }, + { + "type": "text", + "bbox": [ + 0.588, + 0.711, + 0.777, + 0.723 + ], + "angle": 0, + "content": "Tuned on validation 18.4 256.8" + }, + { + "type": "text", + "bbox": [ + 0.578, + 0.731, + 0.787, + 0.782 + ], + "angle": 0, + "content": "(b) Segmentation metrics for the pneumothorax mask predicted by DiffEdit [10], for hyperparameters tuned on the validation set (bottom) and tuned per image (top; which requires ground-truth masks)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.794, + 0.788, + 0.837 + ], + "angle": 0, + "content": "Fig. 6: Evaluating pneumothorax masks predicted using DiffEdit [10]. (a) Predicted masks (red) are noisy, with chest drains often incorrectly segmented as well as or instead of the pneumothorax (blue); (b) this is demonstrated quantitatively with low Dice score and high AHD." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.495, + 0.129 + ], + "angle": 0, + "content": "F. Pérez-García, S. Bond-Taylor et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.487, + 0.162 + ], + "angle": 0, + "content": "6 Limitations and future work" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.173, + 0.786, + 0.324 + ], + "angle": 0, + "content": "Despite the encouraging results presented in the paper, RadEdit is not without limitations and more work is needed to extend it to more applications. Currently, training datasets and models must be manually analysed to predict potential failure cases, simulate these failures to test the hypothesis, and finally quantitatively evaluate the model; future work could automate such failure mode discovery. Another limitation is that current editing techniques do not enable all types of stress-testing; for example, with current approaches, we are unable to test segmentation models' behaviour to cardiomegaly (enlarged heart) since this would require segmentation maps to be adjusted after editing. However, this could potentially be enabled by enlarging heart segmentations to simulate cardiomegaly and adjusting the ground-truth lung segmentation accordingly." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.325, + 0.786, + 0.445 + ], + "angle": 0, + "content": "When using generative editing, it is not possible to guarantee that unwanted changes will not occur. With RadEdit, we minimise this by forcing spuriously correlated regions to remain the same, only using classifier-free guidance within the editing mask, and filtering via image-text alignment. Future work to better maintain structure when editing will help with this issue, but masks will still be necessary to bypass spurious correlations. Furthermore, due to potential overlap of 3D structures, editing 2D X-rays can be limited. In such cases, editing single structures using RadEdit may not be successful. To address this, 3D CT images could be edited, then projected to synthetic X-rays [18]." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.445, + 0.786, + 0.52 + ], + "angle": 0, + "content": "When producing simulated stress test sets, several factors affect edit quality including classifier-free guidance weight, number of inference steps, and time step to encode to. Additionally, the text encoder must well understand specified pathologies to provide informative features to condition the generative model on; similarly, the diffusion model must be able to capture fine details and well cover the data distribution." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.52, + 0.786, + 0.611 + ], + "angle": 0, + "content": "Finally, more research is required to develop better approaches for quantifying edit quality for downstream tasks. In particular, observing a change in downstream performance is not necessarily indicative of real-world performance as edit quality may be poor. While the introduced BioViL-T editing score can be used to quantify edit quality, this introduces reliance on a potentially biased model. Additionally, the BioViL-T editing score is not suited to detect the artefacts introduced by LANCE and DiffEdit." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.63, + 0.343, + 0.647 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.659, + 0.786, + 0.84 + ], + "angle": 0, + "content": "In this study, we illustrate the efficacy of generative image editing as a robust tool for stress-testing biomedical vision models. Our focus is on assessing their robustness against three types of dataset shifts commonly encountered in biomedical imaging: acquisition shift, manifestation shift, and population shift. We highlight that one of the significant challenges in biomedical image editing is the correlations learned by the generative model, which can result in artefacts during the editing process. To mitigate these artefacts, RadEdit relies on various types of masks to restrict the effects of the editing to certain areas while ensuring the consistency of the edited images. This approach enables us to generate high fidelity synthetic test sets that exhibit common dataset shifts. These synthetic test sets are used to identify and quantify the failure modes of biomedical classification and segmentation models. This provides a valuable supplement to explainable AI approaches such as Grad-CAM [63] and saliency maps [66, 1]." + } + ], + [ + { + "type": "title", + "bbox": [ + 0.434, + 0.143, + 0.57, + 0.164 + ], + "angle": 0, + "content": "Bibliography" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.212, + 0.787, + 0.258 + ], + "angle": 0, + "content": "[1] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. Advances in neural information processing systems, 31, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.257, + 0.787, + 0.302 + ], + "angle": 0, + "content": "[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.302, + 0.787, + 0.332 + ], + "angle": 0, + "content": "[3] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM Transactions on Graphics (TOG), 42(4):1-11, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.332, + 0.787, + 0.362 + ], + "angle": 0, + "content": "[4] Samaneh Azadi, Catherine Olsson, Trevor Darrell, Ian Goodfellow, and Augustus Odena. Discriminator rejection sampling. arXiv preprint arXiv:1810.06758, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.362, + 0.787, + 0.467 + ], + "angle": 0, + "content": "[5] Shruthi Bannur, Stephanie Hyland, Qianchu Liu, Fernando Perez-Garcia, Maximilian Ilse, Daniel C. Castro, Benedikt Boecking, Harshita Sharma, Kenya Bouzid, Anja Thieme, Anton Schwaighofer, Maria Wetscherek, Matthew P. Lungren, Aditya Nori, Javier Alvarez-Valle, and Ozan Oktay. Learning to Exploit Temporal Structure for Biomedical Vision-Language Processing. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 1 2023. https://doi.org/10.48550/arxiv.2301.04558." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.467, + 0.787, + 0.526 + ], + "angle": 0, + "content": "[6] Riccardo Barbano, Alexander Denker, Hyungjin Chung, Tae Hoon Roh, Simon Arridge, Peter Maass, Bangti Jin, and Jong Chul Ye. Steerable conditional diffusion for out-of-distribution adaptation in imaging inverse problems. arXiv preprint arXiv:2308.14409, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.526, + 0.787, + 0.556 + ], + "angle": 0, + "content": "[7] Daniel C Castro, Ian Walker, and Ben Glocker. Causality matters in medical imaging. Nature Communications, 11(1):3673, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.556, + 0.787, + 0.63 + ], + "angle": 0, + "content": "[8] Pierre Chambon, Christian Bluethgen, Jean-Benoit Delbrouck, Rogier Van der Sluijs, Malgorzata Polacin, Juan Manuel Zambrano Chaves, Tanishq Mathew Abraham, Shivanshu Purohit, Curtis P Langlotz, and Akshay Chaudhari. Roentgen: vision-language foundation model for chest x-ray generation. arXiv preprint arXiv:2211.12737, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.63, + 0.787, + 0.675 + ], + "angle": 0, + "content": "[9] Li-Ching Chen, Po-Chih Kuo, Ryan Wang, Judy Gichoya, and Leo Anthony Celi. Chest X-ray segmentation images based on MIMIC-CXR (version 1.0.0). PhysioNet, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.676, + 0.787, + 0.706 + ], + "angle": 0, + "content": "[10] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. DiffEdit: Diffusion-based semantic image editing with mask guidance, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.706, + 0.787, + 0.764 + ], + "angle": 0, + "content": "[11] Alex J. DeGrave, Joseph D. Janizek, and Su-In Lee. AI for radiographic COVID-19 detection selects shortcuts over signal. Nature Machine Intelligence, 3(7):610-619, 2021. ISSN 2522-5839. https://doi.org/10.1038/s42256-021-00338-7." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.764, + 0.787, + 0.811 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Jost Tobias Springenberg, and Thomas Brox. Learning to generate chairs with convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1538-1546, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "[13] Sijing Feng, Damian Azzollini, Ji Soo Kim, Cheng-Kai Jin, Simon P. Gordon, Jason Yeoh, Eve Kim, Mina Han, Andrew Lee, Aakash Patel, Joy Wu, Martin" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.212, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.495, + 0.129 + ], + "angle": 0, + "content": "F. Pérez-García, S. Bond-Taylor et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.249, + 0.146, + 0.786, + 0.207 + ], + "angle": 0, + "content": "Urschler, Amy Fong, Cameron Simmers, Gregory P. Tarr, Stuart Barnard, and Ben Wilson. Curation of the CANDID-PTX dataset with free-text reports. *Radiology: Artificial Intelligence*, 3(6):e210136, 2021. ISSN 2638-6100. https://doi.org/10.1148/ryai.2021210136." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.209, + 0.787, + 0.268 + ], + "angle": 0, + "content": "[14] Virginia Fernandez, Pedro Sanchez, Walter Hugo Lopez Pinaya, Grzegorz Jacenków, Sotirios A Tsaftaris, and Jorge Cardoso. Privacy distillation: Reducing re-identification risk of multimodal diffusion models. arXiv preprint arXiv:2306.01322, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.787, + 0.315 + ], + "angle": 0, + "content": "[15] Alessandro Fontanella, Grant Mair, Joanna Wardlaw, Emanuele Trucco, and Amos Storkey. Diffusion models for counterfactual generation and anomaly detection in brain images. URL http://arxiv.org/abs/2308.02062." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.317, + 0.787, + 0.375 + ], + "angle": 0, + "content": "[16] Nicolas Gaggion, Candelaria Mosquera, Lucas Mansilla, Martina Aineseder, Diego H Milone, and Enzo Ferrante. CheXmask: a large-scale dataset of anatomical segmentation masks for multi-center chest X-ray images. arXiv preprint arXiv:2307.03293, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.378, + 0.787, + 0.437 + ], + "angle": 0, + "content": "[17] Rinon Gal, Or Patashnik, Haggai Maron, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. StyleGAN-NADA: CLIP-guided domain adaptation of image generators. ACM Transactions on Graphics, 41(4), 2022. https://doi.org/10.1145/3528223.3530164." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.44, + 0.787, + 0.5 + ], + "angle": 0, + "content": "[18] Cong Gao, Benjamin D Killeen, Yicheng Hu, Robert B Grupp, Russell H Taylor, Mehran Armand, and Mathias Unberath. Synthetic data accelerates the development of generalizable learning-based algorithms for x-ray image analysis. Nature Machine Intelligence, 5(3):294-308, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.787, + 0.546 + ], + "angle": 0, + "content": "[19] Timur Garipov, Sebastiaan De Peuter, Ge Yang, Vikas Garg, Samuel Kaski, and Tommi Jaakkola. Compositional sculpting of iterative generative processes. arXiv preprint arXiv:2309.16115, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.548, + 0.787, + 0.577 + ], + "angle": 0, + "content": "[20] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.579, + 0.787, + 0.639 + ], + "angle": 0, + "content": "[21] Yu Gu, Jianwei Yang, Naoto Usuyama, Chunyuan Li, Sheng Zhang, Matthew P Lungren, Jianfeng Gao, and Hoifung Poon. Biomedjourney: Counterfactual biomedical image generation by instruction-learning from multimodal patient journeys. arXiv preprint arXiv:2310.10765, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.641, + 0.787, + 0.67 + ], + "angle": 0, + "content": "[22] Will Douglas Heaven. Hundreds of ai tools have been built to catch covid. none of them helped. MIT Technology Review. Retrieved December 2023, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.672, + 0.787, + 0.716 + ], + "angle": 0, + "content": "[23] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.787, + 0.763 + ], + "angle": 0, + "content": "[24] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.765, + 0.787, + 0.824 + ], + "angle": 0, + "content": "[25] Irina Higgins, Loic Matthew, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. \\(\\beta\\)-VAE: Learning basic visual concepts with a constrained variational framework. In International Conference on Learning Representations, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.826, + 0.733, + 0.84 + ], + "angle": 0, + "content": "[26] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022." + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.146, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.28, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "RadEdit: stress-testing biomedical vision models via diffusion image editing" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.146, + 0.785, + 0.176 + ], + "angle": 0, + "content": "[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.178, + 0.787, + 0.223 + ], + "angle": 0, + "content": "[28] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research, 23(1):2249-2281, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.223, + 0.786, + 0.252 + ], + "angle": 0, + "content": "[29] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly DDPM noise space: Inversion and manipulations, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.253, + 0.787, + 0.342 + ], + "angle": 0, + "content": "[30] Jeremy Irvin, Pranav Rajpurkar, Michael Ko, Yifan Yu, Silviana Ciurea-Ilicus, Chris Chute, Henrik Marklund, Behzad Haghgoo, Robyn Ball, Katie Shpanskaya, Jayne Seekins, David A. Mong, Safwan S. Halabi, Jesse K. Sandberg, Ricky Jones, David B. Larson, Curtis P. Langlotz, Bhavik N. Patel, Matthew P. Lungren, and Andrew Y. Ng. CheXpert: A large chest radiograph dataset with uncertainty labels and expert comparison, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.787, + 0.403 + ], + "angle": 0, + "content": "[31] Stefan Jaeger, Sema Candemir, Sameer Antani, Yi-Xiang J Wang, Pu-Xuan Lu, and George Thoma. Two public chest x-ray datasets for computer-aided screening of pulmonary diseases. Quantitative imaging in medicine and surgery, 4(6):475, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.403, + 0.787, + 0.433 + ], + "angle": 0, + "content": "[32] Priyank Jaini, Kevin Clark, and Robert Geirhos. Intriguing properties of generative classifiers. arXiv preprint arXiv:2309.16779, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.787, + 0.508 + ], + "angle": 0, + "content": "[33] Alistair E. W. Johnson, Tom J. Pollard, Seth J. Berkowitz, Nathaniel R. Greenbaum, Matthew P. Lungren, Chih-ying Deng, Roger G. Mark, and Steven Horng. MIMIC-CXR, a de-identified publicly available database of chest radiographs with free-text reports. Scientific Data, 6(1):317, 2019. ISSN 2052-4463. https://doi.org/10.1038/s41597-019-0322-0." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.787, + 0.554 + ], + "angle": 0, + "content": "[34] Charles Jones, Daniel C Castro, Fabio De Sousa Ribeiro, Ozan Oktay, Melissa McCradden, and Ben Glocker. No fair lunch: A causal perspective on dataset bias in machine learning for medical imaging. arXiv preprint arXiv:2307.16526, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.787, + 0.613 + ], + "angle": 0, + "content": "[35] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10124–10134, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.614, + 0.787, + 0.644 + ], + "angle": 0, + "content": "[36] Priyatham Kattakinda, Alexander Levine, and Soheil Feizi. Invariant learning via diffusion dreamed distribution shifts. arXiv preprint arXiv:2211.10370, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.645, + 0.787, + 0.659 + ], + "angle": 0, + "content": "[37] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.787, + 0.719 + ], + "angle": 0, + "content": "[38] Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas Phillips, Irena Gao, et al. Wilds: A benchmark of in-the-wild distribution shifts. In International Conference on Machine Learning, pages 5637-5664. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.72, + 0.787, + 0.78 + ], + "angle": 0, + "content": "[39] Ira Ktena, Olivia Wiles, Isabela Albuquerque, Sylvestre-Alvise Rebuffi, Ryutaro Tanno, Abhijit Guha Roy, Shekoofeh Azizi, Danielle Belgrave, Pushmeet Kohli, Alan Karthikesalingam, et al. Generative models improve fairness of medical classifiers under distribution shifts. arXiv preprint arXiv:2304.09218, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.781, + 0.787, + 0.84 + ], + "angle": 0, + "content": "[40] Agostina J Larrazabal, Nicolás Nieto, Victoria Peterson, Diego H Milone, and Enzo Ferrante. Gender imbalance in medical imaging datasets produces biased classifiers for computer-aided diagnosis. Proceedings of the National Academy of Sciences, 117(23):12592–12594, 2020." + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.146, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.495, + 0.129 + ], + "angle": 0, + "content": "F. Pérez-García, S. Bond-Taylor et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.177 + ], + "angle": 0, + "content": "[41] Choong Ho Lee and Hyung-Jin Yoon. Medical big data: promise and challenges. *Kidney research and clinical practice*, 36(1):3, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.179, + 0.785, + 0.207 + ], + "angle": 0, + "content": "[42] Xiaodan Li, Yuefeng Chen, Yao Zhu, Shuhui Wang, Rong Zhang, and Hui Xue. ImageNet-e: Benchmarking neural network robustness via attribute editing, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.208, + 0.785, + 0.252 + ], + "angle": 0, + "content": "[43] Wufeng Liu, Jiaxin Luo, Yan Yang, Wenlian Wang, Junkui Deng, and Liang Yu. Automatic lung segmentation in chest x-ray images using improved u-net. Scientific Reports, 12(1):8649, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.253, + 0.785, + 0.312 + ], + "angle": 0, + "content": "[44] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. Advances in Neural Information Processing Systems, 33:11525-11538, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.313, + 0.786, + 0.584 + ], + "angle": 0, + "content": "[45] Lena Maier-Hein, Annika Reinke, Patrick Godau, Minu D. Tizabi, Florian Buettner, Evangelia Christodoulou, Ben Glocker, Fabian Isensee, Jens Kleesiek, Michal Kozubek, Mauricio Reyes, Michael A. Riegler, Manuel Wiesenfarth, A. Emre Kavur, Carole H. Sudre, Michael Baumgartner, Matthias Eisenmann, Doreen Heckmann-Nötzel, A. Tim Rädsch, Laura Acion, Michela Antonelli, Tal Arbel, Spyridon Bakas, Arrielenis, Matthew Blaschko, M. Jorge Cardoso, Veronika Cheplygina, Beth A. Cimini, Gary S. Collins, Keyvan Farahani, Luciana Ferrer, Adrian Galdran, Bram van Ginneken, Robert Haase, Daniel A. Hashimoto, Michael M. Hoffman, Merel Huisman, Pierre Jannin, Charles E. Kahn, Dagmar Kainmueller, Bernhard Kainz, Alexandros Karargyris, Alan Karthikesalingam, Hannes Kenngott, Florian Kofler, Annette Kopp-Schneider, Anna Kreshuk, Tahsin Kurc, Bennett A. Landman, Geert Litjens, Amin Madani, Klaus Maier-Hein, Anne L. Martel, Peter Mattson, Erik Meijering, Bjoern Menze, Karel G. M. Moons, Henning Müller, Brennan Nichyporuk, Felix Nickel, Jens Petersen, Nasir Rajpoot, Nicola Rieke, Julio Saez-Rodriguez, Clara I. Sánchez, Shravya Shetty, Maarten van Smeden, Ronald M. Summers, Abdel A. Taha, Aleksei Tiulpin, Sotirios A. Tsaftaris, Ben Van Calster, Gaël Varoquaux, and Paul F. Jäger. Metrics reloaded: Recommendations for image analysis validation, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.585, + 0.786, + 0.629 + ], + "angle": 0, + "content": "[46] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.63, + 0.786, + 0.69 + ], + "angle": 0, + "content": "[47] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Nulltext inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.69, + 0.786, + 0.75 + ], + "angle": 0, + "content": "[48] Gustav Müller-Franzes, Jan Moritz Niehues, Firas Khader, Soroosh Tayebi Arasteh, Christoph Haarburger, Christiane Kuhl, Tianci Wang, Tianyu Han, Sven Nebelung, Jakob Nikolas Kather, et al. Diffusion probabilistic models beat gans on medical images. arXiv preprint arXiv:2212.07501, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.751, + 0.786, + 0.795 + ], + "angle": 0, + "content": "[49] Nick Pawlowski, Daniel C. Castro, and Ben Glocker. Deep structural causal models for tractable counterfactual inference. In Advances in Neural Information Processing Systems, volume 33, pages 857-869, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.796, + 0.786, + 0.84 + ], + "angle": 0, + "content": "[50] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis, 2023." + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.28, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "RadEdit: stress-testing biomedical vision models via diffusion image editing" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.146, + 0.788, + 0.193 + ], + "angle": 0, + "content": "[51] Viraj Prabhu, Sriram Yenamandra, Prithvijit Chattopadhyay, and Judy Hoffman. LANCE: Stress-testing visual models by generating language-guided counterfactual images, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.194, + 0.788, + 0.239 + ], + "angle": 0, + "content": "[52] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.24, + 0.788, + 0.302 + ], + "angle": 0, + "content": "[53] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.302, + 0.788, + 0.346 + ], + "angle": 0, + "content": "[54] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.348, + 0.788, + 0.409 + ], + "angle": 0, + "content": "[55] Jacob C. Reinhold, Aaron Carass, and Jerry L. Prince. A structural causal model for MR images of multiple sclerosis. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2021, volume 12905 of LNCS, pages 782–792, 2021. https://doi.org/10.1007/978-3-030-87240-3_75." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.41, + 0.788, + 0.486 + ], + "angle": 0, + "content": "[56] Michael Roberts, Derek Driggs, Matthew Thorpe, Julian Gilbey, Michael Yeung, Stephan Ursprung, Angelica I Aviles-Rivero, Christian Etmann, Cathal McCague, Lucian Beer, et al. Common pitfalls and recommendations for using machine learning to detect and prognosticate for Covid-19 using chest radiographs and ct scans. Nature Machine Intelligence, 3(3):199-217, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.487, + 0.788, + 0.518 + ], + "angle": 0, + "content": "[57] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.519, + 0.788, + 0.609 + ], + "angle": 0, + "content": "[58] Johannes Rueckel, Lena Trappmann, Balthasar Schachtner, Philipp Wesp, Boj Friedrich Hoppe, Nicola Fink, Jens Ricke, Julien Dinkel, Michael Ingrisch, and Bastian Oliver Sabel. Impact of confounding thoracic tubes and pleural dehiscence extent on artificial intelligence pneumothorax detection in chest radiographs. Investigative Radiology, 55(12):792-798, July 2020. ISSN 0020-9996. https://doi.org/10.1097/rli.0000000000000707." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.788, + 0.715 + ], + "angle": 0, + "content": "[59] Johannes Rueckel, Christian Huemmer, Andreas Fieselmann, Florin-Cristian Gheu, Awais Mansoor, Balthasar Schachtner, Philipp Wesp, Lena Trappmann, Basel Munawwar, Jens Ricke, Michael Ingrisch, and Bastian O. Sabel. Pneumothorax detection in chest radiographs: optimizing artificial intelligence system for accuracy and confounding bias reduction using in-image annotations in algorithm training. European Radiology, 31(10):7888-7900, 2021. https://doi.org/10.1007/s00330-021-07833-w." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.788, + 0.762 + ], + "angle": 0, + "content": "[60] Chitwan Sahara, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.763, + 0.788, + 0.808 + ], + "angle": 0, + "content": "[61] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Semantic foggy scene understanding with synthetic data. International Journal of Computer Vision, 126: 973-992, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "[62] Pedro Sanchez, Antanas Kascenas, Xiao Liu, Alison Q O'Neil, and Sotirios A Tsaftaris. What is healthy? generative counterfactual diffusion for lesion localization" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.146, + 0.788, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.495, + 0.129 + ], + "angle": 0, + "content": "F. Pérez-García, S. Bond-Taylor et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.249, + 0.146, + 0.786, + 0.175 + ], + "angle": 0, + "content": "tion. In MICCAI Workshop on Deep Generative Models, pages 34-44. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.178, + 0.787, + 0.237 + ], + "angle": 0, + "content": "[63] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.239, + 0.787, + 0.283 + ], + "angle": 0, + "content": "[64] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9243–9252, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.284, + 0.787, + 0.373 + ], + "angle": 0, + "content": "[65] Junji Shiraishi, Shigehiko Katsuragawa, Junpei Ikezoe, Tsuneo Matsumoto, Takeshi Kobayashi, Ken-ichi Komatsu, Mitate Matsui, Hiroshi Fujita, Yoshie Kodera, and Kunio Doi. Development of a digital image database for chest radiographs with and without a lung nodule: receiver operating characteristic analysis of radiologists' detection of pulmonary nodules. American Journal of Roentgenology, 174(1):71-74, 2000." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.374, + 0.787, + 0.418 + ], + "angle": 0, + "content": "[66] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.42, + 0.787, + 0.464 + ], + "angle": 0, + "content": "[67] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.465, + 0.787, + 0.493 + ], + "angle": 0, + "content": "[68] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.495, + 0.787, + 0.538 + ], + "angle": 0, + "content": "[69] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.539, + 0.787, + 0.598 + ], + "angle": 0, + "content": "[70] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.6, + 0.787, + 0.659 + ], + "angle": 0, + "content": "[71] Paul Upchurch, Jacob Gardner, Geoff Pleiss, Robert Pless, Noah Snavely, Kavita Bala, and Kilian Weinberger. Deep feature interpolation for image content changes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7064-7073, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.661, + 0.787, + 0.703 + ], + "angle": 0, + "content": "[72] Boris Van Breugel, Nabeel Seedat, Fergus Imrie, and Mihaela van der Schaar. Can you rely on your model evaluation? improving model evaluation with synthetic test data, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.705, + 0.787, + 0.78 + ], + "angle": 0, + "content": "[73] Maria de la Iglesia Vayá, Jose Manuel Saborit, Joaquim Angel Montell, Antonio Pertusa, Aurelia Bustos, Miguel Cazorla, Joaquin Galant, Xavier Barber, Domingo Orozco-Beltrán, Francisco García-García, Marisa Caparrós, Germán González, and Jose María Salinas. BIMCV COVID-19+: a large annotated dataset of RX and CT images from COVID-19 patients, 2020. version: 3." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.781, + 0.787, + 0.81 + ], + "angle": 0, + "content": "[74] I von Borzyskowski, A Mazumder, B Mateen, and M Wooldridge. Data science and ai in the age of covid-19, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.84 + ], + "angle": 0, + "content": "[75] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M. Summers. ChestX-ray8: Hospital-scale chest x-ray database and" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.146, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.28, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "RadEdit: stress-testing biomedical vision models via diffusion image editing" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "21" + }, + { + "type": "ref_text", + "bbox": [ + 0.252, + 0.147, + 0.786, + 0.207 + ], + "angle": 0, + "content": "benchmarks on weakly-supervised classification and localization of common thorax diseases. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3462-3471, 2017. https://doi.org/10.1109/CVPR.2017.369." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.208, + 0.787, + 0.268 + ], + "angle": 0, + "content": "[76] Laure Wynants, Ben Van Calster, Gary S Collins, Richard D Riley, Georg Heinze, Ewoud Schuit, Elena Albu, Banafsheh Arshi, Vanessa Bellou, Marc MJ Bonten, et al. Prediction models for diagnosis and prognosis of Covid-19: systematic review and critical appraisal. bmj, 369, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.268, + 0.787, + 0.314 + ], + "angle": 0, + "content": "[77] Anna Zawacki, Carol Wu, George Shih, Julia Elliott, Mikhail Fomitchev, Mohannad ParasLakhani Hussain, Phil Culliton, and Shunxing Bao. Siim-acr pneumothorax segmentation, 2019." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.208, + 0.787, + 0.314 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/RadEdit_ stress-testing biomedical vision models via diffusion image editing/730452c5-1675-40d8-8c46-f903c2fa2703_origin.pdf b/2024/RadEdit_ stress-testing biomedical vision models via diffusion image editing/730452c5-1675-40d8-8c46-f903c2fa2703_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bcc61b9297fdc6b398f580a02cedf8d75c79f428 --- /dev/null +++ b/2024/RadEdit_ stress-testing biomedical vision models via diffusion image editing/730452c5-1675-40d8-8c46-f903c2fa2703_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a4e8b61d270ea734811b98382c9de22bb2d06ae920f58e40ee74d5b2cabaa7c7 +size 4667760 diff --git a/2024/RadEdit_ stress-testing biomedical vision models via diffusion image editing/full.md b/2024/RadEdit_ stress-testing biomedical vision models via diffusion image editing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2dff9f224db23b53ae3e59e5e0af21b685b66ac2 --- /dev/null +++ b/2024/RadEdit_ stress-testing biomedical vision models via diffusion image editing/full.md @@ -0,0 +1,381 @@ +# RadEdit: stress-testing biomedical vision models via diffusion image editing + +Fernando Pérez-García\*,1, Sam Bond-Taylor\*,1, Pedro P. Sanchez+,2, Boris van Breugel+,3, Daniel C. Castro1, Harshita Sharma1, Valentina Salvatelli1, Maria T.A. Wetscherek1, Hannah Richardson1, Matthew P. Lungren1,4,5, Aditya Nori1, Javier Alvarez-Valle1, Ozan Oktay†,1, and Maximilian Ilse†,1 + +1 Microsoft Health Futures +2 University of Edinburgh +3 University of Cambridge +4 University of California +5 Stanford University +* Shared first author + ++ Work done at Microsoft Health Futures + +† Shared last author + +Abstract. Biomedical imaging datasets are often small and biased, meaning that real-world performance of predictive models can be substantially lower than expected from internal testing. This work proposes using generative image editing to simulate dataset shifts and diagnose failure modes of biomedical vision models; this can be used in advance of deployment to assess readiness, potentially reducing cost and patient harm. Existing editing methods can produce undesirable changes, with spurious correlations learned due to the co-occurrence of disease and treatment interventions, limiting practical applicability. To address this, we train a text-to-image diffusion model on multiple chest X-ray datasets and introduce a new editing method, RadEdit, that uses multiple image masks, if present, to constrain changes and ensure consistency in the edited images, minimising bias. We consider three types of dataset shifts: acquisition shift, manifestation shift, and population shift, and demonstrate that our approach can diagnose failures and quantify model robustness without additional data collection, complementing more qualitative tools for explainable AI. + +Keywords: Image editing $\cdot$ diffusion models $\cdot$ biomedical imaging + +# 1 Introduction + +Developing accurate and robust models for biomedical image analysis requires large and diverse datasets that are often difficult to obtain due to ethical, legal, geographical, and financial constraints [41]. This leads to biased training datasets that affect the performance of trained models and generalisation to real-world scenarios [59, 40]. Such data mismatch may arise from genuine differences in upstream data acquisition as well as from the selection criteria for dataset creation, which materialise as various forms of dataset shifts (population, acquisition, annotation, prevalence, manifestation) [7]. + +Top: editing out COVID-19 features results in false positives since the classifier relies on acquisition differences, e.g., radiographic markers (white arrow). + +Middle: editing out a pneumothorax (PTX) results in false positives since the classifier instead detects chest drains. + +Bottom: editing abnormalities into lungs causes a lung segmentation model to mislabel (blue: ground-truth segmentation; red: model prediction). + +![](images/d2a8b5c9a55a2a7c69814c666b6e44508a804669c7e9df2845998298be920b81.jpg) +Fig.1: Stress-testing models by simulating dataset shifts via image editing. + +Biomedical vision models, when put into real-world use, can be unhelpful or potentially even harmful to patients if they are affected by dataset shifts, leading to missed diagnoses [22, 74, 76, 56]. For example, the COVID-19 pandemic led to hundreds of detection tools being developed, with some put into use in hospitals; yet Roberts et al. [56] found that "none of the models identified are of potential clinical use due to methodological flaws and/or underlying biases." It is therefore crucial to properly assess models for bias, prior to real-world use. + +Recent deep generative models have made remarkable improvements in terms of sample quality, diversity, and steerability [57, 48, 35, 28]. These models have been shown to generalise to out-of-distribution domains [42, 6, 32, 19], opening up avenues for new applications. One such application is generating synthetic data for stress-testing models [51, 42, 72]. This involves creating data that is realistic, yet can represent settings, domains, or populations that do not appear (enough) in the real training/test data. + +In this work, we investigate how deep generative models can be used for stress-testing biomedical imaging models. We consider three dataset shift scenarios: + +1. Acquisition shift: classifying COVID-19 cases when the positive and negative cases were acquired at different hospitals (Sec. 5.2). +2. Manifestation shift: detecting if pneumothorax1 was resolved when chest drains (inserted to treat pneumothorax) are present (Sec. 5.3). +3. Population shift: segmenting lungs in the presence of abnormalities rarely or never seen in the training dataset (Sec. 5.4). + +For each of these scenarios, we simulate dataset shifts, producing stress-test sets which can occur in the real world but do not appear or are underrepresented in the original training/test sets. Following prior work, these test sets are synthesised using generative image editing, which unlike generating images from scratch, only minimally modifies the images, hence, better retains fidelity and diversity [51, 42]. For the above scenarios, we use generative editing to 1. remove only COVID-19 features while keeping visual indicators of the different hospitals; 2. remove only pneumothorax while keeping the chest drain; and 3. add abnormalities that occlude lung structures in the image. + +We train a diffusion model [27, 57] on a large collection of chest X-rays from a variety of biomedical imaging datasets (Sec. 5.1), enabling us to add and remove a wide variety of pathologies and support devices when editing. Despite the diversity within these datasets, substantial biases are still present, some of which are learned by the generative model; as a result, correlated features may also be modified. For example, in Scenario 2, removing the pneumothorax might also remove the chest drains as both features typically co-occur in datasets [58], since chest drains are used to treat pneumothorax. Furthermore, when editing only within editing masks, artefacts often appear at the border of the masks. Lastly, artefacts occur when editing images outside of the training dataset domain of the diffusion model used for editing. To overcome these challenges, we propose using multiple masks to break existing correlations. This involves defining which regions must change, and explicitly forcing correlated regions to remain unchanged. In addition, we allow the area outside of the masks to be modified by the diffusion model to ensure image consistency. Since our proposed editing method, which we call RadEdit, leads to only minimal overall changes of chest X-rays, we are able to generate synthetic datasets that can be used to stress-test segmentation models (Scenario 3), which, to the best of our knowledge, we are the first to demonstrate. + +In summary, our contributions are as follows: + +- We introduce a novel editing approach that reduces the presence of artefacts in edited images and simplifies prompt construction compared to prior work [10, 51]. +- Our editing approach allows us to construct synthetic datasets with specific data shifts by performing zero-shot edits on datasets/abnormalities not seen in training. +- We conduct a broad set of experiments using these synthetic datasets to stress-test and expose biases in biomedical classification and, for the first time, segmentation models, introducing a new use case of synthetic data into the medical setting. + +# 2 Preliminaries + +In this section, we introduce background context for stress-testing biomedical imaging models: failure modes of biomedical imaging models caused by different dataset shifts; diffusion models as versatile generative models; and diffusion-based image editing. + +# 2.1 Dataset shifts + +Dataset shift refers to a discrepancy between the training and test data distributions due to external factors [7, 34]. Such shifts are regularly observed in machine learning for biomedical imaging, often due to data scarcity. For example, collected training datasets might consist primarily of healthy patients. However, when the model is used in practice after training, there could be a shift towards unhealthy patients. A taxonomy of different types of dataset shifts in the context of biomedical imaging was developed by Castro et al. [7]. In this paper, we consider three dataset shifts of particular interest. + +Acquisition shift results from the use of different scanners (manufacturer, hardware, and software) or imaging protocols as often encountered when using data from multiple cohorts. These changes affect factors such as image resolution, contrast, patient positioning, and image markings. + +Manifestation shift results from the way the prediction targets physically manifest in anatomy changes between domains. For example, training datasets could consist of more severe pathological cases than observed in practice, or a pathology may co-occur with different visual features, e.g., support devices. + +Population shift results from differences in intrinsic characteristics of the populations under study, changing the anatomical appearance distribution. This definition encompasses examples such as age, sex, ethnicity, and comorbidities, but also abnormalities such as pleural effusion and support devices. In contrast to manifestation shift, the shift in anatomical appearance is not affected by prediction targets. + +# 2.2 Diffusion models + +Denoising diffusion probabilistic models (DDPMs) [27, 67] are a versatile and effective class of generative models that enable sampling from the data distribution by learning to denoise samples corrupted with Gaussian noise. DDPMs are formed by defining a forward time process that gradually adds noise to data points $x_0$ through the recursion + +$$ +x _ {t} = \sqrt {1 - \beta_ {t}} x _ {t - 1} + \sqrt {\beta_ {t}} \epsilon_ {t}, \quad t = 1, \dots , T, \quad \text {s . t .} x _ {t} = \sqrt {\bar {\alpha} _ {t}} x _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \bar {\epsilon} _ {t}, \tag {1} +$$ + +where $\epsilon_{1:T},\bar{\epsilon}_{1:T}\sim \mathcal{N}(0,I),\beta_{1:T}$ is a predefined noise schedule that determines how quickly to corrupt the data and ensures that $x_{T}$ contains little to no information about $x_0$ , and $\bar{\alpha}_t = \prod_{s = 1}^t (1 - \beta_s)$ . To form a generative model, the process is reversed in time, gradually transforming Gaussian noise into samples from the learned distribution. While the exact reversal is intractable, a variational approximation is defined by [68]: + +$$ +x _ {t - 1} = \hat {\mu} _ {t} \left(x _ {t}, f _ {\theta} \left(x _ {t}, t, c\right)\right) + \sigma_ {t} z _ {t}, \tag {2} +$$ + +$$ +\hat {\mu} _ {t} \left(x _ {t}, \epsilon_ {t}\right) = \sqrt {\bar {\alpha} _ {t - 1}} \frac {x _ {t} - \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {t}}{\sqrt {\bar {\alpha} _ {t}}} + \sqrt {1 - \bar {\alpha} _ {t - 1} - \sigma_ {t} ^ {2}} \epsilon_ {t}, \tag {3} +$$ + +where $c$ is a conditioning signal such as a text description, $f_{\theta}(x_t,t,c)$ is a learned approximation of the noise $\bar{\epsilon}_t$ that corrupted the image $x_0$ to obtain $x_{t}$ , $z_{1:T} \sim \mathcal{N}(0,I)$ , and $\sigma_{1:T}$ controls how much noise is introduced. The process is Markovian and known as a DDPM [27] when $\sigma_t = \sqrt{(1 - \bar{\alpha}_{t - 1}) / (1 - \bar{\alpha}_t)}\sqrt{1 - \bar{\alpha}_t / \bar{\alpha}_{t - 1}}$ , while for $\sigma_t = 0$ the process is deterministic and is called a denoising diffusion implicit model (DDIM) [68]. + +# 2.3 Image editing + +Since DDIMs have a one-to-one correspondence with latent vectors $x_{T}$ , we can deterministically map data points to latents by running the DDIM generative process in reverse [68], called DDIM inversion. Several approaches [10, 46] have shown that images can be edited by running the reverse diffusion process augmented by the latent vectors and a modified prompt $c$ . However, this method can lead to undesired artefacts in the edited images. For example, structures unrelated to the desired edit may also change shape, size, or location. To address this, Huberman-Spiegelglas et al. [29] propose DDPM inversion. Here, the original forward process defined in Eq. (1) is adapted, replacing the correlated vectors $\bar{\epsilon}_{1:T}$ with statistically independent vectors $\tilde{\epsilon}_{1:T}$ (more details in Appendix B). These noise vectors are then used in the generative process, retaining the structure of the original image better than DDIM inversion. + +# 3 Related work + +# 3.1 Generative image editing + +With advances in deep generative modelling, several approaches to image editing have emerged. Many of these early approaches use compressed latent manipulation [12, 52, 64, 71] where fine-grained edits are difficult to achieve and can result in unwanted changes. More recently, the unparalleled flexibility of diffusion models, together with advances in plain text conditioning, have opened up new avenues for editing techniques. + +Here, we describe some notable diffusion editing methods. SDEdit [46] shows that diffusion models trained solely on real images can be used to generate images from sketches by perturbing sketches with noise, then running the reverse diffusion process. Palette [60] is a diffusion model trained for inpainting by filling regions with noise and denoising. Blended diffusion [2, 3] uses masks with CLIP [53] conditioning to guide local edits. Multiple works show that injecting U-Net activations, obtained by encoding the original image into the generation process, better retains image structure [24, 70]. DiffEdit [10] uses text prompts to determine the appropriate region to edit. Mokady et al. [47] improve diffusion inversion quality by optimising the diffusion trajectory. + +Crucially, in the works which use masks for editing, a single type of mask is always used to define the region of interest. In this work, we argue that a second type of mask is required to avoid the loss of features caused by spurious correlations. As better editing approaches are developed, this requirement should be kept in mind. + +# 3.2 Biomedical imaging counterfactuals + +Generative models have also been applied to biomedical counterfactual generation. Reinhold et al. [55] manipulate causes of multiple sclerosis in brain MRI with deep structural causal models [49]. Sanchez et al. [62] and Fontanella et al. [15] use editing to remove pathologies for abnormality detection. Ktena et al. [39] generate out-of-distribution samples to improve classifier performance. Gu et al. [21] train a diffusion model to model disease progression by conditioning on a prior X-ray and text progression description. Unlike our approach, these methods do not use masks to enforce which regions may or may not be edited, meaning that spurious correlations might affect edits. Additionally, these methods use synthetic data to augment and improve model performance whereas we focus on using synthetic medical data for stress-testing. + +# 3.3 Stress-testing + +Several approaches have used non-deep-generative-model methods to stress-test networks. Hendrycks and Dietterich [23] evaluate classification models' robustness to corruptions such as blurring, Gaussian noise, and JPEG artefacts. Sakaridis et al. [61] stress-test a segmentation model for roads by using an optical model to add synthetic fog to scenes. Koh et al. [38] collate a dataset presenting various distribution shifts. + +More recent models have made use of conditional generative models to simulate shifts. Prabhu et al. [51] propose LANCE, which stress-tests classification models by using diffusion-based image editing to modify image subjects via caption editing with + +a large language model (LLM); Kattakinda et al. [36] do similar, but instead modify the background. Li et al. [42] use diffusion models with a single subject mask to separately edit backgrounds and subjects. Van Breugel et al. [72] use generative adversarial networks to simulate distribution shifts on tabular data. This line of research is partially related to adversarial attacks [20], where the focus is on minimally modifying images such that they are visually indistinguishable to a human, but the attacked model fails. + +# 3.4 Limitations of existing editing-based stress-testing methods + +Recent advancements in diffusion modelling have drastically improved image editing. However, two prevalent approaches, LANCE [51] and DiffEdit [10], produce artefacts in medical images, making them unsuitable for stress-testing biomedical vision models. + +LANCE only uses a global prompt (no mask) for image editing. While effective in the natural image domain, it leads to artefacts in the biomedical domain. For example in Sec. 5.4, we add pathologies and support devices to images of healthy lungs to stress-test lung segmentation models; in such cases we must ensure that the position and shape of the lung borders are not altered during editing. However, we find that LANCE changes the position and shape of the lung border thus making edited images unsuitable for stress-testing such models (Sec. 5.5). In addition, we find that LANCE potentially removes support devices when prompted to remove pathologies, a direct effect of the correlations in the diffusion model's training datasets (Sec. 5.1), making LANCE unsuited for testing the robustness of models to manifestation shift. + +DiffEdit addresses these issues by editing only inside an automatically predicted mask $m_{\mathrm{edit}}$ . However, these predicted masks often mismatch the manually annotated ground-truth, especially for small and complex abnormalities like pneumothorax1 (see Sec. 5.5). Moreover, spurious correlations learned by the diffusion model can lead to the inclusion of support devices in the automatically predicted masks. Furthermore, even when relying on manually annotated masks, DiffEdit can introduce sharp discrepancies at mask boundaries, leading to unrealistic artefacts, such as when adding consolidation that should partially occlude the lung border (Fig. 10b in the Appendix). + +# 4 Method + +Our objective is to create synthetic test data through image editing that simulates specific data shifts, to rigorously evaluate biomedical imaging models. This synthetic data is used to predict model robustness, eliminating need for additional real-world test data. + +# 4.1 Improved editing with RadEdit + +To address the issues outlined in Sec. 3.4, we propose RadEdit: by introducing 'keep' and 'edit' masks into the editing process, RadEdit explicitly specifies which areas must remain unchanged (keep) and which should be actively modified based on the conditioning signal (edit). Crucially, these masks need not be mutually exclusive, allowing changes in the unmasked regions to ensure global consistency. Using masks, we assume that spurious correlations are mostly non-overlapping [44]. + +Algorithm 1 RadEdit (ours) uses multiple masks to decouple spurious correlations +Require: original image $x_0$ , inversion prompt $c_{\mathrm{inv}}$ , editing prompt c, edit mask medit, keep mask $m_{\mathrm{keep}}$ CFG weight w, diffusion model $f_{\theta}$ $(\hat{x}_{1:T},z_{1:T})\leftarrow$ DDPMINVERSION(xo,cinv) $\triangleright$ Encode image. Procedure in Appendix B + $x_{T}\gets \hat{x}_{T}$ +for $t\gets T$ to 1 do + $\epsilon_{\mathrm{cond},t}\gets f_{\theta}(xt,t,c)$ $\triangleright$ Predict conditional noise + $\epsilon_{\mathrm{uncond},t}\gets f_{\theta}(xt,t,c = \emptyset)$ $\triangleright$ Predict unconditional noise + $\epsilon_t\gets \epsilon_{\mathrm{uncond},t} + w(\epsilon_{\mathrm{cond},t} - \epsilon_{\mathrm{uncond},t})$ $\triangleright$ Combine noise predictions with CFG + $\epsilon_t\gets m_{\mathrm{edit}}\odot \epsilon_t + (1 - m_{\mathrm{edit}})\odot \epsilon_{\mathrm{uncond},t}$ $\triangleright$ Use CFG only within medit + $x_{t - 1}\gets \hat{\mu}_t(x_t,\epsilon_t) + \sigma_t z_t$ $\triangleright$ Move to next time step + $x_{t - 1}\gets m_{\mathrm{keep}}\odot \hat{x}_{t - 1} + (1 - m_{\mathrm{keep}})\odot x_{t - 1}$ $\triangleright$ Undo edits within mkeep +return edited version of $x_0$ + +RadEdit is detailed in Algorithm 1, where a number of key properties make RadEdit more suitable for biomedical image editing than prior editing methods. Firstly, since we aim to edit only within the edit mask $m_{\mathrm{edit}}$ , classifier-free guidance (CFG) [26] is used only within this region, with high guidance values (following [29], we use a value of 15) ensuring that pathologies are completely removed without drastically changing the rest of the image. This approach also simplifies choosing a prompt for editing since we do not have to take into account the effect of the prompt on the rest of the image. Secondly, we allow the area outside $m_{\mathrm{edit}}$ to be modified via unconditional generation to ensure image consistency. Lastly, the region of the keep mask $m_{\mathrm{keep}}$ is reverted to the encoding, ensuring that this region remains the same. Instead of initiating our generating process from pure noise we set $x_{T} = \hat{x}_{T}$ , where $\hat{x}_T$ is the last output of the DDPM inversion. + +In Fig. 3c, 10c, we show that RadEdit enables artefact-free editing while preserving structures of interest. Because the anatomical layout remains intact after editing, masks still correspond to the same structures, therefore the same masks can be reused to stress-test segmentation models (Sec. 5.4). In practice, we use a latent diffusion model [57], therefore all operations in Algorithm 1 are performed in the latent space of a variational autoencoder (VAE) [57]; this does not limit the generality of the approach. + +# 4.2 Using synthetic images to uncover bias + +Despite advancements in biomedical computer vision, recent studies have shown that bias in training and test data can lead to unrealistically high performance of machine learning models on the test set [59, 11]. In our experiments, we use RadEdit to create high quality synthetic test datasets that realistically capture specific dataset shifts, allowing us to quantify the robustness of models to these dataset shifts. By using masks, we can precisely edit the original training data to represent either acquisition shift, manifestation shift, or population shift [7] (Secs. 5.2 to 5.4). These synthetic test sets are used to stress-test (potentially biased) biomedical vision models by comparing performance to the real (biased) test set; a significant drop in performance indicates that the vision model is not robust to the dataset shift that can occur in clinical settings. This serves as a complementary tool to visual explainable AI tools like Grad-CAM [63] and saliency maps [66, 1], which offer qualitative insight into the robustness of models. + +# 4.3 BioViL-T editing score + +Since generative models result in samples of varying quality, poor-quality samples can be filtered out using image-text alignment scores, which quantitatively assess how closely related image-text pairs are via a pre-trained model that embeds similar images and text to nearby vectors [4, 54, 53, 14]. For image editing, we instead assess how similar the change in text and image embeddings are after editing: for a real image-text pair $(I_{\mathrm{real}}, T_{\mathrm{real}})$ , edited image-text pair $(I_{\mathrm{edit}}, T_{\mathrm{edit}})$ , image encoder $E_I$ , and text encoder $E_T$ , the editing score is defined based on directional similarity [17]: + +$$ +S _ {\mathrm {B i o V i L - T}} = \frac {\Delta I \cdot \Delta T}{\| \Delta I \| \| \Delta T \|}, \quad \text {w h e r e} \quad \begin{array}{l} \Delta I = E _ {I} \left(I _ {\mathrm {e d i t}}\right) - E _ {I} \left(I _ {\mathrm {r e a l}}\right), \text {a n d} \\ \Delta T = E _ {T} \left(T _ {\mathrm {e d i t}}\right) - E _ {T} \left(T _ {\mathrm {r e a l}}\right). \end{array} \tag {4} +$$ + +Given the focus on biomedical data, we use the BioViL-T [5] image and text encoders: domain-specific vision-language models trained to analyse chest X-rays and radiology reports, therefore well suited to measure changes in the edited image, such as removed pathologies. Following Prabhu et al. [51], we discard images with $S_{\mathrm{BioViL - T}} < 0.2$ . This is not only effective for filtering out poor quality edits but is also able to detect whether the original image $I_{\mathrm{real}}$ does not match the original text description $T_{\mathrm{real}}$ well. + +# 5 Experiments + +# 5.1 Diffusion model + +Our editing method is heavily dependent on a latent diffusion model [57] that can generate realistic chest X-rays. We use the VAE [37, 25] of SDXL [50] which can adequately reconstruct chest X-rays [8]. The VAE is frozen, and the denoising U-Net is trained on three datasets downsampled and centre-cropped to $512 \times 512$ pixels: MIMIC-CXR [33], ChestX-ray8 [75], and CheXpert [30], totalling 487680 training images. This data diversity allows us to perform zero-shot edits on datasets not seen during training. + +For MIMIC-CXR, we only include frontal view chest X-rays, and condition the denoising U-Net on the corresponding impression section in the radiology report (a short clinically actionable outline of the main findings). We employ the tokeniser and frozen text encoder from BioViL-T [5]. For ChestX-ray8 and CheXpert, we condition on a list of all abnormalities present in an image as indicated by the labels, e.g., 'Cardiomegaly. Pneumothorax'. If the list of abnormalities is empty, we use the string 'No findings'. More details on the datasets and diffusion model training can be found in Appendix E. + +# 5.2 Acquisition shift + +Background To show how RadEdit can be used to quantify the robustness of models to acquisition shift, we closely follow the experimental setup of DeGrave et al. [11], who show that deep learning systems built to detect COVID-19 from chest X-rays rely on confounding factors rather than pathology features. This problem arises when COVID-19-positive and -negative images come from disparate sources. In our setup, all COVID-19-positive cases come from the BIMCV dataset [73] (denoted BIMCV+), + +![](images/9111de6a974fbc3b2c5917a8c2007f88b2d20157edd2638845627203f073c19f.jpg) +Fig.2: Removing COVID-19 features with LANCE $^2$ (b) also changes the laterality markers and reduces contrast. In contrast, RadEdit (c; ours) preserves anatomical structures and laterality markers, and retains the original contrast. +(a) Original Image + +![](images/fc36787c107450cc0af8b9e69ae3edd6634b3472011cf568d7520bb5ec21c319.jpg) +(b) LANCE [51] + +![](images/354372db538e0480126b35eabce74802facc1997cc588ad2fe8ec5b211b46288.jpg) +(c) RadEdit (ours) + +and all COVID-19-negative cases from MIMIC-CXR [33] (see Fig. 11). A classifier trained on these datasets will rely on spurious features indicative of the data's origin, e.g., laterality markers or the field of view, instead of features caused by the pathology. + +Setup A synthetic test set is created by applying RadEdit to remove COVID-19 features1 from BIMCV+ images using the prompt 'No acute cardiopulmonary process'3 (Fig. 2); the included bounding boxes of COVID-19 features are used as the edit mask $m_{\mathrm{edit}}$ . Since this is the only mask available, we set the keep mask as $m_{\mathrm{keep}} = 1 - m_{\mathrm{edit}}$ . After filtering using the BioViL-T editing (Sec. 4.3), this results in a synthetic dataset of 2774 COVID-19-negative images containing the same spurious features as BIMCV+. Both the weak and strong predictor models are classifiers trained using a ResNet50 architecture (see Appendix F for more implementation details). + +Findings Tab. 1, shows the performance of a COVID-19 classifier (weak predictor) trained on BIMCV+ and MIMIC-CXR. In accordance with DeGrave et al. [11], we find that the weak predictor performs exceptionally well on the real test set (i.e. test splits of both datasets) since the model learned to distinguish the two data sources instead of learning visual features related to COVID-19. However, in the second row of Tab. 1, we see a drop of $95\%$ in accuracy meaning that the model fails to classify the synthetic images as COVID-19-negative. The weak predictor is not robust to a shift in acquisition. + +To show that the decreased performance of the weak predictor is not caused by artefacts in the edited images, we train a more robust COVID-19 classifier (strong predictor), using the BIMCV+ and BIMCV- datasets, as in [11], where the BIMCV- dataset consists of only COVID-19-negative cases from BIMCV, and test on the same two test + +Table 1: Quantifying robustness of COVID-19 detectors to acquisition shift. We train a weak predictor on the 'Biased' dataset—a combination of BIMCV+ [73] and MIMIC-CXR [33]; and a strong predictor on an unbiased dataset—a combination of BIMCV+ and BIMCV; the 'Synthetic' test set consists of 2774 COVID-19-negative images with the same spurious features as the BIMCV+ datasets, e.g. laterality markers. We report mean accuracy and standard deviation across 5 runs. + +
PredictorTest dataAccuracy
WeakBiased99.1 ± 0.2
WeakSynthetic5.5 ± 2.1
StrongBiased74.4 ± 3.0
StrongSynthetic76.0 ± 7.7
+ +![](images/d8b08b200a2c06a0e26d7200596a7eaa7c8afb6de45548faa4a8189e7973fe82.jpg) +Fig.3: Removing pneumothorax (red) with LANCE $^2$ (b) also removes the spuriously correlated chest drain (blue) and reduces contrast. In contrast, RadEdit (c; ours) preserves the chest drain and better preserves anatomical structures. +(a) Original Image + +![](images/fc158dd38debd7237738a6c816ded24d5d19c7a6cf421d03d19ba7f44547e67f.jpg) +(b) LANCE [51] + +![](images/bb6ce92710ee1361641cde70c894f33ba86a2916f4f05e6cf11783410a8f5ebf.jpg) +(c) RadEdit (ours) + +datasets. Comparing rows one and three of Tab. 1, we find that the strong predictor performs worse on the test set containing samples from BIMCV+ and MIMIC-CXR than the weak predictor (row one). This is expected as the strong predictor relies on actual pathology features. Lastly, rows three and four of Tab. 1 show that the strong predictor performs similarly on the real and synthetic test sets, attesting to the quality of our edits. + +# 5.3 Manifestation shift + +Background In this section, we show how RadEdit can be used to quantify the robustness of biomedical vision models to manifestation shift. We closely follow the experimental setup of Rueckel et al. [59], who demonstrate that pneumothorax classification models are strongly biased by the presence of chest drains: while the average performance of pneumothorax classifiers is high, performance on the subset of images with a chest drain but no pneumothorax is significantly lower. This is due to chest drains being a common treatment for pneumothorax, resulting in the majority of images in datasets like CANDID-PTX [13] containing a chest drain only if there is a pneumothorax. As a result, only $1\%$ of images in CANDID-PTX contain a chest drain but no pneumothorax. + +Setup We use RadEdit to create a synthetic dataset containing images with chest drains but no pneumothorax, by editing out the pneumothorax from CANDID-PTX images using the prompt 'No acute cardiopulmonary process' (Fig. 3). The edit mask $m_{\mathrm{edit}}$ is set as a mask of the pneumothorax, and the keep mask $m_{\mathrm{keep}}$ set as the chest drain mask. This ensures that the chest drain is not removed, while preventing border artefacts. After filtering using the BioViL-T editing score (Sec. 4.3), 628 images are left; in contrast, the real test set contains only 16 of cases with drains but no pneumothorax. The weak predictor is a ResNet50 trained to classify pneumothorax. Following Rueckel et al. [59], the strong predictor is an EfficientNet U-Net [69] trained on SIIM-ACR [77] to both segment pneumothorax and classify pneumothorax (more details in Appendix G). + +Findings In accordance with [59], we show in Tab. 2 that a pneumothorax classifier (weak predictor) trained on CANDID-PTX performs exceptionally well on the test split of CANDID-PTX, since very few images contain a chest drain and no pneumothorax. However, in row two of Tab. 2, we show a drastic drop in performance on the synthetic test set, i.e., the weak predictor is not robust to manifestation shift. To show that the drop in performance on the synthetic dataset does not come from editing artefacts, we evaluate the more robust segmentation model (strong predictor) on the same test datasets + +Table 2: Quantifying robustness of pneumothorax detectors to manifestation shift. The weak predictor is trained on the biased CANDID-PTX [13] dataset to classify pneumothorax; the strong predictor is trained on SIIM-ACR [77] to classify and segment the pneumothorax. Real 'Biased' test data comes from CANDID-PTX which exhibits strong confounding between the pneumothorax and chest tubes; 'Synthetic' test data is 629 solely edited images containing chest drains but no pneumothorax. We report mean accuracy and standard deviation across 5 runs. + +
PredictorTest dataAccuracy
WeakBiased93.3 ± 0.6
WeakSynthetic17.9 ± 3.7
StrongBiased93.7 ± 1.3
StrongSynthetic81.7 ± 7.1
+ +Rows three and four of Tab. 2), we find that the strong predictor performs on par with the weak predictor in row one; however, the strong predictor closes the majority of the gap between the real test set and the synthetic one, attesting to the quality of our edits. In agreement with Rueckel et al. [59], there is still a performance gap, indicating that the strong predictor still suffers from mild manifestation shift. + +# 5.4 Population shift + +Background In this section, we show how RadEdit can be used to quantify the robustness of lung segmentation models to population shifts. Manually segmenting X-ray images is labour intensive and requires high expertise, leading to small datasets often limited to single pathologies or healthy patients [65, 31], e.g., MIMIC-Seg [9]. These models are thus sensitive to occlusions such as medical devices or pathologies, which typically appear as white regions on X-rays [43]. Evaluating model robustness requires further image collection for each occlusion type, which is time-consuming and costly. + +Setup Here, abnormalities are added to the lung region in healthy X-rays from MIMIC-Seg (Fig. 4). The edit mask $m_{\mathrm{edit}}$ is set as a mask of the lung(s). When editing a single + +![](images/d1509fd8ad578531c7ac80d7555bd446cc49b7e0b76c72db5475b17b150e4d51.jpg) +Fig.4: Adding pulmonary edema (top), pacemakers (middle), and consolidation (bottom) with RadEdit. The 'strong predictor' (d), a segmentation model trained on CheXmask [16] (a large dataset containing various abnormalities) is more robust to these abnormalities than the 'weak predictor' (c), a segmentation model trained on MIMIC-Seg [9] (a small set of mostly healthy patients): the weak predictor traces around the pacemaker and poorly annotates the consolidated lung. Blue: ground-truth annotation; red: predicted segmentation. + +Table 3: Quantifying robustness of lung segmentation models to population shift. The 'weak predictor' is trained on MIMIC-Seg (a small set of predominantly healthy patients); the 'strong predictor' is trained on CheXmask (a large mixed set of patients with various abnormalities). Synthetic test data is created by using RadEdit to add edema, pacemakers, and consolidation. We report the change $(\Delta)$ in Dice score and AHD with respect to the segmentation models evaluated on the ground-truth test set. + +
Test dataWeak PredictorStrong Predictor
Dice↑Δ↓AHD↓Δ↓Dice↑Δ↓AHD↓Δ↓
Real data97.4—6.1—95.5—11.6—
Healthy → edema93.83.621.815.793.91.622.811.2
Healthy → pacemaker85.012.449.843.787.38.229.517.9
Healthy → consolidation85.911.544.138.188.17.429.417.8
+ +lung, the keep mask $m_{\mathrm{keep}}$ corresponds to the lung which must not change, while when editing both lungs we set $m_{\mathrm{keep}} = 0$ to allow opacity adjustments, or for elements to be added outside of the lungs. Stress-test sets are generated for three abnormalities: pulmonary edema, pacemakers, and consolidation1. Prompts are phrased to match similar impressions in the training data (see Appendix H). Both predictors are EfficientNet U-Net [69] models; the weak predictor is trained on MIMIC-Seg, while the strong predictor is trained on CheXmask [16], a larger dataset with various lung abnormalities (more details in Appendix H). We evaluate segmentation quality using Dice similarity coefficient, which is the harmonic mean of the precision and recall, and 95th percentile average Hausdorff distance (AHD), a measure of the distance between two sets [45]. + +Findings Tab. 3 shows that the weak predictor model performs well on the real biased test data, since it is mostly composed of healthy subjects. However, testing on the synthetic lung abnormality datasets (rows two to four), causes performance to drop substantially, i.e. the weak predictor is not robust to population shift. To show that this drop in performance does not come from editing artefacts, we evaluate the strong predictor on the synthetic test sets and see considerably smaller changes in performance. This can be seen in Fig. 4: for pulmonary edema, both models can accurately segment, despite the abnormality; for pacemakers, the weak predictor incorrectly segments around the pacemakers, while the strong predictor more accurately segments the lungs; and for consolidation, both models are less able to segment the lungs accurately, however, the strong predictor gets closer to the ground-truth. See Appendix H for more examples. + +# 5.5 Quantifying the limitations of existing editing methods + +LANCE As seen in the second row of Tab. 3, adding edema leads only to a small drop in performance of the strong predictor. We hypothesise that further drops in performances stem from a mismatch of the original mask and the edited images. We therefore use this setup to quantify how well LANCE and RadEdit preserve the shape and position of the lung borders. Additionally, we study the difference between results using + +![](images/375fc1d5fb456937dfea497ad62d2705686a2b443a6a0a34aa4834acdb4f4e09.jpg) +(a) + +![](images/d4bd5f26608b23783c54424f27891562ec463a3ab4e5603f08ea6e297c24f377.jpg) +(b) + +![](images/9f465dfb4e44cb0f02d1cd45eb24d43345bf71b88185e79f866293b897d8461f.jpg) +(c) +Fig. 5: Comparison of LANCE $^2$ and RadEdit. We measure how well the strong predictor from Tab. 3's outputs matches the ground-truth lung masks (blue) for four synthetic datasets created by adding edema using LANCE and RadEdit with DDIM or DDPM inversion. High Dice / low AHD indicates that the editing method well preserves the lung border's location and shape. + +![](images/1b3c5ac8ffcec0dc13caeb3a545a3d2c7b44d250909b74a47e72414330f29ad5.jpg) +(d) + +![](images/af6befdacdf2b5fb7813977631443e3169b0f3008a4cd3cd061e61b24aa5cff7.jpg) +(e) + +
Editing MethodDice↑AHD↓
(a) Original data95.511.6
(b) LANCE w/ DDIM78.965.1
(c) LANCE w/ DDPM80.169.5
(d) RadEdit w/ DDIM86.239.8
(e) RadEdit w/ DDPM93.922.8
+ +DDIM or DDPM inversion. For all four methods in Fig. 5, we use the same setup as in Sec. 5.4: we first edit the original image with the prompt 'Moderate pulmonary edema. The heart size is normal', and then compare the outputs of the strong predictor with the original ground-truth lung masks. Here, we find that using masks and DDPM inversion is necessary for RadEdit to preserve the shape and position of the lung border. + +DiffEdit We quantify how well DiffEdit's automatically predicted masks match the manual ground-truth using the same setup as in Sec. 5.3: we take an image containing pneumothorax and a chest drain, and try to remove only the pneumothorax. The editing prompt is created by keeping only the parts of the impressions related to pneumothorax and chest drains, and replacing the description of the pneumothorax with 'No pneumothorax'. DiffEdit should therefore predict a mask containing only the pneumothorax. We perform a grid search on the MIMIC-Seg [9] validation set over DiffEdit's hyperparameters (noise strength and binarising threshold) to optimise pneumothorax segmentation metrics, then evaluate on the training set. In Fig. 6 we see that DiffEdit's predicted masks obtain poor quantitative metrics where parts of the pneumothorax are often missing, and the spuriously correlated chest drain is often included in the predicted mask. As a result, DiffEdit's predicted masks are unsuitable for stress-testing. + +![](images/7f9e7b069302971b2be8f0e75a6ff32d5734c9ba94f8c8563f98d4282702d961.jpg) +(a) Examples of pneumothorax masks predicted using DiffEdit [10]. Blue: ground-truth annotation; red: predicted editing mask. + +![](images/9a69c7837b7f1cfa312e9da9f6d9a67e5d85557e42ec66badfe8ccd74cad0ccb.jpg) + +![](images/b3f352f725da320efcb78cb78f5944d6971e44d5b73a91c410a4c774baa79c52.jpg) + +# Hyperparameters Dice AHD + +Tuned per image 33.8 97.7 + +Tuned on validation 18.4 256.8 + +(b) Segmentation metrics for the pneumothorax mask predicted by DiffEdit [10], for hyperparameters tuned on the validation set (bottom) and tuned per image (top; which requires ground-truth masks). + +Fig. 6: Evaluating pneumothorax masks predicted using DiffEdit [10]. (a) Predicted masks (red) are noisy, with chest drains often incorrectly segmented as well as or instead of the pneumothorax (blue); (b) this is demonstrated quantitatively with low Dice score and high AHD. + +# 6 Limitations and future work + +Despite the encouraging results presented in the paper, RadEdit is not without limitations and more work is needed to extend it to more applications. Currently, training datasets and models must be manually analysed to predict potential failure cases, simulate these failures to test the hypothesis, and finally quantitatively evaluate the model; future work could automate such failure mode discovery. Another limitation is that current editing techniques do not enable all types of stress-testing; for example, with current approaches, we are unable to test segmentation models' behaviour to cardiomegaly (enlarged heart) since this would require segmentation maps to be adjusted after editing. However, this could potentially be enabled by enlarging heart segmentations to simulate cardiomegaly and adjusting the ground-truth lung segmentation accordingly. + +When using generative editing, it is not possible to guarantee that unwanted changes will not occur. With RadEdit, we minimise this by forcing spuriously correlated regions to remain the same, only using classifier-free guidance within the editing mask, and filtering via image-text alignment. Future work to better maintain structure when editing will help with this issue, but masks will still be necessary to bypass spurious correlations. Furthermore, due to potential overlap of 3D structures, editing 2D X-rays can be limited. In such cases, editing single structures using RadEdit may not be successful. To address this, 3D CT images could be edited, then projected to synthetic X-rays [18]. + +When producing simulated stress test sets, several factors affect edit quality including classifier-free guidance weight, number of inference steps, and time step to encode to. Additionally, the text encoder must well understand specified pathologies to provide informative features to condition the generative model on; similarly, the diffusion model must be able to capture fine details and well cover the data distribution. + +Finally, more research is required to develop better approaches for quantifying edit quality for downstream tasks. In particular, observing a change in downstream performance is not necessarily indicative of real-world performance as edit quality may be poor. While the introduced BioViL-T editing score can be used to quantify edit quality, this introduces reliance on a potentially biased model. Additionally, the BioViL-T editing score is not suited to detect the artefacts introduced by LANCE and DiffEdit. + +# 7 Conclusion + +In this study, we illustrate the efficacy of generative image editing as a robust tool for stress-testing biomedical vision models. Our focus is on assessing their robustness against three types of dataset shifts commonly encountered in biomedical imaging: acquisition shift, manifestation shift, and population shift. We highlight that one of the significant challenges in biomedical image editing is the correlations learned by the generative model, which can result in artefacts during the editing process. To mitigate these artefacts, RadEdit relies on various types of masks to restrict the effects of the editing to certain areas while ensuring the consistency of the edited images. This approach enables us to generate high fidelity synthetic test sets that exhibit common dataset shifts. These synthetic test sets are used to identify and quantify the failure modes of biomedical classification and segmentation models. This provides a valuable supplement to explainable AI approaches such as Grad-CAM [63] and saliency maps [66, 1]. + +# Bibliography + +[1] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. Advances in neural information processing systems, 31, 2018. +[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. +[3] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM Transactions on Graphics (TOG), 42(4):1-11, 2023. +[4] Samaneh Azadi, Catherine Olsson, Trevor Darrell, Ian Goodfellow, and Augustus Odena. Discriminator rejection sampling. arXiv preprint arXiv:1810.06758, 2018. +[5] Shruthi Bannur, Stephanie Hyland, Qianchu Liu, Fernando Perez-Garcia, Maximilian Ilse, Daniel C. Castro, Benedikt Boecking, Harshita Sharma, Kenya Bouzid, Anja Thieme, Anton Schwaighofer, Maria Wetscherek, Matthew P. Lungren, Aditya Nori, Javier Alvarez-Valle, and Ozan Oktay. Learning to Exploit Temporal Structure for Biomedical Vision-Language Processing. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 1 2023. https://doi.org/10.48550/arxiv.2301.04558. +[6] Riccardo Barbano, Alexander Denker, Hyungjin Chung, Tae Hoon Roh, Simon Arridge, Peter Maass, Bangti Jin, and Jong Chul Ye. Steerable conditional diffusion for out-of-distribution adaptation in imaging inverse problems. arXiv preprint arXiv:2308.14409, 2023. +[7] Daniel C Castro, Ian Walker, and Ben Glocker. Causality matters in medical imaging. Nature Communications, 11(1):3673, 2020. +[8] Pierre Chambon, Christian Bluethgen, Jean-Benoit Delbrouck, Rogier Van der Sluijs, Malgorzata Polacin, Juan Manuel Zambrano Chaves, Tanishq Mathew Abraham, Shivanshu Purohit, Curtis P Langlotz, and Akshay Chaudhari. Roentgen: vision-language foundation model for chest x-ray generation. arXiv preprint arXiv:2211.12737, 2022. +[9] Li-Ching Chen, Po-Chih Kuo, Ryan Wang, Judy Gichoya, and Leo Anthony Celi. Chest X-ray segmentation images based on MIMIC-CXR (version 1.0.0). PhysioNet, 2022. +[10] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. DiffEdit: Diffusion-based semantic image editing with mask guidance, 2022. +[11] Alex J. DeGrave, Joseph D. Janizek, and Su-In Lee. AI for radiographic COVID-19 detection selects shortcuts over signal. Nature Machine Intelligence, 3(7):610-619, 2021. ISSN 2522-5839. https://doi.org/10.1038/s42256-021-00338-7. +[12] Alexey Dosovitskiy, Jost Tobias Springenberg, and Thomas Brox. Learning to generate chairs with convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1538-1546, 2015. +[13] Sijing Feng, Damian Azzollini, Ji Soo Kim, Cheng-Kai Jin, Simon P. Gordon, Jason Yeoh, Eve Kim, Mina Han, Andrew Lee, Aakash Patel, Joy Wu, Martin + +Urschler, Amy Fong, Cameron Simmers, Gregory P. Tarr, Stuart Barnard, and Ben Wilson. Curation of the CANDID-PTX dataset with free-text reports. *Radiology: Artificial Intelligence*, 3(6):e210136, 2021. ISSN 2638-6100. https://doi.org/10.1148/ryai.2021210136. +[14] Virginia Fernandez, Pedro Sanchez, Walter Hugo Lopez Pinaya, Grzegorz Jacenków, Sotirios A Tsaftaris, and Jorge Cardoso. Privacy distillation: Reducing re-identification risk of multimodal diffusion models. arXiv preprint arXiv:2306.01322, 2023. +[15] Alessandro Fontanella, Grant Mair, Joanna Wardlaw, Emanuele Trucco, and Amos Storkey. Diffusion models for counterfactual generation and anomaly detection in brain images. URL http://arxiv.org/abs/2308.02062. +[16] Nicolas Gaggion, Candelaria Mosquera, Lucas Mansilla, Martina Aineseder, Diego H Milone, and Enzo Ferrante. CheXmask: a large-scale dataset of anatomical segmentation masks for multi-center chest X-ray images. arXiv preprint arXiv:2307.03293, 2023. +[17] Rinon Gal, Or Patashnik, Haggai Maron, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. StyleGAN-NADA: CLIP-guided domain adaptation of image generators. ACM Transactions on Graphics, 41(4), 2022. https://doi.org/10.1145/3528223.3530164. +[18] Cong Gao, Benjamin D Killeen, Yicheng Hu, Robert B Grupp, Russell H Taylor, Mehran Armand, and Mathias Unberath. Synthetic data accelerates the development of generalizable learning-based algorithms for x-ray image analysis. Nature Machine Intelligence, 5(3):294-308, 2023. +[19] Timur Garipov, Sebastiaan De Peuter, Ge Yang, Vikas Garg, Samuel Kaski, and Tommi Jaakkola. Compositional sculpting of iterative generative processes. arXiv preprint arXiv:2309.16115, 2023. +[20] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples, 2015. +[21] Yu Gu, Jianwei Yang, Naoto Usuyama, Chunyuan Li, Sheng Zhang, Matthew P Lungren, Jianfeng Gao, and Hoifung Poon. Biomedjourney: Counterfactual biomedical image generation by instruction-learning from multimodal patient journeys. arXiv preprint arXiv:2310.10765, 2023. +[22] Will Douglas Heaven. Hundreds of ai tools have been built to catch covid. none of them helped. MIT Technology Review. Retrieved December 2023, 2021. +[23] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2018. +[24] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations, 2022. +[25] Irina Higgins, Loic Matthew, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. $\beta$ -VAE: Learning basic visual concepts with a constrained variational framework. In International Conference on Learning Representations, 2016. +[26] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. + +[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020. +[28] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research, 23(1):2249-2281, 2022. +[29] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly DDPM noise space: Inversion and manipulations, 2023. +[30] Jeremy Irvin, Pranav Rajpurkar, Michael Ko, Yifan Yu, Silviana Ciurea-Ilicus, Chris Chute, Henrik Marklund, Behzad Haghgoo, Robyn Ball, Katie Shpanskaya, Jayne Seekins, David A. Mong, Safwan S. Halabi, Jesse K. Sandberg, Ricky Jones, David B. Larson, Curtis P. Langlotz, Bhavik N. Patel, Matthew P. Lungren, and Andrew Y. Ng. CheXpert: A large chest radiograph dataset with uncertainty labels and expert comparison, 2019. +[31] Stefan Jaeger, Sema Candemir, Sameer Antani, Yi-Xiang J Wang, Pu-Xuan Lu, and George Thoma. Two public chest x-ray datasets for computer-aided screening of pulmonary diseases. Quantitative imaging in medicine and surgery, 4(6):475, 2014. +[32] Priyank Jaini, Kevin Clark, and Robert Geirhos. Intriguing properties of generative classifiers. arXiv preprint arXiv:2309.16779, 2023. +[33] Alistair E. W. Johnson, Tom J. Pollard, Seth J. Berkowitz, Nathaniel R. Greenbaum, Matthew P. Lungren, Chih-ying Deng, Roger G. Mark, and Steven Horng. MIMIC-CXR, a de-identified publicly available database of chest radiographs with free-text reports. Scientific Data, 6(1):317, 2019. ISSN 2052-4463. https://doi.org/10.1038/s41597-019-0322-0. +[34] Charles Jones, Daniel C Castro, Fabio De Sousa Ribeiro, Ozan Oktay, Melissa McCradden, and Ben Glocker. No fair lunch: A causal perspective on dataset bias in machine learning for medical imaging. arXiv preprint arXiv:2307.16526, 2023. +[35] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10124–10134, 2023. +[36] Priyatham Kattakinda, Alexander Levine, and Soheil Feizi. Invariant learning via diffusion dreamed distribution shifts. arXiv preprint arXiv:2211.10370, 2022. +[37] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes, 2022. +[38] Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas Phillips, Irena Gao, et al. Wilds: A benchmark of in-the-wild distribution shifts. In International Conference on Machine Learning, pages 5637-5664. PMLR, 2021. +[39] Ira Ktena, Olivia Wiles, Isabela Albuquerque, Sylvestre-Alvise Rebuffi, Ryutaro Tanno, Abhijit Guha Roy, Shekoofeh Azizi, Danielle Belgrave, Pushmeet Kohli, Alan Karthikesalingam, et al. Generative models improve fairness of medical classifiers under distribution shifts. arXiv preprint arXiv:2304.09218, 2023. +[40] Agostina J Larrazabal, Nicolás Nieto, Victoria Peterson, Diego H Milone, and Enzo Ferrante. Gender imbalance in medical imaging datasets produces biased classifiers for computer-aided diagnosis. Proceedings of the National Academy of Sciences, 117(23):12592–12594, 2020. + +[41] Choong Ho Lee and Hyung-Jin Yoon. Medical big data: promise and challenges. *Kidney research and clinical practice*, 36(1):3, 2017. +[42] Xiaodan Li, Yuefeng Chen, Yao Zhu, Shuhui Wang, Rong Zhang, and Hui Xue. ImageNet-e: Benchmarking neural network robustness via attribute editing, 2023. +[43] Wufeng Liu, Jiaxin Luo, Yan Yang, Wenlian Wang, Junkui Deng, and Liang Yu. Automatic lung segmentation in chest x-ray images using improved u-net. Scientific Reports, 12(1):8649, 2022. +[44] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. Advances in Neural Information Processing Systems, 33:11525-11538, 2020. +[45] Lena Maier-Hein, Annika Reinke, Patrick Godau, Minu D. Tizabi, Florian Buettner, Evangelia Christodoulou, Ben Glocker, Fabian Isensee, Jens Kleesiek, Michal Kozubek, Mauricio Reyes, Michael A. Riegler, Manuel Wiesenfarth, A. Emre Kavur, Carole H. Sudre, Michael Baumgartner, Matthias Eisenmann, Doreen Heckmann-Nötzel, A. Tim Rädsch, Laura Acion, Michela Antonelli, Tal Arbel, Spyridon Bakas, Arrielenis, Matthew Blaschko, M. Jorge Cardoso, Veronika Cheplygina, Beth A. Cimini, Gary S. Collins, Keyvan Farahani, Luciana Ferrer, Adrian Galdran, Bram van Ginneken, Robert Haase, Daniel A. Hashimoto, Michael M. Hoffman, Merel Huisman, Pierre Jannin, Charles E. Kahn, Dagmar Kainmueller, Bernhard Kainz, Alexandros Karargyris, Alan Karthikesalingam, Hannes Kenngott, Florian Kofler, Annette Kopp-Schneider, Anna Kreshuk, Tahsin Kurc, Bennett A. Landman, Geert Litjens, Amin Madani, Klaus Maier-Hein, Anne L. Martel, Peter Mattson, Erik Meijering, Bjoern Menze, Karel G. M. Moons, Henning Müller, Brennan Nichyporuk, Felix Nickel, Jens Petersen, Nasir Rajpoot, Nicola Rieke, Julio Saez-Rodriguez, Clara I. Sánchez, Shravya Shetty, Maarten van Smeden, Ronald M. Summers, Abdel A. Taha, Aleksei Tiulpin, Sotirios A. Tsaftaris, Ben Van Calster, Gaël Varoquaux, and Paul F. Jäger. Metrics reloaded: Recommendations for image analysis validation, 2023. +[46] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations, 2022. +[47] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Nulltext inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. +[48] Gustav Müller-Franzes, Jan Moritz Niehues, Firas Khader, Soroosh Tayebi Arasteh, Christoph Haarburger, Christiane Kuhl, Tianci Wang, Tianyu Han, Sven Nebelung, Jakob Nikolas Kather, et al. Diffusion probabilistic models beat gans on medical images. arXiv preprint arXiv:2212.07501, 2022. +[49] Nick Pawlowski, Daniel C. Castro, and Ben Glocker. Deep structural causal models for tractable counterfactual inference. In Advances in Neural Information Processing Systems, volume 33, pages 857-869, 2020. +[50] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis, 2023. + +[51] Viraj Prabhu, Sriram Yenamandra, Prithvijit Chattopadhyay, and Judy Hoffman. LANCE: Stress-testing visual models by generating language-guided counterfactual images, 2023. +[52] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015. +[53] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. +[54] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. +[55] Jacob C. Reinhold, Aaron Carass, and Jerry L. Prince. A structural causal model for MR images of multiple sclerosis. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2021, volume 12905 of LNCS, pages 782–792, 2021. https://doi.org/10.1007/978-3-030-87240-3_75. +[56] Michael Roberts, Derek Driggs, Matthew Thorpe, Julian Gilbey, Michael Yeung, Stephan Ursprung, Angelica I Aviles-Rivero, Christian Etmann, Cathal McCague, Lucian Beer, et al. Common pitfalls and recommendations for using machine learning to detect and prognosticate for Covid-19 using chest radiographs and ct scans. Nature Machine Intelligence, 3(3):199-217, 2021. +[57] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. +[58] Johannes Rueckel, Lena Trappmann, Balthasar Schachtner, Philipp Wesp, Boj Friedrich Hoppe, Nicola Fink, Jens Ricke, Julien Dinkel, Michael Ingrisch, and Bastian Oliver Sabel. Impact of confounding thoracic tubes and pleural dehiscence extent on artificial intelligence pneumothorax detection in chest radiographs. Investigative Radiology, 55(12):792-798, July 2020. ISSN 0020-9996. https://doi.org/10.1097/rli.0000000000000707. +[59] Johannes Rueckel, Christian Huemmer, Andreas Fieselmann, Florin-Cristian Gheu, Awais Mansoor, Balthasar Schachtner, Philipp Wesp, Lena Trappmann, Basel Munawwar, Jens Ricke, Michael Ingrisch, and Bastian O. Sabel. Pneumothorax detection in chest radiographs: optimizing artificial intelligence system for accuracy and confounding bias reduction using in-image annotations in algorithm training. European Radiology, 31(10):7888-7900, 2021. https://doi.org/10.1007/s00330-021-07833-w. +[60] Chitwan Sahara, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models, 2022. +[61] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Semantic foggy scene understanding with synthetic data. International Journal of Computer Vision, 126: 973-992, 2018. +[62] Pedro Sanchez, Antanas Kascenas, Xiao Liu, Alison Q O'Neil, and Sotirios A Tsaftaris. What is healthy? generative counterfactual diffusion for lesion localization + +tion. In MICCAI Workshop on Deep Generative Models, pages 34-44. Springer, 2022. +[63] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. +[64] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9243–9252, 2020. +[65] Junji Shiraishi, Shigehiko Katsuragawa, Junpei Ikezoe, Tsuneo Matsumoto, Takeshi Kobayashi, Ken-ichi Komatsu, Mitate Matsui, Hiroshi Fujita, Yoshie Kodera, and Kunio Doi. Development of a digital image database for chest radiographs with and without a lung nodule: receiver operating characteristic analysis of radiologists' detection of pulmonary nodules. American Journal of Roentgenology, 174(1):71-74, 2000. +[66] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034, 2013. +[67] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. +[68] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models, 2022. +[69] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. +[70] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. +[71] Paul Upchurch, Jacob Gardner, Geoff Pleiss, Robert Pless, Noah Snavely, Kavita Bala, and Kilian Weinberger. Deep feature interpolation for image content changes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7064-7073, 2017. +[72] Boris Van Breugel, Nabeel Seedat, Fergus Imrie, and Mihaela van der Schaar. Can you rely on your model evaluation? improving model evaluation with synthetic test data, 2023. +[73] Maria de la Iglesia Vayá, Jose Manuel Saborit, Joaquim Angel Montell, Antonio Pertusa, Aurelia Bustos, Miguel Cazorla, Joaquin Galant, Xavier Barber, Domingo Orozco-Beltrán, Francisco García-García, Marisa Caparrós, Germán González, and Jose María Salinas. BIMCV COVID-19+: a large annotated dataset of RX and CT images from COVID-19 patients, 2020. version: 3. +[74] I von Borzyskowski, A Mazumder, B Mateen, and M Wooldridge. Data science and ai in the age of covid-19, 2021. +[75] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M. Summers. ChestX-ray8: Hospital-scale chest x-ray database and + +benchmarks on weakly-supervised classification and localization of common thorax diseases. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3462-3471, 2017. https://doi.org/10.1109/CVPR.2017.369. + +[76] Laure Wynants, Ben Van Calster, Gary S Collins, Richard D Riley, Georg Heinze, Ewoud Schuit, Elena Albu, Banafsheh Arshi, Vanessa Bellou, Marc MJ Bonten, et al. Prediction models for diagnosis and prognosis of Covid-19: systematic review and critical appraisal. bmj, 369, 2020. +[77] Anna Zawacki, Carol Wu, George Shih, Julia Elliott, Mikhail Fomitchev, Mohannad ParasLakhani Hussain, Phil Culliton, and Shunxing Bao. Siim-acr pneumothorax segmentation, 2019. \ No newline at end of file diff --git a/2024/RadEdit_ stress-testing biomedical vision models via diffusion image editing/images.zip b/2024/RadEdit_ stress-testing biomedical vision models via diffusion image editing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ce71c865300c9e50909a70bef33b3943934489d8 --- /dev/null +++ b/2024/RadEdit_ stress-testing biomedical vision models via diffusion image editing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8900b3170d3f2af61ee2d74e79fabeced354f8a472b525d4913deb9dda137c9 +size 271921 diff --git a/2024/RadEdit_ stress-testing biomedical vision models via diffusion image editing/layout.json b/2024/RadEdit_ stress-testing biomedical vision models via diffusion image editing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a1b265e16e5fa15f77424ca08e5f540223224294 --- /dev/null +++ b/2024/RadEdit_ stress-testing biomedical vision models via diffusion image editing/layout.json @@ -0,0 +1,10945 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 112, + 467, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 112, + 467, + 148 + ], + "spans": [ + { + "bbox": [ + 146, + 112, + 467, + 148 + ], + "type": "text", + "content": "RadEdit: stress-testing biomedical vision models via diffusion image editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 139, + 168, + 474, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 168, + 474, + 217 + ], + "spans": [ + { + "bbox": [ + 139, + 168, + 474, + 217 + ], + "type": "text", + "content": "Fernando Pérez-García\\*,1, Sam Bond-Taylor\\*,1, Pedro P. Sanchez+,2, Boris van Breugel+,3, Daniel C. Castro1, Harshita Sharma1, Valentina Salvatelli1, Maria T.A. Wetscherek1, Hannah Richardson1, Matthew P. Lungren1,4,5, Aditya Nori1, Javier Alvarez-Valle1, Ozan Oktay†,1, and Maximilian Ilse†,1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 255, + 225, + 357, + 291 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 255, + 225, + 357, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 225, + 357, + 236 + ], + "spans": [ + { + "bbox": [ + 255, + 225, + 357, + 236 + ], + "type": "text", + "content": "1 Microsoft Health Futures" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 258, + 237, + 356, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 237, + 356, + 247 + ], + "spans": [ + { + "bbox": [ + 258, + 237, + 356, + 247 + ], + "type": "text", + "content": "2 University of Edinburgh" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 258, + 247, + 356, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 247, + 356, + 258 + ], + "spans": [ + { + "bbox": [ + 258, + 247, + 356, + 258 + ], + "type": "text", + "content": "3 University of Cambridge" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 258, + 259, + 355, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 259, + 355, + 270 + ], + "spans": [ + { + "bbox": [ + 258, + 259, + 355, + 270 + ], + "type": "text", + "content": "4 University of California" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 267, + 270, + 347, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 270, + 347, + 281 + ], + "spans": [ + { + "bbox": [ + 267, + 270, + 347, + 281 + ], + "type": "text", + "content": "5 Stanford University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 270, + 281, + 345, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 281, + 345, + 291 + ], + "spans": [ + { + "bbox": [ + 270, + 281, + 345, + 291 + ], + "type": "text", + "content": "* Shared first author" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 231, + 292, + 382, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 292, + 382, + 302 + ], + "spans": [ + { + "bbox": [ + 231, + 292, + 382, + 302 + ], + "type": "text", + "content": "+ Work done at Microsoft Health Futures" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 272, + 303, + 343, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 303, + 343, + 313 + ], + "spans": [ + { + "bbox": [ + 272, + 303, + 343, + 313 + ], + "type": "text", + "content": "† Shared last author" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 159, + 346, + 453, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 346, + 453, + 511 + ], + "spans": [ + { + "bbox": [ + 159, + 346, + 453, + 511 + ], + "type": "text", + "content": "Abstract. Biomedical imaging datasets are often small and biased, meaning that real-world performance of predictive models can be substantially lower than expected from internal testing. This work proposes using generative image editing to simulate dataset shifts and diagnose failure modes of biomedical vision models; this can be used in advance of deployment to assess readiness, potentially reducing cost and patient harm. Existing editing methods can produce undesirable changes, with spurious correlations learned due to the co-occurrence of disease and treatment interventions, limiting practical applicability. To address this, we train a text-to-image diffusion model on multiple chest X-ray datasets and introduce a new editing method, RadEdit, that uses multiple image masks, if present, to constrain changes and ensure consistency in the edited images, minimising bias. We consider three types of dataset shifts: acquisition shift, manifestation shift, and population shift, and demonstrate that our approach can diagnose failures and quantify model robustness without additional data collection, complementing more qualitative tools for explainable AI." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 160, + 522, + 404, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 522, + 404, + 533 + ], + "spans": [ + { + "bbox": [ + 160, + 522, + 404, + 533 + ], + "type": "text", + "content": "Keywords: Image editing " + }, + { + "bbox": [ + 160, + 522, + 404, + 533 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 522, + 404, + 533 + ], + "type": "text", + "content": " diffusion models " + }, + { + "bbox": [ + 160, + 522, + 404, + 533 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 522, + 404, + 533 + ], + "type": "text", + "content": " biomedical imaging" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 555, + 218, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 218, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 218, + 567 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Developing accurate and robust models for biomedical image analysis requires large and diverse datasets that are often difficult to obtain due to ethical, legal, geographical, and financial constraints [41]. This leads to biased training datasets that affect the performance of trained models and generalisation to real-world scenarios [59, 40]. Such data mismatch may arise from genuine differences in upstream data acquisition as well as from the selection criteria for dataset creation, which materialise as various forms of dataset shifts (population, acquisition, annotation, prevalence, manifestation) [7]." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 137, + 280, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 137, + 280, + 179 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 280, + 179 + ], + "type": "text", + "content": "Top: editing out COVID-19 features results in false positives since the classifier relies on acquisition differences, e.g., radiographic markers (white arrow)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 180, + 280, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 180, + 280, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 180, + 280, + 213 + ], + "type": "text", + "content": "Middle: editing out a pneumothorax (PTX) results in false positives since the classifier instead detects chest drains." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 213, + 280, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 213, + 280, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 213, + 280, + 257 + ], + "type": "text", + "content": "Bottom: editing abnormalities into lungs causes a lung segmentation model to mislabel (blue: ground-truth segmentation; red: model prediction)." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 289, + 113, + 482, + 260 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 280, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 280, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 280, + 137 + ], + "type": "text", + "content": "Fig.1: Stress-testing models by simulating dataset shifts via image editing." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 289, + 113, + 482, + 260 + ], + "lines": [ + { + "bbox": [ + 289, + 113, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 289, + 113, + 482, + 260 + ], + "type": "image", + "image_path": "d2a8b5c9a55a2a7c69814c666b6e44508a804669c7e9df2845998298be920b81.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 291, + 482, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 291, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 291, + 482, + 374 + ], + "type": "text", + "content": "Biomedical vision models, when put into real-world use, can be unhelpful or potentially even harmful to patients if they are affected by dataset shifts, leading to missed diagnoses [22, 74, 76, 56]. For example, the COVID-19 pandemic led to hundreds of detection tools being developed, with some put into use in hospitals; yet Roberts et al. [56] found that \"none of the models identified are of potential clinical use due to methodological flaws and/or underlying biases.\" It is therefore crucial to properly assess models for bias, prior to real-world use." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 375, + 482, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 482, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 482, + 447 + ], + "type": "text", + "content": "Recent deep generative models have made remarkable improvements in terms of sample quality, diversity, and steerability [57, 48, 35, 28]. These models have been shown to generalise to out-of-distribution domains [42, 6, 32, 19], opening up avenues for new applications. One such application is generating synthetic data for stress-testing models [51, 42, 72]. This involves creating data that is realistic, yet can represent settings, domains, or populations that do not appear (enough) in the real training/test data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 447, + 481, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 481, + 471 + ], + "type": "text", + "content": "In this work, we investigate how deep generative models can be used for stress-testing biomedical imaging models. We consider three dataset shift scenarios:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 136, + 474, + 480, + 546 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 136, + 474, + 480, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 474, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 136, + 474, + 480, + 498 + ], + "type": "text", + "content": "1. Acquisition shift: classifying COVID-19 cases when the positive and negative cases were acquired at different hospitals (Sec. 5.2)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 136, + 498, + 480, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 498, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 136, + 498, + 480, + 521 + ], + "type": "text", + "content": "2. Manifestation shift: detecting if pneumothorax1 was resolved when chest drains (inserted to treat pneumothorax) are present (Sec. 5.3)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 136, + 522, + 480, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 522, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 136, + 522, + 480, + 546 + ], + "type": "text", + "content": "3. Population shift: segmenting lungs in the presence of abnormalities rarely or never seen in the training dataset (Sec. 5.4)." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "type": "text", + "content": "For each of these scenarios, we simulate dataset shifts, producing stress-test sets which can occur in the real world but do not appear or are underrepresented in the original training/test sets. Following prior work, these test sets are synthesised using generative image editing, which unlike generating images from scratch, only minimally modifies the images, hence, better retains fidelity and diversity [51, 42]. For the above scenarios, we use generative editing to 1. remove only COVID-19 features while keeping visual indicators of the different hospitals; 2. remove only pneumothorax while keeping the chest drain; and 3. add abnormalities that occlude lung structures in the image." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 303, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 303, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 303, + 102 + ], + "type": "text", + "content": "F. Pérez-García, S. Bond-Taylor et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 146, + 653, + 470, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 653, + 470, + 666 + ], + "spans": [ + { + "bbox": [ + 146, + 653, + 470, + 666 + ], + "type": "text", + "content": "1 We provide descriptions of the medical terms used throughout the paper in Appendix A." + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 319 + ], + "type": "text", + "content": "We train a diffusion model [27, 57] on a large collection of chest X-rays from a variety of biomedical imaging datasets (Sec. 5.1), enabling us to add and remove a wide variety of pathologies and support devices when editing. Despite the diversity within these datasets, substantial biases are still present, some of which are learned by the generative model; as a result, correlated features may also be modified. For example, in Scenario 2, removing the pneumothorax might also remove the chest drains as both features typically co-occur in datasets [58], since chest drains are used to treat pneumothorax. Furthermore, when editing only within editing masks, artefacts often appear at the border of the masks. Lastly, artefacts occur when editing images outside of the training dataset domain of the diffusion model used for editing. To overcome these challenges, we propose using multiple masks to break existing correlations. This involves defining which regions must change, and explicitly forcing correlated regions to remain unchanged. In addition, we allow the area outside of the masks to be modified by the diffusion model to ensure image consistency. Since our proposed editing method, which we call RadEdit, leads to only minimal overall changes of chest X-rays, we are able to generate synthetic datasets that can be used to stress-test segmentation models (Scenario 3), which, to the best of our knowledge, we are the first to demonstrate." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 320, + 330, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 320, + 330, + 331 + ], + "spans": [ + { + "bbox": [ + 146, + 320, + 330, + 331 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 332, + 479, + 413 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 332, + 479, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 479, + 354 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 479, + 354 + ], + "type": "text", + "content": "- We introduce a novel editing approach that reduces the presence of artefacts in edited images and simplifies prompt construction compared to prior work [10, 51]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 356, + 479, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 356, + 479, + 377 + ], + "spans": [ + { + "bbox": [ + 138, + 356, + 479, + 377 + ], + "type": "text", + "content": "- Our editing approach allows us to construct synthetic datasets with specific data shifts by performing zero-shot edits on datasets/abnormalities not seen in training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 380, + 479, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 479, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 479, + 413 + ], + "type": "text", + "content": "- We conduct a broad set of experiments using these synthetic datasets to stress-test and expose biases in biomedical classification and, for the first time, segmentation models, introducing a new use case of synthetic data into the medical setting." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 133, + 431, + 221, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 431, + 221, + 443 + ], + "spans": [ + { + "bbox": [ + 133, + 431, + 221, + 443 + ], + "type": "text", + "content": "2 Preliminaries" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 454, + 481, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 454, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 133, + 454, + 481, + 490 + ], + "type": "text", + "content": "In this section, we introduce background context for stress-testing biomedical imaging models: failure modes of biomedical imaging models caused by different dataset shifts; diffusion models as versatile generative models; and diffusion-based image editing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 506, + 214, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 506, + 214, + 517 + ], + "spans": [ + { + "bbox": [ + 133, + 506, + 214, + 517 + ], + "type": "text", + "content": "2.1 Dataset shifts" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 524, + 481, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 524, + 481, + 607 + ], + "spans": [ + { + "bbox": [ + 133, + 524, + 481, + 607 + ], + "type": "text", + "content": "Dataset shift refers to a discrepancy between the training and test data distributions due to external factors [7, 34]. Such shifts are regularly observed in machine learning for biomedical imaging, often due to data scarcity. For example, collected training datasets might consist primarily of healthy patients. However, when the model is used in practice after training, there could be a shift towards unhealthy patients. A taxonomy of different types of dataset shifts in the context of biomedical imaging was developed by Castro et al. [7]. In this paper, we consider three dataset shifts of particular interest." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 617, + 481, + 665 + ], + "type": "text", + "content": "Acquisition shift results from the use of different scanners (manufacturer, hardware, and software) or imaging protocols as often encountered when using data from multiple cohorts. These changes affect factors such as image resolution, contrast, patient positioning, and image markings." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "type": "text", + "content": "RadEdit: stress-testing biomedical vision models via diffusion image editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 164 + ], + "type": "text", + "content": "Manifestation shift results from the way the prediction targets physically manifest in anatomy changes between domains. For example, training datasets could consist of more severe pathological cases than observed in practice, or a pathology may co-occur with different visual features, e.g., support devices." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 172, + 482, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 482, + 233 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 482, + 233 + ], + "type": "text", + "content": "Population shift results from differences in intrinsic characteristics of the populations under study, changing the anatomical appearance distribution. This definition encompasses examples such as age, sex, ethnicity, and comorbidities, but also abnormalities such as pleural effusion and support devices. In contrast to manifestation shift, the shift in anatomical appearance is not affected by prediction targets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 246, + 230, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 230, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 230, + 257 + ], + "type": "text", + "content": "2.2 Diffusion models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 262, + 481, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 481, + 310 + ], + "type": "text", + "content": "Denoising diffusion probabilistic models (DDPMs) [27, 67] are a versatile and effective class of generative models that enable sampling from the data distribution by learning to denoise samples corrupted with Gaussian noise. DDPMs are formed by defining a forward time process that gradually adds noise to data points " + }, + { + "bbox": [ + 132, + 262, + 481, + 310 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 132, + 262, + 481, + 310 + ], + "type": "text", + "content": " through the recursion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 314, + 481, + 329 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 314, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 138, + 314, + 481, + 329 + ], + "type": "interline_equation", + "content": "x _ {t} = \\sqrt {1 - \\beta_ {t}} x _ {t - 1} + \\sqrt {\\beta_ {t}} \\epsilon_ {t}, \\quad t = 1, \\dots , T, \\quad \\text {s . t .} x _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\bar {\\epsilon} _ {t}, \\tag {1}", + "image_path": "ae364c4f3cac5855cbadab7214541477a527e37023aaac43df003042d51ef80c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 332, + 482, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 482, + 393 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 482, + 393 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 332, + 482, + 393 + ], + "type": "inline_equation", + "content": "\\epsilon_{1:T},\\bar{\\epsilon}_{1:T}\\sim \\mathcal{N}(0,I),\\beta_{1:T}" + }, + { + "bbox": [ + 132, + 332, + 482, + 393 + ], + "type": "text", + "content": " is a predefined noise schedule that determines how quickly to corrupt the data and ensures that " + }, + { + "bbox": [ + 132, + 332, + 482, + 393 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 132, + 332, + 482, + 393 + ], + "type": "text", + "content": " contains little to no information about " + }, + { + "bbox": [ + 132, + 332, + 482, + 393 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 132, + 332, + 482, + 393 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 132, + 332, + 482, + 393 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t = \\prod_{s = 1}^t (1 - \\beta_s)" + }, + { + "bbox": [ + 132, + 332, + 482, + 393 + ], + "type": "text", + "content": ". To form a generative model, the process is reversed in time, gradually transforming Gaussian noise into samples from the learned distribution. While the exact reversal is intractable, a variational approximation is defined by [68]:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 236, + 399, + 481, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 399, + 481, + 411 + ], + "spans": [ + { + "bbox": [ + 236, + 399, + 481, + 411 + ], + "type": "interline_equation", + "content": "x _ {t - 1} = \\hat {\\mu} _ {t} \\left(x _ {t}, f _ {\\theta} \\left(x _ {t}, t, c\\right)\\right) + \\sigma_ {t} z _ {t}, \\tag {2}", + "image_path": "c62c8e34601a1e8f2fdb584cb31d10a332b9c91ebfd50129cf2ba384c72bc626.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 194, + 415, + 481, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 415, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 194, + 415, + 481, + 441 + ], + "type": "interline_equation", + "content": "\\hat {\\mu} _ {t} \\left(x _ {t}, \\epsilon_ {t}\\right) = \\sqrt {\\bar {\\alpha} _ {t - 1}} \\frac {x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t}}{\\sqrt {\\bar {\\alpha} _ {t}}} + \\sqrt {1 - \\bar {\\alpha} _ {t - 1} - \\sigma_ {t} ^ {2}} \\epsilon_ {t}, \\tag {3}", + "image_path": "1b93e740b83adf68938fffd06ebf5e511206c2f8e02afeb20a1f06960d355594.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 444, + 482, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 444, + 482, + 505 + ], + "spans": [ + { + "bbox": [ + 132, + 444, + 482, + 505 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 444, + 482, + 505 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 132, + 444, + 482, + 505 + ], + "type": "text", + "content": " is a conditioning signal such as a text description, " + }, + { + "bbox": [ + 132, + 444, + 482, + 505 + ], + "type": "inline_equation", + "content": "f_{\\theta}(x_t,t,c)" + }, + { + "bbox": [ + 132, + 444, + 482, + 505 + ], + "type": "text", + "content": " is a learned approximation of the noise " + }, + { + "bbox": [ + 132, + 444, + 482, + 505 + ], + "type": "inline_equation", + "content": "\\bar{\\epsilon}_t" + }, + { + "bbox": [ + 132, + 444, + 482, + 505 + ], + "type": "text", + "content": " that corrupted the image " + }, + { + "bbox": [ + 132, + 444, + 482, + 505 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 132, + 444, + 482, + 505 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 132, + 444, + 482, + 505 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 132, + 444, + 482, + 505 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 444, + 482, + 505 + ], + "type": "inline_equation", + "content": "z_{1:T} \\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 132, + 444, + 482, + 505 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 132, + 444, + 482, + 505 + ], + "type": "inline_equation", + "content": "\\sigma_{1:T}" + }, + { + "bbox": [ + 132, + 444, + 482, + 505 + ], + "type": "text", + "content": " controls how much noise is introduced. The process is Markovian and known as a DDPM [27] when " + }, + { + "bbox": [ + 132, + 444, + 482, + 505 + ], + "type": "inline_equation", + "content": "\\sigma_t = \\sqrt{(1 - \\bar{\\alpha}_{t - 1}) / (1 - \\bar{\\alpha}_t)}\\sqrt{1 - \\bar{\\alpha}_t / \\bar{\\alpha}_{t - 1}}" + }, + { + "bbox": [ + 132, + 444, + 482, + 505 + ], + "type": "text", + "content": ", while for " + }, + { + "bbox": [ + 132, + 444, + 482, + 505 + ], + "type": "inline_equation", + "content": "\\sigma_t = 0" + }, + { + "bbox": [ + 132, + 444, + 482, + 505 + ], + "type": "text", + "content": " the process is deterministic and is called a denoising diffusion implicit model (DDIM) [68]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 517, + 217, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 517, + 217, + 529 + ], + "spans": [ + { + "bbox": [ + 132, + 517, + 217, + 529 + ], + "type": "text", + "content": "2.3 Image editing" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 482, + 666 + ], + "type": "text", + "content": "Since DDIMs have a one-to-one correspondence with latent vectors " + }, + { + "bbox": [ + 132, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 132, + 533, + 482, + 666 + ], + "type": "text", + "content": ", we can deterministically map data points to latents by running the DDIM generative process in reverse [68], called DDIM inversion. Several approaches [10, 46] have shown that images can be edited by running the reverse diffusion process augmented by the latent vectors and a modified prompt " + }, + { + "bbox": [ + 132, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 132, + 533, + 482, + 666 + ], + "type": "text", + "content": ". However, this method can lead to undesired artefacts in the edited images. For example, structures unrelated to the desired edit may also change shape, size, or location. To address this, Huberman-Spiegelglas et al. [29] propose DDPM inversion. Here, the original forward process defined in Eq. (1) is adapted, replacing the correlated vectors " + }, + { + "bbox": [ + 132, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\bar{\\epsilon}_{1:T}" + }, + { + "bbox": [ + 132, + 533, + 482, + 666 + ], + "type": "text", + "content": " with statistically independent vectors " + }, + { + "bbox": [ + 132, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tilde{\\epsilon}_{1:T}" + }, + { + "bbox": [ + 132, + 533, + 482, + 666 + ], + "type": "text", + "content": " (more details in Appendix B). These noise vectors are then used in the generative process, retaining the structure of the original image better than DDIM inversion." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 303, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 303, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 303, + 102 + ], + "type": "text", + "content": "F. Pérez-García, S. Bond-Taylor et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 221, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 221, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 221, + 127 + ], + "type": "text", + "content": "3 Related work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 263, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 263, + 152 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 263, + 152 + ], + "type": "text", + "content": "3.1 Generative image editing" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 159, + 481, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 159, + 481, + 219 + ], + "spans": [ + { + "bbox": [ + 133, + 159, + 481, + 219 + ], + "type": "text", + "content": "With advances in deep generative modelling, several approaches to image editing have emerged. Many of these early approaches use compressed latent manipulation [12, 52, 64, 71] where fine-grained edits are difficult to achieve and can result in unwanted changes. More recently, the unparalleled flexibility of diffusion models, together with advances in plain text conditioning, have opened up new avenues for editing techniques." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 220, + 481, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 220, + 481, + 327 + ], + "spans": [ + { + "bbox": [ + 133, + 220, + 481, + 327 + ], + "type": "text", + "content": "Here, we describe some notable diffusion editing methods. SDEdit [46] shows that diffusion models trained solely on real images can be used to generate images from sketches by perturbing sketches with noise, then running the reverse diffusion process. Palette [60] is a diffusion model trained for inpainting by filling regions with noise and denoising. Blended diffusion [2, 3] uses masks with CLIP [53] conditioning to guide local edits. Multiple works show that injecting U-Net activations, obtained by encoding the original image into the generation process, better retains image structure [24, 70]. DiffEdit [10] uses text prompts to determine the appropriate region to edit. Mokady et al. [47] improve diffusion inversion quality by optimising the diffusion trajectory." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 327, + 481, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 327, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 133, + 327, + 481, + 374 + ], + "type": "text", + "content": "Crucially, in the works which use masks for editing, a single type of mask is always used to define the region of interest. In this work, we argue that a second type of mask is required to avoid the loss of features caused by spurious correlations. As better editing approaches are developed, this requirement should be kept in mind." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 392, + 309, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 392, + 309, + 405 + ], + "spans": [ + { + "bbox": [ + 133, + 392, + 309, + 405 + ], + "type": "text", + "content": "3.2 Biomedical imaging counterfactuals" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 412, + 481, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 412, + 481, + 531 + ], + "spans": [ + { + "bbox": [ + 133, + 412, + 481, + 531 + ], + "type": "text", + "content": "Generative models have also been applied to biomedical counterfactual generation. Reinhold et al. [55] manipulate causes of multiple sclerosis in brain MRI with deep structural causal models [49]. Sanchez et al. [62] and Fontanella et al. [15] use editing to remove pathologies for abnormality detection. Ktena et al. [39] generate out-of-distribution samples to improve classifier performance. Gu et al. [21] train a diffusion model to model disease progression by conditioning on a prior X-ray and text progression description. Unlike our approach, these methods do not use masks to enforce which regions may or may not be edited, meaning that spurious correlations might affect edits. Additionally, these methods use synthetic data to augment and improve model performance whereas we focus on using synthetic medical data for stress-testing." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 548, + 214, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 548, + 214, + 562 + ], + "spans": [ + { + "bbox": [ + 133, + 548, + 214, + 562 + ], + "type": "text", + "content": "3.3 Stress-testing" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 570, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 570, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 133, + 570, + 481, + 628 + ], + "type": "text", + "content": "Several approaches have used non-deep-generative-model methods to stress-test networks. Hendrycks and Dietterich [23] evaluate classification models' robustness to corruptions such as blurring, Gaussian noise, and JPEG artefacts. Sakaridis et al. [61] stress-test a segmentation model for roads by using an optical model to add synthetic fog to scenes. Koh et al. [38] collate a dataset presenting various distribution shifts." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 630, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 630, + 481, + 665 + ], + "type": "text", + "content": "More recent models have made use of conditional generative models to simulate shifts. Prabhu et al. [51] propose LANCE, which stress-tests classification models by using diffusion-based image editing to modify image subjects via caption editing with" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "type": "text", + "content": "RadEdit: stress-testing biomedical vision models via diffusion image editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 188 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 188 + ], + "type": "text", + "content": "a large language model (LLM); Kattakinda et al. [36] do similar, but instead modify the background. Li et al. [42] use diffusion models with a single subject mask to separately edit backgrounds and subjects. Van Breugel et al. [72] use generative adversarial networks to simulate distribution shifts on tabular data. This line of research is partially related to adversarial attacks [20], where the focus is on minimally modifying images such that they are visually indistinguishable to a human, but the attacked model fails." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 205, + 408, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 408, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 408, + 217 + ], + "type": "text", + "content": "3.4 Limitations of existing editing-based stress-testing methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 225, + 481, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 225, + 481, + 261 + ], + "spans": [ + { + "bbox": [ + 133, + 225, + 481, + 261 + ], + "type": "text", + "content": "Recent advancements in diffusion modelling have drastically improved image editing. However, two prevalent approaches, LANCE [51] and DiffEdit [10], produce artefacts in medical images, making them unsuitable for stress-testing biomedical vision models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 262, + 481, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 262, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 133, + 262, + 481, + 380 + ], + "type": "text", + "content": "LANCE only uses a global prompt (no mask) for image editing. While effective in the natural image domain, it leads to artefacts in the biomedical domain. For example in Sec. 5.4, we add pathologies and support devices to images of healthy lungs to stress-test lung segmentation models; in such cases we must ensure that the position and shape of the lung borders are not altered during editing. However, we find that LANCE changes the position and shape of the lung border thus making edited images unsuitable for stress-testing such models (Sec. 5.5). In addition, we find that LANCE potentially removes support devices when prompted to remove pathologies, a direct effect of the correlations in the diffusion model's training datasets (Sec. 5.1), making LANCE unsuited for testing the robustness of models to manifestation shift." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 381, + 481, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 381, + 481, + 476 + ], + "spans": [ + { + "bbox": [ + 133, + 381, + 481, + 476 + ], + "type": "text", + "content": "DiffEdit addresses these issues by editing only inside an automatically predicted mask " + }, + { + "bbox": [ + 133, + 381, + 481, + 476 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{edit}}" + }, + { + "bbox": [ + 133, + 381, + 481, + 476 + ], + "type": "text", + "content": ". However, these predicted masks often mismatch the manually annotated ground-truth, especially for small and complex abnormalities like pneumothorax1 (see Sec. 5.5). Moreover, spurious correlations learned by the diffusion model can lead to the inclusion of support devices in the automatically predicted masks. Furthermore, even when relying on manually annotated masks, DiffEdit can introduce sharp discrepancies at mask boundaries, leading to unrealistic artefacts, such as when adding consolidation that should partially occlude the lung border (Fig. 10b in the Appendix)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 495, + 192, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 495, + 192, + 506 + ], + "spans": [ + { + "bbox": [ + 133, + 495, + 192, + 506 + ], + "type": "text", + "content": "4 Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 520, + 481, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 520, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 133, + 520, + 481, + 556 + ], + "type": "text", + "content": "Our objective is to create synthetic test data through image editing that simulates specific data shifts, to rigorously evaluate biomedical imaging models. This synthetic data is used to predict model robustness, eliminating need for additional real-world test data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 573, + 290, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 573, + 290, + 586 + ], + "spans": [ + { + "bbox": [ + 133, + 573, + 290, + 586 + ], + "type": "text", + "content": "4.1 Improved editing with RadEdit" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 594, + 481, + 665 + ], + "type": "text", + "content": "To address the issues outlined in Sec. 3.4, we propose RadEdit: by introducing 'keep' and 'edit' masks into the editing process, RadEdit explicitly specifies which areas must remain unchanged (keep) and which should be actively modified based on the conditioning signal (edit). Crucially, these masks need not be mutually exclusive, allowing changes in the unmasked regions to ensure global consistency. Using masks, we assume that spurious correlations are mostly non-overlapping [44]." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 303, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 303, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 303, + 102 + ], + "type": "text", + "content": "F. Pérez-García, S. Bond-Taylor et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 132, + 130, + 481, + 264 + ], + "blocks": [ + { + "bbox": [ + 132, + 115, + 432, + 127 + ], + "lines": [ + { + "bbox": [ + 132, + 115, + 432, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 432, + 127 + ], + "type": "text", + "content": "Algorithm 1 RadEdit (ours) uses multiple masks to decouple spurious correlations" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "lines": [ + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "spans": [ + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "text", + "content": "Require: original image " + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "text", + "content": " , inversion prompt " + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{inv}}" + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "text", + "content": " , editing prompt c, edit mask medit, keep mask " + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{keep}}" + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "text", + "content": " CFG weight w, diffusion model " + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "inline_equation", + "content": "(\\hat{x}_{1:T},z_{1:T})\\leftarrow" + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "text", + "content": " DDPMINVERSION(xo,cinv) " + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "text", + "content": " Encode image. Procedure in Appendix B \n" + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "inline_equation", + "content": "x_{T}\\gets \\hat{x}_{T}" + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "text", + "content": " \nfor " + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "inline_equation", + "content": "t\\gets T" + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "text", + "content": " to 1 do \n" + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{cond},t}\\gets f_{\\theta}(xt,t,c)" + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "text", + "content": " Predict conditional noise \n" + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{uncond},t}\\gets f_{\\theta}(xt,t,c = \\emptyset)" + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "text", + "content": " Predict unconditional noise \n" + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "inline_equation", + "content": "\\epsilon_t\\gets \\epsilon_{\\mathrm{uncond},t} + w(\\epsilon_{\\mathrm{cond},t} - \\epsilon_{\\mathrm{uncond},t})" + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "text", + "content": " Combine noise predictions with CFG \n" + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "inline_equation", + "content": "\\epsilon_t\\gets m_{\\mathrm{edit}}\\odot \\epsilon_t + (1 - m_{\\mathrm{edit}})\\odot \\epsilon_{\\mathrm{uncond},t}" + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "text", + "content": " Use CFG only within medit \n" + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "inline_equation", + "content": "x_{t - 1}\\gets \\hat{\\mu}_t(x_t,\\epsilon_t) + \\sigma_t z_t" + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "text", + "content": " Move to next time step \n" + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "inline_equation", + "content": "x_{t - 1}\\gets m_{\\mathrm{keep}}\\odot \\hat{x}_{t - 1} + (1 - m_{\\mathrm{keep}})\\odot x_{t - 1}" + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "text", + "content": " Undo edits within mkeep \nreturn edited version of " + }, + { + "bbox": [ + 132, + 130, + 481, + 264 + ], + "type": "inline_equation", + "content": "x_0" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 285, + 482, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 417 + ], + "type": "text", + "content": "RadEdit is detailed in Algorithm 1, where a number of key properties make RadEdit more suitable for biomedical image editing than prior editing methods. Firstly, since we aim to edit only within the edit mask " + }, + { + "bbox": [ + 130, + 285, + 482, + 417 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{edit}}" + }, + { + "bbox": [ + 130, + 285, + 482, + 417 + ], + "type": "text", + "content": ", classifier-free guidance (CFG) [26] is used only within this region, with high guidance values (following [29], we use a value of 15) ensuring that pathologies are completely removed without drastically changing the rest of the image. This approach also simplifies choosing a prompt for editing since we do not have to take into account the effect of the prompt on the rest of the image. Secondly, we allow the area outside " + }, + { + "bbox": [ + 130, + 285, + 482, + 417 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{edit}}" + }, + { + "bbox": [ + 130, + 285, + 482, + 417 + ], + "type": "text", + "content": " to be modified via unconditional generation to ensure image consistency. Lastly, the region of the keep mask " + }, + { + "bbox": [ + 130, + 285, + 482, + 417 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{keep}}" + }, + { + "bbox": [ + 130, + 285, + 482, + 417 + ], + "type": "text", + "content": " is reverted to the encoding, ensuring that this region remains the same. Instead of initiating our generating process from pure noise we set " + }, + { + "bbox": [ + 130, + 285, + 482, + 417 + ], + "type": "inline_equation", + "content": "x_{T} = \\hat{x}_{T}" + }, + { + "bbox": [ + 130, + 285, + 482, + 417 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 285, + 482, + 417 + ], + "type": "inline_equation", + "content": "\\hat{x}_T" + }, + { + "bbox": [ + 130, + 285, + 482, + 417 + ], + "type": "text", + "content": " is the last output of the DDPM inversion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 418, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 489 + ], + "type": "text", + "content": "In Fig. 3c, 10c, we show that RadEdit enables artefact-free editing while preserving structures of interest. Because the anatomical layout remains intact after editing, masks still correspond to the same structures, therefore the same masks can be reused to stress-test segmentation models (Sec. 5.4). In practice, we use a latent diffusion model [57], therefore all operations in Algorithm 1 are performed in the latent space of a variational autoencoder (VAE) [57]; this does not limit the generality of the approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 504, + 322, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 504, + 322, + 516 + ], + "spans": [ + { + "bbox": [ + 131, + 504, + 322, + 516 + ], + "type": "text", + "content": "4.2 Using synthetic images to uncover bias" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "content": "Despite advancements in biomedical computer vision, recent studies have shown that bias in training and test data can lead to unrealistically high performance of machine learning models on the test set [59, 11]. In our experiments, we use RadEdit to create high quality synthetic test datasets that realistically capture specific dataset shifts, allowing us to quantify the robustness of models to these dataset shifts. By using masks, we can precisely edit the original training data to represent either acquisition shift, manifestation shift, or population shift [7] (Secs. 5.2 to 5.4). These synthetic test sets are used to stress-test (potentially biased) biomedical vision models by comparing performance to the real (biased) test set; a significant drop in performance indicates that the vision model is not robust to the dataset shift that can occur in clinical settings. This serves as a complementary tool to visual explainable AI tools like Grad-CAM [63] and saliency maps [66, 1], which offer qualitative insight into the robustness of models." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 171, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 171, + 91, + 448, + 102 + ], + "type": "text", + "content": "RadEdit: stress-testing biomedical vision models via diffusion image editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 255, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 255, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 255, + 128 + ], + "type": "text", + "content": "4.3 BioViL-T editing score" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 220 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 220 + ], + "type": "text", + "content": "Since generative models result in samples of varying quality, poor-quality samples can be filtered out using image-text alignment scores, which quantitatively assess how closely related image-text pairs are via a pre-trained model that embeds similar images and text to nearby vectors [4, 54, 53, 14]. For image editing, we instead assess how similar the change in text and image embeddings are after editing: for a real image-text pair " + }, + { + "bbox": [ + 130, + 137, + 482, + 220 + ], + "type": "inline_equation", + "content": "(I_{\\mathrm{real}}, T_{\\mathrm{real}})" + }, + { + "bbox": [ + 130, + 137, + 482, + 220 + ], + "type": "text", + "content": ", edited image-text pair " + }, + { + "bbox": [ + 130, + 137, + 482, + 220 + ], + "type": "inline_equation", + "content": "(I_{\\mathrm{edit}}, T_{\\mathrm{edit}})" + }, + { + "bbox": [ + 130, + 137, + 482, + 220 + ], + "type": "text", + "content": ", image encoder " + }, + { + "bbox": [ + 130, + 137, + 482, + 220 + ], + "type": "inline_equation", + "content": "E_I" + }, + { + "bbox": [ + 130, + 137, + 482, + 220 + ], + "type": "text", + "content": ", and text encoder " + }, + { + "bbox": [ + 130, + 137, + 482, + 220 + ], + "type": "inline_equation", + "content": "E_T" + }, + { + "bbox": [ + 130, + 137, + 482, + 220 + ], + "type": "text", + "content": ", the editing score is defined based on directional similarity [17]:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 164, + 228, + 481, + 257 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 228, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 164, + 228, + 481, + 257 + ], + "type": "interline_equation", + "content": "S _ {\\mathrm {B i o V i L - T}} = \\frac {\\Delta I \\cdot \\Delta T}{\\| \\Delta I \\| \\| \\Delta T \\|}, \\quad \\text {w h e r e} \\quad \\begin{array}{l} \\Delta I = E _ {I} \\left(I _ {\\mathrm {e d i t}}\\right) - E _ {I} \\left(I _ {\\mathrm {r e a l}}\\right), \\text {a n d} \\\\ \\Delta T = E _ {T} \\left(T _ {\\mathrm {e d i t}}\\right) - E _ {T} \\left(T _ {\\mathrm {r e a l}}\\right). \\end{array} \\tag {4}", + "image_path": "7e794edb87fd464d9c1bdd3887986754da10832bbfbe8cfef42a18d5329b908a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 265, + 482, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 265, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 482, + 338 + ], + "type": "text", + "content": "Given the focus on biomedical data, we use the BioViL-T [5] image and text encoders: domain-specific vision-language models trained to analyse chest X-rays and radiology reports, therefore well suited to measure changes in the edited image, such as removed pathologies. Following Prabhu et al. [51], we discard images with " + }, + { + "bbox": [ + 130, + 265, + 482, + 338 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{BioViL - T}} < 0.2" + }, + { + "bbox": [ + 130, + 265, + 482, + 338 + ], + "type": "text", + "content": ". This is not only effective for filtering out poor quality edits but is also able to detect whether the original image " + }, + { + "bbox": [ + 130, + 265, + 482, + 338 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{real}}" + }, + { + "bbox": [ + 130, + 265, + 482, + 338 + ], + "type": "text", + "content": " does not match the original text description " + }, + { + "bbox": [ + 130, + 265, + 482, + 338 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{real}}" + }, + { + "bbox": [ + 130, + 265, + 482, + 338 + ], + "type": "text", + "content": " well." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 356, + 218, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 356, + 218, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 356, + 218, + 369 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 380, + 226, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 226, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 226, + 392 + ], + "type": "text", + "content": "5.1 Diffusion model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 400, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 472 + ], + "type": "text", + "content": "Our editing method is heavily dependent on a latent diffusion model [57] that can generate realistic chest X-rays. We use the VAE [37, 25] of SDXL [50] which can adequately reconstruct chest X-rays [8]. The VAE is frozen, and the denoising U-Net is trained on three datasets downsampled and centre-cropped to " + }, + { + "bbox": [ + 130, + 400, + 482, + 472 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 130, + 400, + 482, + 472 + ], + "type": "text", + "content": " pixels: MIMIC-CXR [33], ChestX-ray8 [75], and CheXpert [30], totalling 487680 training images. This data diversity allows us to perform zero-shot edits on datasets not seen during training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 472, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 556 + ], + "type": "text", + "content": "For MIMIC-CXR, we only include frontal view chest X-rays, and condition the denoising U-Net on the corresponding impression section in the radiology report (a short clinically actionable outline of the main findings). We employ the tokeniser and frozen text encoder from BioViL-T [5]. For ChestX-ray8 and CheXpert, we condition on a list of all abnormalities present in an image as indicated by the labels, e.g., 'Cardiomegaly. Pneumothorax'. If the list of abnormalities is empty, we use the string 'No findings'. More details on the datasets and diffusion model training can be found in Appendix E." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 574, + 228, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 228, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 228, + 586 + ], + "type": "text", + "content": "5.2 Acquisition shift" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Background To show how RadEdit can be used to quantify the robustness of models to acquisition shift, we closely follow the experimental setup of DeGrave et al. [11], who show that deep learning systems built to detect COVID-19 from chest X-rays rely on confounding factors rather than pathology features. This problem arises when COVID-19-positive and -negative images come from disparate sources. In our setup, all COVID-19-positive cases come from the BIMCV dataset [73] (denoted BIMCV+)," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 303, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 303, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 303, + 102 + ], + "type": "text", + "content": "F. Pérez-García, S. Bond-Taylor et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 249, + 118, + 328, + 197 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 242, + 212 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 242, + 212 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 242, + 212 + ], + "type": "text", + "content": "Fig.2: Removing COVID-19 features with LANCE" + }, + { + "bbox": [ + 132, + 114, + 242, + 212 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 132, + 114, + 242, + 212 + ], + "type": "text", + "content": " (b) also changes the laterality markers and reduces contrast. In contrast, RadEdit (c; ours) preserves anatomical structures and laterality markers, and retains the original contrast." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 249, + 118, + 328, + 197 + ], + "lines": [ + { + "bbox": [ + 249, + 118, + 328, + 197 + ], + "spans": [ + { + "bbox": [ + 249, + 118, + 328, + 197 + ], + "type": "image", + "image_path": "9111de6a974fbc3b2c5917a8c2007f88b2d20157edd2638845627203f073c19f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 200, + 314, + 209 + ], + "lines": [ + { + "bbox": [ + 260, + 200, + 314, + 209 + ], + "spans": [ + { + "bbox": [ + 260, + 200, + 314, + 209 + ], + "type": "text", + "content": "(a) Original Image" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 328, + 118, + 402, + 197 + ], + "blocks": [ + { + "bbox": [ + 328, + 118, + 402, + 197 + ], + "lines": [ + { + "bbox": [ + 328, + 118, + 402, + 197 + ], + "spans": [ + { + "bbox": [ + 328, + 118, + 402, + 197 + ], + "type": "image", + "image_path": "fc36787c107450cc0af8b9e69ae3edd6634b3472011cf568d7520bb5ec21c319.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 200, + 388, + 209 + ], + "lines": [ + { + "bbox": [ + 339, + 200, + 388, + 209 + ], + "spans": [ + { + "bbox": [ + 339, + 200, + 388, + 209 + ], + "type": "text", + "content": "(b) LANCE [51]" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 402, + 118, + 480, + 197 + ], + "blocks": [ + { + "bbox": [ + 402, + 118, + 480, + 197 + ], + "lines": [ + { + "bbox": [ + 402, + 118, + 480, + 197 + ], + "spans": [ + { + "bbox": [ + 402, + 118, + 480, + 197 + ], + "type": "image", + "image_path": "354372db538e0480126b35eabce74802facc1997cc588ad2fe8ec5b211b46288.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 200, + 466, + 209 + ], + "lines": [ + { + "bbox": [ + 414, + 200, + 466, + 209 + ], + "spans": [ + { + "bbox": [ + 414, + 200, + 466, + 209 + ], + "type": "text", + "content": "(c) RadEdit (ours)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 236, + 479, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 236, + 479, + 274 + ], + "spans": [ + { + "bbox": [ + 131, + 236, + 479, + 274 + ], + "type": "text", + "content": "and all COVID-19-negative cases from MIMIC-CXR [33] (see Fig. 11). A classifier trained on these datasets will rely on spurious features indicative of the data's origin, e.g., laterality markers or the field of view, instead of features caused by the pathology." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 285, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 285, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 131, + 285, + 482, + 380 + ], + "type": "text", + "content": "Setup A synthetic test set is created by applying RadEdit to remove COVID-19 features1 from BIMCV+ images using the prompt 'No acute cardiopulmonary process'3 (Fig. 2); the included bounding boxes of COVID-19 features are used as the edit mask " + }, + { + "bbox": [ + 131, + 285, + 482, + 380 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{edit}}" + }, + { + "bbox": [ + 131, + 285, + 482, + 380 + ], + "type": "text", + "content": ". Since this is the only mask available, we set the keep mask as " + }, + { + "bbox": [ + 131, + 285, + 482, + 380 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{keep}} = 1 - m_{\\mathrm{edit}}" + }, + { + "bbox": [ + 131, + 285, + 482, + 380 + ], + "type": "text", + "content": ". After filtering using the BioViL-T editing (Sec. 4.3), this results in a synthetic dataset of 2774 COVID-19-negative images containing the same spurious features as BIMCV+. Both the weak and strong predictor models are classifiers trained using a ResNet50 architecture (see Appendix F for more implementation details)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 392, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 392, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 131, + 392, + 482, + 475 + ], + "type": "text", + "content": "Findings Tab. 1, shows the performance of a COVID-19 classifier (weak predictor) trained on BIMCV+ and MIMIC-CXR. In accordance with DeGrave et al. [11], we find that the weak predictor performs exceptionally well on the real test set (i.e. test splits of both datasets) since the model learned to distinguish the two data sources instead of learning visual features related to COVID-19. However, in the second row of Tab. 1, we see a drop of " + }, + { + "bbox": [ + 131, + 392, + 482, + 475 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 131, + 392, + 482, + 475 + ], + "type": "text", + "content": " in accuracy meaning that the model fails to classify the synthetic images as COVID-19-negative. The weak predictor is not robust to a shift in acquisition." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 476, + 482, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 476, + 482, + 523 + ], + "spans": [ + { + "bbox": [ + 131, + 476, + 482, + 523 + ], + "type": "text", + "content": "To show that the decreased performance of the weak predictor is not caused by artefacts in the edited images, we train a more robust COVID-19 classifier (strong predictor), using the BIMCV+ and BIMCV- datasets, as in [11], where the BIMCV- dataset consists of only COVID-19-negative cases from BIMCV, and test on the same two test" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 363, + 555, + 480, + 627 + ], + "blocks": [ + { + "bbox": [ + 132, + 544, + 357, + 632 + ], + "lines": [ + { + "bbox": [ + 132, + 544, + 357, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 357, + 632 + ], + "type": "text", + "content": "Table 1: Quantifying robustness of COVID-19 detectors to acquisition shift. We train a weak predictor on the 'Biased' dataset—a combination of BIMCV+ [73] and MIMIC-CXR [33]; and a strong predictor on an unbiased dataset—a combination of BIMCV+ and BIMCV; the 'Synthetic' test set consists of 2774 COVID-19-negative images with the same spurious features as the BIMCV+ datasets, e.g. laterality markers. We report mean accuracy and standard deviation across 5 runs." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 363, + 555, + 480, + 627 + ], + "lines": [ + { + "bbox": [ + 363, + 555, + 480, + 627 + ], + "spans": [ + { + "bbox": [ + 363, + 555, + 480, + 627 + ], + "type": "table", + "html": "
PredictorTest dataAccuracy
WeakBiased99.1 ± 0.2
WeakSynthetic5.5 ± 2.1
StrongBiased74.4 ± 3.0
StrongSynthetic76.0 ± 7.7
", + "image_path": "a1be34d50db1492e81c3452b6ecfd8e322d799b6a444cbd7f0b0f481c12a6f00.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 171, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 171, + 91, + 448, + 102 + ], + "type": "text", + "content": "RadEdit: stress-testing biomedical vision models via diffusion image editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 146, + 642, + 355, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 642, + 355, + 653 + ], + "spans": [ + { + "bbox": [ + 146, + 642, + 355, + 653 + ], + "type": "text", + "content": "2For LANCE, we perform the text perturbation manually." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 147, + 653, + 398, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 653, + 398, + 665 + ], + "spans": [ + { + "bbox": [ + 147, + 653, + 398, + 665 + ], + "type": "text", + "content": "This is a common radiological description of a 'normal' chest X-ray." + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 254, + 115, + 328, + 192 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 250, + 201 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 250, + 201 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 250, + 201 + ], + "type": "text", + "content": "Fig.3: Removing pneumothorax (red) with LANCE" + }, + { + "bbox": [ + 132, + 114, + 250, + 201 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 132, + 114, + 250, + 201 + ], + "type": "text", + "content": " (b) also removes the spuriously correlated chest drain (blue) and reduces contrast. In contrast, RadEdit (c; ours) preserves the chest drain and better preserves anatomical structures." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 254, + 115, + 328, + 192 + ], + "lines": [ + { + "bbox": [ + 254, + 115, + 328, + 192 + ], + "spans": [ + { + "bbox": [ + 254, + 115, + 328, + 192 + ], + "type": "image", + "image_path": "d8b08b200a2c06a0e26d7200596a7eaa7c8afb6de45548faa4a8189e7973fe82.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 263, + 195, + 317, + 205 + ], + "lines": [ + { + "bbox": [ + 263, + 195, + 317, + 205 + ], + "spans": [ + { + "bbox": [ + 263, + 195, + 317, + 205 + ], + "type": "text", + "content": "(a) Original Image" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 328, + 116, + 402, + 192 + ], + "blocks": [ + { + "bbox": [ + 328, + 116, + 402, + 192 + ], + "lines": [ + { + "bbox": [ + 328, + 116, + 402, + 192 + ], + "spans": [ + { + "bbox": [ + 328, + 116, + 402, + 192 + ], + "type": "image", + "image_path": "fc158dd38debd7237738a6c816ded24d5d19c7a6cf421d03d19ba7f44547e67f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 195, + 389, + 205 + ], + "lines": [ + { + "bbox": [ + 340, + 195, + 389, + 205 + ], + "spans": [ + { + "bbox": [ + 340, + 195, + 389, + 205 + ], + "type": "text", + "content": "(b) LANCE [51]" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 402, + 116, + 480, + 192 + ], + "blocks": [ + { + "bbox": [ + 402, + 116, + 480, + 192 + ], + "lines": [ + { + "bbox": [ + 402, + 116, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 402, + 116, + 480, + 192 + ], + "type": "image", + "image_path": "bb6ce92710ee1361641cde70c894f33ba86a2916f4f05e6cf11783410a8f5ebf.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 195, + 466, + 205 + ], + "lines": [ + { + "bbox": [ + 413, + 195, + 466, + 205 + ], + "spans": [ + { + "bbox": [ + 413, + 195, + 466, + 205 + ], + "type": "text", + "content": "(c) RadEdit (ours)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 230, + 479, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 230, + 479, + 289 + ], + "spans": [ + { + "bbox": [ + 131, + 230, + 479, + 289 + ], + "type": "text", + "content": "datasets. Comparing rows one and three of Tab. 1, we find that the strong predictor performs worse on the test set containing samples from BIMCV+ and MIMIC-CXR than the weak predictor (row one). This is expected as the strong predictor relies on actual pathology features. Lastly, rows three and four of Tab. 1 show that the strong predictor performs similarly on the real and synthetic test sets, attesting to the quality of our edits." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 308, + 238, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 238, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 238, + 319 + ], + "type": "text", + "content": "5.3 Manifestation shift" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 328, + 480, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 328, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 131, + 328, + 480, + 437 + ], + "type": "text", + "content": "Background In this section, we show how RadEdit can be used to quantify the robustness of biomedical vision models to manifestation shift. We closely follow the experimental setup of Rueckel et al. [59], who demonstrate that pneumothorax classification models are strongly biased by the presence of chest drains: while the average performance of pneumothorax classifiers is high, performance on the subset of images with a chest drain but no pneumothorax is significantly lower. This is due to chest drains being a common treatment for pneumothorax, resulting in the majority of images in datasets like CANDID-PTX [13] containing a chest drain only if there is a pneumothorax. As a result, only " + }, + { + "bbox": [ + 131, + 328, + 480, + 437 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 131, + 328, + 480, + 437 + ], + "type": "text", + "content": " of images in CANDID-PTX contain a chest drain but no pneumothorax." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 449, + 480, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 449, + 480, + 570 + ], + "spans": [ + { + "bbox": [ + 131, + 449, + 480, + 570 + ], + "type": "text", + "content": "Setup We use RadEdit to create a synthetic dataset containing images with chest drains but no pneumothorax, by editing out the pneumothorax from CANDID-PTX images using the prompt 'No acute cardiopulmonary process' (Fig. 3). The edit mask " + }, + { + "bbox": [ + 131, + 449, + 480, + 570 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{edit}}" + }, + { + "bbox": [ + 131, + 449, + 480, + 570 + ], + "type": "text", + "content": " is set as a mask of the pneumothorax, and the keep mask " + }, + { + "bbox": [ + 131, + 449, + 480, + 570 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{keep}}" + }, + { + "bbox": [ + 131, + 449, + 480, + 570 + ], + "type": "text", + "content": " set as the chest drain mask. This ensures that the chest drain is not removed, while preventing border artefacts. After filtering using the BioViL-T editing score (Sec. 4.3), 628 images are left; in contrast, the real test set contains only 16 of cases with drains but no pneumothorax. The weak predictor is a ResNet50 trained to classify pneumothorax. Following Rueckel et al. [59], the strong predictor is an EfficientNet U-Net [69] trained on SIIM-ACR [77] to both segment pneumothorax and classify pneumothorax (more details in Appendix G)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 582, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 582, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 582, + 479, + 665 + ], + "type": "text", + "content": "Findings In accordance with [59], we show in Tab. 2 that a pneumothorax classifier (weak predictor) trained on CANDID-PTX performs exceptionally well on the test split of CANDID-PTX, since very few images contain a chest drain and no pneumothorax. However, in row two of Tab. 2, we show a drastic drop in performance on the synthetic test set, i.e., the weak predictor is not robust to manifestation shift. To show that the drop in performance on the synthetic dataset does not come from editing artefacts, we evaluate the more robust segmentation model (strong predictor) on the same test datasets" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 302, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 302, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 302, + 102 + ], + "type": "text", + "content": "F. Pérez-García, S. Bond-Taylor et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 364, + 138, + 483, + 213 + ], + "blocks": [ + { + "bbox": [ + 131, + 125, + 354, + 234 + ], + "lines": [ + { + "bbox": [ + 131, + 125, + 354, + 234 + ], + "spans": [ + { + "bbox": [ + 131, + 125, + 354, + 234 + ], + "type": "text", + "content": "Table 2: Quantifying robustness of pneumothorax detectors to manifestation shift. The weak predictor is trained on the biased CANDID-PTX [13] dataset to classify pneumothorax; the strong predictor is trained on SIIM-ACR [77] to classify and segment the pneumothorax. Real 'Biased' test data comes from CANDID-PTX which exhibits strong confounding between the pneumothorax and chest tubes; 'Synthetic' test data is 629 solely edited images containing chest drains but no pneumothorax. We report mean accuracy and standard deviation across 5 runs." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 364, + 138, + 483, + 213 + ], + "lines": [ + { + "bbox": [ + 364, + 138, + 483, + 213 + ], + "spans": [ + { + "bbox": [ + 364, + 138, + 483, + 213 + ], + "type": "table", + "html": "
PredictorTest dataAccuracy
WeakBiased93.3 ± 0.6
WeakSynthetic17.9 ± 3.7
StrongBiased93.7 ± 1.3
StrongSynthetic81.7 ± 7.1
", + "image_path": "ea7c63fdedf15994d67632ed6840aa312c2aa8288c0ff54f200e75f16485ec7e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 257, + 482, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 257, + 482, + 318 + ], + "spans": [ + { + "bbox": [ + 131, + 257, + 482, + 318 + ], + "type": "text", + "content": " Rows three and four of Tab. 2), we find that the strong predictor performs on par with the weak predictor in row one; however, the strong predictor closes the majority of the gap between the real test set and the synthetic one, attesting to the quality of our edits. In agreement with Rueckel et al. [59], there is still a performance gap, indicating that the strong predictor still suffers from mild manifestation shift." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 334, + 227, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 334, + 227, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 334, + 227, + 346 + ], + "type": "text", + "content": "5.4 Population shift" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 353, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 353, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 131, + 353, + 482, + 437 + ], + "type": "text", + "content": "Background In this section, we show how RadEdit can be used to quantify the robustness of lung segmentation models to population shifts. Manually segmenting X-ray images is labour intensive and requires high expertise, leading to small datasets often limited to single pathologies or healthy patients [65, 31], e.g., MIMIC-Seg [9]. These models are thus sensitive to occlusions such as medical devices or pathologies, which typically appear as white regions on X-rays [43]. Evaluating model robustness requires further image collection for each occlusion type, which is time-consuming and costly." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 447, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 447, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 131, + 447, + 482, + 472 + ], + "type": "text", + "content": "Setup Here, abnormalities are added to the lung region in healthy X-rays from MIMIC-Seg (Fig. 4). The edit mask " + }, + { + "bbox": [ + 131, + 447, + 482, + 472 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{edit}}" + }, + { + "bbox": [ + 131, + 447, + 482, + 472 + ], + "type": "text", + "content": " is set as a mask of the lung(s). When editing a single" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 285, + 492, + 482, + 660 + ], + "blocks": [ + { + "bbox": [ + 131, + 492, + 274, + 657 + ], + "lines": [ + { + "bbox": [ + 131, + 492, + 274, + 657 + ], + "spans": [ + { + "bbox": [ + 131, + 492, + 274, + 657 + ], + "type": "text", + "content": "Fig.4: Adding pulmonary edema (top), pacemakers (middle), and consolidation (bottom) with RadEdit. The 'strong predictor' (d), a segmentation model trained on CheXmask [16] (a large dataset containing various abnormalities) is more robust to these abnormalities than the 'weak predictor' (c), a segmentation model trained on MIMIC-Seg [9] (a small set of mostly healthy patients): the weak predictor traces around the pacemaker and poorly annotates the consolidated lung. Blue: ground-truth annotation; red: predicted segmentation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 285, + 492, + 482, + 660 + ], + "lines": [ + { + "bbox": [ + 285, + 492, + 482, + 660 + ], + "spans": [ + { + "bbox": [ + 285, + 492, + 482, + 660 + ], + "type": "image", + "image_path": "d1509fd8ad578531c7ac80d7555bd446cc49b7e0b76c72db5475b17b150e4d51.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 171, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 171, + 91, + 448, + 102 + ], + "type": "text", + "content": "RadEdit: stress-testing biomedical vision models via diffusion image editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 154, + 190, + 459, + 281 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": "Table 3: Quantifying robustness of lung segmentation models to population shift. The 'weak predictor' is trained on MIMIC-Seg (a small set of predominantly healthy patients); the 'strong predictor' is trained on CheXmask (a large mixed set of patients with various abnormalities). Synthetic test data is created by using RadEdit to add edema, pacemakers, and consolidation. We report the change " + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "(\\Delta)" + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": " in Dice score and AHD with respect to the segmentation models evaluated on the ground-truth test set." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 154, + 190, + 459, + 281 + ], + "lines": [ + { + "bbox": [ + 154, + 190, + 459, + 281 + ], + "spans": [ + { + "bbox": [ + 154, + 190, + 459, + 281 + ], + "type": "table", + "html": "
Test dataWeak PredictorStrong Predictor
Dice↑Δ↓AHD↓Δ↓Dice↑Δ↓AHD↓Δ↓
Real data97.4—6.1—95.5—11.6—
Healthy → edema93.83.621.815.793.91.622.811.2
Healthy → pacemaker85.012.449.843.787.38.229.517.9
Healthy → consolidation85.911.544.138.188.17.429.417.8
", + "image_path": "7dde52658bee57b516c34c59ca481e039836d98bfda9979ce6734f3668ef1ba4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 303, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 422 + ], + "type": "text", + "content": "lung, the keep mask " + }, + { + "bbox": [ + 130, + 303, + 482, + 422 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{keep}}" + }, + { + "bbox": [ + 130, + 303, + 482, + 422 + ], + "type": "text", + "content": " corresponds to the lung which must not change, while when editing both lungs we set " + }, + { + "bbox": [ + 130, + 303, + 482, + 422 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{keep}} = 0" + }, + { + "bbox": [ + 130, + 303, + 482, + 422 + ], + "type": "text", + "content": " to allow opacity adjustments, or for elements to be added outside of the lungs. Stress-test sets are generated for three abnormalities: pulmonary edema, pacemakers, and consolidation1. Prompts are phrased to match similar impressions in the training data (see Appendix H). Both predictors are EfficientNet U-Net [69] models; the weak predictor is trained on MIMIC-Seg, while the strong predictor is trained on CheXmask [16], a larger dataset with various lung abnormalities (more details in Appendix H). We evaluate segmentation quality using Dice similarity coefficient, which is the harmonic mean of the precision and recall, and 95th percentile average Hausdorff distance (AHD), a measure of the distance between two sets [45]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 434, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 567 + ], + "type": "text", + "content": "Findings Tab. 3 shows that the weak predictor model performs well on the real biased test data, since it is mostly composed of healthy subjects. However, testing on the synthetic lung abnormality datasets (rows two to four), causes performance to drop substantially, i.e. the weak predictor is not robust to population shift. To show that this drop in performance does not come from editing artefacts, we evaluate the strong predictor on the synthetic test sets and see considerably smaller changes in performance. This can be seen in Fig. 4: for pulmonary edema, both models can accurately segment, despite the abnormality; for pacemakers, the weak predictor incorrectly segments around the pacemakers, while the strong predictor more accurately segments the lungs; and for consolidation, both models are less able to segment the lungs accurately, however, the strong predictor gets closer to the ground-truth. See Appendix H for more examples." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 585, + 389, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 585, + 389, + 597 + ], + "spans": [ + { + "bbox": [ + 131, + 585, + 389, + 597 + ], + "type": "text", + "content": "5.5 Quantifying the limitations of existing editing methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "LANCE As seen in the second row of Tab. 3, adding edema leads only to a small drop in performance of the strong predictor. We hypothesise that further drops in performances stem from a mismatch of the original mask and the edited images. We therefore use this setup to quantify how well LANCE and RadEdit preserve the shape and position of the lung borders. Additionally, we study the difference between results using" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 302, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 302, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 302, + 102 + ], + "type": "text", + "content": "F. Pérez-García, S. Bond-Taylor et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 200, + 185 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 200, + 185 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 200, + 185 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 200, + 185 + ], + "type": "image", + "image_path": "375fc1d5fb456937dfea497ad62d2705686a2b443a6a0a34aa4834acdb4f4e09.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 188, + 171, + 196 + ], + "lines": [ + { + "bbox": [ + 162, + 188, + 171, + 196 + ], + "spans": [ + { + "bbox": [ + 162, + 188, + 171, + 196 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 202, + 115, + 272, + 184 + ], + "blocks": [ + { + "bbox": [ + 202, + 115, + 272, + 184 + ], + "lines": [ + { + "bbox": [ + 202, + 115, + 272, + 184 + ], + "spans": [ + { + "bbox": [ + 202, + 115, + 272, + 184 + ], + "type": "image", + "image_path": "d4bd5f26608b23783c54424f27891562ec463a3ab4e5603f08ea6e297c24f377.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 188, + 241, + 196 + ], + "lines": [ + { + "bbox": [ + 231, + 188, + 241, + 196 + ], + "spans": [ + { + "bbox": [ + 231, + 188, + 241, + 196 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 272, + 115, + 341, + 184 + ], + "blocks": [ + { + "bbox": [ + 272, + 115, + 341, + 184 + ], + "lines": [ + { + "bbox": [ + 272, + 115, + 341, + 184 + ], + "spans": [ + { + "bbox": [ + 272, + 115, + 341, + 184 + ], + "type": "image", + "image_path": "9f465dfb4e44cb0f02d1cd45eb24d43345bf71b88185e79f866293b897d8461f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 301, + 188, + 310, + 196 + ], + "lines": [ + { + "bbox": [ + 301, + 188, + 310, + 196 + ], + "spans": [ + { + "bbox": [ + 301, + 188, + 310, + 196 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 202, + 332, + 280 + ], + "lines": [ + { + "bbox": [ + 132, + 202, + 332, + 280 + ], + "spans": [ + { + "bbox": [ + 132, + 202, + 332, + 280 + ], + "type": "text", + "content": "Fig. 5: Comparison of LANCE" + }, + { + "bbox": [ + 132, + 202, + 332, + 280 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 132, + 202, + 332, + 280 + ], + "type": "text", + "content": " and RadEdit. We measure how well the strong predictor from Tab. 3's outputs matches the ground-truth lung masks (blue) for four synthetic datasets created by adding edema using LANCE and RadEdit with DDIM or DDPM inversion. High Dice / low AHD indicates that the editing method well preserves the lung border's location and shape." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 341, + 115, + 409, + 184 + ], + "blocks": [ + { + "bbox": [ + 341, + 115, + 409, + 184 + ], + "lines": [ + { + "bbox": [ + 341, + 115, + 409, + 184 + ], + "spans": [ + { + "bbox": [ + 341, + 115, + 409, + 184 + ], + "type": "image", + "image_path": "1b3c5ac8ffcec0dc13caeb3a545a3d2c7b44d250909b74a47e72414330f29ad5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 188, + 380, + 196 + ], + "lines": [ + { + "bbox": [ + 369, + 188, + 380, + 196 + ], + "spans": [ + { + "bbox": [ + 369, + 188, + 380, + 196 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 410, + 115, + 481, + 185 + ], + "blocks": [ + { + "bbox": [ + 410, + 115, + 481, + 185 + ], + "lines": [ + { + "bbox": [ + 410, + 115, + 481, + 185 + ], + "spans": [ + { + "bbox": [ + 410, + 115, + 481, + 185 + ], + "type": "image", + "image_path": "af6befdacdf2b5fb7813977631443e3169b0f3008a4cd3cd061e61b24aa5cff7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 439, + 188, + 448, + 196 + ], + "lines": [ + { + "bbox": [ + 439, + 188, + 448, + 196 + ], + "spans": [ + { + "bbox": [ + 439, + 188, + 448, + 196 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 337, + 199, + 475, + 278 + ], + "blocks": [ + { + "bbox": [ + 337, + 199, + 475, + 278 + ], + "lines": [ + { + "bbox": [ + 337, + 199, + 475, + 278 + ], + "spans": [ + { + "bbox": [ + 337, + 199, + 475, + 278 + ], + "type": "table", + "html": "
Editing MethodDice↑AHD↓
(a) Original data95.511.6
(b) LANCE w/ DDIM78.965.1
(c) LANCE w/ DDPM80.169.5
(d) RadEdit w/ DDIM86.239.8
(e) RadEdit w/ DDPM93.922.8
", + "image_path": "4d5a72cfccddfa19ddfe3a6600b2b90edaa7b9a21ed2602d9341520e8093e60b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 296, + 482, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 296, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 131, + 296, + 482, + 357 + ], + "type": "text", + "content": "DDIM or DDPM inversion. For all four methods in Fig. 5, we use the same setup as in Sec. 5.4: we first edit the original image with the prompt 'Moderate pulmonary edema. The heart size is normal', and then compare the outputs of the strong predictor with the original ground-truth lung masks. Here, we find that using masks and DDPM inversion is necessary for RadEdit to preserve the shape and position of the lung border." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 367, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 367, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 131, + 367, + 482, + 511 + ], + "type": "text", + "content": "DiffEdit We quantify how well DiffEdit's automatically predicted masks match the manual ground-truth using the same setup as in Sec. 5.3: we take an image containing pneumothorax and a chest drain, and try to remove only the pneumothorax. The editing prompt is created by keeping only the parts of the impressions related to pneumothorax and chest drains, and replacing the description of the pneumothorax with 'No pneumothorax'. DiffEdit should therefore predict a mask containing only the pneumothorax. We perform a grid search on the MIMIC-Seg [9] validation set over DiffEdit's hyperparameters (noise strength and binarising threshold) to optimise pneumothorax segmentation metrics, then evaluate on the training set. In Fig. 6 we see that DiffEdit's predicted masks obtain poor quantitative metrics where parts of the pneumothorax are often missing, and the spuriously correlated chest drain is often included in the predicted mask. As a result, DiffEdit's predicted masks are unsuitable for stress-testing." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 133, + 533, + 201, + 604 + ], + "blocks": [ + { + "bbox": [ + 133, + 533, + 201, + 604 + ], + "lines": [ + { + "bbox": [ + 133, + 533, + 201, + 604 + ], + "spans": [ + { + "bbox": [ + 133, + 533, + 201, + 604 + ], + "type": "image", + "image_path": "7f9e7b069302971b2be8f0e75a6ff32d5734c9ba94f8c8563f98d4282702d961.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 605, + 349, + 623 + ], + "lines": [ + { + "bbox": [ + 132, + 605, + 349, + 623 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 349, + 623 + ], + "type": "text", + "content": "(a) Examples of pneumothorax masks predicted using DiffEdit [10]. Blue: ground-truth annotation; red: predicted editing mask." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 209, + 533, + 277, + 604 + ], + "blocks": [ + { + "bbox": [ + 209, + 533, + 277, + 604 + ], + "lines": [ + { + "bbox": [ + 209, + 533, + 277, + 604 + ], + "spans": [ + { + "bbox": [ + 209, + 533, + 277, + 604 + ], + "type": "image", + "image_path": "9a69c7837b7f1cfa312e9da9f6d9a67e5d85557e42ec66badfe8ccd74cad0ccb.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 280, + 533, + 349, + 604 + ], + "blocks": [ + { + "bbox": [ + 280, + 533, + 349, + 604 + ], + "lines": [ + { + "bbox": [ + 280, + 533, + 349, + 604 + ], + "spans": [ + { + "bbox": [ + 280, + 533, + 349, + 604 + ], + "type": "image", + "image_path": "b3f352f725da320efcb78cb78f5944d6971e44d5b73a91c410a4c774baa79c52.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 359, + 536, + 476, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 536, + 476, + 546 + ], + "spans": [ + { + "bbox": [ + 359, + 536, + 476, + 546 + ], + "type": "text", + "content": "Hyperparameters Dice AHD" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 359, + 552, + 476, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 552, + 476, + 561 + ], + "spans": [ + { + "bbox": [ + 359, + 552, + 476, + 561 + ], + "type": "text", + "content": "Tuned per image 33.8 97.7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 359, + 563, + 475, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 563, + 475, + 572 + ], + "spans": [ + { + "bbox": [ + 359, + 563, + 475, + 572 + ], + "type": "text", + "content": "Tuned on validation 18.4 256.8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 353, + 578, + 481, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 578, + 481, + 619 + ], + "spans": [ + { + "bbox": [ + 353, + 578, + 481, + 619 + ], + "type": "text", + "content": "(b) Segmentation metrics for the pneumothorax mask predicted by DiffEdit [10], for hyperparameters tuned on the validation set (bottom) and tuned per image (top; which requires ground-truth masks)." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 131, + 628, + 482, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 628, + 482, + 662 + ], + "spans": [ + { + "bbox": [ + 131, + 628, + 482, + 662 + ], + "type": "text", + "content": "Fig. 6: Evaluating pneumothorax masks predicted using DiffEdit [10]. (a) Predicted masks (red) are noisy, with chest drains often incorrectly segmented as well as or instead of the pneumothorax (blue); (b) this is demonstrated quantitatively with low Dice score and high AHD." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 171, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 171, + 91, + 448, + 102 + ], + "type": "text", + "content": "RadEdit: stress-testing biomedical vision models via diffusion image editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 298, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 298, + 128 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 298, + 128 + ], + "type": "text", + "content": "6 Limitations and future work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 137, + 481, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 137, + 481, + 256 + ], + "spans": [ + { + "bbox": [ + 133, + 137, + 481, + 256 + ], + "type": "text", + "content": "Despite the encouraging results presented in the paper, RadEdit is not without limitations and more work is needed to extend it to more applications. Currently, training datasets and models must be manually analysed to predict potential failure cases, simulate these failures to test the hypothesis, and finally quantitatively evaluate the model; future work could automate such failure mode discovery. Another limitation is that current editing techniques do not enable all types of stress-testing; for example, with current approaches, we are unable to test segmentation models' behaviour to cardiomegaly (enlarged heart) since this would require segmentation maps to be adjusted after editing. However, this could potentially be enabled by enlarging heart segmentations to simulate cardiomegaly and adjusting the ground-truth lung segmentation accordingly." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 257, + 481, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 257, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 133, + 257, + 481, + 352 + ], + "type": "text", + "content": "When using generative editing, it is not possible to guarantee that unwanted changes will not occur. With RadEdit, we minimise this by forcing spuriously correlated regions to remain the same, only using classifier-free guidance within the editing mask, and filtering via image-text alignment. Future work to better maintain structure when editing will help with this issue, but masks will still be necessary to bypass spurious correlations. Furthermore, due to potential overlap of 3D structures, editing 2D X-rays can be limited. In such cases, editing single structures using RadEdit may not be successful. To address this, 3D CT images could be edited, then projected to synthetic X-rays [18]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 352, + 481, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 352, + 481, + 411 + ], + "spans": [ + { + "bbox": [ + 133, + 352, + 481, + 411 + ], + "type": "text", + "content": "When producing simulated stress test sets, several factors affect edit quality including classifier-free guidance weight, number of inference steps, and time step to encode to. Additionally, the text encoder must well understand specified pathologies to provide informative features to condition the generative model on; similarly, the diffusion model must be able to capture fine details and well cover the data distribution." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 411, + 481, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 411, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 133, + 411, + 481, + 483 + ], + "type": "text", + "content": "Finally, more research is required to develop better approaches for quantifying edit quality for downstream tasks. In particular, observing a change in downstream performance is not necessarily indicative of real-world performance as edit quality may be poor. While the introduced BioViL-T editing score can be used to quantify edit quality, this introduces reliance on a potentially biased model. Additionally, the BioViL-T editing score is not suited to detect the artefacts introduced by LANCE and DiffEdit." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 498, + 209, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 498, + 209, + 512 + ], + "spans": [ + { + "bbox": [ + 133, + 498, + 209, + 512 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 521, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 521, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 521, + 481, + 665 + ], + "type": "text", + "content": "In this study, we illustrate the efficacy of generative image editing as a robust tool for stress-testing biomedical vision models. Our focus is on assessing their robustness against three types of dataset shifts commonly encountered in biomedical imaging: acquisition shift, manifestation shift, and population shift. We highlight that one of the significant challenges in biomedical image editing is the correlations learned by the generative model, which can result in artefacts during the editing process. To mitigate these artefacts, RadEdit relies on various types of masks to restrict the effects of the editing to certain areas while ensuring the consistency of the edited images. This approach enables us to generate high fidelity synthetic test sets that exhibit common dataset shifts. These synthetic test sets are used to identify and quantify the failure modes of biomedical classification and segmentation models. This provides a valuable supplement to explainable AI approaches such as Grad-CAM [63] and saliency maps [66, 1]." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 302, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 302, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 302, + 102 + ], + "type": "text", + "content": "F. Pérez-García, S. Bond-Taylor et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 265, + 113, + 348, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 113, + 348, + 129 + ], + "spans": [ + { + "bbox": [ + 265, + 113, + 348, + 129 + ], + "type": "text", + "content": "Bibliography" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 133, + 167, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 138, + 167, + 481, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 167, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 138, + 167, + 481, + 204 + ], + "type": "text", + "content": "[1] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. Advances in neural information processing systems, 31, 2018." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 138, + 203, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 203, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 138, + 203, + 481, + 239 + ], + "type": "text", + "content": "[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 239, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 481, + 262 + ], + "type": "text", + "content": "[3] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM Transactions on Graphics (TOG), 42(4):1-11, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 262, + 481, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 481, + 286 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 481, + 286 + ], + "type": "text", + "content": "[4] Samaneh Azadi, Catherine Olsson, Trevor Darrell, Ian Goodfellow, and Augustus Odena. Discriminator rejection sampling. arXiv preprint arXiv:1810.06758, 2018." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 286, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 286, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 286, + 481, + 369 + ], + "type": "text", + "content": "[5] Shruthi Bannur, Stephanie Hyland, Qianchu Liu, Fernando Perez-Garcia, Maximilian Ilse, Daniel C. Castro, Benedikt Boecking, Harshita Sharma, Kenya Bouzid, Anja Thieme, Anton Schwaighofer, Maria Wetscherek, Matthew P. Lungren, Aditya Nori, Javier Alvarez-Valle, and Ozan Oktay. Learning to Exploit Temporal Structure for Biomedical Vision-Language Processing. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 1 2023. https://doi.org/10.48550/arxiv.2301.04558." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 369, + 481, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 369, + 481, + 416 + ], + "spans": [ + { + "bbox": [ + 138, + 369, + 481, + 416 + ], + "type": "text", + "content": "[6] Riccardo Barbano, Alexander Denker, Hyungjin Chung, Tae Hoon Roh, Simon Arridge, Peter Maass, Bangti Jin, and Jong Chul Ye. Steerable conditional diffusion for out-of-distribution adaptation in imaging inverse problems. arXiv preprint arXiv:2308.14409, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 416, + 481, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 416, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 138, + 416, + 481, + 440 + ], + "type": "text", + "content": "[7] Daniel C Castro, Ian Walker, and Ben Glocker. Causality matters in medical imaging. Nature Communications, 11(1):3673, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 440, + 481, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 440, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 440, + 481, + 498 + ], + "type": "text", + "content": "[8] Pierre Chambon, Christian Bluethgen, Jean-Benoit Delbrouck, Rogier Van der Sluijs, Malgorzata Polacin, Juan Manuel Zambrano Chaves, Tanishq Mathew Abraham, Shivanshu Purohit, Curtis P Langlotz, and Akshay Chaudhari. Roentgen: vision-language foundation model for chest x-ray generation. arXiv preprint arXiv:2211.12737, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 498, + 481, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 498, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 138, + 498, + 481, + 534 + ], + "type": "text", + "content": "[9] Li-Ching Chen, Po-Chih Kuo, Ryan Wang, Judy Gichoya, and Leo Anthony Celi. Chest X-ray segmentation images based on MIMIC-CXR (version 1.0.0). PhysioNet, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 535, + 481, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 535, + 481, + 559 + ], + "spans": [ + { + "bbox": [ + 133, + 535, + 481, + 559 + ], + "type": "text", + "content": "[10] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. DiffEdit: Diffusion-based semantic image editing with mask guidance, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 559, + 481, + 605 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 559, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 133, + 559, + 481, + 605 + ], + "type": "text", + "content": "[11] Alex J. DeGrave, Joseph D. Janizek, and Su-In Lee. AI for radiographic COVID-19 detection selects shortcuts over signal. Nature Machine Intelligence, 3(7):610-619, 2021. ISSN 2522-5839. https://doi.org/10.1038/s42256-021-00338-7." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 605, + 481, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 605, + 481, + 642 + ], + "spans": [ + { + "bbox": [ + 133, + 605, + 481, + 642 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Jost Tobias Springenberg, and Thomas Brox. Learning to generate chairs with convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1538-1546, 2015." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 641, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 641, + 481, + 665 + ], + "type": "text", + "content": "[13] Sijing Feng, Damian Azzollini, Ji Soo Kim, Cheng-Kai Jin, Simon P. Gordon, Jason Yeoh, Eve Kim, Mina Han, Andrew Lee, Aakash Patel, Joy Wu, Martin" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 152, + 115, + 481, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 115, + 481, + 163 + ], + "spans": [ + { + "bbox": [ + 152, + 115, + 481, + 163 + ], + "type": "text", + "content": "Urschler, Amy Fong, Cameron Simmers, Gregory P. Tarr, Stuart Barnard, and Ben Wilson. Curation of the CANDID-PTX dataset with free-text reports. *Radiology: Artificial Intelligence*, 3(6):e210136, 2021. ISSN 2638-6100. https://doi.org/10.1148/ryai.2021210136." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 165, + 481, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 165, + 481, + 212 + ], + "spans": [ + { + "bbox": [ + 133, + 165, + 481, + 212 + ], + "type": "text", + "content": "[14] Virginia Fernandez, Pedro Sanchez, Walter Hugo Lopez Pinaya, Grzegorz Jacenków, Sotirios A Tsaftaris, and Jorge Cardoso. Privacy distillation: Reducing re-identification risk of multimodal diffusion models. arXiv preprint arXiv:2306.01322, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 214, + 481, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 481, + 249 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 481, + 249 + ], + "type": "text", + "content": "[15] Alessandro Fontanella, Grant Mair, Joanna Wardlaw, Emanuele Trucco, and Amos Storkey. Diffusion models for counterfactual generation and anomaly detection in brain images. URL http://arxiv.org/abs/2308.02062." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 251, + 481, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 481, + 297 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 481, + 297 + ], + "type": "text", + "content": "[16] Nicolas Gaggion, Candelaria Mosquera, Lucas Mansilla, Martina Aineseder, Diego H Milone, and Enzo Ferrante. CheXmask: a large-scale dataset of anatomical segmentation masks for multi-center chest X-ray images. arXiv preprint arXiv:2307.03293, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 299, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 299, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 299, + 481, + 346 + ], + "type": "text", + "content": "[17] Rinon Gal, Or Patashnik, Haggai Maron, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. StyleGAN-NADA: CLIP-guided domain adaptation of image generators. ACM Transactions on Graphics, 41(4), 2022. https://doi.org/10.1145/3528223.3530164." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 348, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 481, + 396 + ], + "type": "text", + "content": "[18] Cong Gao, Benjamin D Killeen, Yicheng Hu, Robert B Grupp, Russell H Taylor, Mehran Armand, and Mathias Unberath. Synthetic data accelerates the development of generalizable learning-based algorithms for x-ray image analysis. Nature Machine Intelligence, 5(3):294-308, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 397, + 481, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 481, + 432 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 481, + 432 + ], + "type": "text", + "content": "[19] Timur Garipov, Sebastiaan De Peuter, Ge Yang, Vikas Garg, Samuel Kaski, and Tommi Jaakkola. Compositional sculpting of iterative generative processes. arXiv preprint arXiv:2309.16115, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 434, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 481, + 456 + ], + "type": "text", + "content": "[20] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples, 2015." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 458, + 481, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 458, + 481, + 506 + ], + "spans": [ + { + "bbox": [ + 132, + 458, + 481, + 506 + ], + "type": "text", + "content": "[21] Yu Gu, Jianwei Yang, Naoto Usuyama, Chunyuan Li, Sheng Zhang, Matthew P Lungren, Jianfeng Gao, and Hoifung Poon. Biomedjourney: Counterfactual biomedical image generation by instruction-learning from multimodal patient journeys. arXiv preprint arXiv:2310.10765, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 507, + 481, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 507, + 481, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 507, + 481, + 530 + ], + "type": "text", + "content": "[22] Will Douglas Heaven. Hundreds of ai tools have been built to catch covid. none of them helped. MIT Technology Review. Retrieved December 2023, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 532, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 481, + 567 + ], + "type": "text", + "content": "[23] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2018." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 568, + 481, + 604 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 481, + 604 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 481, + 604 + ], + "type": "text", + "content": "[24] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations, 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 605, + 481, + 652 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 605, + 481, + 652 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 481, + 652 + ], + "type": "text", + "content": "[25] Irina Higgins, Loic Matthew, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. " + }, + { + "bbox": [ + 132, + 605, + 481, + 652 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 132, + 605, + 481, + 652 + ], + "type": "text", + "content": "-VAE: Learning basic visual concepts with a constrained variational framework. In International Conference on Learning Representations, 2016." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 654, + 448, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 654, + 448, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 654, + 448, + 665 + ], + "type": "text", + "content": "[26] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 302, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 302, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 302, + 102 + ], + "type": "text", + "content": "F. Pérez-García, S. Bond-Taylor et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 132, + 115, + 480, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 480, + 139 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 480, + 139 + ], + "type": "text", + "content": "[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 481, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 481, + 176 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 481, + 176 + ], + "type": "text", + "content": "[28] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research, 23(1):2249-2281, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 176, + 481, + 199 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 176, + 481, + 199 + ], + "spans": [ + { + "bbox": [ + 132, + 176, + 481, + 199 + ], + "type": "text", + "content": "[29] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly DDPM noise space: Inversion and manipulations, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 200, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 200, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 200, + 481, + 270 + ], + "type": "text", + "content": "[30] Jeremy Irvin, Pranav Rajpurkar, Michael Ko, Yifan Yu, Silviana Ciurea-Ilicus, Chris Chute, Henrik Marklund, Behzad Haghgoo, Robyn Ball, Katie Shpanskaya, Jayne Seekins, David A. Mong, Safwan S. Halabi, Jesse K. Sandberg, Ricky Jones, David B. Larson, Curtis P. Langlotz, Bhavik N. Patel, Matthew P. Lungren, and Andrew Y. Ng. CheXpert: A large chest radiograph dataset with uncertainty labels and expert comparison, 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 319 + ], + "type": "text", + "content": "[31] Stefan Jaeger, Sema Candemir, Sameer Antani, Yi-Xiang J Wang, Pu-Xuan Lu, and George Thoma. Two public chest x-ray datasets for computer-aided screening of pulmonary diseases. Quantitative imaging in medicine and surgery, 4(6):475, 2014." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 319, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 481, + 342 + ], + "type": "text", + "content": "[32] Priyank Jaini, Kevin Clark, and Robert Geirhos. Intriguing properties of generative classifiers. arXiv preprint arXiv:2309.16779, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 343, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 481, + 402 + ], + "type": "text", + "content": "[33] Alistair E. W. Johnson, Tom J. Pollard, Seth J. Berkowitz, Nathaniel R. Greenbaum, Matthew P. Lungren, Chih-ying Deng, Roger G. Mark, and Steven Horng. MIMIC-CXR, a de-identified publicly available database of chest radiographs with free-text reports. Scientific Data, 6(1):317, 2019. ISSN 2052-4463. https://doi.org/10.1038/s41597-019-0322-0." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 403, + 481, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 438 + ], + "type": "text", + "content": "[34] Charles Jones, Daniel C Castro, Fabio De Sousa Ribeiro, Ozan Oktay, Melissa McCradden, and Ben Glocker. No fair lunch: A causal perspective on dataset bias in machine learning for medical imaging. arXiv preprint arXiv:2307.16526, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 439, + 481, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 481, + 485 + ], + "type": "text", + "content": "[35] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10124–10134, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 486, + 481, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 481, + 510 + ], + "type": "text", + "content": "[36] Priyatham Kattakinda, Alexander Levine, and Soheil Feizi. Invariant learning via diffusion dreamed distribution shifts. arXiv preprint arXiv:2211.10370, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 510, + 481, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 481, + 521 + ], + "type": "text", + "content": "[37] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 522, + 481, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 481, + 569 + ], + "type": "text", + "content": "[38] Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas Phillips, Irena Gao, et al. Wilds: A benchmark of in-the-wild distribution shifts. In International Conference on Machine Learning, pages 5637-5664. PMLR, 2021." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 570, + 481, + 617 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 570, + 481, + 617 + ], + "spans": [ + { + "bbox": [ + 132, + 570, + 481, + 617 + ], + "type": "text", + "content": "[39] Ira Ktena, Olivia Wiles, Isabela Albuquerque, Sylvestre-Alvise Rebuffi, Ryutaro Tanno, Abhijit Guha Roy, Shekoofeh Azizi, Danielle Belgrave, Pushmeet Kohli, Alan Karthikesalingam, et al. Generative models improve fairness of medical classifiers under distribution shifts. arXiv preprint arXiv:2304.09218, 2023." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 618, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 618, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 618, + 481, + 665 + ], + "type": "text", + "content": "[40] Agostina J Larrazabal, Nicolás Nieto, Victoria Peterson, Diego H Milone, and Enzo Ferrante. Gender imbalance in medical imaging datasets produces biased classifiers for computer-aided diagnosis. Proceedings of the National Academy of Sciences, 117(23):12592–12594, 2020." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 171, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 171, + 91, + 448, + 102 + ], + "type": "text", + "content": "RadEdit: stress-testing biomedical vision models via diffusion image editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 140 + ], + "type": "text", + "content": "[41] Choong Ho Lee and Hyung-Jin Yoon. Medical big data: promise and challenges. *Kidney research and clinical practice*, 36(1):3, 2017." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 141, + 480, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 141, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 132, + 141, + 480, + 163 + ], + "type": "text", + "content": "[42] Xiaodan Li, Yuefeng Chen, Yao Zhu, Shuhui Wang, Rong Zhang, and Hui Xue. ImageNet-e: Benchmarking neural network robustness via attribute editing, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 164, + 480, + 199 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 164, + 480, + 199 + ], + "spans": [ + { + "bbox": [ + 133, + 164, + 480, + 199 + ], + "type": "text", + "content": "[43] Wufeng Liu, Jiaxin Luo, Yan Yang, Wenlian Wang, Junkui Deng, and Liang Yu. Automatic lung segmentation in chest x-ray images using improved u-net. Scientific Reports, 12(1):8649, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 200, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 200, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 134, + 200, + 480, + 247 + ], + "type": "text", + "content": "[44] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. Advances in Neural Information Processing Systems, 33:11525-11538, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 247, + 481, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 247, + 481, + 462 + ], + "spans": [ + { + "bbox": [ + 134, + 247, + 481, + 462 + ], + "type": "text", + "content": "[45] Lena Maier-Hein, Annika Reinke, Patrick Godau, Minu D. Tizabi, Florian Buettner, Evangelia Christodoulou, Ben Glocker, Fabian Isensee, Jens Kleesiek, Michal Kozubek, Mauricio Reyes, Michael A. Riegler, Manuel Wiesenfarth, A. Emre Kavur, Carole H. Sudre, Michael Baumgartner, Matthias Eisenmann, Doreen Heckmann-Nötzel, A. Tim Rädsch, Laura Acion, Michela Antonelli, Tal Arbel, Spyridon Bakas, Arrielenis, Matthew Blaschko, M. Jorge Cardoso, Veronika Cheplygina, Beth A. Cimini, Gary S. Collins, Keyvan Farahani, Luciana Ferrer, Adrian Galdran, Bram van Ginneken, Robert Haase, Daniel A. Hashimoto, Michael M. Hoffman, Merel Huisman, Pierre Jannin, Charles E. Kahn, Dagmar Kainmueller, Bernhard Kainz, Alexandros Karargyris, Alan Karthikesalingam, Hannes Kenngott, Florian Kofler, Annette Kopp-Schneider, Anna Kreshuk, Tahsin Kurc, Bennett A. Landman, Geert Litjens, Amin Madani, Klaus Maier-Hein, Anne L. Martel, Peter Mattson, Erik Meijering, Bjoern Menze, Karel G. M. Moons, Henning Müller, Brennan Nichyporuk, Felix Nickel, Jens Petersen, Nasir Rajpoot, Nicola Rieke, Julio Saez-Rodriguez, Clara I. Sánchez, Shravya Shetty, Maarten van Smeden, Ronald M. Summers, Abdel A. Taha, Aleksei Tiulpin, Sotirios A. Tsaftaris, Ben Van Calster, Gaël Varoquaux, and Paul F. Jäger. Metrics reloaded: Recommendations for image analysis validation, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 463, + 481, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 463, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 133, + 463, + 481, + 498 + ], + "type": "text", + "content": "[46] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 498, + 481, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 498, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 133, + 498, + 481, + 546 + ], + "type": "text", + "content": "[47] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Nulltext inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 546, + 481, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 546, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 133, + 546, + 481, + 594 + ], + "type": "text", + "content": "[48] Gustav Müller-Franzes, Jan Moritz Niehues, Firas Khader, Soroosh Tayebi Arasteh, Christoph Haarburger, Christiane Kuhl, Tianci Wang, Tianyu Han, Sven Nebelung, Jakob Nikolas Kather, et al. Diffusion probabilistic models beat gans on medical images. arXiv preprint arXiv:2212.07501, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 594, + 481, + 629 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 594, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 133, + 594, + 481, + 629 + ], + "type": "text", + "content": "[49] Nick Pawlowski, Daniel C. Castro, and Ben Glocker. Deep structural causal models for tractable counterfactual inference. In Advances in Neural Information Processing Systems, volume 33, pages 857-869, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 630, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 630, + 481, + 665 + ], + "type": "text", + "content": "[50] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis, 2023." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 302, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 302, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 302, + 102 + ], + "type": "text", + "content": "F. Pérez-García, S. Bond-Taylor et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 132, + 115, + 482, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 482, + 152 + ], + "type": "text", + "content": "[51] Viraj Prabhu, Sriram Yenamandra, Prithvijit Chattopadhyay, and Judy Hoffman. LANCE: Stress-testing visual models by generating language-guided counterfactual images, 2023." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 153, + 482, + 189 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 153, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 132, + 153, + 482, + 189 + ], + "type": "text", + "content": "[52] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 190, + 482, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 190, + 482, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 190, + 482, + 239 + ], + "type": "text", + "content": "[53] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 239, + 482, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 239, + 482, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 239, + 482, + 274 + ], + "type": "text", + "content": "[54] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 275, + 482, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 275, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 132, + 275, + 482, + 323 + ], + "type": "text", + "content": "[55] Jacob C. Reinhold, Aaron Carass, and Jerry L. Prince. A structural causal model for MR images of multiple sclerosis. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2021, volume 12905 of LNCS, pages 782–792, 2021. https://doi.org/10.1007/978-3-030-87240-3_75." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 324, + 482, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 324, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 324, + 482, + 384 + ], + "type": "text", + "content": "[56] Michael Roberts, Derek Driggs, Matthew Thorpe, Julian Gilbey, Michael Yeung, Stephan Ursprung, Angelica I Aviles-Rivero, Christian Etmann, Cathal McCague, Lucian Beer, et al. Common pitfalls and recommendations for using machine learning to detect and prognosticate for Covid-19 using chest radiographs and ct scans. Nature Machine Intelligence, 3(3):199-217, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 385, + 482, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 482, + 410 + ], + "type": "text", + "content": "[57] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 411, + 482, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 411, + 482, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 411, + 482, + 482 + ], + "type": "text", + "content": "[58] Johannes Rueckel, Lena Trappmann, Balthasar Schachtner, Philipp Wesp, Boj Friedrich Hoppe, Nicola Fink, Jens Ricke, Julien Dinkel, Michael Ingrisch, and Bastian Oliver Sabel. Impact of confounding thoracic tubes and pleural dehiscence extent on artificial intelligence pneumothorax detection in chest radiographs. Investigative Radiology, 55(12):792-798, July 2020. ISSN 0020-9996. https://doi.org/10.1097/rli.0000000000000707." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 483, + 482, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 482, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 482, + 566 + ], + "type": "text", + "content": "[59] Johannes Rueckel, Christian Huemmer, Andreas Fieselmann, Florin-Cristian Gheu, Awais Mansoor, Balthasar Schachtner, Philipp Wesp, Lena Trappmann, Basel Munawwar, Jens Ricke, Michael Ingrisch, and Bastian O. Sabel. Pneumothorax detection in chest radiographs: optimizing artificial intelligence system for accuracy and confounding bias reduction using in-image annotations in algorithm training. European Radiology, 31(10):7888-7900, 2021. https://doi.org/10.1007/s00330-021-07833-w." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 567, + 482, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 482, + 603 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 482, + 603 + ], + "type": "text", + "content": "[60] Chitwan Sahara, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 604, + 482, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 604, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 132, + 604, + 482, + 639 + ], + "type": "text", + "content": "[61] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Semantic foggy scene understanding with synthetic data. International Journal of Computer Vision, 126: 973-992, 2018." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 641, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 482, + 666 + ], + "type": "text", + "content": "[62] Pedro Sanchez, Antanas Kascenas, Xiao Liu, Alison Q O'Neil, and Sotirios A Tsaftaris. What is healthy? generative counterfactual diffusion for lesion localization" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 171, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 171, + 91, + 448, + 102 + ], + "type": "text", + "content": "RadEdit: stress-testing biomedical vision models via diffusion image editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 152, + 115, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 115, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 152, + 115, + 481, + 138 + ], + "type": "text", + "content": "tion. In MICCAI Workshop on Deep Generative Models, pages 34-44. Springer, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 481, + 187 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 481, + 187 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 481, + 187 + ], + "type": "text", + "content": "[63] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 189, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 189, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 189, + 481, + 224 + ], + "type": "text", + "content": "[64] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9243–9252, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 224, + 481, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 224, + 481, + 295 + ], + "spans": [ + { + "bbox": [ + 133, + 224, + 481, + 295 + ], + "type": "text", + "content": "[65] Junji Shiraishi, Shigehiko Katsuragawa, Junpei Ikezoe, Tsuneo Matsumoto, Takeshi Kobayashi, Ken-ichi Komatsu, Mitate Matsui, Hiroshi Fujita, Yoshie Kodera, and Kunio Doi. Development of a digital image database for chest radiographs with and without a lung nodule: receiver operating characteristic analysis of radiologists' detection of pulmonary nodules. American Journal of Roentgenology, 174(1):71-74, 2000." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 296, + 481, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 296, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 133, + 296, + 481, + 331 + ], + "type": "text", + "content": "[66] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034, 2013." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 332, + 481, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 332, + 481, + 367 + ], + "spans": [ + { + "bbox": [ + 133, + 332, + 481, + 367 + ], + "type": "text", + "content": "[67] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 368, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 368, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 133, + 368, + 481, + 390 + ], + "type": "text", + "content": "[68] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 392, + 481, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 392, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 133, + 392, + 481, + 426 + ], + "type": "text", + "content": "[69] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 426, + 481, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 426, + 481, + 473 + ], + "spans": [ + { + "bbox": [ + 133, + 426, + 481, + 473 + ], + "type": "text", + "content": "[70] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 475, + 481, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 475, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 133, + 475, + 481, + 521 + ], + "type": "text", + "content": "[71] Paul Upchurch, Jacob Gardner, Geoff Pleiss, Robert Pless, Noah Snavely, Kavita Bala, and Kilian Weinberger. Deep feature interpolation for image content changes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7064-7073, 2017." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 523, + 481, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 523, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 133, + 523, + 481, + 556 + ], + "type": "text", + "content": "[72] Boris Van Breugel, Nabeel Seedat, Fergus Imrie, and Mihaela van der Schaar. Can you rely on your model evaluation? improving model evaluation with synthetic test data, 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 558, + 481, + 617 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 558, + 481, + 617 + ], + "spans": [ + { + "bbox": [ + 133, + 558, + 481, + 617 + ], + "type": "text", + "content": "[73] Maria de la Iglesia Vayá, Jose Manuel Saborit, Joaquim Angel Montell, Antonio Pertusa, Aurelia Bustos, Miguel Cazorla, Joaquin Galant, Xavier Barber, Domingo Orozco-Beltrán, Francisco García-García, Marisa Caparrós, Germán González, and Jose María Salinas. BIMCV COVID-19+: a large annotated dataset of RX and CT images from COVID-19 patients, 2020. version: 3." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 618, + 481, + 641 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 618, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 133, + 618, + 481, + 641 + ], + "type": "text", + "content": "[74] I von Borzyskowski, A Mazumder, B Mateen, and M Wooldridge. Data science and ai in the age of covid-19, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "text", + "content": "[75] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M. Summers. ChestX-ray8: Hospital-scale chest x-ray database and" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 302, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 302, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 302, + 102 + ], + "type": "text", + "content": "F. Pérez-García, S. Bond-Taylor et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + }, + { + "para_blocks": [ + { + "bbox": [ + 154, + 116, + 481, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 116, + 481, + 163 + ], + "spans": [ + { + "bbox": [ + 154, + 116, + 481, + 163 + ], + "type": "text", + "content": "benchmarks on weakly-supervised classification and localization of common thorax diseases. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3462-3471, 2017. https://doi.org/10.1109/CVPR.2017.369." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 164, + 481, + 248 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 133, + 164, + 481, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 164, + 481, + 212 + ], + "spans": [ + { + "bbox": [ + 133, + 164, + 481, + 212 + ], + "type": "text", + "content": "[76] Laure Wynants, Ben Van Calster, Gary S Collins, Richard D Riley, Georg Heinze, Ewoud Schuit, Elena Albu, Banafsheh Arshi, Vanessa Bellou, Marc MJ Bonten, et al. Prediction models for diagnosis and prognosis of Covid-19: systematic review and critical appraisal. bmj, 369, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 212, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 212, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 133, + 212, + 481, + 248 + ], + "type": "text", + "content": "[77] Anna Zawacki, Carol Wu, George Shih, Julia Elliott, Mikhail Fomitchev, Mohannad ParasLakhani Hussain, Phil Culliton, and Shunxing Bao. Siim-acr pneumothorax segmentation, 2019." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "type": "text", + "content": "RadEdit: stress-testing biomedical vision models via diffusion image editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "21" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 20 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis/f615a129-2ee2-41a8-9f1c-82a796edae0e_content_list.json b/2024/Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis/f615a129-2ee2-41a8-9f1c-82a796edae0e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..04fcd43674c8cdc32181af4ba5583ca1ec84ecc1 --- /dev/null +++ b/2024/Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis/f615a129-2ee2-41a8-9f1c-82a796edae0e_content_list.json @@ -0,0 +1,1815 @@ +[ + { + "type": "text", + "text": "Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis", + "text_level": 1, + "bbox": [ + 269, + 141, + 733, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuanhao Cai $^{1}$ , Yixun Liang $^{2}$ , Jiahao Wang $^{1}$ , Angtian Wang $^{1}$ , Yulun Zhang $^{3,\\dagger}$ , Xiaokang Yang $^{3}$ , Zongwei Zhou $^{1,\\dagger}$ , and Alan Yuille $^{1}$", + "bbox": [ + 241, + 210, + 759, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Johns Hopkins University, 2 HKUST(GZ), 3 Shanghai Jiao Tong University", + "bbox": [ + 241, + 253, + 759, + 270 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. X-ray is widely applied for transmission imaging due to its stronger penetration than natural light. When rendering novel view X-ray projections, existing methods mainly based on NeRF suffer from long training time and slow inference speed. In this paper, we propose a 3D Gaussian splatting-based method, namely X-Gaussian, for X-ray novel view synthesis. Firstly, we redesign a radiative Gaussian point cloud model inspired by the isotropic nature of X-ray imaging. Our model excludes the influence of view direction when learning to predict the radiation intensity of 3D points. Based on this model, we develop a Differentiable Radiative Rasterization (DRR) with CUDA implementation. Secondly, we customize an Angle-pose Cuboid Uniform Initialization (ACUI) strategy that directly uses the parameters of the X-ray scanner to compute the camera information and then uniformly samples point positions within a cuboid enclosing the scanned object. Experiments show that our X-Gaussian outperforms state-of-the-art methods by 6.5 dB while enjoying less than $15\\%$ training time and over $73\\times$ inference speed. The application on CT reconstruction also reveals the practical values of our method. Code is at https://github.com/caiyuanhao1998/X-Gaussian", + "bbox": [ + 259, + 300, + 743, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Medical Imaging $\\cdot$ View Synthesis $\\cdot$ Gaussian Splatting", + "bbox": [ + 261, + 563, + 720, + 578 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 621, + 375, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "X-ray novel view synthesis (NVS) aims to create X-ray projections of an object from new viewpoints that are not originally captured, using only existing projections scanned from different view directions. As we know, X-ray has stronger penetrating power to capture internal structures of imaged objects and is therefore widely applied in medical imaging [12,13,15,18,19,43]. Nonetheless, X-ray is harmful to human body due to its powerful ionizing radiation, especially when the dose of X-ray increases. Improving NVS techniques can help reduce the exposure to X-rays and provide comprehensive viewpoints of imaged parts for doctors and downstream tasks such as sparse-view CT reconstruction. Hence, X-ray NVS is very important and valuable. We study this task in the circular cone beam X-ray scanning scenario [7,8,10,35,39,44,56,59].", + "bbox": [ + 212, + 650, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger =$ corresponding authors.", + "bbox": [ + 230, + 825, + 415, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/97b2aca403e0517f00163ab81701ee073b786d63c5262cb1690b7816622656fd.jpg", + "image_caption": [ + "Fig. 1: Point cloud visualization of the original 3DGS [25] (top) and our X-Gaussian (bottom). We visualize the positions and opacities of the Gaussian point clouds at different training iterations. We also visualize the volume of foot as a reference. Note that the volume is not the ground truth of point clouds. Our X-Gaussian can better represent the detailed structures than 3DGS, showing faster and better convergence." + ], + "image_footnote": [], + "bbox": [ + 212, + 146, + 777, + 309 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing methods are mainly based on neural radiance fields (NeRF) [37]. They usually employ a multi-layer perceptron (MLP) to learn the mapping from the point position to the radiodensity and then create projections by volume rendering along rays. This ray tracing scheme is time-consuming because it needs to sample many 3D points and then compute them for every single ray, slowing down the training and inference processes. Even the recent most efficient NeRF-based method [60] still requires over an hour for training and yields suboptimal results at a slow inference speed of 2 fps. This increases the waiting time of patients and doctors, leading to low diagnostic efficiency.", + "bbox": [ + 212, + 443, + 787, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, 3D Gaussian splattering (3DGS) [25] has demonstrated promising reconstruction quality while enjoying much faster inference speed than NeRF-based algorithms in RGB domain, which motivates us to follow this technical route. However, due to the fundamental differences between X-ray and natural light imaging, directly applying the original 3DGS to X-ray NVS may encounter two issues. Firstly, the spherical harmonics (SH) in RGB 3DGS is not suitable for modeling the X-ray radiation intensity of 3D points. Specifically, natural light imaging relies on the reflection off the surface. The color of a 3D point is anisotropic and view-dependent. Based on this nature, the original Gaussian point cloud model uses SH to fit the illumination distribution. In contrast, X-rays penetrate the object and attenuate, thereby forming an image. Given specific X-rays, the radiation intensity of a 3D point depends on its radiodensity and is independent to the view direction, which means the point radiation intensity is isotropic. Secondly, the original point cloud initialization algorithm, structure-from-motion (SfM) [46], is also not suitable for X-ray imaging. Compared to RGB images, X-ray images are grayscale and their contrast is lower. Additionally, different layers of an object may overlap on the same position of the projection", + "bbox": [ + 212, + 583, + 787, + 843 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Yuanhao Cai et al.", + "bbox": [ + 271, + 114, + 397, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "due to the transmission imaging nature of X-rays. These two problems degrade the accuracy of feature detection and matching in SfM. Meanwhile, running SfM is time-consuming, which prolongs the training process of Gaussian point clouds.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the above issues, we propose a novel 3DGS-based method, X-Gaussian, for X-ray NVs. Our X-Gaussian composes two key techniques. Firstly, we redesign a radiative Gaussian point cloud model inspired by the isotropic property of X-ray imaging. We present a Radiation Intensity Response Function (RIRF) to replace the SH function of the original 3DGS. Different from SH, our RIRF excludes the influence of view direction. To this end, it adopts the inner product between a learnable vector representing the inherent point features and a set of basis weights to fit the radiation intensity of a 3D point. Based on this point cloud model, we further develop a Differentiable Radiative Rasterization (DRR) with a CUDA implementation to render novel projections. Secondly, we customize an Angle-pose Cuboid Uniform Initialization (ACUI) strategy for camera calibration parameters and Gaussian point clouds. Our ACUI first exploits the parameters of the X-ray scanner to compute the intrinsic and extrinsic matrices. Then we set up a cuboid that can completely enclose the scanned object. Within this cuboid, we uniformly sample 3D points at intervals to initialize the center positions of the Gaussian point clouds. Free from running the SfM algorithm, our ACUI significantly reduces the training time. Equipped with the two proposed techniques, our X-Gaussian outperforms SOTA methods by 6.5 dB while enjoying $73 \\times$ inference speed and $7 \\times$ training speed.", + "bbox": [ + 212, + 191, + 787, + 478 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main contributions of this work can be summarized as follows:", + "bbox": [ + 238, + 479, + 723, + 493 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel 3D Gaussian splatting-based framework, X-Gaussian, for X-ray novel view synthesis. To our knowledge, this is the first attempt to explore the potential of Gaussian splatting in X-ray neural rendering.", + "- We design a radiative Gaussian point cloud model with a differentiable radiative rasterization based on the isotropic nature of X-ray imaging.", + "- We present an angle-pose cuboid uniform initialization strategy for Gaussian point clouds and camera calibration in circular cone beam X-ray scanning.", + "- Our X-Gaussian significantly outperforms SOTA NeRF-based methods with much faster speed. Experiments also show that our method can improve the performance of sparse-view CT reconstruction, showing its practical values." + ], + "bbox": [ + 223, + 503, + 782, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 676, + 387, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Neural Radiance Field", + "text_level": 1, + "bbox": [ + 215, + 709, + 447, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "NeRF [37] learns an implicit neural scene representation of color and volume density, given the position of a 3D point and view direction. It has achieved great success in NVS and inspired an explosion of follow-up papers to improve its quality [3-5,22,48] and speed [9,11,21,30,38,41,54]. For example, Instant-NGP [38] adopts hash tables as the encoder to allow small MLP for fast training and inference. Some later works extend the application domain of NeRF from natural light to X-rays [14,58,60]. For instance, NAF [60] follows the settings", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Radiative Gaussian Splitting for Efficient X-ray Novel View Synthesis", + "bbox": [ + 263, + 114, + 732, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/16f43c0f96a51489bfb4872285d3dfff2b5ec580c759d311106f60a8a62b4425.jpg", + "image_caption": [ + "Fig. 2: Pipeline of our method. (a) Angle-posed Cuboid Uniform Initialization (ACUI) strategy uses the parameters of X-ray scanner to compute intrinsic and extrinsic matrices, and samples center points for 3D Gaussians. (b) Our radiative Gaussian point cloud model learns to predict the radiation intensity of 3D points. (c) Based on our Gaussian model, we develop a GPU-friendly Differentiable Radiative Rasterization (DRR)." + ], + "image_footnote": [], + "bbox": [ + 210, + 148, + 782, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of Instant-NGP to learn the implicit mapping from 3D position to attenuation. Yet, the ray tracing and volume rendering schemes are time-consuming, which limits the training and inference speed of NeRF-based X-ray NVS algorithms.", + "bbox": [ + 212, + 417, + 787, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Gaussian Splitting", + "text_level": 1, + "bbox": [ + 214, + 486, + 419, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3DGS [25] represents scenes using millions of 3D Gaussian point clouds. This approach is fundamentally different from NeRF-based algorithms by employing an explicit representation coupled with highly parallelized rasterization workflows. These features enable more efficient computation and rendering processes. Hence, 3DGS has achieved great success in several fields, including 3D Generation [20,28,31,33,47,55,61], Dynamic Scene Modeling [34,50,53], SLAM [24,36,52,57], Inverse Rendering [23,32,51], etc. However, most applications of 3DGS are focused on natural scenes with RGB colors. The potential of 3DGS in X-ray imaging still remains under-explored. Our goal is to fill this research gap.", + "bbox": [ + 212, + 511, + 787, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 214, + 671, + 330, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The pipeline of our X-Gaussian is shown in Fig. 2. Firstly, we design an Angle-posed Cuboid Uniform Initialization (ACUI) to compute the intrinsic and extrinsic matrices from the parameters of X-ray scanner, as illustrated in Fig. 2 (a). Then ACUI uniformly samples 3D points within a cuboid that can completely enclose the scanned object to initialize the center positions of our radiative Gaussian point clouds in Fig. 2 (b). Given a view direction, the 3D point clouds undergo our Differentiable Radiative Rasterization (DRR) to derive the rendered image, as depicted in Fig. 2 (c). In this section, we will introduce our radiative Gaussian point cloud model and DRR processing first and then the ACUI strategy.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Yuanhao Cai et al.", + "bbox": [ + 271, + 114, + 397, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3265cdb3c1db691484126c0b557ae8f76e0819232c02b24238280df5e57b12b9.jpg", + "image_caption": [ + "(a) Original RGB Gaussian Point Cloud" + ], + "image_footnote": [], + "bbox": [ + 218, + 148, + 480, + 262 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d8fb9744312a4a78a7b5c5945597d095b461df18b1a031ce95131cf733984593.jpg", + "image_caption": [ + "(b) Our Radiative Gaussian Point Cloud", + "Fig. 3: Comparison between the Gaussian point cloud models of the original 3DGS and our X-Gaussian. (a) The original RGB Gaussian point cloud model uses spherical harmonics (SH) to simulate the anisotropic natural light distribution and view-dependent color. (b) Our radiative Gaussian point cloud model employs the weighted sum of point features to fit the isotropic X-ray penetration and view-independent radiation intensity." + ], + "image_footnote": [], + "bbox": [ + 537, + 147, + 777, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Radiative Gaussian Point Cloud Model", + "text_level": 1, + "bbox": [ + 215, + 409, + 584, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "An object can be represented by a set of basic Gaussian point clouds $\\mathcal{G}$ as", + "bbox": [ + 215, + 433, + 750, + 448 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} = \\left\\{G _ {i} \\left(\\boldsymbol {\\mu} _ {i}, \\boldsymbol {\\Sigma} _ {i}, \\alpha_ {i}\\right) \\mid i = 1, 2, \\dots , N _ {p} \\right\\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 460, + 784, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $G_{i}$ refers to the $i$ -th Gaussian point cloud. Its center position, covariance, and opacity are defined as $\\pmb{\\mu}_{i} \\in \\mathbb{R}^{3}$ , $\\pmb{\\Sigma}_{i} \\in \\mathbb{R}^{3 \\times 3}$ , and $\\alpha_{i} \\in \\mathbb{R}$ . $\\pmb{\\Sigma}_{i}$ is represented by a rotation matrix $\\mathbf{R}_i \\in \\mathbb{R}^3$ and a scaling matrix $\\mathbf{S}_i \\in \\mathbb{R}^3$ as $\\pmb{\\Sigma}_{i} = \\mathbf{R}_{i}\\mathbf{S}_{i}\\mathbf{S}_{i}^{\\top}\\mathbf{R}_{i}^{\\top}$ . $\\pmb{\\mu}_{i}, \\pmb{\\Sigma}_{i}, \\alpha_{i}, \\mathbf{R}_{i}$ , and $\\mathbf{S}_i$ are learnable parameters. Besides these basic attributes, each Gaussian point cloud also employs additional learnable parameters to fit different imaging scenarios, e.g., natural light imaging and X-ray imaging.", + "bbox": [ + 214, + 487, + 782, + 577 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first review the original RGB Gaussian point cloud model [25] in natural light imaging. As shown in Fig. 3 (a), the color of a 3D point is represented by spherical harmonics (SH). The point color is anisotropic and changes with the view direction. Each Gaussian point cloud learns to predict the SH coefficients $\\mathbf{k} = \\{k_l^m |0\\leq l\\leq L, - l\\leq m\\leq l\\} \\in \\mathbb{R}^{(L + 1)^2\\times 3}$ , where each $k_{l}^{m}\\in \\mathbb{R}^{3}$ is a set of 3 coefficients corresponding to the RGB components. $L$ is the degree of SH. Then the point color $\\mathbf{c}\\in \\mathbb{R}^3$ at the view direction $\\mathbf{d} = (\\theta ,\\phi)$ is derived by", + "bbox": [ + 214, + 578, + 784, + 684 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} (\\mathbf {d}, \\mathbf {k}) = \\sum_ {l = 0} ^ {L} \\sum_ {m = - l} ^ {l} k _ {l} ^ {m} Y _ {l} ^ {m} (\\theta , \\phi), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 695, + 784, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Y_{l}^{m}:\\mathbb{S}^{2}\\to \\mathbb{R}$ is the SH function that maps points on the sphere to real numbers. Please refer to the supplementary for its detailed formulation.", + "bbox": [ + 214, + 750, + 782, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although 3DGS [25] achieves fast inference speed and good performance in natural light imaging, the RGB Gaussian point cloud model is not suitable for X-ray scenarios due to the fundamental differences between natural light imaging and X-ray imaging. Natural light imaging relies on the reflection off the surface of", + "bbox": [ + 214, + 779, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Radiative Gaussian Splitting for Efficient X-ray Novel View Synthesis", + "bbox": [ + 263, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "object. The anisotropic color modeled by SH is view-dependent. e.g., in Fig. 3 (a), the point color is blue from the left viewpoint and green from the right viewpoint. In contrast, X-ray imaging is based on the attenuation when penetrating the object. The degree of attenuation depends on the isotropic radiodensity property. Thus, the radiation intensity of a 3D point is view-independent.", + "bbox": [ + 212, + 146, + 782, + 222 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In light of the above analysis, we redesign our radiative Gaussian point cloud model. Different from the original 3DGS that uses SH to fit the color information for each point, our model introduces a Radiation Intensity Response Function (RIRF) to predict the radiation intensity of the 3D point. As illustrated in Fig. 3 (b), each Gaussian point cloud learns a feature vector $\\mathbf{f} \\in \\mathbb{R}^{N_f}$ to represent its inherent radiative properties. Subsequently, the radiation intensity $\\mathbf{i} \\in \\mathbb{R}$ of the center point of a 3D Gaussian at any view direction is modeled by RIRF as", + "bbox": [ + 212, + 222, + 784, + 328 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {i} (\\mathbf {f}) = \\operatorname {R I R F} (\\mathbf {f}) = \\operatorname {S i g m o i d} (\\boldsymbol {\\lambda} \\odot \\mathbf {f}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 339, + 784, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the Sigmoid function activates and normalizes the radiation intensity. $\\lambda \\in \\mathbb{R}^{N_f}$ is a set of constant weights controlling the importance of each component of $\\mathbf{f}$ . Then the set of our radiative Gaussian point clouds $\\mathcal{G}_x$ is formulated as", + "bbox": [ + 214, + 364, + 782, + 410 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {x} = \\left\\{G _ {i} \\left(\\boldsymbol {\\mu} _ {i}, \\boldsymbol {\\Sigma} _ {i}, \\alpha_ {i}, \\mathbf {f} _ {i}\\right) \\mid i = 1, 2, \\dots , N _ {p} \\right\\}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 421, + 784, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{f}_i\\in \\mathbb{R}^{N_f}$ denotes the feature vector of the $i$ -th Gaussian point cloud. Please note that Eq. (3) excludes the influence of the view direction $\\mathbf{d} = (\\theta ,\\phi)$ , which matches the isotropic nature of X-ray imaging. Meanwhile, Eq. (3) is free from the complex computation of SH function. Hence, the forward and backward processes of our X-Gaussian are much faster than those of the original 3DGS.", + "bbox": [ + 214, + 446, + 782, + 522 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Differentiable Radiative Rasterization", + "text_level": 1, + "bbox": [ + 215, + 542, + 573, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on our radiative Gaussian point cloud, we develop a Differentiable Radiative Rasterization (DRR), as shown in Fig. 2 (c). We first summarize the overall DRR processing $F_{\\mathrm{DRR}}$ and then describe its details. DRR is represented as", + "bbox": [ + 214, + 566, + 782, + 612 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} = F _ {\\mathrm {D R R}} \\left(\\mathbf {M} _ {e x t}, \\mathbf {M} _ {i n t}, \\left\\{G _ {i} \\left(\\boldsymbol {\\mu} _ {i}, \\boldsymbol {\\Sigma} _ {i}, \\alpha_ {i}, \\mathbf {f} _ {i}\\right) \\mid i = 1, 2, \\dots , N _ {p} \\right\\}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 287, + 625, + 784, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{I} \\in \\mathbb{R}^{H \\times W}$ denotes the rendered image, $\\mathbf{M}_{ext} \\in \\mathbb{R}^{4 \\times 4}$ represents the extrinsic matrix, and $\\mathbf{M}_{int} \\in \\mathbb{R}^{4 \\times 3}$ refers to the intrinsic matrix. Subsequently, we introduce the details of $F_{\\mathrm{DRR}}$ . To begin with, the possibility value of the $i$ -th Gaussian distribution at the 3D point position $\\mathbf{x} \\in \\mathbb{R}^3$ is formulated as", + "bbox": [ + 214, + 648, + 782, + 710 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP (\\mathbf {x} \\mid \\boldsymbol {\\mu} _ {i}, \\boldsymbol {\\Sigma} _ {i}) = \\exp \\left(- \\frac {1}{2} (\\mathbf {x} - \\boldsymbol {\\mu} _ {i}) ^ {\\top} \\boldsymbol {\\Sigma} _ {i} ^ {- 1} (\\mathbf {x} - \\boldsymbol {\\mu} _ {i})\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 720, + 784, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then we project the 3D Gaussians to the 2D detector plane for subsequent rendering. $\\pmb{\\mu}_i$ is firstly transferred from the world coordinate system to the camera coordinate system and then projected to the image coordinate system as", + "bbox": [ + 214, + 755, + 782, + 801 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {\\mathbf {t}} _ {i} = \\left[ \\begin{array}{l} \\mathbf {t} _ {i} \\\\ 1 \\end{array} \\right] = \\mathbf {M} _ {e x t} \\widetilde {\\boldsymbol {\\mu}} _ {i} = \\mathbf {M} _ {e x t} \\left[ \\begin{array}{l} \\boldsymbol {\\mu} _ {i} \\\\ 1 \\end{array} \\right], \\quad \\widetilde {\\mathbf {u}} _ {i} = \\left[ \\begin{array}{l} \\mathbf {u} _ {i} \\\\ 1 \\end{array} \\right] = \\mathbf {M} _ {i n t} \\widetilde {\\mathbf {t}} _ {i} = \\mathbf {M} _ {i n t} \\left[ \\begin{array}{l} \\mathbf {t} _ {i} \\\\ 1 \\end{array} \\right], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 811, + 782, + 844 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Yuanhao Cai et al.", + "bbox": [ + 271, + 114, + 397, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{t}_i = (t_x, t_y, t_z) \\in \\mathbb{R}^3$ is the camera coordinate of $\\pmb{\\mu}_i$ and $\\mathbf{u}_i \\in \\mathbb{R}^2$ is the image coordinate of $\\pmb{\\mu}_i$ . $\\widetilde{\\mathbf{u}}_i$ , $\\widetilde{\\mathbf{t}}_i$ , and $\\widetilde{\\pmb{\\mu}}_i$ are the homogeneous coordinates of $\\mathbf{u}_i$ , $\\mathbf{t}_i$ , and $\\pmb{\\mu}_i$ , respectively. Subsequently, we transfer the 3D covariance matrix $\\pmb{\\Sigma}_i$ to its counterpart $\\pmb{\\Sigma}_i' \\in \\mathbb{R}^{3 \\times 3}$ in the camera coordinate system as", + "bbox": [ + 214, + 145, + 787, + 209 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} _ {i} ^ {\\prime} = \\mathbf {J} _ {i} \\mathbf {W} _ {i} \\boldsymbol {\\Sigma} _ {i} \\mathbf {W} _ {i} ^ {\\top} \\mathbf {J} _ {i} ^ {\\top}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 220, + 785, + 239 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{J}_i\\in \\mathbb{R}^{3\\times 3}$ is the Jacobian of the affine approximation of the projective transformation. $\\mathbf{W}_i\\in \\mathbb{R}^{3\\times 3}$ is the viewing transformation. We derive them by", + "bbox": [ + 214, + 250, + 785, + 280 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {J} _ {i} = \\left[ \\begin{array}{c c c} \\frac {L _ {S D}}{t _ {z}} & 0 & - \\frac {L _ {S D}}{t _ {z} ^ {2}} \\frac {t _ {x}}{t _ {z}} \\\\ 0 & \\frac {L _ {S D}}{t _ {z}} & - \\frac {L _ {S D}}{t _ {z} ^ {2}} \\frac {t _ {y}}{t _ {z}} \\\\ 0 & 0 & 0 \\end{array} \\right], \\quad \\mathbf {W} _ {i} = \\left[ \\begin{array}{c c c} - \\sin \\phi & \\cos \\phi & 0 \\\\ 0 & 0 & - 1 \\\\ - \\cos \\phi & - \\sin \\phi & 0 \\end{array} \\right], \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 284, + 292, + 785, + 347 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $L_{SD}$ represents the distance between the X-ray source and detector. $\\phi$ refers to the azimuth angle of the source. Following [25,29,62], we obtain the 2D covariance matrix $\\pmb{\\Sigma}_i^{\\prime \\prime} \\in \\mathbb{R}^{2 \\times 2}$ by skipping the third row and column of $\\pmb{\\Sigma}_i^{\\prime}$ . Then the 2D projection is partitioned into non-overlapping tiles. The 3D Gaussians $(\\pmb{\\mu}_i, \\pmb{\\Sigma}_i)$ are assigned to different tiles according to their 2D projections $(\\pmb{u}_i, \\pmb{\\Sigma}_i^{\\prime \\prime})$ , as shown in the left image of Fig. 2 (c). These 3D Gaussians are sorted by the distances to the 2D detector. Then the intensity $\\mathbf{I}(p) \\in \\mathbb{R}$ at pixel $p$ is obtained by blending $\\mathcal{N}$ ordered points overlapping the pixel in the corresponding tile as", + "bbox": [ + 214, + 356, + 787, + 479 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} (p) = \\sum_ {j \\in \\mathcal {N}} \\mathbf {i} _ {j} \\sigma_ {j} \\prod_ {k = 1} ^ {j - 1} (1 - \\sigma_ {k}), \\quad \\sigma_ {j} = \\alpha_ {j} P (\\mathbf {x} _ {j} | \\boldsymbol {\\mu} _ {j}, \\boldsymbol {\\Sigma} _ {j}), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 310, + 491, + 785, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_j$ is the $j$ -th intersection 3D point of the X-ray landing on pixel $p$ and the Gaussian point clouds in 3D space. $\\mathbf{i}_j$ is the radiation intensity of $\\mathbf{x}_j$ .", + "bbox": [ + 214, + 542, + 785, + 573 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Optimization. Eventually, the training objective $\\mathcal{L}$ is the weighted sum of $\\mathcal{L}_1$ loss and SSIM loss between the rendered and ground-truth projection images as", + "bbox": [ + 214, + 573, + 785, + 603 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = (1 - \\gamma) \\mathcal {L} _ {1} + \\gamma \\mathcal {L} _ {\\text {S S I M}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 619, + 785, + 635 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\gamma$ is a hyperparameter balancing the importances of the two loss terms. By minimizing Eq. (11), we can optimize the attributes of 3D Gaussians, i.e., $\\pmb{\\mu}_i, \\pmb{\\Sigma}_i, \\alpha_i$ , and $\\mathbf{f}_i$ in Eq. (5). $N_p$ is adjusted by the adaptive control [25]. The optimization process is visualized in Fig. 1 and the video file in supplementary.", + "bbox": [ + 214, + 641, + 785, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compared to the RGB rasterization in 3DGS [25], our DRR avoids the complex computations related to the view direction in the forward and backward processes, thereby enjoying cheaper training costs and faster inference speed.", + "bbox": [ + 214, + 702, + 785, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Angle-pose Cuboid Uniform Initialization", + "text_level": 1, + "bbox": [ + 214, + 768, + 606, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "At the beginning of training, we need to initialize the parameters in Eq. (5) for rasterization. Specifically, $\\mathbf{\\Sigma}_i$ , $\\alpha_i$ , and $\\mathbf{f}_i$ are randomly initialized. In natural light imaging, the original 3DGS [25] adopts the SfM [46] algorithm to compute the", + "bbox": [ + 214, + 794, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Radiative Gaussian Splitting for Efficient X-ray Novel View Synthesis", + "bbox": [ + 263, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "initial $\\pmb{\\mu}_i$ , $N_p$ , $\\mathbf{M}_{ext}$ , and $\\mathbf{M}_{int}$ . SfM detects and matches features from multiview images. It is not suitable for X-ray imaging due to two reasons. Firstly, X-ray images are grayscale and low-contrast. Secondly, different layers of an object may overlap on the same positions of the projection. These two problems degrade the accuracy of feature detection and matching in SfM. Besides, running the SfM algorithm usually requires a long time, which prolongs the training process.", + "bbox": [ + 212, + 146, + 782, + 237 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To address these issues, we customize an Angle-posed Cuboid Uniform Initialization (ACUI) strategy for circular cone beam X-ray scanning scenario where a scanner emits cone-shaped X-ray beams and captures projections at equal angular intervals. As shown in Fig. 2 (a), ACUI uses the parameters of X-ray scanner to compute the extrinsic matrix $\\mathbf{M}_{ext}$ and intrinsic matrix $\\mathbf{M}_{int}$ as", + "bbox": [ + 212, + 238, + 784, + 314 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {e x t} = \\left[ \\begin{array}{c c c c} - \\sin \\phi & \\cos \\phi & 0 & 0 \\\\ 0 & 0 & - 1 & 0 \\\\ - \\cos \\phi & - \\sin \\phi & 0 & L _ {S O} \\\\ 0 & 0 & 0 & 1 \\end{array} \\right], \\quad \\mathbf {M} _ {i n t} = \\left[ \\begin{array}{c c c c} L _ {S D} & 0 & W / 2 & 0 \\\\ 0 & L _ {S D} & H / 2 & 0 \\\\ 0 & 0 & 1 & 0 \\end{array} \\right], \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 238, + 328, + 784, + 391 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $L_{SO}$ represents the distance between the X-ray source and the scanned object. The elevation angle of the X-ray source is set to zero and remains unchanged. The next step of ACUI is to initialize the center positions of 3D Gaussians. Although the precise shape of the scanned object is not given at the beginning, the scanning space can be approximated. We set up a cuboid with size $S_{1} \\times S_{2} \\times S_{3}$ (mm) that can completely enclose the object. The center of this cuboid is also the center of the object and the origin of the world coordinate system. We divide this cuboid by a grid with size $M_{1} \\times M_{2} \\times M_{3}$ (voxel). Then we uniformly sample points within the grid at interval $d \\in \\mathbb{R}$ as", + "bbox": [ + 212, + 405, + 784, + 541 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} = \\left\\{\\left(\\frac {n _ {1} S _ {1} d}{M _ {1}}, \\frac {n _ {2} S _ {2} d}{M _ {2}}, \\frac {n _ {3} S _ {3} d}{M _ {3}}\\right) \\mid - \\left[ \\frac {M _ {i}}{2 d} \\right] - 1 \\leq n _ {i} \\leq \\left[ \\frac {M _ {i}}{2 d} \\right] + 1, i = 1, 2, 3 \\right\\}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 554, + 784, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $n_i \\in \\mathbb{Z}$ . Then we use the size and elements of $\\mathcal{P}$ to initialize $N_p$ and $\\pmb{\\mu}_i$ . Avoiding running SfM, ACUI allows X-Gaussian to enjoy a faster training speed.", + "bbox": [ + 214, + 598, + 782, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 660, + 375, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Experimental Settings", + "text_level": 1, + "bbox": [ + 214, + 700, + 444, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dataset. Following NAF [60], we adopt the public datasets of human organ CTs, i.e., LIDC-IDRI [2] and the open scientific visualization dataset [27], to evaluate our method. The test scenes include chest, foot, head, abdomen, and pancreas. We adopt the open-source tomographic toolbox TIGRE [6] to capture 100 projections with $3\\%$ noise in the range of $0 \\sim 180^{\\circ}$ . In the NVS task, 50 projections are used for training and the other 50 projections are used for testing. The CT volumes are used for testing in the sparse-view CT reconstruction task.", + "bbox": [ + 212, + 733, + 782, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Yuanhao Cai et al.", + "bbox": [ + 271, + 114, + 398, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/557d50bc5802d32c245254be3795a0e91f46a231319c2291e4cf3ccfc1f21f4c.jpg", + "table_caption": [ + "Table 1: Quantitative results on the novel view synthesis task. The average inference speed and training time of all scenes evaluated on an RTX 8000 GPU are reported. In the cell of the results of each scene, PSNR (upper) and SSIM (lower) are listed." + ], + "table_footnote": [], + "table_body": "
MethodInfer SpeedTrain TimeChestFootHeadAbdomenPancreasAverage
InTomo [58]0.62 fps125 min28.94839.48234.83227.64120.03130.187
0.99150.99790.99770.96460.85370.9611
NeRF [37]0.14 fps313 min36.15741.05329.76024.62019.85330.289
0.99880.99890.99910.95590.85600.9617
TensoRF [9]0.77 fps178 min23.60937.72834.42927.38229.23530.477
0.94020.99290.98790.87300.80310.9194
NeAT [42]1.78 fps69 min40.76538.23627.73826.74137.52634.201
0.99900.99630.92950.85630.90170.9366
NAF [60]2.01 fps63 min42.36638.35330.17437.59036.22836.942
0.99930.99130.95310.98550.88440.9627
X-Gaussian148 fps9 min43.88742.15341.57945.76243.64043.404
0.99980.99970.99970.99990.99760.9993
", + "bbox": [ + 218, + 198, + 784, + 426 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details. Our X-Gaussian is implemented by PyTorch [40] and CUDA [17]. The model is trained with the Adam optimizer [26] ( $\\beta_{1} = 0.9$ , $\\beta_{2} = 0.999$ , and $\\epsilon = 1\\times 10^{-15}$ ) for $2\\times 10^{4}$ iterations. The learning rate for point cloud position is initially set to $1.9\\times 10^{-4}$ and exponentially decays to $1.9\\times 10^{-6}$ . The learning rates for point feature, opacity, scaling, and rotation are set to $2\\times 10^{-3}$ , $8\\times 10^{-3}$ , $5\\times 10^{-3}$ , and $1\\times 10^{-3}$ . $\\gamma$ in Eq. (11) is set to 0.2. We adopt peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) [49] to evaluate the performance. Frames per second (fps) is used to measure the inference speed. Experiments are conducted on an RTX 8000 GPU.", + "bbox": [ + 212, + 455, + 787, + 593 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Novel View Synthesis", + "text_level": 1, + "bbox": [ + 214, + 616, + 441, + 633 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Quantitative Results. Tab. 1 shows the quantitative comparisons between our X-Gaussian and five SOTA NeRF-based algorithms, including InTomo [58], NeRF [37], TensoRF [9], NeAT [42], and NAF [60] on the NVs task.", + "bbox": [ + 212, + 642, + 785, + 688 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We report the average inference speed and training time of different methods on all scenes. In the cell of the results of each scene, PSNR (upper entry in the cell) and SSIM (lower entry in the cell) are listed. As can be observed that our X-Gaussian not only surpasses SOTA methods by large margins in performance but also enjoys much faster inference speed and cheaper training costs. More specifically, compared with the recent best X-ray NeRF-based method NAF, our X-Gaussian outperforms it by 6.5 dB on average and is $73 \\times$ faster in inference while requiring less than $15\\%$ training time. When compared with the SOTA RGB NeRF-based method TensoRF, our X-Gaussian is 12.93 dB higher while enjoying $192 \\times$ inference speed and $20 \\times$ training speed.", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Radiative Gaussian Splitting for Efficient X-ray Novel View Synthesis", + "bbox": [ + 263, + 113, + 730, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/704a3c2ce3cce0b91c4a450be244bf2b5a2d6b6b689c15768c55d3dbcb1202ca.jpg", + "image_caption": [ + "Fig. 4: Qualitative results of novel view synthesis on the scenes of pancreas (top) and chest (bottom). Our X-Gaussian yields clearer results. Please zoom in for a better view." + ], + "image_footnote": [], + "bbox": [ + 217, + 145, + 785, + 358 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Qualitative Results. Figs. 4 depict the qualitative comparisons of NVS on the scenes of pancreas and chest. It can be observed from the zoomed-in patches that previous NeRF-based algorithms fail to render clear novel views. They either introduce undesired artifacts or produce blurry textures such as the toe bones of the foot. In contrast, our method yields visually realistic images by rendering more fine-grained details and clearer structural contents.", + "bbox": [ + 212, + 441, + 787, + 532 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Sparse-View CT Reconstruction", + "text_level": 1, + "bbox": [ + 215, + 556, + 529, + 571 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We compare our method with SOTA NeRF-based algorithms on sparse-view CT reconstruction. Since the Gaussian point clouds cannot directly infer the radiodensities of the CT volume, we evaluate different NeRF-based methods and our X-Gaussian by using them to create novel-view projections for three learning-free CT reconstruction methods, including an analytical method (FDK [16]) and two iterative methods (SART [1] and ASD-POCS [45]). Specifically, these three methods reconstruct the CTs from 5 original projections and 95 novel-view projections rendered by different NVS algorithms. The quantitative results are listed in Tab. 2. When only using 5 original views (+ None), FDK, SART, and ASD-POCS achieve 7.41, 17.24, and $17.03\\mathrm{dB}$ in PSNR, respectively. They fail to reconstruct the CT volumes. When employing our X-Gaussian to create novel X-ray projections for FDK, SART, and ASD-POCS, they yield the most significant improvements of 15.19, 13.01, and $13.53\\mathrm{dB}$ in PSNR. These improvements are 1.32, 2.41, and $2.65\\mathrm{dB}$ higher than the improvements of using NAF.", + "bbox": [ + 212, + 583, + 787, + 792 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Figs. 5 shows the qualitative results of sparse-view CT reconstruction on the scenes of foot and chest. Without using rendered novel-view projections, SART fails in reconstructing the CT slices. When using SOTA X-ray NeRF-", + "bbox": [ + 212, + 795, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Yuanhao Cai et al.", + "bbox": [ + 271, + 114, + 397, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/706e072ff6a9b0ec1c7bd9347ed5b5029314d4cc826ae064dbd819de5298a273.jpg", + "table_caption": [ + "Table 2: Results on sparse-view CT reconstruction. NeRF-based methods and our X-Gaussian are used to create novel views for FDK [16], SART [1], and ASD-POCS [45]." + ], + "table_footnote": [], + "table_body": "
Method Metric+ None+ InTomo [58]+ NeRF [37]+ TensorRF [9]+ NeAT [42]+ NAF [60]+ X-Gaussian
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
FDK7.410.09320.310.49820.570.50220.610.50120.940.51121.280.52322.600.584
Δ FDK0012.900.40513.160.40913.200.40813.520.41813.870.43015.190.491
SART17.240.52826.280.85926.780.85327.060.86727.310.86927.840.87930.250.907
Δ SART009.040.3319.540.3259.820.33910.070.34110.600.35113.010.379
ASD-POCS17.030.52525.440.84726.580.85726.930.86826.950.86527.910.88030.560.926
Δ ASD-POCS008.410.3229.550.3329.900.3439.920.34010.880.35513.530.401
", + "bbox": [ + 210, + 185, + 787, + 320 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "based methods to create novel views, SART produces over-smooth CT slices with blurry structural contents. On the contrary, when using our X-Gaussian to assist SART and ASD-POCS, they can reconstruct much clearer CT slices with more high-frequency textures and fine-grained structural details, such as the vessels in the chest (Fig. 5). These results clearly demonstrate the potential practical values of our method on the sparse-view CT reconstruction task.", + "bbox": [ + 212, + 348, + 787, + 441 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 464, + 388, + 479 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Break-down Ablation. We first conduct a break-down ablation experiment to study the effect of each proposed technique towards higher performance and faster speed. We adopt the original 3DGS [25] as the baseline model and naively average the RGB channels to represent the value of radiation intensity. The results are listed in Tab. 3a. The baseline model yields 37.21 dB PSNR in performance. Its average training time and inference speed are 31 min 38 s and 64 fps, respectively. We can observe from Tab. 3a: (i) When using ACUI to replace the time-consuming SfM [46] algorithm for initialization, the training time is significantly reduced by $34\\%$ while the performance yields an improvement of 1.66 dB. This evidence suggests that our ACUI strategy is much faster than the SfM [46] algorithm used in the original 3DGS and can compute more accurate camera calibration parameters for 3D Gaussians and subsequent rendering. (ii) Then we apply our radiative Gaussian point cloud model equipped with the proposed Differentiable Radiative Rasterization (DRR) to replace the original RGB Gaussian point cloud model and its RGB rasterization. As analyzed in Sec. 3.1 and compared in Fig. 3, the anisotropic spherical harmonics (SH) are not suitable for X-ray imaging because X-ray imaging based on penetration is isotropic. In contrast, our radiative Gaussian point cloud model can better fit the view-independent radiation intensity in 3D space. Therefore, the performance is significantly improved by 4.53 dB in PSNR. Besides, removing the computation related to the view direction from the forward and backward processes of rasterization can further accelerate the training and inference speed. Thus, the training time is reduced by $54.10\\%$ and the inference speed is $2.1\\times$ faster.", + "bbox": [ + 212, + 492, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Radiative Gaussian Splitting for Efficient X-ray Novel View Synthesis", + "bbox": [ + 263, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c93447777a2c2d8286aa2487c68acfc3a3cf1ca13ce3f955a41c7fa0aff3296f.jpg", + "image_caption": [ + "Fig. 5: Visual results of sparse-view CT reconstruction on the scenes of foot and chest. NeRF-based methods and our X-Gaussian are used to create novel views for SART [1]." + ], + "image_footnote": [], + "bbox": [ + 217, + 146, + 779, + 333 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Initialization of Point Position. We compare different initialization strategies for the center positions of Gaussian point clouds including random, spherical, FDK [16], and cubic initialization. To be specific, Random initialization means randomly sampling points within the scanned area in 3D space. Spherical initialization uniformly samples point positions within a sphere that can completely enclose the scanned object. FDK [16] initialization adopts the FDK algorithm to back-project the given projections into 3D point positions. Cuboid initialization is our ACUI. Please note that we keep the computed intrinsic and extrinsic matrices the same for fair comparison between different strategies. The results are reported in Tab. 3b. FDK initialization slightly outperforms our ACUI by 0.066 dB but its training time is $2.59 \\times$ longer and its inference speed is 55 fps slower than ACUI. This is because the back-projection in FDK is time-consuming and initializes redundant points. The random and spherical initialization strategies yield lower PSNR and slower speed than cubic initialization. To achieve a better trade-off, we adopt the cubic initialization, i.e., ACUI, which enjoys good performance, the cheapest training cost, and the fastest inference speed.", + "bbox": [ + 212, + 417, + 787, + 660 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Parameter Analysis. We conduct parameter analysis of the number of features $N_{f}$ and the sampling interval $d$ . The results are shown in Tab. 3c and Tab. 3d. In Tab. 3c, (i) When increasing $N_{f}$ , the performance gradually improves but the magnitude of the improvement decreases. In particular, $N_{f} = 32$ achieves the best results of 43.42 dB in PSNR. $N_{f} = 16$ achieves on-par results with $N_{f} = 32$ , only 0.013 dB lower. (ii) We notice that the training time and inference speed do not change monotonically. This is because the Gaussian point clouds with various feature dimensions have different representing ability and computational complexity. The number of final 3D Gaussians after training also varies substantially. When $N_{f} = 16$ , the training time reaches a local minimum", + "bbox": [ + 215, + 686, + 787, + 843 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Yuanhao Cai et al.", + "bbox": [ + 271, + 114, + 397, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/eec9fa853229df89eff6de68d849d125417ccdb8c06887d192d53a8920e61bd4.jpg", + "table_caption": [ + "Table 3: Ablation study. PSNR, SSIM, training time, and inference speed are reported." + ], + "table_footnote": [], + "table_body": "
Method3DGS [25]+ ACUI+ DRR
PSNR37.21338.87243.404
SSIM0.98130.98710.9993
Train time (s)18981172538
Infer speed (fps)6472148
", + "bbox": [ + 223, + 170, + 473, + 224 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/d02fcb290b344fdd5205d9e64e8862a0d954d5c48c51291a22c9f2bbe680b753.jpg", + "table_caption": [ + "(a) Break-down ablation study" + ], + "table_footnote": [], + "table_body": "
MethodRandomSphericalFDK [16]Cubic
PSNR41.32942.83743.47043.404
SSIM0.99420.99880.99930.9993
Train time (s)6015751394538
Infer speed (fps)11213693148
", + "bbox": [ + 478, + 171, + 777, + 224 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/bc82404cb21d2cde92219c7f0456db974f0c33102e1b9cbbbf58f65baf473fa6.jpg", + "table_caption": [ + "(b) Ablation of point position initialization" + ], + "table_footnote": [], + "table_body": "
Number12481632
PSNR38.81840.20542.13042.86843.40443.417
SSIM0.98400.99310.99830.99910.99930.9993
Train time (s)511525521553538752
Speed (fps)153152158127148101
", + "bbox": [ + 220, + 251, + 496, + 306 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/3bed1f177afb93c94a3c47dccd6eab0f21d3265717d1806bb68e9f4bfdaa5da4.jpg", + "table_caption": [ + "(c) Analysis of the number of features $N_{f}$" + ], + "table_footnote": [], + "table_body": "
Interval12481632
PSNR42.85342.97943.21543.40443.31143.294
SSIM0.99890.99900.99920.99930.99920.9992
Train time (s)785593545538534566
Speed (fps)869411414813597
", + "bbox": [ + 501, + 251, + 777, + 306 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "(d) Analysis of the initialized interval $d$", + "bbox": [ + 517, + 308, + 750, + 320 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "and the inference speed is at its local maximum. Hence, we eventually adopt $N_{f} = 16$ to reach a more optimal balance between performance and speed.", + "bbox": [ + 215, + 366, + 784, + 396 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Tab. 3d, the best reconstruction performance and the fastest inference speed are achieved at $d = 8$ . The training time (538 s) at $d = 8$ is almost the same as the shortest one (534 s) at $d = 16$ . Thus, we eventually set $d$ to 8.", + "bbox": [ + 212, + 398, + 784, + 443 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Convergence Analysis. We conduct two visual analyses to compare the convergence between our X-Gaussian and original 3DGS [25] in Fig. 1, and between X-Gaussian and the SOTA X-ray NeRF-based method NAF [60] in Fig. 6.", + "bbox": [ + 212, + 474, + 784, + 518 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Specifically, we adopt the same ACUI strategy for the original 3DGS to focus on comparing the Gaussian point cloud model and the rasterization. For fairness, we train 3DGS and our X-Gaussian on the scene of foot with the same settings and visualize the positions and opacities of Gaussian point clouds at the 100-th, 1000-th, 5000-th, 10000-th, and 20000-th iterations of the training process in Fig. 1. We also visualize the CT volume of foot as a reference. As can be seen that 3DGS with RGB rasterization converges slowly and suffers from more noisy point clouds. Plus, the final trained model of 3DGS at the 20000-th iteration contains more redundant Gaussians that are unnecessary to represent the 3D structure of the foot, which reduces the model's inference speed. In contrast, our X-Gaussian equipped with the proposed DRR shows faster and better convergence. In particular, as early as the 1000-th iteration, our radiative Gaussian point clouds have essentially formed the basic shape of the foot. Besides, the final trained X-Gaussian at the 20000-th iteration can better represent the 3D geometry and more accurate structural contents than the original 3DGS.", + "bbox": [ + 212, + 521, + 785, + 747 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Additionally, in Fig. 6, we visualize the rendered novel projections with the same angle $\\phi$ of NAF and our X-Gaussian at 20s, 60s, and 180s of the training process on the scene of chest. NAF produces blurry images with severe noises on the background regions within the first three minutes of the training. In contrast, our X-Gaussian can reconstruct clearer structural details like the ribs and blood vessels with cleaner background of the chest at the first minute of the training.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Radiative Gaussian Splitting for Efficient X-ray Novel View Synthesis", + "bbox": [ + 263, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0736b903be6058f5f5a2d54ceffdc55e2f223d8a0b811795ae658ac8ba443dea.jpg", + "image_caption": [ + "NAF at 20s" + ], + "image_footnote": [], + "bbox": [ + 217, + 154, + 295, + 214 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/80050c19b38b34b78e5727fbc48268c4df08f044a17e65cf254c034226472562.jpg", + "image_caption": [ + "NAF at 60s" + ], + "image_footnote": [], + "bbox": [ + 295, + 155, + 372, + 214 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/2fc4386cb4049a3985d704dfc2a99f5a40a441bcd58eb20415092ea092d1ca60.jpg", + "image_caption": [ + "NAF at 180s" + ], + "image_footnote": [], + "bbox": [ + 374, + 155, + 450, + 214 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/cb5fbba27a6b1712ba952f139e001024fe0a215cef1d0f59b41bf07f06b227d1.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 450, + 155, + 527, + 214 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/9bd14f35032f1bf570daab932c842dec9da505c590669d8829e8fd99e3e909b0.jpg", + "image_caption": [ + "X-Gaussian at 20s" + ], + "image_footnote": [], + "bbox": [ + 217, + 223, + 295, + 282 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/432a86bc726fa19eebd1e9632e74775fe693a390ef901c499e8a60eeafa95415.jpg", + "image_caption": [ + "X-Gaussian at 60s" + ], + "image_footnote": [], + "bbox": [ + 295, + 223, + 372, + 282 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/e017efe17023ff92f513f599338661ae4cfa4d56190a242a313309412bffc47f.jpg", + "image_caption": [ + "X-Gaussian at 180s" + ], + "image_footnote": [], + "bbox": [ + 374, + 223, + 450, + 282 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/35cf8f70b02f21ebca3633d16cbee4ca267e15cd7de87c303bc8921e2f16518c.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 450, + 223, + 527, + 282 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/f051d1c3945bd126fa7106fb93b0c540bc07e0069f5239e01f306ef58925e6b6.jpg", + "image_caption": [ + "Fig.6: Convergence analysis of NAF [60] vs. our X-Gaussian. We visualize the rendered projections at 20s, 60s, and 180s of training. Our X-Gaussian shows faster and better convergence.", + "Fig. 7: Analysis of covariance. The mean covariance of 3D Gaussians on different scenes decreases when the number of training views increases." + ], + "image_footnote": [], + "bbox": [ + 537, + 155, + 769, + 292 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Analysis of Covariance. We study how the shape of the 3D Gaussian point cloud changes with the number of training views in Fig. 7. As the number of training views increases, the mean covariance of 3D Gaussians that control the size of Gaussian point clouds decreases. This indicates that the 3D Gaussian point clouds gradually change from coarse to fine, thereby being more capable of representing fine-grained structures, such as a small tumor in the abdomen.", + "bbox": [ + 212, + 385, + 787, + 476 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Limitation", + "text_level": 1, + "bbox": [ + 214, + 497, + 354, + 513 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The main limitation of our method is that its memory cost is non-trivial, about 51.25 MB on average. It is challenging to deploy it on low-RAM mobile devices.", + "bbox": [ + 212, + 527, + 784, + 558 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 578, + 359, + 594 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we propose the first 3DGS-based framework, X-Gaussian, for X-ray novel view synthesis. Firstly, we design a radiative Gaussian point cloud model based on X-ray imaging properties. This model excludes the influence of view direction when fitting radiation intensity. For this model, we develop a GPU-friendly differentiable radiative rasterization CUDA kernel that renders projections at a faster speed than RGB rasterization. Secondly, we customize an initialization strategy, ACUI, that does not need to execute the SfM algorithm. Instead, ACUI uses the parameters of X-ray scanner to compute the extrinsic and intrinsic matrices, and then uniformly samples center points for 3D Gaussians within a cuboid enclosing the scanned object. Experiments demonstrate that our X-Gaussian significantly outperforms SOTA methods by over $6.5\\mathrm{dB}$ while enjoying $73\\times$ faster inference speed and only requiring $15\\%$ of training time.", + "bbox": [ + 212, + 607, + 787, + 790 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements: This work was supported by the Lustgarten Foundation for Pancreatic Cancer Research and the Patrick J. McGovern Foundation Award.", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Yuanhao Cai et al.", + "bbox": [ + 271, + 114, + 397, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Andersen, A.H., Kak, A.C.: Simultaneous algebraic reconstruction technique (sart): a superior implementation of the art algorithm. Ultrasonic imaging (1984)", + "2. Armato III, S.G., McLennan, G., Bidaut, L., McNitt-Gray, M.F., Meyer, C.R., Reeves, A.P., Zhao, B., Aberle, D.R., Henschke, C.I., Hoffman, E.A., et al.: The lung image database consortium (lidc) and image database resource initiative (idri): a completed reference database of lung nodules on ct scans. Medical physics (2011)", + "3. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: ICCV (2021)", + "4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In: CVPR (2022)", + "5. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-nerf: Anti-aliased grid-based neural radiance fields. In: ICCV (2023)", + "6. Biguri, A., Dosanjh, M., Hancock, S., Soleimani, M.: Tigre: a matlab-gpu toolbox for cbct image reconstruction. Biomedical Physics & Engineering Express (2016)", + "7. Boone, J., Shah, N., Nelson, T.: A comprehensive analysis of coefficients for pendant-geometry cone-beam breast computed tomography. Medical physics (2004)", + "8. Boone, J.M., Nelson, T.R., Lindfors, K.K., Seibert, J.A.: Dedicated breast ct: radiation dose and image quality evaluation. Radiology (2001)", + "9. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: ECCV (2022)", + "10. Chen, B., Ning, R.: Cone-beam volume ct breast imaging: Feasibility study. Medical physics (2002)", + "1. Chen, Z., Funkhouser, T., Hedman, P., Tagliasacchi, A.: Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In: CVPR (2023)", + "2. Cormack, A.M.: Representation of a function by its line integrals, with some radiological applications. Journal of applied physics (1963)", + "3. Cormack, A.M.: Representation of a function by its line integrals, with some radiological applications. ii. Journal of Applied Physics (1964)", + "4. Corona-Figueroa, A., Frawley, J., Bond-Taylor, S., Bethapudi, S., Shum, H.P., Willcocks, C.G.: Mednerf: Medical neural radiance fields for reconstructing 3d-aware ct-projections from a single x-ray. In: International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC) (2022)", + "5. Elbakri, I.A., Fessler, J.A.: Segmentation-free statistical image reconstruction for polyenergetic x-ray computed tomography with experimental validation. Physics in Medicine & Biology (2003)", + "6. Feldkamp, L.A., Davis, L.C., Kress, J.W.: Practical cone-beam algorithm. Josa a (1984)", + "7. Guide, D.: Cuda c programming guide. NVIDIA (2013)", + "8. Hounsfield, G.N.: Computerized transverse axial scanning (tomography): Part 1. description of system. The British journal of radiology (1973)", + "9. Hounsfield, G.N.: Computed medical imaging. Science (1980)", + "20. Hu, S., Liu, Z.: Gauhuman: Articulated gaussian splatting from monocular human videos. arXiv preprint arXiv: (2023)", + "21. Hu, T., Liu, S., Chen, Y., Shen, T., Jia, J.: Efficientnerf efficient neural radian fields. In: CVPR (2023)" + ], + "bbox": [ + 225, + 176, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Radiative Gaussian Splitting for Efficient X-ray Novel View Synthesis", + "bbox": [ + 263, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "22. Hu, W., Wang, Y., Ma, L., Yang, B., Gao, L., Liu, X., Ma, Y.: Tri-miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields. In: ICCV (2023)", + "23. Jiang, Y., Tu, J., Liu, Y., Gao, X., Long, X., Wang, W., Ma, Y.: Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. arXiv preprint arXiv:2311.17977 (2023)", + "24. Keetha, N., Karhade, J., Jatavallabhula, K.M., Yang, G., Scherer, S., Ramanan, D., Luiten, J.: Splatam: Splat, track & map 3d gaussians for dense rgb-d slam. arXiv preprint arXiv:2312.02126 (2023)", + "25. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (2023)", + "26. Kingma, D.P., Ba, J.L.: Adam: A method for stochastic optimization. In: ICLR (2015)", + "27. Klacansky, P.: Scientific visualization datasets (2022), https://klacansky.com/open-scivis-datasets/", + "28. Kocabas, M., Chang, J.H.R., Gabriel, J., Tuzel, O., Ranjan, A.: Hugs: Human gaussian splats. arXiv preprint arXiv:2311.17910 (2023)", + "29. Kopanas, G., Philip, J., Leimkuhler, T., Drettakis, G.: Point-based neural rendering with per-view optimization. In: Computer Graphics Forum (2021)", + "30. Li, R., Gao, H., Tancik, M., Kanazawa, A.: Nerfacc: Efficient sampling accelerates nerfs. In: ICCV (2023)", + "31. Liang, Y., Yang, X., Lin, J., Li, H., Xu, X., Chen, Y.: Luciddreamer: Towards high-fidelity text-to-3d generation via interval score matching. arXiv preprint arXiv:2311.11284 (2023)", + "32. Liang, Z., Zhang, Q., Feng, Y., Shan, Y., Jia, K.: Gs-ir: 3d gaussian splatting for inverse rendering. arXiv preprint arXiv:2311.16473 (2023)", + "33. Liu, X., Zhan, X., Tang, J., Shan, Y., Zeng, G., Lin, D., Liu, X., Liu, Z.: Human-gaussian: Text-driven 3d human generation with gaussian splatting. arXiv preprint arXiv:2311.17061 (2023)", + "34. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713 (2023)", + "35. Manglos, S.H., Gagne, G.M., Krol, A., Thomas, F.D., Narayanaswamy, R.: Transmission maximum-likelihood reconstruction with ordered subsets for cone beam ct. Physics in Medicine & Biology (1995)", + "36. Matsuki, H., Murai, R., Kelly, P.H., Davison, A.J.: Gaussian splatting slam. arXiv preprint arXiv:2312.06741 (2023)", + "37. Mildenhall, B., Srinivasan, P., Tancik, M., Barron, J., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)", + "38. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM ToG (2022)", + "39. Pan, J., Zhou, T., Han, Y., Jiang, M., et al.: Variable weighted ordered subset image reconstruction algorithm. International Journal of Biomedical Imaging (2006)", + "40. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: NeurIPS (2019)", + "41. Reiser, C., Szeliski, R., Verbin, D., Srinivasan, P., Mildenhall, B., Geiger, A., Barron, J., Hedman, P.: Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. TOG (2023)", + "42. Rückert, D., Wang, Y., Li, R., Idoughi, R., Heidrich, W.: Neat: Neural adaptive tomography. TOG (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Yuanhao Cai et al.", + "bbox": [ + 271, + 114, + 397, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "43. Sauer, K., Bouman, C.: A local update strategy for iterative reconstruction from projections. TIP (1993)", + "44. Scarfe, W.C., Farman, A.G., Sukovic, P., et al.: Clinical applications of cone-beam computed tomography in dental practice. Journal-Canadian Dental Association (2006)", + "45. Sidky, E.Y., Pan, X.: Image reconstruction in circular cone-beam computed tomography by constrained, total-variation minimization. Physics in Medicine & Biology (2008)", + "46. Snavely, N., Seitz, S.M., Szeliski, R.: Photo tourism: exploring photo collections in 3d. In: SIGGRAPH (2006)", + "47. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)", + "48. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: CVPR (2022)", + "49. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncell, E.P.: Image quality assessment: from error visibility to structural similarity. TIP (2004)", + "50. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528 (2023)", + "51. Xie, T., Zong, Z., Qiu, Y., Li, X., Feng, Y., Yang, Y., Jiang, C.: Physgaus-sian: Physics-integrated 3d gaussians for generative dynamics. arXiv preprint arXiv:2311.12198 (2023)", + "52. Yan, C., Qu, D., Wang, D., Xu, D., Wang, Z., Zhao, B., Li, X.: Gs-slam: Dense visual slam with 3d gaussian splatting. arXiv preprint arXiv:2311.11700 (2023)", + "53. Yang, Z., Yang, H., Pan, Z., Zhu, X., Zhang, L.: Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642 (2023)", + "54. Yariv, L., Hedman, P., Reiser, C., Verbin, D., Srinivasan, P.P., Szeliski, R., Barron, J.T., Mildenhall, B.: Bakedsdf: Meshing neural sdfs for real-time view synthesis. In: SIGGRAPH (2023)", + "55. Yi, T., Fang, J., Wu, G., Xie, L., Zhang, X., Liu, W., Tian, Q., Wang, X.: Gaussian dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529 (2023)", + "56. Yu, L., Zou, Y., Sidky, E.Y., Pelizzari, C.A., Munro, P., Pan, X.: Region of interest reconstruction from truncated data in circular cone-beam ct. TMI (2006)", + "57. Yugay, V., Li, Y., Gevers, T., Oswald, M.R.: Gaussian-slam: Photo-realistic dense slam with gaussian splatting. arXiv preprint arXiv:2312.10070 (2023)", + "58. Zang, G., Idoughi, R., Li, R., Wonka, P., Heidrich, W.: Intratomo: self-supervised learning-based tomography via sinogram synthesis and prediction. In: CVPR (2021)", + "59. Zbijewski, W., Defrise, M., Viergever, M.A., Beekman, F.J.: Statistical reconstruction for x-ray ct systems with non-continuous detectors. Physics in Medicine & Biology (2006)", + "60. Zha, R., Zhang, Y., Li, H.: Naf: neural attenuation fields for sparse-view cbct reconstruction. In: MICCAI (2022)", + "61. Zhang, T., Yu, H.X., Wu, R., Feng, B.Y., Zheng, C., Snavely, N., Wu, J., Freeman, W.T.: PhysDreamer: Physics-based interaction with 3d objects via video generation. In: ECCV (2024)", + "62. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: Ewa volume splatting. In: Proceedings Visualization, 2001. VIS'01. IEEE (2001)" + ], + "bbox": [ + 215, + 147, + 784, + 832 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Radiative Gaussian Splitting for Efficient X-ray Novel View Synthesis", + "bbox": [ + 264, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis/f615a129-2ee2-41a8-9f1c-82a796edae0e_model.json b/2024/Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis/f615a129-2ee2-41a8-9f1c-82a796edae0e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2105ff15d0b6e1403d3e3eba2b16c81825336e20 --- /dev/null +++ b/2024/Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis/f615a129-2ee2-41a8-9f1c-82a796edae0e_model.json @@ -0,0 +1,2632 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.271, + 0.142, + 0.734, + 0.187 + ], + "angle": 0, + "content": "Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.212, + 0.761, + 0.245 + ], + "angle": 0, + "content": "Yuanhao Cai \\(^{1}\\), Yixun Liang \\(^{2}\\), Jiahao Wang \\(^{1}\\), Angtian Wang \\(^{1}\\), Yulun Zhang \\(^{3,\\dagger}\\), Xiaokang Yang \\(^{3}\\), Zongwei Zhou \\(^{1,\\dagger}\\), and Alan Yuille \\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.254, + 0.761, + 0.271 + ], + "angle": 0, + "content": "1 Johns Hopkins University, 2 HKUST(GZ), 3 Shanghai Jiao Tong University" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.301, + 0.744, + 0.552 + ], + "angle": 0, + "content": "Abstract. X-ray is widely applied for transmission imaging due to its stronger penetration than natural light. When rendering novel view X-ray projections, existing methods mainly based on NeRF suffer from long training time and slow inference speed. In this paper, we propose a 3D Gaussian splatting-based method, namely X-Gaussian, for X-ray novel view synthesis. Firstly, we redesign a radiative Gaussian point cloud model inspired by the isotropic nature of X-ray imaging. Our model excludes the influence of view direction when learning to predict the radiation intensity of 3D points. Based on this model, we develop a Differentiable Radiative Rasterization (DRR) with CUDA implementation. Secondly, we customize an Angle-pose Cuboid Uniform Initialization (ACUI) strategy that directly uses the parameters of the X-ray scanner to compute the camera information and then uniformly samples point positions within a cuboid enclosing the scanned object. Experiments show that our X-Gaussian outperforms state-of-the-art methods by 6.5 dB while enjoying less than \\(15\\%\\) training time and over \\(73\\times\\) inference speed. The application on CT reconstruction also reveals the practical values of our method. Code is at https://github.com/caiyuanhao1998/X-Gaussian" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.564, + 0.721, + 0.579 + ], + "angle": 0, + "content": "Keywords: Medical Imaging \\(\\cdot\\) View Synthesis \\(\\cdot\\) Gaussian Splatting" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.622, + 0.377, + 0.637 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.651, + 0.788, + 0.819 + ], + "angle": 0, + "content": "X-ray novel view synthesis (NVS) aims to create X-ray projections of an object from new viewpoints that are not originally captured, using only existing projections scanned from different view directions. As we know, X-ray has stronger penetrating power to capture internal structures of imaged objects and is therefore widely applied in medical imaging [12,13,15,18,19,43]. Nonetheless, X-ray is harmful to human body due to its powerful ionizing radiation, especially when the dose of X-ray increases. Improving NVS techniques can help reduce the exposure to X-rays and provide comprehensive viewpoints of imaged parts for doctors and downstream tasks such as sparse-view CT reconstruction. Hence, X-ray NVS is very important and valuable. We study this task in the circular cone beam X-ray scanning scenario [7,8,10,35,39,44,56,59]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.827, + 0.416, + 0.841 + ], + "angle": 0, + "content": "\\(\\dagger =\\) corresponding authors." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.398, + 0.127 + ], + "angle": 0, + "content": "Yuanhao Cai et al." + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.147, + 0.778, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.338, + 0.788, + 0.409 + ], + "angle": 0, + "content": "Fig. 1: Point cloud visualization of the original 3DGS [25] (top) and our X-Gaussian (bottom). We visualize the positions and opacities of the Gaussian point clouds at different training iterations. We also visualize the volume of foot as a reference. Note that the volume is not the ground truth of point clouds. Our X-Gaussian can better represent the detailed structures than 3DGS, showing faster and better convergence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.788, + 0.581 + ], + "angle": 0, + "content": "Existing methods are mainly based on neural radiance fields (NeRF) [37]. They usually employ a multi-layer perceptron (MLP) to learn the mapping from the point position to the radiodensity and then create projections by volume rendering along rays. This ray tracing scheme is time-consuming because it needs to sample many 3D points and then compute them for every single ray, slowing down the training and inference processes. Even the recent most efficient NeRF-based method [60] still requires over an hour for training and yields suboptimal results at a slow inference speed of 2 fps. This increases the waiting time of patients and doctors, leading to low diagnostic efficiency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.844 + ], + "angle": 0, + "content": "Recently, 3D Gaussian splattering (3DGS) [25] has demonstrated promising reconstruction quality while enjoying much faster inference speed than NeRF-based algorithms in RGB domain, which motivates us to follow this technical route. However, due to the fundamental differences between X-ray and natural light imaging, directly applying the original 3DGS to X-ray NVS may encounter two issues. Firstly, the spherical harmonics (SH) in RGB 3DGS is not suitable for modeling the X-ray radiation intensity of 3D points. Specifically, natural light imaging relies on the reflection off the surface. The color of a 3D point is anisotropic and view-dependent. Based on this nature, the original Gaussian point cloud model uses SH to fit the illumination distribution. In contrast, X-rays penetrate the object and attenuate, thereby forming an image. Given specific X-rays, the radiation intensity of a 3D point depends on its radiodensity and is independent to the view direction, which means the point radiation intensity is isotropic. Secondly, the original point cloud initialization algorithm, structure-from-motion (SfM) [46], is also not suitable for X-ray imaging. Compared to RGB images, X-ray images are grayscale and their contrast is lower. Additionally, different layers of an object may overlap on the same position of the projection" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.264, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Radiative Gaussian Splitting for Efficient X-ray Novel View Synthesis" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "due to the transmission imaging nature of X-rays. These two problems degrade the accuracy of feature detection and matching in SfM. Meanwhile, running SfM is time-consuming, which prolongs the training process of Gaussian point clouds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.788, + 0.479 + ], + "angle": 0, + "content": "To address the above issues, we propose a novel 3DGS-based method, X-Gaussian, for X-ray NVs. Our X-Gaussian composes two key techniques. Firstly, we redesign a radiative Gaussian point cloud model inspired by the isotropic property of X-ray imaging. We present a Radiation Intensity Response Function (RIRF) to replace the SH function of the original 3DGS. Different from SH, our RIRF excludes the influence of view direction. To this end, it adopts the inner product between a learnable vector representing the inherent point features and a set of basis weights to fit the radiation intensity of a 3D point. Based on this point cloud model, we further develop a Differentiable Radiative Rasterization (DRR) with a CUDA implementation to render novel projections. Secondly, we customize an Angle-pose Cuboid Uniform Initialization (ACUI) strategy for camera calibration parameters and Gaussian point clouds. Our ACUI first exploits the parameters of the X-ray scanner to compute the intrinsic and extrinsic matrices. Then we set up a cuboid that can completely enclose the scanned object. Within this cuboid, we uniformly sample 3D points at intervals to initialize the center positions of the Gaussian point clouds. Free from running the SfM algorithm, our ACUI significantly reduces the training time. Equipped with the two proposed techniques, our X-Gaussian outperforms SOTA methods by 6.5 dB while enjoying \\(73 \\times\\) inference speed and \\(7 \\times\\) training speed." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.48, + 0.724, + 0.494 + ], + "angle": 0, + "content": "The main contributions of this work can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.504, + 0.784, + 0.548 + ], + "angle": 0, + "content": "- We propose a novel 3D Gaussian splatting-based framework, X-Gaussian, for X-ray novel view synthesis. To our knowledge, this is the first attempt to explore the potential of Gaussian splatting in X-ray neural rendering." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.55, + 0.784, + 0.579 + ], + "angle": 0, + "content": "- We design a radiative Gaussian point cloud model with a differentiable radiative rasterization based on the isotropic nature of X-ray imaging." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.58, + 0.784, + 0.609 + ], + "angle": 0, + "content": "- We present an angle-pose cuboid uniform initialization strategy for Gaussian point clouds and camera calibration in circular cone beam X-ray scanning." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.61, + 0.784, + 0.654 + ], + "angle": 0, + "content": "- Our X-Gaussian significantly outperforms SOTA NeRF-based methods with much faster speed. Experiments also show that our method can improve the performance of sparse-view CT reconstruction, showing its practical values." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.504, + 0.784, + 0.654 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.678, + 0.388, + 0.695 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.71, + 0.449, + 0.724 + ], + "angle": 0, + "content": "2.1 Neural Radiance Field" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "NeRF [37] learns an implicit neural scene representation of color and volume density, given the position of a 3D point and view direction. It has achieved great success in NVS and inspired an explosion of follow-up papers to improve its quality [3-5,22,48] and speed [9,11,21,30,38,41,54]. For example, Instant-NGP [38] adopts hash tables as the encoder to allow small MLP for fast training and inference. Some later works extend the application domain of NeRF from natural light to X-rays [14,58,60]. For instance, NAF [60] follows the settings" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.398, + 0.127 + ], + "angle": 0, + "content": "Yuanhao Cai et al." + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.149, + 0.784, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.318, + 0.788, + 0.39 + ], + "angle": 0, + "content": "Fig. 2: Pipeline of our method. (a) Angle-posed Cuboid Uniform Initialization (ACUI) strategy uses the parameters of X-ray scanner to compute intrinsic and extrinsic matrices, and samples center points for 3D Gaussians. (b) Our radiative Gaussian point cloud model learns to predict the radiation intensity of 3D points. (c) Based on our Gaussian model, we develop a GPU-friendly Differentiable Radiative Rasterization (DRR)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.788, + 0.466 + ], + "angle": 0, + "content": "of Instant-NGP to learn the implicit mapping from 3D position to attenuation. Yet, the ray tracing and volume rendering schemes are time-consuming, which limits the training and inference speed of NeRF-based X-ray NVS algorithms." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.487, + 0.42, + 0.503 + ], + "angle": 0, + "content": "2.2 Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.512, + 0.788, + 0.65 + ], + "angle": 0, + "content": "3DGS [25] represents scenes using millions of 3D Gaussian point clouds. This approach is fundamentally different from NeRF-based algorithms by employing an explicit representation coupled with highly parallelized rasterization workflows. These features enable more efficient computation and rendering processes. Hence, 3DGS has achieved great success in several fields, including 3D Generation [20,28,31,33,47,55,61], Dynamic Scene Modeling [34,50,53], SLAM [24,36,52,57], Inverse Rendering [23,32,51], etc. However, most applications of 3DGS are focused on natural scenes with RGB colors. The potential of 3DGS in X-ray imaging still remains under-explored. Our goal is to fill this research gap." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.672, + 0.331, + 0.687 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The pipeline of our X-Gaussian is shown in Fig. 2. Firstly, we design an Angle-posed Cuboid Uniform Initialization (ACUI) to compute the intrinsic and extrinsic matrices from the parameters of X-ray scanner, as illustrated in Fig. 2 (a). Then ACUI uniformly samples 3D points within a cuboid that can completely enclose the scanned object to initialize the center positions of our radiative Gaussian point clouds in Fig. 2 (b). Given a view direction, the 3D point clouds undergo our Differentiable Radiative Rasterization (DRR) to derive the rendered image, as depicted in Fig. 2 (c). In this section, we will introduce our radiative Gaussian point cloud model and DRR processing first and then the ACUI strategy." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.264, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Radiative Gaussian Splitting for Efficient X-ray Novel View Synthesis" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.149, + 0.481, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.27, + 0.267, + 0.436, + 0.278 + ], + "angle": 0, + "content": "(a) Original RGB Gaussian Point Cloud" + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.148, + 0.779, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.267, + 0.738, + 0.278 + ], + "angle": 0, + "content": "(b) Our Radiative Gaussian Point Cloud" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.31, + 0.785, + 0.38 + ], + "angle": 0, + "content": "Fig. 3: Comparison between the Gaussian point cloud models of the original 3DGS and our X-Gaussian. (a) The original RGB Gaussian point cloud model uses spherical harmonics (SH) to simulate the anisotropic natural light distribution and view-dependent color. (b) Our radiative Gaussian point cloud model employs the weighted sum of point features to fit the isotropic X-ray penetration and view-independent radiation intensity." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.41, + 0.586, + 0.423 + ], + "angle": 0, + "content": "3.1 Radiative Gaussian Point Cloud Model" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.434, + 0.75, + 0.449 + ], + "angle": 0, + "content": "An object can be represented by a set of basic Gaussian point clouds \\(\\mathcal{G}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.36, + 0.461, + 0.785, + 0.478 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} = \\left\\{G _ {i} \\left(\\boldsymbol {\\mu} _ {i}, \\boldsymbol {\\Sigma} _ {i}, \\alpha_ {i}\\right) \\mid i = 1, 2, \\dots , N _ {p} \\right\\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.488, + 0.784, + 0.578 + ], + "angle": 0, + "content": "where \\(G_{i}\\) refers to the \\(i\\)-th Gaussian point cloud. Its center position, covariance, and opacity are defined as \\(\\pmb{\\mu}_{i} \\in \\mathbb{R}^{3}\\), \\(\\pmb{\\Sigma}_{i} \\in \\mathbb{R}^{3 \\times 3}\\), and \\(\\alpha_{i} \\in \\mathbb{R}\\). \\(\\pmb{\\Sigma}_{i}\\) is represented by a rotation matrix \\(\\mathbf{R}_i \\in \\mathbb{R}^3\\) and a scaling matrix \\(\\mathbf{S}_i \\in \\mathbb{R}^3\\) as \\(\\pmb{\\Sigma}_{i} = \\mathbf{R}_{i}\\mathbf{S}_{i}\\mathbf{S}_{i}^{\\top}\\mathbf{R}_{i}^{\\top}\\). \\(\\pmb{\\mu}_{i}, \\pmb{\\Sigma}_{i}, \\alpha_{i}, \\mathbf{R}_{i}\\), and \\(\\mathbf{S}_i\\) are learnable parameters. Besides these basic attributes, each Gaussian point cloud also employs additional learnable parameters to fit different imaging scenarios, e.g., natural light imaging and X-ray imaging." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.579, + 0.785, + 0.685 + ], + "angle": 0, + "content": "We first review the original RGB Gaussian point cloud model [25] in natural light imaging. As shown in Fig. 3 (a), the color of a 3D point is represented by spherical harmonics (SH). The point color is anisotropic and changes with the view direction. Each Gaussian point cloud learns to predict the SH coefficients \\(\\mathbf{k} = \\{k_l^m |0\\leq l\\leq L, - l\\leq m\\leq l\\} \\in \\mathbb{R}^{(L + 1)^2\\times 3}\\), where each \\(k_{l}^{m}\\in \\mathbb{R}^{3}\\) is a set of 3 coefficients corresponding to the RGB components. \\(L\\) is the degree of SH. Then the point color \\(\\mathbf{c}\\in \\mathbb{R}^3\\) at the view direction \\(\\mathbf{d} = (\\theta ,\\phi)\\) is derived by" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.696, + 0.785, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} (\\mathbf {d}, \\mathbf {k}) = \\sum_ {l = 0} ^ {L} \\sum_ {m = - l} ^ {l} k _ {l} ^ {m} Y _ {l} ^ {m} (\\theta , \\phi), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.75, + 0.784, + 0.779 + ], + "angle": 0, + "content": "where \\( Y_{l}^{m}:\\mathbb{S}^{2}\\to \\mathbb{R} \\) is the SH function that maps points on the sphere to real numbers. Please refer to the supplementary for its detailed formulation." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Although 3DGS [25] achieves fast inference speed and good performance in natural light imaging, the RGB Gaussian point cloud model is not suitable for X-ray scenarios due to the fundamental differences between natural light imaging and X-ray imaging. Natural light imaging relies on the reflection off the surface of" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.398, + 0.127 + ], + "angle": 0, + "content": "Yuanhao Cai et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.223 + ], + "angle": 0, + "content": "object. The anisotropic color modeled by SH is view-dependent. e.g., in Fig. 3 (a), the point color is blue from the left viewpoint and green from the right viewpoint. In contrast, X-ray imaging is based on the attenuation when penetrating the object. The degree of attenuation depends on the isotropic radiodensity property. Thus, the radiation intensity of a 3D point is view-independent." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.785, + 0.329 + ], + "angle": 0, + "content": "In light of the above analysis, we redesign our radiative Gaussian point cloud model. Different from the original 3DGS that uses SH to fit the color information for each point, our model introduces a Radiation Intensity Response Function (RIRF) to predict the radiation intensity of the 3D point. As illustrated in Fig. 3 (b), each Gaussian point cloud learns a feature vector \\(\\mathbf{f} \\in \\mathbb{R}^{N_f}\\) to represent its inherent radiative properties. Subsequently, the radiation intensity \\(\\mathbf{i} \\in \\mathbb{R}\\) of the center point of a 3D Gaussian at any view direction is modeled by RIRF as" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.34, + 0.785, + 0.355 + ], + "angle": 0, + "content": "\\[\n\\mathbf {i} (\\mathbf {f}) = \\operatorname {R I R F} (\\mathbf {f}) = \\operatorname {S i g m o i d} (\\boldsymbol {\\lambda} \\odot \\mathbf {f}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.365, + 0.784, + 0.411 + ], + "angle": 0, + "content": "where the Sigmoid function activates and normalizes the radiation intensity. \\(\\lambda \\in \\mathbb{R}^{N_f}\\) is a set of constant weights controlling the importance of each component of \\(\\mathbf{f}\\). Then the set of our radiative Gaussian point clouds \\(\\mathcal{G}_x\\) is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.353, + 0.422, + 0.785, + 0.437 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {x} = \\left\\{G _ {i} \\left(\\boldsymbol {\\mu} _ {i}, \\boldsymbol {\\Sigma} _ {i}, \\alpha_ {i}, \\mathbf {f} _ {i}\\right) \\mid i = 1, 2, \\dots , N _ {p} \\right\\}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.447, + 0.784, + 0.523 + ], + "angle": 0, + "content": "where \\(\\mathbf{f}_i\\in \\mathbb{R}^{N_f}\\) denotes the feature vector of the \\(i\\)-th Gaussian point cloud. Please note that Eq. (3) excludes the influence of the view direction \\(\\mathbf{d} = (\\theta ,\\phi)\\), which matches the isotropic nature of X-ray imaging. Meanwhile, Eq. (3) is free from the complex computation of SH function. Hence, the forward and backward processes of our X-Gaussian are much faster than those of the original 3DGS." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.544, + 0.574, + 0.558 + ], + "angle": 0, + "content": "3.2 Differentiable Radiative Rasterization" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.568, + 0.784, + 0.613 + ], + "angle": 0, + "content": "Based on our radiative Gaussian point cloud, we develop a Differentiable Radiative Rasterization (DRR), as shown in Fig. 2 (c). We first summarize the overall DRR processing \\( F_{\\mathrm{DRR}} \\) and then describe its details. DRR is represented as" + }, + { + "type": "equation", + "bbox": [ + 0.288, + 0.625, + 0.785, + 0.641 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} = F _ {\\mathrm {D R R}} \\left(\\mathbf {M} _ {e x t}, \\mathbf {M} _ {i n t}, \\left\\{G _ {i} \\left(\\boldsymbol {\\mu} _ {i}, \\boldsymbol {\\Sigma} _ {i}, \\alpha_ {i}, \\mathbf {f} _ {i}\\right) \\mid i = 1, 2, \\dots , N _ {p} \\right\\}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.65, + 0.784, + 0.711 + ], + "angle": 0, + "content": "where \\(\\mathbf{I} \\in \\mathbb{R}^{H \\times W}\\) denotes the rendered image, \\(\\mathbf{M}_{ext} \\in \\mathbb{R}^{4 \\times 4}\\) represents the extrinsic matrix, and \\(\\mathbf{M}_{int} \\in \\mathbb{R}^{4 \\times 3}\\) refers to the intrinsic matrix. Subsequently, we introduce the details of \\(F_{\\mathrm{DRR}}\\). To begin with, the possibility value of the \\(i\\)-th Gaussian distribution at the 3D point position \\(\\mathbf{x} \\in \\mathbb{R}^3\\) is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.332, + 0.722, + 0.785, + 0.748 + ], + "angle": 0, + "content": "\\[\nP (\\mathbf {x} \\mid \\boldsymbol {\\mu} _ {i}, \\boldsymbol {\\Sigma} _ {i}) = \\exp \\left(- \\frac {1}{2} (\\mathbf {x} - \\boldsymbol {\\mu} _ {i}) ^ {\\top} \\boldsymbol {\\Sigma} _ {i} ^ {- 1} (\\mathbf {x} - \\boldsymbol {\\mu} _ {i})\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.756, + 0.784, + 0.802 + ], + "angle": 0, + "content": "Then we project the 3D Gaussians to the 2D detector plane for subsequent rendering. \\(\\pmb{\\mu}_i\\) is firstly transferred from the world coordinate system to the camera coordinate system and then projected to the image coordinate system as" + }, + { + "type": "equation", + "bbox": [ + 0.227, + 0.812, + 0.784, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\widetilde {\\mathbf {t}} _ {i} = \\left[ \\begin{array}{l} \\mathbf {t} _ {i} \\\\ 1 \\end{array} \\right] = \\mathbf {M} _ {e x t} \\widetilde {\\boldsymbol {\\mu}} _ {i} = \\mathbf {M} _ {e x t} \\left[ \\begin{array}{l} \\boldsymbol {\\mu} _ {i} \\\\ 1 \\end{array} \\right], \\quad \\widetilde {\\mathbf {u}} _ {i} = \\left[ \\begin{array}{l} \\mathbf {u} _ {i} \\\\ 1 \\end{array} \\right] = \\mathbf {M} _ {i n t} \\widetilde {\\mathbf {t}} _ {i} = \\mathbf {M} _ {i n t} \\left[ \\begin{array}{l} \\mathbf {t} _ {i} \\\\ 1 \\end{array} \\right], \\tag {7}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.264, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Radiative Gaussian Splitting for Efficient X-ray Novel View Synthesis" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.146, + 0.788, + 0.21 + ], + "angle": 0, + "content": "where \\(\\mathbf{t}_i = (t_x, t_y, t_z) \\in \\mathbb{R}^3\\) is the camera coordinate of \\(\\pmb{\\mu}_i\\) and \\(\\mathbf{u}_i \\in \\mathbb{R}^2\\) is the image coordinate of \\(\\pmb{\\mu}_i\\). \\(\\widetilde{\\mathbf{u}}_i\\), \\(\\widetilde{\\mathbf{t}}_i\\), and \\(\\widetilde{\\pmb{\\mu}}_i\\) are the homogeneous coordinates of \\(\\mathbf{u}_i\\), \\(\\mathbf{t}_i\\), and \\(\\pmb{\\mu}_i\\), respectively. Subsequently, we transfer the 3D covariance matrix \\(\\pmb{\\Sigma}_i\\) to its counterpart \\(\\pmb{\\Sigma}_i' \\in \\mathbb{R}^{3 \\times 3}\\) in the camera coordinate system as" + }, + { + "type": "equation", + "bbox": [ + 0.425, + 0.222, + 0.786, + 0.241 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} _ {i} ^ {\\prime} = \\mathbf {J} _ {i} \\mathbf {W} _ {i} \\boldsymbol {\\Sigma} _ {i} \\mathbf {W} _ {i} ^ {\\top} \\mathbf {J} _ {i} ^ {\\top}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.25, + 0.786, + 0.281 + ], + "angle": 0, + "content": "where \\(\\mathbf{J}_i\\in \\mathbb{R}^{3\\times 3}\\) is the Jacobian of the affine approximation of the projective transformation. \\(\\mathbf{W}_i\\in \\mathbb{R}^{3\\times 3}\\) is the viewing transformation. We derive them by" + }, + { + "type": "equation", + "bbox": [ + 0.285, + 0.293, + 0.786, + 0.348 + ], + "angle": 0, + "content": "\\[\n\\mathbf {J} _ {i} = \\left[ \\begin{array}{c c c} \\frac {L _ {S D}}{t _ {z}} & 0 & - \\frac {L _ {S D}}{t _ {z} ^ {2}} \\frac {t _ {x}}{t _ {z}} \\\\ 0 & \\frac {L _ {S D}}{t _ {z}} & - \\frac {L _ {S D}}{t _ {z} ^ {2}} \\frac {t _ {y}}{t _ {z}} \\\\ 0 & 0 & 0 \\end{array} \\right], \\quad \\mathbf {W} _ {i} = \\left[ \\begin{array}{c c c} - \\sin \\phi & \\cos \\phi & 0 \\\\ 0 & 0 & - 1 \\\\ - \\cos \\phi & - \\sin \\phi & 0 \\end{array} \\right], \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.357, + 0.788, + 0.48 + ], + "angle": 0, + "content": "where \\(L_{SD}\\) represents the distance between the X-ray source and detector. \\(\\phi\\) refers to the azimuth angle of the source. Following [25,29,62], we obtain the 2D covariance matrix \\(\\pmb{\\Sigma}_i^{\\prime \\prime} \\in \\mathbb{R}^{2 \\times 2}\\) by skipping the third row and column of \\(\\pmb{\\Sigma}_i^{\\prime}\\). Then the 2D projection is partitioned into non-overlapping tiles. The 3D Gaussians \\((\\pmb{\\mu}_i, \\pmb{\\Sigma}_i)\\) are assigned to different tiles according to their 2D projections \\((\\pmb{u}_i, \\pmb{\\Sigma}_i^{\\prime \\prime})\\), as shown in the left image of Fig. 2 (c). These 3D Gaussians are sorted by the distances to the 2D detector. Then the intensity \\(\\mathbf{I}(p) \\in \\mathbb{R}\\) at pixel \\(p\\) is obtained by blending \\(\\mathcal{N}\\) ordered points overlapping the pixel in the corresponding tile as" + }, + { + "type": "equation", + "bbox": [ + 0.312, + 0.492, + 0.786, + 0.533 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} (p) = \\sum_ {j \\in \\mathcal {N}} \\mathbf {i} _ {j} \\sigma_ {j} \\prod_ {k = 1} ^ {j - 1} (1 - \\sigma_ {k}), \\quad \\sigma_ {j} = \\alpha_ {j} P (\\mathbf {x} _ {j} | \\boldsymbol {\\mu} _ {j}, \\boldsymbol {\\Sigma} _ {j}), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.544, + 0.786, + 0.574 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_j\\) is the \\(j\\)-th intersection 3D point of the X-ray landing on pixel \\(p\\) and the Gaussian point clouds in 3D space. \\(\\mathbf{i}_j\\) is the radiation intensity of \\(\\mathbf{x}_j\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.574, + 0.786, + 0.604 + ], + "angle": 0, + "content": "Optimization. Eventually, the training objective \\(\\mathcal{L}\\) is the weighted sum of \\(\\mathcal{L}_1\\) loss and SSIM loss between the rendered and ground-truth projection images as" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.621, + 0.786, + 0.636 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = (1 - \\gamma) \\mathcal {L} _ {1} + \\gamma \\mathcal {L} _ {\\text {S S I M}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.642, + 0.786, + 0.702 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is a hyperparameter balancing the importances of the two loss terms. By minimizing Eq. (11), we can optimize the attributes of 3D Gaussians, i.e., \\(\\pmb{\\mu}_i, \\pmb{\\Sigma}_i, \\alpha_i\\), and \\(\\mathbf{f}_i\\) in Eq. (5). \\(N_p\\) is adjusted by the adaptive control [25]. The optimization process is visualized in Fig. 1 and the video file in supplementary." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.703, + 0.786, + 0.748 + ], + "angle": 0, + "content": "Compared to the RGB rasterization in 3DGS [25], our DRR avoids the complex computations related to the view direction in the forward and backward processes, thereby enjoying cheaper training costs and faster inference speed." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.77, + 0.607, + 0.785 + ], + "angle": 0, + "content": "3.3 Angle-pose Cuboid Uniform Initialization" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "At the beginning of training, we need to initialize the parameters in Eq. (5) for rasterization. Specifically, \\(\\mathbf{\\Sigma}_i\\), \\(\\alpha_i\\), and \\(\\mathbf{f}_i\\) are randomly initialized. In natural light imaging, the original 3DGS [25] adopts the SfM [46] algorithm to compute the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.399, + 0.128 + ], + "angle": 0, + "content": "Yuanhao Cai et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.238 + ], + "angle": 0, + "content": "initial \\(\\pmb{\\mu}_i\\), \\(N_p\\), \\(\\mathbf{M}_{ext}\\), and \\(\\mathbf{M}_{int}\\). SfM detects and matches features from multiview images. It is not suitable for X-ray imaging due to two reasons. Firstly, X-ray images are grayscale and low-contrast. Secondly, different layers of an object may overlap on the same positions of the projection. These two problems degrade the accuracy of feature detection and matching in SfM. Besides, running the SfM algorithm usually requires a long time, which prolongs the training process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.239, + 0.785, + 0.315 + ], + "angle": 0, + "content": "To address these issues, we customize an Angle-posed Cuboid Uniform Initialization (ACUI) strategy for circular cone beam X-ray scanning scenario where a scanner emits cone-shaped X-ray beams and captures projections at equal angular intervals. As shown in Fig. 2 (a), ACUI uses the parameters of X-ray scanner to compute the extrinsic matrix \\(\\mathbf{M}_{ext}\\) and intrinsic matrix \\(\\mathbf{M}_{int}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.24, + 0.329, + 0.785, + 0.392 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {e x t} = \\left[ \\begin{array}{c c c c} - \\sin \\phi & \\cos \\phi & 0 & 0 \\\\ 0 & 0 & - 1 & 0 \\\\ - \\cos \\phi & - \\sin \\phi & 0 & L _ {S O} \\\\ 0 & 0 & 0 & 1 \\end{array} \\right], \\quad \\mathbf {M} _ {i n t} = \\left[ \\begin{array}{c c c c} L _ {S D} & 0 & W / 2 & 0 \\\\ 0 & L _ {S D} & H / 2 & 0 \\\\ 0 & 0 & 1 & 0 \\end{array} \\right], \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.785, + 0.542 + ], + "angle": 0, + "content": "where \\( L_{SO} \\) represents the distance between the X-ray source and the scanned object. The elevation angle of the X-ray source is set to zero and remains unchanged. The next step of ACUI is to initialize the center positions of 3D Gaussians. Although the precise shape of the scanned object is not given at the beginning, the scanning space can be approximated. We set up a cuboid with size \\( S_{1} \\times S_{2} \\times S_{3} \\) (mm) that can completely enclose the object. The center of this cuboid is also the center of the object and the origin of the world coordinate system. We divide this cuboid by a grid with size \\( M_{1} \\times M_{2} \\times M_{3} \\) (voxel). Then we uniformly sample points within the grid at interval \\( d \\in \\mathbb{R} \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.227, + 0.555, + 0.785, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} = \\left\\{\\left(\\frac {n _ {1} S _ {1} d}{M _ {1}}, \\frac {n _ {2} S _ {2} d}{M _ {2}}, \\frac {n _ {3} S _ {3} d}{M _ {3}}\\right) \\mid - \\left[ \\frac {M _ {i}}{2 d} \\right] - 1 \\leq n _ {i} \\leq \\left[ \\frac {M _ {i}}{2 d} \\right] + 1, i = 1, 2, 3 \\right\\}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.599, + 0.784, + 0.631 + ], + "angle": 0, + "content": "where \\(n_i \\in \\mathbb{Z}\\). Then we use the size and elements of \\(\\mathcal{P}\\) to initialize \\(N_p\\) and \\(\\pmb{\\mu}_i\\). Avoiding running SfM, ACUI allows X-Gaussian to enjoy a faster training speed." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.661, + 0.376, + 0.679 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.702, + 0.446, + 0.717 + ], + "angle": 0, + "content": "4.1 Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Dataset. Following NAF [60], we adopt the public datasets of human organ CTs, i.e., LIDC-IDRI [2] and the open scientific visualization dataset [27], to evaluate our method. The test scenes include chest, foot, head, abdomen, and pancreas. We adopt the open-source tomographic toolbox TIGRE [6] to capture 100 projections with \\(3\\%\\) noise in the range of \\(0 \\sim 180^{\\circ}\\). In the NVS task, 50 projections are used for training and the other 50 projections are used for testing. The CT volumes are used for testing in the sparse-view CT reconstruction task." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.264, + 0.114, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Radiative Gaussian Splitting for Efficient X-ray Novel View Synthesis" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 1: Quantitative results on the novel view synthesis task. The average inference speed and training time of all scenes evaluated on an RTX 8000 GPU are reported. In the cell of the results of each scene, PSNR (upper) and SSIM (lower) are listed." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.199, + 0.785, + 0.428 + ], + "angle": 0, + "content": "
MethodInfer SpeedTrain TimeChestFootHeadAbdomenPancreasAverage
InTomo [58]0.62 fps125 min28.94839.48234.83227.64120.03130.187
0.99150.99790.99770.96460.85370.9611
NeRF [37]0.14 fps313 min36.15741.05329.76024.62019.85330.289
0.99880.99890.99910.95590.85600.9617
TensoRF [9]0.77 fps178 min23.60937.72834.42927.38229.23530.477
0.94020.99290.98790.87300.80310.9194
NeAT [42]1.78 fps69 min40.76538.23627.73826.74137.52634.201
0.99900.99630.92950.85630.90170.9366
NAF [60]2.01 fps63 min42.36638.35330.17437.59036.22836.942
0.99930.99130.95310.98550.88440.9627
X-Gaussian148 fps9 min43.88742.15341.57945.76243.64043.404
0.99980.99970.99970.99990.99760.9993
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.594 + ], + "angle": 0, + "content": "Implementation Details. Our X-Gaussian is implemented by PyTorch [40] and CUDA [17]. The model is trained with the Adam optimizer [26] (\\(\\beta_{1} = 0.9\\), \\(\\beta_{2} = 0.999\\), and \\(\\epsilon = 1\\times 10^{-15}\\)) for \\(2\\times 10^{4}\\) iterations. The learning rate for point cloud position is initially set to \\(1.9\\times 10^{-4}\\) and exponentially decays to \\(1.9\\times 10^{-6}\\). The learning rates for point feature, opacity, scaling, and rotation are set to \\(2\\times 10^{-3}\\), \\(8\\times 10^{-3}\\), \\(5\\times 10^{-3}\\), and \\(1\\times 10^{-3}\\). \\(\\gamma\\) in Eq. (11) is set to 0.2. We adopt peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) [49] to evaluate the performance. Frames per second (fps) is used to measure the inference speed. Experiments are conducted on an RTX 8000 GPU." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.617, + 0.442, + 0.634 + ], + "angle": 0, + "content": "4.2 Novel View Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.787, + 0.689 + ], + "angle": 0, + "content": "Quantitative Results. Tab. 1 shows the quantitative comparisons between our X-Gaussian and five SOTA NeRF-based algorithms, including InTomo [58], NeRF [37], TensoRF [9], NeAT [42], and NAF [60] on the NVs task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We report the average inference speed and training time of different methods on all scenes. In the cell of the results of each scene, PSNR (upper entry in the cell) and SSIM (lower entry in the cell) are listed. As can be observed that our X-Gaussian not only surpasses SOTA methods by large margins in performance but also enjoys much faster inference speed and cheaper training costs. More specifically, compared with the recent best X-ray NeRF-based method NAF, our X-Gaussian outperforms it by 6.5 dB on average and is \\(73 \\times\\) faster in inference while requiring less than \\(15\\%\\) training time. When compared with the SOTA RGB NeRF-based method TensoRF, our X-Gaussian is 12.93 dB higher while enjoying \\(192 \\times\\) inference speed and \\(20 \\times\\) training speed." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.398, + 0.127 + ], + "angle": 0, + "content": "Yuanhao Cai et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.146, + 0.787, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.384, + 0.788, + 0.414 + ], + "angle": 0, + "content": "Fig. 4: Qualitative results of novel view synthesis on the scenes of pancreas (top) and chest (bottom). Our X-Gaussian yields clearer results. Please zoom in for a better view." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.442, + 0.788, + 0.534 + ], + "angle": 0, + "content": "Qualitative Results. Figs. 4 depict the qualitative comparisons of NVS on the scenes of pancreas and chest. It can be observed from the zoomed-in patches that previous NeRF-based algorithms fail to render clear novel views. They either introduce undesired artifacts or produce blurry textures such as the toe bones of the foot. In contrast, our method yields visually realistic images by rendering more fine-grained details and clearer structural contents." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.557, + 0.53, + 0.573 + ], + "angle": 0, + "content": "4.3 Sparse-View CT Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.794 + ], + "angle": 0, + "content": "We compare our method with SOTA NeRF-based algorithms on sparse-view CT reconstruction. Since the Gaussian point clouds cannot directly infer the radiodensities of the CT volume, we evaluate different NeRF-based methods and our X-Gaussian by using them to create novel-view projections for three learning-free CT reconstruction methods, including an analytical method (FDK [16]) and two iterative methods (SART [1] and ASD-POCS [45]). Specifically, these three methods reconstruct the CTs from 5 original projections and 95 novel-view projections rendered by different NVS algorithms. The quantitative results are listed in Tab. 2. When only using 5 original views (+ None), FDK, SART, and ASD-POCS achieve 7.41, 17.24, and \\(17.03\\mathrm{dB}\\) in PSNR, respectively. They fail to reconstruct the CT volumes. When employing our X-Gaussian to create novel X-ray projections for FDK, SART, and ASD-POCS, they yield the most significant improvements of 15.19, 13.01, and \\(13.53\\mathrm{dB}\\) in PSNR. These improvements are 1.32, 2.41, and \\(2.65\\mathrm{dB}\\) higher than the improvements of using NAF." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Figs. 5 shows the qualitative results of sparse-view CT reconstruction on the scenes of foot and chest. Without using rendered novel-view projections, SART fails in reconstructing the CT slices. When using SOTA X-ray NeRF-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.264, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Radiative Gaussian Splitting for Efficient X-ray Novel View Synthesis" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 2: Results on sparse-view CT reconstruction. NeRF-based methods and our X-Gaussian are used to create novel views for FDK [16], SART [1], and ASD-POCS [45]." + }, + { + "type": "table", + "bbox": [ + 0.212, + 0.186, + 0.789, + 0.321 + ], + "angle": 0, + "content": "
Method Metric+ None+ InTomo [58]+ NeRF [37]+ TensorRF [9]+ NeAT [42]+ NAF [60]+ X-Gaussian
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
FDK7.410.09320.310.49820.570.50220.610.50120.940.51121.280.52322.600.584
Δ FDK0012.900.40513.160.40913.200.40813.520.41813.870.43015.190.491
SART17.240.52826.280.85926.780.85327.060.86727.310.86927.840.87930.250.907
Δ SART009.040.3319.540.3259.820.33910.070.34110.600.35113.010.379
ASD-POCS17.030.52525.440.84726.580.85726.930.86826.950.86527.910.88030.560.926
Δ ASD-POCS008.410.3229.550.3329.900.3439.920.34010.880.35513.530.401
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.349, + 0.788, + 0.442 + ], + "angle": 0, + "content": "based methods to create novel views, SART produces over-smooth CT slices with blurry structural contents. On the contrary, when using our X-Gaussian to assist SART and ASD-POCS, they can reconstruct much clearer CT slices with more high-frequency textures and fine-grained structural details, such as the vessels in the chest (Fig. 5). These results clearly demonstrate the potential practical values of our method on the sparse-view CT reconstruction task." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.465, + 0.39, + 0.48 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Break-down Ablation. We first conduct a break-down ablation experiment to study the effect of each proposed technique towards higher performance and faster speed. We adopt the original 3DGS [25] as the baseline model and naively average the RGB channels to represent the value of radiation intensity. The results are listed in Tab. 3a. The baseline model yields 37.21 dB PSNR in performance. Its average training time and inference speed are 31 min 38 s and 64 fps, respectively. We can observe from Tab. 3a: (i) When using ACUI to replace the time-consuming SfM [46] algorithm for initialization, the training time is significantly reduced by \\(34\\%\\) while the performance yields an improvement of 1.66 dB. This evidence suggests that our ACUI strategy is much faster than the SfM [46] algorithm used in the original 3DGS and can compute more accurate camera calibration parameters for 3D Gaussians and subsequent rendering. (ii) Then we apply our radiative Gaussian point cloud model equipped with the proposed Differentiable Radiative Rasterization (DRR) to replace the original RGB Gaussian point cloud model and its RGB rasterization. As analyzed in Sec. 3.1 and compared in Fig. 3, the anisotropic spherical harmonics (SH) are not suitable for X-ray imaging because X-ray imaging based on penetration is isotropic. In contrast, our radiative Gaussian point cloud model can better fit the view-independent radiation intensity in 3D space. Therefore, the performance is significantly improved by 4.53 dB in PSNR. Besides, removing the computation related to the view direction from the forward and backward processes of rasterization can further accelerate the training and inference speed. Thus, the training time is reduced by \\(54.10\\%\\) and the inference speed is \\(2.1\\times\\) faster." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.398, + 0.127 + ], + "angle": 0, + "content": "Yuanhao Cai et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.147, + 0.781, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.359, + 0.785, + 0.389 + ], + "angle": 0, + "content": "Fig. 5: Visual results of sparse-view CT reconstruction on the scenes of foot and chest. NeRF-based methods and our X-Gaussian are used to create novel views for SART [1]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.789, + 0.661 + ], + "angle": 0, + "content": "Initialization of Point Position. We compare different initialization strategies for the center positions of Gaussian point clouds including random, spherical, FDK [16], and cubic initialization. To be specific, Random initialization means randomly sampling points within the scanned area in 3D space. Spherical initialization uniformly samples point positions within a sphere that can completely enclose the scanned object. FDK [16] initialization adopts the FDK algorithm to back-project the given projections into 3D point positions. Cuboid initialization is our ACUI. Please note that we keep the computed intrinsic and extrinsic matrices the same for fair comparison between different strategies. The results are reported in Tab. 3b. FDK initialization slightly outperforms our ACUI by 0.066 dB but its training time is \\(2.59 \\times\\) longer and its inference speed is 55 fps slower than ACUI. This is because the back-projection in FDK is time-consuming and initializes redundant points. The random and spherical initialization strategies yield lower PSNR and slower speed than cubic initialization. To achieve a better trade-off, we adopt the cubic initialization, i.e., ACUI, which enjoys good performance, the cheapest training cost, and the fastest inference speed." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.688, + 0.788, + 0.844 + ], + "angle": 0, + "content": "Parameter Analysis. We conduct parameter analysis of the number of features \\( N_{f} \\) and the sampling interval \\( d \\). The results are shown in Tab. 3c and Tab. 3d. In Tab. 3c, (i) When increasing \\( N_{f} \\), the performance gradually improves but the magnitude of the improvement decreases. In particular, \\( N_{f} = 32 \\) achieves the best results of 43.42 dB in PSNR. \\( N_{f} = 16 \\) achieves on-par results with \\( N_{f} = 32 \\), only 0.013 dB lower. (ii) We notice that the training time and inference speed do not change monotonically. This is because the Gaussian point clouds with various feature dimensions have different representing ability and computational complexity. The number of final 3D Gaussians after training also varies substantially. When \\( N_{f} = 16 \\), the training time reaches a local minimum" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.264, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Radiative Gaussian Splitting for Efficient X-ray Novel View Synthesis" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.784, + 0.16 + ], + "angle": 0, + "content": "Table 3: Ablation study. PSNR, SSIM, training time, and inference speed are reported." + }, + { + "type": "table", + "bbox": [ + 0.225, + 0.171, + 0.475, + 0.226 + ], + "angle": 0, + "content": "
Method3DGS [25]+ ACUI+ DRR
PSNR37.21338.87243.404
SSIM0.98130.98710.9993
Train time (s)18981172538
Infer speed (fps)6472148
" + }, + { + "type": "table_caption", + "bbox": [ + 0.254, + 0.228, + 0.434, + 0.24 + ], + "angle": 0, + "content": "(a) Break-down ablation study" + }, + { + "type": "table", + "bbox": [ + 0.48, + 0.172, + 0.778, + 0.226 + ], + "angle": 0, + "content": "
MethodRandomSphericalFDK [16]Cubic
PSNR41.32942.83743.47043.404
SSIM0.99420.99880.99930.9993
Train time (s)6015751394538
Infer speed (fps)11213693148
" + }, + { + "type": "table_caption", + "bbox": [ + 0.497, + 0.228, + 0.748, + 0.24 + ], + "angle": 0, + "content": "(b) Ablation of point position initialization" + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.252, + 0.497, + 0.307 + ], + "angle": 0, + "content": "
Number12481632
PSNR38.81840.20542.13042.86843.40443.417
SSIM0.98400.99310.99830.99910.99930.9993
Train time (s)511525521553538752
Speed (fps)153152158127148101
" + }, + { + "type": "table_caption", + "bbox": [ + 0.23, + 0.309, + 0.474, + 0.321 + ], + "angle": 0, + "content": "(c) Analysis of the number of features \\(N_{f}\\)" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.252, + 0.779, + 0.308 + ], + "angle": 0, + "content": "
Interval12481632
PSNR42.85342.97943.21543.40443.31143.294
SSIM0.99890.99900.99920.99930.99920.9992
Train time (s)785593545538534566
Speed (fps)869411414813597
" + }, + { + "type": "table_caption", + "bbox": [ + 0.518, + 0.309, + 0.751, + 0.321 + ], + "angle": 0, + "content": "(d) Analysis of the initialized interval \\(d\\)" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.367, + 0.785, + 0.397 + ], + "angle": 0, + "content": "and the inference speed is at its local maximum. Hence, we eventually adopt \\( N_{f} = 16 \\) to reach a more optimal balance between performance and speed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.785, + 0.444 + ], + "angle": 0, + "content": "In Tab. 3d, the best reconstruction performance and the fastest inference speed are achieved at \\( d = 8 \\). The training time (538 s) at \\( d = 8 \\) is almost the same as the shortest one (534 s) at \\( d = 16 \\). Thus, we eventually set \\( d \\) to 8." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.785, + 0.52 + ], + "angle": 0, + "content": "Convergence Analysis. We conduct two visual analyses to compare the convergence between our X-Gaussian and original 3DGS [25] in Fig. 1, and between X-Gaussian and the SOTA X-ray NeRF-based method NAF [60] in Fig. 6." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.787, + 0.748 + ], + "angle": 0, + "content": "Specifically, we adopt the same ACUI strategy for the original 3DGS to focus on comparing the Gaussian point cloud model and the rasterization. For fairness, we train 3DGS and our X-Gaussian on the scene of foot with the same settings and visualize the positions and opacities of Gaussian point clouds at the 100-th, 1000-th, 5000-th, 10000-th, and 20000-th iterations of the training process in Fig. 1. We also visualize the CT volume of foot as a reference. As can be seen that 3DGS with RGB rasterization converges slowly and suffers from more noisy point clouds. Plus, the final trained model of 3DGS at the 20000-th iteration contains more redundant Gaussians that are unnecessary to represent the 3D structure of the foot, which reduces the model's inference speed. In contrast, our X-Gaussian equipped with the proposed DRR shows faster and better convergence. In particular, as early as the 1000-th iteration, our radiative Gaussian point clouds have essentially formed the basic shape of the foot. Besides, the final trained X-Gaussian at the 20000-th iteration can better represent the 3D geometry and more accurate structural contents than the original 3DGS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Additionally, in Fig. 6, we visualize the rendered novel projections with the same angle \\(\\phi\\) of NAF and our X-Gaussian at 20s, 60s, and 180s of the training process on the scene of chest. NAF produces blurry images with severe noises on the background regions within the first three minutes of the training. In contrast, our X-Gaussian can reconstruct clearer structural details like the ribs and blood vessels with cleaner background of the chest at the first minute of the training." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.398, + 0.127 + ], + "angle": 0, + "content": "Yuanhao Cai et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.155, + 0.296, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.238, + 0.216, + 0.277, + 0.222 + ], + "angle": 0, + "content": "NAF at 20s" + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.156, + 0.373, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.316, + 0.216, + 0.354, + 0.222 + ], + "angle": 0, + "content": "NAF at 60s" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.156, + 0.451, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.391, + 0.216, + 0.433, + 0.222 + ], + "angle": 0, + "content": "NAF at 180s" + }, + { + "type": "image", + "bbox": [ + 0.452, + 0.156, + 0.528, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.467, + 0.216, + 0.512, + 0.222 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.224, + 0.296, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.284, + 0.289, + 0.292 + ], + "angle": 0, + "content": "X-Gaussian at 20s" + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.224, + 0.373, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.303, + 0.284, + 0.366, + 0.292 + ], + "angle": 0, + "content": "X-Gaussian at 60s" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.224, + 0.451, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.378, + 0.284, + 0.445, + 0.292 + ], + "angle": 0, + "content": "X-Gaussian at 180s" + }, + { + "type": "image", + "bbox": [ + 0.452, + 0.224, + 0.528, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.467, + 0.284, + 0.512, + 0.292 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.302, + 0.528, + 0.358 + ], + "angle": 0, + "content": "Fig.6: Convergence analysis of NAF [60] vs. our X-Gaussian. We visualize the rendered projections at 20s, 60s, and 180s of training. Our X-Gaussian shows faster and better convergence." + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.156, + 0.77, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.302, + 0.776, + 0.358 + ], + "angle": 0, + "content": "Fig. 7: Analysis of covariance. The mean covariance of 3D Gaussians on different scenes decreases when the number of training views increases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.386, + 0.788, + 0.477 + ], + "angle": 0, + "content": "Analysis of Covariance. We study how the shape of the 3D Gaussian point cloud changes with the number of training views in Fig. 7. As the number of training views increases, the mean covariance of 3D Gaussians that control the size of Gaussian point clouds decreases. This indicates that the 3D Gaussian point clouds gradually change from coarse to fine, thereby being more capable of representing fine-grained structures, such as a small tumor in the abdomen." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.498, + 0.356, + 0.515 + ], + "angle": 0, + "content": "5 Limitation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.528, + 0.785, + 0.559 + ], + "angle": 0, + "content": "The main limitation of our method is that its memory cost is non-trivial, about 51.25 MB on average. It is challenging to deploy it on low-RAM mobile devices." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.579, + 0.36, + 0.595 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.788, + 0.791 + ], + "angle": 0, + "content": "In this paper, we propose the first 3DGS-based framework, X-Gaussian, for X-ray novel view synthesis. Firstly, we design a radiative Gaussian point cloud model based on X-ray imaging properties. This model excludes the influence of view direction when fitting radiation intensity. For this model, we develop a GPU-friendly differentiable radiative rasterization CUDA kernel that renders projections at a faster speed than RGB rasterization. Secondly, we customize an initialization strategy, ACUI, that does not need to execute the SfM algorithm. Instead, ACUI uses the parameters of X-ray scanner to compute the extrinsic and intrinsic matrices, and then uniformly samples center points for 3D Gaussians within a cuboid enclosing the scanned object. Experiments demonstrate that our X-Gaussian significantly outperforms SOTA methods by over \\(6.5\\mathrm{dB}\\) while enjoying \\(73\\times\\) faster inference speed and only requiring \\(15\\%\\) of training time." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Acknowledgements: This work was supported by the Lustgarten Foundation for Pancreatic Cancer Research and the Patrick J. McGovern Foundation Award." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.264, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Radiative Gaussian Splitting for Efficient X-ray Novel View Synthesis" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "1. Andersen, A.H., Kak, A.C.: Simultaneous algebraic reconstruction technique (sart): a superior implementation of the art algorithm. Ultrasonic imaging (1984)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.785, + 0.26 + ], + "angle": 0, + "content": "2. Armato III, S.G., McLennan, G., Bidaut, L., McNitt-Gray, M.F., Meyer, C.R., Reeves, A.P., Zhao, B., Aberle, D.R., Henschke, C.I., Hoffman, E.A., et al.: The lung image database consortium (lidc) and image database resource initiative (idri): a completed reference database of lung nodules on ct scans. Medical physics (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.785, + 0.3 + ], + "angle": 0, + "content": "3. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.301, + 0.787, + 0.328 + ], + "angle": 0, + "content": "4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.33, + 0.785, + 0.356 + ], + "angle": 0, + "content": "5. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-nerf: Anti-aliased grid-based neural radiance fields. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.358, + 0.785, + 0.384 + ], + "angle": 0, + "content": "6. Biguri, A., Dosanjh, M., Hancock, S., Soleimani, M.: Tigre: a matlab-gpu toolbox for cbct image reconstruction. Biomedical Physics & Engineering Express (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.385, + 0.785, + 0.425 + ], + "angle": 0, + "content": "7. Boone, J., Shah, N., Nelson, T.: A comprehensive analysis of coefficients for pendant-geometry cone-beam breast computed tomography. Medical physics (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.426, + 0.785, + 0.453 + ], + "angle": 0, + "content": "8. Boone, J.M., Nelson, T.R., Lindfors, K.K., Seibert, J.A.: Dedicated breast ct: radiation dose and image quality evaluation. Radiology (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.455, + 0.785, + 0.48 + ], + "angle": 0, + "content": "9. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.481, + 0.785, + 0.508 + ], + "angle": 0, + "content": "10. Chen, B., Ning, R.: Cone-beam volume ct breast imaging: Feasibility study. Medical physics (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "1. Chen, Z., Funkhouser, T., Hedman, P., Tagliasacchi, A.: Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.551, + 0.785, + 0.578 + ], + "angle": 0, + "content": "2. Cormack, A.M.: Representation of a function by its line integrals, with some radiological applications. Journal of applied physics (1963)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.579, + 0.785, + 0.605 + ], + "angle": 0, + "content": "3. Cormack, A.M.: Representation of a function by its line integrals, with some radiological applications. ii. Journal of Applied Physics (1964)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.606, + 0.785, + 0.661 + ], + "angle": 0, + "content": "4. Corona-Figueroa, A., Frawley, J., Bond-Taylor, S., Bethapudi, S., Shum, H.P., Willcocks, C.G.: Mednerf: Medical neural radiance fields for reconstructing 3d-aware ct-projections from a single x-ray. In: International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "5. Elbakri, I.A., Fessler, J.A.: Segmentation-free statistical image reconstruction for polyenergetic x-ray computed tomography with experimental validation. Physics in Medicine & Biology (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.703, + 0.785, + 0.729 + ], + "angle": 0, + "content": "6. Feldkamp, L.A., Davis, L.C., Kress, J.W.: Practical cone-beam algorithm. Josa a (1984)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.73, + 0.613, + 0.744 + ], + "angle": 0, + "content": "7. Guide, D.: Cuda c programming guide. NVIDIA (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.785, + 0.771 + ], + "angle": 0, + "content": "8. Hounsfield, G.N.: Computerized transverse axial scanning (tomography): Part 1. description of system. The British journal of radiology (1973)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.65, + 0.785 + ], + "angle": 0, + "content": "9. Hounsfield, G.N.: Computed medical imaging. Science (1980)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "20. Hu, S., Liu, Z.: Gauhuman: Articulated gaussian splatting from monocular human videos. arXiv preprint arXiv: (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "21. Hu, T., Liu, S., Chen, Y., Shen, T., Jia, J.: Efficientnerf efficient neural radian fields. In: CVPR (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.398, + 0.127 + ], + "angle": 0, + "content": "Yuanhao Cai et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "22. Hu, W., Wang, Y., Ma, L., Yang, B., Gao, L., Liu, X., Ma, Y.: Tri-miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.217 + ], + "angle": 0, + "content": "23. Jiang, Y., Tu, J., Liu, Y., Gao, X., Long, X., Wang, W., Ma, Y.: Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. arXiv preprint arXiv:2311.17977 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "24. Keetha, N., Karhade, J., Jatavallabhula, K.M., Yang, G., Scherer, S., Ramanan, D., Luiten, J.: Splatam: Splat, track & map 3d gaussians for dense rgb-d slam. arXiv preprint arXiv:2312.02126 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.286 + ], + "angle": 0, + "content": "25. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "26. Kingma, D.P., Ba, J.L.: Adam: A method for stochastic optimization. In: ICLR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "27. Klacansky, P.: Scientific visualization datasets (2022), https://klacansky.com/open-scivis-datasets/" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "28. Kocabas, M., Chang, J.H.R., Gabriel, J., Tuzel, O., Ranjan, A.: Hugs: Human gaussian splats. arXiv preprint arXiv:2311.17910 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.397 + ], + "angle": 0, + "content": "29. Kopanas, G., Philip, J., Leimkuhler, T., Drettakis, G.: Point-based neural rendering with per-view optimization. In: Computer Graphics Forum (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "30. Li, R., Gao, H., Tancik, M., Kanazawa, A.: Nerfacc: Efficient sampling accelerates nerfs. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.466 + ], + "angle": 0, + "content": "31. Liang, Y., Yang, X., Lin, J., Li, H., Xu, X., Chen, Y.: Luciddreamer: Towards high-fidelity text-to-3d generation via interval score matching. arXiv preprint arXiv:2311.11284 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.494 + ], + "angle": 0, + "content": "32. Liang, Z., Zhang, Q., Feng, Y., Shan, Y., Jia, K.: Gs-ir: 3d gaussian splatting for inverse rendering. arXiv preprint arXiv:2311.16473 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.535 + ], + "angle": 0, + "content": "33. Liu, X., Zhan, X., Tang, J., Shan, Y., Zeng, G., Lin, D., Liu, X., Liu, Z.: Human-gaussian: Text-driven 3d human generation with gaussian splatting. arXiv preprint arXiv:2311.17061 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.563 + ], + "angle": 0, + "content": "34. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.605 + ], + "angle": 0, + "content": "35. Manglos, S.H., Gagne, G.M., Krol, A., Thomas, F.D., Narayanaswamy, R.: Transmission maximum-likelihood reconstruction with ordered subsets for cone beam ct. Physics in Medicine & Biology (1995)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.633 + ], + "angle": 0, + "content": "36. Matsuki, H., Murai, R., Kelly, P.H., Davison, A.J.: Gaussian splatting slam. arXiv preprint arXiv:2312.06741 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.634, + 0.785, + 0.673 + ], + "angle": 0, + "content": "37. Mildenhall, B., Srinivasan, P., Tancik, M., Barron, J., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "38. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM ToG (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.703, + 0.785, + 0.73 + ], + "angle": 0, + "content": "39. Pan, J., Zhou, T., Han, Y., Jiang, M., et al.: Variable weighted ordered subset image reconstruction algorithm. International Journal of Biomedical Imaging (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.731, + 0.785, + 0.771 + ], + "angle": 0, + "content": "40. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: NeurIPS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "41. Reiser, C., Szeliski, R., Verbin, D., Srinivasan, P., Mildenhall, B., Geiger, A., Barron, J., Hedman, P.: Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. TOG (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "42. Rückert, D., Wang, Y., Li, R., Idoughi, R., Heidrich, W.: Neat: Neural adaptive tomography. TOG (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.265, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Radiative Gaussian Splitting for Efficient X-ray Novel View Synthesis" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "43. Sauer, K., Bouman, C.: A local update strategy for iterative reconstruction from projections. TIP (1993)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.215 + ], + "angle": 0, + "content": "44. Scarfe, W.C., Farman, A.G., Sukovic, P., et al.: Clinical applications of cone-beam computed tomography in dental practice. Journal-Canadian Dental Association (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.217, + 0.785, + 0.256 + ], + "angle": 0, + "content": "45. Sidky, E.Y., Pan, X.: Image reconstruction in circular cone-beam computed tomography by constrained, total-variation minimization. Physics in Medicine & Biology (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.257, + 0.785, + 0.282 + ], + "angle": 0, + "content": "46. Snavely, N., Seitz, S.M., Szeliski, R.: Photo tourism: exploring photo collections in 3d. In: SIGGRAPH (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.283, + 0.785, + 0.31 + ], + "angle": 0, + "content": "47. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.311, + 0.785, + 0.35 + ], + "angle": 0, + "content": "48. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.351, + 0.785, + 0.376 + ], + "angle": 0, + "content": "49. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncell, E.P.: Image quality assessment: from error visibility to structural similarity. TIP (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.377, + 0.785, + 0.416 + ], + "angle": 0, + "content": "50. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.417, + 0.785, + 0.457 + ], + "angle": 0, + "content": "51. Xie, T., Zong, Z., Qiu, Y., Li, X., Feng, Y., Yang, Y., Jiang, C.: Physgaus-sian: Physics-integrated 3d gaussians for generative dynamics. arXiv preprint arXiv:2311.12198 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.458, + 0.785, + 0.483 + ], + "angle": 0, + "content": "52. Yan, C., Qu, D., Wang, D., Xu, D., Wang, Z., Zhao, B., Li, X.: Gs-slam: Dense visual slam with 3d gaussian splatting. arXiv preprint arXiv:2311.11700 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.484, + 0.785, + 0.524 + ], + "angle": 0, + "content": "53. Yang, Z., Yang, H., Pan, Z., Zhu, X., Zhang, L.: Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.525, + 0.785, + 0.564 + ], + "angle": 0, + "content": "54. Yariv, L., Hedman, P., Reiser, C., Verbin, D., Srinivasan, P.P., Szeliski, R., Barron, J.T., Mildenhall, B.: Bakedsdf: Meshing neural sdfs for real-time view synthesis. In: SIGGRAPH (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.565, + 0.785, + 0.605 + ], + "angle": 0, + "content": "55. Yi, T., Fang, J., Wu, G., Xie, L., Zhang, X., Liu, W., Tian, Q., Wang, X.: Gaussian dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.606, + 0.785, + 0.632 + ], + "angle": 0, + "content": "56. Yu, L., Zou, Y., Sidky, E.Y., Pelizzari, C.A., Munro, P., Pan, X.: Region of interest reconstruction from truncated data in circular cone-beam ct. TMI (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.633, + 0.785, + 0.657 + ], + "angle": 0, + "content": "57. Yugay, V., Li, Y., Gevers, T., Oswald, M.R.: Gaussian-slam: Photo-realistic dense slam with gaussian splatting. arXiv preprint arXiv:2312.10070 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.659, + 0.785, + 0.697 + ], + "angle": 0, + "content": "58. Zang, G., Idoughi, R., Li, R., Wonka, P., Heidrich, W.: Intratomo: self-supervised learning-based tomography via sinogram synthesis and prediction. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.699, + 0.785, + 0.739 + ], + "angle": 0, + "content": "59. Zbijewski, W., Defrise, M., Viergever, M.A., Beekman, F.J.: Statistical reconstruction for x-ray ct systems with non-continuous detectors. Physics in Medicine & Biology (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.74, + 0.785, + 0.765 + ], + "angle": 0, + "content": "60. Zha, R., Zhang, Y., Li, H.: Naf: neural attenuation fields for sparse-view cbct reconstruction. In: MICCAI (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.766, + 0.785, + 0.805 + ], + "angle": 0, + "content": "61. Zhang, T., Yu, H.X., Wu, R., Feng, B.Y., Zheng, C., Snavely, N., Wu, J., Freeman, W.T.: PhysDreamer: Physics-based interaction with 3d objects via video generation. In: ECCV (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.806, + 0.785, + 0.833 + ], + "angle": 0, + "content": "62. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: Ewa volume splatting. In: Proceedings Visualization, 2001. VIS'01. IEEE (2001)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.833 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis/f615a129-2ee2-41a8-9f1c-82a796edae0e_origin.pdf b/2024/Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis/f615a129-2ee2-41a8-9f1c-82a796edae0e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d8e22f05165a8aaed9da3b1595ce4fb46f816f5e --- /dev/null +++ b/2024/Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis/f615a129-2ee2-41a8-9f1c-82a796edae0e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a8fccd6ef3245272fcb0f3344216b84162f216c88a5523035b32b2a1dfc3b3e7 +size 5999831 diff --git a/2024/Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis/full.md b/2024/Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b52c1f791f1928939f01643037fd9ac4bacb3555 --- /dev/null +++ b/2024/Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis/full.md @@ -0,0 +1,344 @@ +# Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis + +Yuanhao Cai $^{1}$ , Yixun Liang $^{2}$ , Jiahao Wang $^{1}$ , Angtian Wang $^{1}$ , Yulun Zhang $^{3,\dagger}$ , Xiaokang Yang $^{3}$ , Zongwei Zhou $^{1,\dagger}$ , and Alan Yuille $^{1}$ + +1 Johns Hopkins University, 2 HKUST(GZ), 3 Shanghai Jiao Tong University + +Abstract. X-ray is widely applied for transmission imaging due to its stronger penetration than natural light. When rendering novel view X-ray projections, existing methods mainly based on NeRF suffer from long training time and slow inference speed. In this paper, we propose a 3D Gaussian splatting-based method, namely X-Gaussian, for X-ray novel view synthesis. Firstly, we redesign a radiative Gaussian point cloud model inspired by the isotropic nature of X-ray imaging. Our model excludes the influence of view direction when learning to predict the radiation intensity of 3D points. Based on this model, we develop a Differentiable Radiative Rasterization (DRR) with CUDA implementation. Secondly, we customize an Angle-pose Cuboid Uniform Initialization (ACUI) strategy that directly uses the parameters of the X-ray scanner to compute the camera information and then uniformly samples point positions within a cuboid enclosing the scanned object. Experiments show that our X-Gaussian outperforms state-of-the-art methods by 6.5 dB while enjoying less than $15\%$ training time and over $73\times$ inference speed. The application on CT reconstruction also reveals the practical values of our method. Code is at https://github.com/caiyuanhao1998/X-Gaussian + +Keywords: Medical Imaging $\cdot$ View Synthesis $\cdot$ Gaussian Splatting + +# 1 Introduction + +X-ray novel view synthesis (NVS) aims to create X-ray projections of an object from new viewpoints that are not originally captured, using only existing projections scanned from different view directions. As we know, X-ray has stronger penetrating power to capture internal structures of imaged objects and is therefore widely applied in medical imaging [12,13,15,18,19,43]. Nonetheless, X-ray is harmful to human body due to its powerful ionizing radiation, especially when the dose of X-ray increases. Improving NVS techniques can help reduce the exposure to X-rays and provide comprehensive viewpoints of imaged parts for doctors and downstream tasks such as sparse-view CT reconstruction. Hence, X-ray NVS is very important and valuable. We study this task in the circular cone beam X-ray scanning scenario [7,8,10,35,39,44,56,59]. + +![](images/97b2aca403e0517f00163ab81701ee073b786d63c5262cb1690b7816622656fd.jpg) +Fig. 1: Point cloud visualization of the original 3DGS [25] (top) and our X-Gaussian (bottom). We visualize the positions and opacities of the Gaussian point clouds at different training iterations. We also visualize the volume of foot as a reference. Note that the volume is not the ground truth of point clouds. Our X-Gaussian can better represent the detailed structures than 3DGS, showing faster and better convergence. + +Existing methods are mainly based on neural radiance fields (NeRF) [37]. They usually employ a multi-layer perceptron (MLP) to learn the mapping from the point position to the radiodensity and then create projections by volume rendering along rays. This ray tracing scheme is time-consuming because it needs to sample many 3D points and then compute them for every single ray, slowing down the training and inference processes. Even the recent most efficient NeRF-based method [60] still requires over an hour for training and yields suboptimal results at a slow inference speed of 2 fps. This increases the waiting time of patients and doctors, leading to low diagnostic efficiency. + +Recently, 3D Gaussian splattering (3DGS) [25] has demonstrated promising reconstruction quality while enjoying much faster inference speed than NeRF-based algorithms in RGB domain, which motivates us to follow this technical route. However, due to the fundamental differences between X-ray and natural light imaging, directly applying the original 3DGS to X-ray NVS may encounter two issues. Firstly, the spherical harmonics (SH) in RGB 3DGS is not suitable for modeling the X-ray radiation intensity of 3D points. Specifically, natural light imaging relies on the reflection off the surface. The color of a 3D point is anisotropic and view-dependent. Based on this nature, the original Gaussian point cloud model uses SH to fit the illumination distribution. In contrast, X-rays penetrate the object and attenuate, thereby forming an image. Given specific X-rays, the radiation intensity of a 3D point depends on its radiodensity and is independent to the view direction, which means the point radiation intensity is isotropic. Secondly, the original point cloud initialization algorithm, structure-from-motion (SfM) [46], is also not suitable for X-ray imaging. Compared to RGB images, X-ray images are grayscale and their contrast is lower. Additionally, different layers of an object may overlap on the same position of the projection + +due to the transmission imaging nature of X-rays. These two problems degrade the accuracy of feature detection and matching in SfM. Meanwhile, running SfM is time-consuming, which prolongs the training process of Gaussian point clouds. + +To address the above issues, we propose a novel 3DGS-based method, X-Gaussian, for X-ray NVs. Our X-Gaussian composes two key techniques. Firstly, we redesign a radiative Gaussian point cloud model inspired by the isotropic property of X-ray imaging. We present a Radiation Intensity Response Function (RIRF) to replace the SH function of the original 3DGS. Different from SH, our RIRF excludes the influence of view direction. To this end, it adopts the inner product between a learnable vector representing the inherent point features and a set of basis weights to fit the radiation intensity of a 3D point. Based on this point cloud model, we further develop a Differentiable Radiative Rasterization (DRR) with a CUDA implementation to render novel projections. Secondly, we customize an Angle-pose Cuboid Uniform Initialization (ACUI) strategy for camera calibration parameters and Gaussian point clouds. Our ACUI first exploits the parameters of the X-ray scanner to compute the intrinsic and extrinsic matrices. Then we set up a cuboid that can completely enclose the scanned object. Within this cuboid, we uniformly sample 3D points at intervals to initialize the center positions of the Gaussian point clouds. Free from running the SfM algorithm, our ACUI significantly reduces the training time. Equipped with the two proposed techniques, our X-Gaussian outperforms SOTA methods by 6.5 dB while enjoying $73 \times$ inference speed and $7 \times$ training speed. + +The main contributions of this work can be summarized as follows: + +- We propose a novel 3D Gaussian splatting-based framework, X-Gaussian, for X-ray novel view synthesis. To our knowledge, this is the first attempt to explore the potential of Gaussian splatting in X-ray neural rendering. +- We design a radiative Gaussian point cloud model with a differentiable radiative rasterization based on the isotropic nature of X-ray imaging. +- We present an angle-pose cuboid uniform initialization strategy for Gaussian point clouds and camera calibration in circular cone beam X-ray scanning. +- Our X-Gaussian significantly outperforms SOTA NeRF-based methods with much faster speed. Experiments also show that our method can improve the performance of sparse-view CT reconstruction, showing its practical values. + +# 2 Related Work + +# 2.1 Neural Radiance Field + +NeRF [37] learns an implicit neural scene representation of color and volume density, given the position of a 3D point and view direction. It has achieved great success in NVS and inspired an explosion of follow-up papers to improve its quality [3-5,22,48] and speed [9,11,21,30,38,41,54]. For example, Instant-NGP [38] adopts hash tables as the encoder to allow small MLP for fast training and inference. Some later works extend the application domain of NeRF from natural light to X-rays [14,58,60]. For instance, NAF [60] follows the settings + +![](images/16f43c0f96a51489bfb4872285d3dfff2b5ec580c759d311106f60a8a62b4425.jpg) +Fig. 2: Pipeline of our method. (a) Angle-posed Cuboid Uniform Initialization (ACUI) strategy uses the parameters of X-ray scanner to compute intrinsic and extrinsic matrices, and samples center points for 3D Gaussians. (b) Our radiative Gaussian point cloud model learns to predict the radiation intensity of 3D points. (c) Based on our Gaussian model, we develop a GPU-friendly Differentiable Radiative Rasterization (DRR). + +of Instant-NGP to learn the implicit mapping from 3D position to attenuation. Yet, the ray tracing and volume rendering schemes are time-consuming, which limits the training and inference speed of NeRF-based X-ray NVS algorithms. + +# 2.2 Gaussian Splitting + +3DGS [25] represents scenes using millions of 3D Gaussian point clouds. This approach is fundamentally different from NeRF-based algorithms by employing an explicit representation coupled with highly parallelized rasterization workflows. These features enable more efficient computation and rendering processes. Hence, 3DGS has achieved great success in several fields, including 3D Generation [20,28,31,33,47,55,61], Dynamic Scene Modeling [34,50,53], SLAM [24,36,52,57], Inverse Rendering [23,32,51], etc. However, most applications of 3DGS are focused on natural scenes with RGB colors. The potential of 3DGS in X-ray imaging still remains under-explored. Our goal is to fill this research gap. + +# 3 Method + +The pipeline of our X-Gaussian is shown in Fig. 2. Firstly, we design an Angle-posed Cuboid Uniform Initialization (ACUI) to compute the intrinsic and extrinsic matrices from the parameters of X-ray scanner, as illustrated in Fig. 2 (a). Then ACUI uniformly samples 3D points within a cuboid that can completely enclose the scanned object to initialize the center positions of our radiative Gaussian point clouds in Fig. 2 (b). Given a view direction, the 3D point clouds undergo our Differentiable Radiative Rasterization (DRR) to derive the rendered image, as depicted in Fig. 2 (c). In this section, we will introduce our radiative Gaussian point cloud model and DRR processing first and then the ACUI strategy. + +![](images/3265cdb3c1db691484126c0b557ae8f76e0819232c02b24238280df5e57b12b9.jpg) +(a) Original RGB Gaussian Point Cloud + +![](images/d8fb9744312a4a78a7b5c5945597d095b461df18b1a031ce95131cf733984593.jpg) +(b) Our Radiative Gaussian Point Cloud +Fig. 3: Comparison between the Gaussian point cloud models of the original 3DGS and our X-Gaussian. (a) The original RGB Gaussian point cloud model uses spherical harmonics (SH) to simulate the anisotropic natural light distribution and view-dependent color. (b) Our radiative Gaussian point cloud model employs the weighted sum of point features to fit the isotropic X-ray penetration and view-independent radiation intensity. + +# 3.1 Radiative Gaussian Point Cloud Model + +An object can be represented by a set of basic Gaussian point clouds $\mathcal{G}$ as + +$$ +\mathcal {G} = \left\{G _ {i} \left(\boldsymbol {\mu} _ {i}, \boldsymbol {\Sigma} _ {i}, \alpha_ {i}\right) \mid i = 1, 2, \dots , N _ {p} \right\}, \tag {1} +$$ + +where $G_{i}$ refers to the $i$ -th Gaussian point cloud. Its center position, covariance, and opacity are defined as $\pmb{\mu}_{i} \in \mathbb{R}^{3}$ , $\pmb{\Sigma}_{i} \in \mathbb{R}^{3 \times 3}$ , and $\alpha_{i} \in \mathbb{R}$ . $\pmb{\Sigma}_{i}$ is represented by a rotation matrix $\mathbf{R}_i \in \mathbb{R}^3$ and a scaling matrix $\mathbf{S}_i \in \mathbb{R}^3$ as $\pmb{\Sigma}_{i} = \mathbf{R}_{i}\mathbf{S}_{i}\mathbf{S}_{i}^{\top}\mathbf{R}_{i}^{\top}$ . $\pmb{\mu}_{i}, \pmb{\Sigma}_{i}, \alpha_{i}, \mathbf{R}_{i}$ , and $\mathbf{S}_i$ are learnable parameters. Besides these basic attributes, each Gaussian point cloud also employs additional learnable parameters to fit different imaging scenarios, e.g., natural light imaging and X-ray imaging. + +We first review the original RGB Gaussian point cloud model [25] in natural light imaging. As shown in Fig. 3 (a), the color of a 3D point is represented by spherical harmonics (SH). The point color is anisotropic and changes with the view direction. Each Gaussian point cloud learns to predict the SH coefficients $\mathbf{k} = \{k_l^m |0\leq l\leq L, - l\leq m\leq l\} \in \mathbb{R}^{(L + 1)^2\times 3}$ , where each $k_{l}^{m}\in \mathbb{R}^{3}$ is a set of 3 coefficients corresponding to the RGB components. $L$ is the degree of SH. Then the point color $\mathbf{c}\in \mathbb{R}^3$ at the view direction $\mathbf{d} = (\theta ,\phi)$ is derived by + +$$ +\mathbf {c} (\mathbf {d}, \mathbf {k}) = \sum_ {l = 0} ^ {L} \sum_ {m = - l} ^ {l} k _ {l} ^ {m} Y _ {l} ^ {m} (\theta , \phi), \tag {2} +$$ + +where $Y_{l}^{m}:\mathbb{S}^{2}\to \mathbb{R}$ is the SH function that maps points on the sphere to real numbers. Please refer to the supplementary for its detailed formulation. + +Although 3DGS [25] achieves fast inference speed and good performance in natural light imaging, the RGB Gaussian point cloud model is not suitable for X-ray scenarios due to the fundamental differences between natural light imaging and X-ray imaging. Natural light imaging relies on the reflection off the surface of + +object. The anisotropic color modeled by SH is view-dependent. e.g., in Fig. 3 (a), the point color is blue from the left viewpoint and green from the right viewpoint. In contrast, X-ray imaging is based on the attenuation when penetrating the object. The degree of attenuation depends on the isotropic radiodensity property. Thus, the radiation intensity of a 3D point is view-independent. + +In light of the above analysis, we redesign our radiative Gaussian point cloud model. Different from the original 3DGS that uses SH to fit the color information for each point, our model introduces a Radiation Intensity Response Function (RIRF) to predict the radiation intensity of the 3D point. As illustrated in Fig. 3 (b), each Gaussian point cloud learns a feature vector $\mathbf{f} \in \mathbb{R}^{N_f}$ to represent its inherent radiative properties. Subsequently, the radiation intensity $\mathbf{i} \in \mathbb{R}$ of the center point of a 3D Gaussian at any view direction is modeled by RIRF as + +$$ +\mathbf {i} (\mathbf {f}) = \operatorname {R I R F} (\mathbf {f}) = \operatorname {S i g m o i d} (\boldsymbol {\lambda} \odot \mathbf {f}), \tag {3} +$$ + +where the Sigmoid function activates and normalizes the radiation intensity. $\lambda \in \mathbb{R}^{N_f}$ is a set of constant weights controlling the importance of each component of $\mathbf{f}$ . Then the set of our radiative Gaussian point clouds $\mathcal{G}_x$ is formulated as + +$$ +\mathcal {G} _ {x} = \left\{G _ {i} \left(\boldsymbol {\mu} _ {i}, \boldsymbol {\Sigma} _ {i}, \alpha_ {i}, \mathbf {f} _ {i}\right) \mid i = 1, 2, \dots , N _ {p} \right\}, \tag {4} +$$ + +where $\mathbf{f}_i\in \mathbb{R}^{N_f}$ denotes the feature vector of the $i$ -th Gaussian point cloud. Please note that Eq. (3) excludes the influence of the view direction $\mathbf{d} = (\theta ,\phi)$ , which matches the isotropic nature of X-ray imaging. Meanwhile, Eq. (3) is free from the complex computation of SH function. Hence, the forward and backward processes of our X-Gaussian are much faster than those of the original 3DGS. + +# 3.2 Differentiable Radiative Rasterization + +Based on our radiative Gaussian point cloud, we develop a Differentiable Radiative Rasterization (DRR), as shown in Fig. 2 (c). We first summarize the overall DRR processing $F_{\mathrm{DRR}}$ and then describe its details. DRR is represented as + +$$ +\mathbf {I} = F _ {\mathrm {D R R}} \left(\mathbf {M} _ {e x t}, \mathbf {M} _ {i n t}, \left\{G _ {i} \left(\boldsymbol {\mu} _ {i}, \boldsymbol {\Sigma} _ {i}, \alpha_ {i}, \mathbf {f} _ {i}\right) \mid i = 1, 2, \dots , N _ {p} \right\}\right), \tag {5} +$$ + +where $\mathbf{I} \in \mathbb{R}^{H \times W}$ denotes the rendered image, $\mathbf{M}_{ext} \in \mathbb{R}^{4 \times 4}$ represents the extrinsic matrix, and $\mathbf{M}_{int} \in \mathbb{R}^{4 \times 3}$ refers to the intrinsic matrix. Subsequently, we introduce the details of $F_{\mathrm{DRR}}$ . To begin with, the possibility value of the $i$ -th Gaussian distribution at the 3D point position $\mathbf{x} \in \mathbb{R}^3$ is formulated as + +$$ +P (\mathbf {x} \mid \boldsymbol {\mu} _ {i}, \boldsymbol {\Sigma} _ {i}) = \exp \left(- \frac {1}{2} (\mathbf {x} - \boldsymbol {\mu} _ {i}) ^ {\top} \boldsymbol {\Sigma} _ {i} ^ {- 1} (\mathbf {x} - \boldsymbol {\mu} _ {i})\right). \tag {6} +$$ + +Then we project the 3D Gaussians to the 2D detector plane for subsequent rendering. $\pmb{\mu}_i$ is firstly transferred from the world coordinate system to the camera coordinate system and then projected to the image coordinate system as + +$$ +\widetilde {\mathbf {t}} _ {i} = \left[ \begin{array}{l} \mathbf {t} _ {i} \\ 1 \end{array} \right] = \mathbf {M} _ {e x t} \widetilde {\boldsymbol {\mu}} _ {i} = \mathbf {M} _ {e x t} \left[ \begin{array}{l} \boldsymbol {\mu} _ {i} \\ 1 \end{array} \right], \quad \widetilde {\mathbf {u}} _ {i} = \left[ \begin{array}{l} \mathbf {u} _ {i} \\ 1 \end{array} \right] = \mathbf {M} _ {i n t} \widetilde {\mathbf {t}} _ {i} = \mathbf {M} _ {i n t} \left[ \begin{array}{l} \mathbf {t} _ {i} \\ 1 \end{array} \right], \tag {7} +$$ + +where $\mathbf{t}_i = (t_x, t_y, t_z) \in \mathbb{R}^3$ is the camera coordinate of $\pmb{\mu}_i$ and $\mathbf{u}_i \in \mathbb{R}^2$ is the image coordinate of $\pmb{\mu}_i$ . $\widetilde{\mathbf{u}}_i$ , $\widetilde{\mathbf{t}}_i$ , and $\widetilde{\pmb{\mu}}_i$ are the homogeneous coordinates of $\mathbf{u}_i$ , $\mathbf{t}_i$ , and $\pmb{\mu}_i$ , respectively. Subsequently, we transfer the 3D covariance matrix $\pmb{\Sigma}_i$ to its counterpart $\pmb{\Sigma}_i' \in \mathbb{R}^{3 \times 3}$ in the camera coordinate system as + +$$ +\boldsymbol {\Sigma} _ {i} ^ {\prime} = \mathbf {J} _ {i} \mathbf {W} _ {i} \boldsymbol {\Sigma} _ {i} \mathbf {W} _ {i} ^ {\top} \mathbf {J} _ {i} ^ {\top}, \tag {8} +$$ + +where $\mathbf{J}_i\in \mathbb{R}^{3\times 3}$ is the Jacobian of the affine approximation of the projective transformation. $\mathbf{W}_i\in \mathbb{R}^{3\times 3}$ is the viewing transformation. We derive them by + +$$ +\mathbf {J} _ {i} = \left[ \begin{array}{c c c} \frac {L _ {S D}}{t _ {z}} & 0 & - \frac {L _ {S D}}{t _ {z} ^ {2}} \frac {t _ {x}}{t _ {z}} \\ 0 & \frac {L _ {S D}}{t _ {z}} & - \frac {L _ {S D}}{t _ {z} ^ {2}} \frac {t _ {y}}{t _ {z}} \\ 0 & 0 & 0 \end{array} \right], \quad \mathbf {W} _ {i} = \left[ \begin{array}{c c c} - \sin \phi & \cos \phi & 0 \\ 0 & 0 & - 1 \\ - \cos \phi & - \sin \phi & 0 \end{array} \right], \tag {9} +$$ + +where $L_{SD}$ represents the distance between the X-ray source and detector. $\phi$ refers to the azimuth angle of the source. Following [25,29,62], we obtain the 2D covariance matrix $\pmb{\Sigma}_i^{\prime \prime} \in \mathbb{R}^{2 \times 2}$ by skipping the third row and column of $\pmb{\Sigma}_i^{\prime}$ . Then the 2D projection is partitioned into non-overlapping tiles. The 3D Gaussians $(\pmb{\mu}_i, \pmb{\Sigma}_i)$ are assigned to different tiles according to their 2D projections $(\pmb{u}_i, \pmb{\Sigma}_i^{\prime \prime})$ , as shown in the left image of Fig. 2 (c). These 3D Gaussians are sorted by the distances to the 2D detector. Then the intensity $\mathbf{I}(p) \in \mathbb{R}$ at pixel $p$ is obtained by blending $\mathcal{N}$ ordered points overlapping the pixel in the corresponding tile as + +$$ +\mathbf {I} (p) = \sum_ {j \in \mathcal {N}} \mathbf {i} _ {j} \sigma_ {j} \prod_ {k = 1} ^ {j - 1} (1 - \sigma_ {k}), \quad \sigma_ {j} = \alpha_ {j} P (\mathbf {x} _ {j} | \boldsymbol {\mu} _ {j}, \boldsymbol {\Sigma} _ {j}), \tag {10} +$$ + +where $\mathbf{x}_j$ is the $j$ -th intersection 3D point of the X-ray landing on pixel $p$ and the Gaussian point clouds in 3D space. $\mathbf{i}_j$ is the radiation intensity of $\mathbf{x}_j$ . + +Optimization. Eventually, the training objective $\mathcal{L}$ is the weighted sum of $\mathcal{L}_1$ loss and SSIM loss between the rendered and ground-truth projection images as + +$$ +\mathcal {L} = (1 - \gamma) \mathcal {L} _ {1} + \gamma \mathcal {L} _ {\text {S S I M}}, \tag {11} +$$ + +where $\gamma$ is a hyperparameter balancing the importances of the two loss terms. By minimizing Eq. (11), we can optimize the attributes of 3D Gaussians, i.e., $\pmb{\mu}_i, \pmb{\Sigma}_i, \alpha_i$ , and $\mathbf{f}_i$ in Eq. (5). $N_p$ is adjusted by the adaptive control [25]. The optimization process is visualized in Fig. 1 and the video file in supplementary. + +Compared to the RGB rasterization in 3DGS [25], our DRR avoids the complex computations related to the view direction in the forward and backward processes, thereby enjoying cheaper training costs and faster inference speed. + +# 3.3 Angle-pose Cuboid Uniform Initialization + +At the beginning of training, we need to initialize the parameters in Eq. (5) for rasterization. Specifically, $\mathbf{\Sigma}_i$ , $\alpha_i$ , and $\mathbf{f}_i$ are randomly initialized. In natural light imaging, the original 3DGS [25] adopts the SfM [46] algorithm to compute the + +initial $\pmb{\mu}_i$ , $N_p$ , $\mathbf{M}_{ext}$ , and $\mathbf{M}_{int}$ . SfM detects and matches features from multiview images. It is not suitable for X-ray imaging due to two reasons. Firstly, X-ray images are grayscale and low-contrast. Secondly, different layers of an object may overlap on the same positions of the projection. These two problems degrade the accuracy of feature detection and matching in SfM. Besides, running the SfM algorithm usually requires a long time, which prolongs the training process. + +To address these issues, we customize an Angle-posed Cuboid Uniform Initialization (ACUI) strategy for circular cone beam X-ray scanning scenario where a scanner emits cone-shaped X-ray beams and captures projections at equal angular intervals. As shown in Fig. 2 (a), ACUI uses the parameters of X-ray scanner to compute the extrinsic matrix $\mathbf{M}_{ext}$ and intrinsic matrix $\mathbf{M}_{int}$ as + +$$ +\mathbf {M} _ {e x t} = \left[ \begin{array}{c c c c} - \sin \phi & \cos \phi & 0 & 0 \\ 0 & 0 & - 1 & 0 \\ - \cos \phi & - \sin \phi & 0 & L _ {S O} \\ 0 & 0 & 0 & 1 \end{array} \right], \quad \mathbf {M} _ {i n t} = \left[ \begin{array}{c c c c} L _ {S D} & 0 & W / 2 & 0 \\ 0 & L _ {S D} & H / 2 & 0 \\ 0 & 0 & 1 & 0 \end{array} \right], \tag {12} +$$ + +where $L_{SO}$ represents the distance between the X-ray source and the scanned object. The elevation angle of the X-ray source is set to zero and remains unchanged. The next step of ACUI is to initialize the center positions of 3D Gaussians. Although the precise shape of the scanned object is not given at the beginning, the scanning space can be approximated. We set up a cuboid with size $S_{1} \times S_{2} \times S_{3}$ (mm) that can completely enclose the object. The center of this cuboid is also the center of the object and the origin of the world coordinate system. We divide this cuboid by a grid with size $M_{1} \times M_{2} \times M_{3}$ (voxel). Then we uniformly sample points within the grid at interval $d \in \mathbb{R}$ as + +$$ +\mathcal {P} = \left\{\left(\frac {n _ {1} S _ {1} d}{M _ {1}}, \frac {n _ {2} S _ {2} d}{M _ {2}}, \frac {n _ {3} S _ {3} d}{M _ {3}}\right) \mid - \left[ \frac {M _ {i}}{2 d} \right] - 1 \leq n _ {i} \leq \left[ \frac {M _ {i}}{2 d} \right] + 1, i = 1, 2, 3 \right\}, \tag {13} +$$ + +where $n_i \in \mathbb{Z}$ . Then we use the size and elements of $\mathcal{P}$ to initialize $N_p$ and $\pmb{\mu}_i$ . Avoiding running SfM, ACUI allows X-Gaussian to enjoy a faster training speed. + +# 4 Experiments + +# 4.1 Experimental Settings + +Dataset. Following NAF [60], we adopt the public datasets of human organ CTs, i.e., LIDC-IDRI [2] and the open scientific visualization dataset [27], to evaluate our method. The test scenes include chest, foot, head, abdomen, and pancreas. We adopt the open-source tomographic toolbox TIGRE [6] to capture 100 projections with $3\%$ noise in the range of $0 \sim 180^{\circ}$ . In the NVS task, 50 projections are used for training and the other 50 projections are used for testing. The CT volumes are used for testing in the sparse-view CT reconstruction task. + +Table 1: Quantitative results on the novel view synthesis task. The average inference speed and training time of all scenes evaluated on an RTX 8000 GPU are reported. In the cell of the results of each scene, PSNR (upper) and SSIM (lower) are listed. + +
MethodInfer SpeedTrain TimeChestFootHeadAbdomenPancreasAverage
InTomo [58]0.62 fps125 min28.94839.48234.83227.64120.03130.187
0.99150.99790.99770.96460.85370.9611
NeRF [37]0.14 fps313 min36.15741.05329.76024.62019.85330.289
0.99880.99890.99910.95590.85600.9617
TensoRF [9]0.77 fps178 min23.60937.72834.42927.38229.23530.477
0.94020.99290.98790.87300.80310.9194
NeAT [42]1.78 fps69 min40.76538.23627.73826.74137.52634.201
0.99900.99630.92950.85630.90170.9366
NAF [60]2.01 fps63 min42.36638.35330.17437.59036.22836.942
0.99930.99130.95310.98550.88440.9627
X-Gaussian148 fps9 min43.88742.15341.57945.76243.64043.404
0.99980.99970.99970.99990.99760.9993
+ +Implementation Details. Our X-Gaussian is implemented by PyTorch [40] and CUDA [17]. The model is trained with the Adam optimizer [26] ( $\beta_{1} = 0.9$ , $\beta_{2} = 0.999$ , and $\epsilon = 1\times 10^{-15}$ ) for $2\times 10^{4}$ iterations. The learning rate for point cloud position is initially set to $1.9\times 10^{-4}$ and exponentially decays to $1.9\times 10^{-6}$ . The learning rates for point feature, opacity, scaling, and rotation are set to $2\times 10^{-3}$ , $8\times 10^{-3}$ , $5\times 10^{-3}$ , and $1\times 10^{-3}$ . $\gamma$ in Eq. (11) is set to 0.2. We adopt peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) [49] to evaluate the performance. Frames per second (fps) is used to measure the inference speed. Experiments are conducted on an RTX 8000 GPU. + +# 4.2 Novel View Synthesis + +Quantitative Results. Tab. 1 shows the quantitative comparisons between our X-Gaussian and five SOTA NeRF-based algorithms, including InTomo [58], NeRF [37], TensoRF [9], NeAT [42], and NAF [60] on the NVs task. + +We report the average inference speed and training time of different methods on all scenes. In the cell of the results of each scene, PSNR (upper entry in the cell) and SSIM (lower entry in the cell) are listed. As can be observed that our X-Gaussian not only surpasses SOTA methods by large margins in performance but also enjoys much faster inference speed and cheaper training costs. More specifically, compared with the recent best X-ray NeRF-based method NAF, our X-Gaussian outperforms it by 6.5 dB on average and is $73 \times$ faster in inference while requiring less than $15\%$ training time. When compared with the SOTA RGB NeRF-based method TensoRF, our X-Gaussian is 12.93 dB higher while enjoying $192 \times$ inference speed and $20 \times$ training speed. + +![](images/704a3c2ce3cce0b91c4a450be244bf2b5a2d6b6b689c15768c55d3dbcb1202ca.jpg) +Fig. 4: Qualitative results of novel view synthesis on the scenes of pancreas (top) and chest (bottom). Our X-Gaussian yields clearer results. Please zoom in for a better view. + +Qualitative Results. Figs. 4 depict the qualitative comparisons of NVS on the scenes of pancreas and chest. It can be observed from the zoomed-in patches that previous NeRF-based algorithms fail to render clear novel views. They either introduce undesired artifacts or produce blurry textures such as the toe bones of the foot. In contrast, our method yields visually realistic images by rendering more fine-grained details and clearer structural contents. + +# 4.3 Sparse-View CT Reconstruction + +We compare our method with SOTA NeRF-based algorithms on sparse-view CT reconstruction. Since the Gaussian point clouds cannot directly infer the radiodensities of the CT volume, we evaluate different NeRF-based methods and our X-Gaussian by using them to create novel-view projections for three learning-free CT reconstruction methods, including an analytical method (FDK [16]) and two iterative methods (SART [1] and ASD-POCS [45]). Specifically, these three methods reconstruct the CTs from 5 original projections and 95 novel-view projections rendered by different NVS algorithms. The quantitative results are listed in Tab. 2. When only using 5 original views (+ None), FDK, SART, and ASD-POCS achieve 7.41, 17.24, and $17.03\mathrm{dB}$ in PSNR, respectively. They fail to reconstruct the CT volumes. When employing our X-Gaussian to create novel X-ray projections for FDK, SART, and ASD-POCS, they yield the most significant improvements of 15.19, 13.01, and $13.53\mathrm{dB}$ in PSNR. These improvements are 1.32, 2.41, and $2.65\mathrm{dB}$ higher than the improvements of using NAF. + +Figs. 5 shows the qualitative results of sparse-view CT reconstruction on the scenes of foot and chest. Without using rendered novel-view projections, SART fails in reconstructing the CT slices. When using SOTA X-ray NeRF- + +Table 2: Results on sparse-view CT reconstruction. NeRF-based methods and our X-Gaussian are used to create novel views for FDK [16], SART [1], and ASD-POCS [45]. + +
Method Metric+ None+ InTomo [58]+ NeRF [37]+ TensorRF [9]+ NeAT [42]+ NAF [60]+ X-Gaussian
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
FDK7.410.09320.310.49820.570.50220.610.50120.940.51121.280.52322.600.584
Δ FDK0012.900.40513.160.40913.200.40813.520.41813.870.43015.190.491
SART17.240.52826.280.85926.780.85327.060.86727.310.86927.840.87930.250.907
Δ SART009.040.3319.540.3259.820.33910.070.34110.600.35113.010.379
ASD-POCS17.030.52525.440.84726.580.85726.930.86826.950.86527.910.88030.560.926
Δ ASD-POCS008.410.3229.550.3329.900.3439.920.34010.880.35513.530.401
+ +based methods to create novel views, SART produces over-smooth CT slices with blurry structural contents. On the contrary, when using our X-Gaussian to assist SART and ASD-POCS, they can reconstruct much clearer CT slices with more high-frequency textures and fine-grained structural details, such as the vessels in the chest (Fig. 5). These results clearly demonstrate the potential practical values of our method on the sparse-view CT reconstruction task. + +# 4.4 Ablation Study + +Break-down Ablation. We first conduct a break-down ablation experiment to study the effect of each proposed technique towards higher performance and faster speed. We adopt the original 3DGS [25] as the baseline model and naively average the RGB channels to represent the value of radiation intensity. The results are listed in Tab. 3a. The baseline model yields 37.21 dB PSNR in performance. Its average training time and inference speed are 31 min 38 s and 64 fps, respectively. We can observe from Tab. 3a: (i) When using ACUI to replace the time-consuming SfM [46] algorithm for initialization, the training time is significantly reduced by $34\%$ while the performance yields an improvement of 1.66 dB. This evidence suggests that our ACUI strategy is much faster than the SfM [46] algorithm used in the original 3DGS and can compute more accurate camera calibration parameters for 3D Gaussians and subsequent rendering. (ii) Then we apply our radiative Gaussian point cloud model equipped with the proposed Differentiable Radiative Rasterization (DRR) to replace the original RGB Gaussian point cloud model and its RGB rasterization. As analyzed in Sec. 3.1 and compared in Fig. 3, the anisotropic spherical harmonics (SH) are not suitable for X-ray imaging because X-ray imaging based on penetration is isotropic. In contrast, our radiative Gaussian point cloud model can better fit the view-independent radiation intensity in 3D space. Therefore, the performance is significantly improved by 4.53 dB in PSNR. Besides, removing the computation related to the view direction from the forward and backward processes of rasterization can further accelerate the training and inference speed. Thus, the training time is reduced by $54.10\%$ and the inference speed is $2.1\times$ faster. + +![](images/c93447777a2c2d8286aa2487c68acfc3a3cf1ca13ce3f955a41c7fa0aff3296f.jpg) +Fig. 5: Visual results of sparse-view CT reconstruction on the scenes of foot and chest. NeRF-based methods and our X-Gaussian are used to create novel views for SART [1]. + +Initialization of Point Position. We compare different initialization strategies for the center positions of Gaussian point clouds including random, spherical, FDK [16], and cubic initialization. To be specific, Random initialization means randomly sampling points within the scanned area in 3D space. Spherical initialization uniformly samples point positions within a sphere that can completely enclose the scanned object. FDK [16] initialization adopts the FDK algorithm to back-project the given projections into 3D point positions. Cuboid initialization is our ACUI. Please note that we keep the computed intrinsic and extrinsic matrices the same for fair comparison between different strategies. The results are reported in Tab. 3b. FDK initialization slightly outperforms our ACUI by 0.066 dB but its training time is $2.59 \times$ longer and its inference speed is 55 fps slower than ACUI. This is because the back-projection in FDK is time-consuming and initializes redundant points. The random and spherical initialization strategies yield lower PSNR and slower speed than cubic initialization. To achieve a better trade-off, we adopt the cubic initialization, i.e., ACUI, which enjoys good performance, the cheapest training cost, and the fastest inference speed. + +Parameter Analysis. We conduct parameter analysis of the number of features $N_{f}$ and the sampling interval $d$ . The results are shown in Tab. 3c and Tab. 3d. In Tab. 3c, (i) When increasing $N_{f}$ , the performance gradually improves but the magnitude of the improvement decreases. In particular, $N_{f} = 32$ achieves the best results of 43.42 dB in PSNR. $N_{f} = 16$ achieves on-par results with $N_{f} = 32$ , only 0.013 dB lower. (ii) We notice that the training time and inference speed do not change monotonically. This is because the Gaussian point clouds with various feature dimensions have different representing ability and computational complexity. The number of final 3D Gaussians after training also varies substantially. When $N_{f} = 16$ , the training time reaches a local minimum + +Table 3: Ablation study. PSNR, SSIM, training time, and inference speed are reported. + +
Method3DGS [25]+ ACUI+ DRR
PSNR37.21338.87243.404
SSIM0.98130.98710.9993
Train time (s)18981172538
Infer speed (fps)6472148
+ +(a) Break-down ablation study + +
MethodRandomSphericalFDK [16]Cubic
PSNR41.32942.83743.47043.404
SSIM0.99420.99880.99930.9993
Train time (s)6015751394538
Infer speed (fps)11213693148
+ +(b) Ablation of point position initialization + +
Number12481632
PSNR38.81840.20542.13042.86843.40443.417
SSIM0.98400.99310.99830.99910.99930.9993
Train time (s)511525521553538752
Speed (fps)153152158127148101
+ +(c) Analysis of the number of features $N_{f}$ + +
Interval12481632
PSNR42.85342.97943.21543.40443.31143.294
SSIM0.99890.99900.99920.99930.99920.9992
Train time (s)785593545538534566
Speed (fps)869411414813597
+ +(d) Analysis of the initialized interval $d$ + +and the inference speed is at its local maximum. Hence, we eventually adopt $N_{f} = 16$ to reach a more optimal balance between performance and speed. + +In Tab. 3d, the best reconstruction performance and the fastest inference speed are achieved at $d = 8$ . The training time (538 s) at $d = 8$ is almost the same as the shortest one (534 s) at $d = 16$ . Thus, we eventually set $d$ to 8. + +Convergence Analysis. We conduct two visual analyses to compare the convergence between our X-Gaussian and original 3DGS [25] in Fig. 1, and between X-Gaussian and the SOTA X-ray NeRF-based method NAF [60] in Fig. 6. + +Specifically, we adopt the same ACUI strategy for the original 3DGS to focus on comparing the Gaussian point cloud model and the rasterization. For fairness, we train 3DGS and our X-Gaussian on the scene of foot with the same settings and visualize the positions and opacities of Gaussian point clouds at the 100-th, 1000-th, 5000-th, 10000-th, and 20000-th iterations of the training process in Fig. 1. We also visualize the CT volume of foot as a reference. As can be seen that 3DGS with RGB rasterization converges slowly and suffers from more noisy point clouds. Plus, the final trained model of 3DGS at the 20000-th iteration contains more redundant Gaussians that are unnecessary to represent the 3D structure of the foot, which reduces the model's inference speed. In contrast, our X-Gaussian equipped with the proposed DRR shows faster and better convergence. In particular, as early as the 1000-th iteration, our radiative Gaussian point clouds have essentially formed the basic shape of the foot. Besides, the final trained X-Gaussian at the 20000-th iteration can better represent the 3D geometry and more accurate structural contents than the original 3DGS. + +Additionally, in Fig. 6, we visualize the rendered novel projections with the same angle $\phi$ of NAF and our X-Gaussian at 20s, 60s, and 180s of the training process on the scene of chest. NAF produces blurry images with severe noises on the background regions within the first three minutes of the training. In contrast, our X-Gaussian can reconstruct clearer structural details like the ribs and blood vessels with cleaner background of the chest at the first minute of the training. + +![](images/0736b903be6058f5f5a2d54ceffdc55e2f223d8a0b811795ae658ac8ba443dea.jpg) +NAF at 20s + +![](images/80050c19b38b34b78e5727fbc48268c4df08f044a17e65cf254c034226472562.jpg) +NAF at 60s + +![](images/2fc4386cb4049a3985d704dfc2a99f5a40a441bcd58eb20415092ea092d1ca60.jpg) +NAF at 180s + +![](images/cb5fbba27a6b1712ba952f139e001024fe0a215cef1d0f59b41bf07f06b227d1.jpg) +Ground Truth + +![](images/9bd14f35032f1bf570daab932c842dec9da505c590669d8829e8fd99e3e909b0.jpg) +X-Gaussian at 20s + +![](images/432a86bc726fa19eebd1e9632e74775fe693a390ef901c499e8a60eeafa95415.jpg) +X-Gaussian at 60s + +![](images/e017efe17023ff92f513f599338661ae4cfa4d56190a242a313309412bffc47f.jpg) +X-Gaussian at 180s + +![](images/35cf8f70b02f21ebca3633d16cbee4ca267e15cd7de87c303bc8921e2f16518c.jpg) +Ground Truth + +![](images/f051d1c3945bd126fa7106fb93b0c540bc07e0069f5239e01f306ef58925e6b6.jpg) +Fig.6: Convergence analysis of NAF [60] vs. our X-Gaussian. We visualize the rendered projections at 20s, 60s, and 180s of training. Our X-Gaussian shows faster and better convergence. +Fig. 7: Analysis of covariance. The mean covariance of 3D Gaussians on different scenes decreases when the number of training views increases. + +Analysis of Covariance. We study how the shape of the 3D Gaussian point cloud changes with the number of training views in Fig. 7. As the number of training views increases, the mean covariance of 3D Gaussians that control the size of Gaussian point clouds decreases. This indicates that the 3D Gaussian point clouds gradually change from coarse to fine, thereby being more capable of representing fine-grained structures, such as a small tumor in the abdomen. + +# 5 Limitation + +The main limitation of our method is that its memory cost is non-trivial, about 51.25 MB on average. It is challenging to deploy it on low-RAM mobile devices. + +# 6 Conclusion + +In this paper, we propose the first 3DGS-based framework, X-Gaussian, for X-ray novel view synthesis. Firstly, we design a radiative Gaussian point cloud model based on X-ray imaging properties. This model excludes the influence of view direction when fitting radiation intensity. For this model, we develop a GPU-friendly differentiable radiative rasterization CUDA kernel that renders projections at a faster speed than RGB rasterization. Secondly, we customize an initialization strategy, ACUI, that does not need to execute the SfM algorithm. Instead, ACUI uses the parameters of X-ray scanner to compute the extrinsic and intrinsic matrices, and then uniformly samples center points for 3D Gaussians within a cuboid enclosing the scanned object. Experiments demonstrate that our X-Gaussian significantly outperforms SOTA methods by over $6.5\mathrm{dB}$ while enjoying $73\times$ faster inference speed and only requiring $15\%$ of training time. + +Acknowledgements: This work was supported by the Lustgarten Foundation for Pancreatic Cancer Research and the Patrick J. McGovern Foundation Award. + +# References + +1. Andersen, A.H., Kak, A.C.: Simultaneous algebraic reconstruction technique (sart): a superior implementation of the art algorithm. Ultrasonic imaging (1984) +2. Armato III, S.G., McLennan, G., Bidaut, L., McNitt-Gray, M.F., Meyer, C.R., Reeves, A.P., Zhao, B., Aberle, D.R., Henschke, C.I., Hoffman, E.A., et al.: The lung image database consortium (lidc) and image database resource initiative (idri): a completed reference database of lung nodules on ct scans. Medical physics (2011) +3. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: ICCV (2021) +4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In: CVPR (2022) +5. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-nerf: Anti-aliased grid-based neural radiance fields. In: ICCV (2023) +6. Biguri, A., Dosanjh, M., Hancock, S., Soleimani, M.: Tigre: a matlab-gpu toolbox for cbct image reconstruction. Biomedical Physics & Engineering Express (2016) +7. Boone, J., Shah, N., Nelson, T.: A comprehensive analysis of coefficients for pendant-geometry cone-beam breast computed tomography. Medical physics (2004) +8. Boone, J.M., Nelson, T.R., Lindfors, K.K., Seibert, J.A.: Dedicated breast ct: radiation dose and image quality evaluation. Radiology (2001) +9. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: ECCV (2022) +10. Chen, B., Ning, R.: Cone-beam volume ct breast imaging: Feasibility study. Medical physics (2002) +1. Chen, Z., Funkhouser, T., Hedman, P., Tagliasacchi, A.: Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In: CVPR (2023) +2. Cormack, A.M.: Representation of a function by its line integrals, with some radiological applications. Journal of applied physics (1963) +3. Cormack, A.M.: Representation of a function by its line integrals, with some radiological applications. ii. Journal of Applied Physics (1964) +4. Corona-Figueroa, A., Frawley, J., Bond-Taylor, S., Bethapudi, S., Shum, H.P., Willcocks, C.G.: Mednerf: Medical neural radiance fields for reconstructing 3d-aware ct-projections from a single x-ray. In: International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC) (2022) +5. Elbakri, I.A., Fessler, J.A.: Segmentation-free statistical image reconstruction for polyenergetic x-ray computed tomography with experimental validation. Physics in Medicine & Biology (2003) +6. Feldkamp, L.A., Davis, L.C., Kress, J.W.: Practical cone-beam algorithm. Josa a (1984) +7. Guide, D.: Cuda c programming guide. NVIDIA (2013) +8. Hounsfield, G.N.: Computerized transverse axial scanning (tomography): Part 1. description of system. The British journal of radiology (1973) +9. Hounsfield, G.N.: Computed medical imaging. Science (1980) +20. Hu, S., Liu, Z.: Gauhuman: Articulated gaussian splatting from monocular human videos. arXiv preprint arXiv: (2023) +21. Hu, T., Liu, S., Chen, Y., Shen, T., Jia, J.: Efficientnerf efficient neural radian fields. In: CVPR (2023) + +22. Hu, W., Wang, Y., Ma, L., Yang, B., Gao, L., Liu, X., Ma, Y.: Tri-miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields. In: ICCV (2023) +23. Jiang, Y., Tu, J., Liu, Y., Gao, X., Long, X., Wang, W., Ma, Y.: Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. arXiv preprint arXiv:2311.17977 (2023) +24. Keetha, N., Karhade, J., Jatavallabhula, K.M., Yang, G., Scherer, S., Ramanan, D., Luiten, J.: Splatam: Splat, track & map 3d gaussians for dense rgb-d slam. arXiv preprint arXiv:2312.02126 (2023) +25. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (2023) +26. Kingma, D.P., Ba, J.L.: Adam: A method for stochastic optimization. In: ICLR (2015) +27. Klacansky, P.: Scientific visualization datasets (2022), https://klacansky.com/open-scivis-datasets/ +28. Kocabas, M., Chang, J.H.R., Gabriel, J., Tuzel, O., Ranjan, A.: Hugs: Human gaussian splats. arXiv preprint arXiv:2311.17910 (2023) +29. Kopanas, G., Philip, J., Leimkuhler, T., Drettakis, G.: Point-based neural rendering with per-view optimization. In: Computer Graphics Forum (2021) +30. Li, R., Gao, H., Tancik, M., Kanazawa, A.: Nerfacc: Efficient sampling accelerates nerfs. In: ICCV (2023) +31. Liang, Y., Yang, X., Lin, J., Li, H., Xu, X., Chen, Y.: Luciddreamer: Towards high-fidelity text-to-3d generation via interval score matching. arXiv preprint arXiv:2311.11284 (2023) +32. Liang, Z., Zhang, Q., Feng, Y., Shan, Y., Jia, K.: Gs-ir: 3d gaussian splatting for inverse rendering. arXiv preprint arXiv:2311.16473 (2023) +33. Liu, X., Zhan, X., Tang, J., Shan, Y., Zeng, G., Lin, D., Liu, X., Liu, Z.: Human-gaussian: Text-driven 3d human generation with gaussian splatting. arXiv preprint arXiv:2311.17061 (2023) +34. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713 (2023) +35. Manglos, S.H., Gagne, G.M., Krol, A., Thomas, F.D., Narayanaswamy, R.: Transmission maximum-likelihood reconstruction with ordered subsets for cone beam ct. Physics in Medicine & Biology (1995) +36. Matsuki, H., Murai, R., Kelly, P.H., Davison, A.J.: Gaussian splatting slam. arXiv preprint arXiv:2312.06741 (2023) +37. Mildenhall, B., Srinivasan, P., Tancik, M., Barron, J., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) +38. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM ToG (2022) +39. Pan, J., Zhou, T., Han, Y., Jiang, M., et al.: Variable weighted ordered subset image reconstruction algorithm. International Journal of Biomedical Imaging (2006) +40. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: NeurIPS (2019) +41. Reiser, C., Szeliski, R., Verbin, D., Srinivasan, P., Mildenhall, B., Geiger, A., Barron, J., Hedman, P.: Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. TOG (2023) +42. Rückert, D., Wang, Y., Li, R., Idoughi, R., Heidrich, W.: Neat: Neural adaptive tomography. TOG (2022) + +43. Sauer, K., Bouman, C.: A local update strategy for iterative reconstruction from projections. TIP (1993) +44. Scarfe, W.C., Farman, A.G., Sukovic, P., et al.: Clinical applications of cone-beam computed tomography in dental practice. Journal-Canadian Dental Association (2006) +45. Sidky, E.Y., Pan, X.: Image reconstruction in circular cone-beam computed tomography by constrained, total-variation minimization. Physics in Medicine & Biology (2008) +46. Snavely, N., Seitz, S.M., Szeliski, R.: Photo tourism: exploring photo collections in 3d. In: SIGGRAPH (2006) +47. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023) +48. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: CVPR (2022) +49. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncell, E.P.: Image quality assessment: from error visibility to structural similarity. TIP (2004) +50. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528 (2023) +51. Xie, T., Zong, Z., Qiu, Y., Li, X., Feng, Y., Yang, Y., Jiang, C.: Physgaus-sian: Physics-integrated 3d gaussians for generative dynamics. arXiv preprint arXiv:2311.12198 (2023) +52. Yan, C., Qu, D., Wang, D., Xu, D., Wang, Z., Zhao, B., Li, X.: Gs-slam: Dense visual slam with 3d gaussian splatting. arXiv preprint arXiv:2311.11700 (2023) +53. Yang, Z., Yang, H., Pan, Z., Zhu, X., Zhang, L.: Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642 (2023) +54. Yariv, L., Hedman, P., Reiser, C., Verbin, D., Srinivasan, P.P., Szeliski, R., Barron, J.T., Mildenhall, B.: Bakedsdf: Meshing neural sdfs for real-time view synthesis. In: SIGGRAPH (2023) +55. Yi, T., Fang, J., Wu, G., Xie, L., Zhang, X., Liu, W., Tian, Q., Wang, X.: Gaussian dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529 (2023) +56. Yu, L., Zou, Y., Sidky, E.Y., Pelizzari, C.A., Munro, P., Pan, X.: Region of interest reconstruction from truncated data in circular cone-beam ct. TMI (2006) +57. Yugay, V., Li, Y., Gevers, T., Oswald, M.R.: Gaussian-slam: Photo-realistic dense slam with gaussian splatting. arXiv preprint arXiv:2312.10070 (2023) +58. Zang, G., Idoughi, R., Li, R., Wonka, P., Heidrich, W.: Intratomo: self-supervised learning-based tomography via sinogram synthesis and prediction. In: CVPR (2021) +59. Zbijewski, W., Defrise, M., Viergever, M.A., Beekman, F.J.: Statistical reconstruction for x-ray ct systems with non-continuous detectors. Physics in Medicine & Biology (2006) +60. Zha, R., Zhang, Y., Li, H.: Naf: neural attenuation fields for sparse-view cbct reconstruction. In: MICCAI (2022) +61. Zhang, T., Yu, H.X., Wu, R., Feng, B.Y., Zheng, C., Snavely, N., Wu, J., Freeman, W.T.: PhysDreamer: Physics-based interaction with 3d objects via video generation. In: ECCV (2024) +62. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: Ewa volume splatting. In: Proceedings Visualization, 2001. VIS'01. IEEE (2001) \ No newline at end of file diff --git a/2024/Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis/images.zip b/2024/Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f414403fd0e8c3af8038b5441b6d19d7e8e73c50 --- /dev/null +++ b/2024/Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0fc7c5c53a90a1fa0f5e5ad16b15f37c99c1947da76ee43c008ba50494cbe54b +size 554155 diff --git a/2024/Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis/layout.json b/2024/Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fed8e4c123477b4bd65b7cbb02dd0cddc33c138f --- /dev/null +++ b/2024/Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis/layout.json @@ -0,0 +1,10976 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 165, + 112, + 449, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 112, + 449, + 148 + ], + "spans": [ + { + "bbox": [ + 165, + 112, + 449, + 148 + ], + "type": "text", + "content": "Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "spans": [ + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "content": "Yuanhao Cai " + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Yixun Liang " + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Jiahao Wang " + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Angtian Wang " + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Yulun Zhang " + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{3,\\dagger}" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Xiaokang Yang " + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Zongwei Zhou " + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "content": ", and Alan Yuille " + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 148, + 201, + 465, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 201, + 465, + 214 + ], + "spans": [ + { + "bbox": [ + 148, + 201, + 465, + 214 + ], + "type": "text", + "content": "1 Johns Hopkins University, 2 HKUST(GZ), 3 Shanghai Jiao Tong University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 238, + 455, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 238, + 455, + 437 + ], + "spans": [ + { + "bbox": [ + 159, + 238, + 455, + 437 + ], + "type": "text", + "content": "Abstract. X-ray is widely applied for transmission imaging due to its stronger penetration than natural light. When rendering novel view X-ray projections, existing methods mainly based on NeRF suffer from long training time and slow inference speed. In this paper, we propose a 3D Gaussian splatting-based method, namely X-Gaussian, for X-ray novel view synthesis. Firstly, we redesign a radiative Gaussian point cloud model inspired by the isotropic nature of X-ray imaging. Our model excludes the influence of view direction when learning to predict the radiation intensity of 3D points. Based on this model, we develop a Differentiable Radiative Rasterization (DRR) with CUDA implementation. Secondly, we customize an Angle-pose Cuboid Uniform Initialization (ACUI) strategy that directly uses the parameters of the X-ray scanner to compute the camera information and then uniformly samples point positions within a cuboid enclosing the scanned object. Experiments show that our X-Gaussian outperforms state-of-the-art methods by 6.5 dB while enjoying less than " + }, + { + "bbox": [ + 159, + 238, + 455, + 437 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 159, + 238, + 455, + 437 + ], + "type": "text", + "content": " training time and over " + }, + { + "bbox": [ + 159, + 238, + 455, + 437 + ], + "type": "inline_equation", + "content": "73\\times" + }, + { + "bbox": [ + 159, + 238, + 455, + 437 + ], + "type": "text", + "content": " inference speed. The application on CT reconstruction also reveals the practical values of our method. Code is at https://github.com/caiyuanhao1998/X-Gaussian" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 446, + 441, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 446, + 441, + 458 + ], + "spans": [ + { + "bbox": [ + 160, + 446, + 441, + 458 + ], + "type": "text", + "content": "Keywords: Medical Imaging " + }, + { + "bbox": [ + 160, + 446, + 441, + 458 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 446, + 441, + 458 + ], + "type": "text", + "content": " View Synthesis " + }, + { + "bbox": [ + 160, + 446, + 441, + 458 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 446, + 441, + 458 + ], + "type": "text", + "content": " Gaussian Splatting" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 492, + 230, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 492, + 230, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 492, + 230, + 504 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 515, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 515, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 515, + 482, + 648 + ], + "type": "text", + "content": "X-ray novel view synthesis (NVS) aims to create X-ray projections of an object from new viewpoints that are not originally captured, using only existing projections scanned from different view directions. As we know, X-ray has stronger penetrating power to capture internal structures of imaged objects and is therefore widely applied in medical imaging [12,13,15,18,19,43]. Nonetheless, X-ray is harmful to human body due to its powerful ionizing radiation, especially when the dose of X-ray increases. Improving NVS techniques can help reduce the exposure to X-rays and provide comprehensive viewpoints of imaged parts for doctors and downstream tasks such as sparse-view CT reconstruction. Hence, X-ray NVS is very important and valuable. We study this task in the circular cone beam X-ray scanning scenario [7,8,10,35,39,44,56,59]." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 654, + 254, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 254, + 666 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 254, + 666 + ], + "type": "inline_equation", + "content": "\\dagger =" + }, + { + "bbox": [ + 141, + 654, + 254, + 666 + ], + "type": "text", + "content": " corresponding authors." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 130, + 116, + 476, + 245 + ], + "blocks": [ + { + "bbox": [ + 130, + 116, + 476, + 245 + ], + "lines": [ + { + "bbox": [ + 130, + 116, + 476, + 245 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 476, + 245 + ], + "type": "image", + "image_path": "97b2aca403e0517f00163ab81701ee073b786d63c5262cb1690b7816622656fd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 267, + 482, + 323 + ], + "lines": [ + { + "bbox": [ + 130, + 267, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 482, + 323 + ], + "type": "text", + "content": "Fig. 1: Point cloud visualization of the original 3DGS [25] (top) and our X-Gaussian (bottom). We visualize the positions and opacities of the Gaussian point clouds at different training iterations. We also visualize the volume of foot as a reference. Note that the volume is not the ground truth of point clouds. Our X-Gaussian can better represent the detailed structures than 3DGS, showing faster and better convergence." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 351, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 482, + 460 + ], + "type": "text", + "content": "Existing methods are mainly based on neural radiance fields (NeRF) [37]. They usually employ a multi-layer perceptron (MLP) to learn the mapping from the point position to the radiodensity and then create projections by volume rendering along rays. This ray tracing scheme is time-consuming because it needs to sample many 3D points and then compute them for every single ray, slowing down the training and inference processes. Even the recent most efficient NeRF-based method [60] still requires over an hour for training and yields suboptimal results at a slow inference speed of 2 fps. This increases the waiting time of patients and doctors, leading to low diagnostic efficiency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": "Recently, 3D Gaussian splattering (3DGS) [25] has demonstrated promising reconstruction quality while enjoying much faster inference speed than NeRF-based algorithms in RGB domain, which motivates us to follow this technical route. However, due to the fundamental differences between X-ray and natural light imaging, directly applying the original 3DGS to X-ray NVS may encounter two issues. Firstly, the spherical harmonics (SH) in RGB 3DGS is not suitable for modeling the X-ray radiation intensity of 3D points. Specifically, natural light imaging relies on the reflection off the surface. The color of a 3D point is anisotropic and view-dependent. Based on this nature, the original Gaussian point cloud model uses SH to fit the illumination distribution. In contrast, X-rays penetrate the object and attenuate, thereby forming an image. Given specific X-rays, the radiation intensity of a 3D point depends on its radiodensity and is independent to the view direction, which means the point radiation intensity is isotropic. Secondly, the original point cloud initialization algorithm, structure-from-motion (SfM) [46], is also not suitable for X-ray imaging. Compared to RGB images, X-ray images are grayscale and their contrast is lower. Additionally, different layers of an object may overlap on the same position of the projection" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 243, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 243, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 243, + 100 + ], + "type": "text", + "content": "Yuanhao Cai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "due to the transmission imaging nature of X-rays. These two problems degrade the accuracy of feature detection and matching in SfM. Meanwhile, running SfM is time-consuming, which prolongs the training process of Gaussian point clouds." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 379 + ], + "type": "text", + "content": "To address the above issues, we propose a novel 3DGS-based method, X-Gaussian, for X-ray NVs. Our X-Gaussian composes two key techniques. Firstly, we redesign a radiative Gaussian point cloud model inspired by the isotropic property of X-ray imaging. We present a Radiation Intensity Response Function (RIRF) to replace the SH function of the original 3DGS. Different from SH, our RIRF excludes the influence of view direction. To this end, it adopts the inner product between a learnable vector representing the inherent point features and a set of basis weights to fit the radiation intensity of a 3D point. Based on this point cloud model, we further develop a Differentiable Radiative Rasterization (DRR) with a CUDA implementation to render novel projections. Secondly, we customize an Angle-pose Cuboid Uniform Initialization (ACUI) strategy for camera calibration parameters and Gaussian point clouds. Our ACUI first exploits the parameters of the X-ray scanner to compute the intrinsic and extrinsic matrices. Then we set up a cuboid that can completely enclose the scanned object. Within this cuboid, we uniformly sample 3D points at intervals to initialize the center positions of the Gaussian point clouds. Free from running the SfM algorithm, our ACUI significantly reduces the training time. Equipped with the two proposed techniques, our X-Gaussian outperforms SOTA methods by 6.5 dB while enjoying " + }, + { + "bbox": [ + 130, + 152, + 482, + 379 + ], + "type": "inline_equation", + "content": "73 \\times" + }, + { + "bbox": [ + 130, + 152, + 482, + 379 + ], + "type": "text", + "content": " inference speed and " + }, + { + "bbox": [ + 130, + 152, + 482, + 379 + ], + "type": "inline_equation", + "content": "7 \\times" + }, + { + "bbox": [ + 130, + 152, + 482, + 379 + ], + "type": "text", + "content": " training speed." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 380, + 443, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 380, + 443, + 391 + ], + "spans": [ + { + "bbox": [ + 146, + 380, + 443, + 391 + ], + "type": "text", + "content": "The main contributions of this work can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 399, + 479, + 517 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 399, + 479, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 479, + 434 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 479, + 434 + ], + "type": "text", + "content": "- We propose a novel 3D Gaussian splatting-based framework, X-Gaussian, for X-ray novel view synthesis. To our knowledge, this is the first attempt to explore the potential of Gaussian splatting in X-ray neural rendering." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 435, + 479, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 435, + 479, + 458 + ], + "spans": [ + { + "bbox": [ + 137, + 435, + 479, + 458 + ], + "type": "text", + "content": "- We design a radiative Gaussian point cloud model with a differentiable radiative rasterization based on the isotropic nature of X-ray imaging." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 459, + 479, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 459, + 479, + 482 + ], + "spans": [ + { + "bbox": [ + 137, + 459, + 479, + 482 + ], + "type": "text", + "content": "- We present an angle-pose cuboid uniform initialization strategy for Gaussian point clouds and camera calibration in circular cone beam X-ray scanning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 483, + 479, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 483, + 479, + 517 + ], + "spans": [ + { + "bbox": [ + 137, + 483, + 479, + 517 + ], + "type": "text", + "content": "- Our X-Gaussian significantly outperforms SOTA NeRF-based methods with much faster speed. Experiments also show that our method can improve the performance of sparse-view CT reconstruction, showing its practical values." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 536, + 237, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 237, + 550 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 237, + 550 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 562, + 274, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 562, + 274, + 573 + ], + "spans": [ + { + "bbox": [ + 132, + 562, + 274, + 573 + ], + "type": "text", + "content": "2.1 Neural Radiance Field" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "NeRF [37] learns an implicit neural scene representation of color and volume density, given the position of a 3D point and view direction. It has achieved great success in NVS and inspired an explosion of follow-up papers to improve its quality [3-5,22,48] and speed [9,11,21,30,38,41,54]. For example, Instant-NGP [38] adopts hash tables as the encoder to allow small MLP for fast training and inference. Some later works extend the application domain of NeRF from natural light to X-rays [14,58,60]. For instance, NAF [60] follows the settings" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 161, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 161, + 91, + 448, + 103 + ], + "type": "text", + "content": "Radiative Gaussian Splitting for Efficient X-ray Novel View Synthesis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 129, + 118, + 479, + 224 + ], + "blocks": [ + { + "bbox": [ + 129, + 118, + 479, + 224 + ], + "lines": [ + { + "bbox": [ + 129, + 118, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 129, + 118, + 479, + 224 + ], + "type": "image", + "image_path": "16f43c0f96a51489bfb4872285d3dfff2b5ec580c759d311106f60a8a62b4425.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 251, + 482, + 308 + ], + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 308 + ], + "type": "text", + "content": "Fig. 2: Pipeline of our method. (a) Angle-posed Cuboid Uniform Initialization (ACUI) strategy uses the parameters of X-ray scanner to compute intrinsic and extrinsic matrices, and samples center points for 3D Gaussians. (b) Our radiative Gaussian point cloud model learns to predict the radiation intensity of 3D points. (c) Based on our Gaussian model, we develop a GPU-friendly Differentiable Radiative Rasterization (DRR)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 331, + 482, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 369 + ], + "type": "text", + "content": "of Instant-NGP to learn the implicit mapping from 3D position to attenuation. Yet, the ray tracing and volume rendering schemes are time-consuming, which limits the training and inference speed of NeRF-based X-ray NVS algorithms." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 385, + 257, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 385, + 257, + 398 + ], + "spans": [ + { + "bbox": [ + 131, + 385, + 257, + 398 + ], + "type": "text", + "content": "2.2 Gaussian Splitting" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 405, + 482, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 405, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 482, + 514 + ], + "type": "text", + "content": "3DGS [25] represents scenes using millions of 3D Gaussian point clouds. This approach is fundamentally different from NeRF-based algorithms by employing an explicit representation coupled with highly parallelized rasterization workflows. These features enable more efficient computation and rendering processes. Hence, 3DGS has achieved great success in several fields, including 3D Generation [20,28,31,33,47,55,61], Dynamic Scene Modeling [34,50,53], SLAM [24,36,52,57], Inverse Rendering [23,32,51], etc. However, most applications of 3DGS are focused on natural scenes with RGB colors. The potential of 3DGS in X-ray imaging still remains under-explored. Our goal is to fill this research gap." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 532, + 202, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 532, + 202, + 544 + ], + "spans": [ + { + "bbox": [ + 131, + 532, + 202, + 544 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "The pipeline of our X-Gaussian is shown in Fig. 2. Firstly, we design an Angle-posed Cuboid Uniform Initialization (ACUI) to compute the intrinsic and extrinsic matrices from the parameters of X-ray scanner, as illustrated in Fig. 2 (a). Then ACUI uniformly samples 3D points within a cuboid that can completely enclose the scanned object to initialize the center positions of our radiative Gaussian point clouds in Fig. 2 (b). Given a view direction, the 3D point clouds undergo our Differentiable Radiative Rasterization (DRR) to derive the rendered image, as depicted in Fig. 2 (c). In this section, we will introduce our radiative Gaussian point cloud model and DRR processing first and then the ACUI strategy." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 243, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 243, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 243, + 100 + ], + "type": "text", + "content": "Yuanhao Cai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 118, + 294, + 208 + ], + "blocks": [ + { + "bbox": [ + 134, + 118, + 294, + 208 + ], + "lines": [ + { + "bbox": [ + 134, + 118, + 294, + 208 + ], + "spans": [ + { + "bbox": [ + 134, + 118, + 294, + 208 + ], + "type": "image", + "image_path": "3265cdb3c1db691484126c0b557ae8f76e0819232c02b24238280df5e57b12b9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 211, + 266, + 220 + ], + "lines": [ + { + "bbox": [ + 165, + 211, + 266, + 220 + ], + "spans": [ + { + "bbox": [ + 165, + 211, + 266, + 220 + ], + "type": "text", + "content": "(a) Original RGB Gaussian Point Cloud" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 329, + 117, + 476, + 208 + ], + "blocks": [ + { + "bbox": [ + 329, + 117, + 476, + 208 + ], + "lines": [ + { + "bbox": [ + 329, + 117, + 476, + 208 + ], + "spans": [ + { + "bbox": [ + 329, + 117, + 476, + 208 + ], + "type": "image", + "image_path": "d8fb9744312a4a78a7b5c5945597d095b461df18b1a031ce95131cf733984593.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 211, + 451, + 220 + ], + "lines": [ + { + "bbox": [ + 349, + 211, + 451, + 220 + ], + "spans": [ + { + "bbox": [ + 349, + 211, + 451, + 220 + ], + "type": "text", + "content": "(b) Our Radiative Gaussian Point Cloud" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 245, + 480, + 300 + ], + "lines": [ + { + "bbox": [ + 131, + 245, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 131, + 245, + 480, + 300 + ], + "type": "text", + "content": "Fig. 3: Comparison between the Gaussian point cloud models of the original 3DGS and our X-Gaussian. (a) The original RGB Gaussian point cloud model uses spherical harmonics (SH) to simulate the anisotropic natural light distribution and view-dependent color. (b) Our radiative Gaussian point cloud model employs the weighted sum of point features to fit the isotropic X-ray penetration and view-independent radiation intensity." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 324, + 358, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 324, + 358, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 324, + 358, + 335 + ], + "type": "text", + "content": "3.1 Radiative Gaussian Point Cloud Model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 343, + 459, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 459, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 459, + 355 + ], + "type": "text", + "content": "An object can be represented by a set of basic Gaussian point clouds " + }, + { + "bbox": [ + 132, + 343, + 459, + 355 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 132, + 343, + 459, + 355 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 220, + 365, + 480, + 378 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 365, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 220, + 365, + 480, + 378 + ], + "type": "interline_equation", + "content": "\\mathcal {G} = \\left\\{G _ {i} \\left(\\boldsymbol {\\mu} _ {i}, \\boldsymbol {\\Sigma} _ {i}, \\alpha_ {i}\\right) \\mid i = 1, 2, \\dots , N _ {p} \\right\\}, \\tag {1}", + "image_path": "91ba0910b2a808d320be52e57ba66c4961dde9541b74635c26d73bb14a82bb77.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 386, + 479, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 386, + 479, + 457 + ], + "spans": [ + { + "bbox": [ + 131, + 386, + 479, + 457 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 386, + 479, + 457 + ], + "type": "inline_equation", + "content": "G_{i}" + }, + { + "bbox": [ + 131, + 386, + 479, + 457 + ], + "type": "text", + "content": " refers to the " + }, + { + "bbox": [ + 131, + 386, + 479, + 457 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 386, + 479, + 457 + ], + "type": "text", + "content": "-th Gaussian point cloud. Its center position, covariance, and opacity are defined as " + }, + { + "bbox": [ + 131, + 386, + 479, + 457 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_{i} \\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 131, + 386, + 479, + 457 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 386, + 479, + 457 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_{i} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 131, + 386, + 479, + 457 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 386, + 479, + 457 + ], + "type": "inline_equation", + "content": "\\alpha_{i} \\in \\mathbb{R}" + }, + { + "bbox": [ + 131, + 386, + 479, + 457 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 386, + 479, + 457 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_{i}" + }, + { + "bbox": [ + 131, + 386, + 479, + 457 + ], + "type": "text", + "content": " is represented by a rotation matrix " + }, + { + "bbox": [ + 131, + 386, + 479, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 131, + 386, + 479, + 457 + ], + "type": "text", + "content": " and a scaling matrix " + }, + { + "bbox": [ + 131, + 386, + 479, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_i \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 131, + 386, + 479, + 457 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 131, + 386, + 479, + 457 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_{i} = \\mathbf{R}_{i}\\mathbf{S}_{i}\\mathbf{S}_{i}^{\\top}\\mathbf{R}_{i}^{\\top}" + }, + { + "bbox": [ + 131, + 386, + 479, + 457 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 386, + 479, + 457 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_{i}, \\pmb{\\Sigma}_{i}, \\alpha_{i}, \\mathbf{R}_{i}" + }, + { + "bbox": [ + 131, + 386, + 479, + 457 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 386, + 479, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_i" + }, + { + "bbox": [ + 131, + 386, + 479, + 457 + ], + "type": "text", + "content": " are learnable parameters. Besides these basic attributes, each Gaussian point cloud also employs additional learnable parameters to fit different imaging scenarios, e.g., natural light imaging and X-ray imaging." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 458, + 480, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 458, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 131, + 458, + 480, + 542 + ], + "type": "text", + "content": "We first review the original RGB Gaussian point cloud model [25] in natural light imaging. As shown in Fig. 3 (a), the color of a 3D point is represented by spherical harmonics (SH). The point color is anisotropic and changes with the view direction. Each Gaussian point cloud learns to predict the SH coefficients " + }, + { + "bbox": [ + 131, + 458, + 480, + 542 + ], + "type": "inline_equation", + "content": "\\mathbf{k} = \\{k_l^m |0\\leq l\\leq L, - l\\leq m\\leq l\\} \\in \\mathbb{R}^{(L + 1)^2\\times 3}" + }, + { + "bbox": [ + 131, + 458, + 480, + 542 + ], + "type": "text", + "content": ", where each " + }, + { + "bbox": [ + 131, + 458, + 480, + 542 + ], + "type": "inline_equation", + "content": "k_{l}^{m}\\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 131, + 458, + 480, + 542 + ], + "type": "text", + "content": " is a set of 3 coefficients corresponding to the RGB components. " + }, + { + "bbox": [ + 131, + 458, + 480, + 542 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 131, + 458, + 480, + 542 + ], + "type": "text", + "content": " is the degree of SH. Then the point color " + }, + { + "bbox": [ + 131, + 458, + 480, + 542 + ], + "type": "inline_equation", + "content": "\\mathbf{c}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 131, + 458, + 480, + 542 + ], + "type": "text", + "content": " at the view direction " + }, + { + "bbox": [ + 131, + 458, + 480, + 542 + ], + "type": "inline_equation", + "content": "\\mathbf{d} = (\\theta ,\\phi)" + }, + { + "bbox": [ + 131, + 458, + 480, + 542 + ], + "type": "text", + "content": " is derived by" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 233, + 551, + 480, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 551, + 480, + 583 + ], + "spans": [ + { + "bbox": [ + 233, + 551, + 480, + 583 + ], + "type": "interline_equation", + "content": "\\mathbf {c} (\\mathbf {d}, \\mathbf {k}) = \\sum_ {l = 0} ^ {L} \\sum_ {m = - l} ^ {l} k _ {l} ^ {m} Y _ {l} ^ {m} (\\theta , \\phi), \\tag {2}", + "image_path": "9fad02609b55999896d1ffed112898f409972426b4e10275713d43061c453657.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 594, + 479, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 594, + 479, + 616 + ], + "spans": [ + { + "bbox": [ + 131, + 594, + 479, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 594, + 479, + 616 + ], + "type": "inline_equation", + "content": "Y_{l}^{m}:\\mathbb{S}^{2}\\to \\mathbb{R}" + }, + { + "bbox": [ + 131, + 594, + 479, + 616 + ], + "type": "text", + "content": " is the SH function that maps points on the sphere to real numbers. Please refer to the supplementary for its detailed formulation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 481, + 665 + ], + "type": "text", + "content": "Although 3DGS [25] achieves fast inference speed and good performance in natural light imaging, the RGB Gaussian point cloud model is not suitable for X-ray scenarios due to the fundamental differences between natural light imaging and X-ray imaging. Natural light imaging relies on the reflection off the surface of" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 161, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 161, + 91, + 448, + 102 + ], + "type": "text", + "content": "Radiative Gaussian Splitting for Efficient X-ray Novel View Synthesis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": "object. The anisotropic color modeled by SH is view-dependent. e.g., in Fig. 3 (a), the point color is blue from the left viewpoint and green from the right viewpoint. In contrast, X-ray imaging is based on the attenuation when penetrating the object. The degree of attenuation depends on the isotropic radiodensity property. Thus, the radiation intensity of a 3D point is view-independent." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 480, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 480, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 480, + 260 + ], + "type": "text", + "content": "In light of the above analysis, we redesign our radiative Gaussian point cloud model. Different from the original 3DGS that uses SH to fit the color information for each point, our model introduces a Radiation Intensity Response Function (RIRF) to predict the radiation intensity of the 3D point. As illustrated in Fig. 3 (b), each Gaussian point cloud learns a feature vector " + }, + { + "bbox": [ + 130, + 176, + 480, + 260 + ], + "type": "inline_equation", + "content": "\\mathbf{f} \\in \\mathbb{R}^{N_f}" + }, + { + "bbox": [ + 130, + 176, + 480, + 260 + ], + "type": "text", + "content": " to represent its inherent radiative properties. Subsequently, the radiation intensity " + }, + { + "bbox": [ + 130, + 176, + 480, + 260 + ], + "type": "inline_equation", + "content": "\\mathbf{i} \\in \\mathbb{R}" + }, + { + "bbox": [ + 130, + 176, + 480, + 260 + ], + "type": "text", + "content": " of the center point of a 3D Gaussian at any view direction is modeled by RIRF as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 233, + 269, + 480, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 269, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 233, + 269, + 480, + 281 + ], + "type": "interline_equation", + "content": "\\mathbf {i} (\\mathbf {f}) = \\operatorname {R I R F} (\\mathbf {f}) = \\operatorname {S i g m o i d} (\\boldsymbol {\\lambda} \\odot \\mathbf {f}), \\tag {3}", + "image_path": "b456b4ac949aa36b0f8e93b5aa2f0f9da6c44fbbccffeb32b45512446c01f4a2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 289, + 479, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 289, + 479, + 325 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 479, + 325 + ], + "type": "text", + "content": "where the Sigmoid function activates and normalizes the radiation intensity. " + }, + { + "bbox": [ + 131, + 289, + 479, + 325 + ], + "type": "inline_equation", + "content": "\\lambda \\in \\mathbb{R}^{N_f}" + }, + { + "bbox": [ + 131, + 289, + 479, + 325 + ], + "type": "text", + "content": " is a set of constant weights controlling the importance of each component of " + }, + { + "bbox": [ + 131, + 289, + 479, + 325 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 131, + 289, + 479, + 325 + ], + "type": "text", + "content": ". Then the set of our radiative Gaussian point clouds " + }, + { + "bbox": [ + 131, + 289, + 479, + 325 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_x" + }, + { + "bbox": [ + 131, + 289, + 479, + 325 + ], + "type": "text", + "content": " is formulated as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 216, + 334, + 480, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 334, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 216, + 334, + 480, + 346 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {x} = \\left\\{G _ {i} \\left(\\boldsymbol {\\mu} _ {i}, \\boldsymbol {\\Sigma} _ {i}, \\alpha_ {i}, \\mathbf {f} _ {i}\\right) \\mid i = 1, 2, \\dots , N _ {p} \\right\\}, \\tag {4}", + "image_path": "cbf3944f1257165d234dc6accf23b75077cde7420137d08568b8e5f3af7e69dc.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 354, + 479, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 354, + 479, + 414 + ], + "spans": [ + { + "bbox": [ + 131, + 354, + 479, + 414 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 354, + 479, + 414 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i\\in \\mathbb{R}^{N_f}" + }, + { + "bbox": [ + 131, + 354, + 479, + 414 + ], + "type": "text", + "content": " denotes the feature vector of the " + }, + { + "bbox": [ + 131, + 354, + 479, + 414 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 354, + 479, + 414 + ], + "type": "text", + "content": "-th Gaussian point cloud. Please note that Eq. (3) excludes the influence of the view direction " + }, + { + "bbox": [ + 131, + 354, + 479, + 414 + ], + "type": "inline_equation", + "content": "\\mathbf{d} = (\\theta ,\\phi)" + }, + { + "bbox": [ + 131, + 354, + 479, + 414 + ], + "type": "text", + "content": ", which matches the isotropic nature of X-ray imaging. Meanwhile, Eq. (3) is free from the complex computation of SH function. Hence, the forward and backward processes of our X-Gaussian are much faster than those of the original 3DGS." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 430, + 351, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 351, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 351, + 441 + ], + "type": "text", + "content": "3.2 Differentiable Radiative Rasterization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 449, + 479, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 449, + 479, + 485 + ], + "spans": [ + { + "bbox": [ + 131, + 449, + 479, + 485 + ], + "type": "text", + "content": "Based on our radiative Gaussian point cloud, we develop a Differentiable Radiative Rasterization (DRR), as shown in Fig. 2 (c). We first summarize the overall DRR processing " + }, + { + "bbox": [ + 131, + 449, + 479, + 485 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{DRR}}" + }, + { + "bbox": [ + 131, + 449, + 479, + 485 + ], + "type": "text", + "content": " and then describe its details. DRR is represented as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 176, + 495, + 480, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 495, + 480, + 507 + ], + "spans": [ + { + "bbox": [ + 176, + 495, + 480, + 507 + ], + "type": "interline_equation", + "content": "\\mathbf {I} = F _ {\\mathrm {D R R}} \\left(\\mathbf {M} _ {e x t}, \\mathbf {M} _ {i n t}, \\left\\{G _ {i} \\left(\\boldsymbol {\\mu} _ {i}, \\boldsymbol {\\Sigma} _ {i}, \\alpha_ {i}, \\mathbf {f} _ {i}\\right) \\mid i = 1, 2, \\dots , N _ {p} \\right\\}\\right), \\tag {5}", + "image_path": "80d214403097bc06c2ab28833a781e236bae1b3566efff6168a954eca3fcdba3.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 514, + 479, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 514, + 479, + 563 + ], + "spans": [ + { + "bbox": [ + 131, + 514, + 479, + 563 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 514, + 479, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 131, + 514, + 479, + 563 + ], + "type": "text", + "content": " denotes the rendered image, " + }, + { + "bbox": [ + 131, + 514, + 479, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{ext} \\in \\mathbb{R}^{4 \\times 4}" + }, + { + "bbox": [ + 131, + 514, + 479, + 563 + ], + "type": "text", + "content": " represents the extrinsic matrix, and " + }, + { + "bbox": [ + 131, + 514, + 479, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{int} \\in \\mathbb{R}^{4 \\times 3}" + }, + { + "bbox": [ + 131, + 514, + 479, + 563 + ], + "type": "text", + "content": " refers to the intrinsic matrix. Subsequently, we introduce the details of " + }, + { + "bbox": [ + 131, + 514, + 479, + 563 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{DRR}}" + }, + { + "bbox": [ + 131, + 514, + 479, + 563 + ], + "type": "text", + "content": ". To begin with, the possibility value of the " + }, + { + "bbox": [ + 131, + 514, + 479, + 563 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 514, + 479, + 563 + ], + "type": "text", + "content": "-th Gaussian distribution at the 3D point position " + }, + { + "bbox": [ + 131, + 514, + 479, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 131, + 514, + 479, + 563 + ], + "type": "text", + "content": " is formulated as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 203, + 571, + 480, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 571, + 480, + 592 + ], + "spans": [ + { + "bbox": [ + 203, + 571, + 480, + 592 + ], + "type": "interline_equation", + "content": "P (\\mathbf {x} \\mid \\boldsymbol {\\mu} _ {i}, \\boldsymbol {\\Sigma} _ {i}) = \\exp \\left(- \\frac {1}{2} (\\mathbf {x} - \\boldsymbol {\\mu} _ {i}) ^ {\\top} \\boldsymbol {\\Sigma} _ {i} ^ {- 1} (\\mathbf {x} - \\boldsymbol {\\mu} _ {i})\\right). \\tag {6}", + "image_path": "2af9f1ee99ecfaaa62bd3773f4df2b4e32d694c8e68df98f54c8a3da064f9dfd.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 598, + 479, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 598, + 479, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 598, + 479, + 635 + ], + "type": "text", + "content": "Then we project the 3D Gaussians to the 2D detector plane for subsequent rendering. " + }, + { + "bbox": [ + 131, + 598, + 479, + 635 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_i" + }, + { + "bbox": [ + 131, + 598, + 479, + 635 + ], + "type": "text", + "content": " is firstly transferred from the world coordinate system to the camera coordinate system and then projected to the image coordinate system as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 643, + 479, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 479, + 669 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 479, + 669 + ], + "type": "interline_equation", + "content": "\\widetilde {\\mathbf {t}} _ {i} = \\left[ \\begin{array}{l} \\mathbf {t} _ {i} \\\\ 1 \\end{array} \\right] = \\mathbf {M} _ {e x t} \\widetilde {\\boldsymbol {\\mu}} _ {i} = \\mathbf {M} _ {e x t} \\left[ \\begin{array}{l} \\boldsymbol {\\mu} _ {i} \\\\ 1 \\end{array} \\right], \\quad \\widetilde {\\mathbf {u}} _ {i} = \\left[ \\begin{array}{l} \\mathbf {u} _ {i} \\\\ 1 \\end{array} \\right] = \\mathbf {M} _ {i n t} \\widetilde {\\mathbf {t}} _ {i} = \\mathbf {M} _ {i n t} \\left[ \\begin{array}{l} \\mathbf {t} _ {i} \\\\ 1 \\end{array} \\right], \\tag {7}", + "image_path": "fd12079913b9336bf1e6b6456cf98e5c640ce850a2217dc4a0049ea1ccc4fde4.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 243, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 243, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 243, + 100 + ], + "type": "text", + "content": "Yuanhao Cai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 115, + 482, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 115, + 482, + 166 + ], + "spans": [ + { + "bbox": [ + 131, + 115, + 482, + 166 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 115, + 482, + 166 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i = (t_x, t_y, t_z) \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 131, + 115, + 482, + 166 + ], + "type": "text", + "content": " is the camera coordinate of " + }, + { + "bbox": [ + 131, + 115, + 482, + 166 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_i" + }, + { + "bbox": [ + 131, + 115, + 482, + 166 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 115, + 482, + 166 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_i \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 131, + 115, + 482, + 166 + ], + "type": "text", + "content": " is the image coordinate of " + }, + { + "bbox": [ + 131, + 115, + 482, + 166 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_i" + }, + { + "bbox": [ + 131, + 115, + 482, + 166 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 115, + 482, + 166 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{u}}_i" + }, + { + "bbox": [ + 131, + 115, + 482, + 166 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 115, + 482, + 166 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{t}}_i" + }, + { + "bbox": [ + 131, + 115, + 482, + 166 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 115, + 482, + 166 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pmb{\\mu}}_i" + }, + { + "bbox": [ + 131, + 115, + 482, + 166 + ], + "type": "text", + "content": " are the homogeneous coordinates of " + }, + { + "bbox": [ + 131, + 115, + 482, + 166 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_i" + }, + { + "bbox": [ + 131, + 115, + 482, + 166 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 115, + 482, + 166 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i" + }, + { + "bbox": [ + 131, + 115, + 482, + 166 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 115, + 482, + 166 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_i" + }, + { + "bbox": [ + 131, + 115, + 482, + 166 + ], + "type": "text", + "content": ", respectively. Subsequently, we transfer the 3D covariance matrix " + }, + { + "bbox": [ + 131, + 115, + 482, + 166 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_i" + }, + { + "bbox": [ + 131, + 115, + 482, + 166 + ], + "type": "text", + "content": " to its counterpart " + }, + { + "bbox": [ + 131, + 115, + 482, + 166 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_i' \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 131, + 115, + 482, + 166 + ], + "type": "text", + "content": " in the camera coordinate system as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 260, + 175, + 481, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 175, + 481, + 190 + ], + "spans": [ + { + "bbox": [ + 260, + 175, + 481, + 190 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} _ {i} ^ {\\prime} = \\mathbf {J} _ {i} \\mathbf {W} _ {i} \\boldsymbol {\\Sigma} _ {i} \\mathbf {W} _ {i} ^ {\\top} \\mathbf {J} _ {i} ^ {\\top}, \\tag {8}", + "image_path": "9c360f92c6b222cfab47976b4a750f2ece93c8adececd81018195ac9b7434b39.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 198, + 481, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 198, + 481, + 222 + ], + "spans": [ + { + "bbox": [ + 131, + 198, + 481, + 222 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 198, + 481, + 222 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_i\\in \\mathbb{R}^{3\\times 3}" + }, + { + "bbox": [ + 131, + 198, + 481, + 222 + ], + "type": "text", + "content": " is the Jacobian of the affine approximation of the projective transformation. " + }, + { + "bbox": [ + 131, + 198, + 481, + 222 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_i\\in \\mathbb{R}^{3\\times 3}" + }, + { + "bbox": [ + 131, + 198, + 481, + 222 + ], + "type": "text", + "content": " is the viewing transformation. We derive them by" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 174, + 232, + 481, + 275 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 232, + 481, + 275 + ], + "spans": [ + { + "bbox": [ + 174, + 232, + 481, + 275 + ], + "type": "interline_equation", + "content": "\\mathbf {J} _ {i} = \\left[ \\begin{array}{c c c} \\frac {L _ {S D}}{t _ {z}} & 0 & - \\frac {L _ {S D}}{t _ {z} ^ {2}} \\frac {t _ {x}}{t _ {z}} \\\\ 0 & \\frac {L _ {S D}}{t _ {z}} & - \\frac {L _ {S D}}{t _ {z} ^ {2}} \\frac {t _ {y}}{t _ {z}} \\\\ 0 & 0 & 0 \\end{array} \\right], \\quad \\mathbf {W} _ {i} = \\left[ \\begin{array}{c c c} - \\sin \\phi & \\cos \\phi & 0 \\\\ 0 & 0 & - 1 \\\\ - \\cos \\phi & - \\sin \\phi & 0 \\end{array} \\right], \\tag {9}", + "image_path": "654d42e8604d3fd646db69abe41c7d2b48f3f57168cdfd70f2ed325f29698678.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 282, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 282, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 131, + 282, + 482, + 380 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 282, + 482, + 380 + ], + "type": "inline_equation", + "content": "L_{SD}" + }, + { + "bbox": [ + 131, + 282, + 482, + 380 + ], + "type": "text", + "content": " represents the distance between the X-ray source and detector. " + }, + { + "bbox": [ + 131, + 282, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 131, + 282, + 482, + 380 + ], + "type": "text", + "content": " refers to the azimuth angle of the source. Following [25,29,62], we obtain the 2D covariance matrix " + }, + { + "bbox": [ + 131, + 282, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_i^{\\prime \\prime} \\in \\mathbb{R}^{2 \\times 2}" + }, + { + "bbox": [ + 131, + 282, + 482, + 380 + ], + "type": "text", + "content": " by skipping the third row and column of " + }, + { + "bbox": [ + 131, + 282, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_i^{\\prime}" + }, + { + "bbox": [ + 131, + 282, + 482, + 380 + ], + "type": "text", + "content": ". Then the 2D projection is partitioned into non-overlapping tiles. The 3D Gaussians " + }, + { + "bbox": [ + 131, + 282, + 482, + 380 + ], + "type": "inline_equation", + "content": "(\\pmb{\\mu}_i, \\pmb{\\Sigma}_i)" + }, + { + "bbox": [ + 131, + 282, + 482, + 380 + ], + "type": "text", + "content": " are assigned to different tiles according to their 2D projections " + }, + { + "bbox": [ + 131, + 282, + 482, + 380 + ], + "type": "inline_equation", + "content": "(\\pmb{u}_i, \\pmb{\\Sigma}_i^{\\prime \\prime})" + }, + { + "bbox": [ + 131, + 282, + 482, + 380 + ], + "type": "text", + "content": ", as shown in the left image of Fig. 2 (c). These 3D Gaussians are sorted by the distances to the 2D detector. Then the intensity " + }, + { + "bbox": [ + 131, + 282, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{I}(p) \\in \\mathbb{R}" + }, + { + "bbox": [ + 131, + 282, + 482, + 380 + ], + "type": "text", + "content": " at pixel " + }, + { + "bbox": [ + 131, + 282, + 482, + 380 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 131, + 282, + 482, + 380 + ], + "type": "text", + "content": " is obtained by blending " + }, + { + "bbox": [ + 131, + 282, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 131, + 282, + 482, + 380 + ], + "type": "text", + "content": " ordered points overlapping the pixel in the corresponding tile as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 190, + 389, + 481, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 389, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 190, + 389, + 481, + 422 + ], + "type": "interline_equation", + "content": "\\mathbf {I} (p) = \\sum_ {j \\in \\mathcal {N}} \\mathbf {i} _ {j} \\sigma_ {j} \\prod_ {k = 1} ^ {j - 1} (1 - \\sigma_ {k}), \\quad \\sigma_ {j} = \\alpha_ {j} P (\\mathbf {x} _ {j} | \\boldsymbol {\\mu} _ {j}, \\boldsymbol {\\Sigma} _ {j}), \\tag {10}", + "image_path": "562b83a995960bcc892798331b694e4c1961f948caab88f6d0db053510644b25.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 430, + 481, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 430, + 481, + 454 + ], + "spans": [ + { + "bbox": [ + 131, + 430, + 481, + 454 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 430, + 481, + 454 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j" + }, + { + "bbox": [ + 131, + 430, + 481, + 454 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 131, + 430, + 481, + 454 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 131, + 430, + 481, + 454 + ], + "type": "text", + "content": "-th intersection 3D point of the X-ray landing on pixel " + }, + { + "bbox": [ + 131, + 430, + 481, + 454 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 131, + 430, + 481, + 454 + ], + "type": "text", + "content": " and the Gaussian point clouds in 3D space. " + }, + { + "bbox": [ + 131, + 430, + 481, + 454 + ], + "type": "inline_equation", + "content": "\\mathbf{i}_j" + }, + { + "bbox": [ + 131, + 430, + 481, + 454 + ], + "type": "text", + "content": " is the radiation intensity of " + }, + { + "bbox": [ + 131, + 430, + 481, + 454 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j" + }, + { + "bbox": [ + 131, + 430, + 481, + 454 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 454, + 481, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 454, + 481, + 478 + ], + "spans": [ + { + "bbox": [ + 131, + 454, + 481, + 478 + ], + "type": "text", + "content": "Optimization. Eventually, the training objective " + }, + { + "bbox": [ + 131, + 454, + 481, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 131, + 454, + 481, + 478 + ], + "type": "text", + "content": " is the weighted sum of " + }, + { + "bbox": [ + 131, + 454, + 481, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 131, + 454, + 481, + 478 + ], + "type": "text", + "content": " loss and SSIM loss between the rendered and ground-truth projection images as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 252, + 491, + 481, + 503 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 491, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 252, + 491, + 481, + 503 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = (1 - \\gamma) \\mathcal {L} _ {1} + \\gamma \\mathcal {L} _ {\\text {S S I M}}, \\tag {11}", + "image_path": "5f964cafe47339111bff48a43b413a471f4d887fe9ffcc9e62e37f6d08a5286b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 508, + 481, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 508, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 131, + 508, + 481, + 555 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 508, + 481, + 555 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 131, + 508, + 481, + 555 + ], + "type": "text", + "content": " is a hyperparameter balancing the importances of the two loss terms. By minimizing Eq. (11), we can optimize the attributes of 3D Gaussians, i.e., " + }, + { + "bbox": [ + 131, + 508, + 481, + 555 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_i, \\pmb{\\Sigma}_i, \\alpha_i" + }, + { + "bbox": [ + 131, + 508, + 481, + 555 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 508, + 481, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 131, + 508, + 481, + 555 + ], + "type": "text", + "content": " in Eq. (5). " + }, + { + "bbox": [ + 131, + 508, + 481, + 555 + ], + "type": "inline_equation", + "content": "N_p" + }, + { + "bbox": [ + 131, + 508, + 481, + 555 + ], + "type": "text", + "content": " is adjusted by the adaptive control [25]. The optimization process is visualized in Fig. 1 and the video file in supplementary." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 556, + 481, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 556, + 481, + 592 + ], + "spans": [ + { + "bbox": [ + 131, + 556, + 481, + 592 + ], + "type": "text", + "content": "Compared to the RGB rasterization in 3DGS [25], our DRR avoids the complex computations related to the view direction in the forward and backward processes, thereby enjoying cheaper training costs and faster inference speed." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 609, + 371, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 371, + 621 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 371, + 621 + ], + "type": "text", + "content": "3.3 Angle-pose Cuboid Uniform Initialization" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "type": "text", + "content": "At the beginning of training, we need to initialize the parameters in Eq. (5) for rasterization. Specifically, " + }, + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{\\Sigma}_i" + }, + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\alpha_i" + }, + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "type": "text", + "content": " are randomly initialized. In natural light imaging, the original 3DGS [25] adopts the SfM [46] algorithm to compute the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 161, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 161, + 91, + 448, + 103 + ], + "type": "text", + "content": "Radiative Gaussian Splitting for Efficient X-ray Novel View Synthesis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "text", + "content": "initial " + }, + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_i" + }, + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "inline_equation", + "content": "N_p" + }, + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{ext}" + }, + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{int}" + }, + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "text", + "content": ". SfM detects and matches features from multiview images. It is not suitable for X-ray imaging due to two reasons. Firstly, X-ray images are grayscale and low-contrast. Secondly, different layers of an object may overlap on the same positions of the projection. These two problems degrade the accuracy of feature detection and matching in SfM. Besides, running the SfM algorithm usually requires a long time, which prolongs the training process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 189, + 480, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 189, + 480, + 249 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 480, + 249 + ], + "type": "text", + "content": "To address these issues, we customize an Angle-posed Cuboid Uniform Initialization (ACUI) strategy for circular cone beam X-ray scanning scenario where a scanner emits cone-shaped X-ray beams and captures projections at equal angular intervals. As shown in Fig. 2 (a), ACUI uses the parameters of X-ray scanner to compute the extrinsic matrix " + }, + { + "bbox": [ + 130, + 189, + 480, + 249 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{ext}" + }, + { + "bbox": [ + 130, + 189, + 480, + 249 + ], + "type": "text", + "content": " and intrinsic matrix " + }, + { + "bbox": [ + 130, + 189, + 480, + 249 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{int}" + }, + { + "bbox": [ + 130, + 189, + 480, + 249 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 260, + 480, + 310 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 260, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 146, + 260, + 480, + 310 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {e x t} = \\left[ \\begin{array}{c c c c} - \\sin \\phi & \\cos \\phi & 0 & 0 \\\\ 0 & 0 & - 1 & 0 \\\\ - \\cos \\phi & - \\sin \\phi & 0 & L _ {S O} \\\\ 0 & 0 & 0 & 1 \\end{array} \\right], \\quad \\mathbf {M} _ {i n t} = \\left[ \\begin{array}{c c c c} L _ {S D} & 0 & W / 2 & 0 \\\\ 0 & L _ {S D} & H / 2 & 0 \\\\ 0 & 0 & 1 & 0 \\end{array} \\right], \\tag {12}", + "image_path": "b5f836418a72a4b1dbf9d182d17ecc199e4f285e6488122fe0842d2efe0804a2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 321, + 480, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 480, + 429 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 321, + 480, + 429 + ], + "type": "inline_equation", + "content": "L_{SO}" + }, + { + "bbox": [ + 130, + 321, + 480, + 429 + ], + "type": "text", + "content": " represents the distance between the X-ray source and the scanned object. The elevation angle of the X-ray source is set to zero and remains unchanged. The next step of ACUI is to initialize the center positions of 3D Gaussians. Although the precise shape of the scanned object is not given at the beginning, the scanning space can be approximated. We set up a cuboid with size " + }, + { + "bbox": [ + 130, + 321, + 480, + 429 + ], + "type": "inline_equation", + "content": "S_{1} \\times S_{2} \\times S_{3}" + }, + { + "bbox": [ + 130, + 321, + 480, + 429 + ], + "type": "text", + "content": " (mm) that can completely enclose the object. The center of this cuboid is also the center of the object and the origin of the world coordinate system. We divide this cuboid by a grid with size " + }, + { + "bbox": [ + 130, + 321, + 480, + 429 + ], + "type": "inline_equation", + "content": "M_{1} \\times M_{2} \\times M_{3}" + }, + { + "bbox": [ + 130, + 321, + 480, + 429 + ], + "type": "text", + "content": " (voxel). Then we uniformly sample points within the grid at interval " + }, + { + "bbox": [ + 130, + 321, + 480, + 429 + ], + "type": "inline_equation", + "content": "d \\in \\mathbb{R}" + }, + { + "bbox": [ + 130, + 321, + 480, + 429 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 439, + 480, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 439, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 439, + 480, + 464 + ], + "type": "interline_equation", + "content": "\\mathcal {P} = \\left\\{\\left(\\frac {n _ {1} S _ {1} d}{M _ {1}}, \\frac {n _ {2} S _ {2} d}{M _ {2}}, \\frac {n _ {3} S _ {3} d}{M _ {3}}\\right) \\mid - \\left[ \\frac {M _ {i}}{2 d} \\right] - 1 \\leq n _ {i} \\leq \\left[ \\frac {M _ {i}}{2 d} \\right] + 1, i = 1, 2, 3 \\right\\}, \\tag {13}", + "image_path": "eafbc70a0b7ec06d8922c909de3e39d9dd5fda610ba31cb3484c0f3075f75c20.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 474, + 479, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 474, + 479, + 499 + ], + "spans": [ + { + "bbox": [ + 131, + 474, + 479, + 499 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 474, + 479, + 499 + ], + "type": "inline_equation", + "content": "n_i \\in \\mathbb{Z}" + }, + { + "bbox": [ + 131, + 474, + 479, + 499 + ], + "type": "text", + "content": ". Then we use the size and elements of " + }, + { + "bbox": [ + 131, + 474, + 479, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 131, + 474, + 479, + 499 + ], + "type": "text", + "content": " to initialize " + }, + { + "bbox": [ + 131, + 474, + 479, + 499 + ], + "type": "inline_equation", + "content": "N_p" + }, + { + "bbox": [ + 131, + 474, + 479, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 474, + 479, + 499 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_i" + }, + { + "bbox": [ + 131, + 474, + 479, + 499 + ], + "type": "text", + "content": ". Avoiding running SfM, ACUI allows X-Gaussian to enjoy a faster training speed." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 523, + 230, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 523, + 230, + 537 + ], + "spans": [ + { + "bbox": [ + 131, + 523, + 230, + 537 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 555, + 272, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 555, + 272, + 567 + ], + "spans": [ + { + "bbox": [ + 131, + 555, + 272, + 567 + ], + "type": "text", + "content": "4.1 Experimental Settings" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "content": "Dataset. Following NAF [60], we adopt the public datasets of human organ CTs, i.e., LIDC-IDRI [2] and the open scientific visualization dataset [27], to evaluate our method. The test scenes include chest, foot, head, abdomen, and pancreas. We adopt the open-source tomographic toolbox TIGRE [6] to capture 100 projections with " + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "content": " noise in the range of " + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "inline_equation", + "content": "0 \\sim 180^{\\circ}" + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "content": ". In the NVS task, 50 projections are used for training and the other 50 projections are used for testing. The CT volumes are used for testing in the sparse-view CT reconstruction task." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 244, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 244, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 244, + 101 + ], + "type": "text", + "content": "Yuanhao Cai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 157, + 480, + 338 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 1: Quantitative results on the novel view synthesis task. The average inference speed and training time of all scenes evaluated on an RTX 8000 GPU are reported. In the cell of the results of each scene, PSNR (upper) and SSIM (lower) are listed." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 157, + 480, + 338 + ], + "lines": [ + { + "bbox": [ + 134, + 157, + 480, + 338 + ], + "spans": [ + { + "bbox": [ + 134, + 157, + 480, + 338 + ], + "type": "table", + "html": "
MethodInfer SpeedTrain TimeChestFootHeadAbdomenPancreasAverage
InTomo [58]0.62 fps125 min28.94839.48234.83227.64120.03130.187
0.99150.99790.99770.96460.85370.9611
NeRF [37]0.14 fps313 min36.15741.05329.76024.62019.85330.289
0.99880.99890.99910.95590.85600.9617
TensoRF [9]0.77 fps178 min23.60937.72834.42927.38229.23530.477
0.94020.99290.98790.87300.80310.9194
NeAT [42]1.78 fps69 min40.76538.23627.73826.74137.52634.201
0.99900.99630.92950.85630.90170.9366
NAF [60]2.01 fps63 min42.36638.35330.17437.59036.22836.942
0.99930.99130.95310.98550.88440.9627
X-Gaussian148 fps9 min43.88742.15341.57945.76243.64043.404
0.99980.99970.99970.99990.99760.9993
", + "image_path": "557d50bc5802d32c245254be3795a0e91f46a231319c2291e4cf3ccfc1f21f4c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 361, + 482, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 470 + ], + "type": "text", + "content": "Implementation Details. Our X-Gaussian is implemented by PyTorch [40] and CUDA [17]. The model is trained with the Adam optimizer [26] (" + }, + { + "bbox": [ + 130, + 361, + 482, + 470 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 130, + 361, + 482, + 470 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 361, + 482, + 470 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 130, + 361, + 482, + 470 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 361, + 482, + 470 + ], + "type": "inline_equation", + "content": "\\epsilon = 1\\times 10^{-15}" + }, + { + "bbox": [ + 130, + 361, + 482, + 470 + ], + "type": "text", + "content": ") for " + }, + { + "bbox": [ + 130, + 361, + 482, + 470 + ], + "type": "inline_equation", + "content": "2\\times 10^{4}" + }, + { + "bbox": [ + 130, + 361, + 482, + 470 + ], + "type": "text", + "content": " iterations. The learning rate for point cloud position is initially set to " + }, + { + "bbox": [ + 130, + 361, + 482, + 470 + ], + "type": "inline_equation", + "content": "1.9\\times 10^{-4}" + }, + { + "bbox": [ + 130, + 361, + 482, + 470 + ], + "type": "text", + "content": " and exponentially decays to " + }, + { + "bbox": [ + 130, + 361, + 482, + 470 + ], + "type": "inline_equation", + "content": "1.9\\times 10^{-6}" + }, + { + "bbox": [ + 130, + 361, + 482, + 470 + ], + "type": "text", + "content": ". The learning rates for point feature, opacity, scaling, and rotation are set to " + }, + { + "bbox": [ + 130, + 361, + 482, + 470 + ], + "type": "inline_equation", + "content": "2\\times 10^{-3}" + }, + { + "bbox": [ + 130, + 361, + 482, + 470 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 361, + 482, + 470 + ], + "type": "inline_equation", + "content": "8\\times 10^{-3}" + }, + { + "bbox": [ + 130, + 361, + 482, + 470 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 361, + 482, + 470 + ], + "type": "inline_equation", + "content": "5\\times 10^{-3}" + }, + { + "bbox": [ + 130, + 361, + 482, + 470 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 361, + 482, + 470 + ], + "type": "inline_equation", + "content": "1\\times 10^{-3}" + }, + { + "bbox": [ + 130, + 361, + 482, + 470 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 361, + 482, + 470 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 361, + 482, + 470 + ], + "type": "text", + "content": " in Eq. (11) is set to 0.2. We adopt peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) [49] to evaluate the performance. Frames per second (fps) is used to measure the inference speed. Experiments are conducted on an RTX 8000 GPU." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 488, + 270, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 488, + 270, + 502 + ], + "spans": [ + { + "bbox": [ + 131, + 488, + 270, + 502 + ], + "type": "text", + "content": "4.2 Novel View Synthesis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 509, + 481, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 481, + 545 + ], + "type": "text", + "content": "Quantitative Results. Tab. 1 shows the quantitative comparisons between our X-Gaussian and five SOTA NeRF-based algorithms, including InTomo [58], NeRF [37], TensoRF [9], NeAT [42], and NAF [60] on the NVs task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "We report the average inference speed and training time of different methods on all scenes. In the cell of the results of each scene, PSNR (upper entry in the cell) and SSIM (lower entry in the cell) are listed. As can be observed that our X-Gaussian not only surpasses SOTA methods by large margins in performance but also enjoys much faster inference speed and cheaper training costs. More specifically, compared with the recent best X-ray NeRF-based method NAF, our X-Gaussian outperforms it by 6.5 dB on average and is " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "73 \\times" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " faster in inference while requiring less than " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " training time. When compared with the SOTA RGB NeRF-based method TensoRF, our X-Gaussian is 12.93 dB higher while enjoying " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "192 \\times" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " inference speed and " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "20 \\times" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " training speed." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 161, + 90, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 90, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 161, + 90, + 447, + 103 + ], + "type": "text", + "content": "Radiative Gaussian Splitting for Efficient X-ray Novel View Synthesis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 115, + 481, + 284 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 481, + 284 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 481, + 284 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 481, + 284 + ], + "type": "image", + "image_path": "704a3c2ce3cce0b91c4a450be244bf2b5a2d6b6b689c15768c55d3dbcb1202ca.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 304, + 482, + 327 + ], + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 327 + ], + "type": "text", + "content": "Fig. 4: Qualitative results of novel view synthesis on the scenes of pancreas (top) and chest (bottom). Our X-Gaussian yields clearer results. Please zoom in for a better view." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 350, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 422 + ], + "type": "text", + "content": "Qualitative Results. Figs. 4 depict the qualitative comparisons of NVS on the scenes of pancreas and chest. It can be observed from the zoomed-in patches that previous NeRF-based algorithms fail to render clear novel views. They either introduce undesired artifacts or produce blurry textures such as the toe bones of the foot. In contrast, our method yields visually realistic images by rendering more fine-grained details and clearer structural contents." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 441, + 324, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 324, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 324, + 453 + ], + "type": "text", + "content": "4.3 Sparse-View CT Reconstruction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 462, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 628 + ], + "type": "text", + "content": "We compare our method with SOTA NeRF-based algorithms on sparse-view CT reconstruction. Since the Gaussian point clouds cannot directly infer the radiodensities of the CT volume, we evaluate different NeRF-based methods and our X-Gaussian by using them to create novel-view projections for three learning-free CT reconstruction methods, including an analytical method (FDK [16]) and two iterative methods (SART [1] and ASD-POCS [45]). Specifically, these three methods reconstruct the CTs from 5 original projections and 95 novel-view projections rendered by different NVS algorithms. The quantitative results are listed in Tab. 2. When only using 5 original views (+ None), FDK, SART, and ASD-POCS achieve 7.41, 17.24, and " + }, + { + "bbox": [ + 130, + 462, + 482, + 628 + ], + "type": "inline_equation", + "content": "17.03\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 462, + 482, + 628 + ], + "type": "text", + "content": " in PSNR, respectively. They fail to reconstruct the CT volumes. When employing our X-Gaussian to create novel X-ray projections for FDK, SART, and ASD-POCS, they yield the most significant improvements of 15.19, 13.01, and " + }, + { + "bbox": [ + 130, + 462, + 482, + 628 + ], + "type": "inline_equation", + "content": "13.53\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 462, + 482, + 628 + ], + "type": "text", + "content": " in PSNR. These improvements are 1.32, 2.41, and " + }, + { + "bbox": [ + 130, + 462, + 482, + 628 + ], + "type": "inline_equation", + "content": "2.65\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 462, + 482, + 628 + ], + "type": "text", + "content": " higher than the improvements of using NAF." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": "Figs. 5 shows the qualitative results of sparse-view CT reconstruction on the scenes of foot and chest. Without using rendered novel-view projections, SART fails in reconstructing the CT slices. When using SOTA X-ray NeRF-" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 243, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 243, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 243, + 100 + ], + "type": "text", + "content": "Yuanhao Cai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 129, + 147, + 482, + 254 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 2: Results on sparse-view CT reconstruction. NeRF-based methods and our X-Gaussian are used to create novel views for FDK [16], SART [1], and ASD-POCS [45]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 129, + 147, + 482, + 254 + ], + "lines": [ + { + "bbox": [ + 129, + 147, + 482, + 254 + ], + "spans": [ + { + "bbox": [ + 129, + 147, + 482, + 254 + ], + "type": "table", + "html": "
Method Metric+ None+ InTomo [58]+ NeRF [37]+ TensorRF [9]+ NeAT [42]+ NAF [60]+ X-Gaussian
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
FDK7.410.09320.310.49820.570.50220.610.50120.940.51121.280.52322.600.584
Δ FDK0012.900.40513.160.40913.200.40813.520.41813.870.43015.190.491
SART17.240.52826.280.85926.780.85327.060.86727.310.86927.840.87930.250.907
Δ SART009.040.3319.540.3259.820.33910.070.34110.600.35113.010.379
ASD-POCS17.030.52525.440.84726.580.85726.930.86826.950.86527.910.88030.560.926
Δ ASD-POCS008.410.3229.550.3329.900.3439.920.34010.880.35513.530.401
", + "image_path": "706e072ff6a9b0ec1c7bd9347ed5b5029314d4cc826ae064dbd819de5298a273.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 276, + 482, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 276, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 276, + 482, + 350 + ], + "type": "text", + "content": "based methods to create novel views, SART produces over-smooth CT slices with blurry structural contents. On the contrary, when using our X-Gaussian to assist SART and ASD-POCS, they can reconstruct much clearer CT slices with more high-frequency textures and fine-grained structural details, such as the vessels in the chest (Fig. 5). These results clearly demonstrate the potential practical values of our method on the sparse-view CT reconstruction task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 368, + 238, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 368, + 238, + 380 + ], + "spans": [ + { + "bbox": [ + 131, + 368, + 238, + 380 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 390, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 666 + ], + "type": "text", + "content": "Break-down Ablation. We first conduct a break-down ablation experiment to study the effect of each proposed technique towards higher performance and faster speed. We adopt the original 3DGS [25] as the baseline model and naively average the RGB channels to represent the value of radiation intensity. The results are listed in Tab. 3a. The baseline model yields 37.21 dB PSNR in performance. Its average training time and inference speed are 31 min 38 s and 64 fps, respectively. We can observe from Tab. 3a: (i) When using ACUI to replace the time-consuming SfM [46] algorithm for initialization, the training time is significantly reduced by " + }, + { + "bbox": [ + 130, + 390, + 482, + 666 + ], + "type": "inline_equation", + "content": "34\\%" + }, + { + "bbox": [ + 130, + 390, + 482, + 666 + ], + "type": "text", + "content": " while the performance yields an improvement of 1.66 dB. This evidence suggests that our ACUI strategy is much faster than the SfM [46] algorithm used in the original 3DGS and can compute more accurate camera calibration parameters for 3D Gaussians and subsequent rendering. (ii) Then we apply our radiative Gaussian point cloud model equipped with the proposed Differentiable Radiative Rasterization (DRR) to replace the original RGB Gaussian point cloud model and its RGB rasterization. As analyzed in Sec. 3.1 and compared in Fig. 3, the anisotropic spherical harmonics (SH) are not suitable for X-ray imaging because X-ray imaging based on penetration is isotropic. In contrast, our radiative Gaussian point cloud model can better fit the view-independent radiation intensity in 3D space. Therefore, the performance is significantly improved by 4.53 dB in PSNR. Besides, removing the computation related to the view direction from the forward and backward processes of rasterization can further accelerate the training and inference speed. Thus, the training time is reduced by " + }, + { + "bbox": [ + 130, + 390, + 482, + 666 + ], + "type": "inline_equation", + "content": "54.10\\%" + }, + { + "bbox": [ + 130, + 390, + 482, + 666 + ], + "type": "text", + "content": " and the inference speed is " + }, + { + "bbox": [ + 130, + 390, + 482, + 666 + ], + "type": "inline_equation", + "content": "2.1\\times" + }, + { + "bbox": [ + 130, + 390, + 482, + 666 + ], + "type": "text", + "content": " faster." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 161, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 161, + 91, + 447, + 102 + ], + "type": "text", + "content": "Radiative Gaussian Splitting for Efficient X-ray Novel View Synthesis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 116, + 477, + 264 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 477, + 264 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 477, + 264 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 477, + 264 + ], + "type": "image", + "image_path": "c93447777a2c2d8286aa2487c68acfc3a3cf1ca13ce3f955a41c7fa0aff3296f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 284, + 480, + 308 + ], + "lines": [ + { + "bbox": [ + 132, + 284, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 284, + 480, + 308 + ], + "type": "text", + "content": "Fig. 5: Visual results of sparse-view CT reconstruction on the scenes of foot and chest. NeRF-based methods and our X-Gaussian are used to create novel views for SART [1]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 331, + 482, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 523 + ], + "type": "text", + "content": "Initialization of Point Position. We compare different initialization strategies for the center positions of Gaussian point clouds including random, spherical, FDK [16], and cubic initialization. To be specific, Random initialization means randomly sampling points within the scanned area in 3D space. Spherical initialization uniformly samples point positions within a sphere that can completely enclose the scanned object. FDK [16] initialization adopts the FDK algorithm to back-project the given projections into 3D point positions. Cuboid initialization is our ACUI. Please note that we keep the computed intrinsic and extrinsic matrices the same for fair comparison between different strategies. The results are reported in Tab. 3b. FDK initialization slightly outperforms our ACUI by 0.066 dB but its training time is " + }, + { + "bbox": [ + 130, + 331, + 482, + 523 + ], + "type": "inline_equation", + "content": "2.59 \\times" + }, + { + "bbox": [ + 130, + 331, + 482, + 523 + ], + "type": "text", + "content": " longer and its inference speed is 55 fps slower than ACUI. This is because the back-projection in FDK is time-consuming and initializes redundant points. The random and spherical initialization strategies yield lower PSNR and slower speed than cubic initialization. To achieve a better trade-off, we adopt the cubic initialization, i.e., ACUI, which enjoys good performance, the cheapest training cost, and the fastest inference speed." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 544, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 482, + 668 + ], + "type": "text", + "content": "Parameter Analysis. We conduct parameter analysis of the number of features " + }, + { + "bbox": [ + 132, + 544, + 482, + 668 + ], + "type": "inline_equation", + "content": "N_{f}" + }, + { + "bbox": [ + 132, + 544, + 482, + 668 + ], + "type": "text", + "content": " and the sampling interval " + }, + { + "bbox": [ + 132, + 544, + 482, + 668 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 132, + 544, + 482, + 668 + ], + "type": "text", + "content": ". The results are shown in Tab. 3c and Tab. 3d. In Tab. 3c, (i) When increasing " + }, + { + "bbox": [ + 132, + 544, + 482, + 668 + ], + "type": "inline_equation", + "content": "N_{f}" + }, + { + "bbox": [ + 132, + 544, + 482, + 668 + ], + "type": "text", + "content": ", the performance gradually improves but the magnitude of the improvement decreases. In particular, " + }, + { + "bbox": [ + 132, + 544, + 482, + 668 + ], + "type": "inline_equation", + "content": "N_{f} = 32" + }, + { + "bbox": [ + 132, + 544, + 482, + 668 + ], + "type": "text", + "content": " achieves the best results of 43.42 dB in PSNR. " + }, + { + "bbox": [ + 132, + 544, + 482, + 668 + ], + "type": "inline_equation", + "content": "N_{f} = 16" + }, + { + "bbox": [ + 132, + 544, + 482, + 668 + ], + "type": "text", + "content": " achieves on-par results with " + }, + { + "bbox": [ + 132, + 544, + 482, + 668 + ], + "type": "inline_equation", + "content": "N_{f} = 32" + }, + { + "bbox": [ + 132, + 544, + 482, + 668 + ], + "type": "text", + "content": ", only 0.013 dB lower. (ii) We notice that the training time and inference speed do not change monotonically. This is because the Gaussian point clouds with various feature dimensions have different representing ability and computational complexity. The number of final 3D Gaussians after training also varies substantially. When " + }, + { + "bbox": [ + 132, + 544, + 482, + 668 + ], + "type": "inline_equation", + "content": "N_{f} = 16" + }, + { + "bbox": [ + 132, + 544, + 482, + 668 + ], + "type": "text", + "content": ", the training time reaches a local minimum" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 243, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 243, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 243, + 100 + ], + "type": "text", + "content": "Yuanhao Cai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 135, + 290, + 178 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 126 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 126 + ], + "type": "text", + "content": "Table 3: Ablation study. PSNR, SSIM, training time, and inference speed are reported." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 135, + 290, + 178 + ], + "lines": [ + { + "bbox": [ + 137, + 135, + 290, + 178 + ], + "spans": [ + { + "bbox": [ + 137, + 135, + 290, + 178 + ], + "type": "table", + "html": "
Method3DGS [25]+ ACUI+ DRR
PSNR37.21338.87243.404
SSIM0.98130.98710.9993
Train time (s)18981172538
Infer speed (fps)6472148
", + "image_path": "eec9fa853229df89eff6de68d849d125417ccdb8c06887d192d53a8920e61bd4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 293, + 136, + 476, + 178 + ], + "blocks": [ + { + "bbox": [ + 155, + 180, + 265, + 190 + ], + "lines": [ + { + "bbox": [ + 155, + 180, + 265, + 190 + ], + "spans": [ + { + "bbox": [ + 155, + 180, + 265, + 190 + ], + "type": "text", + "content": "(a) Break-down ablation study" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 293, + 136, + 476, + 178 + ], + "lines": [ + { + "bbox": [ + 293, + 136, + 476, + 178 + ], + "spans": [ + { + "bbox": [ + 293, + 136, + 476, + 178 + ], + "type": "table", + "html": "
MethodRandomSphericalFDK [16]Cubic
PSNR41.32942.83743.47043.404
SSIM0.99420.99880.99930.9993
Train time (s)6015751394538
Infer speed (fps)11213693148
", + "image_path": "d02fcb290b344fdd5205d9e64e8862a0d954d5c48c51291a22c9f2bbe680b753.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 135, + 199, + 304, + 243 + ], + "blocks": [ + { + "bbox": [ + 304, + 180, + 457, + 190 + ], + "lines": [ + { + "bbox": [ + 304, + 180, + 457, + 190 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 457, + 190 + ], + "type": "text", + "content": "(b) Ablation of point position initialization" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 199, + 304, + 243 + ], + "lines": [ + { + "bbox": [ + 135, + 199, + 304, + 243 + ], + "spans": [ + { + "bbox": [ + 135, + 199, + 304, + 243 + ], + "type": "table", + "html": "
Number12481632
PSNR38.81840.20542.13042.86843.40443.417
SSIM0.98400.99310.99830.99910.99930.9993
Train time (s)511525521553538752
Speed (fps)153152158127148101
", + "image_path": "bc82404cb21d2cde92219c7f0456db974f0c33102e1b9cbbbf58f65baf473fa6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 307, + 199, + 476, + 243 + ], + "blocks": [ + { + "bbox": [ + 140, + 244, + 290, + 254 + ], + "lines": [ + { + "bbox": [ + 140, + 244, + 290, + 254 + ], + "spans": [ + { + "bbox": [ + 140, + 244, + 290, + 254 + ], + "type": "text", + "content": "(c) Analysis of the number of features " + }, + { + "bbox": [ + 140, + 244, + 290, + 254 + ], + "type": "inline_equation", + "content": "N_{f}" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 199, + 476, + 243 + ], + "lines": [ + { + "bbox": [ + 307, + 199, + 476, + 243 + ], + "spans": [ + { + "bbox": [ + 307, + 199, + 476, + 243 + ], + "type": "table", + "html": "
Interval12481632
PSNR42.85342.97943.21543.40443.31143.294
SSIM0.99890.99900.99920.99930.99920.9992
Train time (s)785593545538534566
Speed (fps)869411414813597
", + "image_path": "3bed1f177afb93c94a3c47dccd6eab0f21d3265717d1806bb68e9f4bfdaa5da4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 244, + 459, + 254 + ], + "lines": [ + { + "bbox": [ + 317, + 244, + 459, + 254 + ], + "spans": [ + { + "bbox": [ + 317, + 244, + 459, + 254 + ], + "type": "text", + "content": "(d) Analysis of the initialized interval " + }, + { + "bbox": [ + 317, + 244, + 459, + 254 + ], + "type": "inline_equation", + "content": "d" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 290, + 480, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 290, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 290, + 480, + 314 + ], + "type": "text", + "content": "and the inference speed is at its local maximum. Hence, we eventually adopt " + }, + { + "bbox": [ + 132, + 290, + 480, + 314 + ], + "type": "inline_equation", + "content": "N_{f} = 16" + }, + { + "bbox": [ + 132, + 290, + 480, + 314 + ], + "type": "text", + "content": " to reach a more optimal balance between performance and speed." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 316, + 480, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 480, + 351 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 480, + 351 + ], + "type": "text", + "content": "In Tab. 3d, the best reconstruction performance and the fastest inference speed are achieved at " + }, + { + "bbox": [ + 130, + 316, + 480, + 351 + ], + "type": "inline_equation", + "content": "d = 8" + }, + { + "bbox": [ + 130, + 316, + 480, + 351 + ], + "type": "text", + "content": ". The training time (538 s) at " + }, + { + "bbox": [ + 130, + 316, + 480, + 351 + ], + "type": "inline_equation", + "content": "d = 8" + }, + { + "bbox": [ + 130, + 316, + 480, + 351 + ], + "type": "text", + "content": " is almost the same as the shortest one (534 s) at " + }, + { + "bbox": [ + 130, + 316, + 480, + 351 + ], + "type": "inline_equation", + "content": "d = 16" + }, + { + "bbox": [ + 130, + 316, + 480, + 351 + ], + "type": "text", + "content": ". Thus, we eventually set " + }, + { + "bbox": [ + 130, + 316, + 480, + 351 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 316, + 480, + 351 + ], + "type": "text", + "content": " to 8." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 376, + 480, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 480, + 411 + ], + "type": "text", + "content": "Convergence Analysis. We conduct two visual analyses to compare the convergence between our X-Gaussian and original 3DGS [25] in Fig. 1, and between X-Gaussian and the SOTA X-ray NeRF-based method NAF [60] in Fig. 6." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 413, + 481, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 481, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 481, + 592 + ], + "type": "text", + "content": "Specifically, we adopt the same ACUI strategy for the original 3DGS to focus on comparing the Gaussian point cloud model and the rasterization. For fairness, we train 3DGS and our X-Gaussian on the scene of foot with the same settings and visualize the positions and opacities of Gaussian point clouds at the 100-th, 1000-th, 5000-th, 10000-th, and 20000-th iterations of the training process in Fig. 1. We also visualize the CT volume of foot as a reference. As can be seen that 3DGS with RGB rasterization converges slowly and suffers from more noisy point clouds. Plus, the final trained model of 3DGS at the 20000-th iteration contains more redundant Gaussians that are unnecessary to represent the 3D structure of the foot, which reduces the model's inference speed. In contrast, our X-Gaussian equipped with the proposed DRR shows faster and better convergence. In particular, as early as the 1000-th iteration, our radiative Gaussian point clouds have essentially formed the basic shape of the foot. Besides, the final trained X-Gaussian at the 20000-th iteration can better represent the 3D geometry and more accurate structural contents than the original 3DGS." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "Additionally, in Fig. 6, we visualize the rendered novel projections with the same angle " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " of NAF and our X-Gaussian at 20s, 60s, and 180s of the training process on the scene of chest. NAF produces blurry images with severe noises on the background regions within the first three minutes of the training. In contrast, our X-Gaussian can reconstruct clearer structural details like the ribs and blood vessels with cleaner background of the chest at the first minute of the training." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 161, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 161, + 91, + 448, + 102 + ], + "type": "text", + "content": "Radiative Gaussian Splitting for Efficient X-ray Novel View Synthesis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 122, + 181, + 170 + ], + "blocks": [ + { + "bbox": [ + 133, + 122, + 181, + 170 + ], + "lines": [ + { + "bbox": [ + 133, + 122, + 181, + 170 + ], + "spans": [ + { + "bbox": [ + 133, + 122, + 181, + 170 + ], + "type": "image", + "image_path": "0736b903be6058f5f5a2d54ceffdc55e2f223d8a0b811795ae658ac8ba443dea.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 171, + 169, + 175 + ], + "lines": [ + { + "bbox": [ + 145, + 171, + 169, + 175 + ], + "spans": [ + { + "bbox": [ + 145, + 171, + 169, + 175 + ], + "type": "text", + "content": "NAF at 20s" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 181, + 123, + 228, + 170 + ], + "blocks": [ + { + "bbox": [ + 181, + 123, + 228, + 170 + ], + "lines": [ + { + "bbox": [ + 181, + 123, + 228, + 170 + ], + "spans": [ + { + "bbox": [ + 181, + 123, + 228, + 170 + ], + "type": "image", + "image_path": "80050c19b38b34b78e5727fbc48268c4df08f044a17e65cf254c034226472562.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 193, + 171, + 216, + 175 + ], + "lines": [ + { + "bbox": [ + 193, + 171, + 216, + 175 + ], + "spans": [ + { + "bbox": [ + 193, + 171, + 216, + 175 + ], + "type": "text", + "content": "NAF at 60s" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 229, + 123, + 276, + 170 + ], + "blocks": [ + { + "bbox": [ + 229, + 123, + 276, + 170 + ], + "lines": [ + { + "bbox": [ + 229, + 123, + 276, + 170 + ], + "spans": [ + { + "bbox": [ + 229, + 123, + 276, + 170 + ], + "type": "image", + "image_path": "2fc4386cb4049a3985d704dfc2a99f5a40a441bcd58eb20415092ea092d1ca60.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 171, + 264, + 175 + ], + "lines": [ + { + "bbox": [ + 239, + 171, + 264, + 175 + ], + "spans": [ + { + "bbox": [ + 239, + 171, + 264, + 175 + ], + "type": "text", + "content": "NAF at 180s" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 276, + 123, + 323, + 170 + ], + "blocks": [ + { + "bbox": [ + 276, + 123, + 323, + 170 + ], + "lines": [ + { + "bbox": [ + 276, + 123, + 323, + 170 + ], + "spans": [ + { + "bbox": [ + 276, + 123, + 323, + 170 + ], + "type": "image", + "image_path": "cb5fbba27a6b1712ba952f139e001024fe0a215cef1d0f59b41bf07f06b227d1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 285, + 171, + 313, + 175 + ], + "lines": [ + { + "bbox": [ + 285, + 171, + 313, + 175 + ], + "spans": [ + { + "bbox": [ + 285, + 171, + 313, + 175 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 133, + 177, + 181, + 224 + ], + "blocks": [ + { + "bbox": [ + 133, + 177, + 181, + 224 + ], + "lines": [ + { + "bbox": [ + 133, + 177, + 181, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 177, + 181, + 224 + ], + "type": "image", + "image_path": "9bd14f35032f1bf570daab932c842dec9da505c590669d8829e8fd99e3e909b0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 224, + 176, + 231 + ], + "lines": [ + { + "bbox": [ + 138, + 224, + 176, + 231 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 176, + 231 + ], + "type": "text", + "content": "X-Gaussian at 20s" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 181, + 177, + 228, + 224 + ], + "blocks": [ + { + "bbox": [ + 181, + 177, + 228, + 224 + ], + "lines": [ + { + "bbox": [ + 181, + 177, + 228, + 224 + ], + "spans": [ + { + "bbox": [ + 181, + 177, + 228, + 224 + ], + "type": "image", + "image_path": "432a86bc726fa19eebd1e9632e74775fe693a390ef901c499e8a60eeafa95415.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 185, + 224, + 223, + 231 + ], + "lines": [ + { + "bbox": [ + 185, + 224, + 223, + 231 + ], + "spans": [ + { + "bbox": [ + 185, + 224, + 223, + 231 + ], + "type": "text", + "content": "X-Gaussian at 60s" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 229, + 177, + 276, + 224 + ], + "blocks": [ + { + "bbox": [ + 229, + 177, + 276, + 224 + ], + "lines": [ + { + "bbox": [ + 229, + 177, + 276, + 224 + ], + "spans": [ + { + "bbox": [ + 229, + 177, + 276, + 224 + ], + "type": "image", + "image_path": "e017efe17023ff92f513f599338661ae4cfa4d56190a242a313309412bffc47f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 224, + 272, + 231 + ], + "lines": [ + { + "bbox": [ + 231, + 224, + 272, + 231 + ], + "spans": [ + { + "bbox": [ + 231, + 224, + 272, + 231 + ], + "type": "text", + "content": "X-Gaussian at 180s" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 276, + 177, + 323, + 224 + ], + "blocks": [ + { + "bbox": [ + 276, + 177, + 323, + 224 + ], + "lines": [ + { + "bbox": [ + 276, + 177, + 323, + 224 + ], + "spans": [ + { + "bbox": [ + 276, + 177, + 323, + 224 + ], + "type": "image", + "image_path": "35cf8f70b02f21ebca3633d16cbee4ca267e15cd7de87c303bc8921e2f16518c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 285, + 224, + 313, + 231 + ], + "lines": [ + { + "bbox": [ + 285, + 224, + 313, + 231 + ], + "spans": [ + { + "bbox": [ + 285, + 224, + 313, + 231 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 329, + 123, + 471, + 232 + ], + "blocks": [ + { + "bbox": [ + 131, + 239, + 323, + 283 + ], + "lines": [ + { + "bbox": [ + 131, + 239, + 323, + 283 + ], + "spans": [ + { + "bbox": [ + 131, + 239, + 323, + 283 + ], + "type": "text", + "content": "Fig.6: Convergence analysis of NAF [60] vs. our X-Gaussian. We visualize the rendered projections at 20s, 60s, and 180s of training. Our X-Gaussian shows faster and better convergence." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 329, + 123, + 471, + 232 + ], + "lines": [ + { + "bbox": [ + 329, + 123, + 471, + 232 + ], + "spans": [ + { + "bbox": [ + 329, + 123, + 471, + 232 + ], + "type": "image", + "image_path": "f051d1c3945bd126fa7106fb93b0c540bc07e0069f5239e01f306ef58925e6b6.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 239, + 474, + 283 + ], + "lines": [ + { + "bbox": [ + 327, + 239, + 474, + 283 + ], + "spans": [ + { + "bbox": [ + 327, + 239, + 474, + 283 + ], + "type": "text", + "content": "Fig. 7: Analysis of covariance. The mean covariance of 3D Gaussians on different scenes decreases when the number of training views increases." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 305, + 482, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 305, + 482, + 377 + ], + "spans": [ + { + "bbox": [ + 130, + 305, + 482, + 377 + ], + "type": "text", + "content": "Analysis of Covariance. We study how the shape of the 3D Gaussian point cloud changes with the number of training views in Fig. 7. As the number of training views increases, the mean covariance of 3D Gaussians that control the size of Gaussian point clouds decreases. This indicates that the 3D Gaussian point clouds gradually change from coarse to fine, thereby being more capable of representing fine-grained structures, such as a small tumor in the abdomen." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 131, + 394, + 217, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 394, + 217, + 407 + ], + "spans": [ + { + "bbox": [ + 131, + 394, + 217, + 407 + ], + "type": "text", + "content": "5 Limitation" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 130, + 418, + 480, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 480, + 442 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 480, + 442 + ], + "type": "text", + "content": "The main limitation of our method is that its memory cost is non-trivial, about 51.25 MB on average. It is challenging to deploy it on low-RAM mobile devices." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 131, + 458, + 220, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 458, + 220, + 471 + ], + "spans": [ + { + "bbox": [ + 131, + 458, + 220, + 471 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 130, + 481, + 482, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 482, + 626 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 482, + 626 + ], + "type": "text", + "content": "In this paper, we propose the first 3DGS-based framework, X-Gaussian, for X-ray novel view synthesis. Firstly, we design a radiative Gaussian point cloud model based on X-ray imaging properties. This model excludes the influence of view direction when fitting radiation intensity. For this model, we develop a GPU-friendly differentiable radiative rasterization CUDA kernel that renders projections at a faster speed than RGB rasterization. Secondly, we customize an initialization strategy, ACUI, that does not need to execute the SfM algorithm. Instead, ACUI uses the parameters of X-ray scanner to compute the extrinsic and intrinsic matrices, and then uniformly samples center points for 3D Gaussians within a cuboid enclosing the scanned object. Experiments demonstrate that our X-Gaussian significantly outperforms SOTA methods by over " + }, + { + "bbox": [ + 130, + 481, + 482, + 626 + ], + "type": "inline_equation", + "content": "6.5\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 481, + 482, + 626 + ], + "type": "text", + "content": " while enjoying " + }, + { + "bbox": [ + 130, + 481, + 482, + 626 + ], + "type": "inline_equation", + "content": "73\\times" + }, + { + "bbox": [ + 130, + 481, + 482, + 626 + ], + "type": "text", + "content": " faster inference speed and only requiring " + }, + { + "bbox": [ + 130, + 481, + 482, + 626 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 130, + 481, + 482, + 626 + ], + "type": "text", + "content": " of training time." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "Acknowledgements: This work was supported by the Lustgarten Foundation for Pancreatic Cancer Research and the Patrick J. McGovern Foundation Award." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 243, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 243, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 243, + 100 + ], + "type": "text", + "content": "Yuanhao Cai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "text", + "content": "1. Andersen, A.H., Kak, A.C.: Simultaneous algebraic reconstruction technique (sart): a superior implementation of the art algorithm. Ultrasonic imaging (1984)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 162, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 480, + 205 + ], + "type": "text", + "content": "2. Armato III, S.G., McLennan, G., Bidaut, L., McNitt-Gray, M.F., Meyer, C.R., Reeves, A.P., Zhao, B., Aberle, D.R., Henschke, C.I., Hoffman, E.A., et al.: The lung image database consortium (lidc) and image database resource initiative (idri): a completed reference database of lung nodules on ct scans. Medical physics (2011)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 206, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 480, + 237 + ], + "type": "text", + "content": "3. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: ICCV (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 238, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 238, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 238, + 481, + 259 + ], + "type": "text", + "content": "4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In: CVPR (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 261, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 261, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 138, + 261, + 480, + 281 + ], + "type": "text", + "content": "5. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-nerf: Anti-aliased grid-based neural radiance fields. In: ICCV (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 283, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 283, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 283, + 480, + 304 + ], + "type": "text", + "content": "6. Biguri, A., Dosanjh, M., Hancock, S., Soleimani, M.: Tigre: a matlab-gpu toolbox for cbct image reconstruction. Biomedical Physics & Engineering Express (2016)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 304, + 480, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 480, + 336 + ], + "type": "text", + "content": "7. Boone, J., Shah, N., Nelson, T.: A comprehensive analysis of coefficients for pendant-geometry cone-beam breast computed tomography. Medical physics (2004)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 337, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 337, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 337, + 480, + 358 + ], + "type": "text", + "content": "8. Boone, J.M., Nelson, T.R., Lindfors, K.K., Seibert, J.A.: Dedicated breast ct: radiation dose and image quality evaluation. Radiology (2001)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 360, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 360, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 360, + 480, + 380 + ], + "type": "text", + "content": "9. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: ECCV (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 480, + 402 + ], + "type": "text", + "content": "10. Chen, B., Ning, R.: Cone-beam volume ct breast imaging: Feasibility study. Medical physics (2002)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "type": "text", + "content": "1. Chen, Z., Funkhouser, T., Hedman, P., Tagliasacchi, A.: Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In: CVPR (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 436, + 480, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 436, + 480, + 457 + ], + "spans": [ + { + "bbox": [ + 138, + 436, + 480, + 457 + ], + "type": "text", + "content": "2. Cormack, A.M.: Representation of a function by its line integrals, with some radiological applications. Journal of applied physics (1963)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 458, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 458, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 458, + 480, + 479 + ], + "type": "text", + "content": "3. Cormack, A.M.: Representation of a function by its line integrals, with some radiological applications. ii. Journal of Applied Physics (1964)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 479, + 480, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 523 + ], + "type": "text", + "content": "4. Corona-Figueroa, A., Frawley, J., Bond-Taylor, S., Bethapudi, S., Shum, H.P., Willcocks, C.G.: Mednerf: Medical neural radiance fields for reconstructing 3d-aware ct-projections from a single x-ray. In: International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC) (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "text", + "content": "5. Elbakri, I.A., Fessler, J.A.: Segmentation-free statistical image reconstruction for polyenergetic x-ray computed tomography with experimental validation. Physics in Medicine & Biology (2003)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 556, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 556, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 138, + 556, + 480, + 577 + ], + "type": "text", + "content": "6. Feldkamp, L.A., Davis, L.C., Kress, J.W.: Practical cone-beam algorithm. Josa a (1984)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 578, + 375, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 375, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 375, + 589 + ], + "type": "text", + "content": "7. Guide, D.: Cuda c programming guide. NVIDIA (2013)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 590, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 610 + ], + "type": "text", + "content": "8. Hounsfield, G.N.: Computerized transverse axial scanning (tomography): Part 1. description of system. The British journal of radiology (1973)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 611, + 397, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 397, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 397, + 621 + ], + "type": "text", + "content": "9. Hounsfield, G.N.: Computed medical imaging. Science (1980)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "text", + "content": "20. Hu, S., Liu, Z.: Gauhuman: Articulated gaussian splatting from monocular human videos. arXiv preprint arXiv: (2023)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "text", + "content": "21. Hu, T., Liu, S., Chen, Y., Shen, T., Jia, J.: Efficientnerf efficient neural radian fields. In: CVPR (2023)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 161, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 161, + 91, + 447, + 102 + ], + "type": "text", + "content": "Radiative Gaussian Splitting for Efficient X-ray Novel View Synthesis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "22. Hu, W., Wang, Y., Ma, L., Yang, B., Gao, L., Liu, X., Ma, Y.: Tri-miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields. In: ICCV (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "type": "text", + "content": "23. Jiang, Y., Tu, J., Liu, Y., Gao, X., Long, X., Wang, W., Ma, Y.: Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. arXiv preprint arXiv:2311.17977 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "text", + "content": "24. Keetha, N., Karhade, J., Jatavallabhula, K.M., Yang, G., Scherer, S., Ramanan, D., Luiten, J.: Splatam: Splat, track & map 3d gaussians for dense rgb-d slam. arXiv preprint arXiv:2312.02126 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "text", + "content": "25. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "text", + "content": "26. Kingma, D.P., Ba, J.L.: Adam: A method for stochastic optimization. In: ICLR (2015)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "27. Klacansky, P.: Scientific visualization datasets (2022), https://klacansky.com/open-scivis-datasets/" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "text", + "content": "28. Kocabas, M., Chang, J.H.R., Gabriel, J., Tuzel, O., Ranjan, A.: Hugs: Human gaussian splats. arXiv preprint arXiv:2311.17910 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "text", + "content": "29. Kopanas, G., Philip, J., Leimkuhler, T., Drettakis, G.: Point-based neural rendering with per-view optimization. In: Computer Graphics Forum (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "text", + "content": "30. Li, R., Gao, H., Tancik, M., Kanazawa, A.: Nerfacc: Efficient sampling accelerates nerfs. In: ICCV (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "text", + "content": "31. Liang, Y., Yang, X., Lin, J., Li, H., Xu, X., Chen, Y.: Luciddreamer: Towards high-fidelity text-to-3d generation via interval score matching. arXiv preprint arXiv:2311.11284 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "text", + "content": "32. Liang, Z., Zhang, Q., Feng, Y., Shan, Y., Jia, K.: Gs-ir: 3d gaussian splatting for inverse rendering. arXiv preprint arXiv:2311.16473 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "text", + "content": "33. Liu, X., Zhan, X., Tang, J., Shan, Y., Zeng, G., Lin, D., Liu, X., Liu, Z.: Human-gaussian: Text-driven 3d human generation with gaussian splatting. arXiv preprint arXiv:2311.17061 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "text", + "content": "34. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "type": "text", + "content": "35. Manglos, S.H., Gagne, G.M., Krol, A., Thomas, F.D., Narayanaswamy, R.: Transmission maximum-likelihood reconstruction with ordered subsets for cone beam ct. Physics in Medicine & Biology (1995)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "text", + "content": "36. Matsuki, H., Murai, R., Kelly, P.H., Davison, A.J.: Gaussian splatting slam. arXiv preprint arXiv:2312.06741 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "text", + "content": "37. Mildenhall, B., Srinivasan, P., Tancik, M., Barron, J., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "text", + "content": "38. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM ToG (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 556, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 480, + 578 + ], + "type": "text", + "content": "39. Pan, J., Zhou, T., Han, Y., Jiang, M., et al.: Variable weighted ordered subset image reconstruction algorithm. International Journal of Biomedical Imaging (2006)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "text", + "content": "40. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: NeurIPS (2019)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "text", + "content": "41. Reiser, C., Szeliski, R., Verbin, D., Srinivasan, P., Mildenhall, B., Geiger, A., Barron, J., Hedman, P.: Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. TOG (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "42. Rückert, D., Wang, Y., Li, R., Idoughi, R., Heidrich, W.: Neat: Neural adaptive tomography. TOG (2022)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 243, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 243, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 243, + 100 + ], + "type": "text", + "content": "Yuanhao Cai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 659 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "43. Sauer, K., Bouman, C.: A local update strategy for iterative reconstruction from projections. TIP (1993)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 170 + ], + "type": "text", + "content": "44. Scarfe, W.C., Farman, A.G., Sukovic, P., et al.: Clinical applications of cone-beam computed tomography in dental practice. Journal-Canadian Dental Association (2006)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 480, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 480, + 202 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 480, + 202 + ], + "type": "text", + "content": "45. Sidky, E.Y., Pan, X.: Image reconstruction in circular cone-beam computed tomography by constrained, total-variation minimization. Physics in Medicine & Biology (2008)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 203, + 480, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 203, + 480, + 223 + ], + "spans": [ + { + "bbox": [ + 134, + 203, + 480, + 223 + ], + "type": "text", + "content": "46. Snavely, N., Seitz, S.M., Szeliski, R.: Photo tourism: exploring photo collections in 3d. In: SIGGRAPH (2006)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 224, + 480, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 224, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 134, + 224, + 480, + 245 + ], + "type": "text", + "content": "47. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 246, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 246, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 134, + 246, + 480, + 277 + ], + "type": "text", + "content": "48. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: CVPR (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 277, + 480, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 277, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 134, + 277, + 480, + 297 + ], + "type": "text", + "content": "49. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncell, E.P.: Image quality assessment: from error visibility to structural similarity. TIP (2004)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 298, + 480, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 298, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 134, + 298, + 480, + 329 + ], + "type": "text", + "content": "50. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 330, + 480, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 330, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 134, + 330, + 480, + 361 + ], + "type": "text", + "content": "51. Xie, T., Zong, Z., Qiu, Y., Li, X., Feng, Y., Yang, Y., Jiang, C.: Physgaus-sian: Physics-integrated 3d gaussians for generative dynamics. arXiv preprint arXiv:2311.12198 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 362, + 480, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 362, + 480, + 382 + ], + "spans": [ + { + "bbox": [ + 134, + 362, + 480, + 382 + ], + "type": "text", + "content": "52. Yan, C., Qu, D., Wang, D., Xu, D., Wang, Z., Zhao, B., Li, X.: Gs-slam: Dense visual slam with 3d gaussian splatting. arXiv preprint arXiv:2311.11700 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 383, + 480, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 383, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 134, + 383, + 480, + 415 + ], + "type": "text", + "content": "53. Yang, Z., Yang, H., Pan, Z., Zhu, X., Zhang, L.: Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 415, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 415, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 134, + 415, + 480, + 446 + ], + "type": "text", + "content": "54. Yariv, L., Hedman, P., Reiser, C., Verbin, D., Srinivasan, P.P., Szeliski, R., Barron, J.T., Mildenhall, B.: Bakedsdf: Meshing neural sdfs for real-time view synthesis. In: SIGGRAPH (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 447, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 447, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 134, + 447, + 480, + 479 + ], + "type": "text", + "content": "55. Yi, T., Fang, J., Wu, G., Xie, L., Zhang, X., Liu, W., Tian, Q., Wang, X.: Gaussian dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 479, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 479, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 134, + 479, + 480, + 500 + ], + "type": "text", + "content": "56. Yu, L., Zou, Y., Sidky, E.Y., Pelizzari, C.A., Munro, P., Pan, X.: Region of interest reconstruction from truncated data in circular cone-beam ct. TMI (2006)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 501, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 501, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 134, + 501, + 480, + 520 + ], + "type": "text", + "content": "57. Yugay, V., Li, Y., Gevers, T., Oswald, M.R.: Gaussian-slam: Photo-realistic dense slam with gaussian splatting. arXiv preprint arXiv:2312.10070 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 521, + 480, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 521, + 480, + 552 + ], + "spans": [ + { + "bbox": [ + 134, + 521, + 480, + 552 + ], + "type": "text", + "content": "58. Zang, G., Idoughi, R., Li, R., Wonka, P., Heidrich, W.: Intratomo: self-supervised learning-based tomography via sinogram synthesis and prediction. In: CVPR (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 553, + 480, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 553, + 480, + 585 + ], + "spans": [ + { + "bbox": [ + 134, + 553, + 480, + 585 + ], + "type": "text", + "content": "59. Zbijewski, W., Defrise, M., Viergever, M.A., Beekman, F.J.: Statistical reconstruction for x-ray ct systems with non-continuous detectors. Physics in Medicine & Biology (2006)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 586, + 480, + 605 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 586, + 480, + 605 + ], + "spans": [ + { + "bbox": [ + 134, + 586, + 480, + 605 + ], + "type": "text", + "content": "60. Zha, R., Zhang, Y., Li, H.: Naf: neural attenuation fields for sparse-view cbct reconstruction. In: MICCAI (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 134, + 606, + 480, + 637 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 606, + 480, + 637 + ], + "spans": [ + { + "bbox": [ + 134, + 606, + 480, + 637 + ], + "type": "text", + "content": "61. Zhang, T., Yu, H.X., Wu, R., Feng, B.Y., Zheng, C., Snavely, N., Wu, J., Freeman, W.T.: PhysDreamer: Physics-based interaction with 3d objects via video generation. In: ECCV (2024)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 134, + 638, + 480, + 659 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 638, + 480, + 659 + ], + "spans": [ + { + "bbox": [ + 134, + 638, + 480, + 659 + ], + "type": "text", + "content": "62. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: Ewa volume splatting. In: Proceedings Visualization, 2001. VIS'01. IEEE (2001)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "text", + "content": "Radiative Gaussian Splitting for Efficient X-ray Novel View Synthesis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Raindrop Clarity_ A Dual-Focused Dataset for Day and Night Raindrop Removal/64be914f-41c4-4c4e-9cfc-b492879688db_content_list.json b/2024/Raindrop Clarity_ A Dual-Focused Dataset for Day and Night Raindrop Removal/64be914f-41c4-4c4e-9cfc-b492879688db_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..dccdadb2865fec8a87f12d3425190487a3867fee --- /dev/null +++ b/2024/Raindrop Clarity_ A Dual-Focused Dataset for Day and Night Raindrop Removal/64be914f-41c4-4c4e-9cfc-b492879688db_content_list.json @@ -0,0 +1,1632 @@ +[ + { + "type": "text", + "text": "Raindrop Clarity: A Dual-Focused Dataset for Day and Night Raindrop Removal", + "text_level": 1, + "bbox": [ + 236, + 140, + 767, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yeying Jin $^{1}$ , Xin Li $^{2}$ , Jiadong Wang $^{1}$ , Yan Zhang $^{1}$ , and Malu Zhang $^{3}$", + "bbox": [ + 217, + 210, + 781, + 228 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{1}$ National University of Singapore", + "2 University of Science and Technology of China", + "3 University of Electronic Science and Technology of China jinyeying@u.nus.edu, xin.li@ustc.edu.cn, maluzhang@uestc.edu.cn" + ], + "bbox": [ + 256, + 238, + 743, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Existing raindrop removal datasets have two shortcomings. First, they consist of images captured by cameras with a focus on the background, leading to the presence of blurry raindrops. To our knowledge, none of these datasets include images where the focus is specifically on raindrops, which results in a blurry background. Second, these datasets predominantly consist of daytime images, thereby lacking nighttime raindrop scenarios. Consequently, algorithms trained on these datasets may struggle to perform effectively in raindrop-focused or nighttime scenarios. The absence of datasets specifically designed for raindrop-focused and nighttime raindrops constrains research in this area. In this paper, we introduce a large-scale, real-world raindrop removal dataset called Raindrop Clarity. Raindrop Clarity comprises 15,186 high-quality pairs/triplets (raindrops, blur, and background) of images with raindrops and the corresponding clear background images. There are 5,442 daytime raindrop images and 9,744 nighttime raindrop images. Specifically, the 5,442 daytime images include 3,606 raindrop- and 1,836 background-focused images. While the 9,744 nighttime images contain 4,838 raindrop- and 4,906 background-focused images. Our dataset will enable the community to explore background-focused and raindrop-focused images, including challenges unique to daytime and nighttime conditions.", + "bbox": [ + 259, + 325, + 754, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Raindrop-focused $\\cdot$ Background-focused $\\cdot$ Nighttime", + "bbox": [ + 261, + 616, + 697, + 631 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 652, + 375, + 669 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adherent raindrops on lenses or windscreens can significantly reduce visibility. Raindrop removal is crucial for surveillance, self-driving cars, object detection, outdoor photography, filmmaking, augmented reality, etc. Recently, a few rain-drop datasets (e.g., [24-26, 30]) have been proposed. In Fig. 1, these datasets primarily contain daytime background scenes with blurry raindrops, as the images are taken with the camera's focus on the background. They have impacts in the raindrop removal area, inspiring several successful methods [24, 25, 27].", + "bbox": [ + 212, + 683, + 787, + 790 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 217, + 797, + 383, + 811 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Our data and code are available at: https://github.com/jinyeying/RaindropClarity", + "bbox": [ + 218, + 811, + 785, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2a28196a7e79f658d09f317b322114916a74e7f5ea7307afe6b112d0824e465c.jpg", + "image_caption": [ + "(a) Qian" + ], + "image_footnote": [], + "bbox": [ + 303, + 143, + 411, + 258 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2ab26cdc364edff5e64ea4b27c258ce6fd0a412e0217b41a1d9ab5b85b670ad1.jpg", + "image_caption": [ + "(b) RainDS Quan" + ], + "image_footnote": [], + "bbox": [ + 413, + 146, + 522, + 258 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/827e2afd6998367cec9353397a2f8cf683a58932f2422a2228faaedd4abcef03.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 146, + 604, + 202 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/051bf678be361065f2bc54067f67a02eb79a431d7e4088d8b25235e2f57a1591.jpg", + "image_caption": [ + "(c) RobotCar Porav", + "(d) Windshield" + ], + "image_footnote": [], + "bbox": [ + 522, + 203, + 604, + 258 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f80e7ec8b1c18f776fbc2d80ce54fd4287a1a095c868abee17d129ad6a6c6e36.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 146, + 697, + 202 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5de40a39599e5a1658d34b98f25e8a791fff58dbc087a43d8d48803620d25c2c.jpg", + "image_caption": [ + "Fig. 1. Existing raindrop datasets (e.g., [24-26,30]) exhibit two limitations: first, they do not include raindrop-focused images (Fig. 2 right); and second, they lack night raindrops (Fig. 2 bottom).", + "Fig. 2. Motivation for Raindrop Clarity: real images sourced from the Internet feature scenarios overlooked by existing datasets, including raindrop-focused (right) and nighttime raindrops (bottom). In this figure, $\\mathrm{Rd} =$ Raindrop, $\\mathrm{Bg} =$ Background." + ], + "image_footnote": [], + "bbox": [ + 292, + 340, + 709, + 465 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While these datasets are valuable, they have two limitations. First, a significant issue arises when the camera is in automatic focus mode. In Fig. 2 (right), the camera may unintentionally focus on raindrops that are attached to the lens, windshield, or glass, resulting in images with sharp raindrops and a blurry background. Consequently, methods designed for images with blurry raindrops and a sharp background may not perform optimally. Second, existing datasets consist only of daytime images. Therefore, existing methods, especially those reliant on training data, may struggle to effectively handle nighttime raindrop images because of the inherent domain gap between day and night conditions.", + "bbox": [ + 212, + 551, + 787, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In Fig. 2, we observe that besides daytime images with blurry raindrops and a sharp background, it is common to come across images with sharp raindrops and a blurry background (right), as well as nighttime raindrops (bottom). Thus, we collected varied raindrop images as shown in Fig. 3: the top-left with daytime blurry raindrops against a sharp background, the top-right with daytime sharp raindrops and a blurry background, the bottom-left with nighttime blurry raindrops and a clear background, and the bottom-right with nighttime sharp raindrops with a blurry background.", + "bbox": [ + 212, + 688, + 787, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Physical solutions like rain covers or wipers can reduce the presence of raindrops, but their effectiveness is limited. In certain situations where wipers cannot", + "bbox": [ + 214, + 810, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Jin, X. Li et al.", + "bbox": [ + 271, + 114, + 395, + 127 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3e45e130b4dbc90f0eccd49d28b114afa6310ad369f1cba5d88c984f7c5caa5a.jpg", + "image_caption": [ + "Background-focused (Blurry raindrop)", + "Fig. 3. Raindrop Clarity: Examples of our pairs $(\\tilde{\\mathbf{x}}, b_0)$ and triplets $(\\tilde{\\mathbf{x}}, x_0, b_0)$ of a raindrop image $\\tilde{\\mathbf{x}}$ , the corresponding blurry background image $x_0$ and the corresponding clear background image $b_0$ , in day and night. For the pairs $(\\tilde{\\mathbf{x}}, b_0)$ , there is no $x_0$ , since in this case $x_0 = b_0$ . The reason we have the blurry background image $x_0$ in triplets is to have the raindrop difference maps $\\tilde{\\mathbf{m}}$ (see Fig. 8)." + ], + "image_footnote": [], + "bbox": [ + 218, + 157, + 506, + 265 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/56d15b79ab79137bc18d8d135fae2c13ebbb686963fcb4ed55f97210e4b15528.jpg", + "image_caption": [ + "Raindrop-focused (Blurry background)" + ], + "image_footnote": [], + "bbox": [ + 513, + 157, + 782, + 265 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "be used, or wind blows rain onto the lens, these physical measures become impractical. Physics-based methods [1,8,23,44,45] remove raindrops by leveraging the physics of raindrops or utilizing generated synthetic data. However, the effectiveness of methods trained on synthetic data relies on the quality of raindrop rendering models. Currently, raindrop rendering models commonly employ surface modeling and ray tracing [8,28,29]. Unfortunately, these methods may not fully consider factors such as scene depth, lighting conditions, and the 3D shapes of raindrops [23,28], which are crucial for realistic raindrop rendering. Additionally, raindrops can introduce refraction effects [7,44], distorting the shapes and positions of objects, which further complicates accurate raindrop rendering. To the best of our knowledge, developing a rendering approach that can produce physically accurate raindrops adaptable to various scenes is challenging.", + "bbox": [ + 212, + 381, + 787, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivated by the aforementioned challenges and the observation that existing raindrop datasets predominantly consist of background-focused images while neglecting raindrop-focused images and nighttime raindrops, we collect a novel dataset, called \"Raindrop Clarity\". Our dataset comprises 6,742 (1,836 daytime, 4,906 nighttime) pairs and 8,444 (3,606 daytime, 4,838 nighttime) triplets of raindrop-focused images and background-focused images. All of them have the corresponding clear background images. The triplets include one more raindrop-free blurry background images. Fig. 3 shows samples and explanation on the meaning of our pairs and triplets. From the 15,186 raindrops pairs/triplet, 5,442 are captured during daytime and the 9,744 at nighttime. Raindrops Clarity exhibits diverse sizes and shapes, and raindrop appearance varies based on the scene and lighting conditions. To our knowledge, Raindrop Clarity is the largest real-world raindrop dataset that contains both the pair and triplets, day and night. Our main contributions are summarized as follows:", + "bbox": [ + 212, + 565, + 787, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We present the Raindrop Clarity dataset, a large-scale real-world dataset that encompasses both raindrop-focused triplets and background-focused pairs. For the pairs, we provide images with blurry raindrops and their cor", + "bbox": [ + 223, + 794, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Raindrop Clarity: Dual-Focused Dataset for Day Night Raindrop Removal", + "bbox": [ + 236, + 113, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "responding clear backgrounds. For the triplets, we additionally provide the raindrop-free blurry background.", + "bbox": [ + 240, + 146, + 784, + 176 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Raindrop Clarity includes images collected under both daytime and nighttime. It aims to remove raindrops and recover clear backgrounds regardless of day or night, whether the raindrops or background are blurry.", + "- Our experimental results indicate that even with the application of state-of-the-art raindrop removal methods, there are instances where these methods fail. This suggests the existence of unresolved challenges in the field of raindrop removal, which await further attention and investigation." + ], + "bbox": [ + 225, + 176, + 782, + 280 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 306, + 387, + 323 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Raindrop Dataset. Qian et al. [25] introduce a raindrop dataset consisting of raindrop and corresponding ground truths. Porav et al. [24] propose a stereo RobotCar dataset, using a rain-making device that causes one lens to be affected by raindrops, while the other lens remains clear. Hao et al. [8] create a synthetic dataset of adherent raindrops with pixel-level masks for training deraining algorithms. Quan et al. [26] introduce RainDS, a dataset that includes raindrops, rain steaks, and their corresponding ground truths. Soboleva and Shipitko [30] introduce Windshields, a dataset containing raindrop images on both camera lenses and windshields, annotated with binary masks representing the raindrop areas. However, all of these datasets primarily consist of daytime images taken by cameras focusing on the background scenes and thus have blurry raindrops.", + "bbox": [ + 212, + 340, + 784, + 505 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Single Image Raindrop Removal. Most image deraining methods have primarily focused on removing rain streaks rather than raindrops. However, recent learning-based methods have been proposed specifically for single-image raindrop removal. Eigen et al. [6] introduce a supervised CNN-based method to remove raindrops, using a shallow CNN. Qian et al. [25] propose an AttentiveGAN for raindrop removal, that relies on attention maps but has a limited focus on local spatial information. Quan et al. [27] propose RaindropAttention, which employs an attention mechanism to generate attention maps based on mathematical raindrop shapes, yet it lacks exploration of long-range information. Liu et al. [18] demonstrate a dual residual network, by leveraging the complementary nature of paired residual blocks. Hao et al. [8] propose a raindrop synthesis algorithm, used their synthetic dataset to train a multi-task network for raindrop detection and removal. Porav et al. [24] propose a Pix2PixHD-based deraining network architecture. Subsequent work by Quan et al. [26] propose a cascaded network architecture for simultaneous removal of rain streaks and raindrops, achieved through a neural architecture search. Yasarla and Patel [41] propose the uncertainty guided multi-scale residual learning network to effectively remove rain from images. Zhang et al. [47] propose the Dual Attention-in-Attention Model to remove rain streaks and raindrops.", + "bbox": [ + 212, + 507, + 784, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Restoration Backbone. There are image restoration and enhancement tasks that utilize transformer-based models, taking advantage of the long-range dependencies, including IPT [2] (vanilla transformer), Uformer [36], SwinIR [17]", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Jin, X. Li et al.", + "bbox": [ + 271, + 114, + 395, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e0cc4a2196a527ed30aa7e521e72ab8fdc4ff802fd543412780a6ac3e07ba60c.jpg", + "image_caption": [ + "Background-focused (Blurry raindrop)", + "Fig. 4. Pairs $(\\tilde{\\mathbf{x}},\\pmb{b}_0)$ show various sizes, shapes, and densities." + ], + "image_footnote": [], + "bbox": [ + 303, + 157, + 699, + 305 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b9a641c25144839a2280d1c9d874fd44c468af638a5b4c748589affc2fe33c90.jpg", + "image_caption": [ + "Raindrop-focused (Blurry background)", + "Fig. 5. The triplets $(\\tilde{\\mathbf{x}}, \\mathbf{x}_0, \\mathbf{b}_0)$ cover diverse scenes and shapes." + ], + "image_footnote": [], + "bbox": [ + 302, + 407, + 699, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and Restormer [46]. Xiao et al. [40] propose an image deraining transformer (IDT) for rain streak and raindrop removal with window-based and spatial-based dual Transformer. RaindropDiffusion [21] develops a UNet-based diffusion framework, which demonstrates robust generation capabilities for raindrops, rain streaks, and snow. RainDiffusion [38], an unsupervised diffusion model, utilizes unpaired real-world data instead of weakly adversarial training. DiT [22] is a transformer-based diffusion model.", + "bbox": [ + 212, + 733, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Raindrop Clarity: Dual-Focused Dataset for Day Night Raindrop Removal", + "bbox": [ + 236, + 113, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Raindrop Clarity Dataset", + "text_level": 1, + "bbox": [ + 215, + 143, + 498, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Data Collection", + "text_level": 1, + "bbox": [ + 215, + 176, + 392, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To create Raindrop Clarity, we employed a sphere pan-tilt platform to keep our camera stationary. The images were captured using Sony FDR-AX33 4K Ultra HD Handycam Camcorder, Sony alpha 7R III digital camera, the back camera of an iPhone 14 Pro and iPhone 15 Pro Max, ensuring the highest quality possible. The 4K Camcorder captured RGB images at a resolution $3840 \\times 2164$ .", + "bbox": [ + 212, + 204, + 785, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Using the optical refraction model (raindrops appear upside-down), we projected raindrops by spraying water on glass and adjusting the distance, angle, speed, and patterns of the nozzle/wipers. We also captured in rain conditions.", + "bbox": [ + 212, + 280, + 785, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our collection procedure was designed to ensure diversity in the raindrop (such as focuses, shapes and sizes), background scenes and lighting conditions. First, we fixed the camera position on a tripod and installed a glass plate. The distance between the camera and the glass plate ranged from $5\\mathrm{cm}$ to $25\\mathrm{cm}$ . Second, we sprayed waterdrops/raindrops onto the glass plate and focused the camera on the raindrops. Third, we sprayed the waterdrops/raindrops to capture images of them under different shapes and sizes, ranging from sparse to dense. The captured images include elliptical raindrops and raindrop flow traces $(\\tilde{\\mathbf{x}})$ . Fourth, we removed the glass plate, allowing the camera to focus on the very close plane. The duration is kept very short, within milliseconds, to minimize lighting changes. In this way, we obtained raindrop-free blurry background images $(\\boldsymbol{x}_0)$ . Finally, we adjusted the camera focus to the background, directing the camera to focus on the distant background $(\\boldsymbol{b}_0)$ .", + "bbox": [ + 212, + 325, + 787, + 522 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Characteristics of Raindrop Clarity", + "text_level": 1, + "bbox": [ + 215, + 544, + 552, + 560 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Raindrop Clarity is a new dataset consisting of around 15,186 high-quality image raindrop pairs and triplets. In Fig. 3, our dataset introduces a unique element of triplet data, which includes raindrop-focused images, corresponding blurry raindrop-free background images, and their clear background equivalents, distinguishing it from existing datasets. The second feature that distinguishes the Raindrop Clarity from the existing ones is the collection of night raindrops.", + "bbox": [ + 212, + 571, + 785, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Objectives. Our dataset is designed to support research in the elimination of raindrops and the restoration of a clear background, irrespective of whether the camera's focus is on the raindrops (leading to a blurry background) or on the background (resulting in blurry raindrops). This holds true regardless of whether it is daytime or nighttime.", + "bbox": [ + 212, + 675, + 785, + 751 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Raindrops and Diversity. Our Raindrop Clarity dataset offers an extensive coverage of diverse raindrop sizes, shapes, and occlusion types in various scenes and lighting conditions, making it an exceptionally comprehensive resource for advancing raindrop removal research. In Fig. 4 and Fig. 5, the dataset captures raindrops and rain flows of different sizes and shapes, including heavy raindrops", + "bbox": [ + 212, + 763, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Jin, X. Li et al.", + "bbox": [ + 271, + 114, + 395, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Background-focused (Blurry raindrop)", + "text_level": 1, + "bbox": [ + 387, + 143, + 622, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/47bce7102b8f68aee4e427203c83661aa5c5a3c9c56a0d532d369558d5d3a1f9.jpg", + "image_caption": [ + "Fig. 6. The pairs $(\\tilde{\\mathbf{x}}, b_0)$ under different lighting conditions." + ], + "image_footnote": [], + "bbox": [ + 303, + 157, + 699, + 303 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Raindrop-focused (Blurry background)", + "text_level": 1, + "bbox": [ + 387, + 344, + 625, + 358 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d249440cd2567605a4ec72dcfa257fba87508684a1e32cfca34a75bbe140253c.jpg", + "image_caption": [ + "Fig. 7. Triplets $(\\tilde{\\mathbf{x}},\\pmb {x}_0,\\pmb {b}_0)$ under different lighting conditions." + ], + "image_footnote": [], + "bbox": [ + 300, + 358, + 697, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "that result in unpredictable water flows and blurry backgrounds when they hit window panes or car windshields. Two primary types of occlusions caused by raindrops, namely partial and complete occlusions, are included in our dataset. Both types present unique challenges for raindrop removal algorithms.", + "bbox": [ + 212, + 635, + 787, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scenes and Large-scale. The dataset encompasses a wide range of background scenes, such as urban, countryside, campus, wilderness, road, and aerial views, covering both day and night. This variation allows for the exploration of different lighting and weather conditions in different scenes. With approximately 38,816 images in total, including 15,186 high-quality pairs/triplets of images, the Raindrop Clarity dataset is one of the largest datasets of its kind.", + "bbox": [ + 212, + 707, + 787, + 799 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Lighting Conditions. The appearance of raindrops highly depends on the surrounding environment. A distinctive feature of our dataset is its comprehen", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Raindrop Clarity: Dual-Focused Dataset for Day Night Raindrop Removal", + "bbox": [ + 236, + 113, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cc2e87abdb21276293405bcef7c1766af32b89157de6ac621f6a76fe7f400430.jpg", + "table_caption": [ + "Table 1. Summary of comparisons between our Raindrop Clarity and existing datasets. Bg-focus = Background-focused, Rd-focus = Raindrop-focused, RD = Raindrop. Raindrop Clarity uniquely features raindrop-focused triplets and the nighttime raindrops." + ], + "table_footnote": [], + "table_body": "
DatasetsComponentsN. ImagesSyn/RealBg-focusRd-focus
DayNightDayNight
Our RD Clarity (Day)3,606 Triplets+1,836 Pairs14,490Real✓×✓×
Our RD Clarity (Night)4,838 Triplets+4,906 Pairs24,326Real×✓×✓
Raindrop Qian [25]919 Pairs1,838Real✓×××
RainDS-Real [26]248 Quadruplets992Real✓×××
RainDS-Syn [26]1,200 Quadruplets4,800Syn✓×××
RobotCar Porav [24]4,818 Pairs9,636Real✓×××
Windshield [30]3,3903,390Real✓×××
", + "bbox": [ + 220, + 198, + 787, + 323 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/af482284e479f43a9d1193198ddcecedba50c33dfd8f36183ff77d445dfba4df.jpg", + "image_caption": [ + "Fig. 8. We show $(\\tilde{\\mathbf{x}},\\pmb{x}_0,\\tilde{\\mathbf{m}},\\pmb{b}_0)$ , difference map $\\tilde{\\mathbf{m}}$ is derived by distinguishing raindrop-focused $\\tilde{\\mathbf{x}}$ from blurry background $\\pmb{x}_0$ ." + ], + "image_footnote": [], + "bbox": [ + 302, + 334, + 702, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sive coverage of various environmental conditions, setting it apart from other datasets. The dataset's variation in day and night conditions makes it valuable for developing raindrop removal algorithms.", + "bbox": [ + 212, + 558, + 784, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During the daytime, raindrop images exhibit a diverse array of appearances influenced by multiple factors such as sunlight direction, intensity, and shadowing. In Figs. 4-5, our dataset captures this variety, with images taken at different times of the day and under varying weather conditions (such as light rain, heavy rain, and rain combined with wind), providing rich information for algorithms to learn.", + "bbox": [ + 212, + 604, + 784, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unlike daytime, nighttime raindrops present unique challenges due to artificial lighting conditions and low light. The appearance of raindrops can drastically change under street lights, headlights, or neon lights. In Fig. 6 and Fig. 7, our dataset includes night images with varied artificial lights, contributing to the complexity and comprehensiveness of the dataset. These night images are challenging due to the higher contrast between raindrops and the background, and potential overexposure or underexposure issues.", + "bbox": [ + 212, + 694, + 785, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Background- and Raindrop-focused. When the raindrop image $\\tilde{\\mathbf{x}}$ is background-focused, we decompose a raindrop degraded image $\\tilde{\\mathbf{x}}$ into a clear background", + "bbox": [ + 214, + 809, + 812, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Jin, X. Li et al.", + "bbox": [ + 271, + 114, + 395, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7f5b04c3855e7d9c0ef331fe024be4eaf97c211097e8eb0f93e4f506c4dc70e5.jpg", + "image_caption": [ + "Fig. 9. Comparison of existing methods on daytime raindrop-focused images, covering both raindrop algorithms and restoration backbones." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 450 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "image $\\pmb{b}_0$ by following this model [25]: $\\tilde{\\mathbf{x}} = (1 - \\mathbf{M})\\odot \\pmb{x}_0 + \\mathbf{D}$ , where $\\mathbf{D}$ is the effect of the blurry raindrops, referring to the localized blurry imagery formed by the light reflected from the surroundings. The operator $\\odot$ denotes element-wise multiplication. When binary mask $\\mathbf{M}(\\mathbf{x})$ equals 1, it signifies that the pixel $x$ belongs to the area affected by raindrops. Conversely, if $\\mathbf{M}(\\mathbf{x})$ does not equal 1, it indicates that the pixel is within the background areas. In the case where $\\tilde{\\mathbf{x}}$ is background-focused raindrop images, $\\pmb{x}_0$ is identical to the clear, raindrop-free image $\\pmb{b}_0$ , namely $\\pmb{x}_0 = \\pmb{b}_0$ . Therefore, we use $(\\tilde{\\mathbf{x}},\\pmb{b}_0)$ to stand for background-focused (resulting in blurry raindrops) pairs, shown in the left part of Fig. 3.", + "bbox": [ + 212, + 521, + 787, + 656 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In the right part of Fig. 3 and Fig. 8, when $\\tilde{\\mathbf{x}}$ is the raindrop-focused image, $\\pmb{x}_0$ is the corresponding raindrop-free blurry background image, while $\\pmb{b}_0$ is the clear background image. In this case, $\\pmb{x}_0 \\neq \\pmb{b}_0$ . We can calculate the pixel-level difference map $\\tilde{\\mathbf{m}}$ between the raindrop-free blurry background $\\pmb{x}_0$ and the raindrop-focused $\\tilde{\\mathbf{x}}$ as: $\\tilde{\\mathbf{m}} = \\tilde{\\mathbf{x}} - \\pmb{x}_0$ . Therefore, we use $(\\tilde{\\mathbf{x}}, \\pmb{x}_0, \\pmb{b}_0)$ to stand for raindrop-focused (resulting in blurry background) triplets.", + "bbox": [ + 212, + 657, + 787, + 748 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Since the camera is stationary, neither the pairs $(\\tilde{\\mathbf{x}},\\boldsymbol{b}_0)$ nor the triplets $(\\tilde{\\mathbf{x}},\\boldsymbol{x}_0,\\boldsymbol{b}_0)$ have alignment problems. This is also evident in the raindrop difference maps $\\tilde{\\mathbf{m}}$ , which accurately assign values to raindrop regions across multiple images. Specifically, for each pixel $\\mathbf{x}$ in the input image, any discrepancies in the map's assigned values for the static background would indicate alignment issues; however, these are not observed, as $\\tilde{\\mathbf{m}}$ indicates only the raindrop regions.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Raindrop Clarity: Dual-Focused Dataset for Day Night Raindrop Removal", + "bbox": [ + 236, + 113, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/2834ab7e804f67a3e48f994a78ae66bc59942b962fde84518dc14de3647a90e4.jpg", + "table_caption": [ + "Table 2. Quantitative evaluation of daytime and nighttime raindrop data with a benchmark for raindrop algorithms and restoration backbones." + ], + "table_footnote": [], + "table_body": "
Data\\Method InputRaindrop AlgorithmsRestoration Backbones
AtGAN [25]Robot [24]RainDS [26]IDT [40]RDdiff [21]Restor. [46]Uform. [36]DiT [22]
DayPSNR↑ 21.9223.6223.8225.3926.0525.5226.0825.7826.03
SSIM↑ 0.5600.6580.6700.7170.7360.7340.7480.7270.752
LPIPS↓ 0.2470.2000.1910.1470.1410.1110.1310.1480.106
NightPSNR↑ 24.7824.3824.5725.3926.8126.4826.8725.2626.23
SSIM↑ 0.7260.7730.7870.8200.8510.8310.8480.8160.826
LPIPS↓ 0.2090.1850.1820.1390.1250.1120.1230.1420.111
", + "bbox": [ + 218, + 184, + 787, + 311 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/331fa48fad1a23041aba7ee1afa5253c7288218d5e315086edca017ac6b088a6.jpg", + "image_caption": [ + "Fig. 10. Comparison of methods on nighttime raindrop-focused images, covering both raindrop algorithms and restoration backbones." + ], + "image_footnote": [], + "bbox": [ + 217, + 343, + 785, + 651 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Since the pairs/triplets of images are captured in quick succession, there are no problems with changes in illumination.", + "bbox": [ + 212, + 739, + 782, + 768 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The captured images do not contain any personal information such as phone numbers, or human faces. We will release the dataset publicly with car license plates obscured, as we believe that our Raindrop Clarity dataset will benefit the progress of the raindrop removal field.", + "bbox": [ + 212, + 779, + 784, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Jin, X. Li et al.", + "bbox": [ + 271, + 114, + 395, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/9f4e1ebd7e957d23f62ae571182b474f35f4ec8565594f6d54da0c0268d81c1d.jpg", + "image_caption": [ + "Fig. 11. Comparison of methods on daytime background-focused images, covering both raindrop algorithms and restoration backbones." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 378 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 445, + 375, + 463 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Experimental Setup. Our benchmark includes two parts: Daytime and Night-time Raindrop Removal. As shown in Table 1, our daytime raindrop dataset contains a total of 5,442 paired/triplets images, where 4,713 paired images are used for the training set, and the remaining 729 paired images are regarded as the test set. Our nighttime raindrop dataset consists of 9,744 paired/triplets images, where 8,655 paired images are selected as the training set and the rest 1,089 paired images are used for evaluation.", + "bbox": [ + 212, + 478, + 787, + 585 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Baselines. To create the benchmark, we retrain existing state-of-the-art methods on our daytime and nighttime raindrop datasets. Our benchmark consists of five raindrop algorithms [21, 24-26, 40] and three restoration methods [22, 36, 46]. Specifically, IDT [40], RDiffusion [21], RainDS [26], Robot Car [24] and AtGAN [25] are raindrop algorithms, while Restormer [46], Uformer [36] and DiT [22] are restoration backbones. For a fair comparison, we conduct experiments using the original implementation of these methods. All experiments are implemented in PyTorch and are trained on four NVIDIA RTX A5000 GPUs.", + "bbox": [ + 212, + 598, + 787, + 720 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison with Existing Datasets. The difference between our Raindrop Clarity dataset and the existing datasets is shown in Table. 1 and Fig. 2.", + "bbox": [ + 212, + 733, + 785, + 762 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The stereo RobotCar dataset [24] includes 4,818 image pairs, randomly selected for training, validation, and testing. However, the raindrops in this dataset are disproportionately large and lack realism.", + "bbox": [ + 212, + 763, + 785, + 808 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The raindrop dataset proposed by Qian et al. [25] consists of 861 training pairs and 58 testing pairs. While it features common raindrop shapes, it lacks the", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Raindrop Clarity: Dual-Focused Dataset for Day Night Raindrop Removal", + "bbox": [ + 236, + 113, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ca0b3af94eff70f0db6b6db57dcccd8e2d0dab2ea7ae8bf2aad48a14a0a26c7f.jpg", + "image_caption": [ + "Fig. 12. Comparison of methods on nighttime background-focused images, covering both raindrop algorithms and restoration backbones." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 522 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "diversity of real-world scenes [44]. Also, the backgrounds between the raindrop images and the ground truth are not perfectly aligned [39].", + "bbox": [ + 212, + 594, + 785, + 625 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The RainDS-Real dataset, proposed by Quan et al. [26], comprises 248 quadruplets, including raindrops, rain streaks, and their corresponding ground truths. The RainDS-Synthetic dataset has a domain gap with the real raindrop images. Moreover, the real-world raindrops are consistently small, obscuring only a fraction of the image details, making recovery less challenging.", + "bbox": [ + 212, + 626, + 789, + 700 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The Windshield dataset, proposed by Soboleva and Shipitko [30], was captured by a camera attached to the vehicle during its movement in urban areas and highway environments. The Windshield does not provide the corresponding clean pairs, making this dataset not available for robust raindrop removal method testing. Additionally, frequent moving objects further complicate the dataset.", + "bbox": [ + 212, + 702, + 785, + 792 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Existing raindrop datasets, though valuable, have certain limitations. They do not include raindrop-focused and nighttime raindrops. In addition, we have noticed reflection errors in the datasets from Qian [25] and RainDS-Real [26].", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Jin, X. Li et al.", + "bbox": [ + 271, + 114, + 395, + 127 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison with Synthetic Raindrops. In Fig. 13, we show the existing state-of-the-art synthetic physical model [23], although it can generate raindrops on blurry and clear backgrounds (Fig. 13 d generated on b, Fig. 13 e generated on c). It still has a gap with real-world collected raindrops. Real-world collected raindrops display diverse appearances, including various shapes such as round, elliptical, and long strips. Also, the transparency and glare effects of raindrops collected in the real-world varies.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Quantitative Comparison. To effectively evaluate the results on our proposed datasets, we use three popular metrics PSNR, SSIM [37], and LPIPS [48] in generation tasks [3-5, 9-16, 19, 20, 31-35, 42, 43].", + "bbox": [ + 212, + 266, + 785, + 311 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The qualitative results are shown in Table 2. For the daytime dataset, the PSNR and SSIM of the original input images are 21.92 and 0.560. Raindrop algorithm IDT [40] achieves the best performance (26.05 PSNR of and 0.736 of SSIM) compared with other raindrop algorithms. For the restoration backbones, DiT [22] achieves a PSNR of 26.03 and a SSIM of 0.752. Restormer [46] achieves a PSNR of 26.08 and a SSIM of 0.748, which also performs competitively.", + "bbox": [ + 212, + 311, + 787, + 402 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "It can be observed that, for the nighttime dataset, PSNR and SSIM of the input images are 24.78 and 0.726. Among the raindrop algorithms, IDT [40] achieves better performance than other methods, with a PSNR of 26.81 and a SSIM of 0.851. For the restoration backbones, Restormer [46] achieves the highest PSNR of 26.87, while DiT [22] achieves the lowest LPIPS of 0.111, which outperforms other backbones.", + "bbox": [ + 212, + 402, + 787, + 494 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Qualitative Comparison. Figs. 9-12 show the qualitative results of state-of-the-art methods. To be specific, Fig. 9 and Fig. 11 show the results of daytime raindrop removal. The input images in Fig. 9 and Fig. 11 are daytime raindrop-focused and background-focused, respectively. Fig. 10 and Fig. 12 show the results of nighttime raindrop-focused and background-focused removal, respectively.", + "bbox": [ + 212, + 507, + 787, + 597 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "It can be found that, for daytime and nighttime raindrop-focused removal in Fig. 9 and Fig. 10, existing state-of-the-art methods fail to remove raindrops and recover the background. This is because most of them are designed for background-focused scenes and thus neglect the blurry problem. To our knowledge, Restormer [46] and Uformer [36] are specifically designed to address deblurring problem. However, they still struggle with recovering details and textures. As a result, these methods are struggling to simultaneously remove raindrops and recover the background.", + "bbox": [ + 212, + 598, + 787, + 718 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Fig. 11, for daytime background-focused inputs, we can observe that existing state-of-the-art methods can handle most raindrops. However, they are struggling to address long strips of raindrops. This is because their methods may not detect long strips of raindrops and thus fail to remove them. Furthermore, in Fig. 12, we find that existing state-of-the-art methods are struggling to handle nighttime background-focused samples. To be specific, raindrops in these samples cannot be totally removed and details of the restored images are not clear. This is because nighttime raindrops usually reflect artificial light and thus", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Raindrop Clarity: Dual-Focused Dataset for Day Night Raindrop Removal", + "bbox": [ + 236, + 113, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/dee761b1207cfd6fcf196b74c1496cceaaeac553810491ff83e01575d5da90bc.jpg", + "image_caption": [ + "a Raindrop-focusedd $\\tilde{\\mathbf{x}}$" + ], + "image_footnote": [], + "bbox": [ + 218, + 145, + 330, + 239 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/ffacb94dee9a2664638cdb81258fb4a0bac29230b2eb3de4d5b0ae5516dadfb8.jpg", + "image_caption": [ + "b Blurry background $\\tilde{\\pmb{x}}$" + ], + "image_footnote": [], + "bbox": [ + 333, + 146, + 441, + 239 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/6086c7a2dec2d32060d361d0477405b2d20a3356077595350cafd12ece48e753.jpg", + "image_caption": [ + "c Clear background $\\mathbf{b}_0$", + "Fig. 13. We show the differences between the synthetic raindrops [23] and real-world raindrops from Raindrop Clarity." + ], + "image_footnote": [], + "bbox": [ + 442, + 146, + 553, + 239 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/c401688e2e5e8b4f030a28a47f5b357e392d44d889e71c79d06037b9a389285a.jpg", + "image_caption": [ + "d Synthetic Rd on $\\tilde{\\mathbf{x}}$" + ], + "image_footnote": [], + "bbox": [ + 558, + 148, + 671, + 239 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/129be1559b6919f978fa7536daf88bb212d0b0aed5e525eaf03108b6cf71556d.jpg", + "image_caption": [ + "e Synthetic Rd on $\\mathbf{b}_0$" + ], + "image_footnote": [], + "bbox": [ + 671, + 152, + 782, + 239 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "are more complex than daytime raindrops. Meanwhile, existing state-of-the-art methods are designed for daytime raindrop removal and neglect the effect of complex nighttime conditions. Thus, existing state-of-the-art methods cannot achieve promising results.", + "bbox": [ + 212, + 321, + 784, + 383 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 409, + 357, + 425 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we propose a new raindrop dataset, named Raindrop Clarity. It includes a total of 15,186 high-quality pairs/triplets of images with raindrops and the corresponding clear background images. Moreover, our dataset is collected under various conditions, including different scenes, times (daytime and nighttime), and focuses (raindrop- and background-focused). To be specific, 3,606 daytime raindrop-focused, 1,836 daytime background-focused, 4,838 nighttime raindrop-focused and 4,906 nighttime background-focused images are collected. With the proposed dataset, we re-analyze existing state-of-the-art methods, including raindrop algorithms and restoration backbones, and point out several open problems: raindrop-focus raindrop removal and nighttime raindrop removal. We wish to draw the attention of the community to the open problems.", + "bbox": [ + 212, + 445, + 787, + 612 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Social Impacts", + "text_level": 1, + "bbox": [ + 215, + 638, + 393, + 655 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Existing raindrop datasets focus on daytime background-focused raindrop removal. However, in the real world, raindrop images may focus on raindrops, resulting in a blurry background. Also, raindrops may occur in nighttime images. To draw the attention of the community to the open problems, we propose a new raindrop dataset, named Raindrop Clarity. It mainly collected from four conditions: daytime raindrop-focused, daytime background-focused, nighttime raindrop-focused and nighttime background-focused. Experimental results on our proposed dataset show that existing state-of-the-art methods cannot effectively remove raindrops under complex conditions. As a result, we believe that further exploration of our Raindrop Clarity dataset can help the community to address more complex raindrops.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Jin, X. Li et al.", + "bbox": [ + 271, + 114, + 395, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 145, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This research is supported by the National Research Foundation, Singapore, under its AI Singapore Programme (AISG Award No: AISG2-PhD/2022-01-037[T]). Xin Li's research is supported by NSFC under Grant 623B2098. Malu Zhang's work is supported by the National Science Foundation of China under Grant 62106038, and in part by the Sichuan Science and Technology Program under Grant 2023YFG0259. We thank Robby T. Tan and Beibei Lin for their valuable discussions, revisions, and feedback.", + "bbox": [ + 212, + 174, + 787, + 282 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 301, + 321, + 318 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. [EB/OL], http://liuyunfei.net/Projs/PBRR/, title = PBRR: Physically Based Raindrop Rendering, author = Yunfei Liu, Zhixiang Hao, Shadi You, Yu Li, and Feng Lu", + "2. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12299-12310 (2021)", + "3. Chen, S., Ye, T., Bai, J., Chen, E., Shi, J., Zhu, L.: Sparse sampling transformer with uncertainty-driven ranking for unified removal of raindrops and rain streaks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13106-13117 (2023)", + "4. Chen, S., Ye, T., Liu, Y., Chen, E.: Snowformer: Context interaction transformer with scale-awareness for single image desnowing. arXiv preprint arXiv:2208.09703 (2022)", + "5. Chen, S., Ye, T., Liu, Y., Liao, T., Jiang, J., Chen, E., Chen, P.: Msp-former: Multi-scale projection transformer for single image desnowing. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023)", + "6. Eigen, D., Krishnan, D., Fergus, R.: Restoring an image taken through a window covered with dirt or rain. In: Proceedings of the IEEE international conference on computer vision. pp. 633-640 (2013)", + "7. Halimeh, J.C., Roser, M.: Raindrop detection on car windshields using geometric-photometric environment construction and intensity-based correlation. In: 2009 IEEE Intelligent Vehicles Symposium. pp. 610-615. IEEE (2009)", + "8. Hao, Z., You, S., Li, Y., Li, K., Lu, F.: Learning from synthetic photorealistic raindrop for single image raindrop removal. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. pp. 0-0 (2019)", + "9. Jin, Y., Lin, B., Yan, W., Yuan, Y., Ye, W., Tan, R.T.: Enhancing visibility in nighttime haze images using guided apsf and gradient adaptive convolution. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 2446-2457 (2023)", + "10. Jin, Y., Sharma, A., Tan, R.T.: Dc-shadownet: Single-image hard and soft shadow removal using unsupervised domain-classifier guided network. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5027-5036 (2021)", + "11. Jin, Y., Yang, W., Tan, R.T.: Unsupervised night image enhancement: When layer decomposition meets light-effects suppression. In: European Conference on Computer Vision. pp. 404-421. Springer (2022)" + ], + "bbox": [ + 218, + 330, + 784, + 838 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Raindrop Clarity: Dual-Focused Dataset for Day Night Raindrop Removal", + "bbox": [ + 236, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "12. Jin, Y., Ye, W., Yang, W., Yuan, Y., Tan, R.T.: Des3: Adaptive attention-driven self and soft shadow removal using vit similarity. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 2634-2642 (2024)", + "13. Li, B., Li, X., Zhu, H., Jin, Y., Feng, R., Zhang, Z., Chen, Z.: Sed: Semantic-aware discriminator for image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024)", + "14. Li, X., Jin, X., Lin, J., Liu, S., Wu, Y., Yu, T., Zhou, W., Chen, Z.: Learning disentangled feature representation for hybrid-distorted image restoration. In: ECCV. pp. 313-329. Springer (2020)", + "15. Li, X., Li, B., Jin, X., Lan, C., Chen, Z.: Learning distortion invariant representation for image restoration from a causality perspective. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1714-1724 (2023)", + "16. Li, X., Ren, Y., Jin, X., Lan, C., Wang, X., Zeng, W., Wang, X., Chen, Z.: Diffusion models for image restoration and enhancement-a comprehensive survey. arXiv preprint arXiv:2308.09388 (2023)", + "17. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. arXiv preprint arXiv:2108.10257 (2021)", + "18. Liu, X., Suganuma, M., Sun, Z., Okatani, T.: Dual residual networks leveraging the potential of paired operations for image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7007-7016 (2019)", + "19. Luo, J., Ceylan, D., Yoon, J.S., Zhao, N., Philip, J., Frühstück, A., Li, W., Richardt, C., Wang, T.: Intrinsic diffusion: Joint intrinsic layers from latent diffusion models. In: ACM SIGGRAPH 2024 Conference Papers. SIGGRAPH '24, Association for Computing Machinery, New York, NY, USA (2024). https://doi.org/10.1145/3641519.3657472, https://doi.org/10.1145/3641519.3657472", + "20. Luo, J., Zhao, N., Li, W., Richardt, C.: Crefnet: Learning consistent reflectance estimation with a decoder-sharing transformer. IEEE Transactions on Visualization and Computer Graphics (2023)", + "21. Ozdenizci, O., Legenstein, R.: Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)", + "22. Peebles, W., Xie, S.: Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748 (2022)", + "23. Pizzati, F., Cerri, P., de Charette, R.: Physics-informed guided disentanglement in generative networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)", + "24. Porav, H., Bruls, T., Newman, P.: I can see clearly now: Image restoration via deraining. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 7087-7093. IEEE (2019)", + "25. Qian, R., Tan, R.T., Yang, W., Su, J., Liu, J.: Attentive generative adversarial network for raindrop removal from a single image pp. 2482-2491 (2018)", + "26. Quan, R., Yu, X., Liang, Y., Yang, Y.: Removing raindrops and rain streaks in one go. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9147-9156 (2021)", + "27. Quan, Y., Deng, S., Chen, Y., Ji, H.: Deep learning for seeing through window with raindrops. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2463-2471 (2019)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Jin, X. Li et al.", + "bbox": [ + 271, + 114, + 395, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "28. Roser, M., Geiger, A.: Video-based raindrop detection for improved image registration. In: 2009 IEEE 12th International Conference on Computer Vision Workshops, ICCV Workshops. pp. 570-577. IEEE (2009)", + "29. Roser, M., Kurz, J., Geiger, A.: Realistic modeling of water droplets for monocular adherent raindrop recognition usingBezier curves. In: Asian conference on computer vision. pp. 235-244. Springer (2010)", + "30. Soboleva, V., Shipitko, O.: Raindrops on windshield: Dataset and lightweight gradient-based detection algorithm. In: 2021 IEEE Symposium Series on Computational Intelligence (SSCI). pp. 1-7. IEEE (2021)", + "31. Wang, C., Pan, J., Lin, W., Dong, J., Wang, W., Wu, X.M.: Selfpromer: Self-prompt dehazing transformers with depth-consistency. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 5327-5335 (2024)", + "32. Wang, C., Pan, J., Wang, W., Dong, J., Wang, M., Ju, Y., Chen, J.: Promptstorer: A prompting image restoration method with degradation perception. Advances in Neural Information Processing Systems 36, 8898-8912 (2023)", + "33. Wang, C., Pan, J., Wang, W., Fu, G., Liang, S., Wang, M., Wu, X.M., Liu, J.: Correlation matching transformation transformers for uhd image restoration. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 5336-5344 (2024)", + "34. Wang, J., Pan, Z., Zhang, M., Tan, R.T., Li, H.: Restoring speaking lips from occlusion for audio-visual speech recognition. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 19144-19152 (2024)", + "35. Wang, J., Qian, X., Zhang, M., Tan, R.T., Li, H.: Seeing what you said: Talking face generation guided by a lip reading expert. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14653-14662 (2023)", + "36. Wang, Z., Cun, X., Bao, J., Zhou, W., Liu, J., Li, H.: Uformer: A general u-shaped transformer for image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17683-17693 (2022)", + "37. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)", + "38. Wei, M., Shen, Y., Wang, Y., Xie, H., Wang, F.L.: Raindiffusion: When unsupervised learning meets diffusion models for real-world image deraining. arXiv preprint arXiv:2301.09430 (2023)", + "39. Wen, Q., Wu, Y., Chen, Q.: Video waterfall removal via spatio-temporal fusion in driving scenes. arXiv preprint arXiv:2302.05916 (2023)", + "40. Xiao, J., Fu, X., Liu, A., Wu, F., Zha, Z.J.: Image de-raining transformer. IEEE Transactions on Pattern Analysis and Machine Intelligence pp. 1-18 (2022)", + "41. Yasarla, R., Patel, V.M.: Uncertainty guided multi-scale residual learning-using a cycle spinning cnn for single image de-raining. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8405-8414 (2019)", + "42. Ye, T., Chen, S., Bai, J., Shi, J., Xue, C., Jiang, J., Yin, J., Chen, E., Liu, Y.: Adverse weather removal with codebook priors. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12653-12664 (2023)", + "43. Ye, T., Zhang, Y., Jiang, M., Chen, L., Liu, Y., Chen, S., Chen, E.: Perceiving and modeling density for image dehazing. In: European conference on computer vision. pp. 130-145. Springer (2022)", + "44. You, S., Tan, R.T., Kawakami, R., Ikeuchi, K.: Adherent raindrop detection and removal in video. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1035-1042 (2013)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Raindrop Clarity: Dual-Focused Dataset for Day Night Raindrop Removal", + "bbox": [ + 236, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "45. You, S., Tan, R.T., Kawakami, R., Mukaigawa, Y., Ikeuchi, K.: Adherent raindrop modeling, detection and removal in video. IEEE transactions on pattern analysis and machine intelligence 38(9), 1721-1733 (2015)", + "46. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR (2022)", + "47. Zhang, K., Li, D., Luo, W., Ren, W.: Dual attention-in-attention model for joint rain streak and raindrop removal. IEEE Transactions on Image Processing 30, 7608-7619 (2021)", + "48. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + ], + "bbox": [ + 215, + 146, + 784, + 286 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y. Jin, X. Li et al.", + "bbox": [ + 271, + 114, + 395, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Raindrop Clarity_ A Dual-Focused Dataset for Day and Night Raindrop Removal/64be914f-41c4-4c4e-9cfc-b492879688db_model.json b/2024/Raindrop Clarity_ A Dual-Focused Dataset for Day and Night Raindrop Removal/64be914f-41c4-4c4e-9cfc-b492879688db_model.json new file mode 100644 index 0000000000000000000000000000000000000000..80f81425b8414f2467aff964def86873d9e61981 --- /dev/null +++ b/2024/Raindrop Clarity_ A Dual-Focused Dataset for Day and Night Raindrop Removal/64be914f-41c4-4c4e-9cfc-b492879688db_model.json @@ -0,0 +1,2392 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.237, + 0.141, + 0.768, + 0.187 + ], + "angle": 0, + "content": "Raindrop Clarity: A Dual-Focused Dataset for Day and Night Raindrop Removal" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.212, + 0.782, + 0.229 + ], + "angle": 0, + "content": "Yeying Jin\\(^{1}\\), Xin Li\\(^{2}\\), Jiadong Wang\\(^{1}\\), Yan Zhang\\(^{1}\\), and Malu Zhang\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.383, + 0.239, + 0.619, + 0.254 + ], + "angle": 0, + "content": "\\(^{1}\\) National University of Singapore" + }, + { + "type": "text", + "bbox": [ + 0.339, + 0.254, + 0.664, + 0.268 + ], + "angle": 0, + "content": "2 University of Science and Technology of China" + }, + { + "type": "text", + "bbox": [ + 0.257, + 0.268, + 0.745, + 0.297 + ], + "angle": 0, + "content": "3 University of Electronic Science and Technology of China jinyeying@u.nus.edu, xin.li@ustc.edu.cn, maluzhang@uestc.edu.cn" + }, + { + "type": "list", + "bbox": [ + 0.257, + 0.239, + 0.745, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.327, + 0.756, + 0.604 + ], + "angle": 0, + "content": "Abstract. Existing raindrop removal datasets have two shortcomings. First, they consist of images captured by cameras with a focus on the background, leading to the presence of blurry raindrops. To our knowledge, none of these datasets include images where the focus is specifically on raindrops, which results in a blurry background. Second, these datasets predominantly consist of daytime images, thereby lacking nighttime raindrop scenarios. Consequently, algorithms trained on these datasets may struggle to perform effectively in raindrop-focused or nighttime scenarios. The absence of datasets specifically designed for raindrop-focused and nighttime raindrops constrains research in this area. In this paper, we introduce a large-scale, real-world raindrop removal dataset called Raindrop Clarity. Raindrop Clarity comprises 15,186 high-quality pairs/triplets (raindrops, blur, and background) of images with raindrops and the corresponding clear background images. There are 5,442 daytime raindrop images and 9,744 nighttime raindrop images. Specifically, the 5,442 daytime images include 3,606 raindrop- and 1,836 background-focused images. While the 9,744 nighttime images contain 4,838 raindrop- and 4,906 background-focused images. Our dataset will enable the community to explore background-focused and raindrop-focused images, including challenges unique to daytime and nighttime conditions." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.617, + 0.698, + 0.632 + ], + "angle": 0, + "content": "Keywords: Raindrop-focused \\(\\cdot\\) Background-focused \\(\\cdot\\) Nighttime" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.654, + 0.377, + 0.67 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.684, + 0.788, + 0.791 + ], + "angle": 0, + "content": "Adherent raindrops on lenses or windscreens can significantly reduce visibility. Raindrop removal is crucial for surveillance, self-driving cars, object detection, outdoor photography, filmmaking, augmented reality, etc. Recently, a few rain-drop datasets (e.g., [24-26, 30]) have been proposed. In Fig. 1, these datasets primarily contain daytime background scenes with blurry raindrops, as the images are taken with the camera's focus on the background. They have impacts in the raindrop removal area, inspiring several successful methods [24, 25, 27]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.798, + 0.384, + 0.812 + ], + "angle": 0, + "content": "* Corresponding author" + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.812, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Our data and code are available at: https://github.com/jinyeying/RaindropClarity" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.798, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.128 + ], + "angle": 0, + "content": "Y. Jin, X. Li et al." + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.145, + 0.413, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.341, + 0.262, + 0.378, + 0.272 + ], + "angle": 0, + "content": "(a) Qian" + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.147, + 0.523, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.431, + 0.262, + 0.507, + 0.272 + ], + "angle": 0, + "content": "(b) RainDS Quan" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.147, + 0.605, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.204, + 0.605, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.262, + 0.606, + 0.271 + ], + "angle": 0, + "content": "(c) RobotCar Porav" + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.147, + 0.698, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.62, + 0.262, + 0.685, + 0.272 + ], + "angle": 0, + "content": "(d) Windshield" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.284, + 0.787, + 0.327 + ], + "angle": 0, + "content": "Fig. 1. Existing raindrop datasets (e.g., [24-26,30]) exhibit two limitations: first, they do not include raindrop-focused images (Fig. 2 right); and second, they lack night raindrops (Fig. 2 bottom)." + }, + { + "type": "image", + "bbox": [ + 0.294, + 0.342, + 0.71, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.478, + 0.787, + 0.521 + ], + "angle": 0, + "content": "Fig. 2. Motivation for Raindrop Clarity: real images sourced from the Internet feature scenarios overlooked by existing datasets, including raindrop-focused (right) and nighttime raindrops (bottom). In this figure, \\(\\mathrm{Rd} =\\) Raindrop, \\(\\mathrm{Bg} =\\) Background." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.788, + 0.688 + ], + "angle": 0, + "content": "While these datasets are valuable, they have two limitations. First, a significant issue arises when the camera is in automatic focus mode. In Fig. 2 (right), the camera may unintentionally focus on raindrops that are attached to the lens, windshield, or glass, resulting in images with sharp raindrops and a blurry background. Consequently, methods designed for images with blurry raindrops and a sharp background may not perform optimally. Second, existing datasets consist only of daytime images. Therefore, existing methods, especially those reliant on training data, may struggle to effectively handle nighttime raindrop images because of the inherent domain gap between day and night conditions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.809 + ], + "angle": 0, + "content": "In Fig. 2, we observe that besides daytime images with blurry raindrops and a sharp background, it is common to come across images with sharp raindrops and a blurry background (right), as well as nighttime raindrops (bottom). Thus, we collected varied raindrop images as shown in Fig. 3: the top-left with daytime blurry raindrops against a sharp background, the top-right with daytime sharp raindrops and a blurry background, the bottom-left with nighttime blurry raindrops and a clear background, and the bottom-right with nighttime sharp raindrops with a blurry background." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.811, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Physical solutions like rain covers or wipers can reduce the presence of raindrops, but their effectiveness is limited. In certain situations where wipers cannot" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.237, + 0.114, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Raindrop Clarity: Dual-Focused Dataset for Day Night Raindrop Removal" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image_caption", + "bbox": [ + 0.284, + 0.146, + 0.477, + 0.159 + ], + "angle": 0, + "content": "Background-focused (Blurry raindrop)" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.159, + 0.508, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.147, + 0.744, + 0.159 + ], + "angle": 0, + "content": "Raindrop-focused (Blurry background)" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.159, + 0.783, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.278, + 0.788, + 0.349 + ], + "angle": 0, + "content": "Fig. 3. Raindrop Clarity: Examples of our pairs \\((\\tilde{\\mathbf{x}}, b_0)\\) and triplets \\((\\tilde{\\mathbf{x}}, x_0, b_0)\\) of a raindrop image \\(\\tilde{\\mathbf{x}}\\), the corresponding blurry background image \\(x_0\\) and the corresponding clear background image \\(b_0\\), in day and night. For the pairs \\((\\tilde{\\mathbf{x}}, b_0)\\), there is no \\(x_0\\), since in this case \\(x_0 = b_0\\). The reason we have the blurry background image \\(x_0\\) in triplets is to have the raindrop difference maps \\(\\tilde{\\mathbf{m}}\\) (see Fig. 8)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.788, + 0.563 + ], + "angle": 0, + "content": "be used, or wind blows rain onto the lens, these physical measures become impractical. Physics-based methods [1,8,23,44,45] remove raindrops by leveraging the physics of raindrops or utilizing generated synthetic data. However, the effectiveness of methods trained on synthetic data relies on the quality of raindrop rendering models. Currently, raindrop rendering models commonly employ surface modeling and ray tracing [8,28,29]. Unfortunately, these methods may not fully consider factors such as scene depth, lighting conditions, and the 3D shapes of raindrops [23,28], which are crucial for realistic raindrop rendering. Additionally, raindrops can introduce refraction effects [7,44], distorting the shapes and positions of objects, which further complicates accurate raindrop rendering. To the best of our knowledge, developing a rendering approach that can produce physically accurate raindrops adaptable to various scenes is challenging." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.788, + 0.777 + ], + "angle": 0, + "content": "Motivated by the aforementioned challenges and the observation that existing raindrop datasets predominantly consist of background-focused images while neglecting raindrop-focused images and nighttime raindrops, we collect a novel dataset, called \"Raindrop Clarity\". Our dataset comprises 6,742 (1,836 daytime, 4,906 nighttime) pairs and 8,444 (3,606 daytime, 4,838 nighttime) triplets of raindrop-focused images and background-focused images. All of them have the corresponding clear background images. The triplets include one more raindrop-free blurry background images. Fig. 3 shows samples and explanation on the meaning of our pairs and triplets. From the 15,186 raindrops pairs/triplet, 5,442 are captured during daytime and the 9,744 at nighttime. Raindrops Clarity exhibits diverse sizes and shapes, and raindrop appearance varies based on the scene and lighting conditions. To our knowledge, Raindrop Clarity is the largest real-world raindrop dataset that contains both the pair and triplets, day and night. Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "- We present the Raindrop Clarity dataset, a large-scale real-world dataset that encompasses both raindrop-focused triplets and background-focused pairs. For the pairs, we provide images with blurry raindrops and their cor" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.128 + ], + "angle": 0, + "content": "Y. Jin, X. Li et al." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.147, + 0.785, + 0.177 + ], + "angle": 0, + "content": "responding clear backgrounds. For the triplets, we additionally provide the raindrop-free blurry background." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.178, + 0.784, + 0.221 + ], + "angle": 0, + "content": "- Raindrop Clarity includes images collected under both daytime and nighttime. It aims to remove raindrops and recover clear backgrounds regardless of day or night, whether the raindrops or background are blurry." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.223, + 0.784, + 0.281 + ], + "angle": 0, + "content": "- Our experimental results indicate that even with the application of state-of-the-art raindrop removal methods, there are instances where these methods fail. This suggests the existence of unresolved challenges in the field of raindrop removal, which await further attention and investigation." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.178, + 0.784, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.308, + 0.388, + 0.324 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.785, + 0.506 + ], + "angle": 0, + "content": "Raindrop Dataset. Qian et al. [25] introduce a raindrop dataset consisting of raindrop and corresponding ground truths. Porav et al. [24] propose a stereo RobotCar dataset, using a rain-making device that causes one lens to be affected by raindrops, while the other lens remains clear. Hao et al. [8] create a synthetic dataset of adherent raindrops with pixel-level masks for training deraining algorithms. Quan et al. [26] introduce RainDS, a dataset that includes raindrops, rain steaks, and their corresponding ground truths. Soboleva and Shipitko [30] introduce Windshields, a dataset containing raindrop images on both camera lenses and windshields, annotated with binary masks representing the raindrop areas. However, all of these datasets primarily consist of daytime images taken by cameras focusing on the background scenes and thus have blurry raindrops." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.785, + 0.793 + ], + "angle": 0, + "content": "Single Image Raindrop Removal. Most image deraining methods have primarily focused on removing rain streaks rather than raindrops. However, recent learning-based methods have been proposed specifically for single-image raindrop removal. Eigen et al. [6] introduce a supervised CNN-based method to remove raindrops, using a shallow CNN. Qian et al. [25] propose an AttentiveGAN for raindrop removal, that relies on attention maps but has a limited focus on local spatial information. Quan et al. [27] propose RaindropAttention, which employs an attention mechanism to generate attention maps based on mathematical raindrop shapes, yet it lacks exploration of long-range information. Liu et al. [18] demonstrate a dual residual network, by leveraging the complementary nature of paired residual blocks. Hao et al. [8] propose a raindrop synthesis algorithm, used their synthetic dataset to train a multi-task network for raindrop detection and removal. Porav et al. [24] propose a Pix2PixHD-based deraining network architecture. Subsequent work by Quan et al. [26] propose a cascaded network architecture for simultaneous removal of rain streaks and raindrops, achieved through a neural architecture search. Yasarla and Patel [41] propose the uncertainty guided multi-scale residual learning network to effectively remove rain from images. Zhang et al. [47] propose the Dual Attention-in-Attention Model to remove rain streaks and raindrops." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Restoration Backbone. There are image restoration and enhancement tasks that utilize transformer-based models, taking advantage of the long-range dependencies, including IPT [2] (vanilla transformer), Uformer [36], SwinIR [17]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.237, + 0.114, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Raindrop Clarity: Dual-Focused Dataset for Day Night Raindrop Removal" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image_caption", + "bbox": [ + 0.388, + 0.145, + 0.623, + 0.159 + ], + "angle": 0, + "content": "Background-focused (Blurry raindrop)" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.159, + 0.7, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.318, + 0.706, + 0.333 + ], + "angle": 0, + "content": "Fig. 4. Pairs \\((\\tilde{\\mathbf{x}},\\pmb{b}_0)\\) show various sizes, shapes, and densities." + }, + { + "type": "image_caption", + "bbox": [ + 0.388, + 0.396, + 0.627, + 0.409 + ], + "angle": 0, + "content": "Raindrop-focused (Blurry background)" + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.409, + 0.7, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.29, + 0.644, + 0.711, + 0.659 + ], + "angle": 0, + "content": "Fig. 5. The triplets \\((\\tilde{\\mathbf{x}}, \\mathbf{x}_0, \\mathbf{b}_0)\\) cover diverse scenes and shapes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.84 + ], + "angle": 0, + "content": "and Restormer [46]. Xiao et al. [40] propose an image deraining transformer (IDT) for rain streak and raindrop removal with window-based and spatial-based dual Transformer. RaindropDiffusion [21] develops a UNet-based diffusion framework, which demonstrates robust generation capabilities for raindrops, rain streaks, and snow. RainDiffusion [38], an unsupervised diffusion model, utilizes unpaired real-world data instead of weakly adversarial training. DiT [22] is a transformer-based diffusion model." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.128 + ], + "angle": 0, + "content": "Y. Jin, X. Li et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.499, + 0.163 + ], + "angle": 0, + "content": "3 Raindrop Clarity Dataset" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.178, + 0.393, + 0.192 + ], + "angle": 0, + "content": "3.1 Data Collection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.205, + 0.787, + 0.279 + ], + "angle": 0, + "content": "To create Raindrop Clarity, we employed a sphere pan-tilt platform to keep our camera stationary. The images were captured using Sony FDR-AX33 4K Ultra HD Handycam Camcorder, Sony alpha 7R III digital camera, the back camera of an iPhone 14 Pro and iPhone 15 Pro Max, ensuring the highest quality possible. The 4K Camcorder captured RGB images at a resolution \\(3840 \\times 2164\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.281, + 0.787, + 0.326 + ], + "angle": 0, + "content": "Using the optical refraction model (raindrops appear upside-down), we projected raindrops by spraying water on glass and adjusting the distance, angle, speed, and patterns of the nozzle/wipers. We also captured in rain conditions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.326, + 0.789, + 0.523 + ], + "angle": 0, + "content": "Our collection procedure was designed to ensure diversity in the raindrop (such as focuses, shapes and sizes), background scenes and lighting conditions. First, we fixed the camera position on a tripod and installed a glass plate. The distance between the camera and the glass plate ranged from \\(5\\mathrm{cm}\\) to \\(25\\mathrm{cm}\\). Second, we sprayed waterdrops/raindrops onto the glass plate and focused the camera on the raindrops. Third, we sprayed the waterdrops/raindrops to capture images of them under different shapes and sizes, ranging from sparse to dense. The captured images include elliptical raindrops and raindrop flow traces \\((\\tilde{\\mathbf{x}})\\). Fourth, we removed the glass plate, allowing the camera to focus on the very close plane. The duration is kept very short, within milliseconds, to minimize lighting changes. In this way, we obtained raindrop-free blurry background images \\((\\boldsymbol{x}_0)\\). Finally, we adjusted the camera focus to the background, directing the camera to focus on the distant background \\((\\boldsymbol{b}_0)\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.545, + 0.553, + 0.561 + ], + "angle": 0, + "content": "3.2 Characteristics of Raindrop Clarity" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.572, + 0.787, + 0.663 + ], + "angle": 0, + "content": "Raindrop Clarity is a new dataset consisting of around 15,186 high-quality image raindrop pairs and triplets. In Fig. 3, our dataset introduces a unique element of triplet data, which includes raindrop-focused images, corresponding blurry raindrop-free background images, and their clear background equivalents, distinguishing it from existing datasets. The second feature that distinguishes the Raindrop Clarity from the existing ones is the collection of night raindrops." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.676, + 0.787, + 0.752 + ], + "angle": 0, + "content": "Objectives. Our dataset is designed to support research in the elimination of raindrops and the restoration of a clear background, irrespective of whether the camera's focus is on the raindrops (leading to a blurry background) or on the background (resulting in blurry raindrops). This holds true regardless of whether it is daytime or nighttime." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Raindrops and Diversity. Our Raindrop Clarity dataset offers an extensive coverage of diverse raindrop sizes, shapes, and occlusion types in various scenes and lighting conditions, making it an exceptionally comprehensive resource for advancing raindrop removal research. In Fig. 4 and Fig. 5, the dataset captures raindrops and rain flows of different sizes and shapes, including heavy raindrops" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.237, + 0.114, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Raindrop Clarity: Dual-Focused Dataset for Day Night Raindrop Removal" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "title", + "bbox": [ + 0.388, + 0.145, + 0.623, + 0.158 + ], + "angle": 0, + "content": "Background-focused (Blurry raindrop)" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.158, + 0.7, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.302, + 0.317, + 0.7, + 0.332 + ], + "angle": 0, + "content": "Fig. 6. The pairs \\((\\tilde{\\mathbf{x}}, b_0)\\) under different lighting conditions." + }, + { + "type": "title", + "bbox": [ + 0.388, + 0.345, + 0.626, + 0.359 + ], + "angle": 0, + "content": "Raindrop-focused (Blurry background)" + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.359, + 0.699, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.593, + 0.706, + 0.607 + ], + "angle": 0, + "content": "Fig. 7. Triplets \\((\\tilde{\\mathbf{x}},\\pmb {x}_0,\\pmb {b}_0)\\) under different lighting conditions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.636, + 0.788, + 0.698 + ], + "angle": 0, + "content": "that result in unpredictable water flows and blurry backgrounds when they hit window panes or car windshields. Two primary types of occlusions caused by raindrops, namely partial and complete occlusions, are included in our dataset. Both types present unique challenges for raindrop removal algorithms." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.708, + 0.788, + 0.8 + ], + "angle": 0, + "content": "Scenes and Large-scale. The dataset encompasses a wide range of background scenes, such as urban, countryside, campus, wilderness, road, and aerial views, covering both day and night. This variation allows for the exploration of different lighting and weather conditions in different scenes. With approximately 38,816 images in total, including 15,186 high-quality pairs/triplets of images, the Raindrop Clarity dataset is one of the largest datasets of its kind." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Lighting Conditions. The appearance of raindrops highly depends on the surrounding environment. A distinctive feature of our dataset is its comprehen" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.128 + ], + "angle": 0, + "content": "Y. Jin, X. Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 1. Summary of comparisons between our Raindrop Clarity and existing datasets. Bg-focus = Background-focused, Rd-focus = Raindrop-focused, RD = Raindrop. Raindrop Clarity uniquely features raindrop-focused triplets and the nighttime raindrops." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.199, + 0.789, + 0.324 + ], + "angle": 0, + "content": "
DatasetsComponentsN. ImagesSyn/RealBg-focusRd-focus
DayNightDayNight
Our RD Clarity (Day)3,606 Triplets+1,836 Pairs14,490Real✓×✓×
Our RD Clarity (Night)4,838 Triplets+4,906 Pairs24,326Real×✓×✓
Raindrop Qian [25]919 Pairs1,838Real✓×××
RainDS-Real [26]248 Quadruplets992Real✓×××
RainDS-Syn [26]1,200 Quadruplets4,800Syn✓×××
RobotCar Porav [24]4,818 Pairs9,636Real✓×××
Windshield [30]3,3903,390Real✓×××
" + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.335, + 0.703, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.503, + 0.787, + 0.531 + ], + "angle": 0, + "content": "Fig. 8. We show \\((\\tilde{\\mathbf{x}},\\pmb{x}_0,\\tilde{\\mathbf{m}},\\pmb{b}_0)\\), difference map \\(\\tilde{\\mathbf{m}}\\) is derived by distinguishing raindrop-focused \\(\\tilde{\\mathbf{x}}\\) from blurry background \\(\\pmb{x}_0\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.559, + 0.785, + 0.604 + ], + "angle": 0, + "content": "sive coverage of various environmental conditions, setting it apart from other datasets. The dataset's variation in day and night conditions makes it valuable for developing raindrop removal algorithms." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.605, + 0.785, + 0.693 + ], + "angle": 0, + "content": "During the daytime, raindrop images exhibit a diverse array of appearances influenced by multiple factors such as sunlight direction, intensity, and shadowing. In Figs. 4-5, our dataset captures this variety, with images taken at different times of the day and under varying weather conditions (such as light rain, heavy rain, and rain combined with wind), providing rich information for algorithms to learn." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.695, + 0.787, + 0.801 + ], + "angle": 0, + "content": "Unlike daytime, nighttime raindrops present unique challenges due to artificial lighting conditions and low light. The appearance of raindrops can drastically change under street lights, headlights, or neon lights. In Fig. 6 and Fig. 7, our dataset includes night images with varied artificial lights, contributing to the complexity and comprehensiveness of the dataset. These night images are challenging due to the higher contrast between raindrops and the background, and potential overexposure or underexposure issues." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.813, + 0.842 + ], + "angle": 0, + "content": "Background- and Raindrop-focused. When the raindrop image \\(\\tilde{\\mathbf{x}}\\) is background-focused, we decompose a raindrop degraded image \\(\\tilde{\\mathbf{x}}\\) into a clear background" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.237, + 0.114, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Raindrop Clarity: Dual-Focused Dataset for Day Night Raindrop Removal" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.462, + 0.788, + 0.49 + ], + "angle": 0, + "content": "Fig. 9. Comparison of existing methods on daytime raindrop-focused images, covering both raindrop algorithms and restoration backbones." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.657 + ], + "angle": 0, + "content": "image \\(\\pmb{b}_0\\) by following this model [25]: \\(\\tilde{\\mathbf{x}} = (1 - \\mathbf{M})\\odot \\pmb{x}_0 + \\mathbf{D}\\), where \\(\\mathbf{D}\\) is the effect of the blurry raindrops, referring to the localized blurry imagery formed by the light reflected from the surroundings. The operator \\(\\odot\\) denotes element-wise multiplication. When binary mask \\(\\mathbf{M}(\\mathbf{x})\\) equals 1, it signifies that the pixel \\(x\\) belongs to the area affected by raindrops. Conversely, if \\(\\mathbf{M}(\\mathbf{x})\\) does not equal 1, it indicates that the pixel is within the background areas. In the case where \\(\\tilde{\\mathbf{x}}\\) is background-focused raindrop images, \\(\\pmb{x}_0\\) is identical to the clear, raindrop-free image \\(\\pmb{b}_0\\), namely \\(\\pmb{x}_0 = \\pmb{b}_0\\). Therefore, we use \\((\\tilde{\\mathbf{x}},\\pmb{b}_0)\\) to stand for background-focused (resulting in blurry raindrops) pairs, shown in the left part of Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.788, + 0.749 + ], + "angle": 0, + "content": "In the right part of Fig. 3 and Fig. 8, when \\(\\tilde{\\mathbf{x}}\\) is the raindrop-focused image, \\(\\pmb{x}_0\\) is the corresponding raindrop-free blurry background image, while \\(\\pmb{b}_0\\) is the clear background image. In this case, \\(\\pmb{x}_0 \\neq \\pmb{b}_0\\). We can calculate the pixel-level difference map \\(\\tilde{\\mathbf{m}}\\) between the raindrop-free blurry background \\(\\pmb{x}_0\\) and the raindrop-focused \\(\\tilde{\\mathbf{x}}\\) as: \\(\\tilde{\\mathbf{m}} = \\tilde{\\mathbf{x}} - \\pmb{x}_0\\). Therefore, we use \\((\\tilde{\\mathbf{x}}, \\pmb{x}_0, \\pmb{b}_0)\\) to stand for raindrop-focused (resulting in blurry background) triplets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Since the camera is stationary, neither the pairs \\((\\tilde{\\mathbf{x}},\\boldsymbol{b}_0)\\) nor the triplets \\((\\tilde{\\mathbf{x}},\\boldsymbol{x}_0,\\boldsymbol{b}_0)\\) have alignment problems. This is also evident in the raindrop difference maps \\(\\tilde{\\mathbf{m}}\\), which accurately assign values to raindrop regions across multiple images. Specifically, for each pixel \\(\\mathbf{x}\\) in the input image, any discrepancies in the map's assigned values for the static background would indicate alignment issues; however, these are not observed, as \\(\\tilde{\\mathbf{m}}\\) indicates only the raindrop regions." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.396, + 0.128 + ], + "angle": 0, + "content": "Y. Jin, X. Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.173 + ], + "angle": 0, + "content": "Table 2. Quantitative evaluation of daytime and nighttime raindrop data with a benchmark for raindrop algorithms and restoration backbones." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.185, + 0.788, + 0.312 + ], + "angle": 0, + "content": "
Data\\Method InputRaindrop AlgorithmsRestoration Backbones
AtGAN [25]Robot [24]RainDS [26]IDT [40]RDdiff [21]Restor. [46]Uform. [36]DiT [22]
DayPSNR↑ 21.9223.6223.8225.3926.0525.5226.0825.7826.03
SSIM↑ 0.5600.6580.6700.7170.7360.7340.7480.7270.752
LPIPS↓ 0.2470.2000.1910.1470.1410.1110.1310.1480.106
NightPSNR↑ 24.7824.3824.5725.3926.8126.4826.8725.2626.23
SSIM↑ 0.7260.7730.7870.8200.8510.8310.8480.8160.826
LPIPS↓ 0.2090.1850.1820.1390.1250.1120.1230.1420.111
" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.344, + 0.787, + 0.652 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.661, + 0.784, + 0.689 + ], + "angle": 0, + "content": "Fig. 10. Comparison of methods on nighttime raindrop-focused images, covering both raindrop algorithms and restoration backbones." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.74, + 0.784, + 0.77 + ], + "angle": 0, + "content": "Since the pairs/triplets of images are captured in quick succession, there are no problems with changes in illumination." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.84 + ], + "angle": 0, + "content": "The captured images do not contain any personal information such as phone numbers, or human faces. We will release the dataset publicly with car license plates obscured, as we believe that our Raindrop Clarity dataset will benefit the progress of the raindrop removal field." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.237, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Raindrop Clarity: Dual-Focused Dataset for Day Night Raindrop Removal" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.388, + 0.785, + 0.417 + ], + "angle": 0, + "content": "Fig. 11. Comparison of methods on daytime background-focused images, covering both raindrop algorithms and restoration backbones." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.446, + 0.377, + 0.464 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.788, + 0.586 + ], + "angle": 0, + "content": "Experimental Setup. Our benchmark includes two parts: Daytime and Night-time Raindrop Removal. As shown in Table 1, our daytime raindrop dataset contains a total of 5,442 paired/triplets images, where 4,713 paired images are used for the training set, and the remaining 729 paired images are regarded as the test set. Our nighttime raindrop dataset consists of 9,744 paired/triplets images, where 8,655 paired images are selected as the training set and the rest 1,089 paired images are used for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.721 + ], + "angle": 0, + "content": "Baselines. To create the benchmark, we retrain existing state-of-the-art methods on our daytime and nighttime raindrop datasets. Our benchmark consists of five raindrop algorithms [21, 24-26, 40] and three restoration methods [22, 36, 46]. Specifically, IDT [40], RDiffusion [21], RainDS [26], Robot Car [24] and AtGAN [25] are raindrop algorithms, while Restormer [46], Uformer [36] and DiT [22] are restoration backbones. For a fair comparison, we conduct experiments using the original implementation of these methods. All experiments are implemented in PyTorch and are trained on four NVIDIA RTX A5000 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.787, + 0.763 + ], + "angle": 0, + "content": "Comparison with Existing Datasets. The difference between our Raindrop Clarity dataset and the existing datasets is shown in Table. 1 and Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.809 + ], + "angle": 0, + "content": "The stereo RobotCar dataset [24] includes 4,818 image pairs, randomly selected for training, validation, and testing. However, the raindrops in this dataset are disproportionately large and lack realism." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "The raindrop dataset proposed by Qian et al. [25] consists of 861 training pairs and 58 testing pairs. While it features common raindrop shapes, it lacks the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.396, + 0.128 + ], + "angle": 0, + "content": "Y. Jin, X. Li et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.535, + 0.788, + 0.564 + ], + "angle": 0, + "content": "Fig. 12. Comparison of methods on nighttime background-focused images, covering both raindrop algorithms and restoration backbones." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.787, + 0.626 + ], + "angle": 0, + "content": "diversity of real-world scenes [44]. Also, the backgrounds between the raindrop images and the ground truth are not perfectly aligned [39]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.79, + 0.702 + ], + "angle": 0, + "content": "The RainDS-Real dataset, proposed by Quan et al. [26], comprises 248 quadruplets, including raindrops, rain streaks, and their corresponding ground truths. The RainDS-Synthetic dataset has a domain gap with the real raindrop images. Moreover, the real-world raindrops are consistently small, obscuring only a fraction of the image details, making recovery less challenging." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.703, + 0.787, + 0.793 + ], + "angle": 0, + "content": "The Windshield dataset, proposed by Soboleva and Shipitko [30], was captured by a camera attached to the vehicle during its movement in urban areas and highway environments. The Windshield does not provide the corresponding clean pairs, making this dataset not available for robust raindrop removal method testing. Additionally, frequent moving objects further complicate the dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Existing raindrop datasets, though valuable, have certain limitations. They do not include raindrop-focused and nighttime raindrops. In addition, we have noticed reflection errors in the datasets from Qian [25] and RainDS-Real [26]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.237, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Raindrop Clarity: Dual-Focused Dataset for Day Night Raindrop Removal" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "Comparison with Synthetic Raindrops. In Fig. 13, we show the existing state-of-the-art synthetic physical model [23], although it can generate raindrops on blurry and clear backgrounds (Fig. 13 d generated on b, Fig. 13 e generated on c). It still has a gap with real-world collected raindrops. Real-world collected raindrops display diverse appearances, including various shapes such as round, elliptical, and long strips. Also, the transparency and glare effects of raindrops collected in the real-world varies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.267, + 0.787, + 0.312 + ], + "angle": 0, + "content": "Quantitative Comparison. To effectively evaluate the results on our proposed datasets, we use three popular metrics PSNR, SSIM [37], and LPIPS [48] in generation tasks [3-5, 9-16, 19, 20, 31-35, 42, 43]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.788, + 0.403 + ], + "angle": 0, + "content": "The qualitative results are shown in Table 2. For the daytime dataset, the PSNR and SSIM of the original input images are 21.92 and 0.560. Raindrop algorithm IDT [40] achieves the best performance (26.05 PSNR of and 0.736 of SSIM) compared with other raindrop algorithms. For the restoration backbones, DiT [22] achieves a PSNR of 26.03 and a SSIM of 0.752. Restormer [46] achieves a PSNR of 26.08 and a SSIM of 0.748, which also performs competitively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.788, + 0.495 + ], + "angle": 0, + "content": "It can be observed that, for the nighttime dataset, PSNR and SSIM of the input images are 24.78 and 0.726. Among the raindrop algorithms, IDT [40] achieves better performance than other methods, with a PSNR of 26.81 and a SSIM of 0.851. For the restoration backbones, Restormer [46] achieves the highest PSNR of 26.87, while DiT [22] achieves the lowest LPIPS of 0.111, which outperforms other backbones." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.598 + ], + "angle": 0, + "content": "Qualitative Comparison. Figs. 9-12 show the qualitative results of state-of-the-art methods. To be specific, Fig. 9 and Fig. 11 show the results of daytime raindrop removal. The input images in Fig. 9 and Fig. 11 are daytime raindrop-focused and background-focused, respectively. Fig. 10 and Fig. 12 show the results of nighttime raindrop-focused and background-focused removal, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.719 + ], + "angle": 0, + "content": "It can be found that, for daytime and nighttime raindrop-focused removal in Fig. 9 and Fig. 10, existing state-of-the-art methods fail to remove raindrops and recover the background. This is because most of them are designed for background-focused scenes and thus neglect the blurry problem. To our knowledge, Restormer [46] and Uformer [36] are specifically designed to address deblurring problem. However, they still struggle with recovering details and textures. As a result, these methods are struggling to simultaneously remove raindrops and recover the background." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In Fig. 11, for daytime background-focused inputs, we can observe that existing state-of-the-art methods can handle most raindrops. However, they are struggling to address long strips of raindrops. This is because their methods may not detect long strips of raindrops and thus fail to remove them. Furthermore, in Fig. 12, we find that existing state-of-the-art methods are struggling to handle nighttime background-focused samples. To be specific, raindrops in these samples cannot be totally removed and details of the restored images are not clear. This is because nighttime raindrops usually reflect artificial light and thus" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.396, + 0.128 + ], + "angle": 0, + "content": "Y. Jin, X. Li et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.146, + 0.331, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.242, + 0.326, + 0.251 + ], + "angle": 0, + "content": "a Raindrop-focusedd \\(\\tilde{\\mathbf{x}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.147, + 0.442, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.242, + 0.442, + 0.251 + ], + "angle": 0, + "content": "b Blurry background \\(\\tilde{\\pmb{x}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.444, + 0.147, + 0.555, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.449, + 0.242, + 0.556, + 0.251 + ], + "angle": 0, + "content": "c Clear background \\(\\mathbf{b}_0\\)" + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.149, + 0.672, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.242, + 0.662, + 0.25 + ], + "angle": 0, + "content": "d Synthetic Rd on \\(\\tilde{\\mathbf{x}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.153, + 0.784, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.68, + 0.241, + 0.778, + 0.25 + ], + "angle": 0, + "content": "e Synthetic Rd on \\(\\mathbf{b}_0\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.263, + 0.785, + 0.292 + ], + "angle": 0, + "content": "Fig. 13. We show the differences between the synthetic raindrops [23] and real-world raindrops from Raindrop Clarity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.323, + 0.785, + 0.385 + ], + "angle": 0, + "content": "are more complex than daytime raindrops. Meanwhile, existing state-of-the-art methods are designed for daytime raindrop removal and neglect the effect of complex nighttime conditions. Thus, existing state-of-the-art methods cannot achieve promising results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.41, + 0.358, + 0.426 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.788, + 0.613 + ], + "angle": 0, + "content": "In this paper, we propose a new raindrop dataset, named Raindrop Clarity. It includes a total of 15,186 high-quality pairs/triplets of images with raindrops and the corresponding clear background images. Moreover, our dataset is collected under various conditions, including different scenes, times (daytime and nighttime), and focuses (raindrop- and background-focused). To be specific, 3,606 daytime raindrop-focused, 1,836 daytime background-focused, 4,838 nighttime raindrop-focused and 4,906 nighttime background-focused images are collected. With the proposed dataset, we re-analyze existing state-of-the-art methods, including raindrop algorithms and restoration backbones, and point out several open problems: raindrop-focus raindrop removal and nighttime raindrop removal. We wish to draw the attention of the community to the open problems." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.639, + 0.395, + 0.656 + ], + "angle": 0, + "content": "6 Social Impacts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Existing raindrop datasets focus on daytime background-focused raindrop removal. However, in the real world, raindrop images may focus on raindrops, resulting in a blurry background. Also, raindrops may occur in nighttime images. To draw the attention of the community to the open problems, we propose a new raindrop dataset, named Raindrop Clarity. It mainly collected from four conditions: daytime raindrop-focused, daytime background-focused, nighttime raindrop-focused and nighttime background-focused. Experimental results on our proposed dataset show that existing state-of-the-art methods cannot effectively remove raindrops under complex conditions. As a result, we believe that further exploration of our Raindrop Clarity dataset can help the community to address more complex raindrops." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.237, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Raindrop Clarity: Dual-Focused Dataset for Day Night Raindrop Removal" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.146, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.283 + ], + "angle": 0, + "content": "This research is supported by the National Research Foundation, Singapore, under its AI Singapore Programme (AISG Award No: AISG2-PhD/2022-01-037[T]). Xin Li's research is supported by NSFC under Grant 623B2098. Malu Zhang's work is supported by the National Science Foundation of China under Grant 62106038, and in part by the Sichuan Science and Technology Program under Grant 2023YFG0259. We thank Robby T. Tan and Beibei Lin for their valuable discussions, revisions, and feedback." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.303, + 0.323, + 0.319 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.332, + 0.785, + 0.371 + ], + "angle": 0, + "content": "1. [EB/OL], http://liuyunfei.net/Projs/PBRR/, title = PBRR: Physically Based Raindrop Rendering, author = Yunfei Liu, Zhixiang Hao, Shadi You, Yu Li, and Feng Lu" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.373, + 0.785, + 0.413 + ], + "angle": 0, + "content": "2. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12299-12310 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.415, + 0.785, + 0.467 + ], + "angle": 0, + "content": "3. Chen, S., Ye, T., Bai, J., Chen, E., Shi, J., Zhu, L.: Sparse sampling transformer with uncertainty-driven ranking for unified removal of raindrops and rain streaks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13106-13117 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.47, + 0.785, + 0.509 + ], + "angle": 0, + "content": "4. Chen, S., Ye, T., Liu, Y., Chen, E.: Snowformer: Context interaction transformer with scale-awareness for single image desnowing. arXiv preprint arXiv:2208.09703 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.511, + 0.785, + 0.564 + ], + "angle": 0, + "content": "5. Chen, S., Ye, T., Liu, Y., Liao, T., Jiang, J., Chen, E., Chen, P.: Msp-former: Multi-scale projection transformer for single image desnowing. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.567, + 0.785, + 0.605 + ], + "angle": 0, + "content": "6. Eigen, D., Krishnan, D., Fergus, R.: Restoring an image taken through a window covered with dirt or rain. In: Proceedings of the IEEE international conference on computer vision. pp. 633-640 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.607, + 0.785, + 0.646 + ], + "angle": 0, + "content": "7. Halimeh, J.C., Roser, M.: Raindrop detection on car windshields using geometric-photometric environment construction and intensity-based correlation. In: 2009 IEEE Intelligent Vehicles Symposium. pp. 610-615. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.649, + 0.785, + 0.687 + ], + "angle": 0, + "content": "8. Hao, Z., You, S., Li, Y., Li, K., Lu, F.: Learning from synthetic photorealistic raindrop for single image raindrop removal. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. pp. 0-0 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.741 + ], + "angle": 0, + "content": "9. Jin, Y., Lin, B., Yan, W., Yuan, Y., Ye, W., Tan, R.T.: Enhancing visibility in nighttime haze images using guided apsf and gradient adaptive convolution. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 2446-2457 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.744, + 0.785, + 0.797 + ], + "angle": 0, + "content": "10. Jin, Y., Sharma, A., Tan, R.T.: Dc-shadownet: Single-image hard and soft shadow removal using unsupervised domain-classifier guided network. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5027-5036 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.785, + 0.839 + ], + "angle": 0, + "content": "11. Jin, Y., Yang, W., Tan, R.T.: Unsupervised night image enhancement: When layer decomposition meets light-effects suppression. In: European Conference on Computer Vision. pp. 404-421. Springer (2022)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.332, + 0.785, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.396, + 0.128 + ], + "angle": 0, + "content": "Y. Jin, X. Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "12. Jin, Y., Ye, W., Yang, W., Yuan, Y., Tan, R.T.: Des3: Adaptive attention-driven self and soft shadow removal using vit similarity. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 2634-2642 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.191, + 0.787, + 0.232 + ], + "angle": 0, + "content": "13. Li, B., Li, X., Zhu, H., Jin, Y., Feng, R., Zhang, Z., Chen, Z.: Sed: Semantic-aware discriminator for image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.787, + 0.274 + ], + "angle": 0, + "content": "14. Li, X., Jin, X., Lin, J., Liu, S., Wu, Y., Yu, T., Zhou, W., Chen, Z.: Learning disentangled feature representation for hybrid-distorted image restoration. In: ECCV. pp. 313-329. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.787, + 0.33 + ], + "angle": 0, + "content": "15. Li, X., Li, B., Jin, X., Lan, C., Chen, Z.: Learning distortion invariant representation for image restoration from a causality perspective. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1714-1724 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.332, + 0.787, + 0.374 + ], + "angle": 0, + "content": "16. Li, X., Ren, Y., Jin, X., Lan, C., Wang, X., Zeng, W., Wang, X., Chen, Z.: Diffusion models for image restoration and enhancement-a comprehensive survey. arXiv preprint arXiv:2308.09388 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.787, + 0.402 + ], + "angle": 0, + "content": "17. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. arXiv preprint arXiv:2108.10257 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.403, + 0.787, + 0.458 + ], + "angle": 0, + "content": "18. Liu, X., Suganuma, M., Sun, Z., Okatani, T.: Dual residual networks leveraging the potential of paired operations for image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7007-7016 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.459, + 0.787, + 0.528 + ], + "angle": 0, + "content": "19. Luo, J., Ceylan, D., Yoon, J.S., Zhao, N., Philip, J., Frühstück, A., Li, W., Richardt, C., Wang, T.: Intrinsic diffusion: Joint intrinsic layers from latent diffusion models. In: ACM SIGGRAPH 2024 Conference Papers. SIGGRAPH '24, Association for Computing Machinery, New York, NY, USA (2024). https://doi.org/10.1145/3641519.3657472, https://doi.org/10.1145/3641519.3657472" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.787, + 0.571 + ], + "angle": 0, + "content": "20. Luo, J., Zhao, N., Li, W., Richardt, C.: Crefnet: Learning consistent reflectance estimation with a decoder-sharing transformer. IEEE Transactions on Visualization and Computer Graphics (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.572, + 0.787, + 0.613 + ], + "angle": 0, + "content": "21. Ozdenizci, O., Legenstein, R.: Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.614, + 0.787, + 0.641 + ], + "angle": 0, + "content": "22. Peebles, W., Xie, S.: Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.643, + 0.787, + 0.684 + ], + "angle": 0, + "content": "23. Pizzati, F., Cerri, P., de Charette, R.: Physics-informed guided disentanglement in generative networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.787, + 0.727 + ], + "angle": 0, + "content": "24. Porav, H., Bruls, T., Newman, P.: I can see clearly now: Image restoration via deraining. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 7087-7093. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.728, + 0.787, + 0.755 + ], + "angle": 0, + "content": "25. Qian, R., Tan, R.T., Yang, W., Su, J., Liu, J.: Attentive generative adversarial network for raindrop removal from a single image pp. 2482-2491 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.787, + 0.798 + ], + "angle": 0, + "content": "26. Quan, R., Yu, X., Liang, Y., Yang, Y.: Removing raindrops and rain streaks in one go. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9147-9156 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "27. Quan, Y., Deng, S., Chen, Y., Ji, H.: Deep learning for seeing through window with raindrops. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2463-2471 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.237, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Raindrop Clarity: Dual-Focused Dataset for Day Night Raindrop Removal" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "28. Roser, M., Geiger, A.: Video-based raindrop detection for improved image registration. In: 2009 IEEE 12th International Conference on Computer Vision Workshops, ICCV Workshops. pp. 570-577. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "29. Roser, M., Kurz, J., Geiger, A.: Realistic modeling of water droplets for monocular adherent raindrop recognition usingBezier curves. In: Asian conference on computer vision. pp. 235-244. Springer (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.273 + ], + "angle": 0, + "content": "30. Soboleva, V., Shipitko, O.: Raindrops on windshield: Dataset and lightweight gradient-based detection algorithm. In: 2021 IEEE Symposium Series on Computational Intelligence (SSCI). pp. 1-7. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.314 + ], + "angle": 0, + "content": "31. Wang, C., Pan, J., Lin, W., Dong, J., Wang, W., Wu, X.M.: Selfpromer: Self-prompt dehazing transformers with depth-consistency. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 5327-5335 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.356 + ], + "angle": 0, + "content": "32. Wang, C., Pan, J., Wang, W., Dong, J., Wang, M., Ju, Y., Chen, J.: Promptstorer: A prompting image restoration method with degradation perception. Advances in Neural Information Processing Systems 36, 8898-8912 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.41 + ], + "angle": 0, + "content": "33. Wang, C., Pan, J., Wang, W., Fu, G., Liang, S., Wang, M., Wu, X.M., Liu, J.: Correlation matching transformation transformers for uhd image restoration. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 5336-5344 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.453 + ], + "angle": 0, + "content": "34. Wang, J., Pan, Z., Zhang, M., Tan, R.T., Li, H.: Restoring speaking lips from occlusion for audio-visual speech recognition. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 19144-19152 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.495 + ], + "angle": 0, + "content": "35. Wang, J., Qian, X., Zhang, M., Tan, R.T., Li, H.: Seeing what you said: Talking face generation guided by a lip reading expert. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14653-14662 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.536 + ], + "angle": 0, + "content": "36. Wang, Z., Cun, X., Bao, J., Zhou, W., Liu, J., Li, H.: Uformer: A general u-shaped transformer for image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17683-17693 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "37. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.619 + ], + "angle": 0, + "content": "38. Wei, M., Shen, Y., Wang, Y., Xie, H., Wang, F.L.: Raindiffusion: When unsupervised learning meets diffusion models for real-world image deraining. arXiv preprint arXiv:2301.09430 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.785, + 0.646 + ], + "angle": 0, + "content": "39. Wen, Q., Wu, Y., Chen, Q.: Video waterfall removal via spatio-temporal fusion in driving scenes. arXiv preprint arXiv:2302.05916 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.674 + ], + "angle": 0, + "content": "40. Xiao, J., Fu, X., Liu, A., Wu, F., Zha, Z.J.: Image de-raining transformer. IEEE Transactions on Pattern Analysis and Machine Intelligence pp. 1-18 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "41. Yasarla, R., Patel, V.M.: Uncertainty guided multi-scale residual learning-using a cycle spinning cnn for single image de-raining. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8405-8414 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.758 + ], + "angle": 0, + "content": "42. Ye, T., Chen, S., Bai, J., Shi, J., Xue, C., Jiang, J., Yin, J., Chen, E., Liu, Y.: Adverse weather removal with codebook priors. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12653-12664 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "43. Ye, T., Zhang, Y., Jiang, M., Chen, L., Liu, Y., Chen, S., Chen, E.: Perceiving and modeling density for image dehazing. In: European conference on computer vision. pp. 130-145. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "44. You, S., Tan, R.T., Kawakami, R., Ikeuchi, K.: Adherent raindrop detection and removal in video. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1035-1042 (2013)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.128 + ], + "angle": 0, + "content": "Y. Jin, X. Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "45. You, S., Tan, R.T., Kawakami, R., Mukaigawa, Y., Ikeuchi, K.: Adherent raindrop modeling, detection and removal in video. IEEE transactions on pattern analysis and machine intelligence 38(9), 1721-1733 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "46. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.258 + ], + "angle": 0, + "content": "47. Zhang, K., Li, D., Luo, W., Ren, W.: Dual attention-in-attention model for joint rain streak and raindrop removal. IEEE Transactions on Image Processing 30, 7608-7619 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.287 + ], + "angle": 0, + "content": "48. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.287 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Raindrop Clarity_ A Dual-Focused Dataset for Day and Night Raindrop Removal/64be914f-41c4-4c4e-9cfc-b492879688db_origin.pdf b/2024/Raindrop Clarity_ A Dual-Focused Dataset for Day and Night Raindrop Removal/64be914f-41c4-4c4e-9cfc-b492879688db_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..18abc633bb350d42bf2e403171890bba4a309c93 --- /dev/null +++ b/2024/Raindrop Clarity_ A Dual-Focused Dataset for Day and Night Raindrop Removal/64be914f-41c4-4c4e-9cfc-b492879688db_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b41681b66376b8ec0577952a669fdf57730a3862914341a9cd8d7eb4a6df6645 +size 44298155 diff --git a/2024/Raindrop Clarity_ A Dual-Focused Dataset for Day and Night Raindrop Removal/full.md b/2024/Raindrop Clarity_ A Dual-Focused Dataset for Day and Night Raindrop Removal/full.md new file mode 100644 index 0000000000000000000000000000000000000000..30a94653fe74cf26cb9bca716daf2d5d5c9cd8cc --- /dev/null +++ b/2024/Raindrop Clarity_ A Dual-Focused Dataset for Day and Night Raindrop Removal/full.md @@ -0,0 +1,268 @@ +# Raindrop Clarity: A Dual-Focused Dataset for Day and Night Raindrop Removal + +Yeying Jin $^{1}$ , Xin Li $^{2}$ , Jiadong Wang $^{1}$ , Yan Zhang $^{1}$ , and Malu Zhang $^{3}$ + +$^{1}$ National University of Singapore +2 University of Science and Technology of China +3 University of Electronic Science and Technology of China jinyeying@u.nus.edu, xin.li@ustc.edu.cn, maluzhang@uestc.edu.cn + +Abstract. Existing raindrop removal datasets have two shortcomings. First, they consist of images captured by cameras with a focus on the background, leading to the presence of blurry raindrops. To our knowledge, none of these datasets include images where the focus is specifically on raindrops, which results in a blurry background. Second, these datasets predominantly consist of daytime images, thereby lacking nighttime raindrop scenarios. Consequently, algorithms trained on these datasets may struggle to perform effectively in raindrop-focused or nighttime scenarios. The absence of datasets specifically designed for raindrop-focused and nighttime raindrops constrains research in this area. In this paper, we introduce a large-scale, real-world raindrop removal dataset called Raindrop Clarity. Raindrop Clarity comprises 15,186 high-quality pairs/triplets (raindrops, blur, and background) of images with raindrops and the corresponding clear background images. There are 5,442 daytime raindrop images and 9,744 nighttime raindrop images. Specifically, the 5,442 daytime images include 3,606 raindrop- and 1,836 background-focused images. While the 9,744 nighttime images contain 4,838 raindrop- and 4,906 background-focused images. Our dataset will enable the community to explore background-focused and raindrop-focused images, including challenges unique to daytime and nighttime conditions. + +Keywords: Raindrop-focused $\cdot$ Background-focused $\cdot$ Nighttime + +# 1 Introduction + +Adherent raindrops on lenses or windscreens can significantly reduce visibility. Raindrop removal is crucial for surveillance, self-driving cars, object detection, outdoor photography, filmmaking, augmented reality, etc. Recently, a few rain-drop datasets (e.g., [24-26, 30]) have been proposed. In Fig. 1, these datasets primarily contain daytime background scenes with blurry raindrops, as the images are taken with the camera's focus on the background. They have impacts in the raindrop removal area, inspiring several successful methods [24, 25, 27]. + +![](images/2a28196a7e79f658d09f317b322114916a74e7f5ea7307afe6b112d0824e465c.jpg) +(a) Qian + +![](images/2ab26cdc364edff5e64ea4b27c258ce6fd0a412e0217b41a1d9ab5b85b670ad1.jpg) +(b) RainDS Quan + +![](images/827e2afd6998367cec9353397a2f8cf683a58932f2422a2228faaedd4abcef03.jpg) + +![](images/051bf678be361065f2bc54067f67a02eb79a431d7e4088d8b25235e2f57a1591.jpg) +(c) RobotCar Porav +(d) Windshield + +![](images/f80e7ec8b1c18f776fbc2d80ce54fd4287a1a095c868abee17d129ad6a6c6e36.jpg) + +![](images/5de40a39599e5a1658d34b98f25e8a791fff58dbc087a43d8d48803620d25c2c.jpg) +Fig. 1. Existing raindrop datasets (e.g., [24-26,30]) exhibit two limitations: first, they do not include raindrop-focused images (Fig. 2 right); and second, they lack night raindrops (Fig. 2 bottom). +Fig. 2. Motivation for Raindrop Clarity: real images sourced from the Internet feature scenarios overlooked by existing datasets, including raindrop-focused (right) and nighttime raindrops (bottom). In this figure, $\mathrm{Rd} =$ Raindrop, $\mathrm{Bg} =$ Background. + +While these datasets are valuable, they have two limitations. First, a significant issue arises when the camera is in automatic focus mode. In Fig. 2 (right), the camera may unintentionally focus on raindrops that are attached to the lens, windshield, or glass, resulting in images with sharp raindrops and a blurry background. Consequently, methods designed for images with blurry raindrops and a sharp background may not perform optimally. Second, existing datasets consist only of daytime images. Therefore, existing methods, especially those reliant on training data, may struggle to effectively handle nighttime raindrop images because of the inherent domain gap between day and night conditions. + +In Fig. 2, we observe that besides daytime images with blurry raindrops and a sharp background, it is common to come across images with sharp raindrops and a blurry background (right), as well as nighttime raindrops (bottom). Thus, we collected varied raindrop images as shown in Fig. 3: the top-left with daytime blurry raindrops against a sharp background, the top-right with daytime sharp raindrops and a blurry background, the bottom-left with nighttime blurry raindrops and a clear background, and the bottom-right with nighttime sharp raindrops with a blurry background. + +Physical solutions like rain covers or wipers can reduce the presence of raindrops, but their effectiveness is limited. In certain situations where wipers cannot + +![](images/3e45e130b4dbc90f0eccd49d28b114afa6310ad369f1cba5d88c984f7c5caa5a.jpg) +Background-focused (Blurry raindrop) +Fig. 3. Raindrop Clarity: Examples of our pairs $(\tilde{\mathbf{x}}, b_0)$ and triplets $(\tilde{\mathbf{x}}, x_0, b_0)$ of a raindrop image $\tilde{\mathbf{x}}$ , the corresponding blurry background image $x_0$ and the corresponding clear background image $b_0$ , in day and night. For the pairs $(\tilde{\mathbf{x}}, b_0)$ , there is no $x_0$ , since in this case $x_0 = b_0$ . The reason we have the blurry background image $x_0$ in triplets is to have the raindrop difference maps $\tilde{\mathbf{m}}$ (see Fig. 8). + +![](images/56d15b79ab79137bc18d8d135fae2c13ebbb686963fcb4ed55f97210e4b15528.jpg) +Raindrop-focused (Blurry background) + +be used, or wind blows rain onto the lens, these physical measures become impractical. Physics-based methods [1,8,23,44,45] remove raindrops by leveraging the physics of raindrops or utilizing generated synthetic data. However, the effectiveness of methods trained on synthetic data relies on the quality of raindrop rendering models. Currently, raindrop rendering models commonly employ surface modeling and ray tracing [8,28,29]. Unfortunately, these methods may not fully consider factors such as scene depth, lighting conditions, and the 3D shapes of raindrops [23,28], which are crucial for realistic raindrop rendering. Additionally, raindrops can introduce refraction effects [7,44], distorting the shapes and positions of objects, which further complicates accurate raindrop rendering. To the best of our knowledge, developing a rendering approach that can produce physically accurate raindrops adaptable to various scenes is challenging. + +Motivated by the aforementioned challenges and the observation that existing raindrop datasets predominantly consist of background-focused images while neglecting raindrop-focused images and nighttime raindrops, we collect a novel dataset, called "Raindrop Clarity". Our dataset comprises 6,742 (1,836 daytime, 4,906 nighttime) pairs and 8,444 (3,606 daytime, 4,838 nighttime) triplets of raindrop-focused images and background-focused images. All of them have the corresponding clear background images. The triplets include one more raindrop-free blurry background images. Fig. 3 shows samples and explanation on the meaning of our pairs and triplets. From the 15,186 raindrops pairs/triplet, 5,442 are captured during daytime and the 9,744 at nighttime. Raindrops Clarity exhibits diverse sizes and shapes, and raindrop appearance varies based on the scene and lighting conditions. To our knowledge, Raindrop Clarity is the largest real-world raindrop dataset that contains both the pair and triplets, day and night. Our main contributions are summarized as follows: + +- We present the Raindrop Clarity dataset, a large-scale real-world dataset that encompasses both raindrop-focused triplets and background-focused pairs. For the pairs, we provide images with blurry raindrops and their cor + +responding clear backgrounds. For the triplets, we additionally provide the raindrop-free blurry background. + +- Raindrop Clarity includes images collected under both daytime and nighttime. It aims to remove raindrops and recover clear backgrounds regardless of day or night, whether the raindrops or background are blurry. +- Our experimental results indicate that even with the application of state-of-the-art raindrop removal methods, there are instances where these methods fail. This suggests the existence of unresolved challenges in the field of raindrop removal, which await further attention and investigation. + +# 2 Related Work + +Raindrop Dataset. Qian et al. [25] introduce a raindrop dataset consisting of raindrop and corresponding ground truths. Porav et al. [24] propose a stereo RobotCar dataset, using a rain-making device that causes one lens to be affected by raindrops, while the other lens remains clear. Hao et al. [8] create a synthetic dataset of adherent raindrops with pixel-level masks for training deraining algorithms. Quan et al. [26] introduce RainDS, a dataset that includes raindrops, rain steaks, and their corresponding ground truths. Soboleva and Shipitko [30] introduce Windshields, a dataset containing raindrop images on both camera lenses and windshields, annotated with binary masks representing the raindrop areas. However, all of these datasets primarily consist of daytime images taken by cameras focusing on the background scenes and thus have blurry raindrops. + +Single Image Raindrop Removal. Most image deraining methods have primarily focused on removing rain streaks rather than raindrops. However, recent learning-based methods have been proposed specifically for single-image raindrop removal. Eigen et al. [6] introduce a supervised CNN-based method to remove raindrops, using a shallow CNN. Qian et al. [25] propose an AttentiveGAN for raindrop removal, that relies on attention maps but has a limited focus on local spatial information. Quan et al. [27] propose RaindropAttention, which employs an attention mechanism to generate attention maps based on mathematical raindrop shapes, yet it lacks exploration of long-range information. Liu et al. [18] demonstrate a dual residual network, by leveraging the complementary nature of paired residual blocks. Hao et al. [8] propose a raindrop synthesis algorithm, used their synthetic dataset to train a multi-task network for raindrop detection and removal. Porav et al. [24] propose a Pix2PixHD-based deraining network architecture. Subsequent work by Quan et al. [26] propose a cascaded network architecture for simultaneous removal of rain streaks and raindrops, achieved through a neural architecture search. Yasarla and Patel [41] propose the uncertainty guided multi-scale residual learning network to effectively remove rain from images. Zhang et al. [47] propose the Dual Attention-in-Attention Model to remove rain streaks and raindrops. + +Restoration Backbone. There are image restoration and enhancement tasks that utilize transformer-based models, taking advantage of the long-range dependencies, including IPT [2] (vanilla transformer), Uformer [36], SwinIR [17] + +![](images/e0cc4a2196a527ed30aa7e521e72ab8fdc4ff802fd543412780a6ac3e07ba60c.jpg) +Background-focused (Blurry raindrop) +Fig. 4. Pairs $(\tilde{\mathbf{x}},\pmb{b}_0)$ show various sizes, shapes, and densities. + +![](images/b9a641c25144839a2280d1c9d874fd44c468af638a5b4c748589affc2fe33c90.jpg) +Raindrop-focused (Blurry background) +Fig. 5. The triplets $(\tilde{\mathbf{x}}, \mathbf{x}_0, \mathbf{b}_0)$ cover diverse scenes and shapes. + +and Restormer [46]. Xiao et al. [40] propose an image deraining transformer (IDT) for rain streak and raindrop removal with window-based and spatial-based dual Transformer. RaindropDiffusion [21] develops a UNet-based diffusion framework, which demonstrates robust generation capabilities for raindrops, rain streaks, and snow. RainDiffusion [38], an unsupervised diffusion model, utilizes unpaired real-world data instead of weakly adversarial training. DiT [22] is a transformer-based diffusion model. + +# 3 Raindrop Clarity Dataset + +# 3.1 Data Collection + +To create Raindrop Clarity, we employed a sphere pan-tilt platform to keep our camera stationary. The images were captured using Sony FDR-AX33 4K Ultra HD Handycam Camcorder, Sony alpha 7R III digital camera, the back camera of an iPhone 14 Pro and iPhone 15 Pro Max, ensuring the highest quality possible. The 4K Camcorder captured RGB images at a resolution $3840 \times 2164$ . + +Using the optical refraction model (raindrops appear upside-down), we projected raindrops by spraying water on glass and adjusting the distance, angle, speed, and patterns of the nozzle/wipers. We also captured in rain conditions. + +Our collection procedure was designed to ensure diversity in the raindrop (such as focuses, shapes and sizes), background scenes and lighting conditions. First, we fixed the camera position on a tripod and installed a glass plate. The distance between the camera and the glass plate ranged from $5\mathrm{cm}$ to $25\mathrm{cm}$ . Second, we sprayed waterdrops/raindrops onto the glass plate and focused the camera on the raindrops. Third, we sprayed the waterdrops/raindrops to capture images of them under different shapes and sizes, ranging from sparse to dense. The captured images include elliptical raindrops and raindrop flow traces $(\tilde{\mathbf{x}})$ . Fourth, we removed the glass plate, allowing the camera to focus on the very close plane. The duration is kept very short, within milliseconds, to minimize lighting changes. In this way, we obtained raindrop-free blurry background images $(\boldsymbol{x}_0)$ . Finally, we adjusted the camera focus to the background, directing the camera to focus on the distant background $(\boldsymbol{b}_0)$ . + +# 3.2 Characteristics of Raindrop Clarity + +Raindrop Clarity is a new dataset consisting of around 15,186 high-quality image raindrop pairs and triplets. In Fig. 3, our dataset introduces a unique element of triplet data, which includes raindrop-focused images, corresponding blurry raindrop-free background images, and their clear background equivalents, distinguishing it from existing datasets. The second feature that distinguishes the Raindrop Clarity from the existing ones is the collection of night raindrops. + +Objectives. Our dataset is designed to support research in the elimination of raindrops and the restoration of a clear background, irrespective of whether the camera's focus is on the raindrops (leading to a blurry background) or on the background (resulting in blurry raindrops). This holds true regardless of whether it is daytime or nighttime. + +Raindrops and Diversity. Our Raindrop Clarity dataset offers an extensive coverage of diverse raindrop sizes, shapes, and occlusion types in various scenes and lighting conditions, making it an exceptionally comprehensive resource for advancing raindrop removal research. In Fig. 4 and Fig. 5, the dataset captures raindrops and rain flows of different sizes and shapes, including heavy raindrops + +# Background-focused (Blurry raindrop) + +![](images/47bce7102b8f68aee4e427203c83661aa5c5a3c9c56a0d532d369558d5d3a1f9.jpg) +Fig. 6. The pairs $(\tilde{\mathbf{x}}, b_0)$ under different lighting conditions. + +# Raindrop-focused (Blurry background) + +![](images/d249440cd2567605a4ec72dcfa257fba87508684a1e32cfca34a75bbe140253c.jpg) +Fig. 7. Triplets $(\tilde{\mathbf{x}},\pmb {x}_0,\pmb {b}_0)$ under different lighting conditions. + +that result in unpredictable water flows and blurry backgrounds when they hit window panes or car windshields. Two primary types of occlusions caused by raindrops, namely partial and complete occlusions, are included in our dataset. Both types present unique challenges for raindrop removal algorithms. + +Scenes and Large-scale. The dataset encompasses a wide range of background scenes, such as urban, countryside, campus, wilderness, road, and aerial views, covering both day and night. This variation allows for the exploration of different lighting and weather conditions in different scenes. With approximately 38,816 images in total, including 15,186 high-quality pairs/triplets of images, the Raindrop Clarity dataset is one of the largest datasets of its kind. + +Lighting Conditions. The appearance of raindrops highly depends on the surrounding environment. A distinctive feature of our dataset is its comprehen + +Table 1. Summary of comparisons between our Raindrop Clarity and existing datasets. Bg-focus = Background-focused, Rd-focus = Raindrop-focused, RD = Raindrop. Raindrop Clarity uniquely features raindrop-focused triplets and the nighttime raindrops. + +
DatasetsComponentsN. ImagesSyn/RealBg-focusRd-focus
DayNightDayNight
Our RD Clarity (Day)3,606 Triplets+1,836 Pairs14,490Real✓×✓×
Our RD Clarity (Night)4,838 Triplets+4,906 Pairs24,326Real×✓×✓
Raindrop Qian [25]919 Pairs1,838Real✓×××
RainDS-Real [26]248 Quadruplets992Real✓×××
RainDS-Syn [26]1,200 Quadruplets4,800Syn✓×××
RobotCar Porav [24]4,818 Pairs9,636Real✓×××
Windshield [30]3,3903,390Real✓×××
+ +![](images/af482284e479f43a9d1193198ddcecedba50c33dfd8f36183ff77d445dfba4df.jpg) +Fig. 8. We show $(\tilde{\mathbf{x}},\pmb{x}_0,\tilde{\mathbf{m}},\pmb{b}_0)$ , difference map $\tilde{\mathbf{m}}$ is derived by distinguishing raindrop-focused $\tilde{\mathbf{x}}$ from blurry background $\pmb{x}_0$ . + +sive coverage of various environmental conditions, setting it apart from other datasets. The dataset's variation in day and night conditions makes it valuable for developing raindrop removal algorithms. + +During the daytime, raindrop images exhibit a diverse array of appearances influenced by multiple factors such as sunlight direction, intensity, and shadowing. In Figs. 4-5, our dataset captures this variety, with images taken at different times of the day and under varying weather conditions (such as light rain, heavy rain, and rain combined with wind), providing rich information for algorithms to learn. + +Unlike daytime, nighttime raindrops present unique challenges due to artificial lighting conditions and low light. The appearance of raindrops can drastically change under street lights, headlights, or neon lights. In Fig. 6 and Fig. 7, our dataset includes night images with varied artificial lights, contributing to the complexity and comprehensiveness of the dataset. These night images are challenging due to the higher contrast between raindrops and the background, and potential overexposure or underexposure issues. + +Background- and Raindrop-focused. When the raindrop image $\tilde{\mathbf{x}}$ is background-focused, we decompose a raindrop degraded image $\tilde{\mathbf{x}}$ into a clear background + +![](images/7f5b04c3855e7d9c0ef331fe024be4eaf97c211097e8eb0f93e4f506c4dc70e5.jpg) +Fig. 9. Comparison of existing methods on daytime raindrop-focused images, covering both raindrop algorithms and restoration backbones. + +image $\pmb{b}_0$ by following this model [25]: $\tilde{\mathbf{x}} = (1 - \mathbf{M})\odot \pmb{x}_0 + \mathbf{D}$ , where $\mathbf{D}$ is the effect of the blurry raindrops, referring to the localized blurry imagery formed by the light reflected from the surroundings. The operator $\odot$ denotes element-wise multiplication. When binary mask $\mathbf{M}(\mathbf{x})$ equals 1, it signifies that the pixel $x$ belongs to the area affected by raindrops. Conversely, if $\mathbf{M}(\mathbf{x})$ does not equal 1, it indicates that the pixel is within the background areas. In the case where $\tilde{\mathbf{x}}$ is background-focused raindrop images, $\pmb{x}_0$ is identical to the clear, raindrop-free image $\pmb{b}_0$ , namely $\pmb{x}_0 = \pmb{b}_0$ . Therefore, we use $(\tilde{\mathbf{x}},\pmb{b}_0)$ to stand for background-focused (resulting in blurry raindrops) pairs, shown in the left part of Fig. 3. + +In the right part of Fig. 3 and Fig. 8, when $\tilde{\mathbf{x}}$ is the raindrop-focused image, $\pmb{x}_0$ is the corresponding raindrop-free blurry background image, while $\pmb{b}_0$ is the clear background image. In this case, $\pmb{x}_0 \neq \pmb{b}_0$ . We can calculate the pixel-level difference map $\tilde{\mathbf{m}}$ between the raindrop-free blurry background $\pmb{x}_0$ and the raindrop-focused $\tilde{\mathbf{x}}$ as: $\tilde{\mathbf{m}} = \tilde{\mathbf{x}} - \pmb{x}_0$ . Therefore, we use $(\tilde{\mathbf{x}}, \pmb{x}_0, \pmb{b}_0)$ to stand for raindrop-focused (resulting in blurry background) triplets. + +Since the camera is stationary, neither the pairs $(\tilde{\mathbf{x}},\boldsymbol{b}_0)$ nor the triplets $(\tilde{\mathbf{x}},\boldsymbol{x}_0,\boldsymbol{b}_0)$ have alignment problems. This is also evident in the raindrop difference maps $\tilde{\mathbf{m}}$ , which accurately assign values to raindrop regions across multiple images. Specifically, for each pixel $\mathbf{x}$ in the input image, any discrepancies in the map's assigned values for the static background would indicate alignment issues; however, these are not observed, as $\tilde{\mathbf{m}}$ indicates only the raindrop regions. + +Table 2. Quantitative evaluation of daytime and nighttime raindrop data with a benchmark for raindrop algorithms and restoration backbones. + +
Data\Method InputRaindrop AlgorithmsRestoration Backbones
AtGAN [25]Robot [24]RainDS [26]IDT [40]RDdiff [21]Restor. [46]Uform. [36]DiT [22]
DayPSNR↑ 21.9223.6223.8225.3926.0525.5226.0825.7826.03
SSIM↑ 0.5600.6580.6700.7170.7360.7340.7480.7270.752
LPIPS↓ 0.2470.2000.1910.1470.1410.1110.1310.1480.106
NightPSNR↑ 24.7824.3824.5725.3926.8126.4826.8725.2626.23
SSIM↑ 0.7260.7730.7870.8200.8510.8310.8480.8160.826
LPIPS↓ 0.2090.1850.1820.1390.1250.1120.1230.1420.111
+ +![](images/331fa48fad1a23041aba7ee1afa5253c7288218d5e315086edca017ac6b088a6.jpg) +Fig. 10. Comparison of methods on nighttime raindrop-focused images, covering both raindrop algorithms and restoration backbones. + +Since the pairs/triplets of images are captured in quick succession, there are no problems with changes in illumination. + +The captured images do not contain any personal information such as phone numbers, or human faces. We will release the dataset publicly with car license plates obscured, as we believe that our Raindrop Clarity dataset will benefit the progress of the raindrop removal field. + +![](images/9f4e1ebd7e957d23f62ae571182b474f35f4ec8565594f6d54da0c0268d81c1d.jpg) +Fig. 11. Comparison of methods on daytime background-focused images, covering both raindrop algorithms and restoration backbones. + +# 4 Experiments + +Experimental Setup. Our benchmark includes two parts: Daytime and Night-time Raindrop Removal. As shown in Table 1, our daytime raindrop dataset contains a total of 5,442 paired/triplets images, where 4,713 paired images are used for the training set, and the remaining 729 paired images are regarded as the test set. Our nighttime raindrop dataset consists of 9,744 paired/triplets images, where 8,655 paired images are selected as the training set and the rest 1,089 paired images are used for evaluation. + +Baselines. To create the benchmark, we retrain existing state-of-the-art methods on our daytime and nighttime raindrop datasets. Our benchmark consists of five raindrop algorithms [21, 24-26, 40] and three restoration methods [22, 36, 46]. Specifically, IDT [40], RDiffusion [21], RainDS [26], Robot Car [24] and AtGAN [25] are raindrop algorithms, while Restormer [46], Uformer [36] and DiT [22] are restoration backbones. For a fair comparison, we conduct experiments using the original implementation of these methods. All experiments are implemented in PyTorch and are trained on four NVIDIA RTX A5000 GPUs. + +Comparison with Existing Datasets. The difference between our Raindrop Clarity dataset and the existing datasets is shown in Table. 1 and Fig. 2. + +The stereo RobotCar dataset [24] includes 4,818 image pairs, randomly selected for training, validation, and testing. However, the raindrops in this dataset are disproportionately large and lack realism. + +The raindrop dataset proposed by Qian et al. [25] consists of 861 training pairs and 58 testing pairs. While it features common raindrop shapes, it lacks the + +![](images/ca0b3af94eff70f0db6b6db57dcccd8e2d0dab2ea7ae8bf2aad48a14a0a26c7f.jpg) +Fig. 12. Comparison of methods on nighttime background-focused images, covering both raindrop algorithms and restoration backbones. + +diversity of real-world scenes [44]. Also, the backgrounds between the raindrop images and the ground truth are not perfectly aligned [39]. + +The RainDS-Real dataset, proposed by Quan et al. [26], comprises 248 quadruplets, including raindrops, rain streaks, and their corresponding ground truths. The RainDS-Synthetic dataset has a domain gap with the real raindrop images. Moreover, the real-world raindrops are consistently small, obscuring only a fraction of the image details, making recovery less challenging. + +The Windshield dataset, proposed by Soboleva and Shipitko [30], was captured by a camera attached to the vehicle during its movement in urban areas and highway environments. The Windshield does not provide the corresponding clean pairs, making this dataset not available for robust raindrop removal method testing. Additionally, frequent moving objects further complicate the dataset. + +Existing raindrop datasets, though valuable, have certain limitations. They do not include raindrop-focused and nighttime raindrops. In addition, we have noticed reflection errors in the datasets from Qian [25] and RainDS-Real [26]. + +Comparison with Synthetic Raindrops. In Fig. 13, we show the existing state-of-the-art synthetic physical model [23], although it can generate raindrops on blurry and clear backgrounds (Fig. 13 d generated on b, Fig. 13 e generated on c). It still has a gap with real-world collected raindrops. Real-world collected raindrops display diverse appearances, including various shapes such as round, elliptical, and long strips. Also, the transparency and glare effects of raindrops collected in the real-world varies. + +Quantitative Comparison. To effectively evaluate the results on our proposed datasets, we use three popular metrics PSNR, SSIM [37], and LPIPS [48] in generation tasks [3-5, 9-16, 19, 20, 31-35, 42, 43]. + +The qualitative results are shown in Table 2. For the daytime dataset, the PSNR and SSIM of the original input images are 21.92 and 0.560. Raindrop algorithm IDT [40] achieves the best performance (26.05 PSNR of and 0.736 of SSIM) compared with other raindrop algorithms. For the restoration backbones, DiT [22] achieves a PSNR of 26.03 and a SSIM of 0.752. Restormer [46] achieves a PSNR of 26.08 and a SSIM of 0.748, which also performs competitively. + +It can be observed that, for the nighttime dataset, PSNR and SSIM of the input images are 24.78 and 0.726. Among the raindrop algorithms, IDT [40] achieves better performance than other methods, with a PSNR of 26.81 and a SSIM of 0.851. For the restoration backbones, Restormer [46] achieves the highest PSNR of 26.87, while DiT [22] achieves the lowest LPIPS of 0.111, which outperforms other backbones. + +Qualitative Comparison. Figs. 9-12 show the qualitative results of state-of-the-art methods. To be specific, Fig. 9 and Fig. 11 show the results of daytime raindrop removal. The input images in Fig. 9 and Fig. 11 are daytime raindrop-focused and background-focused, respectively. Fig. 10 and Fig. 12 show the results of nighttime raindrop-focused and background-focused removal, respectively. + +It can be found that, for daytime and nighttime raindrop-focused removal in Fig. 9 and Fig. 10, existing state-of-the-art methods fail to remove raindrops and recover the background. This is because most of them are designed for background-focused scenes and thus neglect the blurry problem. To our knowledge, Restormer [46] and Uformer [36] are specifically designed to address deblurring problem. However, they still struggle with recovering details and textures. As a result, these methods are struggling to simultaneously remove raindrops and recover the background. + +In Fig. 11, for daytime background-focused inputs, we can observe that existing state-of-the-art methods can handle most raindrops. However, they are struggling to address long strips of raindrops. This is because their methods may not detect long strips of raindrops and thus fail to remove them. Furthermore, in Fig. 12, we find that existing state-of-the-art methods are struggling to handle nighttime background-focused samples. To be specific, raindrops in these samples cannot be totally removed and details of the restored images are not clear. This is because nighttime raindrops usually reflect artificial light and thus + +![](images/dee761b1207cfd6fcf196b74c1496cceaaeac553810491ff83e01575d5da90bc.jpg) +a Raindrop-focusedd $\tilde{\mathbf{x}}$ + +![](images/ffacb94dee9a2664638cdb81258fb4a0bac29230b2eb3de4d5b0ae5516dadfb8.jpg) +b Blurry background $\tilde{\pmb{x}}$ + +![](images/6086c7a2dec2d32060d361d0477405b2d20a3356077595350cafd12ece48e753.jpg) +c Clear background $\mathbf{b}_0$ +Fig. 13. We show the differences between the synthetic raindrops [23] and real-world raindrops from Raindrop Clarity. + +![](images/c401688e2e5e8b4f030a28a47f5b357e392d44d889e71c79d06037b9a389285a.jpg) +d Synthetic Rd on $\tilde{\mathbf{x}}$ + +![](images/129be1559b6919f978fa7536daf88bb212d0b0aed5e525eaf03108b6cf71556d.jpg) +e Synthetic Rd on $\mathbf{b}_0$ + +are more complex than daytime raindrops. Meanwhile, existing state-of-the-art methods are designed for daytime raindrop removal and neglect the effect of complex nighttime conditions. Thus, existing state-of-the-art methods cannot achieve promising results. + +# 5 Conclusion + +In this paper, we propose a new raindrop dataset, named Raindrop Clarity. It includes a total of 15,186 high-quality pairs/triplets of images with raindrops and the corresponding clear background images. Moreover, our dataset is collected under various conditions, including different scenes, times (daytime and nighttime), and focuses (raindrop- and background-focused). To be specific, 3,606 daytime raindrop-focused, 1,836 daytime background-focused, 4,838 nighttime raindrop-focused and 4,906 nighttime background-focused images are collected. With the proposed dataset, we re-analyze existing state-of-the-art methods, including raindrop algorithms and restoration backbones, and point out several open problems: raindrop-focus raindrop removal and nighttime raindrop removal. We wish to draw the attention of the community to the open problems. + +# 6 Social Impacts + +Existing raindrop datasets focus on daytime background-focused raindrop removal. However, in the real world, raindrop images may focus on raindrops, resulting in a blurry background. Also, raindrops may occur in nighttime images. To draw the attention of the community to the open problems, we propose a new raindrop dataset, named Raindrop Clarity. It mainly collected from four conditions: daytime raindrop-focused, daytime background-focused, nighttime raindrop-focused and nighttime background-focused. Experimental results on our proposed dataset show that existing state-of-the-art methods cannot effectively remove raindrops under complex conditions. As a result, we believe that further exploration of our Raindrop Clarity dataset can help the community to address more complex raindrops. + +# Acknowledgments + +This research is supported by the National Research Foundation, Singapore, under its AI Singapore Programme (AISG Award No: AISG2-PhD/2022-01-037[T]). Xin Li's research is supported by NSFC under Grant 623B2098. Malu Zhang's work is supported by the National Science Foundation of China under Grant 62106038, and in part by the Sichuan Science and Technology Program under Grant 2023YFG0259. We thank Robby T. Tan and Beibei Lin for their valuable discussions, revisions, and feedback. + +# References + +1. [EB/OL], http://liuyunfei.net/Projs/PBRR/, title = PBRR: Physically Based Raindrop Rendering, author = Yunfei Liu, Zhixiang Hao, Shadi You, Yu Li, and Feng Lu +2. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12299-12310 (2021) +3. Chen, S., Ye, T., Bai, J., Chen, E., Shi, J., Zhu, L.: Sparse sampling transformer with uncertainty-driven ranking for unified removal of raindrops and rain streaks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13106-13117 (2023) +4. Chen, S., Ye, T., Liu, Y., Chen, E.: Snowformer: Context interaction transformer with scale-awareness for single image desnowing. arXiv preprint arXiv:2208.09703 (2022) +5. Chen, S., Ye, T., Liu, Y., Liao, T., Jiang, J., Chen, E., Chen, P.: Msp-former: Multi-scale projection transformer for single image desnowing. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023) +6. Eigen, D., Krishnan, D., Fergus, R.: Restoring an image taken through a window covered with dirt or rain. In: Proceedings of the IEEE international conference on computer vision. pp. 633-640 (2013) +7. Halimeh, J.C., Roser, M.: Raindrop detection on car windshields using geometric-photometric environment construction and intensity-based correlation. In: 2009 IEEE Intelligent Vehicles Symposium. pp. 610-615. IEEE (2009) +8. Hao, Z., You, S., Li, Y., Li, K., Lu, F.: Learning from synthetic photorealistic raindrop for single image raindrop removal. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. pp. 0-0 (2019) +9. Jin, Y., Lin, B., Yan, W., Yuan, Y., Ye, W., Tan, R.T.: Enhancing visibility in nighttime haze images using guided apsf and gradient adaptive convolution. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 2446-2457 (2023) +10. Jin, Y., Sharma, A., Tan, R.T.: Dc-shadownet: Single-image hard and soft shadow removal using unsupervised domain-classifier guided network. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5027-5036 (2021) +11. Jin, Y., Yang, W., Tan, R.T.: Unsupervised night image enhancement: When layer decomposition meets light-effects suppression. In: European Conference on Computer Vision. pp. 404-421. Springer (2022) + +12. Jin, Y., Ye, W., Yang, W., Yuan, Y., Tan, R.T.: Des3: Adaptive attention-driven self and soft shadow removal using vit similarity. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 2634-2642 (2024) +13. Li, B., Li, X., Zhu, H., Jin, Y., Feng, R., Zhang, Z., Chen, Z.: Sed: Semantic-aware discriminator for image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024) +14. Li, X., Jin, X., Lin, J., Liu, S., Wu, Y., Yu, T., Zhou, W., Chen, Z.: Learning disentangled feature representation for hybrid-distorted image restoration. In: ECCV. pp. 313-329. Springer (2020) +15. Li, X., Li, B., Jin, X., Lan, C., Chen, Z.: Learning distortion invariant representation for image restoration from a causality perspective. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1714-1724 (2023) +16. Li, X., Ren, Y., Jin, X., Lan, C., Wang, X., Zeng, W., Wang, X., Chen, Z.: Diffusion models for image restoration and enhancement-a comprehensive survey. arXiv preprint arXiv:2308.09388 (2023) +17. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. arXiv preprint arXiv:2108.10257 (2021) +18. Liu, X., Suganuma, M., Sun, Z., Okatani, T.: Dual residual networks leveraging the potential of paired operations for image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7007-7016 (2019) +19. Luo, J., Ceylan, D., Yoon, J.S., Zhao, N., Philip, J., Frühstück, A., Li, W., Richardt, C., Wang, T.: Intrinsic diffusion: Joint intrinsic layers from latent diffusion models. In: ACM SIGGRAPH 2024 Conference Papers. SIGGRAPH '24, Association for Computing Machinery, New York, NY, USA (2024). https://doi.org/10.1145/3641519.3657472, https://doi.org/10.1145/3641519.3657472 +20. Luo, J., Zhao, N., Li, W., Richardt, C.: Crefnet: Learning consistent reflectance estimation with a decoder-sharing transformer. IEEE Transactions on Visualization and Computer Graphics (2023) +21. Ozdenizci, O., Legenstein, R.: Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) +22. Peebles, W., Xie, S.: Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748 (2022) +23. Pizzati, F., Cerri, P., de Charette, R.: Physics-informed guided disentanglement in generative networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) +24. Porav, H., Bruls, T., Newman, P.: I can see clearly now: Image restoration via deraining. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 7087-7093. IEEE (2019) +25. Qian, R., Tan, R.T., Yang, W., Su, J., Liu, J.: Attentive generative adversarial network for raindrop removal from a single image pp. 2482-2491 (2018) +26. Quan, R., Yu, X., Liang, Y., Yang, Y.: Removing raindrops and rain streaks in one go. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9147-9156 (2021) +27. Quan, Y., Deng, S., Chen, Y., Ji, H.: Deep learning for seeing through window with raindrops. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2463-2471 (2019) + +28. Roser, M., Geiger, A.: Video-based raindrop detection for improved image registration. In: 2009 IEEE 12th International Conference on Computer Vision Workshops, ICCV Workshops. pp. 570-577. IEEE (2009) +29. Roser, M., Kurz, J., Geiger, A.: Realistic modeling of water droplets for monocular adherent raindrop recognition usingBezier curves. In: Asian conference on computer vision. pp. 235-244. Springer (2010) +30. Soboleva, V., Shipitko, O.: Raindrops on windshield: Dataset and lightweight gradient-based detection algorithm. In: 2021 IEEE Symposium Series on Computational Intelligence (SSCI). pp. 1-7. IEEE (2021) +31. Wang, C., Pan, J., Lin, W., Dong, J., Wang, W., Wu, X.M.: Selfpromer: Self-prompt dehazing transformers with depth-consistency. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 5327-5335 (2024) +32. Wang, C., Pan, J., Wang, W., Dong, J., Wang, M., Ju, Y., Chen, J.: Promptstorer: A prompting image restoration method with degradation perception. Advances in Neural Information Processing Systems 36, 8898-8912 (2023) +33. Wang, C., Pan, J., Wang, W., Fu, G., Liang, S., Wang, M., Wu, X.M., Liu, J.: Correlation matching transformation transformers for uhd image restoration. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 5336-5344 (2024) +34. Wang, J., Pan, Z., Zhang, M., Tan, R.T., Li, H.: Restoring speaking lips from occlusion for audio-visual speech recognition. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 19144-19152 (2024) +35. Wang, J., Qian, X., Zhang, M., Tan, R.T., Li, H.: Seeing what you said: Talking face generation guided by a lip reading expert. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14653-14662 (2023) +36. Wang, Z., Cun, X., Bao, J., Zhou, W., Liu, J., Li, H.: Uformer: A general u-shaped transformer for image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17683-17693 (2022) +37. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004) +38. Wei, M., Shen, Y., Wang, Y., Xie, H., Wang, F.L.: Raindiffusion: When unsupervised learning meets diffusion models for real-world image deraining. arXiv preprint arXiv:2301.09430 (2023) +39. Wen, Q., Wu, Y., Chen, Q.: Video waterfall removal via spatio-temporal fusion in driving scenes. arXiv preprint arXiv:2302.05916 (2023) +40. Xiao, J., Fu, X., Liu, A., Wu, F., Zha, Z.J.: Image de-raining transformer. IEEE Transactions on Pattern Analysis and Machine Intelligence pp. 1-18 (2022) +41. Yasarla, R., Patel, V.M.: Uncertainty guided multi-scale residual learning-using a cycle spinning cnn for single image de-raining. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8405-8414 (2019) +42. Ye, T., Chen, S., Bai, J., Shi, J., Xue, C., Jiang, J., Yin, J., Chen, E., Liu, Y.: Adverse weather removal with codebook priors. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12653-12664 (2023) +43. Ye, T., Zhang, Y., Jiang, M., Chen, L., Liu, Y., Chen, S., Chen, E.: Perceiving and modeling density for image dehazing. In: European conference on computer vision. pp. 130-145. Springer (2022) +44. You, S., Tan, R.T., Kawakami, R., Ikeuchi, K.: Adherent raindrop detection and removal in video. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1035-1042 (2013) + +45. You, S., Tan, R.T., Kawakami, R., Mukaigawa, Y., Ikeuchi, K.: Adherent raindrop modeling, detection and removal in video. IEEE transactions on pattern analysis and machine intelligence 38(9), 1721-1733 (2015) +46. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR (2022) +47. Zhang, K., Li, D., Luo, W., Ren, W.: Dual attention-in-attention model for joint rain streak and raindrop removal. IEEE Transactions on Image Processing 30, 7608-7619 (2021) +48. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018) \ No newline at end of file diff --git a/2024/Raindrop Clarity_ A Dual-Focused Dataset for Day and Night Raindrop Removal/images.zip b/2024/Raindrop Clarity_ A Dual-Focused Dataset for Day and Night Raindrop Removal/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8cdbec437a2479ac195378acc97e3b4e67ec1f48 --- /dev/null +++ b/2024/Raindrop Clarity_ A Dual-Focused Dataset for Day and Night Raindrop Removal/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6009d5180e5b0d9af4ce10d501b87d0772b228a4c8156ec0d6005f02649f021d +size 1059781 diff --git a/2024/Raindrop Clarity_ A Dual-Focused Dataset for Day and Night Raindrop Removal/layout.json b/2024/Raindrop Clarity_ A Dual-Focused Dataset for Day and Night Raindrop Removal/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9cfb6596684966c62cd466d6e2dbcae29e64d708 --- /dev/null +++ b/2024/Raindrop Clarity_ A Dual-Focused Dataset for Day and Night Raindrop Removal/layout.json @@ -0,0 +1,8679 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 145, + 111, + 470, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 111, + 470, + 148 + ], + "spans": [ + { + "bbox": [ + 145, + 111, + 470, + 148 + ], + "type": "text", + "content": "Raindrop Clarity: A Dual-Focused Dataset for Day and Night Raindrop Removal" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 133, + 167, + 478, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 167, + 478, + 181 + ], + "spans": [ + { + "bbox": [ + 133, + 167, + 478, + 181 + ], + "type": "text", + "content": "Yeying Jin" + }, + { + "bbox": [ + 133, + 167, + 478, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 167, + 478, + 181 + ], + "type": "text", + "content": ", Xin Li" + }, + { + "bbox": [ + 133, + 167, + 478, + 181 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 167, + 478, + 181 + ], + "type": "text", + "content": ", Jiadong Wang" + }, + { + "bbox": [ + 133, + 167, + 478, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 167, + 478, + 181 + ], + "type": "text", + "content": ", Yan Zhang" + }, + { + "bbox": [ + 133, + 167, + 478, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 167, + 478, + 181 + ], + "type": "text", + "content": ", and Malu Zhang" + }, + { + "bbox": [ + 133, + 167, + 478, + 181 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 157, + 189, + 455, + 235 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 234, + 189, + 378, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 189, + 378, + 201 + ], + "spans": [ + { + "bbox": [ + 234, + 189, + 378, + 201 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 234, + 189, + 378, + 201 + ], + "type": "text", + "content": " National University of Singapore" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 207, + 201, + 406, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 201, + 406, + 212 + ], + "spans": [ + { + "bbox": [ + 207, + 201, + 406, + 212 + ], + "type": "text", + "content": "2 University of Science and Technology of China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 157, + 212, + 455, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 212, + 455, + 235 + ], + "spans": [ + { + "bbox": [ + 157, + 212, + 455, + 235 + ], + "type": "text", + "content": "3 University of Electronic Science and Technology of China jinyeying@u.nus.edu, xin.li@ustc.edu.cn, maluzhang@uestc.edu.cn" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 159, + 258, + 462, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 258, + 462, + 478 + ], + "spans": [ + { + "bbox": [ + 159, + 258, + 462, + 478 + ], + "type": "text", + "content": "Abstract. Existing raindrop removal datasets have two shortcomings. First, they consist of images captured by cameras with a focus on the background, leading to the presence of blurry raindrops. To our knowledge, none of these datasets include images where the focus is specifically on raindrops, which results in a blurry background. Second, these datasets predominantly consist of daytime images, thereby lacking nighttime raindrop scenarios. Consequently, algorithms trained on these datasets may struggle to perform effectively in raindrop-focused or nighttime scenarios. The absence of datasets specifically designed for raindrop-focused and nighttime raindrops constrains research in this area. In this paper, we introduce a large-scale, real-world raindrop removal dataset called Raindrop Clarity. Raindrop Clarity comprises 15,186 high-quality pairs/triplets (raindrops, blur, and background) of images with raindrops and the corresponding clear background images. There are 5,442 daytime raindrop images and 9,744 nighttime raindrop images. Specifically, the 5,442 daytime images include 3,606 raindrop- and 1,836 background-focused images. While the 9,744 nighttime images contain 4,838 raindrop- and 4,906 background-focused images. Our dataset will enable the community to explore background-focused and raindrop-focused images, including challenges unique to daytime and nighttime conditions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 488, + 427, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 488, + 427, + 500 + ], + "spans": [ + { + "bbox": [ + 160, + 488, + 427, + 500 + ], + "type": "text", + "content": "Keywords: Raindrop-focused " + }, + { + "bbox": [ + 160, + 488, + 427, + 500 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 488, + 427, + 500 + ], + "type": "text", + "content": " Background-focused " + }, + { + "bbox": [ + 160, + 488, + 427, + 500 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 488, + 427, + 500 + ], + "type": "text", + "content": " Nighttime" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 517, + 230, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 517, + 230, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 517, + 230, + 530 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 541, + 482, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 541, + 482, + 626 + ], + "spans": [ + { + "bbox": [ + 130, + 541, + 482, + 626 + ], + "type": "text", + "content": "Adherent raindrops on lenses or windscreens can significantly reduce visibility. Raindrop removal is crucial for surveillance, self-driving cars, object detection, outdoor photography, filmmaking, augmented reality, etc. Recently, a few rain-drop datasets (e.g., [24-26, 30]) have been proposed. In Fig. 1, these datasets primarily contain daytime background scenes with blurry raindrops, as the images are taken with the camera's focus on the background. They have impacts in the raindrop removal area, inspiring several successful methods [24, 25, 27]." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 632, + 235, + 643 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 235, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 235, + 643 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "text", + "content": "Our data and code are available at: https://github.com/jinyeying/RaindropClarity" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 186, + 114, + 252, + 205 + ], + "blocks": [ + { + "bbox": [ + 186, + 114, + 252, + 205 + ], + "lines": [ + { + "bbox": [ + 186, + 114, + 252, + 205 + ], + "spans": [ + { + "bbox": [ + 186, + 114, + 252, + 205 + ], + "type": "image", + "image_path": "2a28196a7e79f658d09f317b322114916a74e7f5ea7307afe6b112d0824e465c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 207, + 231, + 215 + ], + "lines": [ + { + "bbox": [ + 208, + 207, + 231, + 215 + ], + "spans": [ + { + "bbox": [ + 208, + 207, + 231, + 215 + ], + "type": "text", + "content": "(a) Qian" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 253, + 116, + 320, + 205 + ], + "blocks": [ + { + "bbox": [ + 253, + 116, + 320, + 205 + ], + "lines": [ + { + "bbox": [ + 253, + 116, + 320, + 205 + ], + "spans": [ + { + "bbox": [ + 253, + 116, + 320, + 205 + ], + "type": "image", + "image_path": "2ab26cdc364edff5e64ea4b27c258ce6fd0a412e0217b41a1d9ab5b85b670ad1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 263, + 207, + 310, + 215 + ], + "lines": [ + { + "bbox": [ + 263, + 207, + 310, + 215 + ], + "spans": [ + { + "bbox": [ + 263, + 207, + 310, + 215 + ], + "type": "text", + "content": "(b) RainDS Quan" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 320, + 116, + 370, + 160 + ], + "blocks": [ + { + "bbox": [ + 320, + 116, + 370, + 160 + ], + "lines": [ + { + "bbox": [ + 320, + 116, + 370, + 160 + ], + "spans": [ + { + "bbox": [ + 320, + 116, + 370, + 160 + ], + "type": "image", + "image_path": "827e2afd6998367cec9353397a2f8cf683a58932f2422a2228faaedd4abcef03.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 320, + 161, + 370, + 205 + ], + "blocks": [ + { + "bbox": [ + 320, + 161, + 370, + 205 + ], + "lines": [ + { + "bbox": [ + 320, + 161, + 370, + 205 + ], + "spans": [ + { + "bbox": [ + 320, + 161, + 370, + 205 + ], + "type": "image", + "image_path": "051bf678be361065f2bc54067f67a02eb79a431d7e4088d8b25235e2f57a1591.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 207, + 370, + 214 + ], + "lines": [ + { + "bbox": [ + 320, + 207, + 370, + 214 + ], + "spans": [ + { + "bbox": [ + 320, + 207, + 370, + 214 + ], + "type": "text", + "content": "(c) RobotCar Porav" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 379, + 207, + 419, + 215 + ], + "lines": [ + { + "bbox": [ + 379, + 207, + 419, + 215 + ], + "spans": [ + { + "bbox": [ + 379, + 207, + 419, + 215 + ], + "type": "text", + "content": "(d) Windshield" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 370, + 116, + 427, + 160 + ], + "blocks": [ + { + "bbox": [ + 370, + 116, + 427, + 160 + ], + "lines": [ + { + "bbox": [ + 370, + 116, + 427, + 160 + ], + "spans": [ + { + "bbox": [ + 370, + 116, + 427, + 160 + ], + "type": "image", + "image_path": "f80e7ec8b1c18f776fbc2d80ce54fd4287a1a095c868abee17d129ad6a6c6e36.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 179, + 270, + 434, + 369 + ], + "blocks": [ + { + "bbox": [ + 131, + 224, + 481, + 258 + ], + "lines": [ + { + "bbox": [ + 131, + 224, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 131, + 224, + 481, + 258 + ], + "type": "text", + "content": "Fig. 1. Existing raindrop datasets (e.g., [24-26,30]) exhibit two limitations: first, they do not include raindrop-focused images (Fig. 2 right); and second, they lack night raindrops (Fig. 2 bottom)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 179, + 270, + 434, + 369 + ], + "lines": [ + { + "bbox": [ + 179, + 270, + 434, + 369 + ], + "spans": [ + { + "bbox": [ + 179, + 270, + 434, + 369 + ], + "type": "image", + "image_path": "5de40a39599e5a1658d34b98f25e8a791fff58dbc087a43d8d48803620d25c2c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 378, + 481, + 412 + ], + "lines": [ + { + "bbox": [ + 131, + 378, + 481, + 412 + ], + "spans": [ + { + "bbox": [ + 131, + 378, + 481, + 412 + ], + "type": "text", + "content": "Fig. 2. Motivation for Raindrop Clarity: real images sourced from the Internet feature scenarios overlooked by existing datasets, including raindrop-focused (right) and nighttime raindrops (bottom). In this figure, " + }, + { + "bbox": [ + 131, + 378, + 481, + 412 + ], + "type": "inline_equation", + "content": "\\mathrm{Rd} =" + }, + { + "bbox": [ + 131, + 378, + 481, + 412 + ], + "type": "text", + "content": " Raindrop, " + }, + { + "bbox": [ + 131, + 378, + 481, + 412 + ], + "type": "inline_equation", + "content": "\\mathrm{Bg} =" + }, + { + "bbox": [ + 131, + 378, + 481, + 412 + ], + "type": "text", + "content": " Background." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 437, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 544 + ], + "type": "text", + "content": "While these datasets are valuable, they have two limitations. First, a significant issue arises when the camera is in automatic focus mode. In Fig. 2 (right), the camera may unintentionally focus on raindrops that are attached to the lens, windshield, or glass, resulting in images with sharp raindrops and a blurry background. Consequently, methods designed for images with blurry raindrops and a sharp background may not perform optimally. Second, existing datasets consist only of daytime images. Therefore, existing methods, especially those reliant on training data, may struggle to effectively handle nighttime raindrop images because of the inherent domain gap between day and night conditions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "type": "text", + "content": "In Fig. 2, we observe that besides daytime images with blurry raindrops and a sharp background, it is common to come across images with sharp raindrops and a blurry background (right), as well as nighttime raindrops (bottom). Thus, we collected varied raindrop images as shown in Fig. 3: the top-left with daytime blurry raindrops against a sharp background, the top-right with daytime sharp raindrops and a blurry background, the bottom-left with nighttime blurry raindrops and a clear background, and the bottom-right with nighttime sharp raindrops with a blurry background." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 642, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 642, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 642, + 481, + 666 + ], + "type": "text", + "content": "Physical solutions like rain covers or wipers can reduce the presence of raindrops, but their effectiveness is limited. In certain situations where wipers cannot" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "text", + "content": "Y. Jin, X. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 125, + 310, + 210 + ], + "blocks": [ + { + "bbox": [ + 173, + 115, + 291, + 125 + ], + "lines": [ + { + "bbox": [ + 173, + 115, + 291, + 125 + ], + "spans": [ + { + "bbox": [ + 173, + 115, + 291, + 125 + ], + "type": "text", + "content": "Background-focused (Blurry raindrop)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 134, + 125, + 310, + 210 + ], + "lines": [ + { + "bbox": [ + 134, + 125, + 310, + 210 + ], + "spans": [ + { + "bbox": [ + 134, + 125, + 310, + 210 + ], + "type": "image", + "image_path": "3e45e130b4dbc90f0eccd49d28b114afa6310ad369f1cba5d88c984f7c5caa5a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 220, + 482, + 276 + ], + "lines": [ + { + "bbox": [ + 130, + 220, + 482, + 276 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 482, + 276 + ], + "type": "text", + "content": "Fig. 3. Raindrop Clarity: Examples of our pairs " + }, + { + "bbox": [ + 130, + 220, + 482, + 276 + ], + "type": "inline_equation", + "content": "(\\tilde{\\mathbf{x}}, b_0)" + }, + { + "bbox": [ + 130, + 220, + 482, + 276 + ], + "type": "text", + "content": " and triplets " + }, + { + "bbox": [ + 130, + 220, + 482, + 276 + ], + "type": "inline_equation", + "content": "(\\tilde{\\mathbf{x}}, x_0, b_0)" + }, + { + "bbox": [ + 130, + 220, + 482, + 276 + ], + "type": "text", + "content": " of a raindrop image " + }, + { + "bbox": [ + 130, + 220, + 482, + 276 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 220, + 482, + 276 + ], + "type": "text", + "content": ", the corresponding blurry background image " + }, + { + "bbox": [ + 130, + 220, + 482, + 276 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 130, + 220, + 482, + 276 + ], + "type": "text", + "content": " and the corresponding clear background image " + }, + { + "bbox": [ + 130, + 220, + 482, + 276 + ], + "type": "inline_equation", + "content": "b_0" + }, + { + "bbox": [ + 130, + 220, + 482, + 276 + ], + "type": "text", + "content": ", in day and night. For the pairs " + }, + { + "bbox": [ + 130, + 220, + 482, + 276 + ], + "type": "inline_equation", + "content": "(\\tilde{\\mathbf{x}}, b_0)" + }, + { + "bbox": [ + 130, + 220, + 482, + 276 + ], + "type": "text", + "content": ", there is no " + }, + { + "bbox": [ + 130, + 220, + 482, + 276 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 130, + 220, + 482, + 276 + ], + "type": "text", + "content": ", since in this case " + }, + { + "bbox": [ + 130, + 220, + 482, + 276 + ], + "type": "inline_equation", + "content": "x_0 = b_0" + }, + { + "bbox": [ + 130, + 220, + 482, + 276 + ], + "type": "text", + "content": ". The reason we have the blurry background image " + }, + { + "bbox": [ + 130, + 220, + 482, + 276 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 130, + 220, + 482, + 276 + ], + "type": "text", + "content": " in triplets is to have the raindrop difference maps " + }, + { + "bbox": [ + 130, + 220, + 482, + 276 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{m}}" + }, + { + "bbox": [ + 130, + 220, + 482, + 276 + ], + "type": "text", + "content": " (see Fig. 8)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 314, + 125, + 479, + 210 + ], + "blocks": [ + { + "bbox": [ + 336, + 116, + 455, + 125 + ], + "lines": [ + { + "bbox": [ + 336, + 116, + 455, + 125 + ], + "spans": [ + { + "bbox": [ + 336, + 116, + 455, + 125 + ], + "type": "text", + "content": "Raindrop-focused (Blurry background)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 314, + 125, + 479, + 210 + ], + "lines": [ + { + "bbox": [ + 314, + 125, + 479, + 210 + ], + "spans": [ + { + "bbox": [ + 314, + 125, + 479, + 210 + ], + "type": "image", + "image_path": "56d15b79ab79137bc18d8d135fae2c13ebbb686963fcb4ed55f97210e4b15528.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 302, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 482, + 445 + ], + "type": "text", + "content": "be used, or wind blows rain onto the lens, these physical measures become impractical. Physics-based methods [1,8,23,44,45] remove raindrops by leveraging the physics of raindrops or utilizing generated synthetic data. However, the effectiveness of methods trained on synthetic data relies on the quality of raindrop rendering models. Currently, raindrop rendering models commonly employ surface modeling and ray tracing [8,28,29]. Unfortunately, these methods may not fully consider factors such as scene depth, lighting conditions, and the 3D shapes of raindrops [23,28], which are crucial for realistic raindrop rendering. Additionally, raindrops can introduce refraction effects [7,44], distorting the shapes and positions of objects, which further complicates accurate raindrop rendering. To the best of our knowledge, developing a rendering approach that can produce physically accurate raindrops adaptable to various scenes is challenging." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 448, + 482, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 615 + ], + "type": "text", + "content": "Motivated by the aforementioned challenges and the observation that existing raindrop datasets predominantly consist of background-focused images while neglecting raindrop-focused images and nighttime raindrops, we collect a novel dataset, called \"Raindrop Clarity\". Our dataset comprises 6,742 (1,836 daytime, 4,906 nighttime) pairs and 8,444 (3,606 daytime, 4,838 nighttime) triplets of raindrop-focused images and background-focused images. All of them have the corresponding clear background images. The triplets include one more raindrop-free blurry background images. Fig. 3 shows samples and explanation on the meaning of our pairs and triplets. From the 15,186 raindrops pairs/triplet, 5,442 are captured during daytime and the 9,744 at nighttime. Raindrops Clarity exhibits diverse sizes and shapes, and raindrop appearance varies based on the scene and lighting conditions. To our knowledge, Raindrop Clarity is the largest real-world raindrop dataset that contains both the pair and triplets, day and night. Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 137, + 629, + 482, + 666 + ], + "type": "text", + "content": "- We present the Raindrop Clarity dataset, a large-scale real-world dataset that encompasses both raindrop-focused triplets and background-focused pairs. For the pairs, we provide images with blurry raindrops and their cor" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "type": "text", + "content": "Raindrop Clarity: Dual-Focused Dataset for Day Night Raindrop Removal" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 147, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 480, + 140 + ], + "type": "text", + "content": "responding clear backgrounds. For the triplets, we additionally provide the raindrop-free blurry background." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 479, + 222 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 138, + 140, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 479, + 175 + ], + "type": "text", + "content": "- Raindrop Clarity includes images collected under both daytime and nighttime. It aims to remove raindrops and recover clear backgrounds regardless of day or night, whether the raindrops or background are blurry." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 176, + 479, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 176, + 479, + 222 + ], + "spans": [ + { + "bbox": [ + 138, + 176, + 479, + 222 + ], + "type": "text", + "content": "- Our experimental results indicate that even with the application of state-of-the-art raindrop removal methods, there are instances where these methods fail. This suggests the existence of unresolved challenges in the field of raindrop removal, which await further attention and investigation." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 243, + 237, + 256 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 243, + 237, + 256 + ], + "spans": [ + { + "bbox": [ + 132, + 243, + 237, + 256 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 270, + 480, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 480, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 480, + 400 + ], + "type": "text", + "content": "Raindrop Dataset. Qian et al. [25] introduce a raindrop dataset consisting of raindrop and corresponding ground truths. Porav et al. [24] propose a stereo RobotCar dataset, using a rain-making device that causes one lens to be affected by raindrops, while the other lens remains clear. Hao et al. [8] create a synthetic dataset of adherent raindrops with pixel-level masks for training deraining algorithms. Quan et al. [26] introduce RainDS, a dataset that includes raindrops, rain steaks, and their corresponding ground truths. Soboleva and Shipitko [30] introduce Windshields, a dataset containing raindrop images on both camera lenses and windshields, annotated with binary masks representing the raindrop areas. However, all of these datasets primarily consist of daytime images taken by cameras focusing on the background scenes and thus have blurry raindrops." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 402, + 480, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 480, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 480, + 628 + ], + "type": "text", + "content": "Single Image Raindrop Removal. Most image deraining methods have primarily focused on removing rain streaks rather than raindrops. However, recent learning-based methods have been proposed specifically for single-image raindrop removal. Eigen et al. [6] introduce a supervised CNN-based method to remove raindrops, using a shallow CNN. Qian et al. [25] propose an AttentiveGAN for raindrop removal, that relies on attention maps but has a limited focus on local spatial information. Quan et al. [27] propose RaindropAttention, which employs an attention mechanism to generate attention maps based on mathematical raindrop shapes, yet it lacks exploration of long-range information. Liu et al. [18] demonstrate a dual residual network, by leveraging the complementary nature of paired residual blocks. Hao et al. [8] propose a raindrop synthesis algorithm, used their synthetic dataset to train a multi-task network for raindrop detection and removal. Porav et al. [24] propose a Pix2PixHD-based deraining network architecture. Subsequent work by Quan et al. [26] propose a cascaded network architecture for simultaneous removal of rain streaks and raindrops, achieved through a neural architecture search. Yasarla and Patel [41] propose the uncertainty guided multi-scale residual learning network to effectively remove rain from images. Zhang et al. [47] propose the Dual Attention-in-Attention Model to remove rain streaks and raindrops." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Restoration Backbone. There are image restoration and enhancement tasks that utilize transformer-based models, taking advantage of the long-range dependencies, including IPT [2] (vanilla transformer), Uformer [36], SwinIR [17]" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "text", + "content": "Y. Jin, X. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 186, + 125, + 428, + 242 + ], + "blocks": [ + { + "bbox": [ + 237, + 114, + 381, + 125 + ], + "lines": [ + { + "bbox": [ + 237, + 114, + 381, + 125 + ], + "spans": [ + { + "bbox": [ + 237, + 114, + 381, + 125 + ], + "type": "text", + "content": "Background-focused (Blurry raindrop)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 186, + 125, + 428, + 242 + ], + "lines": [ + { + "bbox": [ + 186, + 125, + 428, + 242 + ], + "spans": [ + { + "bbox": [ + 186, + 125, + 428, + 242 + ], + "type": "image", + "image_path": "e0cc4a2196a527ed30aa7e521e72ab8fdc4ff802fd543412780a6ac3e07ba60c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 180, + 251, + 432, + 263 + ], + "lines": [ + { + "bbox": [ + 180, + 251, + 432, + 263 + ], + "spans": [ + { + "bbox": [ + 180, + 251, + 432, + 263 + ], + "type": "text", + "content": "Fig. 4. Pairs " + }, + { + "bbox": [ + 180, + 251, + 432, + 263 + ], + "type": "inline_equation", + "content": "(\\tilde{\\mathbf{x}},\\pmb{b}_0)" + }, + { + "bbox": [ + 180, + 251, + 432, + 263 + ], + "type": "text", + "content": " show various sizes, shapes, and densities." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 185, + 323, + 428, + 499 + ], + "blocks": [ + { + "bbox": [ + 237, + 313, + 383, + 323 + ], + "lines": [ + { + "bbox": [ + 237, + 313, + 383, + 323 + ], + "spans": [ + { + "bbox": [ + 237, + 313, + 383, + 323 + ], + "type": "text", + "content": "Raindrop-focused (Blurry background)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 185, + 323, + 428, + 499 + ], + "lines": [ + { + "bbox": [ + 185, + 323, + 428, + 499 + ], + "spans": [ + { + "bbox": [ + 185, + 323, + 428, + 499 + ], + "type": "image", + "image_path": "b9a641c25144839a2280d1c9d874fd44c468af638a5b4c748589affc2fe33c90.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 177, + 510, + 435, + 521 + ], + "lines": [ + { + "bbox": [ + 177, + 510, + 435, + 521 + ], + "spans": [ + { + "bbox": [ + 177, + 510, + 435, + 521 + ], + "type": "text", + "content": "Fig. 5. The triplets " + }, + { + "bbox": [ + 177, + 510, + 435, + 521 + ], + "type": "inline_equation", + "content": "(\\tilde{\\mathbf{x}}, \\mathbf{x}_0, \\mathbf{b}_0)" + }, + { + "bbox": [ + 177, + 510, + 435, + 521 + ], + "type": "text", + "content": " cover diverse scenes and shapes." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "content": "and Restormer [46]. Xiao et al. [40] propose an image deraining transformer (IDT) for rain streak and raindrop removal with window-based and spatial-based dual Transformer. RaindropDiffusion [21] develops a UNet-based diffusion framework, which demonstrates robust generation capabilities for raindrops, rain streaks, and snow. RainDiffusion [38], an unsupervised diffusion model, utilizes unpaired real-world data instead of weakly adversarial training. DiT [22] is a transformer-based diffusion model." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "type": "text", + "content": "Raindrop Clarity: Dual-Focused Dataset for Day Night Raindrop Removal" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 305, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 305, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 305, + 129 + ], + "type": "text", + "content": "3 Raindrop Clarity Dataset" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 240, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 240, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 240, + 152 + ], + "type": "text", + "content": "3.1 Data Collection" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 162, + 481, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 162, + 481, + 220 + ], + "spans": [ + { + "bbox": [ + 130, + 162, + 481, + 220 + ], + "type": "text", + "content": "To create Raindrop Clarity, we employed a sphere pan-tilt platform to keep our camera stationary. The images were captured using Sony FDR-AX33 4K Ultra HD Handycam Camcorder, Sony alpha 7R III digital camera, the back camera of an iPhone 14 Pro and iPhone 15 Pro Max, ensuring the highest quality possible. The 4K Camcorder captured RGB images at a resolution " + }, + { + "bbox": [ + 130, + 162, + 481, + 220 + ], + "type": "inline_equation", + "content": "3840 \\times 2164" + }, + { + "bbox": [ + 130, + 162, + 481, + 220 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 222, + 481, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 222, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 222, + 481, + 258 + ], + "type": "text", + "content": "Using the optical refraction model (raindrops appear upside-down), we projected raindrops by spraying water on glass and adjusting the distance, angle, speed, and patterns of the nozzle/wipers. We also captured in rain conditions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 258, + 482, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 414 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 414 + ], + "type": "text", + "content": "Our collection procedure was designed to ensure diversity in the raindrop (such as focuses, shapes and sizes), background scenes and lighting conditions. First, we fixed the camera position on a tripod and installed a glass plate. The distance between the camera and the glass plate ranged from " + }, + { + "bbox": [ + 130, + 258, + 482, + 414 + ], + "type": "inline_equation", + "content": "5\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 258, + 482, + 414 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 258, + 482, + 414 + ], + "type": "inline_equation", + "content": "25\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 258, + 482, + 414 + ], + "type": "text", + "content": ". Second, we sprayed waterdrops/raindrops onto the glass plate and focused the camera on the raindrops. Third, we sprayed the waterdrops/raindrops to capture images of them under different shapes and sizes, ranging from sparse to dense. The captured images include elliptical raindrops and raindrop flow traces " + }, + { + "bbox": [ + 130, + 258, + 482, + 414 + ], + "type": "inline_equation", + "content": "(\\tilde{\\mathbf{x}})" + }, + { + "bbox": [ + 130, + 258, + 482, + 414 + ], + "type": "text", + "content": ". Fourth, we removed the glass plate, allowing the camera to focus on the very close plane. The duration is kept very short, within milliseconds, to minimize lighting changes. In this way, we obtained raindrop-free blurry background images " + }, + { + "bbox": [ + 130, + 258, + 482, + 414 + ], + "type": "inline_equation", + "content": "(\\boldsymbol{x}_0)" + }, + { + "bbox": [ + 130, + 258, + 482, + 414 + ], + "type": "text", + "content": ". Finally, we adjusted the camera focus to the background, directing the camera to focus on the distant background " + }, + { + "bbox": [ + 130, + 258, + 482, + 414 + ], + "type": "inline_equation", + "content": "(\\boldsymbol{b}_0)" + }, + { + "bbox": [ + 130, + 258, + 482, + 414 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 431, + 338, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 338, + 444 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 338, + 444 + ], + "type": "text", + "content": "3.2 Characteristics of Raindrop Clarity" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 453, + 481, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 481, + 525 + ], + "type": "text", + "content": "Raindrop Clarity is a new dataset consisting of around 15,186 high-quality image raindrop pairs and triplets. In Fig. 3, our dataset introduces a unique element of triplet data, which includes raindrop-focused images, corresponding blurry raindrop-free background images, and their clear background equivalents, distinguishing it from existing datasets. The second feature that distinguishes the Raindrop Clarity from the existing ones is the collection of night raindrops." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 535, + 481, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 535, + 481, + 595 + ], + "spans": [ + { + "bbox": [ + 130, + 535, + 481, + 595 + ], + "type": "text", + "content": "Objectives. Our dataset is designed to support research in the elimination of raindrops and the restoration of a clear background, irrespective of whether the camera's focus is on the raindrops (leading to a blurry background) or on the background (resulting in blurry raindrops). This holds true regardless of whether it is daytime or nighttime." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "Raindrops and Diversity. Our Raindrop Clarity dataset offers an extensive coverage of diverse raindrop sizes, shapes, and occlusion types in various scenes and lighting conditions, making it an exceptionally comprehensive resource for advancing raindrop removal research. In Fig. 4 and Fig. 5, the dataset captures raindrops and rain flows of different sizes and shapes, including heavy raindrops" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "text", + "content": "Y. Jin, X. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 237, + 114, + 381, + 125 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 114, + 381, + 125 + ], + "spans": [ + { + "bbox": [ + 237, + 114, + 381, + 125 + ], + "type": "text", + "content": "Background-focused (Blurry raindrop)" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 186, + 125, + 428, + 240 + ], + "blocks": [ + { + "bbox": [ + 186, + 125, + 428, + 240 + ], + "lines": [ + { + "bbox": [ + 186, + 125, + 428, + 240 + ], + "spans": [ + { + "bbox": [ + 186, + 125, + 428, + 240 + ], + "type": "image", + "image_path": "47bce7102b8f68aee4e427203c83661aa5c5a3c9c56a0d532d369558d5d3a1f9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 251, + 428, + 262 + ], + "lines": [ + { + "bbox": [ + 184, + 251, + 428, + 262 + ], + "spans": [ + { + "bbox": [ + 184, + 251, + 428, + 262 + ], + "type": "text", + "content": "Fig. 6. The pairs " + }, + { + "bbox": [ + 184, + 251, + 428, + 262 + ], + "type": "inline_equation", + "content": "(\\tilde{\\mathbf{x}}, b_0)" + }, + { + "bbox": [ + 184, + 251, + 428, + 262 + ], + "type": "text", + "content": " under different lighting conditions." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 237, + 273, + 383, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 273, + 383, + 284 + ], + "spans": [ + { + "bbox": [ + 237, + 273, + 383, + 284 + ], + "type": "text", + "content": "Raindrop-focused (Blurry background)" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 184, + 284, + 427, + 460 + ], + "blocks": [ + { + "bbox": [ + 184, + 284, + 427, + 460 + ], + "lines": [ + { + "bbox": [ + 184, + 284, + 427, + 460 + ], + "spans": [ + { + "bbox": [ + 184, + 284, + 427, + 460 + ], + "type": "image", + "image_path": "d249440cd2567605a4ec72dcfa257fba87508684a1e32cfca34a75bbe140253c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 469, + 432, + 480 + ], + "lines": [ + { + "bbox": [ + 181, + 469, + 432, + 480 + ], + "spans": [ + { + "bbox": [ + 181, + 469, + 432, + 480 + ], + "type": "text", + "content": "Fig. 7. Triplets " + }, + { + "bbox": [ + 181, + 469, + 432, + 480 + ], + "type": "inline_equation", + "content": "(\\tilde{\\mathbf{x}},\\pmb {x}_0,\\pmb {b}_0)" + }, + { + "bbox": [ + 181, + 469, + 432, + 480 + ], + "type": "text", + "content": " under different lighting conditions." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 503, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 482, + 552 + ], + "type": "text", + "content": "that result in unpredictable water flows and blurry backgrounds when they hit window panes or car windshields. Two primary types of occlusions caused by raindrops, namely partial and complete occlusions, are included in our dataset. Both types present unique challenges for raindrop removal algorithms." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 560, + 482, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 560, + 482, + 633 + ], + "spans": [ + { + "bbox": [ + 130, + 560, + 482, + 633 + ], + "type": "text", + "content": "Scenes and Large-scale. The dataset encompasses a wide range of background scenes, such as urban, countryside, campus, wilderness, road, and aerial views, covering both day and night. This variation allows for the exploration of different lighting and weather conditions in different scenes. With approximately 38,816 images in total, including 15,186 high-quality pairs/triplets of images, the Raindrop Clarity dataset is one of the largest datasets of its kind." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Lighting Conditions. The appearance of raindrops highly depends on the surrounding environment. A distinctive feature of our dataset is its comprehen" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "type": "text", + "content": "Raindrop Clarity: Dual-Focused Dataset for Day Night Raindrop Removal" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 157, + 482, + 256 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 1. Summary of comparisons between our Raindrop Clarity and existing datasets. Bg-focus = Background-focused, Rd-focus = Raindrop-focused, RD = Raindrop. Raindrop Clarity uniquely features raindrop-focused triplets and the nighttime raindrops." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 157, + 482, + 256 + ], + "lines": [ + { + "bbox": [ + 135, + 157, + 482, + 256 + ], + "spans": [ + { + "bbox": [ + 135, + 157, + 482, + 256 + ], + "type": "table", + "html": "
DatasetsComponentsN. ImagesSyn/RealBg-focusRd-focus
DayNightDayNight
Our RD Clarity (Day)3,606 Triplets+1,836 Pairs14,490Real✓×✓×
Our RD Clarity (Night)4,838 Triplets+4,906 Pairs24,326Real×✓×✓
Raindrop Qian [25]919 Pairs1,838Real✓×××
RainDS-Real [26]248 Quadruplets992Real✓×××
RainDS-Syn [26]1,200 Quadruplets4,800Syn✓×××
RobotCar Porav [24]4,818 Pairs9,636Real✓×××
Windshield [30]3,3903,390Real✓×××
", + "image_path": "cc2e87abdb21276293405bcef7c1766af32b89157de6ac621f6a76fe7f400430.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 185, + 265, + 430, + 390 + ], + "blocks": [ + { + "bbox": [ + 185, + 265, + 430, + 390 + ], + "lines": [ + { + "bbox": [ + 185, + 265, + 430, + 390 + ], + "spans": [ + { + "bbox": [ + 185, + 265, + 430, + 390 + ], + "type": "image", + "image_path": "af482284e479f43a9d1193198ddcecedba50c33dfd8f36183ff77d445dfba4df.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 398, + 481, + 420 + ], + "lines": [ + { + "bbox": [ + 131, + 398, + 481, + 420 + ], + "spans": [ + { + "bbox": [ + 131, + 398, + 481, + 420 + ], + "type": "text", + "content": "Fig. 8. We show " + }, + { + "bbox": [ + 131, + 398, + 481, + 420 + ], + "type": "inline_equation", + "content": "(\\tilde{\\mathbf{x}},\\pmb{x}_0,\\tilde{\\mathbf{m}},\\pmb{b}_0)" + }, + { + "bbox": [ + 131, + 398, + 481, + 420 + ], + "type": "text", + "content": ", difference map " + }, + { + "bbox": [ + 131, + 398, + 481, + 420 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{m}}" + }, + { + "bbox": [ + 131, + 398, + 481, + 420 + ], + "type": "text", + "content": " is derived by distinguishing raindrop-focused " + }, + { + "bbox": [ + 131, + 398, + 481, + 420 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}" + }, + { + "bbox": [ + 131, + 398, + 481, + 420 + ], + "type": "text", + "content": " from blurry background " + }, + { + "bbox": [ + 131, + 398, + 481, + 420 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 131, + 398, + 481, + 420 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 442, + 480, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 442, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 442, + 480, + 478 + ], + "type": "text", + "content": "sive coverage of various environmental conditions, setting it apart from other datasets. The dataset's variation in day and night conditions makes it valuable for developing raindrop removal algorithms." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 479, + 480, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 480, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 480, + 548 + ], + "type": "text", + "content": "During the daytime, raindrop images exhibit a diverse array of appearances influenced by multiple factors such as sunlight direction, intensity, and shadowing. In Figs. 4-5, our dataset captures this variety, with images taken at different times of the day and under varying weather conditions (such as light rain, heavy rain, and rain combined with wind), providing rich information for algorithms to learn." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 550, + 481, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 550, + 481, + 634 + ], + "spans": [ + { + "bbox": [ + 130, + 550, + 481, + 634 + ], + "type": "text", + "content": "Unlike daytime, nighttime raindrops present unique challenges due to artificial lighting conditions and low light. The appearance of raindrops can drastically change under street lights, headlights, or neon lights. In Fig. 6 and Fig. 7, our dataset includes night images with varied artificial lights, contributing to the complexity and comprehensiveness of the dataset. These night images are challenging due to the higher contrast between raindrops and the background, and potential overexposure or underexposure issues." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 641, + 497, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 497, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 497, + 666 + ], + "type": "text", + "content": "Background- and Raindrop-focused. When the raindrop image " + }, + { + "bbox": [ + 131, + 641, + 497, + 666 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}" + }, + { + "bbox": [ + 131, + 641, + 497, + 666 + ], + "type": "text", + "content": " is background-focused, we decompose a raindrop degraded image " + }, + { + "bbox": [ + 131, + 641, + 497, + 666 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}" + }, + { + "bbox": [ + 131, + 641, + 497, + 666 + ], + "type": "text", + "content": " into a clear background" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "text", + "content": "Y. Jin, X. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 357 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 357 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 357 + ], + "type": "image", + "image_path": "7f5b04c3855e7d9c0ef331fe024be4eaf97c211097e8eb0f93e4f506c4dc70e5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 365, + 482, + 388 + ], + "lines": [ + { + "bbox": [ + 132, + 365, + 482, + 388 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 482, + 388 + ], + "type": "text", + "content": "Fig. 9. Comparison of existing methods on daytime raindrop-focused images, covering both raindrop algorithms and restoration backbones." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 413, + 482, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 520 + ], + "type": "text", + "content": "image " + }, + { + "bbox": [ + 130, + 413, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\pmb{b}_0" + }, + { + "bbox": [ + 130, + 413, + 482, + 520 + ], + "type": "text", + "content": " by following this model [25]: " + }, + { + "bbox": [ + 130, + 413, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}} = (1 - \\mathbf{M})\\odot \\pmb{x}_0 + \\mathbf{D}" + }, + { + "bbox": [ + 130, + 413, + 482, + 520 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 413, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 130, + 413, + 482, + 520 + ], + "type": "text", + "content": " is the effect of the blurry raindrops, referring to the localized blurry imagery formed by the light reflected from the surroundings. The operator " + }, + { + "bbox": [ + 130, + 413, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 413, + 482, + 520 + ], + "type": "text", + "content": " denotes element-wise multiplication. When binary mask " + }, + { + "bbox": [ + 130, + 413, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\mathbf{M}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 413, + 482, + 520 + ], + "type": "text", + "content": " equals 1, it signifies that the pixel " + }, + { + "bbox": [ + 130, + 413, + 482, + 520 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 413, + 482, + 520 + ], + "type": "text", + "content": " belongs to the area affected by raindrops. Conversely, if " + }, + { + "bbox": [ + 130, + 413, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\mathbf{M}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 413, + 482, + 520 + ], + "type": "text", + "content": " does not equal 1, it indicates that the pixel is within the background areas. In the case where " + }, + { + "bbox": [ + 130, + 413, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 413, + 482, + 520 + ], + "type": "text", + "content": " is background-focused raindrop images, " + }, + { + "bbox": [ + 130, + 413, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 130, + 413, + 482, + 520 + ], + "type": "text", + "content": " is identical to the clear, raindrop-free image " + }, + { + "bbox": [ + 130, + 413, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\pmb{b}_0" + }, + { + "bbox": [ + 130, + 413, + 482, + 520 + ], + "type": "text", + "content": ", namely " + }, + { + "bbox": [ + 130, + 413, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0 = \\pmb{b}_0" + }, + { + "bbox": [ + 130, + 413, + 482, + 520 + ], + "type": "text", + "content": ". Therefore, we use " + }, + { + "bbox": [ + 130, + 413, + 482, + 520 + ], + "type": "inline_equation", + "content": "(\\tilde{\\mathbf{x}},\\pmb{b}_0)" + }, + { + "bbox": [ + 130, + 413, + 482, + 520 + ], + "type": "text", + "content": " to stand for background-focused (resulting in blurry raindrops) pairs, shown in the left part of Fig. 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "text", + "content": "In the right part of Fig. 3 and Fig. 8, when " + }, + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "text", + "content": " is the raindrop-focused image, " + }, + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "text", + "content": " is the corresponding raindrop-free blurry background image, while " + }, + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\pmb{b}_0" + }, + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "text", + "content": " is the clear background image. In this case, " + }, + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0 \\neq \\pmb{b}_0" + }, + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "text", + "content": ". We can calculate the pixel-level difference map " + }, + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{m}}" + }, + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "text", + "content": " between the raindrop-free blurry background " + }, + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "text", + "content": " and the raindrop-focused " + }, + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "text", + "content": " as: " + }, + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{m}} = \\tilde{\\mathbf{x}} - \\pmb{x}_0" + }, + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "text", + "content": ". Therefore, we use " + }, + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "inline_equation", + "content": "(\\tilde{\\mathbf{x}}, \\pmb{x}_0, \\pmb{b}_0)" + }, + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "text", + "content": " to stand for raindrop-focused (resulting in blurry background) triplets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Since the camera is stationary, neither the pairs " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\tilde{\\mathbf{x}},\\boldsymbol{b}_0)" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " nor the triplets " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\tilde{\\mathbf{x}},\\boldsymbol{x}_0,\\boldsymbol{b}_0)" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " have alignment problems. This is also evident in the raindrop difference maps " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{m}}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", which accurately assign values to raindrop regions across multiple images. Specifically, for each pixel " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " in the input image, any discrepancies in the map's assigned values for the static background would indicate alignment issues; however, these are not observed, as " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{m}}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " indicates only the raindrop regions." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "type": "text", + "content": "Raindrop Clarity: Dual-Focused Dataset for Day Night Raindrop Removal" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 146, + 482, + 247 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 2. Quantitative evaluation of daytime and nighttime raindrop data with a benchmark for raindrop algorithms and restoration backbones." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 146, + 482, + 247 + ], + "lines": [ + { + "bbox": [ + 134, + 146, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 134, + 146, + 482, + 247 + ], + "type": "table", + "html": "
Data\\Method InputRaindrop AlgorithmsRestoration Backbones
AtGAN [25]Robot [24]RainDS [26]IDT [40]RDdiff [21]Restor. [46]Uform. [36]DiT [22]
DayPSNR↑ 21.9223.6223.8225.3926.0525.5226.0825.7826.03
SSIM↑ 0.5600.6580.6700.7170.7360.7340.7480.7270.752
LPIPS↓ 0.2470.2000.1910.1470.1410.1110.1310.1480.106
NightPSNR↑ 24.7824.3824.5725.3926.8126.4826.8725.2626.23
SSIM↑ 0.7260.7730.7870.8200.8510.8310.8480.8160.826
LPIPS↓ 0.2090.1850.1820.1390.1250.1120.1230.1420.111
", + "image_path": "2834ab7e804f67a3e48f994a78ae66bc59942b962fde84518dc14de3647a90e4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 272, + 481, + 516 + ], + "blocks": [ + { + "bbox": [ + 133, + 272, + 481, + 516 + ], + "lines": [ + { + "bbox": [ + 133, + 272, + 481, + 516 + ], + "spans": [ + { + "bbox": [ + 133, + 272, + 481, + 516 + ], + "type": "image", + "image_path": "331fa48fad1a23041aba7ee1afa5253c7288218d5e315086edca017ac6b088a6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 523, + 479, + 545 + ], + "lines": [ + { + "bbox": [ + 132, + 523, + 479, + 545 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 479, + 545 + ], + "type": "text", + "content": "Fig. 10. Comparison of methods on nighttime raindrop-focused images, covering both raindrop algorithms and restoration backbones." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 586, + 479, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 479, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 479, + 609 + ], + "type": "text", + "content": "Since the pairs/triplets of images are captured in quick succession, there are no problems with changes in illumination." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "content": "The captured images do not contain any personal information such as phone numbers, or human faces. We will release the dataset publicly with car license plates obscured, as we believe that our Raindrop Clarity dataset will benefit the progress of the raindrop removal field." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "text", + "content": "Y. Jin, X. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 300 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 300 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 300 + ], + "type": "image", + "image_path": "9f4e1ebd7e957d23f62ae571182b474f35f4ec8565594f6d54da0c0268d81c1d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 307, + 480, + 330 + ], + "lines": [ + { + "bbox": [ + 132, + 307, + 480, + 330 + ], + "spans": [ + { + "bbox": [ + 132, + 307, + 480, + 330 + ], + "type": "text", + "content": "Fig. 11. Comparison of methods on daytime background-focused images, covering both raindrop algorithms and restoration backbones." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 353, + 230, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 230, + 367 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 230, + 367 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 379, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 482, + 464 + ], + "type": "text", + "content": "Experimental Setup. Our benchmark includes two parts: Daytime and Night-time Raindrop Removal. As shown in Table 1, our daytime raindrop dataset contains a total of 5,442 paired/triplets images, where 4,713 paired images are used for the training set, and the remaining 729 paired images are regarded as the test set. Our nighttime raindrop dataset consists of 9,744 paired/triplets images, where 8,655 paired images are selected as the training set and the rest 1,089 paired images are used for evaluation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "text", + "content": "Baselines. To create the benchmark, we retrain existing state-of-the-art methods on our daytime and nighttime raindrop datasets. Our benchmark consists of five raindrop algorithms [21, 24-26, 40] and three restoration methods [22, 36, 46]. Specifically, IDT [40], RDiffusion [21], RainDS [26], Robot Car [24] and AtGAN [25] are raindrop algorithms, while Restormer [46], Uformer [36] and DiT [22] are restoration backbones. For a fair comparison, we conduct experiments using the original implementation of these methods. All experiments are implemented in PyTorch and are trained on four NVIDIA RTX A5000 GPUs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 581, + 481, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 481, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 481, + 604 + ], + "type": "text", + "content": "Comparison with Existing Datasets. The difference between our Raindrop Clarity dataset and the existing datasets is shown in Table. 1 and Fig. 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 640 + ], + "type": "text", + "content": "The stereo RobotCar dataset [24] includes 4,818 image pairs, randomly selected for training, validation, and testing. However, the raindrops in this dataset are disproportionately large and lack realism." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "The raindrop dataset proposed by Qian et al. [25] consists of 861 training pairs and 58 testing pairs. While it features common raindrop shapes, it lacks the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "type": "text", + "content": "Raindrop Clarity: Dual-Focused Dataset for Day Night Raindrop Removal" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 414 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 414 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 414 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 414 + ], + "type": "image", + "image_path": "ca0b3af94eff70f0db6b6db57dcccd8e2d0dab2ea7ae8bf2aad48a14a0a26c7f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 423, + 482, + 446 + ], + "lines": [ + { + "bbox": [ + 132, + 423, + 482, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 423, + 482, + 446 + ], + "type": "text", + "content": "Fig. 12. Comparison of methods on nighttime background-focused images, covering both raindrop algorithms and restoration backbones." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 471, + 481, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 481, + 495 + ], + "type": "text", + "content": "diversity of real-world scenes [44]. Also, the backgrounds between the raindrop images and the ground truth are not perfectly aligned [39]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 496, + 483, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 483, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 483, + 555 + ], + "type": "text", + "content": "The RainDS-Real dataset, proposed by Quan et al. [26], comprises 248 quadruplets, including raindrops, rain streaks, and their corresponding ground truths. The RainDS-Synthetic dataset has a domain gap with the real raindrop images. Moreover, the real-world raindrops are consistently small, obscuring only a fraction of the image details, making recovery less challenging." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 556, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 556, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 556, + 481, + 628 + ], + "type": "text", + "content": "The Windshield dataset, proposed by Soboleva and Shipitko [30], was captured by a camera attached to the vehicle during its movement in urban areas and highway environments. The Windshield does not provide the corresponding clean pairs, making this dataset not available for robust raindrop removal method testing. Additionally, frequent moving objects further complicate the dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Existing raindrop datasets, though valuable, have certain limitations. They do not include raindrop-focused and nighttime raindrops. In addition, we have noticed reflection errors in the datasets from Qian [25] and RainDS-Real [26]." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "text", + "content": "Y. Jin, X. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "Comparison with Synthetic Raindrops. In Fig. 13, we show the existing state-of-the-art synthetic physical model [23], although it can generate raindrops on blurry and clear backgrounds (Fig. 13 d generated on b, Fig. 13 e generated on c). It still has a gap with real-world collected raindrops. Real-world collected raindrops display diverse appearances, including various shapes such as round, elliptical, and long strips. Also, the transparency and glare effects of raindrops collected in the real-world varies." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 211, + 481, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 211, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 211, + 481, + 247 + ], + "type": "text", + "content": "Quantitative Comparison. To effectively evaluate the results on our proposed datasets, we use three popular metrics PSNR, SSIM [37], and LPIPS [48] in generation tasks [3-5, 9-16, 19, 20, 31-35, 42, 43]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 247, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 319 + ], + "type": "text", + "content": "The qualitative results are shown in Table 2. For the daytime dataset, the PSNR and SSIM of the original input images are 21.92 and 0.560. Raindrop algorithm IDT [40] achieves the best performance (26.05 PSNR of and 0.736 of SSIM) compared with other raindrop algorithms. For the restoration backbones, DiT [22] achieves a PSNR of 26.03 and a SSIM of 0.752. Restormer [46] achieves a PSNR of 26.08 and a SSIM of 0.748, which also performs competitively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 319, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 392 + ], + "type": "text", + "content": "It can be observed that, for the nighttime dataset, PSNR and SSIM of the input images are 24.78 and 0.726. Among the raindrop algorithms, IDT [40] achieves better performance than other methods, with a PSNR of 26.81 and a SSIM of 0.851. For the restoration backbones, Restormer [46] achieves the highest PSNR of 26.87, while DiT [22] achieves the lowest LPIPS of 0.111, which outperforms other backbones." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 402, + 482, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 473 + ], + "type": "text", + "content": "Qualitative Comparison. Figs. 9-12 show the qualitative results of state-of-the-art methods. To be specific, Fig. 9 and Fig. 11 show the results of daytime raindrop removal. The input images in Fig. 9 and Fig. 11 are daytime raindrop-focused and background-focused, respectively. Fig. 10 and Fig. 12 show the results of nighttime raindrop-focused and background-focused removal, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 569 + ], + "type": "text", + "content": "It can be found that, for daytime and nighttime raindrop-focused removal in Fig. 9 and Fig. 10, existing state-of-the-art methods fail to remove raindrops and recover the background. This is because most of them are designed for background-focused scenes and thus neglect the blurry problem. To our knowledge, Restormer [46] and Uformer [36] are specifically designed to address deblurring problem. However, they still struggle with recovering details and textures. As a result, these methods are struggling to simultaneously remove raindrops and recover the background." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "In Fig. 11, for daytime background-focused inputs, we can observe that existing state-of-the-art methods can handle most raindrops. However, they are struggling to address long strips of raindrops. This is because their methods may not detect long strips of raindrops and thus fail to remove them. Furthermore, in Fig. 12, we find that existing state-of-the-art methods are struggling to handle nighttime background-focused samples. To be specific, raindrops in these samples cannot be totally removed and details of the restored images are not clear. This is because nighttime raindrops usually reflect artificial light and thus" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 145, + 90, + 447, + 102 + ], + "type": "text", + "content": "Raindrop Clarity: Dual-Focused Dataset for Day Night Raindrop Removal" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 202, + 190 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 202, + 190 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 202, + 190 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 202, + 190 + ], + "type": "image", + "image_path": "dee761b1207cfd6fcf196b74c1496cceaaeac553810491ff83e01575d5da90bc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 191, + 199, + 198 + ], + "lines": [ + { + "bbox": [ + 136, + 191, + 199, + 198 + ], + "spans": [ + { + "bbox": [ + 136, + 191, + 199, + 198 + ], + "type": "text", + "content": "a Raindrop-focusedd " + }, + { + "bbox": [ + 136, + 191, + 199, + 198 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 204, + 116, + 270, + 190 + ], + "blocks": [ + { + "bbox": [ + 204, + 116, + 270, + 190 + ], + "lines": [ + { + "bbox": [ + 204, + 116, + 270, + 190 + ], + "spans": [ + { + "bbox": [ + 204, + 116, + 270, + 190 + ], + "type": "image", + "image_path": "ffacb94dee9a2664638cdb81258fb4a0bac29230b2eb3de4d5b0ae5516dadfb8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 191, + 270, + 198 + ], + "lines": [ + { + "bbox": [ + 206, + 191, + 270, + 198 + ], + "spans": [ + { + "bbox": [ + 206, + 191, + 270, + 198 + ], + "type": "text", + "content": "b Blurry background " + }, + { + "bbox": [ + 206, + 191, + 270, + 198 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{x}}" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 271, + 116, + 339, + 190 + ], + "blocks": [ + { + "bbox": [ + 271, + 116, + 339, + 190 + ], + "lines": [ + { + "bbox": [ + 271, + 116, + 339, + 190 + ], + "spans": [ + { + "bbox": [ + 271, + 116, + 339, + 190 + ], + "type": "image", + "image_path": "6086c7a2dec2d32060d361d0477405b2d20a3356077595350cafd12ece48e753.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 274, + 191, + 340, + 198 + ], + "lines": [ + { + "bbox": [ + 274, + 191, + 340, + 198 + ], + "spans": [ + { + "bbox": [ + 274, + 191, + 340, + 198 + ], + "type": "text", + "content": "c Clear background " + }, + { + "bbox": [ + 274, + 191, + 340, + 198 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_0" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 208, + 480, + 231 + ], + "lines": [ + { + "bbox": [ + 131, + 208, + 480, + 231 + ], + "spans": [ + { + "bbox": [ + 131, + 208, + 480, + 231 + ], + "type": "text", + "content": "Fig. 13. We show the differences between the synthetic raindrops [23] and real-world raindrops from Raindrop Clarity." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 342, + 118, + 411, + 190 + ], + "blocks": [ + { + "bbox": [ + 342, + 118, + 411, + 190 + ], + "lines": [ + { + "bbox": [ + 342, + 118, + 411, + 190 + ], + "spans": [ + { + "bbox": [ + 342, + 118, + 411, + 190 + ], + "type": "image", + "image_path": "c401688e2e5e8b4f030a28a47f5b357e392d44d889e71c79d06037b9a389285a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 191, + 405, + 198 + ], + "lines": [ + { + "bbox": [ + 345, + 191, + 405, + 198 + ], + "spans": [ + { + "bbox": [ + 345, + 191, + 405, + 198 + ], + "type": "text", + "content": "d Synthetic Rd on " + }, + { + "bbox": [ + 345, + 191, + 405, + 198 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 411, + 121, + 479, + 190 + ], + "blocks": [ + { + "bbox": [ + 411, + 121, + 479, + 190 + ], + "lines": [ + { + "bbox": [ + 411, + 121, + 479, + 190 + ], + "spans": [ + { + "bbox": [ + 411, + 121, + 479, + 190 + ], + "type": "image", + "image_path": "129be1559b6919f978fa7536daf88bb212d0b0aed5e525eaf03108b6cf71556d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 190, + 476, + 198 + ], + "lines": [ + { + "bbox": [ + 416, + 190, + 476, + 198 + ], + "spans": [ + { + "bbox": [ + 416, + 190, + 476, + 198 + ], + "type": "text", + "content": "e Synthetic Rd on " + }, + { + "bbox": [ + 416, + 190, + 476, + 198 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_0" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 255, + 480, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 480, + 304 + ], + "type": "text", + "content": "are more complex than daytime raindrops. Meanwhile, existing state-of-the-art methods are designed for daytime raindrop removal and neglect the effect of complex nighttime conditions. Thus, existing state-of-the-art methods cannot achieve promising results." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 324, + 219, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 324, + 219, + 337 + ], + "spans": [ + { + "bbox": [ + 132, + 324, + 219, + 337 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 353, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 482, + 485 + ], + "type": "text", + "content": "In this paper, we propose a new raindrop dataset, named Raindrop Clarity. It includes a total of 15,186 high-quality pairs/triplets of images with raindrops and the corresponding clear background images. Moreover, our dataset is collected under various conditions, including different scenes, times (daytime and nighttime), and focuses (raindrop- and background-focused). To be specific, 3,606 daytime raindrop-focused, 1,836 daytime background-focused, 4,838 nighttime raindrop-focused and 4,906 nighttime background-focused images are collected. With the proposed dataset, we re-analyze existing state-of-the-art methods, including raindrop algorithms and restoration backbones, and point out several open problems: raindrop-focus raindrop removal and nighttime raindrop removal. We wish to draw the attention of the community to the open problems." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 506, + 241, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 506, + 241, + 519 + ], + "spans": [ + { + "bbox": [ + 132, + 506, + 241, + 519 + ], + "type": "text", + "content": "6 Social Impacts" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Existing raindrop datasets focus on daytime background-focused raindrop removal. However, in the real world, raindrop images may focus on raindrops, resulting in a blurry background. Also, raindrops may occur in nighttime images. To draw the attention of the community to the open problems, we propose a new raindrop dataset, named Raindrop Clarity. It mainly collected from four conditions: daytime raindrop-focused, daytime background-focused, nighttime raindrop-focused and nighttime background-focused. Experimental results on our proposed dataset show that existing state-of-the-art methods cannot effectively remove raindrops under complex conditions. As a result, we believe that further exploration of our Raindrop Clarity dataset can help the community to address more complex raindrops." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "text", + "content": "Y. Jin, X. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 115, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 115, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 224 + ], + "type": "text", + "content": "This research is supported by the National Research Foundation, Singapore, under its AI Singapore Programme (AISG Award No: AISG2-PhD/2022-01-037[T]). Xin Li's research is supported by NSFC under Grant 623B2098. Malu Zhang's work is supported by the National Science Foundation of China under Grant 62106038, and in part by the Sichuan Science and Technology Program under Grant 2023YFG0259. We thank Robby T. Tan and Beibei Lin for their valuable discussions, revisions, and feedback." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 239, + 197, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 239, + 197, + 252 + ], + "spans": [ + { + "bbox": [ + 133, + 239, + 197, + 252 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 262, + 480, + 664 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 262, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 480, + 293 + ], + "type": "text", + "content": "1. [EB/OL], http://liuyunfei.net/Projs/PBRR/, title = PBRR: Physically Based Raindrop Rendering, author = Yunfei Liu, Zhixiang Hao, Shadi You, Yu Li, and Feng Lu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 295, + 480, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 295, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 138, + 295, + 480, + 327 + ], + "type": "text", + "content": "2. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12299-12310 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 328, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 328, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 328, + 480, + 369 + ], + "type": "text", + "content": "3. Chen, S., Ye, T., Bai, J., Chen, E., Shi, J., Zhu, L.: Sparse sampling transformer with uncertainty-driven ranking for unified removal of raindrops and rain streaks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13106-13117 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 372, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 372, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 372, + 480, + 403 + ], + "type": "text", + "content": "4. Chen, S., Ye, T., Liu, Y., Chen, E.: Snowformer: Context interaction transformer with scale-awareness for single image desnowing. arXiv preprint arXiv:2208.09703 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 404, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 404, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 404, + 480, + 446 + ], + "type": "text", + "content": "5. Chen, S., Ye, T., Liu, Y., Liao, T., Jiang, J., Chen, E., Chen, P.: Msp-former: Multi-scale projection transformer for single image desnowing. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 449, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 480, + 479 + ], + "type": "text", + "content": "6. Eigen, D., Krishnan, D., Fergus, R.: Restoring an image taken through a window covered with dirt or rain. In: Proceedings of the IEEE international conference on computer vision. pp. 633-640 (2013)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 480, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 480, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 480, + 480, + 511 + ], + "type": "text", + "content": "7. Halimeh, J.C., Roser, M.: Raindrop detection on car windshields using geometric-photometric environment construction and intensity-based correlation. In: 2009 IEEE Intelligent Vehicles Symposium. pp. 610-615. IEEE (2009)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 514, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 480, + 544 + ], + "type": "text", + "content": "8. Hao, Z., You, S., Li, Y., Li, K., Lu, F.: Learning from synthetic photorealistic raindrop for single image raindrop removal. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. pp. 0-0 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 545, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 586 + ], + "type": "text", + "content": "9. Jin, Y., Lin, B., Yan, W., Yuan, Y., Ye, W., Tan, R.T.: Enhancing visibility in nighttime haze images using guided apsf and gradient adaptive convolution. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 2446-2457 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 589, + 480, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 589, + 480, + 631 + ], + "spans": [ + { + "bbox": [ + 134, + 589, + 480, + 631 + ], + "type": "text", + "content": "10. Jin, Y., Sharma, A., Tan, R.T.: Dc-shadownet: Single-image hard and soft shadow removal using unsupervised domain-classifier guided network. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5027-5036 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 632, + 480, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 480, + 664 + ], + "type": "text", + "content": "11. Jin, Y., Yang, W., Tan, R.T.: Unsupervised night image enhancement: When layer decomposition meets light-effects suppression. In: European Conference on Computer Vision. pp. 404-421. Springer (2022)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "type": "text", + "content": "Raindrop Clarity: Dual-Focused Dataset for Day Night Raindrop Removal" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "12. Jin, Y., Ye, W., Yang, W., Yuan, Y., Tan, R.T.: Des3: Adaptive attention-driven self and soft shadow removal using vit similarity. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 2634-2642 (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 151, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 151, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 151, + 481, + 183 + ], + "type": "text", + "content": "13. Li, B., Li, X., Zhu, H., Jin, Y., Feng, R., Zhang, Z., Chen, Z.: Sed: Semantic-aware discriminator for image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "type": "text", + "content": "14. Li, X., Jin, X., Lin, J., Liu, S., Wu, Y., Yu, T., Zhou, W., Chen, Z.: Learning disentangled feature representation for hybrid-distorted image restoration. In: ECCV. pp. 313-329. Springer (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 481, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 261 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 261 + ], + "type": "text", + "content": "15. Li, X., Li, B., Jin, X., Lan, C., Chen, Z.: Learning distortion invariant representation for image restoration from a causality perspective. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1714-1724 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "type": "text", + "content": "16. Li, X., Ren, Y., Jin, X., Lan, C., Wang, X., Zeng, W., Wang, X., Chen, Z.: Diffusion models for image restoration and enhancement-a comprehensive survey. arXiv preprint arXiv:2308.09388 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 297, + 481, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 481, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 481, + 318 + ], + "type": "text", + "content": "17. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. arXiv preprint arXiv:2108.10257 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 319, + 481, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 481, + 362 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 481, + 362 + ], + "type": "text", + "content": "18. Liu, X., Suganuma, M., Sun, Z., Okatani, T.: Dual residual networks leveraging the potential of paired operations for image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7007-7016 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 363, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 481, + 418 + ], + "type": "text", + "content": "19. Luo, J., Ceylan, D., Yoon, J.S., Zhao, N., Philip, J., Frühstück, A., Li, W., Richardt, C., Wang, T.: Intrinsic diffusion: Joint intrinsic layers from latent diffusion models. In: ACM SIGGRAPH 2024 Conference Papers. SIGGRAPH '24, Association for Computing Machinery, New York, NY, USA (2024). https://doi.org/10.1145/3641519.3657472, https://doi.org/10.1145/3641519.3657472" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 419, + 481, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 481, + 452 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 481, + 452 + ], + "type": "text", + "content": "20. Luo, J., Zhao, N., Li, W., Richardt, C.: Crefnet: Learning consistent reflectance estimation with a decoder-sharing transformer. IEEE Transactions on Visualization and Computer Graphics (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 453, + 481, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 481, + 485 + ], + "type": "text", + "content": "21. Ozdenizci, O., Legenstein, R.: Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 486, + 481, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 481, + 507 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 481, + 507 + ], + "type": "text", + "content": "22. Peebles, W., Xie, S.: Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 509, + 481, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 481, + 541 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 481, + 541 + ], + "type": "text", + "content": "23. Pizzati, F., Cerri, P., de Charette, R.: Physics-informed guided disentanglement in generative networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 543, + 481, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 481, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 481, + 575 + ], + "type": "text", + "content": "24. Porav, H., Bruls, T., Newman, P.: I can see clearly now: Image restoration via deraining. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 7087-7093. IEEE (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 576, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 481, + 597 + ], + "type": "text", + "content": "25. Qian, R., Tan, R.T., Yang, W., Su, J., Liu, J.: Attentive generative adversarial network for raindrop removal from a single image pp. 2482-2491 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "text", + "content": "26. Quan, R., Yu, X., Liang, Y., Yang, Y.: Removing raindrops and rain streaks in one go. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9147-9156 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "27. Quan, Y., Deng, S., Chen, Y., Ji, H.: Deep learning for seeing through window with raindrops. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2463-2471 (2019)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "text", + "content": "Y. Jin, X. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 149 + ], + "type": "text", + "content": "28. Roser, M., Geiger, A.: Video-based raindrop detection for improved image registration. In: 2009 IEEE 12th International Conference on Computer Vision Workshops, ICCV Workshops. pp. 570-577. IEEE (2009)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "text", + "content": "29. Roser, M., Kurz, J., Geiger, A.: Realistic modeling of water droplets for monocular adherent raindrop recognition usingBezier curves. In: Asian conference on computer vision. pp. 235-244. Springer (2010)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 216 + ], + "type": "text", + "content": "30. Soboleva, V., Shipitko, O.: Raindrops on windshield: Dataset and lightweight gradient-based detection algorithm. In: 2021 IEEE Symposium Series on Computational Intelligence (SSCI). pp. 1-7. IEEE (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "text", + "content": "31. Wang, C., Pan, J., Lin, W., Dong, J., Wang, W., Wu, X.M.: Selfpromer: Self-prompt dehazing transformers with depth-consistency. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 5327-5335 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "text", + "content": "32. Wang, C., Pan, J., Wang, W., Dong, J., Wang, M., Ju, Y., Chen, J.: Promptstorer: A prompting image restoration method with degradation perception. Advances in Neural Information Processing Systems 36, 8898-8912 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 480, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 324 + ], + "type": "text", + "content": "33. Wang, C., Pan, J., Wang, W., Fu, G., Liang, S., Wang, M., Wu, X.M., Liu, J.: Correlation matching transformation transformers for uhd image restoration. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 5336-5344 (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 325, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 358 + ], + "type": "text", + "content": "34. Wang, J., Pan, Z., Zhang, M., Tan, R.T., Li, H.: Restoring speaking lips from occlusion for audio-visual speech recognition. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 19144-19152 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "type": "text", + "content": "35. Wang, J., Qian, X., Zhang, M., Tan, R.T., Li, H.: Seeing what you said: Talking face generation guided by a lip reading expert. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14653-14662 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "type": "text", + "content": "36. Wang, Z., Cun, X., Bao, J., Zhou, W., Liu, J., Li, H.: Uformer: A general u-shaped transformer for image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17683-17693 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "text", + "content": "37. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 457, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 490 + ], + "type": "text", + "content": "38. Wei, M., Shen, Y., Wang, Y., Xie, H., Wang, F.L.: Raindiffusion: When unsupervised learning meets diffusion models for real-world image deraining. arXiv preprint arXiv:2301.09430 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "type": "text", + "content": "39. Wen, Q., Wu, Y., Chen, Q.: Video waterfall removal via spatio-temporal fusion in driving scenes. arXiv preprint arXiv:2302.05916 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "text", + "content": "40. Xiao, J., Fu, X., Liu, A., Wu, F., Zha, Z.J.: Image de-raining transformer. IEEE Transactions on Pattern Analysis and Machine Intelligence pp. 1-18 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "text", + "content": "41. Yasarla, R., Patel, V.M.: Uncertainty guided multi-scale residual learning-using a cycle spinning cnn for single image de-raining. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8405-8414 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "text", + "content": "42. Ye, T., Chen, S., Bai, J., Shi, J., Xue, C., Jiang, J., Yin, J., Chen, E., Liu, Y.: Adverse weather removal with codebook priors. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12653-12664 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "43. Ye, T., Zhang, Y., Jiang, M., Chen, L., Liu, Y., Chen, S., Chen, E.: Perceiving and modeling density for image dehazing. In: European conference on computer vision. pp. 130-145. Springer (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "44. You, S., Tan, R.T., Kawakami, R., Ikeuchi, K.: Adherent raindrop detection and removal in video. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1035-1042 (2013)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 145, + 91, + 447, + 102 + ], + "type": "text", + "content": "Raindrop Clarity: Dual-Focused Dataset for Day Night Raindrop Removal" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 227 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "45. You, S., Tan, R.T., Kawakami, R., Mukaigawa, Y., Ikeuchi, K.: Adherent raindrop modeling, detection and removal in video. IEEE transactions on pattern analysis and machine intelligence 38(9), 1721-1733 (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "text", + "content": "46. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 204 + ], + "type": "text", + "content": "47. Zhang, K., Li, D., Luo, W., Ren, W.: Dual attention-in-attention model for joint rain streak and raindrop removal. IEEE Transactions on Image Processing 30, 7608-7619 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 227 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 227 + ], + "type": "text", + "content": "48. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "text", + "content": "Y. Jin, X. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Raising the Ceiling_ Conflict-Free Local Feature Matching with Dynamic View Switching/f888146a-2080-4739-ac33-923c8980cbb0_content_list.json b/2024/Raising the Ceiling_ Conflict-Free Local Feature Matching with Dynamic View Switching/f888146a-2080-4739-ac33-923c8980cbb0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b4490e72ccc9778196fab87bea957a1a2cd8e22b --- /dev/null +++ b/2024/Raising the Ceiling_ Conflict-Free Local Feature Matching with Dynamic View Switching/f888146a-2080-4739-ac33-923c8980cbb0_content_list.json @@ -0,0 +1,1800 @@ +[ + { + "type": "text", + "text": "Raising the Ceiling: Conflict-Free Local Feature Matching with Dynamic View Switching", + "text_level": 1, + "bbox": [ + 228, + 140, + 774, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaoyong Lu and Songlin Du\\*", + "bbox": [ + 390, + 212, + 609, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "School of Automation, Southeast University, Nanjing, China {luxiaoyong,sdu}@seu.edu.cn", + "bbox": [ + 299, + 239, + 700, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Current feature matching methods prioritize improving modeling capabilities to better align outputs with ground-truth matches, which are the theoretical upper bound on matching results, metaphorically depicted as the \"ceiling\". However, these enhancements fail to address the underlying issues that directly hinder ground-truth matches, including the scarcity of matchable points in small scale images, matching conflicts in dense methods, and the keypoint-repeatability reliance in sparse methods. We propose a novel feature matching method named RCM, which Raises the Ceiling of Matching from three aspects. 1) RCM introduces a dynamic view switching mechanism to address the scarcity of matchable points in source images by strategically switching image pairs. 2) RCM proposes a conflict-free coarse matching module, addressing matching conflicts in the target image through a many-to-one matching strategy. 3) By integrating the semi-sparse paradigm and the coarse-to-fine architecture, RCM preserves the benefits of both high efficiency and global search, mitigating the reliance on keypoint repeatability. As a result, RCM enables more matchable points in the source image to be matched in an exhaustive and conflict-free manner in the target image, leading to a substantial $260\\%$ increase in ground-truth matches. Comprehensive experiments show that RCM exhibits remarkable performance and efficiency in comparison to state-of-the-art methods.", + "bbox": [ + 261, + 297, + 738, + 587 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Feature Matching $\\cdot$ Camera Pose Estimation $\\cdot$ Transformer", + "bbox": [ + 261, + 601, + 738, + 614 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 637, + 375, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Feature matching, which aims to establish accurate correspondences between two images, underpins various pivotal tasks in computer vision, such as object recognition [25], structure from motion (SfM) [37], and simultaneous localization and mapping (SLAM) [13]. However, challenges arise from factors such as viewpoint change, illumination variation, scale variation, occlusion and motion blur between images. These complexities elevate the difficulty of feature matching, particularly when real-time performance is required for practical applications.", + "bbox": [ + 212, + 666, + 785, + 772 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Presently, methods for feature matching can be broadly categorized into two groups: sparse matching methods and dense matching methods. Sparse methods [5, 28, 36, 38] are designed to establish correspondences between two sets of", + "bbox": [ + 212, + 772, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 218, + 825, + 382, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b000a93906636589f5fefc3ff86b1534db7feba3276d062781b789ff99085c5d.jpg", + "image_caption": [ + "Fig.1: Comparison among RCM, $\\mathrm{RCM}_{\\mathrm{Lite}}$ , SuperGlue [36], LoFTR [39] and MatchFormer [42]. Full RCM significantly accelerates LoFTR by $2.2\\times$ while achieving superior performance. The lightweight $\\mathrm{RCM}_{\\mathrm{Lite}}$ outperforms SuperGlue by $+7.5\\%$ while maintaining a speed advantage. The same color indicates the matched features in the visualization." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 395, + 277 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bc961c5aeabb66e8a5b7385cc55b39ed1a6bd10ce8780e845444600d5a19cdf2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 143, + 784, + 266 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "keypoints. Consequently, the incorporation of detectors [9, 27, 31, 34] becomes essential in this process. In contrast, dense methods [20, 33, 39, 47] undertake a direct pursuit of matches between the dense features.", + "bbox": [ + 212, + 392, + 784, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Through a comprehensive examination of prevailing sparse and dense matching methods, we identify and subsequently address three key issues that constrain the theoretical upper bound, i.e., the ceiling, on the number of matches.", + "bbox": [ + 212, + 439, + 785, + 484 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) Sparse and dense methods: dramatic decrease in the number of matchable keypoints in small scale images. Overlapping regions in small scale images are limited in size, with the majority of keypoints or grid points situated in irrelevant background, which not only causes redundancy in the message passing but also radically limits the number of ground-truth matches.", + "For the semi-sparse [16, 17] and dense [39] matching paradigms, this challenge only arises in the source image as they perform a dense search in the target image. To tackle this challenge, we propose a dynamic view switcher that learns to switch the larger scale image to the source image. This strategic switch significantly augments the number of matchable keypoints within the overlapping region to radically raise the theoretical upper bound of the matcher, as illustrated in Fig. 6(b).", + "(2) Dense method: matching conflicts in large scale variation scenes caused by the one-to-one matching strategy. As shown in Fig. 2(b), the case of many-to-one matching arises when dealing with significant scale variations, as more pixels in large scale images correspond to fewer pixels in small scale images. However, existing semi-dense methods struggle with this challenge as they employ a one-to-one matching strategy in the coarse matching stage, generating at most one coarse match within each $8 \\times 8$ region in the target image while discarding other valid matches.", + "In our method, correspondences are established between the source and target image in a many-to-one manner. Each matchable point can independently search for correspondences without conflicts, as depicted in Fig. 2(c) and Fig. 6(a)." + ], + "bbox": [ + 212, + 486, + 787, + 842 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "X. Lu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/74a42b21438e3071e8987226b3e027aa81a2c3f70425473f87bc84d56f547560.jpg", + "image_caption": [ + "Fig. 2: Comparison of three matching methods. (a) The sparse methods rely on precise keypoint detection at corresponding points in both images. (b) We highlight the issues of matching conflicts introduced in Sec. 1. (c) RCM establishes conflict-free coarse matches and adjusts them independently in the fine matching stage." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 782, + 251 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(3) Sparse method: heavy reliance on the repeatability of keypoint. Sparse methods strategically narrow the search scope for matchers by leveraging detectors, yet they concurrently establish a severe reliance on the repeatability of keypoints. When input keypoints are not precisely detected at corresponding positions in both images and result in few ground-truth matches in challenging scenes, even a perfect sparse matcher fails to \"create something out of nothing\".", + "bbox": [ + 212, + 348, + 782, + 438 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This challenge inspires the semi-sparse paradigm [16,17] to detect keypoints in the source image and perform a dense search in the target image, eliminating the requirement for precise detection of keypoints in both images. We adopt the semi-sparse paradigm and integrate it seamlessly with the coarse-to-fine architecture [39], allowing RCM to maintain efficiency while globally searching for sub-pixel matches in the target image.", + "bbox": [ + 212, + 439, + 784, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To summarize, this paper presents the following contributions:", + "bbox": [ + 238, + 530, + 689, + 545 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We seamlessly integrate the semi-sparse paradigm with the coarse-to-fine architecture, establishing a strong foundation for our subsequent efforts to raise the theoretical upper bound of feature matching.", + "- We design a view switcher that dynamically switches the source and target images based on their scales. This view switcher offers a direct and effective method for increasing the number of matchable points in the source image.", + "- We propose a conflict-free coarse matching module, which matches two sets of features in a many-to-one fashion, further breaking through the match quantity barrier in large scale variation scenes.", + "- The proposed RCM and $\\mathrm{RCM}_{\\mathrm{Lite}}$ achieve an excellent performance-efficiency balance, rendering the matchers suitable for a wider range of applications." + ], + "bbox": [ + 225, + 554, + 784, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 753, + 395, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Local Feature Matching", + "text_level": 1, + "bbox": [ + 215, + 784, + 459, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Local feature matching methods are categorized as sparse and dense methods. Sparse methods involve the extraction of sparse features with a keypoint detec", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "RCM: Conflict-Free Local Feature Matching with Dynamic View Switching", + "bbox": [ + 230, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tor, followed by establishing correspondences through a matching process. Dense methods directly establish correspondences between dense features.", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sparse matching contains well-known methods like SIFT [27], SURF [3], BRIEF [4] and ORB [34], which are widely utilized. The advent of deep learning introduces learning-based descriptors as a dominant approach. Learning-based detectors and descriptors [9-11,18,31,46] harness the power of convolutional neural networks (CNN) for improved robustness. SuperGlue [36] pioneered global relationship modeling among sparse features using Transformer [41], which enriches features via self- and cross-attention mechanisms.", + "bbox": [ + 212, + 176, + 785, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dense matching methods [6,39,42] eschew the detector and instead establish dense correspondences between images, resulting in a significant improvement over sparse methods. We attribute this advancement to the increase in ground-truth matches facilitated by coarse-to-fine matching, which effectively covers the entire target image. This success in raising the feature matching ceiling inspires RCM to inherit the coarse-to-fine matching approach and pursue further enhancements in ground-truth matches.", + "bbox": [ + 212, + 284, + 785, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semi-Sparse matching methods, matching sparse and dense features asymmetrically, are proposed by S2DHM [16] and S2DNet [17]. These methods, while preserving sparse keypoints in the source image, use dense features at full image resolution in the target image, resulting in excessive memory usage. In our approach, we integrate the semi-sparse matching paradigm with coarse-to-fine matching [39] to maintain high efficiency. Additionally, the proposed dynamic view switching and many-to-one matching empower the semi-sparse paradigm to achieve performance comparable to modern sparse and dense paradigms.", + "bbox": [ + 212, + 391, + 785, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Feature Matching with Scale Invariance", + "text_level": 1, + "bbox": [ + 214, + 537, + 589, + 553 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Feature matching struggles in scenes with large scale variations, primarily due to the scarcity of matchable points in small scale images and matching conflicts due to one-to-one matching. Many methods are proposed to align scales before matching. OETR [7] identifies overlapping regions between images with object detection. ScaleNet [2] and Scale-Net [15] directly predict the scale variation. To address the scale challenge, we divide scenes with large scale variations into two cases: source and target images with (1) small and large scales and (2) large and small scales. In the first case, both sparse [5,36] and dense [39,42] methods face a reduction in matchable points in the source image, which is solved by the proposed dynamic view switcher. In the second case, dense methods encounter matching conflicts in small scale target images, a challenge mitigated by the proposed many-to-one matching strategy. Similar to many-to-one matching, AdaMatcher [19] addresses the matching conflict in the second case through adaptive assignment. However, it is limited to one-to-one matching in the first case, where the scarcity of matchable points in the source image persists. With the combination of dynamic view switcher and many-to-one matching, we can effectively obtain more matchable points in the source image and perform conflict-free matching in the target image in all scenes with large scale variations.", + "bbox": [ + 212, + 568, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "X. Lu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8192036d02f81b484985c511c3596d1adc80580b752026974d62e7beda6a1325.jpg", + "image_caption": [ + "Fig. 3: RCM is comprised of four primary modules: 1. The U-Net Extraction Module extracts multi-scale coarse and fine features. 2. The View Switching Module dynamically switches the larger scale image to the sparse branch, where dense coarse features are transformed into sparse features through the detection head. 3. The Conflict-Free Coarse Matching Module involves attention layers processing two sets of coarse features, which are subsequently matched via the many-to-one matching layer. 4. The Fine Matching Module further refines coarse matches based on fine features." + ], + "image_footnote": [], + "bbox": [ + 236, + 143, + 769, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 493, + 379, + 510 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Overview", + "text_level": 1, + "bbox": [ + 215, + 521, + 339, + 535 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall architecture of RCM is illustrated in Fig. 3. The U-Net extraction module initiates the process by employing CNN to extract coarse and fine features from both images. Subsequently, the coarse features of the two images undergo processing within the view switcher. In this module, the coarse features of larger scale images are switched to the sparse branch and extracted by the detection head as sparse features. And the coarse features of smaller scale images are retained as dense features. The sparse and dense features are then matched in a many-to-one fashion through the conflict-free coarse matching module. Finally the fine matching module crops out the fine features in the coarse matching positions and refines coarse matches through the correlation-based approach [39].", + "bbox": [ + 212, + 542, + 787, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Local Feature Extraction", + "text_level": 1, + "bbox": [ + 215, + 712, + 468, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We extend the VGG-style encoder from SuperPoint [9] to a U-Net architecture to extract multi-scale features from images $I_{A}$ and $I_{B}$ . The decoder output is employed as fine features, denoted as $F_{A}^{f}\\in \\mathbb{R}^{H_{f}\\times W_{f}\\times C_{2}}$ and $F_{B}^{f}\\in \\mathbb{R}^{H_{f}\\times W_{f}\\times C_{2}}$ . For brevity, we omit the $A$ and $B$ subscripts in the rest of the paper. On the other hand, the coarse features, denoted as $F^c\\in \\mathbb{R}^{H_c\\times W_c\\times C_1}$ , are formed by merging features from multiple resolutions, enhancing the robustness of the conflict-free coarse matching module.", + "bbox": [ + 212, + 733, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "RCM: Conflict-Free Local Feature Matching with Dynamic View Switching", + "bbox": [ + 230, + 113, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 View Switching Module", + "text_level": 1, + "bbox": [ + 215, + 146, + 459, + 161 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Fig. 6(b), employing a small scale image as the source image leads to a scarcity of matchable points, which consequently leads to fewer matches. As shown in Fig. 4, we propose a binary classification network named view switcher to determine whether to switch the source and target images, where a lightweight CNN is employed to process the correlation map of coarse features $F_{A}^{c}$ and $F_{B}^{c}$ to perceive scale variation. The result of view switcher $VS$ is computed as", + "bbox": [ + 212, + 172, + 460, + 400 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/65ed5ebf5c532371f7409796faec5700437d297da6b9264eddb9912fbc34ba0d.jpg", + "image_caption": [ + "Fig. 4: The view switcher perceives the scale variation between image pairs to predict a binary classification result, indicating whether to switch the images." + ], + "image_footnote": [], + "bbox": [ + 478, + 204, + 781, + 314 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nV S = \\operatorname {S o f t m a x} (\\operatorname {C N N} (\\operatorname {C o r r} (\\operatorname {P o o l} (F _ {A} ^ {c}), \\operatorname {P o o l} (F _ {B} ^ {c})))) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 320, + 414, + 785, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathrm{Pool}(\\cdot)$ represents average pooling and $\\mathrm{Corr}(\\cdot, \\cdot)$ signifies similarity calculation through the inner product. The view switcher is trained to keep sparse branch processing larger scale images and dense branch processing smaller scale images. In the sparse branch, we employ the detection head in [9] to detect keypoints and sample the feature map at the keypoint positions as sparse features $F_{s}^{c}$ . The U-Net encoder and detection head are inherited from SuperPoint [9] and are frozen, making the detection equivalent to SuperPoint. Assuming that $N$ keypoints are detected, features in the sparse branch are reduced from $\\mathbb{R}^{H_c \\times W_c \\times C_1}$ to $\\mathbb{R}^{N \\times C_1}$ .", + "bbox": [ + 212, + 445, + 787, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4 Conflict-Free Coarse Matching", + "text_level": 1, + "bbox": [ + 215, + 604, + 516, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Position Encoder Given that the attention mechanism lacks awareness of the positions of the features, we embed positional information into the coarse features $F^c$ through a position encoder PE, in which a lightweight MLP progressively expands the positional information $P \\in \\mathbb{R}^{N \\times 3}$ into the coarse feature dimension $C_1$ and updates the coarse features. $P$ encompasses the $x$ and $y$ coordinates of keypoints, and their confidence scores $s$ generated by the detection head.", + "bbox": [ + 212, + 632, + 787, + 723 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Attention Layer The self- and cross-attention layers are employed for the global enhancement of sparse and dense features. In contrast to dense methods that use linear attention [21] for efficiency, the semi-sparse matching paradigm allows us to adopt vanilla attention [41], which is denoted as", + "bbox": [ + 212, + 747, + 787, + 808 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nM = \\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {S o f t m a x} \\left(Q K ^ {T}\\right) V. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 821, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "X. Lu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The attentional message is denoted as $M$ . In self-attention, $Q, K, V$ are three linear projections of a single feature set. In cross-attention, $Q$ is projected from one feature set, while $K$ and $V$ are projected from another feature set.", + "bbox": [ + 212, + 146, + 784, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The feed-forward network $\\mathrm{MLP_f}$ processes the attentional message $M$ to update the features as", + "bbox": [ + 212, + 191, + 785, + 222 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nF ^ {i + 1} = F ^ {i} + \\mathrm {M L P} _ {\\mathrm {f}} \\left(F ^ {i} | | M\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 250, + 785, + 268 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $F^i$ is the feature in layer $i$ and $||$ is the concatenation operation. We perform self- and cross-attention on sparse and dense features for $L_1$ iterations, resulting in enhanced features $\\tilde{F}_s^c$ and $\\tilde{F}_d^c$ that capture global information.", + "bbox": [ + 212, + 279, + 785, + 327 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Many-to-One Coarse Matching with Dustbin With globally enhanced features, we perform matching between sparse and dense features in a many-to-one manner. Initially, we utilize a set of learnable parameters as a dustbin, denoted as $bin \\in \\mathbb{R}^{1 \\times C_1}$ , and concatenate it with the dense features $\\hat{F}_d^c = \\tilde{F}_d^c ||bin \\in \\mathbb{R}^{(H_cW_c + 1) \\times C_1}$ . The dustbin is introduced to match with non-matchable sparse features, enabling our matcher to handle non-overlapping regions and occlusion. We compute the score matrix with dustbin $S_{bin}$ between the sparse feature $\\tilde{F}_s^c \\in \\mathbb{R}^{N \\times C_1}$ and dense feature with dustbin $\\hat{F}_d^c \\in \\mathbb{R}^{(H_cW_c + 1) \\times C_1}$ as", + "bbox": [ + 212, + 347, + 787, + 470 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nS _ {b i n} = \\tilde {F} _ {s} ^ {c} \\hat {F} _ {d} ^ {c} / \\tau \\in \\mathbb {R} ^ {N \\times (H _ {c} W _ {c} + 1)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 479, + 785, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\tau$ is a learnable temperature parameter.", + "bbox": [ + 212, + 510, + 552, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Current dense methods utilize the Dual-Softmax operator to compute the coarse matching probability $P_{c}$ , defined as $P_{c} = \\mathrm{Softmax}_{row}(S) \\cdot \\mathrm{Softmax}_{col}(S) \\in \\mathbb{R}^{N \\times H_{c}W_{c}}$ . The Dual-Softmax operator keeps the summation of rows and columns of $P_{c}$ less than or equal to 1. This limits each feature to match at most one feature, that is, one-to-one matching. However multiple points may correspond to one point in scenes with large scale variations, which leads to matching conflicts as shown in Fig. 2(b). To address this issue, a single Softmax operation on the score with dustbin $S_{bin}$ is employed to release the constraint on target images while maintaining the constraint on source images as", + "bbox": [ + 212, + 526, + 787, + 662 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nP _ {c} = \\operatorname {S o f t m a x} _ {r o w} \\left(S _ {b i n}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 418, + 676, + 785, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where each sparse feature can only match one dense feature or the dustbin, while one dense feature can match multiple sparse features. The strategy of many-to-one matching with dustbin can be comprehended as independently assigning each sparse feature to either its corresponding dense feature or to the dustbin. As shown in Fig. 2(c), following conflict-free coarse matching, the positions of sparse features, which correspond to the same dense feature, are independently adjusted during the fine matching stage.", + "bbox": [ + 212, + 704, + 787, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ultimately, features matched to the dustbin are discarded, and we proceed to select matches with matching probabilities surpassing the threshold $\\theta_c$ .", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "RCM: Conflict-Free Local Feature Matching with Dynamic View Switching", + "bbox": [ + 228, + 113, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.5 Fine Matching", + "text_level": 1, + "bbox": [ + 215, + 146, + 383, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For each coarse match, we sample a single feature at the source fine feature map and crop a feature window of size $w \\times w$ at the target fine feature map. Given the collection of $K$ coarse matches, we obtain fine features denoted as $\\hat{F}_s^f \\in \\mathbb{R}^{K \\times 1 \\times C_2}$ and $\\hat{F}_d^f \\in \\mathbb{R}^{K \\times w^2 \\times C_2}$ for the sparse and dense branches, respectively.", + "bbox": [ + 212, + 167, + 782, + 228 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Subsequently, $\\hat{F}_s^f$ and $\\hat{F}_d^f$ are processed by the cross-attention module for $L_{2}$ iterations. Inspired by [6], in order to capture the local context of dense features, we substitute the self-attention module with the integration of $3\\times 3$ convolutions into the feed-forward network. For fine matching, the features are updated by", + "bbox": [ + 212, + 229, + 782, + 291 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nF ^ {i + 1} = F ^ {i} + \\operatorname {C o n v} _ {3} \\left(F ^ {i} | | M\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 299, + 784, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compute the correlation map between the centroid of the source feature window and all the features in the target feature window, which represents the matching probability, and then get the fine matching position by computing the expectation of the probability distribution, as introduced in [39].", + "bbox": [ + 212, + 321, + 782, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.6 Supervision", + "text_level": 1, + "bbox": [ + 215, + 400, + 359, + 416 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our loss function encompasses four components: the coarse matching loss $\\mathcal{L}_c^m$ , the coarse dustbin loss $\\mathcal{L}_c^d$ , the fine matching loss $\\mathcal{L}_f$ , and the view switching loss $\\mathcal{L}_{vs}$ . Ground-truth matches $M_{gt} = \\{(i_m,j_m)\\}$ are generated from camera poses and depth maps, establishing many-to-one correspondence between sparse and dense features. And non-matchable sparse features $U_{gt} = \\{i_u\\}$ , which are keypoints without any ground-truth match, are also obtained.", + "bbox": [ + 212, + 420, + 782, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The coarse matching loss $\\mathcal{L}_c^m$ is computed as the negative log-likelihood loss of the ground-truth matches $M_{gt}$ over the matching probability matrix $P_c$ of each attention layer,", + "bbox": [ + 212, + 511, + 784, + 556 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} ^ {m} = - \\frac {1}{L} \\sum_ {l} \\left[ \\frac {1}{| M _ {g t} |} \\sum_ {(i, j) \\in M _ {g t}} \\log P _ {c} ^ {l} (i, j) \\right]. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 561, + 784, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The coarse dustbin loss $\\mathcal{L}_c^d$ refers to the negative log-likelihood loss of the non-matchable sparse features $U_{gt}$ over the last column of the matching probability matrix $P_c$ , which signifies the matching probabilities with the dustbin,", + "bbox": [ + 212, + 604, + 782, + 651 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} ^ {d} = - \\frac {1}{| U _ {g t} |} \\sum_ {i \\in U _ {g t}} \\log P _ {c} (i, H _ {c} W _ {c} + 1). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 656, + 784, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The fine matching loss $\\mathcal{L}_f$ is computed by the L2-distance between each refined position and the ground-truth position, as introduced in [39].", + "bbox": [ + 212, + 696, + 782, + 727 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To supervise the view switcher, we randomly select 500 ground-truth matches in each image pair and compute the distance between points in each image. The image with the larger average distance is identified as the larger scale image. If the original source image is the larger scale image, the ground-truth $VS_{gt}$ is false, and vice versa. The binary cross-entropy loss of the switcher result $VS$ is subsequently computed as", + "bbox": [ + 212, + 728, + 782, + 816 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v s} = - \\left[ V S _ {g t} \\log (V S) + \\left(1 - V S _ {g t}\\right) \\log (1 - V S) \\right]. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 825, + 784, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "X. Lu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fa3ddd7a904c7d6a2c796c4c3f7c0ae5b1a83ce93f2768ebc3b12fdcef5cc839.jpg", + "table_caption": [ + "Table 1: Evaluation on HPatches [1] for homography estimation. We report the percentage of correctly estimated homographies with mean corner errors below $1/3/5$ pixels. We mark the best accuracy in bold." + ], + "table_footnote": [], + "table_body": "
CategoryMethodOverallIlluminationViewpoint
Accuracy (%,ε < 1/3/5 px)
SparseSuperPoint [9] + NN0.46 / 0.78 / 0.850.57 / 0.92 / 0.970.35 / 0.65 / 0.74
D2Net [10] + NN0.38 / 0.72 / 0.810.65 / 0.95 / 0.980.13 / 0.51 / 0.65
R2D2 [31] + NN0.47 / 0.78 / 0.830.63 / 0.93 / 0.980.33 / 0.64 / 0.70
ASLFeat [29] + NN0.48 / 0.81 / 0.880.63 / 0.94 / 0.980.34 / 0.69 / 0.78
SuperPoint + SuperGlue [36]0.51 / 0.83 / 0.890.62 / 0.93 / 0.980.41 / 0.73 / 0.81
SuperPoint + CAPS [43] + NN0.49 / 0.79 / 0.860.62 / 0.93 / 0.980.36 / 0.65 / 0.75
SIFT + CAPS [43] + NN0.36 / 0.76 / 0.850.48 / 0.89 / 0.950.26 / 0.65 / 0.76
SuperPoint + LightGlue [24]0.39 / 0.85 / 0.910.52 / 0.96 / 0.980.28 / 0.75 / 0.84
(Semi-)DenseSparseNCNet [32]0.36 / 0.66 / 0.760.62 / 0.92 / 0.970.13 / 0.41 / 0.57
NCNet [33]0.48 / 0.61 / 0.710.98 / 0.98 / 0.980.02 / 0.28 / 0.46
Patch2Pix [47]0.50 / 0.79 / 0.870.71 / 0.95 / 0.980.30 / 0.64 / 0.76
LoFTR [39]0.55 / 0.81 / 0.860.74 / 0.95 / 0.980.38 / 0.69 / 0.76
MatchFormer [42]0.55 / 0.81 / 0.870.75 / 0.95 / 0.980.37 / 0.68 / 0.78
Semi-SparseRCMLite0.50 / 0.86 / 0.910.55 / 0.95 / 0.980.45 / 0.79 / 0.85
RCM0.56 / 0.88 / 0.930.65 / 0.97 / 0.990.47 / 0.80 / 0.88
", + "bbox": [ + 220, + 198, + 779, + 388 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 417, + 375, + 434 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 452, + 457, + 468 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Both RCM and $\\mathrm{RCM}_{\\mathrm{Lite}}$ consist of $L_{1} = 5$ layers of coarse attention and $L_{2} = 2$ layers of fine attention. The dimensions for coarse and fine features are $C_1 = 256$ and $C_2 = 64$ , respectively. The key distinction between RCM and $\\mathrm{RCM}_{\\mathrm{Lite}}$ lies in their coarse feature resolution and fine matching window size. For RCM, the coarse feature resolution is $1/8$ , and the fine matching window size is $w = 5$ . For $\\mathrm{RCM}_{\\mathrm{Lite}}$ , the coarse feature resolution is $1/16$ , and the fine matching window size is $w = 9$ . The learnable temperature $\\tau$ is initialized to 0.1, and the coarse matching threshold $\\theta_c$ is set to 0.2. The proposed method is implemented using Pytorch [30]. During the training phase, images are resized and padded to the size of $832 \\times 832$ . We detect 1024 keypoints in the source image, and if there are insufficient keypoints, random keypoints are added to ensure efficient batching. RCM and $\\mathrm{RCM}_{\\mathrm{Lite}}$ are trained on the MegaDepth dataset [23] for 30 epochs with a batch size of 4 by AdamW optimizer [26] and are not fine-tuned for any other tasks. The initial learning rate is set to 0.0002, and the training employs the cosine decay learning rate scheduler with 1 epoch of linear warm-up. More details can be found in the Supplementary Material.", + "bbox": [ + 212, + 481, + 787, + 726 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Homography Estimation", + "text_level": 1, + "bbox": [ + 215, + 750, + 465, + 766 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Dataset. The HPatches [1] dataset is widely recognized for evaluating local features and matchers. In line with the evaluation protocol in [10], we assess the performance of our method across 108 HPatches sequences, which comprise 52 sequences with illumination variations and 56 sequences with viewpoint changes.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "RCM: Conflict-Free Local Feature Matching with Dynamic View Switching", + "bbox": [ + 230, + 113, + 730, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Metric. Following the metrics employed in [9, 36, 47], we report the percentage of correctly estimated homographies with mean corner errors below $1/3/5$ pixels. Vanilla RANSAC [14] is applied to estimate the homography for all methods.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results. As shown in Tab. 1, RCM achieves the best overall performance at all thresholds. Comparing the dense methods, RCM shows a significant advantage in accuracy in large viewpoint change scenarios, leading LoFTR and MatchFormer by $(+9\\%, +11\\%, +12\\%)$ and $(+10\\%, +12\\%, +10\\%)$ , respectively. We attribute this performance improvement to the increased matchable points in the source image facilitated by the view switcher, and the conflict-free matching achieved in the target image through many-to-one matching. Compared to the sparse method SuperGlue, RCM achieves an overall performance lead of $(+5\\%, +5\\%, +4\\%)$ , benefiting from the dense search due to the semi-sparse paradigm and the sub-pixel accuracy due to the coarse-to-fine architecture.", + "bbox": [ + 212, + 191, + 784, + 343 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/eee4d8d385a176815a66111f9875d486270dfc991605787f40dd8b240ef97331.jpg", + "table_caption": [ + "Table 2: Two-view pose estimation results on MegaDepth [23] in outdoor scenes." + ], + "table_footnote": [], + "table_body": "
CategoryMethodPose estimation
@5°@10°@20°
SparseSuperGlue [36]42.259.073.6
SGMNet [5]40.559.073.6
ClusterGNN [38]44.258.570.3
LightGlue [24]46.464.378.2
DeDoDe-B [11]49.165.077.1
DeDoDe-G [11]51.268.481.2
DenseDRC-Net [22]27.042.958.3
PDC-Net+(H) [40]43.161.976.1
LoFTR [39]52.869.281.2
RoMa [12]63.076.986.4
Semi-SparseRCMLite49.766.579.3
RCM53.269.481.5
", + "bbox": [ + 220, + 412, + 493, + 560 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/bc2a4a2adbc1b915ae4e218dbb40ed25b913b1284812622c67b74c6ae8a39b0c.jpg", + "table_caption": [ + "Table 3: Two-view pose estimation results on ScanNet [8] in indoor scenes." + ], + "table_footnote": [], + "table_body": "
CategoryMethodPose estimationAUC
@5°@10°@20°
SparseOANet [44]11.826.943.9
SuperGlue [36]16.233.851.8
SGMNet [5]15.432.148.3
LightGlue [24]*13.929.645.7
DenseDRC-Net [22]*7.717.930.5
MatchFormer [42]*15.832.048.0
LoFTR-OT [39]*16.933.650.6
Semi-SparseRCMLite*15.231.548.8
RCM*17.334.652.1
", + "bbox": [ + 511, + 426, + 777, + 545 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Relative Pose Estimation", + "text_level": 1, + "bbox": [ + 215, + 614, + 470, + 628 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Dataset. We evaluate matchers on the MegaDepth [23] and the ScanNet [8] dataset for outdoor and indoor pose estimation, respectively. For the MegaDepth dataset, all test images are resized such that their longer dimensions are 1216. For the ScanNet dataset, all test images are resized to $640 \\times 480$ pixels.", + "bbox": [ + 212, + 642, + 782, + 702 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Metric. Based on the matching results, we apply vanilla RANSAC to solve the fundamental matrices and relative poses of the image pairs. The AUC of the pose error at thresholds $(5^{\\circ}, 10^{\\circ}, 20^{\\circ})$ are reported, where pose error is the maximum angular error in rotation and translation.", + "bbox": [ + 212, + 703, + 784, + 762 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results. As shown in Tab. 2, RCM achieves the second best performance on the MegaDepth dataset. RCM outperforms the state-of-the-art method LoFTR, demonstrating the benefits of the view switcher and many-to-one matching. The lightweight $\\mathrm{RCM}_{\\mathrm{Lite}}$ achieves a performance boost $(+7.5\\%, +7.5\\%, +5.7\\%)$ compared to SuperGlue. Fig. 5 illustrates qualitative comparisons of the SuperGlue,", + "bbox": [ + 212, + 763, + 784, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "X. Lu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/9148033fa40eeb7596386c3b80ff5c80d3ba8236b4df17b55762abea518745a8.jpg", + "image_caption": [ + "Fig. 5: Qualitative comparison of SuperGlue [36], LoFTR [39] and proposed RCM on MegaDepth and ScanNet." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 782, + 398 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "LoFTR and RCM. More visualizations and analyses can be found in the Supplementary Material.", + "bbox": [ + 212, + 465, + 782, + 494 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As shown in Tab. 3, RCM achieves a $(0.4\\%, 1\\%, 1.5\\%)$ performance improvement over LoFTR in the ScanNet dataset. It's worth noting that the superscript * indicates zero-shot performance, and the model is trained on MegaDepth. Nevertheless, our methods achieve performance comparable to state-of-the-art sparse methods trained on ScanNet. Dense and semi-sparse matching methods find matches in the target image exhaustively, mitigating the challenge of nonrepeatability in indoor scenes, a key limitation of sparse methods.", + "bbox": [ + 212, + 496, + 784, + 602 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Visual Localization", + "text_level": 1, + "bbox": [ + 215, + 621, + 419, + 635 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Dataset. We assess the performance of our method using the Aachen DayNight v1.1 benchmark [45], a well-established dataset for visual localization. This dataset comprises 824 day-time images and 191 night-time images, chosen as queries for outdoor visual localization.", + "bbox": [ + 212, + 643, + 782, + 702 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Metric. We utilize the open-source localization pipeline HLoc [35] for conducting localization with RCM. We report the percentage of successfully localized images within three error thresholds $(0.25m, 2^{\\circ}) / (0.5m, 5^{\\circ}) / (1m, 10^{\\circ})$ .", + "bbox": [ + 212, + 704, + 782, + 748 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results. As shown in Tab. 4, RCM demonstrates comparable or superior performance compared to the dense methods LoFTR and ASpanFormer in outdoor visual localization tasks. Our approach exhibits strong generalization across various downstream tasks without the need for fine-tuning. It is noteworthy that the semi-sparse method RCM is significantly more efficient than the dense methods, achieving a runtime speedup of over twofold.", + "bbox": [ + 212, + 750, + 784, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "RCM: Conflict-Free Local Feature Matching with Dynamic View Switching", + "bbox": [ + 230, + 113, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/f142fa070020cb597e3d4a378fae261f407e54fc30296cc0ec6be5d6c031e726.jpg", + "table_caption": [ + "Table 4: Visual localization on the Aachen Day-Night benchmark v1.1 [45]." + ], + "table_footnote": [], + "table_body": "
MethodDayNight
(0.25m, 2°) / (0.5m, 5°) / (1m, 10°)
Localization with matching pairs generated by HLoc
LoFTR [39]88.7 / 95.6 / 99.078.5 / 90.6 / 99.0
ASpanFormer [6]89.4 / 95.6 / 99.077.5 / 91.6 / 99.0
DeDoDe-B [11]87.4 / 94.7 / 98.570.7 / 88.0 / 97.9
DeDoDe-G [11]88.5 / 95.4 / 99.074.9 / 89.5 / 99.5
SP [9]+LightGlue [24]89.6 / 95.8 / 99.272.8 / 88.0 / 99.0
RCMLite87.7 / 94.7 / 98.472.3 / 91.1 / 98.4
RCM89.7 / 96.0 / 98.772.8 / 91.6 / 99.0
", + "bbox": [ + 305, + 170, + 697, + 292 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f6616bcbb4d0320346338474f8bfffc51e74ef86b5a8a3e62d1f2d96c8f012fc.jpg", + "image_caption": [ + "Fig. 6: Ablation study on view switcher and many-to-one matching. (a) (b) shows the actual matching results and (c) (d) shows the ground-truth coarse matches. The harmonious combination of view switcher and many-to-one matching allows more matchable points in the source image and conflict-free matching in the target image." + ], + "image_footnote": [], + "bbox": [ + 222, + 304, + 527, + 503 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f1afd91a353ab4d556a8cece7cd08f2ad21cb1064bdf954fc6ae8e64438aad18.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 303, + 784, + 502 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 599, + 388, + 614 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this subsection, we conduct qualitative and quantitative ablation experiments on MegaDepth. All models in Tab. 5 are trained for 15 epochs at 544 resolution and tested at 1216 resolution. For the baseline model (1), we apply the original SuperPoint [9] to generate keypoints and descriptors for both images. O2O and M2O in Tab. 5 are abbreviations for one-to-one and many-to-one, respectively.", + "bbox": [ + 212, + 627, + 787, + 702 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ground-Truth Match Analysis. Fig. 6(d) illustrates the core idea of RCM, which is to increase the theoretical upper bound on the matcher, i.e., ground-truth matches, through view switcher and many-to-one matching. Fig. 6(c) illustrates a quantitative comparison of ground-truth matches, showing an increase as the view switching accuracy improves during training. Eventually, RCM and $\\mathrm{RCM}_{\\mathrm{Lite}}$ achieve view switching accuracies of $90.9\\%$ and $87.8\\%$ , respectively. Through the integration of many-to-one matching and the view switcher, the ground-truth matches surge to $260\\%$ compared to one-to-one matching, which is a compelling demonstration of raising the ceiling of feature matching.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "X. Lu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/a2265a2cc9d474b98d9135c6d2e6279bf231566af98ce5b9ed104cdeeb513131.jpg", + "table_caption": [ + "Table 5: Ablation study of each proposal on MegaDepth." + ], + "table_footnote": [], + "table_body": "
Method#MatchesPrec.
(1) Sparse sampling451.798.0
(2) Dense sampling932.193.5
(3) Semi-Sparse sampling492.393.7
(4) + U-Net Extraction507.894.2
(5) O2O → M2O770.7(+51.8%)97.2
(6) + Switcher (RCMLite)951.4(+23.4%)97.8
(7) 1/16 → 1/8 (RCM)961.498.1
", + "bbox": [ + 222, + 188, + 493, + 306 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/f7b405c952bfe8438fc2a9cfd97dc4076d9f480b9f0bb12f864ff47000185062.jpg", + "table_caption": [ + "Table 6: Evaluation on extreme-scale MegaDepth." + ], + "table_footnote": [], + "table_body": "
Method#MatchesPrec.
SP [9] + SuperGlue [36]26175.5
SP [9] + LightGlue [24]31376.0
LoFTR [39]42453.5
MatchFormer [42]46477.0
ASpanFormer [6]42577.1
RCM88778.9
wo switcher65375.4
flipped switcher7654.2
", + "bbox": [ + 535, + 184, + 754, + 310 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Sampling Strategy. As shown in Tab. 5, the baseline models (1), (2), and (3) differ solely in sampling strategies for coarse features. Precisely, (1) detects keypoints in both images, (2) employs dense feature as the coarse feature, and (3) detects keypoints in source image and employs dense feature in target image.", + "bbox": [ + 212, + 348, + 782, + 411 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Many-to-One Matching. As shown in Fig. 6(a), in contrast to the one-to-one matching strategy, the many-to-one matching strategy pursues conflict-free matching, resulting in more detailed matches. As shown in Tab. 5(5), substituting one-to-one matching with many-to-one matching improves the number of matches by $51.8\\%$ and precision by $3\\%$ .", + "bbox": [ + 212, + 414, + 784, + 491 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "View Switcher. As shown in Fig. 6(b), without the view switcher, the matcher can only produce a small number of matches in small scale images because only a limited number of keypoints are detected in the overlapping region. The view switcher enables the detection of more matchable keypoints in the source image. As shown in Tab. 5(6), the average number of matches increases by $23.4\\%$ with the view switcher. More analyses about view switcher and many-to-one matching can be found in the Supplementary Material.", + "bbox": [ + 212, + 494, + 787, + 602 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "RCM and $\\mathrm{RCM}_{\\mathrm{Lite}}$ . The distinctions between RCM and $\\mathrm{RCM}_{\\mathrm{Lite}}$ lie in the coarse feature resolution and fine matching window. RCM employs coarse features at 1/8 resolution, consistent with the dense attention-based methods. We apply 1/16 resolution coarse features in $\\mathrm{RCM}_{\\mathrm{Lite}}$ to trade performance degradation for faster runtime, facilitating applications in real-time scenarios. Because of the additional layers at 1/16 resolution in the feature extraction network, $\\mathrm{RCM}_{\\mathrm{Lite}}$ consists of more parameters than RCM. Also to cover larger coarse feature patches, we design a larger fine matching window for $\\mathrm{RCM}_{\\mathrm{Lite}}$ .", + "bbox": [ + 212, + 604, + 787, + 728 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Robustness to Scale Variation. We evaluate matchers on the MegaDepth subset defined in OETR [7], where the overlap scale ratio is in $[4, +\\infty]$ . As shown in Tab. 6, RCM achieves the best performance with $2.8\\times$ and $2.1\\times$ more matches than LightGlue and ASpanFormer respectively. The performance of RCM is severely degraded without the view switcher, demonstrating the significance of the view switcher in large scale variation scenarios. Flipping the switching result, i.e., actively taking small scale image as source image, leads to the worst results.", + "bbox": [ + 212, + 731, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "RCM: Conflict-Free Local Feature Matching with Dynamic View Switching", + "bbox": [ + 230, + 113, + 730, + 130 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 127 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Efficiency. The parameters, runtimes, and performances of state-of-the-art matchers are presented in Tab. 7. Notably, RCM and $\\mathrm{RCM}_{\\mathrm{Lite}}$ achieve the optimal balance between performance and efficiency. Although the performance of RCM lags behind the most precise matcher", + "bbox": [ + 215, + 146, + 431, + 296 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/d8daae1f4bb59da5206070271f90084477d9aa665252581b2005fc6a632a773a.jpg", + "table_caption": [ + "Table 7: Efficiency Analysis on RTX3090. The 1st, 2nd, and 3rd-best methods are highlighted." + ], + "table_footnote": [], + "table_body": "
CategoryMethodAUC@10°#Param.Time (ms)Avg. Rank
SparseDeDoDe [11]65.028.1M3224.7
LightGlue [24]64.311.9M723.7
DenseLoFTR [39]69.211.6M4543.7
RoMa [12]76.9111M9134.3
Semi-SparseRCMLite66.511.4M1342.7
RCM69.49.8M2082
", + "bbox": [ + 444, + 191, + 785, + 280 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "RoMa, RCM achieves respectable performance with only $8.8\\%$ of the parameters and $23\\%$ of the runtime, making it more suitable for real-time applications.", + "bbox": [ + 215, + 297, + 784, + 328 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Limitation and Discussion", + "text_level": 1, + "bbox": [ + 215, + 349, + 503, + 364 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The view switcher assumes that one image has a relatively large scale, while both images being small in scale remains a tricky case in feature matching. A potential approach involves dynamically scaling the image pairs according to the overlapping regions, ensuring that the overlapping regions in two images remain simultaneously large and comparable in scale.", + "bbox": [ + 212, + 378, + 784, + 455 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Since the matching points in the target image are regression-based points rather than pre-detected points, RCM suffers from the lack of discrete tracks in the SfM task and requires additional adaptation strategy.", + "bbox": [ + 212, + 455, + 784, + 500 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The semi-sparse paradigm mitigates the keypoint-repeatability reliance, addressing only a portion of the broader keypoint dependence. Challenges in detection persist, such as low quantity and poor distribution, influencing the performance of RCM in texture-less scenes. We believe that the integration of the semi-sparse paradigm with advanced detection methods holds promising potential as a path forward for effective and efficient feature matching.", + "bbox": [ + 212, + 500, + 784, + 590 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 613, + 359, + 628 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper introduces a novel perspective to enhance matchers by raising the theoretical upper bound. We propose the dynamic view switching mechanism for the source image and the conflict-free coarse matching layer for the target image, respectively. Through the integration of the view switcher and many-to-one matching layer, we enhance the number of matchable points in the source image while achieving conflict-free matching in the target image. By essentially augmenting the number of ground-truth matches between images, RCM raises the theoretical upper bound on the actual matching results. The results from qualitative and quantitative experiments demonstrate a significant enhancement in the practical performance of the matcher with a higher ceiling, particularly evident in challenging scenarios characterized by large scale variations. With its outstanding performance and efficiency, we believe that RCM holds considerable potential for research and application.", + "bbox": [ + 212, + 643, + 785, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "X. Lu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by the SEU Innovation Capability Enhancement Plan for Doctoral Students under grant CXJH SEU 24128.", + "bbox": [ + 215, + 176, + 784, + 205 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 229, + 321, + 244 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Balntas, V., Lenc, K., Vedaldi, A., Mikolajczyk, K.: Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In: Proceedings of the CVPR. pp. 5173-5182 (2017)", + "2. Barroso-Laguna, A., Tian, Y., Mikolajczyk, K.: Scalenet: A shallow architecture for scale estimation. In: Proceedings of the CVPR. pp. 12808-12818 (2022)", + "3. Bay, H., Tuytelaars, T., Van G., L.: SURF: Speeded Up Robust Features. In: Proceedings of the ECCV. pp. 404-417 (2006)", + "4. Calonder, M., Lepetit, V., Strecha, C., Fua, P.: BRIEF: Binary Robust Independent Elementary Features. In: Proceedings of the ECCV. pp. 778-792 (2010)", + "5. Chen, H., Luo, Z., Zhang, J., Zhou, L., Bai, X., Hu, Z., Tai, C., Quan, L.: Learning to match features with seeded graph matching network. In: Proceedings of the ICCV. pp. 6301-6310 (2021)", + "6. Chen, H., Luo, Z., Zhou, L., Tian, Y., Zhen, M., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Aspanformer: Detector-free image matching with adaptive span transformer. In: Proceedings of the ECCV. pp. 20-36 (2022)", + "7. Chen, Y., Huang, D., Xu, S., Liu, J., Liu, Y.: Guide local feature matching by overlap estimation. In: Proceedings of the AAAI. pp. 365-373 (2022)", + "8. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the CVPR. pp. 5828-5839 (2017)", + "9. DeTone, D., Malisiewicz, T., Rabinovich, A.: SuperPoint: Self-Supervised Interest Point Detection and Description. In: Proceedings of the CVPRW. pp. 224-236 (2018)", + "0. Dusmanu, M., Rocco, I., Pajdla, T., Pollefeys, M., Sivic, J., Torii, A., Sattler, T.: D2-Net: A Trainable CNN for Joint Description and Detection of Local Features. In: Proceedings of the CVPR. pp. 8092–8101 (2019)", + "1. Edstedt, J., Bokman, G., Wadenback, M., Felsberg, M.: Dedode: Detect, don't describe—describe, don't detect for local feature matching. In: Proceedings of the 3DV. pp. 148-157 (2024)", + "2. Edstedt, J., Sun, Q., Bokman, G., Wadenback, M., Felsberg, M.: Roma: Robust dense feature matching. In: Proceedings of the CVPR. pp. 19790-19800 (2024)", + "3. Engel, J., Koltun, V., Cremers, D.: Direct sparse odometry. IEEE Transactions on Pattern Analysis and Machine Intelligence 40(3), 611-625 (2017)", + "4. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM pp. 381-395 (1981)", + "5. Fu, Y., Wu, Y.: Scale-net: Learning to reduce scale differences for large-scale invariant image matching. arXiv preprint arXiv:2112.10485 (2021)", + "6. Germain, H., Bourmaud, G., Lepetit, V.: Sparse-to-dense hypercolumn matching for long-term visual localization. In: Proceedings of the 3DV. pp. 513-523 (2019)", + "7. Germain, H., Bourmaud, G., Lepetit, V.: S2dnet: Learning image features for accurate sparse-to-dense matching. In: Proceedings of the ECCV. pp. 626-643 (2020)" + ], + "bbox": [ + 225, + 258, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "RCM: Conflict-Free Local Feature Matching with Dynamic View Switching", + "bbox": [ + 230, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Gleize, P., Wang, W., Feiszli, M.: Silk: Simple learned keypoints. In: Proceedings of the CVPR. pp. 22499-22508 (2023)", + "19. Huang, D., Chen, Y., Liu, Y., Liu, J., Xu, S., Wu, W., Ding, Y., Tang, F., Wang, C.: Adaptive assignment for geometry aware local feature matching. In: Proceedings of the CVPR. pp. 5425-5434 (2023)", + "20. Jiang, W., Trulls, E., Hosang, J., Tagliasacchi, A., Yi, K.M.: Cotr: Correspondence transformer for matching across images. In: Proceedings of the ICCV. pp. 6207-6217 (2021)", + "21. Katharopoulos, A., Vyas, A., Pappas, N., Fleuret, F.: Transformers are rnns: Fast autoregressive transformers with linear attention. In: Proceedings of the ICML. pp. 5156-5165 (2020)", + "22. Li, X., Han, K., Li, S., Prisacariu, V.: Dual-resolution correspondence networks. In: Proceedings of the NeurIPS. pp. 17346-17357 (2020)", + "23. Li, Z., Snavely, N.: MegaDepth: Learning Single-View Depth Prediction from Internet Photos. In: Proceedings of the CVPR. pp. 2041–2050 (2018)", + "24. Lindenberger, P., Sarlin, P., Pollefeys, M.: Lightglue: Local feature matching at light speed. In: Proceedings of the ICCV. pp. 17627-17638 (2023)", + "25. Liu, C., Yuen, J., Torralba, A., Sivic, J., Freeman, W.T.: Sift flow: Dense correspondence across different scenes. In: Proceedings of the ECCV. pp. 28-42 (2008)", + "26. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)", + "27. Lowe, D.G.: Distinctive Image Features from Scale-Invariant Keypoints. International Journal of Computer Vision 60(2), 91-110 (2004)", + "28. Lu, X., Yan, Y., Kang, B., Du, S.: Paraformer: Parallel attention transformer for efficient feature matching. arXiv preprint arXiv:2303.00941 (2023)", + "29. Luo, Z., Zhou, L., Bai, X., Chen, H., Zhang, J., Yao, Y., Li, S., Fang, T., Quan, L.: Aslfeat: Learning local features of accurate shape and localization. In: Proceedings of the CVPR. pp. 6589-6598 (2020)", + "30. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. Proceedings of the NeurIPS 32 (2019)", + "31. Revaud, J., De Souza, C., Humenberger, M., Weinzaepfel, P.: R2D2: Reliable and Repeatable Detector and Descriptor. In: Proceedings of the NeurIPS. pp. 12414-12424 (2019)", + "32. Rocco, I., Arandjelovic, R., Sivic, J.: Efficient neighbourhood consensus networks via submanifold sparse convolutions. In: Proceedings of the ECCV. pp. 605-621 (2020)", + "33. Rocco, I., Cimpoi, M., Arandjelović, R., Torii, A., Pajdla, T., Sivic, J.: Neighbourhood consensus networks. In: Proceedings of the NeurIPS. pp. 1658-1669 (2018)", + "34. Rublee, E., Rabaud, V., Konolige, K., Bradski, G.: ORB: An efficient alternative to SIFT or SURF. In: Proceedings of the ICCV. pp. 2564-2571 (2011)", + "35. Sarlin, P., Cadena, C. and Siegwart, R., Dymczyk, M.: From coarse to fine: Robust hierarchical localization at large scale. In: Proceedings of the CVPR. pp. 12716-12725 (2019)", + "36. Sarlin, P., DeTone, D., Malisiewicz, T., Rabinovich, A.: SuperGlue: Learning Feature Matching With Graph Neural Networks. In: Proceedings of the CVPR. pp. 4937-4946 (2020)", + "37. Schonberger, J.L., Frahm, J.: Structure-from-motion revisited. In: Proceedings of the CVPR. pp. 4104-4113 (2016)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "X. Lu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "38. Shi, Y., Cai, J., Shavit, Y., Mu, T., Feng, W., Zhang, K.: Clustergnn: Cluster-based coarse-to-fine graph neural network for efficient feature matching. In: Proceedings of the CVPR. pp. 12517-12526 (June 2022)", + "39. Sun, J., Shen, Z., Wang, Y., Bao, H., Zhou, X.: LoFTR: Detector-Free Local Feature Matching With Transformers. In: Proceedings of the CVPR. pp. 8922-8931 (2021)", + "40. Truong, P., Danelljan, M., Timofte, R., Van Gool, L.: Pdc-net+: Enhanced probabilistic dense correspondence network. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)", + "41. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is All you Need. In: Proceedings of the NeurIPS. pp. 5998-6008 (2017)", + "42. Wang, Q., Zhang, J., Yang, K., Peng, K., Stiefelhagen, R.: Matchformer: Interleaving attention in transformers for feature matching. In: Proceedings of the ACCV. pp. 2746-2762 (2022)", + "43. Wang, Q., Zhou, X., Hariharan, B., Snavely, N.: Learning feature descriptors using camera pose supervision. In: Proceedings of the ECCV. pp. 757-774 (2020)", + "44. Zhang, J., Sun, D., Luo, Z., Yao, A., Zhou, L., Shen, T., Chen, Y., Quan, L., Liao, H.: Learning two-view correspondences and geometry using order-aware network. In: Proceedings of the ICCV. pp. 5845-5854 (2019)", + "45. Zhang, Z., Sattler, T., Scaramuzza, D.: Reference pose generation for long-term visual localization via learned features and view synthesis. International Journal of Computer Vision pp. 821-844 (2021)", + "46. Zhao, X., Wu, X., Chen, W., Chen, P.C., Xu, Q., Li, Z.: A liked: A lighter keypoint and descriptor extraction network via deformable transformation. IEEE Transactions on Instrumentation and Measurement 72, 1-16 (2023)", + "47. Zhou, Q., Sattler, T., Leal-Taixe, L.: Patch2pix: Epipolar-guided pixel-level correspondences. In: Proceedings of the CVPR. pp. 4669-4678 (2021)" + ], + "bbox": [ + 215, + 146, + 787, + 535 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "RCM: Conflict-Free Local Feature Matching with Dynamic View Switching", + "bbox": [ + 230, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Raising the Ceiling_ Conflict-Free Local Feature Matching with Dynamic View Switching/f888146a-2080-4739-ac33-923c8980cbb0_model.json b/2024/Raising the Ceiling_ Conflict-Free Local Feature Matching with Dynamic View Switching/f888146a-2080-4739-ac33-923c8980cbb0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5a7293d5af886ec3e82e38893395787482bc44bb --- /dev/null +++ b/2024/Raising the Ceiling_ Conflict-Free Local Feature Matching with Dynamic View Switching/f888146a-2080-4739-ac33-923c8980cbb0_model.json @@ -0,0 +1,2423 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.23, + 0.141, + 0.776, + 0.187 + ], + "angle": 0, + "content": "Raising the Ceiling: Conflict-Free Local Feature Matching with Dynamic View Switching" + }, + { + "type": "text", + "bbox": [ + 0.392, + 0.213, + 0.61, + 0.228 + ], + "angle": 0, + "content": "Xiaoyong Lu and Songlin Du\\*" + }, + { + "type": "text", + "bbox": [ + 0.3, + 0.24, + 0.702, + 0.268 + ], + "angle": 0, + "content": "School of Automation, Southeast University, Nanjing, China {luxiaoyong,sdu}@seu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.298, + 0.74, + 0.588 + ], + "angle": 0, + "content": "Abstract. Current feature matching methods prioritize improving modeling capabilities to better align outputs with ground-truth matches, which are the theoretical upper bound on matching results, metaphorically depicted as the \"ceiling\". However, these enhancements fail to address the underlying issues that directly hinder ground-truth matches, including the scarcity of matchable points in small scale images, matching conflicts in dense methods, and the keypoint-repeatability reliance in sparse methods. We propose a novel feature matching method named RCM, which Raises the Ceiling of Matching from three aspects. 1) RCM introduces a dynamic view switching mechanism to address the scarcity of matchable points in source images by strategically switching image pairs. 2) RCM proposes a conflict-free coarse matching module, addressing matching conflicts in the target image through a many-to-one matching strategy. 3) By integrating the semi-sparse paradigm and the coarse-to-fine architecture, RCM preserves the benefits of both high efficiency and global search, mitigating the reliance on keypoint repeatability. As a result, RCM enables more matchable points in the source image to be matched in an exhaustive and conflict-free manner in the target image, leading to a substantial \\(260\\%\\) increase in ground-truth matches. Comprehensive experiments show that RCM exhibits remarkable performance and efficiency in comparison to state-of-the-art methods." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.602, + 0.74, + 0.615 + ], + "angle": 0, + "content": "Keywords: Feature Matching \\(\\cdot\\) Camera Pose Estimation \\(\\cdot\\) Transformer" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.638, + 0.377, + 0.654 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.667, + 0.787, + 0.773 + ], + "angle": 0, + "content": "Feature matching, which aims to establish accurate correspondences between two images, underpins various pivotal tasks in computer vision, such as object recognition [25], structure from motion (SfM) [37], and simultaneous localization and mapping (SLAM) [13]. However, challenges arise from factors such as viewpoint change, illumination variation, scale variation, occlusion and motion blur between images. These complexities elevate the difficulty of feature matching, particularly when real-time performance is required for practical applications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.773, + 0.788, + 0.819 + ], + "angle": 0, + "content": "Presently, methods for feature matching can be broadly categorized into two groups: sparse matching methods and dense matching methods. Sparse methods [5, 28, 36, 38] are designed to establish correspondences between two sets of" + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.826, + 0.383, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "X. Lu et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.144, + 0.396, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.397, + 0.145, + 0.785, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.289, + 0.788, + 0.359 + ], + "angle": 0, + "content": "Fig.1: Comparison among RCM, \\(\\mathrm{RCM}_{\\mathrm{Lite}}\\), SuperGlue [36], LoFTR [39] and MatchFormer [42]. Full RCM significantly accelerates LoFTR by \\(2.2\\times\\) while achieving superior performance. The lightweight \\(\\mathrm{RCM}_{\\mathrm{Lite}}\\) outperforms SuperGlue by \\(+7.5\\%\\) while maintaining a speed advantage. The same color indicates the matched features in the visualization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.393, + 0.785, + 0.438 + ], + "angle": 0, + "content": "keypoints. Consequently, the incorporation of detectors [9, 27, 31, 34] becomes essential in this process. In contrast, dense methods [20, 33, 39, 47] undertake a direct pursuit of matches between the dense features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.44, + 0.787, + 0.486 + ], + "angle": 0, + "content": "Through a comprehensive examination of prevailing sparse and dense matching methods, we identify and subsequently address three key issues that constrain the theoretical upper bound, i.e., the ceiling, on the number of matches." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.789, + 0.563 + ], + "angle": 0, + "content": "(1) Sparse and dense methods: dramatic decrease in the number of matchable keypoints in small scale images. Overlapping regions in small scale images are limited in size, with the majority of keypoints or grid points situated in irrelevant background, which not only causes redundancy in the message passing but also radically limits the number of ground-truth matches." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.789, + 0.672 + ], + "angle": 0, + "content": "For the semi-sparse [16, 17] and dense [39] matching paradigms, this challenge only arises in the source image as they perform a dense search in the target image. To tackle this challenge, we propose a dynamic view switcher that learns to switch the larger scale image to the source image. This strategic switch significantly augments the number of matchable keypoints within the overlapping region to radically raise the theoretical upper bound of the matcher, as illustrated in Fig. 6(b)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.672, + 0.789, + 0.794 + ], + "angle": 0, + "content": "(2) Dense method: matching conflicts in large scale variation scenes caused by the one-to-one matching strategy. As shown in Fig. 2(b), the case of many-to-one matching arises when dealing with significant scale variations, as more pixels in large scale images correspond to fewer pixels in small scale images. However, existing semi-dense methods struggle with this challenge as they employ a one-to-one matching strategy in the coarse matching stage, generating at most one coarse match within each \\(8 \\times 8\\) region in the target image while discarding other valid matches." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.843 + ], + "angle": 0, + "content": "In our method, correspondences are established between the source and target image in a many-to-one manner. Each matchable point can independently search for correspondences without conflicts, as depicted in Fig. 2(c) and Fig. 6(a)." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.487, + 0.789, + 0.843 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.231, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "RCM: Conflict-Free Local Feature Matching with Dynamic View Switching" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.784, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.263, + 0.788, + 0.32 + ], + "angle": 0, + "content": "Fig. 2: Comparison of three matching methods. (a) The sparse methods rely on precise keypoint detection at corresponding points in both images. (b) We highlight the issues of matching conflicts introduced in Sec. 1. (c) RCM establishes conflict-free coarse matches and adjusts them independently in the fine matching stage." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.349, + 0.784, + 0.439 + ], + "angle": 0, + "content": "(3) Sparse method: heavy reliance on the repeatability of keypoint. Sparse methods strategically narrow the search scope for matchers by leveraging detectors, yet they concurrently establish a severe reliance on the repeatability of keypoints. When input keypoints are not precisely detected at corresponding positions in both images and result in few ground-truth matches in challenging scenes, even a perfect sparse matcher fails to \"create something out of nothing\"." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.44, + 0.785, + 0.53 + ], + "angle": 0, + "content": "This challenge inspires the semi-sparse paradigm [16,17] to detect keypoints in the source image and perform a dense search in the target image, eliminating the requirement for precise detection of keypoints in both images. We adopt the semi-sparse paradigm and integrate it seamlessly with the coarse-to-fine architecture [39], allowing RCM to maintain efficiency while globally searching for sub-pixel matches in the target image." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.531, + 0.691, + 0.546 + ], + "angle": 0, + "content": "To summarize, this paper presents the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.555, + 0.785, + 0.6 + ], + "angle": 0, + "content": "- We seamlessly integrate the semi-sparse paradigm with the coarse-to-fine architecture, establishing a strong foundation for our subsequent efforts to raise the theoretical upper bound of feature matching." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.601, + 0.785, + 0.645 + ], + "angle": 0, + "content": "- We design a view switcher that dynamically switches the source and target images based on their scales. This view switcher offers a direct and effective method for increasing the number of matchable points in the source image." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.646, + 0.785, + 0.69 + ], + "angle": 0, + "content": "- We propose a conflict-free coarse matching module, which matches two sets of features in a many-to-one fashion, further breaking through the match quantity barrier in large scale variation scenes." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.691, + 0.785, + 0.721 + ], + "angle": 0, + "content": "- The proposed RCM and \\(\\mathrm{RCM}_{\\mathrm{Lite}}\\) achieve an excellent performance-efficiency balance, rendering the matchers suitable for a wider range of applications." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.555, + 0.785, + 0.721 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.754, + 0.396, + 0.769 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.785, + 0.46, + 0.801 + ], + "angle": 0, + "content": "2.1 Local Feature Matching" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Local feature matching methods are categorized as sparse and dense methods. Sparse methods involve the extraction of sparse features with a keypoint detec" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "X. Lu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "tor, followed by establishing correspondences through a matching process. Dense methods directly establish correspondences between dense features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.787, + 0.284 + ], + "angle": 0, + "content": "Sparse matching contains well-known methods like SIFT [27], SURF [3], BRIEF [4] and ORB [34], which are widely utilized. The advent of deep learning introduces learning-based descriptors as a dominant approach. Learning-based detectors and descriptors [9-11,18,31,46] harness the power of convolutional neural networks (CNN) for improved robustness. SuperGlue [36] pioneered global relationship modeling among sparse features using Transformer [41], which enriches features via self- and cross-attention mechanisms." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.285, + 0.787, + 0.39 + ], + "angle": 0, + "content": "Dense matching methods [6,39,42] eschew the detector and instead establish dense correspondences between images, resulting in a significant improvement over sparse methods. We attribute this advancement to the increase in ground-truth matches facilitated by coarse-to-fine matching, which effectively covers the entire target image. This success in raising the feature matching ceiling inspires RCM to inherit the coarse-to-fine matching approach and pursue further enhancements in ground-truth matches." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.392, + 0.787, + 0.513 + ], + "angle": 0, + "content": "Semi-Sparse matching methods, matching sparse and dense features asymmetrically, are proposed by S2DHM [16] and S2DNet [17]. These methods, while preserving sparse keypoints in the source image, use dense features at full image resolution in the target image, resulting in excessive memory usage. In our approach, we integrate the semi-sparse matching paradigm with coarse-to-fine matching [39] to maintain high efficiency. Additionally, the proposed dynamic view switching and many-to-one matching empower the semi-sparse paradigm to achieve performance comparable to modern sparse and dense paradigms." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.539, + 0.59, + 0.554 + ], + "angle": 0, + "content": "2.2 Feature Matching with Scale Invariance" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Feature matching struggles in scenes with large scale variations, primarily due to the scarcity of matchable points in small scale images and matching conflicts due to one-to-one matching. Many methods are proposed to align scales before matching. OETR [7] identifies overlapping regions between images with object detection. ScaleNet [2] and Scale-Net [15] directly predict the scale variation. To address the scale challenge, we divide scenes with large scale variations into two cases: source and target images with (1) small and large scales and (2) large and small scales. In the first case, both sparse [5,36] and dense [39,42] methods face a reduction in matchable points in the source image, which is solved by the proposed dynamic view switcher. In the second case, dense methods encounter matching conflicts in small scale target images, a challenge mitigated by the proposed many-to-one matching strategy. Similar to many-to-one matching, AdaMatcher [19] addresses the matching conflict in the second case through adaptive assignment. However, it is limited to one-to-one matching in the first case, where the scarcity of matchable points in the source image persists. With the combination of dynamic view switcher and many-to-one matching, we can effectively obtain more matchable points in the source image and perform conflict-free matching in the target image in all scenes with large scale variations." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.231, + 0.114, + 0.733, + 0.13 + ], + "angle": 0, + "content": "RCM: Conflict-Free Local Feature Matching with Dynamic View Switching" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.144, + 0.77, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.371, + 0.788, + 0.469 + ], + "angle": 0, + "content": "Fig. 3: RCM is comprised of four primary modules: 1. The U-Net Extraction Module extracts multi-scale coarse and fine features. 2. The View Switching Module dynamically switches the larger scale image to the sparse branch, where dense coarse features are transformed into sparse features through the detection head. 3. The Conflict-Free Coarse Matching Module involves attention layers processing two sets of coarse features, which are subsequently matched via the many-to-one matching layer. 4. The Fine Matching Module further refines coarse matches based on fine features." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.494, + 0.38, + 0.511 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.522, + 0.341, + 0.536 + ], + "angle": 0, + "content": "3.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.543, + 0.788, + 0.696 + ], + "angle": 0, + "content": "The overall architecture of RCM is illustrated in Fig. 3. The U-Net extraction module initiates the process by employing CNN to extract coarse and fine features from both images. Subsequently, the coarse features of the two images undergo processing within the view switcher. In this module, the coarse features of larger scale images are switched to the sparse branch and extracted by the detection head as sparse features. And the coarse features of smaller scale images are retained as dense features. The sparse and dense features are then matched in a many-to-one fashion through the conflict-free coarse matching module. Finally the fine matching module crops out the fine features in the coarse matching positions and refines coarse matches through the correlation-based approach [39]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.713, + 0.469, + 0.727 + ], + "angle": 0, + "content": "3.2 Local Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We extend the VGG-style encoder from SuperPoint [9] to a U-Net architecture to extract multi-scale features from images \\( I_{A} \\) and \\( I_{B} \\). The decoder output is employed as fine features, denoted as \\( F_{A}^{f}\\in \\mathbb{R}^{H_{f}\\times W_{f}\\times C_{2}} \\) and \\( F_{B}^{f}\\in \\mathbb{R}^{H_{f}\\times W_{f}\\times C_{2}} \\). For brevity, we omit the \\( A \\) and \\( B \\) subscripts in the rest of the paper. On the other hand, the coarse features, denoted as \\( F^c\\in \\mathbb{R}^{H_c\\times W_c\\times C_1} \\), are formed by merging features from multiple resolutions, enhancing the robustness of the conflict-free coarse matching module." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "X. Lu et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.46, + 0.162 + ], + "angle": 0, + "content": "3.3 View Switching Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.461, + 0.401 + ], + "angle": 0, + "content": "As shown in Fig. 6(b), employing a small scale image as the source image leads to a scarcity of matchable points, which consequently leads to fewer matches. As shown in Fig. 4, we propose a binary classification network named view switcher to determine whether to switch the source and target images, where a lightweight CNN is employed to process the correlation map of coarse features \\( F_{A}^{c} \\) and \\( F_{B}^{c} \\) to perceive scale variation. The result of view switcher \\( VS \\) is computed as" + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.205, + 0.782, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.469, + 0.324, + 0.788, + 0.38 + ], + "angle": 0, + "content": "Fig. 4: The view switcher perceives the scale variation between image pairs to predict a binary classification result, indicating whether to switch the images." + }, + { + "type": "equation", + "bbox": [ + 0.321, + 0.415, + 0.786, + 0.433 + ], + "angle": 0, + "content": "\\[\nV S = \\operatorname {S o f t m a x} (\\operatorname {C N N} (\\operatorname {C o r r} (\\operatorname {P o o l} (F _ {A} ^ {c}), \\operatorname {P o o l} (F _ {B} ^ {c})))) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.788, + 0.581 + ], + "angle": 0, + "content": "where \\(\\mathrm{Pool}(\\cdot)\\) represents average pooling and \\(\\mathrm{Corr}(\\cdot, \\cdot)\\) signifies similarity calculation through the inner product. The view switcher is trained to keep sparse branch processing larger scale images and dense branch processing smaller scale images. In the sparse branch, we employ the detection head in [9] to detect keypoints and sample the feature map at the keypoint positions as sparse features \\(F_{s}^{c}\\). The U-Net encoder and detection head are inherited from SuperPoint [9] and are frozen, making the detection equivalent to SuperPoint. Assuming that \\(N\\) keypoints are detected, features in the sparse branch are reduced from \\(\\mathbb{R}^{H_c \\times W_c \\times C_1}\\) to \\(\\mathbb{R}^{N \\times C_1}\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.606, + 0.517, + 0.621 + ], + "angle": 0, + "content": "3.4 Conflict-Free Coarse Matching" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.633, + 0.788, + 0.724 + ], + "angle": 0, + "content": "Position Encoder Given that the attention mechanism lacks awareness of the positions of the features, we embed positional information into the coarse features \\( F^c \\) through a position encoder PE, in which a lightweight MLP progressively expands the positional information \\( P \\in \\mathbb{R}^{N \\times 3} \\) into the coarse feature dimension \\( C_1 \\) and updates the coarse features. \\( P \\) encompasses the \\( x \\) and \\( y \\) coordinates of keypoints, and their confidence scores \\( s \\) generated by the detection head." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.748, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Attention Layer The self- and cross-attention layers are employed for the global enhancement of sparse and dense features. In contrast to dense methods that use linear attention [21] for efficiency, the semi-sparse matching paradigm allows us to adopt vanilla attention [41], which is denoted as" + }, + { + "type": "equation", + "bbox": [ + 0.338, + 0.822, + 0.786, + 0.84 + ], + "angle": 0, + "content": "\\[\nM = \\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {S o f t m a x} \\left(Q K ^ {T}\\right) V. \\tag {2}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.23, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "RCM: Conflict-Free Local Feature Matching with Dynamic View Switching" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.192 + ], + "angle": 0, + "content": "The attentional message is denoted as \\( M \\). In self-attention, \\( Q, K, V \\) are three linear projections of a single feature set. In cross-attention, \\( Q \\) is projected from one feature set, while \\( K \\) and \\( V \\) are projected from another feature set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.787, + 0.223 + ], + "angle": 0, + "content": "The feed-forward network \\(\\mathrm{MLP_f}\\) processes the attentional message \\(M\\) to update the features as" + }, + { + "type": "equation", + "bbox": [ + 0.403, + 0.251, + 0.786, + 0.269 + ], + "angle": 0, + "content": "\\[\nF ^ {i + 1} = F ^ {i} + \\mathrm {M L P} _ {\\mathrm {f}} \\left(F ^ {i} | | M\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.28, + 0.787, + 0.328 + ], + "angle": 0, + "content": "where \\( F^i \\) is the feature in layer \\( i \\) and \\( || \\) is the concatenation operation. We perform self- and cross-attention on sparse and dense features for \\( L_1 \\) iterations, resulting in enhanced features \\( \\tilde{F}_s^c \\) and \\( \\tilde{F}_d^c \\) that capture global information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.348, + 0.788, + 0.472 + ], + "angle": 0, + "content": "Many-to-One Coarse Matching with Dustbin With globally enhanced features, we perform matching between sparse and dense features in a many-to-one manner. Initially, we utilize a set of learnable parameters as a dustbin, denoted as \\( bin \\in \\mathbb{R}^{1 \\times C_1} \\), and concatenate it with the dense features \\( \\hat{F}_d^c = \\tilde{F}_d^c ||bin \\in \\mathbb{R}^{(H_cW_c + 1) \\times C_1} \\). The dustbin is introduced to match with non-matchable sparse features, enabling our matcher to handle non-overlapping regions and occlusion. We compute the score matrix with dustbin \\( S_{bin} \\) between the sparse feature \\( \\tilde{F}_s^c \\in \\mathbb{R}^{N \\times C_1} \\) and dense feature with dustbin \\( \\hat{F}_d^c \\in \\mathbb{R}^{(H_cW_c + 1) \\times C_1} \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.481, + 0.786, + 0.5 + ], + "angle": 0, + "content": "\\[\nS _ {b i n} = \\tilde {F} _ {s} ^ {c} \\hat {F} _ {d} ^ {c} / \\tau \\in \\mathbb {R} ^ {N \\times (H _ {c} W _ {c} + 1)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.511, + 0.553, + 0.526 + ], + "angle": 0, + "content": "where \\(\\tau\\) is a learnable temperature parameter." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.527, + 0.788, + 0.663 + ], + "angle": 0, + "content": "Current dense methods utilize the Dual-Softmax operator to compute the coarse matching probability \\( P_{c} \\), defined as \\( P_{c} = \\mathrm{Softmax}_{row}(S) \\cdot \\mathrm{Softmax}_{col}(S) \\in \\mathbb{R}^{N \\times H_{c}W_{c}} \\). The Dual-Softmax operator keeps the summation of rows and columns of \\( P_{c} \\) less than or equal to 1. This limits each feature to match at most one feature, that is, one-to-one matching. However multiple points may correspond to one point in scenes with large scale variations, which leads to matching conflicts as shown in Fig. 2(b). To address this issue, a single Softmax operation on the score with dustbin \\( S_{bin} \\) is employed to release the constraint on target images while maintaining the constraint on source images as" + }, + { + "type": "equation", + "bbox": [ + 0.419, + 0.677, + 0.786, + 0.693 + ], + "angle": 0, + "content": "\\[\nP _ {c} = \\operatorname {S o f t m a x} _ {r o w} \\left(S _ {b i n}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.81 + ], + "angle": 0, + "content": "where each sparse feature can only match one dense feature or the dustbin, while one dense feature can match multiple sparse features. The strategy of many-to-one matching with dustbin can be comprehended as independently assigning each sparse feature to either its corresponding dense feature or to the dustbin. As shown in Fig. 2(c), following conflict-free coarse matching, the positions of sparse features, which correspond to the same dense feature, are independently adjusted during the fine matching stage." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Ultimately, features matched to the dustbin are discarded, and we proceed to select matches with matching probabilities surpassing the threshold \\(\\theta_c\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "X. Lu et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.384, + 0.163 + ], + "angle": 0, + "content": "3.5 Fine Matching" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.168, + 0.784, + 0.229 + ], + "angle": 0, + "content": "For each coarse match, we sample a single feature at the source fine feature map and crop a feature window of size \\( w \\times w \\) at the target fine feature map. Given the collection of \\( K \\) coarse matches, we obtain fine features denoted as \\( \\hat{F}_s^f \\in \\mathbb{R}^{K \\times 1 \\times C_2} \\) and \\( \\hat{F}_d^f \\in \\mathbb{R}^{K \\times w^2 \\times C_2} \\) for the sparse and dense branches, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.231, + 0.784, + 0.292 + ], + "angle": 0, + "content": "Subsequently, \\(\\hat{F}_s^f\\) and \\(\\hat{F}_d^f\\) are processed by the cross-attention module for \\(L_{2}\\) iterations. Inspired by [6], in order to capture the local context of dense features, we substitute the self-attention module with the integration of \\(3\\times 3\\) convolutions into the feed-forward network. For fine matching, the features are updated by" + }, + { + "type": "equation", + "bbox": [ + 0.403, + 0.3, + 0.785, + 0.317 + ], + "angle": 0, + "content": "\\[\nF ^ {i + 1} = F ^ {i} + \\operatorname {C o n v} _ {3} \\left(F ^ {i} | | M\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.323, + 0.784, + 0.384 + ], + "angle": 0, + "content": "We compute the correlation map between the centroid of the source feature window and all the features in the target feature window, which represents the matching probability, and then get the fine matching position by computing the expectation of the probability distribution, as introduced in [39]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.401, + 0.36, + 0.417 + ], + "angle": 0, + "content": "3.6 Supervision" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.784, + 0.51 + ], + "angle": 0, + "content": "Our loss function encompasses four components: the coarse matching loss \\(\\mathcal{L}_c^m\\), the coarse dustbin loss \\(\\mathcal{L}_c^d\\), the fine matching loss \\(\\mathcal{L}_f\\), and the view switching loss \\(\\mathcal{L}_{vs}\\). Ground-truth matches \\(M_{gt} = \\{(i_m,j_m)\\}\\) are generated from camera poses and depth maps, establishing many-to-one correspondence between sparse and dense features. And non-matchable sparse features \\(U_{gt} = \\{i_u\\}\\), which are keypoints without any ground-truth match, are also obtained." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.512, + 0.785, + 0.557 + ], + "angle": 0, + "content": "The coarse matching loss \\(\\mathcal{L}_c^m\\) is computed as the negative log-likelihood loss of the ground-truth matches \\(M_{gt}\\) over the matching probability matrix \\(P_c\\) of each attention layer," + }, + { + "type": "equation", + "bbox": [ + 0.356, + 0.562, + 0.785, + 0.6 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} ^ {m} = - \\frac {1}{L} \\sum_ {l} \\left[ \\frac {1}{| M _ {g t} |} \\sum_ {(i, j) \\in M _ {g t}} \\log P _ {c} ^ {l} (i, j) \\right]. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.606, + 0.784, + 0.652 + ], + "angle": 0, + "content": "The coarse dustbin loss \\(\\mathcal{L}_c^d\\) refers to the negative log-likelihood loss of the non-matchable sparse features \\(U_{gt}\\) over the last column of the matching probability matrix \\(P_c\\), which signifies the matching probabilities with the dustbin," + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.657, + 0.785, + 0.693 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} ^ {d} = - \\frac {1}{| U _ {g t} |} \\sum_ {i \\in U _ {g t}} \\log P _ {c} (i, H _ {c} W _ {c} + 1). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.698, + 0.784, + 0.728 + ], + "angle": 0, + "content": "The fine matching loss \\(\\mathcal{L}_f\\) is computed by the L2-distance between each refined position and the ground-truth position, as introduced in [39]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.729, + 0.784, + 0.818 + ], + "angle": 0, + "content": "To supervise the view switcher, we randomly select 500 ground-truth matches in each image pair and compute the distance between points in each image. The image with the larger average distance is identified as the larger scale image. If the original source image is the larger scale image, the ground-truth \\( VS_{gt} \\) is false, and vice versa. The binary cross-entropy loss of the switcher result \\( VS \\) is subsequently computed as" + }, + { + "type": "equation", + "bbox": [ + 0.325, + 0.826, + 0.785, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v s} = - \\left[ V S _ {g t} \\log (V S) + \\left(1 - V S _ {g t}\\right) \\log (1 - V S) \\right]. \\tag {9}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.231, + 0.114, + 0.732, + 0.131 + ], + "angle": 0, + "content": "RCM: Conflict-Free Local Feature Matching with Dynamic View Switching" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 1: Evaluation on HPatches [1] for homography estimation. We report the percentage of correctly estimated homographies with mean corner errors below \\(1/3/5\\) pixels. We mark the best accuracy in bold." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.199, + 0.78, + 0.389 + ], + "angle": 0, + "content": "
CategoryMethodOverallIlluminationViewpoint
Accuracy (%,ε < 1/3/5 px)
SparseSuperPoint [9] + NN0.46 / 0.78 / 0.850.57 / 0.92 / 0.970.35 / 0.65 / 0.74
D2Net [10] + NN0.38 / 0.72 / 0.810.65 / 0.95 / 0.980.13 / 0.51 / 0.65
R2D2 [31] + NN0.47 / 0.78 / 0.830.63 / 0.93 / 0.980.33 / 0.64 / 0.70
ASLFeat [29] + NN0.48 / 0.81 / 0.880.63 / 0.94 / 0.980.34 / 0.69 / 0.78
SuperPoint + SuperGlue [36]0.51 / 0.83 / 0.890.62 / 0.93 / 0.980.41 / 0.73 / 0.81
SuperPoint + CAPS [43] + NN0.49 / 0.79 / 0.860.62 / 0.93 / 0.980.36 / 0.65 / 0.75
SIFT + CAPS [43] + NN0.36 / 0.76 / 0.850.48 / 0.89 / 0.950.26 / 0.65 / 0.76
SuperPoint + LightGlue [24]0.39 / 0.85 / 0.910.52 / 0.96 / 0.980.28 / 0.75 / 0.84
(Semi-)DenseSparseNCNet [32]0.36 / 0.66 / 0.760.62 / 0.92 / 0.970.13 / 0.41 / 0.57
NCNet [33]0.48 / 0.61 / 0.710.98 / 0.98 / 0.980.02 / 0.28 / 0.46
Patch2Pix [47]0.50 / 0.79 / 0.870.71 / 0.95 / 0.980.30 / 0.64 / 0.76
LoFTR [39]0.55 / 0.81 / 0.860.74 / 0.95 / 0.980.38 / 0.69 / 0.76
MatchFormer [42]0.55 / 0.81 / 0.870.75 / 0.95 / 0.980.37 / 0.68 / 0.78
Semi-SparseRCMLite0.50 / 0.86 / 0.910.55 / 0.95 / 0.980.45 / 0.79 / 0.85
RCM0.56 / 0.88 / 0.930.65 / 0.97 / 0.990.47 / 0.80 / 0.88
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.418, + 0.377, + 0.435 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.453, + 0.458, + 0.469 + ], + "angle": 0, + "content": "4.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.482, + 0.789, + 0.727 + ], + "angle": 0, + "content": "Both RCM and \\(\\mathrm{RCM}_{\\mathrm{Lite}}\\) consist of \\(L_{1} = 5\\) layers of coarse attention and \\(L_{2} = 2\\) layers of fine attention. The dimensions for coarse and fine features are \\(C_1 = 256\\) and \\(C_2 = 64\\), respectively. The key distinction between RCM and \\(\\mathrm{RCM}_{\\mathrm{Lite}}\\) lies in their coarse feature resolution and fine matching window size. For RCM, the coarse feature resolution is \\(1/8\\), and the fine matching window size is \\(w = 5\\). For \\(\\mathrm{RCM}_{\\mathrm{Lite}}\\), the coarse feature resolution is \\(1/16\\), and the fine matching window size is \\(w = 9\\). The learnable temperature \\(\\tau\\) is initialized to 0.1, and the coarse matching threshold \\(\\theta_c\\) is set to 0.2. The proposed method is implemented using Pytorch [30]. During the training phase, images are resized and padded to the size of \\(832 \\times 832\\). We detect 1024 keypoints in the source image, and if there are insufficient keypoints, random keypoints are added to ensure efficient batching. RCM and \\(\\mathrm{RCM}_{\\mathrm{Lite}}\\) are trained on the MegaDepth dataset [23] for 30 epochs with a batch size of 4 by AdamW optimizer [26] and are not fine-tuned for any other tasks. The initial learning rate is set to 0.0002, and the training employs the cosine decay learning rate scheduler with 1 epoch of linear warm-up. More details can be found in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.751, + 0.466, + 0.767 + ], + "angle": 0, + "content": "4.2 Homography Estimation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Dataset. The HPatches [1] dataset is widely recognized for evaluating local features and matchers. In line with the evaluation protocol in [10], we assess the performance of our method across 108 HPatches sequences, which comprise 52 sequences with illumination variations and 56 sequences with viewpoint changes." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "X. Lu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "Metric. Following the metrics employed in [9, 36, 47], we report the percentage of correctly estimated homographies with mean corner errors below \\(1/3/5\\) pixels. Vanilla RANSAC [14] is applied to estimate the homography for all methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.344 + ], + "angle": 0, + "content": "Results. As shown in Tab. 1, RCM achieves the best overall performance at all thresholds. Comparing the dense methods, RCM shows a significant advantage in accuracy in large viewpoint change scenarios, leading LoFTR and MatchFormer by \\((+9\\%, +11\\%, +12\\%)\\) and \\((+10\\%, +12\\%, +10\\%)\\), respectively. We attribute this performance improvement to the increased matchable points in the source image facilitated by the view switcher, and the conflict-free matching achieved in the target image through many-to-one matching. Compared to the sparse method SuperGlue, RCM achieves an overall performance lead of \\((+5\\%, +5\\%, +4\\%)\\), benefiting from the dense search due to the semi-sparse paradigm and the sub-pixel accuracy due to the coarse-to-fine architecture." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.372, + 0.498, + 0.4 + ], + "angle": 0, + "content": "Table 2: Two-view pose estimation results on MegaDepth [23] in outdoor scenes." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.413, + 0.495, + 0.561 + ], + "angle": 0, + "content": "
CategoryMethodPose estimation
@5°@10°@20°
SparseSuperGlue [36]42.259.073.6
SGMNet [5]40.559.073.6
ClusterGNN [38]44.258.570.3
LightGlue [24]46.464.378.2
DeDoDe-B [11]49.165.077.1
DeDoDe-G [11]51.268.481.2
DenseDRC-Net [22]27.042.958.3
PDC-Net+(H) [40]43.161.976.1
LoFTR [39]52.869.281.2
RoMa [12]63.076.986.4
Semi-SparseRCMLite49.766.579.3
RCM53.269.481.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.504, + 0.386, + 0.785, + 0.414 + ], + "angle": 0, + "content": "Table 3: Two-view pose estimation results on ScanNet [8] in indoor scenes." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.427, + 0.779, + 0.546 + ], + "angle": 0, + "content": "
CategoryMethodPose estimationAUC
@5°@10°@20°
SparseOANet [44]11.826.943.9
SuperGlue [36]16.233.851.8
SGMNet [5]15.432.148.3
LightGlue [24]*13.929.645.7
DenseDRC-Net [22]*7.717.930.5
MatchFormer [42]*15.832.048.0
LoFTR-OT [39]*16.933.650.6
Semi-SparseRCMLite*15.231.548.8
RCM*17.334.652.1
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.615, + 0.471, + 0.629 + ], + "angle": 0, + "content": "4.3 Relative Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.784, + 0.703 + ], + "angle": 0, + "content": "Dataset. We evaluate matchers on the MegaDepth [23] and the ScanNet [8] dataset for outdoor and indoor pose estimation, respectively. For the MegaDepth dataset, all test images are resized such that their longer dimensions are 1216. For the ScanNet dataset, all test images are resized to \\(640 \\times 480\\) pixels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.785, + 0.763 + ], + "angle": 0, + "content": "Metric. Based on the matching results, we apply vanilla RANSAC to solve the fundamental matrices and relative poses of the image pairs. The AUC of the pose error at thresholds \\((5^{\\circ}, 10^{\\circ}, 20^{\\circ})\\) are reported, where pose error is the maximum angular error in rotation and translation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Results. As shown in Tab. 2, RCM achieves the second best performance on the MegaDepth dataset. RCM outperforms the state-of-the-art method LoFTR, demonstrating the benefits of the view switcher and many-to-one matching. The lightweight \\(\\mathrm{RCM}_{\\mathrm{Lite}}\\) achieves a performance boost \\((+7.5\\%, +7.5\\%, +5.7\\%)\\) compared to SuperGlue. Fig. 5 illustrates qualitative comparisons of the SuperGlue," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.231, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RCM: Conflict-Free Local Feature Matching with Dynamic View Switching" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.784, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.41, + 0.785, + 0.439 + ], + "angle": 0, + "content": "Fig. 5: Qualitative comparison of SuperGlue [36], LoFTR [39] and proposed RCM on MegaDepth and ScanNet." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.466, + 0.784, + 0.496 + ], + "angle": 0, + "content": "LoFTR and RCM. More visualizations and analyses can be found in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.497, + 0.785, + 0.603 + ], + "angle": 0, + "content": "As shown in Tab. 3, RCM achieves a \\((0.4\\%, 1\\%, 1.5\\%)\\) performance improvement over LoFTR in the ScanNet dataset. It's worth noting that the superscript * indicates zero-shot performance, and the model is trained on MegaDepth. Nevertheless, our methods achieve performance comparable to state-of-the-art sparse methods trained on ScanNet. Dense and semi-sparse matching methods find matches in the target image exhaustively, mitigating the challenge of nonrepeatability in indoor scenes, a key limitation of sparse methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.622, + 0.421, + 0.636 + ], + "angle": 0, + "content": "4.4 Visual Localization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.784, + 0.703 + ], + "angle": 0, + "content": "Dataset. We assess the performance of our method using the Aachen DayNight v1.1 benchmark [45], a well-established dataset for visual localization. This dataset comprises 824 day-time images and 191 night-time images, chosen as queries for outdoor visual localization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.784, + 0.749 + ], + "angle": 0, + "content": "Metric. We utilize the open-source localization pipeline HLoc [35] for conducting localization with RCM. We report the percentage of successfully localized images within three error thresholds \\((0.25m, 2^{\\circ}) / (0.5m, 5^{\\circ}) / (1m, 10^{\\circ})\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Results. As shown in Tab. 4, RCM demonstrates comparable or superior performance compared to the dense methods LoFTR and ASpanFormer in outdoor visual localization tasks. Our approach exhibits strong generalization across various downstream tasks without the need for fine-tuning. It is noteworthy that the semi-sparse method RCM is significantly more efficient than the dense methods, achieving a runtime speedup of over twofold." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "X. Lu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.247, + 0.145, + 0.754, + 0.16 + ], + "angle": 0, + "content": "Table 4: Visual localization on the Aachen Day-Night benchmark v1.1 [45]." + }, + { + "type": "table", + "bbox": [ + 0.306, + 0.171, + 0.699, + 0.293 + ], + "angle": 0, + "content": "
MethodDayNight
(0.25m, 2°) / (0.5m, 5°) / (1m, 10°)
Localization with matching pairs generated by HLoc
LoFTR [39]88.7 / 95.6 / 99.078.5 / 90.6 / 99.0
ASpanFormer [6]89.4 / 95.6 / 99.077.5 / 91.6 / 99.0
DeDoDe-B [11]87.4 / 94.7 / 98.570.7 / 88.0 / 97.9
DeDoDe-G [11]88.5 / 95.4 / 99.074.9 / 89.5 / 99.5
SP [9]+LightGlue [24]89.6 / 95.8 / 99.272.8 / 88.0 / 99.0
RCMLite87.7 / 94.7 / 98.472.3 / 91.1 / 98.4
RCM89.7 / 96.0 / 98.772.8 / 91.6 / 99.0
" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.305, + 0.528, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.304, + 0.785, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.513, + 0.788, + 0.571 + ], + "angle": 0, + "content": "Fig. 6: Ablation study on view switcher and many-to-one matching. (a) (b) shows the actual matching results and (c) (d) shows the ground-truth coarse matches. The harmonious combination of view switcher and many-to-one matching allows more matchable points in the source image and conflict-free matching in the target image." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.6, + 0.39, + 0.616 + ], + "angle": 0, + "content": "4.5 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.703 + ], + "angle": 0, + "content": "In this subsection, we conduct qualitative and quantitative ablation experiments on MegaDepth. All models in Tab. 5 are trained for 15 epochs at 544 resolution and tested at 1216 resolution. For the baseline model (1), we apply the original SuperPoint [9] to generate keypoints and descriptors for both images. O2O and M2O in Tab. 5 are abbreviations for one-to-one and many-to-one, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Ground-Truth Match Analysis. Fig. 6(d) illustrates the core idea of RCM, which is to increase the theoretical upper bound on the matcher, i.e., ground-truth matches, through view switcher and many-to-one matching. Fig. 6(c) illustrates a quantitative comparison of ground-truth matches, showing an increase as the view switching accuracy improves during training. Eventually, RCM and \\(\\mathrm{RCM}_{\\mathrm{Lite}}\\) achieve view switching accuracies of \\(90.9\\%\\) and \\(87.8\\%\\), respectively. Through the integration of many-to-one matching and the view switcher, the ground-truth matches surge to \\(260\\%\\) compared to one-to-one matching, which is a compelling demonstration of raising the ceiling of feature matching." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.231, + 0.114, + 0.732, + 0.131 + ], + "angle": 0, + "content": "RCM: Conflict-Free Local Feature Matching with Dynamic View Switching" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.128 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.149, + 0.499, + 0.178 + ], + "angle": 0, + "content": "Table 5: Ablation study of each proposal on MegaDepth." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.189, + 0.495, + 0.308 + ], + "angle": 0, + "content": "
Method#MatchesPrec.
(1) Sparse sampling451.798.0
(2) Dense sampling932.193.5
(3) Semi-Sparse sampling492.393.7
(4) + U-Net Extraction507.894.2
(5) O2O → M2O770.7(+51.8%)97.2
(6) + Switcher (RCMLite)951.4(+23.4%)97.8
(7) 1/16 → 1/8 (RCM)961.498.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.504, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 6: Evaluation on extreme-scale MegaDepth." + }, + { + "type": "table", + "bbox": [ + 0.537, + 0.185, + 0.756, + 0.311 + ], + "angle": 0, + "content": "
Method#MatchesPrec.
SP [9] + SuperGlue [36]26175.5
SP [9] + LightGlue [24]31376.0
LoFTR [39]42453.5
MatchFormer [42]46477.0
ASpanFormer [6]42577.1
RCM88778.9
wo switcher65375.4
flipped switcher7654.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.349, + 0.784, + 0.412 + ], + "angle": 0, + "content": "Sampling Strategy. As shown in Tab. 5, the baseline models (1), (2), and (3) differ solely in sampling strategies for coarse features. Precisely, (1) detects keypoints in both images, (2) employs dense feature as the coarse feature, and (3) detects keypoints in source image and employs dense feature in target image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.785, + 0.492 + ], + "angle": 0, + "content": "Many-to-One Matching. As shown in Fig. 6(a), in contrast to the one-to-one matching strategy, the many-to-one matching strategy pursues conflict-free matching, resulting in more detailed matches. As shown in Tab. 5(5), substituting one-to-one matching with many-to-one matching improves the number of matches by \\(51.8\\%\\) and precision by \\(3\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.788, + 0.603 + ], + "angle": 0, + "content": "View Switcher. As shown in Fig. 6(b), without the view switcher, the matcher can only produce a small number of matches in small scale images because only a limited number of keypoints are detected in the overlapping region. The view switcher enables the detection of more matchable keypoints in the source image. As shown in Tab. 5(6), the average number of matches increases by \\(23.4\\%\\) with the view switcher. More analyses about view switcher and many-to-one matching can be found in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.606, + 0.788, + 0.729 + ], + "angle": 0, + "content": "RCM and \\(\\mathrm{RCM}_{\\mathrm{Lite}}\\). The distinctions between RCM and \\(\\mathrm{RCM}_{\\mathrm{Lite}}\\) lie in the coarse feature resolution and fine matching window. RCM employs coarse features at 1/8 resolution, consistent with the dense attention-based methods. We apply 1/16 resolution coarse features in \\(\\mathrm{RCM}_{\\mathrm{Lite}}\\) to trade performance degradation for faster runtime, facilitating applications in real-time scenarios. Because of the additional layers at 1/16 resolution in the feature extraction network, \\(\\mathrm{RCM}_{\\mathrm{Lite}}\\) consists of more parameters than RCM. Also to cover larger coarse feature patches, we design a larger fine matching window for \\(\\mathrm{RCM}_{\\mathrm{Lite}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.732, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Robustness to Scale Variation. We evaluate matchers on the MegaDepth subset defined in OETR [7], where the overlap scale ratio is in \\([4, +\\infty]\\). As shown in Tab. 6, RCM achieves the best performance with \\(2.8\\times\\) and \\(2.1\\times\\) more matches than LightGlue and ASpanFormer respectively. The performance of RCM is severely degraded without the view switcher, demonstrating the significance of the view switcher in large scale variation scenarios. Flipping the switching result, i.e., actively taking small scale image as source image, leads to the worst results." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "X. Lu et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.432, + 0.297 + ], + "angle": 0, + "content": "Efficiency. The parameters, runtimes, and performances of state-of-the-art matchers are presented in Tab. 7. Notably, RCM and \\(\\mathrm{RCM}_{\\mathrm{Lite}}\\) achieve the optimal balance between performance and efficiency. Although the performance of RCM lags behind the most precise matcher" + }, + { + "type": "table_caption", + "bbox": [ + 0.443, + 0.163, + 0.788, + 0.19 + ], + "angle": 0, + "content": "Table 7: Efficiency Analysis on RTX3090. The 1st, 2nd, and 3rd-best methods are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.446, + 0.192, + 0.787, + 0.281 + ], + "angle": 0, + "content": "
CategoryMethodAUC@10°#Param.Time (ms)Avg. Rank
SparseDeDoDe [11]65.028.1M3224.7
LightGlue [24]64.311.9M723.7
DenseLoFTR [39]69.211.6M4543.7
RoMa [12]76.9111M9134.3
Semi-SparseRCMLite66.511.4M1342.7
RCM69.49.8M2082
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.298, + 0.785, + 0.329 + ], + "angle": 0, + "content": "RoMa, RCM achieves respectable performance with only \\(8.8\\%\\) of the parameters and \\(23\\%\\) of the runtime, making it more suitable for real-time applications." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.35, + 0.504, + 0.366 + ], + "angle": 0, + "content": "5 Limitation and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.785, + 0.456 + ], + "angle": 0, + "content": "The view switcher assumes that one image has a relatively large scale, while both images being small in scale remains a tricky case in feature matching. A potential approach involves dynamically scaling the image pairs according to the overlapping regions, ensuring that the overlapping regions in two images remain simultaneously large and comparable in scale." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.456, + 0.785, + 0.501 + ], + "angle": 0, + "content": "Since the matching points in the target image are regression-based points rather than pre-detected points, RCM suffers from the lack of discrete tracks in the SfM task and requires additional adaptation strategy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.501, + 0.785, + 0.592 + ], + "angle": 0, + "content": "The semi-sparse paradigm mitigates the keypoint-repeatability reliance, addressing only a portion of the broader keypoint dependence. Challenges in detection persist, such as low quantity and poor distribution, influencing the performance of RCM in texture-less scenes. We believe that the integration of the semi-sparse paradigm with advanced detection methods holds promising potential as a path forward for effective and efficient feature matching." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.614, + 0.36, + 0.63 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.787, + 0.841 + ], + "angle": 0, + "content": "This paper introduces a novel perspective to enhance matchers by raising the theoretical upper bound. We propose the dynamic view switching mechanism for the source image and the conflict-free coarse matching layer for the target image, respectively. Through the integration of the view switcher and many-to-one matching layer, we enhance the number of matchable points in the source image while achieving conflict-free matching in the target image. By essentially augmenting the number of ground-truth matches between images, RCM raises the theoretical upper bound on the actual matching results. The results from qualitative and quantitative experiments demonstrate a significant enhancement in the practical performance of the matcher with a higher ceiling, particularly evident in challenging scenarios characterized by large scale variations. With its outstanding performance and efficiency, we believe that RCM holds considerable potential for research and application." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.231, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RCM: Conflict-Free Local Feature Matching with Dynamic View Switching" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.207 + ], + "angle": 0, + "content": "This work was supported by the SEU Innovation Capability Enhancement Plan for Doctoral Students under grant CXJH SEU 24128." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.23, + 0.323, + 0.246 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.26, + 0.786, + 0.302 + ], + "angle": 0, + "content": "1. Balntas, V., Lenc, K., Vedaldi, A., Mikolajczyk, K.: Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In: Proceedings of the CVPR. pp. 5173-5182 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.303, + 0.786, + 0.329 + ], + "angle": 0, + "content": "2. Barroso-Laguna, A., Tian, Y., Mikolajczyk, K.: Scalenet: A shallow architecture for scale estimation. In: Proceedings of the CVPR. pp. 12808-12818 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.33, + 0.786, + 0.357 + ], + "angle": 0, + "content": "3. Bay, H., Tuytelaars, T., Van G., L.: SURF: Speeded Up Robust Features. In: Proceedings of the ECCV. pp. 404-417 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.358, + 0.786, + 0.384 + ], + "angle": 0, + "content": "4. Calonder, M., Lepetit, V., Strecha, C., Fua, P.: BRIEF: Binary Robust Independent Elementary Features. In: Proceedings of the ECCV. pp. 778-792 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.385, + 0.786, + 0.426 + ], + "angle": 0, + "content": "5. Chen, H., Luo, Z., Zhang, J., Zhou, L., Bai, X., Hu, Z., Tai, C., Quan, L.: Learning to match features with seeded graph matching network. In: Proceedings of the ICCV. pp. 6301-6310 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.427, + 0.786, + 0.468 + ], + "angle": 0, + "content": "6. Chen, H., Luo, Z., Zhou, L., Tian, Y., Zhen, M., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Aspanformer: Detector-free image matching with adaptive span transformer. In: Proceedings of the ECCV. pp. 20-36 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.469, + 0.786, + 0.495 + ], + "angle": 0, + "content": "7. Chen, Y., Huang, D., Xu, S., Liu, J., Liu, Y.: Guide local feature matching by overlap estimation. In: Proceedings of the AAAI. pp. 365-373 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.496, + 0.786, + 0.536 + ], + "angle": 0, + "content": "8. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the CVPR. pp. 5828-5839 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.537, + 0.786, + 0.577 + ], + "angle": 0, + "content": "9. DeTone, D., Malisiewicz, T., Rabinovich, A.: SuperPoint: Self-Supervised Interest Point Detection and Description. In: Proceedings of the CVPRW. pp. 224-236 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.578, + 0.786, + 0.619 + ], + "angle": 0, + "content": "0. Dusmanu, M., Rocco, I., Pajdla, T., Pollefeys, M., Sivic, J., Torii, A., Sattler, T.: D2-Net: A Trainable CNN for Joint Description and Detection of Local Features. In: Proceedings of the CVPR. pp. 8092–8101 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.786, + 0.66 + ], + "angle": 0, + "content": "1. Edstedt, J., Bokman, G., Wadenback, M., Felsberg, M.: Dedode: Detect, don't describe—describe, don't detect for local feature matching. In: Proceedings of the 3DV. pp. 148-157 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.786, + 0.688 + ], + "angle": 0, + "content": "2. Edstedt, J., Sun, Q., Bokman, G., Wadenback, M., Felsberg, M.: Roma: Robust dense feature matching. In: Proceedings of the CVPR. pp. 19790-19800 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.786, + 0.716 + ], + "angle": 0, + "content": "3. Engel, J., Koltun, V., Cremers, D.: Direct sparse odometry. IEEE Transactions on Pattern Analysis and Machine Intelligence 40(3), 611-625 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.786, + 0.757 + ], + "angle": 0, + "content": "4. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM pp. 381-395 (1981)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.786, + 0.784 + ], + "angle": 0, + "content": "5. Fu, Y., Wu, Y.: Scale-net: Learning to reduce scale differences for large-scale invariant image matching. arXiv preprint arXiv:2112.10485 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.786, + 0.812 + ], + "angle": 0, + "content": "6. Germain, H., Bourmaud, G., Lepetit, V.: Sparse-to-dense hypercolumn matching for long-term visual localization. In: Proceedings of the 3DV. pp. 513-523 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.786, + 0.84 + ], + "angle": 0, + "content": "7. Germain, H., Bourmaud, G., Lepetit, V.: S2dnet: Learning image features for accurate sparse-to-dense matching. In: Proceedings of the ECCV. pp. 626-643 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.26, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "X. Lu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "18. Gleize, P., Wang, W., Feiszli, M.: Silk: Simple learned keypoints. In: Proceedings of the CVPR. pp. 22499-22508 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "19. Huang, D., Chen, Y., Liu, Y., Liu, J., Xu, S., Wu, W., Ding, Y., Tang, F., Wang, C.: Adaptive assignment for geometry aware local feature matching. In: Proceedings of the CVPR. pp. 5425-5434 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.219, + 0.785, + 0.26 + ], + "angle": 0, + "content": "20. Jiang, W., Trulls, E., Hosang, J., Tagliasacchi, A., Yi, K.M.: Cotr: Correspondence transformer for matching across images. In: Proceedings of the ICCV. pp. 6207-6217 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.261, + 0.785, + 0.302 + ], + "angle": 0, + "content": "21. Katharopoulos, A., Vyas, A., Pappas, N., Fleuret, F.: Transformers are rnns: Fast autoregressive transformers with linear attention. In: Proceedings of the ICML. pp. 5156-5165 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.303, + 0.785, + 0.331 + ], + "angle": 0, + "content": "22. Li, X., Han, K., Li, S., Prisacariu, V.: Dual-resolution correspondence networks. In: Proceedings of the NeurIPS. pp. 17346-17357 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.332, + 0.785, + 0.36 + ], + "angle": 0, + "content": "23. Li, Z., Snavely, N.: MegaDepth: Learning Single-View Depth Prediction from Internet Photos. In: Proceedings of the CVPR. pp. 2041–2050 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.361, + 0.785, + 0.388 + ], + "angle": 0, + "content": "24. Lindenberger, P., Sarlin, P., Pollefeys, M.: Lightglue: Local feature matching at light speed. In: Proceedings of the ICCV. pp. 17627-17638 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.389, + 0.785, + 0.417 + ], + "angle": 0, + "content": "25. Liu, C., Yuen, J., Torralba, A., Sivic, J., Freeman, W.T.: Sift flow: Dense correspondence across different scenes. In: Proceedings of the ECCV. pp. 28-42 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.418, + 0.785, + 0.445 + ], + "angle": 0, + "content": "26. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.446, + 0.785, + 0.473 + ], + "angle": 0, + "content": "27. Lowe, D.G.: Distinctive Image Features from Scale-Invariant Keypoints. International Journal of Computer Vision 60(2), 91-110 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.785, + 0.502 + ], + "angle": 0, + "content": "28. Lu, X., Yan, Y., Kang, B., Du, S.: Paraformer: Parallel attention transformer for efficient feature matching. arXiv preprint arXiv:2303.00941 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.503, + 0.785, + 0.544 + ], + "angle": 0, + "content": "29. Luo, Z., Zhou, L., Bai, X., Chen, H., Zhang, J., Yao, Y., Li, S., Fang, T., Quan, L.: Aslfeat: Learning local features of accurate shape and localization. In: Proceedings of the CVPR. pp. 6589-6598 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.545, + 0.785, + 0.587 + ], + "angle": 0, + "content": "30. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. Proceedings of the NeurIPS 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.588, + 0.785, + 0.628 + ], + "angle": 0, + "content": "31. Revaud, J., De Souza, C., Humenberger, M., Weinzaepfel, P.: R2D2: Reliable and Repeatable Detector and Descriptor. In: Proceedings of the NeurIPS. pp. 12414-12424 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.629, + 0.785, + 0.67 + ], + "angle": 0, + "content": "32. Rocco, I., Arandjelovic, R., Sivic, J.: Efficient neighbourhood consensus networks via submanifold sparse convolutions. In: Proceedings of the ECCV. pp. 605-621 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.671, + 0.785, + 0.699 + ], + "angle": 0, + "content": "33. Rocco, I., Cimpoi, M., Arandjelović, R., Torii, A., Pajdla, T., Sivic, J.: Neighbourhood consensus networks. In: Proceedings of the NeurIPS. pp. 1658-1669 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.785, + 0.728 + ], + "angle": 0, + "content": "34. Rublee, E., Rabaud, V., Konolige, K., Bradski, G.: ORB: An efficient alternative to SIFT or SURF. In: Proceedings of the ICCV. pp. 2564-2571 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.785, + 0.77 + ], + "angle": 0, + "content": "35. Sarlin, P., Cadena, C. and Siegwart, R., Dymczyk, M.: From coarse to fine: Robust hierarchical localization at large scale. In: Proceedings of the CVPR. pp. 12716-12725 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.785, + 0.812 + ], + "angle": 0, + "content": "36. Sarlin, P., DeTone, D., Malisiewicz, T., Rabinovich, A.: SuperGlue: Learning Feature Matching With Graph Neural Networks. In: Proceedings of the CVPR. pp. 4937-4946 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.841 + ], + "angle": 0, + "content": "37. Schonberger, J.L., Frahm, J.: Structure-from-motion revisited. In: Proceedings of the CVPR. pp. 4104-4113 (2016)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.231, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RCM: Conflict-Free Local Feature Matching with Dynamic View Switching" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "38. Shi, Y., Cai, J., Shavit, Y., Mu, T., Feng, W., Zhang, K.: Clustergnn: Cluster-based coarse-to-fine graph neural network for efficient feature matching. In: Proceedings of the CVPR. pp. 12517-12526 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "39. Sun, J., Shen, Z., Wang, Y., Bao, H., Zhou, X.: LoFTR: Detector-Free Local Feature Matching With Transformers. In: Proceedings of the CVPR. pp. 8922-8931 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.788, + 0.272 + ], + "angle": 0, + "content": "40. Truong, P., Danelljan, M., Timofte, R., Van Gool, L.: Pdc-net+: Enhanced probabilistic dense correspondence network. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.788, + 0.313 + ], + "angle": 0, + "content": "41. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is All you Need. In: Proceedings of the NeurIPS. pp. 5998-6008 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.788, + 0.355 + ], + "angle": 0, + "content": "42. Wang, Q., Zhang, J., Yang, K., Peng, K., Stiefelhagen, R.: Matchformer: Interleaving attention in transformers for feature matching. In: Proceedings of the ACCV. pp. 2746-2762 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.788, + 0.384 + ], + "angle": 0, + "content": "43. Wang, Q., Zhou, X., Hariharan, B., Snavely, N.: Learning feature descriptors using camera pose supervision. In: Proceedings of the ECCV. pp. 757-774 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.385, + 0.788, + 0.425 + ], + "angle": 0, + "content": "44. Zhang, J., Sun, D., Luo, Z., Yao, A., Zhou, L., Shen, T., Chen, Y., Quan, L., Liao, H.: Learning two-view correspondences and geometry using order-aware network. In: Proceedings of the ICCV. pp. 5845-5854 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.426, + 0.788, + 0.467 + ], + "angle": 0, + "content": "45. Zhang, Z., Sattler, T., Scaramuzza, D.: Reference pose generation for long-term visual localization via learned features and view synthesis. International Journal of Computer Vision pp. 821-844 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.788, + 0.508 + ], + "angle": 0, + "content": "46. Zhao, X., Wu, X., Chen, W., Chen, P.C., Xu, Q., Li, Z.: A liked: A lighter keypoint and descriptor extraction network via deformable transformation. IEEE Transactions on Instrumentation and Measurement 72, 1-16 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.788, + 0.536 + ], + "angle": 0, + "content": "47. Zhou, Q., Sattler, T., Leal-Taixe, L.: Patch2pix: Epipolar-guided pixel-level correspondences. In: Proceedings of the CVPR. pp. 4669-4678 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.536 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Raising the Ceiling_ Conflict-Free Local Feature Matching with Dynamic View Switching/f888146a-2080-4739-ac33-923c8980cbb0_origin.pdf b/2024/Raising the Ceiling_ Conflict-Free Local Feature Matching with Dynamic View Switching/f888146a-2080-4739-ac33-923c8980cbb0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a75f1788dcc526f8f06a153a73ad508f4d4c041a --- /dev/null +++ b/2024/Raising the Ceiling_ Conflict-Free Local Feature Matching with Dynamic View Switching/f888146a-2080-4739-ac33-923c8980cbb0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69984dff12644e95210c3f6e5e3e13511b4b295668f6ca1d80bb698e93039430 +size 5709829 diff --git a/2024/Raising the Ceiling_ Conflict-Free Local Feature Matching with Dynamic View Switching/full.md b/2024/Raising the Ceiling_ Conflict-Free Local Feature Matching with Dynamic View Switching/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4b5302a1406644677b9df74e407106ee12fa2d54 --- /dev/null +++ b/2024/Raising the Ceiling_ Conflict-Free Local Feature Matching with Dynamic View Switching/full.md @@ -0,0 +1,319 @@ +# Raising the Ceiling: Conflict-Free Local Feature Matching with Dynamic View Switching + +Xiaoyong Lu and Songlin Du\* + +School of Automation, Southeast University, Nanjing, China {luxiaoyong,sdu}@seu.edu.cn + +Abstract. Current feature matching methods prioritize improving modeling capabilities to better align outputs with ground-truth matches, which are the theoretical upper bound on matching results, metaphorically depicted as the "ceiling". However, these enhancements fail to address the underlying issues that directly hinder ground-truth matches, including the scarcity of matchable points in small scale images, matching conflicts in dense methods, and the keypoint-repeatability reliance in sparse methods. We propose a novel feature matching method named RCM, which Raises the Ceiling of Matching from three aspects. 1) RCM introduces a dynamic view switching mechanism to address the scarcity of matchable points in source images by strategically switching image pairs. 2) RCM proposes a conflict-free coarse matching module, addressing matching conflicts in the target image through a many-to-one matching strategy. 3) By integrating the semi-sparse paradigm and the coarse-to-fine architecture, RCM preserves the benefits of both high efficiency and global search, mitigating the reliance on keypoint repeatability. As a result, RCM enables more matchable points in the source image to be matched in an exhaustive and conflict-free manner in the target image, leading to a substantial $260\%$ increase in ground-truth matches. Comprehensive experiments show that RCM exhibits remarkable performance and efficiency in comparison to state-of-the-art methods. + +Keywords: Feature Matching $\cdot$ Camera Pose Estimation $\cdot$ Transformer + +# 1 Introduction + +Feature matching, which aims to establish accurate correspondences between two images, underpins various pivotal tasks in computer vision, such as object recognition [25], structure from motion (SfM) [37], and simultaneous localization and mapping (SLAM) [13]. However, challenges arise from factors such as viewpoint change, illumination variation, scale variation, occlusion and motion blur between images. These complexities elevate the difficulty of feature matching, particularly when real-time performance is required for practical applications. + +Presently, methods for feature matching can be broadly categorized into two groups: sparse matching methods and dense matching methods. Sparse methods [5, 28, 36, 38] are designed to establish correspondences between two sets of + +![](images/b000a93906636589f5fefc3ff86b1534db7feba3276d062781b789ff99085c5d.jpg) +Fig.1: Comparison among RCM, $\mathrm{RCM}_{\mathrm{Lite}}$ , SuperGlue [36], LoFTR [39] and MatchFormer [42]. Full RCM significantly accelerates LoFTR by $2.2\times$ while achieving superior performance. The lightweight $\mathrm{RCM}_{\mathrm{Lite}}$ outperforms SuperGlue by $+7.5\%$ while maintaining a speed advantage. The same color indicates the matched features in the visualization. + +![](images/bc961c5aeabb66e8a5b7385cc55b39ed1a6bd10ce8780e845444600d5a19cdf2.jpg) + +keypoints. Consequently, the incorporation of detectors [9, 27, 31, 34] becomes essential in this process. In contrast, dense methods [20, 33, 39, 47] undertake a direct pursuit of matches between the dense features. + +Through a comprehensive examination of prevailing sparse and dense matching methods, we identify and subsequently address three key issues that constrain the theoretical upper bound, i.e., the ceiling, on the number of matches. + +(1) Sparse and dense methods: dramatic decrease in the number of matchable keypoints in small scale images. Overlapping regions in small scale images are limited in size, with the majority of keypoints or grid points situated in irrelevant background, which not only causes redundancy in the message passing but also radically limits the number of ground-truth matches. +For the semi-sparse [16, 17] and dense [39] matching paradigms, this challenge only arises in the source image as they perform a dense search in the target image. To tackle this challenge, we propose a dynamic view switcher that learns to switch the larger scale image to the source image. This strategic switch significantly augments the number of matchable keypoints within the overlapping region to radically raise the theoretical upper bound of the matcher, as illustrated in Fig. 6(b). +(2) Dense method: matching conflicts in large scale variation scenes caused by the one-to-one matching strategy. As shown in Fig. 2(b), the case of many-to-one matching arises when dealing with significant scale variations, as more pixels in large scale images correspond to fewer pixels in small scale images. However, existing semi-dense methods struggle with this challenge as they employ a one-to-one matching strategy in the coarse matching stage, generating at most one coarse match within each $8 \times 8$ region in the target image while discarding other valid matches. +In our method, correspondences are established between the source and target image in a many-to-one manner. Each matchable point can independently search for correspondences without conflicts, as depicted in Fig. 2(c) and Fig. 6(a). + +![](images/74a42b21438e3071e8987226b3e027aa81a2c3f70425473f87bc84d56f547560.jpg) +Fig. 2: Comparison of three matching methods. (a) The sparse methods rely on precise keypoint detection at corresponding points in both images. (b) We highlight the issues of matching conflicts introduced in Sec. 1. (c) RCM establishes conflict-free coarse matches and adjusts them independently in the fine matching stage. + +(3) Sparse method: heavy reliance on the repeatability of keypoint. Sparse methods strategically narrow the search scope for matchers by leveraging detectors, yet they concurrently establish a severe reliance on the repeatability of keypoints. When input keypoints are not precisely detected at corresponding positions in both images and result in few ground-truth matches in challenging scenes, even a perfect sparse matcher fails to "create something out of nothing". + +This challenge inspires the semi-sparse paradigm [16,17] to detect keypoints in the source image and perform a dense search in the target image, eliminating the requirement for precise detection of keypoints in both images. We adopt the semi-sparse paradigm and integrate it seamlessly with the coarse-to-fine architecture [39], allowing RCM to maintain efficiency while globally searching for sub-pixel matches in the target image. + +To summarize, this paper presents the following contributions: + +- We seamlessly integrate the semi-sparse paradigm with the coarse-to-fine architecture, establishing a strong foundation for our subsequent efforts to raise the theoretical upper bound of feature matching. +- We design a view switcher that dynamically switches the source and target images based on their scales. This view switcher offers a direct and effective method for increasing the number of matchable points in the source image. +- We propose a conflict-free coarse matching module, which matches two sets of features in a many-to-one fashion, further breaking through the match quantity barrier in large scale variation scenes. +- The proposed RCM and $\mathrm{RCM}_{\mathrm{Lite}}$ achieve an excellent performance-efficiency balance, rendering the matchers suitable for a wider range of applications. + +# 2 Related Works + +# 2.1 Local Feature Matching + +Local feature matching methods are categorized as sparse and dense methods. Sparse methods involve the extraction of sparse features with a keypoint detec + +tor, followed by establishing correspondences through a matching process. Dense methods directly establish correspondences between dense features. + +Sparse matching contains well-known methods like SIFT [27], SURF [3], BRIEF [4] and ORB [34], which are widely utilized. The advent of deep learning introduces learning-based descriptors as a dominant approach. Learning-based detectors and descriptors [9-11,18,31,46] harness the power of convolutional neural networks (CNN) for improved robustness. SuperGlue [36] pioneered global relationship modeling among sparse features using Transformer [41], which enriches features via self- and cross-attention mechanisms. + +Dense matching methods [6,39,42] eschew the detector and instead establish dense correspondences between images, resulting in a significant improvement over sparse methods. We attribute this advancement to the increase in ground-truth matches facilitated by coarse-to-fine matching, which effectively covers the entire target image. This success in raising the feature matching ceiling inspires RCM to inherit the coarse-to-fine matching approach and pursue further enhancements in ground-truth matches. + +Semi-Sparse matching methods, matching sparse and dense features asymmetrically, are proposed by S2DHM [16] and S2DNet [17]. These methods, while preserving sparse keypoints in the source image, use dense features at full image resolution in the target image, resulting in excessive memory usage. In our approach, we integrate the semi-sparse matching paradigm with coarse-to-fine matching [39] to maintain high efficiency. Additionally, the proposed dynamic view switching and many-to-one matching empower the semi-sparse paradigm to achieve performance comparable to modern sparse and dense paradigms. + +# 2.2 Feature Matching with Scale Invariance + +Feature matching struggles in scenes with large scale variations, primarily due to the scarcity of matchable points in small scale images and matching conflicts due to one-to-one matching. Many methods are proposed to align scales before matching. OETR [7] identifies overlapping regions between images with object detection. ScaleNet [2] and Scale-Net [15] directly predict the scale variation. To address the scale challenge, we divide scenes with large scale variations into two cases: source and target images with (1) small and large scales and (2) large and small scales. In the first case, both sparse [5,36] and dense [39,42] methods face a reduction in matchable points in the source image, which is solved by the proposed dynamic view switcher. In the second case, dense methods encounter matching conflicts in small scale target images, a challenge mitigated by the proposed many-to-one matching strategy. Similar to many-to-one matching, AdaMatcher [19] addresses the matching conflict in the second case through adaptive assignment. However, it is limited to one-to-one matching in the first case, where the scarcity of matchable points in the source image persists. With the combination of dynamic view switcher and many-to-one matching, we can effectively obtain more matchable points in the source image and perform conflict-free matching in the target image in all scenes with large scale variations. + +![](images/8192036d02f81b484985c511c3596d1adc80580b752026974d62e7beda6a1325.jpg) +Fig. 3: RCM is comprised of four primary modules: 1. The U-Net Extraction Module extracts multi-scale coarse and fine features. 2. The View Switching Module dynamically switches the larger scale image to the sparse branch, where dense coarse features are transformed into sparse features through the detection head. 3. The Conflict-Free Coarse Matching Module involves attention layers processing two sets of coarse features, which are subsequently matched via the many-to-one matching layer. 4. The Fine Matching Module further refines coarse matches based on fine features. + +# 3 Methodology + +# 3.1 Overview + +The overall architecture of RCM is illustrated in Fig. 3. The U-Net extraction module initiates the process by employing CNN to extract coarse and fine features from both images. Subsequently, the coarse features of the two images undergo processing within the view switcher. In this module, the coarse features of larger scale images are switched to the sparse branch and extracted by the detection head as sparse features. And the coarse features of smaller scale images are retained as dense features. The sparse and dense features are then matched in a many-to-one fashion through the conflict-free coarse matching module. Finally the fine matching module crops out the fine features in the coarse matching positions and refines coarse matches through the correlation-based approach [39]. + +# 3.2 Local Feature Extraction + +We extend the VGG-style encoder from SuperPoint [9] to a U-Net architecture to extract multi-scale features from images $I_{A}$ and $I_{B}$ . The decoder output is employed as fine features, denoted as $F_{A}^{f}\in \mathbb{R}^{H_{f}\times W_{f}\times C_{2}}$ and $F_{B}^{f}\in \mathbb{R}^{H_{f}\times W_{f}\times C_{2}}$ . For brevity, we omit the $A$ and $B$ subscripts in the rest of the paper. On the other hand, the coarse features, denoted as $F^c\in \mathbb{R}^{H_c\times W_c\times C_1}$ , are formed by merging features from multiple resolutions, enhancing the robustness of the conflict-free coarse matching module. + +# 3.3 View Switching Module + +As shown in Fig. 6(b), employing a small scale image as the source image leads to a scarcity of matchable points, which consequently leads to fewer matches. As shown in Fig. 4, we propose a binary classification network named view switcher to determine whether to switch the source and target images, where a lightweight CNN is employed to process the correlation map of coarse features $F_{A}^{c}$ and $F_{B}^{c}$ to perceive scale variation. The result of view switcher $VS$ is computed as + +![](images/65ed5ebf5c532371f7409796faec5700437d297da6b9264eddb9912fbc34ba0d.jpg) +Fig. 4: The view switcher perceives the scale variation between image pairs to predict a binary classification result, indicating whether to switch the images. + +$$ +V S = \operatorname {S o f t m a x} (\operatorname {C N N} (\operatorname {C o r r} (\operatorname {P o o l} (F _ {A} ^ {c}), \operatorname {P o o l} (F _ {B} ^ {c})))) \tag {1} +$$ + +where $\mathrm{Pool}(\cdot)$ represents average pooling and $\mathrm{Corr}(\cdot, \cdot)$ signifies similarity calculation through the inner product. The view switcher is trained to keep sparse branch processing larger scale images and dense branch processing smaller scale images. In the sparse branch, we employ the detection head in [9] to detect keypoints and sample the feature map at the keypoint positions as sparse features $F_{s}^{c}$ . The U-Net encoder and detection head are inherited from SuperPoint [9] and are frozen, making the detection equivalent to SuperPoint. Assuming that $N$ keypoints are detected, features in the sparse branch are reduced from $\mathbb{R}^{H_c \times W_c \times C_1}$ to $\mathbb{R}^{N \times C_1}$ . + +# 3.4 Conflict-Free Coarse Matching + +Position Encoder Given that the attention mechanism lacks awareness of the positions of the features, we embed positional information into the coarse features $F^c$ through a position encoder PE, in which a lightweight MLP progressively expands the positional information $P \in \mathbb{R}^{N \times 3}$ into the coarse feature dimension $C_1$ and updates the coarse features. $P$ encompasses the $x$ and $y$ coordinates of keypoints, and their confidence scores $s$ generated by the detection head. + +Attention Layer The self- and cross-attention layers are employed for the global enhancement of sparse and dense features. In contrast to dense methods that use linear attention [21] for efficiency, the semi-sparse matching paradigm allows us to adopt vanilla attention [41], which is denoted as + +$$ +M = \operatorname {A t t e n t i o n} (Q, K, V) = \operatorname {S o f t m a x} \left(Q K ^ {T}\right) V. \tag {2} +$$ + +The attentional message is denoted as $M$ . In self-attention, $Q, K, V$ are three linear projections of a single feature set. In cross-attention, $Q$ is projected from one feature set, while $K$ and $V$ are projected from another feature set. + +The feed-forward network $\mathrm{MLP_f}$ processes the attentional message $M$ to update the features as + +$$ +F ^ {i + 1} = F ^ {i} + \mathrm {M L P} _ {\mathrm {f}} \left(F ^ {i} | | M\right), \tag {3} +$$ + +where $F^i$ is the feature in layer $i$ and $||$ is the concatenation operation. We perform self- and cross-attention on sparse and dense features for $L_1$ iterations, resulting in enhanced features $\tilde{F}_s^c$ and $\tilde{F}_d^c$ that capture global information. + +Many-to-One Coarse Matching with Dustbin With globally enhanced features, we perform matching between sparse and dense features in a many-to-one manner. Initially, we utilize a set of learnable parameters as a dustbin, denoted as $bin \in \mathbb{R}^{1 \times C_1}$ , and concatenate it with the dense features $\hat{F}_d^c = \tilde{F}_d^c ||bin \in \mathbb{R}^{(H_cW_c + 1) \times C_1}$ . The dustbin is introduced to match with non-matchable sparse features, enabling our matcher to handle non-overlapping regions and occlusion. We compute the score matrix with dustbin $S_{bin}$ between the sparse feature $\tilde{F}_s^c \in \mathbb{R}^{N \times C_1}$ and dense feature with dustbin $\hat{F}_d^c \in \mathbb{R}^{(H_cW_c + 1) \times C_1}$ as + +$$ +S _ {b i n} = \tilde {F} _ {s} ^ {c} \hat {F} _ {d} ^ {c} / \tau \in \mathbb {R} ^ {N \times (H _ {c} W _ {c} + 1)}, \tag {4} +$$ + +where $\tau$ is a learnable temperature parameter. + +Current dense methods utilize the Dual-Softmax operator to compute the coarse matching probability $P_{c}$ , defined as $P_{c} = \mathrm{Softmax}_{row}(S) \cdot \mathrm{Softmax}_{col}(S) \in \mathbb{R}^{N \times H_{c}W_{c}}$ . The Dual-Softmax operator keeps the summation of rows and columns of $P_{c}$ less than or equal to 1. This limits each feature to match at most one feature, that is, one-to-one matching. However multiple points may correspond to one point in scenes with large scale variations, which leads to matching conflicts as shown in Fig. 2(b). To address this issue, a single Softmax operation on the score with dustbin $S_{bin}$ is employed to release the constraint on target images while maintaining the constraint on source images as + +$$ +P _ {c} = \operatorname {S o f t m a x} _ {r o w} \left(S _ {b i n}\right), \tag {5} +$$ + +where each sparse feature can only match one dense feature or the dustbin, while one dense feature can match multiple sparse features. The strategy of many-to-one matching with dustbin can be comprehended as independently assigning each sparse feature to either its corresponding dense feature or to the dustbin. As shown in Fig. 2(c), following conflict-free coarse matching, the positions of sparse features, which correspond to the same dense feature, are independently adjusted during the fine matching stage. + +Ultimately, features matched to the dustbin are discarded, and we proceed to select matches with matching probabilities surpassing the threshold $\theta_c$ . + +# 3.5 Fine Matching + +For each coarse match, we sample a single feature at the source fine feature map and crop a feature window of size $w \times w$ at the target fine feature map. Given the collection of $K$ coarse matches, we obtain fine features denoted as $\hat{F}_s^f \in \mathbb{R}^{K \times 1 \times C_2}$ and $\hat{F}_d^f \in \mathbb{R}^{K \times w^2 \times C_2}$ for the sparse and dense branches, respectively. + +Subsequently, $\hat{F}_s^f$ and $\hat{F}_d^f$ are processed by the cross-attention module for $L_{2}$ iterations. Inspired by [6], in order to capture the local context of dense features, we substitute the self-attention module with the integration of $3\times 3$ convolutions into the feed-forward network. For fine matching, the features are updated by + +$$ +F ^ {i + 1} = F ^ {i} + \operatorname {C o n v} _ {3} \left(F ^ {i} | | M\right). \tag {6} +$$ + +We compute the correlation map between the centroid of the source feature window and all the features in the target feature window, which represents the matching probability, and then get the fine matching position by computing the expectation of the probability distribution, as introduced in [39]. + +# 3.6 Supervision + +Our loss function encompasses four components: the coarse matching loss $\mathcal{L}_c^m$ , the coarse dustbin loss $\mathcal{L}_c^d$ , the fine matching loss $\mathcal{L}_f$ , and the view switching loss $\mathcal{L}_{vs}$ . Ground-truth matches $M_{gt} = \{(i_m,j_m)\}$ are generated from camera poses and depth maps, establishing many-to-one correspondence between sparse and dense features. And non-matchable sparse features $U_{gt} = \{i_u\}$ , which are keypoints without any ground-truth match, are also obtained. + +The coarse matching loss $\mathcal{L}_c^m$ is computed as the negative log-likelihood loss of the ground-truth matches $M_{gt}$ over the matching probability matrix $P_c$ of each attention layer, + +$$ +\mathcal {L} _ {c} ^ {m} = - \frac {1}{L} \sum_ {l} \left[ \frac {1}{| M _ {g t} |} \sum_ {(i, j) \in M _ {g t}} \log P _ {c} ^ {l} (i, j) \right]. \tag {7} +$$ + +The coarse dustbin loss $\mathcal{L}_c^d$ refers to the negative log-likelihood loss of the non-matchable sparse features $U_{gt}$ over the last column of the matching probability matrix $P_c$ , which signifies the matching probabilities with the dustbin, + +$$ +\mathcal {L} _ {c} ^ {d} = - \frac {1}{| U _ {g t} |} \sum_ {i \in U _ {g t}} \log P _ {c} (i, H _ {c} W _ {c} + 1). \tag {8} +$$ + +The fine matching loss $\mathcal{L}_f$ is computed by the L2-distance between each refined position and the ground-truth position, as introduced in [39]. + +To supervise the view switcher, we randomly select 500 ground-truth matches in each image pair and compute the distance between points in each image. The image with the larger average distance is identified as the larger scale image. If the original source image is the larger scale image, the ground-truth $VS_{gt}$ is false, and vice versa. The binary cross-entropy loss of the switcher result $VS$ is subsequently computed as + +$$ +\mathcal {L} _ {v s} = - \left[ V S _ {g t} \log (V S) + \left(1 - V S _ {g t}\right) \log (1 - V S) \right]. \tag {9} +$$ + +Table 1: Evaluation on HPatches [1] for homography estimation. We report the percentage of correctly estimated homographies with mean corner errors below $1/3/5$ pixels. We mark the best accuracy in bold. + +
CategoryMethodOverallIlluminationViewpoint
Accuracy (%,ε < 1/3/5 px)
SparseSuperPoint [9] + NN0.46 / 0.78 / 0.850.57 / 0.92 / 0.970.35 / 0.65 / 0.74
D2Net [10] + NN0.38 / 0.72 / 0.810.65 / 0.95 / 0.980.13 / 0.51 / 0.65
R2D2 [31] + NN0.47 / 0.78 / 0.830.63 / 0.93 / 0.980.33 / 0.64 / 0.70
ASLFeat [29] + NN0.48 / 0.81 / 0.880.63 / 0.94 / 0.980.34 / 0.69 / 0.78
SuperPoint + SuperGlue [36]0.51 / 0.83 / 0.890.62 / 0.93 / 0.980.41 / 0.73 / 0.81
SuperPoint + CAPS [43] + NN0.49 / 0.79 / 0.860.62 / 0.93 / 0.980.36 / 0.65 / 0.75
SIFT + CAPS [43] + NN0.36 / 0.76 / 0.850.48 / 0.89 / 0.950.26 / 0.65 / 0.76
SuperPoint + LightGlue [24]0.39 / 0.85 / 0.910.52 / 0.96 / 0.980.28 / 0.75 / 0.84
(Semi-)DenseSparseNCNet [32]0.36 / 0.66 / 0.760.62 / 0.92 / 0.970.13 / 0.41 / 0.57
NCNet [33]0.48 / 0.61 / 0.710.98 / 0.98 / 0.980.02 / 0.28 / 0.46
Patch2Pix [47]0.50 / 0.79 / 0.870.71 / 0.95 / 0.980.30 / 0.64 / 0.76
LoFTR [39]0.55 / 0.81 / 0.860.74 / 0.95 / 0.980.38 / 0.69 / 0.76
MatchFormer [42]0.55 / 0.81 / 0.870.75 / 0.95 / 0.980.37 / 0.68 / 0.78
Semi-SparseRCMLite0.50 / 0.86 / 0.910.55 / 0.95 / 0.980.45 / 0.79 / 0.85
RCM0.56 / 0.88 / 0.930.65 / 0.97 / 0.990.47 / 0.80 / 0.88
+ +# 4 Experiments + +# 4.1 Implementation Details + +Both RCM and $\mathrm{RCM}_{\mathrm{Lite}}$ consist of $L_{1} = 5$ layers of coarse attention and $L_{2} = 2$ layers of fine attention. The dimensions for coarse and fine features are $C_1 = 256$ and $C_2 = 64$ , respectively. The key distinction between RCM and $\mathrm{RCM}_{\mathrm{Lite}}$ lies in their coarse feature resolution and fine matching window size. For RCM, the coarse feature resolution is $1/8$ , and the fine matching window size is $w = 5$ . For $\mathrm{RCM}_{\mathrm{Lite}}$ , the coarse feature resolution is $1/16$ , and the fine matching window size is $w = 9$ . The learnable temperature $\tau$ is initialized to 0.1, and the coarse matching threshold $\theta_c$ is set to 0.2. The proposed method is implemented using Pytorch [30]. During the training phase, images are resized and padded to the size of $832 \times 832$ . We detect 1024 keypoints in the source image, and if there are insufficient keypoints, random keypoints are added to ensure efficient batching. RCM and $\mathrm{RCM}_{\mathrm{Lite}}$ are trained on the MegaDepth dataset [23] for 30 epochs with a batch size of 4 by AdamW optimizer [26] and are not fine-tuned for any other tasks. The initial learning rate is set to 0.0002, and the training employs the cosine decay learning rate scheduler with 1 epoch of linear warm-up. More details can be found in the Supplementary Material. + +# 4.2 Homography Estimation + +Dataset. The HPatches [1] dataset is widely recognized for evaluating local features and matchers. In line with the evaluation protocol in [10], we assess the performance of our method across 108 HPatches sequences, which comprise 52 sequences with illumination variations and 56 sequences with viewpoint changes. + +Metric. Following the metrics employed in [9, 36, 47], we report the percentage of correctly estimated homographies with mean corner errors below $1/3/5$ pixels. Vanilla RANSAC [14] is applied to estimate the homography for all methods. + +Results. As shown in Tab. 1, RCM achieves the best overall performance at all thresholds. Comparing the dense methods, RCM shows a significant advantage in accuracy in large viewpoint change scenarios, leading LoFTR and MatchFormer by $(+9\%, +11\%, +12\%)$ and $(+10\%, +12\%, +10\%)$ , respectively. We attribute this performance improvement to the increased matchable points in the source image facilitated by the view switcher, and the conflict-free matching achieved in the target image through many-to-one matching. Compared to the sparse method SuperGlue, RCM achieves an overall performance lead of $(+5\%, +5\%, +4\%)$ , benefiting from the dense search due to the semi-sparse paradigm and the sub-pixel accuracy due to the coarse-to-fine architecture. + +Table 2: Two-view pose estimation results on MegaDepth [23] in outdoor scenes. + +
CategoryMethodPose estimation
@5°@10°@20°
SparseSuperGlue [36]42.259.073.6
SGMNet [5]40.559.073.6
ClusterGNN [38]44.258.570.3
LightGlue [24]46.464.378.2
DeDoDe-B [11]49.165.077.1
DeDoDe-G [11]51.268.481.2
DenseDRC-Net [22]27.042.958.3
PDC-Net+(H) [40]43.161.976.1
LoFTR [39]52.869.281.2
RoMa [12]63.076.986.4
Semi-SparseRCMLite49.766.579.3
RCM53.269.481.5
+ +Table 3: Two-view pose estimation results on ScanNet [8] in indoor scenes. + +
CategoryMethodPose estimationAUC
@5°@10°@20°
SparseOANet [44]11.826.943.9
SuperGlue [36]16.233.851.8
SGMNet [5]15.432.148.3
LightGlue [24]*13.929.645.7
DenseDRC-Net [22]*7.717.930.5
MatchFormer [42]*15.832.048.0
LoFTR-OT [39]*16.933.650.6
Semi-SparseRCMLite*15.231.548.8
RCM*17.334.652.1
+ +# 4.3 Relative Pose Estimation + +Dataset. We evaluate matchers on the MegaDepth [23] and the ScanNet [8] dataset for outdoor and indoor pose estimation, respectively. For the MegaDepth dataset, all test images are resized such that their longer dimensions are 1216. For the ScanNet dataset, all test images are resized to $640 \times 480$ pixels. + +Metric. Based on the matching results, we apply vanilla RANSAC to solve the fundamental matrices and relative poses of the image pairs. The AUC of the pose error at thresholds $(5^{\circ}, 10^{\circ}, 20^{\circ})$ are reported, where pose error is the maximum angular error in rotation and translation. + +Results. As shown in Tab. 2, RCM achieves the second best performance on the MegaDepth dataset. RCM outperforms the state-of-the-art method LoFTR, demonstrating the benefits of the view switcher and many-to-one matching. The lightweight $\mathrm{RCM}_{\mathrm{Lite}}$ achieves a performance boost $(+7.5\%, +7.5\%, +5.7\%)$ compared to SuperGlue. Fig. 5 illustrates qualitative comparisons of the SuperGlue, + +![](images/9148033fa40eeb7596386c3b80ff5c80d3ba8236b4df17b55762abea518745a8.jpg) +Fig. 5: Qualitative comparison of SuperGlue [36], LoFTR [39] and proposed RCM on MegaDepth and ScanNet. + +LoFTR and RCM. More visualizations and analyses can be found in the Supplementary Material. + +As shown in Tab. 3, RCM achieves a $(0.4\%, 1\%, 1.5\%)$ performance improvement over LoFTR in the ScanNet dataset. It's worth noting that the superscript * indicates zero-shot performance, and the model is trained on MegaDepth. Nevertheless, our methods achieve performance comparable to state-of-the-art sparse methods trained on ScanNet. Dense and semi-sparse matching methods find matches in the target image exhaustively, mitigating the challenge of nonrepeatability in indoor scenes, a key limitation of sparse methods. + +# 4.4 Visual Localization + +Dataset. We assess the performance of our method using the Aachen DayNight v1.1 benchmark [45], a well-established dataset for visual localization. This dataset comprises 824 day-time images and 191 night-time images, chosen as queries for outdoor visual localization. + +Metric. We utilize the open-source localization pipeline HLoc [35] for conducting localization with RCM. We report the percentage of successfully localized images within three error thresholds $(0.25m, 2^{\circ}) / (0.5m, 5^{\circ}) / (1m, 10^{\circ})$ . + +Results. As shown in Tab. 4, RCM demonstrates comparable or superior performance compared to the dense methods LoFTR and ASpanFormer in outdoor visual localization tasks. Our approach exhibits strong generalization across various downstream tasks without the need for fine-tuning. It is noteworthy that the semi-sparse method RCM is significantly more efficient than the dense methods, achieving a runtime speedup of over twofold. + +Table 4: Visual localization on the Aachen Day-Night benchmark v1.1 [45]. + +
MethodDayNight
(0.25m, 2°) / (0.5m, 5°) / (1m, 10°)
Localization with matching pairs generated by HLoc
LoFTR [39]88.7 / 95.6 / 99.078.5 / 90.6 / 99.0
ASpanFormer [6]89.4 / 95.6 / 99.077.5 / 91.6 / 99.0
DeDoDe-B [11]87.4 / 94.7 / 98.570.7 / 88.0 / 97.9
DeDoDe-G [11]88.5 / 95.4 / 99.074.9 / 89.5 / 99.5
SP [9]+LightGlue [24]89.6 / 95.8 / 99.272.8 / 88.0 / 99.0
RCMLite87.7 / 94.7 / 98.472.3 / 91.1 / 98.4
RCM89.7 / 96.0 / 98.772.8 / 91.6 / 99.0
+ +![](images/f6616bcbb4d0320346338474f8bfffc51e74ef86b5a8a3e62d1f2d96c8f012fc.jpg) +Fig. 6: Ablation study on view switcher and many-to-one matching. (a) (b) shows the actual matching results and (c) (d) shows the ground-truth coarse matches. The harmonious combination of view switcher and many-to-one matching allows more matchable points in the source image and conflict-free matching in the target image. + +![](images/f1afd91a353ab4d556a8cece7cd08f2ad21cb1064bdf954fc6ae8e64438aad18.jpg) + +# 4.5 Ablation Study + +In this subsection, we conduct qualitative and quantitative ablation experiments on MegaDepth. All models in Tab. 5 are trained for 15 epochs at 544 resolution and tested at 1216 resolution. For the baseline model (1), we apply the original SuperPoint [9] to generate keypoints and descriptors for both images. O2O and M2O in Tab. 5 are abbreviations for one-to-one and many-to-one, respectively. + +Ground-Truth Match Analysis. Fig. 6(d) illustrates the core idea of RCM, which is to increase the theoretical upper bound on the matcher, i.e., ground-truth matches, through view switcher and many-to-one matching. Fig. 6(c) illustrates a quantitative comparison of ground-truth matches, showing an increase as the view switching accuracy improves during training. Eventually, RCM and $\mathrm{RCM}_{\mathrm{Lite}}$ achieve view switching accuracies of $90.9\%$ and $87.8\%$ , respectively. Through the integration of many-to-one matching and the view switcher, the ground-truth matches surge to $260\%$ compared to one-to-one matching, which is a compelling demonstration of raising the ceiling of feature matching. + +Table 5: Ablation study of each proposal on MegaDepth. + +
Method#MatchesPrec.
(1) Sparse sampling451.798.0
(2) Dense sampling932.193.5
(3) Semi-Sparse sampling492.393.7
(4) + U-Net Extraction507.894.2
(5) O2O → M2O770.7(+51.8%)97.2
(6) + Switcher (RCMLite)951.4(+23.4%)97.8
(7) 1/16 → 1/8 (RCM)961.498.1
+ +Table 6: Evaluation on extreme-scale MegaDepth. + +
Method#MatchesPrec.
SP [9] + SuperGlue [36]26175.5
SP [9] + LightGlue [24]31376.0
LoFTR [39]42453.5
MatchFormer [42]46477.0
ASpanFormer [6]42577.1
RCM88778.9
wo switcher65375.4
flipped switcher7654.2
+ +Sampling Strategy. As shown in Tab. 5, the baseline models (1), (2), and (3) differ solely in sampling strategies for coarse features. Precisely, (1) detects keypoints in both images, (2) employs dense feature as the coarse feature, and (3) detects keypoints in source image and employs dense feature in target image. + +Many-to-One Matching. As shown in Fig. 6(a), in contrast to the one-to-one matching strategy, the many-to-one matching strategy pursues conflict-free matching, resulting in more detailed matches. As shown in Tab. 5(5), substituting one-to-one matching with many-to-one matching improves the number of matches by $51.8\%$ and precision by $3\%$ . + +View Switcher. As shown in Fig. 6(b), without the view switcher, the matcher can only produce a small number of matches in small scale images because only a limited number of keypoints are detected in the overlapping region. The view switcher enables the detection of more matchable keypoints in the source image. As shown in Tab. 5(6), the average number of matches increases by $23.4\%$ with the view switcher. More analyses about view switcher and many-to-one matching can be found in the Supplementary Material. + +RCM and $\mathrm{RCM}_{\mathrm{Lite}}$ . The distinctions between RCM and $\mathrm{RCM}_{\mathrm{Lite}}$ lie in the coarse feature resolution and fine matching window. RCM employs coarse features at 1/8 resolution, consistent with the dense attention-based methods. We apply 1/16 resolution coarse features in $\mathrm{RCM}_{\mathrm{Lite}}$ to trade performance degradation for faster runtime, facilitating applications in real-time scenarios. Because of the additional layers at 1/16 resolution in the feature extraction network, $\mathrm{RCM}_{\mathrm{Lite}}$ consists of more parameters than RCM. Also to cover larger coarse feature patches, we design a larger fine matching window for $\mathrm{RCM}_{\mathrm{Lite}}$ . + +Robustness to Scale Variation. We evaluate matchers on the MegaDepth subset defined in OETR [7], where the overlap scale ratio is in $[4, +\infty]$ . As shown in Tab. 6, RCM achieves the best performance with $2.8\times$ and $2.1\times$ more matches than LightGlue and ASpanFormer respectively. The performance of RCM is severely degraded without the view switcher, demonstrating the significance of the view switcher in large scale variation scenarios. Flipping the switching result, i.e., actively taking small scale image as source image, leads to the worst results. + +Efficiency. The parameters, runtimes, and performances of state-of-the-art matchers are presented in Tab. 7. Notably, RCM and $\mathrm{RCM}_{\mathrm{Lite}}$ achieve the optimal balance between performance and efficiency. Although the performance of RCM lags behind the most precise matcher + +Table 7: Efficiency Analysis on RTX3090. The 1st, 2nd, and 3rd-best methods are highlighted. + +
CategoryMethodAUC@10°#Param.Time (ms)Avg. Rank
SparseDeDoDe [11]65.028.1M3224.7
LightGlue [24]64.311.9M723.7
DenseLoFTR [39]69.211.6M4543.7
RoMa [12]76.9111M9134.3
Semi-SparseRCMLite66.511.4M1342.7
RCM69.49.8M2082
+ +RoMa, RCM achieves respectable performance with only $8.8\%$ of the parameters and $23\%$ of the runtime, making it more suitable for real-time applications. + +# 5 Limitation and Discussion + +The view switcher assumes that one image has a relatively large scale, while both images being small in scale remains a tricky case in feature matching. A potential approach involves dynamically scaling the image pairs according to the overlapping regions, ensuring that the overlapping regions in two images remain simultaneously large and comparable in scale. + +Since the matching points in the target image are regression-based points rather than pre-detected points, RCM suffers from the lack of discrete tracks in the SfM task and requires additional adaptation strategy. + +The semi-sparse paradigm mitigates the keypoint-repeatability reliance, addressing only a portion of the broader keypoint dependence. Challenges in detection persist, such as low quantity and poor distribution, influencing the performance of RCM in texture-less scenes. We believe that the integration of the semi-sparse paradigm with advanced detection methods holds promising potential as a path forward for effective and efficient feature matching. + +# 6 Conclusion + +This paper introduces a novel perspective to enhance matchers by raising the theoretical upper bound. We propose the dynamic view switching mechanism for the source image and the conflict-free coarse matching layer for the target image, respectively. Through the integration of the view switcher and many-to-one matching layer, we enhance the number of matchable points in the source image while achieving conflict-free matching in the target image. By essentially augmenting the number of ground-truth matches between images, RCM raises the theoretical upper bound on the actual matching results. The results from qualitative and quantitative experiments demonstrate a significant enhancement in the practical performance of the matcher with a higher ceiling, particularly evident in challenging scenarios characterized by large scale variations. With its outstanding performance and efficiency, we believe that RCM holds considerable potential for research and application. + +# Acknowledgments + +This work was supported by the SEU Innovation Capability Enhancement Plan for Doctoral Students under grant CXJH SEU 24128. + +# References + +1. Balntas, V., Lenc, K., Vedaldi, A., Mikolajczyk, K.: Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In: Proceedings of the CVPR. pp. 5173-5182 (2017) +2. Barroso-Laguna, A., Tian, Y., Mikolajczyk, K.: Scalenet: A shallow architecture for scale estimation. In: Proceedings of the CVPR. pp. 12808-12818 (2022) +3. Bay, H., Tuytelaars, T., Van G., L.: SURF: Speeded Up Robust Features. In: Proceedings of the ECCV. pp. 404-417 (2006) +4. Calonder, M., Lepetit, V., Strecha, C., Fua, P.: BRIEF: Binary Robust Independent Elementary Features. In: Proceedings of the ECCV. pp. 778-792 (2010) +5. Chen, H., Luo, Z., Zhang, J., Zhou, L., Bai, X., Hu, Z., Tai, C., Quan, L.: Learning to match features with seeded graph matching network. In: Proceedings of the ICCV. pp. 6301-6310 (2021) +6. Chen, H., Luo, Z., Zhou, L., Tian, Y., Zhen, M., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Aspanformer: Detector-free image matching with adaptive span transformer. In: Proceedings of the ECCV. pp. 20-36 (2022) +7. Chen, Y., Huang, D., Xu, S., Liu, J., Liu, Y.: Guide local feature matching by overlap estimation. In: Proceedings of the AAAI. pp. 365-373 (2022) +8. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the CVPR. pp. 5828-5839 (2017) +9. DeTone, D., Malisiewicz, T., Rabinovich, A.: SuperPoint: Self-Supervised Interest Point Detection and Description. In: Proceedings of the CVPRW. pp. 224-236 (2018) +0. Dusmanu, M., Rocco, I., Pajdla, T., Pollefeys, M., Sivic, J., Torii, A., Sattler, T.: D2-Net: A Trainable CNN for Joint Description and Detection of Local Features. In: Proceedings of the CVPR. pp. 8092–8101 (2019) +1. Edstedt, J., Bokman, G., Wadenback, M., Felsberg, M.: Dedode: Detect, don't describe—describe, don't detect for local feature matching. In: Proceedings of the 3DV. pp. 148-157 (2024) +2. Edstedt, J., Sun, Q., Bokman, G., Wadenback, M., Felsberg, M.: Roma: Robust dense feature matching. In: Proceedings of the CVPR. pp. 19790-19800 (2024) +3. Engel, J., Koltun, V., Cremers, D.: Direct sparse odometry. IEEE Transactions on Pattern Analysis and Machine Intelligence 40(3), 611-625 (2017) +4. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM pp. 381-395 (1981) +5. Fu, Y., Wu, Y.: Scale-net: Learning to reduce scale differences for large-scale invariant image matching. arXiv preprint arXiv:2112.10485 (2021) +6. Germain, H., Bourmaud, G., Lepetit, V.: Sparse-to-dense hypercolumn matching for long-term visual localization. In: Proceedings of the 3DV. pp. 513-523 (2019) +7. Germain, H., Bourmaud, G., Lepetit, V.: S2dnet: Learning image features for accurate sparse-to-dense matching. In: Proceedings of the ECCV. pp. 626-643 (2020) + +18. Gleize, P., Wang, W., Feiszli, M.: Silk: Simple learned keypoints. In: Proceedings of the CVPR. pp. 22499-22508 (2023) +19. Huang, D., Chen, Y., Liu, Y., Liu, J., Xu, S., Wu, W., Ding, Y., Tang, F., Wang, C.: Adaptive assignment for geometry aware local feature matching. In: Proceedings of the CVPR. pp. 5425-5434 (2023) +20. Jiang, W., Trulls, E., Hosang, J., Tagliasacchi, A., Yi, K.M.: Cotr: Correspondence transformer for matching across images. In: Proceedings of the ICCV. pp. 6207-6217 (2021) +21. Katharopoulos, A., Vyas, A., Pappas, N., Fleuret, F.: Transformers are rnns: Fast autoregressive transformers with linear attention. In: Proceedings of the ICML. pp. 5156-5165 (2020) +22. Li, X., Han, K., Li, S., Prisacariu, V.: Dual-resolution correspondence networks. In: Proceedings of the NeurIPS. pp. 17346-17357 (2020) +23. Li, Z., Snavely, N.: MegaDepth: Learning Single-View Depth Prediction from Internet Photos. In: Proceedings of the CVPR. pp. 2041–2050 (2018) +24. Lindenberger, P., Sarlin, P., Pollefeys, M.: Lightglue: Local feature matching at light speed. In: Proceedings of the ICCV. pp. 17627-17638 (2023) +25. Liu, C., Yuen, J., Torralba, A., Sivic, J., Freeman, W.T.: Sift flow: Dense correspondence across different scenes. In: Proceedings of the ECCV. pp. 28-42 (2008) +26. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017) +27. Lowe, D.G.: Distinctive Image Features from Scale-Invariant Keypoints. International Journal of Computer Vision 60(2), 91-110 (2004) +28. Lu, X., Yan, Y., Kang, B., Du, S.: Paraformer: Parallel attention transformer for efficient feature matching. arXiv preprint arXiv:2303.00941 (2023) +29. Luo, Z., Zhou, L., Bai, X., Chen, H., Zhang, J., Yao, Y., Li, S., Fang, T., Quan, L.: Aslfeat: Learning local features of accurate shape and localization. In: Proceedings of the CVPR. pp. 6589-6598 (2020) +30. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. Proceedings of the NeurIPS 32 (2019) +31. Revaud, J., De Souza, C., Humenberger, M., Weinzaepfel, P.: R2D2: Reliable and Repeatable Detector and Descriptor. In: Proceedings of the NeurIPS. pp. 12414-12424 (2019) +32. Rocco, I., Arandjelovic, R., Sivic, J.: Efficient neighbourhood consensus networks via submanifold sparse convolutions. In: Proceedings of the ECCV. pp. 605-621 (2020) +33. Rocco, I., Cimpoi, M., Arandjelović, R., Torii, A., Pajdla, T., Sivic, J.: Neighbourhood consensus networks. In: Proceedings of the NeurIPS. pp. 1658-1669 (2018) +34. Rublee, E., Rabaud, V., Konolige, K., Bradski, G.: ORB: An efficient alternative to SIFT or SURF. In: Proceedings of the ICCV. pp. 2564-2571 (2011) +35. Sarlin, P., Cadena, C. and Siegwart, R., Dymczyk, M.: From coarse to fine: Robust hierarchical localization at large scale. In: Proceedings of the CVPR. pp. 12716-12725 (2019) +36. Sarlin, P., DeTone, D., Malisiewicz, T., Rabinovich, A.: SuperGlue: Learning Feature Matching With Graph Neural Networks. In: Proceedings of the CVPR. pp. 4937-4946 (2020) +37. Schonberger, J.L., Frahm, J.: Structure-from-motion revisited. In: Proceedings of the CVPR. pp. 4104-4113 (2016) + +38. Shi, Y., Cai, J., Shavit, Y., Mu, T., Feng, W., Zhang, K.: Clustergnn: Cluster-based coarse-to-fine graph neural network for efficient feature matching. In: Proceedings of the CVPR. pp. 12517-12526 (June 2022) +39. Sun, J., Shen, Z., Wang, Y., Bao, H., Zhou, X.: LoFTR: Detector-Free Local Feature Matching With Transformers. In: Proceedings of the CVPR. pp. 8922-8931 (2021) +40. Truong, P., Danelljan, M., Timofte, R., Van Gool, L.: Pdc-net+: Enhanced probabilistic dense correspondence network. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) +41. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is All you Need. In: Proceedings of the NeurIPS. pp. 5998-6008 (2017) +42. Wang, Q., Zhang, J., Yang, K., Peng, K., Stiefelhagen, R.: Matchformer: Interleaving attention in transformers for feature matching. In: Proceedings of the ACCV. pp. 2746-2762 (2022) +43. Wang, Q., Zhou, X., Hariharan, B., Snavely, N.: Learning feature descriptors using camera pose supervision. In: Proceedings of the ECCV. pp. 757-774 (2020) +44. Zhang, J., Sun, D., Luo, Z., Yao, A., Zhou, L., Shen, T., Chen, Y., Quan, L., Liao, H.: Learning two-view correspondences and geometry using order-aware network. In: Proceedings of the ICCV. pp. 5845-5854 (2019) +45. Zhang, Z., Sattler, T., Scaramuzza, D.: Reference pose generation for long-term visual localization via learned features and view synthesis. International Journal of Computer Vision pp. 821-844 (2021) +46. Zhao, X., Wu, X., Chen, W., Chen, P.C., Xu, Q., Li, Z.: A liked: A lighter keypoint and descriptor extraction network via deformable transformation. IEEE Transactions on Instrumentation and Measurement 72, 1-16 (2023) +47. Zhou, Q., Sattler, T., Leal-Taixe, L.: Patch2pix: Epipolar-guided pixel-level correspondences. In: Proceedings of the CVPR. pp. 4669-4678 (2021) \ No newline at end of file diff --git a/2024/Raising the Ceiling_ Conflict-Free Local Feature Matching with Dynamic View Switching/images.zip b/2024/Raising the Ceiling_ Conflict-Free Local Feature Matching with Dynamic View Switching/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4d809fc5601b39d493b8c4b4b0ea3bd6bccf449a --- /dev/null +++ b/2024/Raising the Ceiling_ Conflict-Free Local Feature Matching with Dynamic View Switching/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7fe72dd3beee4ef1a2b7bd63d9ef64307533a990c7b2414b4783563701258cc7 +size 701775 diff --git a/2024/Raising the Ceiling_ Conflict-Free Local Feature Matching with Dynamic View Switching/layout.json b/2024/Raising the Ceiling_ Conflict-Free Local Feature Matching with Dynamic View Switching/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9aeef9fa84c4ffa30885af784dde6f70b48a3e53 --- /dev/null +++ b/2024/Raising the Ceiling_ Conflict-Free Local Feature Matching with Dynamic View Switching/layout.json @@ -0,0 +1,9915 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "spans": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "text", + "content": "Raising the Ceiling: Conflict-Free Local Feature Matching with Dynamic View Switching" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 239, + 168, + 373, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 168, + 373, + 180 + ], + "spans": [ + { + "bbox": [ + 239, + 168, + 373, + 180 + ], + "type": "text", + "content": "Xiaoyong Lu and Songlin Du\\*" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 183, + 190, + 429, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 190, + 429, + 212 + ], + "spans": [ + { + "bbox": [ + 183, + 190, + 429, + 212 + ], + "type": "text", + "content": "School of Automation, Southeast University, Nanjing, China {luxiaoyong,sdu}@seu.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 236, + 452, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 236, + 452, + 465 + ], + "spans": [ + { + "bbox": [ + 160, + 236, + 452, + 465 + ], + "type": "text", + "content": "Abstract. Current feature matching methods prioritize improving modeling capabilities to better align outputs with ground-truth matches, which are the theoretical upper bound on matching results, metaphorically depicted as the \"ceiling\". However, these enhancements fail to address the underlying issues that directly hinder ground-truth matches, including the scarcity of matchable points in small scale images, matching conflicts in dense methods, and the keypoint-repeatability reliance in sparse methods. We propose a novel feature matching method named RCM, which Raises the Ceiling of Matching from three aspects. 1) RCM introduces a dynamic view switching mechanism to address the scarcity of matchable points in source images by strategically switching image pairs. 2) RCM proposes a conflict-free coarse matching module, addressing matching conflicts in the target image through a many-to-one matching strategy. 3) By integrating the semi-sparse paradigm and the coarse-to-fine architecture, RCM preserves the benefits of both high efficiency and global search, mitigating the reliance on keypoint repeatability. As a result, RCM enables more matchable points in the source image to be matched in an exhaustive and conflict-free manner in the target image, leading to a substantial " + }, + { + "bbox": [ + 160, + 236, + 452, + 465 + ], + "type": "inline_equation", + "content": "260\\%" + }, + { + "bbox": [ + 160, + 236, + 452, + 465 + ], + "type": "text", + "content": " increase in ground-truth matches. Comprehensive experiments show that RCM exhibits remarkable performance and efficiency in comparison to state-of-the-art methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 476, + 452, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 476, + 452, + 487 + ], + "spans": [ + { + "bbox": [ + 160, + 476, + 452, + 487 + ], + "type": "text", + "content": "Keywords: Feature Matching " + }, + { + "bbox": [ + 160, + 476, + 452, + 487 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 476, + 452, + 487 + ], + "type": "text", + "content": " Camera Pose Estimation " + }, + { + "bbox": [ + 160, + 476, + 452, + 487 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 476, + 452, + 487 + ], + "type": "text", + "content": " Transformer" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 505, + 230, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 505, + 230, + 517 + ], + "spans": [ + { + "bbox": [ + 133, + 505, + 230, + 517 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 528, + 481, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 481, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 481, + 612 + ], + "type": "text", + "content": "Feature matching, which aims to establish accurate correspondences between two images, underpins various pivotal tasks in computer vision, such as object recognition [25], structure from motion (SfM) [37], and simultaneous localization and mapping (SLAM) [13]. However, challenges arise from factors such as viewpoint change, illumination variation, scale variation, occlusion and motion blur between images. These complexities elevate the difficulty of feature matching, particularly when real-time performance is required for practical applications." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 612, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 612, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 612, + 482, + 648 + ], + "type": "text", + "content": "Presently, methods for feature matching can be broadly categorized into two groups: sparse matching methods and dense matching methods. Sparse methods [5, 28, 36, 38] are designed to establish correspondences between two sets of" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 654, + 234, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 654, + 234, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 654, + 234, + 665 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 242, + 220 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 242, + 220 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 242, + 220 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 242, + 220 + ], + "type": "image", + "image_path": "b000a93906636589f5fefc3ff86b1534db7feba3276d062781b789ff99085c5d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 228, + 482, + 284 + ], + "lines": [ + { + "bbox": [ + 130, + 228, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 482, + 284 + ], + "type": "text", + "content": "Fig.1: Comparison among RCM, " + }, + { + "bbox": [ + 130, + 228, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\mathrm{RCM}_{\\mathrm{Lite}}" + }, + { + "bbox": [ + 130, + 228, + 482, + 284 + ], + "type": "text", + "content": ", SuperGlue [36], LoFTR [39] and MatchFormer [42]. Full RCM significantly accelerates LoFTR by " + }, + { + "bbox": [ + 130, + 228, + 482, + 284 + ], + "type": "inline_equation", + "content": "2.2\\times" + }, + { + "bbox": [ + 130, + 228, + 482, + 284 + ], + "type": "text", + "content": " while achieving superior performance. The lightweight " + }, + { + "bbox": [ + 130, + 228, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\mathrm{RCM}_{\\mathrm{Lite}}" + }, + { + "bbox": [ + 130, + 228, + 482, + 284 + ], + "type": "text", + "content": " outperforms SuperGlue by " + }, + { + "bbox": [ + 130, + 228, + 482, + 284 + ], + "type": "inline_equation", + "content": "+7.5\\%" + }, + { + "bbox": [ + 130, + 228, + 482, + 284 + ], + "type": "text", + "content": " while maintaining a speed advantage. The same color indicates the matched features in the visualization." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 242, + 114, + 480, + 211 + ], + "blocks": [ + { + "bbox": [ + 242, + 114, + 480, + 211 + ], + "lines": [ + { + "bbox": [ + 242, + 114, + 480, + 211 + ], + "spans": [ + { + "bbox": [ + 242, + 114, + 480, + 211 + ], + "type": "image", + "image_path": "bc961c5aeabb66e8a5b7385cc55b39ed1a6bd10ce8780e845444600d5a19cdf2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 311, + 480, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 480, + 346 + ], + "type": "text", + "content": "keypoints. Consequently, the incorporation of detectors [9, 27, 31, 34] becomes essential in this process. In contrast, dense methods [20, 33, 39, 47] undertake a direct pursuit of matches between the dense features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 348, + 481, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 481, + 384 + ], + "type": "text", + "content": "Through a comprehensive examination of prevailing sparse and dense matching methods, we identify and subsequently address three key issues that constrain the theoretical upper bound, i.e., the ceiling, on the number of matches." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 385, + 482, + 667 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 130, + 385, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 482, + 445 + ], + "type": "text", + "content": "(1) Sparse and dense methods: dramatic decrease in the number of matchable keypoints in small scale images. Overlapping regions in small scale images are limited in size, with the majority of keypoints or grid points situated in irrelevant background, which not only causes redundancy in the message passing but also radically limits the number of ground-truth matches." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 447, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 482, + 532 + ], + "type": "text", + "content": "For the semi-sparse [16, 17] and dense [39] matching paradigms, this challenge only arises in the source image as they perform a dense search in the target image. To tackle this challenge, we propose a dynamic view switcher that learns to switch the larger scale image to the source image. This strategic switch significantly augments the number of matchable keypoints within the overlapping region to radically raise the theoretical upper bound of the matcher, as illustrated in Fig. 6(b)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 532, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 532, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 532, + 482, + 628 + ], + "type": "text", + "content": "(2) Dense method: matching conflicts in large scale variation scenes caused by the one-to-one matching strategy. As shown in Fig. 2(b), the case of many-to-one matching arises when dealing with significant scale variations, as more pixels in large scale images correspond to fewer pixels in small scale images. However, existing semi-dense methods struggle with this challenge as they employ a one-to-one matching strategy in the coarse matching stage, generating at most one coarse match within each " + }, + { + "bbox": [ + 130, + 532, + 482, + 628 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 130, + 532, + 482, + 628 + ], + "type": "text", + "content": " region in the target image while discarding other valid matches." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "In our method, correspondences are established between the source and target image in a many-to-one manner. Each matchable point can independently search for correspondences without conflicts, as depicted in Fig. 2(c) and Fig. 6(a)." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "X. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 479, + 199 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 479, + 199 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 479, + 199 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 479, + 199 + ], + "type": "image", + "image_path": "74a42b21438e3071e8987226b3e027aa81a2c3f70425473f87bc84d56f547560.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 208, + 482, + 253 + ], + "lines": [ + { + "bbox": [ + 130, + 208, + 482, + 253 + ], + "spans": [ + { + "bbox": [ + 130, + 208, + 482, + 253 + ], + "type": "text", + "content": "Fig. 2: Comparison of three matching methods. (a) The sparse methods rely on precise keypoint detection at corresponding points in both images. (b) We highlight the issues of matching conflicts introduced in Sec. 1. (c) RCM establishes conflict-free coarse matches and adjusts them independently in the fine matching stage." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 276, + 479, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 276, + 479, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 276, + 479, + 347 + ], + "type": "text", + "content": "(3) Sparse method: heavy reliance on the repeatability of keypoint. Sparse methods strategically narrow the search scope for matchers by leveraging detectors, yet they concurrently establish a severe reliance on the repeatability of keypoints. When input keypoints are not precisely detected at corresponding positions in both images and result in few ground-truth matches in challenging scenes, even a perfect sparse matcher fails to \"create something out of nothing\"." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 348, + 480, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 480, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 480, + 419 + ], + "type": "text", + "content": "This challenge inspires the semi-sparse paradigm [16,17] to detect keypoints in the source image and perform a dense search in the target image, eliminating the requirement for precise detection of keypoints in both images. We adopt the semi-sparse paradigm and integrate it seamlessly with the coarse-to-fine architecture [39], allowing RCM to maintain efficiency while globally searching for sub-pixel matches in the target image." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 146, + 420, + 422, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 420, + 422, + 432 + ], + "spans": [ + { + "bbox": [ + 146, + 420, + 422, + 432 + ], + "type": "text", + "content": "To summarize, this paper presents the following contributions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 439, + 480, + 571 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 138, + 439, + 480, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 439, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 138, + 439, + 480, + 475 + ], + "type": "text", + "content": "- We seamlessly integrate the semi-sparse paradigm with the coarse-to-fine architecture, establishing a strong foundation for our subsequent efforts to raise the theoretical upper bound of feature matching." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 475, + 480, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 475, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 475, + 480, + 510 + ], + "type": "text", + "content": "- We design a view switcher that dynamically switches the source and target images based on their scales. This view switcher offers a direct and effective method for increasing the number of matchable points in the source image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 511, + 480, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 511, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 138, + 511, + 480, + 546 + ], + "type": "text", + "content": "- We propose a conflict-free coarse matching module, which matches two sets of features in a many-to-one fashion, further breaking through the match quantity barrier in large scale variation scenes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 547, + 480, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 480, + 571 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 480, + 571 + ], + "type": "text", + "content": "- The proposed RCM and " + }, + { + "bbox": [ + 138, + 547, + 480, + 571 + ], + "type": "inline_equation", + "content": "\\mathrm{RCM}_{\\mathrm{Lite}}" + }, + { + "bbox": [ + 138, + 547, + 480, + 571 + ], + "type": "text", + "content": " achieve an excellent performance-efficiency balance, rendering the matchers suitable for a wider range of applications." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 597, + 242, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 242, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 242, + 609 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 621, + 281, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 281, + 634 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 281, + 634 + ], + "type": "text", + "content": "2.1 Local Feature Matching" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Local feature matching methods are categorized as sparse and dense methods. Sparse methods involve the extraction of sparse features with a keypoint detec" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 141, + 91, + 448, + 102 + ], + "type": "text", + "content": "RCM: Conflict-Free Local Feature Matching with Dynamic View Switching" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "tor, followed by establishing correspondences through a matching process. Dense methods directly establish correspondences between dense features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 224 + ], + "type": "text", + "content": "Sparse matching contains well-known methods like SIFT [27], SURF [3], BRIEF [4] and ORB [34], which are widely utilized. The advent of deep learning introduces learning-based descriptors as a dominant approach. Learning-based detectors and descriptors [9-11,18,31,46] harness the power of convolutional neural networks (CNN) for improved robustness. SuperGlue [36] pioneered global relationship modeling among sparse features using Transformer [41], which enriches features via self- and cross-attention mechanisms." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 225, + 481, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 225, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 481, + 308 + ], + "type": "text", + "content": "Dense matching methods [6,39,42] eschew the detector and instead establish dense correspondences between images, resulting in a significant improvement over sparse methods. We attribute this advancement to the increase in ground-truth matches facilitated by coarse-to-fine matching, which effectively covers the entire target image. This success in raising the feature matching ceiling inspires RCM to inherit the coarse-to-fine matching approach and pursue further enhancements in ground-truth matches." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 310, + 481, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 481, + 406 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 481, + 406 + ], + "type": "text", + "content": "Semi-Sparse matching methods, matching sparse and dense features asymmetrically, are proposed by S2DHM [16] and S2DNet [17]. These methods, while preserving sparse keypoints in the source image, use dense features at full image resolution in the target image, resulting in excessive memory usage. In our approach, we integrate the semi-sparse matching paradigm with coarse-to-fine matching [39] to maintain high efficiency. Additionally, the proposed dynamic view switching and many-to-one matching empower the semi-sparse paradigm to achieve performance comparable to modern sparse and dense paradigms." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 426, + 361, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 426, + 361, + 438 + ], + "spans": [ + { + "bbox": [ + 131, + 426, + 361, + 438 + ], + "type": "text", + "content": "2.2 Feature Matching with Scale Invariance" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": "Feature matching struggles in scenes with large scale variations, primarily due to the scarcity of matchable points in small scale images and matching conflicts due to one-to-one matching. Many methods are proposed to align scales before matching. OETR [7] identifies overlapping regions between images with object detection. ScaleNet [2] and Scale-Net [15] directly predict the scale variation. To address the scale challenge, we divide scenes with large scale variations into two cases: source and target images with (1) small and large scales and (2) large and small scales. In the first case, both sparse [5,36] and dense [39,42] methods face a reduction in matchable points in the source image, which is solved by the proposed dynamic view switcher. In the second case, dense methods encounter matching conflicts in small scale target images, a challenge mitigated by the proposed many-to-one matching strategy. Similar to many-to-one matching, AdaMatcher [19] addresses the matching conflict in the second case through adaptive assignment. However, it is limited to one-to-one matching in the first case, where the scarcity of matchable points in the source image persists. With the combination of dynamic view switcher and many-to-one matching, we can effectively obtain more matchable points in the source image and perform conflict-free matching in the target image in all scenes with large scale variations." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "X. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 114, + 471, + 284 + ], + "blocks": [ + { + "bbox": [ + 145, + 114, + 471, + 284 + ], + "lines": [ + { + "bbox": [ + 145, + 114, + 471, + 284 + ], + "spans": [ + { + "bbox": [ + 145, + 114, + 471, + 284 + ], + "type": "image", + "image_path": "8192036d02f81b484985c511c3596d1adc80580b752026974d62e7beda6a1325.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 293, + 482, + 371 + ], + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 371 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 371 + ], + "type": "text", + "content": "Fig. 3: RCM is comprised of four primary modules: 1. The U-Net Extraction Module extracts multi-scale coarse and fine features. 2. The View Switching Module dynamically switches the larger scale image to the sparse branch, where dense coarse features are transformed into sparse features through the detection head. 3. The Conflict-Free Coarse Matching Module involves attention layers processing two sets of coarse features, which are subsequently matched via the many-to-one matching layer. 4. The Fine Matching Module further refines coarse matches based on fine features." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 391, + 232, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 232, + 404 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 232, + 404 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 413, + 208, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 208, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 208, + 424 + ], + "type": "text", + "content": "3.1 Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 430, + 482, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 551 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 551 + ], + "type": "text", + "content": "The overall architecture of RCM is illustrated in Fig. 3. The U-Net extraction module initiates the process by employing CNN to extract coarse and fine features from both images. Subsequently, the coarse features of the two images undergo processing within the view switcher. In this module, the coarse features of larger scale images are switched to the sparse branch and extracted by the detection head as sparse features. And the coarse features of smaller scale images are retained as dense features. The sparse and dense features are then matched in a many-to-one fashion through the conflict-free coarse matching module. Finally the fine matching module crops out the fine features in the coarse matching positions and refines coarse matches through the correlation-based approach [39]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 564, + 287, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 564, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 564, + 287, + 575 + ], + "type": "text", + "content": "3.2 Local Feature Extraction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "content": "We extend the VGG-style encoder from SuperPoint [9] to a U-Net architecture to extract multi-scale features from images " + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "inline_equation", + "content": "I_{A}" + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "inline_equation", + "content": "I_{B}" + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "content": ". The decoder output is employed as fine features, denoted as " + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "inline_equation", + "content": "F_{A}^{f}\\in \\mathbb{R}^{H_{f}\\times W_{f}\\times C_{2}}" + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "inline_equation", + "content": "F_{B}^{f}\\in \\mathbb{R}^{H_{f}\\times W_{f}\\times C_{2}}" + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "content": ". For brevity, we omit the " + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "content": " subscripts in the rest of the paper. On the other hand, the coarse features, denoted as " + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "inline_equation", + "content": "F^c\\in \\mathbb{R}^{H_c\\times W_c\\times C_1}" + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "content": ", are formed by merging features from multiple resolutions, enhancing the robustness of the conflict-free coarse matching module." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 141, + 90, + 448, + 102 + ], + "type": "text", + "content": "RCM: Conflict-Free Local Feature Matching with Dynamic View Switching" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 281, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 281, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 281, + 128 + ], + "type": "text", + "content": "3.3 View Switching Module" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 282, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 282, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 282, + 317 + ], + "type": "text", + "content": "As shown in Fig. 6(b), employing a small scale image as the source image leads to a scarcity of matchable points, which consequently leads to fewer matches. As shown in Fig. 4, we propose a binary classification network named view switcher to determine whether to switch the source and target images, where a lightweight CNN is employed to process the correlation map of coarse features " + }, + { + "bbox": [ + 130, + 137, + 282, + 317 + ], + "type": "inline_equation", + "content": "F_{A}^{c}" + }, + { + "bbox": [ + 130, + 137, + 282, + 317 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 137, + 282, + 317 + ], + "type": "inline_equation", + "content": "F_{B}^{c}" + }, + { + "bbox": [ + 130, + 137, + 282, + 317 + ], + "type": "text", + "content": " to perceive scale variation. The result of view switcher " + }, + { + "bbox": [ + 130, + 137, + 282, + 317 + ], + "type": "inline_equation", + "content": "VS" + }, + { + "bbox": [ + 130, + 137, + 282, + 317 + ], + "type": "text", + "content": " is computed as" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 293, + 162, + 478, + 249 + ], + "blocks": [ + { + "bbox": [ + 293, + 162, + 478, + 249 + ], + "lines": [ + { + "bbox": [ + 293, + 162, + 478, + 249 + ], + "spans": [ + { + "bbox": [ + 293, + 162, + 478, + 249 + ], + "type": "image", + "image_path": "65ed5ebf5c532371f7409796faec5700437d297da6b9264eddb9912fbc34ba0d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 287, + 256, + 482, + 300 + ], + "lines": [ + { + "bbox": [ + 287, + 256, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 287, + 256, + 482, + 300 + ], + "type": "text", + "content": "Fig. 4: The view switcher perceives the scale variation between image pairs to predict a binary classification result, indicating whether to switch the images." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 196, + 328, + 481, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 328, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 196, + 328, + 481, + 342 + ], + "type": "interline_equation", + "content": "V S = \\operatorname {S o f t m a x} (\\operatorname {C N N} (\\operatorname {C o r r} (\\operatorname {P o o l} (F _ {A} ^ {c}), \\operatorname {P o o l} (F _ {B} ^ {c})))) \\tag {1}", + "image_path": "e6912762d980449cd272b0d52a7adfaed55b91ae5a21212f39d2c7840df85e4c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 353, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 482, + 460 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 353, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\mathrm{Pool}(\\cdot)" + }, + { + "bbox": [ + 130, + 353, + 482, + 460 + ], + "type": "text", + "content": " represents average pooling and " + }, + { + "bbox": [ + 130, + 353, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\mathrm{Corr}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 353, + 482, + 460 + ], + "type": "text", + "content": " signifies similarity calculation through the inner product. The view switcher is trained to keep sparse branch processing larger scale images and dense branch processing smaller scale images. In the sparse branch, we employ the detection head in [9] to detect keypoints and sample the feature map at the keypoint positions as sparse features " + }, + { + "bbox": [ + 130, + 353, + 482, + 460 + ], + "type": "inline_equation", + "content": "F_{s}^{c}" + }, + { + "bbox": [ + 130, + 353, + 482, + 460 + ], + "type": "text", + "content": ". The U-Net encoder and detection head are inherited from SuperPoint [9] and are frozen, making the detection equivalent to SuperPoint. Assuming that " + }, + { + "bbox": [ + 130, + 353, + 482, + 460 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 353, + 482, + 460 + ], + "type": "text", + "content": " keypoints are detected, features in the sparse branch are reduced from " + }, + { + "bbox": [ + 130, + 353, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{H_c \\times W_c \\times C_1}" + }, + { + "bbox": [ + 130, + 353, + 482, + 460 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 353, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{N \\times C_1}" + }, + { + "bbox": [ + 130, + 353, + 482, + 460 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 479, + 316, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 316, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 316, + 491 + ], + "type": "text", + "content": "3.4 Conflict-Free Coarse Matching" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 501, + 482, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 482, + 573 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 482, + 573 + ], + "type": "text", + "content": "Position Encoder Given that the attention mechanism lacks awareness of the positions of the features, we embed positional information into the coarse features " + }, + { + "bbox": [ + 130, + 501, + 482, + 573 + ], + "type": "inline_equation", + "content": "F^c" + }, + { + "bbox": [ + 130, + 501, + 482, + 573 + ], + "type": "text", + "content": " through a position encoder PE, in which a lightweight MLP progressively expands the positional information " + }, + { + "bbox": [ + 130, + 501, + 482, + 573 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 130, + 501, + 482, + 573 + ], + "type": "text", + "content": " into the coarse feature dimension " + }, + { + "bbox": [ + 130, + 501, + 482, + 573 + ], + "type": "inline_equation", + "content": "C_1" + }, + { + "bbox": [ + 130, + 501, + 482, + 573 + ], + "type": "text", + "content": " and updates the coarse features. " + }, + { + "bbox": [ + 130, + 501, + 482, + 573 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 501, + 482, + 573 + ], + "type": "text", + "content": " encompasses the " + }, + { + "bbox": [ + 130, + 501, + 482, + 573 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 501, + 482, + 573 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 501, + 482, + 573 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 501, + 482, + 573 + ], + "type": "text", + "content": " coordinates of keypoints, and their confidence scores " + }, + { + "bbox": [ + 130, + 501, + 482, + 573 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 501, + 482, + 573 + ], + "type": "text", + "content": " generated by the detection head." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 592, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 592, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 592, + 482, + 640 + ], + "type": "text", + "content": "Attention Layer The self- and cross-attention layers are employed for the global enhancement of sparse and dense features. In contrast to dense methods that use linear attention [21] for efficiency, the semi-sparse matching paradigm allows us to adopt vanilla attention [41], which is denoted as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 206, + 651, + 481, + 665 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 651, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 206, + 651, + 481, + 665 + ], + "type": "interline_equation", + "content": "M = \\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {S o f t m a x} \\left(Q K ^ {T}\\right) V. \\tag {2}", + "image_path": "7a435812e9c51cca8013911a338625efd3e6f9ab7a738df84ac75cfc1cbd1dbb.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "X. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "content": "The attentional message is denoted as " + }, + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "content": ". In self-attention, " + }, + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "inline_equation", + "content": "Q, K, V" + }, + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "content": " are three linear projections of a single feature set. In cross-attention, " + }, + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "content": " is projected from one feature set, while " + }, + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "content": " are projected from another feature set." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 481, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 481, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 481, + 176 + ], + "type": "text", + "content": "The feed-forward network " + }, + { + "bbox": [ + 130, + 152, + 481, + 176 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP_f}" + }, + { + "bbox": [ + 130, + 152, + 481, + 176 + ], + "type": "text", + "content": " processes the attentional message " + }, + { + "bbox": [ + 130, + 152, + 481, + 176 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 152, + 481, + 176 + ], + "type": "text", + "content": " to update the features as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 246, + 198, + 481, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 198, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 246, + 198, + 481, + 213 + ], + "type": "interline_equation", + "content": "F ^ {i + 1} = F ^ {i} + \\mathrm {M L P} _ {\\mathrm {f}} \\left(F ^ {i} | | M\\right), \\tag {3}", + "image_path": "5f40078406cd6fd88d87c79057afe74ae291a3345457f809e95437d8b0e453af.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 221, + 481, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 221, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 221, + 481, + 259 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 221, + 481, + 259 + ], + "type": "inline_equation", + "content": "F^i" + }, + { + "bbox": [ + 130, + 221, + 481, + 259 + ], + "type": "text", + "content": " is the feature in layer " + }, + { + "bbox": [ + 130, + 221, + 481, + 259 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 221, + 481, + 259 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 221, + 481, + 259 + ], + "type": "inline_equation", + "content": "||" + }, + { + "bbox": [ + 130, + 221, + 481, + 259 + ], + "type": "text", + "content": " is the concatenation operation. We perform self- and cross-attention on sparse and dense features for " + }, + { + "bbox": [ + 130, + 221, + 481, + 259 + ], + "type": "inline_equation", + "content": "L_1" + }, + { + "bbox": [ + 130, + 221, + 481, + 259 + ], + "type": "text", + "content": " iterations, resulting in enhanced features " + }, + { + "bbox": [ + 130, + 221, + 481, + 259 + ], + "type": "inline_equation", + "content": "\\tilde{F}_s^c" + }, + { + "bbox": [ + 130, + 221, + 481, + 259 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 221, + 481, + 259 + ], + "type": "inline_equation", + "content": "\\tilde{F}_d^c" + }, + { + "bbox": [ + 130, + 221, + 481, + 259 + ], + "type": "text", + "content": " that capture global information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 275, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 275, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 482, + 373 + ], + "type": "text", + "content": "Many-to-One Coarse Matching with Dustbin With globally enhanced features, we perform matching between sparse and dense features in a many-to-one manner. Initially, we utilize a set of learnable parameters as a dustbin, denoted as " + }, + { + "bbox": [ + 130, + 275, + 482, + 373 + ], + "type": "inline_equation", + "content": "bin \\in \\mathbb{R}^{1 \\times C_1}" + }, + { + "bbox": [ + 130, + 275, + 482, + 373 + ], + "type": "text", + "content": ", and concatenate it with the dense features " + }, + { + "bbox": [ + 130, + 275, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\hat{F}_d^c = \\tilde{F}_d^c ||bin \\in \\mathbb{R}^{(H_cW_c + 1) \\times C_1}" + }, + { + "bbox": [ + 130, + 275, + 482, + 373 + ], + "type": "text", + "content": ". The dustbin is introduced to match with non-matchable sparse features, enabling our matcher to handle non-overlapping regions and occlusion. We compute the score matrix with dustbin " + }, + { + "bbox": [ + 130, + 275, + 482, + 373 + ], + "type": "inline_equation", + "content": "S_{bin}" + }, + { + "bbox": [ + 130, + 275, + 482, + 373 + ], + "type": "text", + "content": " between the sparse feature " + }, + { + "bbox": [ + 130, + 275, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\tilde{F}_s^c \\in \\mathbb{R}^{N \\times C_1}" + }, + { + "bbox": [ + 130, + 275, + 482, + 373 + ], + "type": "text", + "content": " and dense feature with dustbin " + }, + { + "bbox": [ + 130, + 275, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\hat{F}_d^c \\in \\mathbb{R}^{(H_cW_c + 1) \\times C_1}" + }, + { + "bbox": [ + 130, + 275, + 482, + 373 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 238, + 380, + 481, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 380, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 238, + 380, + 481, + 396 + ], + "type": "interline_equation", + "content": "S _ {b i n} = \\tilde {F} _ {s} ^ {c} \\hat {F} _ {d} ^ {c} / \\tau \\in \\mathbb {R} ^ {N \\times (H _ {c} W _ {c} + 1)}, \\tag {4}", + "image_path": "b775304597dd0fe719ebdc9acda6c0fc5e7badcfd617301171a7e33aed5cb7fe.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 404, + 338, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 404, + 338, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 338, + 416 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 404, + 338, + 416 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 404, + 338, + 416 + ], + "type": "text", + "content": " is a learnable temperature parameter." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 417, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 417, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 417, + 482, + 525 + ], + "type": "text", + "content": "Current dense methods utilize the Dual-Softmax operator to compute the coarse matching probability " + }, + { + "bbox": [ + 130, + 417, + 482, + 525 + ], + "type": "inline_equation", + "content": "P_{c}" + }, + { + "bbox": [ + 130, + 417, + 482, + 525 + ], + "type": "text", + "content": ", defined as " + }, + { + "bbox": [ + 130, + 417, + 482, + 525 + ], + "type": "inline_equation", + "content": "P_{c} = \\mathrm{Softmax}_{row}(S) \\cdot \\mathrm{Softmax}_{col}(S) \\in \\mathbb{R}^{N \\times H_{c}W_{c}}" + }, + { + "bbox": [ + 130, + 417, + 482, + 525 + ], + "type": "text", + "content": ". The Dual-Softmax operator keeps the summation of rows and columns of " + }, + { + "bbox": [ + 130, + 417, + 482, + 525 + ], + "type": "inline_equation", + "content": "P_{c}" + }, + { + "bbox": [ + 130, + 417, + 482, + 525 + ], + "type": "text", + "content": " less than or equal to 1. This limits each feature to match at most one feature, that is, one-to-one matching. However multiple points may correspond to one point in scenes with large scale variations, which leads to matching conflicts as shown in Fig. 2(b). To address this issue, a single Softmax operation on the score with dustbin " + }, + { + "bbox": [ + 130, + 417, + 482, + 525 + ], + "type": "inline_equation", + "content": "S_{bin}" + }, + { + "bbox": [ + 130, + 417, + 482, + 525 + ], + "type": "text", + "content": " is employed to release the constraint on target images while maintaining the constraint on source images as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 256, + 536, + 481, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 536, + 481, + 548 + ], + "spans": [ + { + "bbox": [ + 256, + 536, + 481, + 548 + ], + "type": "interline_equation", + "content": "P _ {c} = \\operatorname {S o f t m a x} _ {r o w} \\left(S _ {b i n}\\right), \\tag {5}", + "image_path": "3aef7ada578664c23719152a3f0cb46af683c95df23f1025d22544f549473028.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 558, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 641 + ], + "type": "text", + "content": "where each sparse feature can only match one dense feature or the dustbin, while one dense feature can match multiple sparse features. The strategy of many-to-one matching with dustbin can be comprehended as independently assigning each sparse feature to either its corresponding dense feature or to the dustbin. As shown in Fig. 2(c), following conflict-free coarse matching, the positions of sparse features, which correspond to the same dense feature, are independently adjusted during the fine matching stage." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Ultimately, features matched to the dustbin are discarded, and we proceed to select matches with matching probabilities surpassing the threshold " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\theta_c" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 140, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 140, + 90, + 448, + 103 + ], + "type": "text", + "content": "RCM: Conflict-Free Local Feature Matching with Dynamic View Switching" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 235, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 235, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 235, + 129 + ], + "type": "text", + "content": "3.5 Fine Matching" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 133, + 479, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 133, + 479, + 181 + ], + "spans": [ + { + "bbox": [ + 130, + 133, + 479, + 181 + ], + "type": "text", + "content": "For each coarse match, we sample a single feature at the source fine feature map and crop a feature window of size " + }, + { + "bbox": [ + 130, + 133, + 479, + 181 + ], + "type": "inline_equation", + "content": "w \\times w" + }, + { + "bbox": [ + 130, + 133, + 479, + 181 + ], + "type": "text", + "content": " at the target fine feature map. Given the collection of " + }, + { + "bbox": [ + 130, + 133, + 479, + 181 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 133, + 479, + 181 + ], + "type": "text", + "content": " coarse matches, we obtain fine features denoted as " + }, + { + "bbox": [ + 130, + 133, + 479, + 181 + ], + "type": "inline_equation", + "content": "\\hat{F}_s^f \\in \\mathbb{R}^{K \\times 1 \\times C_2}" + }, + { + "bbox": [ + 130, + 133, + 479, + 181 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 133, + 479, + 181 + ], + "type": "inline_equation", + "content": "\\hat{F}_d^f \\in \\mathbb{R}^{K \\times w^2 \\times C_2}" + }, + { + "bbox": [ + 130, + 133, + 479, + 181 + ], + "type": "text", + "content": " for the sparse and dense branches, respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 479, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 479, + 231 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 479, + 231 + ], + "type": "text", + "content": "Subsequently, " + }, + { + "bbox": [ + 130, + 182, + 479, + 231 + ], + "type": "inline_equation", + "content": "\\hat{F}_s^f" + }, + { + "bbox": [ + 130, + 182, + 479, + 231 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 182, + 479, + 231 + ], + "type": "inline_equation", + "content": "\\hat{F}_d^f" + }, + { + "bbox": [ + 130, + 182, + 479, + 231 + ], + "type": "text", + "content": " are processed by the cross-attention module for " + }, + { + "bbox": [ + 130, + 182, + 479, + 231 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 130, + 182, + 479, + 231 + ], + "type": "text", + "content": " iterations. Inspired by [6], in order to capture the local context of dense features, we substitute the self-attention module with the integration of " + }, + { + "bbox": [ + 130, + 182, + 479, + 231 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 130, + 182, + 479, + 231 + ], + "type": "text", + "content": " convolutions into the feed-forward network. For fine matching, the features are updated by" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 246, + 237, + 480, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 237, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 246, + 237, + 480, + 251 + ], + "type": "interline_equation", + "content": "F ^ {i + 1} = F ^ {i} + \\operatorname {C o n v} _ {3} \\left(F ^ {i} | | M\\right). \\tag {6}", + "image_path": "a34c78f6c50d96166b0b7ff666f9bb7e0faf74ad69e6c2df48703ec31e79104d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 255, + 479, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 479, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 479, + 304 + ], + "type": "text", + "content": "We compute the correlation map between the centroid of the source feature window and all the features in the target feature window, which represents the matching probability, and then get the fine matching position by computing the expectation of the probability distribution, as introduced in [39]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 317, + 220, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 317, + 220, + 330 + ], + "spans": [ + { + "bbox": [ + 132, + 317, + 220, + 330 + ], + "type": "text", + "content": "3.6 Supervision" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 333, + 479, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 479, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 479, + 403 + ], + "type": "text", + "content": "Our loss function encompasses four components: the coarse matching loss " + }, + { + "bbox": [ + 130, + 333, + 479, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c^m" + }, + { + "bbox": [ + 130, + 333, + 479, + 403 + ], + "type": "text", + "content": ", the coarse dustbin loss " + }, + { + "bbox": [ + 130, + 333, + 479, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c^d" + }, + { + "bbox": [ + 130, + 333, + 479, + 403 + ], + "type": "text", + "content": ", the fine matching loss " + }, + { + "bbox": [ + 130, + 333, + 479, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_f" + }, + { + "bbox": [ + 130, + 333, + 479, + 403 + ], + "type": "text", + "content": ", and the view switching loss " + }, + { + "bbox": [ + 130, + 333, + 479, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{vs}" + }, + { + "bbox": [ + 130, + 333, + 479, + 403 + ], + "type": "text", + "content": ". Ground-truth matches " + }, + { + "bbox": [ + 130, + 333, + 479, + 403 + ], + "type": "inline_equation", + "content": "M_{gt} = \\{(i_m,j_m)\\}" + }, + { + "bbox": [ + 130, + 333, + 479, + 403 + ], + "type": "text", + "content": " are generated from camera poses and depth maps, establishing many-to-one correspondence between sparse and dense features. And non-matchable sparse features " + }, + { + "bbox": [ + 130, + 333, + 479, + 403 + ], + "type": "inline_equation", + "content": "U_{gt} = \\{i_u\\}" + }, + { + "bbox": [ + 130, + 333, + 479, + 403 + ], + "type": "text", + "content": ", which are keypoints without any ground-truth match, are also obtained." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 405, + 480, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 405, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 480, + 441 + ], + "type": "text", + "content": "The coarse matching loss " + }, + { + "bbox": [ + 130, + 405, + 480, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c^m" + }, + { + "bbox": [ + 130, + 405, + 480, + 441 + ], + "type": "text", + "content": " is computed as the negative log-likelihood loss of the ground-truth matches " + }, + { + "bbox": [ + 130, + 405, + 480, + 441 + ], + "type": "inline_equation", + "content": "M_{gt}" + }, + { + "bbox": [ + 130, + 405, + 480, + 441 + ], + "type": "text", + "content": " over the matching probability matrix " + }, + { + "bbox": [ + 130, + 405, + 480, + 441 + ], + "type": "inline_equation", + "content": "P_c" + }, + { + "bbox": [ + 130, + 405, + 480, + 441 + ], + "type": "text", + "content": " of each attention layer," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 217, + 445, + 480, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 445, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 217, + 445, + 480, + 475 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} ^ {m} = - \\frac {1}{L} \\sum_ {l} \\left[ \\frac {1}{| M _ {g t} |} \\sum_ {(i, j) \\in M _ {g t}} \\log P _ {c} ^ {l} (i, j) \\right]. \\tag {7}", + "image_path": "10153830cadb8a431e0eb2eebed457c17cf0437cba0d0e4d050bd2acb4fa33d6.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 479, + 479, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 479, + 516 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 479, + 516 + ], + "type": "text", + "content": "The coarse dustbin loss " + }, + { + "bbox": [ + 130, + 479, + 479, + 516 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c^d" + }, + { + "bbox": [ + 130, + 479, + 479, + 516 + ], + "type": "text", + "content": " refers to the negative log-likelihood loss of the non-matchable sparse features " + }, + { + "bbox": [ + 130, + 479, + 479, + 516 + ], + "type": "inline_equation", + "content": "U_{gt}" + }, + { + "bbox": [ + 130, + 479, + 479, + 516 + ], + "type": "text", + "content": " over the last column of the matching probability matrix " + }, + { + "bbox": [ + 130, + 479, + 479, + 516 + ], + "type": "inline_equation", + "content": "P_c" + }, + { + "bbox": [ + 130, + 479, + 479, + 516 + ], + "type": "text", + "content": ", which signifies the matching probabilities with the dustbin," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 224, + 520, + 480, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 520, + 480, + 548 + ], + "spans": [ + { + "bbox": [ + 224, + 520, + 480, + 548 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} ^ {d} = - \\frac {1}{| U _ {g t} |} \\sum_ {i \\in U _ {g t}} \\log P _ {c} (i, H _ {c} W _ {c} + 1). \\tag {8}", + "image_path": "258b41a17b8196542c35f9a7d17049f9f959752e2b3cb908d0a6efa0cd6e29f7.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 552, + 479, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 479, + 576 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 479, + 576 + ], + "type": "text", + "content": "The fine matching loss " + }, + { + "bbox": [ + 130, + 552, + 479, + 576 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_f" + }, + { + "bbox": [ + 130, + 552, + 479, + 576 + ], + "type": "text", + "content": " is computed by the L2-distance between each refined position and the ground-truth position, as introduced in [39]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 577, + 479, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 577, + 479, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 577, + 479, + 647 + ], + "type": "text", + "content": "To supervise the view switcher, we randomly select 500 ground-truth matches in each image pair and compute the distance between points in each image. The image with the larger average distance is identified as the larger scale image. If the original source image is the larger scale image, the ground-truth " + }, + { + "bbox": [ + 130, + 577, + 479, + 647 + ], + "type": "inline_equation", + "content": "VS_{gt}" + }, + { + "bbox": [ + 130, + 577, + 479, + 647 + ], + "type": "text", + "content": " is false, and vice versa. The binary cross-entropy loss of the switcher result " + }, + { + "bbox": [ + 130, + 577, + 479, + 647 + ], + "type": "inline_equation", + "content": "VS" + }, + { + "bbox": [ + 130, + 577, + 479, + 647 + ], + "type": "text", + "content": " is subsequently computed as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 198, + 654, + 480, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 654, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 198, + 654, + 480, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v s} = - \\left[ V S _ {g t} \\log (V S) + \\left(1 - V S _ {g t}\\right) \\log (1 - V S) \\right]. \\tag {9}", + "image_path": "9ccdb6d12913552143e5e220885ec5b69a0bbe4799702f0e7582623757bf8409.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "X. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 157, + 477, + 308 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 1: Evaluation on HPatches [1] for homography estimation. We report the percentage of correctly estimated homographies with mean corner errors below " + }, + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "inline_equation", + "content": "1/3/5" + }, + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": " pixels. We mark the best accuracy in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 157, + 477, + 308 + ], + "lines": [ + { + "bbox": [ + 135, + 157, + 477, + 308 + ], + "spans": [ + { + "bbox": [ + 135, + 157, + 477, + 308 + ], + "type": "table", + "html": "
CategoryMethodOverallIlluminationViewpoint
Accuracy (%,ε < 1/3/5 px)
SparseSuperPoint [9] + NN0.46 / 0.78 / 0.850.57 / 0.92 / 0.970.35 / 0.65 / 0.74
D2Net [10] + NN0.38 / 0.72 / 0.810.65 / 0.95 / 0.980.13 / 0.51 / 0.65
R2D2 [31] + NN0.47 / 0.78 / 0.830.63 / 0.93 / 0.980.33 / 0.64 / 0.70
ASLFeat [29] + NN0.48 / 0.81 / 0.880.63 / 0.94 / 0.980.34 / 0.69 / 0.78
SuperPoint + SuperGlue [36]0.51 / 0.83 / 0.890.62 / 0.93 / 0.980.41 / 0.73 / 0.81
SuperPoint + CAPS [43] + NN0.49 / 0.79 / 0.860.62 / 0.93 / 0.980.36 / 0.65 / 0.75
SIFT + CAPS [43] + NN0.36 / 0.76 / 0.850.48 / 0.89 / 0.950.26 / 0.65 / 0.76
SuperPoint + LightGlue [24]0.39 / 0.85 / 0.910.52 / 0.96 / 0.980.28 / 0.75 / 0.84
(Semi-)DenseSparseNCNet [32]0.36 / 0.66 / 0.760.62 / 0.92 / 0.970.13 / 0.41 / 0.57
NCNet [33]0.48 / 0.61 / 0.710.98 / 0.98 / 0.980.02 / 0.28 / 0.46
Patch2Pix [47]0.50 / 0.79 / 0.870.71 / 0.95 / 0.980.30 / 0.64 / 0.76
LoFTR [39]0.55 / 0.81 / 0.860.74 / 0.95 / 0.980.38 / 0.69 / 0.76
MatchFormer [42]0.55 / 0.81 / 0.870.75 / 0.95 / 0.980.37 / 0.68 / 0.78
Semi-SparseRCMLite0.50 / 0.86 / 0.910.55 / 0.95 / 0.980.45 / 0.79 / 0.85
RCM0.56 / 0.88 / 0.930.65 / 0.97 / 0.990.47 / 0.80 / 0.88
", + "image_path": "fa3ddd7a904c7d6a2c796c4c3f7c0ae5b1a83ce93f2768ebc3b12fdcef5cc839.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 331, + 230, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 230, + 344 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 230, + 344 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 358, + 280, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 280, + 371 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 280, + 371 + ], + "type": "text", + "content": "4.1 Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "text", + "content": "Both RCM and " + }, + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "inline_equation", + "content": "\\mathrm{RCM}_{\\mathrm{Lite}}" + }, + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "text", + "content": " consist of " + }, + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "inline_equation", + "content": "L_{1} = 5" + }, + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "text", + "content": " layers of coarse attention and " + }, + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "inline_equation", + "content": "L_{2} = 2" + }, + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "text", + "content": " layers of fine attention. The dimensions for coarse and fine features are " + }, + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "inline_equation", + "content": "C_1 = 256" + }, + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "inline_equation", + "content": "C_2 = 64" + }, + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "text", + "content": ", respectively. The key distinction between RCM and " + }, + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "inline_equation", + "content": "\\mathrm{RCM}_{\\mathrm{Lite}}" + }, + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "text", + "content": " lies in their coarse feature resolution and fine matching window size. For RCM, the coarse feature resolution is " + }, + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "inline_equation", + "content": "1/8" + }, + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "text", + "content": ", and the fine matching window size is " + }, + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "inline_equation", + "content": "w = 5" + }, + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "inline_equation", + "content": "\\mathrm{RCM}_{\\mathrm{Lite}}" + }, + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "text", + "content": ", the coarse feature resolution is " + }, + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "inline_equation", + "content": "1/16" + }, + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "text", + "content": ", and the fine matching window size is " + }, + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "inline_equation", + "content": "w = 9" + }, + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "text", + "content": ". The learnable temperature " + }, + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "text", + "content": " is initialized to 0.1, and the coarse matching threshold " + }, + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "inline_equation", + "content": "\\theta_c" + }, + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "text", + "content": " is set to 0.2. The proposed method is implemented using Pytorch [30]. During the training phase, images are resized and padded to the size of " + }, + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "inline_equation", + "content": "832 \\times 832" + }, + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "text", + "content": ". We detect 1024 keypoints in the source image, and if there are insufficient keypoints, random keypoints are added to ensure efficient batching. RCM and " + }, + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "inline_equation", + "content": "\\mathrm{RCM}_{\\mathrm{Lite}}" + }, + { + "bbox": [ + 130, + 381, + 482, + 575 + ], + "type": "text", + "content": " are trained on the MegaDepth dataset [23] for 30 epochs with a batch size of 4 by AdamW optimizer [26] and are not fine-tuned for any other tasks. The initial learning rate is set to 0.0002, and the training employs the cosine decay learning rate scheduler with 1 epoch of linear warm-up. More details can be found in the Supplementary Material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 594, + 285, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 285, + 607 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 285, + 607 + ], + "type": "text", + "content": "4.2 Homography Estimation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Dataset. The HPatches [1] dataset is widely recognized for evaluating local features and matchers. In line with the evaluation protocol in [10], we assess the performance of our method across 108 HPatches sequences, which comprise 52 sequences with illumination variations and 56 sequences with viewpoint changes." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 90, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 90, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 141, + 90, + 447, + 103 + ], + "type": "text", + "content": "RCM: Conflict-Free Local Feature Matching with Dynamic View Switching" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "Metric. Following the metrics employed in [9, 36, 47], we report the percentage of correctly estimated homographies with mean corner errors below " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "1/3/5" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " pixels. Vanilla RANSAC [14] is applied to estimate the homography for all methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 272 + ], + "type": "text", + "content": "Results. As shown in Tab. 1, RCM achieves the best overall performance at all thresholds. Comparing the dense methods, RCM shows a significant advantage in accuracy in large viewpoint change scenarios, leading LoFTR and MatchFormer by " + }, + { + "bbox": [ + 130, + 152, + 480, + 272 + ], + "type": "inline_equation", + "content": "(+9\\%, +11\\%, +12\\%)" + }, + { + "bbox": [ + 130, + 152, + 480, + 272 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 152, + 480, + 272 + ], + "type": "inline_equation", + "content": "(+10\\%, +12\\%, +10\\%)" + }, + { + "bbox": [ + 130, + 152, + 480, + 272 + ], + "type": "text", + "content": ", respectively. We attribute this performance improvement to the increased matchable points in the source image facilitated by the view switcher, and the conflict-free matching achieved in the target image through many-to-one matching. Compared to the sparse method SuperGlue, RCM achieves an overall performance lead of " + }, + { + "bbox": [ + 130, + 152, + 480, + 272 + ], + "type": "inline_equation", + "content": "(+5\\%, +5\\%, +4\\%)" + }, + { + "bbox": [ + 130, + 152, + 480, + 272 + ], + "type": "text", + "content": ", benefiting from the dense search due to the semi-sparse paradigm and the sub-pixel accuracy due to the coarse-to-fine architecture." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 327, + 302, + 444 + ], + "blocks": [ + { + "bbox": [ + 132, + 294, + 304, + 316 + ], + "lines": [ + { + "bbox": [ + 132, + 294, + 304, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 294, + 304, + 316 + ], + "type": "text", + "content": "Table 2: Two-view pose estimation results on MegaDepth [23] in outdoor scenes." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 327, + 302, + 444 + ], + "lines": [ + { + "bbox": [ + 135, + 327, + 302, + 444 + ], + "spans": [ + { + "bbox": [ + 135, + 327, + 302, + 444 + ], + "type": "table", + "html": "
CategoryMethodPose estimation
@5°@10°@20°
SparseSuperGlue [36]42.259.073.6
SGMNet [5]40.559.073.6
ClusterGNN [38]44.258.570.3
LightGlue [24]46.464.378.2
DeDoDe-B [11]49.165.077.1
DeDoDe-G [11]51.268.481.2
DenseDRC-Net [22]27.042.958.3
PDC-Net+(H) [40]43.161.976.1
LoFTR [39]52.869.281.2
RoMa [12]63.076.986.4
Semi-SparseRCMLite49.766.579.3
RCM53.269.481.5
", + "image_path": "eee4d8d385a176815a66111f9875d486270dfc991605787f40dd8b240ef97331.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 313, + 338, + 476, + 432 + ], + "blocks": [ + { + "bbox": [ + 308, + 305, + 480, + 327 + ], + "lines": [ + { + "bbox": [ + 308, + 305, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 480, + 327 + ], + "type": "text", + "content": "Table 3: Two-view pose estimation results on ScanNet [8] in indoor scenes." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 338, + 476, + 432 + ], + "lines": [ + { + "bbox": [ + 313, + 338, + 476, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 476, + 432 + ], + "type": "table", + "html": "
CategoryMethodPose estimationAUC
@5°@10°@20°
SparseOANet [44]11.826.943.9
SuperGlue [36]16.233.851.8
SGMNet [5]15.432.148.3
LightGlue [24]*13.929.645.7
DenseDRC-Net [22]*7.717.930.5
MatchFormer [42]*15.832.048.0
LoFTR-OT [39]*16.933.650.6
Semi-SparseRCMLite*15.231.548.8
RCM*17.334.652.1
", + "image_path": "bc2a4a2adbc1b915ae4e218dbb40ed25b913b1284812622c67b74c6ae8a39b0c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 487, + 288, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 288, + 498 + ], + "type": "text", + "content": "4.3 Relative Pose Estimation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 509, + 479, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 479, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 479, + 556 + ], + "type": "text", + "content": "Dataset. We evaluate matchers on the MegaDepth [23] and the ScanNet [8] dataset for outdoor and indoor pose estimation, respectively. For the MegaDepth dataset, all test images are resized such that their longer dimensions are 1216. For the ScanNet dataset, all test images are resized to " + }, + { + "bbox": [ + 130, + 509, + 479, + 556 + ], + "type": "inline_equation", + "content": "640 \\times 480" + }, + { + "bbox": [ + 130, + 509, + 479, + 556 + ], + "type": "text", + "content": " pixels." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 557, + 480, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 480, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 480, + 604 + ], + "type": "text", + "content": "Metric. Based on the matching results, we apply vanilla RANSAC to solve the fundamental matrices and relative poses of the image pairs. The AUC of the pose error at thresholds " + }, + { + "bbox": [ + 130, + 557, + 480, + 604 + ], + "type": "inline_equation", + "content": "(5^{\\circ}, 10^{\\circ}, 20^{\\circ})" + }, + { + "bbox": [ + 130, + 557, + 480, + 604 + ], + "type": "text", + "content": " are reported, where pose error is the maximum angular error in rotation and translation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "content": "Results. As shown in Tab. 2, RCM achieves the second best performance on the MegaDepth dataset. RCM outperforms the state-of-the-art method LoFTR, demonstrating the benefits of the view switcher and many-to-one matching. The lightweight " + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{RCM}_{\\mathrm{Lite}}" + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "content": " achieves a performance boost " + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "inline_equation", + "content": "(+7.5\\%, +7.5\\%, +5.7\\%)" + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "content": " compared to SuperGlue. Fig. 5 illustrates qualitative comparisons of the SuperGlue," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "X. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 479, + 316 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 479, + 316 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 479, + 316 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 479, + 316 + ], + "type": "image", + "image_path": "9148033fa40eeb7596386c3b80ff5c80d3ba8236b4df17b55762abea518745a8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 324, + 480, + 347 + ], + "lines": [ + { + "bbox": [ + 131, + 324, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 131, + 324, + 480, + 347 + ], + "type": "text", + "content": "Fig. 5: Qualitative comparison of SuperGlue [36], LoFTR [39] and proposed RCM on MegaDepth and ScanNet." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 369, + 479, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 479, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 479, + 392 + ], + "type": "text", + "content": "LoFTR and RCM. More visualizations and analyses can be found in the Supplementary Material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 393, + 480, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 393, + 480, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 393, + 480, + 477 + ], + "type": "text", + "content": "As shown in Tab. 3, RCM achieves a " + }, + { + "bbox": [ + 130, + 393, + 480, + 477 + ], + "type": "inline_equation", + "content": "(0.4\\%, 1\\%, 1.5\\%)" + }, + { + "bbox": [ + 130, + 393, + 480, + 477 + ], + "type": "text", + "content": " performance improvement over LoFTR in the ScanNet dataset. It's worth noting that the superscript * indicates zero-shot performance, and the model is trained on MegaDepth. Nevertheless, our methods achieve performance comparable to state-of-the-art sparse methods trained on ScanNet. Dense and semi-sparse matching methods find matches in the target image exhaustively, mitigating the challenge of nonrepeatability in indoor scenes, a key limitation of sparse methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 492, + 257, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 492, + 257, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 492, + 257, + 503 + ], + "type": "text", + "content": "4.4 Visual Localization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 479, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 479, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 479, + 556 + ], + "type": "text", + "content": "Dataset. We assess the performance of our method using the Aachen DayNight v1.1 benchmark [45], a well-established dataset for visual localization. This dataset comprises 824 day-time images and 191 night-time images, chosen as queries for outdoor visual localization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 479, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 479, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 479, + 593 + ], + "type": "text", + "content": "Metric. We utilize the open-source localization pipeline HLoc [35] for conducting localization with RCM. We report the percentage of successfully localized images within three error thresholds " + }, + { + "bbox": [ + 130, + 558, + 479, + 593 + ], + "type": "inline_equation", + "content": "(0.25m, 2^{\\circ}) / (0.5m, 5^{\\circ}) / (1m, 10^{\\circ})" + }, + { + "bbox": [ + 130, + 558, + 479, + 593 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "content": "Results. As shown in Tab. 4, RCM demonstrates comparable or superior performance compared to the dense methods LoFTR and ASpanFormer in outdoor visual localization tasks. Our approach exhibits strong generalization across various downstream tasks without the need for fine-tuning. It is noteworthy that the semi-sparse method RCM is significantly more efficient than the dense methods, achieving a runtime speedup of over twofold." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 141, + 90, + 447, + 102 + ], + "type": "text", + "content": "RCM: Conflict-Free Local Feature Matching with Dynamic View Switching" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 187, + 135, + 427, + 232 + ], + "blocks": [ + { + "bbox": [ + 151, + 114, + 461, + 126 + ], + "lines": [ + { + "bbox": [ + 151, + 114, + 461, + 126 + ], + "spans": [ + { + "bbox": [ + 151, + 114, + 461, + 126 + ], + "type": "text", + "content": "Table 4: Visual localization on the Aachen Day-Night benchmark v1.1 [45]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 187, + 135, + 427, + 232 + ], + "lines": [ + { + "bbox": [ + 187, + 135, + 427, + 232 + ], + "spans": [ + { + "bbox": [ + 187, + 135, + 427, + 232 + ], + "type": "table", + "html": "
MethodDayNight
(0.25m, 2°) / (0.5m, 5°) / (1m, 10°)
Localization with matching pairs generated by HLoc
LoFTR [39]88.7 / 95.6 / 99.078.5 / 90.6 / 99.0
ASpanFormer [6]89.4 / 95.6 / 99.077.5 / 91.6 / 99.0
DeDoDe-B [11]87.4 / 94.7 / 98.570.7 / 88.0 / 97.9
DeDoDe-G [11]88.5 / 95.4 / 99.074.9 / 89.5 / 99.5
SP [9]+LightGlue [24]89.6 / 95.8 / 99.272.8 / 88.0 / 99.0
RCMLite87.7 / 94.7 / 98.472.3 / 91.1 / 98.4
RCM89.7 / 96.0 / 98.772.8 / 91.6 / 99.0
", + "image_path": "f142fa070020cb597e3d4a378fae261f407e54fc30296cc0ec6be5d6c031e726.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 136, + 241, + 323, + 399 + ], + "blocks": [ + { + "bbox": [ + 136, + 241, + 323, + 399 + ], + "lines": [ + { + "bbox": [ + 136, + 241, + 323, + 399 + ], + "spans": [ + { + "bbox": [ + 136, + 241, + 323, + 399 + ], + "type": "image", + "image_path": "f6616bcbb4d0320346338474f8bfffc51e74ef86b5a8a3e62d1f2d96c8f012fc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 406, + 482, + 452 + ], + "lines": [ + { + "bbox": [ + 130, + 406, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 406, + 482, + 452 + ], + "type": "text", + "content": "Fig. 6: Ablation study on view switcher and many-to-one matching. (a) (b) shows the actual matching results and (c) (d) shows the ground-truth coarse matches. The harmonious combination of view switcher and many-to-one matching allows more matchable points in the source image and conflict-free matching in the target image." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 324, + 240, + 480, + 398 + ], + "blocks": [ + { + "bbox": [ + 324, + 240, + 480, + 398 + ], + "lines": [ + { + "bbox": [ + 324, + 240, + 480, + 398 + ], + "spans": [ + { + "bbox": [ + 324, + 240, + 480, + 398 + ], + "type": "image", + "image_path": "f1afd91a353ab4d556a8cece7cd08f2ad21cb1064bdf954fc6ae8e64438aad18.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 475, + 238, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 238, + 487 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 238, + 487 + ], + "type": "text", + "content": "4.5 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 497, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 556 + ], + "type": "text", + "content": "In this subsection, we conduct qualitative and quantitative ablation experiments on MegaDepth. All models in Tab. 5 are trained for 15 epochs at 544 resolution and tested at 1216 resolution. For the baseline model (1), we apply the original SuperPoint [9] to generate keypoints and descriptors for both images. O2O and M2O in Tab. 5 are abbreviations for one-to-one and many-to-one, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Ground-Truth Match Analysis. Fig. 6(d) illustrates the core idea of RCM, which is to increase the theoretical upper bound on the matcher, i.e., ground-truth matches, through view switcher and many-to-one matching. Fig. 6(c) illustrates a quantitative comparison of ground-truth matches, showing an increase as the view switching accuracy improves during training. Eventually, RCM and " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{RCM}_{\\mathrm{Lite}}" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " achieve view switching accuracies of " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "90.9\\%" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "87.8\\%" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": ", respectively. Through the integration of many-to-one matching and the view switcher, the ground-truth matches surge to " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "260\\%" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " compared to one-to-one matching, which is a compelling demonstration of raising the ceiling of feature matching." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "X. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 149, + 302, + 243 + ], + "blocks": [ + { + "bbox": [ + 132, + 118, + 305, + 140 + ], + "lines": [ + { + "bbox": [ + 132, + 118, + 305, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 118, + 305, + 140 + ], + "type": "text", + "content": "Table 5: Ablation study of each proposal on MegaDepth." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 149, + 302, + 243 + ], + "lines": [ + { + "bbox": [ + 136, + 149, + 302, + 243 + ], + "spans": [ + { + "bbox": [ + 136, + 149, + 302, + 243 + ], + "type": "table", + "html": "
Method#MatchesPrec.
(1) Sparse sampling451.798.0
(2) Dense sampling932.193.5
(3) Semi-Sparse sampling492.393.7
(4) + U-Net Extraction507.894.2
(5) O2O → M2O770.7(+51.8%)97.2
(6) + Switcher (RCMLite)951.4(+23.4%)97.8
(7) 1/16 → 1/8 (RCM)961.498.1
", + "image_path": "a2265a2cc9d474b98d9135c6d2e6279bf231566af98ce5b9ed104cdeeb513131.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 328, + 146, + 462, + 246 + ], + "blocks": [ + { + "bbox": [ + 308, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 308, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 308, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 6: Evaluation on extreme-scale MegaDepth." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 146, + 462, + 246 + ], + "lines": [ + { + "bbox": [ + 328, + 146, + 462, + 246 + ], + "spans": [ + { + "bbox": [ + 328, + 146, + 462, + 246 + ], + "type": "table", + "html": "
Method#MatchesPrec.
SP [9] + SuperGlue [36]26175.5
SP [9] + LightGlue [24]31376.0
LoFTR [39]42453.5
MatchFormer [42]46477.0
ASpanFormer [6]42577.1
RCM88778.9
wo switcher65375.4
flipped switcher7654.2
", + "image_path": "f7b405c952bfe8438fc2a9cfd97dc4076d9f480b9f0bb12f864ff47000185062.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 276, + 479, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 276, + 479, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 276, + 479, + 326 + ], + "type": "text", + "content": "Sampling Strategy. As shown in Tab. 5, the baseline models (1), (2), and (3) differ solely in sampling strategies for coarse features. Precisely, (1) detects keypoints in both images, (2) employs dense feature as the coarse feature, and (3) detects keypoints in source image and employs dense feature in target image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 328, + 480, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 480, + 389 + ], + "type": "text", + "content": "Many-to-One Matching. As shown in Fig. 6(a), in contrast to the one-to-one matching strategy, the many-to-one matching strategy pursues conflict-free matching, resulting in more detailed matches. As shown in Tab. 5(5), substituting one-to-one matching with many-to-one matching improves the number of matches by " + }, + { + "bbox": [ + 130, + 328, + 480, + 389 + ], + "type": "inline_equation", + "content": "51.8\\%" + }, + { + "bbox": [ + 130, + 328, + 480, + 389 + ], + "type": "text", + "content": " and precision by " + }, + { + "bbox": [ + 130, + 328, + 480, + 389 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 130, + 328, + 480, + 389 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 392, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 477 + ], + "type": "text", + "content": "View Switcher. As shown in Fig. 6(b), without the view switcher, the matcher can only produce a small number of matches in small scale images because only a limited number of keypoints are detected in the overlapping region. The view switcher enables the detection of more matchable keypoints in the source image. As shown in Tab. 5(6), the average number of matches increases by " + }, + { + "bbox": [ + 130, + 392, + 482, + 477 + ], + "type": "inline_equation", + "content": "23.4\\%" + }, + { + "bbox": [ + 130, + 392, + 482, + 477 + ], + "type": "text", + "content": " with the view switcher. More analyses about view switcher and many-to-one matching can be found in the Supplementary Material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 479, + 482, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 577 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 577 + ], + "type": "text", + "content": "RCM and " + }, + { + "bbox": [ + 130, + 479, + 482, + 577 + ], + "type": "inline_equation", + "content": "\\mathrm{RCM}_{\\mathrm{Lite}}" + }, + { + "bbox": [ + 130, + 479, + 482, + 577 + ], + "type": "text", + "content": ". The distinctions between RCM and " + }, + { + "bbox": [ + 130, + 479, + 482, + 577 + ], + "type": "inline_equation", + "content": "\\mathrm{RCM}_{\\mathrm{Lite}}" + }, + { + "bbox": [ + 130, + 479, + 482, + 577 + ], + "type": "text", + "content": " lie in the coarse feature resolution and fine matching window. RCM employs coarse features at 1/8 resolution, consistent with the dense attention-based methods. We apply 1/16 resolution coarse features in " + }, + { + "bbox": [ + 130, + 479, + 482, + 577 + ], + "type": "inline_equation", + "content": "\\mathrm{RCM}_{\\mathrm{Lite}}" + }, + { + "bbox": [ + 130, + 479, + 482, + 577 + ], + "type": "text", + "content": " to trade performance degradation for faster runtime, facilitating applications in real-time scenarios. Because of the additional layers at 1/16 resolution in the feature extraction network, " + }, + { + "bbox": [ + 130, + 479, + 482, + 577 + ], + "type": "inline_equation", + "content": "\\mathrm{RCM}_{\\mathrm{Lite}}" + }, + { + "bbox": [ + 130, + 479, + 482, + 577 + ], + "type": "text", + "content": " consists of more parameters than RCM. Also to cover larger coarse feature patches, we design a larger fine matching window for " + }, + { + "bbox": [ + 130, + 479, + 482, + 577 + ], + "type": "inline_equation", + "content": "\\mathrm{RCM}_{\\mathrm{Lite}}" + }, + { + "bbox": [ + 130, + 479, + 482, + 577 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 579, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 579, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 579, + 482, + 665 + ], + "type": "text", + "content": "Robustness to Scale Variation. We evaluate matchers on the MegaDepth subset defined in OETR [7], where the overlap scale ratio is in " + }, + { + "bbox": [ + 130, + 579, + 482, + 665 + ], + "type": "inline_equation", + "content": "[4, +\\infty]" + }, + { + "bbox": [ + 130, + 579, + 482, + 665 + ], + "type": "text", + "content": ". As shown in Tab. 6, RCM achieves the best performance with " + }, + { + "bbox": [ + 130, + 579, + 482, + 665 + ], + "type": "inline_equation", + "content": "2.8\\times" + }, + { + "bbox": [ + 130, + 579, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 579, + 482, + 665 + ], + "type": "inline_equation", + "content": "2.1\\times" + }, + { + "bbox": [ + 130, + 579, + 482, + 665 + ], + "type": "text", + "content": " more matches than LightGlue and ASpanFormer respectively. The performance of RCM is severely degraded without the view switcher, demonstrating the significance of the view switcher in large scale variation scenarios. Flipping the switching result, i.e., actively taking small scale image as source image, leads to the worst results." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 90, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 90, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 141, + 90, + 447, + 103 + ], + "type": "text", + "content": "RCM: Conflict-Free Local Feature Matching with Dynamic View Switching" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 264, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 264, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 264, + 235 + ], + "type": "text", + "content": "Efficiency. The parameters, runtimes, and performances of state-of-the-art matchers are presented in Tab. 7. Notably, RCM and " + }, + { + "bbox": [ + 132, + 116, + 264, + 235 + ], + "type": "inline_equation", + "content": "\\mathrm{RCM}_{\\mathrm{Lite}}" + }, + { + "bbox": [ + 132, + 116, + 264, + 235 + ], + "type": "text", + "content": " achieve the optimal balance between performance and efficiency. Although the performance of RCM lags behind the most precise matcher" + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 272, + 152, + 481, + 222 + ], + "blocks": [ + { + "bbox": [ + 271, + 129, + 482, + 150 + ], + "lines": [ + { + "bbox": [ + 271, + 129, + 482, + 150 + ], + "spans": [ + { + "bbox": [ + 271, + 129, + 482, + 150 + ], + "type": "text", + "content": "Table 7: Efficiency Analysis on RTX3090. The 1st, 2nd, and 3rd-best methods are highlighted." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 272, + 152, + 481, + 222 + ], + "lines": [ + { + "bbox": [ + 272, + 152, + 481, + 222 + ], + "spans": [ + { + "bbox": [ + 272, + 152, + 481, + 222 + ], + "type": "table", + "html": "
CategoryMethodAUC@10°#Param.Time (ms)Avg. Rank
SparseDeDoDe [11]65.028.1M3224.7
LightGlue [24]64.311.9M723.7
DenseLoFTR [39]69.211.6M4543.7
RoMa [12]76.9111M9134.3
Semi-SparseRCMLite66.511.4M1342.7
RCM69.49.8M2082
", + "image_path": "d8daae1f4bb59da5206070271f90084477d9aa665252581b2005fc6a632a773a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 236, + 480, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 260 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 260 + ], + "type": "text", + "content": "RoMa, RCM achieves respectable performance with only " + }, + { + "bbox": [ + 132, + 236, + 480, + 260 + ], + "type": "inline_equation", + "content": "8.8\\%" + }, + { + "bbox": [ + 132, + 236, + 480, + 260 + ], + "type": "text", + "content": " of the parameters and " + }, + { + "bbox": [ + 132, + 236, + 480, + 260 + ], + "type": "inline_equation", + "content": "23\\%" + }, + { + "bbox": [ + 132, + 236, + 480, + 260 + ], + "type": "text", + "content": " of the runtime, making it more suitable for real-time applications." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 277, + 308, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 277, + 308, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 308, + 289 + ], + "type": "text", + "content": "5 Limitation and Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 300, + 480, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 480, + 361 + ], + "type": "text", + "content": "The view switcher assumes that one image has a relatively large scale, while both images being small in scale remains a tricky case in feature matching. A potential approach involves dynamically scaling the image pairs according to the overlapping regions, ensuring that the overlapping regions in two images remain simultaneously large and comparable in scale." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 361, + 480, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 480, + 396 + ], + "type": "text", + "content": "Since the matching points in the target image are regression-based points rather than pre-detected points, RCM suffers from the lack of discrete tracks in the SfM task and requires additional adaptation strategy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 396, + 480, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 480, + 468 + ], + "type": "text", + "content": "The semi-sparse paradigm mitigates the keypoint-repeatability reliance, addressing only a portion of the broader keypoint dependence. Challenges in detection persist, such as low quantity and poor distribution, influencing the performance of RCM in texture-less scenes. We believe that the integration of the semi-sparse paradigm with advanced detection methods holds promising potential as a path forward for effective and efficient feature matching." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 486, + 220, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 220, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 220, + 498 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 510, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 481, + 666 + ], + "type": "text", + "content": "This paper introduces a novel perspective to enhance matchers by raising the theoretical upper bound. We propose the dynamic view switching mechanism for the source image and the conflict-free coarse matching layer for the target image, respectively. Through the integration of the view switcher and many-to-one matching layer, we enhance the number of matchable points in the source image while achieving conflict-free matching in the target image. By essentially augmenting the number of ground-truth matches between images, RCM raises the theoretical upper bound on the actual matching results. The results from qualitative and quantitative experiments demonstrate a significant enhancement in the practical performance of the matcher with a higher ceiling, particularly evident in challenging scenarios characterized by large scale variations. With its outstanding performance and efficiency, we believe that RCM holds considerable potential for research and application." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "X. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 163 + ], + "type": "text", + "content": "This work was supported by the SEU Innovation Capability Enhancement Plan for Doctoral Students under grant CXJH SEU 24128." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 205, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 205, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 481, + 239 + ], + "type": "text", + "content": "1. Balntas, V., Lenc, K., Vedaldi, A., Mikolajczyk, K.: Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In: Proceedings of the CVPR. pp. 5173-5182 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 239, + 481, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 481, + 260 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 481, + 260 + ], + "type": "text", + "content": "2. Barroso-Laguna, A., Tian, Y., Mikolajczyk, K.: Scalenet: A shallow architecture for scale estimation. In: Proceedings of the CVPR. pp. 12808-12818 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 261, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 261, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 138, + 261, + 481, + 282 + ], + "type": "text", + "content": "3. Bay, H., Tuytelaars, T., Van G., L.: SURF: Speeded Up Robust Features. In: Proceedings of the ECCV. pp. 404-417 (2006)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 283, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 283, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 283, + 481, + 304 + ], + "type": "text", + "content": "4. Calonder, M., Lepetit, V., Strecha, C., Fua, P.: BRIEF: Binary Robust Independent Elementary Features. In: Proceedings of the ECCV. pp. 778-792 (2010)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 304, + 481, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 481, + 337 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 481, + 337 + ], + "type": "text", + "content": "5. Chen, H., Luo, Z., Zhang, J., Zhou, L., Bai, X., Hu, Z., Tai, C., Quan, L.: Learning to match features with seeded graph matching network. In: Proceedings of the ICCV. pp. 6301-6310 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 338, + 481, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 338, + 481, + 370 + ], + "spans": [ + { + "bbox": [ + 138, + 338, + 481, + 370 + ], + "type": "text", + "content": "6. Chen, H., Luo, Z., Zhou, L., Tian, Y., Zhen, M., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Aspanformer: Detector-free image matching with adaptive span transformer. In: Proceedings of the ECCV. pp. 20-36 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 371, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 371, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 371, + 481, + 392 + ], + "type": "text", + "content": "7. Chen, Y., Huang, D., Xu, S., Liu, J., Liu, Y.: Guide local feature matching by overlap estimation. In: Proceedings of the AAAI. pp. 365-373 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 392, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 481, + 424 + ], + "type": "text", + "content": "8. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the CVPR. pp. 5828-5839 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 425, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 481, + 456 + ], + "type": "text", + "content": "9. DeTone, D., Malisiewicz, T., Rabinovich, A.: SuperPoint: Self-Supervised Interest Point Detection and Description. In: Proceedings of the CVPRW. pp. 224-236 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 457, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 457, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 138, + 457, + 481, + 490 + ], + "type": "text", + "content": "0. Dusmanu, M., Rocco, I., Pajdla, T., Pollefeys, M., Sivic, J., Torii, A., Sattler, T.: D2-Net: A Trainable CNN for Joint Description and Detection of Local Features. In: Proceedings of the CVPR. pp. 8092–8101 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 491, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 481, + 522 + ], + "type": "text", + "content": "1. Edstedt, J., Bokman, G., Wadenback, M., Felsberg, M.: Dedode: Detect, don't describe—describe, don't detect for local feature matching. In: Proceedings of the 3DV. pp. 148-157 (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 523, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 481, + 544 + ], + "type": "text", + "content": "2. Edstedt, J., Sun, Q., Bokman, G., Wadenback, M., Felsberg, M.: Roma: Robust dense feature matching. In: Proceedings of the CVPR. pp. 19790-19800 (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 545, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 481, + 567 + ], + "type": "text", + "content": "3. Engel, J., Koltun, V., Cremers, D.: Direct sparse odometry. IEEE Transactions on Pattern Analysis and Machine Intelligence 40(3), 611-625 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 481, + 599 + ], + "type": "text", + "content": "4. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM pp. 381-395 (1981)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 600, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 481, + 620 + ], + "type": "text", + "content": "5. Fu, Y., Wu, Y.: Scale-net: Learning to reduce scale differences for large-scale invariant image matching. arXiv preprint arXiv:2112.10485 (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 621, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 481, + 643 + ], + "type": "text", + "content": "6. Germain, H., Bourmaud, G., Lepetit, V.: Sparse-to-dense hypercolumn matching for long-term visual localization. In: Proceedings of the 3DV. pp. 513-523 (2019)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "text", + "content": "7. Germain, H., Bourmaud, G., Lepetit, V.: S2dnet: Learning image features for accurate sparse-to-dense matching. In: Proceedings of the ECCV. pp. 626-643 (2020)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 141, + 91, + 447, + 102 + ], + "type": "text", + "content": "RCM: Conflict-Free Local Feature Matching with Dynamic View Switching" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "18. Gleize, P., Wang, W., Feiszli, M.: Silk: Simple learned keypoints. In: Proceedings of the CVPR. pp. 22499-22508 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "type": "text", + "content": "19. Huang, D., Chen, Y., Liu, Y., Liu, J., Xu, S., Wu, W., Ding, Y., Tang, F., Wang, C.: Adaptive assignment for geometry aware local feature matching. In: Proceedings of the CVPR. pp. 5425-5434 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 173, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 173, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 173, + 480, + 205 + ], + "type": "text", + "content": "20. Jiang, W., Trulls, E., Hosang, J., Tagliasacchi, A., Yi, K.M.: Cotr: Correspondence transformer for matching across images. In: Proceedings of the ICCV. pp. 6207-6217 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 206, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 480, + 239 + ], + "type": "text", + "content": "21. Katharopoulos, A., Vyas, A., Pappas, N., Fleuret, F.: Transformers are rnns: Fast autoregressive transformers with linear attention. In: Proceedings of the ICML. pp. 5156-5165 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 239, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 239, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 239, + 480, + 262 + ], + "type": "text", + "content": "22. Li, X., Han, K., Li, S., Prisacariu, V.: Dual-resolution correspondence networks. In: Proceedings of the NeurIPS. pp. 17346-17357 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 262, + 480, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 480, + 285 + ], + "type": "text", + "content": "23. Li, Z., Snavely, N.: MegaDepth: Learning Single-View Depth Prediction from Internet Photos. In: Proceedings of the CVPR. pp. 2041–2050 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 285, + 480, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 480, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 480, + 307 + ], + "type": "text", + "content": "24. Lindenberger, P., Sarlin, P., Pollefeys, M.: Lightglue: Local feature matching at light speed. In: Proceedings of the ICCV. pp. 17627-17638 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 308, + 480, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 480, + 330 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 480, + 330 + ], + "type": "text", + "content": "25. Liu, C., Yuen, J., Torralba, A., Sivic, J., Freeman, W.T.: Sift flow: Dense correspondence across different scenes. In: Proceedings of the ECCV. pp. 28-42 (2008)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 331, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 480, + 352 + ], + "type": "text", + "content": "26. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 353, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 480, + 374 + ], + "type": "text", + "content": "27. Lowe, D.G.: Distinctive Image Features from Scale-Invariant Keypoints. International Journal of Computer Vision 60(2), 91-110 (2004)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 375, + 480, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 480, + 397 + ], + "type": "text", + "content": "28. Lu, X., Yan, Y., Kang, B., Du, S.: Paraformer: Parallel attention transformer for efficient feature matching. arXiv preprint arXiv:2303.00941 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 398, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 398, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 398, + 480, + 430 + ], + "type": "text", + "content": "29. Luo, Z., Zhou, L., Bai, X., Chen, H., Zhang, J., Yao, Y., Li, S., Fang, T., Quan, L.: Aslfeat: Learning local features of accurate shape and localization. In: Proceedings of the CVPR. pp. 6589-6598 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 431, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 480, + 464 + ], + "type": "text", + "content": "30. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. Proceedings of the NeurIPS 32 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 465, + 480, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 465, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 465, + 480, + 497 + ], + "type": "text", + "content": "31. Revaud, J., De Souza, C., Humenberger, M., Weinzaepfel, P.: R2D2: Reliable and Repeatable Detector and Descriptor. In: Proceedings of the NeurIPS. pp. 12414-12424 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 498, + 480, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 480, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 480, + 530 + ], + "type": "text", + "content": "32. Rocco, I., Arandjelovic, R., Sivic, J.: Efficient neighbourhood consensus networks via submanifold sparse convolutions. In: Proceedings of the ECCV. pp. 605-621 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 531, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 480, + 553 + ], + "type": "text", + "content": "33. Rocco, I., Cimpoi, M., Arandjelović, R., Torii, A., Pajdla, T., Sivic, J.: Neighbourhood consensus networks. In: Proceedings of the NeurIPS. pp. 1658-1669 (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 554, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 480, + 576 + ], + "type": "text", + "content": "34. Rublee, E., Rabaud, V., Konolige, K., Bradski, G.: ORB: An efficient alternative to SIFT or SURF. In: Proceedings of the ICCV. pp. 2564-2571 (2011)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 577, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 480, + 609 + ], + "type": "text", + "content": "35. Sarlin, P., Cadena, C. and Siegwart, R., Dymczyk, M.: From coarse to fine: Robust hierarchical localization at large scale. In: Proceedings of the CVPR. pp. 12716-12725 (2019)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 610, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 480, + 643 + ], + "type": "text", + "content": "36. Sarlin, P., DeTone, D., Malisiewicz, T., Rabinovich, A.: SuperGlue: Learning Feature Matching With Graph Neural Networks. In: Proceedings of the CVPR. pp. 4937-4946 (2020)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "text", + "content": "37. Schonberger, J.L., Frahm, J.: Structure-from-motion revisited. In: Proceedings of the CVPR. pp. 4104-4113 (2016)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "X. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 424 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "38. Shi, Y., Cai, J., Shavit, Y., Mu, T., Feng, W., Zhang, K.: Clustergnn: Cluster-based coarse-to-fine graph neural network for efficient feature matching. In: Proceedings of the CVPR. pp. 12517-12526 (June 2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "type": "text", + "content": "39. Sun, J., Shen, Z., Wang, Y., Bao, H., Zhou, X.: LoFTR: Detector-Free Local Feature Matching With Transformers. In: Proceedings of the CVPR. pp. 8922-8931 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 482, + 215 + ], + "type": "text", + "content": "40. Truong, P., Danelljan, M., Timofte, R., Van Gool, L.: Pdc-net+: Enhanced probabilistic dense correspondence network. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 482, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 482, + 247 + ], + "type": "text", + "content": "41. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is All you Need. In: Proceedings of the NeurIPS. pp. 5998-6008 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 482, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 482, + 281 + ], + "type": "text", + "content": "42. Wang, Q., Zhang, J., Yang, K., Peng, K., Stiefelhagen, R.: Matchformer: Interleaving attention in transformers for feature matching. In: Proceedings of the ACCV. pp. 2746-2762 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 482, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 482, + 304 + ], + "type": "text", + "content": "43. Wang, Q., Zhou, X., Hariharan, B., Snavely, N.: Learning feature descriptors using camera pose supervision. In: Proceedings of the ECCV. pp. 757-774 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 482, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 482, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 482, + 336 + ], + "type": "text", + "content": "44. Zhang, J., Sun, D., Luo, Z., Yao, A., Zhou, L., Shen, T., Chen, Y., Quan, L., Liao, H.: Learning two-view correspondences and geometry using order-aware network. In: Proceedings of the ICCV. pp. 5845-5854 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 337, + 482, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 337, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 337, + 482, + 369 + ], + "type": "text", + "content": "45. Zhang, Z., Sattler, T., Scaramuzza, D.: Reference pose generation for long-term visual localization via learned features and view synthesis. International Journal of Computer Vision pp. 821-844 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 370, + 482, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 482, + 402 + ], + "type": "text", + "content": "46. Zhao, X., Wu, X., Chen, W., Chen, P.C., Xu, Q., Li, Z.: A liked: A lighter keypoint and descriptor extraction network via deformable transformation. IEEE Transactions on Instrumentation and Measurement 72, 1-16 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 482, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 482, + 424 + ], + "type": "text", + "content": "47. Zhou, Q., Sattler, T., Leal-Taixe, L.: Patch2pix: Epipolar-guided pixel-level correspondences. In: Proceedings of the CVPR. pp. 4669-4678 (2021)" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 141, + 91, + 447, + 102 + ], + "type": "text", + "content": "RCM: Conflict-Free Local Feature Matching with Dynamic View Switching" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Random Walk on Pixel Manifolds for Anomaly Segmentation of Complex Driving Scenes/40e3d166-d87c-43d0-aeb7-8851ed372fb1_content_list.json b/2024/Random Walk on Pixel Manifolds for Anomaly Segmentation of Complex Driving Scenes/40e3d166-d87c-43d0-aeb7-8851ed372fb1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..df9ab7025fe8b772d5a8d9e9ddfbb006cb8f8d1b --- /dev/null +++ b/2024/Random Walk on Pixel Manifolds for Anomaly Segmentation of Complex Driving Scenes/40e3d166-d87c-43d0-aeb7-8851ed372fb1_content_list.json @@ -0,0 +1,1519 @@ +[ + { + "type": "text", + "text": "Random Walk on Pixel Manifolds for Anomaly Segmentation of Complex Driving Scenes", + "text_level": 1, + "bbox": [ + 233, + 140, + 772, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zelong Zeng1 and Kaname Tomite1", + "bbox": [ + 364, + 210, + 633, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "SenseTime Japan, Tokyo, Japan {zengzelong,tomite}@sensetime.jp", + "bbox": [ + 375, + 239, + 625, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In anomaly segmentation for complex driving scenes, state-of-the-art approaches utilize anomaly scoring functions to calculate anomaly scores. For these functions, accurately predicting the logits of inlier classes for each pixel is crucial for precisely inferring the anomaly score. However, in real-world driving scenarios, the diversity of scenes often results in distorted manifolds of pixel embeddings in the space. This effect is not conducive to directly using the pixel embeddings for the logit prediction during inference, a concern overlooked by existing methods. To address this problem, we propose a novel method called Random Walk on Pixel Manifolds (RWPM). RWPM utilizes random walks to reveal the intrinsic relationships among pixels to refine the pixel embeddings. The refined pixel embeddings alleviate the distortion of manifolds, improving the accuracy of anomaly scores. Our extensive experiments show that RWPM consistently improve the performance of the existing anomaly segmentation methods and achieve the best results1.", + "bbox": [ + 259, + 301, + 756, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Autonomous vehicles $\\cdot$ Anomaly segmentation $\\cdot$ Semantic segmentation", + "bbox": [ + 259, + 522, + 743, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 575, + 375, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semantic segmentation [9-11,13,39] plays a crucial role in the field of automated driving, which can aid the automated driving system in perceiving and understanding the surrounding environment, thereby facilitating the system make accurate decision. Typically, the semantic segmentation models are trained on closed-datasets with fixed semantic categories, showing good recognition capabilities for inlier classes (e.g., roads, vehicles, etc.). However, these models often fail to effectively detect potential outlier class objects (e.g., animals, fallen tires, etc.) in real-world scenarios. Due to the critical importance of correctly detecting outlier class targets for the safety of autonomous driving, there is a growing focus on addressing this task, namely, anomaly segmentation.", + "bbox": [ + 212, + 604, + 787, + 755 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The standard paradigm for anomaly segmentation is assigning an anomaly score to each pixel of an input image, where a higher score indicates a greater probability that the pixel belongs to an outlier. Thus, the way used for estimating the anomaly score is crucial for anomaly segmentation methods. In", + "bbox": [ + 212, + 757, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1 Code is available at: https://github.com/ZelongZeng/RwPM", + "bbox": [ + 217, + 824, + 640, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/66465a8b974307d9e4d39db93e496fd578b716b3a4fdf1ea9f19da4f91b66e54.jpg", + "image_caption": [ + "(i) An ideal manifold structure", + "Fig. 1: A toy example about manifolds of pixel embeddings. The dashed circles indicate regions with high predicted logit to the corresponding inlier class. For the pixel in these regions, its anomaly score is small. The prototype is the vector of network weights of the classifier for the corresponding inlier class. (i) An ideal manifold structure. All inlier pixels are within the corresponding high logit region, whereas all outlier pixels are outside the areas. The anomaly score derived from the anomaly scoring function effectively discriminates between them. (ii) The manifolds are affected by the diversity of data, causing some inlier pixels to deviate from the corresponding high logit region, while some outlier pixels approach the regions. This results in false positives/negatives in the anomaly scores. However, pixels of the same class are still in the same manifold, indicating that the manifold structure can be utilized to reveal the intrinsic relationships between pixels. (iii) Our RWPM utilizes random walks to capture the manifold structure to diffuse and update pixel embeddings. The embeddings within the same manifold tend to become more similar after the updating. Fig 3 and Fig 4 respectively present the visualization of embedding distributions and the qualitative results of real examples, consistent with the description of this toy example." + ], + "image_footnote": [ + "Inlier pixel" + ], + "bbox": [ + 246, + 142, + 390, + 233 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e5dd3250608f82a6b1a86386c513f141d9085137879f723e452084cd3cb51639.jpg", + "image_caption": [ + "(ii) A distorted manifold structure" + ], + "image_footnote": [ + "$\\times \\times$ Inlier class prototype" + ], + "bbox": [ + 413, + 142, + 586, + 232 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7eb03a33c34f0e318dc50b70b00fe594d25cdc69d3ad192468b00ec5e72d85c5.jpg", + "image_caption": [ + "(iii) A refined manifold structure" + ], + "image_footnote": [ + "Outlier pixel" + ], + "bbox": [ + 607, + 142, + 754, + 233 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "early approaches, the anomaly score is often given by estimating the prediction uncertainty [16, 21, 27, 29], or by comparing the discrepancy between the original image and its reconstructed image generated based on the semantic segmentation predictions [14, 15, 31, 35]. Different from early methods, recent methods [12, 19, 32, 34, 36, 37] adopt Outlier Exposure (OE) strategy [22], which directly learn a pixel-level anomaly class to detect anomalous pixels that deviate from inlier classes. Specifically, these methods propose anomaly scoring functions to infer the anomaly scores and, based on the functions, leverage auxiliary outlier data to fine-tune a semantic segmentation model. The auxiliary outlier data is typically generated by extracting outlier-class objects from other datasets (e.g., COCO [30]) and incorporating them into driving scene images [34, 36, 37]. Furthermore, some of them [19, 34, 36] adopt mask-based segmentation networks instead of pixel-based segmentation networks, achieving state-of-the-art (SOTA) performance in anomaly segmentation. In this paper, we concentrate on the current methods using OE strategy. These methods indicate superior performance to the early methods using prediction uncertainty or image reconstruction.", + "bbox": [ + 212, + 537, + 787, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The anomaly scoring functions in OE methods infer anomaly scores based on the inlier class logit predictions from a semantic segmentation model. Specifically, the logic of these functions can be represented as follows: when the predicted logits of a pixel embedding are small across all inlier classes, indicating that the", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z. Zeng et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "pixel is dissimilar to all inlier classes and is an outlier with high probability, the function yields a higher anomaly score. Therefore, accurately measuring the logits of each pixel is crucial for precisely evaluating the anomaly score. However, in practical open-world driving scenarios, various environmental conditions (such as lighting, road materials, etc.) and a diverse range of unknown class objects are present. During the inference phase, the diversity of driving scenes often results in distorted manifolds of pixel embedding in the embedding space, causing pixel embeddings to deviate from the ideal distribution. This phenomenon has been validated in previous works of metric learning [3,25,40,41]. Since the logits are calculated from the inner product (i.e., cosine distance) of the pixel embeddings and the vectors of classifiers' network weights, the distorted manifolds are not conducive to directly using the pixel embeddings for the logit prediction. However, current anomaly segmentation methods ignore this issue and directly use the pixel embeddings for the logit prediction in the inference phase, so the results often fail to accurately reflect the anomaly score for the corresponding pixels, causing high false positive/negative rate (see Fig. 1(i) and Fig. 1(ii)).", + "bbox": [ + 212, + 145, + 787, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we propose a novel method, called Random Walk on Pixel Manifolds (RWPM), to alleviate the effect of data diversity on the manifolds. RWPM utilizes random walks to capture the manifold structure of the pixel embeddings to diffuse and update the embedding of each pixel. Since the structure of data manifolds implicitly contains the intrinsic relationship among data points (see Fig. 1(ii)) [3,24,40,43], using random walks to measure the similarity between pixels on manifolds (i.e., distance on manifolds rather than Euclidean distance) better reflects the intrinsic relationships between pixels. Consequently, diffusing and updating pixel embeddings based on the similarity on manifolds results in high similarity among pixels on the same manifold and low similarity among pixels on different manifolds. In the embedding space, this manifests as pixel embeddings on the same manifold forming more compact clusters, thereby mitigating the distortion effect of data diversity on the manifolds, which facilitates the logit prediction (see Fig. 1(iii)). Furthermore, since the typically large number of pixels in an image, utilizing all pixels for the random walks results in significant memory and computational overhead, which damages the deployment and real-time execution of our RWPM. To address this issue, we introduce the Partial Random Walk strategy for RWPM, reducing memory consumption and enhancing the operational efficiency of RWPM. Note that, RWPM is proposed to use in the inference phase, thus no any additional training is required. Additionally, RWPM can be directly integrated into existing anomaly segmentation framework without the need for any extra modification to the network structure. Extensive experiments in various anomaly segmentation benchmarks of road scenes, namely Fishyscapes [5], Road Anomaly [31], and SMIYC [6], demonstrate that RWPM consistently improves the performance of existing methods, achieving SOTA results. To summarise, our contributions are the following:", + "bbox": [ + 212, + 391, + 787, + 785 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "RWPM", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "- We introduce that the diversity of driving scenes results in distorted manifolds of pixel embedding, thereby affecting the accuracy of anomaly scoring", + "bbox": [ + 225, + 809, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "functions in inferring anomaly scores. This problem is ignored by existing anomaly segmentation methods.", + "bbox": [ + 240, + 146, + 784, + 174 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- We propose the Random Walk on Pixel Manifolds (RWPM) that utilizes random walks to reveal the intrinsic relationships among pixels to refine the pixel embeddings. This way alleviate the effect on manifolds caused by the driving scene diversity.", + "bbox": [ + 225, + 176, + 782, + 233 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- We propose the Partial Random Walk strategy for RWPM to reduce the memory consumption and improve the operational efficiency.", + "bbox": [ + 225, + 234, + 782, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 282, + 387, + 300 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Anomaly Segmentation. Anomaly segmentation aims to detect the outlier class object in some specific scenes, such as complex driving scenes [6, 13, 31]. Existing anomaly segmentation can be broadly divided into three categories: (a) Uncertainty-based methods, (b) Reconstruction-based methods and (c) Outlier Exposure methods. Uncertainty-based methods assume that outlier samples result in low-confidence predictions. Based on this assumption, they focus on measuring the pixel-wise anomaly score by estimating the prediction uncertainty through various ways, such as ensembles [27], Bayesian estimation [16], maximum softmax probability [21, 29], and logit [20, 26]. Since segmentation models are trained on close-set, they may misclassify unseen classes with high confidence, leading uncertainty-based methods fail to detect anomalies. Reconstruction-based methods detect anomaly pixel by comparing the discrepancy between the original image and its reconstructed image generated based on the semantic segmentation predictions [14, 15, 31, 35]. These methods are challenged by the accuracy of segmentation prediction and the difficulties of training. Recent methods [7, 12, 19, 34, 36, 37] introduce the Outlier Exposure (OE) strategy [22] that utilizes auxiliary outlier data and anomaly scoring functions to learn models. The auxiliary outlier data is the image from other datasets (e.g., COCO [30]), or cutting the outlier objects from other datasets and pasting them into the inlier scenes, and the anomaly scoring functions are used to calculate the anomaly score for each pixel. Compared with previous works, OE strategy can reach better performance. More recently, some OE methods [19, 34, 36] adopt mask-based segmentation networks [10, 11], instead of pixel-based segmentation networks [8, 9], as their backbones. Based on the powerful capability of mask-based networks in segmentation tasks, these methods significantly reduces false detection rates, achieving SOTA results. In this paper, we only concentrate on the OE methods. The anomaly scoring functions employed in OE methods are based on the predicted logits of inlier classes. However, during the inference phase, the diversity of driving scenes results in distorted manifolds of pixel embeddings. As a result, the scoring functions may fail to use the embeddings to accurately calculate logits, thereby affects the accuracy of measuring anomaly scores. Current OE methods ignore this problem. Our proposed RWPM addresses this problem by utilizing random walks on pixel manifolds to reveal the intrinsic relationships among pixels and refine the pixel embeddings. To the best of our knowledge, RWPM is the first work to propose and resolve this problem.", + "bbox": [ + 212, + 311, + 787, + 843 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z. Zeng et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mining on Data Manifolds. Mining on data manifolds, also known as diffusion process, is a well-established technique commonly used in retrieval tasks [3, 4, 25, 40, 43]. These works utilize random walks on data manifolds to capture the data manifold structure for revealing the intrinsic relationship between data points. The data manifold is constructed by a graph, where each node represents a sample and each edge presents the connection between a node and its neighborhoods, weighted proportionally according to their distance. For retrieval tasks, the images of the same object may be shown in different conditions (e.g., lighting and angles). This diversity often leads to the manifold structure is not conducive to accurately ranking using distance metrics directly [25, 40]. Thus, diffusion process is used to diffuse the state (typically initialized to 1) of a query image on the data manifolds. If a sample is on the same data manifold with the query (the distance on manifolds is close), then more energy of the state will flow to that sample. Therefore, the final state of all samples can be seen as the accurate ranking scores. Moreover, diffusion process is also adopted in semi-supervised learning [24, 42] to propagate the label of labeled samples to unlabeled samples. Similarly, some unsupervised segmentation works [17, 23] utilize diffusion process to propagate the labels of manually annotated pixels to the unlabeled pixels. These existing works utilize random walks on data manifolds for ranking or label propagation. Different from these works, our target is to solve the problem occurring in pixel manifolds for anomaly scoring functions. We propose innovatively using random walks to diffuse and update pixel embeddings, eliminating the effect of distorted manifolds directly at the embedding-level, thereby improving the accuracy of subsequent processing.", + "bbox": [ + 212, + 145, + 787, + 510 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Proposed Method", + "text_level": 1, + "bbox": [ + 215, + 530, + 426, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 559, + 372, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We denote an image space $\\mathcal{X} \\subset \\mathbb{R}^{3 \\times H \\times W}$ , where $H$ and $W$ are the height and width, respectively. During the inference phase, the goal for an anomaly segmentation model $f$ is to map an input image $\\pmb{x} \\in \\mathcal{X}$ to an anomaly space, i.e., $f: \\pmb{x} \\mapsto \\mathcal{Z} \\subset \\mathbb{R}^{H \\times W}$ . Specifically, an anomaly segmentation model typically consists of the backbone encoder-decoder $F$ and the classifier $C$ . For an input image $\\pmb{x}$ , firstly, its pixel embedding map is encoded as $\\pmb{p} = F(\\pmb{x}) \\in \\mathbb{R}^{d \\times H \\times W}$ , where $d$ is the dimension of the pixel embedding. Then, its logit map $\\pmb{l}$ is predicted by $\\pmb{l} = C(F(\\pmb{x})) \\in \\mathbb{R}^{K \\times H \\times W}$ , where $K$ is the number of inlier classes. Note that, depending on the backbone network, the size of the embedding map and logit map may sometimes be smaller than the original image. For the sake of convenience in this paper, we do not make a special distinction. Finally, a scoring function is used to calculate the anomaly score map $\\pmb{z} \\in \\mathcal{Z}$ based on the logit map $\\pmb{l}$ .", + "bbox": [ + 212, + 582, + 787, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For anomaly scoring functions, their logic can be represented as follows: when the predicted logits of a pixel embedding are small across all inlier classes, indicating that the pixel is dissimilar to all inlier classes and is likely an outlier, the function generates a higher anomaly score. For example, let's define $l_{h,w}(k)$", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "RWPM", + "bbox": [ + 679, + 113, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0cc1a045b91634a7ad204f533266b3087e3c813bfb6a897fc97866f134f6e246.jpg", + "image_caption": [ + "Fig. 2: Overview. This figure illustrates the application of RWPM to an existing anomaly segmentation framework during the inference phase. The red dashed box highlights the RWPM part. First, the encoder-decoder is used to extract the pixel embedding map of the input image. The pixel embedding map are subsequently partitioned into $n^2$ sub-maps. For each sub-map, we update its pixel embeddings by using random walks to obtain a refined sub-map. Finally, the refined sub-maps are concatenated to form the refined embedding map, which are then input into the subsequent network structure. Notably, RWPM can be directly integrated into existing frameworks without requiring extra training or changes to the network structure." + ], + "image_footnote": [], + "bbox": [ + 230, + 145, + 774, + 252 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "is the logit of a pixel $\\pmb{x}_{h,w}$ for class $k$ . The scoring function in PEBAL [37] is $E_{h,w} = -\\log \\sum_{k \\in \\{1, \\dots, K\\}} \\exp(\\pmb{l}_{h,w}(k))$ . Obviously, when the anomaly score $E_{h,w}$ is large, it indicates that the logits of the pixel across all inlier classes are small, i.e., the pixel $\\pmb{x}_{h,w}$ is dissimilar from all inlier classes (see Section S.2 of the supplementary material for the analysis of other anomaly scoring functions).", + "bbox": [ + 212, + 416, + 787, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Graph Construction", + "text_level": 1, + "bbox": [ + 215, + 515, + 429, + 531 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We begin by constructing a graph to represent the manifolds of pixel embeddings. In our approach, we utilize an affinity matrix based on cosine similarity to characterize the relationships between pixels. Specifically, we first use the encoder-decoder component of an anomaly segmentation model to extract the pixel embeddings $\\pmb{p} \\in \\mathbb{R}^{d \\times H \\times W}$ . For ease of description and computation, we reshape the $\\pmb{p}$ as $\\pmb{p}^r \\in \\mathbb{R}^{HW \\times d}$ , where each row represents a pixel's embedding. We then normalize each row of $\\pmb{p}^r$ to obtain $\\widehat{\\pmb{p}}^r$ , i.e., $\\widehat{\\pmb{p}}_i^r = \\pmb{p}_i^r / \\| \\pmb{p}_i^r \\|$ . The affinity matrix is defined as $\\mathbf{W} \\in \\mathbb{R}^{HW \\times HW}$ , where each component is obtained by:", + "bbox": [ + 212, + 541, + 787, + 662 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {W} _ {i j} = \\left\\{ \\begin{array}{c c} \\langle \\hat {\\pmb {p}} _ {i} ^ {r}, \\hat {\\pmb {p}} _ {j} ^ {r} \\rangle & i \\neq j \\\\ 0 & i = j \\end{array} , \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 672, + 785, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\langle \\cdot, \\cdot \\rangle$ denotes inner product, $\\mathbf{W}_{ii} = 0$ is guaranteed to circumvent self-loops in the graph represented by $\\mathbf{W}$ .", + "bbox": [ + 212, + 718, + 785, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To mitigate the influence of noise, locally constrained random walk methods [24, 40, 41] adopt local constraints to construct the graph, i.e., each point is only connected to its locally nearest neighbors in the graph. Determining the nearest neighbors for each point traditionally involves applying nearest neighbor search algorithms, such as the $k$ -NN algorithm. However, in anomaly segmentation tasks, the size of the graph is large and the graph is constructed online,", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z. Zeng et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "making the search algorithms impractical. To benefit from the local constraints, we utilize a softmax function to normalize each row of $\\mathbf{W}$ to obtain locally constrained graph $\\mathbf{S}$ :", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} _ {i j} = \\left\\{ \\begin{array}{c c} \\frac {\\exp \\left(\\mathbf {W} _ {i j} / \\tau\\right)}{\\sum_ {i = 1 , i \\neq j} ^ {H W} \\exp \\left(\\mathbf {W} _ {i j} / \\tau\\right)} & i \\neq j \\\\ 0 & i = j \\end{array} , \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 382, + 203, + 785, + 242 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\tau$ is a temperature hyper-parameter, we set its value less than 1.0 to remove the effect of non-neighbor points. Compared to typical search algorithms, Eq. 2 can impose local constraints more efficiently and effectively (see Section S.4 of the supplementary material), and keep the scale of embedding in subsequent random walks.", + "bbox": [ + 212, + 253, + 784, + 328 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Random Walk Process", + "text_level": 1, + "bbox": [ + 215, + 351, + 449, + 366 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Now we have constructed the graph $\\mathbf{S}$ , we can start the random walks. Different from previous mining on data manifolds works utilizing random walks to propagate the states of known samples to other unknown samples, we use a random walk to propagate the embedding between pixels on manifolds.", + "bbox": [ + 212, + 377, + 782, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first define $\\mathbf{m}^t\\in \\mathbb{R}^{HW\\times d}$ as the updated pixel embeddings in the $t$ -th step of the random walks. $\\mathbf{m}^0$ is initialized as $\\pmb{p}^r$ . Then, each iteration step of the random walks is described as:", + "bbox": [ + 212, + 436, + 784, + 481 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {m} ^ {t + 1} = \\alpha \\mathbf {S} \\mathbf {m} ^ {t} + (1 - \\alpha) \\mathbf {m} ^ {0}, \\alpha \\in (0, 1), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 494, + 784, + 511 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\alpha$ is the probability continuing the random walk from the current state $\\mathbf{m}^t$ , while $(1 - \\alpha)$ represents the probability of restarting from the initial state $\\mathbf{m}^0$ . At each iteration, guided by the manifold structure, each pixel diffuses its embedding to other pixels while also receiving embedding from other pixels for updating its own embedding. The updated pixel embedding can be seen as a weighted ensemble of other pixel embeddings. If pixels are similar on manifolds, their embeddings will gradually become similar. Mathematically speaking, the whole iteration process can be written as a closed-from as follows [43]:", + "bbox": [ + 212, + 523, + 784, + 643 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {m} ^ {\\infty} = (1 - \\alpha) (\\mathbf {I} - \\alpha \\mathbf {S}) ^ {- 1} \\mathbf {m} ^ {0}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 657, + 784, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{I} \\in \\mathbb{R}^{HW \\times HW}$ is an identity matrix. Finally, $\\mathbf{m}^{\\infty}$ is used as the refined pixel embeddings for the subsequent anomaly segmentation process.", + "bbox": [ + 212, + 684, + 782, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Partial Random Walk", + "text_level": 1, + "bbox": [ + 215, + 738, + 442, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "When the size of the image is large, the size of the relevant matrices used for the calculation of random walks becomes extremely huge as well. For instance, when the image size is $(512 \\times 1024)$ , the size of matrix $S$ reaches $(524288 \\times 524288)$ . This significantly increases the memory consumption and computational time of RWPM, which is not conducive to online operation. Unfortunately, in the field", + "bbox": [ + 212, + 763, + 784, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "RWPM", + "bbox": [ + 679, + 113, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of autonomous driving, the image size of driving scenes is generally large. To address this problem, we propose Partial Random Walk, which consists of the following two aspects: embedding map partitioning and limited iteration.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To improve efficiency, we first propose embedding map partitioning strategy to reduce the size of the pixel embedding map used for random walks. Specifically, for a pixel embedding map $\\pmb{p} \\in \\mathbb{R}^{d \\times H \\times W}$ , we equally divide it into $n \\times n$ sub-maps $\\left\\{\\pmb{p}^{(1)}, \\dots, \\pmb{p}^{(n^2)}\\right\\} \\in \\mathbb{R}^{d \\times \\frac{H}{n} \\times \\frac{W}{n}}$ . Then, for each sub-map in $\\left\\{\\pmb{p}^{(1)}, \\dots, \\pmb{p}^{(n^2)}\\right\\}$ , we follow the steps of Eq. 1 and Eq. 2 to construct their sub-graphs $\\left\\{\\pmb{S}_{\\pmb{p}^{(1)}}, \\dots, \\pmb{S}_{\\pmb{p}^{(n^2)}}\\right\\} \\in \\mathbb{R}^{d \\times \\frac{HW}{n^2} \\times \\frac{HW}{n^2}}$ . We also attempted other methods to reduce the number of pixel embeddings, such as resize or superpixel [1]. Experimental results indicate that embedding map partitioning is the optimal solution among them (see Section 4.2).", + "bbox": [ + 212, + 191, + 784, + 351 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Next, for each sub-map $\\pmb{p}^{(i)}$ , we perform random walks with the limited iteration strategy in their corresponding sub-graphs $S_{\\pmb{p}^{(i)}}$ . That is, we use Eq. 3 for a small number of iterations instead of using the closed-form Eq. 4. We found that a few iterations by using Eq. 3 are enough to yield high quality results. The time complexity of Eq. 3 with $T$ iterations for each sub-map is $O(Td(\\frac{HW}{n^2})^2)$ , while the time complexity of Eq. 4 is $O((\\frac{HW}{n^2})^3) + O(d(\\frac{HW}{n^2})^2)$ . Since $Td$ is much smaller than $\\frac{HW}{n^2}$ , the limited iteration strategy can significantly improve the efficiency of RWPM. After the random walks, we obtain the refined pixel embeddings sub-map $\\mathbf{m}_{\\pmb{p}^{(i)}}^T$ for each sub-map $\\pmb{p}^{(i)}$ . Finally, we concatenate the refined sub-maps based on their positions to obtain the final refined pixel embedding map. Additionally, when $n$ is large (observed experimentally as $n > 2$ ), due to spatial constraints, the baseline values of anomaly scores outputted on each sub-map may vary. Hence, a calibration is performed. For adjacent refined sub-maps $\\pmb{p}^{(i)'}$ and $\\pmb{p}^{(j)'}$ , we compute the average anomaly scores $I$ and $J$ of all pixels on their adjacent edges, respectively. Subsequently, using $\\pmb{p}^{(i)'}$ as the reference, we calculate the ratio $\\frac{I}{J}$ , and then multiply all anomaly scores on sub-map $\\pmb{p}^{(j)'}$ by this ratio to achieve the calibration.", + "bbox": [ + 212, + 352, + 785, + 613 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 2 shows the workflow of the proposed RWPM.", + "bbox": [ + 238, + 614, + 607, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 215, + 655, + 366, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dataset: We evaluate our approach on three standard benchmark datasets for anomaly segmentation, i.e., Fishyscapes Lost&Found [5], Road Anomaly [31] and Segment Me If You Can (SMIYC) [6]. Fishyscapes Lost&Found has 100 validation images. The domain of this dataset is similar to that of Cityscapes [13] dataset. Road Anomaly dataset contains 60 images of real-world road scenes. Unlike the Fishyscapes Lost&Found dataset, the images in Road Anomaly contain various classes and sizes of anomaly objects, making it more challenging. SMIYC consists of two subsets, Anomaly track and Obstacle track. The Anomaly track focuses on detecting large anomalous objects, while the Obstacle track emphasizes detecting small anomalous objects on the road. Anomaly track contains 10", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z. Zeng et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "validation images and 100 online test images, while Obstacle track contains 30 validation images and 327 online test images. The Road Anomaly and SMIYC datasets both exhibit high domain shifts compared to the Cityscapes dataset.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation Metrics: Following the previous works, we evaluate our method on the standard metrics for anomaly segmentation. On the Fishyscapes Lost&Found and Road Anomaly, we adopt the area under receiver operating characteristics (AUROC), average precision (AP), and the false positive rate at a true positive rate of $95\\%$ (FPR95). On the SMIYC, in addition to reporting per-pixel metrics such as AP and FPR95, we also calculate results for component-level metrics [6] such as the component-wise intersection over union (sIoU), positive predictive value (PPV), and mean F1.", + "bbox": [ + 212, + 191, + 785, + 311 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details: To demonstrate the effectiveness of RWPM for anomaly segmentation methods, we selected representative methods from both pixel-based and mask-based networks. For pixel-based methods, we chose PE-BAL [37] and Balanced Energy [12]. These methods all utilize DeepLabv3+ [9] with WideResnet38 as the backbone and were trained on the Cityscapes dataset and COCO OE images [37]. For mask-based methods, we selected RbA [34] and Mask2Anomaly [36]. These methods use Mask2Former [10] with Swin-L [33] as the backbone and were also trained on the Cityscapes and COCO OE images. All experiments are conducted on a single NVIDIA RTX A6000 GPU. To ensure efficient implementation on a single GPU for all experiments, we adopt embedding map partitioning and limited iteration in all experiments. Specifically, for pixel-based methods, we set the partitioning parameter $n = 4$ . For mask-based methods, as their embedding map size is smaller, we set $n = 2$ . Additionally, for the random walk, if not specifically emphasized, we set the number of iterations $T = 20 / T = 5$ for the Road Anomaly/the other datasets and the transition probability $\\alpha = 0.99$ for all datasets. The $\\tau$ is set as 0.01. It is important to note that our method is used in the inference phase. In other words, RWPM is directly applied to the pre-trained models of other existing anomaly segmentation methods, thus requiring no additional training.", + "bbox": [ + 212, + 313, + 787, + 602 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Effectiveness of RWPM", + "text_level": 1, + "bbox": [ + 215, + 623, + 457, + 638 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "First, to validate that RWPM can improve the performance of existing anomaly segmentation methods, we integrate our approach with four recent representative methods, namely PEBAL [37], Balanced Energy [12], Mask2Anomaly [36] and RbA [34]. PEBAL and Balanced Energy employ pixel-based networks as their backbones, whereas Mask2Anomaly and RbA utilize mask-based networks as their backbones. We evaluate the results on both the Fishyscapes Lost&Found validation set and Road Anomaly test set. The experimental results are reported in Table 1. These results demonstrate that our proposed RWPM consistently and significantly enhances the performance of existing anomaly segmentation methods, across different backbone architectures. Additional results and parameters details are shown in Section S.3 of the supplementary material.", + "bbox": [ + 212, + 648, + 787, + 816 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "RWPM", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "2 We carefully reproduce the experiments by using the official code.", + "bbox": [ + 217, + 824, + 671, + 839 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/6e9b9d9f65e060ef502e8a8c488e8f6e22c8fa50a180a1d3fa2524f9d3b17d31.jpg", + "table_caption": [ + "Table 1: Comparison with strong and representative anomaly segmentation baselines across different backbone Architectures (Arch). Bold denotes the better results between using RWPM and not using RWPM with the same anomaly segmentation baselines. $\\dagger$ indicates that use the calibration in sub-map concatenation as described in Section 3.4." + ], + "table_footnote": [], + "table_body": "
Benchmark→ Fishscapes Lost&Found Road Anomaly
Method ↓ArchAuROC↑AP↑FPR95↓AuROC↑AP↑FPR95↓
PEBAL [37]Pixel-based98.9658.814.7787.6345.1044.58
PEBAL + RWPM† (Ours)Pixel-base99.2066.853.6889.4850.2936.81
Balanced Energy [12]Pixel-based99.0367.072.9388.3141.4841.46
Balanced Energy + RWPM† (Ours)Pixel-base99.2972.952.3890.5148.2632.10
Mask2Anomaly2 [36]Mask-based95.4765.277.7996.5480.0413.95
Mask2Anomaly + RWPM (Ours)Mask-base95.4865.387.3397.4480.097.45
RbA [34]Mask-based98.6270.816.3097.9985.426.92
RbA + RWPM (Ours)Mask-base98.8271.166.1298.0487.345.27
", + "bbox": [ + 272, + 226, + 728, + 329 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/f4602b08a7368d13dce289755f65213222d6f4abf80d21cfe1cdda10c36152e5.jpg", + "table_caption": [ + "Table 2: Comparison with state-of-the-art anomaly segmentation methods across different backbone Architectures (Arch). Bold and underline denote the best and the second best results, respectively." + ], + "table_footnote": [], + "table_body": "
Benchmark→Fishscapes L&FRoad AnomalyAnomaly TrackObstacle TrackAverage
Method ↓ArchAP↑FPR95↓AP↑FPR95↓AP↑FPR95↓AP↑FPR95↓AP↑FPR95↓
Synboost [15] (CVPR'21)Pixel-based60.5831.0241.8459.7256.4461.8671.343.1557.4638.94
SML [26] (ICCV'21)Pixel-based22.7433.4925.8249.4746.839.53.436.824.6939.82
Meta-OOD [7] (ICCV'21)Pixel-based41.3137.6948.8431.7785.4715.0085.070.7565.1321.30
Learning Embedding [5](IJCV'21)Pixel-based4.6524.36--37.5270.760.8246.38--
Void Classifier [5](IJCV'21)Pixel-based10.2922.11--36.6163.4910.4441.54--
GMMSeg-DL [28] (NeurIPS'22)Pixel-based43.4713.1134.4247.90------
DenseHybrid [18] (ECCV'22)Pixel-based69.795.0931.3963.9777.969.8187.080.2465.5619.78
PEBAL [37] (ECCV'22)Pixel-based58.814.7745.1044.5849.1440.824.9812.6839.5825.71
Balanced Energy [12] (CVPR'23)Pixel-based67.072.9341.4841.46------
RPL-CoroCL [32] (ICCV'23)Pixel-based70.612.5271.6117.7483.4911.6885.930.5877.918.13
Mask2Anomaly [36] (ICCV'23)Mask-based69.449.2280.0413.9588.6214.5793.100.2082.879.49
M2F-EAM [19] (CVPRW'23)Mask-based52.0320.5166.6713.4276.393.966.917.965.4836.43
RbA [34] (ICCV'23)Mask-based70.816.3085.426.9290.8611.5991.850.4684.736.33
RbA + RWPM (Ours)Mask-based71.166.1287.345.2792.0010.1593.300.2886.005.46
", + "bbox": [ + 218, + 401, + 787, + 560 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Second, to further demonstrate the effectiveness of our RWPM, we compare it with other state-of-the-art (SOTA) methods. Specifically, we adopt the combination of RbA with RWPM as our instance approach for the comparison. In addition to the Fishyscapes and Road Anomaly datasets, we also evaluate our approach on a more challenging dataset, i.e., SMIYC online test set (Anomaly track and Obstacle track). SMIYC is characterized by its high domain shift and diversity of objects. Table 2 presents the results. The results indicate that our method can consistently achieve high performance across various datasets, while the previous SOTA methods fail to do so. We also compare all approaches by computing the average results of AP and FPR95. The results show that our approach outperforms the previous methods and achieve the best results.", + "bbox": [ + 212, + 595, + 787, + 762 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Furthermore, in addition to pixel-level metrics, we also utilize component-level metrics [6] to evaluate performance on the SMIYC online test set. Compared to pixel-level metrics, component-level metrics provide a better evaluation of the detection results for all anomalous objects in the scene. We report the results in Table 3. The experimental results show that RWPM significantly improves", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z. Zeng et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "the performance of RbA in component-level metrics, surpassing other existing SOTA methods by a large margin. This indicates that RWPM can assist models in better detecting anomalous objects and achieving more precise localization of these anomalies [6].", + "bbox": [ + 212, + 146, + 787, + 208 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/ece8203819a604e5d32cd9147eb9bb8e1c9073c8b09f7d107620d63535ce826f.jpg", + "table_caption": [ + "Table 3: Component level evaluation: Comparison with state-of-the-art anomaly segmentation methods on the SMIYC online test set. Bold denote the best results." + ], + "table_footnote": [], + "table_body": "
Benchmark→Anomaly TrackObstacle Track
Method ↓ArchsIoU ↑PPV↑mean F1↑sIoU ↑PPV↑mean F1↑
Synboost [15] (CVPR'21)Pixel-based34.6817.819.9944.2841.7537.57
SML [26] (ICCV'21)Pixel-based26.0024.7012.205.1013.303.00
Meta-OOD [7] (ICCV'21)Pixel-based49.3139.5128.7247.8763.6448.51
Void Classifier [5](IJCV'21)Pixel-based21.1422.136.496.3420.275.41
Learning Embedding [5](IJCV'21)Pixel-based33.8620.547.9035.642.872.31
JSRNet [38] (ICCV'21)Pixel-based20.2029.2713.6618.5524.4611.02
DenseHybrid [18] (ECCV'22)Pixel-based54.1724.1331.0845.7450.1050.72
PEBAL [37] (ECCV'22)Pixel-based38.8827.2014.4829.917.555.54
RPL-CoroCL [32] (ICCV'23)Pixel-based49.7629.9630.1652.6156.6556.69
Mask2Fromer [36] (ICCV'23)Mask-based25.2018.2015.305.0021.904.80
Maskomaly [2] (BMVC'23)Mask-based55.451.650.0---
RbA [34] (ICCV'23)Mask-based55.6952.1446.8058.3658.7860.85
RbA + RWPM (Ours)Mask-based57.0061.2558.4458.8972.5169.85
", + "bbox": [ + 220, + 273, + 787, + 431 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 482, + 388, + 498 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "If not specifically emphasized, all the results reported in this section are conducted on a RTX A6000 GPU.", + "bbox": [ + 212, + 507, + 785, + 536 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The effect of the partitioning parameter $n$ and calibration: We first test the effect of the embedding map partitioning parameter $n$ and the calibration, and we report the results in Table 4. FPS denotes the frames-per-second results. Row $n = 1$ displays the RWPM results without utilizing the embedding map partitioning strategy, indicating significant computational and memory overhead. However, the partitioning strategy ( $n = 2, 4, 8$ ) significantly improves the running speed and reduces memory consumption while maintaining effectiveness. Additionally, as $n$ increases, the running efficiency also improves, albeit with a slight decrease in performance. However, when $n > 2$ , the calibration can enhance the performance, particularly for the FPR95 metrics.", + "bbox": [ + 212, + 537, + 787, + 688 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The effect of iteration number $T$ : Table 5 shows that, compared to using Eq. 4 ( $T = \\infty$ ), utilizing the limited iteration strategy ( $T = 5 \\sim 100$ ) can lead to higher efficiency, resulting in even better performance for RWPM. Additionally, as the number of iterations increases, performance tends to improve, but running efficiency (FPS) decreases accordingly.", + "bbox": [ + 212, + 688, + 787, + 763 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Different strategies for reducing the number of pixel embedding: We also attempt other strategies to reduce the number of pixel embedding for RWPM, such as resizing the embedding map or use SLIC [1] algorithm to extract superpixels. Table 6(Left) presents the results. The resize strategy achieves good performance but requires preserving a larger map size. Therefore, it fails", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "RWPM", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/e7034d61e2de45c479c1e29f42489f1365e9cdfed056d6be79516809a414ff66.jpg", + "table_caption": [ + "Table 4: The effect of the partitioning parameter $n$ and calibration. We set $T = {20}$ for all experiments. ${}^{ \\dagger }$ indicates that use the calibration in sub-map concatenation." + ], + "table_footnote": [], + "table_body": "
Benchmark→Road Anomaly
AP↑FPR95↓FPS↑GPU Memory Usage↓
RbA w/o RWPM85.426.9211.123.49GiB
n=1 (two GPUs)87.905.170.3274.96GiB
n=287.345.272.047.24GiB
n=2†87.185.252.047.24GiB
n=487.056.314.353.49GiB
n=4†87.175.324.353.49GiB
n=886.676.426.263.55GiB
n=8†86.915.416.263.55GiB
", + "bbox": [ + 328, + 184, + 671, + 295 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/2590a223a96f8ebbe925195f3ed69338d07de91b36c75eeba15eb9c7f2eec250.jpg", + "table_caption": [ + "Table 5: The effect of iteration number $T$ . We set $n = 2$ for all experiments." + ], + "table_footnote": [], + "table_body": "
Benchmark→ Road Anomaly
AP↑FPR95↓FPS↑GPU Memory Usage↓
RbA w/o RWPM85.426.9211.123.49GiB
T=∞87.515.240.298.88GiB
T=586.895.983.707.24GiB
T=1087.085.772.787.24GiB
T=2087.345.272.047.24GiB
T=5087.365.080.927.24GiB
T=10087.545.020.507.24GiB
", + "bbox": [ + 328, + 335, + 671, + 431 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "to substantially reduce the number of pixel embeddings, resulting in significant GPU memory usage. On the other hand, the superpixel strategy requires a considerable amount of runtime for the superpixel extraction. In comparison, our proposed partitioning strategy can achieve good results with higher efficiency and lower resource consumption. This makes it more suitable for applications on practical platforms.", + "bbox": [ + 212, + 460, + 787, + 551 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In-distribution segmentation performance on Cityscapes validation set: We report the results on the Cityscapes validation set in Table 6(Right). All methods use Mask2Former [10] as their backbones. We observe that most SOTA methods affect the in-distribution segmentation performance of Mask2Former, such as RbA (drop from 82.25 to 81.93). However, RWPM can alleviate this problem. Additionally, even for the SOTA segmentation method (Mask2Former), RWPM can further enhance its in-distribution segmentation performance.", + "bbox": [ + 212, + 551, + 787, + 659 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/c94fa5e457112aed6076439fd0cdd0f7c2a43759761404ada407ad65b49f769f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Benchmark→ Road Anomaly
Strategy↓AP↑FPR95↓FPS↑Memory
Resize (0.9 full size)87.805.391.4123.82GIB
Resize (0.8 full size)87.245.442.1515.98GiB
Superpixel (30000 pixels)87.515.500.1911.46GiB
Superpixel (20000 pixels)87.035.980.326.31GiB
Partitioning (n = 2)87.345.272.047.24GiB
", + "bbox": [ + 218, + 758, + 504, + 832 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/6f9c75e324ea804073c35260d96260aa04a21cd1edc52fbe0b42bd4b3c96067c.jpg", + "table_caption": [ + "Table 6: Ablation Tables: (Left) shows the results of different strategies for reducing the number of pixel embedding. We find that partitioning is the best. (Right) shows the segmentation performance on the Cityscapes validation set. RWPM can also improve the segmentation performance of in-distribution." + ], + "table_footnote": [], + "table_body": "
Benchmark→Cityscapes
Method↓mIoU↑
PEBAL-Mask [34]75.32
DenseHybrid-Mask [34]80.27
RbA [34]81.93
RbA+RWPM(Ours)82.16
Mask2Former [10]82.25
Mask2Former+RWPM(Ours)82.43
", + "bbox": [ + 545, + 753, + 743, + 835 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z. Zeng et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Visualization of Pixel Embedding Distribution", + "text_level": 1, + "bbox": [ + 215, + 146, + 643, + 161 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To investigate the effect of RWPM in the embedding space, we visualized the distribution of pixel embeddings by using t-SNE. We selected an image from the Road Anomaly dataset (as shown in Fig 3(a)). Due to the gravel road surface in this image, there is a significant domain gap between the road in this image and the roads in the training set, resulting in many false positives on the road (as shown in the PEBAL results in Fig 3(a)). Then, we randomly selected 300 inlier pixels and 150 outlier pixels from the red dashed box region of this image for visualization. Fig 3(b) and Fig 3(c) respectively show the distributions of these pixels in the space without and with applying RWPM. In Fig 3(b) and Fig 3(c), deep blue circles represent inlier pixels (road), light blue circles represent outlier pixels (boar), and X symbols represent the inlier class prototype (the vector of network weights of classifier). From Fig 3(b), we observe that without using RWPM, due to the influence of data diversity, the distribution of inliers and outliers is distorted. Consequently, it is difficult for an anomaly scoring function to distinguish between inlier and outlier pixels based on their logits. However, from Fig 3(c), we find that RWPM can significantly alleviate the problem. Specifically, the pixel embeddings of the same category form more compact clusters, enabling the anomaly scoring function to clearly discriminate between inliers and outliers based on the logit. This is consistent with the description in the toy sample in Fig 1. Note that, in the top left corner of Fig 3(c), some inlier points become very close to outlier points after applying RWPM. This is because these points are located in the boundaries of anomalous objects, making it very difficult to discriminate them.", + "bbox": [ + 212, + 169, + 787, + 517 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/6dd5b6908bd5cb09c02c161f71c5dfd35a89c4d5e7069e3b5ab17ee3a1d54722.jpg", + "image_caption": [ + "(a) Sample Image" + ], + "image_footnote": [], + "bbox": [ + 222, + 545, + 377, + 645 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e2e53be7e629c7892031c9d2987f959c7d9189311c0793d88e366dbe9e2abd20.jpg", + "image_caption": [ + "Fig. 3: The visualization of pixel embedding distribution. All embeddings are extracted from PEBAL model. We use cosine distance as the distance metric. We observed that RWPM can optimize the distribution of pixel embeddings in the space." + ], + "image_footnote": [], + "bbox": [ + 393, + 550, + 571, + 641 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ed945554d75c59edb7f825bfe123f491a8baa0a13b4cf05cac18a6fb9a3371d6.jpg", + "image_caption": [ + "(b) Distribution w/o RWPM", + "(c) Distribution with RWPM" + ], + "image_footnote": [], + "bbox": [ + 602, + 551, + 777, + 641 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Qualitative Result", + "text_level": 1, + "bbox": [ + 215, + 770, + 413, + 786 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Finally, to visually demonstrate the effectiveness of RWPM, we present qualitative results. Specifically, we compare the results of RbA on the Road Anomaly dataset without and with RWPM. The qualitative results are illustrated in Fig 4.", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "RWPM", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Fig 4 shows that RWPM significantly improves the performance of RbA, reducing both the false positive rate and false negative rate. For example, in the first row, RbA generates some false positive results on the road and some false negative results on the manhole region (third column of the first row). However, after applying RWPM, these false positive/negative results are effectively eliminated (fourth column of the first row). Additionally, RWPM assists the RbA model to accurately detect and localize each anomalous object at the component level, such as more precise edge detection. This observation aligns with the results of the component-level metrics presented in Table 3.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/cc0f8c32e7c4447fe9c9c6d6ca48e6b2e3e8f9f23f94724ed4996462dff56e83.jpg", + "image_caption": [ + "Fig. 4: Qualitative results: The first column displays the input image, while the second column shows its corresponding ground truth. The third column displays the anomaly score results generated by RbA, and the fourth column presents the anomaly score results after applying RWPM to RbA. Yellow indicates high anomaly scores." + ], + "image_footnote": [], + "bbox": [ + 233, + 306, + 772, + 486 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusions and Discussions", + "text_level": 1, + "bbox": [ + 214, + 599, + 527, + 617 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we propose a simple yet effective method, called Random Walk on Pixel Manifolds (RWPM), to improve the performance of existing anomaly segmentation methods. RWPM is a post-processing method that can be directly integrated into existing anomaly detection frameworks without requiring additional training or modifications to the network structure. RWPM mitigates the impact of data diversity on manifold structure by diffusing and updating pixel embeddings on pixel manifolds using random walk. This process enables anomaly scoring functions to obtain more accurate estimations of anomaly scores. Additionally, we propose Partial Random Walk strategy, which significantly improves the efficiency of RWPM.", + "bbox": [ + 212, + 628, + 787, + 779 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "While RWPM achieves significant results in most cases, it is limited when the pre-trained model itself exhibits significant inaccuracies. To address this limitation, we plan to enhance the model's generalization ability during training to accurately detect a broader range of diverse and unknown anomalies.", + "bbox": [ + 212, + 779, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z. Zeng et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by SenseTime Japan.", + "bbox": [ + 215, + 174, + 555, + 189 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 210, + 321, + 224 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Achanta, R., Shaji, A., Smith, K., Lucchi, A., Fua, P., Süsstrunk, S.: Slic superpixels compared to state-of-the-art superpixel methods. IEEE transactions on pattern analysis and machine intelligence 34(11), 2274-2282 (2012)", + "2. Ackermann, J., Sakaridis, C., Yu, F.: Maskomaly: Zero-shot mask anomaly segmentation. arXiv preprint arXiv:2305.16972 (2023)", + "3. Bai, S., Bai, X., Tian, Q., Latecki, L.J.: Regularized diffusion process for visual retrieval. In: Proceedings of the AAAI conference on artificial intelligence. vol. 31 (2017). https://doi.org/10.1609/aaai.v31i1.11216", + "4. Bai, S., Zhou, Z., Wang, J., Bai, X., Jan Latecki, L., Tian, Q.: Ensemble diffusion for retrieval. In: Proceedings of the IEEE International conference on computer vision. pp. 774-783 (2017)", + "5. Blum, H., Sarlin, P.E., Nieto, J., Siegwart, R., Cadena, C.: The fishyscapes benchmark: Measuring blind spots in semantic segmentation. International Journal of Computer Vision 129, 3119-3135 (2021). https://doi.org/10.1007/s11263-021-01511-6", + "6. Chan, R., Lis, K., Uhlemeyer, S., Blum, H., Honari, S., Siegwart, R., Fua, P., Salzmann, M., Rottmann, M.: Segmentmeifyoucan: A benchmark for anomaly segmentation. arXiv preprint arXiv:2104.14812 (2021)", + "7. Chan, R., Rottmann, M., Gottschalk, H.: Entropy maximization and meta classification for out-of-distribution detection in semantic segmentation. In: Proceedings of the ieee/cvf international conference on computer vision. pp. 5128-5137 (2021)", + "8. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence 40(4), 834-848 (2017)", + "9. Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 801-818 (2018). https://doi.org/10.1007/978-3-030-01234-2_49", + "0. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290–1299 (2022). https://doi.org/10.1109/CVPR52688.2022.00135", + "1. Cheng, B., Schwing, A., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems 34, 17864-17875 (2021)", + "2. Choi, H., Jeong, H., Choi, J.Y.: Balanced energy regularization loss for out-of-distribution detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15691-15700 (2023)", + "3. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3213-3223 (2016). https://doi.org/10.1109/CVPR.2016.350" + ], + "bbox": [ + 225, + 238, + 784, + 837 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "RWPM", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Creusot, C., Munawar, A.: Real-time small obstacle detection on highways using compressive rbm road reconstruction. In: 2015 IEEE Intelligent Vehicles Symposium (IV). pp. 162-167. IEEE (2015)", + "15. Di Biase, G., Blum, H., Siegwart, R., Cadena, C.: Pixel-wise anomaly detection in complex driving scenes. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16918-16927 (2021). https://doi.org/10.1109/CVPR46437.2021.01664", + "16. Gal, Y., Ghahramani, Z.: Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In: international conference on machine learning. pp. 1050-1059. PMLR (2016)", + "17. Grady, L.: Random walks for image segmentation. IEEE transactions on pattern analysis and machine intelligence 28(11), 1768-1783 (2006)", + "18. Grcic, M., Bevandić, P., Šegvic, S.: Densehybrid: Hybrid anomaly detection for dense open-set recognition. In: European Conference on Computer Vision. pp. 500-517. Springer (2022)", + "19. Grcic, M., Saric, J., Segvic, S.: On advantages of mask-level recognition for outlier-aware segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2936-2946 (2023)", + "20. Hendrycks, D., Basart, S., Mazeika, M., Zou, A., Kwon, J., Mostajabi, M., Steinhardt, J., Song, D.: Scaling out-of-distribution detection for real-world settings. arXiv preprint arXiv:1911.11132 (2019)", + "21. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. In: International Conference on Learning Representations (2017), https://openreview.net/forum?id=Hkg4TI9x1", + "22. Hendrycks, D., Mazeika, M., Dietterich, T.: Deep anomaly detection with outlier exposure. arXiv preprint arXiv:1812.04606 (2018)", + "23. Huang, Z., Li, S.: From difference to similarity: A manifold ranking-based hyperspectral anomaly detection framework. IEEE Transactions on Geoscience and Remote Sensing 57(10), 8118-8130 (2019)", + "24. Iscen, A., Tolias, G., Avrithis, Y., Chum, O.: Mining on manifolds: Metric learning without labels. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7642-7651 (2018). https://doi.org/10.1109/CVPR.2018.00797", + "25. Iscen, A., Tobias, G., Avrithis, Y., Furon, T., Chum, O.: Efficient diffusion on region manifolds: Recovering small objects with compact cnn representations. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2077-2086 (2017). https://doi.org/10.1109/CVPR.2017.105", + "26. Jung, S., Lee, J., Gwak, D., Choi, S., Choo, J.: Standardized max logits: A simple yet effective approach for identifying unexpected road obstacles in urban-scene segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15425-15434 (2021)", + "27. Lakshminarayanan, B., Pritzel, A., Blundell, C.: Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information processing systems 30 (2017)", + "28. Liang, C., Wang, W., Miao, J., Yang, Y.: Gmmseg: Gaussian mixture based generative semantic segmentation models. Advances in Neural Information Processing Systems 35, 31360-31375 (2022)", + "29. Liang, S., Li, Y., Srikant, R.: Enhancing the reliability of out-of-distribution image detection in neural networks. In: International Conference on Learning Representations (2018), https://openreview.net/forum?id=H1VGkIxRZ" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z. Zeng et al.", + "bbox": [ + 271, + 114, + 361, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "30. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014). https://doi.org/10.1007/978-3-319-10602-1_48", + "31. Lis, K., Nakka, K., Fua, P., Salzmann, M.: Detecting the unexpected via image resynthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2152-2161 (2019). https://doi.org/10.1109/ICCV.2019.00224", + "32. Liu, Y., Ding, C., Tian, Y., Pang, G., Belagiannis, V., Reid, I., Carneiro, G.: Residual pattern learning for pixel-wise out-of-distribution detection in semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1151-1161 (2023)", + "33. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)", + "34. Nayal, N., Yavuz, M., Henriques, J.F., Güney, F.: Rba: Segmenting unknown regions rejected by all. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 711-722 (2023)", + "35. Ohgushi, T., Horiguchi, K., Yamanaka, M.: Road obstacle detection method based on an autoencoder with semantic segmentation. In: Proceedings of the Asian Conference on Computer Vision (2020)", + "36. Rai, S.N., Cermelli, F., Fontanel, D., Masone, C., Caputo, B.: Unmasking anomalies in road-scene segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4037-4046 (2023)", + "37. Tian, Y., Liu, Y., Pang, G., Liu, F., Chen, Y., Carneiro, G.: Pixel-wise energy-biased abstention learning for anomaly segmentation on complex urban driving scenes. In: European Conference on Computer Vision. pp. 246-263. Springer (2022). https://doi.org/10.1007/978-3-031-19842-7_15", + "38. Vojir, T., Šipka, T., Aljundi, R., Chumerin, N., Reino, D.O., Matas, J.: Road anomaly detection by partial image reconstruction with segmentation coupling. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15651-15660 (2021)", + "39. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems 34, 12077-12090 (2021)", + "40. Yang, F., Hinami, R., Matsui, Y., Ly, S., Satoh, S.: Efficient image retrieval via decoupling diffusion into online and offline processing. In: Proceedings of the AAAI conference on artificial intelligence. vol. 33, pp. 9087-9094 (2019). https://doi.org/10.1609/aaai.v33i01.33019087", + "41. Yang, F., Wang, Z., Xiao, J., Satoh, S.: Mining on heterogeneous manifolds for zero-shot cross-modal image retrieval. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 12589-12596 (2020)", + "42. Zhou, D., Bousquet, O., Lal, T., Weston, J., Scholkopf, B.: Learning with local and global consistency. Advances in neural information processing systems 16 (2003)", + "43. Zhou, D., Weston, J., Gretton, A., Bousquet, O., Schölkopf, B.: Ranking on data manifolds. Advances in neural information processing systems 16 (2003)" + ], + "bbox": [ + 212, + 146, + 787, + 797 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "RWPM", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Random Walk on Pixel Manifolds for Anomaly Segmentation of Complex Driving Scenes/40e3d166-d87c-43d0-aeb7-8851ed372fb1_model.json b/2024/Random Walk on Pixel Manifolds for Anomaly Segmentation of Complex Driving Scenes/40e3d166-d87c-43d0-aeb7-8851ed372fb1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1214f16d6cc4e8855c42007746a1ada35bab2bfc --- /dev/null +++ b/2024/Random Walk on Pixel Manifolds for Anomaly Segmentation of Complex Driving Scenes/40e3d166-d87c-43d0-aeb7-8851ed372fb1_model.json @@ -0,0 +1,2093 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.234, + 0.141, + 0.773, + 0.187 + ], + "angle": 0, + "content": "Random Walk on Pixel Manifolds for Anomaly Segmentation of Complex Driving Scenes" + }, + { + "type": "text", + "bbox": [ + 0.366, + 0.212, + 0.634, + 0.228 + ], + "angle": 0, + "content": "Zelong Zeng1 and Kaname Tomite1" + }, + { + "type": "text", + "bbox": [ + 0.377, + 0.24, + 0.627, + 0.269 + ], + "angle": 0, + "content": "SenseTime Japan, Tokyo, Japan {zengzelong,tomite}@sensetime.jp" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.302, + 0.758, + 0.51 + ], + "angle": 0, + "content": "Abstract. In anomaly segmentation for complex driving scenes, state-of-the-art approaches utilize anomaly scoring functions to calculate anomaly scores. For these functions, accurately predicting the logits of inlier classes for each pixel is crucial for precisely inferring the anomaly score. However, in real-world driving scenarios, the diversity of scenes often results in distorted manifolds of pixel embeddings in the space. This effect is not conducive to directly using the pixel embeddings for the logit prediction during inference, a concern overlooked by existing methods. To address this problem, we propose a novel method called Random Walk on Pixel Manifolds (RWPM). RWPM utilizes random walks to reveal the intrinsic relationships among pixels to refine the pixel embeddings. The refined pixel embeddings alleviate the distortion of manifolds, improving the accuracy of anomaly scores. Our extensive experiments show that RWPM consistently improve the performance of the existing anomaly segmentation methods and achieve the best results1." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.523, + 0.744, + 0.552 + ], + "angle": 0, + "content": "Keywords: Autonomous vehicles \\(\\cdot\\) Anomaly segmentation \\(\\cdot\\) Semantic segmentation" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.576, + 0.377, + 0.592 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.606, + 0.788, + 0.756 + ], + "angle": 0, + "content": "Semantic segmentation [9-11,13,39] plays a crucial role in the field of automated driving, which can aid the automated driving system in perceiving and understanding the surrounding environment, thereby facilitating the system make accurate decision. Typically, the semantic segmentation models are trained on closed-datasets with fixed semantic categories, showing good recognition capabilities for inlier classes (e.g., roads, vehicles, etc.). However, these models often fail to effectively detect potential outlier class objects (e.g., animals, fallen tires, etc.) in real-world scenarios. Due to the critical importance of correctly detecting outlier class targets for the safety of autonomous driving, there is a growing focus on addressing this task, namely, anomaly segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.758, + 0.788, + 0.817 + ], + "angle": 0, + "content": "The standard paradigm for anomaly segmentation is assigning an anomaly score to each pixel of an input image, where a higher score indicates a greater probability that the pixel belongs to an outlier. Thus, the way used for estimating the anomaly score is crucial for anomaly segmentation methods. In" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.641, + 0.841 + ], + "angle": 0, + "content": "1 Code is available at: https://github.com/ZelongZeng/RwPM" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "Z. Zeng et al." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.143, + 0.391, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.234, + 0.377, + 0.242 + ], + "angle": 0, + "content": "(i) An ideal manifold structure" + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.143, + 0.588, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.234, + 0.566, + 0.243 + ], + "angle": 0, + "content": "(ii) A distorted manifold structure" + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.143, + 0.756, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.234, + 0.751, + 0.242 + ], + "angle": 0, + "content": "(iii) A refined manifold structure" + }, + { + "type": "image_footnote", + "bbox": [ + 0.354, + 0.248, + 0.418, + 0.257 + ], + "angle": 0, + "content": "Inlier pixel" + }, + { + "type": "image_footnote", + "bbox": [ + 0.439, + 0.247, + 0.551, + 0.258 + ], + "angle": 0, + "content": "\\(\\times \\times\\) Inlier class prototype" + }, + { + "type": "image_footnote", + "bbox": [ + 0.565, + 0.248, + 0.63, + 0.257 + ], + "angle": 0, + "content": "Outlier pixel" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.272, + 0.789, + 0.51 + ], + "angle": 0, + "content": "Fig. 1: A toy example about manifolds of pixel embeddings. The dashed circles indicate regions with high predicted logit to the corresponding inlier class. For the pixel in these regions, its anomaly score is small. The prototype is the vector of network weights of the classifier for the corresponding inlier class. (i) An ideal manifold structure. All inlier pixels are within the corresponding high logit region, whereas all outlier pixels are outside the areas. The anomaly score derived from the anomaly scoring function effectively discriminates between them. (ii) The manifolds are affected by the diversity of data, causing some inlier pixels to deviate from the corresponding high logit region, while some outlier pixels approach the regions. This results in false positives/negatives in the anomaly scores. However, pixels of the same class are still in the same manifold, indicating that the manifold structure can be utilized to reveal the intrinsic relationships between pixels. (iii) Our RWPM utilizes random walks to capture the manifold structure to diffuse and update pixel embeddings. The embeddings within the same manifold tend to become more similar after the updating. Fig 3 and Fig 4 respectively present the visualization of embedding distributions and the qualitative results of real examples, consistent with the description of this toy example." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.789, + 0.78 + ], + "angle": 0, + "content": "early approaches, the anomaly score is often given by estimating the prediction uncertainty [16, 21, 27, 29], or by comparing the discrepancy between the original image and its reconstructed image generated based on the semantic segmentation predictions [14, 15, 31, 35]. Different from early methods, recent methods [12, 19, 32, 34, 36, 37] adopt Outlier Exposure (OE) strategy [22], which directly learn a pixel-level anomaly class to detect anomalous pixels that deviate from inlier classes. Specifically, these methods propose anomaly scoring functions to infer the anomaly scores and, based on the functions, leverage auxiliary outlier data to fine-tune a semantic segmentation model. The auxiliary outlier data is typically generated by extracting outlier-class objects from other datasets (e.g., COCO [30]) and incorporating them into driving scene images [34, 36, 37]. Furthermore, some of them [19, 34, 36] adopt mask-based segmentation networks instead of pixel-based segmentation networks, achieving state-of-the-art (SOTA) performance in anomaly segmentation. In this paper, we concentrate on the current methods using OE strategy. These methods indicate superior performance to the early methods using prediction uncertainty or image reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.842 + ], + "angle": 0, + "content": "The anomaly scoring functions in OE methods infer anomaly scores based on the inlier class logit predictions from a semantic segmentation model. Specifically, the logic of these functions can be represented as follows: when the predicted logits of a pixel embedding are small across all inlier classes, indicating that the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.68, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RWPM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.39 + ], + "angle": 0, + "content": "pixel is dissimilar to all inlier classes and is an outlier with high probability, the function yields a higher anomaly score. Therefore, accurately measuring the logits of each pixel is crucial for precisely evaluating the anomaly score. However, in practical open-world driving scenarios, various environmental conditions (such as lighting, road materials, etc.) and a diverse range of unknown class objects are present. During the inference phase, the diversity of driving scenes often results in distorted manifolds of pixel embedding in the embedding space, causing pixel embeddings to deviate from the ideal distribution. This phenomenon has been validated in previous works of metric learning [3,25,40,41]. Since the logits are calculated from the inner product (i.e., cosine distance) of the pixel embeddings and the vectors of classifiers' network weights, the distorted manifolds are not conducive to directly using the pixel embeddings for the logit prediction. However, current anomaly segmentation methods ignore this issue and directly use the pixel embeddings for the logit prediction in the inference phase, so the results often fail to accurately reflect the anomaly score for the corresponding pixels, causing high false positive/negative rate (see Fig. 1(i) and Fig. 1(ii))." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.392, + 0.788, + 0.786 + ], + "angle": 0, + "content": "In this paper, we propose a novel method, called Random Walk on Pixel Manifolds (RWPM), to alleviate the effect of data diversity on the manifolds. RWPM utilizes random walks to capture the manifold structure of the pixel embeddings to diffuse and update the embedding of each pixel. Since the structure of data manifolds implicitly contains the intrinsic relationship among data points (see Fig. 1(ii)) [3,24,40,43], using random walks to measure the similarity between pixels on manifolds (i.e., distance on manifolds rather than Euclidean distance) better reflects the intrinsic relationships between pixels. Consequently, diffusing and updating pixel embeddings based on the similarity on manifolds results in high similarity among pixels on the same manifold and low similarity among pixels on different manifolds. In the embedding space, this manifests as pixel embeddings on the same manifold forming more compact clusters, thereby mitigating the distortion effect of data diversity on the manifolds, which facilitates the logit prediction (see Fig. 1(iii)). Furthermore, since the typically large number of pixels in an image, utilizing all pixels for the random walks results in significant memory and computational overhead, which damages the deployment and real-time execution of our RWPM. To address this issue, we introduce the Partial Random Walk strategy for RWPM, reducing memory consumption and enhancing the operational efficiency of RWPM. Note that, RWPM is proposed to use in the inference phase, thus no any additional training is required. Additionally, RWPM can be directly integrated into existing anomaly segmentation framework without the need for any extra modification to the network structure. Extensive experiments in various anomaly segmentation benchmarks of road scenes, namely Fishyscapes [5], Road Anomaly [31], and SMIYC [6], demonstrate that RWPM consistently improves the performance of existing methods, achieving SOTA results. To summarise, our contributions are the following:" + }, + { + "type": "page_footnote", + "bbox": [ + 0.226, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "- We introduce that the diversity of driving scenes results in distorted manifolds of pixel embedding, thereby affecting the accuracy of anomaly scoring" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "Z. Zeng et al." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "functions in inferring anomaly scores. This problem is ignored by existing anomaly segmentation methods." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.177, + 0.784, + 0.234 + ], + "angle": 0, + "content": "- We propose the Random Walk on Pixel Manifolds (RWPM) that utilizes random walks to reveal the intrinsic relationships among pixels to refine the pixel embeddings. This way alleviate the effect on manifolds caused by the driving scene diversity." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.236, + 0.784, + 0.264 + ], + "angle": 0, + "content": "- We propose the Partial Random Walk strategy for RWPM to reduce the memory consumption and improve the operational efficiency." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.284, + 0.388, + 0.301 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.788, + 0.844 + ], + "angle": 0, + "content": "Anomaly Segmentation. Anomaly segmentation aims to detect the outlier class object in some specific scenes, such as complex driving scenes [6, 13, 31]. Existing anomaly segmentation can be broadly divided into three categories: (a) Uncertainty-based methods, (b) Reconstruction-based methods and (c) Outlier Exposure methods. Uncertainty-based methods assume that outlier samples result in low-confidence predictions. Based on this assumption, they focus on measuring the pixel-wise anomaly score by estimating the prediction uncertainty through various ways, such as ensembles [27], Bayesian estimation [16], maximum softmax probability [21, 29], and logit [20, 26]. Since segmentation models are trained on close-set, they may misclassify unseen classes with high confidence, leading uncertainty-based methods fail to detect anomalies. Reconstruction-based methods detect anomaly pixel by comparing the discrepancy between the original image and its reconstructed image generated based on the semantic segmentation predictions [14, 15, 31, 35]. These methods are challenged by the accuracy of segmentation prediction and the difficulties of training. Recent methods [7, 12, 19, 34, 36, 37] introduce the Outlier Exposure (OE) strategy [22] that utilizes auxiliary outlier data and anomaly scoring functions to learn models. The auxiliary outlier data is the image from other datasets (e.g., COCO [30]), or cutting the outlier objects from other datasets and pasting them into the inlier scenes, and the anomaly scoring functions are used to calculate the anomaly score for each pixel. Compared with previous works, OE strategy can reach better performance. More recently, some OE methods [19, 34, 36] adopt mask-based segmentation networks [10, 11], instead of pixel-based segmentation networks [8, 9], as their backbones. Based on the powerful capability of mask-based networks in segmentation tasks, these methods significantly reduces false detection rates, achieving SOTA results. In this paper, we only concentrate on the OE methods. The anomaly scoring functions employed in OE methods are based on the predicted logits of inlier classes. However, during the inference phase, the diversity of driving scenes results in distorted manifolds of pixel embeddings. As a result, the scoring functions may fail to use the embeddings to accurately calculate logits, thereby affects the accuracy of measuring anomaly scores. Current OE methods ignore this problem. Our proposed RWPM addresses this problem by utilizing random walks on pixel manifolds to reveal the intrinsic relationships among pixels and refine the pixel embeddings. To the best of our knowledge, RWPM is the first work to propose and resolve this problem." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.68, + 0.114, + 0.732, + 0.128 + ], + "angle": 0, + "content": "RWPM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.511 + ], + "angle": 0, + "content": "Mining on Data Manifolds. Mining on data manifolds, also known as diffusion process, is a well-established technique commonly used in retrieval tasks [3, 4, 25, 40, 43]. These works utilize random walks on data manifolds to capture the data manifold structure for revealing the intrinsic relationship between data points. The data manifold is constructed by a graph, where each node represents a sample and each edge presents the connection between a node and its neighborhoods, weighted proportionally according to their distance. For retrieval tasks, the images of the same object may be shown in different conditions (e.g., lighting and angles). This diversity often leads to the manifold structure is not conducive to accurately ranking using distance metrics directly [25, 40]. Thus, diffusion process is used to diffuse the state (typically initialized to 1) of a query image on the data manifolds. If a sample is on the same data manifold with the query (the distance on manifolds is close), then more energy of the state will flow to that sample. Therefore, the final state of all samples can be seen as the accurate ranking scores. Moreover, diffusion process is also adopted in semi-supervised learning [24, 42] to propagate the label of labeled samples to unlabeled samples. Similarly, some unsupervised segmentation works [17, 23] utilize diffusion process to propagate the labels of manually annotated pixels to the unlabeled pixels. These existing works utilize random walks on data manifolds for ranking or label propagation. Different from these works, our target is to solve the problem occurring in pixel manifolds for anomaly scoring functions. We propose innovatively using random walks to diffuse and update pixel embeddings, eliminating the effect of distorted manifolds directly at the embedding-level, thereby improving the accuracy of subsequent processing." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.531, + 0.427, + 0.549 + ], + "angle": 0, + "content": "3 Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.56, + 0.373, + 0.575 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.789, + 0.78 + ], + "angle": 0, + "content": "We denote an image space \\(\\mathcal{X} \\subset \\mathbb{R}^{3 \\times H \\times W}\\), where \\(H\\) and \\(W\\) are the height and width, respectively. During the inference phase, the goal for an anomaly segmentation model \\(f\\) is to map an input image \\(\\pmb{x} \\in \\mathcal{X}\\) to an anomaly space, i.e., \\(f: \\pmb{x} \\mapsto \\mathcal{Z} \\subset \\mathbb{R}^{H \\times W}\\). Specifically, an anomaly segmentation model typically consists of the backbone encoder-decoder \\(F\\) and the classifier \\(C\\). For an input image \\(\\pmb{x}\\), firstly, its pixel embedding map is encoded as \\(\\pmb{p} = F(\\pmb{x}) \\in \\mathbb{R}^{d \\times H \\times W}\\), where \\(d\\) is the dimension of the pixel embedding. Then, its logit map \\(\\pmb{l}\\) is predicted by \\(\\pmb{l} = C(F(\\pmb{x})) \\in \\mathbb{R}^{K \\times H \\times W}\\), where \\(K\\) is the number of inlier classes. Note that, depending on the backbone network, the size of the embedding map and logit map may sometimes be smaller than the original image. For the sake of convenience in this paper, we do not make a special distinction. Finally, a scoring function is used to calculate the anomaly score map \\(\\pmb{z} \\in \\mathcal{Z}\\) based on the logit map \\(\\pmb{l}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.842 + ], + "angle": 0, + "content": "For anomaly scoring functions, their logic can be represented as follows: when the predicted logits of a pixel embedding are small across all inlier classes, indicating that the pixel is dissimilar to all inlier classes and is likely an outlier, the function generates a higher anomaly score. For example, let's define \\( l_{h,w}(k) \\)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "Z. Zeng et al." + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.146, + 0.776, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.261, + 0.788, + 0.388 + ], + "angle": 0, + "content": "Fig. 2: Overview. This figure illustrates the application of RWPM to an existing anomaly segmentation framework during the inference phase. The red dashed box highlights the RWPM part. First, the encoder-decoder is used to extract the pixel embedding map of the input image. The pixel embedding map are subsequently partitioned into \\( n^2 \\) sub-maps. For each sub-map, we update its pixel embeddings by using random walks to obtain a refined sub-map. Finally, the refined sub-maps are concatenated to form the refined embedding map, which are then input into the subsequent network structure. Notably, RWPM can be directly integrated into existing frameworks without requiring extra training or changes to the network structure." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.788, + 0.495 + ], + "angle": 0, + "content": "is the logit of a pixel \\( \\pmb{x}_{h,w} \\) for class \\( k \\). The scoring function in PEBAL [37] is \\( E_{h,w} = -\\log \\sum_{k \\in \\{1, \\dots, K\\}} \\exp(\\pmb{l}_{h,w}(k)) \\). Obviously, when the anomaly score \\( E_{h,w} \\) is large, it indicates that the logits of the pixel across all inlier classes are small, i.e., the pixel \\( \\pmb{x}_{h,w} \\) is dissimilar from all inlier classes (see Section S.2 of the supplementary material for the analysis of other anomaly scoring functions)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.516, + 0.43, + 0.532 + ], + "angle": 0, + "content": "3.2 Graph Construction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.542, + 0.788, + 0.664 + ], + "angle": 0, + "content": "We begin by constructing a graph to represent the manifolds of pixel embeddings. In our approach, we utilize an affinity matrix based on cosine similarity to characterize the relationships between pixels. Specifically, we first use the encoder-decoder component of an anomaly segmentation model to extract the pixel embeddings \\(\\pmb{p} \\in \\mathbb{R}^{d \\times H \\times W}\\). For ease of description and computation, we reshape the \\(\\pmb{p}\\) as \\(\\pmb{p}^r \\in \\mathbb{R}^{HW \\times d}\\), where each row represents a pixel's embedding. We then normalize each row of \\(\\pmb{p}^r\\) to obtain \\(\\widehat{\\pmb{p}}^r\\), i.e., \\(\\widehat{\\pmb{p}}_i^r = \\pmb{p}_i^r / \\| \\pmb{p}_i^r \\|\\). The affinity matrix is defined as \\(\\mathbf{W} \\in \\mathbb{R}^{HW \\times HW}\\), where each component is obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.674, + 0.787, + 0.711 + ], + "angle": 0, + "content": "\\[\n\\mathbf {W} _ {i j} = \\left\\{ \\begin{array}{c c} \\langle \\hat {\\pmb {p}} _ {i} ^ {r}, \\hat {\\pmb {p}} _ {j} ^ {r} \\rangle & i \\neq j \\\\ 0 & i = j \\end{array} , \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.787, + 0.749 + ], + "angle": 0, + "content": "where \\(\\langle \\cdot, \\cdot \\rangle\\) denotes inner product, \\(\\mathbf{W}_{ii} = 0\\) is guaranteed to circumvent self-loops in the graph represented by \\(\\mathbf{W}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To mitigate the influence of noise, locally constrained random walk methods [24, 40, 41] adopt local constraints to construct the graph, i.e., each point is only connected to its locally nearest neighbors in the graph. Determining the nearest neighbors for each point traditionally involves applying nearest neighbor search algorithms, such as the \\(k\\)-NN algorithm. However, in anomaly segmentation tasks, the size of the graph is large and the graph is constructed online," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.68, + 0.114, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RWPM" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "making the search algorithms impractical. To benefit from the local constraints, we utilize a softmax function to normalize each row of \\(\\mathbf{W}\\) to obtain locally constrained graph \\(\\mathbf{S}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.383, + 0.204, + 0.786, + 0.243 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} _ {i j} = \\left\\{ \\begin{array}{c c} \\frac {\\exp \\left(\\mathbf {W} _ {i j} / \\tau\\right)}{\\sum_ {i = 1 , i \\neq j} ^ {H W} \\exp \\left(\\mathbf {W} _ {i j} / \\tau\\right)} & i \\neq j \\\\ 0 & i = j \\end{array} , \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.254, + 0.785, + 0.329 + ], + "angle": 0, + "content": "where \\(\\tau\\) is a temperature hyper-parameter, we set its value less than 1.0 to remove the effect of non-neighbor points. Compared to typical search algorithms, Eq. 2 can impose local constraints more efficiently and effectively (see Section S.4 of the supplementary material), and keep the scale of embedding in subsequent random walks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.352, + 0.45, + 0.367 + ], + "angle": 0, + "content": "3.3 Random Walk Process" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.378, + 0.784, + 0.437 + ], + "angle": 0, + "content": "Now we have constructed the graph \\(\\mathbf{S}\\), we can start the random walks. Different from previous mining on data manifolds works utilizing random walks to propagate the states of known samples to other unknown samples, we use a random walk to propagate the embedding between pixels on manifolds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.785, + 0.482 + ], + "angle": 0, + "content": "We first define \\(\\mathbf{m}^t\\in \\mathbb{R}^{HW\\times d}\\) as the updated pixel embeddings in the \\(t\\) -th step of the random walks. \\(\\mathbf{m}^0\\) is initialized as \\(\\pmb{p}^r\\). Then, each iteration step of the random walks is described as:" + }, + { + "type": "equation", + "bbox": [ + 0.363, + 0.496, + 0.785, + 0.512 + ], + "angle": 0, + "content": "\\[\n\\mathbf {m} ^ {t + 1} = \\alpha \\mathbf {S} \\mathbf {m} ^ {t} + (1 - \\alpha) \\mathbf {m} ^ {0}, \\alpha \\in (0, 1), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.785, + 0.644 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is the probability continuing the random walk from the current state \\(\\mathbf{m}^t\\), while \\((1 - \\alpha)\\) represents the probability of restarting from the initial state \\(\\mathbf{m}^0\\). At each iteration, guided by the manifold structure, each pixel diffuses its embedding to other pixels while also receiving embedding from other pixels for updating its own embedding. The updated pixel embedding can be seen as a weighted ensemble of other pixel embeddings. If pixels are similar on manifolds, their embeddings will gradually become similar. Mathematically speaking, the whole iteration process can be written as a closed-from as follows [43]:" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.658, + 0.785, + 0.674 + ], + "angle": 0, + "content": "\\[\n\\mathbf {m} ^ {\\infty} = (1 - \\alpha) (\\mathbf {I} - \\alpha \\mathbf {S}) ^ {- 1} \\mathbf {m} ^ {0}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.685, + 0.784, + 0.717 + ], + "angle": 0, + "content": "where \\(\\mathbf{I} \\in \\mathbb{R}^{HW \\times HW}\\) is an identity matrix. Finally, \\(\\mathbf{m}^{\\infty}\\) is used as the refined pixel embeddings for the subsequent anomaly segmentation process." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.739, + 0.444, + 0.754 + ], + "angle": 0, + "content": "3.4 Partial Random Walk" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.785, + 0.84 + ], + "angle": 0, + "content": "When the size of the image is large, the size of the relevant matrices used for the calculation of random walks becomes extremely huge as well. For instance, when the image size is \\((512 \\times 1024)\\), the size of matrix \\(S\\) reaches \\((524288 \\times 524288)\\). This significantly increases the memory consumption and computational time of RWPM, which is not conducive to online operation. Unfortunately, in the field" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "Z. Zeng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "of autonomous driving, the image size of driving scenes is generally large. To address this problem, we propose Partial Random Walk, which consists of the following two aspects: embedding map partitioning and limited iteration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.352 + ], + "angle": 0, + "content": "To improve efficiency, we first propose embedding map partitioning strategy to reduce the size of the pixel embedding map used for random walks. Specifically, for a pixel embedding map \\(\\pmb{p} \\in \\mathbb{R}^{d \\times H \\times W}\\), we equally divide it into \\(n \\times n\\) sub-maps \\(\\left\\{\\pmb{p}^{(1)}, \\dots, \\pmb{p}^{(n^2)}\\right\\} \\in \\mathbb{R}^{d \\times \\frac{H}{n} \\times \\frac{W}{n}}\\). Then, for each sub-map in \\(\\left\\{\\pmb{p}^{(1)}, \\dots, \\pmb{p}^{(n^2)}\\right\\}\\), we follow the steps of Eq. 1 and Eq. 2 to construct their sub-graphs \\(\\left\\{\\pmb{S}_{\\pmb{p}^{(1)}}, \\dots, \\pmb{S}_{\\pmb{p}^{(n^2)}}\\right\\} \\in \\mathbb{R}^{d \\times \\frac{HW}{n^2} \\times \\frac{HW}{n^2}}\\). We also attempted other methods to reduce the number of pixel embeddings, such as resize or superpixel [1]. Experimental results indicate that embedding map partitioning is the optimal solution among them (see Section 4.2)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.353, + 0.787, + 0.614 + ], + "angle": 0, + "content": "Next, for each sub-map \\( \\pmb{p}^{(i)} \\), we perform random walks with the limited iteration strategy in their corresponding sub-graphs \\( S_{\\pmb{p}^{(i)}} \\). That is, we use Eq. 3 for a small number of iterations instead of using the closed-form Eq. 4. We found that a few iterations by using Eq. 3 are enough to yield high quality results. The time complexity of Eq. 3 with \\( T \\) iterations for each sub-map is \\( O(Td(\\frac{HW}{n^2})^2) \\), while the time complexity of Eq. 4 is \\( O((\\frac{HW}{n^2})^3) + O(d(\\frac{HW}{n^2})^2) \\). Since \\( Td \\) is much smaller than \\( \\frac{HW}{n^2} \\), the limited iteration strategy can significantly improve the efficiency of RWPM. After the random walks, we obtain the refined pixel embeddings sub-map \\( \\mathbf{m}_{\\pmb{p}^{(i)}}^T \\) for each sub-map \\( \\pmb{p}^{(i)} \\). Finally, we concatenate the refined sub-maps based on their positions to obtain the final refined pixel embedding map. Additionally, when \\( n \\) is large (observed experimentally as \\( n > 2 \\)), due to spatial constraints, the baseline values of anomaly scores outputted on each sub-map may vary. Hence, a calibration is performed. For adjacent refined sub-maps \\( \\pmb{p}^{(i)'} \\) and \\( \\pmb{p}^{(j)'} \\), we compute the average anomaly scores \\( I \\) and \\( J \\) of all pixels on their adjacent edges, respectively. Subsequently, using \\( \\pmb{p}^{(i)'} \\) as the reference, we calculate the ratio \\( \\frac{I}{J} \\), and then multiply all anomaly scores on sub-map \\( \\pmb{p}^{(j)'} \\) by this ratio to achieve the calibration." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.616, + 0.608, + 0.631 + ], + "angle": 0, + "content": "Fig. 2 shows the workflow of the proposed RWPM." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.656, + 0.367, + 0.673 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Dataset: We evaluate our approach on three standard benchmark datasets for anomaly segmentation, i.e., Fishyscapes Lost&Found [5], Road Anomaly [31] and Segment Me If You Can (SMIYC) [6]. Fishyscapes Lost&Found has 100 validation images. The domain of this dataset is similar to that of Cityscapes [13] dataset. Road Anomaly dataset contains 60 images of real-world road scenes. Unlike the Fishyscapes Lost&Found dataset, the images in Road Anomaly contain various classes and sizes of anomaly objects, making it more challenging. SMIYC consists of two subsets, Anomaly track and Obstacle track. The Anomaly track focuses on detecting large anomalous objects, while the Obstacle track emphasizes detecting small anomalous objects on the road. Anomaly track contains 10" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.68, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RWPM" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "validation images and 100 online test images, while Obstacle track contains 30 validation images and 327 online test images. The Road Anomaly and SMIYC datasets both exhibit high domain shifts compared to the Cityscapes dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.787, + 0.312 + ], + "angle": 0, + "content": "Evaluation Metrics: Following the previous works, we evaluate our method on the standard metrics for anomaly segmentation. On the Fishyscapes Lost&Found and Road Anomaly, we adopt the area under receiver operating characteristics (AUROC), average precision (AP), and the false positive rate at a true positive rate of \\(95\\%\\) (FPR95). On the SMIYC, in addition to reporting per-pixel metrics such as AP and FPR95, we also calculate results for component-level metrics [6] such as the component-wise intersection over union (sIoU), positive predictive value (PPV), and mean F1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.603 + ], + "angle": 0, + "content": "Implementation Details: To demonstrate the effectiveness of RWPM for anomaly segmentation methods, we selected representative methods from both pixel-based and mask-based networks. For pixel-based methods, we chose PE-BAL [37] and Balanced Energy [12]. These methods all utilize DeepLabv3+ [9] with WideResnet38 as the backbone and were trained on the Cityscapes dataset and COCO OE images [37]. For mask-based methods, we selected RbA [34] and Mask2Anomaly [36]. These methods use Mask2Former [10] with Swin-L [33] as the backbone and were also trained on the Cityscapes and COCO OE images. All experiments are conducted on a single NVIDIA RTX A6000 GPU. To ensure efficient implementation on a single GPU for all experiments, we adopt embedding map partitioning and limited iteration in all experiments. Specifically, for pixel-based methods, we set the partitioning parameter \\( n = 4 \\). For mask-based methods, as their embedding map size is smaller, we set \\( n = 2 \\). Additionally, for the random walk, if not specifically emphasized, we set the number of iterations \\( T = 20 / T = 5 \\) for the Road Anomaly/the other datasets and the transition probability \\( \\alpha = 0.99 \\) for all datasets. The \\( \\tau \\) is set as 0.01. It is important to note that our method is used in the inference phase. In other words, RWPM is directly applied to the pre-trained models of other existing anomaly segmentation methods, thus requiring no additional training." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.624, + 0.458, + 0.639 + ], + "angle": 0, + "content": "4.1 Effectiveness of RWPM" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.65, + 0.788, + 0.817 + ], + "angle": 0, + "content": "First, to validate that RWPM can improve the performance of existing anomaly segmentation methods, we integrate our approach with four recent representative methods, namely PEBAL [37], Balanced Energy [12], Mask2Anomaly [36] and RbA [34]. PEBAL and Balanced Energy employ pixel-based networks as their backbones, whereas Mask2Anomaly and RbA utilize mask-based networks as their backbones. We evaluate the results on both the Fishyscapes Lost&Found validation set and Road Anomaly test set. The experimental results are reported in Table 1. These results demonstrate that our proposed RWPM consistently and significantly enhances the performance of existing anomaly segmentation methods, across different backbone architectures. Additional results and parameters details are shown in Section S.3 of the supplementary material." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.672, + 0.84 + ], + "angle": 0, + "content": "2 We carefully reproduce the experiments by using the official code." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "Z. Zeng et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.214 + ], + "angle": 0, + "content": "Table 1: Comparison with strong and representative anomaly segmentation baselines across different backbone Architectures (Arch). Bold denotes the better results between using RWPM and not using RWPM with the same anomaly segmentation baselines. \\(\\dagger\\) indicates that use the calibration in sub-map concatenation as described in Section 3.4." + }, + { + "type": "table", + "bbox": [ + 0.274, + 0.227, + 0.73, + 0.33 + ], + "angle": 0, + "content": "
Benchmark→ Fishscapes Lost&Found Road Anomaly
Method ↓ArchAuROC↑AP↑FPR95↓AuROC↑AP↑FPR95↓
PEBAL [37]Pixel-based98.9658.814.7787.6345.1044.58
PEBAL + RWPM† (Ours)Pixel-base99.2066.853.6889.4850.2936.81
Balanced Energy [12]Pixel-based99.0367.072.9388.3141.4841.46
Balanced Energy + RWPM† (Ours)Pixel-base99.2972.952.3890.5148.2632.10
Mask2Anomaly2 [36]Mask-based95.4765.277.7996.5480.0413.95
Mask2Anomaly + RWPM (Ours)Mask-base95.4865.387.3397.4480.097.45
RbA [34]Mask-based98.6270.816.3097.9985.426.92
RbA + RWPM (Ours)Mask-base98.8271.166.1298.0487.345.27
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.348, + 0.788, + 0.391 + ], + "angle": 0, + "content": "Table 2: Comparison with state-of-the-art anomaly segmentation methods across different backbone Architectures (Arch). Bold and underline denote the best and the second best results, respectively." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.402, + 0.789, + 0.561 + ], + "angle": 0, + "content": "
Benchmark→Fishscapes L&FRoad AnomalyAnomaly TrackObstacle TrackAverage
Method ↓ArchAP↑FPR95↓AP↑FPR95↓AP↑FPR95↓AP↑FPR95↓AP↑FPR95↓
Synboost [15] (CVPR'21)Pixel-based60.5831.0241.8459.7256.4461.8671.343.1557.4638.94
SML [26] (ICCV'21)Pixel-based22.7433.4925.8249.4746.839.53.436.824.6939.82
Meta-OOD [7] (ICCV'21)Pixel-based41.3137.6948.8431.7785.4715.0085.070.7565.1321.30
Learning Embedding [5](IJCV'21)Pixel-based4.6524.36--37.5270.760.8246.38--
Void Classifier [5](IJCV'21)Pixel-based10.2922.11--36.6163.4910.4441.54--
GMMSeg-DL [28] (NeurIPS'22)Pixel-based43.4713.1134.4247.90------
DenseHybrid [18] (ECCV'22)Pixel-based69.795.0931.3963.9777.969.8187.080.2465.5619.78
PEBAL [37] (ECCV'22)Pixel-based58.814.7745.1044.5849.1440.824.9812.6839.5825.71
Balanced Energy [12] (CVPR'23)Pixel-based67.072.9341.4841.46------
RPL-CoroCL [32] (ICCV'23)Pixel-based70.612.5271.6117.7483.4911.6885.930.5877.918.13
Mask2Anomaly [36] (ICCV'23)Mask-based69.449.2280.0413.9588.6214.5793.100.2082.879.49
M2F-EAM [19] (CVPRW'23)Mask-based52.0320.5166.6713.4276.393.966.917.965.4836.43
RbA [34] (ICCV'23)Mask-based70.816.3085.426.9290.8611.5991.850.4684.736.33
RbA + RWPM (Ours)Mask-based71.166.1287.345.2792.0010.1593.300.2886.005.46
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.788, + 0.763 + ], + "angle": 0, + "content": "Second, to further demonstrate the effectiveness of our RWPM, we compare it with other state-of-the-art (SOTA) methods. Specifically, we adopt the combination of RbA with RWPM as our instance approach for the comparison. In addition to the Fishyscapes and Road Anomaly datasets, we also evaluate our approach on a more challenging dataset, i.e., SMIYC online test set (Anomaly track and Obstacle track). SMIYC is characterized by its high domain shift and diversity of objects. Table 2 presents the results. The results indicate that our method can consistently achieve high performance across various datasets, while the previous SOTA methods fail to do so. We also compare all approaches by computing the average results of AP and FPR95. The results show that our approach outperforms the previous methods and achieve the best results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Furthermore, in addition to pixel-level metrics, we also utilize component-level metrics [6] to evaluate performance on the SMIYC online test set. Compared to pixel-level metrics, component-level metrics provide a better evaluation of the detection results for all anomalous objects in the scene. We report the results in Table 3. The experimental results show that RWPM significantly improves" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.68, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RWPM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.209 + ], + "angle": 0, + "content": "the performance of RbA in component-level metrics, surpassing other existing SOTA methods by a large margin. This indicates that RWPM can assist models in better detecting anomalous objects and achieving more precise localization of these anomalies [6]." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.234, + 0.787, + 0.263 + ], + "angle": 0, + "content": "Table 3: Component level evaluation: Comparison with state-of-the-art anomaly segmentation methods on the SMIYC online test set. Bold denote the best results." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.274, + 0.789, + 0.432 + ], + "angle": 0, + "content": "
Benchmark→Anomaly TrackObstacle Track
Method ↓ArchsIoU ↑PPV↑mean F1↑sIoU ↑PPV↑mean F1↑
Synboost [15] (CVPR'21)Pixel-based34.6817.819.9944.2841.7537.57
SML [26] (ICCV'21)Pixel-based26.0024.7012.205.1013.303.00
Meta-OOD [7] (ICCV'21)Pixel-based49.3139.5128.7247.8763.6448.51
Void Classifier [5](IJCV'21)Pixel-based21.1422.136.496.3420.275.41
Learning Embedding [5](IJCV'21)Pixel-based33.8620.547.9035.642.872.31
JSRNet [38] (ICCV'21)Pixel-based20.2029.2713.6618.5524.4611.02
DenseHybrid [18] (ECCV'22)Pixel-based54.1724.1331.0845.7450.1050.72
PEBAL [37] (ECCV'22)Pixel-based38.8827.2014.4829.917.555.54
RPL-CoroCL [32] (ICCV'23)Pixel-based49.7629.9630.1652.6156.6556.69
Mask2Fromer [36] (ICCV'23)Mask-based25.2018.2015.305.0021.904.80
Maskomaly [2] (BMVC'23)Mask-based55.451.650.0---
RbA [34] (ICCV'23)Mask-based55.6952.1446.8058.3658.7860.85
RbA + RWPM (Ours)Mask-based57.0061.2558.4458.8972.5169.85
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.483, + 0.389, + 0.499 + ], + "angle": 0, + "content": "4.2 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.787, + 0.537 + ], + "angle": 0, + "content": "If not specifically emphasized, all the results reported in this section are conducted on a RTX A6000 GPU." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.689 + ], + "angle": 0, + "content": "The effect of the partitioning parameter \\( n \\) and calibration: We first test the effect of the embedding map partitioning parameter \\( n \\) and the calibration, and we report the results in Table 4. FPS denotes the frames-per-second results. Row \\( n = 1 \\) displays the RWPM results without utilizing the embedding map partitioning strategy, indicating significant computational and memory overhead. However, the partitioning strategy (\\( n = 2, 4, 8 \\)) significantly improves the running speed and reduces memory consumption while maintaining effectiveness. Additionally, as \\( n \\) increases, the running efficiency also improves, albeit with a slight decrease in performance. However, when \\( n > 2 \\), the calibration can enhance the performance, particularly for the FPR95 metrics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.764 + ], + "angle": 0, + "content": "The effect of iteration number \\( T \\): Table 5 shows that, compared to using Eq. 4 (\\( T = \\infty \\)), utilizing the limited iteration strategy (\\( T = 5 \\sim 100 \\)) can lead to higher efficiency, resulting in even better performance for RWPM. Additionally, as the number of iterations increases, performance tends to improve, but running efficiency (FPS) decreases accordingly." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Different strategies for reducing the number of pixel embedding: We also attempt other strategies to reduce the number of pixel embedding for RWPM, such as resizing the embedding map or use SLIC [1] algorithm to extract superpixels. Table 6(Left) presents the results. The resize strategy achieves good performance but requires preserving a larger map size. Therefore, it fails" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "Z. Zeng et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 4: The effect of the partitioning parameter \\( n \\) and calibration. We set \\( T = {20} \\) for all experiments. \\( {}^{ \\dagger } \\) indicates that use the calibration in sub-map concatenation." + }, + { + "type": "table", + "bbox": [ + 0.33, + 0.185, + 0.672, + 0.296 + ], + "angle": 0, + "content": "
Benchmark→Road Anomaly
AP↑FPR95↓FPS↑GPU Memory Usage↓
RbA w/o RWPM85.426.9211.123.49GiB
n=1 (two GPUs)87.905.170.3274.96GiB
n=287.345.272.047.24GiB
n=2†87.185.252.047.24GiB
n=487.056.314.353.49GiB
n=4†87.175.324.353.49GiB
n=886.676.426.263.55GiB
n=8†86.915.416.263.55GiB
" + }, + { + "type": "table_caption", + "bbox": [ + 0.244, + 0.31, + 0.757, + 0.326 + ], + "angle": 0, + "content": "Table 5: The effect of iteration number \\( T \\) . We set \\( n = 2 \\) for all experiments." + }, + { + "type": "table", + "bbox": [ + 0.33, + 0.337, + 0.672, + 0.433 + ], + "angle": 0, + "content": "
Benchmark→ Road Anomaly
AP↑FPR95↓FPS↑GPU Memory Usage↓
RbA w/o RWPM85.426.9211.123.49GiB
T=∞87.515.240.298.88GiB
T=586.895.983.707.24GiB
T=1087.085.772.787.24GiB
T=2087.345.272.047.24GiB
T=5087.365.080.927.24GiB
T=10087.545.020.507.24GiB
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.552 + ], + "angle": 0, + "content": "to substantially reduce the number of pixel embeddings, resulting in significant GPU memory usage. On the other hand, the superpixel strategy requires a considerable amount of runtime for the superpixel extraction. In comparison, our proposed partitioning strategy can achieve good results with higher efficiency and lower resource consumption. This makes it more suitable for applications on practical platforms." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.66 + ], + "angle": 0, + "content": "In-distribution segmentation performance on Cityscapes validation set: We report the results on the Cityscapes validation set in Table 6(Right). All methods use Mask2Former [10] as their backbones. We observe that most SOTA methods affect the in-distribution segmentation performance of Mask2Former, such as RbA (drop from 82.25 to 81.93). However, RWPM can alleviate this problem. Additionally, even for the SOTA segmentation method (Mask2Former), RWPM can further enhance its in-distribution segmentation performance." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.687, + 0.788, + 0.743 + ], + "angle": 0, + "content": "Table 6: Ablation Tables: (Left) shows the results of different strategies for reducing the number of pixel embedding. We find that partitioning is the best. (Right) shows the segmentation performance on the Cityscapes validation set. RWPM can also improve the segmentation performance of in-distribution." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.759, + 0.505, + 0.833 + ], + "angle": 0, + "content": "
Benchmark→ Road Anomaly
Strategy↓AP↑FPR95↓FPS↑Memory
Resize (0.9 full size)87.805.391.4123.82GIB
Resize (0.8 full size)87.245.442.1515.98GiB
Superpixel (30000 pixels)87.515.500.1911.46GiB
Superpixel (20000 pixels)87.035.980.326.31GiB
Partitioning (n = 2)87.345.272.047.24GiB
" + }, + { + "type": "table", + "bbox": [ + 0.547, + 0.755, + 0.744, + 0.837 + ], + "angle": 0, + "content": "
Benchmark→Cityscapes
Method↓mIoU↑
PEBAL-Mask [34]75.32
DenseHybrid-Mask [34]80.27
RbA [34]81.93
RbA+RWPM(Ours)82.16
Mask2Former [10]82.25
Mask2Former+RWPM(Ours)82.43
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.68, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RWPM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.644, + 0.162 + ], + "angle": 0, + "content": "4.3 Visualization of Pixel Embedding Distribution" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.788, + 0.518 + ], + "angle": 0, + "content": "To investigate the effect of RWPM in the embedding space, we visualized the distribution of pixel embeddings by using t-SNE. We selected an image from the Road Anomaly dataset (as shown in Fig 3(a)). Due to the gravel road surface in this image, there is a significant domain gap between the road in this image and the roads in the training set, resulting in many false positives on the road (as shown in the PEBAL results in Fig 3(a)). Then, we randomly selected 300 inlier pixels and 150 outlier pixels from the red dashed box region of this image for visualization. Fig 3(b) and Fig 3(c) respectively show the distributions of these pixels in the space without and with applying RWPM. In Fig 3(b) and Fig 3(c), deep blue circles represent inlier pixels (road), light blue circles represent outlier pixels (boar), and X symbols represent the inlier class prototype (the vector of network weights of classifier). From Fig 3(b), we observe that without using RWPM, due to the influence of data diversity, the distribution of inliers and outliers is distorted. Consequently, it is difficult for an anomaly scoring function to distinguish between inlier and outlier pixels based on their logits. However, from Fig 3(c), we find that RWPM can significantly alleviate the problem. Specifically, the pixel embeddings of the same category form more compact clusters, enabling the anomaly scoring function to clearly discriminate between inliers and outliers based on the logit. This is consistent with the description in the toy sample in Fig 1. Note that, in the top left corner of Fig 3(c), some inlier points become very close to outlier points after applying RWPM. This is because these points are located in the boundaries of anomalous objects, making it very difficult to discriminate them." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.546, + 0.379, + 0.646 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.65, + 0.361, + 0.664 + ], + "angle": 0, + "content": "(a) Sample Image" + }, + { + "type": "image", + "bbox": [ + 0.395, + 0.551, + 0.573, + 0.642 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.391, + 0.65, + 0.582, + 0.664 + ], + "angle": 0, + "content": "(b) Distribution w/o RWPM" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.553, + 0.779, + 0.642 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.65, + 0.792, + 0.665 + ], + "angle": 0, + "content": "(c) Distribution with RWPM" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.676, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Fig. 3: The visualization of pixel embedding distribution. All embeddings are extracted from PEBAL model. We use cosine distance as the distance metric. We observed that RWPM can optimize the distribution of pixel embeddings in the space." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.771, + 0.414, + 0.787 + ], + "angle": 0, + "content": "4.4 Qualitative Result" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Finally, to visually demonstrate the effectiveness of RWPM, we present qualitative results. Specifically, we compare the results of RbA on the Road Anomaly dataset without and with RWPM. The qualitative results are illustrated in Fig 4." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "Z. Zeng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "Fig 4 shows that RWPM significantly improves the performance of RbA, reducing both the false positive rate and false negative rate. For example, in the first row, RbA generates some false positive results on the road and some false negative results on the manhole region (third column of the first row). However, after applying RWPM, these false positive/negative results are effectively eliminated (fourth column of the first row). Additionally, RWPM assists the RbA model to accurately detect and localize each anomalous object at the component level, such as more precise edge detection. This observation aligns with the results of the component-level metrics presented in Table 3." + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.307, + 0.773, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.498, + 0.788, + 0.555 + ], + "angle": 0, + "content": "Fig. 4: Qualitative results: The first column displays the input image, while the second column shows its corresponding ground truth. The third column displays the anomaly score results generated by RbA, and the fourth column presents the anomaly score results after applying RWPM to RbA. Yellow indicates high anomaly scores." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.601, + 0.528, + 0.618 + ], + "angle": 0, + "content": "5 Conclusions and Discussions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.78 + ], + "angle": 0, + "content": "In this paper, we propose a simple yet effective method, called Random Walk on Pixel Manifolds (RWPM), to improve the performance of existing anomaly segmentation methods. RWPM is a post-processing method that can be directly integrated into existing anomaly detection frameworks without requiring additional training or modifications to the network structure. RWPM mitigates the impact of data diversity on manifold structure by diffusing and updating pixel embeddings on pixel manifolds using random walk. This process enables anomaly scoring functions to obtain more accurate estimations of anomaly scores. Additionally, we propose Partial Random Walk strategy, which significantly improves the efficiency of RWPM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "While RWPM achieves significant results in most cases, it is limited when the pre-trained model itself exhibits significant inaccuracies. To address this limitation, we plan to enhance the model's generalization ability during training to accurately detect a broader range of diverse and unknown anomalies." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.68, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RWPM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.175, + 0.556, + 0.19 + ], + "angle": 0, + "content": "This work was supported by SenseTime Japan." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.211, + 0.323, + 0.226 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.239, + 0.785, + 0.28 + ], + "angle": 0, + "content": "1. Achanta, R., Shaji, A., Smith, K., Lucchi, A., Fua, P., Süsstrunk, S.: Slic superpixels compared to state-of-the-art superpixel methods. IEEE transactions on pattern analysis and machine intelligence 34(11), 2274-2282 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.281, + 0.785, + 0.307 + ], + "angle": 0, + "content": "2. Ackermann, J., Sakaridis, C., Yu, F.: Maskomaly: Zero-shot mask anomaly segmentation. arXiv preprint arXiv:2305.16972 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.308, + 0.785, + 0.348 + ], + "angle": 0, + "content": "3. Bai, S., Bai, X., Tian, Q., Latecki, L.J.: Regularized diffusion process for visual retrieval. In: Proceedings of the AAAI conference on artificial intelligence. vol. 31 (2017). https://doi.org/10.1609/aaai.v31i1.11216" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.349, + 0.785, + 0.389 + ], + "angle": 0, + "content": "4. Bai, S., Zhou, Z., Wang, J., Bai, X., Jan Latecki, L., Tian, Q.: Ensemble diffusion for retrieval. In: Proceedings of the IEEE International conference on computer vision. pp. 774-783 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.39, + 0.785, + 0.442 + ], + "angle": 0, + "content": "5. Blum, H., Sarlin, P.E., Nieto, J., Siegwart, R., Cadena, C.: The fishyscapes benchmark: Measuring blind spots in semantic segmentation. International Journal of Computer Vision 129, 3119-3135 (2021). https://doi.org/10.1007/s11263-021-01511-6" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.444, + 0.785, + 0.485 + ], + "angle": 0, + "content": "6. Chan, R., Lis, K., Uhlemeyer, S., Blum, H., Honari, S., Siegwart, R., Fua, P., Salzmann, M., Rottmann, M.: Segmentmeifyoucan: A benchmark for anomaly segmentation. arXiv preprint arXiv:2104.14812 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.485, + 0.785, + 0.526 + ], + "angle": 0, + "content": "7. Chan, R., Rottmann, M., Gottschalk, H.: Entropy maximization and meta classification for out-of-distribution detection in semantic segmentation. In: Proceedings of the ieee/cvf international conference on computer vision. pp. 5128-5137 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.527, + 0.785, + 0.58 + ], + "angle": 0, + "content": "8. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence 40(4), 834-848 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.581, + 0.785, + 0.635 + ], + "angle": 0, + "content": "9. Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 801-818 (2018). https://doi.org/10.1007/978-3-030-01234-2_49" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.636, + 0.785, + 0.689 + ], + "angle": 0, + "content": "0. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290–1299 (2022). https://doi.org/10.1109/CVPR52688.2022.00135" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.69, + 0.785, + 0.731 + ], + "angle": 0, + "content": "1. Cheng, B., Schwing, A., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems 34, 17864-17875 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.785, + 0.772 + ], + "angle": 0, + "content": "2. Choi, H., Jeong, H., Choi, J.Y.: Balanced energy regularization loss for out-of-distribution detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15691-15700 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.838 + ], + "angle": 0, + "content": "3. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3213-3223 (2016). https://doi.org/10.1109/CVPR.2016.350" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.239, + 0.785, + 0.838 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.128 + ], + "angle": 0, + "content": "Z. Zeng et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "14. Creusot, C., Munawar, A.: Real-time small obstacle detection on highways using compressive rbm road reconstruction. In: 2015 IEEE Intelligent Vehicles Symposium (IV). pp. 162-167. IEEE (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.243 + ], + "angle": 0, + "content": "15. Di Biase, G., Blum, H., Siegwart, R., Cadena, C.: Pixel-wise anomaly detection in complex driving scenes. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16918-16927 (2021). https://doi.org/10.1109/CVPR46437.2021.01664" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "16. Gal, Y., Ghahramani, Z.: Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In: international conference on machine learning. pp. 1050-1059. PMLR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.288, + 0.785, + 0.314 + ], + "angle": 0, + "content": "17. Grady, L.: Random walks for image segmentation. IEEE transactions on pattern analysis and machine intelligence 28(11), 1768-1783 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "18. Grcic, M., Bevandić, P., Šegvic, S.: Densehybrid: Hybrid anomaly detection for dense open-set recognition. In: European Conference on Computer Vision. pp. 500-517. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.397 + ], + "angle": 0, + "content": "19. Grcic, M., Saric, J., Segvic, S.: On advantages of mask-level recognition for outlier-aware segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2936-2946 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "20. Hendrycks, D., Basart, S., Mazeika, M., Zou, A., Kwon, J., Mostajabi, M., Steinhardt, J., Song, D.: Scaling out-of-distribution detection for real-world settings. arXiv preprint arXiv:1911.11132 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.44, + 0.785, + 0.48 + ], + "angle": 0, + "content": "21. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. In: International Conference on Learning Representations (2017), https://openreview.net/forum?id=Hkg4TI9x1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.508 + ], + "angle": 0, + "content": "22. Hendrycks, D., Mazeika, M., Dietterich, T.: Deep anomaly detection with outlier exposure. arXiv preprint arXiv:1812.04606 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.549 + ], + "angle": 0, + "content": "23. Huang, Z., Li, S.: From difference to similarity: A manifold ranking-based hyperspectral anomaly detection framework. IEEE Transactions on Geoscience and Remote Sensing 57(10), 8118-8130 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.603 + ], + "angle": 0, + "content": "24. Iscen, A., Tolias, G., Avrithis, Y., Chum, O.: Mining on manifolds: Metric learning without labels. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7642-7651 (2018). https://doi.org/10.1109/CVPR.2018.00797" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.605, + 0.785, + 0.66 + ], + "angle": 0, + "content": "25. Iscen, A., Tobias, G., Avrithis, Y., Furon, T., Chum, O.: Efficient diffusion on region manifolds: Recovering small objects with compact cnn representations. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2077-2086 (2017). https://doi.org/10.1109/CVPR.2017.105" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.715 + ], + "angle": 0, + "content": "26. Jung, S., Lee, J., Gwak, D., Choi, S., Choo, J.: Standardized max logits: A simple yet effective approach for identifying unexpected road obstacles in urban-scene segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15425-15434 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.785, + 0.757 + ], + "angle": 0, + "content": "27. Lakshminarayanan, B., Pritzel, A., Blundell, C.: Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.798 + ], + "angle": 0, + "content": "28. Liang, C., Wang, W., Miao, J., Yang, Y.: Gmmseg: Gaussian mixture based generative semantic segmentation models. Advances in Neural Information Processing Systems 35, 31360-31375 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "29. Liang, S., Li, Y., Srikant, R.: Enhancing the reliability of out-of-distribution image detection in neural networks. In: International Conference on Learning Representations (2018), https://openreview.net/forum?id=H1VGkIxRZ" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.68, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RWPM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.218 + ], + "angle": 0, + "content": "30. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014). https://doi.org/10.1007/978-3-319-10602-1_48" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.259 + ], + "angle": 0, + "content": "31. Lis, K., Nakka, K., Fua, P., Salzmann, M.: Detecting the unexpected via image resynthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2152-2161 (2019). https://doi.org/10.1109/ICCV.2019.00224" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.787, + 0.313 + ], + "angle": 0, + "content": "32. Liu, Y., Ding, C., Tian, Y., Pang, G., Belagiannis, V., Reid, I., Carneiro, G.: Residual pattern learning for pixel-wise out-of-distribution detection in semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1151-1161 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.314, + 0.787, + 0.369 + ], + "angle": 0, + "content": "33. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.787, + 0.411 + ], + "angle": 0, + "content": "34. Nayal, N., Yavuz, M., Henriques, J.F., Güney, F.: Rba: Segmenting unknown regions rejected by all. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 711-722 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.787, + 0.453 + ], + "angle": 0, + "content": "35. Ohgushi, T., Horiguchi, K., Yamanaka, M.: Road obstacle detection method based on an autoencoder with semantic segmentation. In: Proceedings of the Asian Conference on Computer Vision (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.493 + ], + "angle": 0, + "content": "36. Rai, S.N., Cermelli, F., Fontanel, D., Masone, C., Caputo, B.: Unmasking anomalies in road-scene segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4037-4046 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.787, + 0.549 + ], + "angle": 0, + "content": "37. Tian, Y., Liu, Y., Pang, G., Liu, F., Chen, Y., Carneiro, G.: Pixel-wise energy-biased abstention learning for anomaly segmentation on complex urban driving scenes. In: European Conference on Computer Vision. pp. 246-263. Springer (2022). https://doi.org/10.1007/978-3-031-19842-7_15" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.549, + 0.787, + 0.605 + ], + "angle": 0, + "content": "38. Vojir, T., Šipka, T., Aljundi, R., Chumerin, N., Reino, D.O., Matas, J.: Road anomaly detection by partial image reconstruction with segmentation coupling. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15651-15660 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.646 + ], + "angle": 0, + "content": "39. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems 34, 12077-12090 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.646, + 0.787, + 0.701 + ], + "angle": 0, + "content": "40. Yang, F., Hinami, R., Matsui, Y., Ly, S., Satoh, S.: Efficient image retrieval via decoupling diffusion into online and offline processing. In: Proceedings of the AAAI conference on artificial intelligence. vol. 33, pp. 9087-9094 (2019). https://doi.org/10.1609/aaai.v33i01.33019087" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.787, + 0.742 + ], + "angle": 0, + "content": "41. Yang, F., Wang, Z., Xiao, J., Satoh, S.: Mining on heterogeneous manifolds for zero-shot cross-modal image retrieval. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 12589-12596 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.787, + 0.772 + ], + "angle": 0, + "content": "42. Zhou, D., Bousquet, O., Lal, T., Weston, J., Scholkopf, B.: Learning with local and global consistency. Advances in neural information processing systems 16 (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.799 + ], + "angle": 0, + "content": "43. Zhou, D., Weston, J., Gretton, A., Bousquet, O., Schölkopf, B.: Ranking on data manifolds. Advances in neural information processing systems 16 (2003)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.799 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Random Walk on Pixel Manifolds for Anomaly Segmentation of Complex Driving Scenes/40e3d166-d87c-43d0-aeb7-8851ed372fb1_origin.pdf b/2024/Random Walk on Pixel Manifolds for Anomaly Segmentation of Complex Driving Scenes/40e3d166-d87c-43d0-aeb7-8851ed372fb1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ce429fc882b2b43047c73f8b8ca49f447accade1 --- /dev/null +++ b/2024/Random Walk on Pixel Manifolds for Anomaly Segmentation of Complex Driving Scenes/40e3d166-d87c-43d0-aeb7-8851ed372fb1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:866b2fb9d7ac249b5c18654c0ba3a10a01d2605611e9fdff3dad251820ee5a41 +size 1659874 diff --git a/2024/Random Walk on Pixel Manifolds for Anomaly Segmentation of Complex Driving Scenes/full.md b/2024/Random Walk on Pixel Manifolds for Anomaly Segmentation of Complex Driving Scenes/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b19c4c21e66aad4d65698b4c4fee3b1991b12543 --- /dev/null +++ b/2024/Random Walk on Pixel Manifolds for Anomaly Segmentation of Complex Driving Scenes/full.md @@ -0,0 +1,252 @@ +# Random Walk on Pixel Manifolds for Anomaly Segmentation of Complex Driving Scenes + +Zelong Zeng1 and Kaname Tomite1 + +SenseTime Japan, Tokyo, Japan {zengzelong,tomite}@sensetime.jp + +Abstract. In anomaly segmentation for complex driving scenes, state-of-the-art approaches utilize anomaly scoring functions to calculate anomaly scores. For these functions, accurately predicting the logits of inlier classes for each pixel is crucial for precisely inferring the anomaly score. However, in real-world driving scenarios, the diversity of scenes often results in distorted manifolds of pixel embeddings in the space. This effect is not conducive to directly using the pixel embeddings for the logit prediction during inference, a concern overlooked by existing methods. To address this problem, we propose a novel method called Random Walk on Pixel Manifolds (RWPM). RWPM utilizes random walks to reveal the intrinsic relationships among pixels to refine the pixel embeddings. The refined pixel embeddings alleviate the distortion of manifolds, improving the accuracy of anomaly scores. Our extensive experiments show that RWPM consistently improve the performance of the existing anomaly segmentation methods and achieve the best results1. + +Keywords: Autonomous vehicles $\cdot$ Anomaly segmentation $\cdot$ Semantic segmentation + +# 1 Introduction + +Semantic segmentation [9-11,13,39] plays a crucial role in the field of automated driving, which can aid the automated driving system in perceiving and understanding the surrounding environment, thereby facilitating the system make accurate decision. Typically, the semantic segmentation models are trained on closed-datasets with fixed semantic categories, showing good recognition capabilities for inlier classes (e.g., roads, vehicles, etc.). However, these models often fail to effectively detect potential outlier class objects (e.g., animals, fallen tires, etc.) in real-world scenarios. Due to the critical importance of correctly detecting outlier class targets for the safety of autonomous driving, there is a growing focus on addressing this task, namely, anomaly segmentation. + +The standard paradigm for anomaly segmentation is assigning an anomaly score to each pixel of an input image, where a higher score indicates a greater probability that the pixel belongs to an outlier. Thus, the way used for estimating the anomaly score is crucial for anomaly segmentation methods. In + +(i) An ideal manifold structure +Fig. 1: A toy example about manifolds of pixel embeddings. The dashed circles indicate regions with high predicted logit to the corresponding inlier class. For the pixel in these regions, its anomaly score is small. The prototype is the vector of network weights of the classifier for the corresponding inlier class. (i) An ideal manifold structure. All inlier pixels are within the corresponding high logit region, whereas all outlier pixels are outside the areas. The anomaly score derived from the anomaly scoring function effectively discriminates between them. (ii) The manifolds are affected by the diversity of data, causing some inlier pixels to deviate from the corresponding high logit region, while some outlier pixels approach the regions. This results in false positives/negatives in the anomaly scores. However, pixels of the same class are still in the same manifold, indicating that the manifold structure can be utilized to reveal the intrinsic relationships between pixels. (iii) Our RWPM utilizes random walks to capture the manifold structure to diffuse and update pixel embeddings. The embeddings within the same manifold tend to become more similar after the updating. Fig 3 and Fig 4 respectively present the visualization of embedding distributions and the qualitative results of real examples, consistent with the description of this toy example. +![](images/66465a8b974307d9e4d39db93e496fd578b716b3a4fdf1ea9f19da4f91b66e54.jpg) +Inlier pixel + +(ii) A distorted manifold structure +![](images/e5dd3250608f82a6b1a86386c513f141d9085137879f723e452084cd3cb51639.jpg) +$\times \times$ Inlier class prototype + +(iii) A refined manifold structure +![](images/7eb03a33c34f0e318dc50b70b00fe594d25cdc69d3ad192468b00ec5e72d85c5.jpg) +Outlier pixel + +early approaches, the anomaly score is often given by estimating the prediction uncertainty [16, 21, 27, 29], or by comparing the discrepancy between the original image and its reconstructed image generated based on the semantic segmentation predictions [14, 15, 31, 35]. Different from early methods, recent methods [12, 19, 32, 34, 36, 37] adopt Outlier Exposure (OE) strategy [22], which directly learn a pixel-level anomaly class to detect anomalous pixels that deviate from inlier classes. Specifically, these methods propose anomaly scoring functions to infer the anomaly scores and, based on the functions, leverage auxiliary outlier data to fine-tune a semantic segmentation model. The auxiliary outlier data is typically generated by extracting outlier-class objects from other datasets (e.g., COCO [30]) and incorporating them into driving scene images [34, 36, 37]. Furthermore, some of them [19, 34, 36] adopt mask-based segmentation networks instead of pixel-based segmentation networks, achieving state-of-the-art (SOTA) performance in anomaly segmentation. In this paper, we concentrate on the current methods using OE strategy. These methods indicate superior performance to the early methods using prediction uncertainty or image reconstruction. + +The anomaly scoring functions in OE methods infer anomaly scores based on the inlier class logit predictions from a semantic segmentation model. Specifically, the logic of these functions can be represented as follows: when the predicted logits of a pixel embedding are small across all inlier classes, indicating that the + +pixel is dissimilar to all inlier classes and is an outlier with high probability, the function yields a higher anomaly score. Therefore, accurately measuring the logits of each pixel is crucial for precisely evaluating the anomaly score. However, in practical open-world driving scenarios, various environmental conditions (such as lighting, road materials, etc.) and a diverse range of unknown class objects are present. During the inference phase, the diversity of driving scenes often results in distorted manifolds of pixel embedding in the embedding space, causing pixel embeddings to deviate from the ideal distribution. This phenomenon has been validated in previous works of metric learning [3,25,40,41]. Since the logits are calculated from the inner product (i.e., cosine distance) of the pixel embeddings and the vectors of classifiers' network weights, the distorted manifolds are not conducive to directly using the pixel embeddings for the logit prediction. However, current anomaly segmentation methods ignore this issue and directly use the pixel embeddings for the logit prediction in the inference phase, so the results often fail to accurately reflect the anomaly score for the corresponding pixels, causing high false positive/negative rate (see Fig. 1(i) and Fig. 1(ii)). + +In this paper, we propose a novel method, called Random Walk on Pixel Manifolds (RWPM), to alleviate the effect of data diversity on the manifolds. RWPM utilizes random walks to capture the manifold structure of the pixel embeddings to diffuse and update the embedding of each pixel. Since the structure of data manifolds implicitly contains the intrinsic relationship among data points (see Fig. 1(ii)) [3,24,40,43], using random walks to measure the similarity between pixels on manifolds (i.e., distance on manifolds rather than Euclidean distance) better reflects the intrinsic relationships between pixels. Consequently, diffusing and updating pixel embeddings based on the similarity on manifolds results in high similarity among pixels on the same manifold and low similarity among pixels on different manifolds. In the embedding space, this manifests as pixel embeddings on the same manifold forming more compact clusters, thereby mitigating the distortion effect of data diversity on the manifolds, which facilitates the logit prediction (see Fig. 1(iii)). Furthermore, since the typically large number of pixels in an image, utilizing all pixels for the random walks results in significant memory and computational overhead, which damages the deployment and real-time execution of our RWPM. To address this issue, we introduce the Partial Random Walk strategy for RWPM, reducing memory consumption and enhancing the operational efficiency of RWPM. Note that, RWPM is proposed to use in the inference phase, thus no any additional training is required. Additionally, RWPM can be directly integrated into existing anomaly segmentation framework without the need for any extra modification to the network structure. Extensive experiments in various anomaly segmentation benchmarks of road scenes, namely Fishyscapes [5], Road Anomaly [31], and SMIYC [6], demonstrate that RWPM consistently improves the performance of existing methods, achieving SOTA results. To summarise, our contributions are the following: + +functions in inferring anomaly scores. This problem is ignored by existing anomaly segmentation methods. + +- We propose the Random Walk on Pixel Manifolds (RWPM) that utilizes random walks to reveal the intrinsic relationships among pixels to refine the pixel embeddings. This way alleviate the effect on manifolds caused by the driving scene diversity. + +- We propose the Partial Random Walk strategy for RWPM to reduce the memory consumption and improve the operational efficiency. + +# 2 Related Work + +Anomaly Segmentation. Anomaly segmentation aims to detect the outlier class object in some specific scenes, such as complex driving scenes [6, 13, 31]. Existing anomaly segmentation can be broadly divided into three categories: (a) Uncertainty-based methods, (b) Reconstruction-based methods and (c) Outlier Exposure methods. Uncertainty-based methods assume that outlier samples result in low-confidence predictions. Based on this assumption, they focus on measuring the pixel-wise anomaly score by estimating the prediction uncertainty through various ways, such as ensembles [27], Bayesian estimation [16], maximum softmax probability [21, 29], and logit [20, 26]. Since segmentation models are trained on close-set, they may misclassify unseen classes with high confidence, leading uncertainty-based methods fail to detect anomalies. Reconstruction-based methods detect anomaly pixel by comparing the discrepancy between the original image and its reconstructed image generated based on the semantic segmentation predictions [14, 15, 31, 35]. These methods are challenged by the accuracy of segmentation prediction and the difficulties of training. Recent methods [7, 12, 19, 34, 36, 37] introduce the Outlier Exposure (OE) strategy [22] that utilizes auxiliary outlier data and anomaly scoring functions to learn models. The auxiliary outlier data is the image from other datasets (e.g., COCO [30]), or cutting the outlier objects from other datasets and pasting them into the inlier scenes, and the anomaly scoring functions are used to calculate the anomaly score for each pixel. Compared with previous works, OE strategy can reach better performance. More recently, some OE methods [19, 34, 36] adopt mask-based segmentation networks [10, 11], instead of pixel-based segmentation networks [8, 9], as their backbones. Based on the powerful capability of mask-based networks in segmentation tasks, these methods significantly reduces false detection rates, achieving SOTA results. In this paper, we only concentrate on the OE methods. The anomaly scoring functions employed in OE methods are based on the predicted logits of inlier classes. However, during the inference phase, the diversity of driving scenes results in distorted manifolds of pixel embeddings. As a result, the scoring functions may fail to use the embeddings to accurately calculate logits, thereby affects the accuracy of measuring anomaly scores. Current OE methods ignore this problem. Our proposed RWPM addresses this problem by utilizing random walks on pixel manifolds to reveal the intrinsic relationships among pixels and refine the pixel embeddings. To the best of our knowledge, RWPM is the first work to propose and resolve this problem. + +Mining on Data Manifolds. Mining on data manifolds, also known as diffusion process, is a well-established technique commonly used in retrieval tasks [3, 4, 25, 40, 43]. These works utilize random walks on data manifolds to capture the data manifold structure for revealing the intrinsic relationship between data points. The data manifold is constructed by a graph, where each node represents a sample and each edge presents the connection between a node and its neighborhoods, weighted proportionally according to their distance. For retrieval tasks, the images of the same object may be shown in different conditions (e.g., lighting and angles). This diversity often leads to the manifold structure is not conducive to accurately ranking using distance metrics directly [25, 40]. Thus, diffusion process is used to diffuse the state (typically initialized to 1) of a query image on the data manifolds. If a sample is on the same data manifold with the query (the distance on manifolds is close), then more energy of the state will flow to that sample. Therefore, the final state of all samples can be seen as the accurate ranking scores. Moreover, diffusion process is also adopted in semi-supervised learning [24, 42] to propagate the label of labeled samples to unlabeled samples. Similarly, some unsupervised segmentation works [17, 23] utilize diffusion process to propagate the labels of manually annotated pixels to the unlabeled pixels. These existing works utilize random walks on data manifolds for ranking or label propagation. Different from these works, our target is to solve the problem occurring in pixel manifolds for anomaly scoring functions. We propose innovatively using random walks to diffuse and update pixel embeddings, eliminating the effect of distorted manifolds directly at the embedding-level, thereby improving the accuracy of subsequent processing. + +# 3 Proposed Method + +# 3.1 Preliminaries + +We denote an image space $\mathcal{X} \subset \mathbb{R}^{3 \times H \times W}$ , where $H$ and $W$ are the height and width, respectively. During the inference phase, the goal for an anomaly segmentation model $f$ is to map an input image $\pmb{x} \in \mathcal{X}$ to an anomaly space, i.e., $f: \pmb{x} \mapsto \mathcal{Z} \subset \mathbb{R}^{H \times W}$ . Specifically, an anomaly segmentation model typically consists of the backbone encoder-decoder $F$ and the classifier $C$ . For an input image $\pmb{x}$ , firstly, its pixel embedding map is encoded as $\pmb{p} = F(\pmb{x}) \in \mathbb{R}^{d \times H \times W}$ , where $d$ is the dimension of the pixel embedding. Then, its logit map $\pmb{l}$ is predicted by $\pmb{l} = C(F(\pmb{x})) \in \mathbb{R}^{K \times H \times W}$ , where $K$ is the number of inlier classes. Note that, depending on the backbone network, the size of the embedding map and logit map may sometimes be smaller than the original image. For the sake of convenience in this paper, we do not make a special distinction. Finally, a scoring function is used to calculate the anomaly score map $\pmb{z} \in \mathcal{Z}$ based on the logit map $\pmb{l}$ . + +For anomaly scoring functions, their logic can be represented as follows: when the predicted logits of a pixel embedding are small across all inlier classes, indicating that the pixel is dissimilar to all inlier classes and is likely an outlier, the function generates a higher anomaly score. For example, let's define $l_{h,w}(k)$ + +![](images/0cc1a045b91634a7ad204f533266b3087e3c813bfb6a897fc97866f134f6e246.jpg) +Fig. 2: Overview. This figure illustrates the application of RWPM to an existing anomaly segmentation framework during the inference phase. The red dashed box highlights the RWPM part. First, the encoder-decoder is used to extract the pixel embedding map of the input image. The pixel embedding map are subsequently partitioned into $n^2$ sub-maps. For each sub-map, we update its pixel embeddings by using random walks to obtain a refined sub-map. Finally, the refined sub-maps are concatenated to form the refined embedding map, which are then input into the subsequent network structure. Notably, RWPM can be directly integrated into existing frameworks without requiring extra training or changes to the network structure. + +is the logit of a pixel $\pmb{x}_{h,w}$ for class $k$ . The scoring function in PEBAL [37] is $E_{h,w} = -\log \sum_{k \in \{1, \dots, K\}} \exp(\pmb{l}_{h,w}(k))$ . Obviously, when the anomaly score $E_{h,w}$ is large, it indicates that the logits of the pixel across all inlier classes are small, i.e., the pixel $\pmb{x}_{h,w}$ is dissimilar from all inlier classes (see Section S.2 of the supplementary material for the analysis of other anomaly scoring functions). + +# 3.2 Graph Construction + +We begin by constructing a graph to represent the manifolds of pixel embeddings. In our approach, we utilize an affinity matrix based on cosine similarity to characterize the relationships between pixels. Specifically, we first use the encoder-decoder component of an anomaly segmentation model to extract the pixel embeddings $\pmb{p} \in \mathbb{R}^{d \times H \times W}$ . For ease of description and computation, we reshape the $\pmb{p}$ as $\pmb{p}^r \in \mathbb{R}^{HW \times d}$ , where each row represents a pixel's embedding. We then normalize each row of $\pmb{p}^r$ to obtain $\widehat{\pmb{p}}^r$ , i.e., $\widehat{\pmb{p}}_i^r = \pmb{p}_i^r / \| \pmb{p}_i^r \|$ . The affinity matrix is defined as $\mathbf{W} \in \mathbb{R}^{HW \times HW}$ , where each component is obtained by: + +$$ +\mathbf {W} _ {i j} = \left\{ \begin{array}{c c} \langle \hat {\pmb {p}} _ {i} ^ {r}, \hat {\pmb {p}} _ {j} ^ {r} \rangle & i \neq j \\ 0 & i = j \end{array} , \right. \tag {1} +$$ + +where $\langle \cdot, \cdot \rangle$ denotes inner product, $\mathbf{W}_{ii} = 0$ is guaranteed to circumvent self-loops in the graph represented by $\mathbf{W}$ . + +To mitigate the influence of noise, locally constrained random walk methods [24, 40, 41] adopt local constraints to construct the graph, i.e., each point is only connected to its locally nearest neighbors in the graph. Determining the nearest neighbors for each point traditionally involves applying nearest neighbor search algorithms, such as the $k$ -NN algorithm. However, in anomaly segmentation tasks, the size of the graph is large and the graph is constructed online, + +making the search algorithms impractical. To benefit from the local constraints, we utilize a softmax function to normalize each row of $\mathbf{W}$ to obtain locally constrained graph $\mathbf{S}$ : + +$$ +\mathbf {S} _ {i j} = \left\{ \begin{array}{c c} \frac {\exp \left(\mathbf {W} _ {i j} / \tau\right)}{\sum_ {i = 1 , i \neq j} ^ {H W} \exp \left(\mathbf {W} _ {i j} / \tau\right)} & i \neq j \\ 0 & i = j \end{array} , \right. \tag {2} +$$ + +where $\tau$ is a temperature hyper-parameter, we set its value less than 1.0 to remove the effect of non-neighbor points. Compared to typical search algorithms, Eq. 2 can impose local constraints more efficiently and effectively (see Section S.4 of the supplementary material), and keep the scale of embedding in subsequent random walks. + +# 3.3 Random Walk Process + +Now we have constructed the graph $\mathbf{S}$ , we can start the random walks. Different from previous mining on data manifolds works utilizing random walks to propagate the states of known samples to other unknown samples, we use a random walk to propagate the embedding between pixels on manifolds. + +We first define $\mathbf{m}^t\in \mathbb{R}^{HW\times d}$ as the updated pixel embeddings in the $t$ -th step of the random walks. $\mathbf{m}^0$ is initialized as $\pmb{p}^r$ . Then, each iteration step of the random walks is described as: + +$$ +\mathbf {m} ^ {t + 1} = \alpha \mathbf {S} \mathbf {m} ^ {t} + (1 - \alpha) \mathbf {m} ^ {0}, \alpha \in (0, 1), \tag {3} +$$ + +where $\alpha$ is the probability continuing the random walk from the current state $\mathbf{m}^t$ , while $(1 - \alpha)$ represents the probability of restarting from the initial state $\mathbf{m}^0$ . At each iteration, guided by the manifold structure, each pixel diffuses its embedding to other pixels while also receiving embedding from other pixels for updating its own embedding. The updated pixel embedding can be seen as a weighted ensemble of other pixel embeddings. If pixels are similar on manifolds, their embeddings will gradually become similar. Mathematically speaking, the whole iteration process can be written as a closed-from as follows [43]: + +$$ +\mathbf {m} ^ {\infty} = (1 - \alpha) (\mathbf {I} - \alpha \mathbf {S}) ^ {- 1} \mathbf {m} ^ {0}, \tag {4} +$$ + +where $\mathbf{I} \in \mathbb{R}^{HW \times HW}$ is an identity matrix. Finally, $\mathbf{m}^{\infty}$ is used as the refined pixel embeddings for the subsequent anomaly segmentation process. + +# 3.4 Partial Random Walk + +When the size of the image is large, the size of the relevant matrices used for the calculation of random walks becomes extremely huge as well. For instance, when the image size is $(512 \times 1024)$ , the size of matrix $S$ reaches $(524288 \times 524288)$ . This significantly increases the memory consumption and computational time of RWPM, which is not conducive to online operation. Unfortunately, in the field + +of autonomous driving, the image size of driving scenes is generally large. To address this problem, we propose Partial Random Walk, which consists of the following two aspects: embedding map partitioning and limited iteration. + +To improve efficiency, we first propose embedding map partitioning strategy to reduce the size of the pixel embedding map used for random walks. Specifically, for a pixel embedding map $\pmb{p} \in \mathbb{R}^{d \times H \times W}$ , we equally divide it into $n \times n$ sub-maps $\left\{\pmb{p}^{(1)}, \dots, \pmb{p}^{(n^2)}\right\} \in \mathbb{R}^{d \times \frac{H}{n} \times \frac{W}{n}}$ . Then, for each sub-map in $\left\{\pmb{p}^{(1)}, \dots, \pmb{p}^{(n^2)}\right\}$ , we follow the steps of Eq. 1 and Eq. 2 to construct their sub-graphs $\left\{\pmb{S}_{\pmb{p}^{(1)}}, \dots, \pmb{S}_{\pmb{p}^{(n^2)}}\right\} \in \mathbb{R}^{d \times \frac{HW}{n^2} \times \frac{HW}{n^2}}$ . We also attempted other methods to reduce the number of pixel embeddings, such as resize or superpixel [1]. Experimental results indicate that embedding map partitioning is the optimal solution among them (see Section 4.2). + +Next, for each sub-map $\pmb{p}^{(i)}$ , we perform random walks with the limited iteration strategy in their corresponding sub-graphs $S_{\pmb{p}^{(i)}}$ . That is, we use Eq. 3 for a small number of iterations instead of using the closed-form Eq. 4. We found that a few iterations by using Eq. 3 are enough to yield high quality results. The time complexity of Eq. 3 with $T$ iterations for each sub-map is $O(Td(\frac{HW}{n^2})^2)$ , while the time complexity of Eq. 4 is $O((\frac{HW}{n^2})^3) + O(d(\frac{HW}{n^2})^2)$ . Since $Td$ is much smaller than $\frac{HW}{n^2}$ , the limited iteration strategy can significantly improve the efficiency of RWPM. After the random walks, we obtain the refined pixel embeddings sub-map $\mathbf{m}_{\pmb{p}^{(i)}}^T$ for each sub-map $\pmb{p}^{(i)}$ . Finally, we concatenate the refined sub-maps based on their positions to obtain the final refined pixel embedding map. Additionally, when $n$ is large (observed experimentally as $n > 2$ ), due to spatial constraints, the baseline values of anomaly scores outputted on each sub-map may vary. Hence, a calibration is performed. For adjacent refined sub-maps $\pmb{p}^{(i)'}$ and $\pmb{p}^{(j)'}$ , we compute the average anomaly scores $I$ and $J$ of all pixels on their adjacent edges, respectively. Subsequently, using $\pmb{p}^{(i)'}$ as the reference, we calculate the ratio $\frac{I}{J}$ , and then multiply all anomaly scores on sub-map $\pmb{p}^{(j)'}$ by this ratio to achieve the calibration. + +Fig. 2 shows the workflow of the proposed RWPM. + +# 4 Experiment + +Dataset: We evaluate our approach on three standard benchmark datasets for anomaly segmentation, i.e., Fishyscapes Lost&Found [5], Road Anomaly [31] and Segment Me If You Can (SMIYC) [6]. Fishyscapes Lost&Found has 100 validation images. The domain of this dataset is similar to that of Cityscapes [13] dataset. Road Anomaly dataset contains 60 images of real-world road scenes. Unlike the Fishyscapes Lost&Found dataset, the images in Road Anomaly contain various classes and sizes of anomaly objects, making it more challenging. SMIYC consists of two subsets, Anomaly track and Obstacle track. The Anomaly track focuses on detecting large anomalous objects, while the Obstacle track emphasizes detecting small anomalous objects on the road. Anomaly track contains 10 + +validation images and 100 online test images, while Obstacle track contains 30 validation images and 327 online test images. The Road Anomaly and SMIYC datasets both exhibit high domain shifts compared to the Cityscapes dataset. + +Evaluation Metrics: Following the previous works, we evaluate our method on the standard metrics for anomaly segmentation. On the Fishyscapes Lost&Found and Road Anomaly, we adopt the area under receiver operating characteristics (AUROC), average precision (AP), and the false positive rate at a true positive rate of $95\%$ (FPR95). On the SMIYC, in addition to reporting per-pixel metrics such as AP and FPR95, we also calculate results for component-level metrics [6] such as the component-wise intersection over union (sIoU), positive predictive value (PPV), and mean F1. + +Implementation Details: To demonstrate the effectiveness of RWPM for anomaly segmentation methods, we selected representative methods from both pixel-based and mask-based networks. For pixel-based methods, we chose PE-BAL [37] and Balanced Energy [12]. These methods all utilize DeepLabv3+ [9] with WideResnet38 as the backbone and were trained on the Cityscapes dataset and COCO OE images [37]. For mask-based methods, we selected RbA [34] and Mask2Anomaly [36]. These methods use Mask2Former [10] with Swin-L [33] as the backbone and were also trained on the Cityscapes and COCO OE images. All experiments are conducted on a single NVIDIA RTX A6000 GPU. To ensure efficient implementation on a single GPU for all experiments, we adopt embedding map partitioning and limited iteration in all experiments. Specifically, for pixel-based methods, we set the partitioning parameter $n = 4$ . For mask-based methods, as their embedding map size is smaller, we set $n = 2$ . Additionally, for the random walk, if not specifically emphasized, we set the number of iterations $T = 20 / T = 5$ for the Road Anomaly/the other datasets and the transition probability $\alpha = 0.99$ for all datasets. The $\tau$ is set as 0.01. It is important to note that our method is used in the inference phase. In other words, RWPM is directly applied to the pre-trained models of other existing anomaly segmentation methods, thus requiring no additional training. + +# 4.1 Effectiveness of RWPM + +First, to validate that RWPM can improve the performance of existing anomaly segmentation methods, we integrate our approach with four recent representative methods, namely PEBAL [37], Balanced Energy [12], Mask2Anomaly [36] and RbA [34]. PEBAL and Balanced Energy employ pixel-based networks as their backbones, whereas Mask2Anomaly and RbA utilize mask-based networks as their backbones. We evaluate the results on both the Fishyscapes Lost&Found validation set and Road Anomaly test set. The experimental results are reported in Table 1. These results demonstrate that our proposed RWPM consistently and significantly enhances the performance of existing anomaly segmentation methods, across different backbone architectures. Additional results and parameters details are shown in Section S.3 of the supplementary material. + +Table 1: Comparison with strong and representative anomaly segmentation baselines across different backbone Architectures (Arch). Bold denotes the better results between using RWPM and not using RWPM with the same anomaly segmentation baselines. $\dagger$ indicates that use the calibration in sub-map concatenation as described in Section 3.4. + +
Benchmark→ Fishscapes Lost&Found Road Anomaly
Method ↓ArchAuROC↑AP↑FPR95↓AuROC↑AP↑FPR95↓
PEBAL [37]Pixel-based98.9658.814.7787.6345.1044.58
PEBAL + RWPM† (Ours)Pixel-base99.2066.853.6889.4850.2936.81
Balanced Energy [12]Pixel-based99.0367.072.9388.3141.4841.46
Balanced Energy + RWPM† (Ours)Pixel-base99.2972.952.3890.5148.2632.10
Mask2Anomaly2 [36]Mask-based95.4765.277.7996.5480.0413.95
Mask2Anomaly + RWPM (Ours)Mask-base95.4865.387.3397.4480.097.45
RbA [34]Mask-based98.6270.816.3097.9985.426.92
RbA + RWPM (Ours)Mask-base98.8271.166.1298.0487.345.27
+ +Table 2: Comparison with state-of-the-art anomaly segmentation methods across different backbone Architectures (Arch). Bold and underline denote the best and the second best results, respectively. + +
Benchmark→Fishscapes L&FRoad AnomalyAnomaly TrackObstacle TrackAverage
Method ↓ArchAP↑FPR95↓AP↑FPR95↓AP↑FPR95↓AP↑FPR95↓AP↑FPR95↓
Synboost [15] (CVPR'21)Pixel-based60.5831.0241.8459.7256.4461.8671.343.1557.4638.94
SML [26] (ICCV'21)Pixel-based22.7433.4925.8249.4746.839.53.436.824.6939.82
Meta-OOD [7] (ICCV'21)Pixel-based41.3137.6948.8431.7785.4715.0085.070.7565.1321.30
Learning Embedding [5](IJCV'21)Pixel-based4.6524.36--37.5270.760.8246.38--
Void Classifier [5](IJCV'21)Pixel-based10.2922.11--36.6163.4910.4441.54--
GMMSeg-DL [28] (NeurIPS'22)Pixel-based43.4713.1134.4247.90------
DenseHybrid [18] (ECCV'22)Pixel-based69.795.0931.3963.9777.969.8187.080.2465.5619.78
PEBAL [37] (ECCV'22)Pixel-based58.814.7745.1044.5849.1440.824.9812.6839.5825.71
Balanced Energy [12] (CVPR'23)Pixel-based67.072.9341.4841.46------
RPL-CoroCL [32] (ICCV'23)Pixel-based70.612.5271.6117.7483.4911.6885.930.5877.918.13
Mask2Anomaly [36] (ICCV'23)Mask-based69.449.2280.0413.9588.6214.5793.100.2082.879.49
M2F-EAM [19] (CVPRW'23)Mask-based52.0320.5166.6713.4276.393.966.917.965.4836.43
RbA [34] (ICCV'23)Mask-based70.816.3085.426.9290.8611.5991.850.4684.736.33
RbA + RWPM (Ours)Mask-based71.166.1287.345.2792.0010.1593.300.2886.005.46
+ +Second, to further demonstrate the effectiveness of our RWPM, we compare it with other state-of-the-art (SOTA) methods. Specifically, we adopt the combination of RbA with RWPM as our instance approach for the comparison. In addition to the Fishyscapes and Road Anomaly datasets, we also evaluate our approach on a more challenging dataset, i.e., SMIYC online test set (Anomaly track and Obstacle track). SMIYC is characterized by its high domain shift and diversity of objects. Table 2 presents the results. The results indicate that our method can consistently achieve high performance across various datasets, while the previous SOTA methods fail to do so. We also compare all approaches by computing the average results of AP and FPR95. The results show that our approach outperforms the previous methods and achieve the best results. + +Furthermore, in addition to pixel-level metrics, we also utilize component-level metrics [6] to evaluate performance on the SMIYC online test set. Compared to pixel-level metrics, component-level metrics provide a better evaluation of the detection results for all anomalous objects in the scene. We report the results in Table 3. The experimental results show that RWPM significantly improves + +the performance of RbA in component-level metrics, surpassing other existing SOTA methods by a large margin. This indicates that RWPM can assist models in better detecting anomalous objects and achieving more precise localization of these anomalies [6]. + +Table 3: Component level evaluation: Comparison with state-of-the-art anomaly segmentation methods on the SMIYC online test set. Bold denote the best results. + +
Benchmark→Anomaly TrackObstacle Track
Method ↓ArchsIoU ↑PPV↑mean F1↑sIoU ↑PPV↑mean F1↑
Synboost [15] (CVPR'21)Pixel-based34.6817.819.9944.2841.7537.57
SML [26] (ICCV'21)Pixel-based26.0024.7012.205.1013.303.00
Meta-OOD [7] (ICCV'21)Pixel-based49.3139.5128.7247.8763.6448.51
Void Classifier [5](IJCV'21)Pixel-based21.1422.136.496.3420.275.41
Learning Embedding [5](IJCV'21)Pixel-based33.8620.547.9035.642.872.31
JSRNet [38] (ICCV'21)Pixel-based20.2029.2713.6618.5524.4611.02
DenseHybrid [18] (ECCV'22)Pixel-based54.1724.1331.0845.7450.1050.72
PEBAL [37] (ECCV'22)Pixel-based38.8827.2014.4829.917.555.54
RPL-CoroCL [32] (ICCV'23)Pixel-based49.7629.9630.1652.6156.6556.69
Mask2Fromer [36] (ICCV'23)Mask-based25.2018.2015.305.0021.904.80
Maskomaly [2] (BMVC'23)Mask-based55.451.650.0---
RbA [34] (ICCV'23)Mask-based55.6952.1446.8058.3658.7860.85
RbA + RWPM (Ours)Mask-based57.0061.2558.4458.8972.5169.85
+ +# 4.2 Ablation Study + +If not specifically emphasized, all the results reported in this section are conducted on a RTX A6000 GPU. + +The effect of the partitioning parameter $n$ and calibration: We first test the effect of the embedding map partitioning parameter $n$ and the calibration, and we report the results in Table 4. FPS denotes the frames-per-second results. Row $n = 1$ displays the RWPM results without utilizing the embedding map partitioning strategy, indicating significant computational and memory overhead. However, the partitioning strategy ( $n = 2, 4, 8$ ) significantly improves the running speed and reduces memory consumption while maintaining effectiveness. Additionally, as $n$ increases, the running efficiency also improves, albeit with a slight decrease in performance. However, when $n > 2$ , the calibration can enhance the performance, particularly for the FPR95 metrics. + +The effect of iteration number $T$ : Table 5 shows that, compared to using Eq. 4 ( $T = \infty$ ), utilizing the limited iteration strategy ( $T = 5 \sim 100$ ) can lead to higher efficiency, resulting in even better performance for RWPM. Additionally, as the number of iterations increases, performance tends to improve, but running efficiency (FPS) decreases accordingly. + +Different strategies for reducing the number of pixel embedding: We also attempt other strategies to reduce the number of pixel embedding for RWPM, such as resizing the embedding map or use SLIC [1] algorithm to extract superpixels. Table 6(Left) presents the results. The resize strategy achieves good performance but requires preserving a larger map size. Therefore, it fails + +Table 4: The effect of the partitioning parameter $n$ and calibration. We set $T = {20}$ for all experiments. ${}^{ \dagger }$ indicates that use the calibration in sub-map concatenation. + +
Benchmark→Road Anomaly
AP↑FPR95↓FPS↑GPU Memory Usage↓
RbA w/o RWPM85.426.9211.123.49GiB
n=1 (two GPUs)87.905.170.3274.96GiB
n=287.345.272.047.24GiB
n=2†87.185.252.047.24GiB
n=487.056.314.353.49GiB
n=4†87.175.324.353.49GiB
n=886.676.426.263.55GiB
n=8†86.915.416.263.55GiB
+ +Table 5: The effect of iteration number $T$ . We set $n = 2$ for all experiments. + +
Benchmark→ Road Anomaly
AP↑FPR95↓FPS↑GPU Memory Usage↓
RbA w/o RWPM85.426.9211.123.49GiB
T=∞87.515.240.298.88GiB
T=586.895.983.707.24GiB
T=1087.085.772.787.24GiB
T=2087.345.272.047.24GiB
T=5087.365.080.927.24GiB
T=10087.545.020.507.24GiB
+ +to substantially reduce the number of pixel embeddings, resulting in significant GPU memory usage. On the other hand, the superpixel strategy requires a considerable amount of runtime for the superpixel extraction. In comparison, our proposed partitioning strategy can achieve good results with higher efficiency and lower resource consumption. This makes it more suitable for applications on practical platforms. + +In-distribution segmentation performance on Cityscapes validation set: We report the results on the Cityscapes validation set in Table 6(Right). All methods use Mask2Former [10] as their backbones. We observe that most SOTA methods affect the in-distribution segmentation performance of Mask2Former, such as RbA (drop from 82.25 to 81.93). However, RWPM can alleviate this problem. Additionally, even for the SOTA segmentation method (Mask2Former), RWPM can further enhance its in-distribution segmentation performance. + +
Benchmark→ Road Anomaly
Strategy↓AP↑FPR95↓FPS↑Memory
Resize (0.9 full size)87.805.391.4123.82GIB
Resize (0.8 full size)87.245.442.1515.98GiB
Superpixel (30000 pixels)87.515.500.1911.46GiB
Superpixel (20000 pixels)87.035.980.326.31GiB
Partitioning (n = 2)87.345.272.047.24GiB
+ +Table 6: Ablation Tables: (Left) shows the results of different strategies for reducing the number of pixel embedding. We find that partitioning is the best. (Right) shows the segmentation performance on the Cityscapes validation set. RWPM can also improve the segmentation performance of in-distribution. + +
Benchmark→Cityscapes
Method↓mIoU↑
PEBAL-Mask [34]75.32
DenseHybrid-Mask [34]80.27
RbA [34]81.93
RbA+RWPM(Ours)82.16
Mask2Former [10]82.25
Mask2Former+RWPM(Ours)82.43
+ +# 4.3 Visualization of Pixel Embedding Distribution + +To investigate the effect of RWPM in the embedding space, we visualized the distribution of pixel embeddings by using t-SNE. We selected an image from the Road Anomaly dataset (as shown in Fig 3(a)). Due to the gravel road surface in this image, there is a significant domain gap between the road in this image and the roads in the training set, resulting in many false positives on the road (as shown in the PEBAL results in Fig 3(a)). Then, we randomly selected 300 inlier pixels and 150 outlier pixels from the red dashed box region of this image for visualization. Fig 3(b) and Fig 3(c) respectively show the distributions of these pixels in the space without and with applying RWPM. In Fig 3(b) and Fig 3(c), deep blue circles represent inlier pixels (road), light blue circles represent outlier pixels (boar), and X symbols represent the inlier class prototype (the vector of network weights of classifier). From Fig 3(b), we observe that without using RWPM, due to the influence of data diversity, the distribution of inliers and outliers is distorted. Consequently, it is difficult for an anomaly scoring function to distinguish between inlier and outlier pixels based on their logits. However, from Fig 3(c), we find that RWPM can significantly alleviate the problem. Specifically, the pixel embeddings of the same category form more compact clusters, enabling the anomaly scoring function to clearly discriminate between inliers and outliers based on the logit. This is consistent with the description in the toy sample in Fig 1. Note that, in the top left corner of Fig 3(c), some inlier points become very close to outlier points after applying RWPM. This is because these points are located in the boundaries of anomalous objects, making it very difficult to discriminate them. + +![](images/6dd5b6908bd5cb09c02c161f71c5dfd35a89c4d5e7069e3b5ab17ee3a1d54722.jpg) +(a) Sample Image + +![](images/e2e53be7e629c7892031c9d2987f959c7d9189311c0793d88e366dbe9e2abd20.jpg) +Fig. 3: The visualization of pixel embedding distribution. All embeddings are extracted from PEBAL model. We use cosine distance as the distance metric. We observed that RWPM can optimize the distribution of pixel embeddings in the space. + +![](images/ed945554d75c59edb7f825bfe123f491a8baa0a13b4cf05cac18a6fb9a3371d6.jpg) +(b) Distribution w/o RWPM +(c) Distribution with RWPM + +# 4.4 Qualitative Result + +Finally, to visually demonstrate the effectiveness of RWPM, we present qualitative results. Specifically, we compare the results of RbA on the Road Anomaly dataset without and with RWPM. The qualitative results are illustrated in Fig 4. + +Fig 4 shows that RWPM significantly improves the performance of RbA, reducing both the false positive rate and false negative rate. For example, in the first row, RbA generates some false positive results on the road and some false negative results on the manhole region (third column of the first row). However, after applying RWPM, these false positive/negative results are effectively eliminated (fourth column of the first row). Additionally, RWPM assists the RbA model to accurately detect and localize each anomalous object at the component level, such as more precise edge detection. This observation aligns with the results of the component-level metrics presented in Table 3. + +![](images/cc0f8c32e7c4447fe9c9c6d6ca48e6b2e3e8f9f23f94724ed4996462dff56e83.jpg) +Fig. 4: Qualitative results: The first column displays the input image, while the second column shows its corresponding ground truth. The third column displays the anomaly score results generated by RbA, and the fourth column presents the anomaly score results after applying RWPM to RbA. Yellow indicates high anomaly scores. + +# 5 Conclusions and Discussions + +In this paper, we propose a simple yet effective method, called Random Walk on Pixel Manifolds (RWPM), to improve the performance of existing anomaly segmentation methods. RWPM is a post-processing method that can be directly integrated into existing anomaly detection frameworks without requiring additional training or modifications to the network structure. RWPM mitigates the impact of data diversity on manifold structure by diffusing and updating pixel embeddings on pixel manifolds using random walk. This process enables anomaly scoring functions to obtain more accurate estimations of anomaly scores. Additionally, we propose Partial Random Walk strategy, which significantly improves the efficiency of RWPM. + +While RWPM achieves significant results in most cases, it is limited when the pre-trained model itself exhibits significant inaccuracies. To address this limitation, we plan to enhance the model's generalization ability during training to accurately detect a broader range of diverse and unknown anomalies. + +# Acknowledgements + +This work was supported by SenseTime Japan. + +# References + +1. Achanta, R., Shaji, A., Smith, K., Lucchi, A., Fua, P., Süsstrunk, S.: Slic superpixels compared to state-of-the-art superpixel methods. IEEE transactions on pattern analysis and machine intelligence 34(11), 2274-2282 (2012) +2. Ackermann, J., Sakaridis, C., Yu, F.: Maskomaly: Zero-shot mask anomaly segmentation. arXiv preprint arXiv:2305.16972 (2023) +3. Bai, S., Bai, X., Tian, Q., Latecki, L.J.: Regularized diffusion process for visual retrieval. In: Proceedings of the AAAI conference on artificial intelligence. vol. 31 (2017). https://doi.org/10.1609/aaai.v31i1.11216 +4. Bai, S., Zhou, Z., Wang, J., Bai, X., Jan Latecki, L., Tian, Q.: Ensemble diffusion for retrieval. In: Proceedings of the IEEE International conference on computer vision. pp. 774-783 (2017) +5. Blum, H., Sarlin, P.E., Nieto, J., Siegwart, R., Cadena, C.: The fishyscapes benchmark: Measuring blind spots in semantic segmentation. International Journal of Computer Vision 129, 3119-3135 (2021). https://doi.org/10.1007/s11263-021-01511-6 +6. Chan, R., Lis, K., Uhlemeyer, S., Blum, H., Honari, S., Siegwart, R., Fua, P., Salzmann, M., Rottmann, M.: Segmentmeifyoucan: A benchmark for anomaly segmentation. arXiv preprint arXiv:2104.14812 (2021) +7. Chan, R., Rottmann, M., Gottschalk, H.: Entropy maximization and meta classification for out-of-distribution detection in semantic segmentation. In: Proceedings of the ieee/cvf international conference on computer vision. pp. 5128-5137 (2021) +8. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence 40(4), 834-848 (2017) +9. Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 801-818 (2018). https://doi.org/10.1007/978-3-030-01234-2_49 +0. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290–1299 (2022). https://doi.org/10.1109/CVPR52688.2022.00135 +1. Cheng, B., Schwing, A., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems 34, 17864-17875 (2021) +2. Choi, H., Jeong, H., Choi, J.Y.: Balanced energy regularization loss for out-of-distribution detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15691-15700 (2023) +3. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3213-3223 (2016). https://doi.org/10.1109/CVPR.2016.350 + +14. Creusot, C., Munawar, A.: Real-time small obstacle detection on highways using compressive rbm road reconstruction. In: 2015 IEEE Intelligent Vehicles Symposium (IV). pp. 162-167. IEEE (2015) +15. Di Biase, G., Blum, H., Siegwart, R., Cadena, C.: Pixel-wise anomaly detection in complex driving scenes. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16918-16927 (2021). https://doi.org/10.1109/CVPR46437.2021.01664 +16. Gal, Y., Ghahramani, Z.: Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In: international conference on machine learning. pp. 1050-1059. PMLR (2016) +17. Grady, L.: Random walks for image segmentation. IEEE transactions on pattern analysis and machine intelligence 28(11), 1768-1783 (2006) +18. Grcic, M., Bevandić, P., Šegvic, S.: Densehybrid: Hybrid anomaly detection for dense open-set recognition. In: European Conference on Computer Vision. pp. 500-517. Springer (2022) +19. Grcic, M., Saric, J., Segvic, S.: On advantages of mask-level recognition for outlier-aware segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2936-2946 (2023) +20. Hendrycks, D., Basart, S., Mazeika, M., Zou, A., Kwon, J., Mostajabi, M., Steinhardt, J., Song, D.: Scaling out-of-distribution detection for real-world settings. arXiv preprint arXiv:1911.11132 (2019) +21. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. In: International Conference on Learning Representations (2017), https://openreview.net/forum?id=Hkg4TI9x1 +22. Hendrycks, D., Mazeika, M., Dietterich, T.: Deep anomaly detection with outlier exposure. arXiv preprint arXiv:1812.04606 (2018) +23. Huang, Z., Li, S.: From difference to similarity: A manifold ranking-based hyperspectral anomaly detection framework. IEEE Transactions on Geoscience and Remote Sensing 57(10), 8118-8130 (2019) +24. Iscen, A., Tolias, G., Avrithis, Y., Chum, O.: Mining on manifolds: Metric learning without labels. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7642-7651 (2018). https://doi.org/10.1109/CVPR.2018.00797 +25. Iscen, A., Tobias, G., Avrithis, Y., Furon, T., Chum, O.: Efficient diffusion on region manifolds: Recovering small objects with compact cnn representations. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2077-2086 (2017). https://doi.org/10.1109/CVPR.2017.105 +26. Jung, S., Lee, J., Gwak, D., Choi, S., Choo, J.: Standardized max logits: A simple yet effective approach for identifying unexpected road obstacles in urban-scene segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15425-15434 (2021) +27. Lakshminarayanan, B., Pritzel, A., Blundell, C.: Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information processing systems 30 (2017) +28. Liang, C., Wang, W., Miao, J., Yang, Y.: Gmmseg: Gaussian mixture based generative semantic segmentation models. Advances in Neural Information Processing Systems 35, 31360-31375 (2022) +29. Liang, S., Li, Y., Srikant, R.: Enhancing the reliability of out-of-distribution image detection in neural networks. In: International Conference on Learning Representations (2018), https://openreview.net/forum?id=H1VGkIxRZ + +30. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014). https://doi.org/10.1007/978-3-319-10602-1_48 +31. Lis, K., Nakka, K., Fua, P., Salzmann, M.: Detecting the unexpected via image resynthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2152-2161 (2019). https://doi.org/10.1109/ICCV.2019.00224 +32. Liu, Y., Ding, C., Tian, Y., Pang, G., Belagiannis, V., Reid, I., Carneiro, G.: Residual pattern learning for pixel-wise out-of-distribution detection in semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1151-1161 (2023) +33. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021) +34. Nayal, N., Yavuz, M., Henriques, J.F., Güney, F.: Rba: Segmenting unknown regions rejected by all. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 711-722 (2023) +35. Ohgushi, T., Horiguchi, K., Yamanaka, M.: Road obstacle detection method based on an autoencoder with semantic segmentation. In: Proceedings of the Asian Conference on Computer Vision (2020) +36. Rai, S.N., Cermelli, F., Fontanel, D., Masone, C., Caputo, B.: Unmasking anomalies in road-scene segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4037-4046 (2023) +37. Tian, Y., Liu, Y., Pang, G., Liu, F., Chen, Y., Carneiro, G.: Pixel-wise energy-biased abstention learning for anomaly segmentation on complex urban driving scenes. In: European Conference on Computer Vision. pp. 246-263. Springer (2022). https://doi.org/10.1007/978-3-031-19842-7_15 +38. Vojir, T., Šipka, T., Aljundi, R., Chumerin, N., Reino, D.O., Matas, J.: Road anomaly detection by partial image reconstruction with segmentation coupling. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15651-15660 (2021) +39. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems 34, 12077-12090 (2021) +40. Yang, F., Hinami, R., Matsui, Y., Ly, S., Satoh, S.: Efficient image retrieval via decoupling diffusion into online and offline processing. In: Proceedings of the AAAI conference on artificial intelligence. vol. 33, pp. 9087-9094 (2019). https://doi.org/10.1609/aaai.v33i01.33019087 +41. Yang, F., Wang, Z., Xiao, J., Satoh, S.: Mining on heterogeneous manifolds for zero-shot cross-modal image retrieval. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 12589-12596 (2020) +42. Zhou, D., Bousquet, O., Lal, T., Weston, J., Scholkopf, B.: Learning with local and global consistency. Advances in neural information processing systems 16 (2003) +43. Zhou, D., Weston, J., Gretton, A., Bousquet, O., Schölkopf, B.: Ranking on data manifolds. Advances in neural information processing systems 16 (2003) \ No newline at end of file diff --git a/2024/Random Walk on Pixel Manifolds for Anomaly Segmentation of Complex Driving Scenes/images.zip b/2024/Random Walk on Pixel Manifolds for Anomaly Segmentation of Complex Driving Scenes/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1f840a7488a4737a3f998c8347971f3550a677b6 --- /dev/null +++ b/2024/Random Walk on Pixel Manifolds for Anomaly Segmentation of Complex Driving Scenes/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:32c5f56d4aef4c95ed000e670e7d7c22ea15443b5be79065dd0c6e68f501549a +size 414989 diff --git a/2024/Random Walk on Pixel Manifolds for Anomaly Segmentation of Complex Driving Scenes/layout.json b/2024/Random Walk on Pixel Manifolds for Anomaly Segmentation of Complex Driving Scenes/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0ee3e55cff1995b9ae0837063cf93c19fa62ebef --- /dev/null +++ b/2024/Random Walk on Pixel Manifolds for Anomaly Segmentation of Complex Driving Scenes/layout.json @@ -0,0 +1,8438 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 111, + 473, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 111, + 473, + 148 + ], + "spans": [ + { + "bbox": [ + 143, + 111, + 473, + 148 + ], + "type": "text", + "content": "Random Walk on Pixel Manifolds for Anomaly Segmentation of Complex Driving Scenes" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 223, + 167, + 388, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 167, + 388, + 180 + ], + "spans": [ + { + "bbox": [ + 223, + 167, + 388, + 180 + ], + "type": "text", + "content": "Zelong Zeng1 and Kaname Tomite1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 230, + 190, + 383, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 190, + 383, + 213 + ], + "spans": [ + { + "bbox": [ + 230, + 190, + 383, + 213 + ], + "type": "text", + "content": "SenseTime Japan, Tokyo, Japan {zengzelong,tomite}@sensetime.jp" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 239, + 463, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 239, + 463, + 403 + ], + "spans": [ + { + "bbox": [ + 159, + 239, + 463, + 403 + ], + "type": "text", + "content": "Abstract. In anomaly segmentation for complex driving scenes, state-of-the-art approaches utilize anomaly scoring functions to calculate anomaly scores. For these functions, accurately predicting the logits of inlier classes for each pixel is crucial for precisely inferring the anomaly score. However, in real-world driving scenarios, the diversity of scenes often results in distorted manifolds of pixel embeddings in the space. This effect is not conducive to directly using the pixel embeddings for the logit prediction during inference, a concern overlooked by existing methods. To address this problem, we propose a novel method called Random Walk on Pixel Manifolds (RWPM). RWPM utilizes random walks to reveal the intrinsic relationships among pixels to refine the pixel embeddings. The refined pixel embeddings alleviate the distortion of manifolds, improving the accuracy of anomaly scores. Our extensive experiments show that RWPM consistently improve the performance of the existing anomaly segmentation methods and achieve the best results1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 414, + 455, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 414, + 455, + 437 + ], + "spans": [ + { + "bbox": [ + 159, + 414, + 455, + 437 + ], + "type": "text", + "content": "Keywords: Autonomous vehicles " + }, + { + "bbox": [ + 159, + 414, + 455, + 437 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 414, + 455, + 437 + ], + "type": "text", + "content": " Anomaly segmentation " + }, + { + "bbox": [ + 159, + 414, + 455, + 437 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 414, + 455, + 437 + ], + "type": "text", + "content": " Semantic segmentation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 456, + 230, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 230, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 230, + 468 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 479, + 482, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 598 + ], + "type": "text", + "content": "Semantic segmentation [9-11,13,39] plays a crucial role in the field of automated driving, which can aid the automated driving system in perceiving and understanding the surrounding environment, thereby facilitating the system make accurate decision. Typically, the semantic segmentation models are trained on closed-datasets with fixed semantic categories, showing good recognition capabilities for inlier classes (e.g., roads, vehicles, etc.). However, these models often fail to effectively detect potential outlier class objects (e.g., animals, fallen tires, etc.) in real-world scenarios. Due to the critical importance of correctly detecting outlier class targets for the safety of autonomous driving, there is a growing focus on addressing this task, namely, anomaly segmentation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 600, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 600, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 600, + 482, + 647 + ], + "type": "text", + "content": "The standard paradigm for anomaly segmentation is assigning an anomaly score to each pixel of an input image, where a higher score indicates a greater probability that the pixel belongs to an outlier. Thus, the way used for estimating the anomaly score is crucial for anomaly segmentation methods. In" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 392, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 392, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 392, + 666 + ], + "type": "text", + "content": "1 Code is available at: https://github.com/ZelongZeng/RwPM" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 113, + 239, + 185 + ], + "blocks": [ + { + "bbox": [ + 151, + 113, + 239, + 185 + ], + "lines": [ + { + "bbox": [ + 151, + 113, + 239, + 185 + ], + "spans": [ + { + "bbox": [ + 151, + 113, + 239, + 185 + ], + "type": "image", + "image_path": "66465a8b974307d9e4d39db93e496fd578b716b3a4fdf1ea9f19da4f91b66e54.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 185, + 230, + 191 + ], + "lines": [ + { + "bbox": [ + 154, + 185, + 230, + 191 + ], + "spans": [ + { + "bbox": [ + 154, + 185, + 230, + 191 + ], + "type": "text", + "content": "(i) An ideal manifold structure" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 216, + 196, + 255, + 203 + ], + "lines": [ + { + "bbox": [ + 216, + 196, + 255, + 203 + ], + "spans": [ + { + "bbox": [ + 216, + 196, + 255, + 203 + ], + "type": "text", + "content": "Inlier pixel" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 130, + 215, + 482, + 403 + ], + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 403 + ], + "type": "text", + "content": "Fig. 1: A toy example about manifolds of pixel embeddings. The dashed circles indicate regions with high predicted logit to the corresponding inlier class. For the pixel in these regions, its anomaly score is small. The prototype is the vector of network weights of the classifier for the corresponding inlier class. (i) An ideal manifold structure. All inlier pixels are within the corresponding high logit region, whereas all outlier pixels are outside the areas. The anomaly score derived from the anomaly scoring function effectively discriminates between them. (ii) The manifolds are affected by the diversity of data, causing some inlier pixels to deviate from the corresponding high logit region, while some outlier pixels approach the regions. This results in false positives/negatives in the anomaly scores. However, pixels of the same class are still in the same manifold, indicating that the manifold structure can be utilized to reveal the intrinsic relationships between pixels. (iii) Our RWPM utilizes random walks to capture the manifold structure to diffuse and update pixel embeddings. The embeddings within the same manifold tend to become more similar after the updating. Fig 3 and Fig 4 respectively present the visualization of embedding distributions and the qualitative results of real examples, consistent with the description of this toy example." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 253, + 113, + 359, + 184 + ], + "blocks": [ + { + "bbox": [ + 253, + 113, + 359, + 184 + ], + "lines": [ + { + "bbox": [ + 253, + 113, + 359, + 184 + ], + "spans": [ + { + "bbox": [ + 253, + 113, + 359, + 184 + ], + "type": "image", + "image_path": "e5dd3250608f82a6b1a86386c513f141d9085137879f723e452084cd3cb51639.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 185, + 346, + 192 + ], + "lines": [ + { + "bbox": [ + 260, + 185, + 346, + 192 + ], + "spans": [ + { + "bbox": [ + 260, + 185, + 346, + 192 + ], + "type": "text", + "content": "(ii) A distorted manifold structure" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 268, + 195, + 337, + 204 + ], + "lines": [ + { + "bbox": [ + 268, + 195, + 337, + 204 + ], + "spans": [ + { + "bbox": [ + 268, + 195, + 337, + 204 + ], + "type": "inline_equation", + "content": "\\times \\times" + }, + { + "bbox": [ + 268, + 195, + 337, + 204 + ], + "type": "text", + "content": " Inlier class prototype" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 372, + 113, + 462, + 185 + ], + "blocks": [ + { + "bbox": [ + 372, + 113, + 462, + 185 + ], + "lines": [ + { + "bbox": [ + 372, + 113, + 462, + 185 + ], + "spans": [ + { + "bbox": [ + 372, + 113, + 462, + 185 + ], + "type": "image", + "image_path": "7eb03a33c34f0e318dc50b70b00fe594d25cdc69d3ad192468b00ec5e72d85c5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 377, + 185, + 459, + 191 + ], + "lines": [ + { + "bbox": [ + 377, + 185, + 459, + 191 + ], + "spans": [ + { + "bbox": [ + 377, + 185, + 459, + 191 + ], + "type": "text", + "content": "(iii) A refined manifold structure" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 345, + 196, + 385, + 203 + ], + "lines": [ + { + "bbox": [ + 345, + 196, + 385, + 203 + ], + "spans": [ + { + "bbox": [ + 345, + 196, + 385, + 203 + ], + "type": "text", + "content": "Outlier pixel" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 426, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 617 + ], + "type": "text", + "content": "early approaches, the anomaly score is often given by estimating the prediction uncertainty [16, 21, 27, 29], or by comparing the discrepancy between the original image and its reconstructed image generated based on the semantic segmentation predictions [14, 15, 31, 35]. Different from early methods, recent methods [12, 19, 32, 34, 36, 37] adopt Outlier Exposure (OE) strategy [22], which directly learn a pixel-level anomaly class to detect anomalous pixels that deviate from inlier classes. Specifically, these methods propose anomaly scoring functions to infer the anomaly scores and, based on the functions, leverage auxiliary outlier data to fine-tune a semantic segmentation model. The auxiliary outlier data is typically generated by extracting outlier-class objects from other datasets (e.g., COCO [30]) and incorporating them into driving scene images [34, 36, 37]. Furthermore, some of them [19, 34, 36] adopt mask-based segmentation networks instead of pixel-based segmentation networks, achieving state-of-the-art (SOTA) performance in anomaly segmentation. In this paper, we concentrate on the current methods using OE strategy. These methods indicate superior performance to the early methods using prediction uncertainty or image reconstruction." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "The anomaly scoring functions in OE methods infer anomaly scores based on the inlier class logit predictions from a semantic segmentation model. Specifically, the logic of these functions can be represented as follows: when the predicted logits of a pixel embedding are small across all inlier classes, indicating that the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "Z. Zeng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 308 + ], + "type": "text", + "content": "pixel is dissimilar to all inlier classes and is an outlier with high probability, the function yields a higher anomaly score. Therefore, accurately measuring the logits of each pixel is crucial for precisely evaluating the anomaly score. However, in practical open-world driving scenarios, various environmental conditions (such as lighting, road materials, etc.) and a diverse range of unknown class objects are present. During the inference phase, the diversity of driving scenes often results in distorted manifolds of pixel embedding in the embedding space, causing pixel embeddings to deviate from the ideal distribution. This phenomenon has been validated in previous works of metric learning [3,25,40,41]. Since the logits are calculated from the inner product (i.e., cosine distance) of the pixel embeddings and the vectors of classifiers' network weights, the distorted manifolds are not conducive to directly using the pixel embeddings for the logit prediction. However, current anomaly segmentation methods ignore this issue and directly use the pixel embeddings for the logit prediction in the inference phase, so the results often fail to accurately reflect the anomaly score for the corresponding pixels, causing high false positive/negative rate (see Fig. 1(i) and Fig. 1(ii))." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 310, + 482, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 482, + 622 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 482, + 622 + ], + "type": "text", + "content": "In this paper, we propose a novel method, called Random Walk on Pixel Manifolds (RWPM), to alleviate the effect of data diversity on the manifolds. RWPM utilizes random walks to capture the manifold structure of the pixel embeddings to diffuse and update the embedding of each pixel. Since the structure of data manifolds implicitly contains the intrinsic relationship among data points (see Fig. 1(ii)) [3,24,40,43], using random walks to measure the similarity between pixels on manifolds (i.e., distance on manifolds rather than Euclidean distance) better reflects the intrinsic relationships between pixels. Consequently, diffusing and updating pixel embeddings based on the similarity on manifolds results in high similarity among pixels on the same manifold and low similarity among pixels on different manifolds. In the embedding space, this manifests as pixel embeddings on the same manifold forming more compact clusters, thereby mitigating the distortion effect of data diversity on the manifolds, which facilitates the logit prediction (see Fig. 1(iii)). Furthermore, since the typically large number of pixels in an image, utilizing all pixels for the random walks results in significant memory and computational overhead, which damages the deployment and real-time execution of our RWPM. To address this issue, we introduce the Partial Random Walk strategy for RWPM, reducing memory consumption and enhancing the operational efficiency of RWPM. Note that, RWPM is proposed to use in the inference phase, thus no any additional training is required. Additionally, RWPM can be directly integrated into existing anomaly segmentation framework without the need for any extra modification to the network structure. Extensive experiments in various anomaly segmentation benchmarks of road scenes, namely Fishyscapes [5], Road Anomaly [31], and SMIYC [6], demonstrate that RWPM consistently improves the performance of existing methods, achieving SOTA results. To summarise, our contributions are the following:" + } + ] + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "RWPM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 138, + 641, + 481, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 641, + 481, + 666 + ], + "type": "text", + "content": "- We introduce that the diversity of driving scenes results in distorted manifolds of pixel embedding, thereby affecting the accuracy of anomaly scoring" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 147, + 116, + 480, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 480, + 138 + ], + "type": "text", + "content": "functions in inferring anomaly scores. This problem is ignored by existing anomaly segmentation methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 479, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 479, + 185 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 479, + 185 + ], + "type": "text", + "content": "- We propose the Random Walk on Pixel Manifolds (RWPM) that utilizes random walks to reveal the intrinsic relationships among pixels to refine the pixel embeddings. This way alleviate the effect on manifolds caused by the driving scene diversity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 186, + 479, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 186, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 138, + 186, + 479, + 209 + ], + "type": "text", + "content": "- We propose the Partial Random Walk strategy for RWPM to reduce the memory consumption and improve the operational efficiency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 224, + 237, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 237, + 238 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 237, + 238 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 247, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 668 + ], + "type": "text", + "content": "Anomaly Segmentation. Anomaly segmentation aims to detect the outlier class object in some specific scenes, such as complex driving scenes [6, 13, 31]. Existing anomaly segmentation can be broadly divided into three categories: (a) Uncertainty-based methods, (b) Reconstruction-based methods and (c) Outlier Exposure methods. Uncertainty-based methods assume that outlier samples result in low-confidence predictions. Based on this assumption, they focus on measuring the pixel-wise anomaly score by estimating the prediction uncertainty through various ways, such as ensembles [27], Bayesian estimation [16], maximum softmax probability [21, 29], and logit [20, 26]. Since segmentation models are trained on close-set, they may misclassify unseen classes with high confidence, leading uncertainty-based methods fail to detect anomalies. Reconstruction-based methods detect anomaly pixel by comparing the discrepancy between the original image and its reconstructed image generated based on the semantic segmentation predictions [14, 15, 31, 35]. These methods are challenged by the accuracy of segmentation prediction and the difficulties of training. Recent methods [7, 12, 19, 34, 36, 37] introduce the Outlier Exposure (OE) strategy [22] that utilizes auxiliary outlier data and anomaly scoring functions to learn models. The auxiliary outlier data is the image from other datasets (e.g., COCO [30]), or cutting the outlier objects from other datasets and pasting them into the inlier scenes, and the anomaly scoring functions are used to calculate the anomaly score for each pixel. Compared with previous works, OE strategy can reach better performance. More recently, some OE methods [19, 34, 36] adopt mask-based segmentation networks [10, 11], instead of pixel-based segmentation networks [8, 9], as their backbones. Based on the powerful capability of mask-based networks in segmentation tasks, these methods significantly reduces false detection rates, achieving SOTA results. In this paper, we only concentrate on the OE methods. The anomaly scoring functions employed in OE methods are based on the predicted logits of inlier classes. However, during the inference phase, the diversity of driving scenes results in distorted manifolds of pixel embeddings. As a result, the scoring functions may fail to use the embeddings to accurately calculate logits, thereby affects the accuracy of measuring anomaly scores. Current OE methods ignore this problem. Our proposed RWPM addresses this problem by utilizing random walks on pixel manifolds to reveal the intrinsic relationships among pixels and refine the pixel embeddings. To the best of our knowledge, RWPM is the first work to propose and resolve this problem." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "Z. Zeng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 404 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 404 + ], + "type": "text", + "content": "Mining on Data Manifolds. Mining on data manifolds, also known as diffusion process, is a well-established technique commonly used in retrieval tasks [3, 4, 25, 40, 43]. These works utilize random walks on data manifolds to capture the data manifold structure for revealing the intrinsic relationship between data points. The data manifold is constructed by a graph, where each node represents a sample and each edge presents the connection between a node and its neighborhoods, weighted proportionally according to their distance. For retrieval tasks, the images of the same object may be shown in different conditions (e.g., lighting and angles). This diversity often leads to the manifold structure is not conducive to accurately ranking using distance metrics directly [25, 40]. Thus, diffusion process is used to diffuse the state (typically initialized to 1) of a query image on the data manifolds. If a sample is on the same data manifold with the query (the distance on manifolds is close), then more energy of the state will flow to that sample. Therefore, the final state of all samples can be seen as the accurate ranking scores. Moreover, diffusion process is also adopted in semi-supervised learning [24, 42] to propagate the label of labeled samples to unlabeled samples. Similarly, some unsupervised segmentation works [17, 23] utilize diffusion process to propagate the labels of manually annotated pixels to the unlabeled pixels. These existing works utilize random walks on data manifolds for ranking or label propagation. Different from these works, our target is to solve the problem occurring in pixel manifolds for anomaly scoring functions. We propose innovatively using random walks to diffuse and update pixel embeddings, eliminating the effect of distorted manifolds directly at the embedding-level, thereby improving the accuracy of subsequent processing." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 420, + 261, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 261, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 261, + 434 + ], + "type": "text", + "content": "3 Proposed Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 443, + 228, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 443, + 228, + 455 + ], + "spans": [ + { + "bbox": [ + 132, + 443, + 228, + 455 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "text", + "content": "We denote an image space " + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{X} \\subset \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "text", + "content": " are the height and width, respectively. During the inference phase, the goal for an anomaly segmentation model " + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "text", + "content": " is to map an input image " + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathcal{X}" + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "text", + "content": " to an anomaly space, i.e., " + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "inline_equation", + "content": "f: \\pmb{x} \\mapsto \\mathcal{Z} \\subset \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "text", + "content": ". Specifically, an anomaly segmentation model typically consists of the backbone encoder-decoder " + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "text", + "content": " and the classifier " + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "text", + "content": ". For an input image " + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "text", + "content": ", firstly, its pixel embedding map is encoded as " + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\pmb{p} = F(\\pmb{x}) \\in \\mathbb{R}^{d \\times H \\times W}" + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "text", + "content": " is the dimension of the pixel embedding. Then, its logit map " + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\pmb{l}" + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "text", + "content": " is predicted by " + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\pmb{l} = C(F(\\pmb{x})) \\in \\mathbb{R}^{K \\times H \\times W}" + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "text", + "content": " is the number of inlier classes. Note that, depending on the backbone network, the size of the embedding map and logit map may sometimes be smaller than the original image. For the sake of convenience in this paper, we do not make a special distinction. Finally, a scoring function is used to calculate the anomaly score map " + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\pmb{z} \\in \\mathcal{Z}" + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "text", + "content": " based on the logit map " + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\pmb{l}" + }, + { + "bbox": [ + 130, + 461, + 482, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "For anomaly scoring functions, their logic can be represented as follows: when the predicted logits of a pixel embedding are small across all inlier classes, indicating that the pixel is dissimilar to all inlier classes and is likely an outlier, the function generates a higher anomaly score. For example, let's define " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "l_{h,w}(k)" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 90, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 90, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 416, + 90, + 447, + 101 + ], + "type": "text", + "content": "RWPM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 115, + 474, + 200 + ], + "blocks": [ + { + "bbox": [ + 141, + 115, + 474, + 200 + ], + "lines": [ + { + "bbox": [ + 141, + 115, + 474, + 200 + ], + "spans": [ + { + "bbox": [ + 141, + 115, + 474, + 200 + ], + "type": "image", + "image_path": "0cc1a045b91634a7ad204f533266b3087e3c813bfb6a897fc97866f134f6e246.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 206, + 482, + 307 + ], + "lines": [ + { + "bbox": [ + 130, + 206, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 206, + 482, + 307 + ], + "type": "text", + "content": "Fig. 2: Overview. This figure illustrates the application of RWPM to an existing anomaly segmentation framework during the inference phase. The red dashed box highlights the RWPM part. First, the encoder-decoder is used to extract the pixel embedding map of the input image. The pixel embedding map are subsequently partitioned into " + }, + { + "bbox": [ + 130, + 206, + 482, + 307 + ], + "type": "inline_equation", + "content": "n^2" + }, + { + "bbox": [ + 130, + 206, + 482, + 307 + ], + "type": "text", + "content": " sub-maps. For each sub-map, we update its pixel embeddings by using random walks to obtain a refined sub-map. Finally, the refined sub-maps are concatenated to form the refined embedding map, which are then input into the subsequent network structure. Notably, RWPM can be directly integrated into existing frameworks without requiring extra training or changes to the network structure." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 330, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 482, + 392 + ], + "type": "text", + "content": "is the logit of a pixel " + }, + { + "bbox": [ + 130, + 330, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{h,w}" + }, + { + "bbox": [ + 130, + 330, + 482, + 392 + ], + "type": "text", + "content": " for class " + }, + { + "bbox": [ + 130, + 330, + 482, + 392 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 330, + 482, + 392 + ], + "type": "text", + "content": ". The scoring function in PEBAL [37] is " + }, + { + "bbox": [ + 130, + 330, + 482, + 392 + ], + "type": "inline_equation", + "content": "E_{h,w} = -\\log \\sum_{k \\in \\{1, \\dots, K\\}} \\exp(\\pmb{l}_{h,w}(k))" + }, + { + "bbox": [ + 130, + 330, + 482, + 392 + ], + "type": "text", + "content": ". Obviously, when the anomaly score " + }, + { + "bbox": [ + 130, + 330, + 482, + 392 + ], + "type": "inline_equation", + "content": "E_{h,w}" + }, + { + "bbox": [ + 130, + 330, + 482, + 392 + ], + "type": "text", + "content": " is large, it indicates that the logits of the pixel across all inlier classes are small, i.e., the pixel " + }, + { + "bbox": [ + 130, + 330, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{h,w}" + }, + { + "bbox": [ + 130, + 330, + 482, + 392 + ], + "type": "text", + "content": " is dissimilar from all inlier classes (see Section S.2 of the supplementary material for the analysis of other anomaly scoring functions)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 408, + 263, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 263, + 421 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 263, + 421 + ], + "type": "text", + "content": "3.2 Graph Construction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 429, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 482, + 525 + ], + "type": "text", + "content": "We begin by constructing a graph to represent the manifolds of pixel embeddings. In our approach, we utilize an affinity matrix based on cosine similarity to characterize the relationships between pixels. Specifically, we first use the encoder-decoder component of an anomaly segmentation model to extract the pixel embeddings " + }, + { + "bbox": [ + 130, + 429, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\pmb{p} \\in \\mathbb{R}^{d \\times H \\times W}" + }, + { + "bbox": [ + 130, + 429, + 482, + 525 + ], + "type": "text", + "content": ". For ease of description and computation, we reshape the " + }, + { + "bbox": [ + 130, + 429, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 130, + 429, + 482, + 525 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 429, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\pmb{p}^r \\in \\mathbb{R}^{HW \\times d}" + }, + { + "bbox": [ + 130, + 429, + 482, + 525 + ], + "type": "text", + "content": ", where each row represents a pixel's embedding. We then normalize each row of " + }, + { + "bbox": [ + 130, + 429, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\pmb{p}^r" + }, + { + "bbox": [ + 130, + 429, + 482, + 525 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 130, + 429, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\widehat{\\pmb{p}}^r" + }, + { + "bbox": [ + 130, + 429, + 482, + 525 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 429, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\widehat{\\pmb{p}}_i^r = \\pmb{p}_i^r / \\| \\pmb{p}_i^r \\|" + }, + { + "bbox": [ + 130, + 429, + 482, + 525 + ], + "type": "text", + "content": ". The affinity matrix is defined as " + }, + { + "bbox": [ + 130, + 429, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{HW \\times HW}" + }, + { + "bbox": [ + 130, + 429, + 482, + 525 + ], + "type": "text", + "content": ", where each component is obtained by:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 252, + 533, + 481, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 533, + 481, + 563 + ], + "spans": [ + { + "bbox": [ + 252, + 533, + 481, + 563 + ], + "type": "interline_equation", + "content": "\\mathbf {W} _ {i j} = \\left\\{ \\begin{array}{c c} \\langle \\hat {\\pmb {p}} _ {i} ^ {r}, \\hat {\\pmb {p}} _ {j} ^ {r} \\rangle & i \\neq j \\\\ 0 & i = j \\end{array} , \\right. \\tag {1}", + "image_path": "935411a96720712dcbf5b2158734968d55d9b822daccddd590b9f1b51e9bdedb.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 569, + 481, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 481, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 569, + 481, + 593 + ], + "type": "inline_equation", + "content": "\\langle \\cdot, \\cdot \\rangle" + }, + { + "bbox": [ + 130, + 569, + 481, + 593 + ], + "type": "text", + "content": " denotes inner product, " + }, + { + "bbox": [ + 130, + 569, + 481, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{ii} = 0" + }, + { + "bbox": [ + 130, + 569, + 481, + 593 + ], + "type": "text", + "content": " is guaranteed to circumvent self-loops in the graph represented by " + }, + { + "bbox": [ + 130, + 569, + 481, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 130, + 569, + 481, + 593 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "To mitigate the influence of noise, locally constrained random walk methods [24, 40, 41] adopt local constraints to construct the graph, i.e., each point is only connected to its locally nearest neighbors in the graph. Determining the nearest neighbors for each point traditionally involves applying nearest neighbor search algorithms, such as the " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "-NN algorithm. However, in anomaly segmentation tasks, the size of the graph is large and the graph is constructed online," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "Z. Zeng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "making the search algorithms impractical. To benefit from the local constraints, we utilize a softmax function to normalize each row of " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " to obtain locally constrained graph " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 234, + 161, + 481, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 161, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 234, + 161, + 481, + 192 + ], + "type": "interline_equation", + "content": "\\mathbf {S} _ {i j} = \\left\\{ \\begin{array}{c c} \\frac {\\exp \\left(\\mathbf {W} _ {i j} / \\tau\\right)}{\\sum_ {i = 1 , i \\neq j} ^ {H W} \\exp \\left(\\mathbf {W} _ {i j} / \\tau\\right)} & i \\neq j \\\\ 0 & i = j \\end{array} , \\right. \\tag {2}", + "image_path": "fbddca577e9cf127617cbc1456139ada5da5ea3fb4624945a6db128bbd54b714.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 201, + 480, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 480, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 480, + 260 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 201, + 480, + 260 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 201, + 480, + 260 + ], + "type": "text", + "content": " is a temperature hyper-parameter, we set its value less than 1.0 to remove the effect of non-neighbor points. Compared to typical search algorithms, Eq. 2 can impose local constraints more efficiently and effectively (see Section S.4 of the supplementary material), and keep the scale of embedding in subsequent random walks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 278, + 275, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 275, + 290 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 275, + 290 + ], + "type": "text", + "content": "3.3 Random Walk Process" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 299, + 479, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 299, + 479, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 299, + 479, + 346 + ], + "type": "text", + "content": "Now we have constructed the graph " + }, + { + "bbox": [ + 130, + 299, + 479, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 130, + 299, + 479, + 346 + ], + "type": "text", + "content": ", we can start the random walks. Different from previous mining on data manifolds works utilizing random walks to propagate the states of known samples to other unknown samples, we use a random walk to propagate the embedding between pixels on manifolds." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 346, + 480, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 480, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 480, + 381 + ], + "type": "text", + "content": "We first define " + }, + { + "bbox": [ + 130, + 346, + 480, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^t\\in \\mathbb{R}^{HW\\times d}" + }, + { + "bbox": [ + 130, + 346, + 480, + 381 + ], + "type": "text", + "content": " as the updated pixel embeddings in the " + }, + { + "bbox": [ + 130, + 346, + 480, + 381 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 346, + 480, + 381 + ], + "type": "text", + "content": " -th step of the random walks. " + }, + { + "bbox": [ + 130, + 346, + 480, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^0" + }, + { + "bbox": [ + 130, + 346, + 480, + 381 + ], + "type": "text", + "content": " is initialized as " + }, + { + "bbox": [ + 130, + 346, + 480, + 381 + ], + "type": "inline_equation", + "content": "\\pmb{p}^r" + }, + { + "bbox": [ + 130, + 346, + 480, + 381 + ], + "type": "text", + "content": ". Then, each iteration step of the random walks is described as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 222, + 392, + 480, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 392, + 480, + 405 + ], + "spans": [ + { + "bbox": [ + 222, + 392, + 480, + 405 + ], + "type": "interline_equation", + "content": "\\mathbf {m} ^ {t + 1} = \\alpha \\mathbf {S} \\mathbf {m} ^ {t} + (1 - \\alpha) \\mathbf {m} ^ {0}, \\alpha \\in (0, 1), \\tag {3}", + "image_path": "c48549b9b563215db2da04a8600e068f7ebec66ae63ec491ff1b30972105f2a1.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 415, + 480, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 480, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 415, + 480, + 510 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 415, + 480, + 510 + ], + "type": "text", + "content": " is the probability continuing the random walk from the current state " + }, + { + "bbox": [ + 130, + 415, + 480, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^t" + }, + { + "bbox": [ + 130, + 415, + 480, + 510 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 130, + 415, + 480, + 510 + ], + "type": "inline_equation", + "content": "(1 - \\alpha)" + }, + { + "bbox": [ + 130, + 415, + 480, + 510 + ], + "type": "text", + "content": " represents the probability of restarting from the initial state " + }, + { + "bbox": [ + 130, + 415, + 480, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^0" + }, + { + "bbox": [ + 130, + 415, + 480, + 510 + ], + "type": "text", + "content": ". At each iteration, guided by the manifold structure, each pixel diffuses its embedding to other pixels while also receiving embedding from other pixels for updating its own embedding. The updated pixel embedding can be seen as a weighted ensemble of other pixel embeddings. If pixels are similar on manifolds, their embeddings will gradually become similar. Mathematically speaking, the whole iteration process can be written as a closed-from as follows [43]:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 241, + 521, + 480, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 521, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 241, + 521, + 480, + 533 + ], + "type": "interline_equation", + "content": "\\mathbf {m} ^ {\\infty} = (1 - \\alpha) (\\mathbf {I} - \\alpha \\mathbf {S}) ^ {- 1} \\mathbf {m} ^ {0}, \\tag {4}", + "image_path": "834167b3f24ec2ade27879bc221ac69d3b7488b51753c4257069610d4ea87322.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 542, + 479, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 542, + 479, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 542, + 479, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 542, + 479, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{HW \\times HW}" + }, + { + "bbox": [ + 130, + 542, + 479, + 567 + ], + "type": "text", + "content": " is an identity matrix. Finally, " + }, + { + "bbox": [ + 130, + 542, + 479, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^{\\infty}" + }, + { + "bbox": [ + 130, + 542, + 479, + 567 + ], + "type": "text", + "content": " is used as the refined pixel embeddings for the subsequent anomaly segmentation process." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 585, + 271, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 585, + 271, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 585, + 271, + 597 + ], + "type": "text", + "content": "3.4 Partial Random Walk" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "content": "When the size of the image is large, the size of the relevant matrices used for the calculation of random walks becomes extremely huge as well. For instance, when the image size is " + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "inline_equation", + "content": "(512 \\times 1024)" + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "content": ", the size of matrix " + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "content": " reaches " + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "inline_equation", + "content": "(524288 \\times 524288)" + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "content": ". This significantly increases the memory consumption and computational time of RWPM, which is not conducive to online operation. Unfortunately, in the field" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 90, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 90, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 90, + 447, + 100 + ], + "type": "text", + "content": "RWPM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "of autonomous driving, the image size of driving scenes is generally large. To address this problem, we propose Partial Random Walk, which consists of the following two aspects: embedding map partitioning and limited iteration." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 278 + ], + "type": "text", + "content": "To improve efficiency, we first propose embedding map partitioning strategy to reduce the size of the pixel embedding map used for random walks. Specifically, for a pixel embedding map " + }, + { + "bbox": [ + 130, + 152, + 480, + 278 + ], + "type": "inline_equation", + "content": "\\pmb{p} \\in \\mathbb{R}^{d \\times H \\times W}" + }, + { + "bbox": [ + 130, + 152, + 480, + 278 + ], + "type": "text", + "content": ", we equally divide it into " + }, + { + "bbox": [ + 130, + 152, + 480, + 278 + ], + "type": "inline_equation", + "content": "n \\times n" + }, + { + "bbox": [ + 130, + 152, + 480, + 278 + ], + "type": "text", + "content": " sub-maps " + }, + { + "bbox": [ + 130, + 152, + 480, + 278 + ], + "type": "inline_equation", + "content": "\\left\\{\\pmb{p}^{(1)}, \\dots, \\pmb{p}^{(n^2)}\\right\\} \\in \\mathbb{R}^{d \\times \\frac{H}{n} \\times \\frac{W}{n}}" + }, + { + "bbox": [ + 130, + 152, + 480, + 278 + ], + "type": "text", + "content": ". Then, for each sub-map in " + }, + { + "bbox": [ + 130, + 152, + 480, + 278 + ], + "type": "inline_equation", + "content": "\\left\\{\\pmb{p}^{(1)}, \\dots, \\pmb{p}^{(n^2)}\\right\\}" + }, + { + "bbox": [ + 130, + 152, + 480, + 278 + ], + "type": "text", + "content": ", we follow the steps of Eq. 1 and Eq. 2 to construct their sub-graphs " + }, + { + "bbox": [ + 130, + 152, + 480, + 278 + ], + "type": "inline_equation", + "content": "\\left\\{\\pmb{S}_{\\pmb{p}^{(1)}}, \\dots, \\pmb{S}_{\\pmb{p}^{(n^2)}}\\right\\} \\in \\mathbb{R}^{d \\times \\frac{HW}{n^2} \\times \\frac{HW}{n^2}}" + }, + { + "bbox": [ + 130, + 152, + 480, + 278 + ], + "type": "text", + "content": ". We also attempted other methods to reduce the number of pixel embeddings, such as resize or superpixel [1]. Experimental results indicate that embedding map partitioning is the optimal solution among them (see Section 4.2)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "text", + "content": "Next, for each sub-map " + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "inline_equation", + "content": "\\pmb{p}^{(i)}" + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "text", + "content": ", we perform random walks with the limited iteration strategy in their corresponding sub-graphs " + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "inline_equation", + "content": "S_{\\pmb{p}^{(i)}}" + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "text", + "content": ". That is, we use Eq. 3 for a small number of iterations instead of using the closed-form Eq. 4. We found that a few iterations by using Eq. 3 are enough to yield high quality results. The time complexity of Eq. 3 with " + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "text", + "content": " iterations for each sub-map is " + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "inline_equation", + "content": "O(Td(\\frac{HW}{n^2})^2)" + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "text", + "content": ", while the time complexity of Eq. 4 is " + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "inline_equation", + "content": "O((\\frac{HW}{n^2})^3) + O(d(\\frac{HW}{n^2})^2)" + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "inline_equation", + "content": "Td" + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "text", + "content": " is much smaller than " + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "inline_equation", + "content": "\\frac{HW}{n^2}" + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "text", + "content": ", the limited iteration strategy can significantly improve the efficiency of RWPM. After the random walks, we obtain the refined pixel embeddings sub-map " + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_{\\pmb{p}^{(i)}}^T" + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "text", + "content": " for each sub-map " + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "inline_equation", + "content": "\\pmb{p}^{(i)}" + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "text", + "content": ". Finally, we concatenate the refined sub-maps based on their positions to obtain the final refined pixel embedding map. Additionally, when " + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "text", + "content": " is large (observed experimentally as " + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "inline_equation", + "content": "n > 2" + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "text", + "content": "), due to spatial constraints, the baseline values of anomaly scores outputted on each sub-map may vary. Hence, a calibration is performed. For adjacent refined sub-maps " + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "inline_equation", + "content": "\\pmb{p}^{(i)'}" + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "inline_equation", + "content": "\\pmb{p}^{(j)'}" + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "text", + "content": ", we compute the average anomaly scores " + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "text", + "content": " of all pixels on their adjacent edges, respectively. Subsequently, using " + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "inline_equation", + "content": "\\pmb{p}^{(i)'}" + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "text", + "content": " as the reference, we calculate the ratio " + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "inline_equation", + "content": "\\frac{I}{J}" + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "text", + "content": ", and then multiply all anomaly scores on sub-map " + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "inline_equation", + "content": "\\pmb{p}^{(j)'}" + }, + { + "bbox": [ + 130, + 279, + 481, + 486 + ], + "type": "text", + "content": " by this ratio to achieve the calibration." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 487, + 372, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 487, + 372, + 499 + ], + "spans": [ + { + "bbox": [ + 146, + 487, + 372, + 499 + ], + "type": "text", + "content": "Fig. 2 shows the workflow of the proposed RWPM." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 519, + 224, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 519, + 224, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 519, + 224, + 533 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Dataset: We evaluate our approach on three standard benchmark datasets for anomaly segmentation, i.e., Fishyscapes Lost&Found [5], Road Anomaly [31] and Segment Me If You Can (SMIYC) [6]. Fishyscapes Lost&Found has 100 validation images. The domain of this dataset is similar to that of Cityscapes [13] dataset. Road Anomaly dataset contains 60 images of real-world road scenes. Unlike the Fishyscapes Lost&Found dataset, the images in Road Anomaly contain various classes and sizes of anomaly objects, making it more challenging. SMIYC consists of two subsets, Anomaly track and Obstacle track. The Anomaly track focuses on detecting large anomalous objects, while the Obstacle track emphasizes detecting small anomalous objects on the road. Anomaly track contains 10" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "Z. Zeng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "validation images and 100 online test images, while Obstacle track contains 30 validation images and 327 online test images. The Road Anomaly and SMIYC datasets both exhibit high domain shifts compared to the Cityscapes dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 481, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 481, + 247 + ], + "type": "text", + "content": "Evaluation Metrics: Following the previous works, we evaluate our method on the standard metrics for anomaly segmentation. On the Fishyscapes Lost&Found and Road Anomaly, we adopt the area under receiver operating characteristics (AUROC), average precision (AP), and the false positive rate at a true positive rate of " + }, + { + "bbox": [ + 130, + 152, + 481, + 247 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 152, + 481, + 247 + ], + "type": "text", + "content": " (FPR95). On the SMIYC, in addition to reporting per-pixel metrics such as AP and FPR95, we also calculate results for component-level metrics [6] such as the component-wise intersection over union (sIoU), positive predictive value (PPV), and mean F1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 248, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 477 + ], + "type": "text", + "content": "Implementation Details: To demonstrate the effectiveness of RWPM for anomaly segmentation methods, we selected representative methods from both pixel-based and mask-based networks. For pixel-based methods, we chose PE-BAL [37] and Balanced Energy [12]. These methods all utilize DeepLabv3+ [9] with WideResnet38 as the backbone and were trained on the Cityscapes dataset and COCO OE images [37]. For mask-based methods, we selected RbA [34] and Mask2Anomaly [36]. These methods use Mask2Former [10] with Swin-L [33] as the backbone and were also trained on the Cityscapes and COCO OE images. All experiments are conducted on a single NVIDIA RTX A6000 GPU. To ensure efficient implementation on a single GPU for all experiments, we adopt embedding map partitioning and limited iteration in all experiments. Specifically, for pixel-based methods, we set the partitioning parameter " + }, + { + "bbox": [ + 130, + 248, + 482, + 477 + ], + "type": "inline_equation", + "content": "n = 4" + }, + { + "bbox": [ + 130, + 248, + 482, + 477 + ], + "type": "text", + "content": ". For mask-based methods, as their embedding map size is smaller, we set " + }, + { + "bbox": [ + 130, + 248, + 482, + 477 + ], + "type": "inline_equation", + "content": "n = 2" + }, + { + "bbox": [ + 130, + 248, + 482, + 477 + ], + "type": "text", + "content": ". Additionally, for the random walk, if not specifically emphasized, we set the number of iterations " + }, + { + "bbox": [ + 130, + 248, + 482, + 477 + ], + "type": "inline_equation", + "content": "T = 20 / T = 5" + }, + { + "bbox": [ + 130, + 248, + 482, + 477 + ], + "type": "text", + "content": " for the Road Anomaly/the other datasets and the transition probability " + }, + { + "bbox": [ + 130, + 248, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\alpha = 0.99" + }, + { + "bbox": [ + 130, + 248, + 482, + 477 + ], + "type": "text", + "content": " for all datasets. The " + }, + { + "bbox": [ + 130, + 248, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 248, + 482, + 477 + ], + "type": "text", + "content": " is set as 0.01. It is important to note that our method is used in the inference phase. In other words, RWPM is directly applied to the pre-trained models of other existing anomaly segmentation methods, thus requiring no additional training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 494, + 280, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 280, + 506 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 280, + 506 + ], + "type": "text", + "content": "4.1 Effectiveness of RWPM" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 514, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 647 + ], + "type": "text", + "content": "First, to validate that RWPM can improve the performance of existing anomaly segmentation methods, we integrate our approach with four recent representative methods, namely PEBAL [37], Balanced Energy [12], Mask2Anomaly [36] and RbA [34]. PEBAL and Balanced Energy employ pixel-based networks as their backbones, whereas Mask2Anomaly and RbA utilize mask-based networks as their backbones. We evaluate the results on both the Fishyscapes Lost&Found validation set and Road Anomaly test set. The experimental results are reported in Table 1. These results demonstrate that our proposed RWPM consistently and significantly enhances the performance of existing anomaly segmentation methods, across different backbone architectures. Additional results and parameters details are shown in Section S.3 of the supplementary material." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "RWPM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 411, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 411, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 411, + 665 + ], + "type": "text", + "content": "2 We carefully reproduce the experiments by using the official code." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 167, + 179, + 446, + 261 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": "Table 1: Comparison with strong and representative anomaly segmentation baselines across different backbone Architectures (Arch). Bold denotes the better results between using RWPM and not using RWPM with the same anomaly segmentation baselines. " + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": " indicates that use the calibration in sub-map concatenation as described in Section 3.4." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 179, + 446, + 261 + ], + "lines": [ + { + "bbox": [ + 167, + 179, + 446, + 261 + ], + "spans": [ + { + "bbox": [ + 167, + 179, + 446, + 261 + ], + "type": "table", + "html": "
Benchmark→ Fishscapes Lost&Found Road Anomaly
Method ↓ArchAuROC↑AP↑FPR95↓AuROC↑AP↑FPR95↓
PEBAL [37]Pixel-based98.9658.814.7787.6345.1044.58
PEBAL + RWPM† (Ours)Pixel-base99.2066.853.6889.4850.2936.81
Balanced Energy [12]Pixel-based99.0367.072.9388.3141.4841.46
Balanced Energy + RWPM† (Ours)Pixel-base99.2972.952.3890.5148.2632.10
Mask2Anomaly2 [36]Mask-based95.4765.277.7996.5480.0413.95
Mask2Anomaly + RWPM (Ours)Mask-base95.4865.387.3397.4480.097.45
RbA [34]Mask-based98.6270.816.3097.9985.426.92
RbA + RWPM (Ours)Mask-base98.8271.166.1298.0487.345.27
", + "image_path": "6e9b9d9f65e060ef502e8a8c488e8f6e22c8fa50a180a1d3fa2524f9d3b17d31.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 134, + 318, + 482, + 444 + ], + "blocks": [ + { + "bbox": [ + 130, + 275, + 482, + 309 + ], + "lines": [ + { + "bbox": [ + 130, + 275, + 482, + 309 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 482, + 309 + ], + "type": "text", + "content": "Table 2: Comparison with state-of-the-art anomaly segmentation methods across different backbone Architectures (Arch). Bold and underline denote the best and the second best results, respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 318, + 482, + 444 + ], + "lines": [ + { + "bbox": [ + 134, + 318, + 482, + 444 + ], + "spans": [ + { + "bbox": [ + 134, + 318, + 482, + 444 + ], + "type": "table", + "html": "
Benchmark→Fishscapes L&FRoad AnomalyAnomaly TrackObstacle TrackAverage
Method ↓ArchAP↑FPR95↓AP↑FPR95↓AP↑FPR95↓AP↑FPR95↓AP↑FPR95↓
Synboost [15] (CVPR'21)Pixel-based60.5831.0241.8459.7256.4461.8671.343.1557.4638.94
SML [26] (ICCV'21)Pixel-based22.7433.4925.8249.4746.839.53.436.824.6939.82
Meta-OOD [7] (ICCV'21)Pixel-based41.3137.6948.8431.7785.4715.0085.070.7565.1321.30
Learning Embedding [5](IJCV'21)Pixel-based4.6524.36--37.5270.760.8246.38--
Void Classifier [5](IJCV'21)Pixel-based10.2922.11--36.6163.4910.4441.54--
GMMSeg-DL [28] (NeurIPS'22)Pixel-based43.4713.1134.4247.90------
DenseHybrid [18] (ECCV'22)Pixel-based69.795.0931.3963.9777.969.8187.080.2465.5619.78
PEBAL [37] (ECCV'22)Pixel-based58.814.7745.1044.5849.1440.824.9812.6839.5825.71
Balanced Energy [12] (CVPR'23)Pixel-based67.072.9341.4841.46------
RPL-CoroCL [32] (ICCV'23)Pixel-based70.612.5271.6117.7483.4911.6885.930.5877.918.13
Mask2Anomaly [36] (ICCV'23)Mask-based69.449.2280.0413.9588.6214.5793.100.2082.879.49
M2F-EAM [19] (CVPRW'23)Mask-based52.0320.5166.6713.4276.393.966.917.965.4836.43
RbA [34] (ICCV'23)Mask-based70.816.3085.426.9290.8611.5991.850.4684.736.33
RbA + RWPM (Ours)Mask-based71.166.1287.345.2792.0010.1593.300.2886.005.46
", + "image_path": "f4602b08a7368d13dce289755f65213222d6f4abf80d21cfe1cdda10c36152e5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "type": "text", + "content": "Second, to further demonstrate the effectiveness of our RWPM, we compare it with other state-of-the-art (SOTA) methods. Specifically, we adopt the combination of RbA with RWPM as our instance approach for the comparison. In addition to the Fishyscapes and Road Anomaly datasets, we also evaluate our approach on a more challenging dataset, i.e., SMIYC online test set (Anomaly track and Obstacle track). SMIYC is characterized by its high domain shift and diversity of objects. Table 2 presents the results. The results indicate that our method can consistently achieve high performance across various datasets, while the previous SOTA methods fail to do so. We also compare all approaches by computing the average results of AP and FPR95. The results show that our approach outperforms the previous methods and achieve the best results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Furthermore, in addition to pixel-level metrics, we also utilize component-level metrics [6] to evaluate performance on the SMIYC online test set. Compared to pixel-level metrics, component-level metrics provide a better evaluation of the detection results for all anomalous objects in the scene. We report the results in Table 3. The experimental results show that RWPM significantly improves" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "Z. Zeng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": "the performance of RbA in component-level metrics, surpassing other existing SOTA methods by a large margin. This indicates that RWPM can assist models in better detecting anomalous objects and achieving more precise localization of these anomalies [6]." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 135, + 217, + 482, + 342 + ], + "blocks": [ + { + "bbox": [ + 132, + 185, + 481, + 208 + ], + "lines": [ + { + "bbox": [ + 132, + 185, + 481, + 208 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 481, + 208 + ], + "type": "text", + "content": "Table 3: Component level evaluation: Comparison with state-of-the-art anomaly segmentation methods on the SMIYC online test set. Bold denote the best results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 217, + 482, + 342 + ], + "lines": [ + { + "bbox": [ + 135, + 217, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 135, + 217, + 482, + 342 + ], + "type": "table", + "html": "
Benchmark→Anomaly TrackObstacle Track
Method ↓ArchsIoU ↑PPV↑mean F1↑sIoU ↑PPV↑mean F1↑
Synboost [15] (CVPR'21)Pixel-based34.6817.819.9944.2841.7537.57
SML [26] (ICCV'21)Pixel-based26.0024.7012.205.1013.303.00
Meta-OOD [7] (ICCV'21)Pixel-based49.3139.5128.7247.8763.6448.51
Void Classifier [5](IJCV'21)Pixel-based21.1422.136.496.3420.275.41
Learning Embedding [5](IJCV'21)Pixel-based33.8620.547.9035.642.872.31
JSRNet [38] (ICCV'21)Pixel-based20.2029.2713.6618.5524.4611.02
DenseHybrid [18] (ECCV'22)Pixel-based54.1724.1331.0845.7450.1050.72
PEBAL [37] (ECCV'22)Pixel-based38.8827.2014.4829.917.555.54
RPL-CoroCL [32] (ICCV'23)Pixel-based49.7629.9630.1652.6156.6556.69
Mask2Fromer [36] (ICCV'23)Mask-based25.2018.2015.305.0021.904.80
Maskomaly [2] (BMVC'23)Mask-based55.451.650.0---
RbA [34] (ICCV'23)Mask-based55.6952.1446.8058.3658.7860.85
RbA + RWPM (Ours)Mask-based57.0061.2558.4458.8972.5169.85
", + "image_path": "ece8203819a604e5d32cd9147eb9bb8e1c9073c8b09f7d107620d63535ce826f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 382, + 238, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 382, + 238, + 395 + ], + "spans": [ + { + "bbox": [ + 132, + 382, + 238, + 395 + ], + "type": "text", + "content": "4.2 Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 402, + 481, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 481, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 481, + 425 + ], + "type": "text", + "content": "If not specifically emphasized, all the results reported in this section are conducted on a RTX A6000 GPU." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": "The effect of the partitioning parameter " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": " and calibration: We first test the effect of the embedding map partitioning parameter " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": " and the calibration, and we report the results in Table 4. FPS denotes the frames-per-second results. Row " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "n = 1" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": " displays the RWPM results without utilizing the embedding map partitioning strategy, indicating significant computational and memory overhead. However, the partitioning strategy (" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "n = 2, 4, 8" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": ") significantly improves the running speed and reduces memory consumption while maintaining effectiveness. Additionally, as " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": " increases, the running efficiency also improves, albeit with a slight decrease in performance. However, when " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "n > 2" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": ", the calibration can enhance the performance, particularly for the FPR95 metrics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "type": "text", + "content": "The effect of iteration number " + }, + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "type": "text", + "content": ": Table 5 shows that, compared to using Eq. 4 (" + }, + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "type": "inline_equation", + "content": "T = \\infty" + }, + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "type": "text", + "content": "), utilizing the limited iteration strategy (" + }, + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "type": "inline_equation", + "content": "T = 5 \\sim 100" + }, + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "type": "text", + "content": ") can lead to higher efficiency, resulting in even better performance for RWPM. Additionally, as the number of iterations increases, performance tends to improve, but running efficiency (FPS) decreases accordingly." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Different strategies for reducing the number of pixel embedding: We also attempt other strategies to reduce the number of pixel embedding for RWPM, such as resizing the embedding map or use SLIC [1] algorithm to extract superpixels. Table 6(Left) presents the results. The resize strategy achieves good performance but requires preserving a larger map size. Therefore, it fails" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "RWPM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 201, + 146, + 411, + 234 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 4: The effect of the partitioning parameter " + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": " and calibration. We set " + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "T = {20}" + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": " for all experiments. " + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": " indicates that use the calibration in sub-map concatenation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 201, + 146, + 411, + 234 + ], + "lines": [ + { + "bbox": [ + 201, + 146, + 411, + 234 + ], + "spans": [ + { + "bbox": [ + 201, + 146, + 411, + 234 + ], + "type": "table", + "html": "
Benchmark→Road Anomaly
AP↑FPR95↓FPS↑GPU Memory Usage↓
RbA w/o RWPM85.426.9211.123.49GiB
n=1 (two GPUs)87.905.170.3274.96GiB
n=287.345.272.047.24GiB
n=2†87.185.252.047.24GiB
n=487.056.314.353.49GiB
n=4†87.175.324.353.49GiB
n=886.676.426.263.55GiB
n=8†86.915.416.263.55GiB
", + "image_path": "e7034d61e2de45c479c1e29f42489f1365e9cdfed056d6be79516809a414ff66.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 201, + 266, + 411, + 342 + ], + "blocks": [ + { + "bbox": [ + 149, + 245, + 463, + 258 + ], + "lines": [ + { + "bbox": [ + 149, + 245, + 463, + 258 + ], + "spans": [ + { + "bbox": [ + 149, + 245, + 463, + 258 + ], + "type": "text", + "content": "Table 5: The effect of iteration number " + }, + { + "bbox": [ + 149, + 245, + 463, + 258 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 149, + 245, + 463, + 258 + ], + "type": "text", + "content": " . We set " + }, + { + "bbox": [ + 149, + 245, + 463, + 258 + ], + "type": "inline_equation", + "content": "n = 2" + }, + { + "bbox": [ + 149, + 245, + 463, + 258 + ], + "type": "text", + "content": " for all experiments." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 201, + 266, + 411, + 342 + ], + "lines": [ + { + "bbox": [ + 201, + 266, + 411, + 342 + ], + "spans": [ + { + "bbox": [ + 201, + 266, + 411, + 342 + ], + "type": "table", + "html": "
Benchmark→ Road Anomaly
AP↑FPR95↓FPS↑GPU Memory Usage↓
RbA w/o RWPM85.426.9211.123.49GiB
T=∞87.515.240.298.88GiB
T=586.895.983.707.24GiB
T=1087.085.772.787.24GiB
T=2087.345.272.047.24GiB
T=5087.365.080.927.24GiB
T=10087.545.020.507.24GiB
", + "image_path": "2590a223a96f8ebbe925195f3ed69338d07de91b36c75eeba15eb9c7f2eec250.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "text", + "content": "to substantially reduce the number of pixel embeddings, resulting in significant GPU memory usage. On the other hand, the superpixel strategy requires a considerable amount of runtime for the superpixel extraction. In comparison, our proposed partitioning strategy can achieve good results with higher efficiency and lower resource consumption. This makes it more suitable for applications on practical platforms." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 437, + 482, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 522 + ], + "type": "text", + "content": "In-distribution segmentation performance on Cityscapes validation set: We report the results on the Cityscapes validation set in Table 6(Right). All methods use Mask2Former [10] as their backbones. We observe that most SOTA methods affect the in-distribution segmentation performance of Mask2Former, such as RbA (drop from 82.25 to 81.93). However, RWPM can alleviate this problem. Additionally, even for the SOTA segmentation method (Mask2Former), RWPM can further enhance its in-distribution segmentation performance." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 134, + 601, + 309, + 659 + ], + "blocks": [ + { + "bbox": [ + 134, + 601, + 309, + 659 + ], + "lines": [ + { + "bbox": [ + 134, + 601, + 309, + 659 + ], + "spans": [ + { + "bbox": [ + 134, + 601, + 309, + 659 + ], + "type": "table", + "html": "
Benchmark→ Road Anomaly
Strategy↓AP↑FPR95↓FPS↑Memory
Resize (0.9 full size)87.805.391.4123.82GIB
Resize (0.8 full size)87.245.442.1515.98GiB
Superpixel (30000 pixels)87.515.500.1911.46GiB
Superpixel (20000 pixels)87.035.980.326.31GiB
Partitioning (n = 2)87.345.272.047.24GiB
", + "image_path": "c94fa5e457112aed6076439fd0cdd0f7c2a43759761404ada407ad65b49f769f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 334, + 597, + 455, + 662 + ], + "blocks": [ + { + "bbox": [ + 130, + 544, + 482, + 588 + ], + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 588 + ], + "type": "text", + "content": "Table 6: Ablation Tables: (Left) shows the results of different strategies for reducing the number of pixel embedding. We find that partitioning is the best. (Right) shows the segmentation performance on the Cityscapes validation set. RWPM can also improve the segmentation performance of in-distribution." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 334, + 597, + 455, + 662 + ], + "lines": [ + { + "bbox": [ + 334, + 597, + 455, + 662 + ], + "spans": [ + { + "bbox": [ + 334, + 597, + 455, + 662 + ], + "type": "table", + "html": "
Benchmark→Cityscapes
Method↓mIoU↑
PEBAL-Mask [34]75.32
DenseHybrid-Mask [34]80.27
RbA [34]81.93
RbA+RWPM(Ours)82.16
Mask2Former [10]82.25
Mask2Former+RWPM(Ours)82.43
", + "image_path": "6f9c75e324ea804073c35260d96260aa04a21cd1edc52fbe0b42bd4b3c96067c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "Z. Zeng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 394, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 394, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 394, + 128 + ], + "type": "text", + "content": "4.3 Visualization of Pixel Embedding Distribution" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 482, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 482, + 410 + ], + "type": "text", + "content": "To investigate the effect of RWPM in the embedding space, we visualized the distribution of pixel embeddings by using t-SNE. We selected an image from the Road Anomaly dataset (as shown in Fig 3(a)). Due to the gravel road surface in this image, there is a significant domain gap between the road in this image and the roads in the training set, resulting in many false positives on the road (as shown in the PEBAL results in Fig 3(a)). Then, we randomly selected 300 inlier pixels and 150 outlier pixels from the red dashed box region of this image for visualization. Fig 3(b) and Fig 3(c) respectively show the distributions of these pixels in the space without and with applying RWPM. In Fig 3(b) and Fig 3(c), deep blue circles represent inlier pixels (road), light blue circles represent outlier pixels (boar), and X symbols represent the inlier class prototype (the vector of network weights of classifier). From Fig 3(b), we observe that without using RWPM, due to the influence of data diversity, the distribution of inliers and outliers is distorted. Consequently, it is difficult for an anomaly scoring function to distinguish between inlier and outlier pixels based on their logits. However, from Fig 3(c), we find that RWPM can significantly alleviate the problem. Specifically, the pixel embeddings of the same category form more compact clusters, enabling the anomaly scoring function to clearly discriminate between inliers and outliers based on the logit. This is consistent with the description in the toy sample in Fig 1. Note that, in the top left corner of Fig 3(c), some inlier points become very close to outlier points after applying RWPM. This is because these points are located in the boundaries of anomalous objects, making it very difficult to discriminate them." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 136, + 432, + 231, + 511 + ], + "blocks": [ + { + "bbox": [ + 136, + 432, + 231, + 511 + ], + "lines": [ + { + "bbox": [ + 136, + 432, + 231, + 511 + ], + "spans": [ + { + "bbox": [ + 136, + 432, + 231, + 511 + ], + "type": "image", + "image_path": "6dd5b6908bd5cb09c02c161f71c5dfd35a89c4d5e7069e3b5ab17ee3a1d54722.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 147, + 514, + 220, + 525 + ], + "lines": [ + { + "bbox": [ + 147, + 514, + 220, + 525 + ], + "spans": [ + { + "bbox": [ + 147, + 514, + 220, + 525 + ], + "type": "text", + "content": "(a) Sample Image" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 241, + 436, + 350, + 508 + ], + "blocks": [ + { + "bbox": [ + 241, + 436, + 350, + 508 + ], + "lines": [ + { + "bbox": [ + 241, + 436, + 350, + 508 + ], + "spans": [ + { + "bbox": [ + 241, + 436, + 350, + 508 + ], + "type": "image", + "image_path": "e2e53be7e629c7892031c9d2987f959c7d9189311c0793d88e366dbe9e2abd20.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 535, + 482, + 568 + ], + "lines": [ + { + "bbox": [ + 130, + 535, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 535, + 482, + 568 + ], + "type": "text", + "content": "Fig. 3: The visualization of pixel embedding distribution. All embeddings are extracted from PEBAL model. We use cosine distance as the distance metric. We observed that RWPM can optimize the distribution of pixel embeddings in the space." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 369, + 437, + 476, + 508 + ], + "blocks": [ + { + "bbox": [ + 239, + 514, + 356, + 525 + ], + "lines": [ + { + "bbox": [ + 239, + 514, + 356, + 525 + ], + "spans": [ + { + "bbox": [ + 239, + 514, + 356, + 525 + ], + "type": "text", + "content": "(b) Distribution w/o RWPM" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 369, + 437, + 476, + 508 + ], + "lines": [ + { + "bbox": [ + 369, + 437, + 476, + 508 + ], + "spans": [ + { + "bbox": [ + 369, + 437, + 476, + 508 + ], + "type": "image", + "image_path": "ed945554d75c59edb7f825bfe123f491a8baa0a13b4cf05cac18a6fb9a3371d6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 514, + 484, + 526 + ], + "lines": [ + { + "bbox": [ + 365, + 514, + 484, + 526 + ], + "spans": [ + { + "bbox": [ + 365, + 514, + 484, + 526 + ], + "type": "text", + "content": "(c) Distribution with RWPM" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 610, + 253, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 253, + 623 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 253, + 623 + ], + "type": "text", + "content": "4.4 Qualitative Result" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "Finally, to visually demonstrate the effectiveness of RWPM, we present qualitative results. Specifically, we compare the results of RbA on the Road Anomaly dataset without and with RWPM. The qualitative results are illustrated in Fig 4." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "RWPM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "Fig 4 shows that RWPM significantly improves the performance of RbA, reducing both the false positive rate and false negative rate. For example, in the first row, RbA generates some false positive results on the road and some false negative results on the manhole region (third column of the first row). However, after applying RWPM, these false positive/negative results are effectively eliminated (fourth column of the first row). Additionally, RWPM assists the RbA model to accurately detect and localize each anomalous object at the component level, such as more precise edge detection. This observation aligns with the results of the component-level metrics presented in Table 3." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 143, + 243, + 473, + 385 + ], + "blocks": [ + { + "bbox": [ + 143, + 243, + 473, + 385 + ], + "lines": [ + { + "bbox": [ + 143, + 243, + 473, + 385 + ], + "spans": [ + { + "bbox": [ + 143, + 243, + 473, + 385 + ], + "type": "image", + "image_path": "cc0f8c32e7c4447fe9c9c6d6ca48e6b2e3e8f9f23f94724ed4996462dff56e83.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 394, + 482, + 439 + ], + "lines": [ + { + "bbox": [ + 130, + 394, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 482, + 439 + ], + "type": "text", + "content": "Fig. 4: Qualitative results: The first column displays the input image, while the second column shows its corresponding ground truth. The third column displays the anomaly score results generated by RbA, and the fourth column presents the anomaly score results after applying RWPM to RbA. Yellow indicates high anomaly scores." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 475, + 323, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 475, + 323, + 489 + ], + "spans": [ + { + "bbox": [ + 131, + 475, + 323, + 489 + ], + "type": "text", + "content": "5 Conclusions and Discussions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 617 + ], + "type": "text", + "content": "In this paper, we propose a simple yet effective method, called Random Walk on Pixel Manifolds (RWPM), to improve the performance of existing anomaly segmentation methods. RWPM is a post-processing method that can be directly integrated into existing anomaly detection frameworks without requiring additional training or modifications to the network structure. RWPM mitigates the impact of data diversity on manifold structure by diffusing and updating pixel embeddings on pixel manifolds using random walk. This process enables anomaly scoring functions to obtain more accurate estimations of anomaly scores. Additionally, we propose Partial Random Walk strategy, which significantly improves the efficiency of RWPM." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": "While RWPM achieves significant results in most cases, it is limited when the pre-trained model itself exhibits significant inaccuracies. To address this limitation, we plan to enhance the model's generalization ability during training to accurately detect a broader range of diverse and unknown anomalies." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "Z. Zeng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 340, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 340, + 150 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 340, + 150 + ], + "type": "text", + "content": "This work was supported by SenseTime Japan." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 167, + 197, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 167, + 197, + 178 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 197, + 178 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 189, + 480, + 663 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 189, + 480, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 189, + 480, + 221 + ], + "spans": [ + { + "bbox": [ + 138, + 189, + 480, + 221 + ], + "type": "text", + "content": "1. Achanta, R., Shaji, A., Smith, K., Lucchi, A., Fua, P., Süsstrunk, S.: Slic superpixels compared to state-of-the-art superpixel methods. IEEE transactions on pattern analysis and machine intelligence 34(11), 2274-2282 (2012)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 222, + 480, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 222, + 480, + 243 + ], + "spans": [ + { + "bbox": [ + 138, + 222, + 480, + 243 + ], + "type": "text", + "content": "2. Ackermann, J., Sakaridis, C., Yu, F.: Maskomaly: Zero-shot mask anomaly segmentation. arXiv preprint arXiv:2305.16972 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 243, + 480, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 243, + 480, + 275 + ], + "spans": [ + { + "bbox": [ + 138, + 243, + 480, + 275 + ], + "type": "text", + "content": "3. Bai, S., Bai, X., Tian, Q., Latecki, L.J.: Regularized diffusion process for visual retrieval. In: Proceedings of the AAAI conference on artificial intelligence. vol. 31 (2017). https://doi.org/10.1609/aaai.v31i1.11216" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 276, + 480, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 276, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 276, + 480, + 308 + ], + "type": "text", + "content": "4. Bai, S., Zhou, Z., Wang, J., Bai, X., Jan Latecki, L., Tian, Q.: Ensemble diffusion for retrieval. In: Proceedings of the IEEE International conference on computer vision. pp. 774-783 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 308, + 480, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 308, + 480, + 350 + ], + "spans": [ + { + "bbox": [ + 138, + 308, + 480, + 350 + ], + "type": "text", + "content": "5. Blum, H., Sarlin, P.E., Nieto, J., Siegwart, R., Cadena, C.: The fishyscapes benchmark: Measuring blind spots in semantic segmentation. International Journal of Computer Vision 129, 3119-3135 (2021). https://doi.org/10.1007/s11263-021-01511-6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 351, + 480, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 351, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 138, + 351, + 480, + 384 + ], + "type": "text", + "content": "6. Chan, R., Lis, K., Uhlemeyer, S., Blum, H., Honari, S., Siegwart, R., Fua, P., Salzmann, M., Rottmann, M.: Segmentmeifyoucan: A benchmark for anomaly segmentation. arXiv preprint arXiv:2104.14812 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 384, + 480, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 384, + 480, + 416 + ], + "spans": [ + { + "bbox": [ + 138, + 384, + 480, + 416 + ], + "type": "text", + "content": "7. Chan, R., Rottmann, M., Gottschalk, H.: Entropy maximization and meta classification for out-of-distribution detection in semantic segmentation. In: Proceedings of the ieee/cvf international conference on computer vision. pp. 5128-5137 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 417, + 480, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 417, + 480, + 459 + ], + "spans": [ + { + "bbox": [ + 138, + 417, + 480, + 459 + ], + "type": "text", + "content": "8. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence 40(4), 834-848 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 460, + 480, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 480, + 502 + ], + "type": "text", + "content": "9. Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 801-818 (2018). https://doi.org/10.1007/978-3-030-01234-2_49" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 503, + 480, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 503, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 138, + 503, + 480, + 545 + ], + "type": "text", + "content": "0. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290–1299 (2022). https://doi.org/10.1109/CVPR52688.2022.00135" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 546, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 546, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 546, + 480, + 578 + ], + "type": "text", + "content": "1. Cheng, B., Schwing, A., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems 34, 17864-17875 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 578, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 611 + ], + "type": "text", + "content": "2. Choi, H., Jeong, H., Choi, J.Y.: Balanced energy regularization loss for out-of-distribution detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15691-15700 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 611, + 480, + 663 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 663 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 663 + ], + "type": "text", + "content": "3. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3213-3223 (2016). https://doi.org/10.1109/CVPR.2016.350" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "RWPM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "14. Creusot, C., Munawar, A.: Real-time small obstacle detection on highways using compressive rbm road reconstruction. In: 2015 IEEE Intelligent Vehicles Symposium (IV). pp. 162-167. IEEE (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 192 + ], + "type": "text", + "content": "15. Di Biase, G., Blum, H., Siegwart, R., Cadena, C.: Pixel-wise anomaly detection in complex driving scenes. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16918-16927 (2021). https://doi.org/10.1109/CVPR46437.2021.01664" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "type": "text", + "content": "16. Gal, Y., Ghahramani, Z.: Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In: international conference on machine learning. pp. 1050-1059. PMLR (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 228, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 480, + 248 + ], + "type": "text", + "content": "17. Grady, L.: Random walks for image segmentation. IEEE transactions on pattern analysis and machine intelligence 28(11), 1768-1783 (2006)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "text", + "content": "18. Grcic, M., Bevandić, P., Šegvic, S.: Densehybrid: Hybrid anomaly detection for dense open-set recognition. In: European Conference on Computer Vision. pp. 500-517. Springer (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 314 + ], + "type": "text", + "content": "19. Grcic, M., Saric, J., Segvic, S.: On advantages of mask-level recognition for outlier-aware segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2936-2946 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "text", + "content": "20. Hendrycks, D., Basart, S., Mazeika, M., Zou, A., Kwon, J., Mostajabi, M., Steinhardt, J., Song, D.: Scaling out-of-distribution detection for real-world settings. arXiv preprint arXiv:1911.11132 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 348, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 480, + 380 + ], + "type": "text", + "content": "21. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. In: International Conference on Learning Representations (2017), https://openreview.net/forum?id=Hkg4TI9x1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "text", + "content": "22. Hendrycks, D., Mazeika, M., Dietterich, T.: Deep anomaly detection with outlier exposure. arXiv preprint arXiv:1812.04606 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 434 + ], + "type": "text", + "content": "23. Huang, Z., Li, S.: From difference to similarity: A manifold ranking-based hyperspectral anomaly detection framework. IEEE Transactions on Geoscience and Remote Sensing 57(10), 8118-8130 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 480, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 477 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 477 + ], + "type": "text", + "content": "24. Iscen, A., Tolias, G., Avrithis, Y., Chum, O.: Mining on manifolds: Metric learning without labels. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7642-7651 (2018). https://doi.org/10.1109/CVPR.2018.00797" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 522 + ], + "type": "text", + "content": "25. Iscen, A., Tobias, G., Avrithis, Y., Furon, T., Chum, O.: Efficient diffusion on region manifolds: Recovering small objects with compact cnn representations. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2077-2086 (2017). https://doi.org/10.1109/CVPR.2017.105" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 523, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 566 + ], + "type": "text", + "content": "26. Jung, S., Lee, J., Gwak, D., Choi, S., Choo, J.: Standardized max logits: A simple yet effective approach for identifying unexpected road obstacles in urban-scene segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15425-15434 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "text", + "content": "27. Lakshminarayanan, B., Pritzel, A., Blundell, C.: Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "28. Liang, C., Wang, W., Miao, J., Yang, Y.: Gmmseg: Gaussian mixture based generative semantic segmentation models. Advances in Neural Information Processing Systems 35, 31360-31375 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "29. Liang, S., Li, Y., Srikant, R.: Enhancing the reliability of out-of-distribution image detection in neural networks. In: International Conference on Learning Representations (2018), https://openreview.net/forum?id=H1VGkIxRZ" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "text", + "content": "Z. Zeng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 632 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 172 + ], + "type": "text", + "content": "30. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014). https://doi.org/10.1007/978-3-319-10602-1_48" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "text", + "content": "31. Lis, K., Nakka, K., Fua, P., Salzmann, M.: Detecting the unexpected via image resynthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2152-2161 (2019). https://doi.org/10.1109/ICCV.2019.00224" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 247 + ], + "type": "text", + "content": "32. Liu, Y., Ding, C., Tian, Y., Pang, G., Belagiannis, V., Reid, I., Carneiro, G.: Residual pattern learning for pixel-wise out-of-distribution detection in semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1151-1161 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 248, + 481, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 248, + 481, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 248, + 481, + 292 + ], + "type": "text", + "content": "33. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "text", + "content": "34. Nayal, N., Yavuz, M., Henriques, J.F., Güney, F.: Rba: Segmenting unknown regions rejected by all. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 711-722 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "type": "text", + "content": "35. Ohgushi, T., Horiguchi, K., Yamanaka, M.: Road obstacle detection method based on an autoencoder with semantic segmentation. In: Proceedings of the Asian Conference on Computer Vision (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "text", + "content": "36. Rai, S.N., Cermelli, F., Fontanel, D., Masone, C., Caputo, B.: Unmasking anomalies in road-scene segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4037-4046 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 391, + 481, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 481, + 434 + ], + "type": "text", + "content": "37. Tian, Y., Liu, Y., Pang, G., Liu, F., Chen, Y., Carneiro, G.: Pixel-wise energy-biased abstention learning for anomaly segmentation on complex urban driving scenes. In: European Conference on Computer Vision. pp. 246-263. Springer (2022). https://doi.org/10.1007/978-3-031-19842-7_15" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 434, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 481, + 479 + ], + "type": "text", + "content": "38. Vojir, T., Šipka, T., Aljundi, R., Chumerin, N., Reino, D.O., Matas, J.: Road anomaly detection by partial image reconstruction with segmentation coupling. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15651-15660 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "39. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems 34, 12077-12090 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 511, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 511, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 511, + 481, + 555 + ], + "type": "text", + "content": "40. Yang, F., Hinami, R., Matsui, Y., Ly, S., Satoh, S.: Efficient image retrieval via decoupling diffusion into online and offline processing. In: Proceedings of the AAAI conference on artificial intelligence. vol. 33, pp. 9087-9094 (2019). https://doi.org/10.1609/aaai.v33i01.33019087" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 555, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 587 + ], + "type": "text", + "content": "41. Yang, F., Wang, Z., Xiao, J., Satoh, S.: Mining on heterogeneous manifolds for zero-shot cross-modal image retrieval. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 12589-12596 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 588, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 481, + 611 + ], + "type": "text", + "content": "42. Zhou, D., Bousquet, O., Lal, T., Weston, J., Scholkopf, B.: Learning with local and global consistency. Advances in neural information processing systems 16 (2003)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "text", + "content": "43. Zhou, D., Weston, J., Gretton, A., Bousquet, O., Schölkopf, B.: Ranking on data manifolds. Advances in neural information processing systems 16 (2003)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "RWPM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/RangeLDM_ Fast Realistic LiDAR Point Cloud Generation/ec4004eb-e76a-4f41-9891-f055a3eeaf7c_content_list.json b/2024/RangeLDM_ Fast Realistic LiDAR Point Cloud Generation/ec4004eb-e76a-4f41-9891-f055a3eeaf7c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2326a7f3e8cdb7ef4fabeec28cef5f317759f0e7 --- /dev/null +++ b/2024/RangeLDM_ Fast Realistic LiDAR Point Cloud Generation/ec4004eb-e76a-4f41-9891-f055a3eeaf7c_content_list.json @@ -0,0 +1,2019 @@ +[ + { + "type": "text", + "text": "RangeLDM: Fast Realistic LiDAR Point Cloud Generation", + "text_level": 1, + "bbox": [ + 233, + 140, + 772, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qianjiang Hu, Zhimin Zhang, and Wei Hu", + "bbox": [ + 321, + 213, + 673, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wangxuan Institute of Computer Technology, Peking University, Beijing, China hqjpu@pku.edu.cn zm_zhang@stu.pku.edu.cn forhuwei@pku.edu.cn", + "bbox": [ + 235, + 241, + 767, + 270 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bec2e48447c9c338c2bbaeaff95f297df8162460662d1d078c38a0ffad5f51d5.jpg", + "image_caption": [ + "(a) Unconditional Generation" + ], + "image_footnote": [], + "bbox": [ + 217, + 305, + 344, + 421 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/454315a4e9010a1ee50a3bec5d60c4f2c2ccecdffce1caba0a9f0534c53bcad9.jpg", + "image_caption": [ + "(b) Conditional Generation" + ], + "image_footnote": [], + "bbox": [ + 351, + 305, + 604, + 421 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2555aee352ba07603cd717c1df9adb4d8080dfcf1c5a8735f801122ecd200ee3.jpg", + "image_caption": [ + "(c) Quality vs. Speed", + "Fig. 1: (a). Unconditional LiDAR point cloud generation with realistic global structure. (b). Conditional LiDAR point cloud generation, including LiDAR point cloud upsampling and inpainting. (c). Generation quality (Maximum Mean Discrepancy, abbr. MMD) vs. generation speed (samples/s) of competitive LiDAR point cloud generation methods on the KITTI-360 [38] dataset. The proposed method outperforms the state-of-the-art methods LiDARGen [95] and UltraLiDAR [79] in both generation quality and generation speed. All speeds are evaluated on a single RTX 3090 GPU." + ], + "image_footnote": [], + "bbox": [ + 607, + 309, + 782, + 421 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Autonomous driving demands high-quality LiDAR data, yet the cost of physical LiDAR sensors presents a significant scaling-up challenge. While recent efforts have explored deep generative models to address this issue, they often consume substantial computational resources with slow generation speeds while suffering from a lack of realism. To address these limitations, we introduce RangeLDM, a novel approach for rapidly generating high-quality range-view LiDAR point clouds via latent diffusion models. We achieve this by correcting range-view data distribution for accurate projection from point clouds to range images via Hough voting, which has a critical impact on generative learning. We then compress the range images into a latent space with a variational autoencoder, and leverage a diffusion model to enhance expressivity. Additionally, we instruct the model to preserve 3D structural fidelity by devising a range-guided discriminator. Experimental results on KITTI-360 and nuScenes datasets demonstrate both the robust expressiveness and fast speed of our LiDAR point cloud generation.", + "bbox": [ + 259, + 564, + 740, + 785 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Diffusion models $\\cdot$ Autonomous systems $\\cdot$ Point clouds", + "bbox": [ + 261, + 800, + 712, + 814 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\mathbb{O}$ Corresponding author: W. Hu.", + "bbox": [ + 230, + 824, + 452, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 143, + 375, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Autonomous systems have drawn great attention in both the academia and the industry community. Numerous autonomous systems leverage the power of various sensors and deep learning to enhance the perception of the 3D world. LiDAR (Light Detection And Ranging), with its ability to provide precise 3D geometric information about the surroundings, has become a popular sensor choice for autonomous systems, including self-driving cars [61,62,86], surveying drones [52,90] and robots [42,78].", + "bbox": [ + 212, + 175, + 785, + 280 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, while LiDAR offers accurate geometric measurements, it comes with a significant limitation: the data collection process is exceedingly expensive and challenging to scale up. This renders physical sensors impractical for scalable and customizable data collection. Additionally, it is difficult to collect data in corner cases such as car accidents and extreme weather conditions in the field.", + "bbox": [ + 212, + 281, + 785, + 356 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One approach to mitigate this issue is to employ existing LiDAR simulation toolkits [37, 43] to synthesize more data. Nevertheless, these systems typically demand manual scene creation or rely on multiple prior scans of the real world. Another approach is leveraging deep generative models to generate LiDAR point clouds. This is challenging due to the high degree of unstructuredness, sparsity, and non-uniformity in LiDAR point clouds. Caccia et al. [5] proposed LiDAR GAN and LiDAR VAE to generate LiDAR point clouds with generative adversarial networks (GANs) [21] and variational autoencoders (VAEs) [32], respectively. However, the generation results often suffer from issues such as fuzzy or missing details. LiDARGen [95] introduced a novel score-based model to synthesis LiDAR point clouds, which however sampled slowly and failed to generate high-quality geometric details at a far range. UltraLiDAR [79] proposed to synthesize voxelized point clouds in bird's-eyes-view (BEV) with vector quantized VAE (VQ-VAE) [74], resulting in more realistic point clouds than previous methods. Nevertheless, due to the large size and sparsity of point clouds in BEV, most computing power of UltraLiDAR is devoted to generating empty voxels, which results in a low generation speed.", + "bbox": [ + 212, + 357, + 785, + 612 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose RangeLDM, a novel approach based on latent diffusion models (LDMs) [55], which improves both the quality and speed of LiDAR point cloud generation significantly. Firstly, we introduce a range image view to represent point clouds. The reasons are twofold: 1) Range images are compact and closely mimic the sampling conditions of LiDAR, making them a suitable representation. A LiDAR point cloud essentially constructs a 2.5D scene from a single viewpoint instead of a full 3D point cloud [27]. Consequently, organizing the point cloud in range view ensures that no information is overlooked. 2) 2D image generation techniques are relatively mature, providing a solid foundation for our work. While LiDARGen [95] also represents point clouds as range images, it suffers from blurriness caused by inaccurate projection. In contrast, we provide insights that the correct range-view data distribution has a critical impact on such range-view-based generative model learning, and accurately project point clouds onto range images using parameters estimated by Hough Voting. Then, to achieve a faster sampling speed [55] and greater expressivity [55,72], we compress", + "bbox": [ + 212, + 613, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Q. Hu et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "range images to the latent space with a VAE and develop a diffusion model operating on the latent space, leveraging the successful paradigm of LDMs [55,72,89]. Further, in order to enhance the ability of the VAE in reconstructing 3D structures, we introduce a range-guided discriminator, which is supervised from the spherical coordinates and thus geometry-sensitive. This plays a crucial role in guiding the decoder to generate high-quality range images, ultimately preserving the fidelity of the 3D structure. Extensive experimental results show that the proposed RangeLDM achieves the state-of-the-art performance on KITTI-360 [38] and nuScenes [6] datasets in terms of both the generation quality and generation speed, as demonstrated in Figure 1 (c). The LiDAR upsampling and LiDAR inpainting results also demonstrate the potential of the proposed method for conditional generation, as presented in Figure 1 (a) and (b).", + "bbox": [ + 212, + 146, + 787, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To summarize, the main contributions of this paper include:", + "bbox": [ + 240, + 329, + 671, + 343 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a latent diffusion model to capture the distribution of range-view LiDAR point clouds, aiming to generate realistic point cloud scenes at a fast speed.", + "- We enlighten the significance of the correct range-view data distribution for range-view-based generative models and achieve high-quality range image projection via Hough Voting.", + "- We exploit a range-guided discriminator to ensure preserving the fidelity of the 3D geometric structure in generated point clouds.", + "- Experimental results on the KITTI-360 and nuScenes datasets highlight that our approach outperforms state-of-the-art methods in terms of both visual quality and especially generation speed." + ], + "bbox": [ + 225, + 354, + 784, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 547, + 387, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LiDAR Representation. LiDAR point clouds can be represented in various forms, encompassing raw point clouds, voxels, range views, and multi-view fusion. Point-based models [49, 62, 64, 76, 84, 85, 91] directly encode 3D objects from raw points using the PointNet [50] encoder and subsequently perform detection or segmentation based on point features. These methods wholly preserve the irregularity and locality of a point cloud but have relatively higher latency. Voxel-based methods [13, 14, 17, 34, 35, 60, 63, 67, 77, 81, 82, 86, 94] voxelize point clouds for convolutional neural networks to efficiently capture features. They are computationally effective but the desertion of fine-grained patterns degrades further refinement. Given that the range view is compact and compatible with the LiDAR sensor's sampling process, range-image-based methods [2, 8, 10-12, 18, 33, 36, 44, 45, 68, 71, 80, 92] have been proposed to directly process range images for LiDAR perception. Multi-view fusion methods [22, 26, 61] amalgamate multiple representations, which yield better results at the expense of processing speed. In this paper, we adopt the range-view representation, which is congruent with the LiDAR sensor's sampling process and can be efficiently encoded by image generation models.", + "bbox": [ + 212, + 582, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "RangeLDM: Fast Realistic LiDAR Point Cloud Generation", + "bbox": [ + 339, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generative Models for Point Clouds. Given observed samples of interest, generative models aim to learn the underlying distribution of the data and generate new samples. The first 3D generative models [1,65,73] for point clouds are based on GANs [21]. They operate by generating point clouds and discriminating them from real samples in an adversarial manner. The second category of methods [19,20,31,40,46,47,70] employ VAEs [32] to explore the probabilistic latent space of 3D shapes. Auto-regressive models [4] have also been introduced to generate point clouds [69] from scratch or semantic contexts. PointFlow [83] and SoftFlow [30] leverage normalizing flows [53] to capture the likelihood of shapes and points. Notably, the recent success of denoising diffusion models (DDMs) [25] for image synthesis [55] has also been extended to the domain of 3D point cloud generation [7,28,41,48,89,93]. However, these previous works often concentrate on the object level and are not well-suited for handling large scenes.", + "bbox": [ + 212, + 146, + 787, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this paper, our focus lies in the generation of LiDAR point clouds. While [5] and [57] employ VAE or GANs for LiDAR point cloud generation, the realism achieved in their results is relatively limited. UltraLiDAR [79], on the other hand, utilizes VQ-VAE to generate voxelized LiDAR point clouds, but at the expense of introducing quantified losses and slow generation speed. LiDARGen [95] proposed a novel score-matching energy-based model to generate higher-quality Li-DAR point clouds, but it samples very slowly and has degraded geometric details at a far range. A contemporary work LiDM [51] also utilized LDMs to generate LiDAR point clouds. However, while they focused on generating LiDAR point clouds conditional on multimodal data, we focus on improving the quality of point cloud generation directly.", + "bbox": [ + 212, + 357, + 787, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Background on Denoising Diffusion Models", + "text_level": 1, + "bbox": [ + 215, + 545, + 671, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "DDMs [25] are latent variable models that employ a pre-defined posterior distribution, known as the forward diffusion process, and are trained with a denoising objective. More specifically, given samples $\\mathbf{x}_0\\sim q(\\mathbf{x}_0)$ from a data distribution, DDMs follow the form $p_{\\theta}(\\mathbf{x}_0)\\coloneqq \\int p_{\\theta}(\\mathbf{x}_{0:T})d\\mathbf{x}_{1:T}$ , where $T$ denotes the number of steps, $\\mathbf{x}_1,\\dots,\\mathbf{x}_T$ are latent variables that gradually add noise to the data $\\mathbf{x}_0$ , and $\\theta$ denotes the parameter set of the DDM decoder.", + "bbox": [ + 212, + 575, + 787, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "DDMs are trained by minimizing the evidence lower bound (ELBO) of the data $\\mathbf{x}_0$ under $p_{\\theta}(\\mathbf{x}_{0:T})$ . This objective can be simplified to [25]:", + "bbox": [ + 212, + 666, + 787, + 698 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {\\mathbf {x}, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {x} _ {t} + \\epsilon , t) \\| _ {2} ^ {2} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 705, + 785, + 732 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $t$ uniformly sampled from $\\{1,\\dots,T\\}$ .", + "bbox": [ + 214, + 742, + 516, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 The Proposed Method", + "text_level": 1, + "bbox": [ + 215, + 779, + 472, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 2, we first project point clouds onto high-quality range images (as detailed in Section 4.1), considering the compact nature of the range view,", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Q. Hu et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/81e720c0c8cc65c583bd34340169d7c339a66c8720c285c53ecfb5c22fc94402.jpg", + "image_caption": [ + "Fig. 2: The framework of the proposed RangeLDM. Firstly, we project point clouds onto high-quality range images via Hough Voting (Section 4.1). Subsequently, we train a VAE to compress the range images into low-dimensional latent features $\\mathbf{z}_0$ , which encodes the range images with the encoder $\\mathcal{E}_{\\zeta}$ and reconstructs range images from latent features with the decoder $\\mathcal{G}_{\\eta}$ (Section 4.2). Here, a range-guided discriminator $\\mathcal{D}_{\\tau}$ is introduced to guide the decoder in the reconstruction of 3D structures. We finally train a latent diffusion model to capture the distribution of the latent features (Section 4.2). With optional conditional inputs, the proposed method is applicable to tasks such as point cloud upsampling and inpainting (Section 4.3)." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 787, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which aligns seamlessly with the sampling process of LiDAR sensors. Our model training process is then focused on these projected 2D range images. We then compress range images into a low-dimensional latent space through a VAE (as detailed in Section 4.2). Subsequently, we proceed to train DDMs within the reduced-dimensional latent space (as discussed in Section 4.2). We discuss applications of the proposed model in Section 4.3, which involves unconditional generation and conditional generation. Implementation details are explained in Section 4.4.", + "bbox": [ + 212, + 397, + 787, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 High-Quality Range Projection", + "text_level": 1, + "bbox": [ + 215, + 541, + 519, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The range image presents LiDAR data compactly and intuitively, with rows indicating the laser beams and columns representing the yaw angles. We convert point clouds to range images using spherical projection. Typically, for a point $\\mathbf{p}$ in Cartesian coordinates $(x,y,z)$ , we calculate its spherical coordinates $(r,\\theta,\\phi)$ using:", + "bbox": [ + 212, + 566, + 787, + 643 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr = \\sqrt {x ^ {2} + y ^ {2} + z ^ {2}}, \\theta = \\mathrm {a t a n} (y, x), \\phi = \\mathrm {a t a n} \\left(z, \\sqrt {x ^ {2} + y ^ {2}}\\right). \\qquad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 284, + 654, + 785, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, in most current datasets such as KITTI-360, multiple lasers from the Velodyne LiDAR system do not share a common origin for their measurements. This may introduce errors in the direct conversion from Cartesian points to spherical points, resulting in incorrect range-view data distribution and thus low-quality range images, as shown at the top of Fig. 3.", + "bbox": [ + 212, + 685, + 787, + 761 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address this issue, we adopt Hough Voting to estimate heights and pitch angles $\\{h_j,\\phi_j\\}_{j = 1,\\dots ,N}$ for Velodyne sensors [3]. We then adjust the point cloud transformation to a range image using", + "bbox": [ + 212, + 761, + 787, + 808 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr = \\sqrt {x ^ {2} + y ^ {2} + (z - h _ {j}) ^ {2}}, \\theta = \\mathrm {a t a n} (y, x), \\phi = \\phi_ {j}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 818, + 785, + 844 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "RangeLDM: Fast Realistic LiDAR Point Cloud Generation", + "bbox": [ + 338, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6a8c524874c6ef0c93606503b68a308b8287dbb1a00c4117a9abdee57d58ca3e.jpg", + "image_caption": [ + "Fig. 3: Comparison of range projection by the typical method described in Eq. 2 and our method with Hough Voting as in Eq. 3." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 529, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c1c194f7f013d7decc561a9f467f07ec1311cdf6a8edfb270d0c0e7f7b2d0307.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 545, + 148, + 782, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $h_j$ and $\\phi_j$ refer to the $j$ -th Velodyne sensor.", + "bbox": [ + 215, + 277, + 580, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We then rasterize points $(r,\\theta ,\\phi)$ into a 2D cylindrical projection $R(u,v)$ (a.k.a., range image) of size $H\\times W$ with $u = ((\\theta +\\pi) / 2\\pi)W,v = j$ , where $(u,v)$ denotes the grid coordinate of a point in the range image. Thus, we obtain high-quality range images as illustrated at the bottom of Fig. 3. We denote the obtained range image as $\\mathbf{x}\\in \\mathbb{R}^{H\\times W\\times 2}$ , which comprises $H\\times W$ pixels associated with both range and intensity $\\{r,i\\}$ .", + "bbox": [ + 212, + 294, + 784, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To verify the impact of range projection, we trained range-based LiDAR generation methods like LiDARGen [95] with our obtained high-quality range images. Table 1 shows substantial improvement, which sheds light on the significance of the correct range-view data distribution for such range-based generative model learning.", + "bbox": [ + 212, + 383, + 784, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Training", + "text_level": 1, + "bbox": [ + 215, + 484, + 333, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As a range image is often of high dimensionality at the scale of $64 \\times 1024$ , it would be computationally expensive to learn the distribution directly. Instead, we compress the range image into a low-dimensional latent space via a VAE and then model the distribution via an LDM. Consequently, the training of the proposed RangeLDM includes two distinct stages. In the first stage, we train a regular VAE to compress the range image into latent features $\\mathbf{z}_0 \\in \\mathbb{R}^{h \\times w \\times c}$ by a downsampling factor $f = H / h = W / w$ . In the second stage, we train an LDM to learn the distribution of the latent encoding $\\mathbf{z}_0$ .", + "bbox": [ + 212, + 512, + 784, + 632 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "First-Stage: Dimensionality Reduction. In the first stage, we reduce the dimensionality of the obtained range images by a VAE, which removes imperceptible high-frequency details and leads to low-dimensional latent features that encode prominent information in the original range image.", + "bbox": [ + 212, + 633, + 784, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A standard VAE consists of two main components: 1) an encoder $\\mathcal{E}_{\\zeta}$ that transforms an input range image $\\mathbf{x} \\in \\mathbb{R}^{H \\times W \\times 2}$ into latent features $\\mathbf{z}_0 = \\mathcal{E}_{\\zeta}(\\mathbf{x}) \\in \\mathbb{R}^{h \\times w \\times c}$ and 2) a decoder $\\mathcal{G}_{\\eta}$ that takes the latent features $\\mathbf{z}_0$ as input and generates the reconstructed range image $\\hat{\\mathbf{x}} = \\mathcal{G}_{\\eta}(\\mathbf{z}_0)$ . The entire VAE is trained by maximizing a modified ELBO with respect to the parameters $\\zeta$ and $\\eta$ [32,54]:", + "bbox": [ + 212, + 694, + 784, + 771 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {E L B O} (\\zeta , \\eta) = \\mathbb {E} _ {q _ {\\zeta} (\\mathbf {z} _ {0} | \\mathbf {x})} \\left[ \\log p _ {\\boldsymbol {\\eta}} (\\mathbf {x} \\mid \\mathbf {z} _ {0}) \\right] - \\lambda_ {1} D _ {\\mathrm {K L}} \\left(q _ {\\zeta} (\\mathbf {z} _ {0} \\mid \\mathbf {x}) \\| p (\\mathbf {z} _ {0})\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 259, + 785, + 784, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the above equation, the first term evaluates the reconstruction likelihood of the decoder from the latent $\\mathbf{z}_0$ and corresponds to an $L_{1}$ reconstruction loss,", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Q. Hu et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "while the second term quantifies how closely the learned distribution of $\\mathbf{z}_0$ resembles the prior distribution $p(\\mathbf{z}_0) = \\mathcal{N}(\\mathbf{z}_0; \\mathbf{0}, \\mathbf{I})$ held over latent variables. The hyperparameter $\\lambda_1$ balances the reconstruction accuracy and Kullback-Leibler regularization.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Range-Guided Discriminator. In order to mitigate the blurriness introduced by relying solely on the pixel-space $L_{1}$ reconstruction loss, we integrate the VAE with a patch-based adversarial discriminator [15, 29, 55, 87]. Furthermore, to ensure that the reconstruction remains confined within the manifold underlying the range image and exploit the geometric information from the spherical coordinates, we propose a range-guided discriminator $\\mathcal{D}_{\\tau}$ , with parameters $\\tau$ . Specifically, we adapt the Meta-Kernel [18] to replace the standard convolution in the discriminator, aiming to learn convolution weights from relative spherical coordinates. The Meta-Kernel is formulated as", + "bbox": [ + 212, + 207, + 784, + 340 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} _ {i} ^ {\\prime} = \\mathcal {W} \\left(\\mathcal {A} _ {j \\in \\mathcal {N} (i)} \\left(\\Phi (\\gamma (\\mathbf {p} _ {j}, \\mathbf {p} _ {i})) \\odot \\mathbf {h} _ {j}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 349, + 784, + 383 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{h}$ and $\\mathbf{h}'$ represent the input feature and output feature, respectively. The function $\\varPhi$ denotes a Multi-Layer Perceptron (MLP) with two fully-connected layers, $\\mathcal{A}$ is a concatenation operation, $\\mathcal{N}(i)$ denotes the 2D-grid neighborhood around the center pixel $i$ , and $\\mathcal{W}$ is a fully-connected layer. The term $\\gamma (\\mathbf{p}_j,\\mathbf{p}_i)$ represents the distance between points $\\mathbf{p}_j$ and $\\mathbf{p}_i$ in spherical coordinates, which is expressed as", + "bbox": [ + 214, + 393, + 787, + 483 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\gamma \\left(\\mathbf {p} _ {j}, \\mathbf {p} _ {i}\\right) := \\left\\{r _ {j} \\cos (\\Delta \\theta) \\cos (\\Delta \\phi) - r _ {i}, r _ {j} \\cos (\\Delta \\theta) \\sin (\\Delta \\phi), r _ {j} \\sin (\\Delta \\theta) \\right\\}, \\\\ \\Delta \\theta = \\theta_ {j} - \\theta_ {i}, \\quad \\Delta \\phi = \\phi_ {j} - \\phi_ {i}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 243, + 494, + 738, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The Meta-Kernel is aware of local 3D structures by adaptively adjusting convolution kernel weights from relative spherical coordinates. This makes it challenging for the decoder to deceive the range-guided discriminator. This, in turn, encourages the decoder to generate more realistic range images.", + "bbox": [ + 212, + 537, + 782, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With the range-guided discriminator, the final objective of the first-stage training is:", + "bbox": [ + 212, + 599, + 782, + 628 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\zeta , \\eta = \\arg \\min _ {\\zeta , \\eta} \\max _ {\\tau} \\mathcal {L} _ {E L B O} (\\zeta , \\eta) + \\lambda_ {2} \\mathcal {L} _ {a d v} (\\zeta , \\eta , \\tau), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 641, + 784, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{adv}$ is the GAN Hinge loss [39] and $\\lambda_{2}$ balances the two losses.", + "bbox": [ + 212, + 672, + 720, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Second-Stage: Latent Diffusion Modeling In the second stage, we freeze the encoder and decoder of the learned VAE and train an LDM on the encoded latent feature $\\mathbf{z}_0$ for distribution modeling of the range image. During the diffusion process, noise is added to the initial latent $\\mathbf{z}_0$ , resulting in a noisy latent $\\mathbf{z}_t$ , where the noise level increases over time steps $t \\in T$ .", + "bbox": [ + 212, + 688, + 784, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Regarding the modeling of the unconditional distribution $p(\\mathbf{z}_0)$ , based on Eq. 1, we learn a time-conditional UNet [56] $\\epsilon_{\\theta}(z_t,t)$ that predicts the noise added to the noisy latent $\\mathbf{z}_t$ :", + "bbox": [ + 212, + 763, + 782, + 808 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {L D M}} := \\mathbb {E} _ {\\mathcal {E} (x), \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t\\right) \\| _ {2} ^ {2} \\right]. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 348, + 819, + 784, + 844 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "RangeLDM: Fast Realistic LiDAR Point Cloud Generation", + "bbox": [ + 338, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As to the modeling of the conditional distribution $p(\\mathbf{z}_0|\\mathbf{y})$ , where $\\mathbf{y}$ is a condition such as sparse or masked point clouds, we learn a conditional LDM via", + "bbox": [ + 212, + 146, + 782, + 189 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {L D M}} := \\mathbb {E} _ {\\mathcal {E} (x), \\mathbf {y}, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {z} _ {t}, t, \\tau_ {\\theta} (\\mathbf {y})) \\| _ {2} ^ {2} \\right], \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 191, + 785, + 217 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\tau_{\\theta}$ is a condition encoder that projects $\\mathbf{y}$ into an intermediate representation $\\tau_{\\theta}(\\mathbf{y})$ , and $\\epsilon_{\\theta}(\\mathbf{z}_t, t, \\tau_{\\theta}(\\mathbf{y}))$ is a UNet interpolated with cross-attention layers [55, 75].", + "bbox": [ + 212, + 227, + 782, + 273 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3 Generation", + "text_level": 1, + "bbox": [ + 215, + 304, + 354, + 318 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unconditional Generation. With the decoder $\\mathcal{G}_{\\eta}$ and the LDM $\\epsilon_{\\theta}$ , we establish a hierarchical generative model $p_{\\eta, \\theta}(\\mathbf{x}, \\mathbf{z}_0) = p_{\\eta}(\\mathbf{x} | \\mathbf{z}_0) p_{\\theta}(\\mathbf{z}_0)$ . That is, we first generate latent features with the LDM, and then map the latent features back to the original range image space using the decoder $\\mathcal{D}_{\\eta}$ for LiDAR point cloud generation.", + "bbox": [ + 212, + 337, + 782, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conditional Generation. Given a condition $\\mathbf{y}$ , the conditional generative model is defined as $p_{\\eta, \\theta}(\\mathbf{x}, \\mathbf{z}_0 | \\mathbf{y}) = p_{\\eta}(\\mathbf{x} | \\mathbf{z}_0, \\mathbf{y}) p_{\\theta}(\\mathbf{z}_0 | \\mathbf{y})$ . There could be various conditions for LiDAR point cloud generation, such as sparse point clouds, incomplete point clouds, camera or even text. Here, we illustrate two applications of conditional generation: LiDAR point cloud upsampling and inpainting.", + "bbox": [ + 212, + 415, + 782, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "LiDAR Point Cloud Upsampling. This task takes a sparse point cloud $^1$ $\\mathbf{y} \\in \\mathbb{R}^{h \\times W \\times 2}$ as input (following the setting of LiDARGen) and expects the model to produce a denser counterpart $\\hat{\\mathbf{x}} \\in \\mathbb{R}^{H \\times W \\times 2}$ , which is necessary for accurate LiDAR-based perception and density-insensitive domain adaptation. We train LDM following Eq. 9 with the sparse point cloud $\\mathbf{y}$ and its ground-truth dense variant $\\mathbf{x}$ . The condition encoder $\\tau_{\\theta}$ is configured as a reshaper that transforms the sparse point cloud $\\mathbf{y}$ into $\\mathbf{y}' \\in \\mathbb{R}^{h \\times w \\times 2f}$ . The reshaped $\\mathbf{y}'$ is then concatenated with noisy latent variable $\\mathbf{z}_t$ and fed into the UNet $\\epsilon_{\\theta}$ for noise prediction.", + "bbox": [ + 212, + 491, + 784, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "LiDAR Point Cloud Inpainting. This task is required when LiDAR sensors cannot capture the entire scene due to object occlusion or sensor limitations. We train an inpainting LDM from the ground truth point cloud $\\mathbf{x}$ , a masked point cloud $\\mathbf{x}'$ , and the corresponding mask $\\mathbf{m} \\in \\{0,1\\}^{H \\times W}$ . The masked point cloud $\\mathbf{x}'$ is encoded with the encoder $\\mathcal{E}_{\\zeta}$ from the VAE and is then concatenated with a downsampled mask $\\mathbf{m}' \\in \\{0,1\\}^{h \\times w}$ as the condition term in Eq. 9. Mathematically, we represent the condition $\\mathbf{y}$ as $\\{\\mathbf{x}',\\mathbf{m}\\}$ , and deploy the condition encoder to project it into $\\tau_{\\theta}(\\mathbf{y}) = \\mathcal{E}_{\\zeta}(\\mathbf{x}') \\| ds(\\mathbf{m})$ , where $\\|\\cdot\\|$ denotes the concatenation operation and $ds(\\cdot)$ indicates the downsampling operation. The projected condition is then concatenated with the noisy latent variable $\\mathbf{z}_t$ and input into the UNet $\\epsilon_{\\theta}$ for noise prediction.", + "bbox": [ + 212, + 628, + 785, + 796 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Q. Hu et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "1 Without causing ambiguity, the point clouds are converted to range images by default.", + "bbox": [ + 217, + 810, + 782, + 838 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a928bec7ff81d9dcf368aa8a104b347ccb364abccb06399f799be61df6d7c880.jpg", + "image_caption": [ + "Fig. 4: Qualitative results comparing against baselines for unconditional Li-DAR generation on KITTI-360. Real point clouds are only for reference. Our model produces results that closely resemble real-world data, which excels in generating road scenes, such as cars (the first row), road bollards (the second row) and crossroads (the last row)." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 782, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.4 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 362, + 455, + 376 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Circular Convolution. Range images are inherently circular, which means that the left boundary of a range image is actually connected to its right boundary. However, standard convolutions employ zero-padding or symmetry padding and do not take into account such constraints. Thus, as in LiDArgen [95], we replace standard convolutions in both the VAE and LDM with circular convolutions [59], which treat the left and right boundaries as connected neighbors in the topology. Since the 2D convolution operator satisfies translation invariance, it is evident that the entire network is invariant to horizontal shifting of the range image (i.e., rotation in the xy-plane of the point cloud).", + "bbox": [ + 212, + 386, + 784, + 521 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Conditioning the Model with Direction. Real LiDAR point clouds collected from driving scenes usually have a certain directionality. For example, since vehicles typically follow the direction of the road, the x-axis direction of the LiDAR point cloud usually aligns with the road direction. However, since the whole network of our model is invariant to rotation in the xy-plane of the point cloud, the point clouds generated from random noise could result in arbitrary directions. To avoid this, we generate point clouds with directional conditioning. Specifically, we use an $h \\times w$ matrix as the condition, in which only the values from the first column are set to 1, while the rest are all set to 0. This 1-valued column corresponds to the x-axis of the point cloud. Such simple conditioning demonstrates effective control over the direction of the generated point cloud, as presented in the qualitative results of the supplementary material.", + "bbox": [ + 212, + 522, + 784, + 703 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Network Architectures and Model Hyperparameters. For simplicity, we adopted a similar architecture to that in [55]. The downsampling factor $f$ is set as 4 to strike a good balance between efficiency and realistic results. The VAE comprises three encoder blocks and three decoder blocks. The LDM consists of four downsampling blocks with the last three followed by a transformer layer, and four upsampling blocks with the first three blocks also followed by a transformer layer. In the generation process, we employed 50 denoising steps with DDIM sampler [66] for point cloud generation. More implementation details can be found in released code https://github.com/WoodwindHu/RangeLDM.", + "bbox": [ + 212, + 704, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "RangeLDM: Fast Realistic LiDAR Point Cloud Generation", + "bbox": [ + 339, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c1f715f0ec81ea3370f4aec310bbc2f0abae528348e19494642888e0efad7f3b.jpg", + "image_caption": [ + "Fig. 5: Comparison of upsampling results on KITTI-360. We downsampled the ground truth by a factor of four as the input, and demonstrated the results of different methods on $4 \\times$ -upsampling of the input." + ], + "image_footnote": [], + "bbox": [ + 217, + 146, + 785, + 296 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 345, + 374, + 363 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 377, + 426, + 393 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. We evaluate our model on two challenging datasets, KITTI-360 and nuScenes. KITTI-360 has 81,106 LiDAR readings from 9 sequences in Germany, covering diverse scenes. We used the first two sequences for validation and trained on the rest. NuScenes is a public dataset with 297,737 LiDAR sweeps in the training set and 52,423 in the testing set, collected in Boston and Singapore.", + "bbox": [ + 212, + 401, + 785, + 477 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation metrics. Following [95], we employed three metrics to perform quantitative analysis: Maximum Mean Discrepancy (MMD), Jensen-Shannon divergence (JSD), and Frechet Range Distance (FRD score). We use a $100 \\times 100$ 2D histogram on the BEV plane to calculate MMD and JSD metrics. The Frechet Range Distance (FRD) score [95] is a metric used to evaluate the quality of samples acquired by a generative model, inspired by the FID score for images [24]. To compute the FRD score, we use RangeNet++ [45], an encoder-decoder-based network for segmentation, which is pre-trained on KITTI-360.", + "bbox": [ + 212, + 478, + 785, + 598 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baselines. We evaluated our approach for LiDAR point cloud generation against several competitive methods, including LiDAR VAE [5], LiDAR GAN [5], Projected GAN [58], LiDARGen [95] and UltraLiDAR [79].", + "bbox": [ + 212, + 598, + 785, + 643 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Unconditional Generation", + "text_level": 1, + "bbox": [ + 215, + 666, + 475, + 679 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/9ae5726bb744bd142bcfbd8751aaf882cabed043b96c1fd26ca8f891f5a581a1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodYears\\( MMD_{BEV} \\)↓FRD ↓\\( JSD_{BEV} \\)↓
LiDAR GAN [5]IROS 2019\\( 3.06 \\times 10^{-3} \\)3003.8-
LiDAR VAE [5]IROS 2019\\( 1.00 \\times 10^{-3} \\)2261.50.161
Projected GAN [58]NeurIPS 2021\\( 3.47 \\times 10^{-4} \\)2117.20.085
LiDARGen [95]ECCV 2022\\( 3.87 \\times 10^{-4} \\)2040.10.067
UltraLiDAR [79]CVPR 2023\\( 1.96 \\times 10^{-4} \\)-0.071
Ours\\( 3.07 \\times 10^{-5} \\)1074.90.045
", + "bbox": [ + 305, + 686, + 697, + 777 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 2: Unconditional generation results on KITTI-360 [38].", + "bbox": [ + 292, + 777, + 707, + 791 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Quantitative Results on KITTI-360. Table 2 displays the quantitative results of the proposed RangeLDM and competing algorithms. The results clearly", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Q. Hu et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "illustrate that our method significantly outperforms the baselines across all metrics. Notably, our results on the MMD metric are an order of magnitude lower than those of other methods. This underscores the remarkable expressive capabilities of our model in the context of LiDAR point cloud generation.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Qualitative Results on KITTI-360. Figure 4 displays a set of randomly generated samples from competing algorithms, alongside real point cloud samples extracted from the KITTI-360 dataset for comparison. We observe that the Projected GAN [58] exhibits noticeable artifacts in the distant range. LiDARGen [95] effectively captures the general layout, but produces rather noisy outputs and lacks the same degree of straight walls as actual samples. UltraLiDAR [79] generates structured and reasonable scenes, yet it only creates voxelized point clouds (as magnified in the first row) without intensity features. In contrast, our model consistently surpasses the baseline models and yields results closely resembling real-world data. For example, we can generate cars (emphasized by red boxes in the first row of Figure 4) and road bollards (red boxes in the second row of Figure 4) on the road. Additionally, our model generates various scenes such as straightway (the first two rows of Figure 4) and crossroads (the last row of Figure 4). The supplementary materials provide more visualization results.", + "bbox": [ + 212, + 209, + 787, + 421 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Human study on KITTI-360. To assess the perceptual quality, we conducted an A/B test involving a group of 14 researchers with LiDAR expertise. Following the same evaluation system as [95] and [79], we present pairs of randomly selected images", + "bbox": [ + 212, + 422, + 486, + 529 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/db4c042a681e0ac93dd682265290e759d1640b550b45b8980b803d4394592926.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPercent prefer ours
Ours vs. VAE [5]98.8%
Ours vs. GAN [5]98.0%
Ours vs. ProjectedGAN [58]93.2%
Ours vs. LiDARGen [95]90.2%
Ours vs. UltraLiDAR [79]86.5%
", + "bbox": [ + 503, + 426, + 781, + 503 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "from two point clouds and ask participants to determine which one appeared more realistic. The results, as displayed in Table 3, unequivocally demonstrate the superior visual quality of generation results by our model. In most of the cases, testers favored our results over the baselines.", + "bbox": [ + 212, + 529, + 787, + 589 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation on the NuScenes Dataset. The results listed in Table 4 demonstrate that our model outperforms LiDAR VAE and LiDARGen significantly over nuScenes. In particular, our results on the MMD metric are an order of magnitude lower than those of other methods, while we reduce to about $1/3$ of the", + "bbox": [ + 212, + 593, + 542, + 698 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "results of competitive methods in terms of the JSD metric. Also, Figure 6 illustrates the superiority of our model over LiDAR VAE and LiDARGen in terms of qualitative comparison.", + "bbox": [ + 212, + 699, + 782, + 743 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/41bab5708cb0874f32ee37fcf4c282c2aed1d95d301caf60073acdf91628d1b7.jpg", + "table_caption": [ + "Table 3: Human study on KITTI-360." + ], + "table_footnote": [], + "table_body": "
Method\\( MMD_{BEV} \\)↓\\( JSD_{BEV} \\)↓
LiDAR VAE [5]\\( 1.1 \\times 10^{-3} \\)-
\\( LiDARGen^† \\)[95]\\( 1.9 \\times 10^{-3} \\)0.160
Ours\\( 1.9 \\times 10^{-4} \\)0.054
", + "bbox": [ + 555, + 595, + 782, + 647 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 4: Unconditional generation results on nuScenes dataset [6]. †: Reproduced by us.", + "bbox": [ + 553, + 647, + 784, + 688 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Model Size. The number of parameters in LiDARGen [95] and UltraLiDAR [79] are 29.7M and 40.3M, respectively. To ensure fair competition, we limited the capacity of our model to be similar to other approaches. Our model comprises 41.4M parameters, consisting of 12.7M parameters for VAE and 28.7M parameters for LDM. As illustrated in Table 2, our model outperforms the baselines by a large margin under similar model sizes.", + "bbox": [ + 212, + 746, + 787, + 838 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "RangeLDM: Fast Realistic LiDAR Point Cloud Generation", + "bbox": [ + 338, + 113, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f95bf65998d7447a178c9b8f34403bdb06caba45f1541d1d6839153fb0bd9016.jpg", + "image_caption": [ + "Fig. 6: Qualitative results comparing against LiDAR VAE and LiDARGen for unconditional LiDAR generation on nuScenes. Real point clouds are only for reference." + ], + "image_footnote": [], + "bbox": [ + 303, + 143, + 699, + 253 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/53f8439d4f3cbdbd3bda733ce4cdbf5643e427800f3f4fe151dc37d4641133ac.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Hough Range-Guided \nVoting DiscriminatorCircular \nConvolutionDirection \nConditionedMMDBEV ↓FRD ↓JSDBEV ↓
(a)××××3.95 × 10-41536.70.067
(b)✓×××6.57 × 10-51229.30.056
(c)✓✓××4.72 × 10-51103.10.051
(d)✓✓✓×3.90 × 10-41797.20.078
(e)✓✓✓✓3.07 × 10-51074.90.045
", + "bbox": [ + 250, + 287, + 753, + 380 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Generation Efficiency. The proposed dimensionality reduction offers RangeLDM a considerable advantage in terms of generation speed. As indicated in Table 5, the proposed method significantly outperforms LiDARGen [95] and UltraLi-DAR [79] in terms of generation speed. In particular, our model is 200 times faster than LiDARGen and 30 times faster than UltraLi-DAR.", + "bbox": [ + 215, + 412, + 571, + 534 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/cf3634e45c8488387553b6321675678312548ae926b19bb0f47453f9a84313d2.jpg", + "table_caption": [ + "Table 6: Main ablation study." + ], + "table_footnote": [], + "table_body": "
MethodThroughput ↑ (samples/s)
LiDARGen [95]0.02
UltraLiDAR [79]0.16
Ours4.86
", + "bbox": [ + 586, + 416, + 782, + 481 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Conditional Generation", + "text_level": 1, + "bbox": [ + 215, + 556, + 455, + 571 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To evaluate the conditional generation performance of RangeLDM, we conducted experiments on the KITTI-360 dataset over two tasks: LiDAR point cloud upsampling and inpainting.", + "bbox": [ + 212, + 583, + 784, + 627 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "LiDAR Point Cloud Upsampling. We obtain sparse input point clouds by selecting a subset of 16 beams from the raw 64-beam sensors, in line with LiDARGen. We compared our approach against PUNet [88], DeepRS [9], Grad-PU [23], bicubic interpolation, Nearest Neighbor (NN) interpolation and LiDARGen.", + "bbox": [ + 212, + 628, + 784, + 688 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Quantitatively, in addition to measuring Mean Absolute Error (MAE) in the range view, we also employed RangeNet++ semantic segmentation to evaluate the quality of upsampled results. As shown in Table 7, the proposed method outperforms competing methods in all metrics, including", + "bbox": [ + 215, + 689, + 496, + 809 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/966fd642ac6b200f9e8a89c7a312cbc7d2f014c748a05400bc3695d789b8b0d6.jpg", + "table_caption": [ + "Table 5: Inference speed on a single RTX 3090 GPU." + ], + "table_footnote": [], + "table_body": "
MethodYearsMAE ↓Accuracy ↑ IoU ↑
PUNet [88]20186.88--
DeepRS [9]20223.96--
Grad-PU [23]20235.09--
Bicubic-2.600.2650.166
NN-2.180.5460.394
LiDARGen [95]20221.230.6080.449
Ours0.890.7220.566
", + "bbox": [ + 503, + 693, + 782, + 792 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 7: LiDAR upsampling.", + "bbox": [ + 540, + 792, + 741, + 806 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "MAE, per-point segmentation accuracy, and segmentation Intersection over Union (IoU).", + "bbox": [ + 215, + 810, + 784, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Q. Hu et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/cc21bf27bd40cbe214ff1e23d79ef5777a37a5bfaaeaef5548d3b3ab4716a33d.jpg", + "image_caption": [ + "Fig. 7: Inpainting results. Left: ground truth point clouds. Middle and right (red): input point clouds. Middle and right (blue): recovered point clouds." + ], + "image_footnote": [], + "bbox": [ + 303, + 143, + 702, + 241 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/59aec08343dbaccee62d95b22d4da83d42eda9f7fa840c17a9bb4fc184e794d2.jpg", + "image_caption": [ + "Fig. 8: VAE reconstruction results. The proposed VAE with a range-guided discriminator reconstructs point clouds with less noise and more precise object structures." + ], + "image_footnote": [], + "bbox": [ + 225, + 276, + 781, + 333 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Qualitatively, the visualization comparison with several competitive methods presented in Figure 5 indicates that the proposed upsampling method exhibits results highly similar to the ground truth data, showing clear object details and consistent LiDAR scan lines. We provide additional visualization results in the supplementary materials.", + "bbox": [ + 212, + 375, + 784, + 448 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "LiDAR Point Cloud Inpainting. To mimic the case of missing data in LiDAR point clouds, we mask point clouds in front of the vehicle within a $22.5^{\\circ}$ range and perform inpainting using LiDARGen and our method. Our results achieve much less MAE (0.190) between the reconstructed results and ground truth compared to that of LiDARGen (0.367). As shown in Figure 7, LiDARGen generates roads but fails to recover the masked car (the first row) and the wall (the second row), while we are able to reconstruct the details. More results are presented in the supplementary materials.", + "bbox": [ + 212, + 450, + 785, + 571 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 598, + 387, + 614 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "All ablation studies are conducted on the KITTI-360 dataset.", + "bbox": [ + 214, + 627, + 658, + 641 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Main Ablation. As demonstrated in Table 6, we investigate the contribution of each component. Starting from the backbone model (a) without any component, we gradually add each component for evaluation. By applying Hough Voting to the training, the performance of variant (b) improves significantly, demonstrating the huge impact of the correct range-view data distribution. With the addition of the range-guided discriminator component, model (c) achieves better performance, thus clarifying the effectiveness of the geometry-sensitive module. Compared to model (c), the performance of model (d) decreases because circular convolution generates LiDAR scenes with arbitrary directions due to rotation invariance. However, when combined with direction-conditioned generation that guides the model to generate point clouds in the correct direction, the overall performance improves, as demonstrated by model (e). We refer the readers to the supplementary materials for qualitative results.", + "bbox": [ + 212, + 643, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "RangeLDM: Fast Realistic LiDAR Point Cloud Generation", + "bbox": [ + 338, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/f8e3e089e149013022e75c2fff79a2d9c560caa6af846c065f95823e2d440412.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelVAEModel Size (M)\\( MMD_{BEV} \\)↓FRD ↓\\( JSD_{BEV} \\)↓Throughput ↑(samples/s)
Ours-DMS×40.3\\( 1.03 × 10^{-4} \\)1392.20.0621.52
Ours-DM×114.7\\( 4.14 × 10^{-5} \\)899.00.0400.64
Ours✓41.4\\( 3.07 × 10^{-5} \\)1074.90.0454.86
", + "bbox": [ + 277, + 143, + 725, + 210 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Latent diffusion. We explore the contribution of latent diffusion to the generation quality and generation speed of the model. As shown in Table 8, we constructed two variants of different sizes that directly generate range images with diffusion models, denoted as \"Ours-DMS\" (with a smaller size) and \"Ours-DM\" (with a larger size), for comparison. It turns out that a much larger model size is required for \"Ours-DM\" to achieve competitive performance with \"Ours\", due to the multitude of high-frequency details contained in the range image. Additionally, the superior throughput of \"Ours\" demonstrates the efficiency of latent diffusion.", + "bbox": [ + 212, + 272, + 787, + 407 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "VAE Architecture. We investigate the contribution of the proposed range-guided discriminator by comparing our VAE with two baselines: 1) VAE without dis", + "bbox": [ + 215, + 409, + 459, + 483 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/7fea9b7e1fffd3289957b2731578b6a4f951434ef1dec4b73998cd67d4aa3e86.jpg", + "table_caption": [ + "Table 8: Ablation of latent diffusion. Model without latent diffusion requires more parameters to achieve similar performance with the latent diffusion model. We generated 1000 samples for calculating the throughput." + ], + "table_footnote": [], + "table_body": "
DiscriminatorPSNRrange↑ MAE↓ FRD↓ CD
No26.770.0195532.90.0808
Vanilla26.700.0189496.70.0726
Range-Guided27.190.0186483.60.0676
", + "bbox": [ + 475, + 412, + 784, + 465 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 9: Ablation on the VAE architecture.", + "bbox": [ + 478, + 467, + 777, + 477 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "criminalitor; and 2) VAE with vanilla discriminator (i.e., using the original 2D convolution in the discriminator). We evaluate the reconstruction performance of VAEs with 2D and 3D metrics, including PSNR, MAE, FRD and Chamfer Distance (CD) [16]. As listed in Table 9, the VAE with the proposed range-guided discriminator outperforms the baselines in both 2D and 3D reconstruction metrics. This gives credits to the ability of perceiving local 3D structures by the range-guided discriminator. The qualitative results presented in Figure 8 also indicate that the proposed range-guided discriminator ensures the point clouds are reconstructed with less noise and more precise object structures.", + "bbox": [ + 212, + 484, + 787, + 619 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 642, + 369, + 657 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We propose a novel RangeLDM model to generate realistic range-view LiDAR point clouds at a fast speed. We ensure the quality of projection from point clouds to range images with correct distribution via Hough Voting. Then we compress range images to latent features with a VAE, and train a diffusion model in the lower-dimensional latent space. Additionally, we enhance the range-image reconstruction quality of the VAE with a range-guided discriminator. Experiments conducted on KITTI-360 and nuScenes datasets demonstrate the superior generation quality and sampling efficiency of our method. In future, we will explore generating labeled point clouds and creating corner-case data, such as in car accidents and extreme weather conditions, for potential applications in robust self-driving.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Q. Hu et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Achlioptas, P., Diamanti, O., Mitlagkas, I., Guibas, L.: Learning representations and generative models for 3D point clouds. In: International Conference on Machine Learning. pp. 40-49. PMLR (2018)", + "2. Ando, A., Gidaris, S., Bursuc, A., Puy, G., Boulch, A., Marlet, R.: RangeViT: Towards vision transformers for 3D semantic segmentation in autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5240-5250 (2023)", + "3. Bewley, A., Sun, P., Mensink, T., Anguelov, D., Sminchisescu, C.: Range conditioned dilated convolutions for scale invariant 3D object detection. In: Conference on Robot Learning (2020)", + "4. Box, G.E., Jenkins, G.M., Reinsel, G.C., Ljung, G.M.: Time series analysis: forecasting and control. John Wiley & Sons (2015)", + "5. Caccia, L., Van Hoof, H., Courville, A., Pineau, J.: Deep generative modeling of LiDAR data. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 5034-5040. IEEE (2019)", + "6. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: NuScenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11621-11631 (2020)", + "7. Cai, R., Yang, G., Averbuch-Elor, H., Hao, Z., Belongie, S., Snavely, N., Hariharan, B.: Learning gradient fields for shape generation. In: European Conference on Computer Vision. pp. 364-381 (2020)", + "8. Chai, Y., Sun, P., Ngiam, J., Wang, W., Caine, B., Vasudevan, V., Zhang, X., Anguelov, D.: To the point: Efficient 3D object detection in the range image with graph convolution kernels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16000-16009 (2021)", + "9. Chen, H., Luo, S., Hu, W., et al.: Deep point set resampling via gradient fields. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(3), 2913-2930 (2022)", + "0. Chen, X., Vizzo, I., Läbe, T., Behley, J., Stachniss, C.: Range image-based LiDAR localization for autonomous vehicles. In: IEEE International Conference on Robotics and Automation (ICRA). pp. 5802-5808. IEEE (2021)", + "1. Cheng, H., Han, X., Xiao, G.: CeNet: Toward concise and efficient LiDAR semantic segmentation for autonomous driving. In: IEEE International Conference on Multimedia and Expo (ICME). pp. 01-06. IEEE (2022)", + "2. Cortinhal, T., Tzelepis, G., Erdal Aksoy, E.: SalsaNext: Fast, uncertainty-aware semantic segmentation of LiDAR point clouds. In: Advances in Visual Computing. pp. 207-222. Springer International Publishing, Cham (2020)", + "3. Deng, J., Shi, S., Li, P., Zhou, W., Zhang, Y., Li, H.: Voxel R-CNN: Towards high performance voxel-based 3D object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1201-1209 (2021)", + "4. Engelcke, M., Rao, D., Wang, D.Z., Tong, C.H., Posner, I.: Vote3Deep: Fast object detection in 3D point clouds using efficient convolutional neural networks. In: 2017 IEEE International Conference on Robotics and Automation (ICRA). pp. 1355-1361. IEEE (2017)", + "5. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12873-12883 (2021)" + ], + "bbox": [ + 225, + 175, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "RangeLDM: Fast Realistic LiDAR Point Cloud Generation", + "bbox": [ + 339, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Fan, H., Su, H., Guibas, L.J.: A point set generation network for 3D object reconstruction from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 605-613 (2017)", + "17. Fan, L., Pang, Z., Zhang, T., Wang, Y.X., Zhao, H., Wang, F., Wang, N., Zhang, Z.: Embracing single stride 3D object detector with sparse transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8458-8468 (2022)", + "18. Fan, L., Xiong, X., Wang, F., Wang, N., Zhang, Z.: RangeDet: In defense of range view for LiDAR-based 3D object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2918-2927 (2021)", + "19. Gao, L., Wu, T., Yuan, Y.J., Lin, M.X., Lai, Y.K., Zhang, H.: TM-NET: Deep generative networks for textured meshes. ACM Transactions on Graphics (TOG) 40(6), 1-15 (2021)", + "20. Gao, L., Yang, J., Wu, T., Yuan, Y.J., Fu, H., Lai, Y.K., Zhang, H.: SDM-NET: Deep generative network for structured deformable mesh. ACM Transactions on Graphics (TOG) 38(6), 1-15 (2019)", + "21. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020)", + "22. He, C., Li, R., Li, S., Zhang, L.: Voxel set transformer: A set-to-set approach to 3D object detection from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8417-8427 (2022)", + "23. He, Y., Tang, D., Zhang, Y., Xue, X., Fu, Y.: Grad-PU: Arbitrary-scale point cloud upsampling via gradient descent with learned distance functions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5354-5363 (2023)", + "24. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: GANs trained by a two time-scale update rule converge to a local nash equilibrium. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017)", + "25. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33, 6840-6851 (2020)", + "26. Hu, J.S., Kuai, T., Waslander, S.L.: Point density-aware voxels for LiDAR 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8469-8478 (2022)", + "27. Hu, P., Ziglar, J., Held, D., Ramanan, D.: What you see is what you get: Exploiting visibility for 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11001-11009 (2020)", + "28. Hui, K.H., Li, R., Hu, J., Fu, C.W.: Neural wavelet-domain diffusion for 3D shape generation. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)", + "29. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1125-1134 (2017)", + "30. Kim, H., Lee, H., Kang, W.H., Lee, J.Y., Kim, N.S.: SoftFlow: Probabilistic framework for normalizing flow on manifolds. Advances in Neural Information Processing Systems 33, 16388-16397 (2020)", + "31. Kim, J., Yoo, J., Lee, J., Hong, S.: SetVAE: Learning hierarchical composition for generative modeling of set-structured data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15059–15068 (2021)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Q. Hu et al.", + "bbox": [ + 271, + 114, + 352, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)", + "33. Kochanov, D., Nejadasl, F.K., Booij, O.: KprNet: Improving projection-based LiDAR semantic segmentation. arXiv preprint arXiv:2007.12668 (2020)", + "34. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: PointPillars: Fast encoders for object detection from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12697-12705 (2019)", + "35. Li, B.: 3D fully convolutional network for vehicle detection in point cloud. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1513-1518. IEEE (2017)", + "36. Li, B., Zhang, T., Xia, T.: Vehicle detection from 3D LiDAR using fully convolutional network. arXiv preprint arXiv:1608.07916 (2016)", + "37. Li, C., Ren, Y., Liu, B.: Pcgen: Point cloud generator for lidar simulation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 11676-11682. IEEE (2023)", + "38. Liao, Y., Xie, J., Geiger, A.: KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2D and 3D. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(3), 3292-3310 (2022)", + "39. Lim, J.H., Ye, J.C.: Geometric gan. arXiv preprint arXiv:1705.02894 (2017)", + "40. Litany, O., Bronstein, A., Bronstein, M., Makadia, A.: Deformable shape completion with graph convolutional autoencoders. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1886-1895 (2018)", + "41. Luo, S., Hu, W.: Diffusion probabilistic models for 3D point cloud generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2837-2845 (2021)", + "42. Malavazi, F.B., Guyonneau, R., Fasquel, J.B., Lagrange, S., Mercier, F.: LiDAR-only based navigation algorithm for an autonomous agricultural robot. Computers and Electronics in Agriculture 154, 71-79 (2018)", + "43. Manivasagam, S., Wang, S., Wong, K., Zeng, W., Sazanovich, M., Tan, S., Yang, B., Ma, W.C., Urtasun, R.: LiDARsim: Realistic LiDAR simulation by leveraging the real world. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11167-11176 (2020)", + "44. Meyer, G.P., Laddha, A., Kee, E., Vallespi-Gonzalez, C., Wellington, C.K.: Laser-Net: An efficient probabilistic 3D object detector for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12677-12686 (2019)", + "45. Milioto, A., Vizzo, I., Behley, J., Stachniss, C.: RangeNet++: Fast and accurate LiDAR semantic segmentation. In: IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 4213-4220. IEEE (2019)", + "46. Mittal, P., Cheng, Y.C., Singh, M., Tulsiani, S.: AutoSDF: Shape priors for 3D completion, reconstruction and generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 306-315 (2022)", + "47. Mo, K., Guerrero, P., Yi, L., Su, H., Wonka, P., Mitra, N., Guibas, L.J.: StructureNet: Hierarchical graph networks for 3D shape generation. arXiv preprint arXiv:1908.00575 (2019)", + "48. Nichol, A., Jun, H., Dhariwal, P., Mishkin, P., Chen, M.: Point-E: A system for generating 3D point clouds from complex prompts. arXiv preprint arXiv:2212.08751 (2022)", + "49. Pan, X., Xia, Z., Song, S., Li, L.E., Huang, G.: 3D object detection with pointformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7463-7472 (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "RangeLDM: Fast Realistic LiDAR Point Cloud Generation", + "bbox": [ + 338, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "50. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: PointNet: Deep learning on point sets for 3D classification and segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 652-660 (2017)", + "51. Ran, H., Guizilini, V., Wang, Y.: Towards realistic scene generation with lidar diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024)", + "52. Resop, J.P., Lehmann, L., Hession, W.C.: Drone laser scanning for modeling river-scape topography and vegetation: Comparison with traditional aerial LiDAR. Drones 3(2), 35 (2019)", + "53. Rezende, D., Mohamed, S.: Variational inference with normalizing flows. In: International Conference on Machine Learning. pp. 1530-1538. PMLR (2015)", + "54. Rezende, D.J., Mohamed, S., Wierstra, D.: Stochastic backpropagation and approximate inference in deep generative models. In: International Conference on Machine Learning. pp. 1278-1286. PMLR (2014)", + "55. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022)", + "56. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention. pp. 234-241. Springer (2015)", + "57. Sallab, A.E., Sobh, I., Zahran, M., Essam, N.: LiDAR sensor modeling and data augmentation with GANs for autonomous driving. arXiv preprint arXiv:1905.07290 (2019)", + "58. Sauer, A., Chitta, K., Müller, J., Geiger, A.: Projected GANs converge faster. Advances in Neural Information Processing Systems 34, 17480-17492 (2021)", + "59. Schubert, S., Neubert, P., Pöschmann, J., Protzel, P.: Circular convolutional neural networks for panoramic images and laser data. In: IEEE Intelligent Vehicles Symposium (IV). pp. 653-660. IEEE (2019)", + "60. Shi, G., Li, R., Ma, C.: PillarNet: Real-time and high-performance pillar-based 3D object detection. In: European Conference on Computer Vision. pp. 35-52. Springer (2022)", + "61. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: PV-RCNN: Point-voxel feature set abstraction for 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10529-10538 (2020)", + "62. Shi, S., Wang, X., Li, H.: PointRCNN: 3D object proposal generation and detection from point cloud. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 770-779 (2019)", + "63. Shi, S., Wang, Z., Shi, J., Wang, X., Li, H.: From points to parts: 3D object detection from point cloud with part-aware and part-aggregation network. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(8), 2647-2664 (2020)", + "64. Shi, W., Rajkumar, R.: Point-GNN: Graph neural network for 3D object detection in a point cloud. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1711-1719 (2020)", + "65. Shu, D.W., Park, S.W., Kwon, J.: 3D point cloud generative adversarial network based on tree structured graph convolutions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3859-3868 (2019)", + "66. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Q. Hu et al.", + "bbox": [ + 271, + 114, + 352, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "67. Sun, P., Tan, M., Wang, W., Liu, C., Xia, F., Leng, Z., Anguelov, D.: SWFormer: Sparse window transformer for 3D object detection in point clouds. In: European Conference on Computer Vision. pp. 426-442. Springer (2022)", + "68. Sun, P., Wang, W., Chai, Y., Elsayed, G., Bewley, A., Zhang, X., Sminchisescu, C., Anguelov, D.: RSN: Range sparse net for efficient, accurate LiDAR 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5725-5734 (2021)", + "69. Sun, Y., Wang, Y., Liu, Z., Siegel, J., Sarma, S.: PointGrow: Autoregressively learned point cloud generation with self-attention. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 61-70 (2020)", + "70. Tan, Q., Gao, L., Lai, Y.K., Xia, S.: Variational autoencoders for deforming 3D mesh models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5841-5850 (2018)", + "71. Tian, Z., Chu, X., Wang, X., Wei, X., Shen, C.: Fully convolutional one-stage 3D object detection on LiDAR range images. Advances in Neural Information Processing Systems 35, 34899-34911 (2022)", + "72. Vahdat, A., Kreis, K., Kautz, J.: Score-based generative modeling in latent space. Advances in Neural Information Processing Systems 34, 11287-11302 (2021)", + "73. Valsesia, D., Fracastoro, G., Magli, E.: Learning localized generative models for 3D point clouds via graph convolution. In: International Conference on Learning Representations (2018)", + "74. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in Neural Information Processing Systems 30 (2017)", + "75. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in Neural Information Processing Systems 30 (2017)", + "76. Wang, H., Shi, S., Yang, Z., Fang, R., Qian, Q., Li, H., Schiele, B., Wang, L.: RBGNet: Ray-based grouping for 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1110-1119 (2022)", + "77. Wang, Y., Fathi, A., Kundu, A., Ross, D.A., Pantofaru, C., Funkhouser, T., Solomon, J.: Pillar-based object detection for autonomous driving. In: European Conference on Computer Vision. pp. 18-34. Springer (2020)", + "78. Weiss, U., Biber, P.: Plant detection and mapping for agricultural robots using a 3D LiDAR sensor. Robotics and Autonomous Systems 59(5), 265-273 (2011)", + "79. Xiong, Y., Ma, W.C., Wang, J., Urtasun, R.: Learning compact representations for LiDAR completion and generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1074-1083 (2023)", + "80. Xu, C., Wu, B., Wang, Z., Zhan, W., Vajda, P., Keutzer, K., Tomizuka, M.: SqueezeSegV3: Spatially-adaptive convolution for efficient point-cloud segmentation. In: European Conference on Computer Vision. pp. 1-19. Springer (2020)", + "81. Yan Yan, Y.M., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018)", + "82. Yang, B., Luo, W., Urtasun, R.: Pixor: Real-time 3D object detection from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7652-7660 (2018)", + "83. Yang, G., Huang, X., Hao, Z., Liu, M.Y., Belongie, S., Hariharan, B.: PointFlow: 3D point cloud generation with continuous normalizing flows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4541-4550 (2019)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "RangeLDM: Fast Realistic LiDAR Point Cloud Generation", + "bbox": [ + 338, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "84. Yang, Z., Sun, Y., Liu, S., Jia, J.: 3DSSD: Point-based 3D single stage object detector. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11040-11048 (2020)", + "85. Yang, Z., Sun, Y., Liu, S., Shen, X., Jia, J.: STD: Sparse-to-dense 3D object detector for point cloud. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1951-1960 (2019)", + "86. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3D object detection and tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11784-11793 (2021)", + "87. Yu, J., Li, X., Koh, J.Y., Zhang, H., Pang, R., Qin, J., Ku, A., Xu, Y., Baldridge, J., Wu, Y.: Vector-quantized image modeling with improved VQGAN. arXiv preprint arXiv:2110.04627 (2021)", + "88. Yu, L., Li, X., Fu, C.W., Cohen-Or, D., Heng, P.A.: Pu-Net: Point cloud upsampling network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2790-2799 (2018)", + "89. Zeng, X., Vahdat, A., Williams, F., Gojcic, Z., Litany, O., Fidler, S., Kreis, K.: LION: Latent point diffusion models for 3D shape generation. Advances in Neural Information Processing Systems 35, 10021-10039 (2022)", + "90. Zhang, J., Singh, S.: LOAM: LiDAR odometry and mapping in real-time. In: Robotics: Science and Systems. vol. 2, pp. 1-9. Berkeley, CA (2014)", + "91. Zhang, Y., Hu, Q., Xu, G., Ma, Y., Wan, J., Guo, Y.: Not all points are equal: Learning highly efficient point-based detectors for 3D LiDAR point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18953-18962 (2022)", + "92. Zhao, Y., Bai, L., Huang, X.: FidNet: LiDAR point cloud semantic segmentation with fully interpolation decoding. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 4453-4458. IEEE (2021)", + "93. Zhou, L., Du, Y., Wu, J.: 3D shape generation and completion through point-voxel diffusion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5826-5835 (2021)", + "94. Zhou, Y., Tuzel, O.: VoxelNet: End-to-end learning for point cloud based 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4490-4499 (2018)", + "95. Zyrianov, V., Zhu, X., Wang, S.: Learning to generate realistic LiDAR point clouds. In: European Conference on Computer Vision. pp. 17-35. Springer (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 632 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "Q. Hu et al.", + "bbox": [ + 271, + 114, + 352, + 128 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/2024/RangeLDM_ Fast Realistic LiDAR Point Cloud Generation/ec4004eb-e76a-4f41-9891-f055a3eeaf7c_model.json b/2024/RangeLDM_ Fast Realistic LiDAR Point Cloud Generation/ec4004eb-e76a-4f41-9891-f055a3eeaf7c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3e842d971cec9848f59b628ff0914f94c99c0c55 --- /dev/null +++ b/2024/RangeLDM_ Fast Realistic LiDAR Point Cloud Generation/ec4004eb-e76a-4f41-9891-f055a3eeaf7c_model.json @@ -0,0 +1,3100 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.234, + 0.141, + 0.773, + 0.185 + ], + "angle": 0, + "content": "RangeLDM: Fast Realistic LiDAR Point Cloud Generation" + }, + { + "type": "text", + "bbox": [ + 0.323, + 0.214, + 0.674, + 0.23 + ], + "angle": 0, + "content": "Qianjiang Hu, Zhimin Zhang, and Wei Hu" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.242, + 0.768, + 0.271 + ], + "angle": 0, + "content": "Wangxuan Institute of Computer Technology, Peking University, Beijing, China hqjpu@pku.edu.cn zm_zhang@stu.pku.edu.cn forhuwei@pku.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.306, + 0.346, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.424, + 0.338, + 0.432 + ], + "angle": 0, + "content": "(a) Unconditional Generation" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.306, + 0.605, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.418, + 0.424, + 0.523, + 0.432 + ], + "angle": 0, + "content": "(b) Conditional Generation" + }, + { + "type": "image", + "bbox": [ + 0.608, + 0.31, + 0.783, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.655, + 0.424, + 0.734, + 0.433 + ], + "angle": 0, + "content": "(c) Quality vs. Speed" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.437, + 0.788, + 0.534 + ], + "angle": 0, + "content": "Fig. 1: (a). Unconditional LiDAR point cloud generation with realistic global structure. (b). Conditional LiDAR point cloud generation, including LiDAR point cloud upsampling and inpainting. (c). Generation quality (Maximum Mean Discrepancy, abbr. MMD) vs. generation speed (samples/s) of competitive LiDAR point cloud generation methods on the KITTI-360 [38] dataset. The proposed method outperforms the state-of-the-art methods LiDARGen [95] and UltraLiDAR [79] in both generation quality and generation speed. All speeds are evaluated on a single RTX 3090 GPU." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.565, + 0.741, + 0.786 + ], + "angle": 0, + "content": "Abstract. Autonomous driving demands high-quality LiDAR data, yet the cost of physical LiDAR sensors presents a significant scaling-up challenge. While recent efforts have explored deep generative models to address this issue, they often consume substantial computational resources with slow generation speeds while suffering from a lack of realism. To address these limitations, we introduce RangeLDM, a novel approach for rapidly generating high-quality range-view LiDAR point clouds via latent diffusion models. We achieve this by correcting range-view data distribution for accurate projection from point clouds to range images via Hough voting, which has a critical impact on generative learning. We then compress the range images into a latent space with a variational autoencoder, and leverage a diffusion model to enhance expressivity. Additionally, we instruct the model to preserve 3D structural fidelity by devising a range-guided discriminator. Experimental results on KITTI-360 and nuScenes datasets demonstrate both the robust expressiveness and fast speed of our LiDAR point cloud generation." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.801, + 0.714, + 0.815 + ], + "angle": 0, + "content": "Keywords: Diffusion models \\(\\cdot\\) Autonomous systems \\(\\cdot\\) Point clouds" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.825, + 0.454, + 0.84 + ], + "angle": 0, + "content": "\\(\\mathbb{O}\\) Corresponding author: W. Hu." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Q. Hu et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.377, + 0.161 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.787, + 0.281 + ], + "angle": 0, + "content": "Autonomous systems have drawn great attention in both the academia and the industry community. Numerous autonomous systems leverage the power of various sensors and deep learning to enhance the perception of the 3D world. LiDAR (Light Detection And Ranging), with its ability to provide precise 3D geometric information about the surroundings, has become a popular sensor choice for autonomous systems, including self-driving cars [61,62,86], surveying drones [52,90] and robots [42,78]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.282, + 0.787, + 0.357 + ], + "angle": 0, + "content": "However, while LiDAR offers accurate geometric measurements, it comes with a significant limitation: the data collection process is exceedingly expensive and challenging to scale up. This renders physical sensors impractical for scalable and customizable data collection. Additionally, it is difficult to collect data in corner cases such as car accidents and extreme weather conditions in the field." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.787, + 0.613 + ], + "angle": 0, + "content": "One approach to mitigate this issue is to employ existing LiDAR simulation toolkits [37, 43] to synthesize more data. Nevertheless, these systems typically demand manual scene creation or rely on multiple prior scans of the real world. Another approach is leveraging deep generative models to generate LiDAR point clouds. This is challenging due to the high degree of unstructuredness, sparsity, and non-uniformity in LiDAR point clouds. Caccia et al. [5] proposed LiDAR GAN and LiDAR VAE to generate LiDAR point clouds with generative adversarial networks (GANs) [21] and variational autoencoders (VAEs) [32], respectively. However, the generation results often suffer from issues such as fuzzy or missing details. LiDARGen [95] introduced a novel score-based model to synthesis LiDAR point clouds, which however sampled slowly and failed to generate high-quality geometric details at a far range. UltraLiDAR [79] proposed to synthesize voxelized point clouds in bird's-eyes-view (BEV) with vector quantized VAE (VQ-VAE) [74], resulting in more realistic point clouds than previous methods. Nevertheless, due to the large size and sparsity of point clouds in BEV, most computing power of UltraLiDAR is devoted to generating empty voxels, which results in a low generation speed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.787, + 0.842 + ], + "angle": 0, + "content": "To this end, we propose RangeLDM, a novel approach based on latent diffusion models (LDMs) [55], which improves both the quality and speed of LiDAR point cloud generation significantly. Firstly, we introduce a range image view to represent point clouds. The reasons are twofold: 1) Range images are compact and closely mimic the sampling conditions of LiDAR, making them a suitable representation. A LiDAR point cloud essentially constructs a 2.5D scene from a single viewpoint instead of a full 3D point cloud [27]. Consequently, organizing the point cloud in range view ensures that no information is overlooked. 2) 2D image generation techniques are relatively mature, providing a solid foundation for our work. While LiDARGen [95] also represents point clouds as range images, it suffers from blurriness caused by inaccurate projection. In contrast, we provide insights that the correct range-view data distribution has a critical impact on such range-view-based generative model learning, and accurately project point clouds onto range images using parameters estimated by Hough Voting. Then, to achieve a faster sampling speed [55] and greater expressivity [55,72], we compress" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.34, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RangeLDM: Fast Realistic LiDAR Point Cloud Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": "range images to the latent space with a VAE and develop a diffusion model operating on the latent space, leveraging the successful paradigm of LDMs [55,72,89]. Further, in order to enhance the ability of the VAE in reconstructing 3D structures, we introduce a range-guided discriminator, which is supervised from the spherical coordinates and thus geometry-sensitive. This plays a crucial role in guiding the decoder to generate high-quality range images, ultimately preserving the fidelity of the 3D structure. Extensive experimental results show that the proposed RangeLDM achieves the state-of-the-art performance on KITTI-360 [38] and nuScenes [6] datasets in terms of both the generation quality and generation speed, as demonstrated in Figure 1 (c). The LiDAR upsampling and LiDAR inpainting results also demonstrate the potential of the proposed method for conditional generation, as presented in Figure 1 (a) and (b)." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.33, + 0.673, + 0.344 + ], + "angle": 0, + "content": "To summarize, the main contributions of this paper include:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.356, + 0.784, + 0.4 + ], + "angle": 0, + "content": "- We propose a latent diffusion model to capture the distribution of range-view LiDAR point clouds, aiming to generate realistic point cloud scenes at a fast speed." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.402, + 0.785, + 0.446 + ], + "angle": 0, + "content": "- We enlighten the significance of the correct range-view data distribution for range-view-based generative models and achieve high-quality range image projection via Hough Voting." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.448, + 0.785, + 0.478 + ], + "angle": 0, + "content": "- We exploit a range-guided discriminator to ensure preserving the fidelity of the 3D geometric structure in generated point clouds." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.479, + 0.785, + 0.524 + ], + "angle": 0, + "content": "- Experimental results on the KITTI-360 and nuScenes datasets highlight that our approach outperforms state-of-the-art methods in terms of both visual quality and especially generation speed." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.356, + 0.785, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.549, + 0.388, + 0.566 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.789, + 0.842 + ], + "angle": 0, + "content": "LiDAR Representation. LiDAR point clouds can be represented in various forms, encompassing raw point clouds, voxels, range views, and multi-view fusion. Point-based models [49, 62, 64, 76, 84, 85, 91] directly encode 3D objects from raw points using the PointNet [50] encoder and subsequently perform detection or segmentation based on point features. These methods wholly preserve the irregularity and locality of a point cloud but have relatively higher latency. Voxel-based methods [13, 14, 17, 34, 35, 60, 63, 67, 77, 81, 82, 86, 94] voxelize point clouds for convolutional neural networks to efficiently capture features. They are computationally effective but the desertion of fine-grained patterns degrades further refinement. Given that the range view is compact and compatible with the LiDAR sensor's sampling process, range-image-based methods [2, 8, 10-12, 18, 33, 36, 44, 45, 68, 71, 80, 92] have been proposed to directly process range images for LiDAR perception. Multi-view fusion methods [22, 26, 61] amalgamate multiple representations, which yield better results at the expense of processing speed. In this paper, we adopt the range-view representation, which is congruent with the LiDAR sensor's sampling process and can be efficiently encoded by image generation models." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Q. Hu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.356 + ], + "angle": 0, + "content": "Generative Models for Point Clouds. Given observed samples of interest, generative models aim to learn the underlying distribution of the data and generate new samples. The first 3D generative models [1,65,73] for point clouds are based on GANs [21]. They operate by generating point clouds and discriminating them from real samples in an adversarial manner. The second category of methods [19,20,31,40,46,47,70] employ VAEs [32] to explore the probabilistic latent space of 3D shapes. Auto-regressive models [4] have also been introduced to generate point clouds [69] from scratch or semantic contexts. PointFlow [83] and SoftFlow [30] leverage normalizing flows [53] to capture the likelihood of shapes and points. Notably, the recent success of denoising diffusion models (DDMs) [25] for image synthesis [55] has also been extended to the domain of 3D point cloud generation [7,28,41,48,89,93]. However, these previous works often concentrate on the object level and are not well-suited for handling large scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.788, + 0.526 + ], + "angle": 0, + "content": "In this paper, our focus lies in the generation of LiDAR point clouds. While [5] and [57] employ VAE or GANs for LiDAR point cloud generation, the realism achieved in their results is relatively limited. UltraLiDAR [79], on the other hand, utilizes VQ-VAE to generate voxelized LiDAR point clouds, but at the expense of introducing quantified losses and slow generation speed. LiDARGen [95] proposed a novel score-matching energy-based model to generate higher-quality Li-DAR point clouds, but it samples very slowly and has degraded geometric details at a far range. A contemporary work LiDM [51] also utilized LDMs to generate LiDAR point clouds. However, while they focused on generating LiDAR point clouds conditional on multimodal data, we focus on improving the quality of point cloud generation directly." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.546, + 0.673, + 0.566 + ], + "angle": 0, + "content": "3 Background on Denoising Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.577, + 0.788, + 0.667 + ], + "angle": 0, + "content": "DDMs [25] are latent variable models that employ a pre-defined posterior distribution, known as the forward diffusion process, and are trained with a denoising objective. More specifically, given samples \\(\\mathbf{x}_0\\sim q(\\mathbf{x}_0)\\) from a data distribution, DDMs follow the form \\(p_{\\theta}(\\mathbf{x}_0)\\coloneqq \\int p_{\\theta}(\\mathbf{x}_{0:T})d\\mathbf{x}_{1:T}\\), where \\(T\\) denotes the number of steps, \\(\\mathbf{x}_1,\\dots,\\mathbf{x}_T\\) are latent variables that gradually add noise to the data \\(\\mathbf{x}_0\\), and \\(\\theta\\) denotes the parameter set of the DDM decoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.667, + 0.788, + 0.699 + ], + "angle": 0, + "content": "DDMs are trained by minimizing the evidence lower bound (ELBO) of the data \\(\\mathbf{x}_0\\) under \\(p_{\\theta}(\\mathbf{x}_{0:T})\\). This objective can be simplified to [25]:" + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.707, + 0.787, + 0.733 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {\\mathbf {x}, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {x} _ {t} + \\epsilon , t) \\| _ {2} ^ {2} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.743, + 0.517, + 0.759 + ], + "angle": 0, + "content": "with \\(t\\) uniformly sampled from \\(\\{1,\\dots,T\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.78, + 0.473, + 0.798 + ], + "angle": 0, + "content": "4 The Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "As shown in Fig. 2, we first project point clouds onto high-quality range images (as detailed in Section 4.1), considering the compact nature of the range view," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.339, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "RangeLDM: Fast Realistic LiDAR Point Cloud Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.788, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.264, + 0.788, + 0.389 + ], + "angle": 0, + "content": "Fig. 2: The framework of the proposed RangeLDM. Firstly, we project point clouds onto high-quality range images via Hough Voting (Section 4.1). Subsequently, we train a VAE to compress the range images into low-dimensional latent features \\(\\mathbf{z}_0\\), which encodes the range images with the encoder \\(\\mathcal{E}_{\\zeta}\\) and reconstructs range images from latent features with the decoder \\(\\mathcal{G}_{\\eta}\\) (Section 4.2). Here, a range-guided discriminator \\(\\mathcal{D}_{\\tau}\\) is introduced to guide the decoder in the reconstruction of 3D structures. We finally train a latent diffusion model to capture the distribution of the latent features (Section 4.2). With optional conditional inputs, the proposed method is applicable to tasks such as point cloud upsampling and inpainting (Section 4.3)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.788, + 0.518 + ], + "angle": 0, + "content": "which aligns seamlessly with the sampling process of LiDAR sensors. Our model training process is then focused on these projected 2D range images. We then compress range images into a low-dimensional latent space through a VAE (as detailed in Section 4.2). Subsequently, we proceed to train DDMs within the reduced-dimensional latent space (as discussed in Section 4.2). We discuss applications of the proposed model in Section 4.3, which involves unconditional generation and conditional generation. Implementation details are explained in Section 4.4." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.542, + 0.52, + 0.558 + ], + "angle": 0, + "content": "4.1 High-Quality Range Projection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.645 + ], + "angle": 0, + "content": "The range image presents LiDAR data compactly and intuitively, with rows indicating the laser beams and columns representing the yaw angles. We convert point clouds to range images using spherical projection. Typically, for a point \\(\\mathbf{p}\\) in Cartesian coordinates \\((x,y,z)\\), we calculate its spherical coordinates \\((r,\\theta,\\phi)\\) using:" + }, + { + "type": "equation", + "bbox": [ + 0.285, + 0.655, + 0.787, + 0.681 + ], + "angle": 0, + "content": "\\[\nr = \\sqrt {x ^ {2} + y ^ {2} + z ^ {2}}, \\theta = \\mathrm {a t a n} (y, x), \\phi = \\mathrm {a t a n} \\left(z, \\sqrt {x ^ {2} + y ^ {2}}\\right). \\qquad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.686, + 0.788, + 0.762 + ], + "angle": 0, + "content": "However, in most current datasets such as KITTI-360, multiple lasers from the Velodyne LiDAR system do not share a common origin for their measurements. This may introduce errors in the direct conversion from Cartesian points to spherical points, resulting in incorrect range-view data distribution and thus low-quality range images, as shown at the top of Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.762, + 0.788, + 0.809 + ], + "angle": 0, + "content": "To address this issue, we adopt Hough Voting to estimate heights and pitch angles \\(\\{h_j,\\phi_j\\}_{j = 1,\\dots ,N}\\) for Velodyne sensors [3]. We then adjust the point cloud transformation to a range image using" + }, + { + "type": "equation", + "bbox": [ + 0.322, + 0.819, + 0.787, + 0.845 + ], + "angle": 0, + "content": "\\[\nr = \\sqrt {x ^ {2} + y ^ {2} + (z - h _ {j}) ^ {2}}, \\theta = \\mathrm {a t a n} (y, x), \\phi = \\phi_ {j}, \\tag {3}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Q. Hu et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.531, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.227, + 0.532, + 0.269 + ], + "angle": 0, + "content": "Fig. 3: Comparison of range projection by the typical method described in Eq. 2 and our method with Hough Voting as in Eq. 3." + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.149, + 0.784, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.279, + 0.581, + 0.294 + ], + "angle": 0, + "content": "where \\(h_j\\) and \\(\\phi_j\\) refer to the \\(j\\)-th Velodyne sensor." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.295, + 0.785, + 0.385 + ], + "angle": 0, + "content": "We then rasterize points \\((r,\\theta ,\\phi)\\) into a 2D cylindrical projection \\(R(u,v)\\) (a.k.a., range image) of size \\(H\\times W\\) with \\(u = ((\\theta +\\pi) / 2\\pi)W,v = j\\), where \\((u,v)\\) denotes the grid coordinate of a point in the range image. Thus, we obtain high-quality range images as illustrated at the bottom of Fig. 3. We denote the obtained range image as \\(\\mathbf{x}\\in \\mathbb{R}^{H\\times W\\times 2}\\), which comprises \\(H\\times W\\) pixels associated with both range and intensity \\(\\{r,i\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.785, + 0.461 + ], + "angle": 0, + "content": "To verify the impact of range projection, we trained range-based LiDAR generation methods like LiDARGen [95] with our obtained high-quality range images. Table 1 shows substantial improvement, which sheds light on the significance of the correct range-view data distribution for such range-based generative model learning." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.485, + 0.334, + 0.501 + ], + "angle": 0, + "content": "4.2 Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.513, + 0.785, + 0.633 + ], + "angle": 0, + "content": "As a range image is often of high dimensionality at the scale of \\(64 \\times 1024\\), it would be computationally expensive to learn the distribution directly. Instead, we compress the range image into a low-dimensional latent space via a VAE and then model the distribution via an LDM. Consequently, the training of the proposed RangeLDM includes two distinct stages. In the first stage, we train a regular VAE to compress the range image into latent features \\(\\mathbf{z}_0 \\in \\mathbb{R}^{h \\times w \\times c}\\) by a downsampling factor \\(f = H / h = W / w\\). In the second stage, we train an LDM to learn the distribution of the latent encoding \\(\\mathbf{z}_0\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.634, + 0.785, + 0.695 + ], + "angle": 0, + "content": "First-Stage: Dimensionality Reduction. In the first stage, we reduce the dimensionality of the obtained range images by a VAE, which removes imperceptible high-frequency details and leads to low-dimensional latent features that encode prominent information in the original range image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.695, + 0.785, + 0.772 + ], + "angle": 0, + "content": "A standard VAE consists of two main components: 1) an encoder \\(\\mathcal{E}_{\\zeta}\\) that transforms an input range image \\(\\mathbf{x} \\in \\mathbb{R}^{H \\times W \\times 2}\\) into latent features \\(\\mathbf{z}_0 = \\mathcal{E}_{\\zeta}(\\mathbf{x}) \\in \\mathbb{R}^{h \\times w \\times c}\\) and 2) a decoder \\(\\mathcal{G}_{\\eta}\\) that takes the latent features \\(\\mathbf{z}_0\\) as input and generates the reconstructed range image \\(\\hat{\\mathbf{x}} = \\mathcal{G}_{\\eta}(\\mathbf{z}_0)\\). The entire VAE is trained by maximizing a modified ELBO with respect to the parameters \\(\\zeta\\) and \\(\\eta\\) [32,54]:" + }, + { + "type": "equation", + "bbox": [ + 0.26, + 0.786, + 0.785, + 0.804 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {E L B O} (\\zeta , \\eta) = \\mathbb {E} _ {q _ {\\zeta} (\\mathbf {z} _ {0} | \\mathbf {x})} \\left[ \\log p _ {\\boldsymbol {\\eta}} (\\mathbf {x} \\mid \\mathbf {z} _ {0}) \\right] - \\lambda_ {1} D _ {\\mathrm {K L}} \\left(q _ {\\zeta} (\\mathbf {z} _ {0} \\mid \\mathbf {x}) \\| p (\\mathbf {z} _ {0})\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In the above equation, the first term evaluates the reconstruction likelihood of the decoder from the latent \\(\\mathbf{z}_0\\) and corresponds to an \\(L_{1}\\) reconstruction loss," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.339, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "RangeLDM: Fast Realistic LiDAR Point Cloud Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "while the second term quantifies how closely the learned distribution of \\(\\mathbf{z}_0\\) resembles the prior distribution \\(p(\\mathbf{z}_0) = \\mathcal{N}(\\mathbf{z}_0; \\mathbf{0}, \\mathbf{I})\\) held over latent variables. The hyperparameter \\(\\lambda_1\\) balances the reconstruction accuracy and Kullback-Leibler regularization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.342 + ], + "angle": 0, + "content": "Range-Guided Discriminator. In order to mitigate the blurriness introduced by relying solely on the pixel-space \\( L_{1} \\) reconstruction loss, we integrate the VAE with a patch-based adversarial discriminator [15, 29, 55, 87]. Furthermore, to ensure that the reconstruction remains confined within the manifold underlying the range image and exploit the geometric information from the spherical coordinates, we propose a range-guided discriminator \\( \\mathcal{D}_{\\tau} \\), with parameters \\( \\tau \\). Specifically, we adapt the Meta-Kernel [18] to replace the standard convolution in the discriminator, aiming to learn convolution weights from relative spherical coordinates. The Meta-Kernel is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.365, + 0.351, + 0.785, + 0.384 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} _ {i} ^ {\\prime} = \\mathcal {W} \\left(\\mathcal {A} _ {j \\in \\mathcal {N} (i)} \\left(\\Phi (\\gamma (\\mathbf {p} _ {j}, \\mathbf {p} _ {i})) \\odot \\mathbf {h} _ {j}\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.394, + 0.788, + 0.484 + ], + "angle": 0, + "content": "where \\(\\mathbf{h}\\) and \\(\\mathbf{h}'\\) represent the input feature and output feature, respectively. The function \\(\\varPhi\\) denotes a Multi-Layer Perceptron (MLP) with two fully-connected layers, \\(\\mathcal{A}\\) is a concatenation operation, \\(\\mathcal{N}(i)\\) denotes the 2D-grid neighborhood around the center pixel \\(i\\), and \\(\\mathcal{W}\\) is a fully-connected layer. The term \\(\\gamma (\\mathbf{p}_j,\\mathbf{p}_i)\\) represents the distance between points \\(\\mathbf{p}_j\\) and \\(\\mathbf{p}_i\\) in spherical coordinates, which is expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.244, + 0.495, + 0.74, + 0.529 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\gamma \\left(\\mathbf {p} _ {j}, \\mathbf {p} _ {i}\\right) := \\left\\{r _ {j} \\cos (\\Delta \\theta) \\cos (\\Delta \\phi) - r _ {i}, r _ {j} \\cos (\\Delta \\theta) \\sin (\\Delta \\phi), r _ {j} \\sin (\\Delta \\theta) \\right\\}, \\\\ \\Delta \\theta = \\theta_ {j} - \\theta_ {i}, \\quad \\Delta \\phi = \\phi_ {j} - \\phi_ {i}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.784, + 0.599 + ], + "angle": 0, + "content": "The Meta-Kernel is aware of local 3D structures by adaptively adjusting convolution kernel weights from relative spherical coordinates. This makes it challenging for the decoder to deceive the range-guided discriminator. This, in turn, encourages the decoder to generate more realistic range images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.6, + 0.784, + 0.629 + ], + "angle": 0, + "content": "With the range-guided discriminator, the final objective of the first-stage training is:" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.642, + 0.785, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\zeta , \\eta = \\arg \\min _ {\\zeta , \\eta} \\max _ {\\tau} \\mathcal {L} _ {E L B O} (\\zeta , \\eta) + \\lambda_ {2} \\mathcal {L} _ {a d v} (\\zeta , \\eta , \\tau), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.722, + 0.688 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{adv}\\) is the GAN Hinge loss [39] and \\(\\lambda_{2}\\) balances the two losses." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.785, + 0.763 + ], + "angle": 0, + "content": "Second-Stage: Latent Diffusion Modeling In the second stage, we freeze the encoder and decoder of the learned VAE and train an LDM on the encoded latent feature \\(\\mathbf{z}_0\\) for distribution modeling of the range image. During the diffusion process, noise is added to the initial latent \\(\\mathbf{z}_0\\), resulting in a noisy latent \\(\\mathbf{z}_t\\), where the noise level increases over time steps \\(t \\in T\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.784, + 0.809 + ], + "angle": 0, + "content": "Regarding the modeling of the unconditional distribution \\( p(\\mathbf{z}_0) \\), based on Eq. 1, we learn a time-conditional UNet [56] \\( \\epsilon_{\\theta}(z_t,t) \\) that predicts the noise added to the noisy latent \\( \\mathbf{z}_t \\):" + }, + { + "type": "equation", + "bbox": [ + 0.349, + 0.82, + 0.785, + 0.845 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {L D M}} := \\mathbb {E} _ {\\mathcal {E} (x), \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t\\right) \\| _ {2} ^ {2} \\right]. \\tag {8}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Q. Hu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.19 + ], + "angle": 0, + "content": "As to the modeling of the conditional distribution \\( p(\\mathbf{z}_0|\\mathbf{y}) \\), where \\( \\mathbf{y} \\) is a condition such as sparse or masked point clouds, we learn a conditional LDM via" + }, + { + "type": "equation", + "bbox": [ + 0.32, + 0.193, + 0.786, + 0.218 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {L D M}} := \\mathbb {E} _ {\\mathcal {E} (x), \\mathbf {y}, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {z} _ {t}, t, \\tau_ {\\theta} (\\mathbf {y})) \\| _ {2} ^ {2} \\right], \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.228, + 0.784, + 0.275 + ], + "angle": 0, + "content": "where \\(\\tau_{\\theta}\\) is a condition encoder that projects \\(\\mathbf{y}\\) into an intermediate representation \\(\\tau_{\\theta}(\\mathbf{y})\\), and \\(\\epsilon_{\\theta}(\\mathbf{z}_t, t, \\tau_{\\theta}(\\mathbf{y}))\\) is a UNet interpolated with cross-attention layers [55, 75]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.305, + 0.356, + 0.319 + ], + "angle": 0, + "content": "4.3 Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.338, + 0.784, + 0.413 + ], + "angle": 0, + "content": "Unconditional Generation. With the decoder \\(\\mathcal{G}_{\\eta}\\) and the LDM \\(\\epsilon_{\\theta}\\), we establish a hierarchical generative model \\(p_{\\eta, \\theta}(\\mathbf{x}, \\mathbf{z}_0) = p_{\\eta}(\\mathbf{x} | \\mathbf{z}_0) p_{\\theta}(\\mathbf{z}_0)\\). That is, we first generate latent features with the LDM, and then map the latent features back to the original range image space using the decoder \\(\\mathcal{D}_{\\eta}\\) for LiDAR point cloud generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.784, + 0.492 + ], + "angle": 0, + "content": "Conditional Generation. Given a condition \\(\\mathbf{y}\\), the conditional generative model is defined as \\(p_{\\eta, \\theta}(\\mathbf{x}, \\mathbf{z}_0 | \\mathbf{y}) = p_{\\eta}(\\mathbf{x} | \\mathbf{z}_0, \\mathbf{y}) p_{\\theta}(\\mathbf{z}_0 | \\mathbf{y})\\). There could be various conditions for LiDAR point cloud generation, such as sparse point clouds, incomplete point clouds, camera or even text. Here, we illustrate two applications of conditional generation: LiDAR point cloud upsampling and inpainting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.785, + 0.629 + ], + "angle": 0, + "content": "LiDAR Point Cloud Upsampling. This task takes a sparse point cloud\\(^1\\) \\(\\mathbf{y} \\in \\mathbb{R}^{h \\times W \\times 2}\\) as input (following the setting of LiDARGen) and expects the model to produce a denser counterpart \\(\\hat{\\mathbf{x}} \\in \\mathbb{R}^{H \\times W \\times 2}\\), which is necessary for accurate LiDAR-based perception and density-insensitive domain adaptation. We train LDM following Eq. 9 with the sparse point cloud \\(\\mathbf{y}\\) and its ground-truth dense variant \\(\\mathbf{x}\\). The condition encoder \\(\\tau_{\\theta}\\) is configured as a reshaper that transforms the sparse point cloud \\(\\mathbf{y}\\) into \\(\\mathbf{y}' \\in \\mathbb{R}^{h \\times w \\times 2f}\\). The reshaped \\(\\mathbf{y}'\\) is then concatenated with noisy latent variable \\(\\mathbf{z}_t\\) and fed into the UNet \\(\\epsilon_{\\theta}\\) for noise prediction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.786, + 0.797 + ], + "angle": 0, + "content": "LiDAR Point Cloud Inpainting. This task is required when LiDAR sensors cannot capture the entire scene due to object occlusion or sensor limitations. We train an inpainting LDM from the ground truth point cloud \\(\\mathbf{x}\\), a masked point cloud \\(\\mathbf{x}'\\), and the corresponding mask \\(\\mathbf{m} \\in \\{0,1\\}^{H \\times W}\\). The masked point cloud \\(\\mathbf{x}'\\) is encoded with the encoder \\(\\mathcal{E}_{\\zeta}\\) from the VAE and is then concatenated with a downsampled mask \\(\\mathbf{m}' \\in \\{0,1\\}^{h \\times w}\\) as the condition term in Eq. 9. Mathematically, we represent the condition \\(\\mathbf{y}\\) as \\(\\{\\mathbf{x}',\\mathbf{m}\\}\\), and deploy the condition encoder to project it into \\(\\tau_{\\theta}(\\mathbf{y}) = \\mathcal{E}_{\\zeta}(\\mathbf{x}') \\| ds(\\mathbf{m})\\), where \\(\\|\\cdot\\|\\) denotes the concatenation operation and \\(ds(\\cdot)\\) indicates the downsampling operation. The projected condition is then concatenated with the noisy latent variable \\(\\mathbf{z}_t\\) and input into the UNet \\(\\epsilon_{\\theta}\\) for noise prediction." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.784, + 0.839 + ], + "angle": 0, + "content": "1 Without causing ambiguity, the point clouds are converted to range images by default." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.34, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RangeLDM: Fast Realistic LiDAR Point Cloud Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.784, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.285, + 0.785, + 0.354 + ], + "angle": 0, + "content": "Fig. 4: Qualitative results comparing against baselines for unconditional Li-DAR generation on KITTI-360. Real point clouds are only for reference. Our model produces results that closely resemble real-world data, which excels in generating road scenes, such as cars (the first row), road bollards (the second row) and crossroads (the last row)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.363, + 0.456, + 0.377 + ], + "angle": 0, + "content": "4.4 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.785, + 0.522 + ], + "angle": 0, + "content": "Circular Convolution. Range images are inherently circular, which means that the left boundary of a range image is actually connected to its right boundary. However, standard convolutions employ zero-padding or symmetry padding and do not take into account such constraints. Thus, as in LiDArgen [95], we replace standard convolutions in both the VAE and LDM with circular convolutions [59], which treat the left and right boundaries as connected neighbors in the topology. Since the 2D convolution operator satisfies translation invariance, it is evident that the entire network is invariant to horizontal shifting of the range image (i.e., rotation in the xy-plane of the point cloud)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.785, + 0.704 + ], + "angle": 0, + "content": "Conditioning the Model with Direction. Real LiDAR point clouds collected from driving scenes usually have a certain directionality. For example, since vehicles typically follow the direction of the road, the x-axis direction of the LiDAR point cloud usually aligns with the road direction. However, since the whole network of our model is invariant to rotation in the xy-plane of the point cloud, the point clouds generated from random noise could result in arbitrary directions. To avoid this, we generate point clouds with directional conditioning. Specifically, we use an \\( h \\times w \\) matrix as the condition, in which only the values from the first column are set to 1, while the rest are all set to 0. This 1-valued column corresponds to the x-axis of the point cloud. Such simple conditioning demonstrates effective control over the direction of the generated point cloud, as presented in the qualitative results of the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Network Architectures and Model Hyperparameters. For simplicity, we adopted a similar architecture to that in [55]. The downsampling factor \\( f \\) is set as 4 to strike a good balance between efficiency and realistic results. The VAE comprises three encoder blocks and three decoder blocks. The LDM consists of four downsampling blocks with the last three followed by a transformer layer, and four upsampling blocks with the first three blocks also followed by a transformer layer. In the generation process, we employed 50 denoising steps with DDIM sampler [66] for point cloud generation. More implementation details can be found in released code https://github.com/WoodwindHu/RangeLDM." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Q. Hu et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.298, + 0.785, + 0.339 + ], + "angle": 0, + "content": "Fig. 5: Comparison of upsampling results on KITTI-360. We downsampled the ground truth by a factor of four as the input, and demonstrated the results of different methods on \\(4 \\times\\)-upsampling of the input." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.346, + 0.375, + 0.364 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.378, + 0.427, + 0.394 + ], + "angle": 0, + "content": "5.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.787, + 0.478 + ], + "angle": 0, + "content": "Datasets. We evaluate our model on two challenging datasets, KITTI-360 and nuScenes. KITTI-360 has 81,106 LiDAR readings from 9 sequences in Germany, covering diverse scenes. We used the first two sequences for validation and trained on the rest. NuScenes is a public dataset with 297,737 LiDAR sweeps in the training set and 52,423 in the testing set, collected in Boston and Singapore." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.787, + 0.599 + ], + "angle": 0, + "content": "Evaluation metrics. Following [95], we employed three metrics to perform quantitative analysis: Maximum Mean Discrepancy (MMD), Jensen-Shannon divergence (JSD), and Frechet Range Distance (FRD score). We use a \\(100 \\times 100\\) 2D histogram on the BEV plane to calculate MMD and JSD metrics. The Frechet Range Distance (FRD) score [95] is a metric used to evaluate the quality of samples acquired by a generative model, inspired by the FID score for images [24]. To compute the FRD score, we use RangeNet++ [45], an encoder-decoder-based network for segmentation, which is pre-trained on KITTI-360." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.645 + ], + "angle": 0, + "content": "Baselines. We evaluated our approach for LiDAR point cloud generation against several competitive methods, including LiDAR VAE [5], LiDAR GAN [5], Projected GAN [58], LiDARGen [95] and UltraLiDAR [79]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.667, + 0.476, + 0.68 + ], + "angle": 0, + "content": "5.2 Unconditional Generation" + }, + { + "type": "table", + "bbox": [ + 0.306, + 0.688, + 0.699, + 0.778 + ], + "angle": 0, + "content": "
MethodYears\\( MMD_{BEV} \\)↓FRD ↓\\( JSD_{BEV} \\)↓
LiDAR GAN [5]IROS 2019\\( 3.06 \\times 10^{-3} \\)3003.8-
LiDAR VAE [5]IROS 2019\\( 1.00 \\times 10^{-3} \\)2261.50.161
Projected GAN [58]NeurIPS 2021\\( 3.47 \\times 10^{-4} \\)2117.20.085
LiDARGen [95]ECCV 2022\\( 3.87 \\times 10^{-4} \\)2040.10.067
UltraLiDAR [79]CVPR 2023\\( 1.96 \\times 10^{-4} \\)-0.071
Ours\\( 3.07 \\times 10^{-5} \\)1074.90.045
" + }, + { + "type": "table_caption", + "bbox": [ + 0.293, + 0.778, + 0.709, + 0.792 + ], + "angle": 0, + "content": "Table 2: Unconditional generation results on KITTI-360 [38]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Quantitative Results on KITTI-360. Table 2 displays the quantitative results of the proposed RangeLDM and competing algorithms. The results clearly" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.339, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RangeLDM: Fast Realistic LiDAR Point Cloud Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "illustrate that our method significantly outperforms the baselines across all metrics. Notably, our results on the MMD metric are an order of magnitude lower than those of other methods. This underscores the remarkable expressive capabilities of our model in the context of LiDAR point cloud generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.21, + 0.788, + 0.422 + ], + "angle": 0, + "content": "Qualitative Results on KITTI-360. Figure 4 displays a set of randomly generated samples from competing algorithms, alongside real point cloud samples extracted from the KITTI-360 dataset for comparison. We observe that the Projected GAN [58] exhibits noticeable artifacts in the distant range. LiDARGen [95] effectively captures the general layout, but produces rather noisy outputs and lacks the same degree of straight walls as actual samples. UltraLiDAR [79] generates structured and reasonable scenes, yet it only creates voxelized point clouds (as magnified in the first row) without intensity features. In contrast, our model consistently surpasses the baseline models and yields results closely resembling real-world data. For example, we can generate cars (emphasized by red boxes in the first row of Figure 4) and road bollards (red boxes in the second row of Figure 4) on the road. Additionally, our model generates various scenes such as straightway (the first two rows of Figure 4) and crossroads (the last row of Figure 4). The supplementary materials provide more visualization results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.424, + 0.488, + 0.53 + ], + "angle": 0, + "content": "Human study on KITTI-360. To assess the perceptual quality, we conducted an A/B test involving a group of 14 researchers with LiDAR expertise. Following the same evaluation system as [95] and [79], we present pairs of randomly selected images" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.428, + 0.782, + 0.505 + ], + "angle": 0, + "content": "
MethodPercent prefer ours
Ours vs. VAE [5]98.8%
Ours vs. GAN [5]98.0%
Ours vs. ProjectedGAN [58]93.2%
Ours vs. LiDARGen [95]90.2%
Ours vs. UltraLiDAR [79]86.5%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.511, + 0.505, + 0.774, + 0.518 + ], + "angle": 0, + "content": "Table 3: Human study on KITTI-360." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.53, + 0.788, + 0.59 + ], + "angle": 0, + "content": "from two point clouds and ask participants to determine which one appeared more realistic. The results, as displayed in Table 3, unequivocally demonstrate the superior visual quality of generation results by our model. In most of the cases, testers favored our results over the baselines." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.544, + 0.699 + ], + "angle": 0, + "content": "Evaluation on the NuScenes Dataset. The results listed in Table 4 demonstrate that our model outperforms LiDAR VAE and LiDARGen significantly over nuScenes. In particular, our results on the MMD metric are an order of magnitude lower than those of other methods, while we reduce to about \\(1/3\\) of the" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.7, + 0.784, + 0.744 + ], + "angle": 0, + "content": "results of competitive methods in terms of the JSD metric. Also, Figure 6 illustrates the superiority of our model over LiDAR VAE and LiDARGen in terms of qualitative comparison." + }, + { + "type": "table", + "bbox": [ + 0.557, + 0.597, + 0.784, + 0.648 + ], + "angle": 0, + "content": "
Method\\( MMD_{BEV} \\)↓\\( JSD_{BEV} \\)↓
LiDAR VAE [5]\\( 1.1 \\times 10^{-3} \\)-
\\( LiDARGen^† \\)[95]\\( 1.9 \\times 10^{-3} \\)0.160
Ours\\( 1.9 \\times 10^{-4} \\)0.054
" + }, + { + "type": "table_caption", + "bbox": [ + 0.555, + 0.648, + 0.785, + 0.689 + ], + "angle": 0, + "content": "Table 4: Unconditional generation results on nuScenes dataset [6]. †: Reproduced by us." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.747, + 0.788, + 0.839 + ], + "angle": 0, + "content": "Model Size. The number of parameters in LiDARGen [95] and UltraLiDAR [79] are 29.7M and 40.3M, respectively. To ensure fair competition, we limited the capacity of our model to be similar to other approaches. Our model comprises 41.4M parameters, consisting of 12.7M parameters for VAE and 28.7M parameters for LDM. As illustrated in Table 2, our model outperforms the baselines by a large margin under similar model sizes." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Q. Hu et al." + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.144, + 0.7, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.255, + 0.784, + 0.283 + ], + "angle": 0, + "content": "Fig. 6: Qualitative results comparing against LiDAR VAE and LiDARGen for unconditional LiDAR generation on nuScenes. Real point clouds are only for reference." + }, + { + "type": "table", + "bbox": [ + 0.25, + 0.288, + 0.754, + 0.381 + ], + "angle": 0, + "content": "
Hough Range-Guided \nVoting DiscriminatorCircular \nConvolutionDirection \nConditionedMMDBEV ↓FRD ↓JSDBEV ↓
(a)××××3.95 × 10-41536.70.067
(b)✓×××6.57 × 10-51229.30.056
(c)✓✓××4.72 × 10-51103.10.051
(d)✓✓✓×3.90 × 10-41797.20.078
(e)✓✓✓✓3.07 × 10-51074.90.045
" + }, + { + "type": "table_caption", + "bbox": [ + 0.398, + 0.381, + 0.605, + 0.394 + ], + "angle": 0, + "content": "Table 6: Main ablation study." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.414, + 0.573, + 0.535 + ], + "angle": 0, + "content": "Generation Efficiency. The proposed dimensionality reduction offers RangeLDM a considerable advantage in terms of generation speed. As indicated in Table 5, the proposed method significantly outperforms LiDARGen [95] and UltraLi-DAR [79] in terms of generation speed. In particular, our model is 200 times faster than LiDARGen and 30 times faster than UltraLi-DAR." + }, + { + "type": "table", + "bbox": [ + 0.587, + 0.417, + 0.784, + 0.482 + ], + "angle": 0, + "content": "
MethodThroughput ↑ (samples/s)
LiDARGen [95]0.02
UltraLiDAR [79]0.16
Ours4.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.584, + 0.483, + 0.787, + 0.51 + ], + "angle": 0, + "content": "Table 5: Inference speed on a single RTX 3090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.558, + 0.457, + 0.572 + ], + "angle": 0, + "content": "5.3 Conditional Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.785, + 0.628 + ], + "angle": 0, + "content": "To evaluate the conditional generation performance of RangeLDM, we conducted experiments on the KITTI-360 dataset over two tasks: LiDAR point cloud upsampling and inpainting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.785, + 0.689 + ], + "angle": 0, + "content": "LiDAR Point Cloud Upsampling. We obtain sparse input point clouds by selecting a subset of 16 beams from the raw 64-beam sensors, in line with LiDARGen. We compared our approach against PUNet [88], DeepRS [9], Grad-PU [23], bicubic interpolation, Nearest Neighbor (NN) interpolation and LiDARGen." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.69, + 0.497, + 0.81 + ], + "angle": 0, + "content": "Quantitatively, in addition to measuring Mean Absolute Error (MAE) in the range view, we also employed RangeNet++ semantic segmentation to evaluate the quality of upsampled results. As shown in Table 7, the proposed method outperforms competing methods in all metrics, including" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.694, + 0.783, + 0.794 + ], + "angle": 0, + "content": "
MethodYearsMAE ↓Accuracy ↑ IoU ↑
PUNet [88]20186.88--
DeepRS [9]20223.96--
Grad-PU [23]20235.09--
Bicubic-2.600.2650.166
NN-2.180.5460.394
LiDARGen [95]20221.230.6080.449
Ours0.890.7220.566
" + }, + { + "type": "table_caption", + "bbox": [ + 0.542, + 0.794, + 0.743, + 0.807 + ], + "angle": 0, + "content": "Table 7: LiDAR upsampling." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.811, + 0.785, + 0.84 + ], + "angle": 0, + "content": "MAE, per-point segmentation accuracy, and segmentation Intersection over Union (IoU)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.339, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RangeLDM: Fast Realistic LiDAR Point Cloud Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.144, + 0.703, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.242, + 0.784, + 0.27 + ], + "angle": 0, + "content": "Fig. 7: Inpainting results. Left: ground truth point clouds. Middle and right (red): input point clouds. Middle and right (blue): recovered point clouds." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.277, + 0.782, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.336, + 0.784, + 0.364 + ], + "angle": 0, + "content": "Fig. 8: VAE reconstruction results. The proposed VAE with a range-guided discriminator reconstructs point clouds with less noise and more precise object structures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.785, + 0.449 + ], + "angle": 0, + "content": "Qualitatively, the visualization comparison with several competitive methods presented in Figure 5 indicates that the proposed upsampling method exhibits results highly similar to the ground truth data, showing clear object details and consistent LiDAR scan lines. We provide additional visualization results in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.451, + 0.787, + 0.572 + ], + "angle": 0, + "content": "LiDAR Point Cloud Inpainting. To mimic the case of missing data in LiDAR point clouds, we mask point clouds in front of the vehicle within a \\(22.5^{\\circ}\\) range and perform inpainting using LiDARGen and our method. Our results achieve much less MAE (0.190) between the reconstructed results and ground truth compared to that of LiDARGen (0.367). As shown in Figure 7, LiDARGen generates roads but fails to recover the masked car (the first row) and the wall (the second row), while we are able to reconstruct the details. More results are presented in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.599, + 0.388, + 0.615 + ], + "angle": 0, + "content": "5.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.628, + 0.66, + 0.642 + ], + "angle": 0, + "content": "All ablation studies are conducted on the KITTI-360 dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Main Ablation. As demonstrated in Table 6, we investigate the contribution of each component. Starting from the backbone model (a) without any component, we gradually add each component for evaluation. By applying Hough Voting to the training, the performance of variant (b) improves significantly, demonstrating the huge impact of the correct range-view data distribution. With the addition of the range-guided discriminator component, model (c) achieves better performance, thus clarifying the effectiveness of the geometry-sensitive module. Compared to model (c), the performance of model (d) decreases because circular convolution generates LiDAR scenes with arbitrary directions due to rotation invariance. However, when combined with direction-conditioned generation that guides the model to generate point clouds in the correct direction, the overall performance improves, as demonstrated by model (e). We refer the readers to the supplementary materials for qualitative results." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Q. Hu et al." + }, + { + "type": "table", + "bbox": [ + 0.278, + 0.144, + 0.726, + 0.212 + ], + "angle": 0, + "content": "
ModelVAEModel Size (M)\\( MMD_{BEV} \\)↓FRD ↓\\( JSD_{BEV} \\)↓Throughput ↑(samples/s)
Ours-DMS×40.3\\( 1.03 × 10^{-4} \\)1392.20.0621.52
Ours-DM×114.7\\( 4.14 × 10^{-5} \\)899.00.0400.64
Ours✓41.4\\( 3.07 × 10^{-5} \\)1074.90.0454.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.213, + 0.787, + 0.255 + ], + "angle": 0, + "content": "Table 8: Ablation of latent diffusion. Model without latent diffusion requires more parameters to achieve similar performance with the latent diffusion model. We generated 1000 samples for calculating the throughput." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.408 + ], + "angle": 0, + "content": "Latent diffusion. We explore the contribution of latent diffusion to the generation quality and generation speed of the model. As shown in Table 8, we constructed two variants of different sizes that directly generate range images with diffusion models, denoted as \"Ours-DMS\" (with a smaller size) and \"Ours-DM\" (with a larger size), for comparison. It turns out that a much larger model size is required for \"Ours-DM\" to achieve competitive performance with \"Ours\", due to the multitude of high-frequency details contained in the range image. Additionally, the superior throughput of \"Ours\" demonstrates the efficiency of latent diffusion." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.41, + 0.46, + 0.484 + ], + "angle": 0, + "content": "VAE Architecture. We investigate the contribution of the proposed range-guided discriminator by comparing our VAE with two baselines: 1) VAE without dis" + }, + { + "type": "table", + "bbox": [ + 0.476, + 0.413, + 0.785, + 0.467 + ], + "angle": 0, + "content": "
DiscriminatorPSNRrange↑ MAE↓ FRD↓ CD
No26.770.0195532.90.0808
Vanilla26.700.0189496.70.0726
Range-Guided27.190.0186483.60.0676
" + }, + { + "type": "table_caption", + "bbox": [ + 0.479, + 0.468, + 0.778, + 0.478 + ], + "angle": 0, + "content": "Table 9: Ablation on the VAE architecture." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.485, + 0.789, + 0.621 + ], + "angle": 0, + "content": "criminalitor; and 2) VAE with vanilla discriminator (i.e., using the original 2D convolution in the discriminator). We evaluate the reconstruction performance of VAEs with 2D and 3D metrics, including PSNR, MAE, FRD and Chamfer Distance (CD) [16]. As listed in Table 9, the VAE with the proposed range-guided discriminator outperforms the baselines in both 2D and 3D reconstruction metrics. This gives credits to the ability of perceiving local 3D structures by the range-guided discriminator. The qualitative results presented in Figure 8 also indicate that the proposed range-guided discriminator ensures the point clouds are reconstructed with less noise and more precise object structures." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.643, + 0.37, + 0.659 + ], + "angle": 0, + "content": "6 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We propose a novel RangeLDM model to generate realistic range-view LiDAR point clouds at a fast speed. We ensure the quality of projection from point clouds to range images with correct distribution via Hough Voting. Then we compress range images to latent features with a VAE, and train a diffusion model in the lower-dimensional latent space. Additionally, we enhance the range-image reconstruction quality of the VAE with a range-guided discriminator. Experiments conducted on KITTI-360 and nuScenes datasets demonstrate the superior generation quality and sampling efficiency of our method. In future, we will explore generating labeled point clouds and creating corner-case data, such as in car accidents and extreme weather conditions, for potential applications in robust self-driving." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.34, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RangeLDM: Fast Realistic LiDAR Point Cloud Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.176, + 0.786, + 0.218 + ], + "angle": 0, + "content": "1. Achlioptas, P., Diamanti, O., Mitlagkas, I., Guibas, L.: Learning representations and generative models for 3D point clouds. In: International Conference on Machine Learning. pp. 40-49. PMLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.786, + 0.273 + ], + "angle": 0, + "content": "2. Ando, A., Gidaris, S., Bursuc, A., Puy, G., Boulch, A., Marlet, R.: RangeViT: Towards vision transformers for 3D semantic segmentation in autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5240-5250 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.273, + 0.786, + 0.315 + ], + "angle": 0, + "content": "3. Bewley, A., Sun, P., Mensink, T., Anguelov, D., Sminchisescu, C.: Range conditioned dilated convolutions for scale invariant 3D object detection. In: Conference on Robot Learning (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.315, + 0.786, + 0.342 + ], + "angle": 0, + "content": "4. Box, G.E., Jenkins, G.M., Reinsel, G.C., Ljung, G.M.: Time series analysis: forecasting and control. John Wiley & Sons (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.786, + 0.384 + ], + "angle": 0, + "content": "5. Caccia, L., Van Hoof, H., Courville, A., Pineau, J.: Deep generative modeling of LiDAR data. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 5034-5040. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.384, + 0.786, + 0.439 + ], + "angle": 0, + "content": "6. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: NuScenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11621-11631 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.439, + 0.786, + 0.481 + ], + "angle": 0, + "content": "7. Cai, R., Yang, G., Averbuch-Elor, H., Hao, Z., Belongie, S., Snavely, N., Hariharan, B.: Learning gradient fields for shape generation. In: European Conference on Computer Vision. pp. 364-381 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.481, + 0.786, + 0.536 + ], + "angle": 0, + "content": "8. Chai, Y., Sun, P., Ngiam, J., Wang, W., Caine, B., Vasudevan, V., Zhang, X., Anguelov, D.: To the point: Efficient 3D object detection in the range image with graph convolution kernels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16000-16009 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.536, + 0.786, + 0.577 + ], + "angle": 0, + "content": "9. Chen, H., Luo, S., Hu, W., et al.: Deep point set resampling via gradient fields. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(3), 2913-2930 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.578, + 0.786, + 0.619 + ], + "angle": 0, + "content": "0. Chen, X., Vizzo, I., Läbe, T., Behley, J., Stachniss, C.: Range image-based LiDAR localization for autonomous vehicles. In: IEEE International Conference on Robotics and Automation (ICRA). pp. 5802-5808. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.786, + 0.66 + ], + "angle": 0, + "content": "1. Cheng, H., Han, X., Xiao, G.: CeNet: Toward concise and efficient LiDAR semantic segmentation for autonomous driving. In: IEEE International Conference on Multimedia and Expo (ICME). pp. 01-06. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.786, + 0.702 + ], + "angle": 0, + "content": "2. Cortinhal, T., Tzelepis, G., Erdal Aksoy, E.: SalsaNext: Fast, uncertainty-aware semantic segmentation of LiDAR point clouds. In: Advances in Visual Computing. pp. 207-222. Springer International Publishing, Cham (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.786, + 0.743 + ], + "angle": 0, + "content": "3. Deng, J., Shi, S., Li, P., Zhou, W., Zhang, Y., Li, H.: Voxel R-CNN: Towards high performance voxel-based 3D object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1201-1209 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.786, + 0.799 + ], + "angle": 0, + "content": "4. Engelcke, M., Rao, D., Wang, D.Z., Tong, C.H., Posner, I.: Vote3Deep: Fast object detection in 3D point clouds using efficient convolutional neural networks. In: 2017 IEEE International Conference on Robotics and Automation (ICRA). pp. 1355-1361. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "5. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12873-12883 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.176, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.128 + ], + "angle": 0, + "content": "Q. Hu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "16. Fan, H., Su, H., Guibas, L.J.: A point set generation network for 3D object reconstruction from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 605-613 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.245 + ], + "angle": 0, + "content": "17. Fan, L., Pang, Z., Zhang, T., Wang, Y.X., Zhao, H., Wang, F., Wang, N., Zhang, Z.: Embracing single stride 3D object detector with sparse transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8458-8468 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "18. Fan, L., Xiong, X., Wang, F., Wang, N., Zhang, Z.: RangeDet: In defense of range view for LiDAR-based 3D object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2918-2927 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "19. Gao, L., Wu, T., Yuan, Y.J., Lin, M.X., Lai, Y.K., Zhang, H.: TM-NET: Deep generative networks for textured meshes. ACM Transactions on Graphics (TOG) 40(6), 1-15 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "20. Gao, L., Yang, J., Wu, T., Yuan, Y.J., Fu, H., Lai, Y.K., Zhang, H.: SDM-NET: Deep generative network for structured deformable mesh. ACM Transactions on Graphics (TOG) 38(6), 1-15 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.411 + ], + "angle": 0, + "content": "21. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.453 + ], + "angle": 0, + "content": "22. He, C., Li, R., Li, S., Zhang, L.: Voxel set transformer: A set-to-set approach to 3D object detection from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8417-8427 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.508 + ], + "angle": 0, + "content": "23. He, Y., Tang, D., Zhang, Y., Xue, X., Fu, Y.: Grad-PU: Arbitrary-scale point cloud upsampling via gradient descent with learned distance functions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5354-5363 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.785, + 0.577 + ], + "angle": 0, + "content": "24. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: GANs trained by a two time-scale update rule converge to a local nash equilibrium. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.605 + ], + "angle": 0, + "content": "25. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33, 6840-6851 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.646 + ], + "angle": 0, + "content": "26. Hu, J.S., Kuai, T., Waslander, S.L.: Point density-aware voxels for LiDAR 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8469-8478 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "27. Hu, P., Ziglar, J., Held, D., Ramanan, D.: What you see is what you get: Exploiting visibility for 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11001-11009 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.716 + ], + "angle": 0, + "content": "28. Hui, K.H., Li, R., Hu, J., Fu, C.W.: Neural wavelet-domain diffusion for 3D shape generation. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.785, + 0.758 + ], + "angle": 0, + "content": "29. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1125-1134 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "30. Kim, H., Lee, H., Kang, W.H., Lee, J.Y., Kim, N.S.: SoftFlow: Probabilistic framework for normalizing flow on manifolds. Advances in Neural Information Processing Systems 33, 16388-16397 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "31. Kim, J., Yoo, J., Lee, J., Hong, S.: SetVAE: Learning hierarchical composition for generative modeling of set-structured data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15059–15068 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.339, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RangeLDM: Fast Realistic LiDAR Point Cloud Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "32. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.787, + 0.202 + ], + "angle": 0, + "content": "33. Kochanov, D., Nejadasl, F.K., Booij, O.: KprNet: Improving projection-based LiDAR semantic segmentation. arXiv preprint arXiv:2007.12668 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.787, + 0.243 + ], + "angle": 0, + "content": "34. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: PointPillars: Fast encoders for object detection from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12697-12705 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.243, + 0.787, + 0.284 + ], + "angle": 0, + "content": "35. Li, B.: 3D fully convolutional network for vehicle detection in point cloud. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1513-1518. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.284, + 0.787, + 0.31 + ], + "angle": 0, + "content": "36. Li, B., Zhang, T., Xia, T.: Vehicle detection from 3D LiDAR using fully convolutional network. arXiv preprint arXiv:1608.07916 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.311, + 0.787, + 0.351 + ], + "angle": 0, + "content": "37. Li, C., Ren, Y., Liu, B.: Pcgen: Point cloud generator for lidar simulation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 11676-11682. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.351, + 0.787, + 0.392 + ], + "angle": 0, + "content": "38. Liao, Y., Xie, J., Geiger, A.: KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2D and 3D. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(3), 3292-3310 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.392, + 0.749, + 0.405 + ], + "angle": 0, + "content": "39. Lim, J.H., Ye, J.C.: Geometric gan. arXiv preprint arXiv:1705.02894 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.406, + 0.787, + 0.446 + ], + "angle": 0, + "content": "40. Litany, O., Bronstein, A., Bronstein, M., Makadia, A.: Deformable shape completion with graph convolutional autoencoders. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1886-1895 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.446, + 0.787, + 0.487 + ], + "angle": 0, + "content": "41. Luo, S., Hu, W.: Diffusion probabilistic models for 3D point cloud generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2837-2845 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.487, + 0.787, + 0.527 + ], + "angle": 0, + "content": "42. Malavazi, F.B., Guyonneau, R., Fasquel, J.B., Lagrange, S., Mercier, F.: LiDAR-only based navigation algorithm for an autonomous agricultural robot. Computers and Electronics in Agriculture 154, 71-79 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.528, + 0.787, + 0.581 + ], + "angle": 0, + "content": "43. Manivasagam, S., Wang, S., Wong, K., Zeng, W., Sazanovich, M., Tan, S., Yang, B., Ma, W.C., Urtasun, R.: LiDARsim: Realistic LiDAR simulation by leveraging the real world. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11167-11176 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.582, + 0.787, + 0.636 + ], + "angle": 0, + "content": "44. Meyer, G.P., Laddha, A., Kee, E., Vallespi-Gonzalez, C., Wellington, C.K.: Laser-Net: An efficient probabilistic 3D object detector for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12677-12686 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.636, + 0.787, + 0.677 + ], + "angle": 0, + "content": "45. Milioto, A., Vizzo, I., Behley, J., Stachniss, C.: RangeNet++: Fast and accurate LiDAR semantic segmentation. In: IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 4213-4220. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.677, + 0.787, + 0.717 + ], + "angle": 0, + "content": "46. Mittal, P., Cheng, Y.C., Singh, M., Tulsiani, S.: AutoSDF: Shape priors for 3D completion, reconstruction and generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 306-315 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.718, + 0.787, + 0.758 + ], + "angle": 0, + "content": "47. Mo, K., Guerrero, P., Yi, L., Su, H., Wonka, P., Mitra, N., Guibas, L.J.: StructureNet: Hierarchical graph networks for 3D shape generation. arXiv preprint arXiv:1908.00575 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.759, + 0.787, + 0.799 + ], + "angle": 0, + "content": "48. Nichol, A., Jun, H., Dhariwal, P., Mishkin, P., Chen, M.: Point-E: A system for generating 3D point clouds from complex prompts. arXiv preprint arXiv:2212.08751 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "49. Pan, X., Xia, Z., Song, S., Li, L.E., Huang, G.: 3D object detection with pointformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7463-7472 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.354, + 0.129 + ], + "angle": 0, + "content": "Q. Hu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.19 + ], + "angle": 0, + "content": "50. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: PointNet: Deep learning on point sets for 3D classification and segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 652-660 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.233 + ], + "angle": 0, + "content": "51. Ran, H., Guizilini, V., Wang, Y.: Towards realistic scene generation with lidar diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.787, + 0.274 + ], + "angle": 0, + "content": "52. Resop, J.P., Lehmann, L., Hession, W.C.: Drone laser scanning for modeling river-scape topography and vegetation: Comparison with traditional aerial LiDAR. Drones 3(2), 35 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.274, + 0.787, + 0.303 + ], + "angle": 0, + "content": "53. Rezende, D., Mohamed, S.: Variational inference with normalizing flows. In: International Conference on Machine Learning. pp. 1530-1538. PMLR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.303, + 0.787, + 0.346 + ], + "angle": 0, + "content": "54. Rezende, D.J., Mohamed, S., Wierstra, D.: Stochastic backpropagation and approximate inference in deep generative models. In: International Conference on Machine Learning. pp. 1278-1286. PMLR (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.787, + 0.389 + ], + "angle": 0, + "content": "55. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.389, + 0.787, + 0.431 + ], + "angle": 0, + "content": "56. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention. pp. 234-241. Springer (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.431, + 0.787, + 0.472 + ], + "angle": 0, + "content": "57. Sallab, A.E., Sobh, I., Zahran, M., Essam, N.: LiDAR sensor modeling and data augmentation with GANs for autonomous driving. arXiv preprint arXiv:1905.07290 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.472, + 0.787, + 0.502 + ], + "angle": 0, + "content": "58. Sauer, A., Chitta, K., Müller, J., Geiger, A.: Projected GANs converge faster. Advances in Neural Information Processing Systems 34, 17480-17492 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.787, + 0.544 + ], + "angle": 0, + "content": "59. Schubert, S., Neubert, P., Pöschmann, J., Protzel, P.: Circular convolutional neural networks for panoramic images and laser data. In: IEEE Intelligent Vehicles Symposium (IV). pp. 653-660. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.544, + 0.787, + 0.587 + ], + "angle": 0, + "content": "60. Shi, G., Li, R., Ma, C.: PillarNet: Real-time and high-performance pillar-based 3D object detection. In: European Conference on Computer Vision. pp. 35-52. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.787, + 0.642 + ], + "angle": 0, + "content": "61. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: PV-RCNN: Point-voxel feature set abstraction for 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10529-10538 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.642, + 0.787, + 0.685 + ], + "angle": 0, + "content": "62. Shi, S., Wang, X., Li, H.: PointRCNN: 3D object proposal generation and detection from point cloud. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 770-779 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.685, + 0.787, + 0.728 + ], + "angle": 0, + "content": "63. Shi, S., Wang, Z., Shi, J., Wang, X., Li, H.: From points to parts: 3D object detection from point cloud with part-aware and part-aggregation network. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(8), 2647-2664 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.728, + 0.787, + 0.77 + ], + "angle": 0, + "content": "64. Shi, W., Rajkumar, R.: Point-GNN: Graph neural network for 3D object detection in a point cloud. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1711-1719 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.77, + 0.787, + 0.812 + ], + "angle": 0, + "content": "65. Shu, D.W., Park, S.W., Kwon, J.: 3D point cloud generative adversarial network based on tree structured graph convolutions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3859-3868 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.812, + 0.787, + 0.841 + ], + "angle": 0, + "content": "66. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.339, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RangeLDM: Fast Realistic LiDAR Point Cloud Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "67. Sun, P., Tan, M., Wang, W., Liu, C., Xia, F., Leng, Z., Anguelov, D.: SWFormer: Sparse window transformer for 3D object detection in point clouds. In: European Conference on Computer Vision. pp. 426-442. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.245 + ], + "angle": 0, + "content": "68. Sun, P., Wang, W., Chai, Y., Elsayed, G., Bewley, A., Zhang, X., Sminchisescu, C., Anguelov, D.: RSN: Range sparse net for efficient, accurate LiDAR 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5725-5734 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.246, + 0.788, + 0.3 + ], + "angle": 0, + "content": "69. Sun, Y., Wang, Y., Liu, Z., Siegel, J., Sarma, S.: PointGrow: Autoregressively learned point cloud generation with self-attention. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 61-70 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.342 + ], + "angle": 0, + "content": "70. Tan, Q., Gao, L., Lai, Y.K., Xia, S.: Variational autoencoders for deforming 3D mesh models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5841-5850 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.384 + ], + "angle": 0, + "content": "71. Tian, Z., Chu, X., Wang, X., Wei, X., Shen, C.: Fully convolutional one-stage 3D object detection on LiDAR range images. Advances in Neural Information Processing Systems 35, 34899-34911 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.384, + 0.788, + 0.412 + ], + "angle": 0, + "content": "72. Vahdat, A., Kreis, K., Kautz, J.: Score-based generative modeling in latent space. Advances in Neural Information Processing Systems 34, 11287-11302 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.412, + 0.788, + 0.453 + ], + "angle": 0, + "content": "73. Valsesia, D., Fracastoro, G., Magli, E.: Learning localized generative models for 3D point clouds via graph convolution. In: International Conference on Learning Representations (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.453, + 0.788, + 0.481 + ], + "angle": 0, + "content": "74. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in Neural Information Processing Systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.481, + 0.788, + 0.522 + ], + "angle": 0, + "content": "75. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in Neural Information Processing Systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.577 + ], + "angle": 0, + "content": "76. Wang, H., Shi, S., Yang, Z., Fang, R., Qian, Q., Li, H., Schiele, B., Wang, L.: RBGNet: Ray-based grouping for 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1110-1119 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.577, + 0.788, + 0.62 + ], + "angle": 0, + "content": "77. Wang, Y., Fathi, A., Kundu, A., Ross, D.A., Pantofaru, C., Funkhouser, T., Solomon, J.: Pillar-based object detection for autonomous driving. In: European Conference on Computer Vision. pp. 18-34. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.62, + 0.788, + 0.647 + ], + "angle": 0, + "content": "78. Weiss, U., Biber, P.: Plant detection and mapping for agricultural robots using a 3D LiDAR sensor. Robotics and Autonomous Systems 59(5), 265-273 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.647, + 0.788, + 0.689 + ], + "angle": 0, + "content": "79. Xiong, Y., Ma, W.C., Wang, J., Urtasun, R.: Learning compact representations for LiDAR completion and generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1074-1083 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.731 + ], + "angle": 0, + "content": "80. Xu, C., Wu, B., Wang, Z., Zhan, W., Vajda, P., Keutzer, K., Tomizuka, M.: SqueezeSegV3: Spatially-adaptive convolution for efficient point-cloud segmentation. In: European Conference on Computer Vision. pp. 1-19. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.731, + 0.788, + 0.758 + ], + "angle": 0, + "content": "81. Yan Yan, Y.M., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.758, + 0.788, + 0.799 + ], + "angle": 0, + "content": "82. Yang, B., Luo, W., Urtasun, R.: Pixor: Real-time 3D object detection from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7652-7660 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.788, + 0.842 + ], + "angle": 0, + "content": "83. Yang, G., Huang, X., Hao, Z., Liu, M.Y., Belongie, S., Hariharan, B.: PointFlow: 3D point cloud generation with continuous normalizing flows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4541-4550 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.129 + ], + "angle": 0, + "content": "Q. Hu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "84. Yang, Z., Sun, Y., Liu, S., Jia, J.: 3DSSD: Point-based 3D single stage object detector. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11040-11048 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.231 + ], + "angle": 0, + "content": "85. Yang, Z., Sun, Y., Liu, S., Shen, X., Jia, J.: STD: Sparse-to-dense 3D object detector for point cloud. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1951-1960 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "86. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3D object detection and tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11784-11793 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.313 + ], + "angle": 0, + "content": "87. Yu, J., Li, X., Koh, J.Y., Zhang, H., Pang, R., Qin, J., Ku, A., Xu, Y., Baldridge, J., Wu, Y.: Vector-quantized image modeling with improved VQGAN. arXiv preprint arXiv:2110.04627 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "88. Yu, L., Li, X., Fu, C.W., Cohen-Or, D., Heng, P.A.: Pu-Net: Point cloud upsampling network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2790-2799 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.397 + ], + "angle": 0, + "content": "89. Zeng, X., Vahdat, A., Williams, F., Gojcic, Z., Litany, O., Fidler, S., Kreis, K.: LION: Latent point diffusion models for 3D shape generation. Advances in Neural Information Processing Systems 35, 10021-10039 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "90. Zhang, J., Singh, S.: LOAM: LiDAR odometry and mapping in real-time. In: Robotics: Science and Systems. vol. 2, pp. 1-9. Berkeley, CA (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.426, + 0.785, + 0.48 + ], + "angle": 0, + "content": "91. Zhang, Y., Hu, Q., Xu, G., Ma, Y., Wan, J., Guo, Y.: Not all points are equal: Learning highly efficient point-based detectors for 3D LiDAR point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18953-18962 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "92. Zhao, Y., Bai, L., Huang, X.: FidNet: LiDAR point cloud semantic segmentation with fully interpolation decoding. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 4453-4458. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.563 + ], + "angle": 0, + "content": "93. Zhou, L., Du, Y., Wu, J.: 3D shape generation and completion through point-voxel diffusion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5826-5835 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.605 + ], + "angle": 0, + "content": "94. Zhou, Y., Tuzel, O.: VoxelNet: End-to-end learning for point cloud based 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4490-4499 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.633 + ], + "angle": 0, + "content": "95. Zyrianov, V., Zhu, X., Wang, S.: Learning to generate realistic LiDAR point clouds. In: European Conference on Computer Vision. pp. 17-35. Springer (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.633 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/RangeLDM_ Fast Realistic LiDAR Point Cloud Generation/ec4004eb-e76a-4f41-9891-f055a3eeaf7c_origin.pdf b/2024/RangeLDM_ Fast Realistic LiDAR Point Cloud Generation/ec4004eb-e76a-4f41-9891-f055a3eeaf7c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2f4f66a53e4177f4e091d64d9e4efadcbe0e5ca4 --- /dev/null +++ b/2024/RangeLDM_ Fast Realistic LiDAR Point Cloud Generation/ec4004eb-e76a-4f41-9891-f055a3eeaf7c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:917ed7904782dee5375ffcd9a9a88b2d1a35534e88ab64929661de87225bef47 +size 7941855 diff --git a/2024/RangeLDM_ Fast Realistic LiDAR Point Cloud Generation/full.md b/2024/RangeLDM_ Fast Realistic LiDAR Point Cloud Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..784a5a0cb0ccb73e71aef17b6056461bff28d82a --- /dev/null +++ b/2024/RangeLDM_ Fast Realistic LiDAR Point Cloud Generation/full.md @@ -0,0 +1,379 @@ +# RangeLDM: Fast Realistic LiDAR Point Cloud Generation + +Qianjiang Hu, Zhimin Zhang, and Wei Hu + +Wangxuan Institute of Computer Technology, Peking University, Beijing, China hqjpu@pku.edu.cn zm_zhang@stu.pku.edu.cn forhuwei@pku.edu.cn + +![](images/bec2e48447c9c338c2bbaeaff95f297df8162460662d1d078c38a0ffad5f51d5.jpg) +(a) Unconditional Generation + +![](images/454315a4e9010a1ee50a3bec5d60c4f2c2ccecdffce1caba0a9f0534c53bcad9.jpg) +(b) Conditional Generation + +![](images/2555aee352ba07603cd717c1df9adb4d8080dfcf1c5a8735f801122ecd200ee3.jpg) +(c) Quality vs. Speed +Fig. 1: (a). Unconditional LiDAR point cloud generation with realistic global structure. (b). Conditional LiDAR point cloud generation, including LiDAR point cloud upsampling and inpainting. (c). Generation quality (Maximum Mean Discrepancy, abbr. MMD) vs. generation speed (samples/s) of competitive LiDAR point cloud generation methods on the KITTI-360 [38] dataset. The proposed method outperforms the state-of-the-art methods LiDARGen [95] and UltraLiDAR [79] in both generation quality and generation speed. All speeds are evaluated on a single RTX 3090 GPU. + +Abstract. Autonomous driving demands high-quality LiDAR data, yet the cost of physical LiDAR sensors presents a significant scaling-up challenge. While recent efforts have explored deep generative models to address this issue, they often consume substantial computational resources with slow generation speeds while suffering from a lack of realism. To address these limitations, we introduce RangeLDM, a novel approach for rapidly generating high-quality range-view LiDAR point clouds via latent diffusion models. We achieve this by correcting range-view data distribution for accurate projection from point clouds to range images via Hough voting, which has a critical impact on generative learning. We then compress the range images into a latent space with a variational autoencoder, and leverage a diffusion model to enhance expressivity. Additionally, we instruct the model to preserve 3D structural fidelity by devising a range-guided discriminator. Experimental results on KITTI-360 and nuScenes datasets demonstrate both the robust expressiveness and fast speed of our LiDAR point cloud generation. + +Keywords: Diffusion models $\cdot$ Autonomous systems $\cdot$ Point clouds + +# 1 Introduction + +Autonomous systems have drawn great attention in both the academia and the industry community. Numerous autonomous systems leverage the power of various sensors and deep learning to enhance the perception of the 3D world. LiDAR (Light Detection And Ranging), with its ability to provide precise 3D geometric information about the surroundings, has become a popular sensor choice for autonomous systems, including self-driving cars [61,62,86], surveying drones [52,90] and robots [42,78]. + +However, while LiDAR offers accurate geometric measurements, it comes with a significant limitation: the data collection process is exceedingly expensive and challenging to scale up. This renders physical sensors impractical for scalable and customizable data collection. Additionally, it is difficult to collect data in corner cases such as car accidents and extreme weather conditions in the field. + +One approach to mitigate this issue is to employ existing LiDAR simulation toolkits [37, 43] to synthesize more data. Nevertheless, these systems typically demand manual scene creation or rely on multiple prior scans of the real world. Another approach is leveraging deep generative models to generate LiDAR point clouds. This is challenging due to the high degree of unstructuredness, sparsity, and non-uniformity in LiDAR point clouds. Caccia et al. [5] proposed LiDAR GAN and LiDAR VAE to generate LiDAR point clouds with generative adversarial networks (GANs) [21] and variational autoencoders (VAEs) [32], respectively. However, the generation results often suffer from issues such as fuzzy or missing details. LiDARGen [95] introduced a novel score-based model to synthesis LiDAR point clouds, which however sampled slowly and failed to generate high-quality geometric details at a far range. UltraLiDAR [79] proposed to synthesize voxelized point clouds in bird's-eyes-view (BEV) with vector quantized VAE (VQ-VAE) [74], resulting in more realistic point clouds than previous methods. Nevertheless, due to the large size and sparsity of point clouds in BEV, most computing power of UltraLiDAR is devoted to generating empty voxels, which results in a low generation speed. + +To this end, we propose RangeLDM, a novel approach based on latent diffusion models (LDMs) [55], which improves both the quality and speed of LiDAR point cloud generation significantly. Firstly, we introduce a range image view to represent point clouds. The reasons are twofold: 1) Range images are compact and closely mimic the sampling conditions of LiDAR, making them a suitable representation. A LiDAR point cloud essentially constructs a 2.5D scene from a single viewpoint instead of a full 3D point cloud [27]. Consequently, organizing the point cloud in range view ensures that no information is overlooked. 2) 2D image generation techniques are relatively mature, providing a solid foundation for our work. While LiDARGen [95] also represents point clouds as range images, it suffers from blurriness caused by inaccurate projection. In contrast, we provide insights that the correct range-view data distribution has a critical impact on such range-view-based generative model learning, and accurately project point clouds onto range images using parameters estimated by Hough Voting. Then, to achieve a faster sampling speed [55] and greater expressivity [55,72], we compress + +range images to the latent space with a VAE and develop a diffusion model operating on the latent space, leveraging the successful paradigm of LDMs [55,72,89]. Further, in order to enhance the ability of the VAE in reconstructing 3D structures, we introduce a range-guided discriminator, which is supervised from the spherical coordinates and thus geometry-sensitive. This plays a crucial role in guiding the decoder to generate high-quality range images, ultimately preserving the fidelity of the 3D structure. Extensive experimental results show that the proposed RangeLDM achieves the state-of-the-art performance on KITTI-360 [38] and nuScenes [6] datasets in terms of both the generation quality and generation speed, as demonstrated in Figure 1 (c). The LiDAR upsampling and LiDAR inpainting results also demonstrate the potential of the proposed method for conditional generation, as presented in Figure 1 (a) and (b). + +To summarize, the main contributions of this paper include: + +- We propose a latent diffusion model to capture the distribution of range-view LiDAR point clouds, aiming to generate realistic point cloud scenes at a fast speed. +- We enlighten the significance of the correct range-view data distribution for range-view-based generative models and achieve high-quality range image projection via Hough Voting. +- We exploit a range-guided discriminator to ensure preserving the fidelity of the 3D geometric structure in generated point clouds. +- Experimental results on the KITTI-360 and nuScenes datasets highlight that our approach outperforms state-of-the-art methods in terms of both visual quality and especially generation speed. + +# 2 Related Work + +LiDAR Representation. LiDAR point clouds can be represented in various forms, encompassing raw point clouds, voxels, range views, and multi-view fusion. Point-based models [49, 62, 64, 76, 84, 85, 91] directly encode 3D objects from raw points using the PointNet [50] encoder and subsequently perform detection or segmentation based on point features. These methods wholly preserve the irregularity and locality of a point cloud but have relatively higher latency. Voxel-based methods [13, 14, 17, 34, 35, 60, 63, 67, 77, 81, 82, 86, 94] voxelize point clouds for convolutional neural networks to efficiently capture features. They are computationally effective but the desertion of fine-grained patterns degrades further refinement. Given that the range view is compact and compatible with the LiDAR sensor's sampling process, range-image-based methods [2, 8, 10-12, 18, 33, 36, 44, 45, 68, 71, 80, 92] have been proposed to directly process range images for LiDAR perception. Multi-view fusion methods [22, 26, 61] amalgamate multiple representations, which yield better results at the expense of processing speed. In this paper, we adopt the range-view representation, which is congruent with the LiDAR sensor's sampling process and can be efficiently encoded by image generation models. + +Generative Models for Point Clouds. Given observed samples of interest, generative models aim to learn the underlying distribution of the data and generate new samples. The first 3D generative models [1,65,73] for point clouds are based on GANs [21]. They operate by generating point clouds and discriminating them from real samples in an adversarial manner. The second category of methods [19,20,31,40,46,47,70] employ VAEs [32] to explore the probabilistic latent space of 3D shapes. Auto-regressive models [4] have also been introduced to generate point clouds [69] from scratch or semantic contexts. PointFlow [83] and SoftFlow [30] leverage normalizing flows [53] to capture the likelihood of shapes and points. Notably, the recent success of denoising diffusion models (DDMs) [25] for image synthesis [55] has also been extended to the domain of 3D point cloud generation [7,28,41,48,89,93]. However, these previous works often concentrate on the object level and are not well-suited for handling large scenes. + +In this paper, our focus lies in the generation of LiDAR point clouds. While [5] and [57] employ VAE or GANs for LiDAR point cloud generation, the realism achieved in their results is relatively limited. UltraLiDAR [79], on the other hand, utilizes VQ-VAE to generate voxelized LiDAR point clouds, but at the expense of introducing quantified losses and slow generation speed. LiDARGen [95] proposed a novel score-matching energy-based model to generate higher-quality Li-DAR point clouds, but it samples very slowly and has degraded geometric details at a far range. A contemporary work LiDM [51] also utilized LDMs to generate LiDAR point clouds. However, while they focused on generating LiDAR point clouds conditional on multimodal data, we focus on improving the quality of point cloud generation directly. + +# 3 Background on Denoising Diffusion Models + +DDMs [25] are latent variable models that employ a pre-defined posterior distribution, known as the forward diffusion process, and are trained with a denoising objective. More specifically, given samples $\mathbf{x}_0\sim q(\mathbf{x}_0)$ from a data distribution, DDMs follow the form $p_{\theta}(\mathbf{x}_0)\coloneqq \int p_{\theta}(\mathbf{x}_{0:T})d\mathbf{x}_{1:T}$ , where $T$ denotes the number of steps, $\mathbf{x}_1,\dots,\mathbf{x}_T$ are latent variables that gradually add noise to the data $\mathbf{x}_0$ , and $\theta$ denotes the parameter set of the DDM decoder. + +DDMs are trained by minimizing the evidence lower bound (ELBO) of the data $\mathbf{x}_0$ under $p_{\theta}(\mathbf{x}_{0:T})$ . This objective can be simplified to [25]: + +$$ +\min _ {\theta} \mathbb {E} _ {\mathbf {x}, \epsilon \sim \mathcal {N} (0, 1), t} \left[ \| \epsilon - \epsilon_ {\theta} (\mathbf {x} _ {t} + \epsilon , t) \| _ {2} ^ {2} \right], \tag {1} +$$ + +with $t$ uniformly sampled from $\{1,\dots,T\}$ . + +# 4 The Proposed Method + +As shown in Fig. 2, we first project point clouds onto high-quality range images (as detailed in Section 4.1), considering the compact nature of the range view, + +![](images/81e720c0c8cc65c583bd34340169d7c339a66c8720c285c53ecfb5c22fc94402.jpg) +Fig. 2: The framework of the proposed RangeLDM. Firstly, we project point clouds onto high-quality range images via Hough Voting (Section 4.1). Subsequently, we train a VAE to compress the range images into low-dimensional latent features $\mathbf{z}_0$ , which encodes the range images with the encoder $\mathcal{E}_{\zeta}$ and reconstructs range images from latent features with the decoder $\mathcal{G}_{\eta}$ (Section 4.2). Here, a range-guided discriminator $\mathcal{D}_{\tau}$ is introduced to guide the decoder in the reconstruction of 3D structures. We finally train a latent diffusion model to capture the distribution of the latent features (Section 4.2). With optional conditional inputs, the proposed method is applicable to tasks such as point cloud upsampling and inpainting (Section 4.3). + +which aligns seamlessly with the sampling process of LiDAR sensors. Our model training process is then focused on these projected 2D range images. We then compress range images into a low-dimensional latent space through a VAE (as detailed in Section 4.2). Subsequently, we proceed to train DDMs within the reduced-dimensional latent space (as discussed in Section 4.2). We discuss applications of the proposed model in Section 4.3, which involves unconditional generation and conditional generation. Implementation details are explained in Section 4.4. + +# 4.1 High-Quality Range Projection + +The range image presents LiDAR data compactly and intuitively, with rows indicating the laser beams and columns representing the yaw angles. We convert point clouds to range images using spherical projection. Typically, for a point $\mathbf{p}$ in Cartesian coordinates $(x,y,z)$ , we calculate its spherical coordinates $(r,\theta,\phi)$ using: + +$$ +r = \sqrt {x ^ {2} + y ^ {2} + z ^ {2}}, \theta = \mathrm {a t a n} (y, x), \phi = \mathrm {a t a n} \left(z, \sqrt {x ^ {2} + y ^ {2}}\right). \qquad (2) +$$ + +However, in most current datasets such as KITTI-360, multiple lasers from the Velodyne LiDAR system do not share a common origin for their measurements. This may introduce errors in the direct conversion from Cartesian points to spherical points, resulting in incorrect range-view data distribution and thus low-quality range images, as shown at the top of Fig. 3. + +To address this issue, we adopt Hough Voting to estimate heights and pitch angles $\{h_j,\phi_j\}_{j = 1,\dots ,N}$ for Velodyne sensors [3]. We then adjust the point cloud transformation to a range image using + +$$ +r = \sqrt {x ^ {2} + y ^ {2} + (z - h _ {j}) ^ {2}}, \theta = \mathrm {a t a n} (y, x), \phi = \phi_ {j}, \tag {3} +$$ + +![](images/6a8c524874c6ef0c93606503b68a308b8287dbb1a00c4117a9abdee57d58ca3e.jpg) +Fig. 3: Comparison of range projection by the typical method described in Eq. 2 and our method with Hough Voting as in Eq. 3. + +![](images/c1c194f7f013d7decc561a9f467f07ec1311cdf6a8edfb270d0c0e7f7b2d0307.jpg) + +where $h_j$ and $\phi_j$ refer to the $j$ -th Velodyne sensor. + +We then rasterize points $(r,\theta ,\phi)$ into a 2D cylindrical projection $R(u,v)$ (a.k.a., range image) of size $H\times W$ with $u = ((\theta +\pi) / 2\pi)W,v = j$ , where $(u,v)$ denotes the grid coordinate of a point in the range image. Thus, we obtain high-quality range images as illustrated at the bottom of Fig. 3. We denote the obtained range image as $\mathbf{x}\in \mathbb{R}^{H\times W\times 2}$ , which comprises $H\times W$ pixels associated with both range and intensity $\{r,i\}$ . + +To verify the impact of range projection, we trained range-based LiDAR generation methods like LiDARGen [95] with our obtained high-quality range images. Table 1 shows substantial improvement, which sheds light on the significance of the correct range-view data distribution for such range-based generative model learning. + +# 4.2 Training + +As a range image is often of high dimensionality at the scale of $64 \times 1024$ , it would be computationally expensive to learn the distribution directly. Instead, we compress the range image into a low-dimensional latent space via a VAE and then model the distribution via an LDM. Consequently, the training of the proposed RangeLDM includes two distinct stages. In the first stage, we train a regular VAE to compress the range image into latent features $\mathbf{z}_0 \in \mathbb{R}^{h \times w \times c}$ by a downsampling factor $f = H / h = W / w$ . In the second stage, we train an LDM to learn the distribution of the latent encoding $\mathbf{z}_0$ . + +First-Stage: Dimensionality Reduction. In the first stage, we reduce the dimensionality of the obtained range images by a VAE, which removes imperceptible high-frequency details and leads to low-dimensional latent features that encode prominent information in the original range image. + +A standard VAE consists of two main components: 1) an encoder $\mathcal{E}_{\zeta}$ that transforms an input range image $\mathbf{x} \in \mathbb{R}^{H \times W \times 2}$ into latent features $\mathbf{z}_0 = \mathcal{E}_{\zeta}(\mathbf{x}) \in \mathbb{R}^{h \times w \times c}$ and 2) a decoder $\mathcal{G}_{\eta}$ that takes the latent features $\mathbf{z}_0$ as input and generates the reconstructed range image $\hat{\mathbf{x}} = \mathcal{G}_{\eta}(\mathbf{z}_0)$ . The entire VAE is trained by maximizing a modified ELBO with respect to the parameters $\zeta$ and $\eta$ [32,54]: + +$$ +\mathcal {L} _ {E L B O} (\zeta , \eta) = \mathbb {E} _ {q _ {\zeta} (\mathbf {z} _ {0} | \mathbf {x})} \left[ \log p _ {\boldsymbol {\eta}} (\mathbf {x} \mid \mathbf {z} _ {0}) \right] - \lambda_ {1} D _ {\mathrm {K L}} \left(q _ {\zeta} (\mathbf {z} _ {0} \mid \mathbf {x}) \| p (\mathbf {z} _ {0})\right). \tag {4} +$$ + +In the above equation, the first term evaluates the reconstruction likelihood of the decoder from the latent $\mathbf{z}_0$ and corresponds to an $L_{1}$ reconstruction loss, + +while the second term quantifies how closely the learned distribution of $\mathbf{z}_0$ resembles the prior distribution $p(\mathbf{z}_0) = \mathcal{N}(\mathbf{z}_0; \mathbf{0}, \mathbf{I})$ held over latent variables. The hyperparameter $\lambda_1$ balances the reconstruction accuracy and Kullback-Leibler regularization. + +Range-Guided Discriminator. In order to mitigate the blurriness introduced by relying solely on the pixel-space $L_{1}$ reconstruction loss, we integrate the VAE with a patch-based adversarial discriminator [15, 29, 55, 87]. Furthermore, to ensure that the reconstruction remains confined within the manifold underlying the range image and exploit the geometric information from the spherical coordinates, we propose a range-guided discriminator $\mathcal{D}_{\tau}$ , with parameters $\tau$ . Specifically, we adapt the Meta-Kernel [18] to replace the standard convolution in the discriminator, aiming to learn convolution weights from relative spherical coordinates. The Meta-Kernel is formulated as + +$$ +\mathbf {h} _ {i} ^ {\prime} = \mathcal {W} \left(\mathcal {A} _ {j \in \mathcal {N} (i)} \left(\Phi (\gamma (\mathbf {p} _ {j}, \mathbf {p} _ {i})) \odot \mathbf {h} _ {j}\right)\right), \tag {5} +$$ + +where $\mathbf{h}$ and $\mathbf{h}'$ represent the input feature and output feature, respectively. The function $\varPhi$ denotes a Multi-Layer Perceptron (MLP) with two fully-connected layers, $\mathcal{A}$ is a concatenation operation, $\mathcal{N}(i)$ denotes the 2D-grid neighborhood around the center pixel $i$ , and $\mathcal{W}$ is a fully-connected layer. The term $\gamma (\mathbf{p}_j,\mathbf{p}_i)$ represents the distance between points $\mathbf{p}_j$ and $\mathbf{p}_i$ in spherical coordinates, which is expressed as + +$$ +\begin{array}{l} \gamma \left(\mathbf {p} _ {j}, \mathbf {p} _ {i}\right) := \left\{r _ {j} \cos (\Delta \theta) \cos (\Delta \phi) - r _ {i}, r _ {j} \cos (\Delta \theta) \sin (\Delta \phi), r _ {j} \sin (\Delta \theta) \right\}, \\ \Delta \theta = \theta_ {j} - \theta_ {i}, \quad \Delta \phi = \phi_ {j} - \phi_ {i}. \\ \end{array} +$$ + +The Meta-Kernel is aware of local 3D structures by adaptively adjusting convolution kernel weights from relative spherical coordinates. This makes it challenging for the decoder to deceive the range-guided discriminator. This, in turn, encourages the decoder to generate more realistic range images. + +With the range-guided discriminator, the final objective of the first-stage training is: + +$$ +\zeta , \eta = \arg \min _ {\zeta , \eta} \max _ {\tau} \mathcal {L} _ {E L B O} (\zeta , \eta) + \lambda_ {2} \mathcal {L} _ {a d v} (\zeta , \eta , \tau), \tag {7} +$$ + +where $\mathcal{L}_{adv}$ is the GAN Hinge loss [39] and $\lambda_{2}$ balances the two losses. + +Second-Stage: Latent Diffusion Modeling In the second stage, we freeze the encoder and decoder of the learned VAE and train an LDM on the encoded latent feature $\mathbf{z}_0$ for distribution modeling of the range image. During the diffusion process, noise is added to the initial latent $\mathbf{z}_0$ , resulting in a noisy latent $\mathbf{z}_t$ , where the noise level increases over time steps $t \in T$ . + +Regarding the modeling of the unconditional distribution $p(\mathbf{z}_0)$ , based on Eq. 1, we learn a time-conditional UNet [56] $\epsilon_{\theta}(z_t,t)$ that predicts the noise added to the noisy latent $\mathbf{z}_t$ : + +$$ +L _ {\mathrm {L D M}} := \mathbb {E} _ {\mathcal {E} (x), \epsilon \sim \mathcal {N} (0, 1), t} \left[ \| \epsilon - \epsilon_ {\theta} \left(z _ {t}, t\right) \| _ {2} ^ {2} \right]. \tag {8} +$$ + +As to the modeling of the conditional distribution $p(\mathbf{z}_0|\mathbf{y})$ , where $\mathbf{y}$ is a condition such as sparse or masked point clouds, we learn a conditional LDM via + +$$ +L _ {\mathrm {L D M}} := \mathbb {E} _ {\mathcal {E} (x), \mathbf {y}, \epsilon \sim \mathcal {N} (0, 1), t} \left[ \| \epsilon - \epsilon_ {\theta} (\mathbf {z} _ {t}, t, \tau_ {\theta} (\mathbf {y})) \| _ {2} ^ {2} \right], \tag {9} +$$ + +where $\tau_{\theta}$ is a condition encoder that projects $\mathbf{y}$ into an intermediate representation $\tau_{\theta}(\mathbf{y})$ , and $\epsilon_{\theta}(\mathbf{z}_t, t, \tau_{\theta}(\mathbf{y}))$ is a UNet interpolated with cross-attention layers [55, 75]. + +# 4.3 Generation + +Unconditional Generation. With the decoder $\mathcal{G}_{\eta}$ and the LDM $\epsilon_{\theta}$ , we establish a hierarchical generative model $p_{\eta, \theta}(\mathbf{x}, \mathbf{z}_0) = p_{\eta}(\mathbf{x} | \mathbf{z}_0) p_{\theta}(\mathbf{z}_0)$ . That is, we first generate latent features with the LDM, and then map the latent features back to the original range image space using the decoder $\mathcal{D}_{\eta}$ for LiDAR point cloud generation. + +Conditional Generation. Given a condition $\mathbf{y}$ , the conditional generative model is defined as $p_{\eta, \theta}(\mathbf{x}, \mathbf{z}_0 | \mathbf{y}) = p_{\eta}(\mathbf{x} | \mathbf{z}_0, \mathbf{y}) p_{\theta}(\mathbf{z}_0 | \mathbf{y})$ . There could be various conditions for LiDAR point cloud generation, such as sparse point clouds, incomplete point clouds, camera or even text. Here, we illustrate two applications of conditional generation: LiDAR point cloud upsampling and inpainting. + +LiDAR Point Cloud Upsampling. This task takes a sparse point cloud $^1$ $\mathbf{y} \in \mathbb{R}^{h \times W \times 2}$ as input (following the setting of LiDARGen) and expects the model to produce a denser counterpart $\hat{\mathbf{x}} \in \mathbb{R}^{H \times W \times 2}$ , which is necessary for accurate LiDAR-based perception and density-insensitive domain adaptation. We train LDM following Eq. 9 with the sparse point cloud $\mathbf{y}$ and its ground-truth dense variant $\mathbf{x}$ . The condition encoder $\tau_{\theta}$ is configured as a reshaper that transforms the sparse point cloud $\mathbf{y}$ into $\mathbf{y}' \in \mathbb{R}^{h \times w \times 2f}$ . The reshaped $\mathbf{y}'$ is then concatenated with noisy latent variable $\mathbf{z}_t$ and fed into the UNet $\epsilon_{\theta}$ for noise prediction. + +LiDAR Point Cloud Inpainting. This task is required when LiDAR sensors cannot capture the entire scene due to object occlusion or sensor limitations. We train an inpainting LDM from the ground truth point cloud $\mathbf{x}$ , a masked point cloud $\mathbf{x}'$ , and the corresponding mask $\mathbf{m} \in \{0,1\}^{H \times W}$ . The masked point cloud $\mathbf{x}'$ is encoded with the encoder $\mathcal{E}_{\zeta}$ from the VAE and is then concatenated with a downsampled mask $\mathbf{m}' \in \{0,1\}^{h \times w}$ as the condition term in Eq. 9. Mathematically, we represent the condition $\mathbf{y}$ as $\{\mathbf{x}',\mathbf{m}\}$ , and deploy the condition encoder to project it into $\tau_{\theta}(\mathbf{y}) = \mathcal{E}_{\zeta}(\mathbf{x}') \| ds(\mathbf{m})$ , where $\|\cdot\|$ denotes the concatenation operation and $ds(\cdot)$ indicates the downsampling operation. The projected condition is then concatenated with the noisy latent variable $\mathbf{z}_t$ and input into the UNet $\epsilon_{\theta}$ for noise prediction. + +![](images/a928bec7ff81d9dcf368aa8a104b347ccb364abccb06399f799be61df6d7c880.jpg) +Fig. 4: Qualitative results comparing against baselines for unconditional Li-DAR generation on KITTI-360. Real point clouds are only for reference. Our model produces results that closely resemble real-world data, which excels in generating road scenes, such as cars (the first row), road bollards (the second row) and crossroads (the last row). + +# 4.4 Implementation Details + +Circular Convolution. Range images are inherently circular, which means that the left boundary of a range image is actually connected to its right boundary. However, standard convolutions employ zero-padding or symmetry padding and do not take into account such constraints. Thus, as in LiDArgen [95], we replace standard convolutions in both the VAE and LDM with circular convolutions [59], which treat the left and right boundaries as connected neighbors in the topology. Since the 2D convolution operator satisfies translation invariance, it is evident that the entire network is invariant to horizontal shifting of the range image (i.e., rotation in the xy-plane of the point cloud). + +Conditioning the Model with Direction. Real LiDAR point clouds collected from driving scenes usually have a certain directionality. For example, since vehicles typically follow the direction of the road, the x-axis direction of the LiDAR point cloud usually aligns with the road direction. However, since the whole network of our model is invariant to rotation in the xy-plane of the point cloud, the point clouds generated from random noise could result in arbitrary directions. To avoid this, we generate point clouds with directional conditioning. Specifically, we use an $h \times w$ matrix as the condition, in which only the values from the first column are set to 1, while the rest are all set to 0. This 1-valued column corresponds to the x-axis of the point cloud. Such simple conditioning demonstrates effective control over the direction of the generated point cloud, as presented in the qualitative results of the supplementary material. + +Network Architectures and Model Hyperparameters. For simplicity, we adopted a similar architecture to that in [55]. The downsampling factor $f$ is set as 4 to strike a good balance between efficiency and realistic results. The VAE comprises three encoder blocks and three decoder blocks. The LDM consists of four downsampling blocks with the last three followed by a transformer layer, and four upsampling blocks with the first three blocks also followed by a transformer layer. In the generation process, we employed 50 denoising steps with DDIM sampler [66] for point cloud generation. More implementation details can be found in released code https://github.com/WoodwindHu/RangeLDM. + +![](images/c1f715f0ec81ea3370f4aec310bbc2f0abae528348e19494642888e0efad7f3b.jpg) +Fig. 5: Comparison of upsampling results on KITTI-360. We downsampled the ground truth by a factor of four as the input, and demonstrated the results of different methods on $4 \times$ -upsampling of the input. + +# 5 Experiments + +# 5.1 Experimental Setup + +Datasets. We evaluate our model on two challenging datasets, KITTI-360 and nuScenes. KITTI-360 has 81,106 LiDAR readings from 9 sequences in Germany, covering diverse scenes. We used the first two sequences for validation and trained on the rest. NuScenes is a public dataset with 297,737 LiDAR sweeps in the training set and 52,423 in the testing set, collected in Boston and Singapore. + +Evaluation metrics. Following [95], we employed three metrics to perform quantitative analysis: Maximum Mean Discrepancy (MMD), Jensen-Shannon divergence (JSD), and Frechet Range Distance (FRD score). We use a $100 \times 100$ 2D histogram on the BEV plane to calculate MMD and JSD metrics. The Frechet Range Distance (FRD) score [95] is a metric used to evaluate the quality of samples acquired by a generative model, inspired by the FID score for images [24]. To compute the FRD score, we use RangeNet++ [45], an encoder-decoder-based network for segmentation, which is pre-trained on KITTI-360. + +Baselines. We evaluated our approach for LiDAR point cloud generation against several competitive methods, including LiDAR VAE [5], LiDAR GAN [5], Projected GAN [58], LiDARGen [95] and UltraLiDAR [79]. + +# 5.2 Unconditional Generation + +
MethodYears\( MMD_{BEV} \)↓FRD ↓\( JSD_{BEV} \)↓
LiDAR GAN [5]IROS 2019\( 3.06 \times 10^{-3} \)3003.8-
LiDAR VAE [5]IROS 2019\( 1.00 \times 10^{-3} \)2261.50.161
Projected GAN [58]NeurIPS 2021\( 3.47 \times 10^{-4} \)2117.20.085
LiDARGen [95]ECCV 2022\( 3.87 \times 10^{-4} \)2040.10.067
UltraLiDAR [79]CVPR 2023\( 1.96 \times 10^{-4} \)-0.071
Ours\( 3.07 \times 10^{-5} \)1074.90.045
+ +Table 2: Unconditional generation results on KITTI-360 [38]. + +Quantitative Results on KITTI-360. Table 2 displays the quantitative results of the proposed RangeLDM and competing algorithms. The results clearly + +illustrate that our method significantly outperforms the baselines across all metrics. Notably, our results on the MMD metric are an order of magnitude lower than those of other methods. This underscores the remarkable expressive capabilities of our model in the context of LiDAR point cloud generation. + +Qualitative Results on KITTI-360. Figure 4 displays a set of randomly generated samples from competing algorithms, alongside real point cloud samples extracted from the KITTI-360 dataset for comparison. We observe that the Projected GAN [58] exhibits noticeable artifacts in the distant range. LiDARGen [95] effectively captures the general layout, but produces rather noisy outputs and lacks the same degree of straight walls as actual samples. UltraLiDAR [79] generates structured and reasonable scenes, yet it only creates voxelized point clouds (as magnified in the first row) without intensity features. In contrast, our model consistently surpasses the baseline models and yields results closely resembling real-world data. For example, we can generate cars (emphasized by red boxes in the first row of Figure 4) and road bollards (red boxes in the second row of Figure 4) on the road. Additionally, our model generates various scenes such as straightway (the first two rows of Figure 4) and crossroads (the last row of Figure 4). The supplementary materials provide more visualization results. + +Human study on KITTI-360. To assess the perceptual quality, we conducted an A/B test involving a group of 14 researchers with LiDAR expertise. Following the same evaluation system as [95] and [79], we present pairs of randomly selected images + +
MethodPercent prefer ours
Ours vs. VAE [5]98.8%
Ours vs. GAN [5]98.0%
Ours vs. ProjectedGAN [58]93.2%
Ours vs. LiDARGen [95]90.2%
Ours vs. UltraLiDAR [79]86.5%
+ +from two point clouds and ask participants to determine which one appeared more realistic. The results, as displayed in Table 3, unequivocally demonstrate the superior visual quality of generation results by our model. In most of the cases, testers favored our results over the baselines. + +Evaluation on the NuScenes Dataset. The results listed in Table 4 demonstrate that our model outperforms LiDAR VAE and LiDARGen significantly over nuScenes. In particular, our results on the MMD metric are an order of magnitude lower than those of other methods, while we reduce to about $1/3$ of the + +results of competitive methods in terms of the JSD metric. Also, Figure 6 illustrates the superiority of our model over LiDAR VAE and LiDARGen in terms of qualitative comparison. + +Table 3: Human study on KITTI-360. + +
Method\( MMD_{BEV} \)↓\( JSD_{BEV} \)↓
LiDAR VAE [5]\( 1.1 \times 10^{-3} \)-
\( LiDARGen^† \)[95]\( 1.9 \times 10^{-3} \)0.160
Ours\( 1.9 \times 10^{-4} \)0.054
+ +Table 4: Unconditional generation results on nuScenes dataset [6]. †: Reproduced by us. + +Model Size. The number of parameters in LiDARGen [95] and UltraLiDAR [79] are 29.7M and 40.3M, respectively. To ensure fair competition, we limited the capacity of our model to be similar to other approaches. Our model comprises 41.4M parameters, consisting of 12.7M parameters for VAE and 28.7M parameters for LDM. As illustrated in Table 2, our model outperforms the baselines by a large margin under similar model sizes. + +![](images/f95bf65998d7447a178c9b8f34403bdb06caba45f1541d1d6839153fb0bd9016.jpg) +Fig. 6: Qualitative results comparing against LiDAR VAE and LiDARGen for unconditional LiDAR generation on nuScenes. Real point clouds are only for reference. + +
Hough Range-Guided +Voting DiscriminatorCircular +ConvolutionDirection +ConditionedMMDBEV ↓FRD ↓JSDBEV ↓
(a)××××3.95 × 10-41536.70.067
(b)✓×××6.57 × 10-51229.30.056
(c)✓✓××4.72 × 10-51103.10.051
(d)✓✓✓×3.90 × 10-41797.20.078
(e)✓✓✓✓3.07 × 10-51074.90.045
+ +Generation Efficiency. The proposed dimensionality reduction offers RangeLDM a considerable advantage in terms of generation speed. As indicated in Table 5, the proposed method significantly outperforms LiDARGen [95] and UltraLi-DAR [79] in terms of generation speed. In particular, our model is 200 times faster than LiDARGen and 30 times faster than UltraLi-DAR. + +Table 6: Main ablation study. + +
MethodThroughput ↑ (samples/s)
LiDARGen [95]0.02
UltraLiDAR [79]0.16
Ours4.86
+ +# 5.3 Conditional Generation + +To evaluate the conditional generation performance of RangeLDM, we conducted experiments on the KITTI-360 dataset over two tasks: LiDAR point cloud upsampling and inpainting. + +LiDAR Point Cloud Upsampling. We obtain sparse input point clouds by selecting a subset of 16 beams from the raw 64-beam sensors, in line with LiDARGen. We compared our approach against PUNet [88], DeepRS [9], Grad-PU [23], bicubic interpolation, Nearest Neighbor (NN) interpolation and LiDARGen. + +Quantitatively, in addition to measuring Mean Absolute Error (MAE) in the range view, we also employed RangeNet++ semantic segmentation to evaluate the quality of upsampled results. As shown in Table 7, the proposed method outperforms competing methods in all metrics, including + +Table 5: Inference speed on a single RTX 3090 GPU. + +
MethodYearsMAE ↓Accuracy ↑ IoU ↑
PUNet [88]20186.88--
DeepRS [9]20223.96--
Grad-PU [23]20235.09--
Bicubic-2.600.2650.166
NN-2.180.5460.394
LiDARGen [95]20221.230.6080.449
Ours0.890.7220.566
+ +Table 7: LiDAR upsampling. + +MAE, per-point segmentation accuracy, and segmentation Intersection over Union (IoU). + +![](images/cc21bf27bd40cbe214ff1e23d79ef5777a37a5bfaaeaef5548d3b3ab4716a33d.jpg) +Fig. 7: Inpainting results. Left: ground truth point clouds. Middle and right (red): input point clouds. Middle and right (blue): recovered point clouds. + +![](images/59aec08343dbaccee62d95b22d4da83d42eda9f7fa840c17a9bb4fc184e794d2.jpg) +Fig. 8: VAE reconstruction results. The proposed VAE with a range-guided discriminator reconstructs point clouds with less noise and more precise object structures. + +Qualitatively, the visualization comparison with several competitive methods presented in Figure 5 indicates that the proposed upsampling method exhibits results highly similar to the ground truth data, showing clear object details and consistent LiDAR scan lines. We provide additional visualization results in the supplementary materials. + +LiDAR Point Cloud Inpainting. To mimic the case of missing data in LiDAR point clouds, we mask point clouds in front of the vehicle within a $22.5^{\circ}$ range and perform inpainting using LiDARGen and our method. Our results achieve much less MAE (0.190) between the reconstructed results and ground truth compared to that of LiDARGen (0.367). As shown in Figure 7, LiDARGen generates roads but fails to recover the masked car (the first row) and the wall (the second row), while we are able to reconstruct the details. More results are presented in the supplementary materials. + +# 5.4 Ablation Study + +All ablation studies are conducted on the KITTI-360 dataset. + +Main Ablation. As demonstrated in Table 6, we investigate the contribution of each component. Starting from the backbone model (a) without any component, we gradually add each component for evaluation. By applying Hough Voting to the training, the performance of variant (b) improves significantly, demonstrating the huge impact of the correct range-view data distribution. With the addition of the range-guided discriminator component, model (c) achieves better performance, thus clarifying the effectiveness of the geometry-sensitive module. Compared to model (c), the performance of model (d) decreases because circular convolution generates LiDAR scenes with arbitrary directions due to rotation invariance. However, when combined with direction-conditioned generation that guides the model to generate point clouds in the correct direction, the overall performance improves, as demonstrated by model (e). We refer the readers to the supplementary materials for qualitative results. + +
ModelVAEModel Size (M)\( MMD_{BEV} \)↓FRD ↓\( JSD_{BEV} \)↓Throughput ↑(samples/s)
Ours-DMS×40.3\( 1.03 × 10^{-4} \)1392.20.0621.52
Ours-DM×114.7\( 4.14 × 10^{-5} \)899.00.0400.64
Ours✓41.4\( 3.07 × 10^{-5} \)1074.90.0454.86
+ +Latent diffusion. We explore the contribution of latent diffusion to the generation quality and generation speed of the model. As shown in Table 8, we constructed two variants of different sizes that directly generate range images with diffusion models, denoted as "Ours-DMS" (with a smaller size) and "Ours-DM" (with a larger size), for comparison. It turns out that a much larger model size is required for "Ours-DM" to achieve competitive performance with "Ours", due to the multitude of high-frequency details contained in the range image. Additionally, the superior throughput of "Ours" demonstrates the efficiency of latent diffusion. + +VAE Architecture. We investigate the contribution of the proposed range-guided discriminator by comparing our VAE with two baselines: 1) VAE without dis + +Table 8: Ablation of latent diffusion. Model without latent diffusion requires more parameters to achieve similar performance with the latent diffusion model. We generated 1000 samples for calculating the throughput. + +
DiscriminatorPSNRrange↑ MAE↓ FRD↓ CD
No26.770.0195532.90.0808
Vanilla26.700.0189496.70.0726
Range-Guided27.190.0186483.60.0676
+ +Table 9: Ablation on the VAE architecture. + +criminalitor; and 2) VAE with vanilla discriminator (i.e., using the original 2D convolution in the discriminator). We evaluate the reconstruction performance of VAEs with 2D and 3D metrics, including PSNR, MAE, FRD and Chamfer Distance (CD) [16]. As listed in Table 9, the VAE with the proposed range-guided discriminator outperforms the baselines in both 2D and 3D reconstruction metrics. This gives credits to the ability of perceiving local 3D structures by the range-guided discriminator. The qualitative results presented in Figure 8 also indicate that the proposed range-guided discriminator ensures the point clouds are reconstructed with less noise and more precise object structures. + +# 6 Conclusions + +We propose a novel RangeLDM model to generate realistic range-view LiDAR point clouds at a fast speed. We ensure the quality of projection from point clouds to range images with correct distribution via Hough Voting. Then we compress range images to latent features with a VAE, and train a diffusion model in the lower-dimensional latent space. Additionally, we enhance the range-image reconstruction quality of the VAE with a range-guided discriminator. Experiments conducted on KITTI-360 and nuScenes datasets demonstrate the superior generation quality and sampling efficiency of our method. In future, we will explore generating labeled point clouds and creating corner-case data, such as in car accidents and extreme weather conditions, for potential applications in robust self-driving. + +# References + +1. Achlioptas, P., Diamanti, O., Mitlagkas, I., Guibas, L.: Learning representations and generative models for 3D point clouds. In: International Conference on Machine Learning. pp. 40-49. PMLR (2018) +2. Ando, A., Gidaris, S., Bursuc, A., Puy, G., Boulch, A., Marlet, R.: RangeViT: Towards vision transformers for 3D semantic segmentation in autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5240-5250 (2023) +3. Bewley, A., Sun, P., Mensink, T., Anguelov, D., Sminchisescu, C.: Range conditioned dilated convolutions for scale invariant 3D object detection. In: Conference on Robot Learning (2020) +4. Box, G.E., Jenkins, G.M., Reinsel, G.C., Ljung, G.M.: Time series analysis: forecasting and control. John Wiley & Sons (2015) +5. Caccia, L., Van Hoof, H., Courville, A., Pineau, J.: Deep generative modeling of LiDAR data. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 5034-5040. IEEE (2019) +6. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: NuScenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11621-11631 (2020) +7. Cai, R., Yang, G., Averbuch-Elor, H., Hao, Z., Belongie, S., Snavely, N., Hariharan, B.: Learning gradient fields for shape generation. In: European Conference on Computer Vision. pp. 364-381 (2020) +8. Chai, Y., Sun, P., Ngiam, J., Wang, W., Caine, B., Vasudevan, V., Zhang, X., Anguelov, D.: To the point: Efficient 3D object detection in the range image with graph convolution kernels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16000-16009 (2021) +9. Chen, H., Luo, S., Hu, W., et al.: Deep point set resampling via gradient fields. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(3), 2913-2930 (2022) +0. Chen, X., Vizzo, I., Läbe, T., Behley, J., Stachniss, C.: Range image-based LiDAR localization for autonomous vehicles. In: IEEE International Conference on Robotics and Automation (ICRA). pp. 5802-5808. IEEE (2021) +1. Cheng, H., Han, X., Xiao, G.: CeNet: Toward concise and efficient LiDAR semantic segmentation for autonomous driving. In: IEEE International Conference on Multimedia and Expo (ICME). pp. 01-06. IEEE (2022) +2. Cortinhal, T., Tzelepis, G., Erdal Aksoy, E.: SalsaNext: Fast, uncertainty-aware semantic segmentation of LiDAR point clouds. In: Advances in Visual Computing. pp. 207-222. Springer International Publishing, Cham (2020) +3. Deng, J., Shi, S., Li, P., Zhou, W., Zhang, Y., Li, H.: Voxel R-CNN: Towards high performance voxel-based 3D object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1201-1209 (2021) +4. Engelcke, M., Rao, D., Wang, D.Z., Tong, C.H., Posner, I.: Vote3Deep: Fast object detection in 3D point clouds using efficient convolutional neural networks. In: 2017 IEEE International Conference on Robotics and Automation (ICRA). pp. 1355-1361. IEEE (2017) +5. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12873-12883 (2021) + +16. Fan, H., Su, H., Guibas, L.J.: A point set generation network for 3D object reconstruction from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 605-613 (2017) +17. Fan, L., Pang, Z., Zhang, T., Wang, Y.X., Zhao, H., Wang, F., Wang, N., Zhang, Z.: Embracing single stride 3D object detector with sparse transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8458-8468 (2022) +18. Fan, L., Xiong, X., Wang, F., Wang, N., Zhang, Z.: RangeDet: In defense of range view for LiDAR-based 3D object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2918-2927 (2021) +19. Gao, L., Wu, T., Yuan, Y.J., Lin, M.X., Lai, Y.K., Zhang, H.: TM-NET: Deep generative networks for textured meshes. ACM Transactions on Graphics (TOG) 40(6), 1-15 (2021) +20. Gao, L., Yang, J., Wu, T., Yuan, Y.J., Fu, H., Lai, Y.K., Zhang, H.: SDM-NET: Deep generative network for structured deformable mesh. ACM Transactions on Graphics (TOG) 38(6), 1-15 (2019) +21. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020) +22. He, C., Li, R., Li, S., Zhang, L.: Voxel set transformer: A set-to-set approach to 3D object detection from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8417-8427 (2022) +23. He, Y., Tang, D., Zhang, Y., Xue, X., Fu, Y.: Grad-PU: Arbitrary-scale point cloud upsampling via gradient descent with learned distance functions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5354-5363 (2023) +24. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: GANs trained by a two time-scale update rule converge to a local nash equilibrium. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017) +25. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33, 6840-6851 (2020) +26. Hu, J.S., Kuai, T., Waslander, S.L.: Point density-aware voxels for LiDAR 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8469-8478 (2022) +27. Hu, P., Ziglar, J., Held, D., Ramanan, D.: What you see is what you get: Exploiting visibility for 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11001-11009 (2020) +28. Hui, K.H., Li, R., Hu, J., Fu, C.W.: Neural wavelet-domain diffusion for 3D shape generation. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022) +29. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1125-1134 (2017) +30. Kim, H., Lee, H., Kang, W.H., Lee, J.Y., Kim, N.S.: SoftFlow: Probabilistic framework for normalizing flow on manifolds. Advances in Neural Information Processing Systems 33, 16388-16397 (2020) +31. Kim, J., Yoo, J., Lee, J., Hong, S.: SetVAE: Learning hierarchical composition for generative modeling of set-structured data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15059–15068 (2021) + +32. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013) +33. Kochanov, D., Nejadasl, F.K., Booij, O.: KprNet: Improving projection-based LiDAR semantic segmentation. arXiv preprint arXiv:2007.12668 (2020) +34. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: PointPillars: Fast encoders for object detection from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12697-12705 (2019) +35. Li, B.: 3D fully convolutional network for vehicle detection in point cloud. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1513-1518. IEEE (2017) +36. Li, B., Zhang, T., Xia, T.: Vehicle detection from 3D LiDAR using fully convolutional network. arXiv preprint arXiv:1608.07916 (2016) +37. Li, C., Ren, Y., Liu, B.: Pcgen: Point cloud generator for lidar simulation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 11676-11682. IEEE (2023) +38. Liao, Y., Xie, J., Geiger, A.: KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2D and 3D. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(3), 3292-3310 (2022) +39. Lim, J.H., Ye, J.C.: Geometric gan. arXiv preprint arXiv:1705.02894 (2017) +40. Litany, O., Bronstein, A., Bronstein, M., Makadia, A.: Deformable shape completion with graph convolutional autoencoders. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1886-1895 (2018) +41. Luo, S., Hu, W.: Diffusion probabilistic models for 3D point cloud generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2837-2845 (2021) +42. Malavazi, F.B., Guyonneau, R., Fasquel, J.B., Lagrange, S., Mercier, F.: LiDAR-only based navigation algorithm for an autonomous agricultural robot. Computers and Electronics in Agriculture 154, 71-79 (2018) +43. Manivasagam, S., Wang, S., Wong, K., Zeng, W., Sazanovich, M., Tan, S., Yang, B., Ma, W.C., Urtasun, R.: LiDARsim: Realistic LiDAR simulation by leveraging the real world. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11167-11176 (2020) +44. Meyer, G.P., Laddha, A., Kee, E., Vallespi-Gonzalez, C., Wellington, C.K.: Laser-Net: An efficient probabilistic 3D object detector for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12677-12686 (2019) +45. Milioto, A., Vizzo, I., Behley, J., Stachniss, C.: RangeNet++: Fast and accurate LiDAR semantic segmentation. In: IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 4213-4220. IEEE (2019) +46. Mittal, P., Cheng, Y.C., Singh, M., Tulsiani, S.: AutoSDF: Shape priors for 3D completion, reconstruction and generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 306-315 (2022) +47. Mo, K., Guerrero, P., Yi, L., Su, H., Wonka, P., Mitra, N., Guibas, L.J.: StructureNet: Hierarchical graph networks for 3D shape generation. arXiv preprint arXiv:1908.00575 (2019) +48. Nichol, A., Jun, H., Dhariwal, P., Mishkin, P., Chen, M.: Point-E: A system for generating 3D point clouds from complex prompts. arXiv preprint arXiv:2212.08751 (2022) +49. Pan, X., Xia, Z., Song, S., Li, L.E., Huang, G.: 3D object detection with pointformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7463-7472 (2021) + +50. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: PointNet: Deep learning on point sets for 3D classification and segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 652-660 (2017) +51. Ran, H., Guizilini, V., Wang, Y.: Towards realistic scene generation with lidar diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024) +52. Resop, J.P., Lehmann, L., Hession, W.C.: Drone laser scanning for modeling river-scape topography and vegetation: Comparison with traditional aerial LiDAR. Drones 3(2), 35 (2019) +53. Rezende, D., Mohamed, S.: Variational inference with normalizing flows. In: International Conference on Machine Learning. pp. 1530-1538. PMLR (2015) +54. Rezende, D.J., Mohamed, S., Wierstra, D.: Stochastic backpropagation and approximate inference in deep generative models. In: International Conference on Machine Learning. pp. 1278-1286. PMLR (2014) +55. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022) +56. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention. pp. 234-241. Springer (2015) +57. Sallab, A.E., Sobh, I., Zahran, M., Essam, N.: LiDAR sensor modeling and data augmentation with GANs for autonomous driving. arXiv preprint arXiv:1905.07290 (2019) +58. Sauer, A., Chitta, K., Müller, J., Geiger, A.: Projected GANs converge faster. Advances in Neural Information Processing Systems 34, 17480-17492 (2021) +59. Schubert, S., Neubert, P., Pöschmann, J., Protzel, P.: Circular convolutional neural networks for panoramic images and laser data. In: IEEE Intelligent Vehicles Symposium (IV). pp. 653-660. IEEE (2019) +60. Shi, G., Li, R., Ma, C.: PillarNet: Real-time and high-performance pillar-based 3D object detection. In: European Conference on Computer Vision. pp. 35-52. Springer (2022) +61. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: PV-RCNN: Point-voxel feature set abstraction for 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10529-10538 (2020) +62. Shi, S., Wang, X., Li, H.: PointRCNN: 3D object proposal generation and detection from point cloud. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 770-779 (2019) +63. Shi, S., Wang, Z., Shi, J., Wang, X., Li, H.: From points to parts: 3D object detection from point cloud with part-aware and part-aggregation network. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(8), 2647-2664 (2020) +64. Shi, W., Rajkumar, R.: Point-GNN: Graph neural network for 3D object detection in a point cloud. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1711-1719 (2020) +65. Shu, D.W., Park, S.W., Kwon, J.: 3D point cloud generative adversarial network based on tree structured graph convolutions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3859-3868 (2019) +66. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020) + +67. Sun, P., Tan, M., Wang, W., Liu, C., Xia, F., Leng, Z., Anguelov, D.: SWFormer: Sparse window transformer for 3D object detection in point clouds. In: European Conference on Computer Vision. pp. 426-442. Springer (2022) +68. Sun, P., Wang, W., Chai, Y., Elsayed, G., Bewley, A., Zhang, X., Sminchisescu, C., Anguelov, D.: RSN: Range sparse net for efficient, accurate LiDAR 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5725-5734 (2021) +69. Sun, Y., Wang, Y., Liu, Z., Siegel, J., Sarma, S.: PointGrow: Autoregressively learned point cloud generation with self-attention. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 61-70 (2020) +70. Tan, Q., Gao, L., Lai, Y.K., Xia, S.: Variational autoencoders for deforming 3D mesh models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5841-5850 (2018) +71. Tian, Z., Chu, X., Wang, X., Wei, X., Shen, C.: Fully convolutional one-stage 3D object detection on LiDAR range images. Advances in Neural Information Processing Systems 35, 34899-34911 (2022) +72. Vahdat, A., Kreis, K., Kautz, J.: Score-based generative modeling in latent space. Advances in Neural Information Processing Systems 34, 11287-11302 (2021) +73. Valsesia, D., Fracastoro, G., Magli, E.: Learning localized generative models for 3D point clouds via graph convolution. In: International Conference on Learning Representations (2018) +74. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in Neural Information Processing Systems 30 (2017) +75. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in Neural Information Processing Systems 30 (2017) +76. Wang, H., Shi, S., Yang, Z., Fang, R., Qian, Q., Li, H., Schiele, B., Wang, L.: RBGNet: Ray-based grouping for 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1110-1119 (2022) +77. Wang, Y., Fathi, A., Kundu, A., Ross, D.A., Pantofaru, C., Funkhouser, T., Solomon, J.: Pillar-based object detection for autonomous driving. In: European Conference on Computer Vision. pp. 18-34. Springer (2020) +78. Weiss, U., Biber, P.: Plant detection and mapping for agricultural robots using a 3D LiDAR sensor. Robotics and Autonomous Systems 59(5), 265-273 (2011) +79. Xiong, Y., Ma, W.C., Wang, J., Urtasun, R.: Learning compact representations for LiDAR completion and generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1074-1083 (2023) +80. Xu, C., Wu, B., Wang, Z., Zhan, W., Vajda, P., Keutzer, K., Tomizuka, M.: SqueezeSegV3: Spatially-adaptive convolution for efficient point-cloud segmentation. In: European Conference on Computer Vision. pp. 1-19. Springer (2020) +81. Yan Yan, Y.M., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018) +82. Yang, B., Luo, W., Urtasun, R.: Pixor: Real-time 3D object detection from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7652-7660 (2018) +83. Yang, G., Huang, X., Hao, Z., Liu, M.Y., Belongie, S., Hariharan, B.: PointFlow: 3D point cloud generation with continuous normalizing flows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4541-4550 (2019) + +84. Yang, Z., Sun, Y., Liu, S., Jia, J.: 3DSSD: Point-based 3D single stage object detector. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11040-11048 (2020) +85. Yang, Z., Sun, Y., Liu, S., Shen, X., Jia, J.: STD: Sparse-to-dense 3D object detector for point cloud. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1951-1960 (2019) +86. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3D object detection and tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11784-11793 (2021) +87. Yu, J., Li, X., Koh, J.Y., Zhang, H., Pang, R., Qin, J., Ku, A., Xu, Y., Baldridge, J., Wu, Y.: Vector-quantized image modeling with improved VQGAN. arXiv preprint arXiv:2110.04627 (2021) +88. Yu, L., Li, X., Fu, C.W., Cohen-Or, D., Heng, P.A.: Pu-Net: Point cloud upsampling network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2790-2799 (2018) +89. Zeng, X., Vahdat, A., Williams, F., Gojcic, Z., Litany, O., Fidler, S., Kreis, K.: LION: Latent point diffusion models for 3D shape generation. Advances in Neural Information Processing Systems 35, 10021-10039 (2022) +90. Zhang, J., Singh, S.: LOAM: LiDAR odometry and mapping in real-time. In: Robotics: Science and Systems. vol. 2, pp. 1-9. Berkeley, CA (2014) +91. Zhang, Y., Hu, Q., Xu, G., Ma, Y., Wan, J., Guo, Y.: Not all points are equal: Learning highly efficient point-based detectors for 3D LiDAR point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18953-18962 (2022) +92. Zhao, Y., Bai, L., Huang, X.: FidNet: LiDAR point cloud semantic segmentation with fully interpolation decoding. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 4453-4458. IEEE (2021) +93. Zhou, L., Du, Y., Wu, J.: 3D shape generation and completion through point-voxel diffusion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5826-5835 (2021) +94. Zhou, Y., Tuzel, O.: VoxelNet: End-to-end learning for point cloud based 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4490-4499 (2018) +95. Zyrianov, V., Zhu, X., Wang, S.: Learning to generate realistic LiDAR point clouds. In: European Conference on Computer Vision. pp. 17-35. Springer (2022) \ No newline at end of file diff --git a/2024/RangeLDM_ Fast Realistic LiDAR Point Cloud Generation/images.zip b/2024/RangeLDM_ Fast Realistic LiDAR Point Cloud Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c4521045ba8bce3b2ff9ec962cedea5766a650f3 --- /dev/null +++ b/2024/RangeLDM_ Fast Realistic LiDAR Point Cloud Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:64d3842b8e7a37dce1e51b55c0cb196a59b2eb462c888c278c5b90fe27510d00 +size 603642 diff --git a/2024/RangeLDM_ Fast Realistic LiDAR Point Cloud Generation/layout.json b/2024/RangeLDM_ Fast Realistic LiDAR Point Cloud Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e00c12e9630297af1a8a26e1f3143f286a9b6a09 --- /dev/null +++ b/2024/RangeLDM_ Fast Realistic LiDAR Point Cloud Generation/layout.json @@ -0,0 +1,11752 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 111, + 473, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 111, + 473, + 146 + ], + "spans": [ + { + "bbox": [ + 143, + 111, + 473, + 146 + ], + "type": "text", + "content": "RangeLDM: Fast Realistic LiDAR Point Cloud Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 197, + 169, + 412, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 169, + 412, + 182 + ], + "spans": [ + { + "bbox": [ + 197, + 169, + 412, + 182 + ], + "type": "text", + "content": "Qianjiang Hu, Zhimin Zhang, and Wei Hu" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 144, + 191, + 470, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 191, + 470, + 214 + ], + "spans": [ + { + "bbox": [ + 144, + 191, + 470, + 214 + ], + "type": "text", + "content": "Wangxuan Institute of Computer Technology, Peking University, Beijing, China hqjpu@pku.edu.cn zm_zhang@stu.pku.edu.cn forhuwei@pku.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 242, + 211, + 334 + ], + "blocks": [ + { + "bbox": [ + 133, + 242, + 211, + 334 + ], + "lines": [ + { + "bbox": [ + 133, + 242, + 211, + 334 + ], + "spans": [ + { + "bbox": [ + 133, + 242, + 211, + 334 + ], + "type": "image", + "image_path": "bec2e48447c9c338c2bbaeaff95f297df8162460662d1d078c38a0ffad5f51d5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 335, + 206, + 342 + ], + "lines": [ + { + "bbox": [ + 137, + 335, + 206, + 342 + ], + "spans": [ + { + "bbox": [ + 137, + 335, + 206, + 342 + ], + "type": "text", + "content": "(a) Unconditional Generation" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 215, + 242, + 370, + 334 + ], + "blocks": [ + { + "bbox": [ + 215, + 242, + 370, + 334 + ], + "lines": [ + { + "bbox": [ + 215, + 242, + 370, + 334 + ], + "spans": [ + { + "bbox": [ + 215, + 242, + 370, + 334 + ], + "type": "image", + "image_path": "454315a4e9010a1ee50a3bec5d60c4f2c2ccecdffce1caba0a9f0534c53bcad9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 335, + 320, + 342 + ], + "lines": [ + { + "bbox": [ + 255, + 335, + 320, + 342 + ], + "spans": [ + { + "bbox": [ + 255, + 335, + 320, + 342 + ], + "type": "text", + "content": "(b) Conditional Generation" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 372, + 245, + 479, + 334 + ], + "blocks": [ + { + "bbox": [ + 372, + 245, + 479, + 334 + ], + "lines": [ + { + "bbox": [ + 372, + 245, + 479, + 334 + ], + "spans": [ + { + "bbox": [ + 372, + 245, + 479, + 334 + ], + "type": "image", + "image_path": "2555aee352ba07603cd717c1df9adb4d8080dfcf1c5a8735f801122ecd200ee3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 335, + 449, + 342 + ], + "lines": [ + { + "bbox": [ + 400, + 335, + 449, + 342 + ], + "spans": [ + { + "bbox": [ + 400, + 335, + 449, + 342 + ], + "type": "text", + "content": "(c) Quality vs. Speed" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 346, + 482, + 422 + ], + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 422 + ], + "type": "text", + "content": "Fig. 1: (a). Unconditional LiDAR point cloud generation with realistic global structure. (b). Conditional LiDAR point cloud generation, including LiDAR point cloud upsampling and inpainting. (c). Generation quality (Maximum Mean Discrepancy, abbr. MMD) vs. generation speed (samples/s) of competitive LiDAR point cloud generation methods on the KITTI-360 [38] dataset. The proposed method outperforms the state-of-the-art methods LiDARGen [95] and UltraLiDAR [79] in both generation quality and generation speed. All speeds are evaluated on a single RTX 3090 GPU." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 447, + 453, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 447, + 453, + 622 + ], + "spans": [ + { + "bbox": [ + 159, + 447, + 453, + 622 + ], + "type": "text", + "content": "Abstract. Autonomous driving demands high-quality LiDAR data, yet the cost of physical LiDAR sensors presents a significant scaling-up challenge. While recent efforts have explored deep generative models to address this issue, they often consume substantial computational resources with slow generation speeds while suffering from a lack of realism. To address these limitations, we introduce RangeLDM, a novel approach for rapidly generating high-quality range-view LiDAR point clouds via latent diffusion models. We achieve this by correcting range-view data distribution for accurate projection from point clouds to range images via Hough voting, which has a critical impact on generative learning. We then compress the range images into a latent space with a variational autoencoder, and leverage a diffusion model to enhance expressivity. Additionally, we instruct the model to preserve 3D structural fidelity by devising a range-guided discriminator. Experimental results on KITTI-360 and nuScenes datasets demonstrate both the robust expressiveness and fast speed of our LiDAR point cloud generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 160, + 634, + 436, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 634, + 436, + 645 + ], + "spans": [ + { + "bbox": [ + 160, + 634, + 436, + 645 + ], + "type": "text", + "content": "Keywords: Diffusion models " + }, + { + "bbox": [ + 160, + 634, + 436, + 645 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 634, + 436, + 645 + ], + "type": "text", + "content": " Autonomous systems " + }, + { + "bbox": [ + 160, + 634, + 436, + 645 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 634, + 436, + 645 + ], + "type": "text", + "content": " Point clouds" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 653, + 277, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 653, + 277, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 653, + 277, + 665 + ], + "type": "inline_equation", + "content": "\\mathbb{O}" + }, + { + "bbox": [ + 141, + 653, + 277, + 665 + ], + "type": "text", + "content": " Corresponding author: W. Hu." + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 230, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 230, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 230, + 127 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 481, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 481, + 222 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 481, + 222 + ], + "type": "text", + "content": "Autonomous systems have drawn great attention in both the academia and the industry community. Numerous autonomous systems leverage the power of various sensors and deep learning to enhance the perception of the 3D world. LiDAR (Light Detection And Ranging), with its ability to provide precise 3D geometric information about the surroundings, has become a popular sensor choice for autonomous systems, including self-driving cars [61,62,86], surveying drones [52,90] and robots [42,78]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 223, + 481, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 223, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 223, + 481, + 282 + ], + "type": "text", + "content": "However, while LiDAR offers accurate geometric measurements, it comes with a significant limitation: the data collection process is exceedingly expensive and challenging to scale up. This renders physical sensors impractical for scalable and customizable data collection. Additionally, it is difficult to collect data in corner cases such as car accidents and extreme weather conditions in the field." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 283, + 481, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 481, + 485 + ], + "type": "text", + "content": "One approach to mitigate this issue is to employ existing LiDAR simulation toolkits [37, 43] to synthesize more data. Nevertheless, these systems typically demand manual scene creation or rely on multiple prior scans of the real world. Another approach is leveraging deep generative models to generate LiDAR point clouds. This is challenging due to the high degree of unstructuredness, sparsity, and non-uniformity in LiDAR point clouds. Caccia et al. [5] proposed LiDAR GAN and LiDAR VAE to generate LiDAR point clouds with generative adversarial networks (GANs) [21] and variational autoencoders (VAEs) [32], respectively. However, the generation results often suffer from issues such as fuzzy or missing details. LiDARGen [95] introduced a novel score-based model to synthesis LiDAR point clouds, which however sampled slowly and failed to generate high-quality geometric details at a far range. UltraLiDAR [79] proposed to synthesize voxelized point clouds in bird's-eyes-view (BEV) with vector quantized VAE (VQ-VAE) [74], resulting in more realistic point clouds than previous methods. Nevertheless, due to the large size and sparsity of point clouds in BEV, most computing power of UltraLiDAR is devoted to generating empty voxels, which results in a low generation speed." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "text", + "content": "To this end, we propose RangeLDM, a novel approach based on latent diffusion models (LDMs) [55], which improves both the quality and speed of LiDAR point cloud generation significantly. Firstly, we introduce a range image view to represent point clouds. The reasons are twofold: 1) Range images are compact and closely mimic the sampling conditions of LiDAR, making them a suitable representation. A LiDAR point cloud essentially constructs a 2.5D scene from a single viewpoint instead of a full 3D point cloud [27]. Consequently, organizing the point cloud in range view ensures that no information is overlooked. 2) 2D image generation techniques are relatively mature, providing a solid foundation for our work. While LiDARGen [95] also represents point clouds as range images, it suffers from blurriness caused by inaccurate projection. In contrast, we provide insights that the correct range-view data distribution has a critical impact on such range-view-based generative model learning, and accurately project point clouds onto range images using parameters estimated by Hough Voting. Then, to achieve a faster sampling speed [55] and greater expressivity [55,72], we compress" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": "range images to the latent space with a VAE and develop a diffusion model operating on the latent space, leveraging the successful paradigm of LDMs [55,72,89]. Further, in order to enhance the ability of the VAE in reconstructing 3D structures, we introduce a range-guided discriminator, which is supervised from the spherical coordinates and thus geometry-sensitive. This plays a crucial role in guiding the decoder to generate high-quality range images, ultimately preserving the fidelity of the 3D structure. Extensive experimental results show that the proposed RangeLDM achieves the state-of-the-art performance on KITTI-360 [38] and nuScenes [6] datasets in terms of both the generation quality and generation speed, as demonstrated in Figure 1 (c). The LiDAR upsampling and LiDAR inpainting results also demonstrate the potential of the proposed method for conditional generation, as presented in Figure 1 (a) and (b)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 147, + 261, + 411, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 261, + 411, + 272 + ], + "spans": [ + { + "bbox": [ + 147, + 261, + 411, + 272 + ], + "type": "text", + "content": "To summarize, the main contributions of this paper include:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 281, + 480, + 415 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 281, + 479, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 281, + 479, + 316 + ], + "spans": [ + { + "bbox": [ + 138, + 281, + 479, + 316 + ], + "type": "text", + "content": "- We propose a latent diffusion model to capture the distribution of range-view LiDAR point clouds, aiming to generate realistic point cloud scenes at a fast speed." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 318, + 480, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 318, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 138, + 318, + 480, + 353 + ], + "type": "text", + "content": "- We enlighten the significance of the correct range-view data distribution for range-view-based generative models and achieve high-quality range image projection via Hough Voting." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 354, + 480, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 480, + 378 + ], + "type": "text", + "content": "- We exploit a range-guided discriminator to ensure preserving the fidelity of the 3D geometric structure in generated point clouds." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 379, + 480, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 379, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 138, + 379, + 480, + 415 + ], + "type": "text", + "content": "- Experimental results on the KITTI-360 and nuScenes datasets highlight that our approach outperforms state-of-the-art methods in terms of both visual quality and especially generation speed." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 434, + 237, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 237, + 448 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 237, + 448 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "text", + "content": "LiDAR Representation. LiDAR point clouds can be represented in various forms, encompassing raw point clouds, voxels, range views, and multi-view fusion. Point-based models [49, 62, 64, 76, 84, 85, 91] directly encode 3D objects from raw points using the PointNet [50] encoder and subsequently perform detection or segmentation based on point features. These methods wholly preserve the irregularity and locality of a point cloud but have relatively higher latency. Voxel-based methods [13, 14, 17, 34, 35, 60, 63, 67, 77, 81, 82, 86, 94] voxelize point clouds for convolutional neural networks to efficiently capture features. They are computationally effective but the desertion of fine-grained patterns degrades further refinement. Given that the range view is compact and compatible with the LiDAR sensor's sampling process, range-image-based methods [2, 8, 10-12, 18, 33, 36, 44, 45, 68, 71, 80, 92] have been proposed to directly process range images for LiDAR perception. Multi-view fusion methods [22, 26, 61] amalgamate multiple representations, which yield better results at the expense of processing speed. In this paper, we adopt the range-view representation, which is congruent with the LiDAR sensor's sampling process and can be efficiently encoded by image generation models." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "type": "text", + "content": "RangeLDM: Fast Realistic LiDAR Point Cloud Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 281 + ], + "type": "text", + "content": "Generative Models for Point Clouds. Given observed samples of interest, generative models aim to learn the underlying distribution of the data and generate new samples. The first 3D generative models [1,65,73] for point clouds are based on GANs [21]. They operate by generating point clouds and discriminating them from real samples in an adversarial manner. The second category of methods [19,20,31,40,46,47,70] employ VAEs [32] to explore the probabilistic latent space of 3D shapes. Auto-regressive models [4] have also been introduced to generate point clouds [69] from scratch or semantic contexts. PointFlow [83] and SoftFlow [30] leverage normalizing flows [53] to capture the likelihood of shapes and points. Notably, the recent success of denoising diffusion models (DDMs) [25] for image synthesis [55] has also been extended to the domain of 3D point cloud generation [7,28,41,48,89,93]. However, these previous works often concentrate on the object level and are not well-suited for handling large scenes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 283, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 416 + ], + "type": "text", + "content": "In this paper, our focus lies in the generation of LiDAR point clouds. While [5] and [57] employ VAE or GANs for LiDAR point cloud generation, the realism achieved in their results is relatively limited. UltraLiDAR [79], on the other hand, utilizes VQ-VAE to generate voxelized LiDAR point clouds, but at the expense of introducing quantified losses and slow generation speed. LiDARGen [95] proposed a novel score-matching energy-based model to generate higher-quality Li-DAR point clouds, but it samples very slowly and has degraded geometric details at a far range. A contemporary work LiDM [51] also utilized LDMs to generate LiDAR point clouds. However, while they focused on generating LiDAR point clouds conditional on multimodal data, we focus on improving the quality of point cloud generation directly." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 432, + 411, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 432, + 411, + 448 + ], + "spans": [ + { + "bbox": [ + 132, + 432, + 411, + 448 + ], + "type": "text", + "content": "3 Background on Denoising Diffusion Models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 456, + 482, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 528 + ], + "type": "text", + "content": "DDMs [25] are latent variable models that employ a pre-defined posterior distribution, known as the forward diffusion process, and are trained with a denoising objective. More specifically, given samples " + }, + { + "bbox": [ + 130, + 456, + 482, + 528 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0\\sim q(\\mathbf{x}_0)" + }, + { + "bbox": [ + 130, + 456, + 482, + 528 + ], + "type": "text", + "content": " from a data distribution, DDMs follow the form " + }, + { + "bbox": [ + 130, + 456, + 482, + 528 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_0)\\coloneqq \\int p_{\\theta}(\\mathbf{x}_{0:T})d\\mathbf{x}_{1:T}" + }, + { + "bbox": [ + 130, + 456, + 482, + 528 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 456, + 482, + 528 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 456, + 482, + 528 + ], + "type": "text", + "content": " denotes the number of steps, " + }, + { + "bbox": [ + 130, + 456, + 482, + 528 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1,\\dots,\\mathbf{x}_T" + }, + { + "bbox": [ + 130, + 456, + 482, + 528 + ], + "type": "text", + "content": " are latent variables that gradually add noise to the data " + }, + { + "bbox": [ + 130, + 456, + 482, + 528 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 130, + 456, + 482, + 528 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 456, + 482, + 528 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 456, + 482, + 528 + ], + "type": "text", + "content": " denotes the parameter set of the DDM decoder." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 528, + 482, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 482, + 553 + ], + "type": "text", + "content": "DDMs are trained by minimizing the evidence lower bound (ELBO) of the data " + }, + { + "bbox": [ + 130, + 528, + 482, + 553 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 130, + 528, + 482, + 553 + ], + "type": "text", + "content": " under " + }, + { + "bbox": [ + 130, + 528, + 482, + 553 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_{0:T})" + }, + { + "bbox": [ + 130, + 528, + 482, + 553 + ], + "type": "text", + "content": ". This objective can be simplified to [25]:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 219, + 559, + 481, + 580 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 559, + 481, + 580 + ], + "spans": [ + { + "bbox": [ + 219, + 559, + 481, + 580 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {\\mathbf {x}, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {x} _ {t} + \\epsilon , t) \\| _ {2} ^ {2} \\right], \\tag {1}", + "image_path": "65e0bb7418bcc401e3d1eab15c355afc7a125b2c65875a21e8a9bd93b67112b3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 588, + 316, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 588, + 316, + 601 + ], + "spans": [ + { + "bbox": [ + 131, + 588, + 316, + 601 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 131, + 588, + 316, + 601 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 588, + 316, + 601 + ], + "type": "text", + "content": " uniformly sampled from " + }, + { + "bbox": [ + 131, + 588, + 316, + 601 + ], + "type": "inline_equation", + "content": "\\{1,\\dots,T\\}" + }, + { + "bbox": [ + 131, + 588, + 316, + 601 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 617, + 289, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 617, + 289, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 289, + 632 + ], + "type": "text", + "content": "4 The Proposed Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "As shown in Fig. 2, we first project point clouds onto high-quality range images (as detailed in Section 4.1), considering the compact nature of the range view," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 482, + 208 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 482, + 208 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 482, + 208 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 482, + 208 + ], + "type": "image", + "image_path": "81e720c0c8cc65c583bd34340169d7c339a66c8720c285c53ecfb5c22fc94402.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 209, + 482, + 308 + ], + "lines": [ + { + "bbox": [ + 130, + 209, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 209, + 482, + 308 + ], + "type": "text", + "content": "Fig. 2: The framework of the proposed RangeLDM. Firstly, we project point clouds onto high-quality range images via Hough Voting (Section 4.1). Subsequently, we train a VAE to compress the range images into low-dimensional latent features " + }, + { + "bbox": [ + 130, + 209, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 130, + 209, + 482, + 308 + ], + "type": "text", + "content": ", which encodes the range images with the encoder " + }, + { + "bbox": [ + 130, + 209, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\zeta}" + }, + { + "bbox": [ + 130, + 209, + 482, + 308 + ], + "type": "text", + "content": " and reconstructs range images from latent features with the decoder " + }, + { + "bbox": [ + 130, + 209, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\eta}" + }, + { + "bbox": [ + 130, + 209, + 482, + 308 + ], + "type": "text", + "content": " (Section 4.2). Here, a range-guided discriminator " + }, + { + "bbox": [ + 130, + 209, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\tau}" + }, + { + "bbox": [ + 130, + 209, + 482, + 308 + ], + "type": "text", + "content": " is introduced to guide the decoder in the reconstruction of 3D structures. We finally train a latent diffusion model to capture the distribution of the latent features (Section 4.2). With optional conditional inputs, the proposed method is applicable to tasks such as point cloud upsampling and inpainting (Section 4.3)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 315, + 482, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 410 + ], + "type": "text", + "content": "which aligns seamlessly with the sampling process of LiDAR sensors. Our model training process is then focused on these projected 2D range images. We then compress range images into a low-dimensional latent space through a VAE (as detailed in Section 4.2). Subsequently, we proceed to train DDMs within the reduced-dimensional latent space (as discussed in Section 4.2). We discuss applications of the proposed model in Section 4.3, which involves unconditional generation and conditional generation. Implementation details are explained in Section 4.4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 429, + 318, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 318, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 318, + 441 + ], + "type": "text", + "content": "4.1 High-Quality Range Projection" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 449, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 510 + ], + "type": "text", + "content": "The range image presents LiDAR data compactly and intuitively, with rows indicating the laser beams and columns representing the yaw angles. We convert point clouds to range images using spherical projection. Typically, for a point " + }, + { + "bbox": [ + 130, + 449, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 449, + 482, + 510 + ], + "type": "text", + "content": " in Cartesian coordinates " + }, + { + "bbox": [ + 130, + 449, + 482, + 510 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 130, + 449, + 482, + 510 + ], + "type": "text", + "content": ", we calculate its spherical coordinates " + }, + { + "bbox": [ + 130, + 449, + 482, + 510 + ], + "type": "inline_equation", + "content": "(r,\\theta,\\phi)" + }, + { + "bbox": [ + 130, + 449, + 482, + 510 + ], + "type": "text", + "content": " using:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 174, + 518, + 481, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 518, + 481, + 539 + ], + "spans": [ + { + "bbox": [ + 174, + 518, + 481, + 539 + ], + "type": "interline_equation", + "content": "r = \\sqrt {x ^ {2} + y ^ {2} + z ^ {2}}, \\theta = \\mathrm {a t a n} (y, x), \\phi = \\mathrm {a t a n} \\left(z, \\sqrt {x ^ {2} + y ^ {2}}\\right). \\qquad (2)", + "image_path": "ff1270a53b585668820b4d4128f896ebdd12a39caff4b460cb96a55e310791d8.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 543, + 482, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 543, + 482, + 603 + ], + "spans": [ + { + "bbox": [ + 130, + 543, + 482, + 603 + ], + "type": "text", + "content": "However, in most current datasets such as KITTI-360, multiple lasers from the Velodyne LiDAR system do not share a common origin for their measurements. This may introduce errors in the direct conversion from Cartesian points to spherical points, resulting in incorrect range-view data distribution and thus low-quality range images, as shown at the top of Fig. 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 603, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 603, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 603, + 482, + 640 + ], + "type": "text", + "content": "To address this issue, we adopt Hough Voting to estimate heights and pitch angles " + }, + { + "bbox": [ + 130, + 603, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\{h_j,\\phi_j\\}_{j = 1,\\dots ,N}" + }, + { + "bbox": [ + 130, + 603, + 482, + 640 + ], + "type": "text", + "content": " for Velodyne sensors [3]. We then adjust the point cloud transformation to a range image using" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 197, + 648, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 648, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 197, + 648, + 481, + 669 + ], + "type": "interline_equation", + "content": "r = \\sqrt {x ^ {2} + y ^ {2} + (z - h _ {j}) ^ {2}}, \\theta = \\mathrm {a t a n} (y, x), \\phi = \\phi_ {j}, \\tag {3}", + "image_path": "ac78850d6f3ac97943a580a3ea5ac3e24cb98e230c18e7caedf61fd2d6476064.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 207, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 207, + 91, + 448, + 102 + ], + "type": "text", + "content": "RangeLDM: Fast Realistic LiDAR Point Cloud Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 324, + 178 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 324, + 178 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 324, + 178 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 324, + 178 + ], + "type": "image", + "image_path": "6a8c524874c6ef0c93606503b68a308b8287dbb1a00c4117a9abdee57d58ca3e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 179, + 325, + 213 + ], + "lines": [ + { + "bbox": [ + 132, + 179, + 325, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 179, + 325, + 213 + ], + "type": "text", + "content": "Fig. 3: Comparison of range projection by the typical method described in Eq. 2 and our method with Hough Voting as in Eq. 3." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 334, + 118, + 479, + 209 + ], + "blocks": [ + { + "bbox": [ + 334, + 118, + 479, + 209 + ], + "lines": [ + { + "bbox": [ + 334, + 118, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 334, + 118, + 479, + 209 + ], + "type": "image", + "image_path": "c1c194f7f013d7decc561a9f467f07ec1311cdf6a8edfb270d0c0e7f7b2d0307.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 220, + 355, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 220, + 355, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 220, + 355, + 232 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 220, + 355, + 232 + ], + "type": "inline_equation", + "content": "h_j" + }, + { + "bbox": [ + 132, + 220, + 355, + 232 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 220, + 355, + 232 + ], + "type": "inline_equation", + "content": "\\phi_j" + }, + { + "bbox": [ + 132, + 220, + 355, + 232 + ], + "type": "text", + "content": " refer to the " + }, + { + "bbox": [ + 132, + 220, + 355, + 232 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 132, + 220, + 355, + 232 + ], + "type": "text", + "content": "-th Velodyne sensor." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 233, + 480, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 233, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 480, + 304 + ], + "type": "text", + "content": "We then rasterize points " + }, + { + "bbox": [ + 130, + 233, + 480, + 304 + ], + "type": "inline_equation", + "content": "(r,\\theta ,\\phi)" + }, + { + "bbox": [ + 130, + 233, + 480, + 304 + ], + "type": "text", + "content": " into a 2D cylindrical projection " + }, + { + "bbox": [ + 130, + 233, + 480, + 304 + ], + "type": "inline_equation", + "content": "R(u,v)" + }, + { + "bbox": [ + 130, + 233, + 480, + 304 + ], + "type": "text", + "content": " (a.k.a., range image) of size " + }, + { + "bbox": [ + 130, + 233, + 480, + 304 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 130, + 233, + 480, + 304 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 233, + 480, + 304 + ], + "type": "inline_equation", + "content": "u = ((\\theta +\\pi) / 2\\pi)W,v = j" + }, + { + "bbox": [ + 130, + 233, + 480, + 304 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 233, + 480, + 304 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 130, + 233, + 480, + 304 + ], + "type": "text", + "content": " denotes the grid coordinate of a point in the range image. Thus, we obtain high-quality range images as illustrated at the bottom of Fig. 3. We denote the obtained range image as " + }, + { + "bbox": [ + 130, + 233, + 480, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{x}\\in \\mathbb{R}^{H\\times W\\times 2}" + }, + { + "bbox": [ + 130, + 233, + 480, + 304 + ], + "type": "text", + "content": ", which comprises " + }, + { + "bbox": [ + 130, + 233, + 480, + 304 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 130, + 233, + 480, + 304 + ], + "type": "text", + "content": " pixels associated with both range and intensity " + }, + { + "bbox": [ + 130, + 233, + 480, + 304 + ], + "type": "inline_equation", + "content": "\\{r,i\\}" + }, + { + "bbox": [ + 130, + 233, + 480, + 304 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 304, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 480, + 365 + ], + "type": "text", + "content": "To verify the impact of range projection, we trained range-based LiDAR generation methods like LiDARGen [95] with our obtained high-quality range images. Table 1 shows substantial improvement, which sheds light on the significance of the correct range-view data distribution for such range-based generative model learning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 384, + 204, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 384, + 204, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 384, + 204, + 396 + ], + "type": "text", + "content": "4.2 Training" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 406, + 480, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 406, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 406, + 480, + 501 + ], + "type": "text", + "content": "As a range image is often of high dimensionality at the scale of " + }, + { + "bbox": [ + 130, + 406, + 480, + 501 + ], + "type": "inline_equation", + "content": "64 \\times 1024" + }, + { + "bbox": [ + 130, + 406, + 480, + 501 + ], + "type": "text", + "content": ", it would be computationally expensive to learn the distribution directly. Instead, we compress the range image into a low-dimensional latent space via a VAE and then model the distribution via an LDM. Consequently, the training of the proposed RangeLDM includes two distinct stages. In the first stage, we train a regular VAE to compress the range image into latent features " + }, + { + "bbox": [ + 130, + 406, + 480, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0 \\in \\mathbb{R}^{h \\times w \\times c}" + }, + { + "bbox": [ + 130, + 406, + 480, + 501 + ], + "type": "text", + "content": " by a downsampling factor " + }, + { + "bbox": [ + 130, + 406, + 480, + 501 + ], + "type": "inline_equation", + "content": "f = H / h = W / w" + }, + { + "bbox": [ + 130, + 406, + 480, + 501 + ], + "type": "text", + "content": ". In the second stage, we train an LDM to learn the distribution of the latent encoding " + }, + { + "bbox": [ + 130, + 406, + 480, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 130, + 406, + 480, + 501 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 502, + 480, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 480, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 480, + 550 + ], + "type": "text", + "content": "First-Stage: Dimensionality Reduction. In the first stage, we reduce the dimensionality of the obtained range images by a VAE, which removes imperceptible high-frequency details and leads to low-dimensional latent features that encode prominent information in the original range image." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 550, + 480, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 550, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 550, + 480, + 611 + ], + "type": "text", + "content": "A standard VAE consists of two main components: 1) an encoder " + }, + { + "bbox": [ + 130, + 550, + 480, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\zeta}" + }, + { + "bbox": [ + 130, + 550, + 480, + 611 + ], + "type": "text", + "content": " that transforms an input range image " + }, + { + "bbox": [ + 130, + 550, + 480, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{H \\times W \\times 2}" + }, + { + "bbox": [ + 130, + 550, + 480, + 611 + ], + "type": "text", + "content": " into latent features " + }, + { + "bbox": [ + 130, + 550, + 480, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0 = \\mathcal{E}_{\\zeta}(\\mathbf{x}) \\in \\mathbb{R}^{h \\times w \\times c}" + }, + { + "bbox": [ + 130, + 550, + 480, + 611 + ], + "type": "text", + "content": " and 2) a decoder " + }, + { + "bbox": [ + 130, + 550, + 480, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\eta}" + }, + { + "bbox": [ + 130, + 550, + 480, + 611 + ], + "type": "text", + "content": " that takes the latent features " + }, + { + "bbox": [ + 130, + 550, + 480, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 130, + 550, + 480, + 611 + ], + "type": "text", + "content": " as input and generates the reconstructed range image " + }, + { + "bbox": [ + 130, + 550, + 480, + 611 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}} = \\mathcal{G}_{\\eta}(\\mathbf{z}_0)" + }, + { + "bbox": [ + 130, + 550, + 480, + 611 + ], + "type": "text", + "content": ". The entire VAE is trained by maximizing a modified ELBO with respect to the parameters " + }, + { + "bbox": [ + 130, + 550, + 480, + 611 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 130, + 550, + 480, + 611 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 550, + 480, + 611 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 550, + 480, + 611 + ], + "type": "text", + "content": " [32,54]:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 159, + 622, + 480, + 636 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 622, + 480, + 636 + ], + "spans": [ + { + "bbox": [ + 159, + 622, + 480, + 636 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {E L B O} (\\zeta , \\eta) = \\mathbb {E} _ {q _ {\\zeta} (\\mathbf {z} _ {0} | \\mathbf {x})} \\left[ \\log p _ {\\boldsymbol {\\eta}} (\\mathbf {x} \\mid \\mathbf {z} _ {0}) \\right] - \\lambda_ {1} D _ {\\mathrm {K L}} \\left(q _ {\\zeta} (\\mathbf {z} _ {0} \\mid \\mathbf {x}) \\| p (\\mathbf {z} _ {0})\\right). \\tag {4}", + "image_path": "b83c3ba17eeba4a98ca3fcc31b79d75d6d2e3d17f6903e08ae7200f223b4f55e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "In the above equation, the first term evaluates the reconstruction likelihood of the decoder from the latent " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": " and corresponds to an " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": " reconstruction loss," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "while the second term quantifies how closely the learned distribution of " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " resembles the prior distribution " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "p(\\mathbf{z}_0) = \\mathcal{N}(\\mathbf{z}_0; \\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " held over latent variables. The hyperparameter " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\lambda_1" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " balances the reconstruction accuracy and Kullback-Leibler regularization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 480, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 270 + ], + "type": "text", + "content": "Range-Guided Discriminator. In order to mitigate the blurriness introduced by relying solely on the pixel-space " + }, + { + "bbox": [ + 130, + 164, + 480, + 270 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 130, + 164, + 480, + 270 + ], + "type": "text", + "content": " reconstruction loss, we integrate the VAE with a patch-based adversarial discriminator [15, 29, 55, 87]. Furthermore, to ensure that the reconstruction remains confined within the manifold underlying the range image and exploit the geometric information from the spherical coordinates, we propose a range-guided discriminator " + }, + { + "bbox": [ + 130, + 164, + 480, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\tau}" + }, + { + "bbox": [ + 130, + 164, + 480, + 270 + ], + "type": "text", + "content": ", with parameters " + }, + { + "bbox": [ + 130, + 164, + 480, + 270 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 164, + 480, + 270 + ], + "type": "text", + "content": ". Specifically, we adapt the Meta-Kernel [18] to replace the standard convolution in the discriminator, aiming to learn convolution weights from relative spherical coordinates. The Meta-Kernel is formulated as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 223, + 277, + 480, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 277, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 223, + 277, + 480, + 304 + ], + "type": "interline_equation", + "content": "\\mathbf {h} _ {i} ^ {\\prime} = \\mathcal {W} \\left(\\mathcal {A} _ {j \\in \\mathcal {N} (i)} \\left(\\Phi (\\gamma (\\mathbf {p} _ {j}, \\mathbf {p} _ {i})) \\odot \\mathbf {h} _ {j}\\right)\\right), \\tag {5}", + "image_path": "7ecdb7bc59204410625e8c9e048849b804cc74dd7d2e031981bee21381a622b2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 312, + 482, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 312, + 482, + 383 + ], + "spans": [ + { + "bbox": [ + 131, + 312, + 482, + 383 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 312, + 482, + 383 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 131, + 312, + 482, + 383 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 312, + 482, + 383 + ], + "type": "inline_equation", + "content": "\\mathbf{h}'" + }, + { + "bbox": [ + 131, + 312, + 482, + 383 + ], + "type": "text", + "content": " represent the input feature and output feature, respectively. The function " + }, + { + "bbox": [ + 131, + 312, + 482, + 383 + ], + "type": "inline_equation", + "content": "\\varPhi" + }, + { + "bbox": [ + 131, + 312, + 482, + 383 + ], + "type": "text", + "content": " denotes a Multi-Layer Perceptron (MLP) with two fully-connected layers, " + }, + { + "bbox": [ + 131, + 312, + 482, + 383 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 131, + 312, + 482, + 383 + ], + "type": "text", + "content": " is a concatenation operation, " + }, + { + "bbox": [ + 131, + 312, + 482, + 383 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(i)" + }, + { + "bbox": [ + 131, + 312, + 482, + 383 + ], + "type": "text", + "content": " denotes the 2D-grid neighborhood around the center pixel " + }, + { + "bbox": [ + 131, + 312, + 482, + 383 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 312, + 482, + 383 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 312, + 482, + 383 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 131, + 312, + 482, + 383 + ], + "type": "text", + "content": " is a fully-connected layer. The term " + }, + { + "bbox": [ + 131, + 312, + 482, + 383 + ], + "type": "inline_equation", + "content": "\\gamma (\\mathbf{p}_j,\\mathbf{p}_i)" + }, + { + "bbox": [ + 131, + 312, + 482, + 383 + ], + "type": "text", + "content": " represents the distance between points " + }, + { + "bbox": [ + 131, + 312, + 482, + 383 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_j" + }, + { + "bbox": [ + 131, + 312, + 482, + 383 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 312, + 482, + 383 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 131, + 312, + 482, + 383 + ], + "type": "text", + "content": " in spherical coordinates, which is expressed as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 149, + 392, + 452, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 392, + 452, + 418 + ], + "spans": [ + { + "bbox": [ + 149, + 392, + 452, + 418 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\gamma \\left(\\mathbf {p} _ {j}, \\mathbf {p} _ {i}\\right) := \\left\\{r _ {j} \\cos (\\Delta \\theta) \\cos (\\Delta \\phi) - r _ {i}, r _ {j} \\cos (\\Delta \\theta) \\sin (\\Delta \\phi), r _ {j} \\sin (\\Delta \\theta) \\right\\}, \\\\ \\Delta \\theta = \\theta_ {j} - \\theta_ {i}, \\quad \\Delta \\phi = \\phi_ {j} - \\phi_ {i}. \\\\ \\end{array}", + "image_path": "8b0c16b41c342a0ceebe293540977f92b8f538178de1352b319ca5f8789a76f1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 426, + 479, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 479, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 479, + 474 + ], + "type": "text", + "content": "The Meta-Kernel is aware of local 3D structures by adaptively adjusting convolution kernel weights from relative spherical coordinates. This makes it challenging for the decoder to deceive the range-guided discriminator. This, in turn, encourages the decoder to generate more realistic range images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 475, + 479, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 479, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 479, + 498 + ], + "type": "text", + "content": "With the range-guided discriminator, the final objective of the first-stage training is:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 200, + 508, + 480, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 508, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 200, + 508, + 480, + 525 + ], + "type": "interline_equation", + "content": "\\zeta , \\eta = \\arg \\min _ {\\zeta , \\eta} \\max _ {\\tau} \\mathcal {L} _ {E L B O} (\\zeta , \\eta) + \\lambda_ {2} \\mathcal {L} _ {a d v} (\\zeta , \\eta , \\tau), \\tag {7}", + "image_path": "b004c59c61d53396d3ba5e66ad2d94096d6a52319b67e3681092012e6e91ab2b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 441, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 441, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 441, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 533, + 441, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{adv}" + }, + { + "bbox": [ + 130, + 533, + 441, + 544 + ], + "type": "text", + "content": " is the GAN Hinge loss [39] and " + }, + { + "bbox": [ + 130, + 533, + 441, + 544 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 130, + 533, + 441, + 544 + ], + "type": "text", + "content": " balances the two losses." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 545, + 480, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 480, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 480, + 604 + ], + "type": "text", + "content": "Second-Stage: Latent Diffusion Modeling In the second stage, we freeze the encoder and decoder of the learned VAE and train an LDM on the encoded latent feature " + }, + { + "bbox": [ + 130, + 545, + 480, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 130, + 545, + 480, + 604 + ], + "type": "text", + "content": " for distribution modeling of the range image. During the diffusion process, noise is added to the initial latent " + }, + { + "bbox": [ + 130, + 545, + 480, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 130, + 545, + 480, + 604 + ], + "type": "text", + "content": ", resulting in a noisy latent " + }, + { + "bbox": [ + 130, + 545, + 480, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 130, + 545, + 480, + 604 + ], + "type": "text", + "content": ", where the noise level increases over time steps " + }, + { + "bbox": [ + 130, + 545, + 480, + 604 + ], + "type": "inline_equation", + "content": "t \\in T" + }, + { + "bbox": [ + 130, + 545, + 480, + 604 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 605, + 479, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 479, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 479, + 640 + ], + "type": "text", + "content": "Regarding the modeling of the unconditional distribution " + }, + { + "bbox": [ + 130, + 605, + 479, + 640 + ], + "type": "inline_equation", + "content": "p(\\mathbf{z}_0)" + }, + { + "bbox": [ + 130, + 605, + 479, + 640 + ], + "type": "text", + "content": ", based on Eq. 1, we learn a time-conditional UNet [56] " + }, + { + "bbox": [ + 130, + 605, + 479, + 640 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_t,t)" + }, + { + "bbox": [ + 130, + 605, + 479, + 640 + ], + "type": "text", + "content": " that predicts the noise added to the noisy latent " + }, + { + "bbox": [ + 130, + 605, + 479, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 130, + 605, + 479, + 640 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 213, + 649, + 480, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 649, + 480, + 669 + ], + "spans": [ + { + "bbox": [ + 213, + 649, + 480, + 669 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {L D M}} := \\mathbb {E} _ {\\mathcal {E} (x), \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t\\right) \\| _ {2} ^ {2} \\right]. \\tag {8}", + "image_path": "8b3055a00cdc96878a243a33c70498c7c3fa20d8dec37ecc11d1a9e1965aa827.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 207, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 207, + 91, + 448, + 102 + ], + "type": "text", + "content": "RangeLDM: Fast Realistic LiDAR Point Cloud Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 150 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 150 + ], + "type": "text", + "content": "As to the modeling of the conditional distribution " + }, + { + "bbox": [ + 130, + 116, + 479, + 150 + ], + "type": "inline_equation", + "content": "p(\\mathbf{z}_0|\\mathbf{y})" + }, + { + "bbox": [ + 130, + 116, + 479, + 150 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 479, + 150 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 116, + 479, + 150 + ], + "type": "text", + "content": " is a condition such as sparse or masked point clouds, we learn a conditional LDM via" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 195, + 152, + 481, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 152, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 195, + 152, + 481, + 172 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {L D M}} := \\mathbb {E} _ {\\mathcal {E} (x), \\mathbf {y}, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {z} _ {t}, t, \\tau_ {\\theta} (\\mathbf {y})) \\| _ {2} ^ {2} \\right], \\tag {9}", + "image_path": "c6499217fd40749388cf90e7420872703a579b54c9e3ec59d46c9ed9dbf17ad1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 180, + 479, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 180, + 479, + 217 + ], + "spans": [ + { + "bbox": [ + 130, + 180, + 479, + 217 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 180, + 479, + 217 + ], + "type": "inline_equation", + "content": "\\tau_{\\theta}" + }, + { + "bbox": [ + 130, + 180, + 479, + 217 + ], + "type": "text", + "content": " is a condition encoder that projects " + }, + { + "bbox": [ + 130, + 180, + 479, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 180, + 479, + 217 + ], + "type": "text", + "content": " into an intermediate representation " + }, + { + "bbox": [ + 130, + 180, + 479, + 217 + ], + "type": "inline_equation", + "content": "\\tau_{\\theta}(\\mathbf{y})" + }, + { + "bbox": [ + 130, + 180, + 479, + 217 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 180, + 479, + 217 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{z}_t, t, \\tau_{\\theta}(\\mathbf{y}))" + }, + { + "bbox": [ + 130, + 180, + 479, + 217 + ], + "type": "text", + "content": " is a UNet interpolated with cross-attention layers [55, 75]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 241, + 217, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 241, + 217, + 252 + ], + "spans": [ + { + "bbox": [ + 132, + 241, + 217, + 252 + ], + "type": "text", + "content": "4.3 Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 267, + 479, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 479, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 479, + 327 + ], + "type": "text", + "content": "Unconditional Generation. With the decoder " + }, + { + "bbox": [ + 130, + 267, + 479, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\eta}" + }, + { + "bbox": [ + 130, + 267, + 479, + 327 + ], + "type": "text", + "content": " and the LDM " + }, + { + "bbox": [ + 130, + 267, + 479, + 327 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 267, + 479, + 327 + ], + "type": "text", + "content": ", we establish a hierarchical generative model " + }, + { + "bbox": [ + 130, + 267, + 479, + 327 + ], + "type": "inline_equation", + "content": "p_{\\eta, \\theta}(\\mathbf{x}, \\mathbf{z}_0) = p_{\\eta}(\\mathbf{x} | \\mathbf{z}_0) p_{\\theta}(\\mathbf{z}_0)" + }, + { + "bbox": [ + 130, + 267, + 479, + 327 + ], + "type": "text", + "content": ". That is, we first generate latent features with the LDM, and then map the latent features back to the original range image space using the decoder " + }, + { + "bbox": [ + 130, + 267, + 479, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\eta}" + }, + { + "bbox": [ + 130, + 267, + 479, + 327 + ], + "type": "text", + "content": " for LiDAR point cloud generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 329, + 479, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 479, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 479, + 389 + ], + "type": "text", + "content": "Conditional Generation. Given a condition " + }, + { + "bbox": [ + 130, + 329, + 479, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 329, + 479, + 389 + ], + "type": "text", + "content": ", the conditional generative model is defined as " + }, + { + "bbox": [ + 130, + 329, + 479, + 389 + ], + "type": "inline_equation", + "content": "p_{\\eta, \\theta}(\\mathbf{x}, \\mathbf{z}_0 | \\mathbf{y}) = p_{\\eta}(\\mathbf{x} | \\mathbf{z}_0, \\mathbf{y}) p_{\\theta}(\\mathbf{z}_0 | \\mathbf{y})" + }, + { + "bbox": [ + 130, + 329, + 479, + 389 + ], + "type": "text", + "content": ". There could be various conditions for LiDAR point cloud generation, such as sparse point clouds, incomplete point clouds, camera or even text. Here, we illustrate two applications of conditional generation: LiDAR point cloud upsampling and inpainting." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 389, + 480, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 480, + 498 + ], + "type": "text", + "content": "LiDAR Point Cloud Upsampling. This task takes a sparse point cloud" + }, + { + "bbox": [ + 130, + 389, + 480, + 498 + ], + "type": "inline_equation", + "content": "^1" + }, + { + "bbox": [ + 130, + 389, + 480, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{y} \\in \\mathbb{R}^{h \\times W \\times 2}" + }, + { + "bbox": [ + 130, + 389, + 480, + 498 + ], + "type": "text", + "content": " as input (following the setting of LiDARGen) and expects the model to produce a denser counterpart " + }, + { + "bbox": [ + 130, + 389, + 480, + 498 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}} \\in \\mathbb{R}^{H \\times W \\times 2}" + }, + { + "bbox": [ + 130, + 389, + 480, + 498 + ], + "type": "text", + "content": ", which is necessary for accurate LiDAR-based perception and density-insensitive domain adaptation. We train LDM following Eq. 9 with the sparse point cloud " + }, + { + "bbox": [ + 130, + 389, + 480, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 389, + 480, + 498 + ], + "type": "text", + "content": " and its ground-truth dense variant " + }, + { + "bbox": [ + 130, + 389, + 480, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 389, + 480, + 498 + ], + "type": "text", + "content": ". The condition encoder " + }, + { + "bbox": [ + 130, + 389, + 480, + 498 + ], + "type": "inline_equation", + "content": "\\tau_{\\theta}" + }, + { + "bbox": [ + 130, + 389, + 480, + 498 + ], + "type": "text", + "content": " is configured as a reshaper that transforms the sparse point cloud " + }, + { + "bbox": [ + 130, + 389, + 480, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 389, + 480, + 498 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 130, + 389, + 480, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{y}' \\in \\mathbb{R}^{h \\times w \\times 2f}" + }, + { + "bbox": [ + 130, + 389, + 480, + 498 + ], + "type": "text", + "content": ". The reshaped " + }, + { + "bbox": [ + 130, + 389, + 480, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{y}'" + }, + { + "bbox": [ + 130, + 389, + 480, + 498 + ], + "type": "text", + "content": " is then concatenated with noisy latent variable " + }, + { + "bbox": [ + 130, + 389, + 480, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 130, + 389, + 480, + 498 + ], + "type": "text", + "content": " and fed into the UNet " + }, + { + "bbox": [ + 130, + 389, + 480, + 498 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 389, + 480, + 498 + ], + "type": "text", + "content": " for noise prediction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 498, + 481, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 631 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 631 + ], + "type": "text", + "content": "LiDAR Point Cloud Inpainting. This task is required when LiDAR sensors cannot capture the entire scene due to object occlusion or sensor limitations. We train an inpainting LDM from the ground truth point cloud " + }, + { + "bbox": [ + 130, + 498, + 481, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 498, + 481, + 631 + ], + "type": "text", + "content": ", a masked point cloud " + }, + { + "bbox": [ + 130, + 498, + 481, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'" + }, + { + "bbox": [ + 130, + 498, + 481, + 631 + ], + "type": "text", + "content": ", and the corresponding mask " + }, + { + "bbox": [ + 130, + 498, + 481, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{m} \\in \\{0,1\\}^{H \\times W}" + }, + { + "bbox": [ + 130, + 498, + 481, + 631 + ], + "type": "text", + "content": ". The masked point cloud " + }, + { + "bbox": [ + 130, + 498, + 481, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'" + }, + { + "bbox": [ + 130, + 498, + 481, + 631 + ], + "type": "text", + "content": " is encoded with the encoder " + }, + { + "bbox": [ + 130, + 498, + 481, + 631 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\zeta}" + }, + { + "bbox": [ + 130, + 498, + 481, + 631 + ], + "type": "text", + "content": " from the VAE and is then concatenated with a downsampled mask " + }, + { + "bbox": [ + 130, + 498, + 481, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{m}' \\in \\{0,1\\}^{h \\times w}" + }, + { + "bbox": [ + 130, + 498, + 481, + 631 + ], + "type": "text", + "content": " as the condition term in Eq. 9. Mathematically, we represent the condition " + }, + { + "bbox": [ + 130, + 498, + 481, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 498, + 481, + 631 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 498, + 481, + 631 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}',\\mathbf{m}\\}" + }, + { + "bbox": [ + 130, + 498, + 481, + 631 + ], + "type": "text", + "content": ", and deploy the condition encoder to project it into " + }, + { + "bbox": [ + 130, + 498, + 481, + 631 + ], + "type": "inline_equation", + "content": "\\tau_{\\theta}(\\mathbf{y}) = \\mathcal{E}_{\\zeta}(\\mathbf{x}') \\| ds(\\mathbf{m})" + }, + { + "bbox": [ + 130, + 498, + 481, + 631 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 498, + 481, + 631 + ], + "type": "inline_equation", + "content": "\\|\\cdot\\|" + }, + { + "bbox": [ + 130, + 498, + 481, + 631 + ], + "type": "text", + "content": " denotes the concatenation operation and " + }, + { + "bbox": [ + 130, + 498, + 481, + 631 + ], + "type": "inline_equation", + "content": "ds(\\cdot)" + }, + { + "bbox": [ + 130, + 498, + 481, + 631 + ], + "type": "text", + "content": " indicates the downsampling operation. The projected condition is then concatenated with the noisy latent variable " + }, + { + "bbox": [ + 130, + 498, + 481, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 130, + 498, + 481, + 631 + ], + "type": "text", + "content": " and input into the UNet " + }, + { + "bbox": [ + 130, + 498, + 481, + 631 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 498, + 481, + 631 + ], + "type": "text", + "content": " for noise prediction." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Hu et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 479, + 664 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 479, + 664 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 479, + 664 + ], + "type": "text", + "content": "1 Without causing ambiguity, the point clouds are converted to range images by default." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 479, + 224 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 479, + 224 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 479, + 224 + ], + "type": "image", + "image_path": "a928bec7ff81d9dcf368aa8a104b347ccb364abccb06399f799be61df6d7c880.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 225, + 480, + 280 + ], + "lines": [ + { + "bbox": [ + 130, + 225, + 480, + 280 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 480, + 280 + ], + "type": "text", + "content": "Fig. 4: Qualitative results comparing against baselines for unconditional Li-DAR generation on KITTI-360. Real point clouds are only for reference. Our model produces results that closely resemble real-world data, which excels in generating road scenes, such as cars (the first row), road bollards (the second row) and crossroads (the last row)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 287, + 279, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 287, + 279, + 298 + ], + "spans": [ + { + "bbox": [ + 132, + 287, + 279, + 298 + ], + "type": "text", + "content": "4.4 Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 306, + 480, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 480, + 413 + ], + "type": "text", + "content": "Circular Convolution. Range images are inherently circular, which means that the left boundary of a range image is actually connected to its right boundary. However, standard convolutions employ zero-padding or symmetry padding and do not take into account such constraints. Thus, as in LiDArgen [95], we replace standard convolutions in both the VAE and LDM with circular convolutions [59], which treat the left and right boundaries as connected neighbors in the topology. Since the 2D convolution operator satisfies translation invariance, it is evident that the entire network is invariant to horizontal shifting of the range image (i.e., rotation in the xy-plane of the point cloud)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 414, + 480, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 480, + 557 + ], + "type": "text", + "content": "Conditioning the Model with Direction. Real LiDAR point clouds collected from driving scenes usually have a certain directionality. For example, since vehicles typically follow the direction of the road, the x-axis direction of the LiDAR point cloud usually aligns with the road direction. However, since the whole network of our model is invariant to rotation in the xy-plane of the point cloud, the point clouds generated from random noise could result in arbitrary directions. To avoid this, we generate point clouds with directional conditioning. Specifically, we use an " + }, + { + "bbox": [ + 130, + 414, + 480, + 557 + ], + "type": "inline_equation", + "content": "h \\times w" + }, + { + "bbox": [ + 130, + 414, + 480, + 557 + ], + "type": "text", + "content": " matrix as the condition, in which only the values from the first column are set to 1, while the rest are all set to 0. This 1-valued column corresponds to the x-axis of the point cloud. Such simple conditioning demonstrates effective control over the direction of the generated point cloud, as presented in the qualitative results of the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "content": "Network Architectures and Model Hyperparameters. For simplicity, we adopted a similar architecture to that in [55]. The downsampling factor " + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "content": " is set as 4 to strike a good balance between efficiency and realistic results. The VAE comprises three encoder blocks and three decoder blocks. The LDM consists of four downsampling blocks with the last three followed by a transformer layer, and four upsampling blocks with the first three blocks also followed by a transformer layer. In the generation process, we employed 50 denoising steps with DDIM sampler [66] for point cloud generation. More implementation details can be found in released code https://github.com/WoodwindHu/RangeLDM." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "type": "text", + "content": "RangeLDM: Fast Realistic LiDAR Point Cloud Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 116, + 481, + 235 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 235 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 235 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 235 + ], + "type": "image", + "image_path": "c1f715f0ec81ea3370f4aec310bbc2f0abae528348e19494642888e0efad7f3b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 236, + 480, + 268 + ], + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 268 + ], + "type": "text", + "content": "Fig. 5: Comparison of upsampling results on KITTI-360. We downsampled the ground truth by a factor of four as the input, and demonstrated the results of different methods on " + }, + { + "bbox": [ + 132, + 236, + 480, + 268 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 132, + 236, + 480, + 268 + ], + "type": "text", + "content": "-upsampling of the input." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 274, + 229, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 229, + 288 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 229, + 288 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 299, + 261, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 299, + 261, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 299, + 261, + 312 + ], + "type": "text", + "content": "5.1 Experimental Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 318, + 481, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 481, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 481, + 378 + ], + "type": "text", + "content": "Datasets. We evaluate our model on two challenging datasets, KITTI-360 and nuScenes. KITTI-360 has 81,106 LiDAR readings from 9 sequences in Germany, covering diverse scenes. We used the first two sequences for validation and trained on the rest. NuScenes is a public dataset with 297,737 LiDAR sweeps in the training set and 52,423 in the testing set, collected in Boston and Singapore." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 379, + 481, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 481, + 474 + ], + "type": "text", + "content": "Evaluation metrics. Following [95], we employed three metrics to perform quantitative analysis: Maximum Mean Discrepancy (MMD), Jensen-Shannon divergence (JSD), and Frechet Range Distance (FRD score). We use a " + }, + { + "bbox": [ + 130, + 379, + 481, + 474 + ], + "type": "inline_equation", + "content": "100 \\times 100" + }, + { + "bbox": [ + 130, + 379, + 481, + 474 + ], + "type": "text", + "content": " 2D histogram on the BEV plane to calculate MMD and JSD metrics. The Frechet Range Distance (FRD) score [95] is a metric used to evaluate the quality of samples acquired by a generative model, inspired by the FID score for images [24]. To compute the FRD score, we use RangeNet++ [45], an encoder-decoder-based network for segmentation, which is pre-trained on KITTI-360." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 474, + 481, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 510 + ], + "type": "text", + "content": "Baselines. We evaluated our approach for LiDAR point cloud generation against several competitive methods, including LiDAR VAE [5], LiDAR GAN [5], Projected GAN [58], LiDARGen [95] and UltraLiDAR [79]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 528, + 291, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 528, + 291, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 528, + 291, + 538 + ], + "type": "text", + "content": "5.2 Unconditional Generation" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 187, + 544, + 427, + 616 + ], + "blocks": [ + { + "bbox": [ + 187, + 544, + 427, + 616 + ], + "lines": [ + { + "bbox": [ + 187, + 544, + 427, + 616 + ], + "spans": [ + { + "bbox": [ + 187, + 544, + 427, + 616 + ], + "type": "table", + "html": "
MethodYears\\( MMD_{BEV} \\)↓FRD ↓\\( JSD_{BEV} \\)↓
LiDAR GAN [5]IROS 2019\\( 3.06 \\times 10^{-3} \\)3003.8-
LiDAR VAE [5]IROS 2019\\( 1.00 \\times 10^{-3} \\)2261.50.161
Projected GAN [58]NeurIPS 2021\\( 3.47 \\times 10^{-4} \\)2117.20.085
LiDARGen [95]ECCV 2022\\( 3.87 \\times 10^{-4} \\)2040.10.067
UltraLiDAR [79]CVPR 2023\\( 1.96 \\times 10^{-4} \\)-0.071
Ours\\( 3.07 \\times 10^{-5} \\)1074.90.045
", + "image_path": "9ae5726bb744bd142bcfbd8751aaf882cabed043b96c1fd26ca8f891f5a581a1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 179, + 616, + 433, + 627 + ], + "lines": [ + { + "bbox": [ + 179, + 616, + 433, + 627 + ], + "spans": [ + { + "bbox": [ + 179, + 616, + 433, + 627 + ], + "type": "text", + "content": "Table 2: Unconditional generation results on KITTI-360 [38]." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Quantitative Results on KITTI-360. Table 2 displays the quantitative results of the proposed RangeLDM and competing algorithms. The results clearly" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "illustrate that our method significantly outperforms the baselines across all metrics. Notably, our results on the MMD metric are an order of magnitude lower than those of other methods. This underscores the remarkable expressive capabilities of our model in the context of LiDAR point cloud generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 166, + 482, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 166, + 482, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 166, + 482, + 334 + ], + "type": "text", + "content": "Qualitative Results on KITTI-360. Figure 4 displays a set of randomly generated samples from competing algorithms, alongside real point cloud samples extracted from the KITTI-360 dataset for comparison. We observe that the Projected GAN [58] exhibits noticeable artifacts in the distant range. LiDARGen [95] effectively captures the general layout, but produces rather noisy outputs and lacks the same degree of straight walls as actual samples. UltraLiDAR [79] generates structured and reasonable scenes, yet it only creates voxelized point clouds (as magnified in the first row) without intensity features. In contrast, our model consistently surpasses the baseline models and yields results closely resembling real-world data. For example, we can generate cars (emphasized by red boxes in the first row of Figure 4) and road bollards (red boxes in the second row of Figure 4) on the road. Additionally, our model generates various scenes such as straightway (the first two rows of Figure 4) and crossroads (the last row of Figure 4). The supplementary materials provide more visualization results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 335, + 298, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 298, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 298, + 419 + ], + "type": "text", + "content": "Human study on KITTI-360. To assess the perceptual quality, we conducted an A/B test involving a group of 14 researchers with LiDAR expertise. Following the same evaluation system as [95] and [79], we present pairs of randomly selected images" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 308, + 338, + 478, + 399 + ], + "blocks": [ + { + "bbox": [ + 308, + 338, + 478, + 399 + ], + "lines": [ + { + "bbox": [ + 308, + 338, + 478, + 399 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 478, + 399 + ], + "type": "table", + "html": "
MethodPercent prefer ours
Ours vs. VAE [5]98.8%
Ours vs. GAN [5]98.0%
Ours vs. ProjectedGAN [58]93.2%
Ours vs. LiDARGen [95]90.2%
Ours vs. UltraLiDAR [79]86.5%
", + "image_path": "db4c042a681e0ac93dd682265290e759d1640b550b45b8980b803d4394592926.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 419, + 482, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 482, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 482, + 467 + ], + "type": "text", + "content": "from two point clouds and ask participants to determine which one appeared more realistic. The results, as displayed in Table 3, unequivocally demonstrate the superior visual quality of generation results by our model. In most of the cases, testers favored our results over the baselines." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 470, + 332, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 332, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 332, + 553 + ], + "type": "text", + "content": "Evaluation on the NuScenes Dataset. The results listed in Table 4 demonstrate that our model outperforms LiDAR VAE and LiDARGen significantly over nuScenes. In particular, our results on the MMD metric are an order of magnitude lower than those of other methods, while we reduce to about " + }, + { + "bbox": [ + 130, + 470, + 332, + 553 + ], + "type": "inline_equation", + "content": "1/3" + }, + { + "bbox": [ + 130, + 470, + 332, + 553 + ], + "type": "text", + "content": " of the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 554, + 479, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 554, + 479, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 554, + 479, + 589 + ], + "type": "text", + "content": "results of competitive methods in terms of the JSD metric. Also, Figure 6 illustrates the superiority of our model over LiDAR VAE and LiDARGen in terms of qualitative comparison." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 340, + 472, + 479, + 513 + ], + "blocks": [ + { + "bbox": [ + 312, + 399, + 473, + 410 + ], + "lines": [ + { + "bbox": [ + 312, + 399, + 473, + 410 + ], + "spans": [ + { + "bbox": [ + 312, + 399, + 473, + 410 + ], + "type": "text", + "content": "Table 3: Human study on KITTI-360." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 340, + 472, + 479, + 513 + ], + "lines": [ + { + "bbox": [ + 340, + 472, + 479, + 513 + ], + "spans": [ + { + "bbox": [ + 340, + 472, + 479, + 513 + ], + "type": "table", + "html": "
Method\\( MMD_{BEV} \\)↓\\( JSD_{BEV} \\)↓
LiDAR VAE [5]\\( 1.1 \\times 10^{-3} \\)-
\\( LiDARGen^† \\)[95]\\( 1.9 \\times 10^{-3} \\)0.160
Ours\\( 1.9 \\times 10^{-4} \\)0.054
", + "image_path": "41bab5708cb0874f32ee37fcf4c282c2aed1d95d301caf60073acdf91628d1b7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 339, + 513, + 480, + 545 + ], + "lines": [ + { + "bbox": [ + 339, + 513, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 339, + 513, + 480, + 545 + ], + "type": "text", + "content": "Table 4: Unconditional generation results on nuScenes dataset [6]. †: Reproduced by us." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 591, + 482, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 591, + 482, + 664 + ], + "spans": [ + { + "bbox": [ + 130, + 591, + 482, + 664 + ], + "type": "text", + "content": "Model Size. The number of parameters in LiDARGen [95] and UltraLiDAR [79] are 29.7M and 40.3M, respectively. To ensure fair competition, we limited the capacity of our model to be similar to other approaches. Our model comprises 41.4M parameters, consisting of 12.7M parameters for VAE and 28.7M parameters for LDM. As illustrated in Table 2, our model outperforms the baselines by a large margin under similar model sizes." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 207, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 207, + 90, + 447, + 102 + ], + "type": "text", + "content": "RangeLDM: Fast Realistic LiDAR Point Cloud Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 186, + 114, + 428, + 201 + ], + "blocks": [ + { + "bbox": [ + 186, + 114, + 428, + 201 + ], + "lines": [ + { + "bbox": [ + 186, + 114, + 428, + 201 + ], + "spans": [ + { + "bbox": [ + 186, + 114, + 428, + 201 + ], + "type": "image", + "image_path": "f95bf65998d7447a178c9b8f34403bdb06caba45f1541d1d6839153fb0bd9016.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 201, + 479, + 224 + ], + "lines": [ + { + "bbox": [ + 132, + 201, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 201, + 479, + 224 + ], + "type": "text", + "content": "Fig. 6: Qualitative results comparing against LiDAR VAE and LiDARGen for unconditional LiDAR generation on nuScenes. Real point clouds are only for reference." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 153, + 228, + 461, + 301 + ], + "blocks": [ + { + "bbox": [ + 153, + 228, + 461, + 301 + ], + "lines": [ + { + "bbox": [ + 153, + 228, + 461, + 301 + ], + "spans": [ + { + "bbox": [ + 153, + 228, + 461, + 301 + ], + "type": "table", + "html": "
Hough Range-Guided \nVoting DiscriminatorCircular \nConvolutionDirection \nConditionedMMDBEV ↓FRD ↓JSDBEV ↓
(a)××××3.95 × 10-41536.70.067
(b)✓×××6.57 × 10-51229.30.056
(c)✓✓××4.72 × 10-51103.10.051
(d)✓✓✓×3.90 × 10-41797.20.078
(e)✓✓✓✓3.07 × 10-51074.90.045
", + "image_path": "53f8439d4f3cbdbd3bda733ce4cdbf5643e427800f3f4fe151dc37d4641133ac.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 327, + 350, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 327, + 350, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 350, + 423 + ], + "type": "text", + "content": "Generation Efficiency. The proposed dimensionality reduction offers RangeLDM a considerable advantage in terms of generation speed. As indicated in Table 5, the proposed method significantly outperforms LiDARGen [95] and UltraLi-DAR [79] in terms of generation speed. In particular, our model is 200 times faster than LiDARGen and 30 times faster than UltraLi-DAR." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 359, + 330, + 479, + 381 + ], + "blocks": [ + { + "bbox": [ + 243, + 301, + 370, + 312 + ], + "lines": [ + { + "bbox": [ + 243, + 301, + 370, + 312 + ], + "spans": [ + { + "bbox": [ + 243, + 301, + 370, + 312 + ], + "type": "text", + "content": "Table 6: Main ablation study." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 359, + 330, + 479, + 381 + ], + "lines": [ + { + "bbox": [ + 359, + 330, + 479, + 381 + ], + "spans": [ + { + "bbox": [ + 359, + 330, + 479, + 381 + ], + "type": "table", + "html": "
MethodThroughput ↑ (samples/s)
LiDARGen [95]0.02
UltraLiDAR [79]0.16
Ours4.86
", + "image_path": "cf3634e45c8488387553b6321675678312548ae926b19bb0f47453f9a84313d2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 441, + 279, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 279, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 279, + 453 + ], + "type": "text", + "content": "5.3 Conditional Generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 462, + 480, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 480, + 497 + ], + "type": "text", + "content": "To evaluate the conditional generation performance of RangeLDM, we conducted experiments on the KITTI-360 dataset over two tasks: LiDAR point cloud upsampling and inpainting." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 498, + 480, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 480, + 545 + ], + "type": "text", + "content": "LiDAR Point Cloud Upsampling. We obtain sparse input point clouds by selecting a subset of 16 beams from the raw 64-beam sensors, in line with LiDARGen. We compared our approach against PUNet [88], DeepRS [9], Grad-PU [23], bicubic interpolation, Nearest Neighbor (NN) interpolation and LiDARGen." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 546, + 304, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 546, + 304, + 641 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 304, + 641 + ], + "type": "text", + "content": "Quantitatively, in addition to measuring Mean Absolute Error (MAE) in the range view, we also employed RangeNet++ semantic segmentation to evaluate the quality of upsampled results. As shown in Table 7, the proposed method outperforms competing methods in all metrics, including" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 308, + 549, + 479, + 628 + ], + "blocks": [ + { + "bbox": [ + 357, + 382, + 481, + 403 + ], + "lines": [ + { + "bbox": [ + 357, + 382, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 357, + 382, + 481, + 403 + ], + "type": "text", + "content": "Table 5: Inference speed on a single RTX 3090 GPU." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 549, + 479, + 628 + ], + "lines": [ + { + "bbox": [ + 308, + 549, + 479, + 628 + ], + "spans": [ + { + "bbox": [ + 308, + 549, + 479, + 628 + ], + "type": "table", + "html": "
MethodYearsMAE ↓Accuracy ↑ IoU ↑
PUNet [88]20186.88--
DeepRS [9]20223.96--
Grad-PU [23]20235.09--
Bicubic-2.600.2650.166
NN-2.180.5460.394
LiDARGen [95]20221.230.6080.449
Ours0.890.7220.566
", + "image_path": "966fd642ac6b200f9e8a89c7a312cbc7d2f014c748a05400bc3695d789b8b0d6.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 331, + 628, + 454, + 639 + ], + "lines": [ + { + "bbox": [ + 331, + 628, + 454, + 639 + ], + "spans": [ + { + "bbox": [ + 331, + 628, + 454, + 639 + ], + "type": "text", + "content": "Table 7: LiDAR upsampling." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 642, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 642, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 642, + 480, + 665 + ], + "type": "text", + "content": "MAE, per-point segmentation accuracy, and segmentation Intersection over Union (IoU)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 186, + 114, + 430, + 191 + ], + "blocks": [ + { + "bbox": [ + 186, + 114, + 430, + 191 + ], + "lines": [ + { + "bbox": [ + 186, + 114, + 430, + 191 + ], + "spans": [ + { + "bbox": [ + 186, + 114, + 430, + 191 + ], + "type": "image", + "image_path": "cc21bf27bd40cbe214ff1e23d79ef5777a37a5bfaaeaef5548d3b3ab4716a33d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 191, + 479, + 213 + ], + "lines": [ + { + "bbox": [ + 131, + 191, + 479, + 213 + ], + "spans": [ + { + "bbox": [ + 131, + 191, + 479, + 213 + ], + "type": "text", + "content": "Fig. 7: Inpainting results. Left: ground truth point clouds. Middle and right (red): input point clouds. Middle and right (blue): recovered point clouds." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 138, + 219, + 478, + 264 + ], + "blocks": [ + { + "bbox": [ + 138, + 219, + 478, + 264 + ], + "lines": [ + { + "bbox": [ + 138, + 219, + 478, + 264 + ], + "spans": [ + { + "bbox": [ + 138, + 219, + 478, + 264 + ], + "type": "image", + "image_path": "59aec08343dbaccee62d95b22d4da83d42eda9f7fa840c17a9bb4fc184e794d2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 266, + 479, + 288 + ], + "lines": [ + { + "bbox": [ + 131, + 266, + 479, + 288 + ], + "spans": [ + { + "bbox": [ + 131, + 266, + 479, + 288 + ], + "type": "text", + "content": "Fig. 8: VAE reconstruction results. The proposed VAE with a range-guided discriminator reconstructs point clouds with less noise and more precise object structures." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 297, + 480, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 480, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 480, + 355 + ], + "type": "text", + "content": "Qualitatively, the visualization comparison with several competitive methods presented in Figure 5 indicates that the proposed upsampling method exhibits results highly similar to the ground truth data, showing clear object details and consistent LiDAR scan lines. We provide additional visualization results in the supplementary materials." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 357, + 481, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 481, + 453 + ], + "type": "text", + "content": "LiDAR Point Cloud Inpainting. To mimic the case of missing data in LiDAR point clouds, we mask point clouds in front of the vehicle within a " + }, + { + "bbox": [ + 130, + 357, + 481, + 453 + ], + "type": "inline_equation", + "content": "22.5^{\\circ}" + }, + { + "bbox": [ + 130, + 357, + 481, + 453 + ], + "type": "text", + "content": " range and perform inpainting using LiDARGen and our method. Our results achieve much less MAE (0.190) between the reconstructed results and ground truth compared to that of LiDARGen (0.367). As shown in Figure 7, LiDARGen generates roads but fails to recover the masked car (the first row) and the wall (the second row), while we are able to reconstruct the details. More results are presented in the supplementary materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 474, + 237, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 474, + 237, + 487 + ], + "spans": [ + { + "bbox": [ + 132, + 474, + 237, + 487 + ], + "type": "text", + "content": "5.4 Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 497, + 403, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 497, + 403, + 508 + ], + "spans": [ + { + "bbox": [ + 131, + 497, + 403, + 508 + ], + "type": "text", + "content": "All ablation studies are conducted on the KITTI-360 dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 481, + 666 + ], + "type": "text", + "content": "Main Ablation. As demonstrated in Table 6, we investigate the contribution of each component. Starting from the backbone model (a) without any component, we gradually add each component for evaluation. By applying Hough Voting to the training, the performance of variant (b) improves significantly, demonstrating the huge impact of the correct range-view data distribution. With the addition of the range-guided discriminator component, model (c) achieves better performance, thus clarifying the effectiveness of the geometry-sensitive module. Compared to model (c), the performance of model (d) decreases because circular convolution generates LiDAR scenes with arbitrary directions due to rotation invariance. However, when combined with direction-conditioned generation that guides the model to generate point clouds in the correct direction, the overall performance improves, as demonstrated by model (e). We refer the readers to the supplementary materials for qualitative results." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "type": "text", + "content": "RangeLDM: Fast Realistic LiDAR Point Cloud Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 170, + 114, + 444, + 167 + ], + "blocks": [ + { + "bbox": [ + 170, + 114, + 444, + 167 + ], + "lines": [ + { + "bbox": [ + 170, + 114, + 444, + 167 + ], + "spans": [ + { + "bbox": [ + 170, + 114, + 444, + 167 + ], + "type": "table", + "html": "
ModelVAEModel Size (M)\\( MMD_{BEV} \\)↓FRD ↓\\( JSD_{BEV} \\)↓Throughput ↑(samples/s)
Ours-DMS×40.3\\( 1.03 × 10^{-4} \\)1392.20.0621.52
Ours-DM×114.7\\( 4.14 × 10^{-5} \\)899.00.0400.64
Ours✓41.4\\( 3.07 × 10^{-5} \\)1074.90.0454.86
", + "image_path": "f8e3e089e149013022e75c2fff79a2d9c560caa6af846c065f95823e2d440412.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 216, + 482, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 323 + ], + "type": "text", + "content": "Latent diffusion. We explore the contribution of latent diffusion to the generation quality and generation speed of the model. As shown in Table 8, we constructed two variants of different sizes that directly generate range images with diffusion models, denoted as \"Ours-DMS\" (with a smaller size) and \"Ours-DM\" (with a larger size), for comparison. It turns out that a much larger model size is required for \"Ours-DM\" to achieve competitive performance with \"Ours\", due to the multitude of high-frequency details contained in the range image. Additionally, the superior throughput of \"Ours\" demonstrates the efficiency of latent diffusion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 324, + 281, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 324, + 281, + 383 + ], + "spans": [ + { + "bbox": [ + 132, + 324, + 281, + 383 + ], + "type": "text", + "content": "VAE Architecture. We investigate the contribution of the proposed range-guided discriminator by comparing our VAE with two baselines: 1) VAE without dis" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 291, + 327, + 480, + 369 + ], + "blocks": [ + { + "bbox": [ + 132, + 168, + 481, + 201 + ], + "lines": [ + { + "bbox": [ + 132, + 168, + 481, + 201 + ], + "spans": [ + { + "bbox": [ + 132, + 168, + 481, + 201 + ], + "type": "text", + "content": "Table 8: Ablation of latent diffusion. Model without latent diffusion requires more parameters to achieve similar performance with the latent diffusion model. We generated 1000 samples for calculating the throughput." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 291, + 327, + 480, + 369 + ], + "lines": [ + { + "bbox": [ + 291, + 327, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 291, + 327, + 480, + 369 + ], + "type": "table", + "html": "
DiscriminatorPSNRrange↑ MAE↓ FRD↓ CD
No26.770.0195532.90.0808
Vanilla26.700.0189496.70.0726
Range-Guided27.190.0186483.60.0676
", + "image_path": "7fea9b7e1fffd3289957b2731578b6a4f951434ef1dec4b73998cd67d4aa3e86.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 293, + 370, + 476, + 378 + ], + "lines": [ + { + "bbox": [ + 293, + 370, + 476, + 378 + ], + "spans": [ + { + "bbox": [ + 293, + 370, + 476, + 378 + ], + "type": "text", + "content": "Table 9: Ablation on the VAE architecture." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 384, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 491 + ], + "type": "text", + "content": "criminalitor; and 2) VAE with vanilla discriminator (i.e., using the original 2D convolution in the discriminator). We evaluate the reconstruction performance of VAEs with 2D and 3D metrics, including PSNR, MAE, FRD and Chamfer Distance (CD) [16]. As listed in Table 9, the VAE with the proposed range-guided discriminator outperforms the baselines in both 2D and 3D reconstruction metrics. This gives credits to the ability of perceiving local 3D structures by the range-guided discriminator. The qualitative results presented in Figure 8 also indicate that the proposed range-guided discriminator ensures the point clouds are reconstructed with less noise and more precise object structures." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 509, + 226, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 226, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 226, + 521 + ], + "type": "text", + "content": "6 Conclusions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "We propose a novel RangeLDM model to generate realistic range-view LiDAR point clouds at a fast speed. We ensure the quality of projection from point clouds to range images with correct distribution via Hough Voting. Then we compress range images to latent features with a VAE, and train a diffusion model in the lower-dimensional latent space. Additionally, we enhance the range-image reconstruction quality of the VAE with a range-guided discriminator. Experiments conducted on KITTI-360 and nuScenes datasets demonstrate the superior generation quality and sampling efficiency of our method. In future, we will explore generating labeled point clouds and creating corner-case data, such as in car accidents and extreme weather conditions, for potential applications in robust self-driving." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 139, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 139, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 139, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 139, + 481, + 172 + ], + "type": "text", + "content": "1. Achlioptas, P., Diamanti, O., Mitlagkas, I., Guibas, L.: Learning representations and generative models for 3D point clouds. In: International Conference on Machine Learning. pp. 40-49. PMLR (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 481, + 216 + ], + "type": "text", + "content": "2. Ando, A., Gidaris, S., Bursuc, A., Puy, G., Boulch, A., Marlet, R.: RangeViT: Towards vision transformers for 3D semantic segmentation in autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5240-5250 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 216, + 481, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 216, + 481, + 249 + ], + "spans": [ + { + "bbox": [ + 138, + 216, + 481, + 249 + ], + "type": "text", + "content": "3. Bewley, A., Sun, P., Mensink, T., Anguelov, D., Sminchisescu, C.: Range conditioned dilated convolutions for scale invariant 3D object detection. In: Conference on Robot Learning (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 249, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 249, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 249, + 481, + 270 + ], + "type": "text", + "content": "4. Box, G.E., Jenkins, G.M., Reinsel, G.C., Ljung, G.M.: Time series analysis: forecasting and control. John Wiley & Sons (2015)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 271, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 481, + 304 + ], + "type": "text", + "content": "5. Caccia, L., Van Hoof, H., Courville, A., Pineau, J.: Deep generative modeling of LiDAR data. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 5034-5040. IEEE (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 304, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 481, + 347 + ], + "type": "text", + "content": "6. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: NuScenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11621-11631 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 347, + 481, + 380 + ], + "type": "text", + "content": "7. Cai, R., Yang, G., Averbuch-Elor, H., Hao, Z., Belongie, S., Snavely, N., Hariharan, B.: Learning gradient fields for shape generation. In: European Conference on Computer Vision. pp. 364-381 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 380, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 481, + 424 + ], + "type": "text", + "content": "8. Chai, Y., Sun, P., Ngiam, J., Wang, W., Caine, B., Vasudevan, V., Zhang, X., Anguelov, D.: To the point: Efficient 3D object detection in the range image with graph convolution kernels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16000-16009 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 424, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 424, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 424, + 481, + 456 + ], + "type": "text", + "content": "9. Chen, H., Luo, S., Hu, W., et al.: Deep point set resampling via gradient fields. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(3), 2913-2930 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 457, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 457, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 138, + 457, + 481, + 490 + ], + "type": "text", + "content": "0. Chen, X., Vizzo, I., Läbe, T., Behley, J., Stachniss, C.: Range image-based LiDAR localization for autonomous vehicles. In: IEEE International Conference on Robotics and Automation (ICRA). pp. 5802-5808. IEEE (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 491, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 481, + 522 + ], + "type": "text", + "content": "1. Cheng, H., Han, X., Xiao, G.: CeNet: Toward concise and efficient LiDAR semantic segmentation for autonomous driving. In: IEEE International Conference on Multimedia and Expo (ICME). pp. 01-06. IEEE (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 523, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 481, + 555 + ], + "type": "text", + "content": "2. Cortinhal, T., Tzelepis, G., Erdal Aksoy, E.: SalsaNext: Fast, uncertainty-aware semantic segmentation of LiDAR point clouds. In: Advances in Visual Computing. pp. 207-222. Springer International Publishing, Cham (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 555, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 481, + 588 + ], + "type": "text", + "content": "3. Deng, J., Shi, S., Li, P., Zhou, W., Zhang, Y., Li, H.: Voxel R-CNN: Towards high performance voxel-based 3D object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1201-1209 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 589, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 481, + 632 + ], + "type": "text", + "content": "4. Engelcke, M., Rao, D., Wang, D.Z., Tong, C.H., Posner, I.: Vote3Deep: Fast object detection in 3D point clouds using efficient convolutional neural networks. In: 2017 IEEE International Conference on Robotics and Automation (ICRA). pp. 1355-1361. IEEE (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "5. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12873-12883 (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 208, + 91, + 447, + 102 + ], + "type": "text", + "content": "RangeLDM: Fast Realistic LiDAR Point Cloud Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "16. Fan, H., Su, H., Guibas, L.J.: A point set generation network for 3D object reconstruction from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 605-613 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 194 + ], + "type": "text", + "content": "17. Fan, L., Pang, Z., Zhang, T., Wang, Y.X., Zhao, H., Wang, F., Wang, N., Zhang, Z.: Embracing single stride 3D object detector with sparse transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8458-8468 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "type": "text", + "content": "18. Fan, L., Xiong, X., Wang, F., Wang, N., Zhang, Z.: RangeDet: In defense of range view for LiDAR-based 3D object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2918-2927 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "19. Gao, L., Wu, T., Yuan, Y.J., Lin, M.X., Lai, Y.K., Zhang, H.: TM-NET: Deep generative networks for textured meshes. ACM Transactions on Graphics (TOG) 40(6), 1-15 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "text", + "content": "20. Gao, L., Yang, J., Wu, T., Yuan, Y.J., Fu, H., Lai, Y.K., Zhang, H.: SDM-NET: Deep generative network for structured deformable mesh. ACM Transactions on Graphics (TOG) 38(6), 1-15 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "21. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 358 + ], + "type": "text", + "content": "22. He, C., Li, R., Li, S., Zhang, L.: Voxel set transformer: A set-to-set approach to 3D object detection from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8417-8427 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 402 + ], + "type": "text", + "content": "23. He, Y., Tang, D., Zhang, Y., Xue, X., Fu, Y.: Grad-PU: Arbitrary-scale point cloud upsampling via gradient descent with learned distance functions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5354-5363 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 402, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 480, + 456 + ], + "type": "text", + "content": "24. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: GANs trained by a two time-scale update rule converge to a local nash equilibrium. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "text", + "content": "25. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33, 6840-6851 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "text", + "content": "26. Hu, J.S., Kuai, T., Waslander, S.L.: Point density-aware voxels for LiDAR 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8469-8478 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "27. Hu, P., Ziglar, J., Held, D., Ramanan, D.: What you see is what you get: Exploiting visibility for 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11001-11009 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "type": "text", + "content": "28. Hui, K.H., Li, R., Hu, J., Fu, C.W.: Neural wavelet-domain diffusion for 3D shape generation. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "text", + "content": "29. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1125-1134 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "text", + "content": "30. Kim, H., Lee, H., Kang, W.H., Lee, J.Y., Kim, N.S.: SoftFlow: Probabilistic framework for normalizing flow on manifolds. Advances in Neural Information Processing Systems 33, 16388-16397 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "31. Kim, J., Yoo, J., Lee, J., Hong, S.: SetVAE: Learning hierarchical composition for generative modeling of set-structured data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15059–15068 (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "text", + "content": "Q. Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "32. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 481, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 481, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 481, + 159 + ], + "type": "text", + "content": "33. Kochanov, D., Nejadasl, F.K., Booij, O.: KprNet: Improving projection-based LiDAR semantic segmentation. arXiv preprint arXiv:2007.12668 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 481, + 192 + ], + "type": "text", + "content": "34. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: PointPillars: Fast encoders for object detection from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12697-12705 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 192, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 192, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 192, + 481, + 224 + ], + "type": "text", + "content": "35. Li, B.: 3D fully convolutional network for vehicle detection in point cloud. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1513-1518. IEEE (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 224, + 481, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 481, + 245 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 481, + 245 + ], + "type": "text", + "content": "36. Li, B., Zhang, T., Xia, T.: Vehicle detection from 3D LiDAR using fully convolutional network. arXiv preprint arXiv:1608.07916 (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 246, + 481, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 481, + 277 + ], + "type": "text", + "content": "37. Li, C., Ren, Y., Liu, B.: Pcgen: Point cloud generator for lidar simulation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 11676-11682. IEEE (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 277, + 481, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 481, + 310 + ], + "type": "text", + "content": "38. Liao, Y., Xie, J., Geiger, A.: KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2D and 3D. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(3), 3292-3310 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 310, + 458, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 458, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 458, + 320 + ], + "type": "text", + "content": "39. Lim, J.H., Ye, J.C.: Geometric gan. arXiv preprint arXiv:1705.02894 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 321, + 481, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 481, + 353 + ], + "type": "text", + "content": "40. Litany, O., Bronstein, A., Bronstein, M., Makadia, A.: Deformable shape completion with graph convolutional autoencoders. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1886-1895 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 353, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 481, + 385 + ], + "type": "text", + "content": "41. Luo, S., Hu, W.: Diffusion probabilistic models for 3D point cloud generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2837-2845 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 385, + 481, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 481, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 481, + 417 + ], + "type": "text", + "content": "42. Malavazi, F.B., Guyonneau, R., Fasquel, J.B., Lagrange, S., Mercier, F.: LiDAR-only based navigation algorithm for an autonomous agricultural robot. Computers and Electronics in Agriculture 154, 71-79 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 418, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 481, + 460 + ], + "type": "text", + "content": "43. Manivasagam, S., Wang, S., Wong, K., Zeng, W., Sazanovich, M., Tan, S., Yang, B., Ma, W.C., Urtasun, R.: LiDARsim: Realistic LiDAR simulation by leveraging the real world. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11167-11176 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 460, + 481, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 481, + 503 + ], + "type": "text", + "content": "44. Meyer, G.P., Laddha, A., Kee, E., Vallespi-Gonzalez, C., Wellington, C.K.: Laser-Net: An efficient probabilistic 3D object detector for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12677-12686 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 503, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 481, + 536 + ], + "type": "text", + "content": "45. Milioto, A., Vizzo, I., Behley, J., Stachniss, C.: RangeNet++: Fast and accurate LiDAR semantic segmentation. In: IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 4213-4220. IEEE (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 536, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 481, + 567 + ], + "type": "text", + "content": "46. Mittal, P., Cheng, Y.C., Singh, M., Tulsiani, S.: AutoSDF: Shape priors for 3D completion, reconstruction and generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 306-315 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 568, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 481, + 600 + ], + "type": "text", + "content": "47. Mo, K., Guerrero, P., Yi, L., Su, H., Wonka, P., Mitra, N., Guibas, L.J.: StructureNet: Hierarchical graph networks for 3D shape generation. arXiv preprint arXiv:1908.00575 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 601, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 601, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 601, + 481, + 632 + ], + "type": "text", + "content": "48. Nichol, A., Jun, H., Dhariwal, P., Mishkin, P., Chen, M.: Point-E: A system for generating 3D point clouds from complex prompts. arXiv preprint arXiv:2212.08751 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 481, + 665 + ], + "type": "text", + "content": "49. Pan, X., Xia, Z., Song, S., Li, L.E., Huang, G.: 3D object detection with pointformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7463-7472 (2021)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "type": "text", + "content": "RangeLDM: Fast Realistic LiDAR Point Cloud Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 150 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 150 + ], + "type": "text", + "content": "50. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: PointNet: Deep learning on point sets for 3D classification and segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 652-660 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "type": "text", + "content": "51. Ran, H., Guizilini, V., Wang, Y.: Towards realistic scene generation with lidar diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "type": "text", + "content": "52. Resop, J.P., Lehmann, L., Hession, W.C.: Drone laser scanning for modeling river-scape topography and vegetation: Comparison with traditional aerial LiDAR. Drones 3(2), 35 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 217, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 217, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 481, + 239 + ], + "type": "text", + "content": "53. Rezende, D., Mohamed, S.: Variational inference with normalizing flows. In: International Conference on Machine Learning. pp. 1530-1538. PMLR (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 239, + 481, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 239, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 239, + 481, + 274 + ], + "type": "text", + "content": "54. Rezende, D.J., Mohamed, S., Wierstra, D.: Stochastic backpropagation and approximate inference in deep generative models. In: International Conference on Machine Learning. pp. 1278-1286. PMLR (2014)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "type": "text", + "content": "55. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "type": "text", + "content": "56. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention. pp. 234-241. Springer (2015)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 341, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 481, + 373 + ], + "type": "text", + "content": "57. Sallab, A.E., Sobh, I., Zahran, M., Essam, N.: LiDAR sensor modeling and data augmentation with GANs for autonomous driving. arXiv preprint arXiv:1905.07290 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 373, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 373, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 373, + 481, + 397 + ], + "type": "text", + "content": "58. Sauer, A., Chitta, K., Müller, J., Geiger, A.: Projected GANs converge faster. Advances in Neural Information Processing Systems 34, 17480-17492 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 397, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 481, + 430 + ], + "type": "text", + "content": "59. Schubert, S., Neubert, P., Pöschmann, J., Protzel, P.: Circular convolutional neural networks for panoramic images and laser data. In: IEEE Intelligent Vehicles Symposium (IV). pp. 653-660. IEEE (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 430, + 481, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 481, + 464 + ], + "type": "text", + "content": "60. Shi, G., Li, R., Ma, C.: PillarNet: Real-time and high-performance pillar-based 3D object detection. In: European Conference on Computer Vision. pp. 35-52. Springer (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 464, + 481, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 481, + 508 + ], + "type": "text", + "content": "61. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: PV-RCNN: Point-voxel feature set abstraction for 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10529-10538 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 508, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 508, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 508, + 481, + 542 + ], + "type": "text", + "content": "62. Shi, S., Wang, X., Li, H.: PointRCNN: 3D object proposal generation and detection from point cloud. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 770-779 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 542, + 481, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 542, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 132, + 542, + 481, + 576 + ], + "type": "text", + "content": "63. Shi, S., Wang, Z., Shi, J., Wang, X., Li, H.: From points to parts: 3D object detection from point cloud with part-aware and part-aggregation network. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(8), 2647-2664 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 576, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 481, + 609 + ], + "type": "text", + "content": "64. Shi, W., Rajkumar, R.: Point-GNN: Graph neural network for 3D object detection in a point cloud. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1711-1719 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 609, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 481, + 643 + ], + "type": "text", + "content": "65. Shu, D.W., Park, S.W., Kwon, J.: 3D point cloud generative adversarial network based on tree structured graph convolutions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3859-3868 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "66. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "text", + "content": "Q. Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "67. Sun, P., Tan, M., Wang, W., Liu, C., Xia, F., Leng, Z., Anguelov, D.: SWFormer: Sparse window transformer for 3D object detection in point clouds. In: European Conference on Computer Vision. pp. 426-442. Springer (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "type": "text", + "content": "68. Sun, P., Wang, W., Chai, Y., Elsayed, G., Bewley, A., Zhang, X., Sminchisescu, C., Anguelov, D.: RSN: Range sparse net for efficient, accurate LiDAR 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5725-5734 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 237 + ], + "type": "text", + "content": "69. Sun, Y., Wang, Y., Liu, Z., Siegel, J., Sarma, S.: PointGrow: Autoregressively learned point cloud generation with self-attention. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 61-70 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "type": "text", + "content": "70. Tan, Q., Gao, L., Lai, Y.K., Xia, S.: Variational autoencoders for deforming 3D mesh models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5841-5850 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "type": "text", + "content": "71. Tian, Z., Chu, X., Wang, X., Wei, X., Shen, C.: Fully convolutional one-stage 3D object detection on LiDAR range images. Advances in Neural Information Processing Systems 35, 34899-34911 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 304, + 482, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 326 + ], + "type": "text", + "content": "72. Vahdat, A., Kreis, K., Kautz, J.: Score-based generative modeling in latent space. Advances in Neural Information Processing Systems 34, 11287-11302 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 326, + 482, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 482, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 482, + 358 + ], + "type": "text", + "content": "73. Valsesia, D., Fracastoro, G., Magli, E.: Learning localized generative models for 3D point clouds via graph convolution. In: International Conference on Learning Representations (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 358, + 482, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 380 + ], + "type": "text", + "content": "74. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in Neural Information Processing Systems 30 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "type": "text", + "content": "75. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in Neural Information Processing Systems 30 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 413, + 482, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 456 + ], + "type": "text", + "content": "76. Wang, H., Shi, S., Yang, Z., Fang, R., Qian, Q., Li, H., Schiele, B., Wang, L.: RBGNet: Ray-based grouping for 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1110-1119 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 456, + 482, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 491 + ], + "type": "text", + "content": "77. Wang, Y., Fathi, A., Kundu, A., Ross, D.A., Pantofaru, C., Funkhouser, T., Solomon, J.: Pillar-based object detection for autonomous driving. In: European Conference on Computer Vision. pp. 18-34. Springer (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 491, + 482, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 512 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 512 + ], + "type": "text", + "content": "78. Weiss, U., Biber, P.: Plant detection and mapping for agricultural robots using a 3D LiDAR sensor. Robotics and Autonomous Systems 59(5), 265-273 (2011)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 512, + 482, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 482, + 545 + ], + "type": "text", + "content": "79. Xiong, Y., Ma, W.C., Wang, J., Urtasun, R.: Learning compact representations for LiDAR completion and generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1074-1083 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 545, + 482, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 578 + ], + "type": "text", + "content": "80. Xu, C., Wu, B., Wang, Z., Zhan, W., Vajda, P., Keutzer, K., Tomizuka, M.: SqueezeSegV3: Spatially-adaptive convolution for efficient point-cloud segmentation. In: European Conference on Computer Vision. pp. 1-19. Springer (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 578, + 482, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 578, + 482, + 600 + ], + "spans": [ + { + "bbox": [ + 130, + 578, + 482, + 600 + ], + "type": "text", + "content": "81. Yan Yan, Y.M., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 600, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 600, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 600, + 482, + 632 + ], + "type": "text", + "content": "82. Yang, B., Luo, W., Urtasun, R.: Pixor: Real-time 3D object detection from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7652-7660 (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "text", + "content": "83. Yang, G., Huang, X., Hao, Z., Liu, M.Y., Belongie, S., Hariharan, B.: PointFlow: 3D point cloud generation with continuous normalizing flows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4541-4550 (2019)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "type": "text", + "content": "RangeLDM: Fast Realistic LiDAR Point Cloud Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 501 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "84. Yang, Z., Sun, Y., Liu, S., Jia, J.: 3DSSD: Point-based 3D single stage object detector. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11040-11048 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "text", + "content": "85. Yang, Z., Sun, Y., Liu, S., Shen, X., Jia, J.: STD: Sparse-to-dense 3D object detector for point cloud. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1951-1960 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "text", + "content": "86. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3D object detection and tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11784-11793 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "text", + "content": "87. Yu, J., Li, X., Koh, J.Y., Zhang, H., Pang, R., Qin, J., Ku, A., Xu, Y., Baldridge, J., Wu, Y.: Vector-quantized image modeling with improved VQGAN. arXiv preprint arXiv:2110.04627 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "text", + "content": "88. Yu, L., Li, X., Fu, C.W., Cohen-Or, D., Heng, P.A.: Pu-Net: Point cloud upsampling network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2790-2799 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 314 + ], + "type": "text", + "content": "89. Zeng, X., Vahdat, A., Williams, F., Gojcic, Z., Litany, O., Fidler, S., Kreis, K.: LION: Latent point diffusion models for 3D shape generation. Advances in Neural Information Processing Systems 35, 10021-10039 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "text", + "content": "90. Zhang, J., Singh, S.: LOAM: LiDAR odometry and mapping in real-time. In: Robotics: Science and Systems. vol. 2, pp. 1-9. Berkeley, CA (2014)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 337, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 337, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 337, + 480, + 380 + ], + "type": "text", + "content": "91. Zhang, Y., Hu, Q., Xu, G., Ma, Y., Wan, J., Guo, Y.: Not all points are equal: Learning highly efficient point-based detectors for 3D LiDAR point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18953-18962 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "92. Zhao, Y., Bai, L., Huang, X.: FidNet: LiDAR point cloud semantic segmentation with fully interpolation decoding. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 4453-4458. IEEE (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "type": "text", + "content": "93. Zhou, L., Du, Y., Wu, J.: 3D shape generation and completion through point-voxel diffusion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5826-5835 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "type": "text", + "content": "94. Zhou, Y., Tuzel, O.: VoxelNet: End-to-end learning for point cloud based 3D object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4490-4499 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "text", + "content": "95. Zyrianov, V., Zhu, X., Wang, S.: Learning to generate realistic LiDAR point clouds. In: European Conference on Computer Vision. pp. 17-35. Springer (2022)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "text", + "content": "Q. Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Rasterized Edge Gradients_ Handling Discontinuities Differentially/ac83eb0e-ca7d-4471-8db5-05179e149303_content_list.json b/2024/Rasterized Edge Gradients_ Handling Discontinuities Differentially/ac83eb0e-ca7d-4471-8db5-05179e149303_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..341880ca93f7f81f1bcd2b878cd9aa7c7d7ad023 --- /dev/null +++ b/2024/Rasterized Edge Gradients_ Handling Discontinuities Differentially/ac83eb0e-ca7d-4471-8db5-05179e149303_content_list.json @@ -0,0 +1,2074 @@ +[ + { + "type": "text", + "text": "Rasterized Edge Gradients: Handling Discontinuities Differently", + "text_level": 1, + "bbox": [ + 289, + 140, + 712, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Stanislav Pidhorskyi $^{\\text{①}}$ , Tomas Simon $^{\\text{②}}$ , Gabriel Schwartz $^{\\text{③}}$ , He Wen, Yaser Sheikh, and Jason Saragih", + "bbox": [ + 233, + 212, + 767, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reality Labs,Meta,Pittsburgh, Pennsylvania,USA", + "bbox": [ + 330, + 255, + 671, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Computing the gradients of a rendering process is paramount for diverse applications in computer vision and graphics. However, accurate computation of these gradients is challenging due to discontinuities and rendering approximations, particularly for surface-based representations and rasterization-based rendering. We present a novel method for computing gradients at visibility discontinuities for rasterization-based differentiable renderers. Our method elegantly simplifies the traditionally complex problem through a carefully designed approximation strategy, allowing for a straightforward, effective, and performant solution. We introduce a novel concept of micro-edges, which allows us to treat the rasterized images as outcomes of a differentiable, continuous process aligned with the inherently non-differentiable, discrete-pixel rasterization. This technique eliminates the necessity for rendering approximations or other modifications to the forward pass, preserving the integrity of the rendered image, which makes it applicable to rasterized masks, depth, and normals images where filtering is prohibitive. Utilizing micro-edges simplifies gradient interpretation at discontinuities and enables handling of geometry intersections, offering an advantage over the prior art. We showcase our method in dynamic human head scene reconstruction, demonstrating effective handling of camera images and segmentation masks.", + "bbox": [ + 263, + 306, + 738, + 583 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 611, + 374, + 626 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Significant advances have been made in recent years in modeling real 3D objects from image measurements. Much of this advancement can be credited to improvements in inverse rendering, which enables the automatic inference of 3D scene parameters that best reconstruct the images. While volumetric methods like NeRF [31] simplify inverse rendering by eliminating the need for predefined scene topology, classical mesh-based representations remain widely used due to their efficiency in modeling opaque surfaces. These representations often rely on highly performant rasterization, which requires discrete operations to reason about ordering and coverage, which poses challenges for gradient computations. Our main contribution is a theoretical framework for approximating gradients of a rasterization process that simultaneously elucidates previous constructions, arrives at a simple, fast, and accurate formulation, and further improves gradient approximation accuracy over state-of-the-art methods.", + "bbox": [ + 217, + 643, + 785, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In rasterization, a rendered pixel's value is determined by the foremost triangle covering it. Pixels whose footprint lies entirely within a single triangle are straightforward to handle, as infinitesimal mesh motion does not change their triangle membership. However, pixels on triangle boundaries, either between adjacent triangles, at occlusion boundaries, or at intersections between triangles, must account for how their triangle membership changes in response to infinitesimal changes in the mesh. Existing methods employ different approximations of soft-membership to account for this, enabling gradient computation in those areas, effectively compositing contributions from member triangles via a weighted sum. For example, [20] uses the anti-aliasing approach, [14] averages montecarlo samples during ray-casting, and [40] extends the influence of boundary triangles using a falloff function. In this work, we show that all these methods exhibit approximation errors that can fail to compute correct gradients in one case or another. In particular, none of the existing methods accurately compute the gradient at triangle intersections, which can often happen during inverse rendering, especially when the current estimate is far from the desired solution.", + "bbox": [ + 212, + 146, + 787, + 387 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast to existing works that directly tackle geometric discontinuities and rasterization's discrete nature, we introduce a micro-edge formulation which allows to interpret the rasterized image as an outcome of a continuous process that coincidentally aligns with discrete-pixel rasterization, simplifying gradient computation significantly. Unlike some other works [20, 40], our formulation achieves good gradient estimates without altering the rasterization forward pass, maintaining the rasterized image's integrity. This is crucial for optimizing segmentation masks, depth maps, and normal maps where filtering or smoothing is not feasible, e.g. anisotropic filtering or soft rasterization would mix normals from different surfaces misrepresenting the geometry. The simplifications offered by our micro-edge formulation allow us to seamlessly handle discontinuities caused by geometry intersections, offering an advantage over the prior art.", + "bbox": [ + 212, + 388, + 787, + 569 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our paper introduces a straightforward, accurate, and efficient method for computing gradients in rasterization, comparable to existing techniques but with greater simplicity and the ability to handle self-intersections. We analyze our method's errors from a theoretical standpoint and compare gradients qualitatively and quantitatively with finite differences and other methods across various test cases. We also assess runtime efficiency and accuracy by image size and showcase qualitative and quantitative findings on a synthetic blender dataset [32]. Our method excels in complex applications, such as detailed dynamic human head reconstructions, effectively managing the intricate details of the inner mouth region with significant occlusions and deformations.", + "bbox": [ + 212, + 570, + 787, + 720 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 215, + 746, + 382, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Differentiable rendering is one of the critical building blocks for many existing computer vision problems, such as 3d object reconstruction [10, 17, 45, 50, 51], 3d object prediction [5, 8, 9, 12], pose estimation [1, 6, 11, 38] novel view synthesis [4, 18, 32, 33, 54], as well as newly emerging applications such as text-to-3D", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "S. Pidhorskyi et al.", + "bbox": [ + 271, + 114, + 400, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "generative models [22, 39, 43, 44]. Differentiable rendering employs a wide array of underlying 3D representations including explicit surfaces like mesh-based representations [19, 20, 27], implicit surfaces [2, 16, 37, 46], point clouds [18, 42, 48], explicit or implicit volume representations [4, 23, 26, 32, 35, 50, 52], and hybrid representations [7, 33]. For a comprehensive introduction to differentiable rendering, we refer the reader to Zhao et al.'s excellent SIGGRAPH course notes [55].", + "bbox": [ + 212, + 146, + 782, + 238 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mesh optimization often poses more challenges than implicit and volumetric methods, leading to the latter's prevalence in the field, as noted by Roessle et al. [41]. Despite this, meshes excel in high-performance rendering and imply registration with predefined topologies. Mesh rendering methods can be categorized into ray tracing-based and rasterization-based, with the former including frameworks like Mitsuba 3 [15, 36]. Differentiable ray tracing, aimed at direct illumination or path tracing, is computationally expensive but principled. Key advancements by Li et al. [21] and Zhang et al. [53] split this gradient into discontinuity integrals with Dirac delta functions and differentiable continuous parts. Li et al. replaced Dirac delta integration with boundary line integration, while Zhang et al. applied Reynold's transport theorem for a similar decomposition. Both methods are intricate, necessitating silhouette edge sampling — a significant bottleneck in their application. Loubet et al. [28] proposed a variable change to simplify discontinuity integration, further refined by Bangaru et al. [3] using warped area fields to correct biases, yet practical application challenges persist. Despite their flexibility, the computational expense of such ray tracing frameworks restrict their practicality in optimization-intensive tasks [20].", + "bbox": [ + 212, + 241, + 787, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For rasterization-based methods, rendering meshes differentiably is hindered by non-differentiable visibility at boundaries. Unlike ray tracing, which uses boundary integrals for visibility gradients, rasterization's fixed-grid and z-buffer approach lacks this feature. Solutions typically involve approximating derivatives or approximating rendering to achieve inherent differentiability. De La Gorce et al. [19] and Loper et al. [27] pioneered derivative approximation in rasterization. De La Gorce et al. distinguished between continuous regions and occlusion boundaries, introducing 'occlusion forces' for the latter. Loper et al. simplified this by detecting discontinuities post-rasterization and employing differential filters, like the Sobel filter, for boundary approximation.", + "bbox": [ + 212, + 503, + 787, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The 'Nvdiffrast' approach by Laine et al. [20] tackles point-sampled visibility issues with innovative differentiable analytic antialIASing, which transforms sharp discontinuities into smooth transitions, enabling gradient computation. Analytical antialIASing is achieved by estimating coverage using silhouette edges. A limitation of the method is that if triangles containing silhouette edges do not overlap with any pixel center, they will be overlooked during the antialIASing process. Moreover, analytical antialIASing is inherently approximate, so the obtained gradients are also approximate. Regrettably, the approach is still quite complex, modifies the rasterized image, and requires a specialized data structure for connectivity and detailed pixel coverage computations. On the other hand, methods like Rhodin et al. [40] and Liu et al.'s 'Soft Rasterizer' [24] opt for modifying the rendering model. Rhodin et al. introduce fuzzy edges to soften discontinuities,", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Rasterized Edge Gradients: Handling Discontinuities Differently", + "bbox": [ + 285, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "while 'Soft Rasterizer' further blurs boundaries and averages depth contributions, facilitating gradient propagation across occluded primitives.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we deliberately bypass physically based rendering, global illumination, and lighting/material models, and instead focus exclusively on rasterization-based approaches for their speed. We draw inspiration from the core idea of the edge sampling method [21] and analytical antialIASing [20] and distill this into a much simpler yet effective method that also provides accurate gradients. Notably, our method's simplicity allows direct handling of self-penetrating geometry, an aspect not addressed in previous research. Other methods to handle self-penetrating geometry would require complex and computationally intensive geometry preprocessing at each optimization step, involving intersection detection and differentiable splitting of the faces. In contrast, our approach explicitly manages interpenetrating geometry with virtually no additional overhead.", + "bbox": [ + 212, + 176, + 787, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 361, + 382, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following Li et al. [21], given a 2D pixel filter $k$ and radiance $R$ , a pixel's color can be written:", + "bbox": [ + 212, + 387, + 782, + 415 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI = \\iint_ {D} k (x, y) R (x, y) d x d y. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 415, + 784, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For notational convenience, as in [21], we denote $f(x,y) = k(x,y)R(x,y)$ . For simplicity, let us first consider the case of a single pixel. We are interested in the gradient of $\\frac{\\partial L(I)}{\\partial\\varPhi}$ , where the scalar function $L(I)$ of the rendered pixel $I$ defines our loss, and $\\varPhi$ are the scene parameters that we aim to optimize. According to the chain rule:", + "bbox": [ + 212, + 446, + 787, + 521 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial L (I)}{\\partial \\Phi} = \\frac {\\partial I}{\\partial \\Phi} \\frac {\\partial L}{\\partial I}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 442, + 521, + 784, + 550 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\frac{\\partial L}{\\partial I}$ is the incoming gradient from the loss function and is typically computed by automatic differentiation (AD). So, our goal becomes to find how the pixel value changes with respect to the scene parameters:", + "bbox": [ + 212, + 551, + 782, + 599 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial I}{\\partial \\Phi} = \\frac {\\partial}{\\partial \\Phi} \\iint_ {D} f (x, y) d x d y. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 604, + 784, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the function $f$ may not in fact be differentiable, its integral remains continuous and thus differentiable. Let us assume that $f$ is partitioned into two continuous half-spaces, each represented by two continuous and differentiable functions $f_{a}$ and $f_{b}$ , respectively, as follows:", + "bbox": [ + 212, + 638, + 787, + 700 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf (x, y) = \\theta (\\alpha (x, y)) f _ {a} (x, y) + \\theta (- \\alpha (x, y)) f _ {b} (x, y), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 707, + 784, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ specifies the dividing edge, and $\\theta$ is the Heaviside step function, which selects between $f_{a}$ and $f_{b}$ . The differentiation of the integral can be broken into two parts by applying the product rule. Let us start with the integrand $f_{a}$ :", + "bbox": [ + 212, + 727, + 782, + 773 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\frac {\\partial}{\\partial \\Phi} \\iint_ {D} \\theta (\\alpha (x, y)) f _ {a} (x, y) d x d y \\\\ = \\iint_ {D} \\delta (\\alpha (x, y)) \\frac {\\partial \\alpha (x , y)}{\\partial \\Phi} f _ {a} (x, y) d x d y + \\iint_ {D} \\frac {\\partial f _ {a} (x , y)}{\\partial \\Phi} \\theta (\\alpha (x, y)) d x d y, \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 235, + 777, + 784, + 843 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "S. Pidhorskyi et al.", + "bbox": [ + 271, + 114, + 400, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and a similar derivation can be done for $f_{b}$ . The second term on the right-hand side of Eqn. (5) is the integral of the derivative of the smooth function, $f_{a}(x,y)$ . This derivative is readily computable with AD, and in the context of differentiable rasterizers it typically involves differentiating through the interpolation step, the barycentric coordinates, and the texture sampler. Due to its straightforward computation, we will represent it as $\\Omega$ and omit it from further discussion, focusing instead on the first term on the right-hand side of Eqn. (5). This term involves a Dirac delta function, and can be interpreted as taking the integral over the line $\\alpha(x,y) = 0$ :", + "bbox": [ + 212, + 146, + 782, + 282 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\iint_ {D} \\delta (\\alpha (x, y)) \\frac {\\partial \\alpha (x , y)}{\\partial \\Phi} f _ {a} (x, y) d x d y \\tag {6} \\\\ = \\int_ {\\alpha (x, y) = 0} \\frac {\\partial \\alpha (x , y)}{\\partial \\Phi} \\| \\nabla_ {x, y} \\alpha (x, y) \\| ^ {- 1} f _ {a} (x, y) d t, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 292, + 784, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\| \\nabla_{x,y}\\alpha (x,y)\\|$ is $L^2$ length of the edge equation gradient which accounts for the Jacobian of the variable substitution, and $t$ is the parametrization of the line along which we integrate.", + "bbox": [ + 212, + 367, + 782, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Method", + "text_level": 1, + "bbox": [ + 215, + 433, + 328, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section we describe our method that we call EdgeGrad. First we will focus on the core assumptions and approximations that form the basis of our method. Then we will see that due to our specifically chosen approximation, we can apply the prior method of edge sampling to rasterization. From there, we will focus on edge detection and classification heuristics, and our discretization scheme. We will conclude by describing an extension of our method for handling geometry intersections.", + "bbox": [ + 212, + 462, + 784, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Rasterizing Edges", + "text_level": 1, + "bbox": [ + 215, + 589, + 410, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our primary objective is to differentiate rasterization while preserving its core image formation principles. We aim to calculate the gradients of pixel values relative to the mesh vertex positions, including gradients caused by visibility changes such as occlusion and disocclusion by primitive geometry. We seek to do this with a simple, minimal, and efficient method that does not necessitate modifying the initially rasterized image.", + "bbox": [ + 212, + 611, + 782, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our main insight is that instead of approximating rasterization using a continuous representation, we assume the rasterized image is the result of some other continuous process that coincidentally aligns with our inherently non-differentiable, discrete-pixel rasterization. In essence, we opt for a direct ap", + "bbox": [ + 212, + 702, + 782, + 762 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Rasterized Edge Gradients: Handling Discontinuities Differently", + "bbox": [ + 285, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1 This concept is akin to representing real numbers with floating-point arithmetic. Although there are only 4,294,967,296 single-precision floating-point numbers (a quarter of which lie in the interval [0,1]), they are still employed for continuous computation. We treat these computations as smooth and differentiable, even though they are, in essence, discontinuous.", + "bbox": [ + 217, + 768, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e6b735ec17688fcb4c5fce6d5226cc5ddac903e5c482938d0715501810ca7f43.jpg", + "image_caption": [ + "Fig.1: Micro-edges and Pixel Pair." + ], + "image_footnote": [ + "The figure illustrates construction of the micro-edges and the change in pixel intensities A and B with the edge position $p$ movement, highlighting the C1 discontinuity at $p = 0$ , absent in average intensity." + ], + "bbox": [ + 230, + 142, + 483, + 421 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1797519b671dc6da1b936dd2c896efe585d6957e8332a1c48df4d63e76ed2846.jpg", + "image_caption": [ + "Fig. 2: Gradient computation flow." + ], + "image_footnote": [ + "The figure illustrates the backward propagation process, where gradients from the loss function are traced back and scattered to the fragments." + ], + "bbox": [ + 509, + 143, + 781, + 433 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "proximation to compute gradients, instead of an indirect approach that first necessitates that the rendering be made continuous, for instance, through antialiasing filtering. More specifically, we use a construction where all straight edges, whether they are sides of triangles or intersections between triangles, are comprised of a collection of micro-edges. These micro-edges are assumed to be either strictly vertical or horizontal and are always positioned precisely between two pixels. An illustration of these constructions is shown in Figure 1.", + "bbox": [ + 212, + 549, + 787, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our approximation leads to several key properties: a) boundaries never intersect pixels, only lying exactly between them; b) pixels are always either fully covered or not covered at all; c) there is at most one boundary between any two neighbouring pixels; d) we do not need to access the source geometry to evaluate where the edge is exactly. With this, we can reinterpret the problem as if the rasterized image was in fact obtained by applying an antialIASing filter, albeit without visible effects; the resulting image remains identical to the pre-antialIASing source. Essentially our construct of the micro-edges renders the antialIASing filter as an identity operation.", + "bbox": [ + 212, + 657, + 787, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With this approximation, gradient estimation becomes a local operation, requiring the analysis one edge between two pixels at a time. Note that despite the approximation that all edges reside strictly between pixels, we do not assume", + "bbox": [ + 212, + 795, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "S. Pidhorskyi et al.", + "bbox": [ + 271, + 114, + 400, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "that they are fixed, but rather allow them to freely move within an infinitesimal range, enabling differentiation of pixel values with respect to the edge location. This leads to the greatest advantage of our formulation; it allows us to greatly simplify both the theory and computation, since all edges are axis-aligned.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Pixel Pairs", + "text_level": 1, + "bbox": [ + 215, + 228, + 352, + 242 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Let's analyze two adjacent pixels, A and B, separated by an edge, with their intensities represented as $I_{A}$ and $I_{B}$ respectively (refer to Figure 1 for details). The pixel intensities can be expressed using the Heaviside step function to account for the discontinuity at the edge as in Eqn. (4):", + "bbox": [ + 212, + 252, + 782, + 313 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} I _ {A} = \\iint_ {D _ {A}} \\theta (\\alpha (x, y)) f _ {a} (x, y) + \\theta (- \\alpha (x, y)) f _ {b} (x, y) d x d y \\\\ I _ {B} = \\iint_ {D _ {B}} \\theta (- \\alpha (x, y)) f _ {a} (x, y) + \\theta (\\alpha (x, y)) f _ {b} (x, y) d x d y. \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 299, + 324, + 784, + 390 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Without loss of generality, assuming a horizontal pair we define $\\alpha(x,y) = p_{AB} - x$ , where $p_{AB}$ is the edge location. Edge location $p_{AB}$ is now the scene parameter with respect to which we want to differentiate the loss function. Since $p_{AB}$ effects both $I_A$ and $I_B$ :", + "bbox": [ + 212, + 400, + 782, + 459 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial L}{\\partial p _ {A B}} = \\frac {\\partial L}{\\partial I _ {A}} \\frac {\\partial I _ {A}}{\\partial p _ {A B}} + \\frac {\\partial L}{\\partial I _ {B}} \\frac {\\partial I _ {B}}{\\partial p _ {A B}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 459, + 782, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\frac{\\partial I_A}{\\partial p_{AB}}$ and $\\frac{\\partial I_B}{\\partial p_{AB}}$ derive by applying Eqn. (6) to Eqn. (7). Unfortunately, with the edge strictly on the pixel boundary, we face a $C1$ discontinuity at $p_{AB} = 0$ , making intensity values non-differentiable. Infinitesimal movement of the edge affects $I_B$ when moving right and $I_A$ when moving left, without affecting the other. We still can compute one-sided limits of the derivatives, e.g. $\\frac{\\partial I_A}{\\partial p}^{-}$ from the negative direction and $\\frac{\\partial I_B}{\\partial p}^{+}$ from the positive direction. Applying Eqn. (6) and taking into account that the term $\\| \\nabla_{x,y}\\alpha (x,y)\\|$ becomes 1 as we obtain:", + "bbox": [ + 212, + 494, + 784, + 612 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial I _ {B}}{\\partial p _ {A B}} ^ {-} = \\frac {\\partial I _ {A}}{\\partial p _ {A B}} ^ {+} = 0; \\frac {\\partial I _ {B}}{\\partial p _ {A B}} ^ {+} = \\frac {\\partial I _ {A}}{\\partial p _ {A B}} ^ {-} = \\int_ {x = 0} [ f _ {a} (x, y) - f _ {b} (x, y) ] d y \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 235, + 623, + 782, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We can circumvent the problem of $C_1$ discontinuity by averaging the one-sided limits of the derivatives from the left and right:", + "bbox": [ + 214, + 665, + 782, + 695 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial I _ {B}}{\\partial p _ {A B}} = \\frac {\\partial I _ {A}}{\\partial p _ {A B}} = \\frac {1}{2} \\int_ {x = 0} [ f _ {a} (x, y) - f _ {b} (x, y) ] d y \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 705, + 782, + 737 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Assuming that the pixels have constant value across all of their area (recall the rule that \"pixels are always either fully covered or not covered at all\") and substituting Eqn. (10) into Eqn. (8) we deduce:", + "bbox": [ + 212, + 746, + 782, + 792 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\left. \\frac {\\partial L}{\\partial p _ {A B}} = \\frac {1}{2} \\left(\\frac {\\partial L}{\\partial I _ {A}} + \\frac {\\partial L}{\\partial I _ {B}}\\right) \\left(I _ {A} - I _ {B}\\right) + \\Omega \\right], \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 804, + 782, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Rasterized Edge Gradients: Handling Discontinuities Differently", + "bbox": [ + 285, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "which is one of our main results. Here, $\\frac{\\partial L}{\\partial p_{AB}}$ is the gradient of the loss function with respect to the position of the edge between pixels $A$ and $B$ . Additionally, we reintroduce $\\Omega$ which is the second term on the right-hand side of Eqn. (5) that accounts for differentiating smooth regions of $f$ . This term was omitted in Eqn.(9) and Eqn. (10) for brevity.", + "bbox": [ + 212, + 145, + 782, + 222 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Now, we need to accumulate these gradients into the gradients of the vertex positions of the triangle that created those edges. We do that first by scattering the gradients $\\frac{\\partial L}{\\partial p_{AB}}$ to the fragments of the triangles associated with the pixels forming the edge and then by invoking the backward function of the interpolator module to gather the fragment gradients into the vertex gradients. Next, we will take a look at the gradient scattering mechanism.", + "bbox": [ + 212, + 222, + 784, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3 Edge Classification and Gradient Scattering", + "text_level": 1, + "bbox": [ + 214, + 335, + 624, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To backpropagate the gradients to the scene parameters we need to compute the gradient with respect to fragment positions, while in §4.2 we only derived the gradient $\\frac{\\partial L}{\\partial p_{AB}}$ with respect to the edge position. We scatter each edge gradient $\\frac{\\partial L}{\\partial p_{AB}}$ to adjacent fragments where we accumulate the fragment gradients, as illustrated in Figure 2. Using the chain rule, we obtain fragment gradients be multiplying $\\frac{\\partial L}{\\partial p_{AB}}$ with $\\frac{\\partial p_{AB}}{\\partial r}$ ; the later relates the movement of the edge to the movement of the fragment, where $r$ is the position of the fragment. The computation of $\\frac{\\partial p_{AB}}{\\partial r}$ depends on the type of the edge formed by the fragments. We consider several cases: overhanging primitives (either one another or background); adjacent primitives; intersecting primitives. Unlike Laine et al. [20], we avoid additional data structures like hash maps for connectivity.", + "bbox": [ + 212, + 363, + 784, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our approach examines each neighboring pixel pair, comparing their triangle IDs, see Figure 3 for illustration. Matching IDs indicate no edge; differing IDs prompt further tests. A single valid ID suggests a primitive overhanging the background. Testing a pixel's center against the triangle of the other pixel discerns overhangs and self-intersections. Non-contained pixels in each other's triangles imply adjacency. If both pixel centers are in the triangle of the other pixel, it implies self-intersections, and if in just one - then overhanging. For adjacent primitives we do not scatter gradients at all, since it is likely due to primitives sharing an edge and there is no discontinuity. Please note that even in that case $\\frac{\\partial L}{\\partial p_{AB}}$ may be non-zero due to $\\Omega$ contribution. In the overhanging case, we scatter the gradient only to the primitive on top, as the covered primitive can not affect the position of the edge. Thus, in such cases the covered fragment has $\\frac{\\partial p}{\\partial r} = 0$ while the foreground fragment has $\\frac{\\partial p}{\\partial r} = 1$ reflecting that the fragment movement is rigidly tied to the movement of the edge. Please refer to supplementary material for more details.", + "bbox": [ + 212, + 532, + 785, + 760 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4 Geometry Intersections", + "text_level": 1, + "bbox": [ + 214, + 782, + 457, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This subsection details computing the derivative of edge location with respect to fragment position $\\frac{\\partial p}{\\partial r}$ in scenarios where the edge arises from intersecting", + "bbox": [ + 212, + 809, + 784, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "S. Pidhorskyi et al.", + "bbox": [ + 271, + 114, + 400, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9cc86066d8317d857ce964646a3dc67a00101925ae815eb2251be0ca97ad0293.jpg", + "image_caption": [ + "Fig. 3: Edge classification. The accompanying figure provides a schematic representation of our edge classification process. Edges are categorized into four types: no edge, adjacent primitives, overhanging primitives, and intersecting primitives.", + "Fig. 4: Geometry intersection. A geometrical derivation for the partial derivative $\\frac{\\partial p}{\\partial r}$ , essential for converting edge-related gradients to fragment-related gradients in cases where edges are formed by geometry intersections. This shows the x-z plane, where the render camera is viewing the scene along the z axis so that edges are parallel to the y axis (and perpendicular to the drawing). Projections of both intersecting primitives are transformed into two lines." + ], + "image_footnote": [], + "bbox": [ + 217, + 142, + 495, + 450 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/82348e31d87e5b4a41c69b565e7ff1364fc52a1448726f80d4a788eab7633299.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 148, + 785, + 361 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "geometry. This situation is inherently complex, and a general-case solution would be challenging. However, our micro-edge formulation streamlines the problem because intersection lines projected onto the image plane comprised of microedges, e.i. either vertical or horizontal. This allows us to analyze the problem in 2D, focusing on the x-z or y-z planes, with the z-axis perpendicular to the image plane. We derive the derivative of the edge location by analyzing two fragments of a micro-edge that belongs to the intersection line, focusing on how the edge would move in response to infinitesimal movements of each fragment. Without loss of generality we consider the x-z plane:", + "bbox": [ + 212, + 578, + 787, + 715 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial p}{\\partial r} = - \\mathbf {n} _ {z} ^ {f} \\left[ \\mathbf {n} ^ {f ^ {T}} \\left[ \\begin{array}{l l} 0 & 1 \\\\ - 1 & 0 \\end{array} \\right] \\mathbf {n} ^ {v} \\right] ^ {- 1}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 729, + 785, + 766 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $p$ is edge position, $r$ is the fragment position (in 3D clip space), $\\mathbf{n}^v$ is the varying primitive's normal, and $\\mathbf{n}^f$ is the fixed primitive's normal. In this context, the $x$ -axis lies on the image plane, and the $z$ -axis is perpendicular to it, see Fig. 4. We compute $\\frac{\\partial p}{\\partial r}$ for both fragments that form the intersection edge by", + "bbox": [ + 212, + 779, + 787, + 842 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Rasterized Edge Gradients: Handling Discontinuities Differently", + "bbox": [ + 285, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "first fixing one fragment and varying the other, and then switching them. Please refer to supplementary material for derivation details.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.5 Theoretical analysis of the approximation", + "text_level": 1, + "bbox": [ + 215, + 200, + 606, + 217 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "At first glance, replacing arbitrarily oriented edges with vertical and horizontal micro-edges might seem overly simplistic, raising doubts about accuracy. It could be argued, for example, that the perimeter defined by micro-edges doesn't match the original as micro-edge size approaches zero. However, applying the two-dimensional version of the divergence theorem from vector calculus offers insight:", + "bbox": [ + 212, + 229, + 782, + 305 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\oint_ {C} \\mathbf {f} \\cdot \\mathbf {n} d t = \\iint_ {D} (\\nabla_ {\\Phi} \\cdot \\mathbf {f}) d x d y, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 387, + 318, + 785, + 349 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\mathbf{n}$ is normal to boundary $C$ , within which the integrand function is continuous and differentiable. We are free to pick any such boundaries that includes a discontinuity. In our case, the term $\\frac{\\partial\\alpha(x,y)}{\\partial\\Phi}$ in Eqn. (6) is a normal vector to the boundary $\\alpha (x,y) = 0$ thus the integral is equivalent to the left-hand side of Eqn. (13). According to Eqn. (13), this recasts the boundary derivative integral as an area integral. The difference between the area integrals of micro-edges and original edges is confined to a space defined by the micro-edges size. As pixel size diminishes, the integral approaches the exact value, and the discrepancy area vanishes, confirming the micro-edge approach converges to the same solution as the original formulation in the limit.", + "bbox": [ + 212, + 354, + 787, + 507 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experimental Analysis", + "text_level": 1, + "bbox": [ + 215, + 532, + 472, + 550 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We implement our method in PyTorch with custom CUDA kernels, and largely follow the structure of nvdiffrast [20], with the exception for the antialIASing module which we replace with our EdgeGrad module. For all experiments we utilize Laplacian preconditioning proposed by Nicolet et al. [34]. We find Laplacian preconditioning extremely helpful for stabilizing training and we use it for both, meshes and texture data. While Laplacian preconditioning may reduce self-intersections, this happens only due to much smoother optimization process, but it does not account for them in any way.", + "bbox": [ + 212, + 566, + 784, + 686 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Forward gradients. We first analyze the performance of the proposed approach by evaluating the forward gradients on a variety of test cases. Forward gradients are per-pixel derivatives with respect to some scene parameter, which is depicted with white arrows on the scene images. Fig. 5 shows the forward gradients computed using finite differences, raytracing methods Mitsuba [14,15,36] and redner [21], and rasterization approaches including nvdiffrast [20] and our approach (EdgeGrad). Unlike other methods, our approach still matches finite differences gradients in cases of intersecting geometry. While Mitsuba and Redner, which require multiple samples per pixel (256 in these comparisons), typically offer higher accuracy, they fall short at geometry intersections. Nvdiffrast", + "bbox": [ + 212, + 689, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "S. Pidhorskyi et al.", + "bbox": [ + 271, + 114, + 400, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/f6d8019b5f1778c9efb33293a150d0dd0f86b8cf3e380987e407fb2e9013173f.jpg", + "image_caption": [ + "Fig. 5: Comparison of forward gradient on several test scenes with the numerical solution using finite differences. (a) Synthetic test scene, we show gradients with respect to a parameter. We use (b) finite differences as a reference, and compare with (c) Mitsuba 3 [15,36], (d) redner [21], (e) nvdiffrast [20], and (f) our edge gradient approach." + ], + "image_footnote": [], + "bbox": [ + 232, + 143, + 772, + 426 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "is noisier because it assumes triangles containing silhouette edges always overlap pixel centers, which often isn't the case, leading to omitted gradients.", + "bbox": [ + 212, + 522, + 782, + 551 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Backward gradients. Tab. 1 quantitatively analyzes the accuracy of the backward gradients. We show improved results especially for complex intersecting geometry. Compared to raytracing methods, our rasterization-based approach yields comparable performance at a fraction of the computational cost.", + "bbox": [ + 212, + 551, + 784, + 612 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Runtime performance end accuracy by image size. Fig. 6 shows running time of our method (EdgeGrad) compared to Mitsuba [36] and nvdiffrast [20]. Our method significantly outperforms other approaches in running time on all variety of image sizes and triangle count. Our method has comparable accuracy to Mitsuba [36] (and even better for the case with geometry intersections) at large image sizes, but performs worse for smaller image size due to sub pixel triangles.", + "bbox": [ + 212, + 613, + 784, + 703 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Quantitative scene reconstructions. Tab. 2 shows quantitative results of scene reconstruction on blender dataset [32]. We follow the testing procedure described in [32] and report PNSR, SSIM and LPIPS metrics. We compare our method (EdgeGrad) with a \"continues only\" baseline (which ignores gradients from discontinuities), nvdiffrast [20], and with a variant of our method that does not have a special handling for the geometry intersections. Method optimization procedure is similar to Nicolet et al. [34] we only add MLP for decoding texture values in a view conditioned manner similar to [29]. We initialize geometry with some coarse manually created meshes shown in Fig. 7. To avoid additional complex", + "bbox": [ + 212, + 704, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Rasterized Edge Gradients: Handling Discontinuities Differently", + "bbox": [ + 285, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b643c42ff89145154b5bab67e750934303441922c0807c1c48f7a3cf15489d8b.jpg", + "image_caption": [ + "Scene" + ], + "image_footnote": [], + "bbox": [ + 303, + 167, + 357, + 210 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/090f379491bb8630c9df89a6493fc5c8f5022bd3eb23e6e0ae9f9b2fe088ff27.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 169, + 411, + 210 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ee24749cb0742bd6c8bc88c6930b2ba7c40680db608f7604804287caa8578db2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 169, + 467, + 210 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/2152a1021f050952fa49ab719df62f72b7985c762a744c5c424565a68b56d7b4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
redner [21]1.20%53.52%37.94%
Mitsuba3 [36]0.67%75.83%37.59%
Nvdiffrast [20]45.91%69.33%39.42%
EdgeGrad(our)6.01%3.35%8.35%
", + "bbox": [ + 222, + 210, + 468, + 256 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 1: Accuracy of backward gradients. Relative error, $\\%$ (↓). This table shows relative errors in backward gradient computations for test scenes. Second and third scenes include geometry intersections, emphasizing our method's advantage in managing these complexities.", + "bbox": [ + 215, + 267, + 496, + 364 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/389c0b8eef6169a730b9a8abcf6fda6dec616be3a023323a988728951e109ade.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 145, + 643, + 209 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6485593257a823179f267d5d321d747f931a3337141f962a7b0d8db211654e31.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 146, + 779, + 207 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d90e8eeb0171a38c1b491bfb4c88d81263633763c33fe984056580f10e65f4fd.jpg", + "image_caption": [ + "(a) Running time vs image size", + "(c) Gradient error vs image size self-intersections absent" + ], + "image_footnote": [], + "bbox": [ + 504, + 217, + 643, + 277 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8a12308d9f1178894f10228224d00969ad3d3f2c4643e6f2879325be6ed49e7a.jpg", + "image_caption": [ + "(b) Running time vs triangle count", + "(d) Gradient error vs image size self-intersections present" + ], + "image_footnote": [], + "bbox": [ + 645, + 218, + 779, + 277 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8487633fe2caeb03f444e1e52bf15afd80a1a9cebc2ed6f5cede1d6f6aea9e2e.jpg", + "image_caption": [ + "(a) camera image" + ], + "image_footnote": [], + "bbox": [ + 217, + 401, + 372, + 494 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4ddc715f2aa7b1e89c480bc319b18971613a7ac6ab20ea0b53f2f2a977334999.jpg", + "image_caption": [ + "(c) initialization mesh" + ], + "image_footnote": [], + "bbox": [ + 375, + 407, + 437, + 494 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8d62125d94ee7e410c80a1043a19cf88a70db77957a8acf96ee7a275317ad8ca.jpg", + "image_caption": [ + "(d) optimized mesh" + ], + "image_footnote": [], + "bbox": [ + 449, + 407, + 508, + 492 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/dffc9ff9e2f524a9048ed246965a1c6631296f411bd55bffca9df4c33d233147.jpg", + "image_caption": [ + "Fig. 6: Runtime performance and errors. a) Runtime [ms] by image size [px]; b) runtime [ms] by triangle count; c) gradient error [%] by image size [px] without self-intersections; d) gradient error [%] by image size [px] with self-intersections.", + "(e)rgb estimation", + "(f) segmentation estimation", + "(g) Learned segmentation texture for mouth region" + ], + "image_footnote": [], + "bbox": [ + 514, + 402, + 669, + 493 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/274eb69f9c16bb8ac44903afd294142c7d3ef8e5c41a8215ac8445622ad4b0b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 678, + 406, + 779, + 484 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/88258279ed633583bb90674a385ed9df0235b0c3711379b53a7e4a1f9c96c1e1.jpg", + "image_caption": [ + "Initialization" + ], + "image_footnote": [], + "bbox": [ + 228, + 508, + 267, + 553 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b39e087da88a6328eb84f34e39e8d7a705043941a12e24b2cd917f1f3a5da157.jpg", + "image_caption": [ + "Optimized" + ], + "image_footnote": [], + "bbox": [ + 287, + 508, + 323, + 551 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/248791888a2375df21f3a3a8665e5033f5e575120217a7e35942008366628364.jpg", + "image_caption": [ + "Initialization" + ], + "image_footnote": [], + "bbox": [ + 330, + 512, + 418, + 550 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/560488f97baf3b0cb4e1ca86a411f69d3960957e1f6e6f24ea5b4e14a165ab10.jpg", + "image_caption": [ + "1" + ], + "image_footnote": [], + "bbox": [ + 419, + 512, + 473, + 550 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2c8d35fbfa938274b6c868dd885f44079506b12699137ff86afc5e0bb28d81f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 473, + 512, + 558, + 550 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8d085686284e5bd645fe4691dd3ffa480fcc5611dcd4e7543ab1d93ad0d699f0.jpg", + "image_caption": [ + "Optimized", + "(h) Initialization mesh and reconstructed mesh for synthetic blender dataset" + ], + "image_footnote": [], + "bbox": [ + 488, + 555, + 495, + 560 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a5c41640ef10d4bf8cbf4df6fa7d4ce9ad0a6fda595aa97fb9ad4aa6f6fe7fe0.jpg", + "image_caption": [ + "Initialization" + ], + "image_footnote": [], + "bbox": [ + 568, + 512, + 612, + 550 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e87f739b8c0b08f09e956919173c7cb025c3be4bd014f80975f5c1a94bc1e6da.jpg", + "image_caption": [ + "Optimized" + ], + "image_footnote": [], + "bbox": [ + 622, + 512, + 666, + 550 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e9bc93948f9064931b0977efddcdc988349450976a03e4a2f58f294208c5dfb3.jpg", + "image_caption": [ + "Initialization", + "Optimized" + ], + "image_footnote": [], + "bbox": [ + 669, + 510, + 785, + 551 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/270631c5ca503c37c14aa3c9427544ccfb88f18e097cebcb6a87ead44ba9ee00.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 220, + 666, + 263, + 741 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b8eca095d0ea7e666d6fecef422be1dece8d1cfabca9e37a2c9f0396b88b7560.jpg", + "image_caption": [ + "NvdiffRast" + ], + "image_footnote": [], + "bbox": [ + 267, + 667, + 310, + 742 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/648a873f103fe7c1789773e075752bdf110c8735d327c4768ee1d1a1921ad351.jpg", + "image_caption": [ + "Fig. 7: Experimental setup details. The figure shows: a) the ground truth camera image; b) predicted segmentation mask for mouth region using HRNet [47]; c) initialization mesh from 3DMM; d) optimized avatar mesh; e) rendered image of the optimized model; f) rendered segmentation estimation image; g) learned static segmentation texture; h) initialization meshes and optimized meshes for blender dataset [32].", + "No Edge Handling" + ], + "image_footnote": [], + "bbox": [ + 310, + 665, + 372, + 743 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/44817837ff5de8254c6ce671dfdd446af19918435c0f2ec5acb645579b178cc1.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 375, + 672, + 419, + 738 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5d422cfe9914f436eb1aae59526f6f6d8892d0be75fd0816abfb0fa7b9005d0c.jpg", + "image_caption": [ + "NvdRast" + ], + "image_footnote": [], + "bbox": [ + 424, + 674, + 473, + 738 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e5d35610e079806573a70c835612e36d6bb364ffb6f20d874affdea6727cc56e.jpg", + "image_caption": [ + "No Edge Handling" + ], + "image_footnote": [], + "bbox": [ + 473, + 674, + 531, + 738 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/93b5754078ade38f45e2087b8334616cd999a75c9dbfff144d132b399d7c97b4.jpg", + "image_caption": [ + "Ours", + "Fig. 8: Reconstructions from inverse rendering compare our method with nvdiffrast [20] and rasterization without discontinuity handling. Although nvdiffrast can match the outline of objects well, it suffers from interpenetration (see leaves of Ficus) and does not capture as much detail as ours (see the backrest of the chair and ground of lego). The method with no edge handling fails completely on these objects." + ], + "image_footnote": [], + "bbox": [ + 535, + 678, + 614, + 746 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/17418c45a056fc50c1a8f05c1b7623d4ef36bf41564254015e566cd33678ae3a.jpg", + "image_caption": [ + "NdivdiffRast" + ], + "image_footnote": [], + "bbox": [ + 614, + 678, + 694, + 744 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2b9cd60707d6f6f32529b7e4faaca333e8627c9c39b9a44473a9c98985c0f27b.jpg", + "image_caption": [ + "No Edge Handling" + ], + "image_footnote": [], + "bbox": [ + 696, + 665, + 779, + 741 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "S. Pidhorskyi et al.", + "bbox": [ + 269, + 114, + 400, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/af1dc1de693ba3d88a7be3444fa3a11f174752745efaf2e4e780f171bf889824.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LegoChairShipHotdogFicusDrumsMicMaterials
PSNR ↑Continuous only16.34228.15115.69327.85018.39915.53421.79717.11
Nvdiffrast [20]29.4429.80426.17331.68424.96423.15429.63824.413
EdgeGrad -intersect.29.57433.07726.73134.22526.78124.0831.22425.221
EdgeGrad (our)29.66732.98126.6534.08126.55924.01431.34625.287
LegoChairShipHotdogFicusDrumsMicMaterials
SSIM ↑Continuous only0.75500.95310.69290.93740.82450.77230.91820.7693
Nvdiffrast [20]0.94670.94160.83160.94730.93230.89710.96470.8865
EdgeGrad -intersect.0.94850.97650.84780.97150.94680.92160.97560.8945
EdgeGrad (our)0.95010.97640.84720.97130.94520.92200.97670.8972
LegoChairShipHotdogFicusDrumsMicMaterials
LPIPS ↓Continuous only0.32230.06110.39130.11090.20460.29430.12720.2511
Nvdiffrast [20]0.06530.07710.20200.09360.07810.11160.05490.1291
EdgeGrad -intersect.0.06330.03480.18990.05460.06810.08710.04390.1200
EdgeGrad (our)0.06140.03410.18500.05480.06720.08700.03990.1185
", + "bbox": [ + 233, + 143, + 774, + 301 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 2: Per-scene quantitative results from the realistic synthetic blender dataset [32], featuring scenes with complex geometries and non-Lambertian materials rendered using Blender's Cycles pathtracer.", + "bbox": [ + 215, + 301, + 785, + 344 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1d8b8d3ae3ddd5675c2650aeb5d70c079a873e25269e6360b6ff64cdbbf2eb60.jpg", + "image_caption": [ + "Fig. 9: Example renderings of mesh-based avatars from [29] trained using EdgeGrad, capturing complex inner mouth movements, including teeth and tongue dynamics." + ], + "image_footnote": [], + "bbox": [ + 217, + 369, + 785, + 448 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "ity we opted out automatic remeshing introduced by Nicolet et al. which would necessitate updating UV parametrization in our case.", + "bbox": [ + 212, + 515, + 784, + 545 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Qualitative mesh reconstruction. Qualitatively, we compare the mesh reconstructions obtained on the public NeRF dataset [32] using our approach, a naive approach with no special edge handling, and nvdiffrast [20] in Fig. 8. All methods use the same shading model. Our approach shows improved detail in the recovered meshes.", + "bbox": [ + 212, + 546, + 785, + 619 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6 Application to Dynamic Head Scenes", + "text_level": 1, + "bbox": [ + 215, + 643, + 612, + 660 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We show the importance of correct edge gradients on a head modeling application by fitting a personalized mesh-based avatar representation to facial performances captured in a multiview lightstage. modifying the avatar creation method from [29]. As the facial avatar representation, we use an encoder-decoder architecture to produce mesh vertices and texture-mapped appearance features [25], which are shaded using a pixel-wise MLP producing RGB appearance [29]. We use a coarse face tracking produced using traditional computer vision as initialization for inverse rendering. Importantly, this coarse face template has a small number of vertices (8k) and no geometry in the interior of the mouth.", + "bbox": [ + 212, + 672, + 785, + 809 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Traditionally, fitting the interior of the mouth without strong priors is a complex task due to poor observability and intricately interacting geometry with", + "bbox": [ + 212, + 810, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Rasterized Edge Gradients: Handling Discontinuities Differently", + "bbox": [ + 285, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1970cc83f0a77b518b57863fdc75e12dcd76c03d06f58238ff3249012a5def9d.jpg", + "image_caption": [ + "(a) camera image" + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 305, + 303 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3131f3014cc1620513b16c33ac01caacbfaf04db16b11f7001524f0f1f005299.jpg", + "image_caption": [ + "(b) Nvdiffrast" + ], + "image_footnote": [], + "bbox": [ + 307, + 143, + 392, + 303 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/71b44cfe6577bba41607c6fd59db4b53843b06b01255ea9b960ac0f6c4b138a6.jpg", + "image_caption": [ + "(c) EdgeGrad (our)", + "(d) EdgeGrad (our) w/o segmentation supervision" + ], + "image_footnote": [], + "bbox": [ + 392, + 143, + 475, + 303 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/930f8e889bfb2952980f7dd8dcd1dbff3d9afbc3b2906a565ec9ceb2e7129184.jpg", + "image_caption": [ + "Fig. 10: Qualitative evaluation and ablation on dynamic head scene. We compare: (c) our method, (a) ground truth images, (b) Nvdiffrast [20], and ablated version of our method without including segmentation mask. The avatar model is from [29] trained using EdgeGrad and Nvdiffrast. We highlight artifacts with white arrows." + ], + "image_footnote": [], + "bbox": [ + 477, + 143, + 562, + 303 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "constant contact. Prior work either relies on artists to create alignments, volumetric simulation to prevent self-intersections [13], or on very specific mouth interior capture methods to build priors [30,49]. Instead, we show that by using differentiable rasterization with correct edge gradients, we can achieve fully automatic geometric fitting of the complex geometry of the face and mouth interior including lips, teeth, and tongue using a mesh-based model of fixed topology. We modify [29] in two ways. First, we use the edge gradient rasterization method of Sec. 4. Second, we add supervision in the form of 2D segmentation masks with separate labels for the regions of the teeth, tongue, and lips, for details see Fig. 7. While we can achieve accurate per-frame geometry without the additional segmentation loss supervision, we find that segmentation losses help anchor the teeth and tongue geometry to a consistent region of the UV map. Fig. 9 shows the quality of geometric fitting achievable with this inverse rendering method. Note that we accurately reconstruct the complex contact geometry of the tongue sliding across teeth, teeth-to-teeth contact, as well as the tongue pushing against the lips and cheeks. Fig. 10 demonstrates reconstruction quality of complex inner mouth details, which are challenging due to frequent self-intersections. We also include results without 2D segmentation mask supervision.", + "bbox": [ + 212, + 343, + 787, + 616 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusions and Limitations", + "text_level": 1, + "bbox": [ + 215, + 640, + 527, + 657 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We have presented a fast method for rasterization-based differentiable rendering that has accurate gradients across visibility discontinuities such as triangle edges, occlusions and geometry intersections. Our results demonstrate enhanced mesh reconstruction fidelity across diverse inverse rendering scenarios, comparable to implicit and volumetric methods. Our method also unlocks new applications, accurate mouth interior fitting in dynamic facial performance capture, through fully automatic inverse rendering. Regarding limitations, our method inherits the constraints of rasterization, including the inability to manage transparency, lack of precise antialIASing without multisampling, and the absence of physically based rendering and global illumination. It may also inaccurately interpret adjacent primitives in finely tessellated meshes, leading to sub-pixel triangles.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "S. Pidhorskyi et al.", + "bbox": [ + 271, + 114, + 400, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Baek, S., Kim, K.I., Kim, T.K.: Pushing the envelope for rgb-based dense 3d hand pose estimation via neural rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1067-1076 (2019)", + "2. Bangaru, S.P., Gharbi, M., Luan, F., Li, T.M., Sunkavalli, K., Hasan, M., Bi, S., Xu, Z., Bernstein, G., Durand, F.: Differentiable rendering of neural sdfs through reparameterization. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)", + "3. Bangaru, S.P., Li, T.M., Durand, F.: Unbiased warped-area sampling for differentiable rendering. ACM Transactions on Graphics (TOG) 39(6), 1-18 (2020)", + "4. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)", + "5. Beker, D., Kato, H., Morariu, M.A., Ando, T., Matsuoka, T., Kehl, W., Gaidon, A.: Monocular differentiable rendering for self-supervised 3d object detection. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16. pp. 514-529. Springer (2020)", + "6. Bogo, F., Kanazawa, A., Lassner, C., Gehler, P., Romero, J., Black, M.J.: Keep it spl! Automatic estimation of 3d human pose and shape from a single image. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14. pp. 561-578. Springer (2016)", + "7. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022)", + "8. Chen, W., Ling, H., Gao, J., Smith, E., Lehtinen, J., Jacobson, A., Fidler, S.: Learning to predict 3d objects with an interpolation-based differentiable renderer. Advances in neural information processing systems 32 (2019)", + "9. Chen, W., Litalien, J., Gao, J., Wang, Z., Fuji Tsang, C., Khamis, S., Litany, O., Fidler, S.: Dib-r++: learning to predict lighting and material with a hybrid differentiable renderer. Advances in Neural Information Processing Systems 34, 22834-22848 (2021)", + "0. Choy, C.B., Xu, D., Gwak, J., Chen, K., Savarese, S.: 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14. pp. 628-644. Springer (2016)", + "1. Ge, L., Ren, Z., Li, Y., Xue, Z., Wang, Y., Cai, J., Yuan, J.: 3d hand shape and pose estimation from a single rgb image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10833-10842 (2019)", + "2. Henzler, P., Reizenstein, J., Labatut, P., Shapovalov, R., Ritschel, T., Vedaldi, A., Novotny, D.: Unsupervised learning of 3d object categories from videos in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4700-4709 (2021)", + "3. Ichim, A.E., Kadleček, P., Kavan, L., Pauly, M.: Phace: Physics-based face modeling and animation. ACM Trans. Graph. 36(4) (jul 2017)", + "4. Jakob, W., Speierer, S., Roussel, N., Nimier-David, M., Vicini, D., Zeltner, T., Nicolet, B., Crespo, M., Leroy, V., Zhang, Z.: Mitsuba 3 renderer (2022), https://mitsuba-renderer.org" + ], + "bbox": [ + 225, + 175, + 785, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Rasterized Edge Gradients: Handling Discontinuities Differently", + "bbox": [ + 285, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Jakob, W., Speierer, S., Roussel, N., Vicini, D.: Dr.jit: A just-in-time compiler for differentiable rendering. Transactions on Graphics (Proceedings of SIGGRAPH) 41(4) (Jul 2022). https://doi.org/10.1145/3528223.3530099", + "16. Jiang, Y., Ji, D., Han, Z., Zwicker, M.: Sdiff: Differentiable rendering of signed distance fields for 3d shape optimization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1251-1261 (2020)", + "17. Kato, H., Ushiku, Y., Harada, T.: Neural 3d mesh renderer. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3907-3916 (2018)", + "18. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG) 42(4), 1-14 (2023)", + "19. de La Gorce, M., Fleet, D.J., Paragios, N.: Model-based 3d hand pose estimation from monocular video. IEEE Trans. Pattern Anal. Mach. Intell. 33(9), 1793-1805 (Sep 2011). https://doi.org/10.1109/TPAMI.2011.33, http://dx.doi.org/10.1109/TPAMI.2011.33", + "20. Laine, S., Hellsten, J., Karras, T., Seol, Y., Lehtinen, J., Aila, T.: Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG) 39(6), 1-14 (2020)", + "21. Li, T.M., Aittala, M., Durand, F., Lehtinen, J.: Differentiable monte carlo ray tracing through edge sampling. ACM Trans. Graph. (Proc. SIGGRAPH Asia) 37(6), 222:1-222:11 (2018)", + "22. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 300–309 (2023)", + "23. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020)", + "24. Liu, S., Li, T., Chen, W., Li, H.: Soft rasterizer: A differentiable renderer for image-based 3d reasoning. The IEEE International Conference on Computer Vision (ICCV) (Oct 2019)", + "25. Lombardi, S., Saragih, J., Simon, T., Sheikh, Y.: Deep appearance models for face rendering. ACM Transactions on Graphics (TOG) 37(4), 1-13 (2018)", + "26. Lombardi, S., Simon, T., Saragih, J., Schwartz, G., Lehrmann, A., Sheikh, Y.: Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751 (2019)", + "27. Loper, M.M., Black, M.J.: Opendra: An approximate differentiable renderer. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VII 13. pp. 154-169. Springer (2014)", + "28. Loubet, G., Holzschuch, N., Jakob, W.: Reparameterizing discontinuous integrands for differentiable rendering. Transactions on Graphics (Proceedings of SIGGRAPH Asia) 38(6) (Dec 2019). https://doi.org/10.1145/3355089.3356510", + "29. Ma, S., Simon, T., Saragih, J.M., Wang, D., Li, Y., la Torre, F.D., Sheikh, Y.: Pixel codec avatars. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 64-73 (2021)", + "30. Medina, S., Tomé, D., Stoll, C., Tiede, M., Munhall, K., Hauptmann, A., Matthews, I.: Speech driven tongue animation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE/CVF (2022)", + "31. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "S. Pidhorskyi et al.", + "bbox": [ + 271, + 114, + 400, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)", + "33. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989 (2022)", + "34. Nicolet, B., Jacobson, A., Jakob, W.: Large steps in inverse rendering of geometry. ACM Transactions on Graphics (TOG) 40(6), 1-13 (2021)", + "35. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3504-3515 (2020)", + "36. Nimier-David, M., Vicini, D., Zeltner, T., Jakob, W.: Mitsuba 2: A retargetable forward and inverse renderer. Transactions on Graphics (Proceedings of SIGGRAPH Asia) 38(6) (Dec 2019). https://doi.org/10.1145/3355089.3356498", + "37. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019)", + "38. Pavlakos, G., Zhu, L., Zhou, X., Daniilidis, K.: Learning to estimate 3d human pose and shape from a single color image. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 459-468 (2018)", + "39. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)", + "40. Rhodin, H., Robertini, N., Richardt, C., Seidel, H.P., Theobalt, C.: A versatile scene model with differentiable visibility applied to generative pose estimation. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 765-773 (2015)", + "41. Roessle, B., Barron, J.T., Mildenhall, B., Srinivasan, P.P., Nießner, M.: Dense depth priors for neural radiance fields from sparse input views. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12892-12901 (2022)", + "42. Roveri, R., Öztireli, A.C., Pandele, I., Gross, M.: Pointpronets: Consolidation of point clouds with convolutional neural networks. In: Computer Graphics Forum. vol. 37, pp. 87-99. Wiley Online Library (2018)", + "43. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512 (2023)", + "44. Tsalicoglou, C., Manhardt, F., Tonioni, A., Niemeyer, M., Tombari, F.: Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439 (2023)", + "45. Tulsiani, S., Zhou, T., Efros, A.A., Malik, J.: Multi-view supervision for single-view reconstruction via differentiable ray consistency. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2626-2634 (2017)", + "46. Vicini, D., Speierer, S., Jakob, W.: Differentiable signed distance function rendering. ACM Transactions on Graphics (TOG) 41(4), 1-18 (2022)", + "47. Wang, J., Sun, K., Cheng, T., Jiang, B., Deng, C., Zhao, Y., Liu, D., Mu, Y., Tan, M., Wang, X., et al.: Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence 43(10), 3349-3364 (2020)", + "48. Wiles, O., Gkioxari, G., Szeliski, R., Johnson, J.: Synsin: End-to-end view synthesis from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7467-7477 (2020)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Rasterized Edge Gradients: Handling Discontinuities Differently", + "bbox": [ + 285, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "49. Wu, C., Bradley, D., Garrido, P., Zollhöfer, M., Theobalt, C., Gross, M., Beeler, T.: Model-based teeth reconstruction. ACM Trans. Graph. 35(6) (dec 2016)", + "50. Yan, X., Yang, J., Yumer, E., Guo, Y., Lee, H.: Perspective transformer nets: Learning single-view 3d object reconstruction without 3d supervision. Advances in neural information processing systems 29 (2016)", + "51. Yang, G., Cui, Y., Belongie, S., Hariharan, B.: Learning single-view 3d reconstruction with limited pose supervision. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 86-101 (2018)", + "52. Yu, A., Fridovich-Keil, S., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks (2021)", + "53. Zhang, C., Wu, L., Zheng, C., Gkioulekas, I., Ramamoorthi, R., Zhao, S.: A differential theory of radiative transfer. ACM Trans. Graph. 38(6), 227:1-227:16 (2019)", + "54. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492 (2020)", + "55. Zhao, S., Jakob, W., Li, T.M.: Physics-based differentiable rendering: From theory to implementation. In: ACM SIGGRAPH 2020 Courses. SIGGRAPH '20, Association for Computing Machinery, New York, NY, USA (2020). https://doi.org/10.1145/3388769.3407454, https://doi.org/10.1145/3388769.3407454" + ], + "bbox": [ + 215, + 146, + 784, + 396 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "S. Pidhorskyi et al.", + "bbox": [ + 271, + 114, + 400, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Rasterized Edge Gradients_ Handling Discontinuities Differentially/ac83eb0e-ca7d-4471-8db5-05179e149303_model.json b/2024/Rasterized Edge Gradients_ Handling Discontinuities Differentially/ac83eb0e-ca7d-4471-8db5-05179e149303_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2fa38d6e699b1b4e3108712c2d0d2f1e958001d7 --- /dev/null +++ b/2024/Rasterized Edge Gradients_ Handling Discontinuities Differentially/ac83eb0e-ca7d-4471-8db5-05179e149303_model.json @@ -0,0 +1,2964 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.29, + 0.141, + 0.713, + 0.186 + ], + "angle": 0, + "content": "Rasterized Edge Gradients: Handling Discontinuities Differently" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.213, + 0.768, + 0.243 + ], + "angle": 0, + "content": "Stanislav Pidhorskyi\\(^{\\text{①}}\\), Tomas Simon\\(^{\\text{②}}\\), Gabriel Schwartz\\(^{\\text{③}}\\), He Wen, Yaser Sheikh, and Jason Saragih" + }, + { + "type": "text", + "bbox": [ + 0.331, + 0.256, + 0.672, + 0.269 + ], + "angle": 0, + "content": "Reality Labs,Meta,Pittsburgh, Pennsylvania,USA" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.308, + 0.739, + 0.584 + ], + "angle": 0, + "content": "Abstract. Computing the gradients of a rendering process is paramount for diverse applications in computer vision and graphics. However, accurate computation of these gradients is challenging due to discontinuities and rendering approximations, particularly for surface-based representations and rasterization-based rendering. We present a novel method for computing gradients at visibility discontinuities for rasterization-based differentiable renderers. Our method elegantly simplifies the traditionally complex problem through a carefully designed approximation strategy, allowing for a straightforward, effective, and performant solution. We introduce a novel concept of micro-edges, which allows us to treat the rasterized images as outcomes of a differentiable, continuous process aligned with the inherently non-differentiable, discrete-pixel rasterization. This technique eliminates the necessity for rendering approximations or other modifications to the forward pass, preserving the integrity of the rendered image, which makes it applicable to rasterized masks, depth, and normals images where filtering is prohibitive. Utilizing micro-edges simplifies gradient interpretation at discontinuities and enables handling of geometry intersections, offering an advantage over the prior art. We showcase our method in dynamic human head scene reconstruction, demonstrating effective handling of camera images and segmentation masks." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.612, + 0.375, + 0.627 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.644, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Significant advances have been made in recent years in modeling real 3D objects from image measurements. Much of this advancement can be credited to improvements in inverse rendering, which enables the automatic inference of 3D scene parameters that best reconstruct the images. While volumetric methods like NeRF [31] simplify inverse rendering by eliminating the need for predefined scene topology, classical mesh-based representations remain widely used due to their efficiency in modeling opaque surfaces. These representations often rely on highly performant rasterization, which requires discrete operations to reason about ordering and coverage, which poses challenges for gradient computations. Our main contribution is a theoretical framework for approximating gradients of a rasterization process that simultaneously elucidates previous constructions, arrives at a simple, fast, and accurate formulation, and further improves gradient approximation accuracy over state-of-the-art methods." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.401, + 0.128 + ], + "angle": 0, + "content": "S. Pidhorskyi et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.388 + ], + "angle": 0, + "content": "In rasterization, a rendered pixel's value is determined by the foremost triangle covering it. Pixels whose footprint lies entirely within a single triangle are straightforward to handle, as infinitesimal mesh motion does not change their triangle membership. However, pixels on triangle boundaries, either between adjacent triangles, at occlusion boundaries, or at intersections between triangles, must account for how their triangle membership changes in response to infinitesimal changes in the mesh. Existing methods employ different approximations of soft-membership to account for this, enabling gradient computation in those areas, effectively compositing contributions from member triangles via a weighted sum. For example, [20] uses the anti-aliasing approach, [14] averages montecarlo samples during ray-casting, and [40] extends the influence of boundary triangles using a falloff function. In this work, we show that all these methods exhibit approximation errors that can fail to compute correct gradients in one case or another. In particular, none of the existing methods accurately compute the gradient at triangle intersections, which can often happen during inverse rendering, especially when the current estimate is far from the desired solution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.788, + 0.57 + ], + "angle": 0, + "content": "In contrast to existing works that directly tackle geometric discontinuities and rasterization's discrete nature, we introduce a micro-edge formulation which allows to interpret the rasterized image as an outcome of a continuous process that coincidentally aligns with discrete-pixel rasterization, simplifying gradient computation significantly. Unlike some other works [20, 40], our formulation achieves good gradient estimates without altering the rasterization forward pass, maintaining the rasterized image's integrity. This is crucial for optimizing segmentation masks, depth maps, and normal maps where filtering or smoothing is not feasible, e.g. anisotropic filtering or soft rasterization would mix normals from different surfaces misrepresenting the geometry. The simplifications offered by our micro-edge formulation allow us to seamlessly handle discontinuities caused by geometry intersections, offering an advantage over the prior art." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.571, + 0.788, + 0.722 + ], + "angle": 0, + "content": "In summary, our paper introduces a straightforward, accurate, and efficient method for computing gradients in rasterization, comparable to existing techniques but with greater simplicity and the ability to handle self-intersections. We analyze our method's errors from a theoretical standpoint and compare gradients qualitatively and quantitatively with finite differences and other methods across various test cases. We also assess runtime efficiency and accuracy by image size and showcase qualitative and quantitative findings on a synthetic blender dataset [32]. Our method excels in complex applications, such as detailed dynamic human head reconstructions, effectively managing the intricate details of the inner mouth region with significant occlusions and deformations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.747, + 0.383, + 0.763 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Differentiable rendering is one of the critical building blocks for many existing computer vision problems, such as 3d object reconstruction [10, 17, 45, 50, 51], 3d object prediction [5, 8, 9, 12], pose estimation [1, 6, 11, 38] novel view synthesis [4, 18, 32, 33, 54], as well as newly emerging applications such as text-to-3D" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.287, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rasterized Edge Gradients: Handling Discontinuities Differently" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.239 + ], + "angle": 0, + "content": "generative models [22, 39, 43, 44]. Differentiable rendering employs a wide array of underlying 3D representations including explicit surfaces like mesh-based representations [19, 20, 27], implicit surfaces [2, 16, 37, 46], point clouds [18, 42, 48], explicit or implicit volume representations [4, 23, 26, 32, 35, 50, 52], and hybrid representations [7, 33]. For a comprehensive introduction to differentiable rendering, we refer the reader to Zhao et al.'s excellent SIGGRAPH course notes [55]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.242, + 0.788, + 0.501 + ], + "angle": 0, + "content": "Mesh optimization often poses more challenges than implicit and volumetric methods, leading to the latter's prevalence in the field, as noted by Roessle et al. [41]. Despite this, meshes excel in high-performance rendering and imply registration with predefined topologies. Mesh rendering methods can be categorized into ray tracing-based and rasterization-based, with the former including frameworks like Mitsuba 3 [15, 36]. Differentiable ray tracing, aimed at direct illumination or path tracing, is computationally expensive but principled. Key advancements by Li et al. [21] and Zhang et al. [53] split this gradient into discontinuity integrals with Dirac delta functions and differentiable continuous parts. Li et al. replaced Dirac delta integration with boundary line integration, while Zhang et al. applied Reynold's transport theorem for a similar decomposition. Both methods are intricate, necessitating silhouette edge sampling — a significant bottleneck in their application. Loubet et al. [28] proposed a variable change to simplify discontinuity integration, further refined by Bangaru et al. [3] using warped area fields to correct biases, yet practical application challenges persist. Despite their flexibility, the computational expense of such ray tracing frameworks restrict their practicality in optimization-intensive tasks [20]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.788, + 0.655 + ], + "angle": 0, + "content": "For rasterization-based methods, rendering meshes differentiably is hindered by non-differentiable visibility at boundaries. Unlike ray tracing, which uses boundary integrals for visibility gradients, rasterization's fixed-grid and z-buffer approach lacks this feature. Solutions typically involve approximating derivatives or approximating rendering to achieve inherent differentiability. De La Gorce et al. [19] and Loper et al. [27] pioneered derivative approximation in rasterization. De La Gorce et al. distinguished between continuous regions and occlusion boundaries, introducing 'occlusion forces' for the latter. Loper et al. simplified this by detecting discontinuities post-rasterization and employing differential filters, like the Sobel filter, for boundary approximation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The 'Nvdiffrast' approach by Laine et al. [20] tackles point-sampled visibility issues with innovative differentiable analytic antialIASing, which transforms sharp discontinuities into smooth transitions, enabling gradient computation. Analytical antialIASing is achieved by estimating coverage using silhouette edges. A limitation of the method is that if triangles containing silhouette edges do not overlap with any pixel center, they will be overlooked during the antialIASing process. Moreover, analytical antialIASing is inherently approximate, so the obtained gradients are also approximate. Regrettably, the approach is still quite complex, modifies the rasterized image, and requires a specialized data structure for connectivity and detailed pixel coverage computations. On the other hand, methods like Rhodin et al. [40] and Liu et al.'s 'Soft Rasterizer' [24] opt for modifying the rendering model. Rhodin et al. introduce fuzzy edges to soften discontinuities," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.401, + 0.128 + ], + "angle": 0, + "content": "S. Pidhorskyi et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "while 'Soft Rasterizer' further blurs boundaries and averages depth contributions, facilitating gradient propagation across occluded primitives." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.344 + ], + "angle": 0, + "content": "In this work, we deliberately bypass physically based rendering, global illumination, and lighting/material models, and instead focus exclusively on rasterization-based approaches for their speed. We draw inspiration from the core idea of the edge sampling method [21] and analytical antialIASing [20] and distill this into a much simpler yet effective method that also provides accurate gradients. Notably, our method's simplicity allows direct handling of self-penetrating geometry, an aspect not addressed in previous research. Other methods to handle self-penetrating geometry would require complex and computationally intensive geometry preprocessing at each optimization step, involving intersection detection and differentiable splitting of the faces. In contrast, our approach explicitly manages interpenetrating geometry with virtually no additional overhead." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.362, + 0.383, + 0.379 + ], + "angle": 0, + "content": "3 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.784, + 0.416 + ], + "angle": 0, + "content": "Following Li et al. [21], given a 2D pixel filter \\(k\\) and radiance \\(R\\), a pixel's color can be written:" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.416, + 0.785, + 0.447 + ], + "angle": 0, + "content": "\\[\nI = \\iint_ {D} k (x, y) R (x, y) d x d y. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.522 + ], + "angle": 0, + "content": "For notational convenience, as in [21], we denote \\( f(x,y) = k(x,y)R(x,y) \\). For simplicity, let us first consider the case of a single pixel. We are interested in the gradient of \\( \\frac{\\partial L(I)}{\\partial\\varPhi} \\), where the scalar function \\( L(I) \\) of the rendered pixel \\( I \\) defines our loss, and \\( \\varPhi \\) are the scene parameters that we aim to optimize. According to the chain rule:" + }, + { + "type": "equation", + "bbox": [ + 0.443, + 0.522, + 0.785, + 0.551 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial L (I)}{\\partial \\Phi} = \\frac {\\partial I}{\\partial \\Phi} \\frac {\\partial L}{\\partial I}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.784, + 0.6 + ], + "angle": 0, + "content": "where \\(\\frac{\\partial L}{\\partial I}\\) is the incoming gradient from the loss function and is typically computed by automatic differentiation (AD). So, our goal becomes to find how the pixel value changes with respect to the scene parameters:" + }, + { + "type": "equation", + "bbox": [ + 0.403, + 0.605, + 0.785, + 0.636 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial I}{\\partial \\Phi} = \\frac {\\partial}{\\partial \\Phi} \\iint_ {D} f (x, y) d x d y. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.788, + 0.701 + ], + "angle": 0, + "content": "While the function \\( f \\) may not in fact be differentiable, its integral remains continuous and thus differentiable. Let us assume that \\( f \\) is partitioned into two continuous half-spaces, each represented by two continuous and differentiable functions \\( f_{a} \\) and \\( f_{b} \\), respectively, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.328, + 0.708, + 0.785, + 0.723 + ], + "angle": 0, + "content": "\\[\nf (x, y) = \\theta (\\alpha (x, y)) f _ {a} (x, y) + \\theta (- \\alpha (x, y)) f _ {b} (x, y), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.728, + 0.784, + 0.774 + ], + "angle": 0, + "content": "where \\(\\alpha\\) specifies the dividing edge, and \\(\\theta\\) is the Heaviside step function, which selects between \\(f_{a}\\) and \\(f_{b}\\). The differentiation of the integral can be broken into two parts by applying the product rule. Let us start with the integrand \\(f_{a}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.236, + 0.778, + 0.785, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\frac {\\partial}{\\partial \\Phi} \\iint_ {D} \\theta (\\alpha (x, y)) f _ {a} (x, y) d x d y \\\\ = \\iint_ {D} \\delta (\\alpha (x, y)) \\frac {\\partial \\alpha (x , y)}{\\partial \\Phi} f _ {a} (x, y) d x d y + \\iint_ {D} \\frac {\\partial f _ {a} (x , y)}{\\partial \\Phi} \\theta (\\alpha (x, y)) d x d y, \\tag {5} \\\\ \\end{array}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.286, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Rasterized Edge Gradients: Handling Discontinuities Differently" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.283 + ], + "angle": 0, + "content": "and a similar derivation can be done for \\( f_{b} \\). The second term on the right-hand side of Eqn. (5) is the integral of the derivative of the smooth function, \\( f_{a}(x,y) \\). This derivative is readily computable with AD, and in the context of differentiable rasterizers it typically involves differentiating through the interpolation step, the barycentric coordinates, and the texture sampler. Due to its straightforward computation, we will represent it as \\( \\Omega \\) and omit it from further discussion, focusing instead on the first term on the right-hand side of Eqn. (5). This term involves a Dirac delta function, and can be interpreted as taking the integral over the line \\( \\alpha(x,y) = 0 \\):" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.293, + 0.785, + 0.36 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\iint_ {D} \\delta (\\alpha (x, y)) \\frac {\\partial \\alpha (x , y)}{\\partial \\Phi} f _ {a} (x, y) d x d y \\tag {6} \\\\ = \\int_ {\\alpha (x, y) = 0} \\frac {\\partial \\alpha (x , y)}{\\partial \\Phi} \\| \\nabla_ {x, y} \\alpha (x, y) \\| ^ {- 1} f _ {a} (x, y) d t, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.368, + 0.784, + 0.413 + ], + "angle": 0, + "content": "where \\( \\| \\nabla_{x,y}\\alpha (x,y)\\| \\) is \\( L^2 \\) length of the edge equation gradient which accounts for the Jacobian of the variable substitution, and \\( t \\) is the parametrization of the line along which we integrate." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.434, + 0.33, + 0.451 + ], + "angle": 0, + "content": "4 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.785, + 0.568 + ], + "angle": 0, + "content": "In this section we describe our method that we call EdgeGrad. First we will focus on the core assumptions and approximations that form the basis of our method. Then we will see that due to our specifically chosen approximation, we can apply the prior method of edge sampling to rasterization. From there, we will focus on edge detection and classification heuristics, and our discretization scheme. We will conclude by describing an extension of our method for handling geometry intersections." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.59, + 0.411, + 0.605 + ], + "angle": 0, + "content": "4.1 Rasterizing Edges" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.784, + 0.702 + ], + "angle": 0, + "content": "Our primary objective is to differentiate rasterization while preserving its core image formation principles. We aim to calculate the gradients of pixel values relative to the mesh vertex positions, including gradients caused by visibility changes such as occlusion and disocclusion by primitive geometry. We seek to do this with a simple, minimal, and efficient method that does not necessitate modifying the initially rasterized image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.703, + 0.784, + 0.763 + ], + "angle": 0, + "content": "Our main insight is that instead of approximating rasterization using a continuous representation, we assume the rasterized image is the result of some other continuous process that coincidentally aligns with our inherently non-differentiable, discrete-pixel rasterization. In essence, we opt for a direct ap" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.77, + 0.788, + 0.84 + ], + "angle": 0, + "content": "1 This concept is akin to representing real numbers with floating-point arithmetic. Although there are only 4,294,967,296 single-precision floating-point numbers (a quarter of which lie in the interval [0,1]), they are still employed for continuous computation. We treat these computations as smooth and differentiable, even though they are, in essence, discontinuous." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.401, + 0.128 + ], + "angle": 0, + "content": "S. Pidhorskyi et al." + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.143, + 0.485, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.432, + 0.497, + 0.445 + ], + "angle": 0, + "content": "Fig.1: Micro-edges and Pixel Pair." + }, + { + "type": "image_footnote", + "bbox": [ + 0.216, + 0.446, + 0.499, + 0.515 + ], + "angle": 0, + "content": "The figure illustrates construction of the micro-edges and the change in pixel intensities A and B with the edge position \\( p \\) movement, highlighting the C1 discontinuity at \\( p = 0 \\), absent in average intensity." + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.144, + 0.782, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.444, + 0.788, + 0.457 + ], + "angle": 0, + "content": "Fig. 2: Gradient computation flow." + }, + { + "type": "image_footnote", + "bbox": [ + 0.505, + 0.458, + 0.787, + 0.514 + ], + "angle": 0, + "content": "The figure illustrates the backward propagation process, where gradients from the loss function are traced back and scattered to the fragments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.55, + 0.788, + 0.655 + ], + "angle": 0, + "content": "proximation to compute gradients, instead of an indirect approach that first necessitates that the rendering be made continuous, for instance, through antialiasing filtering. More specifically, we use a construction where all straight edges, whether they are sides of triangles or intersections between triangles, are comprised of a collection of micro-edges. These micro-edges are assumed to be either strictly vertical or horizontal and are always positioned precisely between two pixels. An illustration of these constructions is shown in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.788, + 0.793 + ], + "angle": 0, + "content": "Our approximation leads to several key properties: a) boundaries never intersect pixels, only lying exactly between them; b) pixels are always either fully covered or not covered at all; c) there is at most one boundary between any two neighbouring pixels; d) we do not need to access the source geometry to evaluate where the edge is exactly. With this, we can reinterpret the problem as if the rasterized image was in fact obtained by applying an antialIASing filter, albeit without visible effects; the resulting image remains identical to the pre-antialIASing source. Essentially our construct of the micro-edges renders the antialIASing filter as an identity operation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.84 + ], + "angle": 0, + "content": "With this approximation, gradient estimation becomes a local operation, requiring the analysis one edge between two pixels at a time. Note that despite the approximation that all edges reside strictly between pixels, we do not assume" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.286, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Rasterized Edge Gradients: Handling Discontinuities Differently" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "that they are fixed, but rather allow them to freely move within an infinitesimal range, enabling differentiation of pixel values with respect to the edge location. This leads to the greatest advantage of our formulation; it allows us to greatly simplify both the theory and computation, since all edges are axis-aligned." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.229, + 0.353, + 0.243 + ], + "angle": 0, + "content": "4.2 Pixel Pairs" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.784, + 0.314 + ], + "angle": 0, + "content": "Let's analyze two adjacent pixels, A and B, separated by an edge, with their intensities represented as \\( I_{A} \\) and \\( I_{B} \\) respectively (refer to Figure 1 for details). The pixel intensities can be expressed using the Heaviside step function to account for the discontinuity at the edge as in Eqn. (4):" + }, + { + "type": "equation", + "bbox": [ + 0.3, + 0.325, + 0.785, + 0.391 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} I _ {A} = \\iint_ {D _ {A}} \\theta (\\alpha (x, y)) f _ {a} (x, y) + \\theta (- \\alpha (x, y)) f _ {b} (x, y) d x d y \\\\ I _ {B} = \\iint_ {D _ {B}} \\theta (- \\alpha (x, y)) f _ {a} (x, y) + \\theta (\\alpha (x, y)) f _ {b} (x, y) d x d y. \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.401, + 0.784, + 0.46 + ], + "angle": 0, + "content": "Without loss of generality, assuming a horizontal pair we define \\(\\alpha(x,y) = p_{AB} - x\\), where \\(p_{AB}\\) is the edge location. Edge location \\(p_{AB}\\) is now the scene parameter with respect to which we want to differentiate the loss function. Since \\(p_{AB}\\) effects both \\(I_A\\) and \\(I_B\\):" + }, + { + "type": "equation", + "bbox": [ + 0.385, + 0.46, + 0.784, + 0.49 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial L}{\\partial p _ {A B}} = \\frac {\\partial L}{\\partial I _ {A}} \\frac {\\partial I _ {A}}{\\partial p _ {A B}} + \\frac {\\partial L}{\\partial I _ {B}} \\frac {\\partial I _ {B}}{\\partial p _ {A B}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.496, + 0.785, + 0.613 + ], + "angle": 0, + "content": "where \\(\\frac{\\partial I_A}{\\partial p_{AB}}\\) and \\(\\frac{\\partial I_B}{\\partial p_{AB}}\\) derive by applying Eqn. (6) to Eqn. (7). Unfortunately, with the edge strictly on the pixel boundary, we face a \\(C1\\) discontinuity at \\(p_{AB} = 0\\), making intensity values non-differentiable. Infinitesimal movement of the edge affects \\(I_B\\) when moving right and \\(I_A\\) when moving left, without affecting the other. We still can compute one-sided limits of the derivatives, e.g. \\(\\frac{\\partial I_A}{\\partial p}^{-}\\) from the negative direction and \\(\\frac{\\partial I_B}{\\partial p}^{+}\\) from the positive direction. Applying Eqn. (6) and taking into account that the term \\(\\| \\nabla_{x,y}\\alpha (x,y)\\|\\) becomes 1 as we obtain:" + }, + { + "type": "equation", + "bbox": [ + 0.236, + 0.624, + 0.784, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial I _ {B}}{\\partial p _ {A B}} ^ {-} = \\frac {\\partial I _ {A}}{\\partial p _ {A B}} ^ {+} = 0; \\frac {\\partial I _ {B}}{\\partial p _ {A B}} ^ {+} = \\frac {\\partial I _ {A}}{\\partial p _ {A B}} ^ {-} = \\int_ {x = 0} [ f _ {a} (x, y) - f _ {b} (x, y) ] d y \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.666, + 0.784, + 0.696 + ], + "angle": 0, + "content": "We can circumvent the problem of \\( C_1 \\) discontinuity by averaging the one-sided limits of the derivatives from the left and right:" + }, + { + "type": "equation", + "bbox": [ + 0.336, + 0.707, + 0.784, + 0.738 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial I _ {B}}{\\partial p _ {A B}} = \\frac {\\partial I _ {A}}{\\partial p _ {A B}} = \\frac {1}{2} \\int_ {x = 0} [ f _ {a} (x, y) - f _ {b} (x, y) ] d y \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.747, + 0.784, + 0.793 + ], + "angle": 0, + "content": "Assuming that the pixels have constant value across all of their area (recall the rule that \"pixels are always either fully covered or not covered at all\") and substituting Eqn. (10) into Eqn. (8) we deduce:" + }, + { + "type": "equation", + "bbox": [ + 0.35, + 0.805, + 0.784, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\left. \\frac {\\partial L}{\\partial p _ {A B}} = \\frac {1}{2} \\left(\\frac {\\partial L}{\\partial I _ {A}} + \\frac {\\partial L}{\\partial I _ {B}}\\right) \\left(I _ {A} - I _ {B}\\right) + \\Omega \\right], \\tag {11}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.401, + 0.128 + ], + "angle": 0, + "content": "S. Pidhorskyi et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.223 + ], + "angle": 0, + "content": "which is one of our main results. Here, \\(\\frac{\\partial L}{\\partial p_{AB}}\\) is the gradient of the loss function with respect to the position of the edge between pixels \\(A\\) and \\(B\\). Additionally, we reintroduce \\(\\Omega\\) which is the second term on the right-hand side of Eqn. (5) that accounts for differentiating smooth regions of \\(f\\). This term was omitted in Eqn.(9) and Eqn. (10) for brevity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.785, + 0.314 + ], + "angle": 0, + "content": "Now, we need to accumulate these gradients into the gradients of the vertex positions of the triangle that created those edges. We do that first by scattering the gradients \\(\\frac{\\partial L}{\\partial p_{AB}}\\) to the fragments of the triangles associated with the pixels forming the edge and then by invoking the backward function of the interpolator module to gather the fragment gradients into the vertex gradients. Next, we will take a look at the gradient scattering mechanism." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.337, + 0.625, + 0.353 + ], + "angle": 0, + "content": "4.3 Edge Classification and Gradient Scattering" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.364, + 0.785, + 0.533 + ], + "angle": 0, + "content": "To backpropagate the gradients to the scene parameters we need to compute the gradient with respect to fragment positions, while in §4.2 we only derived the gradient \\(\\frac{\\partial L}{\\partial p_{AB}}\\) with respect to the edge position. We scatter each edge gradient \\(\\frac{\\partial L}{\\partial p_{AB}}\\) to adjacent fragments where we accumulate the fragment gradients, as illustrated in Figure 2. Using the chain rule, we obtain fragment gradients be multiplying \\(\\frac{\\partial L}{\\partial p_{AB}}\\) with \\(\\frac{\\partial p_{AB}}{\\partial r}\\); the later relates the movement of the edge to the movement of the fragment, where \\(r\\) is the position of the fragment. The computation of \\(\\frac{\\partial p_{AB}}{\\partial r}\\) depends on the type of the edge formed by the fragments. We consider several cases: overhanging primitives (either one another or background); adjacent primitives; intersecting primitives. Unlike Laine et al. [20], we avoid additional data structures like hash maps for connectivity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.786, + 0.761 + ], + "angle": 0, + "content": "Our approach examines each neighboring pixel pair, comparing their triangle IDs, see Figure 3 for illustration. Matching IDs indicate no edge; differing IDs prompt further tests. A single valid ID suggests a primitive overhanging the background. Testing a pixel's center against the triangle of the other pixel discerns overhangs and self-intersections. Non-contained pixels in each other's triangles imply adjacency. If both pixel centers are in the triangle of the other pixel, it implies self-intersections, and if in just one - then overhanging. For adjacent primitives we do not scatter gradients at all, since it is likely due to primitives sharing an edge and there is no discontinuity. Please note that even in that case \\(\\frac{\\partial L}{\\partial p_{AB}}\\) may be non-zero due to \\(\\Omega\\) contribution. In the overhanging case, we scatter the gradient only to the primitive on top, as the covered primitive can not affect the position of the edge. Thus, in such cases the covered fragment has \\(\\frac{\\partial p}{\\partial r} = 0\\) while the foreground fragment has \\(\\frac{\\partial p}{\\partial r} = 1\\) reflecting that the fragment movement is rigidly tied to the movement of the edge. Please refer to supplementary material for more details." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.784, + 0.458, + 0.799 + ], + "angle": 0, + "content": "4.4 Geometry Intersections" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.843 + ], + "angle": 0, + "content": "This subsection details computing the derivative of edge location with respect to fragment position \\(\\frac{\\partial p}{\\partial r}\\) in scenarios where the edge arises from intersecting" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.287, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Rasterized Edge Gradients: Handling Discontinuities Differently" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.143, + 0.496, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.462, + 0.498, + 0.545 + ], + "angle": 0, + "content": "Fig. 3: Edge classification. The accompanying figure provides a schematic representation of our edge classification process. Edges are categorized into four types: no edge, adjacent primitives, overhanging primitives, and intersecting primitives." + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.149, + 0.787, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.372, + 0.788, + 0.537 + ], + "angle": 0, + "content": "Fig. 4: Geometry intersection. A geometrical derivation for the partial derivative \\(\\frac{\\partial p}{\\partial r}\\), essential for converting edge-related gradients to fragment-related gradients in cases where edges are formed by geometry intersections. This shows the x-z plane, where the render camera is viewing the scene along the z axis so that edges are parallel to the y axis (and perpendicular to the drawing). Projections of both intersecting primitives are transformed into two lines." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.579, + 0.788, + 0.716 + ], + "angle": 0, + "content": "geometry. This situation is inherently complex, and a general-case solution would be challenging. However, our micro-edge formulation streamlines the problem because intersection lines projected onto the image plane comprised of microedges, e.i. either vertical or horizontal. This allows us to analyze the problem in 2D, focusing on the x-z or y-z planes, with the z-axis perpendicular to the image plane. We derive the derivative of the edge location by analyzing two fragments of a micro-edge that belongs to the intersection line, focusing on how the edge would move in response to infinitesimal movements of each fragment. Without loss of generality we consider the x-z plane:" + }, + { + "type": "equation", + "bbox": [ + 0.387, + 0.73, + 0.786, + 0.767 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial p}{\\partial r} = - \\mathbf {n} _ {z} ^ {f} \\left[ \\mathbf {n} ^ {f ^ {T}} \\left[ \\begin{array}{l l} 0 & 1 \\\\ - 1 & 0 \\end{array} \\right] \\mathbf {n} ^ {v} \\right] ^ {- 1}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.843 + ], + "angle": 0, + "content": "where \\( p \\) is edge position, \\( r \\) is the fragment position (in 3D clip space), \\( \\mathbf{n}^v \\) is the varying primitive's normal, and \\( \\mathbf{n}^f \\) is the fixed primitive's normal. In this context, the \\( x \\)-axis lies on the image plane, and the \\( z \\)-axis is perpendicular to it, see Fig. 4. We compute \\( \\frac{\\partial p}{\\partial r} \\) for both fragments that form the intersection edge by" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.401, + 0.128 + ], + "angle": 0, + "content": "S. Pidhorskyi et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "first fixing one fragment and varying the other, and then switching them. Please refer to supplementary material for derivation details." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.202, + 0.607, + 0.218 + ], + "angle": 0, + "content": "4.5 Theoretical analysis of the approximation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.23, + 0.784, + 0.306 + ], + "angle": 0, + "content": "At first glance, replacing arbitrarily oriented edges with vertical and horizontal micro-edges might seem overly simplistic, raising doubts about accuracy. It could be argued, for example, that the perimeter defined by micro-edges doesn't match the original as micro-edge size approaches zero. However, applying the two-dimensional version of the divergence theorem from vector calculus offers insight:" + }, + { + "type": "equation", + "bbox": [ + 0.388, + 0.319, + 0.786, + 0.35 + ], + "angle": 0, + "content": "\\[\n\\oint_ {C} \\mathbf {f} \\cdot \\mathbf {n} d t = \\iint_ {D} (\\nabla_ {\\Phi} \\cdot \\mathbf {f}) d x d y, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.508 + ], + "angle": 0, + "content": "where \\(\\mathbf{n}\\) is normal to boundary \\(C\\), within which the integrand function is continuous and differentiable. We are free to pick any such boundaries that includes a discontinuity. In our case, the term \\(\\frac{\\partial\\alpha(x,y)}{\\partial\\Phi}\\) in Eqn. (6) is a normal vector to the boundary \\(\\alpha (x,y) = 0\\) thus the integral is equivalent to the left-hand side of Eqn. (13). According to Eqn. (13), this recasts the boundary derivative integral as an area integral. The difference between the area integrals of micro-edges and original edges is confined to a space defined by the micro-edges size. As pixel size diminishes, the integral approaches the exact value, and the discrepancy area vanishes, confirming the micro-edge approach converges to the same solution as the original formulation in the limit." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.534, + 0.473, + 0.551 + ], + "angle": 0, + "content": "5 Experimental Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.785, + 0.688 + ], + "angle": 0, + "content": "We implement our method in PyTorch with custom CUDA kernels, and largely follow the structure of nvdiffrast [20], with the exception for the antialIASing module which we replace with our EdgeGrad module. For all experiments we utilize Laplacian preconditioning proposed by Nicolet et al. [34]. We find Laplacian preconditioning extremely helpful for stabilizing training and we use it for both, meshes and texture data. While Laplacian preconditioning may reduce self-intersections, this happens only due to much smoother optimization process, but it does not account for them in any way." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.786, + 0.841 + ], + "angle": 0, + "content": "Forward gradients. We first analyze the performance of the proposed approach by evaluating the forward gradients on a variety of test cases. Forward gradients are per-pixel derivatives with respect to some scene parameter, which is depicted with white arrows on the scene images. Fig. 5 shows the forward gradients computed using finite differences, raytracing methods Mitsuba [14,15,36] and redner [21], and rasterization approaches including nvdiffrast [20] and our approach (EdgeGrad). Unlike other methods, our approach still matches finite differences gradients in cases of intersecting geometry. While Mitsuba and Redner, which require multiple samples per pixel (256 in these comparisons), typically offer higher accuracy, they fall short at geometry intersections. Nvdiffrast" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.287, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Rasterized Edge Gradients: Handling Discontinuities Differently" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.144, + 0.773, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.437, + 0.785, + 0.493 + ], + "angle": 0, + "content": "Fig. 5: Comparison of forward gradient on several test scenes with the numerical solution using finite differences. (a) Synthetic test scene, we show gradients with respect to a parameter. We use (b) finite differences as a reference, and compare with (c) Mitsuba 3 [15,36], (d) redner [21], (e) nvdiffrast [20], and (f) our edge gradient approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.784, + 0.553 + ], + "angle": 0, + "content": "is noisier because it assumes triangles containing silhouette edges always overlap pixel centers, which often isn't the case, leading to omitted gradients." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.785, + 0.613 + ], + "angle": 0, + "content": "Backward gradients. Tab. 1 quantitatively analyzes the accuracy of the backward gradients. We show improved results especially for complex intersecting geometry. Compared to raytracing methods, our rasterization-based approach yields comparable performance at a fraction of the computational cost." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.785, + 0.704 + ], + "angle": 0, + "content": "Runtime performance end accuracy by image size. Fig. 6 shows running time of our method (EdgeGrad) compared to Mitsuba [36] and nvdiffrast [20]. Our method significantly outperforms other approaches in running time on all variety of image sizes and triangle count. Our method has comparable accuracy to Mitsuba [36] (and even better for the case with geometry intersections) at large image sizes, but performs worse for smaller image size due to sub pixel triangles." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Quantitative scene reconstructions. Tab. 2 shows quantitative results of scene reconstruction on blender dataset [32]. We follow the testing procedure described in [32] and report PNSR, SSIM and LPIPS metrics. We compare our method (EdgeGrad) with a \"continues only\" baseline (which ignores gradients from discontinuities), nvdiffrast [20], and with a variant of our method that does not have a special handling for the geometry intersections. Method optimization procedure is similar to Nicolet et al. [34] we only add MLP for decoding texture values in a view conditioned manner similar to [29]. We initialize geometry with some coarse manually created meshes shown in Fig. 7. To avoid additional complex" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.401, + 0.128 + ], + "angle": 0, + "content": "S. Pidhorskyi et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.244, + 0.185, + 0.301, + 0.195 + ], + "angle": 0, + "content": "Scene" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.169, + 0.359, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.17, + 0.413, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.17, + 0.468, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.212, + 0.47, + 0.257 + ], + "angle": 0, + "content": "
redner [21]1.20%53.52%37.94%
Mitsuba3 [36]0.67%75.83%37.59%
Nvdiffrast [20]45.91%69.33%39.42%
EdgeGrad(our)6.01%3.35%8.35%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.268, + 0.498, + 0.365 + ], + "angle": 0, + "content": "Table 1: Accuracy of backward gradients. Relative error, \\(\\%\\) (↓). This table shows relative errors in backward gradient computations for test scenes. Second and third scenes include geometry intersections, emphasizing our method's advantage in managing these complexities." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.146, + 0.644, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.211, + 0.637, + 0.218 + ], + "angle": 0, + "content": "(a) Running time vs image size" + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.147, + 0.781, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.211, + 0.78, + 0.218 + ], + "angle": 0, + "content": "(b) Running time vs triangle count" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.218, + 0.645, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.28, + 0.632, + 0.293 + ], + "angle": 0, + "content": "(c) Gradient error vs image size self-intersections absent" + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.219, + 0.781, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.28, + 0.771, + 0.293 + ], + "angle": 0, + "content": "(d) Gradient error vs image size self-intersections present" + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.305, + 0.787, + 0.388 + ], + "angle": 0, + "content": "Fig. 6: Runtime performance and errors. a) Runtime [ms] by image size [px]; b) runtime [ms] by triangle count; c) gradient error [%] by image size [px] without self-intersections; d) gradient error [%] by image size [px] with self-intersections." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.402, + 0.374, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.496, + 0.286, + 0.502 + ], + "angle": 0, + "content": "(a) camera image" + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.408, + 0.438, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.378, + 0.495, + 0.434, + 0.509 + ], + "angle": 0, + "content": "(c) initialization mesh" + }, + { + "type": "image", + "bbox": [ + 0.45, + 0.408, + 0.509, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.453, + 0.495, + 0.5, + 0.508 + ], + "angle": 0, + "content": "(d) optimized mesh" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.403, + 0.671, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.495, + 0.579, + 0.501 + ], + "angle": 0, + "content": "(e)rgb estimation" + }, + { + "type": "image_caption", + "bbox": [ + 0.594, + 0.496, + 0.652, + 0.51 + ], + "angle": 0, + "content": "(f) segmentation estimation" + }, + { + "type": "image", + "bbox": [ + 0.679, + 0.407, + 0.781, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.672, + 0.487, + 0.786, + 0.503 + ], + "angle": 0, + "content": "(g) Learned segmentation texture for mouth region" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.509, + 0.268, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.235, + 0.556, + 0.269, + 0.562 + ], + "angle": 0, + "content": "Initialization" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.509, + 0.324, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.556, + 0.324, + 0.562 + ], + "angle": 0, + "content": "Optimized" + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.513, + 0.419, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.347, + 0.555, + 0.381, + 0.561 + ], + "angle": 0, + "content": "Initialization" + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.513, + 0.474, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.434, + 0.555, + 0.463, + 0.561 + ], + "angle": 0, + "content": "1" + }, + { + "type": "image", + "bbox": [ + 0.475, + 0.513, + 0.559, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.556, + 0.496, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.556, + 0.549, + 0.562 + ], + "angle": 0, + "content": "Optimized" + }, + { + "type": "image", + "bbox": [ + 0.569, + 0.513, + 0.613, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.556, + 0.606, + 0.562 + ], + "angle": 0, + "content": "Initialization" + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.513, + 0.667, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.556, + 0.66, + 0.562 + ], + "angle": 0, + "content": "Optimized" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.511, + 0.786, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.556, + 0.717, + 0.562 + ], + "angle": 0, + "content": "Initialization" + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.556, + 0.771, + 0.562 + ], + "angle": 0, + "content": "Optimized" + }, + { + "type": "image_caption", + "bbox": [ + 0.378, + 0.563, + 0.636, + 0.571 + ], + "angle": 0, + "content": "(h) Initialization mesh and reconstructed mesh for synthetic blender dataset" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.582, + 0.788, + 0.653 + ], + "angle": 0, + "content": "Fig. 7: Experimental setup details. The figure shows: a) the ground truth camera image; b) predicted segmentation mask for mouth region using HRNet [47]; c) initialization mesh from 3DMM; d) optimized avatar mesh; e) rendered image of the optimized model; f) rendered segmentation estimation image; g) learned static segmentation texture; h) initialization meshes and optimized meshes for blender dataset [32]." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.667, + 0.264, + 0.742 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.745, + 0.244, + 0.75 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.668, + 0.311, + 0.743 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.271, + 0.745, + 0.295, + 0.75 + ], + "angle": 0, + "content": "NvdiffRast" + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.666, + 0.373, + 0.744 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.316, + 0.745, + 0.355, + 0.75 + ], + "angle": 0, + "content": "No Edge Handling" + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.674, + 0.421, + 0.739 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.744, + 0.417, + 0.75 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.675, + 0.474, + 0.739 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.443, + 0.744, + 0.466, + 0.75 + ], + "angle": 0, + "content": "NvdRast" + }, + { + "type": "image", + "bbox": [ + 0.475, + 0.675, + 0.532, + 0.739 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.487, + 0.744, + 0.527, + 0.75 + ], + "angle": 0, + "content": "No Edge Handling" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.679, + 0.616, + 0.747 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.745, + 0.588, + 0.75 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.679, + 0.695, + 0.745 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.745, + 0.674, + 0.75 + ], + "angle": 0, + "content": "NdivdiffRast" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.666, + 0.78, + 0.742 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.744, + 0.764, + 0.75 + ], + "angle": 0, + "content": "No Edge Handling" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.764, + 0.788, + 0.834 + ], + "angle": 0, + "content": "Fig. 8: Reconstructions from inverse rendering compare our method with nvdiffrast [20] and rasterization without discontinuity handling. Although nvdiffrast can match the outline of objects well, it suffers from interpenetration (see leaves of Ficus) and does not capture as much detail as ours (see the backrest of the chair and ground of lego). The method with no edge handling fails completely on these objects." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.287, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Rasterized Edge Gradients: Handling Discontinuities Differently" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.235, + 0.144, + 0.775, + 0.302 + ], + "angle": 0, + "content": "
LegoChairShipHotdogFicusDrumsMicMaterials
PSNR ↑Continuous only16.34228.15115.69327.85018.39915.53421.79717.11
Nvdiffrast [20]29.4429.80426.17331.68424.96423.15429.63824.413
EdgeGrad -intersect.29.57433.07726.73134.22526.78124.0831.22425.221
EdgeGrad (our)29.66732.98126.6534.08126.55924.01431.34625.287
LegoChairShipHotdogFicusDrumsMicMaterials
SSIM ↑Continuous only0.75500.95310.69290.93740.82450.77230.91820.7693
Nvdiffrast [20]0.94670.94160.83160.94730.93230.89710.96470.8865
EdgeGrad -intersect.0.94850.97650.84780.97150.94680.92160.97560.8945
EdgeGrad (our)0.95010.97640.84720.97130.94520.92200.97670.8972
LegoChairShipHotdogFicusDrumsMicMaterials
LPIPS ↓Continuous only0.32230.06110.39130.11090.20460.29430.12720.2511
Nvdiffrast [20]0.06530.07710.20200.09360.07810.11160.05490.1291
EdgeGrad -intersect.0.06330.03480.18990.05460.06810.08710.04390.1200
EdgeGrad (our)0.06140.03410.18500.05480.06720.08700.03990.1185
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.303, + 0.786, + 0.345 + ], + "angle": 0, + "content": "Table 2: Per-scene quantitative results from the realistic synthetic blender dataset [32], featuring scenes with complex geometries and non-Lambertian materials rendered using Blender's Cycles pathtracer." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.37, + 0.787, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.459, + 0.785, + 0.488 + ], + "angle": 0, + "content": "Fig. 9: Example renderings of mesh-based avatars from [29] trained using EdgeGrad, capturing complex inner mouth movements, including teeth and tongue dynamics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.516, + 0.785, + 0.546 + ], + "angle": 0, + "content": "ity we opted out automatic remeshing introduced by Nicolet et al. which would necessitate updating UV parametrization in our case." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.547, + 0.787, + 0.621 + ], + "angle": 0, + "content": "Qualitative mesh reconstruction. Qualitatively, we compare the mesh reconstructions obtained on the public NeRF dataset [32] using our approach, a naive approach with no special edge handling, and nvdiffrast [20] in Fig. 8. All methods use the same shading model. Our approach shows improved detail in the recovered meshes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.644, + 0.613, + 0.661 + ], + "angle": 0, + "content": "6 Application to Dynamic Head Scenes" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.81 + ], + "angle": 0, + "content": "We show the importance of correct edge gradients on a head modeling application by fitting a personalized mesh-based avatar representation to facial performances captured in a multiview lightstage. modifying the avatar creation method from [29]. As the facial avatar representation, we use an encoder-decoder architecture to produce mesh vertices and texture-mapped appearance features [25], which are shaded using a pixel-wise MLP producing RGB appearance [29]. We use a coarse face tracking produced using traditional computer vision as initialization for inverse rendering. Importantly, this coarse face template has a small number of vertices (8k) and no geometry in the interior of the mouth." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Traditionally, fitting the interior of the mouth without strong priors is a complex task due to poor observability and intricately interacting geometry with" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.401, + 0.128 + ], + "angle": 0, + "content": "S. Pidhorskyi et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.144, + 0.307, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.238, + 0.304, + 0.302, + 0.312 + ], + "angle": 0, + "content": "(a) camera image" + }, + { + "type": "image", + "bbox": [ + 0.308, + 0.145, + 0.393, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.331, + 0.304, + 0.375, + 0.312 + ], + "angle": 0, + "content": "(b) Nvdiffrast" + }, + { + "type": "image", + "bbox": [ + 0.393, + 0.145, + 0.477, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.409, + 0.304, + 0.469, + 0.312 + ], + "angle": 0, + "content": "(c) EdgeGrad (our)" + }, + { + "type": "image", + "bbox": [ + 0.478, + 0.145, + 0.563, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.475, + 0.304, + 0.562, + 0.315 + ], + "angle": 0, + "content": "(d) EdgeGrad (our) w/o segmentation supervision" + }, + { + "type": "image_caption", + "bbox": [ + 0.578, + 0.147, + 0.788, + 0.313 + ], + "angle": 0, + "content": "Fig. 10: Qualitative evaluation and ablation on dynamic head scene. We compare: (c) our method, (a) ground truth images, (b) Nvdiffrast [20], and ablated version of our method without including segmentation mask. The avatar model is from [29] trained using EdgeGrad and Nvdiffrast. We highlight artifacts with white arrows." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.789, + 0.617 + ], + "angle": 0, + "content": "constant contact. Prior work either relies on artists to create alignments, volumetric simulation to prevent self-intersections [13], or on very specific mouth interior capture methods to build priors [30,49]. Instead, we show that by using differentiable rasterization with correct edge gradients, we can achieve fully automatic geometric fitting of the complex geometry of the face and mouth interior including lips, teeth, and tongue using a mesh-based model of fixed topology. We modify [29] in two ways. First, we use the edge gradient rasterization method of Sec. 4. Second, we add supervision in the form of 2D segmentation masks with separate labels for the regions of the teeth, tongue, and lips, for details see Fig. 7. While we can achieve accurate per-frame geometry without the additional segmentation loss supervision, we find that segmentation losses help anchor the teeth and tongue geometry to a consistent region of the UV map. Fig. 9 shows the quality of geometric fitting achievable with this inverse rendering method. Note that we accurately reconstruct the complex contact geometry of the tongue sliding across teeth, teeth-to-teeth contact, as well as the tongue pushing against the lips and cheeks. Fig. 10 demonstrates reconstruction quality of complex inner mouth details, which are challenging due to frequent self-intersections. We also include results without 2D segmentation mask supervision." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.641, + 0.528, + 0.658 + ], + "angle": 0, + "content": "7 Conclusions and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.842 + ], + "angle": 0, + "content": "We have presented a fast method for rasterization-based differentiable rendering that has accurate gradients across visibility discontinuities such as triangle edges, occlusions and geometry intersections. Our results demonstrate enhanced mesh reconstruction fidelity across diverse inverse rendering scenarios, comparable to implicit and volumetric methods. Our method also unlocks new applications, accurate mouth interior fitting in dynamic facial performance capture, through fully automatic inverse rendering. Regarding limitations, our method inherits the constraints of rasterization, including the inability to manage transparency, lack of precise antialIASing without multisampling, and the absence of physically based rendering and global illumination. It may also inaccurately interpret adjacent primitives in finely tessellated meshes, leading to sub-pixel triangles." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.286, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rasterized Edge Gradients: Handling Discontinuities Differently" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.176, + 0.786, + 0.218 + ], + "angle": 0, + "content": "1. Baek, S., Kim, K.I., Kim, T.K.: Pushing the envelope for rgb-based dense 3d hand pose estimation via neural rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1067-1076 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.787, + 0.26 + ], + "angle": 0, + "content": "2. Bangaru, S.P., Gharbi, M., Luan, F., Li, T.M., Sunkavalli, K., Hasan, M., Bi, S., Xu, Z., Bernstein, G., Durand, F.: Differentiable rendering of neural sdfs through reparameterization. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.787, + 0.287 + ], + "angle": 0, + "content": "3. Bangaru, S.P., Li, T.M., Durand, F.: Unbiased warped-area sampling for differentiable rendering. ACM Transactions on Graphics (TOG) 39(6), 1-18 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.787, + 0.342 + ], + "angle": 0, + "content": "4. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.787, + 0.398 + ], + "angle": 0, + "content": "5. Beker, D., Kato, H., Morariu, M.A., Ando, T., Matsuoka, T., Kehl, W., Gaidon, A.: Monocular differentiable rendering for self-supervised 3d object detection. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16. pp. 514-529. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.399, + 0.787, + 0.467 + ], + "angle": 0, + "content": "6. Bogo, F., Kanazawa, A., Lassner, C., Gehler, P., Romero, J., Black, M.J.: Keep it spl! Automatic estimation of 3d human pose and shape from a single image. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14. pp. 561-578. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.468, + 0.787, + 0.522 + ], + "angle": 0, + "content": "7. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.523, + 0.787, + 0.564 + ], + "angle": 0, + "content": "8. Chen, W., Ling, H., Gao, J., Smith, E., Lehtinen, J., Jacobson, A., Fidler, S.: Learning to predict 3d objects with an interpolation-based differentiable renderer. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.787, + 0.619 + ], + "angle": 0, + "content": "9. Chen, W., Litalien, J., Gao, J., Wang, Z., Fuji Tsang, C., Khamis, S., Litany, O., Fidler, S.: Dib-r++: learning to predict lighting and material with a hybrid differentiable renderer. Advances in Neural Information Processing Systems 34, 22834-22848 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.787, + 0.675 + ], + "angle": 0, + "content": "0. Choy, C.B., Xu, D., Gwak, J., Chen, K., Savarese, S.: 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14. pp. 628-644. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.787, + 0.716 + ], + "angle": 0, + "content": "1. Ge, L., Ren, Z., Li, Y., Xue, Z., Wang, Y., Cai, J., Yuan, J.: 3d hand shape and pose estimation from a single rgb image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10833-10842 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.787, + 0.771 + ], + "angle": 0, + "content": "2. Henzler, P., Reizenstein, J., Labatut, P., Shapovalov, R., Ritschel, T., Vedaldi, A., Novotny, D.: Unsupervised learning of 3d object categories from videos in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4700-4709 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.787, + 0.799 + ], + "angle": 0, + "content": "3. Ichim, A.E., Kadleček, P., Kavan, L., Pauly, M.: Phace: Physics-based face modeling and animation. ACM Trans. Graph. 36(4) (jul 2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "4. Jakob, W., Speierer, S., Roussel, N., Nimier-David, M., Vicini, D., Zeltner, T., Nicolet, B., Crespo, M., Leroy, V., Zhang, Z.: Mitsuba 3 renderer (2022), https://mitsuba-renderer.org" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.176, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.401, + 0.128 + ], + "angle": 0, + "content": "S. Pidhorskyi et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "15. Jakob, W., Speierer, S., Roussel, N., Vicini, D.: Dr.jit: A just-in-time compiler for differentiable rendering. Transactions on Graphics (Proceedings of SIGGRAPH) 41(4) (Jul 2022). https://doi.org/10.1145/3528223.3530099" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.23 + ], + "angle": 0, + "content": "16. Jiang, Y., Ji, D., Han, Z., Zwicker, M.: Sdiff: Differentiable rendering of signed distance fields for 3d shape optimization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1251-1261 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.785, + 0.27 + ], + "angle": 0, + "content": "17. Kato, H., Ushiku, Y., Harada, T.: Neural 3d mesh renderer. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3907-3916 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.785, + 0.311 + ], + "angle": 0, + "content": "18. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG) 42(4), 1-14 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.311, + 0.785, + 0.365 + ], + "angle": 0, + "content": "19. de La Gorce, M., Fleet, D.J., Paragios, N.: Model-based 3d hand pose estimation from monocular video. IEEE Trans. Pattern Anal. Mach. Intell. 33(9), 1793-1805 (Sep 2011). https://doi.org/10.1109/TPAMI.2011.33, http://dx.doi.org/10.1109/TPAMI.2011.33" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.365, + 0.785, + 0.407 + ], + "angle": 0, + "content": "20. Laine, S., Hellsten, J., Karras, T., Seol, Y., Lehtinen, J., Aila, T.: Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG) 39(6), 1-14 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.407, + 0.785, + 0.447 + ], + "angle": 0, + "content": "21. Li, T.M., Aittala, M., Durand, F., Lehtinen, J.: Differentiable monte carlo ray tracing through edge sampling. ACM Trans. Graph. (Proc. SIGGRAPH Asia) 37(6), 222:1-222:11 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.785, + 0.502 + ], + "angle": 0, + "content": "22. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 300–309 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.529 + ], + "angle": 0, + "content": "23. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.529, + 0.785, + 0.57 + ], + "angle": 0, + "content": "24. Liu, S., Li, T., Chen, W., Li, H.: Soft rasterizer: A differentiable renderer for image-based 3d reasoning. The IEEE International Conference on Computer Vision (ICCV) (Oct 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.785, + 0.597 + ], + "angle": 0, + "content": "25. Lombardi, S., Saragih, J., Simon, T., Sheikh, Y.: Deep appearance models for face rendering. ACM Transactions on Graphics (TOG) 37(4), 1-13 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.596, + 0.785, + 0.637 + ], + "angle": 0, + "content": "26. Lombardi, S., Simon, T., Saragih, J., Schwartz, G., Lehrmann, A., Sheikh, Y.: Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.678 + ], + "angle": 0, + "content": "27. Loper, M.M., Black, M.J.: Opendra: An approximate differentiable renderer. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VII 13. pp. 154-169. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.785, + 0.718 + ], + "angle": 0, + "content": "28. Loubet, G., Holzschuch, N., Jakob, W.: Reparameterizing discontinuous integrands for differentiable rendering. Transactions on Graphics (Proceedings of SIGGRAPH Asia) 38(6) (Dec 2019). https://doi.org/10.1145/3355089.3356510" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "29. Ma, S., Simon, T., Saragih, J.M., Wang, D., Li, Y., la Torre, F.D., Sheikh, Y.: Pixel codec avatars. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 64-73 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.8 + ], + "angle": 0, + "content": "30. Medina, S., Tomé, D., Stoll, C., Tiede, M., Munhall, K., Hauptmann, A., Matthews, I.: Speech driven tongue animation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE/CVF (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.841 + ], + "angle": 0, + "content": "31. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.287, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rasterized Edge Gradients: Handling Discontinuities Differently" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "32. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.217 + ], + "angle": 0, + "content": "33. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.217, + 0.788, + 0.242 + ], + "angle": 0, + "content": "34. Nicolet, B., Jacobson, A., Jakob, W.: Large steps in inverse rendering of geometry. ACM Transactions on Graphics (TOG) 40(6), 1-13 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.242, + 0.788, + 0.297 + ], + "angle": 0, + "content": "35. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3504-3515 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.297, + 0.788, + 0.338 + ], + "angle": 0, + "content": "36. Nimier-David, M., Vicini, D., Zeltner, T., Jakob, W.: Mitsuba 2: A retargetable forward and inverse renderer. Transactions on Graphics (Proceedings of SIGGRAPH Asia) 38(6) (Dec 2019). https://doi.org/10.1145/3355089.3356498" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.338, + 0.788, + 0.393 + ], + "angle": 0, + "content": "37. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.393, + 0.788, + 0.434 + ], + "angle": 0, + "content": "38. Pavlakos, G., Zhu, L., Zhou, X., Daniilidis, K.: Learning to estimate 3d human pose and shape from a single color image. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 459-468 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.46 + ], + "angle": 0, + "content": "39. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.46, + 0.788, + 0.514 + ], + "angle": 0, + "content": "40. Rhodin, H., Robertini, N., Richardt, C., Seidel, H.P., Theobalt, C.: A versatile scene model with differentiable visibility applied to generative pose estimation. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 765-773 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.514, + 0.788, + 0.569 + ], + "angle": 0, + "content": "41. Roessle, B., Barron, J.T., Mildenhall, B., Srinivasan, P.P., Nießner, M.: Dense depth priors for neural radiance fields from sparse input views. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12892-12901 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.61 + ], + "angle": 0, + "content": "42. Roveri, R., Öztireli, A.C., Pandele, I., Gross, M.: Pointpronets: Consolidation of point clouds with convolutional neural networks. In: Computer Graphics Forum. vol. 37, pp. 87-99. Wiley Online Library (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.61, + 0.788, + 0.637 + ], + "angle": 0, + "content": "43. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.637, + 0.788, + 0.677 + ], + "angle": 0, + "content": "44. Tsalicoglou, C., Manhardt, F., Tonioni, A., Niemeyer, M., Tombari, F.: Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.677, + 0.788, + 0.718 + ], + "angle": 0, + "content": "45. Tulsiani, S., Zhou, T., Efros, A.A., Malik, J.: Multi-view supervision for single-view reconstruction via differentiable ray consistency. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2626-2634 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.718, + 0.788, + 0.745 + ], + "angle": 0, + "content": "46. Vicini, D., Speierer, S., Jakob, W.: Differentiable signed distance function rendering. ACM Transactions on Graphics (TOG) 41(4), 1-18 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.745, + 0.788, + 0.799 + ], + "angle": 0, + "content": "47. Wang, J., Sun, K., Cheng, T., Jiang, B., Deng, C., Zhao, Y., Liu, D., Mu, Y., Tan, M., Wang, X., et al.: Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence 43(10), 3349-3364 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.788, + 0.841 + ], + "angle": 0, + "content": "48. Wiles, O., Gkioxari, G., Szeliski, R., Johnson, J.: Synsin: End-to-end view synthesis from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7467-7477 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.401, + 0.128 + ], + "angle": 0, + "content": "S. Pidhorskyi et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "49. Wu, C., Bradley, D., Garrido, P., Zollhöfer, M., Theobalt, C., Gross, M., Beeler, T.: Model-based teeth reconstruction. ACM Trans. Graph. 35(6) (dec 2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.217 + ], + "angle": 0, + "content": "50. Yan, X., Yang, J., Yumer, E., Guo, Y., Lee, H.: Perspective transformer nets: Learning single-view 3d object reconstruction without 3d supervision. Advances in neural information processing systems 29 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "51. Yang, G., Cui, Y., Belongie, S., Hariharan, B.: Learning single-view 3d reconstruction with limited pose supervision. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 86-101 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.285 + ], + "angle": 0, + "content": "52. Yu, A., Fridovich-Keil, S., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "53. Zhang, C., Wu, L., Zheng, C., Gkioulekas, I., Ramamoorthi, R., Zhao, S.: A differential theory of radiative transfer. ACM Trans. Graph. 38(6), 227:1-227:16 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "54. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.397 + ], + "angle": 0, + "content": "55. Zhao, S., Jakob, W., Li, T.M.: Physics-based differentiable rendering: From theory to implementation. In: ACM SIGGRAPH 2020 Courses. SIGGRAPH '20, Association for Computing Machinery, New York, NY, USA (2020). https://doi.org/10.1145/3388769.3407454, https://doi.org/10.1145/3388769.3407454" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.397 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Rasterized Edge Gradients_ Handling Discontinuities Differentially/ac83eb0e-ca7d-4471-8db5-05179e149303_origin.pdf b/2024/Rasterized Edge Gradients_ Handling Discontinuities Differentially/ac83eb0e-ca7d-4471-8db5-05179e149303_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f4ed3cc241ced74bd9bab9ac7ff5e73fa256e3f8 --- /dev/null +++ b/2024/Rasterized Edge Gradients_ Handling Discontinuities Differentially/ac83eb0e-ca7d-4471-8db5-05179e149303_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb457de98081fac20d892604e5c9fbbb9d3dd816deab2f7fd9f37b3fc08053ee +size 6967789 diff --git a/2024/Rasterized Edge Gradients_ Handling Discontinuities Differentially/full.md b/2024/Rasterized Edge Gradients_ Handling Discontinuities Differentially/full.md new file mode 100644 index 0000000000000000000000000000000000000000..43295ddf73af57140e131602c018e42db27b0848 --- /dev/null +++ b/2024/Rasterized Edge Gradients_ Handling Discontinuities Differentially/full.md @@ -0,0 +1,382 @@ +# Rasterized Edge Gradients: Handling Discontinuities Differently + +Stanislav Pidhorskyi $^{\text{①}}$ , Tomas Simon $^{\text{②}}$ , Gabriel Schwartz $^{\text{③}}$ , He Wen, Yaser Sheikh, and Jason Saragih + +Reality Labs,Meta,Pittsburgh, Pennsylvania,USA + +Abstract. Computing the gradients of a rendering process is paramount for diverse applications in computer vision and graphics. However, accurate computation of these gradients is challenging due to discontinuities and rendering approximations, particularly for surface-based representations and rasterization-based rendering. We present a novel method for computing gradients at visibility discontinuities for rasterization-based differentiable renderers. Our method elegantly simplifies the traditionally complex problem through a carefully designed approximation strategy, allowing for a straightforward, effective, and performant solution. We introduce a novel concept of micro-edges, which allows us to treat the rasterized images as outcomes of a differentiable, continuous process aligned with the inherently non-differentiable, discrete-pixel rasterization. This technique eliminates the necessity for rendering approximations or other modifications to the forward pass, preserving the integrity of the rendered image, which makes it applicable to rasterized masks, depth, and normals images where filtering is prohibitive. Utilizing micro-edges simplifies gradient interpretation at discontinuities and enables handling of geometry intersections, offering an advantage over the prior art. We showcase our method in dynamic human head scene reconstruction, demonstrating effective handling of camera images and segmentation masks. + +# 1 Introduction + +Significant advances have been made in recent years in modeling real 3D objects from image measurements. Much of this advancement can be credited to improvements in inverse rendering, which enables the automatic inference of 3D scene parameters that best reconstruct the images. While volumetric methods like NeRF [31] simplify inverse rendering by eliminating the need for predefined scene topology, classical mesh-based representations remain widely used due to their efficiency in modeling opaque surfaces. These representations often rely on highly performant rasterization, which requires discrete operations to reason about ordering and coverage, which poses challenges for gradient computations. Our main contribution is a theoretical framework for approximating gradients of a rasterization process that simultaneously elucidates previous constructions, arrives at a simple, fast, and accurate formulation, and further improves gradient approximation accuracy over state-of-the-art methods. + +In rasterization, a rendered pixel's value is determined by the foremost triangle covering it. Pixels whose footprint lies entirely within a single triangle are straightforward to handle, as infinitesimal mesh motion does not change their triangle membership. However, pixels on triangle boundaries, either between adjacent triangles, at occlusion boundaries, or at intersections between triangles, must account for how their triangle membership changes in response to infinitesimal changes in the mesh. Existing methods employ different approximations of soft-membership to account for this, enabling gradient computation in those areas, effectively compositing contributions from member triangles via a weighted sum. For example, [20] uses the anti-aliasing approach, [14] averages montecarlo samples during ray-casting, and [40] extends the influence of boundary triangles using a falloff function. In this work, we show that all these methods exhibit approximation errors that can fail to compute correct gradients in one case or another. In particular, none of the existing methods accurately compute the gradient at triangle intersections, which can often happen during inverse rendering, especially when the current estimate is far from the desired solution. + +In contrast to existing works that directly tackle geometric discontinuities and rasterization's discrete nature, we introduce a micro-edge formulation which allows to interpret the rasterized image as an outcome of a continuous process that coincidentally aligns with discrete-pixel rasterization, simplifying gradient computation significantly. Unlike some other works [20, 40], our formulation achieves good gradient estimates without altering the rasterization forward pass, maintaining the rasterized image's integrity. This is crucial for optimizing segmentation masks, depth maps, and normal maps where filtering or smoothing is not feasible, e.g. anisotropic filtering or soft rasterization would mix normals from different surfaces misrepresenting the geometry. The simplifications offered by our micro-edge formulation allow us to seamlessly handle discontinuities caused by geometry intersections, offering an advantage over the prior art. + +In summary, our paper introduces a straightforward, accurate, and efficient method for computing gradients in rasterization, comparable to existing techniques but with greater simplicity and the ability to handle self-intersections. We analyze our method's errors from a theoretical standpoint and compare gradients qualitatively and quantitatively with finite differences and other methods across various test cases. We also assess runtime efficiency and accuracy by image size and showcase qualitative and quantitative findings on a synthetic blender dataset [32]. Our method excels in complex applications, such as detailed dynamic human head reconstructions, effectively managing the intricate details of the inner mouth region with significant occlusions and deformations. + +# 2 Related work + +Differentiable rendering is one of the critical building blocks for many existing computer vision problems, such as 3d object reconstruction [10, 17, 45, 50, 51], 3d object prediction [5, 8, 9, 12], pose estimation [1, 6, 11, 38] novel view synthesis [4, 18, 32, 33, 54], as well as newly emerging applications such as text-to-3D + +generative models [22, 39, 43, 44]. Differentiable rendering employs a wide array of underlying 3D representations including explicit surfaces like mesh-based representations [19, 20, 27], implicit surfaces [2, 16, 37, 46], point clouds [18, 42, 48], explicit or implicit volume representations [4, 23, 26, 32, 35, 50, 52], and hybrid representations [7, 33]. For a comprehensive introduction to differentiable rendering, we refer the reader to Zhao et al.'s excellent SIGGRAPH course notes [55]. + +Mesh optimization often poses more challenges than implicit and volumetric methods, leading to the latter's prevalence in the field, as noted by Roessle et al. [41]. Despite this, meshes excel in high-performance rendering and imply registration with predefined topologies. Mesh rendering methods can be categorized into ray tracing-based and rasterization-based, with the former including frameworks like Mitsuba 3 [15, 36]. Differentiable ray tracing, aimed at direct illumination or path tracing, is computationally expensive but principled. Key advancements by Li et al. [21] and Zhang et al. [53] split this gradient into discontinuity integrals with Dirac delta functions and differentiable continuous parts. Li et al. replaced Dirac delta integration with boundary line integration, while Zhang et al. applied Reynold's transport theorem for a similar decomposition. Both methods are intricate, necessitating silhouette edge sampling — a significant bottleneck in their application. Loubet et al. [28] proposed a variable change to simplify discontinuity integration, further refined by Bangaru et al. [3] using warped area fields to correct biases, yet practical application challenges persist. Despite their flexibility, the computational expense of such ray tracing frameworks restrict their practicality in optimization-intensive tasks [20]. + +For rasterization-based methods, rendering meshes differentiably is hindered by non-differentiable visibility at boundaries. Unlike ray tracing, which uses boundary integrals for visibility gradients, rasterization's fixed-grid and z-buffer approach lacks this feature. Solutions typically involve approximating derivatives or approximating rendering to achieve inherent differentiability. De La Gorce et al. [19] and Loper et al. [27] pioneered derivative approximation in rasterization. De La Gorce et al. distinguished between continuous regions and occlusion boundaries, introducing 'occlusion forces' for the latter. Loper et al. simplified this by detecting discontinuities post-rasterization and employing differential filters, like the Sobel filter, for boundary approximation. + +The 'Nvdiffrast' approach by Laine et al. [20] tackles point-sampled visibility issues with innovative differentiable analytic antialIASing, which transforms sharp discontinuities into smooth transitions, enabling gradient computation. Analytical antialIASing is achieved by estimating coverage using silhouette edges. A limitation of the method is that if triangles containing silhouette edges do not overlap with any pixel center, they will be overlooked during the antialIASing process. Moreover, analytical antialIASing is inherently approximate, so the obtained gradients are also approximate. Regrettably, the approach is still quite complex, modifies the rasterized image, and requires a specialized data structure for connectivity and detailed pixel coverage computations. On the other hand, methods like Rhodin et al. [40] and Liu et al.'s 'Soft Rasterizer' [24] opt for modifying the rendering model. Rhodin et al. introduce fuzzy edges to soften discontinuities, + +while 'Soft Rasterizer' further blurs boundaries and averages depth contributions, facilitating gradient propagation across occluded primitives. + +In this work, we deliberately bypass physically based rendering, global illumination, and lighting/material models, and instead focus exclusively on rasterization-based approaches for their speed. We draw inspiration from the core idea of the edge sampling method [21] and analytical antialIASing [20] and distill this into a much simpler yet effective method that also provides accurate gradients. Notably, our method's simplicity allows direct handling of self-penetrating geometry, an aspect not addressed in previous research. Other methods to handle self-penetrating geometry would require complex and computationally intensive geometry preprocessing at each optimization step, involving intersection detection and differentiable splitting of the faces. In contrast, our approach explicitly manages interpenetrating geometry with virtually no additional overhead. + +# 3 Preliminaries + +Following Li et al. [21], given a 2D pixel filter $k$ and radiance $R$ , a pixel's color can be written: + +$$ +I = \iint_ {D} k (x, y) R (x, y) d x d y. \tag {1} +$$ + +For notational convenience, as in [21], we denote $f(x,y) = k(x,y)R(x,y)$ . For simplicity, let us first consider the case of a single pixel. We are interested in the gradient of $\frac{\partial L(I)}{\partial\varPhi}$ , where the scalar function $L(I)$ of the rendered pixel $I$ defines our loss, and $\varPhi$ are the scene parameters that we aim to optimize. According to the chain rule: + +$$ +\frac {\partial L (I)}{\partial \Phi} = \frac {\partial I}{\partial \Phi} \frac {\partial L}{\partial I}, \tag {2} +$$ + +where $\frac{\partial L}{\partial I}$ is the incoming gradient from the loss function and is typically computed by automatic differentiation (AD). So, our goal becomes to find how the pixel value changes with respect to the scene parameters: + +$$ +\frac {\partial I}{\partial \Phi} = \frac {\partial}{\partial \Phi} \iint_ {D} f (x, y) d x d y. \tag {3} +$$ + +While the function $f$ may not in fact be differentiable, its integral remains continuous and thus differentiable. Let us assume that $f$ is partitioned into two continuous half-spaces, each represented by two continuous and differentiable functions $f_{a}$ and $f_{b}$ , respectively, as follows: + +$$ +f (x, y) = \theta (\alpha (x, y)) f _ {a} (x, y) + \theta (- \alpha (x, y)) f _ {b} (x, y), \tag {4} +$$ + +where $\alpha$ specifies the dividing edge, and $\theta$ is the Heaviside step function, which selects between $f_{a}$ and $f_{b}$ . The differentiation of the integral can be broken into two parts by applying the product rule. Let us start with the integrand $f_{a}$ : + +$$ +\begin{array}{l} \frac {\partial}{\partial \Phi} \iint_ {D} \theta (\alpha (x, y)) f _ {a} (x, y) d x d y \\ = \iint_ {D} \delta (\alpha (x, y)) \frac {\partial \alpha (x , y)}{\partial \Phi} f _ {a} (x, y) d x d y + \iint_ {D} \frac {\partial f _ {a} (x , y)}{\partial \Phi} \theta (\alpha (x, y)) d x d y, \tag {5} \\ \end{array} +$$ + +and a similar derivation can be done for $f_{b}$ . The second term on the right-hand side of Eqn. (5) is the integral of the derivative of the smooth function, $f_{a}(x,y)$ . This derivative is readily computable with AD, and in the context of differentiable rasterizers it typically involves differentiating through the interpolation step, the barycentric coordinates, and the texture sampler. Due to its straightforward computation, we will represent it as $\Omega$ and omit it from further discussion, focusing instead on the first term on the right-hand side of Eqn. (5). This term involves a Dirac delta function, and can be interpreted as taking the integral over the line $\alpha(x,y) = 0$ : + +$$ +\begin{array}{l} \iint_ {D} \delta (\alpha (x, y)) \frac {\partial \alpha (x , y)}{\partial \Phi} f _ {a} (x, y) d x d y \tag {6} \\ = \int_ {\alpha (x, y) = 0} \frac {\partial \alpha (x , y)}{\partial \Phi} \| \nabla_ {x, y} \alpha (x, y) \| ^ {- 1} f _ {a} (x, y) d t, \\ \end{array} +$$ + +where $\| \nabla_{x,y}\alpha (x,y)\|$ is $L^2$ length of the edge equation gradient which accounts for the Jacobian of the variable substitution, and $t$ is the parametrization of the line along which we integrate. + +# 4 Method + +In this section we describe our method that we call EdgeGrad. First we will focus on the core assumptions and approximations that form the basis of our method. Then we will see that due to our specifically chosen approximation, we can apply the prior method of edge sampling to rasterization. From there, we will focus on edge detection and classification heuristics, and our discretization scheme. We will conclude by describing an extension of our method for handling geometry intersections. + +# 4.1 Rasterizing Edges + +Our primary objective is to differentiate rasterization while preserving its core image formation principles. We aim to calculate the gradients of pixel values relative to the mesh vertex positions, including gradients caused by visibility changes such as occlusion and disocclusion by primitive geometry. We seek to do this with a simple, minimal, and efficient method that does not necessitate modifying the initially rasterized image. + +Our main insight is that instead of approximating rasterization using a continuous representation, we assume the rasterized image is the result of some other continuous process that coincidentally aligns with our inherently non-differentiable, discrete-pixel rasterization. In essence, we opt for a direct ap + +Fig.1: Micro-edges and Pixel Pair. +![](images/e6b735ec17688fcb4c5fce6d5226cc5ddac903e5c482938d0715501810ca7f43.jpg) +The figure illustrates construction of the micro-edges and the change in pixel intensities A and B with the edge position $p$ movement, highlighting the C1 discontinuity at $p = 0$ , absent in average intensity. + +Fig. 2: Gradient computation flow. +![](images/1797519b671dc6da1b936dd2c896efe585d6957e8332a1c48df4d63e76ed2846.jpg) +The figure illustrates the backward propagation process, where gradients from the loss function are traced back and scattered to the fragments. + +proximation to compute gradients, instead of an indirect approach that first necessitates that the rendering be made continuous, for instance, through antialiasing filtering. More specifically, we use a construction where all straight edges, whether they are sides of triangles or intersections between triangles, are comprised of a collection of micro-edges. These micro-edges are assumed to be either strictly vertical or horizontal and are always positioned precisely between two pixels. An illustration of these constructions is shown in Figure 1. + +Our approximation leads to several key properties: a) boundaries never intersect pixels, only lying exactly between them; b) pixels are always either fully covered or not covered at all; c) there is at most one boundary between any two neighbouring pixels; d) we do not need to access the source geometry to evaluate where the edge is exactly. With this, we can reinterpret the problem as if the rasterized image was in fact obtained by applying an antialIASing filter, albeit without visible effects; the resulting image remains identical to the pre-antialIASing source. Essentially our construct of the micro-edges renders the antialIASing filter as an identity operation. + +With this approximation, gradient estimation becomes a local operation, requiring the analysis one edge between two pixels at a time. Note that despite the approximation that all edges reside strictly between pixels, we do not assume + +that they are fixed, but rather allow them to freely move within an infinitesimal range, enabling differentiation of pixel values with respect to the edge location. This leads to the greatest advantage of our formulation; it allows us to greatly simplify both the theory and computation, since all edges are axis-aligned. + +# 4.2 Pixel Pairs + +Let's analyze two adjacent pixels, A and B, separated by an edge, with their intensities represented as $I_{A}$ and $I_{B}$ respectively (refer to Figure 1 for details). The pixel intensities can be expressed using the Heaviside step function to account for the discontinuity at the edge as in Eqn. (4): + +$$ +\begin{array}{l} I _ {A} = \iint_ {D _ {A}} \theta (\alpha (x, y)) f _ {a} (x, y) + \theta (- \alpha (x, y)) f _ {b} (x, y) d x d y \\ I _ {B} = \iint_ {D _ {B}} \theta (- \alpha (x, y)) f _ {a} (x, y) + \theta (\alpha (x, y)) f _ {b} (x, y) d x d y. \tag {7} \\ \end{array} +$$ + +Without loss of generality, assuming a horizontal pair we define $\alpha(x,y) = p_{AB} - x$ , where $p_{AB}$ is the edge location. Edge location $p_{AB}$ is now the scene parameter with respect to which we want to differentiate the loss function. Since $p_{AB}$ effects both $I_A$ and $I_B$ : + +$$ +\frac {\partial L}{\partial p _ {A B}} = \frac {\partial L}{\partial I _ {A}} \frac {\partial I _ {A}}{\partial p _ {A B}} + \frac {\partial L}{\partial I _ {B}} \frac {\partial I _ {B}}{\partial p _ {A B}}, \tag {8} +$$ + +where $\frac{\partial I_A}{\partial p_{AB}}$ and $\frac{\partial I_B}{\partial p_{AB}}$ derive by applying Eqn. (6) to Eqn. (7). Unfortunately, with the edge strictly on the pixel boundary, we face a $C1$ discontinuity at $p_{AB} = 0$ , making intensity values non-differentiable. Infinitesimal movement of the edge affects $I_B$ when moving right and $I_A$ when moving left, without affecting the other. We still can compute one-sided limits of the derivatives, e.g. $\frac{\partial I_A}{\partial p}^{-}$ from the negative direction and $\frac{\partial I_B}{\partial p}^{+}$ from the positive direction. Applying Eqn. (6) and taking into account that the term $\| \nabla_{x,y}\alpha (x,y)\|$ becomes 1 as we obtain: + +$$ +\frac {\partial I _ {B}}{\partial p _ {A B}} ^ {-} = \frac {\partial I _ {A}}{\partial p _ {A B}} ^ {+} = 0; \frac {\partial I _ {B}}{\partial p _ {A B}} ^ {+} = \frac {\partial I _ {A}}{\partial p _ {A B}} ^ {-} = \int_ {x = 0} [ f _ {a} (x, y) - f _ {b} (x, y) ] d y \tag {9} +$$ + +We can circumvent the problem of $C_1$ discontinuity by averaging the one-sided limits of the derivatives from the left and right: + +$$ +\frac {\partial I _ {B}}{\partial p _ {A B}} = \frac {\partial I _ {A}}{\partial p _ {A B}} = \frac {1}{2} \int_ {x = 0} [ f _ {a} (x, y) - f _ {b} (x, y) ] d y \tag {10} +$$ + +Assuming that the pixels have constant value across all of their area (recall the rule that "pixels are always either fully covered or not covered at all") and substituting Eqn. (10) into Eqn. (8) we deduce: + +$$ +\left. \frac {\partial L}{\partial p _ {A B}} = \frac {1}{2} \left(\frac {\partial L}{\partial I _ {A}} + \frac {\partial L}{\partial I _ {B}}\right) \left(I _ {A} - I _ {B}\right) + \Omega \right], \tag {11} +$$ + +which is one of our main results. Here, $\frac{\partial L}{\partial p_{AB}}$ is the gradient of the loss function with respect to the position of the edge between pixels $A$ and $B$ . Additionally, we reintroduce $\Omega$ which is the second term on the right-hand side of Eqn. (5) that accounts for differentiating smooth regions of $f$ . This term was omitted in Eqn.(9) and Eqn. (10) for brevity. + +Now, we need to accumulate these gradients into the gradients of the vertex positions of the triangle that created those edges. We do that first by scattering the gradients $\frac{\partial L}{\partial p_{AB}}$ to the fragments of the triangles associated with the pixels forming the edge and then by invoking the backward function of the interpolator module to gather the fragment gradients into the vertex gradients. Next, we will take a look at the gradient scattering mechanism. + +# 4.3 Edge Classification and Gradient Scattering + +To backpropagate the gradients to the scene parameters we need to compute the gradient with respect to fragment positions, while in §4.2 we only derived the gradient $\frac{\partial L}{\partial p_{AB}}$ with respect to the edge position. We scatter each edge gradient $\frac{\partial L}{\partial p_{AB}}$ to adjacent fragments where we accumulate the fragment gradients, as illustrated in Figure 2. Using the chain rule, we obtain fragment gradients be multiplying $\frac{\partial L}{\partial p_{AB}}$ with $\frac{\partial p_{AB}}{\partial r}$ ; the later relates the movement of the edge to the movement of the fragment, where $r$ is the position of the fragment. The computation of $\frac{\partial p_{AB}}{\partial r}$ depends on the type of the edge formed by the fragments. We consider several cases: overhanging primitives (either one another or background); adjacent primitives; intersecting primitives. Unlike Laine et al. [20], we avoid additional data structures like hash maps for connectivity. + +Our approach examines each neighboring pixel pair, comparing their triangle IDs, see Figure 3 for illustration. Matching IDs indicate no edge; differing IDs prompt further tests. A single valid ID suggests a primitive overhanging the background. Testing a pixel's center against the triangle of the other pixel discerns overhangs and self-intersections. Non-contained pixels in each other's triangles imply adjacency. If both pixel centers are in the triangle of the other pixel, it implies self-intersections, and if in just one - then overhanging. For adjacent primitives we do not scatter gradients at all, since it is likely due to primitives sharing an edge and there is no discontinuity. Please note that even in that case $\frac{\partial L}{\partial p_{AB}}$ may be non-zero due to $\Omega$ contribution. In the overhanging case, we scatter the gradient only to the primitive on top, as the covered primitive can not affect the position of the edge. Thus, in such cases the covered fragment has $\frac{\partial p}{\partial r} = 0$ while the foreground fragment has $\frac{\partial p}{\partial r} = 1$ reflecting that the fragment movement is rigidly tied to the movement of the edge. Please refer to supplementary material for more details. + +# 4.4 Geometry Intersections + +This subsection details computing the derivative of edge location with respect to fragment position $\frac{\partial p}{\partial r}$ in scenarios where the edge arises from intersecting + +![](images/9cc86066d8317d857ce964646a3dc67a00101925ae815eb2251be0ca97ad0293.jpg) +Fig. 3: Edge classification. The accompanying figure provides a schematic representation of our edge classification process. Edges are categorized into four types: no edge, adjacent primitives, overhanging primitives, and intersecting primitives. +Fig. 4: Geometry intersection. A geometrical derivation for the partial derivative $\frac{\partial p}{\partial r}$ , essential for converting edge-related gradients to fragment-related gradients in cases where edges are formed by geometry intersections. This shows the x-z plane, where the render camera is viewing the scene along the z axis so that edges are parallel to the y axis (and perpendicular to the drawing). Projections of both intersecting primitives are transformed into two lines. + +![](images/82348e31d87e5b4a41c69b565e7ff1364fc52a1448726f80d4a788eab7633299.jpg) + +geometry. This situation is inherently complex, and a general-case solution would be challenging. However, our micro-edge formulation streamlines the problem because intersection lines projected onto the image plane comprised of microedges, e.i. either vertical or horizontal. This allows us to analyze the problem in 2D, focusing on the x-z or y-z planes, with the z-axis perpendicular to the image plane. We derive the derivative of the edge location by analyzing two fragments of a micro-edge that belongs to the intersection line, focusing on how the edge would move in response to infinitesimal movements of each fragment. Without loss of generality we consider the x-z plane: + +$$ +\frac {\partial p}{\partial r} = - \mathbf {n} _ {z} ^ {f} \left[ \mathbf {n} ^ {f ^ {T}} \left[ \begin{array}{l l} 0 & 1 \\ - 1 & 0 \end{array} \right] \mathbf {n} ^ {v} \right] ^ {- 1}, \tag {12} +$$ + +where $p$ is edge position, $r$ is the fragment position (in 3D clip space), $\mathbf{n}^v$ is the varying primitive's normal, and $\mathbf{n}^f$ is the fixed primitive's normal. In this context, the $x$ -axis lies on the image plane, and the $z$ -axis is perpendicular to it, see Fig. 4. We compute $\frac{\partial p}{\partial r}$ for both fragments that form the intersection edge by + +first fixing one fragment and varying the other, and then switching them. Please refer to supplementary material for derivation details. + +# 4.5 Theoretical analysis of the approximation + +At first glance, replacing arbitrarily oriented edges with vertical and horizontal micro-edges might seem overly simplistic, raising doubts about accuracy. It could be argued, for example, that the perimeter defined by micro-edges doesn't match the original as micro-edge size approaches zero. However, applying the two-dimensional version of the divergence theorem from vector calculus offers insight: + +$$ +\oint_ {C} \mathbf {f} \cdot \mathbf {n} d t = \iint_ {D} (\nabla_ {\Phi} \cdot \mathbf {f}) d x d y, \tag {13} +$$ + +where $\mathbf{n}$ is normal to boundary $C$ , within which the integrand function is continuous and differentiable. We are free to pick any such boundaries that includes a discontinuity. In our case, the term $\frac{\partial\alpha(x,y)}{\partial\Phi}$ in Eqn. (6) is a normal vector to the boundary $\alpha (x,y) = 0$ thus the integral is equivalent to the left-hand side of Eqn. (13). According to Eqn. (13), this recasts the boundary derivative integral as an area integral. The difference between the area integrals of micro-edges and original edges is confined to a space defined by the micro-edges size. As pixel size diminishes, the integral approaches the exact value, and the discrepancy area vanishes, confirming the micro-edge approach converges to the same solution as the original formulation in the limit. + +# 5 Experimental Analysis + +We implement our method in PyTorch with custom CUDA kernels, and largely follow the structure of nvdiffrast [20], with the exception for the antialIASing module which we replace with our EdgeGrad module. For all experiments we utilize Laplacian preconditioning proposed by Nicolet et al. [34]. We find Laplacian preconditioning extremely helpful for stabilizing training and we use it for both, meshes and texture data. While Laplacian preconditioning may reduce self-intersections, this happens only due to much smoother optimization process, but it does not account for them in any way. + +Forward gradients. We first analyze the performance of the proposed approach by evaluating the forward gradients on a variety of test cases. Forward gradients are per-pixel derivatives with respect to some scene parameter, which is depicted with white arrows on the scene images. Fig. 5 shows the forward gradients computed using finite differences, raytracing methods Mitsuba [14,15,36] and redner [21], and rasterization approaches including nvdiffrast [20] and our approach (EdgeGrad). Unlike other methods, our approach still matches finite differences gradients in cases of intersecting geometry. While Mitsuba and Redner, which require multiple samples per pixel (256 in these comparisons), typically offer higher accuracy, they fall short at geometry intersections. Nvdiffrast + +![](images/f6d8019b5f1778c9efb33293a150d0dd0f86b8cf3e380987e407fb2e9013173f.jpg) +Fig. 5: Comparison of forward gradient on several test scenes with the numerical solution using finite differences. (a) Synthetic test scene, we show gradients with respect to a parameter. We use (b) finite differences as a reference, and compare with (c) Mitsuba 3 [15,36], (d) redner [21], (e) nvdiffrast [20], and (f) our edge gradient approach. + +is noisier because it assumes triangles containing silhouette edges always overlap pixel centers, which often isn't the case, leading to omitted gradients. + +Backward gradients. Tab. 1 quantitatively analyzes the accuracy of the backward gradients. We show improved results especially for complex intersecting geometry. Compared to raytracing methods, our rasterization-based approach yields comparable performance at a fraction of the computational cost. + +Runtime performance end accuracy by image size. Fig. 6 shows running time of our method (EdgeGrad) compared to Mitsuba [36] and nvdiffrast [20]. Our method significantly outperforms other approaches in running time on all variety of image sizes and triangle count. Our method has comparable accuracy to Mitsuba [36] (and even better for the case with geometry intersections) at large image sizes, but performs worse for smaller image size due to sub pixel triangles. + +Quantitative scene reconstructions. Tab. 2 shows quantitative results of scene reconstruction on blender dataset [32]. We follow the testing procedure described in [32] and report PNSR, SSIM and LPIPS metrics. We compare our method (EdgeGrad) with a "continues only" baseline (which ignores gradients from discontinuities), nvdiffrast [20], and with a variant of our method that does not have a special handling for the geometry intersections. Method optimization procedure is similar to Nicolet et al. [34] we only add MLP for decoding texture values in a view conditioned manner similar to [29]. We initialize geometry with some coarse manually created meshes shown in Fig. 7. To avoid additional complex + +![](images/b643c42ff89145154b5bab67e750934303441922c0807c1c48f7a3cf15489d8b.jpg) +Scene + +![](images/090f379491bb8630c9df89a6493fc5c8f5022bd3eb23e6e0ae9f9b2fe088ff27.jpg) + +![](images/ee24749cb0742bd6c8bc88c6930b2ba7c40680db608f7604804287caa8578db2.jpg) + +
redner [21]1.20%53.52%37.94%
Mitsuba3 [36]0.67%75.83%37.59%
Nvdiffrast [20]45.91%69.33%39.42%
EdgeGrad(our)6.01%3.35%8.35%
+ +Table 1: Accuracy of backward gradients. Relative error, $\%$ (↓). This table shows relative errors in backward gradient computations for test scenes. Second and third scenes include geometry intersections, emphasizing our method's advantage in managing these complexities. + +![](images/389c0b8eef6169a730b9a8abcf6fda6dec616be3a023323a988728951e109ade.jpg) + +![](images/6485593257a823179f267d5d321d747f931a3337141f962a7b0d8db211654e31.jpg) + +![](images/d90e8eeb0171a38c1b491bfb4c88d81263633763c33fe984056580f10e65f4fd.jpg) +(a) Running time vs image size +(c) Gradient error vs image size self-intersections absent + +![](images/8a12308d9f1178894f10228224d00969ad3d3f2c4643e6f2879325be6ed49e7a.jpg) +(b) Running time vs triangle count +(d) Gradient error vs image size self-intersections present + +![](images/8487633fe2caeb03f444e1e52bf15afd80a1a9cebc2ed6f5cede1d6f6aea9e2e.jpg) +(a) camera image + +![](images/4ddc715f2aa7b1e89c480bc319b18971613a7ac6ab20ea0b53f2f2a977334999.jpg) +(c) initialization mesh + +![](images/8d62125d94ee7e410c80a1043a19cf88a70db77957a8acf96ee7a275317ad8ca.jpg) +(d) optimized mesh + +![](images/dffc9ff9e2f524a9048ed246965a1c6631296f411bd55bffca9df4c33d233147.jpg) +Fig. 6: Runtime performance and errors. a) Runtime [ms] by image size [px]; b) runtime [ms] by triangle count; c) gradient error [%] by image size [px] without self-intersections; d) gradient error [%] by image size [px] with self-intersections. +(e)rgb estimation +(f) segmentation estimation +(g) Learned segmentation texture for mouth region + +![](images/274eb69f9c16bb8ac44903afd294142c7d3ef8e5c41a8215ac8445622ad4b0b1.jpg) + +![](images/88258279ed633583bb90674a385ed9df0235b0c3711379b53a7e4a1f9c96c1e1.jpg) +Initialization + +![](images/b39e087da88a6328eb84f34e39e8d7a705043941a12e24b2cd917f1f3a5da157.jpg) +Optimized + +![](images/248791888a2375df21f3a3a8665e5033f5e575120217a7e35942008366628364.jpg) +Initialization + +![](images/560488f97baf3b0cb4e1ca86a411f69d3960957e1f6e6f24ea5b4e14a165ab10.jpg) +1 + +![](images/2c8d35fbfa938274b6c868dd885f44079506b12699137ff86afc5e0bb28d81f0.jpg) + +![](images/8d085686284e5bd645fe4691dd3ffa480fcc5611dcd4e7543ab1d93ad0d699f0.jpg) +Optimized +(h) Initialization mesh and reconstructed mesh for synthetic blender dataset + +![](images/a5c41640ef10d4bf8cbf4df6fa7d4ce9ad0a6fda595aa97fb9ad4aa6f6fe7fe0.jpg) +Initialization + +![](images/e87f739b8c0b08f09e956919173c7cb025c3be4bd014f80975f5c1a94bc1e6da.jpg) +Optimized + +![](images/e9bc93948f9064931b0977efddcdc988349450976a03e4a2f58f294208c5dfb3.jpg) +Initialization +Optimized + +![](images/270631c5ca503c37c14aa3c9427544ccfb88f18e097cebcb6a87ead44ba9ee00.jpg) +Ours + +![](images/b8eca095d0ea7e666d6fecef422be1dece8d1cfabca9e37a2c9f0396b88b7560.jpg) +NvdiffRast + +![](images/648a873f103fe7c1789773e075752bdf110c8735d327c4768ee1d1a1921ad351.jpg) +Fig. 7: Experimental setup details. The figure shows: a) the ground truth camera image; b) predicted segmentation mask for mouth region using HRNet [47]; c) initialization mesh from 3DMM; d) optimized avatar mesh; e) rendered image of the optimized model; f) rendered segmentation estimation image; g) learned static segmentation texture; h) initialization meshes and optimized meshes for blender dataset [32]. +No Edge Handling + +![](images/44817837ff5de8254c6ce671dfdd446af19918435c0f2ec5acb645579b178cc1.jpg) +Ours + +![](images/5d422cfe9914f436eb1aae59526f6f6d8892d0be75fd0816abfb0fa7b9005d0c.jpg) +NvdRast + +![](images/e5d35610e079806573a70c835612e36d6bb364ffb6f20d874affdea6727cc56e.jpg) +No Edge Handling + +![](images/93b5754078ade38f45e2087b8334616cd999a75c9dbfff144d132b399d7c97b4.jpg) +Ours +Fig. 8: Reconstructions from inverse rendering compare our method with nvdiffrast [20] and rasterization without discontinuity handling. Although nvdiffrast can match the outline of objects well, it suffers from interpenetration (see leaves of Ficus) and does not capture as much detail as ours (see the backrest of the chair and ground of lego). The method with no edge handling fails completely on these objects. + +![](images/17418c45a056fc50c1a8f05c1b7623d4ef36bf41564254015e566cd33678ae3a.jpg) +NdivdiffRast + +![](images/2b9cd60707d6f6f32529b7e4faaca333e8627c9c39b9a44473a9c98985c0f27b.jpg) +No Edge Handling + +
LegoChairShipHotdogFicusDrumsMicMaterials
PSNR ↑Continuous only16.34228.15115.69327.85018.39915.53421.79717.11
Nvdiffrast [20]29.4429.80426.17331.68424.96423.15429.63824.413
EdgeGrad -intersect.29.57433.07726.73134.22526.78124.0831.22425.221
EdgeGrad (our)29.66732.98126.6534.08126.55924.01431.34625.287
LegoChairShipHotdogFicusDrumsMicMaterials
SSIM ↑Continuous only0.75500.95310.69290.93740.82450.77230.91820.7693
Nvdiffrast [20]0.94670.94160.83160.94730.93230.89710.96470.8865
EdgeGrad -intersect.0.94850.97650.84780.97150.94680.92160.97560.8945
EdgeGrad (our)0.95010.97640.84720.97130.94520.92200.97670.8972
LegoChairShipHotdogFicusDrumsMicMaterials
LPIPS ↓Continuous only0.32230.06110.39130.11090.20460.29430.12720.2511
Nvdiffrast [20]0.06530.07710.20200.09360.07810.11160.05490.1291
EdgeGrad -intersect.0.06330.03480.18990.05460.06810.08710.04390.1200
EdgeGrad (our)0.06140.03410.18500.05480.06720.08700.03990.1185
+ +Table 2: Per-scene quantitative results from the realistic synthetic blender dataset [32], featuring scenes with complex geometries and non-Lambertian materials rendered using Blender's Cycles pathtracer. + +![](images/1d8b8d3ae3ddd5675c2650aeb5d70c079a873e25269e6360b6ff64cdbbf2eb60.jpg) +Fig. 9: Example renderings of mesh-based avatars from [29] trained using EdgeGrad, capturing complex inner mouth movements, including teeth and tongue dynamics. + +ity we opted out automatic remeshing introduced by Nicolet et al. which would necessitate updating UV parametrization in our case. + +Qualitative mesh reconstruction. Qualitatively, we compare the mesh reconstructions obtained on the public NeRF dataset [32] using our approach, a naive approach with no special edge handling, and nvdiffrast [20] in Fig. 8. All methods use the same shading model. Our approach shows improved detail in the recovered meshes. + +# 6 Application to Dynamic Head Scenes + +We show the importance of correct edge gradients on a head modeling application by fitting a personalized mesh-based avatar representation to facial performances captured in a multiview lightstage. modifying the avatar creation method from [29]. As the facial avatar representation, we use an encoder-decoder architecture to produce mesh vertices and texture-mapped appearance features [25], which are shaded using a pixel-wise MLP producing RGB appearance [29]. We use a coarse face tracking produced using traditional computer vision as initialization for inverse rendering. Importantly, this coarse face template has a small number of vertices (8k) and no geometry in the interior of the mouth. + +Traditionally, fitting the interior of the mouth without strong priors is a complex task due to poor observability and intricately interacting geometry with + +![](images/1970cc83f0a77b518b57863fdc75e12dcd76c03d06f58238ff3249012a5def9d.jpg) +(a) camera image + +![](images/3131f3014cc1620513b16c33ac01caacbfaf04db16b11f7001524f0f1f005299.jpg) +(b) Nvdiffrast + +![](images/71b44cfe6577bba41607c6fd59db4b53843b06b01255ea9b960ac0f6c4b138a6.jpg) +(c) EdgeGrad (our) +(d) EdgeGrad (our) w/o segmentation supervision + +![](images/930f8e889bfb2952980f7dd8dcd1dbff3d9afbc3b2906a565ec9ceb2e7129184.jpg) +Fig. 10: Qualitative evaluation and ablation on dynamic head scene. We compare: (c) our method, (a) ground truth images, (b) Nvdiffrast [20], and ablated version of our method without including segmentation mask. The avatar model is from [29] trained using EdgeGrad and Nvdiffrast. We highlight artifacts with white arrows. + +constant contact. Prior work either relies on artists to create alignments, volumetric simulation to prevent self-intersections [13], or on very specific mouth interior capture methods to build priors [30,49]. Instead, we show that by using differentiable rasterization with correct edge gradients, we can achieve fully automatic geometric fitting of the complex geometry of the face and mouth interior including lips, teeth, and tongue using a mesh-based model of fixed topology. We modify [29] in two ways. First, we use the edge gradient rasterization method of Sec. 4. Second, we add supervision in the form of 2D segmentation masks with separate labels for the regions of the teeth, tongue, and lips, for details see Fig. 7. While we can achieve accurate per-frame geometry without the additional segmentation loss supervision, we find that segmentation losses help anchor the teeth and tongue geometry to a consistent region of the UV map. Fig. 9 shows the quality of geometric fitting achievable with this inverse rendering method. Note that we accurately reconstruct the complex contact geometry of the tongue sliding across teeth, teeth-to-teeth contact, as well as the tongue pushing against the lips and cheeks. Fig. 10 demonstrates reconstruction quality of complex inner mouth details, which are challenging due to frequent self-intersections. We also include results without 2D segmentation mask supervision. + +# 7 Conclusions and Limitations + +We have presented a fast method for rasterization-based differentiable rendering that has accurate gradients across visibility discontinuities such as triangle edges, occlusions and geometry intersections. Our results demonstrate enhanced mesh reconstruction fidelity across diverse inverse rendering scenarios, comparable to implicit and volumetric methods. Our method also unlocks new applications, accurate mouth interior fitting in dynamic facial performance capture, through fully automatic inverse rendering. Regarding limitations, our method inherits the constraints of rasterization, including the inability to manage transparency, lack of precise antialIASing without multisampling, and the absence of physically based rendering and global illumination. It may also inaccurately interpret adjacent primitives in finely tessellated meshes, leading to sub-pixel triangles. + +# References + +1. Baek, S., Kim, K.I., Kim, T.K.: Pushing the envelope for rgb-based dense 3d hand pose estimation via neural rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1067-1076 (2019) +2. Bangaru, S.P., Gharbi, M., Luan, F., Li, T.M., Sunkavalli, K., Hasan, M., Bi, S., Xu, Z., Bernstein, G., Durand, F.: Differentiable rendering of neural sdfs through reparameterization. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022) +3. Bangaru, S.P., Li, T.M., Durand, F.: Unbiased warped-area sampling for differentiable rendering. ACM Transactions on Graphics (TOG) 39(6), 1-18 (2020) +4. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021) +5. Beker, D., Kato, H., Morariu, M.A., Ando, T., Matsuoka, T., Kehl, W., Gaidon, A.: Monocular differentiable rendering for self-supervised 3d object detection. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16. pp. 514-529. Springer (2020) +6. Bogo, F., Kanazawa, A., Lassner, C., Gehler, P., Romero, J., Black, M.J.: Keep it spl! Automatic estimation of 3d human pose and shape from a single image. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14. pp. 561-578. Springer (2016) +7. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022) +8. Chen, W., Ling, H., Gao, J., Smith, E., Lehtinen, J., Jacobson, A., Fidler, S.: Learning to predict 3d objects with an interpolation-based differentiable renderer. Advances in neural information processing systems 32 (2019) +9. Chen, W., Litalien, J., Gao, J., Wang, Z., Fuji Tsang, C., Khamis, S., Litany, O., Fidler, S.: Dib-r++: learning to predict lighting and material with a hybrid differentiable renderer. Advances in Neural Information Processing Systems 34, 22834-22848 (2021) +0. Choy, C.B., Xu, D., Gwak, J., Chen, K., Savarese, S.: 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14. pp. 628-644. Springer (2016) +1. Ge, L., Ren, Z., Li, Y., Xue, Z., Wang, Y., Cai, J., Yuan, J.: 3d hand shape and pose estimation from a single rgb image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10833-10842 (2019) +2. Henzler, P., Reizenstein, J., Labatut, P., Shapovalov, R., Ritschel, T., Vedaldi, A., Novotny, D.: Unsupervised learning of 3d object categories from videos in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4700-4709 (2021) +3. Ichim, A.E., Kadleček, P., Kavan, L., Pauly, M.: Phace: Physics-based face modeling and animation. ACM Trans. Graph. 36(4) (jul 2017) +4. Jakob, W., Speierer, S., Roussel, N., Nimier-David, M., Vicini, D., Zeltner, T., Nicolet, B., Crespo, M., Leroy, V., Zhang, Z.: Mitsuba 3 renderer (2022), https://mitsuba-renderer.org + +15. Jakob, W., Speierer, S., Roussel, N., Vicini, D.: Dr.jit: A just-in-time compiler for differentiable rendering. Transactions on Graphics (Proceedings of SIGGRAPH) 41(4) (Jul 2022). https://doi.org/10.1145/3528223.3530099 +16. Jiang, Y., Ji, D., Han, Z., Zwicker, M.: Sdiff: Differentiable rendering of signed distance fields for 3d shape optimization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1251-1261 (2020) +17. Kato, H., Ushiku, Y., Harada, T.: Neural 3d mesh renderer. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3907-3916 (2018) +18. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG) 42(4), 1-14 (2023) +19. de La Gorce, M., Fleet, D.J., Paragios, N.: Model-based 3d hand pose estimation from monocular video. IEEE Trans. Pattern Anal. Mach. Intell. 33(9), 1793-1805 (Sep 2011). https://doi.org/10.1109/TPAMI.2011.33, http://dx.doi.org/10.1109/TPAMI.2011.33 +20. Laine, S., Hellsten, J., Karras, T., Seol, Y., Lehtinen, J., Aila, T.: Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG) 39(6), 1-14 (2020) +21. Li, T.M., Aittala, M., Durand, F., Lehtinen, J.: Differentiable monte carlo ray tracing through edge sampling. ACM Trans. Graph. (Proc. SIGGRAPH Asia) 37(6), 222:1-222:11 (2018) +22. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 300–309 (2023) +23. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020) +24. Liu, S., Li, T., Chen, W., Li, H.: Soft rasterizer: A differentiable renderer for image-based 3d reasoning. The IEEE International Conference on Computer Vision (ICCV) (Oct 2019) +25. Lombardi, S., Saragih, J., Simon, T., Sheikh, Y.: Deep appearance models for face rendering. ACM Transactions on Graphics (TOG) 37(4), 1-13 (2018) +26. Lombardi, S., Simon, T., Saragih, J., Schwartz, G., Lehrmann, A., Sheikh, Y.: Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751 (2019) +27. Loper, M.M., Black, M.J.: Opendra: An approximate differentiable renderer. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VII 13. pp. 154-169. Springer (2014) +28. Loubet, G., Holzschuch, N., Jakob, W.: Reparameterizing discontinuous integrands for differentiable rendering. Transactions on Graphics (Proceedings of SIGGRAPH Asia) 38(6) (Dec 2019). https://doi.org/10.1145/3355089.3356510 +29. Ma, S., Simon, T., Saragih, J.M., Wang, D., Li, Y., la Torre, F.D., Sheikh, Y.: Pixel codec avatars. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 64-73 (2021) +30. Medina, S., Tomé, D., Stoll, C., Tiede, M., Munhall, K., Hauptmann, A., Matthews, I.: Speech driven tongue animation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE/CVF (2022) +31. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) + +32. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) +33. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989 (2022) +34. Nicolet, B., Jacobson, A., Jakob, W.: Large steps in inverse rendering of geometry. ACM Transactions on Graphics (TOG) 40(6), 1-13 (2021) +35. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3504-3515 (2020) +36. Nimier-David, M., Vicini, D., Zeltner, T., Jakob, W.: Mitsuba 2: A retargetable forward and inverse renderer. Transactions on Graphics (Proceedings of SIGGRAPH Asia) 38(6) (Dec 2019). https://doi.org/10.1145/3355089.3356498 +37. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019) +38. Pavlakos, G., Zhu, L., Zhou, X., Daniilidis, K.: Learning to estimate 3d human pose and shape from a single color image. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 459-468 (2018) +39. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022) +40. Rhodin, H., Robertini, N., Richardt, C., Seidel, H.P., Theobalt, C.: A versatile scene model with differentiable visibility applied to generative pose estimation. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 765-773 (2015) +41. Roessle, B., Barron, J.T., Mildenhall, B., Srinivasan, P.P., Nießner, M.: Dense depth priors for neural radiance fields from sparse input views. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12892-12901 (2022) +42. Roveri, R., Öztireli, A.C., Pandele, I., Gross, M.: Pointpronets: Consolidation of point clouds with convolutional neural networks. In: Computer Graphics Forum. vol. 37, pp. 87-99. Wiley Online Library (2018) +43. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512 (2023) +44. Tsalicoglou, C., Manhardt, F., Tonioni, A., Niemeyer, M., Tombari, F.: Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439 (2023) +45. Tulsiani, S., Zhou, T., Efros, A.A., Malik, J.: Multi-view supervision for single-view reconstruction via differentiable ray consistency. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2626-2634 (2017) +46. Vicini, D., Speierer, S., Jakob, W.: Differentiable signed distance function rendering. ACM Transactions on Graphics (TOG) 41(4), 1-18 (2022) +47. Wang, J., Sun, K., Cheng, T., Jiang, B., Deng, C., Zhao, Y., Liu, D., Mu, Y., Tan, M., Wang, X., et al.: Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence 43(10), 3349-3364 (2020) +48. Wiles, O., Gkioxari, G., Szeliski, R., Johnson, J.: Synsin: End-to-end view synthesis from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7467-7477 (2020) + +49. Wu, C., Bradley, D., Garrido, P., Zollhöfer, M., Theobalt, C., Gross, M., Beeler, T.: Model-based teeth reconstruction. ACM Trans. Graph. 35(6) (dec 2016) +50. Yan, X., Yang, J., Yumer, E., Guo, Y., Lee, H.: Perspective transformer nets: Learning single-view 3d object reconstruction without 3d supervision. Advances in neural information processing systems 29 (2016) +51. Yang, G., Cui, Y., Belongie, S., Hariharan, B.: Learning single-view 3d reconstruction with limited pose supervision. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 86-101 (2018) +52. Yu, A., Fridovich-Keil, S., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks (2021) +53. Zhang, C., Wu, L., Zheng, C., Gkioulekas, I., Ramamoorthi, R., Zhao, S.: A differential theory of radiative transfer. ACM Trans. Graph. 38(6), 227:1-227:16 (2019) +54. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492 (2020) +55. Zhao, S., Jakob, W., Li, T.M.: Physics-based differentiable rendering: From theory to implementation. In: ACM SIGGRAPH 2020 Courses. SIGGRAPH '20, Association for Computing Machinery, New York, NY, USA (2020). https://doi.org/10.1145/3388769.3407454, https://doi.org/10.1145/3388769.3407454 \ No newline at end of file diff --git a/2024/Rasterized Edge Gradients_ Handling Discontinuities Differentially/images.zip b/2024/Rasterized Edge Gradients_ Handling Discontinuities Differentially/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d2583c00e602814a1ee1e11318975b7346371196 --- /dev/null +++ b/2024/Rasterized Edge Gradients_ Handling Discontinuities Differentially/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4abea891b9046cce2dfb0ec1fa444c4d0ca168a8b087ee5c73bff7702ae45ef7 +size 538819 diff --git a/2024/Rasterized Edge Gradients_ Handling Discontinuities Differentially/layout.json b/2024/Rasterized Edge Gradients_ Handling Discontinuities Differentially/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9e74cb2ea9bf992fe87b127adc11bd5ced3cb4a6 --- /dev/null +++ b/2024/Rasterized Edge Gradients_ Handling Discontinuities Differentially/layout.json @@ -0,0 +1,10964 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 177, + 111, + 436, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 111, + 436, + 147 + ], + "spans": [ + { + "bbox": [ + 177, + 111, + 436, + 147 + ], + "type": "text", + "content": "Rasterized Edge Gradients: Handling Discontinuities Differently" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 143, + 168, + 470, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 168, + 470, + 192 + ], + "spans": [ + { + "bbox": [ + 143, + 168, + 470, + 192 + ], + "type": "text", + "content": "Stanislav Pidhorskyi" + }, + { + "bbox": [ + 143, + 168, + 470, + 192 + ], + "type": "inline_equation", + "content": "^{\\text{①}}" + }, + { + "bbox": [ + 143, + 168, + 470, + 192 + ], + "type": "text", + "content": ", Tomas Simon" + }, + { + "bbox": [ + 143, + 168, + 470, + 192 + ], + "type": "inline_equation", + "content": "^{\\text{②}}" + }, + { + "bbox": [ + 143, + 168, + 470, + 192 + ], + "type": "text", + "content": ", Gabriel Schwartz" + }, + { + "bbox": [ + 143, + 168, + 470, + 192 + ], + "type": "inline_equation", + "content": "^{\\text{③}}" + }, + { + "bbox": [ + 143, + 168, + 470, + 192 + ], + "type": "text", + "content": ", He Wen, Yaser Sheikh, and Jason Saragih" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 202, + 202, + 411, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 202, + 411, + 213 + ], + "spans": [ + { + "bbox": [ + 202, + 202, + 411, + 213 + ], + "type": "text", + "content": "Reality Labs,Meta,Pittsburgh, Pennsylvania,USA" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 161, + 243, + 452, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 243, + 452, + 462 + ], + "spans": [ + { + "bbox": [ + 161, + 243, + 452, + 462 + ], + "type": "text", + "content": "Abstract. Computing the gradients of a rendering process is paramount for diverse applications in computer vision and graphics. However, accurate computation of these gradients is challenging due to discontinuities and rendering approximations, particularly for surface-based representations and rasterization-based rendering. We present a novel method for computing gradients at visibility discontinuities for rasterization-based differentiable renderers. Our method elegantly simplifies the traditionally complex problem through a carefully designed approximation strategy, allowing for a straightforward, effective, and performant solution. We introduce a novel concept of micro-edges, which allows us to treat the rasterized images as outcomes of a differentiable, continuous process aligned with the inherently non-differentiable, discrete-pixel rasterization. This technique eliminates the necessity for rendering approximations or other modifications to the forward pass, preserving the integrity of the rendered image, which makes it applicable to rasterized masks, depth, and normals images where filtering is prohibitive. Utilizing micro-edges simplifies gradient interpretation at discontinuities and enables handling of geometry intersections, offering an advantage over the prior art. We showcase our method in dynamic human head scene reconstruction, demonstrating effective handling of camera images and segmentation masks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 484, + 229, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 484, + 229, + 496 + ], + "spans": [ + { + "bbox": [ + 133, + 484, + 229, + 496 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 510, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 510, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 510, + 481, + 665 + ], + "type": "text", + "content": "Significant advances have been made in recent years in modeling real 3D objects from image measurements. Much of this advancement can be credited to improvements in inverse rendering, which enables the automatic inference of 3D scene parameters that best reconstruct the images. While volumetric methods like NeRF [31] simplify inverse rendering by eliminating the need for predefined scene topology, classical mesh-based representations remain widely used due to their efficiency in modeling opaque surfaces. These representations often rely on highly performant rasterization, which requires discrete operations to reason about ordering and coverage, which poses challenges for gradient computations. Our main contribution is a theoretical framework for approximating gradients of a rasterization process that simultaneously elucidates previous constructions, arrives at a simple, fast, and accurate formulation, and further improves gradient approximation accuracy over state-of-the-art methods." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 307 + ], + "type": "text", + "content": "In rasterization, a rendered pixel's value is determined by the foremost triangle covering it. Pixels whose footprint lies entirely within a single triangle are straightforward to handle, as infinitesimal mesh motion does not change their triangle membership. However, pixels on triangle boundaries, either between adjacent triangles, at occlusion boundaries, or at intersections between triangles, must account for how their triangle membership changes in response to infinitesimal changes in the mesh. Existing methods employ different approximations of soft-membership to account for this, enabling gradient computation in those areas, effectively compositing contributions from member triangles via a weighted sum. For example, [20] uses the anti-aliasing approach, [14] averages montecarlo samples during ray-casting, and [40] extends the influence of boundary triangles using a falloff function. In this work, we show that all these methods exhibit approximation errors that can fail to compute correct gradients in one case or another. In particular, none of the existing methods accurately compute the gradient at triangle intersections, which can often happen during inverse rendering, especially when the current estimate is far from the desired solution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 308, + 482, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 451 + ], + "type": "text", + "content": "In contrast to existing works that directly tackle geometric discontinuities and rasterization's discrete nature, we introduce a micro-edge formulation which allows to interpret the rasterized image as an outcome of a continuous process that coincidentally aligns with discrete-pixel rasterization, simplifying gradient computation significantly. Unlike some other works [20, 40], our formulation achieves good gradient estimates without altering the rasterization forward pass, maintaining the rasterized image's integrity. This is crucial for optimizing segmentation masks, depth maps, and normal maps where filtering or smoothing is not feasible, e.g. anisotropic filtering or soft rasterization would mix normals from different surfaces misrepresenting the geometry. The simplifications offered by our micro-edge formulation allow us to seamlessly handle discontinuities caused by geometry intersections, offering an advantage over the prior art." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 452, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 452, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 452, + 482, + 571 + ], + "type": "text", + "content": "In summary, our paper introduces a straightforward, accurate, and efficient method for computing gradients in rasterization, comparable to existing techniques but with greater simplicity and the ability to handle self-intersections. We analyze our method's errors from a theoretical standpoint and compare gradients qualitatively and quantitatively with finite differences and other methods across various test cases. We also assess runtime efficiency and accuracy by image size and showcase qualitative and quantitative findings on a synthetic blender dataset [32]. Our method excels in complex applications, such as detailed dynamic human head reconstructions, effectively managing the intricate details of the inner mouth region with significant occlusions and deformations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 591, + 234, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 591, + 234, + 604 + ], + "spans": [ + { + "bbox": [ + 132, + 591, + 234, + 604 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Differentiable rendering is one of the critical building blocks for many existing computer vision problems, such as 3d object reconstruction [10, 17, 45, 50, 51], 3d object prediction [5, 8, 9, 12], pose estimation [1, 6, 11, 38] novel view synthesis [4, 18, 32, 33, 54], as well as newly emerging applications such as text-to-3D" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "text", + "content": "S. Pidhorskyi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 189 + ], + "type": "text", + "content": "generative models [22, 39, 43, 44]. Differentiable rendering employs a wide array of underlying 3D representations including explicit surfaces like mesh-based representations [19, 20, 27], implicit surfaces [2, 16, 37, 46], point clouds [18, 42, 48], explicit or implicit volume representations [4, 23, 26, 32, 35, 50, 52], and hybrid representations [7, 33]. For a comprehensive introduction to differentiable rendering, we refer the reader to Zhao et al.'s excellent SIGGRAPH course notes [55]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 191, + 482, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 191, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 191, + 482, + 396 + ], + "type": "text", + "content": "Mesh optimization often poses more challenges than implicit and volumetric methods, leading to the latter's prevalence in the field, as noted by Roessle et al. [41]. Despite this, meshes excel in high-performance rendering and imply registration with predefined topologies. Mesh rendering methods can be categorized into ray tracing-based and rasterization-based, with the former including frameworks like Mitsuba 3 [15, 36]. Differentiable ray tracing, aimed at direct illumination or path tracing, is computationally expensive but principled. Key advancements by Li et al. [21] and Zhang et al. [53] split this gradient into discontinuity integrals with Dirac delta functions and differentiable continuous parts. Li et al. replaced Dirac delta integration with boundary line integration, while Zhang et al. applied Reynold's transport theorem for a similar decomposition. Both methods are intricate, necessitating silhouette edge sampling — a significant bottleneck in their application. Loubet et al. [28] proposed a variable change to simplify discontinuity integration, further refined by Bangaru et al. [3] using warped area fields to correct biases, yet practical application challenges persist. Despite their flexibility, the computational expense of such ray tracing frameworks restrict their practicality in optimization-intensive tasks [20]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 399, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 518 + ], + "type": "text", + "content": "For rasterization-based methods, rendering meshes differentiably is hindered by non-differentiable visibility at boundaries. Unlike ray tracing, which uses boundary integrals for visibility gradients, rasterization's fixed-grid and z-buffer approach lacks this feature. Solutions typically involve approximating derivatives or approximating rendering to achieve inherent differentiability. De La Gorce et al. [19] and Loper et al. [27] pioneered derivative approximation in rasterization. De La Gorce et al. distinguished between continuous regions and occlusion boundaries, introducing 'occlusion forces' for the latter. Loper et al. simplified this by detecting discontinuities post-rasterization and employing differential filters, like the Sobel filter, for boundary approximation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "The 'Nvdiffrast' approach by Laine et al. [20] tackles point-sampled visibility issues with innovative differentiable analytic antialIASing, which transforms sharp discontinuities into smooth transitions, enabling gradient computation. Analytical antialIASing is achieved by estimating coverage using silhouette edges. A limitation of the method is that if triangles containing silhouette edges do not overlap with any pixel center, they will be overlooked during the antialIASing process. Moreover, analytical antialIASing is inherently approximate, so the obtained gradients are also approximate. Regrettably, the approach is still quite complex, modifies the rasterized image, and requires a specialized data structure for connectivity and detailed pixel coverage computations. On the other hand, methods like Rhodin et al. [40] and Liu et al.'s 'Soft Rasterizer' [24] opt for modifying the rendering model. Rhodin et al. introduce fuzzy edges to soften discontinuities," + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rasterized Edge Gradients: Handling Discontinuities Differently" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "while 'Soft Rasterizer' further blurs boundaries and averages depth contributions, facilitating gradient propagation across occluded primitives." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "text", + "content": "In this work, we deliberately bypass physically based rendering, global illumination, and lighting/material models, and instead focus exclusively on rasterization-based approaches for their speed. We draw inspiration from the core idea of the edge sampling method [21] and analytical antialIASing [20] and distill this into a much simpler yet effective method that also provides accurate gradients. Notably, our method's simplicity allows direct handling of self-penetrating geometry, an aspect not addressed in previous research. Other methods to handle self-penetrating geometry would require complex and computationally intensive geometry preprocessing at each optimization step, involving intersection detection and differentiable splitting of the faces. In contrast, our approach explicitly manages interpenetrating geometry with virtually no additional overhead." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 286, + 234, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 286, + 234, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 286, + 234, + 300 + ], + "type": "text", + "content": "3 Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 307, + 479, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 479, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 479, + 329 + ], + "type": "text", + "content": "Following Li et al. [21], given a 2D pixel filter " + }, + { + "bbox": [ + 130, + 307, + 479, + 329 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 307, + 479, + 329 + ], + "type": "text", + "content": " and radiance " + }, + { + "bbox": [ + 130, + 307, + 479, + 329 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 307, + 479, + 329 + ], + "type": "text", + "content": ", a pixel's color can be written:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 244, + 329, + 480, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 329, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 244, + 329, + 480, + 354 + ], + "type": "interline_equation", + "content": "I = \\iint_ {D} k (x, y) R (x, y) d x d y. \\tag {1}", + "image_path": "921cea88dfab59c9bb175e3444e823bae312aa841fe1cd00e116c8c243580d52.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 354, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 413 + ], + "type": "text", + "content": "For notational convenience, as in [21], we denote " + }, + { + "bbox": [ + 130, + 354, + 482, + 413 + ], + "type": "inline_equation", + "content": "f(x,y) = k(x,y)R(x,y)" + }, + { + "bbox": [ + 130, + 354, + 482, + 413 + ], + "type": "text", + "content": ". For simplicity, let us first consider the case of a single pixel. We are interested in the gradient of " + }, + { + "bbox": [ + 130, + 354, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\frac{\\partial L(I)}{\\partial\\varPhi}" + }, + { + "bbox": [ + 130, + 354, + 482, + 413 + ], + "type": "text", + "content": ", where the scalar function " + }, + { + "bbox": [ + 130, + 354, + 482, + 413 + ], + "type": "inline_equation", + "content": "L(I)" + }, + { + "bbox": [ + 130, + 354, + 482, + 413 + ], + "type": "text", + "content": " of the rendered pixel " + }, + { + "bbox": [ + 130, + 354, + 482, + 413 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 354, + 482, + 413 + ], + "type": "text", + "content": " defines our loss, and " + }, + { + "bbox": [ + 130, + 354, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\varPhi" + }, + { + "bbox": [ + 130, + 354, + 482, + 413 + ], + "type": "text", + "content": " are the scene parameters that we aim to optimize. According to the chain rule:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 271, + 413, + 480, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 413, + 480, + 436 + ], + "spans": [ + { + "bbox": [ + 271, + 413, + 480, + 436 + ], + "type": "interline_equation", + "content": "\\frac {\\partial L (I)}{\\partial \\Phi} = \\frac {\\partial I}{\\partial \\Phi} \\frac {\\partial L}{\\partial I}, \\tag {2}", + "image_path": "6b3f12b1e5d812a5e834c56fea25d9b3641c1458ea5851dfa16446f427907fc7.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 437, + 479, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 479, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 479, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 437, + 479, + 475 + ], + "type": "inline_equation", + "content": "\\frac{\\partial L}{\\partial I}" + }, + { + "bbox": [ + 130, + 437, + 479, + 475 + ], + "type": "text", + "content": " is the incoming gradient from the loss function and is typically computed by automatic differentiation (AD). So, our goal becomes to find how the pixel value changes with respect to the scene parameters:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 246, + 479, + 480, + 503 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 479, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 246, + 479, + 480, + 503 + ], + "type": "interline_equation", + "content": "\\frac {\\partial I}{\\partial \\Phi} = \\frac {\\partial}{\\partial \\Phi} \\iint_ {D} f (x, y) d x d y. \\tag {3}", + "image_path": "0d39fc18fca8c4589f181a64642463c5f422781895f1fad07a96ea47368c4f03.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 506, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 555 + ], + "type": "text", + "content": "While the function " + }, + { + "bbox": [ + 130, + 506, + 482, + 555 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 506, + 482, + 555 + ], + "type": "text", + "content": " may not in fact be differentiable, its integral remains continuous and thus differentiable. Let us assume that " + }, + { + "bbox": [ + 130, + 506, + 482, + 555 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 506, + 482, + 555 + ], + "type": "text", + "content": " is partitioned into two continuous half-spaces, each represented by two continuous and differentiable functions " + }, + { + "bbox": [ + 130, + 506, + 482, + 555 + ], + "type": "inline_equation", + "content": "f_{a}" + }, + { + "bbox": [ + 130, + 506, + 482, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 506, + 482, + 555 + ], + "type": "inline_equation", + "content": "f_{b}" + }, + { + "bbox": [ + 130, + 506, + 482, + 555 + ], + "type": "text", + "content": ", respectively, as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 200, + 560, + 480, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 560, + 480, + 572 + ], + "spans": [ + { + "bbox": [ + 200, + 560, + 480, + 572 + ], + "type": "interline_equation", + "content": "f (x, y) = \\theta (\\alpha (x, y)) f _ {a} (x, y) + \\theta (- \\alpha (x, y)) f _ {b} (x, y), \\tag {4}", + "image_path": "118c333a21295d899343decb74833178b854db1800192edd7eeed88e261891c1.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 576, + 479, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 576, + 479, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 576, + 479, + 613 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 576, + 479, + 613 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 576, + 479, + 613 + ], + "type": "text", + "content": " specifies the dividing edge, and " + }, + { + "bbox": [ + 130, + 576, + 479, + 613 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 576, + 479, + 613 + ], + "type": "text", + "content": " is the Heaviside step function, which selects between " + }, + { + "bbox": [ + 130, + 576, + 479, + 613 + ], + "type": "inline_equation", + "content": "f_{a}" + }, + { + "bbox": [ + 130, + 576, + 479, + 613 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 576, + 479, + 613 + ], + "type": "inline_equation", + "content": "f_{b}" + }, + { + "bbox": [ + 130, + 576, + 479, + 613 + ], + "type": "text", + "content": ". The differentiation of the integral can be broken into two parts by applying the product rule. Let us start with the integrand " + }, + { + "bbox": [ + 130, + 576, + 479, + 613 + ], + "type": "inline_equation", + "content": "f_{a}" + }, + { + "bbox": [ + 130, + 576, + 479, + 613 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 144, + 616, + 480, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 616, + 480, + 668 + ], + "spans": [ + { + "bbox": [ + 144, + 616, + 480, + 668 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\frac {\\partial}{\\partial \\Phi} \\iint_ {D} \\theta (\\alpha (x, y)) f _ {a} (x, y) d x d y \\\\ = \\iint_ {D} \\delta (\\alpha (x, y)) \\frac {\\partial \\alpha (x , y)}{\\partial \\Phi} f _ {a} (x, y) d x d y + \\iint_ {D} \\frac {\\partial f _ {a} (x , y)}{\\partial \\Phi} \\theta (\\alpha (x, y)) d x d y, \\tag {5} \\\\ \\end{array}", + "image_path": "eafae01e994472f7ceb6cdebce3796bfdec82790a02b9d2df2ec67465b058f15.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "text", + "content": "S. Pidhorskyi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "content": "and a similar derivation can be done for " + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "inline_equation", + "content": "f_{b}" + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "content": ". The second term on the right-hand side of Eqn. (5) is the integral of the derivative of the smooth function, " + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "inline_equation", + "content": "f_{a}(x,y)" + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "content": ". This derivative is readily computable with AD, and in the context of differentiable rasterizers it typically involves differentiating through the interpolation step, the barycentric coordinates, and the texture sampler. Due to its straightforward computation, we will represent it as " + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "content": " and omit it from further discussion, focusing instead on the first term on the right-hand side of Eqn. (5). This term involves a Dirac delta function, and can be interpreted as taking the integral over the line " + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "inline_equation", + "content": "\\alpha(x,y) = 0" + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 205, + 232, + 480, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 232, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 205, + 232, + 480, + 285 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\iint_ {D} \\delta (\\alpha (x, y)) \\frac {\\partial \\alpha (x , y)}{\\partial \\Phi} f _ {a} (x, y) d x d y \\tag {6} \\\\ = \\int_ {\\alpha (x, y) = 0} \\frac {\\partial \\alpha (x , y)}{\\partial \\Phi} \\| \\nabla_ {x, y} \\alpha (x, y) \\| ^ {- 1} f _ {a} (x, y) d t, \\\\ \\end{array}", + "image_path": "33af76580b2b69af3624a1ce53a03323dcb1449911f0cb39ef5002fcdd9f437e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 291, + 479, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 291, + 479, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 291, + 479, + 327 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 291, + 479, + 327 + ], + "type": "inline_equation", + "content": "\\| \\nabla_{x,y}\\alpha (x,y)\\|" + }, + { + "bbox": [ + 130, + 291, + 479, + 327 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 130, + 291, + 479, + 327 + ], + "type": "inline_equation", + "content": "L^2" + }, + { + "bbox": [ + 130, + 291, + 479, + 327 + ], + "type": "text", + "content": " length of the edge equation gradient which accounts for the Jacobian of the variable substitution, and " + }, + { + "bbox": [ + 130, + 291, + 479, + 327 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 291, + 479, + 327 + ], + "type": "text", + "content": " is the parametrization of the line along which we integrate." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 343, + 201, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 201, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 201, + 357 + ], + "type": "text", + "content": "4 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 366, + 480, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 480, + 449 + ], + "type": "text", + "content": "In this section we describe our method that we call EdgeGrad. First we will focus on the core assumptions and approximations that form the basis of our method. Then we will see that due to our specifically chosen approximation, we can apply the prior method of edge sampling to rasterization. From there, we will focus on edge detection and classification heuristics, and our discretization scheme. We will conclude by describing an extension of our method for handling geometry intersections." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 467, + 251, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 467, + 251, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 467, + 251, + 479 + ], + "type": "text", + "content": "4.1 Rasterizing Edges" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 484, + 479, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 479, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 479, + 555 + ], + "type": "text", + "content": "Our primary objective is to differentiate rasterization while preserving its core image formation principles. We aim to calculate the gradients of pixel values relative to the mesh vertex positions, including gradients caused by visibility changes such as occlusion and disocclusion by primitive geometry. We seek to do this with a simple, minimal, and efficient method that does not necessitate modifying the initially rasterized image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 556, + 479, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 556, + 479, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 556, + 479, + 604 + ], + "type": "text", + "content": "Our main insight is that instead of approximating rasterization using a continuous representation, we assume the rasterized image is the result of some other continuous process that coincidentally aligns with our inherently non-differentiable, discrete-pixel rasterization. In essence, we opt for a direct ap" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rasterized Edge Gradients: Handling Discontinuities Differently" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 609, + 482, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 609, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 609, + 482, + 665 + ], + "type": "text", + "content": "1 This concept is akin to representing real numbers with floating-point arithmetic. Although there are only 4,294,967,296 single-precision floating-point numbers (a quarter of which lie in the interval [0,1]), they are still employed for continuous computation. We treat these computations as smooth and differentiable, even though they are, in essence, discontinuous." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 113, + 296, + 334 + ], + "blocks": [ + { + "bbox": [ + 141, + 113, + 296, + 334 + ], + "lines": [ + { + "bbox": [ + 141, + 113, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 141, + 113, + 296, + 334 + ], + "type": "image", + "image_path": "e6b735ec17688fcb4c5fce6d5226cc5ddac903e5c482938d0715501810ca7f43.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 342, + 304, + 352 + ], + "lines": [ + { + "bbox": [ + 132, + 342, + 304, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 304, + 352 + ], + "type": "text", + "content": "Fig.1: Micro-edges and Pixel Pair." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 353, + 305, + 407 + ], + "lines": [ + { + "bbox": [ + 132, + 353, + 305, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 305, + 407 + ], + "type": "text", + "content": "The figure illustrates construction of the micro-edges and the change in pixel intensities A and B with the edge position " + }, + { + "bbox": [ + 132, + 353, + 305, + 407 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 132, + 353, + 305, + 407 + ], + "type": "text", + "content": " movement, highlighting the C1 discontinuity at " + }, + { + "bbox": [ + 132, + 353, + 305, + 407 + ], + "type": "inline_equation", + "content": "p = 0" + }, + { + "bbox": [ + 132, + 353, + 305, + 407 + ], + "type": "text", + "content": ", absent in average intensity." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 312, + 114, + 478, + 343 + ], + "blocks": [ + { + "bbox": [ + 312, + 114, + 478, + 343 + ], + "lines": [ + { + "bbox": [ + 312, + 114, + 478, + 343 + ], + "spans": [ + { + "bbox": [ + 312, + 114, + 478, + 343 + ], + "type": "image", + "image_path": "1797519b671dc6da1b936dd2c896efe585d6957e8332a1c48df4d63e76ed2846.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 351, + 482, + 361 + ], + "lines": [ + { + "bbox": [ + 309, + 351, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 309, + 351, + 482, + 361 + ], + "type": "text", + "content": "Fig. 2: Gradient computation flow." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 362, + 481, + 407 + ], + "lines": [ + { + "bbox": [ + 309, + 362, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 309, + 362, + 481, + 407 + ], + "type": "text", + "content": "The figure illustrates the backward propagation process, where gradients from the loss function are traced back and scattered to the fragments." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 435, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 518 + ], + "type": "text", + "content": "proximation to compute gradients, instead of an indirect approach that first necessitates that the rendering be made continuous, for instance, through antialiasing filtering. More specifically, we use a construction where all straight edges, whether they are sides of triangles or intersections between triangles, are comprised of a collection of micro-edges. These micro-edges are assumed to be either strictly vertical or horizontal and are always positioned precisely between two pixels. An illustration of these constructions is shown in Figure 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "content": "Our approximation leads to several key properties: a) boundaries never intersect pixels, only lying exactly between them; b) pixels are always either fully covered or not covered at all; c) there is at most one boundary between any two neighbouring pixels; d) we do not need to access the source geometry to evaluate where the edge is exactly. With this, we can reinterpret the problem as if the rasterized image was in fact obtained by applying an antialIASing filter, albeit without visible effects; the resulting image remains identical to the pre-antialIASing source. Essentially our construct of the micro-edges renders the antialIASing filter as an identity operation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": "With this approximation, gradient estimation becomes a local operation, requiring the analysis one edge between two pixels at a time. Note that despite the approximation that all edges reside strictly between pixels, we do not assume" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "text", + "content": "S. Pidhorskyi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "that they are fixed, but rather allow them to freely move within an infinitesimal range, enabling differentiation of pixel values with respect to the edge location. This leads to the greatest advantage of our formulation; it allows us to greatly simplify both the theory and computation, since all edges are axis-aligned." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 181, + 216, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 181, + 216, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 181, + 216, + 192 + ], + "type": "text", + "content": "4.2 Pixel Pairs" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 200, + 479, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 479, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 479, + 248 + ], + "type": "text", + "content": "Let's analyze two adjacent pixels, A and B, separated by an edge, with their intensities represented as " + }, + { + "bbox": [ + 130, + 200, + 479, + 248 + ], + "type": "inline_equation", + "content": "I_{A}" + }, + { + "bbox": [ + 130, + 200, + 479, + 248 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 200, + 479, + 248 + ], + "type": "inline_equation", + "content": "I_{B}" + }, + { + "bbox": [ + 130, + 200, + 479, + 248 + ], + "type": "text", + "content": " respectively (refer to Figure 1 for details). The pixel intensities can be expressed using the Heaviside step function to account for the discontinuity at the edge as in Eqn. (4):" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 183, + 257, + 480, + 309 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 257, + 480, + 309 + ], + "spans": [ + { + "bbox": [ + 183, + 257, + 480, + 309 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} I _ {A} = \\iint_ {D _ {A}} \\theta (\\alpha (x, y)) f _ {a} (x, y) + \\theta (- \\alpha (x, y)) f _ {b} (x, y) d x d y \\\\ I _ {B} = \\iint_ {D _ {B}} \\theta (- \\alpha (x, y)) f _ {a} (x, y) + \\theta (\\alpha (x, y)) f _ {b} (x, y) d x d y. \\tag {7} \\\\ \\end{array}", + "image_path": "2db00a956c50f89942e6e2bc30eb0207e73a8ddb1641898db43d449c611919b7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 317, + 479, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 317, + 479, + 364 + ], + "spans": [ + { + "bbox": [ + 130, + 317, + 479, + 364 + ], + "type": "text", + "content": "Without loss of generality, assuming a horizontal pair we define " + }, + { + "bbox": [ + 130, + 317, + 479, + 364 + ], + "type": "inline_equation", + "content": "\\alpha(x,y) = p_{AB} - x" + }, + { + "bbox": [ + 130, + 317, + 479, + 364 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 317, + 479, + 364 + ], + "type": "inline_equation", + "content": "p_{AB}" + }, + { + "bbox": [ + 130, + 317, + 479, + 364 + ], + "type": "text", + "content": " is the edge location. Edge location " + }, + { + "bbox": [ + 130, + 317, + 479, + 364 + ], + "type": "inline_equation", + "content": "p_{AB}" + }, + { + "bbox": [ + 130, + 317, + 479, + 364 + ], + "type": "text", + "content": " is now the scene parameter with respect to which we want to differentiate the loss function. Since " + }, + { + "bbox": [ + 130, + 317, + 479, + 364 + ], + "type": "inline_equation", + "content": "p_{AB}" + }, + { + "bbox": [ + 130, + 317, + 479, + 364 + ], + "type": "text", + "content": " effects both " + }, + { + "bbox": [ + 130, + 317, + 479, + 364 + ], + "type": "inline_equation", + "content": "I_A" + }, + { + "bbox": [ + 130, + 317, + 479, + 364 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 317, + 479, + 364 + ], + "type": "inline_equation", + "content": "I_B" + }, + { + "bbox": [ + 130, + 317, + 479, + 364 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 235, + 364, + 479, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 364, + 479, + 388 + ], + "spans": [ + { + "bbox": [ + 235, + 364, + 479, + 388 + ], + "type": "interline_equation", + "content": "\\frac {\\partial L}{\\partial p _ {A B}} = \\frac {\\partial L}{\\partial I _ {A}} \\frac {\\partial I _ {A}}{\\partial p _ {A B}} + \\frac {\\partial L}{\\partial I _ {B}} \\frac {\\partial I _ {B}}{\\partial p _ {A B}}, \\tag {8}", + "image_path": "68049070430a189a6441d1c5be00b6dc70e1a8aaaa39318187bd57f81ed3684e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 392, + 480, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 480, + 485 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 392, + 480, + 485 + ], + "type": "inline_equation", + "content": "\\frac{\\partial I_A}{\\partial p_{AB}}" + }, + { + "bbox": [ + 130, + 392, + 480, + 485 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 392, + 480, + 485 + ], + "type": "inline_equation", + "content": "\\frac{\\partial I_B}{\\partial p_{AB}}" + }, + { + "bbox": [ + 130, + 392, + 480, + 485 + ], + "type": "text", + "content": " derive by applying Eqn. (6) to Eqn. (7). Unfortunately, with the edge strictly on the pixel boundary, we face a " + }, + { + "bbox": [ + 130, + 392, + 480, + 485 + ], + "type": "inline_equation", + "content": "C1" + }, + { + "bbox": [ + 130, + 392, + 480, + 485 + ], + "type": "text", + "content": " discontinuity at " + }, + { + "bbox": [ + 130, + 392, + 480, + 485 + ], + "type": "inline_equation", + "content": "p_{AB} = 0" + }, + { + "bbox": [ + 130, + 392, + 480, + 485 + ], + "type": "text", + "content": ", making intensity values non-differentiable. Infinitesimal movement of the edge affects " + }, + { + "bbox": [ + 130, + 392, + 480, + 485 + ], + "type": "inline_equation", + "content": "I_B" + }, + { + "bbox": [ + 130, + 392, + 480, + 485 + ], + "type": "text", + "content": " when moving right and " + }, + { + "bbox": [ + 130, + 392, + 480, + 485 + ], + "type": "inline_equation", + "content": "I_A" + }, + { + "bbox": [ + 130, + 392, + 480, + 485 + ], + "type": "text", + "content": " when moving left, without affecting the other. We still can compute one-sided limits of the derivatives, e.g. " + }, + { + "bbox": [ + 130, + 392, + 480, + 485 + ], + "type": "inline_equation", + "content": "\\frac{\\partial I_A}{\\partial p}^{-}" + }, + { + "bbox": [ + 130, + 392, + 480, + 485 + ], + "type": "text", + "content": " from the negative direction and " + }, + { + "bbox": [ + 130, + 392, + 480, + 485 + ], + "type": "inline_equation", + "content": "\\frac{\\partial I_B}{\\partial p}^{+}" + }, + { + "bbox": [ + 130, + 392, + 480, + 485 + ], + "type": "text", + "content": " from the positive direction. Applying Eqn. (6) and taking into account that the term " + }, + { + "bbox": [ + 130, + 392, + 480, + 485 + ], + "type": "inline_equation", + "content": "\\| \\nabla_{x,y}\\alpha (x,y)\\|" + }, + { + "bbox": [ + 130, + 392, + 480, + 485 + ], + "type": "text", + "content": " becomes 1 as we obtain:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 144, + 494, + 479, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 494, + 479, + 520 + ], + "spans": [ + { + "bbox": [ + 144, + 494, + 479, + 520 + ], + "type": "interline_equation", + "content": "\\frac {\\partial I _ {B}}{\\partial p _ {A B}} ^ {-} = \\frac {\\partial I _ {A}}{\\partial p _ {A B}} ^ {+} = 0; \\frac {\\partial I _ {B}}{\\partial p _ {A B}} ^ {+} = \\frac {\\partial I _ {A}}{\\partial p _ {A B}} ^ {-} = \\int_ {x = 0} [ f _ {a} (x, y) - f _ {b} (x, y) ] d y \\tag {9}", + "image_path": "dda407f514f3317a8b5172f22cf68f8f81cffdb431536aeb2667edd9580dc8ad.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 527, + 479, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 527, + 479, + 551 + ], + "spans": [ + { + "bbox": [ + 131, + 527, + 479, + 551 + ], + "type": "text", + "content": "We can circumvent the problem of " + }, + { + "bbox": [ + 131, + 527, + 479, + 551 + ], + "type": "inline_equation", + "content": "C_1" + }, + { + "bbox": [ + 131, + 527, + 479, + 551 + ], + "type": "text", + "content": " discontinuity by averaging the one-sided limits of the derivatives from the left and right:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 205, + 559, + 479, + 584 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 559, + 479, + 584 + ], + "spans": [ + { + "bbox": [ + 205, + 559, + 479, + 584 + ], + "type": "interline_equation", + "content": "\\frac {\\partial I _ {B}}{\\partial p _ {A B}} = \\frac {\\partial I _ {A}}{\\partial p _ {A B}} = \\frac {1}{2} \\int_ {x = 0} [ f _ {a} (x, y) - f _ {b} (x, y) ] d y \\tag {10}", + "image_path": "e98e8818c1896867b7436e1a6d262d1027290eed75262dbe27ed40fc64dd5a27.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 591, + 479, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 591, + 479, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 591, + 479, + 628 + ], + "type": "text", + "content": "Assuming that the pixels have constant value across all of their area (recall the rule that \"pixels are always either fully covered or not covered at all\") and substituting Eqn. (10) into Eqn. (8) we deduce:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 214, + 637, + 479, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 637, + 479, + 667 + ], + "spans": [ + { + "bbox": [ + 214, + 637, + 479, + 667 + ], + "type": "interline_equation", + "content": "\\left. \\frac {\\partial L}{\\partial p _ {A B}} = \\frac {1}{2} \\left(\\frac {\\partial L}{\\partial I _ {A}} + \\frac {\\partial L}{\\partial I _ {B}}\\right) \\left(I _ {A} - I _ {B}\\right) + \\Omega \\right], \\tag {11}", + "image_path": "f5a4830bb71d09a05e1ed6b1345120d648e0f99bab0c7a0a52e4f55a642c9b1b.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rasterized Edge Gradients: Handling Discontinuities Differently" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 176 + ], + "type": "text", + "content": "which is one of our main results. Here, " + }, + { + "bbox": [ + 130, + 115, + 479, + 176 + ], + "type": "inline_equation", + "content": "\\frac{\\partial L}{\\partial p_{AB}}" + }, + { + "bbox": [ + 130, + 115, + 479, + 176 + ], + "type": "text", + "content": " is the gradient of the loss function with respect to the position of the edge between pixels " + }, + { + "bbox": [ + 130, + 115, + 479, + 176 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 115, + 479, + 176 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 115, + 479, + 176 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 115, + 479, + 176 + ], + "type": "text", + "content": ". Additionally, we reintroduce " + }, + { + "bbox": [ + 130, + 115, + 479, + 176 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 130, + 115, + 479, + 176 + ], + "type": "text", + "content": " which is the second term on the right-hand side of Eqn. (5) that accounts for differentiating smooth regions of " + }, + { + "bbox": [ + 130, + 115, + 479, + 176 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 115, + 479, + 176 + ], + "type": "text", + "content": ". This term was omitted in Eqn.(9) and Eqn. (10) for brevity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 480, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 480, + 248 + ], + "type": "text", + "content": "Now, we need to accumulate these gradients into the gradients of the vertex positions of the triangle that created those edges. We do that first by scattering the gradients " + }, + { + "bbox": [ + 130, + 176, + 480, + 248 + ], + "type": "inline_equation", + "content": "\\frac{\\partial L}{\\partial p_{AB}}" + }, + { + "bbox": [ + 130, + 176, + 480, + 248 + ], + "type": "text", + "content": " to the fragments of the triangles associated with the pixels forming the edge and then by invoking the backward function of the interpolator module to gather the fragment gradients into the vertex gradients. Next, we will take a look at the gradient scattering mechanism." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 266, + 382, + 279 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 266, + 382, + 279 + ], + "spans": [ + { + "bbox": [ + 131, + 266, + 382, + 279 + ], + "type": "text", + "content": "4.3 Edge Classification and Gradient Scattering" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 288, + 480, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 288, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 288, + 480, + 422 + ], + "type": "text", + "content": "To backpropagate the gradients to the scene parameters we need to compute the gradient with respect to fragment positions, while in §4.2 we only derived the gradient " + }, + { + "bbox": [ + 130, + 288, + 480, + 422 + ], + "type": "inline_equation", + "content": "\\frac{\\partial L}{\\partial p_{AB}}" + }, + { + "bbox": [ + 130, + 288, + 480, + 422 + ], + "type": "text", + "content": " with respect to the edge position. We scatter each edge gradient " + }, + { + "bbox": [ + 130, + 288, + 480, + 422 + ], + "type": "inline_equation", + "content": "\\frac{\\partial L}{\\partial p_{AB}}" + }, + { + "bbox": [ + 130, + 288, + 480, + 422 + ], + "type": "text", + "content": " to adjacent fragments where we accumulate the fragment gradients, as illustrated in Figure 2. Using the chain rule, we obtain fragment gradients be multiplying " + }, + { + "bbox": [ + 130, + 288, + 480, + 422 + ], + "type": "inline_equation", + "content": "\\frac{\\partial L}{\\partial p_{AB}}" + }, + { + "bbox": [ + 130, + 288, + 480, + 422 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 288, + 480, + 422 + ], + "type": "inline_equation", + "content": "\\frac{\\partial p_{AB}}{\\partial r}" + }, + { + "bbox": [ + 130, + 288, + 480, + 422 + ], + "type": "text", + "content": "; the later relates the movement of the edge to the movement of the fragment, where " + }, + { + "bbox": [ + 130, + 288, + 480, + 422 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 288, + 480, + 422 + ], + "type": "text", + "content": " is the position of the fragment. The computation of " + }, + { + "bbox": [ + 130, + 288, + 480, + 422 + ], + "type": "inline_equation", + "content": "\\frac{\\partial p_{AB}}{\\partial r}" + }, + { + "bbox": [ + 130, + 288, + 480, + 422 + ], + "type": "text", + "content": " depends on the type of the edge formed by the fragments. We consider several cases: overhanging primitives (either one another or background); adjacent primitives; intersecting primitives. Unlike Laine et al. [20], we avoid additional data structures like hash maps for connectivity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 422, + 481, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 481, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 481, + 602 + ], + "type": "text", + "content": "Our approach examines each neighboring pixel pair, comparing their triangle IDs, see Figure 3 for illustration. Matching IDs indicate no edge; differing IDs prompt further tests. A single valid ID suggests a primitive overhanging the background. Testing a pixel's center against the triangle of the other pixel discerns overhangs and self-intersections. Non-contained pixels in each other's triangles imply adjacency. If both pixel centers are in the triangle of the other pixel, it implies self-intersections, and if in just one - then overhanging. For adjacent primitives we do not scatter gradients at all, since it is likely due to primitives sharing an edge and there is no discontinuity. Please note that even in that case " + }, + { + "bbox": [ + 130, + 422, + 481, + 602 + ], + "type": "inline_equation", + "content": "\\frac{\\partial L}{\\partial p_{AB}}" + }, + { + "bbox": [ + 130, + 422, + 481, + 602 + ], + "type": "text", + "content": " may be non-zero due to " + }, + { + "bbox": [ + 130, + 422, + 481, + 602 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 130, + 422, + 481, + 602 + ], + "type": "text", + "content": " contribution. In the overhanging case, we scatter the gradient only to the primitive on top, as the covered primitive can not affect the position of the edge. Thus, in such cases the covered fragment has " + }, + { + "bbox": [ + 130, + 422, + 481, + 602 + ], + "type": "inline_equation", + "content": "\\frac{\\partial p}{\\partial r} = 0" + }, + { + "bbox": [ + 130, + 422, + 481, + 602 + ], + "type": "text", + "content": " while the foreground fragment has " + }, + { + "bbox": [ + 130, + 422, + 481, + 602 + ], + "type": "inline_equation", + "content": "\\frac{\\partial p}{\\partial r} = 1" + }, + { + "bbox": [ + 130, + 422, + 481, + 602 + ], + "type": "text", + "content": " reflecting that the fragment movement is rigidly tied to the movement of the edge. Please refer to supplementary material for more details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 620, + 280, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 620, + 280, + 632 + ], + "spans": [ + { + "bbox": [ + 131, + 620, + 280, + 632 + ], + "type": "text", + "content": "4.4 Geometry Intersections" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 480, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 667 + ], + "type": "text", + "content": "This subsection details computing the derivative of edge location with respect to fragment position " + }, + { + "bbox": [ + 130, + 641, + 480, + 667 + ], + "type": "inline_equation", + "content": "\\frac{\\partial p}{\\partial r}" + }, + { + "bbox": [ + 130, + 641, + 480, + 667 + ], + "type": "text", + "content": " in scenarios where the edge arises from intersecting" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "text", + "content": "S. Pidhorskyi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 113, + 303, + 357 + ], + "blocks": [ + { + "bbox": [ + 133, + 113, + 303, + 357 + ], + "lines": [ + { + "bbox": [ + 133, + 113, + 303, + 357 + ], + "spans": [ + { + "bbox": [ + 133, + 113, + 303, + 357 + ], + "type": "image", + "image_path": "9cc86066d8317d857ce964646a3dc67a00101925ae815eb2251be0ca97ad0293.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 365, + 304, + 431 + ], + "lines": [ + { + "bbox": [ + 130, + 365, + 304, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 304, + 431 + ], + "type": "text", + "content": "Fig. 3: Edge classification. The accompanying figure provides a schematic representation of our edge classification process. Edges are categorized into four types: no edge, adjacent primitives, overhanging primitives, and intersecting primitives." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 294, + 482, + 425 + ], + "lines": [ + { + "bbox": [ + 307, + 294, + 482, + 425 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 482, + 425 + ], + "type": "text", + "content": "Fig. 4: Geometry intersection. A geometrical derivation for the partial derivative " + }, + { + "bbox": [ + 307, + 294, + 482, + 425 + ], + "type": "inline_equation", + "content": "\\frac{\\partial p}{\\partial r}" + }, + { + "bbox": [ + 307, + 294, + 482, + 425 + ], + "type": "text", + "content": ", essential for converting edge-related gradients to fragment-related gradients in cases where edges are formed by geometry intersections. This shows the x-z plane, where the render camera is viewing the scene along the z axis so that edges are parallel to the y axis (and perpendicular to the drawing). Projections of both intersecting primitives are transformed into two lines." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 118, + 481, + 286 + ], + "blocks": [ + { + "bbox": [ + 309, + 118, + 481, + 286 + ], + "lines": [ + { + "bbox": [ + 309, + 118, + 481, + 286 + ], + "spans": [ + { + "bbox": [ + 309, + 118, + 481, + 286 + ], + "type": "image", + "image_path": "82348e31d87e5b4a41c69b565e7ff1364fc52a1448726f80d4a788eab7633299.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 458, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 482, + 567 + ], + "type": "text", + "content": "geometry. This situation is inherently complex, and a general-case solution would be challenging. However, our micro-edge formulation streamlines the problem because intersection lines projected onto the image plane comprised of microedges, e.i. either vertical or horizontal. This allows us to analyze the problem in 2D, focusing on the x-z or y-z planes, with the z-axis perpendicular to the image plane. We derive the derivative of the edge location by analyzing two fragments of a micro-edge that belongs to the intersection line, focusing on how the edge would move in response to infinitesimal movements of each fragment. Without loss of generality we consider the x-z plane:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 236, + 578, + 481, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 578, + 481, + 607 + ], + "spans": [ + { + "bbox": [ + 236, + 578, + 481, + 607 + ], + "type": "interline_equation", + "content": "\\frac {\\partial p}{\\partial r} = - \\mathbf {n} _ {z} ^ {f} \\left[ \\mathbf {n} ^ {f ^ {T}} \\left[ \\begin{array}{l l} 0 & 1 \\\\ - 1 & 0 \\end{array} \\right] \\mathbf {n} ^ {v} \\right] ^ {- 1}, \\tag {12}", + "image_path": "be84143b4df876bbbf4f63d62782df71cda18452842171104bd104e02630da67.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": " is edge position, " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": " is the fragment position (in 3D clip space), " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{n}^v" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": " is the varying primitive's normal, and " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{n}^f" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": " is the fixed primitive's normal. In this context, the " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": "-axis lies on the image plane, and the " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": "-axis is perpendicular to it, see Fig. 4. We compute " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\frac{\\partial p}{\\partial r}" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": " for both fragments that form the intersection edge by" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rasterized Edge Gradients: Handling Discontinuities Differently" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "first fixing one fragment and varying the other, and then switching them. Please refer to supplementary material for derivation details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 159, + 371, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 159, + 371, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 159, + 371, + 172 + ], + "type": "text", + "content": "4.5 Theoretical analysis of the approximation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 479, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 479, + 242 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 479, + 242 + ], + "type": "text", + "content": "At first glance, replacing arbitrarily oriented edges with vertical and horizontal micro-edges might seem overly simplistic, raising doubts about accuracy. It could be argued, for example, that the perimeter defined by micro-edges doesn't match the original as micro-edge size approaches zero. However, applying the two-dimensional version of the divergence theorem from vector calculus offers insight:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 237, + 252, + 481, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 252, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 237, + 252, + 481, + 277 + ], + "type": "interline_equation", + "content": "\\oint_ {C} \\mathbf {f} \\cdot \\mathbf {n} d t = \\iint_ {D} (\\nabla_ {\\Phi} \\cdot \\mathbf {f}) d x d y, \\tag {13}", + "image_path": "51dd760a1dd0812420363fba6e969a5abeed2d134fcaeb7d9976eeea0f306c94.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 281, + 482, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 402 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 281, + 482, + 402 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 130, + 281, + 482, + 402 + ], + "type": "text", + "content": " is normal to boundary " + }, + { + "bbox": [ + 130, + 281, + 482, + 402 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 281, + 482, + 402 + ], + "type": "text", + "content": ", within which the integrand function is continuous and differentiable. We are free to pick any such boundaries that includes a discontinuity. In our case, the term " + }, + { + "bbox": [ + 130, + 281, + 482, + 402 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\alpha(x,y)}{\\partial\\Phi}" + }, + { + "bbox": [ + 130, + 281, + 482, + 402 + ], + "type": "text", + "content": " in Eqn. (6) is a normal vector to the boundary " + }, + { + "bbox": [ + 130, + 281, + 482, + 402 + ], + "type": "inline_equation", + "content": "\\alpha (x,y) = 0" + }, + { + "bbox": [ + 130, + 281, + 482, + 402 + ], + "type": "text", + "content": " thus the integral is equivalent to the left-hand side of Eqn. (13). According to Eqn. (13), this recasts the boundary derivative integral as an area integral. The difference between the area integrals of micro-edges and original edges is confined to a space defined by the micro-edges size. As pixel size diminishes, the integral approaches the exact value, and the discrepancy area vanishes, confirming the micro-edge approach converges to the same solution as the original formulation in the limit." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 422, + 289, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 422, + 289, + 436 + ], + "spans": [ + { + "bbox": [ + 132, + 422, + 289, + 436 + ], + "type": "text", + "content": "5 Experimental Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 449, + 480, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 480, + 544 + ], + "type": "text", + "content": "We implement our method in PyTorch with custom CUDA kernels, and largely follow the structure of nvdiffrast [20], with the exception for the antialIASing module which we replace with our EdgeGrad module. For all experiments we utilize Laplacian preconditioning proposed by Nicolet et al. [34]. We find Laplacian preconditioning extremely helpful for stabilizing training and we use it for both, meshes and texture data. While Laplacian preconditioning may reduce self-intersections, this happens only due to much smoother optimization process, but it does not account for them in any way." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "type": "text", + "content": "Forward gradients. We first analyze the performance of the proposed approach by evaluating the forward gradients on a variety of test cases. Forward gradients are per-pixel derivatives with respect to some scene parameter, which is depicted with white arrows on the scene images. Fig. 5 shows the forward gradients computed using finite differences, raytracing methods Mitsuba [14,15,36] and redner [21], and rasterization approaches including nvdiffrast [20] and our approach (EdgeGrad). Unlike other methods, our approach still matches finite differences gradients in cases of intersecting geometry. While Mitsuba and Redner, which require multiple samples per pixel (256 in these comparisons), typically offer higher accuracy, they fall short at geometry intersections. Nvdiffrast" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "text", + "content": "S. Pidhorskyi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 114, + 473, + 338 + ], + "blocks": [ + { + "bbox": [ + 142, + 114, + 473, + 338 + ], + "lines": [ + { + "bbox": [ + 142, + 114, + 473, + 338 + ], + "spans": [ + { + "bbox": [ + 142, + 114, + 473, + 338 + ], + "type": "image", + "image_path": "f6d8019b5f1778c9efb33293a150d0dd0f86b8cf3e380987e407fb2e9013173f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 346, + 480, + 390 + ], + "lines": [ + { + "bbox": [ + 130, + 346, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 480, + 390 + ], + "type": "text", + "content": "Fig. 5: Comparison of forward gradient on several test scenes with the numerical solution using finite differences. (a) Synthetic test scene, we show gradients with respect to a parameter. We use (b) finite differences as a reference, and compare with (c) Mitsuba 3 [15,36], (d) redner [21], (e) nvdiffrast [20], and (f) our edge gradient approach." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 414, + 479, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 479, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 479, + 437 + ], + "type": "text", + "content": "is noisier because it assumes triangles containing silhouette edges always overlap pixel centers, which often isn't the case, leading to omitted gradients." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 437, + 480, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 480, + 485 + ], + "type": "text", + "content": "Backward gradients. Tab. 1 quantitatively analyzes the accuracy of the backward gradients. We show improved results especially for complex intersecting geometry. Compared to raytracing methods, our rasterization-based approach yields comparable performance at a fraction of the computational cost." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 480, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 480, + 557 + ], + "type": "text", + "content": "Runtime performance end accuracy by image size. Fig. 6 shows running time of our method (EdgeGrad) compared to Mitsuba [36] and nvdiffrast [20]. Our method significantly outperforms other approaches in running time on all variety of image sizes and triangle count. Our method has comparable accuracy to Mitsuba [36] (and even better for the case with geometry intersections) at large image sizes, but performs worse for smaller image size due to sub pixel triangles." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "content": "Quantitative scene reconstructions. Tab. 2 shows quantitative results of scene reconstruction on blender dataset [32]. We follow the testing procedure described in [32] and report PNSR, SSIM and LPIPS metrics. We compare our method (EdgeGrad) with a \"continues only\" baseline (which ignores gradients from discontinuities), nvdiffrast [20], and with a variant of our method that does not have a special handling for the geometry intersections. Method optimization procedure is similar to Nicolet et al. [34] we only add MLP for decoding texture values in a view conditioned manner similar to [29]. We initialize geometry with some coarse manually created meshes shown in Fig. 7. To avoid additional complex" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rasterized Edge Gradients: Handling Discontinuities Differently" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 186, + 133, + 219, + 167 + ], + "blocks": [ + { + "bbox": [ + 149, + 146, + 184, + 154 + ], + "lines": [ + { + "bbox": [ + 149, + 146, + 184, + 154 + ], + "spans": [ + { + "bbox": [ + 149, + 146, + 184, + 154 + ], + "type": "text", + "content": "Scene" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 186, + 133, + 219, + 167 + ], + "lines": [ + { + "bbox": [ + 186, + 133, + 219, + 167 + ], + "spans": [ + { + "bbox": [ + 186, + 133, + 219, + 167 + ], + "type": "image", + "image_path": "b643c42ff89145154b5bab67e750934303441922c0807c1c48f7a3cf15489d8b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 220, + 134, + 252, + 167 + ], + "blocks": [ + { + "bbox": [ + 220, + 134, + 252, + 167 + ], + "lines": [ + { + "bbox": [ + 220, + 134, + 252, + 167 + ], + "spans": [ + { + "bbox": [ + 220, + 134, + 252, + 167 + ], + "type": "image", + "image_path": "090f379491bb8630c9df89a6493fc5c8f5022bd3eb23e6e0ae9f9b2fe088ff27.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 253, + 134, + 286, + 167 + ], + "blocks": [ + { + "bbox": [ + 253, + 134, + 286, + 167 + ], + "lines": [ + { + "bbox": [ + 253, + 134, + 286, + 167 + ], + "spans": [ + { + "bbox": [ + 253, + 134, + 286, + 167 + ], + "type": "image", + "image_path": "ee24749cb0742bd6c8bc88c6930b2ba7c40680db608f7604804287caa8578db2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 136, + 167, + 287, + 203 + ], + "blocks": [ + { + "bbox": [ + 136, + 167, + 287, + 203 + ], + "lines": [ + { + "bbox": [ + 136, + 167, + 287, + 203 + ], + "spans": [ + { + "bbox": [ + 136, + 167, + 287, + 203 + ], + "type": "table", + "html": "
redner [21]1.20%53.52%37.94%
Mitsuba3 [36]0.67%75.83%37.59%
Nvdiffrast [20]45.91%69.33%39.42%
EdgeGrad(our)6.01%3.35%8.35%
", + "image_path": "2152a1021f050952fa49ab719df62f72b7985c762a744c5c424565a68b56d7b4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 212, + 304, + 289 + ], + "lines": [ + { + "bbox": [ + 132, + 212, + 304, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 212, + 304, + 289 + ], + "type": "text", + "content": "Table 1: Accuracy of backward gradients. Relative error, " + }, + { + "bbox": [ + 132, + 212, + 304, + 289 + ], + "type": "inline_equation", + "content": "\\%" + }, + { + "bbox": [ + 132, + 212, + 304, + 289 + ], + "type": "text", + "content": " (↓). This table shows relative errors in backward gradient computations for test scenes. Second and third scenes include geometry intersections, emphasizing our method's advantage in managing these complexities." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 309, + 115, + 394, + 166 + ], + "blocks": [ + { + "bbox": [ + 309, + 115, + 394, + 166 + ], + "lines": [ + { + "bbox": [ + 309, + 115, + 394, + 166 + ], + "spans": [ + { + "bbox": [ + 309, + 115, + 394, + 166 + ], + "type": "image", + "image_path": "389c0b8eef6169a730b9a8abcf6fda6dec616be3a023323a988728951e109ade.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 394, + 116, + 477, + 164 + ], + "blocks": [ + { + "bbox": [ + 394, + 116, + 477, + 164 + ], + "lines": [ + { + "bbox": [ + 394, + 116, + 477, + 164 + ], + "spans": [ + { + "bbox": [ + 394, + 116, + 477, + 164 + ], + "type": "image", + "image_path": "6485593257a823179f267d5d321d747f931a3337141f962a7b0d8db211654e31.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 172, + 394, + 220 + ], + "blocks": [ + { + "bbox": [ + 318, + 167, + 389, + 172 + ], + "lines": [ + { + "bbox": [ + 318, + 167, + 389, + 172 + ], + "spans": [ + { + "bbox": [ + 318, + 167, + 389, + 172 + ], + "type": "text", + "content": "(a) Running time vs image size" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 172, + 394, + 220 + ], + "lines": [ + { + "bbox": [ + 309, + 172, + 394, + 220 + ], + "spans": [ + { + "bbox": [ + 309, + 172, + 394, + 220 + ], + "type": "image", + "image_path": "d90e8eeb0171a38c1b491bfb4c88d81263633763c33fe984056580f10e65f4fd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 221, + 386, + 232 + ], + "lines": [ + { + "bbox": [ + 315, + 221, + 386, + 232 + ], + "spans": [ + { + "bbox": [ + 315, + 221, + 386, + 232 + ], + "type": "text", + "content": "(c) Gradient error vs image size self-intersections absent" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 395, + 173, + 477, + 220 + ], + "blocks": [ + { + "bbox": [ + 398, + 167, + 477, + 172 + ], + "lines": [ + { + "bbox": [ + 398, + 167, + 477, + 172 + ], + "spans": [ + { + "bbox": [ + 398, + 167, + 477, + 172 + ], + "type": "text", + "content": "(b) Running time vs triangle count" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 395, + 173, + 477, + 220 + ], + "lines": [ + { + "bbox": [ + 395, + 173, + 477, + 220 + ], + "spans": [ + { + "bbox": [ + 395, + 173, + 477, + 220 + ], + "type": "image", + "image_path": "8a12308d9f1178894f10228224d00969ad3d3f2c4643e6f2879325be6ed49e7a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 221, + 471, + 232 + ], + "lines": [ + { + "bbox": [ + 399, + 221, + 471, + 232 + ], + "spans": [ + { + "bbox": [ + 399, + 221, + 471, + 232 + ], + "type": "text", + "content": "(d) Gradient error vs image size self-intersections present" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 133, + 318, + 228, + 392 + ], + "blocks": [ + { + "bbox": [ + 133, + 318, + 228, + 392 + ], + "lines": [ + { + "bbox": [ + 133, + 318, + 228, + 392 + ], + "spans": [ + { + "bbox": [ + 133, + 318, + 228, + 392 + ], + "type": "image", + "image_path": "8487633fe2caeb03f444e1e52bf15afd80a1a9cebc2ed6f5cede1d6f6aea9e2e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 392, + 175, + 397 + ], + "lines": [ + { + "bbox": [ + 138, + 392, + 175, + 397 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 175, + 397 + ], + "type": "text", + "content": "(a) camera image" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 230, + 323, + 268, + 392 + ], + "blocks": [ + { + "bbox": [ + 230, + 323, + 268, + 392 + ], + "lines": [ + { + "bbox": [ + 230, + 323, + 268, + 392 + ], + "spans": [ + { + "bbox": [ + 230, + 323, + 268, + 392 + ], + "type": "image", + "image_path": "4ddc715f2aa7b1e89c480bc319b18971613a7ac6ab20ea0b53f2f2a977334999.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 392, + 265, + 403 + ], + "lines": [ + { + "bbox": [ + 231, + 392, + 265, + 403 + ], + "spans": [ + { + "bbox": [ + 231, + 392, + 265, + 403 + ], + "type": "text", + "content": "(c) initialization mesh" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 275, + 323, + 311, + 390 + ], + "blocks": [ + { + "bbox": [ + 275, + 323, + 311, + 390 + ], + "lines": [ + { + "bbox": [ + 275, + 323, + 311, + 390 + ], + "spans": [ + { + "bbox": [ + 275, + 323, + 311, + 390 + ], + "type": "image", + "image_path": "8d62125d94ee7e410c80a1043a19cf88a70db77957a8acf96ee7a275317ad8ca.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 392, + 306, + 402 + ], + "lines": [ + { + "bbox": [ + 277, + 392, + 306, + 402 + ], + "spans": [ + { + "bbox": [ + 277, + 392, + 306, + 402 + ], + "type": "text", + "content": "(d) optimized mesh" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 315, + 319, + 410, + 391 + ], + "blocks": [ + { + "bbox": [ + 307, + 241, + 481, + 307 + ], + "lines": [ + { + "bbox": [ + 307, + 241, + 481, + 307 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 481, + 307 + ], + "type": "text", + "content": "Fig. 6: Runtime performance and errors. a) Runtime [ms] by image size [px]; b) runtime [ms] by triangle count; c) gradient error [%] by image size [px] without self-intersections; d) gradient error [%] by image size [px] with self-intersections." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 319, + 410, + 391 + ], + "lines": [ + { + "bbox": [ + 315, + 319, + 410, + 391 + ], + "spans": [ + { + "bbox": [ + 315, + 319, + 410, + 391 + ], + "type": "image", + "image_path": "dffc9ff9e2f524a9048ed246965a1c6631296f411bd55bffca9df4c33d233147.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 392, + 354, + 396 + ], + "lines": [ + { + "bbox": [ + 316, + 392, + 354, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 392, + 354, + 396 + ], + "type": "text", + "content": "(e)rgb estimation" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 363, + 392, + 399, + 403 + ], + "lines": [ + { + "bbox": [ + 363, + 392, + 399, + 403 + ], + "spans": [ + { + "bbox": [ + 363, + 392, + 399, + 403 + ], + "type": "text", + "content": "(f) segmentation estimation" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 411, + 385, + 481, + 398 + ], + "lines": [ + { + "bbox": [ + 411, + 385, + 481, + 398 + ], + "spans": [ + { + "bbox": [ + 411, + 385, + 481, + 398 + ], + "type": "text", + "content": "(g) Learned segmentation texture for mouth region" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 415, + 322, + 477, + 384 + ], + "blocks": [ + { + "bbox": [ + 415, + 322, + 477, + 384 + ], + "lines": [ + { + "bbox": [ + 415, + 322, + 477, + 384 + ], + "spans": [ + { + "bbox": [ + 415, + 322, + 477, + 384 + ], + "type": "image", + "image_path": "274eb69f9c16bb8ac44903afd294142c7d3ef8e5c41a8215ac8445622ad4b0b1.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 140, + 403, + 164, + 438 + ], + "blocks": [ + { + "bbox": [ + 140, + 403, + 164, + 438 + ], + "lines": [ + { + "bbox": [ + 140, + 403, + 164, + 438 + ], + "spans": [ + { + "bbox": [ + 140, + 403, + 164, + 438 + ], + "type": "image", + "image_path": "88258279ed633583bb90674a385ed9df0235b0c3711379b53a7e4a1f9c96c1e1.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 440, + 164, + 445 + ], + "lines": [ + { + "bbox": [ + 143, + 440, + 164, + 445 + ], + "spans": [ + { + "bbox": [ + 143, + 440, + 164, + 445 + ], + "type": "text", + "content": "Initialization" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 176, + 403, + 198, + 437 + ], + "blocks": [ + { + "bbox": [ + 176, + 403, + 198, + 437 + ], + "lines": [ + { + "bbox": [ + 176, + 403, + 198, + 437 + ], + "spans": [ + { + "bbox": [ + 176, + 403, + 198, + 437 + ], + "type": "image", + "image_path": "b39e087da88a6328eb84f34e39e8d7a705043941a12e24b2cd917f1f3a5da157.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 180, + 440, + 198, + 445 + ], + "lines": [ + { + "bbox": [ + 180, + 440, + 198, + 445 + ], + "spans": [ + { + "bbox": [ + 180, + 440, + 198, + 445 + ], + "type": "text", + "content": "Optimized" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 202, + 406, + 256, + 436 + ], + "blocks": [ + { + "bbox": [ + 202, + 406, + 256, + 436 + ], + "lines": [ + { + "bbox": [ + 202, + 406, + 256, + 436 + ], + "spans": [ + { + "bbox": [ + 202, + 406, + 256, + 436 + ], + "type": "image", + "image_path": "248791888a2375df21f3a3a8665e5033f5e575120217a7e35942008366628364.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 439, + 233, + 444 + ], + "lines": [ + { + "bbox": [ + 212, + 439, + 233, + 444 + ], + "spans": [ + { + "bbox": [ + 212, + 439, + 233, + 444 + ], + "type": "text", + "content": "Initialization" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 257, + 406, + 290, + 436 + ], + "blocks": [ + { + "bbox": [ + 257, + 406, + 290, + 436 + ], + "lines": [ + { + "bbox": [ + 257, + 406, + 290, + 436 + ], + "spans": [ + { + "bbox": [ + 257, + 406, + 290, + 436 + ], + "type": "image", + "image_path": "560488f97baf3b0cb4e1ca86a411f69d3960957e1f6e6f24ea5b4e14a165ab10.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 265, + 439, + 283, + 444 + ], + "lines": [ + { + "bbox": [ + 265, + 439, + 283, + 444 + ], + "spans": [ + { + "bbox": [ + 265, + 439, + 283, + 444 + ], + "type": "text", + "content": "1" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 290, + 406, + 342, + 436 + ], + "blocks": [ + { + "bbox": [ + 290, + 406, + 342, + 436 + ], + "lines": [ + { + "bbox": [ + 290, + 406, + 342, + 436 + ], + "spans": [ + { + "bbox": [ + 290, + 406, + 342, + 436 + ], + "type": "image", + "image_path": "2c8d35fbfa938274b6c868dd885f44079506b12699137ff86afc5e0bb28d81f0.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 299, + 440, + 303, + 444 + ], + "blocks": [ + { + "bbox": [ + 299, + 440, + 303, + 444 + ], + "lines": [ + { + "bbox": [ + 299, + 440, + 303, + 444 + ], + "spans": [ + { + "bbox": [ + 299, + 440, + 303, + 444 + ], + "type": "image", + "image_path": "8d085686284e5bd645fe4691dd3ffa480fcc5611dcd4e7543ab1d93ad0d699f0.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 440, + 335, + 445 + ], + "lines": [ + { + "bbox": [ + 306, + 440, + 335, + 445 + ], + "spans": [ + { + "bbox": [ + 306, + 440, + 335, + 445 + ], + "type": "text", + "content": "Optimized" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 231, + 445, + 389, + 452 + ], + "lines": [ + { + "bbox": [ + 231, + 445, + 389, + 452 + ], + "spans": [ + { + "bbox": [ + 231, + 445, + 389, + 452 + ], + "type": "text", + "content": "(h) Initialization mesh and reconstructed mesh for synthetic blender dataset" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 348, + 406, + 375, + 436 + ], + "blocks": [ + { + "bbox": [ + 348, + 406, + 375, + 436 + ], + "lines": [ + { + "bbox": [ + 348, + 406, + 375, + 436 + ], + "spans": [ + { + "bbox": [ + 348, + 406, + 375, + 436 + ], + "type": "image", + "image_path": "a5c41640ef10d4bf8cbf4df6fa7d4ce9ad0a6fda595aa97fb9ad4aa6f6fe7fe0.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 440, + 370, + 445 + ], + "lines": [ + { + "bbox": [ + 349, + 440, + 370, + 445 + ], + "spans": [ + { + "bbox": [ + 349, + 440, + 370, + 445 + ], + "type": "text", + "content": "Initialization" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 381, + 406, + 408, + 436 + ], + "blocks": [ + { + "bbox": [ + 381, + 406, + 408, + 436 + ], + "lines": [ + { + "bbox": [ + 381, + 406, + 408, + 436 + ], + "spans": [ + { + "bbox": [ + 381, + 406, + 408, + 436 + ], + "type": "image", + "image_path": "e87f739b8c0b08f09e956919173c7cb025c3be4bd014f80975f5c1a94bc1e6da.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 440, + 403, + 445 + ], + "lines": [ + { + "bbox": [ + 386, + 440, + 403, + 445 + ], + "spans": [ + { + "bbox": [ + 386, + 440, + 403, + 445 + ], + "type": "text", + "content": "Optimized" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 410, + 404, + 481, + 437 + ], + "blocks": [ + { + "bbox": [ + 410, + 404, + 481, + 437 + ], + "lines": [ + { + "bbox": [ + 410, + 404, + 481, + 437 + ], + "spans": [ + { + "bbox": [ + 410, + 404, + 481, + 437 + ], + "type": "image", + "image_path": "e9bc93948f9064931b0977efddcdc988349450976a03e4a2f58f294208c5dfb3.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 440, + 438, + 445 + ], + "lines": [ + { + "bbox": [ + 417, + 440, + 438, + 445 + ], + "spans": [ + { + "bbox": [ + 417, + 440, + 438, + 445 + ], + "type": "text", + "content": "Initialization" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 454, + 440, + 471, + 445 + ], + "lines": [ + { + "bbox": [ + 454, + 440, + 471, + 445 + ], + "spans": [ + { + "bbox": [ + 454, + 440, + 471, + 445 + ], + "type": "text", + "content": "Optimized" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 135, + 528, + 161, + 587 + ], + "blocks": [ + { + "bbox": [ + 135, + 528, + 161, + 587 + ], + "lines": [ + { + "bbox": [ + 135, + 528, + 161, + 587 + ], + "spans": [ + { + "bbox": [ + 135, + 528, + 161, + 587 + ], + "type": "image", + "image_path": "270631c5ca503c37c14aa3c9427544ccfb88f18e097cebcb6a87ead44ba9ee00.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 590, + 149, + 594 + ], + "lines": [ + { + "bbox": [ + 141, + 590, + 149, + 594 + ], + "spans": [ + { + "bbox": [ + 141, + 590, + 149, + 594 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 164, + 529, + 190, + 588 + ], + "blocks": [ + { + "bbox": [ + 164, + 529, + 190, + 588 + ], + "lines": [ + { + "bbox": [ + 164, + 529, + 190, + 588 + ], + "spans": [ + { + "bbox": [ + 164, + 529, + 190, + 588 + ], + "type": "image", + "image_path": "b8eca095d0ea7e666d6fecef422be1dece8d1cfabca9e37a2c9f0396b88b7560.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 590, + 180, + 594 + ], + "lines": [ + { + "bbox": [ + 165, + 590, + 180, + 594 + ], + "spans": [ + { + "bbox": [ + 165, + 590, + 180, + 594 + ], + "type": "text", + "content": "NvdiffRast" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 190, + 527, + 228, + 589 + ], + "blocks": [ + { + "bbox": [ + 131, + 460, + 482, + 517 + ], + "lines": [ + { + "bbox": [ + 131, + 460, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 131, + 460, + 482, + 517 + ], + "type": "text", + "content": "Fig. 7: Experimental setup details. The figure shows: a) the ground truth camera image; b) predicted segmentation mask for mouth region using HRNet [47]; c) initialization mesh from 3DMM; d) optimized avatar mesh; e) rendered image of the optimized model; f) rendered segmentation estimation image; g) learned static segmentation texture; h) initialization meshes and optimized meshes for blender dataset [32]." + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 190, + 527, + 228, + 589 + ], + "lines": [ + { + "bbox": [ + 190, + 527, + 228, + 589 + ], + "spans": [ + { + "bbox": [ + 190, + 527, + 228, + 589 + ], + "type": "image", + "image_path": "648a873f103fe7c1789773e075752bdf110c8735d327c4768ee1d1a1921ad351.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 193, + 590, + 217, + 594 + ], + "lines": [ + { + "bbox": [ + 193, + 590, + 217, + 594 + ], + "spans": [ + { + "bbox": [ + 193, + 590, + 217, + 594 + ], + "type": "text", + "content": "No Edge Handling" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 230, + 533, + 257, + 585 + ], + "blocks": [ + { + "bbox": [ + 230, + 533, + 257, + 585 + ], + "lines": [ + { + "bbox": [ + 230, + 533, + 257, + 585 + ], + "spans": [ + { + "bbox": [ + 230, + 533, + 257, + 585 + ], + "type": "image", + "image_path": "44817837ff5de8254c6ce671dfdd446af19918435c0f2ec5acb645579b178cc1.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 589, + 255, + 594 + ], + "lines": [ + { + "bbox": [ + 247, + 589, + 255, + 594 + ], + "spans": [ + { + "bbox": [ + 247, + 589, + 255, + 594 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 260, + 534, + 290, + 585 + ], + "blocks": [ + { + "bbox": [ + 260, + 534, + 290, + 585 + ], + "lines": [ + { + "bbox": [ + 260, + 534, + 290, + 585 + ], + "spans": [ + { + "bbox": [ + 260, + 534, + 290, + 585 + ], + "type": "image", + "image_path": "5d422cfe9914f436eb1aae59526f6f6d8892d0be75fd0816abfb0fa7b9005d0c.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 271, + 589, + 285, + 594 + ], + "lines": [ + { + "bbox": [ + 271, + 589, + 285, + 594 + ], + "spans": [ + { + "bbox": [ + 271, + 589, + 285, + 594 + ], + "type": "text", + "content": "NvdRast" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_caption" + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 290, + 534, + 325, + 585 + ], + "blocks": [ + { + "bbox": [ + 290, + 534, + 325, + 585 + ], + "lines": [ + { + "bbox": [ + 290, + 534, + 325, + 585 + ], + "spans": [ + { + "bbox": [ + 290, + 534, + 325, + 585 + ], + "type": "image", + "image_path": "e5d35610e079806573a70c835612e36d6bb364ffb6f20d874affdea6727cc56e.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 298, + 589, + 322, + 594 + ], + "lines": [ + { + "bbox": [ + 298, + 589, + 322, + 594 + ], + "spans": [ + { + "bbox": [ + 298, + 589, + 322, + 594 + ], + "type": "text", + "content": "No Edge Handling" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_caption" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 328, + 537, + 376, + 591 + ], + "blocks": [ + { + "bbox": [ + 328, + 537, + 376, + 591 + ], + "lines": [ + { + "bbox": [ + 328, + 537, + 376, + 591 + ], + "spans": [ + { + "bbox": [ + 328, + 537, + 376, + 591 + ], + "type": "image", + "image_path": "93b5754078ade38f45e2087b8334616cd999a75c9dbfff144d132b399d7c97b4.jpg" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 590, + 359, + 594 + ], + "lines": [ + { + "bbox": [ + 352, + 590, + 359, + 594 + ], + "spans": [ + { + "bbox": [ + 352, + 590, + 359, + 594 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 605, + 482, + 660 + ], + "lines": [ + { + "bbox": [ + 131, + 605, + 482, + 660 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 482, + 660 + ], + "type": "text", + "content": "Fig. 8: Reconstructions from inverse rendering compare our method with nvdiffrast [20] and rasterization without discontinuity handling. Although nvdiffrast can match the outline of objects well, it suffers from interpenetration (see leaves of Ficus) and does not capture as much detail as ours (see the backrest of the chair and ground of lego). The method with no edge handling fails completely on these objects." + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_caption" + } + ], + "index": 60 + }, + { + "type": "image", + "bbox": [ + 376, + 537, + 425, + 590 + ], + "blocks": [ + { + "bbox": [ + 376, + 537, + 425, + 590 + ], + "lines": [ + { + "bbox": [ + 376, + 537, + 425, + 590 + ], + "spans": [ + { + "bbox": [ + 376, + 537, + 425, + 590 + ], + "type": "image", + "image_path": "17418c45a056fc50c1a8f05c1b7623d4ef36bf41564254015e566cd33678ae3a.jpg" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 398, + 590, + 412, + 594 + ], + "lines": [ + { + "bbox": [ + 398, + 590, + 412, + 594 + ], + "spans": [ + { + "bbox": [ + 398, + 590, + 412, + 594 + ], + "type": "text", + "content": "NdivdiffRast" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_caption" + } + ], + "index": 62 + }, + { + "type": "image", + "bbox": [ + 426, + 527, + 477, + 587 + ], + "blocks": [ + { + "bbox": [ + 426, + 527, + 477, + 587 + ], + "lines": [ + { + "bbox": [ + 426, + 527, + 477, + 587 + ], + "spans": [ + { + "bbox": [ + 426, + 527, + 477, + 587 + ], + "type": "image", + "image_path": "2b9cd60707d6f6f32529b7e4faaca333e8627c9c39b9a44473a9c98985c0f27b.jpg" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 589, + 467, + 594 + ], + "lines": [ + { + "bbox": [ + 443, + 589, + 467, + 594 + ], + "spans": [ + { + "bbox": [ + 443, + 589, + 467, + 594 + ], + "type": "text", + "content": "No Edge Handling" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_caption" + } + ], + "index": 64 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 245, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 245, + 101 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 245, + 101 + ], + "type": "text", + "content": "S. Pidhorskyi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 114, + 474, + 239 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 474, + 239 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 474, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 474, + 239 + ], + "type": "table", + "html": "
LegoChairShipHotdogFicusDrumsMicMaterials
PSNR ↑Continuous only16.34228.15115.69327.85018.39915.53421.79717.11
Nvdiffrast [20]29.4429.80426.17331.68424.96423.15429.63824.413
EdgeGrad -intersect.29.57433.07726.73134.22526.78124.0831.22425.221
EdgeGrad (our)29.66732.98126.6534.08126.55924.01431.34625.287
LegoChairShipHotdogFicusDrumsMicMaterials
SSIM ↑Continuous only0.75500.95310.69290.93740.82450.77230.91820.7693
Nvdiffrast [20]0.94670.94160.83160.94730.93230.89710.96470.8865
EdgeGrad -intersect.0.94850.97650.84780.97150.94680.92160.97560.8945
EdgeGrad (our)0.95010.97640.84720.97130.94520.92200.97670.8972
LegoChairShipHotdogFicusDrumsMicMaterials
LPIPS ↓Continuous only0.32230.06110.39130.11090.20460.29430.12720.2511
Nvdiffrast [20]0.06530.07710.20200.09360.07810.11160.05490.1291
EdgeGrad -intersect.0.06330.03480.18990.05460.06810.08710.04390.1200
EdgeGrad (our)0.06140.03410.18500.05480.06720.08700.03990.1185
", + "image_path": "af1dc1de693ba3d88a7be3444fa3a11f174752745efaf2e4e780f171bf889824.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 239, + 481, + 273 + ], + "lines": [ + { + "bbox": [ + 132, + 239, + 481, + 273 + ], + "spans": [ + { + "bbox": [ + 132, + 239, + 481, + 273 + ], + "type": "text", + "content": "Table 2: Per-scene quantitative results from the realistic synthetic blender dataset [32], featuring scenes with complex geometries and non-Lambertian materials rendered using Blender's Cycles pathtracer." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 133, + 293, + 481, + 355 + ], + "blocks": [ + { + "bbox": [ + 133, + 293, + 481, + 355 + ], + "lines": [ + { + "bbox": [ + 133, + 293, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 133, + 293, + 481, + 355 + ], + "type": "image", + "image_path": "1d8b8d3ae3ddd5675c2650aeb5d70c079a873e25269e6360b6ff64cdbbf2eb60.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 363, + 480, + 386 + ], + "lines": [ + { + "bbox": [ + 132, + 363, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 480, + 386 + ], + "type": "text", + "content": "Fig. 9: Example renderings of mesh-based avatars from [29] trained using EdgeGrad, capturing complex inner mouth movements, including teeth and tongue dynamics." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 408, + 480, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 480, + 432 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 480, + 432 + ], + "type": "text", + "content": "ity we opted out automatic remeshing introduced by Nicolet et al. which would necessitate updating UV parametrization in our case." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 433, + 481, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 481, + 491 + ], + "type": "text", + "content": "Qualitative mesh reconstruction. Qualitatively, we compare the mesh reconstructions obtained on the public NeRF dataset [32] using our approach, a naive approach with no special edge handling, and nvdiffrast [20] in Fig. 8. All methods use the same shading model. Our approach shows improved detail in the recovered meshes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 510, + 375, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 375, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 375, + 523 + ], + "type": "text", + "content": "6 Application to Dynamic Head Scenes" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 641 + ], + "type": "text", + "content": "We show the importance of correct edge gradients on a head modeling application by fitting a personalized mesh-based avatar representation to facial performances captured in a multiview lightstage. modifying the avatar creation method from [29]. As the facial avatar representation, we use an encoder-decoder architecture to produce mesh vertices and texture-mapped appearance features [25], which are shaded using a pixel-wise MLP producing RGB appearance [29]. We use a coarse face tracking produced using traditional computer vision as initialization for inverse rendering. Importantly, this coarse face template has a small number of vertices (8k) and no geometry in the interior of the mouth." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "text", + "content": "Traditionally, fitting the interior of the mouth without strong priors is a complex task due to poor observability and intricately interacting geometry with" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rasterized Edge Gradients: Handling Discontinuities Differently" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 187, + 240 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 187, + 240 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 187, + 240 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 187, + 240 + ], + "type": "image", + "image_path": "1970cc83f0a77b518b57863fdc75e12dcd76c03d06f58238ff3249012a5def9d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 240, + 184, + 247 + ], + "lines": [ + { + "bbox": [ + 145, + 240, + 184, + 247 + ], + "spans": [ + { + "bbox": [ + 145, + 240, + 184, + 247 + ], + "type": "text", + "content": "(a) camera image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 188, + 114, + 240, + 240 + ], + "blocks": [ + { + "bbox": [ + 188, + 114, + 240, + 240 + ], + "lines": [ + { + "bbox": [ + 188, + 114, + 240, + 240 + ], + "spans": [ + { + "bbox": [ + 188, + 114, + 240, + 240 + ], + "type": "image", + "image_path": "3131f3014cc1620513b16c33ac01caacbfaf04db16b11f7001524f0f1f005299.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 202, + 240, + 229, + 247 + ], + "lines": [ + { + "bbox": [ + 202, + 240, + 229, + 247 + ], + "spans": [ + { + "bbox": [ + 202, + 240, + 229, + 247 + ], + "type": "text", + "content": "(b) Nvdiffrast" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 240, + 114, + 291, + 240 + ], + "blocks": [ + { + "bbox": [ + 240, + 114, + 291, + 240 + ], + "lines": [ + { + "bbox": [ + 240, + 114, + 291, + 240 + ], + "spans": [ + { + "bbox": [ + 240, + 114, + 291, + 240 + ], + "type": "image", + "image_path": "71b44cfe6577bba41607c6fd59db4b53843b06b01255ea9b960ac0f6c4b138a6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 250, + 240, + 287, + 247 + ], + "lines": [ + { + "bbox": [ + 250, + 240, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 250, + 240, + 287, + 247 + ], + "type": "text", + "content": "(c) EdgeGrad (our)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 290, + 240, + 343, + 249 + ], + "lines": [ + { + "bbox": [ + 290, + 240, + 343, + 249 + ], + "spans": [ + { + "bbox": [ + 290, + 240, + 343, + 249 + ], + "type": "text", + "content": "(d) EdgeGrad (our) w/o segmentation supervision" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 292, + 114, + 344, + 240 + ], + "blocks": [ + { + "bbox": [ + 292, + 114, + 344, + 240 + ], + "lines": [ + { + "bbox": [ + 292, + 114, + 344, + 240 + ], + "spans": [ + { + "bbox": [ + 292, + 114, + 344, + 240 + ], + "type": "image", + "image_path": "930f8e889bfb2952980f7dd8dcd1dbff3d9afbc3b2906a565ec9ceb2e7129184.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 116, + 482, + 247 + ], + "lines": [ + { + "bbox": [ + 353, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 353, + 116, + 482, + 247 + ], + "type": "text", + "content": "Fig. 10: Qualitative evaluation and ablation on dynamic head scene. We compare: (c) our method, (a) ground truth images, (b) Nvdiffrast [20], and ablated version of our method without including segmentation mask. The avatar model is from [29] trained using EdgeGrad and Nvdiffrast. We highlight artifacts with white arrows." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 272, + 482, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 488 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 488 + ], + "type": "text", + "content": "constant contact. Prior work either relies on artists to create alignments, volumetric simulation to prevent self-intersections [13], or on very specific mouth interior capture methods to build priors [30,49]. Instead, we show that by using differentiable rasterization with correct edge gradients, we can achieve fully automatic geometric fitting of the complex geometry of the face and mouth interior including lips, teeth, and tongue using a mesh-based model of fixed topology. We modify [29] in two ways. First, we use the edge gradient rasterization method of Sec. 4. Second, we add supervision in the form of 2D segmentation masks with separate labels for the regions of the teeth, tongue, and lips, for details see Fig. 7. While we can achieve accurate per-frame geometry without the additional segmentation loss supervision, we find that segmentation losses help anchor the teeth and tongue geometry to a consistent region of the UV map. Fig. 9 shows the quality of geometric fitting achievable with this inverse rendering method. Note that we accurately reconstruct the complex contact geometry of the tongue sliding across teeth, teeth-to-teeth contact, as well as the tongue pushing against the lips and cheeks. Fig. 10 demonstrates reconstruction quality of complex inner mouth details, which are challenging due to frequent self-intersections. We also include results without 2D segmentation mask supervision." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 507, + 323, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 507, + 323, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 507, + 323, + 521 + ], + "type": "text", + "content": "7 Conclusions and Limitations" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "We have presented a fast method for rasterization-based differentiable rendering that has accurate gradients across visibility discontinuities such as triangle edges, occlusions and geometry intersections. Our results demonstrate enhanced mesh reconstruction fidelity across diverse inverse rendering scenarios, comparable to implicit and volumetric methods. Our method also unlocks new applications, accurate mouth interior fitting in dynamic facial performance capture, through fully automatic inverse rendering. Regarding limitations, our method inherits the constraints of rasterization, including the inability to manage transparency, lack of precise antialIASing without multisampling, and the absence of physically based rendering and global illumination. It may also inaccurately interpret adjacent primitives in finely tessellated meshes, leading to sub-pixel triangles." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "text", + "content": "S. Pidhorskyi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 139, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 139, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 139, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 139, + 481, + 172 + ], + "type": "text", + "content": "1. Baek, S., Kim, K.I., Kim, T.K.: Pushing the envelope for rgb-based dense 3d hand pose estimation via neural rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1067-1076 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 481, + 205 + ], + "type": "text", + "content": "2. Bangaru, S.P., Gharbi, M., Luan, F., Li, T.M., Sunkavalli, K., Hasan, M., Bi, S., Xu, Z., Bernstein, G., Durand, F.: Differentiable rendering of neural sdfs through reparameterization. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 206, + 481, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 481, + 227 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 481, + 227 + ], + "type": "text", + "content": "3. Bangaru, S.P., Li, T.M., Durand, F.: Unbiased warped-area sampling for differentiable rendering. ACM Transactions on Graphics (TOG) 39(6), 1-18 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 228, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 481, + 270 + ], + "type": "text", + "content": "4. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 271, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 481, + 315 + ], + "type": "text", + "content": "5. Beker, D., Kato, H., Morariu, M.A., Ando, T., Matsuoka, T., Kehl, W., Gaidon, A.: Monocular differentiable rendering for self-supervised 3d object detection. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16. pp. 514-529. Springer (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 316, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 481, + 369 + ], + "type": "text", + "content": "6. Bogo, F., Kanazawa, A., Lassner, C., Gehler, P., Romero, J., Black, M.J.: Keep it spl! Automatic estimation of 3d human pose and shape from a single image. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14. pp. 561-578. Springer (2016)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 370, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 481, + 413 + ], + "type": "text", + "content": "7. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 414, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 414, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 414, + 481, + 446 + ], + "type": "text", + "content": "8. Chen, W., Ling, H., Gao, J., Smith, E., Lehtinen, J., Jacobson, A., Fidler, S.: Learning to predict 3d objects with an interpolation-based differentiable renderer. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 447, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 481, + 490 + ], + "type": "text", + "content": "9. Chen, W., Litalien, J., Gao, J., Wang, Z., Fuji Tsang, C., Khamis, S., Litany, O., Fidler, S.: Dib-r++: learning to predict lighting and material with a hybrid differentiable renderer. Advances in Neural Information Processing Systems 34, 22834-22848 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 491, + 481, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 481, + 534 + ], + "type": "text", + "content": "0. Choy, C.B., Xu, D., Gwak, J., Chen, K., Savarese, S.: 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14. pp. 628-644. Springer (2016)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 534, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 481, + 567 + ], + "type": "text", + "content": "1. Ge, L., Ren, Z., Li, Y., Xue, Z., Wang, Y., Cai, J., Yuan, J.: 3d hand shape and pose estimation from a single rgb image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10833-10842 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 567, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 481, + 610 + ], + "type": "text", + "content": "2. Henzler, P., Reizenstein, J., Labatut, P., Shapovalov, R., Ritschel, T., Vedaldi, A., Novotny, D.: Unsupervised learning of 3d object categories from videos in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4700-4709 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "type": "text", + "content": "3. Ichim, A.E., Kadleček, P., Kavan, L., Pauly, M.: Phace: Physics-based face modeling and animation. ACM Trans. Graph. 36(4) (jul 2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "type": "text", + "content": "4. Jakob, W., Speierer, S., Roussel, N., Nimier-David, M., Vicini, D., Zeltner, T., Nicolet, B., Crespo, M., Leroy, V., Zhang, Z.: Mitsuba 3 renderer (2022), https://mitsuba-renderer.org" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rasterized Edge Gradients: Handling Discontinuities Differently" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "15. Jakob, W., Speierer, S., Roussel, N., Vicini, D.: Dr.jit: A just-in-time compiler for differentiable rendering. Transactions on Graphics (Proceedings of SIGGRAPH) 41(4) (Jul 2022). https://doi.org/10.1145/3528223.3530099" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "text", + "content": "16. Jiang, Y., Ji, D., Han, Z., Zwicker, M.: Sdiff: Differentiable rendering of signed distance fields for 3d shape optimization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1251-1261 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 480, + 213 + ], + "type": "text", + "content": "17. Kato, H., Ushiku, Y., Harada, T.: Neural 3d mesh renderer. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3907-3916 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 480, + 246 + ], + "type": "text", + "content": "18. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG) 42(4), 1-14 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 246, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 246, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 133, + 246, + 480, + 289 + ], + "type": "text", + "content": "19. de La Gorce, M., Fleet, D.J., Paragios, N.: Model-based 3d hand pose estimation from monocular video. IEEE Trans. Pattern Anal. Mach. Intell. 33(9), 1793-1805 (Sep 2011). https://doi.org/10.1109/TPAMI.2011.33, http://dx.doi.org/10.1109/TPAMI.2011.33" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 289, + 480, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 322 + ], + "type": "text", + "content": "20. Laine, S., Hellsten, J., Karras, T., Seol, Y., Lehtinen, J., Aila, T.: Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG) 39(6), 1-14 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 322, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 480, + 354 + ], + "type": "text", + "content": "21. Li, T.M., Aittala, M., Durand, F., Lehtinen, J.: Differentiable monte carlo ray tracing through edge sampling. ACM Trans. Graph. (Proc. SIGGRAPH Asia) 37(6), 222:1-222:11 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 354, + 480, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 397 + ], + "type": "text", + "content": "22. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 300–309 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "type": "text", + "content": "23. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 418, + 480, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 451 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 451 + ], + "type": "text", + "content": "24. Liu, S., Li, T., Chen, W., Li, H.: Soft rasterizer: A differentiable renderer for image-based 3d reasoning. The IEEE International Conference on Computer Vision (ICCV) (Oct 2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 450, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 472 + ], + "type": "text", + "content": "25. Lombardi, S., Saragih, J., Simon, T., Sheikh, Y.: Deep appearance models for face rendering. ACM Transactions on Graphics (TOG) 37(4), 1-13 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "type": "text", + "content": "26. Lombardi, S., Simon, T., Saragih, J., Schwartz, G., Lehrmann, A., Sheikh, Y.: Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "text", + "content": "27. Loper, M.M., Black, M.J.: Opendra: An approximate differentiable renderer. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VII 13. pp. 154-169. Springer (2014)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "text", + "content": "28. Loubet, G., Holzschuch, N., Jakob, W.: Reparameterizing discontinuous integrands for differentiable rendering. Transactions on Graphics (Proceedings of SIGGRAPH Asia) 38(6) (Dec 2019). https://doi.org/10.1145/3355089.3356510" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "text", + "content": "29. Ma, S., Simon, T., Saragih, J.M., Wang, D., Li, Y., la Torre, F.D., Sheikh, Y.: Pixel codec avatars. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 64-73 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "text", + "content": "30. Medina, S., Tomé, D., Stoll, C., Tiede, M., Munhall, K., Hauptmann, A., Matthews, I.: Speech driven tongue animation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE/CVF (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 633, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 666 + ], + "type": "text", + "content": "31. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "text", + "content": "S. Pidhorskyi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "32. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 171 + ], + "type": "text", + "content": "33. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 171, + 482, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 171, + 482, + 191 + ], + "spans": [ + { + "bbox": [ + 130, + 171, + 482, + 191 + ], + "type": "text", + "content": "34. Nicolet, B., Jacobson, A., Jakob, W.: Large steps in inverse rendering of geometry. ACM Transactions on Graphics (TOG) 40(6), 1-13 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 191, + 482, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 191, + 482, + 235 + ], + "spans": [ + { + "bbox": [ + 130, + 191, + 482, + 235 + ], + "type": "text", + "content": "35. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3504-3515 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 235, + 482, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 235, + 482, + 267 + ], + "spans": [ + { + "bbox": [ + 130, + 235, + 482, + 267 + ], + "type": "text", + "content": "36. Nimier-David, M., Vicini, D., Zeltner, T., Jakob, W.: Mitsuba 2: A retargetable forward and inverse renderer. Transactions on Graphics (Proceedings of SIGGRAPH Asia) 38(6) (Dec 2019). https://doi.org/10.1145/3355089.3356498" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 267, + 482, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 482, + 311 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 482, + 311 + ], + "type": "text", + "content": "37. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 311, + 482, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 482, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 482, + 343 + ], + "type": "text", + "content": "38. Pavlakos, G., Zhu, L., Zhou, X., Daniilidis, K.: Learning to estimate 3d human pose and shape from a single color image. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 459-468 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 343, + 482, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 364 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 364 + ], + "type": "text", + "content": "39. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 364, + 482, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 482, + 407 + ], + "type": "text", + "content": "40. Rhodin, H., Robertini, N., Richardt, C., Seidel, H.P., Theobalt, C.: A versatile scene model with differentiable visibility applied to generative pose estimation. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 765-773 (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 407, + 482, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 450 + ], + "type": "text", + "content": "41. Roessle, B., Barron, J.T., Mildenhall, B., Srinivasan, P.P., Nießner, M.: Dense depth priors for neural radiance fields from sparse input views. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12892-12901 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 450, + 482, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 483 + ], + "type": "text", + "content": "42. Roveri, R., Öztireli, A.C., Pandele, I., Gross, M.: Pointpronets: Consolidation of point clouds with convolutional neural networks. In: Computer Graphics Forum. vol. 37, pp. 87-99. Wiley Online Library (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 483, + 482, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 482, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 482, + 504 + ], + "type": "text", + "content": "43. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 504, + 482, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 504, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 504, + 482, + 536 + ], + "type": "text", + "content": "44. Tsalicoglou, C., Manhardt, F., Tonioni, A., Niemeyer, M., Tombari, F.: Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 536, + 482, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 482, + 568 + ], + "type": "text", + "content": "45. Tulsiani, S., Zhou, T., Efros, A.A., Malik, J.: Multi-view supervision for single-view reconstruction via differentiable ray consistency. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2626-2634 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 568, + 482, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 590 + ], + "type": "text", + "content": "46. Vicini, D., Speierer, S., Jakob, W.: Differentiable signed distance function rendering. ACM Transactions on Graphics (TOG) 41(4), 1-18 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 590, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 590, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 590, + 482, + 632 + ], + "type": "text", + "content": "47. Wang, J., Sun, K., Cheng, T., Jiang, B., Deng, C., Zhao, Y., Liu, D., Mu, Y., Tan, M., Wang, X., et al.: Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence 43(10), 3349-3364 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "text", + "content": "48. Wiles, O., Gkioxari, G., Szeliski, R., Johnson, J.: Synsin: End-to-end view synthesis from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7467-7477 (2020)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rasterized Edge Gradients: Handling Discontinuities Differently" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 314 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "49. Wu, C., Bradley, D., Garrido, P., Zollhöfer, M., Theobalt, C., Gross, M., Beeler, T.: Model-based teeth reconstruction. ACM Trans. Graph. 35(6) (dec 2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "type": "text", + "content": "50. Yan, X., Yang, J., Yumer, E., Guo, Y., Lee, H.: Perspective transformer nets: Learning single-view 3d object reconstruction without 3d supervision. Advances in neural information processing systems 29 (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "text", + "content": "51. Yang, G., Cui, Y., Belongie, S., Hariharan, B.: Learning single-view 3d reconstruction with limited pose supervision. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 86-101 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 225 + ], + "type": "text", + "content": "52. Yu, A., Fridovich-Keil, S., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "text", + "content": "53. Zhang, C., Wu, L., Zheng, C., Gkioulekas, I., Ramamoorthi, R., Zhao, S.: A differential theory of radiative transfer. ACM Trans. Graph. 38(6), 227:1-227:16 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "54. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 271, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 314 + ], + "type": "text", + "content": "55. Zhao, S., Jakob, W., Li, T.M.: Physics-based differentiable rendering: From theory to implementation. In: ACM SIGGRAPH 2020 Courses. SIGGRAPH '20, Association for Computing Machinery, New York, NY, USA (2020). https://doi.org/10.1145/3388769.3407454, https://doi.org/10.1145/3388769.3407454" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "text", + "content": "S. Pidhorskyi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Rate-Distortion-Cognition Controllable Versatile Neural Image Compression/d4e52d64-79ee-43c9-8953-86dbf5d724e6_content_list.json b/2024/Rate-Distortion-Cognition Controllable Versatile Neural Image Compression/d4e52d64-79ee-43c9-8953-86dbf5d724e6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..57959bf81dcca33f7f186706b276e7f151beab11 --- /dev/null +++ b/2024/Rate-Distortion-Cognition Controllable Versatile Neural Image Compression/d4e52d64-79ee-43c9-8953-86dbf5d724e6_content_list.json @@ -0,0 +1,2035 @@ +[ + { + "type": "text", + "text": "Rate-Distortion-Cognition Controllable Versatile Neural Image Compression", + "text_level": 1, + "bbox": [ + 277, + 140, + 727, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinming Liu $^{1,2}$ , Ruoyu Feng $^{3}$ , Yunpeng Qi $^{3}$ , Qiuyu Chen $^{2}$ , Zhibo Chen $^{3}$ , Wenjun Zeng $^{2}$ , Xin Jin $^{2\\dagger}$", + "bbox": [ + 238, + 210, + 764, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shanghai Jiao Tong University \n $^{2}$ Ningbo Institute of Digital Twin, Eastern Institute of Technology, Ningbo, China \n $^{3}$ University of Science and Technology of China \njmliu206@sjtu.edu.cn jinxin@eitech.edu.cn", + "bbox": [ + 222, + 253, + 779, + 313 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Recently, the field of Image Coding for Machines (ICM) has garnered heightened interest and significant advances thanks to the rapid progress of learning-based techniques for image compression and analysis. Previous studies often require training separate codes to support various bitrate levels, machine tasks, and networks, thus lacking both flexibility and practicality. To address these challenges, we propose a rate-distortion-cognition controllable versatile image compression, which method allows the users to adjust the bitrate (i.e., Rate), image reconstruction quality (i.e., Distortion), and machine task accuracy (i.e., Cognition) with a single neural model, achieving ultra-controllability. Specifically, we first introduce a cognition-oriented loss in the primary compression branch to train a codec for diverse machine tasks. This branch attains variable bitrate by regulating quantization degree through the latent code channels. To further enhance the quality of the reconstructed images, we employ an auxiliary branch to supplement residual information with a scalable bitstream. Ultimately, two branches use a $\\beta x + (1 - \\beta)y$ interpolation strategy to achieve a balanced cognition-distortion trade-off. Extensive experiments demonstrate that our method yields satisfactory ICM performance and flexible Rate-Distortion-Cognition controlling.", + "bbox": [ + 259, + 327, + 743, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Neural Image Compression $\\cdot$ Image Compression for Machine $\\cdot$ Variable-bitrate Compression $\\cdot$ Cognition-distortion Trade-off", + "bbox": [ + 259, + 618, + 740, + 646 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 656, + 375, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Machine vision tasks have become increasingly integral in numerous daily applications, including intelligent traffic systems [27], autonomous driving [32], and facial recognition [62]. As a result, there is a burgeoning consumption of images by machines, which poses significant challenges in the transmission and storage of such extensive data volumes. How to efficiently compress these data while promising machine task accuracy urgently needs to be studied.", + "bbox": [ + 212, + 680, + 787, + 770 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditional image compression standards, such as JPEG [64], HEVC [60], VVC [9], have been widely applied in life. Meanwhile, deep learning-based CODECs have rapidly developed recently, achieving superior rate-distortion performance [17,", + "bbox": [ + 212, + 771, + 800, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 230, + 825, + 397, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/de05175552ea524d0413990280a9a4f964c7d763467e644ab9b13fbaaddb66ad.jpg", + "image_caption": [ + "Fig. 1: Motivation: (a) Traditional image compression tasks require training $K \\times M \\times T$ codecs for $K$ bitrate points, $M$ machine vision tasks, and $T$ cognition-distortion trade-offs. (b) Our method enables controllable rate-distortion-cognition with a single codec." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 542, + 238 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5ec7eced2ac4996a40b3c7490baeaf0f93779d37b714d41f93ce177cfe4626bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 557, + 157, + 785, + 238 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[40, 45, 48]. However, despite success in bit saving and visual fidelity, these methods hardly generalize well to machine tasks w.r.t cognition performance.", + "bbox": [ + 212, + 309, + 782, + 339 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "That's because traditional image compression CODECs are typically optimized to align with human visual perception. They employ distortion metrics, such as PSNR and MS-SSIM [45], as well as realism metrics, including KID and FID [1], to assess the loss incurred by compression procedure. Hence, utilizing these CODEcs directly for machine vision tasks may yield suboptimal outcomes [24, 47].", + "bbox": [ + 212, + 345, + 784, + 421 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this issue, the field of image coding for machines (ICM) has gradually come into focus [24,31,36,41,61]. For example, some works improves the performance of downstream tasks by jointly optimizing the codec and the task network [49]. Choi et al. [18] adopt a scalable method to decouple the features required for downstream machine vision tasks and reconstruction tasks, thereby reducing the required bitrate. Chen et al. [14] fine-tune the codec for different tasks through prompts to achieve better performance. Feng et al. [25] employ a contrastive learning approach to learn an omni-feature to improve the performance on various downstream tasks. However, these methods all remain the following challenges: 1). need to train multiple independent networks to adapt to different bit rates, thus the training cost and memory requirement increase with the desired rate-distortion points proportionally. 2). hard to accommodate diverse machine tasks/networks, bringing extra fine-tuning overhead and bad generalization. 3). fail to meet the practical demand for adjustable distortion-cognition trade-off (decompression fidelity vs. task performance) within a single framework. In summary, as illustrated in Fig. 1(a), previous image compression methods require training different codecs to support various rate-distortion levels and machine tasks, posing a large overhead in practical scene deployment.", + "bbox": [ + 212, + 426, + 787, + 699 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Some variable-bitrate (VBR) image compression methods attempt to achieve compression at different bitrates through a single model [20, 59]. For example, Song et al. [59] use a quality map to indicate compression bitrates, achieving compression at various rates. Cui et al. [20] achieve VBR by scaling different channels of the latent code through a gain unit. However, to our best knowledge, there are currently no VBR methods specifically designed for machine tasks. Most VBR methods are optimized for human vision, and can not be directly applied to machine vision tasks due to the assessment gap.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Jinming Liu et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f1886140281f7a477236101df4633be56c5fbfbffc040d474d770bace4a5bff8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 143, + 356, + 217 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/21dc0a3ed2c0bf008c2f346f3af059845efc27ce5d57877849ae49a654eccefb.jpg", + "image_caption": [ + "(a)", + "Fig. 2: The visualization and spectrum of (a) original image, (b) distortion-oriented compressed image $\\hat{\\pmb{x}}_2$ , (c) cognition-oriented compressed image $\\hat{\\pmb{x}}_1$ and, (d) the residual $|\\hat{\\pmb{x}}_1 - \\hat{\\pmb{x}}_2|$ of (b) and (c)." + ], + "image_footnote": [], + "bbox": [ + 220, + 220, + 356, + 291 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1d81dc35dd0b2a38b7429f6fa8ffc1fba2ab25334cf74de10d912b89890564da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 145, + 500, + 217 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/410381c6ad787e40b23d3211390b8b39f3823075d7d5af2f8dd517f87a484ce3.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 361, + 220, + 500, + 291 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4cad8967b79e68d7f72b72d4a42086628d0e04323af4769f6e15eeed49ad4510.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 145, + 640, + 217 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e3cbd58f7fccb2a541a1d844a675a6006c0e045e50ff3d38fb0f7dfe7fc0546f.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 504, + 220, + 640, + 291 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/472fa1f49c83a6e0dcbb80aa97aac072fe4e59d73cd2670697b41e2f0a3bf0a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 145, + 782, + 217 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/663d44ac5a7c2bec7b70ae9721da776ef3062adbbd1e60d2f48f125b7f2ee5a4.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 643, + 220, + 782, + 291 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Meanwhile, the distortion-realism trade-off of compression has been discussed recently [1,38]. For example, Agustsson et al. [1] achieve adjustable distortion-realism by attaching several layers of neural networks outside the decoder. However, both evaluations of distortion and realism align with human perception intuition, and many assessment aspects of them are largely overlapped. Unlike distortion or realism, cognition requires more valid information needed for downstream tasks. To the best of our knowledge, there are currently no studies discussing how to adjust the distortion-cognition trade-off in a single codec.", + "bbox": [ + 212, + 362, + 787, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In response to the aforementioned issues, we propose a rate-distortion-cognition controllable versatile neural image compression method. Specifically, we first introduce a cognition-oriented loss that combines contrastive constraints in the main compression branch to reconstruct a more general machine task-friendly image. The physical meaning behind this loss is that contrastive learning enables better visual representations for downstream intelligent analysis tasks [10,13,29]. We can see from Fig. 2 that, the normal distortion-oriented compressed images contain more low-frequency information, aligning with the intuition that the human eye is more sensitive to low-frequency components of image [69]. In contrast, the cognition-oriented compressed images contain more valid and critical high-frequency information, making them more suitable for machine vision analysis.", + "bbox": [ + 212, + 486, + 787, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then, we attain variable rate-cognition by regulating quantization degree based on a channel-controllable latent gain unit [20]. This module is end-to-end trainable and can be well compatible with the above contrastive objectives. Moreover, we analyze the difference in the information needed for cognition and distortion, and based on the results we further propose an auxiliary branch to supplement residual information with an extra scalable bitstream. We find only few bits cost could significantly reduce the distortion in compressed images. At last, by controlling interpolation between distortion-oriented and cognition-oriented compressed images, we can adjust the distortion-cognition trade-off in a single codec, thereby achieving controllable rate-distortion-cognition. This effectively enhances the practicality of ICM methods.", + "bbox": [ + 212, + 655, + 787, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The contributions of this paper can be summarized as follows:", + "bbox": [ + 238, + 824, + 687, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Rate-Distortion-Cognition Controllable Versatile Neural Image Compression", + "bbox": [ + 223, + 113, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a rate-distortion-cognition controllable versatile neural image compression model. To the best of our knowledge, we are the first to achieve controllable rate-distortion-cognition within a single codec.", + "- We introduce a cognition-oriented loss based on contrastive learning to generate task-friendlyly compressed images, which main compression process also attains variable-bitrate with a trainable gain unit.", + "- We explore the relationship between the information required for distortion and cognition, and design an auxiliary scalable bitstream to achieve a balanced cognition-distortion trade-off with a two-branch interpolation strategy." + ], + "bbox": [ + 223, + 146, + 784, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Experimental results demonstrate that our approach achieves satisfactory performance in both machine tasks and reconstruction, and outperforms both the latest ICM and learned image compression (LIC) methods in terms of adjustable rate-distortion-cognition practicality.", + "bbox": [ + 212, + 292, + 785, + 353 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 364, + 397, + 381 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Image Compression", + "text_level": 1, + "bbox": [ + 215, + 388, + 424, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image compression seeks to encode original image in a format that is both compact and retains high fidelity. Traditional image codecs, as delineated in numerous studies [9, 54, 60, 64, 66], typically incorporate modules such as intra prediction, discrete cosine transformation or wavelet transformation, quantization, and entropy coding. In contrast, learned-based codecs [6, 17, 45, 51-53] leverage neural networks to minimize distortion and bitrate. Compared to traditional approaches, data-driven learned methods have currently achieved better rate-distortion performance. Moreover, the rate-distortion trade-off in these models is governed by a Lagrange multiplier $\\lambda$ , which typically limits most methods to a single operation point on the rate-distortion curve for a fixed $\\lambda$ value. Recent advancements [19, 20, 37, 63, 70] have introduced various approaches to enable variable-rate compression within a single model framework. For example, Song et al. [59] innovatively propose spatial bit allocation in accordance with a quality map that matches the dimensions of the original image, offering a novel direction in the field of image compression. Cui et al. [20] utilize a scale matrix to scale the different channels of latent code to achieve variable bitrates. But these methods can not be directly applied to machine vision tasks due to the assessment gap.", + "bbox": [ + 212, + 410, + 787, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Image Compression for Machine", + "text_level": 1, + "bbox": [ + 215, + 681, + 529, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image Coding for Machines (ICM) focuses on compressing and transmitting source images to facilitate downstream intelligent tasks, including image classification [21,28,35], object detection [43,55-57], semantic segmentation [5,12,46, 68, 72]. Joint optimization of image compression models with these intelligent tasks [3,33,39,42,65] is a straightforward solution in this domain. Some feature compression methods are also proposed. [4, 15, 16, 22, 23, 25, 50, 58], enhancing both coding efficiency and computing offloading. For example, Feng et al. [25] try to compress an omni-feature for better machine vision task performance. Similarly, methods based on feature selection have been proposed [14, 26, 44, 45, 61],", + "bbox": [ + 212, + 704, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Jinming Liu et al.", + "bbox": [ + 271, + 114, + 392, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d2ce25096ce0780aed51e22792937e02d00c099c2097f6469254eb75bbf1cae1.jpg", + "image_caption": [ + "Fig. 3: The framework of our method. The framework of our method. Our pipeline is divided into a primary branch and an auxiliary branch. The primary branch generates the cognition-oriented image $\\hat{x}_1$ , while the auxiliary branch generates the distortion-oriented residual $r$ . We then use a $\\hat{x}_1 + (1 - \\beta)r$ interpolation strategy to achieve a balanced cognition-distortion trade-off. These two branches are trained separately in two stages: the primary branch uses cognition-oriented loss, and the auxiliary branch uses distortion-oriented loss." + ], + "image_footnote": [], + "bbox": [ + 218, + 147, + 782, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "achieving lower compression rates by compressing local regions instead of all features/images. For instance, Chen [14] et al. fine-tunes the codec via prompts, making the compressed features more suitable for downstream tasks. However, the methods mentioned above require training a separate model for compression at different bitrate points, and currently, there has been no discussion on variable-bitrate coding for machines. On the other hand, the trade-off between distortion and cognition (i.e., machine task performance) has not been publicly explored either.", + "bbox": [ + 212, + 438, + 787, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.3 Realism Image Compression", + "text_level": 1, + "bbox": [ + 214, + 571, + 496, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Traditional image coding goals often aim to achieve the lowest possible distortion. However, previous works [8] have demonstrated that \"low distortion\" is not synonymous with \"high perceptual quality.\" There exists a trade-off among Rate, Distortion, and Realism (human perception). Consequently, GAN networks [52] and Diffusion models [71] have been employed to enhance realism. An attempt [1] was made to use a conditional factor to control the distortion-realism trade-off. EGIC [38] achieving similar effects using two decoders. Based on [1], Iwai et al. [34] introduce a VBR mechanism to allow the control of rate-distortion-realism on a single codec. Distinct from, yet akin to the Rate-Distortion-Realism framework, this paper considers the Rate-Distortion-Cognition issue, targeting the topic of image coding for machine (ICM).", + "bbox": [ + 212, + 590, + 787, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 214, + 771, + 380, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Overview of the Proposed Compression Framework", + "text_level": 1, + "bbox": [ + 214, + 794, + 686, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our methodological approach is illustrated in Fig. 3. The training process is delineated into two stages, wherein the primary branch is dedicated to the gen-", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Rate-Distortion-Cognition Controllable Versatile Neural Image Compression", + "bbox": [ + 223, + 113, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f48430298ab8a04ac840001a0e8d5257c3cb2df1fcc80813fdb47e1cfde3097f.jpg", + "image_caption": [ + "Fig. 4: The training stage I. The parameters of the pretrained backbone remain frozen, with only the parameters of our codec being updated." + ], + "image_footnote": [], + "bbox": [ + 308, + 143, + 694, + 198 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "eration of cognition-oriented compressed images in stage I, while the auxiliary branch focuses on the distortion-oriented compressed images in stage II.", + "bbox": [ + 212, + 255, + 782, + 286 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in stage I of Fig. 4, to enhance the performance of the compressed image $\\hat{\\pmb{x}}_1$ on downstream machine vision tasks, we employ a cognition-oriented loss instead of the conventional distortion metric commonly used in image compression. This loss comprises two parts: 1). Inspired by studies [2, 25] indicating that semantic-level loss helps image cognition and downstream tasks in compression, the first component of the cognition-oriented loss designed herein encompasses a contrastive loss [29] that facilitates the extraction of robust and discriminative representations. 2). Another part of the cognition-oriented loss is MSE of local regions in the reconstructed image where the pixel intensity exceeds [0,1]. This part is used as a penalty term to align the range of normal image values. The stage I process can be described as follows:", + "bbox": [ + 212, + 291, + 784, + 457 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} \\stackrel {\\text {e n c o d e}} {\\longrightarrow} \\boldsymbol {y} \\stackrel {\\text {q u a n t i z a t i o n}} {\\longrightarrow} \\hat {\\boldsymbol {y}} \\stackrel {\\text {d e c o d e}} {\\longrightarrow} \\hat {\\boldsymbol {x}} _ {1} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 382, + 470, + 785, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\pmb{x}$ is the original image, $\\pmb{y}$ and $\\hat{\\pmb{y}}$ are the latent code and quantized latent code extracted from $\\pmb{x}$ . $\\hat{\\pmb{x}}_1$ is the cognition-oriented compressed image. Furthermore, based on the existing structure, we further attain variable rate-cognition by regulating quantization degree using a channel-controllable latent gain unit [20], this module is also trainable and compatible with the cognition-oriented loss.", + "bbox": [ + 212, + 513, + 782, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In stage II, we implement the reconstruction of distortion-oriented compressed images to meet the requirements of human vision. Freezing the codec of the primary branch, we explore to generate distortion-oriented compressed images $\\hat{\\pmb{x}}_2$ in this stage, which is defined as $\\hat{\\pmb{x}}_2 = \\hat{\\pmb{x}}_1 + \\pmb{r}$ : the residual $\\pmb{r}$ between the cognition-oriented and distortion-oriented compressed images is expected to extract from $\\hat{\\pmb{y}}$ at the decoder side. However, since $\\hat{\\pmb{y}}$ is optimized for cognition in stage I, it didn't contain enough distortion-oriented supplemented information. Therefore, we construct an auxiliary branch to enrich the residual $\\pmb{r}$ for the high-quality distortion-oriented compressed image $\\hat{\\pmb{x}}_2$ . As shown in Fig. 3, we first encode the quantization error $(\\pmb{y} - \\hat{\\pmb{y}})$ at the encoder side for transmission. After a scalable transform with variable bitstream, the decoded error $(\\widehat{\\pmb{y} - \\hat{\\pmb{y}}})$ is then added to the $\\hat{\\pmb{y}}$ to get $\\hat{\\pmb{y}}_2$ , and a lightweight residual reconstruction network is used to refine residual $\\pmb{r}$ . Interpolating $\\pmb{r}$ with $\\hat{\\pmb{x}}_1$ using a weight $\\beta$ gets $\\bar{\\pmb{x}} = \\hat{\\pmb{x}}_1 + (1 - \\beta)\\pmb{r}$ , which allows for different distortion-cognition performance. When $\\beta$ equals 0, $\\bar{\\pmb{x}} = \\hat{\\pmb{x}}_2$ the lowest distortion is achieved; conversely, when $\\beta$ equals 1, $\\bar{\\pmb{x}} = \\hat{\\pmb{x}}_1$ the highest cognition performance is attained.", + "bbox": [ + 212, + 595, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Jinming Liu et al.", + "bbox": [ + 271, + 114, + 392, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Primary Branch: Cognition-oriented Variable-bitrate Codec", + "text_level": 1, + "bbox": [ + 215, + 146, + 754, + 161 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A common use of learned image compression (LIC) model [17,45,48,53], can be formulated as the following equation:", + "bbox": [ + 214, + 174, + 782, + 203 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\text {E n c o d i n g}: \\boldsymbol {y} = g _ {a} (\\boldsymbol {x}; \\phi_ {g _ {a}}); \\quad \\hat {\\boldsymbol {y}} = Q (\\boldsymbol {y})\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 205, + 633, + 223 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\text {D e c o d i n g :} \\hat {\\boldsymbol {x}} = g _ {s} (\\hat {\\boldsymbol {y}}; \\phi_ {g _ {s}}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 226, + 785, + 243 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\text {H y p e r - p r i o r :} \\boldsymbol {z} = h _ {a} \\left(\\boldsymbol {y}; \\phi_ {h _ {a}}\\right); \\quad \\hat {\\boldsymbol {z}} = Q (\\boldsymbol {z}); \\quad \\boldsymbol {\\mu} _ {y}, \\boldsymbol {\\sigma} _ {y} = h _ {s} \\left(\\hat {\\boldsymbol {z}}; \\phi_ {h _ {s}}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 245, + 246, + 733, + 265 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\phi_{g_a},\\phi_{g_s},\\phi_{h_a}$ , and $\\phi_{h_s}$ are the parameters of the encoder network $g_{a}$ , the decoder network $g_{s}$ , the hyper-encoder network $h_a$ , and the hyper-decoder network $h_s$ , respectively. $Q$ represents the quantization which is approximated by a noise $\\mathcal{U}\\left(-\\frac{1}{2},\\frac{1}{2}\\right)$ during training, and is a rounding operator during inference. For inference, an image is first input to $g_{a}$ to extract a compact latent code $\\pmb{y}$ , and $\\pmb{y}$ is quantized to obtain $\\hat{\\pmb{y}}$ by quantization operator $Q$ . On the decoder side, $\\hat{\\pmb{y}}$ is input into $g_{s}$ for reconstruction to obtain the reconstructed image $\\hat{\\pmb{x}}$ . Additionally, the hyper path $(h_a$ and $h_s)$ is used to help estimate the probability distribution of $\\pmb{y}$ , which is assumed here to be a gaussian distribution with mean $\\mu_y$ and variance $\\sigma_y$ .", + "bbox": [ + 214, + 268, + 785, + 419 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As discussed before, learned image compression typically requires learning a new model for different bitrate points. Therefore, in this paper, we employ a variable bitrate approach to achieve compression at different bitrates through a single model. Inspired by [20], we scale $\\mathbf{y}$ and $\\mathbf{z}$ using $N$ gain unit $G_{y,n}$ and $G_{z,n}$ , which is a matrix containing $C$ factors, where $C$ is the channel dimension of $\\mathbf{y}$ , and $n \\in [1,N]$ . By scaling different channels, we can adjust the precision of the data we need to transmit, further achieving transmission at $N$ different bitrates. Furthermore, we can interpolate the scaling factors for other bitrate points from these $N$ gain units, allowing us to switch between different bitrates using a single model. Therefore, Equation 2 can be re-formulated as:", + "bbox": [ + 214, + 421, + 787, + 571 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\text {E n c o d i n g}: \\boldsymbol {y} = g _ {a} (\\boldsymbol {x}; \\phi_ {g _ {a}}); \\quad \\hat {\\boldsymbol {y}} = Q (\\boldsymbol {y} \\cdot \\boldsymbol {G} _ {y, n}) / \\boldsymbol {G} _ {y, n}\n$$\n", + "text_format": "latex", + "bbox": [ + 310, + 575, + 687, + 594 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\text {D e c o d i n g}: \\hat {\\boldsymbol {x}} = g _ {s} (\\hat {\\boldsymbol {y}}; \\phi_ {g _ {s}})\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 595, + 596, + 614 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\text {H y p e r - p r i o r :} z = h _ {a} \\left(\\boldsymbol {y} \\cdot \\boldsymbol {G} _ {y, n}; \\phi_ {h _ {a}}\\right); \\quad \\hat {\\boldsymbol {z}} = Q (\\boldsymbol {z} \\cdot \\boldsymbol {G} _ {z, n}) / \\boldsymbol {G} _ {z, n}; \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 272, + 609, + 784, + 633 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} _ {y}, \\sigma_ {y} = h _ {s} (\\hat {z}; \\phi_ {h _ {s}})\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 637, + 573, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For conventional distortion-oriented compression tasks, the loss can be represented as:", + "bbox": [ + 212, + 659, + 782, + 686 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\mathcal {R} (\\hat {\\boldsymbol {y}}) + \\mathcal {R} (\\hat {\\boldsymbol {z}}) + \\lambda_ {n} \\cdot \\mathcal {D} (\\boldsymbol {x}, \\hat {\\boldsymbol {x}}) \\\\ = \\mathbb {E} \\left[ - \\log_ {2} \\left(p _ {\\hat {\\boldsymbol {y}} | \\hat {\\boldsymbol {z}}} (\\hat {\\boldsymbol {y}} \\mid \\hat {\\boldsymbol {z}})\\right) \\right] + \\mathbb {E} \\left[ - \\log_ {2} \\left(p _ {\\hat {\\boldsymbol {z}} | \\psi} (\\hat {\\boldsymbol {z}} \\mid \\psi)\\right) \\right] + \\lambda_ {n} \\cdot \\mathcal {D} (\\boldsymbol {x}, \\hat {\\boldsymbol {x}}) \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 245, + 691, + 785, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where the parameter $\\lambda_{n}$ regulates the balance between rate and distortion, with varying $\\lambda_{n}$ values corresponding to different bit rates. $\\psi$ is a factorized density model used to encode quantized $\\hat{z}$ . During training, we randomly select one of the $N$ bitrate points for each forward pass. The distortion term, denoted by $\\mathcal{D}(\\pmb{x},\\hat{\\pmb{x}})$ , is measured using the Mean Squared Error (MSE) loss, GAN loss or MS-SSIM loss. Additionally, $\\mathcal{R}(\\hat{\\pmb{y}})$ and $\\mathcal{R}(\\hat{z})$ represent the bit rates for the latent representations $\\hat{\\pmb{y}}$ and $\\hat{\\pmb{z}}$ , respectively. To improve the performance of", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Rate-Distortion-Cognition Controllable Versatile Neural Image Compression", + "bbox": [ + 223, + 113, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the reconstructed images for various downstream tasks, here we use cognition-oriented loss as a substitute for the original MSE loss [17, 45, 48]. The first part of the cognition-oriented loss is a contrastive loss computed using a pretrained MoCo [29]. It is defined as follows:", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} = - \\log \\frac {\\exp \\left(q \\cdot k _ {+} / \\tau\\right)}{\\sum_ {i = 0} ^ {K} \\exp \\left(q \\cdot k _ {i} / \\tau\\right)} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 205, + 784, + 237 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $q$ denotes the query vector derived from the current image, $k_{+}$ is a positive key vector generated from a different augmentation of the same image, and $k_{i}$ are the negative key vectors from different images. The denominator aggregates the exponentiated similarities of the query with both the positive and K negative keys, normalized by a temperature parameter $\\tau$ , which adjusts the sharpness of the distribution. This loss function encourages the model to produce embeddings where positive pairs are closer in the feature space than any negative pair, thus learning robust and discriminative feature representations. However, unlike previous work with contrastive loss, which typically aims to learn a better backbone [29] or omni-features through contrastive learning [25], our objective here is to update the codec's parameters to enable the generation of cognition-oriented compressed images. Therefore, we impose constraints on the output values of the codec, to align the range of normal image values and avoid outliers. Here, we compute a local Mean Squared Error (MSE) Loss for values lying outside the [0, 1] range. It can be formulated as:", + "bbox": [ + 212, + 237, + 784, + 460 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {\\text {l o c a l}} (\\boldsymbol {x}, \\hat {\\boldsymbol {x}}) = \\frac {1}{M} \\sum_ {i = 1} ^ {N} \\left[ \\mathbb {I} \\left(\\hat {x} _ {i} < 0 \\vee \\hat {x} _ {i} > 1\\right) \\cdot \\left(\\hat {x} _ {i} - x _ {i}\\right) ^ {2} \\right] \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 310, + 460, + 784, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here, $\\mathbb{I}$ is an indicator function that takes the value of 1 when the predicted value $\\hat{x}_i$ falls outside the interval [0,1], and 0 otherwise. $x_{i}$ and $\\hat{x}_i$ represent the target and predicted values for the $i$ th pixel, respectively, while $M$ denotes the number of pixels. $\\mathcal{D}_{\\mathrm{local}}$ is used as a penalty term to constrain the output range.", + "bbox": [ + 212, + 500, + 784, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As illustrated in Fig. 5, the image distributions of both the original image and the distortion-oriented compressed image primarily fall within the range of 0 to 1. However, if we directly optimize using the contrastive loss, the generated representations would reside within a significantly broader interval (approximately [-4, 6]), which does not align with the inherent characteristics of images. Moreover, directly clipping the generated representations to the [0, 1] range during training before passing them to subsequent networks can lead to the issue of vanishing gradients, resulting in a distribution of representations as shown in Fig. 5(d). Clipping values outside the [0, 1] range during inference would lead to substantial information loss. Employing a local MSE loss as a penalty term can effectively mitigate this problem (see Fig. 5(e)).", + "bbox": [ + 212, + 560, + 785, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Therefore, the overall loss of stage I in the main branch is:", + "bbox": [ + 238, + 726, + 663, + 739 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {I} = \\mathcal {R} (\\hat {\\boldsymbol {y}}) + \\mathcal {R} (\\hat {\\boldsymbol {z}}) + \\lambda_ {n} \\cdot \\mathcal {C} + \\lambda_ {\\text {l o c a l}} \\cdot \\mathcal {D} _ {\\text {l o c a l}} (\\boldsymbol {x}, \\hat {\\boldsymbol {x}}) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 742, + 784, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where the parameter $\\lambda_{\\mathrm{local}}$ serves to modulate the magnitude of influence exerted by the penalty term, denoted as $\\mathcal{D}_{\\mathrm{local}}$ . It is worth noting that the primary function of $\\mathcal{D}_{\\mathrm{local}}$ loss within this context is to enforce constraints on the output range. Consequently, we empirically assign a relatively small fixed value of $10^{-5}$ to $\\lambda_{\\mathrm{local}}$ , acknowledging its role in merely constraining the output range without significantly impacting the primary loss components.", + "bbox": [ + 212, + 756, + 785, + 844 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Jinming Liu et al.", + "bbox": [ + 271, + 114, + 392, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c8b69bab070e5860cef170617247b407042de9f3735d8050d26793ca8bcf4d11.jpg", + "image_caption": [ + "(a)", + "Fig. 5: The histograms for pixel values of: (a) the original image. (b) distortion-oriented compressed image. (c) cognition-oriented compressed image without using local mse loss. (d) cognition-oriented compressed images processed with clipping to the range $(0, 1)$ during training. (e) cognition-oriented compressed images with using local mse loss." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 336, + 224 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b1b0b7df87134551e4e5d9ccd76c0061eba0e412e87acd10037a683c3837308b.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 339, + 143, + 449, + 224 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/2afaa8e768fc42554121d800167e9974c906672ec848cc7c44b51c2ff58c436b.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 452, + 143, + 562, + 224 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/855c5dc16ef4eed1873e5659061c147333487f6cb24483e4e05297ac366e8fa7.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 565, + 143, + 674, + 224 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/636fbb64012d196ee79e65678ce2ea5af79df9fa66a39f777d1f99d46622ab81.jpg", + "image_caption": [ + "(e)" + ], + "image_footnote": [], + "bbox": [ + 674, + 143, + 784, + 224 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/dbd95b25b87ba7638c0e7bb610eda9cb7e098dba0ae2a98dc3aad91ef9da3312.jpg", + "table_caption": [ + "Table 1: Parameter numbers of different Codec modules." + ], + "table_footnote": [], + "table_body": "
Base CodecScalable TransformResidual Reconstruction
Parameters(/M)20.390.191.20
", + "bbox": [ + 264, + 316, + 733, + 344 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.3 Auxiliary Branch: Distortion-oriented Residual Supplement", + "text_level": 1, + "bbox": [ + 214, + 359, + 751, + 376 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To further meet the requirements of human vision, we employ an auxiliary branch to supplement residual information with a scalable bitstream. As shown in Fig. 3, we construct a residual reconstruction network to generate the residual $\\boldsymbol{r}$ . By adding $\\boldsymbol{r}$ to $\\hat{\\boldsymbol{x}}_1$ , we reconstruct an distortion-oriented compressed image $\\hat{\\boldsymbol{x}}_2$ . The residual reconstruction network comprises multiple layers of transposed convolutions and point-wise convolutional layers. The former is to increase the size of the representation, ensuring spatial dimension alignment with the original decoder's representation. The latter is employed to reduce and align the channel dimensions, simultaneously decreasing the model parameters and computational complexity. Additionally, the intermediate layer outputs $f_1, f_2$ , and $f_3$ from the original decoder serve as prior information. They are fed into the residual reconstruction network through a point-wise convolution to help recovery.", + "bbox": [ + 212, + 386, + 784, + 566 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We attempt to input $\\hat{\\pmb{y}}$ to the residual reconstruction network directly to generate $\\pmb{r}$ . However, directly using $\\hat{\\pmb{y}}$ for reconstruction does not work, because $\\hat{\\pmb{y}}$ of the primary branch is optimized for cognitive perception rather than distortion minimization. To prove that, we analyze the quantization operation: as shown in Equation 3, this process involves the multiplication of $\\pmb{y}$ by $G_{y,n}$ followed by quantization. $G_{y,n}$ plays a crucial role in determining the precision of quantization across different channels and, i.e., the extent of quantization error. Given that $G_{y,n}$ is optimized for cognitive perception rather than distortion minimization in stage I, thus the reconstructed representations $\\hat{\\pmb{y}}$ cannot recover the detailed residual. The gradient visualization in Fig. 6(a)(b) reveals the disparities in channel contributions between cognition-oriented and distortion-oriented images. By backpropagating the mean gradients $\\frac{\\partial\\hat{\\pmb{x}}}{\\partial\\hat{\\pmb{y}}}$ from the generated image into $\\hat{\\pmb{y}}$ ’s channels, we observe a marked difference in how $\\hat{\\pmb{y}}$ ’s channels influence the generation of cognition-oriented versus distortion-oriented compressed images.", + "bbox": [ + 212, + 568, + 785, + 779 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To address this issue, we propose to transmit this missed information through a scalable bitstream. Specifically, by using the similar VBR method in Sec. 3.2, we first encode the quantization-induced error $(\\pmb{y} - \\hat{\\pmb{y}})$ into a bitstream via a transformation through three convolutional layers for transmission to the de", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Rate-Distortion-Cognition Controllable Versatile Neural Image Compression", + "bbox": [ + 223, + 113, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/63d200379813159f62c0e3ea3b0d220a23f6f4e31d136dfeb85f5c3ce3fac2bd.jpg", + "image_caption": [ + "Fig. 6: (a) The gradient for $\\hat{\\pmb{y}}$ from cognition-orient image $\\hat{\\pmb{x}}_1$ (b) The gradient for $\\hat{\\pmb{y}}$ from distortion-orient image $\\hat{\\pmb{x}}_2$ (c) The energy of cognition-orient $\\hat{\\pmb{y}}$ and distortion-orient $\\hat{\\pmb{y}}_2$ (Purple represents the overlapping part of the energy between $\\hat{\\pmb{y}}$ and $\\hat{\\pmb{y}}_2$ )." + ], + "image_footnote": [], + "bbox": [ + 236, + 143, + 390, + 233 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/cae51e7aaa09f80faece02c3ceda740f2d49544f330bf3e14365d0e99bba1e58.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 143, + 568, + 233 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/b4b36493df5b6e4f65a28d2653e956179afbf2273cf3cfa18b103abbc152454d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 584, + 159, + 774, + 233 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "coder side. An auxiliary gain unit is used to control the precision and bitrate of the bitstream. Then, we reconstruct bitstream through another three convolutional layers, and the output is added to $\\hat{\\pmb{y}}$ to produce $\\hat{\\pmb{y}}_2$ , which is fed into the residual reconstruction network to generate $\\pmb{r}$ . The scalable transform and the residual reconstruction network together form the auxiliary branch, which exclusively costs fewer parameters compared to the base codec (see Table 1).", + "bbox": [ + 212, + 287, + 782, + 377 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Moreover, we also analyze the differences in energy across various channels between $\\hat{\\pmb{y}}_2$ and the original $\\hat{\\pmb{y}}$ in Fig. 6(c). It is observable that the energy levels across most channels are similar, with only a few channels significantly increased. This suggests that we only need to transmit very few bits to supplement such residuals to enhance decompression quality, which will be further proved in the experiments. Therefore, the overall loss of stage II can be formulated as follows:", + "bbox": [ + 212, + 378, + 784, + 468 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {I I} = \\mathcal {R} (\\hat {\\boldsymbol {s}}) + \\lambda_ {m} \\cdot \\mathcal {D} (\\boldsymbol {x}, \\hat {\\boldsymbol {x}} _ {2}) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 472, + 784, + 489 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\mathcal{R}(\\hat{\\pmb{s}})$ represents the bitrate of the scalable bitstream, $\\mathcal{D}$ is defined as MSE loss, $\\lambda_{m}$ is used to control rate-distortion, $m\\in [1,N]$ , and $\\hat{\\pmb{x}}_2 = \\hat{\\pmb{x}}_1 + \\pmb{r}$ .", + "bbox": [ + 212, + 494, + 782, + 526 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "3.4 Rate-Distortion-Cognition Trade-off", + "text_level": 1, + "bbox": [ + 214, + 540, + 560, + 556 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Through the aforementioned procedures, we are able to achieve two distinct trade-offs: rate-distortion and rate-cognition. To achieve ultra-controllability, by employing interpolation, we can adjust the trade-off between distortion and cognition, thereby enabling the controllable rate-distortion-cognition. The equation is as follows:", + "bbox": [ + 212, + 558, + 782, + 632 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\boldsymbol {x}} = \\hat {\\boldsymbol {x}} _ {1} + (1 - \\beta) \\times \\hat {\\boldsymbol {r}} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 421, + 633, + 784, + 648 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\beta \\in [0,1]$ is used to adjust the balance between cognition and distortion. As $\\beta$ approaches 1, the distortion of $\\bar{\\pmb{x}}$ increases, and the cognition performance of $\\bar{\\pmb{x}}$ is enhanced. Conversely, as $\\beta$ approaches 0, the distortion decreases, and the cognition performance diminishes. When transferring to downstream tasks, $\\beta$ is set to 1, the codec weights are kept frozen, and we only use cognition-oriented compressed images to finetune the network for the downstream tasks.", + "bbox": [ + 212, + 646, + 784, + 736 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 756, + 375, + 773 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experimental Setting", + "text_level": 1, + "bbox": [ + 214, + 776, + 439, + 792 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "**Codec.** We set TIC [48] as our base codec. During training stage I, at each step, we randomly select a number from the set $[0.0625, 0.125, 0.25, 0.5, 1, 2]$ as $\\lambda_{n}$ . Correspondingly, the latent code $\\pmb{y}$ is multiplied by the associated gain unit to", + "bbox": [ + 212, + 794, + 782, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Jinming Liu et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "implement variable bit rates. For convenience, in the subsequent sections, we use $\\alpha \\in [0,1]$ to represent the selected point of the bitrate. When $\\alpha = 0$ , the bitrate is at its minimum, corresponding to the model with $\\lambda_{n} = 0.0625$ . Conversely, when $\\alpha = 1$ , the bitrate is at its maximum, corresponding to the model with $\\lambda_{n} = 2$ . For training stage II, we set $\\lambda_{m} \\in [0.00018, 0.00036, 0.00072, 0.001, 0.0015, 0.002]$ . To assess the rate-distortion performance of the reconstructed image, we employ the Peak Signal-to-Noise Ratio (PSNR) to evaluate the extent of distortion and use bits per pixel (bpp) to measure the compression ratio. Bpp is formulated as: as $\\frac{b}{h \\times w}$ , where $h, w$ are the height and width of the source image, $b$ denotes the total bits cost of the coded feature bit-stream.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Image Classification. we train and validate our model on ImageNet-1k datasets. During the evaluation, before inputting the images into the model, we resize and crop them to 224x224. Meanwhile, ResNet50 [30], serving as the task network, is used to assess the accuracy of classification.", + "bbox": [ + 212, + 301, + 792, + 362 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Object Detection. For the object detection task, we train and validate on the MS COCO2017 train/val dataset. Following the setting in [67], we employ the faster-rcnn-FPN network [57] for object detection. The image scale ranges from [640, 800] pixels during training and is set to 800 at inference by default.", + "bbox": [ + 212, + 366, + 787, + 428 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Semantic Segmentation. For this task, we employ DeepLab V3 [12], a network designed for segmentation tasks. The training and evaluation are conducted on the Cityscapes dataset. The Cityscapes dataset is specifically designed for urban street scenes. It includes 5,000 images with fine annotations. Cross entropy serves as the loss function during the training process. To assess effectiveness, we use the mean intersection over union (mIoU) as the metric.", + "bbox": [ + 212, + 431, + 787, + 523 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Baseline. For evaluating the effectiveness of our method on machine vision tasks, we have selected several different baselines for comparison with our approach. Initially, the traditional image standard VVC (VTM-12.0) [9] was utilized for testing. For learned image compression, we employ TIC [48] to assess its performance across three tasks. Regarding the ICM method, we chose to compare with recent works, namely TransTIC [14] and T-TF [11]. To evaluate the rate-distortion performance of our model, we conducted comparisons with methods such as VTM, TIC [48], Factorized [6] and MBT-Mean [53].", + "bbox": [ + 212, + 527, + 787, + 650 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Quantitative Comparison", + "text_level": 1, + "bbox": [ + 214, + 656, + 470, + 671 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As illustrated in Fig. 7 and 8, we compared our method against recent works. It is first noteworthy that our approach is controllable, by adjusting the values of $\\alpha$ and $\\beta$ , we can achieve the ultra-controllability within the blue area depicted in the figures using a single codec. The bitrate increases with an increase in $\\alpha$ , while the performance for machine vision tasks and image reconstruction tasks respectively increases and decreases with an increase in $\\beta$ . As demonstrated by points ABCD in the Fig. 7 and 8, at $(\\beta = 1, \\alpha = 1)$ (Point A), we achieve the highest bitrate and the optimal machine vision task performance at this bitrate; at $(\\beta = 0, \\alpha = 1)$ (Point B), we obtain the highest bitrate and the lowest distortion at this bitrate; at $(\\beta = 0, \\alpha = 0)$ (Point C), we achieve the lowest bitrate and the lowest distortion at this bitrate; at $(\\beta = 1, \\alpha = 0)$ (Point D), we", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Rate-Distortion-Cognition Controllable Versatile Neural Image Compression", + "bbox": [ + 223, + 113, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/be94c521781b3bcd2e068f4459b85c0f39af3881c638a0e92a01576c3b2e2c2f.jpg", + "image_caption": [ + "(a) Classification on ImageNet" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 405, + 256 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/47db714b07a6b29d70a24124a4a338f9192973c08220a86bce643ea74561cfc9.jpg", + "image_caption": [ + "(b) Segmentation on CityScapes" + ], + "image_footnote": [], + "bbox": [ + 406, + 143, + 594, + 256 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/37e9888803d7edf2c50c0c3103b6528d4df667027efe69729b1de56140826a0a.jpg", + "image_caption": [ + "(c) Detection on COCO" + ], + "image_footnote": [], + "bbox": [ + 596, + 143, + 784, + 256 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a4917e750a7e8f85e0eb6b09a83af5043ef703060a7821f30dc1e6140e004d67.jpg", + "image_caption": [ + "Fig. 7: The rate-machine vision task performance curve ( $\\alpha$ changes from 0 to 1 from left to right, and $\\beta$ changes from 0 to 1 from bottom to top).", + "(a) ImageNet", + "Fig. 8: The rate-PSNR performance curve ( $\\alpha$ changes from 0 to 1 from left to right, and $\\beta$ changes from 0 to 1 from top to bottom)." + ], + "image_footnote": [], + "bbox": [ + 217, + 301, + 408, + 415 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d096978952a45d7da0f12fc226af0c00f76fbbc6c2a46dfa54f1a5e926037cb6.jpg", + "image_caption": [ + "(b) CityScapes" + ], + "image_footnote": [], + "bbox": [ + 408, + 301, + 596, + 415 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5c2eb3b5a357b8a39aedb91241ff74c62babb7fc1cae27a37e1ed4b8b7d52d16.jpg", + "image_caption": [ + "(c) COCO" + ], + "image_footnote": [], + "bbox": [ + 596, + 301, + 784, + 415 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "achieve the lowest bitrate and the best machine vision task performance at this bitrate.", + "bbox": [ + 212, + 479, + 784, + 508 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "As shown in Fig. 7 and Table 2, we compared our method $(\\beta = 1)$ with recent approaches in three tasks: classification, segmentation, and detection. It is evident that, compared with the un-tuned learned image compression method TIC [48] and the traditional image compression standard VVC (VTM-12.0) [9], our method significantly improves machine vision tasks performance, especially at lower bitrates. This demonstrates our model's superior cognition capabilities. Furthermore, compared with some ICM methods, TransTIC [14], T-FT [11] + TIC [48], our approach not only achieves comparable or even superior performance but also provides controllability.", + "bbox": [ + 212, + 510, + 784, + 645 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Fig. 8 and Table 3 illustrate the performance of our method $(\\beta = 0)$ in the image reconstruction task. It is observed that compared with TIC, which is optimized for distortion, our method, being primarily optimized for cognition, leading to a reduction in PSNR. However, it still manages to achieve results comparable to the recent learned image compression methods, such as MBT-Mean [53]. Additionally, as demonstrated in Fig. 8(a), compared with the ICM method TransTIC, our method can achieve a higher upper bound for PSNR.", + "bbox": [ + 212, + 646, + 785, + 752 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 760, + 390, + 776 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In our study, we conducted ablation studies based on the base codec TIC, as shown in Table 4. For the measurements of BD-Acc [7] and BD-PSNR, we set $\\beta = 1$ and $\\beta = 0$ , respectively. We explored the application of the VBR method from [20] to TIC, which allows for varying bitrates with a single model. Directly", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Jinming Liu et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/ad9a8eac15455b1269454141f4e228118c194cbc9c6484be915c48a0a874fe37.jpg", + "table_caption": [ + "Table 2: Comparison of machine vision Table 3: Comparison of image reconstruction task performance." + ], + "table_footnote": [], + "table_body": "
Cls.Det.Seg.
BD-AccBD-APBD-mIoU|Cls.|Det.|Seg.
|BD-PSNR||BD-PSNR||BD-PSNR|
TIC [48]0.6908-0.4421-0.0208TIC [48]0.2352-0.88760.0315
TransTIC [14]8.92181.81370.1128MBT-Mean [53]-0.8492-2.1432-0.9654
T-FT TIC [11]8.34590.70910.1049Factorized [6]-1.5387-2.3248-3.4599
Ours (β=1)9.98761.82590.1133Ours (β=0)-0.4102-1.8293-1.0240
", + "bbox": [ + 243, + 184, + 772, + 273 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/3865c8f70b0aabebe7a23ee242de2f1c110a5a8f6802f9d1724df57240116bd9.jpg", + "table_caption": [ + "Table 4: BD-PSNR and BD-Acc on ImageNet with TIC as anchor. Codec Number accounts for required codecs across $K$ bitrates, $M$ downstream tasks, and $T$ distortion-cognition scenarios." + ], + "table_footnote": [], + "table_body": "
MethodsTIC [48]TIC + VBR [20]TIC + Cognition-oriented lossOurs (w/o scalable bitstream)Ours
BD-PSNR--0.1453-1.5421-1.4640-0.9104
BD-Acc--0.573911.125410.812410.6045
Codec NumbersK × M × TM × TK × T11
", + "bbox": [ + 225, + 323, + 774, + 382 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "transferring VBR methods to machine vision tasks does not yield satisfactory results. At the same time, for other distinct scenarios (different tasks/networks and cognition-distortion trade-off), codec retraining remains requisite. Using the proposed cognition-oriented loss can significantly improve the performance of machine vision tasks. However, for different bitrates and cognition-distortion trade-offs, many codecs are still required. In contrast, our method requires only one codec to be applicable across all scenarios. For our method, when we do not transmit an auxiliary scalable bitstream and directly use $\\hat{\\pmb{y}}$ for the training of residual reconstruction, there is a significant improvement in BD-Acc, but the distortion-oriented compressed images exhibit heavy distortion. As analyzed in Section 3.3 and shown in Fig. 9 and Table 4, we discovered that a minimal scalable bitstream can significantly reduce reconstruction distortion, with BD-PSNR improving by approximately $37.81\\%$ , and the loss in BD-Acc being only around $2\\%$ .", + "bbox": [ + 212, + 404, + 787, + 616 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Visualization", + "text_level": 1, + "bbox": [ + 215, + 635, + 370, + 648 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Fig. 10, we present a visualization of images generated with varying $\\beta$ values, along with their spectrum. It is observable that distortion-oriented compressed images (i.e., $\\beta = 0$ ) lose a significant amount of high-frequency information compared to the original image. However, as the $\\beta$ value increases, the high-frequency information in the images progressively increases. This indicates that cognition-oriented compressed images contain more high-frequency information, making them more conducive for analysis in downstream tasks. For instance, compared to distortion-oriented compressed images, cognition-oriented compressed images depict more detailed wrinkles on the ball held in the statue's hand. On the other hand, cognition-oriented compressed images, compared to both distortion-oriented compressed images and the original, possess a certain degree of color deviation. This suggests to some extent that color restoration is not crucial for some downstream tasks.", + "bbox": [ + 212, + 657, + 787, + 854 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Rate-Distortion-Cognition Controllable Versatile Neural Image Compression", + "bbox": [ + 223, + 113, + 732, + 130 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b9dc8066341995bd9830dd0513cd06347ba476a8924a2d63ae96706230d75e23.jpg", + "image_caption": [ + "(a) Rate-Acc." + ], + "image_footnote": [], + "bbox": [ + 267, + 143, + 472, + 234 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/e5ea83a98108f949b858e663ec3432a50df93416405e10986e772a2f30695468.jpg", + "image_caption": [ + "(b) Rate-PSNR." + ], + "image_footnote": [], + "bbox": [ + 526, + 143, + 728, + 234 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/c88ad9533cb57f15776fcc4e864175ad9d31a075ade933cef2d6096d538e47f0.jpg", + "image_caption": [ + "Fig. 9: The ablation studies.", + "Fig. 10: Visualization results and spectrum with different $\\beta$ values. (a) represents the original image, while (b) to (f) represent the compressed images with $\\beta$ values of 0, 0.25, 0.5, 0.75, and 1.0, respectively, (g) represents the magnification of the corresponding colored frame area." + ], + "image_footnote": [], + "bbox": [ + 218, + 275, + 784, + 415 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 489, + 359, + 506 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we propose a Rate-Distortion-Cognition Controllable Versatile Neural Image Compression model. We introduce a cognition-oriented loss to enhance the machine vision support, and achieve variable-rate by latent channel regulation. To further meet human eye perception, we build an auxiliary branch with a very small and scalable bitrate, decoding distortion-oriented compressed images. Finally, with an interpolation strategy to merge two branches, we also achieve distortion-cognition controlling, thereby realizing controllable rate-distortion-cognition in a single codec. Our experiments demonstrate that our method can achieve excellent performance in classification, segmentation, and detection tasks while also achieving ultra-controllability.", + "bbox": [ + 212, + 513, + 787, + 667 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 215, + 703, + 393, + 720 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work was supported in part by NSFC 62302246 and ZJNSFC under Grant LQ23F010008, and supported by High Performance Computing Center at Eastern Institute of Technology, Ningbo, and Ningbo Institute of Digital Twin.", + "bbox": [ + 212, + 734, + 785, + 781 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Jinming Liu et al.", + "bbox": [ + 271, + 114, + 392, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Agustsson, E., Minnen, D., Toderici, G., Mentzer, F.: Multi-realism image compression with a conditional generator. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22324-22333 (2023) 2, 3, 5", + "2. Agustsson, E., Tschannen, M., Mentzer, F., Timofte, R., Gool, L.V.: Generative adversarial networks for extreme learned image compression. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 221-231 (2019) 6", + "3. Akbari, M., Liang, J., Han, J.: Dsslic: Deep semantic segmentation-based layered image compression. In: ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 2042-2046. IEEE (2019) 4", + "4. Babenko, A., Slesarev, A., Chigorin, A., Lempitsky, V.: Neural codes for image retrieval. In: ECCV. pp. 584-599. Springer (2014) 4", + "5. Badrinarayanan, V., Kendall, A., Cipolla, R.: Segnet: A deep convolutional encoder-decoder architecture for image segmentation. TPAMI 39(12), 2481-2495 (2017) 4", + "6. Balle, J., Laparra, V., Simoncelli, E.P.: End-to-end optimized image compression. In: ICLR (2017) 4, 11, 13", + "7. Bjontegaard, G.: Calculation of average psnr differences between rd-curves. In: VCEG-M33 (2001) 12", + "8. Blau, Y., Michaeli, T.: Rethinking lossy compression: The rate-distortion-perception tradeoff. In: International Conference on Machine Learning. pp. 675-685. PMLR (2019) 5", + "9. Bross, B., Wang, Y.K., Ye, Y., Liu, S., Chen, J., Sullivan, G.J., Ohm, J.R.: Overview of the versatile video coding (vvc) standard and its applications. TCSVT (2021) 1, 4, 11, 12", + "0. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021) 3", + "1. Chamain, L.D., Racapé, F., Bégaint, J., Pushparaja, A., Feltman, S.: End-to-end optimized image compression for machines, a study. In: 2021 Data Compression Conference (DCC). pp. 163-172. IEEE (2021) 11, 12, 13", + "2. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. TPAMI 40(4), 834-848 (2017) 4, 11", + "3. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: ICML. pp. 1597-1607. PMLR (2020) 3", + "4. Chen, Y.H., Weng, Y.C., Kao, C.H., Chien, C., Chiu, W.C., Peng, W.H.: Transtic: Transferring transformer-based image compression from human perception to machine perception. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 23297-23307 (2023) 2, 4, 5, 11, 12, 13", + "5. Chen, Z., Fan, K., Wang, S., Duan, L.Y., Lin, W., Kot, A.: Lossy intermediate deep learning feature compression and evaluation. In: ACM MM. pp. 2414-2422 (2019) 4", + "6. Chen, Z., Fan, K., Wang, S., Duan, L., Lin, W., Kot, A.C.: Toward intelligent sensing: Intermediate deep feature compression. TIP 29, 2230-2243 (2019) 4" + ], + "bbox": [ + 225, + 175, + 785, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Rate-Distortion-Cognition Controllable Versatile Neural Image Compression", + "bbox": [ + 223, + 113, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Cheng, Z., Sun, H., Takeuchi, M., Katto, J.: Learned image compression with discretized gaussian mixture likelihoods and attention modules. In: CVPR. pp. 7939-7948 (2020) 1, 4, 7, 8", + "18. Choi, H., Bajic, I.V.: Scalable image coding for humans and machines. IEEE Transactions on Image Processing 31, 2739-2754 (2022) 2", + "19. Choi, Y., El-Khamy, M., Lee, J.: Variable rate deep image compression with a conditional autoencoder. In: ICCV. pp. 3146-3154 (2019) 4", + "20. Cui, Z., Wang, J., Bai, B., Guo, T., Feng, Y.: G-vae: A continuously variable rate deep image compression framework. arXiv preprint arXiv:2003.02012 (2020) 2, 3, 4, 6, 7, 12, 13", + "21. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2020) 4", + "22. Duan, L.Y., Chandrasekhar, V., Chen, J., Lin, J., Wang, Z., Huang, T., Girod, B., Gao, W.: Overview of thempeg-cdv standard. TIP 25(1), 179-194 (2015) 4", + "23. Duan, L.Y., Lou, Y., Bai, Y., Huang, T., Gao, W., Chandrasekhar, V., Lin, J., Wang, S., Kot, A.C.: Compact descriptors for video analysis: The emerging mpeg standard. IEEE MultiMedia 26(2), 44-54 (2018) 4", + "24. Duan, L., Liu, J., Yang, W., Huang, T., Gao, W.: Video coding for machines: A paradigm of collaborative compression and intelligent analytics. TIP 29, 8680-8695 (2020) 2", + "25. Feng, R., Jin, X., Guo, Z., Feng, R., Gao, Y., He, T., Zhang, Z., Sun, S., Chen, Z.: Image coding for machines with omnipotent feature learning. In: ECCV. pp. 510-528. Springer (2022) 2, 4, 6, 8", + "26. Feng, R., Liu, J., Jin, X., Pan, X., Sun, H., Chen, Z.: Prompt-icm: A unified framework towards image coding for machines with task-driven prompts. arXiv preprint arXiv:2305.02578 (2023) 4", + "27. Gao, W., Ma, S., Duan, L., Tian, Y., Xing, P., Wang, Y., Wang, S., Jia, H., Huang, T.: Digital retina: A way to make the city brain more efficient by visual coding. IEEE Transactions on Circuits and Systems for Video Technology 31(11), 4147-4161 (2021) 1", + "28. Han, K., Xiao, A., Wu, E., Guo, J., Xu, C., Wang, Y.: Transformer in transformer. NeurIPS 34, 15908-15919 (2021) 4", + "29. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: CVPR. pp. 9729-9738 (2020) 3, 6, 8", + "30. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016) 11", + "31. He, T., Sun, S., Guo, Z., Chen, Z.: Beyond coding: Detection-driven image compression with semantically structured bit-stream. In: 2019 Picture Coding Symposium (PCS). pp. 1-5. IEEE (2019) 2", + "32. Hu, Y., Yang, J., Chen, L., Li, K., Sima, C., Zhu, X., Chai, S., Du, S., Lin, T., Wang, W., et al.: Planning-oriented autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17853-17862 (2023) 1", + "33. Hu, Y., Yang, S., Yang, W., Duan, L.Y., Liu, J.: Towards coding for human and machine vision: A scalable image coding approach. In: 2020 IEEE International Conference on Multimedia and Expo (ICME). pp. 1-6. IEEE (2020) 4", + "34. Iwai, S., Miyazaki, T., Omachi, S.: Controlling rate, distortion, and realism: Towards a single comprehensive neural image compression model. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2900-2909 (2024) 5" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Jinming Liu et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. ECCV (2022) 4", + "36. Jin, X., Feng, R., Sun, S., Feng, R., He, T., Chen, Z.: Semantical video coding: Instill static-dynamic clues into structured bitstream for ai tasks. Journal of Visual Communication and Image Representation 93, 103816 (2023) 2", + "37. Johnston, N., Vincent, D., Minnen, D., Covell, M., Singh, S., Chinen, T., Hwang, S.J., Shor, J., Toderici, G.: Improved lossy image compression with priming and spatially adaptive bit rates for recurrent networks. In: CVPR. pp. 4385-4393 (2018) 4", + "38. Körber, N., Kromer, E., Siebert, A., Hauke, S., Mueller-Gritschneider, D.: Egic: Enhanced low-bit-rate generative image compression guided by semantic segmentation. arXiv preprint arXiv:2309.03244 (2023) 3, 5", + "39. Le, N., Zhang, H., Cricri, F., Ghaznavi-Youvalari, R., Rahtu, E.: Image coding for machines: An end-to-end learned approach. In: ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1590-1594. IEEE (2021) 4", + "40. Li, H., Li, S., Dai, W., Li, C., Zou, J., Xiong, H.: Frequency-aware transformer for learned image compression. arXiv preprint arXiv:2310.16387 (2023) 1", + "41. Li, H., Li, S., Ding, S., Dai, W., Cao, M., Li, C., Zou, J., Xiong, H.: Image compression for machine and human vision with spatial-frequency adaptation. In: ECCV. Springer (2024) 2", + "42. Li, X., Shi, J., Chen, Z.: Task-driven semantic coding via reinforcement learning. TIP (2021) 4", + "43. Li, Y., Mao, H., Girshick, R., He, K.: Exploring plain vision transformer backbones for object detection. ECCV (2022) 4", + "44. Liu, J., Sun, H., Katto, J.: Semantic segmentation in learned compressed domain. In: 2022 Picture Coding Symposium (PCS). pp. 181-185. IEEE (2022) 4", + "45. Liu, J., Sun, H., Katto, J.: Learned image compression with mixed transformer-cnn architectures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14388-14397 (2023) 1, 2, 4, 7, 8", + "46. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: CVPR. pp. 3431-3440 (2015) 4", + "47. Lu, G., Ge, X., Zhong, T., Geng, J., Hu, Q.: Preprocessing enhanced image compression for machine vision. arXiv preprint arXiv:2206.05650 (2022) 2", + "48. Lu, M., Guo, P., Shi, H., Cao, C., Ma, Z.: Transformer-based image compression. arXiv preprint arXiv:2111.06707 (2021) 1, 7, 8, 10, 11, 12, 13", + "49. Ma, H., Liu, D., Yan, N., Li, H., Wu, F.: End-to-end optimized versatile image compression with wavelet-like transform. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(3), 1247-1263 (2020) 2", + "50. Ma, S., Zhang, X., Wang, S., Zhang, X., Jia, C., Wang, S.: Joint feature and texture coding: Toward smart video representation via front-end intelligence. TCSVT 29(10), 3095-3105 (2018) 4", + "51. Mentzer, F., Agustsson, E., Tschannen, M., Timofte, R., Van Gool, L.: Conditional probability models for deep image compression. In: CVPR. pp. 4394-4402 (2018) 4", + "52. Mentzer, F., Toderici, G., Tschannen, M., Agustsson, E.: High-fidelity generative image compression. arXiv preprint arXiv:2006.09965 (2020) 4, 5", + "53. Minnen, D., Balle, J., Toderici, G.: Joint autoregressive and hierarchical priors for learned image compression. In: NeurIPS (2018) 4, 7, 11, 12, 13", + "54. Rabbani, M., Joshi, R.: An overview of the jpeg 2000 still image compression standard. Signal processing: Image communication 17(1), 3-48 (2002) 4" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Rate-Distortion-Cognition Controllable Versatile Neural Image Compression", + "bbox": [ + 223, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "55. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified, real-time object detection. In: CVPR. pp. 779-788 (2016) 4", + "56. Redmon, J., Farhadi, A.: Yolo9000: better, faster, stronger. In: CVPR. pp. 7263-7271 (2017) 4", + "57. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS 28, 91-99 (2015) 4, 11", + "58. Singh, S., Abu-El-Haija, S., Johnston, N., Balle, J., Shrivastava, A., Toderici, G.: End-to-end learning of compressible features. In: 2020 IEEE International Conference on Image Processing (ICIP). pp. 3349-3353. IEEE (2020) 4", + "59. Song, M., Choi, J., Han, B.: Variable-rate deep image compression through spatially-adaptive feature transform. In: ICCV. pp. 2380-2389 (2021) 2, 4", + "60. Sullivan, G.J., Ohm, J.R., Han, W.J., Wiegand, T.: Overview of the high efficiency video coding (hevc) standard. TCSVT 22(12), 1649-1668 (2012) 1, 4", + "61. Sun, S., He, T., Chen, Z.: Semantic structured image coding framework for multiple intelligent applications. TCSVT (2020) 2, 4", + "62. Terhörst, P., Ihlefeld, M., Huber, M., Damer, N., Kirchbuchner, F., Raja, K., Kuijper, A.: Qmagface: Simple and accurate quality-aware face recognition. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3484-3494 (2023) 1", + "63. Toderici, G., Vincent, D., Johnston, N., Jin Hwang, S., Minnen, D., Shor, J., Covell, M.: Full resolution image compression with recurrent neural networks. In: CVPR. pp. 5306-5314 (2017) 4", + "64. Wallace, G.K.: Thejpeg still picture compression standard. IEEE transactions on consumer electronics 38(1), xviii-xxxiv (1992) 1, 4", + "65. Wang, S., Wang, S., Yang, W., Zhang, X., Wang, S., Ma, S., Gao, W.: Towards analysis-friendly face representation with scalable feature and texture compression. TMM (2021) 4", + "66. Wiegand, T., Sullivan, G.J., Bjontegaard, G., Luthra, A.: Overview of the h. 264/avc video coding standard. TCSVT 13(7), 560-576 (2003) 4", + "67. Wu, Y., Kirillov, A., Massa, F., Lo, W.Y., Girshick, R.: Detector2. https://github.com/facebookresearch/detectron2 (2019) 11", + "68. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. NeurIPS 34, 12077-12090 (2021) 4", + "69. Yan, F.F., Hou, F., Lu, Z.L., Hu, X., Huang, C.B.: Efficient characterization and classification of contrast sensitivity functions in aging. Scientific reports 7(1), 5045 (2017) 3", + "70. Yang, F., Herranz, L., Van De Weijer, J., Guitián, J.A.I., López, A.M., Mozerov, M.G.: Variable rate deep image compression with modulated autoencoder. IEEE Signal Processing Letters 27, 331-335 (2020) 4", + "71. Yang, R., Mandt, S.: Lossy image compression with conditional diffusion models. arXiv preprint arXiv:2209.06950 (2022) 5", + "72. Zheng, S., Lu, J., Zhao, H., Zhu, X., Luo, Z., Wang, Y., Fu, Y., Feng, J., Xiang, T., Torr, P.H., et al.: Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In: CVPR. pp. 6881-6890 (2021) 4" + ], + "bbox": [ + 215, + 146, + 784, + 770 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Jinming Liu et al.", + "bbox": [ + 271, + 114, + 392, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Rate-Distortion-Cognition Controllable Versatile Neural Image Compression/d4e52d64-79ee-43c9-8953-86dbf5d724e6_model.json b/2024/Rate-Distortion-Cognition Controllable Versatile Neural Image Compression/d4e52d64-79ee-43c9-8953-86dbf5d724e6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..97dcc651820175bfe6e48575880a37c09b80995d --- /dev/null +++ b/2024/Rate-Distortion-Cognition Controllable Versatile Neural Image Compression/d4e52d64-79ee-43c9-8953-86dbf5d724e6_model.json @@ -0,0 +1,2986 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.278, + 0.141, + 0.728, + 0.187 + ], + "angle": 0, + "content": "Rate-Distortion-Cognition Controllable Versatile Neural Image Compression" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.212, + 0.765, + 0.245 + ], + "angle": 0, + "content": "Jinming Liu\\(^{1,2}\\), Ruoyu Feng\\(^{3}\\), Yunpeng Qi\\(^{3}\\), Qiuyu Chen\\(^{2}\\), Zhibo Chen\\(^{3}\\), Wenjun Zeng\\(^{2}\\), Xin Jin\\(^{2\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.254, + 0.78, + 0.314 + ], + "angle": 0, + "content": "\\(^{1}\\) Shanghai Jiao Tong University \n\\(^{2}\\) Ningbo Institute of Digital Twin, Eastern Institute of Technology, Ningbo, China \n\\(^{3}\\) University of Science and Technology of China \njmliu206@sjtu.edu.cn jinxin@eitech.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.328, + 0.744, + 0.606 + ], + "angle": 0, + "content": "Abstract. Recently, the field of Image Coding for Machines (ICM) has garnered heightened interest and significant advances thanks to the rapid progress of learning-based techniques for image compression and analysis. Previous studies often require training separate codes to support various bitrate levels, machine tasks, and networks, thus lacking both flexibility and practicality. To address these challenges, we propose a rate-distortion-cognition controllable versatile image compression, which method allows the users to adjust the bitrate (i.e., Rate), image reconstruction quality (i.e., Distortion), and machine task accuracy (i.e., Cognition) with a single neural model, achieving ultra-controllability. Specifically, we first introduce a cognition-oriented loss in the primary compression branch to train a codec for diverse machine tasks. This branch attains variable bitrate by regulating quantization degree through the latent code channels. To further enhance the quality of the reconstructed images, we employ an auxiliary branch to supplement residual information with a scalable bitstream. Ultimately, two branches use a \\( \\beta x + (1 - \\beta)y \\) interpolation strategy to achieve a balanced cognition-distortion trade-off. Extensive experiments demonstrate that our method yields satisfactory ICM performance and flexible Rate-Distortion-Cognition controlling." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.619, + 0.741, + 0.647 + ], + "angle": 0, + "content": "Keywords: Neural Image Compression \\(\\cdot\\) Image Compression for Machine \\(\\cdot\\) Variable-bitrate Compression \\(\\cdot\\) Cognition-distortion Trade-off" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.657, + 0.377, + 0.673 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.681, + 0.788, + 0.771 + ], + "angle": 0, + "content": "Machine vision tasks have become increasingly integral in numerous daily applications, including intelligent traffic systems [27], autonomous driving [32], and facial recognition [62]. As a result, there is a burgeoning consumption of images by machines, which poses significant challenges in the transmission and storage of such extensive data volumes. How to efficiently compress these data while promising machine task accuracy urgently needs to be studied." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.772, + 0.802, + 0.817 + ], + "angle": 0, + "content": "Traditional image compression standards, such as JPEG [64], HEVC [60], VVC [9], have been widely applied in life. Meanwhile, deep learning-based CODECs have rapidly developed recently, achieving superior rate-distortion performance [17," + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.826, + 0.398, + 0.841 + ], + "angle": 0, + "content": "† Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "Jinming Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.543, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.158, + 0.787, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.244, + 0.788, + 0.287 + ], + "angle": 0, + "content": "Fig. 1: Motivation: (a) Traditional image compression tasks require training \\( K \\times M \\times T \\) codecs for \\( K \\) bitrate points, \\( M \\) machine vision tasks, and \\( T \\) cognition-distortion trade-offs. (b) Our method enables controllable rate-distortion-cognition with a single codec." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.31, + 0.784, + 0.34 + ], + "angle": 0, + "content": "[40, 45, 48]. However, despite success in bit saving and visual fidelity, these methods hardly generalize well to machine tasks w.r.t cognition performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.346, + 0.785, + 0.422 + ], + "angle": 0, + "content": "That's because traditional image compression CODECs are typically optimized to align with human visual perception. They employ distortion metrics, such as PSNR and MS-SSIM [45], as well as realism metrics, including KID and FID [1], to assess the loss incurred by compression procedure. Hence, utilizing these CODEcs directly for machine vision tasks may yield suboptimal outcomes [24, 47]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.427, + 0.789, + 0.7 + ], + "angle": 0, + "content": "To address this issue, the field of image coding for machines (ICM) has gradually come into focus [24,31,36,41,61]. For example, some works improves the performance of downstream tasks by jointly optimizing the codec and the task network [49]. Choi et al. [18] adopt a scalable method to decouple the features required for downstream machine vision tasks and reconstruction tasks, thereby reducing the required bitrate. Chen et al. [14] fine-tune the codec for different tasks through prompts to achieve better performance. Feng et al. [25] employ a contrastive learning approach to learn an omni-feature to improve the performance on various downstream tasks. However, these methods all remain the following challenges: 1). need to train multiple independent networks to adapt to different bit rates, thus the training cost and memory requirement increase with the desired rate-distortion points proportionally. 2). hard to accommodate diverse machine tasks/networks, bringing extra fine-tuning overhead and bad generalization. 3). fail to meet the practical demand for adjustable distortion-cognition trade-off (decompression fidelity vs. task performance) within a single framework. In summary, as illustrated in Fig. 1(a), previous image compression methods require training different codecs to support various rate-distortion levels and machine tasks, posing a large overhead in practical scene deployment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Some variable-bitrate (VBR) image compression methods attempt to achieve compression at different bitrates through a single model [20, 59]. For example, Song et al. [59] use a quality map to indicate compression bitrates, achieving compression at various rates. Cui et al. [20] achieve VBR by scaling different channels of the latent code through a gain unit. However, to our best knowledge, there are currently no VBR methods specifically designed for machine tasks. Most VBR methods are optimized for human vision, and can not be directly applied to machine vision tasks due to the assessment gap." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.225, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Rate-Distortion-Cognition Controllable Versatile Neural Image Compression" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.145, + 0.357, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.221, + 0.357, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.282, + 0.292, + 0.303, + 0.304 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.146, + 0.501, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.221, + 0.501, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.292, + 0.447, + 0.304 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.146, + 0.642, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.221, + 0.642, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.292, + 0.586, + 0.304 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.146, + 0.783, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.221, + 0.783, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.292, + 0.725, + 0.304 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.306, + 0.787, + 0.349 + ], + "angle": 0, + "content": "Fig. 2: The visualization and spectrum of (a) original image, (b) distortion-oriented compressed image \\(\\hat{\\pmb{x}}_2\\), (c) cognition-oriented compressed image \\(\\hat{\\pmb{x}}_1\\) and, (d) the residual \\(|\\hat{\\pmb{x}}_1 - \\hat{\\pmb{x}}_2|\\) of (b) and (c)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.363, + 0.788, + 0.484 + ], + "angle": 0, + "content": "Meanwhile, the distortion-realism trade-off of compression has been discussed recently [1,38]. For example, Agustsson et al. [1] achieve adjustable distortion-realism by attaching several layers of neural networks outside the decoder. However, both evaluations of distortion and realism align with human perception intuition, and many assessment aspects of them are largely overlapped. Unlike distortion or realism, cognition requires more valid information needed for downstream tasks. To the best of our knowledge, there are currently no studies discussing how to adjust the distortion-cognition trade-off in a single codec." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.789, + 0.653 + ], + "angle": 0, + "content": "In response to the aforementioned issues, we propose a rate-distortion-cognition controllable versatile neural image compression method. Specifically, we first introduce a cognition-oriented loss that combines contrastive constraints in the main compression branch to reconstruct a more general machine task-friendly image. The physical meaning behind this loss is that contrastive learning enables better visual representations for downstream intelligent analysis tasks [10,13,29]. We can see from Fig. 2 that, the normal distortion-oriented compressed images contain more low-frequency information, aligning with the intuition that the human eye is more sensitive to low-frequency components of image [69]. In contrast, the cognition-oriented compressed images contain more valid and critical high-frequency information, making them more suitable for machine vision analysis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.656, + 0.788, + 0.822 + ], + "angle": 0, + "content": "Then, we attain variable rate-cognition by regulating quantization degree based on a channel-controllable latent gain unit [20]. This module is end-to-end trainable and can be well compatible with the above contrastive objectives. Moreover, we analyze the difference in the information needed for cognition and distortion, and based on the results we further propose an auxiliary branch to supplement residual information with an extra scalable bitstream. We find only few bits cost could significantly reduce the distortion in compressed images. At last, by controlling interpolation between distortion-oriented and cognition-oriented compressed images, we can adjust the distortion-cognition trade-off in a single codec, thereby achieving controllable rate-distortion-cognition. This effectively enhances the practicality of ICM methods." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.825, + 0.689, + 0.84 + ], + "angle": 0, + "content": "The contributions of this paper can be summarized as follows:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.129 + ], + "angle": 0, + "content": "Jinming Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.147, + 0.784, + 0.191 + ], + "angle": 0, + "content": "- We propose a rate-distortion-cognition controllable versatile neural image compression model. To the best of our knowledge, we are the first to achieve controllable rate-distortion-cognition within a single codec." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.192, + 0.785, + 0.237 + ], + "angle": 0, + "content": "- We introduce a cognition-oriented loss based on contrastive learning to generate task-friendlyly compressed images, which main compression process also attains variable-bitrate with a trainable gain unit." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.238, + 0.785, + 0.283 + ], + "angle": 0, + "content": "- We explore the relationship between the information required for distortion and cognition, and design an auxiliary scalable bitstream to achieve a balanced cognition-distortion trade-off with a two-branch interpolation strategy." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.147, + 0.785, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.293, + 0.786, + 0.354 + ], + "angle": 0, + "content": "Experimental results demonstrate that our approach achieves satisfactory performance in both machine tasks and reconstruction, and outperforms both the latest ICM and learned image compression (LIC) methods in terms of adjustable rate-distortion-cognition practicality." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.366, + 0.398, + 0.382 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.39, + 0.426, + 0.405 + ], + "angle": 0, + "content": "2.1 Image Compression" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.411, + 0.788, + 0.669 + ], + "angle": 0, + "content": "Image compression seeks to encode original image in a format that is both compact and retains high fidelity. Traditional image codecs, as delineated in numerous studies [9, 54, 60, 64, 66], typically incorporate modules such as intra prediction, discrete cosine transformation or wavelet transformation, quantization, and entropy coding. In contrast, learned-based codecs [6, 17, 45, 51-53] leverage neural networks to minimize distortion and bitrate. Compared to traditional approaches, data-driven learned methods have currently achieved better rate-distortion performance. Moreover, the rate-distortion trade-off in these models is governed by a Lagrange multiplier \\(\\lambda\\), which typically limits most methods to a single operation point on the rate-distortion curve for a fixed \\(\\lambda\\) value. Recent advancements [19, 20, 37, 63, 70] have introduced various approaches to enable variable-rate compression within a single model framework. For example, Song et al. [59] innovatively propose spatial bit allocation in accordance with a quality map that matches the dimensions of the original image, offering a novel direction in the field of image compression. Cui et al. [20] utilize a scale matrix to scale the different channels of latent code to achieve variable bitrates. But these methods can not be directly applied to machine vision tasks due to the assessment gap." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.683, + 0.53, + 0.699 + ], + "angle": 0, + "content": "2.2 Image Compression for Machine" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Image Coding for Machines (ICM) focuses on compressing and transmitting source images to facilitate downstream intelligent tasks, including image classification [21,28,35], object detection [43,55-57], semantic segmentation [5,12,46, 68, 72]. Joint optimization of image compression models with these intelligent tasks [3,33,39,42,65] is a straightforward solution in this domain. Some feature compression methods are also proposed. [4, 15, 16, 22, 23, 25, 50, 58], enhancing both coding efficiency and computing offloading. For example, Feng et al. [25] try to compress an omni-feature for better machine vision task performance. Similarly, methods based on feature selection have been proposed [14, 26, 44, 45, 61]," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.225, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Rate-Distortion-Cognition Controllable Versatile Neural Image Compression" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.148, + 0.784, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.329, + 0.788, + 0.427 + ], + "angle": 0, + "content": "Fig. 3: The framework of our method. The framework of our method. Our pipeline is divided into a primary branch and an auxiliary branch. The primary branch generates the cognition-oriented image \\(\\hat{x}_1\\), while the auxiliary branch generates the distortion-oriented residual \\(r\\). We then use a \\(\\hat{x}_1 + (1 - \\beta)r\\) interpolation strategy to achieve a balanced cognition-distortion trade-off. These two branches are trained separately in two stages: the primary branch uses cognition-oriented loss, and the auxiliary branch uses distortion-oriented loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.56 + ], + "angle": 0, + "content": "achieving lower compression rates by compressing local regions instead of all features/images. For instance, Chen [14] et al. fine-tunes the codec via prompts, making the compressed features more suitable for downstream tasks. However, the methods mentioned above require training a separate model for compression at different bitrate points, and currently, there has been no discussion on variable-bitrate coding for machines. On the other hand, the trade-off between distortion and cognition (i.e., machine task performance) has not been publicly explored either." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.573, + 0.498, + 0.589 + ], + "angle": 0, + "content": "2.3 Realism Image Compression" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.789, + 0.76 + ], + "angle": 0, + "content": "Traditional image coding goals often aim to achieve the lowest possible distortion. However, previous works [8] have demonstrated that \"low distortion\" is not synonymous with \"high perceptual quality.\" There exists a trade-off among Rate, Distortion, and Realism (human perception). Consequently, GAN networks [52] and Diffusion models [71] have been employed to enhance realism. An attempt [1] was made to use a conditional factor to control the distortion-realism trade-off. EGIC [38] achieving similar effects using two decoders. Based on [1], Iwai et al. [34] introduce a VBR mechanism to allow the control of rate-distortion-realism on a single codec. Distinct from, yet akin to the Rate-Distortion-Realism framework, this paper considers the Rate-Distortion-Cognition issue, targeting the topic of image coding for machine (ICM)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.772, + 0.382, + 0.79 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.795, + 0.687, + 0.809 + ], + "angle": 0, + "content": "3.1 Overview of the Proposed Compression Framework" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Our methodological approach is illustrated in Fig. 3. The training process is delineated into two stages, wherein the primary branch is dedicated to the gen-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.129 + ], + "angle": 0, + "content": "Jinming Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.145, + 0.696, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.2, + 0.784, + 0.228 + ], + "angle": 0, + "content": "Fig. 4: The training stage I. The parameters of the pretrained backbone remain frozen, with only the parameters of our codec being updated." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.256, + 0.784, + 0.287 + ], + "angle": 0, + "content": "eration of cognition-oriented compressed images in stage I, while the auxiliary branch focuses on the distortion-oriented compressed images in stage II." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.292, + 0.785, + 0.458 + ], + "angle": 0, + "content": "As shown in stage I of Fig. 4, to enhance the performance of the compressed image \\(\\hat{\\pmb{x}}_1\\) on downstream machine vision tasks, we employ a cognition-oriented loss instead of the conventional distortion metric commonly used in image compression. This loss comprises two parts: 1). Inspired by studies [2, 25] indicating that semantic-level loss helps image cognition and downstream tasks in compression, the first component of the cognition-oriented loss designed herein encompasses a contrastive loss [29] that facilitates the extraction of robust and discriminative representations. 2). Another part of the cognition-oriented loss is MSE of local regions in the reconstructed image where the pixel intensity exceeds [0,1]. This part is used as a penalty term to align the range of normal image values. The stage I process can be described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.383, + 0.471, + 0.786, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} \\stackrel {\\text {e n c o d e}} {\\longrightarrow} \\boldsymbol {y} \\stackrel {\\text {q u a n t i z a t i o n}} {\\longrightarrow} \\hat {\\boldsymbol {y}} \\stackrel {\\text {d e c o d e}} {\\longrightarrow} \\hat {\\boldsymbol {x}} _ {1} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.784, + 0.59 + ], + "angle": 0, + "content": "where \\(\\pmb{x}\\) is the original image, \\(\\pmb{y}\\) and \\(\\hat{\\pmb{y}}\\) are the latent code and quantized latent code extracted from \\(\\pmb{x}\\). \\(\\hat{\\pmb{x}}_1\\) is the cognition-oriented compressed image. Furthermore, based on the existing structure, we further attain variable rate-cognition by regulating quantization degree using a channel-controllable latent gain unit [20], this module is also trainable and compatible with the cognition-oriented loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In stage II, we implement the reconstruction of distortion-oriented compressed images to meet the requirements of human vision. Freezing the codec of the primary branch, we explore to generate distortion-oriented compressed images \\(\\hat{\\pmb{x}}_2\\) in this stage, which is defined as \\(\\hat{\\pmb{x}}_2 = \\hat{\\pmb{x}}_1 + \\pmb{r}\\): the residual \\(\\pmb{r}\\) between the cognition-oriented and distortion-oriented compressed images is expected to extract from \\(\\hat{\\pmb{y}}\\) at the decoder side. However, since \\(\\hat{\\pmb{y}}\\) is optimized for cognition in stage I, it didn't contain enough distortion-oriented supplemented information. Therefore, we construct an auxiliary branch to enrich the residual \\(\\pmb{r}\\) for the high-quality distortion-oriented compressed image \\(\\hat{\\pmb{x}}_2\\). As shown in Fig. 3, we first encode the quantization error \\((\\pmb{y} - \\hat{\\pmb{y}})\\) at the encoder side for transmission. After a scalable transform with variable bitstream, the decoded error \\((\\widehat{\\pmb{y} - \\hat{\\pmb{y}}})\\) is then added to the \\(\\hat{\\pmb{y}}\\) to get \\(\\hat{\\pmb{y}}_2\\), and a lightweight residual reconstruction network is used to refine residual \\(\\pmb{r}\\). Interpolating \\(\\pmb{r}\\) with \\(\\hat{\\pmb{x}}_1\\) using a weight \\(\\beta\\) gets \\(\\bar{\\pmb{x}} = \\hat{\\pmb{x}}_1 + (1 - \\beta)\\pmb{r}\\), which allows for different distortion-cognition performance. When \\(\\beta\\) equals 0, \\(\\bar{\\pmb{x}} = \\hat{\\pmb{x}}_2\\) the lowest distortion is achieved; conversely, when \\(\\beta\\) equals 1, \\(\\bar{\\pmb{x}} = \\hat{\\pmb{x}}_1\\) the highest cognition performance is attained." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.225, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Rate-Distortion-Cognition Controllable Versatile Neural Image Compression" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.755, + 0.162 + ], + "angle": 0, + "content": "3.2 Primary Branch: Cognition-oriented Variable-bitrate Codec" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.175, + 0.784, + 0.204 + ], + "angle": 0, + "content": "A common use of learned image compression (LIC) model [17,45,48,53], can be formulated as the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.346, + 0.207, + 0.635, + 0.224 + ], + "angle": 0, + "content": "\\[\n\\text {E n c o d i n g}: \\boldsymbol {y} = g _ {a} (\\boldsymbol {x}; \\phi_ {g _ {a}}); \\quad \\hat {\\boldsymbol {y}} = Q (\\boldsymbol {y})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.227, + 0.786, + 0.244 + ], + "angle": 0, + "content": "\\[\n\\text {D e c o d i n g :} \\hat {\\boldsymbol {x}} = g _ {s} (\\hat {\\boldsymbol {y}}; \\phi_ {g _ {s}}) \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.246, + 0.247, + 0.734, + 0.266 + ], + "angle": 0, + "content": "\\[\n\\text {H y p e r - p r i o r :} \\boldsymbol {z} = h _ {a} \\left(\\boldsymbol {y}; \\phi_ {h _ {a}}\\right); \\quad \\hat {\\boldsymbol {z}} = Q (\\boldsymbol {z}); \\quad \\boldsymbol {\\mu} _ {y}, \\boldsymbol {\\sigma} _ {y} = h _ {s} \\left(\\hat {\\boldsymbol {z}}; \\phi_ {h _ {s}}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.269, + 0.787, + 0.42 + ], + "angle": 0, + "content": "where \\(\\phi_{g_a},\\phi_{g_s},\\phi_{h_a}\\), and \\(\\phi_{h_s}\\) are the parameters of the encoder network \\(g_{a}\\), the decoder network \\(g_{s}\\), the hyper-encoder network \\(h_a\\), and the hyper-decoder network \\(h_s\\), respectively. \\(Q\\) represents the quantization which is approximated by a noise \\(\\mathcal{U}\\left(-\\frac{1}{2},\\frac{1}{2}\\right)\\) during training, and is a rounding operator during inference. For inference, an image is first input to \\(g_{a}\\) to extract a compact latent code \\(\\pmb{y}\\), and \\(\\pmb{y}\\) is quantized to obtain \\(\\hat{\\pmb{y}}\\) by quantization operator \\(Q\\). On the decoder side, \\(\\hat{\\pmb{y}}\\) is input into \\(g_{s}\\) for reconstruction to obtain the reconstructed image \\(\\hat{\\pmb{x}}\\). Additionally, the hyper path \\((h_a\\) and \\(h_s)\\) is used to help estimate the probability distribution of \\(\\pmb{y}\\), which is assumed here to be a gaussian distribution with mean \\(\\mu_y\\) and variance \\(\\sigma_y\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.422, + 0.788, + 0.572 + ], + "angle": 0, + "content": "As discussed before, learned image compression typically requires learning a new model for different bitrate points. Therefore, in this paper, we employ a variable bitrate approach to achieve compression at different bitrates through a single model. Inspired by [20], we scale \\( \\mathbf{y} \\) and \\( \\mathbf{z} \\) using \\( N \\) gain unit \\( G_{y,n} \\) and \\( G_{z,n} \\), which is a matrix containing \\( C \\) factors, where \\( C \\) is the channel dimension of \\( \\mathbf{y} \\), and \\( n \\in [1,N] \\). By scaling different channels, we can adjust the precision of the data we need to transmit, further achieving transmission at \\( N \\) different bitrates. Furthermore, we can interpolate the scaling factors for other bitrate points from these \\( N \\) gain units, allowing us to switch between different bitrates using a single model. Therefore, Equation 2 can be re-formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.311, + 0.577, + 0.689, + 0.595 + ], + "angle": 0, + "content": "\\[\n\\text {E n c o d i n g}: \\boldsymbol {y} = g _ {a} (\\boldsymbol {x}; \\phi_ {g _ {a}}); \\quad \\hat {\\boldsymbol {y}} = Q (\\boldsymbol {y} \\cdot \\boldsymbol {G} _ {y, n}) / \\boldsymbol {G} _ {y, n}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.404, + 0.597, + 0.597, + 0.616 + ], + "angle": 0, + "content": "\\[\n\\text {D e c o d i n g}: \\hat {\\boldsymbol {x}} = g _ {s} (\\hat {\\boldsymbol {y}}; \\phi_ {g _ {s}})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.274, + 0.61, + 0.785, + 0.635 + ], + "angle": 0, + "content": "\\[\n\\text {H y p e r - p r i o r :} z = h _ {a} \\left(\\boldsymbol {y} \\cdot \\boldsymbol {G} _ {y, n}; \\phi_ {h _ {a}}\\right); \\quad \\hat {\\boldsymbol {z}} = Q (\\boldsymbol {z} \\cdot \\boldsymbol {G} _ {z, n}) / \\boldsymbol {G} _ {z, n}; \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.426, + 0.638, + 0.574, + 0.656 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} _ {y}, \\sigma_ {y} = h _ {s} (\\hat {z}; \\phi_ {h _ {s}})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.784, + 0.688 + ], + "angle": 0, + "content": "For conventional distortion-oriented compression tasks, the loss can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.246, + 0.693, + 0.786, + 0.731 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\mathcal {R} (\\hat {\\boldsymbol {y}}) + \\mathcal {R} (\\hat {\\boldsymbol {z}}) + \\lambda_ {n} \\cdot \\mathcal {D} (\\boldsymbol {x}, \\hat {\\boldsymbol {x}}) \\\\ = \\mathbb {E} \\left[ - \\log_ {2} \\left(p _ {\\hat {\\boldsymbol {y}} | \\hat {\\boldsymbol {z}}} (\\hat {\\boldsymbol {y}} \\mid \\hat {\\boldsymbol {z}})\\right) \\right] + \\mathbb {E} \\left[ - \\log_ {2} \\left(p _ {\\hat {\\boldsymbol {z}} | \\psi} (\\hat {\\boldsymbol {z}} \\mid \\psi)\\right) \\right] + \\lambda_ {n} \\cdot \\mathcal {D} (\\boldsymbol {x}, \\hat {\\boldsymbol {x}}) \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "where the parameter \\(\\lambda_{n}\\) regulates the balance between rate and distortion, with varying \\(\\lambda_{n}\\) values corresponding to different bit rates. \\(\\psi\\) is a factorized density model used to encode quantized \\(\\hat{z}\\). During training, we randomly select one of the \\(N\\) bitrate points for each forward pass. The distortion term, denoted by \\(\\mathcal{D}(\\pmb{x},\\hat{\\pmb{x}})\\), is measured using the Mean Squared Error (MSE) loss, GAN loss or MS-SSIM loss. Additionally, \\(\\mathcal{R}(\\hat{\\pmb{y}})\\) and \\(\\mathcal{R}(\\hat{z})\\) represent the bit rates for the latent representations \\(\\hat{\\pmb{y}}\\) and \\(\\hat{\\pmb{z}}\\), respectively. To improve the performance of" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.129 + ], + "angle": 0, + "content": "Jinming Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.206 + ], + "angle": 0, + "content": "the reconstructed images for various downstream tasks, here we use cognition-oriented loss as a substitute for the original MSE loss [17, 45, 48]. The first part of the cognition-oriented loss is a contrastive loss computed using a pretrained MoCo [29]. It is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.206, + 0.785, + 0.238 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} = - \\log \\frac {\\exp \\left(q \\cdot k _ {+} / \\tau\\right)}{\\sum_ {i = 0} ^ {K} \\exp \\left(q \\cdot k _ {i} / \\tau\\right)} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.785, + 0.462 + ], + "angle": 0, + "content": "where \\( q \\) denotes the query vector derived from the current image, \\( k_{+} \\) is a positive key vector generated from a different augmentation of the same image, and \\( k_{i} \\) are the negative key vectors from different images. The denominator aggregates the exponentiated similarities of the query with both the positive and K negative keys, normalized by a temperature parameter \\( \\tau \\), which adjusts the sharpness of the distribution. This loss function encourages the model to produce embeddings where positive pairs are closer in the feature space than any negative pair, thus learning robust and discriminative feature representations. However, unlike previous work with contrastive loss, which typically aims to learn a better backbone [29] or omni-features through contrastive learning [25], our objective here is to update the codec's parameters to enable the generation of cognition-oriented compressed images. Therefore, we impose constraints on the output values of the codec, to align the range of normal image values and avoid outliers. Here, we compute a local Mean Squared Error (MSE) Loss for values lying outside the [0, 1] range. It can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.312, + 0.462, + 0.785, + 0.5 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {\\text {l o c a l}} (\\boldsymbol {x}, \\hat {\\boldsymbol {x}}) = \\frac {1}{M} \\sum_ {i = 1} ^ {N} \\left[ \\mathbb {I} \\left(\\hat {x} _ {i} < 0 \\vee \\hat {x} _ {i} > 1\\right) \\cdot \\left(\\hat {x} _ {i} - x _ {i}\\right) ^ {2} \\right] \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.5, + 0.785, + 0.56 + ], + "angle": 0, + "content": "Here, \\(\\mathbb{I}\\) is an indicator function that takes the value of 1 when the predicted value \\(\\hat{x}_i\\) falls outside the interval [0,1], and 0 otherwise. \\(x_{i}\\) and \\(\\hat{x}_i\\) represent the target and predicted values for the \\(i\\) th pixel, respectively, while \\(M\\) denotes the number of pixels. \\(\\mathcal{D}_{\\mathrm{local}}\\) is used as a penalty term to constrain the output range." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.561, + 0.786, + 0.726 + ], + "angle": 0, + "content": "As illustrated in Fig. 5, the image distributions of both the original image and the distortion-oriented compressed image primarily fall within the range of 0 to 1. However, if we directly optimize using the contrastive loss, the generated representations would reside within a significantly broader interval (approximately [-4, 6]), which does not align with the inherent characteristics of images. Moreover, directly clipping the generated representations to the [0, 1] range during training before passing them to subsequent networks can lead to the issue of vanishing gradients, resulting in a distribution of representations as shown in Fig. 5(d). Clipping values outside the [0, 1] range during inference would lead to substantial information loss. Employing a local MSE loss as a penalty term can effectively mitigate this problem (see Fig. 5(e))." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.727, + 0.664, + 0.741 + ], + "angle": 0, + "content": "Therefore, the overall loss of stage I in the main branch is:" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.743, + 0.785, + 0.758 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {I} = \\mathcal {R} (\\hat {\\boldsymbol {y}}) + \\mathcal {R} (\\hat {\\boldsymbol {z}}) + \\lambda_ {n} \\cdot \\mathcal {C} + \\lambda_ {\\text {l o c a l}} \\cdot \\mathcal {D} _ {\\text {l o c a l}} (\\boldsymbol {x}, \\hat {\\boldsymbol {x}}) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.757, + 0.786, + 0.845 + ], + "angle": 0, + "content": "where the parameter \\(\\lambda_{\\mathrm{local}}\\) serves to modulate the magnitude of influence exerted by the penalty term, denoted as \\(\\mathcal{D}_{\\mathrm{local}}\\). It is worth noting that the primary function of \\(\\mathcal{D}_{\\mathrm{local}}\\) loss within this context is to enforce constraints on the output range. Consequently, we empirically assign a relatively small fixed value of \\(10^{-5}\\) to \\(\\lambda_{\\mathrm{local}}\\), acknowledging its role in merely constraining the output range without significantly impacting the primary loss components." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.225, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Rate-Distortion-Cognition Controllable Versatile Neural Image Compression" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.338, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.269, + 0.226, + 0.288, + 0.238 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.341, + 0.144, + 0.45, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.39, + 0.226, + 0.408, + 0.238 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.144, + 0.563, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.226, + 0.517, + 0.238 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.566, + 0.144, + 0.676, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.611, + 0.226, + 0.629, + 0.238 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.144, + 0.785, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.722, + 0.226, + 0.74, + 0.238 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.242, + 0.785, + 0.298 + ], + "angle": 0, + "content": "Fig. 5: The histograms for pixel values of: (a) the original image. (b) distortion-oriented compressed image. (c) cognition-oriented compressed image without using local mse loss. (d) cognition-oriented compressed images processed with clipping to the range \\((0, 1)\\) during training. (e) cognition-oriented compressed images with using local mse loss." + }, + { + "type": "table_caption", + "bbox": [ + 0.307, + 0.303, + 0.694, + 0.315 + ], + "angle": 0, + "content": "Table 1: Parameter numbers of different Codec modules." + }, + { + "type": "table", + "bbox": [ + 0.265, + 0.317, + 0.734, + 0.345 + ], + "angle": 0, + "content": "
Base CodecScalable TransformResidual Reconstruction
Parameters(/M)20.390.191.20
" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.36, + 0.753, + 0.377 + ], + "angle": 0, + "content": "3.3 Auxiliary Branch: Distortion-oriented Residual Supplement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.785, + 0.568 + ], + "angle": 0, + "content": "To further meet the requirements of human vision, we employ an auxiliary branch to supplement residual information with a scalable bitstream. As shown in Fig. 3, we construct a residual reconstruction network to generate the residual \\( \\boldsymbol{r} \\). By adding \\( \\boldsymbol{r} \\) to \\( \\hat{\\boldsymbol{x}}_1 \\), we reconstruct an distortion-oriented compressed image \\( \\hat{\\boldsymbol{x}}_2 \\). The residual reconstruction network comprises multiple layers of transposed convolutions and point-wise convolutional layers. The former is to increase the size of the representation, ensuring spatial dimension alignment with the original decoder's representation. The latter is employed to reduce and align the channel dimensions, simultaneously decreasing the model parameters and computational complexity. Additionally, the intermediate layer outputs \\( f_1, f_2 \\), and \\( f_3 \\) from the original decoder serve as prior information. They are fed into the residual reconstruction network through a point-wise convolution to help recovery." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.787, + 0.78 + ], + "angle": 0, + "content": "We attempt to input \\(\\hat{\\pmb{y}}\\) to the residual reconstruction network directly to generate \\(\\pmb{r}\\). However, directly using \\(\\hat{\\pmb{y}}\\) for reconstruction does not work, because \\(\\hat{\\pmb{y}}\\) of the primary branch is optimized for cognitive perception rather than distortion minimization. To prove that, we analyze the quantization operation: as shown in Equation 3, this process involves the multiplication of \\(\\pmb{y}\\) by \\(G_{y,n}\\) followed by quantization. \\(G_{y,n}\\) plays a crucial role in determining the precision of quantization across different channels and, i.e., the extent of quantization error. Given that \\(G_{y,n}\\) is optimized for cognitive perception rather than distortion minimization in stage I, thus the reconstructed representations \\(\\hat{\\pmb{y}}\\) cannot recover the detailed residual. The gradient visualization in Fig. 6(a)(b) reveals the disparities in channel contributions between cognition-oriented and distortion-oriented images. By backpropagating the mean gradients \\(\\frac{\\partial\\hat{\\pmb{x}}}{\\partial\\hat{\\pmb{y}}}\\) from the generated image into \\(\\hat{\\pmb{y}}\\)’s channels, we observe a marked difference in how \\(\\hat{\\pmb{y}}\\)’s channels influence the generation of cognition-oriented versus distortion-oriented compressed images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "To address this issue, we propose to transmit this missed information through a scalable bitstream. Specifically, by using the similar VBR method in Sec. 3.2, we first encode the quantization-induced error \\((\\pmb{y} - \\hat{\\pmb{y}})\\) into a bitstream via a transformation through three convolutional layers for transmission to the de" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "Jinming Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.145, + 0.391, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.145, + 0.57, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.586, + 0.16, + 0.776, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.234, + 0.785, + 0.277 + ], + "angle": 0, + "content": "Fig. 6: (a) The gradient for \\(\\hat{\\pmb{y}}\\) from cognition-orient image \\(\\hat{\\pmb{x}}_1\\) (b) The gradient for \\(\\hat{\\pmb{y}}\\) from distortion-orient image \\(\\hat{\\pmb{x}}_2\\) (c) The energy of cognition-orient \\(\\hat{\\pmb{y}}\\) and distortion-orient \\(\\hat{\\pmb{y}}_2\\) (Purple represents the overlapping part of the energy between \\(\\hat{\\pmb{y}}\\) and \\(\\hat{\\pmb{y}}_2\\))." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.288, + 0.784, + 0.378 + ], + "angle": 0, + "content": "coder side. An auxiliary gain unit is used to control the precision and bitrate of the bitstream. Then, we reconstruct bitstream through another three convolutional layers, and the output is added to \\(\\hat{\\pmb{y}}\\) to produce \\(\\hat{\\pmb{y}}_2\\), which is fed into the residual reconstruction network to generate \\(\\pmb{r}\\). The scalable transform and the residual reconstruction network together form the auxiliary branch, which exclusively costs fewer parameters compared to the base codec (see Table 1)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.379, + 0.785, + 0.469 + ], + "angle": 0, + "content": "Moreover, we also analyze the differences in energy across various channels between \\(\\hat{\\pmb{y}}_2\\) and the original \\(\\hat{\\pmb{y}}\\) in Fig. 6(c). It is observable that the energy levels across most channels are similar, with only a few channels significantly increased. This suggests that we only need to transmit very few bits to supplement such residuals to enhance decompression quality, which will be further proved in the experiments. Therefore, the overall loss of stage II can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.473, + 0.785, + 0.49 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {I I} = \\mathcal {R} (\\hat {\\boldsymbol {s}}) + \\lambda_ {m} \\cdot \\mathcal {D} (\\boldsymbol {x}, \\hat {\\boldsymbol {x}} _ {2}) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.784, + 0.527 + ], + "angle": 0, + "content": "where \\(\\mathcal{R}(\\hat{\\pmb{s}})\\) represents the bitrate of the scalable bitstream, \\(\\mathcal{D}\\) is defined as MSE loss, \\(\\lambda_{m}\\) is used to control rate-distortion, \\(m\\in [1,N]\\), and \\(\\hat{\\pmb{x}}_2 = \\hat{\\pmb{x}}_1 + \\pmb{r}\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.541, + 0.561, + 0.557 + ], + "angle": 0, + "content": "3.4 Rate-Distortion-Cognition Trade-off" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.559, + 0.784, + 0.633 + ], + "angle": 0, + "content": "Through the aforementioned procedures, we are able to achieve two distinct trade-offs: rate-distortion and rate-cognition. To achieve ultra-controllability, by employing interpolation, we can adjust the trade-off between distortion and cognition, thereby enabling the controllable rate-distortion-cognition. The equation is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.422, + 0.635, + 0.785, + 0.65 + ], + "angle": 0, + "content": "\\[\n\\bar {\\boldsymbol {x}} = \\hat {\\boldsymbol {x}} _ {1} + (1 - \\beta) \\times \\hat {\\boldsymbol {r}} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.647, + 0.785, + 0.737 + ], + "angle": 0, + "content": "where \\(\\beta \\in [0,1]\\) is used to adjust the balance between cognition and distortion. As \\(\\beta\\) approaches 1, the distortion of \\(\\bar{\\pmb{x}}\\) increases, and the cognition performance of \\(\\bar{\\pmb{x}}\\) is enhanced. Conversely, as \\(\\beta\\) approaches 0, the distortion decreases, and the cognition performance diminishes. When transferring to downstream tasks, \\(\\beta\\) is set to 1, the codec weights are kept frozen, and we only use cognition-oriented compressed images to finetune the network for the downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.757, + 0.376, + 0.774 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.777, + 0.44, + 0.793 + ], + "angle": 0, + "content": "4.1 Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.841 + ], + "angle": 0, + "content": "**Codec.** We set TIC [48] as our base codec. During training stage I, at each step, we randomly select a number from the set \\([0.0625, 0.125, 0.25, 0.5, 1, 2]\\) as \\(\\lambda_{n}\\). Correspondingly, the latent code \\(\\pmb{y}\\) is multiplied by the associated gain unit to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.225, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rate-Distortion-Cognition Controllable Versatile Neural Image Compression" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "implement variable bit rates. For convenience, in the subsequent sections, we use \\(\\alpha \\in [0,1]\\) to represent the selected point of the bitrate. When \\(\\alpha = 0\\), the bitrate is at its minimum, corresponding to the model with \\(\\lambda_{n} = 0.0625\\). Conversely, when \\(\\alpha = 1\\), the bitrate is at its maximum, corresponding to the model with \\(\\lambda_{n} = 2\\). For training stage II, we set \\(\\lambda_{m} \\in [0.00018, 0.00036, 0.00072, 0.001, 0.0015, 0.002]\\). To assess the rate-distortion performance of the reconstructed image, we employ the Peak Signal-to-Noise Ratio (PSNR) to evaluate the extent of distortion and use bits per pixel (bpp) to measure the compression ratio. Bpp is formulated as: as \\(\\frac{b}{h \\times w}\\), where \\(h, w\\) are the height and width of the source image, \\(b\\) denotes the total bits cost of the coded feature bit-stream." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.303, + 0.794, + 0.363 + ], + "angle": 0, + "content": "Image Classification. we train and validate our model on ImageNet-1k datasets. During the evaluation, before inputting the images into the model, we resize and crop them to 224x224. Meanwhile, ResNet50 [30], serving as the task network, is used to assess the accuracy of classification." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.788, + 0.429 + ], + "angle": 0, + "content": "Object Detection. For the object detection task, we train and validate on the MS COCO2017 train/val dataset. Following the setting in [67], we employ the faster-rcnn-FPN network [57] for object detection. The image scale ranges from [640, 800] pixels during training and is set to 800 at inference by default." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.524 + ], + "angle": 0, + "content": "Semantic Segmentation. For this task, we employ DeepLab V3 [12], a network designed for segmentation tasks. The training and evaluation are conducted on the Cityscapes dataset. The Cityscapes dataset is specifically designed for urban street scenes. It includes 5,000 images with fine annotations. Cross entropy serves as the loss function during the training process. To assess effectiveness, we use the mean intersection over union (mIoU) as the metric." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.528, + 0.788, + 0.651 + ], + "angle": 0, + "content": "Baseline. For evaluating the effectiveness of our method on machine vision tasks, we have selected several different baselines for comparison with our approach. Initially, the traditional image standard VVC (VTM-12.0) [9] was utilized for testing. For learned image compression, we employ TIC [48] to assess its performance across three tasks. Regarding the ICM method, we chose to compare with recent works, namely TransTIC [14] and T-TF [11]. To evaluate the rate-distortion performance of our model, we conducted comparisons with methods such as VTM, TIC [48], Factorized [6] and MBT-Mean [53]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.657, + 0.472, + 0.672 + ], + "angle": 0, + "content": "4.2 Quantitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.789, + 0.842 + ], + "angle": 0, + "content": "As illustrated in Fig. 7 and 8, we compared our method against recent works. It is first noteworthy that our approach is controllable, by adjusting the values of \\(\\alpha\\) and \\(\\beta\\), we can achieve the ultra-controllability within the blue area depicted in the figures using a single codec. The bitrate increases with an increase in \\(\\alpha\\), while the performance for machine vision tasks and image reconstruction tasks respectively increases and decreases with an increase in \\(\\beta\\). As demonstrated by points ABCD in the Fig. 7 and 8, at \\((\\beta = 1, \\alpha = 1)\\) (Point A), we achieve the highest bitrate and the optimal machine vision task performance at this bitrate; at \\((\\beta = 0, \\alpha = 1)\\) (Point B), we obtain the highest bitrate and the lowest distortion at this bitrate; at \\((\\beta = 0, \\alpha = 0)\\) (Point C), we achieve the lowest bitrate and the lowest distortion at this bitrate; at \\((\\beta = 1, \\alpha = 0)\\) (Point D), we" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "Jinming Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.406, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.258, + 0.39, + 0.269 + ], + "angle": 0, + "content": "(a) Classification on ImageNet" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.144, + 0.596, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.435, + 0.258, + 0.581, + 0.269 + ], + "angle": 0, + "content": "(b) Segmentation on CityScapes" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.144, + 0.785, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.258, + 0.746, + 0.268 + ], + "angle": 0, + "content": "(c) Detection on COCO" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.273, + 0.786, + 0.301 + ], + "angle": 0, + "content": "Fig. 7: The rate-machine vision task performance curve (\\(\\alpha\\) changes from 0 to 1 from left to right, and \\(\\beta\\) changes from 0 to 1 from bottom to top)." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.303, + 0.409, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.281, + 0.417, + 0.344, + 0.427 + ], + "angle": 0, + "content": "(a) ImageNet" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.303, + 0.597, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.473, + 0.417, + 0.542, + 0.428 + ], + "angle": 0, + "content": "(b) CityScapes" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.302, + 0.785, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.417, + 0.714, + 0.428 + ], + "angle": 0, + "content": "(c) COCO" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.429, + 0.784, + 0.457 + ], + "angle": 0, + "content": "Fig. 8: The rate-PSNR performance curve (\\(\\alpha\\) changes from 0 to 1 from left to right, and \\(\\beta\\) changes from 0 to 1 from top to bottom)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.481, + 0.785, + 0.509 + ], + "angle": 0, + "content": "achieve the lowest bitrate and the best machine vision task performance at this bitrate." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.511, + 0.785, + 0.646 + ], + "angle": 0, + "content": "As shown in Fig. 7 and Table 2, we compared our method \\((\\beta = 1)\\) with recent approaches in three tasks: classification, segmentation, and detection. It is evident that, compared with the un-tuned learned image compression method TIC [48] and the traditional image compression standard VVC (VTM-12.0) [9], our method significantly improves machine vision tasks performance, especially at lower bitrates. This demonstrates our model's superior cognition capabilities. Furthermore, compared with some ICM methods, TransTIC [14], T-FT [11] + TIC [48], our approach not only achieves comparable or even superior performance but also provides controllability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.647, + 0.787, + 0.753 + ], + "angle": 0, + "content": "Fig. 8 and Table 3 illustrate the performance of our method \\((\\beta = 0)\\) in the image reconstruction task. It is observed that compared with TIC, which is optimized for distortion, our method, being primarily optimized for cognition, leading to a reduction in PSNR. However, it still manages to achieve results comparable to the recent learned image compression methods, such as MBT-Mean [53]. Additionally, as demonstrated in Fig. 8(a), compared with the ICM method TransTIC, our method can achieve a higher upper bound for PSNR." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.761, + 0.391, + 0.777 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "In our study, we conducted ablation studies based on the base codec TIC, as shown in Table 4. For the measurements of BD-Acc [7] and BD-PSNR, we set \\(\\beta = 1\\) and \\(\\beta = 0\\), respectively. We explored the application of the VBR method from [20] to TIC, which allows for varying bitrates with a single model. Directly" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.225, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Rate-Distortion-Cognition Controllable Versatile Neural Image Compression" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.782, + 0.174 + ], + "angle": 0, + "content": "Table 2: Comparison of machine vision Table 3: Comparison of image reconstruction task performance." + }, + { + "type": "table", + "bbox": [ + 0.245, + 0.185, + 0.773, + 0.274 + ], + "angle": 0, + "content": "
Cls.Det.Seg.
BD-AccBD-APBD-mIoU|Cls.|Det.|Seg.
|BD-PSNR||BD-PSNR||BD-PSNR|
TIC [48]0.6908-0.4421-0.0208TIC [48]0.2352-0.88760.0315
TransTIC [14]8.92181.81370.1128MBT-Mean [53]-0.8492-2.1432-0.9654
T-FT TIC [11]8.34590.70910.1049Factorized [6]-1.5387-2.3248-3.4599
Ours (β=1)9.98761.82590.1133Ours (β=0)-0.4102-1.8293-1.0240
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.277, + 0.787, + 0.319 + ], + "angle": 0, + "content": "Table 4: BD-PSNR and BD-Acc on ImageNet with TIC as anchor. Codec Number accounts for required codecs across \\( K \\) bitrates, \\( M \\) downstream tasks, and \\( T \\) distortion-cognition scenarios." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.324, + 0.776, + 0.383 + ], + "angle": 0, + "content": "
MethodsTIC [48]TIC + VBR [20]TIC + Cognition-oriented lossOurs (w/o scalable bitstream)Ours
BD-PSNR--0.1453-1.5421-1.4640-0.9104
BD-Acc--0.573911.125410.812410.6045
Codec NumbersK × M × TM × TK × T11
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.789, + 0.617 + ], + "angle": 0, + "content": "transferring VBR methods to machine vision tasks does not yield satisfactory results. At the same time, for other distinct scenarios (different tasks/networks and cognition-distortion trade-off), codec retraining remains requisite. Using the proposed cognition-oriented loss can significantly improve the performance of machine vision tasks. However, for different bitrates and cognition-distortion trade-offs, many codecs are still required. In contrast, our method requires only one codec to be applicable across all scenarios. For our method, when we do not transmit an auxiliary scalable bitstream and directly use \\(\\hat{\\pmb{y}}\\) for the training of residual reconstruction, there is a significant improvement in BD-Acc, but the distortion-oriented compressed images exhibit heavy distortion. As analyzed in Section 3.3 and shown in Fig. 9 and Table 4, we discovered that a minimal scalable bitstream can significantly reduce reconstruction distortion, with BD-PSNR improving by approximately \\(37.81\\%\\), and the loss in BD-Acc being only around \\(2\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.636, + 0.371, + 0.65 + ], + "angle": 0, + "content": "4.4 Visualization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.789, + 0.855 + ], + "angle": 0, + "content": "In Fig. 10, we present a visualization of images generated with varying \\(\\beta\\) values, along with their spectrum. It is observable that distortion-oriented compressed images (i.e., \\(\\beta = 0\\)) lose a significant amount of high-frequency information compared to the original image. However, as the \\(\\beta\\) value increases, the high-frequency information in the images progressively increases. This indicates that cognition-oriented compressed images contain more high-frequency information, making them more conducive for analysis in downstream tasks. For instance, compared to distortion-oriented compressed images, cognition-oriented compressed images depict more detailed wrinkles on the ball held in the statue's hand. On the other hand, cognition-oriented compressed images, compared to both distortion-oriented compressed images and the original, possess a certain degree of color deviation. This suggests to some extent that color restoration is not crucial for some downstream tasks." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.129 + ], + "angle": 0, + "content": "Jinming Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.145, + 0.473, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.329, + 0.239, + 0.411, + 0.251 + ], + "angle": 0, + "content": "(a) Rate-Acc." + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.145, + 0.73, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.239, + 0.678, + 0.251 + ], + "angle": 0, + "content": "(b) Rate-PSNR." + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.255, + 0.599, + 0.27 + ], + "angle": 0, + "content": "Fig. 9: The ablation studies." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.276, + 0.785, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.417, + 0.788, + 0.473 + ], + "angle": 0, + "content": "Fig. 10: Visualization results and spectrum with different \\(\\beta\\) values. (a) represents the original image, while (b) to (f) represent the compressed images with \\(\\beta\\) values of 0, 0.25, 0.5, 0.75, and 1.0, respectively, (g) represents the magnification of the corresponding colored frame area." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.491, + 0.36, + 0.507 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.788, + 0.668 + ], + "angle": 0, + "content": "In this paper, we propose a Rate-Distortion-Cognition Controllable Versatile Neural Image Compression model. We introduce a cognition-oriented loss to enhance the machine vision support, and achieve variable-rate by latent channel regulation. To further meet human eye perception, we build an auxiliary branch with a very small and scalable bitrate, decoding distortion-oriented compressed images. Finally, with an interpolation strategy to merge two branches, we also achieve distortion-cognition controlling, thereby realizing controllable rate-distortion-cognition in a single codec. Our experiments demonstrate that our method can achieve excellent performance in classification, segmentation, and detection tasks while also achieving ultra-controllability." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.704, + 0.395, + 0.722 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.782 + ], + "angle": 0, + "content": "This work was supported in part by NSFC 62302246 and ZJNSFC under Grant LQ23F010008, and supported by High Performance Computing Center at Eastern Institute of Technology, Ningbo, and Ningbo Institute of Digital Twin." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.225, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rate-Distortion-Cognition Controllable Versatile Neural Image Compression" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.176, + 0.786, + 0.218 + ], + "angle": 0, + "content": "1. Agustsson, E., Minnen, D., Toderici, G., Mentzer, F.: Multi-realism image compression with a conditional generator. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22324-22333 (2023) 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.787, + 0.272 + ], + "angle": 0, + "content": "2. Agustsson, E., Tschannen, M., Mentzer, F., Timofte, R., Gool, L.V.: Generative adversarial networks for extreme learned image compression. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 221-231 (2019) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.273, + 0.786, + 0.327 + ], + "angle": 0, + "content": "3. Akbari, M., Liang, J., Han, J.: Dsslic: Deep semantic segmentation-based layered image compression. In: ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 2042-2046. IEEE (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.328, + 0.786, + 0.357 + ], + "angle": 0, + "content": "4. Babenko, A., Slesarev, A., Chigorin, A., Lempitsky, V.: Neural codes for image retrieval. In: ECCV. pp. 584-599. Springer (2014) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.357, + 0.786, + 0.398 + ], + "angle": 0, + "content": "5. Badrinarayanan, V., Kendall, A., Cipolla, R.: Segnet: A deep convolutional encoder-decoder architecture for image segmentation. TPAMI 39(12), 2481-2495 (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.398, + 0.786, + 0.425 + ], + "angle": 0, + "content": "6. Balle, J., Laparra, V., Simoncelli, E.P.: End-to-end optimized image compression. In: ICLR (2017) 4, 11, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.425, + 0.786, + 0.453 + ], + "angle": 0, + "content": "7. Bjontegaard, G.: Calculation of average psnr differences between rd-curves. In: VCEG-M33 (2001) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.453, + 0.786, + 0.495 + ], + "angle": 0, + "content": "8. Blau, Y., Michaeli, T.: Rethinking lossy compression: The rate-distortion-perception tradeoff. In: International Conference on Machine Learning. pp. 675-685. PMLR (2019) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.495, + 0.786, + 0.536 + ], + "angle": 0, + "content": "9. Bross, B., Wang, Y.K., Ye, Y., Liu, S., Chen, J., Sullivan, G.J., Ohm, J.R.: Overview of the versatile video coding (vvc) standard and its applications. TCSVT (2021) 1, 4, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.536, + 0.786, + 0.589 + ], + "angle": 0, + "content": "0. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.59, + 0.786, + 0.633 + ], + "angle": 0, + "content": "1. Chamain, L.D., Racapé, F., Bégaint, J., Pushparaja, A., Feltman, S.: End-to-end optimized image compression for machines, a study. In: 2021 Data Compression Conference (DCC). pp. 163-172. IEEE (2021) 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.786, + 0.675 + ], + "angle": 0, + "content": "2. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. TPAMI 40(4), 834-848 (2017) 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.786, + 0.714 + ], + "angle": 0, + "content": "3. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: ICML. pp. 1597-1607. PMLR (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.715, + 0.786, + 0.77 + ], + "angle": 0, + "content": "4. Chen, Y.H., Weng, Y.C., Kao, C.H., Chien, C., Chiu, W.C., Peng, W.H.: Transtic: Transferring transformer-based image compression from human perception to machine perception. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 23297-23307 (2023) 2, 4, 5, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.786, + 0.813 + ], + "angle": 0, + "content": "5. Chen, Z., Fan, K., Wang, S., Duan, L.Y., Lin, W., Kot, A.: Lossy intermediate deep learning feature compression and evaluation. In: ACM MM. pp. 2414-2422 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.786, + 0.841 + ], + "angle": 0, + "content": "6. Chen, Z., Fan, K., Wang, S., Duan, L., Lin, W., Kot, A.C.: Toward intelligent sensing: Intermediate deep feature compression. TIP 29, 2230-2243 (2019) 4" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.176, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "Jinming Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "17. Cheng, Z., Sun, H., Takeuchi, M., Katto, J.: Learned image compression with discretized gaussian mixture likelihoods and attention modules. In: CVPR. pp. 7939-7948 (2020) 1, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.217 + ], + "angle": 0, + "content": "18. Choi, H., Bajic, I.V.: Scalable image coding for humans and machines. IEEE Transactions on Image Processing 31, 2739-2754 (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.217, + 0.785, + 0.243 + ], + "angle": 0, + "content": "19. Choi, Y., El-Khamy, M., Lee, J.: Variable rate deep image compression with a conditional autoencoder. In: ICCV. pp. 3146-3154 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.243, + 0.785, + 0.283 + ], + "angle": 0, + "content": "20. Cui, Z., Wang, J., Bai, B., Guo, T., Feng, Y.: G-vae: A continuously variable rate deep image compression framework. arXiv preprint arXiv:2003.02012 (2020) 2, 3, 4, 6, 7, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.283, + 0.785, + 0.324 + ], + "angle": 0, + "content": "21. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.324, + 0.785, + 0.352 + ], + "angle": 0, + "content": "22. Duan, L.Y., Chandrasekhar, V., Chen, J., Lin, J., Wang, Z., Huang, T., Girod, B., Gao, W.: Overview of thempeg-cdv standard. TIP 25(1), 179-194 (2015) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.785, + 0.392 + ], + "angle": 0, + "content": "23. Duan, L.Y., Lou, Y., Bai, Y., Huang, T., Gao, W., Chandrasekhar, V., Lin, J., Wang, S., Kot, A.C.: Compact descriptors for video analysis: The emerging mpeg standard. IEEE MultiMedia 26(2), 44-54 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.392, + 0.785, + 0.433 + ], + "angle": 0, + "content": "24. Duan, L., Liu, J., Yang, W., Huang, T., Gao, W.: Video coding for machines: A paradigm of collaborative compression and intelligent analytics. TIP 29, 8680-8695 (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.785, + 0.473 + ], + "angle": 0, + "content": "25. Feng, R., Jin, X., Guo, Z., Feng, R., Gao, Y., He, T., Zhang, Z., Sun, S., Chen, Z.: Image coding for machines with omnipotent feature learning. In: ECCV. pp. 510-528. Springer (2022) 2, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.514 + ], + "angle": 0, + "content": "26. Feng, R., Liu, J., Jin, X., Pan, X., Sun, H., Chen, Z.: Prompt-icm: A unified framework towards image coding for machines with task-driven prompts. arXiv preprint arXiv:2305.02578 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.785, + 0.569 + ], + "angle": 0, + "content": "27. Gao, W., Ma, S., Duan, L., Tian, Y., Xing, P., Wang, Y., Wang, S., Jia, H., Huang, T.: Digital retina: A way to make the city brain more efficient by visual coding. IEEE Transactions on Circuits and Systems for Video Technology 31(11), 4147-4161 (2021) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.785, + 0.596 + ], + "angle": 0, + "content": "28. Han, K., Xiao, A., Wu, E., Guo, J., Xu, C., Wang, Y.: Transformer in transformer. NeurIPS 34, 15908-15919 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.596, + 0.785, + 0.623 + ], + "angle": 0, + "content": "29. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: CVPR. pp. 9729-9738 (2020) 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.785, + 0.649 + ], + "angle": 0, + "content": "30. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.649, + 0.785, + 0.69 + ], + "angle": 0, + "content": "31. He, T., Sun, S., Guo, Z., Chen, Z.: Beyond coding: Detection-driven image compression with semantically structured bit-stream. In: 2019 Picture Coding Symposium (PCS). pp. 1-5. IEEE (2019) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.69, + 0.785, + 0.745 + ], + "angle": 0, + "content": "32. Hu, Y., Yang, J., Chen, L., Li, K., Sima, C., Zhu, X., Chai, S., Du, S., Lin, T., Wang, W., et al.: Planning-oriented autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17853-17862 (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.785, + 0.786 + ], + "angle": 0, + "content": "33. Hu, Y., Yang, S., Yang, W., Duan, L.Y., Liu, J.: Towards coding for human and machine vision: A scalable image coding approach. In: 2020 IEEE International Conference on Multimedia and Expo (ICME). pp. 1-6. IEEE (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "34. Iwai, S., Miyazaki, T., Omachi, S.: Controlling rate, distortion, and realism: Towards a single comprehensive neural image compression model. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2900-2909 (2024) 5" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.225, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rate-Distortion-Cognition Controllable Versatile Neural Image Compression" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "35. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. ECCV (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.785, + 0.217 + ], + "angle": 0, + "content": "36. Jin, X., Feng, R., Sun, S., Feng, R., He, T., Chen, Z.: Semantical video coding: Instill static-dynamic clues into structured bitstream for ai tasks. Journal of Visual Communication and Image Representation 93, 103816 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.217, + 0.785, + 0.269 + ], + "angle": 0, + "content": "37. Johnston, N., Vincent, D., Minnen, D., Covell, M., Singh, S., Chinen, T., Hwang, S.J., Shor, J., Toderici, G.: Improved lossy image compression with priming and spatially adaptive bit rates for recurrent networks. In: CVPR. pp. 4385-4393 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.27, + 0.785, + 0.311 + ], + "angle": 0, + "content": "38. Körber, N., Kromer, E., Siebert, A., Hauke, S., Mueller-Gritschneider, D.: Egic: Enhanced low-bit-rate generative image compression guided by semantic segmentation. arXiv preprint arXiv:2309.03244 (2023) 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.785, + 0.366 + ], + "angle": 0, + "content": "39. Le, N., Zhang, H., Cricri, F., Ghaznavi-Youvalari, R., Rahtu, E.: Image coding for machines: An end-to-end learned approach. In: ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1590-1594. IEEE (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.366, + 0.785, + 0.394 + ], + "angle": 0, + "content": "40. Li, H., Li, S., Dai, W., Li, C., Zou, J., Xiong, H.: Frequency-aware transformer for learned image compression. arXiv preprint arXiv:2310.16387 (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.394, + 0.785, + 0.434 + ], + "angle": 0, + "content": "41. Li, H., Li, S., Ding, S., Dai, W., Cao, M., Li, C., Zou, J., Xiong, H.: Image compression for machine and human vision with spatial-frequency adaptation. In: ECCV. Springer (2024) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.435, + 0.785, + 0.461 + ], + "angle": 0, + "content": "42. Li, X., Shi, J., Chen, Z.: Task-driven semantic coding via reinforcement learning. TIP (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.785, + 0.488 + ], + "angle": 0, + "content": "43. Li, Y., Mao, H., Girshick, R., He, K.: Exploring plain vision transformer backbones for object detection. ECCV (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.516 + ], + "angle": 0, + "content": "44. Liu, J., Sun, H., Katto, J.: Semantic segmentation in learned compressed domain. In: 2022 Picture Coding Symposium (PCS). pp. 181-185. IEEE (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.785, + 0.556 + ], + "angle": 0, + "content": "45. Liu, J., Sun, H., Katto, J.: Learned image compression with mixed transformer-cnn architectures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14388-14397 (2023) 1, 2, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.556, + 0.785, + 0.583 + ], + "angle": 0, + "content": "46. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: CVPR. pp. 3431-3440 (2015) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.785, + 0.61 + ], + "angle": 0, + "content": "47. Lu, G., Ge, X., Zhong, T., Geng, J., Hu, Q.: Preprocessing enhanced image compression for machine vision. arXiv preprint arXiv:2206.05650 (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.785, + 0.637 + ], + "angle": 0, + "content": "48. Lu, M., Guo, P., Shi, H., Cao, C., Ma, Z.: Transformer-based image compression. arXiv preprint arXiv:2111.06707 (2021) 1, 7, 8, 10, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.678 + ], + "angle": 0, + "content": "49. Ma, H., Liu, D., Yan, N., Li, H., Wu, F.: End-to-end optimized versatile image compression with wavelet-like transform. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(3), 1247-1263 (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.785, + 0.718 + ], + "angle": 0, + "content": "50. Ma, S., Zhang, X., Wang, S., Zhang, X., Jia, C., Wang, S.: Joint feature and texture coding: Toward smart video representation via front-end intelligence. TCSVT 29(10), 3095-3105 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.785, + 0.758 + ], + "angle": 0, + "content": "51. Mentzer, F., Agustsson, E., Tschannen, M., Timofte, R., Van Gool, L.: Conditional probability models for deep image compression. In: CVPR. pp. 4394-4402 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.786 + ], + "angle": 0, + "content": "52. Mentzer, F., Toderici, G., Tschannen, M., Agustsson, E.: High-fidelity generative image compression. arXiv preprint arXiv:2006.09965 (2020) 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "53. Minnen, D., Balle, J., Toderici, G.: Joint autoregressive and hierarchical priors for learned image compression. In: NeurIPS (2018) 4, 7, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.841 + ], + "angle": 0, + "content": "54. Rabbani, M., Joshi, R.: An overview of the jpeg 2000 still image compression standard. Signal processing: Image communication 17(1), 3-48 (2002) 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.129 + ], + "angle": 0, + "content": "Jinming Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "55. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified, real-time object detection. In: CVPR. pp. 779-788 (2016) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.203 + ], + "angle": 0, + "content": "56. Redmon, J., Farhadi, A.: Yolo9000: better, faster, stronger. In: CVPR. pp. 7263-7271 (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "57. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS 28, 91-99 (2015) 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "58. Singh, S., Abu-El-Haija, S., Johnston, N., Balle, J., Shrivastava, A., Toderici, G.: End-to-end learning of compressible features. In: 2020 IEEE International Conference on Image Processing (ICIP). pp. 3349-3353. IEEE (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "59. Song, M., Choi, J., Han, B.: Variable-rate deep image compression through spatially-adaptive feature transform. In: ICCV. pp. 2380-2389 (2021) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.785, + 0.328 + ], + "angle": 0, + "content": "60. Sullivan, G.J., Ohm, J.R., Han, W.J., Wiegand, T.: Overview of the high efficiency video coding (hevc) standard. TCSVT 22(12), 1649-1668 (2012) 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.329, + 0.785, + 0.356 + ], + "angle": 0, + "content": "61. Sun, S., He, T., Chen, Z.: Semantic structured image coding framework for multiple intelligent applications. TCSVT (2020) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.411 + ], + "angle": 0, + "content": "62. Terhörst, P., Ihlefeld, M., Huber, M., Damer, N., Kirchbuchner, F., Raja, K., Kuijper, A.: Qmagface: Simple and accurate quality-aware face recognition. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3484-3494 (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "63. Toderici, G., Vincent, D., Johnston, N., Jin Hwang, S., Minnen, D., Shor, J., Covell, M.: Full resolution image compression with recurrent neural networks. In: CVPR. pp. 5306-5314 (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.479 + ], + "angle": 0, + "content": "64. Wallace, G.K.: Thejpeg still picture compression standard. IEEE transactions on consumer electronics 38(1), xviii-xxxiv (1992) 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.48, + 0.785, + 0.522 + ], + "angle": 0, + "content": "65. Wang, S., Wang, S., Yang, W., Zhang, X., Wang, S., Ma, S., Gao, W.: Towards analysis-friendly face representation with scalable feature and texture compression. TMM (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.55 + ], + "angle": 0, + "content": "66. Wiegand, T., Sullivan, G.J., Bjontegaard, G., Luthra, A.: Overview of the h. 264/avc video coding standard. TCSVT 13(7), 560-576 (2003) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.577 + ], + "angle": 0, + "content": "67. Wu, Y., Kirillov, A., Massa, F., Lo, W.Y., Girshick, R.: Detector2. https://github.com/facebookresearch/detectron2 (2019) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.618 + ], + "angle": 0, + "content": "68. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. NeurIPS 34, 12077-12090 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.659 + ], + "angle": 0, + "content": "69. Yan, F.F., Hou, F., Lu, Z.L., Hu, X., Huang, C.B.: Efficient characterization and classification of contrast sensitivity functions in aging. Scientific reports 7(1), 5045 (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.785, + 0.702 + ], + "angle": 0, + "content": "70. Yang, F., Herranz, L., Van De Weijer, J., Guitián, J.A.I., López, A.M., Mozerov, M.G.: Variable rate deep image compression with modulated autoencoder. IEEE Signal Processing Letters 27, 331-335 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.73 + ], + "angle": 0, + "content": "71. Yang, R., Mandt, S.: Lossy image compression with conditional diffusion models. arXiv preprint arXiv:2209.06950 (2022) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.785, + 0.771 + ], + "angle": 0, + "content": "72. Zheng, S., Lu, J., Zhao, H., Zhu, X., Luo, Z., Wang, Y., Fu, Y., Feng, J., Xiang, T., Torr, P.H., et al.: Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In: CVPR. pp. 6881-6890 (2021) 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.771 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Rate-Distortion-Cognition Controllable Versatile Neural Image Compression/d4e52d64-79ee-43c9-8953-86dbf5d724e6_origin.pdf b/2024/Rate-Distortion-Cognition Controllable Versatile Neural Image Compression/d4e52d64-79ee-43c9-8953-86dbf5d724e6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..080329c2383ea0ff50bd9fa5292dfab6a99282c4 --- /dev/null +++ b/2024/Rate-Distortion-Cognition Controllable Versatile Neural Image Compression/d4e52d64-79ee-43c9-8953-86dbf5d724e6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59d90f9a34e07414b35229777ae8b3afce8dad76d882d1ad7e35a4be16051212 +size 3326853 diff --git a/2024/Rate-Distortion-Cognition Controllable Versatile Neural Image Compression/full.md b/2024/Rate-Distortion-Cognition Controllable Versatile Neural Image Compression/full.md new file mode 100644 index 0000000000000000000000000000000000000000..38c7c56c4e7c443ecf9166cb104395bd1dadbb4e --- /dev/null +++ b/2024/Rate-Distortion-Cognition Controllable Versatile Neural Image Compression/full.md @@ -0,0 +1,389 @@ +# Rate-Distortion-Cognition Controllable Versatile Neural Image Compression + +Jinming Liu $^{1,2}$ , Ruoyu Feng $^{3}$ , Yunpeng Qi $^{3}$ , Qiuyu Chen $^{2}$ , Zhibo Chen $^{3}$ , Wenjun Zeng $^{2}$ , Xin Jin $^{2\dagger}$ + +$^{1}$ Shanghai Jiao Tong University + $^{2}$ Ningbo Institute of Digital Twin, Eastern Institute of Technology, Ningbo, China + $^{3}$ University of Science and Technology of China +jmliu206@sjtu.edu.cn jinxin@eitech.edu.cn + +Abstract. Recently, the field of Image Coding for Machines (ICM) has garnered heightened interest and significant advances thanks to the rapid progress of learning-based techniques for image compression and analysis. Previous studies often require training separate codes to support various bitrate levels, machine tasks, and networks, thus lacking both flexibility and practicality. To address these challenges, we propose a rate-distortion-cognition controllable versatile image compression, which method allows the users to adjust the bitrate (i.e., Rate), image reconstruction quality (i.e., Distortion), and machine task accuracy (i.e., Cognition) with a single neural model, achieving ultra-controllability. Specifically, we first introduce a cognition-oriented loss in the primary compression branch to train a codec for diverse machine tasks. This branch attains variable bitrate by regulating quantization degree through the latent code channels. To further enhance the quality of the reconstructed images, we employ an auxiliary branch to supplement residual information with a scalable bitstream. Ultimately, two branches use a $\beta x + (1 - \beta)y$ interpolation strategy to achieve a balanced cognition-distortion trade-off. Extensive experiments demonstrate that our method yields satisfactory ICM performance and flexible Rate-Distortion-Cognition controlling. + +Keywords: Neural Image Compression $\cdot$ Image Compression for Machine $\cdot$ Variable-bitrate Compression $\cdot$ Cognition-distortion Trade-off + +# 1 Introduction + +Machine vision tasks have become increasingly integral in numerous daily applications, including intelligent traffic systems [27], autonomous driving [32], and facial recognition [62]. As a result, there is a burgeoning consumption of images by machines, which poses significant challenges in the transmission and storage of such extensive data volumes. How to efficiently compress these data while promising machine task accuracy urgently needs to be studied. + +Traditional image compression standards, such as JPEG [64], HEVC [60], VVC [9], have been widely applied in life. Meanwhile, deep learning-based CODECs have rapidly developed recently, achieving superior rate-distortion performance [17, + +![](images/de05175552ea524d0413990280a9a4f964c7d763467e644ab9b13fbaaddb66ad.jpg) +Fig. 1: Motivation: (a) Traditional image compression tasks require training $K \times M \times T$ codecs for $K$ bitrate points, $M$ machine vision tasks, and $T$ cognition-distortion trade-offs. (b) Our method enables controllable rate-distortion-cognition with a single codec. + +![](images/5ec7eced2ac4996a40b3c7490baeaf0f93779d37b714d41f93ce177cfe4626bb.jpg) + +[40, 45, 48]. However, despite success in bit saving and visual fidelity, these methods hardly generalize well to machine tasks w.r.t cognition performance. + +That's because traditional image compression CODECs are typically optimized to align with human visual perception. They employ distortion metrics, such as PSNR and MS-SSIM [45], as well as realism metrics, including KID and FID [1], to assess the loss incurred by compression procedure. Hence, utilizing these CODEcs directly for machine vision tasks may yield suboptimal outcomes [24, 47]. + +To address this issue, the field of image coding for machines (ICM) has gradually come into focus [24,31,36,41,61]. For example, some works improves the performance of downstream tasks by jointly optimizing the codec and the task network [49]. Choi et al. [18] adopt a scalable method to decouple the features required for downstream machine vision tasks and reconstruction tasks, thereby reducing the required bitrate. Chen et al. [14] fine-tune the codec for different tasks through prompts to achieve better performance. Feng et al. [25] employ a contrastive learning approach to learn an omni-feature to improve the performance on various downstream tasks. However, these methods all remain the following challenges: 1). need to train multiple independent networks to adapt to different bit rates, thus the training cost and memory requirement increase with the desired rate-distortion points proportionally. 2). hard to accommodate diverse machine tasks/networks, bringing extra fine-tuning overhead and bad generalization. 3). fail to meet the practical demand for adjustable distortion-cognition trade-off (decompression fidelity vs. task performance) within a single framework. In summary, as illustrated in Fig. 1(a), previous image compression methods require training different codecs to support various rate-distortion levels and machine tasks, posing a large overhead in practical scene deployment. + +Some variable-bitrate (VBR) image compression methods attempt to achieve compression at different bitrates through a single model [20, 59]. For example, Song et al. [59] use a quality map to indicate compression bitrates, achieving compression at various rates. Cui et al. [20] achieve VBR by scaling different channels of the latent code through a gain unit. However, to our best knowledge, there are currently no VBR methods specifically designed for machine tasks. Most VBR methods are optimized for human vision, and can not be directly applied to machine vision tasks due to the assessment gap. + +![](images/f1886140281f7a477236101df4633be56c5fbfbffc040d474d770bace4a5bff8.jpg) + +![](images/21dc0a3ed2c0bf008c2f346f3af059845efc27ce5d57877849ae49a654eccefb.jpg) +(a) +Fig. 2: The visualization and spectrum of (a) original image, (b) distortion-oriented compressed image $\hat{\pmb{x}}_2$ , (c) cognition-oriented compressed image $\hat{\pmb{x}}_1$ and, (d) the residual $|\hat{\pmb{x}}_1 - \hat{\pmb{x}}_2|$ of (b) and (c). + +![](images/1d81dc35dd0b2a38b7429f6fa8ffc1fba2ab25334cf74de10d912b89890564da.jpg) + +![](images/410381c6ad787e40b23d3211390b8b39f3823075d7d5af2f8dd517f87a484ce3.jpg) +(b) + +![](images/4cad8967b79e68d7f72b72d4a42086628d0e04323af4769f6e15eeed49ad4510.jpg) + +![](images/e3cbd58f7fccb2a541a1d844a675a6006c0e045e50ff3d38fb0f7dfe7fc0546f.jpg) +(c) + +![](images/472fa1f49c83a6e0dcbb80aa97aac072fe4e59d73cd2670697b41e2f0a3bf0a6.jpg) + +![](images/663d44ac5a7c2bec7b70ae9721da776ef3062adbbd1e60d2f48f125b7f2ee5a4.jpg) +(d) + +Meanwhile, the distortion-realism trade-off of compression has been discussed recently [1,38]. For example, Agustsson et al. [1] achieve adjustable distortion-realism by attaching several layers of neural networks outside the decoder. However, both evaluations of distortion and realism align with human perception intuition, and many assessment aspects of them are largely overlapped. Unlike distortion or realism, cognition requires more valid information needed for downstream tasks. To the best of our knowledge, there are currently no studies discussing how to adjust the distortion-cognition trade-off in a single codec. + +In response to the aforementioned issues, we propose a rate-distortion-cognition controllable versatile neural image compression method. Specifically, we first introduce a cognition-oriented loss that combines contrastive constraints in the main compression branch to reconstruct a more general machine task-friendly image. The physical meaning behind this loss is that contrastive learning enables better visual representations for downstream intelligent analysis tasks [10,13,29]. We can see from Fig. 2 that, the normal distortion-oriented compressed images contain more low-frequency information, aligning with the intuition that the human eye is more sensitive to low-frequency components of image [69]. In contrast, the cognition-oriented compressed images contain more valid and critical high-frequency information, making them more suitable for machine vision analysis. + +Then, we attain variable rate-cognition by regulating quantization degree based on a channel-controllable latent gain unit [20]. This module is end-to-end trainable and can be well compatible with the above contrastive objectives. Moreover, we analyze the difference in the information needed for cognition and distortion, and based on the results we further propose an auxiliary branch to supplement residual information with an extra scalable bitstream. We find only few bits cost could significantly reduce the distortion in compressed images. At last, by controlling interpolation between distortion-oriented and cognition-oriented compressed images, we can adjust the distortion-cognition trade-off in a single codec, thereby achieving controllable rate-distortion-cognition. This effectively enhances the practicality of ICM methods. + +The contributions of this paper can be summarized as follows: + +- We propose a rate-distortion-cognition controllable versatile neural image compression model. To the best of our knowledge, we are the first to achieve controllable rate-distortion-cognition within a single codec. +- We introduce a cognition-oriented loss based on contrastive learning to generate task-friendlyly compressed images, which main compression process also attains variable-bitrate with a trainable gain unit. +- We explore the relationship between the information required for distortion and cognition, and design an auxiliary scalable bitstream to achieve a balanced cognition-distortion trade-off with a two-branch interpolation strategy. + +Experimental results demonstrate that our approach achieves satisfactory performance in both machine tasks and reconstruction, and outperforms both the latest ICM and learned image compression (LIC) methods in terms of adjustable rate-distortion-cognition practicality. + +# 2 Related Works + +# 2.1 Image Compression + +Image compression seeks to encode original image in a format that is both compact and retains high fidelity. Traditional image codecs, as delineated in numerous studies [9, 54, 60, 64, 66], typically incorporate modules such as intra prediction, discrete cosine transformation or wavelet transformation, quantization, and entropy coding. In contrast, learned-based codecs [6, 17, 45, 51-53] leverage neural networks to minimize distortion and bitrate. Compared to traditional approaches, data-driven learned methods have currently achieved better rate-distortion performance. Moreover, the rate-distortion trade-off in these models is governed by a Lagrange multiplier $\lambda$ , which typically limits most methods to a single operation point on the rate-distortion curve for a fixed $\lambda$ value. Recent advancements [19, 20, 37, 63, 70] have introduced various approaches to enable variable-rate compression within a single model framework. For example, Song et al. [59] innovatively propose spatial bit allocation in accordance with a quality map that matches the dimensions of the original image, offering a novel direction in the field of image compression. Cui et al. [20] utilize a scale matrix to scale the different channels of latent code to achieve variable bitrates. But these methods can not be directly applied to machine vision tasks due to the assessment gap. + +# 2.2 Image Compression for Machine + +Image Coding for Machines (ICM) focuses on compressing and transmitting source images to facilitate downstream intelligent tasks, including image classification [21,28,35], object detection [43,55-57], semantic segmentation [5,12,46, 68, 72]. Joint optimization of image compression models with these intelligent tasks [3,33,39,42,65] is a straightforward solution in this domain. Some feature compression methods are also proposed. [4, 15, 16, 22, 23, 25, 50, 58], enhancing both coding efficiency and computing offloading. For example, Feng et al. [25] try to compress an omni-feature for better machine vision task performance. Similarly, methods based on feature selection have been proposed [14, 26, 44, 45, 61], + +![](images/d2ce25096ce0780aed51e22792937e02d00c099c2097f6469254eb75bbf1cae1.jpg) +Fig. 3: The framework of our method. The framework of our method. Our pipeline is divided into a primary branch and an auxiliary branch. The primary branch generates the cognition-oriented image $\hat{x}_1$ , while the auxiliary branch generates the distortion-oriented residual $r$ . We then use a $\hat{x}_1 + (1 - \beta)r$ interpolation strategy to achieve a balanced cognition-distortion trade-off. These two branches are trained separately in two stages: the primary branch uses cognition-oriented loss, and the auxiliary branch uses distortion-oriented loss. + +achieving lower compression rates by compressing local regions instead of all features/images. For instance, Chen [14] et al. fine-tunes the codec via prompts, making the compressed features more suitable for downstream tasks. However, the methods mentioned above require training a separate model for compression at different bitrate points, and currently, there has been no discussion on variable-bitrate coding for machines. On the other hand, the trade-off between distortion and cognition (i.e., machine task performance) has not been publicly explored either. + +# 2.3 Realism Image Compression + +Traditional image coding goals often aim to achieve the lowest possible distortion. However, previous works [8] have demonstrated that "low distortion" is not synonymous with "high perceptual quality." There exists a trade-off among Rate, Distortion, and Realism (human perception). Consequently, GAN networks [52] and Diffusion models [71] have been employed to enhance realism. An attempt [1] was made to use a conditional factor to control the distortion-realism trade-off. EGIC [38] achieving similar effects using two decoders. Based on [1], Iwai et al. [34] introduce a VBR mechanism to allow the control of rate-distortion-realism on a single codec. Distinct from, yet akin to the Rate-Distortion-Realism framework, this paper considers the Rate-Distortion-Cognition issue, targeting the topic of image coding for machine (ICM). + +# 3 Methodology + +# 3.1 Overview of the Proposed Compression Framework + +Our methodological approach is illustrated in Fig. 3. The training process is delineated into two stages, wherein the primary branch is dedicated to the gen- + +![](images/f48430298ab8a04ac840001a0e8d5257c3cb2df1fcc80813fdb47e1cfde3097f.jpg) +Fig. 4: The training stage I. The parameters of the pretrained backbone remain frozen, with only the parameters of our codec being updated. + +eration of cognition-oriented compressed images in stage I, while the auxiliary branch focuses on the distortion-oriented compressed images in stage II. + +As shown in stage I of Fig. 4, to enhance the performance of the compressed image $\hat{\pmb{x}}_1$ on downstream machine vision tasks, we employ a cognition-oriented loss instead of the conventional distortion metric commonly used in image compression. This loss comprises two parts: 1). Inspired by studies [2, 25] indicating that semantic-level loss helps image cognition and downstream tasks in compression, the first component of the cognition-oriented loss designed herein encompasses a contrastive loss [29] that facilitates the extraction of robust and discriminative representations. 2). Another part of the cognition-oriented loss is MSE of local regions in the reconstructed image where the pixel intensity exceeds [0,1]. This part is used as a penalty term to align the range of normal image values. The stage I process can be described as follows: + +$$ +\boldsymbol {x} \stackrel {\text {e n c o d e}} {\longrightarrow} \boldsymbol {y} \stackrel {\text {q u a n t i z a t i o n}} {\longrightarrow} \hat {\boldsymbol {y}} \stackrel {\text {d e c o d e}} {\longrightarrow} \hat {\boldsymbol {x}} _ {1} \tag {1} +$$ + +where $\pmb{x}$ is the original image, $\pmb{y}$ and $\hat{\pmb{y}}$ are the latent code and quantized latent code extracted from $\pmb{x}$ . $\hat{\pmb{x}}_1$ is the cognition-oriented compressed image. Furthermore, based on the existing structure, we further attain variable rate-cognition by regulating quantization degree using a channel-controllable latent gain unit [20], this module is also trainable and compatible with the cognition-oriented loss. + +In stage II, we implement the reconstruction of distortion-oriented compressed images to meet the requirements of human vision. Freezing the codec of the primary branch, we explore to generate distortion-oriented compressed images $\hat{\pmb{x}}_2$ in this stage, which is defined as $\hat{\pmb{x}}_2 = \hat{\pmb{x}}_1 + \pmb{r}$ : the residual $\pmb{r}$ between the cognition-oriented and distortion-oriented compressed images is expected to extract from $\hat{\pmb{y}}$ at the decoder side. However, since $\hat{\pmb{y}}$ is optimized for cognition in stage I, it didn't contain enough distortion-oriented supplemented information. Therefore, we construct an auxiliary branch to enrich the residual $\pmb{r}$ for the high-quality distortion-oriented compressed image $\hat{\pmb{x}}_2$ . As shown in Fig. 3, we first encode the quantization error $(\pmb{y} - \hat{\pmb{y}})$ at the encoder side for transmission. After a scalable transform with variable bitstream, the decoded error $(\widehat{\pmb{y} - \hat{\pmb{y}}})$ is then added to the $\hat{\pmb{y}}$ to get $\hat{\pmb{y}}_2$ , and a lightweight residual reconstruction network is used to refine residual $\pmb{r}$ . Interpolating $\pmb{r}$ with $\hat{\pmb{x}}_1$ using a weight $\beta$ gets $\bar{\pmb{x}} = \hat{\pmb{x}}_1 + (1 - \beta)\pmb{r}$ , which allows for different distortion-cognition performance. When $\beta$ equals 0, $\bar{\pmb{x}} = \hat{\pmb{x}}_2$ the lowest distortion is achieved; conversely, when $\beta$ equals 1, $\bar{\pmb{x}} = \hat{\pmb{x}}_1$ the highest cognition performance is attained. + +# 3.2 Primary Branch: Cognition-oriented Variable-bitrate Codec + +A common use of learned image compression (LIC) model [17,45,48,53], can be formulated as the following equation: + +$$ +\text {E n c o d i n g}: \boldsymbol {y} = g _ {a} (\boldsymbol {x}; \phi_ {g _ {a}}); \quad \hat {\boldsymbol {y}} = Q (\boldsymbol {y}) +$$ + +$$ +\text {D e c o d i n g :} \hat {\boldsymbol {x}} = g _ {s} (\hat {\boldsymbol {y}}; \phi_ {g _ {s}}) \tag {2} +$$ + +$$ +\text {H y p e r - p r i o r :} \boldsymbol {z} = h _ {a} \left(\boldsymbol {y}; \phi_ {h _ {a}}\right); \quad \hat {\boldsymbol {z}} = Q (\boldsymbol {z}); \quad \boldsymbol {\mu} _ {y}, \boldsymbol {\sigma} _ {y} = h _ {s} \left(\hat {\boldsymbol {z}}; \phi_ {h _ {s}}\right) +$$ + +where $\phi_{g_a},\phi_{g_s},\phi_{h_a}$ , and $\phi_{h_s}$ are the parameters of the encoder network $g_{a}$ , the decoder network $g_{s}$ , the hyper-encoder network $h_a$ , and the hyper-decoder network $h_s$ , respectively. $Q$ represents the quantization which is approximated by a noise $\mathcal{U}\left(-\frac{1}{2},\frac{1}{2}\right)$ during training, and is a rounding operator during inference. For inference, an image is first input to $g_{a}$ to extract a compact latent code $\pmb{y}$ , and $\pmb{y}$ is quantized to obtain $\hat{\pmb{y}}$ by quantization operator $Q$ . On the decoder side, $\hat{\pmb{y}}$ is input into $g_{s}$ for reconstruction to obtain the reconstructed image $\hat{\pmb{x}}$ . Additionally, the hyper path $(h_a$ and $h_s)$ is used to help estimate the probability distribution of $\pmb{y}$ , which is assumed here to be a gaussian distribution with mean $\mu_y$ and variance $\sigma_y$ . + +As discussed before, learned image compression typically requires learning a new model for different bitrate points. Therefore, in this paper, we employ a variable bitrate approach to achieve compression at different bitrates through a single model. Inspired by [20], we scale $\mathbf{y}$ and $\mathbf{z}$ using $N$ gain unit $G_{y,n}$ and $G_{z,n}$ , which is a matrix containing $C$ factors, where $C$ is the channel dimension of $\mathbf{y}$ , and $n \in [1,N]$ . By scaling different channels, we can adjust the precision of the data we need to transmit, further achieving transmission at $N$ different bitrates. Furthermore, we can interpolate the scaling factors for other bitrate points from these $N$ gain units, allowing us to switch between different bitrates using a single model. Therefore, Equation 2 can be re-formulated as: + +$$ +\text {E n c o d i n g}: \boldsymbol {y} = g _ {a} (\boldsymbol {x}; \phi_ {g _ {a}}); \quad \hat {\boldsymbol {y}} = Q (\boldsymbol {y} \cdot \boldsymbol {G} _ {y, n}) / \boldsymbol {G} _ {y, n} +$$ + +$$ +\text {D e c o d i n g}: \hat {\boldsymbol {x}} = g _ {s} (\hat {\boldsymbol {y}}; \phi_ {g _ {s}}) +$$ + +$$ +\text {H y p e r - p r i o r :} z = h _ {a} \left(\boldsymbol {y} \cdot \boldsymbol {G} _ {y, n}; \phi_ {h _ {a}}\right); \quad \hat {\boldsymbol {z}} = Q (\boldsymbol {z} \cdot \boldsymbol {G} _ {z, n}) / \boldsymbol {G} _ {z, n}; \tag {3} +$$ + +$$ +\boldsymbol {\mu} _ {y}, \sigma_ {y} = h _ {s} (\hat {z}; \phi_ {h _ {s}}) +$$ + +For conventional distortion-oriented compression tasks, the loss can be represented as: + +$$ +\begin{array}{l} \mathcal {L} = \mathcal {R} (\hat {\boldsymbol {y}}) + \mathcal {R} (\hat {\boldsymbol {z}}) + \lambda_ {n} \cdot \mathcal {D} (\boldsymbol {x}, \hat {\boldsymbol {x}}) \\ = \mathbb {E} \left[ - \log_ {2} \left(p _ {\hat {\boldsymbol {y}} | \hat {\boldsymbol {z}}} (\hat {\boldsymbol {y}} \mid \hat {\boldsymbol {z}})\right) \right] + \mathbb {E} \left[ - \log_ {2} \left(p _ {\hat {\boldsymbol {z}} | \psi} (\hat {\boldsymbol {z}} \mid \psi)\right) \right] + \lambda_ {n} \cdot \mathcal {D} (\boldsymbol {x}, \hat {\boldsymbol {x}}) \tag {4} \\ \end{array} +$$ + +where the parameter $\lambda_{n}$ regulates the balance between rate and distortion, with varying $\lambda_{n}$ values corresponding to different bit rates. $\psi$ is a factorized density model used to encode quantized $\hat{z}$ . During training, we randomly select one of the $N$ bitrate points for each forward pass. The distortion term, denoted by $\mathcal{D}(\pmb{x},\hat{\pmb{x}})$ , is measured using the Mean Squared Error (MSE) loss, GAN loss or MS-SSIM loss. Additionally, $\mathcal{R}(\hat{\pmb{y}})$ and $\mathcal{R}(\hat{z})$ represent the bit rates for the latent representations $\hat{\pmb{y}}$ and $\hat{\pmb{z}}$ , respectively. To improve the performance of + +the reconstructed images for various downstream tasks, here we use cognition-oriented loss as a substitute for the original MSE loss [17, 45, 48]. The first part of the cognition-oriented loss is a contrastive loss computed using a pretrained MoCo [29]. It is defined as follows: + +$$ +\mathcal {C} = - \log \frac {\exp \left(q \cdot k _ {+} / \tau\right)}{\sum_ {i = 0} ^ {K} \exp \left(q \cdot k _ {i} / \tau\right)} \tag {5} +$$ + +where $q$ denotes the query vector derived from the current image, $k_{+}$ is a positive key vector generated from a different augmentation of the same image, and $k_{i}$ are the negative key vectors from different images. The denominator aggregates the exponentiated similarities of the query with both the positive and K negative keys, normalized by a temperature parameter $\tau$ , which adjusts the sharpness of the distribution. This loss function encourages the model to produce embeddings where positive pairs are closer in the feature space than any negative pair, thus learning robust and discriminative feature representations. However, unlike previous work with contrastive loss, which typically aims to learn a better backbone [29] or omni-features through contrastive learning [25], our objective here is to update the codec's parameters to enable the generation of cognition-oriented compressed images. Therefore, we impose constraints on the output values of the codec, to align the range of normal image values and avoid outliers. Here, we compute a local Mean Squared Error (MSE) Loss for values lying outside the [0, 1] range. It can be formulated as: + +$$ +\mathcal {D} _ {\text {l o c a l}} (\boldsymbol {x}, \hat {\boldsymbol {x}}) = \frac {1}{M} \sum_ {i = 1} ^ {N} \left[ \mathbb {I} \left(\hat {x} _ {i} < 0 \vee \hat {x} _ {i} > 1\right) \cdot \left(\hat {x} _ {i} - x _ {i}\right) ^ {2} \right] \tag {6} +$$ + +Here, $\mathbb{I}$ is an indicator function that takes the value of 1 when the predicted value $\hat{x}_i$ falls outside the interval [0,1], and 0 otherwise. $x_{i}$ and $\hat{x}_i$ represent the target and predicted values for the $i$ th pixel, respectively, while $M$ denotes the number of pixels. $\mathcal{D}_{\mathrm{local}}$ is used as a penalty term to constrain the output range. + +As illustrated in Fig. 5, the image distributions of both the original image and the distortion-oriented compressed image primarily fall within the range of 0 to 1. However, if we directly optimize using the contrastive loss, the generated representations would reside within a significantly broader interval (approximately [-4, 6]), which does not align with the inherent characteristics of images. Moreover, directly clipping the generated representations to the [0, 1] range during training before passing them to subsequent networks can lead to the issue of vanishing gradients, resulting in a distribution of representations as shown in Fig. 5(d). Clipping values outside the [0, 1] range during inference would lead to substantial information loss. Employing a local MSE loss as a penalty term can effectively mitigate this problem (see Fig. 5(e)). + +Therefore, the overall loss of stage I in the main branch is: + +$$ +\mathcal {L} _ {I} = \mathcal {R} (\hat {\boldsymbol {y}}) + \mathcal {R} (\hat {\boldsymbol {z}}) + \lambda_ {n} \cdot \mathcal {C} + \lambda_ {\text {l o c a l}} \cdot \mathcal {D} _ {\text {l o c a l}} (\boldsymbol {x}, \hat {\boldsymbol {x}}) \tag {7} +$$ + +where the parameter $\lambda_{\mathrm{local}}$ serves to modulate the magnitude of influence exerted by the penalty term, denoted as $\mathcal{D}_{\mathrm{local}}$ . It is worth noting that the primary function of $\mathcal{D}_{\mathrm{local}}$ loss within this context is to enforce constraints on the output range. Consequently, we empirically assign a relatively small fixed value of $10^{-5}$ to $\lambda_{\mathrm{local}}$ , acknowledging its role in merely constraining the output range without significantly impacting the primary loss components. + +![](images/c8b69bab070e5860cef170617247b407042de9f3735d8050d26793ca8bcf4d11.jpg) +(a) +Fig. 5: The histograms for pixel values of: (a) the original image. (b) distortion-oriented compressed image. (c) cognition-oriented compressed image without using local mse loss. (d) cognition-oriented compressed images processed with clipping to the range $(0, 1)$ during training. (e) cognition-oriented compressed images with using local mse loss. + +![](images/b1b0b7df87134551e4e5d9ccd76c0061eba0e412e87acd10037a683c3837308b.jpg) +(b) + +![](images/2afaa8e768fc42554121d800167e9974c906672ec848cc7c44b51c2ff58c436b.jpg) +(c) + +![](images/855c5dc16ef4eed1873e5659061c147333487f6cb24483e4e05297ac366e8fa7.jpg) +(d) + +![](images/636fbb64012d196ee79e65678ce2ea5af79df9fa66a39f777d1f99d46622ab81.jpg) +(e) + +Table 1: Parameter numbers of different Codec modules. + +
Base CodecScalable TransformResidual Reconstruction
Parameters(/M)20.390.191.20
+ +# 3.3 Auxiliary Branch: Distortion-oriented Residual Supplement + +To further meet the requirements of human vision, we employ an auxiliary branch to supplement residual information with a scalable bitstream. As shown in Fig. 3, we construct a residual reconstruction network to generate the residual $\boldsymbol{r}$ . By adding $\boldsymbol{r}$ to $\hat{\boldsymbol{x}}_1$ , we reconstruct an distortion-oriented compressed image $\hat{\boldsymbol{x}}_2$ . The residual reconstruction network comprises multiple layers of transposed convolutions and point-wise convolutional layers. The former is to increase the size of the representation, ensuring spatial dimension alignment with the original decoder's representation. The latter is employed to reduce and align the channel dimensions, simultaneously decreasing the model parameters and computational complexity. Additionally, the intermediate layer outputs $f_1, f_2$ , and $f_3$ from the original decoder serve as prior information. They are fed into the residual reconstruction network through a point-wise convolution to help recovery. + +We attempt to input $\hat{\pmb{y}}$ to the residual reconstruction network directly to generate $\pmb{r}$ . However, directly using $\hat{\pmb{y}}$ for reconstruction does not work, because $\hat{\pmb{y}}$ of the primary branch is optimized for cognitive perception rather than distortion minimization. To prove that, we analyze the quantization operation: as shown in Equation 3, this process involves the multiplication of $\pmb{y}$ by $G_{y,n}$ followed by quantization. $G_{y,n}$ plays a crucial role in determining the precision of quantization across different channels and, i.e., the extent of quantization error. Given that $G_{y,n}$ is optimized for cognitive perception rather than distortion minimization in stage I, thus the reconstructed representations $\hat{\pmb{y}}$ cannot recover the detailed residual. The gradient visualization in Fig. 6(a)(b) reveals the disparities in channel contributions between cognition-oriented and distortion-oriented images. By backpropagating the mean gradients $\frac{\partial\hat{\pmb{x}}}{\partial\hat{\pmb{y}}}$ from the generated image into $\hat{\pmb{y}}$ ’s channels, we observe a marked difference in how $\hat{\pmb{y}}$ ’s channels influence the generation of cognition-oriented versus distortion-oriented compressed images. + +To address this issue, we propose to transmit this missed information through a scalable bitstream. Specifically, by using the similar VBR method in Sec. 3.2, we first encode the quantization-induced error $(\pmb{y} - \hat{\pmb{y}})$ into a bitstream via a transformation through three convolutional layers for transmission to the de + +![](images/63d200379813159f62c0e3ea3b0d220a23f6f4e31d136dfeb85f5c3ce3fac2bd.jpg) +Fig. 6: (a) The gradient for $\hat{\pmb{y}}$ from cognition-orient image $\hat{\pmb{x}}_1$ (b) The gradient for $\hat{\pmb{y}}$ from distortion-orient image $\hat{\pmb{x}}_2$ (c) The energy of cognition-orient $\hat{\pmb{y}}$ and distortion-orient $\hat{\pmb{y}}_2$ (Purple represents the overlapping part of the energy between $\hat{\pmb{y}}$ and $\hat{\pmb{y}}_2$ ). + +![](images/cae51e7aaa09f80faece02c3ceda740f2d49544f330bf3e14365d0e99bba1e58.jpg) + +![](images/b4b36493df5b6e4f65a28d2653e956179afbf2273cf3cfa18b103abbc152454d.jpg) + +coder side. An auxiliary gain unit is used to control the precision and bitrate of the bitstream. Then, we reconstruct bitstream through another three convolutional layers, and the output is added to $\hat{\pmb{y}}$ to produce $\hat{\pmb{y}}_2$ , which is fed into the residual reconstruction network to generate $\pmb{r}$ . The scalable transform and the residual reconstruction network together form the auxiliary branch, which exclusively costs fewer parameters compared to the base codec (see Table 1). + +Moreover, we also analyze the differences in energy across various channels between $\hat{\pmb{y}}_2$ and the original $\hat{\pmb{y}}$ in Fig. 6(c). It is observable that the energy levels across most channels are similar, with only a few channels significantly increased. This suggests that we only need to transmit very few bits to supplement such residuals to enhance decompression quality, which will be further proved in the experiments. Therefore, the overall loss of stage II can be formulated as follows: + +$$ +\mathcal {L} _ {I I} = \mathcal {R} (\hat {\boldsymbol {s}}) + \lambda_ {m} \cdot \mathcal {D} (\boldsymbol {x}, \hat {\boldsymbol {x}} _ {2}) \tag {8} +$$ + +where $\mathcal{R}(\hat{\pmb{s}})$ represents the bitrate of the scalable bitstream, $\mathcal{D}$ is defined as MSE loss, $\lambda_{m}$ is used to control rate-distortion, $m\in [1,N]$ , and $\hat{\pmb{x}}_2 = \hat{\pmb{x}}_1 + \pmb{r}$ . + +# 3.4 Rate-Distortion-Cognition Trade-off + +Through the aforementioned procedures, we are able to achieve two distinct trade-offs: rate-distortion and rate-cognition. To achieve ultra-controllability, by employing interpolation, we can adjust the trade-off between distortion and cognition, thereby enabling the controllable rate-distortion-cognition. The equation is as follows: + +$$ +\bar {\boldsymbol {x}} = \hat {\boldsymbol {x}} _ {1} + (1 - \beta) \times \hat {\boldsymbol {r}} \tag {9} +$$ + +where $\beta \in [0,1]$ is used to adjust the balance between cognition and distortion. As $\beta$ approaches 1, the distortion of $\bar{\pmb{x}}$ increases, and the cognition performance of $\bar{\pmb{x}}$ is enhanced. Conversely, as $\beta$ approaches 0, the distortion decreases, and the cognition performance diminishes. When transferring to downstream tasks, $\beta$ is set to 1, the codec weights are kept frozen, and we only use cognition-oriented compressed images to finetune the network for the downstream tasks. + +# 4 Experiments + +# 4.1 Experimental Setting + +**Codec.** We set TIC [48] as our base codec. During training stage I, at each step, we randomly select a number from the set $[0.0625, 0.125, 0.25, 0.5, 1, 2]$ as $\lambda_{n}$ . Correspondingly, the latent code $\pmb{y}$ is multiplied by the associated gain unit to + +implement variable bit rates. For convenience, in the subsequent sections, we use $\alpha \in [0,1]$ to represent the selected point of the bitrate. When $\alpha = 0$ , the bitrate is at its minimum, corresponding to the model with $\lambda_{n} = 0.0625$ . Conversely, when $\alpha = 1$ , the bitrate is at its maximum, corresponding to the model with $\lambda_{n} = 2$ . For training stage II, we set $\lambda_{m} \in [0.00018, 0.00036, 0.00072, 0.001, 0.0015, 0.002]$ . To assess the rate-distortion performance of the reconstructed image, we employ the Peak Signal-to-Noise Ratio (PSNR) to evaluate the extent of distortion and use bits per pixel (bpp) to measure the compression ratio. Bpp is formulated as: as $\frac{b}{h \times w}$ , where $h, w$ are the height and width of the source image, $b$ denotes the total bits cost of the coded feature bit-stream. + +Image Classification. we train and validate our model on ImageNet-1k datasets. During the evaluation, before inputting the images into the model, we resize and crop them to 224x224. Meanwhile, ResNet50 [30], serving as the task network, is used to assess the accuracy of classification. + +Object Detection. For the object detection task, we train and validate on the MS COCO2017 train/val dataset. Following the setting in [67], we employ the faster-rcnn-FPN network [57] for object detection. The image scale ranges from [640, 800] pixels during training and is set to 800 at inference by default. + +Semantic Segmentation. For this task, we employ DeepLab V3 [12], a network designed for segmentation tasks. The training and evaluation are conducted on the Cityscapes dataset. The Cityscapes dataset is specifically designed for urban street scenes. It includes 5,000 images with fine annotations. Cross entropy serves as the loss function during the training process. To assess effectiveness, we use the mean intersection over union (mIoU) as the metric. + +Baseline. For evaluating the effectiveness of our method on machine vision tasks, we have selected several different baselines for comparison with our approach. Initially, the traditional image standard VVC (VTM-12.0) [9] was utilized for testing. For learned image compression, we employ TIC [48] to assess its performance across three tasks. Regarding the ICM method, we chose to compare with recent works, namely TransTIC [14] and T-TF [11]. To evaluate the rate-distortion performance of our model, we conducted comparisons with methods such as VTM, TIC [48], Factorized [6] and MBT-Mean [53]. + +# 4.2 Quantitative Comparison + +As illustrated in Fig. 7 and 8, we compared our method against recent works. It is first noteworthy that our approach is controllable, by adjusting the values of $\alpha$ and $\beta$ , we can achieve the ultra-controllability within the blue area depicted in the figures using a single codec. The bitrate increases with an increase in $\alpha$ , while the performance for machine vision tasks and image reconstruction tasks respectively increases and decreases with an increase in $\beta$ . As demonstrated by points ABCD in the Fig. 7 and 8, at $(\beta = 1, \alpha = 1)$ (Point A), we achieve the highest bitrate and the optimal machine vision task performance at this bitrate; at $(\beta = 0, \alpha = 1)$ (Point B), we obtain the highest bitrate and the lowest distortion at this bitrate; at $(\beta = 0, \alpha = 0)$ (Point C), we achieve the lowest bitrate and the lowest distortion at this bitrate; at $(\beta = 1, \alpha = 0)$ (Point D), we + +![](images/be94c521781b3bcd2e068f4459b85c0f39af3881c638a0e92a01576c3b2e2c2f.jpg) +(a) Classification on ImageNet + +![](images/47db714b07a6b29d70a24124a4a338f9192973c08220a86bce643ea74561cfc9.jpg) +(b) Segmentation on CityScapes + +![](images/37e9888803d7edf2c50c0c3103b6528d4df667027efe69729b1de56140826a0a.jpg) +(c) Detection on COCO + +![](images/a4917e750a7e8f85e0eb6b09a83af5043ef703060a7821f30dc1e6140e004d67.jpg) +Fig. 7: The rate-machine vision task performance curve ( $\alpha$ changes from 0 to 1 from left to right, and $\beta$ changes from 0 to 1 from bottom to top). +(a) ImageNet +Fig. 8: The rate-PSNR performance curve ( $\alpha$ changes from 0 to 1 from left to right, and $\beta$ changes from 0 to 1 from top to bottom). + +![](images/d096978952a45d7da0f12fc226af0c00f76fbbc6c2a46dfa54f1a5e926037cb6.jpg) +(b) CityScapes + +![](images/5c2eb3b5a357b8a39aedb91241ff74c62babb7fc1cae27a37e1ed4b8b7d52d16.jpg) +(c) COCO + +achieve the lowest bitrate and the best machine vision task performance at this bitrate. + +As shown in Fig. 7 and Table 2, we compared our method $(\beta = 1)$ with recent approaches in three tasks: classification, segmentation, and detection. It is evident that, compared with the un-tuned learned image compression method TIC [48] and the traditional image compression standard VVC (VTM-12.0) [9], our method significantly improves machine vision tasks performance, especially at lower bitrates. This demonstrates our model's superior cognition capabilities. Furthermore, compared with some ICM methods, TransTIC [14], T-FT [11] + TIC [48], our approach not only achieves comparable or even superior performance but also provides controllability. + +Fig. 8 and Table 3 illustrate the performance of our method $(\beta = 0)$ in the image reconstruction task. It is observed that compared with TIC, which is optimized for distortion, our method, being primarily optimized for cognition, leading to a reduction in PSNR. However, it still manages to achieve results comparable to the recent learned image compression methods, such as MBT-Mean [53]. Additionally, as demonstrated in Fig. 8(a), compared with the ICM method TransTIC, our method can achieve a higher upper bound for PSNR. + +# 4.3 Ablation Study + +In our study, we conducted ablation studies based on the base codec TIC, as shown in Table 4. For the measurements of BD-Acc [7] and BD-PSNR, we set $\beta = 1$ and $\beta = 0$ , respectively. We explored the application of the VBR method from [20] to TIC, which allows for varying bitrates with a single model. Directly + +Table 2: Comparison of machine vision Table 3: Comparison of image reconstruction task performance. + +
Cls.Det.Seg.
BD-AccBD-APBD-mIoU|Cls.|Det.|Seg.
|BD-PSNR||BD-PSNR||BD-PSNR|
TIC [48]0.6908-0.4421-0.0208TIC [48]0.2352-0.88760.0315
TransTIC [14]8.92181.81370.1128MBT-Mean [53]-0.8492-2.1432-0.9654
T-FT TIC [11]8.34590.70910.1049Factorized [6]-1.5387-2.3248-3.4599
Ours (β=1)9.98761.82590.1133Ours (β=0)-0.4102-1.8293-1.0240
+ +Table 4: BD-PSNR and BD-Acc on ImageNet with TIC as anchor. Codec Number accounts for required codecs across $K$ bitrates, $M$ downstream tasks, and $T$ distortion-cognition scenarios. + +
MethodsTIC [48]TIC + VBR [20]TIC + Cognition-oriented lossOurs (w/o scalable bitstream)Ours
BD-PSNR--0.1453-1.5421-1.4640-0.9104
BD-Acc--0.573911.125410.812410.6045
Codec NumbersK × M × TM × TK × T11
+ +transferring VBR methods to machine vision tasks does not yield satisfactory results. At the same time, for other distinct scenarios (different tasks/networks and cognition-distortion trade-off), codec retraining remains requisite. Using the proposed cognition-oriented loss can significantly improve the performance of machine vision tasks. However, for different bitrates and cognition-distortion trade-offs, many codecs are still required. In contrast, our method requires only one codec to be applicable across all scenarios. For our method, when we do not transmit an auxiliary scalable bitstream and directly use $\hat{\pmb{y}}$ for the training of residual reconstruction, there is a significant improvement in BD-Acc, but the distortion-oriented compressed images exhibit heavy distortion. As analyzed in Section 3.3 and shown in Fig. 9 and Table 4, we discovered that a minimal scalable bitstream can significantly reduce reconstruction distortion, with BD-PSNR improving by approximately $37.81\%$ , and the loss in BD-Acc being only around $2\%$ . + +# 4.4 Visualization + +In Fig. 10, we present a visualization of images generated with varying $\beta$ values, along with their spectrum. It is observable that distortion-oriented compressed images (i.e., $\beta = 0$ ) lose a significant amount of high-frequency information compared to the original image. However, as the $\beta$ value increases, the high-frequency information in the images progressively increases. This indicates that cognition-oriented compressed images contain more high-frequency information, making them more conducive for analysis in downstream tasks. For instance, compared to distortion-oriented compressed images, cognition-oriented compressed images depict more detailed wrinkles on the ball held in the statue's hand. On the other hand, cognition-oriented compressed images, compared to both distortion-oriented compressed images and the original, possess a certain degree of color deviation. This suggests to some extent that color restoration is not crucial for some downstream tasks. + +![](images/b9dc8066341995bd9830dd0513cd06347ba476a8924a2d63ae96706230d75e23.jpg) +(a) Rate-Acc. + +![](images/e5ea83a98108f949b858e663ec3432a50df93416405e10986e772a2f30695468.jpg) +(b) Rate-PSNR. + +![](images/c88ad9533cb57f15776fcc4e864175ad9d31a075ade933cef2d6096d538e47f0.jpg) +Fig. 9: The ablation studies. +Fig. 10: Visualization results and spectrum with different $\beta$ values. (a) represents the original image, while (b) to (f) represent the compressed images with $\beta$ values of 0, 0.25, 0.5, 0.75, and 1.0, respectively, (g) represents the magnification of the corresponding colored frame area. + +# 5 Conclusion + +In this paper, we propose a Rate-Distortion-Cognition Controllable Versatile Neural Image Compression model. We introduce a cognition-oriented loss to enhance the machine vision support, and achieve variable-rate by latent channel regulation. To further meet human eye perception, we build an auxiliary branch with a very small and scalable bitrate, decoding distortion-oriented compressed images. Finally, with an interpolation strategy to merge two branches, we also achieve distortion-cognition controlling, thereby realizing controllable rate-distortion-cognition in a single codec. Our experiments demonstrate that our method can achieve excellent performance in classification, segmentation, and detection tasks while also achieving ultra-controllability. + +# Acknowledgments + +This work was supported in part by NSFC 62302246 and ZJNSFC under Grant LQ23F010008, and supported by High Performance Computing Center at Eastern Institute of Technology, Ningbo, and Ningbo Institute of Digital Twin. + +# References + +1. Agustsson, E., Minnen, D., Toderici, G., Mentzer, F.: Multi-realism image compression with a conditional generator. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22324-22333 (2023) 2, 3, 5 +2. Agustsson, E., Tschannen, M., Mentzer, F., Timofte, R., Gool, L.V.: Generative adversarial networks for extreme learned image compression. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 221-231 (2019) 6 +3. Akbari, M., Liang, J., Han, J.: Dsslic: Deep semantic segmentation-based layered image compression. In: ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 2042-2046. IEEE (2019) 4 +4. Babenko, A., Slesarev, A., Chigorin, A., Lempitsky, V.: Neural codes for image retrieval. In: ECCV. pp. 584-599. Springer (2014) 4 +5. Badrinarayanan, V., Kendall, A., Cipolla, R.: Segnet: A deep convolutional encoder-decoder architecture for image segmentation. TPAMI 39(12), 2481-2495 (2017) 4 +6. Balle, J., Laparra, V., Simoncelli, E.P.: End-to-end optimized image compression. In: ICLR (2017) 4, 11, 13 +7. Bjontegaard, G.: Calculation of average psnr differences between rd-curves. In: VCEG-M33 (2001) 12 +8. Blau, Y., Michaeli, T.: Rethinking lossy compression: The rate-distortion-perception tradeoff. In: International Conference on Machine Learning. pp. 675-685. PMLR (2019) 5 +9. Bross, B., Wang, Y.K., Ye, Y., Liu, S., Chen, J., Sullivan, G.J., Ohm, J.R.: Overview of the versatile video coding (vvc) standard and its applications. TCSVT (2021) 1, 4, 11, 12 +0. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021) 3 +1. Chamain, L.D., Racapé, F., Bégaint, J., Pushparaja, A., Feltman, S.: End-to-end optimized image compression for machines, a study. In: 2021 Data Compression Conference (DCC). pp. 163-172. IEEE (2021) 11, 12, 13 +2. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. TPAMI 40(4), 834-848 (2017) 4, 11 +3. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: ICML. pp. 1597-1607. PMLR (2020) 3 +4. Chen, Y.H., Weng, Y.C., Kao, C.H., Chien, C., Chiu, W.C., Peng, W.H.: Transtic: Transferring transformer-based image compression from human perception to machine perception. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 23297-23307 (2023) 2, 4, 5, 11, 12, 13 +5. Chen, Z., Fan, K., Wang, S., Duan, L.Y., Lin, W., Kot, A.: Lossy intermediate deep learning feature compression and evaluation. In: ACM MM. pp. 2414-2422 (2019) 4 +6. Chen, Z., Fan, K., Wang, S., Duan, L., Lin, W., Kot, A.C.: Toward intelligent sensing: Intermediate deep feature compression. TIP 29, 2230-2243 (2019) 4 + +17. Cheng, Z., Sun, H., Takeuchi, M., Katto, J.: Learned image compression with discretized gaussian mixture likelihoods and attention modules. In: CVPR. pp. 7939-7948 (2020) 1, 4, 7, 8 +18. Choi, H., Bajic, I.V.: Scalable image coding for humans and machines. IEEE Transactions on Image Processing 31, 2739-2754 (2022) 2 +19. Choi, Y., El-Khamy, M., Lee, J.: Variable rate deep image compression with a conditional autoencoder. In: ICCV. pp. 3146-3154 (2019) 4 +20. Cui, Z., Wang, J., Bai, B., Guo, T., Feng, Y.: G-vae: A continuously variable rate deep image compression framework. arXiv preprint arXiv:2003.02012 (2020) 2, 3, 4, 6, 7, 12, 13 +21. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2020) 4 +22. Duan, L.Y., Chandrasekhar, V., Chen, J., Lin, J., Wang, Z., Huang, T., Girod, B., Gao, W.: Overview of thempeg-cdv standard. TIP 25(1), 179-194 (2015) 4 +23. Duan, L.Y., Lou, Y., Bai, Y., Huang, T., Gao, W., Chandrasekhar, V., Lin, J., Wang, S., Kot, A.C.: Compact descriptors for video analysis: The emerging mpeg standard. IEEE MultiMedia 26(2), 44-54 (2018) 4 +24. Duan, L., Liu, J., Yang, W., Huang, T., Gao, W.: Video coding for machines: A paradigm of collaborative compression and intelligent analytics. TIP 29, 8680-8695 (2020) 2 +25. Feng, R., Jin, X., Guo, Z., Feng, R., Gao, Y., He, T., Zhang, Z., Sun, S., Chen, Z.: Image coding for machines with omnipotent feature learning. In: ECCV. pp. 510-528. Springer (2022) 2, 4, 6, 8 +26. Feng, R., Liu, J., Jin, X., Pan, X., Sun, H., Chen, Z.: Prompt-icm: A unified framework towards image coding for machines with task-driven prompts. arXiv preprint arXiv:2305.02578 (2023) 4 +27. Gao, W., Ma, S., Duan, L., Tian, Y., Xing, P., Wang, Y., Wang, S., Jia, H., Huang, T.: Digital retina: A way to make the city brain more efficient by visual coding. IEEE Transactions on Circuits and Systems for Video Technology 31(11), 4147-4161 (2021) 1 +28. Han, K., Xiao, A., Wu, E., Guo, J., Xu, C., Wang, Y.: Transformer in transformer. NeurIPS 34, 15908-15919 (2021) 4 +29. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: CVPR. pp. 9729-9738 (2020) 3, 6, 8 +30. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016) 11 +31. He, T., Sun, S., Guo, Z., Chen, Z.: Beyond coding: Detection-driven image compression with semantically structured bit-stream. In: 2019 Picture Coding Symposium (PCS). pp. 1-5. IEEE (2019) 2 +32. Hu, Y., Yang, J., Chen, L., Li, K., Sima, C., Zhu, X., Chai, S., Du, S., Lin, T., Wang, W., et al.: Planning-oriented autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17853-17862 (2023) 1 +33. Hu, Y., Yang, S., Yang, W., Duan, L.Y., Liu, J.: Towards coding for human and machine vision: A scalable image coding approach. In: 2020 IEEE International Conference on Multimedia and Expo (ICME). pp. 1-6. IEEE (2020) 4 +34. Iwai, S., Miyazaki, T., Omachi, S.: Controlling rate, distortion, and realism: Towards a single comprehensive neural image compression model. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2900-2909 (2024) 5 + +35. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. ECCV (2022) 4 +36. Jin, X., Feng, R., Sun, S., Feng, R., He, T., Chen, Z.: Semantical video coding: Instill static-dynamic clues into structured bitstream for ai tasks. Journal of Visual Communication and Image Representation 93, 103816 (2023) 2 +37. Johnston, N., Vincent, D., Minnen, D., Covell, M., Singh, S., Chinen, T., Hwang, S.J., Shor, J., Toderici, G.: Improved lossy image compression with priming and spatially adaptive bit rates for recurrent networks. In: CVPR. pp. 4385-4393 (2018) 4 +38. Körber, N., Kromer, E., Siebert, A., Hauke, S., Mueller-Gritschneider, D.: Egic: Enhanced low-bit-rate generative image compression guided by semantic segmentation. arXiv preprint arXiv:2309.03244 (2023) 3, 5 +39. Le, N., Zhang, H., Cricri, F., Ghaznavi-Youvalari, R., Rahtu, E.: Image coding for machines: An end-to-end learned approach. In: ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1590-1594. IEEE (2021) 4 +40. Li, H., Li, S., Dai, W., Li, C., Zou, J., Xiong, H.: Frequency-aware transformer for learned image compression. arXiv preprint arXiv:2310.16387 (2023) 1 +41. Li, H., Li, S., Ding, S., Dai, W., Cao, M., Li, C., Zou, J., Xiong, H.: Image compression for machine and human vision with spatial-frequency adaptation. In: ECCV. Springer (2024) 2 +42. Li, X., Shi, J., Chen, Z.: Task-driven semantic coding via reinforcement learning. TIP (2021) 4 +43. Li, Y., Mao, H., Girshick, R., He, K.: Exploring plain vision transformer backbones for object detection. ECCV (2022) 4 +44. Liu, J., Sun, H., Katto, J.: Semantic segmentation in learned compressed domain. In: 2022 Picture Coding Symposium (PCS). pp. 181-185. IEEE (2022) 4 +45. Liu, J., Sun, H., Katto, J.: Learned image compression with mixed transformer-cnn architectures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14388-14397 (2023) 1, 2, 4, 7, 8 +46. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: CVPR. pp. 3431-3440 (2015) 4 +47. Lu, G., Ge, X., Zhong, T., Geng, J., Hu, Q.: Preprocessing enhanced image compression for machine vision. arXiv preprint arXiv:2206.05650 (2022) 2 +48. Lu, M., Guo, P., Shi, H., Cao, C., Ma, Z.: Transformer-based image compression. arXiv preprint arXiv:2111.06707 (2021) 1, 7, 8, 10, 11, 12, 13 +49. Ma, H., Liu, D., Yan, N., Li, H., Wu, F.: End-to-end optimized versatile image compression with wavelet-like transform. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(3), 1247-1263 (2020) 2 +50. Ma, S., Zhang, X., Wang, S., Zhang, X., Jia, C., Wang, S.: Joint feature and texture coding: Toward smart video representation via front-end intelligence. TCSVT 29(10), 3095-3105 (2018) 4 +51. Mentzer, F., Agustsson, E., Tschannen, M., Timofte, R., Van Gool, L.: Conditional probability models for deep image compression. In: CVPR. pp. 4394-4402 (2018) 4 +52. Mentzer, F., Toderici, G., Tschannen, M., Agustsson, E.: High-fidelity generative image compression. arXiv preprint arXiv:2006.09965 (2020) 4, 5 +53. Minnen, D., Balle, J., Toderici, G.: Joint autoregressive and hierarchical priors for learned image compression. In: NeurIPS (2018) 4, 7, 11, 12, 13 +54. Rabbani, M., Joshi, R.: An overview of the jpeg 2000 still image compression standard. Signal processing: Image communication 17(1), 3-48 (2002) 4 + +55. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified, real-time object detection. In: CVPR. pp. 779-788 (2016) 4 +56. Redmon, J., Farhadi, A.: Yolo9000: better, faster, stronger. In: CVPR. pp. 7263-7271 (2017) 4 +57. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS 28, 91-99 (2015) 4, 11 +58. Singh, S., Abu-El-Haija, S., Johnston, N., Balle, J., Shrivastava, A., Toderici, G.: End-to-end learning of compressible features. In: 2020 IEEE International Conference on Image Processing (ICIP). pp. 3349-3353. IEEE (2020) 4 +59. Song, M., Choi, J., Han, B.: Variable-rate deep image compression through spatially-adaptive feature transform. In: ICCV. pp. 2380-2389 (2021) 2, 4 +60. Sullivan, G.J., Ohm, J.R., Han, W.J., Wiegand, T.: Overview of the high efficiency video coding (hevc) standard. TCSVT 22(12), 1649-1668 (2012) 1, 4 +61. Sun, S., He, T., Chen, Z.: Semantic structured image coding framework for multiple intelligent applications. TCSVT (2020) 2, 4 +62. Terhörst, P., Ihlefeld, M., Huber, M., Damer, N., Kirchbuchner, F., Raja, K., Kuijper, A.: Qmagface: Simple and accurate quality-aware face recognition. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3484-3494 (2023) 1 +63. Toderici, G., Vincent, D., Johnston, N., Jin Hwang, S., Minnen, D., Shor, J., Covell, M.: Full resolution image compression with recurrent neural networks. In: CVPR. pp. 5306-5314 (2017) 4 +64. Wallace, G.K.: Thejpeg still picture compression standard. IEEE transactions on consumer electronics 38(1), xviii-xxxiv (1992) 1, 4 +65. Wang, S., Wang, S., Yang, W., Zhang, X., Wang, S., Ma, S., Gao, W.: Towards analysis-friendly face representation with scalable feature and texture compression. TMM (2021) 4 +66. Wiegand, T., Sullivan, G.J., Bjontegaard, G., Luthra, A.: Overview of the h. 264/avc video coding standard. TCSVT 13(7), 560-576 (2003) 4 +67. Wu, Y., Kirillov, A., Massa, F., Lo, W.Y., Girshick, R.: Detector2. https://github.com/facebookresearch/detectron2 (2019) 11 +68. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. NeurIPS 34, 12077-12090 (2021) 4 +69. Yan, F.F., Hou, F., Lu, Z.L., Hu, X., Huang, C.B.: Efficient characterization and classification of contrast sensitivity functions in aging. Scientific reports 7(1), 5045 (2017) 3 +70. Yang, F., Herranz, L., Van De Weijer, J., Guitián, J.A.I., López, A.M., Mozerov, M.G.: Variable rate deep image compression with modulated autoencoder. IEEE Signal Processing Letters 27, 331-335 (2020) 4 +71. Yang, R., Mandt, S.: Lossy image compression with conditional diffusion models. arXiv preprint arXiv:2209.06950 (2022) 5 +72. Zheng, S., Lu, J., Zhao, H., Zhu, X., Luo, Z., Wang, Y., Fu, Y., Feng, J., Xiang, T., Torr, P.H., et al.: Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In: CVPR. pp. 6881-6890 (2021) 4 \ No newline at end of file diff --git a/2024/Rate-Distortion-Cognition Controllable Versatile Neural Image Compression/images.zip b/2024/Rate-Distortion-Cognition Controllable Versatile Neural Image Compression/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a8532f9b49af51845c981d8930d802a6f7dd4424 --- /dev/null +++ b/2024/Rate-Distortion-Cognition Controllable Versatile Neural Image Compression/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b28811611174431691308bc946e506edd5381a5c876db4f24dd6015474e79f09 +size 499870 diff --git a/2024/Rate-Distortion-Cognition Controllable Versatile Neural Image Compression/layout.json b/2024/Rate-Distortion-Cognition Controllable Versatile Neural Image Compression/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9eb32ef3698ad48b26d8bd2d042d4faa553e22f3 --- /dev/null +++ b/2024/Rate-Distortion-Cognition Controllable Versatile Neural Image Compression/layout.json @@ -0,0 +1,13194 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 170, + 111, + 445, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 111, + 445, + 148 + ], + "spans": [ + { + "bbox": [ + 170, + 111, + 445, + 148 + ], + "type": "text", + "content": "Rate-Distortion-Cognition Controllable Versatile Neural Image Compression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "spans": [ + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "content": "Jinming Liu" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "content": ", Ruoyu Feng" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "content": ", Yunpeng Qi" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "content": ", Qiuyu Chen" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "content": ", Zhibo Chen" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "content": ", Wenjun Zeng" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "content": ", Xin Jin" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 136, + 201, + 477, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 201, + 477, + 248 + ], + "spans": [ + { + "bbox": [ + 136, + 201, + 477, + 248 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 201, + 477, + 248 + ], + "type": "text", + "content": " Shanghai Jiao Tong University \n" + }, + { + "bbox": [ + 136, + 201, + 477, + 248 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 136, + 201, + 477, + 248 + ], + "type": "text", + "content": " Ningbo Institute of Digital Twin, Eastern Institute of Technology, Ningbo, China \n" + }, + { + "bbox": [ + 136, + 201, + 477, + 248 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 136, + 201, + 477, + 248 + ], + "type": "text", + "content": " University of Science and Technology of China \njmliu206@sjtu.edu.cn jinxin@eitech.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 259, + 455, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 259, + 455, + 479 + ], + "spans": [ + { + "bbox": [ + 159, + 259, + 455, + 479 + ], + "type": "text", + "content": "Abstract. Recently, the field of Image Coding for Machines (ICM) has garnered heightened interest and significant advances thanks to the rapid progress of learning-based techniques for image compression and analysis. Previous studies often require training separate codes to support various bitrate levels, machine tasks, and networks, thus lacking both flexibility and practicality. To address these challenges, we propose a rate-distortion-cognition controllable versatile image compression, which method allows the users to adjust the bitrate (i.e., Rate), image reconstruction quality (i.e., Distortion), and machine task accuracy (i.e., Cognition) with a single neural model, achieving ultra-controllability. Specifically, we first introduce a cognition-oriented loss in the primary compression branch to train a codec for diverse machine tasks. This branch attains variable bitrate by regulating quantization degree through the latent code channels. To further enhance the quality of the reconstructed images, we employ an auxiliary branch to supplement residual information with a scalable bitstream. Ultimately, two branches use a " + }, + { + "bbox": [ + 159, + 259, + 455, + 479 + ], + "type": "inline_equation", + "content": "\\beta x + (1 - \\beta)y" + }, + { + "bbox": [ + 159, + 259, + 455, + 479 + ], + "type": "text", + "content": " interpolation strategy to achieve a balanced cognition-distortion trade-off. Extensive experiments demonstrate that our method yields satisfactory ICM performance and flexible Rate-Distortion-Cognition controlling." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 490, + 453, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 490, + 453, + 512 + ], + "spans": [ + { + "bbox": [ + 159, + 490, + 453, + 512 + ], + "type": "text", + "content": "Keywords: Neural Image Compression " + }, + { + "bbox": [ + 159, + 490, + 453, + 512 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 490, + 453, + 512 + ], + "type": "text", + "content": " Image Compression for Machine " + }, + { + "bbox": [ + 159, + 490, + 453, + 512 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 490, + 453, + 512 + ], + "type": "text", + "content": " Variable-bitrate Compression " + }, + { + "bbox": [ + 159, + 490, + 453, + 512 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 490, + 453, + 512 + ], + "type": "text", + "content": " Cognition-distortion Trade-off" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 520, + 230, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 230, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 230, + 533 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 539, + 482, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 539, + 482, + 610 + ], + "spans": [ + { + "bbox": [ + 130, + 539, + 482, + 610 + ], + "type": "text", + "content": "Machine vision tasks have become increasingly integral in numerous daily applications, including intelligent traffic systems [27], autonomous driving [32], and facial recognition [62]. As a result, there is a burgeoning consumption of images by machines, which poses significant challenges in the transmission and storage of such extensive data volumes. How to efficiently compress these data while promising machine task accuracy urgently needs to be studied." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 611, + 490, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 490, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 490, + 647 + ], + "type": "text", + "content": "Traditional image compression standards, such as JPEG [64], HEVC [60], VVC [9], have been widely applied in life. Meanwhile, deep learning-based CODECs have rapidly developed recently, achieving superior rate-distortion performance [17," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 654, + 243, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 243, + 666 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 243, + 666 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 332, + 189 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 332, + 189 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 332, + 189 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 332, + 189 + ], + "type": "image", + "image_path": "de05175552ea524d0413990280a9a4f964c7d763467e644ab9b13fbaaddb66ad.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 193, + 482, + 227 + ], + "lines": [ + { + "bbox": [ + 130, + 193, + 482, + 227 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 482, + 227 + ], + "type": "text", + "content": "Fig. 1: Motivation: (a) Traditional image compression tasks require training " + }, + { + "bbox": [ + 130, + 193, + 482, + 227 + ], + "type": "inline_equation", + "content": "K \\times M \\times T" + }, + { + "bbox": [ + 130, + 193, + 482, + 227 + ], + "type": "text", + "content": " codecs for " + }, + { + "bbox": [ + 130, + 193, + 482, + 227 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 193, + 482, + 227 + ], + "type": "text", + "content": " bitrate points, " + }, + { + "bbox": [ + 130, + 193, + 482, + 227 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 193, + 482, + 227 + ], + "type": "text", + "content": " machine vision tasks, and " + }, + { + "bbox": [ + 130, + 193, + 482, + 227 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 193, + 482, + 227 + ], + "type": "text", + "content": " cognition-distortion trade-offs. (b) Our method enables controllable rate-distortion-cognition with a single codec." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 341, + 125, + 481, + 189 + ], + "blocks": [ + { + "bbox": [ + 341, + 125, + 481, + 189 + ], + "lines": [ + { + "bbox": [ + 341, + 125, + 481, + 189 + ], + "spans": [ + { + "bbox": [ + 341, + 125, + 481, + 189 + ], + "type": "image", + "image_path": "5ec7eced2ac4996a40b3c7490baeaf0f93779d37b714d41f93ce177cfe4626bb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 245, + 479, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 245, + 479, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 479, + 269 + ], + "type": "text", + "content": "[40, 45, 48]. However, despite success in bit saving and visual fidelity, these methods hardly generalize well to machine tasks w.r.t cognition performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 274, + 480, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 480, + 334 + ], + "type": "text", + "content": "That's because traditional image compression CODECs are typically optimized to align with human visual perception. They employ distortion metrics, such as PSNR and MS-SSIM [45], as well as realism metrics, including KID and FID [1], to assess the loss incurred by compression procedure. Hence, utilizing these CODEcs directly for machine vision tasks may yield suboptimal outcomes [24, 47]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 338, + 482, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 554 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 554 + ], + "type": "text", + "content": "To address this issue, the field of image coding for machines (ICM) has gradually come into focus [24,31,36,41,61]. For example, some works improves the performance of downstream tasks by jointly optimizing the codec and the task network [49]. Choi et al. [18] adopt a scalable method to decouple the features required for downstream machine vision tasks and reconstruction tasks, thereby reducing the required bitrate. Chen et al. [14] fine-tune the codec for different tasks through prompts to achieve better performance. Feng et al. [25] employ a contrastive learning approach to learn an omni-feature to improve the performance on various downstream tasks. However, these methods all remain the following challenges: 1). need to train multiple independent networks to adapt to different bit rates, thus the training cost and memory requirement increase with the desired rate-distortion points proportionally. 2). hard to accommodate diverse machine tasks/networks, bringing extra fine-tuning overhead and bad generalization. 3). fail to meet the practical demand for adjustable distortion-cognition trade-off (decompression fidelity vs. task performance) within a single framework. In summary, as illustrated in Fig. 1(a), previous image compression methods require training different codecs to support various rate-distortion levels and machine tasks, posing a large overhead in practical scene deployment." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Some variable-bitrate (VBR) image compression methods attempt to achieve compression at different bitrates through a single model [20, 59]. For example, Song et al. [59] use a quality map to indicate compression bitrates, achieving compression at various rates. Cui et al. [20] achieve VBR by scaling different channels of the latent code through a gain unit. However, to our best knowledge, there are currently no VBR methods specifically designed for machine tasks. Most VBR methods are optimized for human vision, and can not be directly applied to machine vision tasks due to the assessment gap." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "Jinming Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 218, + 172 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 218, + 172 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 218, + 172 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 218, + 172 + ], + "type": "image", + "image_path": "f1886140281f7a477236101df4633be56c5fbfbffc040d474d770bace4a5bff8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 135, + 175, + 218, + 231 + ], + "blocks": [ + { + "bbox": [ + 135, + 175, + 218, + 231 + ], + "lines": [ + { + "bbox": [ + 135, + 175, + 218, + 231 + ], + "spans": [ + { + "bbox": [ + 135, + 175, + 218, + 231 + ], + "type": "image", + "image_path": "21dc0a3ed2c0bf008c2f346f3af059845efc27ce5d57877849ae49a654eccefb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 172, + 231, + 185, + 240 + ], + "lines": [ + { + "bbox": [ + 172, + 231, + 185, + 240 + ], + "spans": [ + { + "bbox": [ + 172, + 231, + 185, + 240 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 242, + 481, + 276 + ], + "lines": [ + { + "bbox": [ + 131, + 242, + 481, + 276 + ], + "spans": [ + { + "bbox": [ + 131, + 242, + 481, + 276 + ], + "type": "text", + "content": "Fig. 2: The visualization and spectrum of (a) original image, (b) distortion-oriented compressed image " + }, + { + "bbox": [ + 131, + 242, + 481, + 276 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_2" + }, + { + "bbox": [ + 131, + 242, + 481, + 276 + ], + "type": "text", + "content": ", (c) cognition-oriented compressed image " + }, + { + "bbox": [ + 131, + 242, + 481, + 276 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_1" + }, + { + "bbox": [ + 131, + 242, + 481, + 276 + ], + "type": "text", + "content": " and, (d) the residual " + }, + { + "bbox": [ + 131, + 242, + 481, + 276 + ], + "type": "inline_equation", + "content": "|\\hat{\\pmb{x}}_1 - \\hat{\\pmb{x}}_2|" + }, + { + "bbox": [ + 131, + 242, + 481, + 276 + ], + "type": "text", + "content": " of (b) and (c)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 221, + 115, + 306, + 172 + ], + "blocks": [ + { + "bbox": [ + 221, + 115, + 306, + 172 + ], + "lines": [ + { + "bbox": [ + 221, + 115, + 306, + 172 + ], + "spans": [ + { + "bbox": [ + 221, + 115, + 306, + 172 + ], + "type": "image", + "image_path": "1d81dc35dd0b2a38b7429f6fa8ffc1fba2ab25334cf74de10d912b89890564da.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 221, + 175, + 306, + 231 + ], + "blocks": [ + { + "bbox": [ + 221, + 175, + 306, + 231 + ], + "lines": [ + { + "bbox": [ + 221, + 175, + 306, + 231 + ], + "spans": [ + { + "bbox": [ + 221, + 175, + 306, + 231 + ], + "type": "image", + "image_path": "410381c6ad787e40b23d3211390b8b39f3823075d7d5af2f8dd517f87a484ce3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 231, + 273, + 240 + ], + "lines": [ + { + "bbox": [ + 260, + 231, + 273, + 240 + ], + "spans": [ + { + "bbox": [ + 260, + 231, + 273, + 240 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 115, + 392, + 172 + ], + "blocks": [ + { + "bbox": [ + 309, + 115, + 392, + 172 + ], + "lines": [ + { + "bbox": [ + 309, + 115, + 392, + 172 + ], + "spans": [ + { + "bbox": [ + 309, + 115, + 392, + 172 + ], + "type": "image", + "image_path": "4cad8967b79e68d7f72b72d4a42086628d0e04323af4769f6e15eeed49ad4510.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 175, + 392, + 231 + ], + "blocks": [ + { + "bbox": [ + 309, + 175, + 392, + 231 + ], + "lines": [ + { + "bbox": [ + 309, + 175, + 392, + 231 + ], + "spans": [ + { + "bbox": [ + 309, + 175, + 392, + 231 + ], + "type": "image", + "image_path": "e3cbd58f7fccb2a541a1d844a675a6006c0e045e50ff3d38fb0f7dfe7fc0546f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 231, + 358, + 240 + ], + "lines": [ + { + "bbox": [ + 347, + 231, + 358, + 240 + ], + "spans": [ + { + "bbox": [ + 347, + 231, + 358, + 240 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 394, + 115, + 479, + 172 + ], + "blocks": [ + { + "bbox": [ + 394, + 115, + 479, + 172 + ], + "lines": [ + { + "bbox": [ + 394, + 115, + 479, + 172 + ], + "spans": [ + { + "bbox": [ + 394, + 115, + 479, + 172 + ], + "type": "image", + "image_path": "472fa1f49c83a6e0dcbb80aa97aac072fe4e59d73cd2670697b41e2f0a3bf0a6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 394, + 175, + 479, + 231 + ], + "blocks": [ + { + "bbox": [ + 394, + 175, + 479, + 231 + ], + "lines": [ + { + "bbox": [ + 394, + 175, + 479, + 231 + ], + "spans": [ + { + "bbox": [ + 394, + 175, + 479, + 231 + ], + "type": "image", + "image_path": "663d44ac5a7c2bec7b70ae9721da776ef3062adbbd1e60d2f48f125b7f2ee5a4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 231, + 443, + 240 + ], + "lines": [ + { + "bbox": [ + 430, + 231, + 443, + 240 + ], + "spans": [ + { + "bbox": [ + 430, + 231, + 443, + 240 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 287, + 482, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 287, + 482, + 383 + ], + "spans": [ + { + "bbox": [ + 130, + 287, + 482, + 383 + ], + "type": "text", + "content": "Meanwhile, the distortion-realism trade-off of compression has been discussed recently [1,38]. For example, Agustsson et al. [1] achieve adjustable distortion-realism by attaching several layers of neural networks outside the decoder. However, both evaluations of distortion and realism align with human perception intuition, and many assessment aspects of them are largely overlapped. Unlike distortion or realism, cognition requires more valid information needed for downstream tasks. To the best of our knowledge, there are currently no studies discussing how to adjust the distortion-cognition trade-off in a single codec." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 385, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 482, + 517 + ], + "type": "text", + "content": "In response to the aforementioned issues, we propose a rate-distortion-cognition controllable versatile neural image compression method. Specifically, we first introduce a cognition-oriented loss that combines contrastive constraints in the main compression branch to reconstruct a more general machine task-friendly image. The physical meaning behind this loss is that contrastive learning enables better visual representations for downstream intelligent analysis tasks [10,13,29]. We can see from Fig. 2 that, the normal distortion-oriented compressed images contain more low-frequency information, aligning with the intuition that the human eye is more sensitive to low-frequency components of image [69]. In contrast, the cognition-oriented compressed images contain more valid and critical high-frequency information, making them more suitable for machine vision analysis." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 519, + 482, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 519, + 482, + 651 + ], + "spans": [ + { + "bbox": [ + 130, + 519, + 482, + 651 + ], + "type": "text", + "content": "Then, we attain variable rate-cognition by regulating quantization degree based on a channel-controllable latent gain unit [20]. This module is end-to-end trainable and can be well compatible with the above contrastive objectives. Moreover, we analyze the difference in the information needed for cognition and distortion, and based on the results we further propose an auxiliary branch to supplement residual information with an extra scalable bitstream. We find only few bits cost could significantly reduce the distortion in compressed images. At last, by controlling interpolation between distortion-oriented and cognition-oriented compressed images, we can adjust the distortion-cognition trade-off in a single codec, thereby achieving controllable rate-distortion-cognition. This effectively enhances the practicality of ICM methods." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 146, + 653, + 421, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 653, + 421, + 665 + ], + "spans": [ + { + "bbox": [ + 146, + 653, + 421, + 665 + ], + "type": "text", + "content": "The contributions of this paper can be summarized as follows:" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 137, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 137, + 90, + 448, + 102 + ], + "type": "text", + "content": "Rate-Distortion-Cognition Controllable Versatile Neural Image Compression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 480, + 224 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 137, + 116, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 479, + 151 + ], + "type": "text", + "content": "- We propose a rate-distortion-cognition controllable versatile neural image compression model. To the best of our knowledge, we are the first to achieve controllable rate-distortion-cognition within a single codec." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 152, + 480, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 152, + 480, + 187 + ], + "spans": [ + { + "bbox": [ + 138, + 152, + 480, + 187 + ], + "type": "text", + "content": "- We introduce a cognition-oriented loss based on contrastive learning to generate task-friendlyly compressed images, which main compression process also attains variable-bitrate with a trainable gain unit." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 188, + 480, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 188, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 138, + 188, + 480, + 224 + ], + "type": "text", + "content": "- We explore the relationship between the information required for distortion and cognition, and design an auxiliary scalable bitstream to achieve a balanced cognition-distortion trade-off with a two-branch interpolation strategy." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 232, + 481, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 232, + 481, + 280 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 481, + 280 + ], + "type": "text", + "content": "Experimental results demonstrate that our approach achieves satisfactory performance in both machine tasks and reconstruction, and outperforms both the latest ICM and learned image compression (LIC) methods in terms of adjustable rate-distortion-cognition practicality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 289, + 243, + 302 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 243, + 302 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 243, + 302 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 308, + 260, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 260, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 260, + 320 + ], + "type": "text", + "content": "2.1 Image Compression" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 325, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 482, + 529 + ], + "type": "text", + "content": "Image compression seeks to encode original image in a format that is both compact and retains high fidelity. Traditional image codecs, as delineated in numerous studies [9, 54, 60, 64, 66], typically incorporate modules such as intra prediction, discrete cosine transformation or wavelet transformation, quantization, and entropy coding. In contrast, learned-based codecs [6, 17, 45, 51-53] leverage neural networks to minimize distortion and bitrate. Compared to traditional approaches, data-driven learned methods have currently achieved better rate-distortion performance. Moreover, the rate-distortion trade-off in these models is governed by a Lagrange multiplier " + }, + { + "bbox": [ + 130, + 325, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 325, + 482, + 529 + ], + "type": "text", + "content": ", which typically limits most methods to a single operation point on the rate-distortion curve for a fixed " + }, + { + "bbox": [ + 130, + 325, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 325, + 482, + 529 + ], + "type": "text", + "content": " value. Recent advancements [19, 20, 37, 63, 70] have introduced various approaches to enable variable-rate compression within a single model framework. For example, Song et al. [59] innovatively propose spatial bit allocation in accordance with a quality map that matches the dimensions of the original image, offering a novel direction in the field of image compression. Cui et al. [20] utilize a scale matrix to scale the different channels of latent code to achieve variable bitrates. But these methods can not be directly applied to machine vision tasks due to the assessment gap." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 540, + 324, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 540, + 324, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 540, + 324, + 553 + ], + "type": "text", + "content": "2.2 Image Compression for Machine" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 558, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 667 + ], + "type": "text", + "content": "Image Coding for Machines (ICM) focuses on compressing and transmitting source images to facilitate downstream intelligent tasks, including image classification [21,28,35], object detection [43,55-57], semantic segmentation [5,12,46, 68, 72]. Joint optimization of image compression models with these intelligent tasks [3,33,39,42,65] is a straightforward solution in this domain. Some feature compression methods are also proposed. [4, 15, 16, 22, 23, 25, 50, 58], enhancing both coding efficiency and computing offloading. For example, Feng et al. [25] try to compress an omni-feature for better machine vision task performance. Similarly, methods based on feature selection have been proposed [14, 26, 44, 45, 61]," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "text", + "content": "Jinming Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 117, + 479, + 255 + ], + "blocks": [ + { + "bbox": [ + 134, + 117, + 479, + 255 + ], + "lines": [ + { + "bbox": [ + 134, + 117, + 479, + 255 + ], + "spans": [ + { + "bbox": [ + 134, + 117, + 479, + 255 + ], + "type": "image", + "image_path": "d2ce25096ce0780aed51e22792937e02d00c099c2097f6469254eb75bbf1cae1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 260, + 482, + 338 + ], + "lines": [ + { + "bbox": [ + 130, + 260, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 482, + 338 + ], + "type": "text", + "content": "Fig. 3: The framework of our method. The framework of our method. Our pipeline is divided into a primary branch and an auxiliary branch. The primary branch generates the cognition-oriented image " + }, + { + "bbox": [ + 130, + 260, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\hat{x}_1" + }, + { + "bbox": [ + 130, + 260, + 482, + 338 + ], + "type": "text", + "content": ", while the auxiliary branch generates the distortion-oriented residual " + }, + { + "bbox": [ + 130, + 260, + 482, + 338 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 260, + 482, + 338 + ], + "type": "text", + "content": ". We then use a " + }, + { + "bbox": [ + 130, + 260, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\hat{x}_1 + (1 - \\beta)r" + }, + { + "bbox": [ + 130, + 260, + 482, + 338 + ], + "type": "text", + "content": " interpolation strategy to achieve a balanced cognition-distortion trade-off. These two branches are trained separately in two stages: the primary branch uses cognition-oriented loss, and the auxiliary branch uses distortion-oriented loss." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 347, + 482, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 443 + ], + "type": "text", + "content": "achieving lower compression rates by compressing local regions instead of all features/images. For instance, Chen [14] et al. fine-tunes the codec via prompts, making the compressed features more suitable for downstream tasks. However, the methods mentioned above require training a separate model for compression at different bitrate points, and currently, there has been no discussion on variable-bitrate coding for machines. On the other hand, the trade-off between distortion and cognition (i.e., machine task performance) has not been publicly explored either." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 453, + 304, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 453, + 304, + 466 + ], + "spans": [ + { + "bbox": [ + 131, + 453, + 304, + 466 + ], + "type": "text", + "content": "2.3 Realism Image Compression" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 468, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 601 + ], + "type": "text", + "content": "Traditional image coding goals often aim to achieve the lowest possible distortion. However, previous works [8] have demonstrated that \"low distortion\" is not synonymous with \"high perceptual quality.\" There exists a trade-off among Rate, Distortion, and Realism (human perception). Consequently, GAN networks [52] and Diffusion models [71] have been employed to enhance realism. An attempt [1] was made to use a conditional factor to control the distortion-realism trade-off. EGIC [38] achieving similar effects using two decoders. Based on [1], Iwai et al. [34] introduce a VBR mechanism to allow the control of rate-distortion-realism on a single codec. Distinct from, yet akin to the Rate-Distortion-Realism framework, this paper considers the Rate-Distortion-Cognition issue, targeting the topic of image coding for machine (ICM)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 611, + 233, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 611, + 233, + 625 + ], + "spans": [ + { + "bbox": [ + 131, + 611, + 233, + 625 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 629, + 420, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 420, + 640 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 420, + 640 + ], + "type": "text", + "content": "3.1 Overview of the Proposed Compression Framework" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Our methodological approach is illustrated in Fig. 3. The training process is delineated into two stages, wherein the primary branch is dedicated to the gen-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 137, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 137, + 90, + 448, + 103 + ], + "type": "text", + "content": "Rate-Distortion-Cognition Controllable Versatile Neural Image Compression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 189, + 114, + 425, + 157 + ], + "blocks": [ + { + "bbox": [ + 189, + 114, + 425, + 157 + ], + "lines": [ + { + "bbox": [ + 189, + 114, + 425, + 157 + ], + "spans": [ + { + "bbox": [ + 189, + 114, + 425, + 157 + ], + "type": "image", + "image_path": "f48430298ab8a04ac840001a0e8d5257c3cb2df1fcc80813fdb47e1cfde3097f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 158, + 479, + 180 + ], + "lines": [ + { + "bbox": [ + 132, + 158, + 479, + 180 + ], + "spans": [ + { + "bbox": [ + 132, + 158, + 479, + 180 + ], + "type": "text", + "content": "Fig. 4: The training stage I. The parameters of the pretrained backbone remain frozen, with only the parameters of our codec being updated." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 202, + 479, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 202, + 479, + 227 + ], + "spans": [ + { + "bbox": [ + 130, + 202, + 479, + 227 + ], + "type": "text", + "content": "eration of cognition-oriented compressed images in stage I, while the auxiliary branch focuses on the distortion-oriented compressed images in stage II." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 231, + 480, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 231, + 480, + 362 + ], + "spans": [ + { + "bbox": [ + 130, + 231, + 480, + 362 + ], + "type": "text", + "content": "As shown in stage I of Fig. 4, to enhance the performance of the compressed image " + }, + { + "bbox": [ + 130, + 231, + 480, + 362 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_1" + }, + { + "bbox": [ + 130, + 231, + 480, + 362 + ], + "type": "text", + "content": " on downstream machine vision tasks, we employ a cognition-oriented loss instead of the conventional distortion metric commonly used in image compression. This loss comprises two parts: 1). Inspired by studies [2, 25] indicating that semantic-level loss helps image cognition and downstream tasks in compression, the first component of the cognition-oriented loss designed herein encompasses a contrastive loss [29] that facilitates the extraction of robust and discriminative representations. 2). Another part of the cognition-oriented loss is MSE of local regions in the reconstructed image where the pixel intensity exceeds [0,1]. This part is used as a penalty term to align the range of normal image values. The stage I process can be described as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 234, + 373, + 481, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 373, + 481, + 389 + ], + "spans": [ + { + "bbox": [ + 234, + 373, + 481, + 389 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} \\stackrel {\\text {e n c o d e}} {\\longrightarrow} \\boldsymbol {y} \\stackrel {\\text {q u a n t i z a t i o n}} {\\longrightarrow} \\hat {\\boldsymbol {y}} \\stackrel {\\text {d e c o d e}} {\\longrightarrow} \\hat {\\boldsymbol {x}} _ {1} \\tag {1}", + "image_path": "a3330427eea40b1e7651edba72c6c231dbc0435067cbea0b7d0abe4928c1bcfb.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 407, + 479, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 479, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 479, + 467 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 407, + 479, + 467 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 407, + 479, + 467 + ], + "type": "text", + "content": " is the original image, " + }, + { + "bbox": [ + 130, + 407, + 479, + 467 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 130, + 407, + 479, + 467 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 407, + 479, + 467 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 407, + 479, + 467 + ], + "type": "text", + "content": " are the latent code and quantized latent code extracted from " + }, + { + "bbox": [ + 130, + 407, + 479, + 467 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 407, + 479, + 467 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 407, + 479, + 467 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_1" + }, + { + "bbox": [ + 130, + 407, + 479, + 467 + ], + "type": "text", + "content": " is the cognition-oriented compressed image. Furthermore, based on the existing structure, we further attain variable rate-cognition by regulating quantization degree using a channel-controllable latent gain unit [20], this module is also trainable and compatible with the cognition-oriented loss." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "content": "In stage II, we implement the reconstruction of distortion-oriented compressed images to meet the requirements of human vision. Freezing the codec of the primary branch, we explore to generate distortion-oriented compressed images " + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_2" + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "content": " in this stage, which is defined as " + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_2 = \\hat{\\pmb{x}}_1 + \\pmb{r}" + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "content": ": the residual " + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{r}" + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "content": " between the cognition-oriented and distortion-oriented compressed images is expected to extract from " + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "content": " at the decoder side. However, since " + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "content": " is optimized for cognition in stage I, it didn't contain enough distortion-oriented supplemented information. Therefore, we construct an auxiliary branch to enrich the residual " + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{r}" + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "content": " for the high-quality distortion-oriented compressed image " + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_2" + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "content": ". As shown in Fig. 3, we first encode the quantization error " + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\pmb{y} - \\hat{\\pmb{y}})" + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "content": " at the encoder side for transmission. After a scalable transform with variable bitstream, the decoded error " + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\widehat{\\pmb{y} - \\hat{\\pmb{y}}})" + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "content": " is then added to the " + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "content": " to get " + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_2" + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "content": ", and a lightweight residual reconstruction network is used to refine residual " + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{r}" + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "content": ". Interpolating " + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{r}" + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_1" + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "content": " using a weight " + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "content": " gets " + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{x}} = \\hat{\\pmb{x}}_1 + (1 - \\beta)\\pmb{r}" + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "content": ", which allows for different distortion-cognition performance. When " + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "content": " equals 0, " + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{x}} = \\hat{\\pmb{x}}_2" + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "content": " the lowest distortion is achieved; conversely, when " + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "content": " equals 1, " + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{x}} = \\hat{\\pmb{x}}_1" + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "content": " the highest cognition performance is attained." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "text", + "content": "Jinming Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 462, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 462, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 462, + 128 + ], + "type": "text", + "content": "3.2 Primary Branch: Cognition-oriented Variable-bitrate Codec" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 138, + 479, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 138, + 479, + 161 + ], + "spans": [ + { + "bbox": [ + 131, + 138, + 479, + 161 + ], + "type": "text", + "content": "A common use of learned image compression (LIC) model [17,45,48,53], can be formulated as the following equation:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 211, + 163, + 388, + 177 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 163, + 388, + 177 + ], + "spans": [ + { + "bbox": [ + 211, + 163, + 388, + 177 + ], + "type": "interline_equation", + "content": "\\text {E n c o d i n g}: \\boldsymbol {y} = g _ {a} (\\boldsymbol {x}; \\phi_ {g _ {a}}); \\quad \\hat {\\boldsymbol {y}} = Q (\\boldsymbol {y})", + "image_path": "cf0d56d3bcf159c80b429669431d31bdf5ac5427300364f8e037028d4b8c9084.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 240, + 179, + 481, + 193 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 179, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 240, + 179, + 481, + 193 + ], + "type": "interline_equation", + "content": "\\text {D e c o d i n g :} \\hat {\\boldsymbol {x}} = g _ {s} (\\hat {\\boldsymbol {y}}; \\phi_ {g _ {s}}) \\tag {2}", + "image_path": "df5ddc3be625d7f5be7e33aeef294ca13e15022491fe0d5503184a5e0a38a057.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 150, + 195, + 449, + 210 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 195, + 449, + 210 + ], + "spans": [ + { + "bbox": [ + 150, + 195, + 449, + 210 + ], + "type": "interline_equation", + "content": "\\text {H y p e r - p r i o r :} \\boldsymbol {z} = h _ {a} \\left(\\boldsymbol {y}; \\phi_ {h _ {a}}\\right); \\quad \\hat {\\boldsymbol {z}} = Q (\\boldsymbol {z}); \\quad \\boldsymbol {\\mu} _ {y}, \\boldsymbol {\\sigma} _ {y} = h _ {s} \\left(\\hat {\\boldsymbol {z}}; \\phi_ {h _ {s}}\\right)", + "image_path": "3c6f059745fbc7ecc619f85171f020f384150dff634d8b579c445d9131f4a571.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "inline_equation", + "content": "\\phi_{g_a},\\phi_{g_s},\\phi_{h_a}" + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "inline_equation", + "content": "\\phi_{h_s}" + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "text", + "content": " are the parameters of the encoder network " + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "inline_equation", + "content": "g_{a}" + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "text", + "content": ", the decoder network " + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "inline_equation", + "content": "g_{s}" + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "text", + "content": ", the hyper-encoder network " + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "inline_equation", + "content": "h_a" + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "text", + "content": ", and the hyper-decoder network " + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "inline_equation", + "content": "h_s" + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "text", + "content": " represents the quantization which is approximated by a noise " + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "inline_equation", + "content": "\\mathcal{U}\\left(-\\frac{1}{2},\\frac{1}{2}\\right)" + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "text", + "content": " during training, and is a rounding operator during inference. For inference, an image is first input to " + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "inline_equation", + "content": "g_{a}" + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "text", + "content": " to extract a compact latent code " + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "text", + "content": " is quantized to obtain " + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "text", + "content": " by quantization operator " + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "text", + "content": ". On the decoder side, " + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "text", + "content": " is input into " + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "inline_equation", + "content": "g_{s}" + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "text", + "content": " for reconstruction to obtain the reconstructed image " + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}" + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "text", + "content": ". Additionally, the hyper path " + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "inline_equation", + "content": "(h_a" + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "inline_equation", + "content": "h_s)" + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "text", + "content": " is used to help estimate the probability distribution of " + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "text", + "content": ", which is assumed here to be a gaussian distribution with mean " + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "inline_equation", + "content": "\\mu_y" + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "inline_equation", + "content": "\\sigma_y" + }, + { + "bbox": [ + 131, + 213, + 481, + 332 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 334, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 334, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 131, + 334, + 482, + 453 + ], + "type": "text", + "content": "As discussed before, learned image compression typically requires learning a new model for different bitrate points. Therefore, in this paper, we employ a variable bitrate approach to achieve compression at different bitrates through a single model. Inspired by [20], we scale " + }, + { + "bbox": [ + 131, + 334, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 131, + 334, + 482, + 453 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 334, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 131, + 334, + 482, + 453 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 131, + 334, + 482, + 453 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 334, + 482, + 453 + ], + "type": "text", + "content": " gain unit " + }, + { + "bbox": [ + 131, + 334, + 482, + 453 + ], + "type": "inline_equation", + "content": "G_{y,n}" + }, + { + "bbox": [ + 131, + 334, + 482, + 453 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 334, + 482, + 453 + ], + "type": "inline_equation", + "content": "G_{z,n}" + }, + { + "bbox": [ + 131, + 334, + 482, + 453 + ], + "type": "text", + "content": ", which is a matrix containing " + }, + { + "bbox": [ + 131, + 334, + 482, + 453 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 131, + 334, + 482, + 453 + ], + "type": "text", + "content": " factors, where " + }, + { + "bbox": [ + 131, + 334, + 482, + 453 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 131, + 334, + 482, + 453 + ], + "type": "text", + "content": " is the channel dimension of " + }, + { + "bbox": [ + 131, + 334, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 131, + 334, + 482, + 453 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 334, + 482, + 453 + ], + "type": "inline_equation", + "content": "n \\in [1,N]" + }, + { + "bbox": [ + 131, + 334, + 482, + 453 + ], + "type": "text", + "content": ". By scaling different channels, we can adjust the precision of the data we need to transmit, further achieving transmission at " + }, + { + "bbox": [ + 131, + 334, + 482, + 453 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 334, + 482, + 453 + ], + "type": "text", + "content": " different bitrates. Furthermore, we can interpolate the scaling factors for other bitrate points from these " + }, + { + "bbox": [ + 131, + 334, + 482, + 453 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 334, + 482, + 453 + ], + "type": "text", + "content": " gain units, allowing us to switch between different bitrates using a single model. Therefore, Equation 2 can be re-formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 190, + 456, + 421, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 456, + 421, + 471 + ], + "spans": [ + { + "bbox": [ + 190, + 456, + 421, + 471 + ], + "type": "interline_equation", + "content": "\\text {E n c o d i n g}: \\boldsymbol {y} = g _ {a} (\\boldsymbol {x}; \\phi_ {g _ {a}}); \\quad \\hat {\\boldsymbol {y}} = Q (\\boldsymbol {y} \\cdot \\boldsymbol {G} _ {y, n}) / \\boldsymbol {G} _ {y, n}", + "image_path": "aecf4f6f03470a20fa9abad53b8194d286d9ef5a6ee8fad158e81d0d38ffc4a6.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 247, + 472, + 365, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 472, + 365, + 487 + ], + "spans": [ + { + "bbox": [ + 247, + 472, + 365, + 487 + ], + "type": "interline_equation", + "content": "\\text {D e c o d i n g}: \\hat {\\boldsymbol {x}} = g _ {s} (\\hat {\\boldsymbol {y}}; \\phi_ {g _ {s}})", + "image_path": "8afe7e207df7041f0ecba1985425b46aa8e38ed36270e5ae733047e8f94ee04e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 167, + 483, + 480, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 483, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 167, + 483, + 480, + 502 + ], + "type": "interline_equation", + "content": "\\text {H y p e r - p r i o r :} z = h _ {a} \\left(\\boldsymbol {y} \\cdot \\boldsymbol {G} _ {y, n}; \\phi_ {h _ {a}}\\right); \\quad \\hat {\\boldsymbol {z}} = Q (\\boldsymbol {z} \\cdot \\boldsymbol {G} _ {z, n}) / \\boldsymbol {G} _ {z, n}; \\tag {3}", + "image_path": "3ca19345663369c7423cac17082330f2bbd3a3369de3186f63f7407c46a9e79f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 260, + 505, + 351, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 505, + 351, + 519 + ], + "spans": [ + { + "bbox": [ + 260, + 505, + 351, + 519 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} _ {y}, \\sigma_ {y} = h _ {s} (\\hat {z}; \\phi_ {h _ {s}})", + "image_path": "cd0f5e2e643b537fd1cb70767af1f43d561df3059e258fda8c7e5fa4960691f0.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 522, + 479, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 479, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 479, + 544 + ], + "type": "text", + "content": "For conventional distortion-oriented compression tasks, the loss can be represented as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 150, + 548, + 481, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 548, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 150, + 548, + 481, + 578 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\mathcal {R} (\\hat {\\boldsymbol {y}}) + \\mathcal {R} (\\hat {\\boldsymbol {z}}) + \\lambda_ {n} \\cdot \\mathcal {D} (\\boldsymbol {x}, \\hat {\\boldsymbol {x}}) \\\\ = \\mathbb {E} \\left[ - \\log_ {2} \\left(p _ {\\hat {\\boldsymbol {y}} | \\hat {\\boldsymbol {z}}} (\\hat {\\boldsymbol {y}} \\mid \\hat {\\boldsymbol {z}})\\right) \\right] + \\mathbb {E} \\left[ - \\log_ {2} \\left(p _ {\\hat {\\boldsymbol {z}} | \\psi} (\\hat {\\boldsymbol {z}} \\mid \\psi)\\right) \\right] + \\lambda_ {n} \\cdot \\mathcal {D} (\\boldsymbol {x}, \\hat {\\boldsymbol {x}}) \\tag {4} \\\\ \\end{array}", + "image_path": "411951b4db296a86f968fb9b4571b308c140c687161509172f3dd92a915c19aa.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "where the parameter " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{n}" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " regulates the balance between rate and distortion, with varying " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{n}" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " values corresponding to different bit rates. " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " is a factorized density model used to encode quantized " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": ". During training, we randomly select one of the " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " bitrate points for each forward pass. The distortion term, denoted by " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\pmb{x},\\hat{\\pmb{x}})" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": ", is measured using the Mean Squared Error (MSE) loss, GAN loss or MS-SSIM loss. Additionally, " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{R}(\\hat{\\pmb{y}})" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{R}(\\hat{z})" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " represent the bit rates for the latent representations " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{z}}" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": ", respectively. To improve the performance of" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 137, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 137, + 90, + 448, + 103 + ], + "type": "text", + "content": "Rate-Distortion-Cognition Controllable Versatile Neural Image Compression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "the reconstructed images for various downstream tasks, here we use cognition-oriented loss as a substitute for the original MSE loss [17, 45, 48]. The first part of the cognition-oriented loss is a contrastive loss computed using a pretrained MoCo [29]. It is defined as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 242, + 163, + 480, + 188 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 163, + 480, + 188 + ], + "spans": [ + { + "bbox": [ + 242, + 163, + 480, + 188 + ], + "type": "interline_equation", + "content": "\\mathcal {C} = - \\log \\frac {\\exp \\left(q \\cdot k _ {+} / \\tau\\right)}{\\sum_ {i = 0} ^ {K} \\exp \\left(q \\cdot k _ {i} / \\tau\\right)} \\tag {5}", + "image_path": "233afc08047310d0d3584365610ab09c7a98b7c06c81123047a269a5666dd84f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 188, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 480, + 365 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 188, + 480, + 365 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 130, + 188, + 480, + 365 + ], + "type": "text", + "content": " denotes the query vector derived from the current image, " + }, + { + "bbox": [ + 130, + 188, + 480, + 365 + ], + "type": "inline_equation", + "content": "k_{+}" + }, + { + "bbox": [ + 130, + 188, + 480, + 365 + ], + "type": "text", + "content": " is a positive key vector generated from a different augmentation of the same image, and " + }, + { + "bbox": [ + 130, + 188, + 480, + 365 + ], + "type": "inline_equation", + "content": "k_{i}" + }, + { + "bbox": [ + 130, + 188, + 480, + 365 + ], + "type": "text", + "content": " are the negative key vectors from different images. The denominator aggregates the exponentiated similarities of the query with both the positive and K negative keys, normalized by a temperature parameter " + }, + { + "bbox": [ + 130, + 188, + 480, + 365 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 188, + 480, + 365 + ], + "type": "text", + "content": ", which adjusts the sharpness of the distribution. This loss function encourages the model to produce embeddings where positive pairs are closer in the feature space than any negative pair, thus learning robust and discriminative feature representations. However, unlike previous work with contrastive loss, which typically aims to learn a better backbone [29] or omni-features through contrastive learning [25], our objective here is to update the codec's parameters to enable the generation of cognition-oriented compressed images. Therefore, we impose constraints on the output values of the codec, to align the range of normal image values and avoid outliers. Here, we compute a local Mean Squared Error (MSE) Loss for values lying outside the [0, 1] range. It can be formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 190, + 365, + 480, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 365, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 190, + 365, + 480, + 396 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {\\text {l o c a l}} (\\boldsymbol {x}, \\hat {\\boldsymbol {x}}) = \\frac {1}{M} \\sum_ {i = 1} ^ {N} \\left[ \\mathbb {I} \\left(\\hat {x} _ {i} < 0 \\vee \\hat {x} _ {i} > 1\\right) \\cdot \\left(\\hat {x} _ {i} - x _ {i}\\right) ^ {2} \\right] \\tag {6}", + "image_path": "f3fdc68d477755f5b8ef4573bf63260ca4fc0ad7f35cf59d65127593a8e725d9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 396, + 480, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 480, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 480, + 443 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 396, + 480, + 443 + ], + "type": "inline_equation", + "content": "\\mathbb{I}" + }, + { + "bbox": [ + 130, + 396, + 480, + 443 + ], + "type": "text", + "content": " is an indicator function that takes the value of 1 when the predicted value " + }, + { + "bbox": [ + 130, + 396, + 480, + 443 + ], + "type": "inline_equation", + "content": "\\hat{x}_i" + }, + { + "bbox": [ + 130, + 396, + 480, + 443 + ], + "type": "text", + "content": " falls outside the interval [0,1], and 0 otherwise. " + }, + { + "bbox": [ + 130, + 396, + 480, + 443 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 396, + 480, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 396, + 480, + 443 + ], + "type": "inline_equation", + "content": "\\hat{x}_i" + }, + { + "bbox": [ + 130, + 396, + 480, + 443 + ], + "type": "text", + "content": " represent the target and predicted values for the " + }, + { + "bbox": [ + 130, + 396, + 480, + 443 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 396, + 480, + 443 + ], + "type": "text", + "content": " th pixel, respectively, while " + }, + { + "bbox": [ + 130, + 396, + 480, + 443 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 396, + 480, + 443 + ], + "type": "text", + "content": " denotes the number of pixels. " + }, + { + "bbox": [ + 130, + 396, + 480, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{local}}" + }, + { + "bbox": [ + 130, + 396, + 480, + 443 + ], + "type": "text", + "content": " is used as a penalty term to constrain the output range." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 444, + 481, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 444, + 481, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 444, + 481, + 574 + ], + "type": "text", + "content": "As illustrated in Fig. 5, the image distributions of both the original image and the distortion-oriented compressed image primarily fall within the range of 0 to 1. However, if we directly optimize using the contrastive loss, the generated representations would reside within a significantly broader interval (approximately [-4, 6]), which does not align with the inherent characteristics of images. Moreover, directly clipping the generated representations to the [0, 1] range during training before passing them to subsequent networks can lead to the issue of vanishing gradients, resulting in a distribution of representations as shown in Fig. 5(d). Clipping values outside the [0, 1] range during inference would lead to substantial information loss. Employing a local MSE loss as a penalty term can effectively mitigate this problem (see Fig. 5(e))." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 146, + 575, + 406, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 575, + 406, + 586 + ], + "spans": [ + { + "bbox": [ + 146, + 575, + 406, + 586 + ], + "type": "text", + "content": "Therefore, the overall loss of stage I in the main branch is:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 200, + 588, + 480, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 588, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 200, + 588, + 480, + 600 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {I} = \\mathcal {R} (\\hat {\\boldsymbol {y}}) + \\mathcal {R} (\\hat {\\boldsymbol {z}}) + \\lambda_ {n} \\cdot \\mathcal {C} + \\lambda_ {\\text {l o c a l}} \\cdot \\mathcal {D} _ {\\text {l o c a l}} (\\boldsymbol {x}, \\hat {\\boldsymbol {x}}) \\tag {7}", + "image_path": "6a2a585889179cfd51189114145fe6c3d071d9dba15aca2e4d36fa9e5bdb0d26.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 599, + 481, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 599, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 130, + 599, + 481, + 669 + ], + "type": "text", + "content": "where the parameter " + }, + { + "bbox": [ + 130, + 599, + 481, + 669 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{local}}" + }, + { + "bbox": [ + 130, + 599, + 481, + 669 + ], + "type": "text", + "content": " serves to modulate the magnitude of influence exerted by the penalty term, denoted as " + }, + { + "bbox": [ + 130, + 599, + 481, + 669 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{local}}" + }, + { + "bbox": [ + 130, + 599, + 481, + 669 + ], + "type": "text", + "content": ". It is worth noting that the primary function of " + }, + { + "bbox": [ + 130, + 599, + 481, + 669 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{local}}" + }, + { + "bbox": [ + 130, + 599, + 481, + 669 + ], + "type": "text", + "content": " loss within this context is to enforce constraints on the output range. Consequently, we empirically assign a relatively small fixed value of " + }, + { + "bbox": [ + 130, + 599, + 481, + 669 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 130, + 599, + 481, + 669 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 599, + 481, + 669 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{local}}" + }, + { + "bbox": [ + 130, + 599, + 481, + 669 + ], + "type": "text", + "content": ", acknowledging its role in merely constraining the output range without significantly impacting the primary loss components." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "text", + "content": "Jinming Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 206, + 178 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 206, + 178 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 206, + 178 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 206, + 178 + ], + "type": "image", + "image_path": "c8b69bab070e5860cef170617247b407042de9f3735d8050d26793ca8bcf4d11.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 178, + 176, + 188 + ], + "lines": [ + { + "bbox": [ + 164, + 178, + 176, + 188 + ], + "spans": [ + { + "bbox": [ + 164, + 178, + 176, + 188 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 191, + 480, + 236 + ], + "lines": [ + { + "bbox": [ + 130, + 191, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 191, + 480, + 236 + ], + "type": "text", + "content": "Fig. 5: The histograms for pixel values of: (a) the original image. (b) distortion-oriented compressed image. (c) cognition-oriented compressed image without using local mse loss. (d) cognition-oriented compressed images processed with clipping to the range " + }, + { + "bbox": [ + 130, + 191, + 480, + 236 + ], + "type": "inline_equation", + "content": "(0, 1)" + }, + { + "bbox": [ + 130, + 191, + 480, + 236 + ], + "type": "text", + "content": " during training. (e) cognition-oriented compressed images with using local mse loss." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 208, + 114, + 275, + 178 + ], + "blocks": [ + { + "bbox": [ + 208, + 114, + 275, + 178 + ], + "lines": [ + { + "bbox": [ + 208, + 114, + 275, + 178 + ], + "spans": [ + { + "bbox": [ + 208, + 114, + 275, + 178 + ], + "type": "image", + "image_path": "b1b0b7df87134551e4e5d9ccd76c0061eba0e412e87acd10037a683c3837308b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 178, + 249, + 188 + ], + "lines": [ + { + "bbox": [ + 238, + 178, + 249, + 188 + ], + "spans": [ + { + "bbox": [ + 238, + 178, + 249, + 188 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 277, + 114, + 344, + 178 + ], + "blocks": [ + { + "bbox": [ + 277, + 114, + 344, + 178 + ], + "lines": [ + { + "bbox": [ + 277, + 114, + 344, + 178 + ], + "spans": [ + { + "bbox": [ + 277, + 114, + 344, + 178 + ], + "type": "image", + "image_path": "2afaa8e768fc42554121d800167e9974c906672ec848cc7c44b51c2ff58c436b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 178, + 316, + 188 + ], + "lines": [ + { + "bbox": [ + 305, + 178, + 316, + 188 + ], + "spans": [ + { + "bbox": [ + 305, + 178, + 316, + 188 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 346, + 114, + 413, + 178 + ], + "blocks": [ + { + "bbox": [ + 346, + 114, + 413, + 178 + ], + "lines": [ + { + "bbox": [ + 346, + 114, + 413, + 178 + ], + "spans": [ + { + "bbox": [ + 346, + 114, + 413, + 178 + ], + "type": "image", + "image_path": "855c5dc16ef4eed1873e5659061c147333487f6cb24483e4e05297ac366e8fa7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 178, + 384, + 188 + ], + "lines": [ + { + "bbox": [ + 373, + 178, + 384, + 188 + ], + "spans": [ + { + "bbox": [ + 373, + 178, + 384, + 188 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 413, + 114, + 480, + 178 + ], + "blocks": [ + { + "bbox": [ + 413, + 114, + 480, + 178 + ], + "lines": [ + { + "bbox": [ + 413, + 114, + 480, + 178 + ], + "spans": [ + { + "bbox": [ + 413, + 114, + 480, + 178 + ], + "type": "image", + "image_path": "636fbb64012d196ee79e65678ce2ea5af79df9fa66a39f777d1f99d46622ab81.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 441, + 178, + 452, + 188 + ], + "lines": [ + { + "bbox": [ + 441, + 178, + 452, + 188 + ], + "spans": [ + { + "bbox": [ + 441, + 178, + 452, + 188 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 162, + 251, + 449, + 273 + ], + "blocks": [ + { + "bbox": [ + 187, + 239, + 424, + 249 + ], + "lines": [ + { + "bbox": [ + 187, + 239, + 424, + 249 + ], + "spans": [ + { + "bbox": [ + 187, + 239, + 424, + 249 + ], + "type": "text", + "content": "Table 1: Parameter numbers of different Codec modules." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 162, + 251, + 449, + 273 + ], + "lines": [ + { + "bbox": [ + 162, + 251, + 449, + 273 + ], + "spans": [ + { + "bbox": [ + 162, + 251, + 449, + 273 + ], + "type": "table", + "html": "
Base CodecScalable TransformResidual Reconstruction
Parameters(/M)20.390.191.20
", + "image_path": "dbd95b25b87ba7638c0e7bb610eda9cb7e098dba0ae2a98dc3aad91ef9da3312.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 285, + 460, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 285, + 460, + 298 + ], + "spans": [ + { + "bbox": [ + 131, + 285, + 460, + 298 + ], + "type": "text", + "content": "3.3 Auxiliary Branch: Distortion-oriented Residual Supplement" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 306, + 480, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 480, + 449 + ], + "type": "text", + "content": "To further meet the requirements of human vision, we employ an auxiliary branch to supplement residual information with a scalable bitstream. As shown in Fig. 3, we construct a residual reconstruction network to generate the residual " + }, + { + "bbox": [ + 130, + 306, + 480, + 449 + ], + "type": "inline_equation", + "content": "\\boldsymbol{r}" + }, + { + "bbox": [ + 130, + 306, + 480, + 449 + ], + "type": "text", + "content": ". By adding " + }, + { + "bbox": [ + 130, + 306, + 480, + 449 + ], + "type": "inline_equation", + "content": "\\boldsymbol{r}" + }, + { + "bbox": [ + 130, + 306, + 480, + 449 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 306, + 480, + 449 + ], + "type": "inline_equation", + "content": "\\hat{\\boldsymbol{x}}_1" + }, + { + "bbox": [ + 130, + 306, + 480, + 449 + ], + "type": "text", + "content": ", we reconstruct an distortion-oriented compressed image " + }, + { + "bbox": [ + 130, + 306, + 480, + 449 + ], + "type": "inline_equation", + "content": "\\hat{\\boldsymbol{x}}_2" + }, + { + "bbox": [ + 130, + 306, + 480, + 449 + ], + "type": "text", + "content": ". The residual reconstruction network comprises multiple layers of transposed convolutions and point-wise convolutional layers. The former is to increase the size of the representation, ensuring spatial dimension alignment with the original decoder's representation. The latter is employed to reduce and align the channel dimensions, simultaneously decreasing the model parameters and computational complexity. Additionally, the intermediate layer outputs " + }, + { + "bbox": [ + 130, + 306, + 480, + 449 + ], + "type": "inline_equation", + "content": "f_1, f_2" + }, + { + "bbox": [ + 130, + 306, + 480, + 449 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 306, + 480, + 449 + ], + "type": "inline_equation", + "content": "f_3" + }, + { + "bbox": [ + 130, + 306, + 480, + 449 + ], + "type": "text", + "content": " from the original decoder serve as prior information. They are fed into the residual reconstruction network through a point-wise convolution to help recovery." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 450, + 481, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 481, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 481, + 617 + ], + "type": "text", + "content": "We attempt to input " + }, + { + "bbox": [ + 130, + 450, + 481, + 617 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 450, + 481, + 617 + ], + "type": "text", + "content": " to the residual reconstruction network directly to generate " + }, + { + "bbox": [ + 130, + 450, + 481, + 617 + ], + "type": "inline_equation", + "content": "\\pmb{r}" + }, + { + "bbox": [ + 130, + 450, + 481, + 617 + ], + "type": "text", + "content": ". However, directly using " + }, + { + "bbox": [ + 130, + 450, + 481, + 617 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 450, + 481, + 617 + ], + "type": "text", + "content": " for reconstruction does not work, because " + }, + { + "bbox": [ + 130, + 450, + 481, + 617 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 450, + 481, + 617 + ], + "type": "text", + "content": " of the primary branch is optimized for cognitive perception rather than distortion minimization. To prove that, we analyze the quantization operation: as shown in Equation 3, this process involves the multiplication of " + }, + { + "bbox": [ + 130, + 450, + 481, + 617 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 130, + 450, + 481, + 617 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 130, + 450, + 481, + 617 + ], + "type": "inline_equation", + "content": "G_{y,n}" + }, + { + "bbox": [ + 130, + 450, + 481, + 617 + ], + "type": "text", + "content": " followed by quantization. " + }, + { + "bbox": [ + 130, + 450, + 481, + 617 + ], + "type": "inline_equation", + "content": "G_{y,n}" + }, + { + "bbox": [ + 130, + 450, + 481, + 617 + ], + "type": "text", + "content": " plays a crucial role in determining the precision of quantization across different channels and, i.e., the extent of quantization error. Given that " + }, + { + "bbox": [ + 130, + 450, + 481, + 617 + ], + "type": "inline_equation", + "content": "G_{y,n}" + }, + { + "bbox": [ + 130, + 450, + 481, + 617 + ], + "type": "text", + "content": " is optimized for cognitive perception rather than distortion minimization in stage I, thus the reconstructed representations " + }, + { + "bbox": [ + 130, + 450, + 481, + 617 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 450, + 481, + 617 + ], + "type": "text", + "content": " cannot recover the detailed residual. The gradient visualization in Fig. 6(a)(b) reveals the disparities in channel contributions between cognition-oriented and distortion-oriented images. By backpropagating the mean gradients " + }, + { + "bbox": [ + 130, + 450, + 481, + 617 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\hat{\\pmb{x}}}{\\partial\\hat{\\pmb{y}}}" + }, + { + "bbox": [ + 130, + 450, + 481, + 617 + ], + "type": "text", + "content": " from the generated image into " + }, + { + "bbox": [ + 130, + 450, + 481, + 617 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 450, + 481, + 617 + ], + "type": "text", + "content": "’s channels, we observe a marked difference in how " + }, + { + "bbox": [ + 130, + 450, + 481, + 617 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 450, + 481, + 617 + ], + "type": "text", + "content": "’s channels influence the generation of cognition-oriented versus distortion-oriented compressed images." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "To address this issue, we propose to transmit this missed information through a scalable bitstream. Specifically, by using the similar VBR method in Sec. 3.2, we first encode the quantization-induced error " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "(\\pmb{y} - \\hat{\\pmb{y}})" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": " into a bitstream via a transformation through three convolutional layers for transmission to the de" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 137, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 137, + 90, + 448, + 103 + ], + "type": "text", + "content": "Rate-Distortion-Cognition Controllable Versatile Neural Image Compression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 114, + 239, + 185 + ], + "blocks": [ + { + "bbox": [ + 145, + 114, + 239, + 185 + ], + "lines": [ + { + "bbox": [ + 145, + 114, + 239, + 185 + ], + "spans": [ + { + "bbox": [ + 145, + 114, + 239, + 185 + ], + "type": "image", + "image_path": "63d200379813159f62c0e3ea3b0d220a23f6f4e31d136dfeb85f5c3ce3fac2bd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 185, + 480, + 219 + ], + "lines": [ + { + "bbox": [ + 130, + 185, + 480, + 219 + ], + "spans": [ + { + "bbox": [ + 130, + 185, + 480, + 219 + ], + "type": "text", + "content": "Fig. 6: (a) The gradient for " + }, + { + "bbox": [ + 130, + 185, + 480, + 219 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 185, + 480, + 219 + ], + "type": "text", + "content": " from cognition-orient image " + }, + { + "bbox": [ + 130, + 185, + 480, + 219 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_1" + }, + { + "bbox": [ + 130, + 185, + 480, + 219 + ], + "type": "text", + "content": " (b) The gradient for " + }, + { + "bbox": [ + 130, + 185, + 480, + 219 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 185, + 480, + 219 + ], + "type": "text", + "content": " from distortion-orient image " + }, + { + "bbox": [ + 130, + 185, + 480, + 219 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_2" + }, + { + "bbox": [ + 130, + 185, + 480, + 219 + ], + "type": "text", + "content": " (c) The energy of cognition-orient " + }, + { + "bbox": [ + 130, + 185, + 480, + 219 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 185, + 480, + 219 + ], + "type": "text", + "content": " and distortion-orient " + }, + { + "bbox": [ + 130, + 185, + 480, + 219 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_2" + }, + { + "bbox": [ + 130, + 185, + 480, + 219 + ], + "type": "text", + "content": " (Purple represents the overlapping part of the energy between " + }, + { + "bbox": [ + 130, + 185, + 480, + 219 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 185, + 480, + 219 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 185, + 480, + 219 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_2" + }, + { + "bbox": [ + 130, + 185, + 480, + 219 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 255, + 114, + 348, + 185 + ], + "blocks": [ + { + "bbox": [ + 255, + 114, + 348, + 185 + ], + "lines": [ + { + "bbox": [ + 255, + 114, + 348, + 185 + ], + "spans": [ + { + "bbox": [ + 255, + 114, + 348, + 185 + ], + "type": "image", + "image_path": "cae51e7aaa09f80faece02c3ceda740f2d49544f330bf3e14365d0e99bba1e58.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 358, + 126, + 474, + 185 + ], + "blocks": [ + { + "bbox": [ + 358, + 126, + 474, + 185 + ], + "lines": [ + { + "bbox": [ + 358, + 126, + 474, + 185 + ], + "spans": [ + { + "bbox": [ + 358, + 126, + 474, + 185 + ], + "type": "image", + "image_path": "b4b36493df5b6e4f65a28d2653e956179afbf2273cf3cfa18b103abbc152454d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 228, + 479, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 479, + 299 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 479, + 299 + ], + "type": "text", + "content": "coder side. An auxiliary gain unit is used to control the precision and bitrate of the bitstream. Then, we reconstruct bitstream through another three convolutional layers, and the output is added to " + }, + { + "bbox": [ + 130, + 228, + 479, + 299 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 228, + 479, + 299 + ], + "type": "text", + "content": " to produce " + }, + { + "bbox": [ + 130, + 228, + 479, + 299 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_2" + }, + { + "bbox": [ + 130, + 228, + 479, + 299 + ], + "type": "text", + "content": ", which is fed into the residual reconstruction network to generate " + }, + { + "bbox": [ + 130, + 228, + 479, + 299 + ], + "type": "inline_equation", + "content": "\\pmb{r}" + }, + { + "bbox": [ + 130, + 228, + 479, + 299 + ], + "type": "text", + "content": ". The scalable transform and the residual reconstruction network together form the auxiliary branch, which exclusively costs fewer parameters compared to the base codec (see Table 1)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 300, + 480, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 480, + 371 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 480, + 371 + ], + "type": "text", + "content": "Moreover, we also analyze the differences in energy across various channels between " + }, + { + "bbox": [ + 130, + 300, + 480, + 371 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_2" + }, + { + "bbox": [ + 130, + 300, + 480, + 371 + ], + "type": "text", + "content": " and the original " + }, + { + "bbox": [ + 130, + 300, + 480, + 371 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 300, + 480, + 371 + ], + "type": "text", + "content": " in Fig. 6(c). It is observable that the energy levels across most channels are similar, with only a few channels significantly increased. This suggests that we only need to transmit very few bits to supplement such residuals to enhance decompression quality, which will be further proved in the experiments. Therefore, the overall loss of stage II can be formulated as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 244, + 374, + 480, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 374, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 244, + 374, + 480, + 388 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {I I} = \\mathcal {R} (\\hat {\\boldsymbol {s}}) + \\lambda_ {m} \\cdot \\mathcal {D} (\\boldsymbol {x}, \\hat {\\boldsymbol {x}} _ {2}) \\tag {8}", + "image_path": "a690c06b1d2f354baefb66a18987f461bff9185497764d763cebdc8713c02fc1.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 392, + 479, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 479, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 479, + 417 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 392, + 479, + 417 + ], + "type": "inline_equation", + "content": "\\mathcal{R}(\\hat{\\pmb{s}})" + }, + { + "bbox": [ + 130, + 392, + 479, + 417 + ], + "type": "text", + "content": " represents the bitrate of the scalable bitstream, " + }, + { + "bbox": [ + 130, + 392, + 479, + 417 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 392, + 479, + 417 + ], + "type": "text", + "content": " is defined as MSE loss, " + }, + { + "bbox": [ + 130, + 392, + 479, + 417 + ], + "type": "inline_equation", + "content": "\\lambda_{m}" + }, + { + "bbox": [ + 130, + 392, + 479, + 417 + ], + "type": "text", + "content": " is used to control rate-distortion, " + }, + { + "bbox": [ + 130, + 392, + 479, + 417 + ], + "type": "inline_equation", + "content": "m\\in [1,N]" + }, + { + "bbox": [ + 130, + 392, + 479, + 417 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 392, + 479, + 417 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_2 = \\hat{\\pmb{x}}_1 + \\pmb{r}" + }, + { + "bbox": [ + 130, + 392, + 479, + 417 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 428, + 343, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 428, + 343, + 441 + ], + "spans": [ + { + "bbox": [ + 131, + 428, + 343, + 441 + ], + "type": "text", + "content": "3.4 Rate-Distortion-Cognition Trade-off" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 442, + 479, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 442, + 479, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 442, + 479, + 501 + ], + "type": "text", + "content": "Through the aforementioned procedures, we are able to achieve two distinct trade-offs: rate-distortion and rate-cognition. To achieve ultra-controllability, by employing interpolation, we can adjust the trade-off between distortion and cognition, thereby enabling the controllable rate-distortion-cognition. The equation is as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 258, + 502, + 480, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 502, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 258, + 502, + 480, + 514 + ], + "type": "interline_equation", + "content": "\\bar {\\boldsymbol {x}} = \\hat {\\boldsymbol {x}} _ {1} + (1 - \\beta) \\times \\hat {\\boldsymbol {r}} \\tag {9}", + "image_path": "20ba013168b1d390356926829b264b8a4951a8c7c1a6b0f4e1e6d2c28cd1b5ef.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 512, + 480, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 480, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 480, + 583 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 512, + 480, + 583 + ], + "type": "inline_equation", + "content": "\\beta \\in [0,1]" + }, + { + "bbox": [ + 130, + 512, + 480, + 583 + ], + "type": "text", + "content": " is used to adjust the balance between cognition and distortion. As " + }, + { + "bbox": [ + 130, + 512, + 480, + 583 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 512, + 480, + 583 + ], + "type": "text", + "content": " approaches 1, the distortion of " + }, + { + "bbox": [ + 130, + 512, + 480, + 583 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{x}}" + }, + { + "bbox": [ + 130, + 512, + 480, + 583 + ], + "type": "text", + "content": " increases, and the cognition performance of " + }, + { + "bbox": [ + 130, + 512, + 480, + 583 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{x}}" + }, + { + "bbox": [ + 130, + 512, + 480, + 583 + ], + "type": "text", + "content": " is enhanced. Conversely, as " + }, + { + "bbox": [ + 130, + 512, + 480, + 583 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 512, + 480, + 583 + ], + "type": "text", + "content": " approaches 0, the distortion decreases, and the cognition performance diminishes. When transferring to downstream tasks, " + }, + { + "bbox": [ + 130, + 512, + 480, + 583 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 512, + 480, + 583 + ], + "type": "text", + "content": " is set to 1, the codec weights are kept frozen, and we only use cognition-oriented compressed images to finetune the network for the downstream tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 599, + 230, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 599, + 230, + 613 + ], + "spans": [ + { + "bbox": [ + 131, + 599, + 230, + 613 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 615, + 269, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 615, + 269, + 628 + ], + "spans": [ + { + "bbox": [ + 131, + 615, + 269, + 628 + ], + "type": "text", + "content": "4.1 Experimental Setting" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "content": "**Codec.** We set TIC [48] as our base codec. During training stage I, at each step, we randomly select a number from the set " + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "inline_equation", + "content": "[0.0625, 0.125, 0.25, 0.5, 1, 2]" + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\lambda_{n}" + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "content": ". Correspondingly, the latent code " + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "content": " is multiplied by the associated gain unit to" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "Jinming Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "implement variable bit rates. For convenience, in the subsequent sections, we use " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0,1]" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " to represent the selected point of the bitrate. When " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": ", the bitrate is at its minimum, corresponding to the model with " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\lambda_{n} = 0.0625" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": ". Conversely, when " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": ", the bitrate is at its maximum, corresponding to the model with " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\lambda_{n} = 2" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": ". For training stage II, we set " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\lambda_{m} \\in [0.00018, 0.00036, 0.00072, 0.001, 0.0015, 0.002]" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": ". To assess the rate-distortion performance of the reconstructed image, we employ the Peak Signal-to-Noise Ratio (PSNR) to evaluate the extent of distortion and use bits per pixel (bpp) to measure the compression ratio. Bpp is formulated as: as " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\frac{b}{h \\times w}" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "h, w" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " are the height and width of the source image, " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " denotes the total bits cost of the coded feature bit-stream." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 239, + 485, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 485, + 287 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 485, + 287 + ], + "type": "text", + "content": "Image Classification. we train and validate our model on ImageNet-1k datasets. During the evaluation, before inputting the images into the model, we resize and crop them to 224x224. Meanwhile, ResNet50 [30], serving as the task network, is used to assess the accuracy of classification." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 290, + 482, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 339 + ], + "type": "text", + "content": "Object Detection. For the object detection task, we train and validate on the MS COCO2017 train/val dataset. Following the setting in [67], we employ the faster-rcnn-FPN network [57] for object detection. The image scale ranges from [640, 800] pixels during training and is set to 800 at inference by default." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 342, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 415 + ], + "type": "text", + "content": "Semantic Segmentation. For this task, we employ DeepLab V3 [12], a network designed for segmentation tasks. The training and evaluation are conducted on the Cityscapes dataset. The Cityscapes dataset is specifically designed for urban street scenes. It includes 5,000 images with fine annotations. Cross entropy serves as the loss function during the training process. To assess effectiveness, we use the mean intersection over union (mIoU) as the metric." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 418, + 482, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 515 + ], + "type": "text", + "content": "Baseline. For evaluating the effectiveness of our method on machine vision tasks, we have selected several different baselines for comparison with our approach. Initially, the traditional image standard VVC (VTM-12.0) [9] was utilized for testing. For learned image compression, we employ TIC [48] to assess its performance across three tasks. Regarding the ICM method, we chose to compare with recent works, namely TransTIC [14] and T-TF [11]. To evaluate the rate-distortion performance of our model, we conducted comparisons with methods such as VTM, TIC [48], Factorized [6] and MBT-Mean [53]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 520, + 288, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 520, + 288, + 532 + ], + "spans": [ + { + "bbox": [ + 131, + 520, + 288, + 532 + ], + "type": "text", + "content": "4.2 Quantitative Comparison" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "As illustrated in Fig. 7 and 8, we compared our method against recent works. It is first noteworthy that our approach is controllable, by adjusting the values of " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": ", we can achieve the ultra-controllability within the blue area depicted in the figures using a single codec. The bitrate increases with an increase in " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": ", while the performance for machine vision tasks and image reconstruction tasks respectively increases and decreases with an increase in " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": ". As demonstrated by points ABCD in the Fig. 7 and 8, at " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\beta = 1, \\alpha = 1)" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": " (Point A), we achieve the highest bitrate and the optimal machine vision task performance at this bitrate; at " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\beta = 0, \\alpha = 1)" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": " (Point B), we obtain the highest bitrate and the lowest distortion at this bitrate; at " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\beta = 0, \\alpha = 0)" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": " (Point C), we achieve the lowest bitrate and the lowest distortion at this bitrate; at " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\beta = 1, \\alpha = 0)" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": " (Point D), we" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 137, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 137, + 90, + 447, + 102 + ], + "type": "text", + "content": "Rate-Distortion-Cognition Controllable Versatile Neural Image Compression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 248, + 203 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 248, + 203 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 248, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 248, + 203 + ], + "type": "image", + "image_path": "be94c521781b3bcd2e068f4459b85c0f39af3881c638a0e92a01576c3b2e2c2f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 204, + 238, + 213 + ], + "lines": [ + { + "bbox": [ + 153, + 204, + 238, + 213 + ], + "spans": [ + { + "bbox": [ + 153, + 204, + 238, + 213 + ], + "type": "text", + "content": "(a) Classification on ImageNet" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 249, + 114, + 364, + 203 + ], + "blocks": [ + { + "bbox": [ + 249, + 114, + 364, + 203 + ], + "lines": [ + { + "bbox": [ + 249, + 114, + 364, + 203 + ], + "spans": [ + { + "bbox": [ + 249, + 114, + 364, + 203 + ], + "type": "image", + "image_path": "47db714b07a6b29d70a24124a4a338f9192973c08220a86bce643ea74561cfc9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 266, + 204, + 355, + 213 + ], + "lines": [ + { + "bbox": [ + 266, + 204, + 355, + 213 + ], + "spans": [ + { + "bbox": [ + 266, + 204, + 355, + 213 + ], + "type": "text", + "content": "(b) Segmentation on CityScapes" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 365, + 114, + 480, + 203 + ], + "blocks": [ + { + "bbox": [ + 365, + 114, + 480, + 203 + ], + "lines": [ + { + "bbox": [ + 365, + 114, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 365, + 114, + 480, + 203 + ], + "type": "image", + "image_path": "37e9888803d7edf2c50c0c3103b6528d4df667027efe69729b1de56140826a0a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 390, + 204, + 456, + 212 + ], + "lines": [ + { + "bbox": [ + 390, + 204, + 456, + 212 + ], + "spans": [ + { + "bbox": [ + 390, + 204, + 456, + 212 + ], + "type": "text", + "content": "(c) Detection on COCO" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 133, + 239, + 250, + 329 + ], + "blocks": [ + { + "bbox": [ + 131, + 216, + 481, + 238 + ], + "lines": [ + { + "bbox": [ + 131, + 216, + 481, + 238 + ], + "spans": [ + { + "bbox": [ + 131, + 216, + 481, + 238 + ], + "type": "text", + "content": "Fig. 7: The rate-machine vision task performance curve (" + }, + { + "bbox": [ + 131, + 216, + 481, + 238 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 216, + 481, + 238 + ], + "type": "text", + "content": " changes from 0 to 1 from left to right, and " + }, + { + "bbox": [ + 131, + 216, + 481, + 238 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 131, + 216, + 481, + 238 + ], + "type": "text", + "content": " changes from 0 to 1 from bottom to top)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 239, + 250, + 329 + ], + "lines": [ + { + "bbox": [ + 133, + 239, + 250, + 329 + ], + "spans": [ + { + "bbox": [ + 133, + 239, + 250, + 329 + ], + "type": "image", + "image_path": "a4917e750a7e8f85e0eb6b09a83af5043ef703060a7821f30dc1e6140e004d67.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 330, + 210, + 338 + ], + "lines": [ + { + "bbox": [ + 171, + 330, + 210, + 338 + ], + "spans": [ + { + "bbox": [ + 171, + 330, + 210, + 338 + ], + "type": "text", + "content": "(a) ImageNet" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 339, + 479, + 361 + ], + "lines": [ + { + "bbox": [ + 131, + 339, + 479, + 361 + ], + "spans": [ + { + "bbox": [ + 131, + 339, + 479, + 361 + ], + "type": "text", + "content": "Fig. 8: The rate-PSNR performance curve (" + }, + { + "bbox": [ + 131, + 339, + 479, + 361 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 339, + 479, + 361 + ], + "type": "text", + "content": " changes from 0 to 1 from left to right, and " + }, + { + "bbox": [ + 131, + 339, + 479, + 361 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 131, + 339, + 479, + 361 + ], + "type": "text", + "content": " changes from 0 to 1 from top to bottom)." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 250, + 239, + 365, + 329 + ], + "blocks": [ + { + "bbox": [ + 250, + 239, + 365, + 329 + ], + "lines": [ + { + "bbox": [ + 250, + 239, + 365, + 329 + ], + "spans": [ + { + "bbox": [ + 250, + 239, + 365, + 329 + ], + "type": "image", + "image_path": "d096978952a45d7da0f12fc226af0c00f76fbbc6c2a46dfa54f1a5e926037cb6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 289, + 330, + 331, + 338 + ], + "lines": [ + { + "bbox": [ + 289, + 330, + 331, + 338 + ], + "spans": [ + { + "bbox": [ + 289, + 330, + 331, + 338 + ], + "type": "text", + "content": "(b) CityScapes" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 365, + 239, + 480, + 329 + ], + "blocks": [ + { + "bbox": [ + 365, + 239, + 480, + 329 + ], + "lines": [ + { + "bbox": [ + 365, + 239, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 365, + 239, + 480, + 329 + ], + "type": "image", + "image_path": "5c2eb3b5a357b8a39aedb91241ff74c62babb7fc1cae27a37e1ed4b8b7d52d16.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 330, + 436, + 338 + ], + "lines": [ + { + "bbox": [ + 406, + 330, + 436, + 338 + ], + "spans": [ + { + "bbox": [ + 406, + 330, + 436, + 338 + ], + "type": "text", + "content": "(c) COCO" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 380, + 480, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 480, + 403 + ], + "type": "text", + "content": "achieve the lowest bitrate and the best machine vision task performance at this bitrate." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 404, + 480, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 404, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 480, + 511 + ], + "type": "text", + "content": "As shown in Fig. 7 and Table 2, we compared our method " + }, + { + "bbox": [ + 130, + 404, + 480, + 511 + ], + "type": "inline_equation", + "content": "(\\beta = 1)" + }, + { + "bbox": [ + 130, + 404, + 480, + 511 + ], + "type": "text", + "content": " with recent approaches in three tasks: classification, segmentation, and detection. It is evident that, compared with the un-tuned learned image compression method TIC [48] and the traditional image compression standard VVC (VTM-12.0) [9], our method significantly improves machine vision tasks performance, especially at lower bitrates. This demonstrates our model's superior cognition capabilities. Furthermore, compared with some ICM methods, TransTIC [14], T-FT [11] + TIC [48], our approach not only achieves comparable or even superior performance but also provides controllability." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 512, + 481, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 481, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 481, + 596 + ], + "type": "text", + "content": "Fig. 8 and Table 3 illustrate the performance of our method " + }, + { + "bbox": [ + 130, + 512, + 481, + 596 + ], + "type": "inline_equation", + "content": "(\\beta = 0)" + }, + { + "bbox": [ + 130, + 512, + 481, + 596 + ], + "type": "text", + "content": " in the image reconstruction task. It is observed that compared with TIC, which is optimized for distortion, our method, being primarily optimized for cognition, leading to a reduction in PSNR. However, it still manages to achieve results comparable to the recent learned image compression methods, such as MBT-Mean [53]. Additionally, as demonstrated in Fig. 8(a), compared with the ICM method TransTIC, our method can achieve a higher upper bound for PSNR." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 131, + 602, + 239, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 602, + 239, + 615 + ], + "spans": [ + { + "bbox": [ + 131, + 602, + 239, + 615 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "In our study, we conducted ablation studies based on the base codec TIC, as shown in Table 4. For the measurements of BD-Acc [7] and BD-PSNR, we set " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\beta = 1" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\beta = 0" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": ", respectively. We explored the application of the VBR method from [20] to TIC, which allows for varying bitrates with a single model. Directly" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "Jinming Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 149, + 146, + 473, + 217 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 478, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 478, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 478, + 137 + ], + "type": "text", + "content": "Table 2: Comparison of machine vision Table 3: Comparison of image reconstruction task performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 146, + 473, + 217 + ], + "lines": [ + { + "bbox": [ + 149, + 146, + 473, + 217 + ], + "spans": [ + { + "bbox": [ + 149, + 146, + 473, + 217 + ], + "type": "table", + "html": "
Cls.Det.Seg.
BD-AccBD-APBD-mIoU|Cls.|Det.|Seg.
|BD-PSNR||BD-PSNR||BD-PSNR|
TIC [48]0.6908-0.4421-0.0208TIC [48]0.2352-0.88760.0315
TransTIC [14]8.92181.81370.1128MBT-Mean [53]-0.8492-2.1432-0.9654
T-FT TIC [11]8.34590.70910.1049Factorized [6]-1.5387-2.3248-3.4599
Ours (β=1)9.98761.82590.1133Ours (β=0)-0.4102-1.8293-1.0240
", + "image_path": "ad9a8eac15455b1269454141f4e228118c194cbc9c6484be915c48a0a874fe37.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 138, + 256, + 474, + 303 + ], + "blocks": [ + { + "bbox": [ + 132, + 219, + 481, + 252 + ], + "lines": [ + { + "bbox": [ + 132, + 219, + 481, + 252 + ], + "spans": [ + { + "bbox": [ + 132, + 219, + 481, + 252 + ], + "type": "text", + "content": "Table 4: BD-PSNR and BD-Acc on ImageNet with TIC as anchor. Codec Number accounts for required codecs across " + }, + { + "bbox": [ + 132, + 219, + 481, + 252 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 132, + 219, + 481, + 252 + ], + "type": "text", + "content": " bitrates, " + }, + { + "bbox": [ + 132, + 219, + 481, + 252 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 132, + 219, + 481, + 252 + ], + "type": "text", + "content": " downstream tasks, and " + }, + { + "bbox": [ + 132, + 219, + 481, + 252 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 132, + 219, + 481, + 252 + ], + "type": "text", + "content": " distortion-cognition scenarios." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 256, + 474, + 303 + ], + "lines": [ + { + "bbox": [ + 138, + 256, + 474, + 303 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 474, + 303 + ], + "type": "table", + "html": "
MethodsTIC [48]TIC + VBR [20]TIC + Cognition-oriented lossOurs (w/o scalable bitstream)Ours
BD-PSNR--0.1453-1.5421-1.4640-0.9104
BD-Acc--0.573911.125410.812410.6045
Codec NumbersK × M × TM × TK × T11
", + "image_path": "3865c8f70b0aabebe7a23ee242de2f1c110a5a8f6802f9d1724df57240116bd9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "text", + "content": "transferring VBR methods to machine vision tasks does not yield satisfactory results. At the same time, for other distinct scenarios (different tasks/networks and cognition-distortion trade-off), codec retraining remains requisite. Using the proposed cognition-oriented loss can significantly improve the performance of machine vision tasks. However, for different bitrates and cognition-distortion trade-offs, many codecs are still required. In contrast, our method requires only one codec to be applicable across all scenarios. For our method, when we do not transmit an auxiliary scalable bitstream and directly use " + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "text", + "content": " for the training of residual reconstruction, there is a significant improvement in BD-Acc, but the distortion-oriented compressed images exhibit heavy distortion. As analyzed in Section 3.3 and shown in Fig. 9 and Table 4, we discovered that a minimal scalable bitstream can significantly reduce reconstruction distortion, with BD-PSNR improving by approximately " + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "inline_equation", + "content": "37.81\\%" + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "text", + "content": ", and the loss in BD-Acc being only around " + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 503, + 227, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 503, + 227, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 503, + 227, + 514 + ], + "type": "text", + "content": "4.4 Visualization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 677 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 677 + ], + "type": "text", + "content": "In Fig. 10, we present a visualization of images generated with varying " + }, + { + "bbox": [ + 130, + 521, + 482, + 677 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 521, + 482, + 677 + ], + "type": "text", + "content": " values, along with their spectrum. It is observable that distortion-oriented compressed images (i.e., " + }, + { + "bbox": [ + 130, + 521, + 482, + 677 + ], + "type": "inline_equation", + "content": "\\beta = 0" + }, + { + "bbox": [ + 130, + 521, + 482, + 677 + ], + "type": "text", + "content": ") lose a significant amount of high-frequency information compared to the original image. However, as the " + }, + { + "bbox": [ + 130, + 521, + 482, + 677 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 521, + 482, + 677 + ], + "type": "text", + "content": " value increases, the high-frequency information in the images progressively increases. This indicates that cognition-oriented compressed images contain more high-frequency information, making them more conducive for analysis in downstream tasks. For instance, compared to distortion-oriented compressed images, cognition-oriented compressed images depict more detailed wrinkles on the ball held in the statue's hand. On the other hand, cognition-oriented compressed images, compared to both distortion-oriented compressed images and the original, possess a certain degree of color deviation. This suggests to some extent that color restoration is not crucial for some downstream tasks." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 137, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 137, + 90, + 448, + 103 + ], + "type": "text", + "content": "Rate-Distortion-Cognition Controllable Versatile Neural Image Compression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 164, + 114, + 289, + 186 + ], + "blocks": [ + { + "bbox": [ + 164, + 114, + 289, + 186 + ], + "lines": [ + { + "bbox": [ + 164, + 114, + 289, + 186 + ], + "spans": [ + { + "bbox": [ + 164, + 114, + 289, + 186 + ], + "type": "image", + "image_path": "b9dc8066341995bd9830dd0513cd06347ba476a8924a2d63ae96706230d75e23.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 201, + 189, + 251, + 198 + ], + "lines": [ + { + "bbox": [ + 201, + 189, + 251, + 198 + ], + "spans": [ + { + "bbox": [ + 201, + 189, + 251, + 198 + ], + "type": "text", + "content": "(a) Rate-Acc." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 322, + 114, + 446, + 186 + ], + "blocks": [ + { + "bbox": [ + 322, + 114, + 446, + 186 + ], + "lines": [ + { + "bbox": [ + 322, + 114, + 446, + 186 + ], + "spans": [ + { + "bbox": [ + 322, + 114, + 446, + 186 + ], + "type": "image", + "image_path": "e5ea83a98108f949b858e663ec3432a50df93416405e10986e772a2f30695468.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 355, + 189, + 414, + 198 + ], + "lines": [ + { + "bbox": [ + 355, + 189, + 414, + 198 + ], + "spans": [ + { + "bbox": [ + 355, + 189, + 414, + 198 + ], + "type": "text", + "content": "(b) Rate-PSNR." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 134, + 218, + 480, + 329 + ], + "blocks": [ + { + "bbox": [ + 247, + 201, + 366, + 213 + ], + "lines": [ + { + "bbox": [ + 247, + 201, + 366, + 213 + ], + "spans": [ + { + "bbox": [ + 247, + 201, + 366, + 213 + ], + "type": "text", + "content": "Fig. 9: The ablation studies." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 134, + 218, + 480, + 329 + ], + "lines": [ + { + "bbox": [ + 134, + 218, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 134, + 218, + 480, + 329 + ], + "type": "image", + "image_path": "c88ad9533cb57f15776fcc4e864175ad9d31a075ade933cef2d6096d538e47f0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 330, + 482, + 374 + ], + "lines": [ + { + "bbox": [ + 130, + 330, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 482, + 374 + ], + "type": "text", + "content": "Fig. 10: Visualization results and spectrum with different " + }, + { + "bbox": [ + 130, + 330, + 482, + 374 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 330, + 482, + 374 + ], + "type": "text", + "content": " values. (a) represents the original image, while (b) to (f) represent the compressed images with " + }, + { + "bbox": [ + 130, + 330, + 482, + 374 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 330, + 482, + 374 + ], + "type": "text", + "content": " values of 0, 0.25, 0.5, 0.75, and 1.0, respectively, (g) represents the magnification of the corresponding colored frame area." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 388, + 220, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 388, + 220, + 401 + ], + "spans": [ + { + "bbox": [ + 132, + 388, + 220, + 401 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 407, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 529 + ], + "type": "text", + "content": "In this paper, we propose a Rate-Distortion-Cognition Controllable Versatile Neural Image Compression model. We introduce a cognition-oriented loss to enhance the machine vision support, and achieve variable-rate by latent channel regulation. To further meet human eye perception, we build an auxiliary branch with a very small and scalable bitrate, decoding distortion-oriented compressed images. Finally, with an interpolation strategy to merge two branches, we also achieve distortion-cognition controlling, thereby realizing controllable rate-distortion-cognition in a single codec. Our experiments demonstrate that our method can achieve excellent performance in classification, segmentation, and detection tasks while also achieving ultra-controllability." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 557, + 241, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 241, + 571 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 241, + 571 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 582, + 481, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 619 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 619 + ], + "type": "text", + "content": "This work was supported in part by NSFC 62302246 and ZJNSFC under Grant LQ23F010008, and supported by High Performance Computing Center at Eastern Institute of Technology, Ningbo, and Ningbo Institute of Digital Twin." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "text", + "content": "Jinming Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 139, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 139, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 139, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 139, + 481, + 172 + ], + "type": "text", + "content": "1. Agustsson, E., Minnen, D., Toderici, G., Mentzer, F.: Multi-realism image compression with a conditional generator. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22324-22333 (2023) 2, 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 481, + 215 + ], + "type": "text", + "content": "2. Agustsson, E., Tschannen, M., Mentzer, F., Timofte, R., Gool, L.V.: Generative adversarial networks for extreme learned image compression. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 221-231 (2019) 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 216, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 216, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 138, + 216, + 481, + 258 + ], + "type": "text", + "content": "3. Akbari, M., Liang, J., Han, J.: Dsslic: Deep semantic segmentation-based layered image compression. In: ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 2042-2046. IEEE (2019) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 259, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 259, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 138, + 259, + 481, + 282 + ], + "type": "text", + "content": "4. Babenko, A., Slesarev, A., Chigorin, A., Lempitsky, V.: Neural codes for image retrieval. In: ECCV. pp. 584-599. Springer (2014) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 282, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 282, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 138, + 282, + 481, + 315 + ], + "type": "text", + "content": "5. Badrinarayanan, V., Kendall, A., Cipolla, R.: Segnet: A deep convolutional encoder-decoder architecture for image segmentation. TPAMI 39(12), 2481-2495 (2017) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 315, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 315, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 138, + 315, + 481, + 336 + ], + "type": "text", + "content": "6. Balle, J., Laparra, V., Simoncelli, E.P.: End-to-end optimized image compression. In: ICLR (2017) 4, 11, 13" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 336, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 336, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 336, + 481, + 358 + ], + "type": "text", + "content": "7. Bjontegaard, G.: Calculation of average psnr differences between rd-curves. In: VCEG-M33 (2001) 12" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 358, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 358, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 358, + 481, + 392 + ], + "type": "text", + "content": "8. Blau, Y., Michaeli, T.: Rethinking lossy compression: The rate-distortion-perception tradeoff. In: International Conference on Machine Learning. pp. 675-685. PMLR (2019) 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 392, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 481, + 424 + ], + "type": "text", + "content": "9. Bross, B., Wang, Y.K., Ye, Y., Liu, S., Chen, J., Sullivan, G.J., Ohm, J.R.: Overview of the versatile video coding (vvc) standard and its applications. TCSVT (2021) 1, 4, 11, 12" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 424, + 481, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 424, + 481, + 466 + ], + "spans": [ + { + "bbox": [ + 138, + 424, + 481, + 466 + ], + "type": "text", + "content": "0. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 467, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 467, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 467, + 481, + 501 + ], + "type": "text", + "content": "1. Chamain, L.D., Racapé, F., Bégaint, J., Pushparaja, A., Feltman, S.: End-to-end optimized image compression for machines, a study. In: 2021 Data Compression Conference (DCC). pp. 163-172. IEEE (2021) 11, 12, 13" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 501, + 481, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 481, + 534 + ], + "type": "text", + "content": "2. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. TPAMI 40(4), 834-848 (2017) 4, 11" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 534, + 481, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 481, + 565 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 481, + 565 + ], + "type": "text", + "content": "3. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: ICML. pp. 1597-1607. PMLR (2020) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 566, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 566, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 566, + 481, + 609 + ], + "type": "text", + "content": "4. Chen, Y.H., Weng, Y.C., Kao, C.H., Chien, C., Chiu, W.C., Peng, W.H.: Transtic: Transferring transformer-based image compression from human perception to machine perception. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 23297-23307 (2023) 2, 4, 5, 11, 12, 13" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 481, + 643 + ], + "type": "text", + "content": "5. Chen, Z., Fan, K., Wang, S., Duan, L.Y., Lin, W., Kot, A.: Lossy intermediate deep learning feature compression and evaluation. In: ACM MM. pp. 2414-2422 (2019) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 481, + 666 + ], + "type": "text", + "content": "6. Chen, Z., Fan, K., Wang, S., Duan, L., Lin, W., Kot, A.C.: Toward intelligent sensing: Intermediate deep feature compression. TIP 29, 2230-2243 (2019) 4" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 137, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 137, + 90, + 447, + 102 + ], + "type": "text", + "content": "Rate-Distortion-Cognition Controllable Versatile Neural Image Compression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "17. Cheng, Z., Sun, H., Takeuchi, M., Katto, J.: Learned image compression with discretized gaussian mixture likelihoods and attention modules. In: CVPR. pp. 7939-7948 (2020) 1, 4, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 171 + ], + "type": "text", + "content": "18. Choi, H., Bajic, I.V.: Scalable image coding for humans and machines. IEEE Transactions on Image Processing 31, 2739-2754 (2022) 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 480, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 480, + 192 + ], + "type": "text", + "content": "19. Choi, Y., El-Khamy, M., Lee, J.: Variable rate deep image compression with a conditional autoencoder. In: ICCV. pp. 3146-3154 (2019) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 192, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 192, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 192, + 480, + 224 + ], + "type": "text", + "content": "20. Cui, Z., Wang, J., Bai, B., Guo, T., Feng, Y.: G-vae: A continuously variable rate deep image compression framework. arXiv preprint arXiv:2003.02012 (2020) 2, 3, 4, 6, 7, 12, 13" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 480, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 256 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 256 + ], + "type": "text", + "content": "21. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2020) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 256, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 256, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 256, + 480, + 278 + ], + "type": "text", + "content": "22. Duan, L.Y., Chandrasekhar, V., Chen, J., Lin, J., Wang, Z., Huang, T., Girod, B., Gao, W.: Overview of thempeg-cdv standard. TIP 25(1), 179-194 (2015) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 278, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 480, + 310 + ], + "type": "text", + "content": "23. Duan, L.Y., Lou, Y., Bai, Y., Huang, T., Gao, W., Chandrasekhar, V., Lin, J., Wang, S., Kot, A.C.: Compact descriptors for video analysis: The emerging mpeg standard. IEEE MultiMedia 26(2), 44-54 (2018) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 310, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 310, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 480, + 342 + ], + "type": "text", + "content": "24. Duan, L., Liu, J., Yang, W., Huang, T., Gao, W.: Video coding for machines: A paradigm of collaborative compression and intelligent analytics. TIP 29, 8680-8695 (2020) 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "type": "text", + "content": "25. Feng, R., Jin, X., Guo, Z., Feng, R., Gao, Y., He, T., Zhang, Z., Sun, S., Chen, Z.: Image coding for machines with omnipotent feature learning. In: ECCV. pp. 510-528. Springer (2022) 2, 4, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "text", + "content": "26. Feng, R., Liu, J., Jin, X., Pan, X., Sun, H., Chen, Z.: Prompt-icm: A unified framework towards image coding for machines with task-driven prompts. arXiv preprint arXiv:2305.02578 (2023) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 407, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 450 + ], + "type": "text", + "content": "27. Gao, W., Ma, S., Duan, L., Tian, Y., Xing, P., Wang, Y., Wang, S., Jia, H., Huang, T.: Digital retina: A way to make the city brain more efficient by visual coding. IEEE Transactions on Circuits and Systems for Video Technology 31(11), 4147-4161 (2021) 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 450, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 472 + ], + "type": "text", + "content": "28. Han, K., Xiao, A., Wu, E., Guo, J., Xu, C., Wang, Y.: Transformer in transformer. NeurIPS 34, 15908-15919 (2021) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "type": "text", + "content": "29. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: CVPR. pp. 9729-9738 (2020) 3, 6, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 493, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 480, + 514 + ], + "type": "text", + "content": "30. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016) 11" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 514, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 480, + 546 + ], + "type": "text", + "content": "31. He, T., Sun, S., Guo, Z., Chen, Z.: Beyond coding: Detection-driven image compression with semantically structured bit-stream. In: 2019 Picture Coding Symposium (PCS). pp. 1-5. IEEE (2019) 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 546, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 546, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 480, + 590 + ], + "type": "text", + "content": "32. Hu, Y., Yang, J., Chen, L., Li, K., Sima, C., Zhu, X., Chai, S., Du, S., Lin, T., Wang, W., et al.: Planning-oriented autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17853-17862 (2023) 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 590, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 622 + ], + "type": "text", + "content": "33. Hu, Y., Yang, S., Yang, W., Duan, L.Y., Liu, J.: Towards coding for human and machine vision: A scalable image coding approach. In: 2020 IEEE International Conference on Multimedia and Expo (ICME). pp. 1-6. IEEE (2020) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "text", + "content": "34. Iwai, S., Miyazaki, T., Omachi, S.: Controlling rate, distortion, and realism: Towards a single comprehensive neural image compression model. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2900-2909 (2024) 5" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "Jinming Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "35. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. ECCV (2022) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "type": "text", + "content": "36. Jin, X., Feng, R., Sun, S., Feng, R., He, T., Chen, Z.: Semantical video coding: Instill static-dynamic clues into structured bitstream for ai tasks. Journal of Visual Communication and Image Representation 93, 103816 (2023) 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 480, + 213 + ], + "type": "text", + "content": "37. Johnston, N., Vincent, D., Minnen, D., Covell, M., Singh, S., Chinen, T., Hwang, S.J., Shor, J., Toderici, G.: Improved lossy image compression with priming and spatially adaptive bit rates for recurrent networks. In: CVPR. pp. 4385-4393 (2018) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 213, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 213, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 213, + 480, + 246 + ], + "type": "text", + "content": "38. Körber, N., Kromer, E., Siebert, A., Hauke, S., Mueller-Gritschneider, D.: Egic: Enhanced low-bit-rate generative image compression guided by semantic segmentation. arXiv preprint arXiv:2309.03244 (2023) 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 246, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 480, + 289 + ], + "type": "text", + "content": "39. Le, N., Zhang, H., Cricri, F., Ghaznavi-Youvalari, R., Rahtu, E.: Image coding for machines: An end-to-end learned approach. In: ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1590-1594. IEEE (2021) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 289, + 480, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 312 + ], + "type": "text", + "content": "40. Li, H., Li, S., Dai, W., Li, C., Zou, J., Xiong, H.: Frequency-aware transformer for learned image compression. arXiv preprint arXiv:2310.16387 (2023) 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 312, + 480, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 480, + 343 + ], + "type": "text", + "content": "41. Li, H., Li, S., Ding, S., Dai, W., Cao, M., Li, C., Zou, J., Xiong, H.: Image compression for machine and human vision with spatial-frequency adaptation. In: ECCV. Springer (2024) 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 344, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 344, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 344, + 480, + 365 + ], + "type": "text", + "content": "42. Li, X., Shi, J., Chen, Z.: Task-driven semantic coding via reinforcement learning. TIP (2021) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 365, + 480, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 386 + ], + "type": "text", + "content": "43. Li, Y., Mao, H., Girshick, R., He, K.: Exploring plain vision transformer backbones for object detection. ECCV (2022) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 386, + 480, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 408 + ], + "type": "text", + "content": "44. Liu, J., Sun, H., Katto, J.: Semantic segmentation in learned compressed domain. In: 2022 Picture Coding Symposium (PCS). pp. 181-185. IEEE (2022) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 408, + 480, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 480, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 480, + 440 + ], + "type": "text", + "content": "45. Liu, J., Sun, H., Katto, J.: Learned image compression with mixed transformer-cnn architectures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14388-14397 (2023) 1, 2, 4, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 440, + 480, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 480, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 480, + 461 + ], + "type": "text", + "content": "46. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: CVPR. pp. 3431-3440 (2015) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 461, + 480, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 480, + 483 + ], + "type": "text", + "content": "47. Lu, G., Ge, X., Zhong, T., Geng, J., Hu, Q.: Preprocessing enhanced image compression for machine vision. arXiv preprint arXiv:2206.05650 (2022) 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "type": "text", + "content": "48. Lu, M., Guo, P., Shi, H., Cao, C., Ma, Z.: Transformer-based image compression. arXiv preprint arXiv:2111.06707 (2021) 1, 7, 8, 10, 11, 12, 13" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "text", + "content": "49. Ma, H., Liu, D., Yan, N., Li, H., Wu, F.: End-to-end optimized versatile image compression with wavelet-like transform. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(3), 1247-1263 (2020) 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "text", + "content": "50. Ma, S., Zhang, X., Wang, S., Zhang, X., Jia, C., Wang, S.: Joint feature and texture coding: Toward smart video representation via front-end intelligence. TCSVT 29(10), 3095-3105 (2018) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 568, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 480, + 600 + ], + "type": "text", + "content": "51. Mentzer, F., Agustsson, E., Tschannen, M., Timofte, R., Van Gool, L.: Conditional probability models for deep image compression. In: CVPR. pp. 4394-4402 (2018) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "type": "text", + "content": "52. Mentzer, F., Toderici, G., Tschannen, M., Agustsson, E.: High-fidelity generative image compression. arXiv preprint arXiv:2006.09965 (2020) 4, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "text", + "content": "53. Minnen, D., Balle, J., Toderici, G.: Joint autoregressive and hierarchical priors for learned image compression. In: NeurIPS (2018) 4, 7, 11, 12, 13" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "text", + "content": "54. Rabbani, M., Joshi, R.: An overview of the jpeg 2000 still image compression standard. Signal processing: Image communication 17(1), 3-48 (2002) 4" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 137, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 137, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rate-Distortion-Cognition Controllable Versatile Neural Image Compression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 610 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "55. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified, real-time object detection. In: CVPR. pp. 779-788 (2016) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "type": "text", + "content": "56. Redmon, J., Farhadi, A.: Yolo9000: better, faster, stronger. In: CVPR. pp. 7263-7271 (2017) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "text", + "content": "57. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS 28, 91-99 (2015) 4, 11" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "text", + "content": "58. Singh, S., Abu-El-Haija, S., Johnston, N., Balle, J., Shrivastava, A., Toderici, G.: End-to-end learning of compressible features. In: 2020 IEEE International Conference on Image Processing (ICIP). pp. 3349-3353. IEEE (2020) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "text", + "content": "59. Song, M., Choi, J., Han, B.: Variable-rate deep image compression through spatially-adaptive feature transform. In: ICCV. pp. 2380-2389 (2021) 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 238, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 480, + 259 + ], + "type": "text", + "content": "60. Sullivan, G.J., Ohm, J.R., Han, W.J., Wiegand, T.: Overview of the high efficiency video coding (hevc) standard. TCSVT 22(12), 1649-1668 (2012) 1, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 260, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 480, + 281 + ], + "type": "text", + "content": "61. Sun, S., He, T., Chen, Z.: Semantic structured image coding framework for multiple intelligent applications. TCSVT (2020) 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 281, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 325 + ], + "type": "text", + "content": "62. Terhörst, P., Ihlefeld, M., Huber, M., Damer, N., Kirchbuchner, F., Raja, K., Kuijper, A.: Qmagface: Simple and accurate quality-aware face recognition. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3484-3494 (2023) 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "text", + "content": "63. Toderici, G., Vincent, D., Johnston, N., Jin Hwang, S., Minnen, D., Shor, J., Covell, M.: Full resolution image compression with recurrent neural networks. In: CVPR. pp. 5306-5314 (2017) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "text", + "content": "64. Wallace, G.K.: Thejpeg still picture compression standard. IEEE transactions on consumer electronics 38(1), xviii-xxxiv (1992) 1, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "65. Wang, S., Wang, S., Yang, W., Zhang, X., Wang, S., Ma, S., Gao, W.: Towards analysis-friendly face representation with scalable feature and texture compression. TMM (2021) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "type": "text", + "content": "66. Wiegand, T., Sullivan, G.J., Bjontegaard, G., Luthra, A.: Overview of the h. 264/avc video coding standard. TCSVT 13(7), 560-576 (2003) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "text", + "content": "67. Wu, Y., Kirillov, A., Massa, F., Lo, W.Y., Girshick, R.: Detector2. https://github.com/facebookresearch/detectron2 (2019) 11" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "text", + "content": "68. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. NeurIPS 34, 12077-12090 (2021) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 490, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 521 + ], + "type": "text", + "content": "69. Yan, F.F., Hou, F., Lu, Z.L., Hu, X., Huang, C.B.: Efficient characterization and classification of contrast sensitivity functions in aging. Scientific reports 7(1), 5045 (2017) 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "type": "text", + "content": "70. Yang, F., Herranz, L., Van De Weijer, J., Guitián, J.A.I., López, A.M., Mozerov, M.G.: Variable rate deep image compression with modulated autoencoder. IEEE Signal Processing Letters 27, 331-335 (2020) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 555, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 578 + ], + "type": "text", + "content": "71. Yang, R., Mandt, S.: Lossy image compression with conditional diffusion models. arXiv preprint arXiv:2209.06950 (2022) 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "text", + "content": "72. Zheng, S., Lu, J., Zhao, H., Zhu, X., Luo, Z., Wang, Y., Fu, Y., Feng, J., Xiang, T., Torr, P.H., et al.: Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In: CVPR. pp. 6881-6890 (2021) 4" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "text", + "content": "Jinming Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Rawformer_ Unpaired Raw-to-Raw Translation for Learnable Camera ISPs/c3c8340d-834b-4e08-bdff-c9515d0d2660_content_list.json b/2024/Rawformer_ Unpaired Raw-to-Raw Translation for Learnable Camera ISPs/c3c8340d-834b-4e08-bdff-c9515d0d2660_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bcbd4e21698b34f2ea96a02d60f7af1c30c6bf5d --- /dev/null +++ b/2024/Rawformer_ Unpaired Raw-to-Raw Translation for Learnable Camera ISPs/c3c8340d-834b-4e08-bdff-c9515d0d2660_content_list.json @@ -0,0 +1,1883 @@ +[ + { + "type": "text", + "text": "Rawformer: Unpaired Raw-to-Raw Translation for Learnable Camera ISPs", + "text_level": 1, + "bbox": [ + 233, + 140, + 769, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Georgy Perevozhikov1, Nancy Mehta1, Mahmoud Affifi2, and Radu Timofte1", + "bbox": [ + 228, + 212, + 774, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Computer Vision Lab, CAIDAS & IFI, University of Würzburg, John Skilton Str. 4a, 97074 Würzburg, Germany", + "bbox": [ + 279, + 253, + 720, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{georgii.perevozhikov,nancy.mehta,radu.timofte}@uni-wuerzburg.de", + "bbox": [ + 250, + 284, + 751, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ York University, 4700 Keele St, Toronto, Ontario, Canada, M3J 1P3 m.3afifi@gmail.com", + "bbox": [ + 267, + 296, + 733, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Modern smartphone camera quality heavily relies on the image signal processor (ISP) to enhance captured raw images, utilizing carefully designed modules to produce final output images encoded in a standard color space (e.g., sRGB). Neural-based end-to-end learnable ISPs offer promising advancements, potentially replacing traditional ISPs with their ability to adapt without requiring extensive tuning for each new camera model, as is often the case for nearly every module in traditional ISPs. However, the key challenge with the recent learning-based ISPs is the urge to collect large paired datasets for each distinct camera model due to the influence of intrinsic camera characteristics on the formation of input raw images. This paper tackles this challenge by introducing a novel method for unpaired learning of raw-to-raw translation across diverse cameras. Specifically, we propose Rawformer, an unsupervised Transformer-based encoder-decoder method for raw-to-raw translation. It accurately maps raw images captured by a certain camera to the target camera, facilitating the generalization of learnable ISPs to new unseen cameras. Our method demonstrates superior performance on real camera datasets, achieving higher accuracy compared to previous state-of-the-art techniques, and preserving a more robust correlation between the original and translated raw images. The codes and the pretrained models are available at https://github.com/gosha20777/rawformer.", + "bbox": [ + 261, + 362, + 738, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 679, + 374, + 695 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unlike the high-quality digital single-lens reflex (DSLR) cameras, mobile phone cameras possess inherent limitations owing to their smaller sensor size and fixed lenses with limited optical capabilities [55]. The mobile camera image signal processor (ISP) aims to overcome these limitations by applying a series of carefully designed modules (e.g., [1,7,11,18,25,47]) to enhance the quality of the final image rendered in a standard color space, such as sRGB. These modules", + "bbox": [ + 212, + 710, + 787, + 801 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding Author", + "bbox": [ + 215, + 810, + 385, + 825 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "\\*\\* Now at Google", + "bbox": [ + 215, + 825, + 336, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6d4b4324a34ece44df8b3e8175552d1dcb3c71c9c40ec6bee470dfcf63e06f97.jpg", + "image_caption": [ + "Fig. 1: We introduce Rawformer, an unsupervised method for raw-to-raw translation that allows the utilization of pre-trained neural-based ISPs to process raw images captured by previously unseen cameras. Shown are raw and sRGB images processed by a neural-based ISP [54]. a) Raw image from an iPhone X rendered by the neural-based ISP trained on iPhone X's raw images. b) Raw image from an iPhone X rendered by the neural-based ISP trained on Samsung S9's raw images. c) iPhone X raw image translated to Samsung S9's raw space using our method, then processed by the Samsung S9 neural-based ISP. d) Raw image captured by Samsung S9's camera rendered by its native camera ISP, provided as a reference for visual comparison." + ], + "image_footnote": [], + "bbox": [ + 230, + 143, + 356, + 229 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4a11bb1e901a555772de42314ff1b018edd143f81ab5cf8d72c9b18e69d0b52c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 143, + 493, + 228 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d094731a4b6cdc70ad2cafa41d617ddd4e39e9bd24ade8e2bae5d49b1b582cb1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 143, + 633, + 228 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d56d209367b411175300c35c10ca96bbb4793970abf8d2c1dc3b81fb6770607f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 650, + 143, + 772, + 228 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "work in tandem to optimize key parameters, such as sharpness, color accuracy, and noise reduction, resulting in a visually pleasing sRGB image from the raw image captured by the camera [13, 16].", + "bbox": [ + 212, + 397, + 782, + 443 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite yielding promising results, achieving compatibility among camera ISP modules for each new camera device requires extensive tuning and adjustment efforts to attain the desired image quality. This motivates the use of deep neural networks to replace individual camera ISP modules with a single neural-based unit capable of accomplishing the task [12,19,44]. Nonetheless, the training of such neural-based ISPs demands an extensive dataset consisting of raw images taken with a specific camera, accompanied by corresponding \"ground-truth\" sRGB images (typically generated by a high-quality DSLR camera) [19, 22].", + "bbox": [ + 212, + 444, + 787, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "An issue emerges with the introduction of a new camera model, as pre-trained neural-based ISPs might encounter difficulty in accurately rendering raw images captured by the new camera (refer to Fig. 1(b)). This difficulty arises from the potential distinct characteristics of the new camera, such as sensor sensitivities, which influence the formation of its raw images [4,5]. Consequently, inconsistencies arise in the interpretation of raw RGB colors between cameras [4], hindering the generalization of neural-based ISPs to new cameras not encountered during training. To address this issue, re-training or fine-tuning on a new paired dataset becomes necessary, in which the raw images are sourced from the target new camera. This process mirrors similar challenges encountered in traditional ISP development for new cameras.", + "bbox": [ + 212, + 566, + 787, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "An alternative strategy to address this challenge is domain adaptation, where the raw images from the new target camera are mapped to the raw space of the training camera to emulate those captured by the original camera used for training. In light of this, [4] proposed a semi-supervised learning framework for raw image translation, but the approach outlined in [4] inherits the drawbacks associated with the CNN-based architecture, notably the limited receptive field and inefficient encoding of global information. Thus, we present a fully unsupervised", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "G. Perevozhikov et al.", + "bbox": [ + 271, + 114, + 426, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Transformer-based raw-to-raw method, dubbed Rawformer, that maps raw images between camera models by efficiently encoding their global and semantic correlations. Our method eliminates the need for a paired raw-sRGB dataset from each new target camera, enabling the utilization of a neural-based ISP trained on a specific camera's raw images to process raw images taken by new cameras with different characteristics without necessitating any re-training; see Fig. 1(c). Our contributions are summarized as follows:", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose Rawformer, a fully unsupervised encoder-decoder Transformer-based method designed for raw-to-raw translation facilitating the reuse of neural-based ISPs. Rawformer achieves state-of-the-art results in raw-to-raw translation, leading to better generalization for cross-camera learnable ISP rendering.", + "- We introduce contextual-scale aware downsampler and upsampler blocks that efficiently summarize the local-global contextual details in mixed scale representations via its condensed query attention block and scale perceptive feed-forward network.", + "- A novel cross-domain attention-driven discriminator is proposed along with a specialized discriminator head for stabilizing the network training." + ], + "bbox": [ + 225, + 261, + 787, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 450, + 387, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Neural-Based ISP", + "text_level": 1, + "bbox": [ + 215, + 482, + 410, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Several neural-based ISPs methods have been introduced [8,17,19,24,56,59,60], with the majority relying on CNN-based U-Net-like architecture [42]. These include earlier approaches, such as DeepISP [44] and PyNET [22], as well as more recent methods, such as MW-ISPNet [21], AW-Net [12], and LAN [54], which achieve improved image restoration quality by integrating discrete wavelet transform (DWT) and double attention modules (DAM) techniques. Such learnable ISPs are typically trained to map raw images captured by a specific camera model to target high-quality sRGB \"ground-truth\" images produced by some target camera ISP (usually with higher quality than the mobile native ISP; e.g., a DSLR camera [19,22]). A significant challenge for neural-based ISPs is the need for paired datasets of raw-sRGB images. Collecting such datasets is tedious and typically involves using specific equipment. Additionally, it may require post-processing techniques to align the raw images with the target sRGB images [22]. The challenge intensifies when a mobile phone manufacturer introduces a new camera model with different characteristics, which happens frequently [49], requiring the recollection of paired raw-sRGB datasets for each new camera model.", + "bbox": [ + 212, + 507, + 787, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "One solution is to render images to the sRGB space from the device-independent CIE XYZ space [3]. However, this requires calibration for each camera model, and accuracy depends on specific camera conditions (e.g., the Luther condition) [15]. In contrast, our work introduces an unsupervised raw-to-raw mapping technique that eliminates the need for paired raw-sRGB datasets for re-training or camera calibration. Our method maps raw images from new cameras to the camera raw", + "bbox": [ + 212, + 750, + 802, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Rawformer", + "bbox": [ + 655, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "space utilized during the training of neural-based ISPs, enabling the reuse of pre-trained ISPs.", + "bbox": [ + 212, + 146, + 787, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Image-to-Image Domain Adaptation", + "text_level": 1, + "bbox": [ + 214, + 199, + 562, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Numerous generative adversarial network (GAN)-based methods address unpaired image-to-image (I2I) translation challenges [38]. Notably, CycleGAN [62], DualGAN [57], UNIT [32], STARGAN [10], SEAN [63], and U-GAT-IT [28] focus on cycle consistency, employing dual generator networks guided by a cycle-consistency constraint. ACLGAN [61] and CUT [39] represent advancements in this field. ACLGAN relaxes the cycle consistency constraint, favoring an adversarial constraint, potentially enhancing translation quality. CUT utilizes a contrastive loss, eliminating the need for multiple generators and enabling faster training. ITTR further refines CUT's performance by altering the generator architecture. CycleGAN-based model, such as UVCGAN [50] incorporate U-Net-like generators with vision Transformers (ViTs).", + "bbox": [ + 212, + 224, + 785, + 390 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In computational photography, attempts have been made for domain adaptation, including synthesizing raw night images from raw daylight images to enhance neural-ISP rendering quality [41], few-shot domain adaptation for low-light image enhancement [40], and mapping graphics images to a target camera's raw space [45]. Raw-to-raw translation has received even limited attention. Classical approaches require a calibration object to learn global mapping [37], while recent work [4] proposes a semi-supervised CNN learning-based method using a small set of perfectly aligned images. Despite being effective at exploiting the local semantics, this CNN-based method lacks the potential to exploit the long-range/global semantics between different domains. In contrast, our method aims at solving the raw mapping by exploiting both local and global contextual information through fully unsupervised training, achieving state-of-the-art results.", + "bbox": [ + 212, + 391, + 787, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Proposed Method", + "text_level": 1, + "bbox": [ + 214, + 595, + 428, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing raw-to-raw methodologies either rely on supervised approaches requiring paired datasets e.g., [37]) or semi-supervised approaches (e.g., [4]) with limited accuracy. Additionally, domain adaptation techniques (e.g., [39, 62]), which aim to extract local semantics between source and target domains, often fail to capture overall semantics and context-dependent relationships. Our proposed method, Rawformer, addresses these limitations by processing raw images in a fully unsupervised manner. Utilizing an encoder-decoder architecture within a CycleGAN framework, Rawformer effectively captures global semantics and maintains intrinsic consistency between translated images. The generator (Fig. 2(a)) in Rawformer employs contextual-scale aware downsampler and upsampler blocks to enhance content rendering and structural integrity. The discriminator (Fig. 3(a)) stabilizes training by incorporating key discriminator features, thereby preventing model collapse and reducing redundant computations. In the following, we provide a comprehensive elucidation of the proposed components.", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "G. Perevozhikov et al.", + "bbox": [ + 271, + 114, + 426, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/87d13479d97e438cbc6cecdb877aec1e3d210bef0f7116250da0faf5c8f48642.jpg", + "image_caption": [ + "a) Generator" + ], + "image_footnote": [], + "bbox": [ + 215, + 148, + 455, + 308 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/829c774fe30dbaf68490f5134ac9ce23794488ae53868b98badc25b54d2035d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 462, + 148, + 560, + 243 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/48d9ada0364f28b5c3b82ef9e398111ce2c4e07c5c6f2ddde0bb23726fdd8893.jpg", + "image_caption": [ + "c) CQA block", + "Fig.2: a) Overview of the generator architecture (a) of Rawformer. The primary components of the generator: b) contextual-scale aware downsampler block (CSAD), c) condensed query attention block (CQA), d) scale perceptive feed-forward network (SPFN), e) composite downsampling block (CDown), f) composite upsampler (CUp) block, and g) contextual-scale aware upsampler block (CSAU)." + ], + "image_footnote": [], + "bbox": [ + 462, + 244, + 558, + 316 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7ffa760ffe9fb289975c9f9060ae4c653577385a3d9f8d95933507b9458e1208.jpg", + "image_caption": [ + "f) $\\mathrm{CUp}$ block" + ], + "image_footnote": [], + "bbox": [ + 575, + 147, + 681, + 188 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dc4461e5226c53c8efb70affbb69465b38c775bdc558f01eb1c2715c98c2055e.jpg", + "image_caption": [ + "e) CDown block" + ], + "image_footnote": [], + "bbox": [ + 575, + 199, + 679, + 239 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1e459a2f6775c04f9985c274217497f4b8127425b03ac5650483f6ae9cb365c5.jpg", + "image_caption": [ + "g) CSAU block" + ], + "image_footnote": [], + "bbox": [ + 692, + 146, + 784, + 242 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b3ee77e438653845a1e36cf64d406faa3fa2eea0fc0b76ffee8cc305be011c77.jpg", + "image_caption": [ + "d) SPFN block" + ], + "image_footnote": [], + "bbox": [ + 571, + 253, + 715, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Generator's Encoder Network", + "text_level": 1, + "bbox": [ + 215, + 440, + 509, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 2(a), the encoder of the proposed generator in Rawformer comprises stacked contextual-scale aware downsampler (CSAD) blocks to bridge the semantic gap between the source and target domains. Each CSAD block, depicted in Fig. 2(b), offers a holistic approach of understanding the contextual dependencies efficiently between the domains by capturing both the global and local information via its condensed query attention (CQA) block, in parallel with a $3 \\times 3$ depth-wise convolution layer. It further handles the intrinsic correlations among the extracted features in a multi-scale manner via maintaining a synergistic collaboration between the scale perceptive feed-forward network (SPFN), and the composite downsampler (CDown) block.", + "bbox": [ + 212, + 468, + 787, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Condensed Query Attention Block: Self-attention is highly effective in modeling long-range contextual dependencies by comparing each token with all others, facilitating seamless information propagation within the network [58]. However, its computational demands pose challenges for tasks like unpaired raw-to-raw translation, where maintaining both visual quality and semantic consistency is crucial. To address these concerns and overcome the constraints of computation, we introduce the condensed query attention (CQA) block; see Fig. 2(c). Here, we make two key modifications to self-attention to enhance learning efficiency. Firstly, we spatially compress the query vector using average pooling followed by linear projection, and downsample the resulting feature map by a factor of $r$ to confine the crucial information. This results in condensed queries, $(\\mathbf{Q}_c)$ used for similarity comparison, as shown in Eq. 1. Secondly, to balance global and regional modeling with computational efficiency, attention is applied hor", + "bbox": [ + 212, + 643, + 787, + 843 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Rawformer", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "iontally and vertically to eliminate redundant computations. The CQA block operation is defined as:", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nC Q A (\\mathbf {Q} _ {c}, \\mathbf {K}, \\mathbf {V}) = \\mathbf {A} _ {U}. (\\mathbf {A} _ {H} \\cdot \\mathbf {V}); \\quad \\mathbf {Q} _ {c} = W _ {p} (A v g P o o l (\\mathbf {Q}))\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 188, + 733, + 203 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} _ {H} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} _ {c} \\mathbf {K} ^ {T}}{\\sqrt {d}}\\right), \\quad \\mathbf {A} _ {U} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {Q} _ {c} ^ {T}}{\\sqrt {d}}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 203, + 782, + 226 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where, the duet of key $(\\mathbf{K})$ , and value $(\\mathbf{V})$ vectors are obtained through plain linear convolutions, and the dimensions of each of these features are shown in the Fig. 2(c). $W_{p}(.)$ denotes a $1\\times 1$ pointwise convolution, and $\\mathbf{A}_H$ , and $\\mathbf{A}_U$ represents the compressed attention map between the key-condensed query pair horizontally and query-condensed query pair (for focusing on fine grained contextual information) vertically, respectively. The choice of operations and the reason for condensing the queries is investigated in the supplementary of the paper.", + "bbox": [ + 212, + 250, + 784, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Scale Perceptive Feed-forward Network: To proficiently capture the intricacies of the local image structure across various scales by encoding the information from neighboring pixel locations, we propose a scale perceptive feed-forward network (SPFN); see Fig. 2(d). Unlike the conventional feed-forward networks that integrate single-scale depth-wise convolutions [58], we emphasize on enhancing the locality and inter-correlations among the incoming raw features of different domains by inserting multi-scale depth-wise convolutions. Given an input tensor at layer $s$ and scale $i$ , such that $\\mathbf{X}_s^i \\in \\mathbb{R}^{H^i \\times W^i \\times C}$ , SPFN is formulated as:", + "bbox": [ + 212, + 393, + 784, + 526 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{c} \\mathbf {X} _ {s} ^ {1} = \\phi \\left(W _ {d} ^ {3} \\left(W _ {p} ^ {1} (L N (\\mathbf {X} _ {s - 1}))\\right)\\right); \\mathbf {X} _ {s} ^ {2} = \\phi \\left(W _ {d} ^ {5} \\left(W _ {p} ^ {2} (L N (\\mathbf {X} _ {s - 1}))\\right)\\right), \\\\ \\mathbf {Y} _ {s} = W _ {s} ^ {3} \\left(\\mathbf {X} _ {o} ^ {1} (\\widehat {\\mathbf {C}}) \\mathbf {X} _ {s - 1} ^ {2}\\right) + \\mathbf {X} _ {s - 1}. \\end{array} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 277, + 526, + 782, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $\\phi$ represents a LeakyReLU activation function [35], $W_{d}^{k}$ represents a depthwise convolution with filter size $k$ , $LN$ denotes the layer normalization [6], and $\\mathbb{C}$ represents the concatenation operation. This enriches a more nuanced understanding of the underlying image features at different hierarchical levels.", + "bbox": [ + 212, + 566, + 782, + 630 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Composite Down sampler: In general, the essence of downsampling is the retrieval of the multi-scale information in the image. Unlike the common encoder architectures for various image signal processing pipelines [31, 48] that deploy strided convolutions for downsampling, we design a hybrid strategy of leveraging the benefits of information exchange between different resolution features. To adaptively recover the underlying content information from the input domain at multiple scales, we design a composite down (CDown) block that intelligently combines pixel-unshuffle (to prevent the information loss) and strided convolution (to enhance the model's expressiveness) as shown in Fig. 2(e).", + "bbox": [ + 212, + 651, + 784, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Style Modulator: Different cameras have their own distinctive characteristics to be handled or restored. To augment the expressiveness of Rawformer for", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "G. Perevozhikov et al.", + "bbox": [ + 271, + 114, + 424, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "handling camera perturbations, we design a style modulator in the generator using extended pixel-wise ViT [51]. It aids in calibrating the decoder features and subsequently encourages the recovery of crucial details while translation. As demonstrated in Fig. 2(a), the style modulator applies style tokens in each contextual aware upsampler (CSAU) block, where every style token models the weight of the query vector of the CQA block to enrich the style specific context. This addition of style token at each stage in the decoder helps in the flexible adjustment of the feature maps. Unlike StyleGANv2 [27], where the style vectors are generated from the random prior, our approach resorts to learnable style token for conjecturing the important target style directly from the source domain, thus enhancing the consistency and control in raw image translation. More details on style modulation are provided in the supplementary.", + "bbox": [ + 212, + 146, + 787, + 328 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Generator's Decoder Network", + "text_level": 1, + "bbox": [ + 215, + 353, + 509, + 367 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Unlike the popular additive or multiplicative skip connections between the encoder-decoder, we also present a learnable linking layer (LLayer) as shown in the Fig. 2(a) to facilitate the feature preservation in the contextual scale aware upsampler (CSAU) block. The overall design of the CSAU block closely resembles that of the CSAD block, ensuring consistency and coherence in the network architecture in a bottom-up fashion. The primary distinction lies within the composite upsampling block, where for inducing the relevant spatial information into the target domain for the incoming features from the style modulator, we design a composite upsampler block (CUp). It embodies a parallel combination of pixel-shuffle (for preventing the checkerboard artifacts in the target domain) and deconvolution (for combating with aliasing) as shown in Fig. 2(f) and (g).", + "bbox": [ + 212, + 380, + 797, + 561 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/59fa1358f17687797ea4aa3a662abffb9ff0bfe01aaebfcd692437afc70c3cff.jpg", + "image_caption": [ + "Fig. 3: a) The proposed cross-domain attention-guided discriminator. b) Feature map visualizations without (w/o) and with (w) the discriminator head. The inclusion of the discriminator head aids in refining the overall results, and the discriminator training (as shown in c) benefits from the inclusion of the discriminator head, particularly with the cache $(M)$ component." + ], + "image_footnote": [], + "bbox": [ + 261, + 590, + 743, + 726 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Rawformer", + "bbox": [ + 655, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Cross-Domain Attention-Guided Discriminator", + "text_level": 1, + "bbox": [ + 215, + 146, + 648, + 161 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Leveraging batch statistics is an effective strategy to mitigate mode collapse and enhance diversity in GAN models [43]. Traditional minibatch discrimination techniques require large batch sizes [26], posing challenges due to GPU limitations. Inspired by [50], we address this by decoupling batch size from the minibatch discrimination approach. Our core innovation involves using a cache of previous discriminator features as an alternative to larger batch sizes. During training, batch features are stored in four separate caches $(M)$ ; real images from the source and target cameras, and fake images from both domains. This memory bank-style cache utilizes a composite discriminator with a head and body. The head captures batch statistics via batch normalization and two $3 \\times 3$ convolutional layers, processing a concatenated input of current and historical discriminator outputs from the cache, as shown in Fig. 3(a). The discriminator body includes a CQA block followed by stripped convolution to handle detailed global features across domains. This integration of batch statistics, caching, and attention mechanisms significantly enhances the discriminator's ability to manage multi-resolution image data and stabilize the overall training (Figs. 3 b and c).", + "bbox": [ + 212, + 167, + 787, + 424 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Self-Supervised Pre-Training", + "text_level": 1, + "bbox": [ + 215, + 441, + 501, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Self-supervised pre-training optimizes the initial weights of a network by engaging it in auxiliary tasks formulated from the inherent structure of the original data. In general, despite modifications in the cross-domain attention-guided discriminator, GAN training still suffers from mode collapse, where the generator repeatedly tends to produce a limited set of images. Thus, we opt to pre-train the generator of our Rawformer using self-supervised inpainting tasks to enhance initial weight optimization and mitigate this mode collapse problem. More details about the mode of pretraining is provided in the supplementary (Section 1) of the paper.", + "bbox": [ + 212, + 460, + 787, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 616, + 375, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluated our method using two types of datasets: 1) raw-to-raw datasets, which provide unpaired raw images for training and aligned raw images from source and target cameras for assessing unsupervised raw-to-raw translation accuracy (Sec. 4.3), and 2) raw-to-sRGB datasets, offering aligned raw and corresponding sRGB images for evaluating transformed raw image quality followed by learnable neural-based ISP rendering (Sec. 4.4). This evaluation mirrors real-world usage scenarios where raw images from a new camera are translated to the raw space of the camera used for ISP training. In all the experiments, raw images were demosaiced using the Menon algorithm [36] and normalized after black-level subtraction. During our method's training, each raw image was segmented into non-overlapping crops of $224 \\times 224$ pixels. Similarly, we used paired raw-sRGB non-overlapping crops of $224 \\times 224$ pixels to train the camera-specific LAN ISP [54] on each raw-to-sRGB dataset.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "G. Perevozhikov et al.", + "bbox": [ + 271, + 114, + 426, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 215, + 146, + 334, + 159 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We utilized two raw-to-raw datasets: 1) the Raw-to-Raw dataset [4] and 2) the NUS dataset [9], and three raw-to-sRGB datasets: 1) the Zurich raw-toRGB dataset [20], 2) the Samsung S7 dataset [44], and 3) the Mobile AIM21 dataset [19]. The Raw-to-Raw dataset [4] includes raw images from the Samsung Galaxy S9 and the iPhone X, featuring both unpaired and paired sets. The NUS dataset [9] comprises scenes captured by eight DSLR cameras. Following the evaluation in [4], we utilized the Nikon D5200 and Canon EOS 600D DSLR cameras from the NUS dataset [9]. Aligned paired raw images were generated using the same approach as in the Raw-to-Raw dataset [4]. In both datasets, paired raw images are dedicated for testing, while unpaired images are utilized for unsupervised training.", + "bbox": [ + 212, + 169, + 787, + 335 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The Zurich raw-to-RGB dataset [20] includes pairs of raw and sRGB images captured synchronously using a Huawei P20 smartphone camera and a Canon 5D Mark IV DSLR camera. The Samsung S7 dataset [44] comprises paired raw and sRGB images captured by the Samsung S7 smartphone's main camera. The Mobile AIM21 dataset [19] includes images captured using a Sony IMX586 mobile sensor and a Fujimel GFX100 DSLR, employing an advanced dense correspondence algorithm for matching due to imperfect alignment [52].", + "bbox": [ + 212, + 335, + 787, + 444 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Training details", + "text_level": 1, + "bbox": [ + 215, + 462, + 392, + 478 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Inspired by the effective fine-tuning capabilities demonstrated by the pre-trained BERT model for the text completion task [14], we initially trained each generator network in a self-supervised fashion for image completion. Initially, the generator was trained on $32 \\times 32$ pixel patches extracted from training images, with $40\\%$ of these patches being randomly masked out. The objective of the generator during this stage was to reconstruct the original unmasked image from its partially obscured version. Training was performed using the AdamW optimizer [34] with betas set to (0.9, 0.99) and a weight decay of 0.05 for 500 epochs to minimize the following pixel-wise loss function:", + "bbox": [ + 212, + 486, + 787, + 625 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p i x e l - w i s e}} = L _ {1} + (1 - S S I M) + V G G \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 636, + 784, + 655 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $L_{1}$ , $SSIM$ , $VGG$ refer to L1 loss, Structural Similarity Index (SSIM) [53], and VGG perceptual loss [30] between original unmasked image and the reconstructed image. The learning rate was initialized at 0.005, and the learning schedule was managed using the cosine annealing warm restarts strategy [33]. This self-supervised pre-training was shown to improve the raw mapping results (see Table 1).", + "bbox": [ + 212, + 657, + 787, + 748 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Thereafter, the weights of both generators were fine-tuned for our unpaired raw-to-raw translation task for an additional 500 epochs, while simultaneously training the discriminator networks. We use the Adam optimizer [29] with betas set to (0.5, 0.99) and a learning rate of 0.0001 for the discriminator networks and 0.00005 for the generator networks. The discriminator networks were optimized to minimize the following loss functions:", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Rawformer", + "bbox": [ + 653, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c1d10917b3b76d69b428eef7766a51a1a9a98a0241aea81ca61fee702e3d82c6.jpg", + "image_caption": [ + "Fig. 4: Shown are two images captured by Samsung S9 and iPhone X in sRGB (left) and two cropped patches from each image in raw (right). On the right, we show the input raw patch from the corresponding camera and the corresponding ground-truth raw patch from the other camera, along with the results by other methods. Our proposed Rawformer is better at preserving the domain consistent features." + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 756, + 316 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {A} ^ {d i s} = \\ell_ {g a n} \\left(D _ {A} \\left(G _ {B \\rightarrow A} (b)\\right), 0\\right) + \\ell_ {g a n} \\left(D _ {A} (a), 1\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 330, + 443, + 785, + 462 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {B} ^ {d i s} = \\ell_ {g a n} \\left(D _ {B} \\left(G _ {A \\rightarrow B} (a)\\right), 0\\right) + \\ell_ {g a n} \\left(D _ {B} (b), 1\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 330, + 483, + 785, + 501 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $a$ and $b$ represent images from camera $A$ and $B$ , respectively, while $G_{B\\to A}(b)$ and $G_{A\\to B}(a)$ are the translated images produced by each generator network, respectively. The labels 0 and 1 represent \"fake\" and \"real\" raw images, respectively, and $\\ell_{gan}(\\cdot)$ computes the cross-entropy loss.", + "bbox": [ + 212, + 508, + 784, + 568 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "During this stage, the weights of both generators were optimized to minimize the following loss function:", + "bbox": [ + 212, + 569, + 785, + 598 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\text {g e n}} = \\beta_ {1} \\left(\\mathcal {L} _ {A} ^ {\\text {g a n}} + \\mathcal {L} _ {B} ^ {\\text {g a n}}\\right) + \\beta_ {2} \\left(\\mathcal {L} _ {A} ^ {\\text {i d t}} + \\mathcal {L} _ {B} ^ {\\text {i d t}}\\right) + \\beta_ {3} \\left(\\mathcal {L} _ {A} ^ {\\text {c y c}} + \\mathcal {L} _ {B} ^ {\\text {c y c}}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 282, + 613, + 784, + 631 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\beta_{1} = 1, \\beta_{2} = 10, \\beta_{3} = 0.5, \\mathcal{L}_{A}^{gan} = \\ell_{gan}(D_{B}(G_{A\\rightarrow B}(a)),1),$ $\\mathcal{L}_A^{idt} = \\mathcal{L}_{pixel - wise}(G_{B\\to A}(a),a)$ , and $\\mathcal{L}_A^{cyc} = \\mathcal{L}_{pixel - wise}(G_{B\\to A}(G_{A\\to B}(a)),a)$ . Note that at inference time, only a single generator network (e.g., $G_{A\\rightarrow B}$ ) is required to translate raw images captured by the source camera (e.g., $A$ ) to the target camera (e.g., $B$ ).", + "bbox": [ + 212, + 637, + 787, + 715 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Raw-to-Raw Translation Results", + "text_level": 1, + "bbox": [ + 214, + 738, + 532, + 752 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We compare our raw-to-raw translation results with those of the recent semi-supervised raw-to-raw mapping (SSRM) method [4]. Additionally, we compare our method against several established unsupervised generic I2I methodologies, including UVCGANv2 [51], U-GAT-IT [28], UNIT [32], and the classic CycleGAN [62]. All other methods were trained on the same training data used for", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "G. Perevozchikov et al.", + "bbox": [ + 271, + 114, + 424, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/fea6e8e59d189ab7e7719b8f06b6035abd5214f87577e398cfa3875961414df4.jpg", + "table_caption": [ + "Table 1: Raw-to-raw translation results on NUS dataset [9]. The mapping results of Nikon D5200's raw images to Canon EOS 600D DSLR camera and vice versa using our method and other methods are shown. Here, $\\ell_{a} = L_{1}$ , $\\ell_{b} = L_{1} + VGG$ , $\\ell_{c} = \\mathcal{L}_{pixel - wise}$ (Eq. 3), w, and w/o denote the results with and without training on a corresponding configuration, respectively. Additionally, we show the GFLOPs of each method for the $256\\times 256$ input image. Best results are in bold." + ], + "table_footnote": [], + "table_body": "
MethodsCanon-to-NikonNikon-to-Canon
PSNR↑SSIM↑MAE↓(x1e-2)ΔE↓PSNR↑SSIM↑MAE↓(x1e-2)ΔE↓
CycleGAN [62]27.320.833.2115.4026.810.823.3311.4936.8
UNIT [32]25.550.813.1718.2224.730.793.2318.6447.3
U-GAT-IT [28]29.410.873.027.1227.860.853.298.0162.9
SSRM [4]32.360.932.416.2130.810.932.475.9537.1
UVCGANv2 [51]37.110.962.324.3437.290.962.284.2869.6
Ours (w/o self-supervised pre-training; ℓe loss)37.130.962.294.6537.270.962.314.4652.7
Ours (w/o self-supervised pre-training; ℓo loss)37.940.972.213.9637.990.962.233.9352.7
Ours (w/o self-supervised pre-training; ℓe loss)38.730.972.193.6438.260.972.223.7252.7
Ours (w self-supervised pre-training; ℓe loss)41.890.981.462.0441.370.981.412.5352.7
", + "bbox": [ + 259, + 239, + 743, + 330 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/a49a48c4591448973b664eab28aeee10cb9054ff50a1e65afd29602fe959b238.jpg", + "table_caption": [ + "Table 2: Raw-to-raw translation results on the Raw-to-Raw dataset [4]. Shown are the mapping results of Samsung Galaxy S9's raw images to iPhone X and vice versa using our method and other methods. Additionally, the total number of parameters of each method is shown in millions (M), and inference time in milliseconds (ms) without optimization on NVIDIA GeForce RTX 4090 GPU. Results on inference time with optimization are in supplementary." + ], + "table_footnote": [], + "table_body": "
MethodsSamsung-to-iPhoneiPhone-to-Samsung
PSNR↑SSIM↑MAE↓(×1e-2)ΔE↓PSNR↑SSIM↑MAE↓(×1e-2)ΔE↓
CycleGAN [62]24.630.715.4014.7124.370.734.3412.8821.719
UNIT [32]23.910.705.1513.2326.420.784.2713.1632.229
U-GAT-IT [28]28.220.893.115.9831.040.892.395.0434.435
SSRM [4]29.650.892.226.3228.580.902.436.5320.218
UVCGANv2 [51]36.320.942.094.2136.460.922.024.7332.631
Ours40.980.971.332.0941.480.981.211.9926.126
", + "bbox": [ + 259, + 441, + 743, + 523 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "our method, for 500 epochs. Since SSRM is not entirely unsupervised, we incorporated 22 image pairs from the paired subset for its training, as discussed in [4]. Fig. 4 shows qualitative comparisons on the Raw-to-Raw dataset [4], while Tables 1 and 2 present the quantitative results on the NUS dataset [9] and the Raw-to-Raw dataset, respectively. In Tables 1 and 2, we adopted the same quantitative metrics used in [4]: peak signal-to-noise ratio (PSNR), SSIM, mean absolute error (MAE), and $\\Delta E$ 2000 [46]. As can be seen in both the tables, our Rawformer significantly outperforms the alternative methods across all the quantitative metrics by a large margin (around $+12$ dB against SSRM [4]).", + "bbox": [ + 215, + 555, + 787, + 691 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation on Low-Light Images: Translating raw-to-raw for low-light images presents challenges due to various variations among cameras, including differences in raw normalization and noise characteristics [2,4]. To further prove the efficacy of our method against other methods on low-light images, we selected a subset of 15 paired low-light images from the Raw-to-Raw dataset [4]. The results on this subset are presented in Table 3, where our proposed method significantly outperforms SSRM [4], and UVCGANv2 [51] for translating low-light raw images captured by an iPhone X to the Samsung S9's camera.", + "bbox": [ + 215, + 719, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Rawformer", + "bbox": [ + 655, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f24c88d4eb227a1b29742086fbd59bb66d550c827c7849a3bedac3bea665d48b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 171, + 542, + 244 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/fef6e1b78374d75b5182c7477a27f8f62f4afa775fd9514c5a6a503b4cc6e329.jpg", + "table_caption": [ + "Table 3: Raw-to-Raw dataset [4] results on low-light raw images (a subset)." + ], + "table_footnote": [], + "table_body": "
MethodPSNR↑SSIM↑
SSRM [4]23.870.85
UVCGANv2 [51]35.220.94
Ours40.210.97
", + "bbox": [ + 594, + 180, + 751, + 236 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/58319ddbbb8177a9a83b068f576869987d666ac2cb5a54cc82383b5f2fa49bbe.jpg", + "table_caption": [ + "Table 4: Ablation results on the impact of various configurations of the generator model on the NUS dataset [9]. 'CQA', 'Style', 'SPFN' and 'CUp-CDown' refers to the condensed query attention, style modulator, scale perceptive feed-forward network, and the composite up and downsampler, respectively." + ], + "table_footnote": [], + "table_body": "
ConfigurationModulesCanon-to-NikonNikon-to-Canon
BaselineCQAStyleSPFNCUp-CdownPSNR↑SSIM↑PSNR↑SSIM↑
G1✓37.230.9637.310.96
G2✓✓38.210.9637.970.96
G3✓✓✓39.080.9738.990.96
G4✓✓✓✓39.720.9739.680.97
G5✓✓✓✓✓39.930.9739.860.97
", + "bbox": [ + 331, + 325, + 669, + 392 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/823ba6e03b87b699d9223054be375f096edc6cc1dc7c8097f14b3cb0746c80c9.jpg", + "table_caption": [ + "Table 5: Ablation results on the impact of various discriminator model configurations on the NUS dataset [9]. 'Baseline' refers to the $G_{5}$ config. from Tab. 4." + ], + "table_footnote": [], + "table_body": "
ConfigurationModulesCanon-to-NikonNikon-to-Canon
BaselineCQABatchnormDisc. headPSNR↑SSIM↑PSNR↑SSIM↑
D1✓39.930.9739.860.97
D2✓✓40.120.9740.090.97
D3✓✓✓40.240.9740.210.97
D4✓✓✓✓41.890.9841.370.98
", + "bbox": [ + 321, + 445, + 681, + 510 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablation Studies: We conducted a set of ablation studies to validate the impact of the proposed components in the generator and discriminator networks. In this series of experiments, we trained our Rawformer on the NUS dataset [9], in addition to a baseline model trained using the same training settings as our Rawformer. The baseline model comprises a UNet architecture with channel and spatial attention [50] for the generator, and a PatchGAN discriminator [23]. As demonstrated in Table 4, the integration of the proposed modules into the baseline consistently improves the network's capabilities, showcasing an enhancement in overall accuracy. We also conducted further ablation studies that explore various configurations of the discriminator, as shown in Table 5. The results clearly underscore the improvement by the inclusion of each proposed component in the discriminator network (attention, batch normalization, and discriminator head (Disc. head)) in improving the accuracy of our raw-to-raw translation.", + "bbox": [ + 212, + 537, + 787, + 734 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Raw Translation for Learnable ISP", + "text_level": 1, + "bbox": [ + 212, + 755, + 550, + 768 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To comprehend the practical application of our model in real-world scenarios, we utilized three datasets: Zurich raw-to-RGB, Samsung S7 ISP, and Mobile AIM21, encompassing various source and target domain combinations. Specifically, for each dataset, we trained the LAN ISP [54] on a single paired raw-to-sRGB", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "G. Perevozhikov et al.", + "bbox": [ + 271, + 114, + 424, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/76fe4056b37f62850408284510c70eff7e584ff7dfd2bb33cb5eeffcef6861e3.jpg", + "table_caption": [ + "Table 6: Cross-camera raw-to-sRGB results (PSNR/SSIM) comparing: a) UVCGANv2 [51] with b) our Rawformer. Each entry compares ground-truth sRGB images generated by the function: $ISP_{B}(F^{B}(raw_{A}))$ , where $F^{B}$ denotes translation of raw images, $raw_{A}$ , captured by camera $A$ to camera $B$ . $ISP_{B}(\\cdot)$ produces sRGB images using the LAN ISP [54] trained on raw images from camera $B$ . Diagonal results (in red) represent the camera-specific ISP case and are the sRGB images produced by the ISP model tested on the same camera used for training. Best cross-camera results are in bold." + ], + "table_footnote": [], + "table_body": "
Training camera Aa) UVCGANv2 [51]+ LAN ISP [54]b) Our Rawformer+ LAN ISP [54]
Testing camera BTesting camera B
ZRR [20]AIM21 [19]S7 [44]ZRR [20]AIM21 [19]S7 [44]
ZRR [20]19.46/0.7317.72/0.7017.81/0.7019.46/0.7318.97/0.7219.02/0.72
AIM21 [19]22.14/0.7923.48/0.8722.39/0.7923.25/0.8623.48/0.8723.08/0.86
S7 [44]21.07/0.8020.93/0.7922.16/0.8122.08/0.8022.03/0.8022.16/0.81
", + "bbox": [ + 305, + 268, + 699, + 328 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/4ede32badba377a435c3de8c64be3f47c63a858eae6aa6126229fd9af633967a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 263, + 345, + 328, + 396 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3e01e560cde71fd95bd3f329f789760652fded4245422acf2ad280772bb33d92.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 263, + 398, + 326, + 449 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/dcd909ffc277820f7dda4f11df6b9013f8052be6e0a70796e18077d5b49f019c.jpg", + "image_caption": [ + "GT (ZRR)", + "Fig.5: ISP rendering results with our raw translation on various datasets. Each set includes ground-truth (GT) raw-sRGB paired images and LAN ISP [54] results on mapped raw images from different cameras. S7, ZRR, and AIM stand for the Samsung S7 ISP dataset [44], Zurich raw-to-RGB dataset [20], and Mobile AIM21 dataset [19], respectively. The shown results are consistent with the ground-truth, demonstrating the proficiency of our model." + ], + "image_footnote": [], + "bbox": [ + 263, + 450, + 326, + 501 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/cc12e9be4c46b4678bb40efe836fe993c0de6061f6ff54df39251a644a3fcf77.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 345, + 395, + 396 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/583ef5c178ee8ff30e2d53a81b446b9e141d580b87d2bed813cc9ca57ace131c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 398, + 395, + 449 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/67c3c27f66d5c1901334932999c69d01178d2afbdb08a36002b78c426032f964.jpg", + "image_caption": [ + "$\\mathrm{ISP}_{\\mathrm{AIM}}(\\mathrm{F}^{\\mathrm{AIM}}(\\mathrm{raw}_{\\mathrm{RRB}}))$" + ], + "image_footnote": [], + "bbox": [ + 331, + 450, + 395, + 501 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/dfbc321511caa4630efaee585e2a330e6d88e47eb7641538feeffd9cbc89e646.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 400, + 345, + 465, + 396 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e7b343cc1bc9bf1534ede3f48fdba81e500484017e621d12832dc8da9905ad38.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 400, + 398, + 465, + 449 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/dacdc8543859dd5619c7db16497000895bda0bf7d8950fc158c9b40405ec7e42.jpg", + "image_caption": [ + "GT (AIM)" + ], + "image_footnote": [], + "bbox": [ + 400, + 450, + 465, + 501 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/07b4afa3c84b206ae4af3f2bd5f880f010dd2d3490dc2e4da023c872fc36f40b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 467, + 345, + 532, + 396 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/14725fd33bb29868313aeb7e1d722c9cfbac60164554d9f35bdbd6994b88f760.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 467, + 398, + 532, + 449 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/90ae8e7a255b52f2d6b669fcfa778beb479bbc674b03d8775b23a06cde21b718.jpg", + "image_caption": [ + "$\\mathrm{ISP}_{\\mathrm{ZRR}}(\\mathrm{F}^{\\mathrm{ZRR}}(\\mathrm{raw}_{\\mathrm{AIM}}))\\mathrm{ISP}_{\\mathrm{S7}}(\\mathrm{F}^{\\mathrm{ST}}(\\mathrm{raw}_{\\mathrm{AIM}}))$" + ], + "image_footnote": [], + "bbox": [ + 467, + 450, + 532, + 501 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/368226994a734226f180796668fb2963f1018452846f068438ecc11bdcec71b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 345, + 669, + 396 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7353958554e1f0934633aadf26862e41d399ef2e62a85bcd095ac8d9d85de26f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 398, + 669, + 449 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/10526664d7c0d64cbd22034a159e4e93d756bb7bf4de4e0adf7e1bd0014ddc06.jpg", + "image_caption": [ + "GT (S7)" + ], + "image_footnote": [], + "bbox": [ + 604, + 450, + 669, + 501 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b433f06f08ba8e3813ace1980b87e73083f2e0221efdfa4f8060eac4df548018.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 673, + 345, + 738, + 396 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c5e0d7af0302d321cbd6b955f02cfc38d5a140143f2dc9f51ef0c6eca8357b39.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 673, + 398, + 738, + 449 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9b87cb02fc7aaf51264c2638a34a16efb881db4db11da08a3e01c91d8a369b9d.jpg", + "image_caption": [ + "$\\mathrm{ISP}_{\\mathrm{ZRR}}(\\mathrm{F}^{\\mathrm{ZRR}}(\\mathrm{raw}_{S7}))$" + ], + "image_footnote": [], + "bbox": [ + 673, + 450, + 738, + 501 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "dataset, where raw images originated from a single camera. Then, we transferred all raw images from other cameras in the other datasets using our Rawformer before rendering with the trained LAN model. We report the PSNR and SSIM of the rendered sRGB images compared to the ground-truth sRGB images for each set as shown in Table 6. This includes the results of camera-specific LAN ISP models trained and tested on the same datasets. The camera-specific results represent the optimal results achievable by the neural-based ISP, which requires a paired raw-SRGB dataset for training on each camera model. As can be seen, our unsupervised raw-to-raw translation method demonstrates robust improvement, exhibiting only a marginal $3\\%$ reduction in accuracy compared to the camera-specific LAN model trained and tested on the same camera. This underscores the effectiveness of our method in addressing real-world cross-domain broadcasting challenges. See Fig. 5 for qualitative examples.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Rawformer", + "bbox": [ + 655, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Comparison with Few-Shot Domain Adaptation", + "text_level": 1, + "bbox": [ + 215, + 146, + 653, + 162 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We further compared our Rawformer against Prabhakar's Few-Shot Domain Adaptation (PDA) [40]. Specifically, we conducted a 10-shot domain adaptation from the Samsung S7 ISP dataset's camera [44] to the Zurich raw-to-RGB dataset's camera [20], and vice-versa. The PDA model underwent initial training for 10 epochs using the Samsung S7 ISP dataset. Subsequently, domain adaptation was performed on 10 images from the Zurich raw-to-RGB dataset's camera over 4 epochs. This process resulted in performance metrics of 21.07 dB, and 17.81 dB in terms of PSNR and 0.80 and 0.70 for SSIM for our Rawformer, representing approximately $11\\%$ superior performance compared to PDA (see Table 7).", + "bbox": [ + 212, + 170, + 787, + 321 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/61e4669a55fe4b8a6d24e9a7056ea702ab774124a2d414ce677427fd2b02392b.jpg", + "table_caption": [ + "Table 7: Comparison with PDA [40]. S7 and ZRR denote the Samsung S7 ISP [44] and the Zurich Raw-to-RGB [20] datasets, respectively. Results are produced after translating each raw image to the target camera using PDA and our Rawformer, subsequently rendering with a LAN ISP [54] trained on the target camera's raw images." + ], + "table_footnote": [], + "table_body": "
MethodS7-to-ZRRZRR-to-S7
PSNR↑SSIM↑MAE↓ (×1e-2)ΔE↓PSNR↑SSIM↑MAE↓ (×1e-2)ΔE↓
PDA [40]20.330.799.4113.9917.120.6918.0315.86
Ours22.080.808.1810.2719.020.7215.2711.74
", + "bbox": [ + 323, + 417, + 683, + 465 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 215, + 513, + 532, + 530 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we presented a new approach, Rawformer, for raw-to-raw image translation using unpaired data. Rawformer sets a new standard for performance on real camera datasets, showcasing its superior capabilities in raw-to-raw translation when measured against other alternatives. The proposed method reduces the need for the extensive collection of raw-to-sRGB paired datasets to re-train learnable ISPs for new cameras, thereby decreasing associated costs. Our results indicate that existing raw-to-sRGB paired datasets can reliably train neural-based learnable ISPs, which can then be utilized to render raw images taken by unseen cameras, lacking such unpaired data. We believe that our approach provides valuable insights for learnable ISP, raw image translation research, and applications. Our method (without optimization) runs at 26 milliseconds per frame on GPU ( $\\sim$ 1 second on CPU), which may be impractical for real-time rendering in camera previews on devices with limited computational power. Future work will focus on introducing lighter models capable of real-time performance on CPU.", + "bbox": [ + 212, + 545, + 787, + 772 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 215, + 794, + 393, + 811 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work was partly supported by The Alexander von Humboldt Foundation.", + "bbox": [ + 215, + 824, + 779, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "G. Perevozhikov et al.", + "bbox": [ + 271, + 114, + 424, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. A Sharif, S., Naqvi, R.A., Biswas, M.: Beyond joint demosaicking and denoising: An image processing pipeline for a pixel-bin image sensor. In: CVPR (2021)", + "2. Abdelhamed, A., Lin, S., Brown, M.S.: A high-quality denoising dataset for smartphone cameras. In: CVPR (2018)", + "3. Affi, M., Abdelhamed, A., Abuolaim, A., Punnappurath, A., Brown, M.S.: CIE XYZ net: Unprocessing images for low-level computer vision tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(9), 4688-4700 (2021)", + "4. Afifi, M., Abuolaim, A.: Semi-supervised raw-to-raw mapping. In: BMVC (2021)", + "5. Affi, M., Brown, M.S.: Sensor-independent illumination estimation for dnn models. In: BMVC (2019)", + "6. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016)", + "7. Barron, J.T., Tsai, Y.T.: Fast Fourier color constancy. In: CVPR (2017)", + "8. Chen, C., Chen, Q., Xu, J., Koltun, V.: Learning to see in the dark. In: CVPR (2018)", + "9. Cheng, D., Prasad, D.K., Brown, M.S.: Illuminant estimation for color constancy: why spatial-domain methods work and the role of the color distribution. JOSA A 31(5), 1049-1058 (2014)", + "0. Choi, Y., Choi, M., Kim, M., Ha, J.W., Kim, S., Choo, J.: StarGAN: Unified generative adversarial networks for multi-domain image-to-image translation. In: CVPR (2018)", + "1. Conde, M.V., Vasluianu, F., Vazquez-Corral, J., Timofte, R.: Perceptual image enhancement for smartphone real-time applications. In: CVPR (2023)", + "2. Dai, L., Liu, X., Li, C., Chen, J.: AWNet: Attentive wavelet network for image ISP. In: ECCV (2020)", + "3. Delbracio, M., Kelly, D., Brown, M.S., Milanfar, P.: Mobile computational photography: A tour. Annual Review of Vision Science 7, 571-604 (2021)", + "4. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)", + "5. Finlayson, G.D., Zhu, Y.: Designing color filters that make cameras more colorimetric. IEEE Transactions on Image Processing 30, 853-867 (2020)", + "6. Hasinoff, S.W., Sharlet, D., Geiss, R., Adams, A., Barron, J.T., Kainz, F., Chen, J., Levoy, M.: Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM Transactions on Graphics 35(6), 1-12 (2016)", + "7. He, X., Hu, T., Wang, G., Wang, Z., Wang, R., Zhang, Q., Yan, K., Chen, Z., Li, R., Xie, C., et al.: Enhancing RAW-to-sRGB with decoupled style structure in Fourier domain. In: AAAI (2024)", + "8. Herrmann, C., Bowen, R.S., Wadhwa, N., Garg, R., He, Q., Barron, J.T., Zabih, R.: Learning to autofocus. In: CVPR (2020)", + "9. Ignatov, A., Chiang, C.M., Kuo, H.K., Sycheva, A., Timofte, R.: Learned smartphone ISP on mobile NPUs with deep learning, mobile AI 2021 challenge: Report. In: CVPRW (2021)", + "20. Ignatov, A., Timofte, R., Ko, S.J., Kim, S.W., Uhm, K.H., Ji, S.W., Cho, S.J., Hong, J.P., Mei, K., Li, J., et al.: AIM 2019 challenge on raw to rgb mapping: Methods and results. In: ICCVW (2019)", + "21. Ignatov, A., Timofte, R., Zhang, Z., Liu, M., Wang, H., Zuo, W., Zhang, J., Zhang, R., Peng, Z., Ren, S., et al.: AIM 2020 challenge on learned image signal processing pipeline. In: ECCV (2020)" + ], + "bbox": [ + 225, + 172, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Rawformer", + "bbox": [ + 656, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "22. Ignatov, A., Van Gool, L., Timofte, R.: Replacing mobile camera ISP with a single deep learning model. In: CVPRW (2020)", + "23. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: CVPR (2017)", + "24. Jeong, W., Jung, S.W.: RAWtoBit: A fully end-to-end camera isp network. In: ECCV (2022)", + "25. Jiang, Y., Wronski, B., Mildenhall, B., Barron, J.T., Wang, Z., Xue, T.: Fast and high quality image denoising via malleable convolution. In: ECCV (2022)", + "26. Karras, T., Aila, T., Laine, S., Lehtinen, J.: Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196 (2017)", + "27. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: CVPR (2020)", + "28. Kim, J., Kim, M., Kang, H., Lee, K.: U-GAT-IT: Unsupervised generative attentional networks with adaptive layer-instance normalization for image-to-image translation. arXiv preprint arXiv:1907.10830 (2019)", + "29. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)", + "30. Ledig, C., Theis, L., Huszar, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., et al.: Photo-realistic single image superresolution using a generative adversarial network. In: CVPR (2017)", + "31. Liang, Z., Cai, J., Cao, Z., Zhang, L.: Cameranet: A two-stage framework for effective camera isp learning. IEEE Transactions on Image Processing 30, 2248-2262 (2021)", + "32. Liu, M.Y., Breuel, T., Kautz, J.: Unsupervised image-to-image translation networks. In: NeurIPS (2017)", + "33. Loshchilov, I., Hutter, F.: SGDR: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016)", + "34. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)", + "35. Maas, A.L., Hannun, A.Y., Ng, A.Y., et al.: Rectifier nonlinearities improve neural network acoustic models. In: ICML (2013)", + "36. Menon, D., Andriani, S., Calvagno, G.: Demosaicing with directional filtering and a posteriori decision. IEEE Transactions on Image Processing 16(1), 132-141 (2006)", + "37. Nguyen, R., Prasad, D.K., Brown, M.S.: Raw-to-raw: Mapping between image sensor color responses. In: CVPR (2014)", + "38. Pang, Y., Lin, J., Qin, T., Chen, Z.: Image-to-image translation: Methods and applications. IEEE Transactions on Multimedia 24, 3859-3881 (2021)", + "39. Park, T., Efros, A.A., Zhang, R., Zhu, J.Y.: Contrastive learning for unpaired image-to-image translation. In: ECCV (2020)", + "40. Prabhakar, K.R., Vinod, V., Sahoo, N.R., Babu, R.V.: Few-shot domain adaptation for low light raw image enhancement. arXiv preprint arXiv:2303.15528 (2023)", + "41. Punnappurath, A., Abuolaim, A., Abdelhamed, A., Levinshtein, A., Brown, M.S.: Day-to-night image synthesis for training nighttime neural isps. In: CVPR (2022)", + "42. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: International Conference on Medical image computing and computer-assisted intervention (2015)", + "43. Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen, X.: Improved techniques for training gans. Advances in neural information processing systems 29 (2016)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "G. Perevozchikov et al.", + "bbox": [ + 271, + 114, + 424, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "44. Schwartz, E., Giryes, R., Bronstein, A.M.: DeepISP: Toward learning an end-to-end image processing pipeline. IEEE Transactions on Image Processing 28(2), 912-923 (2018)", + "45. Seo, D., Punnappurath, A., Zhao, L., Abdelhamed, A., Tedla, S.K., Park, S., Choe, J., Brown, M.S.: Graphics2RAW: Mapping computer graphics images to sensor raw images. In: ICCV (2023)", + "46. Sharma, G., Wu, W., Dalal, E.N.: The CIEDE2000 color-difference formula: Implementation notes, supplementary test data, and mathematical observations. Color Research & Application 30(1), 21-30 (2005)", + "47. Sindelář, O., Sroubek, F.: Image deblurring in smartphone devices using built-in inertial measurement sensors. Journal of Electronic Imaging 22(1), 011003-011003 (2013)", + "48. Souza, M., Heidrich, W.: Crispnet: Color rendition isp net. arXiv preprint arXiv:2203.10562 (2022)", + "49. Tominaga, S., Nishi, S., Ohtera, R.: Measurement and estimation of spectral sensitivity functions for mobile phone cameras. Sensors 21(15), 4985 (2021)", + "50. Torbunov, D., Huang, Y., Yu, H., Huang, J., Yoo, S., Lin, M., Viren, B., Ren, Y.: UVCGAN: UNet vision transformer cycle-consistent GAN for unpaired image-to-image translation. In: WACV (2023)", + "51. Torbunov, D., Huang, Y., Yu, H., Huang, J., Yoo, S., Lin, M., Viren, B., Ren, Y.: UVCGAN v2: An improved cycle-consistent GAN for unpaired image-to-image translation. arXiv preprint arXiv:2303.16280 (2023)", + "52. Truong, P., Danelljan, M., Van Gool, L., Timofte, R.: Learning accurate dense correspondences and when to trust them. In: CVPR (2021)", + "53. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing 13(4), 600-612 (2004)", + "54. Wirzberger Raimundo, D., Ignatov, A., Timofte, R.: LAN: Lightweight attention-based network for raw-to-rgb smartphone image processing. In: CVPRW. pp. 807-815 (2022)", + "55. Wronski, B., Garcia-Dorado, I., Ernst, M., Kelly, D., Krainin, M., Liang, C.K., Levoy, M., Milanfar, P.: Handheld multi-frame super-resolution. ACM Transactions on Graphics (ToG) 38(4), 1-18 (2019)", + "56. Xing, Y., Qian, Z., Chen, Q.: Invertible image signal processing. In: CVPR (2021)", + "57. Yi, Z., Zhang, H., Tan, P., Gong, M.: DualGAN: Unsupervised dual learning for image-to-image translation. In: ICCV. pp. 2849-2857 (2017)", + "58. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5728-5739 (2022)", + "59. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: CycleISP: Real image restoration via improved data synthesis. In: CVPR (2020)", + "60. Zhang, Z., Wang, H., Liu, M., Wang, R., Zhang, J., Zuo, W.: Learning raw-to-srgb mappings with inaccurately aligned supervision. In: ICCV (2021)", + "61. Zhao, Y., Wu, R., Dong, H.: Unpaired image-to-image translation using adversarial consistency loss. In: ECCV (2020)", + "62. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: ICCV (2017)", + "63. Zhu, P., Abdal, R., Qin, Y., Wonka, P.: SEAN: Image synthesis with semantic region-adaptive normalization. In: CVPR (2020)" + ], + "bbox": [ + 212, + 146, + 787, + 829 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Rawformer", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Rawformer_ Unpaired Raw-to-Raw Translation for Learnable Camera ISPs/c3c8340d-834b-4e08-bdff-c9515d0d2660_model.json b/2024/Rawformer_ Unpaired Raw-to-Raw Translation for Learnable Camera ISPs/c3c8340d-834b-4e08-bdff-c9515d0d2660_model.json new file mode 100644 index 0000000000000000000000000000000000000000..770a1e2a0c422fa11db7cae79f5b306e2e77aee9 --- /dev/null +++ b/2024/Rawformer_ Unpaired Raw-to-Raw Translation for Learnable Camera ISPs/c3c8340d-834b-4e08-bdff-c9515d0d2660_model.json @@ -0,0 +1,2687 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.235, + 0.141, + 0.77, + 0.185 + ], + "angle": 0, + "content": "Rawformer: Unpaired Raw-to-Raw Translation for Learnable Camera ISPs" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.213, + 0.776, + 0.243 + ], + "angle": 0, + "content": "Georgy Perevozhikov1, Nancy Mehta1, Mahmoud Affifi2, and Radu Timofte1" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.255, + 0.721, + 0.284 + ], + "angle": 0, + "content": "1 Computer Vision Lab, CAIDAS & IFI, University of Würzburg, John Skilton Str. 4a, 97074 Würzburg, Germany" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.285, + 0.752, + 0.298 + ], + "angle": 0, + "content": "{georgii.perevozhikov,nancy.mehta,radu.timofte}@uni-wuerzburg.de" + }, + { + "type": "text", + "bbox": [ + 0.268, + 0.297, + 0.735, + 0.326 + ], + "angle": 0, + "content": "\\(^{2}\\) York University, 4700 Keele St, Toronto, Ontario, Canada, M3J 1P3 m.3afifi@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.363, + 0.74, + 0.655 + ], + "angle": 0, + "content": "Abstract. Modern smartphone camera quality heavily relies on the image signal processor (ISP) to enhance captured raw images, utilizing carefully designed modules to produce final output images encoded in a standard color space (e.g., sRGB). Neural-based end-to-end learnable ISPs offer promising advancements, potentially replacing traditional ISPs with their ability to adapt without requiring extensive tuning for each new camera model, as is often the case for nearly every module in traditional ISPs. However, the key challenge with the recent learning-based ISPs is the urge to collect large paired datasets for each distinct camera model due to the influence of intrinsic camera characteristics on the formation of input raw images. This paper tackles this challenge by introducing a novel method for unpaired learning of raw-to-raw translation across diverse cameras. Specifically, we propose Rawformer, an unsupervised Transformer-based encoder-decoder method for raw-to-raw translation. It accurately maps raw images captured by a certain camera to the target camera, facilitating the generalization of learnable ISPs to new unseen cameras. Our method demonstrates superior performance on real camera datasets, achieving higher accuracy compared to previous state-of-the-art techniques, and preserving a more robust correlation between the original and translated raw images. The codes and the pretrained models are available at https://github.com/gosha20777/rawformer." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.68, + 0.375, + 0.696 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.712, + 0.788, + 0.803 + ], + "angle": 0, + "content": "Unlike the high-quality digital single-lens reflex (DSLR) cameras, mobile phone cameras possess inherent limitations owing to their smaller sensor size and fixed lenses with limited optical capabilities [55]. The mobile camera image signal processor (ISP) aims to overcome these limitations by applying a series of carefully designed modules (e.g., [1,7,11,18,25,47]) to enhance the quality of the final image rendered in a standard color space, such as sRGB. These modules" + }, + { + "type": "page_footnote", + "bbox": [ + 0.216, + 0.811, + 0.386, + 0.826 + ], + "angle": 0, + "content": "* Corresponding Author" + }, + { + "type": "page_footnote", + "bbox": [ + 0.216, + 0.826, + 0.337, + 0.84 + ], + "angle": 0, + "content": "\\*\\* Now at Google" + }, + { + "type": "list", + "bbox": [ + 0.216, + 0.811, + 0.386, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.428, + 0.128 + ], + "angle": 0, + "content": "G. Perevozhikov et al." + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.144, + 0.357, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.144, + 0.495, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.144, + 0.634, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.144, + 0.774, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.239, + 0.789, + 0.366 + ], + "angle": 0, + "content": "Fig. 1: We introduce Rawformer, an unsupervised method for raw-to-raw translation that allows the utilization of pre-trained neural-based ISPs to process raw images captured by previously unseen cameras. Shown are raw and sRGB images processed by a neural-based ISP [54]. a) Raw image from an iPhone X rendered by the neural-based ISP trained on iPhone X's raw images. b) Raw image from an iPhone X rendered by the neural-based ISP trained on Samsung S9's raw images. c) iPhone X raw image translated to Samsung S9's raw space using our method, then processed by the Samsung S9 neural-based ISP. d) Raw image captured by Samsung S9's camera rendered by its native camera ISP, provided as a reference for visual comparison." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.784, + 0.444 + ], + "angle": 0, + "content": "work in tandem to optimize key parameters, such as sharpness, color accuracy, and noise reduction, resulting in a visually pleasing sRGB image from the raw image captured by the camera [13, 16]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.567 + ], + "angle": 0, + "content": "Despite yielding promising results, achieving compatibility among camera ISP modules for each new camera device requires extensive tuning and adjustment efforts to attain the desired image quality. This motivates the use of deep neural networks to replace individual camera ISP modules with a single neural-based unit capable of accomplishing the task [12,19,44]. Nonetheless, the training of such neural-based ISPs demands an extensive dataset consisting of raw images taken with a specific camera, accompanied by corresponding \"ground-truth\" sRGB images (typically generated by a high-quality DSLR camera) [19, 22]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.734 + ], + "angle": 0, + "content": "An issue emerges with the introduction of a new camera model, as pre-trained neural-based ISPs might encounter difficulty in accurately rendering raw images captured by the new camera (refer to Fig. 1(b)). This difficulty arises from the potential distinct characteristics of the new camera, such as sensor sensitivities, which influence the formation of its raw images [4,5]. Consequently, inconsistencies arise in the interpretation of raw RGB colors between cameras [4], hindering the generalization of neural-based ISPs to new cameras not encountered during training. To address this issue, re-training or fine-tuning on a new paired dataset becomes necessary, in which the raw images are sourced from the target new camera. This process mirrors similar challenges encountered in traditional ISP development for new cameras." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "An alternative strategy to address this challenge is domain adaptation, where the raw images from the new target camera are mapped to the raw space of the training camera to emulate those captured by the original camera used for training. In light of this, [4] proposed a semi-supervised learning framework for raw image translation, but the approach outlined in [4] inherits the drawbacks associated with the CNN-based architecture, notably the limited receptive field and inefficient encoding of global information. Thus, we present a fully unsupervised" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.656, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Rawformer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "Transformer-based raw-to-raw method, dubbed Rawformer, that maps raw images between camera models by efficiently encoding their global and semantic correlations. Our method eliminates the need for a paired raw-sRGB dataset from each new target camera, enabling the utilization of a neural-based ISP trained on a specific camera's raw images to process raw images taken by new cameras with different characteristics without necessitating any re-training; see Fig. 1(c). Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.262, + 0.788, + 0.337 + ], + "angle": 0, + "content": "- We propose Rawformer, a fully unsupervised encoder-decoder Transformer-based method designed for raw-to-raw translation facilitating the reuse of neural-based ISPs. Rawformer achieves state-of-the-art results in raw-to-raw translation, leading to better generalization for cross-camera learnable ISP rendering." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.338, + 0.788, + 0.397 + ], + "angle": 0, + "content": "- We introduce contextual-scale aware downsampler and upsampler blocks that efficiently summarize the local-global contextual details in mixed scale representations via its condensed query attention block and scale perceptive feed-forward network." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.398, + 0.788, + 0.429 + ], + "angle": 0, + "content": "- A novel cross-domain attention-driven discriminator is proposed along with a specialized discriminator head for stabilizing the network training." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.262, + 0.788, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.451, + 0.388, + 0.468 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.483, + 0.411, + 0.497 + ], + "angle": 0, + "content": "2.1 Neural-Based ISP" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.789, + 0.75 + ], + "angle": 0, + "content": "Several neural-based ISPs methods have been introduced [8,17,19,24,56,59,60], with the majority relying on CNN-based U-Net-like architecture [42]. These include earlier approaches, such as DeepISP [44] and PyNET [22], as well as more recent methods, such as MW-ISPNet [21], AW-Net [12], and LAN [54], which achieve improved image restoration quality by integrating discrete wavelet transform (DWT) and double attention modules (DAM) techniques. Such learnable ISPs are typically trained to map raw images captured by a specific camera model to target high-quality sRGB \"ground-truth\" images produced by some target camera ISP (usually with higher quality than the mobile native ISP; e.g., a DSLR camera [19,22]). A significant challenge for neural-based ISPs is the need for paired datasets of raw-sRGB images. Collecting such datasets is tedious and typically involves using specific equipment. Additionally, it may require post-processing techniques to align the raw images with the target sRGB images [22]. The challenge intensifies when a mobile phone manufacturer introduces a new camera model with different characteristics, which happens frequently [49], requiring the recollection of paired raw-sRGB datasets for each new camera model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.803, + 0.841 + ], + "angle": 0, + "content": "One solution is to render images to the sRGB space from the device-independent CIE XYZ space [3]. However, this requires calibration for each camera model, and accuracy depends on specific camera conditions (e.g., the Luther condition) [15]. In contrast, our work introduces an unsupervised raw-to-raw mapping technique that eliminates the need for paired raw-sRGB datasets for re-training or camera calibration. Our method maps raw images from new cameras to the camera raw" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.428, + 0.129 + ], + "angle": 0, + "content": "G. Perevozhikov et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.178 + ], + "angle": 0, + "content": "space utilized during the training of neural-based ISPs, enabling the reuse of pre-trained ISPs." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.2, + 0.563, + 0.216 + ], + "angle": 0, + "content": "2.2 Image-to-Image Domain Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.226, + 0.787, + 0.391 + ], + "angle": 0, + "content": "Numerous generative adversarial network (GAN)-based methods address unpaired image-to-image (I2I) translation challenges [38]. Notably, CycleGAN [62], DualGAN [57], UNIT [32], STARGAN [10], SEAN [63], and U-GAT-IT [28] focus on cycle consistency, employing dual generator networks guided by a cycle-consistency constraint. ACLGAN [61] and CUT [39] represent advancements in this field. ACLGAN relaxes the cycle consistency constraint, favoring an adversarial constraint, potentially enhancing translation quality. CUT utilizes a contrastive loss, eliminating the need for multiple generators and enabling faster training. ITTR further refines CUT's performance by altering the generator architecture. CycleGAN-based model, such as UVCGAN [50] incorporate U-Net-like generators with vision Transformers (ViTs)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.392, + 0.789, + 0.575 + ], + "angle": 0, + "content": "In computational photography, attempts have been made for domain adaptation, including synthesizing raw night images from raw daylight images to enhance neural-ISP rendering quality [41], few-shot domain adaptation for low-light image enhancement [40], and mapping graphics images to a target camera's raw space [45]. Raw-to-raw translation has received even limited attention. Classical approaches require a calibration object to learn global mapping [37], while recent work [4] proposes a semi-supervised CNN learning-based method using a small set of perfectly aligned images. Despite being effective at exploiting the local semantics, this CNN-based method lacks the potential to exploit the long-range/global semantics between different domains. In contrast, our method aims at solving the raw mapping by exploiting both local and global contextual information through fully unsupervised training, achieving state-of-the-art results." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.597, + 0.429, + 0.615 + ], + "angle": 0, + "content": "3 Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Existing raw-to-raw methodologies either rely on supervised approaches requiring paired datasets e.g., [37]) or semi-supervised approaches (e.g., [4]) with limited accuracy. Additionally, domain adaptation techniques (e.g., [39, 62]), which aim to extract local semantics between source and target domains, often fail to capture overall semantics and context-dependent relationships. Our proposed method, Rawformer, addresses these limitations by processing raw images in a fully unsupervised manner. Utilizing an encoder-decoder architecture within a CycleGAN framework, Rawformer effectively captures global semantics and maintains intrinsic consistency between translated images. The generator (Fig. 2(a)) in Rawformer employs contextual-scale aware downsampler and upsampler blocks to enhance content rendering and structural integrity. The discriminator (Fig. 3(a)) stabilizes training by incorporating key discriminator features, thereby preventing model collapse and reducing redundant computations. In the following, we provide a comprehensive elucidation of the proposed components." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Rawformer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.149, + 0.457, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.317, + 0.315, + 0.366, + 0.323 + ], + "angle": 0, + "content": "a) Generator" + }, + { + "type": "image", + "bbox": [ + 0.464, + 0.149, + 0.561, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.464, + 0.245, + 0.559, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.318, + 0.537, + 0.325 + ], + "angle": 0, + "content": "c) CQA block" + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.148, + 0.682, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.191, + 0.646, + 0.199 + ], + "angle": 0, + "content": "f) \\(\\mathrm{CUp}\\) block" + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.2, + 0.681, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.242, + 0.655, + 0.249 + ], + "angle": 0, + "content": "e) CDown block" + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.147, + 0.785, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.716, + 0.244, + 0.762, + 0.25 + ], + "angle": 0, + "content": "g) CSAU block" + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.255, + 0.716, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.319, + 0.65, + 0.325 + ], + "angle": 0, + "content": "d) SPFN block" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.341, + 0.788, + 0.412 + ], + "angle": 0, + "content": "Fig.2: a) Overview of the generator architecture (a) of Rawformer. The primary components of the generator: b) contextual-scale aware downsampler block (CSAD), c) condensed query attention block (CQA), d) scale perceptive feed-forward network (SPFN), e) composite downsampling block (CDown), f) composite upsampler (CUp) block, and g) contextual-scale aware upsampler block (CSAU)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.441, + 0.51, + 0.456 + ], + "angle": 0, + "content": "3.1 Generator's Encoder Network" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.469, + 0.788, + 0.621 + ], + "angle": 0, + "content": "As shown in Fig. 2(a), the encoder of the proposed generator in Rawformer comprises stacked contextual-scale aware downsampler (CSAD) blocks to bridge the semantic gap between the source and target domains. Each CSAD block, depicted in Fig. 2(b), offers a holistic approach of understanding the contextual dependencies efficiently between the domains by capturing both the global and local information via its condensed query attention (CQA) block, in parallel with a \\(3 \\times 3\\) depth-wise convolution layer. It further handles the intrinsic correlations among the extracted features in a multi-scale manner via maintaining a synergistic collaboration between the scale perceptive feed-forward network (SPFN), and the composite downsampler (CDown) block." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Condensed Query Attention Block: Self-attention is highly effective in modeling long-range contextual dependencies by comparing each token with all others, facilitating seamless information propagation within the network [58]. However, its computational demands pose challenges for tasks like unpaired raw-to-raw translation, where maintaining both visual quality and semantic consistency is crucial. To address these concerns and overcome the constraints of computation, we introduce the condensed query attention (CQA) block; see Fig. 2(c). Here, we make two key modifications to self-attention to enhance learning efficiency. Firstly, we spatially compress the query vector using average pooling followed by linear projection, and downsample the resulting feature map by a factor of \\( r \\) to confine the crucial information. This results in condensed queries, \\( (\\mathbf{Q}_c) \\) used for similarity comparison, as shown in Eq. 1. Secondly, to balance global and regional modeling with computational efficiency, attention is applied hor" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.426, + 0.128 + ], + "angle": 0, + "content": "G. Perevozhikov et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "iontally and vertically to eliminate redundant computations. The CQA block operation is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.315, + 0.189, + 0.735, + 0.204 + ], + "angle": 0, + "content": "\\[\nC Q A (\\mathbf {Q} _ {c}, \\mathbf {K}, \\mathbf {V}) = \\mathbf {A} _ {U}. (\\mathbf {A} _ {H} \\cdot \\mathbf {V}); \\quad \\mathbf {Q} _ {c} = W _ {p} (A v g P o o l (\\mathbf {Q}))\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.315, + 0.204, + 0.784, + 0.227 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} _ {H} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} _ {c} \\mathbf {K} ^ {T}}{\\sqrt {d}}\\right), \\quad \\mathbf {A} _ {U} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {Q} _ {c} ^ {T}}{\\sqrt {d}}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.251, + 0.785, + 0.371 + ], + "angle": 0, + "content": "where, the duet of key \\((\\mathbf{K})\\), and value \\((\\mathbf{V})\\) vectors are obtained through plain linear convolutions, and the dimensions of each of these features are shown in the Fig. 2(c). \\(W_{p}(.)\\) denotes a \\(1\\times 1\\) pointwise convolution, and \\(\\mathbf{A}_H\\), and \\(\\mathbf{A}_U\\) represents the compressed attention map between the key-condensed query pair horizontally and query-condensed query pair (for focusing on fine grained contextual information) vertically, respectively. The choice of operations and the reason for condensing the queries is investigated in the supplementary of the paper." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.785, + 0.527 + ], + "angle": 0, + "content": "Scale Perceptive Feed-forward Network: To proficiently capture the intricacies of the local image structure across various scales by encoding the information from neighboring pixel locations, we propose a scale perceptive feed-forward network (SPFN); see Fig. 2(d). Unlike the conventional feed-forward networks that integrate single-scale depth-wise convolutions [58], we emphasize on enhancing the locality and inter-correlations among the incoming raw features of different domains by inserting multi-scale depth-wise convolutions. Given an input tensor at layer \\(s\\) and scale \\(i\\), such that \\(\\mathbf{X}_s^i \\in \\mathbb{R}^{H^i \\times W^i \\times C}\\), SPFN is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.279, + 0.527, + 0.783, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{c} \\mathbf {X} _ {s} ^ {1} = \\phi \\left(W _ {d} ^ {3} \\left(W _ {p} ^ {1} (L N (\\mathbf {X} _ {s - 1}))\\right)\\right); \\mathbf {X} _ {s} ^ {2} = \\phi \\left(W _ {d} ^ {5} \\left(W _ {p} ^ {2} (L N (\\mathbf {X} _ {s - 1}))\\right)\\right), \\\\ \\mathbf {Y} _ {s} = W _ {s} ^ {3} \\left(\\mathbf {X} _ {o} ^ {1} (\\widehat {\\mathbf {C}}) \\mathbf {X} _ {s - 1} ^ {2}\\right) + \\mathbf {X} _ {s - 1}. \\end{array} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.784, + 0.631 + ], + "angle": 0, + "content": "Here, \\(\\phi\\) represents a LeakyReLU activation function [35], \\(W_{d}^{k}\\) represents a depthwise convolution with filter size \\(k\\), \\(LN\\) denotes the layer normalization [6], and \\(\\mathbb{C}\\) represents the concatenation operation. This enriches a more nuanced understanding of the underlying image features at different hierarchical levels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.652, + 0.785, + 0.789 + ], + "angle": 0, + "content": "Composite Down sampler: In general, the essence of downsampling is the retrieval of the multi-scale information in the image. Unlike the common encoder architectures for various image signal processing pipelines [31, 48] that deploy strided convolutions for downsampling, we design a hybrid strategy of leveraging the benefits of information exchange between different resolution features. To adaptively recover the underlying content information from the input domain at multiple scales, we design a composite down (CDown) block that intelligently combines pixel-unshuffle (to prevent the information loss) and strided convolution (to enhance the model's expressiveness) as shown in Fig. 2(e)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Style Modulator: Different cameras have their own distinctive characteristics to be handled or restored. To augment the expressiveness of Rawformer for" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.656, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Rawformer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.329 + ], + "angle": 0, + "content": "handling camera perturbations, we design a style modulator in the generator using extended pixel-wise ViT [51]. It aids in calibrating the decoder features and subsequently encourages the recovery of crucial details while translation. As demonstrated in Fig. 2(a), the style modulator applies style tokens in each contextual aware upsampler (CSAU) block, where every style token models the weight of the query vector of the CQA block to enrich the style specific context. This addition of style token at each stage in the decoder helps in the flexible adjustment of the feature maps. Unlike StyleGANv2 [27], where the style vectors are generated from the random prior, our approach resorts to learnable style token for conjecturing the important target style directly from the source domain, thus enhancing the consistency and control in raw image translation. More details on style modulation are provided in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.354, + 0.51, + 0.368 + ], + "angle": 0, + "content": "3.2 Generator's Decoder Network" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.799, + 0.563 + ], + "angle": 0, + "content": "Unlike the popular additive or multiplicative skip connections between the encoder-decoder, we also present a learnable linking layer (LLayer) as shown in the Fig. 2(a) to facilitate the feature preservation in the contextual scale aware upsampler (CSAU) block. The overall design of the CSAU block closely resembles that of the CSAD block, ensuring consistency and coherence in the network architecture in a bottom-up fashion. The primary distinction lies within the composite upsampling block, where for inducing the relevant spatial information into the target domain for the incoming features from the style modulator, we design a composite upsampler block (CUp). It embodies a parallel combination of pixel-shuffle (for preventing the checkerboard artifacts in the target domain) and deconvolution (for combating with aliasing) as shown in Fig. 2(f) and (g)." + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.592, + 0.744, + 0.727 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.737, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Fig. 3: a) The proposed cross-domain attention-guided discriminator. b) Feature map visualizations without (w/o) and with (w) the discriminator head. The inclusion of the discriminator head aids in refining the overall results, and the discriminator training (as shown in c) benefits from the inclusion of the discriminator head, particularly with the cache \\((M)\\) component." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.428, + 0.129 + ], + "angle": 0, + "content": "G. Perevozhikov et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.65, + 0.162 + ], + "angle": 0, + "content": "3.3 Cross-Domain Attention-Guided Discriminator" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.168, + 0.788, + 0.425 + ], + "angle": 0, + "content": "Leveraging batch statistics is an effective strategy to mitigate mode collapse and enhance diversity in GAN models [43]. Traditional minibatch discrimination techniques require large batch sizes [26], posing challenges due to GPU limitations. Inspired by [50], we address this by decoupling batch size from the minibatch discrimination approach. Our core innovation involves using a cache of previous discriminator features as an alternative to larger batch sizes. During training, batch features are stored in four separate caches \\((M)\\); real images from the source and target cameras, and fake images from both domains. This memory bank-style cache utilizes a composite discriminator with a head and body. The head captures batch statistics via batch normalization and two \\(3 \\times 3\\) convolutional layers, processing a concatenated input of current and historical discriminator outputs from the cache, as shown in Fig. 3(a). The discriminator body includes a CQA block followed by stripped convolution to handle detailed global features across domains. This integration of batch statistics, caching, and attention mechanisms significantly enhances the discriminator's ability to manage multi-resolution image data and stabilize the overall training (Figs. 3 b and c)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.442, + 0.502, + 0.459 + ], + "angle": 0, + "content": "3.4 Self-Supervised Pre-Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.6 + ], + "angle": 0, + "content": "Self-supervised pre-training optimizes the initial weights of a network by engaging it in auxiliary tasks formulated from the inherent structure of the original data. In general, despite modifications in the cross-domain attention-guided discriminator, GAN training still suffers from mode collapse, where the generator repeatedly tends to produce a limited set of images. Thus, we opt to pre-train the generator of our Rawformer using self-supervised inpainting tasks to enhance initial weight optimization and mitigate this mode collapse problem. More details about the mode of pretraining is provided in the supplementary (Section 1) of the paper." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.617, + 0.377, + 0.634 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We evaluated our method using two types of datasets: 1) raw-to-raw datasets, which provide unpaired raw images for training and aligned raw images from source and target cameras for assessing unsupervised raw-to-raw translation accuracy (Sec. 4.3), and 2) raw-to-sRGB datasets, offering aligned raw and corresponding sRGB images for evaluating transformed raw image quality followed by learnable neural-based ISP rendering (Sec. 4.4). This evaluation mirrors real-world usage scenarios where raw images from a new camera are translated to the raw space of the camera used for ISP training. In all the experiments, raw images were demosaiced using the Menon algorithm [36] and normalized after black-level subtraction. During our method's training, each raw image was segmented into non-overlapping crops of \\(224 \\times 224\\) pixels. Similarly, we used paired raw-sRGB non-overlapping crops of \\(224 \\times 224\\) pixels to train the camera-specific LAN ISP [54] on each raw-to-sRGB dataset." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Rawformer" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.335, + 0.16 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.788, + 0.336 + ], + "angle": 0, + "content": "We utilized two raw-to-raw datasets: 1) the Raw-to-Raw dataset [4] and 2) the NUS dataset [9], and three raw-to-sRGB datasets: 1) the Zurich raw-toRGB dataset [20], 2) the Samsung S7 dataset [44], and 3) the Mobile AIM21 dataset [19]. The Raw-to-Raw dataset [4] includes raw images from the Samsung Galaxy S9 and the iPhone X, featuring both unpaired and paired sets. The NUS dataset [9] comprises scenes captured by eight DSLR cameras. Following the evaluation in [4], we utilized the Nikon D5200 and Canon EOS 600D DSLR cameras from the NUS dataset [9]. Aligned paired raw images were generated using the same approach as in the Raw-to-Raw dataset [4]. In both datasets, paired raw images are dedicated for testing, while unpaired images are utilized for unsupervised training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.336, + 0.788, + 0.445 + ], + "angle": 0, + "content": "The Zurich raw-to-RGB dataset [20] includes pairs of raw and sRGB images captured synchronously using a Huawei P20 smartphone camera and a Canon 5D Mark IV DSLR camera. The Samsung S7 dataset [44] comprises paired raw and sRGB images captured by the Samsung S7 smartphone's main camera. The Mobile AIM21 dataset [19] includes images captured using a Sony IMX586 mobile sensor and a Fujimel GFX100 DSLR, employing an advanced dense correspondence algorithm for matching due to imperfect alignment [52]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.463, + 0.393, + 0.479 + ], + "angle": 0, + "content": "4.2 Training details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.789, + 0.625 + ], + "angle": 0, + "content": "Inspired by the effective fine-tuning capabilities demonstrated by the pre-trained BERT model for the text completion task [14], we initially trained each generator network in a self-supervised fashion for image completion. Initially, the generator was trained on \\(32 \\times 32\\) pixel patches extracted from training images, with \\(40\\%\\) of these patches being randomly masked out. The objective of the generator during this stage was to reconstruct the original unmasked image from its partially obscured version. Training was performed using the AdamW optimizer [34] with betas set to (0.9, 0.99) and a weight decay of 0.05 for 500 epochs to minimize the following pixel-wise loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.637, + 0.785, + 0.656 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p i x e l - w i s e}} = L _ {1} + (1 - S S I M) + V G G \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.749 + ], + "angle": 0, + "content": "where \\( L_{1} \\), \\( SSIM \\), \\( VGG \\) refer to L1 loss, Structural Similarity Index (SSIM) [53], and VGG perceptual loss [30] between original unmasked image and the reconstructed image. The learning rate was initialized at 0.005, and the learning schedule was managed using the cosine annealing warm restarts strategy [33]. This self-supervised pre-training was shown to improve the raw mapping results (see Table 1)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Thereafter, the weights of both generators were fine-tuned for our unpaired raw-to-raw translation task for an additional 500 epochs, while simultaneously training the discriminator networks. We use the Adam optimizer [29] with betas set to (0.5, 0.99) and a learning rate of 0.0001 for the discriminator networks and 0.00005 for the generator networks. The discriminator networks were optimized to minimize the following loss functions:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.426, + 0.127 + ], + "angle": 0, + "content": "G. Perevozchikov et al." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.144, + 0.758, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.329, + 0.788, + 0.4 + ], + "angle": 0, + "content": "Fig. 4: Shown are two images captured by Samsung S9 and iPhone X in sRGB (left) and two cropped patches from each image in raw (right). On the right, we show the input raw patch from the corresponding camera and the corresponding ground-truth raw patch from the other camera, along with the results by other methods. Our proposed Rawformer is better at preserving the domain consistent features." + }, + { + "type": "equation", + "bbox": [ + 0.331, + 0.444, + 0.786, + 0.463 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {A} ^ {d i s} = \\ell_ {g a n} \\left(D _ {A} \\left(G _ {B \\rightarrow A} (b)\\right), 0\\right) + \\ell_ {g a n} \\left(D _ {A} (a), 1\\right) \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.331, + 0.484, + 0.787, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {B} ^ {d i s} = \\ell_ {g a n} \\left(D _ {B} \\left(G _ {A \\rightarrow B} (a)\\right), 0\\right) + \\ell_ {g a n} \\left(D _ {B} (b), 1\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.785, + 0.569 + ], + "angle": 0, + "content": "where \\( a \\) and \\( b \\) represent images from camera \\( A \\) and \\( B \\), respectively, while \\( G_{B\\to A}(b) \\) and \\( G_{A\\to B}(a) \\) are the translated images produced by each generator network, respectively. The labels 0 and 1 represent \"fake\" and \"real\" raw images, respectively, and \\( \\ell_{gan}(\\cdot) \\) computes the cross-entropy loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.786, + 0.599 + ], + "angle": 0, + "content": "During this stage, the weights of both generators were optimized to minimize the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.284, + 0.614, + 0.785, + 0.632 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\text {g e n}} = \\beta_ {1} \\left(\\mathcal {L} _ {A} ^ {\\text {g a n}} + \\mathcal {L} _ {B} ^ {\\text {g a n}}\\right) + \\beta_ {2} \\left(\\mathcal {L} _ {A} ^ {\\text {i d t}} + \\mathcal {L} _ {B} ^ {\\text {i d t}}\\right) + \\beta_ {3} \\left(\\mathcal {L} _ {A} ^ {\\text {c y c}} + \\mathcal {L} _ {B} ^ {\\text {c y c}}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.638, + 0.788, + 0.716 + ], + "angle": 0, + "content": "where \\(\\beta_{1} = 1, \\beta_{2} = 10, \\beta_{3} = 0.5, \\mathcal{L}_{A}^{gan} = \\ell_{gan}(D_{B}(G_{A\\rightarrow B}(a)),1),\\) \\(\\mathcal{L}_A^{idt} = \\mathcal{L}_{pixel - wise}(G_{B\\to A}(a),a)\\), and \\(\\mathcal{L}_A^{cyc} = \\mathcal{L}_{pixel - wise}(G_{B\\to A}(G_{A\\to B}(a)),a)\\). Note that at inference time, only a single generator network (e.g., \\(G_{A\\rightarrow B}\\)) is required to translate raw images captured by the source camera (e.g., \\(A\\)) to the target camera (e.g., \\(B\\))." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.739, + 0.533, + 0.753 + ], + "angle": 0, + "content": "4.3 Raw-to-Raw Translation Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We compare our raw-to-raw translation results with those of the recent semi-supervised raw-to-raw mapping (SSRM) method [4]. Additionally, we compare our method against several established unsupervised generic I2I methodologies, including UVCGANv2 [51], U-GAT-IT [28], UNIT [32], and the classic CycleGAN [62]. All other methods were trained on the same training data used for" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.656, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Rawformer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Table 1: Raw-to-raw translation results on NUS dataset [9]. The mapping results of Nikon D5200's raw images to Canon EOS 600D DSLR camera and vice versa using our method and other methods are shown. Here, \\(\\ell_{a} = L_{1}\\), \\(\\ell_{b} = L_{1} + VGG\\), \\(\\ell_{c} = \\mathcal{L}_{pixel - wise}\\) (Eq. 3), w, and w/o denote the results with and without training on a corresponding configuration, respectively. Additionally, we show the GFLOPs of each method for the \\(256\\times 256\\) input image. Best results are in bold." + }, + { + "type": "table", + "bbox": [ + 0.261, + 0.24, + 0.744, + 0.332 + ], + "angle": 0, + "content": "
MethodsCanon-to-NikonNikon-to-Canon
PSNR↑SSIM↑MAE↓(x1e-2)ΔE↓PSNR↑SSIM↑MAE↓(x1e-2)ΔE↓
CycleGAN [62]27.320.833.2115.4026.810.823.3311.4936.8
UNIT [32]25.550.813.1718.2224.730.793.2318.6447.3
U-GAT-IT [28]29.410.873.027.1227.860.853.298.0162.9
SSRM [4]32.360.932.416.2130.810.932.475.9537.1
UVCGANv2 [51]37.110.962.324.3437.290.962.284.2869.6
Ours (w/o self-supervised pre-training; ℓe loss)37.130.962.294.6537.270.962.314.4652.7
Ours (w/o self-supervised pre-training; ℓo loss)37.940.972.213.9637.990.962.233.9352.7
Ours (w/o self-supervised pre-training; ℓe loss)38.730.972.193.6438.260.972.223.7252.7
Ours (w self-supervised pre-training; ℓe loss)41.890.981.462.0441.370.981.412.5352.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.347, + 0.788, + 0.43 + ], + "angle": 0, + "content": "Table 2: Raw-to-raw translation results on the Raw-to-Raw dataset [4]. Shown are the mapping results of Samsung Galaxy S9's raw images to iPhone X and vice versa using our method and other methods. Additionally, the total number of parameters of each method is shown in millions (M), and inference time in milliseconds (ms) without optimization on NVIDIA GeForce RTX 4090 GPU. Results on inference time with optimization are in supplementary." + }, + { + "type": "table", + "bbox": [ + 0.261, + 0.443, + 0.744, + 0.525 + ], + "angle": 0, + "content": "
MethodsSamsung-to-iPhoneiPhone-to-Samsung
PSNR↑SSIM↑MAE↓(×1e-2)ΔE↓PSNR↑SSIM↑MAE↓(×1e-2)ΔE↓
CycleGAN [62]24.630.715.4014.7124.370.734.3412.8821.719
UNIT [32]23.910.705.1513.2326.420.784.2713.1632.229
U-GAT-IT [28]28.220.893.115.9831.040.892.395.0434.435
SSRM [4]29.650.892.226.3228.580.902.436.5320.218
UVCGANv2 [51]36.320.942.094.2136.460.922.024.7332.631
Ours40.980.971.332.0941.480.981.211.9926.126
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.556, + 0.788, + 0.692 + ], + "angle": 0, + "content": "our method, for 500 epochs. Since SSRM is not entirely unsupervised, we incorporated 22 image pairs from the paired subset for its training, as discussed in [4]. Fig. 4 shows qualitative comparisons on the Raw-to-Raw dataset [4], while Tables 1 and 2 present the quantitative results on the NUS dataset [9] and the Raw-to-Raw dataset, respectively. In Tables 1 and 2, we adopted the same quantitative metrics used in [4]: peak signal-to-noise ratio (PSNR), SSIM, mean absolute error (MAE), and \\(\\Delta E\\) 2000 [46]. As can be seen in both the tables, our Rawformer significantly outperforms the alternative methods across all the quantitative metrics by a large margin (around \\(+12\\) dB against SSRM [4])." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Evaluation on Low-Light Images: Translating raw-to-raw for low-light images presents challenges due to various variations among cameras, including differences in raw normalization and noise characteristics [2,4]. To further prove the efficacy of our method against other methods on low-light images, we selected a subset of 15 paired low-light images from the Raw-to-Raw dataset [4]. The results on this subset are presented in Table 3, where our proposed method significantly outperforms SSRM [4], and UVCGANv2 [51] for translating low-light raw images captured by an iPhone X to the Samsung S9's camera." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.426, + 0.127 + ], + "angle": 0, + "content": "G. Perevozhikov et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.784, + 0.16 + ], + "angle": 0, + "content": "Table 3: Raw-to-Raw dataset [4] results on low-light raw images (a subset)." + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.172, + 0.543, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.596, + 0.181, + 0.753, + 0.237 + ], + "angle": 0, + "content": "
MethodPSNR↑SSIM↑
SSRM [4]23.870.85
UVCGANv2 [51]35.220.94
Ours40.210.97
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.258, + 0.787, + 0.314 + ], + "angle": 0, + "content": "Table 4: Ablation results on the impact of various configurations of the generator model on the NUS dataset [9]. 'CQA', 'Style', 'SPFN' and 'CUp-CDown' refers to the condensed query attention, style modulator, scale perceptive feed-forward network, and the composite up and downsampler, respectively." + }, + { + "type": "table", + "bbox": [ + 0.333, + 0.326, + 0.67, + 0.393 + ], + "angle": 0, + "content": "
ConfigurationModulesCanon-to-NikonNikon-to-Canon
BaselineCQAStyleSPFNCUp-CdownPSNR↑SSIM↑PSNR↑SSIM↑
G1✓37.230.9637.310.96
G2✓✓38.210.9637.970.96
G3✓✓✓39.080.9738.990.96
G4✓✓✓✓39.720.9739.680.97
G5✓✓✓✓✓39.930.9739.860.97
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.406, + 0.784, + 0.434 + ], + "angle": 0, + "content": "Table 5: Ablation results on the impact of various discriminator model configurations on the NUS dataset [9]. 'Baseline' refers to the \\( G_{5} \\) config. from Tab. 4." + }, + { + "type": "table", + "bbox": [ + 0.323, + 0.446, + 0.683, + 0.511 + ], + "angle": 0, + "content": "
ConfigurationModulesCanon-to-NikonNikon-to-Canon
BaselineCQABatchnormDisc. headPSNR↑SSIM↑PSNR↑SSIM↑
D1✓39.930.9739.860.97
D2✓✓40.120.9740.090.97
D3✓✓✓40.240.9740.210.97
D4✓✓✓✓41.890.9841.370.98
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.735 + ], + "angle": 0, + "content": "Ablation Studies: We conducted a set of ablation studies to validate the impact of the proposed components in the generator and discriminator networks. In this series of experiments, we trained our Rawformer on the NUS dataset [9], in addition to a baseline model trained using the same training settings as our Rawformer. The baseline model comprises a UNet architecture with channel and spatial attention [50] for the generator, and a PatchGAN discriminator [23]. As demonstrated in Table 4, the integration of the proposed modules into the baseline consistently improves the network's capabilities, showcasing an enhancement in overall accuracy. We also conducted further ablation studies that explore various configurations of the discriminator, as shown in Table 5. The results clearly underscore the improvement by the inclusion of each proposed component in the discriminator network (attention, batch normalization, and discriminator head (Disc. head)) in improving the accuracy of our raw-to-raw translation." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.756, + 0.551, + 0.77 + ], + "angle": 0, + "content": "4.4 Raw Translation for Learnable ISP" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To comprehend the practical application of our model in real-world scenarios, we utilized three datasets: Zurich raw-to-RGB, Samsung S7 ISP, and Mobile AIM21, encompassing various source and target domain combinations. Specifically, for each dataset, we trained the LAN ISP [54] on a single paired raw-to-sRGB" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.656, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Rawformer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.255 + ], + "angle": 0, + "content": "Table 6: Cross-camera raw-to-sRGB results (PSNR/SSIM) comparing: a) UVCGANv2 [51] with b) our Rawformer. Each entry compares ground-truth sRGB images generated by the function: \\(ISP_{B}(F^{B}(raw_{A}))\\), where \\(F^{B}\\) denotes translation of raw images, \\(raw_{A}\\), captured by camera \\(A\\) to camera \\(B\\). \\(ISP_{B}(\\cdot)\\) produces sRGB images using the LAN ISP [54] trained on raw images from camera \\(B\\). Diagonal results (in red) represent the camera-specific ISP case and are the sRGB images produced by the ISP model tested on the same camera used for training. Best cross-camera results are in bold." + }, + { + "type": "table", + "bbox": [ + 0.306, + 0.269, + 0.7, + 0.329 + ], + "angle": 0, + "content": "
Training camera Aa) UVCGANv2 [51]+ LAN ISP [54]b) Our Rawformer+ LAN ISP [54]
Testing camera BTesting camera B
ZRR [20]AIM21 [19]S7 [44]ZRR [20]AIM21 [19]S7 [44]
ZRR [20]19.46/0.7317.72/0.7017.81/0.7019.46/0.7318.97/0.7219.02/0.72
AIM21 [19]22.14/0.7923.48/0.8722.39/0.7923.25/0.8623.48/0.8723.08/0.86
S7 [44]21.07/0.8020.93/0.7922.16/0.8122.08/0.8022.03/0.8022.16/0.81
" + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.346, + 0.329, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.399, + 0.328, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.452, + 0.328, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.275, + 0.502, + 0.317, + 0.51 + ], + "angle": 0, + "content": "GT (ZRR)" + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.347, + 0.397, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.399, + 0.397, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.452, + 0.397, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.329, + 0.502, + 0.4, + 0.511 + ], + "angle": 0, + "content": "\\(\\mathrm{ISP}_{\\mathrm{AIM}}(\\mathrm{F}^{\\mathrm{AIM}}(\\mathrm{raw}_{\\mathrm{RRB}}))\\)" + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.347, + 0.466, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.399, + 0.466, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.452, + 0.466, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.414, + 0.502, + 0.453, + 0.51 + ], + "angle": 0, + "content": "GT (AIM)" + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.347, + 0.534, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.399, + 0.534, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.452, + 0.534, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.468, + 0.502, + 0.603, + 0.511 + ], + "angle": 0, + "content": "\\(\\mathrm{ISP}_{\\mathrm{ZRR}}(\\mathrm{F}^{\\mathrm{ZRR}}(\\mathrm{raw}_{\\mathrm{AIM}}))\\mathrm{ISP}_{\\mathrm{S7}}(\\mathrm{F}^{\\mathrm{ST}}(\\mathrm{raw}_{\\mathrm{AIM}}))\\)" + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.347, + 0.671, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.399, + 0.671, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.452, + 0.671, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.623, + 0.502, + 0.655, + 0.51 + ], + "angle": 0, + "content": "GT (S7)" + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.347, + 0.74, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.399, + 0.739, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.452, + 0.74, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.674, + 0.502, + 0.741, + 0.511 + ], + "angle": 0, + "content": "\\(\\mathrm{ISP}_{\\mathrm{ZRR}}(\\mathrm{F}^{\\mathrm{ZRR}}(\\mathrm{raw}_{S7}))\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.527, + 0.788, + 0.61 + ], + "angle": 0, + "content": "Fig.5: ISP rendering results with our raw translation on various datasets. Each set includes ground-truth (GT) raw-sRGB paired images and LAN ISP [54] results on mapped raw images from different cameras. S7, ZRR, and AIM stand for the Samsung S7 ISP dataset [44], Zurich raw-to-RGB dataset [20], and Mobile AIM21 dataset [19], respectively. The shown results are consistent with the ground-truth, demonstrating the proficiency of our model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.841 + ], + "angle": 0, + "content": "dataset, where raw images originated from a single camera. Then, we transferred all raw images from other cameras in the other datasets using our Rawformer before rendering with the trained LAN model. We report the PSNR and SSIM of the rendered sRGB images compared to the ground-truth sRGB images for each set as shown in Table 6. This includes the results of camera-specific LAN ISP models trained and tested on the same datasets. The camera-specific results represent the optimal results achievable by the neural-based ISP, which requires a paired raw-SRGB dataset for training on each camera model. As can be seen, our unsupervised raw-to-raw translation method demonstrates robust improvement, exhibiting only a marginal \\(3\\%\\) reduction in accuracy compared to the camera-specific LAN model trained and tested on the same camera. This underscores the effectiveness of our method in addressing real-world cross-domain broadcasting challenges. See Fig. 5 for qualitative examples." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.426, + 0.128 + ], + "angle": 0, + "content": "G. Perevozhikov et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.655, + 0.163 + ], + "angle": 0, + "content": "4.5 Comparison with Few-Shot Domain Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.171, + 0.788, + 0.323 + ], + "angle": 0, + "content": "We further compared our Rawformer against Prabhakar's Few-Shot Domain Adaptation (PDA) [40]. Specifically, we conducted a 10-shot domain adaptation from the Samsung S7 ISP dataset's camera [44] to the Zurich raw-to-RGB dataset's camera [20], and vice-versa. The PDA model underwent initial training for 10 epochs using the Samsung S7 ISP dataset. Subsequently, domain adaptation was performed on 10 images from the Zurich raw-to-RGB dataset's camera over 4 epochs. This process resulted in performance metrics of 21.07 dB, and 17.81 dB in terms of PSNR and 0.80 and 0.70 for SSIM for our Rawformer, representing approximately \\(11\\%\\) superior performance compared to PDA (see Table 7)." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.348, + 0.788, + 0.406 + ], + "angle": 0, + "content": "Table 7: Comparison with PDA [40]. S7 and ZRR denote the Samsung S7 ISP [44] and the Zurich Raw-to-RGB [20] datasets, respectively. Results are produced after translating each raw image to the target camera using PDA and our Rawformer, subsequently rendering with a LAN ISP [54] trained on the target camera's raw images." + }, + { + "type": "table", + "bbox": [ + 0.324, + 0.418, + 0.684, + 0.466 + ], + "angle": 0, + "content": "
MethodS7-to-ZRRZRR-to-S7
PSNR↑SSIM↑MAE↓ (×1e-2)ΔE↓PSNR↑SSIM↑MAE↓ (×1e-2)ΔE↓
PDA [40]20.330.799.4113.9917.120.6918.0315.86
Ours22.080.808.1810.2719.020.7215.2711.74
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.515, + 0.533, + 0.531 + ], + "angle": 0, + "content": "5 Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.546, + 0.788, + 0.773 + ], + "angle": 0, + "content": "In this work, we presented a new approach, Rawformer, for raw-to-raw image translation using unpaired data. Rawformer sets a new standard for performance on real camera datasets, showcasing its superior capabilities in raw-to-raw translation when measured against other alternatives. The proposed method reduces the need for the extensive collection of raw-to-sRGB paired datasets to re-train learnable ISPs for new cameras, thereby decreasing associated costs. Our results indicate that existing raw-to-sRGB paired datasets can reliably train neural-based learnable ISPs, which can then be utilized to render raw images taken by unseen cameras, lacking such unpaired data. We believe that our approach provides valuable insights for learnable ISP, raw image translation research, and applications. Our method (without optimization) runs at 26 milliseconds per frame on GPU (\\(\\sim\\)1 second on CPU), which may be impractical for real-time rendering in camera previews on devices with limited computational power. Future work will focus on introducing lighter models capable of real-time performance on CPU." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.795, + 0.394, + 0.812 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.825, + 0.781, + 0.84 + ], + "angle": 0, + "content": "This work was partly supported by The Alexander von Humboldt Foundation." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.116, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Rawformer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.174, + 0.785, + 0.202 + ], + "angle": 0, + "content": "1. A Sharif, S., Naqvi, R.A., Biswas, M.: Beyond joint demosaicking and denoising: An image processing pipeline for a pixel-bin image sensor. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.203, + 0.785, + 0.228 + ], + "angle": 0, + "content": "2. Abdelhamed, A., Lin, S., Brown, M.S.: A high-quality denoising dataset for smartphone cameras. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.229, + 0.785, + 0.27 + ], + "angle": 0, + "content": "3. Affi, M., Abdelhamed, A., Abuolaim, A., Punnappurath, A., Brown, M.S.: CIE XYZ net: Unprocessing images for low-level computer vision tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(9), 4688-4700 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.271, + 0.779, + 0.283 + ], + "angle": 0, + "content": "4. Afifi, M., Abuolaim, A.: Semi-supervised raw-to-raw mapping. In: BMVC (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.284, + 0.785, + 0.31 + ], + "angle": 0, + "content": "5. Affi, M., Brown, M.S.: Sensor-independent illumination estimation for dnn models. In: BMVC (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.785, + 0.337 + ], + "angle": 0, + "content": "6. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.338, + 0.724, + 0.35 + ], + "angle": 0, + "content": "7. Barron, J.T., Tsai, Y.T.: Fast Fourier color constancy. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.351, + 0.785, + 0.377 + ], + "angle": 0, + "content": "8. Chen, C., Chen, Q., Xu, J., Koltun, V.: Learning to see in the dark. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.785, + 0.418 + ], + "angle": 0, + "content": "9. Cheng, D., Prasad, D.K., Brown, M.S.: Illuminant estimation for color constancy: why spatial-domain methods work and the role of the color distribution. JOSA A 31(5), 1049-1058 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.42, + 0.785, + 0.459 + ], + "angle": 0, + "content": "0. Choi, Y., Choi, M., Kim, M., Ha, J.W., Kim, S., Choo, J.: StarGAN: Unified generative adversarial networks for multi-domain image-to-image translation. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.46, + 0.785, + 0.486 + ], + "angle": 0, + "content": "1. Conde, M.V., Vasluianu, F., Vazquez-Corral, J., Timofte, R.: Perceptual image enhancement for smartphone real-time applications. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.487, + 0.785, + 0.513 + ], + "angle": 0, + "content": "2. Dai, L., Liu, X., Li, C., Chen, J.: AWNet: Attentive wavelet network for image ISP. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.515, + 0.785, + 0.54 + ], + "angle": 0, + "content": "3. Delbracio, M., Kelly, D., Brown, M.S., Milanfar, P.: Mobile computational photography: A tour. Annual Review of Vision Science 7, 571-604 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.541, + 0.785, + 0.581 + ], + "angle": 0, + "content": "4. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.582, + 0.785, + 0.609 + ], + "angle": 0, + "content": "5. Finlayson, G.D., Zhu, Y.: Designing color filters that make cameras more colorimetric. IEEE Transactions on Image Processing 30, 853-867 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.61, + 0.785, + 0.649 + ], + "angle": 0, + "content": "6. Hasinoff, S.W., Sharlet, D., Geiss, R., Adams, A., Barron, J.T., Kainz, F., Chen, J., Levoy, M.: Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM Transactions on Graphics 35(6), 1-12 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.65, + 0.785, + 0.69 + ], + "angle": 0, + "content": "7. He, X., Hu, T., Wang, G., Wang, Z., Wang, R., Zhang, Q., Yan, K., Chen, Z., Li, R., Xie, C., et al.: Enhancing RAW-to-sRGB with decoupled style structure in Fourier domain. In: AAAI (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.785, + 0.718 + ], + "angle": 0, + "content": "8. Herrmann, C., Bowen, R.S., Wadhwa, N., Garg, R., He, Q., Barron, J.T., Zabih, R.: Learning to autofocus. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.719, + 0.785, + 0.758 + ], + "angle": 0, + "content": "9. Ignatov, A., Chiang, C.M., Kuo, H.K., Sycheva, A., Timofte, R.: Learned smartphone ISP on mobile NPUs with deep learning, mobile AI 2021 challenge: Report. In: CVPRW (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.76, + 0.785, + 0.799 + ], + "angle": 0, + "content": "20. Ignatov, A., Timofte, R., Ko, S.J., Kim, S.W., Uhm, K.H., Ji, S.W., Cho, S.J., Hong, J.P., Mei, K., Li, J., et al.: AIM 2019 challenge on raw to rgb mapping: Methods and results. In: ICCVW (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "21. Ignatov, A., Timofte, R., Zhang, Z., Liu, M., Wang, H., Zuo, W., Zhang, J., Zhang, R., Peng, Z., Ren, S., et al.: AIM 2020 challenge on learned image signal processing pipeline. In: ECCV (2020)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.174, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.426, + 0.127 + ], + "angle": 0, + "content": "G. Perevozchikov et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "22. Ignatov, A., Van Gool, L., Timofte, R.: Replacing mobile camera ISP with a single deep learning model. In: CVPRW (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "23. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.205, + 0.785, + 0.232 + ], + "angle": 0, + "content": "24. Jeong, W., Jung, S.W.: RAWtoBit: A fully end-to-end camera isp network. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.785, + 0.261 + ], + "angle": 0, + "content": "25. Jiang, Y., Wronski, B., Mildenhall, B., Barron, J.T., Wang, Z., Xue, T.: Fast and high quality image denoising via malleable convolution. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.262, + 0.785, + 0.289 + ], + "angle": 0, + "content": "26. Karras, T., Aila, T., Laine, S., Lehtinen, J.: Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.29, + 0.785, + 0.318 + ], + "angle": 0, + "content": "27. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.319, + 0.785, + 0.359 + ], + "angle": 0, + "content": "28. Kim, J., Kim, M., Kang, H., Lee, K.: U-GAT-IT: Unsupervised generative attentional networks with adaptive layer-instance normalization for image-to-image translation. arXiv preprint arXiv:1907.10830 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.36, + 0.785, + 0.388 + ], + "angle": 0, + "content": "29. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.389, + 0.785, + 0.431 + ], + "angle": 0, + "content": "30. Ledig, C., Theis, L., Huszar, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., et al.: Photo-realistic single image superresolution using a generative adversarial network. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.431, + 0.785, + 0.471 + ], + "angle": 0, + "content": "31. Liang, Z., Cai, J., Cao, Z., Zhang, L.: Cameranet: A two-stage framework for effective camera isp learning. IEEE Transactions on Image Processing 30, 2248-2262 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.501 + ], + "angle": 0, + "content": "32. Liu, M.Y., Breuel, T., Kautz, J.: Unsupervised image-to-image translation networks. In: NeurIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.529 + ], + "angle": 0, + "content": "33. Loshchilov, I., Hutter, F.: SGDR: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.785, + 0.557 + ], + "angle": 0, + "content": "34. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.558, + 0.785, + 0.586 + ], + "angle": 0, + "content": "35. Maas, A.L., Hannun, A.Y., Ng, A.Y., et al.: Rectifier nonlinearities improve neural network acoustic models. In: ICML (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.785, + 0.614 + ], + "angle": 0, + "content": "36. Menon, D., Andriani, S., Calvagno, G.: Demosaicing with directional filtering and a posteriori decision. IEEE Transactions on Image Processing 16(1), 132-141 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.785, + 0.642 + ], + "angle": 0, + "content": "37. Nguyen, R., Prasad, D.K., Brown, M.S.: Raw-to-raw: Mapping between image sensor color responses. In: CVPR (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.643, + 0.785, + 0.671 + ], + "angle": 0, + "content": "38. Pang, Y., Lin, J., Qin, T., Chen, Z.: Image-to-image translation: Methods and applications. IEEE Transactions on Multimedia 24, 3859-3881 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.672, + 0.785, + 0.699 + ], + "angle": 0, + "content": "39. Park, T., Efros, A.A., Zhang, R., Zhu, J.Y.: Contrastive learning for unpaired image-to-image translation. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.785, + 0.728 + ], + "angle": 0, + "content": "40. Prabhakar, K.R., Vinod, V., Sahoo, N.R., Babu, R.V.: Few-shot domain adaptation for low light raw image enhancement. arXiv preprint arXiv:2303.15528 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.785, + 0.756 + ], + "angle": 0, + "content": "41. Punnappurath, A., Abuolaim, A., Abdelhamed, A., Levinshtein, A., Brown, M.S.: Day-to-night image synthesis for training nighttime neural isps. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "42. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: International Conference on Medical image computing and computer-assisted intervention (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "43. Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen, X.: Improved techniques for training gans. Advances in neural information processing systems 29 (2016)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Rawformer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "44. Schwartz, E., Giryes, R., Bronstein, A.M.: DeepISP: Toward learning an end-to-end image processing pipeline. IEEE Transactions on Image Processing 28(2), 912-923 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.23 + ], + "angle": 0, + "content": "45. Seo, D., Punnappurath, A., Zhao, L., Abdelhamed, A., Tedla, S.K., Park, S., Choe, J., Brown, M.S.: Graphics2RAW: Mapping computer graphics images to sensor raw images. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.23, + 0.788, + 0.272 + ], + "angle": 0, + "content": "46. Sharma, G., Wu, W., Dalal, E.N.: The CIEDE2000 color-difference formula: Implementation notes, supplementary test data, and mathematical observations. Color Research & Application 30(1), 21-30 (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.271, + 0.788, + 0.313 + ], + "angle": 0, + "content": "47. Sindelář, O., Sroubek, F.: Image deblurring in smartphone devices using built-in inertial measurement sensors. Journal of Electronic Imaging 22(1), 011003-011003 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.313, + 0.788, + 0.34 + ], + "angle": 0, + "content": "48. Souza, M., Heidrich, W.: Crispnet: Color rendition isp net. arXiv preprint arXiv:2203.10562 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.34, + 0.788, + 0.368 + ], + "angle": 0, + "content": "49. Tominaga, S., Nishi, S., Ohtera, R.: Measurement and estimation of spectral sensitivity functions for mobile phone cameras. Sensors 21(15), 4985 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.367, + 0.788, + 0.408 + ], + "angle": 0, + "content": "50. Torbunov, D., Huang, Y., Yu, H., Huang, J., Yoo, S., Lin, M., Viren, B., Ren, Y.: UVCGAN: UNet vision transformer cycle-consistent GAN for unpaired image-to-image translation. In: WACV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.408, + 0.788, + 0.449 + ], + "angle": 0, + "content": "51. Torbunov, D., Huang, Y., Yu, H., Huang, J., Yoo, S., Lin, M., Viren, B., Ren, Y.: UVCGAN v2: An improved cycle-consistent GAN for unpaired image-to-image translation. arXiv preprint arXiv:2303.16280 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.449, + 0.788, + 0.476 + ], + "angle": 0, + "content": "52. Truong, P., Danelljan, M., Van Gool, L., Timofte, R.: Learning accurate dense correspondences and when to trust them. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.476, + 0.788, + 0.517 + ], + "angle": 0, + "content": "53. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing 13(4), 600-612 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.517, + 0.788, + 0.557 + ], + "angle": 0, + "content": "54. Wirzberger Raimundo, D., Ignatov, A., Timofte, R.: LAN: Lightweight attention-based network for raw-to-rgb smartphone image processing. In: CVPRW. pp. 807-815 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.557, + 0.788, + 0.599 + ], + "angle": 0, + "content": "55. Wronski, B., Garcia-Dorado, I., Ernst, M., Kelly, D., Krainin, M., Liang, C.K., Levoy, M., Milanfar, P.: Handheld multi-frame super-resolution. ACM Transactions on Graphics (ToG) 38(4), 1-18 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.612 + ], + "angle": 0, + "content": "56. Xing, Y., Qian, Z., Chen, Q.: Invertible image signal processing. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.612, + 0.788, + 0.639 + ], + "angle": 0, + "content": "57. Yi, Z., Zhang, H., Tan, P., Gong, M.: DualGAN: Unsupervised dual learning for image-to-image translation. In: ICCV. pp. 2849-2857 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.639, + 0.788, + 0.693 + ], + "angle": 0, + "content": "58. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5728-5739 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.693, + 0.788, + 0.722 + ], + "angle": 0, + "content": "59. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: CycleISP: Real image restoration via improved data synthesis. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.722, + 0.788, + 0.748 + ], + "angle": 0, + "content": "60. Zhang, Z., Wang, H., Liu, M., Wang, R., Zhang, J., Zuo, W.: Learning raw-to-srgb mappings with inaccurately aligned supervision. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.748, + 0.788, + 0.775 + ], + "angle": 0, + "content": "61. Zhao, Y., Wu, R., Dong, H.: Unpaired image-to-image translation using adversarial consistency loss. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.775, + 0.788, + 0.803 + ], + "angle": 0, + "content": "62. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.803, + 0.788, + 0.83 + ], + "angle": 0, + "content": "63. Zhu, P., Abdal, R., Qin, Y., Wonka, P.: SEAN: Image synthesis with semantic region-adaptive normalization. In: CVPR (2020)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.83 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Rawformer_ Unpaired Raw-to-Raw Translation for Learnable Camera ISPs/c3c8340d-834b-4e08-bdff-c9515d0d2660_origin.pdf b/2024/Rawformer_ Unpaired Raw-to-Raw Translation for Learnable Camera ISPs/c3c8340d-834b-4e08-bdff-c9515d0d2660_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..44a557a808c55c54cd1e4016a70205f200d6c8a1 --- /dev/null +++ b/2024/Rawformer_ Unpaired Raw-to-Raw Translation for Learnable Camera ISPs/c3c8340d-834b-4e08-bdff-c9515d0d2660_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:271722ee0a193f2e1efb3b6d120977357c4187eede8cf4d5e71150f519a2536b +size 7948168 diff --git a/2024/Rawformer_ Unpaired Raw-to-Raw Translation for Learnable Camera ISPs/full.md b/2024/Rawformer_ Unpaired Raw-to-Raw Translation for Learnable Camera ISPs/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dcd45a10fd0851a51eb871a05f12cc5910995d21 --- /dev/null +++ b/2024/Rawformer_ Unpaired Raw-to-Raw Translation for Learnable Camera ISPs/full.md @@ -0,0 +1,338 @@ +# Rawformer: Unpaired Raw-to-Raw Translation for Learnable Camera ISPs + +Georgy Perevozhikov1, Nancy Mehta1, Mahmoud Affifi2, and Radu Timofte1 + +1 Computer Vision Lab, CAIDAS & IFI, University of Würzburg, John Skilton Str. 4a, 97074 Würzburg, Germany + +{georgii.perevozhikov,nancy.mehta,radu.timofte}@uni-wuerzburg.de + +$^{2}$ York University, 4700 Keele St, Toronto, Ontario, Canada, M3J 1P3 m.3afifi@gmail.com + +Abstract. Modern smartphone camera quality heavily relies on the image signal processor (ISP) to enhance captured raw images, utilizing carefully designed modules to produce final output images encoded in a standard color space (e.g., sRGB). Neural-based end-to-end learnable ISPs offer promising advancements, potentially replacing traditional ISPs with their ability to adapt without requiring extensive tuning for each new camera model, as is often the case for nearly every module in traditional ISPs. However, the key challenge with the recent learning-based ISPs is the urge to collect large paired datasets for each distinct camera model due to the influence of intrinsic camera characteristics on the formation of input raw images. This paper tackles this challenge by introducing a novel method for unpaired learning of raw-to-raw translation across diverse cameras. Specifically, we propose Rawformer, an unsupervised Transformer-based encoder-decoder method for raw-to-raw translation. It accurately maps raw images captured by a certain camera to the target camera, facilitating the generalization of learnable ISPs to new unseen cameras. Our method demonstrates superior performance on real camera datasets, achieving higher accuracy compared to previous state-of-the-art techniques, and preserving a more robust correlation between the original and translated raw images. The codes and the pretrained models are available at https://github.com/gosha20777/rawformer. + +# 1 Introduction + +Unlike the high-quality digital single-lens reflex (DSLR) cameras, mobile phone cameras possess inherent limitations owing to their smaller sensor size and fixed lenses with limited optical capabilities [55]. The mobile camera image signal processor (ISP) aims to overcome these limitations by applying a series of carefully designed modules (e.g., [1,7,11,18,25,47]) to enhance the quality of the final image rendered in a standard color space, such as sRGB. These modules + +![](images/6d4b4324a34ece44df8b3e8175552d1dcb3c71c9c40ec6bee470dfcf63e06f97.jpg) +Fig. 1: We introduce Rawformer, an unsupervised method for raw-to-raw translation that allows the utilization of pre-trained neural-based ISPs to process raw images captured by previously unseen cameras. Shown are raw and sRGB images processed by a neural-based ISP [54]. a) Raw image from an iPhone X rendered by the neural-based ISP trained on iPhone X's raw images. b) Raw image from an iPhone X rendered by the neural-based ISP trained on Samsung S9's raw images. c) iPhone X raw image translated to Samsung S9's raw space using our method, then processed by the Samsung S9 neural-based ISP. d) Raw image captured by Samsung S9's camera rendered by its native camera ISP, provided as a reference for visual comparison. + +![](images/4a11bb1e901a555772de42314ff1b018edd143f81ab5cf8d72c9b18e69d0b52c.jpg) + +![](images/d094731a4b6cdc70ad2cafa41d617ddd4e39e9bd24ade8e2bae5d49b1b582cb1.jpg) + +![](images/d56d209367b411175300c35c10ca96bbb4793970abf8d2c1dc3b81fb6770607f.jpg) + +work in tandem to optimize key parameters, such as sharpness, color accuracy, and noise reduction, resulting in a visually pleasing sRGB image from the raw image captured by the camera [13, 16]. + +Despite yielding promising results, achieving compatibility among camera ISP modules for each new camera device requires extensive tuning and adjustment efforts to attain the desired image quality. This motivates the use of deep neural networks to replace individual camera ISP modules with a single neural-based unit capable of accomplishing the task [12,19,44]. Nonetheless, the training of such neural-based ISPs demands an extensive dataset consisting of raw images taken with a specific camera, accompanied by corresponding "ground-truth" sRGB images (typically generated by a high-quality DSLR camera) [19, 22]. + +An issue emerges with the introduction of a new camera model, as pre-trained neural-based ISPs might encounter difficulty in accurately rendering raw images captured by the new camera (refer to Fig. 1(b)). This difficulty arises from the potential distinct characteristics of the new camera, such as sensor sensitivities, which influence the formation of its raw images [4,5]. Consequently, inconsistencies arise in the interpretation of raw RGB colors between cameras [4], hindering the generalization of neural-based ISPs to new cameras not encountered during training. To address this issue, re-training or fine-tuning on a new paired dataset becomes necessary, in which the raw images are sourced from the target new camera. This process mirrors similar challenges encountered in traditional ISP development for new cameras. + +An alternative strategy to address this challenge is domain adaptation, where the raw images from the new target camera are mapped to the raw space of the training camera to emulate those captured by the original camera used for training. In light of this, [4] proposed a semi-supervised learning framework for raw image translation, but the approach outlined in [4] inherits the drawbacks associated with the CNN-based architecture, notably the limited receptive field and inefficient encoding of global information. Thus, we present a fully unsupervised + +Transformer-based raw-to-raw method, dubbed Rawformer, that maps raw images between camera models by efficiently encoding their global and semantic correlations. Our method eliminates the need for a paired raw-sRGB dataset from each new target camera, enabling the utilization of a neural-based ISP trained on a specific camera's raw images to process raw images taken by new cameras with different characteristics without necessitating any re-training; see Fig. 1(c). Our contributions are summarized as follows: + +- We propose Rawformer, a fully unsupervised encoder-decoder Transformer-based method designed for raw-to-raw translation facilitating the reuse of neural-based ISPs. Rawformer achieves state-of-the-art results in raw-to-raw translation, leading to better generalization for cross-camera learnable ISP rendering. +- We introduce contextual-scale aware downsampler and upsampler blocks that efficiently summarize the local-global contextual details in mixed scale representations via its condensed query attention block and scale perceptive feed-forward network. +- A novel cross-domain attention-driven discriminator is proposed along with a specialized discriminator head for stabilizing the network training. + +# 2 Related Work + +# 2.1 Neural-Based ISP + +Several neural-based ISPs methods have been introduced [8,17,19,24,56,59,60], with the majority relying on CNN-based U-Net-like architecture [42]. These include earlier approaches, such as DeepISP [44] and PyNET [22], as well as more recent methods, such as MW-ISPNet [21], AW-Net [12], and LAN [54], which achieve improved image restoration quality by integrating discrete wavelet transform (DWT) and double attention modules (DAM) techniques. Such learnable ISPs are typically trained to map raw images captured by a specific camera model to target high-quality sRGB "ground-truth" images produced by some target camera ISP (usually with higher quality than the mobile native ISP; e.g., a DSLR camera [19,22]). A significant challenge for neural-based ISPs is the need for paired datasets of raw-sRGB images. Collecting such datasets is tedious and typically involves using specific equipment. Additionally, it may require post-processing techniques to align the raw images with the target sRGB images [22]. The challenge intensifies when a mobile phone manufacturer introduces a new camera model with different characteristics, which happens frequently [49], requiring the recollection of paired raw-sRGB datasets for each new camera model. + +One solution is to render images to the sRGB space from the device-independent CIE XYZ space [3]. However, this requires calibration for each camera model, and accuracy depends on specific camera conditions (e.g., the Luther condition) [15]. In contrast, our work introduces an unsupervised raw-to-raw mapping technique that eliminates the need for paired raw-sRGB datasets for re-training or camera calibration. Our method maps raw images from new cameras to the camera raw + +space utilized during the training of neural-based ISPs, enabling the reuse of pre-trained ISPs. + +# 2.2 Image-to-Image Domain Adaptation + +Numerous generative adversarial network (GAN)-based methods address unpaired image-to-image (I2I) translation challenges [38]. Notably, CycleGAN [62], DualGAN [57], UNIT [32], STARGAN [10], SEAN [63], and U-GAT-IT [28] focus on cycle consistency, employing dual generator networks guided by a cycle-consistency constraint. ACLGAN [61] and CUT [39] represent advancements in this field. ACLGAN relaxes the cycle consistency constraint, favoring an adversarial constraint, potentially enhancing translation quality. CUT utilizes a contrastive loss, eliminating the need for multiple generators and enabling faster training. ITTR further refines CUT's performance by altering the generator architecture. CycleGAN-based model, such as UVCGAN [50] incorporate U-Net-like generators with vision Transformers (ViTs). + +In computational photography, attempts have been made for domain adaptation, including synthesizing raw night images from raw daylight images to enhance neural-ISP rendering quality [41], few-shot domain adaptation for low-light image enhancement [40], and mapping graphics images to a target camera's raw space [45]. Raw-to-raw translation has received even limited attention. Classical approaches require a calibration object to learn global mapping [37], while recent work [4] proposes a semi-supervised CNN learning-based method using a small set of perfectly aligned images. Despite being effective at exploiting the local semantics, this CNN-based method lacks the potential to exploit the long-range/global semantics between different domains. In contrast, our method aims at solving the raw mapping by exploiting both local and global contextual information through fully unsupervised training, achieving state-of-the-art results. + +# 3 Proposed Method + +Existing raw-to-raw methodologies either rely on supervised approaches requiring paired datasets e.g., [37]) or semi-supervised approaches (e.g., [4]) with limited accuracy. Additionally, domain adaptation techniques (e.g., [39, 62]), which aim to extract local semantics between source and target domains, often fail to capture overall semantics and context-dependent relationships. Our proposed method, Rawformer, addresses these limitations by processing raw images in a fully unsupervised manner. Utilizing an encoder-decoder architecture within a CycleGAN framework, Rawformer effectively captures global semantics and maintains intrinsic consistency between translated images. The generator (Fig. 2(a)) in Rawformer employs contextual-scale aware downsampler and upsampler blocks to enhance content rendering and structural integrity. The discriminator (Fig. 3(a)) stabilizes training by incorporating key discriminator features, thereby preventing model collapse and reducing redundant computations. In the following, we provide a comprehensive elucidation of the proposed components. + +![](images/87d13479d97e438cbc6cecdb877aec1e3d210bef0f7116250da0faf5c8f48642.jpg) +a) Generator + +![](images/829c774fe30dbaf68490f5134ac9ce23794488ae53868b98badc25b54d2035d5.jpg) + +![](images/48d9ada0364f28b5c3b82ef9e398111ce2c4e07c5c6f2ddde0bb23726fdd8893.jpg) +c) CQA block +Fig.2: a) Overview of the generator architecture (a) of Rawformer. The primary components of the generator: b) contextual-scale aware downsampler block (CSAD), c) condensed query attention block (CQA), d) scale perceptive feed-forward network (SPFN), e) composite downsampling block (CDown), f) composite upsampler (CUp) block, and g) contextual-scale aware upsampler block (CSAU). + +![](images/7ffa760ffe9fb289975c9f9060ae4c653577385a3d9f8d95933507b9458e1208.jpg) +f) $\mathrm{CUp}$ block + +![](images/dc4461e5226c53c8efb70affbb69465b38c775bdc558f01eb1c2715c98c2055e.jpg) +e) CDown block + +![](images/1e459a2f6775c04f9985c274217497f4b8127425b03ac5650483f6ae9cb365c5.jpg) +g) CSAU block + +![](images/b3ee77e438653845a1e36cf64d406faa3fa2eea0fc0b76ffee8cc305be011c77.jpg) +d) SPFN block + +# 3.1 Generator's Encoder Network + +As shown in Fig. 2(a), the encoder of the proposed generator in Rawformer comprises stacked contextual-scale aware downsampler (CSAD) blocks to bridge the semantic gap between the source and target domains. Each CSAD block, depicted in Fig. 2(b), offers a holistic approach of understanding the contextual dependencies efficiently between the domains by capturing both the global and local information via its condensed query attention (CQA) block, in parallel with a $3 \times 3$ depth-wise convolution layer. It further handles the intrinsic correlations among the extracted features in a multi-scale manner via maintaining a synergistic collaboration between the scale perceptive feed-forward network (SPFN), and the composite downsampler (CDown) block. + +Condensed Query Attention Block: Self-attention is highly effective in modeling long-range contextual dependencies by comparing each token with all others, facilitating seamless information propagation within the network [58]. However, its computational demands pose challenges for tasks like unpaired raw-to-raw translation, where maintaining both visual quality and semantic consistency is crucial. To address these concerns and overcome the constraints of computation, we introduce the condensed query attention (CQA) block; see Fig. 2(c). Here, we make two key modifications to self-attention to enhance learning efficiency. Firstly, we spatially compress the query vector using average pooling followed by linear projection, and downsample the resulting feature map by a factor of $r$ to confine the crucial information. This results in condensed queries, $(\mathbf{Q}_c)$ used for similarity comparison, as shown in Eq. 1. Secondly, to balance global and regional modeling with computational efficiency, attention is applied hor + +iontally and vertically to eliminate redundant computations. The CQA block operation is defined as: + +$$ +C Q A (\mathbf {Q} _ {c}, \mathbf {K}, \mathbf {V}) = \mathbf {A} _ {U}. (\mathbf {A} _ {H} \cdot \mathbf {V}); \quad \mathbf {Q} _ {c} = W _ {p} (A v g P o o l (\mathbf {Q})) +$$ + +$$ +\mathbf {A} _ {H} = \operatorname {S o f t m a x} \left(\frac {\mathbf {Q} _ {c} \mathbf {K} ^ {T}}{\sqrt {d}}\right), \quad \mathbf {A} _ {U} = \operatorname {S o f t m a x} \left(\frac {\mathbf {Q} \mathbf {Q} _ {c} ^ {T}}{\sqrt {d}}\right) \tag {1} +$$ + +where, the duet of key $(\mathbf{K})$ , and value $(\mathbf{V})$ vectors are obtained through plain linear convolutions, and the dimensions of each of these features are shown in the Fig. 2(c). $W_{p}(.)$ denotes a $1\times 1$ pointwise convolution, and $\mathbf{A}_H$ , and $\mathbf{A}_U$ represents the compressed attention map between the key-condensed query pair horizontally and query-condensed query pair (for focusing on fine grained contextual information) vertically, respectively. The choice of operations and the reason for condensing the queries is investigated in the supplementary of the paper. + +Scale Perceptive Feed-forward Network: To proficiently capture the intricacies of the local image structure across various scales by encoding the information from neighboring pixel locations, we propose a scale perceptive feed-forward network (SPFN); see Fig. 2(d). Unlike the conventional feed-forward networks that integrate single-scale depth-wise convolutions [58], we emphasize on enhancing the locality and inter-correlations among the incoming raw features of different domains by inserting multi-scale depth-wise convolutions. Given an input tensor at layer $s$ and scale $i$ , such that $\mathbf{X}_s^i \in \mathbb{R}^{H^i \times W^i \times C}$ , SPFN is formulated as: + +$$ +\begin{array}{c} \mathbf {X} _ {s} ^ {1} = \phi \left(W _ {d} ^ {3} \left(W _ {p} ^ {1} (L N (\mathbf {X} _ {s - 1}))\right)\right); \mathbf {X} _ {s} ^ {2} = \phi \left(W _ {d} ^ {5} \left(W _ {p} ^ {2} (L N (\mathbf {X} _ {s - 1}))\right)\right), \\ \mathbf {Y} _ {s} = W _ {s} ^ {3} \left(\mathbf {X} _ {o} ^ {1} (\widehat {\mathbf {C}}) \mathbf {X} _ {s - 1} ^ {2}\right) + \mathbf {X} _ {s - 1}. \end{array} \tag {2} +$$ + +Here, $\phi$ represents a LeakyReLU activation function [35], $W_{d}^{k}$ represents a depthwise convolution with filter size $k$ , $LN$ denotes the layer normalization [6], and $\mathbb{C}$ represents the concatenation operation. This enriches a more nuanced understanding of the underlying image features at different hierarchical levels. + +Composite Down sampler: In general, the essence of downsampling is the retrieval of the multi-scale information in the image. Unlike the common encoder architectures for various image signal processing pipelines [31, 48] that deploy strided convolutions for downsampling, we design a hybrid strategy of leveraging the benefits of information exchange between different resolution features. To adaptively recover the underlying content information from the input domain at multiple scales, we design a composite down (CDown) block that intelligently combines pixel-unshuffle (to prevent the information loss) and strided convolution (to enhance the model's expressiveness) as shown in Fig. 2(e). + +Style Modulator: Different cameras have their own distinctive characteristics to be handled or restored. To augment the expressiveness of Rawformer for + +handling camera perturbations, we design a style modulator in the generator using extended pixel-wise ViT [51]. It aids in calibrating the decoder features and subsequently encourages the recovery of crucial details while translation. As demonstrated in Fig. 2(a), the style modulator applies style tokens in each contextual aware upsampler (CSAU) block, where every style token models the weight of the query vector of the CQA block to enrich the style specific context. This addition of style token at each stage in the decoder helps in the flexible adjustment of the feature maps. Unlike StyleGANv2 [27], where the style vectors are generated from the random prior, our approach resorts to learnable style token for conjecturing the important target style directly from the source domain, thus enhancing the consistency and control in raw image translation. More details on style modulation are provided in the supplementary. + +# 3.2 Generator's Decoder Network + +Unlike the popular additive or multiplicative skip connections between the encoder-decoder, we also present a learnable linking layer (LLayer) as shown in the Fig. 2(a) to facilitate the feature preservation in the contextual scale aware upsampler (CSAU) block. The overall design of the CSAU block closely resembles that of the CSAD block, ensuring consistency and coherence in the network architecture in a bottom-up fashion. The primary distinction lies within the composite upsampling block, where for inducing the relevant spatial information into the target domain for the incoming features from the style modulator, we design a composite upsampler block (CUp). It embodies a parallel combination of pixel-shuffle (for preventing the checkerboard artifacts in the target domain) and deconvolution (for combating with aliasing) as shown in Fig. 2(f) and (g). + +![](images/59fa1358f17687797ea4aa3a662abffb9ff0bfe01aaebfcd692437afc70c3cff.jpg) +Fig. 3: a) The proposed cross-domain attention-guided discriminator. b) Feature map visualizations without (w/o) and with (w) the discriminator head. The inclusion of the discriminator head aids in refining the overall results, and the discriminator training (as shown in c) benefits from the inclusion of the discriminator head, particularly with the cache $(M)$ component. + +# 3.3 Cross-Domain Attention-Guided Discriminator + +Leveraging batch statistics is an effective strategy to mitigate mode collapse and enhance diversity in GAN models [43]. Traditional minibatch discrimination techniques require large batch sizes [26], posing challenges due to GPU limitations. Inspired by [50], we address this by decoupling batch size from the minibatch discrimination approach. Our core innovation involves using a cache of previous discriminator features as an alternative to larger batch sizes. During training, batch features are stored in four separate caches $(M)$ ; real images from the source and target cameras, and fake images from both domains. This memory bank-style cache utilizes a composite discriminator with a head and body. The head captures batch statistics via batch normalization and two $3 \times 3$ convolutional layers, processing a concatenated input of current and historical discriminator outputs from the cache, as shown in Fig. 3(a). The discriminator body includes a CQA block followed by stripped convolution to handle detailed global features across domains. This integration of batch statistics, caching, and attention mechanisms significantly enhances the discriminator's ability to manage multi-resolution image data and stabilize the overall training (Figs. 3 b and c). + +# 3.4 Self-Supervised Pre-Training + +Self-supervised pre-training optimizes the initial weights of a network by engaging it in auxiliary tasks formulated from the inherent structure of the original data. In general, despite modifications in the cross-domain attention-guided discriminator, GAN training still suffers from mode collapse, where the generator repeatedly tends to produce a limited set of images. Thus, we opt to pre-train the generator of our Rawformer using self-supervised inpainting tasks to enhance initial weight optimization and mitigate this mode collapse problem. More details about the mode of pretraining is provided in the supplementary (Section 1) of the paper. + +# 4 Experiments + +We evaluated our method using two types of datasets: 1) raw-to-raw datasets, which provide unpaired raw images for training and aligned raw images from source and target cameras for assessing unsupervised raw-to-raw translation accuracy (Sec. 4.3), and 2) raw-to-sRGB datasets, offering aligned raw and corresponding sRGB images for evaluating transformed raw image quality followed by learnable neural-based ISP rendering (Sec. 4.4). This evaluation mirrors real-world usage scenarios where raw images from a new camera are translated to the raw space of the camera used for ISP training. In all the experiments, raw images were demosaiced using the Menon algorithm [36] and normalized after black-level subtraction. During our method's training, each raw image was segmented into non-overlapping crops of $224 \times 224$ pixels. Similarly, we used paired raw-sRGB non-overlapping crops of $224 \times 224$ pixels to train the camera-specific LAN ISP [54] on each raw-to-sRGB dataset. + +# 4.1 Datasets + +We utilized two raw-to-raw datasets: 1) the Raw-to-Raw dataset [4] and 2) the NUS dataset [9], and three raw-to-sRGB datasets: 1) the Zurich raw-toRGB dataset [20], 2) the Samsung S7 dataset [44], and 3) the Mobile AIM21 dataset [19]. The Raw-to-Raw dataset [4] includes raw images from the Samsung Galaxy S9 and the iPhone X, featuring both unpaired and paired sets. The NUS dataset [9] comprises scenes captured by eight DSLR cameras. Following the evaluation in [4], we utilized the Nikon D5200 and Canon EOS 600D DSLR cameras from the NUS dataset [9]. Aligned paired raw images were generated using the same approach as in the Raw-to-Raw dataset [4]. In both datasets, paired raw images are dedicated for testing, while unpaired images are utilized for unsupervised training. + +The Zurich raw-to-RGB dataset [20] includes pairs of raw and sRGB images captured synchronously using a Huawei P20 smartphone camera and a Canon 5D Mark IV DSLR camera. The Samsung S7 dataset [44] comprises paired raw and sRGB images captured by the Samsung S7 smartphone's main camera. The Mobile AIM21 dataset [19] includes images captured using a Sony IMX586 mobile sensor and a Fujimel GFX100 DSLR, employing an advanced dense correspondence algorithm for matching due to imperfect alignment [52]. + +# 4.2 Training details + +Inspired by the effective fine-tuning capabilities demonstrated by the pre-trained BERT model for the text completion task [14], we initially trained each generator network in a self-supervised fashion for image completion. Initially, the generator was trained on $32 \times 32$ pixel patches extracted from training images, with $40\%$ of these patches being randomly masked out. The objective of the generator during this stage was to reconstruct the original unmasked image from its partially obscured version. Training was performed using the AdamW optimizer [34] with betas set to (0.9, 0.99) and a weight decay of 0.05 for 500 epochs to minimize the following pixel-wise loss function: + +$$ +\mathcal {L} _ {\text {p i x e l - w i s e}} = L _ {1} + (1 - S S I M) + V G G \tag {3} +$$ + +where $L_{1}$ , $SSIM$ , $VGG$ refer to L1 loss, Structural Similarity Index (SSIM) [53], and VGG perceptual loss [30] between original unmasked image and the reconstructed image. The learning rate was initialized at 0.005, and the learning schedule was managed using the cosine annealing warm restarts strategy [33]. This self-supervised pre-training was shown to improve the raw mapping results (see Table 1). + +Thereafter, the weights of both generators were fine-tuned for our unpaired raw-to-raw translation task for an additional 500 epochs, while simultaneously training the discriminator networks. We use the Adam optimizer [29] with betas set to (0.5, 0.99) and a learning rate of 0.0001 for the discriminator networks and 0.00005 for the generator networks. The discriminator networks were optimized to minimize the following loss functions: + +![](images/c1d10917b3b76d69b428eef7766a51a1a9a98a0241aea81ca61fee702e3d82c6.jpg) +Fig. 4: Shown are two images captured by Samsung S9 and iPhone X in sRGB (left) and two cropped patches from each image in raw (right). On the right, we show the input raw patch from the corresponding camera and the corresponding ground-truth raw patch from the other camera, along with the results by other methods. Our proposed Rawformer is better at preserving the domain consistent features. + +$$ +\mathcal {L} _ {A} ^ {d i s} = \ell_ {g a n} \left(D _ {A} \left(G _ {B \rightarrow A} (b)\right), 0\right) + \ell_ {g a n} \left(D _ {A} (a), 1\right) \tag {4} +$$ + +$$ +\mathcal {L} _ {B} ^ {d i s} = \ell_ {g a n} \left(D _ {B} \left(G _ {A \rightarrow B} (a)\right), 0\right) + \ell_ {g a n} \left(D _ {B} (b), 1\right) \tag {5} +$$ + +where $a$ and $b$ represent images from camera $A$ and $B$ , respectively, while $G_{B\to A}(b)$ and $G_{A\to B}(a)$ are the translated images produced by each generator network, respectively. The labels 0 and 1 represent "fake" and "real" raw images, respectively, and $\ell_{gan}(\cdot)$ computes the cross-entropy loss. + +During this stage, the weights of both generators were optimized to minimize the following loss function: + +$$ +\mathcal {L} ^ {\text {g e n}} = \beta_ {1} \left(\mathcal {L} _ {A} ^ {\text {g a n}} + \mathcal {L} _ {B} ^ {\text {g a n}}\right) + \beta_ {2} \left(\mathcal {L} _ {A} ^ {\text {i d t}} + \mathcal {L} _ {B} ^ {\text {i d t}}\right) + \beta_ {3} \left(\mathcal {L} _ {A} ^ {\text {c y c}} + \mathcal {L} _ {B} ^ {\text {c y c}}\right) \tag {6} +$$ + +where $\beta_{1} = 1, \beta_{2} = 10, \beta_{3} = 0.5, \mathcal{L}_{A}^{gan} = \ell_{gan}(D_{B}(G_{A\rightarrow B}(a)),1),$ $\mathcal{L}_A^{idt} = \mathcal{L}_{pixel - wise}(G_{B\to A}(a),a)$ , and $\mathcal{L}_A^{cyc} = \mathcal{L}_{pixel - wise}(G_{B\to A}(G_{A\to B}(a)),a)$ . Note that at inference time, only a single generator network (e.g., $G_{A\rightarrow B}$ ) is required to translate raw images captured by the source camera (e.g., $A$ ) to the target camera (e.g., $B$ ). + +# 4.3 Raw-to-Raw Translation Results + +We compare our raw-to-raw translation results with those of the recent semi-supervised raw-to-raw mapping (SSRM) method [4]. Additionally, we compare our method against several established unsupervised generic I2I methodologies, including UVCGANv2 [51], U-GAT-IT [28], UNIT [32], and the classic CycleGAN [62]. All other methods were trained on the same training data used for + +Table 1: Raw-to-raw translation results on NUS dataset [9]. The mapping results of Nikon D5200's raw images to Canon EOS 600D DSLR camera and vice versa using our method and other methods are shown. Here, $\ell_{a} = L_{1}$ , $\ell_{b} = L_{1} + VGG$ , $\ell_{c} = \mathcal{L}_{pixel - wise}$ (Eq. 3), w, and w/o denote the results with and without training on a corresponding configuration, respectively. Additionally, we show the GFLOPs of each method for the $256\times 256$ input image. Best results are in bold. + +
MethodsCanon-to-NikonNikon-to-Canon
PSNR↑SSIM↑MAE↓(x1e-2)ΔE↓PSNR↑SSIM↑MAE↓(x1e-2)ΔE↓
CycleGAN [62]27.320.833.2115.4026.810.823.3311.4936.8
UNIT [32]25.550.813.1718.2224.730.793.2318.6447.3
U-GAT-IT [28]29.410.873.027.1227.860.853.298.0162.9
SSRM [4]32.360.932.416.2130.810.932.475.9537.1
UVCGANv2 [51]37.110.962.324.3437.290.962.284.2869.6
Ours (w/o self-supervised pre-training; ℓe loss)37.130.962.294.6537.270.962.314.4652.7
Ours (w/o self-supervised pre-training; ℓo loss)37.940.972.213.9637.990.962.233.9352.7
Ours (w/o self-supervised pre-training; ℓe loss)38.730.972.193.6438.260.972.223.7252.7
Ours (w self-supervised pre-training; ℓe loss)41.890.981.462.0441.370.981.412.5352.7
+ +Table 2: Raw-to-raw translation results on the Raw-to-Raw dataset [4]. Shown are the mapping results of Samsung Galaxy S9's raw images to iPhone X and vice versa using our method and other methods. Additionally, the total number of parameters of each method is shown in millions (M), and inference time in milliseconds (ms) without optimization on NVIDIA GeForce RTX 4090 GPU. Results on inference time with optimization are in supplementary. + +
MethodsSamsung-to-iPhoneiPhone-to-Samsung
PSNR↑SSIM↑MAE↓(×1e-2)ΔE↓PSNR↑SSIM↑MAE↓(×1e-2)ΔE↓
CycleGAN [62]24.630.715.4014.7124.370.734.3412.8821.719
UNIT [32]23.910.705.1513.2326.420.784.2713.1632.229
U-GAT-IT [28]28.220.893.115.9831.040.892.395.0434.435
SSRM [4]29.650.892.226.3228.580.902.436.5320.218
UVCGANv2 [51]36.320.942.094.2136.460.922.024.7332.631
Ours40.980.971.332.0941.480.981.211.9926.126
+ +our method, for 500 epochs. Since SSRM is not entirely unsupervised, we incorporated 22 image pairs from the paired subset for its training, as discussed in [4]. Fig. 4 shows qualitative comparisons on the Raw-to-Raw dataset [4], while Tables 1 and 2 present the quantitative results on the NUS dataset [9] and the Raw-to-Raw dataset, respectively. In Tables 1 and 2, we adopted the same quantitative metrics used in [4]: peak signal-to-noise ratio (PSNR), SSIM, mean absolute error (MAE), and $\Delta E$ 2000 [46]. As can be seen in both the tables, our Rawformer significantly outperforms the alternative methods across all the quantitative metrics by a large margin (around $+12$ dB against SSRM [4]). + +Evaluation on Low-Light Images: Translating raw-to-raw for low-light images presents challenges due to various variations among cameras, including differences in raw normalization and noise characteristics [2,4]. To further prove the efficacy of our method against other methods on low-light images, we selected a subset of 15 paired low-light images from the Raw-to-Raw dataset [4]. The results on this subset are presented in Table 3, where our proposed method significantly outperforms SSRM [4], and UVCGANv2 [51] for translating low-light raw images captured by an iPhone X to the Samsung S9's camera. + +![](images/f24c88d4eb227a1b29742086fbd59bb66d550c827c7849a3bedac3bea665d48b.jpg) + +Table 3: Raw-to-Raw dataset [4] results on low-light raw images (a subset). + +
MethodPSNR↑SSIM↑
SSRM [4]23.870.85
UVCGANv2 [51]35.220.94
Ours40.210.97
+ +Table 4: Ablation results on the impact of various configurations of the generator model on the NUS dataset [9]. 'CQA', 'Style', 'SPFN' and 'CUp-CDown' refers to the condensed query attention, style modulator, scale perceptive feed-forward network, and the composite up and downsampler, respectively. + +
ConfigurationModulesCanon-to-NikonNikon-to-Canon
BaselineCQAStyleSPFNCUp-CdownPSNR↑SSIM↑PSNR↑SSIM↑
G1✓37.230.9637.310.96
G2✓✓38.210.9637.970.96
G3✓✓✓39.080.9738.990.96
G4✓✓✓✓39.720.9739.680.97
G5✓✓✓✓✓39.930.9739.860.97
+ +Table 5: Ablation results on the impact of various discriminator model configurations on the NUS dataset [9]. 'Baseline' refers to the $G_{5}$ config. from Tab. 4. + +
ConfigurationModulesCanon-to-NikonNikon-to-Canon
BaselineCQABatchnormDisc. headPSNR↑SSIM↑PSNR↑SSIM↑
D1✓39.930.9739.860.97
D2✓✓40.120.9740.090.97
D3✓✓✓40.240.9740.210.97
D4✓✓✓✓41.890.9841.370.98
+ +Ablation Studies: We conducted a set of ablation studies to validate the impact of the proposed components in the generator and discriminator networks. In this series of experiments, we trained our Rawformer on the NUS dataset [9], in addition to a baseline model trained using the same training settings as our Rawformer. The baseline model comprises a UNet architecture with channel and spatial attention [50] for the generator, and a PatchGAN discriminator [23]. As demonstrated in Table 4, the integration of the proposed modules into the baseline consistently improves the network's capabilities, showcasing an enhancement in overall accuracy. We also conducted further ablation studies that explore various configurations of the discriminator, as shown in Table 5. The results clearly underscore the improvement by the inclusion of each proposed component in the discriminator network (attention, batch normalization, and discriminator head (Disc. head)) in improving the accuracy of our raw-to-raw translation. + +# 4.4 Raw Translation for Learnable ISP + +To comprehend the practical application of our model in real-world scenarios, we utilized three datasets: Zurich raw-to-RGB, Samsung S7 ISP, and Mobile AIM21, encompassing various source and target domain combinations. Specifically, for each dataset, we trained the LAN ISP [54] on a single paired raw-to-sRGB + +Table 6: Cross-camera raw-to-sRGB results (PSNR/SSIM) comparing: a) UVCGANv2 [51] with b) our Rawformer. Each entry compares ground-truth sRGB images generated by the function: $ISP_{B}(F^{B}(raw_{A}))$ , where $F^{B}$ denotes translation of raw images, $raw_{A}$ , captured by camera $A$ to camera $B$ . $ISP_{B}(\cdot)$ produces sRGB images using the LAN ISP [54] trained on raw images from camera $B$ . Diagonal results (in red) represent the camera-specific ISP case and are the sRGB images produced by the ISP model tested on the same camera used for training. Best cross-camera results are in bold. + +
Training camera Aa) UVCGANv2 [51]+ LAN ISP [54]b) Our Rawformer+ LAN ISP [54]
Testing camera BTesting camera B
ZRR [20]AIM21 [19]S7 [44]ZRR [20]AIM21 [19]S7 [44]
ZRR [20]19.46/0.7317.72/0.7017.81/0.7019.46/0.7318.97/0.7219.02/0.72
AIM21 [19]22.14/0.7923.48/0.8722.39/0.7923.25/0.8623.48/0.8723.08/0.86
S7 [44]21.07/0.8020.93/0.7922.16/0.8122.08/0.8022.03/0.8022.16/0.81
+ +![](images/4ede32badba377a435c3de8c64be3f47c63a858eae6aa6126229fd9af633967a.jpg) + +![](images/3e01e560cde71fd95bd3f329f789760652fded4245422acf2ad280772bb33d92.jpg) + +![](images/dcd909ffc277820f7dda4f11df6b9013f8052be6e0a70796e18077d5b49f019c.jpg) +GT (ZRR) +Fig.5: ISP rendering results with our raw translation on various datasets. Each set includes ground-truth (GT) raw-sRGB paired images and LAN ISP [54] results on mapped raw images from different cameras. S7, ZRR, and AIM stand for the Samsung S7 ISP dataset [44], Zurich raw-to-RGB dataset [20], and Mobile AIM21 dataset [19], respectively. The shown results are consistent with the ground-truth, demonstrating the proficiency of our model. + +![](images/cc12e9be4c46b4678bb40efe836fe993c0de6061f6ff54df39251a644a3fcf77.jpg) + +![](images/583ef5c178ee8ff30e2d53a81b446b9e141d580b87d2bed813cc9ca57ace131c.jpg) + +![](images/67c3c27f66d5c1901334932999c69d01178d2afbdb08a36002b78c426032f964.jpg) +$\mathrm{ISP}_{\mathrm{AIM}}(\mathrm{F}^{\mathrm{AIM}}(\mathrm{raw}_{\mathrm{RRB}}))$ + +![](images/dfbc321511caa4630efaee585e2a330e6d88e47eb7641538feeffd9cbc89e646.jpg) + +![](images/e7b343cc1bc9bf1534ede3f48fdba81e500484017e621d12832dc8da9905ad38.jpg) + +![](images/dacdc8543859dd5619c7db16497000895bda0bf7d8950fc158c9b40405ec7e42.jpg) +GT (AIM) + +![](images/07b4afa3c84b206ae4af3f2bd5f880f010dd2d3490dc2e4da023c872fc36f40b.jpg) + +![](images/14725fd33bb29868313aeb7e1d722c9cfbac60164554d9f35bdbd6994b88f760.jpg) + +![](images/90ae8e7a255b52f2d6b669fcfa778beb479bbc674b03d8775b23a06cde21b718.jpg) +$\mathrm{ISP}_{\mathrm{ZRR}}(\mathrm{F}^{\mathrm{ZRR}}(\mathrm{raw}_{\mathrm{AIM}}))\mathrm{ISP}_{\mathrm{S7}}(\mathrm{F}^{\mathrm{ST}}(\mathrm{raw}_{\mathrm{AIM}}))$ + +![](images/368226994a734226f180796668fb2963f1018452846f068438ecc11bdcec71b8.jpg) + +![](images/7353958554e1f0934633aadf26862e41d399ef2e62a85bcd095ac8d9d85de26f.jpg) + +![](images/10526664d7c0d64cbd22034a159e4e93d756bb7bf4de4e0adf7e1bd0014ddc06.jpg) +GT (S7) + +![](images/b433f06f08ba8e3813ace1980b87e73083f2e0221efdfa4f8060eac4df548018.jpg) + +![](images/c5e0d7af0302d321cbd6b955f02cfc38d5a140143f2dc9f51ef0c6eca8357b39.jpg) + +![](images/9b87cb02fc7aaf51264c2638a34a16efb881db4db11da08a3e01c91d8a369b9d.jpg) +$\mathrm{ISP}_{\mathrm{ZRR}}(\mathrm{F}^{\mathrm{ZRR}}(\mathrm{raw}_{S7}))$ + +dataset, where raw images originated from a single camera. Then, we transferred all raw images from other cameras in the other datasets using our Rawformer before rendering with the trained LAN model. We report the PSNR and SSIM of the rendered sRGB images compared to the ground-truth sRGB images for each set as shown in Table 6. This includes the results of camera-specific LAN ISP models trained and tested on the same datasets. The camera-specific results represent the optimal results achievable by the neural-based ISP, which requires a paired raw-SRGB dataset for training on each camera model. As can be seen, our unsupervised raw-to-raw translation method demonstrates robust improvement, exhibiting only a marginal $3\%$ reduction in accuracy compared to the camera-specific LAN model trained and tested on the same camera. This underscores the effectiveness of our method in addressing real-world cross-domain broadcasting challenges. See Fig. 5 for qualitative examples. + +# 4.5 Comparison with Few-Shot Domain Adaptation + +We further compared our Rawformer against Prabhakar's Few-Shot Domain Adaptation (PDA) [40]. Specifically, we conducted a 10-shot domain adaptation from the Samsung S7 ISP dataset's camera [44] to the Zurich raw-to-RGB dataset's camera [20], and vice-versa. The PDA model underwent initial training for 10 epochs using the Samsung S7 ISP dataset. Subsequently, domain adaptation was performed on 10 images from the Zurich raw-to-RGB dataset's camera over 4 epochs. This process resulted in performance metrics of 21.07 dB, and 17.81 dB in terms of PSNR and 0.80 and 0.70 for SSIM for our Rawformer, representing approximately $11\%$ superior performance compared to PDA (see Table 7). + +Table 7: Comparison with PDA [40]. S7 and ZRR denote the Samsung S7 ISP [44] and the Zurich Raw-to-RGB [20] datasets, respectively. Results are produced after translating each raw image to the target camera using PDA and our Rawformer, subsequently rendering with a LAN ISP [54] trained on the target camera's raw images. + +
MethodS7-to-ZRRZRR-to-S7
PSNR↑SSIM↑MAE↓ (×1e-2)ΔE↓PSNR↑SSIM↑MAE↓ (×1e-2)ΔE↓
PDA [40]20.330.799.4113.9917.120.6918.0315.86
Ours22.080.808.1810.2719.020.7215.2711.74
+ +# 5 Conclusion and Future Work + +In this work, we presented a new approach, Rawformer, for raw-to-raw image translation using unpaired data. Rawformer sets a new standard for performance on real camera datasets, showcasing its superior capabilities in raw-to-raw translation when measured against other alternatives. The proposed method reduces the need for the extensive collection of raw-to-sRGB paired datasets to re-train learnable ISPs for new cameras, thereby decreasing associated costs. Our results indicate that existing raw-to-sRGB paired datasets can reliably train neural-based learnable ISPs, which can then be utilized to render raw images taken by unseen cameras, lacking such unpaired data. We believe that our approach provides valuable insights for learnable ISP, raw image translation research, and applications. Our method (without optimization) runs at 26 milliseconds per frame on GPU ( $\sim$ 1 second on CPU), which may be impractical for real-time rendering in camera previews on devices with limited computational power. Future work will focus on introducing lighter models capable of real-time performance on CPU. + +# Acknowledgments + +This work was partly supported by The Alexander von Humboldt Foundation. + +# References + +1. A Sharif, S., Naqvi, R.A., Biswas, M.: Beyond joint demosaicking and denoising: An image processing pipeline for a pixel-bin image sensor. In: CVPR (2021) +2. Abdelhamed, A., Lin, S., Brown, M.S.: A high-quality denoising dataset for smartphone cameras. In: CVPR (2018) +3. Affi, M., Abdelhamed, A., Abuolaim, A., Punnappurath, A., Brown, M.S.: CIE XYZ net: Unprocessing images for low-level computer vision tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(9), 4688-4700 (2021) +4. Afifi, M., Abuolaim, A.: Semi-supervised raw-to-raw mapping. In: BMVC (2021) +5. Affi, M., Brown, M.S.: Sensor-independent illumination estimation for dnn models. In: BMVC (2019) +6. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016) +7. Barron, J.T., Tsai, Y.T.: Fast Fourier color constancy. In: CVPR (2017) +8. Chen, C., Chen, Q., Xu, J., Koltun, V.: Learning to see in the dark. In: CVPR (2018) +9. Cheng, D., Prasad, D.K., Brown, M.S.: Illuminant estimation for color constancy: why spatial-domain methods work and the role of the color distribution. JOSA A 31(5), 1049-1058 (2014) +0. Choi, Y., Choi, M., Kim, M., Ha, J.W., Kim, S., Choo, J.: StarGAN: Unified generative adversarial networks for multi-domain image-to-image translation. In: CVPR (2018) +1. Conde, M.V., Vasluianu, F., Vazquez-Corral, J., Timofte, R.: Perceptual image enhancement for smartphone real-time applications. In: CVPR (2023) +2. Dai, L., Liu, X., Li, C., Chen, J.: AWNet: Attentive wavelet network for image ISP. In: ECCV (2020) +3. Delbracio, M., Kelly, D., Brown, M.S., Milanfar, P.: Mobile computational photography: A tour. Annual Review of Vision Science 7, 571-604 (2021) +4. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) +5. Finlayson, G.D., Zhu, Y.: Designing color filters that make cameras more colorimetric. IEEE Transactions on Image Processing 30, 853-867 (2020) +6. Hasinoff, S.W., Sharlet, D., Geiss, R., Adams, A., Barron, J.T., Kainz, F., Chen, J., Levoy, M.: Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM Transactions on Graphics 35(6), 1-12 (2016) +7. He, X., Hu, T., Wang, G., Wang, Z., Wang, R., Zhang, Q., Yan, K., Chen, Z., Li, R., Xie, C., et al.: Enhancing RAW-to-sRGB with decoupled style structure in Fourier domain. In: AAAI (2024) +8. Herrmann, C., Bowen, R.S., Wadhwa, N., Garg, R., He, Q., Barron, J.T., Zabih, R.: Learning to autofocus. In: CVPR (2020) +9. Ignatov, A., Chiang, C.M., Kuo, H.K., Sycheva, A., Timofte, R.: Learned smartphone ISP on mobile NPUs with deep learning, mobile AI 2021 challenge: Report. In: CVPRW (2021) +20. Ignatov, A., Timofte, R., Ko, S.J., Kim, S.W., Uhm, K.H., Ji, S.W., Cho, S.J., Hong, J.P., Mei, K., Li, J., et al.: AIM 2019 challenge on raw to rgb mapping: Methods and results. In: ICCVW (2019) +21. Ignatov, A., Timofte, R., Zhang, Z., Liu, M., Wang, H., Zuo, W., Zhang, J., Zhang, R., Peng, Z., Ren, S., et al.: AIM 2020 challenge on learned image signal processing pipeline. In: ECCV (2020) + +22. Ignatov, A., Van Gool, L., Timofte, R.: Replacing mobile camera ISP with a single deep learning model. In: CVPRW (2020) +23. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: CVPR (2017) +24. Jeong, W., Jung, S.W.: RAWtoBit: A fully end-to-end camera isp network. In: ECCV (2022) +25. Jiang, Y., Wronski, B., Mildenhall, B., Barron, J.T., Wang, Z., Xue, T.: Fast and high quality image denoising via malleable convolution. In: ECCV (2022) +26. Karras, T., Aila, T., Laine, S., Lehtinen, J.: Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196 (2017) +27. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: CVPR (2020) +28. Kim, J., Kim, M., Kang, H., Lee, K.: U-GAT-IT: Unsupervised generative attentional networks with adaptive layer-instance normalization for image-to-image translation. arXiv preprint arXiv:1907.10830 (2019) +29. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) +30. Ledig, C., Theis, L., Huszar, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., et al.: Photo-realistic single image superresolution using a generative adversarial network. In: CVPR (2017) +31. Liang, Z., Cai, J., Cao, Z., Zhang, L.: Cameranet: A two-stage framework for effective camera isp learning. IEEE Transactions on Image Processing 30, 2248-2262 (2021) +32. Liu, M.Y., Breuel, T., Kautz, J.: Unsupervised image-to-image translation networks. In: NeurIPS (2017) +33. Loshchilov, I., Hutter, F.: SGDR: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016) +34. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017) +35. Maas, A.L., Hannun, A.Y., Ng, A.Y., et al.: Rectifier nonlinearities improve neural network acoustic models. In: ICML (2013) +36. Menon, D., Andriani, S., Calvagno, G.: Demosaicing with directional filtering and a posteriori decision. IEEE Transactions on Image Processing 16(1), 132-141 (2006) +37. Nguyen, R., Prasad, D.K., Brown, M.S.: Raw-to-raw: Mapping between image sensor color responses. In: CVPR (2014) +38. Pang, Y., Lin, J., Qin, T., Chen, Z.: Image-to-image translation: Methods and applications. IEEE Transactions on Multimedia 24, 3859-3881 (2021) +39. Park, T., Efros, A.A., Zhang, R., Zhu, J.Y.: Contrastive learning for unpaired image-to-image translation. In: ECCV (2020) +40. Prabhakar, K.R., Vinod, V., Sahoo, N.R., Babu, R.V.: Few-shot domain adaptation for low light raw image enhancement. arXiv preprint arXiv:2303.15528 (2023) +41. Punnappurath, A., Abuolaim, A., Abdelhamed, A., Levinshtein, A., Brown, M.S.: Day-to-night image synthesis for training nighttime neural isps. In: CVPR (2022) +42. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: International Conference on Medical image computing and computer-assisted intervention (2015) +43. Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen, X.: Improved techniques for training gans. Advances in neural information processing systems 29 (2016) + +44. Schwartz, E., Giryes, R., Bronstein, A.M.: DeepISP: Toward learning an end-to-end image processing pipeline. IEEE Transactions on Image Processing 28(2), 912-923 (2018) +45. Seo, D., Punnappurath, A., Zhao, L., Abdelhamed, A., Tedla, S.K., Park, S., Choe, J., Brown, M.S.: Graphics2RAW: Mapping computer graphics images to sensor raw images. In: ICCV (2023) +46. Sharma, G., Wu, W., Dalal, E.N.: The CIEDE2000 color-difference formula: Implementation notes, supplementary test data, and mathematical observations. Color Research & Application 30(1), 21-30 (2005) +47. Sindelář, O., Sroubek, F.: Image deblurring in smartphone devices using built-in inertial measurement sensors. Journal of Electronic Imaging 22(1), 011003-011003 (2013) +48. Souza, M., Heidrich, W.: Crispnet: Color rendition isp net. arXiv preprint arXiv:2203.10562 (2022) +49. Tominaga, S., Nishi, S., Ohtera, R.: Measurement and estimation of spectral sensitivity functions for mobile phone cameras. Sensors 21(15), 4985 (2021) +50. Torbunov, D., Huang, Y., Yu, H., Huang, J., Yoo, S., Lin, M., Viren, B., Ren, Y.: UVCGAN: UNet vision transformer cycle-consistent GAN for unpaired image-to-image translation. In: WACV (2023) +51. Torbunov, D., Huang, Y., Yu, H., Huang, J., Yoo, S., Lin, M., Viren, B., Ren, Y.: UVCGAN v2: An improved cycle-consistent GAN for unpaired image-to-image translation. arXiv preprint arXiv:2303.16280 (2023) +52. Truong, P., Danelljan, M., Van Gool, L., Timofte, R.: Learning accurate dense correspondences and when to trust them. In: CVPR (2021) +53. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing 13(4), 600-612 (2004) +54. Wirzberger Raimundo, D., Ignatov, A., Timofte, R.: LAN: Lightweight attention-based network for raw-to-rgb smartphone image processing. In: CVPRW. pp. 807-815 (2022) +55. Wronski, B., Garcia-Dorado, I., Ernst, M., Kelly, D., Krainin, M., Liang, C.K., Levoy, M., Milanfar, P.: Handheld multi-frame super-resolution. ACM Transactions on Graphics (ToG) 38(4), 1-18 (2019) +56. Xing, Y., Qian, Z., Chen, Q.: Invertible image signal processing. In: CVPR (2021) +57. Yi, Z., Zhang, H., Tan, P., Gong, M.: DualGAN: Unsupervised dual learning for image-to-image translation. In: ICCV. pp. 2849-2857 (2017) +58. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5728-5739 (2022) +59. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: CycleISP: Real image restoration via improved data synthesis. In: CVPR (2020) +60. Zhang, Z., Wang, H., Liu, M., Wang, R., Zhang, J., Zuo, W.: Learning raw-to-srgb mappings with inaccurately aligned supervision. In: ICCV (2021) +61. Zhao, Y., Wu, R., Dong, H.: Unpaired image-to-image translation using adversarial consistency loss. In: ECCV (2020) +62. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: ICCV (2017) +63. Zhu, P., Abdal, R., Qin, Y., Wonka, P.: SEAN: Image synthesis with semantic region-adaptive normalization. In: CVPR (2020) \ No newline at end of file diff --git a/2024/Rawformer_ Unpaired Raw-to-Raw Translation for Learnable Camera ISPs/images.zip b/2024/Rawformer_ Unpaired Raw-to-Raw Translation for Learnable Camera ISPs/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..04a824ac6369849229260e260dc4cc00444758dc --- /dev/null +++ b/2024/Rawformer_ Unpaired Raw-to-Raw Translation for Learnable Camera ISPs/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f951c49520baa7e6ba107d86c7872c58ebc8c688bb33fcb725e839a78832877 +size 518871 diff --git a/2024/Rawformer_ Unpaired Raw-to-Raw Translation for Learnable Camera ISPs/layout.json b/2024/Rawformer_ Unpaired Raw-to-Raw Translation for Learnable Camera ISPs/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..20dca9d9010cc65d581957b6f98fc9447b1bcdbb --- /dev/null +++ b/2024/Rawformer_ Unpaired Raw-to-Raw Translation for Learnable Camera ISPs/layout.json @@ -0,0 +1,9679 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 111, + 471, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 111, + 471, + 146 + ], + "spans": [ + { + "bbox": [ + 143, + 111, + 471, + 146 + ], + "type": "text", + "content": "Rawformer: Unpaired Raw-to-Raw Translation for Learnable Camera ISPs" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 140, + 168, + 474, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 168, + 474, + 192 + ], + "spans": [ + { + "bbox": [ + 140, + 168, + 474, + 192 + ], + "type": "text", + "content": "Georgy Perevozhikov1, Nancy Mehta1, Mahmoud Affifi2, and Radu Timofte1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 171, + 201, + 441, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 201, + 441, + 224 + ], + "spans": [ + { + "bbox": [ + 171, + 201, + 441, + 224 + ], + "type": "text", + "content": "1 Computer Vision Lab, CAIDAS & IFI, University of Würzburg, John Skilton Str. 4a, 97074 Würzburg, Germany" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 153, + 225, + 460, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 225, + 460, + 236 + ], + "spans": [ + { + "bbox": [ + 153, + 225, + 460, + 236 + ], + "type": "text", + "content": "{georgii.perevozhikov,nancy.mehta,radu.timofte}@uni-wuerzburg.de" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 164, + 235, + 449, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 235, + 449, + 258 + ], + "spans": [ + { + "bbox": [ + 164, + 235, + 449, + 258 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 164, + 235, + 449, + 258 + ], + "type": "text", + "content": " York University, 4700 Keele St, Toronto, Ontario, Canada, M3J 1P3 m.3afifi@gmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 287, + 452, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 287, + 452, + 518 + ], + "spans": [ + { + "bbox": [ + 160, + 287, + 452, + 518 + ], + "type": "text", + "content": "Abstract. Modern smartphone camera quality heavily relies on the image signal processor (ISP) to enhance captured raw images, utilizing carefully designed modules to produce final output images encoded in a standard color space (e.g., sRGB). Neural-based end-to-end learnable ISPs offer promising advancements, potentially replacing traditional ISPs with their ability to adapt without requiring extensive tuning for each new camera model, as is often the case for nearly every module in traditional ISPs. However, the key challenge with the recent learning-based ISPs is the urge to collect large paired datasets for each distinct camera model due to the influence of intrinsic camera characteristics on the formation of input raw images. This paper tackles this challenge by introducing a novel method for unpaired learning of raw-to-raw translation across diverse cameras. Specifically, we propose Rawformer, an unsupervised Transformer-based encoder-decoder method for raw-to-raw translation. It accurately maps raw images captured by a certain camera to the target camera, facilitating the generalization of learnable ISPs to new unseen cameras. Our method demonstrates superior performance on real camera datasets, achieving higher accuracy compared to previous state-of-the-art techniques, and preserving a more robust correlation between the original and translated raw images. The codes and the pretrained models are available at https://github.com/gosha20777/rawformer." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 538, + 229, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 538, + 229, + 551 + ], + "spans": [ + { + "bbox": [ + 133, + 538, + 229, + 551 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "text", + "content": "Unlike the high-quality digital single-lens reflex (DSLR) cameras, mobile phone cameras possess inherent limitations owing to their smaller sensor size and fixed lenses with limited optical capabilities [55]. The mobile camera image signal processor (ISP) aims to overcome these limitations by applying a series of carefully designed modules (e.g., [1,7,11,18,25,47]) to enhance the quality of the final image rendered in a standard color space, such as sRGB. These modules" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 132, + 642, + 236, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 642, + 236, + 654 + ], + "spans": [ + { + "bbox": [ + 132, + 642, + 236, + 654 + ], + "type": "text", + "content": "* Corresponding Author" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 654, + 206, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 654, + 206, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 654, + 206, + 665 + ], + "type": "text", + "content": "\\*\\* Now at Google" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 114, + 218, + 182 + ], + "blocks": [ + { + "bbox": [ + 141, + 114, + 218, + 182 + ], + "lines": [ + { + "bbox": [ + 141, + 114, + 218, + 182 + ], + "spans": [ + { + "bbox": [ + 141, + 114, + 218, + 182 + ], + "type": "image", + "image_path": "6d4b4324a34ece44df8b3e8175552d1dcb3c71c9c40ec6bee470dfcf63e06f97.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 189, + 482, + 289 + ], + "lines": [ + { + "bbox": [ + 130, + 189, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 482, + 289 + ], + "type": "text", + "content": "Fig. 1: We introduce Rawformer, an unsupervised method for raw-to-raw translation that allows the utilization of pre-trained neural-based ISPs to process raw images captured by previously unseen cameras. Shown are raw and sRGB images processed by a neural-based ISP [54]. a) Raw image from an iPhone X rendered by the neural-based ISP trained on iPhone X's raw images. b) Raw image from an iPhone X rendered by the neural-based ISP trained on Samsung S9's raw images. c) iPhone X raw image translated to Samsung S9's raw space using our method, then processed by the Samsung S9 neural-based ISP. d) Raw image captured by Samsung S9's camera rendered by its native camera ISP, provided as a reference for visual comparison." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 227, + 114, + 302, + 181 + ], + "blocks": [ + { + "bbox": [ + 227, + 114, + 302, + 181 + ], + "lines": [ + { + "bbox": [ + 227, + 114, + 302, + 181 + ], + "spans": [ + { + "bbox": [ + 227, + 114, + 302, + 181 + ], + "type": "image", + "image_path": "4a11bb1e901a555772de42314ff1b018edd143f81ab5cf8d72c9b18e69d0b52c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 312, + 114, + 388, + 181 + ], + "blocks": [ + { + "bbox": [ + 312, + 114, + 388, + 181 + ], + "lines": [ + { + "bbox": [ + 312, + 114, + 388, + 181 + ], + "spans": [ + { + "bbox": [ + 312, + 114, + 388, + 181 + ], + "type": "image", + "image_path": "d094731a4b6cdc70ad2cafa41d617ddd4e39e9bd24ade8e2bae5d49b1b582cb1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 398, + 114, + 473, + 181 + ], + "blocks": [ + { + "bbox": [ + 398, + 114, + 473, + 181 + ], + "lines": [ + { + "bbox": [ + 398, + 114, + 473, + 181 + ], + "spans": [ + { + "bbox": [ + 398, + 114, + 473, + 181 + ], + "type": "image", + "image_path": "d56d209367b411175300c35c10ca96bbb4793970abf8d2c1dc3b81fb6770607f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 315, + 479, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 479, + 351 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 479, + 351 + ], + "type": "text", + "content": "work in tandem to optimize key parameters, such as sharpness, color accuracy, and noise reduction, resulting in a visually pleasing sRGB image from the raw image captured by the camera [13, 16]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 352, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 449 + ], + "type": "text", + "content": "Despite yielding promising results, achieving compatibility among camera ISP modules for each new camera device requires extensive tuning and adjustment efforts to attain the desired image quality. This motivates the use of deep neural networks to replace individual camera ISP modules with a single neural-based unit capable of accomplishing the task [12,19,44]. Nonetheless, the training of such neural-based ISPs demands an extensive dataset consisting of raw images taken with a specific camera, accompanied by corresponding \"ground-truth\" sRGB images (typically generated by a high-quality DSLR camera) [19, 22]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 449, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 581 + ], + "type": "text", + "content": "An issue emerges with the introduction of a new camera model, as pre-trained neural-based ISPs might encounter difficulty in accurately rendering raw images captured by the new camera (refer to Fig. 1(b)). This difficulty arises from the potential distinct characteristics of the new camera, such as sensor sensitivities, which influence the formation of its raw images [4,5]. Consequently, inconsistencies arise in the interpretation of raw RGB colors between cameras [4], hindering the generalization of neural-based ISPs to new cameras not encountered during training. To address this issue, re-training or fine-tuning on a new paired dataset becomes necessary, in which the raw images are sourced from the target new camera. This process mirrors similar challenges encountered in traditional ISP development for new cameras." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "An alternative strategy to address this challenge is domain adaptation, where the raw images from the new target camera are mapped to the raw space of the training camera to emulate those captured by the original camera used for training. In light of this, [4] proposed a semi-supervised learning framework for raw image translation, but the approach outlined in [4] inherits the drawbacks associated with the CNN-based architecture, notably the limited receptive field and inefficient encoding of global information. Thus, we present a fully unsupervised" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 261, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 261, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 261, + 101 + ], + "type": "text", + "content": "G. Perevozhikov et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "Transformer-based raw-to-raw method, dubbed Rawformer, that maps raw images between camera models by efficiently encoding their global and semantic correlations. Our method eliminates the need for a paired raw-sRGB dataset from each new target camera, enabling the utilization of a neural-based ISP trained on a specific camera's raw images to process raw images taken by new cameras with different characteristics without necessitating any re-training; see Fig. 1(c). Our contributions are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 207, + 482, + 339 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 207, + 482, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 207, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 207, + 482, + 266 + ], + "type": "text", + "content": "- We propose Rawformer, a fully unsupervised encoder-decoder Transformer-based method designed for raw-to-raw translation facilitating the reuse of neural-based ISPs. Rawformer achieves state-of-the-art results in raw-to-raw translation, leading to better generalization for cross-camera learnable ISP rendering." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 267, + 482, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 482, + 314 + ], + "type": "text", + "content": "- We introduce contextual-scale aware downsampler and upsampler blocks that efficiently summarize the local-global contextual details in mixed scale representations via its condensed query attention block and scale perceptive feed-forward network." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 315, + 482, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 315, + 482, + 339 + ], + "spans": [ + { + "bbox": [ + 138, + 315, + 482, + 339 + ], + "type": "text", + "content": "- A novel cross-domain attention-driven discriminator is proposed along with a specialized discriminator head for stabilizing the network training." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 357, + 237, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 357, + 237, + 370 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 237, + 370 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 382, + 251, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 382, + 251, + 393 + ], + "spans": [ + { + "bbox": [ + 132, + 382, + 251, + 393 + ], + "type": "text", + "content": "2.1 Neural-Based ISP" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 402, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 594 + ], + "type": "text", + "content": "Several neural-based ISPs methods have been introduced [8,17,19,24,56,59,60], with the majority relying on CNN-based U-Net-like architecture [42]. These include earlier approaches, such as DeepISP [44] and PyNET [22], as well as more recent methods, such as MW-ISPNet [21], AW-Net [12], and LAN [54], which achieve improved image restoration quality by integrating discrete wavelet transform (DWT) and double attention modules (DAM) techniques. Such learnable ISPs are typically trained to map raw images captured by a specific camera model to target high-quality sRGB \"ground-truth\" images produced by some target camera ISP (usually with higher quality than the mobile native ISP; e.g., a DSLR camera [19,22]). A significant challenge for neural-based ISPs is the need for paired datasets of raw-sRGB images. Collecting such datasets is tedious and typically involves using specific equipment. Additionally, it may require post-processing techniques to align the raw images with the target sRGB images [22]. The challenge intensifies when a mobile phone manufacturer introduces a new camera model with different characteristics, which happens frequently [49], requiring the recollection of paired raw-sRGB datasets for each new camera model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 491, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 491, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 491, + 666 + ], + "type": "text", + "content": "One solution is to render images to the sRGB space from the device-independent CIE XYZ space [3]. However, this requires calibration for each camera model, and accuracy depends on specific camera conditions (e.g., the Luther condition) [15]. In contrast, our work introduces an unsupervised raw-to-raw mapping technique that eliminates the need for paired raw-sRGB datasets for re-training or camera calibration. Our method maps raw images from new cameras to the camera raw" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 401, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 401, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 401, + 91, + 447, + 101 + ], + "type": "text", + "content": "Rawformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "text", + "content": "space utilized during the training of neural-based ISPs, enabling the reuse of pre-trained ISPs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 158, + 344, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 158, + 344, + 171 + ], + "spans": [ + { + "bbox": [ + 131, + 158, + 344, + 171 + ], + "type": "text", + "content": "2.2 Image-to-Image Domain Adaptation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 178, + 481, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 178, + 481, + 309 + ], + "spans": [ + { + "bbox": [ + 130, + 178, + 481, + 309 + ], + "type": "text", + "content": "Numerous generative adversarial network (GAN)-based methods address unpaired image-to-image (I2I) translation challenges [38]. Notably, CycleGAN [62], DualGAN [57], UNIT [32], STARGAN [10], SEAN [63], and U-GAT-IT [28] focus on cycle consistency, employing dual generator networks guided by a cycle-consistency constraint. ACLGAN [61] and CUT [39] represent advancements in this field. ACLGAN relaxes the cycle consistency constraint, favoring an adversarial constraint, potentially enhancing translation quality. CUT utilizes a contrastive loss, eliminating the need for multiple generators and enabling faster training. ITTR further refines CUT's performance by altering the generator architecture. CycleGAN-based model, such as UVCGAN [50] incorporate U-Net-like generators with vision Transformers (ViTs)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 310, + 482, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 482, + 455 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 482, + 455 + ], + "type": "text", + "content": "In computational photography, attempts have been made for domain adaptation, including synthesizing raw night images from raw daylight images to enhance neural-ISP rendering quality [41], few-shot domain adaptation for low-light image enhancement [40], and mapping graphics images to a target camera's raw space [45]. Raw-to-raw translation has received even limited attention. Classical approaches require a calibration object to learn global mapping [37], while recent work [4] proposes a semi-supervised CNN learning-based method using a small set of perfectly aligned images. Despite being effective at exploiting the local semantics, this CNN-based method lacks the potential to exploit the long-range/global semantics between different domains. In contrast, our method aims at solving the raw mapping by exploiting both local and global contextual information through fully unsupervised training, achieving state-of-the-art results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 472, + 262, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 472, + 262, + 487 + ], + "spans": [ + { + "bbox": [ + 131, + 472, + 262, + 487 + ], + "type": "text", + "content": "3 Proposed Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Existing raw-to-raw methodologies either rely on supervised approaches requiring paired datasets e.g., [37]) or semi-supervised approaches (e.g., [4]) with limited accuracy. Additionally, domain adaptation techniques (e.g., [39, 62]), which aim to extract local semantics between source and target domains, often fail to capture overall semantics and context-dependent relationships. Our proposed method, Rawformer, addresses these limitations by processing raw images in a fully unsupervised manner. Utilizing an encoder-decoder architecture within a CycleGAN framework, Rawformer effectively captures global semantics and maintains intrinsic consistency between translated images. The generator (Fig. 2(a)) in Rawformer employs contextual-scale aware downsampler and upsampler blocks to enhance content rendering and structural integrity. The discriminator (Fig. 3(a)) stabilizes training by incorporating key discriminator features, thereby preventing model collapse and reducing redundant computations. In the following, we provide a comprehensive elucidation of the proposed components." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 261, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 261, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 261, + 102 + ], + "type": "text", + "content": "G. Perevozhikov et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 118, + 279, + 244 + ], + "blocks": [ + { + "bbox": [ + 132, + 118, + 279, + 244 + ], + "lines": [ + { + "bbox": [ + 132, + 118, + 279, + 244 + ], + "spans": [ + { + "bbox": [ + 132, + 118, + 279, + 244 + ], + "type": "image", + "image_path": "87d13479d97e438cbc6cecdb877aec1e3d210bef0f7116250da0faf5c8f48642.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 249, + 223, + 255 + ], + "lines": [ + { + "bbox": [ + 194, + 249, + 223, + 255 + ], + "spans": [ + { + "bbox": [ + 194, + 249, + 223, + 255 + ], + "type": "text", + "content": "a) Generator" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 283, + 118, + 343, + 193 + ], + "blocks": [ + { + "bbox": [ + 283, + 118, + 343, + 193 + ], + "lines": [ + { + "bbox": [ + 283, + 118, + 343, + 193 + ], + "spans": [ + { + "bbox": [ + 283, + 118, + 343, + 193 + ], + "type": "image", + "image_path": "829c774fe30dbaf68490f5134ac9ce23794488ae53868b98badc25b54d2035d5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 283, + 194, + 342, + 251 + ], + "blocks": [ + { + "bbox": [ + 283, + 194, + 342, + 251 + ], + "lines": [ + { + "bbox": [ + 283, + 194, + 342, + 251 + ], + "spans": [ + { + "bbox": [ + 283, + 194, + 342, + 251 + ], + "type": "image", + "image_path": "48d9ada0364f28b5c3b82ef9e398111ce2c4e07c5c6f2ddde0bb23726fdd8893.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 251, + 328, + 257 + ], + "lines": [ + { + "bbox": [ + 302, + 251, + 328, + 257 + ], + "spans": [ + { + "bbox": [ + 302, + 251, + 328, + 257 + ], + "type": "text", + "content": "c) CQA block" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 270, + 482, + 326 + ], + "lines": [ + { + "bbox": [ + 131, + 270, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 131, + 270, + 482, + 326 + ], + "type": "text", + "content": "Fig.2: a) Overview of the generator architecture (a) of Rawformer. The primary components of the generator: b) contextual-scale aware downsampler block (CSAD), c) condensed query attention block (CQA), d) scale perceptive feed-forward network (SPFN), e) composite downsampling block (CDown), f) composite upsampler (CUp) block, and g) contextual-scale aware upsampler block (CSAU)." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 352, + 117, + 417, + 149 + ], + "blocks": [ + { + "bbox": [ + 352, + 117, + 417, + 149 + ], + "lines": [ + { + "bbox": [ + 352, + 117, + 417, + 149 + ], + "spans": [ + { + "bbox": [ + 352, + 117, + 417, + 149 + ], + "type": "image", + "image_path": "7ffa760ffe9fb289975c9f9060ae4c653577385a3d9f8d95933507b9458e1208.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 151, + 395, + 157 + ], + "lines": [ + { + "bbox": [ + 370, + 151, + 395, + 157 + ], + "spans": [ + { + "bbox": [ + 370, + 151, + 395, + 157 + ], + "type": "text", + "content": "f) " + }, + { + "bbox": [ + 370, + 151, + 395, + 157 + ], + "type": "inline_equation", + "content": "\\mathrm{CUp}" + }, + { + "bbox": [ + 370, + 151, + 395, + 157 + ], + "type": "text", + "content": " block" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 352, + 158, + 416, + 190 + ], + "blocks": [ + { + "bbox": [ + 352, + 158, + 416, + 190 + ], + "lines": [ + { + "bbox": [ + 352, + 158, + 416, + 190 + ], + "spans": [ + { + "bbox": [ + 352, + 158, + 416, + 190 + ], + "type": "image", + "image_path": "dc4461e5226c53c8efb70affbb69465b38c775bdc558f01eb1c2715c98c2055e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 191, + 400, + 197 + ], + "lines": [ + { + "bbox": [ + 370, + 191, + 400, + 197 + ], + "spans": [ + { + "bbox": [ + 370, + 191, + 400, + 197 + ], + "type": "text", + "content": "e) CDown block" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 424, + 116, + 480, + 192 + ], + "blocks": [ + { + "bbox": [ + 424, + 116, + 480, + 192 + ], + "lines": [ + { + "bbox": [ + 424, + 116, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 424, + 116, + 480, + 192 + ], + "type": "image", + "image_path": "1e459a2f6775c04f9985c274217497f4b8127425b03ac5650483f6ae9cb365c5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 438, + 193, + 466, + 198 + ], + "lines": [ + { + "bbox": [ + 438, + 193, + 466, + 198 + ], + "spans": [ + { + "bbox": [ + 438, + 193, + 466, + 198 + ], + "type": "text", + "content": "g) CSAU block" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 350, + 201, + 438, + 251 + ], + "blocks": [ + { + "bbox": [ + 350, + 201, + 438, + 251 + ], + "lines": [ + { + "bbox": [ + 350, + 201, + 438, + 251 + ], + "spans": [ + { + "bbox": [ + 350, + 201, + 438, + 251 + ], + "type": "image", + "image_path": "b3ee77e438653845a1e36cf64d406faa3fa2eea0fc0b76ffee8cc305be011c77.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 252, + 397, + 257 + ], + "lines": [ + { + "bbox": [ + 370, + 252, + 397, + 257 + ], + "spans": [ + { + "bbox": [ + 370, + 252, + 397, + 257 + ], + "type": "text", + "content": "d) SPFN block" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 349, + 312, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 349, + 312, + 361 + ], + "spans": [ + { + "bbox": [ + 132, + 349, + 312, + 361 + ], + "type": "text", + "content": "3.1 Generator's Encoder Network" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 371, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 371, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 482, + 491 + ], + "type": "text", + "content": "As shown in Fig. 2(a), the encoder of the proposed generator in Rawformer comprises stacked contextual-scale aware downsampler (CSAD) blocks to bridge the semantic gap between the source and target domains. Each CSAD block, depicted in Fig. 2(b), offers a holistic approach of understanding the contextual dependencies efficiently between the domains by capturing both the global and local information via its condensed query attention (CQA) block, in parallel with a " + }, + { + "bbox": [ + 130, + 371, + 482, + 491 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 371, + 482, + 491 + ], + "type": "text", + "content": " depth-wise convolution layer. It further handles the intrinsic correlations among the extracted features in a multi-scale manner via maintaining a synergistic collaboration between the scale perceptive feed-forward network (SPFN), and the composite downsampler (CDown) block." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "type": "text", + "content": "Condensed Query Attention Block: Self-attention is highly effective in modeling long-range contextual dependencies by comparing each token with all others, facilitating seamless information propagation within the network [58]. However, its computational demands pose challenges for tasks like unpaired raw-to-raw translation, where maintaining both visual quality and semantic consistency is crucial. To address these concerns and overcome the constraints of computation, we introduce the condensed query attention (CQA) block; see Fig. 2(c). Here, we make two key modifications to self-attention to enhance learning efficiency. Firstly, we spatially compress the query vector using average pooling followed by linear projection, and downsample the resulting feature map by a factor of " + }, + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "type": "text", + "content": " to confine the crucial information. This results in condensed queries, " + }, + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "type": "inline_equation", + "content": "(\\mathbf{Q}_c)" + }, + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "type": "text", + "content": " used for similarity comparison, as shown in Eq. 1. Secondly, to balance global and regional modeling with computational efficiency, attention is applied hor" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "Rawformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "iontally and vertically to eliminate redundant computations. The CQA block operation is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 192, + 149, + 449, + 161 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 149, + 449, + 161 + ], + "spans": [ + { + "bbox": [ + 192, + 149, + 449, + 161 + ], + "type": "interline_equation", + "content": "C Q A (\\mathbf {Q} _ {c}, \\mathbf {K}, \\mathbf {V}) = \\mathbf {A} _ {U}. (\\mathbf {A} _ {H} \\cdot \\mathbf {V}); \\quad \\mathbf {Q} _ {c} = W _ {p} (A v g P o o l (\\mathbf {Q}))", + "image_path": "f60c36b523e35b74452cc3695cb04aebd5475cd753fb8cf02f18f5134c81eb64.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 192, + 161, + 479, + 179 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 161, + 479, + 179 + ], + "spans": [ + { + "bbox": [ + 192, + 161, + 479, + 179 + ], + "type": "interline_equation", + "content": "\\mathbf {A} _ {H} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} _ {c} \\mathbf {K} ^ {T}}{\\sqrt {d}}\\right), \\quad \\mathbf {A} _ {U} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {Q} _ {c} ^ {T}}{\\sqrt {d}}\\right) \\tag {1}", + "image_path": "4236c775bf7015fa2250acd8dde64f5e9f9766f833e1e010eec832582bd4f3c9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 198, + 480, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 198, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 480, + 293 + ], + "type": "text", + "content": "where, the duet of key " + }, + { + "bbox": [ + 130, + 198, + 480, + 293 + ], + "type": "inline_equation", + "content": "(\\mathbf{K})" + }, + { + "bbox": [ + 130, + 198, + 480, + 293 + ], + "type": "text", + "content": ", and value " + }, + { + "bbox": [ + 130, + 198, + 480, + 293 + ], + "type": "inline_equation", + "content": "(\\mathbf{V})" + }, + { + "bbox": [ + 130, + 198, + 480, + 293 + ], + "type": "text", + "content": " vectors are obtained through plain linear convolutions, and the dimensions of each of these features are shown in the Fig. 2(c). " + }, + { + "bbox": [ + 130, + 198, + 480, + 293 + ], + "type": "inline_equation", + "content": "W_{p}(.)" + }, + { + "bbox": [ + 130, + 198, + 480, + 293 + ], + "type": "text", + "content": " denotes a " + }, + { + "bbox": [ + 130, + 198, + 480, + 293 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 130, + 198, + 480, + 293 + ], + "type": "text", + "content": " pointwise convolution, and " + }, + { + "bbox": [ + 130, + 198, + 480, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_H" + }, + { + "bbox": [ + 130, + 198, + 480, + 293 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 198, + 480, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_U" + }, + { + "bbox": [ + 130, + 198, + 480, + 293 + ], + "type": "text", + "content": " represents the compressed attention map between the key-condensed query pair horizontally and query-condensed query pair (for focusing on fine grained contextual information) vertically, respectively. The choice of operations and the reason for condensing the queries is investigated in the supplementary of the paper." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 312, + 480, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 480, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 480, + 417 + ], + "type": "text", + "content": "Scale Perceptive Feed-forward Network: To proficiently capture the intricacies of the local image structure across various scales by encoding the information from neighboring pixel locations, we propose a scale perceptive feed-forward network (SPFN); see Fig. 2(d). Unlike the conventional feed-forward networks that integrate single-scale depth-wise convolutions [58], we emphasize on enhancing the locality and inter-correlations among the incoming raw features of different domains by inserting multi-scale depth-wise convolutions. Given an input tensor at layer " + }, + { + "bbox": [ + 130, + 312, + 480, + 417 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 312, + 480, + 417 + ], + "type": "text", + "content": " and scale " + }, + { + "bbox": [ + 130, + 312, + 480, + 417 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 312, + 480, + 417 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 130, + 312, + 480, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_s^i \\in \\mathbb{R}^{H^i \\times W^i \\times C}" + }, + { + "bbox": [ + 130, + 312, + 480, + 417 + ], + "type": "text", + "content": ", SPFN is formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 170, + 417, + 479, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 417, + 479, + 441 + ], + "spans": [ + { + "bbox": [ + 170, + 417, + 479, + 441 + ], + "type": "interline_equation", + "content": "\\begin{array}{c} \\mathbf {X} _ {s} ^ {1} = \\phi \\left(W _ {d} ^ {3} \\left(W _ {p} ^ {1} (L N (\\mathbf {X} _ {s - 1}))\\right)\\right); \\mathbf {X} _ {s} ^ {2} = \\phi \\left(W _ {d} ^ {5} \\left(W _ {p} ^ {2} (L N (\\mathbf {X} _ {s - 1}))\\right)\\right), \\\\ \\mathbf {Y} _ {s} = W _ {s} ^ {3} \\left(\\mathbf {X} _ {o} ^ {1} (\\widehat {\\mathbf {C}}) \\mathbf {X} _ {s - 1} ^ {2}\\right) + \\mathbf {X} _ {s - 1}. \\end{array} \\tag {2}", + "image_path": "eca0a9764f7e16a7db742825e111171ec4d5799cbddb3ae7d09653ea5b9135fb.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 449, + 479, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 479, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 479, + 499 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 449, + 479, + 499 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 449, + 479, + 499 + ], + "type": "text", + "content": " represents a LeakyReLU activation function [35], " + }, + { + "bbox": [ + 130, + 449, + 479, + 499 + ], + "type": "inline_equation", + "content": "W_{d}^{k}" + }, + { + "bbox": [ + 130, + 449, + 479, + 499 + ], + "type": "text", + "content": " represents a depthwise convolution with filter size " + }, + { + "bbox": [ + 130, + 449, + 479, + 499 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 449, + 479, + 499 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 449, + 479, + 499 + ], + "type": "inline_equation", + "content": "LN" + }, + { + "bbox": [ + 130, + 449, + 479, + 499 + ], + "type": "text", + "content": " denotes the layer normalization [6], and " + }, + { + "bbox": [ + 130, + 449, + 479, + 499 + ], + "type": "inline_equation", + "content": "\\mathbb{C}" + }, + { + "bbox": [ + 130, + 449, + 479, + 499 + ], + "type": "text", + "content": " represents the concatenation operation. This enriches a more nuanced understanding of the underlying image features at different hierarchical levels." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 516, + 480, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 516, + 480, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 516, + 480, + 624 + ], + "type": "text", + "content": "Composite Down sampler: In general, the essence of downsampling is the retrieval of the multi-scale information in the image. Unlike the common encoder architectures for various image signal processing pipelines [31, 48] that deploy strided convolutions for downsampling, we design a hybrid strategy of leveraging the benefits of information exchange between different resolution features. To adaptively recover the underlying content information from the input domain at multiple scales, we design a composite down (CDown) block that intelligently combines pixel-unshuffle (to prevent the information loss) and strided convolution (to enhance the model's expressiveness) as shown in Fig. 2(e)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Style Modulator: Different cameras have their own distinctive characteristics to be handled or restored. To augment the expressiveness of Rawformer for" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 260, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 260, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 260, + 101 + ], + "type": "text", + "content": "G. Perevozhikov et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": "handling camera perturbations, we design a style modulator in the generator using extended pixel-wise ViT [51]. It aids in calibrating the decoder features and subsequently encourages the recovery of crucial details while translation. As demonstrated in Fig. 2(a), the style modulator applies style tokens in each contextual aware upsampler (CSAU) block, where every style token models the weight of the query vector of the CQA block to enrich the style specific context. This addition of style token at each stage in the decoder helps in the flexible adjustment of the feature maps. Unlike StyleGANv2 [27], where the style vectors are generated from the random prior, our approach resorts to learnable style token for conjecturing the important target style directly from the source domain, thus enhancing the consistency and control in raw image translation. More details on style modulation are provided in the supplementary." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 280, + 312, + 291 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 280, + 312, + 291 + ], + "spans": [ + { + "bbox": [ + 132, + 280, + 312, + 291 + ], + "type": "text", + "content": "3.2 Generator's Decoder Network" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 301, + 488, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 488, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 488, + 445 + ], + "type": "text", + "content": "Unlike the popular additive or multiplicative skip connections between the encoder-decoder, we also present a learnable linking layer (LLayer) as shown in the Fig. 2(a) to facilitate the feature preservation in the contextual scale aware upsampler (CSAU) block. The overall design of the CSAU block closely resembles that of the CSAD block, ensuring consistency and coherence in the network architecture in a bottom-up fashion. The primary distinction lies within the composite upsampling block, where for inducing the relevant spatial information into the target domain for the incoming features from the style modulator, we design a composite upsampler block (CUp). It embodies a parallel combination of pixel-shuffle (for preventing the checkerboard artifacts in the target domain) and deconvolution (for combating with aliasing) as shown in Fig. 2(f) and (g)." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 160, + 468, + 455, + 575 + ], + "blocks": [ + { + "bbox": [ + 160, + 468, + 455, + 575 + ], + "lines": [ + { + "bbox": [ + 160, + 468, + 455, + 575 + ], + "spans": [ + { + "bbox": [ + 160, + 468, + 455, + 575 + ], + "type": "image", + "image_path": "59fa1358f17687797ea4aa3a662abffb9ff0bfe01aaebfcd692437afc70c3cff.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 583, + 482, + 640 + ], + "lines": [ + { + "bbox": [ + 130, + 583, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 583, + 482, + 640 + ], + "type": "text", + "content": "Fig. 3: a) The proposed cross-domain attention-guided discriminator. b) Feature map visualizations without (w/o) and with (w) the discriminator head. The inclusion of the discriminator head aids in refining the overall results, and the discriminator training (as shown in c) benefits from the inclusion of the discriminator head, particularly with the cache " + }, + { + "bbox": [ + 130, + 583, + 482, + 640 + ], + "type": "inline_equation", + "content": "(M)" + }, + { + "bbox": [ + 130, + 583, + 482, + 640 + ], + "type": "text", + "content": " component." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 401, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 401, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 401, + 91, + 447, + 100 + ], + "type": "text", + "content": "Rawformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 397, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 397, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 397, + 128 + ], + "type": "text", + "content": "3.3 Cross-Domain Attention-Guided Discriminator" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 133, + 482, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 133, + 482, + 336 + ], + "spans": [ + { + "bbox": [ + 130, + 133, + 482, + 336 + ], + "type": "text", + "content": "Leveraging batch statistics is an effective strategy to mitigate mode collapse and enhance diversity in GAN models [43]. Traditional minibatch discrimination techniques require large batch sizes [26], posing challenges due to GPU limitations. Inspired by [50], we address this by decoupling batch size from the minibatch discrimination approach. Our core innovation involves using a cache of previous discriminator features as an alternative to larger batch sizes. During training, batch features are stored in four separate caches " + }, + { + "bbox": [ + 130, + 133, + 482, + 336 + ], + "type": "inline_equation", + "content": "(M)" + }, + { + "bbox": [ + 130, + 133, + 482, + 336 + ], + "type": "text", + "content": "; real images from the source and target cameras, and fake images from both domains. This memory bank-style cache utilizes a composite discriminator with a head and body. The head captures batch statistics via batch normalization and two " + }, + { + "bbox": [ + 130, + 133, + 482, + 336 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 133, + 482, + 336 + ], + "type": "text", + "content": " convolutional layers, processing a concatenated input of current and historical discriminator outputs from the cache, as shown in Fig. 3(a). The discriminator body includes a CQA block followed by stripped convolution to handle detailed global features across domains. This integration of batch statistics, caching, and attention mechanisms significantly enhances the discriminator's ability to manage multi-resolution image data and stabilize the overall training (Figs. 3 b and c)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 350, + 307, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 350, + 307, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 350, + 307, + 363 + ], + "type": "text", + "content": "3.4 Self-Supervised Pre-Training" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 365, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 475 + ], + "type": "text", + "content": "Self-supervised pre-training optimizes the initial weights of a network by engaging it in auxiliary tasks formulated from the inherent structure of the original data. In general, despite modifications in the cross-domain attention-guided discriminator, GAN training still suffers from mode collapse, where the generator repeatedly tends to produce a limited set of images. Thus, we opt to pre-train the generator of our Rawformer using self-supervised inpainting tasks to enhance initial weight optimization and mitigate this mode collapse problem. More details about the mode of pretraining is provided in the supplementary (Section 1) of the paper." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 488, + 230, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 488, + 230, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 488, + 230, + 502 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "We evaluated our method using two types of datasets: 1) raw-to-raw datasets, which provide unpaired raw images for training and aligned raw images from source and target cameras for assessing unsupervised raw-to-raw translation accuracy (Sec. 4.3), and 2) raw-to-sRGB datasets, offering aligned raw and corresponding sRGB images for evaluating transformed raw image quality followed by learnable neural-based ISP rendering (Sec. 4.4). This evaluation mirrors real-world usage scenarios where raw images from a new camera are translated to the raw space of the camera used for ISP training. In all the experiments, raw images were demosaiced using the Menon algorithm [36] and normalized after black-level subtraction. During our method's training, each raw image was segmented into non-overlapping crops of " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " pixels. Similarly, we used paired raw-sRGB non-overlapping crops of " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " pixels to train the camera-specific LAN ISP [54] on each raw-to-sRGB dataset." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 261, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 261, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 261, + 102 + ], + "type": "text", + "content": "G. Perevozhikov et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 205, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 205, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 205, + 126 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 482, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 482, + 266 + ], + "type": "text", + "content": "We utilized two raw-to-raw datasets: 1) the Raw-to-Raw dataset [4] and 2) the NUS dataset [9], and three raw-to-sRGB datasets: 1) the Zurich raw-toRGB dataset [20], 2) the Samsung S7 dataset [44], and 3) the Mobile AIM21 dataset [19]. The Raw-to-Raw dataset [4] includes raw images from the Samsung Galaxy S9 and the iPhone X, featuring both unpaired and paired sets. The NUS dataset [9] comprises scenes captured by eight DSLR cameras. Following the evaluation in [4], we utilized the Nikon D5200 and Canon EOS 600D DSLR cameras from the NUS dataset [9]. Aligned paired raw images were generated using the same approach as in the Raw-to-Raw dataset [4]. In both datasets, paired raw images are dedicated for testing, while unpaired images are utilized for unsupervised training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 266, + 482, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 352 + ], + "type": "text", + "content": "The Zurich raw-to-RGB dataset [20] includes pairs of raw and sRGB images captured synchronously using a Huawei P20 smartphone camera and a Canon 5D Mark IV DSLR camera. The Samsung S7 dataset [44] comprises paired raw and sRGB images captured by the Samsung S7 smartphone's main camera. The Mobile AIM21 dataset [19] includes images captured using a Sony IMX586 mobile sensor and a Fujimel GFX100 DSLR, employing an advanced dense correspondence algorithm for matching due to imperfect alignment [52]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 366, + 240, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 366, + 240, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 366, + 240, + 379 + ], + "type": "text", + "content": "4.2 Training details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 385, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 482, + 495 + ], + "type": "text", + "content": "Inspired by the effective fine-tuning capabilities demonstrated by the pre-trained BERT model for the text completion task [14], we initially trained each generator network in a self-supervised fashion for image completion. Initially, the generator was trained on " + }, + { + "bbox": [ + 130, + 385, + 482, + 495 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 130, + 385, + 482, + 495 + ], + "type": "text", + "content": " pixel patches extracted from training images, with " + }, + { + "bbox": [ + 130, + 385, + 482, + 495 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 130, + 385, + 482, + 495 + ], + "type": "text", + "content": " of these patches being randomly masked out. The objective of the generator during this stage was to reconstruct the original unmasked image from its partially obscured version. Training was performed using the AdamW optimizer [34] with betas set to (0.9, 0.99) and a weight decay of 0.05 for 500 epochs to minimize the following pixel-wise loss function:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 219, + 504, + 480, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 504, + 480, + 519 + ], + "spans": [ + { + "bbox": [ + 219, + 504, + 480, + 519 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p i x e l - w i s e}} = L _ {1} + (1 - S S I M) + V G G \\tag {3}", + "image_path": "1adc3eda63673729d5f17e598c02ffd1963ad0e84aca8bfd51f18eb565d242bc.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "inline_equation", + "content": "SSIM" + }, + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "inline_equation", + "content": "VGG" + }, + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "text", + "content": " refer to L1 loss, Structural Similarity Index (SSIM) [53], and VGG perceptual loss [30] between original unmasked image and the reconstructed image. The learning rate was initialized at 0.005, and the learning schedule was managed using the cosine annealing warm restarts strategy [33]. This self-supervised pre-training was shown to improve the raw mapping results (see Table 1)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Thereafter, the weights of both generators were fine-tuned for our unpaired raw-to-raw translation task for an additional 500 epochs, while simultaneously training the discriminator networks. We use the Adam optimizer [29] with betas set to (0.5, 0.99) and a learning rate of 0.0001 for the discriminator networks and 0.00005 for the generator networks. The discriminator networks were optimized to minimize the following loss functions:" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 101 + ], + "type": "text", + "content": "Rawformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 463, + 251 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 463, + 251 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 463, + 251 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 463, + 251 + ], + "type": "image", + "image_path": "c1d10917b3b76d69b428eef7766a51a1a9a98a0241aea81ca61fee702e3d82c6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 260, + 482, + 316 + ], + "lines": [ + { + "bbox": [ + 130, + 260, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 482, + 316 + ], + "type": "text", + "content": "Fig. 4: Shown are two images captured by Samsung S9 and iPhone X in sRGB (left) and two cropped patches from each image in raw (right). On the right, we show the input raw patch from the corresponding camera and the corresponding ground-truth raw patch from the other camera, along with the results by other methods. Our proposed Rawformer is better at preserving the domain consistent features." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 202, + 351, + 481, + 366 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 351, + 481, + 366 + ], + "spans": [ + { + "bbox": [ + 202, + 351, + 481, + 366 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {A} ^ {d i s} = \\ell_ {g a n} \\left(D _ {A} \\left(G _ {B \\rightarrow A} (b)\\right), 0\\right) + \\ell_ {g a n} \\left(D _ {A} (a), 1\\right) \\tag {4}", + "image_path": "7b9d9916a69b1245c67e3eca4fc5ac71ff51e38e7452d465531b870f193fcc48.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 202, + 383, + 481, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 383, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 202, + 383, + 481, + 397 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {B} ^ {d i s} = \\ell_ {g a n} \\left(D _ {B} \\left(G _ {A \\rightarrow B} (a)\\right), 0\\right) + \\ell_ {g a n} \\left(D _ {B} (b), 1\\right) \\tag {5}", + "image_path": "e0a1de7010a1b89c2287f2ee4ce3003dd33fb71eb861df178651bfba2c8a8b02.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 403, + 480, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 480, + 450 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 403, + 480, + 450 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 130, + 403, + 480, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 403, + 480, + 450 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 403, + 480, + 450 + ], + "type": "text", + "content": " represent images from camera " + }, + { + "bbox": [ + 130, + 403, + 480, + 450 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 403, + 480, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 403, + 480, + 450 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 403, + 480, + 450 + ], + "type": "text", + "content": ", respectively, while " + }, + { + "bbox": [ + 130, + 403, + 480, + 450 + ], + "type": "inline_equation", + "content": "G_{B\\to A}(b)" + }, + { + "bbox": [ + 130, + 403, + 480, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 403, + 480, + 450 + ], + "type": "inline_equation", + "content": "G_{A\\to B}(a)" + }, + { + "bbox": [ + 130, + 403, + 480, + 450 + ], + "type": "text", + "content": " are the translated images produced by each generator network, respectively. The labels 0 and 1 represent \"fake\" and \"real\" raw images, respectively, and " + }, + { + "bbox": [ + 130, + 403, + 480, + 450 + ], + "type": "inline_equation", + "content": "\\ell_{gan}(\\cdot)" + }, + { + "bbox": [ + 130, + 403, + 480, + 450 + ], + "type": "text", + "content": " computes the cross-entropy loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 451, + 481, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 481, + 474 + ], + "type": "text", + "content": "During this stage, the weights of both generators were optimized to minimize the following loss function:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 173, + 486, + 480, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 486, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 173, + 486, + 480, + 500 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\text {g e n}} = \\beta_ {1} \\left(\\mathcal {L} _ {A} ^ {\\text {g a n}} + \\mathcal {L} _ {B} ^ {\\text {g a n}}\\right) + \\beta_ {2} \\left(\\mathcal {L} _ {A} ^ {\\text {i d t}} + \\mathcal {L} _ {B} ^ {\\text {i d t}}\\right) + \\beta_ {3} \\left(\\mathcal {L} _ {A} ^ {\\text {c y c}} + \\mathcal {L} _ {B} ^ {\\text {c y c}}\\right) \\tag {6}", + "image_path": "9931027c8f8fa03a0770727316488db6e61ee8c6282d8d6a7bf67ff93a50a1d8.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 505, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 505, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 505, + 482, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 505, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 1, \\beta_{2} = 10, \\beta_{3} = 0.5, \\mathcal{L}_{A}^{gan} = \\ell_{gan}(D_{B}(G_{A\\rightarrow B}(a)),1)," + }, + { + "bbox": [ + 130, + 505, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_A^{idt} = \\mathcal{L}_{pixel - wise}(G_{B\\to A}(a),a)" + }, + { + "bbox": [ + 130, + 505, + 482, + 567 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 505, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_A^{cyc} = \\mathcal{L}_{pixel - wise}(G_{B\\to A}(G_{A\\to B}(a)),a)" + }, + { + "bbox": [ + 130, + 505, + 482, + 567 + ], + "type": "text", + "content": ". Note that at inference time, only a single generator network (e.g., " + }, + { + "bbox": [ + 130, + 505, + 482, + 567 + ], + "type": "inline_equation", + "content": "G_{A\\rightarrow B}" + }, + { + "bbox": [ + 130, + 505, + 482, + 567 + ], + "type": "text", + "content": ") is required to translate raw images captured by the source camera (e.g., " + }, + { + "bbox": [ + 130, + 505, + 482, + 567 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 505, + 482, + 567 + ], + "type": "text", + "content": ") to the target camera (e.g., " + }, + { + "bbox": [ + 130, + 505, + 482, + 567 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 505, + 482, + 567 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 585, + 326, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 585, + 326, + 596 + ], + "spans": [ + { + "bbox": [ + 131, + 585, + 326, + 596 + ], + "type": "text", + "content": "4.3 Raw-to-Raw Translation Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "We compare our raw-to-raw translation results with those of the recent semi-supervised raw-to-raw mapping (SSRM) method [4]. Additionally, we compare our method against several established unsupervised generic I2I methodologies, including UVCGANv2 [51], U-GAT-IT [28], UNIT [32], and the classic CycleGAN [62]. All other methods were trained on the same training data used for" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 260, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 260, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 260, + 100 + ], + "type": "text", + "content": "G. Perevozchikov et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 159, + 190, + 455, + 262 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 180 + ], + "type": "text", + "content": "Table 1: Raw-to-raw translation results on NUS dataset [9]. The mapping results of Nikon D5200's raw images to Canon EOS 600D DSLR camera and vice versa using our method and other methods are shown. Here, " + }, + { + "bbox": [ + 132, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "\\ell_{a} = L_{1}" + }, + { + "bbox": [ + 132, + 114, + 482, + 180 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "\\ell_{b} = L_{1} + VGG" + }, + { + "bbox": [ + 132, + 114, + 482, + 180 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "\\ell_{c} = \\mathcal{L}_{pixel - wise}" + }, + { + "bbox": [ + 132, + 114, + 482, + 180 + ], + "type": "text", + "content": " (Eq. 3), w, and w/o denote the results with and without training on a corresponding configuration, respectively. Additionally, we show the GFLOPs of each method for the " + }, + { + "bbox": [ + 132, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "256\\times 256" + }, + { + "bbox": [ + 132, + 114, + 482, + 180 + ], + "type": "text", + "content": " input image. Best results are in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 159, + 190, + 455, + 262 + ], + "lines": [ + { + "bbox": [ + 159, + 190, + 455, + 262 + ], + "spans": [ + { + "bbox": [ + 159, + 190, + 455, + 262 + ], + "type": "table", + "html": "
MethodsCanon-to-NikonNikon-to-Canon
PSNR↑SSIM↑MAE↓(x1e-2)ΔE↓PSNR↑SSIM↑MAE↓(x1e-2)ΔE↓
CycleGAN [62]27.320.833.2115.4026.810.823.3311.4936.8
UNIT [32]25.550.813.1718.2224.730.793.2318.6447.3
U-GAT-IT [28]29.410.873.027.1227.860.853.298.0162.9
SSRM [4]32.360.932.416.2130.810.932.475.9537.1
UVCGANv2 [51]37.110.962.324.3437.290.962.284.2869.6
Ours (w/o self-supervised pre-training; ℓe loss)37.130.962.294.6537.270.962.314.4652.7
Ours (w/o self-supervised pre-training; ℓo loss)37.940.972.213.9637.990.962.233.9352.7
Ours (w/o self-supervised pre-training; ℓe loss)38.730.972.193.6438.260.972.223.7252.7
Ours (w self-supervised pre-training; ℓe loss)41.890.981.462.0441.370.981.412.5352.7
", + "image_path": "fea6e8e59d189ab7e7719b8f06b6035abd5214f87577e398cfa3875961414df4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 159, + 350, + 455, + 415 + ], + "blocks": [ + { + "bbox": [ + 132, + 274, + 482, + 340 + ], + "lines": [ + { + "bbox": [ + 132, + 274, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 482, + 340 + ], + "type": "text", + "content": "Table 2: Raw-to-raw translation results on the Raw-to-Raw dataset [4]. Shown are the mapping results of Samsung Galaxy S9's raw images to iPhone X and vice versa using our method and other methods. Additionally, the total number of parameters of each method is shown in millions (M), and inference time in milliseconds (ms) without optimization on NVIDIA GeForce RTX 4090 GPU. Results on inference time with optimization are in supplementary." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 159, + 350, + 455, + 415 + ], + "lines": [ + { + "bbox": [ + 159, + 350, + 455, + 415 + ], + "spans": [ + { + "bbox": [ + 159, + 350, + 455, + 415 + ], + "type": "table", + "html": "
MethodsSamsung-to-iPhoneiPhone-to-Samsung
PSNR↑SSIM↑MAE↓(×1e-2)ΔE↓PSNR↑SSIM↑MAE↓(×1e-2)ΔE↓
CycleGAN [62]24.630.715.4014.7124.370.734.3412.8821.719
UNIT [32]23.910.705.1513.2326.420.784.2713.1632.229
U-GAT-IT [28]28.220.893.115.9831.040.892.395.0434.435
SSRM [4]29.650.892.226.3228.580.902.436.5320.218
UVCGANv2 [51]36.320.942.094.2136.460.922.024.7332.631
Ours40.980.971.332.0941.480.981.211.9926.126
", + "image_path": "a49a48c4591448973b664eab28aeee10cb9054ff50a1e65afd29602fe959b238.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 440, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 482, + 548 + ], + "type": "text", + "content": "our method, for 500 epochs. Since SSRM is not entirely unsupervised, we incorporated 22 image pairs from the paired subset for its training, as discussed in [4]. Fig. 4 shows qualitative comparisons on the Raw-to-Raw dataset [4], while Tables 1 and 2 present the quantitative results on the NUS dataset [9] and the Raw-to-Raw dataset, respectively. In Tables 1 and 2, we adopted the same quantitative metrics used in [4]: peak signal-to-noise ratio (PSNR), SSIM, mean absolute error (MAE), and " + }, + { + "bbox": [ + 132, + 440, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\Delta E" + }, + { + "bbox": [ + 132, + 440, + 482, + 548 + ], + "type": "text", + "content": " 2000 [46]. As can be seen in both the tables, our Rawformer significantly outperforms the alternative methods across all the quantitative metrics by a large margin (around " + }, + { + "bbox": [ + 132, + 440, + 482, + 548 + ], + "type": "inline_equation", + "content": "+12" + }, + { + "bbox": [ + 132, + 440, + 482, + 548 + ], + "type": "text", + "content": " dB against SSRM [4])." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 570, + 482, + 665 + ], + "type": "text", + "content": "Evaluation on Low-Light Images: Translating raw-to-raw for low-light images presents challenges due to various variations among cameras, including differences in raw normalization and noise characteristics [2,4]. To further prove the efficacy of our method against other methods on low-light images, we selected a subset of 15 paired low-light images from the Raw-to-Raw dataset [4]. The results on this subset are presented in Table 3, where our proposed method significantly outperforms SSRM [4], and UVCGANv2 [51] for translating low-light raw images captured by an iPhone X to the Samsung S9's camera." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 401, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 401, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 401, + 91, + 447, + 100 + ], + "type": "text", + "content": "Rawformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 136, + 332, + 194 + ], + "blocks": [ + { + "bbox": [ + 143, + 136, + 332, + 194 + ], + "lines": [ + { + "bbox": [ + 143, + 136, + 332, + 194 + ], + "spans": [ + { + "bbox": [ + 143, + 136, + 332, + 194 + ], + "type": "image", + "image_path": "f24c88d4eb227a1b29742086fbd59bb66d550c827c7849a3bedac3bea665d48b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 364, + 143, + 460, + 187 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 126 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 126 + ], + "type": "text", + "content": "Table 3: Raw-to-Raw dataset [4] results on low-light raw images (a subset)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 364, + 143, + 460, + 187 + ], + "lines": [ + { + "bbox": [ + 364, + 143, + 460, + 187 + ], + "spans": [ + { + "bbox": [ + 364, + 143, + 460, + 187 + ], + "type": "table", + "html": "
MethodPSNR↑SSIM↑
SSRM [4]23.870.85
UVCGANv2 [51]35.220.94
Ours40.210.97
", + "image_path": "fef6e1b78374d75b5182c7477a27f8f62f4afa775fd9514c5a6a503b4cc6e329.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 203, + 258, + 410, + 311 + ], + "blocks": [ + { + "bbox": [ + 130, + 204, + 481, + 248 + ], + "lines": [ + { + "bbox": [ + 130, + 204, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 204, + 481, + 248 + ], + "type": "text", + "content": "Table 4: Ablation results on the impact of various configurations of the generator model on the NUS dataset [9]. 'CQA', 'Style', 'SPFN' and 'CUp-CDown' refers to the condensed query attention, style modulator, scale perceptive feed-forward network, and the composite up and downsampler, respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 203, + 258, + 410, + 311 + ], + "lines": [ + { + "bbox": [ + 203, + 258, + 410, + 311 + ], + "spans": [ + { + "bbox": [ + 203, + 258, + 410, + 311 + ], + "type": "table", + "html": "
ConfigurationModulesCanon-to-NikonNikon-to-Canon
BaselineCQAStyleSPFNCUp-CdownPSNR↑SSIM↑PSNR↑SSIM↑
G1✓37.230.9637.310.96
G2✓✓38.210.9637.970.96
G3✓✓✓39.080.9738.990.96
G4✓✓✓✓39.720.9739.680.97
G5✓✓✓✓✓39.930.9739.860.97
", + "image_path": "58319ddbbb8177a9a83b068f576869987d666ac2cb5a54cc82383b5f2fa49bbe.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 197, + 353, + 417, + 404 + ], + "blocks": [ + { + "bbox": [ + 130, + 321, + 479, + 343 + ], + "lines": [ + { + "bbox": [ + 130, + 321, + 479, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 479, + 343 + ], + "type": "text", + "content": "Table 5: Ablation results on the impact of various discriminator model configurations on the NUS dataset [9]. 'Baseline' refers to the " + }, + { + "bbox": [ + 130, + 321, + 479, + 343 + ], + "type": "inline_equation", + "content": "G_{5}" + }, + { + "bbox": [ + 130, + 321, + 479, + 343 + ], + "type": "text", + "content": " config. from Tab. 4." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 197, + 353, + 417, + 404 + ], + "lines": [ + { + "bbox": [ + 197, + 353, + 417, + 404 + ], + "spans": [ + { + "bbox": [ + 197, + 353, + 417, + 404 + ], + "type": "table", + "html": "
ConfigurationModulesCanon-to-NikonNikon-to-Canon
BaselineCQABatchnormDisc. headPSNR↑SSIM↑PSNR↑SSIM↑
D1✓39.930.9739.860.97
D2✓✓40.120.9740.090.97
D3✓✓✓40.240.9740.210.97
D4✓✓✓✓41.890.9841.370.98
", + "image_path": "823ba6e03b87b699d9223054be375f096edc6cc1dc7c8097f14b3cb0746c80c9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 426, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 582 + ], + "type": "text", + "content": "Ablation Studies: We conducted a set of ablation studies to validate the impact of the proposed components in the generator and discriminator networks. In this series of experiments, we trained our Rawformer on the NUS dataset [9], in addition to a baseline model trained using the same training settings as our Rawformer. The baseline model comprises a UNet architecture with channel and spatial attention [50] for the generator, and a PatchGAN discriminator [23]. As demonstrated in Table 4, the integration of the proposed modules into the baseline consistently improves the network's capabilities, showcasing an enhancement in overall accuracy. We also conducted further ablation studies that explore various configurations of the discriminator, as shown in Table 5. The results clearly underscore the improvement by the inclusion of each proposed component in the discriminator network (attention, batch normalization, and discriminator head (Disc. head)) in improving the accuracy of our raw-to-raw translation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 598, + 337, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 598, + 337, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 598, + 337, + 609 + ], + "type": "text", + "content": "4.4 Raw Translation for Learnable ISP" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "To comprehend the practical application of our model in real-world scenarios, we utilized three datasets: Zurich raw-to-RGB, Samsung S7 ISP, and Mobile AIM21, encompassing various source and target domain combinations. Specifically, for each dataset, we trained the LAN ISP [54] on a single paired raw-to-sRGB" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 260, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 260, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 260, + 100 + ], + "type": "text", + "content": "G. Perevozhikov et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 187, + 213, + 428, + 260 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "text", + "content": "Table 6: Cross-camera raw-to-sRGB results (PSNR/SSIM) comparing: a) UVCGANv2 [51] with b) our Rawformer. Each entry compares ground-truth sRGB images generated by the function: " + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "inline_equation", + "content": "ISP_{B}(F^{B}(raw_{A}))" + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "inline_equation", + "content": "F^{B}" + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "text", + "content": " denotes translation of raw images, " + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "inline_equation", + "content": "raw_{A}" + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "text", + "content": ", captured by camera " + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "text", + "content": " to camera " + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "inline_equation", + "content": "ISP_{B}(\\cdot)" + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "text", + "content": " produces sRGB images using the LAN ISP [54] trained on raw images from camera " + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "text", + "content": ". Diagonal results (in red) represent the camera-specific ISP case and are the sRGB images produced by the ISP model tested on the same camera used for training. Best cross-camera results are in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 187, + 213, + 428, + 260 + ], + "lines": [ + { + "bbox": [ + 187, + 213, + 428, + 260 + ], + "spans": [ + { + "bbox": [ + 187, + 213, + 428, + 260 + ], + "type": "table", + "html": "
Training camera Aa) UVCGANv2 [51]+ LAN ISP [54]b) Our Rawformer+ LAN ISP [54]
Testing camera BTesting camera B
ZRR [20]AIM21 [19]S7 [44]ZRR [20]AIM21 [19]S7 [44]
ZRR [20]19.46/0.7317.72/0.7017.81/0.7019.46/0.7318.97/0.7219.02/0.72
AIM21 [19]22.14/0.7923.48/0.8722.39/0.7923.25/0.8623.48/0.8723.08/0.86
S7 [44]21.07/0.8020.93/0.7922.16/0.8122.08/0.8022.03/0.8022.16/0.81
", + "image_path": "76fe4056b37f62850408284510c70eff7e584ff7dfd2bb33cb5eeffcef6861e3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 161, + 274, + 201, + 314 + ], + "blocks": [ + { + "bbox": [ + 161, + 274, + 201, + 314 + ], + "lines": [ + { + "bbox": [ + 161, + 274, + 201, + 314 + ], + "spans": [ + { + "bbox": [ + 161, + 274, + 201, + 314 + ], + "type": "image", + "image_path": "4ede32badba377a435c3de8c64be3f47c63a858eae6aa6126229fd9af633967a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 161, + 316, + 200, + 356 + ], + "blocks": [ + { + "bbox": [ + 161, + 316, + 200, + 356 + ], + "lines": [ + { + "bbox": [ + 161, + 316, + 200, + 356 + ], + "spans": [ + { + "bbox": [ + 161, + 316, + 200, + 356 + ], + "type": "image", + "image_path": "3e01e560cde71fd95bd3f329f789760652fded4245422acf2ad280772bb33d92.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 161, + 357, + 200, + 397 + ], + "blocks": [ + { + "bbox": [ + 161, + 357, + 200, + 397 + ], + "lines": [ + { + "bbox": [ + 161, + 357, + 200, + 397 + ], + "spans": [ + { + "bbox": [ + 161, + 357, + 200, + 397 + ], + "type": "image", + "image_path": "dcd909ffc277820f7dda4f11df6b9013f8052be6e0a70796e18077d5b49f019c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 397, + 194, + 403 + ], + "lines": [ + { + "bbox": [ + 168, + 397, + 194, + 403 + ], + "spans": [ + { + "bbox": [ + 168, + 397, + 194, + 403 + ], + "type": "text", + "content": "GT (ZRR)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 417, + 482, + 483 + ], + "lines": [ + { + "bbox": [ + 130, + 417, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 417, + 482, + 483 + ], + "type": "text", + "content": "Fig.5: ISP rendering results with our raw translation on various datasets. Each set includes ground-truth (GT) raw-sRGB paired images and LAN ISP [54] results on mapped raw images from different cameras. S7, ZRR, and AIM stand for the Samsung S7 ISP dataset [44], Zurich raw-to-RGB dataset [20], and Mobile AIM21 dataset [19], respectively. The shown results are consistent with the ground-truth, demonstrating the proficiency of our model." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 203, + 274, + 242, + 314 + ], + "blocks": [ + { + "bbox": [ + 203, + 274, + 242, + 314 + ], + "lines": [ + { + "bbox": [ + 203, + 274, + 242, + 314 + ], + "spans": [ + { + "bbox": [ + 203, + 274, + 242, + 314 + ], + "type": "image", + "image_path": "cc12e9be4c46b4678bb40efe836fe993c0de6061f6ff54df39251a644a3fcf77.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 203, + 316, + 242, + 356 + ], + "blocks": [ + { + "bbox": [ + 203, + 316, + 242, + 356 + ], + "lines": [ + { + "bbox": [ + 203, + 316, + 242, + 356 + ], + "spans": [ + { + "bbox": [ + 203, + 316, + 242, + 356 + ], + "type": "image", + "image_path": "583ef5c178ee8ff30e2d53a81b446b9e141d580b87d2bed813cc9ca57ace131c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 203, + 357, + 242, + 397 + ], + "blocks": [ + { + "bbox": [ + 203, + 357, + 242, + 397 + ], + "lines": [ + { + "bbox": [ + 203, + 357, + 242, + 397 + ], + "spans": [ + { + "bbox": [ + 203, + 357, + 242, + 397 + ], + "type": "image", + "image_path": "67c3c27f66d5c1901334932999c69d01178d2afbdb08a36002b78c426032f964.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 201, + 397, + 244, + 404 + ], + "lines": [ + { + "bbox": [ + 201, + 397, + 244, + 404 + ], + "spans": [ + { + "bbox": [ + 201, + 397, + 244, + 404 + ], + "type": "inline_equation", + "content": "\\mathrm{ISP}_{\\mathrm{AIM}}(\\mathrm{F}^{\\mathrm{AIM}}(\\mathrm{raw}_{\\mathrm{RRB}}))" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 245, + 274, + 285, + 314 + ], + "blocks": [ + { + "bbox": [ + 245, + 274, + 285, + 314 + ], + "lines": [ + { + "bbox": [ + 245, + 274, + 285, + 314 + ], + "spans": [ + { + "bbox": [ + 245, + 274, + 285, + 314 + ], + "type": "image", + "image_path": "dfbc321511caa4630efaee585e2a330e6d88e47eb7641538feeffd9cbc89e646.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 245, + 316, + 285, + 356 + ], + "blocks": [ + { + "bbox": [ + 245, + 316, + 285, + 356 + ], + "lines": [ + { + "bbox": [ + 245, + 316, + 285, + 356 + ], + "spans": [ + { + "bbox": [ + 245, + 316, + 285, + 356 + ], + "type": "image", + "image_path": "e7b343cc1bc9bf1534ede3f48fdba81e500484017e621d12832dc8da9905ad38.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 245, + 357, + 285, + 397 + ], + "blocks": [ + { + "bbox": [ + 245, + 357, + 285, + 397 + ], + "lines": [ + { + "bbox": [ + 245, + 357, + 285, + 397 + ], + "spans": [ + { + "bbox": [ + 245, + 357, + 285, + 397 + ], + "type": "image", + "image_path": "dacdc8543859dd5619c7db16497000895bda0bf7d8950fc158c9b40405ec7e42.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 253, + 397, + 277, + 403 + ], + "lines": [ + { + "bbox": [ + 253, + 397, + 277, + 403 + ], + "spans": [ + { + "bbox": [ + 253, + 397, + 277, + 403 + ], + "type": "text", + "content": "GT (AIM)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 286, + 274, + 326, + 314 + ], + "blocks": [ + { + "bbox": [ + 286, + 274, + 326, + 314 + ], + "lines": [ + { + "bbox": [ + 286, + 274, + 326, + 314 + ], + "spans": [ + { + "bbox": [ + 286, + 274, + 326, + 314 + ], + "type": "image", + "image_path": "07b4afa3c84b206ae4af3f2bd5f880f010dd2d3490dc2e4da023c872fc36f40b.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 286, + 316, + 326, + 356 + ], + "blocks": [ + { + "bbox": [ + 286, + 316, + 326, + 356 + ], + "lines": [ + { + "bbox": [ + 286, + 316, + 326, + 356 + ], + "spans": [ + { + "bbox": [ + 286, + 316, + 326, + 356 + ], + "type": "image", + "image_path": "14725fd33bb29868313aeb7e1d722c9cfbac60164554d9f35bdbd6994b88f760.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 286, + 357, + 326, + 397 + ], + "blocks": [ + { + "bbox": [ + 286, + 357, + 326, + 397 + ], + "lines": [ + { + "bbox": [ + 286, + 357, + 326, + 397 + ], + "spans": [ + { + "bbox": [ + 286, + 357, + 326, + 397 + ], + "type": "image", + "image_path": "90ae8e7a255b52f2d6b669fcfa778beb479bbc674b03d8775b23a06cde21b718.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 286, + 397, + 369, + 404 + ], + "lines": [ + { + "bbox": [ + 286, + 397, + 369, + 404 + ], + "spans": [ + { + "bbox": [ + 286, + 397, + 369, + 404 + ], + "type": "inline_equation", + "content": "\\mathrm{ISP}_{\\mathrm{ZRR}}(\\mathrm{F}^{\\mathrm{ZRR}}(\\mathrm{raw}_{\\mathrm{AIM}}))\\mathrm{ISP}_{\\mathrm{S7}}(\\mathrm{F}^{\\mathrm{ST}}(\\mathrm{raw}_{\\mathrm{AIM}}))" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 370, + 274, + 410, + 314 + ], + "blocks": [ + { + "bbox": [ + 370, + 274, + 410, + 314 + ], + "lines": [ + { + "bbox": [ + 370, + 274, + 410, + 314 + ], + "spans": [ + { + "bbox": [ + 370, + 274, + 410, + 314 + ], + "type": "image", + "image_path": "368226994a734226f180796668fb2963f1018452846f068438ecc11bdcec71b8.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 370, + 316, + 410, + 356 + ], + "blocks": [ + { + "bbox": [ + 370, + 316, + 410, + 356 + ], + "lines": [ + { + "bbox": [ + 370, + 316, + 410, + 356 + ], + "spans": [ + { + "bbox": [ + 370, + 316, + 410, + 356 + ], + "type": "image", + "image_path": "7353958554e1f0934633aadf26862e41d399ef2e62a85bcd095ac8d9d85de26f.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 370, + 357, + 410, + 397 + ], + "blocks": [ + { + "bbox": [ + 370, + 357, + 410, + 397 + ], + "lines": [ + { + "bbox": [ + 370, + 357, + 410, + 397 + ], + "spans": [ + { + "bbox": [ + 370, + 357, + 410, + 397 + ], + "type": "image", + "image_path": "10526664d7c0d64cbd22034a159e4e93d756bb7bf4de4e0adf7e1bd0014ddc06.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 397, + 400, + 403 + ], + "lines": [ + { + "bbox": [ + 381, + 397, + 400, + 403 + ], + "spans": [ + { + "bbox": [ + 381, + 397, + 400, + 403 + ], + "type": "text", + "content": "GT (S7)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 412, + 274, + 452, + 314 + ], + "blocks": [ + { + "bbox": [ + 412, + 274, + 452, + 314 + ], + "lines": [ + { + "bbox": [ + 412, + 274, + 452, + 314 + ], + "spans": [ + { + "bbox": [ + 412, + 274, + 452, + 314 + ], + "type": "image", + "image_path": "b433f06f08ba8e3813ace1980b87e73083f2e0221efdfa4f8060eac4df548018.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 412, + 316, + 452, + 356 + ], + "blocks": [ + { + "bbox": [ + 412, + 316, + 452, + 356 + ], + "lines": [ + { + "bbox": [ + 412, + 316, + 452, + 356 + ], + "spans": [ + { + "bbox": [ + 412, + 316, + 452, + 356 + ], + "type": "image", + "image_path": "c5e0d7af0302d321cbd6b955f02cfc38d5a140143f2dc9f51ef0c6eca8357b39.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 412, + 357, + 452, + 397 + ], + "blocks": [ + { + "bbox": [ + 412, + 357, + 452, + 397 + ], + "lines": [ + { + "bbox": [ + 412, + 357, + 452, + 397 + ], + "spans": [ + { + "bbox": [ + 412, + 357, + 452, + 397 + ], + "type": "image", + "image_path": "9b87cb02fc7aaf51264c2638a34a16efb881db4db11da08a3e01c91d8a369b9d.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 412, + 397, + 453, + 404 + ], + "lines": [ + { + "bbox": [ + 412, + 397, + 453, + 404 + ], + "spans": [ + { + "bbox": [ + 412, + 397, + 453, + 404 + ], + "type": "inline_equation", + "content": "\\mathrm{ISP}_{\\mathrm{ZRR}}(\\mathrm{F}^{\\mathrm{ZRR}}(\\mathrm{raw}_{S7}))" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "dataset, where raw images originated from a single camera. Then, we transferred all raw images from other cameras in the other datasets using our Rawformer before rendering with the trained LAN model. We report the PSNR and SSIM of the rendered sRGB images compared to the ground-truth sRGB images for each set as shown in Table 6. This includes the results of camera-specific LAN ISP models trained and tested on the same datasets. The camera-specific results represent the optimal results achievable by the neural-based ISP, which requires a paired raw-SRGB dataset for training on each camera model. As can be seen, our unsupervised raw-to-raw translation method demonstrates robust improvement, exhibiting only a marginal " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " reduction in accuracy compared to the camera-specific LAN model trained and tested on the same camera. This underscores the effectiveness of our method in addressing real-world cross-domain broadcasting challenges. See Fig. 5 for qualitative examples." + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 401, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 401, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 401, + 91, + 447, + 100 + ], + "type": "text", + "content": "Rawformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 400, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 400, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 400, + 129 + ], + "type": "text", + "content": "4.5 Comparison with Few-Shot Domain Adaptation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 135, + 482, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 135, + 482, + 255 + ], + "spans": [ + { + "bbox": [ + 130, + 135, + 482, + 255 + ], + "type": "text", + "content": "We further compared our Rawformer against Prabhakar's Few-Shot Domain Adaptation (PDA) [40]. Specifically, we conducted a 10-shot domain adaptation from the Samsung S7 ISP dataset's camera [44] to the Zurich raw-to-RGB dataset's camera [20], and vice-versa. The PDA model underwent initial training for 10 epochs using the Samsung S7 ISP dataset. Subsequently, domain adaptation was performed on 10 images from the Zurich raw-to-RGB dataset's camera over 4 epochs. This process resulted in performance metrics of 21.07 dB, and 17.81 dB in terms of PSNR and 0.80 and 0.70 for SSIM for our Rawformer, representing approximately " + }, + { + "bbox": [ + 130, + 135, + 482, + 255 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 130, + 135, + 482, + 255 + ], + "type": "text", + "content": " superior performance compared to PDA (see Table 7)." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 198, + 331, + 418, + 369 + ], + "blocks": [ + { + "bbox": [ + 130, + 275, + 482, + 321 + ], + "lines": [ + { + "bbox": [ + 130, + 275, + 482, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 482, + 321 + ], + "type": "text", + "content": "Table 7: Comparison with PDA [40]. S7 and ZRR denote the Samsung S7 ISP [44] and the Zurich Raw-to-RGB [20] datasets, respectively. Results are produced after translating each raw image to the target camera using PDA and our Rawformer, subsequently rendering with a LAN ISP [54] trained on the target camera's raw images." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 198, + 331, + 418, + 369 + ], + "lines": [ + { + "bbox": [ + 198, + 331, + 418, + 369 + ], + "spans": [ + { + "bbox": [ + 198, + 331, + 418, + 369 + ], + "type": "table", + "html": "
MethodS7-to-ZRRZRR-to-S7
PSNR↑SSIM↑MAE↓ (×1e-2)ΔE↓PSNR↑SSIM↑MAE↓ (×1e-2)ΔE↓
PDA [40]20.330.799.4113.9917.120.6918.0315.86
Ours22.080.808.1810.2719.020.7215.2711.74
", + "image_path": "61e4669a55fe4b8a6d24e9a7056ea702ab774124a2d414ce677427fd2b02392b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 407, + 326, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 326, + 420 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 326, + 420 + ], + "type": "text", + "content": "5 Conclusion and Future Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 432, + 482, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 432, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 482, + 612 + ], + "type": "text", + "content": "In this work, we presented a new approach, Rawformer, for raw-to-raw image translation using unpaired data. Rawformer sets a new standard for performance on real camera datasets, showcasing its superior capabilities in raw-to-raw translation when measured against other alternatives. The proposed method reduces the need for the extensive collection of raw-to-sRGB paired datasets to re-train learnable ISPs for new cameras, thereby decreasing associated costs. Our results indicate that existing raw-to-sRGB paired datasets can reliably train neural-based learnable ISPs, which can then be utilized to render raw images taken by unseen cameras, lacking such unpaired data. We believe that our approach provides valuable insights for learnable ISP, raw image translation research, and applications. Our method (without optimization) runs at 26 milliseconds per frame on GPU (" + }, + { + "bbox": [ + 130, + 432, + 482, + 612 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 130, + 432, + 482, + 612 + ], + "type": "text", + "content": "1 second on CPU), which may be impractical for real-time rendering in camera previews on devices with limited computational power. Future work will focus on introducing lighter models capable of real-time performance on CPU." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 629, + 241, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 241, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 241, + 643 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 653, + 477, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 653, + 477, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 653, + 477, + 665 + ], + "type": "text", + "content": "This work was partly supported by The Alexander von Humboldt Foundation." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 260, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 260, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 260, + 101 + ], + "type": "text", + "content": "G. Perevozhikov et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 159 + ], + "type": "text", + "content": "1. A Sharif, S., Naqvi, R.A., Biswas, M.: Beyond joint demosaicking and denoising: An image processing pipeline for a pixel-bin image sensor. In: CVPR (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 160, + 480, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 160, + 480, + 180 + ], + "spans": [ + { + "bbox": [ + 138, + 160, + 480, + 180 + ], + "type": "text", + "content": "2. Abdelhamed, A., Lin, S., Brown, M.S.: A high-quality denoising dataset for smartphone cameras. In: CVPR (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 181, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 181, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 138, + 181, + 480, + 213 + ], + "type": "text", + "content": "3. Affi, M., Abdelhamed, A., Abuolaim, A., Punnappurath, A., Brown, M.S.: CIE XYZ net: Unprocessing images for low-level computer vision tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(9), 4688-4700 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 214, + 476, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 214, + 476, + 224 + ], + "spans": [ + { + "bbox": [ + 138, + 214, + 476, + 224 + ], + "type": "text", + "content": "4. Afifi, M., Abuolaim, A.: Semi-supervised raw-to-raw mapping. In: BMVC (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 224, + 480, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 480, + 245 + ], + "type": "text", + "content": "5. Affi, M., Brown, M.S.: Sensor-independent illumination estimation for dnn models. In: BMVC (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 246, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 480, + 266 + ], + "type": "text", + "content": "6. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 267, + 443, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 443, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 443, + 277 + ], + "type": "text", + "content": "7. Barron, J.T., Tsai, Y.T.: Fast Fourier color constancy. In: CVPR (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 277, + 480, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 480, + 298 + ], + "type": "text", + "content": "8. Chen, C., Chen, Q., Xu, J., Koltun, V.: Learning to see in the dark. In: CVPR (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 299, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 480, + 331 + ], + "type": "text", + "content": "9. Cheng, D., Prasad, D.K., Brown, M.S.: Illuminant estimation for color constancy: why spatial-domain methods work and the role of the color distribution. JOSA A 31(5), 1049-1058 (2014)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 332, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 480, + 363 + ], + "type": "text", + "content": "0. Choi, Y., Choi, M., Kim, M., Ha, J.W., Kim, S., Choo, J.: StarGAN: Unified generative adversarial networks for multi-domain image-to-image translation. In: CVPR (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 364, + 480, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 364, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 138, + 364, + 480, + 384 + ], + "type": "text", + "content": "1. Conde, M.V., Vasluianu, F., Vazquez-Corral, J., Timofte, R.: Perceptual image enhancement for smartphone real-time applications. In: CVPR (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 385, + 480, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 385, + 480, + 406 + ], + "spans": [ + { + "bbox": [ + 138, + 385, + 480, + 406 + ], + "type": "text", + "content": "2. Dai, L., Liu, X., Li, C., Chen, J.: AWNet: Attentive wavelet network for image ISP. In: ECCV (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 407, + 480, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 480, + 427 + ], + "type": "text", + "content": "3. Delbracio, M., Kelly, D., Brown, M.S., Milanfar, P.: Mobile computational photography: A tour. Annual Review of Vision Science 7, 571-604 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 428, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 428, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 428, + 480, + 460 + ], + "type": "text", + "content": "4. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 460, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 480, + 482 + ], + "type": "text", + "content": "5. Finlayson, G.D., Zhu, Y.: Designing color filters that make cameras more colorimetric. IEEE Transactions on Image Processing 30, 853-867 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 483, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 483, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 138, + 483, + 480, + 514 + ], + "type": "text", + "content": "6. Hasinoff, S.W., Sharlet, D., Geiss, R., Adams, A., Barron, J.T., Kainz, F., Chen, J., Levoy, M.: Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM Transactions on Graphics 35(6), 1-12 (2016)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 514, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 480, + 546 + ], + "type": "text", + "content": "7. He, X., Hu, T., Wang, G., Wang, Z., Wang, R., Zhang, Q., Yan, K., Chen, Z., Li, R., Xie, C., et al.: Enhancing RAW-to-sRGB with decoupled style structure in Fourier domain. In: AAAI (2024)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 547, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 480, + 568 + ], + "type": "text", + "content": "8. Herrmann, C., Bowen, R.S., Wadhwa, N., Garg, R., He, Q., Barron, J.T., Zabih, R.: Learning to autofocus. In: CVPR (2020)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 569, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 569, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 569, + 480, + 600 + ], + "type": "text", + "content": "9. Ignatov, A., Chiang, C.M., Kuo, H.K., Sycheva, A., Timofte, R.: Learned smartphone ISP on mobile NPUs with deep learning, mobile AI 2021 challenge: Report. In: CVPRW (2021)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "text", + "content": "20. Ignatov, A., Timofte, R., Ko, S.J., Kim, S.W., Uhm, K.H., Ji, S.W., Cho, S.J., Hong, J.P., Mei, K., Li, J., et al.: AIM 2019 challenge on raw to rgb mapping: Methods and results. In: ICCVW (2019)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "21. Ignatov, A., Timofte, R., Zhang, Z., Liu, M., Wang, H., Zuo, W., Zhang, J., Zhang, R., Peng, Z., Ren, S., et al.: AIM 2020 challenge on learned image signal processing pipeline. In: ECCV (2020)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 101 + ], + "type": "text", + "content": "Rawformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "22. Ignatov, A., Van Gool, L., Timofte, R.: Replacing mobile camera ISP with a single deep learning model. In: CVPRW (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 161 + ], + "type": "text", + "content": "23. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: CVPR (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 162, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 162, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 480, + 183 + ], + "type": "text", + "content": "24. Jeong, W., Jung, S.W.: RAWtoBit: A fully end-to-end camera isp network. In: ECCV (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 184, + 480, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 480, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 480, + 206 + ], + "type": "text", + "content": "25. Jiang, Y., Wronski, B., Mildenhall, B., Barron, J.T., Wang, Z., Xue, T.: Fast and high quality image denoising via malleable convolution. In: ECCV (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 207, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 480, + 228 + ], + "type": "text", + "content": "26. Karras, T., Aila, T., Laine, S., Lehtinen, J.: Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 229, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 480, + 251 + ], + "type": "text", + "content": "27. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: CVPR (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 252, + 480, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 252, + 480, + 284 + ], + "spans": [ + { + "bbox": [ + 132, + 252, + 480, + 284 + ], + "type": "text", + "content": "28. Kim, J., Kim, M., Kang, H., Lee, K.: U-GAT-IT: Unsupervised generative attentional networks with adaptive layer-instance normalization for image-to-image translation. arXiv preprint arXiv:1907.10830 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 285, + 480, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 480, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 480, + 307 + ], + "type": "text", + "content": "29. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 308, + 480, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 480, + 341 + ], + "type": "text", + "content": "30. Ledig, C., Theis, L., Huszar, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., et al.: Photo-realistic single image superresolution using a generative adversarial network. In: CVPR (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 341, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 480, + 373 + ], + "type": "text", + "content": "31. Liang, Z., Cai, J., Cao, Z., Zhang, L.: Cameranet: A two-stage framework for effective camera isp learning. IEEE Transactions on Image Processing 30, 2248-2262 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 374, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 396 + ], + "type": "text", + "content": "32. Liu, M.Y., Breuel, T., Kautz, J.: Unsupervised image-to-image translation networks. In: NeurIPS (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "type": "text", + "content": "33. Loshchilov, I., Hutter, F.: SGDR: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 419, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 480, + 441 + ], + "type": "text", + "content": "34. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 441, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 480, + 464 + ], + "type": "text", + "content": "35. Maas, A.L., Hannun, A.Y., Ng, A.Y., et al.: Rectifier nonlinearities improve neural network acoustic models. In: ICML (2013)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 464, + 480, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 480, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 480, + 486 + ], + "type": "text", + "content": "36. Menon, D., Andriani, S., Calvagno, G.: Demosaicing with directional filtering and a posteriori decision. IEEE Transactions on Image Processing 16(1), 132-141 (2006)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 487, + 480, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 480, + 508 + ], + "type": "text", + "content": "37. Nguyen, R., Prasad, D.K., Brown, M.S.: Raw-to-raw: Mapping between image sensor color responses. In: CVPR (2014)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 509, + 480, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 480, + 531 + ], + "type": "text", + "content": "38. Pang, Y., Lin, J., Qin, T., Chen, Z.: Image-to-image translation: Methods and applications. IEEE Transactions on Multimedia 24, 3859-3881 (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 532, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 480, + 553 + ], + "type": "text", + "content": "39. Park, T., Efros, A.A., Zhang, R., Zhu, J.Y.: Contrastive learning for unpaired image-to-image translation. In: ECCV (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 554, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 480, + 576 + ], + "type": "text", + "content": "40. Prabhakar, K.R., Vinod, V., Sahoo, N.R., Babu, R.V.: Few-shot domain adaptation for low light raw image enhancement. arXiv preprint arXiv:2303.15528 (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 577, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 480, + 598 + ], + "type": "text", + "content": "41. Punnappurath, A., Abuolaim, A., Abdelhamed, A., Levinshtein, A., Brown, M.S.: Day-to-night image synthesis for training nighttime neural isps. In: CVPR (2022)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "text", + "content": "42. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: International Conference on Medical image computing and computer-assisted intervention (2015)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "43. Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen, X.: Improved techniques for training gans. Advances in neural information processing systems 29 (2016)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 260, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 260, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 260, + 100 + ], + "type": "text", + "content": "G. Perevozchikov et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 657 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "44. Schwartz, E., Giryes, R., Bronstein, A.M.: DeepISP: Toward learning an end-to-end image processing pipeline. IEEE Transactions on Image Processing 28(2), 912-923 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "text", + "content": "45. Seo, D., Punnappurath, A., Zhao, L., Abdelhamed, A., Tedla, S.K., Park, S., Choe, J., Brown, M.S.: Graphics2RAW: Mapping computer graphics images to sensor raw images. In: ICCV (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "text", + "content": "46. Sharma, G., Wu, W., Dalal, E.N.: The CIEDE2000 color-difference formula: Implementation notes, supplementary test data, and mathematical observations. Color Research & Application 30(1), 21-30 (2005)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 214, + 482, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 482, + 247 + ], + "type": "text", + "content": "47. Sindelář, O., Sroubek, F.: Image deblurring in smartphone devices using built-in inertial measurement sensors. Journal of Electronic Imaging 22(1), 011003-011003 (2013)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 247, + 482, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 269 + ], + "type": "text", + "content": "48. Souza, M., Heidrich, W.: Crispnet: Color rendition isp net. arXiv preprint arXiv:2203.10562 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 269, + 482, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 482, + 291 + ], + "type": "text", + "content": "49. Tominaga, S., Nishi, S., Ohtera, R.: Measurement and estimation of spectral sensitivity functions for mobile phone cameras. Sensors 21(15), 4985 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 290, + 482, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 323 + ], + "type": "text", + "content": "50. Torbunov, D., Huang, Y., Yu, H., Huang, J., Yoo, S., Lin, M., Viren, B., Ren, Y.: UVCGAN: UNet vision transformer cycle-consistent GAN for unpaired image-to-image translation. In: WACV (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 323, + 482, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 355 + ], + "type": "text", + "content": "51. Torbunov, D., Huang, Y., Yu, H., Huang, J., Yoo, S., Lin, M., Viren, B., Ren, Y.: UVCGAN v2: An improved cycle-consistent GAN for unpaired image-to-image translation. arXiv preprint arXiv:2303.16280 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 355, + 482, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 355, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 355, + 482, + 376 + ], + "type": "text", + "content": "52. Truong, P., Danelljan, M., Van Gool, L., Timofte, R.: Learning accurate dense correspondences and when to trust them. In: CVPR (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 376, + 482, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 409 + ], + "type": "text", + "content": "53. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing 13(4), 600-612 (2004)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 409, + 482, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 409, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 482, + 441 + ], + "type": "text", + "content": "54. Wirzberger Raimundo, D., Ignatov, A., Timofte, R.: LAN: Lightweight attention-based network for raw-to-rgb smartphone image processing. In: CVPRW. pp. 807-815 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 441, + 482, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 482, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 482, + 474 + ], + "type": "text", + "content": "55. Wronski, B., Garcia-Dorado, I., Ernst, M., Kelly, D., Krainin, M., Liang, C.K., Levoy, M., Milanfar, P.: Handheld multi-frame super-resolution. ACM Transactions on Graphics (ToG) 38(4), 1-18 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 474, + 482, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 484 + ], + "type": "text", + "content": "56. Xing, Y., Qian, Z., Chen, Q.: Invertible image signal processing. In: CVPR (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 484, + 482, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 482, + 506 + ], + "type": "text", + "content": "57. Yi, Z., Zhang, H., Tan, P., Gong, M.: DualGAN: Unsupervised dual learning for image-to-image translation. In: ICCV. pp. 2849-2857 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 506, + 482, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 548 + ], + "type": "text", + "content": "58. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5728-5739 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 548, + 482, + 571 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 482, + 571 + ], + "type": "text", + "content": "59. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: CycleISP: Real image restoration via improved data synthesis. In: CVPR (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 571, + 482, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 571, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 571, + 482, + 592 + ], + "type": "text", + "content": "60. Zhang, Z., Wang, H., Liu, M., Wang, R., Zhang, J., Zuo, W.: Learning raw-to-srgb mappings with inaccurately aligned supervision. In: ICCV (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 592, + 482, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 592, + 482, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 592, + 482, + 613 + ], + "type": "text", + "content": "61. Zhao, Y., Wu, R., Dong, H.: Unpaired image-to-image translation using adversarial consistency loss. In: ECCV (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 613, + 482, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 613, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 613, + 482, + 635 + ], + "type": "text", + "content": "62. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: ICCV (2017)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 635, + 482, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 635, + 482, + 657 + ], + "spans": [ + { + "bbox": [ + 130, + 635, + 482, + 657 + ], + "type": "text", + "content": "63. Zhu, P., Abdal, R., Qin, Y., Wonka, P.: SEAN: Image synthesis with semantic region-adaptive normalization. In: CVPR (2020)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "Rawformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Ray Denoising_ Depth-aware Hard Negative Sampling for Multi-view 3D Object Detection/ebbe937b-35fd-41db-b04a-6c0c4e0d90c8_content_list.json b/2024/Ray Denoising_ Depth-aware Hard Negative Sampling for Multi-view 3D Object Detection/ebbe937b-35fd-41db-b04a-6c0c4e0d90c8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7cee94a820fd600c486905f56429ffdb0ebdf569 --- /dev/null +++ b/2024/Ray Denoising_ Depth-aware Hard Negative Sampling for Multi-view 3D Object Detection/ebbe937b-35fd-41db-b04a-6c0c4e0d90c8_content_list.json @@ -0,0 +1,1632 @@ +[ + { + "type": "text", + "text": "Ray Denoising: Depth-aware Hard Negative Sampling for Multi-view 3D Object Detection", + "text_level": 1, + "bbox": [ + 238, + 140, + 763, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Feng Liu1* Tengteng Huang2 Qianjing Zhang2 Haotian Yao2 Chi Zhang2 Fang Wan1 Qixiang Ye1 Yanzhao Zhou1†", + "bbox": [ + 215, + 212, + 785, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of Chinese Academy of Sciences liufeng20@mails.ucas.ac.cn {wanfang,qxye,zhouyanzhao}@ucas.ac.cn", + "bbox": [ + 238, + 253, + 759, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Mach Drive {tengteng.huang,qianjing.zhang,haotian.yao,chi.zhang}@mach-drive.com", + "bbox": [ + 238, + 282, + 763, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Multi-view 3D object detection systems often struggle with generating precise predictions due to the challenges in estimating depth from images, increasing redundant and incorrect detections. Our paper presents Ray Denoising, an innovative method that enhances detection accuracy by strategically sampling along camera rays to construct hard negative examples. These examples, visually challenging to differentiate from true positives, compel the model to learn depth-aware features, thereby improving its capacity to distinguish between true and false positives. Ray Denoising is designed as a plug-and-play module, compatible with any DETR-style multi-view 3D detectors, and it only minimally increases training computational costs without affecting inference speed. Our comprehensive experiments, including detailed ablation studies, consistently demonstrate that Ray Denoising outperforms strong baselines across multiple datasets. It achieves a $1.9\\%$ improvement in mean Average Precision (mAP) over the state-of-the-art StreamPETR method on the NuScenes dataset. It shows significant performance gains on the Argoverse 2 dataset, highlighting its generalization capability. The code is available at https://github.com/LiewFeng/RayDN.", + "bbox": [ + 259, + 353, + 738, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Multi-view 3D Object Detection $\\cdot$ Depth-aware Hard Negative Sampling $\\cdot$ Ray Denoising", + "bbox": [ + 259, + 616, + 738, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 672, + 375, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D object detection is a crucial component in autonomous driving systems, drawing considerable interest from the computer vision community. The field of image-based 3D object detection [9,10,13,14,37,38] is experiencing a surge in research due to its cost-effectiveness compared to LiDAR-based solutions. A key challenge in multi-view 3D object detection, which relies on images from surrounding cameras, is the difficulty in estimating depth from images, leading to duplicate predictions, as shown in Figure 1.", + "bbox": [ + 212, + 708, + 787, + 813 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Work was done during internship at Mach Drive. † Corresponding Author.", + "bbox": [ + 217, + 824, + 730, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a10894823f689f66f34a796554925a1b51b5b203014ba9f4f1e5bccbbcaa396b.jpg", + "image_caption": [ + "Fig. 1: The challenge of estimating depth from images in multi-view 3D object detection leads to duplicate predictions and false positive detections along camera rays. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 330, + 143, + 674, + 313 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the methodological improvements, multi-view 3D object detectors struggle to reduce false positive predictions arising from depth ambiguities. Several recent studies [8, 15, 17, 19, 22, 26, 34, 41, 47] have attempted to tackle this issue by incorporating temporal information. However, these methods do not explicitly account for the 3D structure of the scene, which limits their potential for further enhancement.", + "bbox": [ + 212, + 398, + 787, + 488 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, previous works have explored applying general techniques such as Non-Maximum Suppression (NMS) and Focal Loss to mitigate duplicate predictions. NMS, a post-processing technique, targets false positive predictions with high Intersection over Union (IoU) but is less effective when these predictions are scattered along rays with low IoU. Focal Loss, a loss function designed to reduce high-confidence false positive predictions, has also been implemented. However, it has been observed that multi-view 3D object detectors using Focal Loss still face challenges in effectively resolving the problem of false positive predictions along the same rays.", + "bbox": [ + 212, + 489, + 787, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our quantitative analysis highlighted the importance of addressing false positive predictions along the same ray as Ground Truth. By utilizing the precise positional data of Ground Truth objects, we could identify and eliminate these redundant predictions within the state-of-the-art StreamPETR method [34]. This process significantly enhanced $5.4\\%$ in mean Average Precision (mAP), emphasizing the critical need for models to improve their depth estimation capabilities. The substantial improvement underscores the potential of refining depth estimation to suppress these predictions and enhance overall detection performance.", + "bbox": [ + 212, + 627, + 787, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our key observation is that false positives often occur along camera rays due to the inherent limitation in conventional multi-view object detectors. Since the depth information for each pixel is not accurately estimated, the position embedding can only encode the ray direction. As a result, queries on the same ray will consistently interact with identical visual features from the image, leading to numerous duplicate predictions (false positives) along that ray. This scenario", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "F. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "underscores the model's need to learn depth-aware features that can discern objects in depth despite the visual features being the same for objects along the same ray. We propose a novel method called Ray Denoising (i.e., RayDN). This framework is inherently flexible and does not limit the choice of distribution for sampling depth-aware hard negative samples. Based on our ablation studies, we have chosen the Beta distribution for its effectiveness in capturing the spatial distribution of false positives that models are likely to generate. This choice enables Ray Denoising to create depth-aware hard negative samples used for denoising, thereby enhancing the model's ability to learn more robust features and representations for distinguishing false positives along the ray, as depicted in Figure 2. Ray Denoising introduces only a marginal increase in computational costs during the training stage without affecting inference speed.", + "bbox": [ + 212, + 146, + 787, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To summarize our contributions, we highlight the following key points. Firstly, we have identified the persistent challenge of false positive predictions along the same ray, which acts as a bottleneck in the performance of multi-view 3D object detectors. Secondly, we introduce Ray Denoising, a novel denoising method that utilizes the Beta distribution to create depth-aware hard negative samples along rays. This method explicitly takes into account the 3D structure of the scene, offering a flexible solution compatible with any DETR-style multi-view 3D detector to address the issue of duplicate predictions along rays. Lastly, our method achieves state-of-the-art results on the NuScenes dataset [2], significantly enhancing the performance of multi-view 3D object detectors. Specifically, we have improved upon the current state-of-the-art method, StreamPETR, by $1.9\\%$ in mean Average Precision (mAP), thereby demonstrating the effectiveness of Ray Denoising.", + "bbox": [ + 212, + 329, + 787, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 545, + 387, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Image-based 3D Object Detection", + "text_level": 1, + "bbox": [ + 215, + 574, + 542, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D object detection is a cornerstone task in autonomous driving systems. The domain has witnessed notable advancements in monocular 3D object detection, largely attributed to the KITTI benchmark [5]. This progress has been driven by a range of research studies [1,24,28,31,35,44]. Nevertheless, monocular systems face limitations due to their dependence on a single viewpoint and limited data, which hampers their capacity to manage complex scenarios. To overcome these limitations, extensive benchmarks [2,32,39] have been developed, offering multi-viewpoint data that enriches the field of multi-view 3D object detection and propels the evolution of advanced detection methodologies.", + "bbox": [ + 212, + 598, + 787, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The multi-view 3D object detection research is primarily bifurcated into dense Bird's Eye View (BEV)-based and sparse query-based algorithms. Our work is categorized under the sparse query-based algorithms. Dense BEV-based algorithms convert multi-view image features into a dense BEV representation using the Lift-Splat-Shoot (LSS) technique [27]. To mitigate overfitting from LSS, BEVDet [9] recommends data augmentation on BEV features. BEVDepth [14] introduces depth estimation supervision to enhance LSS precision. Compared to", + "bbox": [ + 212, + 734, + 800, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Ray Denoising", + "bbox": [ + 632, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/de98f7c221cf5f88b61b2009104a629912291a786a023b3f94d3030168fa9010.jpg", + "image_caption": [ + "Fig. 2: The proposed Ray Denoising approach (right) effectively reduces false positive detections along the ray (highlighted by red rectangles) in the previous state-of-the-art method StreamPETR [34] (left). Best viewed by zooming on the screen." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 300 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LSS-like methods, BEVFormer-like and DETR-like methods can leverage the strong capabilities of transformers for spatial-temporal modeling. BEVFormer-like methods use dense BEV features to enable multiple tasks (e.g., detection and segmentation) but also introduce high computational costs. Sparse query-based algorithms such as DETR3D [37] and PETR [21] utilize learnable 3D object queries and position encoding to engage with multi-view 2D features. DAT [43] introduces a pipeline to construct depth-aware position embeddings and performs contrastive learning on BEV features to tackle duplicate predictions. Recent studies [8, 13, 15, 17, 19, 22, 26, 34] have integrated temporal information to refine these methods further.", + "bbox": [ + 212, + 382, + 787, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Hard Negative Samples Mining", + "text_level": 1, + "bbox": [ + 215, + 556, + 522, + 573 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hard negative sample mining is a strategy to minimize false positive predictions in object detection. Early works like R-CNN [6] employed standard hard negative mining methods [4, 33], where misclassified samples were utilized for retraining. OHEM [29] and RetinaNet [16] introduced adaptive sampling and focal loss to address the challenges of heuristics and hyper-parameters in hard negative sample selection. These methods select hard negative samples from existing predictions based on confidence scores and matched ground-truth object classes. Our approach takes a different path by constructing new hard negative samples based on the locations of ground-truth objects rather than selecting from existing predictions or depending on prediction confidence and matched ground-truth object classes.", + "bbox": [ + 212, + 580, + 787, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Denoising in Object Detection", + "text_level": 1, + "bbox": [ + 215, + 768, + 514, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In 2D object detection, models such as DETR [3] and its variants [20,36,46] often face convergence issues, which are partly due to the instability of bipartite graph matching and the inconsistency in optimization goals during the early stages of", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "F. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "training. To tackle this, DN-DETR [12] introduces the concept of feeding noisy ground-truth bounding boxes into the Transformer decoder, with the model being trained to reconstruct the original, clean boxes. This process, known as denoising, involves using queries initialized with these noisy ground-truth bounding boxes, referred to as denoising queries. DINO [42] further refines this approach by generating both positive and negative denoising queries for each ground-truth bounding box, with the negative queries incorporating more noise to improve the model's ability to reject incorrect predictions.", + "bbox": [ + 212, + 146, + 782, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the realm of 3D object detection, DETR-style methods [19,22,34] perpetuate the denoising approach by creating a denoising query for each ground-truth bounding box. Based on the noise level, these queries are classified as either an object or 'no object'. However, these methods neglect to incorporate knowledge of the 3D scene structure when generating denoising queries. In multi-view 3D object detection, the absence of depth information for each pixel results in models producing duplicate predictions along the same ray with different depths. To combat these false positive predictions, we harness the 3D structure knowledge of the scene. We devise multiple negative ray-denoising queries distributed along the rays, spatially distinct from the traditional negative denoising queries surrounding the ground-truth objects. This innovative design markedly improves the detector's capability to differentiate between true positive predictions and false positive predictions along the same ray.", + "bbox": [ + 212, + 267, + 784, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 487, + 380, + 505 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section elaborates on our proposed method, Ray Denoising, for multi-view 3D object detection. We start with a comprehensive overview of the DETR-style multi-view 3D object detector framework in Section 3.1. Following that, we describe the three pivotal steps for integrating Ray Denoising into the framework to address the challenge of duplicate predictions. Section 3.2 outlines the ray generation process. Section 3.3 explains utilizing the Beta distribution family for sampling reference points. Section 3.4 details the construction of spatial denoising queries, referred to as Ray Queries, along each ray. Lastly, Section 3.5 delves into the essential differences between our Ray Queries and the denoising queries employed in prior works and the influence of Ray Denoising on the model training process.", + "bbox": [ + 212, + 518, + 787, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Overview", + "text_level": 1, + "bbox": [ + 215, + 709, + 339, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ray Denoising is designed to be integrated into any DETR-style multi-view 3D object detector [15, 17, 19, 21, 22, 34, 37]. These detectors typically consist of a convolutional network-based feature encoder and a transformer-based decoder. The process begins by feeding $N$ surround-view images $\\pmb{I} = \\{I_i\\in \\mathbb{R}^{3\\times H_I\\times W_I},i = 1,2,\\dots,N\\}$ into the feature encoder (e.g., ResNet101 [7]) to extract image features $\\pmb{F} = \\{F_{i}\\in \\mathbb{R}^{C\\times H_{F}\\times W_{F}},i = 1,2,\\dots,N\\}$ . Here, $H_{I}$ and $W_{I}$ denote the image dimensions, $H_{F}$ and $W_{F}$ are the feature dimensions, and $C$ represents the", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Ray Denoising", + "bbox": [ + 632, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/88714622b6b3ea899694bc720348b5a8c1f9b783061227f1d00f4ecae45bf6ba.jpg", + "image_caption": [ + "Fig. 3: Overall framework of the Ray Denoising approach, a plug-and-play training technique for DETR-style multi-view 3D object detectors, focuses on refining the model's ability to distinguish true positives from false positives in depth. Casting rays and sampling depth-aware denoising queries effectively tackle the challenge of false positives arising from the inherent difficulties in visually estimating depth, leading to substantial improvements in detection performance over strong baselines. Best viewed in color and by zooming on the screen." + ], + "image_footnote": [], + "bbox": [ + 228, + 148, + 777, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "number of channels in the feature. To facilitate 3D perception, multiple points within each camera's frustum are transformed and encoded into position embeddings [21]. These embeddings enable the multi-view image features to interact with 3D queries.", + "bbox": [ + 212, + 473, + 787, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The final step involves $N$ object queries $\\mathbf{Q} \\in \\mathbb{R}^{N \\times 256}$ , which are derived from a set of learnable 3D reference points $\\mathbf{P} \\in \\mathbb{R}^{N \\times 3}$ . These queries interact with the multi-view image features $\\mathbf{F}$ in the transformer decoder, employing multi-layer cross-attention to identify objects. The decoder's output features are then processed by prediction heads (multi-layer perceptron, i.e., MLP) to yield classification scores $(cls)$ , position offsets $(x, y, z)$ , scales $(w, h, l)$ , directions $(\\theta_x, \\theta_y)$ , and velocities $(v_x, v_y)$ .", + "bbox": [ + 212, + 534, + 787, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Ray Casting", + "text_level": 1, + "bbox": [ + 215, + 665, + 366, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The reference points for our ray denoising queries are distributed along the camera ray, which extends from the camera's optical center to the ground-truth object on the image plane. To establish this ray, we project the 3D center of the ground-truth object into the camera frustum space using the following transformation:", + "bbox": [ + 212, + 693, + 787, + 767 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} ^ {\\prime} = \\mathbf {K} \\cdot \\mathbf {C} _ {G T}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 444, + 768, + 785, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{K}$ is the $4\\times 4$ transformation matrix that maps points from 3D world space to camera frustum space. $\\mathbf{C}_{GT} = (x,y,z,1)$ is the center of the groundtruth object in 3D space, and $\\mathbf{C}' = (u\\times d,v\\times d,d,1)$ is the corresponding", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "F. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4c4a2d14581946113ef63f596cdc1d87267b0add529ac47c993cac2426c793e4.jpg", + "image_caption": [ + "(a) Distribution Comparison" + ], + "image_footnote": [], + "bbox": [ + 246, + 147, + 496, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/22781dfede13b14d9e9f299c4676acdea843872bb5f1c03fda5ed7ddba61e0fe.jpg", + "image_caption": [ + "(b) Beta distribution family", + "Fig. 4: (a) Distribution comparison showing that the Beta distribution is bounded between -1 and 1, unlike the Laplace and Gaussian distributions, which are unbounded. (b) The Beta distribution family, with the x-range adjusted from [0, 1] to $[-1, 1]$ using the transformation $y = 2x - 1$ . Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 504, + 145, + 756, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "projected center in the camera frustum, with $(u,v)$ being the pixel coordinates and $d$ the depth value.", + "bbox": [ + 212, + 404, + 784, + 433 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With the depth of the ground-truth object's center known, we can determine the coordinates of valid reference points along the ray as follows:", + "bbox": [ + 212, + 434, + 784, + 465 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {d} _ {i} = d + \\beta_ {i} \\cdot f \\left(\\mathbf {S} _ {G T}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 423, + 476, + 784, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $f$ is a function that encodes the average scale of the ground-truth object, calculated as $f(\\mathbf{S}_{GT}) = k \\cdot \\frac{w + h + l}{6}$ . The radius $k$ defines the valid distribution range for reference points, and $\\beta_{i}$ is the offset for the $i$ -th reference point.", + "bbox": [ + 212, + 506, + 782, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The position of the reference point is then obtained by:", + "bbox": [ + 238, + 551, + 640, + 566 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {P}} _ {i} = \\mathbf {K} ^ {- 1} \\cdot \\hat {\\mathbf {C}} _ {i} ^ {\\prime}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 441, + 577, + 784, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{C}}_i^{\\prime} = (u\\times \\hat{d}_i,v\\times \\hat{d}_i,\\hat{d}_i,1)$", + "bbox": [ + 214, + 607, + 459, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Sample Generation", + "text_level": 1, + "bbox": [ + 214, + 647, + 419, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With the valid reference points along the rays defined, we proceed to sample $N$ reference points to emulate the distribution of false positive predictions that arise from depth ambiguities. A straightforward uniform distribution has proven effective, yielding a $2.2\\%$ mAP improvement as depicted in Figure 4a. To more accurately simulate the distribution of false positives, we consider two key aspects. First, we assess whether models tend to predict objects as being closer or further away from the ego-vehicle relative to the actual positions of the ground-truth objects. Second, we examine whether models are more likely to predict objects near or far from the center of the ground-truth objects.", + "bbox": [ + 212, + 672, + 785, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Distributions like the Beta, Gaussian, and Laplace are suitable for handling complex scenarios, with the Gaussian and Laplace being unbounded. However,", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Ray Denoising", + "bbox": [ + 632, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "samples with significant offsets from the ground truth can result in large regression losses, negatively impacting optimization. As shown in Figure 4a, the Beta distribution slightly outperforms the others. Notably, the uniform distribution is a special case within the Beta family. For a versatile sampling strategy applicable to all scenarios, we select the Beta distribution family, characterized by the probability density function:", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\np (x | \\lambda , \\mu) = \\frac {\\Gamma (\\lambda + \\mu)}{\\Gamma (\\lambda) + \\Gamma (\\mu)} x ^ {\\lambda - 1} (1 - x) ^ {\\mu - 1}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 247, + 785, + 280 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\lambda$ and $\\mu$ are hyper-parameters that shape the distribution. The Gamma function, denoted as $\\Gamma(x)$ , is defined as:", + "bbox": [ + 214, + 287, + 787, + 319 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\Gamma (x) = \\int_ {0} ^ {\\infty} t ^ {x - 1} e ^ {- t} d t. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 419, + 329, + 785, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The beta distribution's probability density function (PDF), visualized in Figure 4b, reveals its behavior under different parameter settings. When $\\lambda$ equals $\\mu$ , the distribution is symmetric, reflecting a scenario where the model predicts objects with equal likelihood at closer or further distances. For instance, with $\\lambda = \\mu = 1$ , the distribution simplifies to a uniform distribution. Adjusting $\\lambda$ and $\\mu$ to be higher shifts the reference points of the denoising queries closer to the ground-truth object's center and vice versa.", + "bbox": [ + 212, + 369, + 787, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When $\\lambda$ exceeds $\\mu$ , such as in the case of $\\lambda = 8$ and $\\mu = 2$ , the distribution suggests that models are more likely to predict objects further away from the ego-vehicle. Conversely, when $\\lambda$ is less than $\\mu$ , for example, $\\lambda = 2$ and $\\mu = 8$ , the distribution indicates a propensity for models to predict objects closer to the ego-vehicle. This flexibility in the Beta distribution allows for tailored sampling strategies that can adapt to the specific challenges of depth estimation in multiview 3D object detection.", + "bbox": [ + 212, + 477, + 789, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Query Denoising", + "text_level": 1, + "bbox": [ + 214, + 604, + 401, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We construct ray denoising queries from the sampled $N$ reference points. Specifically, a multi-layer perceptron (MLP) is utilized to project the normalized 3D coordinates of each point into a latent feature space. These $N$ ray denoising queries $\\pmb{q}_r$ are combined with the learnable object queries $\\pmb{q}_o$ from the baseline detector and are input into the transformer decoder. An attention mask is employed during the self-attention operations to prevent information leakage from the ray denoising queries to the object queries. The query closest to the center of the ground truth (GT) object is designated as the positive query, inheriting the class label of the corresponding GT object. The remaining $N - 1$ queries are labeled as background, i.e., 'no object'. The loss for ray denoising queries adheres to the same criteria as for the learnable object queries, using Focal Loss for classification and L1 Loss for regression. This process ensures that the model is effectively trained to discern between true positives and false positives along the same ray.", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "F. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5 Discussion", + "text_level": 1, + "bbox": [ + 215, + 146, + 349, + 159 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our Ray Denoising approach is based on the pivotal observation that image-based 3D detection systems often struggle to distinguish true positives from false positives along camera rays. DETR-style multi-view 3D object detectors implicitly learn depth estimation from ground truth supervision. However, the randomly distributed reference points of learnable queries do not fully leverage the available ground truth information. While these reference points are updated during training, they fail to provide adequate hard negative samples for each object in every scene. To enhance the utilization of ground truth information, traditional denoising techniques introduce additional reference points distributed uniformly around the ground-truth object during training. These instance-specific reference points have improved detection performance [19,22,34]. However, they overlook the depth ambiguities intrinsic to multi-view 3D object detection. The absence of precise depth information for each pixel means that the position embedding can only encode ray direction, not depth. This leads to queries on the same ray interacting with the same image features, resulting in redundant predictions. Ray Denoising diverges from traditional techniques by strategically generating reference points along rays that extend from cameras to objects. This approach explicitly considers the 3D structure of each object in the scene, providing a sufficient number of hard negative samples. During training, these Ray Queries interact within the self-attention layer, effectively guiding the model to suppress depth-ambiguous spatial hard negative samples. This interaction enhances the detector's ability to differentiate between true positive predictions (objects) and false positive predictions (duplicate detections), improving detection accuracy.", + "bbox": [ + 212, + 167, + 787, + 531 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 215, + 549, + 366, + 566 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Dataset and Metrics", + "text_level": 1, + "bbox": [ + 215, + 577, + 431, + 590 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our model's performance was assessed using two datasets: nuScenes [2] and Argovverse 2 [39]. The nuScenes dataset include 1000 video sequences, which are split into training (700 videos), validation (150 videos), and testing (150 videos) sets, each approximately 20 seconds long with annotations at 0.5-second intervals. The dataset comprises 1.4 million annotated 3D bounding boxes across ten object classes. Evaluation metrics include the mean Average Precision (mAP) and five true positive metrics: ATE, ASE, AOE, AVE, and AAE, which assess translation, scale, orientation, velocity, and attribute errors. The nuScenes Detection Score (NDS) is a comprehensive score derived from these metrics, offering an overall performance evaluation.", + "bbox": [ + 212, + 598, + 787, + 748 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The Argoverse 2 dataset contains 1000 unique scenes, each 15 seconds long, annotated at $10\\mathrm{Hz}$ . The scenes are divided into 700 for training, 150 for validation, and 150 for testing. The evaluation covers 26 categories within a 150-meter range, addressing long-range perception tasks. Metrics include the mAP and the Composite Detection Score (CDS), which integrates three other true positive metrics: ATE, ASE, and AOE.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Ray Denoising", + "bbox": [ + 632, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/5b9fcccbb114ab0b4619f299cd2e0cb4e72dc8950a09b46d5f4bec9f954dae4f.jpg", + "table_caption": [ + "Table 1: Comparison on the nuScenes validation set. $\\dagger$ Indicates methods that benefit from perspective-view pre-training." + ], + "table_footnote": [], + "table_body": "
MethodsBackboneImage SizemAPNDSmATEmASEmAOEmAVEmAAE
BevDet4D [8]ResNet50256×70432.245.70.7030.2780.4950.3540.206
PETRv2 [22]ResNet50256×70434.945.60.7000.2750.5800.4370.187
BEVDepth [14]ResNet50256×70435.147.50.6290.2670.4790.4280.198
BEVStereo [13]ResNet50256×70437.250.00.5980.2700.4380.3670.190
BEVFomerv2 [41]†ResNet50-42.352.90.6180.2730.4130.3330.188
SOLOFusion [26]ResNet50256×70442.753.40.5670.2740.5110.2520.181
SparseBEV [19]†ResNet50256×70444.855.80.5810.2710.3730.2470.190
StreamPETR [34]†ResNet50256×70445.055.00.6130.2670.4130.2650.198
RayDN† (Ours)ResNet50256×70446.956.30.5790.2640.4330.2560.187
BEVDepth [14]ResNet101512×140841.253.50.5650.2660.3580.3310.190
PETRv2 [22]†ResNet101640×160042.152.40.6810.2670.3570.3770.186
Sparse4D [17]†ResNet101900×160043.654.10.6330.2790.3630.3170.177
SOLOFusion [26]ResNet101512×140848.358.20.5030.2640.3810.2460.207
SparseBEV [19]†ResNet101512×140850.159.20.5620.2650.3210.2430.195
StreamPETR [34]†ResNet101512×140850.459.20.5690.2620.3150.2570.199
RayDN† (Ours)ResNet101512×140851.860.40.5410.2600.3150.2360.200
", + "bbox": [ + 215, + 184, + 789, + 377 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/ad5d40f5996ce1df058c171a188420a9c4cf8d68124fa83017c8e7db894cf6b1.jpg", + "table_caption": [ + "Table 2: Comparison on the nuScenes test set." + ], + "table_footnote": [], + "table_body": "
MethodsBackboneImage SizemAPNDSmATEmASEmAOEmAVEmAAE
DETR3D [37]V2-99900×160041.247.90.6410.2550.3940.8450.133
MV2D [38]V2-99640×160046.351.40.5420.2470.4030.8570.127
BEVFormer [15]V2-99900×160048.156.90.5820.2560.3750.3780.126
PETRv2 [22]V2-99640×160049.058.20.5610.2430.3610.3430.120
PolarFormer [10]V2-99900×160049.357.20.5560.2560.3640.4390.127
BEVStereo [13]V2-99900×160052.561.00.4310.2460.3580.3570.138
HoP [47]V2-99640×160052.861.20.4910.2420.3320.3430.109
SparseBEV [19]V2-99640×160054.362.70.5020.2440.3240.2510.126
StreamPETR [34]V2-99640×160055.063.60.4790.2390.3170.2410.119
RayDN (Ours)V2-99640×160056.564.50.4610.2410.3220.2390.114
", + "bbox": [ + 218, + 426, + 782, + 554 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 593, + 455, + 607 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Our experimental setup leverages ResNet50 [7], ResNet101, and V2-99 [11] backbones, each with distinct pre-training configurations. We detail the performance of ResNet50 and ResNet101 models, which are pre-trained on nuImages [2], on the nuScenes validation set. To showcase the scalability of our approach, we also report results on the nuScenes test set using V2-99, initialized from the DD3D [25] checkpoint. Our models are optimized using the AdamW [23] optimizer with a batch size of 16. The learning rate is set to $4 \\times 10^{-4}$ for models trained only on the training set and $3 \\times 10^{-4}$ for those trained on both the training and validation sets. We adopt a cosine annealing policy for learning rate scheduling. For benchmarking against state-of-the-art (SOTA) methods, models are trained for 60 epochs without CBGS [45] and for 24 epochs in ablation studies in nuScenes. For experiments on Argoverse 2, models are trained for 6 epochs. Our implementation is primarily based on the StreamPETR [34] framework.", + "bbox": [ + 212, + 642, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "F. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/f122060a0c270bcc0c56c44a348e556fd4ba2f2119f8cded6579660f5dcda83c.jpg", + "table_caption": [ + "Table 3: Comparisons on the Argoverse 2 validation set. We evaluate across 26 object categories within a range of 150 meters." + ], + "table_footnote": [], + "table_body": "
MethodsBackboneImage SizemAPCDSmATEmASEmAOE
PETR [21]V2-99900×64017.612.20.9110.3390.819
Sparse4Dv2 [18]V2-99900×64018.913.40.8320.3430.723
StreamPETR [34]V2-99900×64020.314.60.8430.3210.650
RayDN (Ours)V2-99900×64022.316.10.8250.3250.629
", + "bbox": [ + 267, + 184, + 736, + 252 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 215, + 280, + 620, + 294 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We compare the proposed Ray Denoising method with other state-of-the-art multi-view 3D object detectors on the validation and test sets of the nuScenes dataset, as well as the validation set of the Argoverse 2 dataset. It's important to note that our method does not employ test time augmentation (TTA).", + "bbox": [ + 215, + 306, + 784, + 364 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "nuScenes Validation Set. Table 1 presents a comparison with state-of-the-art methods on the nuScenes validation set. We evaluate both ResNet-50 and ResNet-101 backbones. With ResNet-50 as the backbone and an image size of $704 \\times 256$ , we achieve $46.9\\%$ mAP and $56.3\\%$ NDS, improving upon the previous state-of-the-art method, StreamPETR, by $1.9\\%$ mAP and $1.3\\%$ NDS. Using a more robust ResNet-101 backbone and increasing the image size to $512 \\times 1408$ , our performance reaches $51.6\\%$ mAP and $59.8\\%$ NDS, outperforming StreamPETR by $1.2\\%$ mAP and $0.6\\%$ NDS. These results demonstrate the scalability of Ray Denoising.", + "bbox": [ + 215, + 367, + 785, + 501 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "nuScenes Test Set. The results evaluated by the test server are detailed in Table 2. Our method includes training on both the training and validation sets. Notably, we achieve $56.5\\%$ mAP and $64.5\\%$ NDS, surpassing StreamPETR by an absolute $1.5\\%$ mAP and $0.9\\%$ NDS.", + "bbox": [ + 215, + 503, + 784, + 561 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Argoverse 2 Validation Set. To assess the generalization capability of Ray Denoising, we conduct additional experiments on the Argoverse 2 dataset, as displayed in Table 3. Our method significantly outperforms previous state-of-the-art methods, achieving a $2.0\\%$ mAP and $1.5\\%$ CDS improvement on the validation set. These metrics underscore the generalization capability of our approach.", + "bbox": [ + 215, + 564, + 784, + 638 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The tables demonstrate that our Ray Denoising approach notably enhances mAP. Considering that mAP is significantly impacted by incorrect false positives, this enhancement firmly validates the effectiveness of Ray Denoising in reducing redundant predictions along the camera rays.", + "bbox": [ + 215, + 638, + 784, + 699 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 723, + 387, + 738 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "This section delves into the ablation studies performed using the validation sets from the nuScenes and Argoverse 2 datasets. Unless specified otherwise, our experiments on nuScenes are based on a ResNet50 backbone, which is pre-trained on the nuImages dataset [2]. Our input data comprises 8 frames, each with a $704 \\times 256$ pixels resolution. The decoder utilizes 428 queries, and the model is trained for 24 epochs, forgoing Class-Balanced Group Sampling (CBGS) [45].", + "bbox": [ + 215, + 750, + 784, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Ray Denoising", + "bbox": [ + 632, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/84c2932491fff4e12ba73da07ba15c6db087ad6e29b3bb98bd4b394ee2312b7b.jpg", + "table_caption": [ + "Table 4: Ablation studies on the radius of ray denoising queries." + ], + "table_footnote": [], + "table_body": "
MethodRadiusmAPNDS
SOTA Baseline [34]-39.148.6
241.149.8
+RayDN (Ours)341.950.8
441.350.1
", + "bbox": [ + 250, + 183, + 493, + 252 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/aa316e7ae10113e8ab4efa0a1c1cfebf7e06d3f870b44d86f2a10996a3fdbc7d.jpg", + "table_caption": [ + "Table 6: Ablation studies on the distribution of ray denoising queries on nuScenes dataset." + ], + "table_footnote": [], + "table_body": "
Methodλ μ mAPNDS
SOTA Baseline [34]-39.148.6
+RayDN (Ours)141.350.4
241.550.1
341.350.5
741.750.7
841.950.8
941.650.4
", + "bbox": [ + 264, + 316, + 483, + 416 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/1d75207de81d4d50ec16d5425f2a13943a2f65e921a5f644695f8781b557444c.jpg", + "table_caption": [ + "Table 5: Ablation studies on the number of ray denoising queries." + ], + "table_footnote": [], + "table_body": "
Method#QmAPNDS
SOTA Baseline [34]-39.148.6
341.650.2
+RayDN (Ours)541.950.8
741.150.5
", + "bbox": [ + 526, + 183, + 748, + 252 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/da0456bb682320c156bf88d462b43d92d5134b9040d3f2243e610755512c83db.jpg", + "table_caption": [ + "Table 7: Ablation studies on the distribution of ray denoising queries for the Argoverse 2 dataset." + ], + "table_footnote": [], + "table_body": "
Methodλ μmAPCDS
SOTA Baseline [34]-20.314.6
+RayDN (Ours)121.315.4
321.715.8
421.315.3
621.515.7
722.316.1
821.515.6
", + "bbox": [ + 516, + 318, + 735, + 416 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Radius of Ray Denoising Queries. Table 4 shows how varying the radius $k$ affects performance. We find that a radius of $k = 3$ yields the optimal results, while $k = 2$ and $k = 4$ also lead to notable enhancements in performance.", + "bbox": [ + 212, + 446, + 784, + 491 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Number of Ray Denoising Queries. We explore the effect of the quantity of ray denoising queries on the model's performance in Table 5. The findings reveal that both mAP and NDS metrics improve with increasing ray queries, reaching a plateau of 5 queries. Surprisingly, upping the count to 7 queries results in a dip in mAP, which might be attributed to an imbalance in the ratio of positive to negative ray denoising queries.", + "bbox": [ + 212, + 491, + 785, + 582 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Distribution of Ray Denoising Queries. We investigate the impact of the distribution of ray denoising queries with various hyper-parameters across the nuScenes and Argoverse 2 datasets. The results are detailed in Table 6 and Table 7. Our findings indicate that employing a uniform distribution, which is dataset-agnostic (i.e., setting $\\lambda = 1$ and $\\mu = 1$ ), leads to substantial performance gains for Ray Denoising. Specifically, we observe a $2.2\\%$ increase in mAP for the nuScenes dataset and a $1.0\\%$ increase in mAP for the Argoverse 2 dataset, highlighting the method's strong generalization capabilities. Fine-tuning these hyper-parameters further enhances the performance, with an additional $0.6\\%$ mAP improvement on nuScenes and $1.0\\%$ mAP on Argoverse. Across both datasets, models configured with $\\lambda > \\mu$ slightly outperform those with $\\lambda < \\mu$ , suggesting that false positives tend to be at greater distance from the ego-vehicle.", + "bbox": [ + 212, + 582, + 787, + 765 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Precision-Recall Analysis. We assess the ability of Ray Denoising to reduce false positive predictions by plotting precision-recall curves and Average Precision (AP) for each object class. Figure 5a reveals that Ray Denoising enhances precision across a range of recall levels and under various distance thresholds.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "F. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d0b91b7a89a031bb1770ad96eab0df73bee095aefb6cb0ea81b5b7748c9927cd.jpg", + "image_caption": [ + "(a) Precision-Recall Curve" + ], + "image_footnote": [], + "bbox": [ + 243, + 145, + 498, + 290 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/245734db36b9a630fd1de0330c5d8c048ad2804fd257f9c15985555fe8cb0374.jpg", + "image_caption": [ + "(b) AP for each class" + ], + "image_footnote": [], + "bbox": [ + 508, + 147, + 753, + 287 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/67bbe4937d086929e32c83c727dd7a8a10923f3469b49e05c6cd55054de44722.jpg", + "image_caption": [ + "Fig. 5: (a) Visualization of the precision-recall curves at various distance thresholds. Ray Denoising consistently enhances precision across nearly all recall levels, effectively suppressing false positives. (b) Class-wise AP comparison. Ray Denoising performs superior over the SOTA StreamPETR in all object classes. Best viewed in color.", + "Fig.6: Visualization of our detection results on the validation set of nuScenes. Ray Denoising effectively mitigates duplicate false positives while maintaining the ability to detect highly occluded objects along the same ray. Best viewed on the screen." + ], + "image_footnote": [], + "bbox": [ + 245, + 388, + 756, + 467 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Similarly, Figure 5b shows that Ray Denoising surpasses StreamPETR in AP across all classes. These results collectively affirm that Ray Denoising effectively mitigates the occurrence of false positive predictions.", + "bbox": [ + 212, + 551, + 784, + 595 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Visualization of Detection Results. Our approach creates multiple spatial hard negative samples along the same ray, potentially interfering with detecting true positives if they are also located along that ray. To test this, we present a visualization of the detection outcomes in Figure 6. The results show that Ray Denoising effectively eliminates duplicate false positives without compromising the detection of true positives, even when they are heavily occluded along the same ray.", + "bbox": [ + 212, + 597, + 787, + 702 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Generalization Ability of Ray Denoising. To assess the versatility of Ray Denoising, we apply it to different models with varied input data. We use the state-of-the-art CMT [40] as a reference and test Ray Denoising with single-frame images and multi-modal data. In the single-frame scenario, we employ a ResNet50 backbone pre-trained on ImageNet [2], with input images sized at $256 \\times 704$ pixels. For the multi-modal setup, the input image resolution is $320 \\times 800$ pixels, and the voxel dimensions are $0.1\\mathrm{m} \\times 0.1\\mathrm{m}$ . The batch size is set to 16 due to memory constraints, and all models are trained for 20 epochs using CBGS [45]. Table 8 shows that Ray Denoising improves upon the baseline", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Ray Denoising", + "bbox": [ + 632, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/1645f47af6ae69c96d16f351fd2a22a58eb330813f8539d1b209bc031e86aa64.jpg", + "table_caption": [ + "Table 8: Ablation studies on generalization ability of Ray Denoising." + ], + "table_footnote": [], + "table_body": "
MethodDescriptionmAPNDS
CMT [40]single frame30.637.7
+ RayDNsingle frame32.139.1
CMT [40]multi-modal67.670.4
+ RayDNmulti-modal68.671.3
", + "bbox": [ + 282, + 183, + 500, + 250 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/cacbb5497133a3442ddbf109fae76fea9274f5692048b8d50a53c95a4526fe67.jpg", + "table_caption": [ + "Table 9: Ablation studies on the effect of depth estimation." + ], + "table_footnote": [], + "table_body": "
3DPPERayDNmAPNDS
39.148.6
✓42.051.1
✓41.950.8
✓✓43.952.9
", + "bbox": [ + 532, + 183, + 715, + 250 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "with a $1.5\\%$ increase in mAP and a $1.4\\%$ increase in NDS for the single-frame setup. Moreover, it achieves a $1.0\\%$ mAP and a $0.9\\%$ NDS enhancement on the already high-performing multi-modal baseline. These experiments confirm the broad applicability of Ray Denoising.", + "bbox": [ + 212, + 277, + 784, + 337 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Effect of Depth Estimation on Ray Denoising. In DETR-style multi-view 3D object detectors, the lack of depth information leads to the use of camera rays as position embeddings for image features, which can be imprecise. To refine these embeddings, 3DPPE [30] suggests encoding an estimated 3D point guided by external supervision for depth estimation. To explore the role of depth estimation in conjunction with Ray Denoising, we integrate 3DPPE into our framework. The results, detailed in Table 9, show a marked improvement in performance with 3DPPE. However, Ray Denoising alone, without external depth supervision, also achieves similar performance. This underscores our core insight: Ray Denoising enhances the model's depth perception by guiding it to learn superior depth perception ability during training. When Ray Denoising is paired with 3DPPE, we further obtain a $+1.9\\%$ increase in mAP and a $+1.7\\%$ increase in NDS. This demonstrates that while depth estimation algorithms can improve the clarity of position embeddings, Ray Denoising's ability to refine the model's understanding of depth through robust training is a powerful complement, even when depth information is available.", + "bbox": [ + 212, + 337, + 785, + 579 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 599, + 359, + 616 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We introduce Ray Denoising, a method designed to overcome the critical challenge of depth estimation in multi-view 3D object detection. Ray Denoising tackles the issue of false detections along camera rays, which are a direct consequence of imprecise depth information from images. By leveraging the 3D structure of the scene, Ray Denoising prompts the model to learn depth-aware features, leading to improved differentiation between false and true positives along the same ray without introducing extra inference costs. Our comprehensive experiments on the NuScenes and Argoverse 2 datasets demonstrate that Ray Denoising consistently and significantly outperforms strong baselines, achieving new state-of-the-art performance in Multi-view 3D Object Detection. Acknowledgment. This work was supported by the Fundamental Research Funds for the Central Universities (E3E41903, E2ET1104, E3ET6201X2), the National Natural Science Foundation of China (NSFC) under Grant 62225208 and 62171431.", + "bbox": [ + 212, + 628, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "F. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Brazil, G., Liu, X.: M3d-rpn: Monocular 3d region proposal network for object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9287-9296 (2019)", + "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)", + "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)", + "4. Felzenszwalb, P.F., Girshick, R.B., McAllester, D., Ramanan, D.: Object detection with discriminatively trained part-based models. IEEE transactions on pattern analysis and machine intelligence 32(9), 1627-1645 (2009)", + "5. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset. The International Journal of Robotics Research 32(11), 1231-1237 (2013)", + "6. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Region-based convolutional networks for accurate object detection and segmentation. IEEE transactions on pattern analysis and machine intelligence 38(1), 142-158 (2015)", + "7. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "8. Huang, J., Huang, G.: Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054 (2022)", + "9. Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multicamera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790 (2021)", + "0. Jiang, Y., Zhang, L., Miao, Z., Zhu, X., Gao, J., Hu, W., Jiang, Y.G.: Polarformer: Multi-camera 3d object detection with polar transformer. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1042-1050 (2023)", + "1. Lee, Y., Hwang, J.w., Lee, S., Bae, Y., Park, J.: An energy andgpu-computation efficient backbone network for real-time object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 0-0 (2019)", + "2. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13619-13627 (2022)", + "3. Li, Y., Bao, H., Ge, Z., Yang, J., Sun, J., Li, Z.: Bevstereo: Enhancing depth estimation in multi-view 3d object detection with temporal stereo. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1486-1494 (2023)", + "4. Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1477-1485 (2023)", + "5. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: European conference on computer vision. pp. 1-18. Springer (2022)", + "6. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)" + ], + "bbox": [ + 225, + 172, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Ray Denoising", + "bbox": [ + 632, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Lin, X., Lin, T., Pei, Z., Huang, L., Su, Z.: Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion. arXiv preprint arXiv:2211.10581 (2022)", + "18. Lin, X., Lin, T., Pei, Z., Huang, L., Su, Z.: Sparse4d v2: Recurrent temporal fusion with sparse model. arXiv preprint arXiv:2305.14018 (2023)", + "19. Liu, H., Teng, Y., Lu, T., Wang, H., Wang, L.: Sparsebev: High-performance sparse 3d object detection from multi-camera videos. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18580-18590 (2023)", + "20. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329 (2022)", + "21. Liu, Y., Wang, T., Zhang, X., Sun, J.: Petr: Position embedding transformation for multi-view 3d object detection. In: European Conference on Computer Vision. pp. 531-548. Springer (2022)", + "22. Liu, Y., Yan, J., Jia, F., Li, S., Gao, A., Wang, T., Zhang, X.: Petrov2: A unified framework for 3d perception from multi-camera images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3262-3272 (2023)", + "23. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2018)", + "24. Ma, X., Zhang, Y., Xu, D., Zhou, D., Yi, S., Li, H., Ouyang, W.: Delving into localization errors for monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4721-4730 (2021)", + "25. Park, D., Ambrus, R., Guizilini, V., Li, J., Gaidon, A.: Is pseudo-lidar needed for monocular 3d object detection? In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3142-3152 (2021)", + "26. Park, J., Xu, C., Yang, S., Keutzer, K., Kitani, K.M., Tomizuka, M., Zhan, W.: Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. In: The Eleventh International Conference on Learning Representations (2022)", + "27. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 194-210. Springer (2020)", + "28. Reading, C., Harakeh, A., Chae, J., Waslander, S.L.: Categorical depth distribution network for monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8555-8564 (2021)", + "29. Shrivastava, A., Gupta, A., Girshick, R.: Training region-based object detectors with online hard example mining. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 761-769 (2016)", + "30. Shu, C., Deng, J., Yu, F., Liu, Y.: 3dppe: 3d point positional encoding for transformer-based multi-camera 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 3580-3589 (October 2023)", + "31. Simonelli, A., Bulo, S.R., Porzi, L., López-Antequera, M., Kontschieder, P.: Disentangling monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1991-1999 (2019)", + "32. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2446-2454 (2020)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "F. Liu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Sung, K.K., Poggio, T.: Example-based learning for view-based human face detection. IEEE Transactions on pattern analysis and machine intelligence 20(1), 39-51 (1998)", + "34. Wang, S., Liu, Y., Wang, T., Li, Y., Zhang, X.: Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 3621-3631 (October 2023)", + "35. Wang, T., Zhu, X., Pang, J., Lin, D.: Fcos3d: Fully convolutional one-stage monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 913-922 (2021)", + "36. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor detr: Query design for transformer-based detector. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 2567-2575 (2022)", + "37. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: Conference on Robot Learning. pp. 180–191. PMLR (2022)", + "38. Wang, Z., Huang, Z., Fu, J., Wang, N., Liu, S.: Object as query: Lifting any 2d object detector to 3d detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3791-3800 (2023)", + "39. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., et al.: Argoverse 2: Next generation datasets for self-driving perception and forecasting. In: Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2) (2021)", + "40. Yan, J., Liu, Y., Sun, J., Jia, F., Li, S., Wang, T., Zhang, X.: Cross modal transformer: Towards fast and robust 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18268-18278 (2023)", + "41. Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z., Huang, G., Li, H., Qiao, Y., Lu, L., et al.: Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17830-17839 (2023)", + "42. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In: The Eleventh International Conference on Learning Representations (2022)", + "43. Zhang, H., Li, H., Zeng, A., Li, F., Liu, S., Liao, X., Zhang, L.: Introducing depth into transformer-based 3d object detection. arXiv preprint arXiv:2302.13002 (2023)", + "44. Zhang, Y., Lu, J., Zhou, J.: Objects are different: Flexible monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3289-3298 (2021)", + "45. Zhu, B., Jiang, Z., Zhou, X., Li, Z., Yu, G.: Class-balanced grouping and sampling for point cloud 3d object detection. arXiv preprint arXiv:1908.09492 (2019)", + "46. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)", + "47. Zong, Z., Jiang, D., Song, G., Xue, Z., Su, J., Li, H., Liu, Y.: Temporal enhanced training of multi-view 3d object detector via historical object prediction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 3781-3790 (October 2023)" + ], + "bbox": [ + 212, + 146, + 787, + 811 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Ray Denoising", + "bbox": [ + 632, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Ray Denoising_ Depth-aware Hard Negative Sampling for Multi-view 3D Object Detection/ebbe937b-35fd-41db-b04a-6c0c4e0d90c8_model.json b/2024/Ray Denoising_ Depth-aware Hard Negative Sampling for Multi-view 3D Object Detection/ebbe937b-35fd-41db-b04a-6c0c4e0d90c8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..21b5d19d02436025fb50ca843923ffb0f9523227 --- /dev/null +++ b/2024/Ray Denoising_ Depth-aware Hard Negative Sampling for Multi-view 3D Object Detection/ebbe937b-35fd-41db-b04a-6c0c4e0d90c8_model.json @@ -0,0 +1,2236 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.24, + 0.141, + 0.764, + 0.187 + ], + "angle": 0, + "content": "Ray Denoising: Depth-aware Hard Negative Sampling for Multi-view 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.213, + 0.786, + 0.245 + ], + "angle": 0, + "content": "Feng Liu1* Tengteng Huang2 Qianjing Zhang2 Haotian Yao2 Chi Zhang2 Fang Wan1 Qixiang Ye1 Yanzhao Zhou1†" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.255, + 0.761, + 0.283 + ], + "angle": 0, + "content": "1 University of Chinese Academy of Sciences liufeng20@mails.ucas.ac.cn {wanfang,qxye,zhouyanzhao}@ucas.ac.cn" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.283, + 0.764, + 0.311 + ], + "angle": 0, + "content": "2 Mach Drive {tengteng.huang,qianjing.zhang,haotian.yao,chi.zhang}@mach-drive.com" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.354, + 0.74, + 0.603 + ], + "angle": 0, + "content": "Abstract. Multi-view 3D object detection systems often struggle with generating precise predictions due to the challenges in estimating depth from images, increasing redundant and incorrect detections. Our paper presents Ray Denoising, an innovative method that enhances detection accuracy by strategically sampling along camera rays to construct hard negative examples. These examples, visually challenging to differentiate from true positives, compel the model to learn depth-aware features, thereby improving its capacity to distinguish between true and false positives. Ray Denoising is designed as a plug-and-play module, compatible with any DETR-style multi-view 3D detectors, and it only minimally increases training computational costs without affecting inference speed. Our comprehensive experiments, including detailed ablation studies, consistently demonstrate that Ray Denoising outperforms strong baselines across multiple datasets. It achieves a \\(1.9\\%\\) improvement in mean Average Precision (mAP) over the state-of-the-art StreamPETR method on the NuScenes dataset. It shows significant performance gains on the Argoverse 2 dataset, highlighting its generalization capability. The code is available at https://github.com/LiewFeng/RayDN." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.617, + 0.74, + 0.645 + ], + "angle": 0, + "content": "Keywords: Multi-view 3D Object Detection \\(\\cdot\\) Depth-aware Hard Negative Sampling \\(\\cdot\\) Ray Denoising" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.674, + 0.377, + 0.691 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.709, + 0.788, + 0.814 + ], + "angle": 0, + "content": "3D object detection is a crucial component in autonomous driving systems, drawing considerable interest from the computer vision community. The field of image-based 3D object detection [9,10,13,14,37,38] is experiencing a surge in research due to its cost-effectiveness compared to LiDAR-based solutions. A key challenge in multi-view 3D object detection, which relies on images from surrounding cameras, is the difficulty in estimating depth from images, leading to duplicate predictions, as shown in Figure 1." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.731, + 0.841 + ], + "angle": 0, + "content": "* Work was done during internship at Mach Drive. † Corresponding Author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "F. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.144, + 0.675, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.324, + 0.788, + 0.366 + ], + "angle": 0, + "content": "Fig. 1: The challenge of estimating depth from images in multi-view 3D object detection leads to duplicate predictions and false positive detections along camera rays. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.789, + 0.489 + ], + "angle": 0, + "content": "Despite the methodological improvements, multi-view 3D object detectors struggle to reduce false positive predictions arising from depth ambiguities. Several recent studies [8, 15, 17, 19, 22, 26, 34, 41, 47] have attempted to tackle this issue by incorporating temporal information. However, these methods do not explicitly account for the 3D structure of the scene, which limits their potential for further enhancement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.789, + 0.627 + ], + "angle": 0, + "content": "Furthermore, previous works have explored applying general techniques such as Non-Maximum Suppression (NMS) and Focal Loss to mitigate duplicate predictions. NMS, a post-processing technique, targets false positive predictions with high Intersection over Union (IoU) but is less effective when these predictions are scattered along rays with low IoU. Focal Loss, a loss function designed to reduce high-confidence false positive predictions, has also been implemented. However, it has been observed that multi-view 3D object detectors using Focal Loss still face challenges in effectively resolving the problem of false positive predictions along the same rays." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.789, + 0.749 + ], + "angle": 0, + "content": "Our quantitative analysis highlighted the importance of addressing false positive predictions along the same ray as Ground Truth. By utilizing the precise positional data of Ground Truth objects, we could identify and eliminate these redundant predictions within the state-of-the-art StreamPETR method [34]. This process significantly enhanced \\(5.4\\%\\) in mean Average Precision (mAP), emphasizing the critical need for models to improve their depth estimation capabilities. The substantial improvement underscores the potential of refining depth estimation to suppress these predictions and enhance overall detection performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Our key observation is that false positives often occur along camera rays due to the inherent limitation in conventional multi-view object detectors. Since the depth information for each pixel is not accurately estimated, the position embedding can only encode the ray direction. As a result, queries on the same ray will consistently interact with identical visual features from the image, leading to numerous duplicate predictions (false positives) along that ray. This scenario" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.633, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Ray Denoising" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": "underscores the model's need to learn depth-aware features that can discern objects in depth despite the visual features being the same for objects along the same ray. We propose a novel method called Ray Denoising (i.e., RayDN). This framework is inherently flexible and does not limit the choice of distribution for sampling depth-aware hard negative samples. Based on our ablation studies, we have chosen the Beta distribution for its effectiveness in capturing the spatial distribution of false positives that models are likely to generate. This choice enables Ray Denoising to create depth-aware hard negative samples used for denoising, thereby enhancing the model's ability to learn more robust features and representations for distinguishing false positives along the ray, as depicted in Figure 2. Ray Denoising introduces only a marginal increase in computational costs during the training stage without affecting inference speed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.525 + ], + "angle": 0, + "content": "To summarize our contributions, we highlight the following key points. Firstly, we have identified the persistent challenge of false positive predictions along the same ray, which acts as a bottleneck in the performance of multi-view 3D object detectors. Secondly, we introduce Ray Denoising, a novel denoising method that utilizes the Beta distribution to create depth-aware hard negative samples along rays. This method explicitly takes into account the 3D structure of the scene, offering a flexible solution compatible with any DETR-style multi-view 3D detector to address the issue of duplicate predictions along rays. Lastly, our method achieves state-of-the-art results on the NuScenes dataset [2], significantly enhancing the performance of multi-view 3D object detectors. Specifically, we have improved upon the current state-of-the-art method, StreamPETR, by \\(1.9\\%\\) in mean Average Precision (mAP), thereby demonstrating the effectiveness of Ray Denoising." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.546, + 0.388, + 0.563 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.575, + 0.543, + 0.592 + ], + "angle": 0, + "content": "2.1 Image-based 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.735 + ], + "angle": 0, + "content": "3D object detection is a cornerstone task in autonomous driving systems. The domain has witnessed notable advancements in monocular 3D object detection, largely attributed to the KITTI benchmark [5]. This progress has been driven by a range of research studies [1,24,28,31,35,44]. Nevertheless, monocular systems face limitations due to their dependence on a single viewpoint and limited data, which hampers their capacity to manage complex scenarios. To overcome these limitations, extensive benchmarks [2,32,39] have been developed, offering multi-viewpoint data that enriches the field of multi-view 3D object detection and propels the evolution of advanced detection methodologies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.802, + 0.842 + ], + "angle": 0, + "content": "The multi-view 3D object detection research is primarily bifurcated into dense Bird's Eye View (BEV)-based and sparse query-based algorithms. Our work is categorized under the sparse query-based algorithms. Dense BEV-based algorithms convert multi-view image features into a dense BEV representation using the Lift-Splat-Shoot (LSS) technique [27]. To mitigate overfitting from LSS, BEVDet [9] recommends data augmentation on BEV features. BEVDepth [14] introduces depth estimation supervision to enhance LSS precision. Compared to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "F. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.312, + 0.788, + 0.356 + ], + "angle": 0, + "content": "Fig. 2: The proposed Ray Denoising approach (right) effectively reduces false positive detections along the ray (highlighted by red rectangles) in the previous state-of-the-art method StreamPETR [34] (left). Best viewed by zooming on the screen." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.383, + 0.789, + 0.534 + ], + "angle": 0, + "content": "LSS-like methods, BEVFormer-like and DETR-like methods can leverage the strong capabilities of transformers for spatial-temporal modeling. BEVFormer-like methods use dense BEV features to enable multiple tasks (e.g., detection and segmentation) but also introduce high computational costs. Sparse query-based algorithms such as DETR3D [37] and PETR [21] utilize learnable 3D object queries and position encoding to engage with multi-view 2D features. DAT [43] introduces a pipeline to construct depth-aware position embeddings and performs contrastive learning on BEV features to tackle duplicate predictions. Recent studies [8, 13, 15, 17, 19, 22, 26, 34] have integrated temporal information to refine these methods further." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.557, + 0.523, + 0.574 + ], + "angle": 0, + "content": "2.2 Hard Negative Samples Mining" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.789, + 0.75 + ], + "angle": 0, + "content": "Hard negative sample mining is a strategy to minimize false positive predictions in object detection. Early works like R-CNN [6] employed standard hard negative mining methods [4, 33], where misclassified samples were utilized for retraining. OHEM [29] and RetinaNet [16] introduced adaptive sampling and focal loss to address the challenges of heuristics and hyper-parameters in hard negative sample selection. These methods select hard negative samples from existing predictions based on confidence scores and matched ground-truth object classes. Our approach takes a different path by constructing new hard negative samples based on the locations of ground-truth objects rather than selecting from existing predictions or depending on prediction confidence and matched ground-truth object classes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.77, + 0.515, + 0.787 + ], + "angle": 0, + "content": "2.3 Denoising in Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.842 + ], + "angle": 0, + "content": "In 2D object detection, models such as DETR [3] and its variants [20,36,46] often face convergence issues, which are partly due to the instability of bipartite graph matching and the inconsistency in optimization goals during the early stages of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.633, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Ray Denoising" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.265 + ], + "angle": 0, + "content": "training. To tackle this, DN-DETR [12] introduces the concept of feeding noisy ground-truth bounding boxes into the Transformer decoder, with the model being trained to reconstruct the original, clean boxes. This process, known as denoising, involves using queries initialized with these noisy ground-truth bounding boxes, referred to as denoising queries. DINO [42] further refines this approach by generating both positive and negative denoising queries for each ground-truth bounding box, with the negative queries incorporating more noise to improve the model's ability to reject incorrect predictions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.785, + 0.465 + ], + "angle": 0, + "content": "In the realm of 3D object detection, DETR-style methods [19,22,34] perpetuate the denoising approach by creating a denoising query for each ground-truth bounding box. Based on the noise level, these queries are classified as either an object or 'no object'. However, these methods neglect to incorporate knowledge of the 3D scene structure when generating denoising queries. In multi-view 3D object detection, the absence of depth information for each pixel results in models producing duplicate predictions along the same ray with different depths. To combat these false positive predictions, we harness the 3D structure knowledge of the scene. We devise multiple negative ray-denoising queries distributed along the rays, spatially distinct from the traditional negative denoising queries surrounding the ground-truth objects. This innovative design markedly improves the detector's capability to differentiate between true positive predictions and false positive predictions along the same ray." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.488, + 0.381, + 0.506 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.788, + 0.687 + ], + "angle": 0, + "content": "This section elaborates on our proposed method, Ray Denoising, for multi-view 3D object detection. We start with a comprehensive overview of the DETR-style multi-view 3D object detector framework in Section 3.1. Following that, we describe the three pivotal steps for integrating Ray Denoising into the framework to address the challenge of duplicate predictions. Section 3.2 outlines the ray generation process. Section 3.3 explains utilizing the Beta distribution family for sampling reference points. Section 3.4 details the construction of spatial denoising queries, referred to as Ray Queries, along each ray. Lastly, Section 3.5 delves into the essential differences between our Ray Queries and the denoising queries employed in prior works and the influence of Ray Denoising on the model training process." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.71, + 0.341, + 0.723 + ], + "angle": 0, + "content": "3.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Ray Denoising is designed to be integrated into any DETR-style multi-view 3D object detector [15, 17, 19, 21, 22, 34, 37]. These detectors typically consist of a convolutional network-based feature encoder and a transformer-based decoder. The process begins by feeding \\(N\\) surround-view images \\(\\pmb{I} = \\{I_i\\in \\mathbb{R}^{3\\times H_I\\times W_I},i = 1,2,\\dots,N\\}\\) into the feature encoder (e.g., ResNet101 [7]) to extract image features \\(\\pmb{F} = \\{F_{i}\\in \\mathbb{R}^{C\\times H_{F}\\times W_{F}},i = 1,2,\\dots,N\\}\\). Here, \\(H_{I}\\) and \\(W_{I}\\) denote the image dimensions, \\(H_{F}\\) and \\(W_{F}\\) are the feature dimensions, and \\(C\\) represents the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "F. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.149, + 0.778, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.346, + 0.788, + 0.445 + ], + "angle": 0, + "content": "Fig. 3: Overall framework of the Ray Denoising approach, a plug-and-play training technique for DETR-style multi-view 3D object detectors, focuses on refining the model's ability to distinguish true positives from false positives in depth. Casting rays and sampling depth-aware denoising queries effectively tackle the challenge of false positives arising from the inherent difficulties in visually estimating depth, leading to substantial improvements in detection performance over strong baselines. Best viewed in color and by zooming on the screen." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.788, + 0.534 + ], + "angle": 0, + "content": "number of channels in the feature. To facilitate 3D perception, multiple points within each camera's frustum are transformed and encoded into position embeddings [21]. These embeddings enable the multi-view image features to interact with 3D queries." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.645 + ], + "angle": 0, + "content": "The final step involves \\(N\\) object queries \\(\\mathbf{Q} \\in \\mathbb{R}^{N \\times 256}\\), which are derived from a set of learnable 3D reference points \\(\\mathbf{P} \\in \\mathbb{R}^{N \\times 3}\\). These queries interact with the multi-view image features \\(\\mathbf{F}\\) in the transformer decoder, employing multi-layer cross-attention to identify objects. The decoder's output features are then processed by prediction heads (multi-layer perceptron, i.e., MLP) to yield classification scores \\((cls)\\), position offsets \\((x, y, z)\\), scales \\((w, h, l)\\), directions \\((\\theta_x, \\theta_y)\\), and velocities \\((v_x, v_y)\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.666, + 0.367, + 0.683 + ], + "angle": 0, + "content": "3.2 Ray Casting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.694, + 0.788, + 0.768 + ], + "angle": 0, + "content": "The reference points for our ray denoising queries are distributed along the camera ray, which extends from the camera's optical center to the ground-truth object on the image plane. To establish this ray, we project the 3D center of the ground-truth object into the camera frustum space using the following transformation:" + }, + { + "type": "equation", + "bbox": [ + 0.446, + 0.77, + 0.787, + 0.788 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} ^ {\\prime} = \\mathbf {K} \\cdot \\mathbf {C} _ {G T}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "where \\(\\mathbf{K}\\) is the \\(4\\times 4\\) transformation matrix that maps points from 3D world space to camera frustum space. \\(\\mathbf{C}_{GT} = (x,y,z,1)\\) is the center of the groundtruth object in 3D space, and \\(\\mathbf{C}' = (u\\times d,v\\times d,d,1)\\) is the corresponding" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.633, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Ray Denoising" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.148, + 0.497, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.295, + 0.456, + 0.308 + ], + "angle": 0, + "content": "(a) Distribution Comparison" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.146, + 0.758, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.295, + 0.714, + 0.308 + ], + "angle": 0, + "content": "(b) Beta distribution family" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.318, + 0.788, + 0.375 + ], + "angle": 0, + "content": "Fig. 4: (a) Distribution comparison showing that the Beta distribution is bounded between -1 and 1, unlike the Laplace and Gaussian distributions, which are unbounded. (b) The Beta distribution family, with the x-range adjusted from [0, 1] to \\([-1, 1]\\) using the transformation \\(y = 2x - 1\\). Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.785, + 0.434 + ], + "angle": 0, + "content": "projected center in the camera frustum, with \\((u,v)\\) being the pixel coordinates and \\(d\\) the depth value." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.785, + 0.466 + ], + "angle": 0, + "content": "With the depth of the ground-truth object's center known, we can determine the coordinates of valid reference points along the ray as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.424, + 0.477, + 0.785, + 0.495 + ], + "angle": 0, + "content": "\\[\n\\hat {d} _ {i} = d + \\beta_ {i} \\cdot f \\left(\\mathbf {S} _ {G T}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.784, + 0.552 + ], + "angle": 0, + "content": "where \\( f \\) is a function that encodes the average scale of the ground-truth object, calculated as \\( f(\\mathbf{S}_{GT}) = k \\cdot \\frac{w + h + l}{6} \\). The radius \\( k \\) defines the valid distribution range for reference points, and \\( \\beta_{i} \\) is the offset for the \\( i \\)-th reference point." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.553, + 0.641, + 0.568 + ], + "angle": 0, + "content": "The position of the reference point is then obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.442, + 0.578, + 0.785, + 0.596 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {P}} _ {i} = \\mathbf {K} ^ {- 1} \\cdot \\hat {\\mathbf {C}} _ {i} ^ {\\prime}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.608, + 0.46, + 0.626 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{C}}_i^{\\prime} = (u\\times \\hat{d}_i,v\\times \\hat{d}_i,\\hat{d}_i,1)\\)" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.648, + 0.421, + 0.663 + ], + "angle": 0, + "content": "3.3 Sample Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.809 + ], + "angle": 0, + "content": "With the valid reference points along the rays defined, we proceed to sample \\(N\\) reference points to emulate the distribution of false positive predictions that arise from depth ambiguities. A straightforward uniform distribution has proven effective, yielding a \\(2.2\\%\\) mAP improvement as depicted in Figure 4a. To more accurately simulate the distribution of false positives, we consider two key aspects. First, we assess whether models tend to predict objects as being closer or further away from the ego-vehicle relative to the actual positions of the ground-truth objects. Second, we examine whether models are more likely to predict objects near or far from the center of the ground-truth objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Distributions like the Beta, Gaussian, and Laplace are suitable for handling complex scenarios, with the Gaussian and Laplace being unbounded. However," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "F. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "samples with significant offsets from the ground truth can result in large regression losses, negatively impacting optimization. As shown in Figure 4a, the Beta distribution slightly outperforms the others. Notably, the uniform distribution is a special case within the Beta family. For a versatile sampling strategy applicable to all scenarios, we select the Beta distribution family, characterized by the probability density function:" + }, + { + "type": "equation", + "bbox": [ + 0.355, + 0.248, + 0.786, + 0.281 + ], + "angle": 0, + "content": "\\[\np (x | \\lambda , \\mu) = \\frac {\\Gamma (\\lambda + \\mu)}{\\Gamma (\\lambda) + \\Gamma (\\mu)} x ^ {\\lambda - 1} (1 - x) ^ {\\mu - 1}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.289, + 0.789, + 0.32 + ], + "angle": 0, + "content": "where \\(\\lambda\\) and \\(\\mu\\) are hyper-parameters that shape the distribution. The Gamma function, denoted as \\(\\Gamma(x)\\), is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.42, + 0.33, + 0.786, + 0.363 + ], + "angle": 0, + "content": "\\[\n\\Gamma (x) = \\int_ {0} ^ {\\infty} t ^ {x - 1} e ^ {- t} d t. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.371, + 0.789, + 0.476 + ], + "angle": 0, + "content": "The beta distribution's probability density function (PDF), visualized in Figure 4b, reveals its behavior under different parameter settings. When \\(\\lambda\\) equals \\(\\mu\\), the distribution is symmetric, reflecting a scenario where the model predicts objects with equal likelihood at closer or further distances. For instance, with \\(\\lambda = \\mu = 1\\), the distribution simplifies to a uniform distribution. Adjusting \\(\\lambda\\) and \\(\\mu\\) to be higher shifts the reference points of the denoising queries closer to the ground-truth object's center and vice versa." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.79, + 0.584 + ], + "angle": 0, + "content": "When \\(\\lambda\\) exceeds \\(\\mu\\), such as in the case of \\(\\lambda = 8\\) and \\(\\mu = 2\\), the distribution suggests that models are more likely to predict objects further away from the ego-vehicle. Conversely, when \\(\\lambda\\) is less than \\(\\mu\\), for example, \\(\\lambda = 2\\) and \\(\\mu = 8\\), the distribution indicates a propensity for models to predict objects closer to the ego-vehicle. This flexibility in the Beta distribution allows for tailored sampling strategies that can adapt to the specific challenges of depth estimation in multiview 3D object detection." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.605, + 0.403, + 0.621 + ], + "angle": 0, + "content": "3.4 Query Denoising" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We construct ray denoising queries from the sampled \\(N\\) reference points. Specifically, a multi-layer perceptron (MLP) is utilized to project the normalized 3D coordinates of each point into a latent feature space. These \\(N\\) ray denoising queries \\(\\pmb{q}_r\\) are combined with the learnable object queries \\(\\pmb{q}_o\\) from the baseline detector and are input into the transformer decoder. An attention mask is employed during the self-attention operations to prevent information leakage from the ray denoising queries to the object queries. The query closest to the center of the ground truth (GT) object is designated as the positive query, inheriting the class label of the corresponding GT object. The remaining \\(N - 1\\) queries are labeled as background, i.e., 'no object'. The loss for ray denoising queries adheres to the same criteria as for the learnable object queries, using Focal Loss for classification and L1 Loss for regression. This process ensures that the model is effectively trained to discern between true positives and false positives along the same ray." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.633, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Ray Denoising" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.35, + 0.16 + ], + "angle": 0, + "content": "3.5 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.168, + 0.788, + 0.532 + ], + "angle": 0, + "content": "Our Ray Denoising approach is based on the pivotal observation that image-based 3D detection systems often struggle to distinguish true positives from false positives along camera rays. DETR-style multi-view 3D object detectors implicitly learn depth estimation from ground truth supervision. However, the randomly distributed reference points of learnable queries do not fully leverage the available ground truth information. While these reference points are updated during training, they fail to provide adequate hard negative samples for each object in every scene. To enhance the utilization of ground truth information, traditional denoising techniques introduce additional reference points distributed uniformly around the ground-truth object during training. These instance-specific reference points have improved detection performance [19,22,34]. However, they overlook the depth ambiguities intrinsic to multi-view 3D object detection. The absence of precise depth information for each pixel means that the position embedding can only encode ray direction, not depth. This leads to queries on the same ray interacting with the same image features, resulting in redundant predictions. Ray Denoising diverges from traditional techniques by strategically generating reference points along rays that extend from cameras to objects. This approach explicitly considers the 3D structure of each object in the scene, providing a sufficient number of hard negative samples. During training, these Ray Queries interact within the self-attention layer, effectively guiding the model to suppress depth-ambiguous spatial hard negative samples. This interaction enhances the detector's ability to differentiate between true positive predictions (objects) and false positive predictions (duplicate detections), improving detection accuracy." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.55, + 0.367, + 0.567 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.578, + 0.433, + 0.592 + ], + "angle": 0, + "content": "4.1 Dataset and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Our model's performance was assessed using two datasets: nuScenes [2] and Argovverse 2 [39]. The nuScenes dataset include 1000 video sequences, which are split into training (700 videos), validation (150 videos), and testing (150 videos) sets, each approximately 20 seconds long with annotations at 0.5-second intervals. The dataset comprises 1.4 million annotated 3D bounding boxes across ten object classes. Evaluation metrics include the mean Average Precision (mAP) and five true positive metrics: ATE, ASE, AOE, AVE, and AAE, which assess translation, scale, orientation, velocity, and attribute errors. The nuScenes Detection Score (NDS) is a comprehensive score derived from these metrics, offering an overall performance evaluation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The Argoverse 2 dataset contains 1000 unique scenes, each 15 seconds long, annotated at \\(10\\mathrm{Hz}\\). The scenes are divided into 700 for training, 150 for validation, and 150 for testing. The evaluation covers 26 categories within a 150-meter range, addressing long-range perception tasks. Metrics include the mAP and the Composite Detection Score (CDS), which integrates three other true positive metrics: ATE, ASE, and AOE." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "F. Liu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.787, + 0.174 + ], + "angle": 0, + "content": "Table 1: Comparison on the nuScenes validation set. \\(\\dagger\\) Indicates methods that benefit from perspective-view pre-training." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.185, + 0.79, + 0.378 + ], + "angle": 0, + "content": "
MethodsBackboneImage SizemAPNDSmATEmASEmAOEmAVEmAAE
BevDet4D [8]ResNet50256×70432.245.70.7030.2780.4950.3540.206
PETRv2 [22]ResNet50256×70434.945.60.7000.2750.5800.4370.187
BEVDepth [14]ResNet50256×70435.147.50.6290.2670.4790.4280.198
BEVStereo [13]ResNet50256×70437.250.00.5980.2700.4380.3670.190
BEVFomerv2 [41]†ResNet50-42.352.90.6180.2730.4130.3330.188
SOLOFusion [26]ResNet50256×70442.753.40.5670.2740.5110.2520.181
SparseBEV [19]†ResNet50256×70444.855.80.5810.2710.3730.2470.190
StreamPETR [34]†ResNet50256×70445.055.00.6130.2670.4130.2650.198
RayDN† (Ours)ResNet50256×70446.956.30.5790.2640.4330.2560.187
BEVDepth [14]ResNet101512×140841.253.50.5650.2660.3580.3310.190
PETRv2 [22]†ResNet101640×160042.152.40.6810.2670.3570.3770.186
Sparse4D [17]†ResNet101900×160043.654.10.6330.2790.3630.3170.177
SOLOFusion [26]ResNet101512×140848.358.20.5030.2640.3810.2460.207
SparseBEV [19]†ResNet101512×140850.159.20.5620.2650.3210.2430.195
StreamPETR [34]†ResNet101512×140850.459.20.5690.2620.3150.2570.199
RayDN† (Ours)ResNet101512×140851.860.40.5410.2600.3150.2360.200
" + }, + { + "type": "table_caption", + "bbox": [ + 0.343, + 0.402, + 0.659, + 0.416 + ], + "angle": 0, + "content": "Table 2: Comparison on the nuScenes test set." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.428, + 0.784, + 0.555 + ], + "angle": 0, + "content": "
MethodsBackboneImage SizemAPNDSmATEmASEmAOEmAVEmAAE
DETR3D [37]V2-99900×160041.247.90.6410.2550.3940.8450.133
MV2D [38]V2-99640×160046.351.40.5420.2470.4030.8570.127
BEVFormer [15]V2-99900×160048.156.90.5820.2560.3750.3780.126
PETRv2 [22]V2-99640×160049.058.20.5610.2430.3610.3430.120
PolarFormer [10]V2-99900×160049.357.20.5560.2560.3640.4390.127
BEVStereo [13]V2-99900×160052.561.00.4310.2460.3580.3570.138
HoP [47]V2-99640×160052.861.20.4910.2420.3320.3430.109
SparseBEV [19]V2-99640×160054.362.70.5020.2440.3240.2510.126
StreamPETR [34]V2-99640×160055.063.60.4790.2390.3170.2410.119
RayDN (Ours)V2-99640×160056.564.50.4610.2410.3220.2390.114
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.594, + 0.456, + 0.608 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Our experimental setup leverages ResNet50 [7], ResNet101, and V2-99 [11] backbones, each with distinct pre-training configurations. We detail the performance of ResNet50 and ResNet101 models, which are pre-trained on nuImages [2], on the nuScenes validation set. To showcase the scalability of our approach, we also report results on the nuScenes test set using V2-99, initialized from the DD3D [25] checkpoint. Our models are optimized using the AdamW [23] optimizer with a batch size of 16. The learning rate is set to \\(4 \\times 10^{-4}\\) for models trained only on the training set and \\(3 \\times 10^{-4}\\) for those trained on both the training and validation sets. We adopt a cosine annealing policy for learning rate scheduling. For benchmarking against state-of-the-art (SOTA) methods, models are trained for 60 epochs without CBGS [45] and for 24 epochs in ablation studies in nuScenes. For experiments on Argoverse 2, models are trained for 6 epochs. Our implementation is primarily based on the StreamPETR [34] framework." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.633, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Ray Denoising" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 3: Comparisons on the Argoverse 2 validation set. We evaluate across 26 object categories within a range of 150 meters." + }, + { + "type": "table", + "bbox": [ + 0.269, + 0.185, + 0.737, + 0.253 + ], + "angle": 0, + "content": "
MethodsBackboneImage SizemAPCDSmATEmASEmAOE
PETR [21]V2-99900×64017.612.20.9110.3390.819
Sparse4Dv2 [18]V2-99900×64018.913.40.8320.3430.723
StreamPETR [34]V2-99900×64020.314.60.8430.3210.650
RayDN (Ours)V2-99900×64022.316.10.8250.3250.629
" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.281, + 0.621, + 0.295 + ], + "angle": 0, + "content": "4.3 Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.307, + 0.785, + 0.366 + ], + "angle": 0, + "content": "We compare the proposed Ray Denoising method with other state-of-the-art multi-view 3D object detectors on the validation and test sets of the nuScenes dataset, as well as the validation set of the Argoverse 2 dataset. It's important to note that our method does not employ test time augmentation (TTA)." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.368, + 0.787, + 0.502 + ], + "angle": 0, + "content": "nuScenes Validation Set. Table 1 presents a comparison with state-of-the-art methods on the nuScenes validation set. We evaluate both ResNet-50 and ResNet-101 backbones. With ResNet-50 as the backbone and an image size of \\(704 \\times 256\\), we achieve \\(46.9\\%\\) mAP and \\(56.3\\%\\) NDS, improving upon the previous state-of-the-art method, StreamPETR, by \\(1.9\\%\\) mAP and \\(1.3\\%\\) NDS. Using a more robust ResNet-101 backbone and increasing the image size to \\(512 \\times 1408\\), our performance reaches \\(51.6\\%\\) mAP and \\(59.8\\%\\) NDS, outperforming StreamPETR by \\(1.2\\%\\) mAP and \\(0.6\\%\\) NDS. These results demonstrate the scalability of Ray Denoising." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.504, + 0.785, + 0.562 + ], + "angle": 0, + "content": "nuScenes Test Set. The results evaluated by the test server are detailed in Table 2. Our method includes training on both the training and validation sets. Notably, we achieve \\(56.5\\%\\) mAP and \\(64.5\\%\\) NDS, surpassing StreamPETR by an absolute \\(1.5\\%\\) mAP and \\(0.9\\%\\) NDS." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.639 + ], + "angle": 0, + "content": "Argoverse 2 Validation Set. To assess the generalization capability of Ray Denoising, we conduct additional experiments on the Argoverse 2 dataset, as displayed in Table 3. Our method significantly outperforms previous state-of-the-art methods, achieving a \\(2.0\\%\\) mAP and \\(1.5\\%\\) CDS improvement on the validation set. These metrics underscore the generalization capability of our approach." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.64, + 0.785, + 0.7 + ], + "angle": 0, + "content": "The tables demonstrate that our Ray Denoising approach notably enhances mAP. Considering that mAP is significantly impacted by incorrect false positives, this enhancement firmly validates the effectiveness of Ray Denoising in reducing redundant predictions along the camera rays." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.724, + 0.388, + 0.739 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.75, + 0.785, + 0.84 + ], + "angle": 0, + "content": "This section delves into the ablation studies performed using the validation sets from the nuScenes and Argoverse 2 datasets. Unless specified otherwise, our experiments on nuScenes are based on a ResNet50 backbone, which is pre-trained on the nuImages dataset [2]. Our input data comprises 8 frames, each with a \\(704 \\times 256\\) pixels resolution. The decoder utilizes 428 queries, and the model is trained for 24 epochs, forgoing Class-Balanced Group Sampling (CBGS) [45]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "F. Liu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.241, + 0.145, + 0.506, + 0.173 + ], + "angle": 0, + "content": "Table 4: Ablation studies on the radius of ray denoising queries." + }, + { + "type": "table", + "bbox": [ + 0.251, + 0.184, + 0.495, + 0.253 + ], + "angle": 0, + "content": "
MethodRadiusmAPNDS
SOTA Baseline [34]-39.148.6
241.149.8
+RayDN (Ours)341.950.8
441.350.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.255, + 0.266, + 0.496, + 0.307 + ], + "angle": 0, + "content": "Table 6: Ablation studies on the distribution of ray denoising queries on nuScenes dataset." + }, + { + "type": "table", + "bbox": [ + 0.265, + 0.318, + 0.485, + 0.417 + ], + "angle": 0, + "content": "
Methodλ μ mAPNDS
SOTA Baseline [34]-39.148.6
+RayDN (Ours)141.350.4
241.550.1
341.350.5
741.750.7
841.950.8
941.650.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.517, + 0.145, + 0.761, + 0.174 + ], + "angle": 0, + "content": "Table 5: Ablation studies on the number of ray denoising queries." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.184, + 0.749, + 0.253 + ], + "angle": 0, + "content": "
Method#QmAPNDS
SOTA Baseline [34]-39.148.6
341.650.2
+RayDN (Ours)541.950.8
741.150.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.507, + 0.266, + 0.748, + 0.308 + ], + "angle": 0, + "content": "Table 7: Ablation studies on the distribution of ray denoising queries for the Argoverse 2 dataset." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.319, + 0.736, + 0.417 + ], + "angle": 0, + "content": "
Methodλ μmAPCDS
SOTA Baseline [34]-20.314.6
+RayDN (Ours)121.315.4
321.715.8
421.315.3
621.515.7
722.316.1
821.515.6
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.785, + 0.492 + ], + "angle": 0, + "content": "Radius of Ray Denoising Queries. Table 4 shows how varying the radius \\( k \\) affects performance. We find that a radius of \\( k = 3 \\) yields the optimal results, while \\( k = 2 \\) and \\( k = 4 \\) also lead to notable enhancements in performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.787, + 0.583 + ], + "angle": 0, + "content": "Number of Ray Denoising Queries. We explore the effect of the quantity of ray denoising queries on the model's performance in Table 5. The findings reveal that both mAP and NDS metrics improve with increasing ray queries, reaching a plateau of 5 queries. Surprisingly, upping the count to 7 queries results in a dip in mAP, which might be attributed to an imbalance in the ratio of positive to negative ray denoising queries." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.766 + ], + "angle": 0, + "content": "Distribution of Ray Denoising Queries. We investigate the impact of the distribution of ray denoising queries with various hyper-parameters across the nuScenes and Argoverse 2 datasets. The results are detailed in Table 6 and Table 7. Our findings indicate that employing a uniform distribution, which is dataset-agnostic (i.e., setting \\(\\lambda = 1\\) and \\(\\mu = 1\\)), leads to substantial performance gains for Ray Denoising. Specifically, we observe a \\(2.2\\%\\) increase in mAP for the nuScenes dataset and a \\(1.0\\%\\) increase in mAP for the Argoverse 2 dataset, highlighting the method's strong generalization capabilities. Fine-tuning these hyper-parameters further enhances the performance, with an additional \\(0.6\\%\\) mAP improvement on nuScenes and \\(1.0\\%\\) mAP on Argoverse. Across both datasets, models configured with \\(\\lambda > \\mu\\) slightly outperform those with \\(\\lambda < \\mu\\), suggesting that false positives tend to be at greater distance from the ego-vehicle." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Precision-Recall Analysis. We assess the ability of Ray Denoising to reduce false positive predictions by plotting precision-recall curves and Average Precision (AP) for each object class. Figure 5a reveals that Ray Denoising enhances precision across a range of recall levels and under various distance thresholds." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.633, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Ray Denoising" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.146, + 0.499, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.296, + 0.451, + 0.307 + ], + "angle": 0, + "content": "(a) Precision-Recall Curve" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.148, + 0.754, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.295, + 0.695, + 0.307 + ], + "angle": 0, + "content": "(b) AP for each class" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.318, + 0.788, + 0.375 + ], + "angle": 0, + "content": "Fig. 5: (a) Visualization of the precision-recall curves at various distance thresholds. Ray Denoising consistently enhances precision across nearly all recall levels, effectively suppressing false positives. (b) Class-wise AP comparison. Ray Denoising performs superior over the SOTA StreamPETR in all object classes. Best viewed in color." + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.389, + 0.758, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.478, + 0.788, + 0.52 + ], + "angle": 0, + "content": "Fig.6: Visualization of our detection results on the validation set of nuScenes. Ray Denoising effectively mitigates duplicate false positives while maintaining the ability to detect highly occluded objects along the same ray. Best viewed on the screen." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.785, + 0.597 + ], + "angle": 0, + "content": "Similarly, Figure 5b shows that Ray Denoising surpasses StreamPETR in AP across all classes. These results collectively affirm that Ray Denoising effectively mitigates the occurrence of false positive predictions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.788, + 0.703 + ], + "angle": 0, + "content": "Visualization of Detection Results. Our approach creates multiple spatial hard negative samples along the same ray, potentially interfering with detecting true positives if they are also located along that ray. To test this, we present a visualization of the detection outcomes in Figure 6. The results show that Ray Denoising effectively eliminates duplicate false positives without compromising the detection of true positives, even when they are heavily occluded along the same ray." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Generalization Ability of Ray Denoising. To assess the versatility of Ray Denoising, we apply it to different models with varied input data. We use the state-of-the-art CMT [40] as a reference and test Ray Denoising with single-frame images and multi-modal data. In the single-frame scenario, we employ a ResNet50 backbone pre-trained on ImageNet [2], with input images sized at \\( 256 \\times 704 \\) pixels. For the multi-modal setup, the input image resolution is \\( 320 \\times 800 \\) pixels, and the voxel dimensions are \\( 0.1\\mathrm{m} \\times 0.1\\mathrm{m} \\). The batch size is set to 16 due to memory constraints, and all models are trained for 20 epochs using CBGS [45]. Table 8 shows that Ray Denoising improves upon the baseline" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "F. Liu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.272, + 0.145, + 0.511, + 0.173 + ], + "angle": 0, + "content": "Table 8: Ablation studies on generalization ability of Ray Denoising." + }, + { + "type": "table", + "bbox": [ + 0.283, + 0.184, + 0.5, + 0.251 + ], + "angle": 0, + "content": "
MethodDescriptionmAPNDS
CMT [40]single frame30.637.7
+ RayDNsingle frame32.139.1
CMT [40]multi-modal67.670.4
+ RayDNmulti-modal68.671.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.524, + 0.145, + 0.729, + 0.173 + ], + "angle": 0, + "content": "Table 9: Ablation studies on the effect of depth estimation." + }, + { + "type": "table", + "bbox": [ + 0.533, + 0.184, + 0.717, + 0.251 + ], + "angle": 0, + "content": "
3DPPERayDNmAPNDS
39.148.6
✓42.051.1
✓41.950.8
✓✓43.952.9
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.278, + 0.785, + 0.338 + ], + "angle": 0, + "content": "with a \\(1.5\\%\\) increase in mAP and a \\(1.4\\%\\) increase in NDS for the single-frame setup. Moreover, it achieves a \\(1.0\\%\\) mAP and a \\(0.9\\%\\) NDS enhancement on the already high-performing multi-modal baseline. These experiments confirm the broad applicability of Ray Denoising." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.338, + 0.787, + 0.58 + ], + "angle": 0, + "content": "Effect of Depth Estimation on Ray Denoising. In DETR-style multi-view 3D object detectors, the lack of depth information leads to the use of camera rays as position embeddings for image features, which can be imprecise. To refine these embeddings, 3DPPE [30] suggests encoding an estimated 3D point guided by external supervision for depth estimation. To explore the role of depth estimation in conjunction with Ray Denoising, we integrate 3DPPE into our framework. The results, detailed in Table 9, show a marked improvement in performance with 3DPPE. However, Ray Denoising alone, without external depth supervision, also achieves similar performance. This underscores our core insight: Ray Denoising enhances the model's depth perception by guiding it to learn superior depth perception ability during training. When Ray Denoising is paired with 3DPPE, we further obtain a \\(+1.9\\%\\) increase in mAP and a \\(+1.7\\%\\) increase in NDS. This demonstrates that while depth estimation algorithms can improve the clarity of position embeddings, Ray Denoising's ability to refine the model's understanding of depth through robust training is a powerful complement, even when depth information is available." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.6, + 0.36, + 0.617 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We introduce Ray Denoising, a method designed to overcome the critical challenge of depth estimation in multi-view 3D object detection. Ray Denoising tackles the issue of false detections along camera rays, which are a direct consequence of imprecise depth information from images. By leveraging the 3D structure of the scene, Ray Denoising prompts the model to learn depth-aware features, leading to improved differentiation between false and true positives along the same ray without introducing extra inference costs. Our comprehensive experiments on the NuScenes and Argoverse 2 datasets demonstrate that Ray Denoising consistently and significantly outperforms strong baselines, achieving new state-of-the-art performance in Multi-view 3D Object Detection. Acknowledgment. This work was supported by the Fundamental Research Funds for the Central Universities (E3E41903, E2ET1104, E3ET6201X2), the National Natural Science Foundation of China (NSFC) under Grant 62225208 and 62171431." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.633, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Ray Denoising" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.214 + ], + "angle": 0, + "content": "1. Brazil, G., Liu, X.: M3d-rpn: Monocular 3d region proposal network for object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9287-9296 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.215, + 0.785, + 0.269 + ], + "angle": 0, + "content": "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.27, + 0.785, + 0.31 + ], + "angle": 0, + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.785, + 0.35 + ], + "angle": 0, + "content": "4. Felzenszwalb, P.F., Girshick, R.B., McAllester, D., Ramanan, D.: Object detection with discriminatively trained part-based models. IEEE transactions on pattern analysis and machine intelligence 32(9), 1627-1645 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.351, + 0.785, + 0.377 + ], + "angle": 0, + "content": "5. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset. The International Journal of Robotics Research 32(11), 1231-1237 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.785, + 0.418 + ], + "angle": 0, + "content": "6. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Region-based convolutional networks for accurate object detection and segmentation. IEEE transactions on pattern analysis and machine intelligence 38(1), 142-158 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.42, + 0.785, + 0.459 + ], + "angle": 0, + "content": "7. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.46, + 0.785, + 0.485 + ], + "angle": 0, + "content": "8. Huang, J., Huang, G.: Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.486, + 0.785, + 0.526 + ], + "angle": 0, + "content": "9. Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multicamera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.528, + 0.785, + 0.567 + ], + "angle": 0, + "content": "0. Jiang, Y., Zhang, L., Miao, Z., Zhu, X., Gao, J., Hu, W., Jiang, Y.G.: Polarformer: Multi-camera 3d object detection with polar transformer. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1042-1050 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.568, + 0.785, + 0.622 + ], + "angle": 0, + "content": "1. Lee, Y., Hwang, J.w., Lee, S., Bae, Y., Park, J.: An energy andgpu-computation efficient backbone network for real-time object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 0-0 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.623, + 0.785, + 0.662 + ], + "angle": 0, + "content": "2. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13619-13627 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.663, + 0.785, + 0.703 + ], + "angle": 0, + "content": "3. Li, Y., Bao, H., Ge, Z., Yang, J., Sun, J., Li, Z.: Bevstereo: Enhancing depth estimation in multi-view 3d object detection with temporal stereo. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1486-1494 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.704, + 0.785, + 0.744 + ], + "angle": 0, + "content": "4. Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1477-1485 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.785, + 0.799 + ], + "angle": 0, + "content": "5. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: European conference on computer vision. pp. 1-18. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "6. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.354, + 0.127 + ], + "angle": 0, + "content": "F. Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "17. Lin, X., Lin, T., Pei, Z., Huang, L., Su, Z.: Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion. arXiv preprint arXiv:2211.10581 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "18. Lin, X., Lin, T., Pei, Z., Huang, L., Su, Z.: Sparse4d v2: Recurrent temporal fusion with sparse model. arXiv preprint arXiv:2305.14018 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.206, + 0.785, + 0.247 + ], + "angle": 0, + "content": "19. Liu, H., Teng, Y., Lu, T., Wang, H., Wang, L.: Sparsebev: High-performance sparse 3d object detection from multi-camera videos. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18580-18590 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.249, + 0.785, + 0.288 + ], + "angle": 0, + "content": "20. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.291, + 0.785, + 0.331 + ], + "angle": 0, + "content": "21. Liu, Y., Wang, T., Zhang, X., Sun, J.: Petr: Position embedding transformation for multi-view 3d object detection. In: European Conference on Computer Vision. pp. 531-548. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.333, + 0.785, + 0.374 + ], + "angle": 0, + "content": "22. Liu, Y., Yan, J., Jia, F., Li, S., Gao, A., Wang, T., Zhang, X.: Petrov2: A unified framework for 3d perception from multi-camera images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3262-3272 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.376, + 0.785, + 0.402 + ], + "angle": 0, + "content": "23. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.404, + 0.785, + 0.445 + ], + "angle": 0, + "content": "24. Ma, X., Zhang, Y., Xu, D., Zhou, D., Yi, S., Li, H., Ouyang, W.: Delving into localization errors for monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4721-4730 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.785, + 0.488 + ], + "angle": 0, + "content": "25. Park, D., Ambrus, R., Guizilini, V., Li, J., Gaidon, A.: Is pseudo-lidar needed for monocular 3d object detection? In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3142-3152 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.489, + 0.785, + 0.543 + ], + "angle": 0, + "content": "26. Park, J., Xu, C., Yang, S., Keutzer, K., Kitani, K.M., Tomizuka, M., Zhan, W.: Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. In: The Eleventh International Conference on Learning Representations (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.546, + 0.785, + 0.6 + ], + "angle": 0, + "content": "27. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 194-210. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.602, + 0.785, + 0.642 + ], + "angle": 0, + "content": "28. Reading, C., Harakeh, A., Chae, J., Waslander, S.L.: Categorical depth distribution network for monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8555-8564 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.644, + 0.785, + 0.685 + ], + "angle": 0, + "content": "29. Shrivastava, A., Gupta, A., Girshick, R.: Training region-based object detectors with online hard example mining. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 761-769 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.687, + 0.785, + 0.741 + ], + "angle": 0, + "content": "30. Shu, C., Deng, J., Yu, F., Liu, Y.: 3dppe: 3d point positional encoding for transformer-based multi-camera 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 3580-3589 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.784 + ], + "angle": 0, + "content": "31. Simonelli, A., Bulo, S.R., Porzi, L., López-Antequera, M., Kontschieder, P.: Disentangling monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1991-1999 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "32. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2446-2454 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.633, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Ray Denoising" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "33. Sung, K.K., Poggio, T.: Example-based learning for view-based human face detection. IEEE Transactions on pattern analysis and machine intelligence 20(1), 39-51 (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "34. Wang, S., Liu, Y., Wang, T., Li, Y., Zhang, X.: Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 3621-3631 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.787, + 0.286 + ], + "angle": 0, + "content": "35. Wang, T., Zhu, X., Pang, J., Lin, D.: Fcos3d: Fully convolutional one-stage monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 913-922 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.328 + ], + "angle": 0, + "content": "36. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor detr: Query design for transformer-based detector. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 2567-2575 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "37. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: Conference on Robot Learning. pp. 180–191. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.412 + ], + "angle": 0, + "content": "38. Wang, Z., Huang, Z., Fu, J., Wang, N., Liu, S.: Object as query: Lifting any 2d object detector to 3d detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3791-3800 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.481 + ], + "angle": 0, + "content": "39. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., et al.: Argoverse 2: Next generation datasets for self-driving perception and forecasting. In: Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "40. Yan, J., Liu, Y., Sun, J., Jia, F., Li, S., Wang, T., Zhang, X.: Cross modal transformer: Towards fast and robust 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18268-18278 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.787, + 0.578 + ], + "angle": 0, + "content": "41. Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z., Huang, G., Li, H., Qiao, Y., Lu, L., et al.: Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17830-17839 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.619 + ], + "angle": 0, + "content": "42. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In: The Eleventh International Conference on Learning Representations (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.787, + 0.647 + ], + "angle": 0, + "content": "43. Zhang, H., Li, H., Zeng, A., Li, F., Liu, S., Liao, X., Zhang, L.: Introducing depth into transformer-based 3d object detection. arXiv preprint arXiv:2302.13002 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.688 + ], + "angle": 0, + "content": "44. Zhang, Y., Lu, J., Zhou, J.: Objects are different: Flexible monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3289-3298 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.716 + ], + "angle": 0, + "content": "45. Zhu, B., Jiang, Z., Zhou, X., Li, Z., Yu, G.: Class-balanced grouping and sampling for point cloud 3d object detection. arXiv preprint arXiv:1908.09492 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.787, + 0.757 + ], + "angle": 0, + "content": "46. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.813 + ], + "angle": 0, + "content": "47. Zong, Z., Jiang, D., Song, G., Xue, Z., Su, J., Li, H., Liu, Y.: Temporal enhanced training of multi-view 3d object detector via historical object prediction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 3781-3790 (October 2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.813 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Ray Denoising_ Depth-aware Hard Negative Sampling for Multi-view 3D Object Detection/ebbe937b-35fd-41db-b04a-6c0c4e0d90c8_origin.pdf b/2024/Ray Denoising_ Depth-aware Hard Negative Sampling for Multi-view 3D Object Detection/ebbe937b-35fd-41db-b04a-6c0c4e0d90c8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5a80508b75f35031c22283be1221b30e58bd5e21 --- /dev/null +++ b/2024/Ray Denoising_ Depth-aware Hard Negative Sampling for Multi-view 3D Object Detection/ebbe937b-35fd-41db-b04a-6c0c4e0d90c8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e6fdc770ea65f15f366853985338037e0a1e266f1bc0326dc7a8981b6daf03b +size 1736931 diff --git a/2024/Ray Denoising_ Depth-aware Hard Negative Sampling for Multi-view 3D Object Detection/full.md b/2024/Ray Denoising_ Depth-aware Hard Negative Sampling for Multi-view 3D Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9e1013369be809cbdf65de00df5251dd2c57cd1a --- /dev/null +++ b/2024/Ray Denoising_ Depth-aware Hard Negative Sampling for Multi-view 3D Object Detection/full.md @@ -0,0 +1,283 @@ +# Ray Denoising: Depth-aware Hard Negative Sampling for Multi-view 3D Object Detection + +Feng Liu1* Tengteng Huang2 Qianjing Zhang2 Haotian Yao2 Chi Zhang2 Fang Wan1 Qixiang Ye1 Yanzhao Zhou1† + +1 University of Chinese Academy of Sciences liufeng20@mails.ucas.ac.cn {wanfang,qxye,zhouyanzhao}@ucas.ac.cn + +2 Mach Drive {tengteng.huang,qianjing.zhang,haotian.yao,chi.zhang}@mach-drive.com + +Abstract. Multi-view 3D object detection systems often struggle with generating precise predictions due to the challenges in estimating depth from images, increasing redundant and incorrect detections. Our paper presents Ray Denoising, an innovative method that enhances detection accuracy by strategically sampling along camera rays to construct hard negative examples. These examples, visually challenging to differentiate from true positives, compel the model to learn depth-aware features, thereby improving its capacity to distinguish between true and false positives. Ray Denoising is designed as a plug-and-play module, compatible with any DETR-style multi-view 3D detectors, and it only minimally increases training computational costs without affecting inference speed. Our comprehensive experiments, including detailed ablation studies, consistently demonstrate that Ray Denoising outperforms strong baselines across multiple datasets. It achieves a $1.9\%$ improvement in mean Average Precision (mAP) over the state-of-the-art StreamPETR method on the NuScenes dataset. It shows significant performance gains on the Argoverse 2 dataset, highlighting its generalization capability. The code is available at https://github.com/LiewFeng/RayDN. + +Keywords: Multi-view 3D Object Detection $\cdot$ Depth-aware Hard Negative Sampling $\cdot$ Ray Denoising + +# 1 Introduction + +3D object detection is a crucial component in autonomous driving systems, drawing considerable interest from the computer vision community. The field of image-based 3D object detection [9,10,13,14,37,38] is experiencing a surge in research due to its cost-effectiveness compared to LiDAR-based solutions. A key challenge in multi-view 3D object detection, which relies on images from surrounding cameras, is the difficulty in estimating depth from images, leading to duplicate predictions, as shown in Figure 1. + +![](images/a10894823f689f66f34a796554925a1b51b5b203014ba9f4f1e5bccbbcaa396b.jpg) +Fig. 1: The challenge of estimating depth from images in multi-view 3D object detection leads to duplicate predictions and false positive detections along camera rays. Best viewed in color. + +Despite the methodological improvements, multi-view 3D object detectors struggle to reduce false positive predictions arising from depth ambiguities. Several recent studies [8, 15, 17, 19, 22, 26, 34, 41, 47] have attempted to tackle this issue by incorporating temporal information. However, these methods do not explicitly account for the 3D structure of the scene, which limits their potential for further enhancement. + +Furthermore, previous works have explored applying general techniques such as Non-Maximum Suppression (NMS) and Focal Loss to mitigate duplicate predictions. NMS, a post-processing technique, targets false positive predictions with high Intersection over Union (IoU) but is less effective when these predictions are scattered along rays with low IoU. Focal Loss, a loss function designed to reduce high-confidence false positive predictions, has also been implemented. However, it has been observed that multi-view 3D object detectors using Focal Loss still face challenges in effectively resolving the problem of false positive predictions along the same rays. + +Our quantitative analysis highlighted the importance of addressing false positive predictions along the same ray as Ground Truth. By utilizing the precise positional data of Ground Truth objects, we could identify and eliminate these redundant predictions within the state-of-the-art StreamPETR method [34]. This process significantly enhanced $5.4\%$ in mean Average Precision (mAP), emphasizing the critical need for models to improve their depth estimation capabilities. The substantial improvement underscores the potential of refining depth estimation to suppress these predictions and enhance overall detection performance. + +Our key observation is that false positives often occur along camera rays due to the inherent limitation in conventional multi-view object detectors. Since the depth information for each pixel is not accurately estimated, the position embedding can only encode the ray direction. As a result, queries on the same ray will consistently interact with identical visual features from the image, leading to numerous duplicate predictions (false positives) along that ray. This scenario + +underscores the model's need to learn depth-aware features that can discern objects in depth despite the visual features being the same for objects along the same ray. We propose a novel method called Ray Denoising (i.e., RayDN). This framework is inherently flexible and does not limit the choice of distribution for sampling depth-aware hard negative samples. Based on our ablation studies, we have chosen the Beta distribution for its effectiveness in capturing the spatial distribution of false positives that models are likely to generate. This choice enables Ray Denoising to create depth-aware hard negative samples used for denoising, thereby enhancing the model's ability to learn more robust features and representations for distinguishing false positives along the ray, as depicted in Figure 2. Ray Denoising introduces only a marginal increase in computational costs during the training stage without affecting inference speed. + +To summarize our contributions, we highlight the following key points. Firstly, we have identified the persistent challenge of false positive predictions along the same ray, which acts as a bottleneck in the performance of multi-view 3D object detectors. Secondly, we introduce Ray Denoising, a novel denoising method that utilizes the Beta distribution to create depth-aware hard negative samples along rays. This method explicitly takes into account the 3D structure of the scene, offering a flexible solution compatible with any DETR-style multi-view 3D detector to address the issue of duplicate predictions along rays. Lastly, our method achieves state-of-the-art results on the NuScenes dataset [2], significantly enhancing the performance of multi-view 3D object detectors. Specifically, we have improved upon the current state-of-the-art method, StreamPETR, by $1.9\%$ in mean Average Precision (mAP), thereby demonstrating the effectiveness of Ray Denoising. + +# 2 Related Work + +# 2.1 Image-based 3D Object Detection + +3D object detection is a cornerstone task in autonomous driving systems. The domain has witnessed notable advancements in monocular 3D object detection, largely attributed to the KITTI benchmark [5]. This progress has been driven by a range of research studies [1,24,28,31,35,44]. Nevertheless, monocular systems face limitations due to their dependence on a single viewpoint and limited data, which hampers their capacity to manage complex scenarios. To overcome these limitations, extensive benchmarks [2,32,39] have been developed, offering multi-viewpoint data that enriches the field of multi-view 3D object detection and propels the evolution of advanced detection methodologies. + +The multi-view 3D object detection research is primarily bifurcated into dense Bird's Eye View (BEV)-based and sparse query-based algorithms. Our work is categorized under the sparse query-based algorithms. Dense BEV-based algorithms convert multi-view image features into a dense BEV representation using the Lift-Splat-Shoot (LSS) technique [27]. To mitigate overfitting from LSS, BEVDet [9] recommends data augmentation on BEV features. BEVDepth [14] introduces depth estimation supervision to enhance LSS precision. Compared to + +![](images/de98f7c221cf5f88b61b2009104a629912291a786a023b3f94d3030168fa9010.jpg) +Fig. 2: The proposed Ray Denoising approach (right) effectively reduces false positive detections along the ray (highlighted by red rectangles) in the previous state-of-the-art method StreamPETR [34] (left). Best viewed by zooming on the screen. + +LSS-like methods, BEVFormer-like and DETR-like methods can leverage the strong capabilities of transformers for spatial-temporal modeling. BEVFormer-like methods use dense BEV features to enable multiple tasks (e.g., detection and segmentation) but also introduce high computational costs. Sparse query-based algorithms such as DETR3D [37] and PETR [21] utilize learnable 3D object queries and position encoding to engage with multi-view 2D features. DAT [43] introduces a pipeline to construct depth-aware position embeddings and performs contrastive learning on BEV features to tackle duplicate predictions. Recent studies [8, 13, 15, 17, 19, 22, 26, 34] have integrated temporal information to refine these methods further. + +# 2.2 Hard Negative Samples Mining + +Hard negative sample mining is a strategy to minimize false positive predictions in object detection. Early works like R-CNN [6] employed standard hard negative mining methods [4, 33], where misclassified samples were utilized for retraining. OHEM [29] and RetinaNet [16] introduced adaptive sampling and focal loss to address the challenges of heuristics and hyper-parameters in hard negative sample selection. These methods select hard negative samples from existing predictions based on confidence scores and matched ground-truth object classes. Our approach takes a different path by constructing new hard negative samples based on the locations of ground-truth objects rather than selecting from existing predictions or depending on prediction confidence and matched ground-truth object classes. + +# 2.3 Denoising in Object Detection + +In 2D object detection, models such as DETR [3] and its variants [20,36,46] often face convergence issues, which are partly due to the instability of bipartite graph matching and the inconsistency in optimization goals during the early stages of + +training. To tackle this, DN-DETR [12] introduces the concept of feeding noisy ground-truth bounding boxes into the Transformer decoder, with the model being trained to reconstruct the original, clean boxes. This process, known as denoising, involves using queries initialized with these noisy ground-truth bounding boxes, referred to as denoising queries. DINO [42] further refines this approach by generating both positive and negative denoising queries for each ground-truth bounding box, with the negative queries incorporating more noise to improve the model's ability to reject incorrect predictions. + +In the realm of 3D object detection, DETR-style methods [19,22,34] perpetuate the denoising approach by creating a denoising query for each ground-truth bounding box. Based on the noise level, these queries are classified as either an object or 'no object'. However, these methods neglect to incorporate knowledge of the 3D scene structure when generating denoising queries. In multi-view 3D object detection, the absence of depth information for each pixel results in models producing duplicate predictions along the same ray with different depths. To combat these false positive predictions, we harness the 3D structure knowledge of the scene. We devise multiple negative ray-denoising queries distributed along the rays, spatially distinct from the traditional negative denoising queries surrounding the ground-truth objects. This innovative design markedly improves the detector's capability to differentiate between true positive predictions and false positive predictions along the same ray. + +# 3 Methodology + +This section elaborates on our proposed method, Ray Denoising, for multi-view 3D object detection. We start with a comprehensive overview of the DETR-style multi-view 3D object detector framework in Section 3.1. Following that, we describe the three pivotal steps for integrating Ray Denoising into the framework to address the challenge of duplicate predictions. Section 3.2 outlines the ray generation process. Section 3.3 explains utilizing the Beta distribution family for sampling reference points. Section 3.4 details the construction of spatial denoising queries, referred to as Ray Queries, along each ray. Lastly, Section 3.5 delves into the essential differences between our Ray Queries and the denoising queries employed in prior works and the influence of Ray Denoising on the model training process. + +# 3.1 Overview + +Ray Denoising is designed to be integrated into any DETR-style multi-view 3D object detector [15, 17, 19, 21, 22, 34, 37]. These detectors typically consist of a convolutional network-based feature encoder and a transformer-based decoder. The process begins by feeding $N$ surround-view images $\pmb{I} = \{I_i\in \mathbb{R}^{3\times H_I\times W_I},i = 1,2,\dots,N\}$ into the feature encoder (e.g., ResNet101 [7]) to extract image features $\pmb{F} = \{F_{i}\in \mathbb{R}^{C\times H_{F}\times W_{F}},i = 1,2,\dots,N\}$ . Here, $H_{I}$ and $W_{I}$ denote the image dimensions, $H_{F}$ and $W_{F}$ are the feature dimensions, and $C$ represents the + +![](images/88714622b6b3ea899694bc720348b5a8c1f9b783061227f1d00f4ecae45bf6ba.jpg) +Fig. 3: Overall framework of the Ray Denoising approach, a plug-and-play training technique for DETR-style multi-view 3D object detectors, focuses on refining the model's ability to distinguish true positives from false positives in depth. Casting rays and sampling depth-aware denoising queries effectively tackle the challenge of false positives arising from the inherent difficulties in visually estimating depth, leading to substantial improvements in detection performance over strong baselines. Best viewed in color and by zooming on the screen. + +number of channels in the feature. To facilitate 3D perception, multiple points within each camera's frustum are transformed and encoded into position embeddings [21]. These embeddings enable the multi-view image features to interact with 3D queries. + +The final step involves $N$ object queries $\mathbf{Q} \in \mathbb{R}^{N \times 256}$ , which are derived from a set of learnable 3D reference points $\mathbf{P} \in \mathbb{R}^{N \times 3}$ . These queries interact with the multi-view image features $\mathbf{F}$ in the transformer decoder, employing multi-layer cross-attention to identify objects. The decoder's output features are then processed by prediction heads (multi-layer perceptron, i.e., MLP) to yield classification scores $(cls)$ , position offsets $(x, y, z)$ , scales $(w, h, l)$ , directions $(\theta_x, \theta_y)$ , and velocities $(v_x, v_y)$ . + +# 3.2 Ray Casting + +The reference points for our ray denoising queries are distributed along the camera ray, which extends from the camera's optical center to the ground-truth object on the image plane. To establish this ray, we project the 3D center of the ground-truth object into the camera frustum space using the following transformation: + +$$ +\mathbf {C} ^ {\prime} = \mathbf {K} \cdot \mathbf {C} _ {G T}, \tag {1} +$$ + +where $\mathbf{K}$ is the $4\times 4$ transformation matrix that maps points from 3D world space to camera frustum space. $\mathbf{C}_{GT} = (x,y,z,1)$ is the center of the groundtruth object in 3D space, and $\mathbf{C}' = (u\times d,v\times d,d,1)$ is the corresponding + +![](images/4c4a2d14581946113ef63f596cdc1d87267b0add529ac47c993cac2426c793e4.jpg) +(a) Distribution Comparison + +![](images/22781dfede13b14d9e9f299c4676acdea843872bb5f1c03fda5ed7ddba61e0fe.jpg) +(b) Beta distribution family +Fig. 4: (a) Distribution comparison showing that the Beta distribution is bounded between -1 and 1, unlike the Laplace and Gaussian distributions, which are unbounded. (b) The Beta distribution family, with the x-range adjusted from [0, 1] to $[-1, 1]$ using the transformation $y = 2x - 1$ . Best viewed in color. + +projected center in the camera frustum, with $(u,v)$ being the pixel coordinates and $d$ the depth value. + +With the depth of the ground-truth object's center known, we can determine the coordinates of valid reference points along the ray as follows: + +$$ +\hat {d} _ {i} = d + \beta_ {i} \cdot f \left(\mathbf {S} _ {G T}\right), \tag {2} +$$ + +where $f$ is a function that encodes the average scale of the ground-truth object, calculated as $f(\mathbf{S}_{GT}) = k \cdot \frac{w + h + l}{6}$ . The radius $k$ defines the valid distribution range for reference points, and $\beta_{i}$ is the offset for the $i$ -th reference point. + +The position of the reference point is then obtained by: + +$$ +\hat {\mathbf {P}} _ {i} = \mathbf {K} ^ {- 1} \cdot \hat {\mathbf {C}} _ {i} ^ {\prime}, \tag {3} +$$ + +where $\hat{\mathbf{C}}_i^{\prime} = (u\times \hat{d}_i,v\times \hat{d}_i,\hat{d}_i,1)$ + +# 3.3 Sample Generation + +With the valid reference points along the rays defined, we proceed to sample $N$ reference points to emulate the distribution of false positive predictions that arise from depth ambiguities. A straightforward uniform distribution has proven effective, yielding a $2.2\%$ mAP improvement as depicted in Figure 4a. To more accurately simulate the distribution of false positives, we consider two key aspects. First, we assess whether models tend to predict objects as being closer or further away from the ego-vehicle relative to the actual positions of the ground-truth objects. Second, we examine whether models are more likely to predict objects near or far from the center of the ground-truth objects. + +Distributions like the Beta, Gaussian, and Laplace are suitable for handling complex scenarios, with the Gaussian and Laplace being unbounded. However, + +samples with significant offsets from the ground truth can result in large regression losses, negatively impacting optimization. As shown in Figure 4a, the Beta distribution slightly outperforms the others. Notably, the uniform distribution is a special case within the Beta family. For a versatile sampling strategy applicable to all scenarios, we select the Beta distribution family, characterized by the probability density function: + +$$ +p (x | \lambda , \mu) = \frac {\Gamma (\lambda + \mu)}{\Gamma (\lambda) + \Gamma (\mu)} x ^ {\lambda - 1} (1 - x) ^ {\mu - 1}, \tag {4} +$$ + +where $\lambda$ and $\mu$ are hyper-parameters that shape the distribution. The Gamma function, denoted as $\Gamma(x)$ , is defined as: + +$$ +\Gamma (x) = \int_ {0} ^ {\infty} t ^ {x - 1} e ^ {- t} d t. \tag {5} +$$ + +The beta distribution's probability density function (PDF), visualized in Figure 4b, reveals its behavior under different parameter settings. When $\lambda$ equals $\mu$ , the distribution is symmetric, reflecting a scenario where the model predicts objects with equal likelihood at closer or further distances. For instance, with $\lambda = \mu = 1$ , the distribution simplifies to a uniform distribution. Adjusting $\lambda$ and $\mu$ to be higher shifts the reference points of the denoising queries closer to the ground-truth object's center and vice versa. + +When $\lambda$ exceeds $\mu$ , such as in the case of $\lambda = 8$ and $\mu = 2$ , the distribution suggests that models are more likely to predict objects further away from the ego-vehicle. Conversely, when $\lambda$ is less than $\mu$ , for example, $\lambda = 2$ and $\mu = 8$ , the distribution indicates a propensity for models to predict objects closer to the ego-vehicle. This flexibility in the Beta distribution allows for tailored sampling strategies that can adapt to the specific challenges of depth estimation in multiview 3D object detection. + +# 3.4 Query Denoising + +We construct ray denoising queries from the sampled $N$ reference points. Specifically, a multi-layer perceptron (MLP) is utilized to project the normalized 3D coordinates of each point into a latent feature space. These $N$ ray denoising queries $\pmb{q}_r$ are combined with the learnable object queries $\pmb{q}_o$ from the baseline detector and are input into the transformer decoder. An attention mask is employed during the self-attention operations to prevent information leakage from the ray denoising queries to the object queries. The query closest to the center of the ground truth (GT) object is designated as the positive query, inheriting the class label of the corresponding GT object. The remaining $N - 1$ queries are labeled as background, i.e., 'no object'. The loss for ray denoising queries adheres to the same criteria as for the learnable object queries, using Focal Loss for classification and L1 Loss for regression. This process ensures that the model is effectively trained to discern between true positives and false positives along the same ray. + +# 3.5 Discussion + +Our Ray Denoising approach is based on the pivotal observation that image-based 3D detection systems often struggle to distinguish true positives from false positives along camera rays. DETR-style multi-view 3D object detectors implicitly learn depth estimation from ground truth supervision. However, the randomly distributed reference points of learnable queries do not fully leverage the available ground truth information. While these reference points are updated during training, they fail to provide adequate hard negative samples for each object in every scene. To enhance the utilization of ground truth information, traditional denoising techniques introduce additional reference points distributed uniformly around the ground-truth object during training. These instance-specific reference points have improved detection performance [19,22,34]. However, they overlook the depth ambiguities intrinsic to multi-view 3D object detection. The absence of precise depth information for each pixel means that the position embedding can only encode ray direction, not depth. This leads to queries on the same ray interacting with the same image features, resulting in redundant predictions. Ray Denoising diverges from traditional techniques by strategically generating reference points along rays that extend from cameras to objects. This approach explicitly considers the 3D structure of each object in the scene, providing a sufficient number of hard negative samples. During training, these Ray Queries interact within the self-attention layer, effectively guiding the model to suppress depth-ambiguous spatial hard negative samples. This interaction enhances the detector's ability to differentiate between true positive predictions (objects) and false positive predictions (duplicate detections), improving detection accuracy. + +# 4 Experiment + +# 4.1 Dataset and Metrics + +Our model's performance was assessed using two datasets: nuScenes [2] and Argovverse 2 [39]. The nuScenes dataset include 1000 video sequences, which are split into training (700 videos), validation (150 videos), and testing (150 videos) sets, each approximately 20 seconds long with annotations at 0.5-second intervals. The dataset comprises 1.4 million annotated 3D bounding boxes across ten object classes. Evaluation metrics include the mean Average Precision (mAP) and five true positive metrics: ATE, ASE, AOE, AVE, and AAE, which assess translation, scale, orientation, velocity, and attribute errors. The nuScenes Detection Score (NDS) is a comprehensive score derived from these metrics, offering an overall performance evaluation. + +The Argoverse 2 dataset contains 1000 unique scenes, each 15 seconds long, annotated at $10\mathrm{Hz}$ . The scenes are divided into 700 for training, 150 for validation, and 150 for testing. The evaluation covers 26 categories within a 150-meter range, addressing long-range perception tasks. Metrics include the mAP and the Composite Detection Score (CDS), which integrates three other true positive metrics: ATE, ASE, and AOE. + +Table 1: Comparison on the nuScenes validation set. $\dagger$ Indicates methods that benefit from perspective-view pre-training. + +
MethodsBackboneImage SizemAPNDSmATEmASEmAOEmAVEmAAE
BevDet4D [8]ResNet50256×70432.245.70.7030.2780.4950.3540.206
PETRv2 [22]ResNet50256×70434.945.60.7000.2750.5800.4370.187
BEVDepth [14]ResNet50256×70435.147.50.6290.2670.4790.4280.198
BEVStereo [13]ResNet50256×70437.250.00.5980.2700.4380.3670.190
BEVFomerv2 [41]†ResNet50-42.352.90.6180.2730.4130.3330.188
SOLOFusion [26]ResNet50256×70442.753.40.5670.2740.5110.2520.181
SparseBEV [19]†ResNet50256×70444.855.80.5810.2710.3730.2470.190
StreamPETR [34]†ResNet50256×70445.055.00.6130.2670.4130.2650.198
RayDN† (Ours)ResNet50256×70446.956.30.5790.2640.4330.2560.187
BEVDepth [14]ResNet101512×140841.253.50.5650.2660.3580.3310.190
PETRv2 [22]†ResNet101640×160042.152.40.6810.2670.3570.3770.186
Sparse4D [17]†ResNet101900×160043.654.10.6330.2790.3630.3170.177
SOLOFusion [26]ResNet101512×140848.358.20.5030.2640.3810.2460.207
SparseBEV [19]†ResNet101512×140850.159.20.5620.2650.3210.2430.195
StreamPETR [34]†ResNet101512×140850.459.20.5690.2620.3150.2570.199
RayDN† (Ours)ResNet101512×140851.860.40.5410.2600.3150.2360.200
+ +Table 2: Comparison on the nuScenes test set. + +
MethodsBackboneImage SizemAPNDSmATEmASEmAOEmAVEmAAE
DETR3D [37]V2-99900×160041.247.90.6410.2550.3940.8450.133
MV2D [38]V2-99640×160046.351.40.5420.2470.4030.8570.127
BEVFormer [15]V2-99900×160048.156.90.5820.2560.3750.3780.126
PETRv2 [22]V2-99640×160049.058.20.5610.2430.3610.3430.120
PolarFormer [10]V2-99900×160049.357.20.5560.2560.3640.4390.127
BEVStereo [13]V2-99900×160052.561.00.4310.2460.3580.3570.138
HoP [47]V2-99640×160052.861.20.4910.2420.3320.3430.109
SparseBEV [19]V2-99640×160054.362.70.5020.2440.3240.2510.126
StreamPETR [34]V2-99640×160055.063.60.4790.2390.3170.2410.119
RayDN (Ours)V2-99640×160056.564.50.4610.2410.3220.2390.114
+ +# 4.2 Implementation Details + +Our experimental setup leverages ResNet50 [7], ResNet101, and V2-99 [11] backbones, each with distinct pre-training configurations. We detail the performance of ResNet50 and ResNet101 models, which are pre-trained on nuImages [2], on the nuScenes validation set. To showcase the scalability of our approach, we also report results on the nuScenes test set using V2-99, initialized from the DD3D [25] checkpoint. Our models are optimized using the AdamW [23] optimizer with a batch size of 16. The learning rate is set to $4 \times 10^{-4}$ for models trained only on the training set and $3 \times 10^{-4}$ for those trained on both the training and validation sets. We adopt a cosine annealing policy for learning rate scheduling. For benchmarking against state-of-the-art (SOTA) methods, models are trained for 60 epochs without CBGS [45] and for 24 epochs in ablation studies in nuScenes. For experiments on Argoverse 2, models are trained for 6 epochs. Our implementation is primarily based on the StreamPETR [34] framework. + +Table 3: Comparisons on the Argoverse 2 validation set. We evaluate across 26 object categories within a range of 150 meters. + +
MethodsBackboneImage SizemAPCDSmATEmASEmAOE
PETR [21]V2-99900×64017.612.20.9110.3390.819
Sparse4Dv2 [18]V2-99900×64018.913.40.8320.3430.723
StreamPETR [34]V2-99900×64020.314.60.8430.3210.650
RayDN (Ours)V2-99900×64022.316.10.8250.3250.629
+ +# 4.3 Comparison with State-of-the-Art Methods + +We compare the proposed Ray Denoising method with other state-of-the-art multi-view 3D object detectors on the validation and test sets of the nuScenes dataset, as well as the validation set of the Argoverse 2 dataset. It's important to note that our method does not employ test time augmentation (TTA). + +nuScenes Validation Set. Table 1 presents a comparison with state-of-the-art methods on the nuScenes validation set. We evaluate both ResNet-50 and ResNet-101 backbones. With ResNet-50 as the backbone and an image size of $704 \times 256$ , we achieve $46.9\%$ mAP and $56.3\%$ NDS, improving upon the previous state-of-the-art method, StreamPETR, by $1.9\%$ mAP and $1.3\%$ NDS. Using a more robust ResNet-101 backbone and increasing the image size to $512 \times 1408$ , our performance reaches $51.6\%$ mAP and $59.8\%$ NDS, outperforming StreamPETR by $1.2\%$ mAP and $0.6\%$ NDS. These results demonstrate the scalability of Ray Denoising. + +nuScenes Test Set. The results evaluated by the test server are detailed in Table 2. Our method includes training on both the training and validation sets. Notably, we achieve $56.5\%$ mAP and $64.5\%$ NDS, surpassing StreamPETR by an absolute $1.5\%$ mAP and $0.9\%$ NDS. + +Argoverse 2 Validation Set. To assess the generalization capability of Ray Denoising, we conduct additional experiments on the Argoverse 2 dataset, as displayed in Table 3. Our method significantly outperforms previous state-of-the-art methods, achieving a $2.0\%$ mAP and $1.5\%$ CDS improvement on the validation set. These metrics underscore the generalization capability of our approach. + +The tables demonstrate that our Ray Denoising approach notably enhances mAP. Considering that mAP is significantly impacted by incorrect false positives, this enhancement firmly validates the effectiveness of Ray Denoising in reducing redundant predictions along the camera rays. + +# 4.4 Ablation Study + +This section delves into the ablation studies performed using the validation sets from the nuScenes and Argoverse 2 datasets. Unless specified otherwise, our experiments on nuScenes are based on a ResNet50 backbone, which is pre-trained on the nuImages dataset [2]. Our input data comprises 8 frames, each with a $704 \times 256$ pixels resolution. The decoder utilizes 428 queries, and the model is trained for 24 epochs, forgoing Class-Balanced Group Sampling (CBGS) [45]. + +Table 4: Ablation studies on the radius of ray denoising queries. + +
MethodRadiusmAPNDS
SOTA Baseline [34]-39.148.6
241.149.8
+RayDN (Ours)341.950.8
441.350.1
+ +Table 6: Ablation studies on the distribution of ray denoising queries on nuScenes dataset. + +
Methodλ μ mAPNDS
SOTA Baseline [34]-39.148.6
+RayDN (Ours)141.350.4
241.550.1
341.350.5
741.750.7
841.950.8
941.650.4
+ +Table 5: Ablation studies on the number of ray denoising queries. + +
Method#QmAPNDS
SOTA Baseline [34]-39.148.6
341.650.2
+RayDN (Ours)541.950.8
741.150.5
+ +Table 7: Ablation studies on the distribution of ray denoising queries for the Argoverse 2 dataset. + +
Methodλ μmAPCDS
SOTA Baseline [34]-20.314.6
+RayDN (Ours)121.315.4
321.715.8
421.315.3
621.515.7
722.316.1
821.515.6
+ +Radius of Ray Denoising Queries. Table 4 shows how varying the radius $k$ affects performance. We find that a radius of $k = 3$ yields the optimal results, while $k = 2$ and $k = 4$ also lead to notable enhancements in performance. + +Number of Ray Denoising Queries. We explore the effect of the quantity of ray denoising queries on the model's performance in Table 5. The findings reveal that both mAP and NDS metrics improve with increasing ray queries, reaching a plateau of 5 queries. Surprisingly, upping the count to 7 queries results in a dip in mAP, which might be attributed to an imbalance in the ratio of positive to negative ray denoising queries. + +Distribution of Ray Denoising Queries. We investigate the impact of the distribution of ray denoising queries with various hyper-parameters across the nuScenes and Argoverse 2 datasets. The results are detailed in Table 6 and Table 7. Our findings indicate that employing a uniform distribution, which is dataset-agnostic (i.e., setting $\lambda = 1$ and $\mu = 1$ ), leads to substantial performance gains for Ray Denoising. Specifically, we observe a $2.2\%$ increase in mAP for the nuScenes dataset and a $1.0\%$ increase in mAP for the Argoverse 2 dataset, highlighting the method's strong generalization capabilities. Fine-tuning these hyper-parameters further enhances the performance, with an additional $0.6\%$ mAP improvement on nuScenes and $1.0\%$ mAP on Argoverse. Across both datasets, models configured with $\lambda > \mu$ slightly outperform those with $\lambda < \mu$ , suggesting that false positives tend to be at greater distance from the ego-vehicle. + +Precision-Recall Analysis. We assess the ability of Ray Denoising to reduce false positive predictions by plotting precision-recall curves and Average Precision (AP) for each object class. Figure 5a reveals that Ray Denoising enhances precision across a range of recall levels and under various distance thresholds. + +![](images/d0b91b7a89a031bb1770ad96eab0df73bee095aefb6cb0ea81b5b7748c9927cd.jpg) +(a) Precision-Recall Curve + +![](images/245734db36b9a630fd1de0330c5d8c048ad2804fd257f9c15985555fe8cb0374.jpg) +(b) AP for each class + +![](images/67bbe4937d086929e32c83c727dd7a8a10923f3469b49e05c6cd55054de44722.jpg) +Fig. 5: (a) Visualization of the precision-recall curves at various distance thresholds. Ray Denoising consistently enhances precision across nearly all recall levels, effectively suppressing false positives. (b) Class-wise AP comparison. Ray Denoising performs superior over the SOTA StreamPETR in all object classes. Best viewed in color. +Fig.6: Visualization of our detection results on the validation set of nuScenes. Ray Denoising effectively mitigates duplicate false positives while maintaining the ability to detect highly occluded objects along the same ray. Best viewed on the screen. + +Similarly, Figure 5b shows that Ray Denoising surpasses StreamPETR in AP across all classes. These results collectively affirm that Ray Denoising effectively mitigates the occurrence of false positive predictions. + +Visualization of Detection Results. Our approach creates multiple spatial hard negative samples along the same ray, potentially interfering with detecting true positives if they are also located along that ray. To test this, we present a visualization of the detection outcomes in Figure 6. The results show that Ray Denoising effectively eliminates duplicate false positives without compromising the detection of true positives, even when they are heavily occluded along the same ray. + +Generalization Ability of Ray Denoising. To assess the versatility of Ray Denoising, we apply it to different models with varied input data. We use the state-of-the-art CMT [40] as a reference and test Ray Denoising with single-frame images and multi-modal data. In the single-frame scenario, we employ a ResNet50 backbone pre-trained on ImageNet [2], with input images sized at $256 \times 704$ pixels. For the multi-modal setup, the input image resolution is $320 \times 800$ pixels, and the voxel dimensions are $0.1\mathrm{m} \times 0.1\mathrm{m}$ . The batch size is set to 16 due to memory constraints, and all models are trained for 20 epochs using CBGS [45]. Table 8 shows that Ray Denoising improves upon the baseline + +Table 8: Ablation studies on generalization ability of Ray Denoising. + +
MethodDescriptionmAPNDS
CMT [40]single frame30.637.7
+ RayDNsingle frame32.139.1
CMT [40]multi-modal67.670.4
+ RayDNmulti-modal68.671.3
+ +Table 9: Ablation studies on the effect of depth estimation. + +
3DPPERayDNmAPNDS
39.148.6
✓42.051.1
✓41.950.8
✓✓43.952.9
+ +with a $1.5\%$ increase in mAP and a $1.4\%$ increase in NDS for the single-frame setup. Moreover, it achieves a $1.0\%$ mAP and a $0.9\%$ NDS enhancement on the already high-performing multi-modal baseline. These experiments confirm the broad applicability of Ray Denoising. + +Effect of Depth Estimation on Ray Denoising. In DETR-style multi-view 3D object detectors, the lack of depth information leads to the use of camera rays as position embeddings for image features, which can be imprecise. To refine these embeddings, 3DPPE [30] suggests encoding an estimated 3D point guided by external supervision for depth estimation. To explore the role of depth estimation in conjunction with Ray Denoising, we integrate 3DPPE into our framework. The results, detailed in Table 9, show a marked improvement in performance with 3DPPE. However, Ray Denoising alone, without external depth supervision, also achieves similar performance. This underscores our core insight: Ray Denoising enhances the model's depth perception by guiding it to learn superior depth perception ability during training. When Ray Denoising is paired with 3DPPE, we further obtain a $+1.9\%$ increase in mAP and a $+1.7\%$ increase in NDS. This demonstrates that while depth estimation algorithms can improve the clarity of position embeddings, Ray Denoising's ability to refine the model's understanding of depth through robust training is a powerful complement, even when depth information is available. + +# 5 Conclusion + +We introduce Ray Denoising, a method designed to overcome the critical challenge of depth estimation in multi-view 3D object detection. Ray Denoising tackles the issue of false detections along camera rays, which are a direct consequence of imprecise depth information from images. By leveraging the 3D structure of the scene, Ray Denoising prompts the model to learn depth-aware features, leading to improved differentiation between false and true positives along the same ray without introducing extra inference costs. Our comprehensive experiments on the NuScenes and Argoverse 2 datasets demonstrate that Ray Denoising consistently and significantly outperforms strong baselines, achieving new state-of-the-art performance in Multi-view 3D Object Detection. Acknowledgment. This work was supported by the Fundamental Research Funds for the Central Universities (E3E41903, E2ET1104, E3ET6201X2), the National Natural Science Foundation of China (NSFC) under Grant 62225208 and 62171431. + +# References + +1. Brazil, G., Liu, X.: M3d-rpn: Monocular 3d region proposal network for object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9287-9296 (2019) +2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020) +3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020) +4. Felzenszwalb, P.F., Girshick, R.B., McAllester, D., Ramanan, D.: Object detection with discriminatively trained part-based models. IEEE transactions on pattern analysis and machine intelligence 32(9), 1627-1645 (2009) +5. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset. The International Journal of Robotics Research 32(11), 1231-1237 (2013) +6. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Region-based convolutional networks for accurate object detection and segmentation. IEEE transactions on pattern analysis and machine intelligence 38(1), 142-158 (2015) +7. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +8. Huang, J., Huang, G.: Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054 (2022) +9. Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multicamera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790 (2021) +0. Jiang, Y., Zhang, L., Miao, Z., Zhu, X., Gao, J., Hu, W., Jiang, Y.G.: Polarformer: Multi-camera 3d object detection with polar transformer. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1042-1050 (2023) +1. Lee, Y., Hwang, J.w., Lee, S., Bae, Y., Park, J.: An energy andgpu-computation efficient backbone network for real-time object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 0-0 (2019) +2. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13619-13627 (2022) +3. Li, Y., Bao, H., Ge, Z., Yang, J., Sun, J., Li, Z.: Bevstereo: Enhancing depth estimation in multi-view 3d object detection with temporal stereo. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1486-1494 (2023) +4. Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1477-1485 (2023) +5. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: European conference on computer vision. pp. 1-18. Springer (2022) +6. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017) + +17. Lin, X., Lin, T., Pei, Z., Huang, L., Su, Z.: Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion. arXiv preprint arXiv:2211.10581 (2022) +18. Lin, X., Lin, T., Pei, Z., Huang, L., Su, Z.: Sparse4d v2: Recurrent temporal fusion with sparse model. arXiv preprint arXiv:2305.14018 (2023) +19. Liu, H., Teng, Y., Lu, T., Wang, H., Wang, L.: Sparsebev: High-performance sparse 3d object detection from multi-camera videos. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18580-18590 (2023) +20. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329 (2022) +21. Liu, Y., Wang, T., Zhang, X., Sun, J.: Petr: Position embedding transformation for multi-view 3d object detection. In: European Conference on Computer Vision. pp. 531-548. Springer (2022) +22. Liu, Y., Yan, J., Jia, F., Li, S., Gao, A., Wang, T., Zhang, X.: Petrov2: A unified framework for 3d perception from multi-camera images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3262-3272 (2023) +23. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2018) +24. Ma, X., Zhang, Y., Xu, D., Zhou, D., Yi, S., Li, H., Ouyang, W.: Delving into localization errors for monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4721-4730 (2021) +25. Park, D., Ambrus, R., Guizilini, V., Li, J., Gaidon, A.: Is pseudo-lidar needed for monocular 3d object detection? In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3142-3152 (2021) +26. Park, J., Xu, C., Yang, S., Keutzer, K., Kitani, K.M., Tomizuka, M., Zhan, W.: Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. In: The Eleventh International Conference on Learning Representations (2022) +27. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 194-210. Springer (2020) +28. Reading, C., Harakeh, A., Chae, J., Waslander, S.L.: Categorical depth distribution network for monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8555-8564 (2021) +29. Shrivastava, A., Gupta, A., Girshick, R.: Training region-based object detectors with online hard example mining. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 761-769 (2016) +30. Shu, C., Deng, J., Yu, F., Liu, Y.: 3dppe: 3d point positional encoding for transformer-based multi-camera 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 3580-3589 (October 2023) +31. Simonelli, A., Bulo, S.R., Porzi, L., López-Antequera, M., Kontschieder, P.: Disentangling monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1991-1999 (2019) +32. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2446-2454 (2020) + +33. Sung, K.K., Poggio, T.: Example-based learning for view-based human face detection. IEEE Transactions on pattern analysis and machine intelligence 20(1), 39-51 (1998) +34. Wang, S., Liu, Y., Wang, T., Li, Y., Zhang, X.: Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 3621-3631 (October 2023) +35. Wang, T., Zhu, X., Pang, J., Lin, D.: Fcos3d: Fully convolutional one-stage monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 913-922 (2021) +36. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor detr: Query design for transformer-based detector. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 2567-2575 (2022) +37. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: Conference on Robot Learning. pp. 180–191. PMLR (2022) +38. Wang, Z., Huang, Z., Fu, J., Wang, N., Liu, S.: Object as query: Lifting any 2d object detector to 3d detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3791-3800 (2023) +39. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., et al.: Argoverse 2: Next generation datasets for self-driving perception and forecasting. In: Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2) (2021) +40. Yan, J., Liu, Y., Sun, J., Jia, F., Li, S., Wang, T., Zhang, X.: Cross modal transformer: Towards fast and robust 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18268-18278 (2023) +41. Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z., Huang, G., Li, H., Qiao, Y., Lu, L., et al.: Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17830-17839 (2023) +42. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In: The Eleventh International Conference on Learning Representations (2022) +43. Zhang, H., Li, H., Zeng, A., Li, F., Liu, S., Liao, X., Zhang, L.: Introducing depth into transformer-based 3d object detection. arXiv preprint arXiv:2302.13002 (2023) +44. Zhang, Y., Lu, J., Zhou, J.: Objects are different: Flexible monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3289-3298 (2021) +45. Zhu, B., Jiang, Z., Zhou, X., Li, Z., Yu, G.: Class-balanced grouping and sampling for point cloud 3d object detection. arXiv preprint arXiv:1908.09492 (2019) +46. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020) +47. Zong, Z., Jiang, D., Song, G., Xue, Z., Su, J., Li, H., Liu, Y.: Temporal enhanced training of multi-view 3d object detector via historical object prediction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 3781-3790 (October 2023) \ No newline at end of file diff --git a/2024/Ray Denoising_ Depth-aware Hard Negative Sampling for Multi-view 3D Object Detection/images.zip b/2024/Ray Denoising_ Depth-aware Hard Negative Sampling for Multi-view 3D Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4c623d61b7be703c8ee88cc4c59bc146437d127b --- /dev/null +++ b/2024/Ray Denoising_ Depth-aware Hard Negative Sampling for Multi-view 3D Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3adecef84da97e47f35cb25993907df576a69c75b34840f12d46569b7b7ec3d2 +size 582736 diff --git a/2024/Ray Denoising_ Depth-aware Hard Negative Sampling for Multi-view 3D Object Detection/layout.json b/2024/Ray Denoising_ Depth-aware Hard Negative Sampling for Multi-view 3D Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6b14d509ad4d17ed225862d685dac9f7a62ae956 --- /dev/null +++ b/2024/Ray Denoising_ Depth-aware Hard Negative Sampling for Multi-view 3D Object Detection/layout.json @@ -0,0 +1,8964 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 111, + 467, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 111, + 467, + 148 + ], + "spans": [ + { + "bbox": [ + 146, + 111, + 467, + 148 + ], + "type": "text", + "content": "Ray Denoising: Depth-aware Hard Negative Sampling for Multi-view 3D Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 168, + 481, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 168, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 168, + 481, + 194 + ], + "type": "text", + "content": "Feng Liu1* Tengteng Huang2 Qianjing Zhang2 Haotian Yao2 Chi Zhang2 Fang Wan1 Qixiang Ye1 Yanzhao Zhou1†" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 146, + 201, + 465, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 201, + 465, + 224 + ], + "spans": [ + { + "bbox": [ + 146, + 201, + 465, + 224 + ], + "type": "text", + "content": "1 University of Chinese Academy of Sciences liufeng20@mails.ucas.ac.cn {wanfang,qxye,zhouyanzhao}@ucas.ac.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 224, + 467, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 224, + 467, + 246 + ], + "spans": [ + { + "bbox": [ + 146, + 224, + 467, + 246 + ], + "type": "text", + "content": "2 Mach Drive {tengteng.huang,qianjing.zhang,haotian.yao,chi.zhang}@mach-drive.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 280, + 452, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 280, + 452, + 477 + ], + "spans": [ + { + "bbox": [ + 159, + 280, + 452, + 477 + ], + "type": "text", + "content": "Abstract. Multi-view 3D object detection systems often struggle with generating precise predictions due to the challenges in estimating depth from images, increasing redundant and incorrect detections. Our paper presents Ray Denoising, an innovative method that enhances detection accuracy by strategically sampling along camera rays to construct hard negative examples. These examples, visually challenging to differentiate from true positives, compel the model to learn depth-aware features, thereby improving its capacity to distinguish between true and false positives. Ray Denoising is designed as a plug-and-play module, compatible with any DETR-style multi-view 3D detectors, and it only minimally increases training computational costs without affecting inference speed. Our comprehensive experiments, including detailed ablation studies, consistently demonstrate that Ray Denoising outperforms strong baselines across multiple datasets. It achieves a " + }, + { + "bbox": [ + 159, + 280, + 452, + 477 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 159, + 280, + 452, + 477 + ], + "type": "text", + "content": " improvement in mean Average Precision (mAP) over the state-of-the-art StreamPETR method on the NuScenes dataset. It shows significant performance gains on the Argoverse 2 dataset, highlighting its generalization capability. The code is available at https://github.com/LiewFeng/RayDN." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 488, + 452, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 488, + 452, + 510 + ], + "spans": [ + { + "bbox": [ + 159, + 488, + 452, + 510 + ], + "type": "text", + "content": "Keywords: Multi-view 3D Object Detection " + }, + { + "bbox": [ + 159, + 488, + 452, + 510 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 488, + 452, + 510 + ], + "type": "text", + "content": " Depth-aware Hard Negative Sampling " + }, + { + "bbox": [ + 159, + 488, + 452, + 510 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 488, + 452, + 510 + ], + "type": "text", + "content": " Ray Denoising" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 533, + 230, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 230, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 230, + 547 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 561, + 482, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 561, + 482, + 644 + ], + "spans": [ + { + "bbox": [ + 130, + 561, + 482, + 644 + ], + "type": "text", + "content": "3D object detection is a crucial component in autonomous driving systems, drawing considerable interest from the computer vision community. The field of image-based 3D object detection [9,10,13,14,37,38] is experiencing a surge in research due to its cost-effectiveness compared to LiDAR-based solutions. A key challenge in multi-view 3D object detection, which relies on images from surrounding cameras, is the difficulty in estimating depth from images, leading to duplicate predictions, as shown in Figure 1." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 447, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 447, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 447, + 666 + ], + "type": "text", + "content": "* Work was done during internship at Mach Drive. † Corresponding Author." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 202, + 114, + 413, + 248 + ], + "blocks": [ + { + "bbox": [ + 202, + 114, + 413, + 248 + ], + "lines": [ + { + "bbox": [ + 202, + 114, + 413, + 248 + ], + "spans": [ + { + "bbox": [ + 202, + 114, + 413, + 248 + ], + "type": "image", + "image_path": "a10894823f689f66f34a796554925a1b51b5b203014ba9f4f1e5bccbbcaa396b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 256, + 482, + 289 + ], + "lines": [ + { + "bbox": [ + 132, + 256, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 256, + 482, + 289 + ], + "type": "text", + "content": "Fig. 1: The challenge of estimating depth from images in multi-view 3D object detection leads to duplicate predictions and false positive detections along camera rays. Best viewed in color." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 316, + 482, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 387 + ], + "type": "text", + "content": "Despite the methodological improvements, multi-view 3D object detectors struggle to reduce false positive predictions arising from depth ambiguities. Several recent studies [8, 15, 17, 19, 22, 26, 34, 41, 47] have attempted to tackle this issue by incorporating temporal information. However, these methods do not explicitly account for the 3D structure of the scene, which limits their potential for further enhancement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 388, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 496 + ], + "type": "text", + "content": "Furthermore, previous works have explored applying general techniques such as Non-Maximum Suppression (NMS) and Focal Loss to mitigate duplicate predictions. NMS, a post-processing technique, targets false positive predictions with high Intersection over Union (IoU) but is less effective when these predictions are scattered along rays with low IoU. Focal Loss, a loss function designed to reduce high-confidence false positive predictions, has also been implemented. However, it has been observed that multi-view 3D object detectors using Focal Loss still face challenges in effectively resolving the problem of false positive predictions along the same rays." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "text", + "content": "Our quantitative analysis highlighted the importance of addressing false positive predictions along the same ray as Ground Truth. By utilizing the precise positional data of Ground Truth objects, we could identify and eliminate these redundant predictions within the state-of-the-art StreamPETR method [34]. This process significantly enhanced " + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "inline_equation", + "content": "5.4\\%" + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "text", + "content": " in mean Average Precision (mAP), emphasizing the critical need for models to improve their depth estimation capabilities. The substantial improvement underscores the potential of refining depth estimation to suppress these predictions and enhance overall detection performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Our key observation is that false positives often occur along camera rays due to the inherent limitation in conventional multi-view object detectors. Since the depth information for each pixel is not accurately estimated, the position embedding can only encode the ray direction. As a result, queries on the same ray will consistently interact with identical visual features from the image, leading to numerous duplicate predictions (false positives) along that ray. This scenario" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "F. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": "underscores the model's need to learn depth-aware features that can discern objects in depth despite the visual features being the same for objects along the same ray. We propose a novel method called Ray Denoising (i.e., RayDN). This framework is inherently flexible and does not limit the choice of distribution for sampling depth-aware hard negative samples. Based on our ablation studies, we have chosen the Beta distribution for its effectiveness in capturing the spatial distribution of false positives that models are likely to generate. This choice enables Ray Denoising to create depth-aware hard negative samples used for denoising, thereby enhancing the model's ability to learn more robust features and representations for distinguishing false positives along the ray, as depicted in Figure 2. Ray Denoising introduces only a marginal increase in computational costs during the training stage without affecting inference speed." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 261, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 415 + ], + "type": "text", + "content": "To summarize our contributions, we highlight the following key points. Firstly, we have identified the persistent challenge of false positive predictions along the same ray, which acts as a bottleneck in the performance of multi-view 3D object detectors. Secondly, we introduce Ray Denoising, a novel denoising method that utilizes the Beta distribution to create depth-aware hard negative samples along rays. This method explicitly takes into account the 3D structure of the scene, offering a flexible solution compatible with any DETR-style multi-view 3D detector to address the issue of duplicate predictions along rays. Lastly, our method achieves state-of-the-art results on the NuScenes dataset [2], significantly enhancing the performance of multi-view 3D object detectors. Specifically, we have improved upon the current state-of-the-art method, StreamPETR, by " + }, + { + "bbox": [ + 130, + 261, + 482, + 415 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 130, + 261, + 482, + 415 + ], + "type": "text", + "content": " in mean Average Precision (mAP), thereby demonstrating the effectiveness of Ray Denoising." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 432, + 237, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 432, + 237, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 432, + 237, + 445 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 455, + 332, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 455, + 332, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 455, + 332, + 468 + ], + "type": "text", + "content": "2.1 Image-based 3D Object Detection" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 474, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 582 + ], + "type": "text", + "content": "3D object detection is a cornerstone task in autonomous driving systems. The domain has witnessed notable advancements in monocular 3D object detection, largely attributed to the KITTI benchmark [5]. This progress has been driven by a range of research studies [1,24,28,31,35,44]. Nevertheless, monocular systems face limitations due to their dependence on a single viewpoint and limited data, which hampers their capacity to manage complex scenarios. To overcome these limitations, extensive benchmarks [2,32,39] have been developed, offering multi-viewpoint data that enriches the field of multi-view 3D object detection and propels the evolution of advanced detection methodologies." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 490, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 490, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 490, + 666 + ], + "type": "text", + "content": "The multi-view 3D object detection research is primarily bifurcated into dense Bird's Eye View (BEV)-based and sparse query-based algorithms. Our work is categorized under the sparse query-based algorithms. Dense BEV-based algorithms convert multi-view image features into a dense BEV representation using the Lift-Splat-Shoot (LSS) technique [27]. To mitigate overfitting from LSS, BEVDet [9] recommends data augmentation on BEV features. BEVDepth [14] introduces depth estimation supervision to enhance LSS precision. Compared to" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "type": "text", + "content": "Ray Denoising" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 238 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 238 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 238 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 238 + ], + "type": "image", + "image_path": "de98f7c221cf5f88b61b2009104a629912291a786a023b3f94d3030168fa9010.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 247, + 482, + 281 + ], + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 281 + ], + "type": "text", + "content": "Fig. 2: The proposed Ray Denoising approach (right) effectively reduces false positive detections along the ray (highlighted by red rectangles) in the previous state-of-the-art method StreamPETR [34] (left). Best viewed by zooming on the screen." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 303, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 422 + ], + "type": "text", + "content": "LSS-like methods, BEVFormer-like and DETR-like methods can leverage the strong capabilities of transformers for spatial-temporal modeling. BEVFormer-like methods use dense BEV features to enable multiple tasks (e.g., detection and segmentation) but also introduce high computational costs. Sparse query-based algorithms such as DETR3D [37] and PETR [21] utilize learnable 3D object queries and position encoding to engage with multi-view 2D features. DAT [43] introduces a pipeline to construct depth-aware position embeddings and performs contrastive learning on BEV features to tackle duplicate predictions. Recent studies [8, 13, 15, 17, 19, 22, 26, 34] have integrated temporal information to refine these methods further." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 441, + 320, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 320, + 454 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 320, + 454 + ], + "type": "text", + "content": "2.2 Hard Negative Samples Mining" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 460, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 594 + ], + "type": "text", + "content": "Hard negative sample mining is a strategy to minimize false positive predictions in object detection. Early works like R-CNN [6] employed standard hard negative mining methods [4, 33], where misclassified samples were utilized for retraining. OHEM [29] and RetinaNet [16] introduced adaptive sampling and focal loss to address the challenges of heuristics and hyper-parameters in hard negative sample selection. These methods select hard negative samples from existing predictions based on confidence scores and matched ground-truth object classes. Our approach takes a different path by constructing new hard negative samples based on the locations of ground-truth objects rather than selecting from existing predictions or depending on prediction confidence and matched ground-truth object classes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 609, + 315, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 315, + 623 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 315, + 623 + ], + "type": "text", + "content": "2.3 Denoising in Object Detection" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "In 2D object detection, models such as DETR [3] and its variants [20,36,46] often face convergence issues, which are partly due to the instability of bipartite graph matching and the inconsistency in optimization goals during the early stages of" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "F. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": "training. To tackle this, DN-DETR [12] introduces the concept of feeding noisy ground-truth bounding boxes into the Transformer decoder, with the model being trained to reconstruct the original, clean boxes. This process, known as denoising, involves using queries initialized with these noisy ground-truth bounding boxes, referred to as denoising queries. DINO [42] further refines this approach by generating both positive and negative denoising queries for each ground-truth bounding box, with the negative queries incorporating more noise to improve the model's ability to reject incorrect predictions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 480, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 480, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 480, + 368 + ], + "type": "text", + "content": "In the realm of 3D object detection, DETR-style methods [19,22,34] perpetuate the denoising approach by creating a denoising query for each ground-truth bounding box. Based on the noise level, these queries are classified as either an object or 'no object'. However, these methods neglect to incorporate knowledge of the 3D scene structure when generating denoising queries. In multi-view 3D object detection, the absence of depth information for each pixel results in models producing duplicate predictions along the same ray with different depths. To combat these false positive predictions, we harness the 3D structure knowledge of the scene. We devise multiple negative ray-denoising queries distributed along the rays, spatially distinct from the traditional negative denoising queries surrounding the ground-truth objects. This innovative design markedly improves the detector's capability to differentiate between true positive predictions and false positive predictions along the same ray." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 386, + 233, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 233, + 400 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 233, + 400 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 411, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 544 + ], + "type": "text", + "content": "This section elaborates on our proposed method, Ray Denoising, for multi-view 3D object detection. We start with a comprehensive overview of the DETR-style multi-view 3D object detector framework in Section 3.1. Following that, we describe the three pivotal steps for integrating Ray Denoising into the framework to address the challenge of duplicate predictions. Section 3.2 outlines the ray generation process. Section 3.3 explains utilizing the Beta distribution family for sampling reference points. Section 3.4 details the construction of spatial denoising queries, referred to as Ray Queries, along each ray. Lastly, Section 3.5 delves into the essential differences between our Ray Queries and the denoising queries employed in prior works and the influence of Ray Denoising on the model training process." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 562, + 208, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 562, + 208, + 572 + ], + "spans": [ + { + "bbox": [ + 132, + 562, + 208, + 572 + ], + "type": "text", + "content": "3.1 Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Ray Denoising is designed to be integrated into any DETR-style multi-view 3D object detector [15, 17, 19, 21, 22, 34, 37]. These detectors typically consist of a convolutional network-based feature encoder and a transformer-based decoder. The process begins by feeding " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " surround-view images " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{I} = \\{I_i\\in \\mathbb{R}^{3\\times H_I\\times W_I},i = 1,2,\\dots,N\\}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " into the feature encoder (e.g., ResNet101 [7]) to extract image features " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{F} = \\{F_{i}\\in \\mathbb{R}^{C\\times H_{F}\\times W_{F}},i = 1,2,\\dots,N\\}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "H_{I}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "W_{I}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " denote the image dimensions, " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "H_{F}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "W_{F}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " are the feature dimensions, and " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " represents the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "type": "text", + "content": "Ray Denoising" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 118, + 476, + 262 + ], + "blocks": [ + { + "bbox": [ + 140, + 118, + 476, + 262 + ], + "lines": [ + { + "bbox": [ + 140, + 118, + 476, + 262 + ], + "spans": [ + { + "bbox": [ + 140, + 118, + 476, + 262 + ], + "type": "image", + "image_path": "88714622b6b3ea899694bc720348b5a8c1f9b783061227f1d00f4ecae45bf6ba.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 274, + 482, + 352 + ], + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 352 + ], + "type": "text", + "content": "Fig. 3: Overall framework of the Ray Denoising approach, a plug-and-play training technique for DETR-style multi-view 3D object detectors, focuses on refining the model's ability to distinguish true positives from false positives in depth. Casting rays and sampling depth-aware denoising queries effectively tackle the challenge of false positives arising from the inherent difficulties in visually estimating depth, leading to substantial improvements in detection performance over strong baselines. Best viewed in color and by zooming on the screen." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 375, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 482, + 422 + ], + "type": "text", + "content": "number of channels in the feature. To facilitate 3D perception, multiple points within each camera's frustum are transformed and encoded into position embeddings [21]. These embeddings enable the multi-view image features to interact with 3D queries." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 423, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 510 + ], + "type": "text", + "content": "The final step involves " + }, + { + "bbox": [ + 130, + 423, + 482, + 510 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 423, + 482, + 510 + ], + "type": "text", + "content": " object queries " + }, + { + "bbox": [ + 130, + 423, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} \\in \\mathbb{R}^{N \\times 256}" + }, + { + "bbox": [ + 130, + 423, + 482, + 510 + ], + "type": "text", + "content": ", which are derived from a set of learnable 3D reference points " + }, + { + "bbox": [ + 130, + 423, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{P} \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 130, + 423, + 482, + 510 + ], + "type": "text", + "content": ". These queries interact with the multi-view image features " + }, + { + "bbox": [ + 130, + 423, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 130, + 423, + 482, + 510 + ], + "type": "text", + "content": " in the transformer decoder, employing multi-layer cross-attention to identify objects. The decoder's output features are then processed by prediction heads (multi-layer perceptron, i.e., MLP) to yield classification scores " + }, + { + "bbox": [ + 130, + 423, + 482, + 510 + ], + "type": "inline_equation", + "content": "(cls)" + }, + { + "bbox": [ + 130, + 423, + 482, + 510 + ], + "type": "text", + "content": ", position offsets " + }, + { + "bbox": [ + 130, + 423, + 482, + 510 + ], + "type": "inline_equation", + "content": "(x, y, z)" + }, + { + "bbox": [ + 130, + 423, + 482, + 510 + ], + "type": "text", + "content": ", scales " + }, + { + "bbox": [ + 130, + 423, + 482, + 510 + ], + "type": "inline_equation", + "content": "(w, h, l)" + }, + { + "bbox": [ + 130, + 423, + 482, + 510 + ], + "type": "text", + "content": ", directions " + }, + { + "bbox": [ + 130, + 423, + 482, + 510 + ], + "type": "inline_equation", + "content": "(\\theta_x, \\theta_y)" + }, + { + "bbox": [ + 130, + 423, + 482, + 510 + ], + "type": "text", + "content": ", and velocities " + }, + { + "bbox": [ + 130, + 423, + 482, + 510 + ], + "type": "inline_equation", + "content": "(v_x, v_y)" + }, + { + "bbox": [ + 130, + 423, + 482, + 510 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 527, + 224, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 527, + 224, + 540 + ], + "spans": [ + { + "bbox": [ + 132, + 527, + 224, + 540 + ], + "type": "text", + "content": "3.2 Ray Casting" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 549, + 482, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 549, + 482, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 549, + 482, + 608 + ], + "type": "text", + "content": "The reference points for our ray denoising queries are distributed along the camera ray, which extends from the camera's optical center to the ground-truth object on the image plane. To establish this ray, we project the 3D center of the ground-truth object into the camera frustum space using the following transformation:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 272, + 609, + 481, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 609, + 481, + 624 + ], + "spans": [ + { + "bbox": [ + 272, + 609, + 481, + 624 + ], + "type": "interline_equation", + "content": "\\mathbf {C} ^ {\\prime} = \\mathbf {K} \\cdot \\mathbf {C} _ {G T}, \\tag {1}", + "image_path": "c856a82c688b64afb639eba0e727c9a8996968a2fae36342d3baca4d13e9b797.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "4\\times 4" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " transformation matrix that maps points from 3D world space to camera frustum space. " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{GT} = (x,y,z,1)" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " is the center of the groundtruth object in 3D space, and " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{C}' = (u\\times d,v\\times d,d,1)" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " is the corresponding" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "F. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 117, + 304, + 232 + ], + "blocks": [ + { + "bbox": [ + 151, + 117, + 304, + 232 + ], + "lines": [ + { + "bbox": [ + 151, + 117, + 304, + 232 + ], + "spans": [ + { + "bbox": [ + 151, + 117, + 304, + 232 + ], + "type": "image", + "image_path": "4c4a2d14581946113ef63f596cdc1d87267b0add529ac47c993cac2426c793e4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 233, + 279, + 243 + ], + "lines": [ + { + "bbox": [ + 176, + 233, + 279, + 243 + ], + "spans": [ + { + "bbox": [ + 176, + 233, + 279, + 243 + ], + "type": "text", + "content": "(a) Distribution Comparison" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 115, + 463, + 232 + ], + "blocks": [ + { + "bbox": [ + 309, + 115, + 463, + 232 + ], + "lines": [ + { + "bbox": [ + 309, + 115, + 463, + 232 + ], + "spans": [ + { + "bbox": [ + 309, + 115, + 463, + 232 + ], + "type": "image", + "image_path": "22781dfede13b14d9e9f299c4676acdea843872bb5f1c03fda5ed7ddba61e0fe.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 233, + 436, + 243 + ], + "lines": [ + { + "bbox": [ + 335, + 233, + 436, + 243 + ], + "spans": [ + { + "bbox": [ + 335, + 233, + 436, + 243 + ], + "type": "text", + "content": "(b) Beta distribution family" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 251, + 482, + 297 + ], + "lines": [ + { + "bbox": [ + 131, + 251, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 131, + 251, + 482, + 297 + ], + "type": "text", + "content": "Fig. 4: (a) Distribution comparison showing that the Beta distribution is bounded between -1 and 1, unlike the Laplace and Gaussian distributions, which are unbounded. (b) The Beta distribution family, with the x-range adjusted from [0, 1] to " + }, + { + "bbox": [ + 131, + 251, + 482, + 297 + ], + "type": "inline_equation", + "content": "[-1, 1]" + }, + { + "bbox": [ + 131, + 251, + 482, + 297 + ], + "type": "text", + "content": " using the transformation " + }, + { + "bbox": [ + 131, + 251, + 482, + 297 + ], + "type": "inline_equation", + "content": "y = 2x - 1" + }, + { + "bbox": [ + 131, + 251, + 482, + 297 + ], + "type": "text", + "content": ". Best viewed in color." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 320, + 480, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 480, + 343 + ], + "type": "text", + "content": "projected center in the camera frustum, with " + }, + { + "bbox": [ + 130, + 320, + 480, + 343 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 130, + 320, + 480, + 343 + ], + "type": "text", + "content": " being the pixel coordinates and " + }, + { + "bbox": [ + 130, + 320, + 480, + 343 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 320, + 480, + 343 + ], + "type": "text", + "content": " the depth value." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 344, + 480, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 480, + 369 + ], + "type": "text", + "content": "With the depth of the ground-truth object's center known, we can determine the coordinates of valid reference points along the ray as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 259, + 377, + 480, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 377, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 259, + 377, + 480, + 392 + ], + "type": "interline_equation", + "content": "\\hat {d} _ {i} = d + \\beta_ {i} \\cdot f \\left(\\mathbf {S} _ {G T}\\right), \\tag {2}", + "image_path": "aa21610d3e309e87f436ab72d97e0624885fe9257f6a38d6f034cacaf4f967eb.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 401, + 479, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 479, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 479, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 401, + 479, + 437 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 401, + 479, + 437 + ], + "type": "text", + "content": " is a function that encodes the average scale of the ground-truth object, calculated as " + }, + { + "bbox": [ + 130, + 401, + 479, + 437 + ], + "type": "inline_equation", + "content": "f(\\mathbf{S}_{GT}) = k \\cdot \\frac{w + h + l}{6}" + }, + { + "bbox": [ + 130, + 401, + 479, + 437 + ], + "type": "text", + "content": ". The radius " + }, + { + "bbox": [ + 130, + 401, + 479, + 437 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 401, + 479, + 437 + ], + "type": "text", + "content": " defines the valid distribution range for reference points, and " + }, + { + "bbox": [ + 130, + 401, + 479, + 437 + ], + "type": "inline_equation", + "content": "\\beta_{i}" + }, + { + "bbox": [ + 130, + 401, + 479, + 437 + ], + "type": "text", + "content": " is the offset for the " + }, + { + "bbox": [ + 130, + 401, + 479, + 437 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 401, + 479, + 437 + ], + "type": "text", + "content": "-th reference point." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 146, + 437, + 392, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 437, + 392, + 449 + ], + "spans": [ + { + "bbox": [ + 146, + 437, + 392, + 449 + ], + "type": "text", + "content": "The position of the reference point is then obtained by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 270, + 457, + 480, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 457, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 270, + 457, + 480, + 472 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {P}} _ {i} = \\mathbf {K} ^ {- 1} \\cdot \\hat {\\mathbf {C}} _ {i} ^ {\\prime}, \\tag {3}", + "image_path": "b664ba7a2e522cd9eceac210b9669eb62a58da35e19266fe06d2987f61089165.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 481, + 281, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 481, + 281, + 495 + ], + "spans": [ + { + "bbox": [ + 131, + 481, + 281, + 495 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 481, + 281, + 495 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{C}}_i^{\\prime} = (u\\times \\hat{d}_i,v\\times \\hat{d}_i,\\hat{d}_i,1)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 513, + 257, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 513, + 257, + 525 + ], + "spans": [ + { + "bbox": [ + 131, + 513, + 257, + 525 + ], + "type": "text", + "content": "3.3 Sample Generation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 533, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 640 + ], + "type": "text", + "content": "With the valid reference points along the rays defined, we proceed to sample " + }, + { + "bbox": [ + 130, + 533, + 481, + 640 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 533, + 481, + 640 + ], + "type": "text", + "content": " reference points to emulate the distribution of false positive predictions that arise from depth ambiguities. A straightforward uniform distribution has proven effective, yielding a " + }, + { + "bbox": [ + 130, + 533, + 481, + 640 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 130, + 533, + 481, + 640 + ], + "type": "text", + "content": " mAP improvement as depicted in Figure 4a. To more accurately simulate the distribution of false positives, we consider two key aspects. First, we assess whether models tend to predict objects as being closer or further away from the ego-vehicle relative to the actual positions of the ground-truth objects. Second, we examine whether models are more likely to predict objects near or far from the center of the ground-truth objects." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Distributions like the Beta, Gaussian, and Laplace are suitable for handling complex scenarios, with the Gaussian and Laplace being unbounded. However," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "type": "text", + "content": "Ray Denoising" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "samples with significant offsets from the ground truth can result in large regression losses, negatively impacting optimization. As shown in Figure 4a, the Beta distribution slightly outperforms the others. Notably, the uniform distribution is a special case within the Beta family. For a versatile sampling strategy applicable to all scenarios, we select the Beta distribution family, characterized by the probability density function:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 217, + 196, + 481, + 222 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 196, + 481, + 222 + ], + "spans": [ + { + "bbox": [ + 217, + 196, + 481, + 222 + ], + "type": "interline_equation", + "content": "p (x | \\lambda , \\mu) = \\frac {\\Gamma (\\lambda + \\mu)}{\\Gamma (\\lambda) + \\Gamma (\\mu)} x ^ {\\lambda - 1} (1 - x) ^ {\\mu - 1}, \\tag {4}", + "image_path": "290368c025c12b778486d5547451ae336934aa2ad299e9466b562682ac180127.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 228, + 482, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 228, + 482, + 253 + ], + "spans": [ + { + "bbox": [ + 131, + 228, + 482, + 253 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 228, + 482, + 253 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 131, + 228, + 482, + 253 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 228, + 482, + 253 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 131, + 228, + 482, + 253 + ], + "type": "text", + "content": " are hyper-parameters that shape the distribution. The Gamma function, denoted as " + }, + { + "bbox": [ + 131, + 228, + 482, + 253 + ], + "type": "inline_equation", + "content": "\\Gamma(x)" + }, + { + "bbox": [ + 131, + 228, + 482, + 253 + ], + "type": "text", + "content": ", is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 257, + 261, + 481, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 261, + 481, + 287 + ], + "spans": [ + { + "bbox": [ + 257, + 261, + 481, + 287 + ], + "type": "interline_equation", + "content": "\\Gamma (x) = \\int_ {0} ^ {\\infty} t ^ {x - 1} e ^ {- t} d t. \\tag {5}", + "image_path": "441861e5ec904bf343c280d3e56fb722b193ab5e8dc673e5f92aaa75422d5dc9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "text", + "content": "The beta distribution's probability density function (PDF), visualized in Figure 4b, reveals its behavior under different parameter settings. When " + }, + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "text", + "content": " equals " + }, + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "text", + "content": ", the distribution is symmetric, reflecting a scenario where the model predicts objects with equal likelihood at closer or further distances. For instance, with " + }, + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\lambda = \\mu = 1" + }, + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "text", + "content": ", the distribution simplifies to a uniform distribution. Adjusting " + }, + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "text", + "content": " to be higher shifts the reference points of the denoising queries closer to the ground-truth object's center and vice versa." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 378, + 483, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 483, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 483, + 462 + ], + "type": "text", + "content": "When " + }, + { + "bbox": [ + 130, + 378, + 483, + 462 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 378, + 483, + 462 + ], + "type": "text", + "content": " exceeds " + }, + { + "bbox": [ + 130, + 378, + 483, + 462 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 378, + 483, + 462 + ], + "type": "text", + "content": ", such as in the case of " + }, + { + "bbox": [ + 130, + 378, + 483, + 462 + ], + "type": "inline_equation", + "content": "\\lambda = 8" + }, + { + "bbox": [ + 130, + 378, + 483, + 462 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 378, + 483, + 462 + ], + "type": "inline_equation", + "content": "\\mu = 2" + }, + { + "bbox": [ + 130, + 378, + 483, + 462 + ], + "type": "text", + "content": ", the distribution suggests that models are more likely to predict objects further away from the ego-vehicle. Conversely, when " + }, + { + "bbox": [ + 130, + 378, + 483, + 462 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 378, + 483, + 462 + ], + "type": "text", + "content": " is less than " + }, + { + "bbox": [ + 130, + 378, + 483, + 462 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 378, + 483, + 462 + ], + "type": "text", + "content": ", for example, " + }, + { + "bbox": [ + 130, + 378, + 483, + 462 + ], + "type": "inline_equation", + "content": "\\lambda = 2" + }, + { + "bbox": [ + 130, + 378, + 483, + 462 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 378, + 483, + 462 + ], + "type": "inline_equation", + "content": "\\mu = 8" + }, + { + "bbox": [ + 130, + 378, + 483, + 462 + ], + "type": "text", + "content": ", the distribution indicates a propensity for models to predict objects closer to the ego-vehicle. This flexibility in the Beta distribution allows for tailored sampling strategies that can adapt to the specific challenges of depth estimation in multiview 3D object detection." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 479, + 246, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 479, + 246, + 491 + ], + "spans": [ + { + "bbox": [ + 131, + 479, + 246, + 491 + ], + "type": "text", + "content": "3.4 Query Denoising" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "We construct ray denoising queries from the sampled " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " reference points. Specifically, a multi-layer perceptron (MLP) is utilized to project the normalized 3D coordinates of each point into a latent feature space. These " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " ray denoising queries " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{q}_r" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " are combined with the learnable object queries " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{q}_o" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " from the baseline detector and are input into the transformer decoder. An attention mask is employed during the self-attention operations to prevent information leakage from the ray denoising queries to the object queries. The query closest to the center of the ground truth (GT) object is designated as the positive query, inheriting the class label of the corresponding GT object. The remaining " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " queries are labeled as background, i.e., 'no object'. The loss for ray denoising queries adheres to the same criteria as for the learnable object queries, using Focal Loss for classification and L1 Loss for regression. This process ensures that the model is effectively trained to discern between true positives and false positives along the same ray." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "F. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 214, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 214, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 214, + 126 + ], + "type": "text", + "content": "3.5 Discussion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 133, + 482, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 133, + 482, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 133, + 482, + 421 + ], + "type": "text", + "content": "Our Ray Denoising approach is based on the pivotal observation that image-based 3D detection systems often struggle to distinguish true positives from false positives along camera rays. DETR-style multi-view 3D object detectors implicitly learn depth estimation from ground truth supervision. However, the randomly distributed reference points of learnable queries do not fully leverage the available ground truth information. While these reference points are updated during training, they fail to provide adequate hard negative samples for each object in every scene. To enhance the utilization of ground truth information, traditional denoising techniques introduce additional reference points distributed uniformly around the ground-truth object during training. These instance-specific reference points have improved detection performance [19,22,34]. However, they overlook the depth ambiguities intrinsic to multi-view 3D object detection. The absence of precise depth information for each pixel means that the position embedding can only encode ray direction, not depth. This leads to queries on the same ray interacting with the same image features, resulting in redundant predictions. Ray Denoising diverges from traditional techniques by strategically generating reference points along rays that extend from cameras to objects. This approach explicitly considers the 3D structure of each object in the scene, providing a sufficient number of hard negative samples. During training, these Ray Queries interact within the self-attention layer, effectively guiding the model to suppress depth-ambiguous spatial hard negative samples. This interaction enhances the detector's ability to differentiate between true positive predictions (objects) and false positive predictions (duplicate detections), improving detection accuracy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 435, + 224, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 224, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 224, + 449 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 457, + 264, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 264, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 264, + 468 + ], + "type": "text", + "content": "4.1 Dataset and Metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 474, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 593 + ], + "type": "text", + "content": "Our model's performance was assessed using two datasets: nuScenes [2] and Argovverse 2 [39]. The nuScenes dataset include 1000 video sequences, which are split into training (700 videos), validation (150 videos), and testing (150 videos) sets, each approximately 20 seconds long with annotations at 0.5-second intervals. The dataset comprises 1.4 million annotated 3D bounding boxes across ten object classes. Evaluation metrics include the mean Average Precision (mAP) and five true positive metrics: ATE, ASE, AOE, AVE, and AAE, which assess translation, scale, orientation, velocity, and attribute errors. The nuScenes Detection Score (NDS) is a comprehensive score derived from these metrics, offering an overall performance evaluation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "The Argoverse 2 dataset contains 1000 unique scenes, each 15 seconds long, annotated at " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "10\\mathrm{Hz}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ". The scenes are divided into 700 for training, 150 for validation, and 150 for testing. The evaluation covers 26 categories within a 150-meter range, addressing long-range perception tasks. Metrics include the mAP and the Composite Detection Score (CDS), which integrates three other true positive metrics: ATE, ASE, and AOE." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "type": "text", + "content": "Ray Denoising" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 146, + 483, + 299 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": "Table 1: Comparison on the nuScenes validation set. " + }, + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": " Indicates methods that benefit from perspective-view pre-training." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 146, + 483, + 299 + ], + "lines": [ + { + "bbox": [ + 132, + 146, + 483, + 299 + ], + "spans": [ + { + "bbox": [ + 132, + 146, + 483, + 299 + ], + "type": "table", + "html": "
MethodsBackboneImage SizemAPNDSmATEmASEmAOEmAVEmAAE
BevDet4D [8]ResNet50256×70432.245.70.7030.2780.4950.3540.206
PETRv2 [22]ResNet50256×70434.945.60.7000.2750.5800.4370.187
BEVDepth [14]ResNet50256×70435.147.50.6290.2670.4790.4280.198
BEVStereo [13]ResNet50256×70437.250.00.5980.2700.4380.3670.190
BEVFomerv2 [41]†ResNet50-42.352.90.6180.2730.4130.3330.188
SOLOFusion [26]ResNet50256×70442.753.40.5670.2740.5110.2520.181
SparseBEV [19]†ResNet50256×70444.855.80.5810.2710.3730.2470.190
StreamPETR [34]†ResNet50256×70445.055.00.6130.2670.4130.2650.198
RayDN† (Ours)ResNet50256×70446.956.30.5790.2640.4330.2560.187
BEVDepth [14]ResNet101512×140841.253.50.5650.2660.3580.3310.190
PETRv2 [22]†ResNet101640×160042.152.40.6810.2670.3570.3770.186
Sparse4D [17]†ResNet101900×160043.654.10.6330.2790.3630.3170.177
SOLOFusion [26]ResNet101512×140848.358.20.5030.2640.3810.2460.207
SparseBEV [19]†ResNet101512×140850.159.20.5620.2650.3210.2430.195
StreamPETR [34]†ResNet101512×140850.459.20.5690.2620.3150.2570.199
RayDN† (Ours)ResNet101512×140851.860.40.5410.2600.3150.2360.200
", + "image_path": "5b9fcccbb114ab0b4619f299cd2e0cb4e72dc8950a09b46d5f4bec9f954dae4f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 134, + 338, + 479, + 439 + ], + "blocks": [ + { + "bbox": [ + 209, + 318, + 403, + 329 + ], + "lines": [ + { + "bbox": [ + 209, + 318, + 403, + 329 + ], + "spans": [ + { + "bbox": [ + 209, + 318, + 403, + 329 + ], + "type": "text", + "content": "Table 2: Comparison on the nuScenes test set." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 338, + 479, + 439 + ], + "lines": [ + { + "bbox": [ + 134, + 338, + 479, + 439 + ], + "spans": [ + { + "bbox": [ + 134, + 338, + 479, + 439 + ], + "type": "table", + "html": "
MethodsBackboneImage SizemAPNDSmATEmASEmAOEmAVEmAAE
DETR3D [37]V2-99900×160041.247.90.6410.2550.3940.8450.133
MV2D [38]V2-99640×160046.351.40.5420.2470.4030.8570.127
BEVFormer [15]V2-99900×160048.156.90.5820.2560.3750.3780.126
PETRv2 [22]V2-99640×160049.058.20.5610.2430.3610.3430.120
PolarFormer [10]V2-99900×160049.357.20.5560.2560.3640.4390.127
BEVStereo [13]V2-99900×160052.561.00.4310.2460.3580.3570.138
HoP [47]V2-99640×160052.861.20.4910.2420.3320.3430.109
SparseBEV [19]V2-99640×160054.362.70.5020.2440.3240.2510.126
StreamPETR [34]V2-99640×160055.063.60.4790.2390.3170.2410.119
RayDN (Ours)V2-99640×160056.564.50.4610.2410.3220.2390.114
", + "image_path": "ad5d40f5996ce1df058c171a188420a9c4cf8d68124fa83017c8e7db894cf6b1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 470, + 279, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 470, + 279, + 481 + ], + "spans": [ + { + "bbox": [ + 132, + 470, + 279, + 481 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": "Our experimental setup leverages ResNet50 [7], ResNet101, and V2-99 [11] backbones, each with distinct pre-training configurations. We detail the performance of ResNet50 and ResNet101 models, which are pre-trained on nuImages [2], on the nuScenes validation set. To showcase the scalability of our approach, we also report results on the nuScenes test set using V2-99, initialized from the DD3D [25] checkpoint. Our models are optimized using the AdamW [23] optimizer with a batch size of 16. The learning rate is set to " + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "inline_equation", + "content": "4 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": " for models trained only on the training set and " + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "inline_equation", + "content": "3 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": " for those trained on both the training and validation sets. We adopt a cosine annealing policy for learning rate scheduling. For benchmarking against state-of-the-art (SOTA) methods, models are trained for 60 epochs without CBGS [45] and for 24 epochs in ablation studies in nuScenes. For experiments on Argoverse 2, models are trained for 6 epochs. Our implementation is primarily based on the StreamPETR [34] framework." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "F. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 164, + 146, + 451, + 200 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 3: Comparisons on the Argoverse 2 validation set. We evaluate across 26 object categories within a range of 150 meters." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 164, + 146, + 451, + 200 + ], + "lines": [ + { + "bbox": [ + 164, + 146, + 451, + 200 + ], + "spans": [ + { + "bbox": [ + 164, + 146, + 451, + 200 + ], + "type": "table", + "html": "
MethodsBackboneImage SizemAPCDSmATEmASEmAOE
PETR [21]V2-99900×64017.612.20.9110.3390.819
Sparse4Dv2 [18]V2-99900×64018.913.40.8320.3430.723
StreamPETR [34]V2-99900×64020.314.60.8430.3210.650
RayDN (Ours)V2-99900×64022.316.10.8250.3250.629
", + "image_path": "f122060a0c270bcc0c56c44a348e556fd4ba2f2119f8cded6579660f5dcda83c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 222, + 380, + 233 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 222, + 380, + 233 + ], + "spans": [ + { + "bbox": [ + 132, + 222, + 380, + 233 + ], + "type": "text", + "content": "4.3 Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 243, + 480, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 243, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 243, + 480, + 289 + ], + "type": "text", + "content": "We compare the proposed Ray Denoising method with other state-of-the-art multi-view 3D object detectors on the validation and test sets of the nuScenes dataset, as well as the validation set of the Argoverse 2 dataset. It's important to note that our method does not employ test time augmentation (TTA)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 291, + 481, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 291, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 291, + 481, + 397 + ], + "type": "text", + "content": "nuScenes Validation Set. Table 1 presents a comparison with state-of-the-art methods on the nuScenes validation set. We evaluate both ResNet-50 and ResNet-101 backbones. With ResNet-50 as the backbone and an image size of " + }, + { + "bbox": [ + 132, + 291, + 481, + 397 + ], + "type": "inline_equation", + "content": "704 \\times 256" + }, + { + "bbox": [ + 132, + 291, + 481, + 397 + ], + "type": "text", + "content": ", we achieve " + }, + { + "bbox": [ + 132, + 291, + 481, + 397 + ], + "type": "inline_equation", + "content": "46.9\\%" + }, + { + "bbox": [ + 132, + 291, + 481, + 397 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 132, + 291, + 481, + 397 + ], + "type": "inline_equation", + "content": "56.3\\%" + }, + { + "bbox": [ + 132, + 291, + 481, + 397 + ], + "type": "text", + "content": " NDS, improving upon the previous state-of-the-art method, StreamPETR, by " + }, + { + "bbox": [ + 132, + 291, + 481, + 397 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 132, + 291, + 481, + 397 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 132, + 291, + 481, + 397 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 132, + 291, + 481, + 397 + ], + "type": "text", + "content": " NDS. Using a more robust ResNet-101 backbone and increasing the image size to " + }, + { + "bbox": [ + 132, + 291, + 481, + 397 + ], + "type": "inline_equation", + "content": "512 \\times 1408" + }, + { + "bbox": [ + 132, + 291, + 481, + 397 + ], + "type": "text", + "content": ", our performance reaches " + }, + { + "bbox": [ + 132, + 291, + 481, + 397 + ], + "type": "inline_equation", + "content": "51.6\\%" + }, + { + "bbox": [ + 132, + 291, + 481, + 397 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 132, + 291, + 481, + 397 + ], + "type": "inline_equation", + "content": "59.8\\%" + }, + { + "bbox": [ + 132, + 291, + 481, + 397 + ], + "type": "text", + "content": " NDS, outperforming StreamPETR by " + }, + { + "bbox": [ + 132, + 291, + 481, + 397 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 132, + 291, + 481, + 397 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 132, + 291, + 481, + 397 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 132, + 291, + 481, + 397 + ], + "type": "text", + "content": " NDS. These results demonstrate the scalability of Ray Denoising." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 399, + 480, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 399, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 399, + 480, + 445 + ], + "type": "text", + "content": "nuScenes Test Set. The results evaluated by the test server are detailed in Table 2. Our method includes training on both the training and validation sets. Notably, we achieve " + }, + { + "bbox": [ + 132, + 399, + 480, + 445 + ], + "type": "inline_equation", + "content": "56.5\\%" + }, + { + "bbox": [ + 132, + 399, + 480, + 445 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 132, + 399, + 480, + 445 + ], + "type": "inline_equation", + "content": "64.5\\%" + }, + { + "bbox": [ + 132, + 399, + 480, + 445 + ], + "type": "text", + "content": " NDS, surpassing StreamPETR by an absolute " + }, + { + "bbox": [ + 132, + 399, + 480, + 445 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 132, + 399, + 480, + 445 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 132, + 399, + 480, + 445 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 132, + 399, + 480, + 445 + ], + "type": "text", + "content": " NDS." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 447, + 480, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 506 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 506 + ], + "type": "text", + "content": "Argoverse 2 Validation Set. To assess the generalization capability of Ray Denoising, we conduct additional experiments on the Argoverse 2 dataset, as displayed in Table 3. Our method significantly outperforms previous state-of-the-art methods, achieving a " + }, + { + "bbox": [ + 132, + 447, + 480, + 506 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 132, + 447, + 480, + 506 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 132, + 447, + 480, + 506 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 132, + 447, + 480, + 506 + ], + "type": "text", + "content": " CDS improvement on the validation set. These metrics underscore the generalization capability of our approach." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 506, + 480, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 506, + 480, + 554 + ], + "spans": [ + { + "bbox": [ + 132, + 506, + 480, + 554 + ], + "type": "text", + "content": "The tables demonstrate that our Ray Denoising approach notably enhances mAP. Considering that mAP is significantly impacted by incorrect false positives, this enhancement firmly validates the effectiveness of Ray Denoising in reducing redundant predictions along the camera rays." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 573, + 237, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 573, + 237, + 585 + ], + "spans": [ + { + "bbox": [ + 132, + 573, + 237, + 585 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 594, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 480, + 665 + ], + "type": "text", + "content": "This section delves into the ablation studies performed using the validation sets from the nuScenes and Argoverse 2 datasets. Unless specified otherwise, our experiments on nuScenes are based on a ResNet50 backbone, which is pre-trained on the nuImages dataset [2]. Our input data comprises 8 frames, each with a " + }, + { + "bbox": [ + 132, + 594, + 480, + 665 + ], + "type": "inline_equation", + "content": "704 \\times 256" + }, + { + "bbox": [ + 132, + 594, + 480, + 665 + ], + "type": "text", + "content": " pixels resolution. The decoder utilizes 428 queries, and the model is trained for 24 epochs, forgoing Class-Balanced Group Sampling (CBGS) [45]." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "type": "text", + "content": "Ray Denoising" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 153, + 145, + 302, + 200 + ], + "blocks": [ + { + "bbox": [ + 147, + 114, + 309, + 137 + ], + "lines": [ + { + "bbox": [ + 147, + 114, + 309, + 137 + ], + "spans": [ + { + "bbox": [ + 147, + 114, + 309, + 137 + ], + "type": "text", + "content": "Table 4: Ablation studies on the radius of ray denoising queries." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 153, + 145, + 302, + 200 + ], + "lines": [ + { + "bbox": [ + 153, + 145, + 302, + 200 + ], + "spans": [ + { + "bbox": [ + 153, + 145, + 302, + 200 + ], + "type": "table", + "html": "
MethodRadiusmAPNDS
SOTA Baseline [34]-39.148.6
241.149.8
+RayDN (Ours)341.950.8
441.350.1
", + "image_path": "84c2932491fff4e12ba73da07ba15c6db087ad6e29b3bb98bd4b394ee2312b7b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 162, + 251, + 296, + 330 + ], + "blocks": [ + { + "bbox": [ + 156, + 210, + 303, + 243 + ], + "lines": [ + { + "bbox": [ + 156, + 210, + 303, + 243 + ], + "spans": [ + { + "bbox": [ + 156, + 210, + 303, + 243 + ], + "type": "text", + "content": "Table 6: Ablation studies on the distribution of ray denoising queries on nuScenes dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 162, + 251, + 296, + 330 + ], + "lines": [ + { + "bbox": [ + 162, + 251, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 162, + 251, + 296, + 330 + ], + "type": "table", + "html": "
Methodλ μ mAPNDS
SOTA Baseline [34]-39.148.6
+RayDN (Ours)141.350.4
241.550.1
341.350.5
741.750.7
841.950.8
941.650.4
", + "image_path": "aa316e7ae10113e8ab4efa0a1c1cfebf7e06d3f870b44d86f2a10996a3fdbc7d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 322, + 145, + 458, + 200 + ], + "blocks": [ + { + "bbox": [ + 316, + 114, + 465, + 137 + ], + "lines": [ + { + "bbox": [ + 316, + 114, + 465, + 137 + ], + "spans": [ + { + "bbox": [ + 316, + 114, + 465, + 137 + ], + "type": "text", + "content": "Table 5: Ablation studies on the number of ray denoising queries." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 145, + 458, + 200 + ], + "lines": [ + { + "bbox": [ + 322, + 145, + 458, + 200 + ], + "spans": [ + { + "bbox": [ + 322, + 145, + 458, + 200 + ], + "type": "table", + "html": "
Method#QmAPNDS
SOTA Baseline [34]-39.148.6
341.650.2
+RayDN (Ours)541.950.8
741.150.5
", + "image_path": "1d75207de81d4d50ec16d5425f2a13943a2f65e921a5f644695f8781b557444c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 252, + 450, + 330 + ], + "blocks": [ + { + "bbox": [ + 310, + 210, + 457, + 243 + ], + "lines": [ + { + "bbox": [ + 310, + 210, + 457, + 243 + ], + "spans": [ + { + "bbox": [ + 310, + 210, + 457, + 243 + ], + "type": "text", + "content": "Table 7: Ablation studies on the distribution of ray denoising queries for the Argoverse 2 dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 252, + 450, + 330 + ], + "lines": [ + { + "bbox": [ + 316, + 252, + 450, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 450, + 330 + ], + "type": "table", + "html": "
Methodλ μmAPCDS
SOTA Baseline [34]-20.314.6
+RayDN (Ours)121.315.4
321.715.8
421.315.3
621.515.7
722.316.1
821.515.6
", + "image_path": "da0456bb682320c156bf88d462b43d92d5134b9040d3f2243e610755512c83db.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 354, + 480, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 480, + 389 + ], + "type": "text", + "content": "Radius of Ray Denoising Queries. Table 4 shows how varying the radius " + }, + { + "bbox": [ + 130, + 354, + 480, + 389 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 354, + 480, + 389 + ], + "type": "text", + "content": " affects performance. We find that a radius of " + }, + { + "bbox": [ + 130, + 354, + 480, + 389 + ], + "type": "inline_equation", + "content": "k = 3" + }, + { + "bbox": [ + 130, + 354, + 480, + 389 + ], + "type": "text", + "content": " yields the optimal results, while " + }, + { + "bbox": [ + 130, + 354, + 480, + 389 + ], + "type": "inline_equation", + "content": "k = 2" + }, + { + "bbox": [ + 130, + 354, + 480, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 354, + 480, + 389 + ], + "type": "inline_equation", + "content": "k = 4" + }, + { + "bbox": [ + 130, + 354, + 480, + 389 + ], + "type": "text", + "content": " also lead to notable enhancements in performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 389, + 481, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 481, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 481, + 461 + ], + "type": "text", + "content": "Number of Ray Denoising Queries. We explore the effect of the quantity of ray denoising queries on the model's performance in Table 5. The findings reveal that both mAP and NDS metrics improve with increasing ray queries, reaching a plateau of 5 queries. Surprisingly, upping the count to 7 queries results in a dip in mAP, which might be attributed to an imbalance in the ratio of positive to negative ray denoising queries." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 461, + 482, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 606 + ], + "type": "text", + "content": "Distribution of Ray Denoising Queries. We investigate the impact of the distribution of ray denoising queries with various hyper-parameters across the nuScenes and Argoverse 2 datasets. The results are detailed in Table 6 and Table 7. Our findings indicate that employing a uniform distribution, which is dataset-agnostic (i.e., setting " + }, + { + "bbox": [ + 130, + 461, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\lambda = 1" + }, + { + "bbox": [ + 130, + 461, + 482, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 461, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\mu = 1" + }, + { + "bbox": [ + 130, + 461, + 482, + 606 + ], + "type": "text", + "content": "), leads to substantial performance gains for Ray Denoising. Specifically, we observe a " + }, + { + "bbox": [ + 130, + 461, + 482, + 606 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 130, + 461, + 482, + 606 + ], + "type": "text", + "content": " increase in mAP for the nuScenes dataset and a " + }, + { + "bbox": [ + 130, + 461, + 482, + 606 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 130, + 461, + 482, + 606 + ], + "type": "text", + "content": " increase in mAP for the Argoverse 2 dataset, highlighting the method's strong generalization capabilities. Fine-tuning these hyper-parameters further enhances the performance, with an additional " + }, + { + "bbox": [ + 130, + 461, + 482, + 606 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 130, + 461, + 482, + 606 + ], + "type": "text", + "content": " mAP improvement on nuScenes and " + }, + { + "bbox": [ + 130, + 461, + 482, + 606 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 130, + 461, + 482, + 606 + ], + "type": "text", + "content": " mAP on Argoverse. Across both datasets, models configured with " + }, + { + "bbox": [ + 130, + 461, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\lambda > \\mu" + }, + { + "bbox": [ + 130, + 461, + 482, + 606 + ], + "type": "text", + "content": " slightly outperform those with " + }, + { + "bbox": [ + 130, + 461, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\lambda < \\mu" + }, + { + "bbox": [ + 130, + 461, + 482, + 606 + ], + "type": "text", + "content": ", suggesting that false positives tend to be at greater distance from the ego-vehicle." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Precision-Recall Analysis. We assess the ability of Ray Denoising to reduce false positive predictions by plotting precision-recall curves and Average Precision (AP) for each object class. Figure 5a reveals that Ray Denoising enhances precision across a range of recall levels and under various distance thresholds." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "F. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 115, + 305, + 230 + ], + "blocks": [ + { + "bbox": [ + 149, + 115, + 305, + 230 + ], + "lines": [ + { + "bbox": [ + 149, + 115, + 305, + 230 + ], + "spans": [ + { + "bbox": [ + 149, + 115, + 305, + 230 + ], + "type": "image", + "image_path": "d0b91b7a89a031bb1770ad96eab0df73bee095aefb6cb0ea81b5b7748c9927cd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 180, + 234, + 276, + 243 + ], + "lines": [ + { + "bbox": [ + 180, + 234, + 276, + 243 + ], + "spans": [ + { + "bbox": [ + 180, + 234, + 276, + 243 + ], + "type": "text", + "content": "(a) Precision-Recall Curve" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 311, + 117, + 461, + 228 + ], + "blocks": [ + { + "bbox": [ + 311, + 117, + 461, + 228 + ], + "lines": [ + { + "bbox": [ + 311, + 117, + 461, + 228 + ], + "spans": [ + { + "bbox": [ + 311, + 117, + 461, + 228 + ], + "type": "image", + "image_path": "245734db36b9a630fd1de0330c5d8c048ad2804fd257f9c15985555fe8cb0374.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 233, + 425, + 243 + ], + "lines": [ + { + "bbox": [ + 348, + 233, + 425, + 243 + ], + "spans": [ + { + "bbox": [ + 348, + 233, + 425, + 243 + ], + "type": "text", + "content": "(b) AP for each class" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 150, + 308, + 463, + 370 + ], + "blocks": [ + { + "bbox": [ + 130, + 251, + 482, + 297 + ], + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 297 + ], + "type": "text", + "content": "Fig. 5: (a) Visualization of the precision-recall curves at various distance thresholds. Ray Denoising consistently enhances precision across nearly all recall levels, effectively suppressing false positives. (b) Class-wise AP comparison. Ray Denoising performs superior over the SOTA StreamPETR in all object classes. Best viewed in color." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 150, + 308, + 463, + 370 + ], + "lines": [ + { + "bbox": [ + 150, + 308, + 463, + 370 + ], + "spans": [ + { + "bbox": [ + 150, + 308, + 463, + 370 + ], + "type": "image", + "image_path": "67bbe4937d086929e32c83c727dd7a8a10923f3469b49e05c6cd55054de44722.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 378, + 482, + 411 + ], + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 411 + ], + "type": "text", + "content": "Fig.6: Visualization of our detection results on the validation set of nuScenes. Ray Denoising effectively mitigates duplicate false positives while maintaining the ability to detect highly occluded objects along the same ray. Best viewed on the screen." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 437, + 480, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 480, + 472 + ], + "type": "text", + "content": "Similarly, Figure 5b shows that Ray Denoising surpasses StreamPETR in AP across all classes. These results collectively affirm that Ray Denoising effectively mitigates the occurrence of false positive predictions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 473, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 556 + ], + "type": "text", + "content": "Visualization of Detection Results. Our approach creates multiple spatial hard negative samples along the same ray, potentially interfering with detecting true positives if they are also located along that ray. To test this, we present a visualization of the detection outcomes in Figure 6. The results show that Ray Denoising effectively eliminates duplicate false positives without compromising the detection of true positives, even when they are heavily occluded along the same ray." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Generalization Ability of Ray Denoising. To assess the versatility of Ray Denoising, we apply it to different models with varied input data. We use the state-of-the-art CMT [40] as a reference and test Ray Denoising with single-frame images and multi-modal data. In the single-frame scenario, we employ a ResNet50 backbone pre-trained on ImageNet [2], with input images sized at " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "256 \\times 704" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " pixels. For the multi-modal setup, the input image resolution is " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "320 \\times 800" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " pixels, and the voxel dimensions are " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.1\\mathrm{m} \\times 0.1\\mathrm{m}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ". The batch size is set to 16 due to memory constraints, and all models are trained for 20 epochs using CBGS [45]. Table 8 shows that Ray Denoising improves upon the baseline" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "type": "text", + "content": "Ray Denoising" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 173, + 145, + 306, + 198 + ], + "blocks": [ + { + "bbox": [ + 166, + 114, + 312, + 137 + ], + "lines": [ + { + "bbox": [ + 166, + 114, + 312, + 137 + ], + "spans": [ + { + "bbox": [ + 166, + 114, + 312, + 137 + ], + "type": "text", + "content": "Table 8: Ablation studies on generalization ability of Ray Denoising." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 173, + 145, + 306, + 198 + ], + "lines": [ + { + "bbox": [ + 173, + 145, + 306, + 198 + ], + "spans": [ + { + "bbox": [ + 173, + 145, + 306, + 198 + ], + "type": "table", + "html": "
MethodDescriptionmAPNDS
CMT [40]single frame30.637.7
+ RayDNsingle frame32.139.1
CMT [40]multi-modal67.670.4
+ RayDNmulti-modal68.671.3
", + "image_path": "1645f47af6ae69c96d16f351fd2a22a58eb330813f8539d1b209bc031e86aa64.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 326, + 145, + 438, + 198 + ], + "blocks": [ + { + "bbox": [ + 320, + 114, + 446, + 137 + ], + "lines": [ + { + "bbox": [ + 320, + 114, + 446, + 137 + ], + "spans": [ + { + "bbox": [ + 320, + 114, + 446, + 137 + ], + "type": "text", + "content": "Table 9: Ablation studies on the effect of depth estimation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 145, + 438, + 198 + ], + "lines": [ + { + "bbox": [ + 326, + 145, + 438, + 198 + ], + "spans": [ + { + "bbox": [ + 326, + 145, + 438, + 198 + ], + "type": "table", + "html": "
3DPPERayDNmAPNDS
39.148.6
✓42.051.1
✓41.950.8
✓✓43.952.9
", + "image_path": "cacbb5497133a3442ddbf109fae76fea9274f5692048b8d50a53c95a4526fe67.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 220, + 480, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 220, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 480, + 267 + ], + "type": "text", + "content": "with a " + }, + { + "bbox": [ + 130, + 220, + 480, + 267 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 130, + 220, + 480, + 267 + ], + "type": "text", + "content": " increase in mAP and a " + }, + { + "bbox": [ + 130, + 220, + 480, + 267 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 130, + 220, + 480, + 267 + ], + "type": "text", + "content": " increase in NDS for the single-frame setup. Moreover, it achieves a " + }, + { + "bbox": [ + 130, + 220, + 480, + 267 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 130, + 220, + 480, + 267 + ], + "type": "text", + "content": " mAP and a " + }, + { + "bbox": [ + 130, + 220, + 480, + 267 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 130, + 220, + 480, + 267 + ], + "type": "text", + "content": " NDS enhancement on the already high-performing multi-modal baseline. These experiments confirm the broad applicability of Ray Denoising." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 267, + 481, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 481, + 459 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 481, + 459 + ], + "type": "text", + "content": "Effect of Depth Estimation on Ray Denoising. In DETR-style multi-view 3D object detectors, the lack of depth information leads to the use of camera rays as position embeddings for image features, which can be imprecise. To refine these embeddings, 3DPPE [30] suggests encoding an estimated 3D point guided by external supervision for depth estimation. To explore the role of depth estimation in conjunction with Ray Denoising, we integrate 3DPPE into our framework. The results, detailed in Table 9, show a marked improvement in performance with 3DPPE. However, Ray Denoising alone, without external depth supervision, also achieves similar performance. This underscores our core insight: Ray Denoising enhances the model's depth perception by guiding it to learn superior depth perception ability during training. When Ray Denoising is paired with 3DPPE, we further obtain a " + }, + { + "bbox": [ + 130, + 267, + 481, + 459 + ], + "type": "inline_equation", + "content": "+1.9\\%" + }, + { + "bbox": [ + 130, + 267, + 481, + 459 + ], + "type": "text", + "content": " increase in mAP and a " + }, + { + "bbox": [ + 130, + 267, + 481, + 459 + ], + "type": "inline_equation", + "content": "+1.7\\%" + }, + { + "bbox": [ + 130, + 267, + 481, + 459 + ], + "type": "text", + "content": " increase in NDS. This demonstrates that while depth estimation algorithms can improve the clarity of position embeddings, Ray Denoising's ability to refine the model's understanding of depth through robust training is a powerful complement, even when depth information is available." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 475, + 220, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 220, + 488 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 220, + 488 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 498, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 665 + ], + "type": "text", + "content": "We introduce Ray Denoising, a method designed to overcome the critical challenge of depth estimation in multi-view 3D object detection. Ray Denoising tackles the issue of false detections along camera rays, which are a direct consequence of imprecise depth information from images. By leveraging the 3D structure of the scene, Ray Denoising prompts the model to learn depth-aware features, leading to improved differentiation between false and true positives along the same ray without introducing extra inference costs. Our comprehensive experiments on the NuScenes and Argoverse 2 datasets demonstrate that Ray Denoising consistently and significantly outperforms strong baselines, achieving new state-of-the-art performance in Multi-view 3D Object Detection. Acknowledgment. This work was supported by the Fundamental Research Funds for the Central Universities (E3E41903, E2ET1104, E3ET6201X2), the National Natural Science Foundation of China (NSFC) under Grant 62225208 and 62171431." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "F. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "type": "text", + "content": "1. Brazil, G., Liu, X.: M3d-rpn: Monocular 3d region proposal network for object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9287-9296 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 170, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 170, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 138, + 170, + 480, + 213 + ], + "type": "text", + "content": "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 213, + 480, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 213, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 213, + 480, + 245 + ], + "type": "text", + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 246, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 480, + 277 + ], + "type": "text", + "content": "4. Felzenszwalb, P.F., Girshick, R.B., McAllester, D., Ramanan, D.: Object detection with discriminatively trained part-based models. IEEE transactions on pattern analysis and machine intelligence 32(9), 1627-1645 (2009)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 277, + 480, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 480, + 298 + ], + "type": "text", + "content": "5. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset. The International Journal of Robotics Research 32(11), 1231-1237 (2013)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 299, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 480, + 331 + ], + "type": "text", + "content": "6. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Region-based convolutional networks for accurate object detection and segmentation. IEEE transactions on pattern analysis and machine intelligence 38(1), 142-158 (2015)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 332, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 480, + 363 + ], + "type": "text", + "content": "7. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 364, + 480, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 364, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 138, + 364, + 480, + 384 + ], + "type": "text", + "content": "8. Huang, J., Huang, G.: Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 384, + 480, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 384, + 480, + 416 + ], + "spans": [ + { + "bbox": [ + 138, + 384, + 480, + 416 + ], + "type": "text", + "content": "9. Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multicamera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 418, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 480, + 449 + ], + "type": "text", + "content": "0. Jiang, Y., Zhang, L., Miao, Z., Zhu, X., Gao, J., Hu, W., Jiang, Y.G.: Polarformer: Multi-camera 3d object detection with polar transformer. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1042-1050 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 449, + 480, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 480, + 492 + ], + "type": "text", + "content": "1. Lee, Y., Hwang, J.w., Lee, S., Bae, Y., Park, J.: An energy andgpu-computation efficient backbone network for real-time object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 0-0 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 493, + 480, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 493, + 480, + 524 + ], + "spans": [ + { + "bbox": [ + 138, + 493, + 480, + 524 + ], + "type": "text", + "content": "2. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13619-13627 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 525, + 480, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 480, + 556 + ], + "type": "text", + "content": "3. Li, Y., Bao, H., Ge, Z., Yang, J., Sun, J., Li, Z.: Bevstereo: Enhancing depth estimation in multi-view 3d object detection with temporal stereo. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1486-1494 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 557, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 557, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 557, + 480, + 589 + ], + "type": "text", + "content": "4. Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1477-1485 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 590, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 632 + ], + "type": "text", + "content": "5. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: European conference on computer vision. pp. 1-18. Springer (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "6. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "type": "text", + "content": "Ray Denoising" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "17. Lin, X., Lin, T., Pei, Z., Huang, L., Su, Z.: Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion. arXiv preprint arXiv:2211.10581 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 161 + ], + "type": "text", + "content": "18. Lin, X., Lin, T., Pei, Z., Huang, L., Su, Z.: Sparse4d v2: Recurrent temporal fusion with sparse model. arXiv preprint arXiv:2305.14018 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 163, + 480, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 163, + 480, + 195 + ], + "spans": [ + { + "bbox": [ + 134, + 163, + 480, + 195 + ], + "type": "text", + "content": "19. Liu, H., Teng, Y., Lu, T., Wang, H., Wang, L.: Sparsebev: High-performance sparse 3d object detection from multi-camera videos. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18580-18590 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 197, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 197, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 133, + 197, + 480, + 228 + ], + "type": "text", + "content": "20. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 230, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 230, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 230, + 480, + 262 + ], + "type": "text", + "content": "21. Liu, Y., Wang, T., Zhang, X., Sun, J.: Petr: Position embedding transformation for multi-view 3d object detection. In: European Conference on Computer Vision. pp. 531-548. Springer (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 263, + 480, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 480, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 480, + 296 + ], + "type": "text", + "content": "22. Liu, Y., Yan, J., Jia, F., Li, S., Gao, A., Wang, T., Zhang, X.: Petrov2: A unified framework for 3d perception from multi-camera images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3262-3272 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 297, + 480, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 480, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 480, + 318 + ], + "type": "text", + "content": "23. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 319, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 480, + 352 + ], + "type": "text", + "content": "24. Ma, X., Zhang, Y., Xu, D., Zhou, D., Yi, S., Li, H., Ouyang, W.: Delving into localization errors for monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4721-4730 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 354, + 480, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 386 + ], + "type": "text", + "content": "25. Park, D., Ambrus, R., Guizilini, V., Li, J., Gaidon, A.: Is pseudo-lidar needed for monocular 3d object detection? In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3142-3152 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 387, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 387, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 480, + 430 + ], + "type": "text", + "content": "26. Park, J., Xu, C., Yang, S., Keutzer, K., Kitani, K.M., Tomizuka, M., Zhan, W.: Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. In: The Eleventh International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 432, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 432, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 432, + 480, + 475 + ], + "type": "text", + "content": "27. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 194-210. Springer (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 476, + 480, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 476, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 480, + 508 + ], + "type": "text", + "content": "28. Reading, C., Harakeh, A., Chae, J., Waslander, S.L.: Categorical depth distribution network for monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8555-8564 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 510, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 480, + 542 + ], + "type": "text", + "content": "29. Shrivastava, A., Gupta, A., Girshick, R.: Training region-based object detectors with online hard example mining. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 761-769 (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 544, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 480, + 586 + ], + "type": "text", + "content": "30. Shu, C., Deng, J., Yu, F., Liu, Y.: 3dppe: 3d point positional encoding for transformer-based multi-camera 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 3580-3589 (October 2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 588, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 620 + ], + "type": "text", + "content": "31. Simonelli, A., Bulo, S.R., Porzi, L., López-Antequera, M., Kontschieder, P.: Disentangling monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1991-1999 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "text", + "content": "32. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2446-2454 (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "F. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 643 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "33. Sung, K.K., Poggio, T.: Example-based learning for view-based human face detection. IEEE Transactions on pattern analysis and machine intelligence 20(1), 39-51 (1998)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 194 + ], + "type": "text", + "content": "34. Wang, S., Liu, Y., Wang, T., Li, Y., Zhang, X.: Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 3621-3631 (October 2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "35. Wang, T., Zhu, X., Pang, J., Lin, D.: Fcos3d: Fully convolutional one-stage monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 913-922 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "text", + "content": "36. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor detr: Query design for transformer-based detector. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 2567-2575 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "37. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: Conference on Robot Learning. pp. 180–191. PMLR (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "text", + "content": "38. Wang, Z., Huang, Z., Fu, J., Wang, N., Liu, S.: Object as query: Lifting any 2d object detector to 3d detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3791-3800 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 380 + ], + "type": "text", + "content": "39. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., et al.: Argoverse 2: Next generation datasets for self-driving perception and forecasting. In: Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2) (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "40. Yan, J., Liu, Y., Sun, J., Jia, F., Li, S., Wang, T., Zhang, X.: Cross modal transformer: Towards fast and robust 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18268-18278 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 413, + 481, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 457 + ], + "type": "text", + "content": "41. Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z., Huang, G., Li, H., Qiao, Y., Lu, L., et al.: Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17830-17839 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "text", + "content": "42. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In: The Eleventh International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 490, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 481, + 512 + ], + "type": "text", + "content": "43. Zhang, H., Li, H., Zeng, A., Li, F., Liu, S., Liao, X., Zhang, L.: Introducing depth into transformer-based 3d object detection. arXiv preprint arXiv:2302.13002 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "text", + "content": "44. Zhang, Y., Lu, J., Zhou, J.: Objects are different: Flexible monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3289-3298 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "type": "text", + "content": "45. Zhu, B., Jiang, Z., Zhou, X., Li, Z., Yu, G.: Class-balanced grouping and sampling for point cloud 3d object detection. arXiv preprint arXiv:1908.09492 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "text", + "content": "46. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "type": "text", + "content": "47. Zong, Z., Jiang, D., Song, G., Xue, Z., Su, J., Li, H., Liu, Y.: Temporal enhanced training of multi-view 3d object detector via historical object prediction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 3781-3790 (October 2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 387, + 91, + 447, + 102 + ], + "type": "text", + "content": "Ray Denoising" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Ray-Distance Volume Rendering for Neural Scene Reconstruction/e689642e-0cb9-456e-9b14-7ba1bd1312a3_content_list.json b/2024/Ray-Distance Volume Rendering for Neural Scene Reconstruction/e689642e-0cb9-456e-9b14-7ba1bd1312a3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..153d345a3d2fe76e4211cf71bfea37fa47b627b9 --- /dev/null +++ b/2024/Ray-Distance Volume Rendering for Neural Scene Reconstruction/e689642e-0cb9-456e-9b14-7ba1bd1312a3_content_list.json @@ -0,0 +1,1813 @@ +[ + { + "type": "text", + "text": "Ray-Distance Volume Rendering for Neural Scene Reconstruction", + "text_level": 1, + "bbox": [ + 294, + 140, + 707, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ruihong Yin $^{1}$ , Yunlu Chen $^{2}$ , Sezer Karaoglu $^{1,3}$ , and Theo Gevers $^{1,3}$", + "bbox": [ + 233, + 210, + 769, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of Amsterdam, Amsterdam, The Netherlands", + "bbox": [ + 292, + 238, + 707, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Carnegie Mellon University, Pittsburgh, The United States", + "bbox": [ + 292, + 252, + 704, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 3DUniversum, Amsterdam, The Netherlands", + "bbox": [ + 343, + 267, + 656, + 280 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{r.yin, Th.Gevers}@uva.nl, s.karaoglu@3duniversum.com,", + "bbox": [ + 292, + 281, + 707, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yunluche@andrew.cmu.edu", + "bbox": [ + 410, + 296, + 589, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Existing methods in neural scene reconstruction utilize the Signed Distance Function (SDF) to model the density function. However, in indoor scenes, the density computed from the SDF for a sampled point may not consistently reflect its real importance in volume rendering, often due to the influence of neighboring objects. To tackle this issue, our work proposes a novel approach for indoor scene reconstruction, which instead parameterizes the density function with the Signed Ray Distance Function (SRDF). Firstly, the SRDF is predicted by the network and transformed to a ray-conditioned density function for volume rendering. We argue that the ray-specific SRDF only considers the surface along the camera ray, from which the derived density function is more consistent to the real occupancy than that from the SDF. Secondly, although SRDF and SDF represent different aspects of scene geometries, their values should share the same sign indicating the underlying spatial occupancy. Therefore, this work introduces a SRDF-SDF consistency loss to constrain the signs of the SRDF and SDF outputs. Thirdly, this work proposes a self-supervised visibility task, introducing the physical visibility geometry to the reconstruction task. The visibility task combines prior from predicted SRDF and SDF as pseudo labels, and contributes to generating more accurate 3D geometry. Our method implemented with different representations has been validated on indoor datasets, achieving improved performance in both reconstruction and view synthesis.", + "bbox": [ + 261, + 345, + 738, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Indoor scene reconstruction $\\cdot$ neural radiance fields $\\cdot$ signed ray distance function", + "bbox": [ + 261, + 664, + 738, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 718, + 374, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Indoor 3D scene reconstruction involves using multi-view RGB images as input to generate detailed 3D geometry as output. This task is crucial and has widespread applications in virtual/mixed augmentation, robotics, navigation, and so on. Contrary to a single object, indoor environments contain various elements, such as chairs, walls, tables, and cups, which present a challenge for the task of reconstruction. Traditional reconstruction approaches [1,21,27,33,40,44]", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "use 2D and 3D convolutional neural networks (CNNs) [34, 35] to extract features. They can be classified as depth-based methods [6, 27, 37, 40] and volumetric methods [1, 21, 31, 33, 44]. Depth-based techniques estimate depth for every viewpoint, encountering difficulties like scale uncertainties and inconsistencies in depth across different views. The surfaces created through these techniques can be coarse and imprecise. Unlike depth-based approaches, volumetric techniques utilize 3D CNNs to directly generate 3D geometry, leading to smoother surfaces. Nevertheless, these methods use voxels for scene representation, causing GPU memory usage and computational expenses to grow cubically as voxel resolution increases. Such a limitation restricts the use of volumetric approaches for high-quality scene reconstruction. Moreover, volumetric techniques require 3D annotations for network training, significantly increasing the cost of annotations.", + "bbox": [ + 212, + 146, + 787, + 328 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Implicit representations [18,24,29] encode 3D models as functions of coordinates. Utilizing implicit representations and volume rendering, neural radiance fields (NeRF) [19] achieves remarkable performance in novel view synthesis, even without 3D supervision. Some methods [38,42] exploit the capabilities of NeRF in reconstruction tasks. These methods typically parameterize the volume density as a learnable transformation of the Signed Distance Function (SDF), leading to impressive reconstruction performance. Compared to classic volumetric methods, neural implicit scene representation is more memory-efficient, marking a new trend in reconstruction tasks.", + "bbox": [ + 212, + 330, + 787, + 467 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In multi-object indoor scenes, the SDF from sampled points considers surface points across the entire scene. However, for neural implicit scene reconstruction, along a camera ray, from the camera center to the intersected surface point, the SDF may exhibit fluctuations, leading to several local minimum values. Fig. 1 provides a toy example to illustrate this. In Fig. 1c, volume density yielded from the fluctuating SDF displays several local maximum values, contributing to high weights in Fig. 1d. However, the camera ray $\\overrightarrow{OP}$ targets the green rectangle, while the SDF of the near-surface point $\\mathbf{Q}$ refers to the surface of the blue cylinder. Besides, $\\mathbf{Q}$ is far from the intersected point $\\mathbf{P}$ , but still corresponds to high density and high weight. Nevertheless, during volume rendering that maps the 3D output to 2D space, it is evident that along a ray, points closer to the intersected surface point may carry a higher weight as they are more relevant to actual observations. Consequently, in the multi-object scenes, using the SDF to model the density function along the camera ray may disrupt this relationship, introduce noise, and lead to inaccuracies in 2D results or 3D geometries.", + "bbox": [ + 212, + 470, + 787, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper proposes RS-Recon, a novel indoor scene reconstruction method with the ray-specific density function to address the aforementioned issue. Firstly, our method models the density function as a function of the Signed Ray Distance Function (SRDF). Unlike SDF, SRDF measures the shortest distance to the surface along a camera ray, eliminating the influence of surrounding surfaces that are not located on the given ray. Moreover, the density distribution generated from SRDF attains the local maximum only around the intersected surface, aligning more closely to the concept that points nearer to the surface have more importance. To achieve this, our network predicts both SRDF and SDF, where", + "bbox": [ + 212, + 704, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "R. Yin et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SRDF encompasses the density function, while SDF mainly focuses on describing the 3D surfaces. Secondly, despite SRDF and SDF being defined differently, both of them produce positive values outside the object and negative values within. Therefore, in this paper, a SRDF-SDF consistency loss is devised to ensure they share the same sign. Thirdly, this work introduces a self-supervised visibility task to enhance the prediction of 3D geometry. The visibility task predicts the probability of whether the sampled points are physically visible along the ray in 3D space. The pseudo visibility ground truth is formed by combining the prior of SRDF and SDF from the network itself, without the reliance on multi-view geometry or any additional annotations.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows, (1) This work proposes a novel neural scene reconstruction method that leverages the ray-specific SRDF to model the volume density. Compared to SDF, SRDF better reflects the significance of the real observations. (2) A SRDF-SDF consistency loss is proposed to constrain the alignment between SRDF and SDF, ensuring the same sign inside and outside the object. (3) To enhance the accuracy of 3D geometry, an additional self-supervised visibility task is introduced to predict the visibility probability of sampled points. This task integrates prior information from both the SDF and SRDF predictions as labels. (4) The experiments conducted on both synthetic and real-world indoor datasets illustrate that our method enhances performance in both reconstruction and view synthesis.", + "bbox": [ + 212, + 297, + 789, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 488, + 397, + 505 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Classic Indoor Scene Reconstruction. Traditional methods for indoor scene reconstruction fall into two classes, i.e. depth-based methods [6, 16, 27, 37, 40, 41, 45] and volumetric methods [1, 7, 12, 21, 31, 33, 44]. Depth-based methods predict pixel-level depth for individual frames, and subsequently depth fusion [4] is adopted to generate the 3D scene. To achieve favorable outcomes, these methods often construct a plane sweep cost volume [2, 9] at the feature or image level, leveraging information from multiple views to complement the current frame, e.g. MVSNet [40], DeepVideoMVS [6], and SimpleRecon [27]. Using the depth map as the intermediate representation, these methods encounter challenges such as depth inconsistency and scale ambiguities. In contrast, volumetric methods utilize 3D CNN to directly regress the 3D geometry, which can generate smooth surfaces and reconstruct unobserved regions. For instance, NeuralRecon [33] designs a learnable TSDF fusion module to integrate features from previous frames and predicts the TSDF for sparse volumes. TransformerFusion [1], based on volumetric representation, adopts Transformer to fuse multi-view features and select frames with higher attention. Yin et al. [44] propose to incorporate geometry priors at different levels for the volumetric methods. However, due to the use of 3D CNN in volumetric methods, computational costs escalate significantly with higher resolutions.", + "bbox": [ + 212, + 522, + 787, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural Implicit Scene Representation. With the success of implicit representation and NeRF [19], neural implicit scene representation has emerged as a", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Ray-Distance Volume Rendering for Neural Scene Reconstruction", + "bbox": [ + 294, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0d205758a58093843359625d17fc030815f1eca3d726db456a5906b5306dc3de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 143, + 338, + 205 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/75bd0a2c2b372345578f26a2377ef97ca7ff20f0c644f70e6fd2273b7185103e.jpg", + "image_caption": [ + "(a) Multi-object scene" + ], + "image_footnote": [], + "bbox": [ + 220, + 209, + 344, + 277 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/68ad60a3fce54a38647b70c1fdb4cb37d0440a6c04e04063cc60fe2907447435.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 344, + 143, + 491, + 213 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4d3607c5da3c7578b994bf6b47e117190d1ac2e2699f0aaf50ea6dd49d324810.jpg", + "image_caption": [ + "(b) SDF/SRDF" + ], + "image_footnote": [], + "bbox": [ + 344, + 214, + 491, + 281 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/69282ed5d06f58da24998c4c8a30bbae90c51791cc0878d999b3df40241bdbd8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 143, + 638, + 214 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/09844e35fe2dded36b0d53c850e551c27a62bd2cb180ce8fb9425a33fce69394.jpg", + "image_caption": [ + "(c) Density", + "Fig. 1: Analysis of the density function and weight distribution for a toy scene. (a) Frontal and overhead perspectives of a scene with multiple objects, where a ray originates from the camera center $\\mathbf{O}$ , intersecting with the green rectangle at the point $\\mathbf{P}$ . $\\mathbf{Q}$ is the closest point to the blue cylinder along the ray $\\overrightarrow{OP}$ . (b) The SDF $d_{\\Omega}$ /SRDF $\\tilde{d}_{\\Omega}$ in Eqs. (2) and (4) along the ray $\\overrightarrow{OP}$ . (c) The volume density $\\sigma$ generated from the SDF/SRDF in (b) using Eqs. (3) and (5). (d) The weight distribution $T_{i}\\alpha_{i}$ generated from the density in (c) via Eq. (1). Although $\\mathbf{Q}$ is distant from the surface boundary intersecting with the ray, the density function from the SDF in $\\mathbf{Q}$ generates a high weight in volume rendering, resulting in noisy rendering and reconstruction." + ], + "image_footnote": [], + "bbox": [ + 495, + 214, + 638, + 281 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/31fa9cf154536cc37d162b436657800c9bd388cf03b3e85dce138bb1de71e37e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 143, + 784, + 213 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8df39fee97f9f9281441f8fdffc9be72387a86e21ff05b87a28f63540f2fefb9.jpg", + "image_caption": [ + "(d) Weight" + ], + "image_footnote": [], + "bbox": [ + 640, + 214, + 784, + 281 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "hot spot in the past few years. Generally, reconstruction approaches based on neural implicit scene representation first map point positions to a continuous representation of SDF or occupancy by multi-layer perceptrons (MLPs), then transfer SDF or occupancy to volume density, and finally render the 2D results from the 3D outputs. Despite requiring only 2D supervision during training, these methods can reconstruct high-quality surfaces and are more computationally efficient. For example, VolSDF [42] parameterizes the volume density as a Laplace distribution of SDF, while NeuS [38] models it as the logistic distribution. Some extensions [10,36,39,46,47] introduce extra priors to achieve superior outcomes. For instance, Manhattan-SDF [10] employs planar constraints based on Manhattan-world assumption [3]. NeuRIS [36] proposes to combine normal prior for the reconstruction of texture-less regions, while NeuralRoom [39] implements perturbation-residual restriction to reconstruct the flat region better. MonoSDF [46] exploits both depth and normal cues to improve the reconstruction quality. Although additional priors contribute to the accuracy and completeness of reconstructed surfaces, they employ the SDF to model the volume density, potentially resulting in noisy 2D results in the multi-object scenes. Contrarily, our method advocates for parameterizing the density function with the ray-specific SRDF, resulting in enhanced reconstruction performance.", + "bbox": [ + 212, + 473, + 787, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Some NeRF-based methods also utilize multi-view geometry to introduce additional regularization. For instance, RegNeRF [22] regularizes patched-based depth maps from unobserved viewpoints. Geo-NeuS [8] and ConsistentNeRF [11] use multi-view photometric consistency. SRDF and observation consistencies are used by [48]. GenS [25] introduces multi-view and multi-scale feature consistent", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "R. Yin et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e50b7d5fd484c8d5e91491dca755596776e5e837e5334511d467449d7665368d.jpg", + "image_caption": [ + "Fig. 2: Our framework. A geometry MLP $f_{g}$ is adopted to generate the SDF and geometry features $\\mathbf{F}_{g}$ , utilizing the encoded position (and optionally grid features) as input. Then, $\\mathbf{F}_{g}$ , along with viewing direction, passes through a color MLP $f_{c}$ to predict the color for each point. Notably, (1) our approach advocates for modeling the density function with the ray-specific SRDF in addition to the SDF. For this purpose, a SRDF MLP $f_{s}$ is introduced to generate the SRDF. (2) A SRDF-SDF consistency loss $\\mathcal{L}_{con}$ is devised to align the signs between the generated SRDF and SDF. (3) To enhance the geometry prediction, a self-supervised visibility task is proposed to integrate geometry priors in both SRDF and SDF predicted by the network and generates the pseudo visibility ground truth. The visibility probability is predicted by the SRDF MLP." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 787, + 295 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "cies for neural surface reconstruction. VIP-NeRF [30] adopts a visibility prior, using the plane sweep volume among multiple views to generate visibility labels. However, they depend on the computation of multi-view geometry, some of which may involve processing multiple frames yielding increased computational demands. In contrast, this paper introduces a self-supervised visibility task that requires minimal computational increases and does not depend on multi-view geometry for ground truth generation.", + "bbox": [ + 212, + 476, + 787, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 606, + 330, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we provide the background on NeRF-based reconstruction methods, which typically model the volume density as a transformation of SDF (refer to Sec. 3.1). To overcome the issue of false local maxima in the density from SDF, our approach covered in Sec. 3.2 - Sec. 3.5 proposes to model the density function with the ray-specific SRDF in neural scene reconstruction, which achieves more accurate reconstruction. Fig. 2 presents the framework of our approach.", + "bbox": [ + 212, + 638, + 787, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Background: Volume Rendering Density as Transformed SDF", + "text_level": 1, + "bbox": [ + 212, + 753, + 769, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our goal is to recover the scene geometry given a set of projected 2D views of a 3D scene. To achieve this, NeRF [19] optimizes a continuous 5D function to represent the scene, which predicts the volume density $\\sigma(\\mathbf{p})$ and view-dependent color $\\mathbf{c}(\\mathbf{p},\\mathbf{r})$ for each sampled point $\\mathbf{p} \\in \\mathbb{R}$ and ray direction $\\mathbf{r} \\in S^2$ . With $\\Omega \\subset \\mathbb{R}^3$", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Ray-Distance Volume Rendering for Neural Scene Reconstruction", + "bbox": [ + 294, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "denoting the scene geometry filled with physical presence, the predicted density $\\sigma (\\mathbf{p})$ approximates the ground truth scene density $\\mathbf{1}_{\\varOmega}(\\mathbf{p})$ , such that $\\mathbf{1}_{\\varOmega}(\\mathbf{p}) = 1$ if $\\mathbf{p}\\in \\varOmega$ and 0 otherwise. After that, the classic volume rendering is employed to render the color $\\mathbf{C}(\\mathbf{r})$ , and $N$ points $\\{\\mathbf{p}_i = \\mathbf{o} + z_i\\mathbf{r}, i\\in [1,N]\\}$ are sampled along the camera ray $\\mathbf{r}$ , with $z_{i}$ indicating the depth from the camera center $\\mathbf{o}$ to the sampled point. The color is accumulated along the ray:", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} (\\mathbf {r}) = \\sum_ {i = 1} ^ {N} T _ {i} \\alpha_ {i} \\mathbf {c} \\left(\\mathbf {p} _ {i}, \\mathbf {r}\\right), T _ {i} = \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {i}\\right), \\alpha_ {i} = 1 - \\exp \\left(- \\sigma \\left(\\mathbf {p} _ {i}\\right) \\delta_ {i}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 266, + 247, + 785, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\delta_{i}$ is the interval of neighboring points.", + "bbox": [ + 214, + 296, + 544, + 311 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Despite the great success in novel view synthesis, NeRF has difficulty reconstructing satisfactory actual surfaces from the generated volume density. To address this challenge, recent methods [42,46] propose to predict the SDF and then transform it to volume density, which aligns with the geometry bias that the surface points have a higher density than other non-surface points and corresponds to a higher weight during volume rendering.", + "bbox": [ + 212, + 311, + 782, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For any sampled point $\\mathbf{p} \\in \\mathbb{R}^3$ , the SDF $d_{\\Omega}(\\mathbf{p})$ is with an absolute value representing the shortest distance from $\\mathbf{p}$ to the surface $\\mathcal{M} = \\partial \\Omega$ , while the sign indicates whether the point is outside (positive) or inside (negative) the surface. Formally,", + "bbox": [ + 212, + 402, + 785, + 460 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S D F}: \\quad d _ {\\Omega} (\\mathbf {p}) = (- 1) ^ {\\mathbf {1} _ {\\Omega} (\\mathbf {p})} \\min _ {\\mathbf {p} ^ {*} \\in \\mathcal {M}} \\| \\mathbf {p} - \\mathbf {p} ^ {*} \\| _ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 462, + 784, + 486 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the indicator function $\\mathbf{1}_{\\varOmega}(\\mathbf{p}) = 1$ if $\\mathbf{p}\\in \\varOmega$ and 0 otherwise. This indicates that the SDF considers every surface around the point. To apply SDF to volume rendering, Yariv et al. [42] suggest to derive the volume density $\\sigma^{\\mathrm{SDF}}$ from the signed distance as follows:", + "bbox": [ + 212, + 489, + 787, + 551 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma^ {\\mathrm {S D F}} (\\mathbf {p}) = \\alpha \\Psi_ {\\beta} (- d _ {\\Omega} (\\mathbf {p})) \\text {,} \\Psi_ {\\beta} (s) = \\left\\{ \\begin{array}{l l} \\frac {1}{2} \\exp \\left(\\frac {s}{\\beta}\\right) & \\text {i f} s \\leq 0 \\\\ 1 - \\frac {1}{2} \\exp \\left(- \\frac {s}{\\beta}\\right) & \\text {i f} s > 0 \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 271, + 559, + 785, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in which $\\alpha, \\beta$ are learnable parameters. $\\Psi_{\\beta}$ is the Cumulative Distribution Function (CDF) of the Laplace distribution with zero mean and $\\beta$ scale. Then, $\\sigma^{\\mathrm{SDF}}(\\mathbf{p})$ is applied to Eq. (1) and as the density function $\\sigma$ .", + "bbox": [ + 212, + 614, + 787, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Accurate Scene Volume Density from Ray-Specific SRDF", + "text_level": 1, + "bbox": [ + 214, + 680, + 735, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We illustrate the issues with volume rendering SDF by a toy example in Fig. 1. Indoor scenes typically include multiple objects. Along the camera ray $\\mathbf{r}$ , the SDF may be influenced by surfaces not intersecting the ray, resulting in several ambiguous local maxima of false high volume density and weight. This implies that the 2D observations may have a strong correlation to these distant points, contradicting the fact that the points closer to the surface along the camera ray may carry more significance.", + "bbox": [ + 212, + 703, + 787, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To this end, we propose volume rendering scene density function from SRDF, the distance field local to the ray direction. In contrast to the SDF in Eq. (2),", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "R. Yin et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the SRDF $\\tilde{d_{\\Omega}}$ ( $\\mathbf{p}, \\mathbf{r}$ ) [26,48] computes the shortest distance from the point to the surface along the sampled ray $\\mathbf{r}$ , i.e. the SRDF is ray-dependent. Formally, the SRDF $\\tilde{d_{\\Omega}}$ ( $\\mathbf{p}, \\mathbf{r}$ ) can be represented by", + "bbox": [ + 212, + 146, + 787, + 195 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\text {S R D F :} \\quad \\tilde {d _ {\\Omega}} (\\mathbf {p}, \\mathbf {r}) = (- 1) ^ {\\mathbf {1} _ {\\Omega} (\\mathbf {p})} \\min _ {(\\mathbf {p} + \\rho \\mathbf {r}) \\in \\mathcal {M}; \\rho \\in \\mathbb {R}} | \\rho |, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 205, + 785, + 232 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{1}_{\\Omega}(\\mathbf{p}) = 1$ if $\\mathbf{p} \\in \\Omega$ and $0$ otherwise. In the following sections, $d_{\\Omega}$ from Eq. (2) and $\\tilde{d_{\\Omega}}$ are used to denote SDF and SRDF for brevity.", + "bbox": [ + 212, + 243, + 785, + 275 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig. 1, it can be seen that the weight computed from SRDF in the volume rendering only reaches a local peak around the surface point, aligning more consistently with the nature of the 2D observations. Therefore, this paper proposes to employ the ray-specific SRDF to yield the volume density instead of the SDF. This work predicts both SDF and SRDF, where SDF is mainly used to model the surface. As depicted in Fig. 2, for a sampled point $\\mathbf{p}$ , a geometry MLP $f_{g}$ is applied on the input (including the encoded position or features) to predict SDF $d_{\\Omega}$ and geometry feature $\\mathbf{F}_{g}$ . By definition, SRDF is view-dependent. Hence, the viewing direction $\\mathbf{r}$ , geometry feature $\\mathbf{F}_{g}$ , and point position $\\mathbf{p}$ are concatenated and then passed through a SRDF MLP $f_{s}$ to yield the SRDF. On the other hand, the viewing direction $\\mathbf{r}$ , geometry feature $\\mathbf{F}_{g}$ , point position $\\mathbf{p}$ , and 3D unit normal $\\mathbf{n} = \\nabla d_{\\Omega}$ are used as input to a color MLP $f_{c}$ to output color $\\mathbf{c}$ .", + "bbox": [ + 212, + 276, + 787, + 457 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For volume rendering, our method adopts the predicted SRDF $\\tilde{d_{\\Omega}}$ to derive the volume density $\\sigma$ . This is defined by:", + "bbox": [ + 212, + 458, + 787, + 489 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma^ {\\mathrm {S R D F}} (\\mathbf {p}, \\mathbf {r}) = \\alpha \\Psi_ {\\beta} \\left(- \\tilde {d _ {\\Omega}} (\\mathbf {p}, \\mathbf {r})\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 500, + 785, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Then, $\\sigma^{\\mathrm{SRDF}}(\\mathbf{p},\\mathbf{r})$ is applied to the volume rendering in Eq. (1) as a ray-conditioned density function to obtain the color value $\\mathbf{C}^{\\mathrm{SRDF}}(\\mathbf{r})$ . Besides, during training, we also render color $\\mathbf{C}^{\\mathrm{SDF}}(\\mathbf{r})$ with density functions from SDF to obtain gradient signals to optimize the prediction of SDF, in which the volume rendering processes use the same rays and points as those for $\\mathbf{C}^{\\mathrm{SRDF}}(\\mathbf{r})$ .", + "bbox": [ + 212, + 535, + 787, + 611 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the scene geometry, after the network is trained with NeRF optimization targets, the geometric surface $\\mathcal{M} = \\partial \\Omega$ is extracted as watertight meshes by the marching cubes algorithm [15,24] from the SDF results $d_{\\Omega}$ , which only depends on the query location $\\mathbf{p}$ .", + "bbox": [ + 212, + 612, + 787, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 SRDF-SDF Consistency Loss", + "text_level": 1, + "bbox": [ + 214, + 694, + 503, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following the definitions in Eq. (4), despite SRDF and SDF being defined differently, they share the same sign, which indicates the spatial occupancy — positive outside the surface and negative inside the surface. However, in this paper, where SDF and SRDF are predicted from separate branches, sign consistency is not guaranteed. To deal with this issue, a SRDF-SDF consistency loss is proposed to enforce this constraint.", + "bbox": [ + 212, + 718, + 787, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Although the sign function effectively computes the output's sign, it cannot be differentiated during back-propagation. To a certain degree, the values derived", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Ray-Distance Volume Rendering for Neural Scene Reconstruction", + "bbox": [ + 294, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "from the sigmoid function approximate those calculated by the sign function. Thus, in our SRDF-SDF consistency loss, the sigmoid function is first employed to generate the sign of SRDF and SDF as the approximated sign function $\\varsigma (\\cdot)$ , after which a $\\ell_2$ loss is adopted to quantify the sign difference, as defined by:", + "bbox": [ + 212, + 146, + 787, + 208 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c o n} = \\frac {1}{N _ {r}} \\sum_ {\\mathbf {p}, \\mathbf {r}} M _ {c o n} | | \\varsigma (\\tilde {d} _ {\\Omega}) - \\varsigma (d _ {\\Omega}) | | _ {2} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 219, + 785, + 256 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\varsigma (d) = \\mathsf {S i g m o i d} (k \\cdot d), M _ {c o n} = [ \\widetilde {d} _ {\\Omega} \\cdot d _ {\\Omega} < 0 ]\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 258, + 658, + 277 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in which $k$ is a hyperparameter, controlling the slope of the sigmoid function. The indicator function $[\\tilde{d}_{\\Omega} \\cdot d_{\\Omega} < 0]$ is leveraged to identify whether the generated SDF and SRDF have opposite signs. Therefore, the consistency loss imposes penalties when there is a difference between the signs of SRDF and SDF. $N_r$ represents the number of sampled points in a minibatch that satisfy $[\\tilde{d}_{\\Omega} \\cdot d_{\\Omega} < 0]$ .", + "bbox": [ + 212, + 287, + 787, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Taking the derivative of $\\mathcal{L}_{con}$ with respect to SDF $d_{\\Omega}$ (as an example), is computed as follows:", + "bbox": [ + 212, + 367, + 787, + 397 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\mathcal {L} _ {c o n}}{\\partial d _ {\\Omega}} = - \\frac {2}{N _ {r}} M _ {c o n} | | \\varsigma (\\tilde {d} _ {\\Omega}) - \\varsigma (d _ {\\Omega}) | | _ {1} \\cdot \\frac {\\partial \\varsigma (d _ {\\Omega})}{\\partial d _ {\\Omega}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 409, + 785, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The derivative in Eq. (7) shows that our consistency loss has two advantages: (1) Unlike the sign function, which assigns the same penalty for outputs with different signs of SRDF and SDF, our consistency loss can adjust the penalty according to the extent of the sign discrepancy between SRDF and SDF, referring to $||\\varsigma (\\tilde{d}_{\\varOmega}) - \\varsigma (d_{\\varOmega})||_1$ . (2) The derivative of the sigmoid function $\\frac{\\partial\\varsigma(d_{\\varOmega})}{\\partial d_{\\varOmega}}$ peaks at zero and decreases from 0 to $+\\infty / - \\infty$ . Consequently, predictions in proximity to zero (indicating the surface) may exhibit higher absolute gradient values, providing effective supervision for points near the surface.", + "bbox": [ + 212, + 450, + 787, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Self-supervised Visibility Task", + "text_level": 1, + "bbox": [ + 214, + 597, + 511, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Theoretically, with visibility attributes of 3D points, surface location becomes straightforward. Therefore, to enhance 3D geometry prediction, we propose a self-supervised visibility task in this paper. Along the trajectory of a camera ray, points sampled prior to reaching the first surface point are categorized as visible, whereas points sampled beyond the first surface point are classified as occluded. To find the first surface point along the ray, using SRDF as an example, the multiplication of SRDF between adjacent sampled points is computed. When $\\tilde{d}_{\\Omega}^{i}\\cdot \\tilde{d}_{\\Omega}^{i + 1}\\leq 0$ and $\\tilde{d}_{\\Omega}^{m}\\cdot \\tilde{d}_{\\Omega}^{m + 1} > 0,m\\in [1,i - 1]$ , it indicates that the first surface is within the interval $[\\mathbf{o} + z_i\\mathbf{r},\\mathbf{o} + z_{i + 1}\\mathbf{r}]$ . Consequently, the points between $[\\mathbf{p}_1,\\mathbf{p}_i]$ are considered physically visible while points between $[\\mathbf{p}_{i + 1},\\mathbf{p}_N]$ are regarded as occluded. However, the SRDF or SDF predicted by the network may be inaccurate at some points, introducing potential noise in the visibility task. To improve the accuracy of the visibility labeling, both SDF and SRDF are utilized in determining visibility. Points are labeled as visible or occluded only when there", + "bbox": [ + 212, + 623, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "R. Yin et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/430426dfbbdd3403ac8e18f9c30af6c2477d37d2fe5261d7b07cb31962f98775.jpg", + "table_caption": [ + "Table 1: Evaluation of 3D reconstruction meshes on ScanNet. $\\dagger$ indicates our reproduced results. The best results are bold. $\\downarrow$ denotes that a lower value is preferable while $\\uparrow$ means higher value corresponds to better performance." + ], + "table_footnote": [], + "table_body": "
MethodAcc ↓Comp ↓Prec ↑Recall ↑F-score ↑
COLMAP [28]0.0470.2350.7110.4410.537
UNISURF [23]0.5540.1640.2120.3620.267
NeuS [38]0.1790.2080.3130.2750.291
VolSDF [42]0.4140.1200.3210.3940.346
Manhattan-SDF [10]0.0720.0680.6210.5860.602
S3P [43]0.0550.0590.7090.6600.683
NeuRIS [36]0.0500.0490.7170.6690.692
MonoSDF_Grid [46]0.0720.0570.6600.6010.626
MonoSDF_MLP [46]0.0350.0480.7990.6810.733
HelixSurf [14]0.0380.0440.7860.7270.755
Occ_SDF_Hybrid† [17]0.0400.0410.7830.7480.765
Ours_Grid0.0740.0490.6700.7030.683
Ours_MLP0.0400.0400.8090.7790.794
", + "bbox": [ + 246, + 199, + 754, + 405 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "is agreement between SRDF and SDF on their classification. If the classifications from SRDF and SDF diverge, suggesting difficulties in accurately categorizing these points for the network, such points are omitted from the visibility task. The method for assigning visibility labels is defined as follows", + "bbox": [ + 212, + 431, + 787, + 492 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nV _ {g t} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} (V ^ {\\mathrm {S R D F}} = 1) \\& (V ^ {\\mathrm {S D F}} = 1) \\\\ 0, & \\text {i f} (V ^ {\\mathrm {S R D F}} = 0) \\& (V ^ {\\mathrm {S D F}} = 0) \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 502, + 785, + 542 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $V^{\\mathrm{SRDF}} / V^{\\mathrm{SDF}}$ are the visibility labels determined based on the prediction of SRDF/SDF. 1/0 means label for visible/occluded points. The formation of the visibility label can be inferred as self-supervised, thus eliminating the necessity for multi-view geometry and additional annotations. Additionally, it incorporates information from both SRDF and SDF.", + "bbox": [ + 212, + 551, + 787, + 627 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "From the analysis above, similar to SRDF, the visibility task also relies on the view direction. Therefore, a visibility probability $V_{pred}$ is predicted from the SRDF branch. It is a binary classification task, i.e. visible or occluded, so the binary cross-entropy loss is employed as visibility loss $\\mathcal{L}_{vis}$ during optimization.", + "bbox": [ + 212, + 628, + 787, + 689 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The visibility task can discern the point's visibility along the camera ray, thereby influencing the learning of SRDF. Moreover, the optimization process integrates knowledge of ray-related visibility across the entire space, further impacting the learning of SDF. Notably, unlike 2D supervision, the visibility task provides priors in 3D space, imparting a more direct influence on 3D geometry.", + "bbox": [ + 212, + 689, + 787, + 765 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.5 Optimization", + "text_level": 1, + "bbox": [ + 214, + 785, + 372, + 800 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Loss Functions. During training, the overall loss consists of RGB loss $\\mathcal{L}_c$ , normal loss $\\mathcal{L}_n$ , depth loss $\\mathcal{L}_d$ , Eikonal loss $\\mathcal{L}_e$ , smooth loss $\\mathcal{L}_s$ , SRDF-SDF", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Ray-Distance Volume Rendering for Neural Scene Reconstruction", + "bbox": [ + 294, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "consistency loss $\\mathcal{L}_{con}$ , and visibility loss $\\mathcal{L}_{vis}$", + "bbox": [ + 215, + 146, + 542, + 162 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {c} + \\lambda_ {n} \\mathcal {L} _ {n} + \\lambda_ {d} \\mathcal {L} _ {d} + \\lambda_ {e} \\mathcal {L} _ {e} + \\lambda_ {s} \\mathcal {L} _ {s} + \\lambda_ {c o n} \\mathcal {L} _ {c o n} + \\lambda_ {v i s} \\mathcal {L} _ {v i s} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 284, + 174, + 785, + 189 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "in which $\\lambda_{(\\cdot)}$ are the loss weight. Following MonoSDF [46], the 2.5D depth and normal maps are volume rendered using the density function $\\sigma^{\\mathrm{SRDF}}$ to exploit geometric cues and support training. More details on the computation of loss functions are provided in the supplementary material.", + "bbox": [ + 212, + 199, + 784, + 261 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation Details. Theoretically, our method can be applied to replace SDF-based volume rendering in indoor scene reconstruction. To measure our method, VolSDF-based reconstruction method MonoSDF is considered as our baseline in the following experiments. Furthermore, in the supplementary material, we also verify our method on NeuS-based reconstruction method NeuRIS. We used different spatial encoding techniques as input, including hash features [20] and Fourier features [19], referred to as Grid and MLP respectively in the following experiments. More details are given in the supplementary material.", + "bbox": [ + 212, + 261, + 785, + 383 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 405, + 375, + 422 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 436, + 426, + 452 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. Our proposed method is evaluated on three indoor datasets, comprising: (1) four scenes from the real-world dataset ScanNet [5], (2) eight scenes from the synthetic dataset Replica [32], and (3) four scenes from the advanced set of the real-world large-scale dataset Tanks and Temples [13].", + "bbox": [ + 212, + 460, + 784, + 521 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Metrics. For ScanNet, the assessment of the results involves five reconstruction metrics: accuracy (acc), completeness (comp), precision (prec), recall, and F-score with a threshold of 5cm. For Replica, in alignment with MonoSDF, metrics including normal consistency (normal c.), chamfer L1 distance (Chamfer-L1), and F-score with a threshold of 5cm are utilized. For Tanks and Temples, as evaluated by the official server, results are reported based on the F-score with a 1cm threshold. Additionally, this study assesses the peak signal-to-noise ratio (PSNR) of the rendered 2D RGBs to showcase the effectiveness of our approach. Detailed computations for all metrics are given in the supplementary material.", + "bbox": [ + 212, + 522, + 785, + 656 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baselines. Except for the baseline MonoSDF, our method is also compared to (1) the classic MVS method: COLMAP [28], (2) neural implicit reconstruction methods: UNISURF [23], NeuS [38], VolSDF [42], Manhattan-SDF [10], NeuRIS [36], S3P [43], HelixSurf [14], and Occ_SDF_Hybrid [17].", + "bbox": [ + 214, + 657, + 785, + 719 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Evaluation Results", + "text_level": 1, + "bbox": [ + 215, + 739, + 418, + 753 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Reconstruction on ScanNet [5]. Tab. 1 compares our method against other state-of-the-art (SOTA) methods on ScanNet. When contrasted with the baseline MonoSDF, our method consistently shows better performance across all metrics, regardless of whether utilizing MLP or grid representation. For example, employing grid representation, our method exhibits a $5.7\\%$ higher F-score than", + "bbox": [ + 212, + 763, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "R. Yin et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/3aadb8698266b3f348af8bc1a20a549989e6fb0f3c1195c790e6de5f6b412399.jpg", + "image_caption": [ + "Fig. 3: Qualitative comparisons on ScanNet. It can be seen that our method can reconstruct more surfaces, especially in thin regions." + ], + "image_footnote": [], + "bbox": [ + 233, + 143, + 772, + 306 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/08b9dbfc9bc541e6db6461ca87301e96a223958047e861a36a3ec8920f0fbbfc.jpg", + "image_caption": [ + "Fig.4: Qualitative comparisons on Replica. Compared to MonoSDF and Occ_SDF_Hybrid, our method can generate more accurate surfaces." + ], + "image_footnote": [], + "bbox": [ + 236, + 359, + 767, + 570 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "MonoSDF. Using the MLP representation, our method outperforms MonoSDF by $1.0\\%$ in recall, $9.8\\%$ in precision, and $6.1\\%$ in F-score. In addition, our method with the MLP representation surpasses other SOTA methods on almost all metrics, explaining the effectiveness of our method on reconstruction. Notably, our method achieves a $2.9\\%$ higher F-score compared to Occ_SDF_Hybrid.", + "bbox": [ + 212, + 643, + 787, + 719 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Reconstruction on Replica [32]. The results on the synthetic dataset Replica are shown in Tab. 2. Using the grid representation, our method reports a notable improvement of $9.91\\%$ in chamfer distance and $2.88\\%$ in F-score. Similar increases are also illustrated in our method with the MLP representation across all metrics. Our approach also achieves on-par results with Occ_SDF_Hybrid. Reconstruction on Tanks and Temples [13]. Tab. 3 shows quantitative outcomes on the challenging large-scale dataset Tanks and Temples. The F-score for each scene - Auditorium, Ballroom, Courtroom, and Museum - along with", + "bbox": [ + 212, + 719, + 787, + 842 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Ray-Distance Volume Rendering for Neural Scene Reconstruction", + "bbox": [ + 294, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/3d0b87a6b51ec1dfaf1da2ab1e37bb25319f0772212318fb386dd769af87544c.jpg", + "table_caption": [ + "Table 2: Evaluation of 3D reconstruction meshes on Replica." + ], + "table_footnote": [], + "table_body": "
MethodNormal C. ↑Chamfer-L1 ↓F-score ↑
GridMonoSDF [46]90.933.2385.91
Ours91.662.9188.79
MLPMonoSDF [46]92.112.9486.18
Occ_SDF_Hybrid†[17]93.342.5891.54
Ours93.492.6091.72
", + "bbox": [ + 245, + 170, + 756, + 267 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/d02b0c992fc4062be2766c1c396d7990d17809bbfd520410d52dd8eaa4ccc8e5.jpg", + "table_caption": [ + "Table 3: Evaluation of 3D reconstruction meshes on the advanced set of Tanks and Temples. The F-score is reported in the table." + ], + "table_footnote": [], + "table_body": "
MethodAuditoriumBallroomCourtroomMuseummean
GridMonoSDF [46]3.173.7013.755.686.58
Occ_SDF_Hybrid†[17]3.763.5814.047.277.16
Ours4.844.3114.187.607.73
MLPMonoSDF [46]3.092.4710.005.105.17
Occ_SDF_Hybrid†[17]3.213.3410.294.665.38
Ours4.244.6910.795.916.41
", + "bbox": [ + 217, + 320, + 784, + 431 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "their mean, is reported. Our method, with both the MLP and grid representations, outperforms Occ_SDF_Hybrid and MonoSDF. In particular, our F-score is $1.15\\%$ higher (grid representation) and $1.24\\%$ higher (MLP representation) than MonoSDF across all scenes.", + "bbox": [ + 212, + 459, + 784, + 518 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Visualization of Reconstruction Results. Fig. 3 and Fig. 4 report qualitative results on ScanNet and Replica respectively. In contrast to other methods, our approach can produce superior meshes and recall more regions, particularly for thin regions/small objects, such as chair legs and table legs. Although our method relies on the SRDF for improved density in volume rendering while the surface geometry is extracted by the SDF, both the proposed SRDF-SDF consistency loss and self-supervised visibility task guide the optimization of the SDF to be more consistent with the SRDF, resulting in more accurate surface boundary. Additional visualization is provided in the supplementary material.", + "bbox": [ + 212, + 521, + 784, + 656 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "View Synthesis. Tab. 4 also provides a comparison of view synthesis performance on training views. Notably, 2D views are rendered with density derived", + "bbox": [ + 212, + 657, + 785, + 688 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/e9025e3334bdf47766ab3bf1ae0e77aa5364483421541d8931b27d82bde1d9f4.jpg", + "table_caption": [ + "Table 4: Evaluation of the rendering quality. PSNR is given in this table." + ], + "table_footnote": [], + "table_body": "
MethodScanNetReplicaTanks and Temples
GridMonoSDF [46]28.9640.2528.45
Ours30.0942.0429.52
MLPMonoSDF [46]26.4034.4524.13
Occ_SDF_Hybrid†[17]26.9835.5024.72
Ours27.7736.0625.47
", + "bbox": [ + 230, + 741, + 772, + 838 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "R. Yin et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/c1abc32ebde24c4c854e55fa0bf0ffee651cd00078e0e47a5432f56de0aa0ec8.jpg", + "table_caption": [ + "Table 5: Ablation study of our method on ScanNet." + ], + "table_footnote": [], + "table_body": "
MethodAcc ↓Comp ↓Prec ↑Recall ↑F-score ↑
aBaseline0.0350.0480.7990.6810.733
bRS. Den.0.0400.0440.7720.7200.745
cRS. Den. + Con. L.0.0390.0410.7940.7600.776
dRS. Den. + Con. L. + Vis.(SDF)0.0390.0400.8090.7720.789
eRS. Den. + Con. L. + Vis.(SRDF)0.0420.0400.8040.7740.788
fRS. Den. + Con.L. + Vis.0.0400.0400.8090.7790.794
", + "bbox": [ + 228, + 171, + 776, + 281 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/47c852d8d7fb25b9d71ec8745fe40418576582749129703ee251432cb1116e34.jpg", + "image_caption": [ + "(a) Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 233, + 295, + 352, + 387 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a2595ef98b720e75b2bd727038e66e23d5e0e600ad803fff3f8b25a94331f7c8.jpg", + "image_caption": [ + "(b) RGB from MonoSDF" + ], + "image_footnote": [], + "bbox": [ + 354, + 297, + 470, + 386 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b7db0289271cdaa5b2bbb3423091a8fd4500f567bbf703f076b146de3b417aca.jpg", + "image_caption": [ + "For the red sampled point", + "(c) Weight from MonoSDF" + ], + "image_footnote": [], + "bbox": [ + 472, + 304, + 617, + 385 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f336e9428c36da7a0b2299a287e07165dc751b92244ed3dace9a070b7100cc65.jpg", + "image_caption": [ + "For the yellow sampled point" + ], + "image_footnote": [], + "bbox": [ + 617, + 304, + 766, + 385 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/bd37c51d43d8ca27ac0ca812cd12d6a7cbf3ed7d260c72f21be371fb2fbd4f39.jpg", + "image_caption": [ + "(d) RGB from ours" + ], + "image_footnote": [], + "bbox": [ + 354, + 398, + 472, + 489 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b3c184dcbd608d223a7a827361fde777cec5fa6821ed34ad5ccd6e2e3fa890f3.jpg", + "image_caption": [ + "For the red sampled point", + "(e) Weight from ours", + "Fig.5: Comparison of the yielded weight and image. The weight and image generated by our approach align more closely with the actual observations." + ], + "image_footnote": [], + "bbox": [ + 472, + 410, + 617, + 491 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e09872b627a964405e7318b2d99e091a4aca5b25d2ffb51c68034bdca832ccd7.jpg", + "image_caption": [ + "For the yellow sampled point" + ], + "image_footnote": [], + "bbox": [ + 617, + 410, + 766, + 491 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "from SRDF. On the three datasets, our method outperforms the baseline and Occ_SDF_Hybrid in rendering training views. Overall, our method not only enhances surface reconstruction but also generates more accurate views. The visualization can be found in the supplementary material.", + "bbox": [ + 212, + 575, + 787, + 636 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 661, + 388, + 676 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Tab. 5 presents an ablation study to evaluate the effectiveness of our design on ScanNet with MLP representation. The ablation experiments consist of: (a) Baseline: MonoSDF. (b) RS. Den.: this structure models the volume density using the ray-specific SRDF, without the SRDF-SDF consistency loss and self-supervised visibility task. (c) RS. Den. + Con. L.: Building upon the structure in (b), this setup adopts the proposed SRDF-SDF consistency loss. (d) RS. Den. + Con. L. + Vis.(SDF): based on (c), this configuration adds the self-supervised visibility task, in which only SDF is used to compute the visibility label. (e) RS. Den. + Con. L. + Vis.(SRDF): Different from structure (d), this method relies solely on SRDF to compute the visibility label. (f) RS. Den. + Con. L. + Vis.:", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Ray-Distance Volume Rendering for Neural Scene Reconstruction", + "bbox": [ + 294, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "this structure represents the comprehensive model of our method, wherein the visibility labels integrate priors from both SRDF and SDF predictions.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The comparison between configurations (a) and (b) highlights that our design using SRDF to model the density, without any additional constraints, results in a better performance compared to MonoSDF. The SRDF-SDF consistency loss in (c) surpasses the structure in (b) by $3.1\\%$ in F-score. This shows the significance of sign consistency. Compared to (c), the self-supervised visibility task in (d)-(f) enhances the reconstruction performance. In particular, structure (f) obtains a better F-score than structures (d) and (e). These explain the effectiveness of our self-supervised visibility task and the computation of visibility labels integrating both SRDF and SDF priors. The visualization of the ablation study is given in the supplementary material.", + "bbox": [ + 212, + 176, + 785, + 327 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Furthermore, Fig. 5 compares the weight distribution for the red and yellow sampled points generated by MonoSDF and our method. (1) The analysis of the yellow sampled point demonstrates the benefits of our method for reconstruction: MonoSDF generates high weights and negative SDF values when sampled rays are near the white wall, resulting in false surfaces that significantly deviate from the ground truth. In contrast, our method yields more accurate surfaces and thus enhances both completeness and recall. (2) The analysis of the red sampled point explains the reason why our method can generate accurate views: MonoSDF produces weights with two local maxima, causing an inaccurate RGB value. This occurs because the SDF along the camera ray fluctuates, resulting in varying density and weight. In comparison, our method produces uni-modal weights that peak only around the surface, resulting in more accurate views. The quantitative PSNR in Fig. 5b and Fig. 5d confirms that our rendered image is more accurate compared to MonoSDF. These findings align with our motivation to achieve a more consistent and accurate weight distribution, thereby reconstructing better surfaces and generating high-quality views. This underscores the effectiveness of using ray-specific SRDF to model volume density in our approach.", + "bbox": [ + 212, + 329, + 787, + 585 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 609, + 359, + 626 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work proposes a novel method for neural indoor scene reconstruction, modeling density function with ray-specific SRDF. Firstly, this study analyzed that using the SDF to parameterize the volume density may introduce noises in multi-object indoor scenes, and the SRDF exhibits stronger relationships with the actual 2D observations. Motivated by this, this work employed SRDF to model the volume density, while SDF mainly focuses on the representation of the 3D surface. Secondly, a SRDF-SDF consistency loss has been introduced to align the sign between SRDF and SDF. Thirdly, a self-supervised visibility task has been designed to distinguish whether the sampled 3D point is physically visible or occluded. It combined the prior from SDF and SRDF as visibility labels to improve geometry reconstruction without any other annotations. Experimental results on real-world and synthetic datasets showed that our method excels not only in reconstructing better surfaces but also in generating more accurate views.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "R. Yin et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bozic, A., Palafox, P., Thies, J., Dai, A., Nießner, M.: Transformerfusion: Monocular rgb scene reconstruction using transformers. Advances in Neural Information Processing Systems 34, 1403-1414 (2021)", + "2. Collins, R.T.: A space-sweep approach to true multi-image matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 358-363. IEEE (1996)", + "3. Coughlan, J.M., Yuille, A.L.: Manhattan world: Compass direction from a single image by bayesian inference. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. vol. 2, pp. 941-947. IEEE (1999)", + "4. Curless, B., Levoy, M.: A volumetric method for building complex models from range images. In: Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques. pp. 303-312 (1996)", + "5. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5828-5839 (2017)", + "6. Duzceker, A., Galliani, S., Vogel, C., Speciale, P., Dusmanu, M., Pollefeys, M.: Deepvideomvs: Multi-view stereo on video with recurrent spatio-temporal fusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15324-15333 (2021)", + "7. Feng, Z., Yang, L., Guo, P., Li, B.: Cvrecon: Rethinking 3d geometric feature learning for neural reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17750-17760 (2023)", + "8. Fu, Q., Xu, Q., Ong, Y.S., Tao, W.: Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. Advances in Neural Information Processing Systems 35, 3403-3416 (2022)", + "9. Gallup, D., Frahm, J.M., Mordohai, P., Yang, Q., Pollefeys, M.: Real-time plane-sweeping stereo with multiple sweeping directions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1-8. IEEE (2007)", + "10. Guo, H., Peng, S., Lin, H., Wang, Q., Zhang, G., Bao, H., Zhou, X.: Neural 3d scene reconstruction with the manhattan-world assumption. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5511-5520 (2022)", + "11. Hu, S., Zhou, K., Li, K., Yu, L., Hong, L., Hu, T., Li, Z., Lee, G.H., Liu, Z.: Consistentnerf: Enhancing neural radiance fields with 3d consistency for sparse view synthesis. arXiv preprint arXiv:2305.11031 (2023)", + "12. Ju, J., Tseng, C.W., Bailo, O., Dikov, G., Ghafoorian, M.: Dg-recon: Depth-guided neural 3d scene reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18184-18194 (2023)", + "13. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics 36(4), 1-13 (2017)", + "14. Liang, Z., Huang, Z., Ding, C., Jia, K.: Helixsurf: A robust and efficient neural implicit surface learning of indoor scenes with iterative intertwined regularization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13165-13174 (2023)", + "15. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3d surface construction algorithm. In: Seminal graphics: pioneering efforts that shaped the field, pp. 347-353 (1998)" + ], + "bbox": [ + 218, + 172, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Ray-Distance Volume Rendering for Neural Scene Reconstruction", + "bbox": [ + 294, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Luo, K., Guan, T., Ju, L., Huang, H., Luo, Y.: P-mvsnet: Learning patch-wise matching confidence aggregation for multi-view stereo. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10452-10461 (2019)", + "17. Lyu, X., Dai, P., Li, Z., Yan, D., Lin, Y., Peng, Y., Qi, X.: Learning a room with the occ-sdf hybrid: Signed distance function mingled with occupancy aids scene representation. arXiv preprint arXiv:2303.09152 (2023)", + "18. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4460-4470 (2019)", + "19. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)", + "20. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics 41(4), 1-15 (2022)", + "21. Murez, Z., Van As, T., Bartolozzi, J., Sinha, A., Badrinarayanan, V., Rabinovich, A.: Atlas: End-to-end 3d scene reconstruction from posed images. In: Proceedings of the European Conference on Computer Vision. pp. 414-431. Springer (2020)", + "22. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5480-5490 (2022)", + "23. Oechsle, M., Peng, S., Geiger, A.: Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5589-5599 (2021)", + "24. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 165-174 (2019)", + "25. Peng, R., Gu, X., Tang, L., Shen, S., Yu, F., Wang, R.: Gens: Generalizable neural surface reconstruction from multi-view images. In: Advances in Neural Information Processing Systems (2023)", + "26. Ren, Y., Zhang, T., Pollefeys, M., Susstrunk, S., Wang, F.: Volrecon: Volume rendering of signed ray distance functions for generalizable multi-view reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16685-16695 (2023)", + "27. Sayed, M., Gibson, J., Watson, J., Prisacariu, V., Firman, M., Godard, C.: Simplex: 3d reconstruction without 3d convolutions. In: Proceedings of the European Conference on Computer Vision. pp. 1-19. Springer (2022)", + "28. Schonberger, J.L., Zheng, E., Frahm, J.M., Pollefeys, M.: Pixelwise view selection for unstructured multi-view stereo. In: Proceedings of the European Conference on Computer Vision. pp. 501-518. Springer (2016)", + "29. Sitzmann, V., Martel, J., Bergman, A., Lindell, D., Wetzstein, G.: Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems 33, 7462-7473 (2020)", + "30. Somraj, N., Soundararajan, R.: Vip-nerf: Visibility prior for sparse input neural radiance fields. arXiv preprint arXiv:2305.00041 (2023)", + "31. Stier, N., Rich, A., Sen, P., Höllerer, T.: Vortex: Volumetric 3d reconstruction with transformers for voxelwise view selection and fusion. In: 2021 International Conference on 3D Vision. pp. 320-330. IEEE (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "R. Yin et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019)", + "33. Sun, J., Xie, Y., Chen, L., Zhou, X., Bao, H.: Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15598-15607 (2021)", + "34. Tan, M., Chen, B., Pang, R., Vasudevan, V., Sandler, M., Howard, A., Le, Q.V.: Mnasnet: Platform-aware neural architecture search for mobile. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2820-2828 (2019)", + "35. Tang, H., Liu, Z., Zhao, S., Lin, Y., Lin, J., Wang, H., Han, S.: Searching efficient 3d architectures with sparse point-voxel convolution. In: Proceedings of the European Conference on Computer Vision. pp. 685–702. Springer (2020)", + "36. Wang, J., Wang, P., Long, X., Theobalt, C., Komura, T., Liu, L., Wang, W.: Neuris: Neural reconstruction of indoor scenes using normal priors. In: Proceedings of the European Conference on Computer Vision. pp. 139-155. Springer (2022)", + "37. Wang, K., Shen, S.: Mvdepthnet: Real-time multiview depth estimation neural network. In: 2018 International conference on 3d vision. pp. 248-257. IEEE (2018)", + "38. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689 (2021)", + "39. Wang, Y., Li, Z., Jiang, Y., Zhou, K., Cao, T., Fu, Y., Xiao, C.: Neuralroom: Geometry-constrained neural implicit surfaces for indoor scene reconstruction. ACM Transactions on Graphics 41(6), 1-15 (2022)", + "40. Yao, Y., Luo, Z., Li, S., Fang, T., Quan, L.: Mvsnet: Depth inference for unstructured multi-view stereo. In: Proceedings of the European Conference on Computer Vision. pp. 767-783 (2018)", + "41. Yao, Y., Luo, Z., Li, S., Shen, T., Fang, T., Quan, L.: Recurrent mvsnet for high-resolution multi-view stereo depth inference. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5525-5534 (2019)", + "42. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34, 4805-4815 (2021)", + "43. Ye, B., Liu, S., Li, X., Yang, M.H.: Self-supervised super-plane for neural 3d reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21415-21424 (2023)", + "44. Yin, R., Karaoglu, S., Gevers, T.: Geometry-guided feature learning and fusion for indoor scene reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3652-3661 (2023)", + "45. Yu, Z., Gao, S.: Fast-mvsnet: Sparse-to-dense multi-view stereo with learned propagation and gauss-newton refinement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1949-1958 (2020)", + "46. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in Neural Information Processing Systems 35, 25018-25032 (2022)", + "47. Zhu, Z., Peng, S., Larsson, V., Cui, Z., Oswald, M.R., Geiger, A., Pollefeys, M.: Nicer-slam: Neural implicit scene encoding for rgb slam. arXiv preprint arXiv:2302.03594 (2023)", + "48. Zins, P., Xu, Y., Boyer, E., Wuhrer, S., Tung, T.: Multi-view reconstruction using signed ray distance functions (srdf). In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16696-16706 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 830 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Ray-Distance Volume Rendering for Neural Scene Reconstruction", + "bbox": [ + 294, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Ray-Distance Volume Rendering for Neural Scene Reconstruction/e689642e-0cb9-456e-9b14-7ba1bd1312a3_model.json b/2024/Ray-Distance Volume Rendering for Neural Scene Reconstruction/e689642e-0cb9-456e-9b14-7ba1bd1312a3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1ceeb59db387cda987e35294088d0bb863e2fbcd --- /dev/null +++ b/2024/Ray-Distance Volume Rendering for Neural Scene Reconstruction/e689642e-0cb9-456e-9b14-7ba1bd1312a3_model.json @@ -0,0 +1,2456 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.295, + 0.141, + 0.709, + 0.185 + ], + "angle": 0, + "content": "Ray-Distance Volume Rendering for Neural Scene Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.212, + 0.77, + 0.228 + ], + "angle": 0, + "content": "Ruihong Yin\\(^{1}\\), Yunlu Chen\\(^{2}\\), Sezer Karaoglu\\(^{1,3}\\), and Theo Gevers\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.239, + 0.709, + 0.253 + ], + "angle": 0, + "content": "1 University of Amsterdam, Amsterdam, The Netherlands" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.253, + 0.705, + 0.268 + ], + "angle": 0, + "content": "2 Carnegie Mellon University, Pittsburgh, The United States" + }, + { + "type": "text", + "bbox": [ + 0.344, + 0.268, + 0.658, + 0.281 + ], + "angle": 0, + "content": "3 3DUniversum, Amsterdam, The Netherlands" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.282, + 0.709, + 0.296 + ], + "angle": 0, + "content": "{r.yin, Th.Gevers}@uva.nl, s.karaoglu@3duniversum.com," + }, + { + "type": "text", + "bbox": [ + 0.411, + 0.297, + 0.591, + 0.309 + ], + "angle": 0, + "content": "yunluche@andrew.cmu.edu" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.346, + 0.74, + 0.651 + ], + "angle": 0, + "content": "Abstract. Existing methods in neural scene reconstruction utilize the Signed Distance Function (SDF) to model the density function. However, in indoor scenes, the density computed from the SDF for a sampled point may not consistently reflect its real importance in volume rendering, often due to the influence of neighboring objects. To tackle this issue, our work proposes a novel approach for indoor scene reconstruction, which instead parameterizes the density function with the Signed Ray Distance Function (SRDF). Firstly, the SRDF is predicted by the network and transformed to a ray-conditioned density function for volume rendering. We argue that the ray-specific SRDF only considers the surface along the camera ray, from which the derived density function is more consistent to the real occupancy than that from the SDF. Secondly, although SRDF and SDF represent different aspects of scene geometries, their values should share the same sign indicating the underlying spatial occupancy. Therefore, this work introduces a SRDF-SDF consistency loss to constrain the signs of the SRDF and SDF outputs. Thirdly, this work proposes a self-supervised visibility task, introducing the physical visibility geometry to the reconstruction task. The visibility task combines prior from predicted SRDF and SDF as pseudo labels, and contributes to generating more accurate 3D geometry. Our method implemented with different representations has been validated on indoor datasets, achieving improved performance in both reconstruction and view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.665, + 0.74, + 0.692 + ], + "angle": 0, + "content": "Keywords: Indoor scene reconstruction \\(\\cdot\\) neural radiance fields \\(\\cdot\\) signed ray distance function" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.719, + 0.375, + 0.735 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Indoor 3D scene reconstruction involves using multi-view RGB images as input to generate detailed 3D geometry as output. This task is crucial and has widespread applications in virtual/mixed augmentation, robotics, navigation, and so on. Contrary to a single object, indoor environments contain various elements, such as chairs, walls, tables, and cups, which present a challenge for the task of reconstruction. Traditional reconstruction approaches [1,21,27,33,40,44]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "R. Yin et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.329 + ], + "angle": 0, + "content": "use 2D and 3D convolutional neural networks (CNNs) [34, 35] to extract features. They can be classified as depth-based methods [6, 27, 37, 40] and volumetric methods [1, 21, 31, 33, 44]. Depth-based techniques estimate depth for every viewpoint, encountering difficulties like scale uncertainties and inconsistencies in depth across different views. The surfaces created through these techniques can be coarse and imprecise. Unlike depth-based approaches, volumetric techniques utilize 3D CNNs to directly generate 3D geometry, leading to smoother surfaces. Nevertheless, these methods use voxels for scene representation, causing GPU memory usage and computational expenses to grow cubically as voxel resolution increases. Such a limitation restricts the use of volumetric approaches for high-quality scene reconstruction. Moreover, volumetric techniques require 3D annotations for network training, significantly increasing the cost of annotations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.332, + 0.788, + 0.468 + ], + "angle": 0, + "content": "Implicit representations [18,24,29] encode 3D models as functions of coordinates. Utilizing implicit representations and volume rendering, neural radiance fields (NeRF) [19] achieves remarkable performance in novel view synthesis, even without 3D supervision. Some methods [38,42] exploit the capabilities of NeRF in reconstruction tasks. These methods typically parameterize the volume density as a learnable transformation of the Signed Distance Function (SDF), leading to impressive reconstruction performance. Compared to classic volumetric methods, neural implicit scene representation is more memory-efficient, marking a new trend in reconstruction tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.472, + 0.788, + 0.701 + ], + "angle": 0, + "content": "In multi-object indoor scenes, the SDF from sampled points considers surface points across the entire scene. However, for neural implicit scene reconstruction, along a camera ray, from the camera center to the intersected surface point, the SDF may exhibit fluctuations, leading to several local minimum values. Fig. 1 provides a toy example to illustrate this. In Fig. 1c, volume density yielded from the fluctuating SDF displays several local maximum values, contributing to high weights in Fig. 1d. However, the camera ray \\(\\overrightarrow{OP}\\) targets the green rectangle, while the SDF of the near-surface point \\(\\mathbf{Q}\\) refers to the surface of the blue cylinder. Besides, \\(\\mathbf{Q}\\) is far from the intersected point \\(\\mathbf{P}\\), but still corresponds to high density and high weight. Nevertheless, during volume rendering that maps the 3D output to 2D space, it is evident that along a ray, points closer to the intersected surface point may carry a higher weight as they are more relevant to actual observations. Consequently, in the multi-object scenes, using the SDF to model the density function along the camera ray may disrupt this relationship, introduce noise, and lead to inaccuracies in 2D results or 3D geometries." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.84 + ], + "angle": 0, + "content": "This paper proposes RS-Recon, a novel indoor scene reconstruction method with the ray-specific density function to address the aforementioned issue. Firstly, our method models the density function as a function of the Signed Ray Distance Function (SRDF). Unlike SDF, SRDF measures the shortest distance to the surface along a camera ray, eliminating the influence of surrounding surfaces that are not located on the given ray. Moreover, the density distribution generated from SRDF attains the local maximum only around the intersected surface, aligning more closely to the concept that points nearer to the surface have more importance. To achieve this, our network predicts both SRDF and SDF, where" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.295, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Ray-Distance Volume Rendering for Neural Scene Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "SRDF encompasses the density function, while SDF mainly focuses on describing the 3D surfaces. Secondly, despite SRDF and SDF being defined differently, both of them produce positive values outside the object and negative values within. Therefore, in this paper, a SRDF-SDF consistency loss is devised to ensure they share the same sign. Thirdly, this work introduces a self-supervised visibility task to enhance the prediction of 3D geometry. The visibility task predicts the probability of whether the sampled points are physically visible along the ray in 3D space. The pseudo visibility ground truth is formed by combining the prior of SRDF and SDF from the network itself, without the reliance on multi-view geometry or any additional annotations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.79, + 0.465 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows, (1) This work proposes a novel neural scene reconstruction method that leverages the ray-specific SRDF to model the volume density. Compared to SDF, SRDF better reflects the significance of the real observations. (2) A SRDF-SDF consistency loss is proposed to constrain the alignment between SRDF and SDF, ensuring the same sign inside and outside the object. (3) To enhance the accuracy of 3D geometry, an additional self-supervised visibility task is introduced to predict the visibility probability of sampled points. This task integrates prior information from both the SDF and SRDF predictions as labels. (4) The experiments conducted on both synthetic and real-world indoor datasets illustrate that our method enhances performance in both reconstruction and view synthesis." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.489, + 0.398, + 0.506 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.789, + 0.809 + ], + "angle": 0, + "content": "Classic Indoor Scene Reconstruction. Traditional methods for indoor scene reconstruction fall into two classes, i.e. depth-based methods [6, 16, 27, 37, 40, 41, 45] and volumetric methods [1, 7, 12, 21, 31, 33, 44]. Depth-based methods predict pixel-level depth for individual frames, and subsequently depth fusion [4] is adopted to generate the 3D scene. To achieve favorable outcomes, these methods often construct a plane sweep cost volume [2, 9] at the feature or image level, leveraging information from multiple views to complement the current frame, e.g. MVSNet [40], DeepVideoMVS [6], and SimpleRecon [27]. Using the depth map as the intermediate representation, these methods encounter challenges such as depth inconsistency and scale ambiguities. In contrast, volumetric methods utilize 3D CNN to directly regress the 3D geometry, which can generate smooth surfaces and reconstruct unobserved regions. For instance, NeuralRecon [33] designs a learnable TSDF fusion module to integrate features from previous frames and predicts the TSDF for sparse volumes. TransformerFusion [1], based on volumetric representation, adopts Transformer to fuse multi-view features and select frames with higher attention. Yin et al. [44] propose to incorporate geometry priors at different levels for the volumetric methods. However, due to the use of 3D CNN in volumetric methods, computational costs escalate significantly with higher resolutions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Neural Implicit Scene Representation. With the success of implicit representation and NeRF [19], neural implicit scene representation has emerged as a" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "R. Yin et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.339, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.21, + 0.345, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.284, + 0.324, + 0.294 + ], + "angle": 0, + "content": "(a) Multi-object scene" + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.145, + 0.493, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.215, + 0.493, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.405, + 0.284, + 0.458, + 0.293 + ], + "angle": 0, + "content": "(b) SDF/SRDF" + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.145, + 0.639, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.215, + 0.639, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.555, + 0.284, + 0.6, + 0.294 + ], + "angle": 0, + "content": "(c) Density" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.145, + 0.785, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.215, + 0.785, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.284, + 0.745, + 0.294 + ], + "angle": 0, + "content": "(d) Weight" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.309, + 0.789, + 0.438 + ], + "angle": 0, + "content": "Fig. 1: Analysis of the density function and weight distribution for a toy scene. (a) Frontal and overhead perspectives of a scene with multiple objects, where a ray originates from the camera center \\(\\mathbf{O}\\), intersecting with the green rectangle at the point \\(\\mathbf{P}\\). \\(\\mathbf{Q}\\) is the closest point to the blue cylinder along the ray \\(\\overrightarrow{OP}\\). (b) The SDF \\(d_{\\Omega}\\) /SRDF \\(\\tilde{d}_{\\Omega}\\) in Eqs. (2) and (4) along the ray \\(\\overrightarrow{OP}\\). (c) The volume density \\(\\sigma\\) generated from the SDF/SRDF in (b) using Eqs. (3) and (5). (d) The weight distribution \\(T_{i}\\alpha_{i}\\) generated from the density in (c) via Eq. (1). Although \\(\\mathbf{Q}\\) is distant from the surface boundary intersecting with the ray, the density function from the SDF in \\(\\mathbf{Q}\\) generates a high weight in volume rendering, resulting in noisy rendering and reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.789, + 0.763 + ], + "angle": 0, + "content": "hot spot in the past few years. Generally, reconstruction approaches based on neural implicit scene representation first map point positions to a continuous representation of SDF or occupancy by multi-layer perceptrons (MLPs), then transfer SDF or occupancy to volume density, and finally render the 2D results from the 3D outputs. Despite requiring only 2D supervision during training, these methods can reconstruct high-quality surfaces and are more computationally efficient. For example, VolSDF [42] parameterizes the volume density as a Laplace distribution of SDF, while NeuS [38] models it as the logistic distribution. Some extensions [10,36,39,46,47] introduce extra priors to achieve superior outcomes. For instance, Manhattan-SDF [10] employs planar constraints based on Manhattan-world assumption [3]. NeuRIS [36] proposes to combine normal prior for the reconstruction of texture-less regions, while NeuralRoom [39] implements perturbation-residual restriction to reconstruct the flat region better. MonoSDF [46] exploits both depth and normal cues to improve the reconstruction quality. Although additional priors contribute to the accuracy and completeness of reconstructed surfaces, they employ the SDF to model the volume density, potentially resulting in noisy 2D results in the multi-object scenes. Contrarily, our method advocates for parameterizing the density function with the ray-specific SRDF, resulting in enhanced reconstruction performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Some NeRF-based methods also utilize multi-view geometry to introduce additional regularization. For instance, RegNeRF [22] regularizes patched-based depth maps from unobserved viewpoints. Geo-NeuS [8] and ConsistentNeRF [11] use multi-view photometric consistency. SRDF and observation consistencies are used by [48]. GenS [25] introduces multi-view and multi-scale feature consistent" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.295, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Ray-Distance Volume Rendering for Neural Scene Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.144, + 0.788, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.308, + 0.789, + 0.448 + ], + "angle": 0, + "content": "Fig. 2: Our framework. A geometry MLP \\( f_{g} \\) is adopted to generate the SDF and geometry features \\( \\mathbf{F}_{g} \\), utilizing the encoded position (and optionally grid features) as input. Then, \\( \\mathbf{F}_{g} \\), along with viewing direction, passes through a color MLP \\( f_{c} \\) to predict the color for each point. Notably, (1) our approach advocates for modeling the density function with the ray-specific SRDF in addition to the SDF. For this purpose, a SRDF MLP \\( f_{s} \\) is introduced to generate the SRDF. (2) A SRDF-SDF consistency loss \\( \\mathcal{L}_{con} \\) is devised to align the signs between the generated SRDF and SDF. (3) To enhance the geometry prediction, a self-supervised visibility task is proposed to integrate geometry priors in both SRDF and SDF predicted by the network and generates the pseudo visibility ground truth. The visibility probability is predicted by the SRDF MLP." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.789, + 0.584 + ], + "angle": 0, + "content": "cies for neural surface reconstruction. VIP-NeRF [30] adopts a visibility prior, using the plane sweep volume among multiple views to generate visibility labels. However, they depend on the computation of multi-view geometry, some of which may involve processing multiple frames yielding increased computational demands. In contrast, this paper introduces a self-supervised visibility task that requires minimal computational increases and does not depend on multi-view geometry for ground truth generation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.607, + 0.331, + 0.623 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.788, + 0.732 + ], + "angle": 0, + "content": "In this section, we provide the background on NeRF-based reconstruction methods, which typically model the volume density as a transformation of SDF (refer to Sec. 3.1). To overcome the issue of false local maxima in the density from SDF, our approach covered in Sec. 3.2 - Sec. 3.5 proposes to model the density function with the ray-specific SRDF in neural scene reconstruction, which achieves more accurate reconstruction. Fig. 2 presents the framework of our approach." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.754, + 0.77, + 0.77 + ], + "angle": 0, + "content": "3.1 Background: Volume Rendering Density as Transformed SDF" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Our goal is to recover the scene geometry given a set of projected 2D views of a 3D scene. To achieve this, NeRF [19] optimizes a continuous 5D function to represent the scene, which predicts the volume density \\(\\sigma(\\mathbf{p})\\) and view-dependent color \\(\\mathbf{c}(\\mathbf{p},\\mathbf{r})\\) for each sampled point \\(\\mathbf{p} \\in \\mathbb{R}\\) and ray direction \\(\\mathbf{r} \\in S^2\\). With \\(\\Omega \\subset \\mathbb{R}^3\\)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "R. Yin et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "denoting the scene geometry filled with physical presence, the predicted density \\(\\sigma (\\mathbf{p})\\) approximates the ground truth scene density \\(\\mathbf{1}_{\\varOmega}(\\mathbf{p})\\), such that \\(\\mathbf{1}_{\\varOmega}(\\mathbf{p}) = 1\\) if \\(\\mathbf{p}\\in \\varOmega\\) and 0 otherwise. After that, the classic volume rendering is employed to render the color \\(\\mathbf{C}(\\mathbf{r})\\), and \\(N\\) points \\(\\{\\mathbf{p}_i = \\mathbf{o} + z_i\\mathbf{r}, i\\in [1,N]\\}\\) are sampled along the camera ray \\(\\mathbf{r}\\), with \\(z_{i}\\) indicating the depth from the camera center \\(\\mathbf{o}\\) to the sampled point. The color is accumulated along the ray:" + }, + { + "type": "equation", + "bbox": [ + 0.267, + 0.248, + 0.786, + 0.289 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} (\\mathbf {r}) = \\sum_ {i = 1} ^ {N} T _ {i} \\alpha_ {i} \\mathbf {c} \\left(\\mathbf {p} _ {i}, \\mathbf {r}\\right), T _ {i} = \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {i}\\right), \\alpha_ {i} = 1 - \\exp \\left(- \\sigma \\left(\\mathbf {p} _ {i}\\right) \\delta_ {i}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.297, + 0.545, + 0.312 + ], + "angle": 0, + "content": "where \\(\\delta_{i}\\) is the interval of neighboring points." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.784, + 0.403 + ], + "angle": 0, + "content": "Despite the great success in novel view synthesis, NeRF has difficulty reconstructing satisfactory actual surfaces from the generated volume density. To address this challenge, recent methods [42,46] propose to predict the SDF and then transform it to volume density, which aligns with the geometry bias that the surface points have a higher density than other non-surface points and corresponds to a higher weight during volume rendering." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.786, + 0.462 + ], + "angle": 0, + "content": "For any sampled point \\(\\mathbf{p} \\in \\mathbb{R}^3\\), the SDF \\(d_{\\Omega}(\\mathbf{p})\\) is with an absolute value representing the shortest distance from \\(\\mathbf{p}\\) to the surface \\(\\mathcal{M} = \\partial \\Omega\\), while the sign indicates whether the point is outside (positive) or inside (negative) the surface. Formally," + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.463, + 0.785, + 0.487 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S D F}: \\quad d _ {\\Omega} (\\mathbf {p}) = (- 1) ^ {\\mathbf {1} _ {\\Omega} (\\mathbf {p})} \\min _ {\\mathbf {p} ^ {*} \\in \\mathcal {M}} \\| \\mathbf {p} - \\mathbf {p} ^ {*} \\| _ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.788, + 0.552 + ], + "angle": 0, + "content": "where the indicator function \\(\\mathbf{1}_{\\varOmega}(\\mathbf{p}) = 1\\) if \\(\\mathbf{p}\\in \\varOmega\\) and 0 otherwise. This indicates that the SDF considers every surface around the point. To apply SDF to volume rendering, Yariv et al. [42] suggest to derive the volume density \\(\\sigma^{\\mathrm{SDF}}\\) from the signed distance as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.272, + 0.56, + 0.786, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\sigma^ {\\mathrm {S D F}} (\\mathbf {p}) = \\alpha \\Psi_ {\\beta} (- d _ {\\Omega} (\\mathbf {p})) \\text {,} \\Psi_ {\\beta} (s) = \\left\\{ \\begin{array}{l l} \\frac {1}{2} \\exp \\left(\\frac {s}{\\beta}\\right) & \\text {i f} s \\leq 0 \\\\ 1 - \\frac {1}{2} \\exp \\left(- \\frac {s}{\\beta}\\right) & \\text {i f} s > 0 \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.788, + 0.663 + ], + "angle": 0, + "content": "in which \\(\\alpha, \\beta\\) are learnable parameters. \\(\\Psi_{\\beta}\\) is the Cumulative Distribution Function (CDF) of the Laplace distribution with zero mean and \\(\\beta\\) scale. Then, \\(\\sigma^{\\mathrm{SDF}}(\\mathbf{p})\\) is applied to Eq. (1) and as the density function \\(\\sigma\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.681, + 0.736, + 0.697 + ], + "angle": 0, + "content": "3.2 Accurate Scene Volume Density from Ray-Specific SRDF" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.81 + ], + "angle": 0, + "content": "We illustrate the issues with volume rendering SDF by a toy example in Fig. 1. Indoor scenes typically include multiple objects. Along the camera ray \\(\\mathbf{r}\\), the SDF may be influenced by surfaces not intersecting the ray, resulting in several ambiguous local maxima of false high volume density and weight. This implies that the 2D observations may have a strong correlation to these distant points, contradicting the fact that the points closer to the surface along the camera ray may carry more significance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "To this end, we propose volume rendering scene density function from SRDF, the distance field local to the ray direction. In contrast to the SDF in Eq. (2)," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.295, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Ray-Distance Volume Rendering for Neural Scene Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.196 + ], + "angle": 0, + "content": "the SRDF \\(\\tilde{d_{\\Omega}}\\) (\\(\\mathbf{p}, \\mathbf{r}\\)) [26,48] computes the shortest distance from the point to the surface along the sampled ray \\(\\mathbf{r}\\), i.e. the SRDF is ray-dependent. Formally, the SRDF \\(\\tilde{d_{\\Omega}}\\) (\\(\\mathbf{p}, \\mathbf{r}\\)) can be represented by" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.207, + 0.786, + 0.233 + ], + "angle": 0, + "content": "\\[\n\\text {S R D F :} \\quad \\tilde {d _ {\\Omega}} (\\mathbf {p}, \\mathbf {r}) = (- 1) ^ {\\mathbf {1} _ {\\Omega} (\\mathbf {p})} \\min _ {(\\mathbf {p} + \\rho \\mathbf {r}) \\in \\mathcal {M}; \\rho \\in \\mathbb {R}} | \\rho |, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.244, + 0.786, + 0.276 + ], + "angle": 0, + "content": "where \\(\\mathbf{1}_{\\Omega}(\\mathbf{p}) = 1\\) if \\(\\mathbf{p} \\in \\Omega\\) and \\(0\\) otherwise. In the following sections, \\(d_{\\Omega}\\) from Eq. (2) and \\(\\tilde{d_{\\Omega}}\\) are used to denote SDF and SRDF for brevity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.277, + 0.789, + 0.458 + ], + "angle": 0, + "content": "In Fig. 1, it can be seen that the weight computed from SRDF in the volume rendering only reaches a local peak around the surface point, aligning more consistently with the nature of the 2D observations. Therefore, this paper proposes to employ the ray-specific SRDF to yield the volume density instead of the SDF. This work predicts both SDF and SRDF, where SDF is mainly used to model the surface. As depicted in Fig. 2, for a sampled point \\(\\mathbf{p}\\), a geometry MLP \\(f_{g}\\) is applied on the input (including the encoded position or features) to predict SDF \\(d_{\\Omega}\\) and geometry feature \\(\\mathbf{F}_{g}\\). By definition, SRDF is view-dependent. Hence, the viewing direction \\(\\mathbf{r}\\), geometry feature \\(\\mathbf{F}_{g}\\), and point position \\(\\mathbf{p}\\) are concatenated and then passed through a SRDF MLP \\(f_{s}\\) to yield the SRDF. On the other hand, the viewing direction \\(\\mathbf{r}\\), geometry feature \\(\\mathbf{F}_{g}\\), point position \\(\\mathbf{p}\\), and 3D unit normal \\(\\mathbf{n} = \\nabla d_{\\Omega}\\) are used as input to a color MLP \\(f_{c}\\) to output color \\(\\mathbf{c}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.49 + ], + "angle": 0, + "content": "For volume rendering, our method adopts the predicted SRDF \\(\\tilde{d_{\\Omega}}\\) to derive the volume density \\(\\sigma\\). This is defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.501, + 0.786, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\sigma^ {\\mathrm {S R D F}} (\\mathbf {p}, \\mathbf {r}) = \\alpha \\Psi_ {\\beta} \\left(- \\tilde {d _ {\\Omega}} (\\mathbf {p}, \\mathbf {r})\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.612 + ], + "angle": 0, + "content": "Then, \\(\\sigma^{\\mathrm{SRDF}}(\\mathbf{p},\\mathbf{r})\\) is applied to the volume rendering in Eq. (1) as a ray-conditioned density function to obtain the color value \\(\\mathbf{C}^{\\mathrm{SRDF}}(\\mathbf{r})\\). Besides, during training, we also render color \\(\\mathbf{C}^{\\mathrm{SDF}}(\\mathbf{r})\\) with density functions from SDF to obtain gradient signals to optimize the prediction of SDF, in which the volume rendering processes use the same rays and points as those for \\(\\mathbf{C}^{\\mathrm{SRDF}}(\\mathbf{r})\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.675 + ], + "angle": 0, + "content": "For the scene geometry, after the network is trained with NeRF optimization targets, the geometric surface \\(\\mathcal{M} = \\partial \\Omega\\) is extracted as watertight meshes by the marching cubes algorithm [15,24] from the SDF results \\(d_{\\Omega}\\), which only depends on the query location \\(\\mathbf{p}\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.695, + 0.504, + 0.71 + ], + "angle": 0, + "content": "3.3 SRDF-SDF Consistency Loss" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Following the definitions in Eq. (4), despite SRDF and SDF being defined differently, they share the same sign, which indicates the spatial occupancy — positive outside the surface and negative inside the surface. However, in this paper, where SDF and SRDF are predicted from separate branches, sign consistency is not guaranteed. To deal with this issue, a SRDF-SDF consistency loss is proposed to enforce this constraint." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Although the sign function effectively computes the output's sign, it cannot be differentiated during back-propagation. To a certain degree, the values derived" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "R. Yin et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.209 + ], + "angle": 0, + "content": "from the sigmoid function approximate those calculated by the sign function. Thus, in our SRDF-SDF consistency loss, the sigmoid function is first employed to generate the sign of SRDF and SDF as the approximated sign function \\(\\varsigma (\\cdot)\\), after which a \\(\\ell_2\\) loss is adopted to quantify the sign difference, as defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.22, + 0.786, + 0.257 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c o n} = \\frac {1}{N _ {r}} \\sum_ {\\mathbf {p}, \\mathbf {r}} M _ {c o n} | | \\varsigma (\\tilde {d} _ {\\Omega}) - \\varsigma (d _ {\\Omega}) | | _ {2} \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.35, + 0.259, + 0.66, + 0.279 + ], + "angle": 0, + "content": "\\[\n\\varsigma (d) = \\mathsf {S i g m o i d} (k \\cdot d), M _ {c o n} = [ \\widetilde {d} _ {\\Omega} \\cdot d _ {\\Omega} < 0 ]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.289, + 0.789, + 0.368 + ], + "angle": 0, + "content": "in which \\( k \\) is a hyperparameter, controlling the slope of the sigmoid function. The indicator function \\( [\\tilde{d}_{\\Omega} \\cdot d_{\\Omega} < 0] \\) is leveraged to identify whether the generated SDF and SRDF have opposite signs. Therefore, the consistency loss imposes penalties when there is a difference between the signs of SRDF and SDF. \\( N_r \\) represents the number of sampled points in a minibatch that satisfy \\( [\\tilde{d}_{\\Omega} \\cdot d_{\\Omega} < 0] \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.368, + 0.788, + 0.398 + ], + "angle": 0, + "content": "Taking the derivative of \\(\\mathcal{L}_{con}\\) with respect to SDF \\(d_{\\Omega}\\) (as an example), is computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.337, + 0.41, + 0.786, + 0.442 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\mathcal {L} _ {c o n}}{\\partial d _ {\\Omega}} = - \\frac {2}{N _ {r}} M _ {c o n} | | \\varsigma (\\tilde {d} _ {\\Omega}) - \\varsigma (d _ {\\Omega}) | | _ {1} \\cdot \\frac {\\partial \\varsigma (d _ {\\Omega})}{\\partial d _ {\\Omega}} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.452, + 0.788, + 0.577 + ], + "angle": 0, + "content": "The derivative in Eq. (7) shows that our consistency loss has two advantages: (1) Unlike the sign function, which assigns the same penalty for outputs with different signs of SRDF and SDF, our consistency loss can adjust the penalty according to the extent of the sign discrepancy between SRDF and SDF, referring to \\(||\\varsigma (\\tilde{d}_{\\varOmega}) - \\varsigma (d_{\\varOmega})||_1\\). (2) The derivative of the sigmoid function \\(\\frac{\\partial\\varsigma(d_{\\varOmega})}{\\partial d_{\\varOmega}}\\) peaks at zero and decreases from 0 to \\(+\\infty / - \\infty\\). Consequently, predictions in proximity to zero (indicating the surface) may exhibit higher absolute gradient values, providing effective supervision for points near the surface." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.598, + 0.513, + 0.615 + ], + "angle": 0, + "content": "3.4 Self-supervised Visibility Task" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Theoretically, with visibility attributes of 3D points, surface location becomes straightforward. Therefore, to enhance 3D geometry prediction, we propose a self-supervised visibility task in this paper. Along the trajectory of a camera ray, points sampled prior to reaching the first surface point are categorized as visible, whereas points sampled beyond the first surface point are classified as occluded. To find the first surface point along the ray, using SRDF as an example, the multiplication of SRDF between adjacent sampled points is computed. When \\(\\tilde{d}_{\\Omega}^{i}\\cdot \\tilde{d}_{\\Omega}^{i + 1}\\leq 0\\) and \\(\\tilde{d}_{\\Omega}^{m}\\cdot \\tilde{d}_{\\Omega}^{m + 1} > 0,m\\in [1,i - 1]\\), it indicates that the first surface is within the interval \\([\\mathbf{o} + z_i\\mathbf{r},\\mathbf{o} + z_{i + 1}\\mathbf{r}]\\). Consequently, the points between \\([\\mathbf{p}_1,\\mathbf{p}_i]\\) are considered physically visible while points between \\([\\mathbf{p}_{i + 1},\\mathbf{p}_N]\\) are regarded as occluded. However, the SRDF or SDF predicted by the network may be inaccurate at some points, introducing potential noise in the visibility task. To improve the accuracy of the visibility labeling, both SDF and SRDF are utilized in determining visibility. Points are labeled as visible or occluded only when there" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.295, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Ray-Distance Volume Rendering for Neural Scene Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 1: Evaluation of 3D reconstruction meshes on ScanNet. \\(\\dagger\\) indicates our reproduced results. The best results are bold. \\(\\downarrow\\) denotes that a lower value is preferable while \\(\\uparrow\\) means higher value corresponds to better performance." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.2, + 0.756, + 0.406 + ], + "angle": 0, + "content": "
MethodAcc ↓Comp ↓Prec ↑Recall ↑F-score ↑
COLMAP [28]0.0470.2350.7110.4410.537
UNISURF [23]0.5540.1640.2120.3620.267
NeuS [38]0.1790.2080.3130.2750.291
VolSDF [42]0.4140.1200.3210.3940.346
Manhattan-SDF [10]0.0720.0680.6210.5860.602
S3P [43]0.0550.0590.7090.6600.683
NeuRIS [36]0.0500.0490.7170.6690.692
MonoSDF_Grid [46]0.0720.0570.6600.6010.626
MonoSDF_MLP [46]0.0350.0480.7990.6810.733
HelixSurf [14]0.0380.0440.7860.7270.755
Occ_SDF_Hybrid† [17]0.0400.0410.7830.7480.765
Ours_Grid0.0740.0490.6700.7030.683
Ours_MLP0.0400.0400.8090.7790.794
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.788, + 0.493 + ], + "angle": 0, + "content": "is agreement between SRDF and SDF on their classification. If the classifications from SRDF and SDF diverge, suggesting difficulties in accurately categorizing these points for the network, such points are omitted from the visibility task. The method for assigning visibility labels is defined as follows" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.503, + 0.786, + 0.543 + ], + "angle": 0, + "content": "\\[\nV _ {g t} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} (V ^ {\\mathrm {S R D F}} = 1) \\& (V ^ {\\mathrm {S D F}} = 1) \\\\ 0, & \\text {i f} (V ^ {\\mathrm {S R D F}} = 0) \\& (V ^ {\\mathrm {S D F}} = 0) \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.788, + 0.628 + ], + "angle": 0, + "content": "where \\( V^{\\mathrm{SRDF}} / V^{\\mathrm{SDF}} \\) are the visibility labels determined based on the prediction of SRDF/SDF. 1/0 means label for visible/occluded points. The formation of the visibility label can be inferred as self-supervised, thus eliminating the necessity for multi-view geometry and additional annotations. Additionally, it incorporates information from both SRDF and SDF." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.69 + ], + "angle": 0, + "content": "From the analysis above, similar to SRDF, the visibility task also relies on the view direction. Therefore, a visibility probability \\( V_{pred} \\) is predicted from the SRDF branch. It is a binary classification task, i.e. visible or occluded, so the binary cross-entropy loss is employed as visibility loss \\( \\mathcal{L}_{vis} \\) during optimization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.766 + ], + "angle": 0, + "content": "The visibility task can discern the point's visibility along the camera ray, thereby influencing the learning of SRDF. Moreover, the optimization process integrates knowledge of ray-related visibility across the entire space, further impacting the learning of SDF. Notably, unlike 2D supervision, the visibility task provides priors in 3D space, imparting a more direct influence on 3D geometry." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.786, + 0.373, + 0.801 + ], + "angle": 0, + "content": "3.5 Optimization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Loss Functions. During training, the overall loss consists of RGB loss \\(\\mathcal{L}_c\\), normal loss \\(\\mathcal{L}_n\\), depth loss \\(\\mathcal{L}_d\\), Eikonal loss \\(\\mathcal{L}_e\\), smooth loss \\(\\mathcal{L}_s\\), SRDF-SDF" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "R. Yin et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.544, + 0.163 + ], + "angle": 0, + "content": "consistency loss \\(\\mathcal{L}_{con}\\), and visibility loss \\(\\mathcal{L}_{vis}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.285, + 0.175, + 0.786, + 0.19 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {c} + \\lambda_ {n} \\mathcal {L} _ {n} + \\lambda_ {d} \\mathcal {L} _ {d} + \\lambda_ {e} \\mathcal {L} _ {e} + \\lambda_ {s} \\mathcal {L} _ {s} + \\lambda_ {c o n} \\mathcal {L} _ {c o n} + \\lambda_ {v i s} \\mathcal {L} _ {v i s} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.2, + 0.785, + 0.262 + ], + "angle": 0, + "content": "in which \\(\\lambda_{(\\cdot)}\\) are the loss weight. Following MonoSDF [46], the 2.5D depth and normal maps are volume rendered using the density function \\(\\sigma^{\\mathrm{SRDF}}\\) to exploit geometric cues and support training. More details on the computation of loss functions are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.262, + 0.787, + 0.385 + ], + "angle": 0, + "content": "Implementation Details. Theoretically, our method can be applied to replace SDF-based volume rendering in indoor scene reconstruction. To measure our method, VolSDF-based reconstruction method MonoSDF is considered as our baseline in the following experiments. Furthermore, in the supplementary material, we also verify our method on NeuS-based reconstruction method NeuRIS. We used different spatial encoding techniques as input, including hash features [20] and Fourier features [19], referred to as Grid and MLP respectively in the following experiments. More details are given in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.406, + 0.376, + 0.423 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.438, + 0.428, + 0.453 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.785, + 0.522 + ], + "angle": 0, + "content": "Datasets. Our proposed method is evaluated on three indoor datasets, comprising: (1) four scenes from the real-world dataset ScanNet [5], (2) eight scenes from the synthetic dataset Replica [32], and (3) four scenes from the advanced set of the real-world large-scale dataset Tanks and Temples [13]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.787, + 0.657 + ], + "angle": 0, + "content": "Metrics. For ScanNet, the assessment of the results involves five reconstruction metrics: accuracy (acc), completeness (comp), precision (prec), recall, and F-score with a threshold of 5cm. For Replica, in alignment with MonoSDF, metrics including normal consistency (normal c.), chamfer L1 distance (Chamfer-L1), and F-score with a threshold of 5cm are utilized. For Tanks and Temples, as evaluated by the official server, results are reported based on the F-score with a 1cm threshold. Additionally, this study assesses the peak signal-to-noise ratio (PSNR) of the rendered 2D RGBs to showcase the effectiveness of our approach. Detailed computations for all metrics are given in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.658, + 0.787, + 0.72 + ], + "angle": 0, + "content": "Baselines. Except for the baseline MonoSDF, our method is also compared to (1) the classic MVS method: COLMAP [28], (2) neural implicit reconstruction methods: UNISURF [23], NeuS [38], VolSDF [42], Manhattan-SDF [10], NeuRIS [36], S3P [43], HelixSurf [14], and Occ_SDF_Hybrid [17]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.74, + 0.419, + 0.755 + ], + "angle": 0, + "content": "4.2 Evaluation Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Reconstruction on ScanNet [5]. Tab. 1 compares our method against other state-of-the-art (SOTA) methods on ScanNet. When contrasted with the baseline MonoSDF, our method consistently shows better performance across all metrics, regardless of whether utilizing MLP or grid representation. For example, employing grid representation, our method exhibits a \\(5.7\\%\\) higher F-score than" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.295, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Ray-Distance Volume Rendering for Neural Scene Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.144, + 0.773, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.32, + 0.785, + 0.349 + ], + "angle": 0, + "content": "Fig. 3: Qualitative comparisons on ScanNet. It can be seen that our method can reconstruct more surfaces, especially in thin regions." + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.361, + 0.768, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.585, + 0.788, + 0.616 + ], + "angle": 0, + "content": "Fig.4: Qualitative comparisons on Replica. Compared to MonoSDF and Occ_SDF_Hybrid, our method can generate more accurate surfaces." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.72 + ], + "angle": 0, + "content": "MonoSDF. Using the MLP representation, our method outperforms MonoSDF by \\(1.0\\%\\) in recall, \\(9.8\\%\\) in precision, and \\(6.1\\%\\) in F-score. In addition, our method with the MLP representation surpasses other SOTA methods on almost all metrics, explaining the effectiveness of our method on reconstruction. Notably, our method achieves a \\(2.9\\%\\) higher F-score compared to Occ_SDF_Hybrid." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Reconstruction on Replica [32]. The results on the synthetic dataset Replica are shown in Tab. 2. Using the grid representation, our method reports a notable improvement of \\(9.91\\%\\) in chamfer distance and \\(2.88\\%\\) in F-score. Similar increases are also illustrated in our method with the MLP representation across all metrics. Our approach also achieves on-par results with Occ_SDF_Hybrid. Reconstruction on Tanks and Temples [13]. Tab. 3 shows quantitative outcomes on the challenging large-scale dataset Tanks and Temples. The F-score for each scene - Auditorium, Ballroom, Courtroom, and Museum - along with" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "R. Yin et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.295, + 0.145, + 0.707, + 0.16 + ], + "angle": 0, + "content": "Table 2: Evaluation of 3D reconstruction meshes on Replica." + }, + { + "type": "table", + "bbox": [ + 0.246, + 0.171, + 0.758, + 0.268 + ], + "angle": 0, + "content": "
MethodNormal C. ↑Chamfer-L1 ↓F-score ↑
GridMonoSDF [46]90.933.2385.91
Ours91.662.9188.79
MLPMonoSDF [46]92.112.9486.18
Occ_SDF_Hybrid†[17]93.342.5891.54
Ours93.492.6091.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.28, + 0.787, + 0.31 + ], + "angle": 0, + "content": "Table 3: Evaluation of 3D reconstruction meshes on the advanced set of Tanks and Temples. The F-score is reported in the table." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.321, + 0.785, + 0.432 + ], + "angle": 0, + "content": "
MethodAuditoriumBallroomCourtroomMuseummean
GridMonoSDF [46]3.173.7013.755.686.58
Occ_SDF_Hybrid†[17]3.763.5814.047.277.16
Ours4.844.3114.187.607.73
MLPMonoSDF [46]3.092.4710.005.105.17
Occ_SDF_Hybrid†[17]3.213.3410.294.665.38
Ours4.244.6910.795.916.41
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.46, + 0.785, + 0.52 + ], + "angle": 0, + "content": "their mean, is reported. Our method, with both the MLP and grid representations, outperforms Occ_SDF_Hybrid and MonoSDF. In particular, our F-score is \\(1.15\\%\\) higher (grid representation) and \\(1.24\\%\\) higher (MLP representation) than MonoSDF across all scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.785, + 0.657 + ], + "angle": 0, + "content": "Visualization of Reconstruction Results. Fig. 3 and Fig. 4 report qualitative results on ScanNet and Replica respectively. In contrast to other methods, our approach can produce superior meshes and recall more regions, particularly for thin regions/small objects, such as chair legs and table legs. Although our method relies on the SRDF for improved density in volume rendering while the surface geometry is extracted by the SDF, both the proposed SRDF-SDF consistency loss and self-supervised visibility task guide the optimization of the SDF to be more consistent with the SRDF, resulting in more accurate surface boundary. Additional visualization is provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.787, + 0.689 + ], + "angle": 0, + "content": "View Synthesis. Tab. 4 also provides a comparison of view synthesis performance on training views. Notably, 2D views are rendered with density derived" + }, + { + "type": "table_caption", + "bbox": [ + 0.253, + 0.716, + 0.748, + 0.731 + ], + "angle": 0, + "content": "Table 4: Evaluation of the rendering quality. PSNR is given in this table." + }, + { + "type": "table", + "bbox": [ + 0.231, + 0.742, + 0.774, + 0.839 + ], + "angle": 0, + "content": "
MethodScanNetReplicaTanks and Temples
GridMonoSDF [46]28.9640.2528.45
Ours30.0942.0429.52
MLPMonoSDF [46]26.4034.4524.13
Occ_SDF_Hybrid†[17]26.9835.5024.72
Ours27.7736.0625.47
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.295, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Ray-Distance Volume Rendering for Neural Scene Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.324, + 0.145, + 0.677, + 0.159 + ], + "angle": 0, + "content": "Table 5: Ablation study of our method on ScanNet." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.172, + 0.777, + 0.282 + ], + "angle": 0, + "content": "
MethodAcc ↓Comp ↓Prec ↑Recall ↑F-score ↑
aBaseline0.0350.0480.7990.6810.733
bRS. Den.0.0400.0440.7720.7200.745
cRS. Den. + Con. L.0.0390.0410.7940.7600.776
dRS. Den. + Con. L. + Vis.(SDF)0.0390.0400.8090.7720.789
eRS. Den. + Con. L. + Vis.(SRDF)0.0420.0400.8040.7740.788
fRS. Den. + Con.L. + Vis.0.0400.0400.8090.7790.794
" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.296, + 0.353, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.389, + 0.326, + 0.396 + ], + "angle": 0, + "content": "(a) Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.298, + 0.472, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.37, + 0.389, + 0.463, + 0.396 + ], + "angle": 0, + "content": "(b) RGB from MonoSDF" + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.298, + 0.596, + 0.305 + ], + "angle": 0, + "content": "For the red sampled point" + }, + { + "type": "image", + "bbox": [ + 0.473, + 0.305, + 0.619, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.298, + 0.758, + 0.305 + ], + "angle": 0, + "content": "For the yellow sampled point" + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.305, + 0.767, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.389, + 0.7, + 0.396 + ], + "angle": 0, + "content": "(c) Weight from MonoSDF" + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.399, + 0.473, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.386, + 0.493, + 0.458, + 0.5 + ], + "angle": 0, + "content": "(d) RGB from ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.403, + 0.595, + 0.411 + ], + "angle": 0, + "content": "For the red sampled point" + }, + { + "type": "image", + "bbox": [ + 0.473, + 0.411, + 0.619, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.493, + 0.68, + 0.501 + ], + "angle": 0, + "content": "(e) Weight from ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.403, + 0.758, + 0.411 + ], + "angle": 0, + "content": "For the yellow sampled point" + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.411, + 0.767, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.517, + 0.788, + 0.546 + ], + "angle": 0, + "content": "Fig.5: Comparison of the yielded weight and image. The weight and image generated by our approach align more closely with the actual observations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.576, + 0.788, + 0.637 + ], + "angle": 0, + "content": "from SRDF. On the three datasets, our method outperforms the baseline and Occ_SDF_Hybrid in rendering training views. Overall, our method not only enhances surface reconstruction but also generates more accurate views. The visualization can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.662, + 0.39, + 0.677 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Tab. 5 presents an ablation study to evaluate the effectiveness of our design on ScanNet with MLP representation. The ablation experiments consist of: (a) Baseline: MonoSDF. (b) RS. Den.: this structure models the volume density using the ray-specific SRDF, without the SRDF-SDF consistency loss and self-supervised visibility task. (c) RS. Den. + Con. L.: Building upon the structure in (b), this setup adopts the proposed SRDF-SDF consistency loss. (d) RS. Den. + Con. L. + Vis.(SDF): based on (c), this configuration adds the self-supervised visibility task, in which only SDF is used to compute the visibility label. (e) RS. Den. + Con. L. + Vis.(SRDF): Different from structure (d), this method relies solely on SRDF to compute the visibility label. (f) RS. Den. + Con. L. + Vis.:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "R. Yin et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "this structure represents the comprehensive model of our method, wherein the visibility labels integrate priors from both SRDF and SDF predictions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.787, + 0.328 + ], + "angle": 0, + "content": "The comparison between configurations (a) and (b) highlights that our design using SRDF to model the density, without any additional constraints, results in a better performance compared to MonoSDF. The SRDF-SDF consistency loss in (c) surpasses the structure in (b) by \\(3.1\\%\\) in F-score. This shows the significance of sign consistency. Compared to (c), the self-supervised visibility task in (d)-(f) enhances the reconstruction performance. In particular, structure (f) obtains a better F-score than structures (d) and (e). These explain the effectiveness of our self-supervised visibility task and the computation of visibility labels integrating both SRDF and SDF priors. The visualization of the ablation study is given in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.587 + ], + "angle": 0, + "content": "Furthermore, Fig. 5 compares the weight distribution for the red and yellow sampled points generated by MonoSDF and our method. (1) The analysis of the yellow sampled point demonstrates the benefits of our method for reconstruction: MonoSDF generates high weights and negative SDF values when sampled rays are near the white wall, resulting in false surfaces that significantly deviate from the ground truth. In contrast, our method yields more accurate surfaces and thus enhances both completeness and recall. (2) The analysis of the red sampled point explains the reason why our method can generate accurate views: MonoSDF produces weights with two local maxima, causing an inaccurate RGB value. This occurs because the SDF along the camera ray fluctuates, resulting in varying density and weight. In comparison, our method produces uni-modal weights that peak only around the surface, resulting in more accurate views. The quantitative PSNR in Fig. 5b and Fig. 5d confirms that our rendered image is more accurate compared to MonoSDF. These findings align with our motivation to achieve a more consistent and accurate weight distribution, thereby reconstructing better surfaces and generating high-quality views. This underscores the effectiveness of using ray-specific SRDF to model volume density in our approach." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.61, + 0.36, + 0.627 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This work proposes a novel method for neural indoor scene reconstruction, modeling density function with ray-specific SRDF. Firstly, this study analyzed that using the SDF to parameterize the volume density may introduce noises in multi-object indoor scenes, and the SRDF exhibits stronger relationships with the actual 2D observations. Motivated by this, this work employed SRDF to model the volume density, while SDF mainly focuses on the representation of the 3D surface. Secondly, a SRDF-SDF consistency loss has been introduced to align the sign between SRDF and SDF. Thirdly, a self-supervised visibility task has been designed to distinguish whether the sampled 3D point is physically visible or occluded. It combined the prior from SDF and SRDF as visibility labels to improve geometry reconstruction without any other annotations. Experimental results on real-world and synthetic datasets showed that our method excels not only in reconstructing better surfaces but also in generating more accurate views." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.295, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Ray-Distance Volume Rendering for Neural Scene Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.214 + ], + "angle": 0, + "content": "1. Bozic, A., Palafox, P., Thies, J., Dai, A., Nießner, M.: Transformerfusion: Monocular rgb scene reconstruction using transformers. Advances in Neural Information Processing Systems 34, 1403-1414 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.215, + 0.785, + 0.255 + ], + "angle": 0, + "content": "2. Collins, R.T.: A space-sweep approach to true multi-image matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 358-363. IEEE (1996)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.256, + 0.785, + 0.296 + ], + "angle": 0, + "content": "3. Coughlan, J.M., Yuille, A.L.: Manhattan world: Compass direction from a single image by bayesian inference. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. vol. 2, pp. 941-947. IEEE (1999)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.297, + 0.785, + 0.336 + ], + "angle": 0, + "content": "4. Curless, B., Levoy, M.: A volumetric method for building complex models from range images. In: Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques. pp. 303-312 (1996)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.337, + 0.785, + 0.39 + ], + "angle": 0, + "content": "5. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5828-5839 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.391, + 0.785, + 0.445 + ], + "angle": 0, + "content": "6. Duzceker, A., Galliani, S., Vogel, C., Speciale, P., Dusmanu, M., Pollefeys, M.: Deepvideomvs: Multi-view stereo on video with recurrent spatio-temporal fusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15324-15333 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.446, + 0.785, + 0.487 + ], + "angle": 0, + "content": "7. Feng, Z., Yang, L., Guo, P., Li, B.: Cvrecon: Rethinking 3d geometric feature learning for neural reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17750-17760 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.488, + 0.785, + 0.527 + ], + "angle": 0, + "content": "8. Fu, Q., Xu, Q., Ong, Y.S., Tao, W.: Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. Advances in Neural Information Processing Systems 35, 3403-3416 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.528, + 0.785, + 0.58 + ], + "angle": 0, + "content": "9. Gallup, D., Frahm, J.M., Mordohai, P., Yang, Q., Pollefeys, M.: Real-time plane-sweeping stereo with multiple sweeping directions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1-8. IEEE (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.581, + 0.785, + 0.635 + ], + "angle": 0, + "content": "10. Guo, H., Peng, S., Lin, H., Wang, Q., Zhang, G., Bao, H., Zhou, X.: Neural 3d scene reconstruction with the manhattan-world assumption. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5511-5520 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.636, + 0.785, + 0.677 + ], + "angle": 0, + "content": "11. Hu, S., Zhou, K., Li, K., Yu, L., Hong, L., Hu, T., Li, Z., Lee, G.H., Liu, Z.: Consistentnerf: Enhancing neural radiance fields with 3d consistency for sparse view synthesis. arXiv preprint arXiv:2305.11031 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.678, + 0.785, + 0.718 + ], + "angle": 0, + "content": "12. Ju, J., Tseng, C.W., Bailo, O., Dikov, G., Ghafoorian, M.: Dg-recon: Depth-guided neural 3d scene reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18184-18194 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.719, + 0.785, + 0.745 + ], + "angle": 0, + "content": "13. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics 36(4), 1-13 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.746, + 0.785, + 0.799 + ], + "angle": 0, + "content": "14. Liang, Z., Huang, Z., Ding, C., Jia, K.: Helixsurf: A robust and efficient neural implicit surface learning of indoor scenes with iterative intertwined regularization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13165-13174 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "15. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3d surface construction algorithm. In: Seminal graphics: pioneering efforts that shaped the field, pp. 347-353 (1998)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.173, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "R. Yin et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "16. Luo, K., Guan, T., Ju, L., Huang, H., Luo, Y.: P-mvsnet: Learning patch-wise matching confidence aggregation for multi-view stereo. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10452-10461 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.23 + ], + "angle": 0, + "content": "17. Lyu, X., Dai, P., Li, Z., Yan, D., Lin, Y., Peng, Y., Qi, X.: Learning a room with the occ-sdf hybrid: Signed distance function mingled with occupancy aids scene representation. arXiv preprint arXiv:2303.09152 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.787, + 0.284 + ], + "angle": 0, + "content": "18. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4460-4470 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.787, + 0.325 + ], + "angle": 0, + "content": "19. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.787, + 0.365 + ], + "angle": 0, + "content": "20. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics 41(4), 1-15 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.365, + 0.787, + 0.407 + ], + "angle": 0, + "content": "21. Murez, Z., Van As, T., Bartolozzi, J., Sinha, A., Badrinarayanan, V., Rabinovich, A.: Atlas: End-to-end 3d scene reconstruction from posed images. In: Proceedings of the European Conference on Computer Vision. pp. 414-431. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.407, + 0.787, + 0.461 + ], + "angle": 0, + "content": "22. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5480-5490 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.787, + 0.502 + ], + "angle": 0, + "content": "23. Oechsle, M., Peng, S., Geiger, A.: Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5589-5599 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.787, + 0.555 + ], + "angle": 0, + "content": "24. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 165-174 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.787, + 0.597 + ], + "angle": 0, + "content": "25. Peng, R., Gu, X., Tang, L., Shen, S., Yu, F., Wang, R.: Gens: Generalizable neural surface reconstruction from multi-view images. In: Advances in Neural Information Processing Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.597, + 0.787, + 0.651 + ], + "angle": 0, + "content": "26. Ren, Y., Zhang, T., Pollefeys, M., Susstrunk, S., Wang, F.: Volrecon: Volume rendering of signed ray distance functions for generalizable multi-view reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16685-16695 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.787, + 0.692 + ], + "angle": 0, + "content": "27. Sayed, M., Gibson, J., Watson, J., Prisacariu, V., Firman, M., Godard, C.: Simplex: 3d reconstruction without 3d convolutions. In: Proceedings of the European Conference on Computer Vision. pp. 1-19. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.692, + 0.787, + 0.733 + ], + "angle": 0, + "content": "28. Schonberger, J.L., Zheng, E., Frahm, J.M., Pollefeys, M.: Pixelwise view selection for unstructured multi-view stereo. In: Proceedings of the European Conference on Computer Vision. pp. 501-518. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.733, + 0.787, + 0.773 + ], + "angle": 0, + "content": "29. Sitzmann, V., Martel, J., Bergman, A., Lindell, D., Wetzstein, G.: Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems 33, 7462-7473 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.773, + 0.787, + 0.799 + ], + "angle": 0, + "content": "30. Somraj, N., Soundararajan, R.: Vip-nerf: Visibility prior for sparse input neural radiance fields. arXiv preprint arXiv:2305.00041 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "31. Stier, N., Rich, A., Sen, P., Höllerer, T.: Vortex: Volumetric 3d reconstruction with transformers for voxelwise view selection and fusion. In: 2021 International Conference on 3D Vision. pp. 320-330. IEEE (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.295, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Ray-Distance Volume Rendering for Neural Scene Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "32. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.23 + ], + "angle": 0, + "content": "33. Sun, J., Xie, Y., Chen, L., Zhou, X., Bao, H.: Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15598-15607 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.23, + 0.788, + 0.285 + ], + "angle": 0, + "content": "34. Tan, M., Chen, B., Pang, R., Vasudevan, V., Sandler, M., Howard, A., Le, Q.V.: Mnasnet: Platform-aware neural architecture search for mobile. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2820-2828 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.285, + 0.788, + 0.327 + ], + "angle": 0, + "content": "35. Tang, H., Liu, Z., Zhao, S., Lin, Y., Lin, J., Wang, H., Han, S.: Searching efficient 3d architectures with sparse point-voxel convolution. In: Proceedings of the European Conference on Computer Vision. pp. 685–702. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.326, + 0.788, + 0.368 + ], + "angle": 0, + "content": "36. Wang, J., Wang, P., Long, X., Theobalt, C., Komura, T., Liu, L., Wang, W.: Neuris: Neural reconstruction of indoor scenes using normal priors. In: Proceedings of the European Conference on Computer Vision. pp. 139-155. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.367, + 0.788, + 0.394 + ], + "angle": 0, + "content": "37. Wang, K., Shen, S.: Mvdepthnet: Real-time multiview depth estimation neural network. In: 2018 International conference on 3d vision. pp. 248-257. IEEE (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.394, + 0.788, + 0.435 + ], + "angle": 0, + "content": "38. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.435, + 0.788, + 0.476 + ], + "angle": 0, + "content": "39. Wang, Y., Li, Z., Jiang, Y., Zhou, K., Cao, T., Fu, Y., Xiao, C.: Neuralroom: Geometry-constrained neural implicit surfaces for indoor scene reconstruction. ACM Transactions on Graphics 41(6), 1-15 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.476, + 0.788, + 0.517 + ], + "angle": 0, + "content": "40. Yao, Y., Luo, Z., Li, S., Fang, T., Quan, L.: Mvsnet: Depth inference for unstructured multi-view stereo. In: Proceedings of the European Conference on Computer Vision. pp. 767-783 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.517, + 0.788, + 0.559 + ], + "angle": 0, + "content": "41. Yao, Y., Luo, Z., Li, S., Shen, T., Fang, T., Quan, L.: Recurrent mvsnet for high-resolution multi-view stereo depth inference. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5525-5534 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.558, + 0.788, + 0.585 + ], + "angle": 0, + "content": "42. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34, 4805-4815 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.626 + ], + "angle": 0, + "content": "43. Ye, B., Liu, S., Li, X., Yang, M.H.: Self-supervised super-plane for neural 3d reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21415-21424 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.626, + 0.788, + 0.667 + ], + "angle": 0, + "content": "44. Yin, R., Karaoglu, S., Gevers, T.: Geometry-guided feature learning and fusion for indoor scene reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3652-3661 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.667, + 0.788, + 0.708 + ], + "angle": 0, + "content": "45. Yu, Z., Gao, S.: Fast-mvsnet: Sparse-to-dense multi-view stereo with learned propagation and gauss-newton refinement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1949-1958 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.708, + 0.788, + 0.748 + ], + "angle": 0, + "content": "46. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in Neural Information Processing Systems 35, 25018-25032 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.748, + 0.788, + 0.789 + ], + "angle": 0, + "content": "47. Zhu, Z., Peng, S., Larsson, V., Cui, Z., Oswald, M.R., Geiger, A., Pollefeys, M.: Nicer-slam: Neural implicit scene encoding for rgb slam. arXiv preprint arXiv:2302.03594 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.789, + 0.788, + 0.831 + ], + "angle": 0, + "content": "48. Zins, P., Xu, Y., Boyer, E., Wuhrer, S., Tung, T.: Multi-view reconstruction using signed ray distance functions (srdf). In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16696-16706 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.831 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Ray-Distance Volume Rendering for Neural Scene Reconstruction/e689642e-0cb9-456e-9b14-7ba1bd1312a3_origin.pdf b/2024/Ray-Distance Volume Rendering for Neural Scene Reconstruction/e689642e-0cb9-456e-9b14-7ba1bd1312a3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d9093d52343f444ad3423d0a5f6a006e3f5004df --- /dev/null +++ b/2024/Ray-Distance Volume Rendering for Neural Scene Reconstruction/e689642e-0cb9-456e-9b14-7ba1bd1312a3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aeff6c71a8ce6ce6a547e704862263afed10fd8cdd876ba0ef3274afa37064af +size 4667198 diff --git a/2024/Ray-Distance Volume Rendering for Neural Scene Reconstruction/full.md b/2024/Ray-Distance Volume Rendering for Neural Scene Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cbf15c8fdce64072ab5dbeaeb5a20b828cbc89bb --- /dev/null +++ b/2024/Ray-Distance Volume Rendering for Neural Scene Reconstruction/full.md @@ -0,0 +1,326 @@ +# Ray-Distance Volume Rendering for Neural Scene Reconstruction + +Ruihong Yin $^{1}$ , Yunlu Chen $^{2}$ , Sezer Karaoglu $^{1,3}$ , and Theo Gevers $^{1,3}$ + +1 University of Amsterdam, Amsterdam, The Netherlands + +2 Carnegie Mellon University, Pittsburgh, The United States + +3 3DUniversum, Amsterdam, The Netherlands + +{r.yin, Th.Gevers}@uva.nl, s.karaoglu@3duniversum.com, + +yunluche@andrew.cmu.edu + +Abstract. Existing methods in neural scene reconstruction utilize the Signed Distance Function (SDF) to model the density function. However, in indoor scenes, the density computed from the SDF for a sampled point may not consistently reflect its real importance in volume rendering, often due to the influence of neighboring objects. To tackle this issue, our work proposes a novel approach for indoor scene reconstruction, which instead parameterizes the density function with the Signed Ray Distance Function (SRDF). Firstly, the SRDF is predicted by the network and transformed to a ray-conditioned density function for volume rendering. We argue that the ray-specific SRDF only considers the surface along the camera ray, from which the derived density function is more consistent to the real occupancy than that from the SDF. Secondly, although SRDF and SDF represent different aspects of scene geometries, their values should share the same sign indicating the underlying spatial occupancy. Therefore, this work introduces a SRDF-SDF consistency loss to constrain the signs of the SRDF and SDF outputs. Thirdly, this work proposes a self-supervised visibility task, introducing the physical visibility geometry to the reconstruction task. The visibility task combines prior from predicted SRDF and SDF as pseudo labels, and contributes to generating more accurate 3D geometry. Our method implemented with different representations has been validated on indoor datasets, achieving improved performance in both reconstruction and view synthesis. + +Keywords: Indoor scene reconstruction $\cdot$ neural radiance fields $\cdot$ signed ray distance function + +# 1 Introduction + +Indoor 3D scene reconstruction involves using multi-view RGB images as input to generate detailed 3D geometry as output. This task is crucial and has widespread applications in virtual/mixed augmentation, robotics, navigation, and so on. Contrary to a single object, indoor environments contain various elements, such as chairs, walls, tables, and cups, which present a challenge for the task of reconstruction. Traditional reconstruction approaches [1,21,27,33,40,44] + +use 2D and 3D convolutional neural networks (CNNs) [34, 35] to extract features. They can be classified as depth-based methods [6, 27, 37, 40] and volumetric methods [1, 21, 31, 33, 44]. Depth-based techniques estimate depth for every viewpoint, encountering difficulties like scale uncertainties and inconsistencies in depth across different views. The surfaces created through these techniques can be coarse and imprecise. Unlike depth-based approaches, volumetric techniques utilize 3D CNNs to directly generate 3D geometry, leading to smoother surfaces. Nevertheless, these methods use voxels for scene representation, causing GPU memory usage and computational expenses to grow cubically as voxel resolution increases. Such a limitation restricts the use of volumetric approaches for high-quality scene reconstruction. Moreover, volumetric techniques require 3D annotations for network training, significantly increasing the cost of annotations. + +Implicit representations [18,24,29] encode 3D models as functions of coordinates. Utilizing implicit representations and volume rendering, neural radiance fields (NeRF) [19] achieves remarkable performance in novel view synthesis, even without 3D supervision. Some methods [38,42] exploit the capabilities of NeRF in reconstruction tasks. These methods typically parameterize the volume density as a learnable transformation of the Signed Distance Function (SDF), leading to impressive reconstruction performance. Compared to classic volumetric methods, neural implicit scene representation is more memory-efficient, marking a new trend in reconstruction tasks. + +In multi-object indoor scenes, the SDF from sampled points considers surface points across the entire scene. However, for neural implicit scene reconstruction, along a camera ray, from the camera center to the intersected surface point, the SDF may exhibit fluctuations, leading to several local minimum values. Fig. 1 provides a toy example to illustrate this. In Fig. 1c, volume density yielded from the fluctuating SDF displays several local maximum values, contributing to high weights in Fig. 1d. However, the camera ray $\overrightarrow{OP}$ targets the green rectangle, while the SDF of the near-surface point $\mathbf{Q}$ refers to the surface of the blue cylinder. Besides, $\mathbf{Q}$ is far from the intersected point $\mathbf{P}$ , but still corresponds to high density and high weight. Nevertheless, during volume rendering that maps the 3D output to 2D space, it is evident that along a ray, points closer to the intersected surface point may carry a higher weight as they are more relevant to actual observations. Consequently, in the multi-object scenes, using the SDF to model the density function along the camera ray may disrupt this relationship, introduce noise, and lead to inaccuracies in 2D results or 3D geometries. + +This paper proposes RS-Recon, a novel indoor scene reconstruction method with the ray-specific density function to address the aforementioned issue. Firstly, our method models the density function as a function of the Signed Ray Distance Function (SRDF). Unlike SDF, SRDF measures the shortest distance to the surface along a camera ray, eliminating the influence of surrounding surfaces that are not located on the given ray. Moreover, the density distribution generated from SRDF attains the local maximum only around the intersected surface, aligning more closely to the concept that points nearer to the surface have more importance. To achieve this, our network predicts both SRDF and SDF, where + +SRDF encompasses the density function, while SDF mainly focuses on describing the 3D surfaces. Secondly, despite SRDF and SDF being defined differently, both of them produce positive values outside the object and negative values within. Therefore, in this paper, a SRDF-SDF consistency loss is devised to ensure they share the same sign. Thirdly, this work introduces a self-supervised visibility task to enhance the prediction of 3D geometry. The visibility task predicts the probability of whether the sampled points are physically visible along the ray in 3D space. The pseudo visibility ground truth is formed by combining the prior of SRDF and SDF from the network itself, without the reliance on multi-view geometry or any additional annotations. + +Our contributions can be summarized as follows, (1) This work proposes a novel neural scene reconstruction method that leverages the ray-specific SRDF to model the volume density. Compared to SDF, SRDF better reflects the significance of the real observations. (2) A SRDF-SDF consistency loss is proposed to constrain the alignment between SRDF and SDF, ensuring the same sign inside and outside the object. (3) To enhance the accuracy of 3D geometry, an additional self-supervised visibility task is introduced to predict the visibility probability of sampled points. This task integrates prior information from both the SDF and SRDF predictions as labels. (4) The experiments conducted on both synthetic and real-world indoor datasets illustrate that our method enhances performance in both reconstruction and view synthesis. + +# 2 Related Works + +Classic Indoor Scene Reconstruction. Traditional methods for indoor scene reconstruction fall into two classes, i.e. depth-based methods [6, 16, 27, 37, 40, 41, 45] and volumetric methods [1, 7, 12, 21, 31, 33, 44]. Depth-based methods predict pixel-level depth for individual frames, and subsequently depth fusion [4] is adopted to generate the 3D scene. To achieve favorable outcomes, these methods often construct a plane sweep cost volume [2, 9] at the feature or image level, leveraging information from multiple views to complement the current frame, e.g. MVSNet [40], DeepVideoMVS [6], and SimpleRecon [27]. Using the depth map as the intermediate representation, these methods encounter challenges such as depth inconsistency and scale ambiguities. In contrast, volumetric methods utilize 3D CNN to directly regress the 3D geometry, which can generate smooth surfaces and reconstruct unobserved regions. For instance, NeuralRecon [33] designs a learnable TSDF fusion module to integrate features from previous frames and predicts the TSDF for sparse volumes. TransformerFusion [1], based on volumetric representation, adopts Transformer to fuse multi-view features and select frames with higher attention. Yin et al. [44] propose to incorporate geometry priors at different levels for the volumetric methods. However, due to the use of 3D CNN in volumetric methods, computational costs escalate significantly with higher resolutions. + +Neural Implicit Scene Representation. With the success of implicit representation and NeRF [19], neural implicit scene representation has emerged as a + +![](images/0d205758a58093843359625d17fc030815f1eca3d726db456a5906b5306dc3de.jpg) + +![](images/75bd0a2c2b372345578f26a2377ef97ca7ff20f0c644f70e6fd2273b7185103e.jpg) +(a) Multi-object scene + +![](images/68ad60a3fce54a38647b70c1fdb4cb37d0440a6c04e04063cc60fe2907447435.jpg) + +![](images/4d3607c5da3c7578b994bf6b47e117190d1ac2e2699f0aaf50ea6dd49d324810.jpg) +(b) SDF/SRDF + +![](images/69282ed5d06f58da24998c4c8a30bbae90c51791cc0878d999b3df40241bdbd8.jpg) + +![](images/09844e35fe2dded36b0d53c850e551c27a62bd2cb180ce8fb9425a33fce69394.jpg) +(c) Density +Fig. 1: Analysis of the density function and weight distribution for a toy scene. (a) Frontal and overhead perspectives of a scene with multiple objects, where a ray originates from the camera center $\mathbf{O}$ , intersecting with the green rectangle at the point $\mathbf{P}$ . $\mathbf{Q}$ is the closest point to the blue cylinder along the ray $\overrightarrow{OP}$ . (b) The SDF $d_{\Omega}$ /SRDF $\tilde{d}_{\Omega}$ in Eqs. (2) and (4) along the ray $\overrightarrow{OP}$ . (c) The volume density $\sigma$ generated from the SDF/SRDF in (b) using Eqs. (3) and (5). (d) The weight distribution $T_{i}\alpha_{i}$ generated from the density in (c) via Eq. (1). Although $\mathbf{Q}$ is distant from the surface boundary intersecting with the ray, the density function from the SDF in $\mathbf{Q}$ generates a high weight in volume rendering, resulting in noisy rendering and reconstruction. + +![](images/31fa9cf154536cc37d162b436657800c9bd388cf03b3e85dce138bb1de71e37e.jpg) + +![](images/8df39fee97f9f9281441f8fdffc9be72387a86e21ff05b87a28f63540f2fefb9.jpg) +(d) Weight + +hot spot in the past few years. Generally, reconstruction approaches based on neural implicit scene representation first map point positions to a continuous representation of SDF or occupancy by multi-layer perceptrons (MLPs), then transfer SDF or occupancy to volume density, and finally render the 2D results from the 3D outputs. Despite requiring only 2D supervision during training, these methods can reconstruct high-quality surfaces and are more computationally efficient. For example, VolSDF [42] parameterizes the volume density as a Laplace distribution of SDF, while NeuS [38] models it as the logistic distribution. Some extensions [10,36,39,46,47] introduce extra priors to achieve superior outcomes. For instance, Manhattan-SDF [10] employs planar constraints based on Manhattan-world assumption [3]. NeuRIS [36] proposes to combine normal prior for the reconstruction of texture-less regions, while NeuralRoom [39] implements perturbation-residual restriction to reconstruct the flat region better. MonoSDF [46] exploits both depth and normal cues to improve the reconstruction quality. Although additional priors contribute to the accuracy and completeness of reconstructed surfaces, they employ the SDF to model the volume density, potentially resulting in noisy 2D results in the multi-object scenes. Contrarily, our method advocates for parameterizing the density function with the ray-specific SRDF, resulting in enhanced reconstruction performance. + +Some NeRF-based methods also utilize multi-view geometry to introduce additional regularization. For instance, RegNeRF [22] regularizes patched-based depth maps from unobserved viewpoints. Geo-NeuS [8] and ConsistentNeRF [11] use multi-view photometric consistency. SRDF and observation consistencies are used by [48]. GenS [25] introduces multi-view and multi-scale feature consistent + +![](images/e50b7d5fd484c8d5e91491dca755596776e5e837e5334511d467449d7665368d.jpg) +Fig. 2: Our framework. A geometry MLP $f_{g}$ is adopted to generate the SDF and geometry features $\mathbf{F}_{g}$ , utilizing the encoded position (and optionally grid features) as input. Then, $\mathbf{F}_{g}$ , along with viewing direction, passes through a color MLP $f_{c}$ to predict the color for each point. Notably, (1) our approach advocates for modeling the density function with the ray-specific SRDF in addition to the SDF. For this purpose, a SRDF MLP $f_{s}$ is introduced to generate the SRDF. (2) A SRDF-SDF consistency loss $\mathcal{L}_{con}$ is devised to align the signs between the generated SRDF and SDF. (3) To enhance the geometry prediction, a self-supervised visibility task is proposed to integrate geometry priors in both SRDF and SDF predicted by the network and generates the pseudo visibility ground truth. The visibility probability is predicted by the SRDF MLP. + +cies for neural surface reconstruction. VIP-NeRF [30] adopts a visibility prior, using the plane sweep volume among multiple views to generate visibility labels. However, they depend on the computation of multi-view geometry, some of which may involve processing multiple frames yielding increased computational demands. In contrast, this paper introduces a self-supervised visibility task that requires minimal computational increases and does not depend on multi-view geometry for ground truth generation. + +# 3 Method + +In this section, we provide the background on NeRF-based reconstruction methods, which typically model the volume density as a transformation of SDF (refer to Sec. 3.1). To overcome the issue of false local maxima in the density from SDF, our approach covered in Sec. 3.2 - Sec. 3.5 proposes to model the density function with the ray-specific SRDF in neural scene reconstruction, which achieves more accurate reconstruction. Fig. 2 presents the framework of our approach. + +# 3.1 Background: Volume Rendering Density as Transformed SDF + +Our goal is to recover the scene geometry given a set of projected 2D views of a 3D scene. To achieve this, NeRF [19] optimizes a continuous 5D function to represent the scene, which predicts the volume density $\sigma(\mathbf{p})$ and view-dependent color $\mathbf{c}(\mathbf{p},\mathbf{r})$ for each sampled point $\mathbf{p} \in \mathbb{R}$ and ray direction $\mathbf{r} \in S^2$ . With $\Omega \subset \mathbb{R}^3$ + +denoting the scene geometry filled with physical presence, the predicted density $\sigma (\mathbf{p})$ approximates the ground truth scene density $\mathbf{1}_{\varOmega}(\mathbf{p})$ , such that $\mathbf{1}_{\varOmega}(\mathbf{p}) = 1$ if $\mathbf{p}\in \varOmega$ and 0 otherwise. After that, the classic volume rendering is employed to render the color $\mathbf{C}(\mathbf{r})$ , and $N$ points $\{\mathbf{p}_i = \mathbf{o} + z_i\mathbf{r}, i\in [1,N]\}$ are sampled along the camera ray $\mathbf{r}$ , with $z_{i}$ indicating the depth from the camera center $\mathbf{o}$ to the sampled point. The color is accumulated along the ray: + +$$ +\mathbf {C} (\mathbf {r}) = \sum_ {i = 1} ^ {N} T _ {i} \alpha_ {i} \mathbf {c} \left(\mathbf {p} _ {i}, \mathbf {r}\right), T _ {i} = \prod_ {j = 1} ^ {i - 1} \left(1 - \alpha_ {i}\right), \alpha_ {i} = 1 - \exp \left(- \sigma \left(\mathbf {p} _ {i}\right) \delta_ {i}\right) \tag {1} +$$ + +where $\delta_{i}$ is the interval of neighboring points. + +Despite the great success in novel view synthesis, NeRF has difficulty reconstructing satisfactory actual surfaces from the generated volume density. To address this challenge, recent methods [42,46] propose to predict the SDF and then transform it to volume density, which aligns with the geometry bias that the surface points have a higher density than other non-surface points and corresponds to a higher weight during volume rendering. + +For any sampled point $\mathbf{p} \in \mathbb{R}^3$ , the SDF $d_{\Omega}(\mathbf{p})$ is with an absolute value representing the shortest distance from $\mathbf{p}$ to the surface $\mathcal{M} = \partial \Omega$ , while the sign indicates whether the point is outside (positive) or inside (negative) the surface. Formally, + +$$ +\operatorname {S D F}: \quad d _ {\Omega} (\mathbf {p}) = (- 1) ^ {\mathbf {1} _ {\Omega} (\mathbf {p})} \min _ {\mathbf {p} ^ {*} \in \mathcal {M}} \| \mathbf {p} - \mathbf {p} ^ {*} \| _ {2}, \tag {2} +$$ + +where the indicator function $\mathbf{1}_{\varOmega}(\mathbf{p}) = 1$ if $\mathbf{p}\in \varOmega$ and 0 otherwise. This indicates that the SDF considers every surface around the point. To apply SDF to volume rendering, Yariv et al. [42] suggest to derive the volume density $\sigma^{\mathrm{SDF}}$ from the signed distance as follows: + +$$ +\sigma^ {\mathrm {S D F}} (\mathbf {p}) = \alpha \Psi_ {\beta} (- d _ {\Omega} (\mathbf {p})) \text {,} \Psi_ {\beta} (s) = \left\{ \begin{array}{l l} \frac {1}{2} \exp \left(\frac {s}{\beta}\right) & \text {i f} s \leq 0 \\ 1 - \frac {1}{2} \exp \left(- \frac {s}{\beta}\right) & \text {i f} s > 0 \end{array} \right. \tag {3} +$$ + +in which $\alpha, \beta$ are learnable parameters. $\Psi_{\beta}$ is the Cumulative Distribution Function (CDF) of the Laplace distribution with zero mean and $\beta$ scale. Then, $\sigma^{\mathrm{SDF}}(\mathbf{p})$ is applied to Eq. (1) and as the density function $\sigma$ . + +# 3.2 Accurate Scene Volume Density from Ray-Specific SRDF + +We illustrate the issues with volume rendering SDF by a toy example in Fig. 1. Indoor scenes typically include multiple objects. Along the camera ray $\mathbf{r}$ , the SDF may be influenced by surfaces not intersecting the ray, resulting in several ambiguous local maxima of false high volume density and weight. This implies that the 2D observations may have a strong correlation to these distant points, contradicting the fact that the points closer to the surface along the camera ray may carry more significance. + +To this end, we propose volume rendering scene density function from SRDF, the distance field local to the ray direction. In contrast to the SDF in Eq. (2), + +the SRDF $\tilde{d_{\Omega}}$ ( $\mathbf{p}, \mathbf{r}$ ) [26,48] computes the shortest distance from the point to the surface along the sampled ray $\mathbf{r}$ , i.e. the SRDF is ray-dependent. Formally, the SRDF $\tilde{d_{\Omega}}$ ( $\mathbf{p}, \mathbf{r}$ ) can be represented by + +$$ +\text {S R D F :} \quad \tilde {d _ {\Omega}} (\mathbf {p}, \mathbf {r}) = (- 1) ^ {\mathbf {1} _ {\Omega} (\mathbf {p})} \min _ {(\mathbf {p} + \rho \mathbf {r}) \in \mathcal {M}; \rho \in \mathbb {R}} | \rho |, \tag {4} +$$ + +where $\mathbf{1}_{\Omega}(\mathbf{p}) = 1$ if $\mathbf{p} \in \Omega$ and $0$ otherwise. In the following sections, $d_{\Omega}$ from Eq. (2) and $\tilde{d_{\Omega}}$ are used to denote SDF and SRDF for brevity. + +In Fig. 1, it can be seen that the weight computed from SRDF in the volume rendering only reaches a local peak around the surface point, aligning more consistently with the nature of the 2D observations. Therefore, this paper proposes to employ the ray-specific SRDF to yield the volume density instead of the SDF. This work predicts both SDF and SRDF, where SDF is mainly used to model the surface. As depicted in Fig. 2, for a sampled point $\mathbf{p}$ , a geometry MLP $f_{g}$ is applied on the input (including the encoded position or features) to predict SDF $d_{\Omega}$ and geometry feature $\mathbf{F}_{g}$ . By definition, SRDF is view-dependent. Hence, the viewing direction $\mathbf{r}$ , geometry feature $\mathbf{F}_{g}$ , and point position $\mathbf{p}$ are concatenated and then passed through a SRDF MLP $f_{s}$ to yield the SRDF. On the other hand, the viewing direction $\mathbf{r}$ , geometry feature $\mathbf{F}_{g}$ , point position $\mathbf{p}$ , and 3D unit normal $\mathbf{n} = \nabla d_{\Omega}$ are used as input to a color MLP $f_{c}$ to output color $\mathbf{c}$ . + +For volume rendering, our method adopts the predicted SRDF $\tilde{d_{\Omega}}$ to derive the volume density $\sigma$ . This is defined by: + +$$ +\sigma^ {\mathrm {S R D F}} (\mathbf {p}, \mathbf {r}) = \alpha \Psi_ {\beta} \left(- \tilde {d _ {\Omega}} (\mathbf {p}, \mathbf {r})\right). \tag {5} +$$ + +Then, $\sigma^{\mathrm{SRDF}}(\mathbf{p},\mathbf{r})$ is applied to the volume rendering in Eq. (1) as a ray-conditioned density function to obtain the color value $\mathbf{C}^{\mathrm{SRDF}}(\mathbf{r})$ . Besides, during training, we also render color $\mathbf{C}^{\mathrm{SDF}}(\mathbf{r})$ with density functions from SDF to obtain gradient signals to optimize the prediction of SDF, in which the volume rendering processes use the same rays and points as those for $\mathbf{C}^{\mathrm{SRDF}}(\mathbf{r})$ . + +For the scene geometry, after the network is trained with NeRF optimization targets, the geometric surface $\mathcal{M} = \partial \Omega$ is extracted as watertight meshes by the marching cubes algorithm [15,24] from the SDF results $d_{\Omega}$ , which only depends on the query location $\mathbf{p}$ . + +# 3.3 SRDF-SDF Consistency Loss + +Following the definitions in Eq. (4), despite SRDF and SDF being defined differently, they share the same sign, which indicates the spatial occupancy — positive outside the surface and negative inside the surface. However, in this paper, where SDF and SRDF are predicted from separate branches, sign consistency is not guaranteed. To deal with this issue, a SRDF-SDF consistency loss is proposed to enforce this constraint. + +Although the sign function effectively computes the output's sign, it cannot be differentiated during back-propagation. To a certain degree, the values derived + +from the sigmoid function approximate those calculated by the sign function. Thus, in our SRDF-SDF consistency loss, the sigmoid function is first employed to generate the sign of SRDF and SDF as the approximated sign function $\varsigma (\cdot)$ , after which a $\ell_2$ loss is adopted to quantify the sign difference, as defined by: + +$$ +\mathcal {L} _ {c o n} = \frac {1}{N _ {r}} \sum_ {\mathbf {p}, \mathbf {r}} M _ {c o n} | | \varsigma (\tilde {d} _ {\Omega}) - \varsigma (d _ {\Omega}) | | _ {2} \tag {6} +$$ + +$$ +\varsigma (d) = \mathsf {S i g m o i d} (k \cdot d), M _ {c o n} = [ \widetilde {d} _ {\Omega} \cdot d _ {\Omega} < 0 ] +$$ + +in which $k$ is a hyperparameter, controlling the slope of the sigmoid function. The indicator function $[\tilde{d}_{\Omega} \cdot d_{\Omega} < 0]$ is leveraged to identify whether the generated SDF and SRDF have opposite signs. Therefore, the consistency loss imposes penalties when there is a difference between the signs of SRDF and SDF. $N_r$ represents the number of sampled points in a minibatch that satisfy $[\tilde{d}_{\Omega} \cdot d_{\Omega} < 0]$ . + +Taking the derivative of $\mathcal{L}_{con}$ with respect to SDF $d_{\Omega}$ (as an example), is computed as follows: + +$$ +\frac {\partial \mathcal {L} _ {c o n}}{\partial d _ {\Omega}} = - \frac {2}{N _ {r}} M _ {c o n} | | \varsigma (\tilde {d} _ {\Omega}) - \varsigma (d _ {\Omega}) | | _ {1} \cdot \frac {\partial \varsigma (d _ {\Omega})}{\partial d _ {\Omega}} \tag {7} +$$ + +The derivative in Eq. (7) shows that our consistency loss has two advantages: (1) Unlike the sign function, which assigns the same penalty for outputs with different signs of SRDF and SDF, our consistency loss can adjust the penalty according to the extent of the sign discrepancy between SRDF and SDF, referring to $||\varsigma (\tilde{d}_{\varOmega}) - \varsigma (d_{\varOmega})||_1$ . (2) The derivative of the sigmoid function $\frac{\partial\varsigma(d_{\varOmega})}{\partial d_{\varOmega}}$ peaks at zero and decreases from 0 to $+\infty / - \infty$ . Consequently, predictions in proximity to zero (indicating the surface) may exhibit higher absolute gradient values, providing effective supervision for points near the surface. + +# 3.4 Self-supervised Visibility Task + +Theoretically, with visibility attributes of 3D points, surface location becomes straightforward. Therefore, to enhance 3D geometry prediction, we propose a self-supervised visibility task in this paper. Along the trajectory of a camera ray, points sampled prior to reaching the first surface point are categorized as visible, whereas points sampled beyond the first surface point are classified as occluded. To find the first surface point along the ray, using SRDF as an example, the multiplication of SRDF between adjacent sampled points is computed. When $\tilde{d}_{\Omega}^{i}\cdot \tilde{d}_{\Omega}^{i + 1}\leq 0$ and $\tilde{d}_{\Omega}^{m}\cdot \tilde{d}_{\Omega}^{m + 1} > 0,m\in [1,i - 1]$ , it indicates that the first surface is within the interval $[\mathbf{o} + z_i\mathbf{r},\mathbf{o} + z_{i + 1}\mathbf{r}]$ . Consequently, the points between $[\mathbf{p}_1,\mathbf{p}_i]$ are considered physically visible while points between $[\mathbf{p}_{i + 1},\mathbf{p}_N]$ are regarded as occluded. However, the SRDF or SDF predicted by the network may be inaccurate at some points, introducing potential noise in the visibility task. To improve the accuracy of the visibility labeling, both SDF and SRDF are utilized in determining visibility. Points are labeled as visible or occluded only when there + +Table 1: Evaluation of 3D reconstruction meshes on ScanNet. $\dagger$ indicates our reproduced results. The best results are bold. $\downarrow$ denotes that a lower value is preferable while $\uparrow$ means higher value corresponds to better performance. + +
MethodAcc ↓Comp ↓Prec ↑Recall ↑F-score ↑
COLMAP [28]0.0470.2350.7110.4410.537
UNISURF [23]0.5540.1640.2120.3620.267
NeuS [38]0.1790.2080.3130.2750.291
VolSDF [42]0.4140.1200.3210.3940.346
Manhattan-SDF [10]0.0720.0680.6210.5860.602
S3P [43]0.0550.0590.7090.6600.683
NeuRIS [36]0.0500.0490.7170.6690.692
MonoSDF_Grid [46]0.0720.0570.6600.6010.626
MonoSDF_MLP [46]0.0350.0480.7990.6810.733
HelixSurf [14]0.0380.0440.7860.7270.755
Occ_SDF_Hybrid† [17]0.0400.0410.7830.7480.765
Ours_Grid0.0740.0490.6700.7030.683
Ours_MLP0.0400.0400.8090.7790.794
+ +is agreement between SRDF and SDF on their classification. If the classifications from SRDF and SDF diverge, suggesting difficulties in accurately categorizing these points for the network, such points are omitted from the visibility task. The method for assigning visibility labels is defined as follows + +$$ +V _ {g t} = \left\{ \begin{array}{l l} 1, & \text {i f} (V ^ {\mathrm {S R D F}} = 1) \& (V ^ {\mathrm {S D F}} = 1) \\ 0, & \text {i f} (V ^ {\mathrm {S R D F}} = 0) \& (V ^ {\mathrm {S D F}} = 0) \end{array} \right. \tag {8} +$$ + +where $V^{\mathrm{SRDF}} / V^{\mathrm{SDF}}$ are the visibility labels determined based on the prediction of SRDF/SDF. 1/0 means label for visible/occluded points. The formation of the visibility label can be inferred as self-supervised, thus eliminating the necessity for multi-view geometry and additional annotations. Additionally, it incorporates information from both SRDF and SDF. + +From the analysis above, similar to SRDF, the visibility task also relies on the view direction. Therefore, a visibility probability $V_{pred}$ is predicted from the SRDF branch. It is a binary classification task, i.e. visible or occluded, so the binary cross-entropy loss is employed as visibility loss $\mathcal{L}_{vis}$ during optimization. + +The visibility task can discern the point's visibility along the camera ray, thereby influencing the learning of SRDF. Moreover, the optimization process integrates knowledge of ray-related visibility across the entire space, further impacting the learning of SDF. Notably, unlike 2D supervision, the visibility task provides priors in 3D space, imparting a more direct influence on 3D geometry. + +# 3.5 Optimization + +Loss Functions. During training, the overall loss consists of RGB loss $\mathcal{L}_c$ , normal loss $\mathcal{L}_n$ , depth loss $\mathcal{L}_d$ , Eikonal loss $\mathcal{L}_e$ , smooth loss $\mathcal{L}_s$ , SRDF-SDF + +consistency loss $\mathcal{L}_{con}$ , and visibility loss $\mathcal{L}_{vis}$ + +$$ +\mathcal {L} = \mathcal {L} _ {c} + \lambda_ {n} \mathcal {L} _ {n} + \lambda_ {d} \mathcal {L} _ {d} + \lambda_ {e} \mathcal {L} _ {e} + \lambda_ {s} \mathcal {L} _ {s} + \lambda_ {c o n} \mathcal {L} _ {c o n} + \lambda_ {v i s} \mathcal {L} _ {v i s} \tag {9} +$$ + +in which $\lambda_{(\cdot)}$ are the loss weight. Following MonoSDF [46], the 2.5D depth and normal maps are volume rendered using the density function $\sigma^{\mathrm{SRDF}}$ to exploit geometric cues and support training. More details on the computation of loss functions are provided in the supplementary material. + +Implementation Details. Theoretically, our method can be applied to replace SDF-based volume rendering in indoor scene reconstruction. To measure our method, VolSDF-based reconstruction method MonoSDF is considered as our baseline in the following experiments. Furthermore, in the supplementary material, we also verify our method on NeuS-based reconstruction method NeuRIS. We used different spatial encoding techniques as input, including hash features [20] and Fourier features [19], referred to as Grid and MLP respectively in the following experiments. More details are given in the supplementary material. + +# 4 Experiments + +# 4.1 Experimental Setup + +Datasets. Our proposed method is evaluated on three indoor datasets, comprising: (1) four scenes from the real-world dataset ScanNet [5], (2) eight scenes from the synthetic dataset Replica [32], and (3) four scenes from the advanced set of the real-world large-scale dataset Tanks and Temples [13]. + +Metrics. For ScanNet, the assessment of the results involves five reconstruction metrics: accuracy (acc), completeness (comp), precision (prec), recall, and F-score with a threshold of 5cm. For Replica, in alignment with MonoSDF, metrics including normal consistency (normal c.), chamfer L1 distance (Chamfer-L1), and F-score with a threshold of 5cm are utilized. For Tanks and Temples, as evaluated by the official server, results are reported based on the F-score with a 1cm threshold. Additionally, this study assesses the peak signal-to-noise ratio (PSNR) of the rendered 2D RGBs to showcase the effectiveness of our approach. Detailed computations for all metrics are given in the supplementary material. + +Baselines. Except for the baseline MonoSDF, our method is also compared to (1) the classic MVS method: COLMAP [28], (2) neural implicit reconstruction methods: UNISURF [23], NeuS [38], VolSDF [42], Manhattan-SDF [10], NeuRIS [36], S3P [43], HelixSurf [14], and Occ_SDF_Hybrid [17]. + +# 4.2 Evaluation Results + +Reconstruction on ScanNet [5]. Tab. 1 compares our method against other state-of-the-art (SOTA) methods on ScanNet. When contrasted with the baseline MonoSDF, our method consistently shows better performance across all metrics, regardless of whether utilizing MLP or grid representation. For example, employing grid representation, our method exhibits a $5.7\%$ higher F-score than + +![](images/3aadb8698266b3f348af8bc1a20a549989e6fb0f3c1195c790e6de5f6b412399.jpg) +Fig. 3: Qualitative comparisons on ScanNet. It can be seen that our method can reconstruct more surfaces, especially in thin regions. + +![](images/08b9dbfc9bc541e6db6461ca87301e96a223958047e861a36a3ec8920f0fbbfc.jpg) +Fig.4: Qualitative comparisons on Replica. Compared to MonoSDF and Occ_SDF_Hybrid, our method can generate more accurate surfaces. + +MonoSDF. Using the MLP representation, our method outperforms MonoSDF by $1.0\%$ in recall, $9.8\%$ in precision, and $6.1\%$ in F-score. In addition, our method with the MLP representation surpasses other SOTA methods on almost all metrics, explaining the effectiveness of our method on reconstruction. Notably, our method achieves a $2.9\%$ higher F-score compared to Occ_SDF_Hybrid. + +Reconstruction on Replica [32]. The results on the synthetic dataset Replica are shown in Tab. 2. Using the grid representation, our method reports a notable improvement of $9.91\%$ in chamfer distance and $2.88\%$ in F-score. Similar increases are also illustrated in our method with the MLP representation across all metrics. Our approach also achieves on-par results with Occ_SDF_Hybrid. Reconstruction on Tanks and Temples [13]. Tab. 3 shows quantitative outcomes on the challenging large-scale dataset Tanks and Temples. The F-score for each scene - Auditorium, Ballroom, Courtroom, and Museum - along with + +Table 2: Evaluation of 3D reconstruction meshes on Replica. + +
MethodNormal C. ↑Chamfer-L1 ↓F-score ↑
GridMonoSDF [46]90.933.2385.91
Ours91.662.9188.79
MLPMonoSDF [46]92.112.9486.18
Occ_SDF_Hybrid†[17]93.342.5891.54
Ours93.492.6091.72
+ +Table 3: Evaluation of 3D reconstruction meshes on the advanced set of Tanks and Temples. The F-score is reported in the table. + +
MethodAuditoriumBallroomCourtroomMuseummean
GridMonoSDF [46]3.173.7013.755.686.58
Occ_SDF_Hybrid†[17]3.763.5814.047.277.16
Ours4.844.3114.187.607.73
MLPMonoSDF [46]3.092.4710.005.105.17
Occ_SDF_Hybrid†[17]3.213.3410.294.665.38
Ours4.244.6910.795.916.41
+ +their mean, is reported. Our method, with both the MLP and grid representations, outperforms Occ_SDF_Hybrid and MonoSDF. In particular, our F-score is $1.15\%$ higher (grid representation) and $1.24\%$ higher (MLP representation) than MonoSDF across all scenes. + +Visualization of Reconstruction Results. Fig. 3 and Fig. 4 report qualitative results on ScanNet and Replica respectively. In contrast to other methods, our approach can produce superior meshes and recall more regions, particularly for thin regions/small objects, such as chair legs and table legs. Although our method relies on the SRDF for improved density in volume rendering while the surface geometry is extracted by the SDF, both the proposed SRDF-SDF consistency loss and self-supervised visibility task guide the optimization of the SDF to be more consistent with the SRDF, resulting in more accurate surface boundary. Additional visualization is provided in the supplementary material. + +View Synthesis. Tab. 4 also provides a comparison of view synthesis performance on training views. Notably, 2D views are rendered with density derived + +Table 4: Evaluation of the rendering quality. PSNR is given in this table. + +
MethodScanNetReplicaTanks and Temples
GridMonoSDF [46]28.9640.2528.45
Ours30.0942.0429.52
MLPMonoSDF [46]26.4034.4524.13
Occ_SDF_Hybrid†[17]26.9835.5024.72
Ours27.7736.0625.47
+ +Table 5: Ablation study of our method on ScanNet. + +
MethodAcc ↓Comp ↓Prec ↑Recall ↑F-score ↑
aBaseline0.0350.0480.7990.6810.733
bRS. Den.0.0400.0440.7720.7200.745
cRS. Den. + Con. L.0.0390.0410.7940.7600.776
dRS. Den. + Con. L. + Vis.(SDF)0.0390.0400.8090.7720.789
eRS. Den. + Con. L. + Vis.(SRDF)0.0420.0400.8040.7740.788
fRS. Den. + Con.L. + Vis.0.0400.0400.8090.7790.794
+ +![](images/47c852d8d7fb25b9d71ec8745fe40418576582749129703ee251432cb1116e34.jpg) +(a) Ground Truth + +![](images/a2595ef98b720e75b2bd727038e66e23d5e0e600ad803fff3f8b25a94331f7c8.jpg) +(b) RGB from MonoSDF + +![](images/b7db0289271cdaa5b2bbb3423091a8fd4500f567bbf703f076b146de3b417aca.jpg) +For the red sampled point +(c) Weight from MonoSDF + +![](images/f336e9428c36da7a0b2299a287e07165dc751b92244ed3dace9a070b7100cc65.jpg) +For the yellow sampled point + +![](images/bd37c51d43d8ca27ac0ca812cd12d6a7cbf3ed7d260c72f21be371fb2fbd4f39.jpg) +(d) RGB from ours + +![](images/b3c184dcbd608d223a7a827361fde777cec5fa6821ed34ad5ccd6e2e3fa890f3.jpg) +For the red sampled point +(e) Weight from ours +Fig.5: Comparison of the yielded weight and image. The weight and image generated by our approach align more closely with the actual observations. + +![](images/e09872b627a964405e7318b2d99e091a4aca5b25d2ffb51c68034bdca832ccd7.jpg) +For the yellow sampled point + +from SRDF. On the three datasets, our method outperforms the baseline and Occ_SDF_Hybrid in rendering training views. Overall, our method not only enhances surface reconstruction but also generates more accurate views. The visualization can be found in the supplementary material. + +# 4.3 Ablation Study + +Tab. 5 presents an ablation study to evaluate the effectiveness of our design on ScanNet with MLP representation. The ablation experiments consist of: (a) Baseline: MonoSDF. (b) RS. Den.: this structure models the volume density using the ray-specific SRDF, without the SRDF-SDF consistency loss and self-supervised visibility task. (c) RS. Den. + Con. L.: Building upon the structure in (b), this setup adopts the proposed SRDF-SDF consistency loss. (d) RS. Den. + Con. L. + Vis.(SDF): based on (c), this configuration adds the self-supervised visibility task, in which only SDF is used to compute the visibility label. (e) RS. Den. + Con. L. + Vis.(SRDF): Different from structure (d), this method relies solely on SRDF to compute the visibility label. (f) RS. Den. + Con. L. + Vis.: + +this structure represents the comprehensive model of our method, wherein the visibility labels integrate priors from both SRDF and SDF predictions. + +The comparison between configurations (a) and (b) highlights that our design using SRDF to model the density, without any additional constraints, results in a better performance compared to MonoSDF. The SRDF-SDF consistency loss in (c) surpasses the structure in (b) by $3.1\%$ in F-score. This shows the significance of sign consistency. Compared to (c), the self-supervised visibility task in (d)-(f) enhances the reconstruction performance. In particular, structure (f) obtains a better F-score than structures (d) and (e). These explain the effectiveness of our self-supervised visibility task and the computation of visibility labels integrating both SRDF and SDF priors. The visualization of the ablation study is given in the supplementary material. + +Furthermore, Fig. 5 compares the weight distribution for the red and yellow sampled points generated by MonoSDF and our method. (1) The analysis of the yellow sampled point demonstrates the benefits of our method for reconstruction: MonoSDF generates high weights and negative SDF values when sampled rays are near the white wall, resulting in false surfaces that significantly deviate from the ground truth. In contrast, our method yields more accurate surfaces and thus enhances both completeness and recall. (2) The analysis of the red sampled point explains the reason why our method can generate accurate views: MonoSDF produces weights with two local maxima, causing an inaccurate RGB value. This occurs because the SDF along the camera ray fluctuates, resulting in varying density and weight. In comparison, our method produces uni-modal weights that peak only around the surface, resulting in more accurate views. The quantitative PSNR in Fig. 5b and Fig. 5d confirms that our rendered image is more accurate compared to MonoSDF. These findings align with our motivation to achieve a more consistent and accurate weight distribution, thereby reconstructing better surfaces and generating high-quality views. This underscores the effectiveness of using ray-specific SRDF to model volume density in our approach. + +# 5 Conclusion + +This work proposes a novel method for neural indoor scene reconstruction, modeling density function with ray-specific SRDF. Firstly, this study analyzed that using the SDF to parameterize the volume density may introduce noises in multi-object indoor scenes, and the SRDF exhibits stronger relationships with the actual 2D observations. Motivated by this, this work employed SRDF to model the volume density, while SDF mainly focuses on the representation of the 3D surface. Secondly, a SRDF-SDF consistency loss has been introduced to align the sign between SRDF and SDF. Thirdly, a self-supervised visibility task has been designed to distinguish whether the sampled 3D point is physically visible or occluded. It combined the prior from SDF and SRDF as visibility labels to improve geometry reconstruction without any other annotations. Experimental results on real-world and synthetic datasets showed that our method excels not only in reconstructing better surfaces but also in generating more accurate views. + +# References + +1. Bozic, A., Palafox, P., Thies, J., Dai, A., Nießner, M.: Transformerfusion: Monocular rgb scene reconstruction using transformers. Advances in Neural Information Processing Systems 34, 1403-1414 (2021) +2. Collins, R.T.: A space-sweep approach to true multi-image matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 358-363. IEEE (1996) +3. Coughlan, J.M., Yuille, A.L.: Manhattan world: Compass direction from a single image by bayesian inference. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. vol. 2, pp. 941-947. IEEE (1999) +4. Curless, B., Levoy, M.: A volumetric method for building complex models from range images. In: Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques. pp. 303-312 (1996) +5. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5828-5839 (2017) +6. Duzceker, A., Galliani, S., Vogel, C., Speciale, P., Dusmanu, M., Pollefeys, M.: Deepvideomvs: Multi-view stereo on video with recurrent spatio-temporal fusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15324-15333 (2021) +7. Feng, Z., Yang, L., Guo, P., Li, B.: Cvrecon: Rethinking 3d geometric feature learning for neural reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17750-17760 (2023) +8. Fu, Q., Xu, Q., Ong, Y.S., Tao, W.: Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. Advances in Neural Information Processing Systems 35, 3403-3416 (2022) +9. Gallup, D., Frahm, J.M., Mordohai, P., Yang, Q., Pollefeys, M.: Real-time plane-sweeping stereo with multiple sweeping directions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1-8. IEEE (2007) +10. Guo, H., Peng, S., Lin, H., Wang, Q., Zhang, G., Bao, H., Zhou, X.: Neural 3d scene reconstruction with the manhattan-world assumption. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5511-5520 (2022) +11. Hu, S., Zhou, K., Li, K., Yu, L., Hong, L., Hu, T., Li, Z., Lee, G.H., Liu, Z.: Consistentnerf: Enhancing neural radiance fields with 3d consistency for sparse view synthesis. arXiv preprint arXiv:2305.11031 (2023) +12. Ju, J., Tseng, C.W., Bailo, O., Dikov, G., Ghafoorian, M.: Dg-recon: Depth-guided neural 3d scene reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18184-18194 (2023) +13. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics 36(4), 1-13 (2017) +14. Liang, Z., Huang, Z., Ding, C., Jia, K.: Helixsurf: A robust and efficient neural implicit surface learning of indoor scenes with iterative intertwined regularization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13165-13174 (2023) +15. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3d surface construction algorithm. In: Seminal graphics: pioneering efforts that shaped the field, pp. 347-353 (1998) + +16. Luo, K., Guan, T., Ju, L., Huang, H., Luo, Y.: P-mvsnet: Learning patch-wise matching confidence aggregation for multi-view stereo. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10452-10461 (2019) +17. Lyu, X., Dai, P., Li, Z., Yan, D., Lin, Y., Peng, Y., Qi, X.: Learning a room with the occ-sdf hybrid: Signed distance function mingled with occupancy aids scene representation. arXiv preprint arXiv:2303.09152 (2023) +18. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4460-4470 (2019) +19. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) +20. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics 41(4), 1-15 (2022) +21. Murez, Z., Van As, T., Bartolozzi, J., Sinha, A., Badrinarayanan, V., Rabinovich, A.: Atlas: End-to-end 3d scene reconstruction from posed images. In: Proceedings of the European Conference on Computer Vision. pp. 414-431. Springer (2020) +22. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5480-5490 (2022) +23. Oechsle, M., Peng, S., Geiger, A.: Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5589-5599 (2021) +24. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 165-174 (2019) +25. Peng, R., Gu, X., Tang, L., Shen, S., Yu, F., Wang, R.: Gens: Generalizable neural surface reconstruction from multi-view images. In: Advances in Neural Information Processing Systems (2023) +26. Ren, Y., Zhang, T., Pollefeys, M., Susstrunk, S., Wang, F.: Volrecon: Volume rendering of signed ray distance functions for generalizable multi-view reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16685-16695 (2023) +27. Sayed, M., Gibson, J., Watson, J., Prisacariu, V., Firman, M., Godard, C.: Simplex: 3d reconstruction without 3d convolutions. In: Proceedings of the European Conference on Computer Vision. pp. 1-19. Springer (2022) +28. Schonberger, J.L., Zheng, E., Frahm, J.M., Pollefeys, M.: Pixelwise view selection for unstructured multi-view stereo. In: Proceedings of the European Conference on Computer Vision. pp. 501-518. Springer (2016) +29. Sitzmann, V., Martel, J., Bergman, A., Lindell, D., Wetzstein, G.: Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems 33, 7462-7473 (2020) +30. Somraj, N., Soundararajan, R.: Vip-nerf: Visibility prior for sparse input neural radiance fields. arXiv preprint arXiv:2305.00041 (2023) +31. Stier, N., Rich, A., Sen, P., Höllerer, T.: Vortex: Volumetric 3d reconstruction with transformers for voxelwise view selection and fusion. In: 2021 International Conference on 3D Vision. pp. 320-330. IEEE (2021) + +32. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019) +33. Sun, J., Xie, Y., Chen, L., Zhou, X., Bao, H.: Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15598-15607 (2021) +34. Tan, M., Chen, B., Pang, R., Vasudevan, V., Sandler, M., Howard, A., Le, Q.V.: Mnasnet: Platform-aware neural architecture search for mobile. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2820-2828 (2019) +35. Tang, H., Liu, Z., Zhao, S., Lin, Y., Lin, J., Wang, H., Han, S.: Searching efficient 3d architectures with sparse point-voxel convolution. In: Proceedings of the European Conference on Computer Vision. pp. 685–702. Springer (2020) +36. Wang, J., Wang, P., Long, X., Theobalt, C., Komura, T., Liu, L., Wang, W.: Neuris: Neural reconstruction of indoor scenes using normal priors. In: Proceedings of the European Conference on Computer Vision. pp. 139-155. Springer (2022) +37. Wang, K., Shen, S.: Mvdepthnet: Real-time multiview depth estimation neural network. In: 2018 International conference on 3d vision. pp. 248-257. IEEE (2018) +38. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689 (2021) +39. Wang, Y., Li, Z., Jiang, Y., Zhou, K., Cao, T., Fu, Y., Xiao, C.: Neuralroom: Geometry-constrained neural implicit surfaces for indoor scene reconstruction. ACM Transactions on Graphics 41(6), 1-15 (2022) +40. Yao, Y., Luo, Z., Li, S., Fang, T., Quan, L.: Mvsnet: Depth inference for unstructured multi-view stereo. In: Proceedings of the European Conference on Computer Vision. pp. 767-783 (2018) +41. Yao, Y., Luo, Z., Li, S., Shen, T., Fang, T., Quan, L.: Recurrent mvsnet for high-resolution multi-view stereo depth inference. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5525-5534 (2019) +42. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34, 4805-4815 (2021) +43. Ye, B., Liu, S., Li, X., Yang, M.H.: Self-supervised super-plane for neural 3d reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21415-21424 (2023) +44. Yin, R., Karaoglu, S., Gevers, T.: Geometry-guided feature learning and fusion for indoor scene reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3652-3661 (2023) +45. Yu, Z., Gao, S.: Fast-mvsnet: Sparse-to-dense multi-view stereo with learned propagation and gauss-newton refinement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1949-1958 (2020) +46. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in Neural Information Processing Systems 35, 25018-25032 (2022) +47. Zhu, Z., Peng, S., Larsson, V., Cui, Z., Oswald, M.R., Geiger, A., Pollefeys, M.: Nicer-slam: Neural implicit scene encoding for rgb slam. arXiv preprint arXiv:2302.03594 (2023) +48. Zins, P., Xu, Y., Boyer, E., Wuhrer, S., Tung, T.: Multi-view reconstruction using signed ray distance functions (srdf). In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16696-16706 (2023) \ No newline at end of file diff --git a/2024/Ray-Distance Volume Rendering for Neural Scene Reconstruction/images.zip b/2024/Ray-Distance Volume Rendering for Neural Scene Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..14035391a08055b08c45b90dd1aa9ec8151e7e4c --- /dev/null +++ b/2024/Ray-Distance Volume Rendering for Neural Scene Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:86e9bf7dadaa38f5a85de707b848999070a2c614c6348b389647ea9651cc29fe +size 497603 diff --git a/2024/Ray-Distance Volume Rendering for Neural Scene Reconstruction/layout.json b/2024/Ray-Distance Volume Rendering for Neural Scene Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..09785c83d40de2ef68772d196523d237bf2f2ba5 --- /dev/null +++ b/2024/Ray-Distance Volume Rendering for Neural Scene Reconstruction/layout.json @@ -0,0 +1,10256 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 180, + 111, + 433, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 111, + 433, + 146 + ], + "spans": [ + { + "bbox": [ + 180, + 111, + 433, + 146 + ], + "type": "text", + "content": "Ray-Distance Volume Rendering for Neural Scene Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 143, + 167, + 471, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 167, + 471, + 180 + ], + "spans": [ + { + "bbox": [ + 143, + 167, + 471, + 180 + ], + "type": "text", + "content": "Ruihong Yin" + }, + { + "bbox": [ + 143, + 167, + 471, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 143, + 167, + 471, + 180 + ], + "type": "text", + "content": ", Yunlu Chen" + }, + { + "bbox": [ + 143, + 167, + 471, + 180 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 143, + 167, + 471, + 180 + ], + "type": "text", + "content": ", Sezer Karaoglu" + }, + { + "bbox": [ + 143, + 167, + 471, + 180 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 143, + 167, + 471, + 180 + ], + "type": "text", + "content": ", and Theo Gevers" + }, + { + "bbox": [ + 143, + 167, + 471, + 180 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 179, + 189, + 433, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 189, + 433, + 200 + ], + "spans": [ + { + "bbox": [ + 179, + 189, + 433, + 200 + ], + "type": "text", + "content": "1 University of Amsterdam, Amsterdam, The Netherlands" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 200, + 431, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 200, + 431, + 212 + ], + "spans": [ + { + "bbox": [ + 179, + 200, + 431, + 212 + ], + "type": "text", + "content": "2 Carnegie Mellon University, Pittsburgh, The United States" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 210, + 212, + 402, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 212, + 402, + 222 + ], + "spans": [ + { + "bbox": [ + 210, + 212, + 402, + 222 + ], + "type": "text", + "content": "3 3DUniversum, Amsterdam, The Netherlands" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 179, + 223, + 433, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 223, + 433, + 234 + ], + "spans": [ + { + "bbox": [ + 179, + 223, + 433, + 234 + ], + "type": "text", + "content": "{r.yin, Th.Gevers}@uva.nl, s.karaoglu@3duniversum.com," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 251, + 235, + 361, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 235, + 361, + 244 + ], + "spans": [ + { + "bbox": [ + 251, + 235, + 361, + 244 + ], + "type": "text", + "content": "yunluche@andrew.cmu.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 274, + 452, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 274, + 452, + 515 + ], + "spans": [ + { + "bbox": [ + 160, + 274, + 452, + 515 + ], + "type": "text", + "content": "Abstract. Existing methods in neural scene reconstruction utilize the Signed Distance Function (SDF) to model the density function. However, in indoor scenes, the density computed from the SDF for a sampled point may not consistently reflect its real importance in volume rendering, often due to the influence of neighboring objects. To tackle this issue, our work proposes a novel approach for indoor scene reconstruction, which instead parameterizes the density function with the Signed Ray Distance Function (SRDF). Firstly, the SRDF is predicted by the network and transformed to a ray-conditioned density function for volume rendering. We argue that the ray-specific SRDF only considers the surface along the camera ray, from which the derived density function is more consistent to the real occupancy than that from the SDF. Secondly, although SRDF and SDF represent different aspects of scene geometries, their values should share the same sign indicating the underlying spatial occupancy. Therefore, this work introduces a SRDF-SDF consistency loss to constrain the signs of the SRDF and SDF outputs. Thirdly, this work proposes a self-supervised visibility task, introducing the physical visibility geometry to the reconstruction task. The visibility task combines prior from predicted SRDF and SDF as pseudo labels, and contributes to generating more accurate 3D geometry. Our method implemented with different representations has been validated on indoor datasets, achieving improved performance in both reconstruction and view synthesis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 160, + 526, + 452, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 526, + 452, + 548 + ], + "spans": [ + { + "bbox": [ + 160, + 526, + 452, + 548 + ], + "type": "text", + "content": "Keywords: Indoor scene reconstruction " + }, + { + "bbox": [ + 160, + 526, + 452, + 548 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 526, + 452, + 548 + ], + "type": "text", + "content": " neural radiance fields " + }, + { + "bbox": [ + 160, + 526, + 452, + 548 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 526, + 452, + 548 + ], + "type": "text", + "content": " signed ray distance function" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 569, + 229, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 569, + 229, + 582 + ], + "spans": [ + { + "bbox": [ + 133, + 569, + 229, + 582 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Indoor 3D scene reconstruction involves using multi-view RGB images as input to generate detailed 3D geometry as output. This task is crucial and has widespread applications in virtual/mixed augmentation, robotics, navigation, and so on. Contrary to a single object, indoor environments contain various elements, such as chairs, walls, tables, and cups, which present a challenge for the task of reconstruction. Traditional reconstruction approaches [1,21,27,33,40,44]" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": "use 2D and 3D convolutional neural networks (CNNs) [34, 35] to extract features. They can be classified as depth-based methods [6, 27, 37, 40] and volumetric methods [1, 21, 31, 33, 44]. Depth-based techniques estimate depth for every viewpoint, encountering difficulties like scale uncertainties and inconsistencies in depth across different views. The surfaces created through these techniques can be coarse and imprecise. Unlike depth-based approaches, volumetric techniques utilize 3D CNNs to directly generate 3D geometry, leading to smoother surfaces. Nevertheless, these methods use voxels for scene representation, causing GPU memory usage and computational expenses to grow cubically as voxel resolution increases. Such a limitation restricts the use of volumetric approaches for high-quality scene reconstruction. Moreover, volumetric techniques require 3D annotations for network training, significantly increasing the cost of annotations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 262, + 482, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 370 + ], + "type": "text", + "content": "Implicit representations [18,24,29] encode 3D models as functions of coordinates. Utilizing implicit representations and volume rendering, neural radiance fields (NeRF) [19] achieves remarkable performance in novel view synthesis, even without 3D supervision. Some methods [38,42] exploit the capabilities of NeRF in reconstruction tasks. These methods typically parameterize the volume density as a learnable transformation of the Signed Distance Function (SDF), leading to impressive reconstruction performance. Compared to classic volumetric methods, neural implicit scene representation is more memory-efficient, marking a new trend in reconstruction tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 373, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 555 + ], + "type": "text", + "content": "In multi-object indoor scenes, the SDF from sampled points considers surface points across the entire scene. However, for neural implicit scene reconstruction, along a camera ray, from the camera center to the intersected surface point, the SDF may exhibit fluctuations, leading to several local minimum values. Fig. 1 provides a toy example to illustrate this. In Fig. 1c, volume density yielded from the fluctuating SDF displays several local maximum values, contributing to high weights in Fig. 1d. However, the camera ray " + }, + { + "bbox": [ + 130, + 373, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\overrightarrow{OP}" + }, + { + "bbox": [ + 130, + 373, + 482, + 555 + ], + "type": "text", + "content": " targets the green rectangle, while the SDF of the near-surface point " + }, + { + "bbox": [ + 130, + 373, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 130, + 373, + 482, + 555 + ], + "type": "text", + "content": " refers to the surface of the blue cylinder. Besides, " + }, + { + "bbox": [ + 130, + 373, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 130, + 373, + 482, + 555 + ], + "type": "text", + "content": " is far from the intersected point " + }, + { + "bbox": [ + 130, + 373, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 130, + 373, + 482, + 555 + ], + "type": "text", + "content": ", but still corresponds to high density and high weight. Nevertheless, during volume rendering that maps the 3D output to 2D space, it is evident that along a ray, points closer to the intersected surface point may carry a higher weight as they are more relevant to actual observations. Consequently, in the multi-object scenes, using the SDF to model the density function along the camera ray may disrupt this relationship, introduce noise, and lead to inaccuracies in 2D results or 3D geometries." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "This paper proposes RS-Recon, a novel indoor scene reconstruction method with the ray-specific density function to address the aforementioned issue. Firstly, our method models the density function as a function of the Signed Ray Distance Function (SRDF). Unlike SDF, SRDF measures the shortest distance to the surface along a camera ray, eliminating the influence of surrounding surfaces that are not located on the given ray. Moreover, the density distribution generated from SRDF attains the local maximum only around the intersected surface, aligning more closely to the concept that points nearer to the surface have more importance. To achieve this, our network predicts both SRDF and SDF, where" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "R. Yin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "SRDF encompasses the density function, while SDF mainly focuses on describing the 3D surfaces. Secondly, despite SRDF and SDF being defined differently, both of them produce positive values outside the object and negative values within. Therefore, in this paper, a SRDF-SDF consistency loss is devised to ensure they share the same sign. Thirdly, this work introduces a self-supervised visibility task to enhance the prediction of 3D geometry. The visibility task predicts the probability of whether the sampled points are physically visible along the ray in 3D space. The pseudo visibility ground truth is formed by combining the prior of SRDF and SDF from the network itself, without the reliance on multi-view geometry or any additional annotations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 483, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 483, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 483, + 368 + ], + "type": "text", + "content": "Our contributions can be summarized as follows, (1) This work proposes a novel neural scene reconstruction method that leverages the ray-specific SRDF to model the volume density. Compared to SDF, SRDF better reflects the significance of the real observations. (2) A SRDF-SDF consistency loss is proposed to constrain the alignment between SRDF and SDF, ensuring the same sign inside and outside the object. (3) To enhance the accuracy of 3D geometry, an additional self-supervised visibility task is introduced to predict the visibility probability of sampled points. This task integrates prior information from both the SDF and SRDF predictions as labels. (4) The experiments conducted on both synthetic and real-world indoor datasets illustrate that our method enhances performance in both reconstruction and view synthesis." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 387, + 243, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 387, + 243, + 400 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 243, + 400 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 414, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 640 + ], + "type": "text", + "content": "Classic Indoor Scene Reconstruction. Traditional methods for indoor scene reconstruction fall into two classes, i.e. depth-based methods [6, 16, 27, 37, 40, 41, 45] and volumetric methods [1, 7, 12, 21, 31, 33, 44]. Depth-based methods predict pixel-level depth for individual frames, and subsequently depth fusion [4] is adopted to generate the 3D scene. To achieve favorable outcomes, these methods often construct a plane sweep cost volume [2, 9] at the feature or image level, leveraging information from multiple views to complement the current frame, e.g. MVSNet [40], DeepVideoMVS [6], and SimpleRecon [27]. Using the depth map as the intermediate representation, these methods encounter challenges such as depth inconsistency and scale ambiguities. In contrast, volumetric methods utilize 3D CNN to directly regress the 3D geometry, which can generate smooth surfaces and reconstruct unobserved regions. For instance, NeuralRecon [33] designs a learnable TSDF fusion module to integrate features from previous frames and predicts the TSDF for sparse volumes. TransformerFusion [1], based on volumetric representation, adopts Transformer to fuse multi-view features and select frames with higher attention. Yin et al. [44] propose to incorporate geometry priors at different levels for the volumetric methods. However, due to the use of 3D CNN in volumetric methods, computational costs escalate significantly with higher resolutions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Neural Implicit Scene Representation. With the success of implicit representation and NeRF [19], neural implicit scene representation has emerged as a" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 180, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 180, + 91, + 448, + 102 + ], + "type": "text", + "content": "Ray-Distance Volume Rendering for Neural Scene Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 207, + 163 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 207, + 163 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 207, + 163 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 207, + 163 + ], + "type": "image", + "image_path": "0d205758a58093843359625d17fc030815f1eca3d726db456a5906b5306dc3de.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 135, + 166, + 211, + 220 + ], + "blocks": [ + { + "bbox": [ + 135, + 166, + 211, + 220 + ], + "lines": [ + { + "bbox": [ + 135, + 166, + 211, + 220 + ], + "spans": [ + { + "bbox": [ + 135, + 166, + 211, + 220 + ], + "type": "image", + "image_path": "75bd0a2c2b372345578f26a2377ef97ca7ff20f0c644f70e6fd2273b7185103e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 224, + 198, + 232 + ], + "lines": [ + { + "bbox": [ + 144, + 224, + 198, + 232 + ], + "spans": [ + { + "bbox": [ + 144, + 224, + 198, + 232 + ], + "type": "text", + "content": "(a) Multi-object scene" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 211, + 114, + 301, + 169 + ], + "blocks": [ + { + "bbox": [ + 211, + 114, + 301, + 169 + ], + "lines": [ + { + "bbox": [ + 211, + 114, + 301, + 169 + ], + "spans": [ + { + "bbox": [ + 211, + 114, + 301, + 169 + ], + "type": "image", + "image_path": "68ad60a3fce54a38647b70c1fdb4cb37d0440a6c04e04063cc60fe2907447435.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 211, + 170, + 301, + 223 + ], + "blocks": [ + { + "bbox": [ + 211, + 170, + 301, + 223 + ], + "lines": [ + { + "bbox": [ + 211, + 170, + 301, + 223 + ], + "spans": [ + { + "bbox": [ + 211, + 170, + 301, + 223 + ], + "type": "image", + "image_path": "4d3607c5da3c7578b994bf6b47e117190d1ac2e2699f0aaf50ea6dd49d324810.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 224, + 280, + 232 + ], + "lines": [ + { + "bbox": [ + 247, + 224, + 280, + 232 + ], + "spans": [ + { + "bbox": [ + 247, + 224, + 280, + 232 + ], + "type": "text", + "content": "(b) SDF/SRDF" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 302, + 114, + 391, + 170 + ], + "blocks": [ + { + "bbox": [ + 302, + 114, + 391, + 170 + ], + "lines": [ + { + "bbox": [ + 302, + 114, + 391, + 170 + ], + "spans": [ + { + "bbox": [ + 302, + 114, + 391, + 170 + ], + "type": "image", + "image_path": "69282ed5d06f58da24998c4c8a30bbae90c51791cc0878d999b3df40241bdbd8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 303, + 170, + 391, + 223 + ], + "blocks": [ + { + "bbox": [ + 303, + 170, + 391, + 223 + ], + "lines": [ + { + "bbox": [ + 303, + 170, + 391, + 223 + ], + "spans": [ + { + "bbox": [ + 303, + 170, + 391, + 223 + ], + "type": "image", + "image_path": "09844e35fe2dded36b0d53c850e551c27a62bd2cb180ce8fb9425a33fce69394.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 224, + 367, + 232 + ], + "lines": [ + { + "bbox": [ + 339, + 224, + 367, + 232 + ], + "spans": [ + { + "bbox": [ + 339, + 224, + 367, + 232 + ], + "type": "text", + "content": "(c) Density" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 244, + 482, + 346 + ], + "lines": [ + { + "bbox": [ + 131, + 244, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 131, + 244, + 482, + 346 + ], + "type": "text", + "content": "Fig. 1: Analysis of the density function and weight distribution for a toy scene. (a) Frontal and overhead perspectives of a scene with multiple objects, where a ray originates from the camera center " + }, + { + "bbox": [ + 131, + 244, + 482, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{O}" + }, + { + "bbox": [ + 131, + 244, + 482, + 346 + ], + "type": "text", + "content": ", intersecting with the green rectangle at the point " + }, + { + "bbox": [ + 131, + 244, + 482, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 131, + 244, + 482, + 346 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 244, + 482, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 131, + 244, + 482, + 346 + ], + "type": "text", + "content": " is the closest point to the blue cylinder along the ray " + }, + { + "bbox": [ + 131, + 244, + 482, + 346 + ], + "type": "inline_equation", + "content": "\\overrightarrow{OP}" + }, + { + "bbox": [ + 131, + 244, + 482, + 346 + ], + "type": "text", + "content": ". (b) The SDF " + }, + { + "bbox": [ + 131, + 244, + 482, + 346 + ], + "type": "inline_equation", + "content": "d_{\\Omega}" + }, + { + "bbox": [ + 131, + 244, + 482, + 346 + ], + "type": "text", + "content": " /SRDF " + }, + { + "bbox": [ + 131, + 244, + 482, + 346 + ], + "type": "inline_equation", + "content": "\\tilde{d}_{\\Omega}" + }, + { + "bbox": [ + 131, + 244, + 482, + 346 + ], + "type": "text", + "content": " in Eqs. (2) and (4) along the ray " + }, + { + "bbox": [ + 131, + 244, + 482, + 346 + ], + "type": "inline_equation", + "content": "\\overrightarrow{OP}" + }, + { + "bbox": [ + 131, + 244, + 482, + 346 + ], + "type": "text", + "content": ". (c) The volume density " + }, + { + "bbox": [ + 131, + 244, + 482, + 346 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 131, + 244, + 482, + 346 + ], + "type": "text", + "content": " generated from the SDF/SRDF in (b) using Eqs. (3) and (5). (d) The weight distribution " + }, + { + "bbox": [ + 131, + 244, + 482, + 346 + ], + "type": "inline_equation", + "content": "T_{i}\\alpha_{i}" + }, + { + "bbox": [ + 131, + 244, + 482, + 346 + ], + "type": "text", + "content": " generated from the density in (c) via Eq. (1). Although " + }, + { + "bbox": [ + 131, + 244, + 482, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 131, + 244, + 482, + 346 + ], + "type": "text", + "content": " is distant from the surface boundary intersecting with the ray, the density function from the SDF in " + }, + { + "bbox": [ + 131, + 244, + 482, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 131, + 244, + 482, + 346 + ], + "type": "text", + "content": " generates a high weight in volume rendering, resulting in noisy rendering and reconstruction." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 391, + 114, + 480, + 169 + ], + "blocks": [ + { + "bbox": [ + 391, + 114, + 480, + 169 + ], + "lines": [ + { + "bbox": [ + 391, + 114, + 480, + 169 + ], + "spans": [ + { + "bbox": [ + 391, + 114, + 480, + 169 + ], + "type": "image", + "image_path": "31fa9cf154536cc37d162b436657800c9bd388cf03b3e85dce138bb1de71e37e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 392, + 170, + 480, + 223 + ], + "blocks": [ + { + "bbox": [ + 392, + 170, + 480, + 223 + ], + "lines": [ + { + "bbox": [ + 392, + 170, + 480, + 223 + ], + "spans": [ + { + "bbox": [ + 392, + 170, + 480, + 223 + ], + "type": "image", + "image_path": "8df39fee97f9f9281441f8fdffc9be72387a86e21ff05b87a28f63540f2fefb9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 224, + 455, + 232 + ], + "lines": [ + { + "bbox": [ + 429, + 224, + 455, + 232 + ], + "spans": [ + { + "bbox": [ + 429, + 224, + 455, + 232 + ], + "type": "text", + "content": "(d) Weight" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 375, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 482, + 604 + ], + "type": "text", + "content": "hot spot in the past few years. Generally, reconstruction approaches based on neural implicit scene representation first map point positions to a continuous representation of SDF or occupancy by multi-layer perceptrons (MLPs), then transfer SDF or occupancy to volume density, and finally render the 2D results from the 3D outputs. Despite requiring only 2D supervision during training, these methods can reconstruct high-quality surfaces and are more computationally efficient. For example, VolSDF [42] parameterizes the volume density as a Laplace distribution of SDF, while NeuS [38] models it as the logistic distribution. Some extensions [10,36,39,46,47] introduce extra priors to achieve superior outcomes. For instance, Manhattan-SDF [10] employs planar constraints based on Manhattan-world assumption [3]. NeuRIS [36] proposes to combine normal prior for the reconstruction of texture-less regions, while NeuralRoom [39] implements perturbation-residual restriction to reconstruct the flat region better. MonoSDF [46] exploits both depth and normal cues to improve the reconstruction quality. Although additional priors contribute to the accuracy and completeness of reconstructed surfaces, they employ the SDF to model the volume density, potentially resulting in noisy 2D results in the multi-object scenes. Contrarily, our method advocates for parameterizing the density function with the ray-specific SRDF, resulting in enhanced reconstruction performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Some NeRF-based methods also utilize multi-view geometry to introduce additional regularization. For instance, RegNeRF [22] regularizes patched-based depth maps from unobserved viewpoints. Geo-NeuS [8] and ConsistentNeRF [11] use multi-view photometric consistency. SRDF and observation consistencies are used by [48]. GenS [25] introduces multi-view and multi-scale feature consistent" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "R. Yin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 482, + 234 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 482, + 234 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 482, + 234 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 482, + 234 + ], + "type": "image", + "image_path": "e50b7d5fd484c8d5e91491dca755596776e5e837e5334511d467449d7665368d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 243, + 482, + 354 + ], + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 354 + ], + "type": "text", + "content": "Fig. 2: Our framework. A geometry MLP " + }, + { + "bbox": [ + 130, + 243, + 482, + 354 + ], + "type": "inline_equation", + "content": "f_{g}" + }, + { + "bbox": [ + 130, + 243, + 482, + 354 + ], + "type": "text", + "content": " is adopted to generate the SDF and geometry features " + }, + { + "bbox": [ + 130, + 243, + 482, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{g}" + }, + { + "bbox": [ + 130, + 243, + 482, + 354 + ], + "type": "text", + "content": ", utilizing the encoded position (and optionally grid features) as input. Then, " + }, + { + "bbox": [ + 130, + 243, + 482, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{g}" + }, + { + "bbox": [ + 130, + 243, + 482, + 354 + ], + "type": "text", + "content": ", along with viewing direction, passes through a color MLP " + }, + { + "bbox": [ + 130, + 243, + 482, + 354 + ], + "type": "inline_equation", + "content": "f_{c}" + }, + { + "bbox": [ + 130, + 243, + 482, + 354 + ], + "type": "text", + "content": " to predict the color for each point. Notably, (1) our approach advocates for modeling the density function with the ray-specific SRDF in addition to the SDF. For this purpose, a SRDF MLP " + }, + { + "bbox": [ + 130, + 243, + 482, + 354 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 130, + 243, + 482, + 354 + ], + "type": "text", + "content": " is introduced to generate the SRDF. (2) A SRDF-SDF consistency loss " + }, + { + "bbox": [ + 130, + 243, + 482, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con}" + }, + { + "bbox": [ + 130, + 243, + 482, + 354 + ], + "type": "text", + "content": " is devised to align the signs between the generated SRDF and SDF. (3) To enhance the geometry prediction, a self-supervised visibility task is proposed to integrate geometry priors in both SRDF and SDF predicted by the network and generates the pseudo visibility ground truth. The visibility probability is predicted by the SRDF MLP." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 377, + 482, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 482, + 462 + ], + "type": "text", + "content": "cies for neural surface reconstruction. VIP-NeRF [30] adopts a visibility prior, using the plane sweep volume among multiple views to generate visibility labels. However, they depend on the computation of multi-view geometry, some of which may involve processing multiple frames yielding increased computational demands. In contrast, this paper introduces a self-supervised visibility task that requires minimal computational increases and does not depend on multi-view geometry for ground truth generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 480, + 202, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 480, + 202, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 480, + 202, + 493 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 506, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 579 + ], + "type": "text", + "content": "In this section, we provide the background on NeRF-based reconstruction methods, which typically model the volume density as a transformation of SDF (refer to Sec. 3.1). To overcome the issue of false local maxima in the density from SDF, our approach covered in Sec. 3.2 - Sec. 3.5 proposes to model the density function with the ray-specific SRDF in neural scene reconstruction, which achieves more accurate reconstruction. Fig. 2 presents the framework of our approach." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 597, + 471, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 597, + 471, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 597, + 471, + 609 + ], + "type": "text", + "content": "3.1 Background: Volume Rendering Density as Transformed SDF" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Our goal is to recover the scene geometry given a set of projected 2D views of a 3D scene. To achieve this, NeRF [19] optimizes a continuous 5D function to represent the scene, which predicts the volume density " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sigma(\\mathbf{p})" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " and view-dependent color " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{c}(\\mathbf{p},\\mathbf{r})" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " for each sampled point " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in \\mathbb{R}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " and ray direction " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{r} \\in S^2" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ". With " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Omega \\subset \\mathbb{R}^3" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 180, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 180, + 91, + 448, + 102 + ], + "type": "text", + "content": "Ray-Distance Volume Rendering for Neural Scene Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "denoting the scene geometry filled with physical presence, the predicted density " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\sigma (\\mathbf{p})" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " approximates the ground truth scene density " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathbf{1}_{\\varOmega}(\\mathbf{p})" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathbf{1}_{\\varOmega}(\\mathbf{p}) = 1" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathbf{p}\\in \\varOmega" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " and 0 otherwise. After that, the classic volume rendering is employed to render the color " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathbf{C}(\\mathbf{r})" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " points " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{p}_i = \\mathbf{o} + z_i\\mathbf{r}, i\\in [1,N]\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " are sampled along the camera ray " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " indicating the depth from the camera center " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " to the sampled point. The color is accumulated along the ray:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 163, + 196, + 481, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 163, + 196, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 163, + 196, + 481, + 228 + ], + "type": "interline_equation", + "content": "\\mathbf {C} (\\mathbf {r}) = \\sum_ {i = 1} ^ {N} T _ {i} \\alpha_ {i} \\mathbf {c} \\left(\\mathbf {p} _ {i}, \\mathbf {r}\\right), T _ {i} = \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {i}\\right), \\alpha_ {i} = 1 - \\exp \\left(- \\sigma \\left(\\mathbf {p} _ {i}\\right) \\delta_ {i}\\right) \\tag {1}", + "image_path": "d58352d371946a2585989198ed835c5074fd7434308163fd8d1750f25b057806.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 235, + 333, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 235, + 333, + 247 + ], + "spans": [ + { + "bbox": [ + 131, + 235, + 333, + 247 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 235, + 333, + 247 + ], + "type": "inline_equation", + "content": "\\delta_{i}" + }, + { + "bbox": [ + 131, + 235, + 333, + 247 + ], + "type": "text", + "content": " is the interval of neighboring points." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 247, + 479, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 479, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 479, + 319 + ], + "type": "text", + "content": "Despite the great success in novel view synthesis, NeRF has difficulty reconstructing satisfactory actual surfaces from the generated volume density. To address this challenge, recent methods [42,46] propose to predict the SDF and then transform it to volume density, which aligns with the geometry bias that the surface points have a higher density than other non-surface points and corresponds to a higher weight during volume rendering." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 319, + 481, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 481, + 365 + ], + "type": "text", + "content": "For any sampled point " + }, + { + "bbox": [ + 130, + 319, + 481, + 365 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 319, + 481, + 365 + ], + "type": "text", + "content": ", the SDF " + }, + { + "bbox": [ + 130, + 319, + 481, + 365 + ], + "type": "inline_equation", + "content": "d_{\\Omega}(\\mathbf{p})" + }, + { + "bbox": [ + 130, + 319, + 481, + 365 + ], + "type": "text", + "content": " is with an absolute value representing the shortest distance from " + }, + { + "bbox": [ + 130, + 319, + 481, + 365 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 319, + 481, + 365 + ], + "type": "text", + "content": " to the surface " + }, + { + "bbox": [ + 130, + 319, + 481, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\partial \\Omega" + }, + { + "bbox": [ + 130, + 319, + 481, + 365 + ], + "type": "text", + "content": ", while the sign indicates whether the point is outside (positive) or inside (negative) the surface. Formally," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 212, + 366, + 480, + 385 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 366, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 212, + 366, + 480, + 385 + ], + "type": "interline_equation", + "content": "\\operatorname {S D F}: \\quad d _ {\\Omega} (\\mathbf {p}) = (- 1) ^ {\\mathbf {1} _ {\\Omega} (\\mathbf {p})} \\min _ {\\mathbf {p} ^ {*} \\in \\mathcal {M}} \\| \\mathbf {p} - \\mathbf {p} ^ {*} \\| _ {2}, \\tag {2}", + "image_path": "7726863659be93b876199b9d122808946a7007da55eb0cfa73c62d3caf15613d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 388, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 437 + ], + "type": "text", + "content": "where the indicator function " + }, + { + "bbox": [ + 130, + 388, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{1}_{\\varOmega}(\\mathbf{p}) = 1" + }, + { + "bbox": [ + 130, + 388, + 482, + 437 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 130, + 388, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{p}\\in \\varOmega" + }, + { + "bbox": [ + 130, + 388, + 482, + 437 + ], + "type": "text", + "content": " and 0 otherwise. This indicates that the SDF considers every surface around the point. To apply SDF to volume rendering, Yariv et al. [42] suggest to derive the volume density " + }, + { + "bbox": [ + 130, + 388, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\sigma^{\\mathrm{SDF}}" + }, + { + "bbox": [ + 130, + 388, + 482, + 437 + ], + "type": "text", + "content": " from the signed distance as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 166, + 443, + 481, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 443, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 166, + 443, + 481, + 482 + ], + "type": "interline_equation", + "content": "\\sigma^ {\\mathrm {S D F}} (\\mathbf {p}) = \\alpha \\Psi_ {\\beta} (- d _ {\\Omega} (\\mathbf {p})) \\text {,} \\Psi_ {\\beta} (s) = \\left\\{ \\begin{array}{l l} \\frac {1}{2} \\exp \\left(\\frac {s}{\\beta}\\right) & \\text {i f} s \\leq 0 \\\\ 1 - \\frac {1}{2} \\exp \\left(- \\frac {s}{\\beta}\\right) & \\text {i f} s > 0 \\end{array} \\right. \\tag {3}", + "image_path": "1f1953ebf90a538946bdff166634b96f30eb9389669f5d59baea2612d4be5993.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 487, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 525 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 130, + 487, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta" + }, + { + "bbox": [ + 130, + 487, + 482, + 525 + ], + "type": "text", + "content": " are learnable parameters. " + }, + { + "bbox": [ + 130, + 487, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\Psi_{\\beta}" + }, + { + "bbox": [ + 130, + 487, + 482, + 525 + ], + "type": "text", + "content": " is the Cumulative Distribution Function (CDF) of the Laplace distribution with zero mean and " + }, + { + "bbox": [ + 130, + 487, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 487, + 482, + 525 + ], + "type": "text", + "content": " scale. Then, " + }, + { + "bbox": [ + 130, + 487, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\sigma^{\\mathrm{SDF}}(\\mathbf{p})" + }, + { + "bbox": [ + 130, + 487, + 482, + 525 + ], + "type": "text", + "content": " is applied to Eq. (1) and as the density function " + }, + { + "bbox": [ + 130, + 487, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 487, + 482, + 525 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 539, + 450, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 539, + 450, + 552 + ], + "spans": [ + { + "bbox": [ + 131, + 539, + 450, + 552 + ], + "type": "text", + "content": "3.2 Accurate Scene Volume Density from Ray-Specific SRDF" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "content": "We illustrate the issues with volume rendering SDF by a toy example in Fig. 1. Indoor scenes typically include multiple objects. Along the camera ray " + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "content": ", the SDF may be influenced by surfaces not intersecting the ray, resulting in several ambiguous local maxima of false high volume density and weight. This implies that the 2D observations may have a strong correlation to these distant points, contradicting the fact that the points closer to the surface along the camera ray may carry more significance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "To this end, we propose volume rendering scene density function from SRDF, the distance field local to the ray direction. In contrast to the SDF in Eq. (2)," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "R. Yin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 155 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 155 + ], + "type": "text", + "content": "the SRDF " + }, + { + "bbox": [ + 130, + 116, + 482, + 155 + ], + "type": "inline_equation", + "content": "\\tilde{d_{\\Omega}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 155 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 116, + 482, + 155 + ], + "type": "inline_equation", + "content": "\\mathbf{p}, \\mathbf{r}" + }, + { + "bbox": [ + 130, + 116, + 482, + 155 + ], + "type": "text", + "content": ") [26,48] computes the shortest distance from the point to the surface along the sampled ray " + }, + { + "bbox": [ + 130, + 116, + 482, + 155 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 130, + 116, + 482, + 155 + ], + "type": "text", + "content": ", i.e. the SRDF is ray-dependent. Formally, the SRDF " + }, + { + "bbox": [ + 130, + 116, + 482, + 155 + ], + "type": "inline_equation", + "content": "\\tilde{d_{\\Omega}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 155 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 116, + 482, + 155 + ], + "type": "inline_equation", + "content": "\\mathbf{p}, \\mathbf{r}" + }, + { + "bbox": [ + 130, + 116, + 482, + 155 + ], + "type": "text", + "content": ") can be represented by" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 200, + 163, + 481, + 184 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 163, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 200, + 163, + 481, + 184 + ], + "type": "interline_equation", + "content": "\\text {S R D F :} \\quad \\tilde {d _ {\\Omega}} (\\mathbf {p}, \\mathbf {r}) = (- 1) ^ {\\mathbf {1} _ {\\Omega} (\\mathbf {p})} \\min _ {(\\mathbf {p} + \\rho \\mathbf {r}) \\in \\mathcal {M}; \\rho \\in \\mathbb {R}} | \\rho |, \\tag {4}", + "image_path": "63afe27eab3f99aa0fa88badd2cb73f54f96dc3d49918e8556877fce291a1b83.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 193, + 481, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 193, + 481, + 218 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 481, + 218 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 193, + 481, + 218 + ], + "type": "inline_equation", + "content": "\\mathbf{1}_{\\Omega}(\\mathbf{p}) = 1" + }, + { + "bbox": [ + 130, + 193, + 481, + 218 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 130, + 193, + 481, + 218 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in \\Omega" + }, + { + "bbox": [ + 130, + 193, + 481, + 218 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 193, + 481, + 218 + ], + "type": "inline_equation", + "content": "0" + }, + { + "bbox": [ + 130, + 193, + 481, + 218 + ], + "type": "text", + "content": " otherwise. In the following sections, " + }, + { + "bbox": [ + 130, + 193, + 481, + 218 + ], + "type": "inline_equation", + "content": "d_{\\Omega}" + }, + { + "bbox": [ + 130, + 193, + 481, + 218 + ], + "type": "text", + "content": " from Eq. (2) and " + }, + { + "bbox": [ + 130, + 193, + 481, + 218 + ], + "type": "inline_equation", + "content": "\\tilde{d_{\\Omega}}" + }, + { + "bbox": [ + 130, + 193, + 481, + 218 + ], + "type": "text", + "content": " are used to denote SDF and SRDF for brevity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 219, + 482, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 219, + 482, + 362 + ], + "spans": [ + { + "bbox": [ + 130, + 219, + 482, + 362 + ], + "type": "text", + "content": "In Fig. 1, it can be seen that the weight computed from SRDF in the volume rendering only reaches a local peak around the surface point, aligning more consistently with the nature of the 2D observations. Therefore, this paper proposes to employ the ray-specific SRDF to yield the volume density instead of the SDF. This work predicts both SDF and SRDF, where SDF is mainly used to model the surface. As depicted in Fig. 2, for a sampled point " + }, + { + "bbox": [ + 130, + 219, + 482, + 362 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 219, + 482, + 362 + ], + "type": "text", + "content": ", a geometry MLP " + }, + { + "bbox": [ + 130, + 219, + 482, + 362 + ], + "type": "inline_equation", + "content": "f_{g}" + }, + { + "bbox": [ + 130, + 219, + 482, + 362 + ], + "type": "text", + "content": " is applied on the input (including the encoded position or features) to predict SDF " + }, + { + "bbox": [ + 130, + 219, + 482, + 362 + ], + "type": "inline_equation", + "content": "d_{\\Omega}" + }, + { + "bbox": [ + 130, + 219, + 482, + 362 + ], + "type": "text", + "content": " and geometry feature " + }, + { + "bbox": [ + 130, + 219, + 482, + 362 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{g}" + }, + { + "bbox": [ + 130, + 219, + 482, + 362 + ], + "type": "text", + "content": ". By definition, SRDF is view-dependent. Hence, the viewing direction " + }, + { + "bbox": [ + 130, + 219, + 482, + 362 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 130, + 219, + 482, + 362 + ], + "type": "text", + "content": ", geometry feature " + }, + { + "bbox": [ + 130, + 219, + 482, + 362 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{g}" + }, + { + "bbox": [ + 130, + 219, + 482, + 362 + ], + "type": "text", + "content": ", and point position " + }, + { + "bbox": [ + 130, + 219, + 482, + 362 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 219, + 482, + 362 + ], + "type": "text", + "content": " are concatenated and then passed through a SRDF MLP " + }, + { + "bbox": [ + 130, + 219, + 482, + 362 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 130, + 219, + 482, + 362 + ], + "type": "text", + "content": " to yield the SRDF. On the other hand, the viewing direction " + }, + { + "bbox": [ + 130, + 219, + 482, + 362 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 130, + 219, + 482, + 362 + ], + "type": "text", + "content": ", geometry feature " + }, + { + "bbox": [ + 130, + 219, + 482, + 362 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{g}" + }, + { + "bbox": [ + 130, + 219, + 482, + 362 + ], + "type": "text", + "content": ", point position " + }, + { + "bbox": [ + 130, + 219, + 482, + 362 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 219, + 482, + 362 + ], + "type": "text", + "content": ", and 3D unit normal " + }, + { + "bbox": [ + 130, + 219, + 482, + 362 + ], + "type": "inline_equation", + "content": "\\mathbf{n} = \\nabla d_{\\Omega}" + }, + { + "bbox": [ + 130, + 219, + 482, + 362 + ], + "type": "text", + "content": " are used as input to a color MLP " + }, + { + "bbox": [ + 130, + 219, + 482, + 362 + ], + "type": "inline_equation", + "content": "f_{c}" + }, + { + "bbox": [ + 130, + 219, + 482, + 362 + ], + "type": "text", + "content": " to output color " + }, + { + "bbox": [ + 130, + 219, + 482, + 362 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 219, + 482, + 362 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 363, + 482, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 388 + ], + "type": "text", + "content": "For volume rendering, our method adopts the predicted SRDF " + }, + { + "bbox": [ + 130, + 363, + 482, + 388 + ], + "type": "inline_equation", + "content": "\\tilde{d_{\\Omega}}" + }, + { + "bbox": [ + 130, + 363, + 482, + 388 + ], + "type": "text", + "content": " to derive the volume density " + }, + { + "bbox": [ + 130, + 363, + 482, + 388 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 363, + 482, + 388 + ], + "type": "text", + "content": ". This is defined by:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 231, + 396, + 481, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 396, + 481, + 416 + ], + "spans": [ + { + "bbox": [ + 231, + 396, + 481, + 416 + ], + "type": "interline_equation", + "content": "\\sigma^ {\\mathrm {S R D F}} (\\mathbf {p}, \\mathbf {r}) = \\alpha \\Psi_ {\\beta} \\left(- \\tilde {d _ {\\Omega}} (\\mathbf {p}, \\mathbf {r})\\right). \\tag {5}", + "image_path": "d568032cccbc37017f7ba582721900220e08cd1bae6b69f878dfeafc01ecd7a9.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 424, + 482, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 484 + ], + "type": "text", + "content": "Then, " + }, + { + "bbox": [ + 130, + 424, + 482, + 484 + ], + "type": "inline_equation", + "content": "\\sigma^{\\mathrm{SRDF}}(\\mathbf{p},\\mathbf{r})" + }, + { + "bbox": [ + 130, + 424, + 482, + 484 + ], + "type": "text", + "content": " is applied to the volume rendering in Eq. (1) as a ray-conditioned density function to obtain the color value " + }, + { + "bbox": [ + 130, + 424, + 482, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{C}^{\\mathrm{SRDF}}(\\mathbf{r})" + }, + { + "bbox": [ + 130, + 424, + 482, + 484 + ], + "type": "text", + "content": ". Besides, during training, we also render color " + }, + { + "bbox": [ + 130, + 424, + 482, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{C}^{\\mathrm{SDF}}(\\mathbf{r})" + }, + { + "bbox": [ + 130, + 424, + 482, + 484 + ], + "type": "text", + "content": " with density functions from SDF to obtain gradient signals to optimize the prediction of SDF, in which the volume rendering processes use the same rays and points as those for " + }, + { + "bbox": [ + 130, + 424, + 482, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{C}^{\\mathrm{SRDF}}(\\mathbf{r})" + }, + { + "bbox": [ + 130, + 424, + 482, + 484 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 485, + 482, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 534 + ], + "type": "text", + "content": "For the scene geometry, after the network is trained with NeRF optimization targets, the geometric surface " + }, + { + "bbox": [ + 130, + 485, + 482, + 534 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\partial \\Omega" + }, + { + "bbox": [ + 130, + 485, + 482, + 534 + ], + "type": "text", + "content": " is extracted as watertight meshes by the marching cubes algorithm [15,24] from the SDF results " + }, + { + "bbox": [ + 130, + 485, + 482, + 534 + ], + "type": "inline_equation", + "content": "d_{\\Omega}" + }, + { + "bbox": [ + 130, + 485, + 482, + 534 + ], + "type": "text", + "content": ", which only depends on the query location " + }, + { + "bbox": [ + 130, + 485, + 482, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 485, + 482, + 534 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 550, + 308, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 550, + 308, + 562 + ], + "spans": [ + { + "bbox": [ + 131, + 550, + 308, + 562 + ], + "type": "text", + "content": "3.3 SRDF-SDF Consistency Loss" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 569, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 640 + ], + "type": "text", + "content": "Following the definitions in Eq. (4), despite SRDF and SDF being defined differently, they share the same sign, which indicates the spatial occupancy — positive outside the surface and negative inside the surface. However, in this paper, where SDF and SRDF are predicted from separate branches, sign consistency is not guaranteed. To deal with this issue, a SRDF-SDF consistency loss is proposed to enforce this constraint." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "Although the sign function effectively computes the output's sign, it cannot be differentiated during back-propagation. To a certain degree, the values derived" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 180, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 180, + 91, + 448, + 102 + ], + "type": "text", + "content": "Ray-Distance Volume Rendering for Neural Scene Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": "from the sigmoid function approximate those calculated by the sign function. Thus, in our SRDF-SDF consistency loss, the sigmoid function is first employed to generate the sign of SRDF and SDF as the approximated sign function " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "\\varsigma (\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": ", after which a " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": " loss is adopted to quantify the sign difference, as defined by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 212, + 174, + 481, + 203 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 174, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 212, + 174, + 481, + 203 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c o n} = \\frac {1}{N _ {r}} \\sum_ {\\mathbf {p}, \\mathbf {r}} M _ {c o n} | | \\varsigma (\\tilde {d} _ {\\Omega}) - \\varsigma (d _ {\\Omega}) | | _ {2} \\tag {6}", + "image_path": "d014cbae3f0b06b6568f599e946941c9680216e217f070cbe3cb4a7fe405abc2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 214, + 205, + 403, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 205, + 403, + 220 + ], + "spans": [ + { + "bbox": [ + 214, + 205, + 403, + 220 + ], + "type": "interline_equation", + "content": "\\varsigma (d) = \\mathsf {S i g m o i d} (k \\cdot d), M _ {c o n} = [ \\widetilde {d} _ {\\Omega} \\cdot d _ {\\Omega} < 0 ]", + "image_path": "20c41d9eac6963648bdac4a750c7bbc63ff1b3f7c23c1593d43ef1093e7d3e35.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 228, + 482, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 482, + 291 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 130, + 228, + 482, + 291 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 228, + 482, + 291 + ], + "type": "text", + "content": " is a hyperparameter, controlling the slope of the sigmoid function. The indicator function " + }, + { + "bbox": [ + 130, + 228, + 482, + 291 + ], + "type": "inline_equation", + "content": "[\\tilde{d}_{\\Omega} \\cdot d_{\\Omega} < 0]" + }, + { + "bbox": [ + 130, + 228, + 482, + 291 + ], + "type": "text", + "content": " is leveraged to identify whether the generated SDF and SRDF have opposite signs. Therefore, the consistency loss imposes penalties when there is a difference between the signs of SRDF and SDF. " + }, + { + "bbox": [ + 130, + 228, + 482, + 291 + ], + "type": "inline_equation", + "content": "N_r" + }, + { + "bbox": [ + 130, + 228, + 482, + 291 + ], + "type": "text", + "content": " represents the number of sampled points in a minibatch that satisfy " + }, + { + "bbox": [ + 130, + 228, + 482, + 291 + ], + "type": "inline_equation", + "content": "[\\tilde{d}_{\\Omega} \\cdot d_{\\Omega} < 0]" + }, + { + "bbox": [ + 130, + 228, + 482, + 291 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 291, + 482, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 291, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 291, + 482, + 315 + ], + "type": "text", + "content": "Taking the derivative of " + }, + { + "bbox": [ + 130, + 291, + 482, + 315 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con}" + }, + { + "bbox": [ + 130, + 291, + 482, + 315 + ], + "type": "text", + "content": " with respect to SDF " + }, + { + "bbox": [ + 130, + 291, + 482, + 315 + ], + "type": "inline_equation", + "content": "d_{\\Omega}" + }, + { + "bbox": [ + 130, + 291, + 482, + 315 + ], + "type": "text", + "content": " (as an example), is computed as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 206, + 324, + 481, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 324, + 481, + 350 + ], + "spans": [ + { + "bbox": [ + 206, + 324, + 481, + 350 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\mathcal {L} _ {c o n}}{\\partial d _ {\\Omega}} = - \\frac {2}{N _ {r}} M _ {c o n} | | \\varsigma (\\tilde {d} _ {\\Omega}) - \\varsigma (d _ {\\Omega}) | | _ {1} \\cdot \\frac {\\partial \\varsigma (d _ {\\Omega})}{\\partial d _ {\\Omega}} \\tag {7}", + "image_path": "8abb8277b0c817382cc9a5289cce7c8986f2f3e21a74a7aa621a95bada19a3be.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 357, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 456 + ], + "type": "text", + "content": "The derivative in Eq. (7) shows that our consistency loss has two advantages: (1) Unlike the sign function, which assigns the same penalty for outputs with different signs of SRDF and SDF, our consistency loss can adjust the penalty according to the extent of the sign discrepancy between SRDF and SDF, referring to " + }, + { + "bbox": [ + 130, + 357, + 482, + 456 + ], + "type": "inline_equation", + "content": "||\\varsigma (\\tilde{d}_{\\varOmega}) - \\varsigma (d_{\\varOmega})||_1" + }, + { + "bbox": [ + 130, + 357, + 482, + 456 + ], + "type": "text", + "content": ". (2) The derivative of the sigmoid function " + }, + { + "bbox": [ + 130, + 357, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\varsigma(d_{\\varOmega})}{\\partial d_{\\varOmega}}" + }, + { + "bbox": [ + 130, + 357, + 482, + 456 + ], + "type": "text", + "content": " peaks at zero and decreases from 0 to " + }, + { + "bbox": [ + 130, + 357, + 482, + 456 + ], + "type": "inline_equation", + "content": "+\\infty / - \\infty" + }, + { + "bbox": [ + 130, + 357, + 482, + 456 + ], + "type": "text", + "content": ". Consequently, predictions in proximity to zero (indicating the surface) may exhibit higher absolute gradient values, providing effective supervision for points near the surface." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 473, + 313, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 473, + 313, + 487 + ], + "spans": [ + { + "bbox": [ + 131, + 473, + 313, + 487 + ], + "type": "text", + "content": "3.4 Self-supervised Visibility Task" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 494, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 482, + 666 + ], + "type": "text", + "content": "Theoretically, with visibility attributes of 3D points, surface location becomes straightforward. Therefore, to enhance 3D geometry prediction, we propose a self-supervised visibility task in this paper. Along the trajectory of a camera ray, points sampled prior to reaching the first surface point are categorized as visible, whereas points sampled beyond the first surface point are classified as occluded. To find the first surface point along the ray, using SRDF as an example, the multiplication of SRDF between adjacent sampled points is computed. When " + }, + { + "bbox": [ + 130, + 494, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tilde{d}_{\\Omega}^{i}\\cdot \\tilde{d}_{\\Omega}^{i + 1}\\leq 0" + }, + { + "bbox": [ + 130, + 494, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 494, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tilde{d}_{\\Omega}^{m}\\cdot \\tilde{d}_{\\Omega}^{m + 1} > 0,m\\in [1,i - 1]" + }, + { + "bbox": [ + 130, + 494, + 482, + 666 + ], + "type": "text", + "content": ", it indicates that the first surface is within the interval " + }, + { + "bbox": [ + 130, + 494, + 482, + 666 + ], + "type": "inline_equation", + "content": "[\\mathbf{o} + z_i\\mathbf{r},\\mathbf{o} + z_{i + 1}\\mathbf{r}]" + }, + { + "bbox": [ + 130, + 494, + 482, + 666 + ], + "type": "text", + "content": ". Consequently, the points between " + }, + { + "bbox": [ + 130, + 494, + 482, + 666 + ], + "type": "inline_equation", + "content": "[\\mathbf{p}_1,\\mathbf{p}_i]" + }, + { + "bbox": [ + 130, + 494, + 482, + 666 + ], + "type": "text", + "content": " are considered physically visible while points between " + }, + { + "bbox": [ + 130, + 494, + 482, + 666 + ], + "type": "inline_equation", + "content": "[\\mathbf{p}_{i + 1},\\mathbf{p}_N]" + }, + { + "bbox": [ + 130, + 494, + 482, + 666 + ], + "type": "text", + "content": " are regarded as occluded. However, the SRDF or SDF predicted by the network may be inaccurate at some points, introducing potential noise in the visibility task. To improve the accuracy of the visibility labeling, both SDF and SRDF are utilized in determining visibility. Points are labeled as visible or occluded only when there" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "R. Yin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 158, + 462, + 321 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 1: Evaluation of 3D reconstruction meshes on ScanNet. " + }, + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": " indicates our reproduced results. The best results are bold. " + }, + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": " denotes that a lower value is preferable while " + }, + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": " means higher value corresponds to better performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 158, + 462, + 321 + ], + "lines": [ + { + "bbox": [ + 151, + 158, + 462, + 321 + ], + "spans": [ + { + "bbox": [ + 151, + 158, + 462, + 321 + ], + "type": "table", + "html": "
MethodAcc ↓Comp ↓Prec ↑Recall ↑F-score ↑
COLMAP [28]0.0470.2350.7110.4410.537
UNISURF [23]0.5540.1640.2120.3620.267
NeuS [38]0.1790.2080.3130.2750.291
VolSDF [42]0.4140.1200.3210.3940.346
Manhattan-SDF [10]0.0720.0680.6210.5860.602
S3P [43]0.0550.0590.7090.6600.683
NeuRIS [36]0.0500.0490.7170.6690.692
MonoSDF_Grid [46]0.0720.0570.6600.6010.626
MonoSDF_MLP [46]0.0350.0480.7990.6810.733
HelixSurf [14]0.0380.0440.7860.7270.755
Occ_SDF_Hybrid† [17]0.0400.0410.7830.7480.765
Ours_Grid0.0740.0490.6700.7030.683
Ours_MLP0.0400.0400.8090.7790.794
", + "image_path": "430426dfbbdd3403ac8e18f9c30af6c2477d37d2fe5261d7b07cb31962f98775.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 342, + 482, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 390 + ], + "type": "text", + "content": "is agreement between SRDF and SDF on their classification. If the classifications from SRDF and SDF diverge, suggesting difficulties in accurately categorizing these points for the network, such points are omitted from the visibility task. The method for assigning visibility labels is defined as follows" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 216, + 398, + 481, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 398, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 216, + 398, + 481, + 430 + ], + "type": "interline_equation", + "content": "V _ {g t} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} (V ^ {\\mathrm {S R D F}} = 1) \\& (V ^ {\\mathrm {S D F}} = 1) \\\\ 0, & \\text {i f} (V ^ {\\mathrm {S R D F}} = 0) \\& (V ^ {\\mathrm {S D F}} = 0) \\end{array} \\right. \\tag {8}", + "image_path": "a043396185bb20fe1bca2c5d89db22e5cf197e9411c792bee9001f7b26f5e4a2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 437, + 482, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 497 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 437, + 482, + 497 + ], + "type": "inline_equation", + "content": "V^{\\mathrm{SRDF}} / V^{\\mathrm{SDF}}" + }, + { + "bbox": [ + 130, + 437, + 482, + 497 + ], + "type": "text", + "content": " are the visibility labels determined based on the prediction of SRDF/SDF. 1/0 means label for visible/occluded points. The formation of the visibility label can be inferred as self-supervised, thus eliminating the necessity for multi-view geometry and additional annotations. Additionally, it incorporates information from both SRDF and SDF." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 546 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 546 + ], + "type": "text", + "content": "From the analysis above, similar to SRDF, the visibility task also relies on the view direction. Therefore, a visibility probability " + }, + { + "bbox": [ + 130, + 498, + 482, + 546 + ], + "type": "inline_equation", + "content": "V_{pred}" + }, + { + "bbox": [ + 130, + 498, + 482, + 546 + ], + "type": "text", + "content": " is predicted from the SRDF branch. It is a binary classification task, i.e. visible or occluded, so the binary cross-entropy loss is employed as visibility loss " + }, + { + "bbox": [ + 130, + 498, + 482, + 546 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{vis}" + }, + { + "bbox": [ + 130, + 498, + 482, + 546 + ], + "type": "text", + "content": " during optimization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 546, + 482, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 606 + ], + "type": "text", + "content": "The visibility task can discern the point's visibility along the camera ray, thereby influencing the learning of SRDF. Moreover, the optimization process integrates knowledge of ray-related visibility across the entire space, further impacting the learning of SDF. Notably, unlike 2D supervision, the visibility task provides priors in 3D space, imparting a more direct influence on 3D geometry." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 622, + 228, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 622, + 228, + 634 + ], + "spans": [ + { + "bbox": [ + 131, + 622, + 228, + 634 + ], + "type": "text", + "content": "3.5 Optimization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Loss Functions. During training, the overall loss consists of RGB loss " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": ", normal loss " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_n" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": ", depth loss " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": ", Eikonal loss " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_e" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": ", smooth loss " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": ", SRDF-SDF" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 180, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 180, + 91, + 448, + 102 + ], + "type": "text", + "content": "Ray-Distance Volume Rendering for Neural Scene Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 332, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 332, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 332, + 129 + ], + "type": "text", + "content": "consistency loss " + }, + { + "bbox": [ + 132, + 116, + 332, + 129 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con}" + }, + { + "bbox": [ + 132, + 116, + 332, + 129 + ], + "type": "text", + "content": ", and visibility loss " + }, + { + "bbox": [ + 132, + 116, + 332, + 129 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{vis}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 174, + 138, + 481, + 150 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 138, + 481, + 150 + ], + "spans": [ + { + "bbox": [ + 174, + 138, + 481, + 150 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {c} + \\lambda_ {n} \\mathcal {L} _ {n} + \\lambda_ {d} \\mathcal {L} _ {d} + \\lambda_ {e} \\mathcal {L} _ {e} + \\lambda_ {s} \\mathcal {L} _ {s} + \\lambda_ {c o n} \\mathcal {L} _ {c o n} + \\lambda_ {v i s} \\mathcal {L} _ {v i s} \\tag {9}", + "image_path": "d37a95ad4608fe21219df1976496989bdd6665dd812b9375b694988691f02c0e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 158, + 480, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 158, + 480, + 207 + ], + "spans": [ + { + "bbox": [ + 130, + 158, + 480, + 207 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 130, + 158, + 480, + 207 + ], + "type": "inline_equation", + "content": "\\lambda_{(\\cdot)}" + }, + { + "bbox": [ + 130, + 158, + 480, + 207 + ], + "type": "text", + "content": " are the loss weight. Following MonoSDF [46], the 2.5D depth and normal maps are volume rendered using the density function " + }, + { + "bbox": [ + 130, + 158, + 480, + 207 + ], + "type": "inline_equation", + "content": "\\sigma^{\\mathrm{SRDF}}" + }, + { + "bbox": [ + 130, + 158, + 480, + 207 + ], + "type": "text", + "content": " to exploit geometric cues and support training. More details on the computation of loss functions are provided in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 207, + 481, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 207, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 207, + 481, + 304 + ], + "type": "text", + "content": "Implementation Details. Theoretically, our method can be applied to replace SDF-based volume rendering in indoor scene reconstruction. To measure our method, VolSDF-based reconstruction method MonoSDF is considered as our baseline in the following experiments. Furthermore, in the supplementary material, we also verify our method on NeuS-based reconstruction method NeuRIS. We used different spatial encoding techniques as input, including hash features [20] and Fourier features [19], referred to as Grid and MLP respectively in the following experiments. More details are given in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 321, + 230, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 230, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 230, + 335 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 346, + 261, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 346, + 261, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 346, + 261, + 358 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 365, + 480, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 480, + 413 + ], + "type": "text", + "content": "Datasets. Our proposed method is evaluated on three indoor datasets, comprising: (1) four scenes from the real-world dataset ScanNet [5], (2) eight scenes from the synthetic dataset Replica [32], and (3) four scenes from the advanced set of the real-world large-scale dataset Tanks and Temples [13]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 414, + 481, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 481, + 520 + ], + "type": "text", + "content": "Metrics. For ScanNet, the assessment of the results involves five reconstruction metrics: accuracy (acc), completeness (comp), precision (prec), recall, and F-score with a threshold of 5cm. For Replica, in alignment with MonoSDF, metrics including normal consistency (normal c.), chamfer L1 distance (Chamfer-L1), and F-score with a threshold of 5cm are utilized. For Tanks and Temples, as evaluated by the official server, results are reported based on the F-score with a 1cm threshold. Additionally, this study assesses the peak signal-to-noise ratio (PSNR) of the rendered 2D RGBs to showcase the effectiveness of our approach. Detailed computations for all metrics are given in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 521, + 481, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 521, + 481, + 570 + ], + "spans": [ + { + "bbox": [ + 131, + 521, + 481, + 570 + ], + "type": "text", + "content": "Baselines. Except for the baseline MonoSDF, our method is also compared to (1) the classic MVS method: COLMAP [28], (2) neural implicit reconstruction methods: UNISURF [23], NeuS [38], VolSDF [42], Manhattan-SDF [10], NeuRIS [36], S3P [43], HelixSurf [14], and Occ_SDF_Hybrid [17]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 586, + 256, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 256, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 256, + 597 + ], + "type": "text", + "content": "4.2 Evaluation Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "Reconstruction on ScanNet [5]. Tab. 1 compares our method against other state-of-the-art (SOTA) methods on ScanNet. When contrasted with the baseline MonoSDF, our method consistently shows better performance across all metrics, regardless of whether utilizing MLP or grid representation. For example, employing grid representation, our method exhibits a " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "5.7\\%" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": " higher F-score than" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "R. Yin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 473, + 243 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 473, + 243 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 473, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 473, + 243 + ], + "type": "image", + "image_path": "3aadb8698266b3f348af8bc1a20a549989e6fb0f3c1195c790e6de5f6b412399.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 253, + 480, + 276 + ], + "lines": [ + { + "bbox": [ + 131, + 253, + 480, + 276 + ], + "spans": [ + { + "bbox": [ + 131, + 253, + 480, + 276 + ], + "type": "text", + "content": "Fig. 3: Qualitative comparisons on ScanNet. It can be seen that our method can reconstruct more surfaces, especially in thin regions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 145, + 285, + 470, + 452 + ], + "blocks": [ + { + "bbox": [ + 145, + 285, + 470, + 452 + ], + "lines": [ + { + "bbox": [ + 145, + 285, + 470, + 452 + ], + "spans": [ + { + "bbox": [ + 145, + 285, + 470, + 452 + ], + "type": "image", + "image_path": "08b9dbfc9bc541e6db6461ca87301e96a223958047e861a36a3ec8920f0fbbfc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 463, + 482, + 487 + ], + "lines": [ + { + "bbox": [ + 130, + 463, + 482, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 482, + 487 + ], + "type": "text", + "content": "Fig.4: Qualitative comparisons on Replica. Compared to MonoSDF and Occ_SDF_Hybrid, our method can generate more accurate surfaces." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 510, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 570 + ], + "type": "text", + "content": "MonoSDF. Using the MLP representation, our method outperforms MonoSDF by " + }, + { + "bbox": [ + 130, + 510, + 482, + 570 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 570 + ], + "type": "text", + "content": " in recall, " + }, + { + "bbox": [ + 130, + 510, + 482, + 570 + ], + "type": "inline_equation", + "content": "9.8\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 570 + ], + "type": "text", + "content": " in precision, and " + }, + { + "bbox": [ + 130, + 510, + 482, + 570 + ], + "type": "inline_equation", + "content": "6.1\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 570 + ], + "type": "text", + "content": " in F-score. In addition, our method with the MLP representation surpasses other SOTA methods on almost all metrics, explaining the effectiveness of our method on reconstruction. Notably, our method achieves a " + }, + { + "bbox": [ + 130, + 510, + 482, + 570 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 570 + ], + "type": "text", + "content": " higher F-score compared to Occ_SDF_Hybrid." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": "Reconstruction on Replica [32]. The results on the synthetic dataset Replica are shown in Tab. 2. Using the grid representation, our method reports a notable improvement of " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "9.91\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " in chamfer distance and " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "2.88\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " in F-score. Similar increases are also illustrated in our method with the MLP representation across all metrics. Our approach also achieves on-par results with Occ_SDF_Hybrid. Reconstruction on Tanks and Temples [13]. Tab. 3 shows quantitative outcomes on the challenging large-scale dataset Tanks and Temples. The F-score for each scene - Auditorium, Ballroom, Courtroom, and Museum - along with" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 180, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 180, + 91, + 448, + 102 + ], + "type": "text", + "content": "Ray-Distance Volume Rendering for Neural Scene Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 150, + 135, + 463, + 212 + ], + "blocks": [ + { + "bbox": [ + 180, + 114, + 432, + 126 + ], + "lines": [ + { + "bbox": [ + 180, + 114, + 432, + 126 + ], + "spans": [ + { + "bbox": [ + 180, + 114, + 432, + 126 + ], + "type": "text", + "content": "Table 2: Evaluation of 3D reconstruction meshes on Replica." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 150, + 135, + 463, + 212 + ], + "lines": [ + { + "bbox": [ + 150, + 135, + 463, + 212 + ], + "spans": [ + { + "bbox": [ + 150, + 135, + 463, + 212 + ], + "type": "table", + "html": "
MethodNormal C. ↑Chamfer-L1 ↓F-score ↑
GridMonoSDF [46]90.933.2385.91
Ours91.662.9188.79
MLPMonoSDF [46]92.112.9486.18
Occ_SDF_Hybrid†[17]93.342.5891.54
Ours93.492.6091.72
", + "image_path": "3d0b87a6b51ec1dfaf1da2ab1e37bb25319f0772212318fb386dd769af87544c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 133, + 254, + 480, + 342 + ], + "blocks": [ + { + "bbox": [ + 132, + 221, + 481, + 245 + ], + "lines": [ + { + "bbox": [ + 132, + 221, + 481, + 245 + ], + "spans": [ + { + "bbox": [ + 132, + 221, + 481, + 245 + ], + "type": "text", + "content": "Table 3: Evaluation of 3D reconstruction meshes on the advanced set of Tanks and Temples. The F-score is reported in the table." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 254, + 480, + 342 + ], + "lines": [ + { + "bbox": [ + 133, + 254, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 133, + 254, + 480, + 342 + ], + "type": "table", + "html": "
MethodAuditoriumBallroomCourtroomMuseummean
GridMonoSDF [46]3.173.7013.755.686.58
Occ_SDF_Hybrid†[17]3.763.5814.047.277.16
Ours4.844.3114.187.607.73
MLPMonoSDF [46]3.092.4710.005.105.17
Occ_SDF_Hybrid†[17]3.213.3410.294.665.38
Ours4.244.6910.795.916.41
", + "image_path": "d02b0c992fc4062be2766c1c396d7990d17809bbfd520410d52dd8eaa4ccc8e5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 364, + 480, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 480, + 411 + ], + "type": "text", + "content": "their mean, is reported. Our method, with both the MLP and grid representations, outperforms Occ_SDF_Hybrid and MonoSDF. In particular, our F-score is " + }, + { + "bbox": [ + 130, + 364, + 480, + 411 + ], + "type": "inline_equation", + "content": "1.15\\%" + }, + { + "bbox": [ + 130, + 364, + 480, + 411 + ], + "type": "text", + "content": " higher (grid representation) and " + }, + { + "bbox": [ + 130, + 364, + 480, + 411 + ], + "type": "inline_equation", + "content": "1.24\\%" + }, + { + "bbox": [ + 130, + 364, + 480, + 411 + ], + "type": "text", + "content": " higher (MLP representation) than MonoSDF across all scenes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 413, + 480, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 480, + 520 + ], + "type": "text", + "content": "Visualization of Reconstruction Results. Fig. 3 and Fig. 4 report qualitative results on ScanNet and Replica respectively. In contrast to other methods, our approach can produce superior meshes and recall more regions, particularly for thin regions/small objects, such as chair legs and table legs. Although our method relies on the SRDF for improved density in volume rendering while the surface geometry is extracted by the SDF, both the proposed SRDF-SDF consistency loss and self-supervised visibility task guide the optimization of the SDF to be more consistent with the SRDF, resulting in more accurate surface boundary. Additional visualization is provided in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 481, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 545 + ], + "type": "text", + "content": "View Synthesis. Tab. 4 also provides a comparison of view synthesis performance on training views. Notably, 2D views are rendered with density derived" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 141, + 587, + 473, + 664 + ], + "blocks": [ + { + "bbox": [ + 154, + 567, + 457, + 578 + ], + "lines": [ + { + "bbox": [ + 154, + 567, + 457, + 578 + ], + "spans": [ + { + "bbox": [ + 154, + 567, + 457, + 578 + ], + "type": "text", + "content": "Table 4: Evaluation of the rendering quality. PSNR is given in this table." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 587, + 473, + 664 + ], + "lines": [ + { + "bbox": [ + 141, + 587, + 473, + 664 + ], + "spans": [ + { + "bbox": [ + 141, + 587, + 473, + 664 + ], + "type": "table", + "html": "
MethodScanNetReplicaTanks and Temples
GridMonoSDF [46]28.9640.2528.45
Ours30.0942.0429.52
MLPMonoSDF [46]26.4034.4524.13
Occ_SDF_Hybrid†[17]26.9835.5024.72
Ours27.7736.0625.47
", + "image_path": "e9025e3334bdf47766ab3bf1ae0e77aa5364483421541d8931b27d82bde1d9f4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "R. Yin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 136, + 475, + 223 + ], + "blocks": [ + { + "bbox": [ + 198, + 114, + 414, + 125 + ], + "lines": [ + { + "bbox": [ + 198, + 114, + 414, + 125 + ], + "spans": [ + { + "bbox": [ + 198, + 114, + 414, + 125 + ], + "type": "text", + "content": "Table 5: Ablation study of our method on ScanNet." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 136, + 475, + 223 + ], + "lines": [ + { + "bbox": [ + 140, + 136, + 475, + 223 + ], + "spans": [ + { + "bbox": [ + 140, + 136, + 475, + 223 + ], + "type": "table", + "html": "
MethodAcc ↓Comp ↓Prec ↑Recall ↑F-score ↑
aBaseline0.0350.0480.7990.6810.733
bRS. Den.0.0400.0440.7720.7200.745
cRS. Den. + Con. L.0.0390.0410.7940.7600.776
dRS. Den. + Con. L. + Vis.(SDF)0.0390.0400.8090.7720.789
eRS. Den. + Con. L. + Vis.(SRDF)0.0420.0400.8040.7740.788
fRS. Den. + Con.L. + Vis.0.0400.0400.8090.7790.794
", + "image_path": "c1abc32ebde24c4c854e55fa0bf0ffee651cd00078e0e47a5432f56de0aa0ec8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 143, + 234, + 216, + 307 + ], + "blocks": [ + { + "bbox": [ + 143, + 234, + 216, + 307 + ], + "lines": [ + { + "bbox": [ + 143, + 234, + 216, + 307 + ], + "spans": [ + { + "bbox": [ + 143, + 234, + 216, + 307 + ], + "type": "image", + "image_path": "47c852d8d7fb25b9d71ec8745fe40418576582749129703ee251432cb1116e34.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 308, + 199, + 313 + ], + "lines": [ + { + "bbox": [ + 157, + 308, + 199, + 313 + ], + "spans": [ + { + "bbox": [ + 157, + 308, + 199, + 313 + ], + "type": "text", + "content": "(a) Ground Truth" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 217, + 236, + 288, + 306 + ], + "blocks": [ + { + "bbox": [ + 217, + 236, + 288, + 306 + ], + "lines": [ + { + "bbox": [ + 217, + 236, + 288, + 306 + ], + "spans": [ + { + "bbox": [ + 217, + 236, + 288, + 306 + ], + "type": "image", + "image_path": "a2595ef98b720e75b2bd727038e66e23d5e0e600ad803fff3f8b25a94331f7c8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 226, + 308, + 283, + 313 + ], + "lines": [ + { + "bbox": [ + 226, + 308, + 283, + 313 + ], + "spans": [ + { + "bbox": [ + 226, + 308, + 283, + 313 + ], + "type": "text", + "content": "(b) RGB from MonoSDF" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 289, + 241, + 378, + 305 + ], + "blocks": [ + { + "bbox": [ + 306, + 236, + 364, + 241 + ], + "lines": [ + { + "bbox": [ + 306, + 236, + 364, + 241 + ], + "spans": [ + { + "bbox": [ + 306, + 236, + 364, + 241 + ], + "type": "text", + "content": "For the red sampled point" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 289, + 241, + 378, + 305 + ], + "lines": [ + { + "bbox": [ + 289, + 241, + 378, + 305 + ], + "spans": [ + { + "bbox": [ + 289, + 241, + 378, + 305 + ], + "type": "image", + "image_path": "b7db0289271cdaa5b2bbb3423091a8fd4500f567bbf703f076b146de3b417aca.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 308, + 428, + 313 + ], + "lines": [ + { + "bbox": [ + 365, + 308, + 428, + 313 + ], + "spans": [ + { + "bbox": [ + 365, + 308, + 428, + 313 + ], + "type": "text", + "content": "(c) Weight from MonoSDF" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 378, + 241, + 469, + 305 + ], + "blocks": [ + { + "bbox": [ + 398, + 236, + 463, + 241 + ], + "lines": [ + { + "bbox": [ + 398, + 236, + 463, + 241 + ], + "spans": [ + { + "bbox": [ + 398, + 236, + 463, + 241 + ], + "type": "text", + "content": "For the yellow sampled point" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 378, + 241, + 469, + 305 + ], + "lines": [ + { + "bbox": [ + 378, + 241, + 469, + 305 + ], + "spans": [ + { + "bbox": [ + 378, + 241, + 469, + 305 + ], + "type": "image", + "image_path": "f336e9428c36da7a0b2299a287e07165dc751b92244ed3dace9a070b7100cc65.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 217, + 316, + 289, + 388 + ], + "blocks": [ + { + "bbox": [ + 217, + 316, + 289, + 388 + ], + "lines": [ + { + "bbox": [ + 217, + 316, + 289, + 388 + ], + "spans": [ + { + "bbox": [ + 217, + 316, + 289, + 388 + ], + "type": "image", + "image_path": "bd37c51d43d8ca27ac0ca812cd12d6a7cbf3ed7d260c72f21be371fb2fbd4f39.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 236, + 390, + 280, + 396 + ], + "lines": [ + { + "bbox": [ + 236, + 390, + 280, + 396 + ], + "spans": [ + { + "bbox": [ + 236, + 390, + 280, + 396 + ], + "type": "text", + "content": "(d) RGB from ours" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 289, + 325, + 378, + 389 + ], + "blocks": [ + { + "bbox": [ + 306, + 319, + 364, + 325 + ], + "lines": [ + { + "bbox": [ + 306, + 319, + 364, + 325 + ], + "spans": [ + { + "bbox": [ + 306, + 319, + 364, + 325 + ], + "type": "text", + "content": "For the red sampled point" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 289, + 325, + 378, + 389 + ], + "lines": [ + { + "bbox": [ + 289, + 325, + 378, + 389 + ], + "spans": [ + { + "bbox": [ + 289, + 325, + 378, + 389 + ], + "type": "image", + "image_path": "b3c184dcbd608d223a7a827361fde777cec5fa6821ed34ad5ccd6e2e3fa890f3.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 390, + 416, + 396 + ], + "lines": [ + { + "bbox": [ + 365, + 390, + 416, + 396 + ], + "spans": [ + { + "bbox": [ + 365, + 390, + 416, + 396 + ], + "type": "text", + "content": "(e) Weight from ours" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 409, + 482, + 432 + ], + "lines": [ + { + "bbox": [ + 130, + 409, + 482, + 432 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 482, + 432 + ], + "type": "text", + "content": "Fig.5: Comparison of the yielded weight and image. The weight and image generated by our approach align more closely with the actual observations." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 378, + 325, + 469, + 389 + ], + "blocks": [ + { + "bbox": [ + 398, + 319, + 463, + 325 + ], + "lines": [ + { + "bbox": [ + 398, + 319, + 463, + 325 + ], + "spans": [ + { + "bbox": [ + 398, + 319, + 463, + 325 + ], + "type": "text", + "content": "For the yellow sampled point" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 378, + 325, + 469, + 389 + ], + "lines": [ + { + "bbox": [ + 378, + 325, + 469, + 389 + ], + "spans": [ + { + "bbox": [ + 378, + 325, + 469, + 389 + ], + "type": "image", + "image_path": "e09872b627a964405e7318b2d99e091a4aca5b25d2ffb51c68034bdca832ccd7.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 456, + 482, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 504 + ], + "type": "text", + "content": "from SRDF. On the three datasets, our method outperforms the baseline and Occ_SDF_Hybrid in rendering training views. Overall, our method not only enhances surface reconstruction but also generates more accurate views. The visualization can be found in the supplementary material." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 131, + 524, + 238, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 524, + 238, + 536 + ], + "spans": [ + { + "bbox": [ + 131, + 524, + 238, + 536 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "Tab. 5 presents an ablation study to evaluate the effectiveness of our design on ScanNet with MLP representation. The ablation experiments consist of: (a) Baseline: MonoSDF. (b) RS. Den.: this structure models the volume density using the ray-specific SRDF, without the SRDF-SDF consistency loss and self-supervised visibility task. (c) RS. Den. + Con. L.: Building upon the structure in (b), this setup adopts the proposed SRDF-SDF consistency loss. (d) RS. Den. + Con. L. + Vis.(SDF): based on (c), this configuration adds the self-supervised visibility task, in which only SDF is used to compute the visibility label. (e) RS. Den. + Con. L. + Vis.(SRDF): Different from structure (d), this method relies solely on SRDF to compute the visibility label. (f) RS. Den. + Con. L. + Vis.:" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 180, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 180, + 91, + 448, + 102 + ], + "type": "text", + "content": "Ray-Distance Volume Rendering for Neural Scene Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "this structure represents the comprehensive model of our method, wherein the visibility labels integrate priors from both SRDF and SDF predictions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 259 + ], + "type": "text", + "content": "The comparison between configurations (a) and (b) highlights that our design using SRDF to model the density, without any additional constraints, results in a better performance compared to MonoSDF. The SRDF-SDF consistency loss in (c) surpasses the structure in (b) by " + }, + { + "bbox": [ + 130, + 140, + 481, + 259 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 130, + 140, + 481, + 259 + ], + "type": "text", + "content": " in F-score. This shows the significance of sign consistency. Compared to (c), the self-supervised visibility task in (d)-(f) enhances the reconstruction performance. In particular, structure (f) obtains a better F-score than structures (d) and (e). These explain the effectiveness of our self-supervised visibility task and the computation of visibility labels integrating both SRDF and SDF priors. The visualization of the ablation study is given in the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 261, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 464 + ], + "type": "text", + "content": "Furthermore, Fig. 5 compares the weight distribution for the red and yellow sampled points generated by MonoSDF and our method. (1) The analysis of the yellow sampled point demonstrates the benefits of our method for reconstruction: MonoSDF generates high weights and negative SDF values when sampled rays are near the white wall, resulting in false surfaces that significantly deviate from the ground truth. In contrast, our method yields more accurate surfaces and thus enhances both completeness and recall. (2) The analysis of the red sampled point explains the reason why our method can generate accurate views: MonoSDF produces weights with two local maxima, causing an inaccurate RGB value. This occurs because the SDF along the camera ray fluctuates, resulting in varying density and weight. In comparison, our method produces uni-modal weights that peak only around the surface, resulting in more accurate views. The quantitative PSNR in Fig. 5b and Fig. 5d confirms that our rendered image is more accurate compared to MonoSDF. These findings align with our motivation to achieve a more consistent and accurate weight distribution, thereby reconstructing better surfaces and generating high-quality views. This underscores the effectiveness of using ray-specific SRDF to model volume density in our approach." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 483, + 220, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 220, + 496 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 220, + 496 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "This work proposes a novel method for neural indoor scene reconstruction, modeling density function with ray-specific SRDF. Firstly, this study analyzed that using the SDF to parameterize the volume density may introduce noises in multi-object indoor scenes, and the SRDF exhibits stronger relationships with the actual 2D observations. Motivated by this, this work employed SRDF to model the volume density, while SDF mainly focuses on the representation of the 3D surface. Secondly, a SRDF-SDF consistency loss has been introduced to align the sign between SRDF and SDF. Thirdly, a self-supervised visibility task has been designed to distinguish whether the sampled 3D point is physically visible or occluded. It combined the prior from SDF and SRDF as visibility labels to improve geometry reconstruction without any other annotations. Experimental results on real-world and synthetic datasets showed that our method excels not only in reconstructing better surfaces but also in generating more accurate views." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "R. Yin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 137, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "type": "text", + "content": "1. Bozic, A., Palafox, P., Thies, J., Dai, A., Nießner, M.: Transformerfusion: Monocular rgb scene reconstruction using transformers. Advances in Neural Information Processing Systems 34, 1403-1414 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 170, + 480, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 170, + 480, + 201 + ], + "spans": [ + { + "bbox": [ + 138, + 170, + 480, + 201 + ], + "type": "text", + "content": "2. Collins, R.T.: A space-sweep approach to true multi-image matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 358-363. IEEE (1996)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 202, + 480, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 202, + 480, + 234 + ], + "spans": [ + { + "bbox": [ + 138, + 202, + 480, + 234 + ], + "type": "text", + "content": "3. Coughlan, J.M., Yuille, A.L.: Manhattan world: Compass direction from a single image by bayesian inference. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. vol. 2, pp. 941-947. IEEE (1999)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 235, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 235, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 235, + 480, + 266 + ], + "type": "text", + "content": "4. Curless, B., Levoy, M.: A volumetric method for building complex models from range images. In: Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques. pp. 303-312 (1996)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 266, + 480, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 480, + 308 + ], + "type": "text", + "content": "5. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5828-5839 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 309, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 309, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 309, + 480, + 352 + ], + "type": "text", + "content": "6. Duzceker, A., Galliani, S., Vogel, C., Speciale, P., Dusmanu, M., Pollefeys, M.: Deepvideomvs: Multi-view stereo on video with recurrent spatio-temporal fusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15324-15333 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 353, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 353, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 353, + 480, + 385 + ], + "type": "text", + "content": "7. Feng, Z., Yang, L., Guo, P., Li, B.: Cvrecon: Rethinking 3d geometric feature learning for neural reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17750-17760 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 386, + 480, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 386, + 480, + 417 + ], + "spans": [ + { + "bbox": [ + 138, + 386, + 480, + 417 + ], + "type": "text", + "content": "8. Fu, Q., Xu, Q., Ong, Y.S., Tao, W.: Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. Advances in Neural Information Processing Systems 35, 3403-3416 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 418, + 480, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 480, + 459 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 480, + 459 + ], + "type": "text", + "content": "9. Gallup, D., Frahm, J.M., Mordohai, P., Yang, Q., Pollefeys, M.: Real-time plane-sweeping stereo with multiple sweeping directions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1-8. IEEE (2007)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 460, + 480, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 460, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 134, + 460, + 480, + 502 + ], + "type": "text", + "content": "10. Guo, H., Peng, S., Lin, H., Wang, Q., Zhang, G., Bao, H., Zhou, X.: Neural 3d scene reconstruction with the manhattan-world assumption. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5511-5520 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 503, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 503, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 134, + 503, + 480, + 536 + ], + "type": "text", + "content": "11. Hu, S., Zhou, K., Li, K., Yu, L., Hong, L., Hu, T., Li, Z., Lee, G.H., Liu, Z.: Consistentnerf: Enhancing neural radiance fields with 3d consistency for sparse view synthesis. arXiv preprint arXiv:2305.11031 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 536, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 536, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 134, + 536, + 480, + 568 + ], + "type": "text", + "content": "12. Ju, J., Tseng, C.W., Bailo, O., Dikov, G., Ghafoorian, M.: Dg-recon: Depth-guided neural 3d scene reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18184-18194 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 569, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 569, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 134, + 569, + 480, + 590 + ], + "type": "text", + "content": "13. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics 36(4), 1-13 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 590, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 590, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 590, + 480, + 632 + ], + "type": "text", + "content": "14. Liang, Z., Huang, Z., Ding, C., Jia, K.: Helixsurf: A robust and efficient neural implicit surface learning of indoor scenes with iterative intertwined regularization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13165-13174 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "text", + "content": "15. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3d surface construction algorithm. In: Seminal graphics: pioneering efforts that shaped the field, pp. 347-353 (1998)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 180, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 180, + 91, + 447, + 102 + ], + "type": "text", + "content": "Ray-Distance Volume Rendering for Neural Scene Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "16. Luo, K., Guan, T., Ju, L., Huang, H., Luo, Y.: P-mvsnet: Learning patch-wise matching confidence aggregation for multi-view stereo. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10452-10461 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "text", + "content": "17. Lyu, X., Dai, P., Li, Z., Yan, D., Lin, Y., Peng, Y., Qi, X.: Learning a room with the occ-sdf hybrid: Signed distance function mingled with occupancy aids scene representation. arXiv preprint arXiv:2303.09152 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 481, + 224 + ], + "type": "text", + "content": "18. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4460-4470 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 224, + 481, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 481, + 257 + ], + "type": "text", + "content": "19. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 257, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 481, + 289 + ], + "type": "text", + "content": "20. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics 41(4), 1-15 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 289, + 481, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 481, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 481, + 322 + ], + "type": "text", + "content": "21. Murez, Z., Van As, T., Bartolozzi, J., Sinha, A., Badrinarayanan, V., Rabinovich, A.: Atlas: End-to-end 3d scene reconstruction from posed images. In: Proceedings of the European Conference on Computer Vision. pp. 414-431. Springer (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 322, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 481, + 365 + ], + "type": "text", + "content": "22. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5480-5490 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 365, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 481, + 397 + ], + "type": "text", + "content": "23. Oechsle, M., Peng, S., Geiger, A.: Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5589-5599 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 397, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 481, + 439 + ], + "type": "text", + "content": "24. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 165-174 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "type": "text", + "content": "25. Peng, R., Gu, X., Tang, L., Shen, S., Yu, F., Wang, R.: Gens: Generalizable neural surface reconstruction from multi-view images. In: Advances in Neural Information Processing Systems (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 472, + 481, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 481, + 515 + ], + "type": "text", + "content": "26. Ren, Y., Zhang, T., Pollefeys, M., Susstrunk, S., Wang, F.: Volrecon: Volume rendering of signed ray distance functions for generalizable multi-view reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16685-16695 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 515, + 481, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 481, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 481, + 548 + ], + "type": "text", + "content": "27. Sayed, M., Gibson, J., Watson, J., Prisacariu, V., Firman, M., Godard, C.: Simplex: 3d reconstruction without 3d convolutions. In: Proceedings of the European Conference on Computer Vision. pp. 1-19. Springer (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 548, + 481, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 481, + 580 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 481, + 580 + ], + "type": "text", + "content": "28. Schonberger, J.L., Zheng, E., Frahm, J.M., Pollefeys, M.: Pixelwise view selection for unstructured multi-view stereo. In: Proceedings of the European Conference on Computer Vision. pp. 501-518. Springer (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 580, + 481, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 580, + 481, + 612 + ], + "spans": [ + { + "bbox": [ + 132, + 580, + 481, + 612 + ], + "type": "text", + "content": "29. Sitzmann, V., Martel, J., Bergman, A., Lindell, D., Wetzstein, G.: Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems 33, 7462-7473 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 612, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 612, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 612, + 481, + 632 + ], + "type": "text", + "content": "30. Somraj, N., Soundararajan, R.: Vip-nerf: Visibility prior for sparse input neural radiance fields. arXiv preprint arXiv:2305.00041 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "31. Stier, N., Rich, A., Sen, P., Höllerer, T.: Vortex: Volumetric 3d reconstruction with transformers for voxelwise view selection and fusion. In: 2021 International Conference on 3D Vision. pp. 320-330. IEEE (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "R. Yin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 658 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "32. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "text", + "content": "33. Sun, J., Xie, Y., Chen, L., Zhou, X., Bao, H.: Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15598-15607 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 225 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 225 + ], + "type": "text", + "content": "34. Tan, M., Chen, B., Pang, R., Vasudevan, V., Sandler, M., Howard, A., Le, Q.V.: Mnasnet: Platform-aware neural architecture search for mobile. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2820-2828 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 225, + 482, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 225, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 482, + 258 + ], + "type": "text", + "content": "35. Tang, H., Liu, Z., Zhao, S., Lin, Y., Lin, J., Wang, H., Han, S.: Searching efficient 3d architectures with sparse point-voxel convolution. In: Proceedings of the European Conference on Computer Vision. pp. 685–702. Springer (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 258, + 482, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 291 + ], + "type": "text", + "content": "36. Wang, J., Wang, P., Long, X., Theobalt, C., Komura, T., Liu, L., Wang, W.: Neuris: Neural reconstruction of indoor scenes using normal priors. In: Proceedings of the European Conference on Computer Vision. pp. 139-155. Springer (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 290, + 482, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 312 + ], + "type": "text", + "content": "37. Wang, K., Shen, S.: Mvdepthnet: Real-time multiview depth estimation neural network. In: 2018 International conference on 3d vision. pp. 248-257. IEEE (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 312, + 482, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 344 + ], + "type": "text", + "content": "38. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 344, + 482, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 482, + 376 + ], + "type": "text", + "content": "39. Wang, Y., Li, Z., Jiang, Y., Zhou, K., Cao, T., Fu, Y., Xiao, C.: Neuralroom: Geometry-constrained neural implicit surfaces for indoor scene reconstruction. ACM Transactions on Graphics 41(6), 1-15 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 376, + 482, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 409 + ], + "type": "text", + "content": "40. Yao, Y., Luo, Z., Li, S., Fang, T., Quan, L.: Mvsnet: Depth inference for unstructured multi-view stereo. In: Proceedings of the European Conference on Computer Vision. pp. 767-783 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 409, + 482, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 409, + 482, + 442 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 482, + 442 + ], + "type": "text", + "content": "41. Yao, Y., Luo, Z., Li, S., Shen, T., Fang, T., Quan, L.: Recurrent mvsnet for high-resolution multi-view stereo depth inference. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5525-5534 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 441, + 482, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 482, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 482, + 463 + ], + "type": "text", + "content": "42. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34, 4805-4815 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 462, + 482, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 495 + ], + "type": "text", + "content": "43. Ye, B., Liu, S., Li, X., Yang, M.H.: Self-supervised super-plane for neural 3d reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21415-21424 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 495, + 482, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 528 + ], + "type": "text", + "content": "44. Yin, R., Karaoglu, S., Gevers, T.: Geometry-guided feature learning and fusion for indoor scene reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3652-3661 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 528, + 482, + 560 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 482, + 560 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 482, + 560 + ], + "type": "text", + "content": "45. Yu, Z., Gao, S.: Fast-mvsnet: Sparse-to-dense multi-view stereo with learned propagation and gauss-newton refinement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1949-1958 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 560, + 482, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 560, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 560, + 482, + 592 + ], + "type": "text", + "content": "46. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in Neural Information Processing Systems 35, 25018-25032 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 592, + 482, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 592, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 592, + 482, + 624 + ], + "type": "text", + "content": "47. Zhu, Z., Peng, S., Larsson, V., Cui, Z., Oswald, M.R., Geiger, A., Pollefeys, M.: Nicer-slam: Neural implicit scene encoding for rgb slam. arXiv preprint arXiv:2302.03594 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 624, + 482, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 624, + 482, + 658 + ], + "spans": [ + { + "bbox": [ + 130, + 624, + 482, + 658 + ], + "type": "text", + "content": "48. Zins, P., Xu, Y., Boyer, E., Wuhrer, S., Tung, T.: Multi-view reconstruction using signed ray distance functions (srdf). In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16696-16706 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 180, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 180, + 91, + 447, + 102 + ], + "type": "text", + "content": "Ray-Distance Volume Rendering for Neural Scene Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/ReALFRED_ An Embodied Instruction Following Benchmark in Photo-Realistic Environments/0e4c20a8-1ef4-4401-b5cc-41d9bc314f7b_content_list.json b/2024/ReALFRED_ An Embodied Instruction Following Benchmark in Photo-Realistic Environments/0e4c20a8-1ef4-4401-b5cc-41d9bc314f7b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6426235a149f920accbad9746a5601e4a2dbe2db --- /dev/null +++ b/2024/ReALFRED_ An Embodied Instruction Following Benchmark in Photo-Realistic Environments/0e4c20a8-1ef4-4401-b5cc-41d9bc314f7b_content_list.json @@ -0,0 +1,1756 @@ +[ + { + "type": "text", + "text": "REALFRED: An Embodied Instruction Following Benchmark in Photo-Realistic Environments", + "text_level": 1, + "bbox": [ + 217, + 140, + 785, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Taewoong Kim $^{1,*}$ , Cheolhong Min $^{1,*}$ , Byeonghwi Kim $^{1}$ , Jinyeon Kim $^{1,2}$ , Wonje Jeung $^{1}$ , and Jonghyun Choi $^{1,\\dagger}$", + "bbox": [ + 282, + 210, + 715, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Seoul National University", + "bbox": [ + 406, + 253, + 593, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Yonsei University", + "bbox": [ + 434, + 268, + 566, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{twoongg.kim, cheolhong.min, byeonghwikim}@snu.ac.kr, jinyeonkim@yonsei.ac.kr, {specific0924,jonghyunchoi}@snu.ac.kr", + "bbox": [ + 256, + 282, + 743, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Simulated virtual environments have been widely used to learn robotic agents that perform daily household tasks. These environments encourage research progress by far, but often provide limited object interactability, visual appearance different from real-world environments, or relatively smaller environment sizes. This prevents the learned models in the virtual scenes from being readily deployable. To bridge the gap between these learning environments and deploying (i.e., real) environments, we propose the REALFRED benchmark that employs real-world scenes, objects, and room layouts to learn agents to complete household tasks by understanding free-form language instructions and interacting with objects in large, multi-room and 3D-captured scenes. Specifically, we extend the ALFRED benchmark with updates for larger environmental spaces with smaller visual domain gaps. With REALFRED, we analyze previously crafted methods for the ALFRED benchmark and observe that they consistently yield lower performance in all metrics, encouraging the community to develop methods in more realistic environments. Our code and data are publicly available3.", + "bbox": [ + 259, + 337, + 743, + 574 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Interactive Scanned Environments $\\cdot$ Instruction Following $\\cdot$ Embodied AI $\\cdot$ Reality Gap $\\cdot$ Dataset and Benchmark", + "bbox": [ + 259, + 585, + 740, + 614 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 654, + 375, + 669 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Building autonomous robotic assistants that can perform everyday household tasks has been an elusive aspiration within the research community for decades. To let them learn these intricate tasks, we may provide them with interactive environments where agents can learn task completion skills with numerous interactions with environments. A straightforward approach to train such agents that can carry out real-world activities is to directly deploy robots in real-world environments and let them learn to complete desired tasks. However, this often faces several practical challenges, including cost, time, or safety concerns [3,14,39,68].", + "bbox": [ + 212, + 681, + 787, + 803 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "3 Homepage: https://github.com/snumprlab/realfred", + "bbox": [ + 236, + 810, + 612, + 824 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. †Corresponding author.", + "bbox": [ + 236, + 824, + 537, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3402938b11df9e6b93447085c34b326633deb8fd70164f71dcb53e1f2184e0d2.jpg", + "image_caption": [ + "Goal: Put a cooled slice of banana on the coffee table" + ], + "image_footnote": [], + "bbox": [ + 246, + 157, + 759, + 311 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e0dd9640bb81a4c70c70f16b531072f75286b4cbb3436e8658bb962114a7e7db.jpg", + "image_caption": [ + "Fig. 1: Proposed REALFRED benchmark. The top image provides a perspective view of one of our scenes. The images below represent third-person views at each time step, along with their corresponding descriptions, for better understanding. The agent is required to understand instructions in natural language and then complete the desired tasks by navigating large 3D-captured environments and interacting with objects." + ], + "image_footnote": [], + "bbox": [ + 246, + 311, + 756, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As an alternative, several simulated environments have been introduced [59, 74] which leverage extensive 3D-captured environments obtained from real-world scenes [7, 56]. Compared to real-world deployment, these environments offer agents a faster process of taking actions and observing consequences, and the convenience of resetting the environment and trying again in case of failure, which enables agents to learn the skills to complete desired tasks. Adopting such simulated environments has produced remarkable advancements in various subtasks for embodied AI agents, including visual navigation [9, 47, 50, 55, 73], vision and language navigation [1, 37], and remote object grounding [54]. Due to the inherently static nature of these 3D-captured environments where objects (e.g., books, chairs, etc.) remain non-interactive, however, current benchmarks [1, 9] for these tasks have less focused on object interaction, which might hinder deployability for more complex tasks that require object interaction. Recent studies [44, 57] insert liftable objects in scanned environments for object interaction, but they support limited object interaction such as picking up objects, which might not provide enough deployability for more challenging real-world scenarios such as heating objects using a microwave or cooling objects using a refrigerator.", + "bbox": [ + 212, + 532, + 787, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Meanwhile, virtual game engines such as Unity have been exploited to build object-interactable environments with graphically crafted assets, including walls, floors, ceilings, and objects. These object-interactable environments have led to", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "T. Kim and C. Min et al.", + "bbox": [ + 271, + 114, + 441, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "notable progress in the handling of more intricate object-centric tasks, including rearrangement [67,72] and manipulation [25,26], beyond navigation-centric tasks. In particular, we have observed significant progress in the execution of more complex tasks by natural language instructions [5, 33, 46, 51, 61, 62]. However, object-interactable environments for training and evaluation of such agents often pose several issues, such as visual domain gaps [69] and relatively smaller room sizes compared to their counterparts in 3D-captured environments [7, 56, 66].", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To bridge the gap between limited object interactability and environmental sizes with visual domain discrepancy, we propose the REAL-WORLD ALFRED (REALFRED) benchmark that requires agents to complete long-horizon tasks by understanding free-form language instructions and interacting with objects in large 3D-captured environments, following a similar task setup to the widely used embodied instruction following benchmark, ALFRED [61]. The 3D-captured environments used in the REALFRED benchmark encompass multiple rooms, providing ample space for agents to engage in multiple rooms in a single episode, which adds a sense of realism to the tasks. This resembles real-world scenarios in which agents navigate seamlessly between mul", + "bbox": [ + 212, + 253, + 517, + 541 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/15b4cd8f949c079b9155588ab4edbed0bf4ea81544184566b97f81cb7455f248.jpg", + "image_caption": [ + "Fig. 2: While other benchmarks [7, 12, 19, 37, 41, 43, 48, 49, 56, 57, 61, 67, 72, 74] provide one or two aspects, our proposed REALFRED benchmark addresses all of these aspects." + ], + "image_footnote": [], + "bbox": [ + 540, + 258, + 777, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tiple rooms. Unlike prior benchmarks focusing primarily on single-room activities [61] or 3D-captured environments with limited object interaction [59], REALFRED provides task evaluation in wide, realistic, and interactive environments to mirror the natural expectations of human-robot interactions. Fig. 1. illustrates the household task in one of the scenes in ours.", + "bbox": [ + 212, + 541, + 787, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our experiments, we observe that the models [5,33,46,61-63] proposed in synthetic environments [61] do not perform well in our REALFRED benchmark, implying that models developed for synthetic environments may not easily adapt to realistic environments.", + "bbox": [ + 212, + 618, + 787, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We summarize our contributions as follows:", + "bbox": [ + 214, + 681, + 532, + 696 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose REALFRED, a benchmark for embodied instruction following with 3D-captured multi-room environments and objects.", + "- We collect 3D-captured scenes and objects to reduce the simulation-reality gap and free-form language instructions to support task completion based on agents' language understanding.", + "- We provide analyses on the recent state-of-the-art models in the literature and the relevant Sim2Real transfer to empirically validate the necessity of our REALFRED benchmark." + ], + "bbox": [ + 212, + 714, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "REALFRED", + "bbox": [ + 638, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/a4bee094765f4e7e0eb96a351332041eae3e5cef0785e96a984bf6ddbc054a0d.jpg", + "table_caption": [ + "Table 1: Comparison of REALFRED and other Embodied AI benchmarks. 'Language' column denotes the number of human annotated language directives. 'Environment' column compares spatial characteristics and whether it supports interactivity. 'Inference' column denotes whether their action space includes interactive capability with objects. The REALFRED benchmark is the first benchmark that provides a 3D-captured and interactable environment to solve household tasks that require navigation and interaction at the same time directed by natural language commands. $^\\dagger$ We count the number of dialogue sessions. $^\\ddagger$ We count the number of annotations in English. *Though [16] environment supports interaction, interaction is not required." + ], + "table_footnote": [], + "table_body": "
LanguageEnvironmentInference
Human AnnotationsVisual QualityMulti-Room NavigationMovable ObjectsState ChangesObject Interactability
IQA [24]-SyntheticXX✓✓
ManipulaTHOR [19]; RoomR [72]-SyntheticX✓✓✓
RoboTHOR [16]-Synthetic✓✓✓X*
ProcTHOR [18]-Synthetic✓✓✓✓
ReplicaCAD [67]-Synthetic✓✓✓✓
BEHAVIOR-1K [41]; HSSD-200 [31]-Synthetic✓✓✓✓
OpenRooms [42]-PhotoXXXX
MP3D [7]; HM3D [56]; Gibson [74]-Photo✓XXX
Habitat-Web [57]-Photo✓✓✓✓
ALFRED [61]; TEACH [49]; CHAI [48]25k+; 2.0k+†;12k+SyntheticX✓✓✓
LANI [48]; Walk the Talk [43]28k+; 0.7k+Synthetic✓XXX
Virtualhome [52]2.7k+Synthetic✓XX✓
R2R [1]; RxR [38]; VLN-CE [37]21k+; 42k+‡; 21k+Photo✓XXX
REALFRED30k+Photo✓✓✓✓
", + "bbox": [ + 218, + 281, + 784, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 503, + 388, + 522 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 2 illustrates the comparison of our REALFRED with other benchmarks in three selected aspects. We first review 3D-captured environments and benchmarks that provide visual aesthetics similar to real-world environments. Then, we review simulation environments and benchmarks that support object interaction. Table 1 compares our REALFRED with other indoor datasets.", + "bbox": [ + 212, + 537, + 787, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets and benchmarks with 3D-captured environments. The collection of advanced 3D scans has played a key role in enhancing our understanding of 3D objects [15, 64, 76] and their perception [2, 30, 45]. While these datasets offer valuable insights for a deeper understanding of 3D environments, they lack object interaction for learning interactive embodied agents. To further promote research on embodied agents for real-world applications, training and evaluation of such agents with physical spaces from scanned data [7, 66, 74] have been proposed. They provide a rich source of data for researchers to explore the capabilities of agents operating within real-world-inspired scenarios.", + "bbox": [ + 212, + 613, + 787, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In these environments, notable progress has been achieved, primarily in the realms of navigation and exploration. Extensive research has been conducted on agents capable of navigating complex 3D environments, as evidenced by work such as navigating to a specified object [9, 47, 55], navigating to a certain point [50, 73], and navigating to the shown image [36]. Similarly, exploring novel environments has also yielded valuable insights [8, 13].", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "T. Kim and C. Min et al.", + "bbox": [ + 271, + 114, + 441, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Additionally, there is a work to integrate multi-modal sensory information for developing an agent that can deeply understand environments using inputs such as vision and language [1,38] and audio and vision [10,11]. These advances enable agents to utilize a broader range of sensory data, enhancing their perception and interaction capabilities. Further research explores additional capabilities of embodied agents, such as object localization alongside navigation [54].", + "bbox": [ + 212, + 146, + 782, + 236 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, a fundamental limitation of these scan-based environments remains their static nature, which restricts the potential for object interaction. This limitation has resulted in the majority of prior research primarily focusing on navigation tasks, leaving the broader scope of agent adaptability to more complex, interactive challenges largely unaddressed. Consequently, this limitation hampers the adaptability of agents to more complex tasks.", + "bbox": [ + 212, + 237, + 784, + 327 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Object-interactable simulated environments. Scan-based environments [56, 76] are typically composed of static scene representations, focusing on the visual aspects of real-world spaces. However, they often fail to adequately support interactivity with objects, representing a discrepancy between the dynamic and interactive characteristics of the real world, necessitating the development of environments better aligned with the demands of interactive agent research.", + "bbox": [ + 212, + 328, + 785, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address this gap, various simulators, such as AI2-THOR [35], ManipulaTHOR [19], RoboTHOR [16], VirtualHome [52], TDW [20], iGibson [40], and OmniGibson [41], have emerged as promising solutions. These environments are engineered with a primary focus on enabling interaction between agents and objects. They are built upon game engines, which provide a solid foundation for ensuring realistic interactivity within the virtual environment. On these simulators, researchers publish benchmarks [16, 21, 41, 49, 53, 61, 65, 75] that support interactions. These benchmarks have been promoting research in the field, leading to the development of robotic assistants capable of handling complex tasks.", + "bbox": [ + 212, + 419, + 784, + 554 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, the dataset that provide interactive environments are limited in size of a room because creating a large high-fidelity space is challenging [56]. Furthermore, despite their near-photo realism, agents would face a visual domain gap when deployed in real world environments [69,71,77]. To this end, Habitat-Web [57] proposes an interactive pick-and-place task based on the Habitat simulator [59,67]. Their template-based language instructions, however, might not be sufficient to express the complex nuances of human expression. Innovative methods for acquiring 3D scans from phone-captured layouts have been proposed for learning enviroment-specific policies [17]. Nevertheless, its synthetic assets may lead to visual domain gaps when deployed in the real world. On the contrary, our REALFRED supports photorealism, high interactivity with objects, and freeform language annotations. These features can offer a framework for developing language-driven agents with visual perception for complex household chores.", + "bbox": [ + 212, + 556, + 784, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 The REALFRED Benchmark", + "text_level": 1, + "bbox": [ + 214, + 776, + 540, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To develop agents capable of performing household tasks, substantial progress has been achieved in various domains, including navigation [1, 38], rearrange-", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "REALFRED", + "bbox": [ + 638, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7dbd702b14f1b095ea5db2c1acc45bbe4d7cac0956e9dbce94033e25598cc581.jpg", + "image_caption": [ + "Fig. 3: Top-down view of 3D-captured environments. We provide two examples from our scanned indoor environments. White circles denote where scanners are deployed. By scanning scenes at diverse points, we can prevent including blind spots." + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 480, + 297 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5c4e1ce4af62ed6466e4f9502c01e5596a57e81adffb45c3db7f7372d41057b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 142, + 759, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ment [72], and manipulation tasks [25,26]. In particular, [61] recently introduced the ALFRED benchmark that requires agents to complete long-horizon household tasks by jointly understanding egocentric visual observations and natural language instructions in household environments.", + "bbox": [ + 212, + 375, + 787, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, these environments are restricted to a single room size compared to previously proposed 3D-captured environments [7, 56] consisting of multiple rooms, which could potentially restrict the deployability of agents to larger environments. Furthermore, the environments used in the ALFRED benchmark [61] are built with synthetic CAD assets and therefore could potentially yield visual aesthetics different from those obtained from real-world environments [69], which could eventually cause performance degradation due to visual domain gaps.", + "bbox": [ + 212, + 436, + 789, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To address these issues, we extend the ALFRED benchmark [61] and propose a challenging benchmark, named the REALFRED benchmark, which requires agents to perform household tasks in large indoor environments captured in 3D with object interaction. For training and evaluation, we follow the same protocol as [61] to collect expert demonstrations in the captured large environments.", + "bbox": [ + 212, + 542, + 789, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Object-Interactable 3D-Captured Scenes", + "text_level": 1, + "bbox": [ + 214, + 636, + 598, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To reduce the visual domain gap, a straightforward approach is to use 3D scans of real world environments. However, the captured 3D scans (i.e., meshes) remain static and thus, agents cannot interact with objects in the captured scenes. For object interaction, we manually replace object parts with 3D object assets to support object interaction. For photorealism comparison with previous environments using FID [27] and KID [4] metrics, please refer to Sec A.3.", + "bbox": [ + 212, + 657, + 787, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We detail the process of collecting object-interactable 3D-captured environments and highlight key differences from previously proposed benchmarks below. Data acquisition process. To collect 3D scans of real-world environments, we visit residential properties and employ scanners. Inspired by recent work [2], we collect scans outside of the US to add the diversity with public scanned environments. We utilize the same 3D scanner as [7], equipped with three RGB cameras", + "bbox": [ + 212, + 750, + 789, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "T. Kim and C. Min et al.", + "bbox": [ + 271, + 114, + 442, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a9f07cfaa82e1d03908cbdb4e1dff55c996e268fdb31727e566c774e133c7fd4.jpg", + "image_caption": [ + "(a) Distribution of floor areas by scene." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 491, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e0395aa238e00e249073a9fc97d6591a5ad5252e1e5bf482a31c37c28554a4e3.jpg", + "image_caption": [ + "(b) Distribution of navigable areas by scene.", + "Fig. 4: Distribution of navigable and floor areas in interactive benchmarks by scenes. 'Floor area' denotes the overall size of the scene. 'Navigable area' denotes the size of the space in which the agent can actually navigate. For both metrics, the REALFRED benchmark poses a more even distribution and provides larger areas. For (a), we exclude RoboTHOR since it consists of single-sized floors." + ], + "image_footnote": [], + "bbox": [ + 495, + 143, + 764, + 224 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "along with a depth sensor, and capture images from three distinct perspectives: front-facing and slightly vertical above and below. Panoramic images are acquired through six consecutive captures with horizontal rotation from a fixed viewpoint. We scan each house with 2.5-meter intervals and address blind spots due to furniture with additional scans from different viewpoints. Fig. 3 shows houses where scanners were deployed while capturing scans. It can be discerned that captures were densely concentrated in areas with a high presence of objects.", + "bbox": [ + 212, + 354, + 787, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Object-interactable environments. To construct an environment with numerous interactable objects, separating each object should be preceded. In other words, objects in the scans are initially merged into the background and therefore they remain as the background. Constructing an environment where objects can be interacted with requires the separation of objects from the background and from each other. Therefore, we manually separate the 3D scans into background elements and interactive objects. Furthermore, each object can exist in various states. To visualize changes in an object's state, we add state-relevant textures on objects. For example, we add a stain texture to a clean object when it becomes dirty. Finally, we reconstruct these individual object meshes within the Unity editor, making them compatible with the AI2-THOR simulator [35].", + "bbox": [ + 212, + 460, + 787, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with previous benchmarks. To investigate the spatial characteristics of scenes in the REALFRED benchmark, we compare ours with other benchmarks that support interaction with objects [16, 61, 67]. We observe enhancements in our dataset in both: 1) spatial sizes and 2) spatial complexity.", + "bbox": [ + 212, + 627, + 785, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Spatial size. We compare the REALFRED benchmark with other benchmarks in terms of spatial size [56] by measuring 'Floor area' and 'Navigable area' and provide the result in Fig. 4. 'Floor area' represents the total spatial size $(m^2)$ of a scene, defined by the floor projection. 'Navigable area' measures the spatial size $(m^2)$ of the space in which an agent can actually navigate. 'Navigable area' is smaller than or equal to 'Floor area' since 'Navigable area' excludes areas where the agent collides with any components in the scene from 'Floor area.'", + "bbox": [ + 212, + 688, + 787, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We observe that the REALFRED benchmark yields a diverse distribution of floor areas, compared to previous work [61], with a larger average per scene. Furthermore, the REALFRED benchmark provides a broader range of navigable", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "REALFRED", + "bbox": [ + 638, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/683e351bcc28f1646523522a89c85a12c212416e40ebeaf04eef4bd0e1fdc76d.jpg", + "table_caption": [ + "Table 2: Component sizes of interactive embodied AI benchmarks. Each 'Total floor area' and 'Total navigable area' (Total nav. area) denotes the sum of all floor areas and navigation areas. 'Nav. complex.' denotes navigational complexity. 'Scene clutter' measures the amount of clutter in the scene. We do not compare ReplicaCAD's Navigable area, Navigation complexity, and Scene clutter as agent sizes differ across simulators. The highest value for each metric is shown in bold. $^\\dagger$ We count objects used as target objects in the ObjectNav task. $^\\ddagger$ We count objects used in any of the tasks." + ], + "table_footnote": [], + "table_body": "
SimulatorHabitat2.0AI2THOR
DatasetReplicaCAD [67]RoboTHOR [16]ALFRED [61]REALFRED
# Scenes11175120150
Total floor area (m2)8,824.52,5742,55510,060
Total nav. area (m2)-1,2581,3564,251
Nav. complex.-2.0362.5493.020
Scene clutter-8.0955.1198.072
# object class9214†82‡112‡
", + "bbox": [ + 225, + 253, + 771, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "areas, with larger areas on average. This implies that an agent needs the ability to navigate effectively in spaces of varying sizes, generally wider on average.", + "bbox": [ + 212, + 411, + 782, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also investigate the spatial characteristics of each benchmark [16, 61, 67] and summarize the result in Table 2. 'Total floor area' denotes the sum of all floor areas and 'Total navigable area' denotes the sum of all navigable areas in the dataset. We observe that the REALFRED benchmark has the highest total floor area and navigable area value, implying that REALFRED provides more navigation space for an agent than previous work [16, 61, 67].", + "bbox": [ + 212, + 443, + 787, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Spatial complexity. We now investigate the complexity of spatial structures in the REALFRED benchmark using several metrics. To ensure a fair comparison with other datasets [16, 61, 67], we employ the navigation complexity introduced by [74] and the scene clutter measurement from [56]. A higher navigation complexity indicates an increased difficulty in navigating through the space, while a higher scene clutter implies the presence of more obstacles in the environment. By utilizing these metrics, we conduct a comparative analysis with object-interactable benchmarks [16, 61] and provide the result in Table 2.", + "bbox": [ + 212, + 536, + 787, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We observe that the REALFRED benchmark provides the environment with higher navigation complexity and scene clutter compared to other benchmarks [61, 67]. The high navigation complexity in our scenes stems from their multi-room composition. This setup requires the agent to execute more intricate navigation when moving from one room to another, in comparison to scenarios where the agent operates within a single room. We observe that ours poses the second-highest scene clutter with a marginal gap from [16]. This is because the spaces in [16] are relatively confined with a high density of furniture. We observe that ours has a similar value to [16], meaning that our scenes have a large amount of obstacles with a similar portion of [16] since our scenes' average size is larger than that of [16]. This implies that the REALFRED benchmark provides a complex and challenging space for the agent to explore the environment.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "T. Kim and C. Min et al.", + "bbox": [ + 271, + 114, + 441, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f41c23831bd75a97a3598e7354dbde795a606657acc32dcd95152ffd4b8184fc.jpg", + "image_caption": [ + "Train", + "Fig. 5: The seven types of tasks' distribution in REALFRED. We provide $37.6\\%$ more tasks in valid sets and $19.3\\%$ in total compared to previous benchmark [61]." + ], + "image_footnote": [], + "bbox": [ + 233, + 166, + 390, + 303 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/e0180f8fb7c3f0ef21b0fb068d837a3069c1812cee376ed45c70b4a5467ad558.jpg", + "image_caption": [ + "Valid Seen" + ], + "image_footnote": [], + "bbox": [ + 392, + 167, + 524, + 303 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/073226163b503917b3560c3247eee85d7b00959533c3443860d60d2f0076a1e0.jpg", + "image_caption": [ + "Valid Unseen" + ], + "image_footnote": [], + "bbox": [ + 527, + 161, + 660, + 301 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/82156fb55d9223338d7119ea5a9190dcf6ba80a128fd00e8ba871738d77ff374.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 160, + 769, + 280 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.2 Expert Demonstration Generation", + "text_level": 1, + "bbox": [ + 214, + 371, + 545, + 386 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Each expert demonstration includes a set of an egocentric RGB view and action information with an interaction mask if exists at each time step. Expert demonstrations for each task are generated by a planner [29] with encoded state spaces into Planning Domain Definition Language (PDDL) rules [22]. To generate household tasks, we utilize seven task types introduced in [61].", + "bbox": [ + 212, + 398, + 784, + 473 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Data splits. We split the generated demonstrations into train, validation, and test folds. Specifically, we designate 135 scenes for seen and 15 scenes for unseen fold. Comparison of the amount of each task and the distribution across training and validation folds with previous work [61] is shown in Fig. 5.", + "bbox": [ + 212, + 474, + 517, + 595 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Curating free-form language instructions. Detailed language instructions describe a task that involves a sequence of actions, a point of interest in robotics. The REALFRED benchmark offers 30,696 language directives, each comprising a human-annotated high-level goal and a set of step-by-step instructions. These directives are collected from 93 Amazon Mechanical Turk workers with a 'Master' qualification, ensuring high", + "bbox": [ + 212, + 595, + 517, + 762 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/048bd0d8c9cef05a5b1bc06fcf3e4ed42f72571eff1eb60e49aa647fc0d3418e.jpg", + "image_caption": [ + "Fig.6:Word distribution.The words are arranged starting from the center and extending outwards. The arc length corresponds to the word frequency in the instructions." + ], + "image_footnote": [], + "bbox": [ + 526, + 474, + 787, + 676 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "quality. Collected annotations are validated through an additional voting survey, and invalid instructions are replaced with newly collected instructions. The distribution of language instructions by their first four words is presented in Fig. 6. We provide a detailed annotation process and examples of an expert demonstration with the instruction in Sec. A.1 and A.4 in the supplementary material.", + "bbox": [ + 212, + 762, + 787, + 838 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "REALFRED", + "bbox": [ + 638, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Multi-room embodied instruction following. For household robots to be able to effectively assist humans, it would be more practical to deploy them in complex and diverse environments with various room types, rather than confining them to a single room. Cross-room navigation poses significant challenges, requiring agents to understand room references, plan efficient paths, and overcome obstacles proficiently. This agent is further tasked to interpret visual cues, demonstrate a keen sense of spatial awareness, and process natural language instructions.", + "bbox": [ + 212, + 146, + 514, + 356 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/291eec7c70ead6d17fe86c455078a5e6226156db1f64a610f18d0a844d9a396a.jpg", + "image_caption": [ + "Fig.7: Comparison of expert demonstrations (ALFRED vs. REALFRED). Due to expanded environments, longer steps and planning horizons are required in REALFRED." + ], + "image_footnote": [], + "bbox": [ + 527, + 146, + 787, + 268 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The distribution of the number of steps and the length covered by the expert demonstrations is shown in Fig. 7. We observe that longer steps and trajectories are required to complete our tasks compared to the single-room constraint benchmark [61], meaning that our benchmark provides longer-horizon tasks.", + "bbox": [ + 212, + 357, + 784, + 417 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To broaden a range of tasks, we provide an extensive number of object class types in the REALFRED benchmark. More object classes are added to the REALFRED benchmark, making it a superset of [61] with 86 pickable and 26 receptacle objects. This has resulted in a more diverse set of tasks, with the number of unique tasks being $84.3\\%$ more than the ALFRED benchmark [61], and the proportion of unique tasks being higher in the REALFRED benchmark. We provide detailed information in Sec. A in the supplementary material.", + "bbox": [ + 212, + 417, + 785, + 525 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 544, + 375, + 560 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Metrics. We follow the same evaluation protocol of the ALFRED benchmark [61]. The primary metric is 'Success Rate (SR)' which measures the percentage of completed tasks. 'Goal-Condition Success Rate (GC)' measures the percentage of achieved goal conditions. We also use path-length-weighted metrics to measure how efficiently an agent completes tasks. For more details, kindly refer to [61]. Baselines. We evaluate several recent state-of-the-art methods [5,33,34,46,61-63] with competitive results in [61]. We provide more details in Sec. B.", + "bbox": [ + 212, + 571, + 790, + 678 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Comparison of the State of the Arts", + "text_level": 1, + "bbox": [ + 214, + 696, + 560, + 713 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We evaluate the baselines in the proposed REALFRED benchmark over multiple runs and present the average result in Table 3. We report extended results with path-length-weighted metrics in Sec. C in supplementary.", + "bbox": [ + 212, + 719, + 784, + 763 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For a fair comparison, we separate the baseline into two groups based on the use of additional depth supervision for semantic map reconstruction: 'Imitation Learning' where agents learn direct mapping from visual observations and language instructions to action sequences and 'Spatial Map Reconst.' where agents plan action sequences based on reconstructed semantic spatial representations.", + "bbox": [ + 212, + 765, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "T. Kim and C. Min et al.", + "bbox": [ + 271, + 114, + 441, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/e4a9e400ba27c496621b6d55dc9ad77099f8f46ea37c8fe0e49014c8955e49c8.jpg", + "table_caption": [ + "Table 3: Task and Goal-Condition Success Rate. We train and evaluate recent state-of-the-art methods on our REALFRED benchmark. For a fair comparison, we group these methods into two based on the usage of extra depth supervision: models learned by imitation learning without depth supervision ('Imitation Learning') and ones that maintain semantic spatial representations constructed by predicted depth maps ('Spatial Map Reconst.'). Path-length-weighted (PLW) metrics are reported in Sec. C for each value. $\\dagger$ Authors' implementation as the code is not publicly available." + ], + "table_footnote": [], + "table_body": "
LearningModelValidationTest
SeenUnseenSeenUnseen
SRGCSRGCSRGCSRGC
Imitation LearningSeq2Seq [61]0.77 ± 0.066.93 ± 0.060.00 ± 0.004.03 ± 0.001.10 ± 0.006.60 ± 0.000.00 ± 0.003.50 ± 0.00
MOCA [62]12.64 ± 0.1220.95 ± 0.181.44 ± 0.056.76 ± 0.0414.11 ± 0.0322.84 ± 0.040.62 ± 0.085.14 ± 0.08
ABP† [33]24.71 ± 0.0533.80 ± 0.144.22 ± 0.0511.71 ± 0.2727.44 ± 0.4035.81 ± 0.233.54 ± 0.2310.57 ± 0.22
Spatial Map Reconst.HLSM [5]4.23 ± 0.089.14 ± 0.091.08 ± 0.146.12 ± 0.236.27 ± 0.0410.44 ± 0.130.49 ± 0.164.28 ± 0.13
FILM [46]7.08 ± 0.2811.93 ± 0.234.44 ± 0.179.26 ± 0.138.79 ± 0.0713.03 ± 0.082.15 ± 0.186.56 ± 0.15
LLM-Planner† [63]5.80 ± 0.1911.69 ± 0.353.33 ± 0.228.29 ± 0.198.16 ± 0.2013.20 ± 0.131.90 ± 0.136.33 ± 0.02
CAPEAM† [34]13.45 ± 0.0518.16 ± 0.274.92 ± 0.229.47 ± 0.2315.61 ± 0.1520.22 ± 0.112.87 ± 0.137.36 ± 0.07
Human------85.00 ± 3.5491.30 ± 2.94
", + "bbox": [ + 218, + 253, + 784, + 373 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We observe that all these baselines, 'Imitation Learning' and 'Spatial Map Reconst.' consistently achieve lower performance values for all metrics in both seen and unseen splits compared to performances achieved in [61], implying that our proposed REALFRED provide more challenges compared to [61]. While in [61], the 'Spatial Map Reconst.' baselines [5,34,46] outperform learning-based approaches [33,61,62] by exploiting semantic spatial representation with deterministic algorithms (e.g., obstacle-free navigation path planning [60]), we observe a contrasting result that the 'Imitation Learning' baselines outperforming the 'Spatial Map Reconst.' baselines in our REALFRED benchmark.", + "bbox": [ + 212, + 402, + 787, + 537 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We qualitatively observe that such a confined spatial map reconstruction is due to a limited field of view and map reconstruction methods. The agent's limited field of view often leads to failure in recognizing room corners with doors or narrow aisles, resulting in (single-room-sized) limited map reconstruction. In addition, [34, 46] perceives obstacles larger than they actually are for better obstacle-free path planning by sacrificing navigable area, but this can be quite critical for narrow doors and aisles as they have a small amount of navigable space. This may hinder navigation to other rooms and thus, fail at tasks. We provide a more detailed discussion, supported by figures, on the agents' difficulty in recognizing narrow passages and the resulting impact on the agents, who struggle with spatial map reconstruction in Sec. D in the supplementary material.", + "bbox": [ + 212, + 539, + 787, + 705 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Comparison to the agents with sim-to-real adaptation", + "text_level": 1, + "bbox": [ + 214, + 724, + 705, + 739 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We investigate the transfer from simulation training to real-world scan evaluation (sim-to-real) and from real-world scan training to real-world scan evaluation (real-to-real) with [33]. We train a sim-to-real agent with synthetic visual data and a real-to-real agent with real scanned visual data, respectively. During inference, both agents predict an action and an object mask based on the scanned visual input frame and the given language instruction at every time step.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "REALFRED", + "bbox": [ + 638, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Training and evaluation data selection. For training the sim-to-real agent, we use the training dataset in [61], encompassing $21K$ language annotations. For a fair comparison, we train the real-to-real agent with tasks from the REALFRED benchmark's training fold, specifically involving the manipulation of the same objects that were used in training the sim-to-real agent with $19K$ language annotations. For evaluation, we select tasks from the valid unseen splits of our REALFRED benchmark. We evaluate the agent performance with the tasks that 1) do not require multi-room navigation 2) and those that do. This selection specifically includes tasks that feature objects used in the training phase.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Real-to-sim domain adaptation. The use of Generative Adversarial Networks (GANs) [23,79] for domain adaptation has recently been examined in the literature on robotics [6,28,32,58,77]. These models are employed to adapt input images from real-world domains (i.e. Real $\\rightarrow$ Sim) before they are passed to the agent policy. We train a CycleGan [79] and its off-the-shelf variant [70] with unpaired images collected from the REALFRED (real domain) and ALFRED benchmarks [61] (simulated domain). Among the trained generators, the one performing real to simulation conversion, referred to as the real-to-sim goggle, learns the mapping from the real domain to the simulation domain $G_{S}:\\mathcal{R}\\to \\mathcal{S}$ , where $\\mathcal{R}$ denotes the real domain and $S$ denotes the simulated domain.", + "bbox": [ + 212, + 282, + 787, + 431 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Results. We present the results of the sim-to-real experiments in Table 4. We report the performance of agents when tasks 1) require agents to navigate multirooms, denoted as 'Multi + Single,' and 2) are solvable within a single room, denoted as 'Single only.' Firstly, we observe that the sim-to-real agent significantly underperforms its real-to-real counterpart in all metrics $(\\# (a)$ vs. $\\# (d))$ .We then compare the results with gobbled agents. By comparing agents evaluated in the Multi", + "bbox": [ + 212, + 433, + 517, + 613 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/ec66edd07ae13102485f977aea58e22cd1ce658df2695f350600e686750f691a.jpg", + "table_caption": [ + "Table 4: Comparison to the agents with sim-to-real adaptation. Each 'Sim2Real' and 'Real2Real' denotes an agent trained in simulated environments and 3D-captured scenes. 'Goggle' denotes real-to-sim methods." + ], + "table_footnote": [], + "table_body": "
# SettingGoggleMulti + SingleSingle only
SRPLWSRSRPLWSR
(a) Sim2RealNone0.1150.0120.00.0
(b) Sim2RealCycleGan [79]0.1150.0160.3270.065
(c) Sim2RealUVCGAN-v2 [70]0.1150.0460.3270.187
(d) Real2RealNone2.4050.7852.6140.762
", + "bbox": [ + 529, + 534, + 787, + 598 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "+ Single' tasks, we do not observe improvements in the main metric, success rate (SR). We observe a slight increase in the PLWSR metric for the gobbled sim-to-real agent compared to the vanilla sim-to-real agent (#(a) vs. #(b,c)).", + "bbox": [ + 212, + 614, + 787, + 660 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To isolate the impacts of the visual domain gap, we then compare agents evaluated on 'Single only' tasks. We observe that both gobbled agents show improvements over the vanilla sim-to-real agent which results in zero success rate. However, a noticeable performance gap exists with the real-to-real agent, implying the need for learning in real scanned environments $(\\# (d)$ vs. $\\# (a,b,c))$", + "bbox": [ + 212, + 660, + 787, + 736 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Challenges in REALFRED", + "text_level": 1, + "bbox": [ + 215, + 755, + 486, + 771 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We propose two hypotheses for the low performance observed with state-of-the-art methods on the REALFRED benchmark: (1) navigating within larger scenes and (2) overcoming narrow pathways between rooms. These elements represent challenges of completing tasks within multi-room, household-scale environments.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "T. Kim and C. Min et al.", + "bbox": [ + 271, + 114, + 441, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3eaaf4466a90b882d59174864fcbaf743526bbd11465a420d2d8e479776af21d.jpg", + "image_caption": [ + "Fig. 8: Distribution of collision spots. The heatmap illustrates collision frequency on each spot and the black outline depicts a navigable area. We illustrate collision points of [33] in valid unseen environments to represent the distribution. Specifically, we choose failure episodes from the two largest scenes in the valid unseen split. We observe a remarkable concentration of collisions at the entrance to the rooms. The corresponding egocentric views are presented on the right side of the figure." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 781, + 272 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We analyze failure cases with the best-performing model, ABP [33], on valid unseen split, since it seems the most promising on the REALFRED benchmark. Difficulty in large environments. Understanding the surrounding environment, including the location of objects and the positions of obstacles, can be beneficial for task completion. However, the agent's visual range is bounded, requiring it to expend more steps for exploration in order to perceive larger spaces. This intensifies the challenge of navigating larger spaces with limited steps.", + "bbox": [ + 212, + 398, + 787, + 505 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We conduct an analysis to see the relationship between the success rate (SR) and the size of the space. We follow the interquartile range (IQR) method to set a threshold at $30.44m^2$ that covers the most navigable space sizes in the ALFRED [61]. This threshold marks the upper fence, defining spaces above it as outliers. We classify spaces in the REALFRED benchmark smaller than threshold as smaller scenes, and those larger as larger scenes. Results indicate that the agent [33] with the highest SR in the valid unseen fold showed an average SR of $5.46\\%$ in the smaller scenes and only $1.77\\%$ in the larger scenes. This implies that solving tasks becomes more challenging as the space size increases.", + "bbox": [ + 212, + 506, + 787, + 642 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In addition, we compare the difficulty of navigation between the previous [61] and REALFRED benchmarks with different average spatial sizes. To quantify, we first define milestones for each task, which are spots to be reached to interact with target objects (e.g., Apple, Knife, etc.). We consider navigation to be a success only if all milestones are visited, regardless of whether the actual interaction (e.g., slicing an apple) is performed or not. Consequently, the navigation success rate in the REALFRED benchmark is merely $59.18\\%$ for the valid unseen split while the agent's [33] navigation success rate is $84.82\\%$ in ALFRED [61]. The REALFRED benchmark offering an average space size more than three times larger than the previous work [61] implies the need for model development capable of overcoming the challenges associated with this increased scale.", + "bbox": [ + 212, + 643, + 787, + 808 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Navigation through narrow doorways. The REALFRED benchmark supports environments with multiroom composition, unlike the previous dataset [61]", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "REALFRED", + "bbox": [ + 638, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "providing a single room scale. Specifically, it contains narrow doorways across rooms, which may hinder the agent from navigating through. Here, we hypothesize that the agent would frequently collide with the walls near the door.", + "bbox": [ + 212, + 146, + 784, + 191 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We investigate collision spots and showcase two examples on top of the scene's layout in valid unseen fold in Fig. 8. The number of collisions in failed cases is accumulated and normalized to the maximum number of collisions, respectively, to indicate the collision frequency at each point. The black outline represents the agent's navigable area, as detailed in Sec. 3.1. Each value denotes the normalized collision frequency. In both scenes, we observe that collisions often occur on walls near the door, implying that our spatial characteristic (i.e., including narrow spaces) may hinder the agent from properly navigating without collision.", + "bbox": [ + 212, + 191, + 785, + 311 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Human evaluation. Following [61], we randomly select 100 directives from the test unseen fold and have them evaluated by humans. Five participants are given 20 tasks each, which they complete using a keyboard-and-mouse interface. Before starting, they are given the opportunity to become familiar with the interface.", + "bbox": [ + 212, + 313, + 785, + 373 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Participants achieve a comparable high success rate of $85\\%$ , along with a goal-condition success rate of $91.30\\%$ , in average. We agree that human performance may appear slightly lower compared to the results presented in previous work [57, 61, 78]. This is partly because human participants encountered several difficulties when controlling the agent, particularly in avoiding collisions within narrow corridors. Furthermore, navigating large spaces with a limited egocentric field of view introduced additional challenges, leading to task failures.", + "bbox": [ + 212, + 375, + 787, + 479 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 503, + 359, + 520 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We present REALFRED, a new dataset and benchmark for embodied instruction following task on 3D-captured environments. We capture 150 indoor houses in 3D with interactable objects to enable complex household tasks. The reconstructed indoor scenes provide a larger spatial area and complex multi-room environments that are close to the real-world scenario and challenging for an agent to successfully complete a task. Expert demonstrations are also provided along with free-form human-language instructions.", + "bbox": [ + 212, + 537, + 785, + 643 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In our empirical evaluations, we show that state-of-the-art methods struggle in large multi-room environments, provide analyses of our newly proposed benchmark, and perform Sim2Real transfer experiments. We have released our Embodied AI research data and code for reproducibility. We expect that the REALFRED benchmark will encourage further research on developing robotic agents that execute household tasks by language instructions in the real world.", + "bbox": [ + 212, + 643, + 787, + 734 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitation and future work. Although we support a large number of interactable objects, the types of tasks to be completed are rather limited, considering more complex real-world scenarios. In addition, we currently address natural language in English but users may come from different regions with different languages. We can think of two future research avenues as follows. (1) adding additional complicated types of task that require both hands to complete. (2) supporting a multi-lingual interface for users from different regions.", + "bbox": [ + 212, + 734, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "T. Kim and C. Min et al.", + "bbox": [ + 271, + 114, + 441, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was partly supported by the NRF grant (No.2022R1A2C400230012, $5\\%$ ) and IITP grants (No.RS-2022-II220077 $(5\\%)$ , No.RS-2022-II220113 $(5\\%)$ No.RS-2022-II220959 $(5\\%)$ , No.RS-2022-II220871 $(15\\%)$ , No.RS-2020-II201361 $(5\\%)$ , Yonsei AI), No.RS-2021-II211343 $(5\\%)$ , SNU AI), No.RS-2021-II212068 $(5\\%)$ , AI Innov. Hub), No.RS-2022-II220951 $(50\\%)$ ) funded by the Korea government(MSIT).", + "bbox": [ + 212, + 175, + 787, + 267 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 287, + 321, + 304 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Anderson, P., Wu, Q., Teney, D., Bruce, J., Johnson, M., Sünderhauf, N., Reid, I., Gould, S., van den Hengel, A.: Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In: CVPR (2018)", + "2. Baruch, G., Chen, Z., Dehghan, A., Dimry, T., Feigin, Y., Fu, P., Gebauer, T., Joffe, B., Kurz, D., Schwartz, A., Shulman, E.: ARKitscenes - a diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In: NeurIPS Datasets and Benchmarks Track (2021)", + "3. Berseth, G., Xie, C., Cernek, P., van de Panne, M.: Progressive reinforcement learning with distillation for multi-skilled motion control. In: ICLR (2018)", + "4. Binkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying mmd gans. In: ICLR (2018)", + "5. Blukis, V., Paxton, C., Fox, D., Garg, A., Artzi, Y.: A persistent spatial semantic representation for high-level natural language instruction execution. In: CoRL (2021)", + "6. Bousmalis, K., Irpan, A., Wohlhart, P., Bai, Y., Kelcey, M., Kalakrishnan, M., Downs, L., Ibarz, J., Pastor, P., Konolige, K., et al.: Using simulation and domain adaptation to improve efficiency of deep robotic grasping. In: ICRA (2018)", + "7. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3d: Learning from rgb-d data in indoor environments. arXiv:1709.06158 (2017)", + "8. Chaplot, D.S., Gandhi, D., Gupta, S., Gupta, A., Salakhutdinov, R.: Learning to explore using active neural slam. In: ICLR (2020)", + "9. Chaplot, D.S., Gandhi, D.P., Gupta, A., Salakhutdinov, R.R.: Object goal navigation using goal-oriented semantic exploration. In: NeurIPS (2020)", + "0. Chen, C., Jain, U., Schissler, C., Gari, S.V.A., Al-Halah, Z., Ithapu, V.K., Robinson, P., Grauman, K.: Soundspaces: Audio-visual navigation in 3d environments. In: ECCV (2020)", + "1. Chen, C., Schissler, C., Garg, S., Kobernik, P., Clegg, A., Calamia, P., Batra, D., Robinson, P.W., Grauman, K.: Soundspaces 2.0: A simulation platform for visual-acoustic learning. In: NeurIPS Datasets and Benchmarks Track (2022)", + "2. Chen, H., Suhr, A., Misra, D., Snavely, N., Artzi, Y.: Touchdown: Natural language navigation and spatial reasoning in visual street environments. In: CVPR (2019)", + "3. Chen, T., Gupta, S., Gupta, A.: Learning exploration policies for navigation. In: ICLR (2019)", + "4. Choi, S., Ji, G., Park, J., Kim, H., Mun, J., Lee, J.H., Hwangbo, J.: Learning quadrupedal locomotion on deformable terrain. Science Robotics (2023)", + "5. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017)" + ], + "bbox": [ + 225, + 316, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "REALFRED", + "bbox": [ + 638, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Deitke, M., Han, W., Herrasti, A., Kembhavi, A., Kolve, E., Mottaghi, R., Salvador, J., Schwenk, D., VanderBilt, E., Wallingford, M., et al.: Robothor: An open simulation-to-real embodied ai platform. In: CVPR (2020)", + "17. Deitke, M., Hendrix, R., Farhadi, A., Ehsani, K., Kembhavi, A.: Phone2proc: Bringing robust robots into our chaotic world. In: CVPR (2023)", + "18. Deitke, M., VanderBilt, E., Herrasti, A., Weihs, L., Ehsani, K., Salvador, J., Han, W., Kolve, E., Kembhavi, A., Mottaghi, R.: Procthor: Large-scale embodied ai using procedural generation. In: NeurIPS (2022)", + "19. Ehsani, K., Han, W., Herrasti, A., Vanderbilt, E., Weihs, L., Kolve, E., Kembhavi, A., Mottaghi, R.: Manipulathor: A framework for visual object manipulation. In: CVPR (2021)", + "20. Gan, C., Schwartz, J., Alter, S., Mrowca, D., Schrimpf, M., Traer, J., De Freitas, J., Kubilius, J., Bhandwaldar, A., Haber, N., et al.: Threadworld: A platform for interactive multi-modal physical simulation. In: NeurIPS Datasets and Benchmarks Track (2021)", + "21. Gao, X., Gao, Q., Gong, R., Lin, K., Thattai, G., Sukhatme, G.S.: Dialfred: Dialogue-enabled agents for embodied instruction following. RA-L (2022)", + "22. Ghallab, M., Howe, A., Knoblock, C., McDermott, D., Ram, A., Veloso, M., Weld, D., Wilkins, D.: Pddl the planning domain definition language (1998)", + "23. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: NeurIPS (2014)", + "24. Gordon, D., Kembhavi, A., Rastegari, M., Redmon, J., Fox, D., Farhadi, A.: Iqa: Visual question answering in interactive environments. In: CVPR (2018)", + "25. Gu, J., Xiang, F., Li, X., Ling, Z., Liu, X., Mu, T., Tang, Y., Tao, S., Wei, X., Yao, Y., et al.: Maniskill2: A unified benchmark for generalizable manipulation skills. In: ICLR (2023)", + "26. Heo, M., Lee, Y., Lee, D., Lim, J.J.: Furniturebench: Reproducible real-world benchmark for long-horizon complex manipulation. In: RSS (2023)", + "27. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017)", + "28. Ho, D., Rao, K., Xu, Z., Jang, E., Khansari, M., Bai, Y.: Retinagan: An object-aware approach to sim-to-real transfer. In: ICRA (2021)", + "29. Hoffmann, J., Nebel, B.: The ff planning system: Fast plan generation through heuristic search. In: JAIR (2001)", + "30. Hua, B.S., Pham, Q.H., Nguyen, D.T., Tran, M.K., Yu, L.F., Yeung, S.K.: Scenenn: A scene meshes dataset with annotations. In: 3DV (2016)", + "31. Khanna, M., Mao, Y., Jiang, H., Haresh, S., Shacklett, B., Batra, D., Clegg, A., Undersander, E., Chang, A.X., Savva, M.: Habitat synthetic scenes dataset (hssd200): An analysis of 3d scene scale and realism tradeoffs for objectgoal navigation. In: CVPR (2024)", + "32. Khansari, M., Ho, D., Du, Y., Fuentes, A., Bennice, M., Sievers, N., Kirmani, S., Bai, Y., Jang, E.: Practical visual deep imitation learning via task-level domain consistency. In: ICRA (2023)", + "33. Kim, B., Bhambri, S., Singh, K.P., Mottaghi, R., Choi, J.: Agent with the big picture: Perceiving surroundings for interactive instruction following. In: Embodied AI Workshop @ CVPR (2021)", + "34. Kim, B., Kim, J., Kim, Y., Min, C., Choi, J.: Context-aware planning and environment-aware memory for instruction following embodied agents. In: ICCV (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "T. Kim and C. Min et al.", + "bbox": [ + 271, + 114, + 441, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Kolve, E., Mottaghi, R., Han, W., VanderBilt, E., Weihs, L., Herrasti, A., Gordon, D., Zhu, Y., Gupta, A., Farhadi, A.: Ai2-thor: An interactive 3d environment for visual ai. arXiv:1712.05474 (2017)", + "36. Krantz, J., Gervet, T., Yadav, K., Wang, A., Paxton, C., Mottaghi, R., Batra, D., Malik, J., Lee, S., Chaplot, D.S.: Navigating to objects specified by images. In: ICCV (2023)", + "37. Krantz, J., Wijmans, E., Majumdar, A., Batra, D., Lee, S.: Beyond the nav-graph: Vision-and-language navigation in continuous environments. In: ECCV (2020)", + "38. Ku, A., Anderson, P., Patel, R., Ie, E., Baldridge, J.: Room-Across-Room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. In: EMNLP (2020)", + "39. Kumar, A., Fu, Z., Pathak, D., Malik, J.: Rapid motor adaptation for legged robots. In: RSS (2021)", + "40. Li, C., Xia, F., Martin-Martin, R., Lingelbach, M., Srivastava, S., Shen, B., Vainio, K.E., Gokmen, C., Dharan, G., Jain, T., Kurenkov, A., Liu, K., Gweon, H., Wu, J., Fei-Fei, L., Savarese, S.: igibson 2.0: Object-centric simulation for robot learning of everyday household tasks. In: CoRL (2021)", + "41. Li, C., Zhang, R., Wong, J., Gokmen, C., Srivastava, S., Martin-Martin, R., Wang, C., Levine, G., Lingelbach, M., Sun, J., et al.: Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation. In: CoRL (2022)", + "42. Li, Z., Yu, T.W., Sang, S., Wang, S., Song, M., Liu, Y., Yeh, Y.Y., Zhu, R., Gundavarapu, N., Shi, J., Bi, S., Yu, H.X., Xu, Z., Sunkavalli, K., Hasan, M., Ramamoorthi, R., Chandraker, M.: Open rooms: An open framework for photorealistic indoor scene datasets. In: CVPR (2021)", + "43. MacMahon, M., Stankiewicz, B., Kuipers, B.: Walk the talk: Connecting language, knowledge, and action in route instructions. In: AAAI (2006)", + "44. Majumdar, et al.: Findthis: Language-driven object disambiguation in indoor environments. In: CoRL (2023)", + "45. Mao, Y., Zhang, Y., Jiang, H., Chang, A., Savva, M.: Multiscan: Scalable rgbd scanning for 3d environments with articulated objects. In: NeurIPS (2022)", + "46. Min, S.Y., Chaplot, D.S., Ravikumar, P., Bisk, Y., Salakhutdinov, R.: Film: Following instructions in language with modular methods. In: ICLR (2022)", + "47. Min, S.Y., Tsai, Y.H.H., Ding, W., Farhadi, A., Salakhutdinov, R., Bisk, Y., Zhang, J.: Object goal navigation with end-to-end self-supervision. In: IROS (2023)", + "48. Misra, D., Bennett, A., Blukis, V., Niklasson, E., Shatkhin, M., Artzi, Y.: Mapping instructions to actions in 3d environments with visual goal prediction. In: EMNLP (2018)", + "49. Padmakumar, A., Thomason, J., Shrivastava, A., Lange, P., Narayan-Chen, A., Gella, S., Piramuthu, R., Tur, G., Hakkani-Tur, D.: Teach: Task-driven embodied agents that chat. In: AAAI (2022)", + "50. Partsey, R., Wijmans, E., Yokoyama, N., Dobosevych, O., Batra, D., Maksymets, O.: Is mapping necessary for realistic pointgoal navigation? In: CVPR (2022)", + "51. Pashevich, A., Schmid, C., Sun, Chen: Episodic transformer for vision-and-language navigation. In: ICCV (2021)", + "52. Puig, X., Ra, K., Boben, M., Li, J., Wang, T., Fidler, S., Torralba, A.: Virtualhome: Simulating household activities via programs. In: CVPR (2018)", + "53. Puig, X., Undersander, E., Szot, A., Cote, M.D., Yang, T.Y., Partsey, R., Desai, R., Clegg, A.W., Hlavac, M., Min, S.Y., et al.: Habitat 3.0: A co-habitat for humans, avatars and robots. In: ICLR (2024)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "REALFRED", + "bbox": [ + 638, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "54. Qi, Y., Wu, Q., Anderson, P., Wang, X., Wang, W.Y., Shen, C., Hengel, A.v.d.: Reverie: Remote embodied visual referring expression in real indoor environments. In: CVPR (2020)", + "55. Ramakrishnan, S.K., Chaplot, D.S., Al-Halah, Z., Malik, J., Grauman, K.: Poni: Potential functions for objectgoal navigation with interaction-free learning. In: CVPR (2022)", + "56. Ramakrishnan, S.K., Gokaslan, A., Wijmans, E., Maksymets, O., Clegg, A., Turner, J.M., Undersander, E., Galuba, W., Westbury, A., Chang, A.X., Savva, M., Zhao, Y., Batra, D.: Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI. In: NeurIPS Datasets and Benchmarks Track (2021)", + "57. Ramrakhya, R., Undersander, E., Batra, D., Das, A.: Habitat-web: Learning embodied object-search strategies from human demonstrations at scale. In: CVPR (2022)", + "58. Rao, K., Harris, C., Irpan, A., Levine, S., Ibarz, J., Khansari, M.: Rl-cyclegan: Reinforcement learning aware simulation-to-real. In: CVPR (2020)", + "59. Savva, M., Kadian, A., Maksymets, O., Zhao, Y., Wijmans, E., Jain, B., Straub, J., Liu, J., Koltun, V., Malik, J., Parikh, D., Batra, D.: Habitat: A Platform for Embodied AI Research. In: ICCV (2019)", + "60. Sethian, J.A.: A fast marching level set method for monotonically advancing fronts. In: PNAS (1996)", + "61. Shridhar, M., Thomason, J., Gordon, D., Bisk, Y., Han, W., Mottaghi, R., Zettle-moyer, L., Fox, D.: Alfred: A benchmark for interpreting grounded instructions for everyday tasks. In: CVPR (2020)", + "62. Singh, K.P., Bhambri, S., Kim, B., Mottaghi, R., Choi, J.: Factorizing perception and policy for interactive instruction following. In: ICCV (2021)", + "63. Song, C.H., Wu, J., Washington, C., Sadler, B.M., Chao, W.L., Su, Y.: Llm-planner: Few-shot grounded planning for embodied agents with large language models. In: ICCV (2023)", + "64. Song, S., Lichtenberg, S.P., Xiao, J.: Sun rgb-d: A rgb-d scene understanding benchmark suite. In: CVPR (2015)", + "65. Srivastava, S., Li, C., Lingelbach, M., Martin-Martin, R., Xia, F., Vainio, K.E., Lian, Z., Gokmen, C., Buch, S., Liu, K., et al.: Behavior: Benchmark for everyday household activities in virtual, interactive, and ecological environments. In: CoRL (2021)", + "66. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv:1906.05797 (2019)", + "67. Szot, A., Clegg, A., Undersander, E., Wijmans, E., Zhao, Y., Turner, J., Maestre, N., Mukadam, M., Chaplot, D., Maksymets, O., Gokaslan, A., Vondrus, V., Dharur, S., Meier, F., Galuba, W., Chang, A., Kira, Z., Koltun, V., Malik, J., Savva, M., Batra, D.: Habitat 2.0: Training home assistants to rearrange their habitat. In: NeurIPS (2021)", + "68. Tan, J., Zhang, T., Coumans, E., Iscen, A., Bai, Y., Hafner, D., Bohez, S., Vanhoucke, V.: Sim-to-real: Learning agile locomotion for quadruped robots. In: RSS (2018)", + "69. Tobin, J., Fong, R., Ray, A., Schneider, J., Zaremba, W., Abbeel, P.: Domain randomization for transferring deep neural networks from simulation to the real world. In: IROS (2017)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "T. Kim and C. Min et al.", + "bbox": [ + 271, + 114, + 441, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "70. Torbunov, D., Huang, Y., Tseng, H.H., Yu, H., Huang, J., Yoo, S., Lin, M., Viren, B., Ren, Y.: Uvçgan v2: An improved cycle-consistent gan for unpaired image-to-image translation. arXiv:2303.16280 (2023)", + "71. Truong, J., Chernova, S., Batra, D.: Bi-directional domain adaptation for sim2real transfer of embodied navigation agents. RA-L (2021)", + "72. Weihs, L., Deitke, M., Kembhavi, A., Mottaghi, R.: Visual room rearrangement. In: CVPR (2021)", + "73. Wijmans, E., Kadian, A., Morcos, A., Lee, S., Essa, I., Parikh, D., Savva, M., Batra, D.: Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. In: ICLR (2020)", + "74. Xia, F., Zamir, A.R., He, Z., Sax, A., Malik, J., Savarese, S.: Gibson env: Realworld perception for embodied agents. In: CVPR (2018)", + "75. Yenamandra, S., Ramachandran, A., Yadav, K., Wang, A., Khanna, M., Gervet, T., Yang, T.Y., Jain, V., Clegg, A.W., Turner, J., et al.: Homerobot: Open-vocabulary mobile manipulation. In: CoRL (2023)", + "76. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: ICCV (2023)", + "77. Zhang, J., Tai, L., Yun, P., Xiong, Y., Liu, M., Boedecker, J., Burgard, W.: Vr-goggles for robots: Real-to-sim domain adaptation for visual control. RA-L (2019)", + "78. Zhu, H., Kapoor, R., Min, S.Y., Han, W., Li, J., Geng, K., Neubig, G., Bisk, Y., Kembhavi, A., Weihs, L.: Excalibur: Encouraging and evaluating embodied exploration. In: CVPR (2023)", + "79. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: ICCV (2017)" + ], + "bbox": [ + 212, + 146, + 787, + 479 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "REALFRED", + "bbox": [ + 638, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/ReALFRED_ An Embodied Instruction Following Benchmark in Photo-Realistic Environments/0e4c20a8-1ef4-4401-b5cc-41d9bc314f7b_model.json b/2024/ReALFRED_ An Embodied Instruction Following Benchmark in Photo-Realistic Environments/0e4c20a8-1ef4-4401-b5cc-41d9bc314f7b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..559cdf03168623baea326138881b113c6c4a1e4c --- /dev/null +++ b/2024/ReALFRED_ An Embodied Instruction Following Benchmark in Photo-Realistic Environments/0e4c20a8-1ef4-4401-b5cc-41d9bc314f7b_model.json @@ -0,0 +1,2713 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.218, + 0.141, + 0.787, + 0.185 + ], + "angle": 0, + "content": "REALFRED: An Embodied Instruction Following Benchmark in Photo-Realistic Environments" + }, + { + "type": "text", + "bbox": [ + 0.284, + 0.212, + 0.717, + 0.244 + ], + "angle": 0, + "content": "Taewoong Kim\\(^{1,*}\\), Cheolhong Min\\(^{1,*}\\), Byeonghwi Kim\\(^{1}\\), Jinyeon Kim\\(^{1,2}\\), Wonje Jeung\\(^{1}\\), and Jonghyun Choi\\(^{1,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.408, + 0.255, + 0.594, + 0.269 + ], + "angle": 0, + "content": "\\(^{1}\\) Seoul National University" + }, + { + "type": "text", + "bbox": [ + 0.435, + 0.269, + 0.567, + 0.284 + ], + "angle": 0, + "content": "2 Yonsei University" + }, + { + "type": "text", + "bbox": [ + 0.257, + 0.284, + 0.745, + 0.312 + ], + "angle": 0, + "content": "{twoongg.kim, cheolhong.min, byeonghwikim}@snu.ac.kr, jinyeonkim@yonsei.ac.kr, {specific0924,jonghyunchoi}@snu.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.338, + 0.744, + 0.575 + ], + "angle": 0, + "content": "Abstract. Simulated virtual environments have been widely used to learn robotic agents that perform daily household tasks. These environments encourage research progress by far, but often provide limited object interactability, visual appearance different from real-world environments, or relatively smaller environment sizes. This prevents the learned models in the virtual scenes from being readily deployable. To bridge the gap between these learning environments and deploying (i.e., real) environments, we propose the REALFRED benchmark that employs real-world scenes, objects, and room layouts to learn agents to complete household tasks by understanding free-form language instructions and interacting with objects in large, multi-room and 3D-captured scenes. Specifically, we extend the ALFRED benchmark with updates for larger environmental spaces with smaller visual domain gaps. With REALFRED, we analyze previously crafted methods for the ALFRED benchmark and observe that they consistently yield lower performance in all metrics, encouraging the community to develop methods in more realistic environments. Our code and data are publicly available3." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.587, + 0.741, + 0.615 + ], + "angle": 0, + "content": "Keywords: Interactive Scanned Environments \\(\\cdot\\) Instruction Following \\(\\cdot\\) Embodied AI \\(\\cdot\\) Reality Gap \\(\\cdot\\) Dataset and Benchmark" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.655, + 0.377, + 0.67 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.682, + 0.788, + 0.804 + ], + "angle": 0, + "content": "Building autonomous robotic assistants that can perform everyday household tasks has been an elusive aspiration within the research community for decades. To let them learn these intricate tasks, we may provide them with interactive environments where agents can learn task completion skills with numerous interactions with environments. A straightforward approach to train such agents that can carry out real-world activities is to directly deploy robots in real-world environments and let them learn to complete desired tasks. However, this often faces several practical challenges, including cost, time, or safety concerns [3,14,39,68]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.238, + 0.811, + 0.613, + 0.825 + ], + "angle": 0, + "content": "3 Homepage: https://github.com/snumprlab/realfred" + }, + { + "type": "page_footnote", + "bbox": [ + 0.238, + 0.825, + 0.538, + 0.84 + ], + "angle": 0, + "content": "*Equal contribution. †Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.442, + 0.128 + ], + "angle": 0, + "content": "T. Kim and C. Min et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.144, + 0.597, + 0.157 + ], + "angle": 0, + "content": "Goal: Put a cooled slice of banana on the coffee table" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.159, + 0.761, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.313, + 0.757, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.424, + 0.788, + 0.495 + ], + "angle": 0, + "content": "Fig. 1: Proposed REALFRED benchmark. The top image provides a perspective view of one of our scenes. The images below represent third-person views at each time step, along with their corresponding descriptions, for better understanding. The agent is required to understand instructions in natural language and then complete the desired tasks by navigating large 3D-captured environments and interacting with objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.788, + 0.791 + ], + "angle": 0, + "content": "As an alternative, several simulated environments have been introduced [59, 74] which leverage extensive 3D-captured environments obtained from real-world scenes [7, 56]. Compared to real-world deployment, these environments offer agents a faster process of taking actions and observing consequences, and the convenience of resetting the environment and trying again in case of failure, which enables agents to learn the skills to complete desired tasks. Adopting such simulated environments has produced remarkable advancements in various subtasks for embodied AI agents, including visual navigation [9, 47, 50, 55, 73], vision and language navigation [1, 37], and remote object grounding [54]. Due to the inherently static nature of these 3D-captured environments where objects (e.g., books, chairs, etc.) remain non-interactive, however, current benchmarks [1, 9] for these tasks have less focused on object interaction, which might hinder deployability for more complex tasks that require object interaction. Recent studies [44, 57] insert liftable objects in scanned environments for object interaction, but they support limited object interaction such as picking up objects, which might not provide enough deployability for more challenging real-world scenarios such as heating objects using a microwave or cooling objects using a refrigerator." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Meanwhile, virtual game engines such as Unity have been exploited to build object-interactable environments with graphically crafted assets, including walls, floors, ceilings, and objects. These object-interactable environments have led to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.639, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "REALFRED" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "notable progress in the handling of more intricate object-centric tasks, including rearrangement [67,72] and manipulation [25,26], beyond navigation-centric tasks. In particular, we have observed significant progress in the execution of more complex tasks by natural language instructions [5, 33, 46, 51, 61, 62]. However, object-interactable environments for training and evaluation of such agents often pose several issues, such as visual domain gaps [69] and relatively smaller room sizes compared to their counterparts in 3D-captured environments [7, 56, 66]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.255, + 0.518, + 0.542 + ], + "angle": 0, + "content": "To bridge the gap between limited object interactability and environmental sizes with visual domain discrepancy, we propose the REAL-WORLD ALFRED (REALFRED) benchmark that requires agents to complete long-horizon tasks by understanding free-form language instructions and interacting with objects in large 3D-captured environments, following a similar task setup to the widely used embodied instruction following benchmark, ALFRED [61]. The 3D-captured environments used in the REALFRED benchmark encompass multiple rooms, providing ample space for agents to engage in multiple rooms in a single episode, which adds a sense of realism to the tasks. This resembles real-world scenarios in which agents navigate seamlessly between mul" + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.259, + 0.778, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.467, + 0.788, + 0.537 + ], + "angle": 0, + "content": "Fig. 2: While other benchmarks [7, 12, 19, 37, 41, 43, 48, 49, 56, 57, 61, 67, 72, 74] provide one or two aspects, our proposed REALFRED benchmark addresses all of these aspects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.542, + 0.788, + 0.616 + ], + "angle": 0, + "content": "tiple rooms. Unlike prior benchmarks focusing primarily on single-room activities [61] or 3D-captured environments with limited object interaction [59], REALFRED provides task evaluation in wide, realistic, and interactive environments to mirror the natural expectations of human-robot interactions. Fig. 1. illustrates the household task in one of the scenes in ours." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.619, + 0.788, + 0.679 + ], + "angle": 0, + "content": "In our experiments, we observe that the models [5,33,46,61-63] proposed in synthetic environments [61] do not perform well in our REALFRED benchmark, implying that models developed for synthetic environments may not easily adapt to realistic environments." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.682, + 0.534, + 0.697 + ], + "angle": 0, + "content": "We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.715, + 0.787, + 0.746 + ], + "angle": 0, + "content": "- We propose REALFRED, a benchmark for embodied instruction following with 3D-captured multi-room environments and objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.748, + 0.787, + 0.793 + ], + "angle": 0, + "content": "- We collect 3D-captured scenes and objects to reduce the simulation-reality gap and free-form language instructions to support task completion based on agents' language understanding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "- We provide analyses on the recent state-of-the-art models in the literature and the relevant Sim2Real transfer to empirically validate the necessity of our REALFRED benchmark." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.715, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.442, + 0.128 + ], + "angle": 0, + "content": "T. Kim and C. Min et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.271 + ], + "angle": 0, + "content": "Table 1: Comparison of REALFRED and other Embodied AI benchmarks. 'Language' column denotes the number of human annotated language directives. 'Environment' column compares spatial characteristics and whether it supports interactivity. 'Inference' column denotes whether their action space includes interactive capability with objects. The REALFRED benchmark is the first benchmark that provides a 3D-captured and interactable environment to solve household tasks that require navigation and interaction at the same time directed by natural language commands. \\(^\\dagger\\)We count the number of dialogue sessions. \\(^\\ddagger\\)We count the number of annotations in English. *Though [16] environment supports interaction, interaction is not required." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.282, + 0.785, + 0.478 + ], + "angle": 0, + "content": "
LanguageEnvironmentInference
Human AnnotationsVisual QualityMulti-Room NavigationMovable ObjectsState ChangesObject Interactability
IQA [24]-SyntheticXX✓✓
ManipulaTHOR [19]; RoomR [72]-SyntheticX✓✓✓
RoboTHOR [16]-Synthetic✓✓✓X*
ProcTHOR [18]-Synthetic✓✓✓✓
ReplicaCAD [67]-Synthetic✓✓✓✓
BEHAVIOR-1K [41]; HSSD-200 [31]-Synthetic✓✓✓✓
OpenRooms [42]-PhotoXXXX
MP3D [7]; HM3D [56]; Gibson [74]-Photo✓XXX
Habitat-Web [57]-Photo✓✓✓✓
ALFRED [61]; TEACH [49]; CHAI [48]25k+; 2.0k+†;12k+SyntheticX✓✓✓
LANI [48]; Walk the Talk [43]28k+; 0.7k+Synthetic✓XXX
Virtualhome [52]2.7k+Synthetic✓XX✓
R2R [1]; RxR [38]; VLN-CE [37]21k+; 42k+‡; 21k+Photo✓XXX
REALFRED30k+Photo✓✓✓✓
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.505, + 0.389, + 0.523 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.613 + ], + "angle": 0, + "content": "Fig. 2 illustrates the comparison of our REALFRED with other benchmarks in three selected aspects. We first review 3D-captured environments and benchmarks that provide visual aesthetics similar to real-world environments. Then, we review simulation environments and benchmarks that support object interaction. Table 1 compares our REALFRED with other indoor datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.75 + ], + "angle": 0, + "content": "Datasets and benchmarks with 3D-captured environments. The collection of advanced 3D scans has played a key role in enhancing our understanding of 3D objects [15, 64, 76] and their perception [2, 30, 45]. While these datasets offer valuable insights for a deeper understanding of 3D environments, they lack object interaction for learning interactive embodied agents. To further promote research on embodied agents for real-world applications, training and evaluation of such agents with physical spaces from scanned data [7, 66, 74] have been proposed. They provide a rich source of data for researchers to explore the capabilities of agents operating within real-world-inspired scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In these environments, notable progress has been achieved, primarily in the realms of navigation and exploration. Extensive research has been conducted on agents capable of navigating complex 3D environments, as evidenced by work such as navigating to a specified object [9, 47, 55], navigating to a certain point [50, 73], and navigating to the shown image [36]. Similarly, exploring novel environments has also yielded valuable insights [8, 13]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.639, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "REALFRED" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.237 + ], + "angle": 0, + "content": "Additionally, there is a work to integrate multi-modal sensory information for developing an agent that can deeply understand environments using inputs such as vision and language [1,38] and audio and vision [10,11]. These advances enable agents to utilize a broader range of sensory data, enhancing their perception and interaction capabilities. Further research explores additional capabilities of embodied agents, such as object localization alongside navigation [54]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.785, + 0.328 + ], + "angle": 0, + "content": "However, a fundamental limitation of these scan-based environments remains their static nature, which restricts the potential for object interaction. This limitation has resulted in the majority of prior research primarily focusing on navigation tasks, leaving the broader scope of agent adaptability to more complex, interactive challenges largely unaddressed. Consequently, this limitation hampers the adaptability of agents to more complex tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.787, + 0.418 + ], + "angle": 0, + "content": "Object-interactable simulated environments. Scan-based environments [56, 76] are typically composed of static scene representations, focusing on the visual aspects of real-world spaces. However, they often fail to adequately support interactivity with objects, representing a discrepancy between the dynamic and interactive characteristics of the real world, necessitating the development of environments better aligned with the demands of interactive agent research." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.42, + 0.785, + 0.555 + ], + "angle": 0, + "content": "To address this gap, various simulators, such as AI2-THOR [35], ManipulaTHOR [19], RoboTHOR [16], VirtualHome [52], TDW [20], iGibson [40], and OmniGibson [41], have emerged as promising solutions. These environments are engineered with a primary focus on enabling interaction between agents and objects. They are built upon game engines, which provide a solid foundation for ensuring realistic interactivity within the virtual environment. On these simulators, researchers publish benchmarks [16, 21, 41, 49, 53, 61, 65, 75] that support interactions. These benchmarks have been promoting research in the field, leading to the development of robotic assistants capable of handling complex tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.557, + 0.785, + 0.751 + ], + "angle": 0, + "content": "However, the dataset that provide interactive environments are limited in size of a room because creating a large high-fidelity space is challenging [56]. Furthermore, despite their near-photo realism, agents would face a visual domain gap when deployed in real world environments [69,71,77]. To this end, Habitat-Web [57] proposes an interactive pick-and-place task based on the Habitat simulator [59,67]. Their template-based language instructions, however, might not be sufficient to express the complex nuances of human expression. Innovative methods for acquiring 3D scans from phone-captured layouts have been proposed for learning enviroment-specific policies [17]. Nevertheless, its synthetic assets may lead to visual domain gaps when deployed in the real world. On the contrary, our REALFRED supports photorealism, high interactivity with objects, and freeform language annotations. These features can offer a framework for developing language-driven agents with visual perception for complex household chores." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.777, + 0.542, + 0.793 + ], + "angle": 0, + "content": "3 The REALFRED Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "To develop agents capable of performing household tasks, substantial progress has been achieved in various domains, including navigation [1, 38], rearrange-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.444, + 0.128 + ], + "angle": 0, + "content": "T. Kim and C. Min et al." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.144, + 0.481, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.143, + 0.761, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.307, + 0.789, + 0.351 + ], + "angle": 0, + "content": "Fig. 3: Top-down view of 3D-captured environments. We provide two examples from our scanned indoor environments. White circles denote where scanners are deployed. By scanning scenes at diverse points, we can prevent including blind spots." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.376, + 0.789, + 0.435 + ], + "angle": 0, + "content": "ment [72], and manipulation tasks [25,26]. In particular, [61] recently introduced the ALFRED benchmark that requires agents to complete long-horizon household tasks by jointly understanding egocentric visual observations and natural language instructions in household environments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.437, + 0.79, + 0.542 + ], + "angle": 0, + "content": "However, these environments are restricted to a single room size compared to previously proposed 3D-captured environments [7, 56] consisting of multiple rooms, which could potentially restrict the deployability of agents to larger environments. Furthermore, the environments used in the ALFRED benchmark [61] are built with synthetic CAD assets and therefore could potentially yield visual aesthetics different from those obtained from real-world environments [69], which could eventually cause performance degradation due to visual domain gaps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.543, + 0.79, + 0.618 + ], + "angle": 0, + "content": "To address these issues, we extend the ALFRED benchmark [61] and propose a challenging benchmark, named the REALFRED benchmark, which requires agents to perform household tasks in large indoor environments captured in 3D with object interaction. For training and evaluation, we follow the same protocol as [61] to collect expert demonstrations in the captured large environments." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.637, + 0.599, + 0.654 + ], + "angle": 0, + "content": "3.1 Object-Interactable 3D-Captured Scenes" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.749 + ], + "angle": 0, + "content": "To reduce the visual domain gap, a straightforward approach is to use 3D scans of real world environments. However, the captured 3D scans (i.e., meshes) remain static and thus, agents cannot interact with objects in the captured scenes. For object interaction, we manually replace object parts with 3D object assets to support object interaction. For photorealism comparison with previous environments using FID [27] and KID [4] metrics, please refer to Sec A.3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.79, + 0.841 + ], + "angle": 0, + "content": "We detail the process of collecting object-interactable 3D-captured environments and highlight key differences from previously proposed benchmarks below. Data acquisition process. To collect 3D scans of real-world environments, we visit residential properties and employ scanners. Inspired by recent work [2], we collect scans outside of the US to add the diversity with public scanned environments. We utilize the same 3D scanner as [7], equipped with three RGB cameras" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.639, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "REALFRED" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.493, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.23, + 0.47, + 0.242 + ], + "angle": 0, + "content": "(a) Distribution of floor areas by scene." + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.145, + 0.765, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.502, + 0.23, + 0.76, + 0.242 + ], + "angle": 0, + "content": "(b) Distribution of navigable areas by scene." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.254, + 0.788, + 0.325 + ], + "angle": 0, + "content": "Fig. 4: Distribution of navigable and floor areas in interactive benchmarks by scenes. 'Floor area' denotes the overall size of the scene. 'Navigable area' denotes the size of the space in which the agent can actually navigate. For both metrics, the REALFRED benchmark poses a more even distribution and provides larger areas. For (a), we exclude RoboTHOR since it consists of single-sized floors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.355, + 0.788, + 0.461 + ], + "angle": 0, + "content": "along with a depth sensor, and capture images from three distinct perspectives: front-facing and slightly vertical above and below. Panoramic images are acquired through six consecutive captures with horizontal rotation from a fixed viewpoint. We scan each house with 2.5-meter intervals and address blind spots due to furniture with additional scans from different viewpoints. Fig. 3 shows houses where scanners were deployed while capturing scans. It can be discerned that captures were densely concentrated in areas with a high presence of objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.627 + ], + "angle": 0, + "content": "Object-interactable environments. To construct an environment with numerous interactable objects, separating each object should be preceded. In other words, objects in the scans are initially merged into the background and therefore they remain as the background. Constructing an environment where objects can be interacted with requires the separation of objects from the background and from each other. Therefore, we manually separate the 3D scans into background elements and interactive objects. Furthermore, each object can exist in various states. To visualize changes in an object's state, we add state-relevant textures on objects. For example, we add a stain texture to a clean object when it becomes dirty. Finally, we reconstruct these individual object meshes within the Unity editor, making them compatible with the AI2-THOR simulator [35]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.787, + 0.688 + ], + "angle": 0, + "content": "Comparison with previous benchmarks. To investigate the spatial characteristics of scenes in the REALFRED benchmark, we compare ours with other benchmarks that support interaction with objects [16, 61, 67]. We observe enhancements in our dataset in both: 1) spatial sizes and 2) spatial complexity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Spatial size. We compare the REALFRED benchmark with other benchmarks in terms of spatial size [56] by measuring 'Floor area' and 'Navigable area' and provide the result in Fig. 4. 'Floor area' represents the total spatial size \\((m^2)\\) of a scene, defined by the floor projection. 'Navigable area' measures the spatial size \\((m^2)\\) of the space in which an agent can actually navigate. 'Navigable area' is smaller than or equal to 'Floor area' since 'Navigable area' excludes areas where the agent collides with any components in the scene from 'Floor area.'" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We observe that the REALFRED benchmark yields a diverse distribution of floor areas, compared to previous work [61], with a larger average per scene. Furthermore, the REALFRED benchmark provides a broader range of navigable" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.442, + 0.128 + ], + "angle": 0, + "content": "T. Kim and C. Min et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.244 + ], + "angle": 0, + "content": "Table 2: Component sizes of interactive embodied AI benchmarks. Each 'Total floor area' and 'Total navigable area' (Total nav. area) denotes the sum of all floor areas and navigation areas. 'Nav. complex.' denotes navigational complexity. 'Scene clutter' measures the amount of clutter in the scene. We do not compare ReplicaCAD's Navigable area, Navigation complexity, and Scene clutter as agent sizes differ across simulators. The highest value for each metric is shown in bold. \\(^\\dagger\\)We count objects used as target objects in the ObjectNav task. \\(^\\ddagger\\)We count objects used in any of the tasks." + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.255, + 0.772, + 0.381 + ], + "angle": 0, + "content": "
SimulatorHabitat2.0AI2THOR
DatasetReplicaCAD [67]RoboTHOR [16]ALFRED [61]REALFRED
# Scenes11175120150
Total floor area (m2)8,824.52,5742,55510,060
Total nav. area (m2)-1,2581,3564,251
Nav. complex.-2.0362.5493.020
Scene clutter-8.0955.1198.072
# object class9214†82‡112‡
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.784, + 0.443 + ], + "angle": 0, + "content": "areas, with larger areas on average. This implies that an agent needs the ability to navigate effectively in spaces of varying sizes, generally wider on average." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.788, + 0.536 + ], + "angle": 0, + "content": "We also investigate the spatial characteristics of each benchmark [16, 61, 67] and summarize the result in Table 2. 'Total floor area' denotes the sum of all floor areas and 'Total navigable area' denotes the sum of all navigable areas in the dataset. We observe that the REALFRED benchmark has the highest total floor area and navigable area value, implying that REALFRED provides more navigation space for an agent than previous work [16, 61, 67]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.788, + 0.657 + ], + "angle": 0, + "content": "Spatial complexity. We now investigate the complexity of spatial structures in the REALFRED benchmark using several metrics. To ensure a fair comparison with other datasets [16, 61, 67], we employ the navigation complexity introduced by [74] and the scene clutter measurement from [56]. A higher navigation complexity indicates an increased difficulty in navigating through the space, while a higher scene clutter implies the presence of more obstacles in the environment. By utilizing these metrics, we conduct a comparative analysis with object-interactable benchmarks [16, 61] and provide the result in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We observe that the REALFRED benchmark provides the environment with higher navigation complexity and scene clutter compared to other benchmarks [61, 67]. The high navigation complexity in our scenes stems from their multi-room composition. This setup requires the agent to execute more intricate navigation when moving from one room to another, in comparison to scenarios where the agent operates within a single room. We observe that ours poses the second-highest scene clutter with a marginal gap from [16]. This is because the spaces in [16] are relatively confined with a high density of furniture. We observe that ours has a similar value to [16], meaning that our scenes have a large amount of obstacles with a similar portion of [16] since our scenes' average size is larger than that of [16]. This implies that the REALFRED benchmark provides a complex and challenging space for the agent to explore the environment." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.639, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "REALFRED" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image_caption", + "bbox": [ + 0.316, + 0.145, + 0.35, + 0.157 + ], + "angle": 0, + "content": "Train" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.167, + 0.391, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.434, + 0.146, + 0.501, + 0.157 + ], + "angle": 0, + "content": "Valid Seen" + }, + { + "type": "image", + "bbox": [ + 0.393, + 0.168, + 0.526, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.145, + 0.643, + 0.157 + ], + "angle": 0, + "content": "Valid Unseen" + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.162, + 0.661, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.161, + 0.771, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.313, + 0.788, + 0.342 + ], + "angle": 0, + "content": "Fig. 5: The seven types of tasks' distribution in REALFRED. We provide \\(37.6\\%\\) more tasks in valid sets and \\(19.3\\%\\) in total compared to previous benchmark [61]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.372, + 0.547, + 0.387 + ], + "angle": 0, + "content": "3.2 Expert Demonstration Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.785, + 0.474 + ], + "angle": 0, + "content": "Each expert demonstration includes a set of an egocentric RGB view and action information with an interaction mask if exists at each time step. Expert demonstrations for each task are generated by a planner [29] with encoded state spaces into Planning Domain Definition Language (PDDL) rules [22]. To generate household tasks, we utilize seven task types introduced in [61]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.518, + 0.596 + ], + "angle": 0, + "content": "Data splits. We split the generated demonstrations into train, validation, and test folds. Specifically, we designate 135 scenes for seen and 15 scenes for unseen fold. Comparison of the amount of each task and the distribution across training and validation folds with previous work [61] is shown in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.518, + 0.763 + ], + "angle": 0, + "content": "Curating free-form language instructions. Detailed language instructions describe a task that involves a sequence of actions, a point of interest in robotics. The REALFRED benchmark offers 30,696 language directives, each comprising a human-annotated high-level goal and a set of step-by-step instructions. These directives are collected from 93 Amazon Mechanical Turk workers with a 'Master' qualification, ensuring high" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.476, + 0.789, + 0.678 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.687, + 0.788, + 0.757 + ], + "angle": 0, + "content": "Fig.6:Word distribution.The words are arranged starting from the center and extending outwards. The arc length corresponds to the word frequency in the instructions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.763, + 0.788, + 0.839 + ], + "angle": 0, + "content": "quality. Collected annotations are validated through an additional voting survey, and invalid instructions are replaced with newly collected instructions. The distribution of language instructions by their first four words is presented in Fig. 6. We provide a detailed annotation process and examples of an expert demonstration with the instruction in Sec. A.1 and A.4 in the supplementary material." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.442, + 0.128 + ], + "angle": 0, + "content": "T. Kim and C. Min et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.516, + 0.357 + ], + "angle": 0, + "content": "Multi-room embodied instruction following. For household robots to be able to effectively assist humans, it would be more practical to deploy them in complex and diverse environments with various room types, rather than confining them to a single room. Cross-room navigation poses significant challenges, requiring agents to understand room references, plan efficient paths, and overcome obstacles proficiently. This agent is further tasked to interpret visual cues, demonstrate a keen sense of spatial awareness, and process natural language instructions." + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.147, + 0.788, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.279, + 0.788, + 0.35 + ], + "angle": 0, + "content": "Fig.7: Comparison of expert demonstrations (ALFRED vs. REALFRED). Due to expanded environments, longer steps and planning horizons are required in REALFRED." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.785, + 0.418 + ], + "angle": 0, + "content": "The distribution of the number of steps and the length covered by the expert demonstrations is shown in Fig. 7. We observe that longer steps and trajectories are required to complete our tasks compared to the single-room constraint benchmark [61], meaning that our benchmark provides longer-horizon tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.787, + 0.526 + ], + "angle": 0, + "content": "To broaden a range of tasks, we provide an extensive number of object class types in the REALFRED benchmark. More object classes are added to the REALFRED benchmark, making it a superset of [61] with 86 pickable and 26 receptacle objects. This has resulted in a more diverse set of tasks, with the number of unique tasks being \\(84.3\\%\\) more than the ALFRED benchmark [61], and the proportion of unique tasks being higher in the REALFRED benchmark. We provide detailed information in Sec. A in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.545, + 0.376, + 0.561 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.572, + 0.791, + 0.679 + ], + "angle": 0, + "content": "Metrics. We follow the same evaluation protocol of the ALFRED benchmark [61]. The primary metric is 'Success Rate (SR)' which measures the percentage of completed tasks. 'Goal-Condition Success Rate (GC)' measures the percentage of achieved goal conditions. We also use path-length-weighted metrics to measure how efficiently an agent completes tasks. For more details, kindly refer to [61]. Baselines. We evaluate several recent state-of-the-art methods [5,33,34,46,61-63] with competitive results in [61]. We provide more details in Sec. B." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.698, + 0.562, + 0.714 + ], + "angle": 0, + "content": "4.1 Comparison of the State of the Arts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.785, + 0.765 + ], + "angle": 0, + "content": "We evaluate the baselines in the proposed REALFRED benchmark over multiple runs and present the average result in Table 3. We report extended results with path-length-weighted metrics in Sec. C in supplementary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.787, + 0.84 + ], + "angle": 0, + "content": "For a fair comparison, we separate the baseline into two groups based on the use of additional depth supervision for semantic map reconstruction: 'Imitation Learning' where agents learn direct mapping from visual observations and language instructions to action sequences and 'Spatial Map Reconst.' where agents plan action sequences based on reconstructed semantic spatial representations." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.639, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "REALFRED" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.244 + ], + "angle": 0, + "content": "Table 3: Task and Goal-Condition Success Rate. We train and evaluate recent state-of-the-art methods on our REALFRED benchmark. For a fair comparison, we group these methods into two based on the usage of extra depth supervision: models learned by imitation learning without depth supervision ('Imitation Learning') and ones that maintain semantic spatial representations constructed by predicted depth maps ('Spatial Map Reconst.'). Path-length-weighted (PLW) metrics are reported in Sec. C for each value. \\(\\dagger\\)Authors' implementation as the code is not publicly available." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.255, + 0.785, + 0.374 + ], + "angle": 0, + "content": "
LearningModelValidationTest
SeenUnseenSeenUnseen
SRGCSRGCSRGCSRGC
Imitation LearningSeq2Seq [61]0.77 ± 0.066.93 ± 0.060.00 ± 0.004.03 ± 0.001.10 ± 0.006.60 ± 0.000.00 ± 0.003.50 ± 0.00
MOCA [62]12.64 ± 0.1220.95 ± 0.181.44 ± 0.056.76 ± 0.0414.11 ± 0.0322.84 ± 0.040.62 ± 0.085.14 ± 0.08
ABP† [33]24.71 ± 0.0533.80 ± 0.144.22 ± 0.0511.71 ± 0.2727.44 ± 0.4035.81 ± 0.233.54 ± 0.2310.57 ± 0.22
Spatial Map Reconst.HLSM [5]4.23 ± 0.089.14 ± 0.091.08 ± 0.146.12 ± 0.236.27 ± 0.0410.44 ± 0.130.49 ± 0.164.28 ± 0.13
FILM [46]7.08 ± 0.2811.93 ± 0.234.44 ± 0.179.26 ± 0.138.79 ± 0.0713.03 ± 0.082.15 ± 0.186.56 ± 0.15
LLM-Planner† [63]5.80 ± 0.1911.69 ± 0.353.33 ± 0.228.29 ± 0.198.16 ± 0.2013.20 ± 0.131.90 ± 0.136.33 ± 0.02
CAPEAM† [34]13.45 ± 0.0518.16 ± 0.274.92 ± 0.229.47 ± 0.2315.61 ± 0.1520.22 ± 0.112.87 ± 0.137.36 ± 0.07
Human------85.00 ± 3.5491.30 ± 2.94
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.789, + 0.538 + ], + "angle": 0, + "content": "We observe that all these baselines, 'Imitation Learning' and 'Spatial Map Reconst.' consistently achieve lower performance values for all metrics in both seen and unseen splits compared to performances achieved in [61], implying that our proposed REALFRED provide more challenges compared to [61]. While in [61], the 'Spatial Map Reconst.' baselines [5,34,46] outperform learning-based approaches [33,61,62] by exploiting semantic spatial representation with deterministic algorithms (e.g., obstacle-free navigation path planning [60]), we observe a contrasting result that the 'Imitation Learning' baselines outperforming the 'Spatial Map Reconst.' baselines in our REALFRED benchmark." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.54, + 0.789, + 0.706 + ], + "angle": 0, + "content": "We qualitatively observe that such a confined spatial map reconstruction is due to a limited field of view and map reconstruction methods. The agent's limited field of view often leads to failure in recognizing room corners with doors or narrow aisles, resulting in (single-room-sized) limited map reconstruction. In addition, [34, 46] perceives obstacles larger than they actually are for better obstacle-free path planning by sacrificing navigable area, but this can be quite critical for narrow doors and aisles as they have a small amount of navigable space. This may hinder navigation to other rooms and thus, fail at tasks. We provide a more detailed discussion, supported by figures, on the agents' difficulty in recognizing narrow passages and the resulting impact on the agents, who struggle with spatial map reconstruction in Sec. D in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.726, + 0.707, + 0.741 + ], + "angle": 0, + "content": "4.2 Comparison to the agents with sim-to-real adaptation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We investigate the transfer from simulation training to real-world scan evaluation (sim-to-real) and from real-world scan training to real-world scan evaluation (real-to-real) with [33]. We train a sim-to-real agent with synthetic visual data and a real-to-real agent with real scanned visual data, respectively. During inference, both agents predict an action and an object mask based on the scanned visual input frame and the given language instruction at every time step." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.442, + 0.128 + ], + "angle": 0, + "content": "T. Kim and C. Min et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "Training and evaluation data selection. For training the sim-to-real agent, we use the training dataset in [61], encompassing \\(21K\\) language annotations. For a fair comparison, we train the real-to-real agent with tasks from the REALFRED benchmark's training fold, specifically involving the manipulation of the same objects that were used in training the sim-to-real agent with \\(19K\\) language annotations. For evaluation, we select tasks from the valid unseen splits of our REALFRED benchmark. We evaluate the agent performance with the tasks that 1) do not require multi-room navigation 2) and those that do. This selection specifically includes tasks that feature objects used in the training phase." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.433 + ], + "angle": 0, + "content": "Real-to-sim domain adaptation. The use of Generative Adversarial Networks (GANs) [23,79] for domain adaptation has recently been examined in the literature on robotics [6,28,32,58,77]. These models are employed to adapt input images from real-world domains (i.e. Real \\(\\rightarrow\\) Sim) before they are passed to the agent policy. We train a CycleGan [79] and its off-the-shelf variant [70] with unpaired images collected from the REALFRED (real domain) and ALFRED benchmarks [61] (simulated domain). Among the trained generators, the one performing real to simulation conversion, referred to as the real-to-sim goggle, learns the mapping from the real domain to the simulation domain \\( G_{S}:\\mathcal{R}\\to \\mathcal{S} \\), where \\( \\mathcal{R} \\) denotes the real domain and \\( S \\) denotes the simulated domain." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.518, + 0.614 + ], + "angle": 0, + "content": "Results. We present the results of the sim-to-real experiments in Table 4. We report the performance of agents when tasks 1) require agents to navigate multirooms, denoted as 'Multi + Single,' and 2) are solvable within a single room, denoted as 'Single only.' Firstly, we observe that the sim-to-real agent significantly underperforms its real-to-real counterpart in all metrics \\((\\# (a)\\) vs. \\(\\# (d))\\) .We then compare the results with gobbled agents. By comparing agents evaluated in the Multi" + }, + { + "type": "table_caption", + "bbox": [ + 0.526, + 0.449, + 0.788, + 0.533 + ], + "angle": 0, + "content": "Table 4: Comparison to the agents with sim-to-real adaptation. Each 'Sim2Real' and 'Real2Real' denotes an agent trained in simulated environments and 3D-captured scenes. 'Goggle' denotes real-to-sim methods." + }, + { + "type": "table", + "bbox": [ + 0.53, + 0.535, + 0.788, + 0.599 + ], + "angle": 0, + "content": "
# SettingGoggleMulti + SingleSingle only
SRPLWSRSRPLWSR
(a) Sim2RealNone0.1150.0120.00.0
(b) Sim2RealCycleGan [79]0.1150.0160.3270.065
(c) Sim2RealUVCGAN-v2 [70]0.1150.0460.3270.187
(d) Real2RealNone2.4050.7852.6140.762
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.616, + 0.788, + 0.661 + ], + "angle": 0, + "content": "+ Single' tasks, we do not observe improvements in the main metric, success rate (SR). We observe a slight increase in the PLWSR metric for the gobbled sim-to-real agent compared to the vanilla sim-to-real agent (#(a) vs. #(b,c))." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.661, + 0.788, + 0.737 + ], + "angle": 0, + "content": "To isolate the impacts of the visual domain gap, we then compare agents evaluated on 'Single only' tasks. We observe that both gobbled agents show improvements over the vanilla sim-to-real agent which results in zero success rate. However, a noticeable performance gap exists with the real-to-real agent, implying the need for learning in real scanned environments \\((\\# (d)\\) vs. \\(\\# (a,b,c))\\)" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.756, + 0.488, + 0.772 + ], + "angle": 0, + "content": "4.3 Challenges in REALFRED" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We propose two hypotheses for the low performance observed with state-of-the-art methods on the REALFRED benchmark: (1) navigating within larger scenes and (2) overcoming narrow pathways between rooms. These elements represent challenges of completing tasks within multi-room, household-scale environments." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.639, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "REALFRED" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.144, + 0.782, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.284, + 0.788, + 0.368 + ], + "angle": 0, + "content": "Fig. 8: Distribution of collision spots. The heatmap illustrates collision frequency on each spot and the black outline depicts a navigable area. We illustrate collision points of [33] in valid unseen environments to represent the distribution. Specifically, we choose failure episodes from the two largest scenes in the valid unseen split. We observe a remarkable concentration of collisions at the entrance to the rooms. The corresponding egocentric views are presented on the right side of the figure." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.789, + 0.506 + ], + "angle": 0, + "content": "We analyze failure cases with the best-performing model, ABP [33], on valid unseen split, since it seems the most promising on the REALFRED benchmark. Difficulty in large environments. Understanding the surrounding environment, including the location of objects and the positions of obstacles, can be beneficial for task completion. However, the agent's visual range is bounded, requiring it to expend more steps for exploration in order to perceive larger spaces. This intensifies the challenge of navigating larger spaces with limited steps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.789, + 0.643 + ], + "angle": 0, + "content": "We conduct an analysis to see the relationship between the success rate (SR) and the size of the space. We follow the interquartile range (IQR) method to set a threshold at \\(30.44m^2\\) that covers the most navigable space sizes in the ALFRED [61]. This threshold marks the upper fence, defining spaces above it as outliers. We classify spaces in the REALFRED benchmark smaller than threshold as smaller scenes, and those larger as larger scenes. Results indicate that the agent [33] with the highest SR in the valid unseen fold showed an average SR of \\(5.46\\%\\) in the smaller scenes and only \\(1.77\\%\\) in the larger scenes. This implies that solving tasks becomes more challenging as the space size increases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.809 + ], + "angle": 0, + "content": "In addition, we compare the difficulty of navigation between the previous [61] and REALFRED benchmarks with different average spatial sizes. To quantify, we first define milestones for each task, which are spots to be reached to interact with target objects (e.g., Apple, Knife, etc.). We consider navigation to be a success only if all milestones are visited, regardless of whether the actual interaction (e.g., slicing an apple) is performed or not. Consequently, the navigation success rate in the REALFRED benchmark is merely \\(59.18\\%\\) for the valid unseen split while the agent's [33] navigation success rate is \\(84.82\\%\\) in ALFRED [61]. The REALFRED benchmark offering an average space size more than three times larger than the previous work [61] implies the need for model development capable of overcoming the challenges associated with this increased scale." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Navigation through narrow doorways. The REALFRED benchmark supports environments with multiroom composition, unlike the previous dataset [61]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.442, + 0.128 + ], + "angle": 0, + "content": "T. Kim and C. Min et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.192 + ], + "angle": 0, + "content": "providing a single room scale. Specifically, it contains narrow doorways across rooms, which may hinder the agent from navigating through. Here, we hypothesize that the agent would frequently collide with the walls near the door." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.787, + 0.313 + ], + "angle": 0, + "content": "We investigate collision spots and showcase two examples on top of the scene's layout in valid unseen fold in Fig. 8. The number of collisions in failed cases is accumulated and normalized to the maximum number of collisions, respectively, to indicate the collision frequency at each point. The black outline represents the agent's navigable area, as detailed in Sec. 3.1. Each value denotes the normalized collision frequency. In both scenes, we observe that collisions often occur on walls near the door, implying that our spatial characteristic (i.e., including narrow spaces) may hinder the agent from properly navigating without collision." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.787, + 0.374 + ], + "angle": 0, + "content": "Human evaluation. Following [61], we randomly select 100 directives from the test unseen fold and have them evaluated by humans. Five participants are given 20 tasks each, which they complete using a keyboard-and-mouse interface. Before starting, they are given the opportunity to become familiar with the interface." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.789, + 0.48 + ], + "angle": 0, + "content": "Participants achieve a comparable high success rate of \\(85\\%\\), along with a goal-condition success rate of \\(91.30\\%\\), in average. We agree that human performance may appear slightly lower compared to the results presented in previous work [57, 61, 78]. This is partly because human participants encountered several difficulties when controlling the agent, particularly in avoiding collisions within narrow corridors. Furthermore, navigating large spaces with a limited egocentric field of view introduced additional challenges, leading to task failures." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.505, + 0.36, + 0.521 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.787, + 0.644 + ], + "angle": 0, + "content": "We present REALFRED, a new dataset and benchmark for embodied instruction following task on 3D-captured environments. We capture 150 indoor houses in 3D with interactable objects to enable complex household tasks. The reconstructed indoor scenes provide a larger spatial area and complex multi-room environments that are close to the real-world scenario and challenging for an agent to successfully complete a task. Expert demonstrations are also provided along with free-form human-language instructions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.788, + 0.735 + ], + "angle": 0, + "content": "In our empirical evaluations, we show that state-of-the-art methods struggle in large multi-room environments, provide analyses of our newly proposed benchmark, and perform Sim2Real transfer experiments. We have released our Embodied AI research data and code for reproducibility. We expect that the REALFRED benchmark will encourage further research on developing robotic agents that execute household tasks by language instructions in the real world." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Limitation and future work. Although we support a large number of interactable objects, the types of tasks to be completed are rather limited, considering more complex real-world scenarios. In addition, we currently address natural language in English but users may come from different regions with different languages. We can think of two future research avenues as follows. (1) adding additional complicated types of task that require both hands to complete. (2) supporting a multi-lingual interface for users from different regions." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.639, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "REALFRED" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.268 + ], + "angle": 0, + "content": "This work was partly supported by the NRF grant (No.2022R1A2C400230012, \\(5\\%\\) ) and IITP grants (No.RS-2022-II220077 \\((5\\%)\\) , No.RS-2022-II220113 \\((5\\%)\\) No.RS-2022-II220959 \\((5\\%)\\) , No.RS-2022-II220871 \\((15\\%)\\) , No.RS-2020-II201361 \\((5\\%)\\) , Yonsei AI), No.RS-2021-II211343 \\((5\\%)\\) , SNU AI), No.RS-2021-II212068 \\((5\\%)\\) , AI Innov. Hub), No.RS-2022-II220951 \\((50\\%)\\)) funded by the Korea government(MSIT)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.289, + 0.323, + 0.305 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.785, + 0.36 + ], + "angle": 0, + "content": "1. Anderson, P., Wu, Q., Teney, D., Bruce, J., Johnson, M., Sünderhauf, N., Reid, I., Gould, S., van den Hengel, A.: Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.361, + 0.785, + 0.415 + ], + "angle": 0, + "content": "2. Baruch, G., Chen, Z., Dehghan, A., Dimry, T., Feigin, Y., Fu, P., Gebauer, T., Joffe, B., Kurz, D., Schwartz, A., Shulman, E.: ARKitscenes - a diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In: NeurIPS Datasets and Benchmarks Track (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.416, + 0.785, + 0.442 + ], + "angle": 0, + "content": "3. Berseth, G., Xie, C., Cernek, P., van de Panne, M.: Progressive reinforcement learning with distillation for multi-skilled motion control. In: ICLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.443, + 0.785, + 0.469 + ], + "angle": 0, + "content": "4. Binkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying mmd gans. In: ICLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.47, + 0.785, + 0.511 + ], + "angle": 0, + "content": "5. Blukis, V., Paxton, C., Fox, D., Garg, A., Artzi, Y.: A persistent spatial semantic representation for high-level natural language instruction execution. In: CoRL (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.511, + 0.785, + 0.552 + ], + "angle": 0, + "content": "6. Bousmalis, K., Irpan, A., Wohlhart, P., Bai, Y., Kelcey, M., Kalakrishnan, M., Downs, L., Ibarz, J., Pastor, P., Konolige, K., et al.: Using simulation and domain adaptation to improve efficiency of deep robotic grasping. In: ICRA (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.553, + 0.785, + 0.593 + ], + "angle": 0, + "content": "7. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3d: Learning from rgb-d data in indoor environments. arXiv:1709.06158 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.594, + 0.785, + 0.621 + ], + "angle": 0, + "content": "8. Chaplot, D.S., Gandhi, D., Gupta, S., Gupta, A., Salakhutdinov, R.: Learning to explore using active neural slam. In: ICLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.622, + 0.785, + 0.647 + ], + "angle": 0, + "content": "9. Chaplot, D.S., Gandhi, D.P., Gupta, A., Salakhutdinov, R.R.: Object goal navigation using goal-oriented semantic exploration. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.649, + 0.785, + 0.689 + ], + "angle": 0, + "content": "0. Chen, C., Jain, U., Schissler, C., Gari, S.V.A., Al-Halah, Z., Ithapu, V.K., Robinson, P., Grauman, K.: Soundspaces: Audio-visual navigation in 3d environments. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.69, + 0.785, + 0.731 + ], + "angle": 0, + "content": "1. Chen, C., Schissler, C., Garg, S., Kobernik, P., Clegg, A., Calamia, P., Batra, D., Robinson, P.W., Grauman, K.: Soundspaces 2.0: A simulation platform for visual-acoustic learning. In: NeurIPS Datasets and Benchmarks Track (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.785, + 0.758 + ], + "angle": 0, + "content": "2. Chen, H., Suhr, A., Misra, D., Snavely, N., Artzi, Y.: Touchdown: Natural language navigation and spatial reasoning in visual street environments. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.785 + ], + "angle": 0, + "content": "3. Chen, T., Gupta, S., Gupta, A.: Learning exploration policies for navigation. In: ICLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "4. Choi, S., Ji, G., Park, J., Kim, H., Mun, J., Lee, J.H., Hwangbo, J.: Learning quadrupedal locomotion on deformable terrain. Science Robotics (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "5. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.318, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.442, + 0.128 + ], + "angle": 0, + "content": "T. Kim and C. Min et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "16. Deitke, M., Han, W., Herrasti, A., Kembhavi, A., Kolve, E., Mottaghi, R., Salvador, J., Schwenk, D., VanderBilt, E., Wallingford, M., et al.: Robothor: An open simulation-to-real embodied ai platform. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "17. Deitke, M., Hendrix, R., Farhadi, A., Ehsani, K., Kembhavi, A.: Phone2proc: Bringing robust robots into our chaotic world. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "18. Deitke, M., VanderBilt, E., Herrasti, A., Weihs, L., Ehsani, K., Salvador, J., Han, W., Kolve, E., Kembhavi, A., Mottaghi, R.: Procthor: Large-scale embodied ai using procedural generation. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.259, + 0.785, + 0.3 + ], + "angle": 0, + "content": "19. Ehsani, K., Han, W., Herrasti, A., Vanderbilt, E., Weihs, L., Kolve, E., Kembhavi, A., Mottaghi, R.: Manipulathor: A framework for visual object manipulation. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.3, + 0.785, + 0.357 + ], + "angle": 0, + "content": "20. Gan, C., Schwartz, J., Alter, S., Mrowca, D., Schrimpf, M., Traer, J., De Freitas, J., Kubilius, J., Bhandwaldar, A., Haber, N., et al.: Threadworld: A platform for interactive multi-modal physical simulation. In: NeurIPS Datasets and Benchmarks Track (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.384 + ], + "angle": 0, + "content": "21. Gao, X., Gao, Q., Gong, R., Lin, K., Thattai, G., Sukhatme, G.S.: Dialfred: Dialogue-enabled agents for embodied instruction following. RA-L (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.412 + ], + "angle": 0, + "content": "22. Ghallab, M., Howe, A., Knoblock, C., McDermott, D., Ram, A., Veloso, M., Weld, D., Wilkins, D.: Pddl the planning domain definition language (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.439 + ], + "angle": 0, + "content": "23. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: NeurIPS (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.467 + ], + "angle": 0, + "content": "24. Gordon, D., Kembhavi, A., Rastegari, M., Redmon, J., Fox, D., Farhadi, A.: Iqa: Visual question answering in interactive environments. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "25. Gu, J., Xiang, F., Li, X., Ling, Z., Liu, X., Mu, T., Tang, Y., Tao, S., Wei, X., Yao, Y., et al.: Maniskill2: A unified benchmark for generalizable manipulation skills. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.785, + 0.536 + ], + "angle": 0, + "content": "26. Heo, M., Lee, Y., Lee, D., Lim, J.J.: Furniturebench: Reproducible real-world benchmark for long-horizon complex manipulation. In: RSS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "27. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.577, + 0.785, + 0.605 + ], + "angle": 0, + "content": "28. Ho, D., Rao, K., Xu, Z., Jang, E., Khansari, M., Bai, Y.: Retinagan: An object-aware approach to sim-to-real transfer. In: ICRA (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.605, + 0.785, + 0.633 + ], + "angle": 0, + "content": "29. Hoffmann, J., Nebel, B.: The ff planning system: Fast plan generation through heuristic search. In: JAIR (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.661 + ], + "angle": 0, + "content": "30. Hua, B.S., Pham, Q.H., Nguyen, D.T., Tran, M.K., Yu, L.F., Yeung, S.K.: Scenenn: A scene meshes dataset with annotations. In: 3DV (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.716 + ], + "angle": 0, + "content": "31. Khanna, M., Mao, Y., Jiang, H., Haresh, S., Shacklett, B., Batra, D., Clegg, A., Undersander, E., Chang, A.X., Savva, M.: Habitat synthetic scenes dataset (hssd200): An analysis of 3d scene scale and realism tradeoffs for objectgoal navigation. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.758 + ], + "angle": 0, + "content": "32. Khansari, M., Ho, D., Du, Y., Fuentes, A., Bennice, M., Sievers, N., Kirmani, S., Bai, Y., Jang, E.: Practical visual deep imitation learning via task-level domain consistency. In: ICRA (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "33. Kim, B., Bhambri, S., Singh, K.P., Mottaghi, R., Choi, J.: Agent with the big picture: Perceiving surroundings for interactive instruction following. In: Embodied AI Workshop @ CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "34. Kim, B., Kim, J., Kim, Y., Min, C., Choi, J.: Context-aware planning and environment-aware memory for instruction following embodied agents. In: ICCV (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.639, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "REALFRED" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "35. Kolve, E., Mottaghi, R., Han, W., VanderBilt, E., Weihs, L., Herrasti, A., Gordon, D., Zhu, Y., Gupta, A., Farhadi, A.: Ai2-thor: An interactive 3d environment for visual ai. arXiv:1712.05474 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "36. Krantz, J., Gervet, T., Yadav, K., Wang, A., Paxton, C., Mottaghi, R., Batra, D., Malik, J., Lee, S., Chaplot, D.S.: Navigating to objects specified by images. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.233, + 0.786, + 0.261 + ], + "angle": 0, + "content": "37. Krantz, J., Wijmans, E., Majumdar, A., Batra, D., Lee, S.: Beyond the nav-graph: Vision-and-language navigation in continuous environments. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.262, + 0.787, + 0.302 + ], + "angle": 0, + "content": "38. Ku, A., Anderson, P., Patel, R., Ie, E., Baldridge, J.: Room-Across-Room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. In: EMNLP (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.787, + 0.331 + ], + "angle": 0, + "content": "39. Kumar, A., Fu, Z., Pathak, D., Malik, J.: Rapid motor adaptation for legged robots. In: RSS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.333, + 0.787, + 0.387 + ], + "angle": 0, + "content": "40. Li, C., Xia, F., Martin-Martin, R., Lingelbach, M., Srivastava, S., Shen, B., Vainio, K.E., Gokmen, C., Dharan, G., Jain, T., Kurenkov, A., Liu, K., Gweon, H., Wu, J., Fei-Fei, L., Savarese, S.: igibson 2.0: Object-centric simulation for robot learning of everyday household tasks. In: CoRL (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.389, + 0.787, + 0.43 + ], + "angle": 0, + "content": "41. Li, C., Zhang, R., Wong, J., Gokmen, C., Srivastava, S., Martin-Martin, R., Wang, C., Levine, G., Lingelbach, M., Sun, J., et al.: Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation. In: CoRL (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.431, + 0.787, + 0.486 + ], + "angle": 0, + "content": "42. Li, Z., Yu, T.W., Sang, S., Wang, S., Song, M., Liu, Y., Yeh, Y.Y., Zhu, R., Gundavarapu, N., Shi, J., Bi, S., Yu, H.X., Xu, Z., Sunkavalli, K., Hasan, M., Ramamoorthi, R., Chandraker, M.: Open rooms: An open framework for photorealistic indoor scene datasets. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.488, + 0.787, + 0.514 + ], + "angle": 0, + "content": "43. MacMahon, M., Stankiewicz, B., Kuipers, B.: Walk the talk: Connecting language, knowledge, and action in route instructions. In: AAAI (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.516, + 0.787, + 0.543 + ], + "angle": 0, + "content": "44. Majumdar, et al.: Findthis: Language-driven object disambiguation in indoor environments. In: CoRL (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.545, + 0.787, + 0.571 + ], + "angle": 0, + "content": "45. Mao, Y., Zhang, Y., Jiang, H., Chang, A., Savva, M.: Multiscan: Scalable rgbd scanning for 3d environments with articulated objects. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.573, + 0.787, + 0.599 + ], + "angle": 0, + "content": "46. Min, S.Y., Chaplot, D.S., Ravikumar, P., Bisk, Y., Salakhutdinov, R.: Film: Following instructions in language with modular methods. In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.601, + 0.787, + 0.628 + ], + "angle": 0, + "content": "47. Min, S.Y., Tsai, Y.H.H., Ding, W., Farhadi, A., Salakhutdinov, R., Bisk, Y., Zhang, J.: Object goal navigation with end-to-end self-supervision. In: IROS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.63, + 0.787, + 0.67 + ], + "angle": 0, + "content": "48. Misra, D., Bennett, A., Blukis, V., Niklasson, E., Shatkhin, M., Artzi, Y.: Mapping instructions to actions in 3d environments with visual goal prediction. In: EMNLP (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.672, + 0.787, + 0.712 + ], + "angle": 0, + "content": "49. Padmakumar, A., Thomason, J., Shrivastava, A., Lange, P., Narayan-Chen, A., Gella, S., Piramuthu, R., Tur, G., Hakkani-Tur, D.: Teach: Task-driven embodied agents that chat. In: AAAI (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.715, + 0.787, + 0.741 + ], + "angle": 0, + "content": "50. Partsey, R., Wijmans, E., Yokoyama, N., Dobosevych, O., Batra, D., Maksymets, O.: Is mapping necessary for realistic pointgoal navigation? In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.743, + 0.787, + 0.77 + ], + "angle": 0, + "content": "51. Pashevich, A., Schmid, C., Sun, Chen: Episodic transformer for vision-and-language navigation. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.771, + 0.787, + 0.798 + ], + "angle": 0, + "content": "52. Puig, X., Ra, K., Boben, M., Li, J., Wang, T., Fidler, S., Torralba, A.: Virtualhome: Simulating household activities via programs. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "53. Puig, X., Undersander, E., Szot, A., Cote, M.D., Yang, T.Y., Partsey, R., Desai, R., Clegg, A.W., Hlavac, M., Min, S.Y., et al.: Habitat 3.0: A co-habitat for humans, avatars and robots. In: ICLR (2024)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.442, + 0.128 + ], + "angle": 0, + "content": "T. Kim and C. Min et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "54. Qi, Y., Wu, Q., Anderson, P., Wang, X., Wang, W.Y., Shen, C., Hengel, A.v.d.: Reverie: Remote embodied visual referring expression in real indoor environments. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "55. Ramakrishnan, S.K., Chaplot, D.S., Al-Halah, Z., Malik, J., Grauman, K.: Poni: Potential functions for objectgoal navigation with interaction-free learning. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.785, + 0.302 + ], + "angle": 0, + "content": "56. Ramakrishnan, S.K., Gokaslan, A., Wijmans, E., Maksymets, O., Clegg, A., Turner, J.M., Undersander, E., Galuba, W., Westbury, A., Chang, A.X., Savva, M., Zhao, Y., Batra, D.: Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI. In: NeurIPS Datasets and Benchmarks Track (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.303, + 0.785, + 0.345 + ], + "angle": 0, + "content": "57. Ramrakhya, R., Undersander, E., Batra, D., Das, A.: Habitat-web: Learning embodied object-search strategies from human demonstrations at scale. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.785, + 0.373 + ], + "angle": 0, + "content": "58. Rao, K., Harris, C., Irpan, A., Levine, S., Ibarz, J., Khansari, M.: Rl-cyclegan: Reinforcement learning aware simulation-to-real. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.374, + 0.785, + 0.416 + ], + "angle": 0, + "content": "59. Savva, M., Kadian, A., Maksymets, O., Zhao, Y., Wijmans, E., Jain, B., Straub, J., Liu, J., Koltun, V., Malik, J., Parikh, D., Batra, D.: Habitat: A Platform for Embodied AI Research. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.417, + 0.785, + 0.445 + ], + "angle": 0, + "content": "60. Sethian, J.A.: A fast marching level set method for monotonically advancing fronts. In: PNAS (1996)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.446, + 0.785, + 0.487 + ], + "angle": 0, + "content": "61. Shridhar, M., Thomason, J., Gordon, D., Bisk, Y., Han, W., Mottaghi, R., Zettle-moyer, L., Fox, D.: Alfred: A benchmark for interpreting grounded instructions for everyday tasks. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.516 + ], + "angle": 0, + "content": "62. Singh, K.P., Bhambri, S., Kim, B., Mottaghi, R., Choi, J.: Factorizing perception and policy for interactive instruction following. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.517, + 0.785, + 0.558 + ], + "angle": 0, + "content": "63. Song, C.H., Wu, J., Washington, C., Sadler, B.M., Chao, W.L., Su, Y.: Llm-planner: Few-shot grounded planning for embodied agents with large language models. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.785, + 0.586 + ], + "angle": 0, + "content": "64. Song, S., Lichtenberg, S.P., Xiao, J.: Sun rgb-d: A rgb-d scene understanding benchmark suite. In: CVPR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.588, + 0.785, + 0.643 + ], + "angle": 0, + "content": "65. Srivastava, S., Li, C., Lingelbach, M., Martin-Martin, R., Xia, F., Vainio, K.E., Lian, Z., Gokmen, C., Buch, S., Liu, K., et al.: Behavior: Benchmark for everyday household activities in virtual, interactive, and ecological environments. In: CoRL (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.644, + 0.785, + 0.685 + ], + "angle": 0, + "content": "66. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv:1906.05797 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.785, + 0.755 + ], + "angle": 0, + "content": "67. Szot, A., Clegg, A., Undersander, E., Wijmans, E., Zhao, Y., Turner, J., Maestre, N., Mukadam, M., Chaplot, D., Maksymets, O., Gokaslan, A., Vondrus, V., Dharur, S., Meier, F., Galuba, W., Chang, A., Kira, Z., Koltun, V., Malik, J., Savva, M., Batra, D.: Habitat 2.0: Training home assistants to rearrange their habitat. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "68. Tan, J., Zhang, T., Coumans, E., Iscen, A., Bai, Y., Hafner, D., Bohez, S., Vanhoucke, V.: Sim-to-real: Learning agile locomotion for quadruped robots. In: RSS (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "69. Tobin, J., Fong, R., Ray, A., Schneider, J., Zaremba, W., Abbeel, P.: Domain randomization for transferring deep neural networks from simulation to the real world. In: IROS (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.639, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "REALFRED" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "70. Torbunov, D., Huang, Y., Tseng, H.H., Yu, H., Huang, J., Yoo, S., Lin, M., Viren, B., Ren, Y.: Uvçgan v2: An improved cycle-consistent gan for unpaired image-to-image translation. arXiv:2303.16280 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.217 + ], + "angle": 0, + "content": "71. Truong, J., Chernova, S., Batra, D.: Bi-directional domain adaptation for sim2real transfer of embodied navigation agents. RA-L (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.786, + 0.244 + ], + "angle": 0, + "content": "72. Weihs, L., Deitke, M., Kembhavi, A., Mottaghi, R.: Visual room rearrangement. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.787, + 0.286 + ], + "angle": 0, + "content": "73. Wijmans, E., Kadian, A., Morcos, A., Lee, S., Essa, I., Parikh, D., Savva, M., Batra, D.: Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. In: ICLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.314 + ], + "angle": 0, + "content": "74. Xia, F., Zamir, A.R., He, Z., Sax, A., Malik, J., Savarese, S.: Gibson env: Realworld perception for embodied agents. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "75. Yenamandra, S., Ramachandran, A., Yadav, K., Wang, A., Khanna, M., Gervet, T., Yang, T.Y., Jain, V., Clegg, A.W., Turner, J., et al.: Homerobot: Open-vocabulary mobile manipulation. In: CoRL (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.787, + 0.383 + ], + "angle": 0, + "content": "76. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.411 + ], + "angle": 0, + "content": "77. Zhang, J., Tai, L., Yun, P., Xiong, Y., Liu, M., Boedecker, J., Burgard, W.: Vr-goggles for robots: Real-to-sim domain adaptation for visual control. RA-L (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.787, + 0.452 + ], + "angle": 0, + "content": "78. Zhu, H., Kapoor, R., Min, S.Y., Han, W., Li, J., Geng, K., Neubig, G., Bisk, Y., Kembhavi, A., Weihs, L.: Excalibur: Encouraging and evaluating embodied exploration. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.481 + ], + "angle": 0, + "content": "79. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: ICCV (2017)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.481 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/ReALFRED_ An Embodied Instruction Following Benchmark in Photo-Realistic Environments/0e4c20a8-1ef4-4401-b5cc-41d9bc314f7b_origin.pdf b/2024/ReALFRED_ An Embodied Instruction Following Benchmark in Photo-Realistic Environments/0e4c20a8-1ef4-4401-b5cc-41d9bc314f7b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5acee5c33b42d79caabd9531604c398bd54cd382 --- /dev/null +++ b/2024/ReALFRED_ An Embodied Instruction Following Benchmark in Photo-Realistic Environments/0e4c20a8-1ef4-4401-b5cc-41d9bc314f7b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:696907b2dac8feedd196e9de957a90e90119798a87dead04c4914c77243638b0 +size 4005370 diff --git a/2024/ReALFRED_ An Embodied Instruction Following Benchmark in Photo-Realistic Environments/full.md b/2024/ReALFRED_ An Embodied Instruction Following Benchmark in Photo-Realistic Environments/full.md new file mode 100644 index 0000000000000000000000000000000000000000..75789efeb16cc32ddc8ebf478ca4822be46852d6 --- /dev/null +++ b/2024/ReALFRED_ An Embodied Instruction Following Benchmark in Photo-Realistic Environments/full.md @@ -0,0 +1,307 @@ +# REALFRED: An Embodied Instruction Following Benchmark in Photo-Realistic Environments + +Taewoong Kim $^{1,*}$ , Cheolhong Min $^{1,*}$ , Byeonghwi Kim $^{1}$ , Jinyeon Kim $^{1,2}$ , Wonje Jeung $^{1}$ , and Jonghyun Choi $^{1,\dagger}$ + +$^{1}$ Seoul National University + +2 Yonsei University + +{twoongg.kim, cheolhong.min, byeonghwikim}@snu.ac.kr, jinyeonkim@yonsei.ac.kr, {specific0924,jonghyunchoi}@snu.ac.kr + +Abstract. Simulated virtual environments have been widely used to learn robotic agents that perform daily household tasks. These environments encourage research progress by far, but often provide limited object interactability, visual appearance different from real-world environments, or relatively smaller environment sizes. This prevents the learned models in the virtual scenes from being readily deployable. To bridge the gap between these learning environments and deploying (i.e., real) environments, we propose the REALFRED benchmark that employs real-world scenes, objects, and room layouts to learn agents to complete household tasks by understanding free-form language instructions and interacting with objects in large, multi-room and 3D-captured scenes. Specifically, we extend the ALFRED benchmark with updates for larger environmental spaces with smaller visual domain gaps. With REALFRED, we analyze previously crafted methods for the ALFRED benchmark and observe that they consistently yield lower performance in all metrics, encouraging the community to develop methods in more realistic environments. Our code and data are publicly available3. + +Keywords: Interactive Scanned Environments $\cdot$ Instruction Following $\cdot$ Embodied AI $\cdot$ Reality Gap $\cdot$ Dataset and Benchmark + +# 1 Introduction + +Building autonomous robotic assistants that can perform everyday household tasks has been an elusive aspiration within the research community for decades. To let them learn these intricate tasks, we may provide them with interactive environments where agents can learn task completion skills with numerous interactions with environments. A straightforward approach to train such agents that can carry out real-world activities is to directly deploy robots in real-world environments and let them learn to complete desired tasks. However, this often faces several practical challenges, including cost, time, or safety concerns [3,14,39,68]. + +![](images/3402938b11df9e6b93447085c34b326633deb8fd70164f71dcb53e1f2184e0d2.jpg) +Goal: Put a cooled slice of banana on the coffee table + +![](images/e0dd9640bb81a4c70c70f16b531072f75286b4cbb3436e8658bb962114a7e7db.jpg) +Fig. 1: Proposed REALFRED benchmark. The top image provides a perspective view of one of our scenes. The images below represent third-person views at each time step, along with their corresponding descriptions, for better understanding. The agent is required to understand instructions in natural language and then complete the desired tasks by navigating large 3D-captured environments and interacting with objects. + +As an alternative, several simulated environments have been introduced [59, 74] which leverage extensive 3D-captured environments obtained from real-world scenes [7, 56]. Compared to real-world deployment, these environments offer agents a faster process of taking actions and observing consequences, and the convenience of resetting the environment and trying again in case of failure, which enables agents to learn the skills to complete desired tasks. Adopting such simulated environments has produced remarkable advancements in various subtasks for embodied AI agents, including visual navigation [9, 47, 50, 55, 73], vision and language navigation [1, 37], and remote object grounding [54]. Due to the inherently static nature of these 3D-captured environments where objects (e.g., books, chairs, etc.) remain non-interactive, however, current benchmarks [1, 9] for these tasks have less focused on object interaction, which might hinder deployability for more complex tasks that require object interaction. Recent studies [44, 57] insert liftable objects in scanned environments for object interaction, but they support limited object interaction such as picking up objects, which might not provide enough deployability for more challenging real-world scenarios such as heating objects using a microwave or cooling objects using a refrigerator. + +Meanwhile, virtual game engines such as Unity have been exploited to build object-interactable environments with graphically crafted assets, including walls, floors, ceilings, and objects. These object-interactable environments have led to + +notable progress in the handling of more intricate object-centric tasks, including rearrangement [67,72] and manipulation [25,26], beyond navigation-centric tasks. In particular, we have observed significant progress in the execution of more complex tasks by natural language instructions [5, 33, 46, 51, 61, 62]. However, object-interactable environments for training and evaluation of such agents often pose several issues, such as visual domain gaps [69] and relatively smaller room sizes compared to their counterparts in 3D-captured environments [7, 56, 66]. + +To bridge the gap between limited object interactability and environmental sizes with visual domain discrepancy, we propose the REAL-WORLD ALFRED (REALFRED) benchmark that requires agents to complete long-horizon tasks by understanding free-form language instructions and interacting with objects in large 3D-captured environments, following a similar task setup to the widely used embodied instruction following benchmark, ALFRED [61]. The 3D-captured environments used in the REALFRED benchmark encompass multiple rooms, providing ample space for agents to engage in multiple rooms in a single episode, which adds a sense of realism to the tasks. This resembles real-world scenarios in which agents navigate seamlessly between mul + +![](images/15b4cd8f949c079b9155588ab4edbed0bf4ea81544184566b97f81cb7455f248.jpg) +Fig. 2: While other benchmarks [7, 12, 19, 37, 41, 43, 48, 49, 56, 57, 61, 67, 72, 74] provide one or two aspects, our proposed REALFRED benchmark addresses all of these aspects. + +tiple rooms. Unlike prior benchmarks focusing primarily on single-room activities [61] or 3D-captured environments with limited object interaction [59], REALFRED provides task evaluation in wide, realistic, and interactive environments to mirror the natural expectations of human-robot interactions. Fig. 1. illustrates the household task in one of the scenes in ours. + +In our experiments, we observe that the models [5,33,46,61-63] proposed in synthetic environments [61] do not perform well in our REALFRED benchmark, implying that models developed for synthetic environments may not easily adapt to realistic environments. + +We summarize our contributions as follows: + +- We propose REALFRED, a benchmark for embodied instruction following with 3D-captured multi-room environments and objects. +- We collect 3D-captured scenes and objects to reduce the simulation-reality gap and free-form language instructions to support task completion based on agents' language understanding. +- We provide analyses on the recent state-of-the-art models in the literature and the relevant Sim2Real transfer to empirically validate the necessity of our REALFRED benchmark. + +Table 1: Comparison of REALFRED and other Embodied AI benchmarks. 'Language' column denotes the number of human annotated language directives. 'Environment' column compares spatial characteristics and whether it supports interactivity. 'Inference' column denotes whether their action space includes interactive capability with objects. The REALFRED benchmark is the first benchmark that provides a 3D-captured and interactable environment to solve household tasks that require navigation and interaction at the same time directed by natural language commands. $^\dagger$ We count the number of dialogue sessions. $^\ddagger$ We count the number of annotations in English. *Though [16] environment supports interaction, interaction is not required. + +
LanguageEnvironmentInference
Human AnnotationsVisual QualityMulti-Room NavigationMovable ObjectsState ChangesObject Interactability
IQA [24]-SyntheticXX✓✓
ManipulaTHOR [19]; RoomR [72]-SyntheticX✓✓✓
RoboTHOR [16]-Synthetic✓✓✓X*
ProcTHOR [18]-Synthetic✓✓✓✓
ReplicaCAD [67]-Synthetic✓✓✓✓
BEHAVIOR-1K [41]; HSSD-200 [31]-Synthetic✓✓✓✓
OpenRooms [42]-PhotoXXXX
MP3D [7]; HM3D [56]; Gibson [74]-Photo✓XXX
Habitat-Web [57]-Photo✓✓✓✓
ALFRED [61]; TEACH [49]; CHAI [48]25k+; 2.0k+†;12k+SyntheticX✓✓✓
LANI [48]; Walk the Talk [43]28k+; 0.7k+Synthetic✓XXX
Virtualhome [52]2.7k+Synthetic✓XX✓
R2R [1]; RxR [38]; VLN-CE [37]21k+; 42k+‡; 21k+Photo✓XXX
REALFRED30k+Photo✓✓✓✓
+ +# 2 Related Work + +Fig. 2 illustrates the comparison of our REALFRED with other benchmarks in three selected aspects. We first review 3D-captured environments and benchmarks that provide visual aesthetics similar to real-world environments. Then, we review simulation environments and benchmarks that support object interaction. Table 1 compares our REALFRED with other indoor datasets. + +Datasets and benchmarks with 3D-captured environments. The collection of advanced 3D scans has played a key role in enhancing our understanding of 3D objects [15, 64, 76] and their perception [2, 30, 45]. While these datasets offer valuable insights for a deeper understanding of 3D environments, they lack object interaction for learning interactive embodied agents. To further promote research on embodied agents for real-world applications, training and evaluation of such agents with physical spaces from scanned data [7, 66, 74] have been proposed. They provide a rich source of data for researchers to explore the capabilities of agents operating within real-world-inspired scenarios. + +In these environments, notable progress has been achieved, primarily in the realms of navigation and exploration. Extensive research has been conducted on agents capable of navigating complex 3D environments, as evidenced by work such as navigating to a specified object [9, 47, 55], navigating to a certain point [50, 73], and navigating to the shown image [36]. Similarly, exploring novel environments has also yielded valuable insights [8, 13]. + +Additionally, there is a work to integrate multi-modal sensory information for developing an agent that can deeply understand environments using inputs such as vision and language [1,38] and audio and vision [10,11]. These advances enable agents to utilize a broader range of sensory data, enhancing their perception and interaction capabilities. Further research explores additional capabilities of embodied agents, such as object localization alongside navigation [54]. + +However, a fundamental limitation of these scan-based environments remains their static nature, which restricts the potential for object interaction. This limitation has resulted in the majority of prior research primarily focusing on navigation tasks, leaving the broader scope of agent adaptability to more complex, interactive challenges largely unaddressed. Consequently, this limitation hampers the adaptability of agents to more complex tasks. + +Object-interactable simulated environments. Scan-based environments [56, 76] are typically composed of static scene representations, focusing on the visual aspects of real-world spaces. However, they often fail to adequately support interactivity with objects, representing a discrepancy between the dynamic and interactive characteristics of the real world, necessitating the development of environments better aligned with the demands of interactive agent research. + +To address this gap, various simulators, such as AI2-THOR [35], ManipulaTHOR [19], RoboTHOR [16], VirtualHome [52], TDW [20], iGibson [40], and OmniGibson [41], have emerged as promising solutions. These environments are engineered with a primary focus on enabling interaction between agents and objects. They are built upon game engines, which provide a solid foundation for ensuring realistic interactivity within the virtual environment. On these simulators, researchers publish benchmarks [16, 21, 41, 49, 53, 61, 65, 75] that support interactions. These benchmarks have been promoting research in the field, leading to the development of robotic assistants capable of handling complex tasks. + +However, the dataset that provide interactive environments are limited in size of a room because creating a large high-fidelity space is challenging [56]. Furthermore, despite their near-photo realism, agents would face a visual domain gap when deployed in real world environments [69,71,77]. To this end, Habitat-Web [57] proposes an interactive pick-and-place task based on the Habitat simulator [59,67]. Their template-based language instructions, however, might not be sufficient to express the complex nuances of human expression. Innovative methods for acquiring 3D scans from phone-captured layouts have been proposed for learning enviroment-specific policies [17]. Nevertheless, its synthetic assets may lead to visual domain gaps when deployed in the real world. On the contrary, our REALFRED supports photorealism, high interactivity with objects, and freeform language annotations. These features can offer a framework for developing language-driven agents with visual perception for complex household chores. + +# 3 The REALFRED Benchmark + +To develop agents capable of performing household tasks, substantial progress has been achieved in various domains, including navigation [1, 38], rearrange- + +![](images/7dbd702b14f1b095ea5db2c1acc45bbe4d7cac0956e9dbce94033e25598cc581.jpg) +Fig. 3: Top-down view of 3D-captured environments. We provide two examples from our scanned indoor environments. White circles denote where scanners are deployed. By scanning scenes at diverse points, we can prevent including blind spots. + +![](images/5c4e1ce4af62ed6466e4f9502c01e5596a57e81adffb45c3db7f7372d41057b7.jpg) + +ment [72], and manipulation tasks [25,26]. In particular, [61] recently introduced the ALFRED benchmark that requires agents to complete long-horizon household tasks by jointly understanding egocentric visual observations and natural language instructions in household environments. + +However, these environments are restricted to a single room size compared to previously proposed 3D-captured environments [7, 56] consisting of multiple rooms, which could potentially restrict the deployability of agents to larger environments. Furthermore, the environments used in the ALFRED benchmark [61] are built with synthetic CAD assets and therefore could potentially yield visual aesthetics different from those obtained from real-world environments [69], which could eventually cause performance degradation due to visual domain gaps. + +To address these issues, we extend the ALFRED benchmark [61] and propose a challenging benchmark, named the REALFRED benchmark, which requires agents to perform household tasks in large indoor environments captured in 3D with object interaction. For training and evaluation, we follow the same protocol as [61] to collect expert demonstrations in the captured large environments. + +# 3.1 Object-Interactable 3D-Captured Scenes + +To reduce the visual domain gap, a straightforward approach is to use 3D scans of real world environments. However, the captured 3D scans (i.e., meshes) remain static and thus, agents cannot interact with objects in the captured scenes. For object interaction, we manually replace object parts with 3D object assets to support object interaction. For photorealism comparison with previous environments using FID [27] and KID [4] metrics, please refer to Sec A.3. + +We detail the process of collecting object-interactable 3D-captured environments and highlight key differences from previously proposed benchmarks below. Data acquisition process. To collect 3D scans of real-world environments, we visit residential properties and employ scanners. Inspired by recent work [2], we collect scans outside of the US to add the diversity with public scanned environments. We utilize the same 3D scanner as [7], equipped with three RGB cameras + +![](images/a9f07cfaa82e1d03908cbdb4e1dff55c996e268fdb31727e566c774e133c7fd4.jpg) +(a) Distribution of floor areas by scene. + +![](images/e0395aa238e00e249073a9fc97d6591a5ad5252e1e5bf482a31c37c28554a4e3.jpg) +(b) Distribution of navigable areas by scene. +Fig. 4: Distribution of navigable and floor areas in interactive benchmarks by scenes. 'Floor area' denotes the overall size of the scene. 'Navigable area' denotes the size of the space in which the agent can actually navigate. For both metrics, the REALFRED benchmark poses a more even distribution and provides larger areas. For (a), we exclude RoboTHOR since it consists of single-sized floors. + +along with a depth sensor, and capture images from three distinct perspectives: front-facing and slightly vertical above and below. Panoramic images are acquired through six consecutive captures with horizontal rotation from a fixed viewpoint. We scan each house with 2.5-meter intervals and address blind spots due to furniture with additional scans from different viewpoints. Fig. 3 shows houses where scanners were deployed while capturing scans. It can be discerned that captures were densely concentrated in areas with a high presence of objects. + +Object-interactable environments. To construct an environment with numerous interactable objects, separating each object should be preceded. In other words, objects in the scans are initially merged into the background and therefore they remain as the background. Constructing an environment where objects can be interacted with requires the separation of objects from the background and from each other. Therefore, we manually separate the 3D scans into background elements and interactive objects. Furthermore, each object can exist in various states. To visualize changes in an object's state, we add state-relevant textures on objects. For example, we add a stain texture to a clean object when it becomes dirty. Finally, we reconstruct these individual object meshes within the Unity editor, making them compatible with the AI2-THOR simulator [35]. + +Comparison with previous benchmarks. To investigate the spatial characteristics of scenes in the REALFRED benchmark, we compare ours with other benchmarks that support interaction with objects [16, 61, 67]. We observe enhancements in our dataset in both: 1) spatial sizes and 2) spatial complexity. + +Spatial size. We compare the REALFRED benchmark with other benchmarks in terms of spatial size [56] by measuring 'Floor area' and 'Navigable area' and provide the result in Fig. 4. 'Floor area' represents the total spatial size $(m^2)$ of a scene, defined by the floor projection. 'Navigable area' measures the spatial size $(m^2)$ of the space in which an agent can actually navigate. 'Navigable area' is smaller than or equal to 'Floor area' since 'Navigable area' excludes areas where the agent collides with any components in the scene from 'Floor area.' + +We observe that the REALFRED benchmark yields a diverse distribution of floor areas, compared to previous work [61], with a larger average per scene. Furthermore, the REALFRED benchmark provides a broader range of navigable + +Table 2: Component sizes of interactive embodied AI benchmarks. Each 'Total floor area' and 'Total navigable area' (Total nav. area) denotes the sum of all floor areas and navigation areas. 'Nav. complex.' denotes navigational complexity. 'Scene clutter' measures the amount of clutter in the scene. We do not compare ReplicaCAD's Navigable area, Navigation complexity, and Scene clutter as agent sizes differ across simulators. The highest value for each metric is shown in bold. $^\dagger$ We count objects used as target objects in the ObjectNav task. $^\ddagger$ We count objects used in any of the tasks. + +
SimulatorHabitat2.0AI2THOR
DatasetReplicaCAD [67]RoboTHOR [16]ALFRED [61]REALFRED
# Scenes11175120150
Total floor area (m2)8,824.52,5742,55510,060
Total nav. area (m2)-1,2581,3564,251
Nav. complex.-2.0362.5493.020
Scene clutter-8.0955.1198.072
# object class9214†82‡112‡
+ +areas, with larger areas on average. This implies that an agent needs the ability to navigate effectively in spaces of varying sizes, generally wider on average. + +We also investigate the spatial characteristics of each benchmark [16, 61, 67] and summarize the result in Table 2. 'Total floor area' denotes the sum of all floor areas and 'Total navigable area' denotes the sum of all navigable areas in the dataset. We observe that the REALFRED benchmark has the highest total floor area and navigable area value, implying that REALFRED provides more navigation space for an agent than previous work [16, 61, 67]. + +Spatial complexity. We now investigate the complexity of spatial structures in the REALFRED benchmark using several metrics. To ensure a fair comparison with other datasets [16, 61, 67], we employ the navigation complexity introduced by [74] and the scene clutter measurement from [56]. A higher navigation complexity indicates an increased difficulty in navigating through the space, while a higher scene clutter implies the presence of more obstacles in the environment. By utilizing these metrics, we conduct a comparative analysis with object-interactable benchmarks [16, 61] and provide the result in Table 2. + +We observe that the REALFRED benchmark provides the environment with higher navigation complexity and scene clutter compared to other benchmarks [61, 67]. The high navigation complexity in our scenes stems from their multi-room composition. This setup requires the agent to execute more intricate navigation when moving from one room to another, in comparison to scenarios where the agent operates within a single room. We observe that ours poses the second-highest scene clutter with a marginal gap from [16]. This is because the spaces in [16] are relatively confined with a high density of furniture. We observe that ours has a similar value to [16], meaning that our scenes have a large amount of obstacles with a similar portion of [16] since our scenes' average size is larger than that of [16]. This implies that the REALFRED benchmark provides a complex and challenging space for the agent to explore the environment. + +![](images/f41c23831bd75a97a3598e7354dbde795a606657acc32dcd95152ffd4b8184fc.jpg) +Train +Fig. 5: The seven types of tasks' distribution in REALFRED. We provide $37.6\%$ more tasks in valid sets and $19.3\%$ in total compared to previous benchmark [61]. + +![](images/e0180f8fb7c3f0ef21b0fb068d837a3069c1812cee376ed45c70b4a5467ad558.jpg) +Valid Seen + +![](images/073226163b503917b3560c3247eee85d7b00959533c3443860d60d2f0076a1e0.jpg) +Valid Unseen + +![](images/82156fb55d9223338d7119ea5a9190dcf6ba80a128fd00e8ba871738d77ff374.jpg) + +# 3.2 Expert Demonstration Generation + +Each expert demonstration includes a set of an egocentric RGB view and action information with an interaction mask if exists at each time step. Expert demonstrations for each task are generated by a planner [29] with encoded state spaces into Planning Domain Definition Language (PDDL) rules [22]. To generate household tasks, we utilize seven task types introduced in [61]. + +Data splits. We split the generated demonstrations into train, validation, and test folds. Specifically, we designate 135 scenes for seen and 15 scenes for unseen fold. Comparison of the amount of each task and the distribution across training and validation folds with previous work [61] is shown in Fig. 5. + +Curating free-form language instructions. Detailed language instructions describe a task that involves a sequence of actions, a point of interest in robotics. The REALFRED benchmark offers 30,696 language directives, each comprising a human-annotated high-level goal and a set of step-by-step instructions. These directives are collected from 93 Amazon Mechanical Turk workers with a 'Master' qualification, ensuring high + +![](images/048bd0d8c9cef05a5b1bc06fcf3e4ed42f72571eff1eb60e49aa647fc0d3418e.jpg) +Fig.6:Word distribution.The words are arranged starting from the center and extending outwards. The arc length corresponds to the word frequency in the instructions. + +quality. Collected annotations are validated through an additional voting survey, and invalid instructions are replaced with newly collected instructions. The distribution of language instructions by their first four words is presented in Fig. 6. We provide a detailed annotation process and examples of an expert demonstration with the instruction in Sec. A.1 and A.4 in the supplementary material. + +Multi-room embodied instruction following. For household robots to be able to effectively assist humans, it would be more practical to deploy them in complex and diverse environments with various room types, rather than confining them to a single room. Cross-room navigation poses significant challenges, requiring agents to understand room references, plan efficient paths, and overcome obstacles proficiently. This agent is further tasked to interpret visual cues, demonstrate a keen sense of spatial awareness, and process natural language instructions. + +![](images/291eec7c70ead6d17fe86c455078a5e6226156db1f64a610f18d0a844d9a396a.jpg) +Fig.7: Comparison of expert demonstrations (ALFRED vs. REALFRED). Due to expanded environments, longer steps and planning horizons are required in REALFRED. + +The distribution of the number of steps and the length covered by the expert demonstrations is shown in Fig. 7. We observe that longer steps and trajectories are required to complete our tasks compared to the single-room constraint benchmark [61], meaning that our benchmark provides longer-horizon tasks. + +To broaden a range of tasks, we provide an extensive number of object class types in the REALFRED benchmark. More object classes are added to the REALFRED benchmark, making it a superset of [61] with 86 pickable and 26 receptacle objects. This has resulted in a more diverse set of tasks, with the number of unique tasks being $84.3\%$ more than the ALFRED benchmark [61], and the proportion of unique tasks being higher in the REALFRED benchmark. We provide detailed information in Sec. A in the supplementary material. + +# 4 Experiments + +Metrics. We follow the same evaluation protocol of the ALFRED benchmark [61]. The primary metric is 'Success Rate (SR)' which measures the percentage of completed tasks. 'Goal-Condition Success Rate (GC)' measures the percentage of achieved goal conditions. We also use path-length-weighted metrics to measure how efficiently an agent completes tasks. For more details, kindly refer to [61]. Baselines. We evaluate several recent state-of-the-art methods [5,33,34,46,61-63] with competitive results in [61]. We provide more details in Sec. B. + +# 4.1 Comparison of the State of the Arts + +We evaluate the baselines in the proposed REALFRED benchmark over multiple runs and present the average result in Table 3. We report extended results with path-length-weighted metrics in Sec. C in supplementary. + +For a fair comparison, we separate the baseline into two groups based on the use of additional depth supervision for semantic map reconstruction: 'Imitation Learning' where agents learn direct mapping from visual observations and language instructions to action sequences and 'Spatial Map Reconst.' where agents plan action sequences based on reconstructed semantic spatial representations. + +Table 3: Task and Goal-Condition Success Rate. We train and evaluate recent state-of-the-art methods on our REALFRED benchmark. For a fair comparison, we group these methods into two based on the usage of extra depth supervision: models learned by imitation learning without depth supervision ('Imitation Learning') and ones that maintain semantic spatial representations constructed by predicted depth maps ('Spatial Map Reconst.'). Path-length-weighted (PLW) metrics are reported in Sec. C for each value. $\dagger$ Authors' implementation as the code is not publicly available. + +
LearningModelValidationTest
SeenUnseenSeenUnseen
SRGCSRGCSRGCSRGC
Imitation LearningSeq2Seq [61]0.77 ± 0.066.93 ± 0.060.00 ± 0.004.03 ± 0.001.10 ± 0.006.60 ± 0.000.00 ± 0.003.50 ± 0.00
MOCA [62]12.64 ± 0.1220.95 ± 0.181.44 ± 0.056.76 ± 0.0414.11 ± 0.0322.84 ± 0.040.62 ± 0.085.14 ± 0.08
ABP† [33]24.71 ± 0.0533.80 ± 0.144.22 ± 0.0511.71 ± 0.2727.44 ± 0.4035.81 ± 0.233.54 ± 0.2310.57 ± 0.22
Spatial Map Reconst.HLSM [5]4.23 ± 0.089.14 ± 0.091.08 ± 0.146.12 ± 0.236.27 ± 0.0410.44 ± 0.130.49 ± 0.164.28 ± 0.13
FILM [46]7.08 ± 0.2811.93 ± 0.234.44 ± 0.179.26 ± 0.138.79 ± 0.0713.03 ± 0.082.15 ± 0.186.56 ± 0.15
LLM-Planner† [63]5.80 ± 0.1911.69 ± 0.353.33 ± 0.228.29 ± 0.198.16 ± 0.2013.20 ± 0.131.90 ± 0.136.33 ± 0.02
CAPEAM† [34]13.45 ± 0.0518.16 ± 0.274.92 ± 0.229.47 ± 0.2315.61 ± 0.1520.22 ± 0.112.87 ± 0.137.36 ± 0.07
Human------85.00 ± 3.5491.30 ± 2.94
+ +We observe that all these baselines, 'Imitation Learning' and 'Spatial Map Reconst.' consistently achieve lower performance values for all metrics in both seen and unseen splits compared to performances achieved in [61], implying that our proposed REALFRED provide more challenges compared to [61]. While in [61], the 'Spatial Map Reconst.' baselines [5,34,46] outperform learning-based approaches [33,61,62] by exploiting semantic spatial representation with deterministic algorithms (e.g., obstacle-free navigation path planning [60]), we observe a contrasting result that the 'Imitation Learning' baselines outperforming the 'Spatial Map Reconst.' baselines in our REALFRED benchmark. + +We qualitatively observe that such a confined spatial map reconstruction is due to a limited field of view and map reconstruction methods. The agent's limited field of view often leads to failure in recognizing room corners with doors or narrow aisles, resulting in (single-room-sized) limited map reconstruction. In addition, [34, 46] perceives obstacles larger than they actually are for better obstacle-free path planning by sacrificing navigable area, but this can be quite critical for narrow doors and aisles as they have a small amount of navigable space. This may hinder navigation to other rooms and thus, fail at tasks. We provide a more detailed discussion, supported by figures, on the agents' difficulty in recognizing narrow passages and the resulting impact on the agents, who struggle with spatial map reconstruction in Sec. D in the supplementary material. + +# 4.2 Comparison to the agents with sim-to-real adaptation + +We investigate the transfer from simulation training to real-world scan evaluation (sim-to-real) and from real-world scan training to real-world scan evaluation (real-to-real) with [33]. We train a sim-to-real agent with synthetic visual data and a real-to-real agent with real scanned visual data, respectively. During inference, both agents predict an action and an object mask based on the scanned visual input frame and the given language instruction at every time step. + +Training and evaluation data selection. For training the sim-to-real agent, we use the training dataset in [61], encompassing $21K$ language annotations. For a fair comparison, we train the real-to-real agent with tasks from the REALFRED benchmark's training fold, specifically involving the manipulation of the same objects that were used in training the sim-to-real agent with $19K$ language annotations. For evaluation, we select tasks from the valid unseen splits of our REALFRED benchmark. We evaluate the agent performance with the tasks that 1) do not require multi-room navigation 2) and those that do. This selection specifically includes tasks that feature objects used in the training phase. + +Real-to-sim domain adaptation. The use of Generative Adversarial Networks (GANs) [23,79] for domain adaptation has recently been examined in the literature on robotics [6,28,32,58,77]. These models are employed to adapt input images from real-world domains (i.e. Real $\rightarrow$ Sim) before they are passed to the agent policy. We train a CycleGan [79] and its off-the-shelf variant [70] with unpaired images collected from the REALFRED (real domain) and ALFRED benchmarks [61] (simulated domain). Among the trained generators, the one performing real to simulation conversion, referred to as the real-to-sim goggle, learns the mapping from the real domain to the simulation domain $G_{S}:\mathcal{R}\to \mathcal{S}$ , where $\mathcal{R}$ denotes the real domain and $S$ denotes the simulated domain. + +Results. We present the results of the sim-to-real experiments in Table 4. We report the performance of agents when tasks 1) require agents to navigate multirooms, denoted as 'Multi + Single,' and 2) are solvable within a single room, denoted as 'Single only.' Firstly, we observe that the sim-to-real agent significantly underperforms its real-to-real counterpart in all metrics $(\# (a)$ vs. $\# (d))$ .We then compare the results with gobbled agents. By comparing agents evaluated in the Multi + +Table 4: Comparison to the agents with sim-to-real adaptation. Each 'Sim2Real' and 'Real2Real' denotes an agent trained in simulated environments and 3D-captured scenes. 'Goggle' denotes real-to-sim methods. + +
# SettingGoggleMulti + SingleSingle only
SRPLWSRSRPLWSR
(a) Sim2RealNone0.1150.0120.00.0
(b) Sim2RealCycleGan [79]0.1150.0160.3270.065
(c) Sim2RealUVCGAN-v2 [70]0.1150.0460.3270.187
(d) Real2RealNone2.4050.7852.6140.762
+ ++ Single' tasks, we do not observe improvements in the main metric, success rate (SR). We observe a slight increase in the PLWSR metric for the gobbled sim-to-real agent compared to the vanilla sim-to-real agent (#(a) vs. #(b,c)). + +To isolate the impacts of the visual domain gap, we then compare agents evaluated on 'Single only' tasks. We observe that both gobbled agents show improvements over the vanilla sim-to-real agent which results in zero success rate. However, a noticeable performance gap exists with the real-to-real agent, implying the need for learning in real scanned environments $(\# (d)$ vs. $\# (a,b,c))$ + +# 4.3 Challenges in REALFRED + +We propose two hypotheses for the low performance observed with state-of-the-art methods on the REALFRED benchmark: (1) navigating within larger scenes and (2) overcoming narrow pathways between rooms. These elements represent challenges of completing tasks within multi-room, household-scale environments. + +![](images/3eaaf4466a90b882d59174864fcbaf743526bbd11465a420d2d8e479776af21d.jpg) +Fig. 8: Distribution of collision spots. The heatmap illustrates collision frequency on each spot and the black outline depicts a navigable area. We illustrate collision points of [33] in valid unseen environments to represent the distribution. Specifically, we choose failure episodes from the two largest scenes in the valid unseen split. We observe a remarkable concentration of collisions at the entrance to the rooms. The corresponding egocentric views are presented on the right side of the figure. + +We analyze failure cases with the best-performing model, ABP [33], on valid unseen split, since it seems the most promising on the REALFRED benchmark. Difficulty in large environments. Understanding the surrounding environment, including the location of objects and the positions of obstacles, can be beneficial for task completion. However, the agent's visual range is bounded, requiring it to expend more steps for exploration in order to perceive larger spaces. This intensifies the challenge of navigating larger spaces with limited steps. + +We conduct an analysis to see the relationship between the success rate (SR) and the size of the space. We follow the interquartile range (IQR) method to set a threshold at $30.44m^2$ that covers the most navigable space sizes in the ALFRED [61]. This threshold marks the upper fence, defining spaces above it as outliers. We classify spaces in the REALFRED benchmark smaller than threshold as smaller scenes, and those larger as larger scenes. Results indicate that the agent [33] with the highest SR in the valid unseen fold showed an average SR of $5.46\%$ in the smaller scenes and only $1.77\%$ in the larger scenes. This implies that solving tasks becomes more challenging as the space size increases. + +In addition, we compare the difficulty of navigation between the previous [61] and REALFRED benchmarks with different average spatial sizes. To quantify, we first define milestones for each task, which are spots to be reached to interact with target objects (e.g., Apple, Knife, etc.). We consider navigation to be a success only if all milestones are visited, regardless of whether the actual interaction (e.g., slicing an apple) is performed or not. Consequently, the navigation success rate in the REALFRED benchmark is merely $59.18\%$ for the valid unseen split while the agent's [33] navigation success rate is $84.82\%$ in ALFRED [61]. The REALFRED benchmark offering an average space size more than three times larger than the previous work [61] implies the need for model development capable of overcoming the challenges associated with this increased scale. + +Navigation through narrow doorways. The REALFRED benchmark supports environments with multiroom composition, unlike the previous dataset [61] + +providing a single room scale. Specifically, it contains narrow doorways across rooms, which may hinder the agent from navigating through. Here, we hypothesize that the agent would frequently collide with the walls near the door. + +We investigate collision spots and showcase two examples on top of the scene's layout in valid unseen fold in Fig. 8. The number of collisions in failed cases is accumulated and normalized to the maximum number of collisions, respectively, to indicate the collision frequency at each point. The black outline represents the agent's navigable area, as detailed in Sec. 3.1. Each value denotes the normalized collision frequency. In both scenes, we observe that collisions often occur on walls near the door, implying that our spatial characteristic (i.e., including narrow spaces) may hinder the agent from properly navigating without collision. + +Human evaluation. Following [61], we randomly select 100 directives from the test unseen fold and have them evaluated by humans. Five participants are given 20 tasks each, which they complete using a keyboard-and-mouse interface. Before starting, they are given the opportunity to become familiar with the interface. + +Participants achieve a comparable high success rate of $85\%$ , along with a goal-condition success rate of $91.30\%$ , in average. We agree that human performance may appear slightly lower compared to the results presented in previous work [57, 61, 78]. This is partly because human participants encountered several difficulties when controlling the agent, particularly in avoiding collisions within narrow corridors. Furthermore, navigating large spaces with a limited egocentric field of view introduced additional challenges, leading to task failures. + +# 5 Conclusion + +We present REALFRED, a new dataset and benchmark for embodied instruction following task on 3D-captured environments. We capture 150 indoor houses in 3D with interactable objects to enable complex household tasks. The reconstructed indoor scenes provide a larger spatial area and complex multi-room environments that are close to the real-world scenario and challenging for an agent to successfully complete a task. Expert demonstrations are also provided along with free-form human-language instructions. + +In our empirical evaluations, we show that state-of-the-art methods struggle in large multi-room environments, provide analyses of our newly proposed benchmark, and perform Sim2Real transfer experiments. We have released our Embodied AI research data and code for reproducibility. We expect that the REALFRED benchmark will encourage further research on developing robotic agents that execute household tasks by language instructions in the real world. + +Limitation and future work. Although we support a large number of interactable objects, the types of tasks to be completed are rather limited, considering more complex real-world scenarios. In addition, we currently address natural language in English but users may come from different regions with different languages. We can think of two future research avenues as follows. (1) adding additional complicated types of task that require both hands to complete. (2) supporting a multi-lingual interface for users from different regions. + +# Acknowledgements + +This work was partly supported by the NRF grant (No.2022R1A2C400230012, $5\%$ ) and IITP grants (No.RS-2022-II220077 $(5\%)$ , No.RS-2022-II220113 $(5\%)$ No.RS-2022-II220959 $(5\%)$ , No.RS-2022-II220871 $(15\%)$ , No.RS-2020-II201361 $(5\%)$ , Yonsei AI), No.RS-2021-II211343 $(5\%)$ , SNU AI), No.RS-2021-II212068 $(5\%)$ , AI Innov. Hub), No.RS-2022-II220951 $(50\%)$ ) funded by the Korea government(MSIT). + +# References + +1. Anderson, P., Wu, Q., Teney, D., Bruce, J., Johnson, M., Sünderhauf, N., Reid, I., Gould, S., van den Hengel, A.: Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In: CVPR (2018) +2. Baruch, G., Chen, Z., Dehghan, A., Dimry, T., Feigin, Y., Fu, P., Gebauer, T., Joffe, B., Kurz, D., Schwartz, A., Shulman, E.: ARKitscenes - a diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In: NeurIPS Datasets and Benchmarks Track (2021) +3. Berseth, G., Xie, C., Cernek, P., van de Panne, M.: Progressive reinforcement learning with distillation for multi-skilled motion control. In: ICLR (2018) +4. Binkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying mmd gans. In: ICLR (2018) +5. Blukis, V., Paxton, C., Fox, D., Garg, A., Artzi, Y.: A persistent spatial semantic representation for high-level natural language instruction execution. In: CoRL (2021) +6. Bousmalis, K., Irpan, A., Wohlhart, P., Bai, Y., Kelcey, M., Kalakrishnan, M., Downs, L., Ibarz, J., Pastor, P., Konolige, K., et al.: Using simulation and domain adaptation to improve efficiency of deep robotic grasping. In: ICRA (2018) +7. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3d: Learning from rgb-d data in indoor environments. arXiv:1709.06158 (2017) +8. Chaplot, D.S., Gandhi, D., Gupta, S., Gupta, A., Salakhutdinov, R.: Learning to explore using active neural slam. In: ICLR (2020) +9. Chaplot, D.S., Gandhi, D.P., Gupta, A., Salakhutdinov, R.R.: Object goal navigation using goal-oriented semantic exploration. In: NeurIPS (2020) +0. Chen, C., Jain, U., Schissler, C., Gari, S.V.A., Al-Halah, Z., Ithapu, V.K., Robinson, P., Grauman, K.: Soundspaces: Audio-visual navigation in 3d environments. In: ECCV (2020) +1. Chen, C., Schissler, C., Garg, S., Kobernik, P., Clegg, A., Calamia, P., Batra, D., Robinson, P.W., Grauman, K.: Soundspaces 2.0: A simulation platform for visual-acoustic learning. In: NeurIPS Datasets and Benchmarks Track (2022) +2. Chen, H., Suhr, A., Misra, D., Snavely, N., Artzi, Y.: Touchdown: Natural language navigation and spatial reasoning in visual street environments. In: CVPR (2019) +3. Chen, T., Gupta, S., Gupta, A.: Learning exploration policies for navigation. In: ICLR (2019) +4. Choi, S., Ji, G., Park, J., Kim, H., Mun, J., Lee, J.H., Hwangbo, J.: Learning quadrupedal locomotion on deformable terrain. Science Robotics (2023) +5. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017) + +16. Deitke, M., Han, W., Herrasti, A., Kembhavi, A., Kolve, E., Mottaghi, R., Salvador, J., Schwenk, D., VanderBilt, E., Wallingford, M., et al.: Robothor: An open simulation-to-real embodied ai platform. In: CVPR (2020) +17. Deitke, M., Hendrix, R., Farhadi, A., Ehsani, K., Kembhavi, A.: Phone2proc: Bringing robust robots into our chaotic world. In: CVPR (2023) +18. Deitke, M., VanderBilt, E., Herrasti, A., Weihs, L., Ehsani, K., Salvador, J., Han, W., Kolve, E., Kembhavi, A., Mottaghi, R.: Procthor: Large-scale embodied ai using procedural generation. In: NeurIPS (2022) +19. Ehsani, K., Han, W., Herrasti, A., Vanderbilt, E., Weihs, L., Kolve, E., Kembhavi, A., Mottaghi, R.: Manipulathor: A framework for visual object manipulation. In: CVPR (2021) +20. Gan, C., Schwartz, J., Alter, S., Mrowca, D., Schrimpf, M., Traer, J., De Freitas, J., Kubilius, J., Bhandwaldar, A., Haber, N., et al.: Threadworld: A platform for interactive multi-modal physical simulation. In: NeurIPS Datasets and Benchmarks Track (2021) +21. Gao, X., Gao, Q., Gong, R., Lin, K., Thattai, G., Sukhatme, G.S.: Dialfred: Dialogue-enabled agents for embodied instruction following. RA-L (2022) +22. Ghallab, M., Howe, A., Knoblock, C., McDermott, D., Ram, A., Veloso, M., Weld, D., Wilkins, D.: Pddl the planning domain definition language (1998) +23. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: NeurIPS (2014) +24. Gordon, D., Kembhavi, A., Rastegari, M., Redmon, J., Fox, D., Farhadi, A.: Iqa: Visual question answering in interactive environments. In: CVPR (2018) +25. Gu, J., Xiang, F., Li, X., Ling, Z., Liu, X., Mu, T., Tang, Y., Tao, S., Wei, X., Yao, Y., et al.: Maniskill2: A unified benchmark for generalizable manipulation skills. In: ICLR (2023) +26. Heo, M., Lee, Y., Lee, D., Lim, J.J.: Furniturebench: Reproducible real-world benchmark for long-horizon complex manipulation. In: RSS (2023) +27. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017) +28. Ho, D., Rao, K., Xu, Z., Jang, E., Khansari, M., Bai, Y.: Retinagan: An object-aware approach to sim-to-real transfer. In: ICRA (2021) +29. Hoffmann, J., Nebel, B.: The ff planning system: Fast plan generation through heuristic search. In: JAIR (2001) +30. Hua, B.S., Pham, Q.H., Nguyen, D.T., Tran, M.K., Yu, L.F., Yeung, S.K.: Scenenn: A scene meshes dataset with annotations. In: 3DV (2016) +31. Khanna, M., Mao, Y., Jiang, H., Haresh, S., Shacklett, B., Batra, D., Clegg, A., Undersander, E., Chang, A.X., Savva, M.: Habitat synthetic scenes dataset (hssd200): An analysis of 3d scene scale and realism tradeoffs for objectgoal navigation. In: CVPR (2024) +32. Khansari, M., Ho, D., Du, Y., Fuentes, A., Bennice, M., Sievers, N., Kirmani, S., Bai, Y., Jang, E.: Practical visual deep imitation learning via task-level domain consistency. In: ICRA (2023) +33. Kim, B., Bhambri, S., Singh, K.P., Mottaghi, R., Choi, J.: Agent with the big picture: Perceiving surroundings for interactive instruction following. In: Embodied AI Workshop @ CVPR (2021) +34. Kim, B., Kim, J., Kim, Y., Min, C., Choi, J.: Context-aware planning and environment-aware memory for instruction following embodied agents. In: ICCV (2023) + +35. Kolve, E., Mottaghi, R., Han, W., VanderBilt, E., Weihs, L., Herrasti, A., Gordon, D., Zhu, Y., Gupta, A., Farhadi, A.: Ai2-thor: An interactive 3d environment for visual ai. arXiv:1712.05474 (2017) +36. Krantz, J., Gervet, T., Yadav, K., Wang, A., Paxton, C., Mottaghi, R., Batra, D., Malik, J., Lee, S., Chaplot, D.S.: Navigating to objects specified by images. In: ICCV (2023) +37. Krantz, J., Wijmans, E., Majumdar, A., Batra, D., Lee, S.: Beyond the nav-graph: Vision-and-language navigation in continuous environments. In: ECCV (2020) +38. Ku, A., Anderson, P., Patel, R., Ie, E., Baldridge, J.: Room-Across-Room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. In: EMNLP (2020) +39. Kumar, A., Fu, Z., Pathak, D., Malik, J.: Rapid motor adaptation for legged robots. In: RSS (2021) +40. Li, C., Xia, F., Martin-Martin, R., Lingelbach, M., Srivastava, S., Shen, B., Vainio, K.E., Gokmen, C., Dharan, G., Jain, T., Kurenkov, A., Liu, K., Gweon, H., Wu, J., Fei-Fei, L., Savarese, S.: igibson 2.0: Object-centric simulation for robot learning of everyday household tasks. In: CoRL (2021) +41. Li, C., Zhang, R., Wong, J., Gokmen, C., Srivastava, S., Martin-Martin, R., Wang, C., Levine, G., Lingelbach, M., Sun, J., et al.: Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation. In: CoRL (2022) +42. Li, Z., Yu, T.W., Sang, S., Wang, S., Song, M., Liu, Y., Yeh, Y.Y., Zhu, R., Gundavarapu, N., Shi, J., Bi, S., Yu, H.X., Xu, Z., Sunkavalli, K., Hasan, M., Ramamoorthi, R., Chandraker, M.: Open rooms: An open framework for photorealistic indoor scene datasets. In: CVPR (2021) +43. MacMahon, M., Stankiewicz, B., Kuipers, B.: Walk the talk: Connecting language, knowledge, and action in route instructions. In: AAAI (2006) +44. Majumdar, et al.: Findthis: Language-driven object disambiguation in indoor environments. In: CoRL (2023) +45. Mao, Y., Zhang, Y., Jiang, H., Chang, A., Savva, M.: Multiscan: Scalable rgbd scanning for 3d environments with articulated objects. In: NeurIPS (2022) +46. Min, S.Y., Chaplot, D.S., Ravikumar, P., Bisk, Y., Salakhutdinov, R.: Film: Following instructions in language with modular methods. In: ICLR (2022) +47. Min, S.Y., Tsai, Y.H.H., Ding, W., Farhadi, A., Salakhutdinov, R., Bisk, Y., Zhang, J.: Object goal navigation with end-to-end self-supervision. In: IROS (2023) +48. Misra, D., Bennett, A., Blukis, V., Niklasson, E., Shatkhin, M., Artzi, Y.: Mapping instructions to actions in 3d environments with visual goal prediction. In: EMNLP (2018) +49. Padmakumar, A., Thomason, J., Shrivastava, A., Lange, P., Narayan-Chen, A., Gella, S., Piramuthu, R., Tur, G., Hakkani-Tur, D.: Teach: Task-driven embodied agents that chat. In: AAAI (2022) +50. Partsey, R., Wijmans, E., Yokoyama, N., Dobosevych, O., Batra, D., Maksymets, O.: Is mapping necessary for realistic pointgoal navigation? In: CVPR (2022) +51. Pashevich, A., Schmid, C., Sun, Chen: Episodic transformer for vision-and-language navigation. In: ICCV (2021) +52. Puig, X., Ra, K., Boben, M., Li, J., Wang, T., Fidler, S., Torralba, A.: Virtualhome: Simulating household activities via programs. In: CVPR (2018) +53. Puig, X., Undersander, E., Szot, A., Cote, M.D., Yang, T.Y., Partsey, R., Desai, R., Clegg, A.W., Hlavac, M., Min, S.Y., et al.: Habitat 3.0: A co-habitat for humans, avatars and robots. In: ICLR (2024) + +54. Qi, Y., Wu, Q., Anderson, P., Wang, X., Wang, W.Y., Shen, C., Hengel, A.v.d.: Reverie: Remote embodied visual referring expression in real indoor environments. In: CVPR (2020) +55. Ramakrishnan, S.K., Chaplot, D.S., Al-Halah, Z., Malik, J., Grauman, K.: Poni: Potential functions for objectgoal navigation with interaction-free learning. In: CVPR (2022) +56. Ramakrishnan, S.K., Gokaslan, A., Wijmans, E., Maksymets, O., Clegg, A., Turner, J.M., Undersander, E., Galuba, W., Westbury, A., Chang, A.X., Savva, M., Zhao, Y., Batra, D.: Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI. In: NeurIPS Datasets and Benchmarks Track (2021) +57. Ramrakhya, R., Undersander, E., Batra, D., Das, A.: Habitat-web: Learning embodied object-search strategies from human demonstrations at scale. In: CVPR (2022) +58. Rao, K., Harris, C., Irpan, A., Levine, S., Ibarz, J., Khansari, M.: Rl-cyclegan: Reinforcement learning aware simulation-to-real. In: CVPR (2020) +59. Savva, M., Kadian, A., Maksymets, O., Zhao, Y., Wijmans, E., Jain, B., Straub, J., Liu, J., Koltun, V., Malik, J., Parikh, D., Batra, D.: Habitat: A Platform for Embodied AI Research. In: ICCV (2019) +60. Sethian, J.A.: A fast marching level set method for monotonically advancing fronts. In: PNAS (1996) +61. Shridhar, M., Thomason, J., Gordon, D., Bisk, Y., Han, W., Mottaghi, R., Zettle-moyer, L., Fox, D.: Alfred: A benchmark for interpreting grounded instructions for everyday tasks. In: CVPR (2020) +62. Singh, K.P., Bhambri, S., Kim, B., Mottaghi, R., Choi, J.: Factorizing perception and policy for interactive instruction following. In: ICCV (2021) +63. Song, C.H., Wu, J., Washington, C., Sadler, B.M., Chao, W.L., Su, Y.: Llm-planner: Few-shot grounded planning for embodied agents with large language models. In: ICCV (2023) +64. Song, S., Lichtenberg, S.P., Xiao, J.: Sun rgb-d: A rgb-d scene understanding benchmark suite. In: CVPR (2015) +65. Srivastava, S., Li, C., Lingelbach, M., Martin-Martin, R., Xia, F., Vainio, K.E., Lian, Z., Gokmen, C., Buch, S., Liu, K., et al.: Behavior: Benchmark for everyday household activities in virtual, interactive, and ecological environments. In: CoRL (2021) +66. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv:1906.05797 (2019) +67. Szot, A., Clegg, A., Undersander, E., Wijmans, E., Zhao, Y., Turner, J., Maestre, N., Mukadam, M., Chaplot, D., Maksymets, O., Gokaslan, A., Vondrus, V., Dharur, S., Meier, F., Galuba, W., Chang, A., Kira, Z., Koltun, V., Malik, J., Savva, M., Batra, D.: Habitat 2.0: Training home assistants to rearrange their habitat. In: NeurIPS (2021) +68. Tan, J., Zhang, T., Coumans, E., Iscen, A., Bai, Y., Hafner, D., Bohez, S., Vanhoucke, V.: Sim-to-real: Learning agile locomotion for quadruped robots. In: RSS (2018) +69. Tobin, J., Fong, R., Ray, A., Schneider, J., Zaremba, W., Abbeel, P.: Domain randomization for transferring deep neural networks from simulation to the real world. In: IROS (2017) + +70. Torbunov, D., Huang, Y., Tseng, H.H., Yu, H., Huang, J., Yoo, S., Lin, M., Viren, B., Ren, Y.: Uvçgan v2: An improved cycle-consistent gan for unpaired image-to-image translation. arXiv:2303.16280 (2023) +71. Truong, J., Chernova, S., Batra, D.: Bi-directional domain adaptation for sim2real transfer of embodied navigation agents. RA-L (2021) +72. Weihs, L., Deitke, M., Kembhavi, A., Mottaghi, R.: Visual room rearrangement. In: CVPR (2021) +73. Wijmans, E., Kadian, A., Morcos, A., Lee, S., Essa, I., Parikh, D., Savva, M., Batra, D.: Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. In: ICLR (2020) +74. Xia, F., Zamir, A.R., He, Z., Sax, A., Malik, J., Savarese, S.: Gibson env: Realworld perception for embodied agents. In: CVPR (2018) +75. Yenamandra, S., Ramachandran, A., Yadav, K., Wang, A., Khanna, M., Gervet, T., Yang, T.Y., Jain, V., Clegg, A.W., Turner, J., et al.: Homerobot: Open-vocabulary mobile manipulation. In: CoRL (2023) +76. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: ICCV (2023) +77. Zhang, J., Tai, L., Yun, P., Xiong, Y., Liu, M., Boedecker, J., Burgard, W.: Vr-goggles for robots: Real-to-sim domain adaptation for visual control. RA-L (2019) +78. Zhu, H., Kapoor, R., Min, S.Y., Han, W., Li, J., Geng, K., Neubig, G., Bisk, Y., Kembhavi, A., Weihs, L.: Excalibur: Encouraging and evaluating embodied exploration. In: CVPR (2023) +79. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: ICCV (2017) \ No newline at end of file diff --git a/2024/ReALFRED_ An Embodied Instruction Following Benchmark in Photo-Realistic Environments/images.zip b/2024/ReALFRED_ An Embodied Instruction Following Benchmark in Photo-Realistic Environments/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..695df13a120a0ed9c94b043d889c2cee148b0a6c --- /dev/null +++ b/2024/ReALFRED_ An Embodied Instruction Following Benchmark in Photo-Realistic Environments/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:41b83c4761a3d1500d5ee88ef94c10bbdcd9664ff5cf91fe11c475a3a166bf92 +size 477715 diff --git a/2024/ReALFRED_ An Embodied Instruction Following Benchmark in Photo-Realistic Environments/layout.json b/2024/ReALFRED_ An Embodied Instruction Following Benchmark in Photo-Realistic Environments/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..af4cbb14fcf184b19efecc89cfaa9e63684d37ba --- /dev/null +++ b/2024/ReALFRED_ An Embodied Instruction Following Benchmark in Photo-Realistic Environments/layout.json @@ -0,0 +1,9250 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 111, + 481, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 111, + 481, + 146 + ], + "spans": [ + { + "bbox": [ + 133, + 111, + 481, + 146 + ], + "type": "text", + "content": "REALFRED: An Embodied Instruction Following Benchmark in Photo-Realistic Environments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 173, + 167, + 438, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 167, + 438, + 193 + ], + "spans": [ + { + "bbox": [ + 173, + 167, + 438, + 193 + ], + "type": "text", + "content": "Taewoong Kim" + }, + { + "bbox": [ + 173, + 167, + 438, + 193 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 173, + 167, + 438, + 193 + ], + "type": "text", + "content": ", Cheolhong Min" + }, + { + "bbox": [ + 173, + 167, + 438, + 193 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 173, + 167, + 438, + 193 + ], + "type": "text", + "content": ", Byeonghwi Kim" + }, + { + "bbox": [ + 173, + 167, + 438, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 173, + 167, + 438, + 193 + ], + "type": "text", + "content": ", Jinyeon Kim" + }, + { + "bbox": [ + 173, + 167, + 438, + 193 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 173, + 167, + 438, + 193 + ], + "type": "text", + "content": ", Wonje Jeung" + }, + { + "bbox": [ + 173, + 167, + 438, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 173, + 167, + 438, + 193 + ], + "type": "text", + "content": ", and Jonghyun Choi" + }, + { + "bbox": [ + 173, + 167, + 438, + 193 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 249, + 201, + 363, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 201, + 363, + 213 + ], + "spans": [ + { + "bbox": [ + 249, + 201, + 363, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 249, + 201, + 363, + 213 + ], + "type": "text", + "content": " Seoul National University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 266, + 213, + 347, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 213, + 347, + 224 + ], + "spans": [ + { + "bbox": [ + 266, + 213, + 347, + 224 + ], + "type": "text", + "content": "2 Yonsei University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 157, + 224, + 455, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 224, + 455, + 247 + ], + "spans": [ + { + "bbox": [ + 157, + 224, + 455, + 247 + ], + "type": "text", + "content": "{twoongg.kim, cheolhong.min, byeonghwikim}@snu.ac.kr, jinyeonkim@yonsei.ac.kr, {specific0924,jonghyunchoi}@snu.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 267, + 455, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 267, + 455, + 455 + ], + "spans": [ + { + "bbox": [ + 159, + 267, + 455, + 455 + ], + "type": "text", + "content": "Abstract. Simulated virtual environments have been widely used to learn robotic agents that perform daily household tasks. These environments encourage research progress by far, but often provide limited object interactability, visual appearance different from real-world environments, or relatively smaller environment sizes. This prevents the learned models in the virtual scenes from being readily deployable. To bridge the gap between these learning environments and deploying (i.e., real) environments, we propose the REALFRED benchmark that employs real-world scenes, objects, and room layouts to learn agents to complete household tasks by understanding free-form language instructions and interacting with objects in large, multi-room and 3D-captured scenes. Specifically, we extend the ALFRED benchmark with updates for larger environmental spaces with smaller visual domain gaps. With REALFRED, we analyze previously crafted methods for the ALFRED benchmark and observe that they consistently yield lower performance in all metrics, encouraging the community to develop methods in more realistic environments. Our code and data are publicly available3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 464, + 453, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 464, + 453, + 487 + ], + "spans": [ + { + "bbox": [ + 159, + 464, + 453, + 487 + ], + "type": "text", + "content": "Keywords: Interactive Scanned Environments " + }, + { + "bbox": [ + 159, + 464, + 453, + 487 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 464, + 453, + 487 + ], + "type": "text", + "content": " Instruction Following " + }, + { + "bbox": [ + 159, + 464, + 453, + 487 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 464, + 453, + 487 + ], + "type": "text", + "content": " Embodied AI " + }, + { + "bbox": [ + 159, + 464, + 453, + 487 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 464, + 453, + 487 + ], + "type": "text", + "content": " Reality Gap " + }, + { + "bbox": [ + 159, + 464, + 453, + 487 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 464, + 453, + 487 + ], + "type": "text", + "content": " Dataset and Benchmark" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 518, + 230, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 518, + 230, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 518, + 230, + 530 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 540, + 482, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 540, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 540, + 482, + 636 + ], + "type": "text", + "content": "Building autonomous robotic assistants that can perform everyday household tasks has been an elusive aspiration within the research community for decades. To let them learn these intricate tasks, we may provide them with interactive environments where agents can learn task completion skills with numerous interactions with environments. A straightforward approach to train such agents that can carry out real-world activities is to directly deploy robots in real-world environments and let them learn to complete desired tasks. However, this often faces several practical challenges, including cost, time, or safety concerns [3,14,39,68]." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 642, + 375, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 642, + 375, + 653 + ], + "spans": [ + { + "bbox": [ + 145, + 642, + 375, + 653 + ], + "type": "text", + "content": "3 Homepage: https://github.com/snumprlab/realfred" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 145, + 653, + 329, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 653, + 329, + 665 + ], + "spans": [ + { + "bbox": [ + 145, + 653, + 329, + 665 + ], + "type": "text", + "content": "*Equal contribution. †Corresponding author." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 125, + 465, + 247 + ], + "blocks": [ + { + "bbox": [ + 156, + 114, + 365, + 124 + ], + "lines": [ + { + "bbox": [ + 156, + 114, + 365, + 124 + ], + "spans": [ + { + "bbox": [ + 156, + 114, + 365, + 124 + ], + "type": "text", + "content": "Goal: Put a cooled slice of banana on the coffee table" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 151, + 125, + 465, + 247 + ], + "lines": [ + { + "bbox": [ + 151, + 125, + 465, + 247 + ], + "spans": [ + { + "bbox": [ + 151, + 125, + 465, + 247 + ], + "type": "image", + "image_path": "3402938b11df9e6b93447085c34b326633deb8fd70164f71dcb53e1f2184e0d2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 151, + 247, + 463, + 327 + ], + "blocks": [ + { + "bbox": [ + 151, + 247, + 463, + 327 + ], + "lines": [ + { + "bbox": [ + 151, + 247, + 463, + 327 + ], + "spans": [ + { + "bbox": [ + 151, + 247, + 463, + 327 + ], + "type": "image", + "image_path": "e0dd9640bb81a4c70c70f16b531072f75286b4cbb3436e8658bb962114a7e7db.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 335, + 482, + 392 + ], + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 392 + ], + "type": "text", + "content": "Fig. 1: Proposed REALFRED benchmark. The top image provides a perspective view of one of our scenes. The images below represent third-person views at each time step, along with their corresponding descriptions, for better understanding. The agent is required to understand instructions in natural language and then complete the desired tasks by navigating large 3D-captured environments and interacting with objects." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 422, + 482, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 626 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 626 + ], + "type": "text", + "content": "As an alternative, several simulated environments have been introduced [59, 74] which leverage extensive 3D-captured environments obtained from real-world scenes [7, 56]. Compared to real-world deployment, these environments offer agents a faster process of taking actions and observing consequences, and the convenience of resetting the environment and trying again in case of failure, which enables agents to learn the skills to complete desired tasks. Adopting such simulated environments has produced remarkable advancements in various subtasks for embodied AI agents, including visual navigation [9, 47, 50, 55, 73], vision and language navigation [1, 37], and remote object grounding [54]. Due to the inherently static nature of these 3D-captured environments where objects (e.g., books, chairs, etc.) remain non-interactive, however, current benchmarks [1, 9] for these tasks have less focused on object interaction, which might hinder deployability for more complex tasks that require object interaction. Recent studies [44, 57] insert liftable objects in scanned environments for object interaction, but they support limited object interaction such as picking up objects, which might not provide enough deployability for more challenging real-world scenarios such as heating objects using a microwave or cooling objects using a refrigerator." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Meanwhile, virtual game engines such as Unity have been exploited to build object-interactable environments with graphically crafted assets, including walls, floors, ceilings, and objects. These object-interactable environments have led to" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "text", + "content": "T. Kim and C. Min et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "notable progress in the handling of more intricate object-centric tasks, including rearrangement [67,72] and manipulation [25,26], beyond navigation-centric tasks. In particular, we have observed significant progress in the execution of more complex tasks by natural language instructions [5, 33, 46, 51, 61, 62]. However, object-interactable environments for training and evaluation of such agents often pose several issues, such as visual domain gaps [69] and relatively smaller room sizes compared to their counterparts in 3D-captured environments [7, 56, 66]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 201, + 317, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 317, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 317, + 429 + ], + "type": "text", + "content": "To bridge the gap between limited object interactability and environmental sizes with visual domain discrepancy, we propose the REAL-WORLD ALFRED (REALFRED) benchmark that requires agents to complete long-horizon tasks by understanding free-form language instructions and interacting with objects in large 3D-captured environments, following a similar task setup to the widely used embodied instruction following benchmark, ALFRED [61]. The 3D-captured environments used in the REALFRED benchmark encompass multiple rooms, providing ample space for agents to engage in multiple rooms in a single episode, which adds a sense of realism to the tasks. This resembles real-world scenarios in which agents navigate seamlessly between mul" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 331, + 205, + 476, + 363 + ], + "blocks": [ + { + "bbox": [ + 331, + 205, + 476, + 363 + ], + "lines": [ + { + "bbox": [ + 331, + 205, + 476, + 363 + ], + "spans": [ + { + "bbox": [ + 331, + 205, + 476, + 363 + ], + "type": "image", + "image_path": "15b4cd8f949c079b9155588ab4edbed0bf4ea81544184566b97f81cb7455f248.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 369, + 482, + 425 + ], + "lines": [ + { + "bbox": [ + 321, + 369, + 482, + 425 + ], + "spans": [ + { + "bbox": [ + 321, + 369, + 482, + 425 + ], + "type": "text", + "content": "Fig. 2: While other benchmarks [7, 12, 19, 37, 41, 43, 48, 49, 56, 57, 61, 67, 72, 74] provide one or two aspects, our proposed REALFRED benchmark addresses all of these aspects." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 429, + 482, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 482, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 482, + 487 + ], + "type": "text", + "content": "tiple rooms. Unlike prior benchmarks focusing primarily on single-room activities [61] or 3D-captured environments with limited object interaction [59], REALFRED provides task evaluation in wide, realistic, and interactive environments to mirror the natural expectations of human-robot interactions. Fig. 1. illustrates the household task in one of the scenes in ours." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 490, + 482, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 482, + 537 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 482, + 537 + ], + "type": "text", + "content": "In our experiments, we observe that the models [5,33,46,61-63] proposed in synthetic environments [61] do not perform well in our REALFRED benchmark, implying that models developed for synthetic environments may not easily adapt to realistic environments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 540, + 326, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 540, + 326, + 552 + ], + "spans": [ + { + "bbox": [ + 131, + 540, + 326, + 552 + ], + "type": "text", + "content": "We summarize our contributions as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 566, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 130, + 566, + 481, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 566, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 566, + 481, + 590 + ], + "type": "text", + "content": "- We propose REALFRED, a benchmark for embodied instruction following with 3D-captured multi-room environments and objects." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 592, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 592, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 592, + 481, + 628 + ], + "type": "text", + "content": "- We collect 3D-captured scenes and objects to reduce the simulation-reality gap and free-form language instructions to support task completion based on agents' language understanding." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "- We provide analyses on the recent state-of-the-art models in the literature and the relevant Sim2Real transfer to empirically validate the necessity of our REALFRED benchmark." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "text", + "content": "REALFRED" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 223, + 480, + 378 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 214 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 214 + ], + "type": "text", + "content": "Table 1: Comparison of REALFRED and other Embodied AI benchmarks. 'Language' column denotes the number of human annotated language directives. 'Environment' column compares spatial characteristics and whether it supports interactivity. 'Inference' column denotes whether their action space includes interactive capability with objects. The REALFRED benchmark is the first benchmark that provides a 3D-captured and interactable environment to solve household tasks that require navigation and interaction at the same time directed by natural language commands. " + }, + { + "bbox": [ + 132, + 114, + 482, + 214 + ], + "type": "inline_equation", + "content": "^\\dagger" + }, + { + "bbox": [ + 132, + 114, + 482, + 214 + ], + "type": "text", + "content": "We count the number of dialogue sessions. " + }, + { + "bbox": [ + 132, + 114, + 482, + 214 + ], + "type": "inline_equation", + "content": "^\\ddagger" + }, + { + "bbox": [ + 132, + 114, + 482, + 214 + ], + "type": "text", + "content": "We count the number of annotations in English. *Though [16] environment supports interaction, interaction is not required." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 223, + 480, + 378 + ], + "lines": [ + { + "bbox": [ + 134, + 223, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 134, + 223, + 480, + 378 + ], + "type": "table", + "html": "
LanguageEnvironmentInference
Human AnnotationsVisual QualityMulti-Room NavigationMovable ObjectsState ChangesObject Interactability
IQA [24]-SyntheticXX✓✓
ManipulaTHOR [19]; RoomR [72]-SyntheticX✓✓✓
RoboTHOR [16]-Synthetic✓✓✓X*
ProcTHOR [18]-Synthetic✓✓✓✓
ReplicaCAD [67]-Synthetic✓✓✓✓
BEHAVIOR-1K [41]; HSSD-200 [31]-Synthetic✓✓✓✓
OpenRooms [42]-PhotoXXXX
MP3D [7]; HM3D [56]; Gibson [74]-Photo✓XXX
Habitat-Web [57]-Photo✓✓✓✓
ALFRED [61]; TEACH [49]; CHAI [48]25k+; 2.0k+†;12k+SyntheticX✓✓✓
LANI [48]; Walk the Talk [43]28k+; 0.7k+Synthetic✓XXX
Virtualhome [52]2.7k+Synthetic✓XX✓
R2R [1]; RxR [38]; VLN-CE [37]21k+; 42k+‡; 21k+Photo✓XXX
REALFRED30k+Photo✓✓✓✓
", + "image_path": "a4bee094765f4e7e0eb96a351332041eae3e5cef0785e96a984bf6ddbc054a0d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 399, + 238, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 399, + 238, + 414 + ], + "spans": [ + { + "bbox": [ + 132, + 399, + 238, + 414 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 426, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 485 + ], + "type": "text", + "content": "Fig. 2 illustrates the comparison of our REALFRED with other benchmarks in three selected aspects. We first review 3D-captured environments and benchmarks that provide visual aesthetics similar to real-world environments. Then, we review simulation environments and benchmarks that support object interaction. Table 1 compares our REALFRED with other indoor datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 594 + ], + "type": "text", + "content": "Datasets and benchmarks with 3D-captured environments. The collection of advanced 3D scans has played a key role in enhancing our understanding of 3D objects [15, 64, 76] and their perception [2, 30, 45]. While these datasets offer valuable insights for a deeper understanding of 3D environments, they lack object interaction for learning interactive embodied agents. To further promote research on embodied agents for real-world applications, training and evaluation of such agents with physical spaces from scanned data [7, 66, 74] have been proposed. They provide a rich source of data for researchers to explore the capabilities of agents operating within real-world-inspired scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "In these environments, notable progress has been achieved, primarily in the realms of navigation and exploration. Extensive research has been conducted on agents capable of navigating complex 3D environments, as evidenced by work such as navigating to a specified object [9, 47, 55], navigating to a certain point [50, 73], and navigating to the shown image [36]. Similarly, exploring novel environments has also yielded valuable insights [8, 13]." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "text", + "content": "T. Kim and C. Min et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "content": "Additionally, there is a work to integrate multi-modal sensory information for developing an agent that can deeply understand environments using inputs such as vision and language [1,38] and audio and vision [10,11]. These advances enable agents to utilize a broader range of sensory data, enhancing their perception and interaction capabilities. Further research explores additional capabilities of embodied agents, such as object localization alongside navigation [54]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 480, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 480, + 259 + ], + "type": "text", + "content": "However, a fundamental limitation of these scan-based environments remains their static nature, which restricts the potential for object interaction. This limitation has resulted in the majority of prior research primarily focusing on navigation tasks, leaving the broader scope of agent adaptability to more complex, interactive challenges largely unaddressed. Consequently, this limitation hampers the adaptability of agents to more complex tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 260, + 481, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 481, + 331 + ], + "type": "text", + "content": "Object-interactable simulated environments. Scan-based environments [56, 76] are typically composed of static scene representations, focusing on the visual aspects of real-world spaces. However, they often fail to adequately support interactivity with objects, representing a discrepancy between the dynamic and interactive characteristics of the real world, necessitating the development of environments better aligned with the demands of interactive agent research." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 332, + 480, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 480, + 439 + ], + "type": "text", + "content": "To address this gap, various simulators, such as AI2-THOR [35], ManipulaTHOR [19], RoboTHOR [16], VirtualHome [52], TDW [20], iGibson [40], and OmniGibson [41], have emerged as promising solutions. These environments are engineered with a primary focus on enabling interaction between agents and objects. They are built upon game engines, which provide a solid foundation for ensuring realistic interactivity within the virtual environment. On these simulators, researchers publish benchmarks [16, 21, 41, 49, 53, 61, 65, 75] that support interactions. These benchmarks have been promoting research in the field, leading to the development of robotic assistants capable of handling complex tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 441, + 480, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 480, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 480, + 594 + ], + "type": "text", + "content": "However, the dataset that provide interactive environments are limited in size of a room because creating a large high-fidelity space is challenging [56]. Furthermore, despite their near-photo realism, agents would face a visual domain gap when deployed in real world environments [69,71,77]. To this end, Habitat-Web [57] proposes an interactive pick-and-place task based on the Habitat simulator [59,67]. Their template-based language instructions, however, might not be sufficient to express the complex nuances of human expression. Innovative methods for acquiring 3D scans from phone-captured layouts have been proposed for learning enviroment-specific policies [17]. Nevertheless, its synthetic assets may lead to visual domain gaps when deployed in the real world. On the contrary, our REALFRED supports photorealism, high interactivity with objects, and freeform language annotations. These features can offer a framework for developing language-driven agents with visual perception for complex household chores." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 615, + 331, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 615, + 331, + 628 + ], + "spans": [ + { + "bbox": [ + 131, + 615, + 331, + 628 + ], + "type": "text", + "content": "3 The REALFRED Benchmark" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "To develop agents capable of performing household tasks, substantial progress has been achieved in various domains, including navigation [1, 38], rearrange-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "text", + "content": "REALFRED" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 294, + 236 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 294, + 236 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 294, + 236 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 294, + 236 + ], + "type": "image", + "image_path": "7dbd702b14f1b095ea5db2c1acc45bbe4d7cac0956e9dbce94033e25598cc581.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 243, + 482, + 277 + ], + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 277 + ], + "type": "text", + "content": "Fig. 3: Top-down view of 3D-captured environments. We provide two examples from our scanned indoor environments. White circles denote where scanners are deployed. By scanning scenes at diverse points, we can prevent including blind spots." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 326, + 113, + 465, + 235 + ], + "blocks": [ + { + "bbox": [ + 326, + 113, + 465, + 235 + ], + "lines": [ + { + "bbox": [ + 326, + 113, + 465, + 235 + ], + "spans": [ + { + "bbox": [ + 326, + 113, + 465, + 235 + ], + "type": "image", + "image_path": "5c4e1ce4af62ed6466e4f9502c01e5596a57e81adffb45c3db7f7372d41057b7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 297, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 344 + ], + "type": "text", + "content": "ment [72], and manipulation tasks [25,26]. In particular, [61] recently introduced the ALFRED benchmark that requires agents to complete long-horizon household tasks by jointly understanding egocentric visual observations and natural language instructions in household environments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 346, + 483, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 483, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 483, + 429 + ], + "type": "text", + "content": "However, these environments are restricted to a single room size compared to previously proposed 3D-captured environments [7, 56] consisting of multiple rooms, which could potentially restrict the deployability of agents to larger environments. Furthermore, the environments used in the ALFRED benchmark [61] are built with synthetic CAD assets and therefore could potentially yield visual aesthetics different from those obtained from real-world environments [69], which could eventually cause performance degradation due to visual domain gaps." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 430, + 483, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 483, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 483, + 489 + ], + "type": "text", + "content": "To address these issues, we extend the ALFRED benchmark [61] and propose a challenging benchmark, named the REALFRED benchmark, which requires agents to perform household tasks in large indoor environments captured in 3D with object interaction. For training and evaluation, we follow the same protocol as [61] to collect expert demonstrations in the captured large environments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 504, + 366, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 504, + 366, + 517 + ], + "spans": [ + { + "bbox": [ + 131, + 504, + 366, + 517 + ], + "type": "text", + "content": "3.1 Object-Interactable 3D-Captured Scenes" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "text", + "content": "To reduce the visual domain gap, a straightforward approach is to use 3D scans of real world environments. However, the captured 3D scans (i.e., meshes) remain static and thus, agents cannot interact with objects in the captured scenes. For object interaction, we manually replace object parts with 3D object assets to support object interaction. For photorealism comparison with previous environments using FID [27] and KID [4] metrics, please refer to Sec A.3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 483, + 666 + ], + "type": "text", + "content": "We detail the process of collecting object-interactable 3D-captured environments and highlight key differences from previously proposed benchmarks below. Data acquisition process. To collect 3D scans of real-world environments, we visit residential properties and employ scanners. Inspired by recent work [2], we collect scans outside of the US to add the diversity with public scanned environments. We utilize the same 3D scanner as [7], equipped with three RGB cameras" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 271, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 271, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 271, + 101 + ], + "type": "text", + "content": "T. Kim and C. Min et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 301, + 178 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 301, + 178 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 301, + 178 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 301, + 178 + ], + "type": "image", + "image_path": "a9f07cfaa82e1d03908cbdb4e1dff55c996e268fdb31727e566c774e133c7fd4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 182, + 287, + 191 + ], + "lines": [ + { + "bbox": [ + 146, + 182, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 146, + 182, + 287, + 191 + ], + "type": "text", + "content": "(a) Distribution of floor areas by scene." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 303, + 114, + 468, + 178 + ], + "blocks": [ + { + "bbox": [ + 303, + 114, + 468, + 178 + ], + "lines": [ + { + "bbox": [ + 303, + 114, + 468, + 178 + ], + "spans": [ + { + "bbox": [ + 303, + 114, + 468, + 178 + ], + "type": "image", + "image_path": "e0395aa238e00e249073a9fc97d6591a5ad5252e1e5bf482a31c37c28554a4e3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 182, + 465, + 191 + ], + "lines": [ + { + "bbox": [ + 307, + 182, + 465, + 191 + ], + "spans": [ + { + "bbox": [ + 307, + 182, + 465, + 191 + ], + "type": "text", + "content": "(b) Distribution of navigable areas by scene." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 201, + 482, + 257 + ], + "lines": [ + { + "bbox": [ + 130, + 201, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 482, + 257 + ], + "type": "text", + "content": "Fig. 4: Distribution of navigable and floor areas in interactive benchmarks by scenes. 'Floor area' denotes the overall size of the scene. 'Navigable area' denotes the size of the space in which the agent can actually navigate. For both metrics, the REALFRED benchmark poses a more even distribution and provides larger areas. For (a), we exclude RoboTHOR since it consists of single-sized floors." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 281, + 482, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 365 + ], + "type": "text", + "content": "along with a depth sensor, and capture images from three distinct perspectives: front-facing and slightly vertical above and below. Panoramic images are acquired through six consecutive captures with horizontal rotation from a fixed viewpoint. We scan each house with 2.5-meter intervals and address blind spots due to furniture with additional scans from different viewpoints. Fig. 3 shows houses where scanners were deployed while capturing scans. It can be discerned that captures were densely concentrated in areas with a high presence of objects." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 365, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 496 + ], + "type": "text", + "content": "Object-interactable environments. To construct an environment with numerous interactable objects, separating each object should be preceded. In other words, objects in the scans are initially merged into the background and therefore they remain as the background. Constructing an environment where objects can be interacted with requires the separation of objects from the background and from each other. Therefore, we manually separate the 3D scans into background elements and interactive objects. Furthermore, each object can exist in various states. To visualize changes in an object's state, we add state-relevant textures on objects. For example, we add a stain texture to a clean object when it becomes dirty. Finally, we reconstruct these individual object meshes within the Unity editor, making them compatible with the AI2-THOR simulator [35]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 497, + 481, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 481, + 544 + ], + "type": "text", + "content": "Comparison with previous benchmarks. To investigate the spatial characteristics of scenes in the REALFRED benchmark, we compare ours with other benchmarks that support interaction with objects [16, 61, 67]. We observe enhancements in our dataset in both: 1) spatial sizes and 2) spatial complexity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 545, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 628 + ], + "type": "text", + "content": "Spatial size. We compare the REALFRED benchmark with other benchmarks in terms of spatial size [56] by measuring 'Floor area' and 'Navigable area' and provide the result in Fig. 4. 'Floor area' represents the total spatial size " + }, + { + "bbox": [ + 130, + 545, + 482, + 628 + ], + "type": "inline_equation", + "content": "(m^2)" + }, + { + "bbox": [ + 130, + 545, + 482, + 628 + ], + "type": "text", + "content": " of a scene, defined by the floor projection. 'Navigable area' measures the spatial size " + }, + { + "bbox": [ + 130, + 545, + 482, + 628 + ], + "type": "inline_equation", + "content": "(m^2)" + }, + { + "bbox": [ + 130, + 545, + 482, + 628 + ], + "type": "text", + "content": " of the space in which an agent can actually navigate. 'Navigable area' is smaller than or equal to 'Floor area' since 'Navigable area' excludes areas where the agent collides with any components in the scene from 'Floor area.'" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "We observe that the REALFRED benchmark yields a diverse distribution of floor areas, compared to previous work [61], with a larger average per scene. Furthermore, the REALFRED benchmark provides a broader range of navigable" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "text", + "content": "REALFRED" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 201, + 472, + 301 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "text", + "content": "Table 2: Component sizes of interactive embodied AI benchmarks. Each 'Total floor area' and 'Total navigable area' (Total nav. area) denotes the sum of all floor areas and navigation areas. 'Nav. complex.' denotes navigational complexity. 'Scene clutter' measures the amount of clutter in the scene. We do not compare ReplicaCAD's Navigable area, Navigation complexity, and Scene clutter as agent sizes differ across simulators. The highest value for each metric is shown in bold. " + }, + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "inline_equation", + "content": "^\\dagger" + }, + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "text", + "content": "We count objects used as target objects in the ObjectNav task. " + }, + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "inline_equation", + "content": "^\\ddagger" + }, + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "text", + "content": "We count objects used in any of the tasks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 201, + 472, + 301 + ], + "lines": [ + { + "bbox": [ + 138, + 201, + 472, + 301 + ], + "spans": [ + { + "bbox": [ + 138, + 201, + 472, + 301 + ], + "type": "table", + "html": "
SimulatorHabitat2.0AI2THOR
DatasetReplicaCAD [67]RoboTHOR [16]ALFRED [61]REALFRED
# Scenes11175120150
Total floor area (m2)8,824.52,5742,55510,060
Total nav. area (m2)-1,2581,3564,251
Nav. complex.-2.0362.5493.020
Scene clutter-8.0955.1198.072
# object class9214†82‡112‡
", + "image_path": "683e351bcc28f1646523522a89c85a12c212416e40ebeaf04eef4bd0e1fdc76d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 326, + 479, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 479, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 479, + 350 + ], + "type": "text", + "content": "areas, with larger areas on average. This implies that an agent needs the ability to navigate effectively in spaces of varying sizes, generally wider on average." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 351, + 482, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 482, + 424 + ], + "type": "text", + "content": "We also investigate the spatial characteristics of each benchmark [16, 61, 67] and summarize the result in Table 2. 'Total floor area' denotes the sum of all floor areas and 'Total navigable area' denotes the sum of all navigable areas in the dataset. We observe that the REALFRED benchmark has the highest total floor area and navigable area value, implying that REALFRED provides more navigation space for an agent than previous work [16, 61, 67]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 425, + 482, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 520 + ], + "type": "text", + "content": "Spatial complexity. We now investigate the complexity of spatial structures in the REALFRED benchmark using several metrics. To ensure a fair comparison with other datasets [16, 61, 67], we employ the navigation complexity introduced by [74] and the scene clutter measurement from [56]. A higher navigation complexity indicates an increased difficulty in navigating through the space, while a higher scene clutter implies the presence of more obstacles in the environment. By utilizing these metrics, we conduct a comparative analysis with object-interactable benchmarks [16, 61] and provide the result in Table 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "We observe that the REALFRED benchmark provides the environment with higher navigation complexity and scene clutter compared to other benchmarks [61, 67]. The high navigation complexity in our scenes stems from their multi-room composition. This setup requires the agent to execute more intricate navigation when moving from one room to another, in comparison to scenarios where the agent operates within a single room. We observe that ours poses the second-highest scene clutter with a marginal gap from [16]. This is because the spaces in [16] are relatively confined with a high density of furniture. We observe that ours has a similar value to [16], meaning that our scenes have a large amount of obstacles with a similar portion of [16] since our scenes' average size is larger than that of [16]. This implies that the REALFRED benchmark provides a complex and challenging space for the agent to explore the environment." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "text", + "content": "T. Kim and C. Min et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 132, + 239, + 240 + ], + "blocks": [ + { + "bbox": [ + 193, + 114, + 214, + 124 + ], + "lines": [ + { + "bbox": [ + 193, + 114, + 214, + 124 + ], + "spans": [ + { + "bbox": [ + 193, + 114, + 214, + 124 + ], + "type": "text", + "content": "Train" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 143, + 132, + 239, + 240 + ], + "lines": [ + { + "bbox": [ + 143, + 132, + 239, + 240 + ], + "spans": [ + { + "bbox": [ + 143, + 132, + 239, + 240 + ], + "type": "image", + "image_path": "f41c23831bd75a97a3598e7354dbde795a606657acc32dcd95152ffd4b8184fc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 247, + 482, + 270 + ], + "lines": [ + { + "bbox": [ + 131, + 247, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 131, + 247, + 482, + 270 + ], + "type": "text", + "content": "Fig. 5: The seven types of tasks' distribution in REALFRED. We provide " + }, + { + "bbox": [ + 131, + 247, + 482, + 270 + ], + "type": "inline_equation", + "content": "37.6\\%" + }, + { + "bbox": [ + 131, + 247, + 482, + 270 + ], + "type": "text", + "content": " more tasks in valid sets and " + }, + { + "bbox": [ + 131, + 247, + 482, + 270 + ], + "type": "inline_equation", + "content": "19.3\\%" + }, + { + "bbox": [ + 131, + 247, + 482, + 270 + ], + "type": "text", + "content": " in total compared to previous benchmark [61]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 240, + 133, + 321, + 240 + ], + "blocks": [ + { + "bbox": [ + 265, + 115, + 306, + 124 + ], + "lines": [ + { + "bbox": [ + 265, + 115, + 306, + 124 + ], + "spans": [ + { + "bbox": [ + 265, + 115, + 306, + 124 + ], + "type": "text", + "content": "Valid Seen" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 240, + 133, + 321, + 240 + ], + "lines": [ + { + "bbox": [ + 240, + 133, + 321, + 240 + ], + "spans": [ + { + "bbox": [ + 240, + 133, + 321, + 240 + ], + "type": "image", + "image_path": "e0180f8fb7c3f0ef21b0fb068d837a3069c1812cee376ed45c70b4a5467ad558.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 323, + 128, + 404, + 239 + ], + "blocks": [ + { + "bbox": [ + 342, + 114, + 393, + 124 + ], + "lines": [ + { + "bbox": [ + 342, + 114, + 393, + 124 + ], + "spans": [ + { + "bbox": [ + 342, + 114, + 393, + 124 + ], + "type": "text", + "content": "Valid Unseen" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 323, + 128, + 404, + 239 + ], + "lines": [ + { + "bbox": [ + 323, + 128, + 404, + 239 + ], + "spans": [ + { + "bbox": [ + 323, + 128, + 404, + 239 + ], + "type": "image", + "image_path": "073226163b503917b3560c3247eee85d7b00959533c3443860d60d2f0076a1e0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 408, + 127, + 471, + 222 + ], + "blocks": [ + { + "bbox": [ + 408, + 127, + 471, + 222 + ], + "lines": [ + { + "bbox": [ + 408, + 127, + 471, + 222 + ], + "spans": [ + { + "bbox": [ + 408, + 127, + 471, + 222 + ], + "type": "image", + "image_path": "82156fb55d9223338d7119ea5a9190dcf6ba80a128fd00e8ba871738d77ff374.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 294, + 334, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 294, + 334, + 306 + ], + "spans": [ + { + "bbox": [ + 131, + 294, + 334, + 306 + ], + "type": "text", + "content": "3.2 Expert Demonstration Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 316, + 480, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 480, + 375 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 480, + 375 + ], + "type": "text", + "content": "Each expert demonstration includes a set of an egocentric RGB view and action information with an interaction mask if exists at each time step. Expert demonstrations for each task are generated by a planner [29] with encoded state spaces into Planning Domain Definition Language (PDDL) rules [22]. To generate household tasks, we utilize seven task types introduced in [61]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 376, + 317, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 317, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 317, + 472 + ], + "type": "text", + "content": "Data splits. We split the generated demonstrations into train, validation, and test folds. Specifically, we designate 135 scenes for seen and 15 scenes for unseen fold. Comparison of the amount of each task and the distribution across training and validation folds with previous work [61] is shown in Fig. 5." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 472, + 317, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 317, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 317, + 604 + ], + "type": "text", + "content": "Curating free-form language instructions. Detailed language instructions describe a task that involves a sequence of actions, a point of interest in robotics. The REALFRED benchmark offers 30,696 language directives, each comprising a human-annotated high-level goal and a set of step-by-step instructions. These directives are collected from 93 Amazon Mechanical Turk workers with a 'Master' qualification, ensuring high" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 322, + 376, + 482, + 536 + ], + "blocks": [ + { + "bbox": [ + 322, + 376, + 482, + 536 + ], + "lines": [ + { + "bbox": [ + 322, + 376, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 322, + 376, + 482, + 536 + ], + "type": "image", + "image_path": "048bd0d8c9cef05a5b1bc06fcf3e4ed42f72571eff1eb60e49aa647fc0d3418e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 544, + 482, + 599 + ], + "lines": [ + { + "bbox": [ + 321, + 544, + 482, + 599 + ], + "spans": [ + { + "bbox": [ + 321, + 544, + 482, + 599 + ], + "type": "text", + "content": "Fig.6:Word distribution.The words are arranged starting from the center and extending outwards. The arc length corresponds to the word frequency in the instructions." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 604, + 482, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 604, + 482, + 664 + ], + "spans": [ + { + "bbox": [ + 130, + 604, + 482, + 664 + ], + "type": "text", + "content": "quality. Collected annotations are validated through an additional voting survey, and invalid instructions are replaced with newly collected instructions. The distribution of language instructions by their first four words is presented in Fig. 6. We provide a detailed annotation process and examples of an expert demonstration with the instruction in Sec. A.1 and A.4 in the supplementary material." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "text", + "content": "REALFRED" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 315, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 315, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 315, + 282 + ], + "type": "text", + "content": "Multi-room embodied instruction following. For household robots to be able to effectively assist humans, it would be more practical to deploy them in complex and diverse environments with various room types, rather than confining them to a single room. Cross-room navigation poses significant challenges, requiring agents to understand room references, plan efficient paths, and overcome obstacles proficiently. This agent is further tasked to interpret visual cues, demonstrate a keen sense of spatial awareness, and process natural language instructions." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 323, + 116, + 482, + 213 + ], + "blocks": [ + { + "bbox": [ + 323, + 116, + 482, + 213 + ], + "lines": [ + { + "bbox": [ + 323, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 323, + 116, + 482, + 213 + ], + "type": "image", + "image_path": "291eec7c70ead6d17fe86c455078a5e6226156db1f64a610f18d0a844d9a396a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 220, + 482, + 277 + ], + "lines": [ + { + "bbox": [ + 321, + 220, + 482, + 277 + ], + "spans": [ + { + "bbox": [ + 321, + 220, + 482, + 277 + ], + "type": "text", + "content": "Fig.7: Comparison of expert demonstrations (ALFRED vs. REALFRED). Due to expanded environments, longer steps and planning horizons are required in REALFRED." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 283, + 480, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 480, + 331 + ], + "type": "text", + "content": "The distribution of the number of steps and the length covered by the expert demonstrations is shown in Fig. 7. We observe that longer steps and trajectories are required to complete our tasks compared to the single-room constraint benchmark [61], meaning that our benchmark provides longer-horizon tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 331, + 481, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 481, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 481, + 416 + ], + "type": "text", + "content": "To broaden a range of tasks, we provide an extensive number of object class types in the REALFRED benchmark. More object classes are added to the REALFRED benchmark, making it a superset of [61] with 86 pickable and 26 receptacle objects. This has resulted in a more diverse set of tasks, with the number of unique tasks being " + }, + { + "bbox": [ + 130, + 331, + 481, + 416 + ], + "type": "inline_equation", + "content": "84.3\\%" + }, + { + "bbox": [ + 130, + 331, + 481, + 416 + ], + "type": "text", + "content": " more than the ALFRED benchmark [61], and the proportion of unique tasks being higher in the REALFRED benchmark. We provide detailed information in Sec. A in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 431, + 230, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 230, + 444 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 230, + 444 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 453, + 484, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 484, + 537 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 484, + 537 + ], + "type": "text", + "content": "Metrics. We follow the same evaluation protocol of the ALFRED benchmark [61]. The primary metric is 'Success Rate (SR)' which measures the percentage of completed tasks. 'Goal-Condition Success Rate (GC)' measures the percentage of achieved goal conditions. We also use path-length-weighted metrics to measure how efficiently an agent completes tasks. For more details, kindly refer to [61]. Baselines. We evaluate several recent state-of-the-art methods [5,33,34,46,61-63] with competitive results in [61]. We provide more details in Sec. B." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 552, + 343, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 552, + 343, + 565 + ], + "spans": [ + { + "bbox": [ + 131, + 552, + 343, + 565 + ], + "type": "text", + "content": "4.1 Comparison of the State of the Arts" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 480, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 480, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 480, + 605 + ], + "type": "text", + "content": "We evaluate the baselines in the proposed REALFRED benchmark over multiple runs and present the average result in Table 3. We report extended results with path-length-weighted metrics in Sec. C in supplementary." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 606, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 481, + 665 + ], + "type": "text", + "content": "For a fair comparison, we separate the baseline into two groups based on the use of additional depth supervision for semantic map reconstruction: 'Imitation Learning' where agents learn direct mapping from visual observations and language instructions to action sequences and 'Spatial Map Reconst.' where agents plan action sequences based on reconstructed semantic spatial representations." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "text", + "content": "T. Kim and C. Min et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 201, + 480, + 296 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "text", + "content": "Table 3: Task and Goal-Condition Success Rate. We train and evaluate recent state-of-the-art methods on our REALFRED benchmark. For a fair comparison, we group these methods into two based on the usage of extra depth supervision: models learned by imitation learning without depth supervision ('Imitation Learning') and ones that maintain semantic spatial representations constructed by predicted depth maps ('Spatial Map Reconst.'). Path-length-weighted (PLW) metrics are reported in Sec. C for each value. " + }, + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "text", + "content": "Authors' implementation as the code is not publicly available." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 201, + 480, + 296 + ], + "lines": [ + { + "bbox": [ + 134, + 201, + 480, + 296 + ], + "spans": [ + { + "bbox": [ + 134, + 201, + 480, + 296 + ], + "type": "table", + "html": "
LearningModelValidationTest
SeenUnseenSeenUnseen
SRGCSRGCSRGCSRGC
Imitation LearningSeq2Seq [61]0.77 ± 0.066.93 ± 0.060.00 ± 0.004.03 ± 0.001.10 ± 0.006.60 ± 0.000.00 ± 0.003.50 ± 0.00
MOCA [62]12.64 ± 0.1220.95 ± 0.181.44 ± 0.056.76 ± 0.0414.11 ± 0.0322.84 ± 0.040.62 ± 0.085.14 ± 0.08
ABP† [33]24.71 ± 0.0533.80 ± 0.144.22 ± 0.0511.71 ± 0.2727.44 ± 0.4035.81 ± 0.233.54 ± 0.2310.57 ± 0.22
Spatial Map Reconst.HLSM [5]4.23 ± 0.089.14 ± 0.091.08 ± 0.146.12 ± 0.236.27 ± 0.0410.44 ± 0.130.49 ± 0.164.28 ± 0.13
FILM [46]7.08 ± 0.2811.93 ± 0.234.44 ± 0.179.26 ± 0.138.79 ± 0.0713.03 ± 0.082.15 ± 0.186.56 ± 0.15
LLM-Planner† [63]5.80 ± 0.1911.69 ± 0.353.33 ± 0.228.29 ± 0.198.16 ± 0.2013.20 ± 0.131.90 ± 0.136.33 ± 0.02
CAPEAM† [34]13.45 ± 0.0518.16 ± 0.274.92 ± 0.229.47 ± 0.2315.61 ± 0.1520.22 ± 0.112.87 ± 0.137.36 ± 0.07
Human------85.00 ± 3.5491.30 ± 2.94
", + "image_path": "e4a9e400ba27c496621b6d55dc9ad77099f8f46ea37c8fe0e49014c8955e49c8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 319, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 426 + ], + "type": "text", + "content": "We observe that all these baselines, 'Imitation Learning' and 'Spatial Map Reconst.' consistently achieve lower performance values for all metrics in both seen and unseen splits compared to performances achieved in [61], implying that our proposed REALFRED provide more challenges compared to [61]. While in [61], the 'Spatial Map Reconst.' baselines [5,34,46] outperform learning-based approaches [33,61,62] by exploiting semantic spatial representation with deterministic algorithms (e.g., obstacle-free navigation path planning [60]), we observe a contrasting result that the 'Imitation Learning' baselines outperforming the 'Spatial Map Reconst.' baselines in our REALFRED benchmark." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 427, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 427, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 427, + 482, + 559 + ], + "type": "text", + "content": "We qualitatively observe that such a confined spatial map reconstruction is due to a limited field of view and map reconstruction methods. The agent's limited field of view often leads to failure in recognizing room corners with doors or narrow aisles, resulting in (single-room-sized) limited map reconstruction. In addition, [34, 46] perceives obstacles larger than they actually are for better obstacle-free path planning by sacrificing navigable area, but this can be quite critical for narrow doors and aisles as they have a small amount of navigable space. This may hinder navigation to other rooms and thus, fail at tasks. We provide a more detailed discussion, supported by figures, on the agents' difficulty in recognizing narrow passages and the resulting impact on the agents, who struggle with spatial map reconstruction in Sec. D in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 574, + 432, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 574, + 432, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 574, + 432, + 586 + ], + "type": "text", + "content": "4.2 Comparison to the agents with sim-to-real adaptation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "We investigate the transfer from simulation training to real-world scan evaluation (sim-to-real) and from real-world scan training to real-world scan evaluation (real-to-real) with [33]. We train a sim-to-real agent with synthetic visual data and a real-to-real agent with real scanned visual data, respectively. During inference, both agents predict an action and an object mask based on the scanned visual input frame and the given language instruction at every time step." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "text", + "content": "REALFRED" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "Training and evaluation data selection. For training the sim-to-real agent, we use the training dataset in [61], encompassing " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "21K" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " language annotations. For a fair comparison, we train the real-to-real agent with tasks from the REALFRED benchmark's training fold, specifically involving the manipulation of the same objects that were used in training the sim-to-real agent with " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "19K" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " language annotations. For evaluation, we select tasks from the valid unseen splits of our REALFRED benchmark. We evaluate the agent performance with the tasks that 1) do not require multi-room navigation 2) and those that do. This selection specifically includes tasks that feature objects used in the training phase." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 482, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 342 + ], + "type": "text", + "content": "Real-to-sim domain adaptation. The use of Generative Adversarial Networks (GANs) [23,79] for domain adaptation has recently been examined in the literature on robotics [6,28,32,58,77]. These models are employed to adapt input images from real-world domains (i.e. Real " + }, + { + "bbox": [ + 130, + 224, + 482, + 342 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 224, + 482, + 342 + ], + "type": "text", + "content": " Sim) before they are passed to the agent policy. We train a CycleGan [79] and its off-the-shelf variant [70] with unpaired images collected from the REALFRED (real domain) and ALFRED benchmarks [61] (simulated domain). Among the trained generators, the one performing real to simulation conversion, referred to as the real-to-sim goggle, learns the mapping from the real domain to the simulation domain " + }, + { + "bbox": [ + 130, + 224, + 482, + 342 + ], + "type": "inline_equation", + "content": "G_{S}:\\mathcal{R}\\to \\mathcal{S}" + }, + { + "bbox": [ + 130, + 224, + 482, + 342 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 224, + 482, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 130, + 224, + 482, + 342 + ], + "type": "text", + "content": " denotes the real domain and " + }, + { + "bbox": [ + 130, + 224, + 482, + 342 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 224, + 482, + 342 + ], + "type": "text", + "content": " denotes the simulated domain." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 343, + 317, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 317, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 317, + 486 + ], + "type": "text", + "content": "Results. We present the results of the sim-to-real experiments in Table 4. We report the performance of agents when tasks 1) require agents to navigate multirooms, denoted as 'Multi + Single,' and 2) are solvable within a single room, denoted as 'Single only.' Firstly, we observe that the sim-to-real agent significantly underperforms its real-to-real counterpart in all metrics " + }, + { + "bbox": [ + 130, + 343, + 317, + 486 + ], + "type": "inline_equation", + "content": "(\\# (a)" + }, + { + "bbox": [ + 130, + 343, + 317, + 486 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 130, + 343, + 317, + 486 + ], + "type": "inline_equation", + "content": "\\# (d))" + }, + { + "bbox": [ + 130, + 343, + 317, + 486 + ], + "type": "text", + "content": " .We then compare the results with gobbled agents. By comparing agents evaluated in the Multi" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 324, + 423, + 482, + 474 + ], + "blocks": [ + { + "bbox": [ + 321, + 355, + 482, + 422 + ], + "lines": [ + { + "bbox": [ + 321, + 355, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 321, + 355, + 482, + 422 + ], + "type": "text", + "content": "Table 4: Comparison to the agents with sim-to-real adaptation. Each 'Sim2Real' and 'Real2Real' denotes an agent trained in simulated environments and 3D-captured scenes. 'Goggle' denotes real-to-sim methods." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 423, + 482, + 474 + ], + "lines": [ + { + "bbox": [ + 324, + 423, + 482, + 474 + ], + "spans": [ + { + "bbox": [ + 324, + 423, + 482, + 474 + ], + "type": "table", + "html": "
# SettingGoggleMulti + SingleSingle only
SRPLWSRSRPLWSR
(a) Sim2RealNone0.1150.0120.00.0
(b) Sim2RealCycleGan [79]0.1150.0160.3270.065
(c) Sim2RealUVCGAN-v2 [70]0.1150.0460.3270.187
(d) Real2RealNone2.4050.7852.6140.762
", + "image_path": "ec66edd07ae13102485f977aea58e22cd1ce658df2695f350600e686750f691a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 487, + 482, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 523 + ], + "type": "text", + "content": "+ Single' tasks, we do not observe improvements in the main metric, success rate (SR). We observe a slight increase in the PLWSR metric for the gobbled sim-to-real agent compared to the vanilla sim-to-real agent (#(a) vs. #(b,c))." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 523, + 482, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 523, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 523, + 482, + 583 + ], + "type": "text", + "content": "To isolate the impacts of the visual domain gap, we then compare agents evaluated on 'Single only' tasks. We observe that both gobbled agents show improvements over the vanilla sim-to-real agent which results in zero success rate. However, a noticeable performance gap exists with the real-to-real agent, implying the need for learning in real scanned environments " + }, + { + "bbox": [ + 130, + 523, + 482, + 583 + ], + "type": "inline_equation", + "content": "(\\# (d)" + }, + { + "bbox": [ + 130, + 523, + 482, + 583 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 130, + 523, + 482, + 583 + ], + "type": "inline_equation", + "content": "\\# (a,b,c))" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 598, + 298, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 298, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 298, + 611 + ], + "type": "text", + "content": "4.3 Challenges in REALFRED" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "We propose two hypotheses for the low performance observed with state-of-the-art methods on the REALFRED benchmark: (1) navigating within larger scenes and (2) overcoming narrow pathways between rooms. These elements represent challenges of completing tasks within multi-room, household-scale environments." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "text", + "content": "T. Kim and C. Min et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 478, + 216 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 478, + 216 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 478, + 216 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 478, + 216 + ], + "type": "image", + "image_path": "3eaaf4466a90b882d59174864fcbaf743526bbd11465a420d2d8e479776af21d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 224, + 482, + 291 + ], + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 291 + ], + "type": "text", + "content": "Fig. 8: Distribution of collision spots. The heatmap illustrates collision frequency on each spot and the black outline depicts a navigable area. We illustrate collision points of [33] in valid unseen environments to represent the distribution. Specifically, we choose failure episodes from the two largest scenes in the valid unseen split. We observe a remarkable concentration of collisions at the entrance to the rooms. The corresponding egocentric views are presented on the right side of the figure." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 316, + 482, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 400 + ], + "type": "text", + "content": "We analyze failure cases with the best-performing model, ABP [33], on valid unseen split, since it seems the most promising on the REALFRED benchmark. Difficulty in large environments. Understanding the surrounding environment, including the location of objects and the positions of obstacles, can be beneficial for task completion. However, the agent's visual range is bounded, requiring it to expend more steps for exploration in order to perceive larger spaces. This intensifies the challenge of navigating larger spaces with limited steps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 401, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 509 + ], + "type": "text", + "content": "We conduct an analysis to see the relationship between the success rate (SR) and the size of the space. We follow the interquartile range (IQR) method to set a threshold at " + }, + { + "bbox": [ + 130, + 401, + 482, + 509 + ], + "type": "inline_equation", + "content": "30.44m^2" + }, + { + "bbox": [ + 130, + 401, + 482, + 509 + ], + "type": "text", + "content": " that covers the most navigable space sizes in the ALFRED [61]. This threshold marks the upper fence, defining spaces above it as outliers. We classify spaces in the REALFRED benchmark smaller than threshold as smaller scenes, and those larger as larger scenes. Results indicate that the agent [33] with the highest SR in the valid unseen fold showed an average SR of " + }, + { + "bbox": [ + 130, + 401, + 482, + 509 + ], + "type": "inline_equation", + "content": "5.46\\%" + }, + { + "bbox": [ + 130, + 401, + 482, + 509 + ], + "type": "text", + "content": " in the smaller scenes and only " + }, + { + "bbox": [ + 130, + 401, + 482, + 509 + ], + "type": "inline_equation", + "content": "1.77\\%" + }, + { + "bbox": [ + 130, + 401, + 482, + 509 + ], + "type": "text", + "content": " in the larger scenes. This implies that solving tasks becomes more challenging as the space size increases." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 640 + ], + "type": "text", + "content": "In addition, we compare the difficulty of navigation between the previous [61] and REALFRED benchmarks with different average spatial sizes. To quantify, we first define milestones for each task, which are spots to be reached to interact with target objects (e.g., Apple, Knife, etc.). We consider navigation to be a success only if all milestones are visited, regardless of whether the actual interaction (e.g., slicing an apple) is performed or not. Consequently, the navigation success rate in the REALFRED benchmark is merely " + }, + { + "bbox": [ + 130, + 510, + 482, + 640 + ], + "type": "inline_equation", + "content": "59.18\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 640 + ], + "type": "text", + "content": " for the valid unseen split while the agent's [33] navigation success rate is " + }, + { + "bbox": [ + 130, + 510, + 482, + 640 + ], + "type": "inline_equation", + "content": "84.82\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 640 + ], + "type": "text", + "content": " in ALFRED [61]. The REALFRED benchmark offering an average space size more than three times larger than the previous work [61] implies the need for model development capable of overcoming the challenges associated with this increased scale." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Navigation through narrow doorways. The REALFRED benchmark supports environments with multiroom composition, unlike the previous dataset [61]" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "text", + "content": "REALFRED" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "content": "providing a single room scale. Specifically, it contains narrow doorways across rooms, which may hinder the agent from navigating through. Here, we hypothesize that the agent would frequently collide with the walls near the door." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 481, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 481, + 247 + ], + "type": "text", + "content": "We investigate collision spots and showcase two examples on top of the scene's layout in valid unseen fold in Fig. 8. The number of collisions in failed cases is accumulated and normalized to the maximum number of collisions, respectively, to indicate the collision frequency at each point. The black outline represents the agent's navigable area, as detailed in Sec. 3.1. Each value denotes the normalized collision frequency. In both scenes, we observe that collisions often occur on walls near the door, implying that our spatial characteristic (i.e., including narrow spaces) may hinder the agent from properly navigating without collision." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 248, + 481, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 481, + 296 + ], + "type": "text", + "content": "Human evaluation. Following [61], we randomly select 100 directives from the test unseen fold and have them evaluated by humans. Five participants are given 20 tasks each, which they complete using a keyboard-and-mouse interface. Before starting, they are given the opportunity to become familiar with the interface." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 297, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 380 + ], + "type": "text", + "content": "Participants achieve a comparable high success rate of " + }, + { + "bbox": [ + 130, + 297, + 482, + 380 + ], + "type": "inline_equation", + "content": "85\\%" + }, + { + "bbox": [ + 130, + 297, + 482, + 380 + ], + "type": "text", + "content": ", along with a goal-condition success rate of " + }, + { + "bbox": [ + 130, + 297, + 482, + 380 + ], + "type": "inline_equation", + "content": "91.30\\%" + }, + { + "bbox": [ + 130, + 297, + 482, + 380 + ], + "type": "text", + "content": ", in average. We agree that human performance may appear slightly lower compared to the results presented in previous work [57, 61, 78]. This is partly because human participants encountered several difficulties when controlling the agent, particularly in avoiding collisions within narrow corridors. Furthermore, navigating large spaces with a limited egocentric field of view introduced additional challenges, leading to task failures." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 399, + 220, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 399, + 220, + 412 + ], + "spans": [ + { + "bbox": [ + 132, + 399, + 220, + 412 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 426, + 481, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 481, + 510 + ], + "type": "text", + "content": "We present REALFRED, a new dataset and benchmark for embodied instruction following task on 3D-captured environments. We capture 150 indoor houses in 3D with interactable objects to enable complex household tasks. The reconstructed indoor scenes provide a larger spatial area and complex multi-room environments that are close to the real-world scenario and challenging for an agent to successfully complete a task. Expert demonstrations are also provided along with free-form human-language instructions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 582 + ], + "type": "text", + "content": "In our empirical evaluations, we show that state-of-the-art methods struggle in large multi-room environments, provide analyses of our newly proposed benchmark, and perform Sim2Real transfer experiments. We have released our Embodied AI research data and code for reproducibility. We expect that the REALFRED benchmark will encourage further research on developing robotic agents that execute household tasks by language instructions in the real world." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": "Limitation and future work. Although we support a large number of interactable objects, the types of tasks to be completed are rather limited, considering more complex real-world scenarios. In addition, we currently address natural language in English but users may come from different regions with different languages. We can think of two future research avenues as follows. (1) adding additional complicated types of task that require both hands to complete. (2) supporting a multi-lingual interface for users from different regions." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "text", + "content": "T. Kim and C. Min et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 212 + ], + "type": "text", + "content": "This work was partly supported by the NRF grant (No.2022R1A2C400230012, " + }, + { + "bbox": [ + 130, + 139, + 482, + 212 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 139, + 482, + 212 + ], + "type": "text", + "content": " ) and IITP grants (No.RS-2022-II220077 " + }, + { + "bbox": [ + 130, + 139, + 482, + 212 + ], + "type": "inline_equation", + "content": "(5\\%)" + }, + { + "bbox": [ + 130, + 139, + 482, + 212 + ], + "type": "text", + "content": " , No.RS-2022-II220113 " + }, + { + "bbox": [ + 130, + 139, + 482, + 212 + ], + "type": "inline_equation", + "content": "(5\\%)" + }, + { + "bbox": [ + 130, + 139, + 482, + 212 + ], + "type": "text", + "content": " No.RS-2022-II220959 " + }, + { + "bbox": [ + 130, + 139, + 482, + 212 + ], + "type": "inline_equation", + "content": "(5\\%)" + }, + { + "bbox": [ + 130, + 139, + 482, + 212 + ], + "type": "text", + "content": " , No.RS-2022-II220871 " + }, + { + "bbox": [ + 130, + 139, + 482, + 212 + ], + "type": "inline_equation", + "content": "(15\\%)" + }, + { + "bbox": [ + 130, + 139, + 482, + 212 + ], + "type": "text", + "content": " , No.RS-2020-II201361 " + }, + { + "bbox": [ + 130, + 139, + 482, + 212 + ], + "type": "inline_equation", + "content": "(5\\%)" + }, + { + "bbox": [ + 130, + 139, + 482, + 212 + ], + "type": "text", + "content": " , Yonsei AI), No.RS-2021-II211343 " + }, + { + "bbox": [ + 130, + 139, + 482, + 212 + ], + "type": "inline_equation", + "content": "(5\\%)" + }, + { + "bbox": [ + 130, + 139, + 482, + 212 + ], + "type": "text", + "content": " , SNU AI), No.RS-2021-II212068 " + }, + { + "bbox": [ + 130, + 139, + 482, + 212 + ], + "type": "inline_equation", + "content": "(5\\%)" + }, + { + "bbox": [ + 130, + 139, + 482, + 212 + ], + "type": "text", + "content": " , AI Innov. Hub), No.RS-2022-II220951 " + }, + { + "bbox": [ + 130, + 139, + 482, + 212 + ], + "type": "inline_equation", + "content": "(50\\%)" + }, + { + "bbox": [ + 130, + 139, + 482, + 212 + ], + "type": "text", + "content": ") funded by the Korea government(MSIT)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 228, + 197, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 228, + 197, + 241 + ], + "spans": [ + { + "bbox": [ + 133, + 228, + 197, + 241 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 251, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 251, + 480, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 480, + 285 + ], + "type": "text", + "content": "1. Anderson, P., Wu, Q., Teney, D., Bruce, J., Johnson, M., Sünderhauf, N., Reid, I., Gould, S., van den Hengel, A.: Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In: CVPR (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 285, + 480, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 480, + 328 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 480, + 328 + ], + "type": "text", + "content": "2. Baruch, G., Chen, Z., Dehghan, A., Dimry, T., Feigin, Y., Fu, P., Gebauer, T., Joffe, B., Kurz, D., Schwartz, A., Shulman, E.: ARKitscenes - a diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In: NeurIPS Datasets and Benchmarks Track (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 329, + 480, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 329, + 480, + 350 + ], + "spans": [ + { + "bbox": [ + 138, + 329, + 480, + 350 + ], + "type": "text", + "content": "3. Berseth, G., Xie, C., Cernek, P., van de Panne, M.: Progressive reinforcement learning with distillation for multi-skilled motion control. In: ICLR (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 350, + 480, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 350, + 480, + 371 + ], + "spans": [ + { + "bbox": [ + 138, + 350, + 480, + 371 + ], + "type": "text", + "content": "4. Binkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying mmd gans. In: ICLR (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 372, + 480, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 372, + 480, + 404 + ], + "spans": [ + { + "bbox": [ + 138, + 372, + 480, + 404 + ], + "type": "text", + "content": "5. Blukis, V., Paxton, C., Fox, D., Garg, A., Artzi, Y.: A persistent spatial semantic representation for high-level natural language instruction execution. In: CoRL (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 404, + 480, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 404, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 138, + 404, + 480, + 437 + ], + "type": "text", + "content": "6. Bousmalis, K., Irpan, A., Wohlhart, P., Bai, Y., Kelcey, M., Kalakrishnan, M., Downs, L., Ibarz, J., Pastor, P., Konolige, K., et al.: Using simulation and domain adaptation to improve efficiency of deep robotic grasping. In: ICRA (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 437, + 480, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 437, + 480, + 469 + ], + "spans": [ + { + "bbox": [ + 138, + 437, + 480, + 469 + ], + "type": "text", + "content": "7. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3d: Learning from rgb-d data in indoor environments. arXiv:1709.06158 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 470, + 480, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 470, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 470, + 480, + 491 + ], + "type": "text", + "content": "8. Chaplot, D.S., Gandhi, D., Gupta, S., Gupta, A., Salakhutdinov, R.: Learning to explore using active neural slam. In: ICLR (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 492, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 492, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 492, + 480, + 512 + ], + "type": "text", + "content": "9. Chaplot, D.S., Gandhi, D.P., Gupta, A., Salakhutdinov, R.R.: Object goal navigation using goal-oriented semantic exploration. In: NeurIPS (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 514, + 480, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 480, + 545 + ], + "type": "text", + "content": "0. Chen, C., Jain, U., Schissler, C., Gari, S.V.A., Al-Halah, Z., Ithapu, V.K., Robinson, P., Grauman, K.: Soundspaces: Audio-visual navigation in 3d environments. In: ECCV (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 546, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 546, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 546, + 480, + 578 + ], + "type": "text", + "content": "1. Chen, C., Schissler, C., Garg, S., Kobernik, P., Clegg, A., Calamia, P., Batra, D., Robinson, P.W., Grauman, K.: Soundspaces 2.0: A simulation platform for visual-acoustic learning. In: NeurIPS Datasets and Benchmarks Track (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 578, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 600 + ], + "type": "text", + "content": "2. Chen, H., Suhr, A., Misra, D., Snavely, N., Artzi, Y.: Touchdown: Natural language navigation and spatial reasoning in visual street environments. In: CVPR (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 601, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 621 + ], + "type": "text", + "content": "3. Chen, T., Gupta, S., Gupta, A.: Learning exploration policies for navigation. In: ICLR (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "text", + "content": "4. Choi, S., Ji, G., Park, J., Kim, H., Mun, J., Lee, J.H., Hwangbo, J.: Learning quadrupedal locomotion on deformable terrain. Science Robotics (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "text", + "content": "5. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "text", + "content": "REALFRED" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "16. Deitke, M., Han, W., Herrasti, A., Kembhavi, A., Kolve, E., Mottaghi, R., Salvador, J., Schwenk, D., VanderBilt, E., Wallingford, M., et al.: Robothor: An open simulation-to-real embodied ai platform. In: CVPR (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "text", + "content": "17. Deitke, M., Hendrix, R., Farhadi, A., Ehsani, K., Kembhavi, A.: Phone2proc: Bringing robust robots into our chaotic world. In: CVPR (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "text", + "content": "18. Deitke, M., VanderBilt, E., Herrasti, A., Weihs, L., Ehsani, K., Salvador, J., Han, W., Kolve, E., Kembhavi, A., Mottaghi, R.: Procthor: Large-scale embodied ai using procedural generation. In: NeurIPS (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "type": "text", + "content": "19. Ehsani, K., Han, W., Herrasti, A., Vanderbilt, E., Weihs, L., Kolve, E., Kembhavi, A., Mottaghi, R.: Manipulathor: A framework for visual object manipulation. In: CVPR (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 237, + 480, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 237, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 133, + 237, + 480, + 282 + ], + "type": "text", + "content": "20. Gan, C., Schwartz, J., Alter, S., Mrowca, D., Schrimpf, M., Traer, J., De Freitas, J., Kubilius, J., Bhandwaldar, A., Haber, N., et al.: Threadworld: A platform for interactive multi-modal physical simulation. In: NeurIPS Datasets and Benchmarks Track (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 304 + ], + "type": "text", + "content": "21. Gao, X., Gao, Q., Gong, R., Lin, K., Thattai, G., Sukhatme, G.S.: Dialfred: Dialogue-enabled agents for embodied instruction following. RA-L (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 326 + ], + "type": "text", + "content": "22. Ghallab, M., Howe, A., Knoblock, C., McDermott, D., Ram, A., Veloso, M., Weld, D., Wilkins, D.: Pddl the planning domain definition language (1998)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 347 + ], + "type": "text", + "content": "23. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: NeurIPS (2014)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "text", + "content": "24. Gordon, D., Kembhavi, A., Rastegari, M., Redmon, J., Fox, D., Farhadi, A.: Iqa: Visual question answering in interactive environments. In: CVPR (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "25. Gu, J., Xiang, F., Li, X., Ling, Z., Liu, X., Mu, T., Tang, Y., Tao, S., Wei, X., Yao, Y., et al.: Maniskill2: A unified benchmark for generalizable manipulation skills. In: ICLR (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 402, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 480, + 424 + ], + "type": "text", + "content": "26. Heo, M., Lee, Y., Lee, D., Lim, J.J.: Furniturebench: Reproducible real-world benchmark for long-horizon complex manipulation. In: RSS (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "text", + "content": "27. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 456, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 480, + 479 + ], + "type": "text", + "content": "28. Ho, D., Rao, K., Xu, Z., Jang, E., Khansari, M., Bai, Y.: Retinagan: An object-aware approach to sim-to-real transfer. In: ICRA (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "text", + "content": "29. Hoffmann, J., Nebel, B.: The ff planning system: Fast plan generation through heuristic search. In: JAIR (2001)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 501, + 480, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 523 + ], + "type": "text", + "content": "30. Hua, B.S., Pham, Q.H., Nguyen, D.T., Tran, M.K., Yu, L.F., Yeung, S.K.: Scenenn: A scene meshes dataset with annotations. In: 3DV (2016)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 523, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 567 + ], + "type": "text", + "content": "31. Khanna, M., Mao, Y., Jiang, H., Haresh, S., Shacklett, B., Batra, D., Clegg, A., Undersander, E., Chang, A.X., Savva, M.: Habitat synthetic scenes dataset (hssd200): An analysis of 3d scene scale and realism tradeoffs for objectgoal navigation. In: CVPR (2024)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "text", + "content": "32. Khansari, M., Ho, D., Du, Y., Fuentes, A., Bennice, M., Sievers, N., Kirmani, S., Bai, Y., Jang, E.: Practical visual deep imitation learning via task-level domain consistency. In: ICRA (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "33. Kim, B., Bhambri, S., Singh, K.P., Mottaghi, R., Choi, J.: Agent with the big picture: Perceiving surroundings for interactive instruction following. In: Embodied AI Workshop @ CVPR (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "34. Kim, B., Kim, J., Kim, Y., Min, C., Choi, J.: Context-aware planning and environment-aware memory for instruction following embodied agents. In: ICCV (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "text", + "content": "T. Kim and C. Min et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "35. Kolve, E., Mottaghi, R., Han, W., VanderBilt, E., Weihs, L., Herrasti, A., Gordon, D., Zhu, Y., Gupta, A., Farhadi, A.: Ai2-thor: An interactive 3d environment for visual ai. arXiv:1712.05474 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "text", + "content": "36. Krantz, J., Gervet, T., Yadav, K., Wang, A., Paxton, C., Mottaghi, R., Batra, D., Malik, J., Lee, S., Chaplot, D.S.: Navigating to objects specified by images. In: ICCV (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 481, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 481, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 481, + 206 + ], + "type": "text", + "content": "37. Krantz, J., Wijmans, E., Majumdar, A., Batra, D., Lee, S.: Beyond the nav-graph: Vision-and-language navigation in continuous environments. In: ECCV (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 207, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 481, + 239 + ], + "type": "text", + "content": "38. Ku, A., Anderson, P., Patel, R., Ie, E., Baldridge, J.: Room-Across-Room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. In: EMNLP (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 481, + 262 + ], + "type": "text", + "content": "39. Kumar, A., Fu, Z., Pathak, D., Malik, J.: Rapid motor adaptation for legged robots. In: RSS (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 263, + 481, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 263, + 481, + 306 + ], + "spans": [ + { + "bbox": [ + 133, + 263, + 481, + 306 + ], + "type": "text", + "content": "40. Li, C., Xia, F., Martin-Martin, R., Lingelbach, M., Srivastava, S., Shen, B., Vainio, K.E., Gokmen, C., Dharan, G., Jain, T., Kurenkov, A., Liu, K., Gweon, H., Wu, J., Fei-Fei, L., Savarese, S.: igibson 2.0: Object-centric simulation for robot learning of everyday household tasks. In: CoRL (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 308, + 481, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 308, + 481, + 340 + ], + "spans": [ + { + "bbox": [ + 133, + 308, + 481, + 340 + ], + "type": "text", + "content": "41. Li, C., Zhang, R., Wong, J., Gokmen, C., Srivastava, S., Martin-Martin, R., Wang, C., Levine, G., Lingelbach, M., Sun, J., et al.: Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation. In: CoRL (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 341, + 481, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 341, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 133, + 341, + 481, + 384 + ], + "type": "text", + "content": "42. Li, Z., Yu, T.W., Sang, S., Wang, S., Song, M., Liu, Y., Yeh, Y.Y., Zhu, R., Gundavarapu, N., Shi, J., Bi, S., Yu, H.X., Xu, Z., Sunkavalli, K., Hasan, M., Ramamoorthi, R., Chandraker, M.: Open rooms: An open framework for photorealistic indoor scene datasets. In: CVPR (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 386, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 386, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 133, + 386, + 481, + 407 + ], + "type": "text", + "content": "43. MacMahon, M., Stankiewicz, B., Kuipers, B.: Walk the talk: Connecting language, knowledge, and action in route instructions. In: AAAI (2006)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 408, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 408, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 133, + 408, + 481, + 430 + ], + "type": "text", + "content": "44. Majumdar, et al.: Findthis: Language-driven object disambiguation in indoor environments. In: CoRL (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 431, + 481, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 431, + 481, + 452 + ], + "spans": [ + { + "bbox": [ + 133, + 431, + 481, + 452 + ], + "type": "text", + "content": "45. Mao, Y., Zhang, Y., Jiang, H., Chang, A., Savva, M.: Multiscan: Scalable rgbd scanning for 3d environments with articulated objects. In: NeurIPS (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 453, + 481, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 453, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 133, + 453, + 481, + 474 + ], + "type": "text", + "content": "46. Min, S.Y., Chaplot, D.S., Ravikumar, P., Bisk, Y., Salakhutdinov, R.: Film: Following instructions in language with modular methods. In: ICLR (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 475, + 481, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 475, + 481, + 497 + ], + "spans": [ + { + "bbox": [ + 133, + 475, + 481, + 497 + ], + "type": "text", + "content": "47. Min, S.Y., Tsai, Y.H.H., Ding, W., Farhadi, A., Salakhutdinov, R., Bisk, Y., Zhang, J.: Object goal navigation with end-to-end self-supervision. In: IROS (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 498, + 481, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 498, + 481, + 530 + ], + "spans": [ + { + "bbox": [ + 133, + 498, + 481, + 530 + ], + "type": "text", + "content": "48. Misra, D., Bennett, A., Blukis, V., Niklasson, E., Shatkhin, M., Artzi, Y.: Mapping instructions to actions in 3d environments with visual goal prediction. In: EMNLP (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 532, + 481, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 532, + 481, + 563 + ], + "spans": [ + { + "bbox": [ + 133, + 532, + 481, + 563 + ], + "type": "text", + "content": "49. Padmakumar, A., Thomason, J., Shrivastava, A., Lange, P., Narayan-Chen, A., Gella, S., Piramuthu, R., Tur, G., Hakkani-Tur, D.: Teach: Task-driven embodied agents that chat. In: AAAI (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 566, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 566, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 133, + 566, + 481, + 586 + ], + "type": "text", + "content": "50. Partsey, R., Wijmans, E., Yokoyama, N., Dobosevych, O., Batra, D., Maksymets, O.: Is mapping necessary for realistic pointgoal navigation? In: CVPR (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 588, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 588, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 133, + 588, + 481, + 609 + ], + "type": "text", + "content": "51. Pashevich, A., Schmid, C., Sun, Chen: Episodic transformer for vision-and-language navigation. In: ICCV (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 610, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 610, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 610, + 481, + 632 + ], + "type": "text", + "content": "52. Puig, X., Ra, K., Boben, M., Li, J., Wang, T., Fidler, S., Torralba, A.: Virtualhome: Simulating household activities via programs. In: CVPR (2018)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 133, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 633, + 481, + 665 + ], + "type": "text", + "content": "53. Puig, X., Undersander, E., Szot, A., Cote, M.D., Yang, T.Y., Partsey, R., Desai, R., Clegg, A.W., Hlavac, M., Min, S.Y., et al.: Habitat 3.0: A co-habitat for humans, avatars and robots. In: ICLR (2024)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "text", + "content": "REALFRED" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "54. Qi, Y., Wu, Q., Anderson, P., Wang, X., Wang, W.Y., Shen, C., Hengel, A.v.d.: Reverie: Remote embodied visual referring expression in real indoor environments. In: CVPR (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "text", + "content": "55. Ramakrishnan, S.K., Chaplot, D.S., Al-Halah, Z., Malik, J., Grauman, K.: Poni: Potential functions for objectgoal navigation with interaction-free learning. In: CVPR (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 480, + 239 + ], + "type": "text", + "content": "56. Ramakrishnan, S.K., Gokaslan, A., Wijmans, E., Maksymets, O., Clegg, A., Turner, J.M., Undersander, E., Galuba, W., Westbury, A., Chang, A.X., Savva, M., Zhao, Y., Batra, D.: Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI. In: NeurIPS Datasets and Benchmarks Track (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 239, + 480, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 239, + 480, + 273 + ], + "spans": [ + { + "bbox": [ + 132, + 239, + 480, + 273 + ], + "type": "text", + "content": "57. Ramrakhya, R., Undersander, E., Batra, D., Das, A.: Habitat-web: Learning embodied object-search strategies from human demonstrations at scale. In: CVPR (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 274, + 480, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 480, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 480, + 295 + ], + "type": "text", + "content": "58. Rao, K., Harris, C., Irpan, A., Levine, S., Ibarz, J., Khansari, M.: Rl-cyclegan: Reinforcement learning aware simulation-to-real. In: CVPR (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 296, + 480, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 296, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 296, + 480, + 329 + ], + "type": "text", + "content": "59. Savva, M., Kadian, A., Maksymets, O., Zhao, Y., Wijmans, E., Jain, B., Straub, J., Liu, J., Koltun, V., Malik, J., Parikh, D., Batra, D.: Habitat: A Platform for Embodied AI Research. In: ICCV (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 330, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 480, + 352 + ], + "type": "text", + "content": "60. Sethian, J.A.: A fast marching level set method for monotonically advancing fronts. In: PNAS (1996)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 353, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 480, + 385 + ], + "type": "text", + "content": "61. Shridhar, M., Thomason, J., Gordon, D., Bisk, Y., Han, W., Mottaghi, R., Zettle-moyer, L., Fox, D.: Alfred: A benchmark for interpreting grounded instructions for everyday tasks. In: CVPR (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 386, + 480, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 408 + ], + "type": "text", + "content": "62. Singh, K.P., Bhambri, S., Kim, B., Mottaghi, R., Choi, J.: Factorizing perception and policy for interactive instruction following. In: ICCV (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 409, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 409, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 409, + 480, + 441 + ], + "type": "text", + "content": "63. Song, C.H., Wu, J., Washington, C., Sadler, B.M., Chao, W.L., Su, Y.: Llm-planner: Few-shot grounded planning for embodied agents with large language models. In: ICCV (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 442, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 480, + 464 + ], + "type": "text", + "content": "64. Song, S., Lichtenberg, S.P., Xiao, J.: Sun rgb-d: A rgb-d scene understanding benchmark suite. In: CVPR (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 465, + 480, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 465, + 480, + 509 + ], + "spans": [ + { + "bbox": [ + 132, + 465, + 480, + 509 + ], + "type": "text", + "content": "65. Srivastava, S., Li, C., Lingelbach, M., Martin-Martin, R., Xia, F., Vainio, K.E., Lian, Z., Gokmen, C., Buch, S., Liu, K., et al.: Behavior: Benchmark for everyday household activities in virtual, interactive, and ecological environments. In: CoRL (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 510, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 480, + 542 + ], + "type": "text", + "content": "66. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv:1906.05797 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 543, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 480, + 597 + ], + "type": "text", + "content": "67. Szot, A., Clegg, A., Undersander, E., Wijmans, E., Zhao, Y., Turner, J., Maestre, N., Mukadam, M., Chaplot, D., Maksymets, O., Gokaslan, A., Vondrus, V., Dharur, S., Meier, F., Galuba, W., Chang, A., Kira, Z., Koltun, V., Malik, J., Savva, M., Batra, D.: Habitat 2.0: Training home assistants to rearrange their habitat. In: NeurIPS (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "text", + "content": "68. Tan, J., Zhang, T., Coumans, E., Iscen, A., Bai, Y., Hafner, D., Bohez, S., Vanhoucke, V.: Sim-to-real: Learning agile locomotion for quadruped robots. In: RSS (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "69. Tobin, J., Fong, R., Ray, A., Schneider, J., Zaremba, W., Abbeel, P.: Domain randomization for transferring deep neural networks from simulation to the real world. In: IROS (2017)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 270, + 101 + ], + "type": "text", + "content": "T. Kim and C. Min et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "70. Torbunov, D., Huang, Y., Tseng, H.H., Yu, H., Huang, J., Yoo, S., Lin, M., Viren, B., Ren, Y.: Uvçgan v2: An improved cycle-consistent gan for unpaired image-to-image translation. arXiv:2303.16280 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "type": "text", + "content": "71. Truong, J., Chernova, S., Batra, D.: Bi-directional domain adaptation for sim2real transfer of embodied navigation agents. RA-L (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "text", + "content": "72. Weihs, L., Deitke, M., Kembhavi, A., Mottaghi, R.: Visual room rearrangement. In: CVPR (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "73. Wijmans, E., Kadian, A., Morcos, A., Lee, S., Essa, I., Parikh, D., Savva, M., Batra, D.: Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. In: ICLR (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "text", + "content": "74. Xia, F., Zamir, A.R., He, Z., Sax, A., Malik, J., Savarese, S.: Gibson env: Realworld perception for embodied agents. In: CVPR (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "75. Yenamandra, S., Ramachandran, A., Yadav, K., Wang, A., Khanna, M., Gervet, T., Yang, T.Y., Jain, V., Clegg, A.W., Turner, J., et al.: Homerobot: Open-vocabulary mobile manipulation. In: CoRL (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 303 + ], + "type": "text", + "content": "76. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: ICCV (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "text", + "content": "77. Zhang, J., Tai, L., Yun, P., Xiong, Y., Liu, M., Boedecker, J., Burgard, W.: Vr-goggles for robots: Real-to-sim domain adaptation for visual control. RA-L (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "type": "text", + "content": "78. Zhu, H., Kapoor, R., Min, S.Y., Han, W., Li, J., Geng, K., Neubig, G., Bisk, Y., Kembhavi, A., Weihs, L.: Excalibur: Encouraging and evaluating embodied exploration. In: CVPR (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 380 + ], + "type": "text", + "content": "79. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: ICCV (2017)" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "text", + "content": "REALFRED" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/ReCON_ Training-Free Acceleration for Text-to-Image Synthesis with Retrieval of Concept Prompt Trajectories/3287df3f-9065-4d00-b93c-c529c48f7c92_content_list.json b/2024/ReCON_ Training-Free Acceleration for Text-to-Image Synthesis with Retrieval of Concept Prompt Trajectories/3287df3f-9065-4d00-b93c-c529c48f7c92_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2edd4b49de8f4e6dd8a1d7a5916c31db3613b7bb --- /dev/null +++ b/2024/ReCON_ Training-Free Acceleration for Text-to-Image Synthesis with Retrieval of Concept Prompt Trajectories/3287df3f-9065-4d00-b93c-c529c48f7c92_content_list.json @@ -0,0 +1,1660 @@ +[ + { + "type": "text", + "text": "RECON: Training-Free Acceleration for Text-to-Image Synthesis with Retrieval of Concept Prompt Trajectories", + "text_level": 1, + "bbox": [ + 263, + 140, + 743, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chen-Yi Lu\\*1,2, Shubham Agarwal\\*2, Md Mehrab Tanjim?, Kanak Mahadik?, Anup Rao?, Subrata Mitra?, Shiv Kumar Saini?, Saurabh Bagchi?, and Somali Chaterji", + "bbox": [ + 218, + 233, + 782, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Purdue University", + "bbox": [ + 431, + 292, + 570, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Adobe Research", + "bbox": [ + 439, + 306, + 562, + 318 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Text-to-image diffusion models excel in generating photorealistic images but are hampered by slow processing times. Training-free retrieval-based acceleration methods, which leverage pre-generated \"trajectories,\" have been introduced to address this. Yet, these methods often lack diversity and fidelity as they depend heavily on similarities to stored prompts. To address this, we present RECON (Retrieving Concepts), an innovative retrieval-based diffusion acceleration method that extracts visual \"concepts\" from prompts, forming a knowledge base that facilitates the creation of adaptable trajectories. Consequently, RECON surpasses existing retrieval-based methods, producing high-fidelity images and reducing required Neural Function Evaluations (NFEs) by up to $40\\%$ . Extensive testing on MS-COCO, Pick-a-pick, and DiffusionDB datasets confirms that RECON consistently outperforms established methods across multiple metrics such as Pick Score, CLIP Score, and Aesthetics Score. A user study further indicates that $76\\%$ of images generated by RECON are rated as the highest fidelity, outperforming two competing methods, a purely text-based retrieval and a noise similarity-based retrieval. Project URL: https://stevencylu.github.io/ReCon.", + "bbox": [ + 259, + 356, + 743, + 607 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Text-to-image Generation $\\cdot$ Diffusion Model Acceleration $\\cdot$ Retrieval-based Diffusion", + "bbox": [ + 259, + 618, + 740, + 646 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 694, + 375, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion Probabilistic Models (DMs) are a class of generative models that simulate a process of gradually reversing noise into structured outputs, drawing inspiration from thermodynamics. These models iteratively denoise an initial random noise distribution to the desired data distribution [46]. DMs have proven to be highly effective in a variety of applications, such as text-to-image generation (T2I) [11,17,38,40,41,43,56,65], speech synthesis [10,27,57], and 3D image", + "bbox": [ + 212, + 724, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution.", + "bbox": [ + 230, + 825, + 377, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d625f7c71f67428d7cd4833e4ec0ea8568d1711f874033bad715a4f41e45cc0c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 146, + 334, + 218 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/47173b2ef2e837981066365b792b4d88cb7cc76c4cffeb6d4f0120d61b35bafb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 146, + 416, + 220 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a9cfb3408945354ec4d126241a4622fcc632b6bd318c7bbbd7e0ea41ee871e13.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 146, + 500, + 220 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/30018d2048cb34345c3da5648100c21938aab3ab5ed1dc9c36cb64a1c437145c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 146, + 583, + 220 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d379032f11ddcbf70a29f28e02cc423a32287987a5c5ea443d93b4189a12dde3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 146, + 665, + 220 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/26c46ae7076ed6b84a511f1b108558767965619c5e9d25fe7507ffb48052387a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 146, + 746, + 220 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9d0712dcf5561005d687084ff2feaf5e4400f5dc42bfbefbacaa2524cf6f674b.jpg", + "image_caption": [ + "A large white bear laying on top of a bed with pillows.", + "Fig. 1: Examples of outputs from existing retrieval-based acceleration approaches, namely Text-based (Baseline) and Noise-based (REDi [58]) Retrieval, show their difficulties in accurately representing the input prompts. On the other hand, our proposed Concept-based Retrieval framework (RECON) is able to produce high-fidelity and faithful images compared to other techniques." + ], + "image_footnote": [], + "bbox": [ + 254, + 220, + 416, + 305 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/83a7bd53b2185ebc58c1a404b0087953dc08b92257b8043e5a2370d59fe4e330.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 220, + 500, + 305 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ba601e2486d1e7ec3fd5878202cd7a37d409cac69825f79e57120d4b72db550c.jpg", + "image_caption": [ + "A photo of an old style white oven." + ], + "image_footnote": [], + "bbox": [ + 501, + 220, + 748, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "generation [21, 54, 64]. In particular, DMs like Imagen [43], DALL-E 2 [40], and StableDiffusion XL (SDXL) [38] have excelled in T2I tasks. However, DMs are notably slower than other generative models due to the extensive number of neural function evaluations (NFEs), where each NFE entails a forward network pass. As AI art and related T2I applications gain popularity, enhancing user experience with DMs through faster model inference and reducing NFEs is crucial.", + "bbox": [ + 212, + 428, + 787, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this issue, several approaches have been developed to accelerate the generation process by reducing the number of NFEs in DMs, including both training-based and training-free methods. Training-based methods, such as Distillation and Consistency Models [35,44,50], aim to distill a complex pre-trained model into a more efficient student model, capable of faster image generation with reduced NFEs. While these methods are promising, they often require a substantial amount of training time and data. This challenge becomes even more pronounced as SOTA diffusion models evolve to encompass over a billion parameters [2,38,43], intensifying the high training load.", + "bbox": [ + 212, + 521, + 787, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To mitigate the demand for computing resources, training-free methods prioritize enhancing computational efficiency without necessitating additional model training. These strategies primarily concentrate on identifying more efficient numerical solvers, which reduce the number of denoising steps, thereby decreasing the NFEs required by DMs [33, 59, 62]. Our approach diverges from the more conventional focus on numerical solvers that primarily aim to streamline the denoising process. Rather, it capitalizes on leveraging previously stored intermediate noisy outputs, or \"noises,\" associated with a vast array of known prompts. For any new input prompt, retrieval-based methods retrieve a relevant noise vector from this repository of stored trajectories at a certain intermediate step. This enables the model to resume the denoising process from a later point, effectively skipping the initial steps and thus speeding up image generation. Additionally,", + "bbox": [ + 212, + 659, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "C. Lu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/39e90b29134abceb0168281caa3fef9dbc01fdc9fca5f7aa81a4e3035adff136.jpg", + "image_caption": [ + "Fig. 2: Overview of different approaches. Text-based (Baseline) and noise-based retrieval (REDi [58]) often follow rigid trajectories with extraneous elements (highlighted in red). In contrast, our method (ReCON) parses input prompts into visual concepts and selects closely related ones from a knowledge base (highlighted in green), enabling a malleable trajectory for incorporating input-specific additional details. Here, $0=\\text{Start}$ , $h=\\text{Initial}$ , $k=\\text{Intermediate}$ , and $T=\\text{Final}$ denote the diffusion steps." + ], + "image_footnote": [], + "bbox": [ + 246, + 142, + 759, + 334 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "this retrieval-based approach can be integrated with any diffusion model's numerical solvers, without altering the core denoising mechanics.", + "bbox": [ + 212, + 460, + 784, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To retrieve the nearest trajectory for an input text, a straightforward approach can be based on text-based similarity [3]. Here, we can find the prompt that is most similar to the input prompt and retrieve its associated trajectory to get the noise at an intermediate step before starting the denoising process. In theory, semantically similar prompts (e.g., \"An old man\" vs. \"A man who is old\") should yield comparable trajectories in DMs, allowing the denoising process to resume from an intermediate noise level with analogous outcomes. Different from this simple textual similarity, [58] proposed REDI, a noise-based retrieval framework. This method initiates the denoising sequence for few steps, subsequently identifying the most closely matched trajectory based on noise-wise similarity.", + "bbox": [ + 212, + 492, + 787, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The efficacy of text-based retrieval (Baseline) and noise-based retrieval (REDi) methods in generating images are illustrated in Figure 1. The figure shows that these methods often struggle to produce images that are faithful to the input text. To understand this, we provide a high-level overview of each method in Figure 2. As seen from the illustration, both methods rely on stored prompts, either directly for text similarity or indirectly for noise similarity. This strategy presupposes a diverse repository capable of reflecting the myriad of unique prompts encountered during inference. However, the reality often deviates, as truly analogous prompts with matching concepts are scarce in real-world scenarios. Moreover, even when a prompt is retrieved that seems closely matched in text (Baseline) or initial noise (REDi), it might introduce extraneous concepts into the generated image. Key image attributes, such as layouts and color schemes, are predominantly determined during the early stages of the denoising", + "bbox": [ + 212, + 643, + 789, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "RECON: Concept Retrieval for Diffusion Acceleration", + "bbox": [ + 375, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "process [15]. However, these stages are often skipped in retrieval-based methods, making it difficult to correct the trajectory using the input prompt within the constraints of reduced NFEs post-retrieval. This issue is depicted in Figure 2, where red-colored texts like \"on a bike\" (Baseline) or \"a dog\" (REDI) indicate unwanted details in the final image. Such trajectory deviations present a critical flaw in both techniques, making retrieval-based acceleration challenging.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address these challenges, we present RECON. The central idea of RECON is retrieving concepts from the input prompt, which can be thought of as \"soft templates\" for the image generation process. This method involves collecting intermediate noises corresponding to these concepts and then combining them. The combination results is what we call a \"flexible trajectory,\" which incorporates some details from the retrieved prompts while still leaving scope for additional elements that are unique from the input prompt. This is illustrated in Figure 2 under \"Concept-based Retrieval,\" where the prompt is decomposed into three generic concepts (A surfer, The beach, and A yellow surfboard) retrieved from the knowledge base. This flexibility is central to RECON's approach. Our strategy draws upon NLP research, which suggests that a sentence can be deconstructed into phrases and evaluated for various attributes such as named entities [26], abstractness [5, 47], and visualness [52]. In our method, we primarily use a pretrained language model [31] to extract essential noun phrases and determine their visualness using [52] to identify \"concepts\" (we use \"visual concepts\" and \"concepts\" interchangeably). As illustrated in Figure 1, the adoption of visual concepts allows RECON to generate images that are not only more accurate but also more faithful to the input prompts, surpassing previous methods. To the best of our knowledge, RECON is the first approach to extract visual concepts in T2I tasks for retrieval-based acceleration. Contributions. 1) We introduce RECON, first of its kind to extract concepts for accelerating T2I tasks without the need for additional training. By retrieving concepts similar to the input prompt, RECON crafts a flexible trajectory that closely aligns with the input prompt, allowing for the incorporation of unique details. 2) We conduct a comprehensive experimental study to compare RECON with both text-based (Baseline) and noise-based (REDI) [58] retrieval frameworks across three datasets: MS-COCO [7], Pick-a-Pick [23], and DiffusionDB [55], using various metrics, such as CLIP Score [39], PickScore [23], and Aesthetic Score [1]. Our findings reveal that while text-based retrieval is surprisingly robust relative to noise-based retrieval techniques (namely REDI [58]), especially with SDXL's high-fidelity model, RECON consistently excels in all measured dimensions. Specifically, RECON excels in both fidelity and speed, generating high CLIP Score images in just 6-7 seconds on an A10G GPU—faster than REDI and Baseline, which require more than 10 seconds. 3) Our contribution to the community extends beyond novel methodologies; we have curated a comprehensive concept knowledge base (KB), encompassing over 1.2M visual concepts and their trajectories derived from 1.6M unique prompts across varied datasets. This repository will be made publicly available (upon acceptance) to facilitate further research in the nascent field of retrieval-based T2I acceleration.", + "bbox": [ + 212, + 250, + 787, + 843 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "C. Lu et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 387, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training-based Acceleration. Several attempts have been made to accelerate the sampling process of diffusion models by introducing additional training beyond the base diffusion model. Distillation methods [14,30,34,35,44,63] aim to distill a teacher model into a more efficient student model with fewer denoising steps. Consistency Models [49,50], on the other hand, leverage the inherent properties of ODE samplers within diffusion models to train a new model by minimizing the difference between points along identical trajectories. Despite being able to drastically reduce generation steps (potentially to a single step), these acceleration methods require additional training. For larger base models with billions of parameters, such as SDXL [38], used in our paper, the requisite for training resources becomes even more pronounced. In contrast, our training-free, retrieval-based method can be easily implemented with a single A10G GPU.", + "bbox": [ + 212, + 175, + 787, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training-free Acceleration. Since the introduction of interpreting diffusion models in continuous time [51], the sampling process can be executed by solving reverse SDEs or ODEs, given that the time-dependent distribution is equivalent. This enables a broad spectrum of research to investigate better numerical solvers for the reverse differential equation [22,24,33,60]. Consequently, the total number of required steps (and therefore, NFEs) can be significantly reduced, often to just 10-20. Our contribution, while complementary to the efforts in optimizing numerical solvers, introduces an additional layer of efficiency by exploiting the concept of caching and retrieving intermediate noises. This approach is orthogonal to the direct optimization of solvers because it does not alter the computational process of the samplers. Instead, it leverages the pre-generated data to bypass certain steps of the diffusion process. By caching intermediate noises we can quickly retrieve and resume the generation process from these points for new images. This method is compatible with any recent sampler, providing a versatile and powerful tool for acceleration.", + "bbox": [ + 212, + 357, + 787, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Retrieval-based Diffusion. Current retrieval-based T2I studies typically target generating images of rare entities [6] or adapting to new scenarios [45]. In the training-free acceleration direction, the closest work to ours is REDI [58], where the authors introduced a training-free method for image generation that leverages initial noise to retrieve the most similar generation trajectory. Our approach differs fundamentally from REDI by concurrently retrieving noises from multiple semantically similar concept prompts without the need for initial diffusion steps to obtain the noise vector, thus avoiding additional diffusion steps. Moreover, our study, using the enhanced SDXL model [38] alongside multiple datasets featuring complex prompts, demonstrates a surprising finding: simple text-based retrieval (Baseline in Fig. 2) outperform REDI when tested across extended models and datasets. Nonetheless, both REDI and the baseline approach fall short in fully capturing the intricacies of the input prompts, prompting the exploration of a more effective concept-based retrieval method.", + "bbox": [ + 212, + 583, + 787, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Concept Retrieval for T2I. Several studies have explored synthesizing images with user-provided concepts. These concepts can be either textual concepts [29] or image concepts [8,25,42] to generate personalized or compositionally complex", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "RECON: Concept Retrieval for Diffusion Acceleration", + "bbox": [ + 375, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4a97beaf5867e4097e501d17e2cbc741cf03e223db4c57fcf994aafbe5c39533.jpg", + "image_caption": [ + "Fig. 3: Overview of RECON. For a given input prompt, RECON decomposes it into distinct visual concepts. Our framework then stores unique concepts from known prompts as well as their trajectories for retrieval purpose. During inference, the same concept decomposition happens and the most similar concepts along with their linked noises are retrieved. These concept noise vectors are then aggregated at step $k$ to resume the denoising process from $t = k$ to $T$ , resulting in decreasing $k$ NFEs." + ], + "image_footnote": [], + "bbox": [ + 251, + 143, + 759, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "images. This exploration includes methods for customizing images with specific concept images [8, 25, 42] and composing multiple text-conditioned diffusion models for intricate image compositions [29]. Our work differs in two significant ways. First, we automatically extract visual concepts from text prompts instead of relying on user-provided concepts. Second, we cache and retrieve these automatically derived concepts alongside their corresponding noise vectors for accelerating DM. To achieve automatic extraction of visual concepts, we mainly draw inspiration from existing NLP research on word and phrase abstractness and imageability [5, 19, 32, 47]. More recently, Verma et al. [52] proposed a method for evaluating prompt-level visualness in T2I, which is more relevant for our case. Inspired by this, we accomplish automatic extraction of visual concept prompts by adopting a fine-tuned CLIP [39] to gauge the text-visualness [52]. Using this, we mine around 1.2M unique concept texts, which we later use for retrieval-based acceleration. To the best of our knowledge, this is the first attempt to extract visual concepts specifically for accelerating T2I tasks.", + "bbox": [ + 212, + 439, + 787, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3 Approach", + "text_level": 1, + "bbox": [ + 215, + 690, + 346, + 708 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 723, + 372, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Diffusion Models. We approach diffusion models (DMs) through the lens of differential equations [22, 51], focusing on the \"trajectory\" idea within the diffusion process. Let $\\mathbf{x} \\sim p_{data}(\\mathbf{x})$ represent the data distribution. DMs perturb $\\mathbf{x}$ by iteratively adding i.i.d. Gaussian noise with variance $\\sigma_t^2$ over $t \\in [T,0]$ , resulting in a noise distribution $p(\\mathbf{x}_0; \\sigma_0)$ . When $\\sigma_0^2$ reaches a sufficiently large value $\\sigma_{max}^2$ , the distribution $p(\\mathbf{x}_0; \\sigma_0^2)$ closely resembles isotropic Gaussian noise with", + "bbox": [ + 212, + 748, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "C. Lu et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "variance $\\sigma_{max}^{2}$ . Image generation in DMs starts with a noise-saturated sample $\\mathbf{x}_0$ and progressively denoises to approximate $p(\\mathbf{x}_t; \\sigma_t)$ , where $t$ ranges from 0 to $T$ . In this progression, $\\sigma_{t+1}$ is smaller than $\\sigma_t$ , and $\\sigma_0$ is equal to $\\sigma_{max}$ . The final output at $\\sigma_T = 0$ , should ideally reflect the original data distribution. This iterative denoising procedure is computationally represented by the Probability Flow ODE [51], which tracks the data sample trajectory over time:", + "bbox": [ + 212, + 145, + 787, + 238 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nd \\mathbf {x} = - \\bar {\\sigma} (t) \\sigma (t) \\nabla_ {\\mathbf {x}} \\log p (\\mathbf {x}; \\sigma (t)) d t, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 247, + 785, + 263 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\nabla_{\\mathbf{x}}\\log p(\\mathbf{x};\\sigma)$ denotes the score function, and $\\sigma (t)$ is a user-defined noiselevel schedule with its time derivative represented by $\\bar{\\sigma} (t)$ . In practice, the score function is estimated by training a deep neural network $\\mathbf{s}_{\\theta}(\\mathbf{x};\\sigma)$ parameterized by $\\theta$ . Note that we are using the same notation as Karras et al. [22], where $\\mathbf{x}_0$ is the noisy input and $\\mathbf{x}_T$ is the fully denoised output or generated final image. Throughout this paper, we refer to the trajectory $\\{\\mathbf{x}_t:t\\in \\{0\\dots T\\}\\}$ as the one generated by a conditioned DM (typically using classifier-free guidance [18] for T2I), where $\\mathbf{x}_t$ represents the noisy model outputs (noises) at each time step.", + "bbox": [ + 212, + 268, + 787, + 391 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Retrieval-based Acceleration", + "text_level": 1, + "bbox": [ + 214, + 411, + 500, + 425 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From Equation (1), we can see that the sampling process becomes deterministic once $\\mathbf{x}_0$ is sampled from the Gaussian distribution. This allows for the pre-generation and storage of trajectories corresponding to a diverse range of prompts within a database. During inference, the key challenge lies in retrieving the trajectory that most closely aligns with the trajectory of the input prompt, posing the critical question: How do we select the optimal trajectory to resume the denoising process? To this end, we explore two primary approaches from literature, which we describe below.", + "bbox": [ + 212, + 433, + 787, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Text-based Trajectory Retrieval. Since the input prompt is the initial element, it is intuitive to retrieve trajectories associated with the closest prompts. Specifically, for an input prompt $\\mathcal{P}$ , we encode it with a pre-trained text encoder to obtain its embedding $\\phi(\\mathcal{P})$ . The nearest prompt $\\mathcal{P}'$ is identified by finding the minimum cosine distance between text embeddings, and its trajectory $\\{\\mathbf{x}_t^{P'}\\}_{t=0}^T$ is obtained. The intermediate output of $\\mathcal{P}'$ at step $k$ is looked up in the database and then, the diffusion process is resumed from the intermediate step $k \\in [0,T]$ with the original prompt $\\mathcal{P}$ , effectively skipping $k$ NFEs, as illustrated in Figure 2. We refer to this method as Baseline.", + "bbox": [ + 212, + 553, + 787, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Noise-based Trajectory Retrieval. The alternative method, REDI, retrieves trajectories by searching for the closest noise vector [58]. Specifically, for a given input prompt $\\mathcal{P}$ , REDI first computes $h$ steps in the denoising process and uses the noise $\\mathbf{x}_h^{\\mathcal{P}}$ as a query. It then searches for the nearest noise $\\mathbf{x}_h^{P'}$ in the knowledge base and resumes the denoising process from step $k > h$ , effectively skipping $k - h$ steps. However, this method adds $h$ initial steps to the number of NFEs, as illustrated in Figure 2.", + "bbox": [ + 212, + 688, + 787, + 796 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Our Proposed Framework: RECON", + "text_level": 1, + "bbox": [ + 214, + 814, + 552, + 830 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "RECON: Concept Retrieval for Diffusion Acceleration", + "bbox": [ + 375, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Both Baseline and REDI rely on stored prompts $\\mathcal{P}'$ , either directly through text similarity or indirectly through noise similarity. However, they face challenges in generating images that are faithful to the input prompt $\\mathcal{P}$ . As shown in Figure 2, both the baseline method and REDI introduce extraneous components into the final images, compromising its fidelity. Crucial image attributes, such as layout and color schemes, are defined in the initial denoising stages [15], which are frequently bypassed in retrieval-based approaches, leading to a misalignment from the intended prompt details.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Concept Extraction and KB Creation. To address the challenge of incorporating specific details from input prompts while avoiding the rigidity seen in Baseline and REDI, we introduce a novel approach: decomposing input prompts into visual concepts. However, the challenge lies in distinguishing visual concepts from abstract terms. For example, terms like \"belief,\" \"essentialness,\" and \"spirituality,\" while conceptually significant, often rank high in abstractness [5], but do not translate into visual representations for T2I generation. This distinction is crucial as language models like RoBERTa [9, 31] are proficient in extracting and understanding text, including abstract concepts, the challenge lies in ensuring that these concepts can be visually represented. To tackle this challenge, we adopt a fine-tuned CLIP model to evaluate the visual potential of extracted phrases, ensuring that only those with a clear visual representation are selected [52]. In the process of RECON, each text prompt $\\mathcal{P}$ is decomposed into $N$ visual concepts $\\{c_i\\}_{i=1}^N$ , using a pre-trained RoBERTa [31]. Subsequently, a visualness filtering model $h(\\cdot)$ , fine-tuned to distinguish between visually representable and non-visual text segments, evaluates these extracted noun phrases. Representations. The visualness of a concept $c$ is quantified through cosine similarity $S = 1 - \\langle h(c), h(\\mathcal{I}_{null}) \\rangle$ , where $\\mathcal{I}_{null}$ denotes an image of random RGB values, serving as a sample for non-visual content. Concepts with visualness scores $S$ below a predefined threshold are deemed non-visual and excluded from further consideration. These visually relevant concepts and their trajectories from DMs are then cached in a vector database for retrieval purpose. Figure 3 illustrates these steps in this stage of RECON under the section \"Concept KB and Trajectories Creation.\"", + "bbox": [ + 212, + 268, + 787, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Retrieving and Aggregating Concepts. During inference, given an input prompt $\\mathcal{P}$ , a concept knowledge base, and a pre-trained diffusion model $f_{\\theta}$ , we first decompose the prompt into a set of visual concepts $\\{c_i\\}_{i=1}^N$ . For each concept $c_i$ , we compute its text embedding $\\phi(c_i)$ , which serves as a query for retrieving the closest matching concepts $c_i'$ along with their associated trajectories $\\mathbf{x}_t^{c_i'}$ . Subsequently, at a pre-determined step $k \\in [0,T]$ , we aggregate these trajectories by combining their noise vectors $\\mathbf{x}_k^{c_i'}$ at step $k$ , thus forming a new starting point for the denoising process:", + "bbox": [ + 212, + 633, + 789, + 762 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} _ {t} = \\sum_ {i = 1} ^ {N} w _ {i} \\cdot \\mathbf {x} _ {k} ^ {c _ {i} ^ {\\prime}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 441, + 776, + 785, + 816 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {t + 1} = f _ {\\theta} \\left(\\hat {\\mathbf {x}} _ {t}; t, \\mathcal {P}\\right); t = k, \\dots , T - 1, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 821, + 785, + 837 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "C. Lu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/376c95a8294d684546c0c281399ee70dd8393733f6b89b29b3c949354e7ac95d.jpg", + "image_caption": [ + "Fig. 4: Qualitative comparisons reveal that RECON consistently produces images of quality comparable to those generated by Vanilla, with no retrieval-based acceleration. Here, the numbers of NFEs are: Vanilla=50, Others=35. Full prompts for DiffusionDB appear in Supplementary." + ], + "image_footnote": [], + "bbox": [ + 243, + 151, + 756, + 352 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\{w_{i}:i\\in [1,N]\\}$ corresponds to the weights for each concept. Our method defines weights $w_{i}$ using the cosine similarity between text embeddings of the query and retrieved concepts, normalized across all concepts: $w_{i} = \\frac{\\left\\langle\\phi(c_{i}),\\phi(c_{i}^{\\prime})\\right\\rangle}{\\sum_{i = 1}^{N}\\left\\langle\\phi(c_{i}),\\phi(c_{i}^{\\prime})\\right\\rangle}$ . This approach is grounded in the principle that diffusion models can be regarded as energy-based models (EBMs), facilitating the compositional aggregation of noise vectors [12,28,29] For detailed insights into noise vector compositionality in EBMs, please refer to our Supplementary material. The \"Concept Retrieval\" and \"Concept Noises Aggregation\" stages precede the continuation of the denoising process with the aggregated noise $\\hat{\\mathbf{x}}_t$ and the original prompt $\\mathcal{P}$ to generate the final image $\\mathbf{x}_T$ , as depicted in Figure 3. While only the concept noun phrases are aggregated, the final diffusion steps (steps $k\\dots T$ in Fig. 3) are guided by the original prompt. Thus, essential prompt details beyond nouns still get to influence the resulting image $\\mathbf{x}_T$ . For instance, in col. 3 of Fig. 4, although the concepts extracted from the original prompt \"A herd of sheep standing below very tall buildings\" are \"a herd of sheep\" and \"buildings\" without including \"tall,\" the final image still captures this key attribute.", + "bbox": [ + 212, + 450, + 789, + 705 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 729, + 375, + 748 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Setup", + "text_level": 1, + "bbox": [ + 215, + 765, + 310, + 781 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. For a fair comparison, we first assess the performance of RECON using the same dataset as REDI [58]: MS-COCO [7] (containing 82K and 42K captions in train and validation). However, captions in MS-COCO may not always", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "RECON: Concept Retrieval for Diffusion Acceleration", + "bbox": [ + 375, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/fb70891a5c8405137ee032a2b34c8ea56587fb5d2f759d4ebe8455331a0c7815.jpg", + "image_caption": [ + "Baseline", + "NFE=40", + "NFE = 20" + ], + "image_footnote": [], + "bbox": [ + 303, + 152, + 447, + 191 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/c3715e8be9f343deebf607791b53de4232c1e18a675465d7e0ac800b8f819034.jpg", + "image_caption": [ + "NFE=50" + ], + "image_footnote": [], + "bbox": [ + 450, + 164, + 496, + 212 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/f3c5f1dc9269a48f61d3d343cd10608cdee04201869d7655c081407a95e96baa.jpg", + "image_caption": [ + "Baseline", + "NFE = 40", + "RECON" + ], + "image_footnote": [], + "bbox": [ + 506, + 152, + 651, + 191 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/bfaeba49e76106a84bc9703b9d5c6370e9184348edcb50dea87559f74df3be1d.jpg", + "image_caption": [ + "A dog warrior holding a sword in pink armor.", + "NFE = 20", + "A cat is sitting on a wooden café table with décor.", + "Fig. 5: While other methods struggle to generate faithful images with relatively higher NFEs, ours (RECON) can generate faithful images at NFEs as low as 20." + ], + "image_footnote": [], + "bbox": [ + 504, + 198, + 651, + 237 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/170fe417b8b96c6af0ade37d97f9a01d911f5c6a802afa62c47926861f3765f8.jpg", + "image_caption": [ + "Vanilla", + "NFE=50" + ], + "image_footnote": [], + "bbox": [ + 653, + 171, + 704, + 210 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "reflect real-world T2I prompts, as noted in [23]. Therefore, we also use Pick-a-Pic [23] (35K unique prompts) and DiffusionDB [55] (1.8M unique prompts). These datasets offer a broader range of prompts, more reflective of actual T2I user preferences. For evaluation, we randomly select 10K captions from MS-COCO, 500 prompts from Pick-a-Pic, and 10K from DiffusionDB from their validation/test set (detailed in Supplementary). To our knowledge, such a comprehensive study has not been done before for retrieval-based acceleration.", + "bbox": [ + 212, + 325, + 784, + 430 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Metrics. In unconditional settings, a popular metric to test photorealism is FID [16], which compares feature distributions from real and generated images. However, FID falls short in conditional settings such as T2I. Studies have pointed out FID is negatively correlated with human preferences [23, 38]. Therefore, for a comprehensive evaluation of T2I models, we employ metrics that specifically address the conditional nature of the task. PickScore [23] measures the alignment between text descriptions and generated images. \"CLIP Score\" [39], on the other hand, provides a broader measure of semantic congruence between the text and generated image, useful for understanding how well the generated image captures the gist of the text. \"Aesthetics Score\" [1] complements these by assessing the visual appeal of images. Finally, we employed NFE to quantify RECON's time efficiency, independent of hardware configurations.", + "bbox": [ + 212, + 431, + 784, + 611 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation Details. For building Baseline and REDI (described in Section 3.2), we follow the steps mentioned in [58], but replace the base model with the improved SDXL [38]. We denote the default sampling process in SDXL without any acceleration as \"Vanilla Generation.\" For RECON, we first build a concept knowledge base (KB) from the unique captions and prompts in the training sets of all considered datasets. We extract noun phrases using a pretrained language model, RoBERTa [31] and apply visual filtering with a CLIP model [39] from [52]. To efficiently store and retrieve these extracted concepts and their associated trajectories, we implement a vector database with FAISS [20]. Remarkably, the concept retrieval part of our pipeline merely takes $1.15\\%$ of the total inference time on an A10G GPU. We build the rest of our pipeline based on PyTorch and the HuggingFace Diffusers library [36, 37]. We determine the weights for merging concept noise vectors using the CLIP text similarity scores, as in Equation (2). Unless otherwise mentioned, for all the experiments, we use the Euler sampler [22] as default, setting the number of NFEs to 35 for retrieval-", + "bbox": [ + 212, + 613, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "C. Lu et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/a07421fbf178f5ca0f43e06e01e0885401bafb3866bb03ae2d5314870f86e4e5.jpg", + "image_caption": [ + "\"Two brown teddy bears sitting on top a wooden table.\"" + ], + "image_footnote": [], + "bbox": [ + 292, + 146, + 504, + 311 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/eb2485acb6bc635e380b7156f8f27baa621e67ade6ea23090c24cf12a743ff25.jpg", + "image_caption": [ + "\"renaissance painting of elder in the garden, short silver hair, a wise face, emotions closeup, dressed in roman armour.\"", + "Fig. 6: Baseline and REDI often struggle with 1st Nearest Neighbor (NN) results, leading to poorer outcomes for 2nd and 3rd NNs. Conversely, RECon consistently produces accurate images across all NN concepts, addressing the diversity challenge." + ], + "image_footnote": [], + "bbox": [ + 506, + 146, + 709, + 311 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "based methods and 50 for Vanilla Generation. Further implementation details, including configurations of SDXL, the KB size, specifics of our vector database for caching and retrieval are provided in Supplementary.", + "bbox": [ + 214, + 431, + 787, + 477 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Qualitative Results", + "text_level": 1, + "bbox": [ + 215, + 498, + 421, + 513 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Enhanced Fidelity. In Figure 4, Images generated using RECON demonstrate a higher degree of fidelity to the provided prompts compared to both Baseline and REDI methods. RECON images also closely mirror the results of Vanilla Generation across all datasets. In contrast, images from Baseline and REDI images exhibit notable discrepancies from the input prompt, such as depicting inaccurate objects (animal mouse vs. computer mouse), or omitting key elements described in the prompts (e.g., \"very tall buildings\"). RECON is able to generate uncommon concept combinations like \".cat eating a burger like a person.\", \".raccoon washing blankets.\", and \".cat dressed as a scuba diver..\" . REDI and Baseline often miss the mark with these intricate descriptions, occasionally generating completely irrelevant images, like those for the prompt \"Human abduction by UFO.\" RECON also performs better with rare input prompts. For instance, the phrase \"sparkling diamond gold butterfly\" is absent from KB. Nevertheless, RECON produced a relevant image of a butterfly based on the concept \"a sparkling gold glitter dress,\" thus highlighting the advantages of the flexible trajectory of RECON's design. Figure 5 showcases the outputs from three retrieval-based methods with both high and low number of NFEs. Both Baseline and REDI methods struggle to faithfully capture the input prompts in their images at relatively high number of NFEs. Unsurprisingly, their performance further diminishes at lower number of NFEs. RECON, however, consistently performs better even with NFEs as low as 20.", + "bbox": [ + 214, + 522, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "RECON: Concept Retrieval for Diffusion Acceleration", + "bbox": [ + 375, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/aca80fa86833f53d0925442495975b44509e468647b58537d4b77f93f9ff56cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 143, + 398, + 220 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/eed3dca3f7b5d18f67cf96de738faf0756b752e4df669a1aad160b6d8b1fea7e.jpg", + "image_caption": [ + "Fig.7: RECON can generate images with high CLIP scores in 6-7 seconds using an A10G GPU, whereas both REDI and Baseline require more than 10 seconds to generate images of comparable quality." + ], + "image_footnote": [], + "bbox": [ + 217, + 220, + 334, + 295 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2ab8801be70883fcebcf9ead8cd58d5ea47a77b10a60d4ca8e74c74211e85117.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 339, + 220, + 457, + 295 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/90bf47d605ad220aed51e9779ed3416af8b2dc50f4cc250d8b26bffcfd27ab7b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 143, + 633, + 220 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ad8537c941fbd203567e7688a3251e16a5123364b292627908e9e1254df554af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 143, + 756, + 220 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/95189773cd9e7b733d852cb9e3adc5bab7c65e9150194d319d8dcc92b6735be2.jpg", + "image_caption": [ + "Fig.8: RECON consistently outperforms Baseline and REDI on MS-COCO under varying total number of steps and different percentage of skipped steps w.r.t. total steps taken by Vanilla Generation." + ], + "image_footnote": [], + "bbox": [ + 511, + 220, + 633, + 297 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c5a05bf9dfea237f5c1e1d037d3bc21248c7dc186b5e9e6cb84db1ee6ed50d28.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 220, + 754, + 297 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Diversity. Diversity in T2I generation ensures that the model can produce a range of visually distinct images for a given prompt, enhancing user experience by offering creative variations. The traditional approach for diversity in T2I tasks involves caching different trajectories for multiple seeds. This method, while capable of producing diverse outputs, faces significant scalability challenges. Storing trajectories for a growing number of unique prompts, even for a single seed, quickly becomes prohibitively expensive. A cheaper solution to this problem is to pull the top- $k$ Nearest Neighbor (NN) trajectories. However, Baseline and REDI struggle with accuracy and faithfulness, particularly beyond the first NN, as shown in Figure 6. RECON effectively overcomes this by focusing on similar visual concepts rather than full-length prompts. This approach is more efficient as matching concepts is simpler than entire prompts. For instance, in one of the prompts showcased, the top-three retrieved NN concepts are: 1st: ['two brown teddy bears', 'large wooden table'], 2nd: ['two teddy bears', 'wooden kitchen table'], and 3rd: ['two stuffed teddy bears', 'wooden counter'], all closely related to each other. This similarity among the top concepts enables RECON to maintain diversity without sacrificing fidelity or facing scalability issues.", + "bbox": [ + 212, + 409, + 787, + 667 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Quantitative Results", + "text_level": 1, + "bbox": [ + 215, + 691, + 444, + 707 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison of All Methods. In Figure 8, we show the performance with different percentage of skipped steps with varying the number of total steps (or NFEs) on MS-COCO (results from other datasets are shown in Supplementary). RECON consistently surpasses both REDI and Baseline across different NFE counts, demonstrating more notable improvement at lower number of NFEs. In Figure 7, we demonstrate the CLIP Scores of RECON in comparison to previous works. With just 6 seconds of computation time on an A10G GPU, RECON is able to produce accurate images from prompts across various datasets, while", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "C. Lu et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/41e40f761199cbeaa6582eb77e83ce5a26d2d71982dfc811a56c5344e3c6f55a.jpg", + "table_caption": [ + "Table 1: Our concept-based retrieval RECon surpasses Baseline and REDI across MS-COCO, Pick-a-Pic, and DiffusionDB datasets. NFE is fixed to 35." + ], + "table_footnote": [], + "table_body": "
CLIP Score↑PickScore↑Aesthetics Score↑
COCOPick-a-PicDiffDBCOCOPick-a-PicDiffDBCOCOPick-a-PicDiffDB
Baseline29.1527.6729.6820.7420.4419.815.506.056.26
REDi18.1721.8020.4319.5119.0318.215.516.066.29
ReCon30.1929.9331.4621.3120.7620.085.526.106.31
", + "bbox": [ + 218, + 184, + 787, + 267 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/e220c865fdc2ef149456495aec036b2ab2278b5984f25f28c11c6f34376e4ecb.jpg", + "table_caption": [ + "Table 2: RECON's ablation study where 'Visual-filtering' leverages text-visualness scores [53], while 'weighting' evaluates the similarity between concepts. 'Top-1' prioritizes the highest scoring concept, while $>0.5$ criterion includes only those concepts with similarities above this value." + ], + "table_footnote": [], + "table_body": "
Visual-filtering\nWeighting✓\nX (Top-1)✓\nX (>0.5)✓\nXXX✓\n✓
CLIP Score ↑27.4629.7929.3328.9829.0629.93
PickScore ↑19.8820.5920.5420.1220.3620.84
", + "bbox": [ + 218, + 397, + 504, + 448 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/77260c625e471d6ae09411cede35b6ed143306b259032baf84794994dbc88525.jpg", + "table_caption": [ + "Table 3: Comparison of all methods for different samplers. RECON consistently outperforms Baseline and REDI across different diffusion samplers, showcasing RECON's ability to work with multiple efficient numerical solvers." + ], + "table_footnote": [], + "table_body": "
SchedulerCLIP Score↑PickScore↑
EulerPNDMDDIMDPMEulerPNDMDDIMDPM
Baseline27.6725.9626.7726.7320.4419.8220.1020.00
REDi21.8020.2421.6821.4219.0318.0818.4718.40
ReCON29.9327.6828.9928.5620.7619.8920.2920.06
", + "bbox": [ + 526, + 397, + 784, + 446 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "both REDI and Baseline require over 10 seconds to generate images of similar quality. A detailed study among the retrieval-based methods is presented in Table 1. It is evident that RECON excels over Baseline and REDI in terms of all metrics. We also make a surprising discovery: Baseline is quite competitive compared to the state-of-the-art REDI. This shows that when using an improved base model, a simple approach like the Baseline can perform quite well, especially on T2I alignment scores. For rest of the analysis, we primarily report these two scores (results for the rest are in the Supplementary).", + "bbox": [ + 212, + 484, + 787, + 606 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Aggregation Ablation Study. Table 2 shows the importance of using both weighted aggregation Equation (2) and visual-filtering in RECON. Here, we compare with using only the most similar concept, denoted as Top-1, and excluding low-similarity concepts with a threshold of 0.5, denoted as $(>0.5)$ . The results indicate that both the weighting and visualness filter are essential.", + "bbox": [ + 212, + 609, + 787, + 685 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Concept Retrieval Analysis. Although both Baseline and RECON rely on text similarity scores (either prompt-to-prompt in Baseline or concept-to-concept in RECON), Table 4 shows that the latter consistently outperforms the former. At each similarity score interval, concept-based retrieval in RECON leads to superior results, proving the effectiveness of our approach. The length and complexity of input prompts, often containing multiple concepts, poses the question: how many concepts to break a prompt into? Figure 9 shows that consistently deconstructing prompts into concepts (RECON) generally produces better outcomes than not doing so (Baseline). Also, it shows that there is a sweet spot for the number of concepts to decompose a text prompt into.", + "bbox": [ + 212, + 688, + 789, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "RECON: Concept Retrieval for Diffusion Acceleration", + "bbox": [ + 375, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/72dadcccdf076deac95baad988b175bbe833011a74059eca1b13e04f4d8af906.jpg", + "table_caption": [ + "Table 4: RECON consistently performs better than Baseline at every text similarity score interval and every NFE count (results shown for PickScore†)." + ], + "table_footnote": [], + "table_body": "
TextNFE = 40NFE = 35NFE = 30NFE = 25
SimilarityBaselineReCONBaselineReCONBaselineReCONBaselineReCON
0.5-0.620.3821.0720.1120.9219.9520.5219.6220.18
0.6-0.720.5120.9820.0720.5919.8620.2019.5320.04
0.7-0.820.7720.9620.4420.6420.2520.3619.9220.16
0.8-0.921.2621.4220.9721.1620.8320.8920.4320.65
", + "bbox": [ + 243, + 184, + 754, + 271 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/9ea05ee6f773c04042fc608ef1b551cb74c338a478c38555fdebfc24926d7e11.jpg", + "image_caption": [ + "Fig.9: Prompts can contain an arbitrary number of concepts; so, breaking down a prompt into its constituent concepts is better (RECon) than not doing so (Baseline)." + ], + "image_footnote": [], + "bbox": [ + 300, + 282, + 501, + 376 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/9159d252bec14b720d161954921265999e5d7602d9158c78689a570d34ba5359.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 282, + 700, + 376 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Comparison with Different Samplers. Table 3 showcases the results of incorporating our method into other diffusion samplers, including DDIM [48], DPM-Solver [33], PNDM [22], and Euler [22], with RECON outperforming other retrieval-based techniques across different samplers.", + "bbox": [ + 212, + 445, + 782, + 505 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Human evaluation. We conducted a comprehensive user study involving 214 participants (details provided in Supplementary) to evaluate image generation using RECON, Baseline, and REDI. After filtering for consistency, we retained 200 valid participants. All methods used the same random seed, steps (NFE = 35), and Euler sampler. Results showed RECON's superior performance, with $75.5\\%$ of its generated images ranked first, $16\\%$ second, and $8.5\\%$ third. This preference for RECON was statistically significant (p-value $< 0.05$ ).", + "bbox": [ + 212, + 506, + 787, + 612 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 638, + 359, + 655 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We proposed RECON, a training-free retrieval-based acceleration method for T2I tasks. We observed that previous approaches were hindered by the rigidity of their trajectories, which often resulted in images lacking fidelity and diversity. To address these challenges, RECON retrieves trajectories for the noise by using concept similarity and then merges the noises derived from multiple semantically similar concepts. The visual concepts extracted for this process have broad applicability. While we focus on retrieval-based acceleration in this paper, the potential uses of these concepts can be applied to noise aesthetics classifications [13], noise inversion techniques [61], and latent-to-latent transformations within generative models [4]. We leave the exploration of our concept knowledge base as future work.", + "bbox": [ + 212, + 672, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "C. Lu et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This material is based in part upon work supported by the National Science Foundation under Grant Number CNS-2146449 (NSF CAREER award), gift funding from Adobe Research, and by the Army Research Lab under contract number W911NF-2020-221. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the sponsors.", + "bbox": [ + 212, + 175, + 787, + 266 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 287, + 321, + 304 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. christophschuhmann/improved-aesthetic-predictor: Clip+mlp aesthetic score predictor. https://github.com/christophschuhmann/improved-aesthetic-predictor", + "2. Dall-e 2. https://openai.com/dall-e-2 (2023)", + "3. Agarwal, S., Mitra, S., Chakraborty, S., Karanam, S., Mukherjee, K., Saini, S.K.: Approximate caching for efficiently serving Text-to-Image diffusion models. In: 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI 24). pp. 1173-1189. USENIX Association, Santa Clara, CA (Apr 2024), https://www.usenix.org/conference/nsdi24/presentation/agarwal-shubham", + "4. Bojanowski, P., Joulin, A., Lopez-Paz, D., Szlam, A.: Optimizing the latent space of generative networks. arXiv preprint arXiv:1707.05776 (2017)", + "5. Brysbaert, M., Warriner, A.B., Kuperman, V.: Concreteness ratings for 40 thousandly known english word lemmas. Behavior research methods 46, 904-911 (2014)", + "6. Cai, D., Wang, Y., Liu, L., Shi, S.: Recent advances in retrieval-augmented text generation. In: Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval. pp. 3417-3419 (2022)", + "7. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dollar, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325 (2015)", + "8. Choi, J., Choi, Y., Kim, Y., Kim, J., Yoon, S.: Custom-edit: Text-guided image editing with customized diffusion models. arXiv preprint arXiv:2305.15779 (2023)", + "9. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022)", + "10. Deng, Y., Wu, N., Qiu, C., Luo, Y., Chen, Y.: Mixgan-tts: Efficient and stable speech synthesis based on diffusion model. IEEE Access (2023)", + "1. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021)", + "2. Du, Y., Li, S., Mordatch, I.: Compositional visual generation with energy based models. Advances in Neural Information Processing Systems 33, 6637-6647 (2020)", + "3. Gallego, V.: Personalizing text-to-image generation via aesthetic gradients. arXiv preprint arXiv:2209.12330 (2022)", + "4. Gu, J., Zhai, S., Zhang, Y., Liu, L., Susskind, J.M.: Boot: Data-free distillation of denoising diffusion models with bootstrapping. In: ICML 2023 Workshop on Structured Probabilistic Inference and Generative Modeling (2023)", + "5. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control (2022)" + ], + "bbox": [ + 225, + 316, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "RECON: Concept Retrieval for Diffusion Acceleration", + "bbox": [ + 375, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)", + "17. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)", + "18. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)", + "19. Jeong, J.W., Wang, X.J., Lee, D.H.: Towards measuring the visualness of a concept. In: Proceedings of the 21st ACM international conference on Information and knowledge management. pp. 2415-2418 (2012)", + "20. Johnson, J., Douze, M., Jégou, H.: Billion-scale similarity search with GPUs. IEEE Transactions on Big Data 7(3), 535-547 (2019)", + "21. Karnewar, A., Vedaldi, A., Novotny, D., Mitra, N.J.: Holodiffusion: Training a 3d diffusion model using 2d images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18423-18433 (2023)", + "22. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems 35, 26565-26577 (2022)", + "23. Kirstain, Y., Polyak, A., Singer, U., Matiana, S., Penna, J., Levy, O.: Pick-a-pic: An open dataset of user preferences for text-to-image generation. arXiv preprint arXiv:2305.01569 (2023)", + "24. Kong, Z., Ping, W.: On fast sampling of diffusion probabilistic models. In: ICML Workshop on Invertible Neural Networks, Normalizing Flows, and Explicit Likelihood Models (2021)", + "25. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1931-1941 (2023)", + "26. Li, J., Sun, A., Han, J., Li, C.: A survey on deep learning for named entity recognition. IEEE Transactions on Knowledge and Data Engineering 34(1), 50-70 (2020)", + "27. Liu, J., Li, C., Ren, Y., Chen, F., Zhao, Z.: Diffsinger: Singing voice synthesis via shallow diffusion mechanism. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 11020-11028 (2022)", + "28. Liu, N., Li, S., Du, Y., Tenenbaum, J., Torralba, A.: Learning to compose visual relations. Advances in Neural Information Processing Systems 34, 23166-23178 (2021)", + "29. Liu, N., Li, S., Du, Y., Torralba, A., Tenenbaum, J.B.: Compositional visual generation with composable diffusion models. In: European Conference on Computer Vision. pp. 423-439. Springer (2022)", + "30. Liu, X., Zhang, X., Ma, J., Peng, J., et al.: Instafaflow: One step is enough for high-quality diffusion-based text-to-image generation. In: The Twelfth International Conference on Learning Representations (2023)", + "31. Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., Stoyanov, V.: Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692 (2019)", + "32. Louis, A., Nenkova, A.: What makes writing great? first experiments on article quality prediction in the science journalism domain. Transactions of the Association for Computational Linguistics 1, 341-352 (2013)", + "33. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems 35, 5775-5787 (2022)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "C. Lu et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Luhman, E., Luhman, T.: Knowledge distillation in iterative generative models for improved sampling speed. arXiv preprint arXiv:2101.02388 (2021)", + "35. Meng, C., Rombach, R., Gao, R., Kingma, D., Ermon, S., Ho, J., Salimans, T.: On distillation of guided diffusion models. In: CVPR (2023)", + "36. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems 32 (2019)", + "37. von Platen, P., Patil, S., Lozhkov, A., Cuenca, P., Lambert, N., Rasul, K., Davaadorj, M., Wolf, T.: Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers (2022)", + "38. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: improving latent diffusion models for high-resolution image synthesis. pp. 1-13 (2024)", + "39. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "40. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)", + "41. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)", + "42. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023)", + "43. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)", + "44. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models (2022)", + "45. Sheynin, S., Ashual, O., Polyak, A., Singer, U., Gafni, O., Nachmani, E., Taigman, Y.: Knn-diffusion: Image generation via large-scale retrieval. arXiv preprint arXiv:2204.02849 (2022)", + "46. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256-2265. PMLR (2015)", + "47. Solovyev, V.: Concreteness/abstractness concept: State of the art. In: International Conference on Cognitive Sciences. pp. 275-283. Springer (2020)", + "48. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: International Conference on Learning Representations (2020)", + "49. Song, Y., Dhariwal, P.: Improved techniques for training consistency models. In: The Twelfth International Conference on Learning Representations (2023)", + "50. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models. Proceedings of Machine Learning Research, vol. 202. PMLR (23-29 Jul 2023), https://proceedings.mlr.press/v202/song23a.htm1" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "RECON: Concept Retrieval for Diffusion Acceleration", + "bbox": [ + 375, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "51. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: International Conference on Learning Representations (2020)", + "52. Verma, G., Rossi, R.A., Tensmeyer, C., Gu, J., Nenkova, A.: Learning the visualness of text using large vision-language models. In: The 2023 Conference on Empirical Methods in Natural Language Processing (2023)", + "53. Verma, G., Rossi, R.A., Tensmeyer, C., Gu, J., Nenkova, A.: Learning the visualness of text using large vision-language models. arXiv preprint arXiv:2305.10434 (2023)", + "54. Wang, T., Zhang, B., Zhang, T., Gu, S., Bao, J., Baltrusaitis, T., Shen, J., Chen, D., Wen, F., Chen, Q., et al.: Rodin: A generative model for sculpting 3d digital avatars using diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4563-4573 (2023)", + "55. Wang, Z.J., Montoya, E., Munechika, D., Yang, H., Hoover, B., Chau, D.H.: Diffusiondb: A large-scale prompt gallery dataset for text-to-image generative models. arXiv preprint arXiv:2210.14896 (2022)", + "56. Xu, X., Wang, Z., Zhang, G., Wang, K., Shi, H.: Versatile diffusion: Text, images and variations all in one diffusion model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7754-7765 (2023)", + "57. Yasuda, Y., Toda, T.: Text-to-speech synthesis based on latent variable conversion using diffusion probabilistic model and variational autoencoder. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023)", + "58. Zhang, K., Yang, X., Wang, W.Y., Li, L.: ReDi: Efficient learning-free diffusion inference via trajectory retrieval. In: Proceedings of the 40th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 202, pp. 41770-41785. PMLR (23-29 Jul 2023), https://proceedings.mlr.press/v202/zhang23as.htm1", + "59. Zhang, Q., Chen, Y.: Fast sampling of diffusion models with exponential integrator. In: NeurIPS 2022 Workshop on Score-Based Methods (2022)", + "60. Zhang, Q., Chen, Y.: Fast sampling of diffusion models with exponential integrator. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=Loek7hfb46P", + "61. Zhang, Y., Huang, N., Tang, F., Huang, H., Ma, C., Dong, W., Xu, C.: Inversion-based style transfer with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10146-10156 (2023)", + "62. Zhao, W., Bai, L., Rao, Y., Zhou, J., Lu, J.: Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. arXiv preprint arXiv:2302.04867 (2023)", + "63. Zheng, H., Nie, W., Vahdat, A., Azizzadenesheli, K., Anandkumar, A.: Fast sampling of diffusion models via operator learning. In: International Conference on Machine Learning. pp. 42390-42402. PMLR (2023)", + "64. Zhou, L., Du, Y., Wu, J.: 3d shape generation and completion through point-voxel diffusion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5826-5835 (2021)", + "65. Zhou, Y., Liu, B., Zhu, Y., Yang, X., Chen, C., Xu, J.: Shifted diffusion for text-to-image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10157-10166 (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 811 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "C. Lu et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/ReCON_ Training-Free Acceleration for Text-to-Image Synthesis with Retrieval of Concept Prompt Trajectories/3287df3f-9065-4d00-b93c-c529c48f7c92_model.json b/2024/ReCON_ Training-Free Acceleration for Text-to-Image Synthesis with Retrieval of Concept Prompt Trajectories/3287df3f-9065-4d00-b93c-c529c48f7c92_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7eef99183786717ac3e8a383984a77bddccd93a1 --- /dev/null +++ b/2024/ReCON_ Training-Free Acceleration for Text-to-Image Synthesis with Retrieval of Concept Prompt Trajectories/3287df3f-9065-4d00-b93c-c529c48f7c92_model.json @@ -0,0 +1,2524 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.264, + 0.141, + 0.744, + 0.21 + ], + "angle": 0, + "content": "RECON: Training-Free Acceleration for Text-to-Image Synthesis with Retrieval of Concept Prompt Trajectories" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.234, + 0.783, + 0.283 + ], + "angle": 0, + "content": "Chen-Yi Lu\\*1,2, Shubham Agarwal\\*2, Md Mehrab Tanjim?, Kanak Mahadik?, Anup Rao?, Subrata Mitra?, Shiv Kumar Saini?, Saurabh Bagchi?, and Somali Chaterji" + }, + { + "type": "text", + "bbox": [ + 0.433, + 0.293, + 0.571, + 0.306 + ], + "angle": 0, + "content": "\\(^{1}\\) Purdue University" + }, + { + "type": "text", + "bbox": [ + 0.44, + 0.307, + 0.563, + 0.319 + ], + "angle": 0, + "content": "2 Adobe Research" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.357, + 0.744, + 0.608 + ], + "angle": 0, + "content": "Abstract. Text-to-image diffusion models excel in generating photorealistic images but are hampered by slow processing times. Training-free retrieval-based acceleration methods, which leverage pre-generated \"trajectories,\" have been introduced to address this. Yet, these methods often lack diversity and fidelity as they depend heavily on similarities to stored prompts. To address this, we present RECON (Retrieving Concepts), an innovative retrieval-based diffusion acceleration method that extracts visual \"concepts\" from prompts, forming a knowledge base that facilitates the creation of adaptable trajectories. Consequently, RECON surpasses existing retrieval-based methods, producing high-fidelity images and reducing required Neural Function Evaluations (NFEs) by up to \\(40\\%\\). Extensive testing on MS-COCO, Pick-a-pick, and DiffusionDB datasets confirms that RECON consistently outperforms established methods across multiple metrics such as Pick Score, CLIP Score, and Aesthetics Score. A user study further indicates that \\(76\\%\\) of images generated by RECON are rated as the highest fidelity, outperforming two competing methods, a purely text-based retrieval and a noise similarity-based retrieval. Project URL: https://stevencylu.github.io/ReCon." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.619, + 0.741, + 0.647 + ], + "angle": 0, + "content": "Keywords: Text-to-image Generation \\(\\cdot\\) Diffusion Model Acceleration \\(\\cdot\\) Retrieval-based Diffusion" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.695, + 0.377, + 0.71 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.726, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Diffusion Probabilistic Models (DMs) are a class of generative models that simulate a process of gradually reversing noise into structured outputs, drawing inspiration from thermodynamics. These models iteratively denoise an initial random noise distribution to the desired data distribution [46]. DMs have proven to be highly effective in a variety of applications, such as text-to-image generation (T2I) [11,17,38,40,41,43,56,65], speech synthesis [10,27,57], and 3D image" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.826, + 0.378, + 0.84 + ], + "angle": 0, + "content": "* Equal contribution." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "C. Lu et al." + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.147, + 0.335, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.147, + 0.418, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.147, + 0.5, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.147, + 0.584, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.147, + 0.666, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.147, + 0.747, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.221, + 0.418, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.221, + 0.5, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.221, + 0.749, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.303, + 0.306, + 0.451, + 0.312 + ], + "angle": 0, + "content": "A large white bear laying on top of a bed with pillows." + }, + { + "type": "image_caption", + "bbox": [ + 0.579, + 0.306, + 0.672, + 0.312 + ], + "angle": 0, + "content": "A photo of an old style white oven." + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.326, + 0.788, + 0.397 + ], + "angle": 0, + "content": "Fig. 1: Examples of outputs from existing retrieval-based acceleration approaches, namely Text-based (Baseline) and Noise-based (REDi [58]) Retrieval, show their difficulties in accurately representing the input prompts. On the other hand, our proposed Concept-based Retrieval framework (RECON) is able to produce high-fidelity and faithful images compared to other techniques." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.788, + 0.52 + ], + "angle": 0, + "content": "generation [21, 54, 64]. In particular, DMs like Imagen [43], DALL-E 2 [40], and StableDiffusion XL (SDXL) [38] have excelled in T2I tasks. However, DMs are notably slower than other generative models due to the extensive number of neural function evaluations (NFEs), where each NFE entails a forward network pass. As AI art and related T2I applications gain popularity, enhancing user experience with DMs through faster model inference and reducing NFEs is crucial." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.658 + ], + "angle": 0, + "content": "To address this issue, several approaches have been developed to accelerate the generation process by reducing the number of NFEs in DMs, including both training-based and training-free methods. Training-based methods, such as Distillation and Consistency Models [35,44,50], aim to distill a complex pre-trained model into a more efficient student model, capable of faster image generation with reduced NFEs. While these methods are promising, they often require a substantial amount of training time and data. This challenge becomes even more pronounced as SOTA diffusion models evolve to encompass over a billion parameters [2,38,43], intensifying the high training load." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To mitigate the demand for computing resources, training-free methods prioritize enhancing computational efficiency without necessitating additional model training. These strategies primarily concentrate on identifying more efficient numerical solvers, which reduce the number of denoising steps, thereby decreasing the NFEs required by DMs [33, 59, 62]. Our approach diverges from the more conventional focus on numerical solvers that primarily aim to streamline the denoising process. Rather, it capitalizes on leveraging previously stored intermediate noisy outputs, or \"noises,\" associated with a vast array of known prompts. For any new input prompt, retrieval-based methods retrieve a relevant noise vector from this repository of stored trajectories at a certain intermediate step. This enables the model to resume the denoising process from a later point, effectively skipping the initial steps and thus speeding up image generation. Additionally," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.377, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "RECON: Concept Retrieval for Diffusion Acceleration" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.143, + 0.761, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.346, + 0.788, + 0.43 + ], + "angle": 0, + "content": "Fig. 2: Overview of different approaches. Text-based (Baseline) and noise-based retrieval (REDi [58]) often follow rigid trajectories with extraneous elements (highlighted in red). In contrast, our method (ReCON) parses input prompts into visual concepts and selects closely related ones from a knowledge base (highlighted in green), enabling a malleable trajectory for incorporating input-specific additional details. Here, \\(0=\\text{Start}\\), \\(h=\\text{Initial}\\), \\(k=\\text{Intermediate}\\), and \\(T=\\text{Final}\\) denote the diffusion steps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.785, + 0.492 + ], + "angle": 0, + "content": "this retrieval-based approach can be integrated with any diffusion model's numerical solvers, without altering the core denoising mechanics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.644 + ], + "angle": 0, + "content": "To retrieve the nearest trajectory for an input text, a straightforward approach can be based on text-based similarity [3]. Here, we can find the prompt that is most similar to the input prompt and retrieve its associated trajectory to get the noise at an intermediate step before starting the denoising process. In theory, semantically similar prompts (e.g., \"An old man\" vs. \"A man who is old\") should yield comparable trajectories in DMs, allowing the denoising process to resume from an intermediate noise level with analogous outcomes. Different from this simple textual similarity, [58] proposed REDI, a noise-based retrieval framework. This method initiates the denoising sequence for few steps, subsequently identifying the most closely matched trajectory based on noise-wise similarity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.79, + 0.842 + ], + "angle": 0, + "content": "The efficacy of text-based retrieval (Baseline) and noise-based retrieval (REDi) methods in generating images are illustrated in Figure 1. The figure shows that these methods often struggle to produce images that are faithful to the input text. To understand this, we provide a high-level overview of each method in Figure 2. As seen from the illustration, both methods rely on stored prompts, either directly for text similarity or indirectly for noise similarity. This strategy presupposes a diverse repository capable of reflecting the myriad of unique prompts encountered during inference. However, the reality often deviates, as truly analogous prompts with matching concepts are scarce in real-world scenarios. Moreover, even when a prompt is retrieved that seems closely matched in text (Baseline) or initial noise (REDi), it might introduce extraneous concepts into the generated image. Key image attributes, such as layouts and color schemes, are predominantly determined during the early stages of the denoising" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "C. Lu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "process [15]. However, these stages are often skipped in retrieval-based methods, making it difficult to correct the trajectory using the input prompt within the constraints of reduced NFEs post-retrieval. This issue is depicted in Figure 2, where red-colored texts like \"on a bike\" (Baseline) or \"a dog\" (REDI) indicate unwanted details in the final image. Such trajectory deviations present a critical flaw in both techniques, making retrieval-based acceleration challenging." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.251, + 0.789, + 0.844 + ], + "angle": 0, + "content": "To address these challenges, we present RECON. The central idea of RECON is retrieving concepts from the input prompt, which can be thought of as \"soft templates\" for the image generation process. This method involves collecting intermediate noises corresponding to these concepts and then combining them. The combination results is what we call a \"flexible trajectory,\" which incorporates some details from the retrieved prompts while still leaving scope for additional elements that are unique from the input prompt. This is illustrated in Figure 2 under \"Concept-based Retrieval,\" where the prompt is decomposed into three generic concepts (A surfer, The beach, and A yellow surfboard) retrieved from the knowledge base. This flexibility is central to RECON's approach. Our strategy draws upon NLP research, which suggests that a sentence can be deconstructed into phrases and evaluated for various attributes such as named entities [26], abstractness [5, 47], and visualness [52]. In our method, we primarily use a pretrained language model [31] to extract essential noun phrases and determine their visualness using [52] to identify \"concepts\" (we use \"visual concepts\" and \"concepts\" interchangeably). As illustrated in Figure 1, the adoption of visual concepts allows RECON to generate images that are not only more accurate but also more faithful to the input prompts, surpassing previous methods. To the best of our knowledge, RECON is the first approach to extract visual concepts in T2I tasks for retrieval-based acceleration. Contributions. 1) We introduce RECON, first of its kind to extract concepts for accelerating T2I tasks without the need for additional training. By retrieving concepts similar to the input prompt, RECON crafts a flexible trajectory that closely aligns with the input prompt, allowing for the incorporation of unique details. 2) We conduct a comprehensive experimental study to compare RECON with both text-based (Baseline) and noise-based (REDI) [58] retrieval frameworks across three datasets: MS-COCO [7], Pick-a-Pick [23], and DiffusionDB [55], using various metrics, such as CLIP Score [39], PickScore [23], and Aesthetic Score [1]. Our findings reveal that while text-based retrieval is surprisingly robust relative to noise-based retrieval techniques (namely REDI [58]), especially with SDXL's high-fidelity model, RECON consistently excels in all measured dimensions. Specifically, RECON excels in both fidelity and speed, generating high CLIP Score images in just 6-7 seconds on an A10G GPU—faster than REDI and Baseline, which require more than 10 seconds. 3) Our contribution to the community extends beyond novel methodologies; we have curated a comprehensive concept knowledge base (KB), encompassing over 1.2M visual concepts and their trajectories derived from 1.6M unique prompts across varied datasets. This repository will be made publicly available (upon acceptance) to facilitate further research in the nascent field of retrieval-based T2I acceleration." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.377, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RECON: Concept Retrieval for Diffusion Acceleration" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.388, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.357 + ], + "angle": 0, + "content": "Training-based Acceleration. Several attempts have been made to accelerate the sampling process of diffusion models by introducing additional training beyond the base diffusion model. Distillation methods [14,30,34,35,44,63] aim to distill a teacher model into a more efficient student model with fewer denoising steps. Consistency Models [49,50], on the other hand, leverage the inherent properties of ODE samplers within diffusion models to train a new model by minimizing the difference between points along identical trajectories. Despite being able to drastically reduce generation steps (potentially to a single step), these acceleration methods require additional training. For larger base models with billions of parameters, such as SDXL [38], used in our paper, the requisite for training resources becomes even more pronounced. In contrast, our training-free, retrieval-based method can be easily implemented with a single A10G GPU." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.788, + 0.583 + ], + "angle": 0, + "content": "Training-free Acceleration. Since the introduction of interpreting diffusion models in continuous time [51], the sampling process can be executed by solving reverse SDEs or ODEs, given that the time-dependent distribution is equivalent. This enables a broad spectrum of research to investigate better numerical solvers for the reverse differential equation [22,24,33,60]. Consequently, the total number of required steps (and therefore, NFEs) can be significantly reduced, often to just 10-20. Our contribution, while complementary to the efforts in optimizing numerical solvers, introduces an additional layer of efficiency by exploiting the concept of caching and retrieving intermediate noises. This approach is orthogonal to the direct optimization of solvers because it does not alter the computational process of the samplers. Instead, it leverages the pre-generated data to bypass certain steps of the diffusion process. By caching intermediate noises we can quickly retrieve and resume the generation process from these points for new images. This method is compatible with any recent sampler, providing a versatile and powerful tool for acceleration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Retrieval-based Diffusion. Current retrieval-based T2I studies typically target generating images of rare entities [6] or adapting to new scenarios [45]. In the training-free acceleration direction, the closest work to ours is REDI [58], where the authors introduced a training-free method for image generation that leverages initial noise to retrieve the most similar generation trajectory. Our approach differs fundamentally from REDI by concurrently retrieving noises from multiple semantically similar concept prompts without the need for initial diffusion steps to obtain the noise vector, thus avoiding additional diffusion steps. Moreover, our study, using the enhanced SDXL model [38] alongside multiple datasets featuring complex prompts, demonstrates a surprising finding: simple text-based retrieval (Baseline in Fig. 2) outperform REDI when tested across extended models and datasets. Nonetheless, both REDI and the baseline approach fall short in fully capturing the intricacies of the input prompts, prompting the exploration of a more effective concept-based retrieval method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Concept Retrieval for T2I. Several studies have explored synthesizing images with user-provided concepts. These concepts can be either textual concepts [29] or image concepts [8,25,42] to generate personalized or compositionally complex" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "C. Lu et al." + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.145, + 0.761, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.326, + 0.788, + 0.41 + ], + "angle": 0, + "content": "Fig. 3: Overview of RECON. For a given input prompt, RECON decomposes it into distinct visual concepts. Our framework then stores unique concepts from known prompts as well as their trajectories for retrieval purpose. During inference, the same concept decomposition happens and the most similar concepts along with their linked noises are retrieved. These concept noise vectors are then aggregated at step \\( k \\) to resume the denoising process from \\( t = k \\) to \\( T \\), resulting in decreasing \\( k \\) NFEs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.44, + 0.789, + 0.668 + ], + "angle": 0, + "content": "images. This exploration includes methods for customizing images with specific concept images [8, 25, 42] and composing multiple text-conditioned diffusion models for intricate image compositions [29]. Our work differs in two significant ways. First, we automatically extract visual concepts from text prompts instead of relying on user-provided concepts. Second, we cache and retrieve these automatically derived concepts alongside their corresponding noise vectors for accelerating DM. To achieve automatic extraction of visual concepts, we mainly draw inspiration from existing NLP research on word and phrase abstractness and imageability [5, 19, 32, 47]. More recently, Verma et al. [52] proposed a method for evaluating prompt-level visualness in T2I, which is more relevant for our case. Inspired by this, we accomplish automatic extraction of visual concept prompts by adopting a fine-tuned CLIP [39] to gauge the text-visualness [52]. Using this, we mine around 1.2M unique concept texts, which we later use for retrieval-based acceleration. To the best of our knowledge, this is the first attempt to extract visual concepts specifically for accelerating T2I tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.691, + 0.348, + 0.709 + ], + "angle": 0, + "content": "3 Approach" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.724, + 0.373, + 0.737 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Diffusion Models. We approach diffusion models (DMs) through the lens of differential equations [22, 51], focusing on the \"trajectory\" idea within the diffusion process. Let \\(\\mathbf{x} \\sim p_{data}(\\mathbf{x})\\) represent the data distribution. DMs perturb \\(\\mathbf{x}\\) by iteratively adding i.i.d. Gaussian noise with variance \\(\\sigma_t^2\\) over \\(t \\in [T,0]\\), resulting in a noise distribution \\(p(\\mathbf{x}_0; \\sigma_0)\\). When \\(\\sigma_0^2\\) reaches a sufficiently large value \\(\\sigma_{max}^2\\), the distribution \\(p(\\mathbf{x}_0; \\sigma_0^2)\\) closely resembles isotropic Gaussian noise with" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.377, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "RECON: Concept Retrieval for Diffusion Acceleration" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.239 + ], + "angle": 0, + "content": "variance \\(\\sigma_{max}^{2}\\). Image generation in DMs starts with a noise-saturated sample \\(\\mathbf{x}_0\\) and progressively denoises to approximate \\(p(\\mathbf{x}_t; \\sigma_t)\\), where \\(t\\) ranges from 0 to \\(T\\). In this progression, \\(\\sigma_{t+1}\\) is smaller than \\(\\sigma_t\\), and \\(\\sigma_0\\) is equal to \\(\\sigma_{max}\\). The final output at \\(\\sigma_T = 0\\), should ideally reflect the original data distribution. This iterative denoising procedure is computationally represented by the Probability Flow ODE [51], which tracks the data sample trajectory over time:" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.248, + 0.786, + 0.264 + ], + "angle": 0, + "content": "\\[\nd \\mathbf {x} = - \\bar {\\sigma} (t) \\sigma (t) \\nabla_ {\\mathbf {x}} \\log p (\\mathbf {x}; \\sigma (t)) d t, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.27, + 0.789, + 0.392 + ], + "angle": 0, + "content": "where \\(\\nabla_{\\mathbf{x}}\\log p(\\mathbf{x};\\sigma)\\) denotes the score function, and \\(\\sigma (t)\\) is a user-defined noiselevel schedule with its time derivative represented by \\(\\bar{\\sigma} (t)\\). In practice, the score function is estimated by training a deep neural network \\(\\mathbf{s}_{\\theta}(\\mathbf{x};\\sigma)\\) parameterized by \\(\\theta\\). Note that we are using the same notation as Karras et al. [22], where \\(\\mathbf{x}_0\\) is the noisy input and \\(\\mathbf{x}_T\\) is the fully denoised output or generated final image. Throughout this paper, we refer to the trajectory \\(\\{\\mathbf{x}_t:t\\in \\{0\\dots T\\}\\}\\) as the one generated by a conditioned DM (typically using classifier-free guidance [18] for T2I), where \\(\\mathbf{x}_t\\) represents the noisy model outputs (noises) at each time step." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.412, + 0.5, + 0.426 + ], + "angle": 0, + "content": "3.2 Retrieval-based Acceleration" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.789, + 0.553 + ], + "angle": 0, + "content": "From Equation (1), we can see that the sampling process becomes deterministic once \\(\\mathbf{x}_0\\) is sampled from the Gaussian distribution. This allows for the pre-generation and storage of trajectories corresponding to a diverse range of prompts within a database. During inference, the key challenge lies in retrieving the trajectory that most closely aligns with the trajectory of the input prompt, posing the critical question: How do we select the optimal trajectory to resume the denoising process? To this end, we explore two primary approaches from literature, which we describe below." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.789, + 0.689 + ], + "angle": 0, + "content": "Text-based Trajectory Retrieval. Since the input prompt is the initial element, it is intuitive to retrieve trajectories associated with the closest prompts. Specifically, for an input prompt \\(\\mathcal{P}\\), we encode it with a pre-trained text encoder to obtain its embedding \\(\\phi(\\mathcal{P})\\). The nearest prompt \\(\\mathcal{P}'\\) is identified by finding the minimum cosine distance between text embeddings, and its trajectory \\(\\{\\mathbf{x}_t^{P'}\\}_{t=0}^T\\) is obtained. The intermediate output of \\(\\mathcal{P}'\\) at step \\(k\\) is looked up in the database and then, the diffusion process is resumed from the intermediate step \\(k \\in [0,T]\\) with the original prompt \\(\\mathcal{P}\\), effectively skipping \\(k\\) NFEs, as illustrated in Figure 2. We refer to this method as Baseline." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.797 + ], + "angle": 0, + "content": "Noise-based Trajectory Retrieval. The alternative method, REDI, retrieves trajectories by searching for the closest noise vector [58]. Specifically, for a given input prompt \\(\\mathcal{P}\\), REDI first computes \\(h\\) steps in the denoising process and uses the noise \\(\\mathbf{x}_h^{\\mathcal{P}}\\) as a query. It then searches for the nearest noise \\(\\mathbf{x}_h^{P'}\\) in the knowledge base and resumes the denoising process from step \\(k > h\\), effectively skipping \\(k - h\\) steps. However, this method adds \\(h\\) initial steps to the number of NFEs, as illustrated in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.815, + 0.553, + 0.831 + ], + "angle": 0, + "content": "3.3 Our Proposed Framework: RECON" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "C. Lu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "Both Baseline and REDI rely on stored prompts \\(\\mathcal{P}'\\), either directly through text similarity or indirectly through noise similarity. However, they face challenges in generating images that are faithful to the input prompt \\(\\mathcal{P}\\). As shown in Figure 2, both the baseline method and REDI introduce extraneous components into the final images, compromising its fidelity. Crucial image attributes, such as layout and color schemes, are defined in the initial denoising stages [15], which are frequently bypassed in retrieval-based approaches, leading to a misalignment from the intended prompt details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.27, + 0.789, + 0.632 + ], + "angle": 0, + "content": "Concept Extraction and KB Creation. To address the challenge of incorporating specific details from input prompts while avoiding the rigidity seen in Baseline and REDI, we introduce a novel approach: decomposing input prompts into visual concepts. However, the challenge lies in distinguishing visual concepts from abstract terms. For example, terms like \"belief,\" \"essentialness,\" and \"spirituality,\" while conceptually significant, often rank high in abstractness [5], but do not translate into visual representations for T2I generation. This distinction is crucial as language models like RoBERTa [9, 31] are proficient in extracting and understanding text, including abstract concepts, the challenge lies in ensuring that these concepts can be visually represented. To tackle this challenge, we adopt a fine-tuned CLIP model to evaluate the visual potential of extracted phrases, ensuring that only those with a clear visual representation are selected [52]. In the process of RECON, each text prompt \\(\\mathcal{P}\\) is decomposed into \\(N\\) visual concepts \\(\\{c_i\\}_{i=1}^N\\), using a pre-trained RoBERTa [31]. Subsequently, a visualness filtering model \\(h(\\cdot)\\), fine-tuned to distinguish between visually representable and non-visual text segments, evaluates these extracted noun phrases. Representations. The visualness of a concept \\(c\\) is quantified through cosine similarity \\(S = 1 - \\langle h(c), h(\\mathcal{I}_{null}) \\rangle\\), where \\(\\mathcal{I}_{null}\\) denotes an image of random RGB values, serving as a sample for non-visual content. Concepts with visualness scores \\(S\\) below a predefined threshold are deemed non-visual and excluded from further consideration. These visually relevant concepts and their trajectories from DMs are then cached in a vector database for retrieval purpose. Figure 3 illustrates these steps in this stage of RECON under the section \"Concept KB and Trajectories Creation.\"" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.634, + 0.79, + 0.763 + ], + "angle": 0, + "content": "Retrieving and Aggregating Concepts. During inference, given an input prompt \\(\\mathcal{P}\\), a concept knowledge base, and a pre-trained diffusion model \\(f_{\\theta}\\), we first decompose the prompt into a set of visual concepts \\(\\{c_i\\}_{i=1}^N\\). For each concept \\(c_i\\), we compute its text embedding \\(\\phi(c_i)\\), which serves as a query for retrieving the closest matching concepts \\(c_i'\\) along with their associated trajectories \\(\\mathbf{x}_t^{c_i'}\\). Subsequently, at a pre-determined step \\(k \\in [0,T]\\), we aggregate these trajectories by combining their noise vectors \\(\\mathbf{x}_k^{c_i'}\\) at step \\(k\\), thus forming a new starting point for the denoising process:" + }, + { + "type": "equation", + "bbox": [ + 0.442, + 0.777, + 0.786, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} _ {t} = \\sum_ {i = 1} ^ {N} w _ {i} \\cdot \\mathbf {x} _ {k} ^ {c _ {i} ^ {\\prime}}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.822, + 0.786, + 0.838 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {t + 1} = f _ {\\theta} \\left(\\hat {\\mathbf {x}} _ {t}; t, \\mathcal {P}\\right); t = k, \\dots , T - 1, \\tag {3}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.377, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "RECON: Concept Retrieval for Diffusion Acceleration" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.152, + 0.758, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.364, + 0.79, + 0.422 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparisons reveal that RECON consistently produces images of quality comparable to those generated by Vanilla, with no retrieval-based acceleration. Here, the numbers of NFEs are: Vanilla=50, Others=35. Full prompts for DiffusionDB appear in Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.451, + 0.79, + 0.707 + ], + "angle": 0, + "content": "where \\(\\{w_{i}:i\\in [1,N]\\}\\) corresponds to the weights for each concept. Our method defines weights \\(w_{i}\\) using the cosine similarity between text embeddings of the query and retrieved concepts, normalized across all concepts: \\(w_{i} = \\frac{\\left\\langle\\phi(c_{i}),\\phi(c_{i}^{\\prime})\\right\\rangle}{\\sum_{i = 1}^{N}\\left\\langle\\phi(c_{i}),\\phi(c_{i}^{\\prime})\\right\\rangle}\\). This approach is grounded in the principle that diffusion models can be regarded as energy-based models (EBMs), facilitating the compositional aggregation of noise vectors [12,28,29] For detailed insights into noise vector compositionality in EBMs, please refer to our Supplementary material. The \"Concept Retrieval\" and \"Concept Noises Aggregation\" stages precede the continuation of the denoising process with the aggregated noise \\(\\hat{\\mathbf{x}}_t\\) and the original prompt \\(\\mathcal{P}\\) to generate the final image \\(\\mathbf{x}_T\\), as depicted in Figure 3. While only the concept noun phrases are aggregated, the final diffusion steps (steps \\(k\\dots T\\) in Fig. 3) are guided by the original prompt. Thus, essential prompt details beyond nouns still get to influence the resulting image \\(\\mathbf{x}_T\\). For instance, in col. 3 of Fig. 4, although the concepts extracted from the original prompt \"A herd of sheep standing below very tall buildings\" are \"a herd of sheep\" and \"buildings\" without including \"tall,\" the final image still captures this key attribute." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.731, + 0.376, + 0.749 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.766, + 0.312, + 0.782 + ], + "angle": 0, + "content": "4.1 Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Datasets. For a fair comparison, we first assess the performance of RECON using the same dataset as REDI [58]: MS-COCO [7] (containing 82K and 42K captions in train and validation). However, captions in MS-COCO may not always" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.127 + ], + "angle": 0, + "content": "C. Lu et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.304, + 0.145, + 0.344, + 0.153 + ], + "angle": 0, + "content": "Baseline" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.154, + 0.448, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.354, + 0.193, + 0.39, + 0.198 + ], + "angle": 0, + "content": "NFE=40" + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.165, + 0.498, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.455, + 0.214, + 0.49, + 0.221 + ], + "angle": 0, + "content": "NFE=50" + }, + { + "type": "image_caption", + "bbox": [ + 0.354, + 0.238, + 0.39, + 0.245 + ], + "angle": 0, + "content": "NFE = 20" + }, + { + "type": "image_caption", + "bbox": [ + 0.313, + 0.246, + 0.484, + 0.255 + ], + "angle": 0, + "content": "A dog warrior holding a sword in pink armor." + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.146, + 0.551, + 0.153 + ], + "angle": 0, + "content": "Baseline" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.154, + 0.653, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.193, + 0.593, + 0.199 + ], + "angle": 0, + "content": "NFE = 40" + }, + { + "type": "image_caption", + "bbox": [ + 0.615, + 0.146, + 0.648, + 0.153 + ], + "angle": 0, + "content": "RECON" + }, + { + "type": "image_caption", + "bbox": [ + 0.658, + 0.165, + 0.692, + 0.172 + ], + "angle": 0, + "content": "Vanilla" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.199, + 0.653, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.239, + 0.593, + 0.245 + ], + "angle": 0, + "content": "NFE = 20" + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.172, + 0.705, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.662, + 0.214, + 0.698, + 0.221 + ], + "angle": 0, + "content": "NFE=50" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.246, + 0.7, + 0.255 + ], + "angle": 0, + "content": "A cat is sitting on a wooden café table with décor." + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.267, + 0.785, + 0.294 + ], + "angle": 0, + "content": "Fig. 5: While other methods struggle to generate faithful images with relatively higher NFEs, ours (RECON) can generate faithful images at NFEs as low as 20." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.326, + 0.785, + 0.431 + ], + "angle": 0, + "content": "reflect real-world T2I prompts, as noted in [23]. Therefore, we also use Pick-a-Pic [23] (35K unique prompts) and DiffusionDB [55] (1.8M unique prompts). These datasets offer a broader range of prompts, more reflective of actual T2I user preferences. For evaluation, we randomly select 10K captions from MS-COCO, 500 prompts from Pick-a-Pic, and 10K from DiffusionDB from their validation/test set (detailed in Supplementary). To our knowledge, such a comprehensive study has not been done before for retrieval-based acceleration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.785, + 0.612 + ], + "angle": 0, + "content": "Metrics. In unconditional settings, a popular metric to test photorealism is FID [16], which compares feature distributions from real and generated images. However, FID falls short in conditional settings such as T2I. Studies have pointed out FID is negatively correlated with human preferences [23, 38]. Therefore, for a comprehensive evaluation of T2I models, we employ metrics that specifically address the conditional nature of the task. PickScore [23] measures the alignment between text descriptions and generated images. \"CLIP Score\" [39], on the other hand, provides a broader measure of semantic congruence between the text and generated image, useful for understanding how well the generated image captures the gist of the text. \"Aesthetics Score\" [1] complements these by assessing the visual appeal of images. Finally, we employed NFE to quantify RECON's time efficiency, independent of hardware configurations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Implementation Details. For building Baseline and REDI (described in Section 3.2), we follow the steps mentioned in [58], but replace the base model with the improved SDXL [38]. We denote the default sampling process in SDXL without any acceleration as \"Vanilla Generation.\" For RECON, we first build a concept knowledge base (KB) from the unique captions and prompts in the training sets of all considered datasets. We extract noun phrases using a pretrained language model, RoBERTa [31] and apply visual filtering with a CLIP model [39] from [52]. To efficiently store and retrieve these extracted concepts and their associated trajectories, we implement a vector database with FAISS [20]. Remarkably, the concept retrieval part of our pipeline merely takes \\(1.15\\%\\) of the total inference time on an A10G GPU. We build the rest of our pipeline based on PyTorch and the HuggingFace Diffusers library [36, 37]. We determine the weights for merging concept noise vectors using the CLIP text similarity scores, as in Equation (2). Unless otherwise mentioned, for all the experiments, we use the Euler sampler [22] as default, setting the number of NFEs to 35 for retrieval-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.377, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "RECON: Concept Retrieval for Diffusion Acceleration" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.294, + 0.147, + 0.506, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.314, + 0.495, + 0.333 + ], + "angle": 0, + "content": "\"Two brown teddy bears sitting on top a wooden table.\"" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.147, + 0.71, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.314, + 0.705, + 0.345 + ], + "angle": 0, + "content": "\"renaissance painting of elder in the garden, short silver hair, a wise face, emotions closeup, dressed in roman armour.\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.36, + 0.788, + 0.404 + ], + "angle": 0, + "content": "Fig. 6: Baseline and REDI often struggle with 1st Nearest Neighbor (NN) results, leading to poorer outcomes for 2nd and 3rd NNs. Conversely, RECon consistently produces accurate images across all NN concepts, addressing the diversity challenge." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.432, + 0.788, + 0.478 + ], + "angle": 0, + "content": "based methods and 50 for Vanilla Generation. Further implementation details, including configurations of SDXL, the KB size, specifics of our vector database for caching and retrieval are provided in Supplementary." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.499, + 0.422, + 0.515 + ], + "angle": 0, + "content": "4.2 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.523, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Enhanced Fidelity. In Figure 4, Images generated using RECON demonstrate a higher degree of fidelity to the provided prompts compared to both Baseline and REDI methods. RECON images also closely mirror the results of Vanilla Generation across all datasets. In contrast, images from Baseline and REDI images exhibit notable discrepancies from the input prompt, such as depicting inaccurate objects (animal mouse vs. computer mouse), or omitting key elements described in the prompts (e.g., \"very tall buildings\"). RECON is able to generate uncommon concept combinations like \".cat eating a burger like a person.\", \".raccoon washing blankets.\", and \".cat dressed as a scuba diver..\" . REDI and Baseline often miss the mark with these intricate descriptions, occasionally generating completely irrelevant images, like those for the prompt \"Human abduction by UFO.\" RECON also performs better with rare input prompts. For instance, the phrase \"sparkling diamond gold butterfly\" is absent from KB. Nevertheless, RECON produced a relevant image of a butterfly based on the concept \"a sparkling gold glitter dress,\" thus highlighting the advantages of the flexible trajectory of RECON's design. Figure 5 showcases the outputs from three retrieval-based methods with both high and low number of NFEs. Both Baseline and REDI methods struggle to faithfully capture the input prompts in their images at relatively high number of NFEs. Unsurprisingly, their performance further diminishes at lower number of NFEs. RECON, however, consistently performs better even with NFEs as low as 20." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.127 + ], + "angle": 0, + "content": "C. Lu et al." + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.145, + 0.4, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.222, + 0.336, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.222, + 0.459, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.309, + 0.493, + 0.38 + ], + "angle": 0, + "content": "Fig.7: RECON can generate images with high CLIP scores in 6-7 seconds using an A10G GPU, whereas both REDI and Baseline require more than 10 seconds to generate images of comparable quality." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.145, + 0.635, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.145, + 0.757, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.222, + 0.634, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.222, + 0.756, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.309, + 0.788, + 0.38 + ], + "angle": 0, + "content": "Fig.8: RECON consistently outperforms Baseline and REDI on MS-COCO under varying total number of steps and different percentage of skipped steps w.r.t. total steps taken by Vanilla Generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.789, + 0.668 + ], + "angle": 0, + "content": "Diversity. Diversity in T2I generation ensures that the model can produce a range of visually distinct images for a given prompt, enhancing user experience by offering creative variations. The traditional approach for diversity in T2I tasks involves caching different trajectories for multiple seeds. This method, while capable of producing diverse outputs, faces significant scalability challenges. Storing trajectories for a growing number of unique prompts, even for a single seed, quickly becomes prohibitively expensive. A cheaper solution to this problem is to pull the top-\\(k\\) Nearest Neighbor (NN) trajectories. However, Baseline and REDI struggle with accuracy and faithfulness, particularly beyond the first NN, as shown in Figure 6. RECON effectively overcomes this by focusing on similar visual concepts rather than full-length prompts. This approach is more efficient as matching concepts is simpler than entire prompts. For instance, in one of the prompts showcased, the top-three retrieved NN concepts are: 1st: ['two brown teddy bears', 'large wooden table'], 2nd: ['two teddy bears', 'wooden kitchen table'], and 3rd: ['two stuffed teddy bears', 'wooden counter'], all closely related to each other. This similarity among the top concepts enables RECON to maintain diversity without sacrificing fidelity or facing scalability issues." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.692, + 0.446, + 0.708 + ], + "angle": 0, + "content": "4.3 Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Comparison of All Methods. In Figure 8, we show the performance with different percentage of skipped steps with varying the number of total steps (or NFEs) on MS-COCO (results from other datasets are shown in Supplementary). RECON consistently surpasses both REDI and Baseline across different NFE counts, demonstrating more notable improvement at lower number of NFEs. In Figure 7, we demonstrate the CLIP Scores of RECON in comparison to previous works. With just 6 seconds of computation time on an A10G GPU, RECON is able to produce accurate images from prompts across various datasets, while" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.377, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RECON: Concept Retrieval for Diffusion Acceleration" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.173 + ], + "angle": 0, + "content": "Table 1: Our concept-based retrieval RECon surpasses Baseline and REDI across MS-COCO, Pick-a-Pic, and DiffusionDB datasets. NFE is fixed to 35." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.185, + 0.789, + 0.268 + ], + "angle": 0, + "content": "
CLIP Score↑PickScore↑Aesthetics Score↑
COCOPick-a-PicDiffDBCOCOPick-a-PicDiffDBCOCOPick-a-PicDiffDB
Baseline29.1527.6729.6820.7420.4419.815.506.056.26
REDi18.1721.8020.4319.5119.0318.215.516.066.29
ReCon30.1929.9331.4621.3120.7620.085.526.106.31
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.288, + 0.505, + 0.386 + ], + "angle": 0, + "content": "Table 2: RECON's ablation study where 'Visual-filtering' leverages text-visualness scores [53], while 'weighting' evaluates the similarity between concepts. 'Top-1' prioritizes the highest scoring concept, while \\(>0.5\\) criterion includes only those concepts with similarities above this value." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.398, + 0.505, + 0.449 + ], + "angle": 0, + "content": "
Visual-filtering\nWeighting✓\nX (Top-1)✓\nX (>0.5)✓\nXXX✓\n✓
CLIP Score ↑27.4629.7929.3328.9829.0629.93
PickScore ↑19.8820.5920.5420.1220.3620.84
" + }, + { + "type": "table_caption", + "bbox": [ + 0.525, + 0.288, + 0.788, + 0.384 + ], + "angle": 0, + "content": "Table 3: Comparison of all methods for different samplers. RECON consistently outperforms Baseline and REDI across different diffusion samplers, showcasing RECON's ability to work with multiple efficient numerical solvers." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.398, + 0.785, + 0.448 + ], + "angle": 0, + "content": "
SchedulerCLIP Score↑PickScore↑
EulerPNDMDDIMDPMEulerPNDMDDIMDPM
Baseline27.6725.9626.7726.7320.4419.8220.1020.00
REDi21.8020.2421.6821.4219.0318.0818.4718.40
ReCON29.9327.6828.9928.5620.7619.8920.2920.06
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.485, + 0.789, + 0.607 + ], + "angle": 0, + "content": "both REDI and Baseline require over 10 seconds to generate images of similar quality. A detailed study among the retrieval-based methods is presented in Table 1. It is evident that RECON excels over Baseline and REDI in terms of all metrics. We also make a surprising discovery: Baseline is quite competitive compared to the state-of-the-art REDI. This shows that when using an improved base model, a simple approach like the Baseline can perform quite well, especially on T2I alignment scores. For rest of the analysis, we primarily report these two scores (results for the rest are in the Supplementary)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.61, + 0.789, + 0.686 + ], + "angle": 0, + "content": "Aggregation Ablation Study. Table 2 shows the importance of using both weighted aggregation Equation (2) and visual-filtering in RECON. Here, we compare with using only the most similar concept, denoted as Top-1, and excluding low-similarity concepts with a threshold of 0.5, denoted as \\((>0.5)\\). The results indicate that both the weighting and visualness filter are essential." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.79, + 0.841 + ], + "angle": 0, + "content": "Concept Retrieval Analysis. Although both Baseline and RECON rely on text similarity scores (either prompt-to-prompt in Baseline or concept-to-concept in RECON), Table 4 shows that the latter consistently outperforms the former. At each similarity score interval, concept-based retrieval in RECON leads to superior results, proving the effectiveness of our approach. The length and complexity of input prompts, often containing multiple concepts, poses the question: how many concepts to break a prompt into? Figure 9 shows that consistently deconstructing prompts into concepts (RECON) generally produces better outcomes than not doing so (Baseline). Also, it shows that there is a sweet spot for the number of concepts to decompose a text prompt into." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "C. Lu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.787, + 0.174 + ], + "angle": 0, + "content": "Table 4: RECON consistently performs better than Baseline at every text similarity score interval and every NFE count (results shown for PickScore†)." + }, + { + "type": "table", + "bbox": [ + 0.244, + 0.185, + 0.756, + 0.272 + ], + "angle": 0, + "content": "
TextNFE = 40NFE = 35NFE = 30NFE = 25
SimilarityBaselineReCONBaselineReCONBaselineReCONBaselineReCON
0.5-0.620.3821.0720.1120.9219.9520.5219.6220.18
0.6-0.720.5120.9820.0720.5919.8620.2019.5320.04
0.7-0.820.7720.9620.4420.6420.2520.3619.9220.16
0.8-0.921.2621.4220.9721.1620.8320.8920.4320.65
" + }, + { + "type": "image", + "bbox": [ + 0.301, + 0.284, + 0.502, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.284, + 0.702, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.387, + 0.787, + 0.416 + ], + "angle": 0, + "content": "Fig.9: Prompts can contain an arbitrary number of concepts; so, breaking down a prompt into its constituent concepts is better (RECon) than not doing so (Baseline)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.784, + 0.506 + ], + "angle": 0, + "content": "Comparison with Different Samplers. Table 3 showcases the results of incorporating our method into other diffusion samplers, including DDIM [48], DPM-Solver [33], PNDM [22], and Euler [22], with RECON outperforming other retrieval-based techniques across different samplers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.788, + 0.613 + ], + "angle": 0, + "content": "Human evaluation. We conducted a comprehensive user study involving 214 participants (details provided in Supplementary) to evaluate image generation using RECON, Baseline, and REDI. After filtering for consistency, we retained 200 valid participants. All methods used the same random seed, steps (NFE = 35), and Euler sampler. Results showed RECON's superior performance, with \\(75.5\\%\\) of its generated images ranked first, \\(16\\%\\) second, and \\(8.5\\%\\) third. This preference for RECON was statistically significant (p-value \\(< 0.05\\))." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.639, + 0.36, + 0.656 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We proposed RECON, a training-free retrieval-based acceleration method for T2I tasks. We observed that previous approaches were hindered by the rigidity of their trajectories, which often resulted in images lacking fidelity and diversity. To address these challenges, RECON retrieves trajectories for the noise by using concept similarity and then merges the noises derived from multiple semantically similar concepts. The visual concepts extracted for this process have broad applicability. While we focus on retrieval-based acceleration in this paper, the potential uses of these concepts can be applied to noise aesthetics classifications [13], noise inversion techniques [61], and latent-to-latent transformations within generative models [4]. We leave the exploration of our concept knowledge base as future work." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.377, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RECON: Concept Retrieval for Diffusion Acceleration" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.267 + ], + "angle": 0, + "content": "This material is based in part upon work supported by the National Science Foundation under Grant Number CNS-2146449 (NSF CAREER award), gift funding from Adobe Research, and by the Army Research Lab under contract number W911NF-2020-221. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the sponsors." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.289, + 0.323, + 0.305 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.785, + 0.36 + ], + "angle": 0, + "content": "1. christophschuhmann/improved-aesthetic-predictor: Clip+mlp aesthetic score predictor. https://github.com/christophschuhmann/improved-aesthetic-predictor" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.361, + 0.538, + 0.374 + ], + "angle": 0, + "content": "2. Dall-e 2. https://openai.com/dall-e-2 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.374, + 0.785, + 0.442 + ], + "angle": 0, + "content": "3. Agarwal, S., Mitra, S., Chakraborty, S., Karanam, S., Mukherjee, K., Saini, S.K.: Approximate caching for efficiently serving Text-to-Image diffusion models. In: 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI 24). pp. 1173-1189. USENIX Association, Santa Clara, CA (Apr 2024), https://www.usenix.org/conference/nsdi24/presentation/agarwal-shubham" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.444, + 0.785, + 0.47 + ], + "angle": 0, + "content": "4. Bojanowski, P., Joulin, A., Lopez-Paz, D., Szlam, A.: Optimizing the latent space of generative networks. arXiv preprint arXiv:1707.05776 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.471, + 0.785, + 0.511 + ], + "angle": 0, + "content": "5. Brysbaert, M., Warriner, A.B., Kuperman, V.: Concreteness ratings for 40 thousandly known english word lemmas. Behavior research methods 46, 904-911 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.512, + 0.785, + 0.552 + ], + "angle": 0, + "content": "6. Cai, D., Wang, Y., Liu, L., Shi, S.: Recent advances in retrieval-augmented text generation. In: Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval. pp. 3417-3419 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.553, + 0.785, + 0.593 + ], + "angle": 0, + "content": "7. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dollar, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.594, + 0.785, + 0.621 + ], + "angle": 0, + "content": "8. Choi, J., Choi, Y., Kim, Y., Kim, J., Yoon, S.: Custom-edit: Text-guided image editing with customized diffusion models. arXiv preprint arXiv:2305.15779 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.622, + 0.785, + 0.662 + ], + "angle": 0, + "content": "9. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.663, + 0.785, + 0.689 + ], + "angle": 0, + "content": "10. Deng, Y., Wu, N., Qiu, C., Luo, Y., Chen, Y.: Mixgan-tts: Efficient and stable speech synthesis based on diffusion model. IEEE Access (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.69, + 0.785, + 0.717 + ], + "angle": 0, + "content": "1. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.785, + 0.744 + ], + "angle": 0, + "content": "2. Du, Y., Li, S., Mordatch, I.: Compositional visual generation with energy based models. Advances in Neural Information Processing Systems 33, 6637-6647 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.785, + 0.772 + ], + "angle": 0, + "content": "3. Gallego, V.: Personalizing text-to-image generation via aesthetic gradients. arXiv preprint arXiv:2209.12330 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.773, + 0.785, + 0.812 + ], + "angle": 0, + "content": "4. Gu, J., Zhai, S., Zhang, Y., Liu, L., Susskind, J.M.: Boot: Data-free distillation of denoising diffusion models with bootstrapping. In: ICML 2023 Workshop on Structured Probabilistic Inference and Generative Modeling (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "5. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.318, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "C. Lu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "16. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.217 + ], + "angle": 0, + "content": "17. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "18. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.245, + 0.785, + 0.286 + ], + "angle": 0, + "content": "19. Jeong, J.W., Wang, X.J., Lee, D.H.: Towards measuring the visualness of a concept. In: Proceedings of the 21st ACM international conference on Information and knowledge management. pp. 2415-2418 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "20. Johnson, J., Douze, M., Jégou, H.: Billion-scale similarity search with GPUs. IEEE Transactions on Big Data 7(3), 535-547 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "21. Karnewar, A., Vedaldi, A., Novotny, D., Mitra, N.J.: Holodiffusion: Training a 3d diffusion model using 2d images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18423-18433 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.396 + ], + "angle": 0, + "content": "22. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems 35, 26565-26577 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "23. Kirstain, Y., Polyak, A., Singer, U., Matiana, S., Penna, J., Levy, O.: Pick-a-pic: An open dataset of user preferences for text-to-image generation. arXiv preprint arXiv:2305.01569 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.48 + ], + "angle": 0, + "content": "24. Kong, Z., Ping, W.: On fast sampling of diffusion probabilistic models. In: ICML Workshop on Invertible Neural Networks, Normalizing Flows, and Explicit Likelihood Models (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "25. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1931-1941 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.55 + ], + "angle": 0, + "content": "26. Li, J., Sun, A., Han, J., Li, C.: A survey on deep learning for named entity recognition. IEEE Transactions on Knowledge and Data Engineering 34(1), 50-70 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "27. Liu, J., Li, C., Ren, Y., Chen, F., Zhao, Z.: Diffsinger: Singing voice synthesis via shallow diffusion mechanism. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 11020-11028 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.632 + ], + "angle": 0, + "content": "28. Liu, N., Li, S., Du, Y., Tenenbaum, J., Torralba, A.: Learning to compose visual relations. Advances in Neural Information Processing Systems 34, 23166-23178 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.674 + ], + "angle": 0, + "content": "29. Liu, N., Li, S., Du, Y., Torralba, A., Tenenbaum, J.B.: Compositional visual generation with composable diffusion models. In: European Conference on Computer Vision. pp. 423-439. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "30. Liu, X., Zhang, X., Ma, J., Peng, J., et al.: Instafaflow: One step is enough for high-quality diffusion-based text-to-image generation. In: The Twelfth International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "31. Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., Stoyanov, V.: Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "32. Louis, A., Nenkova, A.: What makes writing great? first experiments on article quality prediction in the science journalism domain. Transactions of the Association for Computational Linguistics 1, 341-352 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "33. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems 35, 5775-5787 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.377, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RECON: Concept Retrieval for Diffusion Acceleration" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.176 + ], + "angle": 0, + "content": "34. Luhman, E., Luhman, T.: Knowledge distillation in iterative generative models for improved sampling speed. arXiv preprint arXiv:2101.02388 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.204 + ], + "angle": 0, + "content": "35. Meng, C., Rombach, R., Gao, R., Kingma, D., Ermon, S., Ho, J., Salimans, T.: On distillation of guided diffusion models. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.205, + 0.788, + 0.26 + ], + "angle": 0, + "content": "36. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.261, + 0.788, + 0.303 + ], + "angle": 0, + "content": "37. von Platen, P., Patil, S., Lozhkov, A., Cuenca, P., Lambert, N., Rasul, K., Davaadorj, M., Wolf, T.: Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.304, + 0.788, + 0.346 + ], + "angle": 0, + "content": "38. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: improving latent diffusion models for high-resolution image synthesis. pp. 1-13 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.346, + 0.788, + 0.401 + ], + "angle": 0, + "content": "39. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.402, + 0.788, + 0.444 + ], + "angle": 0, + "content": "40. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.487 + ], + "angle": 0, + "content": "41. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.487, + 0.788, + 0.543 + ], + "angle": 0, + "content": "42. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.544, + 0.788, + 0.599 + ], + "angle": 0, + "content": "43. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.6, + 0.788, + 0.627 + ], + "angle": 0, + "content": "44. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.67 + ], + "angle": 0, + "content": "45. Sheynin, S., Ashual, O., Polyak, A., Singer, U., Gafni, O., Nachmani, E., Taigman, Y.: Knn-diffusion: Image generation via large-scale retrieval. arXiv preprint arXiv:2204.02849 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.671, + 0.788, + 0.712 + ], + "angle": 0, + "content": "46. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256-2265. PMLR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.713, + 0.788, + 0.741 + ], + "angle": 0, + "content": "47. Solovyev, V.: Concreteness/abstractness concept: State of the art. In: International Conference on Cognitive Sciences. pp. 275-283. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.742, + 0.788, + 0.769 + ], + "angle": 0, + "content": "48. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: International Conference on Learning Representations (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.77, + 0.788, + 0.798 + ], + "angle": 0, + "content": "49. Song, Y., Dhariwal, P.: Improved techniques for training consistency models. In: The Twelfth International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.788, + 0.841 + ], + "angle": 0, + "content": "50. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models. Proceedings of Machine Learning Research, vol. 202. PMLR (23-29 Jul 2023), https://proceedings.mlr.press/v202/song23a.htm1" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "C. Lu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "51. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: International Conference on Learning Representations (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "52. Verma, G., Rossi, R.A., Tensmeyer, C., Gu, J., Nenkova, A.: Learning the visualness of text using large vision-language models. In: The 2023 Conference on Empirical Methods in Natural Language Processing (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "53. Verma, G., Rossi, R.A., Tensmeyer, C., Gu, J., Nenkova, A.: Learning the visualness of text using large vision-language models. arXiv preprint arXiv:2305.10434 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.327 + ], + "angle": 0, + "content": "54. Wang, T., Zhang, B., Zhang, T., Gu, S., Bao, J., Baltrusaitis, T., Shen, J., Chen, D., Wen, F., Chen, Q., et al.: Rodin: A generative model for sculpting 3d digital avatars using diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4563-4573 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "55. Wang, Z.J., Montoya, E., Munechika, D., Yang, H., Hoover, B., Chau, D.H.: Diffusiondb: A large-scale prompt gallery dataset for text-to-image generative models. arXiv preprint arXiv:2210.14896 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.411 + ], + "angle": 0, + "content": "56. Xu, X., Wang, Z., Zhang, G., Wang, K., Shi, H.: Versatile diffusion: Text, images and variations all in one diffusion model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7754-7765 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.466 + ], + "angle": 0, + "content": "57. Yasuda, Y., Toda, T.: Text-to-speech synthesis based on latent variable conversion using diffusion probabilistic model and variational autoencoder. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.535 + ], + "angle": 0, + "content": "58. Zhang, K., Yang, X., Wang, W.Y., Li, L.: ReDi: Efficient learning-free diffusion inference via trajectory retrieval. In: Proceedings of the 40th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 202, pp. 41770-41785. PMLR (23-29 Jul 2023), https://proceedings.mlr.press/v202/zhang23as.htm1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.563 + ], + "angle": 0, + "content": "59. Zhang, Q., Chen, Y.: Fast sampling of diffusion models with exponential integrator. In: NeurIPS 2022 Workshop on Score-Based Methods (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.605 + ], + "angle": 0, + "content": "60. Zhang, Q., Chen, Y.: Fast sampling of diffusion models with exponential integrator. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=Loek7hfb46P" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.647 + ], + "angle": 0, + "content": "61. Zhang, Y., Huang, N., Tang, F., Huang, H., Ma, C., Dong, W., Xu, C.: Inversion-based style transfer with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10146-10156 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "62. Zhao, W., Bai, L., Rao, Y., Zhou, J., Lu, J.: Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. arXiv preprint arXiv:2302.04867 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.73 + ], + "angle": 0, + "content": "63. Zheng, H., Nie, W., Vahdat, A., Azizzadenesheli, K., Anandkumar, A.: Fast sampling of diffusion models via operator learning. In: International Conference on Machine Learning. pp. 42390-42402. PMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.731, + 0.785, + 0.771 + ], + "angle": 0, + "content": "64. Zhou, L., Du, Y., Wu, J.: 3d shape generation and completion through point-voxel diffusion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5826-5835 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.813 + ], + "angle": 0, + "content": "65. Zhou, Y., Liu, B., Zhu, Y., Yang, X., Chen, C., Xu, J.: Shifted diffusion for text-to-image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10157-10166 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.813 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/ReCON_ Training-Free Acceleration for Text-to-Image Synthesis with Retrieval of Concept Prompt Trajectories/3287df3f-9065-4d00-b93c-c529c48f7c92_origin.pdf b/2024/ReCON_ Training-Free Acceleration for Text-to-Image Synthesis with Retrieval of Concept Prompt Trajectories/3287df3f-9065-4d00-b93c-c529c48f7c92_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f812e8f06f32e99b172c0e03ff071fcae88f9f8f --- /dev/null +++ b/2024/ReCON_ Training-Free Acceleration for Text-to-Image Synthesis with Retrieval of Concept Prompt Trajectories/3287df3f-9065-4d00-b93c-c529c48f7c92_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88fbdd9bc86a9ce18c75b6f292df4020319a1c6faea8df7c5781ab98475b69a1 +size 6675456 diff --git a/2024/ReCON_ Training-Free Acceleration for Text-to-Image Synthesis with Retrieval of Concept Prompt Trajectories/full.md b/2024/ReCON_ Training-Free Acceleration for Text-to-Image Synthesis with Retrieval of Concept Prompt Trajectories/full.md new file mode 100644 index 0000000000000000000000000000000000000000..be79a5155b15ac1c7963724305c54ec93f78a03a --- /dev/null +++ b/2024/ReCON_ Training-Free Acceleration for Text-to-Image Synthesis with Retrieval of Concept Prompt Trajectories/full.md @@ -0,0 +1,298 @@ +# RECON: Training-Free Acceleration for Text-to-Image Synthesis with Retrieval of Concept Prompt Trajectories + +Chen-Yi Lu\*1,2, Shubham Agarwal\*2, Md Mehrab Tanjim?, Kanak Mahadik?, Anup Rao?, Subrata Mitra?, Shiv Kumar Saini?, Saurabh Bagchi?, and Somali Chaterji + +$^{1}$ Purdue University + +2 Adobe Research + +Abstract. Text-to-image diffusion models excel in generating photorealistic images but are hampered by slow processing times. Training-free retrieval-based acceleration methods, which leverage pre-generated "trajectories," have been introduced to address this. Yet, these methods often lack diversity and fidelity as they depend heavily on similarities to stored prompts. To address this, we present RECON (Retrieving Concepts), an innovative retrieval-based diffusion acceleration method that extracts visual "concepts" from prompts, forming a knowledge base that facilitates the creation of adaptable trajectories. Consequently, RECON surpasses existing retrieval-based methods, producing high-fidelity images and reducing required Neural Function Evaluations (NFEs) by up to $40\%$ . Extensive testing on MS-COCO, Pick-a-pick, and DiffusionDB datasets confirms that RECON consistently outperforms established methods across multiple metrics such as Pick Score, CLIP Score, and Aesthetics Score. A user study further indicates that $76\%$ of images generated by RECON are rated as the highest fidelity, outperforming two competing methods, a purely text-based retrieval and a noise similarity-based retrieval. Project URL: https://stevencylu.github.io/ReCon. + +Keywords: Text-to-image Generation $\cdot$ Diffusion Model Acceleration $\cdot$ Retrieval-based Diffusion + +# 1 Introduction + +Diffusion Probabilistic Models (DMs) are a class of generative models that simulate a process of gradually reversing noise into structured outputs, drawing inspiration from thermodynamics. These models iteratively denoise an initial random noise distribution to the desired data distribution [46]. DMs have proven to be highly effective in a variety of applications, such as text-to-image generation (T2I) [11,17,38,40,41,43,56,65], speech synthesis [10,27,57], and 3D image + +![](images/d625f7c71f67428d7cd4833e4ec0ea8568d1711f874033bad715a4f41e45cc0c.jpg) + +![](images/47173b2ef2e837981066365b792b4d88cb7cc76c4cffeb6d4f0120d61b35bafb.jpg) + +![](images/a9cfb3408945354ec4d126241a4622fcc632b6bd318c7bbbd7e0ea41ee871e13.jpg) + +![](images/30018d2048cb34345c3da5648100c21938aab3ab5ed1dc9c36cb64a1c437145c.jpg) + +![](images/d379032f11ddcbf70a29f28e02cc423a32287987a5c5ea443d93b4189a12dde3.jpg) + +![](images/26c46ae7076ed6b84a511f1b108558767965619c5e9d25fe7507ffb48052387a.jpg) + +![](images/9d0712dcf5561005d687084ff2feaf5e4400f5dc42bfbefbacaa2524cf6f674b.jpg) +A large white bear laying on top of a bed with pillows. +Fig. 1: Examples of outputs from existing retrieval-based acceleration approaches, namely Text-based (Baseline) and Noise-based (REDi [58]) Retrieval, show their difficulties in accurately representing the input prompts. On the other hand, our proposed Concept-based Retrieval framework (RECON) is able to produce high-fidelity and faithful images compared to other techniques. + +![](images/83a7bd53b2185ebc58c1a404b0087953dc08b92257b8043e5a2370d59fe4e330.jpg) + +![](images/ba601e2486d1e7ec3fd5878202cd7a37d409cac69825f79e57120d4b72db550c.jpg) +A photo of an old style white oven. + +generation [21, 54, 64]. In particular, DMs like Imagen [43], DALL-E 2 [40], and StableDiffusion XL (SDXL) [38] have excelled in T2I tasks. However, DMs are notably slower than other generative models due to the extensive number of neural function evaluations (NFEs), where each NFE entails a forward network pass. As AI art and related T2I applications gain popularity, enhancing user experience with DMs through faster model inference and reducing NFEs is crucial. + +To address this issue, several approaches have been developed to accelerate the generation process by reducing the number of NFEs in DMs, including both training-based and training-free methods. Training-based methods, such as Distillation and Consistency Models [35,44,50], aim to distill a complex pre-trained model into a more efficient student model, capable of faster image generation with reduced NFEs. While these methods are promising, they often require a substantial amount of training time and data. This challenge becomes even more pronounced as SOTA diffusion models evolve to encompass over a billion parameters [2,38,43], intensifying the high training load. + +To mitigate the demand for computing resources, training-free methods prioritize enhancing computational efficiency without necessitating additional model training. These strategies primarily concentrate on identifying more efficient numerical solvers, which reduce the number of denoising steps, thereby decreasing the NFEs required by DMs [33, 59, 62]. Our approach diverges from the more conventional focus on numerical solvers that primarily aim to streamline the denoising process. Rather, it capitalizes on leveraging previously stored intermediate noisy outputs, or "noises," associated with a vast array of known prompts. For any new input prompt, retrieval-based methods retrieve a relevant noise vector from this repository of stored trajectories at a certain intermediate step. This enables the model to resume the denoising process from a later point, effectively skipping the initial steps and thus speeding up image generation. Additionally, + +![](images/39e90b29134abceb0168281caa3fef9dbc01fdc9fca5f7aa81a4e3035adff136.jpg) +Fig. 2: Overview of different approaches. Text-based (Baseline) and noise-based retrieval (REDi [58]) often follow rigid trajectories with extraneous elements (highlighted in red). In contrast, our method (ReCON) parses input prompts into visual concepts and selects closely related ones from a knowledge base (highlighted in green), enabling a malleable trajectory for incorporating input-specific additional details. Here, $0=\text{Start}$ , $h=\text{Initial}$ , $k=\text{Intermediate}$ , and $T=\text{Final}$ denote the diffusion steps. + +this retrieval-based approach can be integrated with any diffusion model's numerical solvers, without altering the core denoising mechanics. + +To retrieve the nearest trajectory for an input text, a straightforward approach can be based on text-based similarity [3]. Here, we can find the prompt that is most similar to the input prompt and retrieve its associated trajectory to get the noise at an intermediate step before starting the denoising process. In theory, semantically similar prompts (e.g., "An old man" vs. "A man who is old") should yield comparable trajectories in DMs, allowing the denoising process to resume from an intermediate noise level with analogous outcomes. Different from this simple textual similarity, [58] proposed REDI, a noise-based retrieval framework. This method initiates the denoising sequence for few steps, subsequently identifying the most closely matched trajectory based on noise-wise similarity. + +The efficacy of text-based retrieval (Baseline) and noise-based retrieval (REDi) methods in generating images are illustrated in Figure 1. The figure shows that these methods often struggle to produce images that are faithful to the input text. To understand this, we provide a high-level overview of each method in Figure 2. As seen from the illustration, both methods rely on stored prompts, either directly for text similarity or indirectly for noise similarity. This strategy presupposes a diverse repository capable of reflecting the myriad of unique prompts encountered during inference. However, the reality often deviates, as truly analogous prompts with matching concepts are scarce in real-world scenarios. Moreover, even when a prompt is retrieved that seems closely matched in text (Baseline) or initial noise (REDi), it might introduce extraneous concepts into the generated image. Key image attributes, such as layouts and color schemes, are predominantly determined during the early stages of the denoising + +process [15]. However, these stages are often skipped in retrieval-based methods, making it difficult to correct the trajectory using the input prompt within the constraints of reduced NFEs post-retrieval. This issue is depicted in Figure 2, where red-colored texts like "on a bike" (Baseline) or "a dog" (REDI) indicate unwanted details in the final image. Such trajectory deviations present a critical flaw in both techniques, making retrieval-based acceleration challenging. + +To address these challenges, we present RECON. The central idea of RECON is retrieving concepts from the input prompt, which can be thought of as "soft templates" for the image generation process. This method involves collecting intermediate noises corresponding to these concepts and then combining them. The combination results is what we call a "flexible trajectory," which incorporates some details from the retrieved prompts while still leaving scope for additional elements that are unique from the input prompt. This is illustrated in Figure 2 under "Concept-based Retrieval," where the prompt is decomposed into three generic concepts (A surfer, The beach, and A yellow surfboard) retrieved from the knowledge base. This flexibility is central to RECON's approach. Our strategy draws upon NLP research, which suggests that a sentence can be deconstructed into phrases and evaluated for various attributes such as named entities [26], abstractness [5, 47], and visualness [52]. In our method, we primarily use a pretrained language model [31] to extract essential noun phrases and determine their visualness using [52] to identify "concepts" (we use "visual concepts" and "concepts" interchangeably). As illustrated in Figure 1, the adoption of visual concepts allows RECON to generate images that are not only more accurate but also more faithful to the input prompts, surpassing previous methods. To the best of our knowledge, RECON is the first approach to extract visual concepts in T2I tasks for retrieval-based acceleration. Contributions. 1) We introduce RECON, first of its kind to extract concepts for accelerating T2I tasks without the need for additional training. By retrieving concepts similar to the input prompt, RECON crafts a flexible trajectory that closely aligns with the input prompt, allowing for the incorporation of unique details. 2) We conduct a comprehensive experimental study to compare RECON with both text-based (Baseline) and noise-based (REDI) [58] retrieval frameworks across three datasets: MS-COCO [7], Pick-a-Pick [23], and DiffusionDB [55], using various metrics, such as CLIP Score [39], PickScore [23], and Aesthetic Score [1]. Our findings reveal that while text-based retrieval is surprisingly robust relative to noise-based retrieval techniques (namely REDI [58]), especially with SDXL's high-fidelity model, RECON consistently excels in all measured dimensions. Specifically, RECON excels in both fidelity and speed, generating high CLIP Score images in just 6-7 seconds on an A10G GPU—faster than REDI and Baseline, which require more than 10 seconds. 3) Our contribution to the community extends beyond novel methodologies; we have curated a comprehensive concept knowledge base (KB), encompassing over 1.2M visual concepts and their trajectories derived from 1.6M unique prompts across varied datasets. This repository will be made publicly available (upon acceptance) to facilitate further research in the nascent field of retrieval-based T2I acceleration. + +# 2 Related Work + +Training-based Acceleration. Several attempts have been made to accelerate the sampling process of diffusion models by introducing additional training beyond the base diffusion model. Distillation methods [14,30,34,35,44,63] aim to distill a teacher model into a more efficient student model with fewer denoising steps. Consistency Models [49,50], on the other hand, leverage the inherent properties of ODE samplers within diffusion models to train a new model by minimizing the difference between points along identical trajectories. Despite being able to drastically reduce generation steps (potentially to a single step), these acceleration methods require additional training. For larger base models with billions of parameters, such as SDXL [38], used in our paper, the requisite for training resources becomes even more pronounced. In contrast, our training-free, retrieval-based method can be easily implemented with a single A10G GPU. + +Training-free Acceleration. Since the introduction of interpreting diffusion models in continuous time [51], the sampling process can be executed by solving reverse SDEs or ODEs, given that the time-dependent distribution is equivalent. This enables a broad spectrum of research to investigate better numerical solvers for the reverse differential equation [22,24,33,60]. Consequently, the total number of required steps (and therefore, NFEs) can be significantly reduced, often to just 10-20. Our contribution, while complementary to the efforts in optimizing numerical solvers, introduces an additional layer of efficiency by exploiting the concept of caching and retrieving intermediate noises. This approach is orthogonal to the direct optimization of solvers because it does not alter the computational process of the samplers. Instead, it leverages the pre-generated data to bypass certain steps of the diffusion process. By caching intermediate noises we can quickly retrieve and resume the generation process from these points for new images. This method is compatible with any recent sampler, providing a versatile and powerful tool for acceleration. + +Retrieval-based Diffusion. Current retrieval-based T2I studies typically target generating images of rare entities [6] or adapting to new scenarios [45]. In the training-free acceleration direction, the closest work to ours is REDI [58], where the authors introduced a training-free method for image generation that leverages initial noise to retrieve the most similar generation trajectory. Our approach differs fundamentally from REDI by concurrently retrieving noises from multiple semantically similar concept prompts without the need for initial diffusion steps to obtain the noise vector, thus avoiding additional diffusion steps. Moreover, our study, using the enhanced SDXL model [38] alongside multiple datasets featuring complex prompts, demonstrates a surprising finding: simple text-based retrieval (Baseline in Fig. 2) outperform REDI when tested across extended models and datasets. Nonetheless, both REDI and the baseline approach fall short in fully capturing the intricacies of the input prompts, prompting the exploration of a more effective concept-based retrieval method. + +Concept Retrieval for T2I. Several studies have explored synthesizing images with user-provided concepts. These concepts can be either textual concepts [29] or image concepts [8,25,42] to generate personalized or compositionally complex + +![](images/4a97beaf5867e4097e501d17e2cbc741cf03e223db4c57fcf994aafbe5c39533.jpg) +Fig. 3: Overview of RECON. For a given input prompt, RECON decomposes it into distinct visual concepts. Our framework then stores unique concepts from known prompts as well as their trajectories for retrieval purpose. During inference, the same concept decomposition happens and the most similar concepts along with their linked noises are retrieved. These concept noise vectors are then aggregated at step $k$ to resume the denoising process from $t = k$ to $T$ , resulting in decreasing $k$ NFEs. + +images. This exploration includes methods for customizing images with specific concept images [8, 25, 42] and composing multiple text-conditioned diffusion models for intricate image compositions [29]. Our work differs in two significant ways. First, we automatically extract visual concepts from text prompts instead of relying on user-provided concepts. Second, we cache and retrieve these automatically derived concepts alongside their corresponding noise vectors for accelerating DM. To achieve automatic extraction of visual concepts, we mainly draw inspiration from existing NLP research on word and phrase abstractness and imageability [5, 19, 32, 47]. More recently, Verma et al. [52] proposed a method for evaluating prompt-level visualness in T2I, which is more relevant for our case. Inspired by this, we accomplish automatic extraction of visual concept prompts by adopting a fine-tuned CLIP [39] to gauge the text-visualness [52]. Using this, we mine around 1.2M unique concept texts, which we later use for retrieval-based acceleration. To the best of our knowledge, this is the first attempt to extract visual concepts specifically for accelerating T2I tasks. + +# 3 Approach + +# 3.1 Preliminaries + +Diffusion Models. We approach diffusion models (DMs) through the lens of differential equations [22, 51], focusing on the "trajectory" idea within the diffusion process. Let $\mathbf{x} \sim p_{data}(\mathbf{x})$ represent the data distribution. DMs perturb $\mathbf{x}$ by iteratively adding i.i.d. Gaussian noise with variance $\sigma_t^2$ over $t \in [T,0]$ , resulting in a noise distribution $p(\mathbf{x}_0; \sigma_0)$ . When $\sigma_0^2$ reaches a sufficiently large value $\sigma_{max}^2$ , the distribution $p(\mathbf{x}_0; \sigma_0^2)$ closely resembles isotropic Gaussian noise with + +variance $\sigma_{max}^{2}$ . Image generation in DMs starts with a noise-saturated sample $\mathbf{x}_0$ and progressively denoises to approximate $p(\mathbf{x}_t; \sigma_t)$ , where $t$ ranges from 0 to $T$ . In this progression, $\sigma_{t+1}$ is smaller than $\sigma_t$ , and $\sigma_0$ is equal to $\sigma_{max}$ . The final output at $\sigma_T = 0$ , should ideally reflect the original data distribution. This iterative denoising procedure is computationally represented by the Probability Flow ODE [51], which tracks the data sample trajectory over time: + +$$ +d \mathbf {x} = - \bar {\sigma} (t) \sigma (t) \nabla_ {\mathbf {x}} \log p (\mathbf {x}; \sigma (t)) d t, \tag {1} +$$ + +where $\nabla_{\mathbf{x}}\log p(\mathbf{x};\sigma)$ denotes the score function, and $\sigma (t)$ is a user-defined noiselevel schedule with its time derivative represented by $\bar{\sigma} (t)$ . In practice, the score function is estimated by training a deep neural network $\mathbf{s}_{\theta}(\mathbf{x};\sigma)$ parameterized by $\theta$ . Note that we are using the same notation as Karras et al. [22], where $\mathbf{x}_0$ is the noisy input and $\mathbf{x}_T$ is the fully denoised output or generated final image. Throughout this paper, we refer to the trajectory $\{\mathbf{x}_t:t\in \{0\dots T\}\}$ as the one generated by a conditioned DM (typically using classifier-free guidance [18] for T2I), where $\mathbf{x}_t$ represents the noisy model outputs (noises) at each time step. + +# 3.2 Retrieval-based Acceleration + +From Equation (1), we can see that the sampling process becomes deterministic once $\mathbf{x}_0$ is sampled from the Gaussian distribution. This allows for the pre-generation and storage of trajectories corresponding to a diverse range of prompts within a database. During inference, the key challenge lies in retrieving the trajectory that most closely aligns with the trajectory of the input prompt, posing the critical question: How do we select the optimal trajectory to resume the denoising process? To this end, we explore two primary approaches from literature, which we describe below. + +Text-based Trajectory Retrieval. Since the input prompt is the initial element, it is intuitive to retrieve trajectories associated with the closest prompts. Specifically, for an input prompt $\mathcal{P}$ , we encode it with a pre-trained text encoder to obtain its embedding $\phi(\mathcal{P})$ . The nearest prompt $\mathcal{P}'$ is identified by finding the minimum cosine distance between text embeddings, and its trajectory $\{\mathbf{x}_t^{P'}\}_{t=0}^T$ is obtained. The intermediate output of $\mathcal{P}'$ at step $k$ is looked up in the database and then, the diffusion process is resumed from the intermediate step $k \in [0,T]$ with the original prompt $\mathcal{P}$ , effectively skipping $k$ NFEs, as illustrated in Figure 2. We refer to this method as Baseline. + +Noise-based Trajectory Retrieval. The alternative method, REDI, retrieves trajectories by searching for the closest noise vector [58]. Specifically, for a given input prompt $\mathcal{P}$ , REDI first computes $h$ steps in the denoising process and uses the noise $\mathbf{x}_h^{\mathcal{P}}$ as a query. It then searches for the nearest noise $\mathbf{x}_h^{P'}$ in the knowledge base and resumes the denoising process from step $k > h$ , effectively skipping $k - h$ steps. However, this method adds $h$ initial steps to the number of NFEs, as illustrated in Figure 2. + +# 3.3 Our Proposed Framework: RECON + +Both Baseline and REDI rely on stored prompts $\mathcal{P}'$ , either directly through text similarity or indirectly through noise similarity. However, they face challenges in generating images that are faithful to the input prompt $\mathcal{P}$ . As shown in Figure 2, both the baseline method and REDI introduce extraneous components into the final images, compromising its fidelity. Crucial image attributes, such as layout and color schemes, are defined in the initial denoising stages [15], which are frequently bypassed in retrieval-based approaches, leading to a misalignment from the intended prompt details. + +Concept Extraction and KB Creation. To address the challenge of incorporating specific details from input prompts while avoiding the rigidity seen in Baseline and REDI, we introduce a novel approach: decomposing input prompts into visual concepts. However, the challenge lies in distinguishing visual concepts from abstract terms. For example, terms like "belief," "essentialness," and "spirituality," while conceptually significant, often rank high in abstractness [5], but do not translate into visual representations for T2I generation. This distinction is crucial as language models like RoBERTa [9, 31] are proficient in extracting and understanding text, including abstract concepts, the challenge lies in ensuring that these concepts can be visually represented. To tackle this challenge, we adopt a fine-tuned CLIP model to evaluate the visual potential of extracted phrases, ensuring that only those with a clear visual representation are selected [52]. In the process of RECON, each text prompt $\mathcal{P}$ is decomposed into $N$ visual concepts $\{c_i\}_{i=1}^N$ , using a pre-trained RoBERTa [31]. Subsequently, a visualness filtering model $h(\cdot)$ , fine-tuned to distinguish between visually representable and non-visual text segments, evaluates these extracted noun phrases. Representations. The visualness of a concept $c$ is quantified through cosine similarity $S = 1 - \langle h(c), h(\mathcal{I}_{null}) \rangle$ , where $\mathcal{I}_{null}$ denotes an image of random RGB values, serving as a sample for non-visual content. Concepts with visualness scores $S$ below a predefined threshold are deemed non-visual and excluded from further consideration. These visually relevant concepts and their trajectories from DMs are then cached in a vector database for retrieval purpose. Figure 3 illustrates these steps in this stage of RECON under the section "Concept KB and Trajectories Creation." + +Retrieving and Aggregating Concepts. During inference, given an input prompt $\mathcal{P}$ , a concept knowledge base, and a pre-trained diffusion model $f_{\theta}$ , we first decompose the prompt into a set of visual concepts $\{c_i\}_{i=1}^N$ . For each concept $c_i$ , we compute its text embedding $\phi(c_i)$ , which serves as a query for retrieving the closest matching concepts $c_i'$ along with their associated trajectories $\mathbf{x}_t^{c_i'}$ . Subsequently, at a pre-determined step $k \in [0,T]$ , we aggregate these trajectories by combining their noise vectors $\mathbf{x}_k^{c_i'}$ at step $k$ , thus forming a new starting point for the denoising process: + +$$ +\hat {\mathbf {x}} _ {t} = \sum_ {i = 1} ^ {N} w _ {i} \cdot \mathbf {x} _ {k} ^ {c _ {i} ^ {\prime}}, \tag {2} +$$ + +$$ +\mathbf {x} _ {t + 1} = f _ {\theta} \left(\hat {\mathbf {x}} _ {t}; t, \mathcal {P}\right); t = k, \dots , T - 1, \tag {3} +$$ + +![](images/376c95a8294d684546c0c281399ee70dd8393733f6b89b29b3c949354e7ac95d.jpg) +Fig. 4: Qualitative comparisons reveal that RECON consistently produces images of quality comparable to those generated by Vanilla, with no retrieval-based acceleration. Here, the numbers of NFEs are: Vanilla=50, Others=35. Full prompts for DiffusionDB appear in Supplementary. + +where $\{w_{i}:i\in [1,N]\}$ corresponds to the weights for each concept. Our method defines weights $w_{i}$ using the cosine similarity between text embeddings of the query and retrieved concepts, normalized across all concepts: $w_{i} = \frac{\left\langle\phi(c_{i}),\phi(c_{i}^{\prime})\right\rangle}{\sum_{i = 1}^{N}\left\langle\phi(c_{i}),\phi(c_{i}^{\prime})\right\rangle}$ . This approach is grounded in the principle that diffusion models can be regarded as energy-based models (EBMs), facilitating the compositional aggregation of noise vectors [12,28,29] For detailed insights into noise vector compositionality in EBMs, please refer to our Supplementary material. The "Concept Retrieval" and "Concept Noises Aggregation" stages precede the continuation of the denoising process with the aggregated noise $\hat{\mathbf{x}}_t$ and the original prompt $\mathcal{P}$ to generate the final image $\mathbf{x}_T$ , as depicted in Figure 3. While only the concept noun phrases are aggregated, the final diffusion steps (steps $k\dots T$ in Fig. 3) are guided by the original prompt. Thus, essential prompt details beyond nouns still get to influence the resulting image $\mathbf{x}_T$ . For instance, in col. 3 of Fig. 4, although the concepts extracted from the original prompt "A herd of sheep standing below very tall buildings" are "a herd of sheep" and "buildings" without including "tall," the final image still captures this key attribute. + +# 4 Experiments + +# 4.1 Setup + +Datasets. For a fair comparison, we first assess the performance of RECON using the same dataset as REDI [58]: MS-COCO [7] (containing 82K and 42K captions in train and validation). However, captions in MS-COCO may not always + +![](images/fb70891a5c8405137ee032a2b34c8ea56587fb5d2f759d4ebe8455331a0c7815.jpg) +Baseline +NFE=40 +NFE = 20 + +![](images/c3715e8be9f343deebf607791b53de4232c1e18a675465d7e0ac800b8f819034.jpg) +NFE=50 + +![](images/f3c5f1dc9269a48f61d3d343cd10608cdee04201869d7655c081407a95e96baa.jpg) +Baseline +NFE = 40 +RECON + +![](images/bfaeba49e76106a84bc9703b9d5c6370e9184348edcb50dea87559f74df3be1d.jpg) +A dog warrior holding a sword in pink armor. +NFE = 20 +A cat is sitting on a wooden café table with décor. +Fig. 5: While other methods struggle to generate faithful images with relatively higher NFEs, ours (RECON) can generate faithful images at NFEs as low as 20. + +![](images/170fe417b8b96c6af0ade37d97f9a01d911f5c6a802afa62c47926861f3765f8.jpg) +Vanilla +NFE=50 + +reflect real-world T2I prompts, as noted in [23]. Therefore, we also use Pick-a-Pic [23] (35K unique prompts) and DiffusionDB [55] (1.8M unique prompts). These datasets offer a broader range of prompts, more reflective of actual T2I user preferences. For evaluation, we randomly select 10K captions from MS-COCO, 500 prompts from Pick-a-Pic, and 10K from DiffusionDB from their validation/test set (detailed in Supplementary). To our knowledge, such a comprehensive study has not been done before for retrieval-based acceleration. + +Metrics. In unconditional settings, a popular metric to test photorealism is FID [16], which compares feature distributions from real and generated images. However, FID falls short in conditional settings such as T2I. Studies have pointed out FID is negatively correlated with human preferences [23, 38]. Therefore, for a comprehensive evaluation of T2I models, we employ metrics that specifically address the conditional nature of the task. PickScore [23] measures the alignment between text descriptions and generated images. "CLIP Score" [39], on the other hand, provides a broader measure of semantic congruence between the text and generated image, useful for understanding how well the generated image captures the gist of the text. "Aesthetics Score" [1] complements these by assessing the visual appeal of images. Finally, we employed NFE to quantify RECON's time efficiency, independent of hardware configurations. + +Implementation Details. For building Baseline and REDI (described in Section 3.2), we follow the steps mentioned in [58], but replace the base model with the improved SDXL [38]. We denote the default sampling process in SDXL without any acceleration as "Vanilla Generation." For RECON, we first build a concept knowledge base (KB) from the unique captions and prompts in the training sets of all considered datasets. We extract noun phrases using a pretrained language model, RoBERTa [31] and apply visual filtering with a CLIP model [39] from [52]. To efficiently store and retrieve these extracted concepts and their associated trajectories, we implement a vector database with FAISS [20]. Remarkably, the concept retrieval part of our pipeline merely takes $1.15\%$ of the total inference time on an A10G GPU. We build the rest of our pipeline based on PyTorch and the HuggingFace Diffusers library [36, 37]. We determine the weights for merging concept noise vectors using the CLIP text similarity scores, as in Equation (2). Unless otherwise mentioned, for all the experiments, we use the Euler sampler [22] as default, setting the number of NFEs to 35 for retrieval- + +![](images/a07421fbf178f5ca0f43e06e01e0885401bafb3866bb03ae2d5314870f86e4e5.jpg) +"Two brown teddy bears sitting on top a wooden table." + +![](images/eb2485acb6bc635e380b7156f8f27baa621e67ade6ea23090c24cf12a743ff25.jpg) +"renaissance painting of elder in the garden, short silver hair, a wise face, emotions closeup, dressed in roman armour." +Fig. 6: Baseline and REDI often struggle with 1st Nearest Neighbor (NN) results, leading to poorer outcomes for 2nd and 3rd NNs. Conversely, RECon consistently produces accurate images across all NN concepts, addressing the diversity challenge. + +based methods and 50 for Vanilla Generation. Further implementation details, including configurations of SDXL, the KB size, specifics of our vector database for caching and retrieval are provided in Supplementary. + +# 4.2 Qualitative Results + +Enhanced Fidelity. In Figure 4, Images generated using RECON demonstrate a higher degree of fidelity to the provided prompts compared to both Baseline and REDI methods. RECON images also closely mirror the results of Vanilla Generation across all datasets. In contrast, images from Baseline and REDI images exhibit notable discrepancies from the input prompt, such as depicting inaccurate objects (animal mouse vs. computer mouse), or omitting key elements described in the prompts (e.g., "very tall buildings"). RECON is able to generate uncommon concept combinations like ".cat eating a burger like a person.", ".raccoon washing blankets.", and ".cat dressed as a scuba diver.." . REDI and Baseline often miss the mark with these intricate descriptions, occasionally generating completely irrelevant images, like those for the prompt "Human abduction by UFO." RECON also performs better with rare input prompts. For instance, the phrase "sparkling diamond gold butterfly" is absent from KB. Nevertheless, RECON produced a relevant image of a butterfly based on the concept "a sparkling gold glitter dress," thus highlighting the advantages of the flexible trajectory of RECON's design. Figure 5 showcases the outputs from three retrieval-based methods with both high and low number of NFEs. Both Baseline and REDI methods struggle to faithfully capture the input prompts in their images at relatively high number of NFEs. Unsurprisingly, their performance further diminishes at lower number of NFEs. RECON, however, consistently performs better even with NFEs as low as 20. + +![](images/aca80fa86833f53d0925442495975b44509e468647b58537d4b77f93f9ff56cc.jpg) + +![](images/eed3dca3f7b5d18f67cf96de738faf0756b752e4df669a1aad160b6d8b1fea7e.jpg) +Fig.7: RECON can generate images with high CLIP scores in 6-7 seconds using an A10G GPU, whereas both REDI and Baseline require more than 10 seconds to generate images of comparable quality. + +![](images/2ab8801be70883fcebcf9ead8cd58d5ea47a77b10a60d4ca8e74c74211e85117.jpg) + +![](images/90bf47d605ad220aed51e9779ed3416af8b2dc50f4cc250d8b26bffcfd27ab7b.jpg) + +![](images/ad8537c941fbd203567e7688a3251e16a5123364b292627908e9e1254df554af.jpg) + +![](images/95189773cd9e7b733d852cb9e3adc5bab7c65e9150194d319d8dcc92b6735be2.jpg) +Fig.8: RECON consistently outperforms Baseline and REDI on MS-COCO under varying total number of steps and different percentage of skipped steps w.r.t. total steps taken by Vanilla Generation. + +![](images/c5a05bf9dfea237f5c1e1d037d3bc21248c7dc186b5e9e6cb84db1ee6ed50d28.jpg) + +Diversity. Diversity in T2I generation ensures that the model can produce a range of visually distinct images for a given prompt, enhancing user experience by offering creative variations. The traditional approach for diversity in T2I tasks involves caching different trajectories for multiple seeds. This method, while capable of producing diverse outputs, faces significant scalability challenges. Storing trajectories for a growing number of unique prompts, even for a single seed, quickly becomes prohibitively expensive. A cheaper solution to this problem is to pull the top- $k$ Nearest Neighbor (NN) trajectories. However, Baseline and REDI struggle with accuracy and faithfulness, particularly beyond the first NN, as shown in Figure 6. RECON effectively overcomes this by focusing on similar visual concepts rather than full-length prompts. This approach is more efficient as matching concepts is simpler than entire prompts. For instance, in one of the prompts showcased, the top-three retrieved NN concepts are: 1st: ['two brown teddy bears', 'large wooden table'], 2nd: ['two teddy bears', 'wooden kitchen table'], and 3rd: ['two stuffed teddy bears', 'wooden counter'], all closely related to each other. This similarity among the top concepts enables RECON to maintain diversity without sacrificing fidelity or facing scalability issues. + +# 4.3 Quantitative Results + +Comparison of All Methods. In Figure 8, we show the performance with different percentage of skipped steps with varying the number of total steps (or NFEs) on MS-COCO (results from other datasets are shown in Supplementary). RECON consistently surpasses both REDI and Baseline across different NFE counts, demonstrating more notable improvement at lower number of NFEs. In Figure 7, we demonstrate the CLIP Scores of RECON in comparison to previous works. With just 6 seconds of computation time on an A10G GPU, RECON is able to produce accurate images from prompts across various datasets, while + +Table 1: Our concept-based retrieval RECon surpasses Baseline and REDI across MS-COCO, Pick-a-Pic, and DiffusionDB datasets. NFE is fixed to 35. + +
CLIP Score↑PickScore↑Aesthetics Score↑
COCOPick-a-PicDiffDBCOCOPick-a-PicDiffDBCOCOPick-a-PicDiffDB
Baseline29.1527.6729.6820.7420.4419.815.506.056.26
REDi18.1721.8020.4319.5119.0318.215.516.066.29
ReCon30.1929.9331.4621.3120.7620.085.526.106.31
+ +Table 2: RECON's ablation study where 'Visual-filtering' leverages text-visualness scores [53], while 'weighting' evaluates the similarity between concepts. 'Top-1' prioritizes the highest scoring concept, while $>0.5$ criterion includes only those concepts with similarities above this value. + +
Visual-filtering +Weighting✓ +X (Top-1)✓ +X (>0.5)✓ +XXX✓ +✓
CLIP Score ↑27.4629.7929.3328.9829.0629.93
PickScore ↑19.8820.5920.5420.1220.3620.84
+ +Table 3: Comparison of all methods for different samplers. RECON consistently outperforms Baseline and REDI across different diffusion samplers, showcasing RECON's ability to work with multiple efficient numerical solvers. + +
SchedulerCLIP Score↑PickScore↑
EulerPNDMDDIMDPMEulerPNDMDDIMDPM
Baseline27.6725.9626.7726.7320.4419.8220.1020.00
REDi21.8020.2421.6821.4219.0318.0818.4718.40
ReCON29.9327.6828.9928.5620.7619.8920.2920.06
+ +both REDI and Baseline require over 10 seconds to generate images of similar quality. A detailed study among the retrieval-based methods is presented in Table 1. It is evident that RECON excels over Baseline and REDI in terms of all metrics. We also make a surprising discovery: Baseline is quite competitive compared to the state-of-the-art REDI. This shows that when using an improved base model, a simple approach like the Baseline can perform quite well, especially on T2I alignment scores. For rest of the analysis, we primarily report these two scores (results for the rest are in the Supplementary). + +Aggregation Ablation Study. Table 2 shows the importance of using both weighted aggregation Equation (2) and visual-filtering in RECON. Here, we compare with using only the most similar concept, denoted as Top-1, and excluding low-similarity concepts with a threshold of 0.5, denoted as $(>0.5)$ . The results indicate that both the weighting and visualness filter are essential. + +Concept Retrieval Analysis. Although both Baseline and RECON rely on text similarity scores (either prompt-to-prompt in Baseline or concept-to-concept in RECON), Table 4 shows that the latter consistently outperforms the former. At each similarity score interval, concept-based retrieval in RECON leads to superior results, proving the effectiveness of our approach. The length and complexity of input prompts, often containing multiple concepts, poses the question: how many concepts to break a prompt into? Figure 9 shows that consistently deconstructing prompts into concepts (RECON) generally produces better outcomes than not doing so (Baseline). Also, it shows that there is a sweet spot for the number of concepts to decompose a text prompt into. + +Table 4: RECON consistently performs better than Baseline at every text similarity score interval and every NFE count (results shown for PickScore†). + +
TextNFE = 40NFE = 35NFE = 30NFE = 25
SimilarityBaselineReCONBaselineReCONBaselineReCONBaselineReCON
0.5-0.620.3821.0720.1120.9219.9520.5219.6220.18
0.6-0.720.5120.9820.0720.5919.8620.2019.5320.04
0.7-0.820.7720.9620.4420.6420.2520.3619.9220.16
0.8-0.921.2621.4220.9721.1620.8320.8920.4320.65
+ +![](images/9ea05ee6f773c04042fc608ef1b551cb74c338a478c38555fdebfc24926d7e11.jpg) +Fig.9: Prompts can contain an arbitrary number of concepts; so, breaking down a prompt into its constituent concepts is better (RECon) than not doing so (Baseline). + +![](images/9159d252bec14b720d161954921265999e5d7602d9158c78689a570d34ba5359.jpg) + +Comparison with Different Samplers. Table 3 showcases the results of incorporating our method into other diffusion samplers, including DDIM [48], DPM-Solver [33], PNDM [22], and Euler [22], with RECON outperforming other retrieval-based techniques across different samplers. + +Human evaluation. We conducted a comprehensive user study involving 214 participants (details provided in Supplementary) to evaluate image generation using RECON, Baseline, and REDI. After filtering for consistency, we retained 200 valid participants. All methods used the same random seed, steps (NFE = 35), and Euler sampler. Results showed RECON's superior performance, with $75.5\%$ of its generated images ranked first, $16\%$ second, and $8.5\%$ third. This preference for RECON was statistically significant (p-value $< 0.05$ ). + +# 5 Conclusion + +We proposed RECON, a training-free retrieval-based acceleration method for T2I tasks. We observed that previous approaches were hindered by the rigidity of their trajectories, which often resulted in images lacking fidelity and diversity. To address these challenges, RECON retrieves trajectories for the noise by using concept similarity and then merges the noises derived from multiple semantically similar concepts. The visual concepts extracted for this process have broad applicability. While we focus on retrieval-based acceleration in this paper, the potential uses of these concepts can be applied to noise aesthetics classifications [13], noise inversion techniques [61], and latent-to-latent transformations within generative models [4]. We leave the exploration of our concept knowledge base as future work. + +# Acknowledgement + +This material is based in part upon work supported by the National Science Foundation under Grant Number CNS-2146449 (NSF CAREER award), gift funding from Adobe Research, and by the Army Research Lab under contract number W911NF-2020-221. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the sponsors. + +# References + +1. christophschuhmann/improved-aesthetic-predictor: Clip+mlp aesthetic score predictor. https://github.com/christophschuhmann/improved-aesthetic-predictor +2. Dall-e 2. https://openai.com/dall-e-2 (2023) +3. Agarwal, S., Mitra, S., Chakraborty, S., Karanam, S., Mukherjee, K., Saini, S.K.: Approximate caching for efficiently serving Text-to-Image diffusion models. In: 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI 24). pp. 1173-1189. USENIX Association, Santa Clara, CA (Apr 2024), https://www.usenix.org/conference/nsdi24/presentation/agarwal-shubham +4. Bojanowski, P., Joulin, A., Lopez-Paz, D., Szlam, A.: Optimizing the latent space of generative networks. arXiv preprint arXiv:1707.05776 (2017) +5. Brysbaert, M., Warriner, A.B., Kuperman, V.: Concreteness ratings for 40 thousandly known english word lemmas. Behavior research methods 46, 904-911 (2014) +6. Cai, D., Wang, Y., Liu, L., Shi, S.: Recent advances in retrieval-augmented text generation. In: Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval. pp. 3417-3419 (2022) +7. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dollar, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325 (2015) +8. Choi, J., Choi, Y., Kim, Y., Kim, J., Yoon, S.: Custom-edit: Text-guided image editing with customized diffusion models. arXiv preprint arXiv:2305.15779 (2023) +9. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022) +10. Deng, Y., Wu, N., Qiu, C., Luo, Y., Chen, Y.: Mixgan-tts: Efficient and stable speech synthesis based on diffusion model. IEEE Access (2023) +1. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021) +2. Du, Y., Li, S., Mordatch, I.: Compositional visual generation with energy based models. Advances in Neural Information Processing Systems 33, 6637-6647 (2020) +3. Gallego, V.: Personalizing text-to-image generation via aesthetic gradients. arXiv preprint arXiv:2209.12330 (2022) +4. Gu, J., Zhai, S., Zhang, Y., Liu, L., Susskind, J.M.: Boot: Data-free distillation of denoising diffusion models with bootstrapping. In: ICML 2023 Workshop on Structured Probabilistic Inference and Generative Modeling (2023) +5. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control (2022) + +16. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017) +17. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020) +18. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022) +19. Jeong, J.W., Wang, X.J., Lee, D.H.: Towards measuring the visualness of a concept. In: Proceedings of the 21st ACM international conference on Information and knowledge management. pp. 2415-2418 (2012) +20. Johnson, J., Douze, M., Jégou, H.: Billion-scale similarity search with GPUs. IEEE Transactions on Big Data 7(3), 535-547 (2019) +21. Karnewar, A., Vedaldi, A., Novotny, D., Mitra, N.J.: Holodiffusion: Training a 3d diffusion model using 2d images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18423-18433 (2023) +22. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems 35, 26565-26577 (2022) +23. Kirstain, Y., Polyak, A., Singer, U., Matiana, S., Penna, J., Levy, O.: Pick-a-pic: An open dataset of user preferences for text-to-image generation. arXiv preprint arXiv:2305.01569 (2023) +24. Kong, Z., Ping, W.: On fast sampling of diffusion probabilistic models. In: ICML Workshop on Invertible Neural Networks, Normalizing Flows, and Explicit Likelihood Models (2021) +25. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1931-1941 (2023) +26. Li, J., Sun, A., Han, J., Li, C.: A survey on deep learning for named entity recognition. IEEE Transactions on Knowledge and Data Engineering 34(1), 50-70 (2020) +27. Liu, J., Li, C., Ren, Y., Chen, F., Zhao, Z.: Diffsinger: Singing voice synthesis via shallow diffusion mechanism. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 11020-11028 (2022) +28. Liu, N., Li, S., Du, Y., Tenenbaum, J., Torralba, A.: Learning to compose visual relations. Advances in Neural Information Processing Systems 34, 23166-23178 (2021) +29. Liu, N., Li, S., Du, Y., Torralba, A., Tenenbaum, J.B.: Compositional visual generation with composable diffusion models. In: European Conference on Computer Vision. pp. 423-439. Springer (2022) +30. Liu, X., Zhang, X., Ma, J., Peng, J., et al.: Instafaflow: One step is enough for high-quality diffusion-based text-to-image generation. In: The Twelfth International Conference on Learning Representations (2023) +31. Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., Stoyanov, V.: Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692 (2019) +32. Louis, A., Nenkova, A.: What makes writing great? first experiments on article quality prediction in the science journalism domain. Transactions of the Association for Computational Linguistics 1, 341-352 (2013) +33. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems 35, 5775-5787 (2022) + +34. Luhman, E., Luhman, T.: Knowledge distillation in iterative generative models for improved sampling speed. arXiv preprint arXiv:2101.02388 (2021) +35. Meng, C., Rombach, R., Gao, R., Kingma, D., Ermon, S., Ho, J., Salimans, T.: On distillation of guided diffusion models. In: CVPR (2023) +36. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems 32 (2019) +37. von Platen, P., Patil, S., Lozhkov, A., Cuenca, P., Lambert, N., Rasul, K., Davaadorj, M., Wolf, T.: Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers (2022) +38. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: improving latent diffusion models for high-resolution image synthesis. pp. 1-13 (2024) +39. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +40. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022) +41. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) +42. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023) +43. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022) +44. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models (2022) +45. Sheynin, S., Ashual, O., Polyak, A., Singer, U., Gafni, O., Nachmani, E., Taigman, Y.: Knn-diffusion: Image generation via large-scale retrieval. arXiv preprint arXiv:2204.02849 (2022) +46. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256-2265. PMLR (2015) +47. Solovyev, V.: Concreteness/abstractness concept: State of the art. In: International Conference on Cognitive Sciences. pp. 275-283. Springer (2020) +48. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: International Conference on Learning Representations (2020) +49. Song, Y., Dhariwal, P.: Improved techniques for training consistency models. In: The Twelfth International Conference on Learning Representations (2023) +50. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models. Proceedings of Machine Learning Research, vol. 202. PMLR (23-29 Jul 2023), https://proceedings.mlr.press/v202/song23a.htm1 + +51. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: International Conference on Learning Representations (2020) +52. Verma, G., Rossi, R.A., Tensmeyer, C., Gu, J., Nenkova, A.: Learning the visualness of text using large vision-language models. In: The 2023 Conference on Empirical Methods in Natural Language Processing (2023) +53. Verma, G., Rossi, R.A., Tensmeyer, C., Gu, J., Nenkova, A.: Learning the visualness of text using large vision-language models. arXiv preprint arXiv:2305.10434 (2023) +54. Wang, T., Zhang, B., Zhang, T., Gu, S., Bao, J., Baltrusaitis, T., Shen, J., Chen, D., Wen, F., Chen, Q., et al.: Rodin: A generative model for sculpting 3d digital avatars using diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4563-4573 (2023) +55. Wang, Z.J., Montoya, E., Munechika, D., Yang, H., Hoover, B., Chau, D.H.: Diffusiondb: A large-scale prompt gallery dataset for text-to-image generative models. arXiv preprint arXiv:2210.14896 (2022) +56. Xu, X., Wang, Z., Zhang, G., Wang, K., Shi, H.: Versatile diffusion: Text, images and variations all in one diffusion model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7754-7765 (2023) +57. Yasuda, Y., Toda, T.: Text-to-speech synthesis based on latent variable conversion using diffusion probabilistic model and variational autoencoder. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023) +58. Zhang, K., Yang, X., Wang, W.Y., Li, L.: ReDi: Efficient learning-free diffusion inference via trajectory retrieval. In: Proceedings of the 40th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 202, pp. 41770-41785. PMLR (23-29 Jul 2023), https://proceedings.mlr.press/v202/zhang23as.htm1 +59. Zhang, Q., Chen, Y.: Fast sampling of diffusion models with exponential integrator. In: NeurIPS 2022 Workshop on Score-Based Methods (2022) +60. Zhang, Q., Chen, Y.: Fast sampling of diffusion models with exponential integrator. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=Loek7hfb46P +61. Zhang, Y., Huang, N., Tang, F., Huang, H., Ma, C., Dong, W., Xu, C.: Inversion-based style transfer with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10146-10156 (2023) +62. Zhao, W., Bai, L., Rao, Y., Zhou, J., Lu, J.: Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. arXiv preprint arXiv:2302.04867 (2023) +63. Zheng, H., Nie, W., Vahdat, A., Azizzadenesheli, K., Anandkumar, A.: Fast sampling of diffusion models via operator learning. In: International Conference on Machine Learning. pp. 42390-42402. PMLR (2023) +64. Zhou, L., Du, Y., Wu, J.: 3d shape generation and completion through point-voxel diffusion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5826-5835 (2021) +65. Zhou, Y., Liu, B., Zhu, Y., Yang, X., Chen, C., Xu, J.: Shifted diffusion for text-to-image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10157-10166 (2023) \ No newline at end of file diff --git a/2024/ReCON_ Training-Free Acceleration for Text-to-Image Synthesis with Retrieval of Concept Prompt Trajectories/images.zip b/2024/ReCON_ Training-Free Acceleration for Text-to-Image Synthesis with Retrieval of Concept Prompt Trajectories/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..132417dc17d9874895235ddcfa0209ab94921b2a --- /dev/null +++ b/2024/ReCON_ Training-Free Acceleration for Text-to-Image Synthesis with Retrieval of Concept Prompt Trajectories/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f7128c6213141c59de13a927ec3ce7c147fd6d7aa9c541d5cf115d3bd82c1c01 +size 549177 diff --git a/2024/ReCON_ Training-Free Acceleration for Text-to-Image Synthesis with Retrieval of Concept Prompt Trajectories/layout.json b/2024/ReCON_ Training-Free Acceleration for Text-to-Image Synthesis with Retrieval of Concept Prompt Trajectories/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..629a2db334b08191046da11d4720c308f2d75aad --- /dev/null +++ b/2024/ReCON_ Training-Free Acceleration for Text-to-Image Synthesis with Retrieval of Concept Prompt Trajectories/layout.json @@ -0,0 +1,9827 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 161, + 111, + 455, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 111, + 455, + 166 + ], + "spans": [ + { + "bbox": [ + 161, + 111, + 455, + 166 + ], + "type": "text", + "content": "RECON: Training-Free Acceleration for Text-to-Image Synthesis with Retrieval of Concept Prompt Trajectories" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 134, + 185, + 479, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 185, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 134, + 185, + 479, + 224 + ], + "type": "text", + "content": "Chen-Yi Lu\\*1,2, Shubham Agarwal\\*2, Md Mehrab Tanjim?, Kanak Mahadik?, Anup Rao?, Subrata Mitra?, Shiv Kumar Saini?, Saurabh Bagchi?, and Somali Chaterji" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 264, + 232, + 349, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 232, + 349, + 242 + ], + "spans": [ + { + "bbox": [ + 264, + 232, + 349, + 242 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 264, + 232, + 349, + 242 + ], + "type": "text", + "content": " Purdue University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 269, + 243, + 344, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 243, + 344, + 252 + ], + "spans": [ + { + "bbox": [ + 269, + 243, + 344, + 252 + ], + "type": "text", + "content": "2 Adobe Research" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 282, + 455, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 282, + 455, + 481 + ], + "spans": [ + { + "bbox": [ + 159, + 282, + 455, + 481 + ], + "type": "text", + "content": "Abstract. Text-to-image diffusion models excel in generating photorealistic images but are hampered by slow processing times. Training-free retrieval-based acceleration methods, which leverage pre-generated \"trajectories,\" have been introduced to address this. Yet, these methods often lack diversity and fidelity as they depend heavily on similarities to stored prompts. To address this, we present RECON (Retrieving Concepts), an innovative retrieval-based diffusion acceleration method that extracts visual \"concepts\" from prompts, forming a knowledge base that facilitates the creation of adaptable trajectories. Consequently, RECON surpasses existing retrieval-based methods, producing high-fidelity images and reducing required Neural Function Evaluations (NFEs) by up to " + }, + { + "bbox": [ + 159, + 282, + 455, + 481 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 159, + 282, + 455, + 481 + ], + "type": "text", + "content": ". Extensive testing on MS-COCO, Pick-a-pick, and DiffusionDB datasets confirms that RECON consistently outperforms established methods across multiple metrics such as Pick Score, CLIP Score, and Aesthetics Score. A user study further indicates that " + }, + { + "bbox": [ + 159, + 282, + 455, + 481 + ], + "type": "inline_equation", + "content": "76\\%" + }, + { + "bbox": [ + 159, + 282, + 455, + 481 + ], + "type": "text", + "content": " of images generated by RECON are rated as the highest fidelity, outperforming two competing methods, a purely text-based retrieval and a noise similarity-based retrieval. Project URL: https://stevencylu.github.io/ReCon." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 490, + 453, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 490, + 453, + 512 + ], + "spans": [ + { + "bbox": [ + 159, + 490, + 453, + 512 + ], + "type": "text", + "content": "Keywords: Text-to-image Generation " + }, + { + "bbox": [ + 159, + 490, + 453, + 512 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 490, + 453, + 512 + ], + "type": "text", + "content": " Diffusion Model Acceleration " + }, + { + "bbox": [ + 159, + 490, + 453, + 512 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 490, + 453, + 512 + ], + "type": "text", + "content": " Retrieval-based Diffusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 550, + 230, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 550, + 230, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 550, + 230, + 562 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "content": "Diffusion Probabilistic Models (DMs) are a class of generative models that simulate a process of gradually reversing noise into structured outputs, drawing inspiration from thermodynamics. These models iteratively denoise an initial random noise distribution to the desired data distribution [46]. DMs have proven to be highly effective in a variety of applications, such as text-to-image generation (T2I) [11,17,38,40,41,43,56,65], speech synthesis [10,27,57], and 3D image" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 654, + 231, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 231, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 231, + 665 + ], + "type": "text", + "content": "* Equal contribution." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 156, + 116, + 205, + 173 + ], + "blocks": [ + { + "bbox": [ + 156, + 116, + 205, + 173 + ], + "lines": [ + { + "bbox": [ + 156, + 116, + 205, + 173 + ], + "spans": [ + { + "bbox": [ + 156, + 116, + 205, + 173 + ], + "type": "image", + "image_path": "d625f7c71f67428d7cd4833e4ec0ea8568d1711f874033bad715a4f41e45cc0c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 206, + 116, + 255, + 175 + ], + "blocks": [ + { + "bbox": [ + 206, + 116, + 255, + 175 + ], + "lines": [ + { + "bbox": [ + 206, + 116, + 255, + 175 + ], + "spans": [ + { + "bbox": [ + 206, + 116, + 255, + 175 + ], + "type": "image", + "image_path": "47173b2ef2e837981066365b792b4d88cb7cc76c4cffeb6d4f0120d61b35bafb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 257, + 116, + 306, + 175 + ], + "blocks": [ + { + "bbox": [ + 257, + 116, + 306, + 175 + ], + "lines": [ + { + "bbox": [ + 257, + 116, + 306, + 175 + ], + "spans": [ + { + "bbox": [ + 257, + 116, + 306, + 175 + ], + "type": "image", + "image_path": "a9cfb3408945354ec4d126241a4622fcc632b6bd318c7bbbd7e0ea41ee871e13.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 116, + 357, + 175 + ], + "blocks": [ + { + "bbox": [ + 307, + 116, + 357, + 175 + ], + "lines": [ + { + "bbox": [ + 307, + 116, + 357, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 116, + 357, + 175 + ], + "type": "image", + "image_path": "30018d2048cb34345c3da5648100c21938aab3ab5ed1dc9c36cb64a1c437145c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 359, + 116, + 407, + 175 + ], + "blocks": [ + { + "bbox": [ + 359, + 116, + 407, + 175 + ], + "lines": [ + { + "bbox": [ + 359, + 116, + 407, + 175 + ], + "spans": [ + { + "bbox": [ + 359, + 116, + 407, + 175 + ], + "type": "image", + "image_path": "d379032f11ddcbf70a29f28e02cc423a32287987a5c5ea443d93b4189a12dde3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 408, + 116, + 457, + 175 + ], + "blocks": [ + { + "bbox": [ + 408, + 116, + 457, + 175 + ], + "lines": [ + { + "bbox": [ + 408, + 116, + 457, + 175 + ], + "spans": [ + { + "bbox": [ + 408, + 116, + 457, + 175 + ], + "type": "image", + "image_path": "26c46ae7076ed6b84a511f1b108558767965619c5e9d25fe7507ffb48052387a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 156, + 175, + 255, + 242 + ], + "blocks": [ + { + "bbox": [ + 156, + 175, + 255, + 242 + ], + "lines": [ + { + "bbox": [ + 156, + 175, + 255, + 242 + ], + "spans": [ + { + "bbox": [ + 156, + 175, + 255, + 242 + ], + "type": "image", + "image_path": "9d0712dcf5561005d687084ff2feaf5e4400f5dc42bfbefbacaa2524cf6f674b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 185, + 242, + 276, + 247 + ], + "lines": [ + { + "bbox": [ + 185, + 242, + 276, + 247 + ], + "spans": [ + { + "bbox": [ + 185, + 242, + 276, + 247 + ], + "type": "text", + "content": "A large white bear laying on top of a bed with pillows." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 258, + 482, + 314 + ], + "lines": [ + { + "bbox": [ + 131, + 258, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 131, + 258, + 482, + 314 + ], + "type": "text", + "content": "Fig. 1: Examples of outputs from existing retrieval-based acceleration approaches, namely Text-based (Baseline) and Noise-based (REDi [58]) Retrieval, show their difficulties in accurately representing the input prompts. On the other hand, our proposed Concept-based Retrieval framework (RECON) is able to produce high-fidelity and faithful images compared to other techniques." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 257, + 175, + 306, + 242 + ], + "blocks": [ + { + "bbox": [ + 257, + 175, + 306, + 242 + ], + "lines": [ + { + "bbox": [ + 257, + 175, + 306, + 242 + ], + "spans": [ + { + "bbox": [ + 257, + 175, + 306, + 242 + ], + "type": "image", + "image_path": "83a7bd53b2185ebc58c1a404b0087953dc08b92257b8043e5a2370d59fe4e330.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 175, + 458, + 242 + ], + "blocks": [ + { + "bbox": [ + 307, + 175, + 458, + 242 + ], + "lines": [ + { + "bbox": [ + 307, + 175, + 458, + 242 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 458, + 242 + ], + "type": "image", + "image_path": "ba601e2486d1e7ec3fd5878202cd7a37d409cac69825f79e57120d4b72db550c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 242, + 411, + 247 + ], + "lines": [ + { + "bbox": [ + 354, + 242, + 411, + 247 + ], + "spans": [ + { + "bbox": [ + 354, + 242, + 411, + 247 + ], + "type": "text", + "content": "A photo of an old style white oven." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 339, + 482, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 482, + 411 + ], + "type": "text", + "content": "generation [21, 54, 64]. In particular, DMs like Imagen [43], DALL-E 2 [40], and StableDiffusion XL (SDXL) [38] have excelled in T2I tasks. However, DMs are notably slower than other generative models due to the extensive number of neural function evaluations (NFEs), where each NFE entails a forward network pass. As AI art and related T2I applications gain popularity, enhancing user experience with DMs through faster model inference and reducing NFEs is crucial." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 413, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 521 + ], + "type": "text", + "content": "To address this issue, several approaches have been developed to accelerate the generation process by reducing the number of NFEs in DMs, including both training-based and training-free methods. Training-based methods, such as Distillation and Consistency Models [35,44,50], aim to distill a complex pre-trained model into a more efficient student model, capable of faster image generation with reduced NFEs. While these methods are promising, they often require a substantial amount of training time and data. This challenge becomes even more pronounced as SOTA diffusion models evolve to encompass over a billion parameters [2,38,43], intensifying the high training load." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "content": "To mitigate the demand for computing resources, training-free methods prioritize enhancing computational efficiency without necessitating additional model training. These strategies primarily concentrate on identifying more efficient numerical solvers, which reduce the number of denoising steps, thereby decreasing the NFEs required by DMs [33, 59, 62]. Our approach diverges from the more conventional focus on numerical solvers that primarily aim to streamline the denoising process. Rather, it capitalizes on leveraging previously stored intermediate noisy outputs, or \"noises,\" associated with a vast array of known prompts. For any new input prompt, retrieval-based methods retrieve a relevant noise vector from this repository of stored trajectories at a certain intermediate step. This enables the model to resume the denoising process from a later point, effectively skipping the initial steps and thus speeding up image generation. Additionally," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "C. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 113, + 465, + 265 + ], + "blocks": [ + { + "bbox": [ + 151, + 113, + 465, + 265 + ], + "lines": [ + { + "bbox": [ + 151, + 113, + 465, + 265 + ], + "spans": [ + { + "bbox": [ + 151, + 113, + 465, + 265 + ], + "type": "image", + "image_path": "39e90b29134abceb0168281caa3fef9dbc01fdc9fca5f7aa81a4e3035adff136.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 274, + 482, + 340 + ], + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 340 + ], + "type": "text", + "content": "Fig. 2: Overview of different approaches. Text-based (Baseline) and noise-based retrieval (REDi [58]) often follow rigid trajectories with extraneous elements (highlighted in red). In contrast, our method (ReCON) parses input prompts into visual concepts and selects closely related ones from a knowledge base (highlighted in green), enabling a malleable trajectory for incorporating input-specific additional details. Here, " + }, + { + "bbox": [ + 130, + 274, + 482, + 340 + ], + "type": "inline_equation", + "content": "0=\\text{Start}" + }, + { + "bbox": [ + 130, + 274, + 482, + 340 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 274, + 482, + 340 + ], + "type": "inline_equation", + "content": "h=\\text{Initial}" + }, + { + "bbox": [ + 130, + 274, + 482, + 340 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 274, + 482, + 340 + ], + "type": "inline_equation", + "content": "k=\\text{Intermediate}" + }, + { + "bbox": [ + 130, + 274, + 482, + 340 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 274, + 482, + 340 + ], + "type": "inline_equation", + "content": "T=\\text{Final}" + }, + { + "bbox": [ + 130, + 274, + 482, + 340 + ], + "type": "text", + "content": " denote the diffusion steps." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 365, + 480, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 480, + 389 + ], + "type": "text", + "content": "this retrieval-based approach can be integrated with any diffusion model's numerical solvers, without altering the core denoising mechanics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 390, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 510 + ], + "type": "text", + "content": "To retrieve the nearest trajectory for an input text, a straightforward approach can be based on text-based similarity [3]. Here, we can find the prompt that is most similar to the input prompt and retrieve its associated trajectory to get the noise at an intermediate step before starting the denoising process. In theory, semantically similar prompts (e.g., \"An old man\" vs. \"A man who is old\") should yield comparable trajectories in DMs, allowing the denoising process to resume from an intermediate noise level with analogous outcomes. Different from this simple textual similarity, [58] proposed REDI, a noise-based retrieval framework. This method initiates the denoising sequence for few steps, subsequently identifying the most closely matched trajectory based on noise-wise similarity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 483, + 666 + ], + "type": "text", + "content": "The efficacy of text-based retrieval (Baseline) and noise-based retrieval (REDi) methods in generating images are illustrated in Figure 1. The figure shows that these methods often struggle to produce images that are faithful to the input text. To understand this, we provide a high-level overview of each method in Figure 2. As seen from the illustration, both methods rely on stored prompts, either directly for text similarity or indirectly for noise similarity. This strategy presupposes a diverse repository capable of reflecting the myriad of unique prompts encountered during inference. However, the reality often deviates, as truly analogous prompts with matching concepts are scarce in real-world scenarios. Moreover, even when a prompt is retrieved that seems closely matched in text (Baseline) or initial noise (REDi), it might introduce extraneous concepts into the generated image. Key image attributes, such as layouts and color schemes, are predominantly determined during the early stages of the denoising" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "type": "text", + "content": "RECON: Concept Retrieval for Diffusion Acceleration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "process [15]. However, these stages are often skipped in retrieval-based methods, making it difficult to correct the trajectory using the input prompt within the constraints of reduced NFEs post-retrieval. This issue is depicted in Figure 2, where red-colored texts like \"on a bike\" (Baseline) or \"a dog\" (REDI) indicate unwanted details in the final image. Such trajectory deviations present a critical flaw in both techniques, making retrieval-based acceleration challenging." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 198, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 198, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 482, + 668 + ], + "type": "text", + "content": "To address these challenges, we present RECON. The central idea of RECON is retrieving concepts from the input prompt, which can be thought of as \"soft templates\" for the image generation process. This method involves collecting intermediate noises corresponding to these concepts and then combining them. The combination results is what we call a \"flexible trajectory,\" which incorporates some details from the retrieved prompts while still leaving scope for additional elements that are unique from the input prompt. This is illustrated in Figure 2 under \"Concept-based Retrieval,\" where the prompt is decomposed into three generic concepts (A surfer, The beach, and A yellow surfboard) retrieved from the knowledge base. This flexibility is central to RECON's approach. Our strategy draws upon NLP research, which suggests that a sentence can be deconstructed into phrases and evaluated for various attributes such as named entities [26], abstractness [5, 47], and visualness [52]. In our method, we primarily use a pretrained language model [31] to extract essential noun phrases and determine their visualness using [52] to identify \"concepts\" (we use \"visual concepts\" and \"concepts\" interchangeably). As illustrated in Figure 1, the adoption of visual concepts allows RECON to generate images that are not only more accurate but also more faithful to the input prompts, surpassing previous methods. To the best of our knowledge, RECON is the first approach to extract visual concepts in T2I tasks for retrieval-based acceleration. Contributions. 1) We introduce RECON, first of its kind to extract concepts for accelerating T2I tasks without the need for additional training. By retrieving concepts similar to the input prompt, RECON crafts a flexible trajectory that closely aligns with the input prompt, allowing for the incorporation of unique details. 2) We conduct a comprehensive experimental study to compare RECON with both text-based (Baseline) and noise-based (REDI) [58] retrieval frameworks across three datasets: MS-COCO [7], Pick-a-Pick [23], and DiffusionDB [55], using various metrics, such as CLIP Score [39], PickScore [23], and Aesthetic Score [1]. Our findings reveal that while text-based retrieval is surprisingly robust relative to noise-based retrieval techniques (namely REDI [58]), especially with SDXL's high-fidelity model, RECON consistently excels in all measured dimensions. Specifically, RECON excels in both fidelity and speed, generating high CLIP Score images in just 6-7 seconds on an A10G GPU—faster than REDI and Baseline, which require more than 10 seconds. 3) Our contribution to the community extends beyond novel methodologies; we have curated a comprehensive concept knowledge base (KB), encompassing over 1.2M visual concepts and their trajectories derived from 1.6M unique prompts across varied datasets. This repository will be made publicly available (upon acceptance) to facilitate further research in the nascent field of retrieval-based T2I acceleration." + } + ] + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "C. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 282 + ], + "type": "text", + "content": "Training-based Acceleration. Several attempts have been made to accelerate the sampling process of diffusion models by introducing additional training beyond the base diffusion model. Distillation methods [14,30,34,35,44,63] aim to distill a teacher model into a more efficient student model with fewer denoising steps. Consistency Models [49,50], on the other hand, leverage the inherent properties of ODE samplers within diffusion models to train a new model by minimizing the difference between points along identical trajectories. Despite being able to drastically reduce generation steps (potentially to a single step), these acceleration methods require additional training. For larger base models with billions of parameters, such as SDXL [38], used in our paper, the requisite for training resources becomes even more pronounced. In contrast, our training-free, retrieval-based method can be easily implemented with a single A10G GPU." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 283, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 461 + ], + "type": "text", + "content": "Training-free Acceleration. Since the introduction of interpreting diffusion models in continuous time [51], the sampling process can be executed by solving reverse SDEs or ODEs, given that the time-dependent distribution is equivalent. This enables a broad spectrum of research to investigate better numerical solvers for the reverse differential equation [22,24,33,60]. Consequently, the total number of required steps (and therefore, NFEs) can be significantly reduced, often to just 10-20. Our contribution, while complementary to the efforts in optimizing numerical solvers, introduces an additional layer of efficiency by exploiting the concept of caching and retrieving intermediate noises. This approach is orthogonal to the direct optimization of solvers because it does not alter the computational process of the samplers. Instead, it leverages the pre-generated data to bypass certain steps of the diffusion process. By caching intermediate noises we can quickly retrieve and resume the generation process from these points for new images. This method is compatible with any recent sampler, providing a versatile and powerful tool for acceleration." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 462, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 628 + ], + "type": "text", + "content": "Retrieval-based Diffusion. Current retrieval-based T2I studies typically target generating images of rare entities [6] or adapting to new scenarios [45]. In the training-free acceleration direction, the closest work to ours is REDI [58], where the authors introduced a training-free method for image generation that leverages initial noise to retrieve the most similar generation trajectory. Our approach differs fundamentally from REDI by concurrently retrieving noises from multiple semantically similar concept prompts without the need for initial diffusion steps to obtain the noise vector, thus avoiding additional diffusion steps. Moreover, our study, using the enhanced SDXL model [38] alongside multiple datasets featuring complex prompts, demonstrates a surprising finding: simple text-based retrieval (Baseline in Fig. 2) outperform REDI when tested across extended models and datasets. Nonetheless, both REDI and the baseline approach fall short in fully capturing the intricacies of the input prompts, prompting the exploration of a more effective concept-based retrieval method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Concept Retrieval for T2I. Several studies have explored synthesizing images with user-provided concepts. These concepts can be either textual concepts [29] or image concepts [8,25,42] to generate personalized or compositionally complex" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 230, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 230, + 91, + 447, + 102 + ], + "type": "text", + "content": "RECON: Concept Retrieval for Diffusion Acceleration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 154, + 114, + 465, + 251 + ], + "blocks": [ + { + "bbox": [ + 154, + 114, + 465, + 251 + ], + "lines": [ + { + "bbox": [ + 154, + 114, + 465, + 251 + ], + "spans": [ + { + "bbox": [ + 154, + 114, + 465, + 251 + ], + "type": "image", + "image_path": "4a97beaf5867e4097e501d17e2cbc741cf03e223db4c57fcf994aafbe5c39533.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 258, + 482, + 324 + ], + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 324 + ], + "type": "text", + "content": "Fig. 3: Overview of RECON. For a given input prompt, RECON decomposes it into distinct visual concepts. Our framework then stores unique concepts from known prompts as well as their trajectories for retrieval purpose. During inference, the same concept decomposition happens and the most similar concepts along with their linked noises are retrieved. These concept noise vectors are then aggregated at step " + }, + { + "bbox": [ + 130, + 258, + 482, + 324 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 258, + 482, + 324 + ], + "type": "text", + "content": " to resume the denoising process from " + }, + { + "bbox": [ + 130, + 258, + 482, + 324 + ], + "type": "inline_equation", + "content": "t = k" + }, + { + "bbox": [ + 130, + 258, + 482, + 324 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 258, + 482, + 324 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 258, + 482, + 324 + ], + "type": "text", + "content": ", resulting in decreasing " + }, + { + "bbox": [ + 130, + 258, + 482, + 324 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 258, + 482, + 324 + ], + "type": "text", + "content": " NFEs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 348, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 482, + 529 + ], + "type": "text", + "content": "images. This exploration includes methods for customizing images with specific concept images [8, 25, 42] and composing multiple text-conditioned diffusion models for intricate image compositions [29]. Our work differs in two significant ways. First, we automatically extract visual concepts from text prompts instead of relying on user-provided concepts. Second, we cache and retrieve these automatically derived concepts alongside their corresponding noise vectors for accelerating DM. To achieve automatic extraction of visual concepts, we mainly draw inspiration from existing NLP research on word and phrase abstractness and imageability [5, 19, 32, 47]. More recently, Verma et al. [52] proposed a method for evaluating prompt-level visualness in T2I, which is more relevant for our case. Inspired by this, we accomplish automatic extraction of visual concept prompts by adopting a fine-tuned CLIP [39] to gauge the text-visualness [52]. Using this, we mine around 1.2M unique concept texts, which we later use for retrieval-based acceleration. To the best of our knowledge, this is the first attempt to extract visual concepts specifically for accelerating T2I tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 547, + 212, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 212, + 561 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 212, + 561 + ], + "type": "text", + "content": "3 Approach" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 573, + 228, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 573, + 228, + 583 + ], + "spans": [ + { + "bbox": [ + 132, + 573, + 228, + 583 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": "Diffusion Models. We approach diffusion models (DMs) through the lens of differential equations [22, 51], focusing on the \"trajectory\" idea within the diffusion process. Let " + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\sim p_{data}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": " represent the data distribution. DMs perturb " + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": " by iteratively adding i.i.d. Gaussian noise with variance " + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\sigma_t^2" + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "inline_equation", + "content": "t \\in [T,0]" + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": ", resulting in a noise distribution " + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0; \\sigma_0)" + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\sigma_0^2" + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": " reaches a sufficiently large value " + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\sigma_{max}^2" + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": ", the distribution " + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0; \\sigma_0^2)" + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": " closely resembles isotropic Gaussian noise with" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "C. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": "variance " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\sigma_{max}^{2}" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": ". Image generation in DMs starts with a noise-saturated sample " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": " and progressively denoises to approximate " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_t; \\sigma_t)" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": " ranges from 0 to " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": ". In this progression, " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\sigma_{t+1}" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": " is smaller than " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\sigma_t" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\sigma_0" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": " is equal to " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\sigma_{max}" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": ". The final output at " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\sigma_T = 0" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": ", should ideally reflect the original data distribution. This iterative denoising procedure is computationally represented by the Probability Flow ODE [51], which tracks the data sample trajectory over time:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 233, + 196, + 481, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 196, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 233, + 196, + 481, + 209 + ], + "type": "interline_equation", + "content": "d \\mathbf {x} = - \\bar {\\sigma} (t) \\sigma (t) \\nabla_ {\\mathbf {x}} \\log p (\\mathbf {x}; \\sigma (t)) d t, \\tag {1}", + "image_path": "9fd8f096051b671f263fd3a90a68442fd1c8e922394d27b5042d437b66220799.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 213, + 482, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 310 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 213, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}}\\log p(\\mathbf{x};\\sigma)" + }, + { + "bbox": [ + 130, + 213, + 482, + 310 + ], + "type": "text", + "content": " denotes the score function, and " + }, + { + "bbox": [ + 130, + 213, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\sigma (t)" + }, + { + "bbox": [ + 130, + 213, + 482, + 310 + ], + "type": "text", + "content": " is a user-defined noiselevel schedule with its time derivative represented by " + }, + { + "bbox": [ + 130, + 213, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\bar{\\sigma} (t)" + }, + { + "bbox": [ + 130, + 213, + 482, + 310 + ], + "type": "text", + "content": ". In practice, the score function is estimated by training a deep neural network " + }, + { + "bbox": [ + 130, + 213, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{\\theta}(\\mathbf{x};\\sigma)" + }, + { + "bbox": [ + 130, + 213, + 482, + 310 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 130, + 213, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 213, + 482, + 310 + ], + "type": "text", + "content": ". Note that we are using the same notation as Karras et al. [22], where " + }, + { + "bbox": [ + 130, + 213, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 130, + 213, + 482, + 310 + ], + "type": "text", + "content": " is the noisy input and " + }, + { + "bbox": [ + 130, + 213, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 130, + 213, + 482, + 310 + ], + "type": "text", + "content": " is the fully denoised output or generated final image. Throughout this paper, we refer to the trajectory " + }, + { + "bbox": [ + 130, + 213, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_t:t\\in \\{0\\dots T\\}\\}" + }, + { + "bbox": [ + 130, + 213, + 482, + 310 + ], + "type": "text", + "content": " as the one generated by a conditioned DM (typically using classifier-free guidance [18] for T2I), where " + }, + { + "bbox": [ + 130, + 213, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 130, + 213, + 482, + 310 + ], + "type": "text", + "content": " represents the noisy model outputs (noises) at each time step." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 326, + 306, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 326, + 306, + 337 + ], + "spans": [ + { + "bbox": [ + 131, + 326, + 306, + 337 + ], + "type": "text", + "content": "3.2 Retrieval-based Acceleration" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 343, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 437 + ], + "type": "text", + "content": "From Equation (1), we can see that the sampling process becomes deterministic once " + }, + { + "bbox": [ + 130, + 343, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 130, + 343, + 482, + 437 + ], + "type": "text", + "content": " is sampled from the Gaussian distribution. This allows for the pre-generation and storage of trajectories corresponding to a diverse range of prompts within a database. During inference, the key challenge lies in retrieving the trajectory that most closely aligns with the trajectory of the input prompt, posing the critical question: How do we select the optimal trajectory to resume the denoising process? To this end, we explore two primary approaches from literature, which we describe below." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 438, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 545 + ], + "type": "text", + "content": "Text-based Trajectory Retrieval. Since the input prompt is the initial element, it is intuitive to retrieve trajectories associated with the closest prompts. Specifically, for an input prompt " + }, + { + "bbox": [ + 130, + 438, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 438, + 482, + 545 + ], + "type": "text", + "content": ", we encode it with a pre-trained text encoder to obtain its embedding " + }, + { + "bbox": [ + 130, + 438, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\phi(\\mathcal{P})" + }, + { + "bbox": [ + 130, + 438, + 482, + 545 + ], + "type": "text", + "content": ". The nearest prompt " + }, + { + "bbox": [ + 130, + 438, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{P}'" + }, + { + "bbox": [ + 130, + 438, + 482, + 545 + ], + "type": "text", + "content": " is identified by finding the minimum cosine distance between text embeddings, and its trajectory " + }, + { + "bbox": [ + 130, + 438, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_t^{P'}\\}_{t=0}^T" + }, + { + "bbox": [ + 130, + 438, + 482, + 545 + ], + "type": "text", + "content": " is obtained. The intermediate output of " + }, + { + "bbox": [ + 130, + 438, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{P}'" + }, + { + "bbox": [ + 130, + 438, + 482, + 545 + ], + "type": "text", + "content": " at step " + }, + { + "bbox": [ + 130, + 438, + 482, + 545 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 438, + 482, + 545 + ], + "type": "text", + "content": " is looked up in the database and then, the diffusion process is resumed from the intermediate step " + }, + { + "bbox": [ + 130, + 438, + 482, + 545 + ], + "type": "inline_equation", + "content": "k \\in [0,T]" + }, + { + "bbox": [ + 130, + 438, + 482, + 545 + ], + "type": "text", + "content": " with the original prompt " + }, + { + "bbox": [ + 130, + 438, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 438, + 482, + 545 + ], + "type": "text", + "content": ", effectively skipping " + }, + { + "bbox": [ + 130, + 438, + 482, + 545 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 438, + 482, + 545 + ], + "type": "text", + "content": " NFEs, as illustrated in Figure 2. We refer to this method as Baseline." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 631 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 631 + ], + "type": "text", + "content": "Noise-based Trajectory Retrieval. The alternative method, REDI, retrieves trajectories by searching for the closest noise vector [58]. Specifically, for a given input prompt " + }, + { + "bbox": [ + 130, + 545, + 482, + 631 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 545, + 482, + 631 + ], + "type": "text", + "content": ", REDI first computes " + }, + { + "bbox": [ + 130, + 545, + 482, + 631 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 545, + 482, + 631 + ], + "type": "text", + "content": " steps in the denoising process and uses the noise " + }, + { + "bbox": [ + 130, + 545, + 482, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_h^{\\mathcal{P}}" + }, + { + "bbox": [ + 130, + 545, + 482, + 631 + ], + "type": "text", + "content": " as a query. It then searches for the nearest noise " + }, + { + "bbox": [ + 130, + 545, + 482, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_h^{P'}" + }, + { + "bbox": [ + 130, + 545, + 482, + 631 + ], + "type": "text", + "content": " in the knowledge base and resumes the denoising process from step " + }, + { + "bbox": [ + 130, + 545, + 482, + 631 + ], + "type": "inline_equation", + "content": "k > h" + }, + { + "bbox": [ + 130, + 545, + 482, + 631 + ], + "type": "text", + "content": ", effectively skipping " + }, + { + "bbox": [ + 130, + 545, + 482, + 631 + ], + "type": "inline_equation", + "content": "k - h" + }, + { + "bbox": [ + 130, + 545, + 482, + 631 + ], + "type": "text", + "content": " steps. However, this method adds " + }, + { + "bbox": [ + 130, + 545, + 482, + 631 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 545, + 482, + 631 + ], + "type": "text", + "content": " initial steps to the number of NFEs, as illustrated in Figure 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 645, + 338, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 645, + 338, + 658 + ], + "spans": [ + { + "bbox": [ + 131, + 645, + 338, + 658 + ], + "type": "text", + "content": "3.3 Our Proposed Framework: RECON" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "type": "text", + "content": "RECON: Concept Retrieval for Diffusion Acceleration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "Both Baseline and REDI rely on stored prompts " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "\\mathcal{P}'" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": ", either directly through text similarity or indirectly through noise similarity. However, they face challenges in generating images that are faithful to the input prompt " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": ". As shown in Figure 2, both the baseline method and REDI introduce extraneous components into the final images, compromising its fidelity. Crucial image attributes, such as layout and color schemes, are defined in the initial denoising stages [15], which are frequently bypassed in retrieval-based approaches, leading to a misalignment from the intended prompt details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 213, + 482, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 500 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 500 + ], + "type": "text", + "content": "Concept Extraction and KB Creation. To address the challenge of incorporating specific details from input prompts while avoiding the rigidity seen in Baseline and REDI, we introduce a novel approach: decomposing input prompts into visual concepts. However, the challenge lies in distinguishing visual concepts from abstract terms. For example, terms like \"belief,\" \"essentialness,\" and \"spirituality,\" while conceptually significant, often rank high in abstractness [5], but do not translate into visual representations for T2I generation. This distinction is crucial as language models like RoBERTa [9, 31] are proficient in extracting and understanding text, including abstract concepts, the challenge lies in ensuring that these concepts can be visually represented. To tackle this challenge, we adopt a fine-tuned CLIP model to evaluate the visual potential of extracted phrases, ensuring that only those with a clear visual representation are selected [52]. In the process of RECON, each text prompt " + }, + { + "bbox": [ + 130, + 213, + 482, + 500 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 213, + 482, + 500 + ], + "type": "text", + "content": " is decomposed into " + }, + { + "bbox": [ + 130, + 213, + 482, + 500 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 213, + 482, + 500 + ], + "type": "text", + "content": " visual concepts " + }, + { + "bbox": [ + 130, + 213, + 482, + 500 + ], + "type": "inline_equation", + "content": "\\{c_i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 213, + 482, + 500 + ], + "type": "text", + "content": ", using a pre-trained RoBERTa [31]. Subsequently, a visualness filtering model " + }, + { + "bbox": [ + 130, + 213, + 482, + 500 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 130, + 213, + 482, + 500 + ], + "type": "text", + "content": ", fine-tuned to distinguish between visually representable and non-visual text segments, evaluates these extracted noun phrases. Representations. The visualness of a concept " + }, + { + "bbox": [ + 130, + 213, + 482, + 500 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 213, + 482, + 500 + ], + "type": "text", + "content": " is quantified through cosine similarity " + }, + { + "bbox": [ + 130, + 213, + 482, + 500 + ], + "type": "inline_equation", + "content": "S = 1 - \\langle h(c), h(\\mathcal{I}_{null}) \\rangle" + }, + { + "bbox": [ + 130, + 213, + 482, + 500 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 213, + 482, + 500 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{null}" + }, + { + "bbox": [ + 130, + 213, + 482, + 500 + ], + "type": "text", + "content": " denotes an image of random RGB values, serving as a sample for non-visual content. Concepts with visualness scores " + }, + { + "bbox": [ + 130, + 213, + 482, + 500 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 213, + 482, + 500 + ], + "type": "text", + "content": " below a predefined threshold are deemed non-visual and excluded from further consideration. These visually relevant concepts and their trajectories from DMs are then cached in a vector database for retrieval purpose. Figure 3 illustrates these steps in this stage of RECON under the section \"Concept KB and Trajectories Creation.\"" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 502, + 483, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 483, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 483, + 604 + ], + "type": "text", + "content": "Retrieving and Aggregating Concepts. During inference, given an input prompt " + }, + { + "bbox": [ + 130, + 502, + 483, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 502, + 483, + 604 + ], + "type": "text", + "content": ", a concept knowledge base, and a pre-trained diffusion model " + }, + { + "bbox": [ + 130, + 502, + 483, + 604 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 130, + 502, + 483, + 604 + ], + "type": "text", + "content": ", we first decompose the prompt into a set of visual concepts " + }, + { + "bbox": [ + 130, + 502, + 483, + 604 + ], + "type": "inline_equation", + "content": "\\{c_i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 502, + 483, + 604 + ], + "type": "text", + "content": ". For each concept " + }, + { + "bbox": [ + 130, + 502, + 483, + 604 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 130, + 502, + 483, + 604 + ], + "type": "text", + "content": ", we compute its text embedding " + }, + { + "bbox": [ + 130, + 502, + 483, + 604 + ], + "type": "inline_equation", + "content": "\\phi(c_i)" + }, + { + "bbox": [ + 130, + 502, + 483, + 604 + ], + "type": "text", + "content": ", which serves as a query for retrieving the closest matching concepts " + }, + { + "bbox": [ + 130, + 502, + 483, + 604 + ], + "type": "inline_equation", + "content": "c_i'" + }, + { + "bbox": [ + 130, + 502, + 483, + 604 + ], + "type": "text", + "content": " along with their associated trajectories " + }, + { + "bbox": [ + 130, + 502, + 483, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t^{c_i'}" + }, + { + "bbox": [ + 130, + 502, + 483, + 604 + ], + "type": "text", + "content": ". Subsequently, at a pre-determined step " + }, + { + "bbox": [ + 130, + 502, + 483, + 604 + ], + "type": "inline_equation", + "content": "k \\in [0,T]" + }, + { + "bbox": [ + 130, + 502, + 483, + 604 + ], + "type": "text", + "content": ", we aggregate these trajectories by combining their noise vectors " + }, + { + "bbox": [ + 130, + 502, + 483, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_k^{c_i'}" + }, + { + "bbox": [ + 130, + 502, + 483, + 604 + ], + "type": "text", + "content": " at step " + }, + { + "bbox": [ + 130, + 502, + 483, + 604 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 502, + 483, + 604 + ], + "type": "text", + "content": ", thus forming a new starting point for the denoising process:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 270, + 615, + 481, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 615, + 481, + 647 + ], + "spans": [ + { + "bbox": [ + 270, + 615, + 481, + 647 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} _ {t} = \\sum_ {i = 1} ^ {N} w _ {i} \\cdot \\mathbf {x} _ {k} ^ {c _ {i} ^ {\\prime}}, \\tag {2}", + "image_path": "8e9aa92e52cde68098f53d064aefc8384d528c2ff1bd01e00985dd6df39df2b2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 228, + 651, + 481, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 651, + 481, + 663 + ], + "spans": [ + { + "bbox": [ + 228, + 651, + 481, + 663 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {t + 1} = f _ {\\theta} \\left(\\hat {\\mathbf {x}} _ {t}; t, \\mathcal {P}\\right); t = k, \\dots , T - 1, \\tag {3}", + "image_path": "6276ad3f326267e8aafa5f9c12f3f2f6ed9b396ed19fe290b1959b90c65ee7b5.jpg" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "C. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 120, + 463, + 279 + ], + "blocks": [ + { + "bbox": [ + 149, + 120, + 463, + 279 + ], + "lines": [ + { + "bbox": [ + 149, + 120, + 463, + 279 + ], + "spans": [ + { + "bbox": [ + 149, + 120, + 463, + 279 + ], + "type": "image", + "image_path": "376c95a8294d684546c0c281399ee70dd8393733f6b89b29b3c949354e7ac95d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 288, + 483, + 334 + ], + "lines": [ + { + "bbox": [ + 130, + 288, + 483, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 288, + 483, + 334 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparisons reveal that RECON consistently produces images of quality comparable to those generated by Vanilla, with no retrieval-based acceleration. Here, the numbers of NFEs are: Vanilla=50, Others=35. Full prompts for DiffusionDB appear in Supplementary." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 357, + 483, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 483, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 483, + 559 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 357, + 483, + 559 + ], + "type": "inline_equation", + "content": "\\{w_{i}:i\\in [1,N]\\}" + }, + { + "bbox": [ + 130, + 357, + 483, + 559 + ], + "type": "text", + "content": " corresponds to the weights for each concept. Our method defines weights " + }, + { + "bbox": [ + 130, + 357, + 483, + 559 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 130, + 357, + 483, + 559 + ], + "type": "text", + "content": " using the cosine similarity between text embeddings of the query and retrieved concepts, normalized across all concepts: " + }, + { + "bbox": [ + 130, + 357, + 483, + 559 + ], + "type": "inline_equation", + "content": "w_{i} = \\frac{\\left\\langle\\phi(c_{i}),\\phi(c_{i}^{\\prime})\\right\\rangle}{\\sum_{i = 1}^{N}\\left\\langle\\phi(c_{i}),\\phi(c_{i}^{\\prime})\\right\\rangle}" + }, + { + "bbox": [ + 130, + 357, + 483, + 559 + ], + "type": "text", + "content": ". This approach is grounded in the principle that diffusion models can be regarded as energy-based models (EBMs), facilitating the compositional aggregation of noise vectors [12,28,29] For detailed insights into noise vector compositionality in EBMs, please refer to our Supplementary material. The \"Concept Retrieval\" and \"Concept Noises Aggregation\" stages precede the continuation of the denoising process with the aggregated noise " + }, + { + "bbox": [ + 130, + 357, + 483, + 559 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_t" + }, + { + "bbox": [ + 130, + 357, + 483, + 559 + ], + "type": "text", + "content": " and the original prompt " + }, + { + "bbox": [ + 130, + 357, + 483, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 357, + 483, + 559 + ], + "type": "text", + "content": " to generate the final image " + }, + { + "bbox": [ + 130, + 357, + 483, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 130, + 357, + 483, + 559 + ], + "type": "text", + "content": ", as depicted in Figure 3. While only the concept noun phrases are aggregated, the final diffusion steps (steps " + }, + { + "bbox": [ + 130, + 357, + 483, + 559 + ], + "type": "inline_equation", + "content": "k\\dots T" + }, + { + "bbox": [ + 130, + 357, + 483, + 559 + ], + "type": "text", + "content": " in Fig. 3) are guided by the original prompt. Thus, essential prompt details beyond nouns still get to influence the resulting image " + }, + { + "bbox": [ + 130, + 357, + 483, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 130, + 357, + 483, + 559 + ], + "type": "text", + "content": ". For instance, in col. 3 of Fig. 4, although the concepts extracted from the original prompt \"A herd of sheep standing below very tall buildings\" are \"a herd of sheep\" and \"buildings\" without including \"tall,\" the final image still captures this key attribute." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 578, + 230, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 230, + 593 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 230, + 593 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 606, + 190, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 606, + 190, + 619 + ], + "spans": [ + { + "bbox": [ + 132, + 606, + 190, + 619 + ], + "type": "text", + "content": "4.1 Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Datasets. For a fair comparison, we first assess the performance of RECON using the same dataset as REDI [58]: MS-COCO [7] (containing 82K and 42K captions in train and validation). However, captions in MS-COCO may not always" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "type": "text", + "content": "RECON: Concept Retrieval for Diffusion Acceleration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 186, + 121, + 274, + 152 + ], + "blocks": [ + { + "bbox": [ + 186, + 114, + 210, + 121 + ], + "lines": [ + { + "bbox": [ + 186, + 114, + 210, + 121 + ], + "spans": [ + { + "bbox": [ + 186, + 114, + 210, + 121 + ], + "type": "text", + "content": "Baseline" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 186, + 121, + 274, + 152 + ], + "lines": [ + { + "bbox": [ + 186, + 121, + 274, + 152 + ], + "spans": [ + { + "bbox": [ + 186, + 121, + 274, + 152 + ], + "type": "image", + "image_path": "fb70891a5c8405137ee032a2b34c8ea56587fb5d2f759d4ebe8455331a0c7815.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 152, + 238, + 156 + ], + "lines": [ + { + "bbox": [ + 216, + 152, + 238, + 156 + ], + "spans": [ + { + "bbox": [ + 216, + 152, + 238, + 156 + ], + "type": "text", + "content": "NFE=40" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 216, + 188, + 238, + 194 + ], + "lines": [ + { + "bbox": [ + 216, + 188, + 238, + 194 + ], + "spans": [ + { + "bbox": [ + 216, + 188, + 238, + 194 + ], + "type": "text", + "content": "NFE = 20" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 276, + 130, + 304, + 168 + ], + "blocks": [ + { + "bbox": [ + 276, + 130, + 304, + 168 + ], + "lines": [ + { + "bbox": [ + 276, + 130, + 304, + 168 + ], + "spans": [ + { + "bbox": [ + 276, + 130, + 304, + 168 + ], + "type": "image", + "image_path": "c3715e8be9f343deebf607791b53de4232c1e18a675465d7e0ac800b8f819034.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 278, + 169, + 299, + 175 + ], + "lines": [ + { + "bbox": [ + 278, + 169, + 299, + 175 + ], + "spans": [ + { + "bbox": [ + 278, + 169, + 299, + 175 + ], + "type": "text", + "content": "NFE=50" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 121, + 399, + 152 + ], + "blocks": [ + { + "bbox": [ + 314, + 115, + 337, + 121 + ], + "lines": [ + { + "bbox": [ + 314, + 115, + 337, + 121 + ], + "spans": [ + { + "bbox": [ + 314, + 115, + 337, + 121 + ], + "type": "text", + "content": "Baseline" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 121, + 399, + 152 + ], + "lines": [ + { + "bbox": [ + 310, + 121, + 399, + 152 + ], + "spans": [ + { + "bbox": [ + 310, + 121, + 399, + 152 + ], + "type": "image", + "image_path": "f3c5f1dc9269a48f61d3d343cd10608cdee04201869d7655c081407a95e96baa.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 152, + 362, + 157 + ], + "lines": [ + { + "bbox": [ + 342, + 152, + 362, + 157 + ], + "spans": [ + { + "bbox": [ + 342, + 152, + 362, + 157 + ], + "type": "text", + "content": "NFE = 40" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 376, + 115, + 396, + 121 + ], + "lines": [ + { + "bbox": [ + 376, + 115, + 396, + 121 + ], + "spans": [ + { + "bbox": [ + 376, + 115, + 396, + 121 + ], + "type": "text", + "content": "RECON" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 157, + 399, + 188 + ], + "blocks": [ + { + "bbox": [ + 191, + 194, + 296, + 201 + ], + "lines": [ + { + "bbox": [ + 191, + 194, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 191, + 194, + 296, + 201 + ], + "type": "text", + "content": "A dog warrior holding a sword in pink armor." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 157, + 399, + 188 + ], + "lines": [ + { + "bbox": [ + 309, + 157, + 399, + 188 + ], + "spans": [ + { + "bbox": [ + 309, + 157, + 399, + 188 + ], + "type": "image", + "image_path": "bfaeba49e76106a84bc9703b9d5c6370e9184348edcb50dea87559f74df3be1d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 189, + 362, + 194 + ], + "lines": [ + { + "bbox": [ + 342, + 189, + 362, + 194 + ], + "spans": [ + { + "bbox": [ + 342, + 189, + 362, + 194 + ], + "type": "text", + "content": "NFE = 20" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 194, + 428, + 201 + ], + "lines": [ + { + "bbox": [ + 313, + 194, + 428, + 201 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 428, + 201 + ], + "type": "text", + "content": "A cat is sitting on a wooden café table with décor." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 211, + 480, + 232 + ], + "lines": [ + { + "bbox": [ + 131, + 211, + 480, + 232 + ], + "spans": [ + { + "bbox": [ + 131, + 211, + 480, + 232 + ], + "type": "text", + "content": "Fig. 5: While other methods struggle to generate faithful images with relatively higher NFEs, ours (RECON) can generate faithful images at NFEs as low as 20." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 400, + 136, + 431, + 167 + ], + "blocks": [ + { + "bbox": [ + 402, + 130, + 423, + 136 + ], + "lines": [ + { + "bbox": [ + 402, + 130, + 423, + 136 + ], + "spans": [ + { + "bbox": [ + 402, + 130, + 423, + 136 + ], + "type": "text", + "content": "Vanilla" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 400, + 136, + 431, + 167 + ], + "lines": [ + { + "bbox": [ + 400, + 136, + 431, + 167 + ], + "spans": [ + { + "bbox": [ + 400, + 136, + 431, + 167 + ], + "type": "image", + "image_path": "170fe417b8b96c6af0ade37d97f9a01d911f5c6a802afa62c47926861f3765f8.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 405, + 169, + 427, + 175 + ], + "lines": [ + { + "bbox": [ + 405, + 169, + 427, + 175 + ], + "spans": [ + { + "bbox": [ + 405, + 169, + 427, + 175 + ], + "type": "text", + "content": "NFE=50" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 258, + 480, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 480, + 341 + ], + "type": "text", + "content": "reflect real-world T2I prompts, as noted in [23]. Therefore, we also use Pick-a-Pic [23] (35K unique prompts) and DiffusionDB [55] (1.8M unique prompts). These datasets offer a broader range of prompts, more reflective of actual T2I user preferences. For evaluation, we randomly select 10K captions from MS-COCO, 500 prompts from Pick-a-Pic, and 10K from DiffusionDB from their validation/test set (detailed in Supplementary). To our knowledge, such a comprehensive study has not been done before for retrieval-based acceleration." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 342, + 480, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 480, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 480, + 484 + ], + "type": "text", + "content": "Metrics. In unconditional settings, a popular metric to test photorealism is FID [16], which compares feature distributions from real and generated images. However, FID falls short in conditional settings such as T2I. Studies have pointed out FID is negatively correlated with human preferences [23, 38]. Therefore, for a comprehensive evaluation of T2I models, we employ metrics that specifically address the conditional nature of the task. PickScore [23] measures the alignment between text descriptions and generated images. \"CLIP Score\" [39], on the other hand, provides a broader measure of semantic congruence between the text and generated image, useful for understanding how well the generated image captures the gist of the text. \"Aesthetics Score\" [1] complements these by assessing the visual appeal of images. Finally, we employed NFE to quantify RECON's time efficiency, independent of hardware configurations." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "text", + "content": "Implementation Details. For building Baseline and REDI (described in Section 3.2), we follow the steps mentioned in [58], but replace the base model with the improved SDXL [38]. We denote the default sampling process in SDXL without any acceleration as \"Vanilla Generation.\" For RECON, we first build a concept knowledge base (KB) from the unique captions and prompts in the training sets of all considered datasets. We extract noun phrases using a pretrained language model, RoBERTa [31] and apply visual filtering with a CLIP model [39] from [52]. To efficiently store and retrieve these extracted concepts and their associated trajectories, we implement a vector database with FAISS [20]. Remarkably, the concept retrieval part of our pipeline merely takes " + }, + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "inline_equation", + "content": "1.15\\%" + }, + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "text", + "content": " of the total inference time on an A10G GPU. We build the rest of our pipeline based on PyTorch and the HuggingFace Diffusers library [36, 37]. We determine the weights for merging concept noise vectors using the CLIP text similarity scores, as in Equation (2). Unless otherwise mentioned, for all the experiments, we use the Euler sampler [22] as default, setting the number of NFEs to 35 for retrieval-" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "C. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 179, + 116, + 309, + 247 + ], + "blocks": [ + { + "bbox": [ + 179, + 116, + 309, + 247 + ], + "lines": [ + { + "bbox": [ + 179, + 116, + 309, + 247 + ], + "spans": [ + { + "bbox": [ + 179, + 116, + 309, + 247 + ], + "type": "image", + "image_path": "a07421fbf178f5ca0f43e06e01e0885401bafb3866bb03ae2d5314870f86e4e5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 248, + 302, + 263 + ], + "lines": [ + { + "bbox": [ + 194, + 248, + 302, + 263 + ], + "spans": [ + { + "bbox": [ + 194, + 248, + 302, + 263 + ], + "type": "text", + "content": "\"Two brown teddy bears sitting on top a wooden table.\"" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 310, + 116, + 434, + 247 + ], + "blocks": [ + { + "bbox": [ + 310, + 116, + 434, + 247 + ], + "lines": [ + { + "bbox": [ + 310, + 116, + 434, + 247 + ], + "spans": [ + { + "bbox": [ + 310, + 116, + 434, + 247 + ], + "type": "image", + "image_path": "eb2485acb6bc635e380b7156f8f27baa621e67ade6ea23090c24cf12a743ff25.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 248, + 431, + 273 + ], + "lines": [ + { + "bbox": [ + 312, + 248, + 431, + 273 + ], + "spans": [ + { + "bbox": [ + 312, + 248, + 431, + 273 + ], + "type": "text", + "content": "\"renaissance painting of elder in the garden, short silver hair, a wise face, emotions closeup, dressed in roman armour.\"" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 285, + 482, + 319 + ], + "lines": [ + { + "bbox": [ + 131, + 285, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 131, + 285, + 482, + 319 + ], + "type": "text", + "content": "Fig. 6: Baseline and REDI often struggle with 1st Nearest Neighbor (NN) results, leading to poorer outcomes for 2nd and 3rd NNs. Conversely, RECon consistently produces accurate images across all NN concepts, addressing the diversity challenge." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 342, + 482, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 342, + 482, + 378 + ], + "spans": [ + { + "bbox": [ + 131, + 342, + 482, + 378 + ], + "type": "text", + "content": "based methods and 50 for Vanilla Generation. Further implementation details, including configurations of SDXL, the KB size, specifics of our vector database for caching and retrieval are provided in Supplementary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 395, + 258, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 395, + 258, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 395, + 258, + 407 + ], + "type": "text", + "content": "4.2 Qualitative Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 414, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 414, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 414, + 482, + 665 + ], + "type": "text", + "content": "Enhanced Fidelity. In Figure 4, Images generated using RECON demonstrate a higher degree of fidelity to the provided prompts compared to both Baseline and REDI methods. RECON images also closely mirror the results of Vanilla Generation across all datasets. In contrast, images from Baseline and REDI images exhibit notable discrepancies from the input prompt, such as depicting inaccurate objects (animal mouse vs. computer mouse), or omitting key elements described in the prompts (e.g., \"very tall buildings\"). RECON is able to generate uncommon concept combinations like \".cat eating a burger like a person.\", \".raccoon washing blankets.\", and \".cat dressed as a scuba diver..\" . REDI and Baseline often miss the mark with these intricate descriptions, occasionally generating completely irrelevant images, like those for the prompt \"Human abduction by UFO.\" RECON also performs better with rare input prompts. For instance, the phrase \"sparkling diamond gold butterfly\" is absent from KB. Nevertheless, RECON produced a relevant image of a butterfly based on the concept \"a sparkling gold glitter dress,\" thus highlighting the advantages of the flexible trajectory of RECON's design. Figure 5 showcases the outputs from three retrieval-based methods with both high and low number of NFEs. Both Baseline and REDI methods struggle to faithfully capture the input prompts in their images at relatively high number of NFEs. Unsurprisingly, their performance further diminishes at lower number of NFEs. RECON, however, consistently performs better even with NFEs as low as 20." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "type": "text", + "content": "RECON: Concept Retrieval for Diffusion Acceleration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 171, + 114, + 244, + 175 + ], + "blocks": [ + { + "bbox": [ + 171, + 114, + 244, + 175 + ], + "lines": [ + { + "bbox": [ + 171, + 114, + 244, + 175 + ], + "spans": [ + { + "bbox": [ + 171, + 114, + 244, + 175 + ], + "type": "image", + "image_path": "aca80fa86833f53d0925442495975b44509e468647b58537d4b77f93f9ff56cc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 175, + 205, + 234 + ], + "blocks": [ + { + "bbox": [ + 133, + 175, + 205, + 234 + ], + "lines": [ + { + "bbox": [ + 133, + 175, + 205, + 234 + ], + "spans": [ + { + "bbox": [ + 133, + 175, + 205, + 234 + ], + "type": "image", + "image_path": "eed3dca3f7b5d18f67cf96de738faf0756b752e4df669a1aad160b6d8b1fea7e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 244, + 301, + 300 + ], + "lines": [ + { + "bbox": [ + 131, + 244, + 301, + 300 + ], + "spans": [ + { + "bbox": [ + 131, + 244, + 301, + 300 + ], + "type": "text", + "content": "Fig.7: RECON can generate images with high CLIP scores in 6-7 seconds using an A10G GPU, whereas both REDI and Baseline require more than 10 seconds to generate images of comparable quality." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 208, + 175, + 280, + 234 + ], + "blocks": [ + { + "bbox": [ + 208, + 175, + 280, + 234 + ], + "lines": [ + { + "bbox": [ + 208, + 175, + 280, + 234 + ], + "spans": [ + { + "bbox": [ + 208, + 175, + 280, + 234 + ], + "type": "image", + "image_path": "2ab8801be70883fcebcf9ead8cd58d5ea47a77b10a60d4ca8e74c74211e85117.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 313, + 114, + 388, + 175 + ], + "blocks": [ + { + "bbox": [ + 313, + 114, + 388, + 175 + ], + "lines": [ + { + "bbox": [ + 313, + 114, + 388, + 175 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 388, + 175 + ], + "type": "image", + "image_path": "90bf47d605ad220aed51e9779ed3416af8b2dc50f4cc250d8b26bffcfd27ab7b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 389, + 114, + 463, + 175 + ], + "blocks": [ + { + "bbox": [ + 389, + 114, + 463, + 175 + ], + "lines": [ + { + "bbox": [ + 389, + 114, + 463, + 175 + ], + "spans": [ + { + "bbox": [ + 389, + 114, + 463, + 175 + ], + "type": "image", + "image_path": "ad8537c941fbd203567e7688a3251e16a5123364b292627908e9e1254df554af.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 313, + 175, + 388, + 236 + ], + "blocks": [ + { + "bbox": [ + 313, + 175, + 388, + 236 + ], + "lines": [ + { + "bbox": [ + 313, + 175, + 388, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 388, + 236 + ], + "type": "image", + "image_path": "95189773cd9e7b733d852cb9e3adc5bab7c65e9150194d319d8dcc92b6735be2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 244, + 482, + 300 + ], + "lines": [ + { + "bbox": [ + 311, + 244, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 311, + 244, + 482, + 300 + ], + "type": "text", + "content": "Fig.8: RECON consistently outperforms Baseline and REDI on MS-COCO under varying total number of steps and different percentage of skipped steps w.r.t. total steps taken by Vanilla Generation." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 389, + 175, + 462, + 236 + ], + "blocks": [ + { + "bbox": [ + 389, + 175, + 462, + 236 + ], + "lines": [ + { + "bbox": [ + 389, + 175, + 462, + 236 + ], + "spans": [ + { + "bbox": [ + 389, + 175, + 462, + 236 + ], + "type": "image", + "image_path": "c5a05bf9dfea237f5c1e1d037d3bc21248c7dc186b5e9e6cb84db1ee6ed50d28.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 324, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 482, + 529 + ], + "type": "text", + "content": "Diversity. Diversity in T2I generation ensures that the model can produce a range of visually distinct images for a given prompt, enhancing user experience by offering creative variations. The traditional approach for diversity in T2I tasks involves caching different trajectories for multiple seeds. This method, while capable of producing diverse outputs, faces significant scalability challenges. Storing trajectories for a growing number of unique prompts, even for a single seed, quickly becomes prohibitively expensive. A cheaper solution to this problem is to pull the top-" + }, + { + "bbox": [ + 130, + 324, + 482, + 529 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 324, + 482, + 529 + ], + "type": "text", + "content": " Nearest Neighbor (NN) trajectories. However, Baseline and REDI struggle with accuracy and faithfulness, particularly beyond the first NN, as shown in Figure 6. RECON effectively overcomes this by focusing on similar visual concepts rather than full-length prompts. This approach is more efficient as matching concepts is simpler than entire prompts. For instance, in one of the prompts showcased, the top-three retrieved NN concepts are: 1st: ['two brown teddy bears', 'large wooden table'], 2nd: ['two teddy bears', 'wooden kitchen table'], and 3rd: ['two stuffed teddy bears', 'wooden counter'], all closely related to each other. This similarity among the top concepts enables RECON to maintain diversity without sacrificing fidelity or facing scalability issues." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 548, + 272, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 272, + 560 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 272, + 560 + ], + "type": "text", + "content": "4.3 Quantitative Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Comparison of All Methods. In Figure 8, we show the performance with different percentage of skipped steps with varying the number of total steps (or NFEs) on MS-COCO (results from other datasets are shown in Supplementary). RECON consistently surpasses both REDI and Baseline across different NFE counts, demonstrating more notable improvement at lower number of NFEs. In Figure 7, we demonstrate the CLIP Scores of RECON in comparison to previous works. With just 6 seconds of computation time on an A10G GPU, RECON is able to produce accurate images from prompts across various datasets, while" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "C. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 146, + 482, + 212 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 1: Our concept-based retrieval RECon surpasses Baseline and REDI across MS-COCO, Pick-a-Pic, and DiffusionDB datasets. NFE is fixed to 35." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 146, + 482, + 212 + ], + "lines": [ + { + "bbox": [ + 134, + 146, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 134, + 146, + 482, + 212 + ], + "type": "table", + "html": "
CLIP Score↑PickScore↑Aesthetics Score↑
COCOPick-a-PicDiffDBCOCOPick-a-PicDiffDBCOCOPick-a-PicDiffDB
Baseline29.1527.6729.6820.7420.4419.815.506.056.26
REDi18.1721.8020.4319.5119.0318.215.516.066.29
ReCon30.1929.9331.4621.3120.7620.085.526.106.31
", + "image_path": "41e40f761199cbeaa6582eb77e83ce5a26d2d71982dfc811a56c5344e3c6f55a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 134, + 315, + 309, + 355 + ], + "blocks": [ + { + "bbox": [ + 132, + 228, + 309, + 305 + ], + "lines": [ + { + "bbox": [ + 132, + 228, + 309, + 305 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 309, + 305 + ], + "type": "text", + "content": "Table 2: RECON's ablation study where 'Visual-filtering' leverages text-visualness scores [53], while 'weighting' evaluates the similarity between concepts. 'Top-1' prioritizes the highest scoring concept, while " + }, + { + "bbox": [ + 132, + 228, + 309, + 305 + ], + "type": "inline_equation", + "content": ">0.5" + }, + { + "bbox": [ + 132, + 228, + 309, + 305 + ], + "type": "text", + "content": " criterion includes only those concepts with similarities above this value." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 315, + 309, + 355 + ], + "lines": [ + { + "bbox": [ + 134, + 315, + 309, + 355 + ], + "spans": [ + { + "bbox": [ + 134, + 315, + 309, + 355 + ], + "type": "table", + "html": "
Visual-filtering\nWeighting✓\nX (Top-1)✓\nX (>0.5)✓\nXXX✓\n✓
CLIP Score ↑27.4629.7929.3328.9829.0629.93
PickScore ↑19.8820.5920.5420.1220.3620.84
", + "image_path": "e220c865fdc2ef149456495aec036b2ab2278b5984f25f28c11c6f34376e4ecb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 322, + 315, + 480, + 354 + ], + "blocks": [ + { + "bbox": [ + 321, + 228, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 321, + 228, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 321, + 228, + 482, + 304 + ], + "type": "text", + "content": "Table 3: Comparison of all methods for different samplers. RECON consistently outperforms Baseline and REDI across different diffusion samplers, showcasing RECON's ability to work with multiple efficient numerical solvers." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 315, + 480, + 354 + ], + "lines": [ + { + "bbox": [ + 322, + 315, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 322, + 315, + 480, + 354 + ], + "type": "table", + "html": "
SchedulerCLIP Score↑PickScore↑
EulerPNDMDDIMDPMEulerPNDMDDIMDPM
Baseline27.6725.9626.7726.7320.4419.8220.1020.00
REDi21.8020.2421.6821.4219.0318.0818.4718.40
ReCON29.9327.6828.9928.5620.7619.8920.2920.06
", + "image_path": "77260c625e471d6ae09411cede35b6ed143306b259032baf84794994dbc88525.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 384, + 482, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 480 + ], + "type": "text", + "content": "both REDI and Baseline require over 10 seconds to generate images of similar quality. A detailed study among the retrieval-based methods is presented in Table 1. It is evident that RECON excels over Baseline and REDI in terms of all metrics. We also make a surprising discovery: Baseline is quite competitive compared to the state-of-the-art REDI. This shows that when using an improved base model, a simple approach like the Baseline can perform quite well, especially on T2I alignment scores. For rest of the analysis, we primarily report these two scores (results for the rest are in the Supplementary)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 483, + 482, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 482, + 543 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 482, + 543 + ], + "type": "text", + "content": "Aggregation Ablation Study. Table 2 shows the importance of using both weighted aggregation Equation (2) and visual-filtering in RECON. Here, we compare with using only the most similar concept, denoted as Top-1, and excluding low-similarity concepts with a threshold of 0.5, denoted as " + }, + { + "bbox": [ + 130, + 483, + 482, + 543 + ], + "type": "inline_equation", + "content": "(>0.5)" + }, + { + "bbox": [ + 130, + 483, + 482, + 543 + ], + "type": "text", + "content": ". The results indicate that both the weighting and visualness filter are essential." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 545, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 483, + 666 + ], + "type": "text", + "content": "Concept Retrieval Analysis. Although both Baseline and RECON rely on text similarity scores (either prompt-to-prompt in Baseline or concept-to-concept in RECON), Table 4 shows that the latter consistently outperforms the former. At each similarity score interval, concept-based retrieval in RECON leads to superior results, proving the effectiveness of our approach. The length and complexity of input prompts, often containing multiple concepts, poses the question: how many concepts to break a prompt into? Figure 9 shows that consistently deconstructing prompts into concepts (RECON) generally produces better outcomes than not doing so (Baseline). Also, it shows that there is a sweet spot for the number of concepts to decompose a text prompt into." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 230, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 230, + 91, + 447, + 102 + ], + "type": "text", + "content": "RECON: Concept Retrieval for Diffusion Acceleration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 149, + 146, + 462, + 215 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": "Table 4: RECON consistently performs better than Baseline at every text similarity score interval and every NFE count (results shown for PickScore†)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 146, + 462, + 215 + ], + "lines": [ + { + "bbox": [ + 149, + 146, + 462, + 215 + ], + "spans": [ + { + "bbox": [ + 149, + 146, + 462, + 215 + ], + "type": "table", + "html": "
TextNFE = 40NFE = 35NFE = 30NFE = 25
SimilarityBaselineReCONBaselineReCONBaselineReCONBaselineReCON
0.5-0.620.3821.0720.1120.9219.9520.5219.6220.18
0.6-0.720.5120.9820.0720.5919.8620.2019.5320.04
0.7-0.820.7720.9620.4420.6420.2520.3619.9220.16
0.8-0.921.2621.4220.9721.1620.8320.8920.4320.65
", + "image_path": "72dadcccdf076deac95baad988b175bbe833011a74059eca1b13e04f4d8af906.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 184, + 224, + 307, + 298 + ], + "blocks": [ + { + "bbox": [ + 184, + 224, + 307, + 298 + ], + "lines": [ + { + "bbox": [ + 184, + 224, + 307, + 298 + ], + "spans": [ + { + "bbox": [ + 184, + 224, + 307, + 298 + ], + "type": "image", + "image_path": "9ea05ee6f773c04042fc608ef1b551cb74c338a478c38555fdebfc24926d7e11.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 306, + 481, + 329 + ], + "lines": [ + { + "bbox": [ + 130, + 306, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 481, + 329 + ], + "type": "text", + "content": "Fig.9: Prompts can contain an arbitrary number of concepts; so, breaking down a prompt into its constituent concepts is better (RECon) than not doing so (Baseline)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 224, + 429, + 298 + ], + "blocks": [ + { + "bbox": [ + 307, + 224, + 429, + 298 + ], + "lines": [ + { + "bbox": [ + 307, + 224, + 429, + 298 + ], + "spans": [ + { + "bbox": [ + 307, + 224, + 429, + 298 + ], + "type": "image", + "image_path": "9159d252bec14b720d161954921265999e5d7602d9158c78689a570d34ba5359.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 353, + 479, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 479, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 479, + 400 + ], + "type": "text", + "content": "Comparison with Different Samplers. Table 3 showcases the results of incorporating our method into other diffusion samplers, including DDIM [48], DPM-Solver [33], PNDM [22], and Euler [22], with RECON outperforming other retrieval-based techniques across different samplers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 401, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 485 + ], + "type": "text", + "content": "Human evaluation. We conducted a comprehensive user study involving 214 participants (details provided in Supplementary) to evaluate image generation using RECON, Baseline, and REDI. After filtering for consistency, we retained 200 valid participants. All methods used the same random seed, steps (NFE = 35), and Euler sampler. Results showed RECON's superior performance, with " + }, + { + "bbox": [ + 130, + 401, + 482, + 485 + ], + "type": "inline_equation", + "content": "75.5\\%" + }, + { + "bbox": [ + 130, + 401, + 482, + 485 + ], + "type": "text", + "content": " of its generated images ranked first, " + }, + { + "bbox": [ + 130, + 401, + 482, + 485 + ], + "type": "inline_equation", + "content": "16\\%" + }, + { + "bbox": [ + 130, + 401, + 482, + 485 + ], + "type": "text", + "content": " second, and " + }, + { + "bbox": [ + 130, + 401, + 482, + 485 + ], + "type": "inline_equation", + "content": "8.5\\%" + }, + { + "bbox": [ + 130, + 401, + 482, + 485 + ], + "type": "text", + "content": " third. This preference for RECON was statistically significant (p-value " + }, + { + "bbox": [ + 130, + 401, + 482, + 485 + ], + "type": "inline_equation", + "content": "< 0.05" + }, + { + "bbox": [ + 130, + 401, + 482, + 485 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 506, + 220, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 506, + 220, + 519 + ], + "spans": [ + { + "bbox": [ + 132, + 506, + 220, + 519 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "content": "We proposed RECON, a training-free retrieval-based acceleration method for T2I tasks. We observed that previous approaches were hindered by the rigidity of their trajectories, which often resulted in images lacking fidelity and diversity. To address these challenges, RECON retrieves trajectories for the noise by using concept similarity and then merges the noises derived from multiple semantically similar concepts. The visual concepts extracted for this process have broad applicability. While we focus on retrieval-based acceleration in this paper, the potential uses of these concepts can be applied to noise aesthetics classifications [13], noise inversion techniques [61], and latent-to-latent transformations within generative models [4]. We leave the exploration of our concept knowledge base as future work." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "C. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 211 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 211 + ], + "type": "text", + "content": "This material is based in part upon work supported by the National Science Foundation under Grant Number CNS-2146449 (NSF CAREER award), gift funding from Adobe Research, and by the Army Research Lab under contract number W911NF-2020-221. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the sponsors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 228, + 197, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 228, + 197, + 241 + ], + "spans": [ + { + "bbox": [ + 133, + 228, + 197, + 241 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 251, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 251, + 480, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 480, + 285 + ], + "type": "text", + "content": "1. christophschuhmann/improved-aesthetic-predictor: Clip+mlp aesthetic score predictor. https://github.com/christophschuhmann/improved-aesthetic-predictor" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 285, + 329, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 329, + 296 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 329, + 296 + ], + "type": "text", + "content": "2. Dall-e 2. https://openai.com/dall-e-2 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 296, + 480, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 296, + 480, + 350 + ], + "spans": [ + { + "bbox": [ + 138, + 296, + 480, + 350 + ], + "type": "text", + "content": "3. Agarwal, S., Mitra, S., Chakraborty, S., Karanam, S., Mukherjee, K., Saini, S.K.: Approximate caching for efficiently serving Text-to-Image diffusion models. In: 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI 24). pp. 1173-1189. USENIX Association, Santa Clara, CA (Apr 2024), https://www.usenix.org/conference/nsdi24/presentation/agarwal-shubham" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 351, + 480, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 351, + 480, + 372 + ], + "spans": [ + { + "bbox": [ + 138, + 351, + 480, + 372 + ], + "type": "text", + "content": "4. Bojanowski, P., Joulin, A., Lopez-Paz, D., Szlam, A.: Optimizing the latent space of generative networks. arXiv preprint arXiv:1707.05776 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 373, + 480, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 373, + 480, + 404 + ], + "spans": [ + { + "bbox": [ + 138, + 373, + 480, + 404 + ], + "type": "text", + "content": "5. Brysbaert, M., Warriner, A.B., Kuperman, V.: Concreteness ratings for 40 thousandly known english word lemmas. Behavior research methods 46, 904-911 (2014)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 405, + 480, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 405, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 138, + 405, + 480, + 437 + ], + "type": "text", + "content": "6. Cai, D., Wang, Y., Liu, L., Shi, S.: Recent advances in retrieval-augmented text generation. In: Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval. pp. 3417-3419 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 437, + 480, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 437, + 480, + 469 + ], + "spans": [ + { + "bbox": [ + 138, + 437, + 480, + 469 + ], + "type": "text", + "content": "7. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dollar, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325 (2015)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 470, + 480, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 470, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 470, + 480, + 491 + ], + "type": "text", + "content": "8. Choi, J., Choi, Y., Kim, Y., Kim, J., Yoon, S.: Custom-edit: Text-guided image editing with customized diffusion models. arXiv preprint arXiv:2305.15779 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 492, + 480, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 492, + 480, + 524 + ], + "spans": [ + { + "bbox": [ + 138, + 492, + 480, + 524 + ], + "type": "text", + "content": "9. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 525, + 480, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 480, + 545 + ], + "type": "text", + "content": "10. Deng, Y., Wu, N., Qiu, C., Luo, Y., Chen, Y.: Mixgan-tts: Efficient and stable speech synthesis based on diffusion model. IEEE Access (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 546, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 546, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 546, + 480, + 567 + ], + "type": "text", + "content": "1. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 568, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 480, + 589 + ], + "type": "text", + "content": "2. Du, Y., Li, S., Mordatch, I.: Compositional visual generation with energy based models. Advances in Neural Information Processing Systems 33, 6637-6647 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 590, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 611 + ], + "type": "text", + "content": "3. Gallego, V.: Personalizing text-to-image generation via aesthetic gradients. arXiv preprint arXiv:2209.12330 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 612, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 612, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 612, + 480, + 643 + ], + "type": "text", + "content": "4. Gu, J., Zhai, S., Zhang, Y., Liu, L., Susskind, J.M.: Boot: Data-free distillation of denoising diffusion models with bootstrapping. In: ICML 2023 Workshop on Structured Probabilistic Inference and Generative Modeling (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "5. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control (2022)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 230, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 230, + 91, + 447, + 102 + ], + "type": "text", + "content": "RECON: Concept Retrieval for Diffusion Acceleration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "16. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 171 + ], + "type": "text", + "content": "17. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 193 + ], + "type": "text", + "content": "18. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "type": "text", + "content": "19. Jeong, J.W., Wang, X.J., Lee, D.H.: Towards measuring the visualness of a concept. In: Proceedings of the 21st ACM international conference on Information and knowledge management. pp. 2415-2418 (2012)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "text", + "content": "20. Johnson, J., Douze, M., Jégou, H.: Billion-scale similarity search with GPUs. IEEE Transactions on Big Data 7(3), 535-547 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "text", + "content": "21. Karnewar, A., Vedaldi, A., Novotny, D., Mitra, N.J.: Holodiffusion: Training a 3d diffusion model using 2d images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18423-18433 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 480, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 313 + ], + "type": "text", + "content": "22. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems 35, 26565-26577 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "text", + "content": "23. Kirstain, Y., Polyak, A., Singer, U., Matiana, S., Penna, J., Levy, O.: Pick-a-pic: An open dataset of user preferences for text-to-image generation. arXiv preprint arXiv:2305.01569 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "24. Kong, Z., Ping, W.: On fast sampling of diffusion probabilistic models. In: ICML Workshop on Invertible Neural Networks, Normalizing Flows, and Explicit Likelihood Models (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "25. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1931-1941 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 414, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 435 + ], + "type": "text", + "content": "26. Li, J., Sun, A., Han, J., Li, C.: A survey on deep learning for named entity recognition. IEEE Transactions on Knowledge and Data Engineering 34(1), 50-70 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "text", + "content": "27. Liu, J., Li, C., Ren, Y., Chen, F., Zhao, Z.: Diffsinger: Singing voice synthesis via shallow diffusion mechanism. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 11020-11028 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "text", + "content": "28. Liu, N., Li, S., Du, Y., Tenenbaum, J., Torralba, A.: Learning to compose visual relations. Advances in Neural Information Processing Systems 34, 23166-23178 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "text", + "content": "29. Liu, N., Li, S., Du, Y., Torralba, A., Tenenbaum, J.B.: Compositional visual generation with composable diffusion models. In: European Conference on Computer Vision. pp. 423-439. Springer (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "text", + "content": "30. Liu, X., Zhang, X., Ma, J., Peng, J., et al.: Instafaflow: One step is enough for high-quality diffusion-based text-to-image generation. In: The Twelfth International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "text", + "content": "31. Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., Stoyanov, V.: Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "32. Louis, A., Nenkova, A.: What makes writing great? first experiments on article quality prediction in the science journalism domain. Transactions of the Association for Computational Linguistics 1, 341-352 (2013)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "33. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems 35, 5775-5787 (2022)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "C. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 139 + ], + "type": "text", + "content": "34. Luhman, E., Luhman, T.: Knowledge distillation in iterative generative models for improved sampling speed. arXiv preprint arXiv:2101.02388 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 161 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 161 + ], + "type": "text", + "content": "35. Meng, C., Rombach, R., Gao, R., Kingma, D., Ermon, S., Ho, J., Salimans, T.: On distillation of guided diffusion models. In: CVPR (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 162, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 162, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 162, + 482, + 205 + ], + "type": "text", + "content": "36. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 206, + 482, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 206, + 482, + 239 + ], + "spans": [ + { + "bbox": [ + 130, + 206, + 482, + 239 + ], + "type": "text", + "content": "37. von Platen, P., Patil, S., Lozhkov, A., Cuenca, P., Lambert, N., Rasul, K., Davaadorj, M., Wolf, T.: Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 240, + 482, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 240, + 482, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 482, + 274 + ], + "type": "text", + "content": "38. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: improving latent diffusion models for high-resolution image synthesis. pp. 1-13 (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 274, + 482, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 317 + ], + "type": "text", + "content": "39. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 318, + 482, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 482, + 351 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 482, + 351 + ], + "type": "text", + "content": "40. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 352, + 482, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 385 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 385 + ], + "type": "text", + "content": "41. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 385, + 482, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 482, + 430 + ], + "type": "text", + "content": "42. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 430, + 482, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 474 + ], + "type": "text", + "content": "43. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 475, + 482, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 496 + ], + "type": "text", + "content": "44. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 497, + 482, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 530 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 530 + ], + "type": "text", + "content": "45. Sheynin, S., Ashual, O., Polyak, A., Singer, U., Gafni, O., Nachmani, E., Taigman, Y.: Knn-diffusion: Image generation via large-scale retrieval. arXiv preprint arXiv:2204.02849 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 531, + 482, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 482, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 482, + 563 + ], + "type": "text", + "content": "46. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256-2265. PMLR (2015)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 564, + 482, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 564, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 564, + 482, + 586 + ], + "type": "text", + "content": "47. Solovyev, V.: Concreteness/abstractness concept: State of the art. In: International Conference on Cognitive Sciences. pp. 275-283. Springer (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 587, + 482, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 587, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 587, + 482, + 609 + ], + "type": "text", + "content": "48. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: International Conference on Learning Representations (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 609, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 609, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 609, + 482, + 632 + ], + "type": "text", + "content": "49. Song, Y., Dhariwal, P.: Improved techniques for training consistency models. In: The Twelfth International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "text", + "content": "50. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models. Proceedings of Machine Learning Research, vol. 202. PMLR (23-29 Jul 2023), https://proceedings.mlr.press/v202/song23a.htm1" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 230, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 230, + 91, + 447, + 102 + ], + "type": "text", + "content": "RECON: Concept Retrieval for Diffusion Acceleration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 643 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "51. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: International Conference on Learning Representations (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "text", + "content": "52. Verma, G., Rossi, R.A., Tensmeyer, C., Gu, J., Nenkova, A.: Learning the visualness of text using large vision-language models. In: The 2023 Conference on Empirical Methods in Natural Language Processing (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "text", + "content": "53. Verma, G., Rossi, R.A., Tensmeyer, C., Gu, J., Nenkova, A.: Learning the visualness of text using large vision-language models. arXiv preprint arXiv:2305.10434 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 258 + ], + "type": "text", + "content": "54. Wang, T., Zhang, B., Zhang, T., Gu, S., Bao, J., Baltrusaitis, T., Shen, J., Chen, D., Wen, F., Chen, Q., et al.: Rodin: A generative model for sculpting 3d digital avatars using diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4563-4573 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "text", + "content": "55. Wang, Z.J., Montoya, E., Munechika, D., Yang, H., Hoover, B., Chau, D.H.: Diffusiondb: A large-scale prompt gallery dataset for text-to-image generative models. arXiv preprint arXiv:2210.14896 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "56. Xu, X., Wang, Z., Zhang, G., Wang, K., Shi, H.: Versatile diffusion: Text, images and variations all in one diffusion model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7754-7765 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 369 + ], + "type": "text", + "content": "57. Yasuda, Y., Toda, T.: Text-to-speech synthesis based on latent variable conversion using diffusion probabilistic model and variational autoencoder. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 369, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 423 + ], + "type": "text", + "content": "58. Zhang, K., Yang, X., Wang, W.Y., Li, L.: ReDi: Efficient learning-free diffusion inference via trajectory retrieval. In: Proceedings of the 40th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 202, pp. 41770-41785. PMLR (23-29 Jul 2023), https://proceedings.mlr.press/v202/zhang23as.htm1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "text", + "content": "59. Zhang, Q., Chen, Y.: Fast sampling of diffusion models with exponential integrator. In: NeurIPS 2022 Workshop on Score-Based Methods (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "type": "text", + "content": "60. Zhang, Q., Chen, Y.: Fast sampling of diffusion models with exponential integrator. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=Loek7hfb46P" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "type": "text", + "content": "61. Zhang, Y., Huang, N., Tang, F., Huang, H., Ma, C., Dong, W., Xu, C.: Inversion-based style transfer with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10146-10156 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "62. Zhao, W., Bai, L., Rao, Y., Zhou, J., Lu, J.: Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. arXiv preprint arXiv:2302.04867 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 545, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 578 + ], + "type": "text", + "content": "63. Zheng, H., Nie, W., Vahdat, A., Azizzadenesheli, K., Anandkumar, A.: Fast sampling of diffusion models via operator learning. In: International Conference on Machine Learning. pp. 42390-42402. PMLR (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "text", + "content": "64. Zhou, L., Du, Y., Wu, J.: 3d shape generation and completion through point-voxel diffusion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5826-5835 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "text", + "content": "65. Zhou, Y., Liu, B., Zhu, Y., Yang, X., Chen, C., Xu, J.: Shifted diffusion for text-to-image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10157-10166 (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "C. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/ReGround_ Improving Textual and Spatial Grounding at No Cost/e369c9ee-bed9-4dbc-92b4-a4d9da7bccb2_content_list.json b/2024/ReGround_ Improving Textual and Spatial Grounding at No Cost/e369c9ee-bed9-4dbc-92b4-a4d9da7bccb2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a8d635a21b364e222eeafaf00c08a0ae6e54a47a --- /dev/null +++ b/2024/ReGround_ Improving Textual and Spatial Grounding at No Cost/e369c9ee-bed9-4dbc-92b4-a4d9da7bccb2_content_list.json @@ -0,0 +1,2290 @@ +[ + { + "type": "text", + "text": "ReGround: Improving Textual and Spatial Grounding at No Cost", + "text_level": 1, + "bbox": [ + 259, + 140, + 743, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Phillip Y. Lee Minhyuk Sung", + "bbox": [ + 370, + 212, + 630, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "KAIST {phillip0701,mhsung}@kaist.ac.kr", + "bbox": [ + 375, + 239, + 625, + 268 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1d0c056f8167f65e44be115b293f34e721c4300ab57e22eeddd617e09ab924ea.jpg", + "image_caption": [ + "Layout" + ], + "image_footnote": [], + "bbox": [ + 217, + 289, + 330, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/131b71cfcede5f2bf278e995186aaa139d9015eec8f10d835a63409c4db35f78.jpg", + "image_caption": [ + "SD" + ], + "image_footnote": [], + "bbox": [ + 331, + 290, + 442, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7ac1edcb020215b834573c83db9cf4616a37af57a4a20bc2be8b7753ddc803e9.jpg", + "image_caption": [ + "GLIGEN $_{\\gamma=1.0}$" + ], + "image_footnote": [], + "bbox": [ + 444, + 290, + 555, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fb658c1a508ade0fb987ceb1b7d7a7a4c961aabc561756948e1df1fc90205279.jpg", + "image_caption": [ + "GLIGEN $\\gamma = 0.2$" + ], + "image_footnote": [], + "bbox": [ + 558, + 290, + 669, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/014e6a24b2cdfa600ad27b111698d1a5a82f6d827861c2d7efcf58c497c57e6f.jpg", + "image_caption": [ + "ReGround" + ], + "image_footnote": [], + "bbox": [ + 671, + 290, + 784, + 375 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"A low poly illustration of a man holding a cell phone in hand.\"", + "bbox": [ + 290, + 376, + 712, + 390 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/32207f22b62516630ca9333c10d1829d11ef4b4923178b68ff933b2fce05f8a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 392, + 330, + 477 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c656866eb405f6fb36e983b7ee9e9de11abfd851d922a8674ba1fa05a25b0a1c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 392, + 442, + 477 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4e49957eca6d7c198c6b075b8f674ccb9985cd69f0e1857e74cf1c90a7fd9681.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 392, + 555, + 477 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1f2226629e40a960990c4ee44625ad1a3992f5826ee58d0b37eb7cc5e528ae17.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 392, + 669, + 477 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ac7698699064277a50e336d9f83c907b3609b2a0fee07de80b1ce7ad2bdec67e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 392, + 784, + 477 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"An elephant walking draped with a colorful blanket.\"", + "bbox": [ + 290, + 478, + 681, + 491 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cdc161ca5382b0ae0e708392168743eafaac4a068e028194e030aa99d1e215e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 493, + 330, + 579 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1ff27e6f5adda68fcdb071a23fbd5c63231df4259846d6bf1721cec30532fc12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 493, + 442, + 579 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3773a220c49c8bf021a97176a8f27e99d7594e3847750a4dc1ffd33e258d8f78.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 493, + 555, + 579 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2884fd53f2d1872e05de42a56061943e924778823667f798fdeabb0afe51f48b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 493, + 669, + 579 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/02a06234d99acea9dbb1851d2d23f2b2ad2fefd396b152cc367f7ea8bf635d8d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 493, + 784, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"One ancient vase [...] the backdrop of a serene, moonlit desert landscape.\"", + "bbox": [ + 250, + 580, + 751, + 594 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3f60db638b8d8f8cd4e4d248b92d3f05bde298a9ce9be200534b72a141fe91f5.jpg", + "image_caption": [ + "Fig. 1: Comparison across Stable Diffusion (SD) [41], GLIGEN [28], and our REGROUND. SD (2nd column) can generate an image aligned with the input prompt (shown below each row), while it does not allow taking spatial constraints such as bounding boxes and labels. GLIGEN (3rd column) enables spatial grounding using gated self-attention, although it often disregards some descriptions in the input prompt due to a bias towards bounding box conditions. Such trends also occur when only activating gated self-attention for 0.2 fraction of the initial denoising steps (4th column). Our REGROUND (last column) resolves the issue of description omission while accurately reflecting the bounding box information." + ], + "image_footnote": [], + "bbox": [ + 218, + 595, + 330, + 681 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eb5bec30d4184918bfec9d8a869b091573b3327d63ee51e92661c11c033a4895.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 595, + 442, + 681 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6de29257702448a1413cf3579dd8fc888b7a2674e461cadd70edebe8c6466950.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 595, + 555, + 681 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e5c0060541006735ca3e0f4eb64824724310d4684354669d107ea9ecf2c5dd0f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 595, + 669, + 681 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1ec0bb9be383b766f86b47560193ef648b2841ca7ef2cd9bd719c79998c7bbd4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 595, + 784, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"A train blowing smoke is coming down the tracks.\"", + "bbox": [ + 330, + 683, + 671, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. When an image generation process is guided by both a text prompt and spatial cues, such as a set of bounding boxes, do these elements work in harmony, or does one dominate the other? Our analysis of a pretrained image diffusion model that integrates gated self-attention into the U-Net reveals that spatial grounding often outweighs textual grounding due to the sequential flow from gated self-attention to cross-attention. We demonstrate that such bias can be significantly mitigated without sacrificing accuracy in either grounding by simply rewiring the network architecture, changing from sequential to parallel for gated self-attention and cross-attention. This surprisingly simple yet effective solution does not require any fine-tuning of the network but substantially reduces the trade-off between the two groundings. Our experiments demonstrate significant improvements from the original GLIGEN to the rewired version in the trade-off between textual grounding and spatial grounding. The project webpage is at https://re-ground.github.io.", + "bbox": [ + 261, + 147, + 743, + 357 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Keywords: Textual Grounding $\\cdot$ Spatial Grounding $\\cdot$ Network Rewiring", + "bbox": [ + 261, + 369, + 740, + 383 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 411, + 375, + 428 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The emergence of diffusion models [17,44,45] has markedly propelled the field of text-to-image (T2I) generation forward, allowing users to generate high-quality images from text prompts. In a bid to further augment the creativity and controllability, recent efforts [2-5,8,11,25,28,32,36,54,55,57] have focused on enabling these models to understand and interpret spatial instructions, such as layouts [4,8,28,32,36,54,55], segmentation masks [2-5,11,25] and sketches [51,57].", + "bbox": [ + 212, + 446, + 787, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Among them, bounding boxes are extensively employed in downstream image generation tasks [4, 8, 28, 32, 36, 55]. GLIGEN [28] is a pioneering work in terms of enhancing existing T2I models with the capability to incorporate additional spatial cues in the form of bounding boxes. Its core component, gated self-attention, is a simple attention module [50] that is plugged into each U-Net [42] layer of a pretrained T2I model such as Stable Diffusion [41], and is trained to accurately position various entities in their designated areas. A notable advantage of GLIGEN is that the original parameters of the underlying model remain unchanged, inheriting the generative capability of the T2I model while introducing the novel functionality of spatial grounding using bounding boxes. This capability has been leveraged by numerous studies to facilitate high-quality, layout-guided image generation [9, 12, 31, 54].", + "bbox": [ + 212, + 537, + 787, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, our analysis reveals that GLIGEN's integration of the gated self-attention into an existing T2I model is not optimal for blending new spatial guidance from bounding boxes with the original textual guidance. It often leads to the omission of specific details from the text prompts. For instance, in the first row and third column of Fig. 1, GLIGEN fails to reflect the description \"low poly illustration\" from the input text prompt. Also in the second row, a crucial detail in the text prompt, \"draped with a colorful blanket\", is neglected in the output image. We refer to this issue as description omission. Such outcomes imply", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "P.Y. Lee and M. Sung", + "bbox": [ + 271, + 114, + 421, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "that the current architectural design of GLIGEN does not effectively harmonize the new spatial guidance and the text conditioning in the given T2I model. Considering the widespread applications of GLIGEN in various layout-based generation tasks [9,12,31,36,53,54,59], these limitations represent a significant bottleneck.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the observed neglect of textual grounding in GLIGEN, we first analyze the root causes. Our investigation reveals that the issue arises from the sequential arrangement of the spatial grounding and textual grounding modules. Specifically, the output of the gated self-attention is directed to a cross-attention module in each layer of the U-Net architecture (Fig. 2-(b)).", + "bbox": [ + 212, + 222, + 784, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Building on this insight, we propose a straightforward yet impactful solution: network rewiring. This approach alters the relationship between the two grounding modules from sequential to parallel (Fig. 2-(c)). Remarkably, this network modification significantly reduces the grounding trade-off between textual and spatial groundings without necessitating any adjustments to the network parameters. Importantly, this rewiring does not require additional network training, extra parameters, or changes in computational load and time. Simply reconfiguring the attention modules of the pretrained GLIGEN, originally trained with the sequential architecture, during inference dramatically enhances performance.", + "bbox": [ + 212, + 297, + 785, + 448 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our experiments on MS-COCO [30] and our newly introduced NSR-1K-GPT datasets, we demonstrate that rewiring the pretrained GLIGEN substantially reduces the trade-off between textual and spatial groundings. This is evidenced by the evaluation of text prompt alignment (measured using CLIP score [38], PickScore [26] and user study) and bounding box alignment (assessed by YOLO score [52]). Furthermore, we show that our rewiring also leads to better outcomes in other frameworks using GLIGEN as a backbone, including BoxDiff [54].", + "bbox": [ + 212, + 449, + 787, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 595, + 387, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Zero-Shot Guidance in Diffusion Models", + "text_level": 1, + "bbox": [ + 214, + 630, + 593, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The progress in diffusion models [17, 44, 45] has significantly elevated the capabilities of text-to-image (T2I) generation, resulting in foundation models [6, 37, 39-41] that exhibit remarkable generative performance. Leveraging the robust performance of these models, recent studies [3-5, 8, 11, 25, 28, 32, 36, 54, 55, 57] have introduced efficient guidance techniques designed to further improve the image generation process. Notably, numerous works [8, 28, 32, 36, 43, 54, 55] focus on the internal architecture (Fig. 2-(a)) of the denoising U-Net of Latent Diffusion Models (LDMs) [41], where self-attention and cross-attention modules are intertwined to facilitate inter-pixel communication and text conditioning. The self-attention of U-Net can be utilized to improve image quality [19] or facilitate image translation [49] and image editing tasks [7]. Since text conditions are integrated via cross-attention, the intermediate attention maps have been leveraged", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "ReGround: Improving Textual and Spatial Grounding at No Cost", + "bbox": [ + 294, + 114, + 732, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to improve text faithfulness [13] or enable spatial manipulation of the generation process [35]. Recently, FreeU [43] analyzed the contributions of the backbone and residuals of the U-Net and proposed a free-lunch strategy to enhance image quality: reweighting the backbone and residual features maps. In contrast to previous works that only deal with self- and cross-attention in standard LDMs, we introduce a method to enhance GLIGEN [28] by reconnecting its gated self-attention with the other attention modules, thereby achieving performance improvement in zero-shot without any tuning of the network parameters.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Layout-Guided Image Generation", + "text_level": 1, + "bbox": [ + 214, + 286, + 539, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The use of layouts, particularly in the form of bounding boxes, has become a popular intermediary to bridge the gap between textual inputs and the images generated [14, 18, 21, 27, 29, 46, 47, 56, 58]. Layout2Im [58] samples object latent codes from a normal distribution, eliminating the need to predict instance masks as done in prior works [18, 21]. LostGAN [46] controls the style of each object by devising an extension of the feature normalization layer used in StyleGAN [22-24], while OC-GAN [47] incorporates the spatial relationships between objects using a scene-graph representation. LAMA [29] introduces a mask adaptation module that mitigates the semantic ambiguity arising from overlaps in the input layout. While these developments have greatly improved user control over image generation, their applicability is confined to the categories found in the training data, such as those of the MS-COCO [30] dataset.", + "bbox": [ + 212, + 308, + 787, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast, recent studies [5, 8, 10, 28, 36, 54, 55, 60] have extended layout-guided image generation towards open-vocabulary, building on the advancements of foundational text-to-image (T2I) models [41]. Training-free approaches [3, 5, 8, 36, 54] aim to improve the spatial grounding of T2I models through straightforward guidance mechanisms. GLIGEN [28], on the other hand, introduces gated self-attention, which is injected into the U-Net architecture of the Latent Diffusion Model [41], and is trained to equip the underlying model with spatial grounding abilities. Given the simple architecture of GLIGEN and its robust grounding accuracy with the input bounding boxes, numerous studies [36, 53, 54, 59] build upon its framework and propose further refinements to increase performance. In this work, we identify and address a significant performance bottleneck in GLIGEN related to description omission and propose a simple yet effective solution.", + "bbox": [ + 212, + 489, + 789, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Background - Latent Diffusion Models [41]", + "text_level": 1, + "bbox": [ + 214, + 690, + 679, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rombach et al. [41] proposed Latent Diffusion Model (LDM), a text-to-image (T2I) diffusion model with a U-Net as the noise prediction network. It is trained to generate an image from an input text prompt by predicting the noise $\\epsilon (\\mathbf{x}_t,t,c)$ conditioned both on the timestep $t$ and the text embedding $c$ . Each layer of LDM's U-Net consists of three core components: a convolutional residual block, followed by a self-attention (SA), and a cross-attention (CA) module (Fig. 2-(a)). In each $l$ -th layer of the U-Net, its residual block first extracts intermediate visual features $F = (f_{1},\\dots,f_{N_{l}})^{T}$ from the output of the previous layer.", + "bbox": [ + 212, + 719, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "P.Y. Lee and M. Sung", + "bbox": [ + 271, + 114, + 421, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/802f6014c8843222c1e0674d593e56802d1180a9b2ff054c97a41caf347e624d.jpg", + "image_caption": [ + "Fig. 2: Comparison between the U-Net architectures of (a) Latent Diffusion Model (LDM) [41], (b) GLIGEN [28] and (c) our REGROUND. From LDM, GLIGEN enables spatial grounding by injecting Gated Self-Attention before cross-attention, forming a sequential flow of them. Based on GLIGEN, our REGROUND changes the relationship of the two attention modules to become parallel, resulting in noticeable improvement in textual grounding while preserving the spatial grounding capability. (The residual block before self-attention is omitted.)" + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 787, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The self-attention module then facilitates interaction between the features in $F$ . Subsequently, the cross-attention module enables the interaction between each visual feature $f_{i}$ and the text embedding $c$ . Throughout this process, the output feature of the previous module is also forwarded through a residual connection, as illustrated in lines 4-5 and 7 of Alg. 1 and also in Fig. 2-(a).", + "bbox": [ + 212, + 381, + 787, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 GLIGEN [28] and Description Omission", + "text_level": 1, + "bbox": [ + 214, + 482, + 638, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we review GLIGEN [28] and its key idea of employing gated self-attention for spatial grounding. Then, we present our key observations on the description omission issue that occurs due to the addition of gated self-attention.", + "bbox": [ + 212, + 515, + 785, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Gated Self-Attention", + "text_level": 1, + "bbox": [ + 214, + 585, + 434, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Li et al. [28] propose a plug-in spatial grounding module, named gated self-attention, which adopts the gated attention mechanism [1] to equip a pretrained T2I diffusion model [41] with spatial grounding capabilities (Fig. 2-(b)). Given a set of bounding boxes and text labels for each of them, let $b_{i}$ be the $xy$ -coordinates of the $i$ -th bounding box's top-left and bottom-right corners, and $p_i$ be the corresponding text label. Then, the $i$ -th grounding token is defined as $g_{i} := \\mathcal{G}(\\mathcal{T}(p_{i}),\\mathcal{F}(b_{i}))$ , where $\\mathcal{T}(\\cdot)$ is a pretrained text encoder [20,38], $\\mathcal{F}(\\cdot)$ is the Fourier embedding [33,48] and $\\mathcal{G}(\\cdot,\\cdot)$ is a shallow MLP network that concatenates the two given embeddings, respectively. Given a set of grounding tokens $\\{g_i\\}$ , gated self-attention learns the self-attention among the unified feature set $(f_1,\\dots,f_{N_l},g_1,\\dots,g_M)$ where $\\{f_i\\}$ is the set of intermediate visual features in the $l$ -th layer of U-Net, and $M$ is the number of bounding boxes.", + "bbox": [ + 212, + 612, + 787, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 2-(b), gated self-attention receives the output of the self-attention along with the residual features as its input and forwards the output features to the cross-attention module. By incorporating gated self-attention into", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "ReGround: Improving Textual and Spatial Grounding at No Cost", + "bbox": [ + 294, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "each layer of the U-Net, the model enables the placement of the entity specified in the text label $p_i$ at the location indicated by the bounding box $b_i$ . Note that the integration of gated self-attention does not require training the network from scratch or fine-tuning it, but can be accomplished simply by training the gated self-attention parameters while keeping all other parameters in the backbone model frozen.", + "bbox": [ + 212, + 146, + 787, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Alg. 1 shows the pseudocode of the U-Net forward-pass including the plug-in of gated self-attention in line 6. Note that $\\beta_{t}$ is set to 1 for GLIGEN. If $\\beta_{t} = 0$ , the algorithm is identical to that of LDM [41].", + "bbox": [ + 212, + 239, + 787, + 287 + ], + "page_idx": 5 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Noise Prediction U-Net with Gated Self-Attention." + ], + "code_body": "Parameters: $\\beta_{t}$ // Weight for GSA. Inputs: $\\mathbf{x}_t, c, \\{g_i\\}_{i=0\\dots N-1}$ ; // Noisy data at timestep $t$ , text condition, and grounding tokens \nOutputs: $\\epsilon_t$ ; // Noise at timestep $t-1$ . \n1 Function U-Net $(\\mathbf{x}_t, c, \\{g_i\\})$ : \n2 $F \\gets \\mathbf{x}_t$ \n3 for $i = 0, \\dots, L-1$ do \n4 $\\begin{array}{c} F_{\\mathrm{RS}} \\gets \\mathrm{Conv}(F) + F; \\\\ F_{\\mathrm{SA}} \\gets \\mathrm{SA}(F_{\\mathrm{RS}}) + F_{\\mathrm{RS}}; \\\\ F_{\\mathrm{GSA}} \\gets \\beta_t \\cdot \\mathrm{GSA}(F_{\\mathrm{SA}}, \\{g_i\\}) + F_{\\mathrm{SA}}; \\\\ F \\gets \\mathrm{CA}(F_{\\mathrm{GSA}}, c) + F_{\\mathrm{GSA}}; \\end{array}$ // Residual block. \n5 // Self-Attention module. \n6 \n7 \n8 $\\epsilon_t \\gets F;$ \n9 return $\\epsilon_t$ ;", + "bbox": [ + 200, + 339, + 787, + 535 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Description Omission", + "text_level": 1, + "bbox": [ + 215, + 604, + 437, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Despite its high accuracy in spatial grounding, GLIGEN [28] frequently struggles to capture essential attributes specified in the input text prompt. As illustrated in Fig. 3, the leftmost image shows \"a person\" and \"a skateboard\" accurately placed in their designated regions. However, a critical detail from the input text prompt, \"black and white photography\", is absent in the output image. This discrepancy often emerges when the input comprises distinct but equally important descriptions regarding the image, presented through text prompts and bounding boxes. Such omissions not only fail to convey the stylistic intent of the image but also tend to overlook significant objects mentioned within the text prompt. Additional examples of this problem are showcased in Fig. 1, where the second row demonstrates the absence of a \"blanket\" in the generated image, a key element from the text prompt. This limitation significantly hampers GLIGEN's fidelity to user-provided text prompts, a challenge we term as description omission.", + "bbox": [ + 212, + 642, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "P.Y. Lee and M. Sung", + "bbox": [ + 271, + 114, + 421, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ddcbaa8b7839a288d6088cc2cee6e9a4d8c2f443f2be629bd09086e49746d7cd.jpg", + "image_caption": [ + "Fig. 3: (a) Images generated by GLIGEN [28] with varying activation duration of gated self-attention $\\gamma$ in scheduled sampling (Sec. 5.1). The red words in the text prompt denote the words used as labels of the input bounding boxes. Note that for GLIGEN to reflect the underlined description in the text prompt in the final image, $\\gamma$ must be decreased to 0.1, which compromises spatial grounding accuracy. (b) In contrast, our REGROUND reflects the underlined phrase even when $\\gamma = 1.0$ , therefore achieving high accuracy in both textual and spatial grounding." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 256 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5 ReGround: Rewiring Attention Modules", + "text_level": 1, + "bbox": [ + 215, + 393, + 643, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Gated self-attention and cross-attention each play a crucial role in enabling spatial and textual groundings, by taking bounding boxes and text prompts as inputs, respectively. To tackle the issue of description omission, we first examine the impact of attention modules on the groundings they do not address: the effect of gated self-attention on textual grounding (Sec. 5.1), and the influence of cross-attention on spatial grounding (Sec. 5.2). Building on this analysis, we propose an approach for network reconfiguration, modifying the connections among self-attention, gated self-attention, and cross-attention modules (Sec. 5.3).", + "bbox": [ + 212, + 425, + 787, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1 Impact of Gated Self-Attention on Textual Grounding", + "text_level": 1, + "bbox": [ + 215, + 569, + 709, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As the issue of description omission arises due to the newly added gated self-attention in GLIGEN [28], we first attempt to mitigate the impact of gated self-attention by using scheduled sampling [28], activating gated self-attention only in a few initial steps of the denoising process. This approach is inspired by the observation that the coarse structure of the final image is established within the first few denoising steps. The scheduling is applied by setting the weight of gated self-attention $\\beta_{t}$ (line 6 of Alg. 1) as", + "bbox": [ + 212, + 595, + 787, + 702 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\beta_ {t} = \\left\\{ \\begin{array}{l l} 1 & (t \\leq \\gamma \\cdot T) \\\\ 0 & (t > \\gamma \\cdot T), \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 419, + 712, + 785, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\gamma \\in [0,1]$ represents the fraction of the initial denoising steps to activate gated self-attention.", + "bbox": [ + 212, + 763, + 785, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 3-(a) shows an example of generated images while incrementally adjusting $\\gamma$ from 1.0 to 0.0. As $\\gamma$ is reduced from 1.0 to 0.0, the details specified in the text prompt, \"a black and white photograph\", begin to be reflected starting", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "ReGround: Improving Textual and Spatial Grounding at No Cost", + "bbox": [ + 294, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "at $\\gamma = 0.1$ , demonstrating that longer activation of gated self-attention may interfere with the alignment of the output image with the text prompt. However, as gated self-attention is activated for shorter durations, the spatial grounding diminishes, as shown in the objects' reduced alignment with the input bounding boxes. This phenomenon illustrates the inherent trade-off between spatial and textual grounding, which cannot be resolved by controlling the duration of gated self-attention activation.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2 Impact of Cross-Attention on Spatial Grounding", + "text_level": 1, + "bbox": [ + 214, + 273, + 663, + 290 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also investigate whether cross-attention has influence on spatial grounding. For this, we conduct a toy experiment by removing cross-attention modules in GLIGEN [28], allowing the output of the gated self-attention to be directly passed to the next layer of the U-Net. This modification is equivalent to changing line 7 of Alg. 1 to $F \\gets F_{GSA}$ .", + "bbox": [ + 212, + 300, + 782, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results are displayed in Fig. 4. Note that, while the appearance of the background and objects changes, the silhouettes of the cat (left) and the individuals (right) remain precisely positioned within their respective bounding boxes without cross-attention. This observation indicates that while gated self-attention that is performed before cross-attention may compromise textual grounding, cross-attention that processes the output of gated self-attention does not affect spatial grounding.", + "bbox": [ + 212, + 375, + 787, + 481 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fdb30a9ad4c464208741327c14aee4da75f9ac9ba57496b212093f5bba5aacb4.jpg", + "image_caption": [ + "Fig. 4: Comparison of the output of GLIGEN [28] with and without cross-attention. While the absence of cross-attention reduces realism and quality of the image, the silhouette of objects remains grounded within the given bounding boxes, as shown in the third column of each case." + ], + "image_footnote": [], + "bbox": [ + 217, + 505, + 772, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3 Network Rewiring: From Sequential to Parallel", + "text_level": 1, + "bbox": [ + 214, + 709, + 650, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Building on the analyses above, we propose a simple yet effective modification to the grounding mechanism, changing the relationship between gated self-attention and cross-attention from sequential to parallel. This change eliminates the placement of gated self-attention before cross-attention, thus preventing the reduction of text grounding caused by gated self-attention. Moreover, in this parallel arrangement, the preservation of spatial grounding is assured, as gated self-attention for spatial grounding does not require subsequent cross-attention.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "P.Y. Lee and M. Sung", + "bbox": [ + 271, + 114, + 421, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Specifically, recall that in GLIGEN [28], the output of gated self-attention is added to the residual from self-attention, which is then passed to the cross-attention module as follows:", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nF _ {G S A} \\gets \\underbrace {\\mathrm {G S A} (F _ {S A} , \\{g _ {i} \\})} + F _ {S A};\n$$\n", + "text_format": "latex", + "bbox": [ + 387, + 195, + 617, + 218 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\ns p a t i a l g r o u n d i n g\n$$\n", + "text_format": "latex", + "bbox": [ + 455, + 220, + 563, + 231 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nF \\leftarrow \\underbrace {\\operatorname {C A} \\left(F _ {G S A} , c\\right)} + F _ {G S A}; \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 227, + 784, + 255 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\text {t e x t u a l g r o u n d i n g}\n$$\n", + "text_format": "latex", + "bbox": [ + 447, + 257, + 557, + 268 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We propose to transform this sequence grounding pipeline into two parallel processes as follows:", + "bbox": [ + 212, + 273, + 782, + 303 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nF \\leftarrow \\underbrace {\\operatorname {G S A} \\left(F _ {S A} , \\{g _ {i} \\}\\right)} + \\underbrace {\\operatorname {C A} \\left(F _ {S A} , c\\right)} + \\underbrace {F _ {S A}}; \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 311, + 784, + 333 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{c c c} \\text {s p a t i a l g r o u n d i n g} & \\text {t e x t u a l g r o u n d i n g} & \\text {r e s i d u a l} \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 334, + 673, + 345 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Refer to Fig. 2 for the visualization of network architecture changes $((\\mathrm{b})\\rightarrow$ (c)). This network rewiring is feasible because the input to gated self-attention remains unchanged, while the input to cross-attention shifts to $F_{SA}$ , for which it was originally designed in the context of Latent Diffusion Models [41].", + "bbox": [ + 212, + 352, + 782, + 411 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "It is important to note that the modification is effective even when applied to the pretrained GLIGEN, which was trained with the sequential structure of the attention modules. Therefore, our rewiring does not require any additional training or fine-tuning, introduces no extra parameters, and does not affect computation time or memory usage during the generation process. The only requirement is the simple reconfiguration of the attention modules at inference time.", + "bbox": [ + 212, + 412, + 784, + 516 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6 Experiments", + "text_level": 1, + "bbox": [ + 214, + 537, + 375, + 553 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we show the effectiveness of our REGROUND by evaluating the spatial grounding on existing layout-caption datasets [12, 30] and the textual grounding on text-image alignment metrics [26, 38]. We use the official GLI-GEN [28] checkpoint which is trained based on Stable Diffusion v1.4 [41].", + "bbox": [ + 212, + 564, + 782, + 625 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6.1 Datasets", + "text_level": 1, + "bbox": [ + 214, + 643, + 334, + 656 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "MS-COCO. We use the validations sets of both MS-COCO-2014 and MS-COCO-2017 datasets [30]. Each dataset provides image-captions pairs and the $xy$ -coordinates of bounding boxes along with their corresponding object categories.", + "bbox": [ + 212, + 662, + 782, + 724 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "NSR-1K-GPT. We also use the NSR-1K benchmark [12] for evaluation. Based on each subset of NSR-1K—Counting and Spatial—we develop a new benchmark, NSR-1K-GPT, augmenting each original caption in NSR-1K using GPT-4 [34]. The instructions for augmentation are to (i) elaborate on the descriptions of each mentioned entity and (ii) provide additional details about the background of the image. More details on the evaluation datasets are provided in the Supplementary (Sec. S1).", + "bbox": [ + 212, + 733, + 784, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "ReGround: Improving Textual and Spatial Grounding at No Cost", + "bbox": [ + 294, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/fcb2b6d78cd34047407c1f0585f925c33cc7af3307104afe66252414890e1c38.jpg", + "image_caption": [ + "Layout" + ], + "image_footnote": [], + "bbox": [ + 218, + 155, + 328, + 239 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/a675bdb566510ebf7cb0b1615466b8a93bc97b7a7346aac8e5f451ffcfded1b8.jpg", + "image_caption": [ + "SD", + "\"A woman in a white shirt standing in front of a fence smiling.\"" + ], + "image_footnote": [], + "bbox": [ + 330, + 155, + 439, + 239 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/d7c222f0bd883054fe1d39fae5abb749c32032149f5aaab54981d698f2458d76.jpg", + "image_caption": [ + "GLIGEN $_{\\gamma = 1.0}$" + ], + "image_footnote": [], + "bbox": [ + 442, + 155, + 552, + 239 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/845b0ae98dbadca274469004f638a143c20fe5a1e36d67c4918fd54e3588b82c.jpg", + "image_caption": [ + "GLIGEN $_{\\gamma=0.2}$" + ], + "image_footnote": [], + "bbox": [ + 555, + 155, + 663, + 239 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/1375ee709e14aa8fb2bd13955715c7044478935ac6fcb298a8d5843676715f05.jpg", + "image_caption": [ + "ReGround" + ], + "image_footnote": [], + "bbox": [ + 666, + 155, + 777, + 239 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/dad2d1026f04d0c6478c87e2d428c721673546bdf6a61026ba3ef059829366b4.jpg", + "image_caption": [ + "\"One spotted cow in the photo, [...] under a clear blue sky.\"" + ], + "image_footnote": [], + "bbox": [ + 218, + 255, + 328, + 340 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/2f6994c3ce78575f1cbcebfa08d16db42204feeab6b5d351a24e504001fb1ae9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 330, + 255, + 439, + 340 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/94a994ac988c5dff30ab9111a6f278f4a51f117fe942d9050ed1435c81e5bd89.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 255, + 552, + 340 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/9e03c7250da3b602519b0b60fa353d6ba1132df0400b9a0a0095eaa72fc4be79.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 255, + 665, + 340 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/cb715a930d882e537f9385274bf29e78ae31fff45227fe614fce9bef8e5cd995.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 668, + 255, + 777, + 340 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/62f7e592193a8e79bfdf41c1b89b4b9c9ae87aa64ba3a9cf87093b935c77522b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 358, + 328, + 443 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/f705d3015702df30eca3851dd4a46b196a401ae91c38651b6cd7169323368348.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 330, + 358, + 439, + 441 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/af32fff613c22d2d3709ebb5026de943aa5f418e18254eed378b0f0d30fe5a27.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 358, + 552, + 441 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/9562212cbb0b7c310b47e1d9cd7fc5ceda3db46108367714a913f0afd80d290b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 358, + 665, + 441 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/b378e9f4021dd9edecf2acd8d107137fab8f87545517f4792bfc3617cad4edbe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 668, + 358, + 777, + 441 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/eb110959bbfc8a2de9baf1402357a140b75c052aef93e50041a9b86aac63f924.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 459, + 328, + 545 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/8bab439c32472a70460c103a943b9b263d145fe58d8940e04743278b5e3ed4fb.jpg", + "image_caption": [ + "\"A foreign airline status board with a clock.\"" + ], + "image_footnote": [], + "bbox": [ + 330, + 459, + 439, + 544 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/1cd6360676e371ed83e2c8bd664bee1913fbdd9a273adf39b9b8dddb5be882b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 459, + 552, + 544 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/810c30771373b86f7389c53a4ad513a7f01be3ba1d5b297512fd544384cb2f46.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 459, + 665, + 544 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/78fa721a1debaec8c10915a39756a78e4f6d754978ef2add0ef5fb711569d219.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 668, + 459, + 777, + 544 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/afd685b2510da2353ac8f4552949cb28efea8873e1058f4ce767bc41dfc6ad9c.jpg", + "image_caption": [ + "\"A palm tree shower curtain is shown in a small bathroom.\"", + "Fig. 5: Qualitative comparisons. Stable Diffusion (SD, 2nd column) generates images that align with the given text descriptions, including the underlined phrase in each row, but cannot take bounding boxes as input. GLIGEN (3rd column) creates images that match the input layouts but suffers from description omission, failing to reflect the underlined descriptions. Scheduled sampling strategy (4th column) can partially address this issue (for instance, in the 5th row, where \"window\" appears in the room), but it results in a noticeable decline in spatial accuracy (as seen in the 1st row, where the tie is not generated). In contrast, our method (last column) accurately incorporates the underlined text descriptions while maintaining precise spatial representation." + ], + "image_footnote": [], + "bbox": [ + 218, + 563, + 328, + 648 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/373851e8d28802d2552548f1c9b2bfd45495b7e8fa1c1b17b6ac2ca8d53eb0f2.jpg", + "image_caption": [ + "\"A view of a living room with large windows.\"" + ], + "image_footnote": [], + "bbox": [ + 330, + 563, + 439, + 647 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/24a5edcf557748b4bfcb3a764ca63faa0441f2bd47b6c937e076fabf60447187.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 563, + 552, + 647 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/bdb1c850f1c5a9cde54fac9fc0c8ea268cde084ba689c1f4bacc568b5df350cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 563, + 665, + 647 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/e1a731f958013e7fcf4c2a0d618735081c0fb616071736486cb0c9cee096f54d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 668, + 563, + 777, + 647 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "P.Y. Lee and M. Sung", + "bbox": [ + 271, + 114, + 421, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/81fb26cead795d08fca56286d50264e9abbb914b963c68023781836f7a44c233.jpg", + "image_caption": [ + "(a) MS-COCO-2014", + "Fig. 6: Comparisons on MS-COCO [30] and NSR-1K-GPT. Each plot shows the relationship between textual grounding (i.e. CLIP score [15]) and spatial grounding (i.e. YOLO score [52]) accuracy of GLIGEN [28] and our method. Note that the plot of our REGROUND is positioned in the top-right quadrant relative to GLIGEN, signifying that it alleviates the inherent trade-off between textual and spatial grounding." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 405, + 253 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7ce3c5555acc970e8f7cfd00eb0d5d73356fad84ded5b23a500cb609be25f85d.jpg", + "image_caption": [ + "(b) MS-COCO-2017" + ], + "image_footnote": [], + "bbox": [ + 405, + 143, + 591, + 253 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ff51c7df2da5ef955b127e09409a2c0cf616d3cf16eb22df3b912091f914057f.jpg", + "image_caption": [ + "(c) NSR-1K-GPT-Counting" + ], + "image_footnote": [], + "bbox": [ + 594, + 143, + 785, + 253 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6.2 Evaluation Metrics", + "text_level": 1, + "bbox": [ + 215, + 377, + 419, + 391 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- YOLO score: Spatial grounding accuracy is assessed usingYOLO score [52]. We employYOLOv7 [52] to detect objects in each generated image and compute the average precision (AP) based on the ground truth bounding box annotations from MS-COCO [30].", + "- CLIP score: Textual grounding accuracy is assessed using CLIP score [15].", + "- FID: Image quality and diversity are evaluated using FID [16].", + "- User Study and PickScore: We conduct a user study to assess human preferences for the generated images based on each input text prompt. Additionally, we use PickScore [26], a human preference predictor, to further analyze the results." + ], + "bbox": [ + 225, + 401, + 785, + 551 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6.3 Comparison with GLIGEN", + "text_level": 1, + "bbox": [ + 215, + 574, + 485, + 589 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Textual-Spatial Grounding Trade-off. We first examine the trade-off between textual and spatial groundings for both GLIGEN [28] and our REGROUND, the rewired version of GLIGEN, while varying the scheduled sampling parameter $\\gamma$ from 1.0 to 0.1.", + "bbox": [ + 214, + 598, + 785, + 657 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Fig. 6-(a), (b) present the graphs of CLIP score [15] and YOLO score [52] measured on the MS-COCO datasets [30]. In MS-COCO-2014, when reducing $\\gamma$ from 1.0 to 0.1, the CLIP score of GLIGEN varies from 30.44 to 31.65, while the YOLO score significantly drops from 58.13 to 22.75 (red in Fig. 6). In contrast, our REGROUND, (blue in Fig. 6), demonstrates a notably superior trade-off between textual and spatial grounding. Specifically, with $\\gamma$ set to 1.0, REGROUND already achieves a CLIP score of 31.29, accounting for $70.25\\%$ of GLIGEN's total improvement in CLIP score when $\\gamma$ is reduced from 1.0 to 0.1. Despite this significant increase in CLIP score, the YOLO score remains largely unchanged, marking 56.96 which represents only a $3.31\\%$ decrease in the range of YOLO score variation for GLIGEN when $\\gamma$ is adjusted from 1.0 to 0.1. Moreover, when varying the $\\gamma$ , the plot for REGROUND (blue) is constantly on the upper right", + "bbox": [ + 214, + 657, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "ReGround: Improving Textual and Spatial Grounding at No Cost", + "bbox": [ + 294, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e4a8403d8740b291024a3725deb0df83fa19d27b12e2ccda242132828f68f2c2.jpg", + "image_caption": [ + "\"A young child looking at a birthday cupcake\"" + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 370, + 252 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/69cd149967164ffc044a8dd9f130cf352eb13e9b08d9609a1f8863e110fa12ef.jpg", + "image_caption": [ + "(a) GLIGEN" + ], + "image_footnote": [], + "bbox": [ + 372, + 145, + 496, + 252 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b025b7e851fd264919931122ef1ef22f9f486b5aa6772388174f4e7efeb0c76a.jpg", + "image_caption": [ + "Fig. 7: Generated images from the text prompt and bounding boxes from the MS-COCO-2017 (left of each column) and our COCO-Drop (right of each column). While GLIGEN [28] fails to generate \"a birthday cupcake\" when the corresponding bounding box is removed, our REGROUND successfully generates a cupcake on the table." + ], + "image_footnote": [], + "bbox": [ + 504, + 143, + 630, + 252 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5063241389f13d6b28a179318402b230ce64fea0ce326c25129d71e4b4c5b1b2.jpg", + "image_caption": [ + "(b) ReGround" + ], + "image_footnote": [], + "bbox": [ + 632, + 145, + 754, + 252 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "side of GLIGEN (red), signifying a more advantageous trade-off across varying $\\gamma$ . The same pattern is observed in MS-COCO-2017, where our REGROUND achieves $68.33\\%$ of the increase in CLIP score of GLIGEN while only compromising YOLO score by $2.62\\%$ compared to the decrease for GLIGEN.", + "bbox": [ + 212, + 380, + 784, + 439 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Fig. 6-(c) further shows a quantitative comparison on the Counting subset of the newly generated NSR-1K-GPT benchmark. The plot reveals a consistent trend with the MS-COCO datasets. By reducing $\\gamma$ from 1.0 to 0.1, GLIGEN's CLIP score is increased from 32.46 to 33.67, while the YOLO score is decreased from 65.36 to 26.38. In contrast, when $\\gamma = 1.0$ , REGROUND achieves a CLIP score of 33.20, which is equal to $61.16\\%$ of GLIGEN's total improvement in CLIP score, while the compromise in YOLO score is equal to only $3.69\\%$ of the total decrease in the YOLO score of GLIGEN from $\\gamma = 1.0$ to $\\gamma = 0.1$ . Moreover, a comparison on the Spatial subset of NSR-1K-GPT is provided in the Supplementary (Sec. S2). These results highlight that the advantage of our REGROUND holds robustly for the realistic image captions provided in the MS-COCO [30], as well as for diverse text prompts generated by GPT-4 [34].", + "bbox": [ + 212, + 441, + 787, + 625 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Random Box Dropping. To further assess the extent of description omission in each method, we modify the MS-COCO-2017 dataset [30] to make the COCO-Drop dataset. In this version, the bounding boxes for $50\\%$ of the categories are randomly removed from each image, thereby preventing every entity described in the text prompt from being included within the bounding boxes.", + "bbox": [ + 214, + 642, + 787, + 717 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Fig. 8 shows the quantitative comparison of REGROUND and GLIGEN on COCO-Drop. In this case, REGROUND shows a larger advantage over GLIGEN in CLIP score, obtaining a gap in CLIP score which is 1.57 times that of the original MS-COCO-2017 dataset before box dropping for $\\gamma = 1.0$ . Such a larger gap in CLIP score demonstrates that compared to GLIGEN, our REGROUND better reflects the text prompts even when some entities in the text prompt are not provided as a bounding box. Fig. 7 displays a representative example, where GLIGEN fails to generate a \"cupcake\" when its corresponding bounding", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "P.Y. Lee and M. Sung", + "bbox": [ + 271, + 114, + 421, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5a6aac1e76d2f7c1516044409b7028fd482dfe931d0d124edc1d7dd5239a3458.jpg", + "image_caption": [ + "Fig. 8: Comparison on the COCO-Drop Fig.9: Comparison of FID [16] on MS-dataset. COCO-2017 [30] dataset." + ], + "image_footnote": [], + "bbox": [ + 230, + 143, + 480, + 287 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0f84585542b5645eb60db586d84e92bd060a1cf141e26fd3b43f91ec8ec89a92.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 143, + 769, + 287 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "box is removed in COCO-Drop, whereas our REGROUND robustly generates the cupcake even when it is not provided as a bounding box.", + "bbox": [ + 212, + 359, + 787, + 392 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Image Quality. Fig. 9 displays the relationship between YOLO score [52] and FID [16] for each method on MS-COCO-2017. Note that the FID of REGROUND is constantly lower than that of GLIGEN [28], meaning that our network rewiring also results in higher image quality and diversity.", + "bbox": [ + 212, + 407, + 787, + 469 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "User Study. We conducted a user study to compare GLIGEN and our ReGround in terms of faithfulness to input text prompts. We used GPT-4 [34] to generate 100 prompts each containing two different objects, along with a bounding box for each object. Participants were given the text prompt along with two images—one from each method—and asked to choose the image that \"better includes all the objects from the prompt.\" Among the 92 out of 100 participants who passed the vigilance tests, our REGROUND surpassed GLIGEN, with a preference rate of $70.05\\%$ compared to $29.95\\%$ . Further details on the user study are provided in the Supplementary (Sec. S1).", + "bbox": [ + 212, + 484, + 787, + 623 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "PickScore. We further compare the PickScore [26] of GLIGEN [28] and our REGROUND given each input text prompt. On MS-COCO-2017, REGROUND is preferred over GLIGEN by $55.66\\%$ to $44.34\\%$ , and on COCO-Drop, REGROUND is preferred by $57.57\\%$ to $42.43\\%$ .", + "bbox": [ + 214, + 636, + 787, + 699 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6.4 Impact of REGROUND as a Backbone", + "text_level": 1, + "bbox": [ + 214, + 722, + 571, + 738 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We demonstrate that applying our rewiring of attention modules can also improve text-image alignment in other layout-guided generation methods that use GLIGEN as a backbone. For instance, BoxDiff [54] is a notable example that uses GLIGEN as its foundation and improves spatial grounding with respect to the bounding boxes by leveraging cross-attention maps as additional spatial cues in a zero-shot manner. Our network rewiring can also be combined with", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "ReGround: Improving Textual and Spatial Grounding at No Cost", + "bbox": [ + 294, + 114, + 732, + 130 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/03b98e25b383ddc4c43d3c259ddeac0171fa48cc95eb2baa45c1405bded64e84.jpg", + "image_caption": [ + "(a) NSR-1K-GPT-Counting" + ], + "image_footnote": [], + "bbox": [ + 217, + 146, + 395, + 246 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/85395662c0e6ddc2922f0b93bfbb39919cb38d383df80e1f937f68d37604ba91.jpg", + "image_caption": [ + "Fig. 10: Comparison of applying BoxDiff [54] on GLIGEN [28] and our REGROUND, respectively. (a) and (b) show that our REGROUND further improves the grounding quality of BoxDiff on NSR-1K-GPT datasets. (c) While BoxDiff with GLIGEN (left) also shows description omission—omitting \"beautiful sunset\" from the text prompt—BoxDiff with our REGROUND contains the sunset in the final image (right)." + ], + "image_footnote": [], + "bbox": [ + 395, + 146, + 563, + 247 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/87b71ec74bd2aa3de3c5b1f08b2ab2b18cc6938d4e601f7375248024dac082c6.jpg", + "image_caption": [ + "(b) NSR-1K-GPT-Spatial", + "(c) BoxDiff Outputs" + ], + "image_footnote": [], + "bbox": [ + 566, + 145, + 782, + 247 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "the zero-shot guidance of BoxDiff. Fig. 10 illustrates the results on the NSR-1K-GPT datasets (a) when BoxDiff uses GLIGEN as the base, and (b) when it uses our REGROUND, the rewired GLIGEN, as the base. It depicts that for the same range of spatial grounding accuracies, REGROUND obtains noticeably higher textual grounding (i.e. CLIP score [15]). Also, as shown in Fig. 10-(c), our network rewiring allows for a more detailed description to accurately appear in the final image, both for the entities in the bounding boxes (\"truck\") and the entities that are given as a text prompt (\"sunset\").", + "bbox": [ + 212, + 369, + 787, + 491 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 511, + 357, + 526 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We have demonstrated that a simple network rewiring of attention modules, making the gated self-attention and cross-attention parallel, surprisingly improves the trade-off between textual and spatial grounding at no additional cost — without introducing any new parameters, any fine-tuning of the network, or any changes in generation time and memory. Using the pretrained GLIGEN [28], which was trained with the original sequential architecture of the two attention modules, the reconfiguration at inference time has led to achieving higher CLIP scores, indicating the noticeable improvement in textual grounding accuracy. Moreover, our REGROUND improves the textual grounding while preserving the spatial grounding accuracy - achieving $70.25\\%$ and $68.33\\%$ of GLIGEN's total improvement with the scheduled sampling in CLIP score while compromising YOLO score only $3.31\\%$ and $2.62\\%$ for the MS-COCO-2014 and MS-COCO-2017 datasets, respectively. We also showcased that this simple yet effective solution for the textual-spatial grounding trade-off can lead to improvements in diverse frameworks using GLIGEN as a base.", + "bbox": [ + 212, + 540, + 787, + 767 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Supplementary. Due to limited space, we provide the following contents in the Supplementary: details on the evaluation setup (Sec. S1), additional quantitative (Sec. S2) and qualitative (Sec. S4) comparisons, and more results of REGROUND as a backbone for other layout-guided generation methods (Sec. S3).", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "P.Y. Lee and M. Sung", + "bbox": [ + 271, + 114, + 421, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by NRF grant (RS-2023-00209723), IITP grants (RS-2019-II190075, RS-2022-II220594, RS-2023-00227592, RS-2024-00399817), and Alchymist Project Program (RS-2024-00423625) funded by the Korean government (MSIT and MOTIE), and grants from the DRB-KAIST SketchTheFuture Research Center, Naver-intel, Adobe Research, Hyundai NGV, KT, and Samsung Electronics.", + "bbox": [ + 212, + 174, + 787, + 266 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 287, + 321, + 304 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: NeurIPS (2022)", + "2. Avrahami, O., Hayes, T., Gafni, O., Gupta, S., Taigman, Y., Parikh, D., Lischinski, D., Fried, O., Yin, X.: Spatext: Spatio-textual representation for controllable image generation. In: CVPR (2023)", + "3. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., et al.: edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022)", + "4. Bansal, A., Chu, H.M., Schwarzschild, A., Sengupta, S., Goldblum, M., Geiping, J., Goldstein, T.: Universal guidance for diffusion models. In: ICLR (2024)", + "5. Bar-Tal, O., Yariv, L., Lipman, Y., Dekel, T.: Multidiffusion: Fusing diffusion paths for controlled image generation. In: ICML (2023)", + "6. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., Manassra, W., Dhariwal, P., Chu, C., Jiao, Y., Ramesh, A.: Improving image generation with better captions (2023), https://cdn.openai.com/papers/dall-e-3.pdf", + "7. Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., Zheng, Y.: Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In: ICCV (2023)", + "8. Chen, M., Laina, I., Vedaldi, A.: Training-free layout control with cross-attention guidance. In: WACV (2024)", + "9. Chen, W.G., Spiridonova, I., Yang, J., Gao, J., Li, C.: Llava-interactive: An all-in-one demo for image chat, segmentation, generation and editing. arXiv preprint arXiv:2311.00571 (2023)", + "0. Cheng, J., Liang, X., Shi, X., He, T., Xiao, T., Li, M.: Layoutdiffuse: Adapting foundational diffusion models for layout-to-image generation. arXiv (2023)", + "1. Couairon, G., Careil, M., Cord, M., Lathuilière, S., Verbeek, J.: Zero-shot spatial layout conditioning for text-to-image diffusion models. In: ICCV (2023)", + "2. Feng, W., Zhu, W., Fu, T.j., Jampani, V., Akula, A., He, X., Basu, S., Wang, X.E., Wang, W.Y.: LayoutGPT: Compositional visual planning and generation with large language models. In: NeurIPS (2023)", + "3. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. In: ICLR (2022)", + "4. Herzig, R., Bar, A., Xu, H., Chechik, G., Darrell, T., Globerson, A.: Learning canonical representations for scene graph to image generation. In: ECCV (2020)", + "5. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: CLIPScore: a reference-free evaluation metric for image captioning. In: EMNLP (2021)" + ], + "bbox": [ + 225, + 316, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "ReGround: Improving Textual and Spatial Grounding at No Cost", + "bbox": [ + 294, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: GANs trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2018)", + "17. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)", + "18. Hong, S., Yang, D., Choi, J., Lee, H.: Inferring semantic layout for hierarchical text-to-image synthesis. In: CVPR (2018)", + "19. Hong, S., Lee, G., Jang, W., Kim, S.: Improving sample quality of diffusion models using self-attention guidance. In: ICCV (2023)", + "20. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (2021). https://doi.org/10.5281/zenodo.5143773, https://doi.org/10.5281/zenodo.5143773", + "21. Johnson, J., Gupta, A., Fei-Fei, L.: Image generation from scene graphs. In: CVPR (2018)", + "22. Karras, T., Aittala, M., Laine, S., Härkönen, E., Hellsten, J., Lehtinen, J., Aila, T.: Alias-free generative adversarial networks. In: NeurIPS (2021)", + "23. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (2019)", + "24. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of StyleGAN. In: CVPR (2020)", + "25. Kim, Y., Lee, J., Kim, J.H., Ha, J.W., Zhu, J.Y.: Dense text-to-image generation with attention modulation. In: ICCV (2023)", + "26. Kirstain, Y., Polyak, A., Singer, U., Matiana, S., Penna, J., Levy, O.: Pick-a-pic: An open dataset of user preferences for text-to-image generation. NeurIPS (2024)", + "27. Li, W., Zhang, P., Zhang, L., Huang, Q., He, X., Lyu, S., Gao, J.: Object-driven text-to-image synthesis via adversarial training. In: CVPR (2019)", + "28. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: GLIGEN: Open-set grounded text-to-image generation. In: CVPR (2023)", + "29. Li, Z., Wu, J., Koh, I., Tang, Y., Sun, L.: Image synthesis from layout with locality-aware mask adaption. In: CVPR (2021)", + "30. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014)", + "31. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)", + "32. Ma, W.D.K., Lewis, J.P., Lahiri, A., Leung, T., Kleijn, W.B.: Directed diffusion: Direct control of object placement through attention guidance. arXiv preprint arXiv:2302.13153 (2023)", + "33. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)", + "34. OpenAI: Chatgpt, https://chat.openai.com/", + "35. Parmar, G., Kumar Singh, K., Zhang, R., Li, Y., Lu, J., Zhu, J.Y.: Zero-shot image-to-image translation. In: ACM TOG (2023)", + "36. Phung, Q., Ge, S., Huang, J.B.: Grounded text-to-image synthesis with attention refocusing. arXiv preprint arXiv:2306.05427 (2023)", + "37. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: SDXL: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "P.Y. Lee and M. Sung", + "bbox": [ + 271, + 114, + 421, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "38. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021)", + "39. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)", + "40. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: ICML (2021)", + "41. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)", + "42. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015)", + "43. Si, C., Huang, Z., Jiang, Y., Liu, Z.: Freeu: Free lunch in diffusion u-net. arXiv preprint arXiv:2309.11497 (2023)", + "44. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models (2021)", + "45. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: ICLR (2021)", + "46. Sun, W., Wu, T.: Image synthesis from reconfigurable layout and style. In: ICCV (2019)", + "47. Sylvain, T., Zhang, P., Bengio, Y., Hjelm, R.D., Sharma, S.: Object-centric image generation from layouts. In: AAAI (2021)", + "48. Tancik, M., Srinivasan, P.P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U., Ramamoorthi, R., Barron, J.T., Ng, R.: Fourier features let networks learn high frequency functions in low dimensional domains. In: NeurIPS (2020)", + "49. Tumanyan, N., Geyer, M., Bagon, S., Dekel, T.: Plug-and-play diffusion features for text-driven image-to-image translation. In: CVPR (2023)", + "50. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)", + "51. Voynov, A., Aberman, K., Cohen-Or, D.: Sketch-guided text-to-image diffusion models. ACM TOG (2023)", + "52. Wang, C.Y., Bochkovskiy, A., Liao, H.Y.M.: YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In: CVPR (2023)", + "53. Xiao, J., Li, L., Lv, H., Wang, S., Huang, Q.: R&b: Region and boundary aware zero-shot grounded text-to-image generation. In: ICLR (2024)", + "54. Xie, J., Li, Y., Huang, Y., Liu, H., Zhang, W., Zheng, Y., Shou, M.Z.: Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In: ICCV (2023)", + "55. Yang, Z., Wang, J., Gan, Z., Li, L., Lin, K., Wu, C., Duan, N., Liu, Z., Liu, C., Zeng, M., et al.: Reco: Region-controlled text-to-image generation. In: CVPR (2023)", + "56. Yang, Z., Liu, D., Wang, C., Yang, J., Tao, D.: Modeling image composition for complex scene generation. In: CVPR (2022)", + "57. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023)", + "58. Zhao, B., Meng, L., Yin, W., Signal, L.: Image generation from layout. In: CVPR (2019)", + "59. Zhao, P., Li, H., Jin, R., Zhou, S.K.: Loco: Locally constrained training-free layout-to-image synthesis. arXiv preprint arXiv:2311.12342 (2023)", + "60. Zheng, G., Zhou, X., Li, X., Qi, Z., Shan, Y., Li, X.: Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In: CVPR (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 825 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "ReGround: Improving Textual and Spatial Grounding at No Cost", + "bbox": [ + 294, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/ReGround_ Improving Textual and Spatial Grounding at No Cost/e369c9ee-bed9-4dbc-92b4-a4d9da7bccb2_model.json b/2024/ReGround_ Improving Textual and Spatial Grounding at No Cost/e369c9ee-bed9-4dbc-92b4-a4d9da7bccb2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1b8a352ebaa407879427a874fcc1ebf37105e6c8 --- /dev/null +++ b/2024/ReGround_ Improving Textual and Spatial Grounding at No Cost/e369c9ee-bed9-4dbc-92b4-a4d9da7bccb2_model.json @@ -0,0 +1,3116 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.26, + 0.141, + 0.745, + 0.187 + ], + "angle": 0, + "content": "ReGround: Improving Textual and Spatial Grounding at No Cost" + }, + { + "type": "text", + "bbox": [ + 0.372, + 0.213, + 0.631, + 0.228 + ], + "angle": 0, + "content": "Phillip Y. Lee Minhyuk Sung" + }, + { + "type": "text", + "bbox": [ + 0.377, + 0.24, + 0.626, + 0.269 + ], + "angle": 0, + "content": "KAIST {phillip0701,mhsung}@kaist.ac.kr" + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.277, + 0.3, + 0.289 + ], + "angle": 0, + "content": "Layout" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.29, + 0.331, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.377, + 0.277, + 0.399, + 0.288 + ], + "angle": 0, + "content": "SD" + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.291, + 0.443, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.455, + 0.277, + 0.548, + 0.29 + ], + "angle": 0, + "content": "GLIGEN\\(_{\\gamma=1.0}\\)" + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.291, + 0.556, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.277, + 0.662, + 0.29 + ], + "angle": 0, + "content": "GLIGEN \\(\\gamma = 0.2\\)" + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.291, + 0.671, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.688, + 0.277, + 0.768, + 0.289 + ], + "angle": 0, + "content": "ReGround" + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.291, + 0.785, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.291, + 0.377, + 0.713, + 0.391 + ], + "angle": 0, + "content": "\"A low poly illustration of a man holding a cell phone in hand.\"" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.393, + 0.331, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.393, + 0.444, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.393, + 0.557, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.393, + 0.671, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.393, + 0.785, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.291, + 0.479, + 0.682, + 0.492 + ], + "angle": 0, + "content": "\"An elephant walking draped with a colorful blanket.\"" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.494, + 0.331, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.494, + 0.444, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.494, + 0.557, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.494, + 0.671, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.494, + 0.785, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.581, + 0.753, + 0.595 + ], + "angle": 0, + "content": "\"One ancient vase [...] the backdrop of a serene, moonlit desert landscape.\"" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.597, + 0.331, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.597, + 0.444, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.597, + 0.557, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.597, + 0.671, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.597, + 0.785, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.331, + 0.684, + 0.673, + 0.698 + ], + "angle": 0, + "content": "\"A train blowing smoke is coming down the tracks.\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.715, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Fig. 1: Comparison across Stable Diffusion (SD) [41], GLIGEN [28], and our REGROUND. SD (2nd column) can generate an image aligned with the input prompt (shown below each row), while it does not allow taking spatial constraints such as bounding boxes and labels. GLIGEN (3rd column) enables spatial grounding using gated self-attention, although it often disregards some descriptions in the input prompt due to a bias towards bounding box conditions. Such trends also occur when only activating gated self-attention for 0.2 fraction of the initial denoising steps (4th column). Our REGROUND (last column) resolves the issue of description omission while accurately reflecting the bounding box information." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.422, + 0.129 + ], + "angle": 0, + "content": "P.Y. Lee and M. Sung" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.148, + 0.744, + 0.358 + ], + "angle": 0, + "content": "Abstract. When an image generation process is guided by both a text prompt and spatial cues, such as a set of bounding boxes, do these elements work in harmony, or does one dominate the other? Our analysis of a pretrained image diffusion model that integrates gated self-attention into the U-Net reveals that spatial grounding often outweighs textual grounding due to the sequential flow from gated self-attention to cross-attention. We demonstrate that such bias can be significantly mitigated without sacrificing accuracy in either grounding by simply rewiring the network architecture, changing from sequential to parallel for gated self-attention and cross-attention. This surprisingly simple yet effective solution does not require any fine-tuning of the network but substantially reduces the trade-off between the two groundings. Our experiments demonstrate significant improvements from the original GLIGEN to the rewired version in the trade-off between textual grounding and spatial grounding. The project webpage is at https://re-ground.github.io." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.37, + 0.741, + 0.385 + ], + "angle": 0, + "content": "Keywords: Textual Grounding \\(\\cdot\\) Spatial Grounding \\(\\cdot\\) Network Rewiring" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.412, + 0.376, + 0.429 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.788, + 0.538 + ], + "angle": 0, + "content": "The emergence of diffusion models [17,44,45] has markedly propelled the field of text-to-image (T2I) generation forward, allowing users to generate high-quality images from text prompts. In a bid to further augment the creativity and controllability, recent efforts [2-5,8,11,25,28,32,36,54,55,57] have focused on enabling these models to understand and interpret spatial instructions, such as layouts [4,8,28,32,36,54,55], segmentation masks [2-5,11,25] and sketches [51,57]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.788, + 0.719 + ], + "angle": 0, + "content": "Among them, bounding boxes are extensively employed in downstream image generation tasks [4, 8, 28, 32, 36, 55]. GLIGEN [28] is a pioneering work in terms of enhancing existing T2I models with the capability to incorporate additional spatial cues in the form of bounding boxes. Its core component, gated self-attention, is a simple attention module [50] that is plugged into each U-Net [42] layer of a pretrained T2I model such as Stable Diffusion [41], and is trained to accurately position various entities in their designated areas. A notable advantage of GLIGEN is that the original parameters of the underlying model remain unchanged, inheriting the generative capability of the T2I model while introducing the novel functionality of spatial grounding using bounding boxes. This capability has been leveraged by numerous studies to facilitate high-quality, layout-guided image generation [9, 12, 31, 54]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "However, our analysis reveals that GLIGEN's integration of the gated self-attention into an existing T2I model is not optimal for blending new spatial guidance from bounding boxes with the original textual guidance. It often leads to the omission of specific details from the text prompts. For instance, in the first row and third column of Fig. 1, GLIGEN fails to reflect the description \"low poly illustration\" from the input text prompt. Also in the second row, a crucial detail in the text prompt, \"draped with a colorful blanket\", is neglected in the output image. We refer to this issue as description omission. Such outcomes imply" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.295, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "ReGround: Improving Textual and Spatial Grounding at No Cost" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.221 + ], + "angle": 0, + "content": "that the current architectural design of GLIGEN does not effectively harmonize the new spatial guidance and the text conditioning in the given T2I model. Considering the widespread applications of GLIGEN in various layout-based generation tasks [9,12,31,36,53,54,59], these limitations represent a significant bottleneck." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.785, + 0.299 + ], + "angle": 0, + "content": "To address the observed neglect of textual grounding in GLIGEN, we first analyze the root causes. Our investigation reveals that the issue arises from the sequential arrangement of the spatial grounding and textual grounding modules. Specifically, the output of the gated self-attention is directed to a cross-attention module in each layer of the U-Net architecture (Fig. 2-(b))." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.787, + 0.449 + ], + "angle": 0, + "content": "Building on this insight, we propose a straightforward yet impactful solution: network rewiring. This approach alters the relationship between the two grounding modules from sequential to parallel (Fig. 2-(c)). Remarkably, this network modification significantly reduces the grounding trade-off between textual and spatial groundings without necessitating any adjustments to the network parameters. Importantly, this rewiring does not require additional network training, extra parameters, or changes in computational load and time. Simply reconfiguring the attention modules of the pretrained GLIGEN, originally trained with the sequential architecture, during inference dramatically enhances performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.571 + ], + "angle": 0, + "content": "In our experiments on MS-COCO [30] and our newly introduced NSR-1K-GPT datasets, we demonstrate that rewiring the pretrained GLIGEN substantially reduces the trade-off between textual and spatial groundings. This is evidenced by the evaluation of text prompt alignment (measured using CLIP score [38], PickScore [26] and user study) and bounding box alignment (assessed by YOLO score [52]). Furthermore, we show that our rewiring also leads to better outcomes in other frameworks using GLIGEN as a backbone, including BoxDiff [54]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.597, + 0.388, + 0.613 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.631, + 0.594, + 0.646 + ], + "angle": 0, + "content": "2.1 Zero-Shot Guidance in Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The progress in diffusion models [17, 44, 45] has significantly elevated the capabilities of text-to-image (T2I) generation, resulting in foundation models [6, 37, 39-41] that exhibit remarkable generative performance. Leveraging the robust performance of these models, recent studies [3-5, 8, 11, 25, 28, 32, 36, 54, 55, 57] have introduced efficient guidance techniques designed to further improve the image generation process. Notably, numerous works [8, 28, 32, 36, 43, 54, 55] focus on the internal architecture (Fig. 2-(a)) of the denoising U-Net of Latent Diffusion Models (LDMs) [41], where self-attention and cross-attention modules are intertwined to facilitate inter-pixel communication and text conditioning. The self-attention of U-Net can be utilized to improve image quality [19] or facilitate image translation [49] and image editing tasks [7]. Since text conditions are integrated via cross-attention, the intermediate attention maps have been leveraged" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.422, + 0.129 + ], + "angle": 0, + "content": "P.Y. Lee and M. Sung" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "to improve text faithfulness [13] or enable spatial manipulation of the generation process [35]. Recently, FreeU [43] analyzed the contributions of the backbone and residuals of the U-Net and proposed a free-lunch strategy to enhance image quality: reweighting the backbone and residual features maps. In contrast to previous works that only deal with self- and cross-attention in standard LDMs, we introduce a method to enhance GLIGEN [28] by reconnecting its gated self-attention with the other attention modules, thereby achieving performance improvement in zero-shot without any tuning of the network parameters." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.287, + 0.54, + 0.302 + ], + "angle": 0, + "content": "2.2 Layout-Guided Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.309, + 0.789, + 0.489 + ], + "angle": 0, + "content": "The use of layouts, particularly in the form of bounding boxes, has become a popular intermediary to bridge the gap between textual inputs and the images generated [14, 18, 21, 27, 29, 46, 47, 56, 58]. Layout2Im [58] samples object latent codes from a normal distribution, eliminating the need to predict instance masks as done in prior works [18, 21]. LostGAN [46] controls the style of each object by devising an extension of the feature normalization layer used in StyleGAN [22-24], while OC-GAN [47] incorporates the spatial relationships between objects using a scene-graph representation. LAMA [29] introduces a mask adaptation module that mitigates the semantic ambiguity arising from overlaps in the input layout. While these developments have greatly improved user control over image generation, their applicability is confined to the categories found in the training data, such as those of the MS-COCO [30] dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.79, + 0.673 + ], + "angle": 0, + "content": "In contrast, recent studies [5, 8, 10, 28, 36, 54, 55, 60] have extended layout-guided image generation towards open-vocabulary, building on the advancements of foundational text-to-image (T2I) models [41]. Training-free approaches [3, 5, 8, 36, 54] aim to improve the spatial grounding of T2I models through straightforward guidance mechanisms. GLIGEN [28], on the other hand, introduces gated self-attention, which is injected into the U-Net architecture of the Latent Diffusion Model [41], and is trained to equip the underlying model with spatial grounding abilities. Given the simple architecture of GLIGEN and its robust grounding accuracy with the input bounding boxes, numerous studies [36, 53, 54, 59] build upon its framework and propose further refinements to increase performance. In this work, we identify and address a significant performance bottleneck in GLIGEN related to description omission and propose a simple yet effective solution." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.691, + 0.681, + 0.71 + ], + "angle": 0, + "content": "3 Background - Latent Diffusion Models [41]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Rombach et al. [41] proposed Latent Diffusion Model (LDM), a text-to-image (T2I) diffusion model with a U-Net as the noise prediction network. It is trained to generate an image from an input text prompt by predicting the noise \\(\\epsilon (\\mathbf{x}_t,t,c)\\) conditioned both on the timestep \\(t\\) and the text embedding \\(c\\). Each layer of LDM's U-Net consists of three core components: a convolutional residual block, followed by a self-attention (SA), and a cross-attention (CA) module (Fig. 2-(a)). In each \\(l\\)-th layer of the U-Net, its residual block first extracts intermediate visual features \\(F = (f_{1},\\dots,f_{N_{l}})^{T}\\) from the output of the previous layer." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.295, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "ReGround: Improving Textual and Spatial Grounding at No Cost" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.144, + 0.788, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.254, + 0.788, + 0.352 + ], + "angle": 0, + "content": "Fig. 2: Comparison between the U-Net architectures of (a) Latent Diffusion Model (LDM) [41], (b) GLIGEN [28] and (c) our REGROUND. From LDM, GLIGEN enables spatial grounding by injecting Gated Self-Attention before cross-attention, forming a sequential flow of them. Based on GLIGEN, our REGROUND changes the relationship of the two attention modules to become parallel, resulting in noticeable improvement in textual grounding while preserving the spatial grounding capability. (The residual block before self-attention is omitted.)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.788, + 0.459 + ], + "angle": 0, + "content": "The self-attention module then facilitates interaction between the features in \\( F \\). Subsequently, the cross-attention module enables the interaction between each visual feature \\( f_{i} \\) and the text embedding \\( c \\). Throughout this process, the output feature of the previous module is also forwarded through a residual connection, as illustrated in lines 4-5 and 7 of Alg. 1 and also in Fig. 2-(a)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.483, + 0.64, + 0.502 + ], + "angle": 0, + "content": "4 GLIGEN [28] and Description Omission" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.516, + 0.787, + 0.563 + ], + "angle": 0, + "content": "In this section, we review GLIGEN [28] and its key idea of employing gated self-attention for spatial grounding. Then, we present our key observations on the description omission issue that occurs due to the addition of gated self-attention." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.587, + 0.436, + 0.601 + ], + "angle": 0, + "content": "4.1 Gated Self-Attention" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Li et al. [28] propose a plug-in spatial grounding module, named gated self-attention, which adopts the gated attention mechanism [1] to equip a pretrained T2I diffusion model [41] with spatial grounding capabilities (Fig. 2-(b)). Given a set of bounding boxes and text labels for each of them, let \\( b_{i} \\) be the \\( xy \\)-coordinates of the \\( i \\)-th bounding box's top-left and bottom-right corners, and \\( p_i \\) be the corresponding text label. Then, the \\( i \\)-th grounding token is defined as \\( g_{i} := \\mathcal{G}(\\mathcal{T}(p_{i}),\\mathcal{F}(b_{i})) \\), where \\( \\mathcal{T}(\\cdot) \\) is a pretrained text encoder [20,38], \\( \\mathcal{F}(\\cdot) \\) is the Fourier embedding [33,48] and \\( \\mathcal{G}(\\cdot,\\cdot) \\) is a shallow MLP network that concatenates the two given embeddings, respectively. Given a set of grounding tokens \\( \\{g_i\\} \\), gated self-attention learns the self-attention among the unified feature set \\( (f_1,\\dots,f_{N_l},g_1,\\dots,g_M) \\) where \\( \\{f_i\\} \\) is the set of intermediate visual features in the \\( l \\)-th layer of U-Net, and \\( M \\) is the number of bounding boxes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "As shown in Fig. 2-(b), gated self-attention receives the output of the self-attention along with the residual features as its input and forwards the output features to the cross-attention module. By incorporating gated self-attention into" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.422, + 0.129 + ], + "angle": 0, + "content": "P.Y. Lee and M. Sung" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.236 + ], + "angle": 0, + "content": "each layer of the U-Net, the model enables the placement of the entity specified in the text label \\( p_i \\) at the location indicated by the bounding box \\( b_i \\). Note that the integration of gated self-attention does not require training the network from scratch or fine-tuning it, but can be accomplished simply by training the gated self-attention parameters while keeping all other parameters in the backbone model frozen." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.24, + 0.788, + 0.288 + ], + "angle": 0, + "content": "Alg. 1 shows the pseudocode of the U-Net forward-pass including the plug-in of gated self-attention in line 6. Note that \\(\\beta_{t}\\) is set to 1 for GLIGEN. If \\(\\beta_{t} = 0\\), the algorithm is identical to that of LDM [41]." + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.323, + 0.692, + 0.338 + ], + "angle": 0, + "content": "Algorithm 1: Noise Prediction U-Net with Gated Self-Attention." + }, + { + "type": "algorithm", + "bbox": [ + 0.202, + 0.34, + 0.788, + 0.536 + ], + "angle": 0, + "content": "Parameters: \\(\\beta_{t}\\) // Weight for GSA. Inputs: \\(\\mathbf{x}_t, c, \\{g_i\\}_{i=0\\dots N-1}\\); // Noisy data at timestep \\(t\\), text condition, and grounding tokens \nOutputs: \\(\\epsilon_t\\); // Noise at timestep \\(t-1\\). \n1 Function U-Net \\((\\mathbf{x}_t, c, \\{g_i\\})\\): \n2 \\(F \\gets \\mathbf{x}_t\\) \n3 for \\(i = 0, \\dots, L-1\\) do \n4 \\(\\begin{array}{c} F_{\\mathrm{RS}} \\gets \\mathrm{Conv}(F) + F; \\\\ F_{\\mathrm{SA}} \\gets \\mathrm{SA}(F_{\\mathrm{RS}}) + F_{\\mathrm{RS}}; \\\\ F_{\\mathrm{GSA}} \\gets \\beta_t \\cdot \\mathrm{GSA}(F_{\\mathrm{SA}}, \\{g_i\\}) + F_{\\mathrm{SA}}; \\\\ F \\gets \\mathrm{CA}(F_{\\mathrm{GSA}}, c) + F_{\\mathrm{GSA}}; \\end{array}\\) // Residual block. \n5 // Self-Attention module. \n6 \n7 \n8 \\(\\epsilon_t \\gets F;\\) \n9 return \\(\\epsilon_t\\);" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.606, + 0.439, + 0.622 + ], + "angle": 0, + "content": "4.2 Description Omission" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Despite its high accuracy in spatial grounding, GLIGEN [28] frequently struggles to capture essential attributes specified in the input text prompt. As illustrated in Fig. 3, the leftmost image shows \"a person\" and \"a skateboard\" accurately placed in their designated regions. However, a critical detail from the input text prompt, \"black and white photography\", is absent in the output image. This discrepancy often emerges when the input comprises distinct but equally important descriptions regarding the image, presented through text prompts and bounding boxes. Such omissions not only fail to convey the stylistic intent of the image but also tend to overlook significant objects mentioned within the text prompt. Additional examples of this problem are showcased in Fig. 1, where the second row demonstrates the absence of a \"blanket\" in the generated image, a key element from the text prompt. This limitation significantly hampers GLIGEN's fidelity to user-provided text prompts, a challenge we term as description omission." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.295, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "ReGround: Improving Textual and Spatial Grounding at No Cost" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.267, + 0.788, + 0.366 + ], + "angle": 0, + "content": "Fig. 3: (a) Images generated by GLIGEN [28] with varying activation duration of gated self-attention \\(\\gamma\\) in scheduled sampling (Sec. 5.1). The red words in the text prompt denote the words used as labels of the input bounding boxes. Note that for GLIGEN to reflect the underlined description in the text prompt in the final image, \\(\\gamma\\) must be decreased to 0.1, which compromises spatial grounding accuracy. (b) In contrast, our REGROUND reflects the underlined phrase even when \\(\\gamma = 1.0\\), therefore achieving high accuracy in both textual and spatial grounding." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.394, + 0.644, + 0.412 + ], + "angle": 0, + "content": "5 ReGround: Rewiring Attention Modules" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.426, + 0.788, + 0.548 + ], + "angle": 0, + "content": "Gated self-attention and cross-attention each play a crucial role in enabling spatial and textual groundings, by taking bounding boxes and text prompts as inputs, respectively. To tackle the issue of description omission, we first examine the impact of attention modules on the groundings they do not address: the effect of gated self-attention on textual grounding (Sec. 5.1), and the influence of cross-attention on spatial grounding (Sec. 5.2). Building on this analysis, we propose an approach for network reconfiguration, modifying the connections among self-attention, gated self-attention, and cross-attention modules (Sec. 5.3)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.57, + 0.71, + 0.586 + ], + "angle": 0, + "content": "5.1 Impact of Gated Self-Attention on Textual Grounding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.788, + 0.703 + ], + "angle": 0, + "content": "As the issue of description omission arises due to the newly added gated self-attention in GLIGEN [28], we first attempt to mitigate the impact of gated self-attention by using scheduled sampling [28], activating gated self-attention only in a few initial steps of the denoising process. This approach is inspired by the observation that the coarse structure of the final image is established within the first few denoising steps. The scheduling is applied by setting the weight of gated self-attention \\(\\beta_{t}\\) (line 6 of Alg. 1) as" + }, + { + "type": "equation", + "bbox": [ + 0.42, + 0.713, + 0.787, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\beta_ {t} = \\left\\{ \\begin{array}{l l} 1 & (t \\leq \\gamma \\cdot T) \\\\ 0 & (t > \\gamma \\cdot T), \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.764, + 0.787, + 0.794 + ], + "angle": 0, + "content": "where \\(\\gamma \\in [0,1]\\) represents the fraction of the initial denoising steps to activate gated self-attention." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Fig. 3-(a) shows an example of generated images while incrementally adjusting \\(\\gamma\\) from 1.0 to 0.0. As \\(\\gamma\\) is reduced from 1.0 to 0.0, the details specified in the text prompt, \"a black and white photograph\", begin to be reflected starting" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.422, + 0.129 + ], + "angle": 0, + "content": "P.Y. Lee and M. Sung" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "at \\(\\gamma = 0.1\\), demonstrating that longer activation of gated self-attention may interfere with the alignment of the output image with the text prompt. However, as gated self-attention is activated for shorter durations, the spatial grounding diminishes, as shown in the objects' reduced alignment with the input bounding boxes. This phenomenon illustrates the inherent trade-off between spatial and textual grounding, which cannot be resolved by controlling the duration of gated self-attention activation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.275, + 0.664, + 0.291 + ], + "angle": 0, + "content": "5.2 Impact of Cross-Attention on Spatial Grounding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.301, + 0.784, + 0.375 + ], + "angle": 0, + "content": "We also investigate whether cross-attention has influence on spatial grounding. For this, we conduct a toy experiment by removing cross-attention modules in GLIGEN [28], allowing the output of the gated self-attention to be directly passed to the next layer of the U-Net. This modification is equivalent to changing line 7 of Alg. 1 to \\( F \\gets F_{GSA} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.376, + 0.788, + 0.482 + ], + "angle": 0, + "content": "The results are displayed in Fig. 4. Note that, while the appearance of the background and objects changes, the silhouettes of the cat (left) and the individuals (right) remain precisely positioned within their respective bounding boxes without cross-attention. This observation indicates that while gated self-attention that is performed before cross-attention may compromise textual grounding, cross-attention that processes the output of gated self-attention does not affect spatial grounding." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.506, + 0.773, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.601, + 0.788, + 0.657 + ], + "angle": 0, + "content": "Fig. 4: Comparison of the output of GLIGEN [28] with and without cross-attention. While the absence of cross-attention reduces realism and quality of the image, the silhouette of objects remains grounded within the given bounding boxes, as shown in the third column of each case." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.71, + 0.651, + 0.726 + ], + "angle": 0, + "content": "5.3 Network Rewiring: From Sequential to Parallel" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Building on the analyses above, we propose a simple yet effective modification to the grounding mechanism, changing the relationship between gated self-attention and cross-attention from sequential to parallel. This change eliminates the placement of gated self-attention before cross-attention, thus preventing the reduction of text grounding caused by gated self-attention. Moreover, in this parallel arrangement, the preservation of spatial grounding is assured, as gated self-attention for spatial grounding does not require subsequent cross-attention." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.295, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "ReGround: Improving Textual and Spatial Grounding at No Cost" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "Specifically, recall that in GLIGEN [28], the output of gated self-attention is added to the residual from self-attention, which is then passed to the cross-attention module as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.388, + 0.196, + 0.619, + 0.219 + ], + "angle": 0, + "content": "\\[\nF _ {G S A} \\gets \\underbrace {\\mathrm {G S A} (F _ {S A} , \\{g _ {i} \\})} + F _ {S A};\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.456, + 0.221, + 0.564, + 0.232 + ], + "angle": 0, + "content": "\\[\ns p a t i a l g r o u n d i n g\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.409, + 0.228, + 0.785, + 0.256 + ], + "angle": 0, + "content": "\\[\nF \\leftarrow \\underbrace {\\operatorname {C A} \\left(F _ {G S A} , c\\right)} + F _ {G S A}; \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.448, + 0.258, + 0.558, + 0.27 + ], + "angle": 0, + "content": "\\[\n\\text {t e x t u a l g r o u n d i n g}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.275, + 0.784, + 0.304 + ], + "angle": 0, + "content": "We propose to transform this sequence grounding pipeline into two parallel processes as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.312, + 0.785, + 0.334 + ], + "angle": 0, + "content": "\\[\nF \\leftarrow \\underbrace {\\operatorname {G S A} \\left(F _ {S A} , \\{g _ {i} \\}\\right)} + \\underbrace {\\operatorname {C A} \\left(F _ {S A} , c\\right)} + \\underbrace {F _ {S A}}; \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.335, + 0.674, + 0.347 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{c c c} \\text {s p a t i a l g r o u n d i n g} & \\text {t e x t u a l g r o u n d i n g} & \\text {r e s i d u a l} \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.353, + 0.784, + 0.412 + ], + "angle": 0, + "content": "Refer to Fig. 2 for the visualization of network architecture changes \\(((\\mathrm{b})\\rightarrow\\) (c)). This network rewiring is feasible because the input to gated self-attention remains unchanged, while the input to cross-attention shifts to \\(F_{SA}\\), for which it was originally designed in the context of Latent Diffusion Models [41]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.785, + 0.517 + ], + "angle": 0, + "content": "It is important to note that the modification is effective even when applied to the pretrained GLIGEN, which was trained with the sequential structure of the attention modules. Therefore, our rewiring does not require any additional training or fine-tuning, introduces no extra parameters, and does not affect computation time or memory usage during the generation process. The only requirement is the simple reconfiguration of the attention modules at inference time." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.538, + 0.376, + 0.554 + ], + "angle": 0, + "content": "6 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.784, + 0.626 + ], + "angle": 0, + "content": "In this section, we show the effectiveness of our REGROUND by evaluating the spatial grounding on existing layout-caption datasets [12, 30] and the textual grounding on text-image alignment metrics [26, 38]. We use the official GLI-GEN [28] checkpoint which is trained based on Stable Diffusion v1.4 [41]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.644, + 0.335, + 0.657 + ], + "angle": 0, + "content": "6.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.663, + 0.784, + 0.725 + ], + "angle": 0, + "content": "MS-COCO. We use the validations sets of both MS-COCO-2014 and MS-COCO-2017 datasets [30]. Each dataset provides image-captions pairs and the \\(xy\\)-coordinates of bounding boxes along with their corresponding object categories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.785, + 0.841 + ], + "angle": 0, + "content": "NSR-1K-GPT. We also use the NSR-1K benchmark [12] for evaluation. Based on each subset of NSR-1K—Counting and Spatial—we develop a new benchmark, NSR-1K-GPT, augmenting each original caption in NSR-1K using GPT-4 [34]. The instructions for augmentation are to (i) elaborate on the descriptions of each mentioned entity and (ii) provide additional details about the background of the image. More details on the evaluation datasets are provided in the Supplementary (Sec. S1)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.422, + 0.129 + ], + "angle": 0, + "content": "P.Y. Lee and M. Sung" + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.144, + 0.3, + 0.155 + ], + "angle": 0, + "content": "Layout" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.156, + 0.329, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.144, + 0.397, + 0.155 + ], + "angle": 0, + "content": "SD" + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.156, + 0.44, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.452, + 0.144, + 0.546, + 0.156 + ], + "angle": 0, + "content": "GLIGEN\\(_{\\gamma = 1.0}\\)" + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.156, + 0.553, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.144, + 0.658, + 0.156 + ], + "angle": 0, + "content": "GLIGEN\\(_{\\gamma=0.2}\\)" + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.156, + 0.665, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.144, + 0.764, + 0.155 + ], + "angle": 0, + "content": "ReGround" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.156, + 0.779, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.242, + 0.707, + 0.255 + ], + "angle": 0, + "content": "\"A woman in a white shirt standing in front of a fence smiling.\"" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.256, + 0.329, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.256, + 0.44, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.256, + 0.553, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.256, + 0.666, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.669, + 0.256, + 0.779, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.306, + 0.343, + 0.696, + 0.357 + ], + "angle": 0, + "content": "\"One spotted cow in the photo, [...] under a clear blue sky.\"" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.359, + 0.329, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.359, + 0.44, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.359, + 0.553, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.359, + 0.666, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.669, + 0.359, + 0.779, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.446, + 0.648, + 0.459 + ], + "angle": 0, + "content": "\"A foreign airline status board with a clock.\"" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.46, + 0.329, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.46, + 0.44, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.46, + 0.553, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.46, + 0.666, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.669, + 0.46, + 0.779, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.304, + 0.548, + 0.696, + 0.562 + ], + "angle": 0, + "content": "\"A palm tree shower curtain is shown in a small bathroom.\"" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.564, + 0.33, + 0.649 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.564, + 0.44, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.564, + 0.553, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.564, + 0.666, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.669, + 0.564, + 0.779, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.348, + 0.651, + 0.652, + 0.665 + ], + "angle": 0, + "content": "\"A view of a living room with large windows.\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.682, + 0.788, + 0.807 + ], + "angle": 0, + "content": "Fig. 5: Qualitative comparisons. Stable Diffusion (SD, 2nd column) generates images that align with the given text descriptions, including the underlined phrase in each row, but cannot take bounding boxes as input. GLIGEN (3rd column) creates images that match the input layouts but suffers from description omission, failing to reflect the underlined descriptions. Scheduled sampling strategy (4th column) can partially address this issue (for instance, in the 5th row, where \"window\" appears in the room), but it results in a noticeable decline in spatial accuracy (as seen in the 1st row, where the tie is not generated). In contrast, our method (last column) accurately incorporates the underlined text descriptions while maintaining precise spatial representation." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.295, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "ReGround: Improving Textual and Spatial Grounding at No Cost" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.145, + 0.406, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.257, + 0.255, + 0.376, + 0.267 + ], + "angle": 0, + "content": "(a) MS-COCO-2014" + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.145, + 0.593, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.447, + 0.255, + 0.566, + 0.267 + ], + "angle": 0, + "content": "(b) MS-COCO-2017" + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.145, + 0.787, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.619, + 0.255, + 0.776, + 0.268 + ], + "angle": 0, + "content": "(c) NSR-1K-GPT-Counting" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.279, + 0.789, + 0.351 + ], + "angle": 0, + "content": "Fig. 6: Comparisons on MS-COCO [30] and NSR-1K-GPT. Each plot shows the relationship between textual grounding (i.e. CLIP score [15]) and spatial grounding (i.e. YOLO score [52]) accuracy of GLIGEN [28] and our method. Note that the plot of our REGROUND is positioned in the top-right quadrant relative to GLIGEN, signifying that it alleviates the inherent trade-off between textual and spatial grounding." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.378, + 0.42, + 0.392 + ], + "angle": 0, + "content": "6.2 Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.402, + 0.787, + 0.463 + ], + "angle": 0, + "content": "- YOLO score: Spatial grounding accuracy is assessed usingYOLO score [52]. We employYOLOv7 [52] to detect objects in each generated image and compute the average precision (AP) based on the ground truth bounding box annotations from MS-COCO [30]." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.463, + 0.786, + 0.478 + ], + "angle": 0, + "content": "- CLIP score: Textual grounding accuracy is assessed using CLIP score [15]." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.478, + 0.699, + 0.492 + ], + "angle": 0, + "content": "- FID: Image quality and diversity are evaluated using FID [16]." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.492, + 0.787, + 0.552 + ], + "angle": 0, + "content": "- User Study and PickScore: We conduct a user study to assess human preferences for the generated images based on each input text prompt. Additionally, we use PickScore [26], a human preference predictor, to further analyze the results." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.402, + 0.787, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.575, + 0.486, + 0.59 + ], + "angle": 0, + "content": "6.3 Comparison with GLIGEN" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.599, + 0.787, + 0.658 + ], + "angle": 0, + "content": "Textual-Spatial Grounding Trade-off. We first examine the trade-off between textual and spatial groundings for both GLIGEN [28] and our REGROUND, the rewired version of GLIGEN, while varying the scheduled sampling parameter \\(\\gamma\\) from 1.0 to 0.1." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Fig. 6-(a), (b) present the graphs of CLIP score [15] and YOLO score [52] measured on the MS-COCO datasets [30]. In MS-COCO-2014, when reducing \\(\\gamma\\) from 1.0 to 0.1, the CLIP score of GLIGEN varies from 30.44 to 31.65, while the YOLO score significantly drops from 58.13 to 22.75 (red in Fig. 6). In contrast, our REGROUND, (blue in Fig. 6), demonstrates a notably superior trade-off between textual and spatial grounding. Specifically, with \\(\\gamma\\) set to 1.0, REGROUND already achieves a CLIP score of 31.29, accounting for \\(70.25\\%\\) of GLIGEN's total improvement in CLIP score when \\(\\gamma\\) is reduced from 1.0 to 0.1. Despite this significant increase in CLIP score, the YOLO score remains largely unchanged, marking 56.96 which represents only a \\(3.31\\%\\) decrease in the range of YOLO score variation for GLIGEN when \\(\\gamma\\) is adjusted from 1.0 to 0.1. Moreover, when varying the \\(\\gamma\\), the plot for REGROUND (blue) is constantly on the upper right" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.422, + 0.129 + ], + "angle": 0, + "content": "P.Y. Lee and M. Sung" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.145, + 0.371, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.146, + 0.497, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.145, + 0.631, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.146, + 0.756, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.254, + 0.642, + 0.267 + ], + "angle": 0, + "content": "\"A young child looking at a birthday cupcake\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.268, + 0.411, + 0.281 + ], + "angle": 0, + "content": "(a) GLIGEN" + }, + { + "type": "image_caption", + "bbox": [ + 0.591, + 0.268, + 0.676, + 0.281 + ], + "angle": 0, + "content": "(b) ReGround" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.293, + 0.787, + 0.35 + ], + "angle": 0, + "content": "Fig. 7: Generated images from the text prompt and bounding boxes from the MS-COCO-2017 (left of each column) and our COCO-Drop (right of each column). While GLIGEN [28] fails to generate \"a birthday cupcake\" when the corresponding bounding box is removed, our REGROUND successfully generates a cupcake on the table." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.785, + 0.44 + ], + "angle": 0, + "content": "side of GLIGEN (red), signifying a more advantageous trade-off across varying \\(\\gamma\\). The same pattern is observed in MS-COCO-2017, where our REGROUND achieves \\(68.33\\%\\) of the increase in CLIP score of GLIGEN while only compromising YOLO score by \\(2.62\\%\\) compared to the decrease for GLIGEN." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.788, + 0.625 + ], + "angle": 0, + "content": "Fig. 6-(c) further shows a quantitative comparison on the Counting subset of the newly generated NSR-1K-GPT benchmark. The plot reveals a consistent trend with the MS-COCO datasets. By reducing \\(\\gamma\\) from 1.0 to 0.1, GLIGEN's CLIP score is increased from 32.46 to 33.67, while the YOLO score is decreased from 65.36 to 26.38. In contrast, when \\(\\gamma = 1.0\\), REGROUND achieves a CLIP score of 33.20, which is equal to \\(61.16\\%\\) of GLIGEN's total improvement in CLIP score, while the compromise in YOLO score is equal to only \\(3.69\\%\\) of the total decrease in the YOLO score of GLIGEN from \\(\\gamma = 1.0\\) to \\(\\gamma = 0.1\\). Moreover, a comparison on the Spatial subset of NSR-1K-GPT is provided in the Supplementary (Sec. S2). These results highlight that the advantage of our REGROUND holds robustly for the realistic image captions provided in the MS-COCO [30], as well as for diverse text prompts generated by GPT-4 [34]." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.643, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Random Box Dropping. To further assess the extent of description omission in each method, we modify the MS-COCO-2017 dataset [30] to make the COCO-Drop dataset. In this version, the bounding boxes for \\(50\\%\\) of the categories are randomly removed from each image, thereby preventing every entity described in the text prompt from being included within the bounding boxes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Fig. 8 shows the quantitative comparison of REGROUND and GLIGEN on COCO-Drop. In this case, REGROUND shows a larger advantage over GLIGEN in CLIP score, obtaining a gap in CLIP score which is 1.57 times that of the original MS-COCO-2017 dataset before box dropping for \\(\\gamma = 1.0\\). Such a larger gap in CLIP score demonstrates that compared to GLIGEN, our REGROUND better reflects the text prompts even when some entities in the text prompt are not provided as a bounding box. Fig. 7 displays a representative example, where GLIGEN fails to generate a \"cupcake\" when its corresponding bounding" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.295, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "ReGround: Improving Textual and Spatial Grounding at No Cost" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.145, + 0.482, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.145, + 0.77, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.301, + 0.788, + 0.331 + ], + "angle": 0, + "content": "Fig. 8: Comparison on the COCO-Drop Fig.9: Comparison of FID [16] on MS-dataset. COCO-2017 [30] dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.788, + 0.393 + ], + "angle": 0, + "content": "box is removed in COCO-Drop, whereas our REGROUND robustly generates the cupcake even when it is not provided as a bounding box." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.408, + 0.788, + 0.47 + ], + "angle": 0, + "content": "Image Quality. Fig. 9 displays the relationship between YOLO score [52] and FID [16] for each method on MS-COCO-2017. Note that the FID of REGROUND is constantly lower than that of GLIGEN [28], meaning that our network rewiring also results in higher image quality and diversity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.485, + 0.788, + 0.624 + ], + "angle": 0, + "content": "User Study. We conducted a user study to compare GLIGEN and our ReGround in terms of faithfulness to input text prompts. We used GPT-4 [34] to generate 100 prompts each containing two different objects, along with a bounding box for each object. Participants were given the text prompt along with two images—one from each method—and asked to choose the image that \"better includes all the objects from the prompt.\" Among the 92 out of 100 participants who passed the vigilance tests, our REGROUND surpassed GLIGEN, with a preference rate of \\(70.05\\%\\) compared to \\(29.95\\%\\). Further details on the user study are provided in the Supplementary (Sec. S1)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.637, + 0.788, + 0.7 + ], + "angle": 0, + "content": "PickScore. We further compare the PickScore [26] of GLIGEN [28] and our REGROUND given each input text prompt. On MS-COCO-2017, REGROUND is preferred over GLIGEN by \\(55.66\\%\\) to \\(44.34\\%\\), and on COCO-Drop, REGROUND is preferred by \\(57.57\\%\\) to \\(42.43\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.723, + 0.573, + 0.739 + ], + "angle": 0, + "content": "6.4 Impact of REGROUND as a Backbone" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We demonstrate that applying our rewiring of attention modules can also improve text-image alignment in other layout-guided generation methods that use GLIGEN as a backbone. For instance, BoxDiff [54] is a notable example that uses GLIGEN as its foundation and improves spatial grounding with respect to the bounding boxes by leveraging cross-attention maps as additional spatial cues in a zero-shot manner. Our network rewiring can also be combined with" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.422, + 0.129 + ], + "angle": 0, + "content": "P.Y. Lee and M. Sung" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.147, + 0.396, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.229, + 0.25, + 0.388, + 0.262 + ], + "angle": 0, + "content": "(a) NSR-1K-GPT-Counting" + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.147, + 0.565, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.25, + 0.559, + 0.262 + ], + "angle": 0, + "content": "(b) NSR-1K-GPT-Spatial" + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.146, + 0.784, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.621, + 0.25, + 0.737, + 0.262 + ], + "angle": 0, + "content": "(c) BoxDiff Outputs" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.273, + 0.787, + 0.344 + ], + "angle": 0, + "content": "Fig. 10: Comparison of applying BoxDiff [54] on GLIGEN [28] and our REGROUND, respectively. (a) and (b) show that our REGROUND further improves the grounding quality of BoxDiff on NSR-1K-GPT datasets. (c) While BoxDiff with GLIGEN (left) also shows description omission—omitting \"beautiful sunset\" from the text prompt—BoxDiff with our REGROUND contains the sunset in the final image (right)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.492 + ], + "angle": 0, + "content": "the zero-shot guidance of BoxDiff. Fig. 10 illustrates the results on the NSR-1K-GPT datasets (a) when BoxDiff uses GLIGEN as the base, and (b) when it uses our REGROUND, the rewired GLIGEN, as the base. It depicts that for the same range of spatial grounding accuracies, REGROUND obtains noticeably higher textual grounding (i.e. CLIP score [15]). Also, as shown in Fig. 10-(c), our network rewiring allows for a more detailed description to accurately appear in the final image, both for the entities in the bounding boxes (\"truck\") and the entities that are given as a text prompt (\"sunset\")." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.512, + 0.359, + 0.527 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.541, + 0.788, + 0.768 + ], + "angle": 0, + "content": "We have demonstrated that a simple network rewiring of attention modules, making the gated self-attention and cross-attention parallel, surprisingly improves the trade-off between textual and spatial grounding at no additional cost — without introducing any new parameters, any fine-tuning of the network, or any changes in generation time and memory. Using the pretrained GLIGEN [28], which was trained with the original sequential architecture of the two attention modules, the reconfiguration at inference time has led to achieving higher CLIP scores, indicating the noticeable improvement in textual grounding accuracy. Moreover, our REGROUND improves the textual grounding while preserving the spatial grounding accuracy - achieving \\(70.25\\%\\) and \\(68.33\\%\\) of GLIGEN's total improvement with the scheduled sampling in CLIP score while compromising YOLO score only \\(3.31\\%\\) and \\(2.62\\%\\) for the MS-COCO-2014 and MS-COCO-2017 datasets, respectively. We also showcased that this simple yet effective solution for the textual-spatial grounding trade-off can lead to improvements in diverse frameworks using GLIGEN as a base." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Supplementary. Due to limited space, we provide the following contents in the Supplementary: details on the evaluation setup (Sec. S1), additional quantitative (Sec. S2) and qualitative (Sec. S4) comparisons, and more results of REGROUND as a backbone for other layout-guided generation methods (Sec. S3)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.295, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ReGround: Improving Textual and Spatial Grounding at No Cost" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.267 + ], + "angle": 0, + "content": "This work was supported by NRF grant (RS-2023-00209723), IITP grants (RS-2019-II190075, RS-2022-II220594, RS-2023-00227592, RS-2024-00399817), and Alchymist Project Program (RS-2024-00423625) funded by the Korean government (MSIT and MOTIE), and grants from the DRB-KAIST SketchTheFuture Research Center, Naver-intel, Adobe Research, Hyundai NGV, KT, and Samsung Electronics." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.289, + 0.323, + 0.305 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.785, + 0.36 + ], + "angle": 0, + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.361, + 0.785, + 0.401 + ], + "angle": 0, + "content": "2. Avrahami, O., Hayes, T., Gafni, O., Gupta, S., Taigman, Y., Parikh, D., Lischinski, D., Fried, O., Yin, X.: Spatext: Spatio-textual representation for controllable image generation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.402, + 0.785, + 0.442 + ], + "angle": 0, + "content": "3. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., et al.: edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.443, + 0.785, + 0.47 + ], + "angle": 0, + "content": "4. Bansal, A., Chu, H.M., Schwarzschild, A., Sengupta, S., Goldblum, M., Geiping, J., Goldstein, T.: Universal guidance for diffusion models. In: ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.471, + 0.785, + 0.497 + ], + "angle": 0, + "content": "5. Bar-Tal, O., Yariv, L., Lipman, Y., Dekel, T.: Multidiffusion: Fusing diffusion paths for controlled image generation. In: ICML (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.498, + 0.785, + 0.553 + ], + "angle": 0, + "content": "6. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., Manassra, W., Dhariwal, P., Chu, C., Jiao, Y., Ramesh, A.: Improving image generation with better captions (2023), https://cdn.openai.com/papers/dall-e-3.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.553, + 0.785, + 0.593 + ], + "angle": 0, + "content": "7. Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., Zheng, Y.: Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.594, + 0.785, + 0.621 + ], + "angle": 0, + "content": "8. Chen, M., Laina, I., Vedaldi, A.: Training-free layout control with cross-attention guidance. In: WACV (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.622, + 0.785, + 0.662 + ], + "angle": 0, + "content": "9. Chen, W.G., Spiridonova, I., Yang, J., Gao, J., Li, C.: Llava-interactive: An all-in-one demo for image chat, segmentation, generation and editing. arXiv preprint arXiv:2311.00571 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.663, + 0.785, + 0.689 + ], + "angle": 0, + "content": "0. Cheng, J., Liang, X., Shi, X., He, T., Xiao, T., Li, M.: Layoutdiffuse: Adapting foundational diffusion models for layout-to-image generation. arXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.69, + 0.785, + 0.717 + ], + "angle": 0, + "content": "1. Couairon, G., Careil, M., Cord, M., Lathuilière, S., Verbeek, J.: Zero-shot spatial layout conditioning for text-to-image diffusion models. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.785, + 0.758 + ], + "angle": 0, + "content": "2. Feng, W., Zhu, W., Fu, T.j., Jampani, V., Akula, A., He, X., Basu, S., Wang, X.E., Wang, W.Y.: LayoutGPT: Compositional visual planning and generation with large language models. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.785 + ], + "angle": 0, + "content": "3. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "4. Herzig, R., Bar, A., Xu, H., Chechik, G., Darrell, T., Globerson, A.: Learning canonical representations for scene graph to image generation. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "5. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: CLIPScore: a reference-free evaluation metric for image captioning. In: EMNLP (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.318, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.422, + 0.129 + ], + "angle": 0, + "content": "P.Y. Lee and M. Sung" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "16. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: GANs trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.786, + 0.218 + ], + "angle": 0, + "content": "17. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.245 + ], + "angle": 0, + "content": "18. Hong, S., Yang, D., Choi, J., Lee, H.: Inferring semantic layout for hierarchical text-to-image synthesis. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.786, + 0.273 + ], + "angle": 0, + "content": "19. Hong, S., Lee, G., Jang, W., Kim, S.: Improving sample quality of diffusion models using self-attention guidance. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.786, + 0.327 + ], + "angle": 0, + "content": "20. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (2021). https://doi.org/10.5281/zenodo.5143773, https://doi.org/10.5281/zenodo.5143773" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.786, + 0.356 + ], + "angle": 0, + "content": "21. Johnson, J., Gupta, A., Fei-Fei, L.: Image generation from scene graphs. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.786, + 0.384 + ], + "angle": 0, + "content": "22. Karras, T., Aittala, M., Laine, S., Härkönen, E., Hellsten, J., Lehtinen, J., Aila, T.: Alias-free generative adversarial networks. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.786, + 0.411 + ], + "angle": 0, + "content": "23. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.786, + 0.439 + ], + "angle": 0, + "content": "24. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of StyleGAN. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.786, + 0.466 + ], + "angle": 0, + "content": "25. Kim, Y., Lee, J., Kim, J.H., Ha, J.W., Zhu, J.Y.: Dense text-to-image generation with attention modulation. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.466, + 0.786, + 0.495 + ], + "angle": 0, + "content": "26. Kirstain, Y., Polyak, A., Singer, U., Matiana, S., Penna, J., Levy, O.: Pick-a-pic: An open dataset of user preferences for text-to-image generation. NeurIPS (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.786, + 0.522 + ], + "angle": 0, + "content": "27. Li, W., Zhang, P., Zhang, L., Huang, Q., He, X., Lyu, S., Gao, J.: Object-driven text-to-image synthesis via adversarial training. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.786, + 0.55 + ], + "angle": 0, + "content": "28. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: GLIGEN: Open-set grounded text-to-image generation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.786, + 0.577 + ], + "angle": 0, + "content": "29. Li, Z., Wu, J., Koh, I., Tang, Y., Sun, L.: Image synthesis from layout with locality-aware mask adaption. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.577, + 0.786, + 0.605 + ], + "angle": 0, + "content": "30. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.605, + 0.786, + 0.647 + ], + "angle": 0, + "content": "31. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.786, + 0.688 + ], + "angle": 0, + "content": "32. Ma, W.D.K., Lewis, J.P., Lahiri, A., Leung, T., Kleijn, W.B.: Directed diffusion: Direct control of object placement through attention guidance. arXiv preprint arXiv:2302.13153 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.688, + 0.786, + 0.73 + ], + "angle": 0, + "content": "33. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.558, + 0.744 + ], + "angle": 0, + "content": "34. OpenAI: Chatgpt, https://chat.openai.com/" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.786, + 0.771 + ], + "angle": 0, + "content": "35. Parmar, G., Kumar Singh, K., Zhang, R., Li, Y., Lu, J., Zhu, J.Y.: Zero-shot image-to-image translation. In: ACM TOG (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.786, + 0.799 + ], + "angle": 0, + "content": "36. Phung, Q., Ge, S., Huang, J.B.: Grounded text-to-image synthesis with attention refocusing. arXiv preprint arXiv:2306.05427 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.786, + 0.841 + ], + "angle": 0, + "content": "37. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: SDXL: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.295, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ReGround: Improving Textual and Spatial Grounding at No Cost" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "38. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "39. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.259 + ], + "angle": 0, + "content": "40. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.26, + 0.788, + 0.286 + ], + "angle": 0, + "content": "41. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.287, + 0.788, + 0.314 + ], + "angle": 0, + "content": "42. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.315, + 0.788, + 0.342 + ], + "angle": 0, + "content": "43. Si, C., Huang, Z., Jiang, Y., Liu, Z.: Freeu: Free lunch in diffusion u-net. arXiv preprint arXiv:2309.11497 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.343, + 0.732, + 0.356 + ], + "angle": 0, + "content": "44. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.357, + 0.788, + 0.396 + ], + "angle": 0, + "content": "45. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.398, + 0.788, + 0.424 + ], + "angle": 0, + "content": "46. Sun, W., Wu, T.: Image synthesis from reconfigurable layout and style. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.425, + 0.788, + 0.452 + ], + "angle": 0, + "content": "47. Sylvain, T., Zhang, P., Bengio, Y., Hjelm, R.D., Sharma, S.: Object-centric image generation from layouts. In: AAAI (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.453, + 0.788, + 0.493 + ], + "angle": 0, + "content": "48. Tancik, M., Srinivasan, P.P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U., Ramamoorthi, R., Barron, J.T., Ng, R.: Fourier features let networks learn high frequency functions in low dimensional domains. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.494, + 0.788, + 0.522 + ], + "angle": 0, + "content": "49. Tumanyan, N., Geyer, M., Bagon, S., Dekel, T.: Plug-and-play diffusion features for text-driven image-to-image translation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.549 + ], + "angle": 0, + "content": "50. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.549, + 0.788, + 0.577 + ], + "angle": 0, + "content": "51. Voynov, A., Aberman, K., Cohen-Or, D.: Sketch-guided text-to-image diffusion models. ACM TOG (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.578, + 0.788, + 0.605 + ], + "angle": 0, + "content": "52. Wang, C.Y., Bochkovskiy, A., Liao, H.Y.M.: YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.606, + 0.788, + 0.632 + ], + "angle": 0, + "content": "53. Xiao, J., Li, L., Lv, H., Wang, S., Huang, Q.: R&b: Region and boundary aware zero-shot grounded text-to-image generation. In: ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.633, + 0.788, + 0.66 + ], + "angle": 0, + "content": "54. Xie, J., Li, Y., Huang, Y., Liu, H., Zhang, W., Zheng, Y., Shou, M.Z.: Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.688 + ], + "angle": 0, + "content": "55. Yang, Z., Wang, J., Gan, Z., Li, L., Lin, K., Wu, C., Duan, N., Liu, Z., Liu, C., Zeng, M., et al.: Reco: Region-controlled text-to-image generation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.716 + ], + "angle": 0, + "content": "56. Yang, Z., Liu, D., Wang, C., Yang, J., Tao, D.: Modeling image composition for complex scene generation. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.717, + 0.788, + 0.743 + ], + "angle": 0, + "content": "57. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.744, + 0.788, + 0.771 + ], + "angle": 0, + "content": "58. Zhao, B., Meng, L., Yin, W., Signal, L.: Image generation from layout. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.772, + 0.788, + 0.799 + ], + "angle": 0, + "content": "59. Zhao, P., Li, H., Jin, R., Zhou, S.K.: Loco: Locally constrained training-free layout-to-image synthesis. arXiv preprint arXiv:2311.12342 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.788, + 0.827 + ], + "angle": 0, + "content": "60. Zheng, G., Zhou, X., Li, X., Qi, Z., Shan, Y., Li, X.: Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In: CVPR (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.827 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/ReGround_ Improving Textual and Spatial Grounding at No Cost/e369c9ee-bed9-4dbc-92b4-a4d9da7bccb2_origin.pdf b/2024/ReGround_ Improving Textual and Spatial Grounding at No Cost/e369c9ee-bed9-4dbc-92b4-a4d9da7bccb2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3549d24b306b7fe40effa2ce8a652372c3da7c99 --- /dev/null +++ b/2024/ReGround_ Improving Textual and Spatial Grounding at No Cost/e369c9ee-bed9-4dbc-92b4-a4d9da7bccb2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f2df7a79203782ccfbc78f4bee2c12d5829a483447bf9ba5d8d2cae4e8f9d4a +size 5623721 diff --git a/2024/ReGround_ Improving Textual and Spatial Grounding at No Cost/full.md b/2024/ReGround_ Improving Textual and Spatial Grounding at No Cost/full.md new file mode 100644 index 0000000000000000000000000000000000000000..75c86dd71bd357a8ba27234dc76329996914d7a3 --- /dev/null +++ b/2024/ReGround_ Improving Textual and Spatial Grounding at No Cost/full.md @@ -0,0 +1,417 @@ +# ReGround: Improving Textual and Spatial Grounding at No Cost + +Phillip Y. Lee Minhyuk Sung + +KAIST {phillip0701,mhsung}@kaist.ac.kr + +![](images/1d0c056f8167f65e44be115b293f34e721c4300ab57e22eeddd617e09ab924ea.jpg) +Layout + +![](images/131b71cfcede5f2bf278e995186aaa139d9015eec8f10d835a63409c4db35f78.jpg) +SD + +![](images/7ac1edcb020215b834573c83db9cf4616a37af57a4a20bc2be8b7753ddc803e9.jpg) +GLIGEN $_{\gamma=1.0}$ + +![](images/fb658c1a508ade0fb987ceb1b7d7a7a4c961aabc561756948e1df1fc90205279.jpg) +GLIGEN $\gamma = 0.2$ + +![](images/014e6a24b2cdfa600ad27b111698d1a5a82f6d827861c2d7efcf58c497c57e6f.jpg) +ReGround + +"A low poly illustration of a man holding a cell phone in hand." + +![](images/32207f22b62516630ca9333c10d1829d11ef4b4923178b68ff933b2fce05f8a7.jpg) + +![](images/c656866eb405f6fb36e983b7ee9e9de11abfd851d922a8674ba1fa05a25b0a1c.jpg) + +![](images/4e49957eca6d7c198c6b075b8f674ccb9985cd69f0e1857e74cf1c90a7fd9681.jpg) + +![](images/1f2226629e40a960990c4ee44625ad1a3992f5826ee58d0b37eb7cc5e528ae17.jpg) + +![](images/ac7698699064277a50e336d9f83c907b3609b2a0fee07de80b1ce7ad2bdec67e.jpg) + +"An elephant walking draped with a colorful blanket." + +![](images/cdc161ca5382b0ae0e708392168743eafaac4a068e028194e030aa99d1e215e7.jpg) + +![](images/1ff27e6f5adda68fcdb071a23fbd5c63231df4259846d6bf1721cec30532fc12.jpg) + +![](images/3773a220c49c8bf021a97176a8f27e99d7594e3847750a4dc1ffd33e258d8f78.jpg) + +![](images/2884fd53f2d1872e05de42a56061943e924778823667f798fdeabb0afe51f48b.jpg) + +![](images/02a06234d99acea9dbb1851d2d23f2b2ad2fefd396b152cc367f7ea8bf635d8d.jpg) + +"One ancient vase [...] the backdrop of a serene, moonlit desert landscape." + +![](images/3f60db638b8d8f8cd4e4d248b92d3f05bde298a9ce9be200534b72a141fe91f5.jpg) +Fig. 1: Comparison across Stable Diffusion (SD) [41], GLIGEN [28], and our REGROUND. SD (2nd column) can generate an image aligned with the input prompt (shown below each row), while it does not allow taking spatial constraints such as bounding boxes and labels. GLIGEN (3rd column) enables spatial grounding using gated self-attention, although it often disregards some descriptions in the input prompt due to a bias towards bounding box conditions. Such trends also occur when only activating gated self-attention for 0.2 fraction of the initial denoising steps (4th column). Our REGROUND (last column) resolves the issue of description omission while accurately reflecting the bounding box information. + +![](images/eb5bec30d4184918bfec9d8a869b091573b3327d63ee51e92661c11c033a4895.jpg) + +![](images/6de29257702448a1413cf3579dd8fc888b7a2674e461cadd70edebe8c6466950.jpg) + +![](images/e5c0060541006735ca3e0f4eb64824724310d4684354669d107ea9ecf2c5dd0f.jpg) + +![](images/1ec0bb9be383b766f86b47560193ef648b2841ca7ef2cd9bd719c79998c7bbd4.jpg) + +"A train blowing smoke is coming down the tracks." + +Abstract. When an image generation process is guided by both a text prompt and spatial cues, such as a set of bounding boxes, do these elements work in harmony, or does one dominate the other? Our analysis of a pretrained image diffusion model that integrates gated self-attention into the U-Net reveals that spatial grounding often outweighs textual grounding due to the sequential flow from gated self-attention to cross-attention. We demonstrate that such bias can be significantly mitigated without sacrificing accuracy in either grounding by simply rewiring the network architecture, changing from sequential to parallel for gated self-attention and cross-attention. This surprisingly simple yet effective solution does not require any fine-tuning of the network but substantially reduces the trade-off between the two groundings. Our experiments demonstrate significant improvements from the original GLIGEN to the rewired version in the trade-off between textual grounding and spatial grounding. The project webpage is at https://re-ground.github.io. + +Keywords: Textual Grounding $\cdot$ Spatial Grounding $\cdot$ Network Rewiring + +# 1 Introduction + +The emergence of diffusion models [17,44,45] has markedly propelled the field of text-to-image (T2I) generation forward, allowing users to generate high-quality images from text prompts. In a bid to further augment the creativity and controllability, recent efforts [2-5,8,11,25,28,32,36,54,55,57] have focused on enabling these models to understand and interpret spatial instructions, such as layouts [4,8,28,32,36,54,55], segmentation masks [2-5,11,25] and sketches [51,57]. + +Among them, bounding boxes are extensively employed in downstream image generation tasks [4, 8, 28, 32, 36, 55]. GLIGEN [28] is a pioneering work in terms of enhancing existing T2I models with the capability to incorporate additional spatial cues in the form of bounding boxes. Its core component, gated self-attention, is a simple attention module [50] that is plugged into each U-Net [42] layer of a pretrained T2I model such as Stable Diffusion [41], and is trained to accurately position various entities in their designated areas. A notable advantage of GLIGEN is that the original parameters of the underlying model remain unchanged, inheriting the generative capability of the T2I model while introducing the novel functionality of spatial grounding using bounding boxes. This capability has been leveraged by numerous studies to facilitate high-quality, layout-guided image generation [9, 12, 31, 54]. + +However, our analysis reveals that GLIGEN's integration of the gated self-attention into an existing T2I model is not optimal for blending new spatial guidance from bounding boxes with the original textual guidance. It often leads to the omission of specific details from the text prompts. For instance, in the first row and third column of Fig. 1, GLIGEN fails to reflect the description "low poly illustration" from the input text prompt. Also in the second row, a crucial detail in the text prompt, "draped with a colorful blanket", is neglected in the output image. We refer to this issue as description omission. Such outcomes imply + +that the current architectural design of GLIGEN does not effectively harmonize the new spatial guidance and the text conditioning in the given T2I model. Considering the widespread applications of GLIGEN in various layout-based generation tasks [9,12,31,36,53,54,59], these limitations represent a significant bottleneck. + +To address the observed neglect of textual grounding in GLIGEN, we first analyze the root causes. Our investigation reveals that the issue arises from the sequential arrangement of the spatial grounding and textual grounding modules. Specifically, the output of the gated self-attention is directed to a cross-attention module in each layer of the U-Net architecture (Fig. 2-(b)). + +Building on this insight, we propose a straightforward yet impactful solution: network rewiring. This approach alters the relationship between the two grounding modules from sequential to parallel (Fig. 2-(c)). Remarkably, this network modification significantly reduces the grounding trade-off between textual and spatial groundings without necessitating any adjustments to the network parameters. Importantly, this rewiring does not require additional network training, extra parameters, or changes in computational load and time. Simply reconfiguring the attention modules of the pretrained GLIGEN, originally trained with the sequential architecture, during inference dramatically enhances performance. + +In our experiments on MS-COCO [30] and our newly introduced NSR-1K-GPT datasets, we demonstrate that rewiring the pretrained GLIGEN substantially reduces the trade-off between textual and spatial groundings. This is evidenced by the evaluation of text prompt alignment (measured using CLIP score [38], PickScore [26] and user study) and bounding box alignment (assessed by YOLO score [52]). Furthermore, we show that our rewiring also leads to better outcomes in other frameworks using GLIGEN as a backbone, including BoxDiff [54]. + +# 2 Related Work + +# 2.1 Zero-Shot Guidance in Diffusion Models + +The progress in diffusion models [17, 44, 45] has significantly elevated the capabilities of text-to-image (T2I) generation, resulting in foundation models [6, 37, 39-41] that exhibit remarkable generative performance. Leveraging the robust performance of these models, recent studies [3-5, 8, 11, 25, 28, 32, 36, 54, 55, 57] have introduced efficient guidance techniques designed to further improve the image generation process. Notably, numerous works [8, 28, 32, 36, 43, 54, 55] focus on the internal architecture (Fig. 2-(a)) of the denoising U-Net of Latent Diffusion Models (LDMs) [41], where self-attention and cross-attention modules are intertwined to facilitate inter-pixel communication and text conditioning. The self-attention of U-Net can be utilized to improve image quality [19] or facilitate image translation [49] and image editing tasks [7]. Since text conditions are integrated via cross-attention, the intermediate attention maps have been leveraged + +to improve text faithfulness [13] or enable spatial manipulation of the generation process [35]. Recently, FreeU [43] analyzed the contributions of the backbone and residuals of the U-Net and proposed a free-lunch strategy to enhance image quality: reweighting the backbone and residual features maps. In contrast to previous works that only deal with self- and cross-attention in standard LDMs, we introduce a method to enhance GLIGEN [28] by reconnecting its gated self-attention with the other attention modules, thereby achieving performance improvement in zero-shot without any tuning of the network parameters. + +# 2.2 Layout-Guided Image Generation + +The use of layouts, particularly in the form of bounding boxes, has become a popular intermediary to bridge the gap between textual inputs and the images generated [14, 18, 21, 27, 29, 46, 47, 56, 58]. Layout2Im [58] samples object latent codes from a normal distribution, eliminating the need to predict instance masks as done in prior works [18, 21]. LostGAN [46] controls the style of each object by devising an extension of the feature normalization layer used in StyleGAN [22-24], while OC-GAN [47] incorporates the spatial relationships between objects using a scene-graph representation. LAMA [29] introduces a mask adaptation module that mitigates the semantic ambiguity arising from overlaps in the input layout. While these developments have greatly improved user control over image generation, their applicability is confined to the categories found in the training data, such as those of the MS-COCO [30] dataset. + +In contrast, recent studies [5, 8, 10, 28, 36, 54, 55, 60] have extended layout-guided image generation towards open-vocabulary, building on the advancements of foundational text-to-image (T2I) models [41]. Training-free approaches [3, 5, 8, 36, 54] aim to improve the spatial grounding of T2I models through straightforward guidance mechanisms. GLIGEN [28], on the other hand, introduces gated self-attention, which is injected into the U-Net architecture of the Latent Diffusion Model [41], and is trained to equip the underlying model with spatial grounding abilities. Given the simple architecture of GLIGEN and its robust grounding accuracy with the input bounding boxes, numerous studies [36, 53, 54, 59] build upon its framework and propose further refinements to increase performance. In this work, we identify and address a significant performance bottleneck in GLIGEN related to description omission and propose a simple yet effective solution. + +# 3 Background - Latent Diffusion Models [41] + +Rombach et al. [41] proposed Latent Diffusion Model (LDM), a text-to-image (T2I) diffusion model with a U-Net as the noise prediction network. It is trained to generate an image from an input text prompt by predicting the noise $\epsilon (\mathbf{x}_t,t,c)$ conditioned both on the timestep $t$ and the text embedding $c$ . Each layer of LDM's U-Net consists of three core components: a convolutional residual block, followed by a self-attention (SA), and a cross-attention (CA) module (Fig. 2-(a)). In each $l$ -th layer of the U-Net, its residual block first extracts intermediate visual features $F = (f_{1},\dots,f_{N_{l}})^{T}$ from the output of the previous layer. + +![](images/802f6014c8843222c1e0674d593e56802d1180a9b2ff054c97a41caf347e624d.jpg) +Fig. 2: Comparison between the U-Net architectures of (a) Latent Diffusion Model (LDM) [41], (b) GLIGEN [28] and (c) our REGROUND. From LDM, GLIGEN enables spatial grounding by injecting Gated Self-Attention before cross-attention, forming a sequential flow of them. Based on GLIGEN, our REGROUND changes the relationship of the two attention modules to become parallel, resulting in noticeable improvement in textual grounding while preserving the spatial grounding capability. (The residual block before self-attention is omitted.) + +The self-attention module then facilitates interaction between the features in $F$ . Subsequently, the cross-attention module enables the interaction between each visual feature $f_{i}$ and the text embedding $c$ . Throughout this process, the output feature of the previous module is also forwarded through a residual connection, as illustrated in lines 4-5 and 7 of Alg. 1 and also in Fig. 2-(a). + +# 4 GLIGEN [28] and Description Omission + +In this section, we review GLIGEN [28] and its key idea of employing gated self-attention for spatial grounding. Then, we present our key observations on the description omission issue that occurs due to the addition of gated self-attention. + +# 4.1 Gated Self-Attention + +Li et al. [28] propose a plug-in spatial grounding module, named gated self-attention, which adopts the gated attention mechanism [1] to equip a pretrained T2I diffusion model [41] with spatial grounding capabilities (Fig. 2-(b)). Given a set of bounding boxes and text labels for each of them, let $b_{i}$ be the $xy$ -coordinates of the $i$ -th bounding box's top-left and bottom-right corners, and $p_i$ be the corresponding text label. Then, the $i$ -th grounding token is defined as $g_{i} := \mathcal{G}(\mathcal{T}(p_{i}),\mathcal{F}(b_{i}))$ , where $\mathcal{T}(\cdot)$ is a pretrained text encoder [20,38], $\mathcal{F}(\cdot)$ is the Fourier embedding [33,48] and $\mathcal{G}(\cdot,\cdot)$ is a shallow MLP network that concatenates the two given embeddings, respectively. Given a set of grounding tokens $\{g_i\}$ , gated self-attention learns the self-attention among the unified feature set $(f_1,\dots,f_{N_l},g_1,\dots,g_M)$ where $\{f_i\}$ is the set of intermediate visual features in the $l$ -th layer of U-Net, and $M$ is the number of bounding boxes. + +As shown in Fig. 2-(b), gated self-attention receives the output of the self-attention along with the residual features as its input and forwards the output features to the cross-attention module. By incorporating gated self-attention into + +each layer of the U-Net, the model enables the placement of the entity specified in the text label $p_i$ at the location indicated by the bounding box $b_i$ . Note that the integration of gated self-attention does not require training the network from scratch or fine-tuning it, but can be accomplished simply by training the gated self-attention parameters while keeping all other parameters in the backbone model frozen. + +Alg. 1 shows the pseudocode of the U-Net forward-pass including the plug-in of gated self-attention in line 6. Note that $\beta_{t}$ is set to 1 for GLIGEN. If $\beta_{t} = 0$ , the algorithm is identical to that of LDM [41]. + +Algorithm 1: Noise Prediction U-Net with Gated Self-Attention. +Parameters: $\beta_{t}$ // Weight for GSA. Inputs: $\mathbf{x}_t, c, \{g_i\}_{i=0\dots N-1}$ ; // Noisy data at timestep $t$ , text condition, and grounding tokens +Outputs: $\epsilon_t$ ; // Noise at timestep $t-1$ . +1 Function U-Net $(\mathbf{x}_t, c, \{g_i\})$ : +2 $F \gets \mathbf{x}_t$ +3 for $i = 0, \dots, L-1$ do +4 $\begin{array}{c} F_{\mathrm{RS}} \gets \mathrm{Conv}(F) + F; \\ F_{\mathrm{SA}} \gets \mathrm{SA}(F_{\mathrm{RS}}) + F_{\mathrm{RS}}; \\ F_{\mathrm{GSA}} \gets \beta_t \cdot \mathrm{GSA}(F_{\mathrm{SA}}, \{g_i\}) + F_{\mathrm{SA}}; \\ F \gets \mathrm{CA}(F_{\mathrm{GSA}}, c) + F_{\mathrm{GSA}}; \end{array}$ // Residual block. +5 // Self-Attention module. +6 +7 +8 $\epsilon_t \gets F;$ +9 return $\epsilon_t$ ; + +# 4.2 Description Omission + +Despite its high accuracy in spatial grounding, GLIGEN [28] frequently struggles to capture essential attributes specified in the input text prompt. As illustrated in Fig. 3, the leftmost image shows "a person" and "a skateboard" accurately placed in their designated regions. However, a critical detail from the input text prompt, "black and white photography", is absent in the output image. This discrepancy often emerges when the input comprises distinct but equally important descriptions regarding the image, presented through text prompts and bounding boxes. Such omissions not only fail to convey the stylistic intent of the image but also tend to overlook significant objects mentioned within the text prompt. Additional examples of this problem are showcased in Fig. 1, where the second row demonstrates the absence of a "blanket" in the generated image, a key element from the text prompt. This limitation significantly hampers GLIGEN's fidelity to user-provided text prompts, a challenge we term as description omission. + +![](images/ddcbaa8b7839a288d6088cc2cee6e9a4d8c2f443f2be629bd09086e49746d7cd.jpg) +Fig. 3: (a) Images generated by GLIGEN [28] with varying activation duration of gated self-attention $\gamma$ in scheduled sampling (Sec. 5.1). The red words in the text prompt denote the words used as labels of the input bounding boxes. Note that for GLIGEN to reflect the underlined description in the text prompt in the final image, $\gamma$ must be decreased to 0.1, which compromises spatial grounding accuracy. (b) In contrast, our REGROUND reflects the underlined phrase even when $\gamma = 1.0$ , therefore achieving high accuracy in both textual and spatial grounding. + +# 5 ReGround: Rewiring Attention Modules + +Gated self-attention and cross-attention each play a crucial role in enabling spatial and textual groundings, by taking bounding boxes and text prompts as inputs, respectively. To tackle the issue of description omission, we first examine the impact of attention modules on the groundings they do not address: the effect of gated self-attention on textual grounding (Sec. 5.1), and the influence of cross-attention on spatial grounding (Sec. 5.2). Building on this analysis, we propose an approach for network reconfiguration, modifying the connections among self-attention, gated self-attention, and cross-attention modules (Sec. 5.3). + +# 5.1 Impact of Gated Self-Attention on Textual Grounding + +As the issue of description omission arises due to the newly added gated self-attention in GLIGEN [28], we first attempt to mitigate the impact of gated self-attention by using scheduled sampling [28], activating gated self-attention only in a few initial steps of the denoising process. This approach is inspired by the observation that the coarse structure of the final image is established within the first few denoising steps. The scheduling is applied by setting the weight of gated self-attention $\beta_{t}$ (line 6 of Alg. 1) as + +$$ +\beta_ {t} = \left\{ \begin{array}{l l} 1 & (t \leq \gamma \cdot T) \\ 0 & (t > \gamma \cdot T), \end{array} \right. \tag {1} +$$ + +where $\gamma \in [0,1]$ represents the fraction of the initial denoising steps to activate gated self-attention. + +Fig. 3-(a) shows an example of generated images while incrementally adjusting $\gamma$ from 1.0 to 0.0. As $\gamma$ is reduced from 1.0 to 0.0, the details specified in the text prompt, "a black and white photograph", begin to be reflected starting + +at $\gamma = 0.1$ , demonstrating that longer activation of gated self-attention may interfere with the alignment of the output image with the text prompt. However, as gated self-attention is activated for shorter durations, the spatial grounding diminishes, as shown in the objects' reduced alignment with the input bounding boxes. This phenomenon illustrates the inherent trade-off between spatial and textual grounding, which cannot be resolved by controlling the duration of gated self-attention activation. + +# 5.2 Impact of Cross-Attention on Spatial Grounding + +We also investigate whether cross-attention has influence on spatial grounding. For this, we conduct a toy experiment by removing cross-attention modules in GLIGEN [28], allowing the output of the gated self-attention to be directly passed to the next layer of the U-Net. This modification is equivalent to changing line 7 of Alg. 1 to $F \gets F_{GSA}$ . + +The results are displayed in Fig. 4. Note that, while the appearance of the background and objects changes, the silhouettes of the cat (left) and the individuals (right) remain precisely positioned within their respective bounding boxes without cross-attention. This observation indicates that while gated self-attention that is performed before cross-attention may compromise textual grounding, cross-attention that processes the output of gated self-attention does not affect spatial grounding. + +![](images/fdb30a9ad4c464208741327c14aee4da75f9ac9ba57496b212093f5bba5aacb4.jpg) +Fig. 4: Comparison of the output of GLIGEN [28] with and without cross-attention. While the absence of cross-attention reduces realism and quality of the image, the silhouette of objects remains grounded within the given bounding boxes, as shown in the third column of each case. + +# 5.3 Network Rewiring: From Sequential to Parallel + +Building on the analyses above, we propose a simple yet effective modification to the grounding mechanism, changing the relationship between gated self-attention and cross-attention from sequential to parallel. This change eliminates the placement of gated self-attention before cross-attention, thus preventing the reduction of text grounding caused by gated self-attention. Moreover, in this parallel arrangement, the preservation of spatial grounding is assured, as gated self-attention for spatial grounding does not require subsequent cross-attention. + +Specifically, recall that in GLIGEN [28], the output of gated self-attention is added to the residual from self-attention, which is then passed to the cross-attention module as follows: + +$$ +F _ {G S A} \gets \underbrace {\mathrm {G S A} (F _ {S A} , \{g _ {i} \})} + F _ {S A}; +$$ + +$$ +s p a t i a l g r o u n d i n g +$$ + +$$ +F \leftarrow \underbrace {\operatorname {C A} \left(F _ {G S A} , c\right)} + F _ {G S A}; \tag {2} +$$ + +$$ +\text {t e x t u a l g r o u n d i n g} +$$ + +We propose to transform this sequence grounding pipeline into two parallel processes as follows: + +$$ +F \leftarrow \underbrace {\operatorname {G S A} \left(F _ {S A} , \{g _ {i} \}\right)} + \underbrace {\operatorname {C A} \left(F _ {S A} , c\right)} + \underbrace {F _ {S A}}; \tag {3} +$$ + +$$ +\begin{array}{c c c} \text {s p a t i a l g r o u n d i n g} & \text {t e x t u a l g r o u n d i n g} & \text {r e s i d u a l} \end{array} +$$ + +Refer to Fig. 2 for the visualization of network architecture changes $((\mathrm{b})\rightarrow$ (c)). This network rewiring is feasible because the input to gated self-attention remains unchanged, while the input to cross-attention shifts to $F_{SA}$ , for which it was originally designed in the context of Latent Diffusion Models [41]. + +It is important to note that the modification is effective even when applied to the pretrained GLIGEN, which was trained with the sequential structure of the attention modules. Therefore, our rewiring does not require any additional training or fine-tuning, introduces no extra parameters, and does not affect computation time or memory usage during the generation process. The only requirement is the simple reconfiguration of the attention modules at inference time. + +# 6 Experiments + +In this section, we show the effectiveness of our REGROUND by evaluating the spatial grounding on existing layout-caption datasets [12, 30] and the textual grounding on text-image alignment metrics [26, 38]. We use the official GLI-GEN [28] checkpoint which is trained based on Stable Diffusion v1.4 [41]. + +# 6.1 Datasets + +MS-COCO. We use the validations sets of both MS-COCO-2014 and MS-COCO-2017 datasets [30]. Each dataset provides image-captions pairs and the $xy$ -coordinates of bounding boxes along with their corresponding object categories. + +NSR-1K-GPT. We also use the NSR-1K benchmark [12] for evaluation. Based on each subset of NSR-1K—Counting and Spatial—we develop a new benchmark, NSR-1K-GPT, augmenting each original caption in NSR-1K using GPT-4 [34]. The instructions for augmentation are to (i) elaborate on the descriptions of each mentioned entity and (ii) provide additional details about the background of the image. More details on the evaluation datasets are provided in the Supplementary (Sec. S1). + +![](images/fcb2b6d78cd34047407c1f0585f925c33cc7af3307104afe66252414890e1c38.jpg) +Layout + +![](images/a675bdb566510ebf7cb0b1615466b8a93bc97b7a7346aac8e5f451ffcfded1b8.jpg) +SD +"A woman in a white shirt standing in front of a fence smiling." + +![](images/d7c222f0bd883054fe1d39fae5abb749c32032149f5aaab54981d698f2458d76.jpg) +GLIGEN $_{\gamma = 1.0}$ + +![](images/845b0ae98dbadca274469004f638a143c20fe5a1e36d67c4918fd54e3588b82c.jpg) +GLIGEN $_{\gamma=0.2}$ + +![](images/1375ee709e14aa8fb2bd13955715c7044478935ac6fcb298a8d5843676715f05.jpg) +ReGround + +![](images/dad2d1026f04d0c6478c87e2d428c721673546bdf6a61026ba3ef059829366b4.jpg) +"One spotted cow in the photo, [...] under a clear blue sky." + +![](images/2f6994c3ce78575f1cbcebfa08d16db42204feeab6b5d351a24e504001fb1ae9.jpg) + +![](images/94a994ac988c5dff30ab9111a6f278f4a51f117fe942d9050ed1435c81e5bd89.jpg) + +![](images/9e03c7250da3b602519b0b60fa353d6ba1132df0400b9a0a0095eaa72fc4be79.jpg) + +![](images/cb715a930d882e537f9385274bf29e78ae31fff45227fe614fce9bef8e5cd995.jpg) + +![](images/62f7e592193a8e79bfdf41c1b89b4b9c9ae87aa64ba3a9cf87093b935c77522b.jpg) + +![](images/f705d3015702df30eca3851dd4a46b196a401ae91c38651b6cd7169323368348.jpg) + +![](images/af32fff613c22d2d3709ebb5026de943aa5f418e18254eed378b0f0d30fe5a27.jpg) + +![](images/9562212cbb0b7c310b47e1d9cd7fc5ceda3db46108367714a913f0afd80d290b.jpg) + +![](images/b378e9f4021dd9edecf2acd8d107137fab8f87545517f4792bfc3617cad4edbe.jpg) + +![](images/eb110959bbfc8a2de9baf1402357a140b75c052aef93e50041a9b86aac63f924.jpg) + +![](images/8bab439c32472a70460c103a943b9b263d145fe58d8940e04743278b5e3ed4fb.jpg) +"A foreign airline status board with a clock." + +![](images/1cd6360676e371ed83e2c8bd664bee1913fbdd9a273adf39b9b8dddb5be882b2.jpg) + +![](images/810c30771373b86f7389c53a4ad513a7f01be3ba1d5b297512fd544384cb2f46.jpg) + +![](images/78fa721a1debaec8c10915a39756a78e4f6d754978ef2add0ef5fb711569d219.jpg) + +![](images/afd685b2510da2353ac8f4552949cb28efea8873e1058f4ce767bc41dfc6ad9c.jpg) +"A palm tree shower curtain is shown in a small bathroom." +Fig. 5: Qualitative comparisons. Stable Diffusion (SD, 2nd column) generates images that align with the given text descriptions, including the underlined phrase in each row, but cannot take bounding boxes as input. GLIGEN (3rd column) creates images that match the input layouts but suffers from description omission, failing to reflect the underlined descriptions. Scheduled sampling strategy (4th column) can partially address this issue (for instance, in the 5th row, where "window" appears in the room), but it results in a noticeable decline in spatial accuracy (as seen in the 1st row, where the tie is not generated). In contrast, our method (last column) accurately incorporates the underlined text descriptions while maintaining precise spatial representation. + +![](images/373851e8d28802d2552548f1c9b2bfd45495b7e8fa1c1b17b6ac2ca8d53eb0f2.jpg) +"A view of a living room with large windows." + +![](images/24a5edcf557748b4bfcb3a764ca63faa0441f2bd47b6c937e076fabf60447187.jpg) + +![](images/bdb1c850f1c5a9cde54fac9fc0c8ea268cde084ba689c1f4bacc568b5df350cf.jpg) + +![](images/e1a731f958013e7fcf4c2a0d618735081c0fb616071736486cb0c9cee096f54d.jpg) + +![](images/81fb26cead795d08fca56286d50264e9abbb914b963c68023781836f7a44c233.jpg) +(a) MS-COCO-2014 +Fig. 6: Comparisons on MS-COCO [30] and NSR-1K-GPT. Each plot shows the relationship between textual grounding (i.e. CLIP score [15]) and spatial grounding (i.e. YOLO score [52]) accuracy of GLIGEN [28] and our method. Note that the plot of our REGROUND is positioned in the top-right quadrant relative to GLIGEN, signifying that it alleviates the inherent trade-off between textual and spatial grounding. + +![](images/7ce3c5555acc970e8f7cfd00eb0d5d73356fad84ded5b23a500cb609be25f85d.jpg) +(b) MS-COCO-2017 + +![](images/ff51c7df2da5ef955b127e09409a2c0cf616d3cf16eb22df3b912091f914057f.jpg) +(c) NSR-1K-GPT-Counting + +# 6.2 Evaluation Metrics + +- YOLO score: Spatial grounding accuracy is assessed usingYOLO score [52]. We employYOLOv7 [52] to detect objects in each generated image and compute the average precision (AP) based on the ground truth bounding box annotations from MS-COCO [30]. +- CLIP score: Textual grounding accuracy is assessed using CLIP score [15]. +- FID: Image quality and diversity are evaluated using FID [16]. +- User Study and PickScore: We conduct a user study to assess human preferences for the generated images based on each input text prompt. Additionally, we use PickScore [26], a human preference predictor, to further analyze the results. + +# 6.3 Comparison with GLIGEN + +Textual-Spatial Grounding Trade-off. We first examine the trade-off between textual and spatial groundings for both GLIGEN [28] and our REGROUND, the rewired version of GLIGEN, while varying the scheduled sampling parameter $\gamma$ from 1.0 to 0.1. + +Fig. 6-(a), (b) present the graphs of CLIP score [15] and YOLO score [52] measured on the MS-COCO datasets [30]. In MS-COCO-2014, when reducing $\gamma$ from 1.0 to 0.1, the CLIP score of GLIGEN varies from 30.44 to 31.65, while the YOLO score significantly drops from 58.13 to 22.75 (red in Fig. 6). In contrast, our REGROUND, (blue in Fig. 6), demonstrates a notably superior trade-off between textual and spatial grounding. Specifically, with $\gamma$ set to 1.0, REGROUND already achieves a CLIP score of 31.29, accounting for $70.25\%$ of GLIGEN's total improvement in CLIP score when $\gamma$ is reduced from 1.0 to 0.1. Despite this significant increase in CLIP score, the YOLO score remains largely unchanged, marking 56.96 which represents only a $3.31\%$ decrease in the range of YOLO score variation for GLIGEN when $\gamma$ is adjusted from 1.0 to 0.1. Moreover, when varying the $\gamma$ , the plot for REGROUND (blue) is constantly on the upper right + +![](images/e4a8403d8740b291024a3725deb0df83fa19d27b12e2ccda242132828f68f2c2.jpg) +"A young child looking at a birthday cupcake" + +![](images/69cd149967164ffc044a8dd9f130cf352eb13e9b08d9609a1f8863e110fa12ef.jpg) +(a) GLIGEN + +![](images/b025b7e851fd264919931122ef1ef22f9f486b5aa6772388174f4e7efeb0c76a.jpg) +Fig. 7: Generated images from the text prompt and bounding boxes from the MS-COCO-2017 (left of each column) and our COCO-Drop (right of each column). While GLIGEN [28] fails to generate "a birthday cupcake" when the corresponding bounding box is removed, our REGROUND successfully generates a cupcake on the table. + +![](images/5063241389f13d6b28a179318402b230ce64fea0ce326c25129d71e4b4c5b1b2.jpg) +(b) ReGround + +side of GLIGEN (red), signifying a more advantageous trade-off across varying $\gamma$ . The same pattern is observed in MS-COCO-2017, where our REGROUND achieves $68.33\%$ of the increase in CLIP score of GLIGEN while only compromising YOLO score by $2.62\%$ compared to the decrease for GLIGEN. + +Fig. 6-(c) further shows a quantitative comparison on the Counting subset of the newly generated NSR-1K-GPT benchmark. The plot reveals a consistent trend with the MS-COCO datasets. By reducing $\gamma$ from 1.0 to 0.1, GLIGEN's CLIP score is increased from 32.46 to 33.67, while the YOLO score is decreased from 65.36 to 26.38. In contrast, when $\gamma = 1.0$ , REGROUND achieves a CLIP score of 33.20, which is equal to $61.16\%$ of GLIGEN's total improvement in CLIP score, while the compromise in YOLO score is equal to only $3.69\%$ of the total decrease in the YOLO score of GLIGEN from $\gamma = 1.0$ to $\gamma = 0.1$ . Moreover, a comparison on the Spatial subset of NSR-1K-GPT is provided in the Supplementary (Sec. S2). These results highlight that the advantage of our REGROUND holds robustly for the realistic image captions provided in the MS-COCO [30], as well as for diverse text prompts generated by GPT-4 [34]. + +Random Box Dropping. To further assess the extent of description omission in each method, we modify the MS-COCO-2017 dataset [30] to make the COCO-Drop dataset. In this version, the bounding boxes for $50\%$ of the categories are randomly removed from each image, thereby preventing every entity described in the text prompt from being included within the bounding boxes. + +Fig. 8 shows the quantitative comparison of REGROUND and GLIGEN on COCO-Drop. In this case, REGROUND shows a larger advantage over GLIGEN in CLIP score, obtaining a gap in CLIP score which is 1.57 times that of the original MS-COCO-2017 dataset before box dropping for $\gamma = 1.0$ . Such a larger gap in CLIP score demonstrates that compared to GLIGEN, our REGROUND better reflects the text prompts even when some entities in the text prompt are not provided as a bounding box. Fig. 7 displays a representative example, where GLIGEN fails to generate a "cupcake" when its corresponding bounding + +![](images/5a6aac1e76d2f7c1516044409b7028fd482dfe931d0d124edc1d7dd5239a3458.jpg) +Fig. 8: Comparison on the COCO-Drop Fig.9: Comparison of FID [16] on MS-dataset. COCO-2017 [30] dataset. + +![](images/0f84585542b5645eb60db586d84e92bd060a1cf141e26fd3b43f91ec8ec89a92.jpg) + +box is removed in COCO-Drop, whereas our REGROUND robustly generates the cupcake even when it is not provided as a bounding box. + +Image Quality. Fig. 9 displays the relationship between YOLO score [52] and FID [16] for each method on MS-COCO-2017. Note that the FID of REGROUND is constantly lower than that of GLIGEN [28], meaning that our network rewiring also results in higher image quality and diversity. + +User Study. We conducted a user study to compare GLIGEN and our ReGround in terms of faithfulness to input text prompts. We used GPT-4 [34] to generate 100 prompts each containing two different objects, along with a bounding box for each object. Participants were given the text prompt along with two images—one from each method—and asked to choose the image that "better includes all the objects from the prompt." Among the 92 out of 100 participants who passed the vigilance tests, our REGROUND surpassed GLIGEN, with a preference rate of $70.05\%$ compared to $29.95\%$ . Further details on the user study are provided in the Supplementary (Sec. S1). + +PickScore. We further compare the PickScore [26] of GLIGEN [28] and our REGROUND given each input text prompt. On MS-COCO-2017, REGROUND is preferred over GLIGEN by $55.66\%$ to $44.34\%$ , and on COCO-Drop, REGROUND is preferred by $57.57\%$ to $42.43\%$ . + +# 6.4 Impact of REGROUND as a Backbone + +We demonstrate that applying our rewiring of attention modules can also improve text-image alignment in other layout-guided generation methods that use GLIGEN as a backbone. For instance, BoxDiff [54] is a notable example that uses GLIGEN as its foundation and improves spatial grounding with respect to the bounding boxes by leveraging cross-attention maps as additional spatial cues in a zero-shot manner. Our network rewiring can also be combined with + +![](images/03b98e25b383ddc4c43d3c259ddeac0171fa48cc95eb2baa45c1405bded64e84.jpg) +(a) NSR-1K-GPT-Counting + +![](images/85395662c0e6ddc2922f0b93bfbb39919cb38d383df80e1f937f68d37604ba91.jpg) +Fig. 10: Comparison of applying BoxDiff [54] on GLIGEN [28] and our REGROUND, respectively. (a) and (b) show that our REGROUND further improves the grounding quality of BoxDiff on NSR-1K-GPT datasets. (c) While BoxDiff with GLIGEN (left) also shows description omission—omitting "beautiful sunset" from the text prompt—BoxDiff with our REGROUND contains the sunset in the final image (right). + +![](images/87b71ec74bd2aa3de3c5b1f08b2ab2b18cc6938d4e601f7375248024dac082c6.jpg) +(b) NSR-1K-GPT-Spatial +(c) BoxDiff Outputs + +the zero-shot guidance of BoxDiff. Fig. 10 illustrates the results on the NSR-1K-GPT datasets (a) when BoxDiff uses GLIGEN as the base, and (b) when it uses our REGROUND, the rewired GLIGEN, as the base. It depicts that for the same range of spatial grounding accuracies, REGROUND obtains noticeably higher textual grounding (i.e. CLIP score [15]). Also, as shown in Fig. 10-(c), our network rewiring allows for a more detailed description to accurately appear in the final image, both for the entities in the bounding boxes ("truck") and the entities that are given as a text prompt ("sunset"). + +# 7 Conclusion + +We have demonstrated that a simple network rewiring of attention modules, making the gated self-attention and cross-attention parallel, surprisingly improves the trade-off between textual and spatial grounding at no additional cost — without introducing any new parameters, any fine-tuning of the network, or any changes in generation time and memory. Using the pretrained GLIGEN [28], which was trained with the original sequential architecture of the two attention modules, the reconfiguration at inference time has led to achieving higher CLIP scores, indicating the noticeable improvement in textual grounding accuracy. Moreover, our REGROUND improves the textual grounding while preserving the spatial grounding accuracy - achieving $70.25\%$ and $68.33\%$ of GLIGEN's total improvement with the scheduled sampling in CLIP score while compromising YOLO score only $3.31\%$ and $2.62\%$ for the MS-COCO-2014 and MS-COCO-2017 datasets, respectively. We also showcased that this simple yet effective solution for the textual-spatial grounding trade-off can lead to improvements in diverse frameworks using GLIGEN as a base. + +Supplementary. Due to limited space, we provide the following contents in the Supplementary: details on the evaluation setup (Sec. S1), additional quantitative (Sec. S2) and qualitative (Sec. S4) comparisons, and more results of REGROUND as a backbone for other layout-guided generation methods (Sec. S3). + +# Acknowledgments + +This work was supported by NRF grant (RS-2023-00209723), IITP grants (RS-2019-II190075, RS-2022-II220594, RS-2023-00227592, RS-2024-00399817), and Alchymist Project Program (RS-2024-00423625) funded by the Korean government (MSIT and MOTIE), and grants from the DRB-KAIST SketchTheFuture Research Center, Naver-intel, Adobe Research, Hyundai NGV, KT, and Samsung Electronics. + +# References + +1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: NeurIPS (2022) +2. Avrahami, O., Hayes, T., Gafni, O., Gupta, S., Taigman, Y., Parikh, D., Lischinski, D., Fried, O., Yin, X.: Spatext: Spatio-textual representation for controllable image generation. In: CVPR (2023) +3. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., et al.: edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022) +4. Bansal, A., Chu, H.M., Schwarzschild, A., Sengupta, S., Goldblum, M., Geiping, J., Goldstein, T.: Universal guidance for diffusion models. In: ICLR (2024) +5. Bar-Tal, O., Yariv, L., Lipman, Y., Dekel, T.: Multidiffusion: Fusing diffusion paths for controlled image generation. In: ICML (2023) +6. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., Manassra, W., Dhariwal, P., Chu, C., Jiao, Y., Ramesh, A.: Improving image generation with better captions (2023), https://cdn.openai.com/papers/dall-e-3.pdf +7. Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., Zheng, Y.: Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In: ICCV (2023) +8. Chen, M., Laina, I., Vedaldi, A.: Training-free layout control with cross-attention guidance. In: WACV (2024) +9. Chen, W.G., Spiridonova, I., Yang, J., Gao, J., Li, C.: Llava-interactive: An all-in-one demo for image chat, segmentation, generation and editing. arXiv preprint arXiv:2311.00571 (2023) +0. Cheng, J., Liang, X., Shi, X., He, T., Xiao, T., Li, M.: Layoutdiffuse: Adapting foundational diffusion models for layout-to-image generation. arXiv (2023) +1. Couairon, G., Careil, M., Cord, M., Lathuilière, S., Verbeek, J.: Zero-shot spatial layout conditioning for text-to-image diffusion models. In: ICCV (2023) +2. Feng, W., Zhu, W., Fu, T.j., Jampani, V., Akula, A., He, X., Basu, S., Wang, X.E., Wang, W.Y.: LayoutGPT: Compositional visual planning and generation with large language models. In: NeurIPS (2023) +3. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. In: ICLR (2022) +4. Herzig, R., Bar, A., Xu, H., Chechik, G., Darrell, T., Globerson, A.: Learning canonical representations for scene graph to image generation. In: ECCV (2020) +5. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: CLIPScore: a reference-free evaluation metric for image captioning. In: EMNLP (2021) + +16. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: GANs trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2018) +17. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020) +18. Hong, S., Yang, D., Choi, J., Lee, H.: Inferring semantic layout for hierarchical text-to-image synthesis. In: CVPR (2018) +19. Hong, S., Lee, G., Jang, W., Kim, S.: Improving sample quality of diffusion models using self-attention guidance. In: ICCV (2023) +20. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (2021). https://doi.org/10.5281/zenodo.5143773, https://doi.org/10.5281/zenodo.5143773 +21. Johnson, J., Gupta, A., Fei-Fei, L.: Image generation from scene graphs. In: CVPR (2018) +22. Karras, T., Aittala, M., Laine, S., Härkönen, E., Hellsten, J., Lehtinen, J., Aila, T.: Alias-free generative adversarial networks. In: NeurIPS (2021) +23. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (2019) +24. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of StyleGAN. In: CVPR (2020) +25. Kim, Y., Lee, J., Kim, J.H., Ha, J.W., Zhu, J.Y.: Dense text-to-image generation with attention modulation. In: ICCV (2023) +26. Kirstain, Y., Polyak, A., Singer, U., Matiana, S., Penna, J., Levy, O.: Pick-a-pic: An open dataset of user preferences for text-to-image generation. NeurIPS (2024) +27. Li, W., Zhang, P., Zhang, L., Huang, Q., He, X., Lyu, S., Gao, J.: Object-driven text-to-image synthesis via adversarial training. In: CVPR (2019) +28. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: GLIGEN: Open-set grounded text-to-image generation. In: CVPR (2023) +29. Li, Z., Wu, J., Koh, I., Tang, Y., Sun, L.: Image synthesis from layout with locality-aware mask adaption. In: CVPR (2021) +30. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014) +31. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023) +32. Ma, W.D.K., Lewis, J.P., Lahiri, A., Leung, T., Kleijn, W.B.: Directed diffusion: Direct control of object placement through attention guidance. arXiv preprint arXiv:2302.13153 (2023) +33. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) +34. OpenAI: Chatgpt, https://chat.openai.com/ +35. Parmar, G., Kumar Singh, K., Zhang, R., Li, Y., Lu, J., Zhu, J.Y.: Zero-shot image-to-image translation. In: ACM TOG (2023) +36. Phung, Q., Ge, S., Huang, J.B.: Grounded text-to-image synthesis with attention refocusing. arXiv preprint arXiv:2306.05427 (2023) +37. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: SDXL: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023) + +38. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021) +39. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022) +40. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: ICML (2021) +41. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) +42. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015) +43. Si, C., Huang, Z., Jiang, Y., Liu, Z.: Freeu: Free lunch in diffusion u-net. arXiv preprint arXiv:2309.11497 (2023) +44. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models (2021) +45. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: ICLR (2021) +46. Sun, W., Wu, T.: Image synthesis from reconfigurable layout and style. In: ICCV (2019) +47. Sylvain, T., Zhang, P., Bengio, Y., Hjelm, R.D., Sharma, S.: Object-centric image generation from layouts. In: AAAI (2021) +48. Tancik, M., Srinivasan, P.P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U., Ramamoorthi, R., Barron, J.T., Ng, R.: Fourier features let networks learn high frequency functions in low dimensional domains. In: NeurIPS (2020) +49. Tumanyan, N., Geyer, M., Bagon, S., Dekel, T.: Plug-and-play diffusion features for text-driven image-to-image translation. In: CVPR (2023) +50. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) +51. Voynov, A., Aberman, K., Cohen-Or, D.: Sketch-guided text-to-image diffusion models. ACM TOG (2023) +52. Wang, C.Y., Bochkovskiy, A., Liao, H.Y.M.: YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In: CVPR (2023) +53. Xiao, J., Li, L., Lv, H., Wang, S., Huang, Q.: R&b: Region and boundary aware zero-shot grounded text-to-image generation. In: ICLR (2024) +54. Xie, J., Li, Y., Huang, Y., Liu, H., Zhang, W., Zheng, Y., Shou, M.Z.: Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In: ICCV (2023) +55. Yang, Z., Wang, J., Gan, Z., Li, L., Lin, K., Wu, C., Duan, N., Liu, Z., Liu, C., Zeng, M., et al.: Reco: Region-controlled text-to-image generation. In: CVPR (2023) +56. Yang, Z., Liu, D., Wang, C., Yang, J., Tao, D.: Modeling image composition for complex scene generation. In: CVPR (2022) +57. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023) +58. Zhao, B., Meng, L., Yin, W., Signal, L.: Image generation from layout. In: CVPR (2019) +59. Zhao, P., Li, H., Jin, R., Zhou, S.K.: Loco: Locally constrained training-free layout-to-image synthesis. arXiv preprint arXiv:2311.12342 (2023) +60. Zheng, G., Zhou, X., Li, X., Qi, Z., Shan, Y., Li, X.: Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In: CVPR (2023) \ No newline at end of file diff --git a/2024/ReGround_ Improving Textual and Spatial Grounding at No Cost/images.zip b/2024/ReGround_ Improving Textual and Spatial Grounding at No Cost/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..29c062383daf0951ce59dd2f077e6bb6793f1c34 --- /dev/null +++ b/2024/ReGround_ Improving Textual and Spatial Grounding at No Cost/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:21dab3d0572369f8cdd0b66b9a78915a9b896afd7f92aa959ee23f28eb27362e +size 636459 diff --git a/2024/ReGround_ Improving Textual and Spatial Grounding at No Cost/layout.json b/2024/ReGround_ Improving Textual and Spatial Grounding at No Cost/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..df109fbf9fb400cde27f875442b060760a227bf1 --- /dev/null +++ b/2024/ReGround_ Improving Textual and Spatial Grounding at No Cost/layout.json @@ -0,0 +1,11736 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 159, + 111, + 455, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 111, + 455, + 148 + ], + "spans": [ + { + "bbox": [ + 159, + 111, + 455, + 148 + ], + "type": "text", + "content": "ReGround: Improving Textual and Spatial Grounding at No Cost" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 227, + 168, + 386, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 168, + 386, + 180 + ], + "spans": [ + { + "bbox": [ + 227, + 168, + 386, + 180 + ], + "type": "text", + "content": "Phillip Y. Lee Minhyuk Sung" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 230, + 190, + 383, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 190, + 383, + 213 + ], + "spans": [ + { + "bbox": [ + 230, + 190, + 383, + 213 + ], + "type": "text", + "content": "KAIST {phillip0701,mhsung}@kaist.ac.kr" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 229, + 202, + 297 + ], + "blocks": [ + { + "bbox": [ + 153, + 219, + 183, + 228 + ], + "lines": [ + { + "bbox": [ + 153, + 219, + 183, + 228 + ], + "spans": [ + { + "bbox": [ + 153, + 219, + 183, + 228 + ], + "type": "text", + "content": "Layout" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 229, + 202, + 297 + ], + "lines": [ + { + "bbox": [ + 133, + 229, + 202, + 297 + ], + "spans": [ + { + "bbox": [ + 133, + 229, + 202, + 297 + ], + "type": "image", + "image_path": "1d0c056f8167f65e44be115b293f34e721c4300ab57e22eeddd617e09ab924ea.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 203, + 230, + 271, + 297 + ], + "blocks": [ + { + "bbox": [ + 230, + 219, + 244, + 228 + ], + "lines": [ + { + "bbox": [ + 230, + 219, + 244, + 228 + ], + "spans": [ + { + "bbox": [ + 230, + 219, + 244, + 228 + ], + "type": "text", + "content": "SD" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 203, + 230, + 271, + 297 + ], + "lines": [ + { + "bbox": [ + 203, + 230, + 271, + 297 + ], + "spans": [ + { + "bbox": [ + 203, + 230, + 271, + 297 + ], + "type": "image", + "image_path": "131b71cfcede5f2bf278e995186aaa139d9015eec8f10d835a63409c4db35f78.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 272, + 230, + 340, + 297 + ], + "blocks": [ + { + "bbox": [ + 278, + 219, + 335, + 229 + ], + "lines": [ + { + "bbox": [ + 278, + 219, + 335, + 229 + ], + "spans": [ + { + "bbox": [ + 278, + 219, + 335, + 229 + ], + "type": "text", + "content": "GLIGEN" + }, + { + "bbox": [ + 278, + 219, + 335, + 229 + ], + "type": "inline_equation", + "content": "_{\\gamma=1.0}" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 272, + 230, + 340, + 297 + ], + "lines": [ + { + "bbox": [ + 272, + 230, + 340, + 297 + ], + "spans": [ + { + "bbox": [ + 272, + 230, + 340, + 297 + ], + "type": "image", + "image_path": "7ac1edcb020215b834573c83db9cf4616a37af57a4a20bc2be8b7753ddc803e9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 342, + 230, + 410, + 297 + ], + "blocks": [ + { + "bbox": [ + 347, + 219, + 405, + 229 + ], + "lines": [ + { + "bbox": [ + 347, + 219, + 405, + 229 + ], + "spans": [ + { + "bbox": [ + 347, + 219, + 405, + 229 + ], + "type": "text", + "content": "GLIGEN " + }, + { + "bbox": [ + 347, + 219, + 405, + 229 + ], + "type": "inline_equation", + "content": "\\gamma = 0.2" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 342, + 230, + 410, + 297 + ], + "lines": [ + { + "bbox": [ + 342, + 230, + 410, + 297 + ], + "spans": [ + { + "bbox": [ + 342, + 230, + 410, + 297 + ], + "type": "image", + "image_path": "fb658c1a508ade0fb987ceb1b7d7a7a4c961aabc561756948e1df1fc90205279.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 411, + 230, + 480, + 297 + ], + "blocks": [ + { + "bbox": [ + 421, + 219, + 470, + 228 + ], + "lines": [ + { + "bbox": [ + 421, + 219, + 470, + 228 + ], + "spans": [ + { + "bbox": [ + 421, + 219, + 470, + 228 + ], + "type": "text", + "content": "ReGround" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 411, + 230, + 480, + 297 + ], + "lines": [ + { + "bbox": [ + 411, + 230, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 411, + 230, + 480, + 297 + ], + "type": "image", + "image_path": "014e6a24b2cdfa600ad27b111698d1a5a82f6d827861c2d7efcf58c497c57e6f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 178, + 298, + 436, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 298, + 436, + 309 + ], + "spans": [ + { + "bbox": [ + 178, + 298, + 436, + 309 + ], + "type": "text", + "content": "\"A low poly illustration of a man holding a cell phone in hand.\"" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 134, + 311, + 202, + 378 + ], + "blocks": [ + { + "bbox": [ + 134, + 311, + 202, + 378 + ], + "lines": [ + { + "bbox": [ + 134, + 311, + 202, + 378 + ], + "spans": [ + { + "bbox": [ + 134, + 311, + 202, + 378 + ], + "type": "image", + "image_path": "32207f22b62516630ca9333c10d1829d11ef4b4923178b68ff933b2fce05f8a7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 203, + 311, + 271, + 378 + ], + "blocks": [ + { + "bbox": [ + 203, + 311, + 271, + 378 + ], + "lines": [ + { + "bbox": [ + 203, + 311, + 271, + 378 + ], + "spans": [ + { + "bbox": [ + 203, + 311, + 271, + 378 + ], + "type": "image", + "image_path": "c656866eb405f6fb36e983b7ee9e9de11abfd851d922a8674ba1fa05a25b0a1c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 272, + 311, + 340, + 378 + ], + "blocks": [ + { + "bbox": [ + 272, + 311, + 340, + 378 + ], + "lines": [ + { + "bbox": [ + 272, + 311, + 340, + 378 + ], + "spans": [ + { + "bbox": [ + 272, + 311, + 340, + 378 + ], + "type": "image", + "image_path": "4e49957eca6d7c198c6b075b8f674ccb9985cd69f0e1857e74cf1c90a7fd9681.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 342, + 311, + 410, + 378 + ], + "blocks": [ + { + "bbox": [ + 342, + 311, + 410, + 378 + ], + "lines": [ + { + "bbox": [ + 342, + 311, + 410, + 378 + ], + "spans": [ + { + "bbox": [ + 342, + 311, + 410, + 378 + ], + "type": "image", + "image_path": "1f2226629e40a960990c4ee44625ad1a3992f5826ee58d0b37eb7cc5e528ae17.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 411, + 311, + 480, + 378 + ], + "blocks": [ + { + "bbox": [ + 411, + 311, + 480, + 378 + ], + "lines": [ + { + "bbox": [ + 411, + 311, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 411, + 311, + 480, + 378 + ], + "type": "image", + "image_path": "ac7698699064277a50e336d9f83c907b3609b2a0fee07de80b1ce7ad2bdec67e.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 178, + 379, + 417, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 379, + 417, + 389 + ], + "spans": [ + { + "bbox": [ + 178, + 379, + 417, + 389 + ], + "type": "text", + "content": "\"An elephant walking draped with a colorful blanket.\"" + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 134, + 391, + 202, + 459 + ], + "blocks": [ + { + "bbox": [ + 134, + 391, + 202, + 459 + ], + "lines": [ + { + "bbox": [ + 134, + 391, + 202, + 459 + ], + "spans": [ + { + "bbox": [ + 134, + 391, + 202, + 459 + ], + "type": "image", + "image_path": "cdc161ca5382b0ae0e708392168743eafaac4a068e028194e030aa99d1e215e7.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 203, + 391, + 271, + 459 + ], + "blocks": [ + { + "bbox": [ + 203, + 391, + 271, + 459 + ], + "lines": [ + { + "bbox": [ + 203, + 391, + 271, + 459 + ], + "spans": [ + { + "bbox": [ + 203, + 391, + 271, + 459 + ], + "type": "image", + "image_path": "1ff27e6f5adda68fcdb071a23fbd5c63231df4259846d6bf1721cec30532fc12.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 272, + 391, + 340, + 459 + ], + "blocks": [ + { + "bbox": [ + 272, + 391, + 340, + 459 + ], + "lines": [ + { + "bbox": [ + 272, + 391, + 340, + 459 + ], + "spans": [ + { + "bbox": [ + 272, + 391, + 340, + 459 + ], + "type": "image", + "image_path": "3773a220c49c8bf021a97176a8f27e99d7594e3847750a4dc1ffd33e258d8f78.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 342, + 391, + 410, + 459 + ], + "blocks": [ + { + "bbox": [ + 342, + 391, + 410, + 459 + ], + "lines": [ + { + "bbox": [ + 342, + 391, + 410, + 459 + ], + "spans": [ + { + "bbox": [ + 342, + 391, + 410, + 459 + ], + "type": "image", + "image_path": "2884fd53f2d1872e05de42a56061943e924778823667f798fdeabb0afe51f48b.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 411, + 391, + 480, + 459 + ], + "blocks": [ + { + "bbox": [ + 411, + 391, + 480, + 459 + ], + "lines": [ + { + "bbox": [ + 411, + 391, + 480, + 459 + ], + "spans": [ + { + "bbox": [ + 411, + 391, + 480, + 459 + ], + "type": "image", + "image_path": "02a06234d99acea9dbb1851d2d23f2b2ad2fefd396b152cc367f7ea8bf635d8d.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 153, + 460, + 460, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 460, + 460, + 471 + ], + "spans": [ + { + "bbox": [ + 153, + 460, + 460, + 471 + ], + "type": "text", + "content": "\"One ancient vase [...] the backdrop of a serene, moonlit desert landscape.\"" + } + ] + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 134, + 472, + 202, + 540 + ], + "blocks": [ + { + "bbox": [ + 134, + 472, + 202, + 540 + ], + "lines": [ + { + "bbox": [ + 134, + 472, + 202, + 540 + ], + "spans": [ + { + "bbox": [ + 134, + 472, + 202, + 540 + ], + "type": "image", + "image_path": "3f60db638b8d8f8cd4e4d248b92d3f05bde298a9ce9be200534b72a141fe91f5.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 566, + 482, + 666 + ], + "lines": [ + { + "bbox": [ + 130, + 566, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 566, + 482, + 666 + ], + "type": "text", + "content": "Fig. 1: Comparison across Stable Diffusion (SD) [41], GLIGEN [28], and our REGROUND. SD (2nd column) can generate an image aligned with the input prompt (shown below each row), while it does not allow taking spatial constraints such as bounding boxes and labels. GLIGEN (3rd column) enables spatial grounding using gated self-attention, although it often disregards some descriptions in the input prompt due to a bias towards bounding box conditions. Such trends also occur when only activating gated self-attention for 0.2 fraction of the initial denoising steps (4th column). Our REGROUND (last column) resolves the issue of description omission while accurately reflecting the bounding box information." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 203, + 472, + 271, + 540 + ], + "blocks": [ + { + "bbox": [ + 203, + 472, + 271, + 540 + ], + "lines": [ + { + "bbox": [ + 203, + 472, + 271, + 540 + ], + "spans": [ + { + "bbox": [ + 203, + 472, + 271, + 540 + ], + "type": "image", + "image_path": "eb5bec30d4184918bfec9d8a869b091573b3327d63ee51e92661c11c033a4895.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 272, + 472, + 340, + 540 + ], + "blocks": [ + { + "bbox": [ + 272, + 472, + 340, + 540 + ], + "lines": [ + { + "bbox": [ + 272, + 472, + 340, + 540 + ], + "spans": [ + { + "bbox": [ + 272, + 472, + 340, + 540 + ], + "type": "image", + "image_path": "6de29257702448a1413cf3579dd8fc888b7a2674e461cadd70edebe8c6466950.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 342, + 472, + 410, + 540 + ], + "blocks": [ + { + "bbox": [ + 342, + 472, + 410, + 540 + ], + "lines": [ + { + "bbox": [ + 342, + 472, + 410, + 540 + ], + "spans": [ + { + "bbox": [ + 342, + 472, + 410, + 540 + ], + "type": "image", + "image_path": "e5c0060541006735ca3e0f4eb64824724310d4684354669d107ea9ecf2c5dd0f.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 411, + 472, + 480, + 540 + ], + "blocks": [ + { + "bbox": [ + 411, + 472, + 480, + 540 + ], + "lines": [ + { + "bbox": [ + 411, + 472, + 480, + 540 + ], + "spans": [ + { + "bbox": [ + 411, + 472, + 480, + 540 + ], + "type": "image", + "image_path": "1ec0bb9be383b766f86b47560193ef648b2841ca7ef2cd9bd719c79998c7bbd4.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "bbox": [ + 202, + 541, + 411, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 541, + 411, + 552 + ], + "spans": [ + { + "bbox": [ + 202, + 541, + 411, + 552 + ], + "type": "text", + "content": "\"A train blowing smoke is coming down the tracks.\"" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 160, + 117, + 455, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 117, + 455, + 283 + ], + "spans": [ + { + "bbox": [ + 160, + 117, + 455, + 283 + ], + "type": "text", + "content": "Abstract. When an image generation process is guided by both a text prompt and spatial cues, such as a set of bounding boxes, do these elements work in harmony, or does one dominate the other? Our analysis of a pretrained image diffusion model that integrates gated self-attention into the U-Net reveals that spatial grounding often outweighs textual grounding due to the sequential flow from gated self-attention to cross-attention. We demonstrate that such bias can be significantly mitigated without sacrificing accuracy in either grounding by simply rewiring the network architecture, changing from sequential to parallel for gated self-attention and cross-attention. This surprisingly simple yet effective solution does not require any fine-tuning of the network but substantially reduces the trade-off between the two groundings. Our experiments demonstrate significant improvements from the original GLIGEN to the rewired version in the trade-off between textual grounding and spatial grounding. The project webpage is at https://re-ground.github.io." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 293, + 453, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 293, + 453, + 304 + ], + "spans": [ + { + "bbox": [ + 160, + 293, + 453, + 304 + ], + "type": "text", + "content": "Keywords: Textual Grounding " + }, + { + "bbox": [ + 160, + 293, + 453, + 304 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 293, + 453, + 304 + ], + "type": "text", + "content": " Spatial Grounding " + }, + { + "bbox": [ + 160, + 293, + 453, + 304 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 293, + 453, + 304 + ], + "type": "text", + "content": " Network Rewiring" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 326, + 230, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 230, + 339 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 230, + 339 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "text", + "content": "The emergence of diffusion models [17,44,45] has markedly propelled the field of text-to-image (T2I) generation forward, allowing users to generate high-quality images from text prompts. In a bid to further augment the creativity and controllability, recent efforts [2-5,8,11,25,28,32,36,54,55,57] have focused on enabling these models to understand and interpret spatial instructions, such as layouts [4,8,28,32,36,54,55], segmentation masks [2-5,11,25] and sketches [51,57]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 569 + ], + "type": "text", + "content": "Among them, bounding boxes are extensively employed in downstream image generation tasks [4, 8, 28, 32, 36, 55]. GLIGEN [28] is a pioneering work in terms of enhancing existing T2I models with the capability to incorporate additional spatial cues in the form of bounding boxes. Its core component, gated self-attention, is a simple attention module [50] that is plugged into each U-Net [42] layer of a pretrained T2I model such as Stable Diffusion [41], and is trained to accurately position various entities in their designated areas. A notable advantage of GLIGEN is that the original parameters of the underlying model remain unchanged, inheriting the generative capability of the T2I model while introducing the novel functionality of spatial grounding using bounding boxes. This capability has been leveraged by numerous studies to facilitate high-quality, layout-guided image generation [9, 12, 31, 54]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "However, our analysis reveals that GLIGEN's integration of the gated self-attention into an existing T2I model is not optimal for blending new spatial guidance from bounding boxes with the original textual guidance. It often leads to the omission of specific details from the text prompts. For instance, in the first row and third column of Fig. 1, GLIGEN fails to reflect the description \"low poly illustration\" from the input text prompt. Also in the second row, a crucial detail in the text prompt, \"draped with a colorful blanket\", is neglected in the output image. We refer to this issue as description omission. Such outcomes imply" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "text", + "content": "P.Y. Lee and M. Sung" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "that the current architectural design of GLIGEN does not effectively harmonize the new spatial guidance and the text conditioning in the given T2I model. Considering the widespread applications of GLIGEN in various layout-based generation tasks [9,12,31,36,53,54,59], these limitations represent a significant bottleneck." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 480, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 480, + 236 + ], + "type": "text", + "content": "To address the observed neglect of textual grounding in GLIGEN, we first analyze the root causes. Our investigation reveals that the issue arises from the sequential arrangement of the spatial grounding and textual grounding modules. Specifically, the output of the gated self-attention is directed to a cross-attention module in each layer of the U-Net architecture (Fig. 2-(b))." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 236, + 481, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 481, + 355 + ], + "type": "text", + "content": "Building on this insight, we propose a straightforward yet impactful solution: network rewiring. This approach alters the relationship between the two grounding modules from sequential to parallel (Fig. 2-(c)). Remarkably, this network modification significantly reduces the grounding trade-off between textual and spatial groundings without necessitating any adjustments to the network parameters. Importantly, this rewiring does not require additional network training, extra parameters, or changes in computational load and time. Simply reconfiguring the attention modules of the pretrained GLIGEN, originally trained with the sequential architecture, during inference dramatically enhances performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 356, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 452 + ], + "type": "text", + "content": "In our experiments on MS-COCO [30] and our newly introduced NSR-1K-GPT datasets, we demonstrate that rewiring the pretrained GLIGEN substantially reduces the trade-off between textual and spatial groundings. This is evidenced by the evaluation of text prompt alignment (measured using CLIP score [38], PickScore [26] and user study) and bounding box alignment (assessed by YOLO score [52]). Furthermore, we show that our rewiring also leads to better outcomes in other frameworks using GLIGEN as a backbone, including BoxDiff [54]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 472, + 237, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 237, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 237, + 485 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 499, + 363, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 499, + 363, + 511 + ], + "spans": [ + { + "bbox": [ + 131, + 499, + 363, + 511 + ], + "type": "text", + "content": "2.1 Zero-Shot Guidance in Diffusion Models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "The progress in diffusion models [17, 44, 45] has significantly elevated the capabilities of text-to-image (T2I) generation, resulting in foundation models [6, 37, 39-41] that exhibit remarkable generative performance. Leveraging the robust performance of these models, recent studies [3-5, 8, 11, 25, 28, 32, 36, 54, 55, 57] have introduced efficient guidance techniques designed to further improve the image generation process. Notably, numerous works [8, 28, 32, 36, 43, 54, 55] focus on the internal architecture (Fig. 2-(a)) of the denoising U-Net of Latent Diffusion Models (LDMs) [41], where self-attention and cross-attention modules are intertwined to facilitate inter-pixel communication and text conditioning. The self-attention of U-Net can be utilized to improve image quality [19] or facilitate image translation [49] and image editing tasks [7]. Since text conditions are integrated via cross-attention, the intermediate attention maps have been leveraged" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 180, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 180, + 91, + 448, + 103 + ], + "type": "text", + "content": "ReGround: Improving Textual and Spatial Grounding at No Cost" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "to improve text faithfulness [13] or enable spatial manipulation of the generation process [35]. Recently, FreeU [43] analyzed the contributions of the backbone and residuals of the U-Net and proposed a free-lunch strategy to enhance image quality: reweighting the backbone and residual features maps. In contrast to previous works that only deal with self- and cross-attention in standard LDMs, we introduce a method to enhance GLIGEN [28] by reconnecting its gated self-attention with the other attention modules, thereby achieving performance improvement in zero-shot without any tuning of the network parameters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 227, + 330, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 227, + 330, + 239 + ], + "spans": [ + { + "bbox": [ + 131, + 227, + 330, + 239 + ], + "type": "text", + "content": "2.2 Layout-Guided Image Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 244, + 482, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 244, + 482, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 244, + 482, + 387 + ], + "type": "text", + "content": "The use of layouts, particularly in the form of bounding boxes, has become a popular intermediary to bridge the gap between textual inputs and the images generated [14, 18, 21, 27, 29, 46, 47, 56, 58]. Layout2Im [58] samples object latent codes from a normal distribution, eliminating the need to predict instance masks as done in prior works [18, 21]. LostGAN [46] controls the style of each object by devising an extension of the feature normalization layer used in StyleGAN [22-24], while OC-GAN [47] incorporates the spatial relationships between objects using a scene-graph representation. LAMA [29] introduces a mask adaptation module that mitigates the semantic ambiguity arising from overlaps in the input layout. While these developments have greatly improved user control over image generation, their applicability is confined to the categories found in the training data, such as those of the MS-COCO [30] dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 388, + 483, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 483, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 483, + 533 + ], + "type": "text", + "content": "In contrast, recent studies [5, 8, 10, 28, 36, 54, 55, 60] have extended layout-guided image generation towards open-vocabulary, building on the advancements of foundational text-to-image (T2I) models [41]. Training-free approaches [3, 5, 8, 36, 54] aim to improve the spatial grounding of T2I models through straightforward guidance mechanisms. GLIGEN [28], on the other hand, introduces gated self-attention, which is injected into the U-Net architecture of the Latent Diffusion Model [41], and is trained to equip the underlying model with spatial grounding abilities. Given the simple architecture of GLIGEN and its robust grounding accuracy with the input bounding boxes, numerous studies [36, 53, 54, 59] build upon its framework and propose further refinements to increase performance. In this work, we identify and address a significant performance bottleneck in GLIGEN related to description omission and propose a simple yet effective solution." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 547, + 416, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 547, + 416, + 562 + ], + "spans": [ + { + "bbox": [ + 131, + 547, + 416, + 562 + ], + "type": "text", + "content": "3 Background - Latent Diffusion Models [41]" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": "Rombach et al. [41] proposed Latent Diffusion Model (LDM), a text-to-image (T2I) diffusion model with a U-Net as the noise prediction network. It is trained to generate an image from an input text prompt by predicting the noise " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\epsilon (\\mathbf{x}_t,t,c)" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " conditioned both on the timestep " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " and the text embedding " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": ". Each layer of LDM's U-Net consists of three core components: a convolutional residual block, followed by a self-attention (SA), and a cross-attention (CA) module (Fig. 2-(a)). In each " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": "-th layer of the U-Net, its residual block first extracts intermediate visual features " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "F = (f_{1},\\dots,f_{N_{l}})^{T}" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " from the output of the previous layer." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "text", + "content": "P.Y. Lee and M. Sung" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 482, + 194 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 194 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 194 + ], + "type": "image", + "image_path": "802f6014c8843222c1e0674d593e56802d1180a9b2ff054c97a41caf347e624d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 201, + 482, + 278 + ], + "lines": [ + { + "bbox": [ + 130, + 201, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 482, + 278 + ], + "type": "text", + "content": "Fig. 2: Comparison between the U-Net architectures of (a) Latent Diffusion Model (LDM) [41], (b) GLIGEN [28] and (c) our REGROUND. From LDM, GLIGEN enables spatial grounding by injecting Gated Self-Attention before cross-attention, forming a sequential flow of them. Based on GLIGEN, our REGROUND changes the relationship of the two attention modules to become parallel, resulting in noticeable improvement in textual grounding while preserving the spatial grounding capability. (The residual block before self-attention is omitted.)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 302, + 482, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 482, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 482, + 363 + ], + "type": "text", + "content": "The self-attention module then facilitates interaction between the features in " + }, + { + "bbox": [ + 130, + 302, + 482, + 363 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 302, + 482, + 363 + ], + "type": "text", + "content": ". Subsequently, the cross-attention module enables the interaction between each visual feature " + }, + { + "bbox": [ + 130, + 302, + 482, + 363 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 130, + 302, + 482, + 363 + ], + "type": "text", + "content": " and the text embedding " + }, + { + "bbox": [ + 130, + 302, + 482, + 363 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 302, + 482, + 363 + ], + "type": "text", + "content": ". Throughout this process, the output feature of the previous module is also forwarded through a residual connection, as illustrated in lines 4-5 and 7 of Alg. 1 and also in Fig. 2-(a)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 382, + 391, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 382, + 391, + 397 + ], + "spans": [ + { + "bbox": [ + 131, + 382, + 391, + 397 + ], + "type": "text", + "content": "4 GLIGEN [28] and Description Omission" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 408, + 481, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 481, + 445 + ], + "type": "text", + "content": "In this section, we review GLIGEN [28] and its key idea of employing gated self-attention for spatial grounding. Then, we present our key observations on the description omission issue that occurs due to the addition of gated self-attention." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 464, + 266, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 464, + 266, + 475 + ], + "spans": [ + { + "bbox": [ + 131, + 464, + 266, + 475 + ], + "type": "text", + "content": "4.1 Gated Self-Attention" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 485, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 628 + ], + "type": "text", + "content": "Li et al. [28] propose a plug-in spatial grounding module, named gated self-attention, which adopts the gated attention mechanism [1] to equip a pretrained T2I diffusion model [41] with spatial grounding capabilities (Fig. 2-(b)). Given a set of bounding boxes and text labels for each of them, let " + }, + { + "bbox": [ + 130, + 485, + 482, + 628 + ], + "type": "inline_equation", + "content": "b_{i}" + }, + { + "bbox": [ + 130, + 485, + 482, + 628 + ], + "type": "text", + "content": " be the " + }, + { + "bbox": [ + 130, + 485, + 482, + 628 + ], + "type": "inline_equation", + "content": "xy" + }, + { + "bbox": [ + 130, + 485, + 482, + 628 + ], + "type": "text", + "content": "-coordinates of the " + }, + { + "bbox": [ + 130, + 485, + 482, + 628 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 485, + 482, + 628 + ], + "type": "text", + "content": "-th bounding box's top-left and bottom-right corners, and " + }, + { + "bbox": [ + 130, + 485, + 482, + 628 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 130, + 485, + 482, + 628 + ], + "type": "text", + "content": " be the corresponding text label. Then, the " + }, + { + "bbox": [ + 130, + 485, + 482, + 628 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 485, + 482, + 628 + ], + "type": "text", + "content": "-th grounding token is defined as " + }, + { + "bbox": [ + 130, + 485, + 482, + 628 + ], + "type": "inline_equation", + "content": "g_{i} := \\mathcal{G}(\\mathcal{T}(p_{i}),\\mathcal{F}(b_{i}))" + }, + { + "bbox": [ + 130, + 485, + 482, + 628 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 485, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\cdot)" + }, + { + "bbox": [ + 130, + 485, + 482, + 628 + ], + "type": "text", + "content": " is a pretrained text encoder [20,38], " + }, + { + "bbox": [ + 130, + 485, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot)" + }, + { + "bbox": [ + 130, + 485, + 482, + 628 + ], + "type": "text", + "content": " is the Fourier embedding [33,48] and " + }, + { + "bbox": [ + 130, + 485, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(\\cdot,\\cdot)" + }, + { + "bbox": [ + 130, + 485, + 482, + 628 + ], + "type": "text", + "content": " is a shallow MLP network that concatenates the two given embeddings, respectively. Given a set of grounding tokens " + }, + { + "bbox": [ + 130, + 485, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\{g_i\\}" + }, + { + "bbox": [ + 130, + 485, + 482, + 628 + ], + "type": "text", + "content": ", gated self-attention learns the self-attention among the unified feature set " + }, + { + "bbox": [ + 130, + 485, + 482, + 628 + ], + "type": "inline_equation", + "content": "(f_1,\\dots,f_{N_l},g_1,\\dots,g_M)" + }, + { + "bbox": [ + 130, + 485, + 482, + 628 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 485, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\{f_i\\}" + }, + { + "bbox": [ + 130, + 485, + 482, + 628 + ], + "type": "text", + "content": " is the set of intermediate visual features in the " + }, + { + "bbox": [ + 130, + 485, + 482, + 628 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 485, + 482, + 628 + ], + "type": "text", + "content": "-th layer of U-Net, and " + }, + { + "bbox": [ + 130, + 485, + 482, + 628 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 485, + 482, + 628 + ], + "type": "text", + "content": " is the number of bounding boxes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "As shown in Fig. 2-(b), gated self-attention receives the output of the self-attention along with the residual features as its input and forwards the output features to the cross-attention module. By incorporating gated self-attention into" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 180, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 180, + 91, + 448, + 102 + ], + "type": "text", + "content": "ReGround: Improving Textual and Spatial Grounding at No Cost" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "type": "text", + "content": "each layer of the U-Net, the model enables the placement of the entity specified in the text label " + }, + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "type": "text", + "content": " at the location indicated by the bounding box " + }, + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "type": "inline_equation", + "content": "b_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "type": "text", + "content": ". Note that the integration of gated self-attention does not require training the network from scratch or fine-tuning it, but can be accomplished simply by training the gated self-attention parameters while keeping all other parameters in the backbone model frozen." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 190, + 482, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 482, + 228 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 482, + 228 + ], + "type": "text", + "content": "Alg. 1 shows the pseudocode of the U-Net forward-pass including the plug-in of gated self-attention in line 6. Note that " + }, + { + "bbox": [ + 130, + 190, + 482, + 228 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 130, + 190, + 482, + 228 + ], + "type": "text", + "content": " is set to 1 for GLIGEN. If " + }, + { + "bbox": [ + 130, + 190, + 482, + 228 + ], + "type": "inline_equation", + "content": "\\beta_{t} = 0" + }, + { + "bbox": [ + 130, + 190, + 482, + 228 + ], + "type": "text", + "content": ", the algorithm is identical to that of LDM [41]." + } + ] + } + ], + "index": 3 + }, + { + "type": "code", + "bbox": [ + 123, + 269, + 482, + 424 + ], + "blocks": [ + { + "bbox": [ + 133, + 255, + 423, + 267 + ], + "lines": [ + { + "bbox": [ + 133, + 255, + 423, + 267 + ], + "spans": [ + { + "bbox": [ + 133, + 255, + 423, + 267 + ], + "type": "text", + "content": "Algorithm 1: Noise Prediction U-Net with Gated Self-Attention." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 123, + 269, + 482, + 424 + ], + "lines": [ + { + "bbox": [ + 123, + 269, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 123, + 269, + 482, + 424 + ], + "type": "text", + "content": "Parameters: " + }, + { + "bbox": [ + 123, + 269, + 482, + 424 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 123, + 269, + 482, + 424 + ], + "type": "text", + "content": " // Weight for GSA. Inputs: " + }, + { + "bbox": [ + 123, + 269, + 482, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t, c, \\{g_i\\}_{i=0\\dots N-1}" + }, + { + "bbox": [ + 123, + 269, + 482, + 424 + ], + "type": "text", + "content": "; // Noisy data at timestep " + }, + { + "bbox": [ + 123, + 269, + 482, + 424 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 123, + 269, + 482, + 424 + ], + "type": "text", + "content": ", text condition, and grounding tokens \nOutputs: " + }, + { + "bbox": [ + 123, + 269, + 482, + 424 + ], + "type": "inline_equation", + "content": "\\epsilon_t" + }, + { + "bbox": [ + 123, + 269, + 482, + 424 + ], + "type": "text", + "content": "; // Noise at timestep " + }, + { + "bbox": [ + 123, + 269, + 482, + 424 + ], + "type": "inline_equation", + "content": "t-1" + }, + { + "bbox": [ + 123, + 269, + 482, + 424 + ], + "type": "text", + "content": ". \n1 Function U-Net " + }, + { + "bbox": [ + 123, + 269, + 482, + 424 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_t, c, \\{g_i\\})" + }, + { + "bbox": [ + 123, + 269, + 482, + 424 + ], + "type": "text", + "content": ": \n2 " + }, + { + "bbox": [ + 123, + 269, + 482, + 424 + ], + "type": "inline_equation", + "content": "F \\gets \\mathbf{x}_t" + }, + { + "bbox": [ + 123, + 269, + 482, + 424 + ], + "type": "text", + "content": " \n3 for " + }, + { + "bbox": [ + 123, + 269, + 482, + 424 + ], + "type": "inline_equation", + "content": "i = 0, \\dots, L-1" + }, + { + "bbox": [ + 123, + 269, + 482, + 424 + ], + "type": "text", + "content": " do \n4 " + }, + { + "bbox": [ + 123, + 269, + 482, + 424 + ], + "type": "inline_equation", + "content": "\\begin{array}{c} F_{\\mathrm{RS}} \\gets \\mathrm{Conv}(F) + F; \\\\ F_{\\mathrm{SA}} \\gets \\mathrm{SA}(F_{\\mathrm{RS}}) + F_{\\mathrm{RS}}; \\\\ F_{\\mathrm{GSA}} \\gets \\beta_t \\cdot \\mathrm{GSA}(F_{\\mathrm{SA}}, \\{g_i\\}) + F_{\\mathrm{SA}}; \\\\ F \\gets \\mathrm{CA}(F_{\\mathrm{GSA}}, c) + F_{\\mathrm{GSA}}; \\end{array}" + }, + { + "bbox": [ + 123, + 269, + 482, + 424 + ], + "type": "text", + "content": " // Residual block. \n5 // Self-Attention module. \n6 \n7 \n8 " + }, + { + "bbox": [ + 123, + 269, + 482, + 424 + ], + "type": "inline_equation", + "content": "\\epsilon_t \\gets F;" + }, + { + "bbox": [ + 123, + 269, + 482, + 424 + ], + "type": "text", + "content": " \n9 return " + }, + { + "bbox": [ + 123, + 269, + 482, + 424 + ], + "type": "inline_equation", + "content": "\\epsilon_t" + }, + { + "bbox": [ + 123, + 269, + 482, + 424 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "code_body" + } + ], + "index": 5, + "sub_type": "algorithm" + }, + { + "bbox": [ + 132, + 479, + 268, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 268, + 492 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 268, + 492 + ], + "type": "text", + "content": "4.2 Description Omission" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": "Despite its high accuracy in spatial grounding, GLIGEN [28] frequently struggles to capture essential attributes specified in the input text prompt. As illustrated in Fig. 3, the leftmost image shows \"a person\" and \"a skateboard\" accurately placed in their designated regions. However, a critical detail from the input text prompt, \"black and white photography\", is absent in the output image. This discrepancy often emerges when the input comprises distinct but equally important descriptions regarding the image, presented through text prompts and bounding boxes. Such omissions not only fail to convey the stylistic intent of the image but also tend to overlook significant objects mentioned within the text prompt. Additional examples of this problem are showcased in Fig. 1, where the second row demonstrates the absence of a \"blanket\" in the generated image, a key element from the text prompt. This limitation significantly hampers GLIGEN's fidelity to user-provided text prompts, a challenge we term as description omission." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "text", + "content": "P.Y. Lee and M. Sung" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 203 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 203 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 203 + ], + "type": "image", + "image_path": "ddcbaa8b7839a288d6088cc2cee6e9a4d8c2f443f2be629bd09086e49746d7cd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 211, + 482, + 289 + ], + "lines": [ + { + "bbox": [ + 130, + 211, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 211, + 482, + 289 + ], + "type": "text", + "content": "Fig. 3: (a) Images generated by GLIGEN [28] with varying activation duration of gated self-attention " + }, + { + "bbox": [ + 130, + 211, + 482, + 289 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 211, + 482, + 289 + ], + "type": "text", + "content": " in scheduled sampling (Sec. 5.1). The red words in the text prompt denote the words used as labels of the input bounding boxes. Note that for GLIGEN to reflect the underlined description in the text prompt in the final image, " + }, + { + "bbox": [ + 130, + 211, + 482, + 289 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 211, + 482, + 289 + ], + "type": "text", + "content": " must be decreased to 0.1, which compromises spatial grounding accuracy. (b) In contrast, our REGROUND reflects the underlined phrase even when " + }, + { + "bbox": [ + 130, + 211, + 482, + 289 + ], + "type": "inline_equation", + "content": "\\gamma = 1.0" + }, + { + "bbox": [ + 130, + 211, + 482, + 289 + ], + "type": "text", + "content": ", therefore achieving high accuracy in both textual and spatial grounding." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 312, + 394, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 394, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 394, + 326 + ], + "type": "text", + "content": "5 ReGround: Rewiring Attention Modules" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 337, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 337, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 337, + 482, + 434 + ], + "type": "text", + "content": "Gated self-attention and cross-attention each play a crucial role in enabling spatial and textual groundings, by taking bounding boxes and text prompts as inputs, respectively. To tackle the issue of description omission, we first examine the impact of attention modules on the groundings they do not address: the effect of gated self-attention on textual grounding (Sec. 5.1), and the influence of cross-attention on spatial grounding (Sec. 5.2). Building on this analysis, we propose an approach for network reconfiguration, modifying the connections among self-attention, gated self-attention, and cross-attention modules (Sec. 5.3)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 451, + 434, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 451, + 434, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 451, + 434, + 464 + ], + "type": "text", + "content": "5.1 Impact of Gated Self-Attention on Textual Grounding" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 472, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 556 + ], + "type": "text", + "content": "As the issue of description omission arises due to the newly added gated self-attention in GLIGEN [28], we first attempt to mitigate the impact of gated self-attention by using scheduled sampling [28], activating gated self-attention only in a few initial steps of the denoising process. This approach is inspired by the observation that the coarse structure of the final image is established within the first few denoising steps. The scheduling is applied by setting the weight of gated self-attention " + }, + { + "bbox": [ + 130, + 472, + 482, + 556 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 130, + 472, + 482, + 556 + ], + "type": "text", + "content": " (line 6 of Alg. 1) as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 257, + 564, + 481, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 564, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 257, + 564, + 481, + 597 + ], + "type": "interline_equation", + "content": "\\beta_ {t} = \\left\\{ \\begin{array}{l l} 1 & (t \\leq \\gamma \\cdot T) \\\\ 0 & (t > \\gamma \\cdot T), \\end{array} \\right. \\tag {1}", + "image_path": "4cbff2cc13a5f7cebefc0808224c9b9eb492fa286fb8d310fc93c1832128f34f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 605, + 481, + 628 + ], + "type": "inline_equation", + "content": "\\gamma \\in [0,1]" + }, + { + "bbox": [ + 130, + 605, + 481, + 628 + ], + "type": "text", + "content": " represents the fraction of the initial denoising steps to activate gated self-attention." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Fig. 3-(a) shows an example of generated images while incrementally adjusting " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": " from 1.0 to 0.0. As " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": " is reduced from 1.0 to 0.0, the details specified in the text prompt, \"a black and white photograph\", begin to be reflected starting" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 180, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 180, + 91, + 448, + 102 + ], + "type": "text", + "content": "ReGround: Improving Textual and Spatial Grounding at No Cost" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "at " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\gamma = 0.1" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ", demonstrating that longer activation of gated self-attention may interfere with the alignment of the output image with the text prompt. However, as gated self-attention is activated for shorter durations, the spatial grounding diminishes, as shown in the objects' reduced alignment with the input bounding boxes. This phenomenon illustrates the inherent trade-off between spatial and textual grounding, which cannot be resolved by controlling the duration of gated self-attention activation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 217, + 406, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 217, + 406, + 230 + ], + "spans": [ + { + "bbox": [ + 131, + 217, + 406, + 230 + ], + "type": "text", + "content": "5.2 Impact of Cross-Attention on Spatial Grounding" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 238, + 479, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 238, + 479, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 238, + 479, + 297 + ], + "type": "text", + "content": "We also investigate whether cross-attention has influence on spatial grounding. For this, we conduct a toy experiment by removing cross-attention modules in GLIGEN [28], allowing the output of the gated self-attention to be directly passed to the next layer of the U-Net. This modification is equivalent to changing line 7 of Alg. 1 to " + }, + { + "bbox": [ + 130, + 238, + 479, + 297 + ], + "type": "inline_equation", + "content": "F \\gets F_{GSA}" + }, + { + "bbox": [ + 130, + 238, + 479, + 297 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 297, + 482, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 381 + ], + "type": "text", + "content": "The results are displayed in Fig. 4. Note that, while the appearance of the background and objects changes, the silhouettes of the cat (left) and the individuals (right) remain precisely positioned within their respective bounding boxes without cross-attention. This observation indicates that while gated self-attention that is performed before cross-attention may compromise textual grounding, cross-attention that processes the output of gated self-attention does not affect spatial grounding." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 133, + 400, + 473, + 466 + ], + "blocks": [ + { + "bbox": [ + 133, + 400, + 473, + 466 + ], + "lines": [ + { + "bbox": [ + 133, + 400, + 473, + 466 + ], + "spans": [ + { + "bbox": [ + 133, + 400, + 473, + 466 + ], + "type": "image", + "image_path": "fdb30a9ad4c464208741327c14aee4da75f9ac9ba57496b212093f5bba5aacb4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 475, + 482, + 520 + ], + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 520 + ], + "type": "text", + "content": "Fig. 4: Comparison of the output of GLIGEN [28] with and without cross-attention. While the absence of cross-attention reduces realism and quality of the image, the silhouette of objects remains grounded within the given bounding boxes, as shown in the third column of each case." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 562, + 398, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 562, + 398, + 574 + ], + "spans": [ + { + "bbox": [ + 131, + 562, + 398, + 574 + ], + "type": "text", + "content": "5.3 Network Rewiring: From Sequential to Parallel" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Building on the analyses above, we propose a simple yet effective modification to the grounding mechanism, changing the relationship between gated self-attention and cross-attention from sequential to parallel. This change eliminates the placement of gated self-attention before cross-attention, thus preventing the reduction of text grounding caused by gated self-attention. Moreover, in this parallel arrangement, the preservation of spatial grounding is assured, as gated self-attention for spatial grounding does not require subsequent cross-attention." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "text", + "content": "P.Y. Lee and M. Sung" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "Specifically, recall that in GLIGEN [28], the output of gated self-attention is added to the residual from self-attention, which is then passed to the cross-attention module as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 237, + 155, + 378, + 173 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 155, + 378, + 173 + ], + "spans": [ + { + "bbox": [ + 237, + 155, + 378, + 173 + ], + "type": "interline_equation", + "content": "F _ {G S A} \\gets \\underbrace {\\mathrm {G S A} (F _ {S A} , \\{g _ {i} \\})} + F _ {S A};", + "image_path": "7d960355d18e45b1e0525c96a8559f37a25bd2bc1558aead2b5b848bf414d174.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 279, + 175, + 345, + 183 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 175, + 345, + 183 + ], + "spans": [ + { + "bbox": [ + 279, + 175, + 345, + 183 + ], + "type": "interline_equation", + "content": "s p a t i a l g r o u n d i n g", + "image_path": "e691f3649efd06935a57176bbe159ab76c6d636a8cf94d6e067260aa5df36a83.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 250, + 180, + 480, + 202 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 180, + 480, + 202 + ], + "spans": [ + { + "bbox": [ + 250, + 180, + 480, + 202 + ], + "type": "interline_equation", + "content": "F \\leftarrow \\underbrace {\\operatorname {C A} \\left(F _ {G S A} , c\\right)} + F _ {G S A}; \\tag {2}", + "image_path": "2014a7bc2c6fd38bdfd9ef87a215d9658e3504a798ac9b390a3028b54fd43c21.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 274, + 204, + 341, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 204, + 341, + 213 + ], + "spans": [ + { + "bbox": [ + 274, + 204, + 341, + 213 + ], + "type": "interline_equation", + "content": "\\text {t e x t u a l g r o u n d i n g}", + "image_path": "e92b538e98fb130f0c2f769bfbbdd827c1cbe5cb98f6699d9acbfcc96b62777d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 217, + 479, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 217, + 479, + 240 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 479, + 240 + ], + "type": "text", + "content": "We propose to transform this sequence grounding pipeline into two parallel processes as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 200, + 247, + 480, + 264 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 247, + 480, + 264 + ], + "spans": [ + { + "bbox": [ + 200, + 247, + 480, + 264 + ], + "type": "interline_equation", + "content": "F \\leftarrow \\underbrace {\\operatorname {G S A} \\left(F _ {S A} , \\{g _ {i} \\}\\right)} + \\underbrace {\\operatorname {C A} \\left(F _ {S A} , c\\right)} + \\underbrace {F _ {S A}}; \\tag {3}", + "image_path": "bb5109fb76535625aae4f874a02cf32d1de8e326bc2dc3f4651394c43e29bf51.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 225, + 265, + 412, + 274 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 265, + 412, + 274 + ], + "spans": [ + { + "bbox": [ + 225, + 265, + 412, + 274 + ], + "type": "interline_equation", + "content": "\\begin{array}{c c c} \\text {s p a t i a l g r o u n d i n g} & \\text {t e x t u a l g r o u n d i n g} & \\text {r e s i d u a l} \\end{array}", + "image_path": "62b7e6256a837be4aefbbffd75528a3e25834d3760bc8044c80c99a7d41be688.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 279, + 479, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 479, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 479, + 326 + ], + "type": "text", + "content": "Refer to Fig. 2 for the visualization of network architecture changes " + }, + { + "bbox": [ + 130, + 279, + 479, + 326 + ], + "type": "inline_equation", + "content": "((\\mathrm{b})\\rightarrow" + }, + { + "bbox": [ + 130, + 279, + 479, + 326 + ], + "type": "text", + "content": " (c)). This network rewiring is feasible because the input to gated self-attention remains unchanged, while the input to cross-attention shifts to " + }, + { + "bbox": [ + 130, + 279, + 479, + 326 + ], + "type": "inline_equation", + "content": "F_{SA}" + }, + { + "bbox": [ + 130, + 279, + 479, + 326 + ], + "type": "text", + "content": ", for which it was originally designed in the context of Latent Diffusion Models [41]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 327, + 480, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 480, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 480, + 409 + ], + "type": "text", + "content": "It is important to note that the modification is effective even when applied to the pretrained GLIGEN, which was trained with the sequential structure of the attention modules. Therefore, our rewiring does not require any additional training or fine-tuning, introduces no extra parameters, and does not affect computation time or memory usage during the generation process. The only requirement is the simple reconfiguration of the attention modules at inference time." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 426, + 230, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 426, + 230, + 438 + ], + "spans": [ + { + "bbox": [ + 131, + 426, + 230, + 438 + ], + "type": "text", + "content": "6 Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 447, + 479, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 479, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 479, + 495 + ], + "type": "text", + "content": "In this section, we show the effectiveness of our REGROUND by evaluating the spatial grounding on existing layout-caption datasets [12, 30] and the textual grounding on text-image alignment metrics [26, 38]. We use the official GLI-GEN [28] checkpoint which is trained based on Stable Diffusion v1.4 [41]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 510, + 205, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 510, + 205, + 520 + ], + "spans": [ + { + "bbox": [ + 131, + 510, + 205, + 520 + ], + "type": "text", + "content": "6.1 Datasets" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 525, + 479, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 479, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 479, + 574 + ], + "type": "text", + "content": "MS-COCO. We use the validations sets of both MS-COCO-2014 and MS-COCO-2017 datasets [30]. Each dataset provides image-captions pairs and the " + }, + { + "bbox": [ + 130, + 525, + 479, + 574 + ], + "type": "inline_equation", + "content": "xy" + }, + { + "bbox": [ + 130, + 525, + 479, + 574 + ], + "type": "text", + "content": "-coordinates of bounding boxes along with their corresponding object categories." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 581, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 480, + 666 + ], + "type": "text", + "content": "NSR-1K-GPT. We also use the NSR-1K benchmark [12] for evaluation. Based on each subset of NSR-1K—Counting and Spatial—we develop a new benchmark, NSR-1K-GPT, augmenting each original caption in NSR-1K using GPT-4 [34]. The instructions for augmentation are to (i) elaborate on the descriptions of each mentioned entity and (ii) provide additional details about the background of the image. More details on the evaluation datasets are provided in the Supplementary (Sec. S1)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 180, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 180, + 91, + 448, + 103 + ], + "type": "text", + "content": "ReGround: Improving Textual and Spatial Grounding at No Cost" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 123, + 201, + 190 + ], + "blocks": [ + { + "bbox": [ + 151, + 114, + 183, + 122 + ], + "lines": [ + { + "bbox": [ + 151, + 114, + 183, + 122 + ], + "spans": [ + { + "bbox": [ + 151, + 114, + 183, + 122 + ], + "type": "text", + "content": "Layout" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 134, + 123, + 201, + 190 + ], + "lines": [ + { + "bbox": [ + 134, + 123, + 201, + 190 + ], + "spans": [ + { + "bbox": [ + 134, + 123, + 201, + 190 + ], + "type": "image", + "image_path": "fcb2b6d78cd34047407c1f0585f925c33cc7af3307104afe66252414890e1c38.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 202, + 123, + 269, + 190 + ], + "blocks": [ + { + "bbox": [ + 229, + 114, + 242, + 122 + ], + "lines": [ + { + "bbox": [ + 229, + 114, + 242, + 122 + ], + "spans": [ + { + "bbox": [ + 229, + 114, + 242, + 122 + ], + "type": "text", + "content": "SD" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 202, + 123, + 269, + 190 + ], + "lines": [ + { + "bbox": [ + 202, + 123, + 269, + 190 + ], + "spans": [ + { + "bbox": [ + 202, + 123, + 269, + 190 + ], + "type": "image", + "image_path": "a675bdb566510ebf7cb0b1615466b8a93bc97b7a7346aac8e5f451ffcfded1b8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 191, + 432, + 201 + ], + "lines": [ + { + "bbox": [ + 179, + 191, + 432, + 201 + ], + "spans": [ + { + "bbox": [ + 179, + 191, + 432, + 201 + ], + "type": "text", + "content": "\"A woman in a white shirt standing in front of a fence smiling.\"" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 271, + 123, + 338, + 190 + ], + "blocks": [ + { + "bbox": [ + 276, + 114, + 334, + 123 + ], + "lines": [ + { + "bbox": [ + 276, + 114, + 334, + 123 + ], + "spans": [ + { + "bbox": [ + 276, + 114, + 334, + 123 + ], + "type": "text", + "content": "GLIGEN" + }, + { + "bbox": [ + 276, + 114, + 334, + 123 + ], + "type": "inline_equation", + "content": "_{\\gamma = 1.0}" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 271, + 123, + 338, + 190 + ], + "lines": [ + { + "bbox": [ + 271, + 123, + 338, + 190 + ], + "spans": [ + { + "bbox": [ + 271, + 123, + 338, + 190 + ], + "type": "image", + "image_path": "d7c222f0bd883054fe1d39fae5abb749c32032149f5aaab54981d698f2458d76.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 340, + 123, + 406, + 190 + ], + "blocks": [ + { + "bbox": [ + 345, + 114, + 402, + 123 + ], + "lines": [ + { + "bbox": [ + 345, + 114, + 402, + 123 + ], + "spans": [ + { + "bbox": [ + 345, + 114, + 402, + 123 + ], + "type": "text", + "content": "GLIGEN" + }, + { + "bbox": [ + 345, + 114, + 402, + 123 + ], + "type": "inline_equation", + "content": "_{\\gamma=0.2}" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 340, + 123, + 406, + 190 + ], + "lines": [ + { + "bbox": [ + 340, + 123, + 406, + 190 + ], + "spans": [ + { + "bbox": [ + 340, + 123, + 406, + 190 + ], + "type": "image", + "image_path": "845b0ae98dbadca274469004f638a143c20fe5a1e36d67c4918fd54e3588b82c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 408, + 123, + 476, + 190 + ], + "blocks": [ + { + "bbox": [ + 417, + 114, + 467, + 122 + ], + "lines": [ + { + "bbox": [ + 417, + 114, + 467, + 122 + ], + "spans": [ + { + "bbox": [ + 417, + 114, + 467, + 122 + ], + "type": "text", + "content": "ReGround" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 408, + 123, + 476, + 190 + ], + "lines": [ + { + "bbox": [ + 408, + 123, + 476, + 190 + ], + "spans": [ + { + "bbox": [ + 408, + 123, + 476, + 190 + ], + "type": "image", + "image_path": "1375ee709e14aa8fb2bd13955715c7044478935ac6fcb298a8d5843676715f05.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 134, + 202, + 201, + 270 + ], + "blocks": [ + { + "bbox": [ + 134, + 202, + 201, + 270 + ], + "lines": [ + { + "bbox": [ + 134, + 202, + 201, + 270 + ], + "spans": [ + { + "bbox": [ + 134, + 202, + 201, + 270 + ], + "type": "image", + "image_path": "dad2d1026f04d0c6478c87e2d428c721673546bdf6a61026ba3ef059829366b4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 271, + 425, + 282 + ], + "lines": [ + { + "bbox": [ + 187, + 271, + 425, + 282 + ], + "spans": [ + { + "bbox": [ + 187, + 271, + 425, + 282 + ], + "type": "text", + "content": "\"One spotted cow in the photo, [...] under a clear blue sky.\"" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 202, + 202, + 269, + 270 + ], + "blocks": [ + { + "bbox": [ + 202, + 202, + 269, + 270 + ], + "lines": [ + { + "bbox": [ + 202, + 202, + 269, + 270 + ], + "spans": [ + { + "bbox": [ + 202, + 202, + 269, + 270 + ], + "type": "image", + "image_path": "2f6994c3ce78575f1cbcebfa08d16db42204feeab6b5d351a24e504001fb1ae9.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 271, + 202, + 338, + 270 + ], + "blocks": [ + { + "bbox": [ + 271, + 202, + 338, + 270 + ], + "lines": [ + { + "bbox": [ + 271, + 202, + 338, + 270 + ], + "spans": [ + { + "bbox": [ + 271, + 202, + 338, + 270 + ], + "type": "image", + "image_path": "94a994ac988c5dff30ab9111a6f278f4a51f117fe942d9050ed1435c81e5bd89.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 340, + 202, + 407, + 270 + ], + "blocks": [ + { + "bbox": [ + 340, + 202, + 407, + 270 + ], + "lines": [ + { + "bbox": [ + 340, + 202, + 407, + 270 + ], + "spans": [ + { + "bbox": [ + 340, + 202, + 407, + 270 + ], + "type": "image", + "image_path": "9e03c7250da3b602519b0b60fa353d6ba1132df0400b9a0a0095eaa72fc4be79.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 409, + 202, + 476, + 270 + ], + "blocks": [ + { + "bbox": [ + 409, + 202, + 476, + 270 + ], + "lines": [ + { + "bbox": [ + 409, + 202, + 476, + 270 + ], + "spans": [ + { + "bbox": [ + 409, + 202, + 476, + 270 + ], + "type": "image", + "image_path": "cb715a930d882e537f9385274bf29e78ae31fff45227fe614fce9bef8e5cd995.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 134, + 284, + 201, + 351 + ], + "blocks": [ + { + "bbox": [ + 134, + 284, + 201, + 351 + ], + "lines": [ + { + "bbox": [ + 134, + 284, + 201, + 351 + ], + "spans": [ + { + "bbox": [ + 134, + 284, + 201, + 351 + ], + "type": "image", + "image_path": "62f7e592193a8e79bfdf41c1b89b4b9c9ae87aa64ba3a9cf87093b935c77522b.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 202, + 284, + 269, + 350 + ], + "blocks": [ + { + "bbox": [ + 202, + 284, + 269, + 350 + ], + "lines": [ + { + "bbox": [ + 202, + 284, + 269, + 350 + ], + "spans": [ + { + "bbox": [ + 202, + 284, + 269, + 350 + ], + "type": "image", + "image_path": "f705d3015702df30eca3851dd4a46b196a401ae91c38651b6cd7169323368348.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 271, + 284, + 338, + 350 + ], + "blocks": [ + { + "bbox": [ + 271, + 284, + 338, + 350 + ], + "lines": [ + { + "bbox": [ + 271, + 284, + 338, + 350 + ], + "spans": [ + { + "bbox": [ + 271, + 284, + 338, + 350 + ], + "type": "image", + "image_path": "af32fff613c22d2d3709ebb5026de943aa5f418e18254eed378b0f0d30fe5a27.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 340, + 284, + 407, + 350 + ], + "blocks": [ + { + "bbox": [ + 340, + 284, + 407, + 350 + ], + "lines": [ + { + "bbox": [ + 340, + 284, + 407, + 350 + ], + "spans": [ + { + "bbox": [ + 340, + 284, + 407, + 350 + ], + "type": "image", + "image_path": "9562212cbb0b7c310b47e1d9cd7fc5ceda3db46108367714a913f0afd80d290b.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 409, + 284, + 476, + 350 + ], + "blocks": [ + { + "bbox": [ + 409, + 284, + 476, + 350 + ], + "lines": [ + { + "bbox": [ + 409, + 284, + 476, + 350 + ], + "spans": [ + { + "bbox": [ + 409, + 284, + 476, + 350 + ], + "type": "image", + "image_path": "b378e9f4021dd9edecf2acd8d107137fab8f87545517f4792bfc3617cad4edbe.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 134, + 364, + 201, + 432 + ], + "blocks": [ + { + "bbox": [ + 134, + 364, + 201, + 432 + ], + "lines": [ + { + "bbox": [ + 134, + 364, + 201, + 432 + ], + "spans": [ + { + "bbox": [ + 134, + 364, + 201, + 432 + ], + "type": "image", + "image_path": "eb110959bbfc8a2de9baf1402357a140b75c052aef93e50041a9b86aac63f924.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 202, + 364, + 269, + 431 + ], + "blocks": [ + { + "bbox": [ + 215, + 353, + 396, + 363 + ], + "lines": [ + { + "bbox": [ + 215, + 353, + 396, + 363 + ], + "spans": [ + { + "bbox": [ + 215, + 353, + 396, + 363 + ], + "type": "text", + "content": "\"A foreign airline status board with a clock.\"" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 202, + 364, + 269, + 431 + ], + "lines": [ + { + "bbox": [ + 202, + 364, + 269, + 431 + ], + "spans": [ + { + "bbox": [ + 202, + 364, + 269, + 431 + ], + "type": "image", + "image_path": "8bab439c32472a70460c103a943b9b263d145fe58d8940e04743278b5e3ed4fb.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 271, + 364, + 338, + 431 + ], + "blocks": [ + { + "bbox": [ + 271, + 364, + 338, + 431 + ], + "lines": [ + { + "bbox": [ + 271, + 364, + 338, + 431 + ], + "spans": [ + { + "bbox": [ + 271, + 364, + 338, + 431 + ], + "type": "image", + "image_path": "1cd6360676e371ed83e2c8bd664bee1913fbdd9a273adf39b9b8dddb5be882b2.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 340, + 364, + 407, + 431 + ], + "blocks": [ + { + "bbox": [ + 340, + 364, + 407, + 431 + ], + "lines": [ + { + "bbox": [ + 340, + 364, + 407, + 431 + ], + "spans": [ + { + "bbox": [ + 340, + 364, + 407, + 431 + ], + "type": "image", + "image_path": "810c30771373b86f7389c53a4ad513a7f01be3ba1d5b297512fd544384cb2f46.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 409, + 364, + 476, + 431 + ], + "blocks": [ + { + "bbox": [ + 409, + 364, + 476, + 431 + ], + "lines": [ + { + "bbox": [ + 409, + 364, + 476, + 431 + ], + "spans": [ + { + "bbox": [ + 409, + 364, + 476, + 431 + ], + "type": "image", + "image_path": "78fa721a1debaec8c10915a39756a78e4f6d754978ef2add0ef5fb711569d219.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 134, + 446, + 201, + 514 + ], + "blocks": [ + { + "bbox": [ + 186, + 434, + 425, + 445 + ], + "lines": [ + { + "bbox": [ + 186, + 434, + 425, + 445 + ], + "spans": [ + { + "bbox": [ + 186, + 434, + 425, + 445 + ], + "type": "text", + "content": "\"A palm tree shower curtain is shown in a small bathroom.\"" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 134, + 446, + 201, + 514 + ], + "lines": [ + { + "bbox": [ + 134, + 446, + 201, + 514 + ], + "spans": [ + { + "bbox": [ + 134, + 446, + 201, + 514 + ], + "type": "image", + "image_path": "afd685b2510da2353ac8f4552949cb28efea8873e1058f4ce767bc41dfc6ad9c.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 540, + 482, + 639 + ], + "lines": [ + { + "bbox": [ + 132, + 540, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 132, + 540, + 482, + 639 + ], + "type": "text", + "content": "Fig. 5: Qualitative comparisons. Stable Diffusion (SD, 2nd column) generates images that align with the given text descriptions, including the underlined phrase in each row, but cannot take bounding boxes as input. GLIGEN (3rd column) creates images that match the input layouts but suffers from description omission, failing to reflect the underlined descriptions. Scheduled sampling strategy (4th column) can partially address this issue (for instance, in the 5th row, where \"window\" appears in the room), but it results in a noticeable decline in spatial accuracy (as seen in the 1st row, where the tie is not generated). In contrast, our method (last column) accurately incorporates the underlined text descriptions while maintaining precise spatial representation." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 202, + 446, + 269, + 513 + ], + "blocks": [ + { + "bbox": [ + 202, + 446, + 269, + 513 + ], + "lines": [ + { + "bbox": [ + 202, + 446, + 269, + 513 + ], + "spans": [ + { + "bbox": [ + 202, + 446, + 269, + 513 + ], + "type": "image", + "image_path": "373851e8d28802d2552548f1c9b2bfd45495b7e8fa1c1b17b6ac2ca8d53eb0f2.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 515, + 399, + 526 + ], + "lines": [ + { + "bbox": [ + 212, + 515, + 399, + 526 + ], + "spans": [ + { + "bbox": [ + 212, + 515, + 399, + 526 + ], + "type": "text", + "content": "\"A view of a living room with large windows.\"" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 271, + 446, + 338, + 513 + ], + "blocks": [ + { + "bbox": [ + 271, + 446, + 338, + 513 + ], + "lines": [ + { + "bbox": [ + 271, + 446, + 338, + 513 + ], + "spans": [ + { + "bbox": [ + 271, + 446, + 338, + 513 + ], + "type": "image", + "image_path": "24a5edcf557748b4bfcb3a764ca63faa0441f2bd47b6c937e076fabf60447187.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 340, + 446, + 407, + 513 + ], + "blocks": [ + { + "bbox": [ + 340, + 446, + 407, + 513 + ], + "lines": [ + { + "bbox": [ + 340, + 446, + 407, + 513 + ], + "spans": [ + { + "bbox": [ + 340, + 446, + 407, + 513 + ], + "type": "image", + "image_path": "bdb1c850f1c5a9cde54fac9fc0c8ea268cde084ba689c1f4bacc568b5df350cf.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 409, + 446, + 476, + 513 + ], + "blocks": [ + { + "bbox": [ + 409, + 446, + 476, + 513 + ], + "lines": [ + { + "bbox": [ + 409, + 446, + 476, + 513 + ], + "spans": [ + { + "bbox": [ + 409, + 446, + 476, + 513 + ], + "type": "image", + "image_path": "e1a731f958013e7fcf4c2a0d618735081c0fb616071736486cb0c9cee096f54d.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "text", + "content": "P.Y. Lee and M. Sung" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 248, + 201 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 248, + 201 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 248, + 201 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 248, + 201 + ], + "type": "image", + "image_path": "81fb26cead795d08fca56286d50264e9abbb914b963c68023781836f7a44c233.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 201, + 230, + 211 + ], + "lines": [ + { + "bbox": [ + 157, + 201, + 230, + 211 + ], + "spans": [ + { + "bbox": [ + 157, + 201, + 230, + 211 + ], + "type": "text", + "content": "(a) MS-COCO-2014" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 220, + 482, + 277 + ], + "lines": [ + { + "bbox": [ + 131, + 220, + 482, + 277 + ], + "spans": [ + { + "bbox": [ + 131, + 220, + 482, + 277 + ], + "type": "text", + "content": "Fig. 6: Comparisons on MS-COCO [30] and NSR-1K-GPT. Each plot shows the relationship between textual grounding (i.e. CLIP score [15]) and spatial grounding (i.e. YOLO score [52]) accuracy of GLIGEN [28] and our method. Note that the plot of our REGROUND is positioned in the top-right quadrant relative to GLIGEN, signifying that it alleviates the inherent trade-off between textual and spatial grounding." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 248, + 114, + 362, + 201 + ], + "blocks": [ + { + "bbox": [ + 248, + 114, + 362, + 201 + ], + "lines": [ + { + "bbox": [ + 248, + 114, + 362, + 201 + ], + "spans": [ + { + "bbox": [ + 248, + 114, + 362, + 201 + ], + "type": "image", + "image_path": "7ce3c5555acc970e8f7cfd00eb0d5d73356fad84ded5b23a500cb609be25f85d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 273, + 201, + 346, + 211 + ], + "lines": [ + { + "bbox": [ + 273, + 201, + 346, + 211 + ], + "spans": [ + { + "bbox": [ + 273, + 201, + 346, + 211 + ], + "type": "text", + "content": "(b) MS-COCO-2017" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 364, + 114, + 481, + 201 + ], + "blocks": [ + { + "bbox": [ + 364, + 114, + 481, + 201 + ], + "lines": [ + { + "bbox": [ + 364, + 114, + 481, + 201 + ], + "spans": [ + { + "bbox": [ + 364, + 114, + 481, + 201 + ], + "type": "image", + "image_path": "ff51c7df2da5ef955b127e09409a2c0cf616d3cf16eb22df3b912091f914057f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 201, + 474, + 212 + ], + "lines": [ + { + "bbox": [ + 378, + 201, + 474, + 212 + ], + "spans": [ + { + "bbox": [ + 378, + 201, + 474, + 212 + ], + "type": "text", + "content": "(c) NSR-1K-GPT-Counting" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 299, + 257, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 299, + 257, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 299, + 257, + 310 + ], + "type": "text", + "content": "6.2 Evaluation Metrics" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 318, + 481, + 437 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 138, + 318, + 481, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 318, + 481, + 366 + ], + "spans": [ + { + "bbox": [ + 138, + 318, + 481, + 366 + ], + "type": "text", + "content": "- YOLO score: Spatial grounding accuracy is assessed usingYOLO score [52]. We employYOLOv7 [52] to detect objects in each generated image and compute the average precision (AP) based on the ground truth bounding box annotations from MS-COCO [30]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 366, + 481, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 366, + 481, + 378 + ], + "spans": [ + { + "bbox": [ + 138, + 366, + 481, + 378 + ], + "type": "text", + "content": "- CLIP score: Textual grounding accuracy is assessed using CLIP score [15]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 378, + 427, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 378, + 427, + 389 + ], + "spans": [ + { + "bbox": [ + 138, + 378, + 427, + 389 + ], + "type": "text", + "content": "- FID: Image quality and diversity are evaluated using FID [16]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 389, + 481, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 389, + 481, + 437 + ], + "spans": [ + { + "bbox": [ + 138, + 389, + 481, + 437 + ], + "type": "text", + "content": "- User Study and PickScore: We conduct a user study to assess human preferences for the generated images based on each input text prompt. Additionally, we use PickScore [26], a human preference predictor, to further analyze the results." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 455, + 297, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 455, + 297, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 455, + 297, + 467 + ], + "type": "text", + "content": "6.3 Comparison with GLIGEN" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 474, + 481, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 474, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 131, + 474, + 481, + 521 + ], + "type": "text", + "content": "Textual-Spatial Grounding Trade-off. We first examine the trade-off between textual and spatial groundings for both GLIGEN [28] and our REGROUND, the rewired version of GLIGEN, while varying the scheduled sampling parameter " + }, + { + "bbox": [ + 131, + 474, + 481, + 521 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 131, + 474, + 481, + 521 + ], + "type": "text", + "content": " from 1.0 to 0.1." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 521, + 482, + 666 + ], + "type": "text", + "content": "Fig. 6-(a), (b) present the graphs of CLIP score [15] and YOLO score [52] measured on the MS-COCO datasets [30]. In MS-COCO-2014, when reducing " + }, + { + "bbox": [ + 131, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 131, + 521, + 482, + 666 + ], + "type": "text", + "content": " from 1.0 to 0.1, the CLIP score of GLIGEN varies from 30.44 to 31.65, while the YOLO score significantly drops from 58.13 to 22.75 (red in Fig. 6). In contrast, our REGROUND, (blue in Fig. 6), demonstrates a notably superior trade-off between textual and spatial grounding. Specifically, with " + }, + { + "bbox": [ + 131, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 131, + 521, + 482, + 666 + ], + "type": "text", + "content": " set to 1.0, REGROUND already achieves a CLIP score of 31.29, accounting for " + }, + { + "bbox": [ + 131, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "70.25\\%" + }, + { + "bbox": [ + 131, + 521, + 482, + 666 + ], + "type": "text", + "content": " of GLIGEN's total improvement in CLIP score when " + }, + { + "bbox": [ + 131, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 131, + 521, + 482, + 666 + ], + "type": "text", + "content": " is reduced from 1.0 to 0.1. Despite this significant increase in CLIP score, the YOLO score remains largely unchanged, marking 56.96 which represents only a " + }, + { + "bbox": [ + 131, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "3.31\\%" + }, + { + "bbox": [ + 131, + 521, + 482, + 666 + ], + "type": "text", + "content": " decrease in the range of YOLO score variation for GLIGEN when " + }, + { + "bbox": [ + 131, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 131, + 521, + 482, + 666 + ], + "type": "text", + "content": " is adjusted from 1.0 to 0.1. Moreover, when varying the " + }, + { + "bbox": [ + 131, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 131, + 521, + 482, + 666 + ], + "type": "text", + "content": ", the plot for REGROUND (blue) is constantly on the upper right" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 180, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 180, + 91, + 448, + 102 + ], + "type": "text", + "content": "ReGround: Improving Textual and Spatial Grounding at No Cost" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 227, + 200 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 227, + 200 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 227, + 200 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 227, + 200 + ], + "type": "image", + "image_path": "e4a8403d8740b291024a3725deb0df83fa19d27b12e2ccda242132828f68f2c2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 201, + 392, + 211 + ], + "lines": [ + { + "bbox": [ + 220, + 201, + 392, + 211 + ], + "spans": [ + { + "bbox": [ + 220, + 201, + 392, + 211 + ], + "type": "text", + "content": "\"A young child looking at a birthday cupcake\"" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 228, + 115, + 304, + 200 + ], + "blocks": [ + { + "bbox": [ + 228, + 115, + 304, + 200 + ], + "lines": [ + { + "bbox": [ + 228, + 115, + 304, + 200 + ], + "spans": [ + { + "bbox": [ + 228, + 115, + 304, + 200 + ], + "type": "image", + "image_path": "69cd149967164ffc044a8dd9f130cf352eb13e9b08d9609a1f8863e110fa12ef.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 204, + 212, + 251, + 222 + ], + "lines": [ + { + "bbox": [ + 204, + 212, + 251, + 222 + ], + "spans": [ + { + "bbox": [ + 204, + 212, + 251, + 222 + ], + "type": "text", + "content": "(a) GLIGEN" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 386, + 200 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 386, + 200 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 386, + 200 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 386, + 200 + ], + "type": "image", + "image_path": "b025b7e851fd264919931122ef1ef22f9f486b5aa6772388174f4e7efeb0c76a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 232, + 481, + 277 + ], + "lines": [ + { + "bbox": [ + 131, + 232, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 131, + 232, + 481, + 277 + ], + "type": "text", + "content": "Fig. 7: Generated images from the text prompt and bounding boxes from the MS-COCO-2017 (left of each column) and our COCO-Drop (right of each column). While GLIGEN [28] fails to generate \"a birthday cupcake\" when the corresponding bounding box is removed, our REGROUND successfully generates a cupcake on the table." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 387, + 115, + 462, + 200 + ], + "blocks": [ + { + "bbox": [ + 387, + 115, + 462, + 200 + ], + "lines": [ + { + "bbox": [ + 387, + 115, + 462, + 200 + ], + "spans": [ + { + "bbox": [ + 387, + 115, + 462, + 200 + ], + "type": "image", + "image_path": "5063241389f13d6b28a179318402b230ce64fea0ce326c25129d71e4b4c5b1b2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 212, + 413, + 222 + ], + "lines": [ + { + "bbox": [ + 361, + 212, + 413, + 222 + ], + "spans": [ + { + "bbox": [ + 361, + 212, + 413, + 222 + ], + "type": "text", + "content": "(b) ReGround" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 301, + 480, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 480, + 348 + ], + "type": "text", + "content": "side of GLIGEN (red), signifying a more advantageous trade-off across varying " + }, + { + "bbox": [ + 130, + 301, + 480, + 348 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 301, + 480, + 348 + ], + "type": "text", + "content": ". The same pattern is observed in MS-COCO-2017, where our REGROUND achieves " + }, + { + "bbox": [ + 130, + 301, + 480, + 348 + ], + "type": "inline_equation", + "content": "68.33\\%" + }, + { + "bbox": [ + 130, + 301, + 480, + 348 + ], + "type": "text", + "content": " of the increase in CLIP score of GLIGEN while only compromising YOLO score by " + }, + { + "bbox": [ + 130, + 301, + 480, + 348 + ], + "type": "inline_equation", + "content": "2.62\\%" + }, + { + "bbox": [ + 130, + 301, + 480, + 348 + ], + "type": "text", + "content": " compared to the decrease for GLIGEN." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 350, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 495 + ], + "type": "text", + "content": "Fig. 6-(c) further shows a quantitative comparison on the Counting subset of the newly generated NSR-1K-GPT benchmark. The plot reveals a consistent trend with the MS-COCO datasets. By reducing " + }, + { + "bbox": [ + 130, + 350, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 350, + 482, + 495 + ], + "type": "text", + "content": " from 1.0 to 0.1, GLIGEN's CLIP score is increased from 32.46 to 33.67, while the YOLO score is decreased from 65.36 to 26.38. In contrast, when " + }, + { + "bbox": [ + 130, + 350, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\gamma = 1.0" + }, + { + "bbox": [ + 130, + 350, + 482, + 495 + ], + "type": "text", + "content": ", REGROUND achieves a CLIP score of 33.20, which is equal to " + }, + { + "bbox": [ + 130, + 350, + 482, + 495 + ], + "type": "inline_equation", + "content": "61.16\\%" + }, + { + "bbox": [ + 130, + 350, + 482, + 495 + ], + "type": "text", + "content": " of GLIGEN's total improvement in CLIP score, while the compromise in YOLO score is equal to only " + }, + { + "bbox": [ + 130, + 350, + 482, + 495 + ], + "type": "inline_equation", + "content": "3.69\\%" + }, + { + "bbox": [ + 130, + 350, + 482, + 495 + ], + "type": "text", + "content": " of the total decrease in the YOLO score of GLIGEN from " + }, + { + "bbox": [ + 130, + 350, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\gamma = 1.0" + }, + { + "bbox": [ + 130, + 350, + 482, + 495 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 350, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\gamma = 0.1" + }, + { + "bbox": [ + 130, + 350, + 482, + 495 + ], + "type": "text", + "content": ". Moreover, a comparison on the Spatial subset of NSR-1K-GPT is provided in the Supplementary (Sec. S2). These results highlight that the advantage of our REGROUND holds robustly for the realistic image captions provided in the MS-COCO [30], as well as for diverse text prompts generated by GPT-4 [34]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 509, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 509, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 131, + 509, + 482, + 568 + ], + "type": "text", + "content": "Random Box Dropping. To further assess the extent of description omission in each method, we modify the MS-COCO-2017 dataset [30] to make the COCO-Drop dataset. In this version, the bounding boxes for " + }, + { + "bbox": [ + 131, + 509, + 482, + 568 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 131, + 509, + 482, + 568 + ], + "type": "text", + "content": " of the categories are randomly removed from each image, thereby preventing every entity described in the text prompt from being included within the bounding boxes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Fig. 8 shows the quantitative comparison of REGROUND and GLIGEN on COCO-Drop. In this case, REGROUND shows a larger advantage over GLIGEN in CLIP score, obtaining a gap in CLIP score which is 1.57 times that of the original MS-COCO-2017 dataset before box dropping for " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\gamma = 1.0" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ". Such a larger gap in CLIP score demonstrates that compared to GLIGEN, our REGROUND better reflects the text prompts even when some entities in the text prompt are not provided as a bounding box. Fig. 7 displays a representative example, where GLIGEN fails to generate a \"cupcake\" when its corresponding bounding" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "text", + "content": "P.Y. Lee and M. Sung" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 114, + 294, + 228 + ], + "blocks": [ + { + "bbox": [ + 141, + 114, + 294, + 228 + ], + "lines": [ + { + "bbox": [ + 141, + 114, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 141, + 114, + 294, + 228 + ], + "type": "image", + "image_path": "5a6aac1e76d2f7c1516044409b7028fd482dfe931d0d124edc1d7dd5239a3458.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 238, + 482, + 262 + ], + "lines": [ + { + "bbox": [ + 131, + 238, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 131, + 238, + 482, + 262 + ], + "type": "text", + "content": "Fig. 8: Comparison on the COCO-Drop Fig.9: Comparison of FID [16] on MS-dataset. COCO-2017 [30] dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 319, + 114, + 471, + 228 + ], + "blocks": [ + { + "bbox": [ + 319, + 114, + 471, + 228 + ], + "lines": [ + { + "bbox": [ + 319, + 114, + 471, + 228 + ], + "spans": [ + { + "bbox": [ + 319, + 114, + 471, + 228 + ], + "type": "image", + "image_path": "0f84585542b5645eb60db586d84e92bd060a1cf141e26fd3b43f91ec8ec89a92.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 285, + 482, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 311 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 311 + ], + "type": "text", + "content": "box is removed in COCO-Drop, whereas our REGROUND robustly generates the cupcake even when it is not provided as a bounding box." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 323, + 482, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 372 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 372 + ], + "type": "text", + "content": "Image Quality. Fig. 9 displays the relationship between YOLO score [52] and FID [16] for each method on MS-COCO-2017. Note that the FID of REGROUND is constantly lower than that of GLIGEN [28], meaning that our network rewiring also results in higher image quality and diversity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 384, + 482, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 494 + ], + "type": "text", + "content": "User Study. We conducted a user study to compare GLIGEN and our ReGround in terms of faithfulness to input text prompts. We used GPT-4 [34] to generate 100 prompts each containing two different objects, along with a bounding box for each object. Participants were given the text prompt along with two images—one from each method—and asked to choose the image that \"better includes all the objects from the prompt.\" Among the 92 out of 100 participants who passed the vigilance tests, our REGROUND surpassed GLIGEN, with a preference rate of " + }, + { + "bbox": [ + 130, + 384, + 482, + 494 + ], + "type": "inline_equation", + "content": "70.05\\%" + }, + { + "bbox": [ + 130, + 384, + 482, + 494 + ], + "type": "text", + "content": " compared to " + }, + { + "bbox": [ + 130, + 384, + 482, + 494 + ], + "type": "inline_equation", + "content": "29.95\\%" + }, + { + "bbox": [ + 130, + 384, + 482, + 494 + ], + "type": "text", + "content": ". Further details on the user study are provided in the Supplementary (Sec. S1)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 504, + 482, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 504, + 482, + 554 + ], + "spans": [ + { + "bbox": [ + 131, + 504, + 482, + 554 + ], + "type": "text", + "content": "PickScore. We further compare the PickScore [26] of GLIGEN [28] and our REGROUND given each input text prompt. On MS-COCO-2017, REGROUND is preferred over GLIGEN by " + }, + { + "bbox": [ + 131, + 504, + 482, + 554 + ], + "type": "inline_equation", + "content": "55.66\\%" + }, + { + "bbox": [ + 131, + 504, + 482, + 554 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 131, + 504, + 482, + 554 + ], + "type": "inline_equation", + "content": "44.34\\%" + }, + { + "bbox": [ + 131, + 504, + 482, + 554 + ], + "type": "text", + "content": ", and on COCO-Drop, REGROUND is preferred by " + }, + { + "bbox": [ + 131, + 504, + 482, + 554 + ], + "type": "inline_equation", + "content": "57.57\\%" + }, + { + "bbox": [ + 131, + 504, + 482, + 554 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 131, + 504, + 482, + 554 + ], + "type": "inline_equation", + "content": "42.43\\%" + }, + { + "bbox": [ + 131, + 504, + 482, + 554 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 572, + 350, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 572, + 350, + 585 + ], + "spans": [ + { + "bbox": [ + 131, + 572, + 350, + 585 + ], + "type": "text", + "content": "6.4 Impact of REGROUND as a Backbone" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "We demonstrate that applying our rewiring of attention modules can also improve text-image alignment in other layout-guided generation methods that use GLIGEN as a backbone. For instance, BoxDiff [54] is a notable example that uses GLIGEN as its foundation and improves spatial grounding with respect to the bounding boxes by leveraging cross-attention maps as additional spatial cues in a zero-shot manner. Our network rewiring can also be combined with" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 180, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 180, + 91, + 448, + 103 + ], + "type": "text", + "content": "ReGround: Improving Textual and Spatial Grounding at No Cost" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 116, + 242, + 195 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 242, + 195 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 242, + 195 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 242, + 195 + ], + "type": "image", + "image_path": "03b98e25b383ddc4c43d3c259ddeac0171fa48cc95eb2baa45c1405bded64e84.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 198, + 237, + 207 + ], + "lines": [ + { + "bbox": [ + 140, + 198, + 237, + 207 + ], + "spans": [ + { + "bbox": [ + 140, + 198, + 237, + 207 + ], + "type": "text", + "content": "(a) NSR-1K-GPT-Counting" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 242, + 116, + 345, + 196 + ], + "blocks": [ + { + "bbox": [ + 242, + 116, + 345, + 196 + ], + "lines": [ + { + "bbox": [ + 242, + 116, + 345, + 196 + ], + "spans": [ + { + "bbox": [ + 242, + 116, + 345, + 196 + ], + "type": "image", + "image_path": "85395662c0e6ddc2922f0b93bfbb39919cb38d383df80e1f937f68d37604ba91.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 216, + 481, + 272 + ], + "lines": [ + { + "bbox": [ + 130, + 216, + 481, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 481, + 272 + ], + "type": "text", + "content": "Fig. 10: Comparison of applying BoxDiff [54] on GLIGEN [28] and our REGROUND, respectively. (a) and (b) show that our REGROUND further improves the grounding quality of BoxDiff on NSR-1K-GPT datasets. (c) While BoxDiff with GLIGEN (left) also shows description omission—omitting \"beautiful sunset\" from the text prompt—BoxDiff with our REGROUND contains the sunset in the final image (right)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 347, + 115, + 479, + 196 + ], + "blocks": [ + { + "bbox": [ + 252, + 198, + 342, + 207 + ], + "lines": [ + { + "bbox": [ + 252, + 198, + 342, + 207 + ], + "spans": [ + { + "bbox": [ + 252, + 198, + 342, + 207 + ], + "type": "text", + "content": "(b) NSR-1K-GPT-Spatial" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 347, + 115, + 479, + 196 + ], + "lines": [ + { + "bbox": [ + 347, + 115, + 479, + 196 + ], + "spans": [ + { + "bbox": [ + 347, + 115, + 479, + 196 + ], + "type": "image", + "image_path": "87b71ec74bd2aa3de3c5b1f08b2ab2b18cc6938d4e601f7375248024dac082c6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 198, + 451, + 207 + ], + "lines": [ + { + "bbox": [ + 380, + 198, + 451, + 207 + ], + "spans": [ + { + "bbox": [ + 380, + 198, + 451, + 207 + ], + "type": "text", + "content": "(c) BoxDiff Outputs" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 293, + 482, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 389 + ], + "type": "text", + "content": "the zero-shot guidance of BoxDiff. Fig. 10 illustrates the results on the NSR-1K-GPT datasets (a) when BoxDiff uses GLIGEN as the base, and (b) when it uses our REGROUND, the rewired GLIGEN, as the base. It depicts that for the same range of spatial grounding accuracies, REGROUND obtains noticeably higher textual grounding (i.e. CLIP score [15]). Also, as shown in Fig. 10-(c), our network rewiring allows for a more detailed description to accurately appear in the final image, both for the entities in the bounding boxes (\"truck\") and the entities that are given as a text prompt (\"sunset\")." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 405, + 219, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 405, + 219, + 417 + ], + "spans": [ + { + "bbox": [ + 132, + 405, + 219, + 417 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 428, + 482, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 428, + 482, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 428, + 482, + 608 + ], + "type": "text", + "content": "We have demonstrated that a simple network rewiring of attention modules, making the gated self-attention and cross-attention parallel, surprisingly improves the trade-off between textual and spatial grounding at no additional cost — without introducing any new parameters, any fine-tuning of the network, or any changes in generation time and memory. Using the pretrained GLIGEN [28], which was trained with the original sequential architecture of the two attention modules, the reconfiguration at inference time has led to achieving higher CLIP scores, indicating the noticeable improvement in textual grounding accuracy. Moreover, our REGROUND improves the textual grounding while preserving the spatial grounding accuracy - achieving " + }, + { + "bbox": [ + 130, + 428, + 482, + 608 + ], + "type": "inline_equation", + "content": "70.25\\%" + }, + { + "bbox": [ + 130, + 428, + 482, + 608 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 428, + 482, + 608 + ], + "type": "inline_equation", + "content": "68.33\\%" + }, + { + "bbox": [ + 130, + 428, + 482, + 608 + ], + "type": "text", + "content": " of GLIGEN's total improvement with the scheduled sampling in CLIP score while compromising YOLO score only " + }, + { + "bbox": [ + 130, + 428, + 482, + 608 + ], + "type": "inline_equation", + "content": "3.31\\%" + }, + { + "bbox": [ + 130, + 428, + 482, + 608 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 428, + 482, + 608 + ], + "type": "inline_equation", + "content": "2.62\\%" + }, + { + "bbox": [ + 130, + 428, + 482, + 608 + ], + "type": "text", + "content": " for the MS-COCO-2014 and MS-COCO-2017 datasets, respectively. We also showcased that this simple yet effective solution for the textual-spatial grounding trade-off can lead to improvements in diverse frameworks using GLIGEN as a base." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Supplementary. Due to limited space, we provide the following contents in the Supplementary: details on the evaluation setup (Sec. S1), additional quantitative (Sec. S2) and qualitative (Sec. S4) comparisons, and more results of REGROUND as a backbone for other layout-guided generation methods (Sec. S3)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "text", + "content": "P.Y. Lee and M. Sung" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 211 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 211 + ], + "type": "text", + "content": "This work was supported by NRF grant (RS-2023-00209723), IITP grants (RS-2019-II190075, RS-2022-II220594, RS-2023-00227592, RS-2024-00399817), and Alchymist Project Program (RS-2024-00423625) funded by the Korean government (MSIT and MOTIE), and grants from the DRB-KAIST SketchTheFuture Research Center, Naver-intel, Adobe Research, Hyundai NGV, KT, and Samsung Electronics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 228, + 197, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 228, + 197, + 241 + ], + "spans": [ + { + "bbox": [ + 133, + 228, + 197, + 241 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 251, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 251, + 480, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 480, + 285 + ], + "type": "text", + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: NeurIPS (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 285, + 480, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 480, + 317 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 480, + 317 + ], + "type": "text", + "content": "2. Avrahami, O., Hayes, T., Gafni, O., Gupta, S., Taigman, Y., Parikh, D., Lischinski, D., Fried, O., Yin, X.: Spatext: Spatio-textual representation for controllable image generation. In: CVPR (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 318, + 480, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 318, + 480, + 350 + ], + "spans": [ + { + "bbox": [ + 138, + 318, + 480, + 350 + ], + "type": "text", + "content": "3. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., et al.: edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 350, + 480, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 350, + 480, + 372 + ], + "spans": [ + { + "bbox": [ + 138, + 350, + 480, + 372 + ], + "type": "text", + "content": "4. Bansal, A., Chu, H.M., Schwarzschild, A., Sengupta, S., Goldblum, M., Geiping, J., Goldstein, T.: Universal guidance for diffusion models. In: ICLR (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 373, + 480, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 373, + 480, + 393 + ], + "spans": [ + { + "bbox": [ + 138, + 373, + 480, + 393 + ], + "type": "text", + "content": "5. Bar-Tal, O., Yariv, L., Lipman, Y., Dekel, T.: Multidiffusion: Fusing diffusion paths for controlled image generation. In: ICML (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 394, + 480, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 394, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 138, + 394, + 480, + 437 + ], + "type": "text", + "content": "6. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., Manassra, W., Dhariwal, P., Chu, C., Jiao, Y., Ramesh, A.: Improving image generation with better captions (2023), https://cdn.openai.com/papers/dall-e-3.pdf" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 437, + 480, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 437, + 480, + 469 + ], + "spans": [ + { + "bbox": [ + 138, + 437, + 480, + 469 + ], + "type": "text", + "content": "7. Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., Zheng, Y.: Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In: ICCV (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 470, + 480, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 470, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 470, + 480, + 491 + ], + "type": "text", + "content": "8. Chen, M., Laina, I., Vedaldi, A.: Training-free layout control with cross-attention guidance. In: WACV (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 492, + 480, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 492, + 480, + 524 + ], + "spans": [ + { + "bbox": [ + 138, + 492, + 480, + 524 + ], + "type": "text", + "content": "9. Chen, W.G., Spiridonova, I., Yang, J., Gao, J., Li, C.: Llava-interactive: An all-in-one demo for image chat, segmentation, generation and editing. arXiv preprint arXiv:2311.00571 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 525, + 480, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 480, + 545 + ], + "type": "text", + "content": "0. Cheng, J., Liang, X., Shi, X., He, T., Xiao, T., Li, M.: Layoutdiffuse: Adapting foundational diffusion models for layout-to-image generation. arXiv (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 546, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 546, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 546, + 480, + 567 + ], + "type": "text", + "content": "1. Couairon, G., Careil, M., Cord, M., Lathuilière, S., Verbeek, J.: Zero-shot spatial layout conditioning for text-to-image diffusion models. In: ICCV (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 568, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 480, + 600 + ], + "type": "text", + "content": "2. Feng, W., Zhu, W., Fu, T.j., Jampani, V., Akula, A., He, X., Basu, S., Wang, X.E., Wang, W.Y.: LayoutGPT: Compositional visual planning and generation with large language models. In: NeurIPS (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 601, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 621 + ], + "type": "text", + "content": "3. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. In: ICLR (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "text", + "content": "4. Herzig, R., Bar, A., Xu, H., Chechik, G., Darrell, T., Globerson, A.: Learning canonical representations for scene graph to image generation. In: ECCV (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "text", + "content": "5. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: CLIPScore: a reference-free evaluation metric for image captioning. In: EMNLP (2021)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 180, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 180, + 91, + 447, + 102 + ], + "type": "text", + "content": "ReGround: Improving Textual and Spatial Grounding at No Cost" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "16. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: GANs trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "text", + "content": "17. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "type": "text", + "content": "18. Hong, S., Yang, D., Choi, J., Lee, H.: Inferring semantic layout for hierarchical text-to-image synthesis. In: CVPR (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "text", + "content": "19. Hong, S., Lee, G., Jang, W., Kim, S.: Improving sample quality of diffusion models using self-attention guidance. In: ICCV (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 258 + ], + "type": "text", + "content": "20. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (2021). https://doi.org/10.5281/zenodo.5143773, https://doi.org/10.5281/zenodo.5143773" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "text", + "content": "21. Johnson, J., Gupta, A., Fei-Fei, L.: Image generation from scene graphs. In: CVPR (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 304 + ], + "type": "text", + "content": "22. Karras, T., Aittala, M., Laine, S., Härkönen, E., Hellsten, J., Lehtinen, J., Aila, T.: Alias-free generative adversarial networks. In: NeurIPS (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "text", + "content": "23. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 325, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 347 + ], + "type": "text", + "content": "24. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of StyleGAN. In: CVPR (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "text", + "content": "25. Kim, Y., Lee, J., Kim, J.H., Ha, J.W., Zhu, J.Y.: Dense text-to-image generation with attention modulation. In: ICCV (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 369, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 392 + ], + "type": "text", + "content": "26. Kirstain, Y., Polyak, A., Singer, U., Matiana, S., Penna, J., Levy, O.: Pick-a-pic: An open dataset of user preferences for text-to-image generation. NeurIPS (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "text", + "content": "27. Li, W., Zhang, P., Zhang, L., Huang, Q., He, X., Lyu, S., Gao, J.: Object-driven text-to-image synthesis via adversarial training. In: CVPR (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "type": "text", + "content": "28. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: GLIGEN: Open-set grounded text-to-image generation. In: CVPR (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "type": "text", + "content": "29. Li, Z., Wu, J., Koh, I., Tang, Y., Sun, L.: Image synthesis from layout with locality-aware mask adaption. In: CVPR (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 456, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 481, + 479 + ], + "type": "text", + "content": "30. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "type": "text", + "content": "31. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "text", + "content": "32. Ma, W.D.K., Lewis, J.P., Lahiri, A., Leung, T., Kleijn, W.B.: Directed diffusion: Direct control of object placement through attention guidance. arXiv preprint arXiv:2302.13153 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 544, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 481, + 578 + ], + "type": "text", + "content": "33. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 578, + 341, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 341, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 341, + 589 + ], + "type": "text", + "content": "34. OpenAI: Chatgpt, https://chat.openai.com/" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "type": "text", + "content": "35. Parmar, G., Kumar Singh, K., Zhang, R., Li, Y., Lu, J., Zhu, J.Y.: Zero-shot image-to-image translation. In: ACM TOG (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "type": "text", + "content": "36. Phung, Q., Ge, S., Huang, J.B.: Grounded text-to-image synthesis with attention refocusing. arXiv preprint arXiv:2306.05427 (2023)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "37. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: SDXL: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "text", + "content": "P.Y. Lee and M. Sung" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 654 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "38. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "39. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "type": "text", + "content": "40. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: ICML (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 205, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 226 + ], + "type": "text", + "content": "41. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 227, + 482, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 227, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 482, + 248 + ], + "type": "text", + "content": "42. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 249, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 482, + 270 + ], + "type": "text", + "content": "43. Si, C., Huang, Z., Jiang, Y., Liu, Z.: Freeu: Free lunch in diffusion u-net. arXiv preprint arXiv:2309.11497 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 271, + 447, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 447, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 447, + 281 + ], + "type": "text", + "content": "44. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 282, + 482, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 482, + 313 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 482, + 313 + ], + "type": "text", + "content": "45. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: ICLR (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 315, + 482, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 335 + ], + "type": "text", + "content": "46. Sun, W., Wu, T.: Image synthesis from reconfigurable layout and style. In: ICCV (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 336, + 482, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 357 + ], + "type": "text", + "content": "47. Sylvain, T., Zhang, P., Bengio, Y., Hjelm, R.D., Sharma, S.: Object-centric image generation from layouts. In: AAAI (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 358, + 482, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 390 + ], + "type": "text", + "content": "48. Tancik, M., Srinivasan, P.P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U., Ramamoorthi, R., Barron, J.T., Ng, R.: Fourier features let networks learn high frequency functions in low dimensional domains. In: NeurIPS (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 391, + 482, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 413 + ], + "type": "text", + "content": "49. Tumanyan, N., Geyer, M., Bagon, S., Dekel, T.: Plug-and-play diffusion features for text-driven image-to-image translation. In: CVPR (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 413, + 482, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 434 + ], + "type": "text", + "content": "50. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 434, + 482, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 456 + ], + "type": "text", + "content": "51. Voynov, A., Aberman, K., Cohen-Or, D.: Sketch-guided text-to-image diffusion models. ACM TOG (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 457, + 482, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 482, + 479 + ], + "type": "text", + "content": "52. Wang, C.Y., Bochkovskiy, A., Liao, H.Y.M.: YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In: CVPR (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 479, + 482, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 500 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 500 + ], + "type": "text", + "content": "53. Xiao, J., Li, L., Lv, H., Wang, S., Huang, Q.: R&b: Region and boundary aware zero-shot grounded text-to-image generation. In: ICLR (2024)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 501, + 482, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 482, + 522 + ], + "type": "text", + "content": "54. Xie, J., Li, Y., Huang, Y., Liu, H., Zhang, W., Zheng, Y., Shou, M.Z.: Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In: ICCV (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 522, + 482, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 544 + ], + "type": "text", + "content": "55. Yang, Z., Wang, J., Gan, Z., Li, L., Lin, K., Wu, C., Duan, N., Liu, Z., Liu, C., Zeng, M., et al.: Reco: Region-controlled text-to-image generation. In: CVPR (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 545, + 482, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 567 + ], + "type": "text", + "content": "56. Yang, Z., Liu, D., Wang, C., Yang, J., Tao, D.: Modeling image composition for complex scene generation. In: CVPR (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 567, + 482, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 482, + 588 + ], + "type": "text", + "content": "57. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 589, + 482, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 589, + 482, + 610 + ], + "spans": [ + { + "bbox": [ + 130, + 589, + 482, + 610 + ], + "type": "text", + "content": "58. Zhao, B., Meng, L., Yin, W., Signal, L.: Image generation from layout. In: CVPR (2019)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 130, + 611, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 482, + 632 + ], + "type": "text", + "content": "59. Zhao, P., Li, H., Jin, R., Zhou, S.K.: Loco: Locally constrained training-free layout-to-image synthesis. arXiv preprint arXiv:2311.12342 (2023)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 130, + 632, + 482, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 654 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 654 + ], + "type": "text", + "content": "60. Zheng, G., Zhou, X., Li, X., Qi, Z., Shan, Y., Li, X.: Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In: CVPR (2023)" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 180, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 180, + 91, + 447, + 102 + ], + "type": "text", + "content": "ReGround: Improving Textual and Spatial Grounding at No Cost" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/ReLoo_ Reconstructing Humans Dressed in Loose Garments from Monocular Video in the Wild/e08d3ae3-38a4-42db-b159-2c5a46fac242_content_list.json b/2024/ReLoo_ Reconstructing Humans Dressed in Loose Garments from Monocular Video in the Wild/e08d3ae3-38a4-42db-b159-2c5a46fac242_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f1560b938685f0933e9884c692cf11a2b068fadb --- /dev/null +++ b/2024/ReLoo_ Reconstructing Humans Dressed in Loose Garments from Monocular Video in the Wild/e08d3ae3-38a4-42db-b159-2c5a46fac242_content_list.json @@ -0,0 +1,1767 @@ +[ + { + "type": "text", + "text": "ReLoo: Reconstructing Humans Dressed in Loose Garments from Monocular Video in the Wild", + "text_level": 1, + "bbox": [ + 218, + 141, + 782, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chen Guo\\*, Tianjian Jiang\\*, Manuel Kaufmann, Chengwei Zheng, Julien Valentin2, Jie Song†, and Otmar Hilliges", + "bbox": [ + 246, + 210, + 754, + 244 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{1}$ ETH Zürich", + "2 Microsoft" + ], + "bbox": [ + 450, + 253, + 550, + 281 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7410daf7c917e3bb4895c227b68fde001bf35658a108c48bb4a453a724c8d7a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 311, + 779, + 453 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. While previous years have seen great progress in the 3D reconstruction of humans from monocular videos, few of the state-of-the-art methods are able to handle loose garments that exhibit large non-rigid surface deformations during articulation. This limits the application of such methods to humans that are dressed in standard pants or T-shirts. Our method, ReLoo, overcomes this limitation and reconstructs high-quality 3D models of humans dressed in loose garments from monocular in-the-wild videos. To tackle this problem, we first establish a layered neural human representation that decomposes clothed humans into a neural inner body and outer clothing. On top of the layered neural representation, we further introduce a non-hierarchical virtual bone deformation module for the clothing layer that can freely move, which allows the accurate recovery of non-rigidly deforming loose clothing. A global optimization jointly optimizes the shape, appearance, and deformations of the human body and clothing via multi-layer differentiable volume rendering. To evaluate ReLoo, we record subjects with dynamically deforming garments in a multi-view capture studio. This evaluation, both on existing and our novel dataset, demonstrates ReLoo's clear superiority over prior art on both indoor datasets and in-the-wild videos.", + "bbox": [ + 259, + 515, + 743, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: human representation $\\cdot$ human reconstruction", + "bbox": [ + 261, + 791, + 656, + 805 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* These authors contributed equally to this work", + "bbox": [ + 220, + 810, + 547, + 824 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author. Now at HKUST(GZ) & HKUST", + "bbox": [ + 220, + 824, + 604, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 143, + 374, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As researchers aim to democratize the creation of realistic human avatars, the reconstruction of 3D clothed humans from casually captured monocular videos has garnered increased attention. While many solutions have been proposed to do so in recent years [1,8,13,18,39,50], they primarily focus on capturing subjects with tight-fitting clothing and perform poorly when reconstructing loose garments whose dynamics are less tightly coupled with body pose. Such loose garments, however, constitute a significant part of a real-life wardrobe and thus failure to capture them is limiting the creation of realistic human avatars from monocular footage. In this paper, we present a method, ReLoo, that overcomes this shortcoming.", + "bbox": [ + 212, + 175, + 787, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Reconstructing humans dressed in loose clothing requires accurate tracking of large, non-rigid deformations and recovering fine-grained details of freely flowing surfaces. Template-based methods [14, 15, 51] have been applied to do so but the acquisition of the template burdens the deployment to unseen subjects and its explicit representation limits its expressive capability to capture dynamically changing surface details. More recently, methods based on neural implicit functions have emerged as a promising remedy for the disadvantages of template-based methods [13, 16, 18, 20, 35, 39, 41, 47, 49, 58-60]. Yet, these methods currently fail to provide convincing reconstructions of humans in loose garments. We observe that few of these methods differentiate between the human body and clothing but model the clothed human as a single entity. This limits the expressiveness and capacity of the underlying model to capture more local features. More importantly, this formulation only allows to drive the off-body garments with skeletal deformations that are derived from the underlying parametric body model (e.g., SMPL [27]). Thus, they are inherently incapable of handling highly dynamic loose garments that are topologically different from the inner body and do not correlate strongly with the bone movement.", + "bbox": [ + 212, + 327, + 787, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we adopt the promising neural implicit shape modeling paradigm, but we argue that a single implicit representation fundamentally limits the representation power and hinders the capability to model complex garment topology that exhibits free-form deformations. To properly model loose garment dynamics that only weakly correlate with skeletal deformations - while still retaining the ability to deform human bodies with skeleton-driven motions - ReLoo takes neural implicit human models to the next level. To do so, our approach is grounded in the following core concepts: i) We establish a layered neural human representation that decomposes clothed humans into the neural inner body and outer clothing. ii) Based on this layered neural representation, we further propose a non-hierarchical virtual bone deformation module for the clothing layer that allows free movement and accurate recovery of highly dynamic loose outfits. iii) In a global optimization, we jointly optimize the shape, appearance, and deformations of both the human body and clothing layer over the entire sequence via multi-layer differentiable volume rendering.", + "bbox": [ + 212, + 584, + 787, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In our experiments, we demonstrate that our framework leads to temporally consistent and high-quality reconstructions of clothed humans dressed in loose", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "C. Guo and T. Jiang et al.", + "bbox": [ + 271, + 114, + 449, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "garments. We also ablate our method to uncover the contribution of its essential components. Furthermore, we conduct comparisons with existing approaches in human surface reconstruction and novel view synthesis, showing that our method outperforms prior art from both the template-based and neural implicit modeling domains. To highlight differences to the prior art, we capture a new dataset, MonoLoose, which puts an emphasis on humans dressed in loose clothing under dynamic motions and contains ground-truth reconstructions captured with a high-end multi-view volumetric recording studio (MVS).", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, in this paper we introduce ReLoo, a method that improves clothed human reconstruction quality and accurately captures human performance dressed in highly dynamic loose garments. Our key contributions are:", + "bbox": [ + 212, + 268, + 787, + 311 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- a novel layered neural human representation, disentangling the inner body and outer clothing; and", + "- a virtual bone deformation module that is built on top of the layered neural human representation and accurately tracks the large surface dynamics; and", + "- a robust framework that leverages multi-layer differentiable volume rendering achieving high-fidelity 3D human reconstructions from monocular in-the-wild videos of humans dressed in highly dynamic loose garments." + ], + "bbox": [ + 225, + 313, + 784, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 443, + 387, + 459 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Single-Layer Human Reconstruction from Monocular Input Template-based monocular human performance capture methods track the pre-defined clothed human template to fit to 2D observations [14,15,51]. They demonstrate robust tracking of human performance even when dressed in loose garments. However, they struggle to generalize to in-the-wild settings due to the reliance on a rigged, personalized, pre-scanned template obtained from a dense capture setup. Follow-up works endeavor to remove this dependency by adding per-vertex offsets on top of the SMPL body [1,12]. Nevertheless, the explicit mesh representation is held back by a fixed resolution and topology, hampering the representation of fine-grained details. Other works have shown compelling results with learning-based methods that learn to regress 3D human geometry and appearance from images [2,16,17,39,49,50,58,60]. A major limitation of these methods is the necessity of high-quality 3D data for supervision. They also often fail to produce space-time coherent reconstructions over frames. Recent works employ neural rendering to fit neural fields to videos to obtain an articulated human model [8,13,18,20,21,26,34,35,37,45,47]. E.g., SelfRecon [18] deploys neural surface rendering [53] to achieve consistent reconstruction over the sequence and Vid2Avatar [13] leverages differentiable volume rendering [52] to eliminate the need for pre-masking thus producing robust 3D human reconstruction. However, all aforementioned methods treat the body and clothing as a single entity, limiting the model's expressiveness and resulting in low-quality reconstructions for loose clothing. In contrast, our method is based on a layered neural human representation that, in conjunction with our novel virtual bone deformation module, enables tracking of highly dynamic loose clothing.", + "bbox": [ + 212, + 477, + 787, + 843 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "ReLoo", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-Layer Human Representation and Reconstruction Several methods exist that investigate how a clothing layer deforms given 3D human motion. They use either physically-simulated training data [3,25,31,32,42,44] or directly deploy physics-informed objectives [10, 40]. These methods have shown compelling results in modeling large deformations of loose outfits. However, they lack generalization to in-the-wild settings and diverse clothing categories due to a reliance on input templates [25, 44] and assume high-quality 3D motion data is already available, not learned from video like in our setting. Our multi-layer representation and deformation modeling are inspired by this line of work, specifically Pan et al. [31], who also make use of virtual bones. Different from [31], our method jointly learns the clothing shape and deformations from only 2D observations, whereas [31] require known clothing templates and 3D simulated data to achieve animation of clothes. Furthermore, [31] define and fix virtual bones using skinning decomposition [24], while in our framework the virtual bone placement evolves naturally with training.", + "bbox": [ + 212, + 146, + 787, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Methods that reconstruct layered clothed humans from videos or images alone are presented in [5, 7, 8, 19, 30, 36, 43, 48, 54]. Some of them either require active multi-view setups [4, 5, 36, 48] or depth information [54] preventing them from being deployed in the wild. BCNet [19] learns to predict clothing geometry draped over the SMPL body from a single image. However, it is limited to pre-defined clothing style templates. SMPLicit [7] and ClothWild [30] extend such learning-based methods to generalize to more general clothing types, but they tend to produce over-smooth results lacking details such as wrinkles. Our closest related work, SCARF [8] is built upon SMPL-X [33] and reconstructs the outer clothing layer using NeRF [29], achieving better reconstruction quality than previous works. Our method differs from SCARF [8] in three regards: 1) ReLoo is a fully implicit representation that is expressive enough to capture detailed body (including faces) and clothing shape jointly, 2) it is not limited to self-rotating motions, and 3) it supports the capture of large non-rigid surface deformations of loose garments thanks to a novel virtual bone deformation module.", + "bbox": [ + 212, + 373, + 787, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In summary, existing monocular-based methods tend to overly rely on parametric body models [27,33] as a human body proxy and thus struggle to model garments whose deformations cannot be easily correlated with the inner body pose under dynamic motion (as is the case for t-shirts or pants that are commonly used as example apparel). Moreover, they can only capture less detailed personalized shape characteristics such as faces. ReLoo overcomes these limitations and when compared to existing single-layer, multi-layer, and template-based methods produces higher fidelity results across the board.", + "bbox": [ + 212, + 599, + 787, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 744, + 330, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce ReLoo, a novel method for detailed geometry and appearance reconstruction of a human performer in highly dynamic, loose clothing from monocular in-the-wild videos. The overview of our method is schematically given in Fig. 1. To enable a temporally consistent, expressive human representation", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "C. Guo and T. Jiang et al.", + "bbox": [ + 271, + 114, + 449, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/20d9f366b461bc6059bc72f9c69b672c6b6109c3631e1124bc429f53c21825dd.jpg", + "image_caption": [ + "Fig. 1: Method Overview. Given an image from a video sequence, we sample points along the camera ray for each neural layer. We warp sampled points for the body layer $\\pmb{x}_d^B$ into canonical space via inverse LBS derived from skeletal deformations. We deform sampled points for the garment layer $\\pmb{x}_d^G$ into canonical space via inverse warping based on the proposed virtual bone deformation module (Sec. 3.2). We then evaluate the respective implicit network to obtain the SDF and radiance values (Sec. 3.1). We apply multi-layer differentiable volume rendering to learn the shape, appearance, and deformations of the layered neural human representation from images (Sec. 3.3). The loss function $\\mathcal{L}$ compares the rendered color predictions with image observations as well as a segmentation mask obtained using SAM [22] (Sec. 3.4)." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 782, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we establish a layered neural implicit representation for the body (inner layer, naked) and garment (outer layer, template-free) as described in Sec. 3.1. On top, we introduce a hybrid deformation strategy which consists of skeletal deformation for the body and a virtual-bone-driven deformation module for the outer loose garment (Sec. 3.2). This allows us to capture dynamic loose garments. We learn the layered human representation and the deformation module jointly by performing multi-layer differentiable volume rendering (Sec. 3.3). The whole process is trained globally to optimize jointly for shape, appearance, and deformations of the inner body and outer garment layer (Sec. 3.4).", + "bbox": [ + 212, + 532, + 787, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Layered Neural Human Representation", + "text_level": 1, + "bbox": [ + 215, + 700, + 584, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We represent the 3D shape of the clothed human with implicit signed-distance fields (SDF) and the appearance with texture fields in a temporally consistent canonical space. The inner body and outer garment are modeled separately. More specifically, we model the geometry and appearance of the body in canonical space with a neural network $f^{B}$ which predicts the signed distance value $s^{B}$ and radiance value $c^{B}$ for any 3D point $\\boldsymbol{x}_{c}^{B}$ in this space. We similarly model the garment with a neural network $f^{G}$ that takes points $\\boldsymbol{x}_{c}^{G}$ in the garment's", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "ReLoo", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "canonical space as input:", + "bbox": [ + 215, + 147, + 397, + 161 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {c} ^ {B}, s ^ {B} = f ^ {B} \\left(\\boldsymbol {x} _ {c} ^ {B}, \\boldsymbol {\\theta}\\right); \\boldsymbol {c} ^ {G}, s ^ {G} = f ^ {G} \\left(\\boldsymbol {x} _ {c} ^ {G}, \\boldsymbol {\\theta}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 172, + 784, + 189 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\theta$ denote the SMPL pose parameters [27], which we concatenate to $\\pmb{x}_c^B$ and $\\pmb{x}_c^G$ to model pose-dependent effects such as facial features and clothing wrinkles. For clothing that is not a single, dress-like garment we use a separate network for the upper and lower garment. For simplicity and without loss of generality, we only discuss a single piece of clothing in the following. Note that if a canonical point is within any of these two surfaces ( $s^B < 0$ or $s^G < 0$ ), it is also within the entire clothed human shape. Thus we can obtain the final clothed human shape by compositing these two neural fields and taking their minimum [38]:", + "bbox": [ + 214, + 198, + 785, + 320 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ns ^ {H} = \\min \\left\\{s ^ {B}, s ^ {G} \\right\\}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 434, + 329, + 784, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Hybrid Deformation Modeling", + "text_level": 1, + "bbox": [ + 215, + 367, + 514, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To find point correspondences between the deformed space in the observed image and our pre-defined canonical space, we devise a hybrid deformation module that treats the body and clothing layers according to their respective levels of rigidity. The human body predominantly depends on skeletal deformation, but the deformation of the loose garments cannot solely be explained by skeletal motion alone. Therefore we propose to drive clothing deformation by a set of additional virtual bones whose transformations are directly learned from video.", + "bbox": [ + 214, + 391, + 785, + 496 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Skeletal Deformation. We follow a standard skeletal deformation based on SMPL to find correspondences in canonical and deformed space for the inner body [13]. Specifically, given the bone transformation matrices $\\mathbf{B}_i$ for joints $i\\in \\{1,\\dots,n_b\\}$ which are derived from the body pose parameters $\\pmb{\\theta}$ , a canonical point $\\pmb{x}_c^B$ is mapped to the corresponding deformed space point $\\pmb{x}_d^B$ via linear blend skinning (LBS):", + "bbox": [ + 214, + 497, + 785, + 587 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {d} ^ {B} = \\sum_ {i = 1} ^ {n _ {s}} w _ {c} ^ {i} \\boldsymbol {B} _ {i} \\boldsymbol {x} _ {c} ^ {B}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 588, + 784, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Conversely, given a point in deformed space $\\pmb{x}_d^B$ , its canonical correspondence can be solved via:", + "bbox": [ + 214, + 632, + 785, + 660 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {c} ^ {B} = \\left(\\sum_ {i = 1} ^ {n _ {s}} w _ {d} ^ {i} \\boldsymbol {B} _ {i}\\right) ^ {- 1} \\boldsymbol {x} _ {d} ^ {B}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 415, + 660, + 784, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $n_s$ denotes the number of skeletal bones in the transformation, and $\\boldsymbol{w}_{(\\cdot)} = \\{w_{(\\cdot)}^1, \\dots, w_{(\\cdot)}^{n_s}\\}$ represents the skinning weights for $\\boldsymbol{x}_{(\\cdot)}^B$ . We assign $\\boldsymbol{w}_d$ to $\\boldsymbol{x}_d^B$ based on the average of neighboring SMPL vertices' skinning weights, weighted by the point-to-point distances in deformed space. The treatment of canonical points $\\boldsymbol{x}_c^B$ follows a similar approach.", + "bbox": [ + 214, + 702, + 785, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Virtual Bone Deformation. The virtual bones correspond to a set of bones $\\mathcal{V} = \\{\\pmb{v}_i\\}_{i=1}^{n_v}$ defined in the canonical space that drives the neighboring 3D garment points $\\pmb{x}_c^G$ using rigid transformations. Different from the skeletons of characters rigged by artists such as SMPL [27], the virtual bones are non-hierarchical", + "bbox": [ + 214, + 779, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "C. Guo and T. Jiang et al.", + "bbox": [ + 271, + 114, + 449, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and are not restricted to rotate in relation to their parents following an anatomical structure. Thus, they can be transformed freely and are applicable to capture the deformations of highly dynamic loose garments.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given a set of virtual bones $\\mathcal{V}$ , their transformations consist of the rotations and translations relative to the SMPL root $\\pmb{\\mathcal{T}}_i = [R_i|T_i]$ which are predicted by a deformation field $\\mathcal{D}^G$ , parameterized via an MLP. $\\mathcal{D}^G$ takes the concatenation of the 3D positions $\\pmb{v}_i$ of the virtual bones, the human body pose $\\pmb{\\theta}$ and a continuous time embedding $\\pmb{t}$ as input, and outputs the axis angles $\\pmb{A}_i$ and translations $\\pmb{T}_i$ . The continuous time embedding $\\pmb{t}$ is included to aid in learning temporal dynamics from videos. The rotations $\\pmb{R}_i$ are further obtained from $\\pmb{A}_i$ via the Rodrigues' formula $f^{Rod}(\\cdot)$ :", + "bbox": [ + 212, + 191, + 787, + 314 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mathcal {T}} _ {i} = \\left[ f ^ {R o d} (\\boldsymbol {A} _ {i}) | \\boldsymbol {T} _ {i} \\right] = \\mathcal {D} ^ {G} (\\boldsymbol {v} _ {i}, \\boldsymbol {\\theta}, t). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 321, + 785, + 340 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To drive a 3D garment point from canonical space $\\pmb{x}_c^G$ to deformed space $\\pmb{x}_d^G$ and vice-versa, we use the neighboring virtual bones' motions and LBS as follows:", + "bbox": [ + 214, + 348, + 782, + 380 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {d} ^ {G} = \\sum_ {i = 1} ^ {n _ {v}} \\delta_ {c} ^ {i} \\boldsymbol {\\mathcal {T}} _ {i} \\boldsymbol {x} _ {c} ^ {G}, \\quad \\boldsymbol {x} _ {c} ^ {G} = \\left(\\sum_ {i = 1} ^ {n _ {v}} \\delta_ {d} ^ {i} \\boldsymbol {\\mathcal {T}} _ {i}\\right) ^ {- 1} \\boldsymbol {x} _ {d} ^ {G} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 390, + 785, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, $n_v$ denotes the number of virtual bones for deforming loose garments, which is a hyperparameter, and $\\pmb{\\delta}_{(\\cdot)} = \\{\\delta_{(\\cdot)}^1,\\dots,\\delta_{(\\cdot)}^{n_v}\\}$ represent the skimming weights of $\\pmb{x}_{(\\cdot)}^G$ w.r.t. each virtual bone $\\pmb{v}_i\\in \\mathcal{V}$ . $\\pmb{\\delta}_{(\\cdot)}$ is calculated based on the inverse of the distance between $\\pmb{x}_c^G$ and each $\\pmb{v}_i$ whereby far away bones are clamped to 0. More details are shown in the Supp. Mat.", + "bbox": [ + 212, + 436, + 782, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Unlike [31], where a garment template is available and the virtual bones are extracted using 3D simulation data, we learn the clothed human model from monocular observations solely and extract the virtual bones on the fly without requiring any specific template prior. We explain the acquisition more in Sec. 3.4.", + "bbox": [ + 212, + 518, + 782, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Multi-Layer Volume Rendering", + "text_level": 1, + "bbox": [ + 214, + 599, + 519, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As we are aiming to jointly reconstruct multiple layers of neural implicit fields, we are required to depart from standard differentiable volume rendering for static scenes (e.g., [29]). We thus introduce multi-layer volume rendering tailored to our multi-layer human representation (Sec. 3.1) to reconstruct both inner body and outer clothing under garment-body occlusions. To do so, we use surface-based volume rendering [52] while re-ordering multiple neural layers [55].", + "bbox": [ + 212, + 625, + 782, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, we shoot a ray $\\pmb{r}$ through every pixel of the image and sample two sets of points in the body layer and the garment layer: $\\{\\pmb{x}_{d,1}^{B},\\dots,\\pmb{x}_{d,N}^{B}\\}$ and $\\{\\pmb{x}_{d,1}^{G},\\dots,\\pmb{x}_{d,N}^{G}\\}$ following the two-stage sampling strategy proposed in [52]. Note that both sets contain the same amount of points. Next, we use the skeletal deformation to warp each sampled point $\\pmb{x}_{d,i}^{B}$ to the body's canonical space $\\pmb{x}_{c,i}^{B}$ and we use our virtual bone deformation module to find canonical correspondences $\\pmb{x}_{c,i}^{G}$ to each sampled point $\\pmb{x}_{d,i}^{G}$ (Sec. 3.2). Then, we obtain the corresponding signed-distance $(s_i^B$ and $s_i^G)$ and radiance values $(\\pmb{c}_i^B$ and $\\pmb{c}_i^G)$ by querying the", + "bbox": [ + 214, + 715, + 787, + 843 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "ReLoo", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "implicit shape and texture fields with the canonical points. We then compute the occupancy for the inner and outer layer at the $i$ -th sampled point as:", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\no _ {i} ^ {B} = \\left(1 - \\exp \\left(- \\sigma_ {i} ^ {B} \\Delta \\boldsymbol {x} _ {i} ^ {B}\\right)\\right); \\sigma_ {i} ^ {B} = \\sigma \\left(s _ {i} ^ {B}, \\boldsymbol {\\theta}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 343, + 189, + 785, + 207 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\no _ {i} ^ {G} = \\left(1 - \\exp \\left(- \\sigma_ {i} ^ {G} \\Delta \\boldsymbol {x} _ {i} ^ {G}\\right)\\right); \\sigma_ {i} ^ {G} = \\sigma \\left(s _ {i} ^ {G}, \\boldsymbol {\\theta}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 210, + 785, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\Delta \\pmb{x}_i^{(\\cdot)}$ is the distance between two adjacent sample points in the respective layer, and $\\sigma (\\cdot)$ represents the scaled Laplace's Cumulative Distribution Function (CDF) to convert $s_i^B$ and $s_i^G$ to volume densities $(\\sigma_{i}^{B}$ and $\\sigma_{i}^{G})$ following [52]. Finally, we integrate the radiance numerically for both layers and obtain the neurally rendered color of the human performer $\\hat{C}^H$ :", + "bbox": [ + 214, + 241, + 782, + 319 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {C} ^ {H} = \\sum_ {i = 1} ^ {N} \\sum_ {p \\in \\{B, G \\}} \\left[ o _ {i} ^ {p} \\boldsymbol {c} _ {i} ^ {p} \\prod_ {q \\in \\{B, G \\}} \\prod_ {j \\in \\mathcal {I} _ {i} ^ {q, p}} \\left(1 - o _ {j} ^ {q}\\right) \\right], \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 332, + 785, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{I}_i^{q,p} = \\{j\\in [1,N]\\mid z(\\pmb{x}_{d,j}^q) < z(\\pmb{x}_{d,i}^p)\\}$ and $z(\\cdot)$ measures the depth of a point w.r.t. the camera origin. In other words, we sort all points according to their depth values from near to far and then conduct the volumetric integration.", + "bbox": [ + 214, + 391, + 787, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Scene Composition. To model the background of the scene we use NeRF++ [56], denoted as $f^S$ , which estimates a color value $\\hat{C}^S$ representing the scene's color. The final pixel color $\\hat{C}$ is a composite of $\\hat{C}^S$ with $\\hat{C}^H$ following [13]. More details are shown in the Supp. Mat.", + "bbox": [ + 214, + 438, + 785, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Global Optimization", + "text_level": 1, + "bbox": [ + 214, + 523, + 431, + 539 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To jointly learn the inner body and outer clothing of clothed humans from monocular videos in a template-free manner, we propose a two-stage training schema. The whole training process is formulated as a global optimization over all estimizable parameters and the entire video sequence.", + "bbox": [ + 212, + 551, + 782, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Two-Stage Training Schema. In the first stage, we leverage skeletal deformation to deform both the body and garment layer. Meanwhile, we warm up the virtual bone deformation field $\\mathcal{D}^G$ by encouraging $\\mathcal{D}^G$ to have similar deformations as the SMPL model around near-body regions. In the second stage, we activate the virtual bone deformation module to drive the garment layer. To obtain virtual bones, we generate garment meshes using Multiresolution IsoSurface Extraction (MISE) [28]. In theory, each of the $M$ vertices of the resulting canonical garment mesh can be made a virtual bone. This will however incur a high computational cost in the deformation module. To mitigate this but still retain expressive capability we employ a quadric mesh simplification algorithm to reduce the number of vertices to $n_v \\ll M$ . The remaining vertices after the simplification are the initial virtual bone locations. We empirically found 80 virtual bones to deliver the best performance-efficiency compromise (see Sec. 4.5). Note that we periodically generate new sets of virtual bones during training to account for changing garment topologies.", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "C. Guo and T. Jiang et al.", + "bbox": [ + 271, + 114, + 450, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Reconstruction Loss. For every ray $\\boldsymbol{r} \\in \\mathcal{R}$ we compute how well the rendered color $\\hat{C}(\\boldsymbol{r})$ matches the image pixel's RGB value $C(\\boldsymbol{r})$ with the $L_{1}$ -distance:", + "bbox": [ + 214, + 146, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r g b}} = \\frac {1}{| \\mathcal {R} |} \\sum_ {\\boldsymbol {r} \\in \\mathcal {R}} | C (\\boldsymbol {r}) - \\hat {C} (\\boldsymbol {r}) |. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 189, + 785, + 223 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Segmentation Loss. We modify Eq. (9) to render the opacity $\\hat{O}^B (\\boldsymbol {r})$ and $\\hat{O}^G (\\boldsymbol {r})$ per pixel for both layers:", + "bbox": [ + 214, + 236, + 782, + 268 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {O} ^ {B} (\\boldsymbol {r}) = \\sum_ {i = 1} ^ {N} \\left[ o _ {i} ^ {B} \\prod_ {q \\in \\{B, G \\}} \\prod_ {j \\in \\mathcal {I} _ {i} ^ {q, B}} \\left(1 - o _ {j} ^ {q}\\right) \\right]; \\hat {O} ^ {G} (\\boldsymbol {r}) = \\sum_ {i = 1} ^ {N} \\left[ o _ {i} ^ {G} \\prod_ {q \\in \\{B, G \\}} \\prod_ {j \\in \\mathcal {I} _ {i} ^ {q, G}} \\left(1 - o _ {j} ^ {q}\\right) \\right]. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 280, + 792, + 338 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The segmentation loss is calculated between the rendered pixel-wise opacity and the segmentation masks extracted using SAM [23]. A robust Geman-McClure error function $\\rho$ [9] is applied to down-weigh potentially erroneous cloth segmentation mask predictions (more details are explained in Supp. Mat):", + "bbox": [ + 214, + 340, + 782, + 398 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s e g}} = \\frac {1}{| \\mathcal {R} |} \\sum_ {\\boldsymbol {r} \\in \\mathcal {R}} \\sum_ {p \\in \\{B, G \\}} \\rho \\left(\\mathcal {M} _ {\\text {s a m}} ^ {p} (\\boldsymbol {r}) - \\hat {O} ^ {p} (\\boldsymbol {r})\\right). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 412, + 782, + 448 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Adaptive Eikonal Loss. We follow IGR [11] and sample points in the canonical space to compute the Eikonal constraint to regularize the validity of our SDFs in each layer. Unlike [11], which randomly samples points in the entire space, we periodically extract the canonical shapes for both layers and sample points around the explicit mesh surfaces:", + "bbox": [ + 214, + 460, + 782, + 535 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {e i k o n a l}} = \\mathbb {E} _ {\\boldsymbol {x} _ {c} ^ {B}} \\left(\\left\\| \\nabla s ^ {B} \\right\\| - 1\\right) ^ {2} + \\mathbb {E} _ {\\boldsymbol {x} _ {c} ^ {G}} \\left(\\left\\| \\nabla s ^ {G} \\right\\| - 1\\right) ^ {2}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 547, + 782, + 566 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Virtual Bone Deformation Regularization. To accelerate the convergence of the virtual bone deformation field in the first stage of the training process, we randomly sample 3D canonical garment points $\\pmb{x}_c^G$ and apply additional SMPL transformation regularization which ensures that the virtual bone deformation $LBS^{G}$ do not deviate excessively from the transformations made by skeletal deformation $LBS^{B}$ during the warm-up stage:", + "bbox": [ + 214, + 578, + 782, + 667 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e g}} = \\left\\| L B S ^ {B} \\left(\\boldsymbol {x} _ {c} ^ {G}\\right) - L B S ^ {G} \\left(\\boldsymbol {x} _ {c} ^ {G}\\right) \\right\\| ^ {2} \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 374, + 680, + 782, + 698 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "See Supp. Mat for more details about the final loss.", + "bbox": [ + 214, + 709, + 584, + 723 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 748, + 374, + 763 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We first introduce the datasets and metrics used for evaluation. Then, we compare our proposed method with state-of-the-art approaches in two tasks: 3D surface reconstruction and novel view synthesis. Ablation studies are then conducted to demonstrate the effectiveness of our core components and design choices.", + "bbox": [ + 214, + 779, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "ReLoo", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/fde61ee00a6471b6575d0a63eef463b3bad94958ed9ac0f78665e389c4a8b259.jpg", + "table_caption": [ + "Table 1: Quantitative evaluation on surface reconstruction. We compute the 3D surface metrics on the MonoLoose dataset. Our method consistently outperforms all baselines on all evaluation metrics (cf. Fig. 2)." + ], + "table_footnote": [], + "table_body": "
MethodC-ℓ2↓NC↑V-IoU↑
SelfRecon [18]2.220.7880.844
Vid2Avatar [13]2.340.7940.776
SCARF [8]3.130.7110.691
Ours w/o Multi-Round Sampl.2.340.7700.879
Ours1.930.8310.881
", + "bbox": [ + 303, + 198, + 697, + 287 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 215, + 315, + 334, + 329 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "MonoLoose Dataset: Due to the lack of datasets that capture dynamic human performance with high-fidelity 3D ground-truth meshes when dressed in loose garments, we captured our own dataset, MonoLoose, with a high-end multiview volumetric capture studio (MVS) [6]. This dataset is specifically curated for evaluating monocular human surface reconstruction and novel view synthesis methods, with a particular focus on subjects dressed in loose attire. It consists of five sequences with different identities, loose garment styles, and motions. For more details on the contents of MonoLoose, please refer to the Supp. Mat.", + "bbox": [ + 212, + 339, + 787, + 460 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "DynaCap [14]: We further evaluate our method on DynaCap, which captures dynamic human performance with a dense multi-view system. We curate two sequences that feature loose garments for novel view synthesis evaluation. Note that DynaCap does not provide dense scans for reconstruction comparison.", + "bbox": [ + 212, + 460, + 787, + 521 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In-the-wild videos: We use in-the-wild videos collected from DeepCap [15] and online videos to demonstrate the robustness and generalization of our method.", + "bbox": [ + 212, + 521, + 785, + 551 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Protocol: We report Chamfer distance $(\\mathbf{C} - \\ell_2)$ [cm], normal consistency (NC), and volumetric IoU for surface reconstruction comparison. Novel view synthesis quality is measured via PSNR, SSIM [46], and LPIPS $(\\times 100)$ [57].", + "bbox": [ + 212, + 551, + 785, + 598 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Surface Reconstruction Comparisons", + "text_level": 1, + "bbox": [ + 215, + 618, + 566, + 633 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We compare our proposed human surface reconstruction method to several state-of-the-art approaches [8, 13, 18] on our MonoLoose dataset. SelfRecon [18] and Vid2Avatar [13] deploy neural rendering to reconstruct the 3D clothed human using a single layer. SCARF [8] reconstructs a hybrid human model based on an explicit inner body and NeRF-based clothing model. All baseline methods rely on SMPL skeleton skinning transformation with additional deformation fields for the garments' motion. Our method outperforms all baselines by a substantial margin on all evaluation metrics (cf. Tab. 1). This disparity becomes more visible in qualitative comparisons shown in Fig. 2. When the dynamic loose garments necessitate large non-skeletal surface deformations, all baseline methods fail to recover complete human surfaces or produce implausible and corrupted reconstructions with visible artifacts (e.g., discontinuities between legs and missing dress parts, see highlights in Fig. 2). Furthermore, they tend to produce less", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "C. Guo and T. Jiang et al.", + "bbox": [ + 271, + 114, + 450, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/806a18f21ad2247b3e03ed01d2cd1908271c82104db77f026bf279271a5a5ccd.jpg", + "image_caption": [ + "Fig. 2: Qualitative 3D surface reconstruction comparison. Baseline methods produce less detailed and implausible 3D clothed human reconstructions with visible artifacts (discontinuities between legs, missing dress parts) due to the strong reliance on skeletal deformations. In contrast, our method correctly recovers the clothing dynamics and generates more detailed and complete 3D human surfaces. Note also that ReLoo produces more detailed facial features." + ], + "image_footnote": [], + "bbox": [ + 230, + 143, + 772, + 431 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "fine-grained details (e.g., the faces shown in the third row and the T-shirts shown in the last row of Fig. 2). In contrast, ReLoo generates complete and plausible 3D human shapes with considerably more details (e.g., clothing wrinkles that fully align with image observations). ReLoo also clearly outperforms the baselines for the surface reconstruction under unseen views (see white background columns in Fig. 2). We attribute this superiority to our proposed neural layered clothed human representation and novel virtual bone deformation module.", + "bbox": [ + 212, + 561, + 787, + 667 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Novel View Synthesis Comparisons", + "text_level": 1, + "bbox": [ + 215, + 691, + 552, + 708 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We compare with the same baselines for the task of novel view synthesis on MonoLoose and DynaCap [14]. We choose an unseen camera from the respective dataset as a novel view for all methods. As shown in Tab. 2, our method outperforms all baseline methods w.r.t. all metrics with an especially large margin on MonoLoose. Fig. 3 shows that SDF-based methods SelfRecon [18] and Vid2Avatar [13] strongly rely on skeletal deformation and cannot accurately recover the correct 3D shapes, leading to similar artifacts in novel view rendering as in surface reconstruction. Although the NeRF-based method SCARF [8] can", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "ReLoo", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f648580d6d622431039a1455a60b5a524abb6efbc9debb02064a20d72dd53d72.jpg", + "image_caption": [ + "Fig. 3: Qualitative novel view synthesis comparison. Our method achieves better rendering quality with detailed texture recovery in e.g., garment patterns and faces. Baseline methods can only produce corrupted and blurry rendering results (dress discontinuities between legs and unsharp texture details)." + ], + "image_footnote": [], + "bbox": [ + 263, + 157, + 743, + 349 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "fit to the coarse shape provided by the contour in image observations, it only produces blurry rendering results. ReLoo, in contrast, produces more plausible and realistic renderings while preserving sharper and fine-grained texture details.", + "bbox": [ + 214, + 450, + 787, + 497 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Qualitative Comparisons with Template-based Method", + "text_level": 1, + "bbox": [ + 214, + 521, + 714, + 536 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Since there is no open-sourced template-based monocular method, we conduct a qualitative comparison with DeepCap [15], which is a learning-based method that predicts the template transformation given image observations. Our method is based on implicit neural fields which are topologically flexible and are not limited to a fixed resolution. As shown in Fig. 4, our method better recovers both the human surface details (e.g., faces and wrinkles) and large non-rigid clothing deformations. More importantly, our representation allows topological changes (empty space between dresses shown in the top row of Fig. 4), while the template-based method is inherently bound to the pre-scanned template mesh.", + "bbox": [ + 212, + 547, + 787, + 684 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 707, + 390, + 722 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Multi-Layer Volume Rendering. To learn the layered human representation, we leverage multi-layer volume rendering that is SDF-based and includes an inverse CDF sampling process. This means we perform multi-round sampling, where we sample two layers individually and combine the samples through sorting. To investigate the effect of this sampling strategy, we compare it to one-round sampling whereby we join the implicit SDFs of the human body and clothing simply by computing the minimum of the joined function. Results in Tab. 1", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "C. Guo and T. Jiang et al.", + "bbox": [ + 271, + 114, + 450, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/9593c6056eb0ddcfa374b2054a2b40153c3e9117b53efc21da1541f390e22d69.jpg", + "table_caption": [ + "Table 2: Quantitative evaluation on novel view synthesis. We report the quantitative results on test views. Our method consistently outperforms other baseline methods on both datasets and all quantitative evaluation metrics, showing more realistic and plausible rendering quality (cf. Fig. 3)." + ], + "table_footnote": [], + "table_body": "
MonoLooseDynaCap [14]
MethodPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
SelfRecon [18]22.50.9536.0826.80.9821.56
Vid2Avatar [13]25.90.9684.6627.10.9831.82
SCARF [8]23.30.9536.5925.50.9792.55
Ours w/o Virtual Bone28.70.9693.8127.30.9821.52
Ours29.20.9703.1527.90.9851.27
", + "bbox": [ + 230, + 212, + 772, + 333 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2abd5b760743873e31c673fcac5a9763e0bcd2ef24f74b0a88a9c257c9daf23b.jpg", + "image_caption": [ + "Image", + "Fig.4: Qualitative comparisons with template-based method." + ], + "image_footnote": [], + "bbox": [ + 217, + 353, + 302, + 479 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f51ac1736881fe74d2120c31dc96656cb85fbb864b97faad0a363f8c2dab800d.jpg", + "image_caption": [ + "DeepCap" + ], + "image_footnote": [], + "bbox": [ + 303, + 353, + 387, + 479 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8fc2e5cb3d04658501f4e017d025edf512105f029dd3d992b963cb327d0eec8b.jpg", + "image_caption": [ + "ReLoo (Ours)" + ], + "image_footnote": [], + "bbox": [ + 388, + 353, + 473, + 479 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Compared to the template-based method, our representation and learning schemes enable more detailed and realistic human surface reconstruction and topological flexibility.", + "bbox": [ + 214, + 534, + 475, + 602 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b0643f7b82465dc7952631e6a090cf99c145092804f4c7e410ee6c77e6894308.jpg", + "image_caption": [ + "Image", + "One-Round Sampl. Multi-Round Sampl.", + "Fig. 5: Importance of multi-round sampling. One-round sampling strategy can lead to physically implausible clothed human reconstructions with severe garment-body interpenetration while multi-round sampling achieves better holistic reconstructions." + ], + "image_footnote": [], + "bbox": [ + 535, + 351, + 772, + 482 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "indicate that multi-round sampling helps to improve the holistic clothed human reconstruction quality and avoids garment-body interpenetration (cf. Fig. 5).", + "bbox": [ + 212, + 638, + 784, + 669 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Virtual Bone Deformation Module. An important hyperparameter in our framework is the number of virtual bones $n_v$ for a garment. We quantitatively analyze the effects by learning the clothed human model with different numbers of virtual bones, i.e. $n_v \\in \\{20, 40, 80, 160, 320\\}$ . We choose a subset of MonoLoose to evaluate novel view synthesis based on the perceptual similarity metric LPIPS [57] and time per training iteration. The quantitative results are reported in Fig. 6. We observe that the time cost linearly increases, while the error decreases with the sharpest drop at $n_v = 80$ bones. We thus select this point as the best performance-efficiency compromise.", + "bbox": [ + 212, + 671, + 787, + 806 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To validate the effectiveness of our virtual bone deformation module, we compare our full model to a version that uses SMPL-based skeletal deformation for", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "ReLoo", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8ef5ee1459db7b6e5b7ca0ed3c3203e69348a8d934ac3d7ed70c51ccd55ea577.jpg", + "image_caption": [ + "Fig.6: Number of virtual bones. LPIPS consistently decreases with the increasing number of virtual bones. We choose 80 virtual bones for the garment layer to balance the method performance and efficiency." + ], + "image_footnote": [], + "bbox": [ + 217, + 145, + 485, + 252 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/52dacf5b7c202a28a8a5648981c6d0d14a11a1a239a0c42e24be072b38c5ad39.jpg", + "image_caption": [ + "Fig.7: Importance of virtual bone. Without the virtual bone deformation, our method is bounded by the expressiveness of skeletal movement and cannot accurately capture the topology and motion of loose garments." + ], + "image_footnote": [], + "bbox": [ + 519, + 152, + 807, + 250 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "the garment layer instead of the virtual bone deformation module. We conducted both quantitative and qualitative analyses as shown in Tab. 2 and Fig. 7. The results demonstrate that integrating the virtual bone deformation module helps to find correct correspondences for points that are not solely controlled by skeletal deformations, leading to plausible and complete novel view synthesis results, while SMPL-based deformation is limited to the hierarchical skeleton structure and struggles with recovering garments that are far away from the inner body.", + "bbox": [ + 212, + 383, + 787, + 491 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 530, + 359, + 546 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We present ReLoo, a novel method that produces temporally consistent 3D reconstructions of humans when dressed in highly dynamic loose garments from monocular in-the-wild videos. Our method does not require any 3D supervision or prior knowledge about the garments. We utilize a carefully designed layered neural implicit human representation to achieve a disentangled reconstruction of the body and the garment. We introduce a non-hierarchical virtual bone deformation module that enables the accurate capture of non-rigidly deforming loose outfits under articulation. A global optimization is formulated to jointly optimize the shape, appearance, and deformations of both inner body and outer clothing from images via multi-layer volume rendering. Our method achieves robust and high-fidelity reconstruction of humans dressed in loose garments.", + "bbox": [ + 212, + 579, + 787, + 744 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations: Although readily available, ReLoo relies on reasonable pose estimates and segmentation masks as inputs. Manual adjustment is occasionally required to obtain SAM masks with sharp boundaries. Our method is mainly deployed to up to two garments. The complexity of ReLoo increases linearly with the number of garments that we aim to reconstruct separately. We discuss more limitations and societal impact in the Supp. Mat.", + "bbox": [ + 212, + 748, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "C. Guo and T. Jiang et al.", + "bbox": [ + 271, + 114, + 450, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was partially supported by the Swiss SERI Consolidation Grant \"AIPERCEIVE\". Chen Guo was partially supported by Microsoft Research Swiss JRC Grant. We are grateful to all our participants for their valued contribution to this research. Computations were carried out in part on the ETH Euler cluster.", + "bbox": [ + 215, + 175, + 785, + 234 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 258, + 321, + 273 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alldieck, T., Magnor, M., Xu, W., Theobalt, C., Pons-Moll, G.: Video based reconstruction of 3d people models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8387-8397 (2018)", + "2. Alldieck, T., Zanfir, M., Sminchisescu, C.: Photorealistic monocular 3d reconstruction of humans wearing clothing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "3. Bertiche, H., Madadi, M., Escalera, S.: Pbns: Physically based neural simulation for unsupervised garment pose space deformation. ACM Trans. Graph. 40(6) (dec 2021)", + "4. Bhatnagar, B.L., Tiwari, G., Theobalt, C., Pons-Moll, G.: Multi-garment net: Learning to dress 3d people from images. In: IEEE International Conference on Computer Vision (ICCV). IEEE (oct 2019)", + "5. Chen, X., Pang, A., Yang, W., Wang, P., Xu, L., Yu, J.: Tightcap: 3d human shape capture with clothing tightness field. ACM Transactions on Graphics (TOG) 41(1), 1-17 (2021)", + "6. Collet, A., Chuang, M., Sweeney, P., Gillett, D., Evseev, D., Calabrese, D., Hoppe, H., Kirk, A., Sullivan, S.: High-quality streamable free-viewpoint video. ACM Trans. Graph. 34(4) (jul 2015). https://doi.org/10.1145/2766945, https://doi.org/10.1145/2766945", + "7. Corona, E., Pumarola, A., Alenyà, G., Pons-Moll, G., Moreno-Noguer, F.: SMPlicit: Topology-aware generative model for clothed people. In: CVPR (2021)", + "8. Feng, Y., Yang, J., Pollefeys, M., Black, M.J., Bolkart, T.: Capturing and animation of body and clothing from monocular video. In: SIGGRAPH Asia 2022 Conference Papers. SA '22 (2022)", + "9. Geman, S., McClure, D.E.: Statistical methods for tomographic image reconstruction (1987)", + "10. Grigorev, A., Thomaszewski, B., Black, M.J., Hilliges, O.: HOOD: Hierarchical graphs for generalized modelling of clothing dynamics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023)", + "11. Gropp, A., Yariv, L., Haim, N., Atzmon, M., Lipman, Y.: Implicit geometric regularization for learning shapes. In: Proceedings of Machine Learning and Systems. pp. 3569-3579 (2020)", + "12. Guo, C., Chen, X., Song, J., Hilliges, O.: Human performance capture from monocular video in the wild. In: 2021 International Conference on 3D Vision (3DV). pp. 889-898. IEEE (2021)", + "13. Guo, C., Jiang, T., Chen, X., Song, J., Hilliges, O.: Vid2 avatar: 3d avatar reconstruction from videos in the wild via self-supervised scene decomposition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023)" + ], + "bbox": [ + 218, + 287, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "ReLoo", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Habermann, M., Liu, L., Xu, W., Zollhoefer, M., Pons-Moll, G., Theobalt, C.: Real-time deep dynamic characters. ACM Transactions on Graphics 40(4) (aug 2021)", + "15. Habermann, M., Xu, W., Zollhoefer, M., Pons-Moll, G., Theobalt, C.: Deepcap: Monocular human performance capture using weak supervision. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (jun 2020)", + "16. He, T., Xu, Y., Saito, S., Soatto, S., Tung, T.: Arch++: Animation-ready clothed human reconstruction revisited. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 11046-11056 (October 2021)", + "17. Huang, Z., Xu, Y., Lassner, C., Li, H., Tung, T.: Arch: Animatable reconstruction of clothed humans. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3093-3102 (2020)", + "18. Jiang, B., Hong, Y., Bao, H., Zhang, J.: Selfrecon: Self reconstruction your digital avatar from monocular video. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "19. Jiang, B., Zhang, J., Hong, Y., Luo, J., Liu, L., Bao, H.: Bcnet: Learning body and cloth shape from a single image. In: European Conference on Computer Vision. Springer (2020)", + "20. Jiang, W., Yi, K.M., Samei, G., Tuzel, O., Ranjan, A.: Neuman: Neural human radiance field from a single video. In: Proceedings of the European conference on computer vision (ECCV) (2022)", + "21. Jiang, Z., Guo, C., Kaufmann, M., Jiang, T., Valentin, J., Hilliges, O., Song, J.: Multiply: Reconstruction of multiple people from monocular video in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2024)", + "22. Ke, L., Ye, M., Danelljan, M., Liu, Y., Tai, Y.W., Tang, C.K., Yu, F.: Segment anything in high quality. In: NeurIPS (2023)", + "23. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4015-4026 (October 2023)", + "24. Le, B.H., Deng, Z.: Smooth skinning decomposition with rigid bones. ACM Trans. Graph. 31(6) (nov 2012). https://doi.org/10.1145/2366145.2366218, https://doi.org/10.1145/2366145.2366218", + "25. Li, Y., Habermann, M., Thomaszewski, B., Coros, S., Beeler, T., Theobalt, C.: Deep physics-aware inference of cloth deformation for monocular human performance capture. In: 2021 International Conference on 3D Vision (3DV). pp. 373-384. IEEE (2021)", + "26. Lin, W., Zheng, C., Yong, J.H., Xu, F.: Relightable and animatable neural avatars from videos. In: Proceedings of the AAAI Conference on Artificial Intelligence (2024)", + "27. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ACM transactions on graphics (TOG) 34(6), 1-16 (2015)", + "28. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4460-4470 (2019)", + "29. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: European conference on computer vision. pp. 405-421. Springer (2020)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "C. Guo and T. Jiang et al.", + "bbox": [ + 271, + 114, + 449, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "30. Moon, G., Nam, H., Shiratori, T., Lee, K.M.: 3d clothed human reconstruction in the wild. In: European Conference on Computer Vision (ECCV) (2022)", + "31. Pan, X., Mai, J., Jiang, X., Tang, D., Li, J., Shao, T., Zhou, K., Jin, X., Manocha, D.: Predicting loose-fitting garment deformations using bone-driven motion networks. In: ACM SIGGRAPH 2022 Conference Proceedings. SIGGRAPH '22, Association for Computing Machinery, New York, NY, USA (2022)", + "32. Patel, C., Liao, Z., Pons-Moll, G.: Tailornet: Predicting clothing in 3d as a function of human pose, shape and garment style. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (jun 2020)", + "33. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3D hands, face, and body from a single image. In: Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR). pp. 10975-10985 (2019)", + "34. Peng, S., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Zhou, X., Bao, H.: Animatable neural radiance fields for modeling dynamic human bodies. In: ICCV (2021)", + "35. Peng, S., Zhang, Y., Xu, Y., Wang, Q., Shuai, Q., Bao, H., Zhou, X.: Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9054-9063 (2021)", + "36. Pons-Moll, G., Pujades, S., Hu, S., Black, M.: Clothcap: Seamless 4d clothing capture and retargeting. ACM Transactions on Graphics, (Proc. SIGGRAPH) 36(4) (2017), two first authors contributed equally", + "37. Qiu, L., Chen, G., Zhou, J., Xu, M., Wang, J., Han, X.: Rec-mv: Reconstructing 3d dynamic cloth from monocular videos. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "38. Ricci, A.: A constructive geometry for computer graphics. Comput. J. 16, 157-160 (1973), https://api_semanticscholar.org/CorpusID:30038820", + "39. Saito, S., Simon, T., Saragih, J., Joo, H.: Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 84-93 (2020)", + "40. Santesteban, I., Otaduy, M.A., Casas, D.: Snug: Self-supervised neural dynamic garments. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8140-8150 (June 2022)", + "41. Su, S.Y., Bagautdinov, T., Rhodin, H.: Danbo: Disentangled articulated neural body representations via graph neural networks. In: European Conference on Computer Vision (2022)", + "42. Su, Z., Wan, W., Yu, T., Liu, L., Fang, L., Wang, W., Liu, Y.: Mulaycap: Multi-layer human performance capture using a monocular video camera. IEEE Transactions on Visualization and Computer Graphics 28(4), 1862-1879 (2022). https://doi.org/10.1109/TVCG.2020.3027763", + "43. Tiwari, G., Bhatnagar, B.L., Tung, T., Pons-Moll, G.: Sizer: A dataset and model for parsing 3d clothing and learning size sensitive 3d clothing. In: European Conference on Computer Vision (ECCV). Springer (August 2020)", + "44. Wang, K., Zhang, G., Cong, S., Yang, J.: Clothed human performance capture with a double-layer neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21098-21107 (June 2023)", + "45. Wang, S., Schwarz, K., Geiger, A., Tang, S.: Arah: Animatable volume rendering of articulated human sdfs. In: European Conference on Computer Vision (ECCV) (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "ReLoo", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "46. Wang, Z., Simoncelli, E.P., Bovik, A.C.: Multiscale structural similarity for image quality assessment. In: The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers. vol. 2 (2003)", + "47. Weng, C.Y., Curless, B., Srinivasan, P.P., Barron, J.T., Kemelmacher-Shlizerman, I.: HumanNeRF: Free-viewpoint rendering of moving people from monocular video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16210-16220 (June 2022)", + "48. Xiang, D., Prada, F., Bagautdinov, T., Xu, W., Dong, Y., Wen, H., Hodgins, J., Wu, C.: Modeling clothing as a separate layer for an animatable human avatar. ACM Trans. Graph. 40(6) (dec 2021)", + "49. Xiu, Y., Yang, J., Cao, X., Tzionas, D., Black, M.J.: ECON: Explicit Clothed humans Optimized via Normal integration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023)", + "50. Xiu, Y., Yang, J., Tzionas, D., Black, M.J.: ICON: Implicit Clothed humans Obtained from Normals. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13296-13306 (June 2022)", + "51. Xu, W., Chatterjee, A., Zollhöfer, M., Rhodin, H., Mehta, D., Seidel, H.P., Theobalt, C.: Monoperfcap: Human performance capture from monocular video. SIGGRAPH 37(2), 27:1-27:15 (May 2018)", + "52. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. In: Advances in Neural Information Processing Systems (2021)", + "53. Yariv, L., Kasten, Y., Moran, D., Galun, M., Atzmon, M., Ronen, B., Lipman, Y.: Multiview neural surface reconstruction by disentangling geometry and appearance. In: Advances in Neural Information Processing Systems (2020)", + "54. Yu, T., Zheng, Z., Zhong, Y., Zhao, J., Dai, Q., Pons-Moll, G., Liu, Y.: Simulcap: Single-view human performance capture with cloth simulation. In: The IEEE International Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (June 2019)", + "55. Zhang, J., Liu, X., Ye, X., Zhao, F., Zhang, Y., Wu, M., Zhang, Y., Xu, L., Yu, J.: Editable free-viewpoint video using a layered neural representation. ACM Transactions on Graphics (TOG) 40(4), 1-18 (2021)", + "56. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv:2010.07492 (2020)", + "57. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)", + "58. Zhang, Z., Sun, L., Yang, Z., Chen, L., Yang, Y.: Global-correlated 3d-decoupling transformer for clothed avatar reconstruction. In: Advances in Neural Information Processing Systems (NeurIPS) (2023)", + "59. Zheng, Z., Huang, H., Yu, T., Zhang, H., Guo, Y., Liu, Y.: Structured local radiance fields for human avatar modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2022)", + "60. Zheng, Z., Yu, T., Liu, Y., Dai, Q.: Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021)" + ], + "bbox": [ + 215, + 146, + 784, + 757 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "C. Guo and T. Jiang et al.", + "bbox": [ + 271, + 114, + 449, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/ReLoo_ Reconstructing Humans Dressed in Loose Garments from Monocular Video in the Wild/e08d3ae3-38a4-42db-b159-2c5a46fac242_model.json b/2024/ReLoo_ Reconstructing Humans Dressed in Loose Garments from Monocular Video in the Wild/e08d3ae3-38a4-42db-b159-2c5a46fac242_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1f2881c893a102601fd2eb95fc973dda11f7d186 --- /dev/null +++ b/2024/ReLoo_ Reconstructing Humans Dressed in Loose Garments from Monocular Video in the Wild/e08d3ae3-38a4-42db-b159-2c5a46fac242_model.json @@ -0,0 +1,2524 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.22, + 0.142, + 0.784, + 0.185 + ], + "angle": 0, + "content": "ReLoo: Reconstructing Humans Dressed in Loose Garments from Monocular Video in the Wild" + }, + { + "type": "text", + "bbox": [ + 0.247, + 0.212, + 0.756, + 0.245 + ], + "angle": 0, + "content": "Chen Guo\\*, Tianjian Jiang\\*, Manuel Kaufmann, Chengwei Zheng, Julien Valentin2, Jie Song†, and Otmar Hilliges" + }, + { + "type": "text", + "bbox": [ + 0.452, + 0.255, + 0.551, + 0.267 + ], + "angle": 0, + "content": "\\(^{1}\\) ETH Zürich" + }, + { + "type": "text", + "bbox": [ + 0.461, + 0.269, + 0.541, + 0.282 + ], + "angle": 0, + "content": "2 Microsoft" + }, + { + "type": "list", + "bbox": [ + 0.452, + 0.255, + 0.551, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.313, + 0.78, + 0.454 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.516, + 0.744, + 0.78 + ], + "angle": 0, + "content": "Abstract. While previous years have seen great progress in the 3D reconstruction of humans from monocular videos, few of the state-of-the-art methods are able to handle loose garments that exhibit large non-rigid surface deformations during articulation. This limits the application of such methods to humans that are dressed in standard pants or T-shirts. Our method, ReLoo, overcomes this limitation and reconstructs high-quality 3D models of humans dressed in loose garments from monocular in-the-wild videos. To tackle this problem, we first establish a layered neural human representation that decomposes clothed humans into a neural inner body and outer clothing. On top of the layered neural representation, we further introduce a non-hierarchical virtual bone deformation module for the clothing layer that can freely move, which allows the accurate recovery of non-rigidly deforming loose clothing. A global optimization jointly optimizes the shape, appearance, and deformations of the human body and clothing via multi-layer differentiable volume rendering. To evaluate ReLoo, we record subjects with dynamically deforming garments in a multi-view capture studio. This evaluation, both on existing and our novel dataset, demonstrates ReLoo's clear superiority over prior art on both indoor datasets and in-the-wild videos." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.792, + 0.658, + 0.806 + ], + "angle": 0, + "content": "Keywords: human representation \\(\\cdot\\) human reconstruction" + }, + { + "type": "page_footnote", + "bbox": [ + 0.221, + 0.811, + 0.548, + 0.825 + ], + "angle": 0, + "content": "* These authors contributed equally to this work" + }, + { + "type": "page_footnote", + "bbox": [ + 0.221, + 0.825, + 0.605, + 0.84 + ], + "angle": 0, + "content": "† Corresponding author. Now at HKUST(GZ) & HKUST" + }, + { + "type": "list", + "bbox": [ + 0.221, + 0.811, + 0.605, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.45, + 0.129 + ], + "angle": 0, + "content": "C. Guo and T. Jiang et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.375, + 0.161 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.326 + ], + "angle": 0, + "content": "As researchers aim to democratize the creation of realistic human avatars, the reconstruction of 3D clothed humans from casually captured monocular videos has garnered increased attention. While many solutions have been proposed to do so in recent years [1,8,13,18,39,50], they primarily focus on capturing subjects with tight-fitting clothing and perform poorly when reconstructing loose garments whose dynamics are less tightly coupled with body pose. Such loose garments, however, constitute a significant part of a real-life wardrobe and thus failure to capture them is limiting the creation of realistic human avatars from monocular footage. In this paper, we present a method, ReLoo, that overcomes this shortcoming." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.583 + ], + "angle": 0, + "content": "Reconstructing humans dressed in loose clothing requires accurate tracking of large, non-rigid deformations and recovering fine-grained details of freely flowing surfaces. Template-based methods [14, 15, 51] have been applied to do so but the acquisition of the template burdens the deployment to unseen subjects and its explicit representation limits its expressive capability to capture dynamically changing surface details. More recently, methods based on neural implicit functions have emerged as a promising remedy for the disadvantages of template-based methods [13, 16, 18, 20, 35, 39, 41, 47, 49, 58-60]. Yet, these methods currently fail to provide convincing reconstructions of humans in loose garments. We observe that few of these methods differentiate between the human body and clothing but model the clothed human as a single entity. This limits the expressiveness and capacity of the underlying model to capture more local features. More importantly, this formulation only allows to drive the off-body garments with skeletal deformations that are derived from the underlying parametric body model (e.g., SMPL [27]). Thus, they are inherently incapable of handling highly dynamic loose garments that are topologically different from the inner body and do not correlate strongly with the bone movement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.788, + 0.81 + ], + "angle": 0, + "content": "In this paper, we adopt the promising neural implicit shape modeling paradigm, but we argue that a single implicit representation fundamentally limits the representation power and hinders the capability to model complex garment topology that exhibits free-form deformations. To properly model loose garment dynamics that only weakly correlate with skeletal deformations - while still retaining the ability to deform human bodies with skeleton-driven motions - ReLoo takes neural implicit human models to the next level. To do so, our approach is grounded in the following core concepts: i) We establish a layered neural human representation that decomposes clothed humans into the neural inner body and outer clothing. ii) Based on this layered neural representation, we further propose a non-hierarchical virtual bone deformation module for the clothing layer that allows free movement and accurate recovery of highly dynamic loose outfits. iii) In a global optimization, we jointly optimize the shape, appearance, and deformations of both the human body and clothing layer over the entire sequence via multi-layer differentiable volume rendering." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In our experiments, we demonstrate that our framework leads to temporally consistent and high-quality reconstructions of clothed humans dressed in loose" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.686, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ReLoo" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "garments. We also ablate our method to uncover the contribution of its essential components. Furthermore, we conduct comparisons with existing approaches in human surface reconstruction and novel view synthesis, showing that our method outperforms prior art from both the template-based and neural implicit modeling domains. To highlight differences to the prior art, we capture a new dataset, MonoLoose, which puts an emphasis on humans dressed in loose clothing under dynamic motions and contains ground-truth reconstructions captured with a high-end multi-view volumetric recording studio (MVS)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.788, + 0.313 + ], + "angle": 0, + "content": "In summary, in this paper we introduce ReLoo, a method that improves clothed human reconstruction quality and accurately captures human performance dressed in highly dynamic loose garments. Our key contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.314, + 0.785, + 0.344 + ], + "angle": 0, + "content": "- a novel layered neural human representation, disentangling the inner body and outer clothing; and" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.344, + 0.785, + 0.373 + ], + "angle": 0, + "content": "- a virtual bone deformation module that is built on top of the layered neural human representation and accurately tracks the large surface dynamics; and" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.375, + 0.785, + 0.419 + ], + "angle": 0, + "content": "- a robust framework that leverages multi-layer differentiable volume rendering achieving high-fidelity 3D human reconstructions from monocular in-the-wild videos of humans dressed in highly dynamic loose garments." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.314, + 0.785, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.444, + 0.388, + 0.46 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Single-Layer Human Reconstruction from Monocular Input Template-based monocular human performance capture methods track the pre-defined clothed human template to fit to 2D observations [14,15,51]. They demonstrate robust tracking of human performance even when dressed in loose garments. However, they struggle to generalize to in-the-wild settings due to the reliance on a rigged, personalized, pre-scanned template obtained from a dense capture setup. Follow-up works endeavor to remove this dependency by adding per-vertex offsets on top of the SMPL body [1,12]. Nevertheless, the explicit mesh representation is held back by a fixed resolution and topology, hampering the representation of fine-grained details. Other works have shown compelling results with learning-based methods that learn to regress 3D human geometry and appearance from images [2,16,17,39,49,50,58,60]. A major limitation of these methods is the necessity of high-quality 3D data for supervision. They also often fail to produce space-time coherent reconstructions over frames. Recent works employ neural rendering to fit neural fields to videos to obtain an articulated human model [8,13,18,20,21,26,34,35,37,45,47]. E.g., SelfRecon [18] deploys neural surface rendering [53] to achieve consistent reconstruction over the sequence and Vid2Avatar [13] leverages differentiable volume rendering [52] to eliminate the need for pre-masking thus producing robust 3D human reconstruction. However, all aforementioned methods treat the body and clothing as a single entity, limiting the model's expressiveness and resulting in low-quality reconstructions for loose clothing. In contrast, our method is based on a layered neural human representation that, in conjunction with our novel virtual bone deformation module, enables tracking of highly dynamic loose clothing." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.45, + 0.129 + ], + "angle": 0, + "content": "C. Guo and T. Jiang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.373 + ], + "angle": 0, + "content": "Multi-Layer Human Representation and Reconstruction Several methods exist that investigate how a clothing layer deforms given 3D human motion. They use either physically-simulated training data [3,25,31,32,42,44] or directly deploy physics-informed objectives [10, 40]. These methods have shown compelling results in modeling large deformations of loose outfits. However, they lack generalization to in-the-wild settings and diverse clothing categories due to a reliance on input templates [25, 44] and assume high-quality 3D motion data is already available, not learned from video like in our setting. Our multi-layer representation and deformation modeling are inspired by this line of work, specifically Pan et al. [31], who also make use of virtual bones. Different from [31], our method jointly learns the clothing shape and deformations from only 2D observations, whereas [31] require known clothing templates and 3D simulated data to achieve animation of clothes. Furthermore, [31] define and fix virtual bones using skinning decomposition [24], while in our framework the virtual bone placement evolves naturally with training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.6 + ], + "angle": 0, + "content": "Methods that reconstruct layered clothed humans from videos or images alone are presented in [5, 7, 8, 19, 30, 36, 43, 48, 54]. Some of them either require active multi-view setups [4, 5, 36, 48] or depth information [54] preventing them from being deployed in the wild. BCNet [19] learns to predict clothing geometry draped over the SMPL body from a single image. However, it is limited to pre-defined clothing style templates. SMPLicit [7] and ClothWild [30] extend such learning-based methods to generalize to more general clothing types, but they tend to produce over-smooth results lacking details such as wrinkles. Our closest related work, SCARF [8] is built upon SMPL-X [33] and reconstructs the outer clothing layer using NeRF [29], achieving better reconstruction quality than previous works. Our method differs from SCARF [8] in three regards: 1) ReLoo is a fully implicit representation that is expressive enough to capture detailed body (including faces) and clothing shape jointly, 2) it is not limited to self-rotating motions, and 3) it supports the capture of large non-rigid surface deformations of loose garments thanks to a novel virtual bone deformation module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.601, + 0.788, + 0.722 + ], + "angle": 0, + "content": "In summary, existing monocular-based methods tend to overly rely on parametric body models [27,33] as a human body proxy and thus struggle to model garments whose deformations cannot be easily correlated with the inner body pose under dynamic motion (as is the case for t-shirts or pants that are commonly used as example apparel). Moreover, they can only capture less detailed personalized shape characteristics such as faces. ReLoo overcomes these limitations and when compared to existing single-layer, multi-layer, and template-based methods produces higher fidelity results across the board." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.746, + 0.331, + 0.763 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "We introduce ReLoo, a novel method for detailed geometry and appearance reconstruction of a human performer in highly dynamic, loose clothing from monocular in-the-wild videos. The overview of our method is schematically given in Fig. 1. To enable a temporally consistent, expressive human representation" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.686, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ReLoo" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.144, + 0.784, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.362, + 0.788, + 0.502 + ], + "angle": 0, + "content": "Fig. 1: Method Overview. Given an image from a video sequence, we sample points along the camera ray for each neural layer. We warp sampled points for the body layer \\(\\pmb{x}_d^B\\) into canonical space via inverse LBS derived from skeletal deformations. We deform sampled points for the garment layer \\(\\pmb{x}_d^G\\) into canonical space via inverse warping based on the proposed virtual bone deformation module (Sec. 3.2). We then evaluate the respective implicit network to obtain the SDF and radiance values (Sec. 3.1). We apply multi-layer differentiable volume rendering to learn the shape, appearance, and deformations of the layered neural human representation from images (Sec. 3.3). The loss function \\(\\mathcal{L}\\) compares the rendered color predictions with image observations as well as a segmentation mask obtained using SAM [22] (Sec. 3.4)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.788, + 0.671 + ], + "angle": 0, + "content": "we establish a layered neural implicit representation for the body (inner layer, naked) and garment (outer layer, template-free) as described in Sec. 3.1. On top, we introduce a hybrid deformation strategy which consists of skeletal deformation for the body and a virtual-bone-driven deformation module for the outer loose garment (Sec. 3.2). This allows us to capture dynamic loose garments. We learn the layered human representation and the deformation module jointly by performing multi-layer differentiable volume rendering (Sec. 3.3). The whole process is trained globally to optimize jointly for shape, appearance, and deformations of the inner body and outer garment layer (Sec. 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.701, + 0.586, + 0.716 + ], + "angle": 0, + "content": "3.1 Layered Neural Human Representation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.842 + ], + "angle": 0, + "content": "We represent the 3D shape of the clothed human with implicit signed-distance fields (SDF) and the appearance with texture fields in a temporally consistent canonical space. The inner body and outer garment are modeled separately. More specifically, we model the geometry and appearance of the body in canonical space with a neural network \\( f^{B} \\) which predicts the signed distance value \\( s^{B} \\) and radiance value \\( c^{B} \\) for any 3D point \\( \\boldsymbol{x}_{c}^{B} \\) in this space. We similarly model the garment with a neural network \\( f^{G} \\) that takes points \\( \\boldsymbol{x}_{c}^{G} \\) in the garment's" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.45, + 0.129 + ], + "angle": 0, + "content": "C. Guo and T. Jiang et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.148, + 0.398, + 0.162 + ], + "angle": 0, + "content": "canonical space as input:" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.173, + 0.785, + 0.19 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {c} ^ {B}, s ^ {B} = f ^ {B} \\left(\\boldsymbol {x} _ {c} ^ {B}, \\boldsymbol {\\theta}\\right); \\boldsymbol {c} ^ {G}, s ^ {G} = f ^ {G} \\left(\\boldsymbol {x} _ {c} ^ {G}, \\boldsymbol {\\theta}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.199, + 0.787, + 0.321 + ], + "angle": 0, + "content": "where \\(\\theta\\) denote the SMPL pose parameters [27], which we concatenate to \\(\\pmb{x}_c^B\\) and \\(\\pmb{x}_c^G\\) to model pose-dependent effects such as facial features and clothing wrinkles. For clothing that is not a single, dress-like garment we use a separate network for the upper and lower garment. For simplicity and without loss of generality, we only discuss a single piece of clothing in the following. Note that if a canonical point is within any of these two surfaces (\\(s^B < 0\\) or \\(s^G < 0\\)), it is also within the entire clothed human shape. Thus we can obtain the final clothed human shape by compositing these two neural fields and taking their minimum [38]:" + }, + { + "type": "equation", + "bbox": [ + 0.435, + 0.33, + 0.785, + 0.348 + ], + "angle": 0, + "content": "\\[\ns ^ {H} = \\min \\left\\{s ^ {B}, s ^ {G} \\right\\}. \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.368, + 0.515, + 0.384 + ], + "angle": 0, + "content": "3.2 Hybrid Deformation Modeling" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.392, + 0.787, + 0.497 + ], + "angle": 0, + "content": "To find point correspondences between the deformed space in the observed image and our pre-defined canonical space, we devise a hybrid deformation module that treats the body and clothing layers according to their respective levels of rigidity. The human body predominantly depends on skeletal deformation, but the deformation of the loose garments cannot solely be explained by skeletal motion alone. Therefore we propose to drive clothing deformation by a set of additional virtual bones whose transformations are directly learned from video." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.498, + 0.787, + 0.588 + ], + "angle": 0, + "content": "Skeletal Deformation. We follow a standard skeletal deformation based on SMPL to find correspondences in canonical and deformed space for the inner body [13]. Specifically, given the bone transformation matrices \\(\\mathbf{B}_i\\) for joints \\(i\\in \\{1,\\dots,n_b\\}\\) which are derived from the body pose parameters \\(\\pmb{\\theta}\\), a canonical point \\(\\pmb{x}_c^B\\) is mapped to the corresponding deformed space point \\(\\pmb{x}_d^B\\) via linear blend skinning (LBS):" + }, + { + "type": "equation", + "bbox": [ + 0.433, + 0.589, + 0.785, + 0.627 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {d} ^ {B} = \\sum_ {i = 1} ^ {n _ {s}} w _ {c} ^ {i} \\boldsymbol {B} _ {i} \\boldsymbol {x} _ {c} ^ {B}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.633, + 0.786, + 0.661 + ], + "angle": 0, + "content": "Conversely, given a point in deformed space \\( \\pmb{x}_d^B \\), its canonical correspondence can be solved via:" + }, + { + "type": "equation", + "bbox": [ + 0.416, + 0.661, + 0.785, + 0.699 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {c} ^ {B} = \\left(\\sum_ {i = 1} ^ {n _ {s}} w _ {d} ^ {i} \\boldsymbol {B} _ {i}\\right) ^ {- 1} \\boldsymbol {x} _ {d} ^ {B}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.703, + 0.787, + 0.78 + ], + "angle": 0, + "content": "Here, \\( n_s \\) denotes the number of skeletal bones in the transformation, and \\( \\boldsymbol{w}_{(\\cdot)} = \\{w_{(\\cdot)}^1, \\dots, w_{(\\cdot)}^{n_s}\\} \\) represents the skinning weights for \\( \\boldsymbol{x}_{(\\cdot)}^B \\). We assign \\( \\boldsymbol{w}_d \\) to \\( \\boldsymbol{x}_d^B \\) based on the average of neighboring SMPL vertices' skinning weights, weighted by the point-to-point distances in deformed space. The treatment of canonical points \\( \\boldsymbol{x}_c^B \\) follows a similar approach." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Virtual Bone Deformation. The virtual bones correspond to a set of bones \\(\\mathcal{V} = \\{\\pmb{v}_i\\}_{i=1}^{n_v}\\) defined in the canonical space that drives the neighboring 3D garment points \\(\\pmb{x}_c^G\\) using rigid transformations. Different from the skeletons of characters rigged by artists such as SMPL [27], the virtual bones are non-hierarchical" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.686, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ReLoo" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "and are not restricted to rotate in relation to their parents following an anatomical structure. Thus, they can be transformed freely and are applicable to capture the deformations of highly dynamic loose garments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.788, + 0.315 + ], + "angle": 0, + "content": "Given a set of virtual bones \\(\\mathcal{V}\\), their transformations consist of the rotations and translations relative to the SMPL root \\(\\pmb{\\mathcal{T}}_i = [R_i|T_i]\\) which are predicted by a deformation field \\(\\mathcal{D}^G\\), parameterized via an MLP. \\(\\mathcal{D}^G\\) takes the concatenation of the 3D positions \\(\\pmb{v}_i\\) of the virtual bones, the human body pose \\(\\pmb{\\theta}\\) and a continuous time embedding \\(\\pmb{t}\\) as input, and outputs the axis angles \\(\\pmb{A}_i\\) and translations \\(\\pmb{T}_i\\). The continuous time embedding \\(\\pmb{t}\\) is included to aid in learning temporal dynamics from videos. The rotations \\(\\pmb{R}_i\\) are further obtained from \\(\\pmb{A}_i\\) via the Rodrigues' formula \\(f^{Rod}(\\cdot)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.323, + 0.786, + 0.341 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mathcal {T}} _ {i} = \\left[ f ^ {R o d} (\\boldsymbol {A} _ {i}) | \\boldsymbol {T} _ {i} \\right] = \\mathcal {D} ^ {G} (\\boldsymbol {v} _ {i}, \\boldsymbol {\\theta}, t). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.349, + 0.784, + 0.381 + ], + "angle": 0, + "content": "To drive a 3D garment point from canonical space \\(\\pmb{x}_c^G\\) to deformed space \\(\\pmb{x}_d^G\\) and vice-versa, we use the neighboring virtual bones' motions and LBS as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.338, + 0.391, + 0.786, + 0.429 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {d} ^ {G} = \\sum_ {i = 1} ^ {n _ {v}} \\delta_ {c} ^ {i} \\boldsymbol {\\mathcal {T}} _ {i} \\boldsymbol {x} _ {c} ^ {G}, \\quad \\boldsymbol {x} _ {c} ^ {G} = \\left(\\sum_ {i = 1} ^ {n _ {v}} \\delta_ {d} ^ {i} \\boldsymbol {\\mathcal {T}} _ {i}\\right) ^ {- 1} \\boldsymbol {x} _ {d} ^ {G} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.784, + 0.519 + ], + "angle": 0, + "content": "Here, \\( n_v \\) denotes the number of virtual bones for deforming loose garments, which is a hyperparameter, and \\( \\pmb{\\delta}_{(\\cdot)} = \\{\\delta_{(\\cdot)}^1,\\dots,\\delta_{(\\cdot)}^{n_v}\\} \\) represent the skimming weights of \\( \\pmb{x}_{(\\cdot)}^G \\) w.r.t. each virtual bone \\( \\pmb{v}_i\\in \\mathcal{V} \\). \\( \\pmb{\\delta}_{(\\cdot)} \\) is calculated based on the inverse of the distance between \\( \\pmb{x}_c^G \\) and each \\( \\pmb{v}_i \\) whereby far away bones are clamped to 0. More details are shown in the Supp. Mat." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.784, + 0.581 + ], + "angle": 0, + "content": "Unlike [31], where a garment template is available and the virtual bones are extracted using 3D simulation data, we learn the clothed human model from monocular observations solely and extract the virtual bones on the fly without requiring any specific template prior. We explain the acquisition more in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.601, + 0.52, + 0.617 + ], + "angle": 0, + "content": "3.3 Multi-Layer Volume Rendering" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.625, + 0.784, + 0.716 + ], + "angle": 0, + "content": "As we are aiming to jointly reconstruct multiple layers of neural implicit fields, we are required to depart from standard differentiable volume rendering for static scenes (e.g., [29]). We thus introduce multi-layer volume rendering tailored to our multi-layer human representation (Sec. 3.1) to reconstruct both inner body and outer clothing under garment-body occlusions. To do so, we use surface-based volume rendering [52] while re-ordering multiple neural layers [55]." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.716, + 0.788, + 0.844 + ], + "angle": 0, + "content": "Specifically, we shoot a ray \\(\\pmb{r}\\) through every pixel of the image and sample two sets of points in the body layer and the garment layer: \\(\\{\\pmb{x}_{d,1}^{B},\\dots,\\pmb{x}_{d,N}^{B}\\}\\) and \\(\\{\\pmb{x}_{d,1}^{G},\\dots,\\pmb{x}_{d,N}^{G}\\}\\) following the two-stage sampling strategy proposed in [52]. Note that both sets contain the same amount of points. Next, we use the skeletal deformation to warp each sampled point \\(\\pmb{x}_{d,i}^{B}\\) to the body's canonical space \\(\\pmb{x}_{c,i}^{B}\\) and we use our virtual bone deformation module to find canonical correspondences \\(\\pmb{x}_{c,i}^{G}\\) to each sampled point \\(\\pmb{x}_{d,i}^{G}\\) (Sec. 3.2). Then, we obtain the corresponding signed-distance \\((s_i^B\\) and \\(s_i^G)\\) and radiance values \\((\\pmb{c}_i^B\\) and \\(\\pmb{c}_i^G)\\) by querying the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.452, + 0.129 + ], + "angle": 0, + "content": "C. Guo and T. Jiang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "implicit shape and texture fields with the canonical points. We then compute the occupancy for the inner and outer layer at the \\(i\\)-th sampled point as:" + }, + { + "type": "equation", + "bbox": [ + 0.344, + 0.19, + 0.786, + 0.208 + ], + "angle": 0, + "content": "\\[\no _ {i} ^ {B} = \\left(1 - \\exp \\left(- \\sigma_ {i} ^ {B} \\Delta \\boldsymbol {x} _ {i} ^ {B}\\right)\\right); \\sigma_ {i} ^ {B} = \\sigma \\left(s _ {i} ^ {B}, \\boldsymbol {\\theta}\\right) \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.211, + 0.786, + 0.23 + ], + "angle": 0, + "content": "\\[\no _ {i} ^ {G} = \\left(1 - \\exp \\left(- \\sigma_ {i} ^ {G} \\Delta \\boldsymbol {x} _ {i} ^ {G}\\right)\\right); \\sigma_ {i} ^ {G} = \\sigma \\left(s _ {i} ^ {G}, \\boldsymbol {\\theta}\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.242, + 0.784, + 0.32 + ], + "angle": 0, + "content": "where \\(\\Delta \\pmb{x}_i^{(\\cdot)}\\) is the distance between two adjacent sample points in the respective layer, and \\(\\sigma (\\cdot)\\) represents the scaled Laplace's Cumulative Distribution Function (CDF) to convert \\(s_i^B\\) and \\(s_i^G\\) to volume densities \\((\\sigma_{i}^{B}\\) and \\(\\sigma_{i}^{G})\\) following [52]. Finally, we integrate the radiance numerically for both layers and obtain the neurally rendered color of the human performer \\(\\hat{C}^H\\):" + }, + { + "type": "equation", + "bbox": [ + 0.328, + 0.333, + 0.786, + 0.381 + ], + "angle": 0, + "content": "\\[\n\\hat {C} ^ {H} = \\sum_ {i = 1} ^ {N} \\sum_ {p \\in \\{B, G \\}} \\left[ o _ {i} ^ {p} \\boldsymbol {c} _ {i} ^ {p} \\prod_ {q \\in \\{B, G \\}} \\prod_ {j \\in \\mathcal {I} _ {i} ^ {q, p}} \\left(1 - o _ {j} ^ {q}\\right) \\right], \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.392, + 0.788, + 0.438 + ], + "angle": 0, + "content": "where \\(\\mathcal{I}_i^{q,p} = \\{j\\in [1,N]\\mid z(\\pmb{x}_{d,j}^q) < z(\\pmb{x}_{d,i}^p)\\}\\) and \\(z(\\cdot)\\) measures the depth of a point w.r.t. the camera origin. In other words, we sort all points according to their depth values from near to far and then conduct the volumetric integration." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.439, + 0.787, + 0.5 + ], + "angle": 0, + "content": "Scene Composition. To model the background of the scene we use NeRF++ [56], denoted as \\( f^S \\), which estimates a color value \\( \\hat{C}^S \\) representing the scene's color. The final pixel color \\( \\hat{C} \\) is a composite of \\( \\hat{C}^S \\) with \\( \\hat{C}^H \\) following [13]. More details are shown in the Supp. Mat." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.525, + 0.432, + 0.54 + ], + "angle": 0, + "content": "3.4 Global Optimization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.784, + 0.613 + ], + "angle": 0, + "content": "To jointly learn the inner body and outer clothing of clothed humans from monocular videos in a template-free manner, we propose a two-stage training schema. The whole training process is formulated as a global optimization over all estimizable parameters and the entire video sequence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Two-Stage Training Schema. In the first stage, we leverage skeletal deformation to deform both the body and garment layer. Meanwhile, we warm up the virtual bone deformation field \\(\\mathcal{D}^G\\) by encouraging \\(\\mathcal{D}^G\\) to have similar deformations as the SMPL model around near-body regions. In the second stage, we activate the virtual bone deformation module to drive the garment layer. To obtain virtual bones, we generate garment meshes using Multiresolution IsoSurface Extraction (MISE) [28]. In theory, each of the \\(M\\) vertices of the resulting canonical garment mesh can be made a virtual bone. This will however incur a high computational cost in the deformation module. To mitigate this but still retain expressive capability we employ a quadric mesh simplification algorithm to reduce the number of vertices to \\(n_v \\ll M\\). The remaining vertices after the simplification are the initial virtual bone locations. We empirically found 80 virtual bones to deliver the best performance-efficiency compromise (see Sec. 4.5). Note that we periodically generate new sets of virtual bones during training to account for changing garment topologies." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.686, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ReLoo" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "Reconstruction Loss. For every ray \\( \\boldsymbol{r} \\in \\mathcal{R} \\) we compute how well the rendered color \\( \\hat{C}(\\boldsymbol{r}) \\) matches the image pixel's RGB value \\( C(\\boldsymbol{r}) \\) with the \\( L_{1} \\)-distance:" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.19, + 0.786, + 0.224 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {r g b}} = \\frac {1}{| \\mathcal {R} |} \\sum_ {\\boldsymbol {r} \\in \\mathcal {R}} | C (\\boldsymbol {r}) - \\hat {C} (\\boldsymbol {r}) |. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.237, + 0.784, + 0.269 + ], + "angle": 0, + "content": "Segmentation Loss. We modify Eq. (9) to render the opacity \\(\\hat{O}^B (\\boldsymbol {r})\\) and \\(\\hat{O}^G (\\boldsymbol {r})\\) per pixel for both layers:" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.281, + 0.793, + 0.339 + ], + "angle": 0, + "content": "\\[\n\\hat {O} ^ {B} (\\boldsymbol {r}) = \\sum_ {i = 1} ^ {N} \\left[ o _ {i} ^ {B} \\prod_ {q \\in \\{B, G \\}} \\prod_ {j \\in \\mathcal {I} _ {i} ^ {q, B}} \\left(1 - o _ {j} ^ {q}\\right) \\right]; \\hat {O} ^ {G} (\\boldsymbol {r}) = \\sum_ {i = 1} ^ {N} \\left[ o _ {i} ^ {G} \\prod_ {q \\in \\{B, G \\}} \\prod_ {j \\in \\mathcal {I} _ {i} ^ {q, G}} \\left(1 - o _ {j} ^ {q}\\right) \\right]. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.341, + 0.784, + 0.4 + ], + "angle": 0, + "content": "The segmentation loss is calculated between the rendered pixel-wise opacity and the segmentation masks extracted using SAM [23]. A robust Geman-McClure error function \\(\\rho\\) [9] is applied to down-weigh potentially erroneous cloth segmentation mask predictions (more details are explained in Supp. Mat):" + }, + { + "type": "equation", + "bbox": [ + 0.343, + 0.413, + 0.784, + 0.449 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s e g}} = \\frac {1}{| \\mathcal {R} |} \\sum_ {\\boldsymbol {r} \\in \\mathcal {R}} \\sum_ {p \\in \\{B, G \\}} \\rho \\left(\\mathcal {M} _ {\\text {s a m}} ^ {p} (\\boldsymbol {r}) - \\hat {O} ^ {p} (\\boldsymbol {r})\\right). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.461, + 0.784, + 0.536 + ], + "angle": 0, + "content": "Adaptive Eikonal Loss. We follow IGR [11] and sample points in the canonical space to compute the Eikonal constraint to regularize the validity of our SDFs in each layer. Unlike [11], which randomly samples points in the entire space, we periodically extract the canonical shapes for both layers and sample points around the explicit mesh surfaces:" + }, + { + "type": "equation", + "bbox": [ + 0.32, + 0.549, + 0.784, + 0.568 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {e i k o n a l}} = \\mathbb {E} _ {\\boldsymbol {x} _ {c} ^ {B}} \\left(\\left\\| \\nabla s ^ {B} \\right\\| - 1\\right) ^ {2} + \\mathbb {E} _ {\\boldsymbol {x} _ {c} ^ {G}} \\left(\\left\\| \\nabla s ^ {G} \\right\\| - 1\\right) ^ {2}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.579, + 0.784, + 0.669 + ], + "angle": 0, + "content": "Virtual Bone Deformation Regularization. To accelerate the convergence of the virtual bone deformation field in the first stage of the training process, we randomly sample 3D canonical garment points \\(\\pmb{x}_c^G\\) and apply additional SMPL transformation regularization which ensures that the virtual bone deformation \\(LBS^{G}\\) do not deviate excessively from the transformations made by skeletal deformation \\(LBS^{B}\\) during the warm-up stage:" + }, + { + "type": "equation", + "bbox": [ + 0.375, + 0.681, + 0.784, + 0.699 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e g}} = \\left\\| L B S ^ {B} \\left(\\boldsymbol {x} _ {c} ^ {G}\\right) - L B S ^ {G} \\left(\\boldsymbol {x} _ {c} ^ {G}\\right) \\right\\| ^ {2} \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.71, + 0.586, + 0.724 + ], + "angle": 0, + "content": "See Supp. Mat for more details about the final loss." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.749, + 0.375, + 0.765 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.784, + 0.84 + ], + "angle": 0, + "content": "We first introduce the datasets and metrics used for evaluation. Then, we compare our proposed method with state-of-the-art approaches in two tasks: 3D surface reconstruction and novel view synthesis. Ablation studies are then conducted to demonstrate the effectiveness of our core components and design choices." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.451, + 0.129 + ], + "angle": 0, + "content": "C. Guo and T. Jiang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 1: Quantitative evaluation on surface reconstruction. We compute the 3D surface metrics on the MonoLoose dataset. Our method consistently outperforms all baselines on all evaluation metrics (cf. Fig. 2)." + }, + { + "type": "table", + "bbox": [ + 0.304, + 0.199, + 0.699, + 0.289 + ], + "angle": 0, + "content": "
MethodC-ℓ2↓NC↑V-IoU↑
SelfRecon [18]2.220.7880.844
Vid2Avatar [13]2.340.7940.776
SCARF [8]3.130.7110.691
Ours w/o Multi-Round Sampl.2.340.7700.879
Ours1.930.8310.881
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.316, + 0.335, + 0.33 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.788, + 0.461 + ], + "angle": 0, + "content": "MonoLoose Dataset: Due to the lack of datasets that capture dynamic human performance with high-fidelity 3D ground-truth meshes when dressed in loose garments, we captured our own dataset, MonoLoose, with a high-end multiview volumetric capture studio (MVS) [6]. This dataset is specifically curated for evaluating monocular human surface reconstruction and novel view synthesis methods, with a particular focus on subjects dressed in loose attire. It consists of five sequences with different identities, loose garment styles, and motions. For more details on the contents of MonoLoose, please refer to the Supp. Mat." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.522 + ], + "angle": 0, + "content": "DynaCap [14]: We further evaluate our method on DynaCap, which captures dynamic human performance with a dense multi-view system. We curate two sequences that feature loose garments for novel view synthesis evaluation. Note that DynaCap does not provide dense scans for reconstruction comparison." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.787, + 0.552 + ], + "angle": 0, + "content": "In-the-wild videos: We use in-the-wild videos collected from DeepCap [15] and online videos to demonstrate the robustness and generalization of our method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.787, + 0.599 + ], + "angle": 0, + "content": "Evaluation Protocol: We report Chamfer distance \\((\\mathbf{C} - \\ell_2)\\) [cm], normal consistency (NC), and volumetric IoU for surface reconstruction comparison. Novel view synthesis quality is measured via PSNR, SSIM [46], and LPIPS \\((\\times 100)\\) [57]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.619, + 0.567, + 0.635 + ], + "angle": 0, + "content": "4.2 Surface Reconstruction Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.842 + ], + "angle": 0, + "content": "We compare our proposed human surface reconstruction method to several state-of-the-art approaches [8, 13, 18] on our MonoLoose dataset. SelfRecon [18] and Vid2Avatar [13] deploy neural rendering to reconstruct the 3D clothed human using a single layer. SCARF [8] reconstructs a hybrid human model based on an explicit inner body and NeRF-based clothing model. All baseline methods rely on SMPL skeleton skinning transformation with additional deformation fields for the garments' motion. Our method outperforms all baselines by a substantial margin on all evaluation metrics (cf. Tab. 1). This disparity becomes more visible in qualitative comparisons shown in Fig. 2. When the dynamic loose garments necessitate large non-skeletal surface deformations, all baseline methods fail to recover complete human surfaces or produce implausible and corrupted reconstructions with visible artifacts (e.g., discontinuities between legs and missing dress parts, see highlights in Fig. 2). Furthermore, they tend to produce less" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.686, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ReLoo" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.145, + 0.773, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.531 + ], + "angle": 0, + "content": "Fig. 2: Qualitative 3D surface reconstruction comparison. Baseline methods produce less detailed and implausible 3D clothed human reconstructions with visible artifacts (discontinuities between legs, missing dress parts) due to the strong reliance on skeletal deformations. In contrast, our method correctly recovers the clothing dynamics and generates more detailed and complete 3D human surfaces. Note also that ReLoo produces more detailed facial features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.563, + 0.788, + 0.669 + ], + "angle": 0, + "content": "fine-grained details (e.g., the faces shown in the third row and the T-shirts shown in the last row of Fig. 2). In contrast, ReLoo generates complete and plausible 3D human shapes with considerably more details (e.g., clothing wrinkles that fully align with image observations). ReLoo also clearly outperforms the baselines for the surface reconstruction under unseen views (see white background columns in Fig. 2). We attribute this superiority to our proposed neural layered clothed human representation and novel virtual bone deformation module." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.692, + 0.553, + 0.709 + ], + "angle": 0, + "content": "4.3 Novel View Synthesis Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We compare with the same baselines for the task of novel view synthesis on MonoLoose and DynaCap [14]. We choose an unseen camera from the respective dataset as a novel view for all methods. As shown in Tab. 2, our method outperforms all baseline methods w.r.t. all metrics with an especially large margin on MonoLoose. Fig. 3 shows that SDF-based methods SelfRecon [18] and Vid2Avatar [13] strongly rely on skeletal deformation and cannot accurately recover the correct 3D shapes, leading to similar artifacts in novel view rendering as in surface reconstruction. Although the NeRF-based method SCARF [8] can" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.451, + 0.129 + ], + "angle": 0, + "content": "C. Guo and T. Jiang et al." + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.159, + 0.744, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.366, + 0.788, + 0.422 + ], + "angle": 0, + "content": "Fig. 3: Qualitative novel view synthesis comparison. Our method achieves better rendering quality with detailed texture recovery in e.g., garment patterns and faces. Baseline methods can only produce corrupted and blurry rendering results (dress discontinuities between legs and unsharp texture details)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.452, + 0.788, + 0.498 + ], + "angle": 0, + "content": "fit to the coarse shape provided by the contour in image observations, it only produces blurry rendering results. ReLoo, in contrast, produces more plausible and realistic renderings while preserving sharper and fine-grained texture details." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.522, + 0.715, + 0.537 + ], + "angle": 0, + "content": "4.4 Qualitative Comparisons with Template-based Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.788, + 0.685 + ], + "angle": 0, + "content": "Since there is no open-sourced template-based monocular method, we conduct a qualitative comparison with DeepCap [15], which is a learning-based method that predicts the template transformation given image observations. Our method is based on implicit neural fields which are topologically flexible and are not limited to a fixed resolution. As shown in Fig. 4, our method better recovers both the human surface details (e.g., faces and wrinkles) and large non-rigid clothing deformations. More importantly, our representation allows topological changes (empty space between dresses shown in the top row of Fig. 4), while the template-based method is inherently bound to the pre-scanned template mesh." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.708, + 0.391, + 0.723 + ], + "angle": 0, + "content": "4.5 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Multi-Layer Volume Rendering. To learn the layered human representation, we leverage multi-layer volume rendering that is SDF-based and includes an inverse CDF sampling process. This means we perform multi-round sampling, where we sample two layers individually and combine the samples through sorting. To investigate the effect of this sampling strategy, we compare it to one-round sampling whereby we join the implicit SDFs of the human body and clothing simply by computing the minimum of the joined function. Results in Tab. 1" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.686, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ReLoo" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 2: Quantitative evaluation on novel view synthesis. We report the quantitative results on test views. Our method consistently outperforms other baseline methods on both datasets and all quantitative evaluation metrics, showing more realistic and plausible rendering quality (cf. Fig. 3)." + }, + { + "type": "table", + "bbox": [ + 0.231, + 0.213, + 0.773, + 0.334 + ], + "angle": 0, + "content": "
MonoLooseDynaCap [14]
MethodPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
SelfRecon [18]22.50.9536.0826.80.9821.56
Vid2Avatar [13]25.90.9684.6627.10.9831.82
SCARF [8]23.30.9536.5925.50.9792.55
Ours w/o Virtual Bone28.70.9693.8127.30.9821.52
Ours29.20.9703.1527.90.9851.27
" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.354, + 0.303, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.481, + 0.277, + 0.49 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.354, + 0.388, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.481, + 0.37, + 0.49 + ], + "angle": 0, + "content": "DeepCap" + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.354, + 0.474, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.402, + 0.481, + 0.463, + 0.49 + ], + "angle": 0, + "content": "ReLoo (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.507, + 0.477, + 0.534 + ], + "angle": 0, + "content": "Fig.4: Qualitative comparisons with template-based method." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.535, + 0.476, + 0.603 + ], + "angle": 0, + "content": "Compared to the template-based method, our representation and learning schemes enable more detailed and realistic human surface reconstruction and topological flexibility." + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.352, + 0.773, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.483, + 0.591, + 0.491 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.618, + 0.483, + 0.772, + 0.491 + ], + "angle": 0, + "content": "One-Round Sampl. Multi-Round Sampl." + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.507, + 0.788, + 0.603 + ], + "angle": 0, + "content": "Fig. 5: Importance of multi-round sampling. One-round sampling strategy can lead to physically implausible clothed human reconstructions with severe garment-body interpenetration while multi-round sampling achieves better holistic reconstructions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.639, + 0.785, + 0.67 + ], + "angle": 0, + "content": "indicate that multi-round sampling helps to improve the holistic clothed human reconstruction quality and avoids garment-body interpenetration (cf. Fig. 5)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.672, + 0.788, + 0.808 + ], + "angle": 0, + "content": "Virtual Bone Deformation Module. An important hyperparameter in our framework is the number of virtual bones \\( n_v \\) for a garment. We quantitatively analyze the effects by learning the clothed human model with different numbers of virtual bones, i.e. \\( n_v \\in \\{20, 40, 80, 160, 320\\} \\). We choose a subset of MonoLoose to evaluate novel view synthesis based on the perceptual similarity metric LPIPS [57] and time per training iteration. The quantitative results are reported in Fig. 6. We observe that the time cost linearly increases, while the error decreases with the sharpest drop at \\( n_v = 80 \\) bones. We thus select this point as the best performance-efficiency compromise." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To validate the effectiveness of our virtual bone deformation module, we compare our full model to a version that uses SMPL-based skeletal deformation for" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.451, + 0.129 + ], + "angle": 0, + "content": "C. Guo and T. Jiang et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.146, + 0.486, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.265, + 0.477, + 0.349 + ], + "angle": 0, + "content": "Fig.6: Number of virtual bones. LPIPS consistently decreases with the increasing number of virtual bones. We choose 80 virtual bones for the garment layer to balance the method performance and efficiency." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.153, + 0.808, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.265, + 0.788, + 0.349 + ], + "angle": 0, + "content": "Fig.7: Importance of virtual bone. Without the virtual bone deformation, our method is bounded by the expressiveness of skeletal movement and cannot accurately capture the topology and motion of loose garments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.788, + 0.492 + ], + "angle": 0, + "content": "the garment layer instead of the virtual bone deformation module. We conducted both quantitative and qualitative analyses as shown in Tab. 2 and Fig. 7. The results demonstrate that integrating the virtual bone deformation module helps to find correct correspondences for points that are not solely controlled by skeletal deformations, leading to plausible and complete novel view synthesis results, while SMPL-based deformation is limited to the hierarchical skeleton structure and struggles with recovering garments that are far away from the inner body." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.531, + 0.36, + 0.547 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.788, + 0.746 + ], + "angle": 0, + "content": "We present ReLoo, a novel method that produces temporally consistent 3D reconstructions of humans when dressed in highly dynamic loose garments from monocular in-the-wild videos. Our method does not require any 3D supervision or prior knowledge about the garments. We utilize a carefully designed layered neural implicit human representation to achieve a disentangled reconstruction of the body and the garment. We introduce a non-hierarchical virtual bone deformation module that enables the accurate capture of non-rigidly deforming loose outfits under articulation. A global optimization is formulated to jointly optimize the shape, appearance, and deformations of both inner body and outer clothing from images via multi-layer volume rendering. Our method achieves robust and high-fidelity reconstruction of humans dressed in loose garments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Limitations: Although readily available, ReLoo relies on reasonable pose estimates and segmentation masks as inputs. Manual adjustment is occasionally required to obtain SAM masks with sharp boundaries. Our method is mainly deployed to up to two garments. The complexity of ReLoo increases linearly with the number of garments that we aim to reconstruct separately. We discuss more limitations and societal impact in the Supp. Mat." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.686, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ReLoo" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.176, + 0.787, + 0.236 + ], + "angle": 0, + "content": "This work was partially supported by the Swiss SERI Consolidation Grant \"AIPERCEIVE\". Chen Guo was partially supported by Microsoft Research Swiss JRC Grant. We are grateful to all our participants for their valued contribution to this research. Computations were carried out in part on the ETH Euler cluster." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.259, + 0.323, + 0.274 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.785, + 0.331 + ], + "angle": 0, + "content": "1. Alldieck, T., Magnor, M., Xu, W., Theobalt, C., Pons-Moll, G.: Video based reconstruction of 3d people models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8387-8397 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.332, + 0.785, + 0.372 + ], + "angle": 0, + "content": "2. Alldieck, T., Zanfir, M., Sminchisescu, C.: Photorealistic monocular 3d reconstruction of humans wearing clothing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.373, + 0.785, + 0.413 + ], + "angle": 0, + "content": "3. Bertiche, H., Madadi, M., Escalera, S.: Pbns: Physically based neural simulation for unsupervised garment pose space deformation. ACM Trans. Graph. 40(6) (dec 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.414, + 0.785, + 0.455 + ], + "angle": 0, + "content": "4. Bhatnagar, B.L., Tiwari, G., Theobalt, C., Pons-Moll, G.: Multi-garment net: Learning to dress 3d people from images. In: IEEE International Conference on Computer Vision (ICCV). IEEE (oct 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.456, + 0.785, + 0.495 + ], + "angle": 0, + "content": "5. Chen, X., Pang, A., Yang, W., Wang, P., Xu, L., Yu, J.: Tightcap: 3d human shape capture with clothing tightness field. ACM Transactions on Graphics (TOG) 41(1), 1-17 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.497, + 0.785, + 0.551 + ], + "angle": 0, + "content": "6. Collet, A., Chuang, M., Sweeney, P., Gillett, D., Evseev, D., Calabrese, D., Hoppe, H., Kirk, A., Sullivan, S.: High-quality streamable free-viewpoint video. ACM Trans. Graph. 34(4) (jul 2015). https://doi.org/10.1145/2766945, https://doi.org/10.1145/2766945" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.552, + 0.785, + 0.579 + ], + "angle": 0, + "content": "7. Corona, E., Pumarola, A., Alenyà, G., Pons-Moll, G., Moreno-Noguer, F.: SMPlicit: Topology-aware generative model for clothed people. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.58, + 0.785, + 0.62 + ], + "angle": 0, + "content": "8. Feng, Y., Yang, J., Pollefeys, M., Black, M.J., Bolkart, T.: Capturing and animation of body and clothing from monocular video. In: SIGGRAPH Asia 2022 Conference Papers. SA '22 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.621, + 0.785, + 0.647 + ], + "angle": 0, + "content": "9. Geman, S., McClure, D.E.: Statistical methods for tomographic image reconstruction (1987)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.648, + 0.785, + 0.702 + ], + "angle": 0, + "content": "10. Grigorev, A., Thomaszewski, B., Black, M.J., Hilliges, O.: HOOD: Hierarchical graphs for generalized modelling of clothing dynamics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.703, + 0.785, + 0.744 + ], + "angle": 0, + "content": "11. Gropp, A., Yariv, L., Haim, N., Atzmon, M., Lipman, Y.: Implicit geometric regularization for learning shapes. In: Proceedings of Machine Learning and Systems. pp. 3569-3579 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.745, + 0.785, + 0.785 + ], + "angle": 0, + "content": "12. Guo, C., Chen, X., Song, J., Hilliges, O.: Human performance capture from monocular video in the wild. In: 2021 International Conference on 3D Vision (3DV). pp. 889-898. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "13. Guo, C., Jiang, T., Chen, X., Song, J., Hilliges, O.: Vid2 avatar: 3d avatar reconstruction from videos in the wild via self-supervised scene decomposition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.288, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.45, + 0.128 + ], + "angle": 0, + "content": "C. Guo and T. Jiang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "14. Habermann, M., Liu, L., Xu, W., Zollhoefer, M., Pons-Moll, G., Theobalt, C.: Real-time deep dynamic characters. ACM Transactions on Graphics 40(4) (aug 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.787, + 0.23 + ], + "angle": 0, + "content": "15. Habermann, M., Xu, W., Zollhoefer, M., Pons-Moll, G., Theobalt, C.: Deepcap: Monocular human performance capture using weak supervision. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (jun 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.787, + 0.271 + ], + "angle": 0, + "content": "16. He, T., Xu, Y., Saito, S., Soatto, S., Tung, T.: Arch++: Animation-ready clothed human reconstruction revisited. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 11046-11056 (October 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.787, + 0.311 + ], + "angle": 0, + "content": "17. Huang, Z., Xu, Y., Lassner, C., Li, H., Tung, T.: Arch: Animatable reconstruction of clothed humans. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3093-3102 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.787, + 0.352 + ], + "angle": 0, + "content": "18. Jiang, B., Hong, Y., Bao, H., Zhang, J.: Selfrecon: Self reconstruction your digital avatar from monocular video. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.787, + 0.393 + ], + "angle": 0, + "content": "19. Jiang, B., Zhang, J., Hong, Y., Luo, J., Liu, L., Bao, H.: Bcnet: Learning body and cloth shape from a single image. In: European Conference on Computer Vision. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.787, + 0.433 + ], + "angle": 0, + "content": "20. Jiang, W., Yi, K.M., Samei, G., Tuzel, O., Ranjan, A.: Neuman: Neural human radiance field from a single video. In: Proceedings of the European conference on computer vision (ECCV) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.787, + 0.488 + ], + "angle": 0, + "content": "21. Jiang, Z., Guo, C., Kaufmann, M., Jiang, T., Valentin, J., Hilliges, O., Song, J.: Multiply: Reconstruction of multiple people from monocular video in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.514 + ], + "angle": 0, + "content": "22. Ke, L., Ye, M., Danelljan, M., Liu, Y., Tai, Y.W., Tang, C.K., Yu, F.: Segment anything in high quality. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.787, + 0.569 + ], + "angle": 0, + "content": "23. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4015-4026 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.787, + 0.609 + ], + "angle": 0, + "content": "24. Le, B.H., Deng, Z.: Smooth skinning decomposition with rigid bones. ACM Trans. Graph. 31(6) (nov 2012). https://doi.org/10.1145/2366145.2366218, https://doi.org/10.1145/2366145.2366218" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.787, + 0.663 + ], + "angle": 0, + "content": "25. Li, Y., Habermann, M., Thomaszewski, B., Coros, S., Beeler, T., Theobalt, C.: Deep physics-aware inference of cloth deformation for monocular human performance capture. In: 2021 International Conference on 3D Vision (3DV). pp. 373-384. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.663, + 0.787, + 0.704 + ], + "angle": 0, + "content": "26. Lin, W., Zheng, C., Yong, J.H., Xu, F.: Relightable and animatable neural avatars from videos. In: Proceedings of the AAAI Conference on Artificial Intelligence (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.787, + 0.745 + ], + "angle": 0, + "content": "27. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ACM transactions on graphics (TOG) 34(6), 1-16 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.787, + 0.799 + ], + "angle": 0, + "content": "28. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4460-4470 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "29. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: European conference on computer vision. pp. 405-421. Springer (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.686, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ReLoo" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "30. Moon, G., Nam, H., Shiratori, T., Lee, K.M.: 3d clothed human reconstruction in the wild. In: European Conference on Computer Vision (ECCV) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.229 + ], + "angle": 0, + "content": "31. Pan, X., Mai, J., Jiang, X., Tang, D., Li, J., Shao, T., Zhou, K., Jin, X., Manocha, D.: Predicting loose-fitting garment deformations using bone-driven motion networks. In: ACM SIGGRAPH 2022 Conference Proceedings. SIGGRAPH '22, Association for Computing Machinery, New York, NY, USA (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.229, + 0.788, + 0.271 + ], + "angle": 0, + "content": "32. Patel, C., Liao, Z., Pons-Moll, G.: Tailornet: Predicting clothing in 3d as a function of human pose, shape and garment style. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (jun 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.271, + 0.788, + 0.325 + ], + "angle": 0, + "content": "33. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3D hands, face, and body from a single image. In: Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR). pp. 10975-10985 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.325, + 0.788, + 0.352 + ], + "angle": 0, + "content": "34. Peng, S., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Zhou, X., Bao, H.: Animatable neural radiance fields for modeling dynamic human bodies. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.352, + 0.788, + 0.406 + ], + "angle": 0, + "content": "35. Peng, S., Zhang, Y., Xu, Y., Wang, Q., Shuai, Q., Bao, H., Zhou, X.: Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9054-9063 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.406, + 0.788, + 0.447 + ], + "angle": 0, + "content": "36. Pons-Moll, G., Pujades, S., Hu, S., Black, M.: Clothcap: Seamless 4d clothing capture and retargeting. ACM Transactions on Graphics, (Proc. SIGGRAPH) 36(4) (2017), two first authors contributed equally" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.447, + 0.788, + 0.488 + ], + "angle": 0, + "content": "37. Qiu, L., Chen, G., Zhou, J., Xu, M., Wang, J., Han, X.: Rec-mv: Reconstructing 3d dynamic cloth from monocular videos. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.488, + 0.788, + 0.514 + ], + "angle": 0, + "content": "38. Ricci, A.: A constructive geometry for computer graphics. Comput. J. 16, 157-160 (1973), https://api_semanticscholar.org/CorpusID:30038820" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.514, + 0.788, + 0.568 + ], + "angle": 0, + "content": "39. Saito, S., Simon, T., Saragih, J., Joo, H.: Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 84-93 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.609 + ], + "angle": 0, + "content": "40. Santesteban, I., Otaduy, M.A., Casas, D.: Snug: Self-supervised neural dynamic garments. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8140-8150 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.609, + 0.788, + 0.65 + ], + "angle": 0, + "content": "41. Su, S.Y., Bagautdinov, T., Rhodin, H.: Danbo: Disentangled articulated neural body representations via graph neural networks. In: European Conference on Computer Vision (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.65, + 0.788, + 0.704 + ], + "angle": 0, + "content": "42. Su, Z., Wan, W., Yu, T., Liu, L., Fang, L., Wang, W., Liu, Y.: Mulaycap: Multi-layer human performance capture using a monocular video camera. IEEE Transactions on Visualization and Computer Graphics 28(4), 1862-1879 (2022). https://doi.org/10.1109/TVCG.2020.3027763" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.745 + ], + "angle": 0, + "content": "43. Tiwari, G., Bhatnagar, B.L., Tung, T., Pons-Moll, G.: Sizer: A dataset and model for parsing 3d clothing and learning size sensitive 3d clothing. In: European Conference on Computer Vision (ECCV). Springer (August 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.745, + 0.788, + 0.799 + ], + "angle": 0, + "content": "44. Wang, K., Zhang, G., Cong, S., Yang, J.: Clothed human performance capture with a double-layer neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21098-21107 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.788, + 0.841 + ], + "angle": 0, + "content": "45. Wang, S., Schwarz, K., Geiger, A., Tang, S.: Arah: Animatable volume rendering of articulated human sdfs. In: European Conference on Computer Vision (ECCV) (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.45, + 0.128 + ], + "angle": 0, + "content": "C. Guo and T. Jiang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "46. Wang, Z., Simoncelli, E.P., Bovik, A.C.: Multiscale structural similarity for image quality assessment. In: The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers. vol. 2 (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.245 + ], + "angle": 0, + "content": "47. Weng, C.Y., Curless, B., Srinivasan, P.P., Barron, J.T., Kemelmacher-Shlizerman, I.: HumanNeRF: Free-viewpoint rendering of moving people from monocular video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16210-16220 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "48. Xiang, D., Prada, F., Bagautdinov, T., Xu, W., Dong, Y., Wen, H., Hodgins, J., Wu, C.: Modeling clothing as a separate layer for an animatable human avatar. ACM Trans. Graph. 40(6) (dec 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.328 + ], + "angle": 0, + "content": "49. Xiu, Y., Yang, J., Cao, X., Tzionas, D., Black, M.J.: ECON: Explicit Clothed humans Optimized via Normal integration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "50. Xiu, Y., Yang, J., Tzionas, D., Black, M.J.: ICON: Implicit Clothed humans Obtained from Normals. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13296-13306 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.411 + ], + "angle": 0, + "content": "51. Xu, W., Chatterjee, A., Zollhöfer, M., Rhodin, H., Mehta, D., Seidel, H.P., Theobalt, C.: Monoperfcap: Human performance capture from monocular video. SIGGRAPH 37(2), 27:1-27:15 (May 2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.439 + ], + "angle": 0, + "content": "52. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. In: Advances in Neural Information Processing Systems (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.44, + 0.785, + 0.481 + ], + "angle": 0, + "content": "53. Yariv, L., Kasten, Y., Moran, D., Galun, M., Atzmon, M., Ronen, B., Lipman, Y.: Multiview neural surface reconstruction by disentangling geometry and appearance. In: Advances in Neural Information Processing Systems (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.535 + ], + "angle": 0, + "content": "54. Yu, T., Zheng, Z., Zhong, Y., Zhao, J., Dai, Q., Pons-Moll, G., Liu, Y.: Simulcap: Single-view human performance capture with cloth simulation. In: The IEEE International Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (June 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.578 + ], + "angle": 0, + "content": "55. Zhang, J., Liu, X., Ye, X., Zhao, F., Zhang, Y., Wu, M., Zhang, Y., Xu, L., Yu, J.: Editable free-viewpoint video using a layered neural representation. ACM Transactions on Graphics (TOG) 40(4), 1-18 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.579, + 0.785, + 0.605 + ], + "angle": 0, + "content": "56. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv:2010.07492 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.633 + ], + "angle": 0, + "content": "57. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.634, + 0.785, + 0.674 + ], + "angle": 0, + "content": "58. Zhang, Z., Sun, L., Yang, Z., Chen, L., Yang, Y.: Global-correlated 3d-decoupling transformer for clothed avatar reconstruction. In: Advances in Neural Information Processing Systems (NeurIPS) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.716 + ], + "angle": 0, + "content": "59. Zheng, Z., Huang, H., Yu, T., Zhang, H., Guo, Y., Liu, Y.: Structured local radiance fields for human avatar modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.785, + 0.758 + ], + "angle": 0, + "content": "60. Zheng, Z., Yu, T., Liu, Y., Dai, Q.: Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.758 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/ReLoo_ Reconstructing Humans Dressed in Loose Garments from Monocular Video in the Wild/e08d3ae3-38a4-42db-b159-2c5a46fac242_origin.pdf b/2024/ReLoo_ Reconstructing Humans Dressed in Loose Garments from Monocular Video in the Wild/e08d3ae3-38a4-42db-b159-2c5a46fac242_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..90e2c41bf36d541f97a4dae662a8ad6282a50c2e --- /dev/null +++ b/2024/ReLoo_ Reconstructing Humans Dressed in Loose Garments from Monocular Video in the Wild/e08d3ae3-38a4-42db-b159-2c5a46fac242_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db275c5961ca22686627036e3391ac55b1a08abc9ac24a9a26cd85425bd443be +size 9650051 diff --git a/2024/ReLoo_ Reconstructing Humans Dressed in Loose Garments from Monocular Video in the Wild/full.md b/2024/ReLoo_ Reconstructing Humans Dressed in Loose Garments from Monocular Video in the Wild/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7b3417fea9415dae1c3662034b5ce024a2332d74 --- /dev/null +++ b/2024/ReLoo_ Reconstructing Humans Dressed in Loose Garments from Monocular Video in the Wild/full.md @@ -0,0 +1,322 @@ +# ReLoo: Reconstructing Humans Dressed in Loose Garments from Monocular Video in the Wild + +Chen Guo\*, Tianjian Jiang\*, Manuel Kaufmann, Chengwei Zheng, Julien Valentin2, Jie Song†, and Otmar Hilliges + +$^{1}$ ETH Zürich +2 Microsoft + +![](images/7410daf7c917e3bb4895c227b68fde001bf35658a108c48bb4a453a724c8d7a8.jpg) + +Abstract. While previous years have seen great progress in the 3D reconstruction of humans from monocular videos, few of the state-of-the-art methods are able to handle loose garments that exhibit large non-rigid surface deformations during articulation. This limits the application of such methods to humans that are dressed in standard pants or T-shirts. Our method, ReLoo, overcomes this limitation and reconstructs high-quality 3D models of humans dressed in loose garments from monocular in-the-wild videos. To tackle this problem, we first establish a layered neural human representation that decomposes clothed humans into a neural inner body and outer clothing. On top of the layered neural representation, we further introduce a non-hierarchical virtual bone deformation module for the clothing layer that can freely move, which allows the accurate recovery of non-rigidly deforming loose clothing. A global optimization jointly optimizes the shape, appearance, and deformations of the human body and clothing via multi-layer differentiable volume rendering. To evaluate ReLoo, we record subjects with dynamically deforming garments in a multi-view capture studio. This evaluation, both on existing and our novel dataset, demonstrates ReLoo's clear superiority over prior art on both indoor datasets and in-the-wild videos. + +Keywords: human representation $\cdot$ human reconstruction + +# 1 Introduction + +As researchers aim to democratize the creation of realistic human avatars, the reconstruction of 3D clothed humans from casually captured monocular videos has garnered increased attention. While many solutions have been proposed to do so in recent years [1,8,13,18,39,50], they primarily focus on capturing subjects with tight-fitting clothing and perform poorly when reconstructing loose garments whose dynamics are less tightly coupled with body pose. Such loose garments, however, constitute a significant part of a real-life wardrobe and thus failure to capture them is limiting the creation of realistic human avatars from monocular footage. In this paper, we present a method, ReLoo, that overcomes this shortcoming. + +Reconstructing humans dressed in loose clothing requires accurate tracking of large, non-rigid deformations and recovering fine-grained details of freely flowing surfaces. Template-based methods [14, 15, 51] have been applied to do so but the acquisition of the template burdens the deployment to unseen subjects and its explicit representation limits its expressive capability to capture dynamically changing surface details. More recently, methods based on neural implicit functions have emerged as a promising remedy for the disadvantages of template-based methods [13, 16, 18, 20, 35, 39, 41, 47, 49, 58-60]. Yet, these methods currently fail to provide convincing reconstructions of humans in loose garments. We observe that few of these methods differentiate between the human body and clothing but model the clothed human as a single entity. This limits the expressiveness and capacity of the underlying model to capture more local features. More importantly, this formulation only allows to drive the off-body garments with skeletal deformations that are derived from the underlying parametric body model (e.g., SMPL [27]). Thus, they are inherently incapable of handling highly dynamic loose garments that are topologically different from the inner body and do not correlate strongly with the bone movement. + +In this paper, we adopt the promising neural implicit shape modeling paradigm, but we argue that a single implicit representation fundamentally limits the representation power and hinders the capability to model complex garment topology that exhibits free-form deformations. To properly model loose garment dynamics that only weakly correlate with skeletal deformations - while still retaining the ability to deform human bodies with skeleton-driven motions - ReLoo takes neural implicit human models to the next level. To do so, our approach is grounded in the following core concepts: i) We establish a layered neural human representation that decomposes clothed humans into the neural inner body and outer clothing. ii) Based on this layered neural representation, we further propose a non-hierarchical virtual bone deformation module for the clothing layer that allows free movement and accurate recovery of highly dynamic loose outfits. iii) In a global optimization, we jointly optimize the shape, appearance, and deformations of both the human body and clothing layer over the entire sequence via multi-layer differentiable volume rendering. + +In our experiments, we demonstrate that our framework leads to temporally consistent and high-quality reconstructions of clothed humans dressed in loose + +garments. We also ablate our method to uncover the contribution of its essential components. Furthermore, we conduct comparisons with existing approaches in human surface reconstruction and novel view synthesis, showing that our method outperforms prior art from both the template-based and neural implicit modeling domains. To highlight differences to the prior art, we capture a new dataset, MonoLoose, which puts an emphasis on humans dressed in loose clothing under dynamic motions and contains ground-truth reconstructions captured with a high-end multi-view volumetric recording studio (MVS). + +In summary, in this paper we introduce ReLoo, a method that improves clothed human reconstruction quality and accurately captures human performance dressed in highly dynamic loose garments. Our key contributions are: + +- a novel layered neural human representation, disentangling the inner body and outer clothing; and +- a virtual bone deformation module that is built on top of the layered neural human representation and accurately tracks the large surface dynamics; and +- a robust framework that leverages multi-layer differentiable volume rendering achieving high-fidelity 3D human reconstructions from monocular in-the-wild videos of humans dressed in highly dynamic loose garments. + +# 2 Related Work + +Single-Layer Human Reconstruction from Monocular Input Template-based monocular human performance capture methods track the pre-defined clothed human template to fit to 2D observations [14,15,51]. They demonstrate robust tracking of human performance even when dressed in loose garments. However, they struggle to generalize to in-the-wild settings due to the reliance on a rigged, personalized, pre-scanned template obtained from a dense capture setup. Follow-up works endeavor to remove this dependency by adding per-vertex offsets on top of the SMPL body [1,12]. Nevertheless, the explicit mesh representation is held back by a fixed resolution and topology, hampering the representation of fine-grained details. Other works have shown compelling results with learning-based methods that learn to regress 3D human geometry and appearance from images [2,16,17,39,49,50,58,60]. A major limitation of these methods is the necessity of high-quality 3D data for supervision. They also often fail to produce space-time coherent reconstructions over frames. Recent works employ neural rendering to fit neural fields to videos to obtain an articulated human model [8,13,18,20,21,26,34,35,37,45,47]. E.g., SelfRecon [18] deploys neural surface rendering [53] to achieve consistent reconstruction over the sequence and Vid2Avatar [13] leverages differentiable volume rendering [52] to eliminate the need for pre-masking thus producing robust 3D human reconstruction. However, all aforementioned methods treat the body and clothing as a single entity, limiting the model's expressiveness and resulting in low-quality reconstructions for loose clothing. In contrast, our method is based on a layered neural human representation that, in conjunction with our novel virtual bone deformation module, enables tracking of highly dynamic loose clothing. + +Multi-Layer Human Representation and Reconstruction Several methods exist that investigate how a clothing layer deforms given 3D human motion. They use either physically-simulated training data [3,25,31,32,42,44] or directly deploy physics-informed objectives [10, 40]. These methods have shown compelling results in modeling large deformations of loose outfits. However, they lack generalization to in-the-wild settings and diverse clothing categories due to a reliance on input templates [25, 44] and assume high-quality 3D motion data is already available, not learned from video like in our setting. Our multi-layer representation and deformation modeling are inspired by this line of work, specifically Pan et al. [31], who also make use of virtual bones. Different from [31], our method jointly learns the clothing shape and deformations from only 2D observations, whereas [31] require known clothing templates and 3D simulated data to achieve animation of clothes. Furthermore, [31] define and fix virtual bones using skinning decomposition [24], while in our framework the virtual bone placement evolves naturally with training. + +Methods that reconstruct layered clothed humans from videos or images alone are presented in [5, 7, 8, 19, 30, 36, 43, 48, 54]. Some of them either require active multi-view setups [4, 5, 36, 48] or depth information [54] preventing them from being deployed in the wild. BCNet [19] learns to predict clothing geometry draped over the SMPL body from a single image. However, it is limited to pre-defined clothing style templates. SMPLicit [7] and ClothWild [30] extend such learning-based methods to generalize to more general clothing types, but they tend to produce over-smooth results lacking details such as wrinkles. Our closest related work, SCARF [8] is built upon SMPL-X [33] and reconstructs the outer clothing layer using NeRF [29], achieving better reconstruction quality than previous works. Our method differs from SCARF [8] in three regards: 1) ReLoo is a fully implicit representation that is expressive enough to capture detailed body (including faces) and clothing shape jointly, 2) it is not limited to self-rotating motions, and 3) it supports the capture of large non-rigid surface deformations of loose garments thanks to a novel virtual bone deformation module. + +In summary, existing monocular-based methods tend to overly rely on parametric body models [27,33] as a human body proxy and thus struggle to model garments whose deformations cannot be easily correlated with the inner body pose under dynamic motion (as is the case for t-shirts or pants that are commonly used as example apparel). Moreover, they can only capture less detailed personalized shape characteristics such as faces. ReLoo overcomes these limitations and when compared to existing single-layer, multi-layer, and template-based methods produces higher fidelity results across the board. + +# 3 Method + +We introduce ReLoo, a novel method for detailed geometry and appearance reconstruction of a human performer in highly dynamic, loose clothing from monocular in-the-wild videos. The overview of our method is schematically given in Fig. 1. To enable a temporally consistent, expressive human representation + +![](images/20d9f366b461bc6059bc72f9c69b672c6b6109c3631e1124bc429f53c21825dd.jpg) +Fig. 1: Method Overview. Given an image from a video sequence, we sample points along the camera ray for each neural layer. We warp sampled points for the body layer $\pmb{x}_d^B$ into canonical space via inverse LBS derived from skeletal deformations. We deform sampled points for the garment layer $\pmb{x}_d^G$ into canonical space via inverse warping based on the proposed virtual bone deformation module (Sec. 3.2). We then evaluate the respective implicit network to obtain the SDF and radiance values (Sec. 3.1). We apply multi-layer differentiable volume rendering to learn the shape, appearance, and deformations of the layered neural human representation from images (Sec. 3.3). The loss function $\mathcal{L}$ compares the rendered color predictions with image observations as well as a segmentation mask obtained using SAM [22] (Sec. 3.4). + +we establish a layered neural implicit representation for the body (inner layer, naked) and garment (outer layer, template-free) as described in Sec. 3.1. On top, we introduce a hybrid deformation strategy which consists of skeletal deformation for the body and a virtual-bone-driven deformation module for the outer loose garment (Sec. 3.2). This allows us to capture dynamic loose garments. We learn the layered human representation and the deformation module jointly by performing multi-layer differentiable volume rendering (Sec. 3.3). The whole process is trained globally to optimize jointly for shape, appearance, and deformations of the inner body and outer garment layer (Sec. 3.4). + +# 3.1 Layered Neural Human Representation + +We represent the 3D shape of the clothed human with implicit signed-distance fields (SDF) and the appearance with texture fields in a temporally consistent canonical space. The inner body and outer garment are modeled separately. More specifically, we model the geometry and appearance of the body in canonical space with a neural network $f^{B}$ which predicts the signed distance value $s^{B}$ and radiance value $c^{B}$ for any 3D point $\boldsymbol{x}_{c}^{B}$ in this space. We similarly model the garment with a neural network $f^{G}$ that takes points $\boldsymbol{x}_{c}^{G}$ in the garment's + +canonical space as input: + +$$ +\boldsymbol {c} ^ {B}, s ^ {B} = f ^ {B} \left(\boldsymbol {x} _ {c} ^ {B}, \boldsymbol {\theta}\right); \boldsymbol {c} ^ {G}, s ^ {G} = f ^ {G} \left(\boldsymbol {x} _ {c} ^ {G}, \boldsymbol {\theta}\right), \tag {1} +$$ + +where $\theta$ denote the SMPL pose parameters [27], which we concatenate to $\pmb{x}_c^B$ and $\pmb{x}_c^G$ to model pose-dependent effects such as facial features and clothing wrinkles. For clothing that is not a single, dress-like garment we use a separate network for the upper and lower garment. For simplicity and without loss of generality, we only discuss a single piece of clothing in the following. Note that if a canonical point is within any of these two surfaces ( $s^B < 0$ or $s^G < 0$ ), it is also within the entire clothed human shape. Thus we can obtain the final clothed human shape by compositing these two neural fields and taking their minimum [38]: + +$$ +s ^ {H} = \min \left\{s ^ {B}, s ^ {G} \right\}. \tag {2} +$$ + +# 3.2 Hybrid Deformation Modeling + +To find point correspondences between the deformed space in the observed image and our pre-defined canonical space, we devise a hybrid deformation module that treats the body and clothing layers according to their respective levels of rigidity. The human body predominantly depends on skeletal deformation, but the deformation of the loose garments cannot solely be explained by skeletal motion alone. Therefore we propose to drive clothing deformation by a set of additional virtual bones whose transformations are directly learned from video. + +Skeletal Deformation. We follow a standard skeletal deformation based on SMPL to find correspondences in canonical and deformed space for the inner body [13]. Specifically, given the bone transformation matrices $\mathbf{B}_i$ for joints $i\in \{1,\dots,n_b\}$ which are derived from the body pose parameters $\pmb{\theta}$ , a canonical point $\pmb{x}_c^B$ is mapped to the corresponding deformed space point $\pmb{x}_d^B$ via linear blend skinning (LBS): + +$$ +\boldsymbol {x} _ {d} ^ {B} = \sum_ {i = 1} ^ {n _ {s}} w _ {c} ^ {i} \boldsymbol {B} _ {i} \boldsymbol {x} _ {c} ^ {B}. \tag {3} +$$ + +Conversely, given a point in deformed space $\pmb{x}_d^B$ , its canonical correspondence can be solved via: + +$$ +\boldsymbol {x} _ {c} ^ {B} = \left(\sum_ {i = 1} ^ {n _ {s}} w _ {d} ^ {i} \boldsymbol {B} _ {i}\right) ^ {- 1} \boldsymbol {x} _ {d} ^ {B}. \tag {4} +$$ + +Here, $n_s$ denotes the number of skeletal bones in the transformation, and $\boldsymbol{w}_{(\cdot)} = \{w_{(\cdot)}^1, \dots, w_{(\cdot)}^{n_s}\}$ represents the skinning weights for $\boldsymbol{x}_{(\cdot)}^B$ . We assign $\boldsymbol{w}_d$ to $\boldsymbol{x}_d^B$ based on the average of neighboring SMPL vertices' skinning weights, weighted by the point-to-point distances in deformed space. The treatment of canonical points $\boldsymbol{x}_c^B$ follows a similar approach. + +Virtual Bone Deformation. The virtual bones correspond to a set of bones $\mathcal{V} = \{\pmb{v}_i\}_{i=1}^{n_v}$ defined in the canonical space that drives the neighboring 3D garment points $\pmb{x}_c^G$ using rigid transformations. Different from the skeletons of characters rigged by artists such as SMPL [27], the virtual bones are non-hierarchical + +and are not restricted to rotate in relation to their parents following an anatomical structure. Thus, they can be transformed freely and are applicable to capture the deformations of highly dynamic loose garments. + +Given a set of virtual bones $\mathcal{V}$ , their transformations consist of the rotations and translations relative to the SMPL root $\pmb{\mathcal{T}}_i = [R_i|T_i]$ which are predicted by a deformation field $\mathcal{D}^G$ , parameterized via an MLP. $\mathcal{D}^G$ takes the concatenation of the 3D positions $\pmb{v}_i$ of the virtual bones, the human body pose $\pmb{\theta}$ and a continuous time embedding $\pmb{t}$ as input, and outputs the axis angles $\pmb{A}_i$ and translations $\pmb{T}_i$ . The continuous time embedding $\pmb{t}$ is included to aid in learning temporal dynamics from videos. The rotations $\pmb{R}_i$ are further obtained from $\pmb{A}_i$ via the Rodrigues' formula $f^{Rod}(\cdot)$ : + +$$ +\boldsymbol {\mathcal {T}} _ {i} = \left[ f ^ {R o d} (\boldsymbol {A} _ {i}) | \boldsymbol {T} _ {i} \right] = \mathcal {D} ^ {G} (\boldsymbol {v} _ {i}, \boldsymbol {\theta}, t). \tag {5} +$$ + +To drive a 3D garment point from canonical space $\pmb{x}_c^G$ to deformed space $\pmb{x}_d^G$ and vice-versa, we use the neighboring virtual bones' motions and LBS as follows: + +$$ +\boldsymbol {x} _ {d} ^ {G} = \sum_ {i = 1} ^ {n _ {v}} \delta_ {c} ^ {i} \boldsymbol {\mathcal {T}} _ {i} \boldsymbol {x} _ {c} ^ {G}, \quad \boldsymbol {x} _ {c} ^ {G} = \left(\sum_ {i = 1} ^ {n _ {v}} \delta_ {d} ^ {i} \boldsymbol {\mathcal {T}} _ {i}\right) ^ {- 1} \boldsymbol {x} _ {d} ^ {G} \tag {6} +$$ + +Here, $n_v$ denotes the number of virtual bones for deforming loose garments, which is a hyperparameter, and $\pmb{\delta}_{(\cdot)} = \{\delta_{(\cdot)}^1,\dots,\delta_{(\cdot)}^{n_v}\}$ represent the skimming weights of $\pmb{x}_{(\cdot)}^G$ w.r.t. each virtual bone $\pmb{v}_i\in \mathcal{V}$ . $\pmb{\delta}_{(\cdot)}$ is calculated based on the inverse of the distance between $\pmb{x}_c^G$ and each $\pmb{v}_i$ whereby far away bones are clamped to 0. More details are shown in the Supp. Mat. + +Unlike [31], where a garment template is available and the virtual bones are extracted using 3D simulation data, we learn the clothed human model from monocular observations solely and extract the virtual bones on the fly without requiring any specific template prior. We explain the acquisition more in Sec. 3.4. + +# 3.3 Multi-Layer Volume Rendering + +As we are aiming to jointly reconstruct multiple layers of neural implicit fields, we are required to depart from standard differentiable volume rendering for static scenes (e.g., [29]). We thus introduce multi-layer volume rendering tailored to our multi-layer human representation (Sec. 3.1) to reconstruct both inner body and outer clothing under garment-body occlusions. To do so, we use surface-based volume rendering [52] while re-ordering multiple neural layers [55]. + +Specifically, we shoot a ray $\pmb{r}$ through every pixel of the image and sample two sets of points in the body layer and the garment layer: $\{\pmb{x}_{d,1}^{B},\dots,\pmb{x}_{d,N}^{B}\}$ and $\{\pmb{x}_{d,1}^{G},\dots,\pmb{x}_{d,N}^{G}\}$ following the two-stage sampling strategy proposed in [52]. Note that both sets contain the same amount of points. Next, we use the skeletal deformation to warp each sampled point $\pmb{x}_{d,i}^{B}$ to the body's canonical space $\pmb{x}_{c,i}^{B}$ and we use our virtual bone deformation module to find canonical correspondences $\pmb{x}_{c,i}^{G}$ to each sampled point $\pmb{x}_{d,i}^{G}$ (Sec. 3.2). Then, we obtain the corresponding signed-distance $(s_i^B$ and $s_i^G)$ and radiance values $(\pmb{c}_i^B$ and $\pmb{c}_i^G)$ by querying the + +implicit shape and texture fields with the canonical points. We then compute the occupancy for the inner and outer layer at the $i$ -th sampled point as: + +$$ +o _ {i} ^ {B} = \left(1 - \exp \left(- \sigma_ {i} ^ {B} \Delta \boldsymbol {x} _ {i} ^ {B}\right)\right); \sigma_ {i} ^ {B} = \sigma \left(s _ {i} ^ {B}, \boldsymbol {\theta}\right) \tag {7} +$$ + +$$ +o _ {i} ^ {G} = \left(1 - \exp \left(- \sigma_ {i} ^ {G} \Delta \boldsymbol {x} _ {i} ^ {G}\right)\right); \sigma_ {i} ^ {G} = \sigma \left(s _ {i} ^ {G}, \boldsymbol {\theta}\right) \tag {8} +$$ + +where $\Delta \pmb{x}_i^{(\cdot)}$ is the distance between two adjacent sample points in the respective layer, and $\sigma (\cdot)$ represents the scaled Laplace's Cumulative Distribution Function (CDF) to convert $s_i^B$ and $s_i^G$ to volume densities $(\sigma_{i}^{B}$ and $\sigma_{i}^{G})$ following [52]. Finally, we integrate the radiance numerically for both layers and obtain the neurally rendered color of the human performer $\hat{C}^H$ : + +$$ +\hat {C} ^ {H} = \sum_ {i = 1} ^ {N} \sum_ {p \in \{B, G \}} \left[ o _ {i} ^ {p} \boldsymbol {c} _ {i} ^ {p} \prod_ {q \in \{B, G \}} \prod_ {j \in \mathcal {I} _ {i} ^ {q, p}} \left(1 - o _ {j} ^ {q}\right) \right], \tag {9} +$$ + +where $\mathcal{I}_i^{q,p} = \{j\in [1,N]\mid z(\pmb{x}_{d,j}^q) < z(\pmb{x}_{d,i}^p)\}$ and $z(\cdot)$ measures the depth of a point w.r.t. the camera origin. In other words, we sort all points according to their depth values from near to far and then conduct the volumetric integration. + +Scene Composition. To model the background of the scene we use NeRF++ [56], denoted as $f^S$ , which estimates a color value $\hat{C}^S$ representing the scene's color. The final pixel color $\hat{C}$ is a composite of $\hat{C}^S$ with $\hat{C}^H$ following [13]. More details are shown in the Supp. Mat. + +# 3.4 Global Optimization + +To jointly learn the inner body and outer clothing of clothed humans from monocular videos in a template-free manner, we propose a two-stage training schema. The whole training process is formulated as a global optimization over all estimizable parameters and the entire video sequence. + +Two-Stage Training Schema. In the first stage, we leverage skeletal deformation to deform both the body and garment layer. Meanwhile, we warm up the virtual bone deformation field $\mathcal{D}^G$ by encouraging $\mathcal{D}^G$ to have similar deformations as the SMPL model around near-body regions. In the second stage, we activate the virtual bone deformation module to drive the garment layer. To obtain virtual bones, we generate garment meshes using Multiresolution IsoSurface Extraction (MISE) [28]. In theory, each of the $M$ vertices of the resulting canonical garment mesh can be made a virtual bone. This will however incur a high computational cost in the deformation module. To mitigate this but still retain expressive capability we employ a quadric mesh simplification algorithm to reduce the number of vertices to $n_v \ll M$ . The remaining vertices after the simplification are the initial virtual bone locations. We empirically found 80 virtual bones to deliver the best performance-efficiency compromise (see Sec. 4.5). Note that we periodically generate new sets of virtual bones during training to account for changing garment topologies. + +Reconstruction Loss. For every ray $\boldsymbol{r} \in \mathcal{R}$ we compute how well the rendered color $\hat{C}(\boldsymbol{r})$ matches the image pixel's RGB value $C(\boldsymbol{r})$ with the $L_{1}$ -distance: + +$$ +\mathcal {L} _ {\mathrm {r g b}} = \frac {1}{| \mathcal {R} |} \sum_ {\boldsymbol {r} \in \mathcal {R}} | C (\boldsymbol {r}) - \hat {C} (\boldsymbol {r}) |. \tag {10} +$$ + +Segmentation Loss. We modify Eq. (9) to render the opacity $\hat{O}^B (\boldsymbol {r})$ and $\hat{O}^G (\boldsymbol {r})$ per pixel for both layers: + +$$ +\hat {O} ^ {B} (\boldsymbol {r}) = \sum_ {i = 1} ^ {N} \left[ o _ {i} ^ {B} \prod_ {q \in \{B, G \}} \prod_ {j \in \mathcal {I} _ {i} ^ {q, B}} \left(1 - o _ {j} ^ {q}\right) \right]; \hat {O} ^ {G} (\boldsymbol {r}) = \sum_ {i = 1} ^ {N} \left[ o _ {i} ^ {G} \prod_ {q \in \{B, G \}} \prod_ {j \in \mathcal {I} _ {i} ^ {q, G}} \left(1 - o _ {j} ^ {q}\right) \right]. \tag {11} +$$ + +The segmentation loss is calculated between the rendered pixel-wise opacity and the segmentation masks extracted using SAM [23]. A robust Geman-McClure error function $\rho$ [9] is applied to down-weigh potentially erroneous cloth segmentation mask predictions (more details are explained in Supp. Mat): + +$$ +\mathcal {L} _ {\text {s e g}} = \frac {1}{| \mathcal {R} |} \sum_ {\boldsymbol {r} \in \mathcal {R}} \sum_ {p \in \{B, G \}} \rho \left(\mathcal {M} _ {\text {s a m}} ^ {p} (\boldsymbol {r}) - \hat {O} ^ {p} (\boldsymbol {r})\right). \tag {12} +$$ + +Adaptive Eikonal Loss. We follow IGR [11] and sample points in the canonical space to compute the Eikonal constraint to regularize the validity of our SDFs in each layer. Unlike [11], which randomly samples points in the entire space, we periodically extract the canonical shapes for both layers and sample points around the explicit mesh surfaces: + +$$ +\mathcal {L} _ {\text {e i k o n a l}} = \mathbb {E} _ {\boldsymbol {x} _ {c} ^ {B}} \left(\left\| \nabla s ^ {B} \right\| - 1\right) ^ {2} + \mathbb {E} _ {\boldsymbol {x} _ {c} ^ {G}} \left(\left\| \nabla s ^ {G} \right\| - 1\right) ^ {2}. \tag {13} +$$ + +Virtual Bone Deformation Regularization. To accelerate the convergence of the virtual bone deformation field in the first stage of the training process, we randomly sample 3D canonical garment points $\pmb{x}_c^G$ and apply additional SMPL transformation regularization which ensures that the virtual bone deformation $LBS^{G}$ do not deviate excessively from the transformations made by skeletal deformation $LBS^{B}$ during the warm-up stage: + +$$ +\mathcal {L} _ {\text {r e g}} = \left\| L B S ^ {B} \left(\boldsymbol {x} _ {c} ^ {G}\right) - L B S ^ {G} \left(\boldsymbol {x} _ {c} ^ {G}\right) \right\| ^ {2} \tag {14} +$$ + +See Supp. Mat for more details about the final loss. + +# 4 Experiments + +We first introduce the datasets and metrics used for evaluation. Then, we compare our proposed method with state-of-the-art approaches in two tasks: 3D surface reconstruction and novel view synthesis. Ablation studies are then conducted to demonstrate the effectiveness of our core components and design choices. + +Table 1: Quantitative evaluation on surface reconstruction. We compute the 3D surface metrics on the MonoLoose dataset. Our method consistently outperforms all baselines on all evaluation metrics (cf. Fig. 2). + +
MethodC-ℓ2↓NC↑V-IoU↑
SelfRecon [18]2.220.7880.844
Vid2Avatar [13]2.340.7940.776
SCARF [8]3.130.7110.691
Ours w/o Multi-Round Sampl.2.340.7700.879
Ours1.930.8310.881
+ +# 4.1 Datasets + +MonoLoose Dataset: Due to the lack of datasets that capture dynamic human performance with high-fidelity 3D ground-truth meshes when dressed in loose garments, we captured our own dataset, MonoLoose, with a high-end multiview volumetric capture studio (MVS) [6]. This dataset is specifically curated for evaluating monocular human surface reconstruction and novel view synthesis methods, with a particular focus on subjects dressed in loose attire. It consists of five sequences with different identities, loose garment styles, and motions. For more details on the contents of MonoLoose, please refer to the Supp. Mat. + +DynaCap [14]: We further evaluate our method on DynaCap, which captures dynamic human performance with a dense multi-view system. We curate two sequences that feature loose garments for novel view synthesis evaluation. Note that DynaCap does not provide dense scans for reconstruction comparison. + +In-the-wild videos: We use in-the-wild videos collected from DeepCap [15] and online videos to demonstrate the robustness and generalization of our method. + +Evaluation Protocol: We report Chamfer distance $(\mathbf{C} - \ell_2)$ [cm], normal consistency (NC), and volumetric IoU for surface reconstruction comparison. Novel view synthesis quality is measured via PSNR, SSIM [46], and LPIPS $(\times 100)$ [57]. + +# 4.2 Surface Reconstruction Comparisons + +We compare our proposed human surface reconstruction method to several state-of-the-art approaches [8, 13, 18] on our MonoLoose dataset. SelfRecon [18] and Vid2Avatar [13] deploy neural rendering to reconstruct the 3D clothed human using a single layer. SCARF [8] reconstructs a hybrid human model based on an explicit inner body and NeRF-based clothing model. All baseline methods rely on SMPL skeleton skinning transformation with additional deformation fields for the garments' motion. Our method outperforms all baselines by a substantial margin on all evaluation metrics (cf. Tab. 1). This disparity becomes more visible in qualitative comparisons shown in Fig. 2. When the dynamic loose garments necessitate large non-skeletal surface deformations, all baseline methods fail to recover complete human surfaces or produce implausible and corrupted reconstructions with visible artifacts (e.g., discontinuities between legs and missing dress parts, see highlights in Fig. 2). Furthermore, they tend to produce less + +![](images/806a18f21ad2247b3e03ed01d2cd1908271c82104db77f026bf279271a5a5ccd.jpg) +Fig. 2: Qualitative 3D surface reconstruction comparison. Baseline methods produce less detailed and implausible 3D clothed human reconstructions with visible artifacts (discontinuities between legs, missing dress parts) due to the strong reliance on skeletal deformations. In contrast, our method correctly recovers the clothing dynamics and generates more detailed and complete 3D human surfaces. Note also that ReLoo produces more detailed facial features. + +fine-grained details (e.g., the faces shown in the third row and the T-shirts shown in the last row of Fig. 2). In contrast, ReLoo generates complete and plausible 3D human shapes with considerably more details (e.g., clothing wrinkles that fully align with image observations). ReLoo also clearly outperforms the baselines for the surface reconstruction under unseen views (see white background columns in Fig. 2). We attribute this superiority to our proposed neural layered clothed human representation and novel virtual bone deformation module. + +# 4.3 Novel View Synthesis Comparisons + +We compare with the same baselines for the task of novel view synthesis on MonoLoose and DynaCap [14]. We choose an unseen camera from the respective dataset as a novel view for all methods. As shown in Tab. 2, our method outperforms all baseline methods w.r.t. all metrics with an especially large margin on MonoLoose. Fig. 3 shows that SDF-based methods SelfRecon [18] and Vid2Avatar [13] strongly rely on skeletal deformation and cannot accurately recover the correct 3D shapes, leading to similar artifacts in novel view rendering as in surface reconstruction. Although the NeRF-based method SCARF [8] can + +![](images/f648580d6d622431039a1455a60b5a524abb6efbc9debb02064a20d72dd53d72.jpg) +Fig. 3: Qualitative novel view synthesis comparison. Our method achieves better rendering quality with detailed texture recovery in e.g., garment patterns and faces. Baseline methods can only produce corrupted and blurry rendering results (dress discontinuities between legs and unsharp texture details). + +fit to the coarse shape provided by the contour in image observations, it only produces blurry rendering results. ReLoo, in contrast, produces more plausible and realistic renderings while preserving sharper and fine-grained texture details. + +# 4.4 Qualitative Comparisons with Template-based Method + +Since there is no open-sourced template-based monocular method, we conduct a qualitative comparison with DeepCap [15], which is a learning-based method that predicts the template transformation given image observations. Our method is based on implicit neural fields which are topologically flexible and are not limited to a fixed resolution. As shown in Fig. 4, our method better recovers both the human surface details (e.g., faces and wrinkles) and large non-rigid clothing deformations. More importantly, our representation allows topological changes (empty space between dresses shown in the top row of Fig. 4), while the template-based method is inherently bound to the pre-scanned template mesh. + +# 4.5 Ablation Study + +Multi-Layer Volume Rendering. To learn the layered human representation, we leverage multi-layer volume rendering that is SDF-based and includes an inverse CDF sampling process. This means we perform multi-round sampling, where we sample two layers individually and combine the samples through sorting. To investigate the effect of this sampling strategy, we compare it to one-round sampling whereby we join the implicit SDFs of the human body and clothing simply by computing the minimum of the joined function. Results in Tab. 1 + +Table 2: Quantitative evaluation on novel view synthesis. We report the quantitative results on test views. Our method consistently outperforms other baseline methods on both datasets and all quantitative evaluation metrics, showing more realistic and plausible rendering quality (cf. Fig. 3). + +
MonoLooseDynaCap [14]
MethodPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
SelfRecon [18]22.50.9536.0826.80.9821.56
Vid2Avatar [13]25.90.9684.6627.10.9831.82
SCARF [8]23.30.9536.5925.50.9792.55
Ours w/o Virtual Bone28.70.9693.8127.30.9821.52
Ours29.20.9703.1527.90.9851.27
+ +![](images/2abd5b760743873e31c673fcac5a9763e0bcd2ef24f74b0a88a9c257c9daf23b.jpg) +Image +Fig.4: Qualitative comparisons with template-based method. + +![](images/f51ac1736881fe74d2120c31dc96656cb85fbb864b97faad0a363f8c2dab800d.jpg) +DeepCap + +![](images/8fc2e5cb3d04658501f4e017d025edf512105f029dd3d992b963cb327d0eec8b.jpg) +ReLoo (Ours) + +Compared to the template-based method, our representation and learning schemes enable more detailed and realistic human surface reconstruction and topological flexibility. + +![](images/b0643f7b82465dc7952631e6a090cf99c145092804f4c7e410ee6c77e6894308.jpg) +Image +One-Round Sampl. Multi-Round Sampl. +Fig. 5: Importance of multi-round sampling. One-round sampling strategy can lead to physically implausible clothed human reconstructions with severe garment-body interpenetration while multi-round sampling achieves better holistic reconstructions. + +indicate that multi-round sampling helps to improve the holistic clothed human reconstruction quality and avoids garment-body interpenetration (cf. Fig. 5). + +Virtual Bone Deformation Module. An important hyperparameter in our framework is the number of virtual bones $n_v$ for a garment. We quantitatively analyze the effects by learning the clothed human model with different numbers of virtual bones, i.e. $n_v \in \{20, 40, 80, 160, 320\}$ . We choose a subset of MonoLoose to evaluate novel view synthesis based on the perceptual similarity metric LPIPS [57] and time per training iteration. The quantitative results are reported in Fig. 6. We observe that the time cost linearly increases, while the error decreases with the sharpest drop at $n_v = 80$ bones. We thus select this point as the best performance-efficiency compromise. + +To validate the effectiveness of our virtual bone deformation module, we compare our full model to a version that uses SMPL-based skeletal deformation for + +![](images/8ef5ee1459db7b6e5b7ca0ed3c3203e69348a8d934ac3d7ed70c51ccd55ea577.jpg) +Fig.6: Number of virtual bones. LPIPS consistently decreases with the increasing number of virtual bones. We choose 80 virtual bones for the garment layer to balance the method performance and efficiency. + +![](images/52dacf5b7c202a28a8a5648981c6d0d14a11a1a239a0c42e24be072b38c5ad39.jpg) +Fig.7: Importance of virtual bone. Without the virtual bone deformation, our method is bounded by the expressiveness of skeletal movement and cannot accurately capture the topology and motion of loose garments. + +the garment layer instead of the virtual bone deformation module. We conducted both quantitative and qualitative analyses as shown in Tab. 2 and Fig. 7. The results demonstrate that integrating the virtual bone deformation module helps to find correct correspondences for points that are not solely controlled by skeletal deformations, leading to plausible and complete novel view synthesis results, while SMPL-based deformation is limited to the hierarchical skeleton structure and struggles with recovering garments that are far away from the inner body. + +# 5 Conclusion + +We present ReLoo, a novel method that produces temporally consistent 3D reconstructions of humans when dressed in highly dynamic loose garments from monocular in-the-wild videos. Our method does not require any 3D supervision or prior knowledge about the garments. We utilize a carefully designed layered neural implicit human representation to achieve a disentangled reconstruction of the body and the garment. We introduce a non-hierarchical virtual bone deformation module that enables the accurate capture of non-rigidly deforming loose outfits under articulation. A global optimization is formulated to jointly optimize the shape, appearance, and deformations of both inner body and outer clothing from images via multi-layer volume rendering. Our method achieves robust and high-fidelity reconstruction of humans dressed in loose garments. + +Limitations: Although readily available, ReLoo relies on reasonable pose estimates and segmentation masks as inputs. Manual adjustment is occasionally required to obtain SAM masks with sharp boundaries. Our method is mainly deployed to up to two garments. The complexity of ReLoo increases linearly with the number of garments that we aim to reconstruct separately. We discuss more limitations and societal impact in the Supp. Mat. + +# Acknowledgements + +This work was partially supported by the Swiss SERI Consolidation Grant "AIPERCEIVE". Chen Guo was partially supported by Microsoft Research Swiss JRC Grant. We are grateful to all our participants for their valued contribution to this research. Computations were carried out in part on the ETH Euler cluster. + +# References + +1. Alldieck, T., Magnor, M., Xu, W., Theobalt, C., Pons-Moll, G.: Video based reconstruction of 3d people models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8387-8397 (2018) +2. Alldieck, T., Zanfir, M., Sminchisescu, C.: Photorealistic monocular 3d reconstruction of humans wearing clothing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +3. Bertiche, H., Madadi, M., Escalera, S.: Pbns: Physically based neural simulation for unsupervised garment pose space deformation. ACM Trans. Graph. 40(6) (dec 2021) +4. Bhatnagar, B.L., Tiwari, G., Theobalt, C., Pons-Moll, G.: Multi-garment net: Learning to dress 3d people from images. In: IEEE International Conference on Computer Vision (ICCV). IEEE (oct 2019) +5. Chen, X., Pang, A., Yang, W., Wang, P., Xu, L., Yu, J.: Tightcap: 3d human shape capture with clothing tightness field. ACM Transactions on Graphics (TOG) 41(1), 1-17 (2021) +6. Collet, A., Chuang, M., Sweeney, P., Gillett, D., Evseev, D., Calabrese, D., Hoppe, H., Kirk, A., Sullivan, S.: High-quality streamable free-viewpoint video. ACM Trans. Graph. 34(4) (jul 2015). https://doi.org/10.1145/2766945, https://doi.org/10.1145/2766945 +7. Corona, E., Pumarola, A., Alenyà, G., Pons-Moll, G., Moreno-Noguer, F.: SMPlicit: Topology-aware generative model for clothed people. In: CVPR (2021) +8. Feng, Y., Yang, J., Pollefeys, M., Black, M.J., Bolkart, T.: Capturing and animation of body and clothing from monocular video. In: SIGGRAPH Asia 2022 Conference Papers. SA '22 (2022) +9. Geman, S., McClure, D.E.: Statistical methods for tomographic image reconstruction (1987) +10. Grigorev, A., Thomaszewski, B., Black, M.J., Hilliges, O.: HOOD: Hierarchical graphs for generalized modelling of clothing dynamics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023) +11. Gropp, A., Yariv, L., Haim, N., Atzmon, M., Lipman, Y.: Implicit geometric regularization for learning shapes. In: Proceedings of Machine Learning and Systems. pp. 3569-3579 (2020) +12. Guo, C., Chen, X., Song, J., Hilliges, O.: Human performance capture from monocular video in the wild. In: 2021 International Conference on 3D Vision (3DV). pp. 889-898. IEEE (2021) +13. Guo, C., Jiang, T., Chen, X., Song, J., Hilliges, O.: Vid2 avatar: 3d avatar reconstruction from videos in the wild via self-supervised scene decomposition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023) + +14. Habermann, M., Liu, L., Xu, W., Zollhoefer, M., Pons-Moll, G., Theobalt, C.: Real-time deep dynamic characters. ACM Transactions on Graphics 40(4) (aug 2021) +15. Habermann, M., Xu, W., Zollhoefer, M., Pons-Moll, G., Theobalt, C.: Deepcap: Monocular human performance capture using weak supervision. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (jun 2020) +16. He, T., Xu, Y., Saito, S., Soatto, S., Tung, T.: Arch++: Animation-ready clothed human reconstruction revisited. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 11046-11056 (October 2021) +17. Huang, Z., Xu, Y., Lassner, C., Li, H., Tung, T.: Arch: Animatable reconstruction of clothed humans. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3093-3102 (2020) +18. Jiang, B., Hong, Y., Bao, H., Zhang, J.: Selfrecon: Self reconstruction your digital avatar from monocular video. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +19. Jiang, B., Zhang, J., Hong, Y., Luo, J., Liu, L., Bao, H.: Bcnet: Learning body and cloth shape from a single image. In: European Conference on Computer Vision. Springer (2020) +20. Jiang, W., Yi, K.M., Samei, G., Tuzel, O., Ranjan, A.: Neuman: Neural human radiance field from a single video. In: Proceedings of the European conference on computer vision (ECCV) (2022) +21. Jiang, Z., Guo, C., Kaufmann, M., Jiang, T., Valentin, J., Hilliges, O., Song, J.: Multiply: Reconstruction of multiple people from monocular video in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2024) +22. Ke, L., Ye, M., Danelljan, M., Liu, Y., Tai, Y.W., Tang, C.K., Yu, F.: Segment anything in high quality. In: NeurIPS (2023) +23. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4015-4026 (October 2023) +24. Le, B.H., Deng, Z.: Smooth skinning decomposition with rigid bones. ACM Trans. Graph. 31(6) (nov 2012). https://doi.org/10.1145/2366145.2366218, https://doi.org/10.1145/2366145.2366218 +25. Li, Y., Habermann, M., Thomaszewski, B., Coros, S., Beeler, T., Theobalt, C.: Deep physics-aware inference of cloth deformation for monocular human performance capture. In: 2021 International Conference on 3D Vision (3DV). pp. 373-384. IEEE (2021) +26. Lin, W., Zheng, C., Yong, J.H., Xu, F.: Relightable and animatable neural avatars from videos. In: Proceedings of the AAAI Conference on Artificial Intelligence (2024) +27. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ACM transactions on graphics (TOG) 34(6), 1-16 (2015) +28. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4460-4470 (2019) +29. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: European conference on computer vision. pp. 405-421. Springer (2020) + +30. Moon, G., Nam, H., Shiratori, T., Lee, K.M.: 3d clothed human reconstruction in the wild. In: European Conference on Computer Vision (ECCV) (2022) +31. Pan, X., Mai, J., Jiang, X., Tang, D., Li, J., Shao, T., Zhou, K., Jin, X., Manocha, D.: Predicting loose-fitting garment deformations using bone-driven motion networks. In: ACM SIGGRAPH 2022 Conference Proceedings. SIGGRAPH '22, Association for Computing Machinery, New York, NY, USA (2022) +32. Patel, C., Liao, Z., Pons-Moll, G.: Tailornet: Predicting clothing in 3d as a function of human pose, shape and garment style. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (jun 2020) +33. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3D hands, face, and body from a single image. In: Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR). pp. 10975-10985 (2019) +34. Peng, S., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Zhou, X., Bao, H.: Animatable neural radiance fields for modeling dynamic human bodies. In: ICCV (2021) +35. Peng, S., Zhang, Y., Xu, Y., Wang, Q., Shuai, Q., Bao, H., Zhou, X.: Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9054-9063 (2021) +36. Pons-Moll, G., Pujades, S., Hu, S., Black, M.: Clothcap: Seamless 4d clothing capture and retargeting. ACM Transactions on Graphics, (Proc. SIGGRAPH) 36(4) (2017), two first authors contributed equally +37. Qiu, L., Chen, G., Zhou, J., Xu, M., Wang, J., Han, X.: Rec-mv: Reconstructing 3d dynamic cloth from monocular videos. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +38. Ricci, A.: A constructive geometry for computer graphics. Comput. J. 16, 157-160 (1973), https://api_semanticscholar.org/CorpusID:30038820 +39. Saito, S., Simon, T., Saragih, J., Joo, H.: Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 84-93 (2020) +40. Santesteban, I., Otaduy, M.A., Casas, D.: Snug: Self-supervised neural dynamic garments. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8140-8150 (June 2022) +41. Su, S.Y., Bagautdinov, T., Rhodin, H.: Danbo: Disentangled articulated neural body representations via graph neural networks. In: European Conference on Computer Vision (2022) +42. Su, Z., Wan, W., Yu, T., Liu, L., Fang, L., Wang, W., Liu, Y.: Mulaycap: Multi-layer human performance capture using a monocular video camera. IEEE Transactions on Visualization and Computer Graphics 28(4), 1862-1879 (2022). https://doi.org/10.1109/TVCG.2020.3027763 +43. Tiwari, G., Bhatnagar, B.L., Tung, T., Pons-Moll, G.: Sizer: A dataset and model for parsing 3d clothing and learning size sensitive 3d clothing. In: European Conference on Computer Vision (ECCV). Springer (August 2020) +44. Wang, K., Zhang, G., Cong, S., Yang, J.: Clothed human performance capture with a double-layer neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21098-21107 (June 2023) +45. Wang, S., Schwarz, K., Geiger, A., Tang, S.: Arah: Animatable volume rendering of articulated human sdfs. In: European Conference on Computer Vision (ECCV) (2022) + +46. Wang, Z., Simoncelli, E.P., Bovik, A.C.: Multiscale structural similarity for image quality assessment. In: The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers. vol. 2 (2003) +47. Weng, C.Y., Curless, B., Srinivasan, P.P., Barron, J.T., Kemelmacher-Shlizerman, I.: HumanNeRF: Free-viewpoint rendering of moving people from monocular video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16210-16220 (June 2022) +48. Xiang, D., Prada, F., Bagautdinov, T., Xu, W., Dong, Y., Wen, H., Hodgins, J., Wu, C.: Modeling clothing as a separate layer for an animatable human avatar. ACM Trans. Graph. 40(6) (dec 2021) +49. Xiu, Y., Yang, J., Cao, X., Tzionas, D., Black, M.J.: ECON: Explicit Clothed humans Optimized via Normal integration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023) +50. Xiu, Y., Yang, J., Tzionas, D., Black, M.J.: ICON: Implicit Clothed humans Obtained from Normals. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13296-13306 (June 2022) +51. Xu, W., Chatterjee, A., Zollhöfer, M., Rhodin, H., Mehta, D., Seidel, H.P., Theobalt, C.: Monoperfcap: Human performance capture from monocular video. SIGGRAPH 37(2), 27:1-27:15 (May 2018) +52. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. In: Advances in Neural Information Processing Systems (2021) +53. Yariv, L., Kasten, Y., Moran, D., Galun, M., Atzmon, M., Ronen, B., Lipman, Y.: Multiview neural surface reconstruction by disentangling geometry and appearance. In: Advances in Neural Information Processing Systems (2020) +54. Yu, T., Zheng, Z., Zhong, Y., Zhao, J., Dai, Q., Pons-Moll, G., Liu, Y.: Simulcap: Single-view human performance capture with cloth simulation. In: The IEEE International Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (June 2019) +55. Zhang, J., Liu, X., Ye, X., Zhao, F., Zhang, Y., Wu, M., Zhang, Y., Xu, L., Yu, J.: Editable free-viewpoint video using a layered neural representation. ACM Transactions on Graphics (TOG) 40(4), 1-18 (2021) +56. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv:2010.07492 (2020) +57. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018) +58. Zhang, Z., Sun, L., Yang, Z., Chen, L., Yang, Y.: Global-correlated 3d-decoupling transformer for clothed avatar reconstruction. In: Advances in Neural Information Processing Systems (NeurIPS) (2023) +59. Zheng, Z., Huang, H., Yu, T., Zhang, H., Guo, Y., Liu, Y.: Structured local radiance fields for human avatar modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2022) +60. Zheng, Z., Yu, T., Liu, Y., Dai, Q.: Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021) \ No newline at end of file diff --git a/2024/ReLoo_ Reconstructing Humans Dressed in Loose Garments from Monocular Video in the Wild/images.zip b/2024/ReLoo_ Reconstructing Humans Dressed in Loose Garments from Monocular Video in the Wild/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..24646e4428c390cc21df3b77e0ddcb0efcfa4c91 --- /dev/null +++ b/2024/ReLoo_ Reconstructing Humans Dressed in Loose Garments from Monocular Video in the Wild/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2cb7d954659b12b51441da909d68fb9b5112ac9655a8956662264266c2ae4bbe +size 469175 diff --git a/2024/ReLoo_ Reconstructing Humans Dressed in Loose Garments from Monocular Video in the Wild/layout.json b/2024/ReLoo_ Reconstructing Humans Dressed in Loose Garments from Monocular Video in the Wild/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3435383e723a5105ee92415fe682c7a5ae42138f --- /dev/null +++ b/2024/ReLoo_ Reconstructing Humans Dressed in Loose Garments from Monocular Video in the Wild/layout.json @@ -0,0 +1,9605 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 112, + 479, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 112, + 479, + 146 + ], + "spans": [ + { + "bbox": [ + 134, + 112, + 479, + 146 + ], + "type": "text", + "content": "ReLoo: Reconstructing Humans Dressed in Loose Garments from Monocular Video in the Wild" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 151, + 167, + 462, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 167, + 462, + 194 + ], + "spans": [ + { + "bbox": [ + 151, + 167, + 462, + 194 + ], + "type": "text", + "content": "Chen Guo\\*, Tianjian Jiang\\*, Manuel Kaufmann, Chengwei Zheng, Julien Valentin2, Jie Song†, and Otmar Hilliges" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 276, + 201, + 337, + 223 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 276, + 201, + 337, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 201, + 337, + 211 + ], + "spans": [ + { + "bbox": [ + 276, + 201, + 337, + 211 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 276, + 201, + 337, + 211 + ], + "type": "text", + "content": " ETH Zürich" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 282, + 213, + 331, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 282, + 213, + 331, + 223 + ], + "spans": [ + { + "bbox": [ + 282, + 213, + 331, + 223 + ], + "type": "text", + "content": "2 Microsoft" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 137, + 247, + 477, + 359 + ], + "blocks": [ + { + "bbox": [ + 137, + 247, + 477, + 359 + ], + "lines": [ + { + "bbox": [ + 137, + 247, + 477, + 359 + ], + "spans": [ + { + "bbox": [ + 137, + 247, + 477, + 359 + ], + "type": "image", + "image_path": "7410daf7c917e3bb4895c227b68fde001bf35658a108c48bb4a453a724c8d7a8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 408, + 455, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 408, + 455, + 617 + ], + "spans": [ + { + "bbox": [ + 159, + 408, + 455, + 617 + ], + "type": "text", + "content": "Abstract. While previous years have seen great progress in the 3D reconstruction of humans from monocular videos, few of the state-of-the-art methods are able to handle loose garments that exhibit large non-rigid surface deformations during articulation. This limits the application of such methods to humans that are dressed in standard pants or T-shirts. Our method, ReLoo, overcomes this limitation and reconstructs high-quality 3D models of humans dressed in loose garments from monocular in-the-wild videos. To tackle this problem, we first establish a layered neural human representation that decomposes clothed humans into a neural inner body and outer clothing. On top of the layered neural representation, we further introduce a non-hierarchical virtual bone deformation module for the clothing layer that can freely move, which allows the accurate recovery of non-rigidly deforming loose clothing. A global optimization jointly optimizes the shape, appearance, and deformations of the human body and clothing via multi-layer differentiable volume rendering. To evaluate ReLoo, we record subjects with dynamically deforming garments in a multi-view capture studio. This evaluation, both on existing and our novel dataset, demonstrates ReLoo's clear superiority over prior art on both indoor datasets and in-the-wild videos." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 627, + 402, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 627, + 402, + 638 + ], + "spans": [ + { + "bbox": [ + 160, + 627, + 402, + 638 + ], + "type": "text", + "content": "Keywords: human representation " + }, + { + "bbox": [ + 160, + 627, + 402, + 638 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 627, + 402, + 638 + ], + "type": "text", + "content": " human reconstruction" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 642, + 335, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 642, + 335, + 653 + ], + "spans": [ + { + "bbox": [ + 135, + 642, + 335, + 653 + ], + "type": "text", + "content": "* These authors contributed equally to this work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 135, + 653, + 370, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 653, + 370, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 653, + 370, + 665 + ], + "type": "text", + "content": "† Corresponding author. Now at HKUST(GZ) & HKUST" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "type": "text", + "content": "As researchers aim to democratize the creation of realistic human avatars, the reconstruction of 3D clothed humans from casually captured monocular videos has garnered increased attention. While many solutions have been proposed to do so in recent years [1,8,13,18,39,50], they primarily focus on capturing subjects with tight-fitting clothing and perform poorly when reconstructing loose garments whose dynamics are less tightly coupled with body pose. Such loose garments, however, constitute a significant part of a real-life wardrobe and thus failure to capture them is limiting the creation of realistic human avatars from monocular footage. In this paper, we present a method, ReLoo, that overcomes this shortcoming." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 259, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 461 + ], + "type": "text", + "content": "Reconstructing humans dressed in loose clothing requires accurate tracking of large, non-rigid deformations and recovering fine-grained details of freely flowing surfaces. Template-based methods [14, 15, 51] have been applied to do so but the acquisition of the template burdens the deployment to unseen subjects and its explicit representation limits its expressive capability to capture dynamically changing surface details. More recently, methods based on neural implicit functions have emerged as a promising remedy for the disadvantages of template-based methods [13, 16, 18, 20, 35, 39, 41, 47, 49, 58-60]. Yet, these methods currently fail to provide convincing reconstructions of humans in loose garments. We observe that few of these methods differentiate between the human body and clothing but model the clothed human as a single entity. This limits the expressiveness and capacity of the underlying model to capture more local features. More importantly, this formulation only allows to drive the off-body garments with skeletal deformations that are derived from the underlying parametric body model (e.g., SMPL [27]). Thus, they are inherently incapable of handling highly dynamic loose garments that are topologically different from the inner body and do not correlate strongly with the bone movement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 463, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 482, + 641 + ], + "type": "text", + "content": "In this paper, we adopt the promising neural implicit shape modeling paradigm, but we argue that a single implicit representation fundamentally limits the representation power and hinders the capability to model complex garment topology that exhibits free-form deformations. To properly model loose garment dynamics that only weakly correlate with skeletal deformations - while still retaining the ability to deform human bodies with skeleton-driven motions - ReLoo takes neural implicit human models to the next level. To do so, our approach is grounded in the following core concepts: i) We establish a layered neural human representation that decomposes clothed humans into the neural inner body and outer clothing. ii) Based on this layered neural representation, we further propose a non-hierarchical virtual bone deformation module for the clothing layer that allows free movement and accurate recovery of highly dynamic loose outfits. iii) In a global optimization, we jointly optimize the shape, appearance, and deformations of both the human body and clothing layer over the entire sequence via multi-layer differentiable volume rendering." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "In our experiments, we demonstrate that our framework leads to temporally consistent and high-quality reconstructions of clothed humans dressed in loose" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 275, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 275, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 275, + 102 + ], + "type": "text", + "content": "C. Guo and T. Jiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "garments. We also ablate our method to uncover the contribution of its essential components. Furthermore, we conduct comparisons with existing approaches in human surface reconstruction and novel view synthesis, showing that our method outperforms prior art from both the template-based and neural implicit modeling domains. To highlight differences to the prior art, we capture a new dataset, MonoLoose, which puts an emphasis on humans dressed in loose clothing under dynamic motions and contains ground-truth reconstructions captured with a high-end multi-view volumetric recording studio (MVS)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 213, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 247 + ], + "type": "text", + "content": "In summary, in this paper we introduce ReLoo, a method that improves clothed human reconstruction quality and accurately captures human performance dressed in highly dynamic loose garments. Our key contributions are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 248, + 480, + 331 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 248, + 480, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 248, + 480, + 272 + ], + "spans": [ + { + "bbox": [ + 138, + 248, + 480, + 272 + ], + "type": "text", + "content": "- a novel layered neural human representation, disentangling the inner body and outer clothing; and" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 272, + 480, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 272, + 480, + 295 + ], + "spans": [ + { + "bbox": [ + 138, + 272, + 480, + 295 + ], + "type": "text", + "content": "- a virtual bone deformation module that is built on top of the layered neural human representation and accurately tracks the large surface dynamics; and" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 297, + 480, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 297, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 297, + 480, + 331 + ], + "type": "text", + "content": "- a robust framework that leverages multi-layer differentiable volume rendering achieving high-fidelity 3D human reconstructions from monocular in-the-wild videos of humans dressed in highly dynamic loose garments." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 351, + 237, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 351, + 237, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 351, + 237, + 364 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 378, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 668 + ], + "type": "text", + "content": "Single-Layer Human Reconstruction from Monocular Input Template-based monocular human performance capture methods track the pre-defined clothed human template to fit to 2D observations [14,15,51]. They demonstrate robust tracking of human performance even when dressed in loose garments. However, they struggle to generalize to in-the-wild settings due to the reliance on a rigged, personalized, pre-scanned template obtained from a dense capture setup. Follow-up works endeavor to remove this dependency by adding per-vertex offsets on top of the SMPL body [1,12]. Nevertheless, the explicit mesh representation is held back by a fixed resolution and topology, hampering the representation of fine-grained details. Other works have shown compelling results with learning-based methods that learn to regress 3D human geometry and appearance from images [2,16,17,39,49,50,58,60]. A major limitation of these methods is the necessity of high-quality 3D data for supervision. They also often fail to produce space-time coherent reconstructions over frames. Recent works employ neural rendering to fit neural fields to videos to obtain an articulated human model [8,13,18,20,21,26,34,35,37,45,47]. E.g., SelfRecon [18] deploys neural surface rendering [53] to achieve consistent reconstruction over the sequence and Vid2Avatar [13] leverages differentiable volume rendering [52] to eliminate the need for pre-masking thus producing robust 3D human reconstruction. However, all aforementioned methods treat the body and clothing as a single entity, limiting the model's expressiveness and resulting in low-quality reconstructions for loose clothing. In contrast, our method is based on a layered neural human representation that, in conjunction with our novel virtual bone deformation module, enables tracking of highly dynamic loose clothing." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "ReLoo" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 295 + ], + "type": "text", + "content": "Multi-Layer Human Representation and Reconstruction Several methods exist that investigate how a clothing layer deforms given 3D human motion. They use either physically-simulated training data [3,25,31,32,42,44] or directly deploy physics-informed objectives [10, 40]. These methods have shown compelling results in modeling large deformations of loose outfits. However, they lack generalization to in-the-wild settings and diverse clothing categories due to a reliance on input templates [25, 44] and assume high-quality 3D motion data is already available, not learned from video like in our setting. Our multi-layer representation and deformation modeling are inspired by this line of work, specifically Pan et al. [31], who also make use of virtual bones. Different from [31], our method jointly learns the clothing shape and deformations from only 2D observations, whereas [31] require known clothing templates and 3D simulated data to achieve animation of clothes. Furthermore, [31] define and fix virtual bones using skinning decomposition [24], while in our framework the virtual bone placement evolves naturally with training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 296, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 475 + ], + "type": "text", + "content": "Methods that reconstruct layered clothed humans from videos or images alone are presented in [5, 7, 8, 19, 30, 36, 43, 48, 54]. Some of them either require active multi-view setups [4, 5, 36, 48] or depth information [54] preventing them from being deployed in the wild. BCNet [19] learns to predict clothing geometry draped over the SMPL body from a single image. However, it is limited to pre-defined clothing style templates. SMPLicit [7] and ClothWild [30] extend such learning-based methods to generalize to more general clothing types, but they tend to produce over-smooth results lacking details such as wrinkles. Our closest related work, SCARF [8] is built upon SMPL-X [33] and reconstructs the outer clothing layer using NeRF [29], achieving better reconstruction quality than previous works. Our method differs from SCARF [8] in three regards: 1) ReLoo is a fully implicit representation that is expressive enough to capture detailed body (including faces) and clothing shape jointly, 2) it is not limited to self-rotating motions, and 3) it supports the capture of large non-rigid surface deformations of loose garments thanks to a novel virtual bone deformation module." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 475, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 571 + ], + "type": "text", + "content": "In summary, existing monocular-based methods tend to overly rely on parametric body models [27,33] as a human body proxy and thus struggle to model garments whose deformations cannot be easily correlated with the inner body pose under dynamic motion (as is the case for t-shirts or pants that are commonly used as example apparel). Moreover, they can only capture less detailed personalized shape characteristics such as faces. ReLoo overcomes these limitations and when compared to existing single-layer, multi-layer, and template-based methods produces higher fidelity results across the board." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 590, + 202, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 202, + 604 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 202, + 604 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "We introduce ReLoo, a novel method for detailed geometry and appearance reconstruction of a human performer in highly dynamic, loose clothing from monocular in-the-wild videos. The overview of our method is schematically given in Fig. 1. To enable a temporally consistent, expressive human representation" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 275, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 275, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 275, + 102 + ], + "type": "text", + "content": "C. Guo and T. Jiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 479, + 277 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 479, + 277 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 479, + 277 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 479, + 277 + ], + "type": "image", + "image_path": "20d9f366b461bc6059bc72f9c69b672c6b6109c3631e1124bc429f53c21825dd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 286, + 482, + 397 + ], + "lines": [ + { + "bbox": [ + 130, + 286, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 286, + 482, + 397 + ], + "type": "text", + "content": "Fig. 1: Method Overview. Given an image from a video sequence, we sample points along the camera ray for each neural layer. We warp sampled points for the body layer " + }, + { + "bbox": [ + 130, + 286, + 482, + 397 + ], + "type": "inline_equation", + "content": "\\pmb{x}_d^B" + }, + { + "bbox": [ + 130, + 286, + 482, + 397 + ], + "type": "text", + "content": " into canonical space via inverse LBS derived from skeletal deformations. We deform sampled points for the garment layer " + }, + { + "bbox": [ + 130, + 286, + 482, + 397 + ], + "type": "inline_equation", + "content": "\\pmb{x}_d^G" + }, + { + "bbox": [ + 130, + 286, + 482, + 397 + ], + "type": "text", + "content": " into canonical space via inverse warping based on the proposed virtual bone deformation module (Sec. 3.2). We then evaluate the respective implicit network to obtain the SDF and radiance values (Sec. 3.1). We apply multi-layer differentiable volume rendering to learn the shape, appearance, and deformations of the layered neural human representation from images (Sec. 3.3). The loss function " + }, + { + "bbox": [ + 130, + 286, + 482, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 130, + 286, + 482, + 397 + ], + "type": "text", + "content": " compares the rendered color predictions with image observations as well as a segmentation mask obtained using SAM [22] (Sec. 3.4)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 422, + 482, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 531 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 531 + ], + "type": "text", + "content": "we establish a layered neural implicit representation for the body (inner layer, naked) and garment (outer layer, template-free) as described in Sec. 3.1. On top, we introduce a hybrid deformation strategy which consists of skeletal deformation for the body and a virtual-bone-driven deformation module for the outer loose garment (Sec. 3.2). This allows us to capture dynamic loose garments. We learn the layered human representation and the deformation module jointly by performing multi-layer differentiable volume rendering (Sec. 3.3). The whole process is trained globally to optimize jointly for shape, appearance, and deformations of the inner body and outer garment layer (Sec. 3.4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 555, + 358, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 358, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 358, + 567 + ], + "type": "text", + "content": "3.1 Layered Neural Human Representation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "We represent the 3D shape of the clothed human with implicit signed-distance fields (SDF) and the appearance with texture fields in a temporally consistent canonical space. The inner body and outer garment are modeled separately. More specifically, we model the geometry and appearance of the body in canonical space with a neural network " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "f^{B}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " which predicts the signed distance value " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "s^{B}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " and radiance value " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "c^{B}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " for any 3D point " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{c}^{B}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " in this space. We similarly model the garment with a neural network " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "f^{G}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " that takes points " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{c}^{G}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " in the garment's" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "ReLoo" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 243, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 243, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 243, + 128 + ], + "type": "text", + "content": "canonical space as input:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 216, + 137, + 480, + 150 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 137, + 480, + 150 + ], + "spans": [ + { + "bbox": [ + 216, + 137, + 480, + 150 + ], + "type": "interline_equation", + "content": "\\boldsymbol {c} ^ {B}, s ^ {B} = f ^ {B} \\left(\\boldsymbol {x} _ {c} ^ {B}, \\boldsymbol {\\theta}\\right); \\boldsymbol {c} ^ {G}, s ^ {G} = f ^ {G} \\left(\\boldsymbol {x} _ {c} ^ {G}, \\boldsymbol {\\theta}\\right), \\tag {1}", + "image_path": "603d83ba7c3912fc5026006b60fa8a02d62606c1f36a511d83257d7ec0b7b45b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 157, + 481, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 157, + 481, + 254 + ], + "spans": [ + { + "bbox": [ + 131, + 157, + 481, + 254 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 157, + 481, + 254 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 131, + 157, + 481, + 254 + ], + "type": "text", + "content": " denote the SMPL pose parameters [27], which we concatenate to " + }, + { + "bbox": [ + 131, + 157, + 481, + 254 + ], + "type": "inline_equation", + "content": "\\pmb{x}_c^B" + }, + { + "bbox": [ + 131, + 157, + 481, + 254 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 157, + 481, + 254 + ], + "type": "inline_equation", + "content": "\\pmb{x}_c^G" + }, + { + "bbox": [ + 131, + 157, + 481, + 254 + ], + "type": "text", + "content": " to model pose-dependent effects such as facial features and clothing wrinkles. For clothing that is not a single, dress-like garment we use a separate network for the upper and lower garment. For simplicity and without loss of generality, we only discuss a single piece of clothing in the following. Note that if a canonical point is within any of these two surfaces (" + }, + { + "bbox": [ + 131, + 157, + 481, + 254 + ], + "type": "inline_equation", + "content": "s^B < 0" + }, + { + "bbox": [ + 131, + 157, + 481, + 254 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 131, + 157, + 481, + 254 + ], + "type": "inline_equation", + "content": "s^G < 0" + }, + { + "bbox": [ + 131, + 157, + 481, + 254 + ], + "type": "text", + "content": "), it is also within the entire clothed human shape. Thus we can obtain the final clothed human shape by compositing these two neural fields and taking their minimum [38]:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 266, + 261, + 480, + 275 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 261, + 480, + 275 + ], + "spans": [ + { + "bbox": [ + 266, + 261, + 480, + 275 + ], + "type": "interline_equation", + "content": "s ^ {H} = \\min \\left\\{s ^ {B}, s ^ {G} \\right\\}. \\tag {2}", + "image_path": "31b3f45a2f4b8c3d144037a7961c88b53f8635de680e8337f814c7a01674a9e3.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 291, + 315, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 291, + 315, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 291, + 315, + 304 + ], + "type": "text", + "content": "3.2 Hybrid Deformation Modeling" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 310, + 481, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 310, + 481, + 393 + ], + "spans": [ + { + "bbox": [ + 131, + 310, + 481, + 393 + ], + "type": "text", + "content": "To find point correspondences between the deformed space in the observed image and our pre-defined canonical space, we devise a hybrid deformation module that treats the body and clothing layers according to their respective levels of rigidity. The human body predominantly depends on skeletal deformation, but the deformation of the loose garments cannot solely be explained by skeletal motion alone. Therefore we propose to drive clothing deformation by a set of additional virtual bones whose transformations are directly learned from video." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 394, + 481, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 394, + 481, + 465 + ], + "spans": [ + { + "bbox": [ + 131, + 394, + 481, + 465 + ], + "type": "text", + "content": "Skeletal Deformation. We follow a standard skeletal deformation based on SMPL to find correspondences in canonical and deformed space for the inner body [13]. Specifically, given the bone transformation matrices " + }, + { + "bbox": [ + 131, + 394, + 481, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_i" + }, + { + "bbox": [ + 131, + 394, + 481, + 465 + ], + "type": "text", + "content": " for joints " + }, + { + "bbox": [ + 131, + 394, + 481, + 465 + ], + "type": "inline_equation", + "content": "i\\in \\{1,\\dots,n_b\\}" + }, + { + "bbox": [ + 131, + 394, + 481, + 465 + ], + "type": "text", + "content": " which are derived from the body pose parameters " + }, + { + "bbox": [ + 131, + 394, + 481, + 465 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 131, + 394, + 481, + 465 + ], + "type": "text", + "content": ", a canonical point " + }, + { + "bbox": [ + 131, + 394, + 481, + 465 + ], + "type": "inline_equation", + "content": "\\pmb{x}_c^B" + }, + { + "bbox": [ + 131, + 394, + 481, + 465 + ], + "type": "text", + "content": " is mapped to the corresponding deformed space point " + }, + { + "bbox": [ + 131, + 394, + 481, + 465 + ], + "type": "inline_equation", + "content": "\\pmb{x}_d^B" + }, + { + "bbox": [ + 131, + 394, + 481, + 465 + ], + "type": "text", + "content": " via linear blend skinning (LBS):" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 264, + 466, + 480, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 466, + 480, + 496 + ], + "spans": [ + { + "bbox": [ + 264, + 466, + 480, + 496 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {d} ^ {B} = \\sum_ {i = 1} ^ {n _ {s}} w _ {c} ^ {i} \\boldsymbol {B} _ {i} \\boldsymbol {x} _ {c} ^ {B}. \\tag {3}", + "image_path": "e7137a88f0cd80a5ad255a8327936be72899762a783a6fb0dea81f2e3d6c0e4d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 501, + 481, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 501, + 481, + 523 + ], + "spans": [ + { + "bbox": [ + 131, + 501, + 481, + 523 + ], + "type": "text", + "content": "Conversely, given a point in deformed space " + }, + { + "bbox": [ + 131, + 501, + 481, + 523 + ], + "type": "inline_equation", + "content": "\\pmb{x}_d^B" + }, + { + "bbox": [ + 131, + 501, + 481, + 523 + ], + "type": "text", + "content": ", its canonical correspondence can be solved via:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 254, + 523, + 480, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 523, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 254, + 523, + 480, + 553 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {c} ^ {B} = \\left(\\sum_ {i = 1} ^ {n _ {s}} w _ {d} ^ {i} \\boldsymbol {B} _ {i}\\right) ^ {- 1} \\boldsymbol {x} _ {d} ^ {B}. \\tag {4}", + "image_path": "2de4bba1d9902ccc696a8984bf7b1c18a2bcb0de97ce8336581da02ed1321d3f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 556, + 481, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 556, + 481, + 617 + ], + "spans": [ + { + "bbox": [ + 131, + 556, + 481, + 617 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 131, + 556, + 481, + 617 + ], + "type": "inline_equation", + "content": "n_s" + }, + { + "bbox": [ + 131, + 556, + 481, + 617 + ], + "type": "text", + "content": " denotes the number of skeletal bones in the transformation, and " + }, + { + "bbox": [ + 131, + 556, + 481, + 617 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_{(\\cdot)} = \\{w_{(\\cdot)}^1, \\dots, w_{(\\cdot)}^{n_s}\\}" + }, + { + "bbox": [ + 131, + 556, + 481, + 617 + ], + "type": "text", + "content": " represents the skinning weights for " + }, + { + "bbox": [ + 131, + 556, + 481, + 617 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{(\\cdot)}^B" + }, + { + "bbox": [ + 131, + 556, + 481, + 617 + ], + "type": "text", + "content": ". We assign " + }, + { + "bbox": [ + 131, + 556, + 481, + 617 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_d" + }, + { + "bbox": [ + 131, + 556, + 481, + 617 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 131, + 556, + 481, + 617 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_d^B" + }, + { + "bbox": [ + 131, + 556, + 481, + 617 + ], + "type": "text", + "content": " based on the average of neighboring SMPL vertices' skinning weights, weighted by the point-to-point distances in deformed space. The treatment of canonical points " + }, + { + "bbox": [ + 131, + 556, + 481, + 617 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_c^B" + }, + { + "bbox": [ + 131, + 556, + 481, + 617 + ], + "type": "text", + "content": " follows a similar approach." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "content": "Virtual Bone Deformation. The virtual bones correspond to a set of bones " + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{\\pmb{v}_i\\}_{i=1}^{n_v}" + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "content": " defined in the canonical space that drives the neighboring 3D garment points " + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{x}_c^G" + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "content": " using rigid transformations. Different from the skeletons of characters rigged by artists such as SMPL [27], the virtual bones are non-hierarchical" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 275, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 275, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 275, + 102 + ], + "type": "text", + "content": "C. Guo and T. Jiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "and are not restricted to rotate in relation to their parents following an anatomical structure. Thus, they can be transformed freely and are applicable to capture the deformations of highly dynamic loose garments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 249 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 249 + ], + "type": "text", + "content": "Given a set of virtual bones " + }, + { + "bbox": [ + 130, + 152, + 482, + 249 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 152, + 482, + 249 + ], + "type": "text", + "content": ", their transformations consist of the rotations and translations relative to the SMPL root " + }, + { + "bbox": [ + 130, + 152, + 482, + 249 + ], + "type": "inline_equation", + "content": "\\pmb{\\mathcal{T}}_i = [R_i|T_i]" + }, + { + "bbox": [ + 130, + 152, + 482, + 249 + ], + "type": "text", + "content": " which are predicted by a deformation field " + }, + { + "bbox": [ + 130, + 152, + 482, + 249 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^G" + }, + { + "bbox": [ + 130, + 152, + 482, + 249 + ], + "type": "text", + "content": ", parameterized via an MLP. " + }, + { + "bbox": [ + 130, + 152, + 482, + 249 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^G" + }, + { + "bbox": [ + 130, + 152, + 482, + 249 + ], + "type": "text", + "content": " takes the concatenation of the 3D positions " + }, + { + "bbox": [ + 130, + 152, + 482, + 249 + ], + "type": "inline_equation", + "content": "\\pmb{v}_i" + }, + { + "bbox": [ + 130, + 152, + 482, + 249 + ], + "type": "text", + "content": " of the virtual bones, the human body pose " + }, + { + "bbox": [ + 130, + 152, + 482, + 249 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 130, + 152, + 482, + 249 + ], + "type": "text", + "content": " and a continuous time embedding " + }, + { + "bbox": [ + 130, + 152, + 482, + 249 + ], + "type": "inline_equation", + "content": "\\pmb{t}" + }, + { + "bbox": [ + 130, + 152, + 482, + 249 + ], + "type": "text", + "content": " as input, and outputs the axis angles " + }, + { + "bbox": [ + 130, + 152, + 482, + 249 + ], + "type": "inline_equation", + "content": "\\pmb{A}_i" + }, + { + "bbox": [ + 130, + 152, + 482, + 249 + ], + "type": "text", + "content": " and translations " + }, + { + "bbox": [ + 130, + 152, + 482, + 249 + ], + "type": "inline_equation", + "content": "\\pmb{T}_i" + }, + { + "bbox": [ + 130, + 152, + 482, + 249 + ], + "type": "text", + "content": ". The continuous time embedding " + }, + { + "bbox": [ + 130, + 152, + 482, + 249 + ], + "type": "inline_equation", + "content": "\\pmb{t}" + }, + { + "bbox": [ + 130, + 152, + 482, + 249 + ], + "type": "text", + "content": " is included to aid in learning temporal dynamics from videos. The rotations " + }, + { + "bbox": [ + 130, + 152, + 482, + 249 + ], + "type": "inline_equation", + "content": "\\pmb{R}_i" + }, + { + "bbox": [ + 130, + 152, + 482, + 249 + ], + "type": "text", + "content": " are further obtained from " + }, + { + "bbox": [ + 130, + 152, + 482, + 249 + ], + "type": "inline_equation", + "content": "\\pmb{A}_i" + }, + { + "bbox": [ + 130, + 152, + 482, + 249 + ], + "type": "text", + "content": " via the Rodrigues' formula " + }, + { + "bbox": [ + 130, + 152, + 482, + 249 + ], + "type": "inline_equation", + "content": "f^{Rod}(\\cdot)" + }, + { + "bbox": [ + 130, + 152, + 482, + 249 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 233, + 255, + 481, + 270 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 255, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 233, + 255, + 481, + 270 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mathcal {T}} _ {i} = \\left[ f ^ {R o d} (\\boldsymbol {A} _ {i}) | \\boldsymbol {T} _ {i} \\right] = \\mathcal {D} ^ {G} (\\boldsymbol {v} _ {i}, \\boldsymbol {\\theta}, t). \\tag {5}", + "image_path": "1e31e9d49238954cec9040d29218176b94b1739ef524af96cf923bb34f49156d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 276, + 479, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 276, + 479, + 301 + ], + "spans": [ + { + "bbox": [ + 131, + 276, + 479, + 301 + ], + "type": "text", + "content": "To drive a 3D garment point from canonical space " + }, + { + "bbox": [ + 131, + 276, + 479, + 301 + ], + "type": "inline_equation", + "content": "\\pmb{x}_c^G" + }, + { + "bbox": [ + 131, + 276, + 479, + 301 + ], + "type": "text", + "content": " to deformed space " + }, + { + "bbox": [ + 131, + 276, + 479, + 301 + ], + "type": "inline_equation", + "content": "\\pmb{x}_d^G" + }, + { + "bbox": [ + 131, + 276, + 479, + 301 + ], + "type": "text", + "content": " and vice-versa, we use the neighboring virtual bones' motions and LBS as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 206, + 309, + 481, + 339 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 309, + 481, + 339 + ], + "spans": [ + { + "bbox": [ + 206, + 309, + 481, + 339 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {d} ^ {G} = \\sum_ {i = 1} ^ {n _ {v}} \\delta_ {c} ^ {i} \\boldsymbol {\\mathcal {T}} _ {i} \\boldsymbol {x} _ {c} ^ {G}, \\quad \\boldsymbol {x} _ {c} ^ {G} = \\left(\\sum_ {i = 1} ^ {n _ {v}} \\delta_ {d} ^ {i} \\boldsymbol {\\mathcal {T}} _ {i}\\right) ^ {- 1} \\boldsymbol {x} _ {d} ^ {G} \\tag {6}", + "image_path": "1de4fe0f9b44ba4c219d1a6e4463e41e1c18b17721c494f6f470c10c8eaaf494.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 346, + 479, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 479, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 479, + 411 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 346, + 479, + 411 + ], + "type": "inline_equation", + "content": "n_v" + }, + { + "bbox": [ + 130, + 346, + 479, + 411 + ], + "type": "text", + "content": " denotes the number of virtual bones for deforming loose garments, which is a hyperparameter, and " + }, + { + "bbox": [ + 130, + 346, + 479, + 411 + ], + "type": "inline_equation", + "content": "\\pmb{\\delta}_{(\\cdot)} = \\{\\delta_{(\\cdot)}^1,\\dots,\\delta_{(\\cdot)}^{n_v}\\}" + }, + { + "bbox": [ + 130, + 346, + 479, + 411 + ], + "type": "text", + "content": " represent the skimming weights of " + }, + { + "bbox": [ + 130, + 346, + 479, + 411 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{(\\cdot)}^G" + }, + { + "bbox": [ + 130, + 346, + 479, + 411 + ], + "type": "text", + "content": " w.r.t. each virtual bone " + }, + { + "bbox": [ + 130, + 346, + 479, + 411 + ], + "type": "inline_equation", + "content": "\\pmb{v}_i\\in \\mathcal{V}" + }, + { + "bbox": [ + 130, + 346, + 479, + 411 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 346, + 479, + 411 + ], + "type": "inline_equation", + "content": "\\pmb{\\delta}_{(\\cdot)}" + }, + { + "bbox": [ + 130, + 346, + 479, + 411 + ], + "type": "text", + "content": " is calculated based on the inverse of the distance between " + }, + { + "bbox": [ + 130, + 346, + 479, + 411 + ], + "type": "inline_equation", + "content": "\\pmb{x}_c^G" + }, + { + "bbox": [ + 130, + 346, + 479, + 411 + ], + "type": "text", + "content": " and each " + }, + { + "bbox": [ + 130, + 346, + 479, + 411 + ], + "type": "inline_equation", + "content": "\\pmb{v}_i" + }, + { + "bbox": [ + 130, + 346, + 479, + 411 + ], + "type": "text", + "content": " whereby far away bones are clamped to 0. More details are shown in the Supp. Mat." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 411, + 479, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 479, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 479, + 460 + ], + "type": "text", + "content": "Unlike [31], where a garment template is available and the virtual bones are extracted using 3D simulation data, we learn the clothed human model from monocular observations solely and extract the virtual bones on the fly without requiring any specific template prior. We explain the acquisition more in Sec. 3.4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 475, + 318, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 475, + 318, + 488 + ], + "spans": [ + { + "bbox": [ + 131, + 475, + 318, + 488 + ], + "type": "text", + "content": "3.3 Multi-Layer Volume Rendering" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 495, + 479, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 479, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 479, + 567 + ], + "type": "text", + "content": "As we are aiming to jointly reconstruct multiple layers of neural implicit fields, we are required to depart from standard differentiable volume rendering for static scenes (e.g., [29]). We thus introduce multi-layer volume rendering tailored to our multi-layer human representation (Sec. 3.1) to reconstruct both inner body and outer clothing under garment-body occlusions. To do so, we use surface-based volume rendering [52] while re-ordering multiple neural layers [55]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 567, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 567, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 131, + 567, + 482, + 668 + ], + "type": "text", + "content": "Specifically, we shoot a ray " + }, + { + "bbox": [ + 131, + 567, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\pmb{r}" + }, + { + "bbox": [ + 131, + 567, + 482, + 668 + ], + "type": "text", + "content": " through every pixel of the image and sample two sets of points in the body layer and the garment layer: " + }, + { + "bbox": [ + 131, + 567, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\{\\pmb{x}_{d,1}^{B},\\dots,\\pmb{x}_{d,N}^{B}\\}" + }, + { + "bbox": [ + 131, + 567, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 567, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\{\\pmb{x}_{d,1}^{G},\\dots,\\pmb{x}_{d,N}^{G}\\}" + }, + { + "bbox": [ + 131, + 567, + 482, + 668 + ], + "type": "text", + "content": " following the two-stage sampling strategy proposed in [52]. Note that both sets contain the same amount of points. Next, we use the skeletal deformation to warp each sampled point " + }, + { + "bbox": [ + 131, + 567, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{d,i}^{B}" + }, + { + "bbox": [ + 131, + 567, + 482, + 668 + ], + "type": "text", + "content": " to the body's canonical space " + }, + { + "bbox": [ + 131, + 567, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{c,i}^{B}" + }, + { + "bbox": [ + 131, + 567, + 482, + 668 + ], + "type": "text", + "content": " and we use our virtual bone deformation module to find canonical correspondences " + }, + { + "bbox": [ + 131, + 567, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{c,i}^{G}" + }, + { + "bbox": [ + 131, + 567, + 482, + 668 + ], + "type": "text", + "content": " to each sampled point " + }, + { + "bbox": [ + 131, + 567, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{d,i}^{G}" + }, + { + "bbox": [ + 131, + 567, + 482, + 668 + ], + "type": "text", + "content": " (Sec. 3.2). Then, we obtain the corresponding signed-distance " + }, + { + "bbox": [ + 131, + 567, + 482, + 668 + ], + "type": "inline_equation", + "content": "(s_i^B" + }, + { + "bbox": [ + 131, + 567, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 567, + 482, + 668 + ], + "type": "inline_equation", + "content": "s_i^G)" + }, + { + "bbox": [ + 131, + 567, + 482, + 668 + ], + "type": "text", + "content": " and radiance values " + }, + { + "bbox": [ + 131, + 567, + 482, + 668 + ], + "type": "inline_equation", + "content": "(\\pmb{c}_i^B" + }, + { + "bbox": [ + 131, + 567, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 567, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\pmb{c}_i^G)" + }, + { + "bbox": [ + 131, + 567, + 482, + 668 + ], + "type": "text", + "content": " by querying the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "ReLoo" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "implicit shape and texture fields with the canonical points. We then compute the occupancy for the inner and outer layer at the " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "-th sampled point as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 210, + 150, + 481, + 164 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 150, + 481, + 164 + ], + "spans": [ + { + "bbox": [ + 210, + 150, + 481, + 164 + ], + "type": "interline_equation", + "content": "o _ {i} ^ {B} = \\left(1 - \\exp \\left(- \\sigma_ {i} ^ {B} \\Delta \\boldsymbol {x} _ {i} ^ {B}\\right)\\right); \\sigma_ {i} ^ {B} = \\sigma \\left(s _ {i} ^ {B}, \\boldsymbol {\\theta}\\right) \\tag {7}", + "image_path": "202aa565014678353138c7f56ad87f94f9e5b5adb70b0d49a3ff808083118a33.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 211, + 167, + 481, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 167, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 211, + 167, + 481, + 182 + ], + "type": "interline_equation", + "content": "o _ {i} ^ {G} = \\left(1 - \\exp \\left(- \\sigma_ {i} ^ {G} \\Delta \\boldsymbol {x} _ {i} ^ {G}\\right)\\right); \\sigma_ {i} ^ {G} = \\sigma \\left(s _ {i} ^ {G}, \\boldsymbol {\\theta}\\right) \\tag {8}", + "image_path": "d6f4827d17eed292d8d3e639a4a009ab2ea796c6535b0ede7adef0a58344724b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 191, + 479, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 191, + 479, + 253 + ], + "spans": [ + { + "bbox": [ + 131, + 191, + 479, + 253 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 191, + 479, + 253 + ], + "type": "inline_equation", + "content": "\\Delta \\pmb{x}_i^{(\\cdot)}" + }, + { + "bbox": [ + 131, + 191, + 479, + 253 + ], + "type": "text", + "content": " is the distance between two adjacent sample points in the respective layer, and " + }, + { + "bbox": [ + 131, + 191, + 479, + 253 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 131, + 191, + 479, + 253 + ], + "type": "text", + "content": " represents the scaled Laplace's Cumulative Distribution Function (CDF) to convert " + }, + { + "bbox": [ + 131, + 191, + 479, + 253 + ], + "type": "inline_equation", + "content": "s_i^B" + }, + { + "bbox": [ + 131, + 191, + 479, + 253 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 191, + 479, + 253 + ], + "type": "inline_equation", + "content": "s_i^G" + }, + { + "bbox": [ + 131, + 191, + 479, + 253 + ], + "type": "text", + "content": " to volume densities " + }, + { + "bbox": [ + 131, + 191, + 479, + 253 + ], + "type": "inline_equation", + "content": "(\\sigma_{i}^{B}" + }, + { + "bbox": [ + 131, + 191, + 479, + 253 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 191, + 479, + 253 + ], + "type": "inline_equation", + "content": "\\sigma_{i}^{G})" + }, + { + "bbox": [ + 131, + 191, + 479, + 253 + ], + "type": "text", + "content": " following [52]. Finally, we integrate the radiance numerically for both layers and obtain the neurally rendered color of the human performer " + }, + { + "bbox": [ + 131, + 191, + 479, + 253 + ], + "type": "inline_equation", + "content": "\\hat{C}^H" + }, + { + "bbox": [ + 131, + 191, + 479, + 253 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 200, + 263, + 481, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 263, + 481, + 301 + ], + "spans": [ + { + "bbox": [ + 200, + 263, + 481, + 301 + ], + "type": "interline_equation", + "content": "\\hat {C} ^ {H} = \\sum_ {i = 1} ^ {N} \\sum_ {p \\in \\{B, G \\}} \\left[ o _ {i} ^ {p} \\boldsymbol {c} _ {i} ^ {p} \\prod_ {q \\in \\{B, G \\}} \\prod_ {j \\in \\mathcal {I} _ {i} ^ {q, p}} \\left(1 - o _ {j} ^ {q}\\right) \\right], \\tag {9}", + "image_path": "586de7779f4067303d75b06ad438868b0affd11b2f607576382094b97a8fc17d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 310, + 482, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 310, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 131, + 310, + 482, + 346 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 310, + 482, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_i^{q,p} = \\{j\\in [1,N]\\mid z(\\pmb{x}_{d,j}^q) < z(\\pmb{x}_{d,i}^p)\\}" + }, + { + "bbox": [ + 131, + 310, + 482, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 310, + 482, + 346 + ], + "type": "inline_equation", + "content": "z(\\cdot)" + }, + { + "bbox": [ + 131, + 310, + 482, + 346 + ], + "type": "text", + "content": " measures the depth of a point w.r.t. the camera origin. In other words, we sort all points according to their depth values from near to far and then conduct the volumetric integration." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 347, + 481, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 347, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 131, + 347, + 481, + 396 + ], + "type": "text", + "content": "Scene Composition. To model the background of the scene we use NeRF++ [56], denoted as " + }, + { + "bbox": [ + 131, + 347, + 481, + 396 + ], + "type": "inline_equation", + "content": "f^S" + }, + { + "bbox": [ + 131, + 347, + 481, + 396 + ], + "type": "text", + "content": ", which estimates a color value " + }, + { + "bbox": [ + 131, + 347, + 481, + 396 + ], + "type": "inline_equation", + "content": "\\hat{C}^S" + }, + { + "bbox": [ + 131, + 347, + 481, + 396 + ], + "type": "text", + "content": " representing the scene's color. The final pixel color " + }, + { + "bbox": [ + 131, + 347, + 481, + 396 + ], + "type": "inline_equation", + "content": "\\hat{C}" + }, + { + "bbox": [ + 131, + 347, + 481, + 396 + ], + "type": "text", + "content": " is a composite of " + }, + { + "bbox": [ + 131, + 347, + 481, + 396 + ], + "type": "inline_equation", + "content": "\\hat{C}^S" + }, + { + "bbox": [ + 131, + 347, + 481, + 396 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 131, + 347, + 481, + 396 + ], + "type": "inline_equation", + "content": "\\hat{C}^H" + }, + { + "bbox": [ + 131, + 347, + 481, + 396 + ], + "type": "text", + "content": " following [13]. More details are shown in the Supp. Mat." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 415, + 264, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 415, + 264, + 427 + ], + "spans": [ + { + "bbox": [ + 131, + 415, + 264, + 427 + ], + "type": "text", + "content": "3.4 Global Optimization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 437, + 479, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 479, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 479, + 485 + ], + "type": "text", + "content": "To jointly learn the inner body and outer clothing of clothed humans from monocular videos in a template-free manner, we propose a two-stage training schema. The whole training process is formulated as a global optimization over all estimizable parameters and the entire video sequence." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "Two-Stage Training Schema. In the first stage, we leverage skeletal deformation to deform both the body and garment layer. Meanwhile, we warm up the virtual bone deformation field " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^G" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " by encouraging " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^G" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " to have similar deformations as the SMPL model around near-body regions. In the second stage, we activate the virtual bone deformation module to drive the garment layer. To obtain virtual bones, we generate garment meshes using Multiresolution IsoSurface Extraction (MISE) [28]. In theory, each of the " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " vertices of the resulting canonical garment mesh can be made a virtual bone. This will however incur a high computational cost in the deformation module. To mitigate this but still retain expressive capability we employ a quadric mesh simplification algorithm to reduce the number of vertices to " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "n_v \\ll M" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ". The remaining vertices after the simplification are the initial virtual bone locations. We empirically found 80 virtual bones to deliver the best performance-efficiency compromise (see Sec. 4.5). Note that we periodically generate new sets of virtual bones during training to account for changing garment topologies." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 276, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 276, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 276, + 102 + ], + "type": "text", + "content": "C. Guo and T. Jiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": "Reconstruction Loss. For every ray " + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\boldsymbol{r} \\in \\mathcal{R}" + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": " we compute how well the rendered color " + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\hat{C}(\\boldsymbol{r})" + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": " matches the image pixel's RGB value " + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "C(\\boldsymbol{r})" + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": " with the " + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": "-distance:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 241, + 150, + 481, + 177 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 150, + 481, + 177 + ], + "spans": [ + { + "bbox": [ + 241, + 150, + 481, + 177 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {r g b}} = \\frac {1}{| \\mathcal {R} |} \\sum_ {\\boldsymbol {r} \\in \\mathcal {R}} | C (\\boldsymbol {r}) - \\hat {C} (\\boldsymbol {r}) |. \\tag {10}", + "image_path": "e7d09cbbeda245ba9de2b728cd5e1c1ba94ea76c84caeef7f757f3f170c7c5e8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 187, + 479, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 187, + 479, + 213 + ], + "spans": [ + { + "bbox": [ + 131, + 187, + 479, + 213 + ], + "type": "text", + "content": "Segmentation Loss. We modify Eq. (9) to render the opacity " + }, + { + "bbox": [ + 131, + 187, + 479, + 213 + ], + "type": "inline_equation", + "content": "\\hat{O}^B (\\boldsymbol {r})" + }, + { + "bbox": [ + 131, + 187, + 479, + 213 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 187, + 479, + 213 + ], + "type": "inline_equation", + "content": "\\hat{O}^G (\\boldsymbol {r})" + }, + { + "bbox": [ + 131, + 187, + 479, + 213 + ], + "type": "text", + "content": " per pixel for both layers:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 222, + 485, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 222, + 485, + 268 + ], + "spans": [ + { + "bbox": [ + 134, + 222, + 485, + 268 + ], + "type": "interline_equation", + "content": "\\hat {O} ^ {B} (\\boldsymbol {r}) = \\sum_ {i = 1} ^ {N} \\left[ o _ {i} ^ {B} \\prod_ {q \\in \\{B, G \\}} \\prod_ {j \\in \\mathcal {I} _ {i} ^ {q, B}} \\left(1 - o _ {j} ^ {q}\\right) \\right]; \\hat {O} ^ {G} (\\boldsymbol {r}) = \\sum_ {i = 1} ^ {N} \\left[ o _ {i} ^ {G} \\prod_ {q \\in \\{B, G \\}} \\prod_ {j \\in \\mathcal {I} _ {i} ^ {q, G}} \\left(1 - o _ {j} ^ {q}\\right) \\right]. \\tag {11}", + "image_path": "7160ee3edd65657d40280526fe2a4c297bf9654529540703a7ecf47ba4fdf816.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 270, + 479, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 270, + 479, + 316 + ], + "spans": [ + { + "bbox": [ + 131, + 270, + 479, + 316 + ], + "type": "text", + "content": "The segmentation loss is calculated between the rendered pixel-wise opacity and the segmentation masks extracted using SAM [23]. A robust Geman-McClure error function " + }, + { + "bbox": [ + 131, + 270, + 479, + 316 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 131, + 270, + 479, + 316 + ], + "type": "text", + "content": " [9] is applied to down-weigh potentially erroneous cloth segmentation mask predictions (more details are explained in Supp. Mat):" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 209, + 327, + 479, + 355 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 327, + 479, + 355 + ], + "spans": [ + { + "bbox": [ + 209, + 327, + 479, + 355 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s e g}} = \\frac {1}{| \\mathcal {R} |} \\sum_ {\\boldsymbol {r} \\in \\mathcal {R}} \\sum_ {p \\in \\{B, G \\}} \\rho \\left(\\mathcal {M} _ {\\text {s a m}} ^ {p} (\\boldsymbol {r}) - \\hat {O} ^ {p} (\\boldsymbol {r})\\right). \\tag {12}", + "image_path": "a74ad07a7e2af1ef6657571bae189b64c35283a6000f622f710829267f8c18da.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 365, + 479, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 365, + 479, + 424 + ], + "spans": [ + { + "bbox": [ + 131, + 365, + 479, + 424 + ], + "type": "text", + "content": "Adaptive Eikonal Loss. We follow IGR [11] and sample points in the canonical space to compute the Eikonal constraint to regularize the validity of our SDFs in each layer. Unlike [11], which randomly samples points in the entire space, we periodically extract the canonical shapes for both layers and sample points around the explicit mesh surfaces:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 195, + 434, + 479, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 434, + 479, + 449 + ], + "spans": [ + { + "bbox": [ + 195, + 434, + 479, + 449 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {e i k o n a l}} = \\mathbb {E} _ {\\boldsymbol {x} _ {c} ^ {B}} \\left(\\left\\| \\nabla s ^ {B} \\right\\| - 1\\right) ^ {2} + \\mathbb {E} _ {\\boldsymbol {x} _ {c} ^ {G}} \\left(\\left\\| \\nabla s ^ {G} \\right\\| - 1\\right) ^ {2}. \\tag {13}", + "image_path": "5a013a89af959031aa49860227bd7ebfafdec6eacc868665eafd38b011ba0aac.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 458, + 479, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 458, + 479, + 529 + ], + "spans": [ + { + "bbox": [ + 131, + 458, + 479, + 529 + ], + "type": "text", + "content": "Virtual Bone Deformation Regularization. To accelerate the convergence of the virtual bone deformation field in the first stage of the training process, we randomly sample 3D canonical garment points " + }, + { + "bbox": [ + 131, + 458, + 479, + 529 + ], + "type": "inline_equation", + "content": "\\pmb{x}_c^G" + }, + { + "bbox": [ + 131, + 458, + 479, + 529 + ], + "type": "text", + "content": " and apply additional SMPL transformation regularization which ensures that the virtual bone deformation " + }, + { + "bbox": [ + 131, + 458, + 479, + 529 + ], + "type": "inline_equation", + "content": "LBS^{G}" + }, + { + "bbox": [ + 131, + 458, + 479, + 529 + ], + "type": "text", + "content": " do not deviate excessively from the transformations made by skeletal deformation " + }, + { + "bbox": [ + 131, + 458, + 479, + 529 + ], + "type": "inline_equation", + "content": "LBS^{B}" + }, + { + "bbox": [ + 131, + 458, + 479, + 529 + ], + "type": "text", + "content": " during the warm-up stage:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 229, + 539, + 479, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 539, + 479, + 553 + ], + "spans": [ + { + "bbox": [ + 229, + 539, + 479, + 553 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e g}} = \\left\\| L B S ^ {B} \\left(\\boldsymbol {x} _ {c} ^ {G}\\right) - L B S ^ {G} \\left(\\boldsymbol {x} _ {c} ^ {G}\\right) \\right\\| ^ {2} \\tag {14}", + "image_path": "36ebaa466201da5944181683879c4c219f9eec5a884d793b81596d29fbd5867f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 562, + 358, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 562, + 358, + 573 + ], + "spans": [ + { + "bbox": [ + 131, + 562, + 358, + 573 + ], + "type": "text", + "content": "See Supp. Mat for more details about the final loss." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 593, + 229, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 593, + 229, + 605 + ], + "spans": [ + { + "bbox": [ + 132, + 593, + 229, + 605 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 617, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 479, + 665 + ], + "type": "text", + "content": "We first introduce the datasets and metrics used for evaluation. Then, we compare our proposed method with state-of-the-art approaches in two tasks: 3D surface reconstruction and novel view synthesis. Ablation studies are then conducted to demonstrate the effectiveness of our core components and design choices." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "ReLoo" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 186, + 157, + 427, + 228 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 1: Quantitative evaluation on surface reconstruction. We compute the 3D surface metrics on the MonoLoose dataset. Our method consistently outperforms all baselines on all evaluation metrics (cf. Fig. 2)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 186, + 157, + 427, + 228 + ], + "lines": [ + { + "bbox": [ + 186, + 157, + 427, + 228 + ], + "spans": [ + { + "bbox": [ + 186, + 157, + 427, + 228 + ], + "type": "table", + "html": "
MethodC-ℓ2↓NC↑V-IoU↑
SelfRecon [18]2.220.7880.844
Vid2Avatar [13]2.340.7940.776
SCARF [8]3.130.7110.691
Ours w/o Multi-Round Sampl.2.340.7700.879
Ours1.930.8310.881
", + "image_path": "fde61ee00a6471b6575d0a63eef463b3bad94958ed9ac0f78665e389c4a8b259.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 250, + 205, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 250, + 205, + 261 + ], + "spans": [ + { + "bbox": [ + 132, + 250, + 205, + 261 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 269, + 482, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 482, + 365 + ], + "type": "text", + "content": "MonoLoose Dataset: Due to the lack of datasets that capture dynamic human performance with high-fidelity 3D ground-truth meshes when dressed in loose garments, we captured our own dataset, MonoLoose, with a high-end multiview volumetric capture studio (MVS) [6]. This dataset is specifically curated for evaluating monocular human surface reconstruction and novel view synthesis methods, with a particular focus on subjects dressed in loose attire. It consists of five sequences with different identities, loose garment styles, and motions. For more details on the contents of MonoLoose, please refer to the Supp. Mat." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 365, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 413 + ], + "type": "text", + "content": "DynaCap [14]: We further evaluate our method on DynaCap, which captures dynamic human performance with a dense multi-view system. We curate two sequences that feature loose garments for novel view synthesis evaluation. Note that DynaCap does not provide dense scans for reconstruction comparison." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 413, + 481, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 481, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 481, + 437 + ], + "type": "text", + "content": "In-the-wild videos: We use in-the-wild videos collected from DeepCap [15] and online videos to demonstrate the robustness and generalization of our method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 437, + 481, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 481, + 474 + ], + "type": "text", + "content": "Evaluation Protocol: We report Chamfer distance " + }, + { + "bbox": [ + 130, + 437, + 481, + 474 + ], + "type": "inline_equation", + "content": "(\\mathbf{C} - \\ell_2)" + }, + { + "bbox": [ + 130, + 437, + 481, + 474 + ], + "type": "text", + "content": " [cm], normal consistency (NC), and volumetric IoU for surface reconstruction comparison. Novel view synthesis quality is measured via PSNR, SSIM [46], and LPIPS " + }, + { + "bbox": [ + 130, + 437, + 481, + 474 + ], + "type": "inline_equation", + "content": "(\\times 100)" + }, + { + "bbox": [ + 130, + 437, + 481, + 474 + ], + "type": "text", + "content": " [57]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 490, + 347, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 347, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 347, + 502 + ], + "type": "text", + "content": "4.2 Surface Reconstruction Comparisons" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "We compare our proposed human surface reconstruction method to several state-of-the-art approaches [8, 13, 18] on our MonoLoose dataset. SelfRecon [18] and Vid2Avatar [13] deploy neural rendering to reconstruct the 3D clothed human using a single layer. SCARF [8] reconstructs a hybrid human model based on an explicit inner body and NeRF-based clothing model. All baseline methods rely on SMPL skeleton skinning transformation with additional deformation fields for the garments' motion. Our method outperforms all baselines by a substantial margin on all evaluation metrics (cf. Tab. 1). This disparity becomes more visible in qualitative comparisons shown in Fig. 2. When the dynamic loose garments necessitate large non-skeletal surface deformations, all baseline methods fail to recover complete human surfaces or produce implausible and corrupted reconstructions with visible artifacts (e.g., discontinuities between legs and missing dress parts, see highlights in Fig. 2). Furthermore, they tend to produce less" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 276, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 276, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 276, + 102 + ], + "type": "text", + "content": "C. Guo and T. Jiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 114, + 473, + 342 + ], + "blocks": [ + { + "bbox": [ + 141, + 114, + 473, + 342 + ], + "lines": [ + { + "bbox": [ + 141, + 114, + 473, + 342 + ], + "spans": [ + { + "bbox": [ + 141, + 114, + 473, + 342 + ], + "type": "image", + "image_path": "806a18f21ad2247b3e03ed01d2cd1908271c82104db77f026bf279271a5a5ccd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 354, + 482, + 420 + ], + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 420 + ], + "type": "text", + "content": "Fig. 2: Qualitative 3D surface reconstruction comparison. Baseline methods produce less detailed and implausible 3D clothed human reconstructions with visible artifacts (discontinuities between legs, missing dress parts) due to the strong reliance on skeletal deformations. In contrast, our method correctly recovers the clothing dynamics and generates more detailed and complete 3D human surfaces. Note also that ReLoo produces more detailed facial features." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 445, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 529 + ], + "type": "text", + "content": "fine-grained details (e.g., the faces shown in the third row and the T-shirts shown in the last row of Fig. 2). In contrast, ReLoo generates complete and plausible 3D human shapes with considerably more details (e.g., clothing wrinkles that fully align with image observations). ReLoo also clearly outperforms the baselines for the surface reconstruction under unseen views (see white background columns in Fig. 2). We attribute this superiority to our proposed neural layered clothed human representation and novel virtual bone deformation module." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 548, + 338, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 338, + 561 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 338, + 561 + ], + "type": "text", + "content": "4.3 Novel View Synthesis Comparisons" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "We compare with the same baselines for the task of novel view synthesis on MonoLoose and DynaCap [14]. We choose an unseen camera from the respective dataset as a novel view for all methods. As shown in Tab. 2, our method outperforms all baseline methods w.r.t. all metrics with an especially large margin on MonoLoose. Fig. 3 shows that SDF-based methods SelfRecon [18] and Vid2Avatar [13] strongly rely on skeletal deformation and cannot accurately recover the correct 3D shapes, leading to similar artifacts in novel view rendering as in surface reconstruction. Although the NeRF-based method SCARF [8] can" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "ReLoo" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 161, + 125, + 455, + 277 + ], + "blocks": [ + { + "bbox": [ + 161, + 125, + 455, + 277 + ], + "lines": [ + { + "bbox": [ + 161, + 125, + 455, + 277 + ], + "spans": [ + { + "bbox": [ + 161, + 125, + 455, + 277 + ], + "type": "image", + "image_path": "f648580d6d622431039a1455a60b5a524abb6efbc9debb02064a20d72dd53d72.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 289, + 482, + 334 + ], + "lines": [ + { + "bbox": [ + 131, + 289, + 482, + 334 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 482, + 334 + ], + "type": "text", + "content": "Fig. 3: Qualitative novel view synthesis comparison. Our method achieves better rendering quality with detailed texture recovery in e.g., garment patterns and faces. Baseline methods can only produce corrupted and blurry rendering results (dress discontinuities between legs and unsharp texture details)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 357, + 482, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 357, + 482, + 394 + ], + "spans": [ + { + "bbox": [ + 131, + 357, + 482, + 394 + ], + "type": "text", + "content": "fit to the coarse shape provided by the contour in image observations, it only produces blurry rendering results. ReLoo, in contrast, produces more plausible and realistic renderings while preserving sharper and fine-grained texture details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 413, + 437, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 413, + 437, + 425 + ], + "spans": [ + { + "bbox": [ + 131, + 413, + 437, + 425 + ], + "type": "text", + "content": "4.4 Qualitative Comparisons with Template-based Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 434, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 542 + ], + "type": "text", + "content": "Since there is no open-sourced template-based monocular method, we conduct a qualitative comparison with DeepCap [15], which is a learning-based method that predicts the template transformation given image observations. Our method is based on implicit neural fields which are topologically flexible and are not limited to a fixed resolution. As shown in Fig. 4, our method better recovers both the human surface details (e.g., faces and wrinkles) and large non-rigid clothing deformations. More importantly, our representation allows topological changes (empty space between dresses shown in the top row of Fig. 4), while the template-based method is inherently bound to the pre-scanned template mesh." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 560, + 239, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 560, + 239, + 572 + ], + "spans": [ + { + "bbox": [ + 131, + 560, + 239, + 572 + ], + "type": "text", + "content": "4.5 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Multi-Layer Volume Rendering. To learn the layered human representation, we leverage multi-layer volume rendering that is SDF-based and includes an inverse CDF sampling process. This means we perform multi-round sampling, where we sample two layers individually and combine the samples through sorting. To investigate the effect of this sampling strategy, we compare it to one-round sampling whereby we join the implicit SDFs of the human body and clothing simply by computing the minimum of the joined function. Results in Tab. 1" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 276, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 276, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 276, + 102 + ], + "type": "text", + "content": "C. Guo and T. Jiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 141, + 168, + 473, + 264 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 160 + ], + "type": "text", + "content": "Table 2: Quantitative evaluation on novel view synthesis. We report the quantitative results on test views. Our method consistently outperforms other baseline methods on both datasets and all quantitative evaluation metrics, showing more realistic and plausible rendering quality (cf. Fig. 3)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 168, + 473, + 264 + ], + "lines": [ + { + "bbox": [ + 141, + 168, + 473, + 264 + ], + "spans": [ + { + "bbox": [ + 141, + 168, + 473, + 264 + ], + "type": "table", + "html": "
MonoLooseDynaCap [14]
MethodPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
SelfRecon [18]22.50.9536.0826.80.9821.56
Vid2Avatar [13]25.90.9684.6627.10.9831.82
SCARF [8]23.30.9536.5925.50.9792.55
Ours w/o Virtual Bone28.70.9693.8127.30.9821.52
Ours29.20.9703.1527.90.9851.27
", + "image_path": "9593c6056eb0ddcfa374b2054a2b40153c3e9117b53efc21da1541f390e22d69.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 280, + 185, + 380 + ], + "blocks": [ + { + "bbox": [ + 133, + 280, + 185, + 380 + ], + "lines": [ + { + "bbox": [ + 133, + 280, + 185, + 380 + ], + "spans": [ + { + "bbox": [ + 133, + 280, + 185, + 380 + ], + "type": "image", + "image_path": "2abd5b760743873e31c673fcac5a9763e0bcd2ef24f74b0a88a9c257c9daf23b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 380, + 169, + 388 + ], + "lines": [ + { + "bbox": [ + 151, + 380, + 169, + 388 + ], + "spans": [ + { + "bbox": [ + 151, + 380, + 169, + 388 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 401, + 291, + 422 + ], + "lines": [ + { + "bbox": [ + 131, + 401, + 291, + 422 + ], + "spans": [ + { + "bbox": [ + 131, + 401, + 291, + 422 + ], + "type": "text", + "content": "Fig.4: Qualitative comparisons with template-based method." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 186, + 280, + 237, + 380 + ], + "blocks": [ + { + "bbox": [ + 186, + 280, + 237, + 380 + ], + "lines": [ + { + "bbox": [ + 186, + 280, + 237, + 380 + ], + "spans": [ + { + "bbox": [ + 186, + 280, + 237, + 380 + ], + "type": "image", + "image_path": "f51ac1736881fe74d2120c31dc96656cb85fbb864b97faad0a363f8c2dab800d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 380, + 226, + 388 + ], + "lines": [ + { + "bbox": [ + 200, + 380, + 226, + 388 + ], + "spans": [ + { + "bbox": [ + 200, + 380, + 226, + 388 + ], + "type": "text", + "content": "DeepCap" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 238, + 280, + 290, + 380 + ], + "blocks": [ + { + "bbox": [ + 238, + 280, + 290, + 380 + ], + "lines": [ + { + "bbox": [ + 238, + 280, + 290, + 380 + ], + "spans": [ + { + "bbox": [ + 238, + 280, + 290, + 380 + ], + "type": "image", + "image_path": "8fc2e5cb3d04658501f4e017d025edf512105f029dd3d992b963cb327d0eec8b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 246, + 380, + 283, + 388 + ], + "lines": [ + { + "bbox": [ + 246, + 380, + 283, + 388 + ], + "spans": [ + { + "bbox": [ + 246, + 380, + 283, + 388 + ], + "type": "text", + "content": "ReLoo (Ours)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 423, + 291, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 423, + 291, + 477 + ], + "spans": [ + { + "bbox": [ + 131, + 423, + 291, + 477 + ], + "type": "text", + "content": "Compared to the template-based method, our representation and learning schemes enable more detailed and realistic human surface reconstruction and topological flexibility." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 328, + 278, + 473, + 382 + ], + "blocks": [ + { + "bbox": [ + 328, + 278, + 473, + 382 + ], + "lines": [ + { + "bbox": [ + 328, + 278, + 473, + 382 + ], + "spans": [ + { + "bbox": [ + 328, + 278, + 473, + 382 + ], + "type": "image", + "image_path": "b0643f7b82465dc7952631e6a090cf99c145092804f4c7e410ee6c77e6894308.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 382, + 361, + 388 + ], + "lines": [ + { + "bbox": [ + 345, + 382, + 361, + 388 + ], + "spans": [ + { + "bbox": [ + 345, + 382, + 361, + 388 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 378, + 382, + 472, + 388 + ], + "lines": [ + { + "bbox": [ + 378, + 382, + 472, + 388 + ], + "spans": [ + { + "bbox": [ + 378, + 382, + 472, + 388 + ], + "type": "text", + "content": "One-Round Sampl. Multi-Round Sampl." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 321, + 401, + 482, + 477 + ], + "lines": [ + { + "bbox": [ + 321, + 401, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 321, + 401, + 482, + 477 + ], + "type": "text", + "content": "Fig. 5: Importance of multi-round sampling. One-round sampling strategy can lead to physically implausible clothed human reconstructions with severe garment-body interpenetration while multi-round sampling achieves better holistic reconstructions." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 506, + 480, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 480, + 530 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 480, + 530 + ], + "type": "text", + "content": "indicate that multi-round sampling helps to improve the holistic clothed human reconstruction quality and avoids garment-body interpenetration (cf. Fig. 5)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 532, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 532, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 532, + 482, + 639 + ], + "type": "text", + "content": "Virtual Bone Deformation Module. An important hyperparameter in our framework is the number of virtual bones " + }, + { + "bbox": [ + 130, + 532, + 482, + 639 + ], + "type": "inline_equation", + "content": "n_v" + }, + { + "bbox": [ + 130, + 532, + 482, + 639 + ], + "type": "text", + "content": " for a garment. We quantitatively analyze the effects by learning the clothed human model with different numbers of virtual bones, i.e. " + }, + { + "bbox": [ + 130, + 532, + 482, + 639 + ], + "type": "inline_equation", + "content": "n_v \\in \\{20, 40, 80, 160, 320\\}" + }, + { + "bbox": [ + 130, + 532, + 482, + 639 + ], + "type": "text", + "content": ". We choose a subset of MonoLoose to evaluate novel view synthesis based on the perceptual similarity metric LPIPS [57] and time per training iteration. The quantitative results are reported in Fig. 6. We observe that the time cost linearly increases, while the error decreases with the sharpest drop at " + }, + { + "bbox": [ + 130, + 532, + 482, + 639 + ], + "type": "inline_equation", + "content": "n_v = 80" + }, + { + "bbox": [ + 130, + 532, + 482, + 639 + ], + "type": "text", + "content": " bones. We thus select this point as the best performance-efficiency compromise." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "To validate the effectiveness of our virtual bone deformation module, we compare our full model to a version that uses SMPL-based skeletal deformation for" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "ReLoo" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 115, + 297, + 200 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 297, + 200 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 297, + 200 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 297, + 200 + ], + "type": "image", + "image_path": "8ef5ee1459db7b6e5b7ca0ed3c3203e69348a8d934ac3d7ed70c51ccd55ea577.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 209, + 291, + 276 + ], + "lines": [ + { + "bbox": [ + 131, + 209, + 291, + 276 + ], + "spans": [ + { + "bbox": [ + 131, + 209, + 291, + 276 + ], + "type": "text", + "content": "Fig.6: Number of virtual bones. LPIPS consistently decreases with the increasing number of virtual bones. We choose 80 virtual bones for the garment layer to balance the method performance and efficiency." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 318, + 121, + 494, + 198 + ], + "blocks": [ + { + "bbox": [ + 318, + 121, + 494, + 198 + ], + "lines": [ + { + "bbox": [ + 318, + 121, + 494, + 198 + ], + "spans": [ + { + "bbox": [ + 318, + 121, + 494, + 198 + ], + "type": "image", + "image_path": "52dacf5b7c202a28a8a5648981c6d0d14a11a1a239a0c42e24be072b38c5ad39.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 209, + 482, + 276 + ], + "lines": [ + { + "bbox": [ + 315, + 209, + 482, + 276 + ], + "spans": [ + { + "bbox": [ + 315, + 209, + 482, + 276 + ], + "type": "text", + "content": "Fig.7: Importance of virtual bone. Without the virtual bone deformation, our method is bounded by the expressiveness of skeletal movement and cannot accurately capture the topology and motion of loose garments." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 304, + 482, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 389 + ], + "type": "text", + "content": "the garment layer instead of the virtual bone deformation module. We conducted both quantitative and qualitative analyses as shown in Tab. 2 and Fig. 7. The results demonstrate that integrating the virtual bone deformation module helps to find correct correspondences for points that are not solely controlled by skeletal deformations, leading to plausible and complete novel view synthesis results, while SMPL-based deformation is limited to the hierarchical skeleton structure and struggles with recovering garments that are far away from the inner body." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 420, + 220, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 220, + 433 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 220, + 433 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 459, + 482, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 482, + 590 + ], + "type": "text", + "content": "We present ReLoo, a novel method that produces temporally consistent 3D reconstructions of humans when dressed in highly dynamic loose garments from monocular in-the-wild videos. Our method does not require any 3D supervision or prior knowledge about the garments. We utilize a carefully designed layered neural implicit human representation to achieve a disentangled reconstruction of the body and the garment. We introduce a non-hierarchical virtual bone deformation module that enables the accurate capture of non-rigidly deforming loose outfits under articulation. A global optimization is formulated to jointly optimize the shape, appearance, and deformations of both inner body and outer clothing from images via multi-layer volume rendering. Our method achieves robust and high-fidelity reconstruction of humans dressed in loose garments." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": "Limitations: Although readily available, ReLoo relies on reasonable pose estimates and segmentation masks as inputs. Manual adjustment is occasionally required to obtain SAM masks with sharp boundaries. Our method is mainly deployed to up to two garments. The complexity of ReLoo increases linearly with the number of garments that we aim to reconstruct separately. We discuss more limitations and societal impact in the Supp. Mat." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 276, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 276, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 276, + 102 + ], + "type": "text", + "content": "C. Guo and T. Jiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 481, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 481, + 186 + ], + "type": "text", + "content": "This work was partially supported by the Swiss SERI Consolidation Grant \"AIPERCEIVE\". Chen Guo was partially supported by Microsoft Research Swiss JRC Grant. We are grateful to all our participants for their valued contribution to this research. Computations were carried out in part on the ETH Euler cluster." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 228, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 228, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 262 + ], + "type": "text", + "content": "1. Alldieck, T., Magnor, M., Xu, W., Theobalt, C., Pons-Moll, G.: Video based reconstruction of 3d people models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8387-8397 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 262, + 480, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 480, + 294 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 480, + 294 + ], + "type": "text", + "content": "2. Alldieck, T., Zanfir, M., Sminchisescu, C.: Photorealistic monocular 3d reconstruction of humans wearing clothing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 295, + 480, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 295, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 138, + 295, + 480, + 327 + ], + "type": "text", + "content": "3. Bertiche, H., Madadi, M., Escalera, S.: Pbns: Physically based neural simulation for unsupervised garment pose space deformation. ACM Trans. Graph. 40(6) (dec 2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 327, + 480, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 327, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 138, + 327, + 480, + 360 + ], + "type": "text", + "content": "4. Bhatnagar, B.L., Tiwari, G., Theobalt, C., Pons-Moll, G.: Multi-garment net: Learning to dress 3d people from images. In: IEEE International Conference on Computer Vision (ICCV). IEEE (oct 2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 361, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 361, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 361, + 480, + 392 + ], + "type": "text", + "content": "5. Chen, X., Pang, A., Yang, W., Wang, P., Xu, L., Yu, J.: Tightcap: 3d human shape capture with clothing tightness field. ACM Transactions on Graphics (TOG) 41(1), 1-17 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 393, + 480, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 393, + 480, + 436 + ], + "spans": [ + { + "bbox": [ + 138, + 393, + 480, + 436 + ], + "type": "text", + "content": "6. Collet, A., Chuang, M., Sweeney, P., Gillett, D., Evseev, D., Calabrese, D., Hoppe, H., Kirk, A., Sullivan, S.: High-quality streamable free-viewpoint video. ACM Trans. Graph. 34(4) (jul 2015). https://doi.org/10.1145/2766945, https://doi.org/10.1145/2766945" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 437, + 480, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 437, + 480, + 458 + ], + "spans": [ + { + "bbox": [ + 138, + 437, + 480, + 458 + ], + "type": "text", + "content": "7. Corona, E., Pumarola, A., Alenyà, G., Pons-Moll, G., Moreno-Noguer, F.: SMPlicit: Topology-aware generative model for clothed people. In: CVPR (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 459, + 480, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 459, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 459, + 480, + 491 + ], + "type": "text", + "content": "8. Feng, Y., Yang, J., Pollefeys, M., Black, M.J., Bolkart, T.: Capturing and animation of body and clothing from monocular video. In: SIGGRAPH Asia 2022 Conference Papers. SA '22 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 491, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 512 + ], + "type": "text", + "content": "9. Geman, S., McClure, D.E.: Statistical methods for tomographic image reconstruction (1987)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 513, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 513, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 134, + 513, + 480, + 555 + ], + "type": "text", + "content": "10. Grigorev, A., Thomaszewski, B., Black, M.J., Hilliges, O.: HOOD: Hierarchical graphs for generalized modelling of clothing dynamics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 556, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 556, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 134, + 556, + 480, + 589 + ], + "type": "text", + "content": "11. Gropp, A., Yariv, L., Haim, N., Atzmon, M., Lipman, Y.: Implicit geometric regularization for learning shapes. In: Proceedings of Machine Learning and Systems. pp. 3569-3579 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 590, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 590, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 134, + 590, + 480, + 621 + ], + "type": "text", + "content": "12. Guo, C., Chen, X., Song, J., Hilliges, O.: Human performance capture from monocular video in the wild. In: 2021 International Conference on 3D Vision (3DV). pp. 889-898. IEEE (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 622, + 480, + 665 + ], + "type": "text", + "content": "13. Guo, C., Jiang, T., Chen, X., Song, J., Hilliges, O.: Vid2 avatar: 3d avatar reconstruction from videos in the wild via self-supervised scene decomposition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "ReLoo" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "14. Habermann, M., Liu, L., Xu, W., Zollhoefer, M., Pons-Moll, G., Theobalt, C.: Real-time deep dynamic characters. ACM Transactions on Graphics 40(4) (aug 2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 481, + 182 + ], + "type": "text", + "content": "15. Habermann, M., Xu, W., Zollhoefer, M., Pons-Moll, G., Theobalt, C.: Deepcap: Monocular human performance capture using weak supervision. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (jun 2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 214 + ], + "type": "text", + "content": "16. He, T., Xu, Y., Saito, S., Soatto, S., Tung, T.: Arch++: Animation-ready clothed human reconstruction revisited. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 11046-11056 (October 2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 214, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 481, + 246 + ], + "type": "text", + "content": "17. Huang, Z., Xu, Y., Lassner, C., Li, H., Tung, T.: Arch: Animatable reconstruction of clothed humans. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3093-3102 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "type": "text", + "content": "18. Jiang, B., Hong, Y., Bao, H., Zhang, J.: Selfrecon: Self reconstruction your digital avatar from monocular video. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 278, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 481, + 311 + ], + "type": "text", + "content": "19. Jiang, B., Zhang, J., Hong, Y., Luo, J., Liu, L., Bao, H.: Bcnet: Learning body and cloth shape from a single image. In: European Conference on Computer Vision. Springer (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 311, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 481, + 342 + ], + "type": "text", + "content": "20. Jiang, W., Yi, K.M., Samei, G., Tuzel, O., Ranjan, A.: Neuman: Neural human radiance field from a single video. In: Proceedings of the European conference on computer vision (ECCV) (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 342, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 481, + 386 + ], + "type": "text", + "content": "21. Jiang, Z., Guo, C., Kaufmann, M., Jiang, T., Valentin, J., Hilliges, O., Song, J.: Multiply: Reconstruction of multiple people from monocular video in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "type": "text", + "content": "22. Ke, L., Ye, M., Danelljan, M., Liu, Y., Tai, Y.W., Tang, C.K., Yu, F.: Segment anything in high quality. In: NeurIPS (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 407, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 450 + ], + "type": "text", + "content": "23. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4015-4026 (October 2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 450, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 481, + 482 + ], + "type": "text", + "content": "24. Le, B.H., Deng, Z.: Smooth skinning decomposition with rigid bones. ACM Trans. Graph. 31(6) (nov 2012). https://doi.org/10.1145/2366145.2366218, https://doi.org/10.1145/2366145.2366218" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 482, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 481, + 525 + ], + "type": "text", + "content": "25. Li, Y., Habermann, M., Thomaszewski, B., Coros, S., Beeler, T., Theobalt, C.: Deep physics-aware inference of cloth deformation for monocular human performance capture. In: 2021 International Conference on 3D Vision (3DV). pp. 373-384. IEEE (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "type": "text", + "content": "26. Lin, W., Zheng, C., Yong, J.H., Xu, F.: Relightable and animatable neural avatars from videos. In: Proceedings of the AAAI Conference on Artificial Intelligence (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 557, + 481, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 481, + 590 + ], + "type": "text", + "content": "27. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ACM transactions on graphics (TOG) 34(6), 1-16 (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 590, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 481, + 632 + ], + "type": "text", + "content": "28. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4460-4470 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "29. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: European conference on computer vision. pp. 405-421. Springer (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 275, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 275, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 275, + 101 + ], + "type": "text", + "content": "C. Guo and T. Jiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "30. Moon, G., Nam, H., Shiratori, T., Lee, K.M.: 3d clothed human reconstruction in the wild. In: European Conference on Computer Vision (ECCV) (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 181 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 181 + ], + "type": "text", + "content": "31. Pan, X., Mai, J., Jiang, X., Tang, D., Li, J., Shao, T., Zhou, K., Jin, X., Manocha, D.: Predicting loose-fitting garment deformations using bone-driven motion networks. In: ACM SIGGRAPH 2022 Conference Proceedings. SIGGRAPH '22, Association for Computing Machinery, New York, NY, USA (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 181, + 482, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 181, + 482, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 181, + 482, + 214 + ], + "type": "text", + "content": "32. Patel, C., Liao, Z., Pons-Moll, G.: Tailornet: Predicting clothing in 3d as a function of human pose, shape and garment style. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (jun 2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 214, + 482, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 482, + 257 + ], + "type": "text", + "content": "33. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3D hands, face, and body from a single image. In: Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR). pp. 10975-10985 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 257, + 482, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 482, + 278 + ], + "type": "text", + "content": "34. Peng, S., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Zhou, X., Bao, H.: Animatable neural radiance fields for modeling dynamic human bodies. In: ICCV (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 278, + 482, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 278, + 482, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 482, + 321 + ], + "type": "text", + "content": "35. Peng, S., Zhang, Y., Xu, Y., Wang, Q., Shuai, Q., Bao, H., Zhou, X.: Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9054-9063 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 321, + 482, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 482, + 354 + ], + "type": "text", + "content": "36. Pons-Moll, G., Pujades, S., Hu, S., Black, M.: Clothcap: Seamless 4d clothing capture and retargeting. ACM Transactions on Graphics, (Proc. SIGGRAPH) 36(4) (2017), two first authors contributed equally" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 354, + 482, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 386 + ], + "type": "text", + "content": "37. Qiu, L., Chen, G., Zhou, J., Xu, M., Wang, J., Han, X.: Rec-mv: Reconstructing 3d dynamic cloth from monocular videos. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 386, + 482, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 407 + ], + "type": "text", + "content": "38. Ricci, A.: A constructive geometry for computer graphics. Comput. J. 16, 157-160 (1973), https://api_semanticscholar.org/CorpusID:30038820" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 407, + 482, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 449 + ], + "type": "text", + "content": "39. Saito, S., Simon, T., Saragih, J., Joo, H.: Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 84-93 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 449, + 482, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 482 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 482 + ], + "type": "text", + "content": "40. Santesteban, I., Otaduy, M.A., Casas, D.: Snug: Self-supervised neural dynamic garments. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8140-8150 (June 2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 482, + 482, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 482, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 482, + 482, + 514 + ], + "type": "text", + "content": "41. Su, S.Y., Bagautdinov, T., Rhodin, H.: Danbo: Disentangled articulated neural body representations via graph neural networks. In: European Conference on Computer Vision (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 514, + 482, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 557 + ], + "type": "text", + "content": "42. Su, Z., Wan, W., Yu, T., Liu, L., Fang, L., Wang, W., Liu, Y.: Mulaycap: Multi-layer human performance capture using a monocular video camera. IEEE Transactions on Visualization and Computer Graphics 28(4), 1862-1879 (2022). https://doi.org/10.1109/TVCG.2020.3027763" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 557, + 482, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 590 + ], + "type": "text", + "content": "43. Tiwari, G., Bhatnagar, B.L., Tung, T., Pons-Moll, G.: Sizer: A dataset and model for parsing 3d clothing and learning size sensitive 3d clothing. In: European Conference on Computer Vision (ECCV). Springer (August 2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 590, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 590, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 590, + 482, + 632 + ], + "type": "text", + "content": "44. Wang, K., Zhang, G., Cong, S., Yang, J.: Clothed human performance capture with a double-layer neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21098-21107 (June 2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "text", + "content": "45. Wang, S., Schwarz, K., Geiger, A., Tang, S.: Arah: Animatable volume rendering of articulated human sdfs. In: European Conference on Computer Vision (ECCV) (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "ReLoo" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 600 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "46. Wang, Z., Simoncelli, E.P., Bovik, A.C.: Multiscale structural similarity for image quality assessment. In: The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers. vol. 2 (2003)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 194 + ], + "type": "text", + "content": "47. Weng, C.Y., Curless, B., Srinivasan, P.P., Barron, J.T., Kemelmacher-Shlizerman, I.: HumanNeRF: Free-viewpoint rendering of moving people from monocular video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16210-16220 (June 2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "text", + "content": "48. Xiang, D., Prada, F., Bagautdinov, T., Xu, W., Dong, Y., Wen, H., Hodgins, J., Wu, C.: Modeling clothing as a separate layer for an animatable human avatar. ACM Trans. Graph. 40(6) (dec 2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 259 + ], + "type": "text", + "content": "49. Xiu, Y., Yang, J., Cao, X., Tzionas, D., Black, M.J.: ECON: Explicit Clothed humans Optimized via Normal integration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "text", + "content": "50. Xiu, Y., Yang, J., Tzionas, D., Black, M.J.: ICON: Implicit Clothed humans Obtained from Normals. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13296-13306 (June 2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "51. Xu, W., Chatterjee, A., Zollhöfer, M., Rhodin, H., Mehta, D., Seidel, H.P., Theobalt, C.: Monoperfcap: Human performance capture from monocular video. SIGGRAPH 37(2), 27:1-27:15 (May 2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 347 + ], + "type": "text", + "content": "52. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. In: Advances in Neural Information Processing Systems (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 348, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 480, + 380 + ], + "type": "text", + "content": "53. Yariv, L., Kasten, Y., Moran, D., Galun, M., Atzmon, M., Ronen, B., Lipman, Y.: Multiview neural surface reconstruction by disentangling geometry and appearance. In: Advances in Neural Information Processing Systems (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 423 + ], + "type": "text", + "content": "54. Yu, T., Zheng, Z., Zhong, Y., Zhao, J., Dai, Q., Pons-Moll, G., Liu, Y.: Simulcap: Single-view human performance capture with cloth simulation. In: The IEEE International Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (June 2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 424, + 480, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 457 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 457 + ], + "type": "text", + "content": "55. Zhang, J., Liu, X., Ye, X., Zhao, F., Zhang, Y., Wu, M., Zhang, Y., Xu, L., Yu, J.: Editable free-viewpoint video using a layered neural representation. ACM Transactions on Graphics (TOG) 40(4), 1-18 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 458, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 458, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 458, + 480, + 479 + ], + "type": "text", + "content": "56. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv:2010.07492 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "text", + "content": "57. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "text", + "content": "58. Zhang, Z., Sun, L., Yang, Z., Chen, L., Yang, Y.: Global-correlated 3d-decoupling transformer for clothed avatar reconstruction. In: Advances in Neural Information Processing Systems (NeurIPS) (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 567 + ], + "type": "text", + "content": "59. Zheng, Z., Huang, H., Yu, T., Zhang, H., Guo, Y., Liu, Y.: Structured local radiance fields for human avatar modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "text", + "content": "60. Zheng, Z., Yu, T., Liu, Y., Dai, Q.: Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 275, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 275, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 275, + 101 + ], + "type": "text", + "content": "C. Guo and T. Jiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/ReMamber_ Referring Image Segmentation with Mamba Twister/04366454-5dd3-4b5b-8a0b-a5b01b0efbcd_content_list.json b/2024/ReMamber_ Referring Image Segmentation with Mamba Twister/04366454-5dd3-4b5b-8a0b-a5b01b0efbcd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2a0351403d4d76a85a5297aaafc803ade8173a02 --- /dev/null +++ b/2024/ReMamber_ Referring Image Segmentation with Mamba Twister/04366454-5dd3-4b5b-8a0b-a5b01b0efbcd_content_list.json @@ -0,0 +1,1794 @@ +[ + { + "type": "text", + "text": "ReMamber: Referring Image Segmentation with Mamba Twister", + "text_level": 1, + "bbox": [ + 225, + 140, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuhuan Yang\\*1, Chaofan Ma\\*1, Jiangchao Yao1, Zhun Zhong\\*2, Ya Zhang1, and Yanfeng Wang\\*1", + "bbox": [ + 223, + 210, + 779, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Shanghai Jiao Tong University", + "bbox": [ + 388, + 260, + 611, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ University of Nottingham", + "bbox": [ + 406, + 276, + 594, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yangyuhuan, chaofanna, sunarker, ya_zhang, wangyanfeng622}@sjtu.edu.cn", + "bbox": [ + 241, + 290, + 759, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhunzhong007@gmail.com", + "bbox": [ + 415, + 305, + 588, + 316 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Referring Image Segmentation (RIS) leveraging transformers has achieved great success on the interpretation of complex visual-language tasks. However, the quadratic computation cost makes it resource-consuming in capturing long-range visual-language dependencies. Fortunately, Mamba addresses this with efficient linear complexity in processing. However, directly applying Mamba to multi-modal interactions presents challenges, primarily due to inadequate channel interactions for the effective fusion of multi-modal data. In this paper, we propose ReMamber, a novel RIS architecture that integrates the power of Mamba with a multi-modal Mamba Twister block. The Mamba Twister explicitly models image-text interaction, and fuses textual and visual features through its unique channel and spatial twisting mechanism. We achieve competitive results on three challenging benchmarks with a simple and efficient architecture. Moreover, we conduct thorough analyses of ReMamber and discuss other fusion designs using Mamba. These provide valuable perspectives for future research. The code has been released at: https://github.com/yyh-rain-song/ReMamber.", + "bbox": [ + 259, + 344, + 754, + 582 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Referring Image Segmentation (RIS) $\\cdot$ Multi-Modal Understanding $\\cdot$ Mamba Architecture", + "bbox": [ + 259, + 592, + 743, + 621 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 661, + 377, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Referring Image Segmentation (RIS) is a crucial yet challenging task in the area of multi-modal understanding [15, 25, 29]. Unlike ordinary image segmentation, RIS involves the identification and segmentation of specific objects in image based on the textual descriptions. This thereby requires the model to be capable of understanding vision-language interactions, which is the core challenge of RIS.", + "bbox": [ + 212, + 688, + 787, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Thanks to the powerful attention mechanisms, it has achieved great success for RIS by leveraging transformers to promote the exact recognition of multimodal information. For example, existing works have designed transformer decoder [4,22] or transformer encoder-decoder [24,35,36,51,55,56] to comprehensively fuse visual and linguistic features, which have achieved great progress.", + "bbox": [ + 212, + 765, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9d93d1e841052cfac5d99205b13358187664222d816e3f973a0ca52663f8b2c6.jpg", + "image_caption": [ + "Fig. 1: We propose ReMamber, a novel referring segmentation architecture with Mamba twister. It consists of several Mamba Twister block. Each block contains several visual state space (VSS) layers and a Twisting layer. The Twisting layer first calculates the interaction between image and text, and then forms a hybrid feature cube. Finally, it \"twists\" the feature cube using the Channel and Spatial Scan along each dimension." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 784, + 369 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nonetheless, it is a quadratic increase in both computation and memory when applying full attention in transformers [5, 49]. This leads to the limitation for resource-intensive scenarios, for example, in capturing long-range visual-language dependencies. And this is particularly important in the context of large-size images with long textual descriptions.", + "bbox": [ + 214, + 470, + 787, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fortunately, recent advances of State Space Models (SSMs) [7,8,11,46] has emerged as promising architectures for solving the above issue. Specifically, Mamba [8] marks a significant advancement for efficient training and inference with linear complexity, which has been incorporated into various visual tasks. However, current efforts primarily pay attention to single-modality settings, such as image classification [33,62], biomedical image segmentation [31,37,44,52], low-level vision [14,60], and point cloud analysis [26,59]. In this paper, we pioneer the exploration of Mamba in multi-modal RIS setting, and identify that the prevalent multi-modality token splicing method in transformers [4,24,56] is no longer effective. Since in Mamba, a fundamental deficiency exists: the interactions are insufficient between channels of different tokens [8], which adversely affects the fusion of multi-modal information in RIS.", + "bbox": [ + 214, + 549, + 787, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this dilemma, we propose ReMamber, a novel referring segmentation architecture with Mamba twister. As shown in Fig. 1, ReMamber comprises several Mamba Twister blocks, which allow the model to be \"aware\" of the textual context at every spatial location. Each block consists of several visual state space (VSS) layers and a Twisting layer. Specifically, the VSS layers initially extract visual features, and the Twisting layer injects the textual information into the visual modality. The Twisting layer is structured into three critical", + "bbox": [ + 214, + 734, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Yang, C. Ma, J. Yao, Z. Zhong, Y. Zhang, and Y. Wang", + "bbox": [ + 269, + 114, + 666, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "components. (1) A vision-language interaction operation is designed to explicitly capture the fine-grained interactions between modalities. This is achieved by evaluating the similarity between visual and textual tokens then mapping them into a shared feature space, thus generating a multi-modal feature. (2) A hybrid feature cube is then created by concatenating the visual feature, multimodal feature, and global textual feature. This ensures that each visual token receives uniform influence from both local and global contexts, preventing the overshadowing of subtle textual cues by predominant visual features. (3) Lastly, to address the inadequate interaction within channels in Mamba, a twisting mechanism is deployed. This process \"twists\" the feature cube channel- and spatial-wise, thereby enhancing interaction within and across modalities. It is accomplished through two consecutive SSMs scanning along channel and spatial dimensions, respectively.", + "bbox": [ + 212, + 146, + 787, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To sum up, the key contributions of our research are highlighted as follows:", + "bbox": [ + 212, + 343, + 785, + 359 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We pioneer the exploration of Mamba in referring image segmentation (RIS), demonstrating Mamba's significant potential for multi-modal understanding.", + "- We design a novel framework, ReMamber, that mainly contains several Mamba Twister blocks. This design effectively captures vision-language interactions using the \"twisting mechanism\".", + "- We achieve competitive results on three challenging benchmarks. Moreover, we conduct thorough analyses of ReMamber and discuss other vision-language fusion designs using Mamba. These provide valuable perspectives for future research." + ], + "bbox": [ + 212, + 359, + 784, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 521, + 387, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Referring Image Segmentation", + "text_level": 1, + "bbox": [ + 215, + 555, + 511, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Referring image segmentation (RIS) aims to segment the entities in the image following natural language instruction. Early approaches [15, 25, 29, 39] leveraged RNNs or LSTMs to encode linguistic representations, and CNNs to extract spatial features from the image at varying levels. These disparate modalities were then integrated using the multi-modal LSTM [2, 15, 21, 25, 29, 39], attention mechanism [6, 16, 45, 57], cycle-consistency [3], and graph convolution [17].", + "bbox": [ + 212, + 582, + 787, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent trends have shifted towards leveraging transformers for enhanced capturing and fusion of vision-language modalities. MDETR [22] and VLT [4] design a transformer decoder for fusing linguistic and visual features. LAVT [56] adopts Swin Transformer as the visual backbone and incorporate vision-language fusion modules within the visual encoder's final layers. Similar strategies are employed by ReSTR [24] and CRIS [51], which utilize dual transformer encoders for initial modality encoding, followed by feature fusion through a multi-modal transformer encoder or decoder. Other models like PolyFormer [30], SeqTr [61] and [43] also adopt a multi-modal transformer for vision-language fusion, but they output masks as sequences of contour points. Meanwhile, GRES [28] and CGFormer [47] consider transformer queries as region proposals, and regarding", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "ReMamber: Referring Image Segmentation with Mamba Twister", + "bbox": [ + 302, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "segmentations as proposal-level classification problems. Distinct from all existing works, we introduce a pioneering multi-modal architecture named ReMamber, based on Mamba [8]. This novel approach underscores the untapped potential of Mamba in advancing the field of referring image segmentation.", + "bbox": [ + 212, + 146, + 787, + 208 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 State Space Models and Visual Applications", + "text_level": 1, + "bbox": [ + 214, + 229, + 627, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "State space models (SSMs), originally derived from control theory, have been effectively combined with deep learning to model the long-range dependencies. Early works like LSSL [12] show potential when combined with HiPPO [9] initialization. S4 [11] was designed to diminish both computational and memory demands associated with state representations. It scales linearly with sequence length, offering a notable advantage over CNNs and transformers. Building on S4, S5 [46] introduces the efficient parallel scan and MIMO SSM, further refining the approach. Later, H3 [7] expanded on these foundations, achieving competitive performance with transformers in language modeling. Recently, Mamba [8] marked a significant advancement with its linear-time inference and efficient training, incorporating a selection mechanism and hardware-aware algorithms building upon prior works [10, 13, 40].", + "bbox": [ + 212, + 253, + 787, + 435 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the demonstrated success of SSMs like Mamba in language modeling, researchers have begun to investigate their applicability to visual tasks. Works such as ViS4mer [19], Selective S4 [50] and TranS4mer [20] directly use the S4's ability to capture long-range sequences for understanding inter-frame relations in video classification and detection. However, they still employ Vision Transformers for intra-frame feature extraction. More recently, Vim [62] and VMamba [33] have shown promising results as fully Mamba-based vision backbones for image classification, detection, and segmentation. At the same time, several studies have explored Mamba-based architecture on various vision tasks, such as biomedical image segmentation [31, 37, 44, 52], low-level vision [14, 60], and point cloud analysis [26, 59]. Nevertheless, all above efforts primarily focus on single-modality vision tasks. In this paper, we pioneer the exploration of Mamba-base architecture in multi-modality RIS task.", + "bbox": [ + 212, + 436, + 787, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methods", + "text_level": 1, + "bbox": [ + 214, + 655, + 339, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This paper aims to develop a multi-modal Mamba-based architecture for RIS task. In Sec. 3.1, we give the preliminary. From Sec. 3.2 to Sec. 3.5, we describe our architecture from coarse to fine. In Sec. 3.6, we introduce other three variants besides our design. And in Sec. 3.7 we detail the training process.", + "bbox": [ + 212, + 686, + 787, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminary: State Space Model", + "text_level": 1, + "bbox": [ + 214, + 768, + 527, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "State Space Model (SSM) is a sequence model inspired by continuous systems. It is designed to capture a mapping relationship between two functions or sequences, expressed as $x(t) \\in \\mathbb{R} \\mapsto y(t) \\in \\mathbb{R}$ , through a hidden state $h(t) \\in \\mathbb{R}^N$ .", + "bbox": [ + 212, + 794, + 785, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Yang, C. Ma, J. Yao, Z. Zhong, Y. Zhang, and Y. Wang", + "bbox": [ + 271, + 113, + 666, + 130 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The evolution of the hidden state over time is governed by specific parameters $\\mathbf{A} \\in \\mathbb{R}^{N \\times N}$ , which directs the state evolution, and $\\mathbf{B} \\in \\mathbb{R}^N$ , $\\mathbf{C} \\in \\mathbb{R}^N$ , which performs the projection. Formally, SSM can be described by the following equations:", + "bbox": [ + 215, + 146, + 787, + 194 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} h ^ {\\prime} (t) = \\mathbf {A} h (t) + \\mathbf {B} x (t), \\\\ (t) = \\mathbf {C} h ^ {\\prime} (t) \\end{array} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 419, + 204, + 785, + 228 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ny (t) = \\mathbf {C} h (t).\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 222, + 521, + 238 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discretizing this system introduces a timescale parameter $\\Delta t$ , which transforms the continuous parameters $\\mathbf{A}, \\mathbf{B}$ into their discrete counterparts $\\overline{\\mathbf{A}}, \\overline{\\mathbf{B}}$ . A common way for this transformation is the zero-order hold (ZOH) approach, defined as:", + "bbox": [ + 212, + 242, + 782, + 299 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\overline {{\\mathbf {A}}} = \\exp (\\boldsymbol {\\Delta} \\mathbf {A}), \\\\ \\overline {{\\mathbf {B}}} = (\\boldsymbol {\\Delta} \\mathbf {A}) ^ {- 1} (\\boldsymbol {\\Delta} \\mathbf {A}) - \\mathbf {I}) - \\boldsymbol {\\Delta} \\mathbf {B}. \\end{array} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 299, + 784, + 325 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\mathbf {B}}} = (\\boldsymbol {\\Delta} \\mathbf {A}) ^ {- 1} (\\exp (\\boldsymbol {\\Delta} \\mathbf {A}) - \\mathbf {I}) \\cdot \\boldsymbol {\\Delta} \\mathbf {B}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 311, + 784, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, the discretized version of this system can be represented as:", + "bbox": [ + 215, + 340, + 692, + 354 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh _ {t} = \\overline {{\\mathbf {A}}} h _ {t - 1} + \\overline {{\\mathbf {B}}} x _ {t}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 433, + 364, + 784, + 390 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ny _ {t} = \\mathbf {C} h _ {t}.\n$$\n", + "text_format": "latex", + "bbox": [ + 434, + 383, + 503, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mamba [8], as a variant of SSM, releases the linearity constraint of the original SSM described in Eq. (1), by making $\\mathbf{B}$ and $\\mathbf{C}$ to be input-dependent, whereas still maintaining the linear complexity during forward process.", + "bbox": [ + 212, + 407, + 782, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Architecture Overview", + "text_level": 1, + "bbox": [ + 215, + 473, + 449, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The pivotal aspect of Referring Image Segmentation (RIS) is the correspondence between image and textual input. To achieve this, we propose ReMamber. Fig. 2 shows the overview of our architecture. The basic block for ReMamber is the Mamba Twister block. It is a multi-modal fusion block that takes both visual and textual feature as input, and outputs the fused multi-modal feature representation. We extract the intermediate feature after each Mamba Twister block, and then feed it into a flexible decoder to generate the final segmentation mask. The decoder is task-invariant, so any downstream architecture can be applied. Below in Sec. 3.3, we will first introduce the key component: Mamba Twister block.", + "bbox": [ + 212, + 494, + 787, + 645 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 Mamba Twister Block", + "text_level": 1, + "bbox": [ + 215, + 666, + 444, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 2, the Mamba Twister Block consists of several visual state space (VSS) layers and a Twisting layer. The VSS layer is designed to process features in the spatial domain. It treats the input feature as a series of image tokens and aims to discern the spatial relationships among these tokens. The Twisting layer aims to inject text condition into the image feature, thus guiding the transformation of image feature. It begins by condensing the textual sequence to compute the cross-correlation between the image and textual modalities. Subsequently, it disseminates this information across each image feature patch via a twisting operation. We'll detail the introduction of VSS layer in Sec. 3.4 and the Twisting layer in Sec. 3.5.", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "ReMamber: Referring Image Segmentation with Mamba Twister", + "bbox": [ + 302, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d425cdebb12538c725842f9fa2a8d47af8f9c7a3b5d5779d3ec499e6f6b593e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 143, + 787, + 282 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c66f5d29467cef28a888958bc867e7063e0607952d56740cde5f43b7569efc97.jpg", + "image_caption": [ + "Fig. 2: Overview architecture of our ReMamber. The basic block for ReMamber is the Mamba Twister block. It consists of several visual state space (VSS) layers and a Twisting layer. The Twisting layer first constructs hybrid feature cube from text, image, and multi-modal features via channel concatenation. Then, it \"twists\" the cube by Channel Scan and Spatial Scan. We extract intermediate features after each Mamba Twister block, and feed it into a flexible decoder for final segmentation." + ], + "image_footnote": [], + "bbox": [ + 241, + 290, + 743, + 404 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4 VSS Layer for Spatial Data Processing", + "text_level": 1, + "bbox": [ + 215, + 512, + 583, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since the SSM is initially designed for processing temporal or causal data, which can not effectively process non-causal data types, i.e., 2-dimensional image in our case. To solve this issue, we adopt the Cross-Scan-Module proposed in [33], which unfolds image patches into sequences and then scans them in four distinct directions, ensuring that the information from all pixels is integrated during feature transformation. We replace the scanning operation in vanilla Mamba block with CSM, forming a visual state space (VSS) layer for our spatial feature transformation.", + "bbox": [ + 212, + 537, + 787, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5 Twisting Layer for Multi-modal Fusion", + "text_level": 1, + "bbox": [ + 215, + 679, + 583, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here we introduce the detailed design of our Twisting layer. Intuitively, the Twisting layer first constructs a feature cube by arranging information from different modalities orderly. Then, pieces of information intertwine with each other when the cube is twisted through two SSM layers along different axis, thereby achieving modality fusion. Formally, given image feature $\\mathbf{F}_i\\in \\mathbb{R}^{h\\times w\\times C_i}$ and text feature $\\mathbf{F}_t\\in \\mathbb{R}^{L\\times C_t}$ , the Mamba twister aims to learn a function $(\\mathbf{F}_i,\\mathbf{F}_t)\\mapsto \\tilde{\\mathbf{F}}\\in \\mathbb{R}^{h\\times w\\times C_o}$ by first forming a hybrid multi-modal feature cube and then conduct two SSM scans on it. $C_i,C_t$ and $C_o$ are the dimensions of visual, textual, and output feature; $h,w$ and $L$ are height, weight, and length.", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Yang, C. Ma, J. Yao, Z. Zhong, Y. Zhang, and Y. Wang", + "bbox": [ + 269, + 114, + 666, + 130 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Forming the Hybrid Feature Cube. To explicitly build the image-text correspondence, we design a novel vision-language interaction operation and then form a hybrid feature cube. This allows more effective modality fusion.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To be specific, we formulated the vision-language interaction operation into two manners: global interaction and local interaction. The global interaction treats the text sequence as a whole, which means that all image patches should be aware of the semantic meaning of the text expression. We pool a global representation $\\mathbf{F}_t^{\\mathrm{CLS}}\\in \\mathbb{R}^{C_t}$ from the text sequence $\\mathbf{F}_t$ to convey this information. The $\\mathbf{F}_t^{\\mathrm{CLS}}$ is further expanded to the same size as the image feature. Formally:", + "bbox": [ + 212, + 191, + 784, + 282 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {F}} _ {t} = \\operatorname {E x p a n d} \\left(\\mathbf {F} _ {t} ^ {\\mathrm {C L S}}\\right) \\in \\mathbb {R} ^ {h \\times w \\times C _ {t}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 290, + 784, + 308 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\operatorname{Expand}(\\cdot)$ denotes the operation that expands the input tensor to the same size as the image feature.", + "bbox": [ + 212, + 316, + 782, + 345 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "However, a global representation may not be enough to capture the intricate relationships between different modalities. For example, some words such as color, shape or location may be more relevant to certain image patches than others. The local interaction aims to capture such correlation in a fine-grained level. Formally, we calculate the local interaction map $\\mathbf{F}_c$ using matrix multiplication:", + "bbox": [ + 212, + 347, + 784, + 434 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {c} = \\mathbf {F} _ {i} \\mathbf {W} _ {i} \\cdot \\left(\\mathbf {F} _ {t} \\mathbf {W} _ {t}\\right) ^ {T} \\in \\mathbb {R} ^ {h \\times w \\times L}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 435, + 784, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_i\\in \\mathbb{R}^{C_i\\times C_c}$ and $\\mathbf{W}_t\\in \\mathbb{R}^{C_t\\times C_c}$ are learnable parameters. To enhance feature processing in high dimension, we use a single convolutional layer to transform $\\mathbf{F}_c$ into $\\tilde{\\mathbf{F}}_c\\in \\mathbb{R}^{h\\times w\\times C_c}$ . We then concatenate $\\mathbf{F}_i$ , $\\tilde{\\mathbf{F}}_t$ , and $\\tilde{\\mathbf{F}}_c$ along the channel dimension to form the hybrid feature cube. Formally:", + "bbox": [ + 212, + 455, + 784, + 518 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {\\text {c u b e}} = \\left[ \\mathbf {F} _ {i}, \\tilde {\\mathbf {F}} _ {t}, \\tilde {\\mathbf {F}} _ {c} \\right] \\in \\mathbb {R} ^ {h \\times w \\times \\left(C _ {i} + C _ {t} + C _ {c}\\right)}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 526, + 784, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Twisting the Hybrid Feature Cube. Generally, the scanning operation for vanilla SSM is almost independent across channels [8]. This is not enough for feature communication, especially when information from different modalities is arranged along the channel dimension, i.e., our hybrid feature cube.", + "bbox": [ + 212, + 561, + 782, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To foster the interactions within and across modalities, here we design a twisting mechanism to \"twist\" the hybrid feature cube along different axis. Specifically, it is composed of the Channel Scan and Spacial Scan successively. The Channel Scan first treats the channel as an ordered sequence, and learns to communicate cross channels, thus foster the modality fusion. Then, the Spatial Scan operates on the spatial dimension, and learns to communicate information cross patches within each modality separately. Formally, this process can be written as:", + "bbox": [ + 212, + 623, + 784, + 741 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {\\text {o u t}} = \\operatorname {S S M} _ {\\text {s p a t i a l}} \\left(\\operatorname {S S M} _ {\\text {c h a n n e l}} \\left(\\mathbf {F} _ {\\text {c u b e}}\\right)\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 744, + 784, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathrm{SSM}_{\\mathrm{channel}}$ refers to the Channel Scan. It is an SSM layer that treats the concatenated feature as an ordered sequence through channel dimension, and presents 1-D selective scan along the channel. While $\\mathrm{SSM}_{\\mathrm{spatial}}$ denotes the Spatial Scan. It is an VSS layer and presents 2D selective scan along the two spatial dimension. $\\mathbf{F}_{\\mathrm{out}}$ is then feed into the next layer for further process.", + "bbox": [ + 212, + 763, + 782, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "ReMamber: Referring Image Segmentation with Mamba Twister", + "bbox": [ + 302, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f6619b0a50cf049cff6d6f9dfc3fe3b7c5b43fd0d01c0616a069983123bb7415.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 143, + 656, + 170 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8f927a0c10a71fd6e58ee3e0efcfedee13aa408c9945f262d0adba7252f423a9.jpg", + "image_caption": [ + "Fig.3: Other multi-modal fusion designs. (a) In-context Conditioning appends text tokens ahead of image tokens. (b) Attention-based Conditioning utilizes cross-attention mechanism for modality fusion. (c) Norm Adaptation learns a scale and bias for the model's normalization layers." + ], + "image_footnote": [], + "bbox": [ + 230, + 172, + 774, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.6 Discussion: Other Variants of Modality Fusion Designs", + "text_level": 1, + "bbox": [ + 214, + 430, + 714, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Besides the proposed Twisting layer, we also explore other variants of multimodal fusion designs. Fig. 3 shows the structure of three other different variants. We here discuss these variants, and later in Sec. 4.3, we analyze the results for these variants.", + "bbox": [ + 212, + 455, + 787, + 516 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In-context Conditioning. As shown in Fig. 3 (a), in this variant, we simply append the text sequence before the image feature, forming a longer sequence than originally being processed. In this way, the model is able to fuse image features with the previous text as a context. This is the most straightforward way to allow the model aware of the text condition.", + "bbox": [ + 212, + 517, + 787, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Attention-based Conditioning. Attention mechanism is usually a strong baseline for sequence modeling. Here we utilize the cross-attention mechanism for two modality fusion, as shown in Fig. 3 (b), To be specific, we use the image feature as query, and the text feature as key and value, hoping to capture the correlation between image and text.", + "bbox": [ + 212, + 593, + 787, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Norm Adaptation. In this variant, we integrate the text input by adapting the scale and bias after norm layer using FiLM [42]. We first pool a global representation from text sequence, and then use a linear projection to transform it into the scale and shift in layer normalization.", + "bbox": [ + 212, + 669, + 787, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.7 Model training", + "text_level": 1, + "bbox": [ + 214, + 753, + 388, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The proposed ReMamber is a general framework for multi-modal fusion. Therefore, there is no restriction on the downstream decoder design nor the loss function. Here we use a simple convolution-based decoder for simplicity. The entire network is trained in an end-to-end manner.", + "bbox": [ + 212, + 779, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Yang, C. Ma, J. Yao, Z. Zhong, Y. Zhang, and Y. Wang", + "bbox": [ + 269, + 113, + 668, + 130 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 143, + 375, + 162 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets and Metrics", + "text_level": 1, + "bbox": [ + 215, + 175, + 439, + 189 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Dataset. To assess the efficacy of our proposed approach, we executed experiments across three widely recognized datasets tailored for the referring image segmentation (RIS) task: RefCOCO, RefCOCO+ [23], and G-Ref [23,38,41]. RefCOCO [23] is a prominent dataset in the RIS domain, comprising 19,994 images and 142,210 referring expressions associated with 50,000 objects derived from the MSCOCO [27] dataset through a two-player game. RefCOCO+ enhances the challenge by excluding expressions with absolute location references, featuring 141,564 expressions for 49,856 objects across 19,992 images. G-Ref [38,41] enriches the dataset diversity with 104,560 referring expressions for 54,882 objects in 26,711 images, showcasing an average sentence length of 8.4 words with a heightened focus on location and appearance descriptions.", + "bbox": [ + 212, + 199, + 787, + 364 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation Metrics. In line with preceding studies [48], we adopt mIoU and oIoU as the main metrics. Furthermore, we incorporate the metric Precision@X $(\\mathrm{X} \\in \\{50, 60, 70, 90\\})$ for a more comprehensive evaluation, whereas Percision@X means the percentage of test images with an IoU score high than X%.", + "bbox": [ + 214, + 364, + 785, + 426 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 446, + 455, + 462 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our model architecture is developed upon the foundations of Mamba [8] and VMamba [33]. We adopt ImageNet pre-trained weights as initialization, and train the model in an end-to-end manner. We set the input image resolution to 480 when compared with the state-of-the-art methods in Tab. 2, following [48,56] for fair comparison. For other experiments (Tabs. 1 and 3 to 4), we set the image resolution to 256 for faster training and ablation study without changing its property. For details about the training settings, please refer to our code. $^{1}$", + "bbox": [ + 212, + 470, + 787, + 575 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Evaluation on Other Variants of Modality Fusion Designs", + "text_level": 1, + "bbox": [ + 215, + 597, + 736, + 613 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As discussed in Sec. 3.6, besides our Mamba Twister, we also provide three other variant architectures for modality fusion, including In-context Conditioning, Attention-based Conditioning and Norm Adaptation. To make a fair comparison, for all architectures, we initialize their common part (VMamba-based parameters) with the same checkpoint. The difference is that the Twisting layer is replaced by the corresponding fusion methods. The parameters of fusion modules are nearly equal, i.e., {103, 113, 118, 116} KB for {In-Context, Attention, Norm Adaptation, Twister}. Here we provide a comprehensive analysis between these architectures, showing the advantages and disadvantages of each variant.", + "bbox": [ + 212, + 619, + 787, + 756 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Tab. 1 shows the comparison result. The Mamba Twister consistently outperforms the other variants across all metrics and datasets, indicating its superior capability in capturing and integrating contextual information for more accurate segmentation.", + "bbox": [ + 212, + 757, + 787, + 816 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "ReMamber: Referring Image Segmentation with Mamba Twister", + "bbox": [ + 302, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "1 https://github.com/yyh-rain-song/ReMamber", + "bbox": [ + 217, + 823, + 540, + 840 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/65969b4ff9432e7e8488ae623456f22498a614a09902ee730c515f71e96a5466.jpg", + "table_caption": [ + "Table 1: Comparison with other modality fusion variants. \"Attention\" means for \"Attention-based Conditioning\", \"In-Context\" for \"In-Context Conditioning\", and \"Adaptation\" for \"Norm Adaptation\". Mamba Twister steadily outperforms other variants across all metrics and datasets, indicating its superior capability in capturing and integrating contextual information for more accurate segmentation." + ], + "table_footnote": [], + "table_body": "
Dataset: RefCOCO
VariantsmIoUvalPr@70mIoUtestAmIoUtestB
oIoUPr@50Pr@70oIoUPr@50Pr@70oIoUPr@50
Attention65.362.375.160.667.564.479.165.461.758.069.653.4
In-Context69.165.979.767.571.268.982.271.866.262.875.061.7
Adaptation70.267.080.769.672.370.283.173.266.862.875.163.1
Mamba Twister71.668.482.170.973.371.584.874.568.464.577.164.9
Dataset: RefCOCO+
VariantsmIoUvalPr@70mIoUtestAmIoUtestB
oIoUPr@50oIoUPr@50Pr@70oIoUPr@50
Attention54.049.760.745.558.755.868.552.946.941.950.636.6
In-Context58.454.066.353.963.259.472.260.551.547.556.444.5
Adaptation60.355.268.257.364.259.973.462.953.948.759.047.2
Mamba Twister61.657.370.058.565.862.175.564.954.049.959.447.8
Dataset: G-Ref
VariantsmIoUvalPr@70Pr@90mIoUoIoUtest
oIoUPr@50Pr@50Pr@60Pr@70Pr@90
Attention50.549.254.246.036.78.150.850.857.349.838.66.4
In-Context54.853.459.552.343.413.355.554.860.753.645.313.3
Adaptation59.356.766.761.053.718.458.656.665.559.852.818.2
Mamba Twister61.158.068.462.855.218.561.259.069.063.455.518.7
", + "bbox": [ + 215, + 224, + 787, + 522 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Surprisingly, despite widely used in transformers, the Attention-based Conditioning performs poorly in our task. This suggests that the cross-attention mechanism is not inherently suitable for Mamba architecture. This may be due to a fundamental discrepancy between the two systems. To be specific, (1) the Mamba model is predicted on ordered sequences that exhibit strict sequential dependencies, where the state at a given time point $t + 1$ is determined by the preceding time point $t$ and a hidden state; (2) In contrast, cross-attention mechanisms treats all tokens within a sequence equally. This discrepancy may undermine the Mamba model's ability to structurally model sequences, as the cross-attention mechanism does not preserve the sequential integrity and the hierarchical dependencies essential for the model's operation.", + "bbox": [ + 212, + 536, + 789, + 700 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Serving as a straightforward baseline, In-context Conditioning performs sub-optimal. This may because it models image-text interaction in an implicit way. In our situation, the length of the image feature is much larger than the textual side, textual information may be diluted during the forward process. Besides, this operation does not differentiate the textual information between image patches, and would be insufficient for multi-modal fusion.", + "bbox": [ + 212, + 702, + 787, + 792 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Finally, Norm Adaptation with explicit image-text interaction modeling appears to be a strong baseline. It achieves better performance than other two variants. However, when calculating scale and bias, it uses only a pooled vector as", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Yang, C. Ma, J. Yao, Z. Zhong, Y. Zhang, and Y. Wang", + "bbox": [ + 271, + 113, + 666, + 130 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "textual representation instead of the entire sequence. This potentially results in the loss of information, making it less effective compared to our Mamba Twister.", + "bbox": [ + 214, + 146, + 787, + 176 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Case-study using Attention Maps. In Fig. 4, we further visualized the attention maps in Attention-based Conditioning as well as the local interaction map in Mamba Twister defined in Eq. (5). The four images in Fig. 4(b) are arranged from left to right with the network going from shallow to deep. It can be observed that in the shallower layers, Mamba Twister focuses more on the lower-level features of the images such as edges. As the layer goes deeper, the cost map mostly concentrating on the target object, indicating that Mamba Twister can progressively guide the image feature towards the target described by the language. In contrast, though the Attention-based Conditioning variant is able to locate the target, its attention map performs poorly. This may suggest that the cross-attention mechanism struggles to accurately capture the correct context information, and the two modalities are not truly fused together.", + "bbox": [ + 212, + 199, + 789, + 381 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/8b7666cefd151e855403be655d58509843b7dc2561f127aaaf01cdfc9768a29b.jpg", + "image_caption": [ + "Fig. 4: Cross-Attention map (up) and our local interaction map (down) comparison. Though both methods are able to predict target correctly, the cross-attention maps don't show correct image-text correlation, while ours are able to capture this relationship accurately, indicating that Mamba Twister is able to gradually fusing the two modality." + ], + "image_footnote": [], + "bbox": [ + 217, + 404, + 787, + 579 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Comparison with the State-of-the-Arts", + "text_level": 1, + "bbox": [ + 214, + 693, + 584, + 709 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Tab. 2 shows the comparison of our ReMamber with state-of-the-art methods on RefCOCO, RefCOCO+, and G-Ref datasets. ReMamber outperforms all other methods across all datasets, demonstrating its superior capability in capturing and integrating contextual information for more accurate segmentation. In particular, Mamba outperforms previous Swin-based methods such as LAVT [56] by a considerable margin, indicating the remarkable capability of the Mamba-based architecture in segmentation tasks. and the efficiency of the newly introduced Mamba Twister.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "ReMamber: Referring Image Segmentation with Mamba Twister", + "bbox": [ + 302, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/dbd2380162b1df3e22cc7b5d74505330d85f752abf1e03e5a2a48a385eaf8ec0.jpg", + "table_caption": [ + "Table 2: Comparison with state-of-the-art methods on RefCOCO, RefCOCO+, and G-Ref. The metric of oIoU is reported. Best results are in bold." + ], + "table_footnote": [], + "table_body": "
ModelBackboneRefCOCORefCOCO+G-Ref
valtestAtestBvaltestAtestBvalUtestU
PCAN [1]ResNet-5069.5171.6464.1858.2563.6848.8959.9860.8
MAAttNet [58]MaskRCNN ResNet-10156.5162.3751.746.6752.3940.0847.6448.61
RMI [29]DeepLab ResNet-10145.1845.6945.5729.8630.4829.50--
RRN [25]DeepLab ResNet-10155.3357.2653.9539.7542.1536.11--
CMSA [57]DeepLab ResNet-10158.3260.6155.0943.7647.637.89--
CAC [3]DeepLab ResNet-10158.9061.7753.81---46.3746.95
STEP [2]DeepLab ResNet-10160.0463.4657.9748.1952.3340.41--
BRINet [16]DeepLab ResNet-10160.9862.9959.2148.1752.3242.11--
CMPC [17]DeepLab ResNet-10161.3664.5359.6449.5653.4443.23--
LSCM [18]DeepLab ResNet-10161.4764.9959.5549.3453.1243.50--
CMPC+ [32]DeepLab ResNet-10162.4765.0860.8250.2554.0443.47--
BUSNet [54]DeepLab ResNet-10163.2766.4161.3951.7656.8744.13--
CGAN [34]DeepLab ResNet-10164.8668.0462.0751.0355.5144.0651.0151.69
EFN [6]Wide ResNet-10162.7665.6959.6751.5055.2443.01--
ETRIS [53]CLIP ResNet-10171.0674.1166.6662.2368.5152.7960.2860.42
ReSTR [24]ViT-B-1667.2269.364.4555.7860.4448.2754.48-
ETRIS [53]ViT-B-1670.5173.5166.6360.1066.8950.1759.8259.91
CRIS [51]CLIP ResNet-5069.5272.7264.761.3967.152.4859.8760.36
CRIS [51]CLIP ResNet-10170.4773.1866.1062.2768.0853.6859.8760.36
LAVT [56]Swin-B72.7375.8268.7962.1468.3855.1061.2462.10
ReMamber (Ours)Mamba-B74.5476.7470.8965.0070.7857.5363.964.0
", + "bbox": [ + 217, + 185, + 787, + 434 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/804f4951d1ca96403a286cebe664842dae69edb4b2ed4baeb11c8582798b7cea.jpg", + "table_caption": [ + "Table 3: Ablation on the combination of two scans. We ablated each scan separately, as well as combining them parallel or swap the order. The occurrence of Spatial Scan affects most to the performance. The three variants perform similarly, with the combination of Spatial-Channel slightly better. Results are evaluated on RefCOCO." + ], + "table_footnote": [], + "table_body": "
Channel ScanSpatial ScanvaltestAtestB
mIoUoIoUPr@50Pr@70mIoUoIoUPr@50Pr@70mIoUoIoUPr@50Pr@70
✓-62.360.170.757.264.963.274.361.959.757.566.351.9
-✓70.065.880.569.172.371.083.773.467.563.575.262.7
Parallel71.068.881.970.073.171.884.774.167.964.676.964.5
Spatial-Channel71.468.882.470.873.371.484.974.267.964.776.564.4
Channel-Spatial71.668.482.170.973.371.584.874.568.464.577.164.9
", + "bbox": [ + 217, + 516, + 787, + 602 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 614, + 388, + 630 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effects of two Scans. Tab. 3 presents a comparative analysis of different scan variants. We provide outcomes of conducting Channel Scan and Spatial Scan independently. Additionally, results from various combinations of these two scans are discussed. The term \"Parallel\" denotes the concurrent execution of both scans, followed by adding the independent output for fusion. \"Channel-Spatial\" refers to the original scanning order in Mamba Twister shown in Fig. 2, where the Channel Scan is executed first, followed by the Spatial Scan. \"Spatial-Channel\" refers to reversing the sequence of the two scans, starting with Spatial Scan and then proceeding to Channel Scan.", + "bbox": [ + 212, + 642, + 785, + 777 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Tab. 3 indicates that utilizing any single scan on its own yields suboptimal results, with the use of Channel Scan alone experiencing the most significant drop in effectiveness. This decline may indicate that a pure Channel Scan alters the data distribution, adversely affecting network stability. Among the three", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Yang, C. Ma, J. Yao, Z. Zhong, Y. Zhang, and Y. Wang", + "bbox": [ + 271, + 113, + 666, + 130 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "hybrid scanning policies, each has its strengths and weaknesses. Overall, the combination of Channel-Spatial Scan appears to offer a considerable advantage.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Distribution Visualization. The visualization of the distribution of the multi-modal input data is depicted in Fig. 5, where we use PCA to map the feature to a 3D space. The image and text data are indicated in red and blue respectively.", + "bbox": [ + 212, + 196, + 782, + 256 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Fig. 5a illustrates the distribution of the input data. We can observe that the text data forms a linear arrangement within the 3D space, as it is duplicated $H \\times W$ times to align with the image size. This suggests that the text instruction should be applied for each pixel within the image. Fig. 5b reveals the data distribution following the application of a Channel Scan. Notably, this process appears to lean towards aggregating different modalities towards a distribution pattern similar with text. Fig. 5c demonstrates the data distribution after a Spatial Scan. Here, the features of the textual and image data disperse, with the modalities intermixed. The Spatial Scan thus reintegrates the previously aligned modalities, distributing them in a manner that reflects their combined influence.", + "bbox": [ + 212, + 257, + 784, + 407 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/02b055069804303b5e24f2058ea3a47655b2a3c38b35a83d3179b7e30772c15b.jpg", + "image_caption": [ + "(a) Input data distribution.", + "Fig.5: Data distribution after Channel Scan and Spacial Scan. Image in red and text data in blue. The Channel Scan tends to aggregate different modalities towards the distribution of textual side. The Spatial Scan reintegrates the previously aligned modalities, distributing them in a manner that reflects their combined influence." + ], + "image_footnote": [], + "bbox": [ + 243, + 434, + 364, + 523 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1543d6f448324b1e8558da17d3d0b12d66dea7473e96bc6a47962de34e2b760f.jpg", + "image_caption": [ + "(b) After Channel Scan." + ], + "image_footnote": [], + "bbox": [ + 439, + 434, + 562, + 522 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a30446d170cd99d08f990f266e2d0557b59c7b68a1bc722635147663ce466922.jpg", + "image_caption": [ + "(c) After Spatial Scan." + ], + "image_footnote": [], + "bbox": [ + 637, + 434, + 759, + 522 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/091bac495cf6387f5aa780b983c519b8071915acf08813f5d9d17a0fdb0f7fce.jpg", + "table_caption": [ + "Table 4: Ablation study on global and local interaction. Both global and local interactions are crucial for modality fusion. Results are evaluated on RefCOCO." + ], + "table_footnote": [], + "table_body": "
ImageGlobalLocalvaltestAtestB
mIoUoIoUPr@50Pr@70mIoUoIoUPr@50Pr@70mIoUoIoUPr@50Pr@70
✓✓-69.166.679.566.371.469.882.370.966.363.774.761.0
✓-✓69.967.980.668.272.270.683.973.166.463.875.361.7
✓✓✓71.668.482.170.973.371.584.874.568.464.577.164.9
", + "bbox": [ + 217, + 700, + 785, + 768 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effects of Global and Local Features. We also ablate the effect of global and local interactions when forming the hybrid feature cube, formally, the $\\tilde{\\mathbf{F}}_t$ and $\\tilde{\\mathbf{F}}_c$ in Eq. (6). The results in Tab. 4 illustrate that the integration of both global and local features significantly enhances the performance.", + "bbox": [ + 212, + 779, + 782, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "ReMamber: Referring Image Segmentation with Mamba Twister", + "bbox": [ + 302, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.6 Visualization Results", + "text_level": 1, + "bbox": [ + 215, + 146, + 437, + 161 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/9fe158de862ee63af9520a2a7fe28128a96fc5a1ecd1b162273f232956b9b989.jpg", + "image_caption": [ + "Fig. 6: Visualization results of our ReMamber and the baseline model LAVT. Our model is able to predict more accurate masks." + ], + "image_footnote": [], + "bbox": [ + 225, + 188, + 789, + 455 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Fig. 6 presents the visualization results of our method compared with the baseline method LAVT [56]. Our ReMamber is capable of producing segmentation results with higher accuracy.", + "bbox": [ + 212, + 500, + 789, + 547 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 565, + 361, + 583 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this study, we introduce ReMamber, a novel architecture utilizing the Mamba framework in Referring Image Segmentation (RIS). It marks a significant advancement in multi-modal understanding. By integrating visual and textual information through innovative Mamba Twister blocks, our approach sets new benchmarks in image-text modality fusion. Achieving competitive results across multiple RIS datasets, our research highlights the potential of Mamba architecture in enhancing the scalability and performance of multi-modal tasks, offering promising directions for future exploration in the field.", + "bbox": [ + 212, + 593, + 789, + 717 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations and Future Works. In our current architecture, the segmentation decoder is constructed by only a few convolutional layers. As shown in Sec. 4.3, the integration of the Cross-Attention mechanism within the Mamba-based architecture demonstrates sub-optimal compatibility, undermining the overall efficacy of our method. In light of these findings, future endeavors will be directed towards the investigation of more sophisticated multi-modal segmentation decoders which best fit Mamba architecture.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 236, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Yang, C. Ma, J. Yao, Z. Zhong, Y. Zhang, and Y. Wang", + "bbox": [ + 271, + 113, + 668, + 130 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Chen, B., Hu, Z., Ji, Z., Bai, J., Zuo, W.: Position-aware contrastive alignment for referring image segmentation. arXiv preprint arXiv:2212.13419 (2022) 12", + "2. Chen, D.J., Jia, S., Lo, Y.C., Chen, H.T., Liu, T.L.: See-through-text grouping for referring image segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7454-7463 (2019) 3, 12", + "3. Chen, Y.W., Tsai, Y.H., Wang, T., Lin, Y.Y., Yang, M.H.: Referring expression object segmentation with caption-aware consistency. arXiv preprint arXiv:1910.04748 (2019) 3, 12", + "4. Ding, H., Liu, C., Wang, S., Jiang, X.: Vision-language transformer and query generation for referring segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16321-16330 (2021) 1, 2, 3", + "5. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 2", + "6. Feng, G., Hu, Z., Zhang, L., Lu, H.: Encoder fusion network with co-attention embedding for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15506-15515 (2021) 3, 12", + "7. Fu, D.Y., Dao, T., Saab, K.K., Thomas, A.W., Rudra, A., Ré, C.: Hungry hungry hippos: Towards language modeling with state space models. arXiv preprint arXiv:2212.14052 (2022) 2, 4", + "8. Gu, A., Dao, T.: Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752 (2023) 2, 4, 5, 7, 9", + "9. Gu, A., Dao, T., Ermon, S., Rudra, A., Ré, C.: Hippo: Recurrent memory with optimal polynomial projections. Advances in neural information processing systems 33, 1474-1487 (2020) 4", + "10. Gu, A., Goel, K., Gupta, A., Ré, C.: On the parameterization and initialization of diagonal state space models. Advances in Neural Information Processing Systems 35, 35971-35983 (2022) 4", + "1. Gu, A., Goel, K., Ré, C.: Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396 (2021) 2, 4", + "2. Gu, A., Johnson, I., Goel, K., Saab, K., Dao, T., Rudra, A., Ré, C.: Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems 34, 572-585 (2021) 4", + "3. Gupta, A., Gu, A., Berant, J.: Diagonal state spaces are as effective as structured state spaces. Advances in Neural Information Processing Systems 35, 22982-22994 (2022) 4", + "4. He, X., Cao, K., Yan, K., Li, R., Xie, C., Zhang, J., Zhou, M.: Pan-mamba: Effective pan-sharpening with state space model. arXiv preprint arXiv:2402.12192 (2024) 2, 4", + "5. Hu, R., Rohrbach, M., Darrell, T.: Segmentation from natural language expressions. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14. pp. 108-124. Springer (2016) 1, 3", + "6. Hu, Z., Feng, G., Sun, J., Zhang, L., Lu, H.: Bi-directional relationship inferring network for referring image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4424-4433 (2020) 3, 12" + ], + "bbox": [ + 225, + 175, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "ReMamber: Referring Image Segmentation with Mamba Twister", + "bbox": [ + 302, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Huang, S., Hui, T., Liu, S., Li, G., Wei, Y., Han, J., Liu, L., Li, B.: Referring image segmentation via cross-modal progressive comprehension. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10488-10497 (2020) 3, 12", + "18. Hui, T., Liu, S., Huang, S., Li, G., Yu, S., Zhang, F., Han, J.: Linguistic structure guided context modeling for referring image segmentation. In: European Conference on Computer Vision. pp. 59-75. Springer (2020) 12", + "19. Islam, M.M., Bertasius, G.: Long movie clip classification with state-space video models. In: European Conference on Computer Vision. pp. 87-104. Springer (2022) 4", + "20. Islam, M.M., Hasan, M., Athrey, K.S., Braskich, T., Bertasius, G.: Efficient movie scene detection using state-space transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18749-18758 (2023) 4", + "21. Jing, Y., Kong, T., Wang, W., Wang, L., Li, L., Tan, T.: Locate then segment: A strong pipeline for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9858-9867 (2021) 3", + "22. Kamath, A., Singh, M., LeCun, Y., Synnaeve, G., Misra, I., Carion, N.: Mdetr-modulated detection for end-to-end multi-modal understanding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1780-1790 (2021) 1, 3", + "23. Kazemzadeh, S., Ordonez, V., Matten, M., Berg, T.: ReferItGame: Referring to objects in photographs of natural scenes. In: Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP) (2014) 9", + "24. Kim, N., Kim, D., Lan, C., Zeng, W., Kwak, S.: Restr: Convolution-free referring image segmentation using transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18145-18154 (2022) 1, 2, 3, 12", + "25. Li, R., Li, K., Kuo, Y.C., Shu, M., Qi, X., Shen, X., Jia, J.: Referring image segmentation via recurrent refinement networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5745-5753 (2018) 1, 3, 12", + "26. Liang, D., Zhou, X., Wang, X., Zhu, X., Xu, W., Zou, Z., Ye, X., Bai, X.: Point-mamba: A simple state space model for point cloud analysis. arXiv preprint arXiv:2402.10739 (2024) 2, 4", + "27. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common Objects in Context. ECCV (2014) 9", + "28. Liu, C., Ding, H., Jiang, X.: Gres: Generalized referring expression segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23592-23601 (2023) 3", + "29. Liu, C., Lin, Z., Shen, X., Yang, J., Lu, X., Yuille, A.: Recurrent multimodal interaction for referring image segmentation. In: Proceedings of the IEEE international conference on computer vision. pp. 1271-1280 (2017) 1, 3, 12", + "30. Liu, J., Ding, H., Cai, Z., Zhang, Y., Satzoda, R.K., Mahadevan, V., Manmatha, R.: Polyformer: Referring image segmentation as sequential polygon generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18653-18663 (2023) 3", + "31. Liu, J., Yang, H., Zhou, H.Y., Xi, Y., Yu, L., Yu, Y., Liang, Y., Shi, G., Zhang, S., Zheng, H., et al.: Swin-umamba: Mamba-based unet with imagenet-based pretraining. arXiv preprint arXiv:2402.03302 (2024) 2, 4" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Yang, C. Ma, J. Yao, Z. Zhong, Y. Zhang, and Y. Wang", + "bbox": [ + 269, + 114, + 666, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Liu, S., Hui, T., Huang, S., Wei, Y., Li, B., Li, G.: Cross-modal progressive comprehension for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(9), 4761-4775 (2021) 12", + "33. Liu, Y., Tian, Y., Zhao, Y., Yu, H., Xie, L., Wang, Y., Ye, Q., Liu, Y.: Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166 (2024) 2, 4, 6, 9", + "34. Luo, G., Zhou, Y., Ji, R., Sun, X., Su, J., Lin, C.W., Tian, Q.: Cascade grouped attention network for referring expression segmentation. Proceedings of the 28th ACM International Conference on Multimedia (2020) 12", + "35. Ma, C., Yang, Y., Ju, C., Zhang, F., Zhang, Y., Wang, Y.: Attrseg: Open-vocabulary semantic segmentation via attribute decomposition-aggregation. NeurIPS (2023) 1", + "36. Ma, C., Yang, Y., Wang, Y., Zhang, Y., Xie, W.: Open-vocabulary semantic segmentation with frozen vision-language models. In: British Machine Vision Conference (2022) 1", + "37. Ma, J., Li, F., Wang, B.: U-mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722 (2024) 2, 4", + "38. Mao, J., Huang, J., Toshev, A., Camburu, O.M., Yuille, A., Murphy, K.: Generation and comprehension of unambiguous object descriptions. CVPR (2015) 9", + "39. Margffoy-Tuay, E., Pérez, J.C., Botero, E., Arbeláez, P.: Dynamic multimodal instance segmentation guided by natural language queries. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 630-645 (2018) 3", + "40. Mehta, H., Gupta, A., Cutkosky, A., Neyshabur, B.: Long range language modeling via gated state spaces. arXiv preprint arXiv:2206.13947 (2022) 4", + "41. Nagaraja, V.K., Morariu, V.I., Davis, L.S.: Modeling context between objects for referring expression understanding. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) ECCV (2016) 9", + "42. Perez, E., Strub, F., de Vries, H., Dumoulin, V., Courville, A.C.: Film: Visual reasoning with a general conditioning layer. In: AAAI (2018) 8", + "43. Qu, M., Wu, Y., Wei, Y., Liu, W., Liang, X., Zhao, Y.: Learning to segment every referring object point by point. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023) 3", + "44. Ruan, J., Xiang, S.: Vm-unet: Vision mamba unet for medical image segmentation. arXiv preprint arXiv:2402.02491 (2024) 2, 4", + "45. Shi, H., Li, H., Meng, F., Wu, Q.: Key-word-aware network for referring expression image segmentation. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 38-54 (2018) 3", + "46. Smith, J.T., Warrington, A., Linderman, S.W.: Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933 (2022) 2, 4", + "47. Tang, J., Zheng, G., Shi, C., Yang, S.: Contrastive grouping with transformer for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23570-23580 (2023) 3", + "48. Tang, J., Zheng, G., Shi, C., Yang, S.: Contrastive grouping with transformer for referring image segmentation. In: CVPR (2023) 9", + "49. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) 2", + "50. Wang, J., Zhu, W., Wang, P., Yu, X., Liu, L., Omar, M., Hamid, R.: Selective structured state-spaces for long-form video understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6387-6397 (2023) 4" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "ReMamber: Referring Image Segmentation with Mamba Twister", + "bbox": [ + 302, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "51. Wang, Z., Lu, Y., Li, Q., Tao, X., Guo, Y., Gong, M., Liu, T.: Cris: Clip-driven referring image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11686-11695 (2022) 1, 3, 12", + "52. Xing, Z., Ye, T., Yang, Y., Liu, G., Zhu, L.: Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. arXiv preprint arXiv:2401.13560 (2024) 2, 4", + "53. Xu, Z., Chen, Z., Zhang, Y., Song, Y., Wan, X., Li, G.: Bridging vision and language encoders: Parameter-efficient tuning for referring image segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17503-17512 (2023) 12", + "54. Yang, S., Xia, M., Li, G., Zhou, H.Y., Yu, Y.: Bottom-up shift and reasoning for referring image segmentation. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11261-11270 (2021) 12", + "55. Yang, Y., Ma, C., Ju, C., Zhang, Y., Wang, Y.: Multi-modal prototypes for open-set semantic segmentation. IJCV (2024) 1", + "56. Yang, Z., Wang, J., Tang, Y., Chen, K., Zhao, H., Torr, P.H.: Lavt: Language-aware vision transformer for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18155-18165 (2022) 1, 2, 3, 9, 11, 12, 14", + "57. Ye, L., Rochan, M., Liu, Z., Wang, Y.: Cross-modal self-attention network for referring image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10502-10511 (2019) 3, 12", + "58. Yu, L., Lin, Z., Shen, X., Yang, J., Lu, X., Bansal, M., Berg, T.L.: Matttnet: Modular attention network for referring expression comprehension. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018) 12", + "59. Zhang, T., Li, X., Yuan, H., Ji, S., Yan, S.: Point could mamba: Point cloud learning via state space model. arXiv preprint arXiv: 2403.00762 (2024) 2, 4", + "60. Zheng, Z., Wu, C.: U-shaped vision mamba for single image dehazing. arXiv preprint arXiv:2402.04139 (2024) 2, 4", + "61. Zhu, C., Zhou, Y., Shen, Y., Luo, G., Pan, X., Lin, M., Chen, C., Cao, L., Sun, X., Ji, R.: Seqtr: A simple yet universal network for visual grounding. In: European Conference on Computer Vision. pp. 598-615. Springer (2022) 3", + "62. Zhu, L., Liao, B., Zhang, Q., Wang, X., Liu, W., Wang, X.: Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417 (2024) 2, 4" + ], + "bbox": [ + 215, + 146, + 784, + 645 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y. Yang, C. Ma, J. Yao, Z. Zhong, Y. Zhang, and Y. Wang", + "bbox": [ + 269, + 113, + 666, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/ReMamber_ Referring Image Segmentation with Mamba Twister/04366454-5dd3-4b5b-8a0b-a5b01b0efbcd_model.json b/2024/ReMamber_ Referring Image Segmentation with Mamba Twister/04366454-5dd3-4b5b-8a0b-a5b01b0efbcd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2125eeb420b49c2453a8f139c599e901822b39f5 --- /dev/null +++ b/2024/ReMamber_ Referring Image Segmentation with Mamba Twister/04366454-5dd3-4b5b-8a0b-a5b01b0efbcd_model.json @@ -0,0 +1,2524 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.227, + 0.141, + 0.78, + 0.187 + ], + "angle": 0, + "content": "ReMamber: Referring Image Segmentation with Mamba Twister" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.211, + 0.78, + 0.252 + ], + "angle": 0, + "content": "Yuhuan Yang\\*1, Chaofan Ma\\*1, Jiangchao Yao1, Zhun Zhong\\*2, Ya Zhang1, and Yanfeng Wang\\*1" + }, + { + "type": "text", + "bbox": [ + 0.39, + 0.261, + 0.612, + 0.277 + ], + "angle": 0, + "content": "1 Shanghai Jiao Tong University" + }, + { + "type": "text", + "bbox": [ + 0.407, + 0.277, + 0.595, + 0.29 + ], + "angle": 0, + "content": "\\(^{2}\\) University of Nottingham" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.291, + 0.761, + 0.305 + ], + "angle": 0, + "content": "{yangyuhuan, chaofanna, sunarker, ya_zhang, wangyanfeng622}@sjtu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.416, + 0.306, + 0.589, + 0.318 + ], + "angle": 0, + "content": "zhunzhong007@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.345, + 0.756, + 0.583 + ], + "angle": 0, + "content": "Abstract. Referring Image Segmentation (RIS) leveraging transformers has achieved great success on the interpretation of complex visual-language tasks. However, the quadratic computation cost makes it resource-consuming in capturing long-range visual-language dependencies. Fortunately, Mamba addresses this with efficient linear complexity in processing. However, directly applying Mamba to multi-modal interactions presents challenges, primarily due to inadequate channel interactions for the effective fusion of multi-modal data. In this paper, we propose ReMamber, a novel RIS architecture that integrates the power of Mamba with a multi-modal Mamba Twister block. The Mamba Twister explicitly models image-text interaction, and fuses textual and visual features through its unique channel and spatial twisting mechanism. We achieve competitive results on three challenging benchmarks with a simple and efficient architecture. Moreover, we conduct thorough analyses of ReMamber and discuss other fusion designs using Mamba. These provide valuable perspectives for future research. The code has been released at: https://github.com/yyh-rain-song/ReMamber." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.593, + 0.744, + 0.622 + ], + "angle": 0, + "content": "Keywords: Referring Image Segmentation (RIS) \\(\\cdot\\) Multi-Modal Understanding \\(\\cdot\\) Mamba Architecture" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.662, + 0.378, + 0.679 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.764 + ], + "angle": 0, + "content": "Referring Image Segmentation (RIS) is a crucial yet challenging task in the area of multi-modal understanding [15, 25, 29]. Unlike ordinary image segmentation, RIS involves the identification and segmentation of specific objects in image based on the textual descriptions. This thereby requires the model to be capable of understanding vision-language interactions, which is the core challenge of RIS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Thanks to the powerful attention mechanisms, it has achieved great success for RIS by leveraging transformers to promote the exact recognition of multimodal information. For example, existing works have designed transformer decoder [4,22] or transformer encoder-decoder [24,35,36,51,55,56] to comprehensively fuse visual and linguistic features, which have achieved great progress." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.668, + 0.13 + ], + "angle": 0, + "content": "Y. Yang, C. Ma, J. Yao, Z. Zhong, Y. Zhang, and Y. Wang" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.785, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.381, + 0.788, + 0.452 + ], + "angle": 0, + "content": "Fig. 1: We propose ReMamber, a novel referring segmentation architecture with Mamba twister. It consists of several Mamba Twister block. Each block contains several visual state space (VSS) layers and a Twisting layer. The Twisting layer first calculates the interaction between image and text, and then forms a hybrid feature cube. Finally, it \"twists\" the feature cube using the Channel and Spatial Scan along each dimension." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.471, + 0.788, + 0.547 + ], + "angle": 0, + "content": "Nonetheless, it is a quadratic increase in both computation and memory when applying full attention in transformers [5, 49]. This leads to the limitation for resource-intensive scenarios, for example, in capturing long-range visual-language dependencies. And this is particularly important in the context of large-size images with long textual descriptions." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.55, + 0.788, + 0.732 + ], + "angle": 0, + "content": "Fortunately, recent advances of State Space Models (SSMs) [7,8,11,46] has emerged as promising architectures for solving the above issue. Specifically, Mamba [8] marks a significant advancement for efficient training and inference with linear complexity, which has been incorporated into various visual tasks. However, current efforts primarily pay attention to single-modality settings, such as image classification [33,62], biomedical image segmentation [31,37,44,52], low-level vision [14,60], and point cloud analysis [26,59]. In this paper, we pioneer the exploration of Mamba in multi-modal RIS setting, and identify that the prevalent multi-modality token splicing method in transformers [4,24,56] is no longer effective. Since in Mamba, a fundamental deficiency exists: the interactions are insufficient between channels of different tokens [8], which adversely affects the fusion of multi-modal information in RIS." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To address this dilemma, we propose ReMamber, a novel referring segmentation architecture with Mamba twister. As shown in Fig. 1, ReMamber comprises several Mamba Twister blocks, which allow the model to be \"aware\" of the textual context at every spatial location. Each block consists of several visual state space (VSS) layers and a Twisting layer. Specifically, the VSS layers initially extract visual features, and the Twisting layer injects the textual information into the visual modality. The Twisting layer is structured into three critical" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.303, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ReMamber: Referring Image Segmentation with Mamba Twister" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.344 + ], + "angle": 0, + "content": "components. (1) A vision-language interaction operation is designed to explicitly capture the fine-grained interactions between modalities. This is achieved by evaluating the similarity between visual and textual tokens then mapping them into a shared feature space, thus generating a multi-modal feature. (2) A hybrid feature cube is then created by concatenating the visual feature, multimodal feature, and global textual feature. This ensures that each visual token receives uniform influence from both local and global contexts, preventing the overshadowing of subtle textual cues by predominant visual features. (3) Lastly, to address the inadequate interaction within channels in Mamba, a twisting mechanism is deployed. This process \"twists\" the feature cube channel- and spatial-wise, thereby enhancing interaction within and across modalities. It is accomplished through two consecutive SSMs scanning along channel and spatial dimensions, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.786, + 0.36 + ], + "angle": 0, + "content": "To sum up, the key contributions of our research are highlighted as follows:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.784, + 0.389 + ], + "angle": 0, + "content": "- We pioneer the exploration of Mamba in referring image segmentation (RIS), demonstrating Mamba's significant potential for multi-modal understanding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.785, + 0.434 + ], + "angle": 0, + "content": "- We design a novel framework, ReMamber, that mainly contains several Mamba Twister blocks. This design effectively captures vision-language interactions using the \"twisting mechanism\"." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.785, + 0.495 + ], + "angle": 0, + "content": "- We achieve competitive results on three challenging benchmarks. Moreover, we conduct thorough analyses of ReMamber and discuss other vision-language fusion designs using Mamba. These provide valuable perspectives for future research." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.36, + 0.785, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.522, + 0.388, + 0.537 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.556, + 0.513, + 0.571 + ], + "angle": 0, + "content": "2.1 Referring Image Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Referring image segmentation (RIS) aims to segment the entities in the image following natural language instruction. Early approaches [15, 25, 29, 39] leveraged RNNs or LSTMs to encode linguistic representations, and CNNs to extract spatial features from the image at varying levels. These disparate modalities were then integrated using the multi-modal LSTM [2, 15, 21, 25, 29, 39], attention mechanism [6, 16, 45, 57], cycle-consistency [3], and graph convolution [17]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Recent trends have shifted towards leveraging transformers for enhanced capturing and fusion of vision-language modalities. MDETR [22] and VLT [4] design a transformer decoder for fusing linguistic and visual features. LAVT [56] adopts Swin Transformer as the visual backbone and incorporate vision-language fusion modules within the visual encoder's final layers. Similar strategies are employed by ReSTR [24] and CRIS [51], which utilize dual transformer encoders for initial modality encoding, followed by feature fusion through a multi-modal transformer encoder or decoder. Other models like PolyFormer [30], SeqTr [61] and [43] also adopt a multi-modal transformer for vision-language fusion, but they output masks as sequences of contour points. Meanwhile, GRES [28] and CGFormer [47] consider transformer queries as region proposals, and regarding" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.114, + 0.668, + 0.131 + ], + "angle": 0, + "content": "Y. Yang, C. Ma, J. Yao, Z. Zhong, Y. Zhang, and Y. Wang" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.209 + ], + "angle": 0, + "content": "segmentations as proposal-level classification problems. Distinct from all existing works, we introduce a pioneering multi-modal architecture named ReMamber, based on Mamba [8]. This novel approach underscores the untapped potential of Mamba in advancing the field of referring image segmentation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.23, + 0.629, + 0.246 + ], + "angle": 0, + "content": "2.2 State Space Models and Visual Applications" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.255, + 0.788, + 0.436 + ], + "angle": 0, + "content": "State space models (SSMs), originally derived from control theory, have been effectively combined with deep learning to model the long-range dependencies. Early works like LSSL [12] show potential when combined with HiPPO [9] initialization. S4 [11] was designed to diminish both computational and memory demands associated with state representations. It scales linearly with sequence length, offering a notable advantage over CNNs and transformers. Building on S4, S5 [46] introduces the efficient parallel scan and MIMO SSM, further refining the approach. Later, H3 [7] expanded on these foundations, achieving competitive performance with transformers in language modeling. Recently, Mamba [8] marked a significant advancement with its linear-time inference and efficient training, incorporating a selection mechanism and hardware-aware algorithms building upon prior works [10, 13, 40]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.437, + 0.789, + 0.634 + ], + "angle": 0, + "content": "With the demonstrated success of SSMs like Mamba in language modeling, researchers have begun to investigate their applicability to visual tasks. Works such as ViS4mer [19], Selective S4 [50] and TranS4mer [20] directly use the S4's ability to capture long-range sequences for understanding inter-frame relations in video classification and detection. However, they still employ Vision Transformers for intra-frame feature extraction. More recently, Vim [62] and VMamba [33] have shown promising results as fully Mamba-based vision backbones for image classification, detection, and segmentation. At the same time, several studies have explored Mamba-based architecture on various vision tasks, such as biomedical image segmentation [31, 37, 44, 52], low-level vision [14, 60], and point cloud analysis [26, 59]. Nevertheless, all above efforts primarily focus on single-modality vision tasks. In this paper, we pioneer the exploration of Mamba-base architecture in multi-modality RIS task." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.656, + 0.341, + 0.673 + ], + "angle": 0, + "content": "3 Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.788, + 0.749 + ], + "angle": 0, + "content": "This paper aims to develop a multi-modal Mamba-based architecture for RIS task. In Sec. 3.1, we give the preliminary. From Sec. 3.2 to Sec. 3.5, we describe our architecture from coarse to fine. In Sec. 3.6, we introduce other three variants besides our design. And in Sec. 3.7 we detail the training process." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.77, + 0.528, + 0.787 + ], + "angle": 0, + "content": "3.1 Preliminary: State Space Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.843 + ], + "angle": 0, + "content": "State Space Model (SSM) is a sequence model inspired by continuous systems. It is designed to capture a mapping relationship between two functions or sequences, expressed as \\( x(t) \\in \\mathbb{R} \\mapsto y(t) \\in \\mathbb{R} \\), through a hidden state \\( h(t) \\in \\mathbb{R}^N \\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.303, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "ReMamber: Referring Image Segmentation with Mamba Twister" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.195 + ], + "angle": 0, + "content": "The evolution of the hidden state over time is governed by specific parameters \\(\\mathbf{A} \\in \\mathbb{R}^{N \\times N}\\), which directs the state evolution, and \\(\\mathbf{B} \\in \\mathbb{R}^N\\), \\(\\mathbf{C} \\in \\mathbb{R}^N\\), which performs the projection. Formally, SSM can be described by the following equations:" + }, + { + "type": "equation", + "bbox": [ + 0.421, + 0.205, + 0.786, + 0.229 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} h ^ {\\prime} (t) = \\mathbf {A} h (t) + \\mathbf {B} x (t), \\\\ (t) = \\mathbf {C} h ^ {\\prime} (t) \\end{array} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.425, + 0.223, + 0.522, + 0.239 + ], + "angle": 0, + "content": "\\[\ny (t) = \\mathbf {C} h (t).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.243, + 0.784, + 0.3 + ], + "angle": 0, + "content": "Discretizing this system introduces a timescale parameter \\(\\Delta t\\), which transforms the continuous parameters \\(\\mathbf{A}, \\mathbf{B}\\) into their discrete counterparts \\(\\overline{\\mathbf{A}}, \\overline{\\mathbf{B}}\\). A common way for this transformation is the zero-order hold (ZOH) approach, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.3, + 0.785, + 0.327 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\overline {{\\mathbf {A}}} = \\exp (\\boldsymbol {\\Delta} \\mathbf {A}), \\\\ \\overline {{\\mathbf {B}}} = (\\boldsymbol {\\Delta} \\mathbf {A}) ^ {- 1} (\\boldsymbol {\\Delta} \\mathbf {A}) - \\mathbf {I}) - \\boldsymbol {\\Delta} \\mathbf {B}. \\end{array} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.378, + 0.312, + 0.785, + 0.337 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\mathbf {B}}} = (\\boldsymbol {\\Delta} \\mathbf {A}) ^ {- 1} (\\exp (\\boldsymbol {\\Delta} \\mathbf {A}) - \\mathbf {I}) \\cdot \\boldsymbol {\\Delta} \\mathbf {B}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.342, + 0.694, + 0.356 + ], + "angle": 0, + "content": "Then, the discretized version of this system can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.434, + 0.366, + 0.785, + 0.391 + ], + "angle": 0, + "content": "\\[\nh _ {t} = \\overline {{\\mathbf {A}}} h _ {t - 1} + \\overline {{\\mathbf {B}}} x _ {t}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.435, + 0.385, + 0.504, + 0.399 + ], + "angle": 0, + "content": "\\[\ny _ {t} = \\mathbf {C} h _ {t}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.409, + 0.784, + 0.455 + ], + "angle": 0, + "content": "Mamba [8], as a variant of SSM, releases the linearity constraint of the original SSM described in Eq. (1), by making \\(\\mathbf{B}\\) and \\(\\mathbf{C}\\) to be input-dependent, whereas still maintaining the linear complexity during forward process." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.474, + 0.45, + 0.488 + ], + "angle": 0, + "content": "3.2 Architecture Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.496, + 0.788, + 0.646 + ], + "angle": 0, + "content": "The pivotal aspect of Referring Image Segmentation (RIS) is the correspondence between image and textual input. To achieve this, we propose ReMamber. Fig. 2 shows the overview of our architecture. The basic block for ReMamber is the Mamba Twister block. It is a multi-modal fusion block that takes both visual and textual feature as input, and outputs the fused multi-modal feature representation. We extract the intermediate feature after each Mamba Twister block, and then feed it into a flexible decoder to generate the final segmentation mask. The decoder is task-invariant, so any downstream architecture can be applied. Below in Sec. 3.3, we will first introduce the key component: Mamba Twister block." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.667, + 0.445, + 0.68 + ], + "angle": 0, + "content": "3.3 Mamba Twister Block" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.841 + ], + "angle": 0, + "content": "As shown in Fig. 2, the Mamba Twister Block consists of several visual state space (VSS) layers and a Twisting layer. The VSS layer is designed to process features in the spatial domain. It treats the input feature as a series of image tokens and aims to discern the spatial relationships among these tokens. The Twisting layer aims to inject text condition into the image feature, thus guiding the transformation of image feature. It begins by condensing the textual sequence to compute the cross-correlation between the image and textual modalities. Subsequently, it disseminates this information across each image feature patch via a twisting operation. We'll detail the introduction of VSS layer in Sec. 3.4 and the Twisting layer in Sec. 3.5." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.668, + 0.131 + ], + "angle": 0, + "content": "Y. Yang, C. Ma, J. Yao, Z. Zhong, Y. Zhang, and Y. Wang" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.144, + 0.789, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.291, + 0.744, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.418, + 0.789, + 0.504 + ], + "angle": 0, + "content": "Fig. 2: Overview architecture of our ReMamber. The basic block for ReMamber is the Mamba Twister block. It consists of several visual state space (VSS) layers and a Twisting layer. The Twisting layer first constructs hybrid feature cube from text, image, and multi-modal features via channel concatenation. Then, it \"twists\" the cube by Channel Scan and Spatial Scan. We extract intermediate features after each Mamba Twister block, and feed it into a flexible decoder for final segmentation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.513, + 0.584, + 0.53 + ], + "angle": 0, + "content": "3.4 VSS Layer for Spatial Data Processing" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.789, + 0.658 + ], + "angle": 0, + "content": "Since the SSM is initially designed for processing temporal or causal data, which can not effectively process non-causal data types, i.e., 2-dimensional image in our case. To solve this issue, we adopt the Cross-Scan-Module proposed in [33], which unfolds image patches into sequences and then scans them in four distinct directions, ensuring that the information from all pixels is integrated during feature transformation. We replace the scanning operation in vanilla Mamba block with CSM, forming a visual state space (VSS) layer for our spatial feature transformation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.68, + 0.584, + 0.696 + ], + "angle": 0, + "content": "3.5 Twisting Layer for Multi-modal Fusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Here we introduce the detailed design of our Twisting layer. Intuitively, the Twisting layer first constructs a feature cube by arranging information from different modalities orderly. Then, pieces of information intertwine with each other when the cube is twisted through two SSM layers along different axis, thereby achieving modality fusion. Formally, given image feature \\(\\mathbf{F}_i\\in \\mathbb{R}^{h\\times w\\times C_i}\\) and text feature \\(\\mathbf{F}_t\\in \\mathbb{R}^{L\\times C_t}\\), the Mamba twister aims to learn a function \\((\\mathbf{F}_i,\\mathbf{F}_t)\\mapsto \\tilde{\\mathbf{F}}\\in \\mathbb{R}^{h\\times w\\times C_o}\\) by first forming a hybrid multi-modal feature cube and then conduct two SSM scans on it. \\(C_i,C_t\\) and \\(C_o\\) are the dimensions of visual, textual, and output feature; \\(h,w\\) and \\(L\\) are height, weight, and length." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.303, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "ReMamber: Referring Image Segmentation with Mamba Twister" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "Forming the Hybrid Feature Cube. To explicitly build the image-text correspondence, we design a novel vision-language interaction operation and then form a hybrid feature cube. This allows more effective modality fusion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.284 + ], + "angle": 0, + "content": "To be specific, we formulated the vision-language interaction operation into two manners: global interaction and local interaction. The global interaction treats the text sequence as a whole, which means that all image patches should be aware of the semantic meaning of the text expression. We pool a global representation \\(\\mathbf{F}_t^{\\mathrm{CLS}}\\in \\mathbb{R}^{C_t}\\) from the text sequence \\(\\mathbf{F}_t\\) to convey this information. The \\(\\mathbf{F}_t^{\\mathrm{CLS}}\\) is further expanded to the same size as the image feature. Formally:" + }, + { + "type": "equation", + "bbox": [ + 0.385, + 0.291, + 0.785, + 0.309 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {F}} _ {t} = \\operatorname {E x p a n d} \\left(\\mathbf {F} _ {t} ^ {\\mathrm {C L S}}\\right) \\in \\mathbb {R} ^ {h \\times w \\times C _ {t}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.317, + 0.784, + 0.347 + ], + "angle": 0, + "content": "where \\(\\operatorname{Expand}(\\cdot)\\) denotes the operation that expands the input tensor to the same size as the image feature." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.348, + 0.785, + 0.435 + ], + "angle": 0, + "content": "However, a global representation may not be enough to capture the intricate relationships between different modalities. For example, some words such as color, shape or location may be more relevant to certain image patches than others. The local interaction aims to capture such correlation in a fine-grained level. Formally, we calculate the local interaction map \\(\\mathbf{F}_c\\) using matrix multiplication:" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.436, + 0.785, + 0.453 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {c} = \\mathbf {F} _ {i} \\mathbf {W} _ {i} \\cdot \\left(\\mathbf {F} _ {t} \\mathbf {W} _ {t}\\right) ^ {T} \\in \\mathbb {R} ^ {h \\times w \\times L}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.785, + 0.519 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_i\\in \\mathbb{R}^{C_i\\times C_c}\\) and \\(\\mathbf{W}_t\\in \\mathbb{R}^{C_t\\times C_c}\\) are learnable parameters. To enhance feature processing in high dimension, we use a single convolutional layer to transform \\(\\mathbf{F}_c\\) into \\(\\tilde{\\mathbf{F}}_c\\in \\mathbb{R}^{h\\times w\\times C_c}\\). We then concatenate \\(\\mathbf{F}_i\\), \\(\\tilde{\\mathbf{F}}_t\\), and \\(\\tilde{\\mathbf{F}}_c\\) along the channel dimension to form the hybrid feature cube. Formally:" + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.527, + 0.785, + 0.544 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {\\text {c u b e}} = \\left[ \\mathbf {F} _ {i}, \\tilde {\\mathbf {F}} _ {t}, \\tilde {\\mathbf {F}} _ {c} \\right] \\in \\mathbb {R} ^ {h \\times w \\times \\left(C _ {i} + C _ {t} + C _ {c}\\right)}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.563, + 0.784, + 0.623 + ], + "angle": 0, + "content": "Twisting the Hybrid Feature Cube. Generally, the scanning operation for vanilla SSM is almost independent across channels [8]. This is not enough for feature communication, especially when information from different modalities is arranged along the channel dimension, i.e., our hybrid feature cube." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.785, + 0.742 + ], + "angle": 0, + "content": "To foster the interactions within and across modalities, here we design a twisting mechanism to \"twist\" the hybrid feature cube along different axis. Specifically, it is composed of the Channel Scan and Spacial Scan successively. The Channel Scan first treats the channel as an ordered sequence, and learns to communicate cross channels, thus foster the modality fusion. Then, the Spatial Scan operates on the spatial dimension, and learns to communicate information cross patches within each modality separately. Formally, this process can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.363, + 0.745, + 0.785, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {\\text {o u t}} = \\operatorname {S S M} _ {\\text {s p a t i a l}} \\left(\\operatorname {S S M} _ {\\text {c h a n n e l}} \\left(\\mathbf {F} _ {\\text {c u b e}}\\right)\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.784, + 0.841 + ], + "angle": 0, + "content": "where \\(\\mathrm{SSM}_{\\mathrm{channel}}\\) refers to the Channel Scan. It is an SSM layer that treats the concatenated feature as an ordered sequence through channel dimension, and presents 1-D selective scan along the channel. While \\(\\mathrm{SSM}_{\\mathrm{spatial}}\\) denotes the Spatial Scan. It is an VSS layer and presents 2D selective scan along the two spatial dimension. \\(\\mathbf{F}_{\\mathrm{out}}\\) is then feed into the next layer for further process." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.114, + 0.669, + 0.131 + ], + "angle": 0, + "content": "Y. Yang, C. Ma, J. Yao, Z. Zhong, Y. Zhang, and Y. Wang" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.144, + 0.658, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.174, + 0.775, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.362, + 0.789, + 0.419 + ], + "angle": 0, + "content": "Fig.3: Other multi-modal fusion designs. (a) In-context Conditioning appends text tokens ahead of image tokens. (b) Attention-based Conditioning utilizes cross-attention mechanism for modality fusion. (c) Norm Adaptation learns a scale and bias for the model's normalization layers." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.431, + 0.715, + 0.447 + ], + "angle": 0, + "content": "3.6 Discussion: Other Variants of Modality Fusion Designs" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.789, + 0.517 + ], + "angle": 0, + "content": "Besides the proposed Twisting layer, we also explore other variants of multimodal fusion designs. Fig. 3 shows the structure of three other different variants. We here discuss these variants, and later in Sec. 4.3, we analyze the results for these variants." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.789, + 0.594 + ], + "angle": 0, + "content": "In-context Conditioning. As shown in Fig. 3 (a), in this variant, we simply append the text sequence before the image feature, forming a longer sequence than originally being processed. In this way, the model is able to fuse image features with the previous text as a context. This is the most straightforward way to allow the model aware of the text condition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.789, + 0.669 + ], + "angle": 0, + "content": "Attention-based Conditioning. Attention mechanism is usually a strong baseline for sequence modeling. Here we utilize the cross-attention mechanism for two modality fusion, as shown in Fig. 3 (b), To be specific, we use the image feature as query, and the text feature as key and value, hoping to capture the correlation between image and text." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.67, + 0.789, + 0.731 + ], + "angle": 0, + "content": "Norm Adaptation. In this variant, we integrate the text input by adapting the scale and bias after norm layer using FiLM [42]. We first pool a global representation from text sequence, and then use a linear projection to transform it into the scale and shift in layer normalization." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.754, + 0.389, + 0.77 + ], + "angle": 0, + "content": "3.7 Model training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.84 + ], + "angle": 0, + "content": "The proposed ReMamber is a general framework for multi-modal fusion. Therefore, there is no restriction on the downstream decoder design nor the loss function. Here we use a simple convolution-based decoder for simplicity. The entire network is trained in an end-to-end manner." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.303, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ReMamber: Referring Image Segmentation with Mamba Twister" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.377, + 0.163 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.176, + 0.44, + 0.19 + ], + "angle": 0, + "content": "4.1 Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.2, + 0.788, + 0.365 + ], + "angle": 0, + "content": "Dataset. To assess the efficacy of our proposed approach, we executed experiments across three widely recognized datasets tailored for the referring image segmentation (RIS) task: RefCOCO, RefCOCO+ [23], and G-Ref [23,38,41]. RefCOCO [23] is a prominent dataset in the RIS domain, comprising 19,994 images and 142,210 referring expressions associated with 50,000 objects derived from the MSCOCO [27] dataset through a two-player game. RefCOCO+ enhances the challenge by excluding expressions with absolute location references, featuring 141,564 expressions for 49,856 objects across 19,992 images. G-Ref [38,41] enriches the dataset diversity with 104,560 referring expressions for 54,882 objects in 26,711 images, showcasing an average sentence length of 8.4 words with a heightened focus on location and appearance descriptions." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.366, + 0.787, + 0.427 + ], + "angle": 0, + "content": "Evaluation Metrics. In line with preceding studies [48], we adopt mIoU and oIoU as the main metrics. Furthermore, we incorporate the metric Precision@X \\((\\mathrm{X} \\in \\{50, 60, 70, 90\\})\\) for a more comprehensive evaluation, whereas Percision@X means the percentage of test images with an IoU score high than X%." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.447, + 0.457, + 0.463 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.471, + 0.788, + 0.577 + ], + "angle": 0, + "content": "Our model architecture is developed upon the foundations of Mamba [8] and VMamba [33]. We adopt ImageNet pre-trained weights as initialization, and train the model in an end-to-end manner. We set the input image resolution to 480 when compared with the state-of-the-art methods in Tab. 2, following [48,56] for fair comparison. For other experiments (Tabs. 1 and 3 to 4), we set the image resolution to 256 for faster training and ablation study without changing its property. For details about the training settings, please refer to our code.\\(^{1}\\)" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.598, + 0.737, + 0.614 + ], + "angle": 0, + "content": "4.3 Evaluation on Other Variants of Modality Fusion Designs" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.621, + 0.788, + 0.757 + ], + "angle": 0, + "content": "As discussed in Sec. 3.6, besides our Mamba Twister, we also provide three other variant architectures for modality fusion, including In-context Conditioning, Attention-based Conditioning and Norm Adaptation. To make a fair comparison, for all architectures, we initialize their common part (VMamba-based parameters) with the same checkpoint. The difference is that the Twisting layer is replaced by the corresponding fusion methods. The parameters of fusion modules are nearly equal, i.e., {103, 113, 118, 116} KB for {In-Context, Attention, Norm Adaptation, Twister}. Here we provide a comprehensive analysis between these architectures, showing the advantages and disadvantages of each variant." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.758, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Tab. 1 shows the comparison result. The Mamba Twister consistently outperforms the other variants across all metrics and datasets, indicating its superior capability in capturing and integrating contextual information for more accurate segmentation." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.824, + 0.542, + 0.841 + ], + "angle": 0, + "content": "1 https://github.com/yyh-rain-song/ReMamber" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.114, + 0.668, + 0.131 + ], + "angle": 0, + "content": "Y. Yang, C. Ma, J. Yao, Z. Zhong, Y. Zhang, and Y. Wang" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 1: Comparison with other modality fusion variants. \"Attention\" means for \"Attention-based Conditioning\", \"In-Context\" for \"In-Context Conditioning\", and \"Adaptation\" for \"Norm Adaptation\". Mamba Twister steadily outperforms other variants across all metrics and datasets, indicating its superior capability in capturing and integrating contextual information for more accurate segmentation." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.226, + 0.789, + 0.523 + ], + "angle": 0, + "content": "
Dataset: RefCOCO
VariantsmIoUvalPr@70mIoUtestAmIoUtestB
oIoUPr@50Pr@70oIoUPr@50Pr@70oIoUPr@50
Attention65.362.375.160.667.564.479.165.461.758.069.653.4
In-Context69.165.979.767.571.268.982.271.866.262.875.061.7
Adaptation70.267.080.769.672.370.283.173.266.862.875.163.1
Mamba Twister71.668.482.170.973.371.584.874.568.464.577.164.9
Dataset: RefCOCO+
VariantsmIoUvalPr@70mIoUtestAmIoUtestB
oIoUPr@50oIoUPr@50Pr@70oIoUPr@50
Attention54.049.760.745.558.755.868.552.946.941.950.636.6
In-Context58.454.066.353.963.259.472.260.551.547.556.444.5
Adaptation60.355.268.257.364.259.973.462.953.948.759.047.2
Mamba Twister61.657.370.058.565.862.175.564.954.049.959.447.8
Dataset: G-Ref
VariantsmIoUvalPr@70Pr@90mIoUoIoUtest
oIoUPr@50Pr@50Pr@60Pr@70Pr@90
Attention50.549.254.246.036.78.150.850.857.349.838.66.4
In-Context54.853.459.552.343.413.355.554.860.753.645.313.3
Adaptation59.356.766.761.053.718.458.656.665.559.852.818.2
Mamba Twister61.158.068.462.855.218.561.259.069.063.455.518.7
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.79, + 0.702 + ], + "angle": 0, + "content": "Surprisingly, despite widely used in transformers, the Attention-based Conditioning performs poorly in our task. This suggests that the cross-attention mechanism is not inherently suitable for Mamba architecture. This may be due to a fundamental discrepancy between the two systems. To be specific, (1) the Mamba model is predicted on ordered sequences that exhibit strict sequential dependencies, where the state at a given time point \\( t + 1 \\) is determined by the preceding time point \\( t \\) and a hidden state; (2) In contrast, cross-attention mechanisms treats all tokens within a sequence equally. This discrepancy may undermine the Mamba model's ability to structurally model sequences, as the cross-attention mechanism does not preserve the sequential integrity and the hierarchical dependencies essential for the model's operation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.703, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Serving as a straightforward baseline, In-context Conditioning performs sub-optimal. This may because it models image-text interaction in an implicit way. In our situation, the length of the image feature is much larger than the textual side, textual information may be diluted during the forward process. Besides, this operation does not differentiate the textual information between image patches, and would be insufficient for multi-modal fusion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Finally, Norm Adaptation with explicit image-text interaction modeling appears to be a strong baseline. It achieves better performance than other two variants. However, when calculating scale and bias, it uses only a pooled vector as" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.303, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "ReMamber: Referring Image Segmentation with Mamba Twister" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.178 + ], + "angle": 0, + "content": "textual representation instead of the entire sequence. This potentially results in the loss of information, making it less effective compared to our Mamba Twister." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.2, + 0.79, + 0.382 + ], + "angle": 0, + "content": "Case-study using Attention Maps. In Fig. 4, we further visualized the attention maps in Attention-based Conditioning as well as the local interaction map in Mamba Twister defined in Eq. (5). The four images in Fig. 4(b) are arranged from left to right with the network going from shallow to deep. It can be observed that in the shallower layers, Mamba Twister focuses more on the lower-level features of the images such as edges. As the layer goes deeper, the cost map mostly concentrating on the target object, indicating that Mamba Twister can progressively guide the image feature towards the target described by the language. In contrast, though the Attention-based Conditioning variant is able to locate the target, its attention map performs poorly. This may suggest that the cross-attention mechanism struggles to accurately capture the correct context information, and the two modalities are not truly fused together." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.405, + 0.789, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.589, + 0.789, + 0.66 + ], + "angle": 0, + "content": "Fig. 4: Cross-Attention map (up) and our local interaction map (down) comparison. Though both methods are able to predict target correctly, the cross-attention maps don't show correct image-text correlation, while ours are able to capture this relationship accurately, indicating that Mamba Twister is able to gradually fusing the two modality." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.694, + 0.585, + 0.71 + ], + "angle": 0, + "content": "4.4 Comparison with the State-of-the-Arts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Tab. 2 shows the comparison of our ReMamber with state-of-the-art methods on RefCOCO, RefCOCO+, and G-Ref datasets. ReMamber outperforms all other methods across all datasets, demonstrating its superior capability in capturing and integrating contextual information for more accurate segmentation. In particular, Mamba outperforms previous Swin-based methods such as LAVT [56] by a considerable margin, indicating the remarkable capability of the Mamba-based architecture in segmentation tasks. and the efficiency of the newly introduced Mamba Twister." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.114, + 0.668, + 0.131 + ], + "angle": 0, + "content": "Y. Yang, C. Ma, J. Yao, Z. Zhong, Y. Zhang, and Y. Wang" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.787, + 0.174 + ], + "angle": 0, + "content": "Table 2: Comparison with state-of-the-art methods on RefCOCO, RefCOCO+, and G-Ref. The metric of oIoU is reported. Best results are in bold." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.186, + 0.789, + 0.435 + ], + "angle": 0, + "content": "
ModelBackboneRefCOCORefCOCO+G-Ref
valtestAtestBvaltestAtestBvalUtestU
PCAN [1]ResNet-5069.5171.6464.1858.2563.6848.8959.9860.8
MAAttNet [58]MaskRCNN ResNet-10156.5162.3751.746.6752.3940.0847.6448.61
RMI [29]DeepLab ResNet-10145.1845.6945.5729.8630.4829.50--
RRN [25]DeepLab ResNet-10155.3357.2653.9539.7542.1536.11--
CMSA [57]DeepLab ResNet-10158.3260.6155.0943.7647.637.89--
CAC [3]DeepLab ResNet-10158.9061.7753.81---46.3746.95
STEP [2]DeepLab ResNet-10160.0463.4657.9748.1952.3340.41--
BRINet [16]DeepLab ResNet-10160.9862.9959.2148.1752.3242.11--
CMPC [17]DeepLab ResNet-10161.3664.5359.6449.5653.4443.23--
LSCM [18]DeepLab ResNet-10161.4764.9959.5549.3453.1243.50--
CMPC+ [32]DeepLab ResNet-10162.4765.0860.8250.2554.0443.47--
BUSNet [54]DeepLab ResNet-10163.2766.4161.3951.7656.8744.13--
CGAN [34]DeepLab ResNet-10164.8668.0462.0751.0355.5144.0651.0151.69
EFN [6]Wide ResNet-10162.7665.6959.6751.5055.2443.01--
ETRIS [53]CLIP ResNet-10171.0674.1166.6662.2368.5152.7960.2860.42
ReSTR [24]ViT-B-1667.2269.364.4555.7860.4448.2754.48-
ETRIS [53]ViT-B-1670.5173.5166.6360.1066.8950.1759.8259.91
CRIS [51]CLIP ResNet-5069.5272.7264.761.3967.152.4859.8760.36
CRIS [51]CLIP ResNet-10170.4773.1866.1062.2768.0853.6859.8760.36
LAVT [56]Swin-B72.7375.8268.7962.1468.3855.1061.2462.10
ReMamber (Ours)Mamba-B74.5476.7470.8965.0070.7857.5363.964.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.449, + 0.788, + 0.507 + ], + "angle": 0, + "content": "Table 3: Ablation on the combination of two scans. We ablated each scan separately, as well as combining them parallel or swap the order. The occurrence of Spatial Scan affects most to the performance. The three variants perform similarly, with the combination of Spatial-Channel slightly better. Results are evaluated on RefCOCO." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.517, + 0.789, + 0.603 + ], + "angle": 0, + "content": "
Channel ScanSpatial ScanvaltestAtestB
mIoUoIoUPr@50Pr@70mIoUoIoUPr@50Pr@70mIoUoIoUPr@50Pr@70
✓-62.360.170.757.264.963.274.361.959.757.566.351.9
-✓70.065.880.569.172.371.083.773.467.563.575.262.7
Parallel71.068.881.970.073.171.884.774.167.964.676.964.5
Spatial-Channel71.468.882.470.873.371.484.974.267.964.776.564.4
Channel-Spatial71.668.482.170.973.371.584.874.568.464.577.164.9
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.615, + 0.39, + 0.631 + ], + "angle": 0, + "content": "4.5 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.787, + 0.779 + ], + "angle": 0, + "content": "Effects of two Scans. Tab. 3 presents a comparative analysis of different scan variants. We provide outcomes of conducting Channel Scan and Spatial Scan independently. Additionally, results from various combinations of these two scans are discussed. The term \"Parallel\" denotes the concurrent execution of both scans, followed by adding the independent output for fusion. \"Channel-Spatial\" refers to the original scanning order in Mamba Twister shown in Fig. 2, where the Channel Scan is executed first, followed by the Spatial Scan. \"Spatial-Channel\" refers to reversing the sequence of the two scans, starting with Spatial Scan and then proceeding to Channel Scan." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Tab. 3 indicates that utilizing any single scan on its own yields suboptimal results, with the use of Channel Scan alone experiencing the most significant drop in effectiveness. This decline may indicate that a pure Channel Scan alters the data distribution, adversely affecting network stability. Among the three" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.303, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ReMamber: Referring Image Segmentation with Mamba Twister" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "hybrid scanning policies, each has its strengths and weaknesses. Overall, the combination of Channel-Spatial Scan appears to offer a considerable advantage." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.197, + 0.784, + 0.257 + ], + "angle": 0, + "content": "Distribution Visualization. The visualization of the distribution of the multi-modal input data is depicted in Fig. 5, where we use PCA to map the feature to a 3D space. The image and text data are indicated in red and blue respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.258, + 0.785, + 0.409 + ], + "angle": 0, + "content": "Fig. 5a illustrates the distribution of the input data. We can observe that the text data forms a linear arrangement within the 3D space, as it is duplicated \\( H \\times W \\) times to align with the image size. This suggests that the text instruction should be applied for each pixel within the image. Fig. 5b reveals the data distribution following the application of a Channel Scan. Notably, this process appears to lean towards aggregating different modalities towards a distribution pattern similar with text. Fig. 5c demonstrates the data distribution after a Spatial Scan. Here, the features of the textual and image data disperse, with the modalities intermixed. The Spatial Scan thus reintegrates the previously aligned modalities, distributing them in a manner that reflects their combined influence." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.435, + 0.365, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.525, + 0.384, + 0.536 + ], + "angle": 0, + "content": "(a) Input data distribution." + }, + { + "type": "image", + "bbox": [ + 0.441, + 0.435, + 0.563, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.431, + 0.525, + 0.572, + 0.536 + ], + "angle": 0, + "content": "(b) After Channel Scan." + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.435, + 0.761, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.633, + 0.525, + 0.766, + 0.536 + ], + "angle": 0, + "content": "(c) After Spatial Scan." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.548, + 0.784, + 0.603 + ], + "angle": 0, + "content": "Fig.5: Data distribution after Channel Scan and Spacial Scan. Image in red and text data in blue. The Channel Scan tends to aggregate different modalities towards the distribution of textual side. The Spatial Scan reintegrates the previously aligned modalities, distributing them in a manner that reflects their combined influence." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.66, + 0.784, + 0.688 + ], + "angle": 0, + "content": "Table 4: Ablation study on global and local interaction. Both global and local interactions are crucial for modality fusion. Results are evaluated on RefCOCO." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.701, + 0.787, + 0.769 + ], + "angle": 0, + "content": "
ImageGlobalLocalvaltestAtestB
mIoUoIoUPr@50Pr@70mIoUoIoUPr@50Pr@70mIoUoIoUPr@50Pr@70
✓✓-69.166.679.566.371.469.882.370.966.363.774.761.0
✓-✓69.967.980.668.272.270.683.973.166.463.875.361.7
✓✓✓71.668.482.170.973.371.584.874.568.464.577.164.9
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Effects of Global and Local Features. We also ablate the effect of global and local interactions when forming the hybrid feature cube, formally, the \\(\\tilde{\\mathbf{F}}_t\\) and \\(\\tilde{\\mathbf{F}}_c\\) in Eq. (6). The results in Tab. 4 illustrate that the integration of both global and local features significantly enhances the performance." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.237, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.114, + 0.669, + 0.131 + ], + "angle": 0, + "content": "Y. Yang, C. Ma, J. Yao, Z. Zhong, Y. Zhang, and Y. Wang" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.439, + 0.162 + ], + "angle": 0, + "content": "4.6 Visualization Results" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.189, + 0.79, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.464, + 0.789, + 0.494 + ], + "angle": 0, + "content": "Fig. 6: Visualization results of our ReMamber and the baseline model LAVT. Our model is able to predict more accurate masks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.5, + 0.79, + 0.548 + ], + "angle": 0, + "content": "Fig. 6 presents the visualization results of our method compared with the baseline method LAVT [56]. Our ReMamber is capable of producing segmentation results with higher accuracy." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.566, + 0.362, + 0.584 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.79, + 0.718 + ], + "angle": 0, + "content": "In this study, we introduce ReMamber, a novel architecture utilizing the Mamba framework in Referring Image Segmentation (RIS). It marks a significant advancement in multi-modal understanding. By integrating visual and textual information through innovative Mamba Twister blocks, our approach sets new benchmarks in image-text modality fusion. Achieving competitive results across multiple RIS datasets, our research highlights the potential of Mamba architecture in enhancing the scalability and performance of multi-modal tasks, offering promising directions for future exploration in the field." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Limitations and Future Works. In our current architecture, the segmentation decoder is constructed by only a few convolutional layers. As shown in Sec. 4.3, the integration of the Cross-Attention mechanism within the Mamba-based architecture demonstrates sub-optimal compatibility, undermining the overall efficacy of our method. In light of these findings, future endeavors will be directed towards the investigation of more sophisticated multi-modal segmentation decoders which best fit Mamba architecture." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.303, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ReMamber: Referring Image Segmentation with Mamba Twister" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.176, + 0.785, + 0.204 + ], + "angle": 0, + "content": "1. Chen, B., Hu, Z., Ji, Z., Bai, J., Zuo, W.: Position-aware contrastive alignment for referring image segmentation. arXiv preprint arXiv:2212.13419 (2022) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.785, + 0.245 + ], + "angle": 0, + "content": "2. Chen, D.J., Jia, S., Lo, Y.C., Chen, H.T., Liu, T.L.: See-through-text grouping for referring image segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7454-7463 (2019) 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.247, + 0.785, + 0.286 + ], + "angle": 0, + "content": "3. Chen, Y.W., Tsai, Y.H., Wang, T., Lin, Y.Y., Yang, M.H.: Referring expression object segmentation with caption-aware consistency. arXiv preprint arXiv:1910.04748 (2019) 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.785, + 0.328 + ], + "angle": 0, + "content": "4. Ding, H., Liu, C., Wang, S., Jiang, X.: Vision-language transformer and query generation for referring segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16321-16330 (2021) 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.33, + 0.785, + 0.383 + ], + "angle": 0, + "content": "5. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.385, + 0.785, + 0.438 + ], + "angle": 0, + "content": "6. Feng, G., Hu, Z., Zhang, L., Lu, H.: Encoder fusion network with co-attention embedding for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15506-15515 (2021) 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.44, + 0.785, + 0.481 + ], + "angle": 0, + "content": "7. Fu, D.Y., Dao, T., Saab, K.K., Thomas, A.W., Rudra, A., Ré, C.: Hungry hungry hippos: Towards language modeling with state space models. arXiv preprint arXiv:2212.14052 (2022) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.482, + 0.785, + 0.508 + ], + "angle": 0, + "content": "8. Gu, A., Dao, T.: Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752 (2023) 2, 4, 5, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "9. Gu, A., Dao, T., Ermon, S., Rudra, A., Ré, C.: Hippo: Recurrent memory with optimal polynomial projections. Advances in neural information processing systems 33, 1474-1487 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.551, + 0.785, + 0.591 + ], + "angle": 0, + "content": "10. Gu, A., Goel, K., Gupta, A., Ré, C.: On the parameterization and initialization of diagonal state space models. Advances in Neural Information Processing Systems 35, 35971-35983 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.785, + 0.619 + ], + "angle": 0, + "content": "1. Gu, A., Goel, K., Ré, C.: Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396 (2021) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.785, + 0.66 + ], + "angle": 0, + "content": "2. Gu, A., Johnson, I., Goel, K., Saab, K., Dao, T., Rudra, A., Ré, C.: Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems 34, 572-585 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "3. Gupta, A., Gu, A., Berant, J.: Diagonal state spaces are as effective as structured state spaces. Advances in Neural Information Processing Systems 35, 22982-22994 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.703, + 0.785, + 0.742 + ], + "angle": 0, + "content": "4. He, X., Cao, K., Yan, K., Li, R., Xie, C., Zhang, J., Zhou, M.: Pan-mamba: Effective pan-sharpening with state space model. arXiv preprint arXiv:2402.12192 (2024) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.785, + 0.798 + ], + "angle": 0, + "content": "5. Hu, R., Rohrbach, M., Darrell, T.: Segmentation from natural language expressions. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14. pp. 108-124. Springer (2016) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "6. Hu, Z., Feng, G., Sun, J., Zhang, L., Lu, H.: Bi-directional relationship inferring network for referring image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4424-4433 (2020) 3, 12" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.176, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.668, + 0.13 + ], + "angle": 0, + "content": "Y. Yang, C. Ma, J. Yao, Z. Zhong, Y. Zhang, and Y. Wang" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "17. Huang, S., Hui, T., Liu, S., Li, G., Wei, Y., Han, J., Liu, L., Li, B.: Referring image segmentation via cross-modal progressive comprehension. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10488-10497 (2020) 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.247 + ], + "angle": 0, + "content": "18. Hui, T., Liu, S., Huang, S., Li, G., Yu, S., Zhang, F., Han, J.: Linguistic structure guided context modeling for referring image segmentation. In: European Conference on Computer Vision. pp. 59-75. Springer (2020) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.248, + 0.787, + 0.288 + ], + "angle": 0, + "content": "19. Islam, M.M., Bertasius, G.: Long movie clip classification with state-space video models. In: European Conference on Computer Vision. pp. 87-104. Springer (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.289, + 0.787, + 0.344 + ], + "angle": 0, + "content": "20. Islam, M.M., Hasan, M., Athrey, K.S., Braskich, T., Bertasius, G.: Efficient movie scene detection using state-space transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18749-18758 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.787, + 0.389 + ], + "angle": 0, + "content": "21. Jing, Y., Kong, T., Wang, W., Wang, L., Li, L., Tan, T.: Locate then segment: A strong pipeline for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9858-9867 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.389, + 0.787, + 0.444 + ], + "angle": 0, + "content": "22. Kamath, A., Singh, M., LeCun, Y., Synnaeve, G., Misra, I., Carion, N.: Mdetr-modulated detection for end-to-end multi-modal understanding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1780-1790 (2021) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.445, + 0.787, + 0.487 + ], + "angle": 0, + "content": "23. Kazemzadeh, S., Ordonez, V., Matten, M., Berg, T.: ReferItGame: Referring to objects in photographs of natural scenes. In: Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP) (2014) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.543 + ], + "angle": 0, + "content": "24. Kim, N., Kim, D., Lan, C., Zeng, W., Kwak, S.: Restr: Convolution-free referring image segmentation using transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18145-18154 (2022) 1, 2, 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.544, + 0.787, + 0.586 + ], + "angle": 0, + "content": "25. Li, R., Li, K., Kuo, Y.C., Shu, M., Qi, X., Shen, X., Jia, J.: Referring image segmentation via recurrent refinement networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5745-5753 (2018) 1, 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.787, + 0.628 + ], + "angle": 0, + "content": "26. Liang, D., Zhou, X., Wang, X., Zhu, X., Xu, W., Zou, Z., Ye, X., Bai, X.: Point-mamba: A simple state space model for point cloud analysis. arXiv preprint arXiv:2402.10739 (2024) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.629, + 0.787, + 0.657 + ], + "angle": 0, + "content": "27. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common Objects in Context. ECCV (2014) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.658, + 0.787, + 0.699 + ], + "angle": 0, + "content": "28. Liu, C., Ding, H., Jiang, X.: Gres: Generalized referring expression segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23592-23601 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.787, + 0.742 + ], + "angle": 0, + "content": "29. Liu, C., Lin, Z., Shen, X., Yang, J., Lu, X., Yuille, A.: Recurrent multimodal interaction for referring image segmentation. In: Proceedings of the IEEE international conference on computer vision. pp. 1271-1280 (2017) 1, 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.787, + 0.798 + ], + "angle": 0, + "content": "30. Liu, J., Ding, H., Cai, Z., Zhang, Y., Satzoda, R.K., Mahadevan, V., Manmatha, R.: Polyformer: Referring image segmentation as sequential polygon generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18653-18663 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "31. Liu, J., Yang, H., Zhou, H.Y., Xi, Y., Yu, L., Yu, Y., Liang, Y., Shi, G., Zhang, S., Zheng, H., et al.: Swin-umamba: Mamba-based unet with imagenet-based pretraining. arXiv preprint arXiv:2402.03302 (2024) 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.303, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ReMamber: Referring Image Segmentation with Mamba Twister" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "32. Liu, S., Hui, T., Huang, S., Wei, Y., Li, B., Li, G.: Cross-modal progressive comprehension for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(9), 4761-4775 (2021) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "33. Liu, Y., Tian, Y., Zhao, Y., Yu, H., Xie, L., Wang, Y., Ye, Q., Liu, Y.: Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166 (2024) 2, 4, 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.786, + 0.259 + ], + "angle": 0, + "content": "34. Luo, G., Zhou, Y., Ji, R., Sun, X., Su, J., Lin, C.W., Tian, Q.: Cascade grouped attention network for referring expression segmentation. Proceedings of the 28th ACM International Conference on Multimedia (2020) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.3 + ], + "angle": 0, + "content": "35. Ma, C., Yang, Y., Ju, C., Zhang, F., Zhang, Y., Wang, Y.: Attrseg: Open-vocabulary semantic segmentation via attribute decomposition-aggregation. NeurIPS (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "36. Ma, C., Yang, Y., Wang, Y., Zhang, Y., Xie, W.: Open-vocabulary semantic segmentation with frozen vision-language models. In: British Machine Vision Conference (2022) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.37 + ], + "angle": 0, + "content": "37. Ma, J., Li, F., Wang, B.: U-mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722 (2024) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.398 + ], + "angle": 0, + "content": "38. Mao, J., Huang, J., Toshev, A., Camburu, O.M., Yuille, A., Murphy, K.: Generation and comprehension of unambiguous object descriptions. CVPR (2015) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.439 + ], + "angle": 0, + "content": "39. Margffoy-Tuay, E., Pérez, J.C., Botero, E., Arbeláez, P.: Dynamic multimodal instance segmentation guided by natural language queries. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 630-645 (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.467 + ], + "angle": 0, + "content": "40. Mehta, H., Gupta, A., Cutkosky, A., Neyshabur, B.: Long range language modeling via gated state spaces. arXiv preprint arXiv:2206.13947 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.509 + ], + "angle": 0, + "content": "41. Nagaraja, V.K., Morariu, V.I., Davis, L.S.: Modeling context between objects for referring expression understanding. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) ECCV (2016) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.785, + 0.537 + ], + "angle": 0, + "content": "42. Perez, E., Strub, F., de Vries, H., Dumoulin, V., Courville, A.C.: Film: Visual reasoning with a general conditioning layer. In: AAAI (2018) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.785, + 0.578 + ], + "angle": 0, + "content": "43. Qu, M., Wu, Y., Wei, Y., Liu, W., Liang, X., Zhao, Y.: Learning to segment every referring object point by point. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.605 + ], + "angle": 0, + "content": "44. Ruan, J., Xiang, S.: Vm-unet: Vision mamba unet for medical image segmentation. arXiv preprint arXiv:2402.02491 (2024) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.605, + 0.785, + 0.647 + ], + "angle": 0, + "content": "45. Shi, H., Li, H., Meng, F., Wu, Q.: Key-word-aware network for referring expression image segmentation. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 38-54 (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.675 + ], + "angle": 0, + "content": "46. Smith, J.T., Warrington, A., Linderman, S.W.: Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933 (2022) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.716 + ], + "angle": 0, + "content": "47. Tang, J., Zheng, G., Shi, C., Yang, S.: Contrastive grouping with transformer for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23570-23580 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.744 + ], + "angle": 0, + "content": "48. Tang, J., Zheng, G., Shi, C., Yang, S.: Contrastive grouping with transformer for referring image segmentation. In: CVPR (2023) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.785 + ], + "angle": 0, + "content": "49. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "50. Wang, J., Zhu, W., Wang, P., Yu, X., Liu, L., Omar, M., Hamid, R.: Selective structured state-spaces for long-form video understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6387-6397 (2023) 4" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.114, + 0.668, + 0.13 + ], + "angle": 0, + "content": "Y. Yang, C. Ma, J. Yao, Z. Zhong, Y. Zhang, and Y. Wang" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "51. Wang, Z., Lu, Y., Li, Q., Tao, X., Guo, Y., Gong, M., Liu, T.: Cris: Clip-driven referring image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11686-11695 (2022) 1, 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.231 + ], + "angle": 0, + "content": "52. Xing, Z., Ye, T., Yang, Y., Liu, G., Zhu, L.: Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. arXiv preprint arXiv:2401.13560 (2024) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.285 + ], + "angle": 0, + "content": "53. Xu, Z., Chen, Z., Zhang, Y., Song, Y., Wan, X., Li, G.: Bridging vision and language encoders: Parameter-efficient tuning for referring image segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17503-17512 (2023) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "54. Yang, S., Xia, M., Li, G., Zhou, H.Y., Yu, Y.: Bottom-up shift and reasoning for referring image segmentation. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11261-11270 (2021) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.355 + ], + "angle": 0, + "content": "55. Yang, Y., Ma, C., Ju, C., Zhang, Y., Wang, Y.: Multi-modal prototypes for open-set semantic segmentation. IJCV (2024) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.41 + ], + "angle": 0, + "content": "56. Yang, Z., Wang, J., Tang, Y., Chen, K., Zhao, H., Torr, P.H.: Lavt: Language-aware vision transformer for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18155-18165 (2022) 1, 2, 3, 9, 11, 12, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.452 + ], + "angle": 0, + "content": "57. Ye, L., Rochan, M., Liu, Z., Wang, Y.: Cross-modal self-attention network for referring image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10502-10511 (2019) 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.507 + ], + "angle": 0, + "content": "58. Yu, L., Lin, Z., Shen, X., Yang, J., Lu, X., Bansal, M., Berg, T.L.: Matttnet: Modular attention network for referring expression comprehension. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.785, + 0.535 + ], + "angle": 0, + "content": "59. Zhang, T., Li, X., Yuan, H., Ji, S., Yan, S.: Point could mamba: Point cloud learning via state space model. arXiv preprint arXiv: 2403.00762 (2024) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.563 + ], + "angle": 0, + "content": "60. Zheng, Z., Wu, C.: U-shaped vision mamba for single image dehazing. arXiv preprint arXiv:2402.04139 (2024) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.605 + ], + "angle": 0, + "content": "61. Zhu, C., Zhou, Y., Shen, Y., Luo, G., Pan, X., Lin, M., Chen, C., Cao, L., Sun, X., Ji, R.: Seqtr: A simple yet universal network for visual grounding. In: European Conference on Computer Vision. pp. 598-615. Springer (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.646 + ], + "angle": 0, + "content": "62. Zhu, L., Liao, B., Zhang, Q., Wang, X., Liu, W., Wang, X.: Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417 (2024) 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.646 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/ReMamber_ Referring Image Segmentation with Mamba Twister/04366454-5dd3-4b5b-8a0b-a5b01b0efbcd_origin.pdf b/2024/ReMamber_ Referring Image Segmentation with Mamba Twister/04366454-5dd3-4b5b-8a0b-a5b01b0efbcd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..702395d0d53c9687b05d02f82075f5f5df47ecf2 --- /dev/null +++ b/2024/ReMamber_ Referring Image Segmentation with Mamba Twister/04366454-5dd3-4b5b-8a0b-a5b01b0efbcd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a538d665679b4bac921b622a55d2227c22bd532d0d692bf205ea16182edeb0b1 +size 3036451 diff --git a/2024/ReMamber_ Referring Image Segmentation with Mamba Twister/full.md b/2024/ReMamber_ Referring Image Segmentation with Mamba Twister/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9703592ab1bb21b5a88f4c372f2865412e71da05 --- /dev/null +++ b/2024/ReMamber_ Referring Image Segmentation with Mamba Twister/full.md @@ -0,0 +1,322 @@ +# ReMamber: Referring Image Segmentation with Mamba Twister + +Yuhuan Yang\*1, Chaofan Ma\*1, Jiangchao Yao1, Zhun Zhong\*2, Ya Zhang1, and Yanfeng Wang\*1 + +1 Shanghai Jiao Tong University + +$^{2}$ University of Nottingham + +{yangyuhuan, chaofanna, sunarker, ya_zhang, wangyanfeng622}@sjtu.edu.cn + +zhunzhong007@gmail.com + +Abstract. Referring Image Segmentation (RIS) leveraging transformers has achieved great success on the interpretation of complex visual-language tasks. However, the quadratic computation cost makes it resource-consuming in capturing long-range visual-language dependencies. Fortunately, Mamba addresses this with efficient linear complexity in processing. However, directly applying Mamba to multi-modal interactions presents challenges, primarily due to inadequate channel interactions for the effective fusion of multi-modal data. In this paper, we propose ReMamber, a novel RIS architecture that integrates the power of Mamba with a multi-modal Mamba Twister block. The Mamba Twister explicitly models image-text interaction, and fuses textual and visual features through its unique channel and spatial twisting mechanism. We achieve competitive results on three challenging benchmarks with a simple and efficient architecture. Moreover, we conduct thorough analyses of ReMamber and discuss other fusion designs using Mamba. These provide valuable perspectives for future research. The code has been released at: https://github.com/yyh-rain-song/ReMamber. + +Keywords: Referring Image Segmentation (RIS) $\cdot$ Multi-Modal Understanding $\cdot$ Mamba Architecture + +# 1 Introduction + +Referring Image Segmentation (RIS) is a crucial yet challenging task in the area of multi-modal understanding [15, 25, 29]. Unlike ordinary image segmentation, RIS involves the identification and segmentation of specific objects in image based on the textual descriptions. This thereby requires the model to be capable of understanding vision-language interactions, which is the core challenge of RIS. + +Thanks to the powerful attention mechanisms, it has achieved great success for RIS by leveraging transformers to promote the exact recognition of multimodal information. For example, existing works have designed transformer decoder [4,22] or transformer encoder-decoder [24,35,36,51,55,56] to comprehensively fuse visual and linguistic features, which have achieved great progress. + +![](images/9d93d1e841052cfac5d99205b13358187664222d816e3f973a0ca52663f8b2c6.jpg) +Fig. 1: We propose ReMamber, a novel referring segmentation architecture with Mamba twister. It consists of several Mamba Twister block. Each block contains several visual state space (VSS) layers and a Twisting layer. The Twisting layer first calculates the interaction between image and text, and then forms a hybrid feature cube. Finally, it "twists" the feature cube using the Channel and Spatial Scan along each dimension. + +Nonetheless, it is a quadratic increase in both computation and memory when applying full attention in transformers [5, 49]. This leads to the limitation for resource-intensive scenarios, for example, in capturing long-range visual-language dependencies. And this is particularly important in the context of large-size images with long textual descriptions. + +Fortunately, recent advances of State Space Models (SSMs) [7,8,11,46] has emerged as promising architectures for solving the above issue. Specifically, Mamba [8] marks a significant advancement for efficient training and inference with linear complexity, which has been incorporated into various visual tasks. However, current efforts primarily pay attention to single-modality settings, such as image classification [33,62], biomedical image segmentation [31,37,44,52], low-level vision [14,60], and point cloud analysis [26,59]. In this paper, we pioneer the exploration of Mamba in multi-modal RIS setting, and identify that the prevalent multi-modality token splicing method in transformers [4,24,56] is no longer effective. Since in Mamba, a fundamental deficiency exists: the interactions are insufficient between channels of different tokens [8], which adversely affects the fusion of multi-modal information in RIS. + +To address this dilemma, we propose ReMamber, a novel referring segmentation architecture with Mamba twister. As shown in Fig. 1, ReMamber comprises several Mamba Twister blocks, which allow the model to be "aware" of the textual context at every spatial location. Each block consists of several visual state space (VSS) layers and a Twisting layer. Specifically, the VSS layers initially extract visual features, and the Twisting layer injects the textual information into the visual modality. The Twisting layer is structured into three critical + +components. (1) A vision-language interaction operation is designed to explicitly capture the fine-grained interactions between modalities. This is achieved by evaluating the similarity between visual and textual tokens then mapping them into a shared feature space, thus generating a multi-modal feature. (2) A hybrid feature cube is then created by concatenating the visual feature, multimodal feature, and global textual feature. This ensures that each visual token receives uniform influence from both local and global contexts, preventing the overshadowing of subtle textual cues by predominant visual features. (3) Lastly, to address the inadequate interaction within channels in Mamba, a twisting mechanism is deployed. This process "twists" the feature cube channel- and spatial-wise, thereby enhancing interaction within and across modalities. It is accomplished through two consecutive SSMs scanning along channel and spatial dimensions, respectively. + +To sum up, the key contributions of our research are highlighted as follows: + +- We pioneer the exploration of Mamba in referring image segmentation (RIS), demonstrating Mamba's significant potential for multi-modal understanding. +- We design a novel framework, ReMamber, that mainly contains several Mamba Twister blocks. This design effectively captures vision-language interactions using the "twisting mechanism". +- We achieve competitive results on three challenging benchmarks. Moreover, we conduct thorough analyses of ReMamber and discuss other vision-language fusion designs using Mamba. These provide valuable perspectives for future research. + +# 2 Related Work + +# 2.1 Referring Image Segmentation + +Referring image segmentation (RIS) aims to segment the entities in the image following natural language instruction. Early approaches [15, 25, 29, 39] leveraged RNNs or LSTMs to encode linguistic representations, and CNNs to extract spatial features from the image at varying levels. These disparate modalities were then integrated using the multi-modal LSTM [2, 15, 21, 25, 29, 39], attention mechanism [6, 16, 45, 57], cycle-consistency [3], and graph convolution [17]. + +Recent trends have shifted towards leveraging transformers for enhanced capturing and fusion of vision-language modalities. MDETR [22] and VLT [4] design a transformer decoder for fusing linguistic and visual features. LAVT [56] adopts Swin Transformer as the visual backbone and incorporate vision-language fusion modules within the visual encoder's final layers. Similar strategies are employed by ReSTR [24] and CRIS [51], which utilize dual transformer encoders for initial modality encoding, followed by feature fusion through a multi-modal transformer encoder or decoder. Other models like PolyFormer [30], SeqTr [61] and [43] also adopt a multi-modal transformer for vision-language fusion, but they output masks as sequences of contour points. Meanwhile, GRES [28] and CGFormer [47] consider transformer queries as region proposals, and regarding + +segmentations as proposal-level classification problems. Distinct from all existing works, we introduce a pioneering multi-modal architecture named ReMamber, based on Mamba [8]. This novel approach underscores the untapped potential of Mamba in advancing the field of referring image segmentation. + +# 2.2 State Space Models and Visual Applications + +State space models (SSMs), originally derived from control theory, have been effectively combined with deep learning to model the long-range dependencies. Early works like LSSL [12] show potential when combined with HiPPO [9] initialization. S4 [11] was designed to diminish both computational and memory demands associated with state representations. It scales linearly with sequence length, offering a notable advantage over CNNs and transformers. Building on S4, S5 [46] introduces the efficient parallel scan and MIMO SSM, further refining the approach. Later, H3 [7] expanded on these foundations, achieving competitive performance with transformers in language modeling. Recently, Mamba [8] marked a significant advancement with its linear-time inference and efficient training, incorporating a selection mechanism and hardware-aware algorithms building upon prior works [10, 13, 40]. + +With the demonstrated success of SSMs like Mamba in language modeling, researchers have begun to investigate their applicability to visual tasks. Works such as ViS4mer [19], Selective S4 [50] and TranS4mer [20] directly use the S4's ability to capture long-range sequences for understanding inter-frame relations in video classification and detection. However, they still employ Vision Transformers for intra-frame feature extraction. More recently, Vim [62] and VMamba [33] have shown promising results as fully Mamba-based vision backbones for image classification, detection, and segmentation. At the same time, several studies have explored Mamba-based architecture on various vision tasks, such as biomedical image segmentation [31, 37, 44, 52], low-level vision [14, 60], and point cloud analysis [26, 59]. Nevertheless, all above efforts primarily focus on single-modality vision tasks. In this paper, we pioneer the exploration of Mamba-base architecture in multi-modality RIS task. + +# 3 Methods + +This paper aims to develop a multi-modal Mamba-based architecture for RIS task. In Sec. 3.1, we give the preliminary. From Sec. 3.2 to Sec. 3.5, we describe our architecture from coarse to fine. In Sec. 3.6, we introduce other three variants besides our design. And in Sec. 3.7 we detail the training process. + +# 3.1 Preliminary: State Space Model + +State Space Model (SSM) is a sequence model inspired by continuous systems. It is designed to capture a mapping relationship between two functions or sequences, expressed as $x(t) \in \mathbb{R} \mapsto y(t) \in \mathbb{R}$ , through a hidden state $h(t) \in \mathbb{R}^N$ . + +The evolution of the hidden state over time is governed by specific parameters $\mathbf{A} \in \mathbb{R}^{N \times N}$ , which directs the state evolution, and $\mathbf{B} \in \mathbb{R}^N$ , $\mathbf{C} \in \mathbb{R}^N$ , which performs the projection. Formally, SSM can be described by the following equations: + +$$ +\begin{array}{l} h ^ {\prime} (t) = \mathbf {A} h (t) + \mathbf {B} x (t), \\ (t) = \mathbf {C} h ^ {\prime} (t) \end{array} \tag {1} +$$ + +$$ +y (t) = \mathbf {C} h (t). +$$ + +Discretizing this system introduces a timescale parameter $\Delta t$ , which transforms the continuous parameters $\mathbf{A}, \mathbf{B}$ into their discrete counterparts $\overline{\mathbf{A}}, \overline{\mathbf{B}}$ . A common way for this transformation is the zero-order hold (ZOH) approach, defined as: + +$$ +\begin{array}{l} \overline {{\mathbf {A}}} = \exp (\boldsymbol {\Delta} \mathbf {A}), \\ \overline {{\mathbf {B}}} = (\boldsymbol {\Delta} \mathbf {A}) ^ {- 1} (\boldsymbol {\Delta} \mathbf {A}) - \mathbf {I}) - \boldsymbol {\Delta} \mathbf {B}. \end{array} \tag {2} +$$ + +$$ +\overline {{\mathbf {B}}} = (\boldsymbol {\Delta} \mathbf {A}) ^ {- 1} (\exp (\boldsymbol {\Delta} \mathbf {A}) - \mathbf {I}) \cdot \boldsymbol {\Delta} \mathbf {B}. \tag {2} +$$ + +Then, the discretized version of this system can be represented as: + +$$ +h _ {t} = \overline {{\mathbf {A}}} h _ {t - 1} + \overline {{\mathbf {B}}} x _ {t}, \tag {3} +$$ + +$$ +y _ {t} = \mathbf {C} h _ {t}. +$$ + +Mamba [8], as a variant of SSM, releases the linearity constraint of the original SSM described in Eq. (1), by making $\mathbf{B}$ and $\mathbf{C}$ to be input-dependent, whereas still maintaining the linear complexity during forward process. + +# 3.2 Architecture Overview + +The pivotal aspect of Referring Image Segmentation (RIS) is the correspondence between image and textual input. To achieve this, we propose ReMamber. Fig. 2 shows the overview of our architecture. The basic block for ReMamber is the Mamba Twister block. It is a multi-modal fusion block that takes both visual and textual feature as input, and outputs the fused multi-modal feature representation. We extract the intermediate feature after each Mamba Twister block, and then feed it into a flexible decoder to generate the final segmentation mask. The decoder is task-invariant, so any downstream architecture can be applied. Below in Sec. 3.3, we will first introduce the key component: Mamba Twister block. + +# 3.3 Mamba Twister Block + +As shown in Fig. 2, the Mamba Twister Block consists of several visual state space (VSS) layers and a Twisting layer. The VSS layer is designed to process features in the spatial domain. It treats the input feature as a series of image tokens and aims to discern the spatial relationships among these tokens. The Twisting layer aims to inject text condition into the image feature, thus guiding the transformation of image feature. It begins by condensing the textual sequence to compute the cross-correlation between the image and textual modalities. Subsequently, it disseminates this information across each image feature patch via a twisting operation. We'll detail the introduction of VSS layer in Sec. 3.4 and the Twisting layer in Sec. 3.5. + +![](images/d425cdebb12538c725842f9fa2a8d47af8f9c7a3b5d5779d3ec499e6f6b593e6.jpg) + +![](images/c66f5d29467cef28a888958bc867e7063e0607952d56740cde5f43b7569efc97.jpg) +Fig. 2: Overview architecture of our ReMamber. The basic block for ReMamber is the Mamba Twister block. It consists of several visual state space (VSS) layers and a Twisting layer. The Twisting layer first constructs hybrid feature cube from text, image, and multi-modal features via channel concatenation. Then, it "twists" the cube by Channel Scan and Spatial Scan. We extract intermediate features after each Mamba Twister block, and feed it into a flexible decoder for final segmentation. + +# 3.4 VSS Layer for Spatial Data Processing + +Since the SSM is initially designed for processing temporal or causal data, which can not effectively process non-causal data types, i.e., 2-dimensional image in our case. To solve this issue, we adopt the Cross-Scan-Module proposed in [33], which unfolds image patches into sequences and then scans them in four distinct directions, ensuring that the information from all pixels is integrated during feature transformation. We replace the scanning operation in vanilla Mamba block with CSM, forming a visual state space (VSS) layer for our spatial feature transformation. + +# 3.5 Twisting Layer for Multi-modal Fusion + +Here we introduce the detailed design of our Twisting layer. Intuitively, the Twisting layer first constructs a feature cube by arranging information from different modalities orderly. Then, pieces of information intertwine with each other when the cube is twisted through two SSM layers along different axis, thereby achieving modality fusion. Formally, given image feature $\mathbf{F}_i\in \mathbb{R}^{h\times w\times C_i}$ and text feature $\mathbf{F}_t\in \mathbb{R}^{L\times C_t}$ , the Mamba twister aims to learn a function $(\mathbf{F}_i,\mathbf{F}_t)\mapsto \tilde{\mathbf{F}}\in \mathbb{R}^{h\times w\times C_o}$ by first forming a hybrid multi-modal feature cube and then conduct two SSM scans on it. $C_i,C_t$ and $C_o$ are the dimensions of visual, textual, and output feature; $h,w$ and $L$ are height, weight, and length. + +Forming the Hybrid Feature Cube. To explicitly build the image-text correspondence, we design a novel vision-language interaction operation and then form a hybrid feature cube. This allows more effective modality fusion. + +To be specific, we formulated the vision-language interaction operation into two manners: global interaction and local interaction. The global interaction treats the text sequence as a whole, which means that all image patches should be aware of the semantic meaning of the text expression. We pool a global representation $\mathbf{F}_t^{\mathrm{CLS}}\in \mathbb{R}^{C_t}$ from the text sequence $\mathbf{F}_t$ to convey this information. The $\mathbf{F}_t^{\mathrm{CLS}}$ is further expanded to the same size as the image feature. Formally: + +$$ +\tilde {\mathbf {F}} _ {t} = \operatorname {E x p a n d} \left(\mathbf {F} _ {t} ^ {\mathrm {C L S}}\right) \in \mathbb {R} ^ {h \times w \times C _ {t}}, \tag {4} +$$ + +where $\operatorname{Expand}(\cdot)$ denotes the operation that expands the input tensor to the same size as the image feature. + +However, a global representation may not be enough to capture the intricate relationships between different modalities. For example, some words such as color, shape or location may be more relevant to certain image patches than others. The local interaction aims to capture such correlation in a fine-grained level. Formally, we calculate the local interaction map $\mathbf{F}_c$ using matrix multiplication: + +$$ +\mathbf {F} _ {c} = \mathbf {F} _ {i} \mathbf {W} _ {i} \cdot \left(\mathbf {F} _ {t} \mathbf {W} _ {t}\right) ^ {T} \in \mathbb {R} ^ {h \times w \times L}, \tag {5} +$$ + +where $\mathbf{W}_i\in \mathbb{R}^{C_i\times C_c}$ and $\mathbf{W}_t\in \mathbb{R}^{C_t\times C_c}$ are learnable parameters. To enhance feature processing in high dimension, we use a single convolutional layer to transform $\mathbf{F}_c$ into $\tilde{\mathbf{F}}_c\in \mathbb{R}^{h\times w\times C_c}$ . We then concatenate $\mathbf{F}_i$ , $\tilde{\mathbf{F}}_t$ , and $\tilde{\mathbf{F}}_c$ along the channel dimension to form the hybrid feature cube. Formally: + +$$ +\mathbf {F} _ {\text {c u b e}} = \left[ \mathbf {F} _ {i}, \tilde {\mathbf {F}} _ {t}, \tilde {\mathbf {F}} _ {c} \right] \in \mathbb {R} ^ {h \times w \times \left(C _ {i} + C _ {t} + C _ {c}\right)}. \tag {6} +$$ + +Twisting the Hybrid Feature Cube. Generally, the scanning operation for vanilla SSM is almost independent across channels [8]. This is not enough for feature communication, especially when information from different modalities is arranged along the channel dimension, i.e., our hybrid feature cube. + +To foster the interactions within and across modalities, here we design a twisting mechanism to "twist" the hybrid feature cube along different axis. Specifically, it is composed of the Channel Scan and Spacial Scan successively. The Channel Scan first treats the channel as an ordered sequence, and learns to communicate cross channels, thus foster the modality fusion. Then, the Spatial Scan operates on the spatial dimension, and learns to communicate information cross patches within each modality separately. Formally, this process can be written as: + +$$ +\mathbf {F} _ {\text {o u t}} = \operatorname {S S M} _ {\text {s p a t i a l}} \left(\operatorname {S S M} _ {\text {c h a n n e l}} \left(\mathbf {F} _ {\text {c u b e}}\right)\right), \tag {7} +$$ + +where $\mathrm{SSM}_{\mathrm{channel}}$ refers to the Channel Scan. It is an SSM layer that treats the concatenated feature as an ordered sequence through channel dimension, and presents 1-D selective scan along the channel. While $\mathrm{SSM}_{\mathrm{spatial}}$ denotes the Spatial Scan. It is an VSS layer and presents 2D selective scan along the two spatial dimension. $\mathbf{F}_{\mathrm{out}}$ is then feed into the next layer for further process. + +![](images/f6619b0a50cf049cff6d6f9dfc3fe3b7c5b43fd0d01c0616a069983123bb7415.jpg) + +![](images/8f927a0c10a71fd6e58ee3e0efcfedee13aa408c9945f262d0adba7252f423a9.jpg) +Fig.3: Other multi-modal fusion designs. (a) In-context Conditioning appends text tokens ahead of image tokens. (b) Attention-based Conditioning utilizes cross-attention mechanism for modality fusion. (c) Norm Adaptation learns a scale and bias for the model's normalization layers. + +# 3.6 Discussion: Other Variants of Modality Fusion Designs + +Besides the proposed Twisting layer, we also explore other variants of multimodal fusion designs. Fig. 3 shows the structure of three other different variants. We here discuss these variants, and later in Sec. 4.3, we analyze the results for these variants. + +In-context Conditioning. As shown in Fig. 3 (a), in this variant, we simply append the text sequence before the image feature, forming a longer sequence than originally being processed. In this way, the model is able to fuse image features with the previous text as a context. This is the most straightforward way to allow the model aware of the text condition. + +Attention-based Conditioning. Attention mechanism is usually a strong baseline for sequence modeling. Here we utilize the cross-attention mechanism for two modality fusion, as shown in Fig. 3 (b), To be specific, we use the image feature as query, and the text feature as key and value, hoping to capture the correlation between image and text. + +Norm Adaptation. In this variant, we integrate the text input by adapting the scale and bias after norm layer using FiLM [42]. We first pool a global representation from text sequence, and then use a linear projection to transform it into the scale and shift in layer normalization. + +# 3.7 Model training + +The proposed ReMamber is a general framework for multi-modal fusion. Therefore, there is no restriction on the downstream decoder design nor the loss function. Here we use a simple convolution-based decoder for simplicity. The entire network is trained in an end-to-end manner. + +# 4 Experiments + +# 4.1 Datasets and Metrics + +Dataset. To assess the efficacy of our proposed approach, we executed experiments across three widely recognized datasets tailored for the referring image segmentation (RIS) task: RefCOCO, RefCOCO+ [23], and G-Ref [23,38,41]. RefCOCO [23] is a prominent dataset in the RIS domain, comprising 19,994 images and 142,210 referring expressions associated with 50,000 objects derived from the MSCOCO [27] dataset through a two-player game. RefCOCO+ enhances the challenge by excluding expressions with absolute location references, featuring 141,564 expressions for 49,856 objects across 19,992 images. G-Ref [38,41] enriches the dataset diversity with 104,560 referring expressions for 54,882 objects in 26,711 images, showcasing an average sentence length of 8.4 words with a heightened focus on location and appearance descriptions. + +Evaluation Metrics. In line with preceding studies [48], we adopt mIoU and oIoU as the main metrics. Furthermore, we incorporate the metric Precision@X $(\mathrm{X} \in \{50, 60, 70, 90\})$ for a more comprehensive evaluation, whereas Percision@X means the percentage of test images with an IoU score high than X%. + +# 4.2 Implementation Details + +Our model architecture is developed upon the foundations of Mamba [8] and VMamba [33]. We adopt ImageNet pre-trained weights as initialization, and train the model in an end-to-end manner. We set the input image resolution to 480 when compared with the state-of-the-art methods in Tab. 2, following [48,56] for fair comparison. For other experiments (Tabs. 1 and 3 to 4), we set the image resolution to 256 for faster training and ablation study without changing its property. For details about the training settings, please refer to our code. $^{1}$ + +# 4.3 Evaluation on Other Variants of Modality Fusion Designs + +As discussed in Sec. 3.6, besides our Mamba Twister, we also provide three other variant architectures for modality fusion, including In-context Conditioning, Attention-based Conditioning and Norm Adaptation. To make a fair comparison, for all architectures, we initialize their common part (VMamba-based parameters) with the same checkpoint. The difference is that the Twisting layer is replaced by the corresponding fusion methods. The parameters of fusion modules are nearly equal, i.e., {103, 113, 118, 116} KB for {In-Context, Attention, Norm Adaptation, Twister}. Here we provide a comprehensive analysis between these architectures, showing the advantages and disadvantages of each variant. + +Tab. 1 shows the comparison result. The Mamba Twister consistently outperforms the other variants across all metrics and datasets, indicating its superior capability in capturing and integrating contextual information for more accurate segmentation. + +Table 1: Comparison with other modality fusion variants. "Attention" means for "Attention-based Conditioning", "In-Context" for "In-Context Conditioning", and "Adaptation" for "Norm Adaptation". Mamba Twister steadily outperforms other variants across all metrics and datasets, indicating its superior capability in capturing and integrating contextual information for more accurate segmentation. + +
Dataset: RefCOCO
VariantsmIoUvalPr@70mIoUtestAmIoUtestB
oIoUPr@50Pr@70oIoUPr@50Pr@70oIoUPr@50
Attention65.362.375.160.667.564.479.165.461.758.069.653.4
In-Context69.165.979.767.571.268.982.271.866.262.875.061.7
Adaptation70.267.080.769.672.370.283.173.266.862.875.163.1
Mamba Twister71.668.482.170.973.371.584.874.568.464.577.164.9
Dataset: RefCOCO+
VariantsmIoUvalPr@70mIoUtestAmIoUtestB
oIoUPr@50oIoUPr@50Pr@70oIoUPr@50
Attention54.049.760.745.558.755.868.552.946.941.950.636.6
In-Context58.454.066.353.963.259.472.260.551.547.556.444.5
Adaptation60.355.268.257.364.259.973.462.953.948.759.047.2
Mamba Twister61.657.370.058.565.862.175.564.954.049.959.447.8
Dataset: G-Ref
VariantsmIoUvalPr@70Pr@90mIoUoIoUtest
oIoUPr@50Pr@50Pr@60Pr@70Pr@90
Attention50.549.254.246.036.78.150.850.857.349.838.66.4
In-Context54.853.459.552.343.413.355.554.860.753.645.313.3
Adaptation59.356.766.761.053.718.458.656.665.559.852.818.2
Mamba Twister61.158.068.462.855.218.561.259.069.063.455.518.7
+ +Surprisingly, despite widely used in transformers, the Attention-based Conditioning performs poorly in our task. This suggests that the cross-attention mechanism is not inherently suitable for Mamba architecture. This may be due to a fundamental discrepancy between the two systems. To be specific, (1) the Mamba model is predicted on ordered sequences that exhibit strict sequential dependencies, where the state at a given time point $t + 1$ is determined by the preceding time point $t$ and a hidden state; (2) In contrast, cross-attention mechanisms treats all tokens within a sequence equally. This discrepancy may undermine the Mamba model's ability to structurally model sequences, as the cross-attention mechanism does not preserve the sequential integrity and the hierarchical dependencies essential for the model's operation. + +Serving as a straightforward baseline, In-context Conditioning performs sub-optimal. This may because it models image-text interaction in an implicit way. In our situation, the length of the image feature is much larger than the textual side, textual information may be diluted during the forward process. Besides, this operation does not differentiate the textual information between image patches, and would be insufficient for multi-modal fusion. + +Finally, Norm Adaptation with explicit image-text interaction modeling appears to be a strong baseline. It achieves better performance than other two variants. However, when calculating scale and bias, it uses only a pooled vector as + +textual representation instead of the entire sequence. This potentially results in the loss of information, making it less effective compared to our Mamba Twister. + +Case-study using Attention Maps. In Fig. 4, we further visualized the attention maps in Attention-based Conditioning as well as the local interaction map in Mamba Twister defined in Eq. (5). The four images in Fig. 4(b) are arranged from left to right with the network going from shallow to deep. It can be observed that in the shallower layers, Mamba Twister focuses more on the lower-level features of the images such as edges. As the layer goes deeper, the cost map mostly concentrating on the target object, indicating that Mamba Twister can progressively guide the image feature towards the target described by the language. In contrast, though the Attention-based Conditioning variant is able to locate the target, its attention map performs poorly. This may suggest that the cross-attention mechanism struggles to accurately capture the correct context information, and the two modalities are not truly fused together. + +![](images/8b7666cefd151e855403be655d58509843b7dc2561f127aaaf01cdfc9768a29b.jpg) +Fig. 4: Cross-Attention map (up) and our local interaction map (down) comparison. Though both methods are able to predict target correctly, the cross-attention maps don't show correct image-text correlation, while ours are able to capture this relationship accurately, indicating that Mamba Twister is able to gradually fusing the two modality. + +# 4.4 Comparison with the State-of-the-Arts + +Tab. 2 shows the comparison of our ReMamber with state-of-the-art methods on RefCOCO, RefCOCO+, and G-Ref datasets. ReMamber outperforms all other methods across all datasets, demonstrating its superior capability in capturing and integrating contextual information for more accurate segmentation. In particular, Mamba outperforms previous Swin-based methods such as LAVT [56] by a considerable margin, indicating the remarkable capability of the Mamba-based architecture in segmentation tasks. and the efficiency of the newly introduced Mamba Twister. + +Table 2: Comparison with state-of-the-art methods on RefCOCO, RefCOCO+, and G-Ref. The metric of oIoU is reported. Best results are in bold. + +
ModelBackboneRefCOCORefCOCO+G-Ref
valtestAtestBvaltestAtestBvalUtestU
PCAN [1]ResNet-5069.5171.6464.1858.2563.6848.8959.9860.8
MAAttNet [58]MaskRCNN ResNet-10156.5162.3751.746.6752.3940.0847.6448.61
RMI [29]DeepLab ResNet-10145.1845.6945.5729.8630.4829.50--
RRN [25]DeepLab ResNet-10155.3357.2653.9539.7542.1536.11--
CMSA [57]DeepLab ResNet-10158.3260.6155.0943.7647.637.89--
CAC [3]DeepLab ResNet-10158.9061.7753.81---46.3746.95
STEP [2]DeepLab ResNet-10160.0463.4657.9748.1952.3340.41--
BRINet [16]DeepLab ResNet-10160.9862.9959.2148.1752.3242.11--
CMPC [17]DeepLab ResNet-10161.3664.5359.6449.5653.4443.23--
LSCM [18]DeepLab ResNet-10161.4764.9959.5549.3453.1243.50--
CMPC+ [32]DeepLab ResNet-10162.4765.0860.8250.2554.0443.47--
BUSNet [54]DeepLab ResNet-10163.2766.4161.3951.7656.8744.13--
CGAN [34]DeepLab ResNet-10164.8668.0462.0751.0355.5144.0651.0151.69
EFN [6]Wide ResNet-10162.7665.6959.6751.5055.2443.01--
ETRIS [53]CLIP ResNet-10171.0674.1166.6662.2368.5152.7960.2860.42
ReSTR [24]ViT-B-1667.2269.364.4555.7860.4448.2754.48-
ETRIS [53]ViT-B-1670.5173.5166.6360.1066.8950.1759.8259.91
CRIS [51]CLIP ResNet-5069.5272.7264.761.3967.152.4859.8760.36
CRIS [51]CLIP ResNet-10170.4773.1866.1062.2768.0853.6859.8760.36
LAVT [56]Swin-B72.7375.8268.7962.1468.3855.1061.2462.10
ReMamber (Ours)Mamba-B74.5476.7470.8965.0070.7857.5363.964.0
+ +Table 3: Ablation on the combination of two scans. We ablated each scan separately, as well as combining them parallel or swap the order. The occurrence of Spatial Scan affects most to the performance. The three variants perform similarly, with the combination of Spatial-Channel slightly better. Results are evaluated on RefCOCO. + +
Channel ScanSpatial ScanvaltestAtestB
mIoUoIoUPr@50Pr@70mIoUoIoUPr@50Pr@70mIoUoIoUPr@50Pr@70
✓-62.360.170.757.264.963.274.361.959.757.566.351.9
-✓70.065.880.569.172.371.083.773.467.563.575.262.7
Parallel71.068.881.970.073.171.884.774.167.964.676.964.5
Spatial-Channel71.468.882.470.873.371.484.974.267.964.776.564.4
Channel-Spatial71.668.482.170.973.371.584.874.568.464.577.164.9
+ +# 4.5 Ablation Study + +Effects of two Scans. Tab. 3 presents a comparative analysis of different scan variants. We provide outcomes of conducting Channel Scan and Spatial Scan independently. Additionally, results from various combinations of these two scans are discussed. The term "Parallel" denotes the concurrent execution of both scans, followed by adding the independent output for fusion. "Channel-Spatial" refers to the original scanning order in Mamba Twister shown in Fig. 2, where the Channel Scan is executed first, followed by the Spatial Scan. "Spatial-Channel" refers to reversing the sequence of the two scans, starting with Spatial Scan and then proceeding to Channel Scan. + +Tab. 3 indicates that utilizing any single scan on its own yields suboptimal results, with the use of Channel Scan alone experiencing the most significant drop in effectiveness. This decline may indicate that a pure Channel Scan alters the data distribution, adversely affecting network stability. Among the three + +hybrid scanning policies, each has its strengths and weaknesses. Overall, the combination of Channel-Spatial Scan appears to offer a considerable advantage. + +Distribution Visualization. The visualization of the distribution of the multi-modal input data is depicted in Fig. 5, where we use PCA to map the feature to a 3D space. The image and text data are indicated in red and blue respectively. + +Fig. 5a illustrates the distribution of the input data. We can observe that the text data forms a linear arrangement within the 3D space, as it is duplicated $H \times W$ times to align with the image size. This suggests that the text instruction should be applied for each pixel within the image. Fig. 5b reveals the data distribution following the application of a Channel Scan. Notably, this process appears to lean towards aggregating different modalities towards a distribution pattern similar with text. Fig. 5c demonstrates the data distribution after a Spatial Scan. Here, the features of the textual and image data disperse, with the modalities intermixed. The Spatial Scan thus reintegrates the previously aligned modalities, distributing them in a manner that reflects their combined influence. + +![](images/02b055069804303b5e24f2058ea3a47655b2a3c38b35a83d3179b7e30772c15b.jpg) +(a) Input data distribution. +Fig.5: Data distribution after Channel Scan and Spacial Scan. Image in red and text data in blue. The Channel Scan tends to aggregate different modalities towards the distribution of textual side. The Spatial Scan reintegrates the previously aligned modalities, distributing them in a manner that reflects their combined influence. + +![](images/1543d6f448324b1e8558da17d3d0b12d66dea7473e96bc6a47962de34e2b760f.jpg) +(b) After Channel Scan. + +![](images/a30446d170cd99d08f990f266e2d0557b59c7b68a1bc722635147663ce466922.jpg) +(c) After Spatial Scan. + +Table 4: Ablation study on global and local interaction. Both global and local interactions are crucial for modality fusion. Results are evaluated on RefCOCO. + +
ImageGlobalLocalvaltestAtestB
mIoUoIoUPr@50Pr@70mIoUoIoUPr@50Pr@70mIoUoIoUPr@50Pr@70
✓✓-69.166.679.566.371.469.882.370.966.363.774.761.0
✓-✓69.967.980.668.272.270.683.973.166.463.875.361.7
✓✓✓71.668.482.170.973.371.584.874.568.464.577.164.9
+ +Effects of Global and Local Features. We also ablate the effect of global and local interactions when forming the hybrid feature cube, formally, the $\tilde{\mathbf{F}}_t$ and $\tilde{\mathbf{F}}_c$ in Eq. (6). The results in Tab. 4 illustrate that the integration of both global and local features significantly enhances the performance. + +# 4.6 Visualization Results + +![](images/9fe158de862ee63af9520a2a7fe28128a96fc5a1ecd1b162273f232956b9b989.jpg) +Fig. 6: Visualization results of our ReMamber and the baseline model LAVT. Our model is able to predict more accurate masks. + +Fig. 6 presents the visualization results of our method compared with the baseline method LAVT [56]. Our ReMamber is capable of producing segmentation results with higher accuracy. + +# 5 Conclusion + +In this study, we introduce ReMamber, a novel architecture utilizing the Mamba framework in Referring Image Segmentation (RIS). It marks a significant advancement in multi-modal understanding. By integrating visual and textual information through innovative Mamba Twister blocks, our approach sets new benchmarks in image-text modality fusion. Achieving competitive results across multiple RIS datasets, our research highlights the potential of Mamba architecture in enhancing the scalability and performance of multi-modal tasks, offering promising directions for future exploration in the field. + +Limitations and Future Works. In our current architecture, the segmentation decoder is constructed by only a few convolutional layers. As shown in Sec. 4.3, the integration of the Cross-Attention mechanism within the Mamba-based architecture demonstrates sub-optimal compatibility, undermining the overall efficacy of our method. In light of these findings, future endeavors will be directed towards the investigation of more sophisticated multi-modal segmentation decoders which best fit Mamba architecture. + +# References + +1. Chen, B., Hu, Z., Ji, Z., Bai, J., Zuo, W.: Position-aware contrastive alignment for referring image segmentation. arXiv preprint arXiv:2212.13419 (2022) 12 +2. Chen, D.J., Jia, S., Lo, Y.C., Chen, H.T., Liu, T.L.: See-through-text grouping for referring image segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7454-7463 (2019) 3, 12 +3. Chen, Y.W., Tsai, Y.H., Wang, T., Lin, Y.Y., Yang, M.H.: Referring expression object segmentation with caption-aware consistency. arXiv preprint arXiv:1910.04748 (2019) 3, 12 +4. Ding, H., Liu, C., Wang, S., Jiang, X.: Vision-language transformer and query generation for referring segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16321-16330 (2021) 1, 2, 3 +5. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 2 +6. Feng, G., Hu, Z., Zhang, L., Lu, H.: Encoder fusion network with co-attention embedding for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15506-15515 (2021) 3, 12 +7. Fu, D.Y., Dao, T., Saab, K.K., Thomas, A.W., Rudra, A., Ré, C.: Hungry hungry hippos: Towards language modeling with state space models. arXiv preprint arXiv:2212.14052 (2022) 2, 4 +8. Gu, A., Dao, T.: Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752 (2023) 2, 4, 5, 7, 9 +9. Gu, A., Dao, T., Ermon, S., Rudra, A., Ré, C.: Hippo: Recurrent memory with optimal polynomial projections. Advances in neural information processing systems 33, 1474-1487 (2020) 4 +10. Gu, A., Goel, K., Gupta, A., Ré, C.: On the parameterization and initialization of diagonal state space models. Advances in Neural Information Processing Systems 35, 35971-35983 (2022) 4 +1. Gu, A., Goel, K., Ré, C.: Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396 (2021) 2, 4 +2. Gu, A., Johnson, I., Goel, K., Saab, K., Dao, T., Rudra, A., Ré, C.: Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems 34, 572-585 (2021) 4 +3. Gupta, A., Gu, A., Berant, J.: Diagonal state spaces are as effective as structured state spaces. Advances in Neural Information Processing Systems 35, 22982-22994 (2022) 4 +4. He, X., Cao, K., Yan, K., Li, R., Xie, C., Zhang, J., Zhou, M.: Pan-mamba: Effective pan-sharpening with state space model. arXiv preprint arXiv:2402.12192 (2024) 2, 4 +5. Hu, R., Rohrbach, M., Darrell, T.: Segmentation from natural language expressions. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14. pp. 108-124. Springer (2016) 1, 3 +6. Hu, Z., Feng, G., Sun, J., Zhang, L., Lu, H.: Bi-directional relationship inferring network for referring image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4424-4433 (2020) 3, 12 + +17. Huang, S., Hui, T., Liu, S., Li, G., Wei, Y., Han, J., Liu, L., Li, B.: Referring image segmentation via cross-modal progressive comprehension. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10488-10497 (2020) 3, 12 +18. Hui, T., Liu, S., Huang, S., Li, G., Yu, S., Zhang, F., Han, J.: Linguistic structure guided context modeling for referring image segmentation. In: European Conference on Computer Vision. pp. 59-75. Springer (2020) 12 +19. Islam, M.M., Bertasius, G.: Long movie clip classification with state-space video models. In: European Conference on Computer Vision. pp. 87-104. Springer (2022) 4 +20. Islam, M.M., Hasan, M., Athrey, K.S., Braskich, T., Bertasius, G.: Efficient movie scene detection using state-space transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18749-18758 (2023) 4 +21. Jing, Y., Kong, T., Wang, W., Wang, L., Li, L., Tan, T.: Locate then segment: A strong pipeline for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9858-9867 (2021) 3 +22. Kamath, A., Singh, M., LeCun, Y., Synnaeve, G., Misra, I., Carion, N.: Mdetr-modulated detection for end-to-end multi-modal understanding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1780-1790 (2021) 1, 3 +23. Kazemzadeh, S., Ordonez, V., Matten, M., Berg, T.: ReferItGame: Referring to objects in photographs of natural scenes. In: Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP) (2014) 9 +24. Kim, N., Kim, D., Lan, C., Zeng, W., Kwak, S.: Restr: Convolution-free referring image segmentation using transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18145-18154 (2022) 1, 2, 3, 12 +25. Li, R., Li, K., Kuo, Y.C., Shu, M., Qi, X., Shen, X., Jia, J.: Referring image segmentation via recurrent refinement networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5745-5753 (2018) 1, 3, 12 +26. Liang, D., Zhou, X., Wang, X., Zhu, X., Xu, W., Zou, Z., Ye, X., Bai, X.: Point-mamba: A simple state space model for point cloud analysis. arXiv preprint arXiv:2402.10739 (2024) 2, 4 +27. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common Objects in Context. ECCV (2014) 9 +28. Liu, C., Ding, H., Jiang, X.: Gres: Generalized referring expression segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23592-23601 (2023) 3 +29. Liu, C., Lin, Z., Shen, X., Yang, J., Lu, X., Yuille, A.: Recurrent multimodal interaction for referring image segmentation. In: Proceedings of the IEEE international conference on computer vision. pp. 1271-1280 (2017) 1, 3, 12 +30. Liu, J., Ding, H., Cai, Z., Zhang, Y., Satzoda, R.K., Mahadevan, V., Manmatha, R.: Polyformer: Referring image segmentation as sequential polygon generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18653-18663 (2023) 3 +31. Liu, J., Yang, H., Zhou, H.Y., Xi, Y., Yu, L., Yu, Y., Liang, Y., Shi, G., Zhang, S., Zheng, H., et al.: Swin-umamba: Mamba-based unet with imagenet-based pretraining. arXiv preprint arXiv:2402.03302 (2024) 2, 4 + +32. Liu, S., Hui, T., Huang, S., Wei, Y., Li, B., Li, G.: Cross-modal progressive comprehension for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(9), 4761-4775 (2021) 12 +33. Liu, Y., Tian, Y., Zhao, Y., Yu, H., Xie, L., Wang, Y., Ye, Q., Liu, Y.: Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166 (2024) 2, 4, 6, 9 +34. Luo, G., Zhou, Y., Ji, R., Sun, X., Su, J., Lin, C.W., Tian, Q.: Cascade grouped attention network for referring expression segmentation. Proceedings of the 28th ACM International Conference on Multimedia (2020) 12 +35. Ma, C., Yang, Y., Ju, C., Zhang, F., Zhang, Y., Wang, Y.: Attrseg: Open-vocabulary semantic segmentation via attribute decomposition-aggregation. NeurIPS (2023) 1 +36. Ma, C., Yang, Y., Wang, Y., Zhang, Y., Xie, W.: Open-vocabulary semantic segmentation with frozen vision-language models. In: British Machine Vision Conference (2022) 1 +37. Ma, J., Li, F., Wang, B.: U-mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722 (2024) 2, 4 +38. Mao, J., Huang, J., Toshev, A., Camburu, O.M., Yuille, A., Murphy, K.: Generation and comprehension of unambiguous object descriptions. CVPR (2015) 9 +39. Margffoy-Tuay, E., Pérez, J.C., Botero, E., Arbeláez, P.: Dynamic multimodal instance segmentation guided by natural language queries. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 630-645 (2018) 3 +40. Mehta, H., Gupta, A., Cutkosky, A., Neyshabur, B.: Long range language modeling via gated state spaces. arXiv preprint arXiv:2206.13947 (2022) 4 +41. Nagaraja, V.K., Morariu, V.I., Davis, L.S.: Modeling context between objects for referring expression understanding. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) ECCV (2016) 9 +42. Perez, E., Strub, F., de Vries, H., Dumoulin, V., Courville, A.C.: Film: Visual reasoning with a general conditioning layer. In: AAAI (2018) 8 +43. Qu, M., Wu, Y., Wei, Y., Liu, W., Liang, X., Zhao, Y.: Learning to segment every referring object point by point. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023) 3 +44. Ruan, J., Xiang, S.: Vm-unet: Vision mamba unet for medical image segmentation. arXiv preprint arXiv:2402.02491 (2024) 2, 4 +45. Shi, H., Li, H., Meng, F., Wu, Q.: Key-word-aware network for referring expression image segmentation. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 38-54 (2018) 3 +46. Smith, J.T., Warrington, A., Linderman, S.W.: Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933 (2022) 2, 4 +47. Tang, J., Zheng, G., Shi, C., Yang, S.: Contrastive grouping with transformer for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23570-23580 (2023) 3 +48. Tang, J., Zheng, G., Shi, C., Yang, S.: Contrastive grouping with transformer for referring image segmentation. In: CVPR (2023) 9 +49. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) 2 +50. Wang, J., Zhu, W., Wang, P., Yu, X., Liu, L., Omar, M., Hamid, R.: Selective structured state-spaces for long-form video understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6387-6397 (2023) 4 + +51. Wang, Z., Lu, Y., Li, Q., Tao, X., Guo, Y., Gong, M., Liu, T.: Cris: Clip-driven referring image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11686-11695 (2022) 1, 3, 12 +52. Xing, Z., Ye, T., Yang, Y., Liu, G., Zhu, L.: Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. arXiv preprint arXiv:2401.13560 (2024) 2, 4 +53. Xu, Z., Chen, Z., Zhang, Y., Song, Y., Wan, X., Li, G.: Bridging vision and language encoders: Parameter-efficient tuning for referring image segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17503-17512 (2023) 12 +54. Yang, S., Xia, M., Li, G., Zhou, H.Y., Yu, Y.: Bottom-up shift and reasoning for referring image segmentation. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11261-11270 (2021) 12 +55. Yang, Y., Ma, C., Ju, C., Zhang, Y., Wang, Y.: Multi-modal prototypes for open-set semantic segmentation. IJCV (2024) 1 +56. Yang, Z., Wang, J., Tang, Y., Chen, K., Zhao, H., Torr, P.H.: Lavt: Language-aware vision transformer for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18155-18165 (2022) 1, 2, 3, 9, 11, 12, 14 +57. Ye, L., Rochan, M., Liu, Z., Wang, Y.: Cross-modal self-attention network for referring image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10502-10511 (2019) 3, 12 +58. Yu, L., Lin, Z., Shen, X., Yang, J., Lu, X., Bansal, M., Berg, T.L.: Matttnet: Modular attention network for referring expression comprehension. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018) 12 +59. Zhang, T., Li, X., Yuan, H., Ji, S., Yan, S.: Point could mamba: Point cloud learning via state space model. arXiv preprint arXiv: 2403.00762 (2024) 2, 4 +60. Zheng, Z., Wu, C.: U-shaped vision mamba for single image dehazing. arXiv preprint arXiv:2402.04139 (2024) 2, 4 +61. Zhu, C., Zhou, Y., Shen, Y., Luo, G., Pan, X., Lin, M., Chen, C., Cao, L., Sun, X., Ji, R.: Seqtr: A simple yet universal network for visual grounding. In: European Conference on Computer Vision. pp. 598-615. Springer (2022) 3 +62. Zhu, L., Liao, B., Zhang, Q., Wang, X., Liu, W., Wang, X.: Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417 (2024) 2, 4 \ No newline at end of file diff --git a/2024/ReMamber_ Referring Image Segmentation with Mamba Twister/images.zip b/2024/ReMamber_ Referring Image Segmentation with Mamba Twister/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..72c5c7a4778392011eb33ce92b41ecbae4d6f32f --- /dev/null +++ b/2024/ReMamber_ Referring Image Segmentation with Mamba Twister/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:58dcc174090a834053e8854a7e1fe6e582fda706d34d659b8c9b2f0dfa455ca5 +size 646957 diff --git a/2024/ReMamber_ Referring Image Segmentation with Mamba Twister/layout.json b/2024/ReMamber_ Referring Image Segmentation with Mamba Twister/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6a43c7d7643e9ed2123d1117b9b70c4987ad24fe --- /dev/null +++ b/2024/ReMamber_ Referring Image Segmentation with Mamba Twister/layout.json @@ -0,0 +1,8568 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "type": "text", + "content": "ReMamber: Referring Image Segmentation with Mamba Twister" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 137, + 167, + 477, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 167, + 477, + 199 + ], + "spans": [ + { + "bbox": [ + 137, + 167, + 477, + 199 + ], + "type": "text", + "content": "Yuhuan Yang\\*1, Chaofan Ma\\*1, Jiangchao Yao1, Zhun Zhong\\*2, Ya Zhang1, and Yanfeng Wang\\*1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 238, + 206, + 374, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 206, + 374, + 219 + ], + "spans": [ + { + "bbox": [ + 238, + 206, + 374, + 219 + ], + "type": "text", + "content": "1 Shanghai Jiao Tong University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 249, + 219, + 364, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 219, + 364, + 229 + ], + "spans": [ + { + "bbox": [ + 249, + 219, + 364, + 229 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 249, + 219, + 364, + 229 + ], + "type": "text", + "content": " University of Nottingham" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 148, + 230, + 465, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 230, + 465, + 241 + ], + "spans": [ + { + "bbox": [ + 148, + 230, + 465, + 241 + ], + "type": "text", + "content": "{yangyuhuan, chaofanna, sunarker, ya_zhang, wangyanfeng622}@sjtu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 254, + 242, + 360, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 242, + 360, + 251 + ], + "spans": [ + { + "bbox": [ + 254, + 242, + 360, + 251 + ], + "type": "text", + "content": "zhunzhong007@gmail.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 273, + 462, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 273, + 462, + 461 + ], + "spans": [ + { + "bbox": [ + 159, + 273, + 462, + 461 + ], + "type": "text", + "content": "Abstract. Referring Image Segmentation (RIS) leveraging transformers has achieved great success on the interpretation of complex visual-language tasks. However, the quadratic computation cost makes it resource-consuming in capturing long-range visual-language dependencies. Fortunately, Mamba addresses this with efficient linear complexity in processing. However, directly applying Mamba to multi-modal interactions presents challenges, primarily due to inadequate channel interactions for the effective fusion of multi-modal data. In this paper, we propose ReMamber, a novel RIS architecture that integrates the power of Mamba with a multi-modal Mamba Twister block. The Mamba Twister explicitly models image-text interaction, and fuses textual and visual features through its unique channel and spatial twisting mechanism. We achieve competitive results on three challenging benchmarks with a simple and efficient architecture. Moreover, we conduct thorough analyses of ReMamber and discuss other fusion designs using Mamba. These provide valuable perspectives for future research. The code has been released at: https://github.com/yyh-rain-song/ReMamber." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 469, + 455, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 469, + 455, + 492 + ], + "spans": [ + { + "bbox": [ + 159, + 469, + 455, + 492 + ], + "type": "text", + "content": "Keywords: Referring Image Segmentation (RIS) " + }, + { + "bbox": [ + 159, + 469, + 455, + 492 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 469, + 455, + 492 + ], + "type": "text", + "content": " Multi-Modal Understanding " + }, + { + "bbox": [ + 159, + 469, + 455, + 492 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 469, + 455, + 492 + ], + "type": "text", + "content": " Mamba Architecture" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 524, + 231, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 524, + 231, + 537 + ], + "spans": [ + { + "bbox": [ + 132, + 524, + 231, + 537 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "type": "text", + "content": "Referring Image Segmentation (RIS) is a crucial yet challenging task in the area of multi-modal understanding [15, 25, 29]. Unlike ordinary image segmentation, RIS involves the identification and segmentation of specific objects in image based on the textual descriptions. This thereby requires the model to be capable of understanding vision-language interactions, which is the core challenge of RIS." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": "Thanks to the powerful attention mechanisms, it has achieved great success for RIS by leveraging transformers to promote the exact recognition of multimodal information. For example, existing works have designed transformer decoder [4,22] or transformer encoder-decoder [24,35,36,51,55,56] to comprehensively fuse visual and linguistic features, which have achieved great progress." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 480, + 293 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 480, + 293 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 480, + 293 + ], + "type": "image", + "image_path": "9d93d1e841052cfac5d99205b13358187664222d816e3f973a0ca52663f8b2c6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 301, + 482, + 357 + ], + "lines": [ + { + "bbox": [ + 131, + 301, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 131, + 301, + 482, + 357 + ], + "type": "text", + "content": "Fig. 1: We propose ReMamber, a novel referring segmentation architecture with Mamba twister. It consists of several Mamba Twister block. Each block contains several visual state space (VSS) layers and a Twisting layer. The Twisting layer first calculates the interaction between image and text, and then forms a hybrid feature cube. Finally, it \"twists\" the feature cube using the Channel and Spatial Scan along each dimension." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 373, + 482, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 373, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 131, + 373, + 482, + 433 + ], + "type": "text", + "content": "Nonetheless, it is a quadratic increase in both computation and memory when applying full attention in transformers [5, 49]. This leads to the limitation for resource-intensive scenarios, for example, in capturing long-range visual-language dependencies. And this is particularly important in the context of large-size images with long textual descriptions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 435, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 435, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 131, + 435, + 482, + 579 + ], + "type": "text", + "content": "Fortunately, recent advances of State Space Models (SSMs) [7,8,11,46] has emerged as promising architectures for solving the above issue. Specifically, Mamba [8] marks a significant advancement for efficient training and inference with linear complexity, which has been incorporated into various visual tasks. However, current efforts primarily pay attention to single-modality settings, such as image classification [33,62], biomedical image segmentation [31,37,44,52], low-level vision [14,60], and point cloud analysis [26,59]. In this paper, we pioneer the exploration of Mamba in multi-modal RIS setting, and identify that the prevalent multi-modality token splicing method in transformers [4,24,56] is no longer effective. Since in Mamba, a fundamental deficiency exists: the interactions are insufficient between channels of different tokens [8], which adversely affects the fusion of multi-modal information in RIS." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 582, + 482, + 666 + ], + "type": "text", + "content": "To address this dilemma, we propose ReMamber, a novel referring segmentation architecture with Mamba twister. As shown in Fig. 1, ReMamber comprises several Mamba Twister blocks, which allow the model to be \"aware\" of the textual context at every spatial location. Each block consists of several visual state space (VSS) layers and a Twisting layer. Specifically, the VSS layers initially extract visual features, and the Twisting layer injects the textual information into the visual modality. The Twisting layer is structured into three critical" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 408, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 408, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 408, + 102 + ], + "type": "text", + "content": "Y. Yang, C. Ma, J. Yao, Z. Zhong, Y. Zhang, and Y. Wang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": "components. (1) A vision-language interaction operation is designed to explicitly capture the fine-grained interactions between modalities. This is achieved by evaluating the similarity between visual and textual tokens then mapping them into a shared feature space, thus generating a multi-modal feature. (2) A hybrid feature cube is then created by concatenating the visual feature, multimodal feature, and global textual feature. This ensures that each visual token receives uniform influence from both local and global contexts, preventing the overshadowing of subtle textual cues by predominant visual features. (3) Lastly, to address the inadequate interaction within channels in Mamba, a twisting mechanism is deployed. This process \"twists\" the feature cube channel- and spatial-wise, thereby enhancing interaction within and across modalities. It is accomplished through two consecutive SSMs scanning along channel and spatial dimensions, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 272, + 481, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 481, + 285 + ], + "type": "text", + "content": "To sum up, the key contributions of our research are highlighted as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 285, + 480, + 392 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 130, + 285, + 479, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 479, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 479, + 308 + ], + "type": "text", + "content": "- We pioneer the exploration of Mamba in referring image segmentation (RIS), demonstrating Mamba's significant potential for multi-modal understanding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 309, + 480, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 480, + 343 + ], + "type": "text", + "content": "- We design a novel framework, ReMamber, that mainly contains several Mamba Twister blocks. This design effectively captures vision-language interactions using the \"twisting mechanism\"." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 345, + 480, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 480, + 392 + ], + "type": "text", + "content": "- We achieve competitive results on three challenging benchmarks. Moreover, we conduct thorough analyses of ReMamber and discuss other vision-language fusion designs using Mamba. These provide valuable perspectives for future research." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 413, + 237, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 237, + 425 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 237, + 425 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 440, + 313, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 313, + 452 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 313, + 452 + ], + "type": "text", + "content": "2.1 Referring Image Segmentation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 461, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 533 + ], + "type": "text", + "content": "Referring image segmentation (RIS) aims to segment the entities in the image following natural language instruction. Early approaches [15, 25, 29, 39] leveraged RNNs or LSTMs to encode linguistic representations, and CNNs to extract spatial features from the image at varying levels. These disparate modalities were then integrated using the multi-modal LSTM [2, 15, 21, 25, 29, 39], attention mechanism [6, 16, 45, 57], cycle-consistency [3], and graph convolution [17]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "Recent trends have shifted towards leveraging transformers for enhanced capturing and fusion of vision-language modalities. MDETR [22] and VLT [4] design a transformer decoder for fusing linguistic and visual features. LAVT [56] adopts Swin Transformer as the visual backbone and incorporate vision-language fusion modules within the visual encoder's final layers. Similar strategies are employed by ReSTR [24] and CRIS [51], which utilize dual transformer encoders for initial modality encoding, followed by feature fusion through a multi-modal transformer encoder or decoder. Other models like PolyFormer [30], SeqTr [61] and [43] also adopt a multi-modal transformer for vision-language fusion, but they output masks as sequences of contour points. Meanwhile, GRES [28] and CGFormer [47] consider transformer queries as region proposals, and regarding" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 185, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 185, + 91, + 447, + 102 + ], + "type": "text", + "content": "ReMamber: Referring Image Segmentation with Mamba Twister" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": "segmentations as proposal-level classification problems. Distinct from all existing works, we introduce a pioneering multi-modal architecture named ReMamber, based on Mamba [8]. This novel approach underscores the untapped potential of Mamba in advancing the field of referring image segmentation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 182, + 384, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 182, + 384, + 194 + ], + "spans": [ + { + "bbox": [ + 131, + 182, + 384, + 194 + ], + "type": "text", + "content": "2.2 State Space Models and Visual Applications" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 201, + 482, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 482, + 345 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 482, + 345 + ], + "type": "text", + "content": "State space models (SSMs), originally derived from control theory, have been effectively combined with deep learning to model the long-range dependencies. Early works like LSSL [12] show potential when combined with HiPPO [9] initialization. S4 [11] was designed to diminish both computational and memory demands associated with state representations. It scales linearly with sequence length, offering a notable advantage over CNNs and transformers. Building on S4, S5 [46] introduces the efficient parallel scan and MIMO SSM, further refining the approach. Later, H3 [7] expanded on these foundations, achieving competitive performance with transformers in language modeling. Recently, Mamba [8] marked a significant advancement with its linear-time inference and efficient training, incorporating a selection mechanism and hardware-aware algorithms building upon prior works [10, 13, 40]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 346, + 482, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 502 + ], + "type": "text", + "content": "With the demonstrated success of SSMs like Mamba in language modeling, researchers have begun to investigate their applicability to visual tasks. Works such as ViS4mer [19], Selective S4 [50] and TranS4mer [20] directly use the S4's ability to capture long-range sequences for understanding inter-frame relations in video classification and detection. However, they still employ Vision Transformers for intra-frame feature extraction. More recently, Vim [62] and VMamba [33] have shown promising results as fully Mamba-based vision backbones for image classification, detection, and segmentation. At the same time, several studies have explored Mamba-based architecture on various vision tasks, such as biomedical image segmentation [31, 37, 44, 52], low-level vision [14, 60], and point cloud analysis [26, 59]. Nevertheless, all above efforts primarily focus on single-modality vision tasks. In this paper, we pioneer the exploration of Mamba-base architecture in multi-modality RIS task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 519, + 208, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 519, + 208, + 533 + ], + "spans": [ + { + "bbox": [ + 131, + 519, + 208, + 533 + ], + "type": "text", + "content": "3 Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 544, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 593 + ], + "type": "text", + "content": "This paper aims to develop a multi-modal Mamba-based architecture for RIS task. In Sec. 3.1, we give the preliminary. From Sec. 3.2 to Sec. 3.5, we describe our architecture from coarse to fine. In Sec. 3.6, we introduce other three variants besides our design. And in Sec. 3.7 we detail the training process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 609, + 323, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 323, + 623 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 323, + 623 + ], + "type": "text", + "content": "3.1 Preliminary: State Space Model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "content": "State Space Model (SSM) is a sequence model inspired by continuous systems. It is designed to capture a mapping relationship between two functions or sequences, expressed as " + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "inline_equation", + "content": "x(t) \\in \\mathbb{R} \\mapsto y(t) \\in \\mathbb{R}" + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "content": ", through a hidden state " + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "inline_equation", + "content": "h(t) \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 90, + 408, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 90, + 408, + 103 + ], + "spans": [ + { + "bbox": [ + 166, + 90, + 408, + 103 + ], + "type": "text", + "content": "Y. Yang, C. Ma, J. Yao, Z. Zhong, Y. Zhang, and Y. Wang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 154 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 154 + ], + "type": "text", + "content": "The evolution of the hidden state over time is governed by specific parameters " + }, + { + "bbox": [ + 132, + 116, + 482, + 154 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 132, + 116, + 482, + 154 + ], + "type": "text", + "content": ", which directs the state evolution, and " + }, + { + "bbox": [ + 132, + 116, + 482, + 154 + ], + "type": "inline_equation", + "content": "\\mathbf{B} \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 132, + 116, + 482, + 154 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 116, + 482, + 154 + ], + "type": "inline_equation", + "content": "\\mathbf{C} \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 132, + 116, + 482, + 154 + ], + "type": "text", + "content": ", which performs the projection. Formally, SSM can be described by the following equations:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 257, + 162, + 481, + 181 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 162, + 481, + 181 + ], + "spans": [ + { + "bbox": [ + 257, + 162, + 481, + 181 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} h ^ {\\prime} (t) = \\mathbf {A} h (t) + \\mathbf {B} x (t), \\\\ (t) = \\mathbf {C} h ^ {\\prime} (t) \\end{array} \\tag {1}", + "image_path": "3a67453e2b53019bdb3799343e08a14e1e4311ec00179565cb8513ceabf41c39.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 260, + 176, + 319, + 189 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 176, + 319, + 189 + ], + "spans": [ + { + "bbox": [ + 260, + 176, + 319, + 189 + ], + "type": "interline_equation", + "content": "y (t) = \\mathbf {C} h (t).", + "image_path": "9001291528ef37cf9cdf08997ea7ee70c93345a6606f9aad8e81fc9ee2a77235.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 192, + 479, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 192, + 479, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 192, + 479, + 237 + ], + "type": "text", + "content": "Discretizing this system introduces a timescale parameter " + }, + { + "bbox": [ + 130, + 192, + 479, + 237 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 130, + 192, + 479, + 237 + ], + "type": "text", + "content": ", which transforms the continuous parameters " + }, + { + "bbox": [ + 130, + 192, + 479, + 237 + ], + "type": "inline_equation", + "content": "\\mathbf{A}, \\mathbf{B}" + }, + { + "bbox": [ + 130, + 192, + 479, + 237 + ], + "type": "text", + "content": " into their discrete counterparts " + }, + { + "bbox": [ + 130, + 192, + 479, + 237 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{A}}, \\overline{\\mathbf{B}}" + }, + { + "bbox": [ + 130, + 192, + 479, + 237 + ], + "type": "text", + "content": ". A common way for this transformation is the zero-order hold (ZOH) approach, defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 231, + 237, + 480, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 237, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 231, + 237, + 480, + 258 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\overline {{\\mathbf {A}}} = \\exp (\\boldsymbol {\\Delta} \\mathbf {A}), \\\\ \\overline {{\\mathbf {B}}} = (\\boldsymbol {\\Delta} \\mathbf {A}) ^ {- 1} (\\boldsymbol {\\Delta} \\mathbf {A}) - \\mathbf {I}) - \\boldsymbol {\\Delta} \\mathbf {B}. \\end{array} \\tag {2}", + "image_path": "f916b69fe045ee5175b978b3096a27b7bc2dc8753a22a7dfa3bd28f814414fde.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 231, + 247, + 480, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 247, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 231, + 247, + 480, + 266 + ], + "type": "interline_equation", + "content": "\\overline {{\\mathbf {B}}} = (\\boldsymbol {\\Delta} \\mathbf {A}) ^ {- 1} (\\exp (\\boldsymbol {\\Delta} \\mathbf {A}) - \\mathbf {I}) \\cdot \\boldsymbol {\\Delta} \\mathbf {B}. \\tag {2}", + "image_path": "3bd496c48204ae8d077e0a543737457bafb0ced400d6b3662fcb5ae67b3553c8.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 270, + 424, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 424, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 424, + 281 + ], + "type": "text", + "content": "Then, the discretized version of this system can be represented as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 265, + 289, + 480, + 309 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 289, + 480, + 309 + ], + "spans": [ + { + "bbox": [ + 265, + 289, + 480, + 309 + ], + "type": "interline_equation", + "content": "h _ {t} = \\overline {{\\mathbf {A}}} h _ {t - 1} + \\overline {{\\mathbf {B}}} x _ {t}, \\tag {3}", + "image_path": "19d049409e2a6dc51f645abead2a6b5a708cebb114cf37a002f2153ab1391ba4.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 266, + 304, + 308, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 304, + 308, + 316 + ], + "spans": [ + { + "bbox": [ + 266, + 304, + 308, + 316 + ], + "type": "interline_equation", + "content": "y _ {t} = \\mathbf {C} h _ {t}.", + "image_path": "2e4bbe843930dc8ce387c59928671f58fcf15088ac4d2a318f62bad9eec20df8.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 323, + 479, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 479, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 479, + 360 + ], + "type": "text", + "content": "Mamba [8], as a variant of SSM, releases the linearity constraint of the original SSM described in Eq. (1), by making " + }, + { + "bbox": [ + 130, + 323, + 479, + 360 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 130, + 323, + 479, + 360 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 323, + 479, + 360 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 130, + 323, + 479, + 360 + ], + "type": "text", + "content": " to be input-dependent, whereas still maintaining the linear complexity during forward process." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 375, + 275, + 386 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 275, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 275, + 386 + ], + "type": "text", + "content": "3.2 Architecture Overview" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "text", + "content": "The pivotal aspect of Referring Image Segmentation (RIS) is the correspondence between image and textual input. To achieve this, we propose ReMamber. Fig. 2 shows the overview of our architecture. The basic block for ReMamber is the Mamba Twister block. It is a multi-modal fusion block that takes both visual and textual feature as input, and outputs the fused multi-modal feature representation. We extract the intermediate feature after each Mamba Twister block, and then feed it into a flexible decoder to generate the final segmentation mask. The decoder is task-invariant, so any downstream architecture can be applied. Below in Sec. 3.3, we will first introduce the key component: Mamba Twister block." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 528, + 272, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 528, + 272, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 528, + 272, + 538 + ], + "type": "text", + "content": "3.3 Mamba Twister Block" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "As shown in Fig. 2, the Mamba Twister Block consists of several visual state space (VSS) layers and a Twisting layer. The VSS layer is designed to process features in the spatial domain. It treats the input feature as a series of image tokens and aims to discern the spatial relationships among these tokens. The Twisting layer aims to inject text condition into the image feature, thus guiding the transformation of image feature. It begins by condensing the textual sequence to compute the cross-correlation between the image and textual modalities. Subsequently, it disseminates this information across each image feature patch via a twisting operation. We'll detail the introduction of VSS layer in Sec. 3.4 and the Twisting layer in Sec. 3.5." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 185, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 185, + 91, + 448, + 103 + ], + "type": "text", + "content": "ReMamber: Referring Image Segmentation with Mamba Twister" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 482, + 224 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 224 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 224 + ], + "type": "image", + "image_path": "d425cdebb12538c725842f9fa2a8d47af8f9c7a3b5d5779d3ec499e6f6b593e6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 148, + 230, + 455, + 320 + ], + "blocks": [ + { + "bbox": [ + 148, + 230, + 455, + 320 + ], + "lines": [ + { + "bbox": [ + 148, + 230, + 455, + 320 + ], + "spans": [ + { + "bbox": [ + 148, + 230, + 455, + 320 + ], + "type": "image", + "image_path": "c66f5d29467cef28a888958bc867e7063e0607952d56740cde5f43b7569efc97.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 331, + 482, + 399 + ], + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 399 + ], + "type": "text", + "content": "Fig. 2: Overview architecture of our ReMamber. The basic block for ReMamber is the Mamba Twister block. It consists of several visual state space (VSS) layers and a Twisting layer. The Twisting layer first constructs hybrid feature cube from text, image, and multi-modal features via channel concatenation. Then, it \"twists\" the cube by Channel Scan and Spatial Scan. We extract intermediate features after each Mamba Twister block, and feed it into a flexible decoder for final segmentation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 406, + 357, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 406, + 357, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 406, + 357, + 419 + ], + "type": "text", + "content": "3.4 VSS Layer for Spatial Data Processing" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 521 + ], + "type": "text", + "content": "Since the SSM is initially designed for processing temporal or causal data, which can not effectively process non-causal data types, i.e., 2-dimensional image in our case. To solve this issue, we adopt the Cross-Scan-Module proposed in [33], which unfolds image patches into sequences and then scans them in four distinct directions, ensuring that the information from all pixels is integrated during feature transformation. We replace the scanning operation in vanilla Mamba block with CSM, forming a visual state space (VSS) layer for our spatial feature transformation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 538, + 357, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 538, + 357, + 551 + ], + "spans": [ + { + "bbox": [ + 132, + 538, + 357, + 551 + ], + "type": "text", + "content": "3.5 Twisting Layer for Multi-modal Fusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Here we introduce the detailed design of our Twisting layer. Intuitively, the Twisting layer first constructs a feature cube by arranging information from different modalities orderly. Then, pieces of information intertwine with each other when the cube is twisted through two SSM layers along different axis, thereby achieving modality fusion. Formally, given image feature " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i\\in \\mathbb{R}^{h\\times w\\times C_i}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " and text feature " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t\\in \\mathbb{R}^{L\\times C_t}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ", the Mamba twister aims to learn a function " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathbf{F}_i,\\mathbf{F}_t)\\mapsto \\tilde{\\mathbf{F}}\\in \\mathbb{R}^{h\\times w\\times C_o}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " by first forming a hybrid multi-modal feature cube and then conduct two SSM scans on it. " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "C_i,C_t" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "C_o" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " are the dimensions of visual, textual, and output feature; " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "h,w" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " are height, weight, and length." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 408, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 408, + 103 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 408, + 103 + ], + "type": "text", + "content": "Y. Yang, C. Ma, J. Yao, Z. Zhong, Y. Zhang, and Y. Wang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "Forming the Hybrid Feature Cube. To explicitly build the image-text correspondence, we design a novel vision-language interaction operation and then form a hybrid feature cube. This allows more effective modality fusion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "content": "To be specific, we formulated the vision-language interaction operation into two manners: global interaction and local interaction. The global interaction treats the text sequence as a whole, which means that all image patches should be aware of the semantic meaning of the text expression. We pool a global representation " + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t^{\\mathrm{CLS}}\\in \\mathbb{R}^{C_t}" + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "content": " from the text sequence " + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t" + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "content": " to convey this information. The " + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t^{\\mathrm{CLS}}" + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "content": " is further expanded to the same size as the image feature. Formally:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 235, + 230, + 480, + 244 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 230, + 480, + 244 + ], + "spans": [ + { + "bbox": [ + 235, + 230, + 480, + 244 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {F}} _ {t} = \\operatorname {E x p a n d} \\left(\\mathbf {F} _ {t} ^ {\\mathrm {C L S}}\\right) \\in \\mathbb {R} ^ {h \\times w \\times C _ {t}}, \\tag {4}", + "image_path": "5ee9e9eb2e422831c6f932fdc0b40ff97004e4b19f6c319ec5e396e5ded63ba9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 251, + 479, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 479, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 479, + 274 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 251, + 479, + 274 + ], + "type": "inline_equation", + "content": "\\operatorname{Expand}(\\cdot)" + }, + { + "bbox": [ + 130, + 251, + 479, + 274 + ], + "type": "text", + "content": " denotes the operation that expands the input tensor to the same size as the image feature." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 275, + 480, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 275, + 480, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 480, + 344 + ], + "type": "text", + "content": "However, a global representation may not be enough to capture the intricate relationships between different modalities. For example, some words such as color, shape or location may be more relevant to certain image patches than others. The local interaction aims to capture such correlation in a fine-grained level. Formally, we calculate the local interaction map " + }, + { + "bbox": [ + 130, + 275, + 480, + 344 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_c" + }, + { + "bbox": [ + 130, + 275, + 480, + 344 + ], + "type": "text", + "content": " using matrix multiplication:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 233, + 345, + 480, + 358 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 345, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 233, + 345, + 480, + 358 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {c} = \\mathbf {F} _ {i} \\mathbf {W} _ {i} \\cdot \\left(\\mathbf {F} _ {t} \\mathbf {W} _ {t}\\right) ^ {T} \\in \\mathbb {R} ^ {h \\times w \\times L}, \\tag {5}", + "image_path": "378d73de61e1445d8077e2a4a2ee3c7d2b269a62d0ff811e67940eb00078adf9.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 361, + 480, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 480, + 411 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 361, + 480, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_i\\in \\mathbb{R}^{C_i\\times C_c}" + }, + { + "bbox": [ + 130, + 361, + 480, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 361, + 480, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_t\\in \\mathbb{R}^{C_t\\times C_c}" + }, + { + "bbox": [ + 130, + 361, + 480, + 411 + ], + "type": "text", + "content": " are learnable parameters. To enhance feature processing in high dimension, we use a single convolutional layer to transform " + }, + { + "bbox": [ + 130, + 361, + 480, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_c" + }, + { + "bbox": [ + 130, + 361, + 480, + 411 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 130, + 361, + 480, + 411 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_c\\in \\mathbb{R}^{h\\times w\\times C_c}" + }, + { + "bbox": [ + 130, + 361, + 480, + 411 + ], + "type": "text", + "content": ". We then concatenate " + }, + { + "bbox": [ + 130, + 361, + 480, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i" + }, + { + "bbox": [ + 130, + 361, + 480, + 411 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 361, + 480, + 411 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_t" + }, + { + "bbox": [ + 130, + 361, + 480, + 411 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 361, + 480, + 411 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_c" + }, + { + "bbox": [ + 130, + 361, + 480, + 411 + ], + "type": "text", + "content": " along the channel dimension to form the hybrid feature cube. Formally:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 219, + 417, + 480, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 417, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 219, + 417, + 480, + 430 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {\\text {c u b e}} = \\left[ \\mathbf {F} _ {i}, \\tilde {\\mathbf {F}} _ {t}, \\tilde {\\mathbf {F}} _ {c} \\right] \\in \\mathbb {R} ^ {h \\times w \\times \\left(C _ {i} + C _ {t} + C _ {c}\\right)}. \\tag {6}", + "image_path": "027fdc097f4e9b32f9890d09d77e0606dbf275b5554b284e929a427a600e85fe.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 445, + 479, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 479, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 479, + 493 + ], + "type": "text", + "content": "Twisting the Hybrid Feature Cube. Generally, the scanning operation for vanilla SSM is almost independent across channels [8]. This is not enough for feature communication, especially when information from different modalities is arranged along the channel dimension, i.e., our hybrid feature cube." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 494, + 480, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 480, + 587 + ], + "type": "text", + "content": "To foster the interactions within and across modalities, here we design a twisting mechanism to \"twist\" the hybrid feature cube along different axis. Specifically, it is composed of the Channel Scan and Spacial Scan successively. The Channel Scan first treats the channel as an ordered sequence, and learns to communicate cross channels, thus foster the modality fusion. Then, the Spatial Scan operates on the spatial dimension, and learns to communicate information cross patches within each modality separately. Formally, this process can be written as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 222, + 590, + 480, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 590, + 480, + 602 + ], + "spans": [ + { + "bbox": [ + 222, + 590, + 480, + 602 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {\\text {o u t}} = \\operatorname {S S M} _ {\\text {s p a t i a l}} \\left(\\operatorname {S S M} _ {\\text {c h a n n e l}} \\left(\\mathbf {F} _ {\\text {c u b e}}\\right)\\right), \\tag {7}", + "image_path": "d328e29ce1d85ca3fb51cf74f722845b03f39644018d72f0f036e9cdb23be024.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 605, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 479, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 605, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{SSM}_{\\mathrm{channel}}" + }, + { + "bbox": [ + 130, + 605, + 479, + 666 + ], + "type": "text", + "content": " refers to the Channel Scan. It is an SSM layer that treats the concatenated feature as an ordered sequence through channel dimension, and presents 1-D selective scan along the channel. While " + }, + { + "bbox": [ + 130, + 605, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{SSM}_{\\mathrm{spatial}}" + }, + { + "bbox": [ + 130, + 605, + 479, + 666 + ], + "type": "text", + "content": " denotes the Spatial Scan. It is an VSS layer and presents 2D selective scan along the two spatial dimension. " + }, + { + "bbox": [ + 130, + 605, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{out}}" + }, + { + "bbox": [ + 130, + 605, + 479, + 666 + ], + "type": "text", + "content": " is then feed into the next layer for further process." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 185, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 185, + 91, + 448, + 102 + ], + "type": "text", + "content": "ReMamber: Referring Image Segmentation with Mamba Twister" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 192, + 114, + 402, + 135 + ], + "blocks": [ + { + "bbox": [ + 192, + 114, + 402, + 135 + ], + "lines": [ + { + "bbox": [ + 192, + 114, + 402, + 135 + ], + "spans": [ + { + "bbox": [ + 192, + 114, + 402, + 135 + ], + "type": "image", + "image_path": "f6619b0a50cf049cff6d6f9dfc3fe3b7c5b43fd0d01c0616a069983123bb7415.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 141, + 137, + 474, + 279 + ], + "blocks": [ + { + "bbox": [ + 141, + 137, + 474, + 279 + ], + "lines": [ + { + "bbox": [ + 141, + 137, + 474, + 279 + ], + "spans": [ + { + "bbox": [ + 141, + 137, + 474, + 279 + ], + "type": "image", + "image_path": "8f927a0c10a71fd6e58ee3e0efcfedee13aa408c9945f262d0adba7252f423a9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 286, + 482, + 331 + ], + "lines": [ + { + "bbox": [ + 130, + 286, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 286, + 482, + 331 + ], + "type": "text", + "content": "Fig.3: Other multi-modal fusion designs. (a) In-context Conditioning appends text tokens ahead of image tokens. (b) Attention-based Conditioning utilizes cross-attention mechanism for modality fusion. (c) Norm Adaptation learns a scale and bias for the model's normalization layers." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 341, + 437, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 341, + 437, + 354 + ], + "spans": [ + { + "bbox": [ + 131, + 341, + 437, + 354 + ], + "type": "text", + "content": "3.6 Discussion: Other Variants of Modality Fusion Designs" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 361, + 482, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 409 + ], + "type": "text", + "content": "Besides the proposed Twisting layer, we also explore other variants of multimodal fusion designs. Fig. 3 shows the structure of three other different variants. We here discuss these variants, and later in Sec. 4.3, we analyze the results for these variants." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 410, + 482, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 482, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 482, + 470 + ], + "type": "text", + "content": "In-context Conditioning. As shown in Fig. 3 (a), in this variant, we simply append the text sequence before the image feature, forming a longer sequence than originally being processed. In this way, the model is able to fuse image features with the previous text as a context. This is the most straightforward way to allow the model aware of the text condition." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 470, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 482, + 529 + ], + "type": "text", + "content": "Attention-based Conditioning. Attention mechanism is usually a strong baseline for sequence modeling. Here we utilize the cross-attention mechanism for two modality fusion, as shown in Fig. 3 (b), To be specific, we use the image feature as query, and the text feature as key and value, hoping to capture the correlation between image and text." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 530, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 530, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 530, + 482, + 578 + ], + "type": "text", + "content": "Norm Adaptation. In this variant, we integrate the text input by adapting the scale and bias after norm layer using FiLM [42]. We first pool a global representation from text sequence, and then use a linear projection to transform it into the scale and shift in layer normalization." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 597, + 238, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 597, + 238, + 609 + ], + "spans": [ + { + "bbox": [ + 131, + 597, + 238, + 609 + ], + "type": "text", + "content": "3.7 Model training" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": "The proposed ReMamber is a general framework for multi-modal fusion. Therefore, there is no restriction on the downstream decoder design nor the loss function. Here we use a simple convolution-based decoder for simplicity. The entire network is trained in an end-to-end manner." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 90, + 409, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 90, + 409, + 103 + ], + "spans": [ + { + "bbox": [ + 165, + 90, + 409, + 103 + ], + "type": "text", + "content": "Y. Yang, C. Ma, J. Yao, Z. Zhong, Y. Zhang, and Y. Wang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 269, + 150 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 269, + 150 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 269, + 150 + ], + "type": "text", + "content": "4.1 Datasets and Metrics" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 158, + 482, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 158, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 158, + 482, + 289 + ], + "type": "text", + "content": "Dataset. To assess the efficacy of our proposed approach, we executed experiments across three widely recognized datasets tailored for the referring image segmentation (RIS) task: RefCOCO, RefCOCO+ [23], and G-Ref [23,38,41]. RefCOCO [23] is a prominent dataset in the RIS domain, comprising 19,994 images and 142,210 referring expressions associated with 50,000 objects derived from the MSCOCO [27] dataset through a two-player game. RefCOCO+ enhances the challenge by excluding expressions with absolute location references, featuring 141,564 expressions for 49,856 objects across 19,992 images. G-Ref [38,41] enriches the dataset diversity with 104,560 referring expressions for 54,882 objects in 26,711 images, showcasing an average sentence length of 8.4 words with a heightened focus on location and appearance descriptions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 289, + 481, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 289, + 481, + 338 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 481, + 338 + ], + "type": "text", + "content": "Evaluation Metrics. In line with preceding studies [48], we adopt mIoU and oIoU as the main metrics. Furthermore, we incorporate the metric Precision@X " + }, + { + "bbox": [ + 131, + 289, + 481, + 338 + ], + "type": "inline_equation", + "content": "(\\mathrm{X} \\in \\{50, 60, 70, 90\\})" + }, + { + "bbox": [ + 131, + 289, + 481, + 338 + ], + "type": "text", + "content": " for a more comprehensive evaluation, whereas Percision@X means the percentage of test images with an IoU score high than X%." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 354, + 279, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 279, + 366 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 279, + 366 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 373, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 456 + ], + "type": "text", + "content": "Our model architecture is developed upon the foundations of Mamba [8] and VMamba [33]. We adopt ImageNet pre-trained weights as initialization, and train the model in an end-to-end manner. We set the input image resolution to 480 when compared with the state-of-the-art methods in Tab. 2, following [48,56] for fair comparison. For other experiments (Tabs. 1 and 3 to 4), we set the image resolution to 256 for faster training and ablation study without changing its property. For details about the training settings, please refer to our code." + }, + { + "bbox": [ + 130, + 373, + 482, + 456 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 473, + 451, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 473, + 451, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 473, + 451, + 486 + ], + "type": "text", + "content": "4.3 Evaluation on Other Variants of Modality Fusion Designs" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 491, + 482, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 599 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 599 + ], + "type": "text", + "content": "As discussed in Sec. 3.6, besides our Mamba Twister, we also provide three other variant architectures for modality fusion, including In-context Conditioning, Attention-based Conditioning and Norm Adaptation. To make a fair comparison, for all architectures, we initialize their common part (VMamba-based parameters) with the same checkpoint. The difference is that the Twisting layer is replaced by the corresponding fusion methods. The parameters of fusion modules are nearly equal, i.e., {103, 113, 118, 116} KB for {In-Context, Attention, Norm Adaptation, Twister}. Here we provide a comprehensive analysis between these architectures, showing the advantages and disadvantages of each variant." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 600, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 600, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 600, + 482, + 647 + ], + "type": "text", + "content": "Tab. 1 shows the comparison result. The Mamba Twister consistently outperforms the other variants across all metrics and datasets, indicating its superior capability in capturing and integrating contextual information for more accurate segmentation." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 185, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 185, + 91, + 447, + 102 + ], + "type": "text", + "content": "ReMamber: Referring Image Segmentation with Mamba Twister" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 652, + 331, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 652, + 331, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 652, + 331, + 666 + ], + "type": "text", + "content": "1 https://github.com/yyh-rain-song/ReMamber" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 178, + 482, + 414 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 1: Comparison with other modality fusion variants. \"Attention\" means for \"Attention-based Conditioning\", \"In-Context\" for \"In-Context Conditioning\", and \"Adaptation\" for \"Norm Adaptation\". Mamba Twister steadily outperforms other variants across all metrics and datasets, indicating its superior capability in capturing and integrating contextual information for more accurate segmentation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 178, + 482, + 414 + ], + "lines": [ + { + "bbox": [ + 132, + 178, + 482, + 414 + ], + "spans": [ + { + "bbox": [ + 132, + 178, + 482, + 414 + ], + "type": "table", + "html": "
Dataset: RefCOCO
VariantsmIoUvalPr@70mIoUtestAmIoUtestB
oIoUPr@50Pr@70oIoUPr@50Pr@70oIoUPr@50
Attention65.362.375.160.667.564.479.165.461.758.069.653.4
In-Context69.165.979.767.571.268.982.271.866.262.875.061.7
Adaptation70.267.080.769.672.370.283.173.266.862.875.163.1
Mamba Twister71.668.482.170.973.371.584.874.568.464.577.164.9
Dataset: RefCOCO+
VariantsmIoUvalPr@70mIoUtestAmIoUtestB
oIoUPr@50oIoUPr@50Pr@70oIoUPr@50
Attention54.049.760.745.558.755.868.552.946.941.950.636.6
In-Context58.454.066.353.963.259.472.260.551.547.556.444.5
Adaptation60.355.268.257.364.259.973.462.953.948.759.047.2
Mamba Twister61.657.370.058.565.862.175.564.954.049.959.447.8
Dataset: G-Ref
VariantsmIoUvalPr@70Pr@90mIoUoIoUtest
oIoUPr@50Pr@50Pr@60Pr@70Pr@90
Attention50.549.254.246.036.78.150.850.857.349.838.66.4
In-Context54.853.459.552.343.413.355.554.860.753.645.313.3
Adaptation59.356.766.761.053.718.458.656.665.559.852.818.2
Mamba Twister61.158.068.462.855.218.561.259.069.063.455.518.7
", + "image_path": "65969b4ff9432e7e8488ae623456f22498a614a09902ee730c515f71e96a5466.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 425, + 483, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 483, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 483, + 555 + ], + "type": "text", + "content": "Surprisingly, despite widely used in transformers, the Attention-based Conditioning performs poorly in our task. This suggests that the cross-attention mechanism is not inherently suitable for Mamba architecture. This may be due to a fundamental discrepancy between the two systems. To be specific, (1) the Mamba model is predicted on ordered sequences that exhibit strict sequential dependencies, where the state at a given time point " + }, + { + "bbox": [ + 130, + 425, + 483, + 555 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 130, + 425, + 483, + 555 + ], + "type": "text", + "content": " is determined by the preceding time point " + }, + { + "bbox": [ + 130, + 425, + 483, + 555 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 425, + 483, + 555 + ], + "type": "text", + "content": " and a hidden state; (2) In contrast, cross-attention mechanisms treats all tokens within a sequence equally. This discrepancy may undermine the Mamba model's ability to structurally model sequences, as the cross-attention mechanism does not preserve the sequential integrity and the hierarchical dependencies essential for the model's operation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 556, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 556, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 556, + 482, + 628 + ], + "type": "text", + "content": "Serving as a straightforward baseline, In-context Conditioning performs sub-optimal. This may because it models image-text interaction in an implicit way. In our situation, the length of the image feature is much larger than the textual side, textual information may be diluted during the forward process. Besides, this operation does not differentiate the textual information between image patches, and would be insufficient for multi-modal fusion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Finally, Norm Adaptation with explicit image-text interaction modeling appears to be a strong baseline. It achieves better performance than other two variants. However, when calculating scale and bias, it uses only a pooled vector as" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 90, + 408, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 90, + 408, + 103 + ], + "spans": [ + { + "bbox": [ + 166, + 90, + 408, + 103 + ], + "type": "text", + "content": "Y. Yang, C. Ma, J. Yao, Z. Zhong, Y. Zhang, and Y. Wang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 140 + ], + "type": "text", + "content": "textual representation instead of the entire sequence. This potentially results in the loss of information, making it less effective compared to our Mamba Twister." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 158, + 483, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 158, + 483, + 302 + ], + "spans": [ + { + "bbox": [ + 130, + 158, + 483, + 302 + ], + "type": "text", + "content": "Case-study using Attention Maps. In Fig. 4, we further visualized the attention maps in Attention-based Conditioning as well as the local interaction map in Mamba Twister defined in Eq. (5). The four images in Fig. 4(b) are arranged from left to right with the network going from shallow to deep. It can be observed that in the shallower layers, Mamba Twister focuses more on the lower-level features of the images such as edges. As the layer goes deeper, the cost map mostly concentrating on the target object, indicating that Mamba Twister can progressively guide the image feature towards the target described by the language. In contrast, though the Attention-based Conditioning variant is able to locate the target, its attention map performs poorly. This may suggest that the cross-attention mechanism struggles to accurately capture the correct context information, and the two modalities are not truly fused together." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 320, + 482, + 459 + ], + "blocks": [ + { + "bbox": [ + 133, + 320, + 482, + 459 + ], + "lines": [ + { + "bbox": [ + 133, + 320, + 482, + 459 + ], + "spans": [ + { + "bbox": [ + 133, + 320, + 482, + 459 + ], + "type": "image", + "image_path": "8b7666cefd151e855403be655d58509843b7dc2561f127aaaf01cdfc9768a29b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 466, + 482, + 522 + ], + "lines": [ + { + "bbox": [ + 130, + 466, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 466, + 482, + 522 + ], + "type": "text", + "content": "Fig. 4: Cross-Attention map (up) and our local interaction map (down) comparison. Though both methods are able to predict target correctly, the cross-attention maps don't show correct image-text correlation, while ours are able to capture this relationship accurately, indicating that Mamba Twister is able to gradually fusing the two modality." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 549, + 358, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 549, + 358, + 562 + ], + "spans": [ + { + "bbox": [ + 131, + 549, + 358, + 562 + ], + "type": "text", + "content": "4.4 Comparison with the State-of-the-Arts" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "Tab. 2 shows the comparison of our ReMamber with state-of-the-art methods on RefCOCO, RefCOCO+, and G-Ref datasets. ReMamber outperforms all other methods across all datasets, demonstrating its superior capability in capturing and integrating contextual information for more accurate segmentation. In particular, Mamba outperforms previous Swin-based methods such as LAVT [56] by a considerable margin, indicating the remarkable capability of the Mamba-based architecture in segmentation tasks. and the efficiency of the newly introduced Mamba Twister." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 185, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 185, + 91, + 448, + 102 + ], + "type": "text", + "content": "ReMamber: Referring Image Segmentation with Mamba Twister" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 147, + 482, + 344 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": "Table 2: Comparison with state-of-the-art methods on RefCOCO, RefCOCO+, and G-Ref. The metric of oIoU is reported. Best results are in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 147, + 482, + 344 + ], + "lines": [ + { + "bbox": [ + 133, + 147, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 133, + 147, + 482, + 344 + ], + "type": "table", + "html": "
ModelBackboneRefCOCORefCOCO+G-Ref
valtestAtestBvaltestAtestBvalUtestU
PCAN [1]ResNet-5069.5171.6464.1858.2563.6848.8959.9860.8
MAAttNet [58]MaskRCNN ResNet-10156.5162.3751.746.6752.3940.0847.6448.61
RMI [29]DeepLab ResNet-10145.1845.6945.5729.8630.4829.50--
RRN [25]DeepLab ResNet-10155.3357.2653.9539.7542.1536.11--
CMSA [57]DeepLab ResNet-10158.3260.6155.0943.7647.637.89--
CAC [3]DeepLab ResNet-10158.9061.7753.81---46.3746.95
STEP [2]DeepLab ResNet-10160.0463.4657.9748.1952.3340.41--
BRINet [16]DeepLab ResNet-10160.9862.9959.2148.1752.3242.11--
CMPC [17]DeepLab ResNet-10161.3664.5359.6449.5653.4443.23--
LSCM [18]DeepLab ResNet-10161.4764.9959.5549.3453.1243.50--
CMPC+ [32]DeepLab ResNet-10162.4765.0860.8250.2554.0443.47--
BUSNet [54]DeepLab ResNet-10163.2766.4161.3951.7656.8744.13--
CGAN [34]DeepLab ResNet-10164.8668.0462.0751.0355.5144.0651.0151.69
EFN [6]Wide ResNet-10162.7665.6959.6751.5055.2443.01--
ETRIS [53]CLIP ResNet-10171.0674.1166.6662.2368.5152.7960.2860.42
ReSTR [24]ViT-B-1667.2269.364.4555.7860.4448.2754.48-
ETRIS [53]ViT-B-1670.5173.5166.6360.1066.8950.1759.8259.91
CRIS [51]CLIP ResNet-5069.5272.7264.761.3967.152.4859.8760.36
CRIS [51]CLIP ResNet-10170.4773.1866.1062.2768.0853.6859.8760.36
LAVT [56]Swin-B72.7375.8268.7962.1468.3855.1061.2462.10
ReMamber (Ours)Mamba-B74.5476.7470.8965.0070.7857.5363.964.0
", + "image_path": "dbd2380162b1df3e22cc7b5d74505330d85f752abf1e03e5a2a48a385eaf8ec0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 133, + 409, + 482, + 477 + ], + "blocks": [ + { + "bbox": [ + 132, + 355, + 482, + 401 + ], + "lines": [ + { + "bbox": [ + 132, + 355, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 132, + 355, + 482, + 401 + ], + "type": "text", + "content": "Table 3: Ablation on the combination of two scans. We ablated each scan separately, as well as combining them parallel or swap the order. The occurrence of Spatial Scan affects most to the performance. The three variants perform similarly, with the combination of Spatial-Channel slightly better. Results are evaluated on RefCOCO." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 409, + 482, + 477 + ], + "lines": [ + { + "bbox": [ + 133, + 409, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 133, + 409, + 482, + 477 + ], + "type": "table", + "html": "
Channel ScanSpatial ScanvaltestAtestB
mIoUoIoUPr@50Pr@70mIoUoIoUPr@50Pr@70mIoUoIoUPr@50Pr@70
✓-62.360.170.757.264.963.274.361.959.757.566.351.9
-✓70.065.880.569.172.371.083.773.467.563.575.262.7
Parallel71.068.881.970.073.171.884.774.167.964.676.964.5
Spatial-Channel71.468.882.470.873.371.484.974.267.964.776.564.4
Channel-Spatial71.668.482.170.973.371.584.874.568.464.577.164.9
", + "image_path": "804f4951d1ca96403a286cebe664842dae69edb4b2ed4baeb11c8582798b7cea.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 487, + 238, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 238, + 499 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 238, + 499 + ], + "type": "text", + "content": "4.5 Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 509, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 481, + 616 + ], + "type": "text", + "content": "Effects of two Scans. Tab. 3 presents a comparative analysis of different scan variants. We provide outcomes of conducting Channel Scan and Spatial Scan independently. Additionally, results from various combinations of these two scans are discussed. The term \"Parallel\" denotes the concurrent execution of both scans, followed by adding the independent output for fusion. \"Channel-Spatial\" refers to the original scanning order in Mamba Twister shown in Fig. 2, where the Channel Scan is executed first, followed by the Spatial Scan. \"Spatial-Channel\" refers to reversing the sequence of the two scans, starting with Spatial Scan and then proceeding to Channel Scan." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Tab. 3 indicates that utilizing any single scan on its own yields suboptimal results, with the use of Channel Scan alone experiencing the most significant drop in effectiveness. This decline may indicate that a pure Channel Scan alters the data distribution, adversely affecting network stability. Among the three" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 90, + 408, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 90, + 408, + 103 + ], + "spans": [ + { + "bbox": [ + 166, + 90, + 408, + 103 + ], + "type": "text", + "content": "Y. Yang, C. Ma, J. Yao, Z. Zhong, Y. Zhang, and Y. Wang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "hybrid scanning policies, each has its strengths and weaknesses. Overall, the combination of Channel-Spatial Scan appears to offer a considerable advantage." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 156, + 479, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 156, + 479, + 203 + ], + "spans": [ + { + "bbox": [ + 130, + 156, + 479, + 203 + ], + "type": "text", + "content": "Distribution Visualization. The visualization of the distribution of the multi-modal input data is depicted in Fig. 5, where we use PCA to map the feature to a 3D space. The image and text data are indicated in red and blue respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 204, + 480, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 204, + 480, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 204, + 480, + 323 + ], + "type": "text", + "content": "Fig. 5a illustrates the distribution of the input data. We can observe that the text data forms a linear arrangement within the 3D space, as it is duplicated " + }, + { + "bbox": [ + 130, + 204, + 480, + 323 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 130, + 204, + 480, + 323 + ], + "type": "text", + "content": " times to align with the image size. This suggests that the text instruction should be applied for each pixel within the image. Fig. 5b reveals the data distribution following the application of a Channel Scan. Notably, this process appears to lean towards aggregating different modalities towards a distribution pattern similar with text. Fig. 5c demonstrates the data distribution after a Spatial Scan. Here, the features of the textual and image data disperse, with the modalities intermixed. The Spatial Scan thus reintegrates the previously aligned modalities, distributing them in a manner that reflects their combined influence." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 149, + 344, + 223, + 415 + ], + "blocks": [ + { + "bbox": [ + 149, + 344, + 223, + 415 + ], + "lines": [ + { + "bbox": [ + 149, + 344, + 223, + 415 + ], + "spans": [ + { + "bbox": [ + 149, + 344, + 223, + 415 + ], + "type": "image", + "image_path": "02b055069804303b5e24f2058ea3a47655b2a3c38b35a83d3179b7e30772c15b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 415, + 235, + 424 + ], + "lines": [ + { + "bbox": [ + 136, + 415, + 235, + 424 + ], + "spans": [ + { + "bbox": [ + 136, + 415, + 235, + 424 + ], + "type": "text", + "content": "(a) Input data distribution." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 434, + 479, + 477 + ], + "lines": [ + { + "bbox": [ + 130, + 434, + 479, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 479, + 477 + ], + "type": "text", + "content": "Fig.5: Data distribution after Channel Scan and Spacial Scan. Image in red and text data in blue. The Channel Scan tends to aggregate different modalities towards the distribution of textual side. The Spatial Scan reintegrates the previously aligned modalities, distributing them in a manner that reflects their combined influence." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 269, + 344, + 344, + 414 + ], + "blocks": [ + { + "bbox": [ + 269, + 344, + 344, + 414 + ], + "lines": [ + { + "bbox": [ + 269, + 344, + 344, + 414 + ], + "spans": [ + { + "bbox": [ + 269, + 344, + 344, + 414 + ], + "type": "image", + "image_path": "1543d6f448324b1e8558da17d3d0b12d66dea7473e96bc6a47962de34e2b760f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 263, + 415, + 350, + 424 + ], + "lines": [ + { + "bbox": [ + 263, + 415, + 350, + 424 + ], + "spans": [ + { + "bbox": [ + 263, + 415, + 350, + 424 + ], + "type": "text", + "content": "(b) After Channel Scan." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 390, + 344, + 465, + 414 + ], + "blocks": [ + { + "bbox": [ + 390, + 344, + 465, + 414 + ], + "lines": [ + { + "bbox": [ + 390, + 344, + 465, + 414 + ], + "spans": [ + { + "bbox": [ + 390, + 344, + 465, + 414 + ], + "type": "image", + "image_path": "a30446d170cd99d08f990f266e2d0557b59c7b68a1bc722635147663ce466922.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 387, + 415, + 468, + 424 + ], + "lines": [ + { + "bbox": [ + 387, + 415, + 468, + 424 + ], + "spans": [ + { + "bbox": [ + 387, + 415, + 468, + 424 + ], + "type": "text", + "content": "(c) After Spatial Scan." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 133, + 555, + 481, + 609 + ], + "blocks": [ + { + "bbox": [ + 130, + 522, + 479, + 544 + ], + "lines": [ + { + "bbox": [ + 130, + 522, + 479, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 479, + 544 + ], + "type": "text", + "content": "Table 4: Ablation study on global and local interaction. Both global and local interactions are crucial for modality fusion. Results are evaluated on RefCOCO." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 555, + 481, + 609 + ], + "lines": [ + { + "bbox": [ + 133, + 555, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 133, + 555, + 481, + 609 + ], + "type": "table", + "html": "
ImageGlobalLocalvaltestAtestB
mIoUoIoUPr@50Pr@70mIoUoIoUPr@50Pr@70mIoUoIoUPr@50Pr@70
✓✓-69.166.679.566.371.469.882.370.966.363.774.761.0
✓-✓69.967.980.668.272.270.683.973.166.463.875.361.7
✓✓✓71.668.482.170.973.371.584.874.568.464.577.164.9
", + "image_path": "091bac495cf6387f5aa780b983c519b8071915acf08813f5d9d17a0fdb0f7fce.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "content": "Effects of Global and Local Features. We also ablate the effect of global and local interactions when forming the hybrid feature cube, formally, the " + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_t" + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_c" + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "content": " in Eq. (6). The results in Tab. 4 illustrate that the integration of both global and local features significantly enhances the performance." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 185, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 185, + 91, + 447, + 102 + ], + "type": "text", + "content": "ReMamber: Referring Image Segmentation with Mamba Twister" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 268, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 268, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 268, + 128 + ], + "type": "text", + "content": "4.6 Visualization Results" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 138, + 149, + 483, + 361 + ], + "blocks": [ + { + "bbox": [ + 138, + 149, + 483, + 361 + ], + "lines": [ + { + "bbox": [ + 138, + 149, + 483, + 361 + ], + "spans": [ + { + "bbox": [ + 138, + 149, + 483, + 361 + ], + "type": "image", + "image_path": "9fe158de862ee63af9520a2a7fe28128a96fc5a1ecd1b162273f232956b9b989.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 367, + 482, + 391 + ], + "lines": [ + { + "bbox": [ + 130, + 367, + 482, + 391 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 482, + 391 + ], + "type": "text", + "content": "Fig. 6: Visualization results of our ReMamber and the baseline model LAVT. Our model is able to predict more accurate masks." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 396, + 483, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 483, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 483, + 434 + ], + "type": "text", + "content": "Fig. 6 presents the visualization results of our method compared with the baseline method LAVT [56]. Our ReMamber is capable of producing segmentation results with higher accuracy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 448, + 221, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 448, + 221, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 448, + 221, + 462 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 470, + 483, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 483, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 483, + 568 + ], + "type": "text", + "content": "In this study, we introduce ReMamber, a novel architecture utilizing the Mamba framework in Referring Image Segmentation (RIS). It marks a significant advancement in multi-modal understanding. By integrating visual and textual information through innovative Mamba Twister blocks, our approach sets new benchmarks in image-text modality fusion. Achieving competitive results across multiple RIS datasets, our research highlights the potential of Mamba architecture in enhancing the scalability and performance of multi-modal tasks, offering promising directions for future exploration in the field." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Limitations and Future Works. In our current architecture, the segmentation decoder is constructed by only a few convolutional layers. As shown in Sec. 4.3, the integration of the Cross-Attention mechanism within the Mamba-based architecture demonstrates sub-optimal compatibility, undermining the overall efficacy of our method. In light of these findings, future endeavors will be directed towards the investigation of more sophisticated multi-modal segmentation decoders which best fit Mamba architecture." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 145, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 145, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 145, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 90, + 409, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 90, + 409, + 103 + ], + "spans": [ + { + "bbox": [ + 166, + 90, + 409, + 103 + ], + "type": "text", + "content": "Y. Yang, C. Ma, J. Yao, Z. Zhong, Y. Zhang, and Y. Wang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 139, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 139, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 139, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 139, + 480, + 161 + ], + "type": "text", + "content": "1. Chen, B., Hu, Z., Ji, Z., Bai, J., Zuo, W.: Position-aware contrastive alignment for referring image segmentation. arXiv preprint arXiv:2212.13419 (2022) 12" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 162, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 480, + 194 + ], + "type": "text", + "content": "2. Chen, D.J., Jia, S., Lo, Y.C., Chen, H.T., Liu, T.L.: See-through-text grouping for referring image segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7454-7463 (2019) 3, 12" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 195, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 195, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 195, + 480, + 226 + ], + "type": "text", + "content": "3. Chen, Y.W., Tsai, Y.H., Wang, T., Lin, Y.Y., Yang, M.H.: Referring expression object segmentation with caption-aware consistency. arXiv preprint arXiv:1910.04748 (2019) 3, 12" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 228, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 259 + ], + "type": "text", + "content": "4. Ding, H., Liu, C., Wang, S., Jiang, X.: Vision-language transformer and query generation for referring segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16321-16330 (2021) 1, 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 261, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 261, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 138, + 261, + 480, + 303 + ], + "type": "text", + "content": "5. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 304, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 480, + 346 + ], + "type": "text", + "content": "6. Feng, G., Hu, Z., Zhang, L., Lu, H.: Encoder fusion network with co-attention embedding for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15506-15515 (2021) 3, 12" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 348, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 480, + 380 + ], + "type": "text", + "content": "7. Fu, D.Y., Dao, T., Saab, K.K., Thomas, A.W., Rudra, A., Ré, C.: Hungry hungry hippos: Towards language modeling with state space models. arXiv preprint arXiv:2212.14052 (2022) 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 381, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 381, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 138, + 381, + 480, + 402 + ], + "type": "text", + "content": "8. Gu, A., Dao, T.: Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752 (2023) 2, 4, 5, 7, 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "type": "text", + "content": "9. Gu, A., Dao, T., Ermon, S., Rudra, A., Ré, C.: Hippo: Recurrent memory with optimal polynomial projections. Advances in neural information processing systems 33, 1474-1487 (2020) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 436, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 436, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 436, + 480, + 468 + ], + "type": "text", + "content": "10. Gu, A., Goel, K., Gupta, A., Ré, C.: On the parameterization and initialization of diagonal state space models. Advances in Neural Information Processing Systems 35, 35971-35983 (2022) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 468, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 490 + ], + "type": "text", + "content": "1. Gu, A., Goel, K., Ré, C.: Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396 (2021) 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "type": "text", + "content": "2. Gu, A., Johnson, I., Goel, K., Saab, K., Dao, T., Rudra, A., Ré, C.: Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems 34, 572-585 (2021) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "text", + "content": "3. Gupta, A., Gu, A., Berant, J.: Diagonal state spaces are as effective as structured state spaces. Advances in Neural Information Processing Systems 35, 22982-22994 (2022) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 556, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 556, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 556, + 480, + 587 + ], + "type": "text", + "content": "4. He, X., Cao, K., Yan, K., Li, R., Xie, C., Zhang, J., Zhou, M.: Pan-mamba: Effective pan-sharpening with state space model. arXiv preprint arXiv:2402.12192 (2024) 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 589, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 480, + 632 + ], + "type": "text", + "content": "5. Hu, R., Rohrbach, M., Darrell, T.: Segmentation from natural language expressions. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14. pp. 108-124. Springer (2016) 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "6. Hu, Z., Feng, G., Sun, J., Zhang, L., Lu, H.: Bi-directional relationship inferring network for referring image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4424-4433 (2020) 3, 12" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 185, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 185, + 91, + 447, + 102 + ], + "type": "text", + "content": "ReMamber: Referring Image Segmentation with Mamba Twister" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "17. Huang, S., Hui, T., Liu, S., Li, G., Wei, Y., Han, J., Liu, L., Li, B.: Referring image segmentation via cross-modal progressive comprehension. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10488-10497 (2020) 3, 12" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 195 + ], + "type": "text", + "content": "18. Hui, T., Liu, S., Huang, S., Li, G., Yu, S., Zhang, F., Han, J.: Linguistic structure guided context modeling for referring image segmentation. In: European Conference on Computer Vision. pp. 59-75. Springer (2020) 12" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 196, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 196, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 481, + 228 + ], + "type": "text", + "content": "19. Islam, M.M., Bertasius, G.: Long movie clip classification with state-space video models. In: European Conference on Computer Vision. pp. 87-104. Springer (2022) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 228, + 481, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 481, + 272 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 481, + 272 + ], + "type": "text", + "content": "20. Islam, M.M., Hasan, M., Athrey, K.S., Braskich, T., Bertasius, G.: Efficient movie scene detection using state-space transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18749-18758 (2023) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "type": "text", + "content": "21. Jing, Y., Kong, T., Wang, W., Wang, L., Li, L., Tan, T.: Locate then segment: A strong pipeline for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9858-9867 (2021) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 308, + 481, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 351 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 351 + ], + "type": "text", + "content": "22. Kamath, A., Singh, M., LeCun, Y., Synnaeve, G., Misra, I., Carion, N.: Mdetr-modulated detection for end-to-end multi-modal understanding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1780-1790 (2021) 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 352, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 481, + 385 + ], + "type": "text", + "content": "23. Kazemzadeh, S., Ordonez, V., Matten, M., Berg, T.: ReferItGame: Referring to objects in photographs of natural scenes. In: Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP) (2014) 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 386, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 430 + ], + "type": "text", + "content": "24. Kim, N., Kim, D., Lan, C., Zeng, W., Kwak, S.: Restr: Convolution-free referring image segmentation using transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18145-18154 (2022) 1, 2, 3, 12" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 430, + 481, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 481, + 464 + ], + "type": "text", + "content": "25. Li, R., Li, K., Kuo, Y.C., Shu, M., Qi, X., Shen, X., Jia, J.: Referring image segmentation via recurrent refinement networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5745-5753 (2018) 1, 3, 12" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 464, + 481, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 481, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 481, + 497 + ], + "type": "text", + "content": "26. Liang, D., Zhou, X., Wang, X., Zhu, X., Xu, W., Zou, Z., Ye, X., Bai, X.: Point-mamba: A simple state space model for point cloud analysis. arXiv preprint arXiv:2402.10739 (2024) 2, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 498, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 520 + ], + "type": "text", + "content": "27. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common Objects in Context. ECCV (2014) 9" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 521, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 481, + 553 + ], + "type": "text", + "content": "28. Liu, C., Ding, H., Jiang, X.: Gres: Generalized referring expression segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23592-23601 (2023) 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 554, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 481, + 587 + ], + "type": "text", + "content": "29. Liu, C., Lin, Z., Shen, X., Yang, J., Lu, X., Yuille, A.: Recurrent multimodal interaction for referring image segmentation. In: Proceedings of the IEEE international conference on computer vision. pp. 1271-1280 (2017) 1, 3, 12" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 588, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 481, + 632 + ], + "type": "text", + "content": "30. Liu, J., Ding, H., Cai, Z., Zhang, Y., Satzoda, R.K., Mahadevan, V., Manmatha, R.: Polyformer: Referring image segmentation as sequential polygon generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18653-18663 (2023) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "31. Liu, J., Yang, H., Zhou, H.Y., Xi, Y., Yu, L., Yu, Y., Liang, Y., Shi, G., Zhang, S., Zheng, H., et al.: Swin-umamba: Mamba-based unet with imagenet-based pretraining. arXiv preprint arXiv:2402.03302 (2024) 2, 4" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 408, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 408, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 408, + 102 + ], + "type": "text", + "content": "Y. Yang, C. Ma, J. Yao, Z. Zhong, Y. Zhang, and Y. Wang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "32. Liu, S., Hui, T., Huang, S., Wei, Y., Li, B., Li, G.: Cross-modal progressive comprehension for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(9), 4761-4775 (2021) 12" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "text", + "content": "33. Liu, Y., Tian, Y., Zhao, Y., Yu, H., Xie, L., Wang, Y., Ye, Q., Liu, Y.: Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166 (2024) 2, 4, 6, 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "text", + "content": "34. Luo, G., Zhou, Y., Ji, R., Sun, X., Su, J., Lin, C.W., Tian, Q.: Cascade grouped attention network for referring expression segmentation. Proceedings of the 28th ACM International Conference on Multimedia (2020) 12" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "35. Ma, C., Yang, Y., Ju, C., Zhang, F., Zhang, Y., Wang, Y.: Attrseg: Open-vocabulary semantic segmentation via attribute decomposition-aggregation. NeurIPS (2023) 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": "36. Ma, C., Yang, Y., Wang, Y., Zhang, Y., Xie, W.: Open-vocabulary semantic segmentation with frozen vision-language models. In: British Machine Vision Conference (2022) 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "text", + "content": "37. Ma, J., Li, F., Wang, B.: U-mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722 (2024) 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "type": "text", + "content": "38. Mao, J., Huang, J., Toshev, A., Camburu, O.M., Yuille, A., Murphy, K.: Generation and comprehension of unambiguous object descriptions. CVPR (2015) 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 347 + ], + "type": "text", + "content": "39. Margffoy-Tuay, E., Pérez, J.C., Botero, E., Arbeláez, P.: Dynamic multimodal instance segmentation guided by natural language queries. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 630-645 (2018) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "text", + "content": "40. Mehta, H., Gupta, A., Cutkosky, A., Neyshabur, B.: Long range language modeling via gated state spaces. arXiv preprint arXiv:2206.13947 (2022) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 403 + ], + "type": "text", + "content": "41. Nagaraja, V.K., Morariu, V.I., Davis, L.S.: Modeling context between objects for referring expression understanding. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) ECCV (2016) 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 402, + 480, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 480, + 425 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 480, + 425 + ], + "type": "text", + "content": "42. Perez, E., Strub, F., de Vries, H., Dumoulin, V., Courville, A.C.: Film: Visual reasoning with a general conditioning layer. In: AAAI (2018) 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 425, + 480, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 480, + 457 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 480, + 457 + ], + "type": "text", + "content": "43. Qu, M., Wu, Y., Wei, Y., Liu, W., Liang, X., Zhao, Y.: Learning to segment every referring object point by point. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023) 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "text", + "content": "44. Ruan, J., Xiang, S.: Vm-unet: Vision mamba unet for medical image segmentation. arXiv preprint arXiv:2402.02491 (2024) 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "type": "text", + "content": "45. Shi, H., Li, H., Meng, F., Wu, Q.: Key-word-aware network for referring expression image segmentation. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 38-54 (2018) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 512, + 480, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 534 + ], + "type": "text", + "content": "46. Smith, J.T., Warrington, A., Linderman, S.W.: Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933 (2022) 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 534, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 567 + ], + "type": "text", + "content": "47. Tang, J., Zheng, G., Shi, C., Yang, S.: Contrastive grouping with transformer for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23570-23580 (2023) 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 567, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 589 + ], + "type": "text", + "content": "48. Tang, J., Zheng, G., Shi, C., Yang, S.: Contrastive grouping with transformer for referring image segmentation. In: CVPR (2023) 9" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "type": "text", + "content": "49. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 480, + 665 + ], + "type": "text", + "content": "50. Wang, J., Zhu, W., Wang, P., Yu, X., Liu, L., Omar, M., Hamid, R.: Selective structured state-spaces for long-form video understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6387-6397 (2023) 4" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 185, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 185, + 91, + 447, + 102 + ], + "type": "text", + "content": "ReMamber: Referring Image Segmentation with Mamba Twister" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 511 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "51. Wang, Z., Lu, Y., Li, Q., Tao, X., Guo, Y., Gong, M., Liu, T.: Cris: Clip-driven referring image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11686-11695 (2022) 1, 3, 12" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "text", + "content": "52. Xing, Z., Ye, T., Yang, Y., Liu, G., Zhu, L.: Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. arXiv preprint arXiv:2401.13560 (2024) 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 225 + ], + "type": "text", + "content": "53. Xu, Z., Chen, Z., Zhang, Y., Song, Y., Wan, X., Li, G.: Bridging vision and language encoders: Parameter-efficient tuning for referring image segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17503-17512 (2023) 12" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "54. Yang, S., Xia, M., Li, G., Zhou, H.Y., Yu, Y.: Bottom-up shift and reasoning for referring image segmentation. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11261-11270 (2021) 12" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "55. Yang, Y., Ma, C., Ju, C., Zhang, Y., Wang, Y.: Multi-modal prototypes for open-set semantic segmentation. IJCV (2024) 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 480, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 324 + ], + "type": "text", + "content": "56. Yang, Z., Wang, J., Tang, Y., Chen, K., Zhao, H., Torr, P.H.: Lavt: Language-aware vision transformer for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18155-18165 (2022) 1, 2, 3, 9, 11, 12, 14" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 325, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 357 + ], + "type": "text", + "content": "57. Ye, L., Rochan, M., Liu, Z., Wang, Y.: Cross-modal self-attention network for referring image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10502-10511 (2019) 3, 12" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 480, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 401 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 401 + ], + "type": "text", + "content": "58. Yu, L., Lin, Z., Shen, X., Yang, J., Lu, X., Bansal, M., Berg, T.L.: Matttnet: Modular attention network for referring expression comprehension. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018) 12" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 402, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 480, + 423 + ], + "type": "text", + "content": "59. Zhang, T., Li, X., Yuan, H., Ji, S., Yan, S.: Point could mamba: Point cloud learning via state space model. arXiv preprint arXiv: 2403.00762 (2024) 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "text", + "content": "60. Zheng, Z., Wu, C.: U-shaped vision mamba for single image dehazing. arXiv preprint arXiv:2402.04139 (2024) 2, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "text", + "content": "61. Zhu, C., Zhou, Y., Shen, Y., Luo, G., Pan, X., Lin, M., Chen, C., Cao, L., Sun, X., Ji, R.: Seqtr: A simple yet universal network for visual grounding. In: European Conference on Computer Vision. pp. 598-615. Springer (2022) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "text", + "content": "62. Zhu, L., Liao, B., Zhang, Q., Wang, X., Liu, W., Wang, X.: Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417 (2024) 2, 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 90, + 408, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 90, + 408, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 90, + 408, + 102 + ], + "type": "text", + "content": "Y. Yang, C. Ma, J. Yao, Z. Zhong, Y. Zhang, and Y. Wang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/ReMatching_ Low-Resolution Representations for Scalable Shape Correspondence/8dd15a7c-48bf-42e4-a752-0b0dbad83c91_content_list.json b/2024/ReMatching_ Low-Resolution Representations for Scalable Shape Correspondence/8dd15a7c-48bf-42e4-a752-0b0dbad83c91_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..78c3edb489265a0468bd97a48c38409e75e04605 --- /dev/null +++ b/2024/ReMatching_ Low-Resolution Representations for Scalable Shape Correspondence/8dd15a7c-48bf-42e4-a752-0b0dbad83c91_content_list.json @@ -0,0 +1,1869 @@ +[ + { + "type": "text", + "text": "ReMatching: Low-Resolution Representations for Scalable Shape Correspondence", + "text_level": 1, + "bbox": [ + 220, + 140, + 782, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Filippo Maggioli1, Daniele Baieri2, Emanuele Rodolà2, and Simone Melzi1", + "bbox": [ + 241, + 210, + 759, + 243 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1 University of Milano-Bicocca, Milan, Italy", + "2 Sapienza - University of Rome, Rome, Italy" + ], + "bbox": [ + 346, + 253, + 653, + 282 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/054bd1a6d36e11904940f38384a53e57a94f58ecc74a27bbc669d4eb29a9846a.jpg", + "image_caption": [ + "Fig. 1: Correspondence between two dense shapes computed with our method, using ZoomOut [38] on low-resolution meshes. The coordinates functions of the dancing faun statue on the left ( $\\sim750\\mathrm{k}$ vertices) are transferred using a functional map to the Aphaea warrior statue on the right ( $\\sim3.5\\mathrm{M}$ vertices). The colors encode the correspondence. Despite the mesh density, shown in the close-up, the computation took $\\sim2$ minutes." + ], + "image_footnote": [], + "bbox": [ + 217, + 323, + 784, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We introduce ReMatching, a novel shape correspondence solution based on the functional maps framework. Our method, by exploiting a new and appropriate re-meshing paradigm, can target shape-matching tasks even on meshes counting millions of vertices, where the original functional maps does not apply or requires a massive computational cost. The core of our procedure is a time-efficient remeshing algorithm which constructs a low-resolution geometry while acting conservatively on the original topology and metric. These properties allow translating the functional maps optimization problem on the resulting low-resolution representation, thus enabling efficient computation of correspondences with functional map approaches. Finally, we propose an efficient technique for extending the estimated correspondence to the original meshes. We show that our method is more efficient and effective through quantitative and qualitative comparisons, outperforming state-of-the-art pipelines in quality and computational cost.", + "bbox": [ + 259, + 632, + 743, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction and related work", + "text_level": 1, + "bbox": [ + 215, + 188, + 542, + 204 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The task of finding a semantically meaningful correspondence between discrete surfaces has always been a fundamental topic in the field of shape analysis. Researchers developed a wealth of solutions for this application, and new methods continue to be designed [10, 57]. Among the various approaches, the functional maps framework [47, 48] received significant attention. Rather than finding a point-wise correspondence, the functional maps framework aims to define a correspondence between functions, encoding it into a small matrix. Following this direction, a variety of works tried to improve the actual computation of the map [13, 38, 53] or to extend the framework to different types of bases [34, 39, 45].", + "bbox": [ + 212, + 220, + 785, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the amount of research outcomes based on the functional maps framework, both exploiting geometric properties and tools [7, 16, 17] and clever optimization techniques [46, 51, 52], the problem of computing the map is still bounded by the time complexity of sparse eigendecompositions [26]. To overcome this problem, researchers started investigating scalable solutions to the Laplace-Beltrami eigenproblem. In this regard, multi-resolution techniques [42, 43, 58, 59] proved to be very effective for scalable computation of spectral quantities for tasks such as retrieval and mesh filtering [30, 54], but their applicability to the functional map framework proved to require additional care [35]. Furthermore, recent research proposed spectral coarsening methods [32] or spectral preserving simplification techniques [28]. However, these techniques usually rely on the computation of the full-resolution spectrum, making them unsuitable for large-scale applications. Gao et al. [19] proposed a mixed-integer programming scalable solution to the matching problem, but they specifically target sparse correspondences only. Alternatively, Marin et al. [36], who provide an approach for human registration that scales across large datasets, but it is unsuitable for dense shapes.", + "bbox": [ + 212, + 356, + 785, + 598 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Closely related to our work is the contribution from Magnet et al. [35], which, for the first time, proposes a solution for scaling the functional maps approach to high-resolution meshes. Their idea is to reduce the dimensionality of the eigenproblem, inducing a linear relationship between the functional space on the mesh and a lower-dimension functional space on a sparse sample. The extreme sparsity of the sampling, combined with its efficient computation, leads to an algorithm that can effectively deal with high-density meshes. Nevertheless, this type of reduction does not mitigate the negative influence that small components and local details at high frequency could have on the alignment of the spectra.", + "bbox": [ + 212, + 598, + 785, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, our pipeline relies on building robust and sparse representations of dense meshes while still keeping an approximately bijective correspondence between the high- and low-resolution shapes, and it is worth mentioning that other works have already attempted to do so. Jiang et al. [23] propose building a lower resolution prismatic shell preserving bijectivity with the original surface. However, their algorithm does not scale to meshes with a high triangle count, gives no control and no guarantees on the final vertex count, and requires many assumption.", + "bbox": [ + 212, + 734, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "F. Maggioli et al.", + "bbox": [ + 271, + 114, + 387, + 128 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Keywords: Shape matching $\\cdot$ Functional maps $\\cdot$ Geometry processing", + "bbox": [ + 261, + 147, + 733, + 161 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tions on the input shape (e.g., manifoldness, orientability, no self-intersections). On the other hand, Liu et al. [31] propose to exploit an intrinsic error metric for decimating a mesh to a given size while keeping track of a geodesic barycentric mapping of each triangle of the resulting shape to a geodesic triangle on the input surface. Despite the robustness of the method, the simplification approach makes it unsuitable for a functional maps setting, where it is required to decimate shapes from millions of vertices to a few thousand. This limitation is further discussed in Section 4.2.", + "bbox": [ + 212, + 146, + 782, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce a new scalable functional map pipeline that efficiently handles meshes with high vertex density (see Figure 1), yields stable results, and is not bounded by the quality of the original triangulation. To summarize:", + "bbox": [ + 212, + 267, + 782, + 311 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- we translate the matching pipeline to low-resolution representations, enabling a fast and scalable computation of functional maps between dense shapes, providing the first alternative to [35];", + "- we propose a new efficient and geometry-preserving remeshing algorithm specifically designed for our pipeline;", + "- we exploit a fast solution for extending scalar maps from the low-resolution remeshed shape to the original surface, thus efficiently obtaining suitable bases for function transfer and shape matching." + ], + "bbox": [ + 222, + 313, + 782, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Background", + "text_level": 1, + "bbox": [ + 215, + 457, + 367, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the discrete setting, we represent a shape as a triangular mesh $\\mathcal{M} = (V,E,T)$ where (i) $V\\subset \\mathbb{R}^3$ is a set of vertices; (ii) $E\\subset V^2$ is a set of edges between vertices; (iii) $T\\subset V^3$ is a set of triangles composing the surface.", + "bbox": [ + 212, + 489, + 782, + 535 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Functional maps", + "text_level": 1, + "bbox": [ + 215, + 556, + 397, + 573 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We discretize a scalar function $f: \\mathcal{M} \\to \\mathbb{R}$ as a signal defined on the vertices $V$ and represented as a vector $\\pmb{f} \\in \\mathbb{R}^{|V|}$ . Hence, the Laplace-Beltrami operator $\\varDelta_{\\mathcal{M}}: \\mathcal{F}(\\mathcal{M}) \\to \\mathcal{F}(\\mathcal{M})$ is discretized as a sparse matrix $\\mathbf{L}_{\\mathcal{M}} \\in \\mathbb{R}^{|V| \\times |V|}$ , generally represented by means of a stiffness matrix $\\mathbf{S}_{\\mathcal{M}}$ and a mass matrix $\\mathbf{A}_{\\mathcal{M}}$ [40, 50]. The eigendecomposition $\\mathbf{S}_{\\mathcal{M}} \\pmb{\\Phi}_{\\mathcal{M}} = \\mathbf{A}_{\\mathcal{M}} \\pmb{\\Phi}_{\\mathcal{M}} \\varLambda_{\\mathcal{M}}$ of the Laplacian yields an orthonormal basis for the functional space on $\\mathcal{M}$ [29,30] which has analogies with the Fourier basis and optimally represents smooth functions when truncated [1].", + "bbox": [ + 212, + 583, + 784, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given two shapes $\\mathcal{M}$ and $\\mathcal{N}$ , respectively with $m$ and $n$ vertices, a point-to-point map $\\Pi : \\mathcal{M} \\to \\mathcal{N}$ can be expressed as a binary matrix $\\Pi \\in \\{0,1\\}^{m \\times n}$ such that $\\Pi(i,j) = 1$ if the vertex $j$ -th of $\\mathcal{N}$ corresponds to the vertex $i$ -th of $\\mathcal{M}$ , and $\\Pi(i,j) = 0$ otherwise. Any point-wise correspondence $\\Pi$ establishes a functional map $T_{\\Pi} : \\mathcal{F}(\\mathcal{N}) \\to \\mathcal{F}(\\mathcal{M})$ as $T_{\\Pi}(f) = f \\circ \\Pi$ , $\\forall f \\in \\mathcal{F}(\\mathcal{N})$ .", + "bbox": [ + 212, + 688, + 784, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As proposed in [47], given $\\Phi_k$ and $\\Psi_k$ the Laplacian bases truncated to size $k$ respectively on $\\mathcal{M}$ and $\\mathcal{N}$ , we can project $T_{\\Pi}$ in these bases and compactly encode the functional map in a $k \\times k$ matrix $\\mathbf{C} = \\Phi_k^\\top \\mathbf{A}_{\\mathcal{M}} \\Pi \\Psi_k$ , which is the linear operator that transfers the coefficients of functions from $\\mathcal{F}(\\mathcal{N})$ in the ones of corresponding functions of $\\mathcal{F}(\\mathcal{M})$ respectively computed with $\\Phi_k$ and $\\Psi_k$ .", + "bbox": [ + 212, + 765, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "ReMatching: Scalable Shape Correspondence", + "bbox": [ + 431, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, given a functional map $T_{\\Pi}$ , or its compact representation $\\mathbf{C}$ , we can retrieve the correspondence $\\Pi$ on which the map is built by the nearest neighbor search in the embedding space of the Laplacian eigenfunctions as detailed in [48].", + "bbox": [ + 212, + 146, + 787, + 193 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Intrinsic Delaunay triangulations (IDT)", + "text_level": 1, + "bbox": [ + 214, + 210, + 589, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The IDT [3, 55] has been introduced for generalizing the Delaunay triangulation [9] to non-Euclidean metric spaces where classical algorithms [5, 60] cannot be used. IDT relies on the duality with the Voronoi diagram [15, 27], as three texels meeting at a point form a triangle whose edges cross the texels boundaries.", + "bbox": [ + 212, + 234, + 782, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 1. Let $(\\mathcal{M},d_{\\mathcal{M}})$ be a 2-dimensional metric space (with distance function $d_{\\mathcal{M}}$ ), and let $S = \\{p_i\\}_{i = 1}^s\\subset \\mathcal{M}$ be a set of $s$ sample points. The Voronoi decomposition (VD) of $\\mathcal{M}$ with respect to $S$ is the collection $\\{P_i\\}_{i = 1}^s$ such that", + "bbox": [ + 214, + 301, + 784, + 349 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {i} = \\left\\{q \\in \\mathcal {M}: p _ {i} = \\underset {p _ {j} \\in S} {\\arg \\min } \\left(d _ {\\mathcal {M}} \\left(p _ {j}, q\\right)\\right) \\right\\}, \\quad \\bigcup_ {i = 1} ^ {s} P _ {i} = \\mathcal {M} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 284, + 357, + 785, + 397 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The points in $S$ are called generators, and the $P_{i}$ are called texels (or cells).", + "bbox": [ + 217, + 404, + 769, + 420 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The intersection between two or more texels could be non-empty. In such a case, we define those texels as adjacent. A VD is general if the intersection of four or more texels is empty. If a VD is general, a Voronoi vertex is a point that belongs to three texels, and a Voronoi edge is a curve $C$ connecting two Voronoi vertices and belonging to two texels.", + "bbox": [ + 212, + 426, + 782, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It can be shown that there is a one-to-one correspondence between a Voronoi edge dividing $P_{i}$ and $P_{j}$ and a Delaunay edge connecting $p_{i}$ to $p_{j}$ [14]. This correspondence gives rise to a necessary and sufficient condition for the IDT to be a proper triangulation (i.e., it realizes a simplicial complex).", + "bbox": [ + 212, + 503, + 782, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 2 ([2,15]). If $S$ induces a general $VD$ , the $VD$ satisfies the closed ball property if:", + "bbox": [ + 212, + 570, + 782, + 599 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- every $P_{i}$ is a closed 2-ball (i.e., a topological disk without holes);", + "- every $P_i \\cap P_j$ is either empty or a closed 1-ball (i.e., a topological segment);", + "- every $P_i \\cap P_j \\cap P_k$ is either empty or a closed 0-ball (i.e., a point)." + ], + "bbox": [ + 225, + 601, + 785, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1 ([2,14,15]). If $S$ induces a general $VD$ , then the $VD$ satisfies the closed ball property if and only if its dual $IDT$ is a proper triangulation.", + "bbox": [ + 214, + 652, + 782, + 684 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 704, + 330, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce a new scalable functional map pipeline handling very dense meshes and yielding stable results independently of the input triangulation's quality.", + "bbox": [ + 212, + 734, + 782, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our method acts through three main steps. At first, we compute two low-resolution meshes that preserve the original metrics of the input shapes (Section 3.1). Then, we apply any existing pipeline, yielding a functional map between the low-resolution eigenspaces (Section 3.2). Finally, we extend the eigenfunctions to the original meshes, allowing us to use them with the functional map", + "bbox": [ + 212, + 765, + 784, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "F. Maggioli et al.", + "bbox": [ + 271, + 114, + 387, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "estimated in the previous step to compute the desired correspondence between them (Section 3.3). Our technique is very general and works with arbitrary manifold triangulations, including meshes with degenerate geometry, non-orientable surfaces, and multiple connected components. The manifoldness requirement does not provide an obstacle to the applicability of our algorithm, as it can be efficiently enforced using existing techniques [18].", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Intrinsic Delaunay remeshing", + "text_level": 1, + "bbox": [ + 214, + 257, + 504, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Front propagation and FPS. Naively computing the VD induced by a set of samples can quickly become very expensive. The exact geodesic algorithm introduced by Mitchell et al. [41] is computationally unfeasible on large meshes. Even using faster solutions, such as the heat method [8], the fast marching algorithm [25], or even the Dijkstra distance [11], searching the closest sample for each vertex is still a $\\mathcal{O}(|V|s)$ operation. Instead, we take inspiration from the front propagation method proposed by Peyre et al. [49]. We start with a decomposition of the mesh induced by a single sample $p_1$ .", + "bbox": [ + 212, + 281, + 488, + 523 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/88eb061469dddfd03768f683ce6015dee27a2fb20c18ea42b41575bc6d24603c.jpg", + "image_caption": [ + "Fig.2: The iterative front propagation from samples inducing a geodesic VD." + ], + "image_footnote": [], + "bbox": [ + 501, + 311, + 784, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This means that every vertex $v$ is part of the texel $P_{1}$ and has distance $D_{1}[v] = d_{\\mathcal{M}}(p_{1}, v)$ from the sample set. We then assume to have a decomposition $P_{1}, \\dots, P_{k-1}$ induced by samples $p_{1}, \\dots, p_{k-1}$ , and a vector $D_{k-1}$ storing the distance of each vertex from the sample set. When adding a sample $p_{k}$ to the VD, we can exploit the fact that $D_{k} = \\min(D_{k-1}, d_{\\mathcal{M}}(p_{k}, V))$ . We start a front from $p_{k}$ and expand it, updating the distances and assigning vertices to texel $P_{k}$ until we reach all vertices such that $D_{k-1}[v] < d_{\\mathcal{M}}(p_{k}, v)$ (see Figure 2). By propagating the front with the fast marching or Dijkstra's algorithm, we can compute a geodesic VD with time $\\mathcal{O}(|V| \\log(|V|) \\log(s))$ .", + "bbox": [ + 212, + 523, + 787, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To obtain the sample points, Peyré et al. [49] propose to use a geodesic farthest point sampling, achieving a geodesic uniform sampling of the surface.", + "bbox": [ + 212, + 659, + 787, + 689 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Definition 3. Let $(\\mathcal{M}, d_{\\mathcal{M}})$ be a 2-dimensional metric space, and let $s \\in \\mathbb{N}$ be an integer. A farthest point sampling (FPS) of size $s$ with respect to $d_{\\mathcal{M}}$ is a set $S_s \\subset \\mathcal{M}$ of $s$ sample points in $\\mathcal{M}$ , incrementally built from a singleton set $S_1 = \\{p_1\\} \\subset \\mathcal{M}$ according to the following rule:", + "bbox": [ + 214, + 696, + 789, + 758 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {\\ell + 1} = S _ {\\ell} \\cup \\left\\{\\underset {q \\in \\mathcal {M}} {\\arg \\max } \\quad \\min _ {p \\in S _ {\\ell}} d _ {\\mathcal {M}} (q, p) \\right\\}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 766, + 785, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, searching for the maximum distance in Equation (2) at every new sample brings again the algorithm to a $\\mathcal{O}(|V|s)$ complexity, really becoming a", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "ReMatching: Scalable Shape Correspondence", + "bbox": [ + 431, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "bottleneck for large meshes counting millions of vertices. In contrast, we propose to introduce a fixed-size binary max-heap in the front propagation algorithm, keeping track of the distances from each vertex to the sample set. During each front propagation, this heap updates the distance of each visited vertex in time $\\mathcal{O}(\\log(|V|))$ , adding no extra complexity to the visit. Furthermore, it can gather and set to zero the farthest vertex at each iteration in just $\\mathcal{O}(\\log(|V|))$ time, totalling a time complexity of $\\mathcal{O}(|V|\\log(|V|)\\log(s))$ for the entire procedure. The fast front propagation algorithm is summarized in Algorithm 1.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 5 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Geodesic FPS and its VD." + ], + "code_body": "1: procedure VORONOIFPS $(\\mathcal{M} = (V,E,T),s)$ \n2: $i\\gets$ random index in $[1,|V|]$ \n3: $S\\gets \\{i\\}$ \n4: $D\\gets d_{\\mathcal{M}}(V,V_i)$ \n5: $H\\gets$ max-heap initialized with $D$ \n6: $P\\gets$ vector of length $|V|$ initialized to $i$ \n7: for $h\\gets 2$ to $s$ do \n8: $p\\gets \\mathrm{FINDMAX}(H)$ \n9: SETKEY $(H,p,0)$ \n10: $D_p\\gets 0$ \n11: $S\\gets S\\cup \\{p\\}$ \n12: Propagate a front from $p$ \n13: Update $D,P,$ and $H$ \n14: end for \n15: end procedure", + "bbox": [ + 215, + 362, + 558, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Flat union property. Even if we are able to compute the VD of a farthest point sampling with high efficiency, we still cannot guarantee that the dual connectivity is a proper IDT. Ensuring the closed ball property can be computationally challenging. In particular, identifying the topology of the boundary between each pair of adjacent texels can become very costly when the number of samples is large. A possibility is to add enough samples so that the surface locally behaves like a plane. This approach, proposed by Leibon et al. [27] and adopted by Peyré", + "bbox": [ + 568, + 282, + 787, + 584 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "et al. [49], makes it easy to lose control over the number of samples, and anyway the requirements are not easy to enforce. For guaranteeing a proper IDT, while still avoiding these issues, we introduce a novel property for a VD. The proofs of our claims are provided in the supplementary materials.", + "bbox": [ + 212, + 584, + 787, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Definition 4 (Flat Union Property (FUP)). Let $(\\mathcal{M}, d_{\\mathcal{M}})$ be a 2-dimensional metric space. Let then $S = \\{p_i\\}_{i=1}^s \\subset \\mathcal{M}$ be a set of $s$ sample points in $\\mathcal{M}$ , inducing a general $VD\\{P_i\\}_{i=1}^s$ . The $VD$ induced by $S$ satisfies the FUP if:", + "bbox": [ + 212, + 652, + 787, + 700 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- every $P_i$ is a closed 2-ball;", + "- if $P_{i} \\cap P_{j}$ is not empty, then $P_{i} \\cup P_{j}$ is a closed 2-ball;", + "- if $P_{i} \\cap P_{j} \\cap P_{k}$ is not empty, then $P_{i} \\cup P_{j} \\cup P_{k}$ is a closed 2-ball." + ], + "bbox": [ + 223, + 699, + 707, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Theorem 2. Let $(\\mathcal{M}, d_{\\mathcal{M}})$ be a 2-dimensional metric space. If a general $VD\\{P_i\\}_{i=1}^s$ of $\\mathcal{M}$ satisfies the FUP, it also satisfies the closed ball property.", + "bbox": [ + 212, + 753, + 787, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The FUP is much easier to verify than the closed ball property, as we only have to ensure that regions are closed 2-balls. Indeed, the following property can be exploited in this regard.", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "F. Maggioli et al.", + "bbox": [ + 271, + 114, + 387, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Proposition 1. Let $\\mathcal{M} = (V,E,F)$ be a manifold polygonal mesh, then $\\mathcal{M}$ is a closed 2-ball if and only if its Euler characteristic $\\chi = |V| - |E| + |F|$ is 1.", + "bbox": [ + 215, + 146, + 787, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dual mesh representation. When we compute a VD of a triangle mesh, we want to partition the vertices into disjoint sets. However, Voronoi texels are defined as regions over the surface, meaning that they should be submanifolds of $\\mathcal{M}$ . By only considering vertices, we likely end up having non-manifold geometries or non well-defined boundaries.", + "bbox": [ + 212, + 190, + 782, + 265 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/52650bede014b0df04c1a9169bff639112145d3a025fa4bab899d031a2a522e4.jpg", + "image_caption": [ + "Fig. 3: Top: the primal and dual connectivities of a triangular mesh. Bottom: regions on the primal mesh are represented as sets of vertices, and the corresponding regions on the dual mesh are composed by faces." + ], + "image_footnote": [], + "bbox": [ + 243, + 297, + 478, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We consider the polygonal mesh $\\tilde{\\mathcal{M}} = (\\tilde{V},\\tilde{E},\\tilde{T})$ to be the dual mesh of $\\mathcal{M} = (V,E,T)$ , built by placing a vertex at each face of $\\mathcal{M}$ and forming a polygonal face for each triangle fan around a vertex of $\\mathcal{M}$ . As shown in the example from Figure 3, dual meshes allow us to define Voronoi texels by means of connected dual faces, making them actual submanifolds of the original surface. Furthermore, we can easily define boundaries between texels as paths made of dual edges.", + "bbox": [ + 511, + 266, + 787, + 462 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This construction has the positive side effect to ensure that every VD is general. Indeed, the only place where three or more texels can meet is a dual vertex. Since texels are made of dual faces (i.e., primal vertices), and the primal mesh is triangular, there is no dual vertex where more than three faces can meet.", + "bbox": [ + 511, + 462, + 787, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given the correspondence between primal and dual geometric elements, we can verify if a region is a closed 2-", + "bbox": [ + 511, + 598, + 785, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ball with Proposition 1 without explicitly constructing the dual mesh. Additional details are provided in the supplementary material. After the initial FPS, we further add samples to enforce the flat union property: if two or more adjacent texels do not form a closed 2-ball, we break the connection adding samples at the Voronoi vertex or along the Voronoi edge; if a texel is not a closed 2-ball, we add a sample along its boundary to reduce its coverage.", + "bbox": [ + 212, + 643, + 787, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Low-resolution matching", + "text_level": 1, + "bbox": [ + 215, + 755, + 465, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Let $\\mathcal{M}$ and $\\mathcal{N}$ be two triangular meshes, and let $\\hat{\\mathcal{M}}$ and $\\hat{\\mathcal{N}}$ be their low-resolution counterparts computed with the algorithm presented above (or any other alternative algorithm). Furthermore, let $\\hat{\\Phi}$ be a basis for the functional space over $\\hat{\\mathcal{M}}$ , and $\\hat{\\Psi}$ a basis for the functional space over $\\hat{\\mathcal{N}}$ . For example these bases can", + "bbox": [ + 212, + 777, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "ReMatching: Scalable Shape Correspondence", + "bbox": [ + 431, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "be the truncated subset of the eigenvectors of the LBO as proposed in [47]. We assume to have some pipeline yielding a functional map $\\mathbf{C}$ , such that", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {\\Psi}} \\mathbf {C} \\approx \\hat {\\boldsymbol {\\Pi}} \\hat {\\boldsymbol {\\Phi}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 444, + 186, + 785, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\hat{\\Pi}:\\hat{\\mathcal{M}}\\to \\hat{\\mathcal{N}}$ is a correspondence between the vertices of $\\hat{\\mathcal{M}}$ and the vertices of $\\hat{\\mathcal{N}}$ . The adopted functional maps solution can be any of the available alternatives, ranging from the original one [47] to the most recent [12]. In our experiments, we consider two of the widely adopted solutions: constrained optimization with product preservation [46] and the iterative procedure ZoomOut [38]. The first is representative of the possible optimization strategies, while the second is probably the most efficient refinement technique for functional maps estimation. Both these methods have been efficiently and effectively applied on meshes with a limited number of vertices (i.e., up to 10 thousands vertices), which is exactly the setting we are in after the proposed remeshing step.", + "bbox": [ + 212, + 215, + 784, + 368 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Extending the correspondence", + "text_level": 1, + "bbox": [ + 215, + 391, + 511, + 406 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To extend the basis from the remeshing to the original surface, we need a mapping that transports scalar fields from the low-resolution mesh $\\hat{\\mathcal{M}}$ to the high-resolution shape $\\mathcal{M}$ . A possible solution would be to map each triangle of $\\hat{\\mathcal{M}}$ to a geodesic triangle in $\\mathcal{M}$ , and then computing the geodesic barycentric coordinates of each vertex inside the triangle [31]. Unfortunately, this approach is very costly, as it requires to compute exact geodesic paths onto $\\mathcal{M}$ [56]. Instead, we approximate this mapping by projecting each vertex of $\\mathcal{M}$ onto the closest surface point of $\\hat{\\mathcal{M}}$ . Despite", + "bbox": [ + 212, + 417, + 431, + 718 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5d3f9704c11ded32c90244c4f873dbea69cd461210c21b679871337bdfecef2f.jpg", + "image_caption": [ + "Fig. 4: Top: remeshing and mapping of the original vertices with the solution from Liu et al. [31] and our pipeline. Bottom: comparison of ground truth geodesic distances from multiple vertices and extended with both methods." + ], + "image_footnote": [], + "bbox": [ + 442, + 446, + 787, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "this approach not being accurate in general, every triangle in $\\hat{\\mathcal{M}}$ corresponds to a geodesic triangle (homeomorphic to a disk) on $\\mathcal{M}$ with the same vertices. Furthermore, starting from a farthest point sampling gives evenly spaced samples, mitigating the geometric complexity of geodesic triangles.", + "bbox": [ + 212, + 719, + 782, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A surface point of $\\hat{\\mathcal{M}}$ is either a vertex, or a weighted average of two (if on an edge) or three vertices (if on a triangle). We build a linear map $\\mathbf{U}_{\\mathcal{M}} \\in \\mathbb{R}^{m \\times s}$ representing the projection of each vertex $v_i \\in \\mathcal{M}$ as a linear combination of at most three vertices in $\\hat{\\mathcal{M}}$ (i.e., the vertices encoding the closest surface point).", + "bbox": [ + 212, + 779, + 784, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "F. Maggioli et al.", + "bbox": [ + 271, + 114, + 387, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 4 shows a comparison between our pipeline and the simplification method proposed by Liu et al. [31]. The top row shows the different geometries produced by the two methods in reducing the mesh size by $90\\%$ and how our mapping compares to the approximate geodesic barycentric coordinates. In the bottom row we compare the two methods in approximating a multiple source geodesic distance field, showing that they do not present appreciable differences. More discussion is provided in the supplementary materials.", + "bbox": [ + 212, + 146, + 782, + 252 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Scalar functions can be evaluated at any surface point by interpolating the values at the vertices. Since every vertex of $\\mathcal{M}$ is associated to a point on the surface of $\\hat{\\mathcal{M}}$ , we can use the linear map $\\mathbf{U}_{\\mathcal{M}}$ for extending scalar functions from the $s$ vertices of $\\hat{\\mathcal{M}}$ to the $m$ vertices of $\\mathcal{M}$ .", + "bbox": [ + 212, + 252, + 782, + 311 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We extend the bases $\\hat{\\Phi}$ and $\\hat{\\Psi}$ to the full-resolution meshes as $\\Phi = \\mathbf{U}_{\\mathcal{M}}\\hat{\\Phi}$ and $\\Psi = \\mathbf{U}_{\\mathcal{N}}\\hat{\\Psi}$ . Then, in a similar fashion as done in Equation (3), we search for a correspondence $\\Pi : \\mathcal{M} \\to \\mathcal{N}$ that maps vertices of $\\mathcal{M}$ to vertices of $\\mathcal{N}$ as", + "bbox": [ + 212, + 311, + 782, + 357 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {U} _ {\\mathcal {N}} \\hat {\\boldsymbol {\\Psi}} \\mathbf {C} = \\boldsymbol {\\Psi} \\mathbf {C} \\approx \\boldsymbol {\\Pi} \\boldsymbol {\\Phi} = \\boldsymbol {\\Pi} \\mathbf {U} _ {\\mathcal {M}} \\hat {\\boldsymbol {\\Phi}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 366, + 784, + 383 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where, for the sake of clarity, we explicitly write the equation in terms of the bases $\\hat{\\Phi}$ and $\\hat{\\Psi}$ . Given the extended bases $\\Phi$ and $\\Psi$ , and the same functional map $\\mathbf{C}$ estimated to solve Equation (3), we obtain the correspondence $\\Pi$ via a nearest neighbor search in the spectral space.", + "bbox": [ + 212, + 391, + 782, + 450 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We notice that nearest neighbor algorithms become very slow on large meshes, effectively becoming a possible bottleneck for the pipeline. Our experiments presented in Figure 10 show that, to the scope of this paper, this solution provides satisfying performance. However, we acknowledge that establishing a relationship between (3) and (4) could lead to a more efficient expression of $\\Pi$ in terms of $\\hat{\\Pi}$ , $\\mathbf{U}_{\\mathcal{M}}$ , and $\\mathbf{U}_{\\mathcal{N}}$ . This avenue warrants future exploration.", + "bbox": [ + 212, + 452, + 784, + 542 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Results", + "text_level": 1, + "bbox": [ + 215, + 563, + 323, + 578 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our goal is to evaluate not only the matching accuracy of our method but also its time performance. For this task, we test it on datasets for dense correspondences, such as the SHREC19 dataset [37] and the TOSCA dataset [6]. We also introduce a novel dataset, BadTOSCA, obtained by randomly altering the vertex positions of the TOSCA meshes, to ensure that our method is stable even under circumstances where strong isometry cannot be assumed3.", + "bbox": [ + 212, + 593, + 782, + 681 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our pipeline relies on an efficient remeshing algorithm that can quickly and drastically reduce the size of a mesh by orders of magnitude while still preserving the underlying geometry and metric. In principle, any remeshing algorithm could be adopted in our pipeline, as the map for extending scalar fields discussed in Section 3.3 only requires having two input meshes. We compare our solution against well-established isotropic and anisotropic remeshing algorithms (respectively, IRM [4,21] and ARM [44]), as well as the scalable functional maps (SFM)", + "bbox": [ + 212, + 683, + 782, + 790 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "ReMatching: Scalable Shape Correspondence", + "bbox": [ + 431, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "3 Details on the dataset generation are provided in the supplementary materials. Code for generating the dataset is available at https://github.com/filthynobleman/bad-tosca", + "bbox": [ + 217, + 796, + 782, + 837 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0df69ac21962c60babee3f06b84dd9565e8bf434c4bbad281971c3f671f993f2.jpg", + "image_caption": [ + "Fig. 5: Left: Accuracy curves on the SHREC19 challenge pairs for the tested methods. We also consider ideal approaches where the initial functional map for ZoomOut is given ground truth dashed lines). Right: Cumulative curves of the execution time for evaluated methods. For all methods, we also consider the time required to run the remeshing or resampling step(s)." + ], + "image_footnote": [], + "bbox": [ + 214, + 143, + 501, + 281 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/8a480c5dff1e07ec217422620535403a8112a65022b65da28812005a941a0a73.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 142, + 785, + 280 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "approach proposed by Magnet et al. [35], which produces a subsampling of the vertices in place of a low-resolution mesh.", + "bbox": [ + 212, + 390, + 784, + 419 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We implemented our remeshing algorithm in $\\mathbf{C}++$ , using Eigen [20] and libigl [22], and the matching pipeline in MATLAB4.", + "bbox": [ + 212, + 420, + 784, + 450 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Quality of matching", + "text_level": 1, + "bbox": [ + 214, + 472, + 426, + 488 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To evaluate our shape matching pipeline, we tested our method on the SHREC19 dataset [37]. We can use different shape correspondence techniques to test the entire pipeline. In particular, we compute the functional map on the low-resolution representation using both the widely adopted approach with products preservation (FMaps) [46]", + "bbox": [ + 212, + 498, + 464, + 679 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "and ZoomOut [38]. We compare our method with the same approaches on the fullresolution meshes, as well as against SFM. For a fair comparison, we remesh all the", + "bbox": [ + 212, + 680, + 431, + 768 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/88c8fc2db9eadb7e6a18f4e3ffd8f2fcd6aec6f9100f44c57bd5b2667c6a8820.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAGE (·10-2) ↓AUC ↑
Ours (ZoomOut)*2.4095.12%
SFM*2.4695.22%
Ours (FMaps)6.2388.21%
Ours (ZoomOut)5.8788.82%
SFM16.8470.19%
ARM (ZoomOut)41.5930.34%
IRM (ZoomOut)8.3084.62%
FMaps28.3156.15%
ZoomOut29.3652.02%
", + "bbox": [ + 433, + 527, + 767, + 670 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 1: Average geodesic error and area under the accuracy curve for each tested method on the SHREC19 challenge pairs. The top rows (denoted by \\*) show the performance of our algorithm and SFM under the assumption of having a ground truth functional map for initializing ZoomOut.", + "bbox": [ + 439, + 681, + 787, + 765 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "shapes to 3k vertices with all methods, as Magnet et al. [35] state that this produces the best performance with SFM. To further ease the alignment of the Laplacian spectra, we also post-process the remeshed shapes (produced with", + "bbox": [ + 212, + 770, + 787, + 816 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "F. Maggioli et al.", + "bbox": [ + 271, + 114, + 387, + 128 + ], + "page_idx": 9 + }, + { + "type": "page_footnote", + "text": "4 Code available at https://github.com/filthynobleman/rematching", + "bbox": [ + 217, + 824, + 694, + 840 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/9264c224dec92854c5c1f2018de35375a4fa180ae419ad4d6e8e9900a5b4c4ff.jpg", + "image_caption": [ + "Fig. 6: Accuracy curves and average geodesic error on the TOSCA dataset (left) and the BadTOSCA dataset (right) for the tested methods." + ], + "image_footnote": [], + "bbox": [ + 215, + 142, + 504, + 244 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1a2a8b2fd554caf4414acdf9387979209afd2081b10d7cc3bb3e12c9df6511b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 142, + 785, + 244 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "IRM, ARM, and our algorithm) to remove small disconnected components made of few triangles before computing the extension map discussed in Section 3.3. This post-processing step cannot be performed for SFM, due to the nature of its sampling strategy. The benefits of this post-processing will be discussed in Section 4.3.", + "bbox": [ + 212, + 318, + 782, + 393 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We summarize the results of our experiment in Figure 5, where we show the accuracy curves for all methods on the SHREC19 connectivity track benchmark, following the paradigm proposed by Kim et al. [24]. To prove the time efficiency of our pipeline, we also measure the time taken by every method on each pair of shapes in the benchmark (including the remeshing step), and we show the cumulative curves of the execution times over the entire dataset in a similar way with respect to the accuracy curves.", + "bbox": [ + 212, + 396, + 784, + 502 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/2133191e7d4ccc87aed022750fe3f300b5436d5204dc336a9b4b1b75c80e2205.jpg", + "image_caption": [ + "Fig.7: Coordinate transfer between nonisometric shapes. The coordinates are used to generate highly complex patterns, whose sensitivity to the input enhances the transfer errors." + ], + "image_footnote": [], + "bbox": [ + 218, + 536, + 501, + 642 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We see that all the low-resolution approaches provide better time and quality performance than the fullresolution methods, as the simpler geometry makes it easier and faster to align the Laplacian eigenfunctions. ARM is the only exception, since the anisotropic remeshing produces many skew triangles and a singular Laplacian matrix for all the shapes. Furthermore, our technique (with both the FMaps and ZoomOut backend) outperforms SFM and achieves better results than using the IRM remeshing strategy. This difference can be better", + "bbox": [ + 511, + 505, + 785, + 731 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "appreciated in Table 1, where we report the average geodesic error (normalized with respect to the shape diameter) and the area under the accuracy curves. In both Figure 5 and Table 1 we have also considered the idealized scenario where our method with ZoomOut as backend and SFM are initialized with a ground truth functional map. In this case, the two approaches obtain comparable results.", + "bbox": [ + 212, + 732, + 784, + 806 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We also compare our method with the other approaches on the TOSCA and BadTOSCA datasets, summarizing the results in Figure 6. In the figure, we show", + "bbox": [ + 214, + 809, + 784, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "ReMatching: Scalable Shape Correspondence", + "bbox": [ + 431, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f9949d10a74651ebcfc43cb1804d82f9380865b255c15d7349e1ccfe06d8d654.jpg", + "image_caption": [ + "Fig.8: Inter-class coordinate transfer from the wolf model to a cat and a dog. Our pipeline is not constrained by the standard Laplacian basis and can be used with other approaches as well." + ], + "image_footnote": [], + "bbox": [ + 225, + 146, + 491, + 247 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2028f44b42e83515274e6cfad997c0df65f4b6cfeb0924d5cd08a9c65abfd6f4.jpg", + "image_caption": [ + "Fig. 9: Comparison of function transfer between very high-resolution models with SFM and our approach." + ], + "image_footnote": [], + "bbox": [ + 506, + 143, + 779, + 247 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "the accuracy curves for the tested methods on both datasets, also providing the average geodesic error normalized by the shape diameter. We see that under the favorable and unrealistic conditions offered by TOSCA, where shapes of the same class are almost perfectly isometric, SFM achieves the best results. However, when we consider the slightly altered geometry of BadTOSCA, SFM suffers a performance drop of about $50\\%$ , while our method shows more stability.", + "bbox": [ + 212, + 363, + 785, + 455 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A significant benefit of the functional map approach is its independence from the difference in resolution between the source and target shape. Transferring a function with functional maps yields a much smoother and neat result than directly a point-wise correspondence. In this regard, we test our method in transferring coordinate functions between isometric and non-isometric shapes with substantial differences in resolution and triangulation. In the example from Figure 7 we generated a complex procedural fractal pattern as a function of the coordinates [33]. This pattern is very sensitive to the input, so even slight errors are highly enhanced. We see that SFM cannot precisely map the coordinates of the source mesh, shifting around and distorting the details of the pattern.", + "bbox": [ + 212, + 455, + 787, + 608 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Furthermore, the example in Figure 8 shows that we can also deal with inter-class function transfer with our pipeline, outperforming SFM in a visual comparison. In the example, we also showcase that we should not necessarily rely on the standard Laplacian basis as a backend. Instead of using ZoomOut to transfer functions between the low-resolution meshes, we rely on the orthogonalized eigenproducts basis introduced by Maggioli et al. [34], extending it with the same technique described in Section 3.2.", + "bbox": [ + 212, + 609, + 787, + 717 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Finally, to ensure that our method can scale to high-resolution meshes, we tested it using two very high-resolution 3D scans of real statues: a dancing faun counting $\\sim 750\\mathrm{k}$ vertices and a warrior from the temple of Aphaea counting $\\sim 3.5\\mathrm{M}$ vertices (see Figures 1 and 9). Since SFM can also deal with dense shapes, we compare it against our method. Figure 9 shows an example of the transfer of a geodesic distance function. While our method can faithfully transfer the function, SFM produces incoherent distances (e.g., wrong isolines on the head and near the shoulder) and evident artefacts (e.g., wrong distance on the hand).", + "bbox": [ + 212, + 719, + 787, + 842 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "F. Maggioli et al.", + "bbox": [ + 271, + 114, + 387, + 128 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 Scalable performance", + "text_level": 1, + "bbox": [ + 215, + 146, + 437, + 161 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Section 4.1 we saw that general-purpose remeshing solutions are not an optimal choice for the matching task. However, Lescoat et al. [28] provided a specialized mesh decimation algorithm (SMD) that reduces the complexity of the mesh while still preserving the Laplacian eigenfunctions. Furthermore, Liu et al. [31] proposed an elegant solution for simplifying a shape based on intrinsic error metric (IEM), which also builds a map for extending scalar functions using geodesic barycentric coordinates. We perform the evaluation on the SHREC19 dataset [37], remeshing the surfaces to 3k vertices with all the methods.", + "bbox": [ + 212, + 176, + 486, + 446 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To achieve the desired scalability, the remeshing step in our pipeline must be fast enough to maintain the improvement obtained by reducing the size of the eigenproblem. In this regard, Figure 10 shows that our technique achieves better time performance than the other remeshing algorithms and the sampling strategy of SFM. In contrast, SMD and IEM are decimation algorithms that iteratively reduce the mesh size, making them orders of magnitude slower and unsuitable for large meshes. Moreover, SMD requires a pre-computation of the Laplacian eigenbasis on the original surface to guide the decimation process and preserve the spectrum, introducing additional time complexity.", + "bbox": [ + 212, + 449, + 488, + 751 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Benefits of remeshing", + "text_level": 1, + "bbox": [ + 215, + 779, + 441, + 794 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/33fa98f4caa616ea99b1f70e11bb472b12ebdee962d007dee1cdc28d9635f51a.jpg", + "image_caption": [ + "Fig.10: Execution time (logarithmic scale on the $y$ -axis) of different remeshing algorithms plotted against the number of vertices of the input shape ( $x$ -axis). For SFM we show the time taken for generating the sampling." + ], + "image_footnote": [], + "bbox": [ + 498, + 207, + 787, + 344 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ff40afdeb38a5cc92852b474a2e754ef6bc15fa1a8e05a97edabb40fe3ada0ee.jpg", + "image_caption": [ + "Fig.11: Top row: coordinate transfer between a pair from SHREC19 using ZoomOut on full-resolution meshes, SFM and our approach. Bottom row: The ground truth $20 \\times 20$ functional map, compared with the alignment computed with FMaps [46] from the full Laplacian eigenbasis, the SFM extended basis and our extended basis." + ], + "image_footnote": [], + "bbox": [ + 500, + 479, + 785, + 652 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Section 4.1, we discussed how we remove small disconnected components after the remeshing step. The example in Figure 11 shows a comparison of mapping", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "ReMatching: Scalable Shape Correspondence", + "bbox": [ + 431, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "between a low-resolution mesh (Source), with less than 7k vertices, to a detailed mesh where eyes are represented with two disconnected components, each composed of 462 vertices out of the 27k of the entire mesh ( $\\sim 2\\%$ of the total vertices). This is enough geometry to attract energy at lower frequencies, and when we try to align the first 20 eigenfunctions with FMaps [46], the resulting functional map turns out to be mostly noise. Initializing ZoomOut with such a map leads to meaningless correspondence, and SFM does not mitigate this issue, as it tries to preserve the original spectrum as much as possible. In contrast, by removing these components and mapping them to their closest surface points, we introduce a small error in the final correspondence, but at the same time, we ease the alignment of the initial eigenbases, resulting in more accurate and meaningful mapping. This is evident in the bottom row of Figure 11, where the ground truth functional map presents a gap in the alignment of the eigenfunctions due to the impossibility of mapping the disconnected eyes of the target shape in any point of the source shape. Using the real eigenfunctions or a close approximation yields an intense noise in the map, producing the meaningless mapping of ZoomOut and SFM. Instead, by mapping the eyes to their nearest point on the head, we can meaningfully align our extended basis.", + "bbox": [ + 212, + 146, + 787, + 420 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 470, + 369, + 486 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We presented a new pipeline for scalable shape matching that relies on a low-resolution representation to compute a functional map between dense shapes efficiently. A core part of this pipeline is our novel remeshing algorithm, which efficiently computes an intrinsic Delaunay triangulation of uniform surface sampling. This procedure is applied to reduce the computational cost of the functional maps framework and make it efficient also on very dense meshes. Furthermore, we use a fast and intuitive technique for extending scalar functions from low-resolution meshes to their dense counterparts, extending the computed correspondence to the original very high-resolution shapes. Our experimental evaluation proved that our procedure is effective on various types of shapes at different resolutions, outperforming other state-of-the-art solutions and solving topological issues related to the alignment of the Laplacian eigenbasis.", + "bbox": [ + 212, + 532, + 787, + 715 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "While using a triangulated mesh as a low-resolution representation could, in principle, be exploited with other types of shape-matching pipelines, we only tested it within the functional maps framework, using the Laplace-Beltrami eigenbasis or bases derived from it. While the current pipeline binds the method's applicability to the setting of functional correspondences, we believe that studying its behaviour with different \"backends\", as well as exploring the relationships between the low-resolution and the high-resolution correspondences, would be an interesting matter for future investigations.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "F. Maggioli et al.", + "bbox": [ + 271, + 114, + 387, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was partially supported by the Italian Ministry of Education, Universities and Research under the grant Dipartimento di Eccellenza 2023-2027 of the Department of Informatics, Systems and Communication of the University of Milano-Bicocca and by the PRIN 2022 project GEOPRIDE Geometric primitive fitting on 3D data for geometric analysis and 3D shapes. This work is also supported by the ERC grant no.802554 SPECGEO and PRIN 2020 project no.2020TA3K9N LEGO.AI. We acknowledge the support of NVIDIA Corporation with the RTX A5000 GPUs granted through the Academic Hardware Grant Program to the University of Milano-Bicocca for the project Learned representations for implicit binary operations on real-world 2D-3D data.", + "bbox": [ + 212, + 174, + 787, + 325 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 347, + 321, + 363 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aflalo, Y., Brezis, H., Kimmel, R.: On the optimality of shape and data representation in the spectral domain. SIAM Journal on Imaging Sciences 8(2), 1141-1160 (2015)", + "2. Amenta, N., Choi, S., Dey, T.K., Leekha, N.: A simple algorithm for homeomorphic surface reconstruction. In: Proceedings of the sixteenth annual symposium on Computational geometry. pp. 213-222 (2000)", + "3. Bobenko, A.I., Springborn, B.A.: A discrete laplace-beltrami operator for simplicial surfaces. Discrete & Computational Geometry 38(4), 740-756 (2007)", + "4. Botsch, M., Pauly, M., Rossl, C., Bischoff, S., Kobbelt, L.: Geometric modeling based on triangle meshes. In: ACM SIGGRAPH 2006 Courses, pp. 1-es (2006)", + "5. Bowyer, A.: Computing Dirichlet tessellations*. The Computer Journal 24(2), 162-166 (01 1981)", + "6. Bronstein, A., Bronstein, M., Kimmel, R.: Numerical Geometry of Non-Rigid Shapes. Springer, New York, NY (2008)", + "7. Burghard, O., Dieckmann, A., Klein, R.: Embedding shapes with green's functions for global shape matching. Computers & Graphics 68, 1-10 (2017)", + "8. Crane, K., Weischedel, C., Wardetzky, M.: Geodesics in heat: A new approach to computing distance based on heat flow. ACM Transactions on Graphics (TOG) 32(5), 1-11 (2013)", + "9. Delaunay, B., et al.: Sur la sphere vide. Izv. Akad. Nauk SSSR, Otdelenie Matematicheskii i Estestvennyka Nauk 7(793-800), 1-2 (1934)", + "0. Deng, B., Yao, Y., Dyke, R.M., Zhang, J.: A survey of non-rigid 3d registration. In: Computer Graphics Forum. vol. 41, pp. 559-589. Wiley Online Library (2022)", + "1. Dijkstra, E.W.: A note on two problems in connexion with graphs. Numerische Mathematik 1, 269-271 (1959)", + "2. Donati, N., Corman, E., Melzi, S., Ovsjanikov, M.: Complex functional maps: A conformal link between tangent bundles. In: Computer Graphics Forum. vol. 41, pp. 317-334. Wiley Online Library (2022)", + "3. Donati, N., Sharma, A., Ovsjanikov, M.: Deep geometric functional maps: Robust feature learning for shape correspondence. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8592-8601 (2020)", + "4. Dyer, R., Zhang, H., Möller, T.: Voronoi-delaunay duality and delaunay meshes. In: Proceedings of the 2007 ACM symposium on Solid and physical modeling. pp. 415-420 (2007)" + ], + "bbox": [ + 225, + 376, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "ReMatching: Scalable Shape Correspondence", + "bbox": [ + 431, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Edelsbrunner, H., Shah, N.R.: Triangulating topological spaces. In: Proceedings of the tenth annual symposium on Computational geometry. pp. 285-292 (1994)", + "16. Eynard, D., Rodola, E., Glashoff, K., Bronstein, M.M.: Coupled functional maps. In: 2016 Fourth International Conference on 3D Vision (3DV). pp. 399-407. IEEE (2016)", + "17. Ezuz, D., Ben-Chen, M.: Deblurring and denoising of maps between shapes. In: Computer Graphics Forum. vol. 36, pp. 165-174. Wiley Online Library (2017)", + "18. Fei, Y., Chen, S., Su, D., Luo, J., Li, M.: A new algorithm for repairing non-manifold surfaces. In: 2013 IEEE 10th International Conference on High Performance Computing and Communications & 2013 IEEE International Conference on Embedded and Ubiquitous Computing. pp. 1704-1708. IEEE (2013)", + "19. Gao, M., Roetzer, P., Eisenberger, M., Lähner, Z., Moeller, M., Cremers, D., Bernard, F.: Sigma: Scale-invariant global sparse shape matching. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 645-654 (2023)", + "20. Guennebaud, G., Jacob, B., et al.: Eigen v3. http://eigen.tuxfamily.org (2010)", + "21. Hoppe, H., DeRose, T., Duchamp, T., McDonald, J., Stuetzle, W.: Mesh optimization. In: Proceedings of the 20th annual conference on Computer graphics and interactive techniques. pp. 19-26 (1993)", + "22. Jacobson, A., Panozzo, D., et al.: libigl: A simple $\\mathrm{C}++$ geometry processing library (2018), https://libigl.github.io/", + "23. Jiang, Z., Schneider, T., Zorin, D., Panozzo, D.: Bijective projection in a shell. ACM Transactions on Graphics (TOG) 39(6), 1-18 (2020)", + "24. Kim, V.G., Lipman, Y., Funkhouser, T.: Blended intrinsic maps. ACM transactions on graphics (TOG) 30(4), 1-12 (2011)", + "25. Kimmel, R., Sethian, J.A.: Computing geodesic paths on manifolds. Proceedings of the national academy of Sciences 95(15), 8431-8435 (1998)", + "26. Kressner, D.: Numerical Methods for General and Structured Eigenvalue Problems. Lecture Notes in Computational Science and Engineering, Springer Berlin Heidelberg (2006)", + "27. Leibon, G., Letscher, D.: Delaunay triangulations and voronoi diagrams for riemannian manifolds. In: Proceedings of the sixteenth annual symposium on Computational geometry. pp. 341-349 (2000)", + "28. Lescoat, T., Liu, H.T.D., Thiery, J.M., Jacobson, A., Boubekeur, T., Ovsjanikov, M.: Spectral mesh simplification. In: Computer Graphics Forum. vol. 39, pp. 315-324. Wiley Online Library (2020)", + "29. Levy, B.: Laplace-beltrami eigenfunctions towards an algorithm that \"understands\" geometry. In: IEEE International Conference on Shape Modeling and Applications 2006 (SMI'06). pp. 13-13 (2006)", + "30. Lévy, B., Zhang, H.: Spectral mesh processing. In: ACM SIGGRAPH 2010 Courses, pp. 1-312 (2010)", + "31. Liu, H.T.D., Gillespie, M., Chislett, B., Sharp, N., Jacobson, A., Crane, K.: Surface simplification using intrinsic error metrics. ACM Trans. Graph. 42(4) (jul 2023). https://doi.org/10.1145/3592403, https://doi.org/10.1145/3592403", + "32. Liu, H.T.D., Jacobson, A., Ovsjanikov, M.: Spectral coarsening of geometric operators. ACM Transactions on Graphics (TOG) 38(4) (jul 2019)", + "33. Maggioli, F., Baieri, D., Melzi, S., Rodolà, E.: Newton's fractals on surfaces via bicomplex algebra. In: ACM SIGGRAPH 2022 Posters, pp. 1-2 (2022)", + "34. Maggioli, F., Melzi, S., Ovsjanikov, M., Bronstein, M.M., Rodola, E.: Orthogonalized fourier polynomials for signal approximation and transfer. In: Computer Graphics Forum. vol. 40, pp. 435-447. Wiley Online Library (2021)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "F. Maggioli et al.", + "bbox": [ + 271, + 114, + 387, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Magnet, R., Ovsjanikov, M.: Scalable and efficient functional map computations on dense meshes. In: Computer Graphics Forum. vol. 42, pp. 89-101. Wiley Online Library (2023)", + "36. Marin, R., Corona, E., Pons-Moll, G.: Nicp: Neural icip for 3d human registration at scale. In: European Conference on Computer Vision. Springer (2024)", + "37. Melzi, S., Marin, R., Rodolà, E., Castellani, U., Ren, J., Poulenard, A., Wonka, P., Ovsjanikov, M.: Shrec 2019: Matching humans with different connectivity. In: Eurographics Workshop on 3D Object Retrieval. vol. 7, p. 3. The Eurographics Association (2019)", + "38. Melzi, S., Ren, J., Rodola, E., Sharma, A., Wonka, P., Ovsjanikov, M.: Zoomout: spectral upsampling for efficient shape correspondence. ACM Transactions on Graphics (TOG) 38(6), 1-14 (2019)", + "39. Melzi, S., Rodolà, E., Castellani, U., Bronstein, M.M.: Localized manifold harmonics for spectral shape analysis. In: Computer Graphics Forum. vol. 37, pp. 20-34. Wiley Online Library (2018)", + "40. Meyer, M., Desbrun, M., Schröder, P., Barr, A.H.: Discrete Differential-Geometry Operators for Triangulated 2-Manifolds. In: Visualization and mathematics III, pp. 35-57. Springer, New York, NY (2003)", + "41. Mitchell, J.S., Mount, D.M., Papadimitriou, C.H.: The discrete geodesic problem. SIAM Journal on Computing 16(4), 647-668 (1987)", + "42. Nasikun, A., Brandt, C., Hildebrandt, K.: Fast approximation of laplace-beltrami eigenproblems. In: Computer Graphics Forum. vol. 37, pp. 121-134. Wiley Online Library (2018)", + "43. Nasikun, A., Hildebrandt, K.: The hierarchical subspace iteration method for laplace-beltrami eigenproblems. ACM Transactions on Graphics (TOG) 41(2), 1-14 (2022)", + "44. Nivoliers, V., Lévy, B., Geuzaine, C.: Anisotropic and feature sensitive triangular remeshing using normal lifting. Journal of Computational and Applied Mathematics 289, 225-240 (2015)", + "45. Nogneng, D., Melzi, S., Rodola, E., Castellani, U., Bronstein, M., Ovsjanikov, M.: Improved functional mappings via product preservation. In: Computer Graphics Forum. vol. 37, pp. 179-190. Wiley Online Library (2018)", + "46. Nogneng, D., Ovsjanikov, M.: Informative descriptor preservation via commutativity for shape matching. In: Computer Graphics Forum. vol. 36, pp. 259-267. Wiley Online Library (2017)", + "47. Ovsjanikov, M., Ben-Chen, M., Solomon, J., Butscher, A., Guibas, L.: Functional maps: a flexible representation of maps between shapes. ACM Transactions on Graphics (ToG) 31(4), 1-11 (2012)", + "48. Ovsjanikov, M., Corman, E., Bronstein, M., Rodola, E., Ben-Chen, M., Guibas, L., Chazal, F., Bronstein, A.: Computing and processing correspondences with functional maps. In: SIGGRAPH ASIA 2016 Courses, pp. 1-60 (2016)", + "49. Peyré, G., Cohen, L.D.: Geodesic remeshing using front propagation. International Journal of Computer Vision 69, 145-156 (2006)", + "50. Pinkall, U., Polthier, K.: Computing Discrete Minimal Surfaces and their Conjugates. Experimental mathematics 2(1), 15-36 (1993)", + "51. Ren, J., Melzi, S., Ovsjanikov, M., Wonka, P., et al.: Maptree: recovering multiple solutions in the space of maps. ACM Transactions on Graphics (TOG) 39(6), 264-1 (2020)", + "52. Ren, J., Melzi, S., Wonka, P., Ovsjanikov, M.: Discrete optimization for shape matching. In: Computer Graphics Forum. vol. 40, pp. 81-96. Wiley Online Library (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "ReMatching: Scalable Shape Correspondence", + "bbox": [ + 431, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "53. Ren, J., Poulenard, A., Wonka, P., Ovsjanikov, M.: Continuous and orientation-preserving correspondences via functional maps. ACM Transactions on Graphics (ToG) 37(6), 1-16 (2018)", + "54. Reuter, M., Wolter, F.E., Peinecke, N.: Laplace-beltrami spectra as 'shape-dna' of surfaces and solids. Computer-Aided Design 38(4), 342-366 (2006)", + "55. Rivin, I.: Euclidean structures on simplicial surfaces and hyperbolic volume. Annals of mathematics 139(3), 553-580 (1994)", + "56. Rustamov, R.M.: Barycentric coordinates on surfaces. In: Computer Graphics Forum. vol. 29, pp. 1507-1516. Wiley Online Library (2010)", + "57. Sahillioğlu, Y.: Recent advances in shape correspondence. The Visual Computer 36(8), 1705-1721 (2020)", + "58. Shoham, M., Vaxman, A., Ben-Chen, M.: Hierarchical functional maps between subdivision surfaces. In: Computer Graphics Forum. vol. 38, pp. 55-73. Wiley Online Library (2019)", + "59. Vaxman, A., Ben-Chen, M., Gotsman, C.: A multi-resolution approach to heat kernels on discrete surfaces. In: ACM SIGGRAPH 2010 papers, pp. 1-10 (2010)", + "60. Watson, D.F.: Computing the n-dimensional Delaunay tessellation with application to Voronoi polytopes*. The Computer Journal 24(2), 167-172 (01 1981)" + ], + "bbox": [ + 215, + 146, + 785, + 396 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "F. Maggioli et al.", + "bbox": [ + 271, + 114, + 387, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/ReMatching_ Low-Resolution Representations for Scalable Shape Correspondence/8dd15a7c-48bf-42e4-a752-0b0dbad83c91_model.json b/2024/ReMatching_ Low-Resolution Representations for Scalable Shape Correspondence/8dd15a7c-48bf-42e4-a752-0b0dbad83c91_model.json new file mode 100644 index 0000000000000000000000000000000000000000..434c45e12e5420e4ad18d2611d8231d179993ac8 --- /dev/null +++ b/2024/ReMatching_ Low-Resolution Representations for Scalable Shape Correspondence/8dd15a7c-48bf-42e4-a752-0b0dbad83c91_model.json @@ -0,0 +1,2656 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.221, + 0.141, + 0.783, + 0.187 + ], + "angle": 0, + "content": "ReMatching: Low-Resolution Representations for Scalable Shape Correspondence" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.212, + 0.761, + 0.244 + ], + "angle": 0, + "content": "Filippo Maggioli1, Daniele Baieri2, Emanuele Rodolà2, and Simone Melzi1" + }, + { + "type": "text", + "bbox": [ + 0.35, + 0.255, + 0.65, + 0.269 + ], + "angle": 0, + "content": "1 University of Milano-Bicocca, Milan, Italy" + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.27, + 0.655, + 0.284 + ], + "angle": 0, + "content": "2 Sapienza - University of Rome, Rome, Italy" + }, + { + "type": "list", + "bbox": [ + 0.348, + 0.255, + 0.655, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.324, + 0.785, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.506, + 0.789, + 0.577 + ], + "angle": 0, + "content": "Fig. 1: Correspondence between two dense shapes computed with our method, using ZoomOut [38] on low-resolution meshes. The coordinates functions of the dancing faun statue on the left (\\(\\sim750\\mathrm{k}\\) vertices) are transferred using a functional map to the Aphaea warrior statue on the right (\\(\\sim3.5\\mathrm{M}\\) vertices). The colors encode the correspondence. Despite the mesh density, shown in the close-up, the computation took \\(\\sim2\\) minutes." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.633, + 0.744, + 0.842 + ], + "angle": 0, + "content": "Abstract. We introduce ReMatching, a novel shape correspondence solution based on the functional maps framework. Our method, by exploiting a new and appropriate re-meshing paradigm, can target shape-matching tasks even on meshes counting millions of vertices, where the original functional maps does not apply or requires a massive computational cost. The core of our procedure is a time-efficient remeshing algorithm which constructs a low-resolution geometry while acting conservatively on the original topology and metric. These properties allow translating the functional maps optimization problem on the resulting low-resolution representation, thus enabling efficient computation of correspondences with functional map approaches. Finally, we propose an efficient technique for extending the estimated correspondence to the original meshes. We show that our method is more efficient and effective through quantitative and qualitative comparisons, outperforming state-of-the-art pipelines in quality and computational cost." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.129 + ], + "angle": 0, + "content": "F. Maggioli et al." + }, + { + "type": "header", + "bbox": [ + 0.262, + 0.148, + 0.735, + 0.162 + ], + "angle": 0, + "content": "Keywords: Shape matching \\(\\cdot\\) Functional maps \\(\\cdot\\) Geometry processing" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.189, + 0.544, + 0.205 + ], + "angle": 0, + "content": "1 Introduction and related work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.221, + 0.786, + 0.356 + ], + "angle": 0, + "content": "The task of finding a semantically meaningful correspondence between discrete surfaces has always been a fundamental topic in the field of shape analysis. Researchers developed a wealth of solutions for this application, and new methods continue to be designed [10, 57]. Among the various approaches, the functional maps framework [47, 48] received significant attention. Rather than finding a point-wise correspondence, the functional maps framework aims to define a correspondence between functions, encoding it into a small matrix. Following this direction, a variety of works tried to improve the actual computation of the map [13, 38, 53] or to extend the framework to different types of bases [34, 39, 45]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.357, + 0.787, + 0.599 + ], + "angle": 0, + "content": "Despite the amount of research outcomes based on the functional maps framework, both exploiting geometric properties and tools [7, 16, 17] and clever optimization techniques [46, 51, 52], the problem of computing the map is still bounded by the time complexity of sparse eigendecompositions [26]. To overcome this problem, researchers started investigating scalable solutions to the Laplace-Beltrami eigenproblem. In this regard, multi-resolution techniques [42, 43, 58, 59] proved to be very effective for scalable computation of spectral quantities for tasks such as retrieval and mesh filtering [30, 54], but their applicability to the functional map framework proved to require additional care [35]. Furthermore, recent research proposed spectral coarsening methods [32] or spectral preserving simplification techniques [28]. However, these techniques usually rely on the computation of the full-resolution spectrum, making them unsuitable for large-scale applications. Gao et al. [19] proposed a mixed-integer programming scalable solution to the matching problem, but they specifically target sparse correspondences only. Alternatively, Marin et al. [36], who provide an approach for human registration that scales across large datasets, but it is unsuitable for dense shapes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.735 + ], + "angle": 0, + "content": "Closely related to our work is the contribution from Magnet et al. [35], which, for the first time, proposes a solution for scaling the functional maps approach to high-resolution meshes. Their idea is to reduce the dimensionality of the eigenproblem, inducing a linear relationship between the functional space on the mesh and a lower-dimension functional space on a sparse sample. The extreme sparsity of the sampling, combined with its efficient computation, leads to an algorithm that can effectively deal with high-density meshes. Nevertheless, this type of reduction does not mitigate the negative influence that small components and local details at high frequency could have on the alignment of the spectra." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Finally, our pipeline relies on building robust and sparse representations of dense meshes while still keeping an approximately bijective correspondence between the high- and low-resolution shapes, and it is worth mentioning that other works have already attempted to do so. Jiang et al. [23] propose building a lower resolution prismatic shell preserving bijectivity with the original surface. However, their algorithm does not scale to meshes with a high triangle count, gives no control and no guarantees on the final vertex count, and requires many assumption." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.432, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ReMatching: Scalable Shape Correspondence" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.265 + ], + "angle": 0, + "content": "tions on the input shape (e.g., manifoldness, orientability, no self-intersections). On the other hand, Liu et al. [31] propose to exploit an intrinsic error metric for decimating a mesh to a given size while keeping track of a geodesic barycentric mapping of each triangle of the resulting shape to a geodesic triangle on the input surface. Despite the robustness of the method, the simplification approach makes it unsuitable for a functional maps setting, where it is required to decimate shapes from millions of vertices to a few thousand. This limitation is further discussed in Section 4.2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.784, + 0.312 + ], + "angle": 0, + "content": "We introduce a new scalable functional map pipeline that efficiently handles meshes with high vertex density (see Figure 1), yields stable results, and is not bounded by the quality of the original triangulation. To summarize:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.314, + 0.784, + 0.358 + ], + "angle": 0, + "content": "- we translate the matching pipeline to low-resolution representations, enabling a fast and scalable computation of functional maps between dense shapes, providing the first alternative to [35];" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.359, + 0.784, + 0.388 + ], + "angle": 0, + "content": "- we propose a new efficient and geometry-preserving remeshing algorithm specifically designed for our pipeline;" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.389, + 0.784, + 0.433 + ], + "angle": 0, + "content": "- we exploit a fast solution for extending scalar maps from the low-resolution remeshed shape to the original surface, thus efficiently obtaining suitable bases for function transfer and shape matching." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.314, + 0.784, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.458, + 0.368, + 0.475 + ], + "angle": 0, + "content": "2 Background" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.784, + 0.536 + ], + "angle": 0, + "content": "In the discrete setting, we represent a shape as a triangular mesh \\(\\mathcal{M} = (V,E,T)\\) where (i) \\(V\\subset \\mathbb{R}^3\\) is a set of vertices; (ii) \\(E\\subset V^2\\) is a set of edges between vertices; (iii) \\(T\\subset V^3\\) is a set of triangles composing the surface." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.558, + 0.398, + 0.574 + ], + "angle": 0, + "content": "2.1 Functional maps" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.785, + 0.688 + ], + "angle": 0, + "content": "We discretize a scalar function \\( f: \\mathcal{M} \\to \\mathbb{R} \\) as a signal defined on the vertices \\( V \\) and represented as a vector \\( \\pmb{f} \\in \\mathbb{R}^{|V|} \\). Hence, the Laplace-Beltrami operator \\( \\varDelta_{\\mathcal{M}}: \\mathcal{F}(\\mathcal{M}) \\to \\mathcal{F}(\\mathcal{M}) \\) is discretized as a sparse matrix \\( \\mathbf{L}_{\\mathcal{M}} \\in \\mathbb{R}^{|V| \\times |V|} \\), generally represented by means of a stiffness matrix \\( \\mathbf{S}_{\\mathcal{M}} \\) and a mass matrix \\( \\mathbf{A}_{\\mathcal{M}} \\) [40, 50]. The eigendecomposition \\( \\mathbf{S}_{\\mathcal{M}} \\pmb{\\Phi}_{\\mathcal{M}} = \\mathbf{A}_{\\mathcal{M}} \\pmb{\\Phi}_{\\mathcal{M}} \\varLambda_{\\mathcal{M}} \\) of the Laplacian yields an orthonormal basis for the functional space on \\( \\mathcal{M} \\) [29,30] which has analogies with the Fourier basis and optimally represents smooth functions when truncated [1]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.785, + 0.764 + ], + "angle": 0, + "content": "Given two shapes \\(\\mathcal{M}\\) and \\(\\mathcal{N}\\), respectively with \\(m\\) and \\(n\\) vertices, a point-to-point map \\(\\Pi : \\mathcal{M} \\to \\mathcal{N}\\) can be expressed as a binary matrix \\(\\Pi \\in \\{0,1\\}^{m \\times n}\\) such that \\(\\Pi(i,j) = 1\\) if the vertex \\(j\\)-th of \\(\\mathcal{N}\\) corresponds to the vertex \\(i\\)-th of \\(\\mathcal{M}\\), and \\(\\Pi(i,j) = 0\\) otherwise. Any point-wise correspondence \\(\\Pi\\) establishes a functional map \\(T_{\\Pi} : \\mathcal{F}(\\mathcal{N}) \\to \\mathcal{F}(\\mathcal{M})\\) as \\(T_{\\Pi}(f) = f \\circ \\Pi\\), \\(\\forall f \\in \\mathcal{F}(\\mathcal{N})\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.784, + 0.84 + ], + "angle": 0, + "content": "As proposed in [47], given \\(\\Phi_k\\) and \\(\\Psi_k\\) the Laplacian bases truncated to size \\(k\\) respectively on \\(\\mathcal{M}\\) and \\(\\mathcal{N}\\), we can project \\(T_{\\Pi}\\) in these bases and compactly encode the functional map in a \\(k \\times k\\) matrix \\(\\mathbf{C} = \\Phi_k^\\top \\mathbf{A}_{\\mathcal{M}} \\Pi \\Psi_k\\), which is the linear operator that transfers the coefficients of functions from \\(\\mathcal{F}(\\mathcal{N})\\) in the ones of corresponding functions of \\(\\mathcal{F}(\\mathcal{M})\\) respectively computed with \\(\\Phi_k\\) and \\(\\Psi_k\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.129 + ], + "angle": 0, + "content": "F. Maggioli et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.194 + ], + "angle": 0, + "content": "Finally, given a functional map \\( T_{\\Pi} \\), or its compact representation \\( \\mathbf{C} \\), we can retrieve the correspondence \\( \\Pi \\) on which the map is built by the nearest neighbor search in the embedding space of the Laplacian eigenfunctions as detailed in [48]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.212, + 0.591, + 0.228 + ], + "angle": 0, + "content": "2.2 Intrinsic Delaunay triangulations (IDT)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.235, + 0.784, + 0.298 + ], + "angle": 0, + "content": "The IDT [3, 55] has been introduced for generalizing the Delaunay triangulation [9] to non-Euclidean metric spaces where classical algorithms [5, 60] cannot be used. IDT relies on the duality with the Voronoi diagram [15, 27], as three texels meeting at a point form a triangle whose edges cross the texels boundaries." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.303, + 0.785, + 0.351 + ], + "angle": 0, + "content": "Definition 1. Let \\((\\mathcal{M},d_{\\mathcal{M}})\\) be a 2-dimensional metric space (with distance function \\(d_{\\mathcal{M}}\\)), and let \\(S = \\{p_i\\}_{i = 1}^s\\subset \\mathcal{M}\\) be a set of \\(s\\) sample points. The Voronoi decomposition (VD) of \\(\\mathcal{M}\\) with respect to \\(S\\) is the collection \\(\\{P_i\\}_{i = 1}^s\\) such that" + }, + { + "type": "equation", + "bbox": [ + 0.285, + 0.358, + 0.786, + 0.398 + ], + "angle": 0, + "content": "\\[\nP _ {i} = \\left\\{q \\in \\mathcal {M}: p _ {i} = \\underset {p _ {j} \\in S} {\\arg \\min } \\left(d _ {\\mathcal {M}} \\left(p _ {j}, q\\right)\\right) \\right\\}, \\quad \\bigcup_ {i = 1} ^ {s} P _ {i} = \\mathcal {M} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.405, + 0.771, + 0.421 + ], + "angle": 0, + "content": "The points in \\( S \\) are called generators, and the \\( P_{i} \\) are called texels (or cells)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.428, + 0.784, + 0.502 + ], + "angle": 0, + "content": "The intersection between two or more texels could be non-empty. In such a case, we define those texels as adjacent. A VD is general if the intersection of four or more texels is empty. If a VD is general, a Voronoi vertex is a point that belongs to three texels, and a Voronoi edge is a curve \\( C \\) connecting two Voronoi vertices and belonging to two texels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.784, + 0.567 + ], + "angle": 0, + "content": "It can be shown that there is a one-to-one correspondence between a Voronoi edge dividing \\( P_{i} \\) and \\( P_{j} \\) and a Delaunay edge connecting \\( p_{i} \\) to \\( p_{j} \\) [14]. This correspondence gives rise to a necessary and sufficient condition for the IDT to be a proper triangulation (i.e., it realizes a simplicial complex)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.571, + 0.784, + 0.601 + ], + "angle": 0, + "content": "Definition 2 ([2,15]). If \\( S \\) induces a general \\( VD \\), the \\( VD \\) satisfies the closed ball property if:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.602, + 0.71, + 0.616 + ], + "angle": 0, + "content": "- every \\( P_{i} \\) is a closed 2-ball (i.e., a topological disk without holes);" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.617, + 0.786, + 0.631 + ], + "angle": 0, + "content": "- every \\( P_i \\cap P_j \\) is either empty or a closed 1-ball (i.e., a topological segment);" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.632, + 0.725, + 0.647 + ], + "angle": 0, + "content": "- every \\( P_i \\cap P_j \\cap P_k \\) is either empty or a closed 0-ball (i.e., a point)." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.602, + 0.786, + 0.647 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.654, + 0.784, + 0.685 + ], + "angle": 0, + "content": "Theorem 1 ([2,14,15]). If \\( S \\) induces a general \\( VD \\), then the \\( VD \\) satisfies the closed ball property if and only if its dual \\( IDT \\) is a proper triangulation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.705, + 0.331, + 0.721 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.784, + 0.765 + ], + "angle": 0, + "content": "We introduce a new scalable functional map pipeline handling very dense meshes and yielding stable results independently of the input triangulation's quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.785, + 0.841 + ], + "angle": 0, + "content": "Our method acts through three main steps. At first, we compute two low-resolution meshes that preserve the original metrics of the input shapes (Section 3.1). Then, we apply any existing pipeline, yielding a functional map between the low-resolution eigenspaces (Section 3.2). Finally, we extend the eigenfunctions to the original meshes, allowing us to use them with the functional map" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.432, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "ReMatching: Scalable Shape Correspondence" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "estimated in the previous step to compute the desired correspondence between them (Section 3.3). Our technique is very general and works with arbitrary manifold triangulations, including meshes with degenerate geometry, non-orientable surfaces, and multiple connected components. The manifoldness requirement does not provide an obstacle to the applicability of our algorithm, as it can be efficiently enforced using existing techniques [18]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.258, + 0.506, + 0.274 + ], + "angle": 0, + "content": "3.1 Intrinsic Delaunay remeshing" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.282, + 0.489, + 0.524 + ], + "angle": 0, + "content": "Front propagation and FPS. Naively computing the VD induced by a set of samples can quickly become very expensive. The exact geodesic algorithm introduced by Mitchell et al. [41] is computationally unfeasible on large meshes. Even using faster solutions, such as the heat method [8], the fast marching algorithm [25], or even the Dijkstra distance [11], searching the closest sample for each vertex is still a \\(\\mathcal{O}(|V|s)\\) operation. Instead, we take inspiration from the front propagation method proposed by Peyre et al. [49]. We start with a decomposition of the mesh induced by a single sample \\(p_1\\)." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.313, + 0.785, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.494, + 0.788, + 0.522 + ], + "angle": 0, + "content": "Fig.2: The iterative front propagation from samples inducing a geodesic VD." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.659 + ], + "angle": 0, + "content": "This means that every vertex \\( v \\) is part of the texel \\( P_{1} \\) and has distance \\( D_{1}[v] = d_{\\mathcal{M}}(p_{1}, v) \\) from the sample set. We then assume to have a decomposition \\( P_{1}, \\dots, P_{k-1} \\) induced by samples \\( p_{1}, \\dots, p_{k-1} \\), and a vector \\( D_{k-1} \\) storing the distance of each vertex from the sample set. When adding a sample \\( p_{k} \\) to the VD, we can exploit the fact that \\( D_{k} = \\min(D_{k-1}, d_{\\mathcal{M}}(p_{k}, V)) \\). We start a front from \\( p_{k} \\) and expand it, updating the distances and assigning vertices to texel \\( P_{k} \\) until we reach all vertices such that \\( D_{k-1}[v] < d_{\\mathcal{M}}(p_{k}, v) \\) (see Figure 2). By propagating the front with the fast marching or Dijkstra's algorithm, we can compute a geodesic VD with time \\( \\mathcal{O}(|V| \\log(|V|) \\log(s)) \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.69 + ], + "angle": 0, + "content": "To obtain the sample points, Peyré et al. [49] propose to use a geodesic farthest point sampling, achieving a geodesic uniform sampling of the surface." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.697, + 0.79, + 0.759 + ], + "angle": 0, + "content": "Definition 3. Let \\((\\mathcal{M}, d_{\\mathcal{M}})\\) be a 2-dimensional metric space, and let \\(s \\in \\mathbb{N}\\) be an integer. A farthest point sampling (FPS) of size \\(s\\) with respect to \\(d_{\\mathcal{M}}\\) is a set \\(S_s \\subset \\mathcal{M}\\) of \\(s\\) sample points in \\(\\mathcal{M}\\), incrementally built from a singleton set \\(S_1 = \\{p_1\\} \\subset \\mathcal{M}\\) according to the following rule:" + }, + { + "type": "equation", + "bbox": [ + 0.355, + 0.767, + 0.787, + 0.802 + ], + "angle": 0, + "content": "\\[\nS _ {\\ell + 1} = S _ {\\ell} \\cup \\left\\{\\underset {q \\in \\mathcal {M}} {\\arg \\max } \\quad \\min _ {p \\in S _ {\\ell}} d _ {\\mathcal {M}} (q, p) \\right\\}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "However, searching for the maximum distance in Equation (2) at every new sample brings again the algorithm to a \\(\\mathcal{O}(|V|s)\\) complexity, really becoming a" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.129 + ], + "angle": 0, + "content": "F. Maggioli et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "bottleneck for large meshes counting millions of vertices. In contrast, we propose to introduce a fixed-size binary max-heap in the front propagation algorithm, keeping track of the distances from each vertex to the sample set. During each front propagation, this heap updates the distance of each visited vertex in time \\(\\mathcal{O}(\\log(|V|))\\), adding no extra complexity to the visit. Furthermore, it can gather and set to zero the farthest vertex at each iteration in just \\(\\mathcal{O}(\\log(|V|))\\) time, totalling a time complexity of \\(\\mathcal{O}(|V|\\log(|V|)\\log(s))\\) for the entire procedure. The fast front propagation algorithm is summarized in Algorithm 1." + }, + { + "type": "code_caption", + "bbox": [ + 0.216, + 0.344, + 0.516, + 0.36 + ], + "angle": 0, + "content": "Algorithm 1 Geodesic FPS and its VD." + }, + { + "type": "algorithm", + "bbox": [ + 0.217, + 0.363, + 0.56, + 0.576 + ], + "angle": 0, + "content": "1: procedure VORONOIFPS \\((\\mathcal{M} = (V,E,T),s)\\) \n2: \\(i\\gets\\) random index in \\([1,|V|]\\) \n3: \\(S\\gets \\{i\\}\\) \n4: \\(D\\gets d_{\\mathcal{M}}(V,V_i)\\) \n5: \\(H\\gets\\) max-heap initialized with \\(D\\) \n6: \\(P\\gets\\) vector of length \\(|V|\\) initialized to \\(i\\) \n7: for \\(h\\gets 2\\) to \\(s\\) do \n8: \\(p\\gets \\mathrm{FINDMAX}(H)\\) \n9: SETKEY \\((H,p,0)\\) \n10: \\(D_p\\gets 0\\) \n11: \\(S\\gets S\\cup \\{p\\}\\) \n12: Propagate a front from \\(p\\) \n13: Update \\(D,P,\\) and \\(H\\) \n14: end for \n15: end procedure" + }, + { + "type": "text", + "bbox": [ + 0.57, + 0.283, + 0.789, + 0.585 + ], + "angle": 0, + "content": "Flat union property. Even if we are able to compute the VD of a farthest point sampling with high efficiency, we still cannot guarantee that the dual connectivity is a proper IDT. Ensuring the closed ball property can be computationally challenging. In particular, identifying the topology of the boundary between each pair of adjacent texels can become very costly when the number of samples is large. A possibility is to add enough samples so that the surface locally behaves like a plane. This approach, proposed by Leibon et al. [27] and adopted by Peyré" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.788, + 0.646 + ], + "angle": 0, + "content": "et al. [49], makes it easy to lose control over the number of samples, and anyway the requirements are not easy to enforce. For guaranteeing a proper IDT, while still avoiding these issues, we introduce a novel property for a VD. The proofs of our claims are provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.654, + 0.789, + 0.701 + ], + "angle": 0, + "content": "Definition 4 (Flat Union Property (FUP)). Let \\((\\mathcal{M}, d_{\\mathcal{M}})\\) be a 2-dimensional metric space. Let then \\(S = \\{p_i\\}_{i=1}^s \\subset \\mathcal{M}\\) be a set of \\(s\\) sample points in \\(\\mathcal{M}\\), inducing a general \\(VD\\{P_i\\}_{i=1}^s\\). The \\(VD\\) induced by \\(S\\) satisfies the FUP if:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.7, + 0.437, + 0.715 + ], + "angle": 0, + "content": "- every \\( P_i \\) is a closed 2-ball;" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.715, + 0.639, + 0.73 + ], + "angle": 0, + "content": "- if \\( P_{i} \\cap P_{j} \\) is not empty, then \\( P_{i} \\cup P_{j} \\) is a closed 2-ball;" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.73, + 0.709, + 0.746 + ], + "angle": 0, + "content": "- if \\( P_{i} \\cap P_{j} \\cap P_{k} \\) is not empty, then \\( P_{i} \\cup P_{j} \\cup P_{k} \\) is a closed 2-ball." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.7, + 0.709, + 0.746 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.755, + 0.788, + 0.787 + ], + "angle": 0, + "content": "Theorem 2. Let \\((\\mathcal{M}, d_{\\mathcal{M}})\\) be a 2-dimensional metric space. If a general \\(VD\\{P_i\\}_{i=1}^s\\) of \\(\\mathcal{M}\\) satisfies the FUP, it also satisfies the closed ball property." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "The FUP is much easier to verify than the closed ball property, as we only have to ensure that regions are closed 2-balls. Indeed, the following property can be exploited in this regard." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.432, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ReMatching: Scalable Shape Correspondence" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.178 + ], + "angle": 0, + "content": "Proposition 1. Let \\(\\mathcal{M} = (V,E,F)\\) be a manifold polygonal mesh, then \\(\\mathcal{M}\\) is a closed 2-ball if and only if its Euler characteristic \\(\\chi = |V| - |E| + |F|\\) is 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.191, + 0.784, + 0.266 + ], + "angle": 0, + "content": "Dual mesh representation. When we compute a VD of a triangle mesh, we want to partition the vertices into disjoint sets. However, Voronoi texels are defined as regions over the surface, meaning that they should be submanifolds of \\(\\mathcal{M}\\). By only considering vertices, we likely end up having non-manifold geometries or non well-defined boundaries." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.298, + 0.48, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.566, + 0.505, + 0.637 + ], + "angle": 0, + "content": "Fig. 3: Top: the primal and dual connectivities of a triangular mesh. Bottom: regions on the primal mesh are represented as sets of vertices, and the corresponding regions on the dual mesh are composed by faces." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.267, + 0.788, + 0.463 + ], + "angle": 0, + "content": "We consider the polygonal mesh \\(\\tilde{\\mathcal{M}} = (\\tilde{V},\\tilde{E},\\tilde{T})\\) to be the dual mesh of \\(\\mathcal{M} = (V,E,T)\\), built by placing a vertex at each face of \\(\\mathcal{M}\\) and forming a polygonal face for each triangle fan around a vertex of \\(\\mathcal{M}\\). As shown in the example from Figure 3, dual meshes allow us to define Voronoi texels by means of connected dual faces, making them actual submanifolds of the original surface. Furthermore, we can easily define boundaries between texels as paths made of dual edges." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.463, + 0.788, + 0.597 + ], + "angle": 0, + "content": "This construction has the positive side effect to ensure that every VD is general. Indeed, the only place where three or more texels can meet is a dual vertex. Since texels are made of dual faces (i.e., primal vertices), and the primal mesh is triangular, there is no dual vertex where more than three faces can meet." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.599, + 0.787, + 0.643 + ], + "angle": 0, + "content": "Given the correspondence between primal and dual geometric elements, we can verify if a region is a closed 2-" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.735 + ], + "angle": 0, + "content": "ball with Proposition 1 without explicitly constructing the dual mesh. Additional details are provided in the supplementary material. After the initial FPS, we further add samples to enforce the flat union property: if two or more adjacent texels do not form a closed 2-ball, we break the connection adding samples at the Voronoi vertex or along the Voronoi edge; if a texel is not a closed 2-ball, we add a sample along its boundary to reduce its coverage." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.756, + 0.467, + 0.772 + ], + "angle": 0, + "content": "3.2 Low-resolution matching" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.779, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Let \\(\\mathcal{M}\\) and \\(\\mathcal{N}\\) be two triangular meshes, and let \\(\\hat{\\mathcal{M}}\\) and \\(\\hat{\\mathcal{N}}\\) be their low-resolution counterparts computed with the algorithm presented above (or any other alternative algorithm). Furthermore, let \\(\\hat{\\Phi}\\) be a basis for the functional space over \\(\\hat{\\mathcal{M}}\\), and \\(\\hat{\\Psi}\\) a basis for the functional space over \\(\\hat{\\mathcal{N}}\\). For example these bases can" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.129 + ], + "angle": 0, + "content": "F. Maggioli et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "be the truncated subset of the eigenvectors of the LBO as proposed in [47]. We assume to have some pipeline yielding a functional map \\(\\mathbf{C}\\), such that" + }, + { + "type": "equation", + "bbox": [ + 0.445, + 0.188, + 0.786, + 0.207 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {\\Psi}} \\mathbf {C} \\approx \\hat {\\boldsymbol {\\Pi}} \\hat {\\boldsymbol {\\Phi}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.217, + 0.785, + 0.369 + ], + "angle": 0, + "content": "where \\(\\hat{\\Pi}:\\hat{\\mathcal{M}}\\to \\hat{\\mathcal{N}}\\) is a correspondence between the vertices of \\(\\hat{\\mathcal{M}}\\) and the vertices of \\(\\hat{\\mathcal{N}}\\). The adopted functional maps solution can be any of the available alternatives, ranging from the original one [47] to the most recent [12]. In our experiments, we consider two of the widely adopted solutions: constrained optimization with product preservation [46] and the iterative procedure ZoomOut [38]. The first is representative of the possible optimization strategies, while the second is probably the most efficient refinement technique for functional maps estimation. Both these methods have been efficiently and effectively applied on meshes with a limited number of vertices (i.e., up to 10 thousands vertices), which is exactly the setting we are in after the proposed remeshing step." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.392, + 0.513, + 0.407 + ], + "angle": 0, + "content": "3.3 Extending the correspondence" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.432, + 0.719 + ], + "angle": 0, + "content": "To extend the basis from the remeshing to the original surface, we need a mapping that transports scalar fields from the low-resolution mesh \\(\\hat{\\mathcal{M}}\\) to the high-resolution shape \\(\\mathcal{M}\\). A possible solution would be to map each triangle of \\(\\hat{\\mathcal{M}}\\) to a geodesic triangle in \\(\\mathcal{M}\\), and then computing the geodesic barycentric coordinates of each vertex inside the triangle [31]. Unfortunately, this approach is very costly, as it requires to compute exact geodesic paths onto \\(\\mathcal{M}\\) [56]. Instead, we approximate this mapping by projecting each vertex of \\(\\mathcal{M}\\) onto the closest surface point of \\(\\hat{\\mathcal{M}}\\). Despite" + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.447, + 0.788, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.441, + 0.647, + 0.788, + 0.717 + ], + "angle": 0, + "content": "Fig. 4: Top: remeshing and mapping of the original vertices with the solution from Liu et al. [31] and our pipeline. Bottom: comparison of ground truth geodesic distances from multiple vertices and extended with both methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.784, + 0.78 + ], + "angle": 0, + "content": "this approach not being accurate in general, every triangle in \\(\\hat{\\mathcal{M}}\\) corresponds to a geodesic triangle (homeomorphic to a disk) on \\(\\mathcal{M}\\) with the same vertices. Furthermore, starting from a farthest point sampling gives evenly spaced samples, mitigating the geometric complexity of geodesic triangles." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.841 + ], + "angle": 0, + "content": "A surface point of \\(\\hat{\\mathcal{M}}\\) is either a vertex, or a weighted average of two (if on an edge) or three vertices (if on a triangle). We build a linear map \\(\\mathbf{U}_{\\mathcal{M}} \\in \\mathbb{R}^{m \\times s}\\) representing the projection of each vertex \\(v_i \\in \\mathcal{M}\\) as a linear combination of at most three vertices in \\(\\hat{\\mathcal{M}}\\) (i.e., the vertices encoding the closest surface point)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.432, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ReMatching: Scalable Shape Correspondence" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.253 + ], + "angle": 0, + "content": "Figure 4 shows a comparison between our pipeline and the simplification method proposed by Liu et al. [31]. The top row shows the different geometries produced by the two methods in reducing the mesh size by \\(90\\%\\) and how our mapping compares to the approximate geodesic barycentric coordinates. In the bottom row we compare the two methods in approximating a multiple source geodesic distance field, showing that they do not present appreciable differences. More discussion is provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.784, + 0.312 + ], + "angle": 0, + "content": "Scalar functions can be evaluated at any surface point by interpolating the values at the vertices. Since every vertex of \\(\\mathcal{M}\\) is associated to a point on the surface of \\(\\hat{\\mathcal{M}}\\), we can use the linear map \\(\\mathbf{U}_{\\mathcal{M}}\\) for extending scalar functions from the \\(s\\) vertices of \\(\\hat{\\mathcal{M}}\\) to the \\(m\\) vertices of \\(\\mathcal{M}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.784, + 0.358 + ], + "angle": 0, + "content": "We extend the bases \\(\\hat{\\Phi}\\) and \\(\\hat{\\Psi}\\) to the full-resolution meshes as \\(\\Phi = \\mathbf{U}_{\\mathcal{M}}\\hat{\\Phi}\\) and \\(\\Psi = \\mathbf{U}_{\\mathcal{N}}\\hat{\\Psi}\\). Then, in a similar fashion as done in Equation (3), we search for a correspondence \\(\\Pi : \\mathcal{M} \\to \\mathcal{N}\\) that maps vertices of \\(\\mathcal{M}\\) to vertices of \\(\\mathcal{N}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.348, + 0.367, + 0.785, + 0.384 + ], + "angle": 0, + "content": "\\[\n\\mathbf {U} _ {\\mathcal {N}} \\hat {\\boldsymbol {\\Psi}} \\mathbf {C} = \\boldsymbol {\\Psi} \\mathbf {C} \\approx \\boldsymbol {\\Pi} \\boldsymbol {\\Phi} = \\boldsymbol {\\Pi} \\mathbf {U} _ {\\mathcal {M}} \\hat {\\boldsymbol {\\Phi}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.392, + 0.784, + 0.452 + ], + "angle": 0, + "content": "where, for the sake of clarity, we explicitly write the equation in terms of the bases \\(\\hat{\\Phi}\\) and \\(\\hat{\\Psi}\\). Given the extended bases \\(\\Phi\\) and \\(\\Psi\\), and the same functional map \\(\\mathbf{C}\\) estimated to solve Equation (3), we obtain the correspondence \\(\\Pi\\) via a nearest neighbor search in the spectral space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.453, + 0.785, + 0.543 + ], + "angle": 0, + "content": "We notice that nearest neighbor algorithms become very slow on large meshes, effectively becoming a possible bottleneck for the pipeline. Our experiments presented in Figure 10 show that, to the scope of this paper, this solution provides satisfying performance. However, we acknowledge that establishing a relationship between (3) and (4) could lead to a more efficient expression of \\(\\Pi\\) in terms of \\(\\hat{\\Pi}\\), \\(\\mathbf{U}_{\\mathcal{M}}\\), and \\(\\mathbf{U}_{\\mathcal{N}}\\). This avenue warrants future exploration." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.564, + 0.325, + 0.579 + ], + "angle": 0, + "content": "4 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.784, + 0.683 + ], + "angle": 0, + "content": "Our goal is to evaluate not only the matching accuracy of our method but also its time performance. For this task, we test it on datasets for dense correspondences, such as the SHREC19 dataset [37] and the TOSCA dataset [6]. We also introduce a novel dataset, BadTOSCA, obtained by randomly altering the vertex positions of the TOSCA meshes, to ensure that our method is stable even under circumstances where strong isometry cannot be assumed3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.684, + 0.784, + 0.791 + ], + "angle": 0, + "content": "Our pipeline relies on an efficient remeshing algorithm that can quickly and drastically reduce the size of a mesh by orders of magnitude while still preserving the underlying geometry and metric. In principle, any remeshing algorithm could be adopted in our pipeline, as the map for extending scalar fields discussed in Section 3.3 only requires having two input meshes. We compare our solution against well-established isotropic and anisotropic remeshing algorithms (respectively, IRM [4,21] and ARM [44]), as well as the scalable functional maps (SFM)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.797, + 0.784, + 0.838 + ], + "angle": 0, + "content": "3 Details on the dataset generation are provided in the supplementary materials. Code for generating the dataset is available at https://github.com/filthynobleman/bad-tosca" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.129 + ], + "angle": 0, + "content": "F. Maggioli et al." + }, + { + "type": "image", + "bbox": [ + 0.215, + 0.144, + 0.502, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.143, + 0.787, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.291, + 0.788, + 0.361 + ], + "angle": 0, + "content": "Fig. 5: Left: Accuracy curves on the SHREC19 challenge pairs for the tested methods. We also consider ideal approaches where the initial functional map for ZoomOut is given ground truth dashed lines). Right: Cumulative curves of the execution time for evaluated methods. For all methods, we also consider the time required to run the remeshing or resampling step(s)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.785, + 0.42 + ], + "angle": 0, + "content": "approach proposed by Magnet et al. [35], which produces a subsampling of the vertices in place of a low-resolution mesh." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.785, + 0.452 + ], + "angle": 0, + "content": "We implemented our remeshing algorithm in \\(\\mathbf{C}++\\), using Eigen [20] and libigl [22], and the matching pipeline in MATLAB4." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.473, + 0.428, + 0.489 + ], + "angle": 0, + "content": "4.1 Quality of matching" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.499, + 0.465, + 0.68 + ], + "angle": 0, + "content": "To evaluate our shape matching pipeline, we tested our method on the SHREC19 dataset [37]. We can use different shape correspondence techniques to test the entire pipeline. In particular, we compute the functional map on the low-resolution representation using both the widely adopted approach with products preservation (FMaps) [46]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.681, + 0.432, + 0.77 + ], + "angle": 0, + "content": "and ZoomOut [38]. We compare our method with the same approaches on the fullresolution meshes, as well as against SFM. For a fair comparison, we remesh all the" + }, + { + "type": "table", + "bbox": [ + 0.434, + 0.529, + 0.769, + 0.671 + ], + "angle": 0, + "content": "
MethodAGE (·10-2) ↓AUC ↑
Ours (ZoomOut)*2.4095.12%
SFM*2.4695.22%
Ours (FMaps)6.2388.21%
Ours (ZoomOut)5.8788.82%
SFM16.8470.19%
ARM (ZoomOut)41.5930.34%
IRM (ZoomOut)8.3084.62%
FMaps28.3156.15%
ZoomOut29.3652.02%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.441, + 0.683, + 0.788, + 0.766 + ], + "angle": 0, + "content": "Table 1: Average geodesic error and area under the accuracy curve for each tested method on the SHREC19 challenge pairs. The top rows (denoted by \\*) show the performance of our algorithm and SFM under the assumption of having a ground truth functional map for initializing ZoomOut." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.771, + 0.788, + 0.817 + ], + "angle": 0, + "content": "shapes to 3k vertices with all methods, as Magnet et al. [35] state that this produces the best performance with SFM. To further ease the alignment of the Laplacian spectra, we also post-process the remeshed shapes (produced with" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.695, + 0.841 + ], + "angle": 0, + "content": "4 Code available at https://github.com/filthynobleman/rematching" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.432, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ReMatching: Scalable Shape Correspondence" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.143, + 0.506, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.143, + 0.787, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.256, + 0.785, + 0.284 + ], + "angle": 0, + "content": "Fig. 6: Accuracy curves and average geodesic error on the TOSCA dataset (left) and the BadTOSCA dataset (right) for the tested methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.319, + 0.784, + 0.394 + ], + "angle": 0, + "content": "IRM, ARM, and our algorithm) to remove small disconnected components made of few triangles before computing the extension map discussed in Section 3.3. This post-processing step cannot be performed for SFM, due to the nature of its sampling strategy. The benefits of this post-processing will be discussed in Section 4.3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.785, + 0.503 + ], + "angle": 0, + "content": "We summarize the results of our experiment in Figure 5, where we show the accuracy curves for all methods on the SHREC19 connectivity track benchmark, following the paradigm proposed by Kim et al. [24]. To prove the time efficiency of our pipeline, we also measure the time taken by every method on each pair of shapes in the benchmark (including the remeshing step), and we show the cumulative curves of the execution times over the entire dataset in a similar way with respect to the accuracy curves." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.537, + 0.502, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.656, + 0.504, + 0.724 + ], + "angle": 0, + "content": "Fig.7: Coordinate transfer between nonisometric shapes. The coordinates are used to generate highly complex patterns, whose sensitivity to the input enhances the transfer errors." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.506, + 0.787, + 0.732 + ], + "angle": 0, + "content": "We see that all the low-resolution approaches provide better time and quality performance than the fullresolution methods, as the simpler geometry makes it easier and faster to align the Laplacian eigenfunctions. ARM is the only exception, since the anisotropic remeshing produces many skew triangles and a singular Laplacian matrix for all the shapes. Furthermore, our technique (with both the FMaps and ZoomOut backend) outperforms SFM and achieves better results than using the IRM remeshing strategy. This difference can be better" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.733, + 0.785, + 0.807 + ], + "angle": 0, + "content": "appreciated in Table 1, where we report the average geodesic error (normalized with respect to the shape diameter) and the area under the accuracy curves. In both Figure 5 and Table 1 we have also considered the idealized scenario where our method with ZoomOut as backend and SFM are initialized with a ground truth functional map. In this case, the two approaches obtain comparable results." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "We also compare our method with the other approaches on the TOSCA and BadTOSCA datasets, summarizing the results in Figure 6. In the figure, we show" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.129 + ], + "angle": 0, + "content": "F. Maggioli et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.147, + 0.492, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.259, + 0.496, + 0.328 + ], + "angle": 0, + "content": "Fig.8: Inter-class coordinate transfer from the wolf model to a cat and a dog. Our pipeline is not constrained by the standard Laplacian basis and can be used with other approaches as well." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.145, + 0.78, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.259, + 0.78, + 0.301 + ], + "angle": 0, + "content": "Fig. 9: Comparison of function transfer between very high-resolution models with SFM and our approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.364, + 0.787, + 0.456 + ], + "angle": 0, + "content": "the accuracy curves for the tested methods on both datasets, also providing the average geodesic error normalized by the shape diameter. We see that under the favorable and unrealistic conditions offered by TOSCA, where shapes of the same class are almost perfectly isometric, SFM achieves the best results. However, when we consider the slightly altered geometry of BadTOSCA, SFM suffers a performance drop of about \\(50\\%\\), while our method shows more stability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.609 + ], + "angle": 0, + "content": "A significant benefit of the functional map approach is its independence from the difference in resolution between the source and target shape. Transferring a function with functional maps yields a much smoother and neat result than directly a point-wise correspondence. In this regard, we test our method in transferring coordinate functions between isometric and non-isometric shapes with substantial differences in resolution and triangulation. In the example from Figure 7 we generated a complex procedural fractal pattern as a function of the coordinates [33]. This pattern is very sensitive to the input, so even slight errors are highly enhanced. We see that SFM cannot precisely map the coordinates of the source mesh, shifting around and distorting the details of the pattern." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.611, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Furthermore, the example in Figure 8 shows that we can also deal with inter-class function transfer with our pipeline, outperforming SFM in a visual comparison. In the example, we also showcase that we should not necessarily rely on the standard Laplacian basis as a backend. Instead of using ZoomOut to transfer functions between the low-resolution meshes, we rely on the orthogonalized eigenproducts basis introduced by Maggioli et al. [34], extending it with the same technique described in Section 3.2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Finally, to ensure that our method can scale to high-resolution meshes, we tested it using two very high-resolution 3D scans of real statues: a dancing faun counting \\(\\sim 750\\mathrm{k}\\) vertices and a warrior from the temple of Aphaea counting \\(\\sim 3.5\\mathrm{M}\\) vertices (see Figures 1 and 9). Since SFM can also deal with dense shapes, we compare it against our method. Figure 9 shows an example of the transfer of a geodesic distance function. While our method can faithfully transfer the function, SFM produces incoherent distances (e.g., wrong isolines on the head and near the shoulder) and evident artefacts (e.g., wrong distance on the hand)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.432, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ReMatching: Scalable Shape Correspondence" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.439, + 0.162 + ], + "angle": 0, + "content": "4.2 Scalable performance" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.487, + 0.447 + ], + "angle": 0, + "content": "In Section 4.1 we saw that general-purpose remeshing solutions are not an optimal choice for the matching task. However, Lescoat et al. [28] provided a specialized mesh decimation algorithm (SMD) that reduces the complexity of the mesh while still preserving the Laplacian eigenfunctions. Furthermore, Liu et al. [31] proposed an elegant solution for simplifying a shape based on intrinsic error metric (IEM), which also builds a map for extending scalar functions using geodesic barycentric coordinates. We perform the evaluation on the SHREC19 dataset [37], remeshing the surfaces to 3k vertices with all the methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.49, + 0.752 + ], + "angle": 0, + "content": "To achieve the desired scalability, the remeshing step in our pipeline must be fast enough to maintain the improvement obtained by reducing the size of the eigenproblem. In this regard, Figure 10 shows that our technique achieves better time performance than the other remeshing algorithms and the sampling strategy of SFM. In contrast, SMD and IEM are decimation algorithms that iteratively reduce the mesh size, making them orders of magnitude slower and unsuitable for large meshes. Moreover, SMD requires a pre-computation of the Laplacian eigenbasis on the original surface to guide the decimation process and preserve the spectrum, introducing additional time complexity." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.78, + 0.442, + 0.795 + ], + "angle": 0, + "content": "4.3 Benefits of remeshing" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.208, + 0.789, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.355, + 0.788, + 0.439 + ], + "angle": 0, + "content": "Fig.10: Execution time (logarithmic scale on the \\(y\\)-axis) of different remeshing algorithms plotted against the number of vertices of the input shape (\\(x\\)-axis). For SFM we show the time taken for generating the sampling." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.481, + 0.787, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.664, + 0.788, + 0.789 + ], + "angle": 0, + "content": "Fig.11: Top row: coordinate transfer between a pair from SHREC19 using ZoomOut on full-resolution meshes, SFM and our approach. Bottom row: The ground truth \\(20 \\times 20\\) functional map, compared with the alignment computed with FMaps [46] from the full Laplacian eigenbasis, the SFM extended basis and our extended basis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In Section 4.1, we discussed how we remove small disconnected components after the remeshing step. The example in Figure 11 shows a comparison of mapping" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.129 + ], + "angle": 0, + "content": "F. Maggioli et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.421 + ], + "angle": 0, + "content": "between a low-resolution mesh (Source), with less than 7k vertices, to a detailed mesh where eyes are represented with two disconnected components, each composed of 462 vertices out of the 27k of the entire mesh (\\(\\sim 2\\%\\) of the total vertices). This is enough geometry to attract energy at lower frequencies, and when we try to align the first 20 eigenfunctions with FMaps [46], the resulting functional map turns out to be mostly noise. Initializing ZoomOut with such a map leads to meaningless correspondence, and SFM does not mitigate this issue, as it tries to preserve the original spectrum as much as possible. In contrast, by removing these components and mapping them to their closest surface points, we introduce a small error in the final correspondence, but at the same time, we ease the alignment of the initial eigenbases, resulting in more accurate and meaningful mapping. This is evident in the bottom row of Figure 11, where the ground truth functional map presents a gap in the alignment of the eigenfunctions due to the impossibility of mapping the disconnected eyes of the target shape in any point of the source shape. Using the real eigenfunctions or a close approximation yields an intense noise in the map, producing the meaningless mapping of ZoomOut and SFM. Instead, by mapping the eyes to their nearest point on the head, we can meaningfully align our extended basis." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.471, + 0.37, + 0.487 + ], + "angle": 0, + "content": "5 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.788, + 0.716 + ], + "angle": 0, + "content": "We presented a new pipeline for scalable shape matching that relies on a low-resolution representation to compute a functional map between dense shapes efficiently. A core part of this pipeline is our novel remeshing algorithm, which efficiently computes an intrinsic Delaunay triangulation of uniform surface sampling. This procedure is applied to reduce the computational cost of the functional maps framework and make it efficient also on very dense meshes. Furthermore, we use a fast and intuitive technique for extending scalar functions from low-resolution meshes to their dense counterparts, extending the computed correspondence to the original very high-resolution shapes. Our experimental evaluation proved that our procedure is effective on various types of shapes at different resolutions, outperforming other state-of-the-art solutions and solving topological issues related to the alignment of the Laplacian eigenbasis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "While using a triangulated mesh as a low-resolution representation could, in principle, be exploited with other types of shape-matching pipelines, we only tested it within the functional maps framework, using the Laplace-Beltrami eigenbasis or bases derived from it. While the current pipeline binds the method's applicability to the setting of functional correspondences, we believe that studying its behaviour with different \"backends\", as well as exploring the relationships between the low-resolution and the high-resolution correspondences, would be an interesting matter for future investigations." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.432, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ReMatching: Scalable Shape Correspondence" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.326 + ], + "angle": 0, + "content": "This work was partially supported by the Italian Ministry of Education, Universities and Research under the grant Dipartimento di Eccellenza 2023-2027 of the Department of Informatics, Systems and Communication of the University of Milano-Bicocca and by the PRIN 2022 project GEOPRIDE Geometric primitive fitting on 3D data for geometric analysis and 3D shapes. This work is also supported by the ERC grant no.802554 SPECGEO and PRIN 2020 project no.2020TA3K9N LEGO.AI. We acknowledge the support of NVIDIA Corporation with the RTX A5000 GPUs granted through the Academic Hardware Grant Program to the University of Milano-Bicocca for the project Learned representations for implicit binary operations on real-world 2D-3D data." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.348, + 0.323, + 0.364 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.377, + 0.786, + 0.417 + ], + "angle": 0, + "content": "1. Aflalo, Y., Brezis, H., Kimmel, R.: On the optimality of shape and data representation in the spectral domain. SIAM Journal on Imaging Sciences 8(2), 1141-1160 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.418, + 0.786, + 0.459 + ], + "angle": 0, + "content": "2. Amenta, N., Choi, S., Dey, T.K., Leekha, N.: A simple algorithm for homeomorphic surface reconstruction. In: Proceedings of the sixteenth annual symposium on Computational geometry. pp. 213-222 (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.459, + 0.786, + 0.486 + ], + "angle": 0, + "content": "3. Bobenko, A.I., Springborn, B.A.: A discrete laplace-beltrami operator for simplicial surfaces. Discrete & Computational Geometry 38(4), 740-756 (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.486, + 0.786, + 0.513 + ], + "angle": 0, + "content": "4. Botsch, M., Pauly, M., Rossl, C., Bischoff, S., Kobbelt, L.: Geometric modeling based on triangle meshes. In: ACM SIGGRAPH 2006 Courses, pp. 1-es (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.513, + 0.786, + 0.54 + ], + "angle": 0, + "content": "5. Bowyer, A.: Computing Dirichlet tessellations*. The Computer Journal 24(2), 162-166 (01 1981)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.54, + 0.786, + 0.568 + ], + "angle": 0, + "content": "6. Bronstein, A., Bronstein, M., Kimmel, R.: Numerical Geometry of Non-Rigid Shapes. Springer, New York, NY (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.568, + 0.786, + 0.595 + ], + "angle": 0, + "content": "7. Burghard, O., Dieckmann, A., Klein, R.: Embedding shapes with green's functions for global shape matching. Computers & Graphics 68, 1-10 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.595, + 0.786, + 0.636 + ], + "angle": 0, + "content": "8. Crane, K., Weischedel, C., Wardetzky, M.: Geodesics in heat: A new approach to computing distance based on heat flow. ACM Transactions on Graphics (TOG) 32(5), 1-11 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.636, + 0.786, + 0.663 + ], + "angle": 0, + "content": "9. Delaunay, B., et al.: Sur la sphere vide. Izv. Akad. Nauk SSSR, Otdelenie Matematicheskii i Estestvennyka Nauk 7(793-800), 1-2 (1934)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.663, + 0.786, + 0.69 + ], + "angle": 0, + "content": "0. Deng, B., Yao, Y., Dyke, R.M., Zhang, J.: A survey of non-rigid 3d registration. In: Computer Graphics Forum. vol. 41, pp. 559-589. Wiley Online Library (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.69, + 0.786, + 0.718 + ], + "angle": 0, + "content": "1. Dijkstra, E.W.: A note on two problems in connexion with graphs. Numerische Mathematik 1, 269-271 (1959)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.786, + 0.758 + ], + "angle": 0, + "content": "2. Donati, N., Corman, E., Melzi, S., Ovsjanikov, M.: Complex functional maps: A conformal link between tangent bundles. In: Computer Graphics Forum. vol. 41, pp. 317-334. Wiley Online Library (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.786, + 0.8 + ], + "angle": 0, + "content": "3. Donati, N., Sharma, A., Ovsjanikov, M.: Deep geometric functional maps: Robust feature learning for shape correspondence. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8592-8601 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.786, + 0.84 + ], + "angle": 0, + "content": "4. Dyer, R., Zhang, H., Möller, T.: Voronoi-delaunay duality and delaunay meshes. In: Proceedings of the 2007 ACM symposium on Solid and physical modeling. pp. 415-420 (2007)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.377, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.129 + ], + "angle": 0, + "content": "F. Maggioli et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.175 + ], + "angle": 0, + "content": "15. Edelsbrunner, H., Shah, N.R.: Triangulating topological spaces. In: Proceedings of the tenth annual symposium on Computational geometry. pp. 285-292 (1994)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.787, + 0.215 + ], + "angle": 0, + "content": "16. Eynard, D., Rodola, E., Glashoff, K., Bronstein, M.M.: Coupled functional maps. In: 2016 Fourth International Conference on 3D Vision (3DV). pp. 399-407. IEEE (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.216, + 0.787, + 0.242 + ], + "angle": 0, + "content": "17. Ezuz, D., Ben-Chen, M.: Deblurring and denoising of maps between shapes. In: Computer Graphics Forum. vol. 36, pp. 165-174. Wiley Online Library (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.243, + 0.787, + 0.297 + ], + "angle": 0, + "content": "18. Fei, Y., Chen, S., Su, D., Luo, J., Li, M.: A new algorithm for repairing non-manifold surfaces. In: 2013 IEEE 10th International Conference on High Performance Computing and Communications & 2013 IEEE International Conference on Embedded and Ubiquitous Computing. pp. 1704-1708. IEEE (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.298, + 0.787, + 0.352 + ], + "angle": 0, + "content": "19. Gao, M., Roetzer, P., Eisenberger, M., Lähner, Z., Moeller, M., Cremers, D., Bernard, F.: Sigma: Scale-invariant global sparse shape matching. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 645-654 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.353, + 0.765, + 0.366 + ], + "angle": 0, + "content": "20. Guennebaud, G., Jacob, B., et al.: Eigen v3. http://eigen.tuxfamily.org (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.366, + 0.787, + 0.406 + ], + "angle": 0, + "content": "21. Hoppe, H., DeRose, T., Duchamp, T., McDonald, J., Stuetzle, W.: Mesh optimization. In: Proceedings of the 20th annual conference on Computer graphics and interactive techniques. pp. 19-26 (1993)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.407, + 0.787, + 0.433 + ], + "angle": 0, + "content": "22. Jacobson, A., Panozzo, D., et al.: libigl: A simple \\(\\mathrm{C}++\\) geometry processing library (2018), https://libigl.github.io/" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.787, + 0.46 + ], + "angle": 0, + "content": "23. Jiang, Z., Schneider, T., Zorin, D., Panozzo, D.: Bijective projection in a shell. ACM Transactions on Graphics (TOG) 39(6), 1-18 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.787, + 0.487 + ], + "angle": 0, + "content": "24. Kim, V.G., Lipman, Y., Funkhouser, T.: Blended intrinsic maps. ACM transactions on graphics (TOG) 30(4), 1-12 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.514 + ], + "angle": 0, + "content": "25. Kimmel, R., Sethian, J.A.: Computing geodesic paths on manifolds. Proceedings of the national academy of Sciences 95(15), 8431-8435 (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.787, + 0.554 + ], + "angle": 0, + "content": "26. Kressner, D.: Numerical Methods for General and Structured Eigenvalue Problems. Lecture Notes in Computational Science and Engineering, Springer Berlin Heidelberg (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.787, + 0.596 + ], + "angle": 0, + "content": "27. Leibon, G., Letscher, D.: Delaunay triangulations and voronoi diagrams for riemannian manifolds. In: Proceedings of the sixteenth annual symposium on Computational geometry. pp. 341-349 (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.597, + 0.787, + 0.637 + ], + "angle": 0, + "content": "28. Lescoat, T., Liu, H.T.D., Thiery, J.M., Jacobson, A., Boubekeur, T., Ovsjanikov, M.: Spectral mesh simplification. In: Computer Graphics Forum. vol. 39, pp. 315-324. Wiley Online Library (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.787, + 0.677 + ], + "angle": 0, + "content": "29. Levy, B.: Laplace-beltrami eigenfunctions towards an algorithm that \"understands\" geometry. In: IEEE International Conference on Shape Modeling and Applications 2006 (SMI'06). pp. 13-13 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.787, + 0.704 + ], + "angle": 0, + "content": "30. Lévy, B., Zhang, H.: Spectral mesh processing. In: ACM SIGGRAPH 2010 Courses, pp. 1-312 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.787, + 0.745 + ], + "angle": 0, + "content": "31. Liu, H.T.D., Gillespie, M., Chislett, B., Sharp, N., Jacobson, A., Crane, K.: Surface simplification using intrinsic error metrics. ACM Trans. Graph. 42(4) (jul 2023). https://doi.org/10.1145/3592403, https://doi.org/10.1145/3592403" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.746, + 0.787, + 0.772 + ], + "angle": 0, + "content": "32. Liu, H.T.D., Jacobson, A., Ovsjanikov, M.: Spectral coarsening of geometric operators. ACM Transactions on Graphics (TOG) 38(4) (jul 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.773, + 0.787, + 0.799 + ], + "angle": 0, + "content": "33. Maggioli, F., Baieri, D., Melzi, S., Rodolà, E.: Newton's fractals on surfaces via bicomplex algebra. In: ACM SIGGRAPH 2022 Posters, pp. 1-2 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "34. Maggioli, F., Melzi, S., Ovsjanikov, M., Bronstein, M.M., Rodola, E.: Orthogonalized fourier polynomials for signal approximation and transfer. In: Computer Graphics Forum. vol. 40, pp. 435-447. Wiley Online Library (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.432, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ReMatching: Scalable Shape Correspondence" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "35. Magnet, R., Ovsjanikov, M.: Scalable and efficient functional map computations on dense meshes. In: Computer Graphics Forum. vol. 42, pp. 89-101. Wiley Online Library (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.787, + 0.217 + ], + "angle": 0, + "content": "36. Marin, R., Corona, E., Pons-Moll, G.: Nicp: Neural icip for 3d human registration at scale. In: European Conference on Computer Vision. Springer (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.217, + 0.787, + 0.271 + ], + "angle": 0, + "content": "37. Melzi, S., Marin, R., Rodolà, E., Castellani, U., Ren, J., Poulenard, A., Wonka, P., Ovsjanikov, M.: Shrec 2019: Matching humans with different connectivity. In: Eurographics Workshop on 3D Object Retrieval. vol. 7, p. 3. The Eurographics Association (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.271, + 0.787, + 0.312 + ], + "angle": 0, + "content": "38. Melzi, S., Ren, J., Rodola, E., Sharma, A., Wonka, P., Ovsjanikov, M.: Zoomout: spectral upsampling for efficient shape correspondence. ACM Transactions on Graphics (TOG) 38(6), 1-14 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.312, + 0.787, + 0.353 + ], + "angle": 0, + "content": "39. Melzi, S., Rodolà, E., Castellani, U., Bronstein, M.M.: Localized manifold harmonics for spectral shape analysis. In: Computer Graphics Forum. vol. 37, pp. 20-34. Wiley Online Library (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.353, + 0.787, + 0.393 + ], + "angle": 0, + "content": "40. Meyer, M., Desbrun, M., Schröder, P., Barr, A.H.: Discrete Differential-Geometry Operators for Triangulated 2-Manifolds. In: Visualization and mathematics III, pp. 35-57. Springer, New York, NY (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.393, + 0.787, + 0.42 + ], + "angle": 0, + "content": "41. Mitchell, J.S., Mount, D.M., Papadimitriou, C.H.: The discrete geodesic problem. SIAM Journal on Computing 16(4), 647-668 (1987)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.42, + 0.787, + 0.461 + ], + "angle": 0, + "content": "42. Nasikun, A., Brandt, C., Hildebrandt, K.: Fast approximation of laplace-beltrami eigenproblems. In: Computer Graphics Forum. vol. 37, pp. 121-134. Wiley Online Library (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.461, + 0.787, + 0.5 + ], + "angle": 0, + "content": "43. Nasikun, A., Hildebrandt, K.: The hierarchical subspace iteration method for laplace-beltrami eigenproblems. ACM Transactions on Graphics (TOG) 41(2), 1-14 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.501, + 0.787, + 0.542 + ], + "angle": 0, + "content": "44. Nivoliers, V., Lévy, B., Geuzaine, C.: Anisotropic and feature sensitive triangular remeshing using normal lifting. Journal of Computational and Applied Mathematics 289, 225-240 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.542, + 0.787, + 0.583 + ], + "angle": 0, + "content": "45. Nogneng, D., Melzi, S., Rodola, E., Castellani, U., Bronstein, M., Ovsjanikov, M.: Improved functional mappings via product preservation. In: Computer Graphics Forum. vol. 37, pp. 179-190. Wiley Online Library (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.583, + 0.787, + 0.624 + ], + "angle": 0, + "content": "46. Nogneng, D., Ovsjanikov, M.: Informative descriptor preservation via commutativity for shape matching. In: Computer Graphics Forum. vol. 36, pp. 259-267. Wiley Online Library (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.624, + 0.787, + 0.665 + ], + "angle": 0, + "content": "47. Ovsjanikov, M., Ben-Chen, M., Solomon, J., Butscher, A., Guibas, L.: Functional maps: a flexible representation of maps between shapes. ACM Transactions on Graphics (ToG) 31(4), 1-11 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.664, + 0.787, + 0.705 + ], + "angle": 0, + "content": "48. Ovsjanikov, M., Corman, E., Bronstein, M., Rodola, E., Ben-Chen, M., Guibas, L., Chazal, F., Bronstein, A.: Computing and processing correspondences with functional maps. In: SIGGRAPH ASIA 2016 Courses, pp. 1-60 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.732 + ], + "angle": 0, + "content": "49. Peyré, G., Cohen, L.D.: Geodesic remeshing using front propagation. International Journal of Computer Vision 69, 145-156 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.732, + 0.787, + 0.759 + ], + "angle": 0, + "content": "50. Pinkall, U., Polthier, K.: Computing Discrete Minimal Surfaces and their Conjugates. Experimental mathematics 2(1), 15-36 (1993)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.759, + 0.787, + 0.8 + ], + "angle": 0, + "content": "51. Ren, J., Melzi, S., Ovsjanikov, M., Wonka, P., et al.: Maptree: recovering multiple solutions in the space of maps. ACM Transactions on Graphics (TOG) 39(6), 264-1 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.8, + 0.787, + 0.841 + ], + "angle": 0, + "content": "52. Ren, J., Melzi, S., Wonka, P., Ovsjanikov, M.: Discrete optimization for shape matching. In: Computer Graphics Forum. vol. 40, pp. 81-96. Wiley Online Library (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.129 + ], + "angle": 0, + "content": "F. Maggioli et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "53. Ren, J., Poulenard, A., Wonka, P., Ovsjanikov, M.: Continuous and orientation-preserving correspondences via functional maps. ACM Transactions on Graphics (ToG) 37(6), 1-16 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.217 + ], + "angle": 0, + "content": "54. Reuter, M., Wolter, F.E., Peinecke, N.: Laplace-beltrami spectra as 'shape-dna' of surfaces and solids. Computer-Aided Design 38(4), 342-366 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "55. Rivin, I.: Euclidean structures on simplicial surfaces and hyperbolic volume. Annals of mathematics 139(3), 553-580 (1994)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.273 + ], + "angle": 0, + "content": "56. Rustamov, R.M.: Barycentric coordinates on surfaces. In: Computer Graphics Forum. vol. 29, pp. 1507-1516. Wiley Online Library (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "57. Sahillioğlu, Y.: Recent advances in shape correspondence. The Visual Computer 36(8), 1705-1721 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "58. Shoham, M., Vaxman, A., Ben-Chen, M.: Hierarchical functional maps between subdivision surfaces. In: Computer Graphics Forum. vol. 38, pp. 55-73. Wiley Online Library (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "59. Vaxman, A., Ben-Chen, M., Gotsman, C.: A multi-resolution approach to heat kernels on discrete surfaces. In: ACM SIGGRAPH 2010 papers, pp. 1-10 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.397 + ], + "angle": 0, + "content": "60. Watson, D.F.: Computing the n-dimensional Delaunay tessellation with application to Voronoi polytopes*. The Computer Journal 24(2), 167-172 (01 1981)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.397 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/ReMatching_ Low-Resolution Representations for Scalable Shape Correspondence/8dd15a7c-48bf-42e4-a752-0b0dbad83c91_origin.pdf b/2024/ReMatching_ Low-Resolution Representations for Scalable Shape Correspondence/8dd15a7c-48bf-42e4-a752-0b0dbad83c91_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b762b79f680c72256081eebff8072737471909ff --- /dev/null +++ b/2024/ReMatching_ Low-Resolution Representations for Scalable Shape Correspondence/8dd15a7c-48bf-42e4-a752-0b0dbad83c91_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d1df02c6d1be27ce0880207281726e7e7798ac56f60fe66ed28e42279537d0a3 +size 10194552 diff --git a/2024/ReMatching_ Low-Resolution Representations for Scalable Shape Correspondence/full.md b/2024/ReMatching_ Low-Resolution Representations for Scalable Shape Correspondence/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fab0377dac514d65b8d6ad07fc31f2140883081b --- /dev/null +++ b/2024/ReMatching_ Low-Resolution Representations for Scalable Shape Correspondence/full.md @@ -0,0 +1,332 @@ +# ReMatching: Low-Resolution Representations for Scalable Shape Correspondence + +Filippo Maggioli1, Daniele Baieri2, Emanuele Rodolà2, and Simone Melzi1 + +1 University of Milano-Bicocca, Milan, Italy +2 Sapienza - University of Rome, Rome, Italy + +![](images/054bd1a6d36e11904940f38384a53e57a94f58ecc74a27bbc669d4eb29a9846a.jpg) +Fig. 1: Correspondence between two dense shapes computed with our method, using ZoomOut [38] on low-resolution meshes. The coordinates functions of the dancing faun statue on the left ( $\sim750\mathrm{k}$ vertices) are transferred using a functional map to the Aphaea warrior statue on the right ( $\sim3.5\mathrm{M}$ vertices). The colors encode the correspondence. Despite the mesh density, shown in the close-up, the computation took $\sim2$ minutes. + +Abstract. We introduce ReMatching, a novel shape correspondence solution based on the functional maps framework. Our method, by exploiting a new and appropriate re-meshing paradigm, can target shape-matching tasks even on meshes counting millions of vertices, where the original functional maps does not apply or requires a massive computational cost. The core of our procedure is a time-efficient remeshing algorithm which constructs a low-resolution geometry while acting conservatively on the original topology and metric. These properties allow translating the functional maps optimization problem on the resulting low-resolution representation, thus enabling efficient computation of correspondences with functional map approaches. Finally, we propose an efficient technique for extending the estimated correspondence to the original meshes. We show that our method is more efficient and effective through quantitative and qualitative comparisons, outperforming state-of-the-art pipelines in quality and computational cost. + +# 1 Introduction and related work + +The task of finding a semantically meaningful correspondence between discrete surfaces has always been a fundamental topic in the field of shape analysis. Researchers developed a wealth of solutions for this application, and new methods continue to be designed [10, 57]. Among the various approaches, the functional maps framework [47, 48] received significant attention. Rather than finding a point-wise correspondence, the functional maps framework aims to define a correspondence between functions, encoding it into a small matrix. Following this direction, a variety of works tried to improve the actual computation of the map [13, 38, 53] or to extend the framework to different types of bases [34, 39, 45]. + +Despite the amount of research outcomes based on the functional maps framework, both exploiting geometric properties and tools [7, 16, 17] and clever optimization techniques [46, 51, 52], the problem of computing the map is still bounded by the time complexity of sparse eigendecompositions [26]. To overcome this problem, researchers started investigating scalable solutions to the Laplace-Beltrami eigenproblem. In this regard, multi-resolution techniques [42, 43, 58, 59] proved to be very effective for scalable computation of spectral quantities for tasks such as retrieval and mesh filtering [30, 54], but their applicability to the functional map framework proved to require additional care [35]. Furthermore, recent research proposed spectral coarsening methods [32] or spectral preserving simplification techniques [28]. However, these techniques usually rely on the computation of the full-resolution spectrum, making them unsuitable for large-scale applications. Gao et al. [19] proposed a mixed-integer programming scalable solution to the matching problem, but they specifically target sparse correspondences only. Alternatively, Marin et al. [36], who provide an approach for human registration that scales across large datasets, but it is unsuitable for dense shapes. + +Closely related to our work is the contribution from Magnet et al. [35], which, for the first time, proposes a solution for scaling the functional maps approach to high-resolution meshes. Their idea is to reduce the dimensionality of the eigenproblem, inducing a linear relationship between the functional space on the mesh and a lower-dimension functional space on a sparse sample. The extreme sparsity of the sampling, combined with its efficient computation, leads to an algorithm that can effectively deal with high-density meshes. Nevertheless, this type of reduction does not mitigate the negative influence that small components and local details at high frequency could have on the alignment of the spectra. + +Finally, our pipeline relies on building robust and sparse representations of dense meshes while still keeping an approximately bijective correspondence between the high- and low-resolution shapes, and it is worth mentioning that other works have already attempted to do so. Jiang et al. [23] propose building a lower resolution prismatic shell preserving bijectivity with the original surface. However, their algorithm does not scale to meshes with a high triangle count, gives no control and no guarantees on the final vertex count, and requires many assumption. + +tions on the input shape (e.g., manifoldness, orientability, no self-intersections). On the other hand, Liu et al. [31] propose to exploit an intrinsic error metric for decimating a mesh to a given size while keeping track of a geodesic barycentric mapping of each triangle of the resulting shape to a geodesic triangle on the input surface. Despite the robustness of the method, the simplification approach makes it unsuitable for a functional maps setting, where it is required to decimate shapes from millions of vertices to a few thousand. This limitation is further discussed in Section 4.2. + +We introduce a new scalable functional map pipeline that efficiently handles meshes with high vertex density (see Figure 1), yields stable results, and is not bounded by the quality of the original triangulation. To summarize: + +- we translate the matching pipeline to low-resolution representations, enabling a fast and scalable computation of functional maps between dense shapes, providing the first alternative to [35]; +- we propose a new efficient and geometry-preserving remeshing algorithm specifically designed for our pipeline; +- we exploit a fast solution for extending scalar maps from the low-resolution remeshed shape to the original surface, thus efficiently obtaining suitable bases for function transfer and shape matching. + +# 2 Background + +In the discrete setting, we represent a shape as a triangular mesh $\mathcal{M} = (V,E,T)$ where (i) $V\subset \mathbb{R}^3$ is a set of vertices; (ii) $E\subset V^2$ is a set of edges between vertices; (iii) $T\subset V^3$ is a set of triangles composing the surface. + +# 2.1 Functional maps + +We discretize a scalar function $f: \mathcal{M} \to \mathbb{R}$ as a signal defined on the vertices $V$ and represented as a vector $\pmb{f} \in \mathbb{R}^{|V|}$ . Hence, the Laplace-Beltrami operator $\varDelta_{\mathcal{M}}: \mathcal{F}(\mathcal{M}) \to \mathcal{F}(\mathcal{M})$ is discretized as a sparse matrix $\mathbf{L}_{\mathcal{M}} \in \mathbb{R}^{|V| \times |V|}$ , generally represented by means of a stiffness matrix $\mathbf{S}_{\mathcal{M}}$ and a mass matrix $\mathbf{A}_{\mathcal{M}}$ [40, 50]. The eigendecomposition $\mathbf{S}_{\mathcal{M}} \pmb{\Phi}_{\mathcal{M}} = \mathbf{A}_{\mathcal{M}} \pmb{\Phi}_{\mathcal{M}} \varLambda_{\mathcal{M}}$ of the Laplacian yields an orthonormal basis for the functional space on $\mathcal{M}$ [29,30] which has analogies with the Fourier basis and optimally represents smooth functions when truncated [1]. + +Given two shapes $\mathcal{M}$ and $\mathcal{N}$ , respectively with $m$ and $n$ vertices, a point-to-point map $\Pi : \mathcal{M} \to \mathcal{N}$ can be expressed as a binary matrix $\Pi \in \{0,1\}^{m \times n}$ such that $\Pi(i,j) = 1$ if the vertex $j$ -th of $\mathcal{N}$ corresponds to the vertex $i$ -th of $\mathcal{M}$ , and $\Pi(i,j) = 0$ otherwise. Any point-wise correspondence $\Pi$ establishes a functional map $T_{\Pi} : \mathcal{F}(\mathcal{N}) \to \mathcal{F}(\mathcal{M})$ as $T_{\Pi}(f) = f \circ \Pi$ , $\forall f \in \mathcal{F}(\mathcal{N})$ . + +As proposed in [47], given $\Phi_k$ and $\Psi_k$ the Laplacian bases truncated to size $k$ respectively on $\mathcal{M}$ and $\mathcal{N}$ , we can project $T_{\Pi}$ in these bases and compactly encode the functional map in a $k \times k$ matrix $\mathbf{C} = \Phi_k^\top \mathbf{A}_{\mathcal{M}} \Pi \Psi_k$ , which is the linear operator that transfers the coefficients of functions from $\mathcal{F}(\mathcal{N})$ in the ones of corresponding functions of $\mathcal{F}(\mathcal{M})$ respectively computed with $\Phi_k$ and $\Psi_k$ . + +Finally, given a functional map $T_{\Pi}$ , or its compact representation $\mathbf{C}$ , we can retrieve the correspondence $\Pi$ on which the map is built by the nearest neighbor search in the embedding space of the Laplacian eigenfunctions as detailed in [48]. + +# 2.2 Intrinsic Delaunay triangulations (IDT) + +The IDT [3, 55] has been introduced for generalizing the Delaunay triangulation [9] to non-Euclidean metric spaces where classical algorithms [5, 60] cannot be used. IDT relies on the duality with the Voronoi diagram [15, 27], as three texels meeting at a point form a triangle whose edges cross the texels boundaries. + +Definition 1. Let $(\mathcal{M},d_{\mathcal{M}})$ be a 2-dimensional metric space (with distance function $d_{\mathcal{M}}$ ), and let $S = \{p_i\}_{i = 1}^s\subset \mathcal{M}$ be a set of $s$ sample points. The Voronoi decomposition (VD) of $\mathcal{M}$ with respect to $S$ is the collection $\{P_i\}_{i = 1}^s$ such that + +$$ +P _ {i} = \left\{q \in \mathcal {M}: p _ {i} = \underset {p _ {j} \in S} {\arg \min } \left(d _ {\mathcal {M}} \left(p _ {j}, q\right)\right) \right\}, \quad \bigcup_ {i = 1} ^ {s} P _ {i} = \mathcal {M} \tag {1} +$$ + +The points in $S$ are called generators, and the $P_{i}$ are called texels (or cells). + +The intersection between two or more texels could be non-empty. In such a case, we define those texels as adjacent. A VD is general if the intersection of four or more texels is empty. If a VD is general, a Voronoi vertex is a point that belongs to three texels, and a Voronoi edge is a curve $C$ connecting two Voronoi vertices and belonging to two texels. + +It can be shown that there is a one-to-one correspondence between a Voronoi edge dividing $P_{i}$ and $P_{j}$ and a Delaunay edge connecting $p_{i}$ to $p_{j}$ [14]. This correspondence gives rise to a necessary and sufficient condition for the IDT to be a proper triangulation (i.e., it realizes a simplicial complex). + +Definition 2 ([2,15]). If $S$ induces a general $VD$ , the $VD$ satisfies the closed ball property if: + +- every $P_{i}$ is a closed 2-ball (i.e., a topological disk without holes); +- every $P_i \cap P_j$ is either empty or a closed 1-ball (i.e., a topological segment); +- every $P_i \cap P_j \cap P_k$ is either empty or a closed 0-ball (i.e., a point). + +Theorem 1 ([2,14,15]). If $S$ induces a general $VD$ , then the $VD$ satisfies the closed ball property if and only if its dual $IDT$ is a proper triangulation. + +# 3 Method + +We introduce a new scalable functional map pipeline handling very dense meshes and yielding stable results independently of the input triangulation's quality. + +Our method acts through three main steps. At first, we compute two low-resolution meshes that preserve the original metrics of the input shapes (Section 3.1). Then, we apply any existing pipeline, yielding a functional map between the low-resolution eigenspaces (Section 3.2). Finally, we extend the eigenfunctions to the original meshes, allowing us to use them with the functional map + +estimated in the previous step to compute the desired correspondence between them (Section 3.3). Our technique is very general and works with arbitrary manifold triangulations, including meshes with degenerate geometry, non-orientable surfaces, and multiple connected components. The manifoldness requirement does not provide an obstacle to the applicability of our algorithm, as it can be efficiently enforced using existing techniques [18]. + +# 3.1 Intrinsic Delaunay remeshing + +Front propagation and FPS. Naively computing the VD induced by a set of samples can quickly become very expensive. The exact geodesic algorithm introduced by Mitchell et al. [41] is computationally unfeasible on large meshes. Even using faster solutions, such as the heat method [8], the fast marching algorithm [25], or even the Dijkstra distance [11], searching the closest sample for each vertex is still a $\mathcal{O}(|V|s)$ operation. Instead, we take inspiration from the front propagation method proposed by Peyre et al. [49]. We start with a decomposition of the mesh induced by a single sample $p_1$ . + +![](images/88eb061469dddfd03768f683ce6015dee27a2fb20c18ea42b41575bc6d24603c.jpg) +Fig.2: The iterative front propagation from samples inducing a geodesic VD. + +This means that every vertex $v$ is part of the texel $P_{1}$ and has distance $D_{1}[v] = d_{\mathcal{M}}(p_{1}, v)$ from the sample set. We then assume to have a decomposition $P_{1}, \dots, P_{k-1}$ induced by samples $p_{1}, \dots, p_{k-1}$ , and a vector $D_{k-1}$ storing the distance of each vertex from the sample set. When adding a sample $p_{k}$ to the VD, we can exploit the fact that $D_{k} = \min(D_{k-1}, d_{\mathcal{M}}(p_{k}, V))$ . We start a front from $p_{k}$ and expand it, updating the distances and assigning vertices to texel $P_{k}$ until we reach all vertices such that $D_{k-1}[v] < d_{\mathcal{M}}(p_{k}, v)$ (see Figure 2). By propagating the front with the fast marching or Dijkstra's algorithm, we can compute a geodesic VD with time $\mathcal{O}(|V| \log(|V|) \log(s))$ . + +To obtain the sample points, Peyré et al. [49] propose to use a geodesic farthest point sampling, achieving a geodesic uniform sampling of the surface. + +Definition 3. Let $(\mathcal{M}, d_{\mathcal{M}})$ be a 2-dimensional metric space, and let $s \in \mathbb{N}$ be an integer. A farthest point sampling (FPS) of size $s$ with respect to $d_{\mathcal{M}}$ is a set $S_s \subset \mathcal{M}$ of $s$ sample points in $\mathcal{M}$ , incrementally built from a singleton set $S_1 = \{p_1\} \subset \mathcal{M}$ according to the following rule: + +$$ +S _ {\ell + 1} = S _ {\ell} \cup \left\{\underset {q \in \mathcal {M}} {\arg \max } \quad \min _ {p \in S _ {\ell}} d _ {\mathcal {M}} (q, p) \right\}. \tag {2} +$$ + +However, searching for the maximum distance in Equation (2) at every new sample brings again the algorithm to a $\mathcal{O}(|V|s)$ complexity, really becoming a + +bottleneck for large meshes counting millions of vertices. In contrast, we propose to introduce a fixed-size binary max-heap in the front propagation algorithm, keeping track of the distances from each vertex to the sample set. During each front propagation, this heap updates the distance of each visited vertex in time $\mathcal{O}(\log(|V|))$ , adding no extra complexity to the visit. Furthermore, it can gather and set to zero the farthest vertex at each iteration in just $\mathcal{O}(\log(|V|))$ time, totalling a time complexity of $\mathcal{O}(|V|\log(|V|)\log(s))$ for the entire procedure. The fast front propagation algorithm is summarized in Algorithm 1. + +Algorithm 1 Geodesic FPS and its VD. +1: procedure VORONOIFPS $(\mathcal{M} = (V,E,T),s)$ +2: $i\gets$ random index in $[1,|V|]$ +3: $S\gets \{i\}$ +4: $D\gets d_{\mathcal{M}}(V,V_i)$ +5: $H\gets$ max-heap initialized with $D$ +6: $P\gets$ vector of length $|V|$ initialized to $i$ +7: for $h\gets 2$ to $s$ do +8: $p\gets \mathrm{FINDMAX}(H)$ +9: SETKEY $(H,p,0)$ +10: $D_p\gets 0$ +11: $S\gets S\cup \{p\}$ +12: Propagate a front from $p$ +13: Update $D,P,$ and $H$ +14: end for +15: end procedure + +Flat union property. Even if we are able to compute the VD of a farthest point sampling with high efficiency, we still cannot guarantee that the dual connectivity is a proper IDT. Ensuring the closed ball property can be computationally challenging. In particular, identifying the topology of the boundary between each pair of adjacent texels can become very costly when the number of samples is large. A possibility is to add enough samples so that the surface locally behaves like a plane. This approach, proposed by Leibon et al. [27] and adopted by Peyré + +et al. [49], makes it easy to lose control over the number of samples, and anyway the requirements are not easy to enforce. For guaranteeing a proper IDT, while still avoiding these issues, we introduce a novel property for a VD. The proofs of our claims are provided in the supplementary materials. + +Definition 4 (Flat Union Property (FUP)). Let $(\mathcal{M}, d_{\mathcal{M}})$ be a 2-dimensional metric space. Let then $S = \{p_i\}_{i=1}^s \subset \mathcal{M}$ be a set of $s$ sample points in $\mathcal{M}$ , inducing a general $VD\{P_i\}_{i=1}^s$ . The $VD$ induced by $S$ satisfies the FUP if: + +- every $P_i$ is a closed 2-ball; +- if $P_{i} \cap P_{j}$ is not empty, then $P_{i} \cup P_{j}$ is a closed 2-ball; +- if $P_{i} \cap P_{j} \cap P_{k}$ is not empty, then $P_{i} \cup P_{j} \cup P_{k}$ is a closed 2-ball. + +Theorem 2. Let $(\mathcal{M}, d_{\mathcal{M}})$ be a 2-dimensional metric space. If a general $VD\{P_i\}_{i=1}^s$ of $\mathcal{M}$ satisfies the FUP, it also satisfies the closed ball property. + +The FUP is much easier to verify than the closed ball property, as we only have to ensure that regions are closed 2-balls. Indeed, the following property can be exploited in this regard. + +Proposition 1. Let $\mathcal{M} = (V,E,F)$ be a manifold polygonal mesh, then $\mathcal{M}$ is a closed 2-ball if and only if its Euler characteristic $\chi = |V| - |E| + |F|$ is 1. + +Dual mesh representation. When we compute a VD of a triangle mesh, we want to partition the vertices into disjoint sets. However, Voronoi texels are defined as regions over the surface, meaning that they should be submanifolds of $\mathcal{M}$ . By only considering vertices, we likely end up having non-manifold geometries or non well-defined boundaries. + +![](images/52650bede014b0df04c1a9169bff639112145d3a025fa4bab899d031a2a522e4.jpg) +Fig. 3: Top: the primal and dual connectivities of a triangular mesh. Bottom: regions on the primal mesh are represented as sets of vertices, and the corresponding regions on the dual mesh are composed by faces. + +We consider the polygonal mesh $\tilde{\mathcal{M}} = (\tilde{V},\tilde{E},\tilde{T})$ to be the dual mesh of $\mathcal{M} = (V,E,T)$ , built by placing a vertex at each face of $\mathcal{M}$ and forming a polygonal face for each triangle fan around a vertex of $\mathcal{M}$ . As shown in the example from Figure 3, dual meshes allow us to define Voronoi texels by means of connected dual faces, making them actual submanifolds of the original surface. Furthermore, we can easily define boundaries between texels as paths made of dual edges. + +This construction has the positive side effect to ensure that every VD is general. Indeed, the only place where three or more texels can meet is a dual vertex. Since texels are made of dual faces (i.e., primal vertices), and the primal mesh is triangular, there is no dual vertex where more than three faces can meet. + +Given the correspondence between primal and dual geometric elements, we can verify if a region is a closed 2- + +ball with Proposition 1 without explicitly constructing the dual mesh. Additional details are provided in the supplementary material. After the initial FPS, we further add samples to enforce the flat union property: if two or more adjacent texels do not form a closed 2-ball, we break the connection adding samples at the Voronoi vertex or along the Voronoi edge; if a texel is not a closed 2-ball, we add a sample along its boundary to reduce its coverage. + +# 3.2 Low-resolution matching + +Let $\mathcal{M}$ and $\mathcal{N}$ be two triangular meshes, and let $\hat{\mathcal{M}}$ and $\hat{\mathcal{N}}$ be their low-resolution counterparts computed with the algorithm presented above (or any other alternative algorithm). Furthermore, let $\hat{\Phi}$ be a basis for the functional space over $\hat{\mathcal{M}}$ , and $\hat{\Psi}$ a basis for the functional space over $\hat{\mathcal{N}}$ . For example these bases can + +be the truncated subset of the eigenvectors of the LBO as proposed in [47]. We assume to have some pipeline yielding a functional map $\mathbf{C}$ , such that + +$$ +\hat {\boldsymbol {\Psi}} \mathbf {C} \approx \hat {\boldsymbol {\Pi}} \hat {\boldsymbol {\Phi}}, \tag {3} +$$ + +where $\hat{\Pi}:\hat{\mathcal{M}}\to \hat{\mathcal{N}}$ is a correspondence between the vertices of $\hat{\mathcal{M}}$ and the vertices of $\hat{\mathcal{N}}$ . The adopted functional maps solution can be any of the available alternatives, ranging from the original one [47] to the most recent [12]. In our experiments, we consider two of the widely adopted solutions: constrained optimization with product preservation [46] and the iterative procedure ZoomOut [38]. The first is representative of the possible optimization strategies, while the second is probably the most efficient refinement technique for functional maps estimation. Both these methods have been efficiently and effectively applied on meshes with a limited number of vertices (i.e., up to 10 thousands vertices), which is exactly the setting we are in after the proposed remeshing step. + +# 3.3 Extending the correspondence + +To extend the basis from the remeshing to the original surface, we need a mapping that transports scalar fields from the low-resolution mesh $\hat{\mathcal{M}}$ to the high-resolution shape $\mathcal{M}$ . A possible solution would be to map each triangle of $\hat{\mathcal{M}}$ to a geodesic triangle in $\mathcal{M}$ , and then computing the geodesic barycentric coordinates of each vertex inside the triangle [31]. Unfortunately, this approach is very costly, as it requires to compute exact geodesic paths onto $\mathcal{M}$ [56]. Instead, we approximate this mapping by projecting each vertex of $\mathcal{M}$ onto the closest surface point of $\hat{\mathcal{M}}$ . Despite + +![](images/5d3f9704c11ded32c90244c4f873dbea69cd461210c21b679871337bdfecef2f.jpg) +Fig. 4: Top: remeshing and mapping of the original vertices with the solution from Liu et al. [31] and our pipeline. Bottom: comparison of ground truth geodesic distances from multiple vertices and extended with both methods. + +this approach not being accurate in general, every triangle in $\hat{\mathcal{M}}$ corresponds to a geodesic triangle (homeomorphic to a disk) on $\mathcal{M}$ with the same vertices. Furthermore, starting from a farthest point sampling gives evenly spaced samples, mitigating the geometric complexity of geodesic triangles. + +A surface point of $\hat{\mathcal{M}}$ is either a vertex, or a weighted average of two (if on an edge) or three vertices (if on a triangle). We build a linear map $\mathbf{U}_{\mathcal{M}} \in \mathbb{R}^{m \times s}$ representing the projection of each vertex $v_i \in \mathcal{M}$ as a linear combination of at most three vertices in $\hat{\mathcal{M}}$ (i.e., the vertices encoding the closest surface point). + +Figure 4 shows a comparison between our pipeline and the simplification method proposed by Liu et al. [31]. The top row shows the different geometries produced by the two methods in reducing the mesh size by $90\%$ and how our mapping compares to the approximate geodesic barycentric coordinates. In the bottom row we compare the two methods in approximating a multiple source geodesic distance field, showing that they do not present appreciable differences. More discussion is provided in the supplementary materials. + +Scalar functions can be evaluated at any surface point by interpolating the values at the vertices. Since every vertex of $\mathcal{M}$ is associated to a point on the surface of $\hat{\mathcal{M}}$ , we can use the linear map $\mathbf{U}_{\mathcal{M}}$ for extending scalar functions from the $s$ vertices of $\hat{\mathcal{M}}$ to the $m$ vertices of $\mathcal{M}$ . + +We extend the bases $\hat{\Phi}$ and $\hat{\Psi}$ to the full-resolution meshes as $\Phi = \mathbf{U}_{\mathcal{M}}\hat{\Phi}$ and $\Psi = \mathbf{U}_{\mathcal{N}}\hat{\Psi}$ . Then, in a similar fashion as done in Equation (3), we search for a correspondence $\Pi : \mathcal{M} \to \mathcal{N}$ that maps vertices of $\mathcal{M}$ to vertices of $\mathcal{N}$ as + +$$ +\mathbf {U} _ {\mathcal {N}} \hat {\boldsymbol {\Psi}} \mathbf {C} = \boldsymbol {\Psi} \mathbf {C} \approx \boldsymbol {\Pi} \boldsymbol {\Phi} = \boldsymbol {\Pi} \mathbf {U} _ {\mathcal {M}} \hat {\boldsymbol {\Phi}}, \tag {4} +$$ + +where, for the sake of clarity, we explicitly write the equation in terms of the bases $\hat{\Phi}$ and $\hat{\Psi}$ . Given the extended bases $\Phi$ and $\Psi$ , and the same functional map $\mathbf{C}$ estimated to solve Equation (3), we obtain the correspondence $\Pi$ via a nearest neighbor search in the spectral space. + +We notice that nearest neighbor algorithms become very slow on large meshes, effectively becoming a possible bottleneck for the pipeline. Our experiments presented in Figure 10 show that, to the scope of this paper, this solution provides satisfying performance. However, we acknowledge that establishing a relationship between (3) and (4) could lead to a more efficient expression of $\Pi$ in terms of $\hat{\Pi}$ , $\mathbf{U}_{\mathcal{M}}$ , and $\mathbf{U}_{\mathcal{N}}$ . This avenue warrants future exploration. + +# 4 Results + +Our goal is to evaluate not only the matching accuracy of our method but also its time performance. For this task, we test it on datasets for dense correspondences, such as the SHREC19 dataset [37] and the TOSCA dataset [6]. We also introduce a novel dataset, BadTOSCA, obtained by randomly altering the vertex positions of the TOSCA meshes, to ensure that our method is stable even under circumstances where strong isometry cannot be assumed3. + +Our pipeline relies on an efficient remeshing algorithm that can quickly and drastically reduce the size of a mesh by orders of magnitude while still preserving the underlying geometry and metric. In principle, any remeshing algorithm could be adopted in our pipeline, as the map for extending scalar fields discussed in Section 3.3 only requires having two input meshes. We compare our solution against well-established isotropic and anisotropic remeshing algorithms (respectively, IRM [4,21] and ARM [44]), as well as the scalable functional maps (SFM) + +![](images/0df69ac21962c60babee3f06b84dd9565e8bf434c4bbad281971c3f671f993f2.jpg) +Fig. 5: Left: Accuracy curves on the SHREC19 challenge pairs for the tested methods. We also consider ideal approaches where the initial functional map for ZoomOut is given ground truth dashed lines). Right: Cumulative curves of the execution time for evaluated methods. For all methods, we also consider the time required to run the remeshing or resampling step(s). + +![](images/8a480c5dff1e07ec217422620535403a8112a65022b65da28812005a941a0a73.jpg) + +approach proposed by Magnet et al. [35], which produces a subsampling of the vertices in place of a low-resolution mesh. + +We implemented our remeshing algorithm in $\mathbf{C}++$ , using Eigen [20] and libigl [22], and the matching pipeline in MATLAB4. + +# 4.1 Quality of matching + +To evaluate our shape matching pipeline, we tested our method on the SHREC19 dataset [37]. We can use different shape correspondence techniques to test the entire pipeline. In particular, we compute the functional map on the low-resolution representation using both the widely adopted approach with products preservation (FMaps) [46] + +and ZoomOut [38]. We compare our method with the same approaches on the fullresolution meshes, as well as against SFM. For a fair comparison, we remesh all the + +
MethodAGE (·10-2) ↓AUC ↑
Ours (ZoomOut)*2.4095.12%
SFM*2.4695.22%
Ours (FMaps)6.2388.21%
Ours (ZoomOut)5.8788.82%
SFM16.8470.19%
ARM (ZoomOut)41.5930.34%
IRM (ZoomOut)8.3084.62%
FMaps28.3156.15%
ZoomOut29.3652.02%
+ +Table 1: Average geodesic error and area under the accuracy curve for each tested method on the SHREC19 challenge pairs. The top rows (denoted by \*) show the performance of our algorithm and SFM under the assumption of having a ground truth functional map for initializing ZoomOut. + +shapes to 3k vertices with all methods, as Magnet et al. [35] state that this produces the best performance with SFM. To further ease the alignment of the Laplacian spectra, we also post-process the remeshed shapes (produced with + +![](images/9264c224dec92854c5c1f2018de35375a4fa180ae419ad4d6e8e9900a5b4c4ff.jpg) +Fig. 6: Accuracy curves and average geodesic error on the TOSCA dataset (left) and the BadTOSCA dataset (right) for the tested methods. + +![](images/1a2a8b2fd554caf4414acdf9387979209afd2081b10d7cc3bb3e12c9df6511b4.jpg) + +IRM, ARM, and our algorithm) to remove small disconnected components made of few triangles before computing the extension map discussed in Section 3.3. This post-processing step cannot be performed for SFM, due to the nature of its sampling strategy. The benefits of this post-processing will be discussed in Section 4.3. + +We summarize the results of our experiment in Figure 5, where we show the accuracy curves for all methods on the SHREC19 connectivity track benchmark, following the paradigm proposed by Kim et al. [24]. To prove the time efficiency of our pipeline, we also measure the time taken by every method on each pair of shapes in the benchmark (including the remeshing step), and we show the cumulative curves of the execution times over the entire dataset in a similar way with respect to the accuracy curves. + +![](images/2133191e7d4ccc87aed022750fe3f300b5436d5204dc336a9b4b1b75c80e2205.jpg) +Fig.7: Coordinate transfer between nonisometric shapes. The coordinates are used to generate highly complex patterns, whose sensitivity to the input enhances the transfer errors. + +We see that all the low-resolution approaches provide better time and quality performance than the fullresolution methods, as the simpler geometry makes it easier and faster to align the Laplacian eigenfunctions. ARM is the only exception, since the anisotropic remeshing produces many skew triangles and a singular Laplacian matrix for all the shapes. Furthermore, our technique (with both the FMaps and ZoomOut backend) outperforms SFM and achieves better results than using the IRM remeshing strategy. This difference can be better + +appreciated in Table 1, where we report the average geodesic error (normalized with respect to the shape diameter) and the area under the accuracy curves. In both Figure 5 and Table 1 we have also considered the idealized scenario where our method with ZoomOut as backend and SFM are initialized with a ground truth functional map. In this case, the two approaches obtain comparable results. + +We also compare our method with the other approaches on the TOSCA and BadTOSCA datasets, summarizing the results in Figure 6. In the figure, we show + +![](images/f9949d10a74651ebcfc43cb1804d82f9380865b255c15d7349e1ccfe06d8d654.jpg) +Fig.8: Inter-class coordinate transfer from the wolf model to a cat and a dog. Our pipeline is not constrained by the standard Laplacian basis and can be used with other approaches as well. + +![](images/2028f44b42e83515274e6cfad997c0df65f4b6cfeb0924d5cd08a9c65abfd6f4.jpg) +Fig. 9: Comparison of function transfer between very high-resolution models with SFM and our approach. + +the accuracy curves for the tested methods on both datasets, also providing the average geodesic error normalized by the shape diameter. We see that under the favorable and unrealistic conditions offered by TOSCA, where shapes of the same class are almost perfectly isometric, SFM achieves the best results. However, when we consider the slightly altered geometry of BadTOSCA, SFM suffers a performance drop of about $50\%$ , while our method shows more stability. + +A significant benefit of the functional map approach is its independence from the difference in resolution between the source and target shape. Transferring a function with functional maps yields a much smoother and neat result than directly a point-wise correspondence. In this regard, we test our method in transferring coordinate functions between isometric and non-isometric shapes with substantial differences in resolution and triangulation. In the example from Figure 7 we generated a complex procedural fractal pattern as a function of the coordinates [33]. This pattern is very sensitive to the input, so even slight errors are highly enhanced. We see that SFM cannot precisely map the coordinates of the source mesh, shifting around and distorting the details of the pattern. + +Furthermore, the example in Figure 8 shows that we can also deal with inter-class function transfer with our pipeline, outperforming SFM in a visual comparison. In the example, we also showcase that we should not necessarily rely on the standard Laplacian basis as a backend. Instead of using ZoomOut to transfer functions between the low-resolution meshes, we rely on the orthogonalized eigenproducts basis introduced by Maggioli et al. [34], extending it with the same technique described in Section 3.2. + +Finally, to ensure that our method can scale to high-resolution meshes, we tested it using two very high-resolution 3D scans of real statues: a dancing faun counting $\sim 750\mathrm{k}$ vertices and a warrior from the temple of Aphaea counting $\sim 3.5\mathrm{M}$ vertices (see Figures 1 and 9). Since SFM can also deal with dense shapes, we compare it against our method. Figure 9 shows an example of the transfer of a geodesic distance function. While our method can faithfully transfer the function, SFM produces incoherent distances (e.g., wrong isolines on the head and near the shoulder) and evident artefacts (e.g., wrong distance on the hand). + +# 4.2 Scalable performance + +In Section 4.1 we saw that general-purpose remeshing solutions are not an optimal choice for the matching task. However, Lescoat et al. [28] provided a specialized mesh decimation algorithm (SMD) that reduces the complexity of the mesh while still preserving the Laplacian eigenfunctions. Furthermore, Liu et al. [31] proposed an elegant solution for simplifying a shape based on intrinsic error metric (IEM), which also builds a map for extending scalar functions using geodesic barycentric coordinates. We perform the evaluation on the SHREC19 dataset [37], remeshing the surfaces to 3k vertices with all the methods. + +To achieve the desired scalability, the remeshing step in our pipeline must be fast enough to maintain the improvement obtained by reducing the size of the eigenproblem. In this regard, Figure 10 shows that our technique achieves better time performance than the other remeshing algorithms and the sampling strategy of SFM. In contrast, SMD and IEM are decimation algorithms that iteratively reduce the mesh size, making them orders of magnitude slower and unsuitable for large meshes. Moreover, SMD requires a pre-computation of the Laplacian eigenbasis on the original surface to guide the decimation process and preserve the spectrum, introducing additional time complexity. + +# 4.3 Benefits of remeshing + +![](images/33fa98f4caa616ea99b1f70e11bb472b12ebdee962d007dee1cdc28d9635f51a.jpg) +Fig.10: Execution time (logarithmic scale on the $y$ -axis) of different remeshing algorithms plotted against the number of vertices of the input shape ( $x$ -axis). For SFM we show the time taken for generating the sampling. + +![](images/ff40afdeb38a5cc92852b474a2e754ef6bc15fa1a8e05a97edabb40fe3ada0ee.jpg) +Fig.11: Top row: coordinate transfer between a pair from SHREC19 using ZoomOut on full-resolution meshes, SFM and our approach. Bottom row: The ground truth $20 \times 20$ functional map, compared with the alignment computed with FMaps [46] from the full Laplacian eigenbasis, the SFM extended basis and our extended basis. + +In Section 4.1, we discussed how we remove small disconnected components after the remeshing step. The example in Figure 11 shows a comparison of mapping + +between a low-resolution mesh (Source), with less than 7k vertices, to a detailed mesh where eyes are represented with two disconnected components, each composed of 462 vertices out of the 27k of the entire mesh ( $\sim 2\%$ of the total vertices). This is enough geometry to attract energy at lower frequencies, and when we try to align the first 20 eigenfunctions with FMaps [46], the resulting functional map turns out to be mostly noise. Initializing ZoomOut with such a map leads to meaningless correspondence, and SFM does not mitigate this issue, as it tries to preserve the original spectrum as much as possible. In contrast, by removing these components and mapping them to their closest surface points, we introduce a small error in the final correspondence, but at the same time, we ease the alignment of the initial eigenbases, resulting in more accurate and meaningful mapping. This is evident in the bottom row of Figure 11, where the ground truth functional map presents a gap in the alignment of the eigenfunctions due to the impossibility of mapping the disconnected eyes of the target shape in any point of the source shape. Using the real eigenfunctions or a close approximation yields an intense noise in the map, producing the meaningless mapping of ZoomOut and SFM. Instead, by mapping the eyes to their nearest point on the head, we can meaningfully align our extended basis. + +# 5 Conclusions + +We presented a new pipeline for scalable shape matching that relies on a low-resolution representation to compute a functional map between dense shapes efficiently. A core part of this pipeline is our novel remeshing algorithm, which efficiently computes an intrinsic Delaunay triangulation of uniform surface sampling. This procedure is applied to reduce the computational cost of the functional maps framework and make it efficient also on very dense meshes. Furthermore, we use a fast and intuitive technique for extending scalar functions from low-resolution meshes to their dense counterparts, extending the computed correspondence to the original very high-resolution shapes. Our experimental evaluation proved that our procedure is effective on various types of shapes at different resolutions, outperforming other state-of-the-art solutions and solving topological issues related to the alignment of the Laplacian eigenbasis. + +While using a triangulated mesh as a low-resolution representation could, in principle, be exploited with other types of shape-matching pipelines, we only tested it within the functional maps framework, using the Laplace-Beltrami eigenbasis or bases derived from it. While the current pipeline binds the method's applicability to the setting of functional correspondences, we believe that studying its behaviour with different "backends", as well as exploring the relationships between the low-resolution and the high-resolution correspondences, would be an interesting matter for future investigations. + +# Acknowledgements + +This work was partially supported by the Italian Ministry of Education, Universities and Research under the grant Dipartimento di Eccellenza 2023-2027 of the Department of Informatics, Systems and Communication of the University of Milano-Bicocca and by the PRIN 2022 project GEOPRIDE Geometric primitive fitting on 3D data for geometric analysis and 3D shapes. This work is also supported by the ERC grant no.802554 SPECGEO and PRIN 2020 project no.2020TA3K9N LEGO.AI. We acknowledge the support of NVIDIA Corporation with the RTX A5000 GPUs granted through the Academic Hardware Grant Program to the University of Milano-Bicocca for the project Learned representations for implicit binary operations on real-world 2D-3D data. + +# References + +1. Aflalo, Y., Brezis, H., Kimmel, R.: On the optimality of shape and data representation in the spectral domain. SIAM Journal on Imaging Sciences 8(2), 1141-1160 (2015) +2. Amenta, N., Choi, S., Dey, T.K., Leekha, N.: A simple algorithm for homeomorphic surface reconstruction. In: Proceedings of the sixteenth annual symposium on Computational geometry. pp. 213-222 (2000) +3. Bobenko, A.I., Springborn, B.A.: A discrete laplace-beltrami operator for simplicial surfaces. Discrete & Computational Geometry 38(4), 740-756 (2007) +4. Botsch, M., Pauly, M., Rossl, C., Bischoff, S., Kobbelt, L.: Geometric modeling based on triangle meshes. In: ACM SIGGRAPH 2006 Courses, pp. 1-es (2006) +5. Bowyer, A.: Computing Dirichlet tessellations*. The Computer Journal 24(2), 162-166 (01 1981) +6. Bronstein, A., Bronstein, M., Kimmel, R.: Numerical Geometry of Non-Rigid Shapes. Springer, New York, NY (2008) +7. Burghard, O., Dieckmann, A., Klein, R.: Embedding shapes with green's functions for global shape matching. Computers & Graphics 68, 1-10 (2017) +8. Crane, K., Weischedel, C., Wardetzky, M.: Geodesics in heat: A new approach to computing distance based on heat flow. ACM Transactions on Graphics (TOG) 32(5), 1-11 (2013) +9. Delaunay, B., et al.: Sur la sphere vide. Izv. Akad. Nauk SSSR, Otdelenie Matematicheskii i Estestvennyka Nauk 7(793-800), 1-2 (1934) +0. Deng, B., Yao, Y., Dyke, R.M., Zhang, J.: A survey of non-rigid 3d registration. In: Computer Graphics Forum. vol. 41, pp. 559-589. Wiley Online Library (2022) +1. Dijkstra, E.W.: A note on two problems in connexion with graphs. Numerische Mathematik 1, 269-271 (1959) +2. Donati, N., Corman, E., Melzi, S., Ovsjanikov, M.: Complex functional maps: A conformal link between tangent bundles. In: Computer Graphics Forum. vol. 41, pp. 317-334. Wiley Online Library (2022) +3. Donati, N., Sharma, A., Ovsjanikov, M.: Deep geometric functional maps: Robust feature learning for shape correspondence. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8592-8601 (2020) +4. Dyer, R., Zhang, H., Möller, T.: Voronoi-delaunay duality and delaunay meshes. In: Proceedings of the 2007 ACM symposium on Solid and physical modeling. pp. 415-420 (2007) + +15. Edelsbrunner, H., Shah, N.R.: Triangulating topological spaces. In: Proceedings of the tenth annual symposium on Computational geometry. pp. 285-292 (1994) +16. Eynard, D., Rodola, E., Glashoff, K., Bronstein, M.M.: Coupled functional maps. In: 2016 Fourth International Conference on 3D Vision (3DV). pp. 399-407. IEEE (2016) +17. Ezuz, D., Ben-Chen, M.: Deblurring and denoising of maps between shapes. In: Computer Graphics Forum. vol. 36, pp. 165-174. Wiley Online Library (2017) +18. Fei, Y., Chen, S., Su, D., Luo, J., Li, M.: A new algorithm for repairing non-manifold surfaces. In: 2013 IEEE 10th International Conference on High Performance Computing and Communications & 2013 IEEE International Conference on Embedded and Ubiquitous Computing. pp. 1704-1708. IEEE (2013) +19. Gao, M., Roetzer, P., Eisenberger, M., Lähner, Z., Moeller, M., Cremers, D., Bernard, F.: Sigma: Scale-invariant global sparse shape matching. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 645-654 (2023) +20. Guennebaud, G., Jacob, B., et al.: Eigen v3. http://eigen.tuxfamily.org (2010) +21. Hoppe, H., DeRose, T., Duchamp, T., McDonald, J., Stuetzle, W.: Mesh optimization. In: Proceedings of the 20th annual conference on Computer graphics and interactive techniques. pp. 19-26 (1993) +22. Jacobson, A., Panozzo, D., et al.: libigl: A simple $\mathrm{C}++$ geometry processing library (2018), https://libigl.github.io/ +23. Jiang, Z., Schneider, T., Zorin, D., Panozzo, D.: Bijective projection in a shell. ACM Transactions on Graphics (TOG) 39(6), 1-18 (2020) +24. Kim, V.G., Lipman, Y., Funkhouser, T.: Blended intrinsic maps. ACM transactions on graphics (TOG) 30(4), 1-12 (2011) +25. Kimmel, R., Sethian, J.A.: Computing geodesic paths on manifolds. Proceedings of the national academy of Sciences 95(15), 8431-8435 (1998) +26. Kressner, D.: Numerical Methods for General and Structured Eigenvalue Problems. Lecture Notes in Computational Science and Engineering, Springer Berlin Heidelberg (2006) +27. Leibon, G., Letscher, D.: Delaunay triangulations and voronoi diagrams for riemannian manifolds. In: Proceedings of the sixteenth annual symposium on Computational geometry. pp. 341-349 (2000) +28. Lescoat, T., Liu, H.T.D., Thiery, J.M., Jacobson, A., Boubekeur, T., Ovsjanikov, M.: Spectral mesh simplification. In: Computer Graphics Forum. vol. 39, pp. 315-324. Wiley Online Library (2020) +29. Levy, B.: Laplace-beltrami eigenfunctions towards an algorithm that "understands" geometry. In: IEEE International Conference on Shape Modeling and Applications 2006 (SMI'06). pp. 13-13 (2006) +30. Lévy, B., Zhang, H.: Spectral mesh processing. In: ACM SIGGRAPH 2010 Courses, pp. 1-312 (2010) +31. Liu, H.T.D., Gillespie, M., Chislett, B., Sharp, N., Jacobson, A., Crane, K.: Surface simplification using intrinsic error metrics. ACM Trans. Graph. 42(4) (jul 2023). https://doi.org/10.1145/3592403, https://doi.org/10.1145/3592403 +32. Liu, H.T.D., Jacobson, A., Ovsjanikov, M.: Spectral coarsening of geometric operators. ACM Transactions on Graphics (TOG) 38(4) (jul 2019) +33. Maggioli, F., Baieri, D., Melzi, S., Rodolà, E.: Newton's fractals on surfaces via bicomplex algebra. In: ACM SIGGRAPH 2022 Posters, pp. 1-2 (2022) +34. Maggioli, F., Melzi, S., Ovsjanikov, M., Bronstein, M.M., Rodola, E.: Orthogonalized fourier polynomials for signal approximation and transfer. In: Computer Graphics Forum. vol. 40, pp. 435-447. Wiley Online Library (2021) + +35. Magnet, R., Ovsjanikov, M.: Scalable and efficient functional map computations on dense meshes. In: Computer Graphics Forum. vol. 42, pp. 89-101. Wiley Online Library (2023) +36. Marin, R., Corona, E., Pons-Moll, G.: Nicp: Neural icip for 3d human registration at scale. In: European Conference on Computer Vision. Springer (2024) +37. Melzi, S., Marin, R., Rodolà, E., Castellani, U., Ren, J., Poulenard, A., Wonka, P., Ovsjanikov, M.: Shrec 2019: Matching humans with different connectivity. In: Eurographics Workshop on 3D Object Retrieval. vol. 7, p. 3. The Eurographics Association (2019) +38. Melzi, S., Ren, J., Rodola, E., Sharma, A., Wonka, P., Ovsjanikov, M.: Zoomout: spectral upsampling for efficient shape correspondence. ACM Transactions on Graphics (TOG) 38(6), 1-14 (2019) +39. Melzi, S., Rodolà, E., Castellani, U., Bronstein, M.M.: Localized manifold harmonics for spectral shape analysis. In: Computer Graphics Forum. vol. 37, pp. 20-34. Wiley Online Library (2018) +40. Meyer, M., Desbrun, M., Schröder, P., Barr, A.H.: Discrete Differential-Geometry Operators for Triangulated 2-Manifolds. In: Visualization and mathematics III, pp. 35-57. Springer, New York, NY (2003) +41. Mitchell, J.S., Mount, D.M., Papadimitriou, C.H.: The discrete geodesic problem. SIAM Journal on Computing 16(4), 647-668 (1987) +42. Nasikun, A., Brandt, C., Hildebrandt, K.: Fast approximation of laplace-beltrami eigenproblems. In: Computer Graphics Forum. vol. 37, pp. 121-134. Wiley Online Library (2018) +43. Nasikun, A., Hildebrandt, K.: The hierarchical subspace iteration method for laplace-beltrami eigenproblems. ACM Transactions on Graphics (TOG) 41(2), 1-14 (2022) +44. Nivoliers, V., Lévy, B., Geuzaine, C.: Anisotropic and feature sensitive triangular remeshing using normal lifting. Journal of Computational and Applied Mathematics 289, 225-240 (2015) +45. Nogneng, D., Melzi, S., Rodola, E., Castellani, U., Bronstein, M., Ovsjanikov, M.: Improved functional mappings via product preservation. In: Computer Graphics Forum. vol. 37, pp. 179-190. Wiley Online Library (2018) +46. Nogneng, D., Ovsjanikov, M.: Informative descriptor preservation via commutativity for shape matching. In: Computer Graphics Forum. vol. 36, pp. 259-267. Wiley Online Library (2017) +47. Ovsjanikov, M., Ben-Chen, M., Solomon, J., Butscher, A., Guibas, L.: Functional maps: a flexible representation of maps between shapes. ACM Transactions on Graphics (ToG) 31(4), 1-11 (2012) +48. Ovsjanikov, M., Corman, E., Bronstein, M., Rodola, E., Ben-Chen, M., Guibas, L., Chazal, F., Bronstein, A.: Computing and processing correspondences with functional maps. In: SIGGRAPH ASIA 2016 Courses, pp. 1-60 (2016) +49. Peyré, G., Cohen, L.D.: Geodesic remeshing using front propagation. International Journal of Computer Vision 69, 145-156 (2006) +50. Pinkall, U., Polthier, K.: Computing Discrete Minimal Surfaces and their Conjugates. Experimental mathematics 2(1), 15-36 (1993) +51. Ren, J., Melzi, S., Ovsjanikov, M., Wonka, P., et al.: Maptree: recovering multiple solutions in the space of maps. ACM Transactions on Graphics (TOG) 39(6), 264-1 (2020) +52. Ren, J., Melzi, S., Wonka, P., Ovsjanikov, M.: Discrete optimization for shape matching. In: Computer Graphics Forum. vol. 40, pp. 81-96. Wiley Online Library (2021) + +53. Ren, J., Poulenard, A., Wonka, P., Ovsjanikov, M.: Continuous and orientation-preserving correspondences via functional maps. ACM Transactions on Graphics (ToG) 37(6), 1-16 (2018) +54. Reuter, M., Wolter, F.E., Peinecke, N.: Laplace-beltrami spectra as 'shape-dna' of surfaces and solids. Computer-Aided Design 38(4), 342-366 (2006) +55. Rivin, I.: Euclidean structures on simplicial surfaces and hyperbolic volume. Annals of mathematics 139(3), 553-580 (1994) +56. Rustamov, R.M.: Barycentric coordinates on surfaces. In: Computer Graphics Forum. vol. 29, pp. 1507-1516. Wiley Online Library (2010) +57. Sahillioğlu, Y.: Recent advances in shape correspondence. The Visual Computer 36(8), 1705-1721 (2020) +58. Shoham, M., Vaxman, A., Ben-Chen, M.: Hierarchical functional maps between subdivision surfaces. In: Computer Graphics Forum. vol. 38, pp. 55-73. Wiley Online Library (2019) +59. Vaxman, A., Ben-Chen, M., Gotsman, C.: A multi-resolution approach to heat kernels on discrete surfaces. In: ACM SIGGRAPH 2010 papers, pp. 1-10 (2010) +60. Watson, D.F.: Computing the n-dimensional Delaunay tessellation with application to Voronoi polytopes*. The Computer Journal 24(2), 167-172 (01 1981) \ No newline at end of file diff --git a/2024/ReMatching_ Low-Resolution Representations for Scalable Shape Correspondence/images.zip b/2024/ReMatching_ Low-Resolution Representations for Scalable Shape Correspondence/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7487eb0b2064578b3188c69c59708b1487e2b0ad --- /dev/null +++ b/2024/ReMatching_ Low-Resolution Representations for Scalable Shape Correspondence/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:615142ec562584c58164c84148b9f4234a5fbc094d4117f57b398851908c6c76 +size 421905 diff --git a/2024/ReMatching_ Low-Resolution Representations for Scalable Shape Correspondence/layout.json b/2024/ReMatching_ Low-Resolution Representations for Scalable Shape Correspondence/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7d579af7ce0cf4b6ee48a7970a52e0d3672223de --- /dev/null +++ b/2024/ReMatching_ Low-Resolution Representations for Scalable Shape Correspondence/layout.json @@ -0,0 +1,12091 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 135, + 111, + 479, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 111, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 135, + 111, + 479, + 148 + ], + "type": "text", + "content": "ReMatching: Low-Resolution Representations for Scalable Shape Correspondence" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 148, + 167, + 465, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 167, + 465, + 193 + ], + "spans": [ + { + "bbox": [ + 148, + 167, + 465, + 193 + ], + "type": "text", + "content": "Filippo Maggioli1, Daniele Baieri2, Emanuele Rodolà2, and Simone Melzi1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 212, + 201, + 400, + 224 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 214, + 201, + 397, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 201, + 397, + 213 + ], + "spans": [ + { + "bbox": [ + 214, + 201, + 397, + 213 + ], + "type": "text", + "content": "1 University of Milano-Bicocca, Milan, Italy" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 212, + 213, + 400, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 213, + 400, + 224 + ], + "spans": [ + { + "bbox": [ + 212, + 213, + 400, + 224 + ], + "type": "text", + "content": "2 Sapienza - University of Rome, Rome, Italy" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 133, + 256, + 480, + 392 + ], + "blocks": [ + { + "bbox": [ + 133, + 256, + 480, + 392 + ], + "lines": [ + { + "bbox": [ + 133, + 256, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 133, + 256, + 480, + 392 + ], + "type": "image", + "image_path": "054bd1a6d36e11904940f38384a53e57a94f58ecc74a27bbc669d4eb29a9846a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 400, + 482, + 456 + ], + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 456 + ], + "type": "text", + "content": "Fig. 1: Correspondence between two dense shapes computed with our method, using ZoomOut [38] on low-resolution meshes. The coordinates functions of the dancing faun statue on the left (" + }, + { + "bbox": [ + 130, + 400, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\sim750\\mathrm{k}" + }, + { + "bbox": [ + 130, + 400, + 482, + 456 + ], + "type": "text", + "content": " vertices) are transferred using a functional map to the Aphaea warrior statue on the right (" + }, + { + "bbox": [ + 130, + 400, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\sim3.5\\mathrm{M}" + }, + { + "bbox": [ + 130, + 400, + 482, + 456 + ], + "type": "text", + "content": " vertices). The colors encode the correspondence. Despite the mesh density, shown in the close-up, the computation took " + }, + { + "bbox": [ + 130, + 400, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\sim2" + }, + { + "bbox": [ + 130, + 400, + 482, + 456 + ], + "type": "text", + "content": " minutes." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 501, + 455, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 501, + 455, + 666 + ], + "spans": [ + { + "bbox": [ + 159, + 501, + 455, + 666 + ], + "type": "text", + "content": "Abstract. We introduce ReMatching, a novel shape correspondence solution based on the functional maps framework. Our method, by exploiting a new and appropriate re-meshing paradigm, can target shape-matching tasks even on meshes counting millions of vertices, where the original functional maps does not apply or requires a massive computational cost. The core of our procedure is a time-efficient remeshing algorithm which constructs a low-resolution geometry while acting conservatively on the original topology and metric. These properties allow translating the functional maps optimization problem on the resulting low-resolution representation, thus enabling efficient computation of correspondences with functional map approaches. Finally, we propose an efficient technique for extending the estimated correspondence to the original meshes. We show that our method is more efficient and effective through quantitative and qualitative comparisons, outperforming state-of-the-art pipelines in quality and computational cost." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 149, + 332, + 162 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 332, + 162 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 332, + 162 + ], + "type": "text", + "content": "1 Introduction and related work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 175, + 481, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 175, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 481, + 281 + ], + "type": "text", + "content": "The task of finding a semantically meaningful correspondence between discrete surfaces has always been a fundamental topic in the field of shape analysis. Researchers developed a wealth of solutions for this application, and new methods continue to be designed [10, 57]. Among the various approaches, the functional maps framework [47, 48] received significant attention. Rather than finding a point-wise correspondence, the functional maps framework aims to define a correspondence between functions, encoding it into a small matrix. Following this direction, a variety of works tried to improve the actual computation of the map [13, 38, 53] or to extend the framework to different types of bases [34, 39, 45]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 282, + 481, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 481, + 474 + ], + "type": "text", + "content": "Despite the amount of research outcomes based on the functional maps framework, both exploiting geometric properties and tools [7, 16, 17] and clever optimization techniques [46, 51, 52], the problem of computing the map is still bounded by the time complexity of sparse eigendecompositions [26]. To overcome this problem, researchers started investigating scalable solutions to the Laplace-Beltrami eigenproblem. In this regard, multi-resolution techniques [42, 43, 58, 59] proved to be very effective for scalable computation of spectral quantities for tasks such as retrieval and mesh filtering [30, 54], but their applicability to the functional map framework proved to require additional care [35]. Furthermore, recent research proposed spectral coarsening methods [32] or spectral preserving simplification techniques [28]. However, these techniques usually rely on the computation of the full-resolution spectrum, making them unsuitable for large-scale applications. Gao et al. [19] proposed a mixed-integer programming scalable solution to the matching problem, but they specifically target sparse correspondences only. Alternatively, Marin et al. [36], who provide an approach for human registration that scales across large datasets, but it is unsuitable for dense shapes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 474, + 481, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 582 + ], + "type": "text", + "content": "Closely related to our work is the contribution from Magnet et al. [35], which, for the first time, proposes a solution for scaling the functional maps approach to high-resolution meshes. Their idea is to reduce the dimensionality of the eigenproblem, inducing a linear relationship between the functional space on the mesh and a lower-dimension functional space on a sparse sample. The extreme sparsity of the sampling, combined with its efficient computation, leads to an algorithm that can effectively deal with high-density meshes. Nevertheless, this type of reduction does not mitigate the negative influence that small components and local details at high frequency could have on the alignment of the spectra." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": "Finally, our pipeline relies on building robust and sparse representations of dense meshes while still keeping an approximately bijective correspondence between the high- and low-resolution shapes, and it is worth mentioning that other works have already attempted to do so. Jiang et al. [23] propose building a lower resolution prismatic shell preserving bijectivity with the original surface. However, their algorithm does not scale to meshes with a high triangle count, gives no control and no guarantees on the final vertex count, and requires many assumption." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "text", + "content": "F. Maggioli et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 160, + 117, + 449, + 128 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 117, + 449, + 128 + ], + "spans": [ + { + "bbox": [ + 160, + 117, + 449, + 128 + ], + "type": "text", + "content": "Keywords: Shape matching " + }, + { + "bbox": [ + 160, + 117, + 449, + 128 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 117, + 449, + 128 + ], + "type": "text", + "content": " Functional maps " + }, + { + "bbox": [ + 160, + 117, + 449, + 128 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 117, + 449, + 128 + ], + "type": "text", + "content": " Geometry processing" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": "tions on the input shape (e.g., manifoldness, orientability, no self-intersections). On the other hand, Liu et al. [31] propose to exploit an intrinsic error metric for decimating a mesh to a given size while keeping track of a geodesic barycentric mapping of each triangle of the resulting shape to a geodesic triangle on the input surface. Despite the robustness of the method, the simplification approach makes it unsuitable for a functional maps setting, where it is required to decimate shapes from millions of vertices to a few thousand. This limitation is further discussed in Section 4.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 479, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 479, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 479, + 247 + ], + "type": "text", + "content": "We introduce a new scalable functional map pipeline that efficiently handles meshes with high vertex density (see Figure 1), yields stable results, and is not bounded by the quality of the original triangulation. To summarize:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 248, + 479, + 342 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 136, + 248, + 479, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 248, + 479, + 283 + ], + "spans": [ + { + "bbox": [ + 136, + 248, + 479, + 283 + ], + "type": "text", + "content": "- we translate the matching pipeline to low-resolution representations, enabling a fast and scalable computation of functional maps between dense shapes, providing the first alternative to [35];" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 284, + 479, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 284, + 479, + 307 + ], + "spans": [ + { + "bbox": [ + 137, + 284, + 479, + 307 + ], + "type": "text", + "content": "- we propose a new efficient and geometry-preserving remeshing algorithm specifically designed for our pipeline;" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 308, + 479, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 308, + 479, + 342 + ], + "spans": [ + { + "bbox": [ + 137, + 308, + 479, + 342 + ], + "type": "text", + "content": "- we exploit a fast solution for extending scalar maps from the low-resolution remeshed shape to the original surface, thus efficiently obtaining suitable bases for function transfer and shape matching." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 362, + 225, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 362, + 225, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 362, + 225, + 376 + ], + "type": "text", + "content": "2 Background" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 388, + 479, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 479, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 479, + 424 + ], + "type": "text", + "content": "In the discrete setting, we represent a shape as a triangular mesh " + }, + { + "bbox": [ + 130, + 388, + 479, + 424 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = (V,E,T)" + }, + { + "bbox": [ + 130, + 388, + 479, + 424 + ], + "type": "text", + "content": " where (i) " + }, + { + "bbox": [ + 130, + 388, + 479, + 424 + ], + "type": "inline_equation", + "content": "V\\subset \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 388, + 479, + 424 + ], + "type": "text", + "content": " is a set of vertices; (ii) " + }, + { + "bbox": [ + 130, + 388, + 479, + 424 + ], + "type": "inline_equation", + "content": "E\\subset V^2" + }, + { + "bbox": [ + 130, + 388, + 479, + 424 + ], + "type": "text", + "content": " is a set of edges between vertices; (iii) " + }, + { + "bbox": [ + 130, + 388, + 479, + 424 + ], + "type": "inline_equation", + "content": "T\\subset V^3" + }, + { + "bbox": [ + 130, + 388, + 479, + 424 + ], + "type": "text", + "content": " is a set of triangles composing the surface." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 441, + 243, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 243, + 454 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 243, + 454 + ], + "type": "text", + "content": "2.1 Functional maps" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 462, + 480, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 480, + 544 + ], + "type": "text", + "content": "We discretize a scalar function " + }, + { + "bbox": [ + 130, + 462, + 480, + 544 + ], + "type": "inline_equation", + "content": "f: \\mathcal{M} \\to \\mathbb{R}" + }, + { + "bbox": [ + 130, + 462, + 480, + 544 + ], + "type": "text", + "content": " as a signal defined on the vertices " + }, + { + "bbox": [ + 130, + 462, + 480, + 544 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 462, + 480, + 544 + ], + "type": "text", + "content": " and represented as a vector " + }, + { + "bbox": [ + 130, + 462, + 480, + 544 + ], + "type": "inline_equation", + "content": "\\pmb{f} \\in \\mathbb{R}^{|V|}" + }, + { + "bbox": [ + 130, + 462, + 480, + 544 + ], + "type": "text", + "content": ". Hence, the Laplace-Beltrami operator " + }, + { + "bbox": [ + 130, + 462, + 480, + 544 + ], + "type": "inline_equation", + "content": "\\varDelta_{\\mathcal{M}}: \\mathcal{F}(\\mathcal{M}) \\to \\mathcal{F}(\\mathcal{M})" + }, + { + "bbox": [ + 130, + 462, + 480, + 544 + ], + "type": "text", + "content": " is discretized as a sparse matrix " + }, + { + "bbox": [ + 130, + 462, + 480, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathcal{M}} \\in \\mathbb{R}^{|V| \\times |V|}" + }, + { + "bbox": [ + 130, + 462, + 480, + 544 + ], + "type": "text", + "content": ", generally represented by means of a stiffness matrix " + }, + { + "bbox": [ + 130, + 462, + 480, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{\\mathcal{M}}" + }, + { + "bbox": [ + 130, + 462, + 480, + 544 + ], + "type": "text", + "content": " and a mass matrix " + }, + { + "bbox": [ + 130, + 462, + 480, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{\\mathcal{M}}" + }, + { + "bbox": [ + 130, + 462, + 480, + 544 + ], + "type": "text", + "content": " [40, 50]. The eigendecomposition " + }, + { + "bbox": [ + 130, + 462, + 480, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{\\mathcal{M}} \\pmb{\\Phi}_{\\mathcal{M}} = \\mathbf{A}_{\\mathcal{M}} \\pmb{\\Phi}_{\\mathcal{M}} \\varLambda_{\\mathcal{M}}" + }, + { + "bbox": [ + 130, + 462, + 480, + 544 + ], + "type": "text", + "content": " of the Laplacian yields an orthonormal basis for the functional space on " + }, + { + "bbox": [ + 130, + 462, + 480, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 462, + 480, + 544 + ], + "type": "text", + "content": " [29,30] which has analogies with the Fourier basis and optimally represents smooth functions when truncated [1]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "text", + "content": "Given two shapes " + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "text", + "content": ", respectively with " + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "text", + "content": " vertices, a point-to-point map " + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "inline_equation", + "content": "\\Pi : \\mathcal{M} \\to \\mathcal{N}" + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "text", + "content": " can be expressed as a binary matrix " + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "inline_equation", + "content": "\\Pi \\in \\{0,1\\}^{m \\times n}" + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "inline_equation", + "content": "\\Pi(i,j) = 1" + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "text", + "content": " if the vertex " + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "text", + "content": "-th of " + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "text", + "content": " corresponds to the vertex " + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "text", + "content": "-th of " + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "inline_equation", + "content": "\\Pi(i,j) = 0" + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "text", + "content": " otherwise. Any point-wise correspondence " + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "inline_equation", + "content": "\\Pi" + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "text", + "content": " establishes a functional map " + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "inline_equation", + "content": "T_{\\Pi} : \\mathcal{F}(\\mathcal{N}) \\to \\mathcal{F}(\\mathcal{M})" + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "inline_equation", + "content": "T_{\\Pi}(f) = f \\circ \\Pi" + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "inline_equation", + "content": "\\forall f \\in \\mathcal{F}(\\mathcal{N})" + }, + { + "bbox": [ + 130, + 545, + 480, + 605 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "text", + "content": "As proposed in [47], given " + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\Phi_k" + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\Psi_k" + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "text", + "content": " the Laplacian bases truncated to size " + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "text", + "content": " respectively on " + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "text", + "content": ", we can project " + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "inline_equation", + "content": "T_{\\Pi}" + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "text", + "content": " in these bases and compactly encode the functional map in a " + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "inline_equation", + "content": "k \\times k" + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "text", + "content": " matrix " + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{C} = \\Phi_k^\\top \\mathbf{A}_{\\mathcal{M}} \\Pi \\Psi_k" + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "text", + "content": ", which is the linear operator that transfers the coefficients of functions from " + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathcal{N})" + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "text", + "content": " in the ones of corresponding functions of " + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathcal{M})" + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "text", + "content": " respectively computed with " + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\Phi_k" + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\Psi_k" + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 264, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 264, + 91, + 447, + 102 + ], + "type": "text", + "content": "ReMatching: Scalable Shape Correspondence" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "content": "Finally, given a functional map " + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "inline_equation", + "content": "T_{\\Pi}" + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "content": ", or its compact representation " + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "content": ", we can retrieve the correspondence " + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "inline_equation", + "content": "\\Pi" + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "content": " on which the map is built by the nearest neighbor search in the embedding space of the Laplacian eigenfunctions as detailed in [48]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 167, + 361, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 167, + 361, + 180 + ], + "spans": [ + { + "bbox": [ + 131, + 167, + 361, + 180 + ], + "type": "text", + "content": "2.2 Intrinsic Delaunay triangulations (IDT)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 186, + 479, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 186, + 479, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 186, + 479, + 236 + ], + "type": "text", + "content": "The IDT [3, 55] has been introduced for generalizing the Delaunay triangulation [9] to non-Euclidean metric spaces where classical algorithms [5, 60] cannot be used. IDT relies on the duality with the Voronoi diagram [15, 27], as three texels meeting at a point form a triangle whose edges cross the texels boundaries." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 239, + 480, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 239, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 131, + 239, + 480, + 277 + ], + "type": "text", + "content": "Definition 1. Let " + }, + { + "bbox": [ + 131, + 239, + 480, + 277 + ], + "type": "inline_equation", + "content": "(\\mathcal{M},d_{\\mathcal{M}})" + }, + { + "bbox": [ + 131, + 239, + 480, + 277 + ], + "type": "text", + "content": " be a 2-dimensional metric space (with distance function " + }, + { + "bbox": [ + 131, + 239, + 480, + 277 + ], + "type": "inline_equation", + "content": "d_{\\mathcal{M}}" + }, + { + "bbox": [ + 131, + 239, + 480, + 277 + ], + "type": "text", + "content": "), and let " + }, + { + "bbox": [ + 131, + 239, + 480, + 277 + ], + "type": "inline_equation", + "content": "S = \\{p_i\\}_{i = 1}^s\\subset \\mathcal{M}" + }, + { + "bbox": [ + 131, + 239, + 480, + 277 + ], + "type": "text", + "content": " be a set of " + }, + { + "bbox": [ + 131, + 239, + 480, + 277 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 131, + 239, + 480, + 277 + ], + "type": "text", + "content": " sample points. The Voronoi decomposition (VD) of " + }, + { + "bbox": [ + 131, + 239, + 480, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 131, + 239, + 480, + 277 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 131, + 239, + 480, + 277 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 131, + 239, + 480, + 277 + ], + "type": "text", + "content": " is the collection " + }, + { + "bbox": [ + 131, + 239, + 480, + 277 + ], + "type": "inline_equation", + "content": "\\{P_i\\}_{i = 1}^s" + }, + { + "bbox": [ + 131, + 239, + 480, + 277 + ], + "type": "text", + "content": " such that" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 174, + 283, + 481, + 315 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 283, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 174, + 283, + 481, + 315 + ], + "type": "interline_equation", + "content": "P _ {i} = \\left\\{q \\in \\mathcal {M}: p _ {i} = \\underset {p _ {j} \\in S} {\\arg \\min } \\left(d _ {\\mathcal {M}} \\left(p _ {j}, q\\right)\\right) \\right\\}, \\quad \\bigcup_ {i = 1} ^ {s} P _ {i} = \\mathcal {M} \\tag {1}", + "image_path": "298f050862e232a4436f1b8f040ecbd6538e91e64c97ebfd984e8891231ad176.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 320, + 471, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 320, + 471, + 333 + ], + "spans": [ + { + "bbox": [ + 133, + 320, + 471, + 333 + ], + "type": "text", + "content": "The points in " + }, + { + "bbox": [ + 133, + 320, + 471, + 333 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 133, + 320, + 471, + 333 + ], + "type": "text", + "content": " are called generators, and the " + }, + { + "bbox": [ + 133, + 320, + 471, + 333 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 133, + 320, + 471, + 333 + ], + "type": "text", + "content": " are called texels (or cells)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 338, + 479, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 479, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 479, + 397 + ], + "type": "text", + "content": "The intersection between two or more texels could be non-empty. In such a case, we define those texels as adjacent. A VD is general if the intersection of four or more texels is empty. If a VD is general, a Voronoi vertex is a point that belongs to three texels, and a Voronoi edge is a curve " + }, + { + "bbox": [ + 130, + 338, + 479, + 397 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 338, + 479, + 397 + ], + "type": "text", + "content": " connecting two Voronoi vertices and belonging to two texels." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 399, + 479, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 479, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 479, + 449 + ], + "type": "text", + "content": "It can be shown that there is a one-to-one correspondence between a Voronoi edge dividing " + }, + { + "bbox": [ + 130, + 399, + 479, + 449 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 130, + 399, + 479, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 399, + 479, + 449 + ], + "type": "inline_equation", + "content": "P_{j}" + }, + { + "bbox": [ + 130, + 399, + 479, + 449 + ], + "type": "text", + "content": " and a Delaunay edge connecting " + }, + { + "bbox": [ + 130, + 399, + 479, + 449 + ], + "type": "inline_equation", + "content": "p_{i}" + }, + { + "bbox": [ + 130, + 399, + 479, + 449 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 399, + 479, + 449 + ], + "type": "inline_equation", + "content": "p_{j}" + }, + { + "bbox": [ + 130, + 399, + 479, + 449 + ], + "type": "text", + "content": " [14]. This correspondence gives rise to a necessary and sufficient condition for the IDT to be a proper triangulation (i.e., it realizes a simplicial complex)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 452, + 479, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 452, + 479, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 452, + 479, + 475 + ], + "type": "text", + "content": "Definition 2 ([2,15]). If " + }, + { + "bbox": [ + 130, + 452, + 479, + 475 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 452, + 479, + 475 + ], + "type": "text", + "content": " induces a general " + }, + { + "bbox": [ + 130, + 452, + 479, + 475 + ], + "type": "inline_equation", + "content": "VD" + }, + { + "bbox": [ + 130, + 452, + 479, + 475 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 130, + 452, + 479, + 475 + ], + "type": "inline_equation", + "content": "VD" + }, + { + "bbox": [ + 130, + 452, + 479, + 475 + ], + "type": "text", + "content": " satisfies the closed ball property if:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 476, + 481, + 512 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 138, + 476, + 434, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 476, + 434, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 476, + 434, + 487 + ], + "type": "text", + "content": "- every " + }, + { + "bbox": [ + 138, + 476, + 434, + 487 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 138, + 476, + 434, + 487 + ], + "type": "text", + "content": " is a closed 2-ball (i.e., a topological disk without holes);" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 488, + 481, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 488, + 481, + 499 + ], + "spans": [ + { + "bbox": [ + 138, + 488, + 481, + 499 + ], + "type": "text", + "content": "- every " + }, + { + "bbox": [ + 138, + 488, + 481, + 499 + ], + "type": "inline_equation", + "content": "P_i \\cap P_j" + }, + { + "bbox": [ + 138, + 488, + 481, + 499 + ], + "type": "text", + "content": " is either empty or a closed 1-ball (i.e., a topological segment);" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 500, + 443, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 500, + 443, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 500, + 443, + 512 + ], + "type": "text", + "content": "- every " + }, + { + "bbox": [ + 138, + 500, + 443, + 512 + ], + "type": "inline_equation", + "content": "P_i \\cap P_j \\cap P_k" + }, + { + "bbox": [ + 138, + 500, + 443, + 512 + ], + "type": "text", + "content": " is either empty or a closed 0-ball (i.e., a point)." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 517, + 479, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 517, + 479, + 542 + ], + "spans": [ + { + "bbox": [ + 131, + 517, + 479, + 542 + ], + "type": "text", + "content": "Theorem 1 ([2,14,15]). If " + }, + { + "bbox": [ + 131, + 517, + 479, + 542 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 131, + 517, + 479, + 542 + ], + "type": "text", + "content": " induces a general " + }, + { + "bbox": [ + 131, + 517, + 479, + 542 + ], + "type": "inline_equation", + "content": "VD" + }, + { + "bbox": [ + 131, + 517, + 479, + 542 + ], + "type": "text", + "content": ", then the " + }, + { + "bbox": [ + 131, + 517, + 479, + 542 + ], + "type": "inline_equation", + "content": "VD" + }, + { + "bbox": [ + 131, + 517, + 479, + 542 + ], + "type": "text", + "content": " satisfies the closed ball property if and only if its dual " + }, + { + "bbox": [ + 131, + 517, + 479, + 542 + ], + "type": "inline_equation", + "content": "IDT" + }, + { + "bbox": [ + 131, + 517, + 479, + 542 + ], + "type": "text", + "content": " is a proper triangulation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 558, + 202, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 202, + 571 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 202, + 571 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 582, + 479, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 479, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 479, + 605 + ], + "type": "text", + "content": "We introduce a new scalable functional map pipeline handling very dense meshes and yielding stable results independently of the input triangulation's quality." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 606, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 480, + 666 + ], + "type": "text", + "content": "Our method acts through three main steps. At first, we compute two low-resolution meshes that preserve the original metrics of the input shapes (Section 3.1). Then, we apply any existing pipeline, yielding a functional map between the low-resolution eigenspaces (Section 3.2). Finally, we extend the eigenfunctions to the original meshes, allowing us to use them with the functional map" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "text", + "content": "F. Maggioli et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "estimated in the previous step to compute the desired correspondence between them (Section 3.3). Our technique is very general and works with arbitrary manifold triangulations, including meshes with degenerate geometry, non-orientable surfaces, and multiple connected components. The manifoldness requirement does not provide an obstacle to the applicability of our algorithm, as it can be efficiently enforced using existing techniques [18]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 204, + 309, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 204, + 309, + 217 + ], + "spans": [ + { + "bbox": [ + 131, + 204, + 309, + 217 + ], + "type": "text", + "content": "3.1 Intrinsic Delaunay remeshing" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 223, + 299, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 223, + 299, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 223, + 299, + 415 + ], + "type": "text", + "content": "Front propagation and FPS. Naively computing the VD induced by a set of samples can quickly become very expensive. The exact geodesic algorithm introduced by Mitchell et al. [41] is computationally unfeasible on large meshes. Even using faster solutions, such as the heat method [8], the fast marching algorithm [25], or even the Dijkstra distance [11], searching the closest sample for each vertex is still a " + }, + { + "bbox": [ + 130, + 223, + 299, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(|V|s)" + }, + { + "bbox": [ + 130, + 223, + 299, + 415 + ], + "type": "text", + "content": " operation. Instead, we take inspiration from the front propagation method proposed by Peyre et al. [49]. We start with a decomposition of the mesh induced by a single sample " + }, + { + "bbox": [ + 130, + 223, + 299, + 415 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 130, + 223, + 299, + 415 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 247, + 480, + 380 + ], + "blocks": [ + { + "bbox": [ + 307, + 247, + 480, + 380 + ], + "lines": [ + { + "bbox": [ + 307, + 247, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 247, + 480, + 380 + ], + "type": "image", + "image_path": "88eb061469dddfd03768f683ce6015dee27a2fb20c18ea42b41575bc6d24603c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 391, + 482, + 413 + ], + "lines": [ + { + "bbox": [ + 304, + 391, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 482, + 413 + ], + "type": "text", + "content": "Fig.2: The iterative front propagation from samples inducing a geodesic VD." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 415, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 521 + ], + "type": "text", + "content": "This means that every vertex " + }, + { + "bbox": [ + 130, + 415, + 482, + 521 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 415, + 482, + 521 + ], + "type": "text", + "content": " is part of the texel " + }, + { + "bbox": [ + 130, + 415, + 482, + 521 + ], + "type": "inline_equation", + "content": "P_{1}" + }, + { + "bbox": [ + 130, + 415, + 482, + 521 + ], + "type": "text", + "content": " and has distance " + }, + { + "bbox": [ + 130, + 415, + 482, + 521 + ], + "type": "inline_equation", + "content": "D_{1}[v] = d_{\\mathcal{M}}(p_{1}, v)" + }, + { + "bbox": [ + 130, + 415, + 482, + 521 + ], + "type": "text", + "content": " from the sample set. We then assume to have a decomposition " + }, + { + "bbox": [ + 130, + 415, + 482, + 521 + ], + "type": "inline_equation", + "content": "P_{1}, \\dots, P_{k-1}" + }, + { + "bbox": [ + 130, + 415, + 482, + 521 + ], + "type": "text", + "content": " induced by samples " + }, + { + "bbox": [ + 130, + 415, + 482, + 521 + ], + "type": "inline_equation", + "content": "p_{1}, \\dots, p_{k-1}" + }, + { + "bbox": [ + 130, + 415, + 482, + 521 + ], + "type": "text", + "content": ", and a vector " + }, + { + "bbox": [ + 130, + 415, + 482, + 521 + ], + "type": "inline_equation", + "content": "D_{k-1}" + }, + { + "bbox": [ + 130, + 415, + 482, + 521 + ], + "type": "text", + "content": " storing the distance of each vertex from the sample set. When adding a sample " + }, + { + "bbox": [ + 130, + 415, + 482, + 521 + ], + "type": "inline_equation", + "content": "p_{k}" + }, + { + "bbox": [ + 130, + 415, + 482, + 521 + ], + "type": "text", + "content": " to the VD, we can exploit the fact that " + }, + { + "bbox": [ + 130, + 415, + 482, + 521 + ], + "type": "inline_equation", + "content": "D_{k} = \\min(D_{k-1}, d_{\\mathcal{M}}(p_{k}, V))" + }, + { + "bbox": [ + 130, + 415, + 482, + 521 + ], + "type": "text", + "content": ". We start a front from " + }, + { + "bbox": [ + 130, + 415, + 482, + 521 + ], + "type": "inline_equation", + "content": "p_{k}" + }, + { + "bbox": [ + 130, + 415, + 482, + 521 + ], + "type": "text", + "content": " and expand it, updating the distances and assigning vertices to texel " + }, + { + "bbox": [ + 130, + 415, + 482, + 521 + ], + "type": "inline_equation", + "content": "P_{k}" + }, + { + "bbox": [ + 130, + 415, + 482, + 521 + ], + "type": "text", + "content": " until we reach all vertices such that " + }, + { + "bbox": [ + 130, + 415, + 482, + 521 + ], + "type": "inline_equation", + "content": "D_{k-1}[v] < d_{\\mathcal{M}}(p_{k}, v)" + }, + { + "bbox": [ + 130, + 415, + 482, + 521 + ], + "type": "text", + "content": " (see Figure 2). By propagating the front with the fast marching or Dijkstra's algorithm, we can compute a geodesic VD with time " + }, + { + "bbox": [ + 130, + 415, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(|V| \\log(|V|) \\log(s))" + }, + { + "bbox": [ + 130, + 415, + 482, + 521 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 522, + 482, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 546 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 546 + ], + "type": "text", + "content": "To obtain the sample points, Peyré et al. [49] propose to use a geodesic farthest point sampling, achieving a geodesic uniform sampling of the surface." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 552, + 483, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 552, + 483, + 601 + ], + "spans": [ + { + "bbox": [ + 131, + 552, + 483, + 601 + ], + "type": "text", + "content": "Definition 3. Let " + }, + { + "bbox": [ + 131, + 552, + 483, + 601 + ], + "type": "inline_equation", + "content": "(\\mathcal{M}, d_{\\mathcal{M}})" + }, + { + "bbox": [ + 131, + 552, + 483, + 601 + ], + "type": "text", + "content": " be a 2-dimensional metric space, and let " + }, + { + "bbox": [ + 131, + 552, + 483, + 601 + ], + "type": "inline_equation", + "content": "s \\in \\mathbb{N}" + }, + { + "bbox": [ + 131, + 552, + 483, + 601 + ], + "type": "text", + "content": " be an integer. A farthest point sampling (FPS) of size " + }, + { + "bbox": [ + 131, + 552, + 483, + 601 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 131, + 552, + 483, + 601 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 131, + 552, + 483, + 601 + ], + "type": "inline_equation", + "content": "d_{\\mathcal{M}}" + }, + { + "bbox": [ + 131, + 552, + 483, + 601 + ], + "type": "text", + "content": " is a set " + }, + { + "bbox": [ + 131, + 552, + 483, + 601 + ], + "type": "inline_equation", + "content": "S_s \\subset \\mathcal{M}" + }, + { + "bbox": [ + 131, + 552, + 483, + 601 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 131, + 552, + 483, + 601 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 131, + 552, + 483, + 601 + ], + "type": "text", + "content": " sample points in " + }, + { + "bbox": [ + 131, + 552, + 483, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 131, + 552, + 483, + 601 + ], + "type": "text", + "content": ", incrementally built from a singleton set " + }, + { + "bbox": [ + 131, + 552, + 483, + 601 + ], + "type": "inline_equation", + "content": "S_1 = \\{p_1\\} \\subset \\mathcal{M}" + }, + { + "bbox": [ + 131, + 552, + 483, + 601 + ], + "type": "text", + "content": " according to the following rule:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 217, + 607, + 481, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 607, + 481, + 635 + ], + "spans": [ + { + "bbox": [ + 217, + 607, + 481, + 635 + ], + "type": "interline_equation", + "content": "S _ {\\ell + 1} = S _ {\\ell} \\cup \\left\\{\\underset {q \\in \\mathcal {M}} {\\arg \\max } \\quad \\min _ {p \\in S _ {\\ell}} d _ {\\mathcal {M}} (q, p) \\right\\}. \\tag {2}", + "image_path": "c2717dcda3e0f02af5c94ebcb9a5d11eda739b274af32e369a04fc150cc0ebd2.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "However, searching for the maximum distance in Equation (2) at every new sample brings again the algorithm to a " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(|V|s)" + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": " complexity, really becoming a" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 264, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 264, + 91, + 448, + 102 + ], + "type": "text", + "content": "ReMatching: Scalable Shape Correspondence" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "bottleneck for large meshes counting millions of vertices. In contrast, we propose to introduce a fixed-size binary max-heap in the front propagation algorithm, keeping track of the distances from each vertex to the sample set. During each front propagation, this heap updates the distance of each visited vertex in time " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(\\log(|V|))" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": ", adding no extra complexity to the visit. Furthermore, it can gather and set to zero the farthest vertex at each iteration in just " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(\\log(|V|))" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " time, totalling a time complexity of " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(|V|\\log(|V|)\\log(s))" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " for the entire procedure. The fast front propagation algorithm is summarized in Algorithm 1." + } + ] + } + ], + "index": 2 + }, + { + "type": "code", + "bbox": [ + 132, + 287, + 342, + 456 + ], + "blocks": [ + { + "bbox": [ + 132, + 272, + 315, + 285 + ], + "lines": [ + { + "bbox": [ + 132, + 272, + 315, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 272, + 315, + 285 + ], + "type": "text", + "content": "Algorithm 1 Geodesic FPS and its VD." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "lines": [ + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "text", + "content": "1: procedure VORONOIFPS " + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "inline_equation", + "content": "(\\mathcal{M} = (V,E,T),s)" + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "text", + "content": " \n2: " + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "inline_equation", + "content": "i\\gets" + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "text", + "content": " random index in " + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "inline_equation", + "content": "[1,|V|]" + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "text", + "content": " \n3: " + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "inline_equation", + "content": "S\\gets \\{i\\}" + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "text", + "content": " \n4: " + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "inline_equation", + "content": "D\\gets d_{\\mathcal{M}}(V,V_i)" + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "text", + "content": " \n5: " + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "inline_equation", + "content": "H\\gets" + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "text", + "content": " max-heap initialized with " + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "text", + "content": " \n6: " + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "inline_equation", + "content": "P\\gets" + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "text", + "content": " vector of length " + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "inline_equation", + "content": "|V|" + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "text", + "content": " initialized to " + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "text", + "content": " \n7: for " + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "inline_equation", + "content": "h\\gets 2" + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "text", + "content": " do \n8: " + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "inline_equation", + "content": "p\\gets \\mathrm{FINDMAX}(H)" + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "text", + "content": " \n9: SETKEY " + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "inline_equation", + "content": "(H,p,0)" + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "text", + "content": " \n10: " + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "inline_equation", + "content": "D_p\\gets 0" + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "text", + "content": " \n11: " + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "inline_equation", + "content": "S\\gets S\\cup \\{p\\}" + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "text", + "content": " \n12: Propagate a front from " + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "text", + "content": " \n13: Update " + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "inline_equation", + "content": "D,P," + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 132, + 287, + 342, + 456 + ], + "type": "text", + "content": " \n14: end for \n15: end procedure" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "code_body" + } + ], + "index": 4, + "sub_type": "algorithm" + }, + { + "bbox": [ + 348, + 224, + 482, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 224, + 482, + 463 + ], + "spans": [ + { + "bbox": [ + 348, + 224, + 482, + 463 + ], + "type": "text", + "content": "Flat union property. Even if we are able to compute the VD of a farthest point sampling with high efficiency, we still cannot guarantee that the dual connectivity is a proper IDT. Ensuring the closed ball property can be computationally challenging. In particular, identifying the topology of the boundary between each pair of adjacent texels can become very costly when the number of samples is large. A possibility is to add enough samples so that the surface locally behaves like a plane. This approach, proposed by Leibon et al. [27] and adopted by Peyré" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 463, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 482, + 511 + ], + "type": "text", + "content": "et al. [49], makes it easy to lose control over the number of samples, and anyway the requirements are not easy to enforce. For guaranteeing a proper IDT, while still avoiding these issues, we introduce a novel property for a VD. The proofs of our claims are provided in the supplementary materials." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "type": "text", + "content": "Definition 4 (Flat Union Property (FUP)). Let " + }, + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "type": "inline_equation", + "content": "(\\mathcal{M}, d_{\\mathcal{M}})" + }, + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "type": "text", + "content": " be a 2-dimensional metric space. Let then " + }, + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "type": "inline_equation", + "content": "S = \\{p_i\\}_{i=1}^s \\subset \\mathcal{M}" + }, + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "type": "text", + "content": " be a set of " + }, + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "type": "text", + "content": " sample points in " + }, + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "type": "text", + "content": ", inducing a general " + }, + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "type": "inline_equation", + "content": "VD\\{P_i\\}_{i=1}^s" + }, + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "type": "inline_equation", + "content": "VD" + }, + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "type": "text", + "content": " induced by " + }, + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 517, + 482, + 555 + ], + "type": "text", + "content": " satisfies the FUP if:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 554, + 433, + 590 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 138, + 554, + 267, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 554, + 267, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 554, + 267, + 566 + ], + "type": "text", + "content": "- every " + }, + { + "bbox": [ + 138, + 554, + 267, + 566 + ], + "type": "inline_equation", + "content": "P_i" + }, + { + "bbox": [ + 138, + 554, + 267, + 566 + ], + "type": "text", + "content": " is a closed 2-ball;" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 566, + 391, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 566, + 391, + 578 + ], + "spans": [ + { + "bbox": [ + 137, + 566, + 391, + 578 + ], + "type": "text", + "content": "- if " + }, + { + "bbox": [ + 137, + 566, + 391, + 578 + ], + "type": "inline_equation", + "content": "P_{i} \\cap P_{j}" + }, + { + "bbox": [ + 137, + 566, + 391, + 578 + ], + "type": "text", + "content": " is not empty, then " + }, + { + "bbox": [ + 137, + 566, + 391, + 578 + ], + "type": "inline_equation", + "content": "P_{i} \\cup P_{j}" + }, + { + "bbox": [ + 137, + 566, + 391, + 578 + ], + "type": "text", + "content": " is a closed 2-ball;" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 578, + 433, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 578, + 433, + 590 + ], + "spans": [ + { + "bbox": [ + 137, + 578, + 433, + 590 + ], + "type": "text", + "content": "- if " + }, + { + "bbox": [ + 137, + 578, + 433, + 590 + ], + "type": "inline_equation", + "content": "P_{i} \\cap P_{j} \\cap P_{k}" + }, + { + "bbox": [ + 137, + 578, + 433, + 590 + ], + "type": "text", + "content": " is not empty, then " + }, + { + "bbox": [ + 137, + 578, + 433, + 590 + ], + "type": "inline_equation", + "content": "P_{i} \\cup P_{j} \\cup P_{k}" + }, + { + "bbox": [ + 137, + 578, + 433, + 590 + ], + "type": "text", + "content": " is a closed 2-ball." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 597, + 482, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 597, + 482, + 623 + ], + "spans": [ + { + "bbox": [ + 130, + 597, + 482, + 623 + ], + "type": "text", + "content": "Theorem 2. Let " + }, + { + "bbox": [ + 130, + 597, + 482, + 623 + ], + "type": "inline_equation", + "content": "(\\mathcal{M}, d_{\\mathcal{M}})" + }, + { + "bbox": [ + 130, + 597, + 482, + 623 + ], + "type": "text", + "content": " be a 2-dimensional metric space. If a general " + }, + { + "bbox": [ + 130, + 597, + 482, + 623 + ], + "type": "inline_equation", + "content": "VD\\{P_i\\}_{i=1}^s" + }, + { + "bbox": [ + 130, + 597, + 482, + 623 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 597, + 482, + 623 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 597, + 482, + 623 + ], + "type": "text", + "content": " satisfies the FUP, it also satisfies the closed ball property." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "The FUP is much easier to verify than the closed ball property, as we only have to ensure that regions are closed 2-balls. Indeed, the following property can be exploited in this regard." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "text", + "content": "F. Maggioli et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "text", + "content": "Proposition 1. Let " + }, + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = (V,E,F)" + }, + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "text", + "content": " be a manifold polygonal mesh, then " + }, + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "text", + "content": " is a closed 2-ball if and only if its Euler characteristic " + }, + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "inline_equation", + "content": "\\chi = |V| - |E| + |F|" + }, + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "text", + "content": " is 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 151, + 479, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 151, + 479, + 210 + ], + "spans": [ + { + "bbox": [ + 130, + 151, + 479, + 210 + ], + "type": "text", + "content": "Dual mesh representation. When we compute a VD of a triangle mesh, we want to partition the vertices into disjoint sets. However, Voronoi texels are defined as regions over the surface, meaning that they should be submanifolds of " + }, + { + "bbox": [ + 130, + 151, + 479, + 210 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 151, + 479, + 210 + ], + "type": "text", + "content": ". By only considering vertices, we likely end up having non-manifold geometries or non well-defined boundaries." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 149, + 236, + 293, + 437 + ], + "blocks": [ + { + "bbox": [ + 149, + 236, + 293, + 437 + ], + "lines": [ + { + "bbox": [ + 149, + 236, + 293, + 437 + ], + "spans": [ + { + "bbox": [ + 149, + 236, + 293, + 437 + ], + "type": "image", + "image_path": "52650bede014b0df04c1a9169bff639112145d3a025fa4bab899d031a2a522e4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 448, + 309, + 504 + ], + "lines": [ + { + "bbox": [ + 130, + 448, + 309, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 309, + 504 + ], + "type": "text", + "content": "Fig. 3: Top: the primal and dual connectivities of a triangular mesh. Bottom: regions on the primal mesh are represented as sets of vertices, and the corresponding regions on the dual mesh are composed by faces." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 211, + 482, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 482, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 482, + 366 + ], + "type": "text", + "content": "We consider the polygonal mesh " + }, + { + "bbox": [ + 313, + 211, + 482, + 366 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{M}} = (\\tilde{V},\\tilde{E},\\tilde{T})" + }, + { + "bbox": [ + 313, + 211, + 482, + 366 + ], + "type": "text", + "content": " to be the dual mesh of " + }, + { + "bbox": [ + 313, + 211, + 482, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = (V,E,T)" + }, + { + "bbox": [ + 313, + 211, + 482, + 366 + ], + "type": "text", + "content": ", built by placing a vertex at each face of " + }, + { + "bbox": [ + 313, + 211, + 482, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 211, + 482, + 366 + ], + "type": "text", + "content": " and forming a polygonal face for each triangle fan around a vertex of " + }, + { + "bbox": [ + 313, + 211, + 482, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 211, + 482, + 366 + ], + "type": "text", + "content": ". As shown in the example from Figure 3, dual meshes allow us to define Voronoi texels by means of connected dual faces, making them actual submanifolds of the original surface. Furthermore, we can easily define boundaries between texels as paths made of dual edges." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 366, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 482, + 472 + ], + "type": "text", + "content": "This construction has the positive side effect to ensure that every VD is general. Indeed, the only place where three or more texels can meet is a dual vertex. Since texels are made of dual faces (i.e., primal vertices), and the primal mesh is triangular, there is no dual vertex where more than three faces can meet." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 474, + 481, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 481, + 509 + ], + "type": "text", + "content": "Given the correspondence between primal and dual geometric elements, we can verify if a region is a closed 2-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 582 + ], + "type": "text", + "content": "ball with Proposition 1 without explicitly constructing the dual mesh. Additional details are provided in the supplementary material. After the initial FPS, we further add samples to enforce the flat union property: if two or more adjacent texels do not form a closed 2-ball, we break the connection adding samples at the Voronoi vertex or along the Voronoi edge; if a texel is not a closed 2-ball, we add a sample along its boundary to reduce its coverage." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 598, + 285, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 285, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 285, + 611 + ], + "type": "text", + "content": "3.2 Low-resolution matching" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": " be two triangular meshes, and let " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}}" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{N}}" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": " be their low-resolution counterparts computed with the algorithm presented above (or any other alternative algorithm). Furthermore, let " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\Phi}" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": " be a basis for the functional space over " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}}" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\Psi}" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": " a basis for the functional space over " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{N}}" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": ". For example these bases can" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 264, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 264, + 91, + 447, + 102 + ], + "type": "text", + "content": "ReMatching: Scalable Shape Correspondence" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "be the truncated subset of the eigenvectors of the LBO as proposed in [47]. We assume to have some pipeline yielding a functional map " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": ", such that" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 272, + 148, + 481, + 163 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 148, + 481, + 163 + ], + "spans": [ + { + "bbox": [ + 272, + 148, + 481, + 163 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {\\Psi}} \\mathbf {C} \\approx \\hat {\\boldsymbol {\\Pi}} \\hat {\\boldsymbol {\\Phi}}, \\tag {3}", + "image_path": "9d0464b1ad64f97f7bb4956faaf9775c6453c8d60e5a8a83e57ad2394e965e68.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 171, + 480, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 171, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 130, + 171, + 480, + 292 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 171, + 480, + 292 + ], + "type": "inline_equation", + "content": "\\hat{\\Pi}:\\hat{\\mathcal{M}}\\to \\hat{\\mathcal{N}}" + }, + { + "bbox": [ + 130, + 171, + 480, + 292 + ], + "type": "text", + "content": " is a correspondence between the vertices of " + }, + { + "bbox": [ + 130, + 171, + 480, + 292 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}}" + }, + { + "bbox": [ + 130, + 171, + 480, + 292 + ], + "type": "text", + "content": " and the vertices of " + }, + { + "bbox": [ + 130, + 171, + 480, + 292 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{N}}" + }, + { + "bbox": [ + 130, + 171, + 480, + 292 + ], + "type": "text", + "content": ". The adopted functional maps solution can be any of the available alternatives, ranging from the original one [47] to the most recent [12]. In our experiments, we consider two of the widely adopted solutions: constrained optimization with product preservation [46] and the iterative procedure ZoomOut [38]. The first is representative of the possible optimization strategies, while the second is probably the most efficient refinement technique for functional maps estimation. Both these methods have been efficiently and effectively applied on meshes with a limited number of vertices (i.e., up to 10 thousands vertices), which is exactly the setting we are in after the proposed remeshing step." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 310, + 313, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 310, + 313, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 313, + 322 + ], + "type": "text", + "content": "3.3 Extending the correspondence" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 331, + 264, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 264, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 264, + 569 + ], + "type": "text", + "content": "To extend the basis from the remeshing to the original surface, we need a mapping that transports scalar fields from the low-resolution mesh " + }, + { + "bbox": [ + 130, + 331, + 264, + 569 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}}" + }, + { + "bbox": [ + 130, + 331, + 264, + 569 + ], + "type": "text", + "content": " to the high-resolution shape " + }, + { + "bbox": [ + 130, + 331, + 264, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 331, + 264, + 569 + ], + "type": "text", + "content": ". A possible solution would be to map each triangle of " + }, + { + "bbox": [ + 130, + 331, + 264, + 569 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}}" + }, + { + "bbox": [ + 130, + 331, + 264, + 569 + ], + "type": "text", + "content": " to a geodesic triangle in " + }, + { + "bbox": [ + 130, + 331, + 264, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 331, + 264, + 569 + ], + "type": "text", + "content": ", and then computing the geodesic barycentric coordinates of each vertex inside the triangle [31]. Unfortunately, this approach is very costly, as it requires to compute exact geodesic paths onto " + }, + { + "bbox": [ + 130, + 331, + 264, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 331, + 264, + 569 + ], + "type": "text", + "content": " [56]. Instead, we approximate this mapping by projecting each vertex of " + }, + { + "bbox": [ + 130, + 331, + 264, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 331, + 264, + 569 + ], + "type": "text", + "content": " onto the closest surface point of " + }, + { + "bbox": [ + 130, + 331, + 264, + 569 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}}" + }, + { + "bbox": [ + 130, + 331, + 264, + 569 + ], + "type": "text", + "content": ". Despite" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 271, + 354, + 482, + 506 + ], + "blocks": [ + { + "bbox": [ + 271, + 354, + 482, + 506 + ], + "lines": [ + { + "bbox": [ + 271, + 354, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 271, + 354, + 482, + 506 + ], + "type": "image", + "image_path": "5d3f9704c11ded32c90244c4f873dbea69cd461210c21b679871337bdfecef2f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 269, + 512, + 482, + 567 + ], + "lines": [ + { + "bbox": [ + 269, + 512, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 269, + 512, + 482, + 567 + ], + "type": "text", + "content": "Fig. 4: Top: remeshing and mapping of the original vertices with the solution from Liu et al. [31] and our pipeline. Bottom: comparison of ground truth geodesic distances from multiple vertices and extended with both methods." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 479, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 479, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 479, + 617 + ], + "type": "text", + "content": "this approach not being accurate in general, every triangle in " + }, + { + "bbox": [ + 130, + 570, + 479, + 617 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}}" + }, + { + "bbox": [ + 130, + 570, + 479, + 617 + ], + "type": "text", + "content": " corresponds to a geodesic triangle (homeomorphic to a disk) on " + }, + { + "bbox": [ + 130, + 570, + 479, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 570, + 479, + 617 + ], + "type": "text", + "content": " with the same vertices. Furthermore, starting from a farthest point sampling gives evenly spaced samples, mitigating the geometric complexity of geodesic triangles." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "type": "text", + "content": "A surface point of " + }, + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}}" + }, + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "type": "text", + "content": " is either a vertex, or a weighted average of two (if on an edge) or three vertices (if on a triangle). We build a linear map " + }, + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{\\mathcal{M}} \\in \\mathbb{R}^{m \\times s}" + }, + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "type": "text", + "content": " representing the projection of each vertex " + }, + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "type": "inline_equation", + "content": "v_i \\in \\mathcal{M}" + }, + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "type": "text", + "content": " as a linear combination of at most three vertices in " + }, + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}}" + }, + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "type": "text", + "content": " (i.e., the vertices encoding the closest surface point)." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "text", + "content": "F. Maggioli et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": "Figure 4 shows a comparison between our pipeline and the simplification method proposed by Liu et al. [31]. The top row shows the different geometries produced by the two methods in reducing the mesh size by " + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": " and how our mapping compares to the approximate geodesic barycentric coordinates. In the bottom row we compare the two methods in approximating a multiple source geodesic distance field, showing that they do not present appreciable differences. More discussion is provided in the supplementary materials." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 479, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 479, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 479, + 247 + ], + "type": "text", + "content": "Scalar functions can be evaluated at any surface point by interpolating the values at the vertices. Since every vertex of " + }, + { + "bbox": [ + 130, + 200, + 479, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 200, + 479, + 247 + ], + "type": "text", + "content": " is associated to a point on the surface of " + }, + { + "bbox": [ + 130, + 200, + 479, + 247 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}}" + }, + { + "bbox": [ + 130, + 200, + 479, + 247 + ], + "type": "text", + "content": ", we can use the linear map " + }, + { + "bbox": [ + 130, + 200, + 479, + 247 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{\\mathcal{M}}" + }, + { + "bbox": [ + 130, + 200, + 479, + 247 + ], + "type": "text", + "content": " for extending scalar functions from the " + }, + { + "bbox": [ + 130, + 200, + 479, + 247 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 200, + 479, + 247 + ], + "type": "text", + "content": " vertices of " + }, + { + "bbox": [ + 130, + 200, + 479, + 247 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}}" + }, + { + "bbox": [ + 130, + 200, + 479, + 247 + ], + "type": "text", + "content": " to the " + }, + { + "bbox": [ + 130, + 200, + 479, + 247 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 200, + 479, + 247 + ], + "type": "text", + "content": " vertices of " + }, + { + "bbox": [ + 130, + 200, + 479, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 200, + 479, + 247 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 247, + 479, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 479, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 479, + 283 + ], + "type": "text", + "content": "We extend the bases " + }, + { + "bbox": [ + 130, + 247, + 479, + 283 + ], + "type": "inline_equation", + "content": "\\hat{\\Phi}" + }, + { + "bbox": [ + 130, + 247, + 479, + 283 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 247, + 479, + 283 + ], + "type": "inline_equation", + "content": "\\hat{\\Psi}" + }, + { + "bbox": [ + 130, + 247, + 479, + 283 + ], + "type": "text", + "content": " to the full-resolution meshes as " + }, + { + "bbox": [ + 130, + 247, + 479, + 283 + ], + "type": "inline_equation", + "content": "\\Phi = \\mathbf{U}_{\\mathcal{M}}\\hat{\\Phi}" + }, + { + "bbox": [ + 130, + 247, + 479, + 283 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 247, + 479, + 283 + ], + "type": "inline_equation", + "content": "\\Psi = \\mathbf{U}_{\\mathcal{N}}\\hat{\\Psi}" + }, + { + "bbox": [ + 130, + 247, + 479, + 283 + ], + "type": "text", + "content": ". Then, in a similar fashion as done in Equation (3), we search for a correspondence " + }, + { + "bbox": [ + 130, + 247, + 479, + 283 + ], + "type": "inline_equation", + "content": "\\Pi : \\mathcal{M} \\to \\mathcal{N}" + }, + { + "bbox": [ + 130, + 247, + 479, + 283 + ], + "type": "text", + "content": " that maps vertices of " + }, + { + "bbox": [ + 130, + 247, + 479, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 247, + 479, + 283 + ], + "type": "text", + "content": " to vertices of " + }, + { + "bbox": [ + 130, + 247, + 479, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 130, + 247, + 479, + 283 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 212, + 290, + 480, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 290, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 212, + 290, + 480, + 304 + ], + "type": "interline_equation", + "content": "\\mathbf {U} _ {\\mathcal {N}} \\hat {\\boldsymbol {\\Psi}} \\mathbf {C} = \\boldsymbol {\\Psi} \\mathbf {C} \\approx \\boldsymbol {\\Pi} \\boldsymbol {\\Phi} = \\boldsymbol {\\Pi} \\mathbf {U} _ {\\mathcal {M}} \\hat {\\boldsymbol {\\Phi}}, \\tag {4}", + "image_path": "d313a3df6d346d61d4b16fd3f1c35a708bd70f2e0141ec694528586937bfd9c2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 310, + 479, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 479, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 479, + 357 + ], + "type": "text", + "content": "where, for the sake of clarity, we explicitly write the equation in terms of the bases " + }, + { + "bbox": [ + 130, + 310, + 479, + 357 + ], + "type": "inline_equation", + "content": "\\hat{\\Phi}" + }, + { + "bbox": [ + 130, + 310, + 479, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 310, + 479, + 357 + ], + "type": "inline_equation", + "content": "\\hat{\\Psi}" + }, + { + "bbox": [ + 130, + 310, + 479, + 357 + ], + "type": "text", + "content": ". Given the extended bases " + }, + { + "bbox": [ + 130, + 310, + 479, + 357 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 130, + 310, + 479, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 310, + 479, + 357 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 130, + 310, + 479, + 357 + ], + "type": "text", + "content": ", and the same functional map " + }, + { + "bbox": [ + 130, + 310, + 479, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 130, + 310, + 479, + 357 + ], + "type": "text", + "content": " estimated to solve Equation (3), we obtain the correspondence " + }, + { + "bbox": [ + 130, + 310, + 479, + 357 + ], + "type": "inline_equation", + "content": "\\Pi" + }, + { + "bbox": [ + 130, + 310, + 479, + 357 + ], + "type": "text", + "content": " via a nearest neighbor search in the spectral space." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 358, + 480, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 480, + 430 + ], + "type": "text", + "content": "We notice that nearest neighbor algorithms become very slow on large meshes, effectively becoming a possible bottleneck for the pipeline. Our experiments presented in Figure 10 show that, to the scope of this paper, this solution provides satisfying performance. However, we acknowledge that establishing a relationship between (3) and (4) could lead to a more efficient expression of " + }, + { + "bbox": [ + 130, + 358, + 480, + 430 + ], + "type": "inline_equation", + "content": "\\Pi" + }, + { + "bbox": [ + 130, + 358, + 480, + 430 + ], + "type": "text", + "content": " in terms of " + }, + { + "bbox": [ + 130, + 358, + 480, + 430 + ], + "type": "inline_equation", + "content": "\\hat{\\Pi}" + }, + { + "bbox": [ + 130, + 358, + 480, + 430 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 358, + 480, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{\\mathcal{M}}" + }, + { + "bbox": [ + 130, + 358, + 480, + 430 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 358, + 480, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{\\mathcal{N}}" + }, + { + "bbox": [ + 130, + 358, + 480, + 430 + ], + "type": "text", + "content": ". This avenue warrants future exploration." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 446, + 198, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 198, + 458 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 198, + 458 + ], + "type": "text", + "content": "4 Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 470, + 479, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 479, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 479, + 540 + ], + "type": "text", + "content": "Our goal is to evaluate not only the matching accuracy of our method but also its time performance. For this task, we test it on datasets for dense correspondences, such as the SHREC19 dataset [37] and the TOSCA dataset [6]. We also introduce a novel dataset, BadTOSCA, obtained by randomly altering the vertex positions of the TOSCA meshes, to ensure that our method is stable even under circumstances where strong isometry cannot be assumed3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 541, + 479, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 541, + 479, + 626 + ], + "spans": [ + { + "bbox": [ + 130, + 541, + 479, + 626 + ], + "type": "text", + "content": "Our pipeline relies on an efficient remeshing algorithm that can quickly and drastically reduce the size of a mesh by orders of magnitude while still preserving the underlying geometry and metric. In principle, any remeshing algorithm could be adopted in our pipeline, as the map for extending scalar fields discussed in Section 3.3 only requires having two input meshes. We compare our solution against well-established isotropic and anisotropic remeshing algorithms (respectively, IRM [4,21] and ARM [44]), as well as the scalable functional maps (SFM)" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 264, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 264, + 91, + 447, + 102 + ], + "type": "text", + "content": "ReMatching: Scalable Shape Correspondence" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 631, + 479, + 663 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 631, + 479, + 663 + ], + "spans": [ + { + "bbox": [ + 133, + 631, + 479, + 663 + ], + "type": "text", + "content": "3 Details on the dataset generation are provided in the supplementary materials. Code for generating the dataset is available at https://github.com/filthynobleman/bad-tosca" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 131, + 114, + 307, + 223 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 307, + 223 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 307, + 223 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 307, + 223 + ], + "type": "image", + "image_path": "0df69ac21962c60babee3f06b84dd9565e8bf434c4bbad281971c3f671f993f2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 230, + 482, + 285 + ], + "lines": [ + { + "bbox": [ + 130, + 230, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 230, + 482, + 285 + ], + "type": "text", + "content": "Fig. 5: Left: Accuracy curves on the SHREC19 challenge pairs for the tested methods. We also consider ideal approaches where the initial functional map for ZoomOut is given ground truth dashed lines). Right: Cumulative curves of the execution time for evaluated methods. For all methods, we also consider the time required to run the remeshing or resampling step(s)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 113, + 481, + 222 + ], + "blocks": [ + { + "bbox": [ + 307, + 113, + 481, + 222 + ], + "lines": [ + { + "bbox": [ + 307, + 113, + 481, + 222 + ], + "spans": [ + { + "bbox": [ + 307, + 113, + 481, + 222 + ], + "type": "image", + "image_path": "8a480c5dff1e07ec217422620535403a8112a65022b65da28812005a941a0a73.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 309, + 480, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 480, + 332 + ], + "type": "text", + "content": "approach proposed by Magnet et al. [35], which produces a subsampling of the vertices in place of a low-resolution mesh." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 333, + 480, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 480, + 357 + ], + "type": "text", + "content": "We implemented our remeshing algorithm in " + }, + { + "bbox": [ + 130, + 333, + 480, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{C}++" + }, + { + "bbox": [ + 130, + 333, + 480, + 357 + ], + "type": "text", + "content": ", using Eigen [20] and libigl [22], and the matching pipeline in MATLAB4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 374, + 261, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 374, + 261, + 387 + ], + "spans": [ + { + "bbox": [ + 131, + 374, + 261, + 387 + ], + "type": "text", + "content": "4.1 Quality of matching" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 395, + 284, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 395, + 284, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 395, + 284, + 538 + ], + "type": "text", + "content": "To evaluate our shape matching pipeline, we tested our method on the SHREC19 dataset [37]. We can use different shape correspondence techniques to test the entire pipeline. In particular, we compute the functional map on the low-resolution representation using both the widely adopted approach with products preservation (FMaps) [46]" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 539, + 264, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 539, + 264, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 539, + 264, + 609 + ], + "type": "text", + "content": "and ZoomOut [38]. We compare our method with the same approaches on the fullresolution meshes, as well as against SFM. For a fair comparison, we remesh all the" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 265, + 418, + 470, + 531 + ], + "blocks": [ + { + "bbox": [ + 265, + 418, + 470, + 531 + ], + "lines": [ + { + "bbox": [ + 265, + 418, + 470, + 531 + ], + "spans": [ + { + "bbox": [ + 265, + 418, + 470, + 531 + ], + "type": "table", + "html": "
MethodAGE (·10-2) ↓AUC ↑
Ours (ZoomOut)*2.4095.12%
SFM*2.4695.22%
Ours (FMaps)6.2388.21%
Ours (ZoomOut)5.8788.82%
SFM16.8470.19%
ARM (ZoomOut)41.5930.34%
IRM (ZoomOut)8.3084.62%
FMaps28.3156.15%
ZoomOut29.3652.02%
", + "image_path": "88c8fc2db9eadb7e6a18f4e3ffd8f2fcd6aec6f9100f44c57bd5b2667c6a8820.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 269, + 540, + 482, + 606 + ], + "lines": [ + { + "bbox": [ + 269, + 540, + 482, + 606 + ], + "spans": [ + { + "bbox": [ + 269, + 540, + 482, + 606 + ], + "type": "text", + "content": "Table 1: Average geodesic error and area under the accuracy curve for each tested method on the SHREC19 challenge pairs. The top rows (denoted by \\*) show the performance of our algorithm and SFM under the assumption of having a ground truth functional map for initializing ZoomOut." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 610, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 610, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 610, + 482, + 647 + ], + "type": "text", + "content": "shapes to 3k vertices with all methods, as Magnet et al. [35] state that this produces the best performance with SFM. To further ease the alignment of the Laplacian spectra, we also post-process the remeshed shapes (produced with" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "text", + "content": "F. Maggioli et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 425, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 425, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 425, + 666 + ], + "type": "text", + "content": "4 Code available at https://github.com/filthynobleman/rematching" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 113, + 309, + 194 + ], + "blocks": [ + { + "bbox": [ + 132, + 113, + 309, + 194 + ], + "lines": [ + { + "bbox": [ + 132, + 113, + 309, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 113, + 309, + 194 + ], + "type": "image", + "image_path": "9264c224dec92854c5c1f2018de35375a4fa180ae419ad4d6e8e9900a5b4c4ff.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 202, + 480, + 224 + ], + "lines": [ + { + "bbox": [ + 131, + 202, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 131, + 202, + 480, + 224 + ], + "type": "text", + "content": "Fig. 6: Accuracy curves and average geodesic error on the TOSCA dataset (left) and the BadTOSCA dataset (right) for the tested methods." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 310, + 113, + 481, + 194 + ], + "blocks": [ + { + "bbox": [ + 310, + 113, + 481, + 194 + ], + "lines": [ + { + "bbox": [ + 310, + 113, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 310, + 113, + 481, + 194 + ], + "type": "image", + "image_path": "1a2a8b2fd554caf4414acdf9387979209afd2081b10d7cc3bb3e12c9df6511b4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 252, + 479, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 252, + 479, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 252, + 479, + 312 + ], + "type": "text", + "content": "IRM, ARM, and our algorithm) to remove small disconnected components made of few triangles before computing the extension map discussed in Section 3.3. This post-processing step cannot be performed for SFM, due to the nature of its sampling strategy. The benefits of this post-processing will be discussed in Section 4.3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 314, + 480, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 480, + 398 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 480, + 398 + ], + "type": "text", + "content": "We summarize the results of our experiment in Figure 5, where we show the accuracy curves for all methods on the SHREC19 connectivity track benchmark, following the paradigm proposed by Kim et al. [24]. To prove the time efficiency of our pipeline, we also measure the time taken by every method on each pair of shapes in the benchmark (including the remeshing step), and we show the cumulative curves of the execution times over the entire dataset in a similar way with respect to the accuracy curves." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 134, + 425, + 307, + 509 + ], + "blocks": [ + { + "bbox": [ + 134, + 425, + 307, + 509 + ], + "lines": [ + { + "bbox": [ + 134, + 425, + 307, + 509 + ], + "spans": [ + { + "bbox": [ + 134, + 425, + 307, + 509 + ], + "type": "image", + "image_path": "2133191e7d4ccc87aed022750fe3f300b5436d5204dc336a9b4b1b75c80e2205.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 519, + 308, + 573 + ], + "lines": [ + { + "bbox": [ + 131, + 519, + 308, + 573 + ], + "spans": [ + { + "bbox": [ + 131, + 519, + 308, + 573 + ], + "type": "text", + "content": "Fig.7: Coordinate transfer between nonisometric shapes. The coordinates are used to generate highly complex patterns, whose sensitivity to the input enhances the transfer errors." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 400, + 481, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 481, + 579 + ], + "type": "text", + "content": "We see that all the low-resolution approaches provide better time and quality performance than the fullresolution methods, as the simpler geometry makes it easier and faster to align the Laplacian eigenfunctions. ARM is the only exception, since the anisotropic remeshing produces many skew triangles and a singular Laplacian matrix for all the shapes. Furthermore, our technique (with both the FMaps and ZoomOut backend) outperforms SFM and achieves better results than using the IRM remeshing strategy. This difference can be better" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 580, + 480, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 580, + 480, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 580, + 480, + 639 + ], + "type": "text", + "content": "appreciated in Table 1, where we report the average geodesic error (normalized with respect to the shape diameter) and the area under the accuracy curves. In both Figure 5 and Table 1 we have also considered the idealized scenario where our method with ZoomOut as backend and SFM are initialized with a ground truth functional map. In this case, the two approaches obtain comparable results." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "type": "text", + "content": "We also compare our method with the other approaches on the TOSCA and BadTOSCA datasets, summarizing the results in Figure 6. In the figure, we show" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 264, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 264, + 91, + 447, + 102 + ], + "type": "text", + "content": "ReMatching: Scalable Shape Correspondence" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 116, + 301, + 196 + ], + "blocks": [ + { + "bbox": [ + 138, + 116, + 301, + 196 + ], + "lines": [ + { + "bbox": [ + 138, + 116, + 301, + 196 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 301, + 196 + ], + "type": "image", + "image_path": "f9949d10a74651ebcfc43cb1804d82f9380865b255c15d7349e1ccfe06d8d654.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 205, + 303, + 259 + ], + "lines": [ + { + "bbox": [ + 135, + 205, + 303, + 259 + ], + "spans": [ + { + "bbox": [ + 135, + 205, + 303, + 259 + ], + "type": "text", + "content": "Fig.8: Inter-class coordinate transfer from the wolf model to a cat and a dog. Our pipeline is not constrained by the standard Laplacian basis and can be used with other approaches as well." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 310, + 114, + 477, + 196 + ], + "blocks": [ + { + "bbox": [ + 310, + 114, + 477, + 196 + ], + "lines": [ + { + "bbox": [ + 310, + 114, + 477, + 196 + ], + "spans": [ + { + "bbox": [ + 310, + 114, + 477, + 196 + ], + "type": "image", + "image_path": "2028f44b42e83515274e6cfad997c0df65f4b6cfeb0924d5cd08a9c65abfd6f4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 205, + 477, + 238 + ], + "lines": [ + { + "bbox": [ + 309, + 205, + 477, + 238 + ], + "spans": [ + { + "bbox": [ + 309, + 205, + 477, + 238 + ], + "type": "text", + "content": "Fig. 9: Comparison of function transfer between very high-resolution models with SFM and our approach." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 288, + 481, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 288, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 288, + 481, + 361 + ], + "type": "text", + "content": "the accuracy curves for the tested methods on both datasets, also providing the average geodesic error normalized by the shape diameter. We see that under the favorable and unrealistic conditions offered by TOSCA, where shapes of the same class are almost perfectly isometric, SFM achieves the best results. However, when we consider the slightly altered geometry of BadTOSCA, SFM suffers a performance drop of about " + }, + { + "bbox": [ + 130, + 288, + 481, + 361 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 288, + 481, + 361 + ], + "type": "text", + "content": ", while our method shows more stability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 361, + 482, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 482 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 482 + ], + "type": "text", + "content": "A significant benefit of the functional map approach is its independence from the difference in resolution between the source and target shape. Transferring a function with functional maps yields a much smoother and neat result than directly a point-wise correspondence. In this regard, we test our method in transferring coordinate functions between isometric and non-isometric shapes with substantial differences in resolution and triangulation. In the example from Figure 7 we generated a complex procedural fractal pattern as a function of the coordinates [33]. This pattern is very sensitive to the input, so even slight errors are highly enhanced. We see that SFM cannot precisely map the coordinates of the source mesh, shifting around and distorting the details of the pattern." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 483, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 482, + 568 + ], + "type": "text", + "content": "Furthermore, the example in Figure 8 shows that we can also deal with inter-class function transfer with our pipeline, outperforming SFM in a visual comparison. In the example, we also showcase that we should not necessarily rely on the standard Laplacian basis as a backend. Instead of using ZoomOut to transfer functions between the low-resolution meshes, we rely on the orthogonalized eigenproducts basis introduced by Maggioli et al. [34], extending it with the same technique described in Section 3.2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": "Finally, to ensure that our method can scale to high-resolution meshes, we tested it using two very high-resolution 3D scans of real statues: a dancing faun counting " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\sim 750\\mathrm{k}" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " vertices and a warrior from the temple of Aphaea counting " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\sim 3.5\\mathrm{M}" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " vertices (see Figures 1 and 9). Since SFM can also deal with dense shapes, we compare it against our method. Figure 9 shows an example of the transfer of a geodesic distance function. While our method can faithfully transfer the function, SFM produces incoherent distances (e.g., wrong isolines on the head and near the shoulder) and evident artefacts (e.g., wrong distance on the hand)." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "text", + "content": "F. Maggioli et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 268, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 268, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 268, + 128 + ], + "type": "text", + "content": "4.2 Scalable performance" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 298, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 298, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 298, + 354 + ], + "type": "text", + "content": "In Section 4.1 we saw that general-purpose remeshing solutions are not an optimal choice for the matching task. However, Lescoat et al. [28] provided a specialized mesh decimation algorithm (SMD) that reduces the complexity of the mesh while still preserving the Laplacian eigenfunctions. Furthermore, Liu et al. [31] proposed an elegant solution for simplifying a shape based on intrinsic error metric (IEM), which also builds a map for extending scalar functions using geodesic barycentric coordinates. We perform the evaluation on the SHREC19 dataset [37], remeshing the surfaces to 3k vertices with all the methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 356, + 299, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 299, + 595 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 299, + 595 + ], + "type": "text", + "content": "To achieve the desired scalability, the remeshing step in our pipeline must be fast enough to maintain the improvement obtained by reducing the size of the eigenproblem. In this regard, Figure 10 shows that our technique achieves better time performance than the other remeshing algorithms and the sampling strategy of SFM. In contrast, SMD and IEM are decimation algorithms that iteratively reduce the mesh size, making them orders of magnitude slower and unsuitable for large meshes. Moreover, SMD requires a pre-computation of the Laplacian eigenbasis on the original surface to guide the decimation process and preserve the spectrum, introducing additional time complexity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 617, + 270, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 617, + 270, + 629 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 270, + 629 + ], + "type": "text", + "content": "4.3 Benefits of remeshing" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 305, + 164, + 482, + 273 + ], + "blocks": [ + { + "bbox": [ + 305, + 164, + 482, + 273 + ], + "lines": [ + { + "bbox": [ + 305, + 164, + 482, + 273 + ], + "spans": [ + { + "bbox": [ + 305, + 164, + 482, + 273 + ], + "type": "image", + "image_path": "33fa98f4caa616ea99b1f70e11bb472b12ebdee962d007dee1cdc28d9635f51a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 281, + 482, + 347 + ], + "lines": [ + { + "bbox": [ + 304, + 281, + 482, + 347 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 482, + 347 + ], + "type": "text", + "content": "Fig.10: Execution time (logarithmic scale on the " + }, + { + "bbox": [ + 304, + 281, + 482, + 347 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 281, + 482, + 347 + ], + "type": "text", + "content": "-axis) of different remeshing algorithms plotted against the number of vertices of the input shape (" + }, + { + "bbox": [ + 304, + 281, + 482, + 347 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 281, + 482, + 347 + ], + "type": "text", + "content": "-axis). For SFM we show the time taken for generating the sampling." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 380, + 481, + 517 + ], + "blocks": [ + { + "bbox": [ + 306, + 380, + 481, + 517 + ], + "lines": [ + { + "bbox": [ + 306, + 380, + 481, + 517 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 481, + 517 + ], + "type": "image", + "image_path": "ff40afdeb38a5cc92852b474a2e754ef6bc15fa1a8e05a97edabb40fe3ada0ee.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 525, + 482, + 624 + ], + "lines": [ + { + "bbox": [ + 304, + 525, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 482, + 624 + ], + "type": "text", + "content": "Fig.11: Top row: coordinate transfer between a pair from SHREC19 using ZoomOut on full-resolution meshes, SFM and our approach. Bottom row: The ground truth " + }, + { + "bbox": [ + 304, + 525, + 482, + 624 + ], + "type": "inline_equation", + "content": "20 \\times 20" + }, + { + "bbox": [ + 304, + 525, + 482, + 624 + ], + "type": "text", + "content": " functional map, compared with the alignment computed with FMaps [46] from the full Laplacian eigenbasis, the SFM extended basis and our extended basis." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "In Section 4.1, we discussed how we remove small disconnected components after the remeshing step. The example in Figure 11 shows a comparison of mapping" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 264, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 264, + 91, + 447, + 102 + ], + "type": "text", + "content": "ReMatching: Scalable Shape Correspondence" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 333 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 333 + ], + "type": "text", + "content": "between a low-resolution mesh (Source), with less than 7k vertices, to a detailed mesh where eyes are represented with two disconnected components, each composed of 462 vertices out of the 27k of the entire mesh (" + }, + { + "bbox": [ + 130, + 116, + 482, + 333 + ], + "type": "inline_equation", + "content": "\\sim 2\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 333 + ], + "type": "text", + "content": " of the total vertices). This is enough geometry to attract energy at lower frequencies, and when we try to align the first 20 eigenfunctions with FMaps [46], the resulting functional map turns out to be mostly noise. Initializing ZoomOut with such a map leads to meaningless correspondence, and SFM does not mitigate this issue, as it tries to preserve the original spectrum as much as possible. In contrast, by removing these components and mapping them to their closest surface points, we introduce a small error in the final correspondence, but at the same time, we ease the alignment of the initial eigenbases, resulting in more accurate and meaningful mapping. This is evident in the bottom row of Figure 11, where the ground truth functional map presents a gap in the alignment of the eigenfunctions due to the impossibility of mapping the disconnected eyes of the target shape in any point of the source shape. Using the real eigenfunctions or a close approximation yields an intense noise in the map, producing the meaningless mapping of ZoomOut and SFM. Instead, by mapping the eyes to their nearest point on the head, we can meaningfully align our extended basis." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 373, + 226, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 373, + 226, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 373, + 226, + 385 + ], + "type": "text", + "content": "5 Conclusions" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 422, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 567 + ], + "type": "text", + "content": "We presented a new pipeline for scalable shape matching that relies on a low-resolution representation to compute a functional map between dense shapes efficiently. A core part of this pipeline is our novel remeshing algorithm, which efficiently computes an intrinsic Delaunay triangulation of uniform surface sampling. This procedure is applied to reduce the computational cost of the functional maps framework and make it efficient also on very dense meshes. Furthermore, we use a fast and intuitive technique for extending scalar functions from low-resolution meshes to their dense counterparts, extending the computed correspondence to the original very high-resolution shapes. Our experimental evaluation proved that our procedure is effective on various types of shapes at different resolutions, outperforming other state-of-the-art solutions and solving topological issues related to the alignment of the Laplacian eigenbasis." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "While using a triangulated mesh as a low-resolution representation could, in principle, be exploited with other types of shape-matching pipelines, we only tested it within the functional maps framework, using the Laplace-Beltrami eigenbasis or bases derived from it. While the current pipeline binds the method's applicability to the setting of functional correspondences, we believe that studying its behaviour with different \"backends\", as well as exploring the relationships between the low-resolution and the high-resolution correspondences, would be an interesting matter for future investigations." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "text", + "content": "F. Maggioli et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 258 + ], + "type": "text", + "content": "This work was partially supported by the Italian Ministry of Education, Universities and Research under the grant Dipartimento di Eccellenza 2023-2027 of the Department of Informatics, Systems and Communication of the University of Milano-Bicocca and by the PRIN 2022 project GEOPRIDE Geometric primitive fitting on 3D data for geometric analysis and 3D shapes. This work is also supported by the ERC grant no.802554 SPECGEO and PRIN 2020 project no.2020TA3K9N LEGO.AI. We acknowledge the support of NVIDIA Corporation with the RTX A5000 GPUs granted through the Academic Hardware Grant Program to the University of Milano-Bicocca for the project Learned representations for implicit binary operations on real-world 2D-3D data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 275, + 197, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 275, + 197, + 288 + ], + "spans": [ + { + "bbox": [ + 133, + 275, + 197, + 288 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 298, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 298, + 481, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 298, + 481, + 330 + ], + "spans": [ + { + "bbox": [ + 138, + 298, + 481, + 330 + ], + "type": "text", + "content": "1. Aflalo, Y., Brezis, H., Kimmel, R.: On the optimality of shape and data representation in the spectral domain. SIAM Journal on Imaging Sciences 8(2), 1141-1160 (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 331, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 481, + 363 + ], + "type": "text", + "content": "2. Amenta, N., Choi, S., Dey, T.K., Leekha, N.: A simple algorithm for homeomorphic surface reconstruction. In: Proceedings of the sixteenth annual symposium on Computational geometry. pp. 213-222 (2000)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 363, + 481, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 363, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 138, + 363, + 481, + 384 + ], + "type": "text", + "content": "3. Bobenko, A.I., Springborn, B.A.: A discrete laplace-beltrami operator for simplicial surfaces. Discrete & Computational Geometry 38(4), 740-756 (2007)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 384, + 481, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 384, + 481, + 406 + ], + "spans": [ + { + "bbox": [ + 138, + 384, + 481, + 406 + ], + "type": "text", + "content": "4. Botsch, M., Pauly, M., Rossl, C., Bischoff, S., Kobbelt, L.: Geometric modeling based on triangle meshes. In: ACM SIGGRAPH 2006 Courses, pp. 1-es (2006)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 406, + 481, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 406, + 481, + 427 + ], + "spans": [ + { + "bbox": [ + 138, + 406, + 481, + 427 + ], + "type": "text", + "content": "5. Bowyer, A.: Computing Dirichlet tessellations*. The Computer Journal 24(2), 162-166 (01 1981)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 427, + 481, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 427, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 138, + 427, + 481, + 449 + ], + "type": "text", + "content": "6. Bronstein, A., Bronstein, M., Kimmel, R.: Numerical Geometry of Non-Rigid Shapes. Springer, New York, NY (2008)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 449, + 481, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 481, + 471 + ], + "type": "text", + "content": "7. Burghard, O., Dieckmann, A., Klein, R.: Embedding shapes with green's functions for global shape matching. Computers & Graphics 68, 1-10 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 471, + 481, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 471, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 138, + 471, + 481, + 503 + ], + "type": "text", + "content": "8. Crane, K., Weischedel, C., Wardetzky, M.: Geodesics in heat: A new approach to computing distance based on heat flow. ACM Transactions on Graphics (TOG) 32(5), 1-11 (2013)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 503, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 503, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 503, + 481, + 525 + ], + "type": "text", + "content": "9. Delaunay, B., et al.: Sur la sphere vide. Izv. Akad. Nauk SSSR, Otdelenie Matematicheskii i Estestvennyka Nauk 7(793-800), 1-2 (1934)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 525, + 481, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 481, + 546 + ], + "type": "text", + "content": "0. Deng, B., Yao, Y., Dyke, R.M., Zhang, J.: A survey of non-rigid 3d registration. In: Computer Graphics Forum. vol. 41, pp. 559-589. Wiley Online Library (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 546, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 546, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 138, + 546, + 481, + 568 + ], + "type": "text", + "content": "1. Dijkstra, E.W.: A note on two problems in connexion with graphs. Numerische Mathematik 1, 269-271 (1959)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 568, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 481, + 600 + ], + "type": "text", + "content": "2. Donati, N., Corman, E., Melzi, S., Ovsjanikov, M.: Complex functional maps: A conformal link between tangent bundles. In: Computer Graphics Forum. vol. 41, pp. 317-334. Wiley Online Library (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 600, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 481, + 633 + ], + "type": "text", + "content": "3. Donati, N., Sharma, A., Ovsjanikov, M.: Deep geometric functional maps: Robust feature learning for shape correspondence. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8592-8601 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "text", + "content": "4. Dyer, R., Zhang, H., Möller, T.: Voronoi-delaunay duality and delaunay meshes. In: Proceedings of the 2007 ACM symposium on Solid and physical modeling. pp. 415-420 (2007)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 264, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 264, + 91, + 447, + 102 + ], + "type": "text", + "content": "ReMatching: Scalable Shape Correspondence" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 481, + 138 + ], + "type": "text", + "content": "15. Edelsbrunner, H., Shah, N.R.: Triangulating topological spaces. In: Proceedings of the tenth annual symposium on Computational geometry. pp. 285-292 (1994)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 170 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 170 + ], + "type": "text", + "content": "16. Eynard, D., Rodola, E., Glashoff, K., Bronstein, M.M.: Coupled functional maps. In: 2016 Fourth International Conference on 3D Vision (3DV). pp. 399-407. IEEE (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 481, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 481, + 191 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 481, + 191 + ], + "type": "text", + "content": "17. Ezuz, D., Ben-Chen, M.: Deblurring and denoising of maps between shapes. In: Computer Graphics Forum. vol. 36, pp. 165-174. Wiley Online Library (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 192, + 481, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 192, + 481, + 235 + ], + "spans": [ + { + "bbox": [ + 133, + 192, + 481, + 235 + ], + "type": "text", + "content": "18. Fei, Y., Chen, S., Su, D., Luo, J., Li, M.: A new algorithm for repairing non-manifold surfaces. In: 2013 IEEE 10th International Conference on High Performance Computing and Communications & 2013 IEEE International Conference on Embedded and Ubiquitous Computing. pp. 1704-1708. IEEE (2013)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 236, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 481, + 278 + ], + "type": "text", + "content": "19. Gao, M., Roetzer, P., Eisenberger, M., Lähner, Z., Moeller, M., Cremers, D., Bernard, F.: Sigma: Scale-invariant global sparse shape matching. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 645-654 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 279, + 468, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 279, + 468, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 279, + 468, + 289 + ], + "type": "text", + "content": "20. Guennebaud, G., Jacob, B., et al.: Eigen v3. http://eigen.tuxfamily.org (2010)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 289, + 481, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 481, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 481, + 321 + ], + "type": "text", + "content": "21. Hoppe, H., DeRose, T., Duchamp, T., McDonald, J., Stuetzle, W.: Mesh optimization. In: Proceedings of the 20th annual conference on Computer graphics and interactive techniques. pp. 19-26 (1993)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 322, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 481, + 342 + ], + "type": "text", + "content": "22. Jacobson, A., Panozzo, D., et al.: libigl: A simple " + }, + { + "bbox": [ + 132, + 322, + 481, + 342 + ], + "type": "inline_equation", + "content": "\\mathrm{C}++" + }, + { + "bbox": [ + 132, + 322, + 481, + 342 + ], + "type": "text", + "content": " geometry processing library (2018), https://libigl.github.io/" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 343, + 481, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 481, + 364 + ], + "type": "text", + "content": "23. Jiang, Z., Schneider, T., Zorin, D., Panozzo, D.: Bijective projection in a shell. ACM Transactions on Graphics (TOG) 39(6), 1-18 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 365, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 481, + 385 + ], + "type": "text", + "content": "24. Kim, V.G., Lipman, Y., Funkhouser, T.: Blended intrinsic maps. ACM transactions on graphics (TOG) 30(4), 1-12 (2011)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "type": "text", + "content": "25. Kimmel, R., Sethian, J.A.: Computing geodesic paths on manifolds. Proceedings of the national academy of Sciences 95(15), 8431-8435 (1998)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 407, + 481, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 438 + ], + "type": "text", + "content": "26. Kressner, D.: Numerical Methods for General and Structured Eigenvalue Problems. Lecture Notes in Computational Science and Engineering, Springer Berlin Heidelberg (2006)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "type": "text", + "content": "27. Leibon, G., Letscher, D.: Delaunay triangulations and voronoi diagrams for riemannian manifolds. In: Proceedings of the sixteenth annual symposium on Computational geometry. pp. 341-349 (2000)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "type": "text", + "content": "28. Lescoat, T., Liu, H.T.D., Thiery, J.M., Jacobson, A., Boubekeur, T., Ovsjanikov, M.: Spectral mesh simplification. In: Computer Graphics Forum. vol. 39, pp. 315-324. Wiley Online Library (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "type": "text", + "content": "29. Levy, B.: Laplace-beltrami eigenfunctions towards an algorithm that \"understands\" geometry. In: IEEE International Conference on Shape Modeling and Applications 2006 (SMI'06). pp. 13-13 (2006)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 536, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 557 + ], + "type": "text", + "content": "30. Lévy, B., Zhang, H.: Spectral mesh processing. In: ACM SIGGRAPH 2010 Courses, pp. 1-312 (2010)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 558, + 481, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 481, + 590 + ], + "type": "text", + "content": "31. Liu, H.T.D., Gillespie, M., Chislett, B., Sharp, N., Jacobson, A., Crane, K.: Surface simplification using intrinsic error metrics. ACM Trans. Graph. 42(4) (jul 2023). https://doi.org/10.1145/3592403, https://doi.org/10.1145/3592403" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 590, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 481, + 611 + ], + "type": "text", + "content": "32. Liu, H.T.D., Jacobson, A., Ovsjanikov, M.: Spectral coarsening of geometric operators. ACM Transactions on Graphics (TOG) 38(4) (jul 2019)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 612, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 612, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 612, + 481, + 632 + ], + "type": "text", + "content": "33. Maggioli, F., Baieri, D., Melzi, S., Rodolà, E.: Newton's fractals on surfaces via bicomplex algebra. In: ACM SIGGRAPH 2022 Posters, pp. 1-2 (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "text", + "content": "34. Maggioli, F., Melzi, S., Ovsjanikov, M., Bronstein, M.M., Rodola, E.: Orthogonalized fourier polynomials for signal approximation and transfer. In: Computer Graphics Forum. vol. 40, pp. 435-447. Wiley Online Library (2021)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "text", + "content": "F. Maggioli et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "35. Magnet, R., Ovsjanikov, M.: Scalable and efficient functional map computations on dense meshes. In: Computer Graphics Forum. vol. 42, pp. 89-101. Wiley Online Library (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 481, + 171 + ], + "type": "text", + "content": "36. Marin, R., Corona, E., Pons-Moll, G.: Nicp: Neural icip for 3d human registration at scale. In: European Conference on Computer Vision. Springer (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 171, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 171, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 171, + 481, + 214 + ], + "type": "text", + "content": "37. Melzi, S., Marin, R., Rodolà, E., Castellani, U., Ren, J., Poulenard, A., Wonka, P., Ovsjanikov, M.: Shrec 2019: Matching humans with different connectivity. In: Eurographics Workshop on 3D Object Retrieval. vol. 7, p. 3. The Eurographics Association (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 214, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 481, + 247 + ], + "type": "text", + "content": "38. Melzi, S., Ren, J., Rodola, E., Sharma, A., Wonka, P., Ovsjanikov, M.: Zoomout: spectral upsampling for efficient shape correspondence. ACM Transactions on Graphics (TOG) 38(6), 1-14 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 247, + 481, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 481, + 279 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 481, + 279 + ], + "type": "text", + "content": "39. Melzi, S., Rodolà, E., Castellani, U., Bronstein, M.M.: Localized manifold harmonics for spectral shape analysis. In: Computer Graphics Forum. vol. 37, pp. 20-34. Wiley Online Library (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 279, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 481, + 311 + ], + "type": "text", + "content": "40. Meyer, M., Desbrun, M., Schröder, P., Barr, A.H.: Discrete Differential-Geometry Operators for Triangulated 2-Manifolds. In: Visualization and mathematics III, pp. 35-57. Springer, New York, NY (2003)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 311, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 481, + 332 + ], + "type": "text", + "content": "41. Mitchell, J.S., Mount, D.M., Papadimitriou, C.H.: The discrete geodesic problem. SIAM Journal on Computing 16(4), 647-668 (1987)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 332, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 481, + 365 + ], + "type": "text", + "content": "42. Nasikun, A., Brandt, C., Hildebrandt, K.: Fast approximation of laplace-beltrami eigenproblems. In: Computer Graphics Forum. vol. 37, pp. 121-134. Wiley Online Library (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 365, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 481, + 396 + ], + "type": "text", + "content": "43. Nasikun, A., Hildebrandt, K.: The hierarchical subspace iteration method for laplace-beltrami eigenproblems. ACM Transactions on Graphics (TOG) 41(2), 1-14 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 396, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 481, + 429 + ], + "type": "text", + "content": "44. Nivoliers, V., Lévy, B., Geuzaine, C.: Anisotropic and feature sensitive triangular remeshing using normal lifting. Journal of Computational and Applied Mathematics 289, 225-240 (2015)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 429, + 481, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 481, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 481, + 461 + ], + "type": "text", + "content": "45. Nogneng, D., Melzi, S., Rodola, E., Castellani, U., Bronstein, M., Ovsjanikov, M.: Improved functional mappings via product preservation. In: Computer Graphics Forum. vol. 37, pp. 179-190. Wiley Online Library (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 461, + 481, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 481, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 481, + 494 + ], + "type": "text", + "content": "46. Nogneng, D., Ovsjanikov, M.: Informative descriptor preservation via commutativity for shape matching. In: Computer Graphics Forum. vol. 36, pp. 259-267. Wiley Online Library (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 494, + 481, + 526 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 481, + 526 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 481, + 526 + ], + "type": "text", + "content": "47. Ovsjanikov, M., Ben-Chen, M., Solomon, J., Butscher, A., Guibas, L.: Functional maps: a flexible representation of maps between shapes. ACM Transactions on Graphics (ToG) 31(4), 1-11 (2012)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 525, + 481, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 481, + 558 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 481, + 558 + ], + "type": "text", + "content": "48. Ovsjanikov, M., Corman, E., Bronstein, M., Rodola, E., Ben-Chen, M., Guibas, L., Chazal, F., Bronstein, A.: Computing and processing correspondences with functional maps. In: SIGGRAPH ASIA 2016 Courses, pp. 1-60 (2016)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 558, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 579 + ], + "type": "text", + "content": "49. Peyré, G., Cohen, L.D.: Geodesic remeshing using front propagation. International Journal of Computer Vision 69, 145-156 (2006)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 579, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 579, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 579, + 481, + 601 + ], + "type": "text", + "content": "50. Pinkall, U., Polthier, K.: Computing Discrete Minimal Surfaces and their Conjugates. Experimental mathematics 2(1), 15-36 (1993)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 130, + 601, + 481, + 633 + ], + "type": "text", + "content": "51. Ren, J., Melzi, S., Ovsjanikov, M., Wonka, P., et al.: Maptree: recovering multiple solutions in the space of maps. ACM Transactions on Graphics (TOG) 39(6), 264-1 (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 633, + 481, + 666 + ], + "type": "text", + "content": "52. Ren, J., Melzi, S., Wonka, P., Ovsjanikov, M.: Discrete optimization for shape matching. In: Computer Graphics Forum. vol. 40, pp. 81-96. Wiley Online Library (2021)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 264, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 264, + 91, + 447, + 102 + ], + "type": "text", + "content": "ReMatching: Scalable Shape Correspondence" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 314 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "53. Ren, J., Poulenard, A., Wonka, P., Ovsjanikov, M.: Continuous and orientation-preserving correspondences via functional maps. ACM Transactions on Graphics (ToG) 37(6), 1-16 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 171 + ], + "type": "text", + "content": "54. Reuter, M., Wolter, F.E., Peinecke, N.: Laplace-beltrami spectra as 'shape-dna' of surfaces and solids. Computer-Aided Design 38(4), 342-366 (2006)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "text", + "content": "55. Rivin, I.: Euclidean structures on simplicial surfaces and hyperbolic volume. Annals of mathematics 139(3), 553-580 (1994)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "type": "text", + "content": "56. Rustamov, R.M.: Barycentric coordinates on surfaces. In: Computer Graphics Forum. vol. 29, pp. 1507-1516. Wiley Online Library (2010)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "text", + "content": "57. Sahillioğlu, Y.: Recent advances in shape correspondence. The Visual Computer 36(8), 1705-1721 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": "58. Shoham, M., Vaxman, A., Ben-Chen, M.: Hierarchical functional maps between subdivision surfaces. In: Computer Graphics Forum. vol. 38, pp. 55-73. Wiley Online Library (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "text", + "content": "59. Vaxman, A., Ben-Chen, M., Gotsman, C.: A multi-resolution approach to heat kernels on discrete surfaces. In: ACM SIGGRAPH 2010 papers, pp. 1-10 (2010)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "text", + "content": "60. Watson, D.F.: Computing the n-dimensional Delaunay tessellation with application to Voronoi polytopes*. The Computer Journal 24(2), 167-172 (01 1981)" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "text", + "content": "F. Maggioli et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/ReMoS_ 3D Motion-Conditioned Reaction Synthesis for Two-Person Interactions/a734f12f-489f-47c4-b3e9-744b3d03eee5_content_list.json b/2024/ReMoS_ 3D Motion-Conditioned Reaction Synthesis for Two-Person Interactions/a734f12f-489f-47c4-b3e9-744b3d03eee5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2fe8e60fc4a56eb83676e1f7e6ecf0aa99d98ee6 --- /dev/null +++ b/2024/ReMoS_ 3D Motion-Conditioned Reaction Synthesis for Two-Person Interactions/a734f12f-489f-47c4-b3e9-744b3d03eee5_content_list.json @@ -0,0 +1,1748 @@ +[ + { + "type": "text", + "text": "REMoS: 3D Motion-Conditioned Reaction Synthesis for Two-Person Interactions", + "text_level": 1, + "bbox": [ + 259, + 140, + 743, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anindita Ghosh $^{1,2}$ , Rishabh Dabral $^{2}$ , Vladislav Golyanik $^{2}$ , Christian Theobalt $^{2}$ , and Philipp Slusallek $^{1,2}$", + "bbox": [ + 238, + 210, + 764, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ German Research Center for Artificial Intelligence (DFKI) \n $^{2}$ Max-Planck Institute for Informatics (MPII) \n $^{3}$ Saarland Informatics Campus", + "bbox": [ + 299, + 253, + 702, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Current approaches for 3D human motion synthesis generate high-quality animations of digital humans performing a wide variety of actions and gestures. However, a notable technological gap exists in addressing the complex dynamics of multi-human interactions within this paradigm. In this work, we present ReMoS, a denoising diffusion-based model that synthesizes full-body reactive motion of a person in a two-person interaction scenario. Given the motion of one person, we employ a combined spatio-temporal cross-attention mechanism to synthesize the reactive body and hand motion of the second person, thereby completing the interactions between the two. We demonstrate ReMoS across challenging two-person scenarios such as pair-dancing, Ninjutsu, kickboxing, and acrobatics, where one person's movements have complex and diverse influences on the other. We also contribute the ReMoCap dataset for two-person interactions containing full-body and finger motions. We evaluate ReMoS through multiple quantitative metrics, qualitative visualizations, and a user study, and also indicate usability in interactive motion editing applications. More details are available on the project page: https://vcai.mpi-inf.mpg.de/projects/remos.", + "bbox": [ + 261, + 329, + 738, + 578 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Reactive Motion Synthesis $\\cdot$ Denoising Diffusion Model", + "bbox": [ + 261, + 590, + 717, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 628, + 374, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Digital 3D character motion synthesis has emerged as the next frontier for animation pipelines [50], particularly through denoising diffusion probabilistic models (DDPMs) [61]. While methods for generating character motion for various tasks such as text or music conditioned motion synthesis [3,4,18,22,23,34,49,80], face and gesture synthesis [5,6,25,47,48,73], human-scene interaction [19,71,82,83] exist, synthesizing interactions between humans is relatively under-explored. Modeling such human-human interactions is essential for designing generative 3D human motion synthesis frameworks supporting the complex physical and social interplay of two interacting persons [66]. It offers new capabilities for character animation tools and software, with applications in commercial and entertainment media [26], interactive mixed and augmented reality [14], and social robotics [76].", + "bbox": [ + 212, + 657, + 787, + 825 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/234de5a2090ce50bb3f0bc7b8e396378ce68b906d95bbf314c3d408a4a4671d5.jpg", + "image_caption": [ + "Fig. 1: Visualizations of reactive 3D motion sequences synthesized with the proposed ReMoS approach. We synthesize the 3D full-body motion of the reactor (blue) conditioned only on the 3D motion of the actor (red), thereby completing the interactions between the two (Ninjutsu practice on the left and Lindy Hop dancing on the right). The synthesized hand interactions are enlarged and highlighted with circles." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Toward this goal, we focus on the task of modeling reactive motions. We aim to automatically generate realistic, temporally-aligned reactive motions of a responding person given the continuous motion of a guiding person (see Fig. 1). This alleviates a common overhead for animators, enabling them to design an acting character and automatically obtain meaningful motions for a reacting character. However, automating such a process poses two key challenges. First, two-person interaction synthesis increases the dimensionality of the already challenging problem of single-person motion synthesis. In our case, the generated reactive motions must align with the conditional signals provided in the form of 3D motion sequences from the actor. Second, doing so in a generative setting without any cues from text prompts or action labels (to obviate additional supervision and data needs) requires a careful trade-off between generating diverse motions and adhering to the narrow manifold of plausible reactions.", + "bbox": [ + 212, + 339, + 787, + 536 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While existing methods for synthesizing interactions [10, 20] and two-person motions [42, 54, 65] are good at generating plausible motions, they rely on additional annotations, such as action labels or text prompts, to specify the motion. Action labels typically depict actions at a high level, such as \"high-five\" or \"salsa\", but do not capture the fine-grained synchronization with the actor's motions. Textual descriptions offer nuance for the motions, but collecting accurate textual annotations is considerably cumbersome, particularly for contact-heavy or fast-paced two-person motions, such as dancing and exercising. Further, performing generative tasks through textual prompting remains challenging for non-experts [79]. Apart from the data, contact-based motion generation also needs to consider hand-based interactions to improve realism. This entails the need for frameworks capable of synthesizing hand movements with higher degrees of freedom, which are challenging to model but crucial for improving motion realism. Balancing focus between the broader full-body reactive movements and the finer hand motions involves addressing motion at two significantly different scales — a challenge exacerbated by the high degrees of freedom of the skeletons.", + "bbox": [ + 212, + 537, + 787, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With this objective in mind, we present ReMoS, a novel approach for Reactive Motion Synthesis with full-body articulations. Inspired by the advancements of DDPMs in 3D human motion synthesis [50,54,65], we develop a DDPM framework (Sec. 3.1) with a cascaded, two-stage generation strategy tailored to our", + "bbox": [ + 212, + 779, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "A. Ghosh et al.", + "bbox": [ + 271, + 114, + 375, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "problem setting (Fig. 2). In the first stage of our diffusion model, we generate the reactive motion for the reactor's body joints conditioned on the actor's body joints. In the second stage, we use the synthesized body motions as an additional parameter to generate masking for appropriate hand motions (Sec. 3.2). We propose a combined spatio-temporal cross-attention (CoST-XA) mapping between the actors' and the reactors' body motion embeddings, that learns the interdependencies in their motions without needing additional annotations. Further, to synthesize hand interactions, we introduce a hand-interaction-aware cross-attention (H-XA) mechanism to ensure the relevant hand joints react to the actors' motions, thus allowing the network to distill localized hand interaction features. We ensure accurate coordination between the actor and the reactor by using contact-based reaction loss (Sec. 3.3) and an inference-time guidance function (Sec. 3.4) that improves the plausibility of the body and hand interactions. To explore reaction synthesis in complex and diverse two-person scenarios, we contribute the REMOCAP dataset consisting of full-body and finger motion sequences for fast-paced swing dance of Lindy Hop [63], and the martial art technique of Ninjutsu [12] (Sec. 4). In summary, our technical contributions are:", + "bbox": [ + 212, + 146, + 787, + 404 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- REMoS. A novel method for reactive 3D human motion synthesis using a cascaded diffusion framework to generate full-body and hand motions. Our framework generates fine-grained reactive motions for complex and dynamic contact-based interactions and derives the reactions directly from the actor's motions, without requiring explicit label or text annotations.", + "- Interaction-Based Attentions. A combined spatio-temporal cross attention (CoST-XA) mechanism to enforce coherence between the body movements of the actor and the reactor, and a hand-interaction-aware cross-attention (H-XA) mechanism to enforce the appropriate hand-based interactions between the two characters.", + "- REMoCAP. A new dataset for two-person interactions under complex scenarios of Lindy Hop dancing and Ninjutsu. The dataset consists of $\\sim 275.7\\mathrm{K}$ motion frames with multiview RGB videos and 3D full-body motion capture of the two interacting persons. It further includes finger-level articulations." + ], + "bbox": [ + 225, + 404, + 787, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We evaluate our approach in different scenarios, including Lindy Hop dancing, Ninjutsu, Acrobatics [24] and Kickboxing [55], and report state-of-the-art performance (Sec. 5.3). We also report a user study on comparing the visual quality of ReMoS compared to the ground truth and the baseline methods (Sec. 5.4).", + "bbox": [ + 212, + 614, + 787, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 700, + 387, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-Person 3D Motion Synthesis. Synthesizing close interactions between two or multiple persons is a challenging task in animation. Early works in multi-person interaction are based on motion graphs [59], interaction patches [58], momentum-based inverse kinematics and motion blending [37], and topologically-based pose representations [28, 29], to name a few. The increasing availability of interaction datasets [16, 39, 43, 53, 55, 78] has led to a rise in data-driven approaches for synthesizing digital partners or opponents in multi-person inter", + "bbox": [ + 212, + 733, + 787, + 842 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "ReMoS", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "action settings. Mousas [46] uses a hidden Markov model to control the dance motions of a digital partner in an immersive setup. Ahuja et al. [1] introduce a residual-attention model to generate body pose in a conversation setting conditioned on two audio signals and the opposite person's body pose. Starke et al. [64] propose a phase function-based network that learns asynchronous movements of each bone and its interaction with external objects. Guo et al. [24] propose the Extreme Pose Interaction dataset with a two-stream network with cross-interaction attention modules for forecasting pose sequences of two interacting characters. GAN-based models [20,45] generate motions for an interacting person conditioned on an input character and class labels depicting the type of reaction performed. These methods leverage the daily interaction datasets such as SBU Kinect [78] and 2C [55]. InterFormer [11] uses an interaction transformer with both spatial and temporal attention to generate reactive motions given some initial seed poses of both characters. It can synthesize sparse-level interactions based on the motions of the K3HI [32] and the DuetDance [39] datasets. Concurrent to our work, Duolando [60] uses an off-policy reinforcement learning model to predict tokenized motion for a leader and follower conditioned on music. The aforementioned methods either depend on seed motion as input or require additional conditions to drive the motion. In contrast, ReMoS focuses on synthesizing well-synchronized full-body and hand motions for the reactor conditioned only on the actor's 3D motions, without using any additional labels or prompts.", + "bbox": [ + 212, + 146, + 787, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion Based 3D Motion Synthesis. Denoising diffusion models [61] have recently demonstrated their high potential in generative human motion modeling, specifically in single-person conditional motion synthesis. Conditional single-person motion generation has been performed using diffusion-based approaches for co-speech gestures [2, 85], audio-driven motion [13, 68, 84], and text-driven motion [54,67,77,81] synthesis. Diffusion-based techniques have also been extensively used for conditional synthesis for human-object interactions [38, 40, 41, 74], hand-object interactions [44, 75], and human-scene interactions [33]. Guided motion diffusion models such as GMD [35] and TraceAndPace [52] incorporate spatial constraints on motion trajectories, to guide the motion towards a goal at inference time. OmniControl [72] proposes spatial and realism guidance to control any joint for diffusion-based human motion generation. For two-person interaction synthesis, RAIG [65] proposes a diffusion-based, role-aware approach for two-person interactions, given separate textual descriptions for each person. BiGraphDiff [10] uses graph transformer denoising diffusion to learn two-person interaction conditioned on action labels. ComMDM [54] uses a communication block between two MDMs [67] to coordinate two-person interaction generation. ContactGen [21] presents a diffusion-based contact prediction module that adaptively estimates potential contact regions between two humans according to the interaction label. Liang et al. [42] propose InterGen, a diffusion-based model to synthesize two-person interactions given text descriptions as input conditions. They also propose the InterHuman dataset that includes diverse two-person interaction scenarios with rich text annotations. Inspired by these recent approaches, we design our proposed method ReMoS as a DDPM-based", + "bbox": [ + 212, + 477, + 787, + 843 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "A. Ghosh et al.", + "bbox": [ + 271, + 114, + 374, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9d441678b2f9f21acf62a1b4370d336ba081f4cc3395a4fbc8383021159f9deb.jpg", + "image_caption": [ + "Fig.2: ReMoS Overview. Given the motion of the actor (bottom-middle, in red), we synthesize a plausible motion for the reactor (bottom-left, in blue). We achieve this using a denoising diffusion-based probabilistic model (center) trained on reactive motion sequences (top-left, in blue)." + ], + "image_footnote": [], + "bbox": [ + 217, + 145, + 504, + 270 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D motion-conditioned reaction synthesis method, consisting of a cascaded diffusion model with a combined spatio-temporal cross-attention mechanism. This allows ReMoS to implicitly learn the fine-grained synchronization between two interacting persons from only one of the person's motions, without any need for additional prompts or labels, and synthesize the corresponding motions of the second person. ReMoS further synthesizes plausible hand motions for the reactor to incorporate realistic hand-based interactions.", + "bbox": [ + 212, + 297, + 787, + 404 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Reactive Motion Synthesis", + "text_level": 1, + "bbox": [ + 215, + 426, + 511, + 443 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We consider a digital, two-person interaction setting where the motions of one character, the actor, are known, and we need to synthesize the motions of the other character, the reactor. ReMoS aims to synthesize the synchronized reactive motions of the reactor in such a setting. We denote the reactor's motion as $X = \\{X_{B},X_{H}\\}$ , where $X_{B}\\in \\mathbb{R}^{N\\times J_{B}\\times 3}$ denotes the 3D full-body joint positions with $J_{B}$ body joints, $X_{H}\\in \\mathbb{R}^{N\\times J_{H}\\times 3}$ denotes the 3D finger joint positions with $J_{H}$ finger joints across both hands, and $N$ denotes the number of frames in the motion sequence. Likewise, we denote the actor's motion as $Y = \\{Y_{B},Y_{H}\\}$ . Our goal is to model the conditional probability distribution $\\mathrm{P}(X|Y)$ from which we can sample plausible reactive motions. We model this distribution using conditional denoising diffusion probabilistic models (DDPMs), owing to their unique strengths in capturing temporal coherence of motion data and handling complex motion dynamics [13,67,68]. We elaborate our methodology below.", + "bbox": [ + 212, + 455, + 787, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 DDPM for Reactive Motion Synthesis", + "text_level": 1, + "bbox": [ + 215, + 672, + 578, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The diffusion process consists of two steps: forward or destructive diffusion, and reverse or denoising diffusion. In the forward process, we progressively corrupt a clean reactive motion sequence $X$ by adding Gaussian noise $\\epsilon$ to it for $T$ steps. With sufficiently small noise and large $T$ , we can get $X^{(T)} \\sim \\mathcal{N}(0,I)$ following closed-form formulation of Ho et al. [30]:", + "bbox": [ + 212, + 696, + 787, + 773 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nX ^ {(t)} = \\sqrt {\\bar {\\alpha} _ {t}} X ^ {(0)} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\quad \\epsilon \\sim \\mathcal {N} (0, I), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 781, + 785, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\bar{\\alpha}_t$ controls the rate of diffusion and $t\\in [0,T]$ . Reversing this diffusion process allows for sampling novel motion sequences from a multivariate Gaussian", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "ReMoS", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "distribution $p\\left(X^{(T)}\\right)\\sim \\mathcal{N}(0,I)$ as", + "bbox": [ + 215, + 143, + 470, + 162 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np \\left(X ^ {(0)}\\right) = p \\left(X ^ {(T)}\\right) \\prod_ {t = 1} ^ {T} p \\left(X ^ {(t - 1)} \\mid X ^ {(t)}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 348, + 175, + 785, + 215 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We approximate the computationally intractable term $p\\left(X^{(t - 1)}|X^{(t)}\\right)$ with a learnable function $f_{\\theta}\\left(X^{(t)},t\\right)$ , and optimize $\\theta$ to encode the space of human reactive motions. Our reactive motion $X$ is also conditioned on the actor's motion $Y$ . Therefore, we modify the learnable function as $X^{(t - 1)} = f_{\\theta}\\left(X^{(t)},t,Y\\right)$ . Following recent works [51,67], we estimate the original motion $X^{(0)} = f_{\\theta}\\left(X^{(t)},t,Y\\right)$ from our diffusion model by iterating through all $t$ denoising steps during inference.", + "bbox": [ + 214, + 227, + 787, + 324 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 ReMoS Framework", + "text_level": 1, + "bbox": [ + 215, + 345, + 421, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To generate fine-grained reactions with appropriate hand motions, ReMoS decodes the reactive motion of $X$ in a cascaded fashion (Fig. 3). It first estimates the full-body joints $X_{B}$ , and then the hand joints $X_{H}$ as", + "bbox": [ + 214, + 371, + 787, + 417 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nX _ {B} ^ {(0)} = f _ {\\theta_ {B}} \\left(X _ {B} ^ {(t)}, t, Y _ {B}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 307, + 428, + 784, + 453 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nX _ {H} ^ {(0)} = f _ {\\theta_ {H}} \\left(X _ {H} ^ {(t)}, t, Y _ {H}, \\mathbb {1} _ {H _ {A}} \\left(Y _ {B}\\right) \\mathbb {1} _ {H _ {R}} \\left(X _ {B} ^ {(0)}\\right)\\right), \\quad \\text {a n d} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 307, + 455, + 784, + 481 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nX ^ {(0)} = \\left\\{X _ {B} ^ {(0)}, X _ {H} ^ {(0)} \\right\\}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 307, + 483, + 784, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbb{1}_{H_A}$ and $\\mathbb{1}_{H_R}$ are binary, spatio-temporal hand-interaction mask functions that determine which hand joints of the reactor and actor interact at any given frame. Our cascaded framework comes from the observation that full-body articulations occur at significantly different scales than hand articulations. Since the body motions affect the hand articulations, our cascaded framework feeds the generated body motions as an additional condition to the hand generation module (Fig. 3). We also train the body and the hand generation modules separately, with additional hand-interaction-aware attention mapping for the hand joints. Previous work [19] used such disjoint training strategy on conditional motion synthesis for single persons. To better accommodate for two-person interactions in our case, we benefit from using a cascaded diffusion strategy [31].", + "bbox": [ + 214, + 518, + 787, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Body Synthesis Module. In the first stage of our cascaded framework, we diffuse the reactor's body motion, $X_B^{(t)}$ , at each diffusion step $t$ , and feed it into a transformer decoder [69] block. To condition the reactor's body motion on the actor's body motion $Y_B$ , we introduce a combined spatio-temporal cross-attention (CoST-XA) mapping. CoST-XA combines the spatial and temporal features of the actor's and the reactor's motions through an attention matrix to simultaneously learn the inter-dependencies between each pair of reactor and actor body joints $(x_{n,j}, y_{n,j})$ , $x_{n,j} \\in X_B^{(t)}$ and $y_{n,j} \\in Y_B$ , respectively, at each frame. Denoting the query from the reactor's motion features as $Q_B \\in \\mathbb{R}^{NJ_B \\times 3}$", + "bbox": [ + 214, + 700, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "A. Ghosh et al.", + "bbox": [ + 271, + 114, + 375, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/91b1996d2b0933da10018d5f4abc699bdb32155efbbd32d044eeffe9d7002479.jpg", + "image_caption": [ + "Fig. 3: ReMoS Framework. Given the full-body sequence of the actor (left, in red), we input noisy body and hand samples (from below) in a cascaded fashion. We synthesize the body samples first, and use them for hand-interaction-aware attention masking (top-center) to synthesize the denoised hand samples (top-right). The full-body reactive motion is a concatenation of the denoised body and hand samples (right, in blue)." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 781, + 265 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and the key-value pair from the actor's motion features as $K_{B} \\in \\mathbb{R}^{NJ_{B} \\times 3}$ and $V_{B} \\in \\mathbb{R}^{NJ_{B} \\times 3}$ , respectively, we formulate an attention matrix of dimensions $NJ_{B} \\times NJ_{B}$ . In contrast, most previous approaches [42,65] use cross-attention only on the temporal features of the motion using an attention matrix of shape $N \\times N$ . InterFormer [11] sequentially uses spatial and temporal cross-attention modules instead of combining them, which may result in a partial loss of information on fine-grained, inter-person interactions across time, especially in the absence of additional annotations (prompts or action labels) to condition the reactions. Our proposed cross-attention module, which we define as", + "bbox": [ + 214, + 378, + 787, + 513 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\text {C o S T - X A} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {B} K _ {B} ^ {T}}{\\sqrt {d _ {K _ {B}}}}\\right) V _ {B}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 526, + 784, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "crucially considers combinations of spatial and temporal interaction features between different body segments of the actor and the reactor to efficiently synchronize their motions. We simultaneously project the diffusion timestep $t$ at each denoising step into the transformer block after the attention blocks and use a final layer of fully-connected network with SiLU activations [15] and batch normalization [7] to generate the body motions. The first stage output is a vector $X_B^{(0)} \\in \\mathbb{R}^{N \\times J_B \\times 3}$ representing the reactor's clean, synthesized body motion.", + "bbox": [ + 214, + 577, + 787, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hands Synthesis Module. The second stage of our cascaded framework synthesizes the reactor's hand joints. We input noisy samples $X_H^{(t)}$ into a similar transformer decoder block conditioned on the actor's hand motions $Y_H$ . Here, we require the attention weights to be high for the hands used in the hand-based interactions and low for the passive hands. We explicitly enforce this using a spatio-temporal hand-interaction-aware cross-attention mapping (H-XA) to encourage learning localized interaction features. For H-XA to work, we introduce binary hand-interaction masks $\\mathbb{1}_{H_A}$ and $\\mathbb{1}_{H_R} \\in \\mathbb{R}^{N \\times J_H}$ . Its entries are calculated by thresholding the distances of the actor's and reactor's wrist joints in $Y_B$ and", + "bbox": [ + 214, + 702, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "ReMoS", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$X_B^{(0)}$ , respectively, to the nearest body joints of the other person. The active entries of $\\mathbb{1}_{H_A}$ and $\\mathbb{1}_{H_R}$ determine which of the actor's and reactor's hands are sufficiently close to the body of the other person (and hence interacting). We then project the encoded $Y_H$ features into each decoder layer using H-XA, as", + "bbox": [ + 215, + 145, + 787, + 209 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {H} - \\mathrm {X A} = \\operatorname {s o f t m a x} \\left(\\frac {\\left(\\mathbb {1} _ {H _ {R}} \\odot Q _ {H}\\right) \\left(\\mathbb {1} _ {H _ {A}} \\odot K _ {H}\\right) ^ {T}}{\\sqrt {d _ {K _ {H}}}}\\right) V _ {H}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 229, + 785, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $K_{H}$ and $V_{H}$ are the hand motion features from the actor, $Q_{H}$ are the hand motion features from the reactor with $d_{K_{H}}$ channels, and $\\odot$ denotes element-wise product for masking the query and key values. We project the diffusion timestep $t$ at each denoising step into the transformer block and use fully-connected networks with SiLU activations to generate the clean, synthesized reactor hand motions $X_{H}^{(0)} \\in \\mathbb{R}^{N \\times J_{H} \\times 3}$ . At the end, to obtain the full reactive motion $X^{(0)}$ , we concatenate $\\left\\{X_{B}^{(0)}, X_{H}^{(0)}\\right\\}$ as in Eqn. 5.", + "bbox": [ + 215, + 277, + 787, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Losses and Training Details", + "text_level": 1, + "bbox": [ + 215, + 412, + 491, + 428 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We train ReMoS to minimize a weighted sum of three loss terms, the reconstruction loss $\\mathcal{L}_c$ , the reaction loss $\\mathcal{L}_r$ , and the kinematic loss $\\mathcal{L}_k$ , as", + "bbox": [ + 215, + 436, + 785, + 467 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {c} \\mathcal {L} _ {c} + \\lambda_ {r} \\mathcal {L} _ {r} + \\lambda_ {k} \\mathcal {L} _ {k}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 410, + 477, + 785, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\lambda_{c},\\lambda_{r}$ and $\\lambda_{k}$ are scalar weights to balance the individual losses.", + "bbox": [ + 215, + 502, + 728, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Reconstruction Loss $\\mathcal{L}_c$ . This is the standard diffusion data term that minimizes the $\\ell_2$ -distance between the ground truth and the synthesized reactive motion, as $\\mathcal{L}_c = \\left\\| X - X^{(0)}\\right\\|_2$ . While the reconstruction loss provides the vital data term to drive the training, it does not enforce interaction synchronization between the actor and the reactor, or any kinematic constraints on the motion.", + "bbox": [ + 215, + 530, + 787, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Reaction loss $\\mathcal{L}_r$ . We introduce the reaction loss to ensure accurate timing and spatial positioning of the reactor's motion with respect to the actor's motion. We calculate the ground-truth Euclidean distance $d(\\cdot, \\cdot)$ between each joint of the actor and the reactor at each frame and minimize the deviations from these distances for the actor and the synthesized reactor's motions. We note that this distance term also implicitly constrains the reactor's joints in frames that are less relevant for interactions, i.e., where the reactor's joints are far from the actor's joints. Further, the Euclidean distances for their hand joints in frames with no hand-based contact become drastically high. To mitigate these concerns, we use an exponentially decaying distance-aware weight $\\exp(-d(x_{n,j}, y_{n,j}))$ at each frame $n$ to focus more on the reactor's joints that are closer to the actor, and therefore more relevant for interaction (see Fig. 4). Thus, we get our reaction loss term as", + "bbox": [ + 215, + 618, + 787, + 797 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r} = \\frac {1}{N J} \\sum_ {n = 1} ^ {N} \\sum_ {j = 1} ^ {J} \\exp \\left(- d \\left(x _ {n, j}, y _ {n, j}\\right)\\right) \\cdot \\left| d \\left(x _ {n, j}, y _ {n, j}\\right) - d \\left(x _ {n, j} ^ {(0)}, y _ {n, j}\\right) \\right|. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 272, + 805, + 785, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "A. Ghosh et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b8bde8858c9488d50183bc0e7fb196dc87da178e8e46dfc08e0bed6d76593f69.jpg", + "table_caption": [ + "Table 1: Dataset Comparisons. Comparing ReMoCap with existing multi-person interaction datasets." + ], + "table_footnote": [], + "table_body": "
Dataset (Chronologically)Motion FramesDuration (hours)Multi-view RGB VideosFinger Articulation
SBU [78]~ 7K0.13XX
NTU-26 [43]~ 22K0.47✓X
2C [55]~ 13K0.06XX
DanceDB [53]~ 1.44M4.00XX
DuetDance [39]~ 196K1.09XX
CHI3D [16]~ 486K2.70✓X
ExPI [24]~ 30K0.33✓X
InterHuman [42]~ 107M6.56XX
DD100 [60]~ 200K1.92X✓
ReMoCap (ours)~ 275.7K2.04✓✓
", + "bbox": [ + 215, + 210, + 529, + 351 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/78f89a1eb4a2a8184ebde828d8e4c8ad8e1b1249baa0032b99b449d940958155.jpg", + "image_caption": [ + "Fig.4: Visualization of Distance Aware Reaction Loss. We use an exponentially decaying distance-aware reaction loss to focus more on the reactor's joints that are closer to the actor." + ], + "image_footnote": [], + "bbox": [ + 578, + 143, + 764, + 266 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Kinematic Loss $\\mathcal{L}_k$ . To maintain the kinematic plausibility of the generated motions, we follow existing literature on regularizers for bone length consistency, foot contacts, and temporal consistency [56,57,68]. Our kinematic loss term is a weighted sum of the joint velocity loss $\\mathcal{L}_{vel}$ , the joint acceleration loss $\\mathcal{L}_{acc}$ , the bone length consistency loss $\\mathcal{L}_{bone}$ , and the foot sliding loss $\\mathcal{L}_{foot}$ , as", + "bbox": [ + 215, + 391, + 787, + 468 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {k} = \\lambda_ {v} \\mathcal {L} _ {v e l} + \\lambda_ {a} \\mathcal {L} _ {a c c} + \\lambda_ {b} \\mathcal {L} _ {b o n e} + \\lambda_ {f} \\mathcal {L} _ {f o o t}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 479, + 784, + 497 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\lambda_v, \\lambda_a, \\lambda_b$ and $\\lambda_f$ are scalar weights (more details in the appendix).", + "bbox": [ + 215, + 507, + 751, + 523 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Inference Time Spatial Guidance", + "text_level": 1, + "bbox": [ + 215, + 545, + 534, + 560 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "While our method synthesizes plausible reactive motions, it can sometimes spatially misalign the reactor's body to the actor's, especially at the arm joints, for fast-paced, contact-heavy interactions. This, in turn, affects finger-joint synthesis in the second stage of our cascaded diffusion. To improve spatial alignment, we leverage guidance functions [35, 52] that can provide gradients to nudge the sampling process towards a certain direction. We design a guidance function $G\\left(\\phi ,\\hat{\\phi}\\right)\\in \\mathbb{R}^{N\\times J_A\\times 3}$ , which minimizes the distance between the $J_{A}$ arm joints of the actor $(\\phi)$ and the $J_{A}$ arm joints of the synthesized reactor $(\\hat{\\phi})$ . Specifically, we re-apply our interaction masks $\\mathbb{1}_{H_A}$ and $\\mathbb{1}_{H_R}$ to determine which of the reactor's hands are more likely to be interacting with the actor, and minimize the distances between the arm joints of the corresponding sides as $G = \\arg \\min_{\\hat{\\phi}}\\left(\\left\\| \\mathbb{1}_{H_A}\\odot \\phi -\\mathbb{1}_{H_R}\\odot \\hat{\\phi}\\right\\|\\right)$ . We then plug $G$ into the denoising pipeline of our body diffusion module, as", + "bbox": [ + 215, + 571, + 787, + 781 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nX _ {B} ^ {(0)} = X _ {B} ^ {(0)} - \\gamma \\nabla_ {X _ {B} ^ {(0)}} G (\\phi , \\hat {\\phi}), \\quad \\gamma = \\text {g u i d a n c e s c a l e}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 295, + 792, + 785, + 818 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "ReMoS", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 ReMoCap Dataset", + "text_level": 1, + "bbox": [ + 215, + 143, + 431, + 162 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We propose the ReMoCap dataset to facilitate research on contact-based two-person interactions with finger-level articulations. It contains complex two-person interactions in two separate scenarios: the fast-paced, swing-style Lindy Hop dancing, and the martial arts of Ninjutsu. We consider these two types of motions for studying physical interactions for two main reasons. First, these motions encompass a significantly diverse spectrum of motions in terms of intensity, speed, and style. Second, choreographed motion generation is developing as an active field of research and we envision this dataset to facilitate research in interactive applications such as performing in virtual reality [9] and remote tutoring [53].", + "bbox": [ + 212, + 178, + 787, + 316 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Data Collection. We invited 4 trained Lindy Hop dancers and 5 trained Ninjutsu artists to perform diverse interactions in a multi-view capture studio. We tracked their motions using a commercially available markerless, multi-view motion-capture system [8], which tracks 93 degrees-of-freedom driving 69 body joints. Capturing motions in a markerless setting enables the performers to move uninhibited, while also making the data suitable for training monocular or multiview motion capture methods under severe inter-person occlusions. We also capture RGB videos from 120 camera views for each sequence. The dataset includes 3D skeleton poses with full-body and finger annotations, foreground-background segmentation masks, and 3D surface reconstructions of the subjects. We capture sequences of multiple lengths totaling $\\sim 275.7\\mathrm{K}$ frames (2.04 hours) of motion data from each view at a frame rate of 50 fps for the Lindy Hop and 25 fps for the Ninjutsu. Out of all the frames, around $150\\mathrm{K}$ frames have hand-based interactions between the two characters where the closest distance between the finger joints of the actor and the reactor is within $50~\\mathrm{mm}$ . We present more dataset statistics in the appendix.", + "bbox": [ + 212, + 330, + 787, + 573 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Dataset Comparison. Table 1 shows a comprehensive comparison of ReMoCap with existing multi-person interaction datasets. Existing datasets consisting of two-person interactions, such as SBU [78], K3HI [32], NTU-26 [43] and 2C [55], are limited in size and motion capture quality. They typically feature simple actions, such as handshakes, punching, pushing and kicking, with weak interactions, and do not capture hand motions. The recent ExPI [24] dataset captures Lindy Hop aerial sequences to model more complex interactions, and the InterHuman [42] dataset features both daily motions (e.g., passing objects, greeting, communicating) and professional motions (e.g., Taekwondo, Latin dance, boxing). However, these datasets do not provide finger-level motion capture data, which is a key requirement to intricately model inter-human activities. Two-person dance datasets, such as DuetDance [39] and DanceDB [53], also lack the hand motion data needed for modeling interactions. Only the concurrent work, Duolando $[60]^*$ proposes the DD100 dataset with strong interactions between two dancers and provides hand motion data.", + "bbox": [ + 212, + 588, + 787, + 816 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "A. Ghosh et al.", + "bbox": [ + 271, + 114, + 374, + 126 + ], + "page_idx": 9 + }, + { + "type": "page_footnote", + "text": "*unpublished at the time of our paper submission", + "bbox": [ + 236, + 825, + 570, + 839 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments and Results", + "text_level": 1, + "bbox": [ + 215, + 143, + 495, + 162 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We conduct comprehensive experiments to evaluate ReMoS on multiple two-person datasets covering a wide range of interaction scenarios. We perform training and evaluation on multiple large-scale datasets, including 2C [55], ExPI [24], and our proposed ReMoCap. We provide the implementation details, report quantitative and qualitative comparisons on standard evaluation metrics and through a user study, report ablations, and show how to apply ReMoS as a motion editing tool. We provide details on the dataset preparation in the appendix.", + "bbox": [ + 212, + 175, + 787, + 282 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 301, + 455, + 316 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As a pre-processing step, we normalize the two-person body poses by translating the actor's motion $Y$ and the reactor's motion $X$ together, such that the root joint of $Y$ is at the global origin for all $N$ frames. We then compute the relative 3D joint coordinates of $X_B$ and $Y_B$ w.r.t. the root joint of $Y_B$ . For each hand, we compute the relative 3D joint coordinates of $X_H$ and $Y_H$ w.r.t. the corresponding wrist joints of $X_H$ and $Y_H$ . Using normalized inter-person, root-relative joint positions benefits the stability and convergence of our model. ReMoS uses $T = 500$ diffusion steps where $\\bar{\\alpha}$ changes linearly from 0.0002 to 0.02. We train for about 64K iterations on both sequences of ReMoCap using Adam [36] with a base learning rate of $10^{-5}$ and a batch size of 64. We decay the learning rate using a Step LR scheduler with a step size of 5 epochs and a decay rate of 0.99. We use $d = 256$ for our latent embedding representation and use 6 layers in our transformer decoder with 4 heads for calculating the attention. The training takes around 8 and 11 hours on an NVIDIA RTX A4000 GPU for Lindy Hop and Ninjutsu, respectively. The inference time is $\\sim 24s$ to generate 50 frames of full-body and finger motions (12.5s for body synthesis and then 11.5s for hand synthesis). We set $\\lambda_c, \\lambda_r, \\lambda_v = 10.0, \\lambda_a, \\lambda_k, \\lambda_b = 1.0$ and $\\lambda_f = 20.0$ (only applied after 100 training epochs) as the loss term weights in Eqns. 8 and 10. We set the guidance scale $\\gamma = 10^{-3}$ in Eqn. 11.", + "bbox": [ + 212, + 325, + 787, + 614 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Baselines and Ablated Versions", + "text_level": 1, + "bbox": [ + 215, + 633, + 519, + 648 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For baselines, we choose the closest motion synthesis methods in a two-person setting, namely, MixNMatch [20], InterFormer [11], ComMDM [54], Role-Aware Interaction Generation (RAIG) [65], and InterGen [42]. InterFormer [11] was originally trained in a reactive motion synthesis setting without additional annotations, such as action labels or text descriptions, and we maintain this setup. For the other methods, we mask out their input text/label embeddings to comply with our annotation-free setting. We re-train these methods on ReMoCap with a thorough hyper-parameter search and report their best performances. We provide more training details in the appendix.", + "bbox": [ + 212, + 657, + 787, + 794 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We also compare our proposed ReMoS model with five of its ablated versions: w/o diffusion. Training ReMoS using a transformer encoder-decoder network without using DDPM strategy.", + "bbox": [ + 215, + 795, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "ReMoS", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/b3bedbb1beb381c1b18a41d4208ea6f32277ecc74e1f7d8b678702f67472c5c9.jpg", + "table_caption": [ + "Table 2: Quantitative Evaluation on ReMoCap. We compare ReMoS with our baselines and ablated versions (Sec. 5.2) on the ReMoCap dataset. We evaluate these methods on metrics such as MPJPE, MPJVE, FID, and Diversity. $\\downarrow$ : lower is better, $\\uparrow$ : higher is better, $\\rightarrow$ : values closer to GT are better. Bold indicates best." + ], + "table_footnote": [], + "table_body": "
MethodLindy HopNinjutsu
MPJPE (mm) ↓MPJVE (mm) ↓FID ↓ (body)FID ↓ (hands)Div →MPJPE (mm) ↓MPJVE (mm) ↓FID ↓ (body)FID ↓ (hands)Div →
GT----7.57----10.51
InterFormer [11]66.68.260.530.654.54270.23.40.570.686.48
MixNMatch [20]70.210.30.770.782.48257.25.20.740.724.83
ComMDM [54]59.44.410.320.537.48201.24.10.340.589.98
RAIG [65]71.24.320.470.638.45199.15.10.210.6310.11
InterGen [42]62.63.920.300.617.21172.63.90.320.579.98
ReMoS (ours)40.72.260.120.267.62139.23.30.160.3510.26
w/o diffusion72.54.910.580.744.04224.54.10.520.646.06
w/o cascading63.94.950.510.557.12223.64.20.420.756.62
w/o CoST-XA44.23.620.210.397.45176.63.60.270.418.91
w/o reaction loss44.63.510.220.387.31144.63.70.230.398.99
w/o spatial guidance41.92.340.120.267.62139.43.40.160.3510.26
", + "bbox": [ + 215, + 212, + 781, + 405 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- w/o cascading. Training ReMoS with an integrated transformer for all the joints instead of a cascaded strategy for body and hand generation.", + "- w/o CoST-XA. Using a spatial followed by a temporal cross-attention mapping instead of the combined spatio-temporal features for body synthesis.", + "- w/o reaction loss. Removing the reaction loss (Eqn. 9) in training.", + "- w/o spatial guidance. Removing spatial guidance $G_{A}$ (Sec. 3.4) in inference." + ], + "bbox": [ + 215, + 431, + 784, + 523 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 215, + 544, + 459, + 559 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We evaluate ReMoS on standard evaluation metrics. We measure the temporal consistency deviation from the ground truth 3D pose following [57] to report the mean per-joint positional error (MPJPE) and the mean per-frame, per-joint velocity error (MPJVE), both in $mm$ , on the synthesized motions. We measure the Fréchet Inception Distance (FID) [27] to compare the distribution gap between the embedding spaces of the generated and ground-truth motions. As part of our method focuses on generating hand motions, we measure the FID score of body and hands separately for ReMoCap. We also compute the latent variance of the generated motions (Diversity) [67, 81]. Table 2 reports the quantitative evaluation of ReMoS with its baselines and ablations on ReMoCap. ReMoS achieves state-of-the-art performance for both the Lindy Hop pair-dance setting and the Ninjutsu setting. We observe that methods using denoising diffusion have higher Diversity compared to transformer-based [11] or GAN-based [20] methods. This enforces the variability claims of denoising diffusion models. We note at least $20\\%$ improvement in the MPJPE and around $40\\%$ improvement in the FID score for the reactor's body motion when using the proposed CoST-XA mechanism, confirming its benefit in learning fine-grained, inter-person dependencies across time. We also note an almost $50\\%$ improvement in the FID scores of", + "bbox": [ + 212, + 566, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "A. Ghosh et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 3: Quantitative Evaluation on the ExPI and 2C datasets. We compare ReMoS with state-of-the-art motion synthesis methods on the ExPI [24] and 2C datasets [55]. $\\downarrow$ : lower is better, $\\uparrow$ : higher is better, $\\rightarrow$ : values closer to GT are better. Bold indicates best.", + "bbox": [ + 215, + 143, + 586, + 213 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/9c6599b67cb5b0b38ee705911180b0d247fc2c330e2b781026ba6b143edb9bca.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodExPI2C
MPJPE (mm) ↓MPJVE (mm) ↓FID ↓ (body) →Div
MPJPE (mm) ↓MPJVE (mm) ↓FID ↓ (body) →Div
GT---2.01---2.22
InterFormer [11]99.13.560.421.3190.75.110.521.45
MixNMatch [20]122.45.560.481.1862.46.010.471.24
ComMDM [54]121.45.410.452.4869.93.340.492.86
RAIG [65]131.23.960.532.5191.64.920.674.45
InterGen [42]100.63.910.432.0967.64.010.472.91
ReMoS (ours)97.93.520.411.9859.13.330.342.07
", + "bbox": [ + 215, + 224, + 588, + 340 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/bf342d683977dbf8a8e51cc8493fa9ddd7547e4b78f9b9677bdca992e685cb1f.jpg", + "table_caption": [ + "Table 4: User Study Results. Mean scores on a five-point Likert scale (scores 1 - 5)." + ], + "table_footnote": [], + "table_body": "
MethodMotion Quality ↑Reaction Plausibility ↑
GT4.86 ± 0.544.72 ± 0.56
InterFormer2.52 ± 0.612.28 ± 0.57
MixNMatch1.92 ± 0.712.18 ± 0.57
ComMDM3.02 ± 0.473.12 ± 0.52
RAIG2.83 ± 0.672.48 ± 0.65
InterGen3.18 ± 0.573.19 ± 0.53
ReMoS3.79 ± 0.553.88 ± 0.54
", + "bbox": [ + 599, + 212, + 785, + 325 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "the reactor's hand motion when using the cascaded strategy and the reaction loss. Further, the spatial guidance function fine-tuning improves MPJPE and MPJVE. The MPJPE values are overall higher in Ninjutsu than in Lindy Hop as the trajectory of the reactor is more diverse for Ninjutsu. We also report the evaluation of ReMoS with its baselines on the ExPI [24] and the 2C [55] datasets in Table 3. These datasets do not provide hand motions, so we only evaluate the reactor's body motions, and report state-of-the-art performance of ReMoS.", + "bbox": [ + 212, + 368, + 787, + 474 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 User Study", + "text_level": 1, + "bbox": [ + 215, + 494, + 356, + 510 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We evaluate the visual quality of our captured ground truth data and the generated reactive motions through a user study. We show participants 26 different interaction sequences across the ground truth, our method, and its baselines. For each sequence, we randomly show them three methods side-by-side and ask them to rate the 3D motions they observe in terms of $(a)$ the reactor's motion quality, irrespective of the actor's motion (Motion Quality), and $(b)$ the plausibility of the reaction given the actor's motion (Reaction Plausibility). We ask the participants to rank these motions from '1' (worst score) to '5' (best score) in a five-point Likert scale. Table 4 reports the mean scores from the responses of 40 participants, excluding the responses that did not pass our validation checks. We notice the ground truth motions in ReMoCap achieve the highest score of around $96\\%$ , indicating that participants perceive our captured ground truth motions to be natural-looking, with realistic interactions. ReMoS has the second best ranking of around $78\\%$ , which is almost $21\\%$ higher than the baselines, showing that it is preferred more over the baselines.", + "bbox": [ + 212, + 520, + 787, + 750 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.5 Qualitative Results and Applications", + "text_level": 1, + "bbox": [ + 215, + 768, + 566, + 784 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Fig. 1 shows qualitative results of ReMoS on ReMoCap and highlights the synthesized hand interactions. We note that our actor and reactor are interchangeable, i.e., depending on the character driving the interaction at a given time,", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "ReMoS", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/bfc4dc9cae03538721721780d790826117b56aa80a9fcc4e5b1b9181996bda9b.jpg", + "image_caption": [ + "Ground-truth" + ], + "image_footnote": [], + "bbox": [ + 274, + 143, + 325, + 195 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/f7579fb4cd9f7dfc2674c4ffc0212fe9efcd7564fc0bcd2f47dc81e88ad2800d.jpg", + "image_caption": [ + "MixNMatch" + ], + "image_footnote": [], + "bbox": [ + 334, + 143, + 380, + 195 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/bb67196bc97a616b275cf401e571591afefe51bb76eb9d5cb05e5dde92b180d6.jpg", + "image_caption": [ + "Interformer" + ], + "image_footnote": [], + "bbox": [ + 398, + 143, + 473, + 195 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/032ff87b8b38f6da2b00e069b9ef40687907ec4c3cce76cb1daeb7d898d9ac56.jpg", + "image_caption": [ + "RAIG ComMDM" + ], + "image_footnote": [], + "bbox": [ + 483, + 143, + 524, + 196 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/14e60147d0f2fdfdf80b1156d739e58f438515ce69ea2bf0236b9145ab838b7f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 535, + 143, + 596, + 196 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "ComMDM", + "bbox": [ + 550, + 196, + 584, + 203 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/85ed6bcffaa4653b19a80a1ae70d21d456f37ca4c5cc98efe65b481a7e1995f1.jpg", + "image_caption": [ + "InterGen" + ], + "image_footnote": [], + "bbox": [ + 604, + 143, + 651, + 196 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/65e249ccf60ca8f66ba665a0da719c9724dd576fe427fc9cd1170bd51f232473.jpg", + "image_caption": [ + "ReMoS" + ], + "image_footnote": [], + "bbox": [ + 663, + 143, + 727, + 196 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/0dc11ec5b0b7ff7d5f43e287f9ffe2ef12eb7a47b3639159275186f32cf27710.jpg", + "image_caption": [ + "(a) Visual comparison of ReMoS with GT and baselines.", + "(b) Pose-completion application.", + "Fig.5: Qualitative Results and Applications. We show some visual results and the application of ReMoS as a motion editing tool. (a) The reactor (in blue) synthesized by ReMoS has the most plausible alignment with the actor (in red) compared to the baselines. (b) We manually control the right-hand wrist joint of the reactor and let ReMoS synthesize the remaining body joints conditioned on the actor. (c) ReMoS synthesizes the reactor's motion in-between the start and end frames." + ], + "image_footnote": [], + "bbox": [ + 215, + 220, + 447, + 277 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/abb142b50b8f604756c834a3d51b1fdd5ce50faacdec7509a23abe58e23de257.jpg", + "image_caption": [ + "(c) Motion in-between application." + ], + "image_footnote": [], + "bbox": [ + 555, + 227, + 785, + 276 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "we can swap the actor and the reactor to produce the relevant reactive motions. Fig. 5a shows a visual comparison of ReMoS with the baselines for one frame of Lindy Hop motion. ReMoS synthesizes reaction with the most plausible alignment with the actor's motion. We provide detailed visual results in our supplementary video. Inspired by the applicability of diffusion-based methods for motion editing and controlling, we demonstrate how ReMoS can also be used as an interactive motion editing tool to control the reactor's motion as desired. In Figs. 5b and 5c, we show results from two different motion editing applications, namely, pose completion with controlled joints and motion in-between. We discuss more details in the appendix.", + "bbox": [ + 212, + 414, + 787, + 565 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 584, + 359, + 599 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "ReMoS brings 3D motion conditioned reaction synthesis to a qualitatively new level by generating diverse, well-synchronized reactions for complex movements and plausible hand motions for contact-based interactions. It outperforms the existing baselines both quantitatively and in a user study. We also highlight some practical applications of our model, such as pose completion and motion in-between, which can lead to the development of useful generative assistants for animators, designers, and creative artists. Even though we utilize joint positions as parameterization due to their ready availability, we can adapt our approach to accommodate mesh-level inter-person contacts by introducing an offset to the contact threshold between the two bodies, thus simulating skin-to-skin interactions. We believe there is scope for improvement in the inference speed, where approaches such as Pro-DDPM [17] and DDIM [62] have shown success. Future directions also involve scaling the problem towards multi-person motion prediction [70] and considering scene-aware interactions for the characters, which would further enhance immersive user experiences.", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "A. Ghosh et al.", + "bbox": [ + 271, + 114, + 375, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This research was supported by the EU Horizon 2020 grant $\\mathrm{Carousel + }$ (101017779), and the ERC Consolidator Grant 4DRepLy (770784). We thank Marc Jahan, Christopher Ruf, Michael Hiery, Thomas Leimkuhler and Sascha Huwer for helping with the Ninjutsu data collection, and Noshaba Cheema for helping with the Lindy Hop data collection. We also thank Janis Sprenger and Duarte David for helping with the motion visualizations.", + "bbox": [ + 212, + 176, + 797, + 268 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 292, + 321, + 308 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ahuja, C., Ma, S., Morency, L.P., Sheikh, Y.: To react or not to react: End-to-end visual pose forecasting for personalized avatar during dyadic conversations. In: 2019 International conference on multimodal interaction (2019)", + "2. Ao, T., Zhang, Z., Liu, L.: GestureDiffuCLIP: Gesture diffusion model with clip latents. SIGGRAPH (2023)", + "3. Aristidou, A., Yiannakidis, A., Aberman, K., Cohen-Or, D., Shamir, A., Chrysanthou, Y.: Rhythm is a dancer: Music-driven motion synthesis with global structure. IEEE Transactions on Visualization and Computer Graphics (TVCG) (2022)", + "4. Athanasiou, N., Petrovich, M., Black, M.J., Varol, G.: Teach: Temporal action composition for 3d humans. In: 2022 International Conference on 3D Vision (3DV) (2022)", + "5. Bhattacharya, U., Childs, E., Rewkowski, N., Manocha, D.: Speech2affectivegestures: Synthesizing co-speech gestures with generative adversarial affective expression learning. In: Proceedings of the 29th ACM International Conference on Multimedia (2021)", + "6. Bhattacharya, U., Rewkowski, N., Banerjee, A., Guhan, P., Bera, A., Manocha, D.: Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In: IEEE Conference on Virtual Reality and 3D User Interfaces (IEEE VR) (2021)", + "7. Bjorck, N., Gomes, C.P., Selman, B., Weinberger, K.Q.: Understanding batch normalization. In: Advances in Neural Information Processing Systems (2018)", + "8. https://capturey.com (2023)", + "9. Chan, J.C., Leung, H., Tang, J.K., Komura, T.: A virtual reality dance training system using motion capture technology. IEEE transactions on learning technologies 4(2) (2010)", + "0. Chopin, B., Tang, H., Daoudi, M.: Bipartite graph diffusion model for human interaction generation. In: Winter Conference on Applications of Computer Vision (WACV) (2024)", + "1. Chopin, B., Tang, H., Otberdout, N., Daoudi, M., Sebe, N.: Interaction transformer for human reaction generation. IEEE Transactions on Multimedia (2023)", + "2. Cummins, A.: In search of the ninja: the historical truth of ninjutsu. The History Press (2012)", + "3. Dabral, R., Mughal, M.H., Golyanik, V., Theobalt, C.: Mofusion: A framework for denoising-diffusion-based motion synthesis. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "4. Egges, A., Papagiannakis, G., Magnenat-Thalmann, N.: Presence and interaction in mixed reality environments. The Visual Computer (2007)" + ], + "bbox": [ + 225, + 324, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "ReMoS", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Elfwing, S., Uchibe, E., Doya, K.: Sigmoid-weighted linear units for neural network function approximation in reinforcement learning. Neural Networks (2018)", + "16. Fieraru, M., Zanfir, M., Oneata, E., Popa, A.I., Olaru, V., Sminchisescu, C.: Three-dimensional reconstruction of human interactions. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2020)", + "17. Gandikota, R., Brown, N.: Pro-ddpm: Progressive growing of variable denoising diffusion probabilistic models for faster convergence. In: 33rd British Machine Vision Conference 2022, BMVC (2022)", + "18. Ghosh, A., Cheema, N., Oguz, C., Theobalt, C., Slusallek, P.: Synthesis of compositional animations from textual descriptions. In: International Conference on Computer Vision (ICCV) (2021)", + "19. Ghosh, A., Dabral, R., Golyanik, V., Theobalt, C., Slusallek, P.: IMoS: Intent-driven full-body motion synthesis for human-object interactions. In: Computer Graphics Forum. vol. 42. Wiley Online Library (2023)", + "20. Goel, A., Men, Q., Ho, E.S.L.: Interaction Mix and Match: Synthesizing Close Interaction using Conditional Hierarchical GAN with Multi-Hot Class Embedding. Computer Graphics Forum (2022)", + "21. Gu, D., Shim, J., Jang, J., Kang, C., Joo, K.: Contactgen: Contact-guided interactive 3d human generation for partners. AAAI (2024)", + "22. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: Conference on Computer Vision and Pattern Recognition (2022)", + "23. Guo, C., Zuo, X., Wang, S., Cheng, L.: Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In: European Conference on Computer Vision (2022)", + "24. Guo, W., Bie, X., Alameda-Pineda, X., Moreno-Noguer, F.: Multi-person extreme motion prediction. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "25. Habibie, I., Elgharib, M., Sarkar, K., Abdullah, A., Nyatsanga, S., Neff, M., Theobalt, C.: A motion matching-based framework for controllable gesture synthesis from speech. In: ACM SIGGRAPH Conference Proceedings (2022)", + "26. Hanser, E., Mc Kevitt, P., Lunney, T., Condell, J.: Scenemaker: Intelligent multimodal visualisation of natural language scripts. In: Irish Conference on Artificial Intelligence and Cognitive Science. Springer (2009)", + "27. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)", + "28. Ho, E.S., Komura, T.: Planning tangling motions for humanoids. In: IEEE-RAS International Conference on Humanoid Robots (2007)", + "29. Ho, E.S., Komura, T.: Character motion synthesis by topology coordinates. In: Computer graphics forum. Wiley Online Library (2009)", + "30. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33 (2020)", + "31. Ho, J., Sahara, C., Chan, W., Fleet, D.J., Norouzi, M., Salimans, T.: Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research 23(1) (2022)", + "32. Hu, T., Zhu, X., Guo, W.: Two-person interaction recognition based on key poses. Journal of Computational Information Systems (2014)", + "33. Huang, S., Wang, Z., Li, P., Jia, B., Liu, T., Zhu, Y., Liang, W., Zhu, S.C.: Diffusion-based generation, optimization, and planning in 3d scenes. In: Proceed-" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "A. Ghosh et al.", + "bbox": [ + 271, + 114, + 374, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)", + "34. Huang, Y., Zhang, J., Liu, S., Bao, Q., Zeng, D., Chen, Z., Liu, W.: Genre-conditioned long-term 3d dance generation driven by music. In: International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE (2022)", + "35. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Guided motion diffusion for controllable human motion synthesis. In: International Conference on Computer Vision (ICCV) (2023)", + "36. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)", + "37. Komura, T., Ho, E.S., Lau, R.W.: Animating reactive motion using momentum-based inverse kinematics. Computer Animation and Virtual Worlds (2005)", + "38. Kulkarni, N., Rempe, D., Genova, K., Kundu, A., Johnson, J., Fouhey, D., Guibas, L.: Nifty: Neural object interaction fields for guided human motion synthesis. arXiv preprint arXiv:2307.07511 (2023)", + "39. Kundu, J.N., Buckchash, H., Mandikal, P., Jamkhandi, A., Radhakrishnan, V.B., et al.: Cross-conditioned recurrent networks for long-term synthesis of inter-person human motion interactions. In: Winter Conference on Applications of Computer Vision (WACV) (2020)", + "40. Li, J., Clegg, A., Mottaghi, R., Wu, J., Puig, X., Liu, C.K.: Controllable human-object interaction synthesis. arXiv preprint arXiv:2312.03913 (2023)", + "41. Li, J., Wu, J., Liu, C.K.: Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG) (2023)", + "42. Liang, H., Zhang, W., Li, W., Yu, J., Xu, L.: InterGen: Diffusion-based multi-human motion generation under complex interactions. International Journal for Computer Vision (IJCV) (2024)", + "43. Liu, J., Shahroudy, A., Perez, M., Wang, G., Duan, L.Y., Kot, A.C.: NTU RGB+D 120: A large-scale benchmark for 3d human activity understanding. IEEE transactions on pattern analysis and machine intelligence (2019)", + "44. Liu, X., Yi, L.: Geneoh diffusion: Towards generalizable hand-object interaction denoising via denoising diffusion. In: International Conference on Learning Representations (ICLR) (2024)", + "45. Men, Q., Shum, H.P., Ho, E.S., Leung, H.: Gan-based reactive motion synthesis with class-aware discriminators for human-human interaction. Computers & Graphics (2022)", + "46. Mousas, C.: Performance-driven dance motion control of a virtual partner character. In: IEEE Conference on Virtual Reality and 3D User Interfaces (VR) (2018)", + "47. Mughal, M.H., Dabral, R., Habibie, I., Donatelli, L., Habermann, M., Theobalt, C.: Convofusion: Multi-modal conversational diffusion for co-speech gesture synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024)", + "48. Ng, E., Joo, H., Hu, L., Li, H., Darrell, T., Kanazawa, A., Ginosar, S.: Learning to listen: Modeling non-deterministic dyadic facial motion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)", + "49. Petrovich, M., Black, M.J., Varol, G.: TEMOS: Generating diverse human motions from textual descriptions. In: European Conference on Computer Vision (2022)", + "50. Po, R., Yifan, W., Golyanik, V., Aberman, K., Barron, J.T., Amit H. Bermano, E.R.C., Dekel, T., Holynski, A., Kanazawa, A., Liu, C.K., Liu, L., Mildenhall, B., Nießner, M., Ommer, B., Theobalt, C., Wonka, P., Wetzstein, G.: State of the art on diffusion models for visual computing. arXiv pre-prints (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "ReMoS", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "51. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. In: International Conference on Virtual Reality (2022)", + "52. Rempe, D., Luo, Z., Bin Peng, X., Yuan, Y., Kitani, K., Kreis, K., Fidler, S., Litany, O.: Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "53. Senecal, S., Nijdam, N.A., Aristidou, A., Magnenat-Thalmann, N.: Salsa dance learning evaluation and motion analysis in gamified virtual reality environment. Multimedia Tools and Applications (2020)", + "54. Shafir, Y., Tevet, G., Kapon, R., Bermano, A.H.: Human motion diffusion as a generative prior. In: International Conference on Learning Representations (ICLR) (2024)", + "55. Shen, Y., Yang, L., Ho, E.S.L., Shum, H.P.H.: Interaction-based human activity comparison. IEEE Transactions on Visualization and Computer Graphics (2020)", + "56. Shimada, S., Golyanik, V., Xu, W., Pérez, P., Theobalt, C.: Neural monocular 3d human motion capture with physical awareness. ACM Transactions on Graphics (ToG) (2021)", + "57. Shimada, S., Golyanik, V., Xu, W., Theobalt, C.: Physcap: Physically plausible monocular 3d motion capture in real time. ACM Transactions on Graphics (ToG) (2020)", + "58. Shum, H.P., Komura, T., Shiraishi, M., Yamazaki, S.: Interaction patches for multi-character animation. ACM transactions on graphics (TOG) 27(5) (2008)", + "59. Shum, H.P., Komura, T., Yamazaki, S.: Simulating competitive interactions using singly captured motions. In: Proceedings of ACM symposium on Virtual reality software and technology (2007)", + "60. Siyao, L., Gu, T., Yang, Z., Lin, Z., Liu, Z., Ding, H., Yang, L., Loy, C.C.: Duolando: Follower gpt with off-policy reinforcement learning for dance accompaniment. In: International Conference on Learning Representations (ICLR) (2024)", + "61. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International Conference on Machine Learning (ICML) (2015)", + "62. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)", + "63. Spring, H.: Swing and the lindy hop: dance, venue, media, and tradition. American Music (1997)", + "64. Starke, S., Zhao, Y., Komura, T., Zaman, K.: Local motion phases for learning multi-contact character movements. ACM Transactions on Graphics (TOG) (2020)", + "65. Tanaka, M., Fujiwara, K.: Role-aware interaction generation from textual description. In: International Conference on Computer Vision (ICCV) (2023)", + "66. Tanke, J., Zhang, L., Zhao, A., Tang, C., Cai, Y., Wang, L., Wu, P.C., Gall, J., Keskin, C.: Social diffusion: Long-term multiple human motion anticipation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023)", + "67. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Bermano, A.H., Cohen-Or, D.: Human motion diffusion model. arXiv preprint arXiv:2209.14916 (2022)", + "68. Tseng, J., Castellon, R., Liu, K.: Edge: Editable dance generation from music. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "69. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems (2017)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "A. Ghosh et al.", + "bbox": [ + 271, + 114, + 374, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "70. Wang, J., Xu, H., Narasimhan, M., Wang, X.: Multi-person 3d motion prediction with multi-range transformers. Advances in Neural Information Processing Systems (2021)", + "71. Wang, Z., Chen, Y., Jia, B., Li, P., Zhang, J., Zhang, J., Liu, T., Zhu, Y., Liang, W., Huang, S.: Move as you say, interact as you can: Language-guided human motion generation with scene affordance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024)", + "72. Xie, Y., Jampani, V., Zhong, L., Sun, D., Jiang, H.: OmniControl: Control any joint at any time for human motion generation. In: International Conference on Learning Representations (ICLR) (2024)", + "73. Xing, J., Xia, M., Zhang, Y., Cun, X., Wang, J., Wong, T.T.: Codetalker: Speech-driven 3d facial animation with discrete motion prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)", + "74. Xu, S., Li, Z., Wang, Y.X., Gui, L.Y.: InterDiff: Generating 3d human-object interactions with physics-informed diffusion. In: International Conference on Computer Vision (ICCV) (2023)", + "75. Ye, Y., Li, X., Gupta, A., De Mello, S., Birchfield, S., Song, J., Tulsiani, S., Liu, S.: Affordance diffusion: Synthesizing hand-object interactions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)", + "76. Yoon, Y., Ko, W.R., Jang, M., Lee, J., Kim, J., Lee, G.: Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In: International Conference on Robotics and Automation (ICRA). IEEE (2019)", + "77. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: Physdiff: Physics-guided human motion diffusion model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023)", + "78. Yun, K., Honorio, J., Chattopadhyay, D., Berg, T.L., Samaras, D.: Two-person interaction detection using body-pose features and multiple instance learning. In: IEEE computer society conference on computer vision and pattern recognition workshops (2012)", + "79. Zamfirescu-Pereira, J., Wong, R.Y., Hartmann, B., Yang, Q.: Why johnny can't prompt: How non-ai experts try (and fail) to design LLM prompts. In: Proceedings of Conference on Human Factors in Computing Systems (CHI) (2023)", + "80. Zhang, J., Zhang, Y., Cun, X., Zhang, Y., Zhao, H., Lu, H., Shen, X., Shan, Y.: T2m-gpt: Generating human motion from textual descriptions with discrete representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)", + "81. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022)", + "82. Zhang, W., Dabral, R., Leimkuhler, T., Golyanik, V., Habermann, M., Theobalt, C.: ROAM: Robust and object-aware motion generation using neural pose descriptors. In: International Conference on 3D Vision (3DV) (2024)", + "83. Zhang, X., Bhatnagar, B.L., Starke, S., Guzov, V., Pons-Moll, G.: Couch: Towards controllable human-chair interactions. In: European Conference on Computer Vision (2022)", + "84. Zhou, Z., Wang, B.: Ude: A unified driving engine for human motion generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)", + "85. Zhu, L., Liu, X., Liu, X., Qian, R., Liu, Z., Yu, L.: Taming diffusion models for audio-driven co-speech gesture generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "ReMoS", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/ReMoS_ 3D Motion-Conditioned Reaction Synthesis for Two-Person Interactions/a734f12f-489f-47c4-b3e9-744b3d03eee5_model.json b/2024/ReMoS_ 3D Motion-Conditioned Reaction Synthesis for Two-Person Interactions/a734f12f-489f-47c4-b3e9-744b3d03eee5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..52937a2ff2fa2f0b2c111bff8fc25e2b96c1bbb7 --- /dev/null +++ b/2024/ReMoS_ 3D Motion-Conditioned Reaction Synthesis for Two-Person Interactions/a734f12f-489f-47c4-b3e9-744b3d03eee5_model.json @@ -0,0 +1,2790 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.26, + 0.141, + 0.745, + 0.187 + ], + "angle": 0, + "content": "REMoS: 3D Motion-Conditioned Reaction Synthesis for Two-Person Interactions" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.212, + 0.765, + 0.244 + ], + "angle": 0, + "content": "Anindita Ghosh\\(^{1,2}\\), Rishabh Dabral\\(^{2}\\), Vladislav Golyanik\\(^{2}\\), Christian Theobalt\\(^{2}\\), and Philipp Slusallek\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.3, + 0.255, + 0.703, + 0.298 + ], + "angle": 0, + "content": "\\(^{1}\\) German Research Center for Artificial Intelligence (DFKI) \n\\(^{2}\\) Max-Planck Institute for Informatics (MPII) \n\\(^{3}\\) Saarland Informatics Campus" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.33, + 0.74, + 0.579 + ], + "angle": 0, + "content": "Abstract. Current approaches for 3D human motion synthesis generate high-quality animations of digital humans performing a wide variety of actions and gestures. However, a notable technological gap exists in addressing the complex dynamics of multi-human interactions within this paradigm. In this work, we present ReMoS, a denoising diffusion-based model that synthesizes full-body reactive motion of a person in a two-person interaction scenario. Given the motion of one person, we employ a combined spatio-temporal cross-attention mechanism to synthesize the reactive body and hand motion of the second person, thereby completing the interactions between the two. We demonstrate ReMoS across challenging two-person scenarios such as pair-dancing, Ninjutsu, kickboxing, and acrobatics, where one person's movements have complex and diverse influences on the other. We also contribute the ReMoCap dataset for two-person interactions containing full-body and finger motions. We evaluate ReMoS through multiple quantitative metrics, qualitative visualizations, and a user study, and also indicate usability in interactive motion editing applications. More details are available on the project page: https://vcai.mpi-inf.mpg.de/projects/remos." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.592, + 0.718, + 0.606 + ], + "angle": 0, + "content": "Keywords: Reactive Motion Synthesis \\(\\cdot\\) Denoising Diffusion Model" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.629, + 0.375, + 0.645 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.827 + ], + "angle": 0, + "content": "Digital 3D character motion synthesis has emerged as the next frontier for animation pipelines [50], particularly through denoising diffusion probabilistic models (DDPMs) [61]. While methods for generating character motion for various tasks such as text or music conditioned motion synthesis [3,4,18,22,23,34,49,80], face and gesture synthesis [5,6,25,47,48,73], human-scene interaction [19,71,82,83] exist, synthesizing interactions between humans is relatively under-explored. Modeling such human-human interactions is essential for designing generative 3D human motion synthesis frameworks supporting the complex physical and social interplay of two interacting persons [66]. It offers new capabilities for character animation tools and software, with applications in commercial and entertainment media [26], interactive mixed and augmented reality [14], and social robotics [76]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.128 + ], + "angle": 0, + "content": "A. Ghosh et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.238, + 0.788, + 0.31 + ], + "angle": 0, + "content": "Fig. 1: Visualizations of reactive 3D motion sequences synthesized with the proposed ReMoS approach. We synthesize the 3D full-body motion of the reactor (blue) conditioned only on the 3D motion of the actor (red), thereby completing the interactions between the two (Ninjutsu practice on the left and Lindy Hop dancing on the right). The synthesized hand interactions are enlarged and highlighted with circles." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.789, + 0.537 + ], + "angle": 0, + "content": "Toward this goal, we focus on the task of modeling reactive motions. We aim to automatically generate realistic, temporally-aligned reactive motions of a responding person given the continuous motion of a guiding person (see Fig. 1). This alleviates a common overhead for animators, enabling them to design an acting character and automatically obtain meaningful motions for a reacting character. However, automating such a process poses two key challenges. First, two-person interaction synthesis increases the dimensionality of the already challenging problem of single-person motion synthesis. In our case, the generated reactive motions must align with the conditional signals provided in the form of 3D motion sequences from the actor. Second, doing so in a generative setting without any cues from text prompts or action labels (to obviate additional supervision and data needs) requires a careful trade-off between generating diverse motions and adhering to the narrow manifold of plausible reactions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.789, + 0.78 + ], + "angle": 0, + "content": "While existing methods for synthesizing interactions [10, 20] and two-person motions [42, 54, 65] are good at generating plausible motions, they rely on additional annotations, such as action labels or text prompts, to specify the motion. Action labels typically depict actions at a high level, such as \"high-five\" or \"salsa\", but do not capture the fine-grained synchronization with the actor's motions. Textual descriptions offer nuance for the motions, but collecting accurate textual annotations is considerably cumbersome, particularly for contact-heavy or fast-paced two-person motions, such as dancing and exercising. Further, performing generative tasks through textual prompting remains challenging for non-experts [79]. Apart from the data, contact-based motion generation also needs to consider hand-based interactions to improve realism. This entails the need for frameworks capable of synthesizing hand movements with higher degrees of freedom, which are challenging to model but crucial for improving motion realism. Balancing focus between the broader full-body reactive movements and the finer hand motions involves addressing motion at two significantly different scales — a challenge exacerbated by the high degrees of freedom of the skeletons." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.79, + 0.842 + ], + "angle": 0, + "content": "With this objective in mind, we present ReMoS, a novel approach for Reactive Motion Synthesis with full-body articulations. Inspired by the advancements of DDPMs in 3D human motion synthesis [50,54,65], we develop a DDPM framework (Sec. 3.1) with a cascaded, two-stage generation strategy tailored to our" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.681, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ReMoS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.405 + ], + "angle": 0, + "content": "problem setting (Fig. 2). In the first stage of our diffusion model, we generate the reactive motion for the reactor's body joints conditioned on the actor's body joints. In the second stage, we use the synthesized body motions as an additional parameter to generate masking for appropriate hand motions (Sec. 3.2). We propose a combined spatio-temporal cross-attention (CoST-XA) mapping between the actors' and the reactors' body motion embeddings, that learns the interdependencies in their motions without needing additional annotations. Further, to synthesize hand interactions, we introduce a hand-interaction-aware cross-attention (H-XA) mechanism to ensure the relevant hand joints react to the actors' motions, thus allowing the network to distill localized hand interaction features. We ensure accurate coordination between the actor and the reactor by using contact-based reaction loss (Sec. 3.3) and an inference-time guidance function (Sec. 3.4) that improves the plausibility of the body and hand interactions. To explore reaction synthesis in complex and diverse two-person scenarios, we contribute the REMOCAP dataset consisting of full-body and finger motion sequences for fast-paced swing dance of Lindy Hop [63], and the martial art technique of Ninjutsu [12] (Sec. 4). In summary, our technical contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.405, + 0.788, + 0.48 + ], + "angle": 0, + "content": "- REMoS. A novel method for reactive 3D human motion synthesis using a cascaded diffusion framework to generate full-body and hand motions. Our framework generates fine-grained reactive motions for complex and dynamic contact-based interactions and derives the reactions directly from the actor's motions, without requiring explicit label or text annotations." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.48, + 0.788, + 0.554 + ], + "angle": 0, + "content": "- Interaction-Based Attentions. A combined spatio-temporal cross attention (CoST-XA) mechanism to enforce coherence between the body movements of the actor and the reactor, and a hand-interaction-aware cross-attention (H-XA) mechanism to enforce the appropriate hand-based interactions between the two characters." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.556, + 0.788, + 0.615 + ], + "angle": 0, + "content": "- REMoCAP. A new dataset for two-person interactions under complex scenarios of Lindy Hop dancing and Ninjutsu. The dataset consists of \\(\\sim 275.7\\mathrm{K}\\) motion frames with multiview RGB videos and 3D full-body motion capture of the two interacting persons. It further includes finger-level articulations." + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.405, + 0.788, + 0.615 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.616, + 0.788, + 0.678 + ], + "angle": 0, + "content": "We evaluate our approach in different scenarios, including Lindy Hop dancing, Ninjutsu, Acrobatics [24] and Kickboxing [55], and report state-of-the-art performance (Sec. 5.3). We also report a user study on comparing the visual quality of ReMoS compared to the ground truth and the baseline methods (Sec. 5.4)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.701, + 0.388, + 0.718 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Multi-Person 3D Motion Synthesis. Synthesizing close interactions between two or multiple persons is a challenging task in animation. Early works in multi-person interaction are based on motion graphs [59], interaction patches [58], momentum-based inverse kinematics and motion blending [37], and topologically-based pose representations [28, 29], to name a few. The increasing availability of interaction datasets [16, 39, 43, 53, 55, 78] has led to a rise in data-driven approaches for synthesizing digital partners or opponents in multi-person inter" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.128 + ], + "angle": 0, + "content": "A. Ghosh et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.466 + ], + "angle": 0, + "content": "action settings. Mousas [46] uses a hidden Markov model to control the dance motions of a digital partner in an immersive setup. Ahuja et al. [1] introduce a residual-attention model to generate body pose in a conversation setting conditioned on two audio signals and the opposite person's body pose. Starke et al. [64] propose a phase function-based network that learns asynchronous movements of each bone and its interaction with external objects. Guo et al. [24] propose the Extreme Pose Interaction dataset with a two-stream network with cross-interaction attention modules for forecasting pose sequences of two interacting characters. GAN-based models [20,45] generate motions for an interacting person conditioned on an input character and class labels depicting the type of reaction performed. These methods leverage the daily interaction datasets such as SBU Kinect [78] and 2C [55]. InterFormer [11] uses an interaction transformer with both spatial and temporal attention to generate reactive motions given some initial seed poses of both characters. It can synthesize sparse-level interactions based on the motions of the K3HI [32] and the DuetDance [39] datasets. Concurrent to our work, Duolando [60] uses an off-policy reinforcement learning model to predict tokenized motion for a leader and follower conditioned on music. The aforementioned methods either depend on seed motion as input or require additional conditions to drive the motion. In contrast, ReMoS focuses on synthesizing well-synchronized full-body and hand motions for the reactor conditioned only on the actor's 3D motions, without using any additional labels or prompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.788, + 0.844 + ], + "angle": 0, + "content": "Diffusion Based 3D Motion Synthesis. Denoising diffusion models [61] have recently demonstrated their high potential in generative human motion modeling, specifically in single-person conditional motion synthesis. Conditional single-person motion generation has been performed using diffusion-based approaches for co-speech gestures [2, 85], audio-driven motion [13, 68, 84], and text-driven motion [54,67,77,81] synthesis. Diffusion-based techniques have also been extensively used for conditional synthesis for human-object interactions [38, 40, 41, 74], hand-object interactions [44, 75], and human-scene interactions [33]. Guided motion diffusion models such as GMD [35] and TraceAndPace [52] incorporate spatial constraints on motion trajectories, to guide the motion towards a goal at inference time. OmniControl [72] proposes spatial and realism guidance to control any joint for diffusion-based human motion generation. For two-person interaction synthesis, RAIG [65] proposes a diffusion-based, role-aware approach for two-person interactions, given separate textual descriptions for each person. BiGraphDiff [10] uses graph transformer denoising diffusion to learn two-person interaction conditioned on action labels. ComMDM [54] uses a communication block between two MDMs [67] to coordinate two-person interaction generation. ContactGen [21] presents a diffusion-based contact prediction module that adaptively estimates potential contact regions between two humans according to the interaction label. Liang et al. [42] propose InterGen, a diffusion-based model to synthesize two-person interactions given text descriptions as input conditions. They also propose the InterHuman dataset that includes diverse two-person interaction scenarios with rich text annotations. Inspired by these recent approaches, we design our proposed method ReMoS as a DDPM-based" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.68, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ReMoS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.146, + 0.506, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.145, + 0.788, + 0.258 + ], + "angle": 0, + "content": "Fig.2: ReMoS Overview. Given the motion of the actor (bottom-middle, in red), we synthesize a plausible motion for the reactor (bottom-left, in blue). We achieve this using a denoising diffusion-based probabilistic model (center) trained on reactive motion sequences (top-left, in blue)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.788, + 0.405 + ], + "angle": 0, + "content": "3D motion-conditioned reaction synthesis method, consisting of a cascaded diffusion model with a combined spatio-temporal cross-attention mechanism. This allows ReMoS to implicitly learn the fine-grained synchronization between two interacting persons from only one of the person's motions, without any need for additional prompts or labels, and synthesize the corresponding motions of the second person. ReMoS further synthesizes plausible hand motions for the reactor to incorporate realistic hand-based interactions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.427, + 0.513, + 0.444 + ], + "angle": 0, + "content": "3 Reactive Motion Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.789, + 0.655 + ], + "angle": 0, + "content": "We consider a digital, two-person interaction setting where the motions of one character, the actor, are known, and we need to synthesize the motions of the other character, the reactor. ReMoS aims to synthesize the synchronized reactive motions of the reactor in such a setting. We denote the reactor's motion as \\( X = \\{X_{B},X_{H}\\} \\), where \\( X_{B}\\in \\mathbb{R}^{N\\times J_{B}\\times 3} \\) denotes the 3D full-body joint positions with \\( J_{B} \\) body joints, \\( X_{H}\\in \\mathbb{R}^{N\\times J_{H}\\times 3} \\) denotes the 3D finger joint positions with \\( J_{H} \\) finger joints across both hands, and \\( N \\) denotes the number of frames in the motion sequence. Likewise, we denote the actor's motion as \\( Y = \\{Y_{B},Y_{H}\\} \\). Our goal is to model the conditional probability distribution \\( \\mathrm{P}(X|Y) \\) from which we can sample plausible reactive motions. We model this distribution using conditional denoising diffusion probabilistic models (DDPMs), owing to their unique strengths in capturing temporal coherence of motion data and handling complex motion dynamics [13,67,68]. We elaborate our methodology below." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.674, + 0.579, + 0.691 + ], + "angle": 0, + "content": "3.1 DDPM for Reactive Motion Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.698, + 0.788, + 0.775 + ], + "angle": 0, + "content": "The diffusion process consists of two steps: forward or destructive diffusion, and reverse or denoising diffusion. In the forward process, we progressively corrupt a clean reactive motion sequence \\( X \\) by adding Gaussian noise \\( \\epsilon \\) to it for \\( T \\) steps. With sufficiently small noise and large \\( T \\), we can get \\( X^{(T)} \\sim \\mathcal{N}(0,I) \\) following closed-form formulation of Ho et al. [30]:" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.782, + 0.787, + 0.801 + ], + "angle": 0, + "content": "\\[\nX ^ {(t)} = \\sqrt {\\bar {\\alpha} _ {t}} X ^ {(0)} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\quad \\epsilon \\sim \\mathcal {N} (0, I), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "where \\(\\bar{\\alpha}_t\\) controls the rate of diffusion and \\(t\\in [0,T]\\). Reversing this diffusion process allows for sampling novel motion sequences from a multivariate Gaussian" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.128 + ], + "angle": 0, + "content": "A. Ghosh et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.145, + 0.471, + 0.163 + ], + "angle": 0, + "content": "distribution \\(p\\left(X^{(T)}\\right)\\sim \\mathcal{N}(0,I)\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.349, + 0.176, + 0.786, + 0.217 + ], + "angle": 0, + "content": "\\[\np \\left(X ^ {(0)}\\right) = p \\left(X ^ {(T)}\\right) \\prod_ {t = 1} ^ {T} p \\left(X ^ {(t - 1)} \\mid X ^ {(t)}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.228, + 0.789, + 0.325 + ], + "angle": 0, + "content": "We approximate the computationally intractable term \\( p\\left(X^{(t - 1)}|X^{(t)}\\right) \\) with a learnable function \\( f_{\\theta}\\left(X^{(t)},t\\right) \\), and optimize \\( \\theta \\) to encode the space of human reactive motions. Our reactive motion \\( X \\) is also conditioned on the actor's motion \\( Y \\). Therefore, we modify the learnable function as \\( X^{(t - 1)} = f_{\\theta}\\left(X^{(t)},t,Y\\right) \\). Following recent works [51,67], we estimate the original motion \\( X^{(0)} = f_{\\theta}\\left(X^{(t)},t,Y\\right) \\) from our diffusion model by iterating through all \\( t \\) denoising steps during inference." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.347, + 0.423, + 0.361 + ], + "angle": 0, + "content": "3.2 ReMoS Framework" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.372, + 0.789, + 0.418 + ], + "angle": 0, + "content": "To generate fine-grained reactions with appropriate hand motions, ReMoS decodes the reactive motion of \\( X \\) in a cascaded fashion (Fig. 3). It first estimates the full-body joints \\( X_{B} \\), and then the hand joints \\( X_{H} \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.308, + 0.429, + 0.785, + 0.454 + ], + "angle": 0, + "content": "\\[\nX _ {B} ^ {(0)} = f _ {\\theta_ {B}} \\left(X _ {B} ^ {(t)}, t, Y _ {B}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.308, + 0.456, + 0.785, + 0.482 + ], + "angle": 0, + "content": "\\[\nX _ {H} ^ {(0)} = f _ {\\theta_ {H}} \\left(X _ {H} ^ {(t)}, t, Y _ {H}, \\mathbb {1} _ {H _ {A}} \\left(Y _ {B}\\right) \\mathbb {1} _ {H _ {R}} \\left(X _ {B} ^ {(0)}\\right)\\right), \\quad \\text {a n d} \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.308, + 0.484, + 0.785, + 0.51 + ], + "angle": 0, + "content": "\\[\nX ^ {(0)} = \\left\\{X _ {B} ^ {(0)}, X _ {H} ^ {(0)} \\right\\}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.52, + 0.788, + 0.687 + ], + "angle": 0, + "content": "where \\(\\mathbb{1}_{H_A}\\) and \\(\\mathbb{1}_{H_R}\\) are binary, spatio-temporal hand-interaction mask functions that determine which hand joints of the reactor and actor interact at any given frame. Our cascaded framework comes from the observation that full-body articulations occur at significantly different scales than hand articulations. Since the body motions affect the hand articulations, our cascaded framework feeds the generated body motions as an additional condition to the hand generation module (Fig. 3). We also train the body and the hand generation modules separately, with additional hand-interaction-aware attention mapping for the hand joints. Previous work [19] used such disjoint training strategy on conditional motion synthesis for single persons. To better accommodate for two-person interactions in our case, we benefit from using a cascaded diffusion strategy [31]." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.701, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Body Synthesis Module. In the first stage of our cascaded framework, we diffuse the reactor's body motion, \\( X_B^{(t)} \\), at each diffusion step \\( t \\), and feed it into a transformer decoder [69] block. To condition the reactor's body motion on the actor's body motion \\( Y_B \\), we introduce a combined spatio-temporal cross-attention (CoST-XA) mapping. CoST-XA combines the spatial and temporal features of the actor's and the reactor's motions through an attention matrix to simultaneously learn the inter-dependencies between each pair of reactor and actor body joints \\( (x_{n,j}, y_{n,j}) \\), \\( x_{n,j} \\in X_B^{(t)} \\) and \\( y_{n,j} \\in Y_B \\), respectively, at each frame. Denoting the query from the reactor's motion features as \\( Q_B \\in \\mathbb{R}^{NJ_B \\times 3} \\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.681, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ReMoS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.144, + 0.782, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.279, + 0.788, + 0.351 + ], + "angle": 0, + "content": "Fig. 3: ReMoS Framework. Given the full-body sequence of the actor (left, in red), we input noisy body and hand samples (from below) in a cascaded fashion. We synthesize the body samples first, and use them for hand-interaction-aware attention masking (top-center) to synthesize the denoised hand samples (top-right). The full-body reactive motion is a concatenation of the denoised body and hand samples (right, in blue)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.379, + 0.788, + 0.515 + ], + "angle": 0, + "content": "and the key-value pair from the actor's motion features as \\( K_{B} \\in \\mathbb{R}^{NJ_{B} \\times 3} \\) and \\( V_{B} \\in \\mathbb{R}^{NJ_{B} \\times 3} \\), respectively, we formulate an attention matrix of dimensions \\( NJ_{B} \\times NJ_{B} \\). In contrast, most previous approaches [42,65] use cross-attention only on the temporal features of the motion using an attention matrix of shape \\( N \\times N \\). InterFormer [11] sequentially uses spatial and temporal cross-attention modules instead of combining them, which may result in a partial loss of information on fine-grained, inter-person interactions across time, especially in the absence of additional annotations (prompts or action labels) to condition the reactions. Our proposed cross-attention module, which we define as" + }, + { + "type": "equation", + "bbox": [ + 0.369, + 0.527, + 0.785, + 0.567 + ], + "angle": 0, + "content": "\\[\n\\text {C o S T - X A} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {B} K _ {B} ^ {T}}{\\sqrt {d _ {K _ {B}}}}\\right) V _ {B}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.578, + 0.788, + 0.687 + ], + "angle": 0, + "content": "crucially considers combinations of spatial and temporal interaction features between different body segments of the actor and the reactor to efficiently synchronize their motions. We simultaneously project the diffusion timestep \\( t \\) at each denoising step into the transformer block after the attention blocks and use a final layer of fully-connected network with SiLU activations [15] and batch normalization [7] to generate the body motions. The first stage output is a vector \\( X_B^{(0)} \\in \\mathbb{R}^{N \\times J_B \\times 3} \\) representing the reactor's clean, synthesized body motion." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.703, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Hands Synthesis Module. The second stage of our cascaded framework synthesizes the reactor's hand joints. We input noisy samples \\( X_H^{(t)} \\) into a similar transformer decoder block conditioned on the actor's hand motions \\( Y_H \\). Here, we require the attention weights to be high for the hands used in the hand-based interactions and low for the passive hands. We explicitly enforce this using a spatio-temporal hand-interaction-aware cross-attention mapping (H-XA) to encourage learning localized interaction features. For H-XA to work, we introduce binary hand-interaction masks \\( \\mathbb{1}_{H_A} \\) and \\( \\mathbb{1}_{H_R} \\in \\mathbb{R}^{N \\times J_H} \\). Its entries are calculated by thresholding the distances of the actor's and reactor's wrist joints in \\( Y_B \\) and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "A. Ghosh et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.146, + 0.788, + 0.21 + ], + "angle": 0, + "content": "\\(X_B^{(0)}\\), respectively, to the nearest body joints of the other person. The active entries of \\(\\mathbb{1}_{H_A}\\) and \\(\\mathbb{1}_{H_R}\\) determine which of the actor's and reactor's hands are sufficiently close to the body of the other person (and hence interacting). We then project the encoded \\(Y_H\\) features into each decoder layer using H-XA, as" + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.23, + 0.786, + 0.27 + ], + "angle": 0, + "content": "\\[\n\\mathrm {H} - \\mathrm {X A} = \\operatorname {s o f t m a x} \\left(\\frac {\\left(\\mathbb {1} _ {H _ {R}} \\odot Q _ {H}\\right) \\left(\\mathbb {1} _ {H _ {A}} \\odot K _ {H}\\right) ^ {T}}{\\sqrt {d _ {K _ {H}}}}\\right) V _ {H}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.278, + 0.788, + 0.395 + ], + "angle": 0, + "content": "where \\( K_{H} \\) and \\( V_{H} \\) are the hand motion features from the actor, \\( Q_{H} \\) are the hand motion features from the reactor with \\( d_{K_{H}} \\) channels, and \\( \\odot \\) denotes element-wise product for masking the query and key values. We project the diffusion timestep \\( t \\) at each denoising step into the transformer block and use fully-connected networks with SiLU activations to generate the clean, synthesized reactor hand motions \\( X_{H}^{(0)} \\in \\mathbb{R}^{N \\times J_{H} \\times 3} \\). At the end, to obtain the full reactive motion \\( X^{(0)} \\), we concatenate \\( \\left\\{X_{B}^{(0)}, X_{H}^{(0)}\\right\\} \\) as in Eqn. 5." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.413, + 0.492, + 0.429 + ], + "angle": 0, + "content": "3.3 Losses and Training Details" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.437, + 0.787, + 0.468 + ], + "angle": 0, + "content": "We train ReMoS to minimize a weighted sum of three loss terms, the reconstruction loss \\(\\mathcal{L}_c\\), the reaction loss \\(\\mathcal{L}_r\\), and the kinematic loss \\(\\mathcal{L}_k\\), as" + }, + { + "type": "equation", + "bbox": [ + 0.411, + 0.478, + 0.786, + 0.493 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {c} \\mathcal {L} _ {c} + \\lambda_ {r} \\mathcal {L} _ {r} + \\lambda_ {k} \\mathcal {L} _ {k}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.503, + 0.729, + 0.519 + ], + "angle": 0, + "content": "where \\(\\lambda_{c},\\lambda_{r}\\) and \\(\\lambda_{k}\\) are scalar weights to balance the individual losses." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.531, + 0.788, + 0.607 + ], + "angle": 0, + "content": "Reconstruction Loss \\(\\mathcal{L}_c\\). This is the standard diffusion data term that minimizes the \\(\\ell_2\\)-distance between the ground truth and the synthesized reactive motion, as \\(\\mathcal{L}_c = \\left\\| X - X^{(0)}\\right\\|_2\\). While the reconstruction loss provides the vital data term to drive the training, it does not enforce interaction synchronization between the actor and the reactor, or any kinematic constraints on the motion." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.619, + 0.788, + 0.799 + ], + "angle": 0, + "content": "Reaction loss \\(\\mathcal{L}_r\\). We introduce the reaction loss to ensure accurate timing and spatial positioning of the reactor's motion with respect to the actor's motion. We calculate the ground-truth Euclidean distance \\(d(\\cdot, \\cdot)\\) between each joint of the actor and the reactor at each frame and minimize the deviations from these distances for the actor and the synthesized reactor's motions. We note that this distance term also implicitly constrains the reactor's joints in frames that are less relevant for interactions, i.e., where the reactor's joints are far from the actor's joints. Further, the Euclidean distances for their hand joints in frames with no hand-based contact become drastically high. To mitigate these concerns, we use an exponentially decaying distance-aware weight \\(\\exp(-d(x_{n,j}, y_{n,j}))\\) at each frame \\(n\\) to focus more on the reactor's joints that are closer to the actor, and therefore more relevant for interaction (see Fig. 4). Thus, we get our reaction loss term as" + }, + { + "type": "equation", + "bbox": [ + 0.273, + 0.806, + 0.786, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r} = \\frac {1}{N J} \\sum_ {n = 1} ^ {N} \\sum_ {j = 1} ^ {J} \\exp \\left(- d \\left(x _ {n, j}, y _ {n, j}\\right)\\right) \\cdot \\left| d \\left(x _ {n, j}, y _ {n, j}\\right) - d \\left(x _ {n, j} ^ {(0)}, y _ {n, j}\\right) \\right|. \\tag {9}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.681, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ReMoS" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.157, + 0.534, + 0.198 + ], + "angle": 0, + "content": "Table 1: Dataset Comparisons. Comparing ReMoCap with existing multi-person interaction datasets." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.211, + 0.53, + 0.352 + ], + "angle": 0, + "content": "
Dataset (Chronologically)Motion FramesDuration (hours)Multi-view RGB VideosFinger Articulation
SBU [78]~ 7K0.13XX
NTU-26 [43]~ 22K0.47✓X
2C [55]~ 13K0.06XX
DanceDB [53]~ 1.44M4.00XX
DuetDance [39]~ 196K1.09XX
CHI3D [16]~ 486K2.70✓X
ExPI [24]~ 30K0.33✓X
InterHuman [42]~ 107M6.56XX
DD100 [60]~ 200K1.92X✓
ReMoCap (ours)~ 275.7K2.04✓✓
" + }, + { + "type": "image", + "bbox": [ + 0.579, + 0.144, + 0.765, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.268, + 0.787, + 0.348 + ], + "angle": 0, + "content": "Fig.4: Visualization of Distance Aware Reaction Loss. We use an exponentially decaying distance-aware reaction loss to focus more on the reactor's joints that are closer to the actor." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.392, + 0.788, + 0.469 + ], + "angle": 0, + "content": "Kinematic Loss \\(\\mathcal{L}_k\\). To maintain the kinematic plausibility of the generated motions, we follow existing literature on regularizers for bone length consistency, foot contacts, and temporal consistency [56,57,68]. Our kinematic loss term is a weighted sum of the joint velocity loss \\(\\mathcal{L}_{vel}\\), the joint acceleration loss \\(\\mathcal{L}_{acc}\\), the bone length consistency loss \\(\\mathcal{L}_{bone}\\), and the foot sliding loss \\(\\mathcal{L}_{foot}\\), as" + }, + { + "type": "equation", + "bbox": [ + 0.343, + 0.48, + 0.785, + 0.498 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {k} = \\lambda_ {v} \\mathcal {L} _ {v e l} + \\lambda_ {a} \\mathcal {L} _ {a c c} + \\lambda_ {b} \\mathcal {L} _ {b o n e} + \\lambda_ {f} \\mathcal {L} _ {f o o t}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.508, + 0.752, + 0.525 + ], + "angle": 0, + "content": "where \\(\\lambda_v, \\lambda_a, \\lambda_b\\) and \\(\\lambda_f\\) are scalar weights (more details in the appendix)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.546, + 0.535, + 0.561 + ], + "angle": 0, + "content": "3.4 Inference Time Spatial Guidance" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.572, + 0.788, + 0.782 + ], + "angle": 0, + "content": "While our method synthesizes plausible reactive motions, it can sometimes spatially misalign the reactor's body to the actor's, especially at the arm joints, for fast-paced, contact-heavy interactions. This, in turn, affects finger-joint synthesis in the second stage of our cascaded diffusion. To improve spatial alignment, we leverage guidance functions [35, 52] that can provide gradients to nudge the sampling process towards a certain direction. We design a guidance function \\( G\\left(\\phi ,\\hat{\\phi}\\right)\\in \\mathbb{R}^{N\\times J_A\\times 3} \\), which minimizes the distance between the \\( J_{A} \\) arm joints of the actor \\( (\\phi) \\) and the \\( J_{A} \\) arm joints of the synthesized reactor \\( (\\hat{\\phi}) \\). Specifically, we re-apply our interaction masks \\( \\mathbb{1}_{H_A} \\) and \\( \\mathbb{1}_{H_R} \\) to determine which of the reactor's hands are more likely to be interacting with the actor, and minimize the distances between the arm joints of the corresponding sides as \\( G = \\arg \\min_{\\hat{\\phi}}\\left(\\left\\| \\mathbb{1}_{H_A}\\odot \\phi -\\mathbb{1}_{H_R}\\odot \\hat{\\phi}\\right\\|\\right) \\). We then plug \\( G \\) into the denoising pipeline of our body diffusion module, as" + }, + { + "type": "equation", + "bbox": [ + 0.297, + 0.793, + 0.786, + 0.819 + ], + "angle": 0, + "content": "\\[\nX _ {B} ^ {(0)} = X _ {B} ^ {(0)} - \\gamma \\nabla_ {X _ {B} ^ {(0)}} G (\\phi , \\hat {\\phi}), \\quad \\gamma = \\text {g u i d a n c e s c a l e}. \\tag {11}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.127 + ], + "angle": 0, + "content": "A. Ghosh et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.432, + 0.163 + ], + "angle": 0, + "content": "4 ReMoCap Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.317 + ], + "angle": 0, + "content": "We propose the ReMoCap dataset to facilitate research on contact-based two-person interactions with finger-level articulations. It contains complex two-person interactions in two separate scenarios: the fast-paced, swing-style Lindy Hop dancing, and the martial arts of Ninjutsu. We consider these two types of motions for studying physical interactions for two main reasons. First, these motions encompass a significantly diverse spectrum of motions in terms of intensity, speed, and style. Second, choreographed motion generation is developing as an active field of research and we envision this dataset to facilitate research in interactive applications such as performing in virtual reality [9] and remote tutoring [53]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.331, + 0.789, + 0.574 + ], + "angle": 0, + "content": "Data Collection. We invited 4 trained Lindy Hop dancers and 5 trained Ninjutsu artists to perform diverse interactions in a multi-view capture studio. We tracked their motions using a commercially available markerless, multi-view motion-capture system [8], which tracks 93 degrees-of-freedom driving 69 body joints. Capturing motions in a markerless setting enables the performers to move uninhibited, while also making the data suitable for training monocular or multiview motion capture methods under severe inter-person occlusions. We also capture RGB videos from 120 camera views for each sequence. The dataset includes 3D skeleton poses with full-body and finger annotations, foreground-background segmentation masks, and 3D surface reconstructions of the subjects. We capture sequences of multiple lengths totaling \\(\\sim 275.7\\mathrm{K}\\) frames (2.04 hours) of motion data from each view at a frame rate of 50 fps for the Lindy Hop and 25 fps for the Ninjutsu. Out of all the frames, around \\(150\\mathrm{K}\\) frames have hand-based interactions between the two characters where the closest distance between the finger joints of the actor and the reactor is within \\(50~\\mathrm{mm}\\). We present more dataset statistics in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.589, + 0.789, + 0.817 + ], + "angle": 0, + "content": "Dataset Comparison. Table 1 shows a comprehensive comparison of ReMoCap with existing multi-person interaction datasets. Existing datasets consisting of two-person interactions, such as SBU [78], K3HI [32], NTU-26 [43] and 2C [55], are limited in size and motion capture quality. They typically feature simple actions, such as handshakes, punching, pushing and kicking, with weak interactions, and do not capture hand motions. The recent ExPI [24] dataset captures Lindy Hop aerial sequences to model more complex interactions, and the InterHuman [42] dataset features both daily motions (e.g., passing objects, greeting, communicating) and professional motions (e.g., Taekwondo, Latin dance, boxing). However, these datasets do not provide finger-level motion capture data, which is a key requirement to intricately model inter-human activities. Two-person dance datasets, such as DuetDance [39] and DanceDB [53], also lack the hand motion data needed for modeling interactions. Only the concurrent work, Duolando \\([60]^*\\) proposes the DD100 dataset with strong interactions between two dancers and provides hand motion data." + }, + { + "type": "page_footnote", + "bbox": [ + 0.238, + 0.826, + 0.571, + 0.84 + ], + "angle": 0, + "content": "*unpublished at the time of our paper submission" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.681, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ReMoS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.496, + 0.163 + ], + "angle": 0, + "content": "5 Experiments and Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.283 + ], + "angle": 0, + "content": "We conduct comprehensive experiments to evaluate ReMoS on multiple two-person datasets covering a wide range of interaction scenarios. We perform training and evaluation on multiple large-scale datasets, including 2C [55], ExPI [24], and our proposed ReMoCap. We provide the implementation details, report quantitative and qualitative comparisons on standard evaluation metrics and through a user study, report ablations, and show how to apply ReMoS as a motion editing tool. We provide details on the dataset preparation in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.303, + 0.457, + 0.318 + ], + "angle": 0, + "content": "5.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.327, + 0.789, + 0.616 + ], + "angle": 0, + "content": "As a pre-processing step, we normalize the two-person body poses by translating the actor's motion \\( Y \\) and the reactor's motion \\( X \\) together, such that the root joint of \\( Y \\) is at the global origin for all \\( N \\) frames. We then compute the relative 3D joint coordinates of \\( X_B \\) and \\( Y_B \\) w.r.t. the root joint of \\( Y_B \\). For each hand, we compute the relative 3D joint coordinates of \\( X_H \\) and \\( Y_H \\) w.r.t. the corresponding wrist joints of \\( X_H \\) and \\( Y_H \\). Using normalized inter-person, root-relative joint positions benefits the stability and convergence of our model. ReMoS uses \\( T = 500 \\) diffusion steps where \\( \\bar{\\alpha} \\) changes linearly from 0.0002 to 0.02. We train for about 64K iterations on both sequences of ReMoCap using Adam [36] with a base learning rate of \\( 10^{-5} \\) and a batch size of 64. We decay the learning rate using a Step LR scheduler with a step size of 5 epochs and a decay rate of 0.99. We use \\( d = 256 \\) for our latent embedding representation and use 6 layers in our transformer decoder with 4 heads for calculating the attention. The training takes around 8 and 11 hours on an NVIDIA RTX A4000 GPU for Lindy Hop and Ninjutsu, respectively. The inference time is \\( \\sim 24s \\) to generate 50 frames of full-body and finger motions (12.5s for body synthesis and then 11.5s for hand synthesis). We set \\( \\lambda_c, \\lambda_r, \\lambda_v = 10.0, \\lambda_a, \\lambda_k, \\lambda_b = 1.0 \\) and \\( \\lambda_f = 20.0 \\) (only applied after 100 training epochs) as the loss term weights in Eqns. 8 and 10. We set the guidance scale \\( \\gamma = 10^{-3} \\) in Eqn. 11." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.635, + 0.521, + 0.649 + ], + "angle": 0, + "content": "5.2 Baselines and Ablated Versions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.795 + ], + "angle": 0, + "content": "For baselines, we choose the closest motion synthesis methods in a two-person setting, namely, MixNMatch [20], InterFormer [11], ComMDM [54], Role-Aware Interaction Generation (RAIG) [65], and InterGen [42]. InterFormer [11] was originally trained in a reactive motion synthesis setting without additional annotations, such as action labels or text descriptions, and we maintain this setup. For the other methods, we mask out their input text/label embeddings to comply with our annotation-free setting. We re-train these methods on ReMoCap with a thorough hyper-parameter search and report their best performances. We provide more training details in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.796, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We also compare our proposed ReMoS model with five of its ablated versions: w/o diffusion. Training ReMoS using a transformer encoder-decoder network without using DDPM strategy." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "A. Ghosh et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 2: Quantitative Evaluation on ReMoCap. We compare ReMoS with our baselines and ablated versions (Sec. 5.2) on the ReMoCap dataset. We evaluate these methods on metrics such as MPJPE, MPJVE, FID, and Diversity. \\(\\downarrow\\): lower is better, \\(\\uparrow\\): higher is better, \\(\\rightarrow\\): values closer to GT are better. Bold indicates best." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.213, + 0.782, + 0.406 + ], + "angle": 0, + "content": "
MethodLindy HopNinjutsu
MPJPE (mm) ↓MPJVE (mm) ↓FID ↓ (body)FID ↓ (hands)Div →MPJPE (mm) ↓MPJVE (mm) ↓FID ↓ (body)FID ↓ (hands)Div →
GT----7.57----10.51
InterFormer [11]66.68.260.530.654.54270.23.40.570.686.48
MixNMatch [20]70.210.30.770.782.48257.25.20.740.724.83
ComMDM [54]59.44.410.320.537.48201.24.10.340.589.98
RAIG [65]71.24.320.470.638.45199.15.10.210.6310.11
InterGen [42]62.63.920.300.617.21172.63.90.320.579.98
ReMoS (ours)40.72.260.120.267.62139.23.30.160.3510.26
w/o diffusion72.54.910.580.744.04224.54.10.520.646.06
w/o cascading63.94.950.510.557.12223.64.20.420.756.62
w/o CoST-XA44.23.620.210.397.45176.63.60.270.418.91
w/o reaction loss44.63.510.220.387.31144.63.70.230.398.99
w/o spatial guidance41.92.340.120.267.62139.43.40.160.3510.26
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.433, + 0.785, + 0.461 + ], + "angle": 0, + "content": "- w/o cascading. Training ReMoS with an integrated transformer for all the joints instead of a cascaded strategy for body and hand generation." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.463, + 0.785, + 0.492 + ], + "angle": 0, + "content": "- w/o CoST-XA. Using a spatial followed by a temporal cross-attention mapping instead of the combined spatio-temporal features for body synthesis." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.494, + 0.73, + 0.507 + ], + "angle": 0, + "content": "- w/o reaction loss. Removing the reaction loss (Eqn. 9) in training." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.509, + 0.784, + 0.524 + ], + "angle": 0, + "content": "- w/o spatial guidance. Removing spatial guidance \\(G_{A}\\) (Sec. 3.4) in inference." + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.433, + 0.785, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.545, + 0.46, + 0.56 + ], + "angle": 0, + "content": "5.3 Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.789, + 0.842 + ], + "angle": 0, + "content": "We evaluate ReMoS on standard evaluation metrics. We measure the temporal consistency deviation from the ground truth 3D pose following [57] to report the mean per-joint positional error (MPJPE) and the mean per-frame, per-joint velocity error (MPJVE), both in \\( mm \\), on the synthesized motions. We measure the Fréchet Inception Distance (FID) [27] to compare the distribution gap between the embedding spaces of the generated and ground-truth motions. As part of our method focuses on generating hand motions, we measure the FID score of body and hands separately for ReMoCap. We also compute the latent variance of the generated motions (Diversity) [67, 81]. Table 2 reports the quantitative evaluation of ReMoS with its baselines and ablations on ReMoCap. ReMoS achieves state-of-the-art performance for both the Lindy Hop pair-dance setting and the Ninjutsu setting. We observe that methods using denoising diffusion have higher Diversity compared to transformer-based [11] or GAN-based [20] methods. This enforces the variability claims of denoising diffusion models. We note at least \\( 20\\% \\) improvement in the MPJPE and around \\( 40\\% \\) improvement in the FID score for the reactor's body motion when using the proposed CoST-XA mechanism, confirming its benefit in learning fine-grained, inter-person dependencies across time. We also note an almost \\( 50\\% \\) improvement in the FID scores of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.681, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ReMoS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.588, + 0.214 + ], + "angle": 0, + "content": "Table 3: Quantitative Evaluation on the ExPI and 2C datasets. We compare ReMoS with state-of-the-art motion synthesis methods on the ExPI [24] and 2C datasets [55]. \\(\\downarrow\\): lower is better, \\(\\uparrow\\): higher is better, \\(\\rightarrow\\): values closer to GT are better. Bold indicates best." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.226, + 0.589, + 0.341 + ], + "angle": 0, + "content": "
MethodExPI2C
MPJPE (mm) ↓MPJVE (mm) ↓FID ↓ (body) →Div
MPJPE (mm) ↓MPJVE (mm) ↓FID ↓ (body) →Div
GT---2.01---2.22
InterFormer [11]99.13.560.421.3190.75.110.521.45
MixNMatch [20]122.45.560.481.1862.46.010.471.24
ComMDM [54]121.45.410.452.4869.93.340.492.86
RAIG [65]131.23.960.532.5191.64.920.674.45
InterGen [42]100.63.910.432.0967.64.010.472.91
ReMoS (ours)97.93.520.411.9859.13.330.342.07
" + }, + { + "type": "table_caption", + "bbox": [ + 0.6, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 4: User Study Results. Mean scores on a five-point Likert scale (scores 1 - 5)." + }, + { + "type": "table", + "bbox": [ + 0.601, + 0.213, + 0.786, + 0.326 + ], + "angle": 0, + "content": "
MethodMotion Quality ↑Reaction Plausibility ↑
GT4.86 ± 0.544.72 ± 0.56
InterFormer2.52 ± 0.612.28 ± 0.57
MixNMatch1.92 ± 0.712.18 ± 0.57
ComMDM3.02 ± 0.473.12 ± 0.52
RAIG2.83 ± 0.672.48 ± 0.65
InterGen3.18 ± 0.573.19 ± 0.53
ReMoS3.79 ± 0.553.88 ± 0.54
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.369, + 0.788, + 0.475 + ], + "angle": 0, + "content": "the reactor's hand motion when using the cascaded strategy and the reaction loss. Further, the spatial guidance function fine-tuning improves MPJPE and MPJVE. The MPJPE values are overall higher in Ninjutsu than in Lindy Hop as the trajectory of the reactor is more diverse for Ninjutsu. We also report the evaluation of ReMoS with its baselines on the ExPI [24] and the 2C [55] datasets in Table 3. These datasets do not provide hand motions, so we only evaluate the reactor's body motions, and report state-of-the-art performance of ReMoS." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.496, + 0.357, + 0.511 + ], + "angle": 0, + "content": "5.4 User Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.789, + 0.75 + ], + "angle": 0, + "content": "We evaluate the visual quality of our captured ground truth data and the generated reactive motions through a user study. We show participants 26 different interaction sequences across the ground truth, our method, and its baselines. For each sequence, we randomly show them three methods side-by-side and ask them to rate the 3D motions they observe in terms of \\((a)\\) the reactor's motion quality, irrespective of the actor's motion (Motion Quality), and \\((b)\\) the plausibility of the reaction given the actor's motion (Reaction Plausibility). We ask the participants to rank these motions from '1' (worst score) to '5' (best score) in a five-point Likert scale. Table 4 reports the mean scores from the responses of 40 participants, excluding the responses that did not pass our validation checks. We notice the ground truth motions in ReMoCap achieve the highest score of around \\(96\\%\\), indicating that participants perceive our captured ground truth motions to be natural-looking, with realistic interactions. ReMoS has the second best ranking of around \\(78\\%\\), which is almost \\(21\\%\\) higher than the baselines, showing that it is preferred more over the baselines." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.77, + 0.567, + 0.785 + ], + "angle": 0, + "content": "5.5 Qualitative Results and Applications" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Fig. 1 shows qualitative results of ReMoS on ReMoCap and highlights the synthesized hand interactions. We note that our actor and reactor are interchangeable, i.e., depending on the character driving the interaction at a given time," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.128 + ], + "angle": 0, + "content": "A. Ghosh et al." + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.144, + 0.326, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.279, + 0.197, + 0.319, + 0.203 + ], + "angle": 0, + "content": "Ground-truth" + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.144, + 0.382, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.197, + 0.378, + 0.203 + ], + "angle": 0, + "content": "MixNMatch" + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.144, + 0.474, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.419, + 0.197, + 0.453, + 0.204 + ], + "angle": 0, + "content": "Interformer" + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.144, + 0.526, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.198, + 0.586, + 0.205 + ], + "angle": 0, + "content": "RAIG ComMDM" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.144, + 0.598, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.198, + 0.586, + 0.204 + ], + "angle": 0, + "content": "ComMDM" + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.144, + 0.653, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.198, + 0.645, + 0.204 + ], + "angle": 0, + "content": "InterGen" + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.144, + 0.728, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.687, + 0.198, + 0.711, + 0.204 + ], + "angle": 0, + "content": "ReMoS" + }, + { + "type": "image_caption", + "bbox": [ + 0.339, + 0.21, + 0.666, + 0.222 + ], + "angle": 0, + "content": "(a) Visual comparison of ReMoS with GT and baselines." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.222, + 0.449, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.237, + 0.28, + 0.428, + 0.293 + ], + "angle": 0, + "content": "(b) Pose-completion application." + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.228, + 0.787, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.28, + 0.781, + 0.293 + ], + "angle": 0, + "content": "(c) Motion in-between application." + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.303, + 0.789, + 0.387 + ], + "angle": 0, + "content": "Fig.5: Qualitative Results and Applications. We show some visual results and the application of ReMoS as a motion editing tool. (a) The reactor (in blue) synthesized by ReMoS has the most plausible alignment with the actor (in red) compared to the baselines. (b) We manually control the right-hand wrist joint of the reactor and let ReMoS synthesize the remaining body joints conditioned on the actor. (c) ReMoS synthesizes the reactor's motion in-between the start and end frames." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.789, + 0.566 + ], + "angle": 0, + "content": "we can swap the actor and the reactor to produce the relevant reactive motions. Fig. 5a shows a visual comparison of ReMoS with the baselines for one frame of Lindy Hop motion. ReMoS synthesizes reaction with the most plausible alignment with the actor's motion. We provide detailed visual results in our supplementary video. Inspired by the applicability of diffusion-based methods for motion editing and controlling, we demonstrate how ReMoS can also be used as an interactive motion editing tool to control the reactor's motion as desired. In Figs. 5b and 5c, we show results from two different motion editing applications, namely, pose completion with controlled joints and motion in-between. We discuss more details in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.585, + 0.36, + 0.601 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.841 + ], + "angle": 0, + "content": "ReMoS brings 3D motion conditioned reaction synthesis to a qualitatively new level by generating diverse, well-synchronized reactions for complex movements and plausible hand motions for contact-based interactions. It outperforms the existing baselines both quantitatively and in a user study. We also highlight some practical applications of our model, such as pose completion and motion in-between, which can lead to the development of useful generative assistants for animators, designers, and creative artists. Even though we utilize joint positions as parameterization due to their ready availability, we can adapt our approach to accommodate mesh-level inter-person contacts by introducing an offset to the contact threshold between the two bodies, thus simulating skin-to-skin interactions. We believe there is scope for improvement in the inference speed, where approaches such as Pro-DDPM [17] and DDIM [62] have shown success. Future directions also involve scaling the problem towards multi-person motion prediction [70] and considering scene-aware interactions for the characters, which would further enhance immersive user experiences." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.681, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ReMoS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.799, + 0.269 + ], + "angle": 0, + "content": "This research was supported by the EU Horizon 2020 grant \\(\\mathrm{Carousel + }\\) (101017779), and the ERC Consolidator Grant 4DRepLy (770784). We thank Marc Jahan, Christopher Ruf, Michael Hiery, Thomas Leimkuhler and Sascha Huwer for helping with the Ninjutsu data collection, and Noshaba Cheema for helping with the Lindy Hop data collection. We also thank Janis Sprenger and Duarte David for helping with the motion visualizations." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.293, + 0.323, + 0.309 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.325, + 0.785, + 0.367 + ], + "angle": 0, + "content": "1. Ahuja, C., Ma, S., Morency, L.P., Sheikh, Y.: To react or not to react: End-to-end visual pose forecasting for personalized avatar during dyadic conversations. In: 2019 International conference on multimodal interaction (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.368, + 0.785, + 0.394 + ], + "angle": 0, + "content": "2. Ao, T., Zhang, Z., Liu, L.: GestureDiffuCLIP: Gesture diffusion model with clip latents. SIGGRAPH (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.396, + 0.785, + 0.436 + ], + "angle": 0, + "content": "3. Aristidou, A., Yiannakidis, A., Aberman, K., Cohen-Or, D., Shamir, A., Chrysanthou, Y.: Rhythm is a dancer: Music-driven motion synthesis with global structure. IEEE Transactions on Visualization and Computer Graphics (TVCG) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.438, + 0.785, + 0.478 + ], + "angle": 0, + "content": "4. Athanasiou, N., Petrovich, M., Black, M.J., Varol, G.: Teach: Temporal action composition for 3d humans. In: 2022 International Conference on 3D Vision (3DV) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.48, + 0.785, + 0.534 + ], + "angle": 0, + "content": "5. Bhattacharya, U., Childs, E., Rewkowski, N., Manocha, D.: Speech2affectivegestures: Synthesizing co-speech gestures with generative adversarial affective expression learning. In: Proceedings of the 29th ACM International Conference on Multimedia (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.535, + 0.785, + 0.589 + ], + "angle": 0, + "content": "6. Bhattacharya, U., Rewkowski, N., Banerjee, A., Guhan, P., Bera, A., Manocha, D.: Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In: IEEE Conference on Virtual Reality and 3D User Interfaces (IEEE VR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.591, + 0.785, + 0.618 + ], + "angle": 0, + "content": "7. Bjorck, N., Gomes, C.P., Selman, B., Weinberger, K.Q.: Understanding batch normalization. In: Advances in Neural Information Processing Systems (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.619, + 0.44, + 0.631 + ], + "angle": 0, + "content": "8. https://capturey.com (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.785, + 0.673 + ], + "angle": 0, + "content": "9. Chan, J.C., Leung, H., Tang, J.K., Komura, T.: A virtual reality dance training system using motion capture technology. IEEE transactions on learning technologies 4(2) (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "0. Chopin, B., Tang, H., Daoudi, M.: Bipartite graph diffusion model for human interaction generation. In: Winter Conference on Applications of Computer Vision (WACV) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.785, + 0.743 + ], + "angle": 0, + "content": "1. Chopin, B., Tang, H., Otberdout, N., Daoudi, M., Sebe, N.: Interaction transformer for human reaction generation. IEEE Transactions on Multimedia (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.785, + 0.77 + ], + "angle": 0, + "content": "2. Cummins, A.: In search of the ninja: the historical truth of ninjutsu. The History Press (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "3. Dabral, R., Mughal, M.H., Golyanik, V., Theobalt, C.: Mofusion: A framework for denoising-diffusion-based motion synthesis. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "4. Egges, A., Papagiannakis, G., Magnenat-Thalmann, N.: Presence and interaction in mixed reality environments. The Visual Computer (2007)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.325, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.127 + ], + "angle": 0, + "content": "A. Ghosh et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "15. Elfwing, S., Uchibe, E., Doya, K.: Sigmoid-weighted linear units for neural network function approximation in reinforcement learning. Neural Networks (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.217 + ], + "angle": 0, + "content": "16. Fieraru, M., Zanfir, M., Oneata, E., Popa, A.I., Olaru, V., Sminchisescu, C.: Three-dimensional reconstruction of human interactions. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.258 + ], + "angle": 0, + "content": "17. Gandikota, R., Brown, N.: Pro-ddpm: Progressive growing of variable denoising diffusion probabilistic models for faster convergence. In: 33rd British Machine Vision Conference 2022, BMVC (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.259, + 0.785, + 0.3 + ], + "angle": 0, + "content": "18. Ghosh, A., Cheema, N., Oguz, C., Theobalt, C., Slusallek, P.: Synthesis of compositional animations from textual descriptions. In: International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "19. Ghosh, A., Dabral, R., Golyanik, V., Theobalt, C., Slusallek, P.: IMoS: Intent-driven full-body motion synthesis for human-object interactions. In: Computer Graphics Forum. vol. 42. Wiley Online Library (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.383 + ], + "angle": 0, + "content": "20. Goel, A., Men, Q., Ho, E.S.L.: Interaction Mix and Match: Synthesizing Close Interaction using Conditional Hierarchical GAN with Multi-Hot Class Embedding. Computer Graphics Forum (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.383, + 0.785, + 0.411 + ], + "angle": 0, + "content": "21. Gu, D., Shim, J., Jang, J., Kang, C., Joo, K.: Contactgen: Contact-guided interactive 3d human generation for partners. AAAI (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.452 + ], + "angle": 0, + "content": "22. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: Conference on Computer Vision and Pattern Recognition (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.493 + ], + "angle": 0, + "content": "23. Guo, C., Zuo, X., Wang, S., Cheng, L.: Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In: European Conference on Computer Vision (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.785, + 0.536 + ], + "angle": 0, + "content": "24. Guo, W., Bie, X., Alameda-Pineda, X., Moreno-Noguer, F.: Multi-person extreme motion prediction. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "25. Habibie, I., Elgharib, M., Sarkar, K., Abdullah, A., Nyatsanga, S., Neff, M., Theobalt, C.: A motion matching-based framework for controllable gesture synthesis from speech. In: ACM SIGGRAPH Conference Proceedings (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.577, + 0.785, + 0.618 + ], + "angle": 0, + "content": "26. Hanser, E., Mc Kevitt, P., Lunney, T., Condell, J.: Scenemaker: Intelligent multimodal visualisation of natural language scripts. In: Irish Conference on Artificial Intelligence and Cognitive Science. Springer (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.66 + ], + "angle": 0, + "content": "27. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.785, + 0.688 + ], + "angle": 0, + "content": "28. Ho, E.S., Komura, T.: Planning tangling motions for humanoids. In: IEEE-RAS International Conference on Humanoid Robots (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.688, + 0.785, + 0.716 + ], + "angle": 0, + "content": "29. Ho, E.S., Komura, T.: Character motion synthesis by topology coordinates. In: Computer graphics forum. Wiley Online Library (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.743 + ], + "angle": 0, + "content": "30. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.785 + ], + "angle": 0, + "content": "31. Ho, J., Sahara, C., Chan, W., Fleet, D.J., Norouzi, M., Salimans, T.: Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research 23(1) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.785, + 0.813 + ], + "angle": 0, + "content": "32. Hu, T., Zhu, X., Guo, W.: Two-person interaction recognition based on key poses. Journal of Computational Information Systems (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "33. Huang, S., Wang, Z., Li, P., Jia, B., Liu, T., Zhu, Y., Liang, W., Zhu, S.C.: Diffusion-based generation, optimization, and planning in 3d scenes. In: Proceed-" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.681, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ReMoS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.147, + 0.786, + 0.175 + ], + "angle": 0, + "content": "ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.787, + 0.218 + ], + "angle": 0, + "content": "34. Huang, Y., Zhang, J., Liu, S., Bao, Q., Zeng, D., Chen, Z., Liu, W.: Genre-conditioned long-term 3d dance generation driven by music. In: International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.787, + 0.259 + ], + "angle": 0, + "content": "35. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Guided motion diffusion for controllable human motion synthesis. In: International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.786, + 0.286 + ], + "angle": 0, + "content": "36. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.786, + 0.314 + ], + "angle": 0, + "content": "37. Komura, T., Ho, E.S., Lau, R.W.: Animating reactive motion using momentum-based inverse kinematics. Computer Animation and Virtual Worlds (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.786, + 0.355 + ], + "angle": 0, + "content": "38. Kulkarni, N., Rempe, D., Genova, K., Kundu, A., Johnson, J., Fouhey, D., Guibas, L.: Nifty: Neural object interaction fields for guided human motion synthesis. arXiv preprint arXiv:2307.07511 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.786, + 0.412 + ], + "angle": 0, + "content": "39. Kundu, J.N., Buckchash, H., Mandikal, P., Jamkhandi, A., Radhakrishnan, V.B., et al.: Cross-conditioned recurrent networks for long-term synthesis of inter-person human motion interactions. In: Winter Conference on Applications of Computer Vision (WACV) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.786, + 0.439 + ], + "angle": 0, + "content": "40. Li, J., Clegg, A., Mottaghi, R., Wu, J., Puig, X., Liu, C.K.: Controllable human-object interaction synthesis. arXiv preprint arXiv:2312.03913 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.786, + 0.466 + ], + "angle": 0, + "content": "41. Li, J., Wu, J., Liu, C.K.: Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.786, + 0.508 + ], + "angle": 0, + "content": "42. Liang, H., Zhang, W., Li, W., Yu, J., Xu, L.: InterGen: Diffusion-based multi-human motion generation under complex interactions. International Journal for Computer Vision (IJCV) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.786, + 0.55 + ], + "angle": 0, + "content": "43. Liu, J., Shahroudy, A., Perez, M., Wang, G., Duan, L.Y., Kot, A.C.: NTU RGB+D 120: A large-scale benchmark for 3d human activity understanding. IEEE transactions on pattern analysis and machine intelligence (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.786, + 0.591 + ], + "angle": 0, + "content": "44. Liu, X., Yi, L.: Geneoh diffusion: Towards generalizable hand-object interaction denoising via denoising diffusion. In: International Conference on Learning Representations (ICLR) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.786, + 0.633 + ], + "angle": 0, + "content": "45. Men, Q., Shum, H.P., Ho, E.S., Leung, H.: Gan-based reactive motion synthesis with class-aware discriminators for human-human interaction. Computers & Graphics (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.786, + 0.661 + ], + "angle": 0, + "content": "46. Mousas, C.: Performance-driven dance motion control of a virtual partner character. In: IEEE Conference on Virtual Reality and 3D User Interfaces (VR) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.786, + 0.715 + ], + "angle": 0, + "content": "47. Mughal, M.H., Dabral, R., Habibie, I., Donatelli, L., Habermann, M., Theobalt, C.: Convofusion: Multi-modal conversational diffusion for co-speech gesture synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.786, + 0.758 + ], + "angle": 0, + "content": "48. Ng, E., Joo, H., Hu, L., Li, H., Darrell, T., Kanazawa, A., Ginosar, S.: Learning to listen: Modeling non-deterministic dyadic facial motion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.786, + 0.785 + ], + "angle": 0, + "content": "49. Petrovich, M., Black, M.J., Varol, G.: TEMOS: Generating diverse human motions from textual descriptions. In: European Conference on Computer Vision (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.786, + 0.84 + ], + "angle": 0, + "content": "50. Po, R., Yifan, W., Golyanik, V., Aberman, K., Barron, J.T., Amit H. Bermano, E.R.C., Dekel, T., Holynski, A., Kanazawa, A., Liu, C.K., Liu, L., Mildenhall, B., Nießner, M., Ommer, B., Theobalt, C., Wonka, P., Wetzstein, G.: State of the art on diffusion models for visual computing. arXiv pre-prints (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.127 + ], + "angle": 0, + "content": "A. Ghosh et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "51. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. In: International Conference on Virtual Reality (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.243 + ], + "angle": 0, + "content": "52. Rempe, D., Luo, Z., Bin Peng, X., Yuan, Y., Kitani, K., Kreis, K., Fidler, S., Litany, O.: Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.785, + 0.285 + ], + "angle": 0, + "content": "53. Senecal, S., Nijdam, N.A., Aristidou, A., Magnenat-Thalmann, N.: Salsa dance learning evaluation and motion analysis in gamified virtual reality environment. Multimedia Tools and Applications (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.785, + 0.325 + ], + "angle": 0, + "content": "54. Shafir, Y., Tevet, G., Kapon, R., Bermano, A.H.: Human motion diffusion as a generative prior. In: International Conference on Learning Representations (ICLR) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.353 + ], + "angle": 0, + "content": "55. Shen, Y., Yang, L., Ho, E.S.L., Shum, H.P.H.: Interaction-based human activity comparison. IEEE Transactions on Visualization and Computer Graphics (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.353, + 0.785, + 0.393 + ], + "angle": 0, + "content": "56. Shimada, S., Golyanik, V., Xu, W., Pérez, P., Theobalt, C.: Neural monocular 3d human motion capture with physical awareness. ACM Transactions on Graphics (ToG) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.785, + 0.434 + ], + "angle": 0, + "content": "57. Shimada, S., Golyanik, V., Xu, W., Theobalt, C.: Physcap: Physically plausible monocular 3d motion capture in real time. ACM Transactions on Graphics (ToG) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.785, + 0.462 + ], + "angle": 0, + "content": "58. Shum, H.P., Komura, T., Shiraishi, M., Yamazaki, S.: Interaction patches for multi-character animation. ACM transactions on graphics (TOG) 27(5) (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.462, + 0.785, + 0.502 + ], + "angle": 0, + "content": "59. Shum, H.P., Komura, T., Yamazaki, S.: Simulating competitive interactions using singly captured motions. In: Proceedings of ACM symposium on Virtual reality software and technology (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.543 + ], + "angle": 0, + "content": "60. Siyao, L., Gu, T., Yang, Z., Lin, Z., Liu, Z., Ding, H., Yang, L., Loy, C.C.: Duolando: Follower gpt with off-policy reinforcement learning for dance accompaniment. In: International Conference on Learning Representations (ICLR) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.543, + 0.785, + 0.583 + ], + "angle": 0, + "content": "61. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International Conference on Machine Learning (ICML) (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.785, + 0.61 + ], + "angle": 0, + "content": "62. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.785, + 0.637 + ], + "angle": 0, + "content": "63. Spring, H.: Swing and the lindy hop: dance, venue, media, and tradition. American Music (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.664 + ], + "angle": 0, + "content": "64. Starke, S., Zhao, Y., Komura, T., Zaman, K.: Local motion phases for learning multi-contact character movements. ACM Transactions on Graphics (TOG) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.785, + 0.692 + ], + "angle": 0, + "content": "65. Tanaka, M., Fujiwara, K.: Role-aware interaction generation from textual description. In: International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.692, + 0.785, + 0.746 + ], + "angle": 0, + "content": "66. Tanke, J., Zhang, L., Zhao, A., Tang, C., Cai, Y., Wang, L., Wu, P.C., Gall, J., Keskin, C.: Social diffusion: Long-term multiple human motion anticipation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.746, + 0.785, + 0.773 + ], + "angle": 0, + "content": "67. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Bermano, A.H., Cohen-Or, D.: Human motion diffusion model. arXiv preprint arXiv:2209.14916 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.773, + 0.785, + 0.8 + ], + "angle": 0, + "content": "68. Tseng, J., Castellon, R., Liu, K.: Edge: Editable dance generation from music. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "69. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.681, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ReMoS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "70. Wang, J., Xu, H., Narasimhan, M., Wang, X.: Multi-person 3d motion prediction with multi-range transformers. Advances in Neural Information Processing Systems (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.244 + ], + "angle": 0, + "content": "71. Wang, Z., Chen, Y., Jia, B., Li, P., Zhang, J., Zhang, J., Liu, T., Zhu, Y., Liang, W., Huang, S.: Move as you say, interact as you can: Language-guided human motion generation with scene affordance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.244, + 0.788, + 0.284 + ], + "angle": 0, + "content": "72. Xie, Y., Jampani, V., Zhong, L., Sun, D., Jiang, H.: OmniControl: Control any joint at any time for human motion generation. In: International Conference on Learning Representations (ICLR) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.284, + 0.788, + 0.325 + ], + "angle": 0, + "content": "73. Xing, J., Xia, M., Zhang, Y., Cun, X., Wang, J., Wong, T.T.: Codetalker: Speech-driven 3d facial animation with discrete motion prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.325, + 0.788, + 0.365 + ], + "angle": 0, + "content": "74. Xu, S., Li, Z., Wang, Y.X., Gui, L.Y.: InterDiff: Generating 3d human-object interactions with physics-informed diffusion. In: International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.406 + ], + "angle": 0, + "content": "75. Ye, Y., Li, X., Gupta, A., De Mello, S., Birchfield, S., Song, J., Tulsiani, S., Liu, S.: Affordance diffusion: Synthesizing hand-object interactions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.406, + 0.788, + 0.447 + ], + "angle": 0, + "content": "76. Yoon, Y., Ko, W.R., Jang, M., Lee, J., Kim, J., Lee, G.: Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In: International Conference on Robotics and Automation (ICRA). IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.447, + 0.788, + 0.488 + ], + "angle": 0, + "content": "77. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: Physdiff: Physics-guided human motion diffusion model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.488, + 0.788, + 0.542 + ], + "angle": 0, + "content": "78. Yun, K., Honorio, J., Chattopadhyay, D., Berg, T.L., Samaras, D.: Two-person interaction detection using body-pose features and multiple instance learning. In: IEEE computer society conference on computer vision and pattern recognition workshops (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.542, + 0.788, + 0.583 + ], + "angle": 0, + "content": "79. Zamfirescu-Pereira, J., Wong, R.Y., Hartmann, B., Yang, Q.: Why johnny can't prompt: How non-ai experts try (and fail) to design LLM prompts. In: Proceedings of Conference on Human Factors in Computing Systems (CHI) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.637 + ], + "angle": 0, + "content": "80. Zhang, J., Zhang, Y., Cun, X., Zhang, Y., Zhao, H., Lu, H., Shen, X., Shan, Y.: T2m-gpt: Generating human motion from textual descriptions with discrete representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.637, + 0.788, + 0.678 + ], + "angle": 0, + "content": "81. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.678, + 0.788, + 0.718 + ], + "angle": 0, + "content": "82. Zhang, W., Dabral, R., Leimkuhler, T., Golyanik, V., Habermann, M., Theobalt, C.: ROAM: Robust and object-aware motion generation using neural pose descriptors. In: International Conference on 3D Vision (3DV) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.718, + 0.788, + 0.759 + ], + "angle": 0, + "content": "83. Zhang, X., Bhatnagar, B.L., Starke, S., Guzov, V., Pons-Moll, G.: Couch: Towards controllable human-chair interactions. In: European Conference on Computer Vision (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.759, + 0.788, + 0.8 + ], + "angle": 0, + "content": "84. Zhou, Z., Wang, B.: Ude: A unified driving engine for human motion generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.8, + 0.788, + 0.841 + ], + "angle": 0, + "content": "85. Zhu, L., Liu, X., Liu, X., Qian, R., Liu, Z., Yu, L.: Taming diffusion models for audio-driven co-speech gesture generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/ReMoS_ 3D Motion-Conditioned Reaction Synthesis for Two-Person Interactions/a734f12f-489f-47c4-b3e9-744b3d03eee5_origin.pdf b/2024/ReMoS_ 3D Motion-Conditioned Reaction Synthesis for Two-Person Interactions/a734f12f-489f-47c4-b3e9-744b3d03eee5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5e2e06396aec5bdc67d5591f901154c21524e79a --- /dev/null +++ b/2024/ReMoS_ 3D Motion-Conditioned Reaction Synthesis for Two-Person Interactions/a734f12f-489f-47c4-b3e9-744b3d03eee5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad93849d724a27acd5056574a65eca822725264ff29364eb7eac26b6572dca09 +size 5630982 diff --git a/2024/ReMoS_ 3D Motion-Conditioned Reaction Synthesis for Two-Person Interactions/full.md b/2024/ReMoS_ 3D Motion-Conditioned Reaction Synthesis for Two-Person Interactions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8907d3d07e4ba92f9ab304788ad8118250d8c6fd --- /dev/null +++ b/2024/ReMoS_ 3D Motion-Conditioned Reaction Synthesis for Two-Person Interactions/full.md @@ -0,0 +1,335 @@ +# REMoS: 3D Motion-Conditioned Reaction Synthesis for Two-Person Interactions + +Anindita Ghosh $^{1,2}$ , Rishabh Dabral $^{2}$ , Vladislav Golyanik $^{2}$ , Christian Theobalt $^{2}$ , and Philipp Slusallek $^{1,2}$ + +$^{1}$ German Research Center for Artificial Intelligence (DFKI) + $^{2}$ Max-Planck Institute for Informatics (MPII) + $^{3}$ Saarland Informatics Campus + +Abstract. Current approaches for 3D human motion synthesis generate high-quality animations of digital humans performing a wide variety of actions and gestures. However, a notable technological gap exists in addressing the complex dynamics of multi-human interactions within this paradigm. In this work, we present ReMoS, a denoising diffusion-based model that synthesizes full-body reactive motion of a person in a two-person interaction scenario. Given the motion of one person, we employ a combined spatio-temporal cross-attention mechanism to synthesize the reactive body and hand motion of the second person, thereby completing the interactions between the two. We demonstrate ReMoS across challenging two-person scenarios such as pair-dancing, Ninjutsu, kickboxing, and acrobatics, where one person's movements have complex and diverse influences on the other. We also contribute the ReMoCap dataset for two-person interactions containing full-body and finger motions. We evaluate ReMoS through multiple quantitative metrics, qualitative visualizations, and a user study, and also indicate usability in interactive motion editing applications. More details are available on the project page: https://vcai.mpi-inf.mpg.de/projects/remos. + +Keywords: Reactive Motion Synthesis $\cdot$ Denoising Diffusion Model + +# 1 Introduction + +Digital 3D character motion synthesis has emerged as the next frontier for animation pipelines [50], particularly through denoising diffusion probabilistic models (DDPMs) [61]. While methods for generating character motion for various tasks such as text or music conditioned motion synthesis [3,4,18,22,23,34,49,80], face and gesture synthesis [5,6,25,47,48,73], human-scene interaction [19,71,82,83] exist, synthesizing interactions between humans is relatively under-explored. Modeling such human-human interactions is essential for designing generative 3D human motion synthesis frameworks supporting the complex physical and social interplay of two interacting persons [66]. It offers new capabilities for character animation tools and software, with applications in commercial and entertainment media [26], interactive mixed and augmented reality [14], and social robotics [76]. + +![](images/234de5a2090ce50bb3f0bc7b8e396378ce68b906d95bbf314c3d408a4a4671d5.jpg) +Fig. 1: Visualizations of reactive 3D motion sequences synthesized with the proposed ReMoS approach. We synthesize the 3D full-body motion of the reactor (blue) conditioned only on the 3D motion of the actor (red), thereby completing the interactions between the two (Ninjutsu practice on the left and Lindy Hop dancing on the right). The synthesized hand interactions are enlarged and highlighted with circles. + +Toward this goal, we focus on the task of modeling reactive motions. We aim to automatically generate realistic, temporally-aligned reactive motions of a responding person given the continuous motion of a guiding person (see Fig. 1). This alleviates a common overhead for animators, enabling them to design an acting character and automatically obtain meaningful motions for a reacting character. However, automating such a process poses two key challenges. First, two-person interaction synthesis increases the dimensionality of the already challenging problem of single-person motion synthesis. In our case, the generated reactive motions must align with the conditional signals provided in the form of 3D motion sequences from the actor. Second, doing so in a generative setting without any cues from text prompts or action labels (to obviate additional supervision and data needs) requires a careful trade-off between generating diverse motions and adhering to the narrow manifold of plausible reactions. + +While existing methods for synthesizing interactions [10, 20] and two-person motions [42, 54, 65] are good at generating plausible motions, they rely on additional annotations, such as action labels or text prompts, to specify the motion. Action labels typically depict actions at a high level, such as "high-five" or "salsa", but do not capture the fine-grained synchronization with the actor's motions. Textual descriptions offer nuance for the motions, but collecting accurate textual annotations is considerably cumbersome, particularly for contact-heavy or fast-paced two-person motions, such as dancing and exercising. Further, performing generative tasks through textual prompting remains challenging for non-experts [79]. Apart from the data, contact-based motion generation also needs to consider hand-based interactions to improve realism. This entails the need for frameworks capable of synthesizing hand movements with higher degrees of freedom, which are challenging to model but crucial for improving motion realism. Balancing focus between the broader full-body reactive movements and the finer hand motions involves addressing motion at two significantly different scales — a challenge exacerbated by the high degrees of freedom of the skeletons. + +With this objective in mind, we present ReMoS, a novel approach for Reactive Motion Synthesis with full-body articulations. Inspired by the advancements of DDPMs in 3D human motion synthesis [50,54,65], we develop a DDPM framework (Sec. 3.1) with a cascaded, two-stage generation strategy tailored to our + +problem setting (Fig. 2). In the first stage of our diffusion model, we generate the reactive motion for the reactor's body joints conditioned on the actor's body joints. In the second stage, we use the synthesized body motions as an additional parameter to generate masking for appropriate hand motions (Sec. 3.2). We propose a combined spatio-temporal cross-attention (CoST-XA) mapping between the actors' and the reactors' body motion embeddings, that learns the interdependencies in their motions without needing additional annotations. Further, to synthesize hand interactions, we introduce a hand-interaction-aware cross-attention (H-XA) mechanism to ensure the relevant hand joints react to the actors' motions, thus allowing the network to distill localized hand interaction features. We ensure accurate coordination between the actor and the reactor by using contact-based reaction loss (Sec. 3.3) and an inference-time guidance function (Sec. 3.4) that improves the plausibility of the body and hand interactions. To explore reaction synthesis in complex and diverse two-person scenarios, we contribute the REMOCAP dataset consisting of full-body and finger motion sequences for fast-paced swing dance of Lindy Hop [63], and the martial art technique of Ninjutsu [12] (Sec. 4). In summary, our technical contributions are: + +- REMoS. A novel method for reactive 3D human motion synthesis using a cascaded diffusion framework to generate full-body and hand motions. Our framework generates fine-grained reactive motions for complex and dynamic contact-based interactions and derives the reactions directly from the actor's motions, without requiring explicit label or text annotations. +- Interaction-Based Attentions. A combined spatio-temporal cross attention (CoST-XA) mechanism to enforce coherence between the body movements of the actor and the reactor, and a hand-interaction-aware cross-attention (H-XA) mechanism to enforce the appropriate hand-based interactions between the two characters. +- REMoCAP. A new dataset for two-person interactions under complex scenarios of Lindy Hop dancing and Ninjutsu. The dataset consists of $\sim 275.7\mathrm{K}$ motion frames with multiview RGB videos and 3D full-body motion capture of the two interacting persons. It further includes finger-level articulations. + +We evaluate our approach in different scenarios, including Lindy Hop dancing, Ninjutsu, Acrobatics [24] and Kickboxing [55], and report state-of-the-art performance (Sec. 5.3). We also report a user study on comparing the visual quality of ReMoS compared to the ground truth and the baseline methods (Sec. 5.4). + +# 2 Related Work + +Multi-Person 3D Motion Synthesis. Synthesizing close interactions between two or multiple persons is a challenging task in animation. Early works in multi-person interaction are based on motion graphs [59], interaction patches [58], momentum-based inverse kinematics and motion blending [37], and topologically-based pose representations [28, 29], to name a few. The increasing availability of interaction datasets [16, 39, 43, 53, 55, 78] has led to a rise in data-driven approaches for synthesizing digital partners or opponents in multi-person inter + +action settings. Mousas [46] uses a hidden Markov model to control the dance motions of a digital partner in an immersive setup. Ahuja et al. [1] introduce a residual-attention model to generate body pose in a conversation setting conditioned on two audio signals and the opposite person's body pose. Starke et al. [64] propose a phase function-based network that learns asynchronous movements of each bone and its interaction with external objects. Guo et al. [24] propose the Extreme Pose Interaction dataset with a two-stream network with cross-interaction attention modules for forecasting pose sequences of two interacting characters. GAN-based models [20,45] generate motions for an interacting person conditioned on an input character and class labels depicting the type of reaction performed. These methods leverage the daily interaction datasets such as SBU Kinect [78] and 2C [55]. InterFormer [11] uses an interaction transformer with both spatial and temporal attention to generate reactive motions given some initial seed poses of both characters. It can synthesize sparse-level interactions based on the motions of the K3HI [32] and the DuetDance [39] datasets. Concurrent to our work, Duolando [60] uses an off-policy reinforcement learning model to predict tokenized motion for a leader and follower conditioned on music. The aforementioned methods either depend on seed motion as input or require additional conditions to drive the motion. In contrast, ReMoS focuses on synthesizing well-synchronized full-body and hand motions for the reactor conditioned only on the actor's 3D motions, without using any additional labels or prompts. + +Diffusion Based 3D Motion Synthesis. Denoising diffusion models [61] have recently demonstrated their high potential in generative human motion modeling, specifically in single-person conditional motion synthesis. Conditional single-person motion generation has been performed using diffusion-based approaches for co-speech gestures [2, 85], audio-driven motion [13, 68, 84], and text-driven motion [54,67,77,81] synthesis. Diffusion-based techniques have also been extensively used for conditional synthesis for human-object interactions [38, 40, 41, 74], hand-object interactions [44, 75], and human-scene interactions [33]. Guided motion diffusion models such as GMD [35] and TraceAndPace [52] incorporate spatial constraints on motion trajectories, to guide the motion towards a goal at inference time. OmniControl [72] proposes spatial and realism guidance to control any joint for diffusion-based human motion generation. For two-person interaction synthesis, RAIG [65] proposes a diffusion-based, role-aware approach for two-person interactions, given separate textual descriptions for each person. BiGraphDiff [10] uses graph transformer denoising diffusion to learn two-person interaction conditioned on action labels. ComMDM [54] uses a communication block between two MDMs [67] to coordinate two-person interaction generation. ContactGen [21] presents a diffusion-based contact prediction module that adaptively estimates potential contact regions between two humans according to the interaction label. Liang et al. [42] propose InterGen, a diffusion-based model to synthesize two-person interactions given text descriptions as input conditions. They also propose the InterHuman dataset that includes diverse two-person interaction scenarios with rich text annotations. Inspired by these recent approaches, we design our proposed method ReMoS as a DDPM-based + +![](images/9d441678b2f9f21acf62a1b4370d336ba081f4cc3395a4fbc8383021159f9deb.jpg) +Fig.2: ReMoS Overview. Given the motion of the actor (bottom-middle, in red), we synthesize a plausible motion for the reactor (bottom-left, in blue). We achieve this using a denoising diffusion-based probabilistic model (center) trained on reactive motion sequences (top-left, in blue). + +3D motion-conditioned reaction synthesis method, consisting of a cascaded diffusion model with a combined spatio-temporal cross-attention mechanism. This allows ReMoS to implicitly learn the fine-grained synchronization between two interacting persons from only one of the person's motions, without any need for additional prompts or labels, and synthesize the corresponding motions of the second person. ReMoS further synthesizes plausible hand motions for the reactor to incorporate realistic hand-based interactions. + +# 3 Reactive Motion Synthesis + +We consider a digital, two-person interaction setting where the motions of one character, the actor, are known, and we need to synthesize the motions of the other character, the reactor. ReMoS aims to synthesize the synchronized reactive motions of the reactor in such a setting. We denote the reactor's motion as $X = \{X_{B},X_{H}\}$ , where $X_{B}\in \mathbb{R}^{N\times J_{B}\times 3}$ denotes the 3D full-body joint positions with $J_{B}$ body joints, $X_{H}\in \mathbb{R}^{N\times J_{H}\times 3}$ denotes the 3D finger joint positions with $J_{H}$ finger joints across both hands, and $N$ denotes the number of frames in the motion sequence. Likewise, we denote the actor's motion as $Y = \{Y_{B},Y_{H}\}$ . Our goal is to model the conditional probability distribution $\mathrm{P}(X|Y)$ from which we can sample plausible reactive motions. We model this distribution using conditional denoising diffusion probabilistic models (DDPMs), owing to their unique strengths in capturing temporal coherence of motion data and handling complex motion dynamics [13,67,68]. We elaborate our methodology below. + +# 3.1 DDPM for Reactive Motion Synthesis + +The diffusion process consists of two steps: forward or destructive diffusion, and reverse or denoising diffusion. In the forward process, we progressively corrupt a clean reactive motion sequence $X$ by adding Gaussian noise $\epsilon$ to it for $T$ steps. With sufficiently small noise and large $T$ , we can get $X^{(T)} \sim \mathcal{N}(0,I)$ following closed-form formulation of Ho et al. [30]: + +$$ +X ^ {(t)} = \sqrt {\bar {\alpha} _ {t}} X ^ {(0)} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon , \quad \epsilon \sim \mathcal {N} (0, I), \tag {1} +$$ + +where $\bar{\alpha}_t$ controls the rate of diffusion and $t\in [0,T]$ . Reversing this diffusion process allows for sampling novel motion sequences from a multivariate Gaussian + +distribution $p\left(X^{(T)}\right)\sim \mathcal{N}(0,I)$ as + +$$ +p \left(X ^ {(0)}\right) = p \left(X ^ {(T)}\right) \prod_ {t = 1} ^ {T} p \left(X ^ {(t - 1)} \mid X ^ {(t)}\right). \tag {2} +$$ + +We approximate the computationally intractable term $p\left(X^{(t - 1)}|X^{(t)}\right)$ with a learnable function $f_{\theta}\left(X^{(t)},t\right)$ , and optimize $\theta$ to encode the space of human reactive motions. Our reactive motion $X$ is also conditioned on the actor's motion $Y$ . Therefore, we modify the learnable function as $X^{(t - 1)} = f_{\theta}\left(X^{(t)},t,Y\right)$ . Following recent works [51,67], we estimate the original motion $X^{(0)} = f_{\theta}\left(X^{(t)},t,Y\right)$ from our diffusion model by iterating through all $t$ denoising steps during inference. + +# 3.2 ReMoS Framework + +To generate fine-grained reactions with appropriate hand motions, ReMoS decodes the reactive motion of $X$ in a cascaded fashion (Fig. 3). It first estimates the full-body joints $X_{B}$ , and then the hand joints $X_{H}$ as + +$$ +X _ {B} ^ {(0)} = f _ {\theta_ {B}} \left(X _ {B} ^ {(t)}, t, Y _ {B}\right), \tag {3} +$$ + +$$ +X _ {H} ^ {(0)} = f _ {\theta_ {H}} \left(X _ {H} ^ {(t)}, t, Y _ {H}, \mathbb {1} _ {H _ {A}} \left(Y _ {B}\right) \mathbb {1} _ {H _ {R}} \left(X _ {B} ^ {(0)}\right)\right), \quad \text {a n d} \tag {4} +$$ + +$$ +X ^ {(0)} = \left\{X _ {B} ^ {(0)}, X _ {H} ^ {(0)} \right\}, \tag {5} +$$ + +where $\mathbb{1}_{H_A}$ and $\mathbb{1}_{H_R}$ are binary, spatio-temporal hand-interaction mask functions that determine which hand joints of the reactor and actor interact at any given frame. Our cascaded framework comes from the observation that full-body articulations occur at significantly different scales than hand articulations. Since the body motions affect the hand articulations, our cascaded framework feeds the generated body motions as an additional condition to the hand generation module (Fig. 3). We also train the body and the hand generation modules separately, with additional hand-interaction-aware attention mapping for the hand joints. Previous work [19] used such disjoint training strategy on conditional motion synthesis for single persons. To better accommodate for two-person interactions in our case, we benefit from using a cascaded diffusion strategy [31]. + +Body Synthesis Module. In the first stage of our cascaded framework, we diffuse the reactor's body motion, $X_B^{(t)}$ , at each diffusion step $t$ , and feed it into a transformer decoder [69] block. To condition the reactor's body motion on the actor's body motion $Y_B$ , we introduce a combined spatio-temporal cross-attention (CoST-XA) mapping. CoST-XA combines the spatial and temporal features of the actor's and the reactor's motions through an attention matrix to simultaneously learn the inter-dependencies between each pair of reactor and actor body joints $(x_{n,j}, y_{n,j})$ , $x_{n,j} \in X_B^{(t)}$ and $y_{n,j} \in Y_B$ , respectively, at each frame. Denoting the query from the reactor's motion features as $Q_B \in \mathbb{R}^{NJ_B \times 3}$ + +![](images/91b1996d2b0933da10018d5f4abc699bdb32155efbbd32d044eeffe9d7002479.jpg) +Fig. 3: ReMoS Framework. Given the full-body sequence of the actor (left, in red), we input noisy body and hand samples (from below) in a cascaded fashion. We synthesize the body samples first, and use them for hand-interaction-aware attention masking (top-center) to synthesize the denoised hand samples (top-right). The full-body reactive motion is a concatenation of the denoised body and hand samples (right, in blue). + +and the key-value pair from the actor's motion features as $K_{B} \in \mathbb{R}^{NJ_{B} \times 3}$ and $V_{B} \in \mathbb{R}^{NJ_{B} \times 3}$ , respectively, we formulate an attention matrix of dimensions $NJ_{B} \times NJ_{B}$ . In contrast, most previous approaches [42,65] use cross-attention only on the temporal features of the motion using an attention matrix of shape $N \times N$ . InterFormer [11] sequentially uses spatial and temporal cross-attention modules instead of combining them, which may result in a partial loss of information on fine-grained, inter-person interactions across time, especially in the absence of additional annotations (prompts or action labels) to condition the reactions. Our proposed cross-attention module, which we define as + +$$ +\text {C o S T - X A} = \operatorname {s o f t m a x} \left(\frac {Q _ {B} K _ {B} ^ {T}}{\sqrt {d _ {K _ {B}}}}\right) V _ {B}, \tag {6} +$$ + +crucially considers combinations of spatial and temporal interaction features between different body segments of the actor and the reactor to efficiently synchronize their motions. We simultaneously project the diffusion timestep $t$ at each denoising step into the transformer block after the attention blocks and use a final layer of fully-connected network with SiLU activations [15] and batch normalization [7] to generate the body motions. The first stage output is a vector $X_B^{(0)} \in \mathbb{R}^{N \times J_B \times 3}$ representing the reactor's clean, synthesized body motion. + +Hands Synthesis Module. The second stage of our cascaded framework synthesizes the reactor's hand joints. We input noisy samples $X_H^{(t)}$ into a similar transformer decoder block conditioned on the actor's hand motions $Y_H$ . Here, we require the attention weights to be high for the hands used in the hand-based interactions and low for the passive hands. We explicitly enforce this using a spatio-temporal hand-interaction-aware cross-attention mapping (H-XA) to encourage learning localized interaction features. For H-XA to work, we introduce binary hand-interaction masks $\mathbb{1}_{H_A}$ and $\mathbb{1}_{H_R} \in \mathbb{R}^{N \times J_H}$ . Its entries are calculated by thresholding the distances of the actor's and reactor's wrist joints in $Y_B$ and + +$X_B^{(0)}$ , respectively, to the nearest body joints of the other person. The active entries of $\mathbb{1}_{H_A}$ and $\mathbb{1}_{H_R}$ determine which of the actor's and reactor's hands are sufficiently close to the body of the other person (and hence interacting). We then project the encoded $Y_H$ features into each decoder layer using H-XA, as + +$$ +\mathrm {H} - \mathrm {X A} = \operatorname {s o f t m a x} \left(\frac {\left(\mathbb {1} _ {H _ {R}} \odot Q _ {H}\right) \left(\mathbb {1} _ {H _ {A}} \odot K _ {H}\right) ^ {T}}{\sqrt {d _ {K _ {H}}}}\right) V _ {H}, \tag {7} +$$ + +where $K_{H}$ and $V_{H}$ are the hand motion features from the actor, $Q_{H}$ are the hand motion features from the reactor with $d_{K_{H}}$ channels, and $\odot$ denotes element-wise product for masking the query and key values. We project the diffusion timestep $t$ at each denoising step into the transformer block and use fully-connected networks with SiLU activations to generate the clean, synthesized reactor hand motions $X_{H}^{(0)} \in \mathbb{R}^{N \times J_{H} \times 3}$ . At the end, to obtain the full reactive motion $X^{(0)}$ , we concatenate $\left\{X_{B}^{(0)}, X_{H}^{(0)}\right\}$ as in Eqn. 5. + +# 3.3 Losses and Training Details + +We train ReMoS to minimize a weighted sum of three loss terms, the reconstruction loss $\mathcal{L}_c$ , the reaction loss $\mathcal{L}_r$ , and the kinematic loss $\mathcal{L}_k$ , as + +$$ +\mathcal {L} = \lambda_ {c} \mathcal {L} _ {c} + \lambda_ {r} \mathcal {L} _ {r} + \lambda_ {k} \mathcal {L} _ {k}, \tag {8} +$$ + +where $\lambda_{c},\lambda_{r}$ and $\lambda_{k}$ are scalar weights to balance the individual losses. + +Reconstruction Loss $\mathcal{L}_c$ . This is the standard diffusion data term that minimizes the $\ell_2$ -distance between the ground truth and the synthesized reactive motion, as $\mathcal{L}_c = \left\| X - X^{(0)}\right\|_2$ . While the reconstruction loss provides the vital data term to drive the training, it does not enforce interaction synchronization between the actor and the reactor, or any kinematic constraints on the motion. + +Reaction loss $\mathcal{L}_r$ . We introduce the reaction loss to ensure accurate timing and spatial positioning of the reactor's motion with respect to the actor's motion. We calculate the ground-truth Euclidean distance $d(\cdot, \cdot)$ between each joint of the actor and the reactor at each frame and minimize the deviations from these distances for the actor and the synthesized reactor's motions. We note that this distance term also implicitly constrains the reactor's joints in frames that are less relevant for interactions, i.e., where the reactor's joints are far from the actor's joints. Further, the Euclidean distances for their hand joints in frames with no hand-based contact become drastically high. To mitigate these concerns, we use an exponentially decaying distance-aware weight $\exp(-d(x_{n,j}, y_{n,j}))$ at each frame $n$ to focus more on the reactor's joints that are closer to the actor, and therefore more relevant for interaction (see Fig. 4). Thus, we get our reaction loss term as + +$$ +\mathcal {L} _ {r} = \frac {1}{N J} \sum_ {n = 1} ^ {N} \sum_ {j = 1} ^ {J} \exp \left(- d \left(x _ {n, j}, y _ {n, j}\right)\right) \cdot \left| d \left(x _ {n, j}, y _ {n, j}\right) - d \left(x _ {n, j} ^ {(0)}, y _ {n, j}\right) \right|. \tag {9} +$$ + +Table 1: Dataset Comparisons. Comparing ReMoCap with existing multi-person interaction datasets. + +
Dataset (Chronologically)Motion FramesDuration (hours)Multi-view RGB VideosFinger Articulation
SBU [78]~ 7K0.13XX
NTU-26 [43]~ 22K0.47✓X
2C [55]~ 13K0.06XX
DanceDB [53]~ 1.44M4.00XX
DuetDance [39]~ 196K1.09XX
CHI3D [16]~ 486K2.70✓X
ExPI [24]~ 30K0.33✓X
InterHuman [42]~ 107M6.56XX
DD100 [60]~ 200K1.92X✓
ReMoCap (ours)~ 275.7K2.04✓✓
+ +![](images/78f89a1eb4a2a8184ebde828d8e4c8ad8e1b1249baa0032b99b449d940958155.jpg) +Fig.4: Visualization of Distance Aware Reaction Loss. We use an exponentially decaying distance-aware reaction loss to focus more on the reactor's joints that are closer to the actor. + +Kinematic Loss $\mathcal{L}_k$ . To maintain the kinematic plausibility of the generated motions, we follow existing literature on regularizers for bone length consistency, foot contacts, and temporal consistency [56,57,68]. Our kinematic loss term is a weighted sum of the joint velocity loss $\mathcal{L}_{vel}$ , the joint acceleration loss $\mathcal{L}_{acc}$ , the bone length consistency loss $\mathcal{L}_{bone}$ , and the foot sliding loss $\mathcal{L}_{foot}$ , as + +$$ +\mathcal {L} _ {k} = \lambda_ {v} \mathcal {L} _ {v e l} + \lambda_ {a} \mathcal {L} _ {a c c} + \lambda_ {b} \mathcal {L} _ {b o n e} + \lambda_ {f} \mathcal {L} _ {f o o t}, \tag {10} +$$ + +where $\lambda_v, \lambda_a, \lambda_b$ and $\lambda_f$ are scalar weights (more details in the appendix). + +# 3.4 Inference Time Spatial Guidance + +While our method synthesizes plausible reactive motions, it can sometimes spatially misalign the reactor's body to the actor's, especially at the arm joints, for fast-paced, contact-heavy interactions. This, in turn, affects finger-joint synthesis in the second stage of our cascaded diffusion. To improve spatial alignment, we leverage guidance functions [35, 52] that can provide gradients to nudge the sampling process towards a certain direction. We design a guidance function $G\left(\phi ,\hat{\phi}\right)\in \mathbb{R}^{N\times J_A\times 3}$ , which minimizes the distance between the $J_{A}$ arm joints of the actor $(\phi)$ and the $J_{A}$ arm joints of the synthesized reactor $(\hat{\phi})$ . Specifically, we re-apply our interaction masks $\mathbb{1}_{H_A}$ and $\mathbb{1}_{H_R}$ to determine which of the reactor's hands are more likely to be interacting with the actor, and minimize the distances between the arm joints of the corresponding sides as $G = \arg \min_{\hat{\phi}}\left(\left\| \mathbb{1}_{H_A}\odot \phi -\mathbb{1}_{H_R}\odot \hat{\phi}\right\|\right)$ . We then plug $G$ into the denoising pipeline of our body diffusion module, as + +$$ +X _ {B} ^ {(0)} = X _ {B} ^ {(0)} - \gamma \nabla_ {X _ {B} ^ {(0)}} G (\phi , \hat {\phi}), \quad \gamma = \text {g u i d a n c e s c a l e}. \tag {11} +$$ + +# 4 ReMoCap Dataset + +We propose the ReMoCap dataset to facilitate research on contact-based two-person interactions with finger-level articulations. It contains complex two-person interactions in two separate scenarios: the fast-paced, swing-style Lindy Hop dancing, and the martial arts of Ninjutsu. We consider these two types of motions for studying physical interactions for two main reasons. First, these motions encompass a significantly diverse spectrum of motions in terms of intensity, speed, and style. Second, choreographed motion generation is developing as an active field of research and we envision this dataset to facilitate research in interactive applications such as performing in virtual reality [9] and remote tutoring [53]. + +Data Collection. We invited 4 trained Lindy Hop dancers and 5 trained Ninjutsu artists to perform diverse interactions in a multi-view capture studio. We tracked their motions using a commercially available markerless, multi-view motion-capture system [8], which tracks 93 degrees-of-freedom driving 69 body joints. Capturing motions in a markerless setting enables the performers to move uninhibited, while also making the data suitable for training monocular or multiview motion capture methods under severe inter-person occlusions. We also capture RGB videos from 120 camera views for each sequence. The dataset includes 3D skeleton poses with full-body and finger annotations, foreground-background segmentation masks, and 3D surface reconstructions of the subjects. We capture sequences of multiple lengths totaling $\sim 275.7\mathrm{K}$ frames (2.04 hours) of motion data from each view at a frame rate of 50 fps for the Lindy Hop and 25 fps for the Ninjutsu. Out of all the frames, around $150\mathrm{K}$ frames have hand-based interactions between the two characters where the closest distance between the finger joints of the actor and the reactor is within $50~\mathrm{mm}$ . We present more dataset statistics in the appendix. + +Dataset Comparison. Table 1 shows a comprehensive comparison of ReMoCap with existing multi-person interaction datasets. Existing datasets consisting of two-person interactions, such as SBU [78], K3HI [32], NTU-26 [43] and 2C [55], are limited in size and motion capture quality. They typically feature simple actions, such as handshakes, punching, pushing and kicking, with weak interactions, and do not capture hand motions. The recent ExPI [24] dataset captures Lindy Hop aerial sequences to model more complex interactions, and the InterHuman [42] dataset features both daily motions (e.g., passing objects, greeting, communicating) and professional motions (e.g., Taekwondo, Latin dance, boxing). However, these datasets do not provide finger-level motion capture data, which is a key requirement to intricately model inter-human activities. Two-person dance datasets, such as DuetDance [39] and DanceDB [53], also lack the hand motion data needed for modeling interactions. Only the concurrent work, Duolando $[60]^*$ proposes the DD100 dataset with strong interactions between two dancers and provides hand motion data. + +# 5 Experiments and Results + +We conduct comprehensive experiments to evaluate ReMoS on multiple two-person datasets covering a wide range of interaction scenarios. We perform training and evaluation on multiple large-scale datasets, including 2C [55], ExPI [24], and our proposed ReMoCap. We provide the implementation details, report quantitative and qualitative comparisons on standard evaluation metrics and through a user study, report ablations, and show how to apply ReMoS as a motion editing tool. We provide details on the dataset preparation in the appendix. + +# 5.1 Implementation Details + +As a pre-processing step, we normalize the two-person body poses by translating the actor's motion $Y$ and the reactor's motion $X$ together, such that the root joint of $Y$ is at the global origin for all $N$ frames. We then compute the relative 3D joint coordinates of $X_B$ and $Y_B$ w.r.t. the root joint of $Y_B$ . For each hand, we compute the relative 3D joint coordinates of $X_H$ and $Y_H$ w.r.t. the corresponding wrist joints of $X_H$ and $Y_H$ . Using normalized inter-person, root-relative joint positions benefits the stability and convergence of our model. ReMoS uses $T = 500$ diffusion steps where $\bar{\alpha}$ changes linearly from 0.0002 to 0.02. We train for about 64K iterations on both sequences of ReMoCap using Adam [36] with a base learning rate of $10^{-5}$ and a batch size of 64. We decay the learning rate using a Step LR scheduler with a step size of 5 epochs and a decay rate of 0.99. We use $d = 256$ for our latent embedding representation and use 6 layers in our transformer decoder with 4 heads for calculating the attention. The training takes around 8 and 11 hours on an NVIDIA RTX A4000 GPU for Lindy Hop and Ninjutsu, respectively. The inference time is $\sim 24s$ to generate 50 frames of full-body and finger motions (12.5s for body synthesis and then 11.5s for hand synthesis). We set $\lambda_c, \lambda_r, \lambda_v = 10.0, \lambda_a, \lambda_k, \lambda_b = 1.0$ and $\lambda_f = 20.0$ (only applied after 100 training epochs) as the loss term weights in Eqns. 8 and 10. We set the guidance scale $\gamma = 10^{-3}$ in Eqn. 11. + +# 5.2 Baselines and Ablated Versions + +For baselines, we choose the closest motion synthesis methods in a two-person setting, namely, MixNMatch [20], InterFormer [11], ComMDM [54], Role-Aware Interaction Generation (RAIG) [65], and InterGen [42]. InterFormer [11] was originally trained in a reactive motion synthesis setting without additional annotations, such as action labels or text descriptions, and we maintain this setup. For the other methods, we mask out their input text/label embeddings to comply with our annotation-free setting. We re-train these methods on ReMoCap with a thorough hyper-parameter search and report their best performances. We provide more training details in the appendix. + +We also compare our proposed ReMoS model with five of its ablated versions: w/o diffusion. Training ReMoS using a transformer encoder-decoder network without using DDPM strategy. + +Table 2: Quantitative Evaluation on ReMoCap. We compare ReMoS with our baselines and ablated versions (Sec. 5.2) on the ReMoCap dataset. We evaluate these methods on metrics such as MPJPE, MPJVE, FID, and Diversity. $\downarrow$ : lower is better, $\uparrow$ : higher is better, $\rightarrow$ : values closer to GT are better. Bold indicates best. + +
MethodLindy HopNinjutsu
MPJPE (mm) ↓MPJVE (mm) ↓FID ↓ (body)FID ↓ (hands)Div →MPJPE (mm) ↓MPJVE (mm) ↓FID ↓ (body)FID ↓ (hands)Div →
GT----7.57----10.51
InterFormer [11]66.68.260.530.654.54270.23.40.570.686.48
MixNMatch [20]70.210.30.770.782.48257.25.20.740.724.83
ComMDM [54]59.44.410.320.537.48201.24.10.340.589.98
RAIG [65]71.24.320.470.638.45199.15.10.210.6310.11
InterGen [42]62.63.920.300.617.21172.63.90.320.579.98
ReMoS (ours)40.72.260.120.267.62139.23.30.160.3510.26
w/o diffusion72.54.910.580.744.04224.54.10.520.646.06
w/o cascading63.94.950.510.557.12223.64.20.420.756.62
w/o CoST-XA44.23.620.210.397.45176.63.60.270.418.91
w/o reaction loss44.63.510.220.387.31144.63.70.230.398.99
w/o spatial guidance41.92.340.120.267.62139.43.40.160.3510.26
+ +- w/o cascading. Training ReMoS with an integrated transformer for all the joints instead of a cascaded strategy for body and hand generation. +- w/o CoST-XA. Using a spatial followed by a temporal cross-attention mapping instead of the combined spatio-temporal features for body synthesis. +- w/o reaction loss. Removing the reaction loss (Eqn. 9) in training. +- w/o spatial guidance. Removing spatial guidance $G_{A}$ (Sec. 3.4) in inference. + +# 5.3 Quantitative Evaluation + +We evaluate ReMoS on standard evaluation metrics. We measure the temporal consistency deviation from the ground truth 3D pose following [57] to report the mean per-joint positional error (MPJPE) and the mean per-frame, per-joint velocity error (MPJVE), both in $mm$ , on the synthesized motions. We measure the Fréchet Inception Distance (FID) [27] to compare the distribution gap between the embedding spaces of the generated and ground-truth motions. As part of our method focuses on generating hand motions, we measure the FID score of body and hands separately for ReMoCap. We also compute the latent variance of the generated motions (Diversity) [67, 81]. Table 2 reports the quantitative evaluation of ReMoS with its baselines and ablations on ReMoCap. ReMoS achieves state-of-the-art performance for both the Lindy Hop pair-dance setting and the Ninjutsu setting. We observe that methods using denoising diffusion have higher Diversity compared to transformer-based [11] or GAN-based [20] methods. This enforces the variability claims of denoising diffusion models. We note at least $20\%$ improvement in the MPJPE and around $40\%$ improvement in the FID score for the reactor's body motion when using the proposed CoST-XA mechanism, confirming its benefit in learning fine-grained, inter-person dependencies across time. We also note an almost $50\%$ improvement in the FID scores of + +Table 3: Quantitative Evaluation on the ExPI and 2C datasets. We compare ReMoS with state-of-the-art motion synthesis methods on the ExPI [24] and 2C datasets [55]. $\downarrow$ : lower is better, $\uparrow$ : higher is better, $\rightarrow$ : values closer to GT are better. Bold indicates best. + +
MethodExPI2C
MPJPE (mm) ↓MPJVE (mm) ↓FID ↓ (body) →Div
MPJPE (mm) ↓MPJVE (mm) ↓FID ↓ (body) →Div
GT---2.01---2.22
InterFormer [11]99.13.560.421.3190.75.110.521.45
MixNMatch [20]122.45.560.481.1862.46.010.471.24
ComMDM [54]121.45.410.452.4869.93.340.492.86
RAIG [65]131.23.960.532.5191.64.920.674.45
InterGen [42]100.63.910.432.0967.64.010.472.91
ReMoS (ours)97.93.520.411.9859.13.330.342.07
+ +Table 4: User Study Results. Mean scores on a five-point Likert scale (scores 1 - 5). + +
MethodMotion Quality ↑Reaction Plausibility ↑
GT4.86 ± 0.544.72 ± 0.56
InterFormer2.52 ± 0.612.28 ± 0.57
MixNMatch1.92 ± 0.712.18 ± 0.57
ComMDM3.02 ± 0.473.12 ± 0.52
RAIG2.83 ± 0.672.48 ± 0.65
InterGen3.18 ± 0.573.19 ± 0.53
ReMoS3.79 ± 0.553.88 ± 0.54
+ +the reactor's hand motion when using the cascaded strategy and the reaction loss. Further, the spatial guidance function fine-tuning improves MPJPE and MPJVE. The MPJPE values are overall higher in Ninjutsu than in Lindy Hop as the trajectory of the reactor is more diverse for Ninjutsu. We also report the evaluation of ReMoS with its baselines on the ExPI [24] and the 2C [55] datasets in Table 3. These datasets do not provide hand motions, so we only evaluate the reactor's body motions, and report state-of-the-art performance of ReMoS. + +# 5.4 User Study + +We evaluate the visual quality of our captured ground truth data and the generated reactive motions through a user study. We show participants 26 different interaction sequences across the ground truth, our method, and its baselines. For each sequence, we randomly show them three methods side-by-side and ask them to rate the 3D motions they observe in terms of $(a)$ the reactor's motion quality, irrespective of the actor's motion (Motion Quality), and $(b)$ the plausibility of the reaction given the actor's motion (Reaction Plausibility). We ask the participants to rank these motions from '1' (worst score) to '5' (best score) in a five-point Likert scale. Table 4 reports the mean scores from the responses of 40 participants, excluding the responses that did not pass our validation checks. We notice the ground truth motions in ReMoCap achieve the highest score of around $96\%$ , indicating that participants perceive our captured ground truth motions to be natural-looking, with realistic interactions. ReMoS has the second best ranking of around $78\%$ , which is almost $21\%$ higher than the baselines, showing that it is preferred more over the baselines. + +# 5.5 Qualitative Results and Applications + +Fig. 1 shows qualitative results of ReMoS on ReMoCap and highlights the synthesized hand interactions. We note that our actor and reactor are interchangeable, i.e., depending on the character driving the interaction at a given time, + +![](images/bfc4dc9cae03538721721780d790826117b56aa80a9fcc4e5b1b9181996bda9b.jpg) +Ground-truth + +![](images/f7579fb4cd9f7dfc2674c4ffc0212fe9efcd7564fc0bcd2f47dc81e88ad2800d.jpg) +MixNMatch + +![](images/bb67196bc97a616b275cf401e571591afefe51bb76eb9d5cb05e5dde92b180d6.jpg) +Interformer + +![](images/032ff87b8b38f6da2b00e069b9ef40687907ec4c3cce76cb1daeb7d898d9ac56.jpg) +RAIG ComMDM + +![](images/14e60147d0f2fdfdf80b1156d739e58f438515ce69ea2bf0236b9145ab838b7f.jpg) + +ComMDM + +![](images/85ed6bcffaa4653b19a80a1ae70d21d456f37ca4c5cc98efe65b481a7e1995f1.jpg) +InterGen + +![](images/65e249ccf60ca8f66ba665a0da719c9724dd576fe427fc9cd1170bd51f232473.jpg) +ReMoS + +![](images/0dc11ec5b0b7ff7d5f43e287f9ffe2ef12eb7a47b3639159275186f32cf27710.jpg) +(a) Visual comparison of ReMoS with GT and baselines. +(b) Pose-completion application. +Fig.5: Qualitative Results and Applications. We show some visual results and the application of ReMoS as a motion editing tool. (a) The reactor (in blue) synthesized by ReMoS has the most plausible alignment with the actor (in red) compared to the baselines. (b) We manually control the right-hand wrist joint of the reactor and let ReMoS synthesize the remaining body joints conditioned on the actor. (c) ReMoS synthesizes the reactor's motion in-between the start and end frames. + +![](images/abb142b50b8f604756c834a3d51b1fdd5ce50faacdec7509a23abe58e23de257.jpg) +(c) Motion in-between application. + +we can swap the actor and the reactor to produce the relevant reactive motions. Fig. 5a shows a visual comparison of ReMoS with the baselines for one frame of Lindy Hop motion. ReMoS synthesizes reaction with the most plausible alignment with the actor's motion. We provide detailed visual results in our supplementary video. Inspired by the applicability of diffusion-based methods for motion editing and controlling, we demonstrate how ReMoS can also be used as an interactive motion editing tool to control the reactor's motion as desired. In Figs. 5b and 5c, we show results from two different motion editing applications, namely, pose completion with controlled joints and motion in-between. We discuss more details in the appendix. + +# 6 Conclusion + +ReMoS brings 3D motion conditioned reaction synthesis to a qualitatively new level by generating diverse, well-synchronized reactions for complex movements and plausible hand motions for contact-based interactions. It outperforms the existing baselines both quantitatively and in a user study. We also highlight some practical applications of our model, such as pose completion and motion in-between, which can lead to the development of useful generative assistants for animators, designers, and creative artists. Even though we utilize joint positions as parameterization due to their ready availability, we can adapt our approach to accommodate mesh-level inter-person contacts by introducing an offset to the contact threshold between the two bodies, thus simulating skin-to-skin interactions. We believe there is scope for improvement in the inference speed, where approaches such as Pro-DDPM [17] and DDIM [62] have shown success. Future directions also involve scaling the problem towards multi-person motion prediction [70] and considering scene-aware interactions for the characters, which would further enhance immersive user experiences. + +# Acknowledgements + +This research was supported by the EU Horizon 2020 grant $\mathrm{Carousel + }$ (101017779), and the ERC Consolidator Grant 4DRepLy (770784). We thank Marc Jahan, Christopher Ruf, Michael Hiery, Thomas Leimkuhler and Sascha Huwer for helping with the Ninjutsu data collection, and Noshaba Cheema for helping with the Lindy Hop data collection. We also thank Janis Sprenger and Duarte David for helping with the motion visualizations. + +# References + +1. Ahuja, C., Ma, S., Morency, L.P., Sheikh, Y.: To react or not to react: End-to-end visual pose forecasting for personalized avatar during dyadic conversations. In: 2019 International conference on multimodal interaction (2019) +2. Ao, T., Zhang, Z., Liu, L.: GestureDiffuCLIP: Gesture diffusion model with clip latents. SIGGRAPH (2023) +3. Aristidou, A., Yiannakidis, A., Aberman, K., Cohen-Or, D., Shamir, A., Chrysanthou, Y.: Rhythm is a dancer: Music-driven motion synthesis with global structure. IEEE Transactions on Visualization and Computer Graphics (TVCG) (2022) +4. Athanasiou, N., Petrovich, M., Black, M.J., Varol, G.: Teach: Temporal action composition for 3d humans. In: 2022 International Conference on 3D Vision (3DV) (2022) +5. Bhattacharya, U., Childs, E., Rewkowski, N., Manocha, D.: Speech2affectivegestures: Synthesizing co-speech gestures with generative adversarial affective expression learning. In: Proceedings of the 29th ACM International Conference on Multimedia (2021) +6. Bhattacharya, U., Rewkowski, N., Banerjee, A., Guhan, P., Bera, A., Manocha, D.: Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In: IEEE Conference on Virtual Reality and 3D User Interfaces (IEEE VR) (2021) +7. Bjorck, N., Gomes, C.P., Selman, B., Weinberger, K.Q.: Understanding batch normalization. In: Advances in Neural Information Processing Systems (2018) +8. https://capturey.com (2023) +9. Chan, J.C., Leung, H., Tang, J.K., Komura, T.: A virtual reality dance training system using motion capture technology. IEEE transactions on learning technologies 4(2) (2010) +0. Chopin, B., Tang, H., Daoudi, M.: Bipartite graph diffusion model for human interaction generation. In: Winter Conference on Applications of Computer Vision (WACV) (2024) +1. Chopin, B., Tang, H., Otberdout, N., Daoudi, M., Sebe, N.: Interaction transformer for human reaction generation. IEEE Transactions on Multimedia (2023) +2. Cummins, A.: In search of the ninja: the historical truth of ninjutsu. The History Press (2012) +3. Dabral, R., Mughal, M.H., Golyanik, V., Theobalt, C.: Mofusion: A framework for denoising-diffusion-based motion synthesis. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +4. Egges, A., Papagiannakis, G., Magnenat-Thalmann, N.: Presence and interaction in mixed reality environments. The Visual Computer (2007) + +15. Elfwing, S., Uchibe, E., Doya, K.: Sigmoid-weighted linear units for neural network function approximation in reinforcement learning. Neural Networks (2018) +16. Fieraru, M., Zanfir, M., Oneata, E., Popa, A.I., Olaru, V., Sminchisescu, C.: Three-dimensional reconstruction of human interactions. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2020) +17. Gandikota, R., Brown, N.: Pro-ddpm: Progressive growing of variable denoising diffusion probabilistic models for faster convergence. In: 33rd British Machine Vision Conference 2022, BMVC (2022) +18. Ghosh, A., Cheema, N., Oguz, C., Theobalt, C., Slusallek, P.: Synthesis of compositional animations from textual descriptions. In: International Conference on Computer Vision (ICCV) (2021) +19. Ghosh, A., Dabral, R., Golyanik, V., Theobalt, C., Slusallek, P.: IMoS: Intent-driven full-body motion synthesis for human-object interactions. In: Computer Graphics Forum. vol. 42. Wiley Online Library (2023) +20. Goel, A., Men, Q., Ho, E.S.L.: Interaction Mix and Match: Synthesizing Close Interaction using Conditional Hierarchical GAN with Multi-Hot Class Embedding. Computer Graphics Forum (2022) +21. Gu, D., Shim, J., Jang, J., Kang, C., Joo, K.: Contactgen: Contact-guided interactive 3d human generation for partners. AAAI (2024) +22. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: Conference on Computer Vision and Pattern Recognition (2022) +23. Guo, C., Zuo, X., Wang, S., Cheng, L.: Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In: European Conference on Computer Vision (2022) +24. Guo, W., Bie, X., Alameda-Pineda, X., Moreno-Noguer, F.: Multi-person extreme motion prediction. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +25. Habibie, I., Elgharib, M., Sarkar, K., Abdullah, A., Nyatsanga, S., Neff, M., Theobalt, C.: A motion matching-based framework for controllable gesture synthesis from speech. In: ACM SIGGRAPH Conference Proceedings (2022) +26. Hanser, E., Mc Kevitt, P., Lunney, T., Condell, J.: Scenemaker: Intelligent multimodal visualisation of natural language scripts. In: Irish Conference on Artificial Intelligence and Cognitive Science. Springer (2009) +27. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017) +28. Ho, E.S., Komura, T.: Planning tangling motions for humanoids. In: IEEE-RAS International Conference on Humanoid Robots (2007) +29. Ho, E.S., Komura, T.: Character motion synthesis by topology coordinates. In: Computer graphics forum. Wiley Online Library (2009) +30. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33 (2020) +31. Ho, J., Sahara, C., Chan, W., Fleet, D.J., Norouzi, M., Salimans, T.: Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research 23(1) (2022) +32. Hu, T., Zhu, X., Guo, W.: Two-person interaction recognition based on key poses. Journal of Computational Information Systems (2014) +33. Huang, S., Wang, Z., Li, P., Jia, B., Liu, T., Zhu, Y., Liang, W., Zhu, S.C.: Diffusion-based generation, optimization, and planning in 3d scenes. In: Proceed- + +ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023) +34. Huang, Y., Zhang, J., Liu, S., Bao, Q., Zeng, D., Chen, Z., Liu, W.: Genre-conditioned long-term 3d dance generation driven by music. In: International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE (2022) +35. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Guided motion diffusion for controllable human motion synthesis. In: International Conference on Computer Vision (ICCV) (2023) +36. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) +37. Komura, T., Ho, E.S., Lau, R.W.: Animating reactive motion using momentum-based inverse kinematics. Computer Animation and Virtual Worlds (2005) +38. Kulkarni, N., Rempe, D., Genova, K., Kundu, A., Johnson, J., Fouhey, D., Guibas, L.: Nifty: Neural object interaction fields for guided human motion synthesis. arXiv preprint arXiv:2307.07511 (2023) +39. Kundu, J.N., Buckchash, H., Mandikal, P., Jamkhandi, A., Radhakrishnan, V.B., et al.: Cross-conditioned recurrent networks for long-term synthesis of inter-person human motion interactions. In: Winter Conference on Applications of Computer Vision (WACV) (2020) +40. Li, J., Clegg, A., Mottaghi, R., Wu, J., Puig, X., Liu, C.K.: Controllable human-object interaction synthesis. arXiv preprint arXiv:2312.03913 (2023) +41. Li, J., Wu, J., Liu, C.K.: Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG) (2023) +42. Liang, H., Zhang, W., Li, W., Yu, J., Xu, L.: InterGen: Diffusion-based multi-human motion generation under complex interactions. International Journal for Computer Vision (IJCV) (2024) +43. Liu, J., Shahroudy, A., Perez, M., Wang, G., Duan, L.Y., Kot, A.C.: NTU RGB+D 120: A large-scale benchmark for 3d human activity understanding. IEEE transactions on pattern analysis and machine intelligence (2019) +44. Liu, X., Yi, L.: Geneoh diffusion: Towards generalizable hand-object interaction denoising via denoising diffusion. In: International Conference on Learning Representations (ICLR) (2024) +45. Men, Q., Shum, H.P., Ho, E.S., Leung, H.: Gan-based reactive motion synthesis with class-aware discriminators for human-human interaction. Computers & Graphics (2022) +46. Mousas, C.: Performance-driven dance motion control of a virtual partner character. In: IEEE Conference on Virtual Reality and 3D User Interfaces (VR) (2018) +47. Mughal, M.H., Dabral, R., Habibie, I., Donatelli, L., Habermann, M., Theobalt, C.: Convofusion: Multi-modal conversational diffusion for co-speech gesture synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024) +48. Ng, E., Joo, H., Hu, L., Li, H., Darrell, T., Kanazawa, A., Ginosar, S.: Learning to listen: Modeling non-deterministic dyadic facial motion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022) +49. Petrovich, M., Black, M.J., Varol, G.: TEMOS: Generating diverse human motions from textual descriptions. In: European Conference on Computer Vision (2022) +50. Po, R., Yifan, W., Golyanik, V., Aberman, K., Barron, J.T., Amit H. Bermano, E.R.C., Dekel, T., Holynski, A., Kanazawa, A., Liu, C.K., Liu, L., Mildenhall, B., Nießner, M., Ommer, B., Theobalt, C., Wonka, P., Wetzstein, G.: State of the art on diffusion models for visual computing. arXiv pre-prints (2023) + +51. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. In: International Conference on Virtual Reality (2022) +52. Rempe, D., Luo, Z., Bin Peng, X., Yuan, Y., Kitani, K., Kreis, K., Fidler, S., Litany, O.: Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +53. Senecal, S., Nijdam, N.A., Aristidou, A., Magnenat-Thalmann, N.: Salsa dance learning evaluation and motion analysis in gamified virtual reality environment. Multimedia Tools and Applications (2020) +54. Shafir, Y., Tevet, G., Kapon, R., Bermano, A.H.: Human motion diffusion as a generative prior. In: International Conference on Learning Representations (ICLR) (2024) +55. Shen, Y., Yang, L., Ho, E.S.L., Shum, H.P.H.: Interaction-based human activity comparison. IEEE Transactions on Visualization and Computer Graphics (2020) +56. Shimada, S., Golyanik, V., Xu, W., Pérez, P., Theobalt, C.: Neural monocular 3d human motion capture with physical awareness. ACM Transactions on Graphics (ToG) (2021) +57. Shimada, S., Golyanik, V., Xu, W., Theobalt, C.: Physcap: Physically plausible monocular 3d motion capture in real time. ACM Transactions on Graphics (ToG) (2020) +58. Shum, H.P., Komura, T., Shiraishi, M., Yamazaki, S.: Interaction patches for multi-character animation. ACM transactions on graphics (TOG) 27(5) (2008) +59. Shum, H.P., Komura, T., Yamazaki, S.: Simulating competitive interactions using singly captured motions. In: Proceedings of ACM symposium on Virtual reality software and technology (2007) +60. Siyao, L., Gu, T., Yang, Z., Lin, Z., Liu, Z., Ding, H., Yang, L., Loy, C.C.: Duolando: Follower gpt with off-policy reinforcement learning for dance accompaniment. In: International Conference on Learning Representations (ICLR) (2024) +61. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International Conference on Machine Learning (ICML) (2015) +62. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020) +63. Spring, H.: Swing and the lindy hop: dance, venue, media, and tradition. American Music (1997) +64. Starke, S., Zhao, Y., Komura, T., Zaman, K.: Local motion phases for learning multi-contact character movements. ACM Transactions on Graphics (TOG) (2020) +65. Tanaka, M., Fujiwara, K.: Role-aware interaction generation from textual description. In: International Conference on Computer Vision (ICCV) (2023) +66. Tanke, J., Zhang, L., Zhao, A., Tang, C., Cai, Y., Wang, L., Wu, P.C., Gall, J., Keskin, C.: Social diffusion: Long-term multiple human motion anticipation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023) +67. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Bermano, A.H., Cohen-Or, D.: Human motion diffusion model. arXiv preprint arXiv:2209.14916 (2022) +68. Tseng, J., Castellon, R., Liu, K.: Edge: Editable dance generation from music. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +69. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems (2017) + +70. Wang, J., Xu, H., Narasimhan, M., Wang, X.: Multi-person 3d motion prediction with multi-range transformers. Advances in Neural Information Processing Systems (2021) +71. Wang, Z., Chen, Y., Jia, B., Li, P., Zhang, J., Zhang, J., Liu, T., Zhu, Y., Liang, W., Huang, S.: Move as you say, interact as you can: Language-guided human motion generation with scene affordance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024) +72. Xie, Y., Jampani, V., Zhong, L., Sun, D., Jiang, H.: OmniControl: Control any joint at any time for human motion generation. In: International Conference on Learning Representations (ICLR) (2024) +73. Xing, J., Xia, M., Zhang, Y., Cun, X., Wang, J., Wong, T.T.: Codetalker: Speech-driven 3d facial animation with discrete motion prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023) +74. Xu, S., Li, Z., Wang, Y.X., Gui, L.Y.: InterDiff: Generating 3d human-object interactions with physics-informed diffusion. In: International Conference on Computer Vision (ICCV) (2023) +75. Ye, Y., Li, X., Gupta, A., De Mello, S., Birchfield, S., Song, J., Tulsiani, S., Liu, S.: Affordance diffusion: Synthesizing hand-object interactions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023) +76. Yoon, Y., Ko, W.R., Jang, M., Lee, J., Kim, J., Lee, G.: Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In: International Conference on Robotics and Automation (ICRA). IEEE (2019) +77. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: Physdiff: Physics-guided human motion diffusion model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023) +78. Yun, K., Honorio, J., Chattopadhyay, D., Berg, T.L., Samaras, D.: Two-person interaction detection using body-pose features and multiple instance learning. In: IEEE computer society conference on computer vision and pattern recognition workshops (2012) +79. Zamfirescu-Pereira, J., Wong, R.Y., Hartmann, B., Yang, Q.: Why johnny can't prompt: How non-ai experts try (and fail) to design LLM prompts. In: Proceedings of Conference on Human Factors in Computing Systems (CHI) (2023) +80. Zhang, J., Zhang, Y., Cun, X., Zhang, Y., Zhao, H., Lu, H., Shen, X., Shan, Y.: T2m-gpt: Generating human motion from textual descriptions with discrete representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023) +81. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022) +82. Zhang, W., Dabral, R., Leimkuhler, T., Golyanik, V., Habermann, M., Theobalt, C.: ROAM: Robust and object-aware motion generation using neural pose descriptors. In: International Conference on 3D Vision (3DV) (2024) +83. Zhang, X., Bhatnagar, B.L., Starke, S., Guzov, V., Pons-Moll, G.: Couch: Towards controllable human-chair interactions. In: European Conference on Computer Vision (2022) +84. Zhou, Z., Wang, B.: Ude: A unified driving engine for human motion generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023) +85. Zhu, L., Liu, X., Liu, X., Qian, R., Liu, Z., Yu, L.: Taming diffusion models for audio-driven co-speech gesture generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023) \ No newline at end of file diff --git a/2024/ReMoS_ 3D Motion-Conditioned Reaction Synthesis for Two-Person Interactions/images.zip b/2024/ReMoS_ 3D Motion-Conditioned Reaction Synthesis for Two-Person Interactions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..30510291463750c8e48cde301c1c2fff58524a74 --- /dev/null +++ b/2024/ReMoS_ 3D Motion-Conditioned Reaction Synthesis for Two-Person Interactions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e43f40ff1405424653494f6c1b011593805f1202c5e023c5b7a2cee63d05fd59 +size 383914 diff --git a/2024/ReMoS_ 3D Motion-Conditioned Reaction Synthesis for Two-Person Interactions/layout.json b/2024/ReMoS_ 3D Motion-Conditioned Reaction Synthesis for Two-Person Interactions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bd1fd575659a43b0987af80a9c7b29a72c4917a6 --- /dev/null +++ b/2024/ReMoS_ 3D Motion-Conditioned Reaction Synthesis for Two-Person Interactions/layout.json @@ -0,0 +1,11270 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 159, + 111, + 455, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 111, + 455, + 148 + ], + "spans": [ + { + "bbox": [ + 159, + 111, + 455, + 148 + ], + "type": "text", + "content": "REMoS: 3D Motion-Conditioned Reaction Synthesis for Two-Person Interactions" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 167, + 468, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 167, + 468, + 193 + ], + "spans": [ + { + "bbox": [ + 146, + 167, + 468, + 193 + ], + "type": "text", + "content": "Anindita Ghosh" + }, + { + "bbox": [ + 146, + 167, + 468, + 193 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 146, + 167, + 468, + 193 + ], + "type": "text", + "content": ", Rishabh Dabral" + }, + { + "bbox": [ + 146, + 167, + 468, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 167, + 468, + 193 + ], + "type": "text", + "content": ", Vladislav Golyanik" + }, + { + "bbox": [ + 146, + 167, + 468, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 167, + 468, + 193 + ], + "type": "text", + "content": ", Christian Theobalt" + }, + { + "bbox": [ + 146, + 167, + 468, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 167, + 468, + 193 + ], + "type": "text", + "content": ", and Philipp Slusallek" + }, + { + "bbox": [ + 146, + 167, + 468, + 193 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 183, + 201, + 430, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 201, + 430, + 236 + ], + "spans": [ + { + "bbox": [ + 183, + 201, + 430, + 236 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 183, + 201, + 430, + 236 + ], + "type": "text", + "content": " German Research Center for Artificial Intelligence (DFKI) \n" + }, + { + "bbox": [ + 183, + 201, + 430, + 236 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 183, + 201, + 430, + 236 + ], + "type": "text", + "content": " Max-Planck Institute for Informatics (MPII) \n" + }, + { + "bbox": [ + 183, + 201, + 430, + 236 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 183, + 201, + 430, + 236 + ], + "type": "text", + "content": " Saarland Informatics Campus" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 261, + 452, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 261, + 452, + 458 + ], + "spans": [ + { + "bbox": [ + 160, + 261, + 452, + 458 + ], + "type": "text", + "content": "Abstract. Current approaches for 3D human motion synthesis generate high-quality animations of digital humans performing a wide variety of actions and gestures. However, a notable technological gap exists in addressing the complex dynamics of multi-human interactions within this paradigm. In this work, we present ReMoS, a denoising diffusion-based model that synthesizes full-body reactive motion of a person in a two-person interaction scenario. Given the motion of one person, we employ a combined spatio-temporal cross-attention mechanism to synthesize the reactive body and hand motion of the second person, thereby completing the interactions between the two. We demonstrate ReMoS across challenging two-person scenarios such as pair-dancing, Ninjutsu, kickboxing, and acrobatics, where one person's movements have complex and diverse influences on the other. We also contribute the ReMoCap dataset for two-person interactions containing full-body and finger motions. We evaluate ReMoS through multiple quantitative metrics, qualitative visualizations, and a user study, and also indicate usability in interactive motion editing applications. More details are available on the project page: https://vcai.mpi-inf.mpg.de/projects/remos." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 468, + 439, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 468, + 439, + 479 + ], + "spans": [ + { + "bbox": [ + 160, + 468, + 439, + 479 + ], + "type": "text", + "content": "Keywords: Reactive Motion Synthesis " + }, + { + "bbox": [ + 160, + 468, + 439, + 479 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 468, + 439, + 479 + ], + "type": "text", + "content": " Denoising Diffusion Model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 498, + 229, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 498, + 229, + 510 + ], + "spans": [ + { + "bbox": [ + 133, + 498, + 229, + 510 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 654 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 654 + ], + "type": "text", + "content": "Digital 3D character motion synthesis has emerged as the next frontier for animation pipelines [50], particularly through denoising diffusion probabilistic models (DDPMs) [61]. While methods for generating character motion for various tasks such as text or music conditioned motion synthesis [3,4,18,22,23,34,49,80], face and gesture synthesis [5,6,25,47,48,73], human-scene interaction [19,71,82,83] exist, synthesizing interactions between humans is relatively under-explored. Modeling such human-human interactions is essential for designing generative 3D human motion synthesis frameworks supporting the complex physical and social interplay of two interacting persons [66]. It offers new capabilities for character animation tools and software, with applications in commercial and entertainment media [26], interactive mixed and augmented reality [14], and social robotics [76]." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 180 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 180 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 180 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 180 + ], + "type": "image", + "image_path": "234de5a2090ce50bb3f0bc7b8e396378ce68b906d95bbf314c3d408a4a4671d5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 188, + 482, + 245 + ], + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 245 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 245 + ], + "type": "text", + "content": "Fig. 1: Visualizations of reactive 3D motion sequences synthesized with the proposed ReMoS approach. We synthesize the 3D full-body motion of the reactor (blue) conditioned only on the 3D motion of the actor (red), thereby completing the interactions between the two (Ninjutsu practice on the left and Lindy Hop dancing on the right). The synthesized hand interactions are enlarged and highlighted with circles." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 269, + 482, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 482, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 482, + 425 + ], + "type": "text", + "content": "Toward this goal, we focus on the task of modeling reactive motions. We aim to automatically generate realistic, temporally-aligned reactive motions of a responding person given the continuous motion of a guiding person (see Fig. 1). This alleviates a common overhead for animators, enabling them to design an acting character and automatically obtain meaningful motions for a reacting character. However, automating such a process poses two key challenges. First, two-person interaction synthesis increases the dimensionality of the already challenging problem of single-person motion synthesis. In our case, the generated reactive motions must align with the conditional signals provided in the form of 3D motion sequences from the actor. Second, doing so in a generative setting without any cues from text prompts or action labels (to obviate additional supervision and data needs) requires a careful trade-off between generating diverse motions and adhering to the narrow manifold of plausible reactions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 426, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 617 + ], + "type": "text", + "content": "While existing methods for synthesizing interactions [10, 20] and two-person motions [42, 54, 65] are good at generating plausible motions, they rely on additional annotations, such as action labels or text prompts, to specify the motion. Action labels typically depict actions at a high level, such as \"high-five\" or \"salsa\", but do not capture the fine-grained synchronization with the actor's motions. Textual descriptions offer nuance for the motions, but collecting accurate textual annotations is considerably cumbersome, particularly for contact-heavy or fast-paced two-person motions, such as dancing and exercising. Further, performing generative tasks through textual prompting remains challenging for non-experts [79]. Apart from the data, contact-based motion generation also needs to consider hand-based interactions to improve realism. This entails the need for frameworks capable of synthesizing hand movements with higher degrees of freedom, which are challenging to model but crucial for improving motion realism. Balancing focus between the broader full-body reactive movements and the finer hand motions involves addressing motion at two significantly different scales — a challenge exacerbated by the high degrees of freedom of the skeletons." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "text", + "content": "With this objective in mind, we present ReMoS, a novel approach for Reactive Motion Synthesis with full-body articulations. Inspired by the advancements of DDPMs in 3D human motion synthesis [50,54,65], we develop a DDPM framework (Sec. 3.1) with a cascaded, two-stage generation strategy tailored to our" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "text", + "content": "A. Ghosh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 320 + ], + "type": "text", + "content": "problem setting (Fig. 2). In the first stage of our diffusion model, we generate the reactive motion for the reactor's body joints conditioned on the actor's body joints. In the second stage, we use the synthesized body motions as an additional parameter to generate masking for appropriate hand motions (Sec. 3.2). We propose a combined spatio-temporal cross-attention (CoST-XA) mapping between the actors' and the reactors' body motion embeddings, that learns the interdependencies in their motions without needing additional annotations. Further, to synthesize hand interactions, we introduce a hand-interaction-aware cross-attention (H-XA) mechanism to ensure the relevant hand joints react to the actors' motions, thus allowing the network to distill localized hand interaction features. We ensure accurate coordination between the actor and the reactor by using contact-based reaction loss (Sec. 3.3) and an inference-time guidance function (Sec. 3.4) that improves the plausibility of the body and hand interactions. To explore reaction synthesis in complex and diverse two-person scenarios, we contribute the REMOCAP dataset consisting of full-body and finger motion sequences for fast-paced swing dance of Lindy Hop [63], and the martial art technique of Ninjutsu [12] (Sec. 4). In summary, our technical contributions are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 320, + 482, + 487 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 320, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 482, + 380 + ], + "type": "text", + "content": "- REMoS. A novel method for reactive 3D human motion synthesis using a cascaded diffusion framework to generate full-body and hand motions. Our framework generates fine-grained reactive motions for complex and dynamic contact-based interactions and derives the reactions directly from the actor's motions, without requiring explicit label or text annotations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 380, + 482, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 380, + 482, + 438 + ], + "spans": [ + { + "bbox": [ + 139, + 380, + 482, + 438 + ], + "type": "text", + "content": "- Interaction-Based Attentions. A combined spatio-temporal cross attention (CoST-XA) mechanism to enforce coherence between the body movements of the actor and the reactor, and a hand-interaction-aware cross-attention (H-XA) mechanism to enforce the appropriate hand-based interactions between the two characters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 440, + 482, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 440, + 482, + 487 + ], + "spans": [ + { + "bbox": [ + 139, + 440, + 482, + 487 + ], + "type": "text", + "content": "- REMoCAP. A new dataset for two-person interactions under complex scenarios of Lindy Hop dancing and Ninjutsu. The dataset consists of " + }, + { + "bbox": [ + 139, + 440, + 482, + 487 + ], + "type": "inline_equation", + "content": "\\sim 275.7\\mathrm{K}" + }, + { + "bbox": [ + 139, + 440, + 482, + 487 + ], + "type": "text", + "content": " motion frames with multiview RGB videos and 3D full-body motion capture of the two interacting persons. It further includes finger-level articulations." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 487, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 536 + ], + "type": "text", + "content": "We evaluate our approach in different scenarios, including Lindy Hop dancing, Ninjutsu, Acrobatics [24] and Kickboxing [55], and report state-of-the-art performance (Sec. 5.3). We also report a user study on comparing the visual quality of ReMoS compared to the ground truth and the baseline methods (Sec. 5.4)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 555, + 237, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 237, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 237, + 568 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "text", + "content": "Multi-Person 3D Motion Synthesis. Synthesizing close interactions between two or multiple persons is a challenging task in animation. Early works in multi-person interaction are based on motion graphs [59], interaction patches [58], momentum-based inverse kinematics and motion blending [37], and topologically-based pose representations [28, 29], to name a few. The increasing availability of interaction datasets [16, 39, 43, 53, 55, 78] has led to a rise in data-driven approaches for synthesizing digital partners or opponents in multi-person inter" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "ReMoS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": "action settings. Mousas [46] uses a hidden Markov model to control the dance motions of a digital partner in an immersive setup. Ahuja et al. [1] introduce a residual-attention model to generate body pose in a conversation setting conditioned on two audio signals and the opposite person's body pose. Starke et al. [64] propose a phase function-based network that learns asynchronous movements of each bone and its interaction with external objects. Guo et al. [24] propose the Extreme Pose Interaction dataset with a two-stream network with cross-interaction attention modules for forecasting pose sequences of two interacting characters. GAN-based models [20,45] generate motions for an interacting person conditioned on an input character and class labels depicting the type of reaction performed. These methods leverage the daily interaction datasets such as SBU Kinect [78] and 2C [55]. InterFormer [11] uses an interaction transformer with both spatial and temporal attention to generate reactive motions given some initial seed poses of both characters. It can synthesize sparse-level interactions based on the motions of the K3HI [32] and the DuetDance [39] datasets. Concurrent to our work, Duolando [60] uses an off-policy reinforcement learning model to predict tokenized motion for a leader and follower conditioned on music. The aforementioned methods either depend on seed motion as input or require additional conditions to drive the motion. In contrast, ReMoS focuses on synthesizing well-synchronized full-body and hand motions for the reactor conditioned only on the actor's 3D motions, without using any additional labels or prompts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 378, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 668 + ], + "type": "text", + "content": "Diffusion Based 3D Motion Synthesis. Denoising diffusion models [61] have recently demonstrated their high potential in generative human motion modeling, specifically in single-person conditional motion synthesis. Conditional single-person motion generation has been performed using diffusion-based approaches for co-speech gestures [2, 85], audio-driven motion [13, 68, 84], and text-driven motion [54,67,77,81] synthesis. Diffusion-based techniques have also been extensively used for conditional synthesis for human-object interactions [38, 40, 41, 74], hand-object interactions [44, 75], and human-scene interactions [33]. Guided motion diffusion models such as GMD [35] and TraceAndPace [52] incorporate spatial constraints on motion trajectories, to guide the motion towards a goal at inference time. OmniControl [72] proposes spatial and realism guidance to control any joint for diffusion-based human motion generation. For two-person interaction synthesis, RAIG [65] proposes a diffusion-based, role-aware approach for two-person interactions, given separate textual descriptions for each person. BiGraphDiff [10] uses graph transformer denoising diffusion to learn two-person interaction conditioned on action labels. ComMDM [54] uses a communication block between two MDMs [67] to coordinate two-person interaction generation. ContactGen [21] presents a diffusion-based contact prediction module that adaptively estimates potential contact regions between two humans according to the interaction label. Liang et al. [42] propose InterGen, a diffusion-based model to synthesize two-person interactions given text descriptions as input conditions. They also propose the InterHuman dataset that includes diverse two-person interaction scenarios with rich text annotations. Inspired by these recent approaches, we design our proposed method ReMoS as a DDPM-based" + } + ] + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 101 + ], + "type": "text", + "content": "A. Ghosh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 115, + 309, + 214 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 309, + 214 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 309, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 309, + 214 + ], + "type": "image", + "image_path": "9d441678b2f9f21acf62a1b4370d336ba081f4cc3395a4fbc8383021159f9deb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 114, + 482, + 204 + ], + "lines": [ + { + "bbox": [ + 321, + 114, + 482, + 204 + ], + "spans": [ + { + "bbox": [ + 321, + 114, + 482, + 204 + ], + "type": "text", + "content": "Fig.2: ReMoS Overview. Given the motion of the actor (bottom-middle, in red), we synthesize a plausible motion for the reactor (bottom-left, in blue). We achieve this using a denoising diffusion-based probabilistic model (center) trained on reactive motion sequences (top-left, in blue)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 482, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 320 + ], + "type": "text", + "content": "3D motion-conditioned reaction synthesis method, consisting of a cascaded diffusion model with a combined spatio-temporal cross-attention mechanism. This allows ReMoS to implicitly learn the fine-grained synchronization between two interacting persons from only one of the person's motions, without any need for additional prompts or labels, and synthesize the corresponding motions of the second person. ReMoS further synthesizes plausible hand motions for the reactor to incorporate realistic hand-based interactions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 338, + 313, + 351 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 338, + 313, + 351 + ], + "spans": [ + { + "bbox": [ + 132, + 338, + 313, + 351 + ], + "type": "text", + "content": "3 Reactive Motion Synthesis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 361, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 518 + ], + "type": "text", + "content": "We consider a digital, two-person interaction setting where the motions of one character, the actor, are known, and we need to synthesize the motions of the other character, the reactor. ReMoS aims to synthesize the synchronized reactive motions of the reactor in such a setting. We denote the reactor's motion as " + }, + { + "bbox": [ + 130, + 361, + 482, + 518 + ], + "type": "inline_equation", + "content": "X = \\{X_{B},X_{H}\\}" + }, + { + "bbox": [ + 130, + 361, + 482, + 518 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 361, + 482, + 518 + ], + "type": "inline_equation", + "content": "X_{B}\\in \\mathbb{R}^{N\\times J_{B}\\times 3}" + }, + { + "bbox": [ + 130, + 361, + 482, + 518 + ], + "type": "text", + "content": " denotes the 3D full-body joint positions with " + }, + { + "bbox": [ + 130, + 361, + 482, + 518 + ], + "type": "inline_equation", + "content": "J_{B}" + }, + { + "bbox": [ + 130, + 361, + 482, + 518 + ], + "type": "text", + "content": " body joints, " + }, + { + "bbox": [ + 130, + 361, + 482, + 518 + ], + "type": "inline_equation", + "content": "X_{H}\\in \\mathbb{R}^{N\\times J_{H}\\times 3}" + }, + { + "bbox": [ + 130, + 361, + 482, + 518 + ], + "type": "text", + "content": " denotes the 3D finger joint positions with " + }, + { + "bbox": [ + 130, + 361, + 482, + 518 + ], + "type": "inline_equation", + "content": "J_{H}" + }, + { + "bbox": [ + 130, + 361, + 482, + 518 + ], + "type": "text", + "content": " finger joints across both hands, and " + }, + { + "bbox": [ + 130, + 361, + 482, + 518 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 361, + 482, + 518 + ], + "type": "text", + "content": " denotes the number of frames in the motion sequence. Likewise, we denote the actor's motion as " + }, + { + "bbox": [ + 130, + 361, + 482, + 518 + ], + "type": "inline_equation", + "content": "Y = \\{Y_{B},Y_{H}\\}" + }, + { + "bbox": [ + 130, + 361, + 482, + 518 + ], + "type": "text", + "content": ". Our goal is to model the conditional probability distribution " + }, + { + "bbox": [ + 130, + 361, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\mathrm{P}(X|Y)" + }, + { + "bbox": [ + 130, + 361, + 482, + 518 + ], + "type": "text", + "content": " from which we can sample plausible reactive motions. We model this distribution using conditional denoising diffusion probabilistic models (DDPMs), owing to their unique strengths in capturing temporal coherence of motion data and handling complex motion dynamics [13,67,68]. We elaborate our methodology below." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 533, + 354, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 354, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 354, + 547 + ], + "type": "text", + "content": "3.1 DDPM for Reactive Motion Synthesis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 552, + 482, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 482, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 482, + 613 + ], + "type": "text", + "content": "The diffusion process consists of two steps: forward or destructive diffusion, and reverse or denoising diffusion. In the forward process, we progressively corrupt a clean reactive motion sequence " + }, + { + "bbox": [ + 130, + 552, + 482, + 613 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 552, + 482, + 613 + ], + "type": "text", + "content": " by adding Gaussian noise " + }, + { + "bbox": [ + 130, + 552, + 482, + 613 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 552, + 482, + 613 + ], + "type": "text", + "content": " to it for " + }, + { + "bbox": [ + 130, + 552, + 482, + 613 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 552, + 482, + 613 + ], + "type": "text", + "content": " steps. With sufficiently small noise and large " + }, + { + "bbox": [ + 130, + 552, + 482, + 613 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 552, + 482, + 613 + ], + "type": "text", + "content": ", we can get " + }, + { + "bbox": [ + 130, + 552, + 482, + 613 + ], + "type": "inline_equation", + "content": "X^{(T)} \\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 130, + 552, + 482, + 613 + ], + "type": "text", + "content": " following closed-form formulation of Ho et al. [30]:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 203, + 619, + 481, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 619, + 481, + 634 + ], + "spans": [ + { + "bbox": [ + 203, + 619, + 481, + 634 + ], + "type": "interline_equation", + "content": "X ^ {(t)} = \\sqrt {\\bar {\\alpha} _ {t}} X ^ {(0)} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\quad \\epsilon \\sim \\mathcal {N} (0, I), \\tag {1}", + "image_path": "5ac4813def2ba65bcb6ee090f1d6069c46eddce040f747702accb3ff96558360.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " controls the rate of diffusion and " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "t\\in [0,T]" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": ". Reversing this diffusion process allows for sampling novel motion sequences from a multivariate Gaussian" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "ReMoS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 288, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 288, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 288, + 129 + ], + "type": "text", + "content": "distribution " + }, + { + "bbox": [ + 132, + 114, + 288, + 129 + ], + "type": "inline_equation", + "content": "p\\left(X^{(T)}\\right)\\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 132, + 114, + 288, + 129 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 213, + 139, + 481, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 139, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 213, + 139, + 481, + 171 + ], + "type": "interline_equation", + "content": "p \\left(X ^ {(0)}\\right) = p \\left(X ^ {(T)}\\right) \\prod_ {t = 1} ^ {T} p \\left(X ^ {(t - 1)} \\mid X ^ {(t)}\\right). \\tag {2}", + "image_path": "785ec16dfa520d0c30d0a2b76078f8c672a85981498c5b5f8dcb09f9682210b8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 180, + 482, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 180, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 131, + 180, + 482, + 257 + ], + "type": "text", + "content": "We approximate the computationally intractable term " + }, + { + "bbox": [ + 131, + 180, + 482, + 257 + ], + "type": "inline_equation", + "content": "p\\left(X^{(t - 1)}|X^{(t)}\\right)" + }, + { + "bbox": [ + 131, + 180, + 482, + 257 + ], + "type": "text", + "content": " with a learnable function " + }, + { + "bbox": [ + 131, + 180, + 482, + 257 + ], + "type": "inline_equation", + "content": "f_{\\theta}\\left(X^{(t)},t\\right)" + }, + { + "bbox": [ + 131, + 180, + 482, + 257 + ], + "type": "text", + "content": ", and optimize " + }, + { + "bbox": [ + 131, + 180, + 482, + 257 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 131, + 180, + 482, + 257 + ], + "type": "text", + "content": " to encode the space of human reactive motions. Our reactive motion " + }, + { + "bbox": [ + 131, + 180, + 482, + 257 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 131, + 180, + 482, + 257 + ], + "type": "text", + "content": " is also conditioned on the actor's motion " + }, + { + "bbox": [ + 131, + 180, + 482, + 257 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 131, + 180, + 482, + 257 + ], + "type": "text", + "content": ". Therefore, we modify the learnable function as " + }, + { + "bbox": [ + 131, + 180, + 482, + 257 + ], + "type": "inline_equation", + "content": "X^{(t - 1)} = f_{\\theta}\\left(X^{(t)},t,Y\\right)" + }, + { + "bbox": [ + 131, + 180, + 482, + 257 + ], + "type": "text", + "content": ". Following recent works [51,67], we estimate the original motion " + }, + { + "bbox": [ + 131, + 180, + 482, + 257 + ], + "type": "inline_equation", + "content": "X^{(0)} = f_{\\theta}\\left(X^{(t)},t,Y\\right)" + }, + { + "bbox": [ + 131, + 180, + 482, + 257 + ], + "type": "text", + "content": " from our diffusion model by iterating through all " + }, + { + "bbox": [ + 131, + 180, + 482, + 257 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 180, + 482, + 257 + ], + "type": "text", + "content": " denoising steps during inference." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 274, + 258, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 258, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 258, + 285 + ], + "type": "text", + "content": "3.2 ReMoS Framework" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 294, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 294, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 131, + 294, + 482, + 331 + ], + "type": "text", + "content": "To generate fine-grained reactions with appropriate hand motions, ReMoS decodes the reactive motion of " + }, + { + "bbox": [ + 131, + 294, + 482, + 331 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 131, + 294, + 482, + 331 + ], + "type": "text", + "content": " in a cascaded fashion (Fig. 3). It first estimates the full-body joints " + }, + { + "bbox": [ + 131, + 294, + 482, + 331 + ], + "type": "inline_equation", + "content": "X_{B}" + }, + { + "bbox": [ + 131, + 294, + 482, + 331 + ], + "type": "text", + "content": ", and then the hand joints " + }, + { + "bbox": [ + 131, + 294, + 482, + 331 + ], + "type": "inline_equation", + "content": "X_{H}" + }, + { + "bbox": [ + 131, + 294, + 482, + 331 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 188, + 339, + 480, + 359 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 339, + 480, + 359 + ], + "spans": [ + { + "bbox": [ + 188, + 339, + 480, + 359 + ], + "type": "interline_equation", + "content": "X _ {B} ^ {(0)} = f _ {\\theta_ {B}} \\left(X _ {B} ^ {(t)}, t, Y _ {B}\\right), \\tag {3}", + "image_path": "b288574a1b070b35b404c7486d5775bad59a5cb8ba983ddd05f0158bc26aaf0e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 188, + 361, + 480, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 361, + 480, + 381 + ], + "spans": [ + { + "bbox": [ + 188, + 361, + 480, + 381 + ], + "type": "interline_equation", + "content": "X _ {H} ^ {(0)} = f _ {\\theta_ {H}} \\left(X _ {H} ^ {(t)}, t, Y _ {H}, \\mathbb {1} _ {H _ {A}} \\left(Y _ {B}\\right) \\mathbb {1} _ {H _ {R}} \\left(X _ {B} ^ {(0)}\\right)\\right), \\quad \\text {a n d} \\tag {4}", + "image_path": "f1ba5697c36014061bdd691f2df89cd5409ade19d6e8dc0dc58dc1015e3015f0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 188, + 383, + 480, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 383, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 188, + 383, + 480, + 403 + ], + "type": "interline_equation", + "content": "X ^ {(0)} = \\left\\{X _ {B} ^ {(0)}, X _ {H} ^ {(0)} \\right\\}, \\tag {5}", + "image_path": "eef5d5000addb840998d8fdbbe087405e233b22947af96d95a2407d456767260.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 411, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 411, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 131, + 411, + 482, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 411, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathbb{1}_{H_A}" + }, + { + "bbox": [ + 131, + 411, + 482, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 411, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathbb{1}_{H_R}" + }, + { + "bbox": [ + 131, + 411, + 482, + 544 + ], + "type": "text", + "content": " are binary, spatio-temporal hand-interaction mask functions that determine which hand joints of the reactor and actor interact at any given frame. Our cascaded framework comes from the observation that full-body articulations occur at significantly different scales than hand articulations. Since the body motions affect the hand articulations, our cascaded framework feeds the generated body motions as an additional condition to the hand generation module (Fig. 3). We also train the body and the hand generation modules separately, with additional hand-interaction-aware attention mapping for the hand joints. Previous work [19] used such disjoint training strategy on conditional motion synthesis for single persons. To better accommodate for two-person interactions in our case, we benefit from using a cascaded diffusion strategy [31]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 555, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 555, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 555, + 482, + 666 + ], + "type": "text", + "content": "Body Synthesis Module. In the first stage of our cascaded framework, we diffuse the reactor's body motion, " + }, + { + "bbox": [ + 131, + 555, + 482, + 666 + ], + "type": "inline_equation", + "content": "X_B^{(t)}" + }, + { + "bbox": [ + 131, + 555, + 482, + 666 + ], + "type": "text", + "content": ", at each diffusion step " + }, + { + "bbox": [ + 131, + 555, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 555, + 482, + 666 + ], + "type": "text", + "content": ", and feed it into a transformer decoder [69] block. To condition the reactor's body motion on the actor's body motion " + }, + { + "bbox": [ + 131, + 555, + 482, + 666 + ], + "type": "inline_equation", + "content": "Y_B" + }, + { + "bbox": [ + 131, + 555, + 482, + 666 + ], + "type": "text", + "content": ", we introduce a combined spatio-temporal cross-attention (CoST-XA) mapping. CoST-XA combines the spatial and temporal features of the actor's and the reactor's motions through an attention matrix to simultaneously learn the inter-dependencies between each pair of reactor and actor body joints " + }, + { + "bbox": [ + 131, + 555, + 482, + 666 + ], + "type": "inline_equation", + "content": "(x_{n,j}, y_{n,j})" + }, + { + "bbox": [ + 131, + 555, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 555, + 482, + 666 + ], + "type": "inline_equation", + "content": "x_{n,j} \\in X_B^{(t)}" + }, + { + "bbox": [ + 131, + 555, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 555, + 482, + 666 + ], + "type": "inline_equation", + "content": "y_{n,j} \\in Y_B" + }, + { + "bbox": [ + 131, + 555, + 482, + 666 + ], + "type": "text", + "content": ", respectively, at each frame. Denoting the query from the reactor's motion features as " + }, + { + "bbox": [ + 131, + 555, + 482, + 666 + ], + "type": "inline_equation", + "content": "Q_B \\in \\mathbb{R}^{NJ_B \\times 3}" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "text", + "content": "A. Ghosh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 478, + 210 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 478, + 210 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 478, + 210 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 478, + 210 + ], + "type": "image", + "image_path": "91b1996d2b0933da10018d5f4abc699bdb32155efbbd32d044eeffe9d7002479.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 220, + 482, + 277 + ], + "lines": [ + { + "bbox": [ + 131, + 220, + 482, + 277 + ], + "spans": [ + { + "bbox": [ + 131, + 220, + 482, + 277 + ], + "type": "text", + "content": "Fig. 3: ReMoS Framework. Given the full-body sequence of the actor (left, in red), we input noisy body and hand samples (from below) in a cascaded fashion. We synthesize the body samples first, and use them for hand-interaction-aware attention masking (top-center) to synthesize the denoised hand samples (top-right). The full-body reactive motion is a concatenation of the denoised body and hand samples (right, in blue)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 300, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 300, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 131, + 300, + 482, + 407 + ], + "type": "text", + "content": "and the key-value pair from the actor's motion features as " + }, + { + "bbox": [ + 131, + 300, + 482, + 407 + ], + "type": "inline_equation", + "content": "K_{B} \\in \\mathbb{R}^{NJ_{B} \\times 3}" + }, + { + "bbox": [ + 131, + 300, + 482, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 300, + 482, + 407 + ], + "type": "inline_equation", + "content": "V_{B} \\in \\mathbb{R}^{NJ_{B} \\times 3}" + }, + { + "bbox": [ + 131, + 300, + 482, + 407 + ], + "type": "text", + "content": ", respectively, we formulate an attention matrix of dimensions " + }, + { + "bbox": [ + 131, + 300, + 482, + 407 + ], + "type": "inline_equation", + "content": "NJ_{B} \\times NJ_{B}" + }, + { + "bbox": [ + 131, + 300, + 482, + 407 + ], + "type": "text", + "content": ". In contrast, most previous approaches [42,65] use cross-attention only on the temporal features of the motion using an attention matrix of shape " + }, + { + "bbox": [ + 131, + 300, + 482, + 407 + ], + "type": "inline_equation", + "content": "N \\times N" + }, + { + "bbox": [ + 131, + 300, + 482, + 407 + ], + "type": "text", + "content": ". InterFormer [11] sequentially uses spatial and temporal cross-attention modules instead of combining them, which may result in a partial loss of information on fine-grained, inter-person interactions across time, especially in the absence of additional annotations (prompts or action labels) to condition the reactions. Our proposed cross-attention module, which we define as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 225, + 417, + 480, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 417, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 225, + 417, + 480, + 449 + ], + "type": "interline_equation", + "content": "\\text {C o S T - X A} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {B} K _ {B} ^ {T}}{\\sqrt {d _ {K _ {B}}}}\\right) V _ {B}, \\tag {6}", + "image_path": "8a41d884e0970eebdb54872b8c8075cf11dbd878dbb8ec48d766f11a9bdbba9d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "text", + "content": "crucially considers combinations of spatial and temporal interaction features between different body segments of the actor and the reactor to efficiently synchronize their motions. We simultaneously project the diffusion timestep " + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "text", + "content": " at each denoising step into the transformer block after the attention blocks and use a final layer of fully-connected network with SiLU activations [15] and batch normalization [7] to generate the body motions. The first stage output is a vector " + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "inline_equation", + "content": "X_B^{(0)} \\in \\mathbb{R}^{N \\times J_B \\times 3}" + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "text", + "content": " representing the reactor's clean, synthesized body motion." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 556, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 556, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 556, + 482, + 666 + ], + "type": "text", + "content": "Hands Synthesis Module. The second stage of our cascaded framework synthesizes the reactor's hand joints. We input noisy samples " + }, + { + "bbox": [ + 131, + 556, + 482, + 666 + ], + "type": "inline_equation", + "content": "X_H^{(t)}" + }, + { + "bbox": [ + 131, + 556, + 482, + 666 + ], + "type": "text", + "content": " into a similar transformer decoder block conditioned on the actor's hand motions " + }, + { + "bbox": [ + 131, + 556, + 482, + 666 + ], + "type": "inline_equation", + "content": "Y_H" + }, + { + "bbox": [ + 131, + 556, + 482, + 666 + ], + "type": "text", + "content": ". Here, we require the attention weights to be high for the hands used in the hand-based interactions and low for the passive hands. We explicitly enforce this using a spatio-temporal hand-interaction-aware cross-attention mapping (H-XA) to encourage learning localized interaction features. For H-XA to work, we introduce binary hand-interaction masks " + }, + { + "bbox": [ + 131, + 556, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{1}_{H_A}" + }, + { + "bbox": [ + 131, + 556, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 556, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{1}_{H_R} \\in \\mathbb{R}^{N \\times J_H}" + }, + { + "bbox": [ + 131, + 556, + 482, + 666 + ], + "type": "text", + "content": ". Its entries are calculated by thresholding the distances of the actor's and reactor's wrist joints in " + }, + { + "bbox": [ + 131, + 556, + 482, + 666 + ], + "type": "inline_equation", + "content": "Y_B" + }, + { + "bbox": [ + 131, + 556, + 482, + 666 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "ReMoS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 482, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 482, + 166 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 482, + 166 + ], + "type": "inline_equation", + "content": "X_B^{(0)}" + }, + { + "bbox": [ + 132, + 115, + 482, + 166 + ], + "type": "text", + "content": ", respectively, to the nearest body joints of the other person. The active entries of " + }, + { + "bbox": [ + 132, + 115, + 482, + 166 + ], + "type": "inline_equation", + "content": "\\mathbb{1}_{H_A}" + }, + { + "bbox": [ + 132, + 115, + 482, + 166 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 115, + 482, + 166 + ], + "type": "inline_equation", + "content": "\\mathbb{1}_{H_R}" + }, + { + "bbox": [ + 132, + 115, + 482, + 166 + ], + "type": "text", + "content": " determine which of the actor's and reactor's hands are sufficiently close to the body of the other person (and hence interacting). We then project the encoded " + }, + { + "bbox": [ + 132, + 115, + 482, + 166 + ], + "type": "inline_equation", + "content": "Y_H" + }, + { + "bbox": [ + 132, + 115, + 482, + 166 + ], + "type": "text", + "content": " features into each decoder layer using H-XA, as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 194, + 182, + 481, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 182, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 194, + 182, + 481, + 213 + ], + "type": "interline_equation", + "content": "\\mathrm {H} - \\mathrm {X A} = \\operatorname {s o f t m a x} \\left(\\frac {\\left(\\mathbb {1} _ {H _ {R}} \\odot Q _ {H}\\right) \\left(\\mathbb {1} _ {H _ {A}} \\odot K _ {H}\\right) ^ {T}}{\\sqrt {d _ {K _ {H}}}}\\right) V _ {H}, \\tag {7}", + "image_path": "7653a91a1e28bb692d9399756fafb4086f4fbc4c3b7bfbaa7abcdaa9622d503c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 220, + 482, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 220, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 220, + 482, + 312 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 220, + 482, + 312 + ], + "type": "inline_equation", + "content": "K_{H}" + }, + { + "bbox": [ + 132, + 220, + 482, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 220, + 482, + 312 + ], + "type": "inline_equation", + "content": "V_{H}" + }, + { + "bbox": [ + 132, + 220, + 482, + 312 + ], + "type": "text", + "content": " are the hand motion features from the actor, " + }, + { + "bbox": [ + 132, + 220, + 482, + 312 + ], + "type": "inline_equation", + "content": "Q_{H}" + }, + { + "bbox": [ + 132, + 220, + 482, + 312 + ], + "type": "text", + "content": " are the hand motion features from the reactor with " + }, + { + "bbox": [ + 132, + 220, + 482, + 312 + ], + "type": "inline_equation", + "content": "d_{K_{H}}" + }, + { + "bbox": [ + 132, + 220, + 482, + 312 + ], + "type": "text", + "content": " channels, and " + }, + { + "bbox": [ + 132, + 220, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 132, + 220, + 482, + 312 + ], + "type": "text", + "content": " denotes element-wise product for masking the query and key values. We project the diffusion timestep " + }, + { + "bbox": [ + 132, + 220, + 482, + 312 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 132, + 220, + 482, + 312 + ], + "type": "text", + "content": " at each denoising step into the transformer block and use fully-connected networks with SiLU activations to generate the clean, synthesized reactor hand motions " + }, + { + "bbox": [ + 132, + 220, + 482, + 312 + ], + "type": "inline_equation", + "content": "X_{H}^{(0)} \\in \\mathbb{R}^{N \\times J_{H} \\times 3}" + }, + { + "bbox": [ + 132, + 220, + 482, + 312 + ], + "type": "text", + "content": ". At the end, to obtain the full reactive motion " + }, + { + "bbox": [ + 132, + 220, + 482, + 312 + ], + "type": "inline_equation", + "content": "X^{(0)}" + }, + { + "bbox": [ + 132, + 220, + 482, + 312 + ], + "type": "text", + "content": ", we concatenate " + }, + { + "bbox": [ + 132, + 220, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\left\\{X_{B}^{(0)}, X_{H}^{(0)}\\right\\}" + }, + { + "bbox": [ + 132, + 220, + 482, + 312 + ], + "type": "text", + "content": " as in Eqn. 5." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 327, + 301, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 327, + 301, + 339 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 301, + 339 + ], + "type": "text", + "content": "3.3 Losses and Training Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 346, + 481, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 346, + 481, + 370 + ], + "spans": [ + { + "bbox": [ + 132, + 346, + 481, + 370 + ], + "type": "text", + "content": "We train ReMoS to minimize a weighted sum of three loss terms, the reconstruction loss " + }, + { + "bbox": [ + 132, + 346, + 481, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 132, + 346, + 481, + 370 + ], + "type": "text", + "content": ", the reaction loss " + }, + { + "bbox": [ + 132, + 346, + 481, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 132, + 346, + 481, + 370 + ], + "type": "text", + "content": ", and the kinematic loss " + }, + { + "bbox": [ + 132, + 346, + 481, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_k" + }, + { + "bbox": [ + 132, + 346, + 481, + 370 + ], + "type": "text", + "content": ", as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 251, + 378, + 481, + 390 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 378, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 251, + 378, + 481, + 390 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {c} \\mathcal {L} _ {c} + \\lambda_ {r} \\mathcal {L} _ {r} + \\lambda_ {k} \\mathcal {L} _ {k}, \\tag {8}", + "image_path": "f67f23cd6e89b61c82036d34cc8d9935e4ea92e65c0758b1d9c96629fe09be90.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 398, + 446, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 398, + 446, + 411 + ], + "spans": [ + { + "bbox": [ + 132, + 398, + 446, + 411 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 398, + 446, + 411 + ], + "type": "inline_equation", + "content": "\\lambda_{c},\\lambda_{r}" + }, + { + "bbox": [ + 132, + 398, + 446, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 398, + 446, + 411 + ], + "type": "inline_equation", + "content": "\\lambda_{k}" + }, + { + "bbox": [ + 132, + 398, + 446, + 411 + ], + "type": "text", + "content": " are scalar weights to balance the individual losses." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 420, + 482, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 482, + 480 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 482, + 480 + ], + "type": "text", + "content": "Reconstruction Loss " + }, + { + "bbox": [ + 132, + 420, + 482, + 480 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 132, + 420, + 482, + 480 + ], + "type": "text", + "content": ". This is the standard diffusion data term that minimizes the " + }, + { + "bbox": [ + 132, + 420, + 482, + 480 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 132, + 420, + 482, + 480 + ], + "type": "text", + "content": "-distance between the ground truth and the synthesized reactive motion, as " + }, + { + "bbox": [ + 132, + 420, + 482, + 480 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c = \\left\\| X - X^{(0)}\\right\\|_2" + }, + { + "bbox": [ + 132, + 420, + 482, + 480 + ], + "type": "text", + "content": ". While the reconstruction loss provides the vital data term to drive the training, it does not enforce interaction synchronization between the actor and the reactor, or any kinematic constraints on the motion." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 490, + 482, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 482, + 632 + ], + "type": "text", + "content": "Reaction loss " + }, + { + "bbox": [ + 132, + 490, + 482, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 132, + 490, + 482, + 632 + ], + "type": "text", + "content": ". We introduce the reaction loss to ensure accurate timing and spatial positioning of the reactor's motion with respect to the actor's motion. We calculate the ground-truth Euclidean distance " + }, + { + "bbox": [ + 132, + 490, + 482, + 632 + ], + "type": "inline_equation", + "content": "d(\\cdot, \\cdot)" + }, + { + "bbox": [ + 132, + 490, + 482, + 632 + ], + "type": "text", + "content": " between each joint of the actor and the reactor at each frame and minimize the deviations from these distances for the actor and the synthesized reactor's motions. We note that this distance term also implicitly constrains the reactor's joints in frames that are less relevant for interactions, i.e., where the reactor's joints are far from the actor's joints. Further, the Euclidean distances for their hand joints in frames with no hand-based contact become drastically high. To mitigate these concerns, we use an exponentially decaying distance-aware weight " + }, + { + "bbox": [ + 132, + 490, + 482, + 632 + ], + "type": "inline_equation", + "content": "\\exp(-d(x_{n,j}, y_{n,j}))" + }, + { + "bbox": [ + 132, + 490, + 482, + 632 + ], + "type": "text", + "content": " at each frame " + }, + { + "bbox": [ + 132, + 490, + 482, + 632 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 132, + 490, + 482, + 632 + ], + "type": "text", + "content": " to focus more on the reactor's joints that are closer to the actor, and therefore more relevant for interaction (see Fig. 4). Thus, we get our reaction loss term as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 167, + 638, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 638, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 167, + 638, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r} = \\frac {1}{N J} \\sum_ {n = 1} ^ {N} \\sum_ {j = 1} ^ {J} \\exp \\left(- d \\left(x _ {n, j}, y _ {n, j}\\right)\\right) \\cdot \\left| d \\left(x _ {n, j}, y _ {n, j}\\right) - d \\left(x _ {n, j} ^ {(0)}, y _ {n, j}\\right) \\right|. \\tag {9}", + "image_path": "e208a63b13f827251b51fed7b49d0b02903a46f779a907a1445d9c2dc7d6069a.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "A. Ghosh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 167, + 324, + 278 + ], + "blocks": [ + { + "bbox": [ + 132, + 124, + 326, + 156 + ], + "lines": [ + { + "bbox": [ + 132, + 124, + 326, + 156 + ], + "spans": [ + { + "bbox": [ + 132, + 124, + 326, + 156 + ], + "type": "text", + "content": "Table 1: Dataset Comparisons. Comparing ReMoCap with existing multi-person interaction datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 167, + 324, + 278 + ], + "lines": [ + { + "bbox": [ + 132, + 167, + 324, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 324, + 278 + ], + "type": "table", + "html": "
Dataset (Chronologically)Motion FramesDuration (hours)Multi-view RGB VideosFinger Articulation
SBU [78]~ 7K0.13XX
NTU-26 [43]~ 22K0.47✓X
2C [55]~ 13K0.06XX
DanceDB [53]~ 1.44M4.00XX
DuetDance [39]~ 196K1.09XX
CHI3D [16]~ 486K2.70✓X
ExPI [24]~ 30K0.33✓X
InterHuman [42]~ 107M6.56XX
DD100 [60]~ 200K1.92X✓
ReMoCap (ours)~ 275.7K2.04✓✓
", + "image_path": "b8bde8858c9488d50183bc0e7fb196dc87da178e8e46dfc08e0bed6d76593f69.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 354, + 114, + 468, + 211 + ], + "blocks": [ + { + "bbox": [ + 354, + 114, + 468, + 211 + ], + "lines": [ + { + "bbox": [ + 354, + 114, + 468, + 211 + ], + "spans": [ + { + "bbox": [ + 354, + 114, + 468, + 211 + ], + "type": "image", + "image_path": "78f89a1eb4a2a8184ebde828d8e4c8ad8e1b1249baa0032b99b449d940958155.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 212, + 481, + 275 + ], + "lines": [ + { + "bbox": [ + 339, + 212, + 481, + 275 + ], + "spans": [ + { + "bbox": [ + 339, + 212, + 481, + 275 + ], + "type": "text", + "content": "Fig.4: Visualization of Distance Aware Reaction Loss. We use an exponentially decaying distance-aware reaction loss to focus more on the reactor's joints that are closer to the actor." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 310, + 482, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 310, + 482, + 371 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 482, + 371 + ], + "type": "text", + "content": "Kinematic Loss " + }, + { + "bbox": [ + 132, + 310, + 482, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_k" + }, + { + "bbox": [ + 132, + 310, + 482, + 371 + ], + "type": "text", + "content": ". To maintain the kinematic plausibility of the generated motions, we follow existing literature on regularizers for bone length consistency, foot contacts, and temporal consistency [56,57,68]. Our kinematic loss term is a weighted sum of the joint velocity loss " + }, + { + "bbox": [ + 132, + 310, + 482, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{vel}" + }, + { + "bbox": [ + 132, + 310, + 482, + 371 + ], + "type": "text", + "content": ", the joint acceleration loss " + }, + { + "bbox": [ + 132, + 310, + 482, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{acc}" + }, + { + "bbox": [ + 132, + 310, + 482, + 371 + ], + "type": "text", + "content": ", the bone length consistency loss " + }, + { + "bbox": [ + 132, + 310, + 482, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bone}" + }, + { + "bbox": [ + 132, + 310, + 482, + 371 + ], + "type": "text", + "content": ", and the foot sliding loss " + }, + { + "bbox": [ + 132, + 310, + 482, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{foot}" + }, + { + "bbox": [ + 132, + 310, + 482, + 371 + ], + "type": "text", + "content": ", as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 209, + 380, + 480, + 394 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 380, + 480, + 394 + ], + "spans": [ + { + "bbox": [ + 209, + 380, + 480, + 394 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {k} = \\lambda_ {v} \\mathcal {L} _ {v e l} + \\lambda_ {a} \\mathcal {L} _ {a c c} + \\lambda_ {b} \\mathcal {L} _ {b o n e} + \\lambda_ {f} \\mathcal {L} _ {f o o t}, \\tag {10}", + "image_path": "5097db32b1afcea03730bf469bf30eab8300be19b2da68dba27d33642b32c634.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 402, + 460, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 460, + 415 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 460, + 415 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 402, + 460, + 415 + ], + "type": "inline_equation", + "content": "\\lambda_v, \\lambda_a, \\lambda_b" + }, + { + "bbox": [ + 132, + 402, + 460, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 402, + 460, + 415 + ], + "type": "inline_equation", + "content": "\\lambda_f" + }, + { + "bbox": [ + 132, + 402, + 460, + 415 + ], + "type": "text", + "content": " are scalar weights (more details in the appendix)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 432, + 327, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 432, + 327, + 444 + ], + "spans": [ + { + "bbox": [ + 132, + 432, + 327, + 444 + ], + "type": "text", + "content": "3.4 Inference Time Spatial Guidance" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 453, + 482, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 482, + 619 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 482, + 619 + ], + "type": "text", + "content": "While our method synthesizes plausible reactive motions, it can sometimes spatially misalign the reactor's body to the actor's, especially at the arm joints, for fast-paced, contact-heavy interactions. This, in turn, affects finger-joint synthesis in the second stage of our cascaded diffusion. To improve spatial alignment, we leverage guidance functions [35, 52] that can provide gradients to nudge the sampling process towards a certain direction. We design a guidance function " + }, + { + "bbox": [ + 132, + 453, + 482, + 619 + ], + "type": "inline_equation", + "content": "G\\left(\\phi ,\\hat{\\phi}\\right)\\in \\mathbb{R}^{N\\times J_A\\times 3}" + }, + { + "bbox": [ + 132, + 453, + 482, + 619 + ], + "type": "text", + "content": ", which minimizes the distance between the " + }, + { + "bbox": [ + 132, + 453, + 482, + 619 + ], + "type": "inline_equation", + "content": "J_{A}" + }, + { + "bbox": [ + 132, + 453, + 482, + 619 + ], + "type": "text", + "content": " arm joints of the actor " + }, + { + "bbox": [ + 132, + 453, + 482, + 619 + ], + "type": "inline_equation", + "content": "(\\phi)" + }, + { + "bbox": [ + 132, + 453, + 482, + 619 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 132, + 453, + 482, + 619 + ], + "type": "inline_equation", + "content": "J_{A}" + }, + { + "bbox": [ + 132, + 453, + 482, + 619 + ], + "type": "text", + "content": " arm joints of the synthesized reactor " + }, + { + "bbox": [ + 132, + 453, + 482, + 619 + ], + "type": "inline_equation", + "content": "(\\hat{\\phi})" + }, + { + "bbox": [ + 132, + 453, + 482, + 619 + ], + "type": "text", + "content": ". Specifically, we re-apply our interaction masks " + }, + { + "bbox": [ + 132, + 453, + 482, + 619 + ], + "type": "inline_equation", + "content": "\\mathbb{1}_{H_A}" + }, + { + "bbox": [ + 132, + 453, + 482, + 619 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 453, + 482, + 619 + ], + "type": "inline_equation", + "content": "\\mathbb{1}_{H_R}" + }, + { + "bbox": [ + 132, + 453, + 482, + 619 + ], + "type": "text", + "content": " to determine which of the reactor's hands are more likely to be interacting with the actor, and minimize the distances between the arm joints of the corresponding sides as " + }, + { + "bbox": [ + 132, + 453, + 482, + 619 + ], + "type": "inline_equation", + "content": "G = \\arg \\min_{\\hat{\\phi}}\\left(\\left\\| \\mathbb{1}_{H_A}\\odot \\phi -\\mathbb{1}_{H_R}\\odot \\hat{\\phi}\\right\\|\\right)" + }, + { + "bbox": [ + 132, + 453, + 482, + 619 + ], + "type": "text", + "content": ". We then plug " + }, + { + "bbox": [ + 132, + 453, + 482, + 619 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 132, + 453, + 482, + 619 + ], + "type": "text", + "content": " into the denoising pipeline of our body diffusion module, as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 181, + 628, + 481, + 648 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 628, + 481, + 648 + ], + "spans": [ + { + "bbox": [ + 181, + 628, + 481, + 648 + ], + "type": "interline_equation", + "content": "X _ {B} ^ {(0)} = X _ {B} ^ {(0)} - \\gamma \\nabla_ {X _ {B} ^ {(0)}} G (\\phi , \\hat {\\phi}), \\quad \\gamma = \\text {g u i d a n c e s c a l e}. \\tag {11}", + "image_path": "a44f662604a9d6101ff1e469e81bbdc395fdc19b59b31297adfd009363db02e3.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "ReMoS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 264, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 264, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 264, + 129 + ], + "type": "text", + "content": "4 ReMoCap Dataset" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 251 + ], + "type": "text", + "content": "We propose the ReMoCap dataset to facilitate research on contact-based two-person interactions with finger-level articulations. It contains complex two-person interactions in two separate scenarios: the fast-paced, swing-style Lindy Hop dancing, and the martial arts of Ninjutsu. We consider these two types of motions for studying physical interactions for two main reasons. First, these motions encompass a significantly diverse spectrum of motions in terms of intensity, speed, and style. Second, choreographed motion generation is developing as an active field of research and we envision this dataset to facilitate research in interactive applications such as performing in virtual reality [9] and remote tutoring [53]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 262, + 482, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 454 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 454 + ], + "type": "text", + "content": "Data Collection. We invited 4 trained Lindy Hop dancers and 5 trained Ninjutsu artists to perform diverse interactions in a multi-view capture studio. We tracked their motions using a commercially available markerless, multi-view motion-capture system [8], which tracks 93 degrees-of-freedom driving 69 body joints. Capturing motions in a markerless setting enables the performers to move uninhibited, while also making the data suitable for training monocular or multiview motion capture methods under severe inter-person occlusions. We also capture RGB videos from 120 camera views for each sequence. The dataset includes 3D skeleton poses with full-body and finger annotations, foreground-background segmentation masks, and 3D surface reconstructions of the subjects. We capture sequences of multiple lengths totaling " + }, + { + "bbox": [ + 130, + 262, + 482, + 454 + ], + "type": "inline_equation", + "content": "\\sim 275.7\\mathrm{K}" + }, + { + "bbox": [ + 130, + 262, + 482, + 454 + ], + "type": "text", + "content": " frames (2.04 hours) of motion data from each view at a frame rate of 50 fps for the Lindy Hop and 25 fps for the Ninjutsu. Out of all the frames, around " + }, + { + "bbox": [ + 130, + 262, + 482, + 454 + ], + "type": "inline_equation", + "content": "150\\mathrm{K}" + }, + { + "bbox": [ + 130, + 262, + 482, + 454 + ], + "type": "text", + "content": " frames have hand-based interactions between the two characters where the closest distance between the finger joints of the actor and the reactor is within " + }, + { + "bbox": [ + 130, + 262, + 482, + 454 + ], + "type": "inline_equation", + "content": "50~\\mathrm{mm}" + }, + { + "bbox": [ + 130, + 262, + 482, + 454 + ], + "type": "text", + "content": ". We present more dataset statistics in the appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 466, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 466, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 466, + 482, + 647 + ], + "type": "text", + "content": "Dataset Comparison. Table 1 shows a comprehensive comparison of ReMoCap with existing multi-person interaction datasets. Existing datasets consisting of two-person interactions, such as SBU [78], K3HI [32], NTU-26 [43] and 2C [55], are limited in size and motion capture quality. They typically feature simple actions, such as handshakes, punching, pushing and kicking, with weak interactions, and do not capture hand motions. The recent ExPI [24] dataset captures Lindy Hop aerial sequences to model more complex interactions, and the InterHuman [42] dataset features both daily motions (e.g., passing objects, greeting, communicating) and professional motions (e.g., Taekwondo, Latin dance, boxing). However, these datasets do not provide finger-level motion capture data, which is a key requirement to intricately model inter-human activities. Two-person dance datasets, such as DuetDance [39] and DanceDB [53], also lack the hand motion data needed for modeling interactions. Only the concurrent work, Duolando " + }, + { + "bbox": [ + 130, + 466, + 482, + 647 + ], + "type": "inline_equation", + "content": "[60]^*" + }, + { + "bbox": [ + 130, + 466, + 482, + 647 + ], + "type": "text", + "content": " proposes the DD100 dataset with strong interactions between two dancers and provides hand motion data." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 100 + ], + "type": "text", + "content": "A. Ghosh et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 145, + 654, + 349, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 654, + 349, + 665 + ], + "spans": [ + { + "bbox": [ + 145, + 654, + 349, + 665 + ], + "type": "text", + "content": "*unpublished at the time of our paper submission" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 303, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 303, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 303, + 129 + ], + "type": "text", + "content": "5 Experiments and Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 224 + ], + "type": "text", + "content": "We conduct comprehensive experiments to evaluate ReMoS on multiple two-person datasets covering a wide range of interaction scenarios. We perform training and evaluation on multiple large-scale datasets, including 2C [55], ExPI [24], and our proposed ReMoCap. We provide the implementation details, report quantitative and qualitative comparisons on standard evaluation metrics and through a user study, report ablations, and show how to apply ReMoS as a motion editing tool. We provide details on the dataset preparation in the appendix." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 239, + 279, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 239, + 279, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 239, + 279, + 251 + ], + "type": "text", + "content": "5.1 Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "text", + "content": "As a pre-processing step, we normalize the two-person body poses by translating the actor's motion " + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "text", + "content": " and the reactor's motion " + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "text", + "content": " together, such that the root joint of " + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "text", + "content": " is at the global origin for all " + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "text", + "content": " frames. We then compute the relative 3D joint coordinates of " + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "inline_equation", + "content": "X_B" + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "inline_equation", + "content": "Y_B" + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "text", + "content": " w.r.t. the root joint of " + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "inline_equation", + "content": "Y_B" + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "text", + "content": ". For each hand, we compute the relative 3D joint coordinates of " + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "inline_equation", + "content": "X_H" + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "inline_equation", + "content": "Y_H" + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "text", + "content": " w.r.t. the corresponding wrist joints of " + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "inline_equation", + "content": "X_H" + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "inline_equation", + "content": "Y_H" + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "text", + "content": ". Using normalized inter-person, root-relative joint positions benefits the stability and convergence of our model. ReMoS uses " + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "inline_equation", + "content": "T = 500" + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "text", + "content": " diffusion steps where " + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}" + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "text", + "content": " changes linearly from 0.0002 to 0.02. We train for about 64K iterations on both sequences of ReMoCap using Adam [36] with a base learning rate of " + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "text", + "content": " and a batch size of 64. We decay the learning rate using a Step LR scheduler with a step size of 5 epochs and a decay rate of 0.99. We use " + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "inline_equation", + "content": "d = 256" + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "text", + "content": " for our latent embedding representation and use 6 layers in our transformer decoder with 4 heads for calculating the attention. The training takes around 8 and 11 hours on an NVIDIA RTX A4000 GPU for Lindy Hop and Ninjutsu, respectively. The inference time is " + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "inline_equation", + "content": "\\sim 24s" + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "text", + "content": " to generate 50 frames of full-body and finger motions (12.5s for body synthesis and then 11.5s for hand synthesis). We set " + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "inline_equation", + "content": "\\lambda_c, \\lambda_r, \\lambda_v = 10.0, \\lambda_a, \\lambda_k, \\lambda_b = 1.0" + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "inline_equation", + "content": "\\lambda_f = 20.0" + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "text", + "content": " (only applied after 100 training epochs) as the loss term weights in Eqns. 8 and 10. We set the guidance scale " + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "inline_equation", + "content": "\\gamma = 10^{-3}" + }, + { + "bbox": [ + 130, + 258, + 482, + 487 + ], + "type": "text", + "content": " in Eqn. 11." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 502, + 318, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 318, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 318, + 514 + ], + "type": "text", + "content": "5.2 Baselines and Ablated Versions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 629 + ], + "type": "text", + "content": "For baselines, we choose the closest motion synthesis methods in a two-person setting, namely, MixNMatch [20], InterFormer [11], ComMDM [54], Role-Aware Interaction Generation (RAIG) [65], and InterGen [42]. InterFormer [11] was originally trained in a reactive motion synthesis setting without additional annotations, such as action labels or text descriptions, and we maintain this setup. For the other methods, we mask out their input text/label embeddings to comply with our annotation-free setting. We re-train these methods on ReMoCap with a thorough hyper-parameter search and report their best performances. We provide more training details in the appendix." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 630, + 482, + 665 + ], + "type": "text", + "content": "We also compare our proposed ReMoS model with five of its ablated versions: w/o diffusion. Training ReMoS using a transformer encoder-decoder network without using DDPM strategy." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "ReMoS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 168, + 478, + 321 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": "Table 2: Quantitative Evaluation on ReMoCap. We compare ReMoS with our baselines and ablated versions (Sec. 5.2) on the ReMoCap dataset. We evaluate these methods on metrics such as MPJPE, MPJVE, FID, and Diversity. " + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": ": lower is better, " + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": ": higher is better, " + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": ": values closer to GT are better. Bold indicates best." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 168, + 478, + 321 + ], + "lines": [ + { + "bbox": [ + 132, + 168, + 478, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 168, + 478, + 321 + ], + "type": "table", + "html": "
MethodLindy HopNinjutsu
MPJPE (mm) ↓MPJVE (mm) ↓FID ↓ (body)FID ↓ (hands)Div →MPJPE (mm) ↓MPJVE (mm) ↓FID ↓ (body)FID ↓ (hands)Div →
GT----7.57----10.51
InterFormer [11]66.68.260.530.654.54270.23.40.570.686.48
MixNMatch [20]70.210.30.770.782.48257.25.20.740.724.83
ComMDM [54]59.44.410.320.537.48201.24.10.340.589.98
RAIG [65]71.24.320.470.638.45199.15.10.210.6310.11
InterGen [42]62.63.920.300.617.21172.63.90.320.579.98
ReMoS (ours)40.72.260.120.267.62139.23.30.160.3510.26
w/o diffusion72.54.910.580.744.04224.54.10.520.646.06
w/o cascading63.94.950.510.557.12223.64.20.420.756.62
w/o CoST-XA44.23.620.210.397.45176.63.60.270.418.91
w/o reaction loss44.63.510.220.387.31144.63.70.230.398.99
w/o spatial guidance41.92.340.120.267.62139.43.40.160.3510.26
", + "image_path": "b3bedbb1beb381c1b18a41d4208ea6f32277ecc74e1f7d8b678702f67472c5c9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 342, + 480, + 415 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 132, + 342, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 365 + ], + "type": "text", + "content": "- w/o cascading. Training ReMoS with an integrated transformer for all the joints instead of a cascaded strategy for body and hand generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 366, + 480, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 366, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 132, + 366, + 480, + 389 + ], + "type": "text", + "content": "- w/o CoST-XA. Using a spatial followed by a temporal cross-attention mapping instead of the combined spatio-temporal features for body synthesis." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 391, + 446, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 446, + 401 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 446, + 401 + ], + "type": "text", + "content": "- w/o reaction loss. Removing the reaction loss (Eqn. 9) in training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 403, + 479, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 479, + 415 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 479, + 415 + ], + "type": "text", + "content": "- w/o spatial guidance. Removing spatial guidance " + }, + { + "bbox": [ + 132, + 403, + 479, + 415 + ], + "type": "inline_equation", + "content": "G_{A}" + }, + { + "bbox": [ + 132, + 403, + 479, + 415 + ], + "type": "text", + "content": " (Sec. 3.4) in inference." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 431, + 281, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 281, + 443 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 281, + 443 + ], + "type": "text", + "content": "5.3 Quantitative Evaluation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 449, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 666 + ], + "type": "text", + "content": "We evaluate ReMoS on standard evaluation metrics. We measure the temporal consistency deviation from the ground truth 3D pose following [57] to report the mean per-joint positional error (MPJPE) and the mean per-frame, per-joint velocity error (MPJVE), both in " + }, + { + "bbox": [ + 130, + 449, + 482, + 666 + ], + "type": "inline_equation", + "content": "mm" + }, + { + "bbox": [ + 130, + 449, + 482, + 666 + ], + "type": "text", + "content": ", on the synthesized motions. We measure the Fréchet Inception Distance (FID) [27] to compare the distribution gap between the embedding spaces of the generated and ground-truth motions. As part of our method focuses on generating hand motions, we measure the FID score of body and hands separately for ReMoCap. We also compute the latent variance of the generated motions (Diversity) [67, 81]. Table 2 reports the quantitative evaluation of ReMoS with its baselines and ablations on ReMoCap. ReMoS achieves state-of-the-art performance for both the Lindy Hop pair-dance setting and the Ninjutsu setting. We observe that methods using denoising diffusion have higher Diversity compared to transformer-based [11] or GAN-based [20] methods. This enforces the variability claims of denoising diffusion models. We note at least " + }, + { + "bbox": [ + 130, + 449, + 482, + 666 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 666 + ], + "type": "text", + "content": " improvement in the MPJPE and around " + }, + { + "bbox": [ + 130, + 449, + 482, + 666 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 666 + ], + "type": "text", + "content": " improvement in the FID score for the reactor's body motion when using the proposed CoST-XA mechanism, confirming its benefit in learning fine-grained, inter-person dependencies across time. We also note an almost " + }, + { + "bbox": [ + 130, + 449, + 482, + 666 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 666 + ], + "type": "text", + "content": " improvement in the FID scores of" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "A. Ghosh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 359, + 169 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 359, + 169 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 359, + 169 + ], + "type": "text", + "content": "Table 3: Quantitative Evaluation on the ExPI and 2C datasets. We compare ReMoS with state-of-the-art motion synthesis methods on the ExPI [24] and 2C datasets [55]. " + }, + { + "bbox": [ + 132, + 114, + 359, + 169 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 132, + 114, + 359, + 169 + ], + "type": "text", + "content": ": lower is better, " + }, + { + "bbox": [ + 132, + 114, + 359, + 169 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 132, + 114, + 359, + 169 + ], + "type": "text", + "content": ": higher is better, " + }, + { + "bbox": [ + 132, + 114, + 359, + 169 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 132, + 114, + 359, + 169 + ], + "type": "text", + "content": ": values closer to GT are better. Bold indicates best." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "type": "table", + "bbox": [ + 132, + 178, + 360, + 270 + ], + "blocks": [ + { + "bbox": [ + 132, + 178, + 360, + 270 + ], + "lines": [ + { + "bbox": [ + 132, + 178, + 360, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 178, + 360, + 270 + ], + "type": "table", + "html": "
MethodExPI2C
MPJPE (mm) ↓MPJVE (mm) ↓FID ↓ (body) →Div
MPJPE (mm) ↓MPJVE (mm) ↓FID ↓ (body) →Div
GT---2.01---2.22
InterFormer [11]99.13.560.421.3190.75.110.521.45
MixNMatch [20]122.45.560.481.1862.46.010.471.24
ComMDM [54]121.45.410.452.4869.93.340.492.86
RAIG [65]131.23.960.532.5191.64.920.674.45
InterGen [42]100.63.910.432.0967.64.010.472.91
ReMoS (ours)97.93.520.411.9859.13.330.342.07
", + "image_path": "9c6599b67cb5b0b38ee705911180b0d247fc2c330e2b781026ba6b143edb9bca.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 367, + 168, + 481, + 258 + ], + "blocks": [ + { + "bbox": [ + 367, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 367, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 367, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 4: User Study Results. Mean scores on a five-point Likert scale (scores 1 - 5)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 367, + 168, + 481, + 258 + ], + "lines": [ + { + "bbox": [ + 367, + 168, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 367, + 168, + 481, + 258 + ], + "type": "table", + "html": "
MethodMotion Quality ↑Reaction Plausibility ↑
GT4.86 ± 0.544.72 ± 0.56
InterFormer2.52 ± 0.612.28 ± 0.57
MixNMatch1.92 ± 0.712.18 ± 0.57
ComMDM3.02 ± 0.473.12 ± 0.52
RAIG2.83 ± 0.672.48 ± 0.65
InterGen3.18 ± 0.573.19 ± 0.53
ReMoS3.79 ± 0.553.88 ± 0.54
", + "image_path": "bf342d683977dbf8a8e51cc8493fa9ddd7547e4b78f9b9677bdca992e685cb1f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 292, + 482, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 292, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 482, + 376 + ], + "type": "text", + "content": "the reactor's hand motion when using the cascaded strategy and the reaction loss. Further, the spatial guidance function fine-tuning improves MPJPE and MPJVE. The MPJPE values are overall higher in Ninjutsu than in Lindy Hop as the trajectory of the reactor is more diverse for Ninjutsu. We also report the evaluation of ReMoS with its baselines on the ExPI [24] and the 2C [55] datasets in Table 3. These datasets do not provide hand motions, so we only evaluate the reactor's body motions, and report state-of-the-art performance of ReMoS." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 392, + 218, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 218, + 404 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 218, + 404 + ], + "type": "text", + "content": "5.4 User Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 412, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 482, + 594 + ], + "type": "text", + "content": "We evaluate the visual quality of our captured ground truth data and the generated reactive motions through a user study. We show participants 26 different interaction sequences across the ground truth, our method, and its baselines. For each sequence, we randomly show them three methods side-by-side and ask them to rate the 3D motions they observe in terms of " + }, + { + "bbox": [ + 130, + 412, + 482, + 594 + ], + "type": "inline_equation", + "content": "(a)" + }, + { + "bbox": [ + 130, + 412, + 482, + 594 + ], + "type": "text", + "content": " the reactor's motion quality, irrespective of the actor's motion (Motion Quality), and " + }, + { + "bbox": [ + 130, + 412, + 482, + 594 + ], + "type": "inline_equation", + "content": "(b)" + }, + { + "bbox": [ + 130, + 412, + 482, + 594 + ], + "type": "text", + "content": " the plausibility of the reaction given the actor's motion (Reaction Plausibility). We ask the participants to rank these motions from '1' (worst score) to '5' (best score) in a five-point Likert scale. Table 4 reports the mean scores from the responses of 40 participants, excluding the responses that did not pass our validation checks. We notice the ground truth motions in ReMoCap achieve the highest score of around " + }, + { + "bbox": [ + 130, + 412, + 482, + 594 + ], + "type": "inline_equation", + "content": "96\\%" + }, + { + "bbox": [ + 130, + 412, + 482, + 594 + ], + "type": "text", + "content": ", indicating that participants perceive our captured ground truth motions to be natural-looking, with realistic interactions. ReMoS has the second best ranking of around " + }, + { + "bbox": [ + 130, + 412, + 482, + 594 + ], + "type": "inline_equation", + "content": "78\\%" + }, + { + "bbox": [ + 130, + 412, + 482, + 594 + ], + "type": "text", + "content": ", which is almost " + }, + { + "bbox": [ + 130, + 412, + 482, + 594 + ], + "type": "inline_equation", + "content": "21\\%" + }, + { + "bbox": [ + 130, + 412, + 482, + 594 + ], + "type": "text", + "content": " higher than the baselines, showing that it is preferred more over the baselines." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 609, + 347, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 347, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 347, + 621 + ], + "type": "text", + "content": "5.5 Qualitative Results and Applications" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Fig. 1 shows qualitative results of ReMoS on ReMoCap and highlights the synthesized hand interactions. We note that our actor and reactor are interchangeable, i.e., depending on the character driving the interaction at a given time," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "ReMoS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 168, + 114, + 199, + 155 + ], + "blocks": [ + { + "bbox": [ + 168, + 114, + 199, + 155 + ], + "lines": [ + { + "bbox": [ + 168, + 114, + 199, + 155 + ], + "spans": [ + { + "bbox": [ + 168, + 114, + 199, + 155 + ], + "type": "image", + "image_path": "bfc4dc9cae03538721721780d790826117b56aa80a9fcc4e5b1b9181996bda9b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 156, + 195, + 160 + ], + "lines": [ + { + "bbox": [ + 170, + 156, + 195, + 160 + ], + "spans": [ + { + "bbox": [ + 170, + 156, + 195, + 160 + ], + "type": "text", + "content": "Ground-truth" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 205, + 114, + 233, + 155 + ], + "blocks": [ + { + "bbox": [ + 205, + 114, + 233, + 155 + ], + "lines": [ + { + "bbox": [ + 205, + 114, + 233, + 155 + ], + "spans": [ + { + "bbox": [ + 205, + 114, + 233, + 155 + ], + "type": "image", + "image_path": "f7579fb4cd9f7dfc2674c4ffc0212fe9efcd7564fc0bcd2f47dc81e88ad2800d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 156, + 231, + 160 + ], + "lines": [ + { + "bbox": [ + 206, + 156, + 231, + 160 + ], + "spans": [ + { + "bbox": [ + 206, + 156, + 231, + 160 + ], + "type": "text", + "content": "MixNMatch" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 244, + 114, + 290, + 155 + ], + "blocks": [ + { + "bbox": [ + 244, + 114, + 290, + 155 + ], + "lines": [ + { + "bbox": [ + 244, + 114, + 290, + 155 + ], + "spans": [ + { + "bbox": [ + 244, + 114, + 290, + 155 + ], + "type": "image", + "image_path": "bb67196bc97a616b275cf401e571591afefe51bb76eb9d5cb05e5dde92b180d6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 256, + 156, + 277, + 161 + ], + "lines": [ + { + "bbox": [ + 256, + 156, + 277, + 161 + ], + "spans": [ + { + "bbox": [ + 256, + 156, + 277, + 161 + ], + "type": "text", + "content": "Interformer" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 296, + 114, + 321, + 156 + ], + "blocks": [ + { + "bbox": [ + 296, + 114, + 321, + 156 + ], + "lines": [ + { + "bbox": [ + 296, + 114, + 321, + 156 + ], + "spans": [ + { + "bbox": [ + 296, + 114, + 321, + 156 + ], + "type": "image", + "image_path": "032ff87b8b38f6da2b00e069b9ef40687907ec4c3cce76cb1daeb7d898d9ac56.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 156, + 358, + 162 + ], + "lines": [ + { + "bbox": [ + 302, + 156, + 358, + 162 + ], + "spans": [ + { + "bbox": [ + 302, + 156, + 358, + 162 + ], + "type": "text", + "content": "RAIG ComMDM" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 328, + 114, + 365, + 156 + ], + "blocks": [ + { + "bbox": [ + 328, + 114, + 365, + 156 + ], + "lines": [ + { + "bbox": [ + 328, + 114, + 365, + 156 + ], + "spans": [ + { + "bbox": [ + 328, + 114, + 365, + 156 + ], + "type": "image", + "image_path": "14e60147d0f2fdfdf80b1156d739e58f438515ce69ea2bf0236b9145ab838b7f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 337, + 156, + 358, + 161 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 156, + 358, + 161 + ], + "spans": [ + { + "bbox": [ + 337, + 156, + 358, + 161 + ], + "type": "text", + "content": "ComMDM" + } + ] + } + ], + "index": 11, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 370, + 114, + 399, + 156 + ], + "blocks": [ + { + "bbox": [ + 370, + 114, + 399, + 156 + ], + "lines": [ + { + "bbox": [ + 370, + 114, + 399, + 156 + ], + "spans": [ + { + "bbox": [ + 370, + 114, + 399, + 156 + ], + "type": "image", + "image_path": "85ed6bcffaa4653b19a80a1ae70d21d456f37ca4c5cc98efe65b481a7e1995f1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 377, + 156, + 394, + 161 + ], + "lines": [ + { + "bbox": [ + 377, + 156, + 394, + 161 + ], + "spans": [ + { + "bbox": [ + 377, + 156, + 394, + 161 + ], + "type": "text", + "content": "InterGen" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 406, + 114, + 445, + 156 + ], + "blocks": [ + { + "bbox": [ + 406, + 114, + 445, + 156 + ], + "lines": [ + { + "bbox": [ + 406, + 114, + 445, + 156 + ], + "spans": [ + { + "bbox": [ + 406, + 114, + 445, + 156 + ], + "type": "image", + "image_path": "65e249ccf60ca8f66ba665a0da719c9724dd576fe427fc9cd1170bd51f232473.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 420, + 156, + 435, + 161 + ], + "lines": [ + { + "bbox": [ + 420, + 156, + 435, + 161 + ], + "spans": [ + { + "bbox": [ + 420, + 156, + 435, + 161 + ], + "type": "text", + "content": "ReMoS" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 132, + 175, + 274, + 220 + ], + "blocks": [ + { + "bbox": [ + 207, + 166, + 407, + 175 + ], + "lines": [ + { + "bbox": [ + 207, + 166, + 407, + 175 + ], + "spans": [ + { + "bbox": [ + 207, + 166, + 407, + 175 + ], + "type": "text", + "content": "(a) Visual comparison of ReMoS with GT and baselines." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 175, + 274, + 220 + ], + "lines": [ + { + "bbox": [ + 132, + 175, + 274, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 175, + 274, + 220 + ], + "type": "image", + "image_path": "0dc11ec5b0b7ff7d5f43e287f9ffe2ef12eb7a47b3639159275186f32cf27710.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 221, + 261, + 232 + ], + "lines": [ + { + "bbox": [ + 145, + 221, + 261, + 232 + ], + "spans": [ + { + "bbox": [ + 145, + 221, + 261, + 232 + ], + "type": "text", + "content": "(b) Pose-completion application." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 239, + 482, + 306 + ], + "lines": [ + { + "bbox": [ + 131, + 239, + 482, + 306 + ], + "spans": [ + { + "bbox": [ + 131, + 239, + 482, + 306 + ], + "type": "text", + "content": "Fig.5: Qualitative Results and Applications. We show some visual results and the application of ReMoS as a motion editing tool. (a) The reactor (in blue) synthesized by ReMoS has the most plausible alignment with the actor (in red) compared to the baselines. (b) We manually control the right-hand wrist joint of the reactor and let ReMoS synthesize the remaining body joints conditioned on the actor. (c) ReMoS synthesizes the reactor's motion in-between the start and end frames." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 340, + 180, + 481, + 219 + ], + "blocks": [ + { + "bbox": [ + 340, + 180, + 481, + 219 + ], + "lines": [ + { + "bbox": [ + 340, + 180, + 481, + 219 + ], + "spans": [ + { + "bbox": [ + 340, + 180, + 481, + 219 + ], + "type": "image", + "image_path": "abb142b50b8f604756c834a3d51b1fdd5ce50faacdec7509a23abe58e23de257.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 221, + 477, + 232 + ], + "lines": [ + { + "bbox": [ + 343, + 221, + 477, + 232 + ], + "spans": [ + { + "bbox": [ + 343, + 221, + 477, + 232 + ], + "type": "text", + "content": "(c) Motion in-between application." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 328, + 482, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 482, + 448 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 482, + 448 + ], + "type": "text", + "content": "we can swap the actor and the reactor to produce the relevant reactive motions. Fig. 5a shows a visual comparison of ReMoS with the baselines for one frame of Lindy Hop motion. ReMoS synthesizes reaction with the most plausible alignment with the actor's motion. We provide detailed visual results in our supplementary video. Inspired by the applicability of diffusion-based methods for motion editing and controlling, we demonstrate how ReMoS can also be used as an interactive motion editing tool to control the reactor's motion as desired. In Figs. 5b and 5c, we show results from two different motion editing applications, namely, pose completion with controlled joints and motion in-between. We discuss more details in the appendix." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 463, + 220, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 463, + 220, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 463, + 220, + 475 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "ReMoS brings 3D motion conditioned reaction synthesis to a qualitatively new level by generating diverse, well-synchronized reactions for complex movements and plausible hand motions for contact-based interactions. It outperforms the existing baselines both quantitatively and in a user study. We also highlight some practical applications of our model, such as pose completion and motion in-between, which can lead to the development of useful generative assistants for animators, designers, and creative artists. Even though we utilize joint positions as parameterization due to their ready availability, we can adapt our approach to accommodate mesh-level inter-person contacts by introducing an offset to the contact threshold between the two bodies, thus simulating skin-to-skin interactions. We believe there is scope for improvement in the inference speed, where approaches such as Pro-DDPM [17] and DDIM [62] have shown success. Future directions also involve scaling the problem towards multi-person motion prediction [70] and considering scene-aware interactions for the characters, which would further enhance immersive user experiences." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "text", + "content": "A. Ghosh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 488, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 488, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 488, + 213 + ], + "type": "text", + "content": "This research was supported by the EU Horizon 2020 grant " + }, + { + "bbox": [ + 130, + 140, + 488, + 213 + ], + "type": "inline_equation", + "content": "\\mathrm{Carousel + }" + }, + { + "bbox": [ + 130, + 140, + 488, + 213 + ], + "type": "text", + "content": " (101017779), and the ERC Consolidator Grant 4DRepLy (770784). We thank Marc Jahan, Christopher Ruf, Michael Hiery, Thomas Leimkuhler and Sascha Huwer for helping with the Ninjutsu data collection, and Noshaba Cheema for helping with the Lindy Hop data collection. We also thank Janis Sprenger and Duarte David for helping with the motion visualizations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 232, + 197, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 232, + 197, + 244 + ], + "spans": [ + { + "bbox": [ + 132, + 232, + 197, + 244 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 257, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 257, + 480, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 257, + 480, + 290 + ], + "spans": [ + { + "bbox": [ + 138, + 257, + 480, + 290 + ], + "type": "text", + "content": "1. Ahuja, C., Ma, S., Morency, L.P., Sheikh, Y.: To react or not to react: End-to-end visual pose forecasting for personalized avatar during dyadic conversations. In: 2019 International conference on multimodal interaction (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 291, + 480, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 291, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 138, + 291, + 480, + 312 + ], + "type": "text", + "content": "2. Ao, T., Zhang, Z., Liu, L.: GestureDiffuCLIP: Gesture diffusion model with clip latents. SIGGRAPH (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 313, + 480, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 313, + 480, + 345 + ], + "spans": [ + { + "bbox": [ + 138, + 313, + 480, + 345 + ], + "type": "text", + "content": "3. Aristidou, A., Yiannakidis, A., Aberman, K., Cohen-Or, D., Shamir, A., Chrysanthou, Y.: Rhythm is a dancer: Music-driven motion synthesis with global structure. IEEE Transactions on Visualization and Computer Graphics (TVCG) (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 346, + 480, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 480, + 378 + ], + "type": "text", + "content": "4. Athanasiou, N., Petrovich, M., Black, M.J., Varol, G.: Teach: Temporal action composition for 3d humans. In: 2022 International Conference on 3D Vision (3DV) (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 380, + 480, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 480, + 422 + ], + "type": "text", + "content": "5. Bhattacharya, U., Childs, E., Rewkowski, N., Manocha, D.: Speech2affectivegestures: Synthesizing co-speech gestures with generative adversarial affective expression learning. In: Proceedings of the 29th ACM International Conference on Multimedia (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 423, + 480, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 423, + 480, + 466 + ], + "spans": [ + { + "bbox": [ + 138, + 423, + 480, + 466 + ], + "type": "text", + "content": "6. Bhattacharya, U., Rewkowski, N., Banerjee, A., Guhan, P., Bera, A., Manocha, D.: Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In: IEEE Conference on Virtual Reality and 3D User Interfaces (IEEE VR) (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 468, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 489 + ], + "type": "text", + "content": "7. Bjorck, N., Gomes, C.P., Selman, B., Weinberger, K.Q.: Understanding batch normalization. In: Advances in Neural Information Processing Systems (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 490, + 269, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 490, + 269, + 499 + ], + "spans": [ + { + "bbox": [ + 138, + 490, + 269, + 499 + ], + "type": "text", + "content": "8. https://capturey.com (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 480, + 533 + ], + "type": "text", + "content": "9. Chan, J.C., Leung, H., Tang, J.K., Komura, T.: A virtual reality dance training system using motion capture technology. IEEE transactions on learning technologies 4(2) (2010)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 480, + 566 + ], + "type": "text", + "content": "0. Chopin, B., Tang, H., Daoudi, M.: Bipartite graph diffusion model for human interaction generation. In: Winter Conference on Applications of Computer Vision (WACV) (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "type": "text", + "content": "1. Chopin, B., Tang, H., Otberdout, N., Daoudi, M., Sebe, N.: Interaction transformer for human reaction generation. IEEE Transactions on Multimedia (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 590, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 609 + ], + "type": "text", + "content": "2. Cummins, A.: In search of the ninja: the historical truth of ninjutsu. The History Press (2012)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "text", + "content": "3. Dabral, R., Mughal, M.H., Golyanik, V., Theobalt, C.: Mofusion: A framework for denoising-diffusion-based motion synthesis. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "text", + "content": "4. Egges, A., Papagiannakis, G., Magnenat-Thalmann, N.: Presence and interaction in mixed reality environments. The Visual Computer (2007)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "ReMoS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "15. Elfwing, S., Uchibe, E., Doya, K.: Sigmoid-weighted linear units for neural network function approximation in reinforcement learning. Neural Networks (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "type": "text", + "content": "16. Fieraru, M., Zanfir, M., Oneata, E., Popa, A.I., Olaru, V., Sminchisescu, C.: Three-dimensional reconstruction of human interactions. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 204 + ], + "type": "text", + "content": "17. Gandikota, R., Brown, N.: Pro-ddpm: Progressive growing of variable denoising diffusion probabilistic models for faster convergence. In: 33rd British Machine Vision Conference 2022, BMVC (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "type": "text", + "content": "18. Ghosh, A., Cheema, N., Oguz, C., Theobalt, C., Slusallek, P.: Synthesis of compositional animations from textual descriptions. In: International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 237, + 480, + 270 + ], + "type": "text", + "content": "19. Ghosh, A., Dabral, R., Golyanik, V., Theobalt, C., Slusallek, P.: IMoS: Intent-driven full-body motion synthesis for human-object interactions. In: Computer Graphics Forum. vol. 42. Wiley Online Library (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 303 + ], + "type": "text", + "content": "20. Goel, A., Men, Q., Ho, E.S.L.: Interaction Mix and Match: Synthesizing Close Interaction using Conditional Hierarchical GAN with Multi-Hot Class Embedding. Computer Graphics Forum (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 303, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 480, + 325 + ], + "type": "text", + "content": "21. Gu, D., Shim, J., Jang, J., Kang, C., Joo, K.: Contactgen: Contact-guided interactive 3d human generation for partners. AAAI (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 357 + ], + "type": "text", + "content": "22. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: Conference on Computer Vision and Pattern Recognition (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "text", + "content": "23. Guo, C., Zuo, X., Wang, S., Cheng, L.: Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In: European Conference on Computer Vision (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 391, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 480, + 424 + ], + "type": "text", + "content": "24. Guo, W., Bie, X., Alameda-Pineda, X., Moreno-Noguer, F.: Multi-person extreme motion prediction. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "text", + "content": "25. Habibie, I., Elgharib, M., Sarkar, K., Abdullah, A., Nyatsanga, S., Neff, M., Theobalt, C.: A motion matching-based framework for controllable gesture synthesis from speech. In: ACM SIGGRAPH Conference Proceedings (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 456, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 480, + 489 + ], + "type": "text", + "content": "26. Hanser, E., Mc Kevitt, P., Lunney, T., Condell, J.: Scenemaker: Intelligent multimodal visualisation of natural language scripts. In: Irish Conference on Artificial Intelligence and Cognitive Science. Springer (2009)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "text", + "content": "27. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "type": "text", + "content": "28. Ho, E.S., Komura, T.: Planning tangling motions for humanoids. In: IEEE-RAS International Conference on Humanoid Robots (2007)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 544, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 480, + 567 + ], + "type": "text", + "content": "29. Ho, E.S., Komura, T.: Character motion synthesis by topology coordinates. In: Computer graphics forum. Wiley Online Library (2009)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "text", + "content": "30. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 588, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 621 + ], + "type": "text", + "content": "31. Ho, J., Sahara, C., Chan, W., Fleet, D.J., Norouzi, M., Salimans, T.: Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research 23(1) (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "text", + "content": "32. Hu, T., Zhu, X., Guo, W.: Two-person interaction recognition based on key poses. Journal of Computational Information Systems (2014)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "33. Huang, S., Wang, Z., Li, P., Jia, B., Liu, T., Zhu, Y., Liang, W., Zhu, S.C.: Diffusion-based generation, optimization, and planning in 3d scenes. In: Proceed-" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 100 + ], + "type": "text", + "content": "A. Ghosh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "type": "text", + "content": "ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 172 + ], + "type": "text", + "content": "34. Huang, Y., Zhang, J., Liu, S., Bao, Q., Zeng, D., Chen, Z., Liu, W.: Genre-conditioned long-term 3d dance generation driven by music. In: International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "type": "text", + "content": "35. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Guided motion diffusion for controllable human motion synthesis. In: International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "text", + "content": "36. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "text", + "content": "37. Komura, T., Ho, E.S., Lau, R.W.: Animating reactive motion using momentum-based inverse kinematics. Computer Animation and Virtual Worlds (2005)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "38. Kulkarni, N., Rempe, D., Genova, K., Kundu, A., Johnson, J., Fouhey, D., Guibas, L.: Nifty: Neural object interaction fields for guided human motion synthesis. arXiv preprint arXiv:2307.07511 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 326 + ], + "type": "text", + "content": "39. Kundu, J.N., Buckchash, H., Mandikal, P., Jamkhandi, A., Radhakrishnan, V.B., et al.: Cross-conditioned recurrent networks for long-term synthesis of inter-person human motion interactions. In: Winter Conference on Applications of Computer Vision (WACV) (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "type": "text", + "content": "40. Li, J., Clegg, A., Mottaghi, R., Wu, J., Puig, X., Liu, C.K.: Controllable human-object interaction synthesis. arXiv preprint arXiv:2312.03913 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "text", + "content": "41. Li, J., Wu, J., Liu, C.K.: Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG) (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "text", + "content": "42. Liang, H., Zhang, W., Li, W., Yu, J., Xu, L.: InterGen: Diffusion-based multi-human motion generation under complex interactions. International Journal for Computer Vision (IJCV) (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "type": "text", + "content": "43. Liu, J., Shahroudy, A., Perez, M., Wang, G., Duan, L.Y., Kot, A.C.: NTU RGB+D 120: A large-scale benchmark for 3d human activity understanding. IEEE transactions on pattern analysis and machine intelligence (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "text", + "content": "44. Liu, X., Yi, L.: Geneoh diffusion: Towards generalizable hand-object interaction denoising via denoising diffusion. In: International Conference on Learning Representations (ICLR) (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "text", + "content": "45. Men, Q., Shum, H.P., Ho, E.S., Leung, H.: Gan-based reactive motion synthesis with class-aware discriminators for human-human interaction. Computers & Graphics (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 501, + 481, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 523 + ], + "type": "text", + "content": "46. Mousas, C.: Performance-driven dance motion control of a virtual partner character. In: IEEE Conference on Virtual Reality and 3D User Interfaces (VR) (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 523, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 481, + 566 + ], + "type": "text", + "content": "47. Mughal, M.H., Dabral, R., Habibie, I., Donatelli, L., Habermann, M., Theobalt, C.: Convofusion: Multi-modal conversational diffusion for co-speech gesture synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "text", + "content": "48. Ng, E., Joo, H., Hu, L., Li, H., Darrell, T., Kanazawa, A., Ginosar, S.: Learning to listen: Modeling non-deterministic dyadic facial motion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 600, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 621 + ], + "type": "text", + "content": "49. Petrovich, M., Black, M.J., Varol, G.: TEMOS: Generating diverse human motions from textual descriptions. In: European Conference on Computer Vision (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "50. Po, R., Yifan, W., Golyanik, V., Aberman, K., Barron, J.T., Amit H. Bermano, E.R.C., Dekel, T., Holynski, A., Kanazawa, A., Liu, C.K., Liu, L., Mildenhall, B., Nießner, M., Ommer, B., Theobalt, C., Wonka, P., Wetzstein, G.: State of the art on diffusion models for visual computing. arXiv pre-prints (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "ReMoS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "51. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. In: International Conference on Virtual Reality (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 192 + ], + "type": "text", + "content": "52. Rempe, D., Luo, Z., Bin Peng, X., Yuan, Y., Kitani, K., Kreis, K., Fidler, S., Litany, O.: Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 192, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 225 + ], + "type": "text", + "content": "53. Senecal, S., Nijdam, N.A., Aristidou, A., Magnenat-Thalmann, N.: Salsa dance learning evaluation and motion analysis in gamified virtual reality environment. Multimedia Tools and Applications (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 225, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 480, + 257 + ], + "type": "text", + "content": "54. Shafir, Y., Tevet, G., Kapon, R., Bermano, A.H.: Human motion diffusion as a generative prior. In: International Conference on Learning Representations (ICLR) (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 257, + 480, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 279 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 279 + ], + "type": "text", + "content": "55. Shen, Y., Yang, L., Ho, E.S.L., Shum, H.P.H.: Interaction-based human activity comparison. IEEE Transactions on Visualization and Computer Graphics (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 279, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 279, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 279, + 480, + 311 + ], + "type": "text", + "content": "56. Shimada, S., Golyanik, V., Xu, W., Pérez, P., Theobalt, C.: Neural monocular 3d human motion capture with physical awareness. ACM Transactions on Graphics (ToG) (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 311, + 480, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 480, + 343 + ], + "type": "text", + "content": "57. Shimada, S., Golyanik, V., Xu, W., Theobalt, C.: Physcap: Physically plausible monocular 3d motion capture in real time. ACM Transactions on Graphics (ToG) (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 343, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 480, + 365 + ], + "type": "text", + "content": "58. Shum, H.P., Komura, T., Shiraishi, M., Yamazaki, S.: Interaction patches for multi-character animation. ACM transactions on graphics (TOG) 27(5) (2008)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 365, + 480, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 397 + ], + "type": "text", + "content": "59. Shum, H.P., Komura, T., Yamazaki, S.: Simulating competitive interactions using singly captured motions. In: Proceedings of ACM symposium on Virtual reality software and technology (2007)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 397, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 430 + ], + "type": "text", + "content": "60. Siyao, L., Gu, T., Yang, Z., Lin, Z., Liu, Z., Ding, H., Yang, L., Loy, C.C.: Duolando: Follower gpt with off-policy reinforcement learning for dance accompaniment. In: International Conference on Learning Representations (ICLR) (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 430, + 480, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 480, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 480, + 461 + ], + "type": "text", + "content": "61. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International Conference on Machine Learning (ICML) (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 461, + 480, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 480, + 483 + ], + "type": "text", + "content": "62. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "type": "text", + "content": "63. Spring, H.: Swing and the lindy hop: dance, venue, media, and tradition. American Music (1997)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "type": "text", + "content": "64. Starke, S., Zhao, Y., Komura, T., Zaman, K.: Local motion phases for learning multi-contact character movements. ACM Transactions on Graphics (TOG) (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 525, + 480, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 548 + ], + "type": "text", + "content": "65. Tanaka, M., Fujiwara, K.: Role-aware interaction generation from textual description. In: International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 548, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 480, + 590 + ], + "type": "text", + "content": "66. Tanke, J., Zhang, L., Zhao, A., Tang, C., Cai, Y., Wang, L., Wu, P.C., Gall, J., Keskin, C.: Social diffusion: Long-term multiple human motion anticipation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 590, + 480, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 612 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 612 + ], + "type": "text", + "content": "67. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Bermano, A.H., Cohen-Or, D.: Human motion diffusion model. arXiv preprint arXiv:2209.14916 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 612, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 612, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 612, + 480, + 633 + ], + "type": "text", + "content": "68. Tseng, J., Castellon, R., Liu, K.: Edge: Editable dance generation from music. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "69. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems (2017)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 100 + ], + "type": "text", + "content": "A. Ghosh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "70. Wang, J., Xu, H., Narasimhan, M., Wang, X.: Multi-person 3d motion prediction with multi-range transformers. Advances in Neural Information Processing Systems (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 193 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 193 + ], + "type": "text", + "content": "71. Wang, Z., Chen, Y., Jia, B., Li, P., Zhang, J., Zhang, J., Liu, T., Zhu, Y., Liang, W., Huang, S.: Move as you say, interact as you can: Language-guided human motion generation with scene affordance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 193, + 482, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 193, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 482, + 224 + ], + "type": "text", + "content": "72. Xie, Y., Jampani, V., Zhong, L., Sun, D., Jiang, H.: OmniControl: Control any joint at any time for human motion generation. In: International Conference on Learning Representations (ICLR) (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 224, + 482, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 257 + ], + "type": "text", + "content": "73. Xing, J., Xia, M., Zhang, Y., Cun, X., Wang, J., Wong, T.T.: Codetalker: Speech-driven 3d facial animation with discrete motion prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 257, + 482, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 482, + 289 + ], + "type": "text", + "content": "74. Xu, S., Li, Z., Wang, Y.X., Gui, L.Y.: InterDiff: Generating 3d human-object interactions with physics-informed diffusion. In: International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 289, + 482, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 321 + ], + "type": "text", + "content": "75. Ye, Y., Li, X., Gupta, A., De Mello, S., Birchfield, S., Song, J., Tulsiani, S., Liu, S.: Affordance diffusion: Synthesizing hand-object interactions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 321, + 482, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 482, + 354 + ], + "type": "text", + "content": "76. Yoon, Y., Ko, W.R., Jang, M., Lee, J., Kim, J., Lee, G.: Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In: International Conference on Robotics and Automation (ICRA). IEEE (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 354, + 482, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 386 + ], + "type": "text", + "content": "77. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: Physdiff: Physics-guided human motion diffusion model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 386, + 482, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 429 + ], + "type": "text", + "content": "78. Yun, K., Honorio, J., Chattopadhyay, D., Berg, T.L., Samaras, D.: Two-person interaction detection using body-pose features and multiple instance learning. In: IEEE computer society conference on computer vision and pattern recognition workshops (2012)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 429, + 482, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 482, + 461 + ], + "type": "text", + "content": "79. Zamfirescu-Pereira, J., Wong, R.Y., Hartmann, B., Yang, Q.: Why johnny can't prompt: How non-ai experts try (and fail) to design LLM prompts. In: Proceedings of Conference on Human Factors in Computing Systems (CHI) (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 461, + 482, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 504 + ], + "type": "text", + "content": "80. Zhang, J., Zhang, Y., Cun, X., Zhang, Y., Zhao, H., Lu, H., Shen, X., Shan, Y.: T2m-gpt: Generating human motion from textual descriptions with discrete representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 504, + 482, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 504, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 504, + 482, + 536 + ], + "type": "text", + "content": "81. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 536, + 482, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 482, + 568 + ], + "type": "text", + "content": "82. Zhang, W., Dabral, R., Leimkuhler, T., Golyanik, V., Habermann, M., Theobalt, C.: ROAM: Robust and object-aware motion generation using neural pose descriptors. In: International Conference on 3D Vision (3DV) (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 568, + 482, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 601 + ], + "type": "text", + "content": "83. Zhang, X., Bhatnagar, B.L., Starke, S., Guzov, V., Pons-Moll, G.: Couch: Towards controllable human-chair interactions. In: European Conference on Computer Vision (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "spans": [ + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "type": "text", + "content": "84. Zhou, Z., Wang, B.: Ude: A unified driving engine for human motion generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 633, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 633, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 633, + 482, + 666 + ], + "type": "text", + "content": "85. Zhu, L., Liu, X., Liu, X., Qian, R., Liu, Z., Yu, L.: Taming diffusion models for audio-driven co-speech gesture generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "ReMoS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/ReNoise_ Real Image Inversion Through Iterative Noising/4a999b7d-d6fa-416b-8345-a221ff72abc9_content_list.json b/2024/ReNoise_ Real Image Inversion Through Iterative Noising/4a999b7d-d6fa-416b-8345-a221ff72abc9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b223444dc1a64bb67c4e0ae34b573e5b7d0c1554 --- /dev/null +++ b/2024/ReNoise_ Real Image Inversion Through Iterative Noising/4a999b7d-d6fa-416b-8345-a221ff72abc9_content_list.json @@ -0,0 +1,1743 @@ +[ + { + "type": "text", + "text": "ReNoise: Real Image Inversion Through Iterative Noising", + "text_level": 1, + "bbox": [ + 218, + 140, + 782, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Daniel Garibi1, Or Patashnik1, Andrey Voynov2, Hadar Averbuch-Elor1, and Daniel Cohen-Or1", + "bbox": [ + 217, + 210, + 782, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Tel Aviv University", + "bbox": [ + 428, + 253, + 573, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Google Research", + "bbox": [ + 437, + 268, + 563, + 282 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/41c090579a293cc48ee4ef721b81c4663f6dcb6c9d3cd0bd9c28676762b66240.jpg", + "image_caption": [ + "Fig. 1: Our ReNoise inversion technique can be applied to various diffusion models, including recent few-step ones. This figure illustrates the performance of our method with SDXL Turbo and LCM models, showing its effectiveness compared to DDIM inversion. Additionally, we demonstrate that the quality of our inversions allows prompt-driven editing. As illustrated on the right, our approach also allows for prompt-driven image edits." + ], + "image_footnote": [], + "bbox": [ + 220, + 315, + 782, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Recent advancements in text-guided diffusion models have unlocked powerful image manipulation capabilities. However, applying these methods to real images necessitates the inversion of the images into the domain of the pretrained diffusion model. Achieving faithful inversion remains a challenge, particularly for more recent models trained to generate images with a small number of denoising steps. In this work, we introduce an inversion method with a high quality-to-operation ratio, enhancing reconstruction accuracy without increasing the number of operations. Building on reversing the diffusion sampling process, our method employs an iterative renoising mechanism at each inversion sampling step. This mechanism refines the approximation of a predicted point along the forward diffusion trajectory, by iteratively applying the pretrained diffusion model, and averaging these predictions. We evaluate the performance of our ReNoise technique using various sampling algorithms and models, including recent accelerated diffusion models. Through comprehensive evaluations and comparisons, we show its effectiveness in terms of both accuracy and speed. Furthermore, we confirm that our method preserves editability by demonstrating text-driven image editing on real images.", + "bbox": [ + 259, + 558, + 740, + 820 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Diffusion Models $\\cdot$ Image Editing $\\cdot$ Inversion", + "bbox": [ + 261, + 825, + 645, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/396a494924ea216dcc23a0b15469fcfd96f59ecf191bad8d9535bf9ae0b534bc.jpg", + "image_caption": [ + "Fig.2: The diffusion process samples a Gaussian noise and iteratively denoises it until reaching the data distribution. At each point along the denoising trajectory, the model predicts a direction, $\\epsilon_{\\theta}(z_t)$ , to step to the next point along the trajectory. To invert a given image from the distribution, the direction from $z_{t}$ to $z_{t + 1}$ is approximated with the inverse of the direction from $z_{t}$ to $z_{t - 1}$ denoted by a dotted blue line." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 493, + 218 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c90c3ee96a4c920e3787eedde0a66016d70d313e0d0b559b867869e1e6f5ca13.jpg", + "image_caption": [ + "Fig.3: Comparing reconstruction results of plain DDIM Inv. with SDXL to DDIM Inv. with one renoising iteration." + ], + "image_footnote": [], + "bbox": [ + 522, + 154, + 794, + 323 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 375, + 375, + 390 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large-scale text-to-image diffusion models have revolutionized the field of image synthesis [19,37,39,40]. In particular, many works have shown that these models can be employed for various types of image manipulation [4,7-9,15,17,23,29,33, 35,47]. To edit real images, many of these techniques often require the inversion of the image into the domain of the diffusion model. That is, given a real image $z_{0}$ , one has to find a Gaussian noise $z_{T}$ , such that denoising $z_{T}$ with the pretrained diffusion model reconstructs the given real image $z_{0}$ . The importance of this task for real image manipulation has prompted many efforts aimed at achieving accurate reconstruction [16,21,30,31].", + "bbox": [ + 212, + 407, + 787, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The diffusion process consists of a series of denoising steps $\\{\\epsilon_{\\theta}(z_t,t)\\}_{t = T}^1$ , which form a trajectory from the Gaussian noise to the model distribution (see Figure 2). Each denoising step is computed by a trained network, typically implemented as a UNet, which predicts $z_{t - 1}$ from $z_{t}$ [19]. The output of the model at each step forms a direction from $z_{t}$ to $z_{t - 1}$ [45]. These steps are not invertible, in the sense that the model was not trained to predict $z_{t}$ from $z_{t - 1}$ . Thus, the problem of inverting a given image is a challenge, and particularly for real images, as they are not necessarily in the model distribution (see Figure 3).", + "bbox": [ + 212, + 542, + 787, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present an inversion method with a high quality-to-operation ratio, which achieves superior reconstruction accuracy for the same number of UNet operations. We build upon the commonly used approach of reversing the diffusion sampling process, which is based on the linearity assumption that the direction from $z_{t}$ to $z_{t+1}$ can be approximated by the negation of the direction from $z_{t}$ to $z_{t-1}$ [12, 43] (see Figure 2). To enhance this approximation, we employ the fixed-point iteration methodology [10]. Specifically, given $z_{t}$ , we begin by using the common approximation to get an initial estimate for $z_{t+1}$ , denoted by $z_{t+1}^{(0)}$ . Then, we iteratively renoise $z_{t}$ , following the direction implied by $z_{t+1}^{(k)}$ to obtain $z_{t+1}^{(k+1)}$ . After repeating this ronoising process several times, we apply an averaging on $z_{t+1}^{(k)}$ to form a more accurate direction from $z_{t}$ to $z_{t+1}$ .", + "bbox": [ + 212, + 664, + 787, + 843 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "D. Garibi et al.", + "bbox": [ + 271, + 114, + 375, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We show that this approach enables longer strides along the inversion trajectory while improving image reconstruction. Therefore, our method can also be effective with diffusion models trained to generate images using a small number of denoising steps [27, 42]. Furthermore, despite the need to repeatedly renoise in each step of the inversion process, the longer strides lead to a more favorable tradeoff of UNet operations for reconstruction quality.", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Through extensive experiments, we demonstrate the effectiveness of our method in both image reconstruction and inversion speed. We validate the versatility of our approach across different samplers and models, including recent time-distilled diffusion models (e.g., SDXL-Turbo [42]). Importantly, we demonstrate that the editability of the inversion achieved by our method allows a wide range of text-driven image manipulations (see Figure 1).", + "bbox": [ + 212, + 237, + 799, + 328 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 332, + 387, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image Editing via Diffusion Models Recent advancements in diffusion models [12, 19] have resulted in unprecedented diversity and fidelity in visual content creation guided by free-form text prompts [36, 37, 39, 40]. Text-to-image models do not directly support text-guided image editing. Therefore, harnessing the power of these models for image editing is a significant research area and many methods have utilized these models for different types of image editing [4, 7-9, 11, 14, 15, 17, 18, 23, 29, 33, 35, 47, 48]. A common approach among these methods requires inversion [21, 31, 43, 49] to edit real images, i.e., obtaining a latent code $z_{T}$ such that denoising it with the pretrained diffusion model returns the original image. Specifically, in this approach two backward processes are done simultaneously using $z_{T}$ . One of the processes reconstructs the image using the original prompt, while the second one injects features from the first process (e.g., attention maps) to preserve some properties of the original image while manipulating other aspects of it.", + "bbox": [ + 212, + 354, + 787, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inversion in Diffusion Models Initial efforts in image inversion for real image editing focused on GANs [2,3,5,6,13,34,38,46,52-54]. The advancements in diffusion models, and in diffusion-based image editing in particular have recently prompted works studying the inversion of a diffusion-based denoising process. This inversion depends on the sampler algorithm used during inference, which can be deterministic [43] or non-deterministic [19, 22]. Inversion methods can be accordingly categorized into two: methods that are suitable for deterministic sampling, and methods suitable for non-deterministic sampling.", + "bbox": [ + 212, + 568, + 787, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Methods that approach the deterministic inversion commonly rely on the DDIM sampling method [43], and build upon DDIM inversion [12, 43]. Mokady et al. [31] observed that the use of classifier-free guidance during inference magnifies the accumulated error of DDIM inversion and therefore leads to poor reconstruction. Following this observation, several works [16, 30, 31] focused on solving this issue by replacing the null text token with a different embedding, which is found after an optimization process or by a closed solution. However, excluding [31] which requires a lengthy optimization, these methods are limited by the reconstruction accuracy of DDIM inversion, which can be poor, especially when a small number of denoising steps is done. In our work, we present a method", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "ReNoise: Real Image Inversion Through Iterative Noising", + "bbox": [ + 349, + 114, + 730, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "that improves the reconstruction quality of DDIM inversion and therefore can be integrated with methods that build on it.", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Another line of work [21, 50] tackles the inversion of DDPM sampler [19]. In these works [21, 50], instead of inverting the image into an initial noise $z_{T}$ , a series of noises $\\{z_{T},\\epsilon_{T},\\dots,\\epsilon_{1}\\}$ is obtained. The definition of this noises series ensures that generating an image with it returns the original input image. However, these methods require a large number of inversion and denoising steps to allow image editing. Applying these methods with an insufficient number of steps leads to too much information encoded in $\\{\\epsilon_T,\\dots,\\epsilon_1\\}$ which limits the ability to edit the generated image. As shall be shown, The editability issue of these methods is particularly evident in few-steps models [26, 27, 42].", + "bbox": [ + 212, + 176, + 787, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Most relevant to our work, two recent inversion methods [28,32] also use the fixed-point iteration technique. Specifically, they improve the reconstruction accuracy of DDIM inversion [43] with Stable Diffusion [39] without introducing a significant computational overhead. In our work, we focus on the problem of real image inversion for recently introduced few-step diffusion models, where the difficulties encountered by previous methods are amplified. Furthermore, we show that our inversion method successfully works with various models and different samplers.", + "bbox": [ + 212, + 314, + 787, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Few Steps Models Recently, new methods [26, 27, 41, 42, 44] that fine-tune text-to-image diffusion models enabled a significant reduction of the number of steps needed for high-quality image generation. While standard diffusion models typically require 50 denoising steps to generate high-quality images, recent accelerated models achieve high-quality synthesis with 1-4 steps only. These new methods pave the way for interactive editing workflows. However, as we show in this paper, using current methods for the inversion of an image with a small number of steps degrades the reconstruction quality in terms of accuracy [12, 43] or editability [21, 50].", + "bbox": [ + 212, + 441, + 787, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 214, + 588, + 331, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 ReNoise Inversion", + "text_level": 1, + "bbox": [ + 214, + 625, + 413, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reversing the Sampler Samplers play a critical role in the diffusion-based image synthesis process. They define the noising and denoising diffusion processes and influence the processes' trajectories and quality of the generated images. While different samplers share the same pre-trained UNet model (denoted by $\\epsilon_{\\theta}$ ) as their backbone, their sampling approaches diverge, leading to nuanced differences in output. The goal of the denoising sampler is to predict the latent code at the previous noise level, $z_{t-1}$ , based on the current noisy data $z_{t}$ , the pretrained UNet model, and a sampled noise, $\\epsilon_{t}$ . Various first-order denoising sampling algorithms adhere to the form:", + "bbox": [ + 212, + 651, + 787, + 787 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t - 1} = \\phi_ {t} z _ {t} + \\psi_ {t} \\epsilon_ {\\theta} (z _ {t}, t, c) + \\rho_ {t} \\epsilon_ {t}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 790, + 785, + 805 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c$ represents a text embedding condition, and $\\phi_t, \\psi_t$ , and $\\rho_t$ denote sampler parameters. At each step, these parameters control the extent to which the", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "D. Garibi et al.", + "bbox": [ + 271, + 114, + 375, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "previous noise is removed $(\\phi_t)$ , the significance assigned to the predicted noise from the UNet $(\\psi_t)$ , and the given weight to the additional noise introduced $(\\rho_t)$ .", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A given image $z_0$ can be inverted by reformulating Equation 1 and applying it iteratively:", + "bbox": [ + 212, + 176, + 784, + 205 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t} = \\frac {z _ {t - 1} - \\psi_ {t} \\epsilon_ {\\theta} \\left(z _ {t} , t , c\\right) - \\rho_ {t} \\epsilon_ {t}}{\\phi_ {t}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 199, + 784, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where for non-deterministic samplers, a series of random noises $\\{\\epsilon_t\\}_{t=1}^T$ is sampled and used during both inversion and image generation processes. However, directly computing $z_t$ from Equation 2 is infeasible since it relies on $\\epsilon_\\theta(z_t, t, c)$ which, in turn, depends on $z_t$ , creating a circular dependency. To solve this implicit function, Dhariwal et al. [12] propose using the approximation $\\epsilon_\\theta(z_t, t, c) \\approx \\epsilon_\\theta(z_{t-1}, t, c)$ :", + "bbox": [ + 212, + 231, + 782, + 321 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t} ^ {(1)} = \\frac {z _ {t - 1} - \\psi_ {t} \\epsilon_ {\\theta} \\left(z _ {t - 1} , t , c\\right) - \\rho_ {t} \\epsilon_ {t}}{\\phi_ {t}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 315, + 784, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This method has several limitations. First, the assumption underlying the approximation used in [12] is that the number of inversion steps is large enough, implying a trajectory close to linear. This assumption restricts the applicability of this inversion method in interactive image editing with recent few-step diffusion models [26, 27, 42, 44], as the inversion process would take significantly longer than inference. Second, this method struggles to produce accurate reconstructions in certain cases, such as highly detailed images or images with large smooth regions, see Figure 3. Moreover, we observe that this inversion method is sensitive to the prompt $c$ and may yield poor results for certain prompts.", + "bbox": [ + 212, + 347, + 784, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ReNoise In a successful inversion trajectory, the direction from $z_{t-1}$ to $z_t$ aligns with the direction from $z_t$ to $z_{t-1}$ in the denoising trajectory. To achieve this, we aim to improve the approximation of $\\epsilon_{\\theta}(z_t,t,c)$ in Eq. 2 compared to the one used in [12]. Building on the fixed-point iteration technique [10], our approach better estimates the instance of $z_t$ that is inputted to the UNet, rather than relying on $z_{t-1}$ .", + "bbox": [ + 212, + 492, + 782, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Intuitively, we utilize the observation that $z_{t}^{(1)}$ (from Eq. 3) offers a more precise estimate of $z_{t}$ compared to $z_{t-1}$ . Therefore, employing $z_{t}^{(1)}$ as the input to the UNet is likely to yield a more accurate direction, thus contributing to reducing the overall error in the inversion step. We illustrate this observation in Figure 5. Iterating this process generates a series of estimations for $z_{t}$ , denoted by $\\{z_{t}^{(k)}\\}_{k=1}^{\\mathcal{K}+1}$ . While the fixed-point iteration technique [10] does not guarantee convergence of this series in the general case, in Section 4, we empirically show that convergence holds in our setting. However, as the convergence is not monotonic, we refine our prediction of $z_{t}$ by averaging several $\\{z_{t}^{(k)}\\}$ , thus considering more than a single estimation of $z_{t}$ . See Figure 6 for an intuitive illustration.", + "bbox": [ + 212, + 583, + 784, + 741 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In more detail, our method iteratively computes estimations of $z_{t}$ during each inversion step $t$ by renoising the noisy latent $z_{t-1}$ multiple times, each with a different noise prediction (see Figure 4). Beginning with $z_{t}^{(1)}$ , in the $k$ -th renoising iteration, the input to the UNet is the result of the previous iteration, $z_{t}^{(k)}$ . Then, $z_{t}^{(k+1)}$ is calculated using the inverted sampler while maintaining $z_{t-1}$ as the starting point of the step. After $\\mathcal{K}$ renoising iterations, we obtain a set of", + "bbox": [ + 212, + 742, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "ReNoise: Real Image Inversion Through Iterative Noising", + "bbox": [ + 349, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5742e3c782e5d830b04cea56a12460576f346867f29852aa2c51d010d3e236e6.jpg", + "image_caption": [ + "Fig. 4: Method overview. Given an image $z_0$ , we iteratively compute $z_1, \\dots, z_T$ , where each $z_t$ is calculated from $z_{t-1}$ . At each time step, we apply the UNet $(\\epsilon_\\theta) \\mathcal{K} + 1$ times, each using a better approximation of $z_t$ as the input. The initial approximation is $z_{t-1}$ . The next one, $z_t^{(1)}$ , is the result of the reversed sampler step (i.e., DDIM). The reversed step begins at $z_{t-1}$ and follows the direction of $\\epsilon_\\theta(z_{t-1}, t)$ . At the $k$ renoising iteration, $z_t^{(k)}$ is the input to the UNet, and we obtain a better $z_t$ approximation. For the last iterations, we optimize $\\epsilon_\\theta(z_t^{(k)}, t)$ to increase editability. As the final denoising direction, we use the average of the UNet predictions of the last few iterations." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 785, + 305 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig.5: Geometric intuition for ReNoise. At each inversion step, we estimate $z_{t}$ (marked with a red star) based on $z_{t-1}$ . The straightforward approach is to use the negated direction of the denoising step from $z_{t-1}$ , assuming the trajectory is approximately linear. However, this assumption is inaccurate, especially in few-step models, where the size of the steps is large. We use the linearity assumption as an initial estimation and keep improving this estimation. We recalculate the denoising step from the previous estimation (which is closer to the target $z_{t}$ ) and then proceed with its negated direction from $z_{t-1}$ (see the orange vectors).", + "bbox": [ + 212, + 422, + 532, + 617 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b9e58c72e7338254c2d815309cd5e7fa86ddf0ec2dbf1dca4521e55476645c43.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 446, + 790, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "estimations $\\{z_t^{(k)}\\}_{k = 1}^{\\mathcal{K} + 1}$ . The next point on the inversion trajectory, $z_{t}$ , is then defined as their weighted average, where $w_{k}$ is the weight assigned to $z_{t}^{(k)}$ . For a detailed description of our method, refer to Algorithm 1.", + "bbox": [ + 212, + 622, + 787, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Reconstruction-Editability Tradeoff", + "text_level": 1, + "bbox": [ + 214, + 679, + 555, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Enhance Editability The goal of inversion is to facilitate the editing of real images using a pretrained image generation model. While the renoising approach attains highly accurate reconstruction results, we observe that the resulting $z_{T}$ lacks editability. This phenomenon can be attributed to the reconstruction-editability tradeoff in image generative models [46]. To address this limitation, we incorporate a technique to enhance the editability of our method.", + "bbox": [ + 212, + 703, + 787, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It has been shown [33] that the noise maps predicted during the inversion process often diverge from the statistical properties of uncorrelated Gaussian white noise, thereby affecting editability. To tackle this challenge, we follow pix2pix-", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "D. Garibi et al.", + "bbox": [ + 271, + 114, + 375, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "zero [33] and regularize the predicted noise at each step, $\\epsilon_{\\theta}(z_t,t,c)$ , using the following loss terms.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "First, we encourage $\\epsilon_{\\theta}(z_t,t,c)$ to follow the same distribution as $\\epsilon_{\\theta}(z_t^{\\prime},t,c)$ , where $z_{t}^{\\prime}$ represents the input image $z_{0}$ with added random noise corresponding to the noise level at timestep $t$ . We do so by dividing $\\epsilon_{\\theta}(z_t,t,c)$ and $\\epsilon_{\\theta}(z_t^{\\prime},t,c)$ into small patches (e.g., $4\\times 4$ ), and computing the KL-divergence between corresponding patches. We denote this loss term by $\\mathcal{L}_{\\mathrm{patch - KL}}$ . Second, we utilize $\\mathcal{L}_{\\mathrm{pair}}$ proposed in pix2pix-zero [33], which penalizes correlations between pairs of pixels. We leverage these losses to enhance the editability of our method, and denote the combination of them as $\\mathcal{L}_{\\mathrm{edit}}$ . For any renoising iteration $k$ where $w_{k} > 0$ , we regularize the UNet's prediction $\\epsilon_{\\theta}(z_t^{(k)},t,c)$ using $\\mathcal{L}_{\\mathrm{edit}}$ before computing $z_{t}^{(k + 1)}$ . See line 9 in Algorithm 1.", + "bbox": [ + 212, + 176, + 787, + 332 + ], + "page_idx": 6 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: ReNoise Inversion" + ], + "code_body": "1 Input: An image $z_{0}$ , number of renoising steps $\\mathcal{K}$ , number of inversion steps $T$ , a series of renoising weights $\\{w_{k}\\}_{k=1}^{\\mathcal{K}+1}$ . \n2 Output: A noisy latent $z_{T}$ and set of noises $\\{\\epsilon_{t}\\}_{t=1}^{T}$ . \n3 for $t = 0,1,\\ldots,T$ do \n4 sample $\\epsilon_{t} \\sim \\mathcal{N}(0,I)$ \n5 $z_{t}^{(0)} \\gets z_{t-1}$ \n6 $z_{t}^{(\\mathrm{avg})} \\gets 0$ \n7 for $k = 0,\\ldots,\\mathcal{K}$ do \n8 $\\delta_{t}^{k} \\gets \\epsilon_{\\theta}(z_{t}^{(k)},t)$ \n9 $\\delta_{t}^{k} \\gets \\text{Enhance-edit}(\\delta_{t}^{k},w_{k+1})$ \n10 $z_{t}^{(k+1)} \\gets \\text{Inverse-Step}(z_{t-1},\\delta_{t}^{k})$ \n11 end \n// Average ReNoised predictions \n23 Function Noise-Corr $(z_{t},t,\\epsilon_{t},z_{t-1})$ : \n12 $z_{t}^{(\\mathrm{avg})} \\gets \\sum_{k=1}^{\\mathcal{K}+1} w_{k} \\cdot z_{t}^{(k)}$ \n24 $\\delta_{t} \\gets \\epsilon_{\\theta}(z_{t},t)$ \n25 $\\epsilon_{t} \\gets \\epsilon_{t} - \\nabla_{\\epsilon_{t}}\\frac{1}{\\rho_{t}}(z_{t-1} - \\phi_{t}z_{t} - \\psi_{t}\\delta_{t})$ \n26 return $\\epsilon_{t}$ \nend \nreturn $(z_{T},\\{\\epsilon_{t}\\}_{t=1}^{T})$", + "bbox": [ + 215, + 359, + 797, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Noise Correction in Non-deterministic Samplers Non-deterministic samplers, in which $\\rho_t > 0$ , introduce noise $(\\epsilon_t)$ at each denoising step. Previous methods [21,50] suggested using $\\epsilon_t$ to bridge the gap between the inversion and denoising trajectories in DDPM inversion. Specifically, given a pair of points $z_{t - 1},z_{t}$ on the inversion trajectory, we denote by $\\hat{z}_{t - 1}$ the point obtained by denoising $z_{t}$ . Ideally, $z_{t - 1}$ and $\\hat{z}_{t - 1}$ should be identical. We define:", + "bbox": [ + 212, + 628, + 787, + 719 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {t} = \\frac {1}{\\rho_ {t}} \\left(z _ {t - 1} - \\phi_ {t} z _ {t} - \\psi_ {t} \\epsilon_ {\\theta} \\left(z _ {t}, t, c\\right)\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 718, + 785, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Integrating this definition into Eq. 1 yields $\\hat{z}_{t - 1} = z_{t - 1}$ . However, we found that replacing $\\epsilon_t$ with the above definition affects editability. Instead, we suggest a more tender approach, optimizing $\\epsilon_t$ based on Eq. 4 as our guiding objective:", + "bbox": [ + 212, + 750, + 782, + 792 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {t} = \\epsilon_ {t} - \\nabla_ {\\epsilon_ {t}} \\frac {1}{\\rho_ {t}} \\left(z _ {t - 1} - \\phi_ {t} z _ {t} - \\psi_ {t} \\epsilon_ {\\theta} \\left(z _ {t}, t, c\\right)\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 792, + 785, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This optimization improves the reconstruction fidelity while preserving the distribution of the noisy-latents.", + "bbox": [ + 212, + 819, + 782, + 849 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "ReNoise: Real Image Inversion Through Iterative Noising", + "bbox": [ + 349, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/425c549fb37a8697b4d1460cd6a27c7a715f365226c14e05a66f37013253ea2f.jpg", + "image_caption": [ + "Fig. 6: Schematic illustration of the ReNoise convergence process to the true inversion of $z_{t-1}$ . While estimates may converge non-monotonically to the unknown target $z_t$ , we found that averaging them improves true value estimation. Typically, the initial iteration exhibits an exponential decrease in the norm between consecutive elements." + ], + "image_footnote": [], + "bbox": [ + 250, + 143, + 509, + 217 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/48283f72459cd1c42321171bc038cb2f7c828989379fbd874f737b41062e1091.jpg", + "image_caption": [ + "Fig.7: Average distance between consequent estimations $z_{t}^{(k)}$ , and $z_{t}^{(k + 1)}$ . Vertical bars indicate the standard deviation. The averages are computed over 32 images and 10 different timesteps." + ], + "image_footnote": [], + "bbox": [ + 555, + 159, + 784, + 231 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Convergence Discussion", + "text_level": 1, + "bbox": [ + 215, + 321, + 483, + 339 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we first express the inversion process as a backward Euler process and our renoising iterations as fixed-point iterations. While these iterations do not converge in the general case, in the supplementary materials we present a toy example where they yield accurate inversions. Then, we analyze the convergence of the renoising iterations in our real-image inversion scenario and empirically verify our method's convergence.", + "bbox": [ + 212, + 345, + 782, + 434 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inversion Process as Backward Euler The denoising process of diffusion models can be mathematically described as solving an ordinary differential equation (ODE). A common method for solving such equations is the Euler method, which takes small steps to approximate the solution. For ODE in the form of $y'(t) = f(t,y(t))$ , Euler solution is defined as:", + "bbox": [ + 214, + 438, + 787, + 513 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\ny _ {n + 1} = y _ {n} + h \\cdot f (t _ {n}, y _ {n}),\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 515, + 594, + 531 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $h$ is the step size. The inversion process can be described as solving ODE using the backward Euler method (or implicit Euler method) [1]. This method is similar to forward Euler, with the difference that $y_{n + 1}$ appears on both sides of the equation:", + "bbox": [ + 212, + 535, + 782, + 594 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\ny _ {n + 1} = y _ {n} + h \\cdot f \\left(t _ {n + 1}, y _ {n + 1}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 594, + 609, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For equations lacking an algebraic solution, several techniques estimate $y_{n+1}$ iteratively. As we described in Section 3.1, the inversion process lacks a closed-form solution, as shown in Equation 2. To address this, the ReNoise method leverages fixed-point iterations, which we refer to as reonising iterations, to progressively refine the estimate of $y_{n+1}$ :", + "bbox": [ + 212, + 612, + 782, + 686 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\ny _ {n + 1} ^ {(0)} = y _ {n}, \\quad y _ {n + 1} ^ {(k + 1)} = y _ {n} + h \\cdot f (t _ {n + 1}, y _ {n + 1} ^ {(k)}.\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 688, + 661, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In our ReNoise method, we average these renoising iterations to mitigate convergence errors, leading to improvement in the reconstruction quality.", + "bbox": [ + 212, + 715, + 782, + 746 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ReNoise Convergence During the inversion process, we aim to find the next noise level inversion, denoted by $\\hat{z}_t$ , such that applying the denoising step to $\\hat{z}_t$ recovers the previous state, $z_{t - 1}$ . Given the noise estimation $\\epsilon_{\\theta}(z_t,t)$ and a given $z_{t - 1}$ , the ReNoise mapping defined in Section 3.1 can be written as $\\mathcal{G}:z_t\\to \\mathrm{InverseStep}(z_{t - 1},\\epsilon_\\theta (z_t,t))$ . For example, in the case of using DDIM sampler the mapping is $\\mathcal{G}(z_t) = \\frac{1}{\\phi_t} (z_{t - 1} - \\psi_t\\epsilon_\\theta (z_t,t))$ . The point $\\hat{z}_t$ , which is", + "bbox": [ + 214, + 750, + 787, + 843 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "D. Garibi et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mapped to $z_{t-1}$ after the denoising step, is a stationary point of this mapping. Given $z_t^{(1)}$ , the first approximation of the next noise level $z_t$ , our goal is to show that the sequence $z_t^{(k)} = \\mathcal{G}^{k-1}(z_t^{(1)})$ , $k \\to \\infty$ converges. As the mapping $\\mathcal{G}$ is continuous, the limit point would be its stationary point. The definition of $\\mathcal{G}$ gives:", + "bbox": [ + 212, + 146, + 782, + 224 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| z _ {t} ^ {(k + 1)} - z _ {t} ^ {(k)} \\right\\| = \\left\\| \\mathcal {G} \\left(z _ {t} ^ {(k)}\\right) - \\mathcal {G} \\left(z _ {t} ^ {(k - 1)}\\right) \\right\\|,\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 218, + 643, + 238 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where the norm is always assumed as the $l_{2}$ -norm. For the ease of the notations, we define $\\varDelta^{(k)}=z_t^{(k)}-z_t^{(k-1)}$ . For convergence proof, it is sufficient to show that the sum of norms of these differences converges, which will imply that $z_{t}^{(k)}$ is the Cauchy sequence. Below we check that in practice $\\| \\varDelta^{(k)}\\|$ decreases exponentially as $k\\to \\infty$ and thus has finite sum. In the assumption that $\\mathcal{G}$ is $\\mathcal{C}^2$ -smooth, the Taylor series conducts:", + "bbox": [ + 212, + 239, + 784, + 332 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\| \\varDelta^ {(k + 1)} \\| = \\| \\mathcal {G} (z _ {t} ^ {(k)}) - \\mathcal {G} (z _ {t} ^ {(k - 1)}) \\| = \\\\ \\left\\| \\mathcal {G} \\left(z _ {t} ^ {(k - 1)}\\right) + \\frac {\\partial \\mathcal {G}}{\\partial z} \\right| _ {z _ {t} ^ {(k - 1)}} \\cdot \\Delta^ {(k)} + O \\left(\\left\\| \\Delta^ {(k)} \\right\\| ^ {2}\\right) - \\mathcal {G} \\left(z _ {t} ^ {(k - 1)}\\right) \\| = \\\\ \\left. \\left\\| \\frac {\\partial \\mathcal {G}}{\\partial z} \\right| _ {z _ {t} ^ {(k - 1)}} \\cdot \\Delta^ {(k)} + O (\\left\\| \\Delta^ {(k)} \\right\\| ^ {2}) \\right\\| \\leq \\left\\| \\frac {\\partial \\mathcal {G}}{\\partial z} \\right| _ {z _ {t} ^ {(k - 1)}} \\| \\cdot \\| \\Delta^ {(k)} \\| + O (\\left\\| \\Delta^ {(k)} \\right\\| ^ {2}) = \\\\ \\frac {\\psi_ {t}}{\\phi_ {t}} \\cdot \\left\\| \\frac {\\partial \\epsilon_ {\\theta}}{\\partial z} \\right| _ {z _ {t} ^ {(k - 1)}} \\| \\cdot \\| \\Delta^ {(k)} \\| + O (\\| \\Delta^ {(k)} \\| ^ {2}) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 235, + 335, + 767, + 449 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Thus, in a sufficiently small neighborhood, the convergence dynamics is defined by the scaled Jacobian norm $\\frac{\\psi_t}{\\phi_t} \\cdot \\left\\| \\frac{\\partial \\epsilon_\\theta}{\\partial z} \\right|_{z_t^{(k-1)}}$ . In the supplementary materials, we show this scaled norm estimation for the SDXL diffusion model for various steps and ReNoise iterations indices $(k)$ . Remarkably, the ReNoise indices minimally impact the scale factor, consistently remaining below 1. This confirms in practice the convergence of the proposed algorithm. Notably, the highest scaled norm values occur at smaller $t$ (excluding the first step) and during the initial renoising iteration. This validates the strategy of not applying ReNoise in early steps, where convergence tends to be slower compared to other noise levels. Additionally, the scaled norm value for the initial $t$ approaches 0, which induces almost immediate convergence.", + "bbox": [ + 212, + 449, + 784, + 617 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Figure 7 illustrates the exponential decrease in distances between consecutive elements $z_{t}^{(k)}$ and $z_{t}^{(k + 1)}$ , which confirms the algorithm's convergence towards the stationary point of the operator $\\mathcal{G}$ . The proposed averaging strategy is aligned with the conclusions described above, and also converges to the desired stationary point. In The supplementary materials, we present a validation for this claim.", + "bbox": [ + 212, + 619, + 782, + 696 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 699, + 375, + 715 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we conduct extensive experiments to validate the effectiveness of our method. We evaluate both the reconstruction quality of our inversion and its editability. To demonstrate the versatility of our approach, we apply it to four models, SD [39], SDXL [36], SDXL Turbo [42], and LCM-LoRA [27], with SDXL Turbo and LCM-LoRA being few-step models. Additionally, we use various sampling algorithms including both deterministic and non-deterministic ones. Implementation details for each model are provided in the supplementary materials. Following previous works [5, 31], we quantitatively evaluate our", + "bbox": [ + 212, + 719, + 782, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "ReNoise: Real Image Inversion Through Iterative Noising", + "bbox": [ + 349, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 1: Image reconstruction results with a fixed number of 100 UNet operations. Each row showcases the results obtained using different combinations of inversion steps, denoising steps, and renoising iterations, totaling 100 operations. As observed, allocating some of the operations to renoising iterations improves the reconstruction quality while maintaining the same execution time.", + "bbox": [ + 212, + 148, + 493, + 286 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/ef2973db3f0bd86bc1978d9448b04e58eb89ae24190f27f2d3cc167f3f5f79be.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Image Reconstruction Results
Inv StepsInf StepsReNoise StepsL2 ↓PSNR ↑LPIPS ↓
505000.0036426.0230.06273
752500.0038225.4660.06605
802000.0040825.0450.07099
901000.0102320.2490.10305
252520.0018229.5690.03637
202030.0016729.8840.03633
101080.0023028.1560.04678
", + "bbox": [ + 214, + 289, + 496, + 412 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 2: Quantitative ablation study on SDXL Turbo. We demonstrate the impact of each component of our inversion method on reconstruction results. The results improve with additional renoising iterations and significant enhancements occur through averaging final estimations. Additionally, we observe a reconstruction-editability trade-off, with edit losses causing degradation that is effectively mitigated by Noise Correction.", + "bbox": [ + 506, + 143, + 787, + 297 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/b3d67507f20d5da40a685fd0b93acdb664b15bbefb2b7ecee9d6fad3578c1d5c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Ablation - Image Reconstruction
L2 ↓PSNR ↑LPIPS↓
Euler Inversion0.070011.7840.20337
+ 1 ReNoise0.055212.7960.20254
+ 4 ReNoise0.024916.5210.14821
+ 9 ReNoise0.012619.7020.10850
+ Averaging ReNoise0.008721.4910.08832
+ Edit Losses0.027618.4320.12616
+ Noise Correction0.019622.0770.08469
", + "bbox": [ + 509, + 299, + 790, + 412 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "method with three metrics: $L_{2}$ , LPIPS [51], and PSNR. Unless stated otherwise, for both inversion and generation we use the prompt obtained from BLIP2 [24].", + "bbox": [ + 212, + 415, + 787, + 446 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Reconstruction and Speed", + "text_level": 1, + "bbox": [ + 214, + 450, + 478, + 465 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We begin by evaluating the reconstruction-speed tradeoff. The main computational cost of both the inversion and denoising processes is the forward pass through the UNet. In each renoising iteration, we perform one forward pass, which makes it computationally equal to a standard inversion step (as done in DDIM Inversion for example). In the following experiments, we compare the results of a sampler reversing with our method, where we match the number of UNet passes between the methods. For example, 8 steps of sampler reversing are compared against 4 steps with one renoising iteration at each step.", + "bbox": [ + 212, + 470, + 787, + 590 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Qualitative Results In Figure 8 we show qualitative results of image reconstruction on SDXL Turbo [42]. Here, we utilize DDIM as the sampler, and apply four denoising steps for all configurations. Each row exhibits results obtained using a different amount of UNet operations. In our method, we apply four inversion steps, and a varying number of renoising iterations. As can be seen, the addition of renoising iterations gradually improves the reconstruction results. Conversely, employing more inversion steps proves insufficient for capturing all details in the image, as evident by the background of the car, or even detrimental to the reconstruction, as observed in the Uluro example.", + "bbox": [ + 212, + 594, + 787, + 731 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Quantitative Results For the quantitative evaluation, we use the MS-COCO 2017 [25] validation dataset. Specifically, we retain images with a resolution greater than $420 \\times 420$ , resulting in a dataset containing 3,865 images.", + "bbox": [ + 212, + 734, + 787, + 779 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We begin by evaluating both the sampler reversing approach and our ReNoise method, while varying the number of UNet operations during the inversion process and keeping the number of denoising steps fixed. This experiment is conducted using various models (SDXL, SDXL Turbo, LCM) and samplers. For all", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "D. Garibi et al.", + "bbox": [ + 271, + 114, + 375, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/06f5e129beec9c5708f7efc9239a1d39b4d7c391a5fa130c3cf4d14d2b0a970e.jpg", + "image_caption": [ + "Fig.8: Qualitative comparison between DDIM Inversion and our ReNoise method on SDXL Turbo (4 denoising steps). The first row displays the original images. The following rows display reconstructions from both methods, using the same number of UNet operations. DDIM employs smaller strides, while ReNoise utilizes more renoising steps." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 488, + 404 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/fefb9658da53f8e6004d929868ba26da842bf11bb0b72813f9e0c438eb103048.jpg", + "image_caption": [ + "Fig. 9: Ablation study on SDXL Turbo. The first row presents the input image. In each subsequent row, we show the reconstruction results using an additional component of our inversion method. The images in the bottom row represent the results obtained by our full method. NC stands for Noise Correction." + ], + "image_footnote": [], + "bbox": [ + 513, + 142, + 781, + 426 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "models, we utilize the DDIM [43] sampler. In addition, we employ the Ancestral-Euler scheduler for SDXL Turbo, and the default LCM sampler for LCM-LoRA. We set the number of denoising steps to 50 for SDXL, and to 4 for SDXL Turbo and LCM-LoRA. Quantitative results, using PSNR as the metric, are presented in Figure 11. We evaluate our method using different configurations. The x-axis refers to the number of UNet operations in the inversion process. Other metrics results are provided in the supplementary materials.", + "bbox": [ + 212, + 547, + 787, + 654 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As depicted in the graphs, incorporating additional renoising iterations proves to be more beneficial for image reconstruction compared to adding more inversion steps. Note that the performance of the Ancestral-Euler and LCM samplers noticeably degrades when the number of inversion steps exceeds the number of denoising steps. Unlike DDIM, these samplers have $\\varPhi_t\\approx 1$ , resulting in an increase in the latent vector's norm beyond what can be effectively denoised in fewer steps. In this experiment, we maintain the same number of UNet operations for both ReNoise and the sampler reversing approach. However, in ReNoise, the number of inversion steps remains fixed, and the additional operations are utilized for renoising iterations, refining each point on the inversion trajectory. Consequently, our method facilitates improved reconstruction when using these noise samplers.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "ReNoise: Real Image Inversion Through Iterative Noising", + "bbox": [ + 349, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/73d002b752f357a55d5bced7a5dd06a7bc6cca1654498a3320305f17127d0034.jpg", + "image_caption": [ + "Fig. 10: LCM Editing Results. We showcase two examples of real images, each followed by three edits. The text below each edited image indicates the specific word or phrase replaced or added to the original prompt for that specific edit." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 789, + 239 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/acff721aa486e340a5a365ddb7f88be1de13d158959fb388ce73e5d484ee4189.jpg", + "image_caption": [ + "Fig. 11: Image reconstruction results comparing sampler reversing inversion techniques across different samplers (e.g., vanilla DDIM inversion) with our ReNoise method using the same sampler. The number of denoising steps remains constant. However, the number of UNet passes varies, with the sampler reversing approach increasing the number of inversion steps, while our method increases the number of renoising iterations. We present various configuration options for our method, including options with or without edit enhancement loss and Noise Correction (NC)." + ], + "image_footnote": [], + "bbox": [ + 220, + 282, + 406, + 377 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/9dd21dbad5845d21963121b5403376123930d2c7f6438b40648649d27366aad9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 282, + 594, + 377 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3eac0152bacc04aa5316a7a9fc35c993e0fd0e7f7e169a34a269fba38abc4b33.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 282, + 785, + 377 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We continue by evaluating both the sampler reversing approach and our method while maintaining a fixed total number of UNet operations for the inversion and denoising processes combined. The results for SDXL with DDIM are presented in Table 1. The table displays various combinations of inversion, denoising, and renoising steps, totaling 100 UNet operations. Despite employing longer strides along the inversion and denoising trajectories, our ReNoise method yields improved reconstruction accuracy, as evident in the table. Furthermore, a reduced number of denoising steps facilitates faster image editing, especially since it commonly involves reusing the same inversion for multiple edits.", + "bbox": [ + 212, + 477, + 787, + 613 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.2 Reconstruction and Editability", + "text_level": 1, + "bbox": [ + 214, + 616, + 517, + 632 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In Figure 10, we illustrate editing results generated by our method with LCM LoRA [27]. These results were obtained by inverting the image using a source prompt and denoising it with a target prompt. Each row exhibits an image followed by three edits accomplished by modifying the original prompt. These edits entail either replacing the object word or adding descriptive adjectives to it. As can be seen, the edited images retain the details present in the original image. For instance, when replacing the cat with a koala, the details in the background are adequately preserved.", + "bbox": [ + 212, + 633, + 787, + 755 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Ablation Studies", + "text_level": 1, + "bbox": [ + 214, + 761, + 401, + 775 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Figure 9 qualitatively demonstrates the effects of each component in our method, highlighting their contribution to the final outcome. Here, we use SDXL Turbo model [42], with the Ancestral-Euler sampler, which is non-deterministic. As our baseline, we simply reverse the sampler process. The reconstruction, while", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "D. Garibi et al.", + "bbox": [ + 271, + 114, + 375, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2abb376c359ed2ac846535570dac97e91daa302ad0e8e68fd08eff26f4882a61.jpg", + "image_caption": [ + "Fig. 12: Comparison with edit-friendly with SDXL Turbo. We inverted the image with the prompt \"a cat is sitting in front of a mirror\" and applied edits.", + "Fig. 13: Image reconstruction comparisons with SD. While NTI and our method achieve comparable results, ours demonstrates significant speed improvement." + ], + "image_footnote": [], + "bbox": [ + 207, + 141, + 794, + 271 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "semantically capturing the main object, fails to reproduce the image's unique details. For example, in the bird image, the reconstruction contains a bird standing on a branch, but the branch is in a different pose and the bird is completely different. Using 9 ReNoise iterations significantly improves the reconstruction, recovering finer details like the bird's original pose and branch texture. However, some subtle details, such as the bird's colors or the color in Brad Pitt's image, remain incomplete. Averaging the final iterations effectively incorporates information from multiple predictions, leading to a more robust reconstruction that captures finer details. Regularize the UNet's noise prediction with $\\mathcal{L}_{\\mathrm{edit}}$ can introduce minor artifacts to the reconstruction, evident in the smoother appearance of the hair of the two people on the left, or in the cake example. Finally, we present our full method by adding the noise correction technique.", + "bbox": [ + 212, + 330, + 787, + 512 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 2 quantitatively showcases the effect each component has on reconstruction results. As can be seen, the best results were obtained by our full method or by averaging the last estimations of $z_{t}$ . Our final method also offers the distinct advantage of getting an editable latent representation.", + "bbox": [ + 212, + 512, + 787, + 571 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In the Appendix, we present an ablation study to justify our editability enhancement and noise correction components.", + "bbox": [ + 212, + 571, + 785, + 602 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Comparisons", + "text_level": 1, + "bbox": [ + 215, + 607, + 370, + 622 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Inversion for Non-deterministic Samplers. In Figure 12 we show a qualitative comparison with \"an edit-friendly DDPM\" [21] where we utilize SDXL Turbo [42]. Specifically, we assess the performance of the edit-friendly method alongside our ReNoise method in terms of both reconstruction and editing.", + "bbox": [ + 212, + 628, + 787, + 688 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We observe that in non-deterministic samplers like DDPM, the parameter $\\rho_0$ in Equation 1 equals zero. This means that in the final denoising step, the random noise addition is skipped to obtain a clean image. In long diffusion processes (e.g., 50-100 steps), the final denoising step often has minimal impact as the majority of image details have already been determined. Conversely, shorter diffusion processes rely on the final denoising step to determine fine details of the image. Due to focusing solely on noise correction to preserve the original image in the inversion process, edit-friendly struggles to reconstruct fine details of the image, such as the shower behind the cat. However, our ReNoise method finds an inversion trajectory that faithfully reconstructs the image and does not rely", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "ReNoise: Real Image Inversion Through Iterative Noising", + "bbox": [ + 349, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "solely on noise corrections. This allows us to better reconstruct fine details such as the shower. Additionally, encoding a significant amount of information within only a few external noise vectors, $\\epsilon_{t}$ , limits editability in certain scenarios. See more examples in the appendix", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Null-prompt Inversion Methods In Figure 13, we present a qualitative comparison between our method and null-text based inversion methods. For this comparison, we utilize Stable Diffusion [39] since these methods rely on a CFG [20] mechanism, which is not employed in SDXL Turbo [42]. Specifically, we compare DDIM Inversion [43] with one renoising iteration to Null-Text Inversion (NTI) [31] and Negative-Prompt Inversion (NPI) [30]. Both NTI and NPI enhance the inversion process by replacing the null-text token embedding when applying CFG. Our method achieves results comparable to NTI, while NPI highlights the limitations of plain DDIM inversion. This is because NPI sets the original prompt as the negative prompt, essentially resulting in an inversion process identical to plain DDIM inversion. Regarding running time, our ReNoise inversion process takes 13 seconds, significantly faster than NTI's 3 minutes. For comparison, plain DDIM inversion and NPI each take 9 seconds.", + "bbox": [ + 212, + 212, + 789, + 411 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 417, + 361, + 434 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we have introduced ReNoise, a universal approach that enhances various inversion algorithms of diffusion models. ReNoise gently guides the inversion curve of a real image towards the source noise from which a denoising process reconstructs the image. ReNoise can be considered as a meta-algorithm that warps the trajectory of any iterative diffusion inversion process. Our experiments demonstrate that averaging the last few renoising iterations significantly enhances reconstruction quality. For a fixed amount of computation, ReNoise shows remarkably higher reconstruction quality and editability. The method is theoretically supported and our experiments reconfirm its effectiveness on a variety of diffusion models and sampling algorithms. Moreover, the method is numerically stable, and always converges to some inversion trajectory that eases hyperparameters adjustment.", + "bbox": [ + 212, + 444, + 787, + 625 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Beyond the net introduction of an effective inversion, the paper presents a twofold important contribution: an effective inversion for few-steps diffusion models, which facilitates effective editing on these models.", + "bbox": [ + 212, + 627, + 787, + 671 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "A limitation of ReNoise is the model-specific hyperparameter tuning required for Edit Enhancement and Noise Correction. While these hyperparameters remain stable for a given model, they may vary across models, and tuning them is necessary to achieve high reconstruction quality while maintaining editability.", + "bbox": [ + 212, + 672, + 787, + 734 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "While ReNoise demonstrates the potential for editing few-step diffusion models, more extensive testing with advanced editing methods is needed. It is worth noting that no such editing has been demonstrated for the few-step diffusion models. We believe and hope that our ReNoise method will pave the way for fast and effective editing methods based on the few-steps models. We also believe that ReNoise can be adapted to the challenging problem of inverting video-diffusion models.", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "D. Garibi et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Numerical Differential Equation Methods, chap. 2, pp. 45-121. John Wiley and Sons, Ltd (2003). https://doi.org/https://doi.org/10.1002/0470868279.ch2, https://onlinelibrary.wiley.com/doi/abs/10.1002/0470868279.ch2", + "2. Abdal, R., Qin, Y., Wonka, P.: Image2stylegan: How to embed images into the stylegan latent space? In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE (Oct 2019). https://doi.org/10.1109/iccv.2019.00453, http://dx.doi.org/10.1109/ICCV.2019.00453", + "3. Abdal, R., Qin, Y., Wonka, P.: Image2stylegan++: How to edit the embedded images? In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (Jun 2020). https://doi.org/10.1109/cvpr42600.2020.00832, http://dx.doi.org/10.1109/cvpr42600.2020.00832", + "4. Alaluf, Y., Garibi, D., Patashnik, O., Averbuch-Elor, H., Cohen-Or, D.: Cross-image attention for zero-shot appearance transfer (2023)", + "5. Alaluf, Y., Patashnik, O., Cohen-Or, D.: Restyle: A residual-based stylegan encoder via iterative refinement. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE (Oct 2021). https://doi.org/10.1109/iccv48922.2021.00664, http://dx.doi.org/10.1109/iccv48922.2021.00664", + "6. Alaluf, Y., Tov, O., Mokady, R., Gal, R., Bermano, A.H.: Hyperstyle: Stylegan inversion with hypernetworks for real image editing (2021)", + "7. Avrahami, O., Fried, O., Lischinski, D.: Blended latent diffusion. arXiv preprint arXiv:2206.02779 (2022)", + "8. Avrahami, O., Lischinski, D., Fried, O.: Blended diffusion for text-driven editing of natural images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18208-18218 (2022)", + "9. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions (2023)", + "0. Burden, R., Faires, J., Burden, A.: Numerical Analysis. Cengage Learning (2015)", + "1. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. ArXiv abs/2210.11427 (2022)", + "2. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis (2021)", + "3. Dinh, T.M., Tran, A.T., Nguyen, R., Hua, B.S.: Hyperinverter: Improving stylegan inversion via hypernetwork. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "4. Epstein, D., Jabri, A., Poole, B., Efros, A.A., Holynski, A.: Diffusion self-guidance for controllable image generation (2023)", + "5. Ge, S., Park, T., Zhu, J.Y., Huang, J.B.: Expressive text-to-image generation with rich text. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023)", + "6. Han, L., Wen, S., Chen, Q., Zhang, Z., Song, K., Ren, M., Gao, R., Stathopoulos, A., He, X., Chen, Y., Liu, D., Zhangli, Q., Jiang, J., Xia, Z., Srivastava, A., Metaxas, D.: Improving tuning-free real image editing with proximal guidance (2023)", + "7. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control (2022)", + "8. Hertz, A., Voynov, A., Fruchter, S., Cohen-Or, D.: Style aligned image generation via shared attention (2023)", + "19. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models (2020)", + "20. Ho, J., Salimans, T.: Classifier-free diffusion guidance (2022)" + ], + "bbox": [ + 225, + 176, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "ReNoise: Real Image Inversion Through Iterative Noising", + "bbox": [ + 349, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "21. Huberman-Spiegelglas, I., Kulikov, V., Michaeli, T.: An edit friendly ddpm noise space: Inversion and manipulations (2023)", + "22. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems 35, 26565-26577 (2022)", + "23. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Magic: Text-based real image editing with diffusion models. In: Conference on Computer Vision and Pattern Recognition 2023 (2023)", + "24. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)", + "25. Lin, T., Maire, M., Belongie, S.J., Bourdev, L.D., Girshick, R.B., Hays, J., Perona, P., Ramanan, D., Doll'a r, P., Zitnick, C.L.: Microsoft COCO: common objects in context. CoRR abs/1405.0312 (2014)", + "26. Luo, S., Tan, Y., Huang, L., Li, J., Zhao, H.: Latent consistency models: Synthesizing high-resolution images with few-step inference (2023)", + "27. Luo, S., Tan, Y., Patil, S., Gu, D., von Platen, P., Passos, A., Huang, L., Li, J., Zhao, H.: Lcm-lora: A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556 (2023)", + "28. Meiri, B., Samuel, D., Darshan, N., Chechik, G., Avidan, S., Ben-Ari, R.: Fixed-point inversion for text-to-image diffusion models (2023)", + "29. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations (2022)", + "30. Miyake, D., Iohara, A., Saito, Y., Tanaka, T.: Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models (2023)", + "31. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models (2022)", + "32. Pan, Z., Gherardi, R., Xie, X., Huang, S.: Effective real image editing with accelerated iterative diffusion inversion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15912-15921 (October 2023)", + "33. Parmar, G., Kumar Singh, K., Zhang, R., Li, Y., Lu, J., Zhu, J.Y.: Zero-shot image-to-image translation. In: Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Proceedings. SIGGRAPH '23, ACM (Jul 2023). https://doi.org/10.1145/3588432.3591513, http://dx.doi.org/10.1145/3588432.3591513", + "34. Parmar, G., Li, Y., Lu, J., Zhang, R., Zhu, J.Y., Singh, K.K.: Spatially-adaptive multilayer selection for gan inversion and editing. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (Jun 2022). https://doi.org/10.1109/cvpr52688.2022.01111, http://dx.doi.org/10.1109/CVPR52688.2022.01111", + "35. Patashnik, O., Garibi, D., Azuri, I., Averbuch-Elor, H., Cohen-Or, D.: Localizing object-level shape variations with text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)", + "36. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis (2023)", + "37. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents (2022)", + "38. Richardson, E., Alaluf, Y., Patashnik, O., Nitzan, Y., Azar, Y., Shapiro, S., Cohen-Or, D.: Encoding in style: a stylegan encoder for image-to-image translation. In:" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "D. Garibi et al.", + "bbox": [ + 271, + 114, + 374, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2021)", + "39. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2022)", + "40. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Ghasemipour, S.K.S., Ayan, B.K., Mahdavi, S.S., Lopes, R.G., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic text-to-image diffusion models with deep language understanding (2022)", + "41. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models. In: International Conference on Learning Representations (2021)", + "42. Sauer, A., Lorenz, D., Blattmann, A., Rombach, R.: Adversarial diffusion distillation (2023)", + "43. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models (2022)", + "44. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models (2023)", + "45. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: International Conference on Learning Representations (2020)", + "46. Tov, O., Alaluf, Y., Nitzan, Y., Patashnik, O., Cohen-Or, D.: Designing an encoder for stylegan image manipulation. ACM Trans. Graph. 40(4) (jul 2021). https://doi.org/10.1145/3450626.3459838", + "47. Tumanyan, N., Geyer, M., Bagon, S., Dekel, T.: Plug-and-play diffusion features for text-driven image-to-image translation pp. 1921-1930 (June 2023)", + "48. Voynov, A., Hertz, A., Arar, M., Fruchter, S., Cohen-Or, D.: Anylens: A generative diffusion model with any rendering lens (2023)", + "49. Wallace, B., Gokul, A., Naik, N.: Edict: Exact diffusion inversion via coupled transformations. arXiv preprint arXiv:2211.12446 (2022)", + "50. Wu, C.H., De la Torre, F.: Unifying diffusion models' latent space, with applications to cyclediffusion and guidance. arXiv preprint arXiv:2210.05559 (2022)", + "51. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)", + "52. Zhu, J., Shen, Y., Zhao, D., Zhou, B.: In-domain gan inversion for real image editing. In: Proceedings of European Conference on Computer Vision (ECCV) (2020)", + "53. Zhu, J.Y., Krahenbihl, P., Shechtman, E., Efros, A.A.: Generative visual manipulation on the natural image manifold. In: Proceedings of European Conference on Computer Vision (ECCV) (2016)", + "54. Zhu, P., Abdal, R., Qin, Y., Femiani, J., Wonka, P.: Improved stylegan embedding: Where are the good latents? (2020)" + ], + "bbox": [ + 215, + 146, + 784, + 672 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "ReNoise: Real Image Inversion Through Iterative Noising", + "bbox": [ + 349, + 114, + 732, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/ReNoise_ Real Image Inversion Through Iterative Noising/4a999b7d-d6fa-416b-8345-a221ff72abc9_model.json b/2024/ReNoise_ Real Image Inversion Through Iterative Noising/4a999b7d-d6fa-416b-8345-a221ff72abc9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..be33d0b4f5c6a2a22a0cc52addafd9d8129d7645 --- /dev/null +++ b/2024/ReNoise_ Real Image Inversion Through Iterative Noising/4a999b7d-d6fa-416b-8345-a221ff72abc9_model.json @@ -0,0 +1,2379 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.22, + 0.141, + 0.784, + 0.187 + ], + "angle": 0, + "content": "ReNoise: Real Image Inversion Through Iterative Noising" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.212, + 0.784, + 0.244 + ], + "angle": 0, + "content": "Daniel Garibi1, Or Patashnik1, Andrey Voynov2, Hadar Averbuch-Elor1, and Daniel Cohen-Or1" + }, + { + "type": "text", + "bbox": [ + 0.429, + 0.255, + 0.575, + 0.269 + ], + "angle": 0, + "content": "1 Tel Aviv University" + }, + { + "type": "text", + "bbox": [ + 0.438, + 0.27, + 0.565, + 0.283 + ], + "angle": 0, + "content": "2 Google Research" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.316, + 0.784, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.479, + 0.788, + 0.559 + ], + "angle": 0, + "content": "Fig. 1: Our ReNoise inversion technique can be applied to various diffusion models, including recent few-step ones. This figure illustrates the performance of our method with SDXL Turbo and LCM models, showing its effectiveness compared to DDIM inversion. Additionally, we demonstrate that the quality of our inversions allows prompt-driven editing. As illustrated on the right, our approach also allows for prompt-driven image edits." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.559, + 0.741, + 0.821 + ], + "angle": 0, + "content": "Abstract. Recent advancements in text-guided diffusion models have unlocked powerful image manipulation capabilities. However, applying these methods to real images necessitates the inversion of the images into the domain of the pretrained diffusion model. Achieving faithful inversion remains a challenge, particularly for more recent models trained to generate images with a small number of denoising steps. In this work, we introduce an inversion method with a high quality-to-operation ratio, enhancing reconstruction accuracy without increasing the number of operations. Building on reversing the diffusion sampling process, our method employs an iterative renoising mechanism at each inversion sampling step. This mechanism refines the approximation of a predicted point along the forward diffusion trajectory, by iteratively applying the pretrained diffusion model, and averaging these predictions. We evaluate the performance of our ReNoise technique using various sampling algorithms and models, including recent accelerated diffusion models. Through comprehensive evaluations and comparisons, we show its effectiveness in terms of both accuracy and speed. Furthermore, we confirm that our method preserves editability by demonstrating text-driven image editing on real images." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.826, + 0.647, + 0.84 + ], + "angle": 0, + "content": "Keywords: Diffusion Models \\(\\cdot\\) Image Editing \\(\\cdot\\) Inversion" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.128 + ], + "angle": 0, + "content": "D. Garibi et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.495, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.228, + 0.505, + 0.366 + ], + "angle": 0, + "content": "Fig.2: The diffusion process samples a Gaussian noise and iteratively denoises it until reaching the data distribution. At each point along the denoising trajectory, the model predicts a direction, \\(\\epsilon_{\\theta}(z_t)\\), to step to the next point along the trajectory. To invert a given image from the distribution, the direction from \\(z_{t}\\) to \\(z_{t + 1}\\) is approximated with the inverse of the direction from \\(z_{t}\\) to \\(z_{t - 1}\\) denoted by a dotted blue line." + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.155, + 0.795, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.325, + 0.788, + 0.367 + ], + "angle": 0, + "content": "Fig.3: Comparing reconstruction results of plain DDIM Inv. with SDXL to DDIM Inv. with one renoising iteration." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.376, + 0.377, + 0.391 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.408, + 0.788, + 0.543 + ], + "angle": 0, + "content": "Large-scale text-to-image diffusion models have revolutionized the field of image synthesis [19,37,39,40]. In particular, many works have shown that these models can be employed for various types of image manipulation [4,7-9,15,17,23,29,33, 35,47]. To edit real images, many of these techniques often require the inversion of the image into the domain of the diffusion model. That is, given a real image \\( z_{0} \\), one has to find a Gaussian noise \\( z_{T} \\), such that denoising \\( z_{T} \\) with the pretrained diffusion model reconstructs the given real image \\( z_{0} \\). The importance of this task for real image manipulation has prompted many efforts aimed at achieving accurate reconstruction [16,21,30,31]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.544, + 0.788, + 0.665 + ], + "angle": 0, + "content": "The diffusion process consists of a series of denoising steps \\(\\{\\epsilon_{\\theta}(z_t,t)\\}_{t = T}^1\\), which form a trajectory from the Gaussian noise to the model distribution (see Figure 2). Each denoising step is computed by a trained network, typically implemented as a UNet, which predicts \\(z_{t - 1}\\) from \\(z_{t}\\) [19]. The output of the model at each step forms a direction from \\(z_{t}\\) to \\(z_{t - 1}\\) [45]. These steps are not invertible, in the sense that the model was not trained to predict \\(z_{t}\\) from \\(z_{t - 1}\\). Thus, the problem of inverting a given image is a challenge, and particularly for real images, as they are not necessarily in the model distribution (see Figure 3)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.665, + 0.789, + 0.844 + ], + "angle": 0, + "content": "In this paper, we present an inversion method with a high quality-to-operation ratio, which achieves superior reconstruction accuracy for the same number of UNet operations. We build upon the commonly used approach of reversing the diffusion sampling process, which is based on the linearity assumption that the direction from \\( z_{t} \\) to \\( z_{t+1} \\) can be approximated by the negation of the direction from \\( z_{t} \\) to \\( z_{t-1} \\) [12, 43] (see Figure 2). To enhance this approximation, we employ the fixed-point iteration methodology [10]. Specifically, given \\( z_{t} \\), we begin by using the common approximation to get an initial estimate for \\( z_{t+1} \\), denoted by \\( z_{t+1}^{(0)} \\). Then, we iteratively renoise \\( z_{t} \\), following the direction implied by \\( z_{t+1}^{(k)} \\) to obtain \\( z_{t+1}^{(k+1)} \\). After repeating this ronoising process several times, we apply an averaging on \\( z_{t+1}^{(k)} \\) to form a more accurate direction from \\( z_{t} \\) to \\( z_{t+1} \\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.35, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "ReNoise: Real Image Inversion Through Iterative Noising" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "We show that this approach enables longer strides along the inversion trajectory while improving image reconstruction. Therefore, our method can also be effective with diffusion models trained to generate images using a small number of denoising steps [27, 42]. Furthermore, despite the need to repeatedly renoise in each step of the inversion process, the longer strides lead to a more favorable tradeoff of UNet operations for reconstruction quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.8, + 0.329 + ], + "angle": 0, + "content": "Through extensive experiments, we demonstrate the effectiveness of our method in both image reconstruction and inversion speed. We validate the versatility of our approach across different samplers and models, including recent time-distilled diffusion models (e.g., SDXL-Turbo [42]). Importantly, we demonstrate that the editability of the inversion achieved by our method allows a wide range of text-driven image manipulations (see Figure 1)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.333, + 0.388, + 0.349 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.355, + 0.788, + 0.567 + ], + "angle": 0, + "content": "Image Editing via Diffusion Models Recent advancements in diffusion models [12, 19] have resulted in unprecedented diversity and fidelity in visual content creation guided by free-form text prompts [36, 37, 39, 40]. Text-to-image models do not directly support text-guided image editing. Therefore, harnessing the power of these models for image editing is a significant research area and many methods have utilized these models for different types of image editing [4, 7-9, 11, 14, 15, 17, 18, 23, 29, 33, 35, 47, 48]. A common approach among these methods requires inversion [21, 31, 43, 49] to edit real images, i.e., obtaining a latent code \\( z_{T} \\) such that denoising it with the pretrained diffusion model returns the original image. Specifically, in this approach two backward processes are done simultaneously using \\( z_{T} \\). One of the processes reconstructs the image using the original prompt, while the second one injects features from the first process (e.g., attention maps) to preserve some properties of the original image while manipulating other aspects of it." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.689 + ], + "angle": 0, + "content": "Inversion in Diffusion Models Initial efforts in image inversion for real image editing focused on GANs [2,3,5,6,13,34,38,46,52-54]. The advancements in diffusion models, and in diffusion-based image editing in particular have recently prompted works studying the inversion of a diffusion-based denoising process. This inversion depends on the sampler algorithm used during inference, which can be deterministic [43] or non-deterministic [19, 22]. Inversion methods can be accordingly categorized into two: methods that are suitable for deterministic sampling, and methods suitable for non-deterministic sampling." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Methods that approach the deterministic inversion commonly rely on the DDIM sampling method [43], and build upon DDIM inversion [12, 43]. Mokady et al. [31] observed that the use of classifier-free guidance during inference magnifies the accumulated error of DDIM inversion and therefore leads to poor reconstruction. Following this observation, several works [16, 30, 31] focused on solving this issue by replacing the null text token with a different embedding, which is found after an optimization process or by a closed solution. However, excluding [31] which requires a lengthy optimization, these methods are limited by the reconstruction accuracy of DDIM inversion, which can be poor, especially when a small number of denoising steps is done. In our work, we present a method" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.128 + ], + "angle": 0, + "content": "D. Garibi et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "that improves the reconstruction quality of DDIM inversion and therefore can be integrated with methods that build on it." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Another line of work [21, 50] tackles the inversion of DDPM sampler [19]. In these works [21, 50], instead of inverting the image into an initial noise \\( z_{T} \\), a series of noises \\( \\{z_{T},\\epsilon_{T},\\dots,\\epsilon_{1}\\} \\) is obtained. The definition of this noises series ensures that generating an image with it returns the original input image. However, these methods require a large number of inversion and denoising steps to allow image editing. Applying these methods with an insufficient number of steps leads to too much information encoded in \\( \\{\\epsilon_T,\\dots,\\epsilon_1\\} \\) which limits the ability to edit the generated image. As shall be shown, The editability issue of these methods is particularly evident in few-steps models [26, 27, 42]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.315, + 0.788, + 0.435 + ], + "angle": 0, + "content": "Most relevant to our work, two recent inversion methods [28,32] also use the fixed-point iteration technique. Specifically, they improve the reconstruction accuracy of DDIM inversion [43] with Stable Diffusion [39] without introducing a significant computational overhead. In our work, we focus on the problem of real image inversion for recently introduced few-step diffusion models, where the difficulties encountered by previous methods are amplified. Furthermore, we show that our inversion method successfully works with various models and different samplers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.788, + 0.581 + ], + "angle": 0, + "content": "Few Steps Models Recently, new methods [26, 27, 41, 42, 44] that fine-tune text-to-image diffusion models enabled a significant reduction of the number of steps needed for high-quality image generation. While standard diffusion models typically require 50 denoising steps to generate high-quality images, recent accelerated models achieve high-quality synthesis with 1-4 steps only. These new methods pave the way for interactive editing workflows. However, as we show in this paper, using current methods for the inversion of an image with a small number of steps degrades the reconstruction quality in terms of accuracy [12, 43] or editability [21, 50]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.589, + 0.332, + 0.605 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.625, + 0.414, + 0.639 + ], + "angle": 0, + "content": "3.1 ReNoise Inversion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.652, + 0.788, + 0.789 + ], + "angle": 0, + "content": "Reversing the Sampler Samplers play a critical role in the diffusion-based image synthesis process. They define the noising and denoising diffusion processes and influence the processes' trajectories and quality of the generated images. While different samplers share the same pre-trained UNet model (denoted by \\(\\epsilon_{\\theta}\\)) as their backbone, their sampling approaches diverge, leading to nuanced differences in output. The goal of the denoising sampler is to predict the latent code at the previous noise level, \\(z_{t-1}\\), based on the current noisy data \\(z_{t}\\), the pretrained UNet model, and a sampled noise, \\(\\epsilon_{t}\\). Various first-order denoising sampling algorithms adhere to the form:" + }, + { + "type": "equation", + "bbox": [ + 0.385, + 0.791, + 0.786, + 0.806 + ], + "angle": 0, + "content": "\\[\nz _ {t - 1} = \\phi_ {t} z _ {t} + \\psi_ {t} \\epsilon_ {\\theta} (z _ {t}, t, c) + \\rho_ {t} \\epsilon_ {t}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "where \\(c\\) represents a text embedding condition, and \\(\\phi_t, \\psi_t\\), and \\(\\rho_t\\) denote sampler parameters. At each step, these parameters control the extent to which the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.35, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "ReNoise: Real Image Inversion Through Iterative Noising" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "previous noise is removed \\((\\phi_t)\\), the significance assigned to the predicted noise from the UNet \\((\\psi_t)\\), and the given weight to the additional noise introduced \\((\\rho_t)\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.206 + ], + "angle": 0, + "content": "A given image \\( z_0 \\) can be inverted by reformulating Equation 1 and applying it iteratively:" + }, + { + "type": "equation", + "bbox": [ + 0.391, + 0.2, + 0.785, + 0.23 + ], + "angle": 0, + "content": "\\[\nz _ {t} = \\frac {z _ {t - 1} - \\psi_ {t} \\epsilon_ {\\theta} \\left(z _ {t} , t , c\\right) - \\rho_ {t} \\epsilon_ {t}}{\\phi_ {t}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.232, + 0.784, + 0.322 + ], + "angle": 0, + "content": "where for non-deterministic samplers, a series of random noises \\(\\{\\epsilon_t\\}_{t=1}^T\\) is sampled and used during both inversion and image generation processes. However, directly computing \\(z_t\\) from Equation 2 is infeasible since it relies on \\(\\epsilon_\\theta(z_t, t, c)\\) which, in turn, depends on \\(z_t\\), creating a circular dependency. To solve this implicit function, Dhariwal et al. [12] propose using the approximation \\(\\epsilon_\\theta(z_t, t, c) \\approx \\epsilon_\\theta(z_{t-1}, t, c)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.376, + 0.316, + 0.785, + 0.346 + ], + "angle": 0, + "content": "\\[\nz _ {t} ^ {(1)} = \\frac {z _ {t - 1} - \\psi_ {t} \\epsilon_ {\\theta} \\left(z _ {t - 1} , t , c\\right) - \\rho_ {t} \\epsilon_ {t}}{\\phi_ {t}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.348, + 0.785, + 0.485 + ], + "angle": 0, + "content": "This method has several limitations. First, the assumption underlying the approximation used in [12] is that the number of inversion steps is large enough, implying a trajectory close to linear. This assumption restricts the applicability of this inversion method in interactive image editing with recent few-step diffusion models [26, 27, 42, 44], as the inversion process would take significantly longer than inference. Second, this method struggles to produce accurate reconstructions in certain cases, such as highly detailed images or images with large smooth regions, see Figure 3. Moreover, we observe that this inversion method is sensitive to the prompt \\( c \\) and may yield poor results for certain prompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.784, + 0.583 + ], + "angle": 0, + "content": "ReNoise In a successful inversion trajectory, the direction from \\( z_{t-1} \\) to \\( z_t \\) aligns with the direction from \\( z_t \\) to \\( z_{t-1} \\) in the denoising trajectory. To achieve this, we aim to improve the approximation of \\( \\epsilon_{\\theta}(z_t,t,c) \\) in Eq. 2 compared to the one used in [12]. Building on the fixed-point iteration technique [10], our approach better estimates the instance of \\( z_t \\) that is inputted to the UNet, rather than relying on \\( z_{t-1} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.785, + 0.742 + ], + "angle": 0, + "content": "Intuitively, we utilize the observation that \\( z_{t}^{(1)} \\) (from Eq. 3) offers a more precise estimate of \\( z_{t} \\) compared to \\( z_{t-1} \\). Therefore, employing \\( z_{t}^{(1)} \\) as the input to the UNet is likely to yield a more accurate direction, thus contributing to reducing the overall error in the inversion step. We illustrate this observation in Figure 5. Iterating this process generates a series of estimations for \\( z_{t} \\), denoted by \\( \\{z_{t}^{(k)}\\}_{k=1}^{\\mathcal{K}+1} \\). While the fixed-point iteration technique [10] does not guarantee convergence of this series in the general case, in Section 4, we empirically show that convergence holds in our setting. However, as the convergence is not monotonic, we refine our prediction of \\( z_{t} \\) by averaging several \\( \\{z_{t}^{(k)}\\} \\), thus considering more than a single estimation of \\( z_{t} \\). See Figure 6 for an intuitive illustration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.743, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In more detail, our method iteratively computes estimations of \\( z_{t} \\) during each inversion step \\( t \\) by renoising the noisy latent \\( z_{t-1} \\) multiple times, each with a different noise prediction (see Figure 4). Beginning with \\( z_{t}^{(1)} \\), in the \\( k \\)-th renoising iteration, the input to the UNet is the result of the previous iteration, \\( z_{t}^{(k)} \\). Then, \\( z_{t}^{(k+1)} \\) is calculated using the inverted sampler while maintaining \\( z_{t-1} \\) as the starting point of the step. After \\( \\mathcal{K} \\) renoising iterations, we obtain a set of" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.128 + ], + "angle": 0, + "content": "D. Garibi et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.144, + 0.787, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.307, + 0.789, + 0.422 + ], + "angle": 0, + "content": "Fig. 4: Method overview. Given an image \\( z_0 \\), we iteratively compute \\( z_1, \\dots, z_T \\), where each \\( z_t \\) is calculated from \\( z_{t-1} \\). At each time step, we apply the UNet \\( (\\epsilon_\\theta) \\mathcal{K} + 1 \\) times, each using a better approximation of \\( z_t \\) as the input. The initial approximation is \\( z_{t-1} \\). The next one, \\( z_t^{(1)} \\), is the result of the reversed sampler step (i.e., DDIM). The reversed step begins at \\( z_{t-1} \\) and follows the direction of \\( \\epsilon_\\theta(z_{t-1}, t) \\). At the \\( k \\) renoising iteration, \\( z_t^{(k)} \\) is the input to the UNet, and we obtain a better \\( z_t \\) approximation. For the last iterations, we optimize \\( \\epsilon_\\theta(z_t^{(k)}, t) \\) to increase editability. As the final denoising direction, we use the average of the UNet predictions of the last few iterations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.424, + 0.533, + 0.618 + ], + "angle": 0, + "content": "Fig.5: Geometric intuition for ReNoise. At each inversion step, we estimate \\( z_{t} \\) (marked with a red star) based on \\( z_{t-1} \\). The straightforward approach is to use the negated direction of the denoising step from \\( z_{t-1} \\), assuming the trajectory is approximately linear. However, this assumption is inaccurate, especially in few-step models, where the size of the steps is large. We use the linearity assumption as an initial estimation and keep improving this estimation. We recalculate the denoising step from the previous estimation (which is closer to the target \\( z_{t} \\)) and then proceed with its negated direction from \\( z_{t-1} \\) (see the orange vectors)." + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.447, + 0.791, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.623, + 0.788, + 0.675 + ], + "angle": 0, + "content": "estimations \\(\\{z_t^{(k)}\\}_{k = 1}^{\\mathcal{K} + 1}\\). The next point on the inversion trajectory, \\(z_{t}\\), is then defined as their weighted average, where \\(w_{k}\\) is the weight assigned to \\(z_{t}^{(k)}\\). For a detailed description of our method, refer to Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.68, + 0.557, + 0.695 + ], + "angle": 0, + "content": "3.2 Reconstruction-Editability Tradeoff" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Enhance Editability The goal of inversion is to facilitate the editing of real images using a pretrained image generation model. While the renoising approach attains highly accurate reconstruction results, we observe that the resulting \\( z_{T} \\) lacks editability. This phenomenon can be attributed to the reconstruction-editability tradeoff in image generative models [46]. To address this limitation, we incorporate a technique to enhance the editability of our method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "It has been shown [33] that the noise maps predicted during the inversion process often diverge from the statistical properties of uncorrelated Gaussian white noise, thereby affecting editability. To tackle this challenge, we follow pix2pix-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.35, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "ReNoise: Real Image Inversion Through Iterative Noising" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "zero [33] and regularize the predicted noise at each step, \\(\\epsilon_{\\theta}(z_t,t,c)\\), using the following loss terms." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.333 + ], + "angle": 0, + "content": "First, we encourage \\(\\epsilon_{\\theta}(z_t,t,c)\\) to follow the same distribution as \\(\\epsilon_{\\theta}(z_t^{\\prime},t,c)\\), where \\(z_{t}^{\\prime}\\) represents the input image \\(z_{0}\\) with added random noise corresponding to the noise level at timestep \\(t\\). We do so by dividing \\(\\epsilon_{\\theta}(z_t,t,c)\\) and \\(\\epsilon_{\\theta}(z_t^{\\prime},t,c)\\) into small patches (e.g., \\(4\\times 4\\)), and computing the KL-divergence between corresponding patches. We denote this loss term by \\(\\mathcal{L}_{\\mathrm{patch - KL}}\\). Second, we utilize \\(\\mathcal{L}_{\\mathrm{pair}}\\) proposed in pix2pix-zero [33], which penalizes correlations between pairs of pixels. We leverage these losses to enhance the editability of our method, and denote the combination of them as \\(\\mathcal{L}_{\\mathrm{edit}}\\). For any renoising iteration \\(k\\) where \\(w_{k} > 0\\), we regularize the UNet's prediction \\(\\epsilon_{\\theta}(z_t^{(k)},t,c)\\) using \\(\\mathcal{L}_{\\mathrm{edit}}\\) before computing \\(z_{t}^{(k + 1)}\\). See line 9 in Algorithm 1." + }, + { + "type": "code_caption", + "bbox": [ + 0.23, + 0.345, + 0.471, + 0.358 + ], + "angle": 0, + "content": "Algorithm 1: ReNoise Inversion" + }, + { + "type": "algorithm", + "bbox": [ + 0.216, + 0.361, + 0.799, + 0.625 + ], + "angle": 0, + "content": "1 Input: An image \\(z_{0}\\), number of renoising steps \\(\\mathcal{K}\\), number of inversion steps \\(T\\), a series of renoising weights \\(\\{w_{k}\\}_{k=1}^{\\mathcal{K}+1}\\). \n2 Output: A noisy latent \\(z_{T}\\) and set of noises \\(\\{\\epsilon_{t}\\}_{t=1}^{T}\\). \n3 for \\(t = 0,1,\\ldots,T\\) do \n4 sample \\(\\epsilon_{t} \\sim \\mathcal{N}(0,I)\\) \n5 \\(z_{t}^{(0)} \\gets z_{t-1}\\) \n6 \\(z_{t}^{(\\mathrm{avg})} \\gets 0\\) \n7 for \\(k = 0,\\ldots,\\mathcal{K}\\) do \n8 \\(\\delta_{t}^{k} \\gets \\epsilon_{\\theta}(z_{t}^{(k)},t)\\) \n9 \\(\\delta_{t}^{k} \\gets \\text{Enhance-edit}(\\delta_{t}^{k},w_{k+1})\\) \n10 \\(z_{t}^{(k+1)} \\gets \\text{Inverse-Step}(z_{t-1},\\delta_{t}^{k})\\) \n11 end \n// Average ReNoised predictions \n23 Function Noise-Corr \\((z_{t},t,\\epsilon_{t},z_{t-1})\\): \n12 \\(z_{t}^{(\\mathrm{avg})} \\gets \\sum_{k=1}^{\\mathcal{K}+1} w_{k} \\cdot z_{t}^{(k)}\\) \n24 \\(\\delta_{t} \\gets \\epsilon_{\\theta}(z_{t},t)\\) \n25 \\(\\epsilon_{t} \\gets \\epsilon_{t} - \\nabla_{\\epsilon_{t}}\\frac{1}{\\rho_{t}}(z_{t-1} - \\phi_{t}z_{t} - \\psi_{t}\\delta_{t})\\) \n26 return \\(\\epsilon_{t}\\) \nend \nreturn \\((z_{T},\\{\\epsilon_{t}\\}_{t=1}^{T})\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.72 + ], + "angle": 0, + "content": "Noise Correction in Non-deterministic Samplers Non-deterministic samplers, in which \\(\\rho_t > 0\\), introduce noise \\((\\epsilon_t)\\) at each denoising step. Previous methods [21,50] suggested using \\(\\epsilon_t\\) to bridge the gap between the inversion and denoising trajectories in DDPM inversion. Specifically, given a pair of points \\(z_{t - 1},z_{t}\\) on the inversion trajectory, we denote by \\(\\hat{z}_{t - 1}\\) the point obtained by denoising \\(z_{t}\\). Ideally, \\(z_{t - 1}\\) and \\(\\hat{z}_{t - 1}\\) should be identical. We define:" + }, + { + "type": "equation", + "bbox": [ + 0.372, + 0.719, + 0.786, + 0.749 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {t} = \\frac {1}{\\rho_ {t}} \\left(z _ {t - 1} - \\phi_ {t} z _ {t} - \\psi_ {t} \\epsilon_ {\\theta} \\left(z _ {t}, t, c\\right)\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.784, + 0.794 + ], + "angle": 0, + "content": "Integrating this definition into Eq. 1 yields \\(\\hat{z}_{t - 1} = z_{t - 1}\\). However, we found that replacing \\(\\epsilon_t\\) with the above definition affects editability. Instead, we suggest a more tender approach, optimizing \\(\\epsilon_t\\) based on Eq. 4 as our guiding objective:" + }, + { + "type": "equation", + "bbox": [ + 0.343, + 0.794, + 0.786, + 0.822 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {t} = \\epsilon_ {t} - \\nabla_ {\\epsilon_ {t}} \\frac {1}{\\rho_ {t}} \\left(z _ {t - 1} - \\phi_ {t} z _ {t} - \\psi_ {t} \\epsilon_ {\\theta} \\left(z _ {t}, t, c\\right)\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.82, + 0.784, + 0.851 + ], + "angle": 0, + "content": "This optimization improves the reconstruction fidelity while preserving the distribution of the noisy-latents." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "D. Garibi et al." + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.145, + 0.51, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.22, + 0.545, + 0.317 + ], + "angle": 0, + "content": "Fig. 6: Schematic illustration of the ReNoise convergence process to the true inversion of \\( z_{t-1} \\). While estimates may converge non-monotonically to the unknown target \\( z_t \\), we found that averaging them improves true value estimation. Typically, the initial iteration exhibits an exponential decrease in the norm between consecutive elements." + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.16, + 0.785, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.233, + 0.788, + 0.317 + ], + "angle": 0, + "content": "Fig.7: Average distance between consequent estimations \\( z_{t}^{(k)} \\), and \\( z_{t}^{(k + 1)} \\). Vertical bars indicate the standard deviation. The averages are computed over 32 images and 10 different timesteps." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.323, + 0.484, + 0.34 + ], + "angle": 0, + "content": "4 Convergence Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.346, + 0.784, + 0.435 + ], + "angle": 0, + "content": "In this section, we first express the inversion process as a backward Euler process and our renoising iterations as fixed-point iterations. While these iterations do not converge in the general case, in the supplementary materials we present a toy example where they yield accurate inversions. Then, we analyze the convergence of the renoising iterations in our real-image inversion scenario and empirically verify our method's convergence." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.439, + 0.788, + 0.515 + ], + "angle": 0, + "content": "Inversion Process as Backward Euler The denoising process of diffusion models can be mathematically described as solving an ordinary differential equation (ODE). A common method for solving such equations is the Euler method, which takes small steps to approximate the solution. For ODE in the form of \\( y'(t) = f(t,y(t)) \\), Euler solution is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.41, + 0.516, + 0.595, + 0.532 + ], + "angle": 0, + "content": "\\[\ny _ {n + 1} = y _ {n} + h \\cdot f (t _ {n}, y _ {n}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.784, + 0.595 + ], + "angle": 0, + "content": "where \\( h \\) is the step size. The inversion process can be described as solving ODE using the backward Euler method (or implicit Euler method) [1]. This method is similar to forward Euler, with the difference that \\( y_{n + 1} \\) appears on both sides of the equation:" + }, + { + "type": "equation", + "bbox": [ + 0.392, + 0.595, + 0.61, + 0.61 + ], + "angle": 0, + "content": "\\[\ny _ {n + 1} = y _ {n} + h \\cdot f \\left(t _ {n + 1}, y _ {n + 1}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.784, + 0.688 + ], + "angle": 0, + "content": "For equations lacking an algebraic solution, several techniques estimate \\( y_{n+1} \\) iteratively. As we described in Section 3.1, the inversion process lacks a closed-form solution, as shown in Equation 2. To address this, the ReNoise method leverages fixed-point iterations, which we refer to as reonising iterations, to progressively refine the estimate of \\( y_{n+1} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.689, + 0.662, + 0.711 + ], + "angle": 0, + "content": "\\[\ny _ {n + 1} ^ {(0)} = y _ {n}, \\quad y _ {n + 1} ^ {(k + 1)} = y _ {n} + h \\cdot f (t _ {n + 1}, y _ {n + 1} ^ {(k)}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.716, + 0.784, + 0.747 + ], + "angle": 0, + "content": "In our ReNoise method, we average these renoising iterations to mitigate convergence errors, leading to improvement in the reconstruction quality." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.75, + 0.788, + 0.844 + ], + "angle": 0, + "content": "ReNoise Convergence During the inversion process, we aim to find the next noise level inversion, denoted by \\(\\hat{z}_t\\), such that applying the denoising step to \\(\\hat{z}_t\\) recovers the previous state, \\(z_{t - 1}\\). Given the noise estimation \\(\\epsilon_{\\theta}(z_t,t)\\) and a given \\(z_{t - 1}\\), the ReNoise mapping defined in Section 3.1 can be written as \\(\\mathcal{G}:z_t\\to \\mathrm{InverseStep}(z_{t - 1},\\epsilon_\\theta (z_t,t))\\). For example, in the case of using DDIM sampler the mapping is \\(\\mathcal{G}(z_t) = \\frac{1}{\\phi_t} (z_{t - 1} - \\psi_t\\epsilon_\\theta (z_t,t))\\). The point \\(\\hat{z}_t\\), which is" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.35, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "ReNoise: Real Image Inversion Through Iterative Noising" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.226 + ], + "angle": 0, + "content": "mapped to \\( z_{t-1} \\) after the denoising step, is a stationary point of this mapping. Given \\( z_t^{(1)} \\), the first approximation of the next noise level \\( z_t \\), our goal is to show that the sequence \\( z_t^{(k)} = \\mathcal{G}^{k-1}(z_t^{(1)}) \\), \\( k \\to \\infty \\) converges. As the mapping \\( \\mathcal{G} \\) is continuous, the limit point would be its stationary point. The definition of \\( \\mathcal{G} \\) gives:" + }, + { + "type": "equation", + "bbox": [ + 0.359, + 0.219, + 0.645, + 0.239 + ], + "angle": 0, + "content": "\\[\n\\left\\| z _ {t} ^ {(k + 1)} - z _ {t} ^ {(k)} \\right\\| = \\left\\| \\mathcal {G} \\left(z _ {t} ^ {(k)}\\right) - \\mathcal {G} \\left(z _ {t} ^ {(k - 1)}\\right) \\right\\|,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.24, + 0.785, + 0.333 + ], + "angle": 0, + "content": "where the norm is always assumed as the \\(l_{2}\\)-norm. For the ease of the notations, we define \\(\\varDelta^{(k)}=z_t^{(k)}-z_t^{(k-1)}\\). For convergence proof, it is sufficient to show that the sum of norms of these differences converges, which will imply that \\(z_{t}^{(k)}\\) is the Cauchy sequence. Below we check that in practice \\(\\| \\varDelta^{(k)}\\|\\) decreases exponentially as \\(k\\to \\infty\\) and thus has finite sum. In the assumption that \\(\\mathcal{G}\\) is \\(\\mathcal{C}^2\\)-smooth, the Taylor series conducts:" + }, + { + "type": "equation", + "bbox": [ + 0.236, + 0.337, + 0.769, + 0.45 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\| \\varDelta^ {(k + 1)} \\| = \\| \\mathcal {G} (z _ {t} ^ {(k)}) - \\mathcal {G} (z _ {t} ^ {(k - 1)}) \\| = \\\\ \\left\\| \\mathcal {G} \\left(z _ {t} ^ {(k - 1)}\\right) + \\frac {\\partial \\mathcal {G}}{\\partial z} \\right| _ {z _ {t} ^ {(k - 1)}} \\cdot \\Delta^ {(k)} + O \\left(\\left\\| \\Delta^ {(k)} \\right\\| ^ {2}\\right) - \\mathcal {G} \\left(z _ {t} ^ {(k - 1)}\\right) \\| = \\\\ \\left. \\left\\| \\frac {\\partial \\mathcal {G}}{\\partial z} \\right| _ {z _ {t} ^ {(k - 1)}} \\cdot \\Delta^ {(k)} + O (\\left\\| \\Delta^ {(k)} \\right\\| ^ {2}) \\right\\| \\leq \\left\\| \\frac {\\partial \\mathcal {G}}{\\partial z} \\right| _ {z _ {t} ^ {(k - 1)}} \\| \\cdot \\| \\Delta^ {(k)} \\| + O (\\left\\| \\Delta^ {(k)} \\right\\| ^ {2}) = \\\\ \\frac {\\psi_ {t}}{\\phi_ {t}} \\cdot \\left\\| \\frac {\\partial \\epsilon_ {\\theta}}{\\partial z} \\right| _ {z _ {t} ^ {(k - 1)}} \\| \\cdot \\| \\Delta^ {(k)} \\| + O (\\| \\Delta^ {(k)} \\| ^ {2}) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.785, + 0.618 + ], + "angle": 0, + "content": "Thus, in a sufficiently small neighborhood, the convergence dynamics is defined by the scaled Jacobian norm \\(\\frac{\\psi_t}{\\phi_t} \\cdot \\left\\| \\frac{\\partial \\epsilon_\\theta}{\\partial z} \\right|_{z_t^{(k-1)}}\\). In the supplementary materials, we show this scaled norm estimation for the SDXL diffusion model for various steps and ReNoise iterations indices \\((k)\\). Remarkably, the ReNoise indices minimally impact the scale factor, consistently remaining below 1. This confirms in practice the convergence of the proposed algorithm. Notably, the highest scaled norm values occur at smaller \\(t\\) (excluding the first step) and during the initial renoising iteration. This validates the strategy of not applying ReNoise in early steps, where convergence tends to be slower compared to other noise levels. Additionally, the scaled norm value for the initial \\(t\\) approaches 0, which induces almost immediate convergence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.62, + 0.784, + 0.697 + ], + "angle": 0, + "content": "Figure 7 illustrates the exponential decrease in distances between consecutive elements \\( z_{t}^{(k)} \\) and \\( z_{t}^{(k + 1)} \\), which confirms the algorithm's convergence towards the stationary point of the operator \\( \\mathcal{G} \\). The proposed averaging strategy is aligned with the conclusions described above, and also converges to the desired stationary point. In The supplementary materials, we present a validation for this claim." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.7, + 0.376, + 0.716 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.784, + 0.841 + ], + "angle": 0, + "content": "In this section, we conduct extensive experiments to validate the effectiveness of our method. We evaluate both the reconstruction quality of our inversion and its editability. To demonstrate the versatility of our approach, we apply it to four models, SD [39], SDXL [36], SDXL Turbo [42], and LCM-LoRA [27], with SDXL Turbo and LCM-LoRA being few-step models. Additionally, we use various sampling algorithms including both deterministic and non-deterministic ones. Implementation details for each model are provided in the supplementary materials. Following previous works [5, 31], we quantitatively evaluate our" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.128 + ], + "angle": 0, + "content": "D. Garibi et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.149, + 0.495, + 0.287 + ], + "angle": 0, + "content": "Table 1: Image reconstruction results with a fixed number of 100 UNet operations. Each row showcases the results obtained using different combinations of inversion steps, denoising steps, and renoising iterations, totaling 100 operations. As observed, allocating some of the operations to renoising iterations improves the reconstruction quality while maintaining the same execution time." + }, + { + "type": "table", + "bbox": [ + 0.215, + 0.29, + 0.498, + 0.414 + ], + "angle": 0, + "content": "
Image Reconstruction Results
Inv StepsInf StepsReNoise StepsL2 ↓PSNR ↑LPIPS ↓
505000.0036426.0230.06273
752500.0038225.4660.06605
802000.0040825.0450.07099
901000.0102320.2490.10305
252520.0018229.5690.03637
202030.0016729.8840.03633
101080.0023028.1560.04678
" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.145, + 0.789, + 0.298 + ], + "angle": 0, + "content": "Table 2: Quantitative ablation study on SDXL Turbo. We demonstrate the impact of each component of our inversion method on reconstruction results. The results improve with additional renoising iterations and significant enhancements occur through averaging final estimations. Additionally, we observe a reconstruction-editability trade-off, with edit losses causing degradation that is effectively mitigated by Noise Correction." + }, + { + "type": "table", + "bbox": [ + 0.51, + 0.3, + 0.791, + 0.413 + ], + "angle": 0, + "content": "
Ablation - Image Reconstruction
L2 ↓PSNR ↑LPIPS↓
Euler Inversion0.070011.7840.20337
+ 1 ReNoise0.055212.7960.20254
+ 4 ReNoise0.024916.5210.14821
+ 9 ReNoise0.012619.7020.10850
+ Averaging ReNoise0.008721.4910.08832
+ Edit Losses0.027618.4320.12616
+ Noise Correction0.019622.0770.08469
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.789, + 0.448 + ], + "angle": 0, + "content": "method with three metrics: \\( L_{2} \\), LPIPS [51], and PSNR. Unless stated otherwise, for both inversion and generation we use the prompt obtained from BLIP2 [24]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.452, + 0.48, + 0.467 + ], + "angle": 0, + "content": "5.1 Reconstruction and Speed" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.471, + 0.789, + 0.592 + ], + "angle": 0, + "content": "We begin by evaluating the reconstruction-speed tradeoff. The main computational cost of both the inversion and denoising processes is the forward pass through the UNet. In each renoising iteration, we perform one forward pass, which makes it computationally equal to a standard inversion step (as done in DDIM Inversion for example). In the following experiments, we compare the results of a sampler reversing with our method, where we match the number of UNet passes between the methods. For example, 8 steps of sampler reversing are compared against 4 steps with one renoising iteration at each step." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.789, + 0.732 + ], + "angle": 0, + "content": "Qualitative Results In Figure 8 we show qualitative results of image reconstruction on SDXL Turbo [42]. Here, we utilize DDIM as the sampler, and apply four denoising steps for all configurations. Each row exhibits results obtained using a different amount of UNet operations. In our method, we apply four inversion steps, and a varying number of renoising iterations. As can be seen, the addition of renoising iterations gradually improves the reconstruction results. Conversely, employing more inversion steps proves insufficient for capturing all details in the image, as evident by the background of the car, or even detrimental to the reconstruction, as observed in the Uluro example." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.789, + 0.78 + ], + "angle": 0, + "content": "Quantitative Results For the quantitative evaluation, we use the MS-COCO 2017 [25] validation dataset. Specifically, we retain images with a resolution greater than \\(420 \\times 420\\), resulting in a dataset containing 3,865 images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We begin by evaluating both the sampler reversing approach and our ReNoise method, while varying the number of UNet operations during the inversion process and keeping the number of denoising steps fixed. This experiment is conducted using various models (SDXL, SDXL Turbo, LCM) and samplers. For all" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.35, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "ReNoise: Real Image Inversion Through Iterative Noising" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.489, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.414, + 0.494, + 0.541 + ], + "angle": 0, + "content": "Fig.8: Qualitative comparison between DDIM Inversion and our ReNoise method on SDXL Turbo (4 denoising steps). The first row displays the original images. The following rows display reconstructions from both methods, using the same number of UNet operations. DDIM employs smaller strides, while ReNoise utilizes more renoising steps." + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.143, + 0.782, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.429, + 0.788, + 0.54 + ], + "angle": 0, + "content": "Fig. 9: Ablation study on SDXL Turbo. The first row presents the input image. In each subsequent row, we show the reconstruction results using an additional component of our inversion method. The images in the bottom row represent the results obtained by our full method. NC stands for Noise Correction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.788, + 0.655 + ], + "angle": 0, + "content": "models, we utilize the DDIM [43] sampler. In addition, we employ the Ancestral-Euler scheduler for SDXL Turbo, and the default LCM sampler for LCM-LoRA. We set the number of denoising steps to 50 for SDXL, and to 4 for SDXL Turbo and LCM-LoRA. Quantitative results, using PSNR as the metric, are presented in Figure 11. We evaluate our method using different configurations. The x-axis refers to the number of UNet operations in the inversion process. Other metrics results are provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.841 + ], + "angle": 0, + "content": "As depicted in the graphs, incorporating additional renoising iterations proves to be more beneficial for image reconstruction compared to adding more inversion steps. Note that the performance of the Ancestral-Euler and LCM samplers noticeably degrades when the number of inversion steps exceeds the number of denoising steps. Unlike DDIM, these samplers have \\(\\varPhi_t\\approx 1\\), resulting in an increase in the latent vector's norm beyond what can be effectively denoised in fewer steps. In this experiment, we maintain the same number of UNet operations for both ReNoise and the sampler reversing approach. However, in ReNoise, the number of inversion steps remains fixed, and the additional operations are utilized for renoising iterations, refining each point on the inversion trajectory. Consequently, our method facilitates improved reconstruction when using these noise samplers." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.128 + ], + "angle": 0, + "content": "D. Garibi et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.79, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.24, + 0.788, + 0.281 + ], + "angle": 0, + "content": "Fig. 10: LCM Editing Results. We showcase two examples of real images, each followed by three edits. The text below each edited image indicates the specific word or phrase replaced or added to the original prompt for that specific edit." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.283, + 0.408, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.283, + 0.596, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.283, + 0.786, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.38, + 0.788, + 0.477 + ], + "angle": 0, + "content": "Fig. 11: Image reconstruction results comparing sampler reversing inversion techniques across different samplers (e.g., vanilla DDIM inversion) with our ReNoise method using the same sampler. The number of denoising steps remains constant. However, the number of UNet passes varies, with the sampler reversing approach increasing the number of inversion steps, while our method increases the number of renoising iterations. We present various configuration options for our method, including options with or without edit enhancement loss and Noise Correction (NC)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.789, + 0.614 + ], + "angle": 0, + "content": "We continue by evaluating both the sampler reversing approach and our method while maintaining a fixed total number of UNet operations for the inversion and denoising processes combined. The results for SDXL with DDIM are presented in Table 1. The table displays various combinations of inversion, denoising, and renoising steps, totaling 100 UNet operations. Despite employing longer strides along the inversion and denoising trajectories, our ReNoise method yields improved reconstruction accuracy, as evident in the table. Furthermore, a reduced number of denoising steps facilitates faster image editing, especially since it commonly involves reusing the same inversion for multiple edits." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.617, + 0.518, + 0.633 + ], + "angle": 0, + "content": "5.2 Reconstruction and Editability" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.635, + 0.788, + 0.756 + ], + "angle": 0, + "content": "In Figure 10, we illustrate editing results generated by our method with LCM LoRA [27]. These results were obtained by inverting the image using a source prompt and denoising it with a target prompt. Each row exhibits an image followed by three edits accomplished by modifying the original prompt. These edits entail either replacing the object word or adding descriptive adjectives to it. As can be seen, the edited images retain the details present in the original image. For instance, when replacing the cat with a koala, the details in the background are adequately preserved." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.762, + 0.402, + 0.776 + ], + "angle": 0, + "content": "5.3 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Figure 9 qualitatively demonstrates the effects of each component in our method, highlighting their contribution to the final outcome. Here, we use SDXL Turbo model [42], with the Ancestral-Euler sampler, which is non-deterministic. As our baseline, we simply reverse the sampler process. The reconstruction, while" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.35, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "ReNoise: Real Image Inversion Through Iterative Noising" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.209, + 0.142, + 0.795, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.274, + 0.493, + 0.329 + ], + "angle": 0, + "content": "Fig. 12: Comparison with edit-friendly with SDXL Turbo. We inverted the image with the prompt \"a cat is sitting in front of a mirror\" and applied edits." + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.274, + 0.788, + 0.329 + ], + "angle": 0, + "content": "Fig. 13: Image reconstruction comparisons with SD. While NTI and our method achieve comparable results, ours demonstrates significant speed improvement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.332, + 0.788, + 0.513 + ], + "angle": 0, + "content": "semantically capturing the main object, fails to reproduce the image's unique details. For example, in the bird image, the reconstruction contains a bird standing on a branch, but the branch is in a different pose and the bird is completely different. Using 9 ReNoise iterations significantly improves the reconstruction, recovering finer details like the bird's original pose and branch texture. However, some subtle details, such as the bird's colors or the color in Brad Pitt's image, remain incomplete. Averaging the final iterations effectively incorporates information from multiple predictions, leading to a more robust reconstruction that captures finer details. Regularize the UNet's noise prediction with \\(\\mathcal{L}_{\\mathrm{edit}}\\) can introduce minor artifacts to the reconstruction, evident in the smoother appearance of the hair of the two people on the left, or in the cake example. Finally, we present our full method by adding the noise correction technique." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.513, + 0.788, + 0.573 + ], + "angle": 0, + "content": "Table 2 quantitatively showcases the effect each component has on reconstruction results. As can be seen, the best results were obtained by our full method or by averaging the last estimations of \\( z_{t} \\). Our final method also offers the distinct advantage of getting an editable latent representation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.573, + 0.787, + 0.603 + ], + "angle": 0, + "content": "In the Appendix, we present an ablation study to justify our editability enhancement and noise correction components." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.608, + 0.371, + 0.623 + ], + "angle": 0, + "content": "5.4 Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.689 + ], + "angle": 0, + "content": "Inversion for Non-deterministic Samplers. In Figure 12 we show a qualitative comparison with \"an edit-friendly DDPM\" [21] where we utilize SDXL Turbo [42]. Specifically, we assess the performance of the edit-friendly method alongside our ReNoise method in terms of both reconstruction and editing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We observe that in non-deterministic samplers like DDPM, the parameter \\(\\rho_0\\) in Equation 1 equals zero. This means that in the final denoising step, the random noise addition is skipped to obtain a clean image. In long diffusion processes (e.g., 50-100 steps), the final denoising step often has minimal impact as the majority of image details have already been determined. Conversely, shorter diffusion processes rely on the final denoising step to determine fine details of the image. Due to focusing solely on noise correction to preserve the original image in the inversion process, edit-friendly struggles to reconstruct fine details of the image, such as the shower behind the cat. However, our ReNoise method finds an inversion trajectory that faithfully reconstructs the image and does not rely" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "D. Garibi et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "solely on noise corrections. This allows us to better reconstruct fine details such as the shower. Additionally, encoding a significant amount of information within only a few external noise vectors, \\(\\epsilon_{t}\\), limits editability in certain scenarios. See more examples in the appendix" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.213, + 0.79, + 0.412 + ], + "angle": 0, + "content": "Null-prompt Inversion Methods In Figure 13, we present a qualitative comparison between our method and null-text based inversion methods. For this comparison, we utilize Stable Diffusion [39] since these methods rely on a CFG [20] mechanism, which is not employed in SDXL Turbo [42]. Specifically, we compare DDIM Inversion [43] with one renoising iteration to Null-Text Inversion (NTI) [31] and Negative-Prompt Inversion (NPI) [30]. Both NTI and NPI enhance the inversion process by replacing the null-text token embedding when applying CFG. Our method achieves results comparable to NTI, while NPI highlights the limitations of plain DDIM inversion. This is because NPI sets the original prompt as the negative prompt, essentially resulting in an inversion process identical to plain DDIM inversion. Regarding running time, our ReNoise inversion process takes 13 seconds, significantly faster than NTI's 3 minutes. For comparison, plain DDIM inversion and NPI each take 9 seconds." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.419, + 0.362, + 0.435 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.789, + 0.626 + ], + "angle": 0, + "content": "In this work, we have introduced ReNoise, a universal approach that enhances various inversion algorithms of diffusion models. ReNoise gently guides the inversion curve of a real image towards the source noise from which a denoising process reconstructs the image. ReNoise can be considered as a meta-algorithm that warps the trajectory of any iterative diffusion inversion process. Our experiments demonstrate that averaging the last few renoising iterations significantly enhances reconstruction quality. For a fixed amount of computation, ReNoise shows remarkably higher reconstruction quality and editability. The method is theoretically supported and our experiments reconfirm its effectiveness on a variety of diffusion models and sampling algorithms. Moreover, the method is numerically stable, and always converges to some inversion trajectory that eases hyperparameters adjustment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.789, + 0.672 + ], + "angle": 0, + "content": "Beyond the net introduction of an effective inversion, the paper presents a twofold important contribution: an effective inversion for few-steps diffusion models, which facilitates effective editing on these models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.789, + 0.735 + ], + "angle": 0, + "content": "A limitation of ReNoise is the model-specific hyperparameter tuning required for Edit Enhancement and Noise Correction. While these hyperparameters remain stable for a given model, they may vary across models, and tuning them is necessary to achieve high reconstruction quality while maintaining editability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.789, + 0.84 + ], + "angle": 0, + "content": "While ReNoise demonstrates the potential for editing few-step diffusion models, more extensive testing with advanced editing methods is needed. It is worth noting that no such editing has been demonstrated for the few-step diffusion models. We believe and hope that our ReNoise method will pave the way for fast and effective editing methods based on the few-steps models. We also believe that ReNoise can be adapted to the challenging problem of inverting video-diffusion models." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.35, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ReNoise: Real Image Inversion Through Iterative Noising" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.786, + 0.218 + ], + "angle": 0, + "content": "1. Numerical Differential Equation Methods, chap. 2, pp. 45-121. John Wiley and Sons, Ltd (2003). https://doi.org/https://doi.org/10.1002/0470868279.ch2, https://onlinelibrary.wiley.com/doi/abs/10.1002/0470868279.ch2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.787, + 0.273 + ], + "angle": 0, + "content": "2. Abdal, R., Qin, Y., Wonka, P.: Image2stylegan: How to embed images into the stylegan latent space? In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE (Oct 2019). https://doi.org/10.1109/iccv.2019.00453, http://dx.doi.org/10.1109/ICCV.2019.00453" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.273, + 0.786, + 0.328 + ], + "angle": 0, + "content": "3. Abdal, R., Qin, Y., Wonka, P.: Image2stylegan++: How to edit the embedded images? In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (Jun 2020). https://doi.org/10.1109/cvpr42600.2020.00832, http://dx.doi.org/10.1109/cvpr42600.2020.00832" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.33, + 0.786, + 0.357 + ], + "angle": 0, + "content": "4. Alaluf, Y., Garibi, D., Patashnik, O., Averbuch-Elor, H., Cohen-Or, D.: Cross-image attention for zero-shot appearance transfer (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.358, + 0.786, + 0.412 + ], + "angle": 0, + "content": "5. Alaluf, Y., Patashnik, O., Cohen-Or, D.: Restyle: A residual-based stylegan encoder via iterative refinement. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE (Oct 2021). https://doi.org/10.1109/iccv48922.2021.00664, http://dx.doi.org/10.1109/iccv48922.2021.00664" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.413, + 0.786, + 0.439 + ], + "angle": 0, + "content": "6. Alaluf, Y., Tov, O., Mokady, R., Gal, R., Bermano, A.H.: Hyperstyle: Stylegan inversion with hypernetworks for real image editing (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.44, + 0.786, + 0.467 + ], + "angle": 0, + "content": "7. Avrahami, O., Fried, O., Lischinski, D.: Blended latent diffusion. arXiv preprint arXiv:2206.02779 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.468, + 0.786, + 0.509 + ], + "angle": 0, + "content": "8. Avrahami, O., Lischinski, D., Fried, O.: Blended diffusion for text-driven editing of natural images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18208-18218 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.509, + 0.786, + 0.536 + ], + "angle": 0, + "content": "9. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.537, + 0.786, + 0.551 + ], + "angle": 0, + "content": "0. Burden, R., Faires, J., Burden, A.: Numerical Analysis. Cengage Learning (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.552, + 0.786, + 0.578 + ], + "angle": 0, + "content": "1. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. ArXiv abs/2210.11427 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.579, + 0.767, + 0.591 + ], + "angle": 0, + "content": "2. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.786, + 0.633 + ], + "angle": 0, + "content": "3. Dinh, T.M., Tran, A.T., Nguyen, R., Hua, B.S.: Hyperinverter: Improving stylegan inversion via hypernetwork. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.786, + 0.66 + ], + "angle": 0, + "content": "4. Epstein, D., Jabri, A., Poole, B., Efros, A.A., Holynski, A.: Diffusion self-guidance for controllable image generation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.786, + 0.702 + ], + "angle": 0, + "content": "5. Ge, S., Park, T., Zhu, J.Y., Huang, J.B.: Expressive text-to-image generation with rich text. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.703, + 0.786, + 0.757 + ], + "angle": 0, + "content": "6. Han, L., Wen, S., Chen, Q., Zhang, Z., Song, K., Ren, M., Gao, R., Stathopoulos, A., He, X., Chen, Y., Liu, D., Zhangli, Q., Jiang, J., Xia, Z., Srivastava, A., Metaxas, D.: Improving tuning-free real image editing with proximal guidance (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.786, + 0.785 + ], + "angle": 0, + "content": "7. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.786, + 0.813 + ], + "angle": 0, + "content": "8. Hertz, A., Voynov, A., Fruchter, S., Cohen-Or, D.: Style aligned image generation via shared attention (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.814, + 0.744, + 0.826 + ], + "angle": 0, + "content": "19. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.827, + 0.647, + 0.84 + ], + "angle": 0, + "content": "20. Ho, J., Salimans, T.: Classifier-free diffusion guidance (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.127 + ], + "angle": 0, + "content": "D. Garibi et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "21. Huberman-Spiegelglas, I., Kulikov, V., Michaeli, T.: An edit friendly ddpm noise space: Inversion and manipulations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.217 + ], + "angle": 0, + "content": "22. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems 35, 26565-26577 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "23. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Magic: Text-based real image editing with diffusion models. In: Conference on Computer Vision and Pattern Recognition 2023 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.26, + 0.785, + 0.3 + ], + "angle": 0, + "content": "24. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.785, + 0.342 + ], + "angle": 0, + "content": "25. Lin, T., Maire, M., Belongie, S.J., Bourdev, L.D., Girshick, R.B., Hays, J., Perona, P., Ramanan, D., Doll'a r, P., Zitnick, C.L.: Microsoft COCO: common objects in context. CoRR abs/1405.0312 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "26. Luo, S., Tan, Y., Huang, L., Li, J., Zhao, H.: Latent consistency models: Synthesizing high-resolution images with few-step inference (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.411 + ], + "angle": 0, + "content": "27. Luo, S., Tan, Y., Patil, S., Gu, D., von Platen, P., Passos, A., Huang, L., Li, J., Zhao, H.: Lcm-lora: A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.439 + ], + "angle": 0, + "content": "28. Meiri, B., Samuel, D., Darshan, N., Chechik, G., Avidan, S., Ben-Ari, R.: Fixed-point inversion for text-to-image diffusion models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.44, + 0.785, + 0.466 + ], + "angle": 0, + "content": "29. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.785, + 0.494 + ], + "angle": 0, + "content": "30. Miyake, D., Iohara, A., Saito, Y., Tanaka, T.: Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.496, + 0.785, + 0.522 + ], + "angle": 0, + "content": "31. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.563 + ], + "angle": 0, + "content": "32. Pan, Z., Gherardi, R., Xie, X., Huang, S.: Effective real image editing with accelerated iterative diffusion inversion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15912-15921 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.631 + ], + "angle": 0, + "content": "33. Parmar, G., Kumar Singh, K., Zhang, R., Li, Y., Lu, J., Zhu, J.Y.: Zero-shot image-to-image translation. In: Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Proceedings. SIGGRAPH '23, ACM (Jul 2023). https://doi.org/10.1145/3588432.3591513, http://dx.doi.org/10.1145/3588432.3591513" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.7 + ], + "angle": 0, + "content": "34. Parmar, G., Li, Y., Lu, J., Zhang, R., Zhu, J.Y., Singh, K.K.: Spatially-adaptive multilayer selection for gan inversion and editing. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (Jun 2022). https://doi.org/10.1109/cvpr52688.2022.01111, http://dx.doi.org/10.1109/CVPR52688.2022.01111" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.743 + ], + "angle": 0, + "content": "35. Patashnik, O., Garibi, D., Azuri, I., Averbuch-Elor, H., Cohen-Or, D.: Localizing object-level shape variations with text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.785, + 0.785 + ], + "angle": 0, + "content": "36. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "37. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "38. Richardson, E., Alaluf, Y., Patashnik, O., Nitzan, Y., Azar, Y., Shapiro, S., Cohen-Or, D.: Encoding in style: a stylegan encoder for image-to-image translation. In:" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.35, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "ReNoise: Real Image Inversion Through Iterative Noising" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.147, + 0.785, + 0.176 + ], + "angle": 0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "39. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.259 + ], + "angle": 0, + "content": "40. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Ghasemipour, S.K.S., Ayan, B.K., Mahdavi, S.S., Lopes, R.G., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic text-to-image diffusion models with deep language understanding (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.286 + ], + "angle": 0, + "content": "41. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models. In: International Conference on Learning Representations (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.288, + 0.785, + 0.314 + ], + "angle": 0, + "content": "42. Sauer, A., Lorenz, D., Blattmann, A., Rombach, R.: Adversarial diffusion distillation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.733, + 0.328 + ], + "angle": 0, + "content": "43. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.329, + 0.741, + 0.342 + ], + "angle": 0, + "content": "44. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.343, + 0.785, + 0.383 + ], + "angle": 0, + "content": "45. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: International Conference on Learning Representations (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.424 + ], + "angle": 0, + "content": "46. Tov, O., Alaluf, Y., Nitzan, Y., Patashnik, O., Cohen-Or, D.: Designing an encoder for stylegan image manipulation. ACM Trans. Graph. 40(4) (jul 2021). https://doi.org/10.1145/3450626.3459838" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.452 + ], + "angle": 0, + "content": "47. Tumanyan, N., Geyer, M., Bagon, S., Dekel, T.: Plug-and-play diffusion features for text-driven image-to-image translation pp. 1921-1930 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.48 + ], + "angle": 0, + "content": "48. Voynov, A., Hertz, A., Arar, M., Fruchter, S., Cohen-Or, D.: Anylens: A generative diffusion model with any rendering lens (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.508 + ], + "angle": 0, + "content": "49. Wallace, B., Gokul, A., Naik, N.: Edict: Exact diffusion inversion via coupled transformations. arXiv preprint arXiv:2211.12446 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "50. Wu, C.H., De la Torre, F.: Unifying diffusion models' latent space, with applications to cyclediffusion and guidance. arXiv preprint arXiv:2210.05559 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.785, + 0.563 + ], + "angle": 0, + "content": "51. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.605 + ], + "angle": 0, + "content": "52. Zhu, J., Shen, Y., Zhao, D., Zhou, B.: In-domain gan inversion for real image editing. In: Proceedings of European Conference on Computer Vision (ECCV) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.646 + ], + "angle": 0, + "content": "53. Zhu, J.Y., Krahenbihl, P., Shechtman, E., Efros, A.A.: Generative visual manipulation on the natural image manifold. In: Proceedings of European Conference on Computer Vision (ECCV) (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.674 + ], + "angle": 0, + "content": "54. Zhu, P., Abdal, R., Qin, Y., Femiani, J., Wonka, P.: Improved stylegan embedding: Where are the good latents? (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.674 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/ReNoise_ Real Image Inversion Through Iterative Noising/4a999b7d-d6fa-416b-8345-a221ff72abc9_origin.pdf b/2024/ReNoise_ Real Image Inversion Through Iterative Noising/4a999b7d-d6fa-416b-8345-a221ff72abc9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..50efd914cdcbbbecd47b0751b354bfb09a4ac255 --- /dev/null +++ b/2024/ReNoise_ Real Image Inversion Through Iterative Noising/4a999b7d-d6fa-416b-8345-a221ff72abc9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ef33230550ac97dd20e1bfa03e3148f85b0348a8ccc90b31541c23098843e10 +size 50499517 diff --git a/2024/ReNoise_ Real Image Inversion Through Iterative Noising/full.md b/2024/ReNoise_ Real Image Inversion Through Iterative Noising/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f104665cb0205b9d500775cc1983ae277944f845 --- /dev/null +++ b/2024/ReNoise_ Real Image Inversion Through Iterative Noising/full.md @@ -0,0 +1,328 @@ +# ReNoise: Real Image Inversion Through Iterative Noising + +Daniel Garibi1, Or Patashnik1, Andrey Voynov2, Hadar Averbuch-Elor1, and Daniel Cohen-Or1 + +1 Tel Aviv University + +2 Google Research + +![](images/41c090579a293cc48ee4ef721b81c4663f6dcb6c9d3cd0bd9c28676762b66240.jpg) +Fig. 1: Our ReNoise inversion technique can be applied to various diffusion models, including recent few-step ones. This figure illustrates the performance of our method with SDXL Turbo and LCM models, showing its effectiveness compared to DDIM inversion. Additionally, we demonstrate that the quality of our inversions allows prompt-driven editing. As illustrated on the right, our approach also allows for prompt-driven image edits. + +Abstract. Recent advancements in text-guided diffusion models have unlocked powerful image manipulation capabilities. However, applying these methods to real images necessitates the inversion of the images into the domain of the pretrained diffusion model. Achieving faithful inversion remains a challenge, particularly for more recent models trained to generate images with a small number of denoising steps. In this work, we introduce an inversion method with a high quality-to-operation ratio, enhancing reconstruction accuracy without increasing the number of operations. Building on reversing the diffusion sampling process, our method employs an iterative renoising mechanism at each inversion sampling step. This mechanism refines the approximation of a predicted point along the forward diffusion trajectory, by iteratively applying the pretrained diffusion model, and averaging these predictions. We evaluate the performance of our ReNoise technique using various sampling algorithms and models, including recent accelerated diffusion models. Through comprehensive evaluations and comparisons, we show its effectiveness in terms of both accuracy and speed. Furthermore, we confirm that our method preserves editability by demonstrating text-driven image editing on real images. + +Keywords: Diffusion Models $\cdot$ Image Editing $\cdot$ Inversion + +![](images/396a494924ea216dcc23a0b15469fcfd96f59ecf191bad8d9535bf9ae0b534bc.jpg) +Fig.2: The diffusion process samples a Gaussian noise and iteratively denoises it until reaching the data distribution. At each point along the denoising trajectory, the model predicts a direction, $\epsilon_{\theta}(z_t)$ , to step to the next point along the trajectory. To invert a given image from the distribution, the direction from $z_{t}$ to $z_{t + 1}$ is approximated with the inverse of the direction from $z_{t}$ to $z_{t - 1}$ denoted by a dotted blue line. + +![](images/c90c3ee96a4c920e3787eedde0a66016d70d313e0d0b559b867869e1e6f5ca13.jpg) +Fig.3: Comparing reconstruction results of plain DDIM Inv. with SDXL to DDIM Inv. with one renoising iteration. + +# 1 Introduction + +Large-scale text-to-image diffusion models have revolutionized the field of image synthesis [19,37,39,40]. In particular, many works have shown that these models can be employed for various types of image manipulation [4,7-9,15,17,23,29,33, 35,47]. To edit real images, many of these techniques often require the inversion of the image into the domain of the diffusion model. That is, given a real image $z_{0}$ , one has to find a Gaussian noise $z_{T}$ , such that denoising $z_{T}$ with the pretrained diffusion model reconstructs the given real image $z_{0}$ . The importance of this task for real image manipulation has prompted many efforts aimed at achieving accurate reconstruction [16,21,30,31]. + +The diffusion process consists of a series of denoising steps $\{\epsilon_{\theta}(z_t,t)\}_{t = T}^1$ , which form a trajectory from the Gaussian noise to the model distribution (see Figure 2). Each denoising step is computed by a trained network, typically implemented as a UNet, which predicts $z_{t - 1}$ from $z_{t}$ [19]. The output of the model at each step forms a direction from $z_{t}$ to $z_{t - 1}$ [45]. These steps are not invertible, in the sense that the model was not trained to predict $z_{t}$ from $z_{t - 1}$ . Thus, the problem of inverting a given image is a challenge, and particularly for real images, as they are not necessarily in the model distribution (see Figure 3). + +In this paper, we present an inversion method with a high quality-to-operation ratio, which achieves superior reconstruction accuracy for the same number of UNet operations. We build upon the commonly used approach of reversing the diffusion sampling process, which is based on the linearity assumption that the direction from $z_{t}$ to $z_{t+1}$ can be approximated by the negation of the direction from $z_{t}$ to $z_{t-1}$ [12, 43] (see Figure 2). To enhance this approximation, we employ the fixed-point iteration methodology [10]. Specifically, given $z_{t}$ , we begin by using the common approximation to get an initial estimate for $z_{t+1}$ , denoted by $z_{t+1}^{(0)}$ . Then, we iteratively renoise $z_{t}$ , following the direction implied by $z_{t+1}^{(k)}$ to obtain $z_{t+1}^{(k+1)}$ . After repeating this ronoising process several times, we apply an averaging on $z_{t+1}^{(k)}$ to form a more accurate direction from $z_{t}$ to $z_{t+1}$ . + +We show that this approach enables longer strides along the inversion trajectory while improving image reconstruction. Therefore, our method can also be effective with diffusion models trained to generate images using a small number of denoising steps [27, 42]. Furthermore, despite the need to repeatedly renoise in each step of the inversion process, the longer strides lead to a more favorable tradeoff of UNet operations for reconstruction quality. + +Through extensive experiments, we demonstrate the effectiveness of our method in both image reconstruction and inversion speed. We validate the versatility of our approach across different samplers and models, including recent time-distilled diffusion models (e.g., SDXL-Turbo [42]). Importantly, we demonstrate that the editability of the inversion achieved by our method allows a wide range of text-driven image manipulations (see Figure 1). + +# 2 Related Work + +Image Editing via Diffusion Models Recent advancements in diffusion models [12, 19] have resulted in unprecedented diversity and fidelity in visual content creation guided by free-form text prompts [36, 37, 39, 40]. Text-to-image models do not directly support text-guided image editing. Therefore, harnessing the power of these models for image editing is a significant research area and many methods have utilized these models for different types of image editing [4, 7-9, 11, 14, 15, 17, 18, 23, 29, 33, 35, 47, 48]. A common approach among these methods requires inversion [21, 31, 43, 49] to edit real images, i.e., obtaining a latent code $z_{T}$ such that denoising it with the pretrained diffusion model returns the original image. Specifically, in this approach two backward processes are done simultaneously using $z_{T}$ . One of the processes reconstructs the image using the original prompt, while the second one injects features from the first process (e.g., attention maps) to preserve some properties of the original image while manipulating other aspects of it. + +Inversion in Diffusion Models Initial efforts in image inversion for real image editing focused on GANs [2,3,5,6,13,34,38,46,52-54]. The advancements in diffusion models, and in diffusion-based image editing in particular have recently prompted works studying the inversion of a diffusion-based denoising process. This inversion depends on the sampler algorithm used during inference, which can be deterministic [43] or non-deterministic [19, 22]. Inversion methods can be accordingly categorized into two: methods that are suitable for deterministic sampling, and methods suitable for non-deterministic sampling. + +Methods that approach the deterministic inversion commonly rely on the DDIM sampling method [43], and build upon DDIM inversion [12, 43]. Mokady et al. [31] observed that the use of classifier-free guidance during inference magnifies the accumulated error of DDIM inversion and therefore leads to poor reconstruction. Following this observation, several works [16, 30, 31] focused on solving this issue by replacing the null text token with a different embedding, which is found after an optimization process or by a closed solution. However, excluding [31] which requires a lengthy optimization, these methods are limited by the reconstruction accuracy of DDIM inversion, which can be poor, especially when a small number of denoising steps is done. In our work, we present a method + +that improves the reconstruction quality of DDIM inversion and therefore can be integrated with methods that build on it. + +Another line of work [21, 50] tackles the inversion of DDPM sampler [19]. In these works [21, 50], instead of inverting the image into an initial noise $z_{T}$ , a series of noises $\{z_{T},\epsilon_{T},\dots,\epsilon_{1}\}$ is obtained. The definition of this noises series ensures that generating an image with it returns the original input image. However, these methods require a large number of inversion and denoising steps to allow image editing. Applying these methods with an insufficient number of steps leads to too much information encoded in $\{\epsilon_T,\dots,\epsilon_1\}$ which limits the ability to edit the generated image. As shall be shown, The editability issue of these methods is particularly evident in few-steps models [26, 27, 42]. + +Most relevant to our work, two recent inversion methods [28,32] also use the fixed-point iteration technique. Specifically, they improve the reconstruction accuracy of DDIM inversion [43] with Stable Diffusion [39] without introducing a significant computational overhead. In our work, we focus on the problem of real image inversion for recently introduced few-step diffusion models, where the difficulties encountered by previous methods are amplified. Furthermore, we show that our inversion method successfully works with various models and different samplers. + +Few Steps Models Recently, new methods [26, 27, 41, 42, 44] that fine-tune text-to-image diffusion models enabled a significant reduction of the number of steps needed for high-quality image generation. While standard diffusion models typically require 50 denoising steps to generate high-quality images, recent accelerated models achieve high-quality synthesis with 1-4 steps only. These new methods pave the way for interactive editing workflows. However, as we show in this paper, using current methods for the inversion of an image with a small number of steps degrades the reconstruction quality in terms of accuracy [12, 43] or editability [21, 50]. + +# 3 Method + +# 3.1 ReNoise Inversion + +Reversing the Sampler Samplers play a critical role in the diffusion-based image synthesis process. They define the noising and denoising diffusion processes and influence the processes' trajectories and quality of the generated images. While different samplers share the same pre-trained UNet model (denoted by $\epsilon_{\theta}$ ) as their backbone, their sampling approaches diverge, leading to nuanced differences in output. The goal of the denoising sampler is to predict the latent code at the previous noise level, $z_{t-1}$ , based on the current noisy data $z_{t}$ , the pretrained UNet model, and a sampled noise, $\epsilon_{t}$ . Various first-order denoising sampling algorithms adhere to the form: + +$$ +z _ {t - 1} = \phi_ {t} z _ {t} + \psi_ {t} \epsilon_ {\theta} (z _ {t}, t, c) + \rho_ {t} \epsilon_ {t}, \tag {1} +$$ + +where $c$ represents a text embedding condition, and $\phi_t, \psi_t$ , and $\rho_t$ denote sampler parameters. At each step, these parameters control the extent to which the + +previous noise is removed $(\phi_t)$ , the significance assigned to the predicted noise from the UNet $(\psi_t)$ , and the given weight to the additional noise introduced $(\rho_t)$ . + +A given image $z_0$ can be inverted by reformulating Equation 1 and applying it iteratively: + +$$ +z _ {t} = \frac {z _ {t - 1} - \psi_ {t} \epsilon_ {\theta} \left(z _ {t} , t , c\right) - \rho_ {t} \epsilon_ {t}}{\phi_ {t}}, \tag {2} +$$ + +where for non-deterministic samplers, a series of random noises $\{\epsilon_t\}_{t=1}^T$ is sampled and used during both inversion and image generation processes. However, directly computing $z_t$ from Equation 2 is infeasible since it relies on $\epsilon_\theta(z_t, t, c)$ which, in turn, depends on $z_t$ , creating a circular dependency. To solve this implicit function, Dhariwal et al. [12] propose using the approximation $\epsilon_\theta(z_t, t, c) \approx \epsilon_\theta(z_{t-1}, t, c)$ : + +$$ +z _ {t} ^ {(1)} = \frac {z _ {t - 1} - \psi_ {t} \epsilon_ {\theta} \left(z _ {t - 1} , t , c\right) - \rho_ {t} \epsilon_ {t}}{\phi_ {t}}. \tag {3} +$$ + +This method has several limitations. First, the assumption underlying the approximation used in [12] is that the number of inversion steps is large enough, implying a trajectory close to linear. This assumption restricts the applicability of this inversion method in interactive image editing with recent few-step diffusion models [26, 27, 42, 44], as the inversion process would take significantly longer than inference. Second, this method struggles to produce accurate reconstructions in certain cases, such as highly detailed images or images with large smooth regions, see Figure 3. Moreover, we observe that this inversion method is sensitive to the prompt $c$ and may yield poor results for certain prompts. + +ReNoise In a successful inversion trajectory, the direction from $z_{t-1}$ to $z_t$ aligns with the direction from $z_t$ to $z_{t-1}$ in the denoising trajectory. To achieve this, we aim to improve the approximation of $\epsilon_{\theta}(z_t,t,c)$ in Eq. 2 compared to the one used in [12]. Building on the fixed-point iteration technique [10], our approach better estimates the instance of $z_t$ that is inputted to the UNet, rather than relying on $z_{t-1}$ . + +Intuitively, we utilize the observation that $z_{t}^{(1)}$ (from Eq. 3) offers a more precise estimate of $z_{t}$ compared to $z_{t-1}$ . Therefore, employing $z_{t}^{(1)}$ as the input to the UNet is likely to yield a more accurate direction, thus contributing to reducing the overall error in the inversion step. We illustrate this observation in Figure 5. Iterating this process generates a series of estimations for $z_{t}$ , denoted by $\{z_{t}^{(k)}\}_{k=1}^{\mathcal{K}+1}$ . While the fixed-point iteration technique [10] does not guarantee convergence of this series in the general case, in Section 4, we empirically show that convergence holds in our setting. However, as the convergence is not monotonic, we refine our prediction of $z_{t}$ by averaging several $\{z_{t}^{(k)}\}$ , thus considering more than a single estimation of $z_{t}$ . See Figure 6 for an intuitive illustration. + +In more detail, our method iteratively computes estimations of $z_{t}$ during each inversion step $t$ by renoising the noisy latent $z_{t-1}$ multiple times, each with a different noise prediction (see Figure 4). Beginning with $z_{t}^{(1)}$ , in the $k$ -th renoising iteration, the input to the UNet is the result of the previous iteration, $z_{t}^{(k)}$ . Then, $z_{t}^{(k+1)}$ is calculated using the inverted sampler while maintaining $z_{t-1}$ as the starting point of the step. After $\mathcal{K}$ renoising iterations, we obtain a set of + +![](images/5742e3c782e5d830b04cea56a12460576f346867f29852aa2c51d010d3e236e6.jpg) +Fig. 4: Method overview. Given an image $z_0$ , we iteratively compute $z_1, \dots, z_T$ , where each $z_t$ is calculated from $z_{t-1}$ . At each time step, we apply the UNet $(\epsilon_\theta) \mathcal{K} + 1$ times, each using a better approximation of $z_t$ as the input. The initial approximation is $z_{t-1}$ . The next one, $z_t^{(1)}$ , is the result of the reversed sampler step (i.e., DDIM). The reversed step begins at $z_{t-1}$ and follows the direction of $\epsilon_\theta(z_{t-1}, t)$ . At the $k$ renoising iteration, $z_t^{(k)}$ is the input to the UNet, and we obtain a better $z_t$ approximation. For the last iterations, we optimize $\epsilon_\theta(z_t^{(k)}, t)$ to increase editability. As the final denoising direction, we use the average of the UNet predictions of the last few iterations. + +Fig.5: Geometric intuition for ReNoise. At each inversion step, we estimate $z_{t}$ (marked with a red star) based on $z_{t-1}$ . The straightforward approach is to use the negated direction of the denoising step from $z_{t-1}$ , assuming the trajectory is approximately linear. However, this assumption is inaccurate, especially in few-step models, where the size of the steps is large. We use the linearity assumption as an initial estimation and keep improving this estimation. We recalculate the denoising step from the previous estimation (which is closer to the target $z_{t}$ ) and then proceed with its negated direction from $z_{t-1}$ (see the orange vectors). + +![](images/b9e58c72e7338254c2d815309cd5e7fa86ddf0ec2dbf1dca4521e55476645c43.jpg) + +estimations $\{z_t^{(k)}\}_{k = 1}^{\mathcal{K} + 1}$ . The next point on the inversion trajectory, $z_{t}$ , is then defined as their weighted average, where $w_{k}$ is the weight assigned to $z_{t}^{(k)}$ . For a detailed description of our method, refer to Algorithm 1. + +# 3.2 Reconstruction-Editability Tradeoff + +Enhance Editability The goal of inversion is to facilitate the editing of real images using a pretrained image generation model. While the renoising approach attains highly accurate reconstruction results, we observe that the resulting $z_{T}$ lacks editability. This phenomenon can be attributed to the reconstruction-editability tradeoff in image generative models [46]. To address this limitation, we incorporate a technique to enhance the editability of our method. + +It has been shown [33] that the noise maps predicted during the inversion process often diverge from the statistical properties of uncorrelated Gaussian white noise, thereby affecting editability. To tackle this challenge, we follow pix2pix- + +zero [33] and regularize the predicted noise at each step, $\epsilon_{\theta}(z_t,t,c)$ , using the following loss terms. + +First, we encourage $\epsilon_{\theta}(z_t,t,c)$ to follow the same distribution as $\epsilon_{\theta}(z_t^{\prime},t,c)$ , where $z_{t}^{\prime}$ represents the input image $z_{0}$ with added random noise corresponding to the noise level at timestep $t$ . We do so by dividing $\epsilon_{\theta}(z_t,t,c)$ and $\epsilon_{\theta}(z_t^{\prime},t,c)$ into small patches (e.g., $4\times 4$ ), and computing the KL-divergence between corresponding patches. We denote this loss term by $\mathcal{L}_{\mathrm{patch - KL}}$ . Second, we utilize $\mathcal{L}_{\mathrm{pair}}$ proposed in pix2pix-zero [33], which penalizes correlations between pairs of pixels. We leverage these losses to enhance the editability of our method, and denote the combination of them as $\mathcal{L}_{\mathrm{edit}}$ . For any renoising iteration $k$ where $w_{k} > 0$ , we regularize the UNet's prediction $\epsilon_{\theta}(z_t^{(k)},t,c)$ using $\mathcal{L}_{\mathrm{edit}}$ before computing $z_{t}^{(k + 1)}$ . See line 9 in Algorithm 1. + +Algorithm 1: ReNoise Inversion +1 Input: An image $z_{0}$ , number of renoising steps $\mathcal{K}$ , number of inversion steps $T$ , a series of renoising weights $\{w_{k}\}_{k=1}^{\mathcal{K}+1}$ . +2 Output: A noisy latent $z_{T}$ and set of noises $\{\epsilon_{t}\}_{t=1}^{T}$ . +3 for $t = 0,1,\ldots,T$ do +4 sample $\epsilon_{t} \sim \mathcal{N}(0,I)$ +5 $z_{t}^{(0)} \gets z_{t-1}$ +6 $z_{t}^{(\mathrm{avg})} \gets 0$ +7 for $k = 0,\ldots,\mathcal{K}$ do +8 $\delta_{t}^{k} \gets \epsilon_{\theta}(z_{t}^{(k)},t)$ +9 $\delta_{t}^{k} \gets \text{Enhance-edit}(\delta_{t}^{k},w_{k+1})$ +10 $z_{t}^{(k+1)} \gets \text{Inverse-Step}(z_{t-1},\delta_{t}^{k})$ +11 end +// Average ReNoised predictions +23 Function Noise-Corr $(z_{t},t,\epsilon_{t},z_{t-1})$ : +12 $z_{t}^{(\mathrm{avg})} \gets \sum_{k=1}^{\mathcal{K}+1} w_{k} \cdot z_{t}^{(k)}$ +24 $\delta_{t} \gets \epsilon_{\theta}(z_{t},t)$ +25 $\epsilon_{t} \gets \epsilon_{t} - \nabla_{\epsilon_{t}}\frac{1}{\rho_{t}}(z_{t-1} - \phi_{t}z_{t} - \psi_{t}\delta_{t})$ +26 return $\epsilon_{t}$ +end +return $(z_{T},\{\epsilon_{t}\}_{t=1}^{T})$ + +Noise Correction in Non-deterministic Samplers Non-deterministic samplers, in which $\rho_t > 0$ , introduce noise $(\epsilon_t)$ at each denoising step. Previous methods [21,50] suggested using $\epsilon_t$ to bridge the gap between the inversion and denoising trajectories in DDPM inversion. Specifically, given a pair of points $z_{t - 1},z_{t}$ on the inversion trajectory, we denote by $\hat{z}_{t - 1}$ the point obtained by denoising $z_{t}$ . Ideally, $z_{t - 1}$ and $\hat{z}_{t - 1}$ should be identical. We define: + +$$ +\epsilon_ {t} = \frac {1}{\rho_ {t}} \left(z _ {t - 1} - \phi_ {t} z _ {t} - \psi_ {t} \epsilon_ {\theta} \left(z _ {t}, t, c\right)\right). \tag {4} +$$ + +Integrating this definition into Eq. 1 yields $\hat{z}_{t - 1} = z_{t - 1}$ . However, we found that replacing $\epsilon_t$ with the above definition affects editability. Instead, we suggest a more tender approach, optimizing $\epsilon_t$ based on Eq. 4 as our guiding objective: + +$$ +\epsilon_ {t} = \epsilon_ {t} - \nabla_ {\epsilon_ {t}} \frac {1}{\rho_ {t}} \left(z _ {t - 1} - \phi_ {t} z _ {t} - \psi_ {t} \epsilon_ {\theta} \left(z _ {t}, t, c\right)\right). \tag {5} +$$ + +This optimization improves the reconstruction fidelity while preserving the distribution of the noisy-latents. + +![](images/425c549fb37a8697b4d1460cd6a27c7a715f365226c14e05a66f37013253ea2f.jpg) +Fig. 6: Schematic illustration of the ReNoise convergence process to the true inversion of $z_{t-1}$ . While estimates may converge non-monotonically to the unknown target $z_t$ , we found that averaging them improves true value estimation. Typically, the initial iteration exhibits an exponential decrease in the norm between consecutive elements. + +![](images/48283f72459cd1c42321171bc038cb2f7c828989379fbd874f737b41062e1091.jpg) +Fig.7: Average distance between consequent estimations $z_{t}^{(k)}$ , and $z_{t}^{(k + 1)}$ . Vertical bars indicate the standard deviation. The averages are computed over 32 images and 10 different timesteps. + +# 4 Convergence Discussion + +In this section, we first express the inversion process as a backward Euler process and our renoising iterations as fixed-point iterations. While these iterations do not converge in the general case, in the supplementary materials we present a toy example where they yield accurate inversions. Then, we analyze the convergence of the renoising iterations in our real-image inversion scenario and empirically verify our method's convergence. + +Inversion Process as Backward Euler The denoising process of diffusion models can be mathematically described as solving an ordinary differential equation (ODE). A common method for solving such equations is the Euler method, which takes small steps to approximate the solution. For ODE in the form of $y'(t) = f(t,y(t))$ , Euler solution is defined as: + +$$ +y _ {n + 1} = y _ {n} + h \cdot f (t _ {n}, y _ {n}), +$$ + +where $h$ is the step size. The inversion process can be described as solving ODE using the backward Euler method (or implicit Euler method) [1]. This method is similar to forward Euler, with the difference that $y_{n + 1}$ appears on both sides of the equation: + +$$ +y _ {n + 1} = y _ {n} + h \cdot f \left(t _ {n + 1}, y _ {n + 1}\right). +$$ + +For equations lacking an algebraic solution, several techniques estimate $y_{n+1}$ iteratively. As we described in Section 3.1, the inversion process lacks a closed-form solution, as shown in Equation 2. To address this, the ReNoise method leverages fixed-point iterations, which we refer to as reonising iterations, to progressively refine the estimate of $y_{n+1}$ : + +$$ +y _ {n + 1} ^ {(0)} = y _ {n}, \quad y _ {n + 1} ^ {(k + 1)} = y _ {n} + h \cdot f (t _ {n + 1}, y _ {n + 1} ^ {(k)}. +$$ + +In our ReNoise method, we average these renoising iterations to mitigate convergence errors, leading to improvement in the reconstruction quality. + +ReNoise Convergence During the inversion process, we aim to find the next noise level inversion, denoted by $\hat{z}_t$ , such that applying the denoising step to $\hat{z}_t$ recovers the previous state, $z_{t - 1}$ . Given the noise estimation $\epsilon_{\theta}(z_t,t)$ and a given $z_{t - 1}$ , the ReNoise mapping defined in Section 3.1 can be written as $\mathcal{G}:z_t\to \mathrm{InverseStep}(z_{t - 1},\epsilon_\theta (z_t,t))$ . For example, in the case of using DDIM sampler the mapping is $\mathcal{G}(z_t) = \frac{1}{\phi_t} (z_{t - 1} - \psi_t\epsilon_\theta (z_t,t))$ . The point $\hat{z}_t$ , which is + +mapped to $z_{t-1}$ after the denoising step, is a stationary point of this mapping. Given $z_t^{(1)}$ , the first approximation of the next noise level $z_t$ , our goal is to show that the sequence $z_t^{(k)} = \mathcal{G}^{k-1}(z_t^{(1)})$ , $k \to \infty$ converges. As the mapping $\mathcal{G}$ is continuous, the limit point would be its stationary point. The definition of $\mathcal{G}$ gives: + +$$ +\left\| z _ {t} ^ {(k + 1)} - z _ {t} ^ {(k)} \right\| = \left\| \mathcal {G} \left(z _ {t} ^ {(k)}\right) - \mathcal {G} \left(z _ {t} ^ {(k - 1)}\right) \right\|, +$$ + +where the norm is always assumed as the $l_{2}$ -norm. For the ease of the notations, we define $\varDelta^{(k)}=z_t^{(k)}-z_t^{(k-1)}$ . For convergence proof, it is sufficient to show that the sum of norms of these differences converges, which will imply that $z_{t}^{(k)}$ is the Cauchy sequence. Below we check that in practice $\| \varDelta^{(k)}\|$ decreases exponentially as $k\to \infty$ and thus has finite sum. In the assumption that $\mathcal{G}$ is $\mathcal{C}^2$ -smooth, the Taylor series conducts: + +$$ +\begin{array}{l} \| \varDelta^ {(k + 1)} \| = \| \mathcal {G} (z _ {t} ^ {(k)}) - \mathcal {G} (z _ {t} ^ {(k - 1)}) \| = \\ \left\| \mathcal {G} \left(z _ {t} ^ {(k - 1)}\right) + \frac {\partial \mathcal {G}}{\partial z} \right| _ {z _ {t} ^ {(k - 1)}} \cdot \Delta^ {(k)} + O \left(\left\| \Delta^ {(k)} \right\| ^ {2}\right) - \mathcal {G} \left(z _ {t} ^ {(k - 1)}\right) \| = \\ \left. \left\| \frac {\partial \mathcal {G}}{\partial z} \right| _ {z _ {t} ^ {(k - 1)}} \cdot \Delta^ {(k)} + O (\left\| \Delta^ {(k)} \right\| ^ {2}) \right\| \leq \left\| \frac {\partial \mathcal {G}}{\partial z} \right| _ {z _ {t} ^ {(k - 1)}} \| \cdot \| \Delta^ {(k)} \| + O (\left\| \Delta^ {(k)} \right\| ^ {2}) = \\ \frac {\psi_ {t}}{\phi_ {t}} \cdot \left\| \frac {\partial \epsilon_ {\theta}}{\partial z} \right| _ {z _ {t} ^ {(k - 1)}} \| \cdot \| \Delta^ {(k)} \| + O (\| \Delta^ {(k)} \| ^ {2}) \\ \end{array} +$$ + +Thus, in a sufficiently small neighborhood, the convergence dynamics is defined by the scaled Jacobian norm $\frac{\psi_t}{\phi_t} \cdot \left\| \frac{\partial \epsilon_\theta}{\partial z} \right|_{z_t^{(k-1)}}$ . In the supplementary materials, we show this scaled norm estimation for the SDXL diffusion model for various steps and ReNoise iterations indices $(k)$ . Remarkably, the ReNoise indices minimally impact the scale factor, consistently remaining below 1. This confirms in practice the convergence of the proposed algorithm. Notably, the highest scaled norm values occur at smaller $t$ (excluding the first step) and during the initial renoising iteration. This validates the strategy of not applying ReNoise in early steps, where convergence tends to be slower compared to other noise levels. Additionally, the scaled norm value for the initial $t$ approaches 0, which induces almost immediate convergence. + +Figure 7 illustrates the exponential decrease in distances between consecutive elements $z_{t}^{(k)}$ and $z_{t}^{(k + 1)}$ , which confirms the algorithm's convergence towards the stationary point of the operator $\mathcal{G}$ . The proposed averaging strategy is aligned with the conclusions described above, and also converges to the desired stationary point. In The supplementary materials, we present a validation for this claim. + +# 5 Experiments + +In this section, we conduct extensive experiments to validate the effectiveness of our method. We evaluate both the reconstruction quality of our inversion and its editability. To demonstrate the versatility of our approach, we apply it to four models, SD [39], SDXL [36], SDXL Turbo [42], and LCM-LoRA [27], with SDXL Turbo and LCM-LoRA being few-step models. Additionally, we use various sampling algorithms including both deterministic and non-deterministic ones. Implementation details for each model are provided in the supplementary materials. Following previous works [5, 31], we quantitatively evaluate our + +Table 1: Image reconstruction results with a fixed number of 100 UNet operations. Each row showcases the results obtained using different combinations of inversion steps, denoising steps, and renoising iterations, totaling 100 operations. As observed, allocating some of the operations to renoising iterations improves the reconstruction quality while maintaining the same execution time. + +
Image Reconstruction Results
Inv StepsInf StepsReNoise StepsL2 ↓PSNR ↑LPIPS ↓
505000.0036426.0230.06273
752500.0038225.4660.06605
802000.0040825.0450.07099
901000.0102320.2490.10305
252520.0018229.5690.03637
202030.0016729.8840.03633
101080.0023028.1560.04678
+ +Table 2: Quantitative ablation study on SDXL Turbo. We demonstrate the impact of each component of our inversion method on reconstruction results. The results improve with additional renoising iterations and significant enhancements occur through averaging final estimations. Additionally, we observe a reconstruction-editability trade-off, with edit losses causing degradation that is effectively mitigated by Noise Correction. + +
Ablation - Image Reconstruction
L2 ↓PSNR ↑LPIPS↓
Euler Inversion0.070011.7840.20337
+ 1 ReNoise0.055212.7960.20254
+ 4 ReNoise0.024916.5210.14821
+ 9 ReNoise0.012619.7020.10850
+ Averaging ReNoise0.008721.4910.08832
+ Edit Losses0.027618.4320.12616
+ Noise Correction0.019622.0770.08469
+ +method with three metrics: $L_{2}$ , LPIPS [51], and PSNR. Unless stated otherwise, for both inversion and generation we use the prompt obtained from BLIP2 [24]. + +# 5.1 Reconstruction and Speed + +We begin by evaluating the reconstruction-speed tradeoff. The main computational cost of both the inversion and denoising processes is the forward pass through the UNet. In each renoising iteration, we perform one forward pass, which makes it computationally equal to a standard inversion step (as done in DDIM Inversion for example). In the following experiments, we compare the results of a sampler reversing with our method, where we match the number of UNet passes between the methods. For example, 8 steps of sampler reversing are compared against 4 steps with one renoising iteration at each step. + +Qualitative Results In Figure 8 we show qualitative results of image reconstruction on SDXL Turbo [42]. Here, we utilize DDIM as the sampler, and apply four denoising steps for all configurations. Each row exhibits results obtained using a different amount of UNet operations. In our method, we apply four inversion steps, and a varying number of renoising iterations. As can be seen, the addition of renoising iterations gradually improves the reconstruction results. Conversely, employing more inversion steps proves insufficient for capturing all details in the image, as evident by the background of the car, or even detrimental to the reconstruction, as observed in the Uluro example. + +Quantitative Results For the quantitative evaluation, we use the MS-COCO 2017 [25] validation dataset. Specifically, we retain images with a resolution greater than $420 \times 420$ , resulting in a dataset containing 3,865 images. + +We begin by evaluating both the sampler reversing approach and our ReNoise method, while varying the number of UNet operations during the inversion process and keeping the number of denoising steps fixed. This experiment is conducted using various models (SDXL, SDXL Turbo, LCM) and samplers. For all + +![](images/06f5e129beec9c5708f7efc9239a1d39b4d7c391a5fa130c3cf4d14d2b0a970e.jpg) +Fig.8: Qualitative comparison between DDIM Inversion and our ReNoise method on SDXL Turbo (4 denoising steps). The first row displays the original images. The following rows display reconstructions from both methods, using the same number of UNet operations. DDIM employs smaller strides, while ReNoise utilizes more renoising steps. + +![](images/fefb9658da53f8e6004d929868ba26da842bf11bb0b72813f9e0c438eb103048.jpg) +Fig. 9: Ablation study on SDXL Turbo. The first row presents the input image. In each subsequent row, we show the reconstruction results using an additional component of our inversion method. The images in the bottom row represent the results obtained by our full method. NC stands for Noise Correction. + +models, we utilize the DDIM [43] sampler. In addition, we employ the Ancestral-Euler scheduler for SDXL Turbo, and the default LCM sampler for LCM-LoRA. We set the number of denoising steps to 50 for SDXL, and to 4 for SDXL Turbo and LCM-LoRA. Quantitative results, using PSNR as the metric, are presented in Figure 11. We evaluate our method using different configurations. The x-axis refers to the number of UNet operations in the inversion process. Other metrics results are provided in the supplementary materials. + +As depicted in the graphs, incorporating additional renoising iterations proves to be more beneficial for image reconstruction compared to adding more inversion steps. Note that the performance of the Ancestral-Euler and LCM samplers noticeably degrades when the number of inversion steps exceeds the number of denoising steps. Unlike DDIM, these samplers have $\varPhi_t\approx 1$ , resulting in an increase in the latent vector's norm beyond what can be effectively denoised in fewer steps. In this experiment, we maintain the same number of UNet operations for both ReNoise and the sampler reversing approach. However, in ReNoise, the number of inversion steps remains fixed, and the additional operations are utilized for renoising iterations, refining each point on the inversion trajectory. Consequently, our method facilitates improved reconstruction when using these noise samplers. + +![](images/73d002b752f357a55d5bced7a5dd06a7bc6cca1654498a3320305f17127d0034.jpg) +Fig. 10: LCM Editing Results. We showcase two examples of real images, each followed by three edits. The text below each edited image indicates the specific word or phrase replaced or added to the original prompt for that specific edit. + +![](images/acff721aa486e340a5a365ddb7f88be1de13d158959fb388ce73e5d484ee4189.jpg) +Fig. 11: Image reconstruction results comparing sampler reversing inversion techniques across different samplers (e.g., vanilla DDIM inversion) with our ReNoise method using the same sampler. The number of denoising steps remains constant. However, the number of UNet passes varies, with the sampler reversing approach increasing the number of inversion steps, while our method increases the number of renoising iterations. We present various configuration options for our method, including options with or without edit enhancement loss and Noise Correction (NC). + +![](images/9dd21dbad5845d21963121b5403376123930d2c7f6438b40648649d27366aad9.jpg) + +![](images/3eac0152bacc04aa5316a7a9fc35c993e0fd0e7f7e169a34a269fba38abc4b33.jpg) + +We continue by evaluating both the sampler reversing approach and our method while maintaining a fixed total number of UNet operations for the inversion and denoising processes combined. The results for SDXL with DDIM are presented in Table 1. The table displays various combinations of inversion, denoising, and renoising steps, totaling 100 UNet operations. Despite employing longer strides along the inversion and denoising trajectories, our ReNoise method yields improved reconstruction accuracy, as evident in the table. Furthermore, a reduced number of denoising steps facilitates faster image editing, especially since it commonly involves reusing the same inversion for multiple edits. + +# 5.2 Reconstruction and Editability + +In Figure 10, we illustrate editing results generated by our method with LCM LoRA [27]. These results were obtained by inverting the image using a source prompt and denoising it with a target prompt. Each row exhibits an image followed by three edits accomplished by modifying the original prompt. These edits entail either replacing the object word or adding descriptive adjectives to it. As can be seen, the edited images retain the details present in the original image. For instance, when replacing the cat with a koala, the details in the background are adequately preserved. + +# 5.3 Ablation Studies + +Figure 9 qualitatively demonstrates the effects of each component in our method, highlighting their contribution to the final outcome. Here, we use SDXL Turbo model [42], with the Ancestral-Euler sampler, which is non-deterministic. As our baseline, we simply reverse the sampler process. The reconstruction, while + +![](images/2abb376c359ed2ac846535570dac97e91daa302ad0e8e68fd08eff26f4882a61.jpg) +Fig. 12: Comparison with edit-friendly with SDXL Turbo. We inverted the image with the prompt "a cat is sitting in front of a mirror" and applied edits. +Fig. 13: Image reconstruction comparisons with SD. While NTI and our method achieve comparable results, ours demonstrates significant speed improvement. + +semantically capturing the main object, fails to reproduce the image's unique details. For example, in the bird image, the reconstruction contains a bird standing on a branch, but the branch is in a different pose and the bird is completely different. Using 9 ReNoise iterations significantly improves the reconstruction, recovering finer details like the bird's original pose and branch texture. However, some subtle details, such as the bird's colors or the color in Brad Pitt's image, remain incomplete. Averaging the final iterations effectively incorporates information from multiple predictions, leading to a more robust reconstruction that captures finer details. Regularize the UNet's noise prediction with $\mathcal{L}_{\mathrm{edit}}$ can introduce minor artifacts to the reconstruction, evident in the smoother appearance of the hair of the two people on the left, or in the cake example. Finally, we present our full method by adding the noise correction technique. + +Table 2 quantitatively showcases the effect each component has on reconstruction results. As can be seen, the best results were obtained by our full method or by averaging the last estimations of $z_{t}$ . Our final method also offers the distinct advantage of getting an editable latent representation. + +In the Appendix, we present an ablation study to justify our editability enhancement and noise correction components. + +# 5.4 Comparisons + +Inversion for Non-deterministic Samplers. In Figure 12 we show a qualitative comparison with "an edit-friendly DDPM" [21] where we utilize SDXL Turbo [42]. Specifically, we assess the performance of the edit-friendly method alongside our ReNoise method in terms of both reconstruction and editing. + +We observe that in non-deterministic samplers like DDPM, the parameter $\rho_0$ in Equation 1 equals zero. This means that in the final denoising step, the random noise addition is skipped to obtain a clean image. In long diffusion processes (e.g., 50-100 steps), the final denoising step often has minimal impact as the majority of image details have already been determined. Conversely, shorter diffusion processes rely on the final denoising step to determine fine details of the image. Due to focusing solely on noise correction to preserve the original image in the inversion process, edit-friendly struggles to reconstruct fine details of the image, such as the shower behind the cat. However, our ReNoise method finds an inversion trajectory that faithfully reconstructs the image and does not rely + +solely on noise corrections. This allows us to better reconstruct fine details such as the shower. Additionally, encoding a significant amount of information within only a few external noise vectors, $\epsilon_{t}$ , limits editability in certain scenarios. See more examples in the appendix + +Null-prompt Inversion Methods In Figure 13, we present a qualitative comparison between our method and null-text based inversion methods. For this comparison, we utilize Stable Diffusion [39] since these methods rely on a CFG [20] mechanism, which is not employed in SDXL Turbo [42]. Specifically, we compare DDIM Inversion [43] with one renoising iteration to Null-Text Inversion (NTI) [31] and Negative-Prompt Inversion (NPI) [30]. Both NTI and NPI enhance the inversion process by replacing the null-text token embedding when applying CFG. Our method achieves results comparable to NTI, while NPI highlights the limitations of plain DDIM inversion. This is because NPI sets the original prompt as the negative prompt, essentially resulting in an inversion process identical to plain DDIM inversion. Regarding running time, our ReNoise inversion process takes 13 seconds, significantly faster than NTI's 3 minutes. For comparison, plain DDIM inversion and NPI each take 9 seconds. + +# 6 Conclusion + +In this work, we have introduced ReNoise, a universal approach that enhances various inversion algorithms of diffusion models. ReNoise gently guides the inversion curve of a real image towards the source noise from which a denoising process reconstructs the image. ReNoise can be considered as a meta-algorithm that warps the trajectory of any iterative diffusion inversion process. Our experiments demonstrate that averaging the last few renoising iterations significantly enhances reconstruction quality. For a fixed amount of computation, ReNoise shows remarkably higher reconstruction quality and editability. The method is theoretically supported and our experiments reconfirm its effectiveness on a variety of diffusion models and sampling algorithms. Moreover, the method is numerically stable, and always converges to some inversion trajectory that eases hyperparameters adjustment. + +Beyond the net introduction of an effective inversion, the paper presents a twofold important contribution: an effective inversion for few-steps diffusion models, which facilitates effective editing on these models. + +A limitation of ReNoise is the model-specific hyperparameter tuning required for Edit Enhancement and Noise Correction. While these hyperparameters remain stable for a given model, they may vary across models, and tuning them is necessary to achieve high reconstruction quality while maintaining editability. + +While ReNoise demonstrates the potential for editing few-step diffusion models, more extensive testing with advanced editing methods is needed. It is worth noting that no such editing has been demonstrated for the few-step diffusion models. We believe and hope that our ReNoise method will pave the way for fast and effective editing methods based on the few-steps models. We also believe that ReNoise can be adapted to the challenging problem of inverting video-diffusion models. + +# References + +1. Numerical Differential Equation Methods, chap. 2, pp. 45-121. John Wiley and Sons, Ltd (2003). https://doi.org/https://doi.org/10.1002/0470868279.ch2, https://onlinelibrary.wiley.com/doi/abs/10.1002/0470868279.ch2 +2. Abdal, R., Qin, Y., Wonka, P.: Image2stylegan: How to embed images into the stylegan latent space? In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE (Oct 2019). https://doi.org/10.1109/iccv.2019.00453, http://dx.doi.org/10.1109/ICCV.2019.00453 +3. Abdal, R., Qin, Y., Wonka, P.: Image2stylegan++: How to edit the embedded images? In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (Jun 2020). https://doi.org/10.1109/cvpr42600.2020.00832, http://dx.doi.org/10.1109/cvpr42600.2020.00832 +4. Alaluf, Y., Garibi, D., Patashnik, O., Averbuch-Elor, H., Cohen-Or, D.: Cross-image attention for zero-shot appearance transfer (2023) +5. Alaluf, Y., Patashnik, O., Cohen-Or, D.: Restyle: A residual-based stylegan encoder via iterative refinement. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE (Oct 2021). https://doi.org/10.1109/iccv48922.2021.00664, http://dx.doi.org/10.1109/iccv48922.2021.00664 +6. Alaluf, Y., Tov, O., Mokady, R., Gal, R., Bermano, A.H.: Hyperstyle: Stylegan inversion with hypernetworks for real image editing (2021) +7. Avrahami, O., Fried, O., Lischinski, D.: Blended latent diffusion. arXiv preprint arXiv:2206.02779 (2022) +8. Avrahami, O., Lischinski, D., Fried, O.: Blended diffusion for text-driven editing of natural images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18208-18218 (2022) +9. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions (2023) +0. Burden, R., Faires, J., Burden, A.: Numerical Analysis. Cengage Learning (2015) +1. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. ArXiv abs/2210.11427 (2022) +2. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis (2021) +3. Dinh, T.M., Tran, A.T., Nguyen, R., Hua, B.S.: Hyperinverter: Improving stylegan inversion via hypernetwork. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +4. Epstein, D., Jabri, A., Poole, B., Efros, A.A., Holynski, A.: Diffusion self-guidance for controllable image generation (2023) +5. Ge, S., Park, T., Zhu, J.Y., Huang, J.B.: Expressive text-to-image generation with rich text. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023) +6. Han, L., Wen, S., Chen, Q., Zhang, Z., Song, K., Ren, M., Gao, R., Stathopoulos, A., He, X., Chen, Y., Liu, D., Zhangli, Q., Jiang, J., Xia, Z., Srivastava, A., Metaxas, D.: Improving tuning-free real image editing with proximal guidance (2023) +7. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control (2022) +8. Hertz, A., Voynov, A., Fruchter, S., Cohen-Or, D.: Style aligned image generation via shared attention (2023) +19. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models (2020) +20. Ho, J., Salimans, T.: Classifier-free diffusion guidance (2022) + +21. Huberman-Spiegelglas, I., Kulikov, V., Michaeli, T.: An edit friendly ddpm noise space: Inversion and manipulations (2023) +22. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems 35, 26565-26577 (2022) +23. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Magic: Text-based real image editing with diffusion models. In: Conference on Computer Vision and Pattern Recognition 2023 (2023) +24. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023) +25. Lin, T., Maire, M., Belongie, S.J., Bourdev, L.D., Girshick, R.B., Hays, J., Perona, P., Ramanan, D., Doll'a r, P., Zitnick, C.L.: Microsoft COCO: common objects in context. CoRR abs/1405.0312 (2014) +26. Luo, S., Tan, Y., Huang, L., Li, J., Zhao, H.: Latent consistency models: Synthesizing high-resolution images with few-step inference (2023) +27. Luo, S., Tan, Y., Patil, S., Gu, D., von Platen, P., Passos, A., Huang, L., Li, J., Zhao, H.: Lcm-lora: A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556 (2023) +28. Meiri, B., Samuel, D., Darshan, N., Chechik, G., Avidan, S., Ben-Ari, R.: Fixed-point inversion for text-to-image diffusion models (2023) +29. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations (2022) +30. Miyake, D., Iohara, A., Saito, Y., Tanaka, T.: Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models (2023) +31. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models (2022) +32. Pan, Z., Gherardi, R., Xie, X., Huang, S.: Effective real image editing with accelerated iterative diffusion inversion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15912-15921 (October 2023) +33. Parmar, G., Kumar Singh, K., Zhang, R., Li, Y., Lu, J., Zhu, J.Y.: Zero-shot image-to-image translation. In: Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Proceedings. SIGGRAPH '23, ACM (Jul 2023). https://doi.org/10.1145/3588432.3591513, http://dx.doi.org/10.1145/3588432.3591513 +34. Parmar, G., Li, Y., Lu, J., Zhang, R., Zhu, J.Y., Singh, K.K.: Spatially-adaptive multilayer selection for gan inversion and editing. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (Jun 2022). https://doi.org/10.1109/cvpr52688.2022.01111, http://dx.doi.org/10.1109/CVPR52688.2022.01111 +35. Patashnik, O., Garibi, D., Azuri, I., Averbuch-Elor, H., Cohen-Or, D.: Localizing object-level shape variations with text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023) +36. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis (2023) +37. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents (2022) +38. Richardson, E., Alaluf, Y., Patashnik, O., Nitzan, Y., Azar, Y., Shapiro, S., Cohen-Or, D.: Encoding in style: a stylegan encoder for image-to-image translation. In: + +IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2021) +39. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2022) +40. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Ghasemipour, S.K.S., Ayan, B.K., Mahdavi, S.S., Lopes, R.G., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic text-to-image diffusion models with deep language understanding (2022) +41. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models. In: International Conference on Learning Representations (2021) +42. Sauer, A., Lorenz, D., Blattmann, A., Rombach, R.: Adversarial diffusion distillation (2023) +43. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models (2022) +44. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models (2023) +45. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: International Conference on Learning Representations (2020) +46. Tov, O., Alaluf, Y., Nitzan, Y., Patashnik, O., Cohen-Or, D.: Designing an encoder for stylegan image manipulation. ACM Trans. Graph. 40(4) (jul 2021). https://doi.org/10.1145/3450626.3459838 +47. Tumanyan, N., Geyer, M., Bagon, S., Dekel, T.: Plug-and-play diffusion features for text-driven image-to-image translation pp. 1921-1930 (June 2023) +48. Voynov, A., Hertz, A., Arar, M., Fruchter, S., Cohen-Or, D.: Anylens: A generative diffusion model with any rendering lens (2023) +49. Wallace, B., Gokul, A., Naik, N.: Edict: Exact diffusion inversion via coupled transformations. arXiv preprint arXiv:2211.12446 (2022) +50. Wu, C.H., De la Torre, F.: Unifying diffusion models' latent space, with applications to cyclediffusion and guidance. arXiv preprint arXiv:2210.05559 (2022) +51. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018) +52. Zhu, J., Shen, Y., Zhao, D., Zhou, B.: In-domain gan inversion for real image editing. In: Proceedings of European Conference on Computer Vision (ECCV) (2020) +53. Zhu, J.Y., Krahenbihl, P., Shechtman, E., Efros, A.A.: Generative visual manipulation on the natural image manifold. In: Proceedings of European Conference on Computer Vision (ECCV) (2016) +54. Zhu, P., Abdal, R., Qin, Y., Femiani, J., Wonka, P.: Improved stylegan embedding: Where are the good latents? (2020) \ No newline at end of file diff --git a/2024/ReNoise_ Real Image Inversion Through Iterative Noising/images.zip b/2024/ReNoise_ Real Image Inversion Through Iterative Noising/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8403c57645db3c1e6445afc5aef2ad9aa60ec969 --- /dev/null +++ b/2024/ReNoise_ Real Image Inversion Through Iterative Noising/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0acf75f60b0d9399115d0816d4624f86778b880bf42291a68a993f29eda4bfc1 +size 599427 diff --git a/2024/ReNoise_ Real Image Inversion Through Iterative Noising/layout.json b/2024/ReNoise_ Real Image Inversion Through Iterative Noising/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e164fbfe5536be9c19f80b76cd88b2ce959ea71a --- /dev/null +++ b/2024/ReNoise_ Real Image Inversion Through Iterative Noising/layout.json @@ -0,0 +1,11271 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "type": "text", + "content": "ReNoise: Real Image Inversion Through Iterative Noising" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 133, + 167, + 479, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 167, + 479, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 167, + 479, + 193 + ], + "type": "text", + "content": "Daniel Garibi1, Or Patashnik1, Andrey Voynov2, Hadar Averbuch-Elor1, and Daniel Cohen-Or1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 262, + 201, + 351, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 201, + 351, + 213 + ], + "spans": [ + { + "bbox": [ + 262, + 201, + 351, + 213 + ], + "type": "text", + "content": "1 Tel Aviv University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 268, + 213, + 345, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 213, + 345, + 224 + ], + "spans": [ + { + "bbox": [ + 268, + 213, + 345, + 224 + ], + "type": "text", + "content": "2 Google Research" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 135, + 250, + 479, + 376 + ], + "blocks": [ + { + "bbox": [ + 135, + 250, + 479, + 376 + ], + "lines": [ + { + "bbox": [ + 135, + 250, + 479, + 376 + ], + "spans": [ + { + "bbox": [ + 135, + 250, + 479, + 376 + ], + "type": "image", + "image_path": "41c090579a293cc48ee4ef721b81c4663f6dcb6c9d3cd0bd9c28676762b66240.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 379, + 482, + 442 + ], + "lines": [ + { + "bbox": [ + 130, + 379, + 482, + 442 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 482, + 442 + ], + "type": "text", + "content": "Fig. 1: Our ReNoise inversion technique can be applied to various diffusion models, including recent few-step ones. This figure illustrates the performance of our method with SDXL Turbo and LCM models, showing its effectiveness compared to DDIM inversion. Additionally, we demonstrate that the quality of our inversions allows prompt-driven editing. As illustrated on the right, our approach also allows for prompt-driven image edits." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 442, + 453, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 442, + 453, + 650 + ], + "spans": [ + { + "bbox": [ + 159, + 442, + 453, + 650 + ], + "type": "text", + "content": "Abstract. Recent advancements in text-guided diffusion models have unlocked powerful image manipulation capabilities. However, applying these methods to real images necessitates the inversion of the images into the domain of the pretrained diffusion model. Achieving faithful inversion remains a challenge, particularly for more recent models trained to generate images with a small number of denoising steps. In this work, we introduce an inversion method with a high quality-to-operation ratio, enhancing reconstruction accuracy without increasing the number of operations. Building on reversing the diffusion sampling process, our method employs an iterative renoising mechanism at each inversion sampling step. This mechanism refines the approximation of a predicted point along the forward diffusion trajectory, by iteratively applying the pretrained diffusion model, and averaging these predictions. We evaluate the performance of our ReNoise technique using various sampling algorithms and models, including recent accelerated diffusion models. Through comprehensive evaluations and comparisons, we show its effectiveness in terms of both accuracy and speed. Furthermore, we confirm that our method preserves editability by demonstrating text-driven image editing on real images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 654, + 395, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 654, + 395, + 665 + ], + "spans": [ + { + "bbox": [ + 160, + 654, + 395, + 665 + ], + "type": "text", + "content": "Keywords: Diffusion Models " + }, + { + "bbox": [ + 160, + 654, + 395, + 665 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 654, + 395, + 665 + ], + "type": "text", + "content": " Image Editing " + }, + { + "bbox": [ + 160, + 654, + 395, + 665 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 654, + 395, + 665 + ], + "type": "text", + "content": " Inversion" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 302, + 173 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 302, + 173 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 302, + 173 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 302, + 173 + ], + "type": "image", + "image_path": "396a494924ea216dcc23a0b15469fcfd96f59ecf191bad8d9535bf9ae0b534bc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 180, + 309, + 289 + ], + "lines": [ + { + "bbox": [ + 132, + 180, + 309, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 180, + 309, + 289 + ], + "type": "text", + "content": "Fig.2: The diffusion process samples a Gaussian noise and iteratively denoises it until reaching the data distribution. At each point along the denoising trajectory, the model predicts a direction, " + }, + { + "bbox": [ + 132, + 180, + 309, + 289 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_t)" + }, + { + "bbox": [ + 132, + 180, + 309, + 289 + ], + "type": "text", + "content": ", to step to the next point along the trajectory. To invert a given image from the distribution, the direction from " + }, + { + "bbox": [ + 132, + 180, + 309, + 289 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 132, + 180, + 309, + 289 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 180, + 309, + 289 + ], + "type": "inline_equation", + "content": "z_{t + 1}" + }, + { + "bbox": [ + 132, + 180, + 309, + 289 + ], + "type": "text", + "content": " is approximated with the inverse of the direction from " + }, + { + "bbox": [ + 132, + 180, + 309, + 289 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 132, + 180, + 309, + 289 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 180, + 309, + 289 + ], + "type": "inline_equation", + "content": "z_{t - 1}" + }, + { + "bbox": [ + 132, + 180, + 309, + 289 + ], + "type": "text", + "content": " denoted by a dotted blue line." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 320, + 122, + 486, + 256 + ], + "blocks": [ + { + "bbox": [ + 320, + 122, + 486, + 256 + ], + "lines": [ + { + "bbox": [ + 320, + 122, + 486, + 256 + ], + "spans": [ + { + "bbox": [ + 320, + 122, + 486, + 256 + ], + "type": "image", + "image_path": "c90c3ee96a4c920e3787eedde0a66016d70d313e0d0b559b867869e1e6f5ca13.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 257, + 482, + 290 + ], + "lines": [ + { + "bbox": [ + 318, + 257, + 482, + 290 + ], + "spans": [ + { + "bbox": [ + 318, + 257, + 482, + 290 + ], + "type": "text", + "content": "Fig.3: Comparing reconstruction results of plain DDIM Inv. with SDXL to DDIM Inv. with one renoising iteration." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 297, + 230, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 230, + 309 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 230, + 309 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "text", + "content": "Large-scale text-to-image diffusion models have revolutionized the field of image synthesis [19,37,39,40]. In particular, many works have shown that these models can be employed for various types of image manipulation [4,7-9,15,17,23,29,33, 35,47]. To edit real images, many of these techniques often require the inversion of the image into the domain of the diffusion model. That is, given a real image " + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "text", + "content": ", one has to find a Gaussian noise " + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "text", + "content": ", such that denoising " + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "text", + "content": " with the pretrained diffusion model reconstructs the given real image " + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "text", + "content": ". The importance of this task for real image manipulation has prompted many efforts aimed at achieving accurate reconstruction [16,21,30,31]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "text", + "content": "The diffusion process consists of a series of denoising steps " + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "inline_equation", + "content": "\\{\\epsilon_{\\theta}(z_t,t)\\}_{t = T}^1" + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "text", + "content": ", which form a trajectory from the Gaussian noise to the model distribution (see Figure 2). Each denoising step is computed by a trained network, typically implemented as a UNet, which predicts " + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "inline_equation", + "content": "z_{t - 1}" + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "text", + "content": " [19]. The output of the model at each step forms a direction from " + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "inline_equation", + "content": "z_{t - 1}" + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "text", + "content": " [45]. These steps are not invertible, in the sense that the model was not trained to predict " + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "inline_equation", + "content": "z_{t - 1}" + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "text", + "content": ". Thus, the problem of inverting a given image is a challenge, and particularly for real images, as they are not necessarily in the model distribution (see Figure 3)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 526, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 526, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 526, + 482, + 668 + ], + "type": "text", + "content": "In this paper, we present an inversion method with a high quality-to-operation ratio, which achieves superior reconstruction accuracy for the same number of UNet operations. We build upon the commonly used approach of reversing the diffusion sampling process, which is based on the linearity assumption that the direction from " + }, + { + "bbox": [ + 130, + 526, + 482, + 668 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 526, + 482, + 668 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 526, + 482, + 668 + ], + "type": "inline_equation", + "content": "z_{t+1}" + }, + { + "bbox": [ + 130, + 526, + 482, + 668 + ], + "type": "text", + "content": " can be approximated by the negation of the direction from " + }, + { + "bbox": [ + 130, + 526, + 482, + 668 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 526, + 482, + 668 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 526, + 482, + 668 + ], + "type": "inline_equation", + "content": "z_{t-1}" + }, + { + "bbox": [ + 130, + 526, + 482, + 668 + ], + "type": "text", + "content": " [12, 43] (see Figure 2). To enhance this approximation, we employ the fixed-point iteration methodology [10]. Specifically, given " + }, + { + "bbox": [ + 130, + 526, + 482, + 668 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 526, + 482, + 668 + ], + "type": "text", + "content": ", we begin by using the common approximation to get an initial estimate for " + }, + { + "bbox": [ + 130, + 526, + 482, + 668 + ], + "type": "inline_equation", + "content": "z_{t+1}" + }, + { + "bbox": [ + 130, + 526, + 482, + 668 + ], + "type": "text", + "content": ", denoted by " + }, + { + "bbox": [ + 130, + 526, + 482, + 668 + ], + "type": "inline_equation", + "content": "z_{t+1}^{(0)}" + }, + { + "bbox": [ + 130, + 526, + 482, + 668 + ], + "type": "text", + "content": ". Then, we iteratively renoise " + }, + { + "bbox": [ + 130, + 526, + 482, + 668 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 526, + 482, + 668 + ], + "type": "text", + "content": ", following the direction implied by " + }, + { + "bbox": [ + 130, + 526, + 482, + 668 + ], + "type": "inline_equation", + "content": "z_{t+1}^{(k)}" + }, + { + "bbox": [ + 130, + 526, + 482, + 668 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 130, + 526, + 482, + 668 + ], + "type": "inline_equation", + "content": "z_{t+1}^{(k+1)}" + }, + { + "bbox": [ + 130, + 526, + 482, + 668 + ], + "type": "text", + "content": ". After repeating this ronoising process several times, we apply an averaging on " + }, + { + "bbox": [ + 130, + 526, + 482, + 668 + ], + "type": "inline_equation", + "content": "z_{t+1}^{(k)}" + }, + { + "bbox": [ + 130, + 526, + 482, + 668 + ], + "type": "text", + "content": " to form a more accurate direction from " + }, + { + "bbox": [ + 130, + 526, + 482, + 668 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 526, + 482, + 668 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 526, + 482, + 668 + ], + "type": "inline_equation", + "content": "z_{t+1}" + }, + { + "bbox": [ + 130, + 526, + 482, + 668 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "text", + "content": "D. Garibi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "We show that this approach enables longer strides along the inversion trajectory while improving image reconstruction. Therefore, our method can also be effective with diffusion models trained to generate images using a small number of denoising steps [27, 42]. Furthermore, despite the need to repeatedly renoise in each step of the inversion process, the longer strides lead to a more favorable tradeoff of UNet operations for reconstruction quality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 489, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 489, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 489, + 260 + ], + "type": "text", + "content": "Through extensive experiments, we demonstrate the effectiveness of our method in both image reconstruction and inversion speed. We validate the versatility of our approach across different samplers and models, including recent time-distilled diffusion models (e.g., SDXL-Turbo [42]). Importantly, we demonstrate that the editability of the inversion achieved by our method allows a wide range of text-driven image manipulations (see Figure 1)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 263, + 237, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 263, + 237, + 276 + ], + "spans": [ + { + "bbox": [ + 131, + 263, + 237, + 276 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 281, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 449 + ], + "type": "text", + "content": "Image Editing via Diffusion Models Recent advancements in diffusion models [12, 19] have resulted in unprecedented diversity and fidelity in visual content creation guided by free-form text prompts [36, 37, 39, 40]. Text-to-image models do not directly support text-guided image editing. Therefore, harnessing the power of these models for image editing is a significant research area and many methods have utilized these models for different types of image editing [4, 7-9, 11, 14, 15, 17, 18, 23, 29, 33, 35, 47, 48]. A common approach among these methods requires inversion [21, 31, 43, 49] to edit real images, i.e., obtaining a latent code " + }, + { + "bbox": [ + 130, + 281, + 482, + 449 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 130, + 281, + 482, + 449 + ], + "type": "text", + "content": " such that denoising it with the pretrained diffusion model returns the original image. Specifically, in this approach two backward processes are done simultaneously using " + }, + { + "bbox": [ + 130, + 281, + 482, + 449 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 130, + 281, + 482, + 449 + ], + "type": "text", + "content": ". One of the processes reconstructs the image using the original prompt, while the second one injects features from the first process (e.g., attention maps) to preserve some properties of the original image while manipulating other aspects of it." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 450, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 545 + ], + "type": "text", + "content": "Inversion in Diffusion Models Initial efforts in image inversion for real image editing focused on GANs [2,3,5,6,13,34,38,46,52-54]. The advancements in diffusion models, and in diffusion-based image editing in particular have recently prompted works studying the inversion of a diffusion-based denoising process. This inversion depends on the sampler algorithm used during inference, which can be deterministic [43] or non-deterministic [19, 22]. Inversion methods can be accordingly categorized into two: methods that are suitable for deterministic sampling, and methods suitable for non-deterministic sampling." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "Methods that approach the deterministic inversion commonly rely on the DDIM sampling method [43], and build upon DDIM inversion [12, 43]. Mokady et al. [31] observed that the use of classifier-free guidance during inference magnifies the accumulated error of DDIM inversion and therefore leads to poor reconstruction. Following this observation, several works [16, 30, 31] focused on solving this issue by replacing the null text token with a different embedding, which is found after an optimization process or by a closed solution. However, excluding [31] which requires a lengthy optimization, these methods are limited by the reconstruction accuracy of DDIM inversion, which can be poor, especially when a small number of denoising steps is done. In our work, we present a method" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 214, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 214, + 91, + 447, + 103 + ], + "type": "text", + "content": "ReNoise: Real Image Inversion Through Iterative Noising" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "that improves the reconstruction quality of DDIM inversion and therefore can be integrated with methods that build on it." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": "Another line of work [21, 50] tackles the inversion of DDPM sampler [19]. In these works [21, 50], instead of inverting the image into an initial noise " + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": ", a series of noises " + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\{z_{T},\\epsilon_{T},\\dots,\\epsilon_{1}\\}" + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": " is obtained. The definition of this noises series ensures that generating an image with it returns the original input image. However, these methods require a large number of inversion and denoising steps to allow image editing. Applying these methods with an insufficient number of steps leads to too much information encoded in " + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\{\\epsilon_T,\\dots,\\epsilon_1\\}" + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": " which limits the ability to edit the generated image. As shall be shown, The editability issue of these methods is particularly evident in few-steps models [26, 27, 42]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 249, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 482, + 344 + ], + "type": "text", + "content": "Most relevant to our work, two recent inversion methods [28,32] also use the fixed-point iteration technique. Specifically, they improve the reconstruction accuracy of DDIM inversion [43] with Stable Diffusion [39] without introducing a significant computational overhead. In our work, we focus on the problem of real image inversion for recently introduced few-step diffusion models, where the difficulties encountered by previous methods are amplified. Furthermore, we show that our inversion method successfully works with various models and different samplers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 350, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 460 + ], + "type": "text", + "content": "Few Steps Models Recently, new methods [26, 27, 41, 42, 44] that fine-tune text-to-image diffusion models enabled a significant reduction of the number of steps needed for high-quality image generation. While standard diffusion models typically require 50 denoising steps to generate high-quality images, recent accelerated models achieve high-quality synthesis with 1-4 steps only. These new methods pave the way for interactive editing workflows. However, as we show in this paper, using current methods for the inversion of an image with a small number of steps degrades the reconstruction quality in terms of accuracy [12, 43] or editability [21, 50]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 466, + 203, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 466, + 203, + 479 + ], + "spans": [ + { + "bbox": [ + 131, + 466, + 203, + 479 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 495, + 253, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 495, + 253, + 506 + ], + "spans": [ + { + "bbox": [ + 131, + 495, + 253, + 506 + ], + "type": "text", + "content": "3.1 ReNoise Inversion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "text", + "content": "Reversing the Sampler Samplers play a critical role in the diffusion-based image synthesis process. They define the noising and denoising diffusion processes and influence the processes' trajectories and quality of the generated images. While different samplers share the same pre-trained UNet model (denoted by " + }, + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "text", + "content": ") as their backbone, their sampling approaches diverge, leading to nuanced differences in output. The goal of the denoising sampler is to predict the latent code at the previous noise level, " + }, + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "inline_equation", + "content": "z_{t-1}" + }, + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "text", + "content": ", based on the current noisy data " + }, + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "text", + "content": ", the pretrained UNet model, and a sampled noise, " + }, + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "inline_equation", + "content": "\\epsilon_{t}" + }, + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "text", + "content": ". Various first-order denoising sampling algorithms adhere to the form:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 235, + 626, + 481, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 626, + 481, + 638 + ], + "spans": [ + { + "bbox": [ + 235, + 626, + 481, + 638 + ], + "type": "interline_equation", + "content": "z _ {t - 1} = \\phi_ {t} z _ {t} + \\psi_ {t} \\epsilon_ {\\theta} (z _ {t}, t, c) + \\rho_ {t} \\epsilon_ {t}, \\tag {1}", + "image_path": "d1deb1a0155e649377799d0edb278fa894a7c771bac5697393c14349e92289a0.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": " represents a text embedding condition, and " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\phi_t, \\psi_t" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\rho_t" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": " denote sampler parameters. At each step, these parameters control the extent to which the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "text", + "content": "D. Garibi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "previous noise is removed " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "(\\phi_t)" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": ", the significance assigned to the predicted noise from the UNet " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "(\\psi_t)" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": ", and the given weight to the additional noise introduced " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "(\\rho_t)" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 163 + ], + "type": "text", + "content": "A given image " + }, + { + "bbox": [ + 130, + 140, + 480, + 163 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 130, + 140, + 480, + 163 + ], + "type": "text", + "content": " can be inverted by reformulating Equation 1 and applying it iteratively:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 239, + 158, + 480, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 158, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 239, + 158, + 480, + 182 + ], + "type": "interline_equation", + "content": "z _ {t} = \\frac {z _ {t - 1} - \\psi_ {t} \\epsilon_ {\\theta} \\left(z _ {t} , t , c\\right) - \\rho_ {t} \\epsilon_ {t}}{\\phi_ {t}}, \\tag {2}", + "image_path": "944338255e6cdcbdb3146857de449440f58305ff8031713211f5d458973a6179.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 183, + 479, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 479, + 255 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 479, + 255 + ], + "type": "text", + "content": "where for non-deterministic samplers, a series of random noises " + }, + { + "bbox": [ + 130, + 183, + 479, + 255 + ], + "type": "inline_equation", + "content": "\\{\\epsilon_t\\}_{t=1}^T" + }, + { + "bbox": [ + 130, + 183, + 479, + 255 + ], + "type": "text", + "content": " is sampled and used during both inversion and image generation processes. However, directly computing " + }, + { + "bbox": [ + 130, + 183, + 479, + 255 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 130, + 183, + 479, + 255 + ], + "type": "text", + "content": " from Equation 2 is infeasible since it relies on " + }, + { + "bbox": [ + 130, + 183, + 479, + 255 + ], + "type": "inline_equation", + "content": "\\epsilon_\\theta(z_t, t, c)" + }, + { + "bbox": [ + 130, + 183, + 479, + 255 + ], + "type": "text", + "content": " which, in turn, depends on " + }, + { + "bbox": [ + 130, + 183, + 479, + 255 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 130, + 183, + 479, + 255 + ], + "type": "text", + "content": ", creating a circular dependency. To solve this implicit function, Dhariwal et al. [12] propose using the approximation " + }, + { + "bbox": [ + 130, + 183, + 479, + 255 + ], + "type": "inline_equation", + "content": "\\epsilon_\\theta(z_t, t, c) \\approx \\epsilon_\\theta(z_{t-1}, t, c)" + }, + { + "bbox": [ + 130, + 183, + 479, + 255 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 230, + 250, + 480, + 274 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 250, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 230, + 250, + 480, + 274 + ], + "type": "interline_equation", + "content": "z _ {t} ^ {(1)} = \\frac {z _ {t - 1} - \\psi_ {t} \\epsilon_ {\\theta} \\left(z _ {t - 1} , t , c\\right) - \\rho_ {t} \\epsilon_ {t}}{\\phi_ {t}}. \\tag {3}", + "image_path": "d3c5bd2faac4655821cf42ad9d7437f5f47333b0c8b15ea52d0fecb3cd1e4e16.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 275, + 480, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 275, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 480, + 384 + ], + "type": "text", + "content": "This method has several limitations. First, the assumption underlying the approximation used in [12] is that the number of inversion steps is large enough, implying a trajectory close to linear. This assumption restricts the applicability of this inversion method in interactive image editing with recent few-step diffusion models [26, 27, 42, 44], as the inversion process would take significantly longer than inference. Second, this method struggles to produce accurate reconstructions in certain cases, such as highly detailed images or images with large smooth regions, see Figure 3. Moreover, we observe that this inversion method is sensitive to the prompt " + }, + { + "bbox": [ + 130, + 275, + 480, + 384 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 275, + 480, + 384 + ], + "type": "text", + "content": " and may yield poor results for certain prompts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 390, + 479, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 479, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 479, + 461 + ], + "type": "text", + "content": "ReNoise In a successful inversion trajectory, the direction from " + }, + { + "bbox": [ + 130, + 390, + 479, + 461 + ], + "type": "inline_equation", + "content": "z_{t-1}" + }, + { + "bbox": [ + 130, + 390, + 479, + 461 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 390, + 479, + 461 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 130, + 390, + 479, + 461 + ], + "type": "text", + "content": " aligns with the direction from " + }, + { + "bbox": [ + 130, + 390, + 479, + 461 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 130, + 390, + 479, + 461 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 390, + 479, + 461 + ], + "type": "inline_equation", + "content": "z_{t-1}" + }, + { + "bbox": [ + 130, + 390, + 479, + 461 + ], + "type": "text", + "content": " in the denoising trajectory. To achieve this, we aim to improve the approximation of " + }, + { + "bbox": [ + 130, + 390, + 479, + 461 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_t,t,c)" + }, + { + "bbox": [ + 130, + 390, + 479, + 461 + ], + "type": "text", + "content": " in Eq. 2 compared to the one used in [12]. Building on the fixed-point iteration technique [10], our approach better estimates the instance of " + }, + { + "bbox": [ + 130, + 390, + 479, + 461 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 130, + 390, + 479, + 461 + ], + "type": "text", + "content": " that is inputted to the UNet, rather than relying on " + }, + { + "bbox": [ + 130, + 390, + 479, + 461 + ], + "type": "inline_equation", + "content": "z_{t-1}" + }, + { + "bbox": [ + 130, + 390, + 479, + 461 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 462, + 480, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 480, + 587 + ], + "type": "text", + "content": "Intuitively, we utilize the observation that " + }, + { + "bbox": [ + 130, + 462, + 480, + 587 + ], + "type": "inline_equation", + "content": "z_{t}^{(1)}" + }, + { + "bbox": [ + 130, + 462, + 480, + 587 + ], + "type": "text", + "content": " (from Eq. 3) offers a more precise estimate of " + }, + { + "bbox": [ + 130, + 462, + 480, + 587 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 462, + 480, + 587 + ], + "type": "text", + "content": " compared to " + }, + { + "bbox": [ + 130, + 462, + 480, + 587 + ], + "type": "inline_equation", + "content": "z_{t-1}" + }, + { + "bbox": [ + 130, + 462, + 480, + 587 + ], + "type": "text", + "content": ". Therefore, employing " + }, + { + "bbox": [ + 130, + 462, + 480, + 587 + ], + "type": "inline_equation", + "content": "z_{t}^{(1)}" + }, + { + "bbox": [ + 130, + 462, + 480, + 587 + ], + "type": "text", + "content": " as the input to the UNet is likely to yield a more accurate direction, thus contributing to reducing the overall error in the inversion step. We illustrate this observation in Figure 5. Iterating this process generates a series of estimations for " + }, + { + "bbox": [ + 130, + 462, + 480, + 587 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 462, + 480, + 587 + ], + "type": "text", + "content": ", denoted by " + }, + { + "bbox": [ + 130, + 462, + 480, + 587 + ], + "type": "inline_equation", + "content": "\\{z_{t}^{(k)}\\}_{k=1}^{\\mathcal{K}+1}" + }, + { + "bbox": [ + 130, + 462, + 480, + 587 + ], + "type": "text", + "content": ". While the fixed-point iteration technique [10] does not guarantee convergence of this series in the general case, in Section 4, we empirically show that convergence holds in our setting. However, as the convergence is not monotonic, we refine our prediction of " + }, + { + "bbox": [ + 130, + 462, + 480, + 587 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 462, + 480, + 587 + ], + "type": "text", + "content": " by averaging several " + }, + { + "bbox": [ + 130, + 462, + 480, + 587 + ], + "type": "inline_equation", + "content": "\\{z_{t}^{(k)}\\}" + }, + { + "bbox": [ + 130, + 462, + 480, + 587 + ], + "type": "text", + "content": ", thus considering more than a single estimation of " + }, + { + "bbox": [ + 130, + 462, + 480, + 587 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 462, + 480, + 587 + ], + "type": "text", + "content": ". See Figure 6 for an intuitive illustration." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 588, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 588, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 588, + 481, + 665 + ], + "type": "text", + "content": "In more detail, our method iteratively computes estimations of " + }, + { + "bbox": [ + 130, + 588, + 481, + 665 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 588, + 481, + 665 + ], + "type": "text", + "content": " during each inversion step " + }, + { + "bbox": [ + 130, + 588, + 481, + 665 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 588, + 481, + 665 + ], + "type": "text", + "content": " by renoising the noisy latent " + }, + { + "bbox": [ + 130, + 588, + 481, + 665 + ], + "type": "inline_equation", + "content": "z_{t-1}" + }, + { + "bbox": [ + 130, + 588, + 481, + 665 + ], + "type": "text", + "content": " multiple times, each with a different noise prediction (see Figure 4). Beginning with " + }, + { + "bbox": [ + 130, + 588, + 481, + 665 + ], + "type": "inline_equation", + "content": "z_{t}^{(1)}" + }, + { + "bbox": [ + 130, + 588, + 481, + 665 + ], + "type": "text", + "content": ", in the " + }, + { + "bbox": [ + 130, + 588, + 481, + 665 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 588, + 481, + 665 + ], + "type": "text", + "content": "-th renoising iteration, the input to the UNet is the result of the previous iteration, " + }, + { + "bbox": [ + 130, + 588, + 481, + 665 + ], + "type": "inline_equation", + "content": "z_{t}^{(k)}" + }, + { + "bbox": [ + 130, + 588, + 481, + 665 + ], + "type": "text", + "content": ". Then, " + }, + { + "bbox": [ + 130, + 588, + 481, + 665 + ], + "type": "inline_equation", + "content": "z_{t}^{(k+1)}" + }, + { + "bbox": [ + 130, + 588, + 481, + 665 + ], + "type": "text", + "content": " is calculated using the inverted sampler while maintaining " + }, + { + "bbox": [ + 130, + 588, + 481, + 665 + ], + "type": "inline_equation", + "content": "z_{t-1}" + }, + { + "bbox": [ + 130, + 588, + 481, + 665 + ], + "type": "text", + "content": " as the starting point of the step. After " + }, + { + "bbox": [ + 130, + 588, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{K}" + }, + { + "bbox": [ + 130, + 588, + 481, + 665 + ], + "type": "text", + "content": " renoising iterations, we obtain a set of" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 214, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 214, + 91, + 448, + 103 + ], + "type": "text", + "content": "ReNoise: Real Image Inversion Through Iterative Noising" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 481, + 242 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 242 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 242 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 242 + ], + "type": "image", + "image_path": "5742e3c782e5d830b04cea56a12460576f346867f29852aa2c51d010d3e236e6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 243, + 482, + 334 + ], + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 334 + ], + "type": "text", + "content": "Fig. 4: Method overview. Given an image " + }, + { + "bbox": [ + 130, + 243, + 482, + 334 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 130, + 243, + 482, + 334 + ], + "type": "text", + "content": ", we iteratively compute " + }, + { + "bbox": [ + 130, + 243, + 482, + 334 + ], + "type": "inline_equation", + "content": "z_1, \\dots, z_T" + }, + { + "bbox": [ + 130, + 243, + 482, + 334 + ], + "type": "text", + "content": ", where each " + }, + { + "bbox": [ + 130, + 243, + 482, + 334 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 130, + 243, + 482, + 334 + ], + "type": "text", + "content": " is calculated from " + }, + { + "bbox": [ + 130, + 243, + 482, + 334 + ], + "type": "inline_equation", + "content": "z_{t-1}" + }, + { + "bbox": [ + 130, + 243, + 482, + 334 + ], + "type": "text", + "content": ". At each time step, we apply the UNet " + }, + { + "bbox": [ + 130, + 243, + 482, + 334 + ], + "type": "inline_equation", + "content": "(\\epsilon_\\theta) \\mathcal{K} + 1" + }, + { + "bbox": [ + 130, + 243, + 482, + 334 + ], + "type": "text", + "content": " times, each using a better approximation of " + }, + { + "bbox": [ + 130, + 243, + 482, + 334 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 130, + 243, + 482, + 334 + ], + "type": "text", + "content": " as the input. The initial approximation is " + }, + { + "bbox": [ + 130, + 243, + 482, + 334 + ], + "type": "inline_equation", + "content": "z_{t-1}" + }, + { + "bbox": [ + 130, + 243, + 482, + 334 + ], + "type": "text", + "content": ". The next one, " + }, + { + "bbox": [ + 130, + 243, + 482, + 334 + ], + "type": "inline_equation", + "content": "z_t^{(1)}" + }, + { + "bbox": [ + 130, + 243, + 482, + 334 + ], + "type": "text", + "content": ", is the result of the reversed sampler step (i.e., DDIM). The reversed step begins at " + }, + { + "bbox": [ + 130, + 243, + 482, + 334 + ], + "type": "inline_equation", + "content": "z_{t-1}" + }, + { + "bbox": [ + 130, + 243, + 482, + 334 + ], + "type": "text", + "content": " and follows the direction of " + }, + { + "bbox": [ + 130, + 243, + 482, + 334 + ], + "type": "inline_equation", + "content": "\\epsilon_\\theta(z_{t-1}, t)" + }, + { + "bbox": [ + 130, + 243, + 482, + 334 + ], + "type": "text", + "content": ". At the " + }, + { + "bbox": [ + 130, + 243, + 482, + 334 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 243, + 482, + 334 + ], + "type": "text", + "content": " renoising iteration, " + }, + { + "bbox": [ + 130, + 243, + 482, + 334 + ], + "type": "inline_equation", + "content": "z_t^{(k)}" + }, + { + "bbox": [ + 130, + 243, + 482, + 334 + ], + "type": "text", + "content": " is the input to the UNet, and we obtain a better " + }, + { + "bbox": [ + 130, + 243, + 482, + 334 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 130, + 243, + 482, + 334 + ], + "type": "text", + "content": " approximation. For the last iterations, we optimize " + }, + { + "bbox": [ + 130, + 243, + 482, + 334 + ], + "type": "inline_equation", + "content": "\\epsilon_\\theta(z_t^{(k)}, t)" + }, + { + "bbox": [ + 130, + 243, + 482, + 334 + ], + "type": "text", + "content": " to increase editability. As the final denoising direction, we use the average of the UNet predictions of the last few iterations." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 335, + 326, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 326, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 326, + 489 + ], + "type": "text", + "content": "Fig.5: Geometric intuition for ReNoise. At each inversion step, we estimate " + }, + { + "bbox": [ + 130, + 335, + 326, + 489 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 335, + 326, + 489 + ], + "type": "text", + "content": " (marked with a red star) based on " + }, + { + "bbox": [ + 130, + 335, + 326, + 489 + ], + "type": "inline_equation", + "content": "z_{t-1}" + }, + { + "bbox": [ + 130, + 335, + 326, + 489 + ], + "type": "text", + "content": ". The straightforward approach is to use the negated direction of the denoising step from " + }, + { + "bbox": [ + 130, + 335, + 326, + 489 + ], + "type": "inline_equation", + "content": "z_{t-1}" + }, + { + "bbox": [ + 130, + 335, + 326, + 489 + ], + "type": "text", + "content": ", assuming the trajectory is approximately linear. However, this assumption is inaccurate, especially in few-step models, where the size of the steps is large. We use the linearity assumption as an initial estimation and keep improving this estimation. We recalculate the denoising step from the previous estimation (which is closer to the target " + }, + { + "bbox": [ + 130, + 335, + 326, + 489 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 335, + 326, + 489 + ], + "type": "text", + "content": ") and then proceed with its negated direction from " + }, + { + "bbox": [ + 130, + 335, + 326, + 489 + ], + "type": "inline_equation", + "content": "z_{t-1}" + }, + { + "bbox": [ + 130, + 335, + 326, + 489 + ], + "type": "text", + "content": " (see the orange vectors)." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 326, + 354, + 484, + 483 + ], + "blocks": [ + { + "bbox": [ + 326, + 354, + 484, + 483 + ], + "lines": [ + { + "bbox": [ + 326, + 354, + 484, + 483 + ], + "spans": [ + { + "bbox": [ + 326, + 354, + 484, + 483 + ], + "type": "image", + "image_path": "b9e58c72e7338254c2d815309cd5e7fa86ddf0ec2dbf1dca4521e55476645c43.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 493, + 482, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 493, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 493, + 482, + 534 + ], + "type": "text", + "content": "estimations " + }, + { + "bbox": [ + 130, + 493, + 482, + 534 + ], + "type": "inline_equation", + "content": "\\{z_t^{(k)}\\}_{k = 1}^{\\mathcal{K} + 1}" + }, + { + "bbox": [ + 130, + 493, + 482, + 534 + ], + "type": "text", + "content": ". The next point on the inversion trajectory, " + }, + { + "bbox": [ + 130, + 493, + 482, + 534 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 493, + 482, + 534 + ], + "type": "text", + "content": ", is then defined as their weighted average, where " + }, + { + "bbox": [ + 130, + 493, + 482, + 534 + ], + "type": "inline_equation", + "content": "w_{k}" + }, + { + "bbox": [ + 130, + 493, + 482, + 534 + ], + "type": "text", + "content": " is the weight assigned to " + }, + { + "bbox": [ + 130, + 493, + 482, + 534 + ], + "type": "inline_equation", + "content": "z_{t}^{(k)}" + }, + { + "bbox": [ + 130, + 493, + 482, + 534 + ], + "type": "text", + "content": ". For a detailed description of our method, refer to Algorithm 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 538, + 340, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 538, + 340, + 550 + ], + "spans": [ + { + "bbox": [ + 131, + 538, + 340, + 550 + ], + "type": "text", + "content": "3.2 Reconstruction-Editability Tradeoff" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "text", + "content": "Enhance Editability The goal of inversion is to facilitate the editing of real images using a pretrained image generation model. While the renoising approach attains highly accurate reconstruction results, we observe that the resulting " + }, + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "text", + "content": " lacks editability. This phenomenon can be attributed to the reconstruction-editability tradeoff in image generative models [46]. To address this limitation, we incorporate a technique to enhance the editability of our method." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "It has been shown [33] that the noise maps predicted during the inversion process often diverge from the statistical properties of uncorrelated Gaussian white noise, thereby affecting editability. To tackle this challenge, we follow pix2pix-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "text", + "content": "D. Garibi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "zero [33] and regularize the predicted noise at each step, " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_t,t,c)" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": ", using the following loss terms." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "text", + "content": "First, we encourage " + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_t,t,c)" + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "text", + "content": " to follow the same distribution as " + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_t^{\\prime},t,c)" + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "inline_equation", + "content": "z_{t}^{\\prime}" + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "text", + "content": " represents the input image " + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "text", + "content": " with added random noise corresponding to the noise level at timestep " + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "text", + "content": ". We do so by dividing " + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_t,t,c)" + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_t^{\\prime},t,c)" + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "text", + "content": " into small patches (e.g., " + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "inline_equation", + "content": "4\\times 4" + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "text", + "content": "), and computing the KL-divergence between corresponding patches. We denote this loss term by " + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{patch - KL}}" + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "text", + "content": ". Second, we utilize " + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pair}}" + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "text", + "content": " proposed in pix2pix-zero [33], which penalizes correlations between pairs of pixels. We leverage these losses to enhance the editability of our method, and denote the combination of them as " + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{edit}}" + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "text", + "content": ". For any renoising iteration " + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "inline_equation", + "content": "w_{k} > 0" + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "text", + "content": ", we regularize the UNet's prediction " + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_t^{(k)},t,c)" + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{edit}}" + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "text", + "content": " before computing " + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "inline_equation", + "content": "z_{t}^{(k + 1)}" + }, + { + "bbox": [ + 130, + 140, + 482, + 263 + ], + "type": "text", + "content": ". See line 9 in Algorithm 1." + } + ] + } + ], + "index": 3 + }, + { + "type": "code", + "bbox": [ + 132, + 285, + 488, + 495 + ], + "blocks": [ + { + "bbox": [ + 140, + 273, + 288, + 283 + ], + "lines": [ + { + "bbox": [ + 140, + 273, + 288, + 283 + ], + "spans": [ + { + "bbox": [ + 140, + 273, + 288, + 283 + ], + "type": "text", + "content": "Algorithm 1: ReNoise Inversion" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "lines": [ + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "text", + "content": "1 Input: An image " + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "text", + "content": ", number of renoising steps " + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{K}" + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "text", + "content": ", number of inversion steps " + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "text", + "content": ", a series of renoising weights " + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "inline_equation", + "content": "\\{w_{k}\\}_{k=1}^{\\mathcal{K}+1}" + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "text", + "content": ". \n2 Output: A noisy latent " + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "text", + "content": " and set of noises " + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "inline_equation", + "content": "\\{\\epsilon_{t}\\}_{t=1}^{T}" + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "text", + "content": ". \n3 for " + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "inline_equation", + "content": "t = 0,1,\\ldots,T" + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "text", + "content": " do \n4 sample " + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "inline_equation", + "content": "\\epsilon_{t} \\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "text", + "content": " \n5 " + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "inline_equation", + "content": "z_{t}^{(0)} \\gets z_{t-1}" + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "text", + "content": " \n6 " + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "inline_equation", + "content": "z_{t}^{(\\mathrm{avg})} \\gets 0" + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "text", + "content": " \n7 for " + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "inline_equation", + "content": "k = 0,\\ldots,\\mathcal{K}" + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "text", + "content": " do \n8 " + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "inline_equation", + "content": "\\delta_{t}^{k} \\gets \\epsilon_{\\theta}(z_{t}^{(k)},t)" + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "text", + "content": " \n9 " + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "inline_equation", + "content": "\\delta_{t}^{k} \\gets \\text{Enhance-edit}(\\delta_{t}^{k},w_{k+1})" + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "text", + "content": " \n10 " + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "inline_equation", + "content": "z_{t}^{(k+1)} \\gets \\text{Inverse-Step}(z_{t-1},\\delta_{t}^{k})" + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "text", + "content": " \n11 end \n// Average ReNoised predictions \n23 Function Noise-Corr " + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "inline_equation", + "content": "(z_{t},t,\\epsilon_{t},z_{t-1})" + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "text", + "content": ": \n12 " + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "inline_equation", + "content": "z_{t}^{(\\mathrm{avg})} \\gets \\sum_{k=1}^{\\mathcal{K}+1} w_{k} \\cdot z_{t}^{(k)}" + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "text", + "content": " \n24 " + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "inline_equation", + "content": "\\delta_{t} \\gets \\epsilon_{\\theta}(z_{t},t)" + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "text", + "content": " \n25 " + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "inline_equation", + "content": "\\epsilon_{t} \\gets \\epsilon_{t} - \\nabla_{\\epsilon_{t}}\\frac{1}{\\rho_{t}}(z_{t-1} - \\phi_{t}z_{t} - \\psi_{t}\\delta_{t})" + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "text", + "content": " \n26 return " + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "inline_equation", + "content": "\\epsilon_{t}" + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "text", + "content": " \nend \nreturn " + }, + { + "bbox": [ + 132, + 285, + 488, + 495 + ], + "type": "inline_equation", + "content": "(z_{T},\\{\\epsilon_{t}\\}_{t=1}^{T})" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "code_body" + } + ], + "index": 5, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "text", + "content": "Noise Correction in Non-deterministic Samplers Non-deterministic samplers, in which " + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\rho_t > 0" + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "text", + "content": ", introduce noise " + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "inline_equation", + "content": "(\\epsilon_t)" + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "text", + "content": " at each denoising step. Previous methods [21,50] suggested using " + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\epsilon_t" + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "text", + "content": " to bridge the gap between the inversion and denoising trajectories in DDPM inversion. Specifically, given a pair of points " + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "inline_equation", + "content": "z_{t - 1},z_{t}" + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "text", + "content": " on the inversion trajectory, we denote by " + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\hat{z}_{t - 1}" + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "text", + "content": " the point obtained by denoising " + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "text", + "content": ". Ideally, " + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "inline_equation", + "content": "z_{t - 1}" + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\hat{z}_{t - 1}" + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "text", + "content": " should be identical. We define:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 227, + 569, + 481, + 593 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 569, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 227, + 569, + 481, + 593 + ], + "type": "interline_equation", + "content": "\\epsilon_ {t} = \\frac {1}{\\rho_ {t}} \\left(z _ {t - 1} - \\phi_ {t} z _ {t} - \\psi_ {t} \\epsilon_ {\\theta} \\left(z _ {t}, t, c\\right)\\right). \\tag {4}", + "image_path": "68e89ff5aaa9e00284a74f5b5327591d466edf5fb9a85985fca69b410d0fe875.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 479, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 479, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 479, + 628 + ], + "type": "text", + "content": "Integrating this definition into Eq. 1 yields " + }, + { + "bbox": [ + 130, + 594, + 479, + 628 + ], + "type": "inline_equation", + "content": "\\hat{z}_{t - 1} = z_{t - 1}" + }, + { + "bbox": [ + 130, + 594, + 479, + 628 + ], + "type": "text", + "content": ". However, we found that replacing " + }, + { + "bbox": [ + 130, + 594, + 479, + 628 + ], + "type": "inline_equation", + "content": "\\epsilon_t" + }, + { + "bbox": [ + 130, + 594, + 479, + 628 + ], + "type": "text", + "content": " with the above definition affects editability. Instead, we suggest a more tender approach, optimizing " + }, + { + "bbox": [ + 130, + 594, + 479, + 628 + ], + "type": "inline_equation", + "content": "\\epsilon_t" + }, + { + "bbox": [ + 130, + 594, + 479, + 628 + ], + "type": "text", + "content": " based on Eq. 4 as our guiding objective:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 209, + 628, + 481, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 628, + 481, + 651 + ], + "spans": [ + { + "bbox": [ + 209, + 628, + 481, + 651 + ], + "type": "interline_equation", + "content": "\\epsilon_ {t} = \\epsilon_ {t} - \\nabla_ {\\epsilon_ {t}} \\frac {1}{\\rho_ {t}} \\left(z _ {t - 1} - \\phi_ {t} z _ {t} - \\psi_ {t} \\epsilon_ {\\theta} \\left(z _ {t}, t, c\\right)\\right). \\tag {5}", + "image_path": "aa04487fec1985aad74f89813be08e076c8f2b959a6fe8f4f7389243668027e4.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 649, + 479, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 649, + 479, + 673 + ], + "spans": [ + { + "bbox": [ + 130, + 649, + 479, + 673 + ], + "type": "text", + "content": "This optimization improves the reconstruction fidelity while preserving the distribution of the noisy-latents." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 214, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 214, + 91, + 448, + 102 + ], + "type": "text", + "content": "ReNoise: Real Image Inversion Through Iterative Noising" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 114, + 312, + 172 + ], + "blocks": [ + { + "bbox": [ + 153, + 114, + 312, + 172 + ], + "lines": [ + { + "bbox": [ + 153, + 114, + 312, + 172 + ], + "spans": [ + { + "bbox": [ + 153, + 114, + 312, + 172 + ], + "type": "image", + "image_path": "425c549fb37a8697b4d1460cd6a27c7a715f365226c14e05a66f37013253ea2f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 174, + 333, + 251 + ], + "lines": [ + { + "bbox": [ + 132, + 174, + 333, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 174, + 333, + 251 + ], + "type": "text", + "content": "Fig. 6: Schematic illustration of the ReNoise convergence process to the true inversion of " + }, + { + "bbox": [ + 132, + 174, + 333, + 251 + ], + "type": "inline_equation", + "content": "z_{t-1}" + }, + { + "bbox": [ + 132, + 174, + 333, + 251 + ], + "type": "text", + "content": ". While estimates may converge non-monotonically to the unknown target " + }, + { + "bbox": [ + 132, + 174, + 333, + 251 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 132, + 174, + 333, + 251 + ], + "type": "text", + "content": ", we found that averaging them improves true value estimation. Typically, the initial iteration exhibits an exponential decrease in the norm between consecutive elements." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 340, + 126, + 480, + 183 + ], + "blocks": [ + { + "bbox": [ + 340, + 126, + 480, + 183 + ], + "lines": [ + { + "bbox": [ + 340, + 126, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 340, + 126, + 480, + 183 + ], + "type": "image", + "image_path": "48283f72459cd1c42321171bc038cb2f7c828989379fbd874f737b41062e1091.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 184, + 482, + 251 + ], + "lines": [ + { + "bbox": [ + 339, + 184, + 482, + 251 + ], + "spans": [ + { + "bbox": [ + 339, + 184, + 482, + 251 + ], + "type": "text", + "content": "Fig.7: Average distance between consequent estimations " + }, + { + "bbox": [ + 339, + 184, + 482, + 251 + ], + "type": "inline_equation", + "content": "z_{t}^{(k)}" + }, + { + "bbox": [ + 339, + 184, + 482, + 251 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 339, + 184, + 482, + 251 + ], + "type": "inline_equation", + "content": "z_{t}^{(k + 1)}" + }, + { + "bbox": [ + 339, + 184, + 482, + 251 + ], + "type": "text", + "content": ". Vertical bars indicate the standard deviation. The averages are computed over 32 images and 10 different timesteps." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 255, + 296, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 255, + 296, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 296, + 269 + ], + "type": "text", + "content": "4 Convergence Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 274, + 479, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 479, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 479, + 344 + ], + "type": "text", + "content": "In this section, we first express the inversion process as a backward Euler process and our renoising iterations as fixed-point iterations. While these iterations do not converge in the general case, in the supplementary materials we present a toy example where they yield accurate inversions. Then, we analyze the convergence of the renoising iterations in our real-image inversion scenario and empirically verify our method's convergence." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 347, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 347, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 131, + 347, + 482, + 407 + ], + "type": "text", + "content": "Inversion Process as Backward Euler The denoising process of diffusion models can be mathematically described as solving an ordinary differential equation (ODE). A common method for solving such equations is the Euler method, which takes small steps to approximate the solution. For ODE in the form of " + }, + { + "bbox": [ + 131, + 347, + 482, + 407 + ], + "type": "inline_equation", + "content": "y'(t) = f(t,y(t))" + }, + { + "bbox": [ + 131, + 347, + 482, + 407 + ], + "type": "text", + "content": ", Euler solution is defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 250, + 408, + 364, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 408, + 364, + 421 + ], + "spans": [ + { + "bbox": [ + 250, + 408, + 364, + 421 + ], + "type": "interline_equation", + "content": "y _ {n + 1} = y _ {n} + h \\cdot f (t _ {n}, y _ {n}),", + "image_path": "9c4a41a980fe386a3eb7328663e90f7e575b23654b430c42560369575eb0e9b9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 424, + 479, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 479, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 479, + 471 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 424, + 479, + 471 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 424, + 479, + 471 + ], + "type": "text", + "content": " is the step size. The inversion process can be described as solving ODE using the backward Euler method (or implicit Euler method) [1]. This method is similar to forward Euler, with the difference that " + }, + { + "bbox": [ + 130, + 424, + 479, + 471 + ], + "type": "inline_equation", + "content": "y_{n + 1}" + }, + { + "bbox": [ + 130, + 424, + 479, + 471 + ], + "type": "text", + "content": " appears on both sides of the equation:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 239, + 471, + 373, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 471, + 373, + 483 + ], + "spans": [ + { + "bbox": [ + 239, + 471, + 373, + 483 + ], + "type": "interline_equation", + "content": "y _ {n + 1} = y _ {n} + h \\cdot f \\left(t _ {n + 1}, y _ {n + 1}\\right).", + "image_path": "22fb0b7c8ddb8434ba5d62e0ed63bb17254dae1fc591969812a900691b7373e9.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 485, + 479, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 479, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 479, + 544 + ], + "type": "text", + "content": "For equations lacking an algebraic solution, several techniques estimate " + }, + { + "bbox": [ + 130, + 485, + 479, + 544 + ], + "type": "inline_equation", + "content": "y_{n+1}" + }, + { + "bbox": [ + 130, + 485, + 479, + 544 + ], + "type": "text", + "content": " iteratively. As we described in Section 3.1, the inversion process lacks a closed-form solution, as shown in Equation 2. To address this, the ReNoise method leverages fixed-point iterations, which we refer to as reonising iterations, to progressively refine the estimate of " + }, + { + "bbox": [ + 130, + 485, + 479, + 544 + ], + "type": "inline_equation", + "content": "y_{n+1}" + }, + { + "bbox": [ + 130, + 485, + 479, + 544 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 207, + 545, + 405, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 545, + 405, + 563 + ], + "spans": [ + { + "bbox": [ + 207, + 545, + 405, + 563 + ], + "type": "interline_equation", + "content": "y _ {n + 1} ^ {(0)} = y _ {n}, \\quad y _ {n + 1} ^ {(k + 1)} = y _ {n} + h \\cdot f (t _ {n + 1}, y _ {n + 1} ^ {(k)}.", + "image_path": "4fd3549a7a1040cd749ba1914677e14b3e0a350bf0e43453631685fdd6574740.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 567, + 479, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 479, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 479, + 591 + ], + "type": "text", + "content": "In our ReNoise method, we average these renoising iterations to mitigate convergence errors, leading to improvement in the reconstruction quality." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 594, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 594, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 131, + 594, + 482, + 668 + ], + "type": "text", + "content": "ReNoise Convergence During the inversion process, we aim to find the next noise level inversion, denoted by " + }, + { + "bbox": [ + 131, + 594, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\hat{z}_t" + }, + { + "bbox": [ + 131, + 594, + 482, + 668 + ], + "type": "text", + "content": ", such that applying the denoising step to " + }, + { + "bbox": [ + 131, + 594, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\hat{z}_t" + }, + { + "bbox": [ + 131, + 594, + 482, + 668 + ], + "type": "text", + "content": " recovers the previous state, " + }, + { + "bbox": [ + 131, + 594, + 482, + 668 + ], + "type": "inline_equation", + "content": "z_{t - 1}" + }, + { + "bbox": [ + 131, + 594, + 482, + 668 + ], + "type": "text", + "content": ". Given the noise estimation " + }, + { + "bbox": [ + 131, + 594, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_t,t)" + }, + { + "bbox": [ + 131, + 594, + 482, + 668 + ], + "type": "text", + "content": " and a given " + }, + { + "bbox": [ + 131, + 594, + 482, + 668 + ], + "type": "inline_equation", + "content": "z_{t - 1}" + }, + { + "bbox": [ + 131, + 594, + 482, + 668 + ], + "type": "text", + "content": ", the ReNoise mapping defined in Section 3.1 can be written as " + }, + { + "bbox": [ + 131, + 594, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{G}:z_t\\to \\mathrm{InverseStep}(z_{t - 1},\\epsilon_\\theta (z_t,t))" + }, + { + "bbox": [ + 131, + 594, + 482, + 668 + ], + "type": "text", + "content": ". For example, in the case of using DDIM sampler the mapping is " + }, + { + "bbox": [ + 131, + 594, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(z_t) = \\frac{1}{\\phi_t} (z_{t - 1} - \\psi_t\\epsilon_\\theta (z_t,t))" + }, + { + "bbox": [ + 131, + 594, + 482, + 668 + ], + "type": "text", + "content": ". The point " + }, + { + "bbox": [ + 131, + 594, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\hat{z}_t" + }, + { + "bbox": [ + 131, + 594, + 482, + 668 + ], + "type": "text", + "content": ", which is" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "D. Garibi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 178 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 178 + ], + "type": "text", + "content": "mapped to " + }, + { + "bbox": [ + 130, + 116, + 479, + 178 + ], + "type": "inline_equation", + "content": "z_{t-1}" + }, + { + "bbox": [ + 130, + 116, + 479, + 178 + ], + "type": "text", + "content": " after the denoising step, is a stationary point of this mapping. Given " + }, + { + "bbox": [ + 130, + 116, + 479, + 178 + ], + "type": "inline_equation", + "content": "z_t^{(1)}" + }, + { + "bbox": [ + 130, + 116, + 479, + 178 + ], + "type": "text", + "content": ", the first approximation of the next noise level " + }, + { + "bbox": [ + 130, + 116, + 479, + 178 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 130, + 116, + 479, + 178 + ], + "type": "text", + "content": ", our goal is to show that the sequence " + }, + { + "bbox": [ + 130, + 116, + 479, + 178 + ], + "type": "inline_equation", + "content": "z_t^{(k)} = \\mathcal{G}^{k-1}(z_t^{(1)})" + }, + { + "bbox": [ + 130, + 116, + 479, + 178 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 479, + 178 + ], + "type": "inline_equation", + "content": "k \\to \\infty" + }, + { + "bbox": [ + 130, + 116, + 479, + 178 + ], + "type": "text", + "content": " converges. As the mapping " + }, + { + "bbox": [ + 130, + 116, + 479, + 178 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 116, + 479, + 178 + ], + "type": "text", + "content": " is continuous, the limit point would be its stationary point. The definition of " + }, + { + "bbox": [ + 130, + 116, + 479, + 178 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 116, + 479, + 178 + ], + "type": "text", + "content": " gives:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 219, + 173, + 394, + 189 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 173, + 394, + 189 + ], + "spans": [ + { + "bbox": [ + 219, + 173, + 394, + 189 + ], + "type": "interline_equation", + "content": "\\left\\| z _ {t} ^ {(k + 1)} - z _ {t} ^ {(k)} \\right\\| = \\left\\| \\mathcal {G} \\left(z _ {t} ^ {(k)}\\right) - \\mathcal {G} \\left(z _ {t} ^ {(k - 1)}\\right) \\right\\|,", + "image_path": "0c64d857e1e7c42837189fb045d5822ba9d93df09fb18e590fb36f102eb03dcc.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 190, + 480, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 480, + 263 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 480, + 263 + ], + "type": "text", + "content": "where the norm is always assumed as the " + }, + { + "bbox": [ + 130, + 190, + 480, + 263 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 130, + 190, + 480, + 263 + ], + "type": "text", + "content": "-norm. For the ease of the notations, we define " + }, + { + "bbox": [ + 130, + 190, + 480, + 263 + ], + "type": "inline_equation", + "content": "\\varDelta^{(k)}=z_t^{(k)}-z_t^{(k-1)}" + }, + { + "bbox": [ + 130, + 190, + 480, + 263 + ], + "type": "text", + "content": ". For convergence proof, it is sufficient to show that the sum of norms of these differences converges, which will imply that " + }, + { + "bbox": [ + 130, + 190, + 480, + 263 + ], + "type": "inline_equation", + "content": "z_{t}^{(k)}" + }, + { + "bbox": [ + 130, + 190, + 480, + 263 + ], + "type": "text", + "content": " is the Cauchy sequence. Below we check that in practice " + }, + { + "bbox": [ + 130, + 190, + 480, + 263 + ], + "type": "inline_equation", + "content": "\\| \\varDelta^{(k)}\\|" + }, + { + "bbox": [ + 130, + 190, + 480, + 263 + ], + "type": "text", + "content": " decreases exponentially as " + }, + { + "bbox": [ + 130, + 190, + 480, + 263 + ], + "type": "inline_equation", + "content": "k\\to \\infty" + }, + { + "bbox": [ + 130, + 190, + 480, + 263 + ], + "type": "text", + "content": " and thus has finite sum. In the assumption that " + }, + { + "bbox": [ + 130, + 190, + 480, + 263 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 190, + 480, + 263 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 130, + 190, + 480, + 263 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^2" + }, + { + "bbox": [ + 130, + 190, + 480, + 263 + ], + "type": "text", + "content": "-smooth, the Taylor series conducts:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 144, + 266, + 470, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 266, + 470, + 356 + ], + "spans": [ + { + "bbox": [ + 144, + 266, + 470, + 356 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\| \\varDelta^ {(k + 1)} \\| = \\| \\mathcal {G} (z _ {t} ^ {(k)}) - \\mathcal {G} (z _ {t} ^ {(k - 1)}) \\| = \\\\ \\left\\| \\mathcal {G} \\left(z _ {t} ^ {(k - 1)}\\right) + \\frac {\\partial \\mathcal {G}}{\\partial z} \\right| _ {z _ {t} ^ {(k - 1)}} \\cdot \\Delta^ {(k)} + O \\left(\\left\\| \\Delta^ {(k)} \\right\\| ^ {2}\\right) - \\mathcal {G} \\left(z _ {t} ^ {(k - 1)}\\right) \\| = \\\\ \\left. \\left\\| \\frac {\\partial \\mathcal {G}}{\\partial z} \\right| _ {z _ {t} ^ {(k - 1)}} \\cdot \\Delta^ {(k)} + O (\\left\\| \\Delta^ {(k)} \\right\\| ^ {2}) \\right\\| \\leq \\left\\| \\frac {\\partial \\mathcal {G}}{\\partial z} \\right| _ {z _ {t} ^ {(k - 1)}} \\| \\cdot \\| \\Delta^ {(k)} \\| + O (\\left\\| \\Delta^ {(k)} \\right\\| ^ {2}) = \\\\ \\frac {\\psi_ {t}}{\\phi_ {t}} \\cdot \\left\\| \\frac {\\partial \\epsilon_ {\\theta}}{\\partial z} \\right| _ {z _ {t} ^ {(k - 1)}} \\| \\cdot \\| \\Delta^ {(k)} \\| + O (\\| \\Delta^ {(k)} \\| ^ {2}) \\\\ \\end{array}", + "image_path": "00499a1cd500ff49af9acf98048c330db78e6c820f0027d65bd2fa3494143510.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 356, + 480, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 480, + 489 + ], + "type": "text", + "content": "Thus, in a sufficiently small neighborhood, the convergence dynamics is defined by the scaled Jacobian norm " + }, + { + "bbox": [ + 130, + 356, + 480, + 489 + ], + "type": "inline_equation", + "content": "\\frac{\\psi_t}{\\phi_t} \\cdot \\left\\| \\frac{\\partial \\epsilon_\\theta}{\\partial z} \\right|_{z_t^{(k-1)}}" + }, + { + "bbox": [ + 130, + 356, + 480, + 489 + ], + "type": "text", + "content": ". In the supplementary materials, we show this scaled norm estimation for the SDXL diffusion model for various steps and ReNoise iterations indices " + }, + { + "bbox": [ + 130, + 356, + 480, + 489 + ], + "type": "inline_equation", + "content": "(k)" + }, + { + "bbox": [ + 130, + 356, + 480, + 489 + ], + "type": "text", + "content": ". Remarkably, the ReNoise indices minimally impact the scale factor, consistently remaining below 1. This confirms in practice the convergence of the proposed algorithm. Notably, the highest scaled norm values occur at smaller " + }, + { + "bbox": [ + 130, + 356, + 480, + 489 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 356, + 480, + 489 + ], + "type": "text", + "content": " (excluding the first step) and during the initial renoising iteration. This validates the strategy of not applying ReNoise in early steps, where convergence tends to be slower compared to other noise levels. Additionally, the scaled norm value for the initial " + }, + { + "bbox": [ + 130, + 356, + 480, + 489 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 356, + 480, + 489 + ], + "type": "text", + "content": " approaches 0, which induces almost immediate convergence." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 491, + 479, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 479, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 479, + 552 + ], + "type": "text", + "content": "Figure 7 illustrates the exponential decrease in distances between consecutive elements " + }, + { + "bbox": [ + 130, + 491, + 479, + 552 + ], + "type": "inline_equation", + "content": "z_{t}^{(k)}" + }, + { + "bbox": [ + 130, + 491, + 479, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 491, + 479, + 552 + ], + "type": "inline_equation", + "content": "z_{t}^{(k + 1)}" + }, + { + "bbox": [ + 130, + 491, + 479, + 552 + ], + "type": "text", + "content": ", which confirms the algorithm's convergence towards the stationary point of the operator " + }, + { + "bbox": [ + 130, + 491, + 479, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 491, + 479, + 552 + ], + "type": "text", + "content": ". The proposed averaging strategy is aligned with the conclusions described above, and also converges to the desired stationary point. In The supplementary materials, we present a validation for this claim." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 554, + 230, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 230, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 230, + 567 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 479, + 666 + ], + "type": "text", + "content": "In this section, we conduct extensive experiments to validate the effectiveness of our method. We evaluate both the reconstruction quality of our inversion and its editability. To demonstrate the versatility of our approach, we apply it to four models, SD [39], SDXL [36], SDXL Turbo [42], and LCM-LoRA [27], with SDXL Turbo and LCM-LoRA being few-step models. Additionally, we use various sampling algorithms including both deterministic and non-deterministic ones. Implementation details for each model are provided in the supplementary materials. Following previous works [5, 31], we quantitatively evaluate our" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 214, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 214, + 91, + 448, + 103 + ], + "type": "text", + "content": "ReNoise: Real Image Inversion Through Iterative Noising" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 118, + 302, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 118, + 302, + 227 + ], + "spans": [ + { + "bbox": [ + 130, + 118, + 302, + 227 + ], + "type": "text", + "content": "Table 1: Image reconstruction results with a fixed number of 100 UNet operations. Each row showcases the results obtained using different combinations of inversion steps, denoising steps, and renoising iterations, totaling 100 operations. As observed, allocating some of the operations to renoising iterations improves the reconstruction quality while maintaining the same execution time." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 131, + 229, + 304, + 327 + ], + "blocks": [ + { + "bbox": [ + 131, + 229, + 304, + 327 + ], + "lines": [ + { + "bbox": [ + 131, + 229, + 304, + 327 + ], + "spans": [ + { + "bbox": [ + 131, + 229, + 304, + 327 + ], + "type": "table", + "html": "
Image Reconstruction Results
Inv StepsInf StepsReNoise StepsL2 ↓PSNR ↑LPIPS ↓
505000.0036426.0230.06273
752500.0038225.4660.06605
802000.0040825.0450.07099
901000.0102320.2490.10305
252520.0018229.5690.03637
202030.0016729.8840.03633
101080.0023028.1560.04678
", + "image_path": "ef2973db3f0bd86bc1978d9448b04e58eb89ae24190f27f2d3cc167f3f5f79be.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 310, + 114, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 114, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 310, + 114, + 482, + 236 + ], + "type": "text", + "content": "Table 2: Quantitative ablation study on SDXL Turbo. We demonstrate the impact of each component of our inversion method on reconstruction results. The results improve with additional renoising iterations and significant enhancements occur through averaging final estimations. Additionally, we observe a reconstruction-editability trade-off, with edit losses causing degradation that is effectively mitigated by Noise Correction." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 312, + 237, + 484, + 327 + ], + "blocks": [ + { + "bbox": [ + 312, + 237, + 484, + 327 + ], + "lines": [ + { + "bbox": [ + 312, + 237, + 484, + 327 + ], + "spans": [ + { + "bbox": [ + 312, + 237, + 484, + 327 + ], + "type": "table", + "html": "
Ablation - Image Reconstruction
L2 ↓PSNR ↑LPIPS↓
Euler Inversion0.070011.7840.20337
+ 1 ReNoise0.055212.7960.20254
+ 4 ReNoise0.024916.5210.14821
+ 9 ReNoise0.012619.7020.10850
+ Averaging ReNoise0.008721.4910.08832
+ Edit Losses0.027618.4320.12616
+ Noise Correction0.019622.0770.08469
", + "image_path": "b3d67507f20d5da40a685fd0b93acdb664b15bbefb2b7ecee9d6fad3578c1d5c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 329, + 482, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 482, + 354 + ], + "type": "text", + "content": "method with three metrics: " + }, + { + "bbox": [ + 130, + 329, + 482, + 354 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 130, + 329, + 482, + 354 + ], + "type": "text", + "content": ", LPIPS [51], and PSNR. Unless stated otherwise, for both inversion and generation we use the prompt obtained from BLIP2 [24]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 357, + 293, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 357, + 293, + 369 + ], + "spans": [ + { + "bbox": [ + 131, + 357, + 293, + 369 + ], + "type": "text", + "content": "5.1 Reconstruction and Speed" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 373, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 468 + ], + "type": "text", + "content": "We begin by evaluating the reconstruction-speed tradeoff. The main computational cost of both the inversion and denoising processes is the forward pass through the UNet. In each renoising iteration, we perform one forward pass, which makes it computationally equal to a standard inversion step (as done in DDIM Inversion for example). In the following experiments, we compare the results of a sampler reversing with our method, where we match the number of UNet passes between the methods. For example, 8 steps of sampler reversing are compared against 4 steps with one renoising iteration at each step." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 471, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 579 + ], + "type": "text", + "content": "Qualitative Results In Figure 8 we show qualitative results of image reconstruction on SDXL Turbo [42]. Here, we utilize DDIM as the sampler, and apply four denoising steps for all configurations. Each row exhibits results obtained using a different amount of UNet operations. In our method, we apply four inversion steps, and a varying number of renoising iterations. As can be seen, the addition of renoising iterations gradually improves the reconstruction results. Conversely, employing more inversion steps proves insufficient for capturing all details in the image, as evident by the background of the car, or even detrimental to the reconstruction, as observed in the Uluro example." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 617 + ], + "type": "text", + "content": "Quantitative Results For the quantitative evaluation, we use the MS-COCO 2017 [25] validation dataset. Specifically, we retain images with a resolution greater than " + }, + { + "bbox": [ + 130, + 582, + 482, + 617 + ], + "type": "inline_equation", + "content": "420 \\times 420" + }, + { + "bbox": [ + 130, + 582, + 482, + 617 + ], + "type": "text", + "content": ", resulting in a dataset containing 3,865 images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "We begin by evaluating both the sampler reversing approach and our ReNoise method, while varying the number of UNet operations during the inversion process and keeping the number of denoising steps fixed. This experiment is conducted using various models (SDXL, SDXL Turbo, LCM) and samplers. For all" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "text", + "content": "D. Garibi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 299, + 320 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 299, + 320 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 299, + 320 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 299, + 320 + ], + "type": "image", + "image_path": "06f5e129beec9c5708f7efc9239a1d39b4d7c391a5fa130c3cf4d14d2b0a970e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 327, + 302, + 428 + ], + "lines": [ + { + "bbox": [ + 132, + 327, + 302, + 428 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 302, + 428 + ], + "type": "text", + "content": "Fig.8: Qualitative comparison between DDIM Inversion and our ReNoise method on SDXL Turbo (4 denoising steps). The first row displays the original images. The following rows display reconstructions from both methods, using the same number of UNet operations. DDIM employs smaller strides, while ReNoise utilizes more renoising steps." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 314, + 113, + 478, + 338 + ], + "blocks": [ + { + "bbox": [ + 314, + 113, + 478, + 338 + ], + "lines": [ + { + "bbox": [ + 314, + 113, + 478, + 338 + ], + "spans": [ + { + "bbox": [ + 314, + 113, + 478, + 338 + ], + "type": "image", + "image_path": "fefb9658da53f8e6004d929868ba26da842bf11bb0b72813f9e0c438eb103048.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 339, + 482, + 427 + ], + "lines": [ + { + "bbox": [ + 311, + 339, + 482, + 427 + ], + "spans": [ + { + "bbox": [ + 311, + 339, + 482, + 427 + ], + "type": "text", + "content": "Fig. 9: Ablation study on SDXL Turbo. The first row presents the input image. In each subsequent row, we show the reconstruction results using an additional component of our inversion method. The images in the bottom row represent the results obtained by our full method. NC stands for Noise Correction." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 434, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 518 + ], + "type": "text", + "content": "models, we utilize the DDIM [43] sampler. In addition, we employ the Ancestral-Euler scheduler for SDXL Turbo, and the default LCM sampler for LCM-LoRA. We set the number of denoising steps to 50 for SDXL, and to 4 for SDXL Turbo and LCM-LoRA. Quantitative results, using PSNR as the metric, are presented in Figure 11. We evaluate our method using different configurations. The x-axis refers to the number of UNet operations in the inversion process. Other metrics results are provided in the supplementary materials." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "As depicted in the graphs, incorporating additional renoising iterations proves to be more beneficial for image reconstruction compared to adding more inversion steps. Note that the performance of the Ancestral-Euler and LCM samplers noticeably degrades when the number of inversion steps exceeds the number of denoising steps. Unlike DDIM, these samplers have " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\varPhi_t\\approx 1" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ", resulting in an increase in the latent vector's norm beyond what can be effectively denoised in fewer steps. In this experiment, we maintain the same number of UNet operations for both ReNoise and the sampler reversing approach. However, in ReNoise, the number of inversion steps remains fixed, and the additional operations are utilized for renoising iterations, refining each point on the inversion trajectory. Consequently, our method facilitates improved reconstruction when using these noise samplers." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 214, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 214, + 91, + 448, + 102 + ], + "type": "text", + "content": "ReNoise: Real Image Inversion Through Iterative Noising" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 483, + 190 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 483, + 190 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 483, + 190 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 483, + 190 + ], + "type": "image", + "image_path": "73d002b752f357a55d5bced7a5dd06a7bc6cca1654498a3320305f17127d0034.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 190, + 482, + 222 + ], + "lines": [ + { + "bbox": [ + 130, + 190, + 482, + 222 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 482, + 222 + ], + "type": "text", + "content": "Fig. 10: LCM Editing Results. We showcase two examples of real images, each followed by three edits. The text below each edited image indicates the specific word or phrase replaced or added to the original prompt for that specific edit." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 135, + 224, + 249, + 299 + ], + "blocks": [ + { + "bbox": [ + 135, + 224, + 249, + 299 + ], + "lines": [ + { + "bbox": [ + 135, + 224, + 249, + 299 + ], + "spans": [ + { + "bbox": [ + 135, + 224, + 249, + 299 + ], + "type": "image", + "image_path": "acff721aa486e340a5a365ddb7f88be1de13d158959fb388ce73e5d484ee4189.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 300, + 482, + 377 + ], + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 377 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 377 + ], + "type": "text", + "content": "Fig. 11: Image reconstruction results comparing sampler reversing inversion techniques across different samplers (e.g., vanilla DDIM inversion) with our ReNoise method using the same sampler. The number of denoising steps remains constant. However, the number of UNet passes varies, with the sampler reversing approach increasing the number of inversion steps, while our method increases the number of renoising iterations. We present various configuration options for our method, including options with or without edit enhancement loss and Noise Correction (NC)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 252, + 224, + 364, + 299 + ], + "blocks": [ + { + "bbox": [ + 252, + 224, + 364, + 299 + ], + "lines": [ + { + "bbox": [ + 252, + 224, + 364, + 299 + ], + "spans": [ + { + "bbox": [ + 252, + 224, + 364, + 299 + ], + "type": "image", + "image_path": "9dd21dbad5845d21963121b5403376123930d2c7f6438b40648649d27366aad9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 367, + 224, + 481, + 299 + ], + "blocks": [ + { + "bbox": [ + 367, + 224, + 481, + 299 + ], + "lines": [ + { + "bbox": [ + 367, + 224, + 481, + 299 + ], + "spans": [ + { + "bbox": [ + 367, + 224, + 481, + 299 + ], + "type": "image", + "image_path": "3eac0152bacc04aa5316a7a9fc35c993e0fd0e7f7e169a34a269fba38abc4b33.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 378, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 486 + ], + "type": "text", + "content": "We continue by evaluating both the sampler reversing approach and our method while maintaining a fixed total number of UNet operations for the inversion and denoising processes combined. The results for SDXL with DDIM are presented in Table 1. The table displays various combinations of inversion, denoising, and renoising steps, totaling 100 UNet operations. Despite employing longer strides along the inversion and denoising trajectories, our ReNoise method yields improved reconstruction accuracy, as evident in the table. Furthermore, a reduced number of denoising steps facilitates faster image editing, especially since it commonly involves reusing the same inversion for multiple edits." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 488, + 317, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 488, + 317, + 501 + ], + "spans": [ + { + "bbox": [ + 131, + 488, + 317, + 501 + ], + "type": "text", + "content": "5.2 Reconstruction and Editability" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 502, + 482, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 482, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 482, + 598 + ], + "type": "text", + "content": "In Figure 10, we illustrate editing results generated by our method with LCM LoRA [27]. These results were obtained by inverting the image using a source prompt and denoising it with a target prompt. Each row exhibits an image followed by three edits accomplished by modifying the original prompt. These edits entail either replacing the object word or adding descriptive adjectives to it. As can be seen, the edited images retain the details present in the original image. For instance, when replacing the cat with a koala, the details in the background are adequately preserved." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 603, + 246, + 614 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 603, + 246, + 614 + ], + "spans": [ + { + "bbox": [ + 131, + 603, + 246, + 614 + ], + "type": "text", + "content": "5.3 Ablation Studies" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Figure 9 qualitatively demonstrates the effects of each component in our method, highlighting their contribution to the final outcome. Here, we use SDXL Turbo model [42], with the Ancestral-Euler sampler, which is non-deterministic. As our baseline, we simply reverse the sampler process. The reconstruction, while" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "text", + "content": "D. Garibi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 127, + 112, + 486, + 215 + ], + "blocks": [ + { + "bbox": [ + 127, + 112, + 486, + 215 + ], + "lines": [ + { + "bbox": [ + 127, + 112, + 486, + 215 + ], + "spans": [ + { + "bbox": [ + 127, + 112, + 486, + 215 + ], + "type": "image", + "image_path": "2abb376c359ed2ac846535570dac97e91daa302ad0e8e68fd08eff26f4882a61.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 217, + 301, + 260 + ], + "lines": [ + { + "bbox": [ + 132, + 217, + 301, + 260 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 301, + 260 + ], + "type": "text", + "content": "Fig. 12: Comparison with edit-friendly with SDXL Turbo. We inverted the image with the prompt \"a cat is sitting in front of a mirror\" and applied edits." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 311, + 217, + 482, + 260 + ], + "lines": [ + { + "bbox": [ + 311, + 217, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 311, + 217, + 482, + 260 + ], + "type": "text", + "content": "Fig. 13: Image reconstruction comparisons with SD. While NTI and our method achieve comparable results, ours demonstrates significant speed improvement." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 262, + 482, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 406 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 406 + ], + "type": "text", + "content": "semantically capturing the main object, fails to reproduce the image's unique details. For example, in the bird image, the reconstruction contains a bird standing on a branch, but the branch is in a different pose and the bird is completely different. Using 9 ReNoise iterations significantly improves the reconstruction, recovering finer details like the bird's original pose and branch texture. However, some subtle details, such as the bird's colors or the color in Brad Pitt's image, remain incomplete. Averaging the final iterations effectively incorporates information from multiple predictions, leading to a more robust reconstruction that captures finer details. Regularize the UNet's noise prediction with " + }, + { + "bbox": [ + 130, + 262, + 482, + 406 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{edit}}" + }, + { + "bbox": [ + 130, + 262, + 482, + 406 + ], + "type": "text", + "content": " can introduce minor artifacts to the reconstruction, evident in the smoother appearance of the hair of the two people on the left, or in the cake example. Finally, we present our full method by adding the noise correction technique." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 406, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 406, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 406, + 482, + 453 + ], + "type": "text", + "content": "Table 2 quantitatively showcases the effect each component has on reconstruction results. As can be seen, the best results were obtained by our full method or by averaging the last estimations of " + }, + { + "bbox": [ + 130, + 406, + 482, + 453 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 406, + 482, + 453 + ], + "type": "text", + "content": ". Our final method also offers the distinct advantage of getting an editable latent representation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 453, + 481, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 481, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 481, + 477 + ], + "type": "text", + "content": "In the Appendix, we present an ablation study to justify our editability enhancement and noise correction components." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 481, + 227, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 481, + 227, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 481, + 227, + 493 + ], + "type": "text", + "content": "5.4 Comparisons" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 498, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 545 + ], + "type": "text", + "content": "Inversion for Non-deterministic Samplers. In Figure 12 we show a qualitative comparison with \"an edit-friendly DDPM\" [21] where we utilize SDXL Turbo [42]. Specifically, we assess the performance of the edit-friendly method alongside our ReNoise method in terms of both reconstruction and editing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "We observe that in non-deterministic samplers like DDPM, the parameter " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\rho_0" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " in Equation 1 equals zero. This means that in the final denoising step, the random noise addition is skipped to obtain a clean image. In long diffusion processes (e.g., 50-100 steps), the final denoising step often has minimal impact as the majority of image details have already been determined. Conversely, shorter diffusion processes rely on the final denoising step to determine fine details of the image. Due to focusing solely on noise correction to preserve the original image in the inversion process, edit-friendly struggles to reconstruct fine details of the image, such as the shower behind the cat. However, our ReNoise method finds an inversion trajectory that faithfully reconstructs the image and does not rely" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 214, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 214, + 91, + 448, + 102 + ], + "type": "text", + "content": "ReNoise: Real Image Inversion Through Iterative Noising" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "solely on noise corrections. This allows us to better reconstruct fine details such as the shower. Additionally, encoding a significant amount of information within only a few external noise vectors, " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "\\epsilon_{t}" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": ", limits editability in certain scenarios. See more examples in the appendix" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 168, + 483, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 168, + 483, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 168, + 483, + 326 + ], + "type": "text", + "content": "Null-prompt Inversion Methods In Figure 13, we present a qualitative comparison between our method and null-text based inversion methods. For this comparison, we utilize Stable Diffusion [39] since these methods rely on a CFG [20] mechanism, which is not employed in SDXL Turbo [42]. Specifically, we compare DDIM Inversion [43] with one renoising iteration to Null-Text Inversion (NTI) [31] and Negative-Prompt Inversion (NPI) [30]. Both NTI and NPI enhance the inversion process by replacing the null-text token embedding when applying CFG. Our method achieves results comparable to NTI, while NPI highlights the limitations of plain DDIM inversion. This is because NPI sets the original prompt as the negative prompt, essentially resulting in an inversion process identical to plain DDIM inversion. Regarding running time, our ReNoise inversion process takes 13 seconds, significantly faster than NTI's 3 minutes. For comparison, plain DDIM inversion and NPI each take 9 seconds." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 331, + 221, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 331, + 221, + 344 + ], + "spans": [ + { + "bbox": [ + 131, + 331, + 221, + 344 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 352, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 495 + ], + "type": "text", + "content": "In this work, we have introduced ReNoise, a universal approach that enhances various inversion algorithms of diffusion models. ReNoise gently guides the inversion curve of a real image towards the source noise from which a denoising process reconstructs the image. ReNoise can be considered as a meta-algorithm that warps the trajectory of any iterative diffusion inversion process. Our experiments demonstrate that averaging the last few renoising iterations significantly enhances reconstruction quality. For a fixed amount of computation, ReNoise shows remarkably higher reconstruction quality and editability. The method is theoretically supported and our experiments reconfirm its effectiveness on a variety of diffusion models and sampling algorithms. Moreover, the method is numerically stable, and always converges to some inversion trajectory that eases hyperparameters adjustment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 497, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 532 + ], + "type": "text", + "content": "Beyond the net introduction of an effective inversion, the paper presents a twofold important contribution: an effective inversion for few-steps diffusion models, which facilitates effective editing on these models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 582 + ], + "type": "text", + "content": "A limitation of ReNoise is the model-specific hyperparameter tuning required for Edit Enhancement and Noise Correction. While these hyperparameters remain stable for a given model, they may vary across models, and tuning them is necessary to achieve high reconstruction quality while maintaining editability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "While ReNoise demonstrates the potential for editing few-step diffusion models, more extensive testing with advanced editing methods is needed. It is worth noting that no such editing has been demonstrated for the few-step diffusion models. We believe and hope that our ReNoise method will pave the way for fast and effective editing methods based on the few-steps models. We also believe that ReNoise can be adapted to the challenging problem of inverting video-diffusion models." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "D. Garibi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 138, + 140, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 481, + 172 + ], + "type": "text", + "content": "1. Numerical Differential Equation Methods, chap. 2, pp. 45-121. John Wiley and Sons, Ltd (2003). https://doi.org/https://doi.org/10.1002/0470868279.ch2, https://onlinelibrary.wiley.com/doi/abs/10.1002/0470868279.ch2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 481, + 216 + ], + "type": "text", + "content": "2. Abdal, R., Qin, Y., Wonka, P.: Image2stylegan: How to embed images into the stylegan latent space? In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE (Oct 2019). https://doi.org/10.1109/iccv.2019.00453, http://dx.doi.org/10.1109/ICCV.2019.00453" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 216, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 216, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 216, + 481, + 259 + ], + "type": "text", + "content": "3. Abdal, R., Qin, Y., Wonka, P.: Image2stylegan++: How to edit the embedded images? In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (Jun 2020). https://doi.org/10.1109/cvpr42600.2020.00832, http://dx.doi.org/10.1109/cvpr42600.2020.00832" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 261, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 261, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 138, + 261, + 481, + 282 + ], + "type": "text", + "content": "4. Alaluf, Y., Garibi, D., Patashnik, O., Averbuch-Elor, H., Cohen-Or, D.: Cross-image attention for zero-shot appearance transfer (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 283, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 283, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 138, + 283, + 481, + 326 + ], + "type": "text", + "content": "5. Alaluf, Y., Patashnik, O., Cohen-Or, D.: Restyle: A residual-based stylegan encoder via iterative refinement. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE (Oct 2021). https://doi.org/10.1109/iccv48922.2021.00664, http://dx.doi.org/10.1109/iccv48922.2021.00664" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 327, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 327, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 327, + 481, + 347 + ], + "type": "text", + "content": "6. Alaluf, Y., Tov, O., Mokady, R., Gal, R., Bermano, A.H.: Hyperstyle: Stylegan inversion with hypernetworks for real image editing (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 348, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 481, + 369 + ], + "type": "text", + "content": "7. Avrahami, O., Fried, O., Lischinski, D.: Blended latent diffusion. arXiv preprint arXiv:2206.02779 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 370, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 481, + 403 + ], + "type": "text", + "content": "8. Avrahami, O., Lischinski, D., Fried, O.: Blended diffusion for text-driven editing of natural images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18208-18218 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 403, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 481, + 424 + ], + "type": "text", + "content": "9. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 425, + 481, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 481, + 436 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 481, + 436 + ], + "type": "text", + "content": "0. Burden, R., Faires, J., Burden, A.: Numerical Analysis. Cengage Learning (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 437, + 481, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 437, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 138, + 437, + 481, + 457 + ], + "type": "text", + "content": "1. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. ArXiv abs/2210.11427 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 458, + 469, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 458, + 469, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 458, + 469, + 468 + ], + "type": "text", + "content": "2. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 481, + 501 + ], + "type": "text", + "content": "3. Dinh, T.M., Tran, A.T., Nguyen, R., Hua, B.S.: Hyperinverter: Improving stylegan inversion via hypernetwork. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 502, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 481, + 522 + ], + "type": "text", + "content": "4. Epstein, D., Jabri, A., Poole, B., Efros, A.A., Holynski, A.: Diffusion self-guidance for controllable image generation (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 523, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 481, + 555 + ], + "type": "text", + "content": "5. Ge, S., Park, T., Zhu, J.Y., Huang, J.B.: Expressive text-to-image generation with rich text. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 556, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 556, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 556, + 481, + 599 + ], + "type": "text", + "content": "6. Han, L., Wen, S., Chen, Q., Zhang, Z., Song, K., Ren, M., Gao, R., Stathopoulos, A., He, X., Chen, Y., Liu, D., Zhangli, Q., Jiang, J., Xia, Z., Srivastava, A., Metaxas, D.: Improving tuning-free real image editing with proximal guidance (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 600, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 481, + 621 + ], + "type": "text", + "content": "7. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 622, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 481, + 643 + ], + "type": "text", + "content": "8. Hertz, A., Voynov, A., Fruchter, S., Cohen-Or, D.: Style aligned image generation via shared attention (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 644, + 455, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 644, + 455, + 654 + ], + "spans": [ + { + "bbox": [ + 138, + 644, + 455, + 654 + ], + "type": "text", + "content": "19. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models (2020)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 654, + 395, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 654, + 395, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 654, + 395, + 665 + ], + "type": "text", + "content": "20. Ho, J., Salimans, T.: Classifier-free diffusion guidance (2022)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 214, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 214, + 91, + 447, + 102 + ], + "type": "text", + "content": "ReNoise: Real Image Inversion Through Iterative Noising" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "21. Huberman-Spiegelglas, I., Kulikov, V., Michaeli, T.: An edit friendly ddpm noise space: Inversion and manipulations (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "type": "text", + "content": "22. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems 35, 26565-26577 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "text", + "content": "23. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Magic: Text-based real image editing with diffusion models. In: Conference on Computer Vision and Pattern Recognition 2023 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 480, + 237 + ], + "type": "text", + "content": "24. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 480, + 270 + ], + "type": "text", + "content": "25. Lin, T., Maire, M., Belongie, S.J., Bourdev, L.D., Girshick, R.B., Hays, J., Perona, P., Ramanan, D., Doll'a r, P., Zitnick, C.L.: Microsoft COCO: common objects in context. CoRR abs/1405.0312 (2014)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "text", + "content": "26. Luo, S., Tan, Y., Huang, L., Li, J., Zhao, H.: Latent consistency models: Synthesizing high-resolution images with few-step inference (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "27. Luo, S., Tan, Y., Patil, S., Gu, D., von Platen, P., Passos, A., Huang, L., Li, J., Zhao, H.: Lcm-lora: A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 347 + ], + "type": "text", + "content": "28. Meiri, B., Samuel, D., Darshan, N., Chechik, G., Avidan, S., Ben-Ari, R.: Fixed-point inversion for text-to-image diffusion models (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 348, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 480, + 369 + ], + "type": "text", + "content": "29. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 370, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 480, + 391 + ], + "type": "text", + "content": "30. Miyake, D., Iohara, A., Saito, Y., Tanaka, T.: Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "text", + "content": "31. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "type": "text", + "content": "32. Pan, Z., Gherardi, R., Xie, X., Huang, S.: Effective real image editing with accelerated iterative diffusion inversion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15912-15921 (October 2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 447, + 480, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 499 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 499 + ], + "type": "text", + "content": "33. Parmar, G., Kumar Singh, K., Zhang, R., Li, Y., Lu, J., Zhu, J.Y.: Zero-shot image-to-image translation. In: Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Proceedings. SIGGRAPH '23, ACM (Jul 2023). https://doi.org/10.1145/3588432.3591513, http://dx.doi.org/10.1145/3588432.3591513" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 501, + 480, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 554 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 554 + ], + "type": "text", + "content": "34. Parmar, G., Li, Y., Lu, J., Zhang, R., Zhu, J.Y., Singh, K.K.: Spatially-adaptive multilayer selection for gan inversion and editing. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (Jun 2022). https://doi.org/10.1109/cvpr52688.2022.01111, http://dx.doi.org/10.1109/CVPR52688.2022.01111" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "text", + "content": "35. Patashnik, O., Garibi, D., Azuri, I., Averbuch-Elor, H., Cohen-Or, D.: Localizing object-level shape variations with text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 590, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 621 + ], + "type": "text", + "content": "36. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "text", + "content": "37. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "text", + "content": "38. Richardson, E., Alaluf, Y., Patashnik, O., Nitzan, Y., Azar, Y., Shapiro, S., Cohen-Or, D.: Encoding in style: a stylegan encoder for image-to-image translation. In:" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 100 + ], + "type": "text", + "content": "D. Garibi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 533 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 149, + 116, + 480, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 116, + 480, + 139 + ], + "spans": [ + { + "bbox": [ + 149, + 116, + 480, + 139 + ], + "type": "text", + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "text", + "content": "39. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 205 + ], + "type": "text", + "content": "40. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Ghasemipour, S.K.S., Ayan, B.K., Mahdavi, S.S., Lopes, R.G., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic text-to-image diffusion models with deep language understanding (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "text", + "content": "41. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models. In: International Conference on Learning Representations (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 228, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 480, + 248 + ], + "type": "text", + "content": "42. Sauer, A., Lorenz, D., Blattmann, A., Rombach, R.: Adversarial diffusion distillation (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 448, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 448, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 448, + 259 + ], + "type": "text", + "content": "43. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 260, + 453, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 453, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 453, + 270 + ], + "type": "text", + "content": "44. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 271, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 271, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 133, + 271, + 480, + 303 + ], + "type": "text", + "content": "45. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: International Conference on Learning Representations (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "text", + "content": "46. Tov, O., Alaluf, Y., Nitzan, Y., Patashnik, O., Cohen-Or, D.: Designing an encoder for stylegan image manipulation. ACM Trans. Graph. 40(4) (jul 2021). https://doi.org/10.1145/3450626.3459838" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "text", + "content": "47. Tumanyan, N., Geyer, M., Bagon, S., Dekel, T.: Plug-and-play diffusion features for text-driven image-to-image translation pp. 1921-1930 (June 2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 358, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 380 + ], + "type": "text", + "content": "48. Voynov, A., Hertz, A., Arar, M., Fruchter, S., Cohen-Or, D.: Anylens: A generative diffusion model with any rendering lens (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "text", + "content": "49. Wallace, B., Gokul, A., Naik, N.: Edict: Exact diffusion inversion via coupled transformations. arXiv preprint arXiv:2211.12446 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "text", + "content": "50. Wu, C.H., De la Torre, F.: Unifying diffusion models' latent space, with applications to cyclediffusion and guidance. arXiv preprint arXiv:2210.05559 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "type": "text", + "content": "51. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "type": "text", + "content": "52. Zhu, J., Shen, Y., Zhao, D., Zhou, B.: In-domain gan inversion for real image editing. In: Proceedings of European Conference on Computer Vision (ECCV) (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "text", + "content": "53. Zhu, J.Y., Krahenbihl, P., Shechtman, E., Efros, A.A.: Generative visual manipulation on the natural image manifold. In: Proceedings of European Conference on Computer Vision (ECCV) (2016)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "text", + "content": "54. Zhu, P., Abdal, R., Qin, Y., Femiani, J., Wonka, P.: Improved stylegan embedding: Where are the good latents? (2020)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 214, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 214, + 91, + 448, + 102 + ], + "type": "text", + "content": "ReNoise: Real Image Inversion Through Iterative Noising" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/RePOSE_ 3D Human Pose Estimation via Spatio-Temporal Depth Relational Consistency/7257623d-d6ee-4c47-b288-fda8cfd3af7e_content_list.json b/2024/RePOSE_ 3D Human Pose Estimation via Spatio-Temporal Depth Relational Consistency/7257623d-d6ee-4c47-b288-fda8cfd3af7e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b0f0c9872b5574f4ffd979eaa798dbec2d3f2b0e --- /dev/null +++ b/2024/RePOSE_ 3D Human Pose Estimation via Spatio-Temporal Depth Relational Consistency/7257623d-d6ee-4c47-b288-fda8cfd3af7e_content_list.json @@ -0,0 +1,1991 @@ +[ + { + "type": "text", + "text": "RePOSE: 3D Human Pose Estimation via Spatio-Temporal Depth Relational Consistency", + "text_level": 1, + "bbox": [ + 230, + 140, + 772, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziming Sun $^{1,2}$ , Yuan Liang $^{1,2\\star}$ , Zejun Ma $^{3}$ , Tianle Zhang $^{1,2}$ , Linchao Bao $^{4}$ , Guiqing Li $^{1}$ , and Shengfeng He $^{2(\\boxtimes)}$", + "bbox": [ + 230, + 210, + 774, + 244 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{1}$ South China University of Technology, China", + "2 Singapore Management University, Singapore" + ], + "bbox": [ + 339, + 253, + 660, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "shengfenghe@smu.edu.sg", + "bbox": [ + 415, + 284, + 586, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Johns Hopkins University, USA", + "bbox": [ + 388, + 296, + 612, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tencent AI Lab, China", + "bbox": [ + 416, + 310, + 586, + 323 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/StupidAdam/RePOSE", + "bbox": [ + 359, + 325, + 640, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We introduce RePOSE, a simple yet effective approach for addressing occlusion challenges in the learning of 3D human pose estimation (HPE) from videos. Conventional approaches typically employ absolute depth signals as supervision, which are adept at discernible keypoints but become less reliable when keypoints are occluded, resulting in vague and inconsistent learning trajectories for the neural network. RePOSE overcomes this limitation by introducing spatio-temporal relational depth consistency into the supervision signals. The core rationale of our method lies in prioritizing the precise sequencing of occluded keypoints. This is achieved by using a relative depth consistency loss that operates in both spatial and temporal domains. By doing so, RePOSE shifts the focus from learning absolute depth values, which can be misleading in occluded scenarios, to relative positioning, which provides a more robust and reliable cue for accurate pose estimation. This subtle yet crucial shift facilitates more consistent and accurate 3D HPE under occlusion conditions. The elegance of our core idea lies in its simplicity and ease of implementation, requiring only a few lines of code. Extensive experiments validate that RePOSE not only outperforms existing state-of-the-art methods but also significantly enhances the robustness and precision of 3D HPE in challenging occluded environments.", + "bbox": [ + 259, + 376, + 743, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D human pose estimation $\\cdot$ depth relational loss functions", + "bbox": [ + 261, + 667, + 740, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 708, + 375, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The estimation of 3D human pose from videos is a foundational task in computer vision, with widespread implications across various applications such as virtual reality [4, 12, 21, 38], human-computer interaction [8, 36, 37], and action recognition [10, 34, 35, 40, 41]. This task involves localizing body joints in a three-dimensional space to create a comprehensive representation of the human body,", + "bbox": [ + 212, + 739, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* The first two authors contributed equally to this work.", + "bbox": [ + 217, + 824, + 598, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1c0744240422b5ab87e1d08e8c0d44299961764cc9c5722792e646faf3ab17d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 143, + 333, + 229 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ee5c2813a09c79f0b6d3f807c0c0d33838a997c3b65d0f3ac5972afdbcdede1c.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 222, + 229, + 331, + 313 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7dcd5bd38044a893f609861415fca27bfdba1859012395069acfeb57eaf5af9b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 339, + 145, + 444, + 229 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f5b726767e5dba04012cabc7e6041a6845e5404b85ca52b5fca53de346d6e88e.jpg", + "image_caption": [ + "(b) GT" + ], + "image_footnote": [], + "bbox": [ + 339, + 229, + 444, + 313 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8e272043c2572a96e9c8ecb996057f97a01fcb710ba3e731e0c56fe53e5ce617.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 452, + 146, + 555, + 229 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d882ac318445d710eea8c43ec415aadab0de81f95b390ea5d435edeb792812e0.jpg", + "image_caption": [ + "(c) MotionBERT", + "Fig. 1: We introduce RePOSE, a novel solution designed to address the occlusion challenges in 3D pose estimation through relational supervision. Unlike existing methods (c) [46] and (d) [44], which struggle with occlusions due to their dependence on absolute depth signals, our approach leverages relational depth consistency, effectively overcoming these obstacles." + ], + "image_footnote": [], + "bbox": [ + 452, + 229, + 555, + 313 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6bda6a8be2c27fa49d56f5e8980e2ec04eb951de7e13bca235d93d2b760e4ca6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 146, + 669, + 229 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/694bf7c0ca073df277bb70a8ccb99aa40fcc167a934bab752673892d379143f3.jpg", + "image_caption": [ + "(d) PoseFormerV2" + ], + "image_footnote": [], + "bbox": [ + 563, + 229, + 669, + 313 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a79564eefe63fb04713494b2d95003caae75f1582bed2ef21ea534e1a9fbcc87.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 146, + 782, + 229 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/06e956d11a28004aab4d74247725c7c16f8b759257fd4192fa47c4ac8cbf8948.jpg", + "image_caption": [ + "(e) Ours" + ], + "image_footnote": [], + "bbox": [ + 676, + 229, + 782, + 313 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "typically in the form of a skeletal model. The methodologies in 3D human pose estimation (HPE) are primarily categorized into two approaches: direct estimation and 2D-to-3D lifting methods. The former [22, 25] entails the direct estimation of human pose from provided videos, whereas the latter [2, 14, 17, 27, 42, 44-46] involves the initial estimation of intermediate 2D poses using off-the-shelf 2D detectors, followed by lifting them to 3D space. 2D-to-3D lifting methods have gained prominence due to the rapid evolution of 2D pose detection technologies, exhibiting superior performance in many scenarios.", + "bbox": [ + 212, + 441, + 787, + 563 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advancements in 2D-to-3D lifting methods often employ Transformer architectures [31], which have shown exceptional ability in capturing global information and modeling sequential data. These methods typically involve transforming detected 2D poses into 3D space using tailored algorithms [13-15,27,42,44-46]. However, this process inherently faces a many-to-one mapping challenge, where multiple distinct 3D poses can correspond to a single 2D projection, particularly under occlusion scenarios, as shown in Fig. 1c and 1d. In such cases, occlusion of key body joints leads to reduced accuracy in 2D pose estimations, creating ambiguities in the 3D reconstruction process. Moreover, relying on these unreliable occluded signals during model training results in vague and uncertain learning pathways. This limitation chiefly stems from the methods' reliance on absolute depth signals, which lose reliability under occluded conditions, thus undermining the overall accuracy of 3D pose estimation.", + "bbox": [ + 212, + 566, + 787, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this challenge, we propose RePOSE, a novel approach specifically designed to enhance the accuracy of 3D HPE in occluded scenarios. RePOSE circumvents the limitations of absolute depth reliance by introducing a spatiotemporal relational depth consistency. Our method emphasizes the correct sequencing of occluded keypoints using a relative depth consistency loss in both", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z. Sun et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "spatial and temporal domains, rather than depending solely on absolute depth values. This shift in approach enables our method to robustly handle occlusions, providing more accurate and reliable pose estimations, as shown in Fig. 1e. Additionally, the simplicity of spatio-temporal relational depth consistency, which can be implemented in just a few lines of code, makes it a practical and effective solution for real-world applications.", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We demonstrate through extensive experiments on the Human3.6M [11] and MPI-INF-3DHP [20] benchmarks that our method outperforms state-of-the-art methods by a large margin, particularly excelling in scenarios involving occlusions. Our findings are further substantiated through visual comparisons and an ablation study, showcasing the effectiveness and innovation of RePOSE in tackling one of the most challenging aspects of 3D human pose estimation. This method lays the groundwork for future research aimed at enhancing pose estimation accuracy and reliability in complex environments.", + "bbox": [ + 212, + 237, + 787, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 383, + 395, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2D-to-3D Lifting HPE. The prevalent approach in 3D HPE involves lifting 2D pose sequences to 3D poses, a method enhanced by advanced 2D detectors like CPN [3], AlphaPose [7], and HRNet [33]. These detectors facilitate the conversion process with various techniques, including Fully Connected Networks (FCN) [6,19], Long Short-Term Memory (LSTM) [9,16,32], Graph Convolutional Networks (GCN) [1,5,39,43], Temporal Convolutional Networks (TCN) [2,17,26], and Transformers [13,14,27,42,44-46]. While SimpleBaseline [19] and LSTM-based methods [9] highlight the initial strides, they exhibit limitations in computational efficiency and temporal information utilization. GCN-based methods [1,5,39,43], though adept at local joint modeling, struggle with computational complexity. Transformer-based architectures, such as PoseFormer [45], bring significant improvements by modeling spatial-temporal dependencies. The multi-solution nature of 2D-to-3D lifting is addressed by methods like MHFormer [14], which generate multiple hypotheses to account for ambiguity in body part locations. Our work challenges this approach, positing a single accurate solution corresponding to the original 2D pose, driven by spatial and temporal relational constraints to optimize results.", + "bbox": [ + 212, + 415, + 787, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Loss Functions in 3D Human Pose Estimation. The accuracy of 3D HPE models heavily relies on the choice of loss functions. The Mean Per Joint Position Error (MPJPE) [14, 27, 44-46] is a prevalent metric, gauging the Euclidean distance between predicted and ground truth joint positions. Its simplicity and direct measurement of joint accuracy make it a popular choice. However, it does not account for joint orientations, potentially overlooking critical aspects of pose estimation. The Weighted-MPJPE (W-MPJPE) [42] attempts to refine this by assigning different weights to various joints, thereby prioritizing certain joints over others based on the application's needs. Yet, this method introduces its own set of challenges, including potential bias in weight assignment and the", + "bbox": [ + 212, + 688, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "3D Human Pose Estimation via Depth Relational Consistency", + "bbox": [ + 316, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "need for expert knowledge in determining these weights. Additionally, the Mean Per Joint Velocity Error (MPJVE) [26,42] has been used to improve the temporal coherence between the predicted pose sequence and the ground truth sequence, which is a crucial aspect in dynamic pose estimation. However, it falls short in leveraging long-range temporal information. To address this gap, Zhang et al. [42] integrated the MPJPE and temporal consistency losses introduced by Hossain et al. [9] to leverage temporal relationships and smooth pose transitions. However, these adaptations insufficiently tackle the issue of relative depth perception among joints. Various studies [29, 46] attempt to quantify limb angle inaccuracies by calculating the Mean Absolute Error between the estimated and ground-truth skeletal angles. Additionally, they explore angle dynamics through the Mean Per-Angle Velocity Error (MPAVE), representing the mean Euclidean distance in the angle's first temporal derivative. Despite these efforts, current methodologies overly focus on the x and y dimensions, derived primarily from 2D pose estimations, neglecting the critical depth (z-axis) disparities. Alternatively, Pavlakos et al. [25] utilize predicted depth values to compute the ordinal depth loss, especially when handling datasets that lack ground truth labels. However, this approach offers a weaker supervisory signal due to the inherent unreliability of the predictions.", + "bbox": [ + 212, + 145, + 787, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In response to the above problems, we introduce two novel loss functions: spatial and temporal relational depth consistency losses. The spatial loss focuses on relative depth differences between joints within a single frame, ensuring accurate depth relationships and mitigating pose estimation inconsistencies. The temporal loss maintains pose smoothness over time, considering relative depth differences of the same joint across frames. This dual focus enhances the overall robustness of 3D HPE, ensuring more accurate, temporally coherent, and visually appealing human pose sequences. Our methodology, utilizing these losses, aims to significantly improve the accuracy and temporal stability of pose estimation, making it more applicable and reliable in various computer vision tasks.", + "bbox": [ + 212, + 435, + 787, + 587 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 RePOSE", + "text_level": 1, + "bbox": [ + 215, + 616, + 339, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Problem Formulation and Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 656, + 591, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our method leverages the 2D-to-3D lifting pipeline to elevate a sequence of 2D skeletons to their corresponding 3D coordinates. Given a sequence of 2D joints with confidence scores, represented as $X \\in \\mathbb{R}^{T \\times J \\times 3}$ where $T$ and $J$ denote the number of frames and joints respectively, this input is transformed into a d-dimensional feature space $F \\in \\mathbb{R}^{T \\times J \\times d}$ by the lifting network, with $d$ varying across different network architectures. The output of this process is the estimated 3D pose sequence, denoted as $\\hat{Y} \\in \\mathbb{R}^{T \\times J \\times 3}$ , obtained through a regression mechanism. In our approach, we specifically treat the third dimension of $\\hat{Y}$ as the depth feature $\\hat{D} \\in \\mathbb{R}^{T \\times J}$ , facilitating the computation of relational depth consistency losses for model training supervision.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z. Sun et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e7032e65c42f3b007f1ebee51fd94053d288f084a8f699b782dee065ae405815.jpg", + "image_caption": [ + "Fig. 2: Our framework pipeline. Our network is designed with a spatio-temporal architecture, enabling simultaneous modeling of both spatial and temporal information. The generated 3D poses are processed through dual loss modules: the spatial loss module focuses on the relative depth disparities among joints within each frame, while the temporal loss module evaluates the relative depth differences of identical joints across consecutive frames. This dual-module approach ensures comprehensive depth accuracy and temporal consistency in the pose estimation process." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 782, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prevailing methodologies in recent studies [13,14,27,42,44-46] predominantly utilize Mean Per Joint Position Error (MPJPE) as loss function, defined as:", + "bbox": [ + 214, + 486, + 784, + 517 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m p j p e} = \\sum_ {t = 1} ^ {T} \\sum_ {i = 1} ^ {J} \\| \\hat {y} _ {t, i} - y _ {t, i} \\| _ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 534, + 784, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $T$ and $J$ represent the total number of frames and joints in each video, $\\hat{y}_{t,i}$ and $y_{t,i}$ correspond to the estimated and ground-truth positions of keypoint $i$ in frame $t$ , respectively.", + "bbox": [ + 214, + 592, + 782, + 637 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While the MPJPE is a widely accepted metric in 3D HPE, it has limitations, particularly in its focus on joint positions without explicit consideration of spatial relationships or pose topology. This becomes a significant concern in scenarios with occlusions, where understanding the relative positions of body parts is vital for accurate pose estimation. To address this, we introduce the spatial depth ranking loss, specifically designed to enhance pose topology awareness.", + "bbox": [ + 212, + 640, + 784, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, conventional techniques often fail to effectively utilize temporal information, as they typically compute loss independently for each frame, neglecting the continuity and interdependence within video sequences. This oversight can significantly limit a model's ability to capture realistic human motion over time, a crucial aspect in dynamic pose estimation tasks. To mitigate this gap, we propose the temporal depth ranking loss, a novel component aimed at incorporating temporal dynamics for more accurate and lifelike pose estimation.", + "bbox": [ + 212, + 734, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "3D Human Pose Estimation via Depth Relational Consistency", + "bbox": [ + 316, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Network Structure", + "text_level": 1, + "bbox": [ + 215, + 146, + 418, + 159 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inspired by [27,45], we recognize the exemplary performance of spatial-temporal transformer architectures in 3D human pose estimation (HPE) due to their robust capability to model both spatial and temporal information in video sequences. In particular, they demonstrate a parallel dual-stream design to be effective in ensuring comprehensive learning of spatial and temporal aspects. Consequently, our network, RePOSE, adopts a bifurcated structure of the spatiotemporal transformer to enhance the capture of spatial and temporal features from input videos.", + "bbox": [ + 212, + 167, + 787, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As depicted in Fig. 2, RePOSE is composed of a spatio-temporal module and two specialized loss modules. The spatial transformer encoding module processes individual body joints as distinct tokens, thereby efficiently capturing the inter-joint relationships within the same frame. In contrast, the temporal transformer encoding module considers cross-frame features, enabling it to discern relationships between joints across different frames. The outputs from these modules are then directed to the spatial and temporal relational depth consistency modules, which compute the respective losses. A crucial aspect of our architecture is the backward pass, which plays a vital role in guiding the model's training process. The nuances and operations of the spatial and temporal relational depth consistency modules are elaborated in subsequent sections.", + "bbox": [ + 212, + 287, + 787, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Spatial Relational Depth Consistency", + "text_level": 1, + "bbox": [ + 215, + 470, + 571, + 486 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In any given frame, the ideal relationship between estimated depths $\\hat{d}$ and ground-truth depths $d$ should adhere to the condition $\\hat{d}_{t,i} < \\hat{d}_{t,j}$ if $d_{t,i} < d_{t,j}$ , where $t$ denotes a specific time step, and $d_{t,i}$ and $d_{t,j}$ represent the depths of different joints $i$ and $j$ at that moment, respectively. The same applies for $\\hat{d}_{t,i}$ and $\\hat{d}_{t,j}$ . However, we observe that this relative depth relationship is often overlooked in existing methods, leading to inaccurate results, especially in scenarios with noise such as self-occlusion. To address this, we introduce a novel loss function to the training procedure, aimed at steering the network optimization in a more accurate direction. The formula is as follows:", + "bbox": [ + 212, + 492, + 491, + 777 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S} = \\sum_ {t = 1} ^ {T} \\sum_ {d _ {t, i} < d _ {t, j}} \\max \\left(0, \\hat {d} _ {t, i} - \\hat {d} _ {t, j}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 782, + 488, + 840 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8b9bd58a00290c6342b649f2a6a062516279bfb3397bfd6c37e7f89ce7dffaab.jpg", + "image_caption": [ + "Fig. 3: Overview of spatio-temporal relational depth consistency loss. The spatial loss (left) enhances relative depth relationships between the in-frame joints. The temporal loss (right) ensures the consistent relative depth relationships of joints across frames. The green dashed lines illustrate examples where the loss is calculated for the relative depth of linked joints against the ground-truth relative depth." + ], + "image_footnote": [], + "bbox": [ + 500, + 502, + 787, + 646 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z. Sun et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Eq. (2) demonstrates that our model penalizes deviations in the ranking of estimated depths from the ground-truth rankings. An illustration of this loss is displayed in the left part of Fig. 3.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In cases where body parts are occluded, standard models often struggle to accurately determine the depth positioning of joints, leading to spatial inconsistencies. By integrating our spatial depth rank loss, we guide the estimations to align more closely with the actual spatial configuration, thereby providing feedback to the model for improved training and enhanced differentiation of obscured body parts or joints. Particularly in self-occlusion scenarios, traditional approaches might incorrectly associate body joints, such as mistaking a hand for an elbow due to overlapping positions. This type of error, which might go unnoticed when solely using MPJPE as the loss function, becomes apparent when the human skeleton is visualized. Our proposed spatial depth rank loss function, by reinforcing the inter-joint relationships within a frame, effectively prevents these errors and ensures a more accurate representation of the human body structure. The code of this loss is displayed in Algorithm 1.", + "bbox": [ + 212, + 193, + 784, + 388 + ], + "page_idx": 6 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [], + "code_body": "Algorithm 1 Spatial Relational Depth Consistency \ndef process(x, J, idx): def process(x, T, idx): x = x.Unsqueeze(idx) return xrepeat_interleave(J, idx) return xrepeat_interleave(T, idx) \ndef loss_spatial(x, gt): def loss_temporal(x, gt): #x, gt: [B, T, J, 3] #x, gt: [B, T, J, 3] _, , J, = x.shape _, T, , = x.shape x = x.permute(0, 2, 1) [... , 2] gt = gt.permute(0, 2, 1) [... , 2] A = process(gt, J, -1) A = process(gt, T, -1) B = process(gt, J, -2) B = process(gt, T, -2) C = process(x, J, -1) C = process(x, T, -1) D = process(x, J, -2) D = process(x, T, -2) Comp_1 = (C - D) * (A < B) Comp_2 = torch.zeros_like(C) Comp_2 = torch.zeros_like(C) SRLoss = torch.maxmin( Comp_1, Comp_2) return torch.mean(SRLoss) return torch.mean(TRLoss)", + "bbox": [ + 212, + 407, + 799, + 699 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Temporal Relational Depth Consistency", + "text_level": 1, + "bbox": [ + 215, + 731, + 589, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In parallel with our spatial considerations, we address depth consistency in the temporal dimension. Temporally, the depth ranking of a specific joint should remain consistent across different time steps to ensure the continuity of the 3D pose sequence. However, challenges arise, particularly during self-occlusion, where the estimated poses may disrupt this continuity. To counteract this, we", + "bbox": [ + 212, + 763, + 784, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "3D Human Pose Estimation via Depth Relational Consistency", + "bbox": [ + 316, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "introduce the following loss function:", + "bbox": [ + 215, + 146, + 483, + 161 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {T} = \\sum_ {j = 1} ^ {J} \\sum_ {d _ {m, j} < d _ {n, j}} \\max \\left(0, \\hat {d} _ {m, j} - \\hat {d} _ {n, j}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 172, + 785, + 217 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $j$ indicates a specific joint, $d_{m,j}$ and $d_{n,j}$ represent the depths of joint $j$ at different times $m$ and $n$ , similarly for $\\hat{d}_{m,j}$ and $\\hat{d}_{n,j}$ . The right part of Fig. 3 shows the idea of this loss.", + "bbox": [ + 212, + 228, + 782, + 273 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The temporal depth ranking loss encourages each joint in the estimated poses to maintain temporal consistency, preventing any joint from deviating suddenly from its original trajectory within the sequence. This loss is particularly crucial in ensuring that the temporal relationships between consecutive frames in video sequences are adequately maintained, which aids in smoothing the motion and making the model less susceptible to outliers. The implementation is detailed in Algorithm 2.", + "bbox": [ + 212, + 275, + 784, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5 Final Objective", + "text_level": 1, + "bbox": [ + 215, + 404, + 390, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our comprehensive approach integrates the spatial and temporal dimensions to significantly enhance the coherence and realism of estimated 3D pose sequences. This integration ensures that the movements captured are not only spatially precise but also exhibit temporal fluidity and lifelike dynamics. Our final objective function combines the conventional MPJPE loss with our spatio-temporal relational depth consistency losses. The final loss function is formulated as follows:", + "bbox": [ + 212, + 430, + 782, + 521 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {m p j p e} + \\lambda_ {S} \\mathcal {L} _ {S} + \\lambda_ {T} \\mathcal {L} _ {T}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 536, + 785, + 554 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{L}mpjpe$ is the standard MPJPE loss, and the parameters $\\lambda_{S}$ and $\\lambda_{T}$ are weighting factors that balance the contribution of the spatial and temporal components in the overall loss function. This unified loss function is designed to optimize the network in a holistic manner, addressing both the individual joint accuracy and the relational depth in both spatial and temporal dimensions.", + "bbox": [ + 212, + 560, + 784, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 660, + 375, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 215, + 691, + 532, + 707 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have conducted training and evaluation of our model using the datasets Human3.6M [11] and MPI-INF-3DHP [20].", + "bbox": [ + 212, + 719, + 782, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Human3.6M is a well-established indoor dataset widely employed in the field of 3D HPE. This dataset comprises 3.6 million video frames captured from four distinct viewing angles, featuring performances by eleven professional actors engaged in diverse activities, including sitting, eating, and walking and more. In line with research [2,17,19,26,45,46], we utilize subjects 1, 5, 6, 7, and 8 as our training dataset, reserving subjects 9 and 11 for testing purposes.", + "bbox": [ + 212, + 750, + 784, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z. Sun et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4c22ed4ec878a98191d4e6f9a52a51fda529caf646c4ff9a1fd9c54586a8c73c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 143, + 331, + 231 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/47930450a25966683445342f95cf83f23d1037c40b33fb3ee0e3204c62896e5e.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 222, + 231, + 331, + 313 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/7a20d2a5405214e98e28acb93750bc5d6445bd313866c22e6b44eeeb40635883.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 145, + 441, + 228 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/3c25018181092d1f98b2488e994461bb62411e32e143746e6b9b4b6cd3594e09.jpg", + "image_caption": [ + "(b) GT" + ], + "image_footnote": [], + "bbox": [ + 336, + 229, + 441, + 313 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/fab57cea4bf05a10c4073afb39931eff242846a39e9a695fc13ebd5710a33319.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 447, + 145, + 553, + 228 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b78b878721b03295a4c4796ae7fc1941daa0fb7df4e91b3d25fed14a40f6a697.jpg", + "image_caption": [ + "(c) MotionBERT", + "Fig. 4: Qualitative comparison on Human3.6M. Our method is evaluated against MotionBERT [46] and PoseFormerV2 [44] in scenarios featuring extensive self-occlusion of the human body. Areas with incorrect relative depth relationships are marked with red circles for emphasis, illustrating the effectiveness of each method in handling complex occlusion challenges." + ], + "image_footnote": [], + "bbox": [ + 449, + 229, + 553, + 313 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/934db952d9968d255e3a775800581501521b0d98911e1028661f8e81831cf519.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 145, + 665, + 228 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/aaadb08ebd0dd2175e757b7cc0312a03bac03415b9c9354c780f469ea052f716.jpg", + "image_caption": [ + "(d) PoseFormerV2" + ], + "image_footnote": [], + "bbox": [ + 560, + 229, + 665, + 313 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0d19798756dd2a26c91ffe0e46824c8a6e4a63fc5bf68828598e21907b0aa15a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 145, + 779, + 228 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/5ae4b58a58bddc1fd7fdaa94c17f877c04a37b4d2b781e8b360b40f895eb1a2a.jpg", + "image_caption": [ + "(e) Ours" + ], + "image_footnote": [], + "bbox": [ + 674, + 229, + 779, + 313 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In our evaluation process, we report two key metrics: MPJPE and Procrustes Aligned-MPJPE (PA-MPJPE), both measured in millimeters. MPJPE assesses the average Euclidean distance between the estimated joint positions and their corresponding ground truth values. In contrast, the PA-MPJPE represents the MPJPE values after performing a rigid alignment procedure involving translation, rotation, and scaling operations to bring the estimated 3D pose as close as possible to the ground truth poses.", + "bbox": [ + 212, + 435, + 787, + 541 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "MPI-INF-3DHP is another large-scale 3D HPE dataset consisting of both constrained indoor and challenging outdoor scenes. It records 8 actors performing 8 activities. It consists of over 1.3 million frames captured from the 14 cameras. To rigorously assess the generalization capabilities of RePOSE, particularly in challenging outdoor environments and scenarios involving occlusions, we conducted a comprehensive evaluation of our model using the MPI-INF-3DHP dataset. Here we utilize 2D pose sequences with a length of 81 frames each as input, as distinct from what it is in Human3.6M. Following previous works [2, 32, 44], we report MPJPE, Percentage of Correct Keypoints (PCK) within the $150\\mathrm{mm}$ range as well as Area Under the Curve (AUC).", + "bbox": [ + 212, + 541, + 787, + 693 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 710, + 455, + 726 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We have implemented our methodology using PyTorch [24] on a computing platform consisting of two NVIDIA RTX 3090 GPUs. Following previous studies [29, 44, 46], we have configured the video sequence length to 243 frames and leveraged the Stacked Hourglass [23] for the extraction of 2D poses during both training and evaluation phases. Additionally, we have applied horizontal flipping augmentation, consistent with [1, 2, 46]. For the optimization of model parameters, we have employed the AdamW [18] optimizer for 100 training epochs,", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "3D Human Pose Estimation via Depth Relational Consistency", + "bbox": [ + 316, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/323ae3a81d550612becd749e720e409a35b85352885d75f1386b9afb3389db78.jpg", + "table_caption": [ + "Table 1: Quantitative Comparison on Human3.6M Dataset. The top and bottom show comparison based on MPJPE and Procrustes-Aligned MPJPE in millimeters respectively. In this table, $T$ represents the number of input frames used in each method. The best-performing methods are highlighted in bold, while the second-best results are underlined for clarity. Different 2D detection methods are indicated as follows: (*) denotes the use of CPN [3], ( $\\dagger$ ) indicates the use of HRNet [33], and ( $\\ddagger$ ) signifies the use of SH [23] for 2D pose detection." + ], + "table_footnote": [], + "table_body": "
MPJPETDir. Disc. Eat. GreetPhonePhotoPosePurch. Sit SitD.SmokeWaitWalkD.WalkWalkT.
*VideoPose3D (CVPR'19) [26]24345.246.743.345.648.155.144.644.357.365.847.144.049.032.833.9
*PoseFormer (ICCV'21) [45]8141.544.839.842.546.551.642.142.053.360.745.543.346.131.832.2
*MHFormer (CVPR'22) [14]35139.243.140.140.944.951.240.641.353.560.343.741.143.829.830.6
†MixSTE (CVPR'22) [42]24336.739.036.539.440.244.939.836.947.954.839.637.839.329.730.6
*P-STMO (ECCV'22) [27]24338.442.139.840.245.248.940.438.353.857.343.941.642.229.329.3
*PoseFormerV2 (CVPR'23) [44]24341.245.441.444.046.653.742.642.655.164.645.742.845.732.332.8
*STCFformer (CVPR'23) [30]24338.441.236.838.042.750.538.738.252.556.841.838.440.226.227.7
*D3DP (ICCV'23) [28]24337.739.935.738.241.948.839.538.350.553.941.639.439.827.427.5
‡MotionBERT (ICCV'23) [46]24336.137.535.832.140.346.336.135.346.953.939.536.335.825.125.3
‡MoAGFormer (WACV'24) [29]24336.838.535.933.041.148.638.034.849.051.440.337.436.327.227.2
‡RePOSE (Ours)24334.636835.031.338.844.735.335.547.550.938.435.534.923.925.036.5
PA-MPJPETDir. Disc. Eat. GreetPhonePhotoPosePurch. Sit SitD.SmokeWaitWalkD.WalkWalkT.
*VideoPose3D (CVPR'19) [26]24334.136.134.437.236.442.234.433.645.052.537.433.837.825.627.3
*PoseFormer (ICCV'21) [45]8132.534.832.634.635.339.532.132.042.848.534.832.435.324.526.0
*MHFormer (CVPR'22) [14]35131.534.932.833.635.339.632.032.243.548.736.432.634.323.925.1
*MixSTE (CVPR'22) [42]24330.833.130.331.833.139.131.130.542.544.534.030.832.722.122.9
*P-STMO (ECCV'22) [27]24331.335.232.933.935.439.332.531.544.648.236.332.934.423.823.9
*PoseFormerV2 (CVPR'23) [44]24332.335.833.735.835.941.133.132.744.251.937.432.835.525.226.5
*STCFformer (CVPR'23) [30]24329.333.030.730.632.738.229.728.842.245.033.329.431.520.922.3
*D3DP (ICCV'23) [28]24330.632.529.131.031.937.630.329.440.643.633.330.531.421.522.4
‡MotionBERT (ICCV'23) [46]24329.531.332.027.032.837.229.829.641.048.234.330.130.221.221.8
‡MoAGFormer (WACV'24) [29]24331.032.631.027.934.038.731.530.041.445.434.830.831.322.823.2
‡RePOSE (Ours)24329.130.929.426.632.135.828.630.639.944.933.329.029.620.121.2
", + "bbox": [ + 217, + 266, + 784, + 554 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "incorporating a weight decay factor of 0.01. The initial learning rate has been set to $5 \\times 10^{-4}$ , accompanied by a learning rate decay strategy with a decay rate of 0.99. Lastly, the hyper-parameters $\\lambda_{S}$ and $\\lambda_{T}$ are all set to 20. When training our model on Human3.6M, We utilize the stride data sample strategy with a one-third interval of the input length to make sure there are fewer overlapping frames between consecutive sequences. But for the MPI-INF-3DHP dataset, the interval is one-nine of the input length because of fewer samples.", + "bbox": [ + 212, + 585, + 787, + 694 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 214, + 719, + 620, + 734 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Human3.6M dataset. The proposed method has been seriously compared with state-of-the-art (SOTA) approaches, utilizing the Human3.6M dataset as the evaluation benchmark. As illustrated in Tab. 1, our method demonstrates exceptional performance, achieving an MPJPE of 36.5 millimeters and a PA-MPJPE of 30.7 millimeters. These results surpass those achieved by leading SOTA methods, indicating a notable performance improvement.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z. Sun et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/46b7cec7a8be7cfc84e78cca1e8f834f77c7d430cadec3fdcc4ef2619fe6d934.jpg", + "table_caption": [ + "Table 2: Computational Complexity Comparison. The MPJPEs here are evaluated on Human3.6M using ground-truth 2D joint locations as input." + ], + "table_footnote": [], + "table_body": "
MethodParam (M)MACs (G)MPJPE
PoseFormer (ICCV'21) [45]9.52.531.3
MHFormer (CVPR'22) [14]30.97.030.5
MixSTE (CVPR'22) [42]33.6139.025.9
P-STMO (ECCV'22) [27]6.20.729.3
PoseFormerV2 (CVPR'23) [44]14.30.5-
STCFormer (CVPR'23) [30]4.719.621.3
MotionBERT (ICCV'23) [46]42.5174.717.8
MotionAGFormer (WACV'24) [29]19.078.317.3
RePOSE (Ours)16.043.715.7
", + "bbox": [ + 285, + 184, + 712, + 330 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/be18d433d79bd0a8b944f5a5fc25e49d914f536739bea6164dffd6f63293e16b.jpg", + "table_caption": [ + "Table 3: Quantitative comparison on MPI-INF-3DHP dataset. The best and second-best results are bolded and underlined." + ], + "table_footnote": [], + "table_body": "
MethodPCK↑AUC↑MPJPE↓
VideoPose3D (CVPR'19) [26]85.551.584.8
PoseFormer (ICCV'21) [45]65.463.257.7
MHFormer (CVPR'22) [14]93.863.358.0
MixSTE (CVPR'22) [42]94.466.554.9
P-STMO (ECCV'22) [27]97.975.832.2
D3DP (ICCV'23) [28]97.777.830.2
PoseFormerV2 (CVPR'23) [44]97.978.827.8
STCFormaler (CVPR'23) [30]98.783.923.1
MotionAGFormer (WACV'24) [29]98.285.316.2
RePOSE (Ours)98.386.715.5
", + "bbox": [ + 292, + 400, + 710, + 550 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "It is worth noting that even when considering the best results reported by each of the prior studies, our method consistently outperforms them, especially in challenging scenarios involving actions such as \"photo,\" \"sitting,\" and \"sitting down.\" This highlights its potential to effectively address challenges related to occlusion. Qualitative results can be found in Fig. 4.", + "bbox": [ + 212, + 594, + 784, + 669 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We also compare the number of parameters, multiply-accumulate operations (MACs), and MPJPE using the ground-truth 2D pose sequences as input, as shown in Tab. 2. Our model attains the best MPJPE out of other works with limited parameters and algorithm complexity, demonstrating the effectiveness and efficiency of our proposed method.", + "bbox": [ + 212, + 671, + 784, + 746 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "MPI-INF-3DHP dataset. Tab. 3 presents a quantitative comparison of our method on the MPI-INF-3DHP dataset. RePOSE achieves impressive results, with an MPJPE of $15.5\\mathrm{mm}$ and an Area Under Curve (AUC) of $86.7\\%$ , significantly outperforming other compared methods by margins of $1.8\\mathrm{mm}$ in MPJPE and $1.4\\%$ in AUC. However, in terms of the Percentage of Correct Keypoints (PCK), RePOSE reaches $98.3\\%$ , which is slightly lower by $0.4\\%$ compared", + "bbox": [ + 212, + 750, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "3D Human Pose Estimation via Depth Relational Consistency", + "bbox": [ + 316, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/5f3f5455109d303d89d43c8c7df680faff1e13a8c0d177ee89254d5af257149f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 142, + 331, + 229 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3a24fc725daba22553b8a94546af8cffd494a94e06b272a28ba4d0f88d1b4a69.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 222, + 229, + 331, + 313 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/02100a0d79ab8aa956b1a9eab01929261c359d6977c1ef2268f446ca8e3b41e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 146, + 439, + 227 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/39863b86a363cbefee185735abc617c38da40bb3e1932ca6b964aac1db9145b0.jpg", + "image_caption": [ + "(b) GT" + ], + "image_footnote": [], + "bbox": [ + 338, + 229, + 439, + 311 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c7970a2ed7a0fcd5e858ff79e1105a51df62fcff18c9efe6c2d58b3703e58f51.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 146, + 552, + 228 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b843a7cab61e261629d0288b08e759075274dc00f6723826f7c7a523ca973d77.jpg", + "image_caption": [ + "(c) D3DP", + "Fig. 5: Qualitative comparison on MPI-INF-3DHP. Our method is compared with D3DP [28] and STCFoer [30]. The areas with incorrect results are pointed out with red arrows and circles to exemplify the importance of relational depth consistency." + ], + "image_footnote": [], + "bbox": [ + 450, + 229, + 553, + 311 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d5141b30a094ea014d444f13a2f8239c7714756d77904571db0d4c8dc040b4fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 146, + 663, + 228 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8912d8606a95cf2b454ce29f8e737086dc0aae73f95dbb8db9f185801f607281.jpg", + "image_caption": [ + "(d) STCFormer" + ], + "image_footnote": [], + "bbox": [ + 563, + 229, + 663, + 311 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/34d63eba5b5b4b6f98cd66b1215985b913750cb602de1205886d5070af862133.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 147, + 777, + 228 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6b2543511f8dd399d8c942ad6c66f1308fe23b25a082bf1cb32f4bbdeeb98e5f.jpg", + "image_caption": [ + "(e) Ours" + ], + "image_footnote": [], + "bbox": [ + 674, + 229, + 777, + 311 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "to the best performance achieved by existing methods, but we excel in the best balance of all metrics. For a more detailed visual assessment, refer to the qualitative comparisons illustrated in Fig. 5.", + "bbox": [ + 212, + 411, + 784, + 455 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 479, + 388, + 496 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To evaluate the distinct contributions of individual components, we incorporated spatial and temporal relational depth consistency loss functions separately into the baseline network. As shown in Tab. 4, it is clear that without these spatial and temporal depth constraints, the baseline model experiences degradation in both the MPJPE and PA-MPJPE metrics. However, the inclusion of either the spatial or temporal loss functions results in a significant enhancement in the model's performance. This observation underscores the effectiveness of these two", + "bbox": [ + 212, + 507, + 491, + 733 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "loss functions in improving the overall quality of the model's predictions.", + "bbox": [ + 212, + 734, + 738, + 748 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Fig. 6 provides a comparative visualization of the performance of the RepoSE with and without the incorporation of two relational depth consistency losses. It is evident from the figure that the omission of these losses unfavorably affects the model's capacity to accurately estimate occluded body parts, as exemplified by the red circles. This inadequacy suggests that the model, without these losses, cannot effectively identify the relative depth of body parts during", + "bbox": [ + 212, + 750, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 4: Ablation Study on Individual Loss Functions. The study reveals that while both spatial and temporal loss functions individually contribute to improved pose estimation, their combined application yields the best results. Performance metrics have been evaluated on the Human3.6M dataset to substantiate this finding.", + "bbox": [ + 496, + 522, + 787, + 633 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/0f9f08c592643690695593f1d759e80560af84a5a020c45314a2e8115c0cac75.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MPJPE Spatial TemporalMPJPE↓PA-MPJPE↓
✓--37.932.5
✓✓-36.931.4
✓-✓37.231.8
✓✓✓36.530.7
", + "bbox": [ + 503, + 657, + 782, + 719 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z. Sun et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/0b32b0329018eb45d3a07e32ac68d60a52ef76cc6bcad2f70a6e9eb4d488ef1a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 143, + 372, + 241 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8e1bd81571567d455c0c0b1e3ad64ddd0f3142de0511a4fcc828b8c0c3136744.jpg", + "image_caption": [ + "(a) Input", + "Fig. 6: Visualized Results of Ablation Study. The second column displays poses generated by a model trained without the spatio-temporal relational depth consistency loss, contrasting with the third column, which shows results from the model including this loss. Joints with less accurate estimations are highlighted in red circles for clarity." + ], + "image_footnote": [], + "bbox": [ + 243, + 242, + 372, + 340 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/6578cb4f6549cacd34706e677dbc302c5522d31096f67e70da93888d6e198cd1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 148, + 500, + 239 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e8fbbfd8791ed3e1bb247458d4f2e5de897c0bc99dc899c2e0a50b4762ff0cea.jpg", + "image_caption": [ + "(b) w/o our losses" + ], + "image_footnote": [], + "bbox": [ + 383, + 244, + 500, + 339 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2af1bcb8aa2e34e1f78294d691e07d32f2aa12beef27932ed505a995f0d6d30f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 148, + 629, + 241 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e3cd59544a8277d065ee6fd435d4ac60cebcfa4354de088be5bba109ba1b80b9.jpg", + "image_caption": [ + "(c) RePOSE" + ], + "image_footnote": [], + "bbox": [ + 511, + 244, + 627, + 337 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/880e335368efc8589ef1c7eeba1669d2dcd5f66f8f706455b8f2eb052bed2a16.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 148, + 756, + 241 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/693e73f2ca705f1f3da685dca86cee04166f884acbd0aa3e87f91ee7857eda80.jpg", + "image_caption": [ + "(d) Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 640, + 244, + 756, + 339 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/068457d406f621e08b77dd027f1bc79520c42b10d846c80c557686a81be88f24.jpg", + "table_caption": [ + "Table 5: Generalization evaluation on Human3.6M with our loss functions. With the integration of our loss functions, all methods exhibit improved performance. $\\Delta$ indicates the difference between 'w/ST' and 'w/o'." + ], + "table_footnote": [], + "table_body": "
MethodTMPJPE↓PA-MPJPE↓
w/Sw/Tw/STw/oΔw/Sw/Tw/STw/oΔ
P-STMO-S [27]2745.345.544.746.1-1.436.436.635.636.8-1.2
STCFformer [30]2743.843.843.344.1-0.834.134.033.734.8-1.1
MotionBERT [46]24337.037.236.737.5-0.831.431.730.831.8-1.0
", + "bbox": [ + 235, + 503, + 769, + 583 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "the training stage, leading to incorrect results. On the contrary, the integration of our proposed losses into the training phase significantly enhances the model's ability to resolve occlusions, thereby improving its estimation accuracy. This improvement is crucial for applications where precise depth perception is necessary, and it emphasizes the value of our contributions to the field of pose estimation in occlusion situations.", + "bbox": [ + 212, + 623, + 787, + 713 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Generalization Evaluation", + "text_level": 1, + "bbox": [ + 215, + 738, + 477, + 752 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To validate the versatility and generalization capacity of our proposed loss functions, we extend their application to similar spatio-temporal transformer-based frameworks, including MotionBERT [46], STCFoer [30], and P-STMO [27]. We directly use their official pre-trained models with default setups and fine-tune them with our loss functions.", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "3D Human Pose Estimation via Depth Relational Consistency", + "bbox": [ + 316, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We detail the generalizability of our individual loss functions in Table 5. Overall, integrating both losses ('w/ST') leads to superior performance metrics compared to the original model configurations without our adjustments. Significantly, even when applied separately, each loss function ('w/S' or 'w/T') yields improvements across all evaluated methods, reinforcing their broad applicability. It is important to note that all reported results were achieved without specific tuning of the hyperparameters $\\lambda_{S}$ and $\\lambda_{T}$ . Therefore, there is an opportunity for further enhancements if these parameters were to be systematically optimized.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 215, + 297, + 509, + 314 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduce RePOSE, an approach designed to enhance 3D pose estimation accuracy in video sequences, specifically under conditions where body parts are obscured or occluded. Diverging from conventional methods that predominantly depend on depth information, which can falter when body parts are not visible, RePOSE adopts an innovative strategy centered on the relative positioning of body segments. This methodology not only exhibits superior efficacy in managing occlusions but also boasts a straightforward implementation, requiring minimal code adjustments. Our comprehensive experiments validate that RePOSE outperforms existing state-of-the-art techniques, particularly in scenarios marked by obstructions. These outcomes underscore RePOSE's efficacy and its promising potential in advancing 3D pose estimation, especially in practical settings where occlusions and partial visibilities are common. Additionally, the simplicity of its integration renders RePOSE a viable and effective solution for researchers aiming to refine pose estimation precision.", + "bbox": [ + 212, + 338, + 787, + 549 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. Despite its significant advantages in occlusion scenarios within 3D human pose estimation, RePOSE encounters limitations, chiefly its dependence on the accuracy of preliminary 2D pose data. Discrepancies in this foundational data can adversely affect subsequent 3D estimations, a challenge inherent in 2D-to-3D conversion techniques. Moreover, although tailored for occlusions, RePOSE's performance may diminish in complex conditions, such as atypical body postures or interactions with objects, and in extremely low-light environments where depth cues are compromised.", + "bbox": [ + 212, + 551, + 787, + 672 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Future Work. The efficacy of our newly proposed loss functions has been established across various spatio-temporal transformer architectures, including MotionBERT [46] and STCFoer [30]. Our next phase of research will explore the extension of these functions to a wider spectrum of models, including those based on GNNs and CNNs. Additionally, we aim to evolve the hyperparameters $\\lambda_{S}$ and $\\lambda_{T}$ from fixed values based on empirical selection to adaptive, learnable parameters optimized through the model's interaction with input data, thereby refining the model's depth dimension understanding in 3D human pose estimation. Future enhancements will focus on augmenting RePOSE's versatility across different input data forms and broadening its efficacy in a more extensive array of challenging conditions.", + "bbox": [ + 212, + 672, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z. Sun et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This project is supported by the Guangdong Basic and Applied Basic Research Foundation under project No. 2024A1515011995, Guangdong Natural Science Funds for Distinguished Young Scholar under Grant 2023B1515020097, and National Research Foundation Singapore under the AI Singapore Programme under Grant AISG3-GV-2023-011.", + "bbox": [ + 212, + 176, + 787, + 253 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 277, + 321, + 292 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Cai, Y., Ge, L., Liu, J., Cai, J., Cham, T.J., Yuan, J., Thalmann, N.M.: Exploiting spatial-temporal relationships for 3d pose estimation via graph convolutional networks. In: ICCV. pp. 2272-2281 (2019) 3, 9", + "2. Chen, T., Fang, C., Shen, X., Zhu, Y., Chen, Z., Luo, J.: Anatomy-aware 3d human pose estimation with bone-based pose decomposition. IEEE TCSVT 32(1), 198-209 (2021) 2, 3, 8, 9", + "3. Chen, Y., Wang, Z., Peng, Y., Zhang, Z., Yu, G., Sun, J.: Cascaded pyramid network for multi-person pose estimation. In: CVPR. pp. 7103-7112 (2018) 3, 10", + "4. Chen, Y., Tu, Z., Ge, L., Zhang, D., Chen, R., Yuan, J.: So-handnet: Self-organizing network for 3d hand pose estimation with semi-supervised learning. In: ICCV. pp. 6961-6970 (2019) 1", + "5. Ci, H., Wang, C., Ma, X., Wang, Y.: Optimizing network structure for 3d human pose estimation. In: ICCV. pp. 2262-2271 (2019) 3", + "6. Ci, H., Wu, M., Zhu, W., Ma, X., Dong, H., Zhong, F., Wang, Y.: Gfpose: Learning 3d human pose prior with gradient fields. In: CVPR. pp. 4800-4810 (2023) 3", + "7. Fang, H.S., Xie, S., Tai, Y.W., Lu, C.: Rumpe: Regional multi-person pose estimation. In: ICCV. pp. 2334-2343 (2017) 3", + "8. Gong, J., Fan, Z., Ke, Q., Rahmani, H., Liu, J.: Meta agent teaming active learning for pose estimation. In: CVPR. pp. 11079-11089 (2022) 1", + "9. Hossain, M.R.I., Little, J.J.: Exploiting temporal information for 3d human pose estimation. In: ECCV. pp. 68-84 (2018) 3, 4", + "0. Huang, S., Wang, W., He, S., Lau, R.W.: Egocentric temporal action proposals. IEEE TIP 27(2), 764-777 (2017) 1", + "1. Ionescu, C., Papava, D., Olaru, V., Sminchisescu, C.: Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE TPAMI 36(7), 1325-1339 (2013) 3, 8", + "2. Jiang, Y., Zhou, Y., Liang, Y., Liu, W., Jiao, J., Quan, Y., He, S.: Diffuse3d: Wide-angle 3d photography via bilateral diffusion. In: ICCV. pp. 8998-9008 (2023) 1", + "3. Li, W., Liu, H., Ding, R., Liu, M., Wang, P., Yang, W.: Exploiting temporal contexts with strided transformer for 3d human pose estimation. IEEE Transactions on Multimedia 25, 1282-1293 (2022) 2, 3, 5", + "4. Li, W., Liu, H., Tang, H., Wang, P., Van Gool, L.: Mhformer: Multi-hypothesis transformer for 3d human pose estimation. In: CVPR. pp. 13147-13156 (2022) 2, 3, 5, 10, 11", + "5. Lin, K., Wang, L., Liu, Z.: End-to-end human pose and mesh reconstruction with transformers. In: CVPR. pp. 1954-1963 (2021) 2", + "6. Lin, M., Lin, L., Liang, X., Wang, K., Cheng, H.: Recurrent 3d pose sequence machines. In: CVPR. pp. 810-819 (2017) 3" + ], + "bbox": [ + 225, + 309, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "3D Human Pose Estimation via Depth Relational Consistency", + "bbox": [ + 316, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Liu, R., Shen, J., Wang, H., Chen, C., Cheung, S.c., Asari, V.: Attention mechanism exploits temporal contexts: Real-time 3d human pose reconstruction. In: CVPR. pp. 5064-5073 (2020) 2, 3, 8", + "18. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019) 9", + "19. Martinez, J., Hossain, R., Romero, J., Little, J.J.: A simple yet effective baseline for 3d human pose estimation. In: ICCV. pp. 2640-2649 (2017) 3, 8", + "20. Mehta, D., Rhodin, H., Casas, D., Fua, P., Sotnychenko, O., Xu, W., Theobalt, C.: Monocular 3d human pose estimation in the wild using improved cnn supervision. In: 2017 international conference on 3D vision (3DV). pp. 506-516. IEEE (2017) 3, 8", + "21. Mehta, D., Sridhar, S., Sotnychenko, O., Rhodin, H., Shafiei, M., Seidel, H.P., Xu, W., Casas, D., Theobalt, C.: Vnct: Real-time 3d human pose estimation with a single rgb camera. ACM TOG 36(4), 1-14 (2017) 1", + "22. Moon, G., Lee, K.M.: I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In: ECCV. pp. 752-768. Springer (2020) 2", + "23. Newell, A., Yang, K., Deng, J.: Stacked hourglass networks for human pose estimation. In: ECCV. pp. 483-499. Springer (2016) 9, 10", + "24. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch. In: NeurIPS (2017) 9", + "25. Pavlakos, G., Zhou, X., Daniilidis, K.: Ordinal depth supervision for 3d human pose estimation. In: CVPR. pp. 7307-7316 (2018) 2, 4", + "26. Pavlo, D., Feichtenhofer, C., Grangier, D., Auli, M.: 3d human pose estimation in video with temporal convolutions and semi-supervised training. In: CVPR. pp. 7753-7762 (2019) 3, 4, 8, 10, 11", + "27. Shan, W., Liu, Z., Zhang, X., Wang, S., Ma, S., Gao, W.: P-stmo: Pre-trained spatial temporal many-to-one model for 3d human pose estimation. In: ECCV. pp. 461-478. Springer (2022) 2, 3, 5, 6, 10, 11, 13", + "28. Shan, W., Liu, Z., Zhang, X., Wang, Z., Han, K., Wang, S., Ma, S., Gao, W.: Diffusion-based 3d human pose estimation with multi-hypothesis aggregation. In: ICCV. pp. 14761-14771 (October 2023) 10, 11, 12", + "29. Soroush Mehraban, Vida Adeli, B.T.: Motionagformer: Enhancing 3d human pose estimation with a transformer-gcnformer network. In: WACV (2024) 4, 9, 10, 11", + "30. Tang, Z., Qiu, Z., Hao, Y., Hong, R., Yao, T.: 3d human pose estimation with spatio-temporal criss-cross attention. In: CVPR. pp. 4790-4799 (2023) 10, 11, 12, 13, 14", + "31. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017) 2", + "32. Wang, J., Yan, S., Xiong, Y., Lin, D.: Motion guided 3d pose estimation from videos. In: ECCV. pp. 764-780. Springer (2020) 3, 9", + "33. Wang, J., Sun, K., Cheng, T., Jiang, B., Deng, C., Zhao, Y., Liu, D., Mu, Y., Tan, M., Wang, X., et al.: Deep high-resolution representation learning for visual recognition. IEEE TPAMI 43(10), 3349-3364 (2020) 3, 10", + "34. Xu, Y., Han, C., Qin, J., Xu, X., Han, G., He, S.: Transductive zero-shot action recognition via visually connected graph convolutional networks. IEEE TNNLS 32(8), 3761-3769 (2020) 1", + "35. Yan, S., Xiong, Y., Lin, D.: Spatial temporal graph convolutional networks for skeleton-based action recognition. In: AAAI. vol. 32 (2018) 1" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z. Sun et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "36. Yang, X., Xu, K., Chen, S., He, S., Yin, B.Y., Lau, R.: Active matting. NeurIPS 31 (2018) 1", + "37. Ye, M., Li, H., Du, B., Shen, J., Shao, L., Hoi, S.C.: Collaborative refining for person re-identification with label noise. IEEE TIP 31, 379-391 (2021) 1", + "38. Yoon, J.S., Liu, L., Golyanik, V., Sarkar, K., Park, H.S., Theobalt, C.: Pose-guided human animation from a single image in the wild. In: CVPR. pp. 15039-15048 (2021) 1", + "39. Yu, B.X., Zhang, Z., Liu, Y., Zhong, S.h., Liu, Y., Chen, C.W.: Gla-gcn: Global-local adaptive graph convolutional network for 3d human pose estimation from monocular video. In: ICCV. pp. 8818-8829 (2023) 3", + "40. Zhang, C., Yang, T., Weng, J., Cao, M., Wang, J., Zou, Y.: Unsupervised pretraining for temporal action localization tasks. In: CVPR. pp. 14031-14041 (2022) 1", + "41. Zhang, J., Ye, G., Tu, Z., Qin, Y., Qin, Q., Zhang, J., Liu, J.: A spatial attentive and temporal dilated (satd) gcn for skeleton-based action recognition. CAAI Transactions on Intelligence Technology 7(1), 46-55 (2022) 1", + "42. Zhang, J., Tu, Z., Yang, J., Chen, Y., Yuan, J.: Mixste: Seq2seq mixed spatiotemporal encoder for 3d human pose estimation in video. In: CVPR. pp. 13232-13242 (2022) 2, 3, 4, 5, 10, 11", + "43. Zhao, L., Peng, X., Tian, Y., Kapadia, M., Metaxas, D.N.: Semantic graph convolutional networks for 3d human pose regression. In: CVPR. pp. 3425-3435 (2019) 3", + "44. Zhao, Q., Zheng, C., Liu, M., Wang, P., Chen, C.: Poseformerv2: Exploring frequency domain for efficient and robust 3d human pose estimation. In: CVPR. pp. 8877-8886 (2023) 2, 3, 5, 9, 10, 11", + "45. Zheng, C., Zhu, S., Mendieta, M., Yang, T., Chen, C., Ding, Z.: 3d human pose estimation with spatial and temporal transformers. In: ICCV. pp. 11656-11665 (2021) 2, 3, 5, 6, 8, 10, 11", + "46. Zhu, W., Ma, X., Liu, Z., Liu, L., Wu, W., Wang, Y.: Motionbert: A unified perspective on learning human motion representations. In: ICCV. pp. 15085-15099 (2023) 2, 3, 4, 5, 8, 9, 10, 11, 13, 14" + ], + "bbox": [ + 212, + 146, + 787, + 575 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "3D Human Pose Estimation via Depth Relational Consistency", + "bbox": [ + 316, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/RePOSE_ 3D Human Pose Estimation via Spatio-Temporal Depth Relational Consistency/7257623d-d6ee-4c47-b288-fda8cfd3af7e_model.json b/2024/RePOSE_ 3D Human Pose Estimation via Spatio-Temporal Depth Relational Consistency/7257623d-d6ee-4c47-b288-fda8cfd3af7e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e4d150379f28a27b613f5a916a6516af7c43ed99 --- /dev/null +++ b/2024/RePOSE_ 3D Human Pose Estimation via Spatio-Temporal Depth Relational Consistency/7257623d-d6ee-4c47-b288-fda8cfd3af7e_model.json @@ -0,0 +1,2643 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.232, + 0.141, + 0.773, + 0.187 + ], + "angle": 0, + "content": "RePOSE: 3D Human Pose Estimation via Spatio-Temporal Depth Relational Consistency" + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.212, + 0.775, + 0.245 + ], + "angle": 0, + "content": "Ziming Sun\\(^{1,2}\\), Yuan Liang\\(^{1,2\\star}\\), Zejun Ma\\(^{3}\\), Tianle Zhang\\(^{1,2}\\), Linchao Bao\\(^{4}\\), Guiqing Li\\(^{1}\\), and Shengfeng He\\(^{2(\\boxtimes)}\\)" + }, + { + "type": "text", + "bbox": [ + 0.341, + 0.255, + 0.66, + 0.269 + ], + "angle": 0, + "content": "\\(^{1}\\) South China University of Technology, China" + }, + { + "type": "text", + "bbox": [ + 0.341, + 0.269, + 0.661, + 0.284 + ], + "angle": 0, + "content": "2 Singapore Management University, Singapore" + }, + { + "type": "list", + "bbox": [ + 0.341, + 0.255, + 0.661, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.416, + 0.285, + 0.588, + 0.298 + ], + "angle": 0, + "content": "shengfenghe@smu.edu.sg" + }, + { + "type": "text", + "bbox": [ + 0.389, + 0.297, + 0.614, + 0.311 + ], + "angle": 0, + "content": "3 Johns Hopkins University, USA" + }, + { + "type": "text", + "bbox": [ + 0.417, + 0.311, + 0.588, + 0.324 + ], + "angle": 0, + "content": "Tencent AI Lab, China" + }, + { + "type": "text", + "bbox": [ + 0.361, + 0.326, + 0.642, + 0.339 + ], + "angle": 0, + "content": "https://github.com/StupidAdam/RePOSE" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.377, + 0.744, + 0.655 + ], + "angle": 0, + "content": "Abstract. We introduce RePOSE, a simple yet effective approach for addressing occlusion challenges in the learning of 3D human pose estimation (HPE) from videos. Conventional approaches typically employ absolute depth signals as supervision, which are adept at discernible keypoints but become less reliable when keypoints are occluded, resulting in vague and inconsistent learning trajectories for the neural network. RePOSE overcomes this limitation by introducing spatio-temporal relational depth consistency into the supervision signals. The core rationale of our method lies in prioritizing the precise sequencing of occluded keypoints. This is achieved by using a relative depth consistency loss that operates in both spatial and temporal domains. By doing so, RePOSE shifts the focus from learning absolute depth values, which can be misleading in occluded scenarios, to relative positioning, which provides a more robust and reliable cue for accurate pose estimation. This subtle yet crucial shift facilitates more consistent and accurate 3D HPE under occlusion conditions. The elegance of our core idea lies in its simplicity and ease of implementation, requiring only a few lines of code. Extensive experiments validate that RePOSE not only outperforms existing state-of-the-art methods but also significantly enhances the robustness and precision of 3D HPE in challenging occluded environments." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.668, + 0.741, + 0.683 + ], + "angle": 0, + "content": "Keywords: 3D human pose estimation \\(\\cdot\\) depth relational loss functions" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.709, + 0.376, + 0.725 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.741, + 0.788, + 0.817 + ], + "angle": 0, + "content": "The estimation of 3D human pose from videos is a foundational task in computer vision, with widespread implications across various applications such as virtual reality [4, 12, 21, 38], human-computer interaction [8, 36, 37], and action recognition [10, 34, 35, 40, 41]. This task involves localizing body joints in a three-dimensional space to create a comprehensive representation of the human body," + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.599, + 0.84 + ], + "angle": 0, + "content": "* The first two authors contributed equally to this work." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "Z. Sun et al." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.144, + 0.334, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.23, + 0.333, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.316, + 0.306, + 0.328 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.341, + 0.146, + 0.445, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.341, + 0.23, + 0.446, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.367, + 0.316, + 0.413, + 0.328 + ], + "angle": 0, + "content": "(b) GT" + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.147, + 0.557, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.23, + 0.556, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.452, + 0.316, + 0.553, + 0.328 + ], + "angle": 0, + "content": "(c) MotionBERT" + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.147, + 0.67, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.23, + 0.67, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.316, + 0.67, + 0.328 + ], + "angle": 0, + "content": "(d) PoseFormerV2" + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.147, + 0.783, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.23, + 0.783, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.316, + 0.753, + 0.328 + ], + "angle": 0, + "content": "(e) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.339, + 0.789, + 0.409 + ], + "angle": 0, + "content": "Fig. 1: We introduce RePOSE, a novel solution designed to address the occlusion challenges in 3D pose estimation through relational supervision. Unlike existing methods (c) [46] and (d) [44], which struggle with occlusions due to their dependence on absolute depth signals, our approach leverages relational depth consistency, effectively overcoming these obstacles." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.788, + 0.564 + ], + "angle": 0, + "content": "typically in the form of a skeletal model. The methodologies in 3D human pose estimation (HPE) are primarily categorized into two approaches: direct estimation and 2D-to-3D lifting methods. The former [22, 25] entails the direct estimation of human pose from provided videos, whereas the latter [2, 14, 17, 27, 42, 44-46] involves the initial estimation of intermediate 2D poses using off-the-shelf 2D detectors, followed by lifting them to 3D space. 2D-to-3D lifting methods have gained prominence due to the rapid evolution of 2D pose detection technologies, exhibiting superior performance in many scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.789, + 0.764 + ], + "angle": 0, + "content": "Recent advancements in 2D-to-3D lifting methods often employ Transformer architectures [31], which have shown exceptional ability in capturing global information and modeling sequential data. These methods typically involve transforming detected 2D poses into 3D space using tailored algorithms [13-15,27,42,44-46]. However, this process inherently faces a many-to-one mapping challenge, where multiple distinct 3D poses can correspond to a single 2D projection, particularly under occlusion scenarios, as shown in Fig. 1c and 1d. In such cases, occlusion of key body joints leads to reduced accuracy in 2D pose estimations, creating ambiguities in the 3D reconstruction process. Moreover, relying on these unreliable occluded signals during model training results in vague and uncertain learning pathways. This limitation chiefly stems from the methods' reliance on absolute depth signals, which lose reliability under occluded conditions, thus undermining the overall accuracy of 3D pose estimation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.841 + ], + "angle": 0, + "content": "To address this challenge, we propose RePOSE, a novel approach specifically designed to enhance the accuracy of 3D HPE in occluded scenarios. RePOSE circumvents the limitations of absolute depth reliance by introducing a spatiotemporal relational depth consistency. Our method emphasizes the correct sequencing of occluded keypoints using a relative depth consistency loss in both" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.317, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "3D Human Pose Estimation via Depth Relational Consistency" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "spatial and temporal domains, rather than depending solely on absolute depth values. This shift in approach enables our method to robustly handle occlusions, providing more accurate and reliable pose estimations, as shown in Fig. 1e. Additionally, the simplicity of spatio-temporal relational depth consistency, which can be implemented in just a few lines of code, makes it a practical and effective solution for real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.788, + 0.358 + ], + "angle": 0, + "content": "We demonstrate through extensive experiments on the Human3.6M [11] and MPI-INF-3DHP [20] benchmarks that our method outperforms state-of-the-art methods by a large margin, particularly excelling in scenarios involving occlusions. Our findings are further substantiated through visual comparisons and an ablation study, showcasing the effectiveness and innovation of RePOSE in tackling one of the most challenging aspects of 3D human pose estimation. This method lays the groundwork for future research aimed at enhancing pose estimation accuracy and reliability in complex environments." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.384, + 0.396, + 0.399 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.788, + 0.688 + ], + "angle": 0, + "content": "2D-to-3D Lifting HPE. The prevalent approach in 3D HPE involves lifting 2D pose sequences to 3D poses, a method enhanced by advanced 2D detectors like CPN [3], AlphaPose [7], and HRNet [33]. These detectors facilitate the conversion process with various techniques, including Fully Connected Networks (FCN) [6,19], Long Short-Term Memory (LSTM) [9,16,32], Graph Convolutional Networks (GCN) [1,5,39,43], Temporal Convolutional Networks (TCN) [2,17,26], and Transformers [13,14,27,42,44-46]. While SimpleBaseline [19] and LSTM-based methods [9] highlight the initial strides, they exhibit limitations in computational efficiency and temporal information utilization. GCN-based methods [1,5,39,43], though adept at local joint modeling, struggle with computational complexity. Transformer-based architectures, such as PoseFormer [45], bring significant improvements by modeling spatial-temporal dependencies. The multi-solution nature of 2D-to-3D lifting is addressed by methods like MHFormer [14], which generate multiple hypotheses to account for ambiguity in body part locations. Our work challenges this approach, positing a single accurate solution corresponding to the original 2D pose, driven by spatial and temporal relational constraints to optimize results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Loss Functions in 3D Human Pose Estimation. The accuracy of 3D HPE models heavily relies on the choice of loss functions. The Mean Per Joint Position Error (MPJPE) [14, 27, 44-46] is a prevalent metric, gauging the Euclidean distance between predicted and ground truth joint positions. Its simplicity and direct measurement of joint accuracy make it a popular choice. However, it does not account for joint orientations, potentially overlooking critical aspects of pose estimation. The Weighted-MPJPE (W-MPJPE) [42] attempts to refine this by assigning different weights to various joints, thereby prioritizing certain joints over others based on the application's needs. Yet, this method introduces its own set of challenges, including potential bias in weight assignment and the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Z. Sun et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.435 + ], + "angle": 0, + "content": "need for expert knowledge in determining these weights. Additionally, the Mean Per Joint Velocity Error (MPJVE) [26,42] has been used to improve the temporal coherence between the predicted pose sequence and the ground truth sequence, which is a crucial aspect in dynamic pose estimation. However, it falls short in leveraging long-range temporal information. To address this gap, Zhang et al. [42] integrated the MPJPE and temporal consistency losses introduced by Hossain et al. [9] to leverage temporal relationships and smooth pose transitions. However, these adaptations insufficiently tackle the issue of relative depth perception among joints. Various studies [29, 46] attempt to quantify limb angle inaccuracies by calculating the Mean Absolute Error between the estimated and ground-truth skeletal angles. Additionally, they explore angle dynamics through the Mean Per-Angle Velocity Error (MPAVE), representing the mean Euclidean distance in the angle's first temporal derivative. Despite these efforts, current methodologies overly focus on the x and y dimensions, derived primarily from 2D pose estimations, neglecting the critical depth (z-axis) disparities. Alternatively, Pavlakos et al. [25] utilize predicted depth values to compute the ordinal depth loss, especially when handling datasets that lack ground truth labels. However, this approach offers a weaker supervisory signal due to the inherent unreliability of the predictions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.788, + 0.588 + ], + "angle": 0, + "content": "In response to the above problems, we introduce two novel loss functions: spatial and temporal relational depth consistency losses. The spatial loss focuses on relative depth differences between joints within a single frame, ensuring accurate depth relationships and mitigating pose estimation inconsistencies. The temporal loss maintains pose smoothness over time, considering relative depth differences of the same joint across frames. This dual focus enhances the overall robustness of 3D HPE, ensuring more accurate, temporally coherent, and visually appealing human pose sequences. Our methodology, utilizing these losses, aims to significantly improve the accuracy and temporal stability of pose estimation, making it more applicable and reliable in various computer vision tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.617, + 0.341, + 0.634 + ], + "angle": 0, + "content": "3 RePOSE" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.657, + 0.592, + 0.673 + ], + "angle": 0, + "content": "3.1 Problem Formulation and Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Our method leverages the 2D-to-3D lifting pipeline to elevate a sequence of 2D skeletons to their corresponding 3D coordinates. Given a sequence of 2D joints with confidence scores, represented as \\( X \\in \\mathbb{R}^{T \\times J \\times 3} \\) where \\( T \\) and \\( J \\) denote the number of frames and joints respectively, this input is transformed into a d-dimensional feature space \\( F \\in \\mathbb{R}^{T \\times J \\times d} \\) by the lifting network, with \\( d \\) varying across different network architectures. The output of this process is the estimated 3D pose sequence, denoted as \\( \\hat{Y} \\in \\mathbb{R}^{T \\times J \\times 3} \\), obtained through a regression mechanism. In our approach, we specifically treat the third dimension of \\( \\hat{Y} \\) as the depth feature \\( \\hat{D} \\in \\mathbb{R}^{T \\times J} \\), facilitating the computation of relational depth consistency losses for model training supervision." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.318, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "3D Human Pose Estimation via Depth Relational Consistency" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.144, + 0.784, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.353, + 0.788, + 0.452 + ], + "angle": 0, + "content": "Fig. 2: Our framework pipeline. Our network is designed with a spatio-temporal architecture, enabling simultaneous modeling of both spatial and temporal information. The generated 3D poses are processed through dual loss modules: the spatial loss module focuses on the relative depth disparities among joints within each frame, while the temporal loss module evaluates the relative depth differences of identical joints across consecutive frames. This dual-module approach ensures comprehensive depth accuracy and temporal consistency in the pose estimation process." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.487, + 0.785, + 0.518 + ], + "angle": 0, + "content": "Prevailing methodologies in recent studies [13,14,27,42,44-46] predominantly utilize Mean Per Joint Position Error (MPJPE) as loss function, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.535, + 0.785, + 0.577 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m p j p e} = \\sum_ {t = 1} ^ {T} \\sum_ {i = 1} ^ {J} \\| \\hat {y} _ {t, i} - y _ {t, i} \\| _ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.593, + 0.784, + 0.638 + ], + "angle": 0, + "content": "where \\( T \\) and \\( J \\) represent the total number of frames and joints in each video, \\( \\hat{y}_{t,i} \\) and \\( y_{t,i} \\) correspond to the estimated and ground-truth positions of keypoint \\( i \\) in frame \\( t \\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.785, + 0.732 + ], + "angle": 0, + "content": "While the MPJPE is a widely accepted metric in 3D HPE, it has limitations, particularly in its focus on joint positions without explicit consideration of spatial relationships or pose topology. This becomes a significant concern in scenarios with occlusions, where understanding the relative positions of body parts is vital for accurate pose estimation. To address this, we introduce the spatial depth ranking loss, specifically designed to enhance pose topology awareness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Additionally, conventional techniques often fail to effectively utilize temporal information, as they typically compute loss independently for each frame, neglecting the continuity and interdependence within video sequences. This oversight can significantly limit a model's ability to capture realistic human motion over time, a crucial aspect in dynamic pose estimation tasks. To mitigate this gap, we propose the temporal depth ranking loss, a novel component aimed at incorporating temporal dynamics for more accurate and lifelike pose estimation." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Z. Sun et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.419, + 0.16 + ], + "angle": 0, + "content": "3.2 Network Structure" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.168, + 0.788, + 0.288 + ], + "angle": 0, + "content": "Inspired by [27,45], we recognize the exemplary performance of spatial-temporal transformer architectures in 3D human pose estimation (HPE) due to their robust capability to model both spatial and temporal information in video sequences. In particular, they demonstrate a parallel dual-stream design to be effective in ensuring comprehensive learning of spatial and temporal aspects. Consequently, our network, RePOSE, adopts a bifurcated structure of the spatiotemporal transformer to enhance the capture of spatial and temporal features from input videos." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.288, + 0.789, + 0.455 + ], + "angle": 0, + "content": "As depicted in Fig. 2, RePOSE is composed of a spatio-temporal module and two specialized loss modules. The spatial transformer encoding module processes individual body joints as distinct tokens, thereby efficiently capturing the inter-joint relationships within the same frame. In contrast, the temporal transformer encoding module considers cross-frame features, enabling it to discern relationships between joints across different frames. The outputs from these modules are then directed to the spatial and temporal relational depth consistency modules, which compute the respective losses. A crucial aspect of our architecture is the backward pass, which plays a vital role in guiding the model's training process. The nuances and operations of the spatial and temporal relational depth consistency modules are elaborated in subsequent sections." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.472, + 0.573, + 0.487 + ], + "angle": 0, + "content": "3.3 Spatial Relational Depth Consistency" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.492, + 0.779 + ], + "angle": 0, + "content": "In any given frame, the ideal relationship between estimated depths \\(\\hat{d}\\) and ground-truth depths \\(d\\) should adhere to the condition \\(\\hat{d}_{t,i} < \\hat{d}_{t,j}\\) if \\(d_{t,i} < d_{t,j}\\), where \\(t\\) denotes a specific time step, and \\(d_{t,i}\\) and \\(d_{t,j}\\) represent the depths of different joints \\(i\\) and \\(j\\) at that moment, respectively. The same applies for \\(\\hat{d}_{t,i}\\) and \\(\\hat{d}_{t,j}\\). However, we observe that this relative depth relationship is often overlooked in existing methods, leading to inaccurate results, especially in scenarios with noise such as self-occlusion. To address this, we introduce a novel loss function to the training procedure, aimed at steering the network optimization in a more accurate direction. The formula is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.783, + 0.49, + 0.841 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S} = \\sum_ {t = 1} ^ {T} \\sum_ {d _ {t, i} < d _ {t, j}} \\max \\left(0, \\hat {d} _ {t, i} - \\hat {d} _ {t, j}\\right), \\tag {2}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.503, + 0.789, + 0.647 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.656, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Fig. 3: Overview of spatio-temporal relational depth consistency loss. The spatial loss (left) enhances relative depth relationships between the in-frame joints. The temporal loss (right) ensures the consistent relative depth relationships of joints across frames. The green dashed lines illustrate examples where the loss is calculated for the relative depth of linked joints against the ground-truth relative depth." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.317, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "3D Human Pose Estimation via Depth Relational Consistency" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "Eq. (2) demonstrates that our model penalizes deviations in the ranking of estimated depths from the ground-truth rankings. An illustration of this loss is displayed in the left part of Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.194, + 0.785, + 0.39 + ], + "angle": 0, + "content": "In cases where body parts are occluded, standard models often struggle to accurately determine the depth positioning of joints, leading to spatial inconsistencies. By integrating our spatial depth rank loss, we guide the estimations to align more closely with the actual spatial configuration, thereby providing feedback to the model for improved training and enhanced differentiation of obscured body parts or joints. Particularly in self-occlusion scenarios, traditional approaches might incorrectly associate body joints, such as mistaking a hand for an elbow due to overlapping positions. This type of error, which might go unnoticed when solely using MPJPE as the loss function, becomes apparent when the human skeleton is visualized. Our proposed spatial depth rank loss function, by reinforcing the inter-joint relationships within a frame, effectively prevents these errors and ensures a more accurate representation of the human body structure. The code of this loss is displayed in Algorithm 1." + }, + { + "type": "algorithm", + "bbox": [ + 0.214, + 0.409, + 0.8, + 0.7 + ], + "angle": 0, + "content": "Algorithm 1 Spatial Relational Depth Consistency \ndef process(x, J, idx): def process(x, T, idx): x = x.Unsqueeze(idx) return xrepeat_interleave(J, idx) return xrepeat_interleave(T, idx) \ndef loss_spatial(x, gt): def loss_temporal(x, gt): #x, gt: [B, T, J, 3] #x, gt: [B, T, J, 3] _, , J, = x.shape _, T, , = x.shape x = x.permute(0, 2, 1) [... , 2] gt = gt.permute(0, 2, 1) [... , 2] A = process(gt, J, -1) A = process(gt, T, -1) B = process(gt, J, -2) B = process(gt, T, -2) C = process(x, J, -1) C = process(x, T, -1) D = process(x, J, -2) D = process(x, T, -2) Comp_1 = (C - D) * (A < B) Comp_2 = torch.zeros_like(C) Comp_2 = torch.zeros_like(C) SRLoss = torch.maxmin( Comp_1, Comp_2) return torch.mean(SRLoss) return torch.mean(TRLoss)" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.732, + 0.591, + 0.748 + ], + "angle": 0, + "content": "3.4 Temporal Relational Depth Consistency" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.785, + 0.84 + ], + "angle": 0, + "content": "In parallel with our spatial considerations, we address depth consistency in the temporal dimension. Temporally, the depth ranking of a specific joint should remain consistent across different time steps to ensure the continuity of the 3D pose sequence. However, challenges arise, particularly during self-occlusion, where the estimated poses may disrupt this continuity. To counteract this, we" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "Z. Sun et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.484, + 0.162 + ], + "angle": 0, + "content": "introduce the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.36, + 0.174, + 0.786, + 0.218 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {T} = \\sum_ {j = 1} ^ {J} \\sum_ {d _ {m, j} < d _ {n, j}} \\max \\left(0, \\hat {d} _ {m, j} - \\hat {d} _ {n, j}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.229, + 0.784, + 0.274 + ], + "angle": 0, + "content": "where \\( j \\) indicates a specific joint, \\( d_{m,j} \\) and \\( d_{n,j} \\) represent the depths of joint \\( j \\) at different times \\( m \\) and \\( n \\), similarly for \\( \\hat{d}_{m,j} \\) and \\( \\hat{d}_{n,j} \\). The right part of Fig. 3 shows the idea of this loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.276, + 0.785, + 0.381 + ], + "angle": 0, + "content": "The temporal depth ranking loss encourages each joint in the estimated poses to maintain temporal consistency, preventing any joint from deviating suddenly from its original trajectory within the sequence. This loss is particularly crucial in ensuring that the temporal relationships between consecutive frames in video sequences are adequately maintained, which aids in smoothing the motion and making the model less susceptible to outliers. The implementation is detailed in Algorithm 2." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.405, + 0.391, + 0.42 + ], + "angle": 0, + "content": "3.5 Final Objective" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.784, + 0.522 + ], + "angle": 0, + "content": "Our comprehensive approach integrates the spatial and temporal dimensions to significantly enhance the coherence and realism of estimated 3D pose sequences. This integration ensures that the movements captured are not only spatially precise but also exhibit temporal fluidity and lifelike dynamics. Our final objective function combines the conventional MPJPE loss with our spatio-temporal relational depth consistency losses. The final loss function is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.537, + 0.786, + 0.555 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {m p j p e} + \\lambda_ {S} \\mathcal {L} _ {S} + \\lambda_ {T} \\mathcal {L} _ {T}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.561, + 0.785, + 0.636 + ], + "angle": 0, + "content": "where \\( \\mathcal{L}mpjpe \\) is the standard MPJPE loss, and the parameters \\( \\lambda_{S} \\) and \\( \\lambda_{T} \\) are weighting factors that balance the contribution of the spatial and temporal components in the overall loss function. This unified loss function is designed to optimize the network in a holistic manner, addressing both the individual joint accuracy and the relational depth in both spatial and temporal dimensions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.661, + 0.376, + 0.679 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.693, + 0.533, + 0.708 + ], + "angle": 0, + "content": "4.1 Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.784, + 0.749 + ], + "angle": 0, + "content": "We have conducted training and evaluation of our model using the datasets Human3.6M [11] and MPI-INF-3DHP [20]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Human3.6M is a well-established indoor dataset widely employed in the field of 3D HPE. This dataset comprises 3.6 million video frames captured from four distinct viewing angles, featuring performances by eleven professional actors engaged in diverse activities, including sitting, eating, and walking and more. In line with research [2,17,19,26,45,46], we utilize subjects 1, 5, 6, 7, and 8 as our training dataset, reserving subjects 9 and 11 for testing purposes." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.317, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "3D Human Pose Estimation via Depth Relational Consistency" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.332, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.232, + 0.332, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.316, + 0.305, + 0.328 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.146, + 0.442, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.23, + 0.442, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.367, + 0.316, + 0.412, + 0.328 + ], + "angle": 0, + "content": "(b) GT" + }, + { + "type": "image", + "bbox": [ + 0.449, + 0.146, + 0.554, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.45, + 0.23, + 0.554, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.453, + 0.316, + 0.552, + 0.328 + ], + "angle": 0, + "content": "(c) MotionBERT" + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.146, + 0.666, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.23, + 0.666, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.561, + 0.316, + 0.669, + 0.328 + ], + "angle": 0, + "content": "(d) PoseFormerV2" + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.146, + 0.78, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.23, + 0.78, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.702, + 0.316, + 0.753, + 0.328 + ], + "angle": 0, + "content": "(e) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.34, + 0.788, + 0.409 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparison on Human3.6M. Our method is evaluated against MotionBERT [46] and PoseFormerV2 [44] in scenarios featuring extensive self-occlusion of the human body. Areas with incorrect relative depth relationships are marked with red circles for emphasis, illustrating the effectiveness of each method in handling complex occlusion challenges." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.788, + 0.542 + ], + "angle": 0, + "content": "In our evaluation process, we report two key metrics: MPJPE and Procrustes Aligned-MPJPE (PA-MPJPE), both measured in millimeters. MPJPE assesses the average Euclidean distance between the estimated joint positions and their corresponding ground truth values. In contrast, the PA-MPJPE represents the MPJPE values after performing a rigid alignment procedure involving translation, rotation, and scaling operations to bring the estimated 3D pose as close as possible to the ground truth poses." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.542, + 0.788, + 0.694 + ], + "angle": 0, + "content": "MPI-INF-3DHP is another large-scale 3D HPE dataset consisting of both constrained indoor and challenging outdoor scenes. It records 8 actors performing 8 activities. It consists of over 1.3 million frames captured from the 14 cameras. To rigorously assess the generalization capabilities of RePOSE, particularly in challenging outdoor environments and scenarios involving occlusions, we conducted a comprehensive evaluation of our model using the MPI-INF-3DHP dataset. Here we utilize 2D pose sequences with a length of 81 frames each as input, as distinct from what it is in Human3.6M. Following previous works [2, 32, 44], we report MPJPE, Percentage of Correct Keypoints (PCK) within the \\(150\\mathrm{mm}\\) range as well as Area Under the Curve (AUC)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.712, + 0.457, + 0.727 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.842 + ], + "angle": 0, + "content": "We have implemented our methodology using PyTorch [24] on a computing platform consisting of two NVIDIA RTX 3090 GPUs. Following previous studies [29, 44, 46], we have configured the video sequence length to 243 frames and leveraged the Stacked Hourglass [23] for the extraction of 2D poses during both training and evaluation phases. Additionally, we have applied horizontal flipping augmentation, consistent with [1, 2, 46]. For the optimization of model parameters, we have employed the AdamW [18] optimizer for 100 training epochs," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Z. Sun et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.244 + ], + "angle": 0, + "content": "Table 1: Quantitative Comparison on Human3.6M Dataset. The top and bottom show comparison based on MPJPE and Procrustes-Aligned MPJPE in millimeters respectively. In this table, \\( T \\) represents the number of input frames used in each method. The best-performing methods are highlighted in bold, while the second-best results are underlined for clarity. Different 2D detection methods are indicated as follows: (*) denotes the use of CPN [3], (\\(\\dagger\\)) indicates the use of HRNet [33], and (\\(\\ddagger\\)) signifies the use of SH [23] for 2D pose detection." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.267, + 0.785, + 0.555 + ], + "angle": 0, + "content": "
MPJPETDir. Disc. Eat. GreetPhonePhotoPosePurch. Sit SitD.SmokeWaitWalkD.WalkWalkT.
*VideoPose3D (CVPR'19) [26]24345.246.743.345.648.155.144.644.357.365.847.144.049.032.833.9
*PoseFormer (ICCV'21) [45]8141.544.839.842.546.551.642.142.053.360.745.543.346.131.832.2
*MHFormer (CVPR'22) [14]35139.243.140.140.944.951.240.641.353.560.343.741.143.829.830.6
†MixSTE (CVPR'22) [42]24336.739.036.539.440.244.939.836.947.954.839.637.839.329.730.6
*P-STMO (ECCV'22) [27]24338.442.139.840.245.248.940.438.353.857.343.941.642.229.329.3
*PoseFormerV2 (CVPR'23) [44]24341.245.441.444.046.653.742.642.655.164.645.742.845.732.332.8
*STCFformer (CVPR'23) [30]24338.441.236.838.042.750.538.738.252.556.841.838.440.226.227.7
*D3DP (ICCV'23) [28]24337.739.935.738.241.948.839.538.350.553.941.639.439.827.427.5
‡MotionBERT (ICCV'23) [46]24336.137.535.832.140.346.336.135.346.953.939.536.335.825.125.3
‡MoAGFormer (WACV'24) [29]24336.838.535.933.041.148.638.034.849.051.440.337.436.327.227.2
‡RePOSE (Ours)24334.636835.031.338.844.735.335.547.550.938.435.534.923.925.036.5
PA-MPJPETDir. Disc. Eat. GreetPhonePhotoPosePurch. Sit SitD.SmokeWaitWalkD.WalkWalkT.
*VideoPose3D (CVPR'19) [26]24334.136.134.437.236.442.234.433.645.052.537.433.837.825.627.3
*PoseFormer (ICCV'21) [45]8132.534.832.634.635.339.532.132.042.848.534.832.435.324.526.0
*MHFormer (CVPR'22) [14]35131.534.932.833.635.339.632.032.243.548.736.432.634.323.925.1
*MixSTE (CVPR'22) [42]24330.833.130.331.833.139.131.130.542.544.534.030.832.722.122.9
*P-STMO (ECCV'22) [27]24331.335.232.933.935.439.332.531.544.648.236.332.934.423.823.9
*PoseFormerV2 (CVPR'23) [44]24332.335.833.735.835.941.133.132.744.251.937.432.835.525.226.5
*STCFformer (CVPR'23) [30]24329.333.030.730.632.738.229.728.842.245.033.329.431.520.922.3
*D3DP (ICCV'23) [28]24330.632.529.131.031.937.630.329.440.643.633.330.531.421.522.4
‡MotionBERT (ICCV'23) [46]24329.531.332.027.032.837.229.829.641.048.234.330.130.221.221.8
‡MoAGFormer (WACV'24) [29]24331.032.631.027.934.038.731.530.041.445.434.830.831.322.823.2
‡RePOSE (Ours)24329.130.929.426.632.135.828.630.639.944.933.329.029.620.121.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.788, + 0.695 + ], + "angle": 0, + "content": "incorporating a weight decay factor of 0.01. The initial learning rate has been set to \\(5 \\times 10^{-4}\\), accompanied by a learning rate decay strategy with a decay rate of 0.99. Lastly, the hyper-parameters \\(\\lambda_{S}\\) and \\(\\lambda_{T}\\) are all set to 20. When training our model on Human3.6M, We utilize the stride data sample strategy with a one-third interval of the input length to make sure there are fewer overlapping frames between consecutive sequences. But for the MPI-INF-3DHP dataset, the interval is one-nine of the input length because of fewer samples." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.72, + 0.622, + 0.736 + ], + "angle": 0, + "content": "4.3 Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Human3.6M dataset. The proposed method has been seriously compared with state-of-the-art (SOTA) approaches, utilizing the Human3.6M dataset as the evaluation benchmark. As illustrated in Tab. 1, our method demonstrates exceptional performance, achieving an MPJPE of 36.5 millimeters and a PA-MPJPE of 30.7 millimeters. These results surpass those achieved by leading SOTA methods, indicating a notable performance improvement." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.317, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "3D Human Pose Estimation via Depth Relational Consistency" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 2: Computational Complexity Comparison. The MPJPEs here are evaluated on Human3.6M using ground-truth 2D joint locations as input." + }, + { + "type": "table", + "bbox": [ + 0.286, + 0.185, + 0.714, + 0.332 + ], + "angle": 0, + "content": "
MethodParam (M)MACs (G)MPJPE
PoseFormer (ICCV'21) [45]9.52.531.3
MHFormer (CVPR'22) [14]30.97.030.5
MixSTE (CVPR'22) [42]33.6139.025.9
P-STMO (ECCV'22) [27]6.20.729.3
PoseFormerV2 (CVPR'23) [44]14.30.5-
STCFormer (CVPR'23) [30]4.719.621.3
MotionBERT (ICCV'23) [46]42.5174.717.8
MotionAGFormer (WACV'24) [29]19.078.317.3
RePOSE (Ours)16.043.715.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.349, + 0.785, + 0.375 + ], + "angle": 0, + "content": "Table 3: Quantitative comparison on MPI-INF-3DHP dataset. The best and second-best results are bolded and underlined." + }, + { + "type": "table", + "bbox": [ + 0.293, + 0.401, + 0.712, + 0.551 + ], + "angle": 0, + "content": "
MethodPCK↑AUC↑MPJPE↓
VideoPose3D (CVPR'19) [26]85.551.584.8
PoseFormer (ICCV'21) [45]65.463.257.7
MHFormer (CVPR'22) [14]93.863.358.0
MixSTE (CVPR'22) [42]94.466.554.9
P-STMO (ECCV'22) [27]97.975.832.2
D3DP (ICCV'23) [28]97.777.830.2
PoseFormerV2 (CVPR'23) [44]97.978.827.8
STCFormaler (CVPR'23) [30]98.783.923.1
MotionAGFormer (WACV'24) [29]98.285.316.2
RePOSE (Ours)98.386.715.5
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.785, + 0.67 + ], + "angle": 0, + "content": "It is worth noting that even when considering the best results reported by each of the prior studies, our method consistently outperforms them, especially in challenging scenarios involving actions such as \"photo,\" \"sitting,\" and \"sitting down.\" This highlights its potential to effectively address challenges related to occlusion. Qualitative results can be found in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.672, + 0.785, + 0.747 + ], + "angle": 0, + "content": "We also compare the number of parameters, multiply-accumulate operations (MACs), and MPJPE using the ground-truth 2D pose sequences as input, as shown in Tab. 2. Our model attains the best MPJPE out of other works with limited parameters and algorithm complexity, demonstrating the effectiveness and efficiency of our proposed method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "MPI-INF-3DHP dataset. Tab. 3 presents a quantitative comparison of our method on the MPI-INF-3DHP dataset. RePOSE achieves impressive results, with an MPJPE of \\(15.5\\mathrm{mm}\\) and an Area Under Curve (AUC) of \\(86.7\\%\\), significantly outperforming other compared methods by margins of \\(1.8\\mathrm{mm}\\) in MPJPE and \\(1.4\\%\\) in AUC. However, in terms of the Percentage of Correct Keypoints (PCK), RePOSE reaches \\(98.3\\%\\), which is slightly lower by \\(0.4\\%\\) compared" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Z. Sun et al." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.143, + 0.333, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.23, + 0.333, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.316, + 0.305, + 0.328 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.147, + 0.44, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.339, + 0.231, + 0.441, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.367, + 0.316, + 0.413, + 0.328 + ], + "angle": 0, + "content": "(b) GT" + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.147, + 0.553, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.231, + 0.554, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.472, + 0.316, + 0.532, + 0.328 + ], + "angle": 0, + "content": "(c) D3DP" + }, + { + "type": "image", + "bbox": [ + 0.564, + 0.147, + 0.665, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.564, + 0.231, + 0.665, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.316, + 0.662, + 0.328 + ], + "angle": 0, + "content": "(d) STCFormer" + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.148, + 0.778, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.231, + 0.778, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.316, + 0.753, + 0.328 + ], + "angle": 0, + "content": "(e) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.339, + 0.785, + 0.381 + ], + "angle": 0, + "content": "Fig. 5: Qualitative comparison on MPI-INF-3DHP. Our method is compared with D3DP [28] and STCFoer [30]. The areas with incorrect results are pointed out with red arrows and circles to exemplify the importance of relational depth consistency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.785, + 0.457 + ], + "angle": 0, + "content": "to the best performance achieved by existing methods, but we excel in the best balance of all metrics. For a more detailed visual assessment, refer to the qualitative comparisons illustrated in Fig. 5." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.481, + 0.389, + 0.497 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.492, + 0.734 + ], + "angle": 0, + "content": "To evaluate the distinct contributions of individual components, we incorporated spatial and temporal relational depth consistency loss functions separately into the baseline network. As shown in Tab. 4, it is clear that without these spatial and temporal depth constraints, the baseline model experiences degradation in both the MPJPE and PA-MPJPE metrics. However, the inclusion of either the spatial or temporal loss functions results in a significant enhancement in the model's performance. This observation underscores the effectiveness of these two" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.74, + 0.749 + ], + "angle": 0, + "content": "loss functions in improving the overall quality of the model's predictions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Fig. 6 provides a comparative visualization of the performance of the RepoSE with and without the incorporation of two relational depth consistency losses. It is evident from the figure that the omission of these losses unfavorably affects the model's capacity to accurately estimate occluded body parts, as exemplified by the red circles. This inadequacy suggests that the model, without these losses, cannot effectively identify the relative depth of body parts during" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.788, + 0.635 + ], + "angle": 0, + "content": "Table 4: Ablation Study on Individual Loss Functions. The study reveals that while both spatial and temporal loss functions individually contribute to improved pose estimation, their combined application yields the best results. Performance metrics have been evaluated on the Human3.6M dataset to substantiate this finding." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.658, + 0.783, + 0.72 + ], + "angle": 0, + "content": "
MPJPE Spatial TemporalMPJPE↓PA-MPJPE↓
✓--37.932.5
✓✓-36.931.4
✓-✓37.231.8
✓✓✓36.530.7
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.317, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "3D Human Pose Estimation via Depth Relational Consistency" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.144, + 0.373, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.243, + 0.373, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.28, + 0.342, + 0.337, + 0.354 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.149, + 0.501, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.245, + 0.501, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.384, + 0.342, + 0.492, + 0.354 + ], + "angle": 0, + "content": "(b) w/o our losses" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.149, + 0.63, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.245, + 0.629, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.342, + 0.605, + 0.354 + ], + "angle": 0, + "content": "(c) RePOSE" + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.149, + 0.758, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.245, + 0.757, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.643, + 0.342, + 0.75, + 0.354 + ], + "angle": 0, + "content": "(d) Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.365, + 0.788, + 0.422 + ], + "angle": 0, + "content": "Fig. 6: Visualized Results of Ablation Study. The second column displays poses generated by a model trained without the spatio-temporal relational depth consistency loss, contrasting with the third column, which shows results from the model including this loss. Joints with less accurate estimations are highlighted in red circles for clarity." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.436, + 0.785, + 0.479 + ], + "angle": 0, + "content": "Table 5: Generalization evaluation on Human3.6M with our loss functions. With the integration of our loss functions, all methods exhibit improved performance. \\(\\Delta\\) indicates the difference between 'w/ST' and 'w/o'." + }, + { + "type": "table", + "bbox": [ + 0.236, + 0.504, + 0.77, + 0.584 + ], + "angle": 0, + "content": "
MethodTMPJPE↓PA-MPJPE↓
w/Sw/Tw/STw/oΔw/Sw/Tw/STw/oΔ
P-STMO-S [27]2745.345.544.746.1-1.436.436.635.636.8-1.2
STCFformer [30]2743.843.843.344.1-0.834.134.033.734.8-1.1
MotionBERT [46]24337.037.236.737.5-0.831.431.730.831.8-1.0
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.788, + 0.714 + ], + "angle": 0, + "content": "the training stage, leading to incorrect results. On the contrary, the integration of our proposed losses into the training phase significantly enhances the model's ability to resolve occlusions, thereby improving its estimation accuracy. This improvement is crucial for applications where precise depth perception is necessary, and it emphasizes the value of our contributions to the field of pose estimation in occlusion situations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.739, + 0.478, + 0.753 + ], + "angle": 0, + "content": "4.5 Generalization Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To validate the versatility and generalization capacity of our proposed loss functions, we extend their application to similar spatio-temporal transformer-based frameworks, including MotionBERT [46], STCFoer [30], and P-STMO [27]. We directly use their official pre-trained models with default setups and fine-tune them with our loss functions." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Z. Sun et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "We detail the generalizability of our individual loss functions in Table 5. Overall, integrating both losses ('w/ST') leads to superior performance metrics compared to the original model configurations without our adjustments. Significantly, even when applied separately, each loss function ('w/S' or 'w/T') yields improvements across all evaluated methods, reinforcing their broad applicability. It is important to note that all reported results were achieved without specific tuning of the hyperparameters \\( \\lambda_{S} \\) and \\( \\lambda_{T} \\). Therefore, there is an opportunity for further enhancements if these parameters were to be systematically optimized." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.299, + 0.511, + 0.315 + ], + "angle": 0, + "content": "5 Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.788, + 0.55 + ], + "angle": 0, + "content": "In this paper, we introduce RePOSE, an approach designed to enhance 3D pose estimation accuracy in video sequences, specifically under conditions where body parts are obscured or occluded. Diverging from conventional methods that predominantly depend on depth information, which can falter when body parts are not visible, RePOSE adopts an innovative strategy centered on the relative positioning of body segments. This methodology not only exhibits superior efficacy in managing occlusions but also boasts a straightforward implementation, requiring minimal code adjustments. Our comprehensive experiments validate that RePOSE outperforms existing state-of-the-art techniques, particularly in scenarios marked by obstructions. These outcomes underscore RePOSE's efficacy and its promising potential in advancing 3D pose estimation, especially in practical settings where occlusions and partial visibilities are common. Additionally, the simplicity of its integration renders RePOSE a viable and effective solution for researchers aiming to refine pose estimation precision." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.788, + 0.673 + ], + "angle": 0, + "content": "Limitations. Despite its significant advantages in occlusion scenarios within 3D human pose estimation, RePOSE encounters limitations, chiefly its dependence on the accuracy of preliminary 2D pose data. Discrepancies in this foundational data can adversely affect subsequent 3D estimations, a challenge inherent in 2D-to-3D conversion techniques. Moreover, although tailored for occlusions, RePOSE's performance may diminish in complex conditions, such as atypical body postures or interactions with objects, and in extremely low-light environments where depth cues are compromised." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Future Work. The efficacy of our newly proposed loss functions has been established across various spatio-temporal transformer architectures, including MotionBERT [46] and STCFoer [30]. Our next phase of research will explore the extension of these functions to a wider spectrum of models, including those based on GNNs and CNNs. Additionally, we aim to evolve the hyperparameters \\(\\lambda_{S}\\) and \\(\\lambda_{T}\\) from fixed values based on empirical selection to adaptive, learnable parameters optimized through the model's interaction with input data, thereby refining the model's depth dimension understanding in 3D human pose estimation. Future enhancements will focus on augmenting RePOSE's versatility across different input data forms and broadening its efficacy in a more extensive array of challenging conditions." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.317, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "3D Human Pose Estimation via Depth Relational Consistency" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.254 + ], + "angle": 0, + "content": "This project is supported by the Guangdong Basic and Applied Basic Research Foundation under project No. 2024A1515011995, Guangdong Natural Science Funds for Distinguished Young Scholar under Grant 2023B1515020097, and National Research Foundation Singapore under the AI Singapore Programme under Grant AISG3-GV-2023-011." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.278, + 0.323, + 0.294 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.787, + 0.351 + ], + "angle": 0, + "content": "1. Cai, Y., Ge, L., Liu, J., Cai, J., Cham, T.J., Yuan, J., Thalmann, N.M.: Exploiting spatial-temporal relationships for 3d pose estimation via graph convolutional networks. In: ICCV. pp. 2272-2281 (2019) 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.354, + 0.787, + 0.393 + ], + "angle": 0, + "content": "2. Chen, T., Fang, C., Shen, X., Zhu, Y., Chen, Z., Luo, J.: Anatomy-aware 3d human pose estimation with bone-based pose decomposition. IEEE TCSVT 32(1), 198-209 (2021) 2, 3, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.395, + 0.787, + 0.422 + ], + "angle": 0, + "content": "3. Chen, Y., Wang, Z., Peng, Y., Zhang, Z., Yu, G., Sun, J.: Cascaded pyramid network for multi-person pose estimation. In: CVPR. pp. 7103-7112 (2018) 3, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.424, + 0.787, + 0.463 + ], + "angle": 0, + "content": "4. Chen, Y., Tu, Z., Ge, L., Zhang, D., Chen, R., Yuan, J.: So-handnet: Self-organizing network for 3d hand pose estimation with semi-supervised learning. In: ICCV. pp. 6961-6970 (2019) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.465, + 0.787, + 0.492 + ], + "angle": 0, + "content": "5. Ci, H., Wang, C., Ma, X., Wang, Y.: Optimizing network structure for 3d human pose estimation. In: ICCV. pp. 2262-2271 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.493, + 0.787, + 0.52 + ], + "angle": 0, + "content": "6. Ci, H., Wu, M., Zhu, W., Ma, X., Dong, H., Zhong, F., Wang, Y.: Gfpose: Learning 3d human pose prior with gradient fields. In: CVPR. pp. 4800-4810 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.521, + 0.787, + 0.547 + ], + "angle": 0, + "content": "7. Fang, H.S., Xie, S., Tai, Y.W., Lu, C.: Rumpe: Regional multi-person pose estimation. In: ICCV. pp. 2334-2343 (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.549, + 0.787, + 0.575 + ], + "angle": 0, + "content": "8. Gong, J., Fan, Z., Ke, Q., Rahmani, H., Liu, J.: Meta agent teaming active learning for pose estimation. In: CVPR. pp. 11079-11089 (2022) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.577, + 0.787, + 0.603 + ], + "angle": 0, + "content": "9. Hossain, M.R.I., Little, J.J.: Exploiting temporal information for 3d human pose estimation. In: ECCV. pp. 68-84 (2018) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.605, + 0.787, + 0.631 + ], + "angle": 0, + "content": "0. Huang, S., Wang, W., He, S., Lau, R.W.: Egocentric temporal action proposals. IEEE TIP 27(2), 764-777 (2017) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.787, + 0.673 + ], + "angle": 0, + "content": "1. Ionescu, C., Papava, D., Olaru, V., Sminchisescu, C.: Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE TPAMI 36(7), 1325-1339 (2013) 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.787, + 0.701 + ], + "angle": 0, + "content": "2. Jiang, Y., Zhou, Y., Liang, Y., Liu, W., Jiao, J., Quan, Y., He, S.: Diffuse3d: Wide-angle 3d photography via bilateral diffusion. In: ICCV. pp. 8998-9008 (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.787, + 0.742 + ], + "angle": 0, + "content": "3. Li, W., Liu, H., Ding, R., Liu, M., Wang, P., Yang, W.: Exploiting temporal contexts with strided transformer for 3d human pose estimation. IEEE Transactions on Multimedia 25, 1282-1293 (2022) 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.787, + 0.783 + ], + "angle": 0, + "content": "4. Li, W., Liu, H., Tang, H., Wang, P., Van Gool, L.: Mhformer: Multi-hypothesis transformer for 3d human pose estimation. In: CVPR. pp. 13147-13156 (2022) 2, 3, 5, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.787, + 0.812 + ], + "angle": 0, + "content": "5. Lin, K., Wang, L., Liu, Z.: End-to-end human pose and mesh reconstruction with transformers. In: CVPR. pp. 1954-1963 (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "6. Lin, M., Lin, L., Liang, X., Wang, K., Cheng, H.: Recurrent 3d pose sequence machines. In: CVPR. pp. 810-819 (2017) 3" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.31, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Z. Sun et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "17. Liu, R., Shen, J., Wang, H., Chen, C., Cheung, S.c., Asari, V.: Attention mechanism exploits temporal contexts: Real-time 3d human pose reconstruction. In: CVPR. pp. 5064-5073 (2020) 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.215 + ], + "angle": 0, + "content": "18. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.217, + 0.785, + 0.244 + ], + "angle": 0, + "content": "19. Martinez, J., Hossain, R., Romero, J., Little, J.J.: A simple yet effective baseline for 3d human pose estimation. In: ICCV. pp. 2640-2649 (2017) 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.298 + ], + "angle": 0, + "content": "20. Mehta, D., Rhodin, H., Casas, D., Fua, P., Sotnychenko, O., Xu, W., Theobalt, C.: Monocular 3d human pose estimation in the wild using improved cnn supervision. In: 2017 international conference on 3D vision (3DV). pp. 506-516. IEEE (2017) 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "21. Mehta, D., Sridhar, S., Sotnychenko, O., Rhodin, H., Shafiei, M., Seidel, H.P., Xu, W., Casas, D., Theobalt, C.: Vnct: Real-time 3d human pose estimation with a single rgb camera. ACM TOG 36(4), 1-14 (2017) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.383 + ], + "angle": 0, + "content": "22. Moon, G., Lee, K.M.: I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In: ECCV. pp. 752-768. Springer (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "23. Newell, A., Yang, K., Deng, J.: Stacked hourglass networks for human pose estimation. In: ECCV. pp. 483-499. Springer (2016) 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "24. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch. In: NeurIPS (2017) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.479 + ], + "angle": 0, + "content": "25. Pavlakos, G., Zhou, X., Daniilidis, K.: Ordinal depth supervision for 3d human pose estimation. In: CVPR. pp. 7307-7316 (2018) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.521 + ], + "angle": 0, + "content": "26. Pavlo, D., Feichtenhofer, C., Grangier, D., Auli, M.: 3d human pose estimation in video with temporal convolutions and semi-supervised training. In: CVPR. pp. 7753-7762 (2019) 3, 4, 8, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.563 + ], + "angle": 0, + "content": "27. Shan, W., Liu, Z., Zhang, X., Wang, S., Ma, S., Gao, W.: P-stmo: Pre-trained spatial temporal many-to-one model for 3d human pose estimation. In: ECCV. pp. 461-478. Springer (2022) 2, 3, 5, 6, 10, 11, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.605 + ], + "angle": 0, + "content": "28. Shan, W., Liu, Z., Zhang, X., Wang, Z., Han, K., Wang, S., Ma, S., Gao, W.: Diffusion-based 3d human pose estimation with multi-hypothesis aggregation. In: ICCV. pp. 14761-14771 (October 2023) 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.633 + ], + "angle": 0, + "content": "29. Soroush Mehraban, Vida Adeli, B.T.: Motionagformer: Enhancing 3d human pose estimation with a transformer-gcnformer network. In: WACV (2024) 4, 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.634, + 0.785, + 0.673 + ], + "angle": 0, + "content": "30. Tang, Z., Qiu, Z., Hao, Y., Hong, R., Yao, T.: 3d human pose estimation with spatio-temporal criss-cross attention. In: CVPR. pp. 4790-4799 (2023) 10, 11, 12, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "31. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.703, + 0.785, + 0.729 + ], + "angle": 0, + "content": "32. Wang, J., Yan, S., Xiong, Y., Lin, D.: Motion guided 3d pose estimation from videos. In: ECCV. pp. 764-780. Springer (2020) 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.785, + 0.77 + ], + "angle": 0, + "content": "33. Wang, J., Sun, K., Cheng, T., Jiang, B., Deng, C., Zhao, Y., Liu, D., Mu, Y., Tan, M., Wang, X., et al.: Deep high-resolution representation learning for visual recognition. IEEE TPAMI 43(10), 3349-3364 (2020) 3, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "34. Xu, Y., Han, C., Qin, J., Xu, X., Han, G., He, S.: Transductive zero-shot action recognition via visually connected graph convolutional networks. IEEE TNNLS 32(8), 3761-3769 (2020) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "35. Yan, S., Xiong, Y., Lin, D.: Spatial temporal graph convolutional networks for skeleton-based action recognition. In: AAAI. vol. 32 (2018) 1" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.317, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "3D Human Pose Estimation via Depth Relational Consistency" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "36. Yang, X., Xu, K., Chen, S., He, S., Yin, B.Y., Lau, R.: Active matting. NeurIPS 31 (2018) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.203 + ], + "angle": 0, + "content": "37. Ye, M., Li, H., Du, B., Shen, J., Shao, L., Hoi, S.C.: Collaborative refining for person re-identification with label noise. IEEE TIP 31, 379-391 (2021) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.788, + 0.245 + ], + "angle": 0, + "content": "38. Yoon, J.S., Liu, L., Golyanik, V., Sarkar, K., Park, H.S., Theobalt, C.: Pose-guided human animation from a single image in the wild. In: CVPR. pp. 15039-15048 (2021) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.245, + 0.788, + 0.286 + ], + "angle": 0, + "content": "39. Yu, B.X., Zhang, Z., Liu, Y., Zhong, S.h., Liu, Y., Chen, C.W.: Gla-gcn: Global-local adaptive graph convolutional network for 3d human pose estimation from monocular video. In: ICCV. pp. 8818-8829 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.327 + ], + "angle": 0, + "content": "40. Zhang, C., Yang, T., Weng, J., Cao, M., Wang, J., Zou, Y.: Unsupervised pretraining for temporal action localization tasks. In: CVPR. pp. 14031-14041 (2022) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.37 + ], + "angle": 0, + "content": "41. Zhang, J., Ye, G., Tu, Z., Qin, Y., Qin, Q., Zhang, J., Liu, J.: A spatial attentive and temporal dilated (satd) gcn for skeleton-based action recognition. CAAI Transactions on Intelligence Technology 7(1), 46-55 (2022) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.411 + ], + "angle": 0, + "content": "42. Zhang, J., Tu, Z., Yang, J., Chen, Y., Yuan, J.: Mixste: Seq2seq mixed spatiotemporal encoder for 3d human pose estimation in video. In: CVPR. pp. 13232-13242 (2022) 2, 3, 4, 5, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.411, + 0.788, + 0.452 + ], + "angle": 0, + "content": "43. Zhao, L., Peng, X., Tian, Y., Kapadia, M., Metaxas, D.N.: Semantic graph convolutional networks for 3d human pose regression. In: CVPR. pp. 3425-3435 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.452, + 0.788, + 0.493 + ], + "angle": 0, + "content": "44. Zhao, Q., Zheng, C., Liu, M., Wang, P., Chen, C.: Poseformerv2: Exploring frequency domain for efficient and robust 3d human pose estimation. In: CVPR. pp. 8877-8886 (2023) 2, 3, 5, 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.535 + ], + "angle": 0, + "content": "45. Zheng, C., Zhu, S., Mendieta, M., Yang, T., Chen, C., Ding, Z.: 3d human pose estimation with spatial and temporal transformers. In: ICCV. pp. 11656-11665 (2021) 2, 3, 5, 6, 8, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.577 + ], + "angle": 0, + "content": "46. Zhu, W., Ma, X., Liu, Z., Liu, L., Wu, W., Wang, Y.: Motionbert: A unified perspective on learning human motion representations. In: ICCV. pp. 15085-15099 (2023) 2, 3, 4, 5, 8, 9, 10, 11, 13, 14" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.577 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/RePOSE_ 3D Human Pose Estimation via Spatio-Temporal Depth Relational Consistency/7257623d-d6ee-4c47-b288-fda8cfd3af7e_origin.pdf b/2024/RePOSE_ 3D Human Pose Estimation via Spatio-Temporal Depth Relational Consistency/7257623d-d6ee-4c47-b288-fda8cfd3af7e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..de2e72a4c669b9727983461a6ac1d03553f65cef --- /dev/null +++ b/2024/RePOSE_ 3D Human Pose Estimation via Spatio-Temporal Depth Relational Consistency/7257623d-d6ee-4c47-b288-fda8cfd3af7e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e1e706d04ec912336ef5460f9d7ffcf9d3cacc6d8cb565175770215291cb6400 +size 4281301 diff --git a/2024/RePOSE_ 3D Human Pose Estimation via Spatio-Temporal Depth Relational Consistency/full.md b/2024/RePOSE_ 3D Human Pose Estimation via Spatio-Temporal Depth Relational Consistency/full.md new file mode 100644 index 0000000000000000000000000000000000000000..76d24f308885ed949416af9ea1336c10f77d5055 --- /dev/null +++ b/2024/RePOSE_ 3D Human Pose Estimation via Spatio-Temporal Depth Relational Consistency/full.md @@ -0,0 +1,340 @@ +# RePOSE: 3D Human Pose Estimation via Spatio-Temporal Depth Relational Consistency + +Ziming Sun $^{1,2}$ , Yuan Liang $^{1,2\star}$ , Zejun Ma $^{3}$ , Tianle Zhang $^{1,2}$ , Linchao Bao $^{4}$ , Guiqing Li $^{1}$ , and Shengfeng He $^{2(\boxtimes)}$ + +$^{1}$ South China University of Technology, China +2 Singapore Management University, Singapore + +shengfenghe@smu.edu.sg + +3 Johns Hopkins University, USA + +Tencent AI Lab, China + +https://github.com/StupidAdam/RePOSE + +Abstract. We introduce RePOSE, a simple yet effective approach for addressing occlusion challenges in the learning of 3D human pose estimation (HPE) from videos. Conventional approaches typically employ absolute depth signals as supervision, which are adept at discernible keypoints but become less reliable when keypoints are occluded, resulting in vague and inconsistent learning trajectories for the neural network. RePOSE overcomes this limitation by introducing spatio-temporal relational depth consistency into the supervision signals. The core rationale of our method lies in prioritizing the precise sequencing of occluded keypoints. This is achieved by using a relative depth consistency loss that operates in both spatial and temporal domains. By doing so, RePOSE shifts the focus from learning absolute depth values, which can be misleading in occluded scenarios, to relative positioning, which provides a more robust and reliable cue for accurate pose estimation. This subtle yet crucial shift facilitates more consistent and accurate 3D HPE under occlusion conditions. The elegance of our core idea lies in its simplicity and ease of implementation, requiring only a few lines of code. Extensive experiments validate that RePOSE not only outperforms existing state-of-the-art methods but also significantly enhances the robustness and precision of 3D HPE in challenging occluded environments. + +Keywords: 3D human pose estimation $\cdot$ depth relational loss functions + +# 1 Introduction + +The estimation of 3D human pose from videos is a foundational task in computer vision, with widespread implications across various applications such as virtual reality [4, 12, 21, 38], human-computer interaction [8, 36, 37], and action recognition [10, 34, 35, 40, 41]. This task involves localizing body joints in a three-dimensional space to create a comprehensive representation of the human body, + +![](images/1c0744240422b5ab87e1d08e8c0d44299961764cc9c5722792e646faf3ab17d1.jpg) + +![](images/ee5c2813a09c79f0b6d3f807c0c0d33838a997c3b65d0f3ac5972afdbcdede1c.jpg) +(a) Input + +![](images/7dcd5bd38044a893f609861415fca27bfdba1859012395069acfeb57eaf5af9b.jpg) + +![](images/f5b726767e5dba04012cabc7e6041a6845e5404b85ca52b5fca53de346d6e88e.jpg) +(b) GT + +![](images/8e272043c2572a96e9c8ecb996057f97a01fcb710ba3e731e0c56fe53e5ce617.jpg) + +![](images/d882ac318445d710eea8c43ec415aadab0de81f95b390ea5d435edeb792812e0.jpg) +(c) MotionBERT +Fig. 1: We introduce RePOSE, a novel solution designed to address the occlusion challenges in 3D pose estimation through relational supervision. Unlike existing methods (c) [46] and (d) [44], which struggle with occlusions due to their dependence on absolute depth signals, our approach leverages relational depth consistency, effectively overcoming these obstacles. + +![](images/6bda6a8be2c27fa49d56f5e8980e2ec04eb951de7e13bca235d93d2b760e4ca6.jpg) + +![](images/694bf7c0ca073df277bb70a8ccb99aa40fcc167a934bab752673892d379143f3.jpg) +(d) PoseFormerV2 + +![](images/a79564eefe63fb04713494b2d95003caae75f1582bed2ef21ea534e1a9fbcc87.jpg) + +![](images/06e956d11a28004aab4d74247725c7c16f8b759257fd4192fa47c4ac8cbf8948.jpg) +(e) Ours + +typically in the form of a skeletal model. The methodologies in 3D human pose estimation (HPE) are primarily categorized into two approaches: direct estimation and 2D-to-3D lifting methods. The former [22, 25] entails the direct estimation of human pose from provided videos, whereas the latter [2, 14, 17, 27, 42, 44-46] involves the initial estimation of intermediate 2D poses using off-the-shelf 2D detectors, followed by lifting them to 3D space. 2D-to-3D lifting methods have gained prominence due to the rapid evolution of 2D pose detection technologies, exhibiting superior performance in many scenarios. + +Recent advancements in 2D-to-3D lifting methods often employ Transformer architectures [31], which have shown exceptional ability in capturing global information and modeling sequential data. These methods typically involve transforming detected 2D poses into 3D space using tailored algorithms [13-15,27,42,44-46]. However, this process inherently faces a many-to-one mapping challenge, where multiple distinct 3D poses can correspond to a single 2D projection, particularly under occlusion scenarios, as shown in Fig. 1c and 1d. In such cases, occlusion of key body joints leads to reduced accuracy in 2D pose estimations, creating ambiguities in the 3D reconstruction process. Moreover, relying on these unreliable occluded signals during model training results in vague and uncertain learning pathways. This limitation chiefly stems from the methods' reliance on absolute depth signals, which lose reliability under occluded conditions, thus undermining the overall accuracy of 3D pose estimation. + +To address this challenge, we propose RePOSE, a novel approach specifically designed to enhance the accuracy of 3D HPE in occluded scenarios. RePOSE circumvents the limitations of absolute depth reliance by introducing a spatiotemporal relational depth consistency. Our method emphasizes the correct sequencing of occluded keypoints using a relative depth consistency loss in both + +spatial and temporal domains, rather than depending solely on absolute depth values. This shift in approach enables our method to robustly handle occlusions, providing more accurate and reliable pose estimations, as shown in Fig. 1e. Additionally, the simplicity of spatio-temporal relational depth consistency, which can be implemented in just a few lines of code, makes it a practical and effective solution for real-world applications. + +We demonstrate through extensive experiments on the Human3.6M [11] and MPI-INF-3DHP [20] benchmarks that our method outperforms state-of-the-art methods by a large margin, particularly excelling in scenarios involving occlusions. Our findings are further substantiated through visual comparisons and an ablation study, showcasing the effectiveness and innovation of RePOSE in tackling one of the most challenging aspects of 3D human pose estimation. This method lays the groundwork for future research aimed at enhancing pose estimation accuracy and reliability in complex environments. + +# 2 Related Works + +2D-to-3D Lifting HPE. The prevalent approach in 3D HPE involves lifting 2D pose sequences to 3D poses, a method enhanced by advanced 2D detectors like CPN [3], AlphaPose [7], and HRNet [33]. These detectors facilitate the conversion process with various techniques, including Fully Connected Networks (FCN) [6,19], Long Short-Term Memory (LSTM) [9,16,32], Graph Convolutional Networks (GCN) [1,5,39,43], Temporal Convolutional Networks (TCN) [2,17,26], and Transformers [13,14,27,42,44-46]. While SimpleBaseline [19] and LSTM-based methods [9] highlight the initial strides, they exhibit limitations in computational efficiency and temporal information utilization. GCN-based methods [1,5,39,43], though adept at local joint modeling, struggle with computational complexity. Transformer-based architectures, such as PoseFormer [45], bring significant improvements by modeling spatial-temporal dependencies. The multi-solution nature of 2D-to-3D lifting is addressed by methods like MHFormer [14], which generate multiple hypotheses to account for ambiguity in body part locations. Our work challenges this approach, positing a single accurate solution corresponding to the original 2D pose, driven by spatial and temporal relational constraints to optimize results. + +Loss Functions in 3D Human Pose Estimation. The accuracy of 3D HPE models heavily relies on the choice of loss functions. The Mean Per Joint Position Error (MPJPE) [14, 27, 44-46] is a prevalent metric, gauging the Euclidean distance between predicted and ground truth joint positions. Its simplicity and direct measurement of joint accuracy make it a popular choice. However, it does not account for joint orientations, potentially overlooking critical aspects of pose estimation. The Weighted-MPJPE (W-MPJPE) [42] attempts to refine this by assigning different weights to various joints, thereby prioritizing certain joints over others based on the application's needs. Yet, this method introduces its own set of challenges, including potential bias in weight assignment and the + +need for expert knowledge in determining these weights. Additionally, the Mean Per Joint Velocity Error (MPJVE) [26,42] has been used to improve the temporal coherence between the predicted pose sequence and the ground truth sequence, which is a crucial aspect in dynamic pose estimation. However, it falls short in leveraging long-range temporal information. To address this gap, Zhang et al. [42] integrated the MPJPE and temporal consistency losses introduced by Hossain et al. [9] to leverage temporal relationships and smooth pose transitions. However, these adaptations insufficiently tackle the issue of relative depth perception among joints. Various studies [29, 46] attempt to quantify limb angle inaccuracies by calculating the Mean Absolute Error between the estimated and ground-truth skeletal angles. Additionally, they explore angle dynamics through the Mean Per-Angle Velocity Error (MPAVE), representing the mean Euclidean distance in the angle's first temporal derivative. Despite these efforts, current methodologies overly focus on the x and y dimensions, derived primarily from 2D pose estimations, neglecting the critical depth (z-axis) disparities. Alternatively, Pavlakos et al. [25] utilize predicted depth values to compute the ordinal depth loss, especially when handling datasets that lack ground truth labels. However, this approach offers a weaker supervisory signal due to the inherent unreliability of the predictions. + +In response to the above problems, we introduce two novel loss functions: spatial and temporal relational depth consistency losses. The spatial loss focuses on relative depth differences between joints within a single frame, ensuring accurate depth relationships and mitigating pose estimation inconsistencies. The temporal loss maintains pose smoothness over time, considering relative depth differences of the same joint across frames. This dual focus enhances the overall robustness of 3D HPE, ensuring more accurate, temporally coherent, and visually appealing human pose sequences. Our methodology, utilizing these losses, aims to significantly improve the accuracy and temporal stability of pose estimation, making it more applicable and reliable in various computer vision tasks. + +# 3 RePOSE + +# 3.1 Problem Formulation and Preliminaries + +Our method leverages the 2D-to-3D lifting pipeline to elevate a sequence of 2D skeletons to their corresponding 3D coordinates. Given a sequence of 2D joints with confidence scores, represented as $X \in \mathbb{R}^{T \times J \times 3}$ where $T$ and $J$ denote the number of frames and joints respectively, this input is transformed into a d-dimensional feature space $F \in \mathbb{R}^{T \times J \times d}$ by the lifting network, with $d$ varying across different network architectures. The output of this process is the estimated 3D pose sequence, denoted as $\hat{Y} \in \mathbb{R}^{T \times J \times 3}$ , obtained through a regression mechanism. In our approach, we specifically treat the third dimension of $\hat{Y}$ as the depth feature $\hat{D} \in \mathbb{R}^{T \times J}$ , facilitating the computation of relational depth consistency losses for model training supervision. + +![](images/e7032e65c42f3b007f1ebee51fd94053d288f084a8f699b782dee065ae405815.jpg) +Fig. 2: Our framework pipeline. Our network is designed with a spatio-temporal architecture, enabling simultaneous modeling of both spatial and temporal information. The generated 3D poses are processed through dual loss modules: the spatial loss module focuses on the relative depth disparities among joints within each frame, while the temporal loss module evaluates the relative depth differences of identical joints across consecutive frames. This dual-module approach ensures comprehensive depth accuracy and temporal consistency in the pose estimation process. + +Prevailing methodologies in recent studies [13,14,27,42,44-46] predominantly utilize Mean Per Joint Position Error (MPJPE) as loss function, defined as: + +$$ +\mathcal {L} _ {m p j p e} = \sum_ {t = 1} ^ {T} \sum_ {i = 1} ^ {J} \| \hat {y} _ {t, i} - y _ {t, i} \| _ {2}, \tag {1} +$$ + +where $T$ and $J$ represent the total number of frames and joints in each video, $\hat{y}_{t,i}$ and $y_{t,i}$ correspond to the estimated and ground-truth positions of keypoint $i$ in frame $t$ , respectively. + +While the MPJPE is a widely accepted metric in 3D HPE, it has limitations, particularly in its focus on joint positions without explicit consideration of spatial relationships or pose topology. This becomes a significant concern in scenarios with occlusions, where understanding the relative positions of body parts is vital for accurate pose estimation. To address this, we introduce the spatial depth ranking loss, specifically designed to enhance pose topology awareness. + +Additionally, conventional techniques often fail to effectively utilize temporal information, as they typically compute loss independently for each frame, neglecting the continuity and interdependence within video sequences. This oversight can significantly limit a model's ability to capture realistic human motion over time, a crucial aspect in dynamic pose estimation tasks. To mitigate this gap, we propose the temporal depth ranking loss, a novel component aimed at incorporating temporal dynamics for more accurate and lifelike pose estimation. + +# 3.2 Network Structure + +Inspired by [27,45], we recognize the exemplary performance of spatial-temporal transformer architectures in 3D human pose estimation (HPE) due to their robust capability to model both spatial and temporal information in video sequences. In particular, they demonstrate a parallel dual-stream design to be effective in ensuring comprehensive learning of spatial and temporal aspects. Consequently, our network, RePOSE, adopts a bifurcated structure of the spatiotemporal transformer to enhance the capture of spatial and temporal features from input videos. + +As depicted in Fig. 2, RePOSE is composed of a spatio-temporal module and two specialized loss modules. The spatial transformer encoding module processes individual body joints as distinct tokens, thereby efficiently capturing the inter-joint relationships within the same frame. In contrast, the temporal transformer encoding module considers cross-frame features, enabling it to discern relationships between joints across different frames. The outputs from these modules are then directed to the spatial and temporal relational depth consistency modules, which compute the respective losses. A crucial aspect of our architecture is the backward pass, which plays a vital role in guiding the model's training process. The nuances and operations of the spatial and temporal relational depth consistency modules are elaborated in subsequent sections. + +# 3.3 Spatial Relational Depth Consistency + +In any given frame, the ideal relationship between estimated depths $\hat{d}$ and ground-truth depths $d$ should adhere to the condition $\hat{d}_{t,i} < \hat{d}_{t,j}$ if $d_{t,i} < d_{t,j}$ , where $t$ denotes a specific time step, and $d_{t,i}$ and $d_{t,j}$ represent the depths of different joints $i$ and $j$ at that moment, respectively. The same applies for $\hat{d}_{t,i}$ and $\hat{d}_{t,j}$ . However, we observe that this relative depth relationship is often overlooked in existing methods, leading to inaccurate results, especially in scenarios with noise such as self-occlusion. To address this, we introduce a novel loss function to the training procedure, aimed at steering the network optimization in a more accurate direction. The formula is as follows: + +$$ +\mathcal {L} _ {S} = \sum_ {t = 1} ^ {T} \sum_ {d _ {t, i} < d _ {t, j}} \max \left(0, \hat {d} _ {t, i} - \hat {d} _ {t, j}\right), \tag {2} +$$ + +![](images/8b9bd58a00290c6342b649f2a6a062516279bfb3397bfd6c37e7f89ce7dffaab.jpg) +Fig. 3: Overview of spatio-temporal relational depth consistency loss. The spatial loss (left) enhances relative depth relationships between the in-frame joints. The temporal loss (right) ensures the consistent relative depth relationships of joints across frames. The green dashed lines illustrate examples where the loss is calculated for the relative depth of linked joints against the ground-truth relative depth. + +Eq. (2) demonstrates that our model penalizes deviations in the ranking of estimated depths from the ground-truth rankings. An illustration of this loss is displayed in the left part of Fig. 3. + +In cases where body parts are occluded, standard models often struggle to accurately determine the depth positioning of joints, leading to spatial inconsistencies. By integrating our spatial depth rank loss, we guide the estimations to align more closely with the actual spatial configuration, thereby providing feedback to the model for improved training and enhanced differentiation of obscured body parts or joints. Particularly in self-occlusion scenarios, traditional approaches might incorrectly associate body joints, such as mistaking a hand for an elbow due to overlapping positions. This type of error, which might go unnoticed when solely using MPJPE as the loss function, becomes apparent when the human skeleton is visualized. Our proposed spatial depth rank loss function, by reinforcing the inter-joint relationships within a frame, effectively prevents these errors and ensures a more accurate representation of the human body structure. The code of this loss is displayed in Algorithm 1. + +Algorithm 1 Spatial Relational Depth Consistency +def process(x, J, idx): def process(x, T, idx): x = x.Unsqueeze(idx) return xrepeat_interleave(J, idx) return xrepeat_interleave(T, idx) +def loss_spatial(x, gt): def loss_temporal(x, gt): #x, gt: [B, T, J, 3] #x, gt: [B, T, J, 3] _, , J, = x.shape _, T, , = x.shape x = x.permute(0, 2, 1) [... , 2] gt = gt.permute(0, 2, 1) [... , 2] A = process(gt, J, -1) A = process(gt, T, -1) B = process(gt, J, -2) B = process(gt, T, -2) C = process(x, J, -1) C = process(x, T, -1) D = process(x, J, -2) D = process(x, T, -2) Comp_1 = (C - D) * (A < B) Comp_2 = torch.zeros_like(C) Comp_2 = torch.zeros_like(C) SRLoss = torch.maxmin( Comp_1, Comp_2) return torch.mean(SRLoss) return torch.mean(TRLoss) + +# 3.4 Temporal Relational Depth Consistency + +In parallel with our spatial considerations, we address depth consistency in the temporal dimension. Temporally, the depth ranking of a specific joint should remain consistent across different time steps to ensure the continuity of the 3D pose sequence. However, challenges arise, particularly during self-occlusion, where the estimated poses may disrupt this continuity. To counteract this, we + +introduce the following loss function: + +$$ +\mathcal {L} _ {T} = \sum_ {j = 1} ^ {J} \sum_ {d _ {m, j} < d _ {n, j}} \max \left(0, \hat {d} _ {m, j} - \hat {d} _ {n, j}\right), \tag {3} +$$ + +where $j$ indicates a specific joint, $d_{m,j}$ and $d_{n,j}$ represent the depths of joint $j$ at different times $m$ and $n$ , similarly for $\hat{d}_{m,j}$ and $\hat{d}_{n,j}$ . The right part of Fig. 3 shows the idea of this loss. + +The temporal depth ranking loss encourages each joint in the estimated poses to maintain temporal consistency, preventing any joint from deviating suddenly from its original trajectory within the sequence. This loss is particularly crucial in ensuring that the temporal relationships between consecutive frames in video sequences are adequately maintained, which aids in smoothing the motion and making the model less susceptible to outliers. The implementation is detailed in Algorithm 2. + +# 3.5 Final Objective + +Our comprehensive approach integrates the spatial and temporal dimensions to significantly enhance the coherence and realism of estimated 3D pose sequences. This integration ensures that the movements captured are not only spatially precise but also exhibit temporal fluidity and lifelike dynamics. Our final objective function combines the conventional MPJPE loss with our spatio-temporal relational depth consistency losses. The final loss function is formulated as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {m p j p e} + \lambda_ {S} \mathcal {L} _ {S} + \lambda_ {T} \mathcal {L} _ {T}, \tag {4} +$$ + +where $\mathcal{L}mpjpe$ is the standard MPJPE loss, and the parameters $\lambda_{S}$ and $\lambda_{T}$ are weighting factors that balance the contribution of the spatial and temporal components in the overall loss function. This unified loss function is designed to optimize the network in a holistic manner, addressing both the individual joint accuracy and the relational depth in both spatial and temporal dimensions. + +# 4 Experiments + +# 4.1 Datasets and Evaluation Metrics + +We have conducted training and evaluation of our model using the datasets Human3.6M [11] and MPI-INF-3DHP [20]. + +Human3.6M is a well-established indoor dataset widely employed in the field of 3D HPE. This dataset comprises 3.6 million video frames captured from four distinct viewing angles, featuring performances by eleven professional actors engaged in diverse activities, including sitting, eating, and walking and more. In line with research [2,17,19,26,45,46], we utilize subjects 1, 5, 6, 7, and 8 as our training dataset, reserving subjects 9 and 11 for testing purposes. + +![](images/4c22ed4ec878a98191d4e6f9a52a51fda529caf646c4ff9a1fd9c54586a8c73c.jpg) + +![](images/47930450a25966683445342f95cf83f23d1037c40b33fb3ee0e3204c62896e5e.jpg) +(a) Input + +![](images/7a20d2a5405214e98e28acb93750bc5d6445bd313866c22e6b44eeeb40635883.jpg) + +![](images/3c25018181092d1f98b2488e994461bb62411e32e143746e6b9b4b6cd3594e09.jpg) +(b) GT + +![](images/fab57cea4bf05a10c4073afb39931eff242846a39e9a695fc13ebd5710a33319.jpg) + +![](images/b78b878721b03295a4c4796ae7fc1941daa0fb7df4e91b3d25fed14a40f6a697.jpg) +(c) MotionBERT +Fig. 4: Qualitative comparison on Human3.6M. Our method is evaluated against MotionBERT [46] and PoseFormerV2 [44] in scenarios featuring extensive self-occlusion of the human body. Areas with incorrect relative depth relationships are marked with red circles for emphasis, illustrating the effectiveness of each method in handling complex occlusion challenges. + +![](images/934db952d9968d255e3a775800581501521b0d98911e1028661f8e81831cf519.jpg) + +![](images/aaadb08ebd0dd2175e757b7cc0312a03bac03415b9c9354c780f469ea052f716.jpg) +(d) PoseFormerV2 + +![](images/0d19798756dd2a26c91ffe0e46824c8a6e4a63fc5bf68828598e21907b0aa15a.jpg) + +![](images/5ae4b58a58bddc1fd7fdaa94c17f877c04a37b4d2b781e8b360b40f895eb1a2a.jpg) +(e) Ours + +In our evaluation process, we report two key metrics: MPJPE and Procrustes Aligned-MPJPE (PA-MPJPE), both measured in millimeters. MPJPE assesses the average Euclidean distance between the estimated joint positions and their corresponding ground truth values. In contrast, the PA-MPJPE represents the MPJPE values after performing a rigid alignment procedure involving translation, rotation, and scaling operations to bring the estimated 3D pose as close as possible to the ground truth poses. + +MPI-INF-3DHP is another large-scale 3D HPE dataset consisting of both constrained indoor and challenging outdoor scenes. It records 8 actors performing 8 activities. It consists of over 1.3 million frames captured from the 14 cameras. To rigorously assess the generalization capabilities of RePOSE, particularly in challenging outdoor environments and scenarios involving occlusions, we conducted a comprehensive evaluation of our model using the MPI-INF-3DHP dataset. Here we utilize 2D pose sequences with a length of 81 frames each as input, as distinct from what it is in Human3.6M. Following previous works [2, 32, 44], we report MPJPE, Percentage of Correct Keypoints (PCK) within the $150\mathrm{mm}$ range as well as Area Under the Curve (AUC). + +# 4.2 Implementation Details + +We have implemented our methodology using PyTorch [24] on a computing platform consisting of two NVIDIA RTX 3090 GPUs. Following previous studies [29, 44, 46], we have configured the video sequence length to 243 frames and leveraged the Stacked Hourglass [23] for the extraction of 2D poses during both training and evaluation phases. Additionally, we have applied horizontal flipping augmentation, consistent with [1, 2, 46]. For the optimization of model parameters, we have employed the AdamW [18] optimizer for 100 training epochs, + +Table 1: Quantitative Comparison on Human3.6M Dataset. The top and bottom show comparison based on MPJPE and Procrustes-Aligned MPJPE in millimeters respectively. In this table, $T$ represents the number of input frames used in each method. The best-performing methods are highlighted in bold, while the second-best results are underlined for clarity. Different 2D detection methods are indicated as follows: (*) denotes the use of CPN [3], ( $\dagger$ ) indicates the use of HRNet [33], and ( $\ddagger$ ) signifies the use of SH [23] for 2D pose detection. + +
MPJPETDir. Disc. Eat. GreetPhonePhotoPosePurch. Sit SitD.SmokeWaitWalkD.WalkWalkT.
*VideoPose3D (CVPR'19) [26]24345.246.743.345.648.155.144.644.357.365.847.144.049.032.833.9
*PoseFormer (ICCV'21) [45]8141.544.839.842.546.551.642.142.053.360.745.543.346.131.832.2
*MHFormer (CVPR'22) [14]35139.243.140.140.944.951.240.641.353.560.343.741.143.829.830.6
†MixSTE (CVPR'22) [42]24336.739.036.539.440.244.939.836.947.954.839.637.839.329.730.6
*P-STMO (ECCV'22) [27]24338.442.139.840.245.248.940.438.353.857.343.941.642.229.329.3
*PoseFormerV2 (CVPR'23) [44]24341.245.441.444.046.653.742.642.655.164.645.742.845.732.332.8
*STCFformer (CVPR'23) [30]24338.441.236.838.042.750.538.738.252.556.841.838.440.226.227.7
*D3DP (ICCV'23) [28]24337.739.935.738.241.948.839.538.350.553.941.639.439.827.427.5
‡MotionBERT (ICCV'23) [46]24336.137.535.832.140.346.336.135.346.953.939.536.335.825.125.3
‡MoAGFormer (WACV'24) [29]24336.838.535.933.041.148.638.034.849.051.440.337.436.327.227.2
‡RePOSE (Ours)24334.636835.031.338.844.735.335.547.550.938.435.534.923.925.036.5
PA-MPJPETDir. Disc. Eat. GreetPhonePhotoPosePurch. Sit SitD.SmokeWaitWalkD.WalkWalkT.
*VideoPose3D (CVPR'19) [26]24334.136.134.437.236.442.234.433.645.052.537.433.837.825.627.3
*PoseFormer (ICCV'21) [45]8132.534.832.634.635.339.532.132.042.848.534.832.435.324.526.0
*MHFormer (CVPR'22) [14]35131.534.932.833.635.339.632.032.243.548.736.432.634.323.925.1
*MixSTE (CVPR'22) [42]24330.833.130.331.833.139.131.130.542.544.534.030.832.722.122.9
*P-STMO (ECCV'22) [27]24331.335.232.933.935.439.332.531.544.648.236.332.934.423.823.9
*PoseFormerV2 (CVPR'23) [44]24332.335.833.735.835.941.133.132.744.251.937.432.835.525.226.5
*STCFformer (CVPR'23) [30]24329.333.030.730.632.738.229.728.842.245.033.329.431.520.922.3
*D3DP (ICCV'23) [28]24330.632.529.131.031.937.630.329.440.643.633.330.531.421.522.4
‡MotionBERT (ICCV'23) [46]24329.531.332.027.032.837.229.829.641.048.234.330.130.221.221.8
‡MoAGFormer (WACV'24) [29]24331.032.631.027.934.038.731.530.041.445.434.830.831.322.823.2
‡RePOSE (Ours)24329.130.929.426.632.135.828.630.639.944.933.329.029.620.121.2
+ +incorporating a weight decay factor of 0.01. The initial learning rate has been set to $5 \times 10^{-4}$ , accompanied by a learning rate decay strategy with a decay rate of 0.99. Lastly, the hyper-parameters $\lambda_{S}$ and $\lambda_{T}$ are all set to 20. When training our model on Human3.6M, We utilize the stride data sample strategy with a one-third interval of the input length to make sure there are fewer overlapping frames between consecutive sequences. But for the MPI-INF-3DHP dataset, the interval is one-nine of the input length because of fewer samples. + +# 4.3 Comparison with State-of-the-Art Methods + +Human3.6M dataset. The proposed method has been seriously compared with state-of-the-art (SOTA) approaches, utilizing the Human3.6M dataset as the evaluation benchmark. As illustrated in Tab. 1, our method demonstrates exceptional performance, achieving an MPJPE of 36.5 millimeters and a PA-MPJPE of 30.7 millimeters. These results surpass those achieved by leading SOTA methods, indicating a notable performance improvement. + +Table 2: Computational Complexity Comparison. The MPJPEs here are evaluated on Human3.6M using ground-truth 2D joint locations as input. + +
MethodParam (M)MACs (G)MPJPE
PoseFormer (ICCV'21) [45]9.52.531.3
MHFormer (CVPR'22) [14]30.97.030.5
MixSTE (CVPR'22) [42]33.6139.025.9
P-STMO (ECCV'22) [27]6.20.729.3
PoseFormerV2 (CVPR'23) [44]14.30.5-
STCFormer (CVPR'23) [30]4.719.621.3
MotionBERT (ICCV'23) [46]42.5174.717.8
MotionAGFormer (WACV'24) [29]19.078.317.3
RePOSE (Ours)16.043.715.7
+ +Table 3: Quantitative comparison on MPI-INF-3DHP dataset. The best and second-best results are bolded and underlined. + +
MethodPCK↑AUC↑MPJPE↓
VideoPose3D (CVPR'19) [26]85.551.584.8
PoseFormer (ICCV'21) [45]65.463.257.7
MHFormer (CVPR'22) [14]93.863.358.0
MixSTE (CVPR'22) [42]94.466.554.9
P-STMO (ECCV'22) [27]97.975.832.2
D3DP (ICCV'23) [28]97.777.830.2
PoseFormerV2 (CVPR'23) [44]97.978.827.8
STCFormaler (CVPR'23) [30]98.783.923.1
MotionAGFormer (WACV'24) [29]98.285.316.2
RePOSE (Ours)98.386.715.5
+ +It is worth noting that even when considering the best results reported by each of the prior studies, our method consistently outperforms them, especially in challenging scenarios involving actions such as "photo," "sitting," and "sitting down." This highlights its potential to effectively address challenges related to occlusion. Qualitative results can be found in Fig. 4. + +We also compare the number of parameters, multiply-accumulate operations (MACs), and MPJPE using the ground-truth 2D pose sequences as input, as shown in Tab. 2. Our model attains the best MPJPE out of other works with limited parameters and algorithm complexity, demonstrating the effectiveness and efficiency of our proposed method. + +MPI-INF-3DHP dataset. Tab. 3 presents a quantitative comparison of our method on the MPI-INF-3DHP dataset. RePOSE achieves impressive results, with an MPJPE of $15.5\mathrm{mm}$ and an Area Under Curve (AUC) of $86.7\%$ , significantly outperforming other compared methods by margins of $1.8\mathrm{mm}$ in MPJPE and $1.4\%$ in AUC. However, in terms of the Percentage of Correct Keypoints (PCK), RePOSE reaches $98.3\%$ , which is slightly lower by $0.4\%$ compared + +![](images/5f3f5455109d303d89d43c8c7df680faff1e13a8c0d177ee89254d5af257149f.jpg) + +![](images/3a24fc725daba22553b8a94546af8cffd494a94e06b272a28ba4d0f88d1b4a69.jpg) +(a) Input + +![](images/02100a0d79ab8aa956b1a9eab01929261c359d6977c1ef2268f446ca8e3b41e6.jpg) + +![](images/39863b86a363cbefee185735abc617c38da40bb3e1932ca6b964aac1db9145b0.jpg) +(b) GT + +![](images/c7970a2ed7a0fcd5e858ff79e1105a51df62fcff18c9efe6c2d58b3703e58f51.jpg) + +![](images/b843a7cab61e261629d0288b08e759075274dc00f6723826f7c7a523ca973d77.jpg) +(c) D3DP +Fig. 5: Qualitative comparison on MPI-INF-3DHP. Our method is compared with D3DP [28] and STCFoer [30]. The areas with incorrect results are pointed out with red arrows and circles to exemplify the importance of relational depth consistency. + +![](images/d5141b30a094ea014d444f13a2f8239c7714756d77904571db0d4c8dc040b4fe.jpg) + +![](images/8912d8606a95cf2b454ce29f8e737086dc0aae73f95dbb8db9f185801f607281.jpg) +(d) STCFormer + +![](images/34d63eba5b5b4b6f98cd66b1215985b913750cb602de1205886d5070af862133.jpg) + +![](images/6b2543511f8dd399d8c942ad6c66f1308fe23b25a082bf1cb32f4bbdeeb98e5f.jpg) +(e) Ours + +to the best performance achieved by existing methods, but we excel in the best balance of all metrics. For a more detailed visual assessment, refer to the qualitative comparisons illustrated in Fig. 5. + +# 4.4 Ablation Study + +To evaluate the distinct contributions of individual components, we incorporated spatial and temporal relational depth consistency loss functions separately into the baseline network. As shown in Tab. 4, it is clear that without these spatial and temporal depth constraints, the baseline model experiences degradation in both the MPJPE and PA-MPJPE metrics. However, the inclusion of either the spatial or temporal loss functions results in a significant enhancement in the model's performance. This observation underscores the effectiveness of these two + +loss functions in improving the overall quality of the model's predictions. + +Fig. 6 provides a comparative visualization of the performance of the RepoSE with and without the incorporation of two relational depth consistency losses. It is evident from the figure that the omission of these losses unfavorably affects the model's capacity to accurately estimate occluded body parts, as exemplified by the red circles. This inadequacy suggests that the model, without these losses, cannot effectively identify the relative depth of body parts during + +Table 4: Ablation Study on Individual Loss Functions. The study reveals that while both spatial and temporal loss functions individually contribute to improved pose estimation, their combined application yields the best results. Performance metrics have been evaluated on the Human3.6M dataset to substantiate this finding. + +
MPJPE Spatial TemporalMPJPE↓PA-MPJPE↓
✓--37.932.5
✓✓-36.931.4
✓-✓37.231.8
✓✓✓36.530.7
+ +![](images/0b32b0329018eb45d3a07e32ac68d60a52ef76cc6bcad2f70a6e9eb4d488ef1a.jpg) + +![](images/8e1bd81571567d455c0c0b1e3ad64ddd0f3142de0511a4fcc828b8c0c3136744.jpg) +(a) Input +Fig. 6: Visualized Results of Ablation Study. The second column displays poses generated by a model trained without the spatio-temporal relational depth consistency loss, contrasting with the third column, which shows results from the model including this loss. Joints with less accurate estimations are highlighted in red circles for clarity. + +![](images/6578cb4f6549cacd34706e677dbc302c5522d31096f67e70da93888d6e198cd1.jpg) + +![](images/e8fbbfd8791ed3e1bb247458d4f2e5de897c0bc99dc899c2e0a50b4762ff0cea.jpg) +(b) w/o our losses + +![](images/2af1bcb8aa2e34e1f78294d691e07d32f2aa12beef27932ed505a995f0d6d30f.jpg) + +![](images/e3cd59544a8277d065ee6fd435d4ac60cebcfa4354de088be5bba109ba1b80b9.jpg) +(c) RePOSE + +![](images/880e335368efc8589ef1c7eeba1669d2dcd5f66f8f706455b8f2eb052bed2a16.jpg) + +![](images/693e73f2ca705f1f3da685dca86cee04166f884acbd0aa3e87f91ee7857eda80.jpg) +(d) Ground Truth + +Table 5: Generalization evaluation on Human3.6M with our loss functions. With the integration of our loss functions, all methods exhibit improved performance. $\Delta$ indicates the difference between 'w/ST' and 'w/o'. + +
MethodTMPJPE↓PA-MPJPE↓
w/Sw/Tw/STw/oΔw/Sw/Tw/STw/oΔ
P-STMO-S [27]2745.345.544.746.1-1.436.436.635.636.8-1.2
STCFformer [30]2743.843.843.344.1-0.834.134.033.734.8-1.1
MotionBERT [46]24337.037.236.737.5-0.831.431.730.831.8-1.0
+ +the training stage, leading to incorrect results. On the contrary, the integration of our proposed losses into the training phase significantly enhances the model's ability to resolve occlusions, thereby improving its estimation accuracy. This improvement is crucial for applications where precise depth perception is necessary, and it emphasizes the value of our contributions to the field of pose estimation in occlusion situations. + +# 4.5 Generalization Evaluation + +To validate the versatility and generalization capacity of our proposed loss functions, we extend their application to similar spatio-temporal transformer-based frameworks, including MotionBERT [46], STCFoer [30], and P-STMO [27]. We directly use their official pre-trained models with default setups and fine-tune them with our loss functions. + +We detail the generalizability of our individual loss functions in Table 5. Overall, integrating both losses ('w/ST') leads to superior performance metrics compared to the original model configurations without our adjustments. Significantly, even when applied separately, each loss function ('w/S' or 'w/T') yields improvements across all evaluated methods, reinforcing their broad applicability. It is important to note that all reported results were achieved without specific tuning of the hyperparameters $\lambda_{S}$ and $\lambda_{T}$ . Therefore, there is an opportunity for further enhancements if these parameters were to be systematically optimized. + +# 5 Conclusion and Discussion + +In this paper, we introduce RePOSE, an approach designed to enhance 3D pose estimation accuracy in video sequences, specifically under conditions where body parts are obscured or occluded. Diverging from conventional methods that predominantly depend on depth information, which can falter when body parts are not visible, RePOSE adopts an innovative strategy centered on the relative positioning of body segments. This methodology not only exhibits superior efficacy in managing occlusions but also boasts a straightforward implementation, requiring minimal code adjustments. Our comprehensive experiments validate that RePOSE outperforms existing state-of-the-art techniques, particularly in scenarios marked by obstructions. These outcomes underscore RePOSE's efficacy and its promising potential in advancing 3D pose estimation, especially in practical settings where occlusions and partial visibilities are common. Additionally, the simplicity of its integration renders RePOSE a viable and effective solution for researchers aiming to refine pose estimation precision. + +Limitations. Despite its significant advantages in occlusion scenarios within 3D human pose estimation, RePOSE encounters limitations, chiefly its dependence on the accuracy of preliminary 2D pose data. Discrepancies in this foundational data can adversely affect subsequent 3D estimations, a challenge inherent in 2D-to-3D conversion techniques. Moreover, although tailored for occlusions, RePOSE's performance may diminish in complex conditions, such as atypical body postures or interactions with objects, and in extremely low-light environments where depth cues are compromised. + +Future Work. The efficacy of our newly proposed loss functions has been established across various spatio-temporal transformer architectures, including MotionBERT [46] and STCFoer [30]. Our next phase of research will explore the extension of these functions to a wider spectrum of models, including those based on GNNs and CNNs. Additionally, we aim to evolve the hyperparameters $\lambda_{S}$ and $\lambda_{T}$ from fixed values based on empirical selection to adaptive, learnable parameters optimized through the model's interaction with input data, thereby refining the model's depth dimension understanding in 3D human pose estimation. Future enhancements will focus on augmenting RePOSE's versatility across different input data forms and broadening its efficacy in a more extensive array of challenging conditions. + +# Acknowledgement + +This project is supported by the Guangdong Basic and Applied Basic Research Foundation under project No. 2024A1515011995, Guangdong Natural Science Funds for Distinguished Young Scholar under Grant 2023B1515020097, and National Research Foundation Singapore under the AI Singapore Programme under Grant AISG3-GV-2023-011. + +# References + +1. Cai, Y., Ge, L., Liu, J., Cai, J., Cham, T.J., Yuan, J., Thalmann, N.M.: Exploiting spatial-temporal relationships for 3d pose estimation via graph convolutional networks. In: ICCV. pp. 2272-2281 (2019) 3, 9 +2. Chen, T., Fang, C., Shen, X., Zhu, Y., Chen, Z., Luo, J.: Anatomy-aware 3d human pose estimation with bone-based pose decomposition. IEEE TCSVT 32(1), 198-209 (2021) 2, 3, 8, 9 +3. Chen, Y., Wang, Z., Peng, Y., Zhang, Z., Yu, G., Sun, J.: Cascaded pyramid network for multi-person pose estimation. In: CVPR. pp. 7103-7112 (2018) 3, 10 +4. Chen, Y., Tu, Z., Ge, L., Zhang, D., Chen, R., Yuan, J.: So-handnet: Self-organizing network for 3d hand pose estimation with semi-supervised learning. In: ICCV. pp. 6961-6970 (2019) 1 +5. Ci, H., Wang, C., Ma, X., Wang, Y.: Optimizing network structure for 3d human pose estimation. In: ICCV. pp. 2262-2271 (2019) 3 +6. Ci, H., Wu, M., Zhu, W., Ma, X., Dong, H., Zhong, F., Wang, Y.: Gfpose: Learning 3d human pose prior with gradient fields. In: CVPR. pp. 4800-4810 (2023) 3 +7. Fang, H.S., Xie, S., Tai, Y.W., Lu, C.: Rumpe: Regional multi-person pose estimation. In: ICCV. pp. 2334-2343 (2017) 3 +8. Gong, J., Fan, Z., Ke, Q., Rahmani, H., Liu, J.: Meta agent teaming active learning for pose estimation. In: CVPR. pp. 11079-11089 (2022) 1 +9. Hossain, M.R.I., Little, J.J.: Exploiting temporal information for 3d human pose estimation. In: ECCV. pp. 68-84 (2018) 3, 4 +0. Huang, S., Wang, W., He, S., Lau, R.W.: Egocentric temporal action proposals. IEEE TIP 27(2), 764-777 (2017) 1 +1. Ionescu, C., Papava, D., Olaru, V., Sminchisescu, C.: Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE TPAMI 36(7), 1325-1339 (2013) 3, 8 +2. Jiang, Y., Zhou, Y., Liang, Y., Liu, W., Jiao, J., Quan, Y., He, S.: Diffuse3d: Wide-angle 3d photography via bilateral diffusion. In: ICCV. pp. 8998-9008 (2023) 1 +3. Li, W., Liu, H., Ding, R., Liu, M., Wang, P., Yang, W.: Exploiting temporal contexts with strided transformer for 3d human pose estimation. IEEE Transactions on Multimedia 25, 1282-1293 (2022) 2, 3, 5 +4. Li, W., Liu, H., Tang, H., Wang, P., Van Gool, L.: Mhformer: Multi-hypothesis transformer for 3d human pose estimation. In: CVPR. pp. 13147-13156 (2022) 2, 3, 5, 10, 11 +5. Lin, K., Wang, L., Liu, Z.: End-to-end human pose and mesh reconstruction with transformers. In: CVPR. pp. 1954-1963 (2021) 2 +6. Lin, M., Lin, L., Liang, X., Wang, K., Cheng, H.: Recurrent 3d pose sequence machines. In: CVPR. pp. 810-819 (2017) 3 + +17. Liu, R., Shen, J., Wang, H., Chen, C., Cheung, S.c., Asari, V.: Attention mechanism exploits temporal contexts: Real-time 3d human pose reconstruction. In: CVPR. pp. 5064-5073 (2020) 2, 3, 8 +18. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019) 9 +19. Martinez, J., Hossain, R., Romero, J., Little, J.J.: A simple yet effective baseline for 3d human pose estimation. In: ICCV. pp. 2640-2649 (2017) 3, 8 +20. Mehta, D., Rhodin, H., Casas, D., Fua, P., Sotnychenko, O., Xu, W., Theobalt, C.: Monocular 3d human pose estimation in the wild using improved cnn supervision. In: 2017 international conference on 3D vision (3DV). pp. 506-516. IEEE (2017) 3, 8 +21. Mehta, D., Sridhar, S., Sotnychenko, O., Rhodin, H., Shafiei, M., Seidel, H.P., Xu, W., Casas, D., Theobalt, C.: Vnct: Real-time 3d human pose estimation with a single rgb camera. ACM TOG 36(4), 1-14 (2017) 1 +22. Moon, G., Lee, K.M.: I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In: ECCV. pp. 752-768. Springer (2020) 2 +23. Newell, A., Yang, K., Deng, J.: Stacked hourglass networks for human pose estimation. In: ECCV. pp. 483-499. Springer (2016) 9, 10 +24. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch. In: NeurIPS (2017) 9 +25. Pavlakos, G., Zhou, X., Daniilidis, K.: Ordinal depth supervision for 3d human pose estimation. In: CVPR. pp. 7307-7316 (2018) 2, 4 +26. Pavlo, D., Feichtenhofer, C., Grangier, D., Auli, M.: 3d human pose estimation in video with temporal convolutions and semi-supervised training. In: CVPR. pp. 7753-7762 (2019) 3, 4, 8, 10, 11 +27. Shan, W., Liu, Z., Zhang, X., Wang, S., Ma, S., Gao, W.: P-stmo: Pre-trained spatial temporal many-to-one model for 3d human pose estimation. In: ECCV. pp. 461-478. Springer (2022) 2, 3, 5, 6, 10, 11, 13 +28. Shan, W., Liu, Z., Zhang, X., Wang, Z., Han, K., Wang, S., Ma, S., Gao, W.: Diffusion-based 3d human pose estimation with multi-hypothesis aggregation. In: ICCV. pp. 14761-14771 (October 2023) 10, 11, 12 +29. Soroush Mehraban, Vida Adeli, B.T.: Motionagformer: Enhancing 3d human pose estimation with a transformer-gcnformer network. In: WACV (2024) 4, 9, 10, 11 +30. Tang, Z., Qiu, Z., Hao, Y., Hong, R., Yao, T.: 3d human pose estimation with spatio-temporal criss-cross attention. In: CVPR. pp. 4790-4799 (2023) 10, 11, 12, 13, 14 +31. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017) 2 +32. Wang, J., Yan, S., Xiong, Y., Lin, D.: Motion guided 3d pose estimation from videos. In: ECCV. pp. 764-780. Springer (2020) 3, 9 +33. Wang, J., Sun, K., Cheng, T., Jiang, B., Deng, C., Zhao, Y., Liu, D., Mu, Y., Tan, M., Wang, X., et al.: Deep high-resolution representation learning for visual recognition. IEEE TPAMI 43(10), 3349-3364 (2020) 3, 10 +34. Xu, Y., Han, C., Qin, J., Xu, X., Han, G., He, S.: Transductive zero-shot action recognition via visually connected graph convolutional networks. IEEE TNNLS 32(8), 3761-3769 (2020) 1 +35. Yan, S., Xiong, Y., Lin, D.: Spatial temporal graph convolutional networks for skeleton-based action recognition. In: AAAI. vol. 32 (2018) 1 + +36. Yang, X., Xu, K., Chen, S., He, S., Yin, B.Y., Lau, R.: Active matting. NeurIPS 31 (2018) 1 +37. Ye, M., Li, H., Du, B., Shen, J., Shao, L., Hoi, S.C.: Collaborative refining for person re-identification with label noise. IEEE TIP 31, 379-391 (2021) 1 +38. Yoon, J.S., Liu, L., Golyanik, V., Sarkar, K., Park, H.S., Theobalt, C.: Pose-guided human animation from a single image in the wild. In: CVPR. pp. 15039-15048 (2021) 1 +39. Yu, B.X., Zhang, Z., Liu, Y., Zhong, S.h., Liu, Y., Chen, C.W.: Gla-gcn: Global-local adaptive graph convolutional network for 3d human pose estimation from monocular video. In: ICCV. pp. 8818-8829 (2023) 3 +40. Zhang, C., Yang, T., Weng, J., Cao, M., Wang, J., Zou, Y.: Unsupervised pretraining for temporal action localization tasks. In: CVPR. pp. 14031-14041 (2022) 1 +41. Zhang, J., Ye, G., Tu, Z., Qin, Y., Qin, Q., Zhang, J., Liu, J.: A spatial attentive and temporal dilated (satd) gcn for skeleton-based action recognition. CAAI Transactions on Intelligence Technology 7(1), 46-55 (2022) 1 +42. Zhang, J., Tu, Z., Yang, J., Chen, Y., Yuan, J.: Mixste: Seq2seq mixed spatiotemporal encoder for 3d human pose estimation in video. In: CVPR. pp. 13232-13242 (2022) 2, 3, 4, 5, 10, 11 +43. Zhao, L., Peng, X., Tian, Y., Kapadia, M., Metaxas, D.N.: Semantic graph convolutional networks for 3d human pose regression. In: CVPR. pp. 3425-3435 (2019) 3 +44. Zhao, Q., Zheng, C., Liu, M., Wang, P., Chen, C.: Poseformerv2: Exploring frequency domain for efficient and robust 3d human pose estimation. In: CVPR. pp. 8877-8886 (2023) 2, 3, 5, 9, 10, 11 +45. Zheng, C., Zhu, S., Mendieta, M., Yang, T., Chen, C., Ding, Z.: 3d human pose estimation with spatial and temporal transformers. In: ICCV. pp. 11656-11665 (2021) 2, 3, 5, 6, 8, 10, 11 +46. Zhu, W., Ma, X., Liu, Z., Liu, L., Wu, W., Wang, Y.: Motionbert: A unified perspective on learning human motion representations. In: ICCV. pp. 15085-15099 (2023) 2, 3, 4, 5, 8, 9, 10, 11, 13, 14 \ No newline at end of file diff --git a/2024/RePOSE_ 3D Human Pose Estimation via Spatio-Temporal Depth Relational Consistency/images.zip b/2024/RePOSE_ 3D Human Pose Estimation via Spatio-Temporal Depth Relational Consistency/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..07b848f781f0a253b0e5594cbe0d3577a96c7eb0 --- /dev/null +++ b/2024/RePOSE_ 3D Human Pose Estimation via Spatio-Temporal Depth Relational Consistency/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e0b5bd842c6fa559c30a8eb86b0a707cc856ce2ae0e4336a2fd1ae4704890472 +size 611826 diff --git a/2024/RePOSE_ 3D Human Pose Estimation via Spatio-Temporal Depth Relational Consistency/layout.json b/2024/RePOSE_ 3D Human Pose Estimation via Spatio-Temporal Depth Relational Consistency/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4b0597508755575f41dd344fec5163143cc87c2d --- /dev/null +++ b/2024/RePOSE_ 3D Human Pose Estimation via Spatio-Temporal Depth Relational Consistency/layout.json @@ -0,0 +1,9740 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 141, + 111, + 473, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 111, + 473, + 148 + ], + "spans": [ + { + "bbox": [ + 141, + 111, + 473, + 148 + ], + "type": "text", + "content": "RePOSE: 3D Human Pose Estimation via Spatio-Temporal Depth Relational Consistency" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 141, + 167, + 474, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 167, + 474, + 194 + ], + "spans": [ + { + "bbox": [ + 141, + 167, + 474, + 194 + ], + "type": "text", + "content": "Ziming Sun" + }, + { + "bbox": [ + 141, + 167, + 474, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 141, + 167, + 474, + 194 + ], + "type": "text", + "content": ", Yuan Liang" + }, + { + "bbox": [ + 141, + 167, + 474, + 194 + ], + "type": "inline_equation", + "content": "^{1,2\\star}" + }, + { + "bbox": [ + 141, + 167, + 474, + 194 + ], + "type": "text", + "content": ", Zejun Ma" + }, + { + "bbox": [ + 141, + 167, + 474, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 141, + 167, + 474, + 194 + ], + "type": "text", + "content": ", Tianle Zhang" + }, + { + "bbox": [ + 141, + 167, + 474, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 141, + 167, + 474, + 194 + ], + "type": "text", + "content": ", Linchao Bao" + }, + { + "bbox": [ + 141, + 167, + 474, + 194 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 141, + 167, + 474, + 194 + ], + "type": "text", + "content": ", Guiqing Li" + }, + { + "bbox": [ + 141, + 167, + 474, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 141, + 167, + 474, + 194 + ], + "type": "text", + "content": ", and Shengfeng He" + }, + { + "bbox": [ + 141, + 167, + 474, + 194 + ], + "type": "inline_equation", + "content": "^{2(\\boxtimes)}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 208, + 201, + 404, + 224 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 208, + 201, + 403, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 201, + 403, + 213 + ], + "spans": [ + { + "bbox": [ + 208, + 201, + 403, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 208, + 201, + 403, + 213 + ], + "type": "text", + "content": " South China University of Technology, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 208, + 213, + 404, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 213, + 404, + 224 + ], + "spans": [ + { + "bbox": [ + 208, + 213, + 404, + 224 + ], + "type": "text", + "content": "2 Singapore Management University, Singapore" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 254, + 225, + 359, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 225, + 359, + 236 + ], + "spans": [ + { + "bbox": [ + 254, + 225, + 359, + 236 + ], + "type": "text", + "content": "shengfenghe@smu.edu.sg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 238, + 235, + 375, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 235, + 375, + 246 + ], + "spans": [ + { + "bbox": [ + 238, + 235, + 375, + 246 + ], + "type": "text", + "content": "3 Johns Hopkins University, USA" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 255, + 246, + 359, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 246, + 359, + 256 + ], + "spans": [ + { + "bbox": [ + 255, + 246, + 359, + 256 + ], + "type": "text", + "content": "Tencent AI Lab, China" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 220, + 258, + 392, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 258, + 392, + 268 + ], + "spans": [ + { + "bbox": [ + 220, + 258, + 392, + 268 + ], + "type": "text", + "content": "https://github.com/StupidAdam/RePOSE" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 159, + 298, + 455, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 298, + 455, + 518 + ], + "spans": [ + { + "bbox": [ + 159, + 298, + 455, + 518 + ], + "type": "text", + "content": "Abstract. We introduce RePOSE, a simple yet effective approach for addressing occlusion challenges in the learning of 3D human pose estimation (HPE) from videos. Conventional approaches typically employ absolute depth signals as supervision, which are adept at discernible keypoints but become less reliable when keypoints are occluded, resulting in vague and inconsistent learning trajectories for the neural network. RePOSE overcomes this limitation by introducing spatio-temporal relational depth consistency into the supervision signals. The core rationale of our method lies in prioritizing the precise sequencing of occluded keypoints. This is achieved by using a relative depth consistency loss that operates in both spatial and temporal domains. By doing so, RePOSE shifts the focus from learning absolute depth values, which can be misleading in occluded scenarios, to relative positioning, which provides a more robust and reliable cue for accurate pose estimation. This subtle yet crucial shift facilitates more consistent and accurate 3D HPE under occlusion conditions. The elegance of our core idea lies in its simplicity and ease of implementation, requiring only a few lines of code. Extensive experiments validate that RePOSE not only outperforms existing state-of-the-art methods but also significantly enhances the robustness and precision of 3D HPE in challenging occluded environments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 160, + 529, + 453, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 529, + 453, + 540 + ], + "spans": [ + { + "bbox": [ + 160, + 529, + 453, + 540 + ], + "type": "text", + "content": "Keywords: 3D human pose estimation " + }, + { + "bbox": [ + 160, + 529, + 453, + 540 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 529, + 453, + 540 + ], + "type": "text", + "content": " depth relational loss functions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 561, + 230, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 561, + 230, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 561, + 230, + 574 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": "The estimation of 3D human pose from videos is a foundational task in computer vision, with widespread implications across various applications such as virtual reality [4, 12, 21, 38], human-computer interaction [8, 36, 37], and action recognition [10, 34, 35, 40, 41]. This task involves localizing body joints in a three-dimensional space to create a comprehensive representation of the human body," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 366, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 366, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 366, + 665 + ], + "type": "text", + "content": "* The first two authors contributed equally to this work." + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 204, + 182 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 204, + 182 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 204, + 182 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 204, + 182 + ], + "type": "image", + "image_path": "1c0744240422b5ab87e1d08e8c0d44299961764cc9c5722792e646faf3ab17d1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 136, + 182, + 203, + 248 + ], + "blocks": [ + { + "bbox": [ + 136, + 182, + 203, + 248 + ], + "lines": [ + { + "bbox": [ + 136, + 182, + 203, + 248 + ], + "spans": [ + { + "bbox": [ + 136, + 182, + 203, + 248 + ], + "type": "image", + "image_path": "ee5c2813a09c79f0b6d3f807c0c0d33838a997c3b65d0f3ac5972afdbcdede1c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 250, + 187, + 259 + ], + "lines": [ + { + "bbox": [ + 153, + 250, + 187, + 259 + ], + "spans": [ + { + "bbox": [ + 153, + 250, + 187, + 259 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 208, + 115, + 272, + 182 + ], + "blocks": [ + { + "bbox": [ + 208, + 115, + 272, + 182 + ], + "lines": [ + { + "bbox": [ + 208, + 115, + 272, + 182 + ], + "spans": [ + { + "bbox": [ + 208, + 115, + 272, + 182 + ], + "type": "image", + "image_path": "7dcd5bd38044a893f609861415fca27bfdba1859012395069acfeb57eaf5af9b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 208, + 182, + 272, + 248 + ], + "blocks": [ + { + "bbox": [ + 208, + 182, + 272, + 248 + ], + "lines": [ + { + "bbox": [ + 208, + 182, + 272, + 248 + ], + "spans": [ + { + "bbox": [ + 208, + 182, + 272, + 248 + ], + "type": "image", + "image_path": "f5b726767e5dba04012cabc7e6041a6845e5404b85ca52b5fca53de346d6e88e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 224, + 250, + 252, + 259 + ], + "lines": [ + { + "bbox": [ + 224, + 250, + 252, + 259 + ], + "spans": [ + { + "bbox": [ + 224, + 250, + 252, + 259 + ], + "type": "text", + "content": "(b) GT" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 277, + 116, + 340, + 182 + ], + "blocks": [ + { + "bbox": [ + 277, + 116, + 340, + 182 + ], + "lines": [ + { + "bbox": [ + 277, + 116, + 340, + 182 + ], + "spans": [ + { + "bbox": [ + 277, + 116, + 340, + 182 + ], + "type": "image", + "image_path": "8e272043c2572a96e9c8ecb996057f97a01fcb710ba3e731e0c56fe53e5ce617.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 277, + 182, + 340, + 248 + ], + "blocks": [ + { + "bbox": [ + 277, + 182, + 340, + 248 + ], + "lines": [ + { + "bbox": [ + 277, + 182, + 340, + 248 + ], + "spans": [ + { + "bbox": [ + 277, + 182, + 340, + 248 + ], + "type": "image", + "image_path": "d882ac318445d710eea8c43ec415aadab0de81f95b390ea5d435edeb792812e0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 250, + 338, + 259 + ], + "lines": [ + { + "bbox": [ + 276, + 250, + 338, + 259 + ], + "spans": [ + { + "bbox": [ + 276, + 250, + 338, + 259 + ], + "type": "text", + "content": "(c) MotionBERT" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 268, + 482, + 323 + ], + "lines": [ + { + "bbox": [ + 131, + 268, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 131, + 268, + 482, + 323 + ], + "type": "text", + "content": "Fig. 1: We introduce RePOSE, a novel solution designed to address the occlusion challenges in 3D pose estimation through relational supervision. Unlike existing methods (c) [46] and (d) [44], which struggle with occlusions due to their dependence on absolute depth signals, our approach leverages relational depth consistency, effectively overcoming these obstacles." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 345, + 116, + 410, + 182 + ], + "blocks": [ + { + "bbox": [ + 345, + 116, + 410, + 182 + ], + "lines": [ + { + "bbox": [ + 345, + 116, + 410, + 182 + ], + "spans": [ + { + "bbox": [ + 345, + 116, + 410, + 182 + ], + "type": "image", + "image_path": "6bda6a8be2c27fa49d56f5e8980e2ec04eb951de7e13bca235d93d2b760e4ca6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 345, + 182, + 410, + 248 + ], + "blocks": [ + { + "bbox": [ + 345, + 182, + 410, + 248 + ], + "lines": [ + { + "bbox": [ + 345, + 182, + 410, + 248 + ], + "spans": [ + { + "bbox": [ + 345, + 182, + 410, + 248 + ], + "type": "image", + "image_path": "694bf7c0ca073df277bb70a8ccb99aa40fcc167a934bab752673892d379143f3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 250, + 410, + 259 + ], + "lines": [ + { + "bbox": [ + 343, + 250, + 410, + 259 + ], + "spans": [ + { + "bbox": [ + 343, + 250, + 410, + 259 + ], + "type": "text", + "content": "(d) PoseFormerV2" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 414, + 116, + 479, + 182 + ], + "blocks": [ + { + "bbox": [ + 414, + 116, + 479, + 182 + ], + "lines": [ + { + "bbox": [ + 414, + 116, + 479, + 182 + ], + "spans": [ + { + "bbox": [ + 414, + 116, + 479, + 182 + ], + "type": "image", + "image_path": "a79564eefe63fb04713494b2d95003caae75f1582bed2ef21ea534e1a9fbcc87.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 414, + 182, + 479, + 248 + ], + "blocks": [ + { + "bbox": [ + 414, + 182, + 479, + 248 + ], + "lines": [ + { + "bbox": [ + 414, + 182, + 479, + 248 + ], + "spans": [ + { + "bbox": [ + 414, + 182, + 479, + 248 + ], + "type": "image", + "image_path": "06e956d11a28004aab4d74247725c7c16f8b759257fd4192fa47c4ac8cbf8948.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 250, + 460, + 259 + ], + "lines": [ + { + "bbox": [ + 429, + 250, + 460, + 259 + ], + "spans": [ + { + "bbox": [ + 429, + 250, + 460, + 259 + ], + "type": "text", + "content": "(e) Ours" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 350, + 482, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 446 + ], + "type": "text", + "content": "typically in the form of a skeletal model. The methodologies in 3D human pose estimation (HPE) are primarily categorized into two approaches: direct estimation and 2D-to-3D lifting methods. The former [22, 25] entails the direct estimation of human pose from provided videos, whereas the latter [2, 14, 17, 27, 42, 44-46] involves the initial estimation of intermediate 2D poses using off-the-shelf 2D detectors, followed by lifting them to 3D space. 2D-to-3D lifting methods have gained prominence due to the rapid evolution of 2D pose detection technologies, exhibiting superior performance in many scenarios." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 449, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 605 + ], + "type": "text", + "content": "Recent advancements in 2D-to-3D lifting methods often employ Transformer architectures [31], which have shown exceptional ability in capturing global information and modeling sequential data. These methods typically involve transforming detected 2D poses into 3D space using tailored algorithms [13-15,27,42,44-46]. However, this process inherently faces a many-to-one mapping challenge, where multiple distinct 3D poses can correspond to a single 2D projection, particularly under occlusion scenarios, as shown in Fig. 1c and 1d. In such cases, occlusion of key body joints leads to reduced accuracy in 2D pose estimations, creating ambiguities in the 3D reconstruction process. Moreover, relying on these unreliable occluded signals during model training results in vague and uncertain learning pathways. This limitation chiefly stems from the methods' reliance on absolute depth signals, which lose reliability under occluded conditions, thus undermining the overall accuracy of 3D pose estimation." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "To address this challenge, we propose RePOSE, a novel approach specifically designed to enhance the accuracy of 3D HPE in occluded scenarios. RePOSE circumvents the limitations of absolute depth reliance by introducing a spatiotemporal relational depth consistency. Our method emphasizes the correct sequencing of occluded keypoints using a relative depth consistency loss in both" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "Z. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "spatial and temporal domains, rather than depending solely on absolute depth values. This shift in approach enables our method to robustly handle occlusions, providing more accurate and reliable pose estimations, as shown in Fig. 1e. Additionally, the simplicity of spatio-temporal relational depth consistency, which can be implemented in just a few lines of code, makes it a practical and effective solution for real-world applications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 482, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 283 + ], + "type": "text", + "content": "We demonstrate through extensive experiments on the Human3.6M [11] and MPI-INF-3DHP [20] benchmarks that our method outperforms state-of-the-art methods by a large margin, particularly excelling in scenarios involving occlusions. Our findings are further substantiated through visual comparisons and an ablation study, showcasing the effectiveness and innovation of RePOSE in tackling one of the most challenging aspects of 3D human pose estimation. This method lays the groundwork for future research aimed at enhancing pose estimation accuracy and reliability in complex environments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 304, + 242, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 242, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 242, + 316 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 329, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 482, + 544 + ], + "type": "text", + "content": "2D-to-3D Lifting HPE. The prevalent approach in 3D HPE involves lifting 2D pose sequences to 3D poses, a method enhanced by advanced 2D detectors like CPN [3], AlphaPose [7], and HRNet [33]. These detectors facilitate the conversion process with various techniques, including Fully Connected Networks (FCN) [6,19], Long Short-Term Memory (LSTM) [9,16,32], Graph Convolutional Networks (GCN) [1,5,39,43], Temporal Convolutional Networks (TCN) [2,17,26], and Transformers [13,14,27,42,44-46]. While SimpleBaseline [19] and LSTM-based methods [9] highlight the initial strides, they exhibit limitations in computational efficiency and temporal information utilization. GCN-based methods [1,5,39,43], though adept at local joint modeling, struggle with computational complexity. Transformer-based architectures, such as PoseFormer [45], bring significant improvements by modeling spatial-temporal dependencies. The multi-solution nature of 2D-to-3D lifting is addressed by methods like MHFormer [14], which generate multiple hypotheses to account for ambiguity in body part locations. Our work challenges this approach, positing a single accurate solution corresponding to the original 2D pose, driven by spatial and temporal relational constraints to optimize results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": "Loss Functions in 3D Human Pose Estimation. The accuracy of 3D HPE models heavily relies on the choice of loss functions. The Mean Per Joint Position Error (MPJPE) [14, 27, 44-46] is a prevalent metric, gauging the Euclidean distance between predicted and ground truth joint positions. Its simplicity and direct measurement of joint accuracy make it a popular choice. However, it does not account for joint orientations, potentially overlooking critical aspects of pose estimation. The Weighted-MPJPE (W-MPJPE) [42] attempts to refine this by assigning different weights to various joints, thereby prioritizing certain joints over others based on the application's needs. Yet, this method introduces its own set of challenges, including potential bias in weight assignment and the" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 194, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 194, + 91, + 447, + 102 + ], + "type": "text", + "content": "3D Human Pose Estimation via Depth Relational Consistency" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 344 + ], + "type": "text", + "content": "need for expert knowledge in determining these weights. Additionally, the Mean Per Joint Velocity Error (MPJVE) [26,42] has been used to improve the temporal coherence between the predicted pose sequence and the ground truth sequence, which is a crucial aspect in dynamic pose estimation. However, it falls short in leveraging long-range temporal information. To address this gap, Zhang et al. [42] integrated the MPJPE and temporal consistency losses introduced by Hossain et al. [9] to leverage temporal relationships and smooth pose transitions. However, these adaptations insufficiently tackle the issue of relative depth perception among joints. Various studies [29, 46] attempt to quantify limb angle inaccuracies by calculating the Mean Absolute Error between the estimated and ground-truth skeletal angles. Additionally, they explore angle dynamics through the Mean Per-Angle Velocity Error (MPAVE), representing the mean Euclidean distance in the angle's first temporal derivative. Despite these efforts, current methodologies overly focus on the x and y dimensions, derived primarily from 2D pose estimations, neglecting the critical depth (z-axis) disparities. Alternatively, Pavlakos et al. [25] utilize predicted depth values to compute the ordinal depth loss, especially when handling datasets that lack ground truth labels. However, this approach offers a weaker supervisory signal due to the inherent unreliability of the predictions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 345, + 482, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 482, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 482, + 465 + ], + "type": "text", + "content": "In response to the above problems, we introduce two novel loss functions: spatial and temporal relational depth consistency losses. The spatial loss focuses on relative depth differences between joints within a single frame, ensuring accurate depth relationships and mitigating pose estimation inconsistencies. The temporal loss maintains pose smoothness over time, considering relative depth differences of the same joint across frames. This dual focus enhances the overall robustness of 3D HPE, ensuring more accurate, temporally coherent, and visually appealing human pose sequences. Our methodology, utilizing these losses, aims to significantly improve the accuracy and temporal stability of pose estimation, making it more applicable and reliable in various computer vision tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 488, + 208, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 488, + 208, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 488, + 208, + 502 + ], + "type": "text", + "content": "3 RePOSE" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 520, + 362, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 362, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 362, + 533 + ], + "type": "text", + "content": "3.1 Problem Formulation and Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Our method leverages the 2D-to-3D lifting pipeline to elevate a sequence of 2D skeletons to their corresponding 3D coordinates. Given a sequence of 2D joints with confidence scores, represented as " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{T \\times J \\times 3}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " denote the number of frames and joints respectively, this input is transformed into a d-dimensional feature space " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "F \\in \\mathbb{R}^{T \\times J \\times d}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " by the lifting network, with " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " varying across different network architectures. The output of this process is the estimated 3D pose sequence, denoted as " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{Y} \\in \\mathbb{R}^{T \\times J \\times 3}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ", obtained through a regression mechanism. In our approach, we specifically treat the third dimension of " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " as the depth feature " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{D} \\in \\mathbb{R}^{T \\times J}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ", facilitating the computation of relational depth consistency losses for model training supervision." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Z. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 479, + 269 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 269 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 269 + ], + "type": "image", + "image_path": "e7032e65c42f3b007f1ebee51fd94053d288f084a8f699b782dee065ae405815.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 279, + 482, + 357 + ], + "lines": [ + { + "bbox": [ + 130, + 279, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 482, + 357 + ], + "type": "text", + "content": "Fig. 2: Our framework pipeline. Our network is designed with a spatio-temporal architecture, enabling simultaneous modeling of both spatial and temporal information. The generated 3D poses are processed through dual loss modules: the spatial loss module focuses on the relative depth disparities among joints within each frame, while the temporal loss module evaluates the relative depth differences of identical joints across consecutive frames. This dual-module approach ensures comprehensive depth accuracy and temporal consistency in the pose estimation process." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 385, + 480, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 385, + 480, + 410 + ], + "spans": [ + { + "bbox": [ + 131, + 385, + 480, + 410 + ], + "type": "text", + "content": "Prevailing methodologies in recent studies [13,14,27,42,44-46] predominantly utilize Mean Per Joint Position Error (MPJPE) as loss function, defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 238, + 423, + 480, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 423, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 238, + 423, + 480, + 456 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m p j p e} = \\sum_ {t = 1} ^ {T} \\sum_ {i = 1} ^ {J} \\| \\hat {y} _ {t, i} - y _ {t, i} \\| _ {2}, \\tag {1}", + "image_path": "c9fb7256c41203d372ab9a317f9474211a742267b224fdf324559fe19da45847.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 469, + 479, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 469, + 479, + 505 + ], + "spans": [ + { + "bbox": [ + 131, + 469, + 479, + 505 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 469, + 479, + 505 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 469, + 479, + 505 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 469, + 479, + 505 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 131, + 469, + 479, + 505 + ], + "type": "text", + "content": " represent the total number of frames and joints in each video, " + }, + { + "bbox": [ + 131, + 469, + 479, + 505 + ], + "type": "inline_equation", + "content": "\\hat{y}_{t,i}" + }, + { + "bbox": [ + 131, + 469, + 479, + 505 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 469, + 479, + 505 + ], + "type": "inline_equation", + "content": "y_{t,i}" + }, + { + "bbox": [ + 131, + 469, + 479, + 505 + ], + "type": "text", + "content": " correspond to the estimated and ground-truth positions of keypoint " + }, + { + "bbox": [ + 131, + 469, + 479, + 505 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 469, + 479, + 505 + ], + "type": "text", + "content": " in frame " + }, + { + "bbox": [ + 131, + 469, + 479, + 505 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 469, + 479, + 505 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 507, + 480, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 480, + 579 + ], + "type": "text", + "content": "While the MPJPE is a widely accepted metric in 3D HPE, it has limitations, particularly in its focus on joint positions without explicit consideration of spatial relationships or pose topology. This becomes a significant concern in scenarios with occlusions, where understanding the relative positions of body parts is vital for accurate pose estimation. To address this, we introduce the spatial depth ranking loss, specifically designed to enhance pose topology awareness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 666 + ], + "type": "text", + "content": "Additionally, conventional techniques often fail to effectively utilize temporal information, as they typically compute loss independently for each frame, neglecting the continuity and interdependence within video sequences. This oversight can significantly limit a model's ability to capture realistic human motion over time, a crucial aspect in dynamic pose estimation tasks. To mitigate this gap, we propose the temporal depth ranking loss, a novel component aimed at incorporating temporal dynamics for more accurate and lifelike pose estimation." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 194, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 194, + 91, + 447, + 102 + ], + "type": "text", + "content": "3D Human Pose Estimation via Depth Relational Consistency" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 256, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 256, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 256, + 126 + ], + "type": "text", + "content": "3.2 Network Structure" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 133, + 482, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 133, + 482, + 228 + ], + "spans": [ + { + "bbox": [ + 130, + 133, + 482, + 228 + ], + "type": "text", + "content": "Inspired by [27,45], we recognize the exemplary performance of spatial-temporal transformer architectures in 3D human pose estimation (HPE) due to their robust capability to model both spatial and temporal information in video sequences. In particular, they demonstrate a parallel dual-stream design to be effective in ensuring comprehensive learning of spatial and temporal aspects. Consequently, our network, RePOSE, adopts a bifurcated structure of the spatiotemporal transformer to enhance the capture of spatial and temporal features from input videos." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 228, + 482, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 482, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 482, + 360 + ], + "type": "text", + "content": "As depicted in Fig. 2, RePOSE is composed of a spatio-temporal module and two specialized loss modules. The spatial transformer encoding module processes individual body joints as distinct tokens, thereby efficiently capturing the inter-joint relationships within the same frame. In contrast, the temporal transformer encoding module considers cross-frame features, enabling it to discern relationships between joints across different frames. The outputs from these modules are then directed to the spatial and temporal relational depth consistency modules, which compute the respective losses. A crucial aspect of our architecture is the backward pass, which plays a vital role in guiding the model's training process. The nuances and operations of the spatial and temporal relational depth consistency modules are elaborated in subsequent sections." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 373, + 350, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 373, + 350, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 373, + 350, + 385 + ], + "type": "text", + "content": "3.3 Spatial Relational Depth Consistency" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 390, + 301, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 301, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 301, + 616 + ], + "type": "text", + "content": "In any given frame, the ideal relationship between estimated depths " + }, + { + "bbox": [ + 130, + 390, + 301, + 616 + ], + "type": "inline_equation", + "content": "\\hat{d}" + }, + { + "bbox": [ + 130, + 390, + 301, + 616 + ], + "type": "text", + "content": " and ground-truth depths " + }, + { + "bbox": [ + 130, + 390, + 301, + 616 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 390, + 301, + 616 + ], + "type": "text", + "content": " should adhere to the condition " + }, + { + "bbox": [ + 130, + 390, + 301, + 616 + ], + "type": "inline_equation", + "content": "\\hat{d}_{t,i} < \\hat{d}_{t,j}" + }, + { + "bbox": [ + 130, + 390, + 301, + 616 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 130, + 390, + 301, + 616 + ], + "type": "inline_equation", + "content": "d_{t,i} < d_{t,j}" + }, + { + "bbox": [ + 130, + 390, + 301, + 616 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 390, + 301, + 616 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 390, + 301, + 616 + ], + "type": "text", + "content": " denotes a specific time step, and " + }, + { + "bbox": [ + 130, + 390, + 301, + 616 + ], + "type": "inline_equation", + "content": "d_{t,i}" + }, + { + "bbox": [ + 130, + 390, + 301, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 390, + 301, + 616 + ], + "type": "inline_equation", + "content": "d_{t,j}" + }, + { + "bbox": [ + 130, + 390, + 301, + 616 + ], + "type": "text", + "content": " represent the depths of different joints " + }, + { + "bbox": [ + 130, + 390, + 301, + 616 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 390, + 301, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 390, + 301, + 616 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 390, + 301, + 616 + ], + "type": "text", + "content": " at that moment, respectively. The same applies for " + }, + { + "bbox": [ + 130, + 390, + 301, + 616 + ], + "type": "inline_equation", + "content": "\\hat{d}_{t,i}" + }, + { + "bbox": [ + 130, + 390, + 301, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 390, + 301, + 616 + ], + "type": "inline_equation", + "content": "\\hat{d}_{t,j}" + }, + { + "bbox": [ + 130, + 390, + 301, + 616 + ], + "type": "text", + "content": ". However, we observe that this relative depth relationship is often overlooked in existing methods, leading to inaccurate results, especially in scenarios with noise such as self-occlusion. To address this, we introduce a novel loss function to the training procedure, aimed at steering the network optimization in a more accurate direction. The formula is as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 620, + 299, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 620, + 299, + 666 + ], + "spans": [ + { + "bbox": [ + 136, + 620, + 299, + 666 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S} = \\sum_ {t = 1} ^ {T} \\sum_ {d _ {t, i} < d _ {t, j}} \\max \\left(0, \\hat {d} _ {t, i} - \\hat {d} _ {t, j}\\right), \\tag {2}", + "image_path": "1f79ea4da83c3c8f61f0b50b22fc24416bd187cc8b7ab94f157b5f1c0866461a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 398, + 482, + 512 + ], + "blocks": [ + { + "bbox": [ + 306, + 398, + 482, + 512 + ], + "lines": [ + { + "bbox": [ + 306, + 398, + 482, + 512 + ], + "spans": [ + { + "bbox": [ + 306, + 398, + 482, + 512 + ], + "type": "image", + "image_path": "8b9bd58a00290c6342b649f2a6a062516279bfb3397bfd6c37e7f89ce7dffaab.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 519, + 482, + 629 + ], + "lines": [ + { + "bbox": [ + 304, + 519, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 519, + 482, + 629 + ], + "type": "text", + "content": "Fig. 3: Overview of spatio-temporal relational depth consistency loss. The spatial loss (left) enhances relative depth relationships between the in-frame joints. The temporal loss (right) ensures the consistent relative depth relationships of joints across frames. The green dashed lines illustrate examples where the loss is calculated for the relative depth of linked joints against the ground-truth relative depth." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Z. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "Eq. (2) demonstrates that our model penalizes deviations in the ranking of estimated depths from the ground-truth rankings. An illustration of this loss is displayed in the left part of Fig. 3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 153, + 480, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 153, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 153, + 480, + 308 + ], + "type": "text", + "content": "In cases where body parts are occluded, standard models often struggle to accurately determine the depth positioning of joints, leading to spatial inconsistencies. By integrating our spatial depth rank loss, we guide the estimations to align more closely with the actual spatial configuration, thereby providing feedback to the model for improved training and enhanced differentiation of obscured body parts or joints. Particularly in self-occlusion scenarios, traditional approaches might incorrectly associate body joints, such as mistaking a hand for an elbow due to overlapping positions. This type of error, which might go unnoticed when solely using MPJPE as the loss function, becomes apparent when the human skeleton is visualized. Our proposed spatial depth rank loss function, by reinforcing the inter-joint relationships within a frame, effectively prevents these errors and ensures a more accurate representation of the human body structure. The code of this loss is displayed in Algorithm 1." + } + ] + } + ], + "index": 3 + }, + { + "type": "code", + "bbox": [ + 130, + 323, + 489, + 554 + ], + "blocks": [ + { + "bbox": [ + 130, + 323, + 489, + 554 + ], + "lines": [ + { + "bbox": [ + 130, + 323, + 489, + 554 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 489, + 554 + ], + "type": "text", + "content": "Algorithm 1 Spatial Relational Depth Consistency \ndef process(x, J, idx): def process(x, T, idx): x = x.Unsqueeze(idx) return xrepeat_interleave(J, idx) return xrepeat_interleave(T, idx) \ndef loss_spatial(x, gt): def loss_temporal(x, gt): #x, gt: [B, T, J, 3] #x, gt: [B, T, J, 3] _, , J, = x.shape _, T, , = x.shape x = x.permute(0, 2, 1) [... , 2] gt = gt.permute(0, 2, 1) [... , 2] A = process(gt, J, -1) A = process(gt, T, -1) B = process(gt, J, -2) B = process(gt, T, -2) C = process(x, J, -1) C = process(x, T, -1) D = process(x, J, -2) D = process(x, T, -2) Comp_1 = (C - D) * (A < B) Comp_2 = torch.zeros_like(C) Comp_2 = torch.zeros_like(C) SRLoss = torch.maxmin( Comp_1, Comp_2) return torch.mean(SRLoss) return torch.mean(TRLoss)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "code_body" + } + ], + "index": 4, + "sub_type": "algorithm" + }, + { + "bbox": [ + 132, + 579, + 361, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 361, + 592 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 361, + 592 + ], + "type": "text", + "content": "3.4 Temporal Relational Depth Consistency" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "content": "In parallel with our spatial considerations, we address depth consistency in the temporal dimension. Temporally, the depth ranking of a specific joint should remain consistent across different time steps to ensure the continuity of the 3D pose sequence. However, challenges arise, particularly during self-occlusion, where the estimated poses may disrupt this continuity. To counteract this, we" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 194, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 194, + 91, + 447, + 102 + ], + "type": "text", + "content": "3D Human Pose Estimation via Depth Relational Consistency" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 296, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 296, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 296, + 128 + ], + "type": "text", + "content": "introduce the following loss function:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 220, + 137, + 481, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 137, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 220, + 137, + 481, + 172 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {T} = \\sum_ {j = 1} ^ {J} \\sum_ {d _ {m, j} < d _ {n, j}} \\max \\left(0, \\hat {d} _ {m, j} - \\hat {d} _ {n, j}\\right), \\tag {3}", + "image_path": "cdcabfe87818809619e84aeec92ec09c19a8122735ddaeaaa7ad671be2e7955a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 181, + 479, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 181, + 479, + 217 + ], + "spans": [ + { + "bbox": [ + 130, + 181, + 479, + 217 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 181, + 479, + 217 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 181, + 479, + 217 + ], + "type": "text", + "content": " indicates a specific joint, " + }, + { + "bbox": [ + 130, + 181, + 479, + 217 + ], + "type": "inline_equation", + "content": "d_{m,j}" + }, + { + "bbox": [ + 130, + 181, + 479, + 217 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 181, + 479, + 217 + ], + "type": "inline_equation", + "content": "d_{n,j}" + }, + { + "bbox": [ + 130, + 181, + 479, + 217 + ], + "type": "text", + "content": " represent the depths of joint " + }, + { + "bbox": [ + 130, + 181, + 479, + 217 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 181, + 479, + 217 + ], + "type": "text", + "content": " at different times " + }, + { + "bbox": [ + 130, + 181, + 479, + 217 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 181, + 479, + 217 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 181, + 479, + 217 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 181, + 479, + 217 + ], + "type": "text", + "content": ", similarly for " + }, + { + "bbox": [ + 130, + 181, + 479, + 217 + ], + "type": "inline_equation", + "content": "\\hat{d}_{m,j}" + }, + { + "bbox": [ + 130, + 181, + 479, + 217 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 181, + 479, + 217 + ], + "type": "inline_equation", + "content": "\\hat{d}_{n,j}" + }, + { + "bbox": [ + 130, + 181, + 479, + 217 + ], + "type": "text", + "content": ". The right part of Fig. 3 shows the idea of this loss." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 218, + 480, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 218, + 480, + 301 + ], + "spans": [ + { + "bbox": [ + 130, + 218, + 480, + 301 + ], + "type": "text", + "content": "The temporal depth ranking loss encourages each joint in the estimated poses to maintain temporal consistency, preventing any joint from deviating suddenly from its original trajectory within the sequence. This loss is particularly crucial in ensuring that the temporal relationships between consecutive frames in video sequences are adequately maintained, which aids in smoothing the motion and making the model less susceptible to outliers. The implementation is detailed in Algorithm 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 320, + 239, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 320, + 239, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 320, + 239, + 332 + ], + "type": "text", + "content": "3.5 Final Objective" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 341, + 479, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 479, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 479, + 413 + ], + "type": "text", + "content": "Our comprehensive approach integrates the spatial and temporal dimensions to significantly enhance the coherence and realism of estimated 3D pose sequences. This integration ensures that the movements captured are not only spatially precise but also exhibit temporal fluidity and lifelike dynamics. Our final objective function combines the conventional MPJPE loss with our spatio-temporal relational depth consistency losses. The final loss function is formulated as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 241, + 425, + 481, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 425, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 241, + 425, + 481, + 439 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {m p j p e} + \\lambda_ {S} \\mathcal {L} _ {S} + \\lambda_ {T} \\mathcal {L} _ {T}, \\tag {4}", + "image_path": "cb5c60fddd9c55d0d806802baab926847cd0b3517cb18d2aadb2873e4df75fe3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 444, + 480, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 444, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 130, + 444, + 480, + 503 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 444, + 480, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{L}mpjpe" + }, + { + "bbox": [ + 130, + 444, + 480, + 503 + ], + "type": "text", + "content": " is the standard MPJPE loss, and the parameters " + }, + { + "bbox": [ + 130, + 444, + 480, + 503 + ], + "type": "inline_equation", + "content": "\\lambda_{S}" + }, + { + "bbox": [ + 130, + 444, + 480, + 503 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 444, + 480, + 503 + ], + "type": "inline_equation", + "content": "\\lambda_{T}" + }, + { + "bbox": [ + 130, + 444, + 480, + 503 + ], + "type": "text", + "content": " are weighting factors that balance the contribution of the spatial and temporal components in the overall loss function. This unified loss function is designed to optimize the network in a holistic manner, addressing both the individual joint accuracy and the relational depth in both spatial and temporal dimensions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 523, + 230, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 230, + 537 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 230, + 537 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 548, + 326, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 326, + 560 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 326, + 560 + ], + "type": "text", + "content": "4.1 Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 570, + 479, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 479, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 479, + 593 + ], + "type": "text", + "content": "We have conducted training and evaluation of our model using the datasets Human3.6M [11] and MPI-INF-3DHP [20]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "content": "Human3.6M is a well-established indoor dataset widely employed in the field of 3D HPE. This dataset comprises 3.6 million video frames captured from four distinct viewing angles, featuring performances by eleven professional actors engaged in diverse activities, including sitting, eating, and walking and more. In line with research [2,17,19,26,45,46], we utilize subjects 1, 5, 6, 7, and 8 as our training dataset, reserving subjects 9 and 11 for testing purposes." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "Z. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 203, + 183 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 203, + 183 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 203, + 183 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 203, + 183 + ], + "type": "image", + "image_path": "4c22ed4ec878a98191d4e6f9a52a51fda529caf646c4ff9a1fd9c54586a8c73c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 136, + 183, + 203, + 248 + ], + "blocks": [ + { + "bbox": [ + 136, + 183, + 203, + 248 + ], + "lines": [ + { + "bbox": [ + 136, + 183, + 203, + 248 + ], + "spans": [ + { + "bbox": [ + 136, + 183, + 203, + 248 + ], + "type": "image", + "image_path": "47930450a25966683445342f95cf83f23d1037c40b33fb3ee0e3204c62896e5e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 250, + 186, + 259 + ], + "lines": [ + { + "bbox": [ + 153, + 250, + 186, + 259 + ], + "spans": [ + { + "bbox": [ + 153, + 250, + 186, + 259 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 206, + 115, + 270, + 181 + ], + "blocks": [ + { + "bbox": [ + 206, + 115, + 270, + 181 + ], + "lines": [ + { + "bbox": [ + 206, + 115, + 270, + 181 + ], + "spans": [ + { + "bbox": [ + 206, + 115, + 270, + 181 + ], + "type": "image", + "image_path": "7a20d2a5405214e98e28acb93750bc5d6445bd313866c22e6b44eeeb40635883.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 206, + 182, + 270, + 248 + ], + "blocks": [ + { + "bbox": [ + 206, + 182, + 270, + 248 + ], + "lines": [ + { + "bbox": [ + 206, + 182, + 270, + 248 + ], + "spans": [ + { + "bbox": [ + 206, + 182, + 270, + 248 + ], + "type": "image", + "image_path": "3c25018181092d1f98b2488e994461bb62411e32e143746e6b9b4b6cd3594e09.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 224, + 250, + 252, + 259 + ], + "lines": [ + { + "bbox": [ + 224, + 250, + 252, + 259 + ], + "spans": [ + { + "bbox": [ + 224, + 250, + 252, + 259 + ], + "type": "text", + "content": "(b) GT" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 274, + 115, + 339, + 181 + ], + "blocks": [ + { + "bbox": [ + 274, + 115, + 339, + 181 + ], + "lines": [ + { + "bbox": [ + 274, + 115, + 339, + 181 + ], + "spans": [ + { + "bbox": [ + 274, + 115, + 339, + 181 + ], + "type": "image", + "image_path": "fab57cea4bf05a10c4073afb39931eff242846a39e9a695fc13ebd5710a33319.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 275, + 182, + 339, + 248 + ], + "blocks": [ + { + "bbox": [ + 275, + 182, + 339, + 248 + ], + "lines": [ + { + "bbox": [ + 275, + 182, + 339, + 248 + ], + "spans": [ + { + "bbox": [ + 275, + 182, + 339, + 248 + ], + "type": "image", + "image_path": "b78b878721b03295a4c4796ae7fc1941daa0fb7df4e91b3d25fed14a40f6a697.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 250, + 337, + 259 + ], + "lines": [ + { + "bbox": [ + 277, + 250, + 337, + 259 + ], + "spans": [ + { + "bbox": [ + 277, + 250, + 337, + 259 + ], + "type": "text", + "content": "(c) MotionBERT" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 269, + 482, + 323 + ], + "lines": [ + { + "bbox": [ + 131, + 269, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 131, + 269, + 482, + 323 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparison on Human3.6M. Our method is evaluated against MotionBERT [46] and PoseFormerV2 [44] in scenarios featuring extensive self-occlusion of the human body. Areas with incorrect relative depth relationships are marked with red circles for emphasis, illustrating the effectiveness of each method in handling complex occlusion challenges." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 343, + 115, + 407, + 181 + ], + "blocks": [ + { + "bbox": [ + 343, + 115, + 407, + 181 + ], + "lines": [ + { + "bbox": [ + 343, + 115, + 407, + 181 + ], + "spans": [ + { + "bbox": [ + 343, + 115, + 407, + 181 + ], + "type": "image", + "image_path": "934db952d9968d255e3a775800581501521b0d98911e1028661f8e81831cf519.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 343, + 182, + 407, + 248 + ], + "blocks": [ + { + "bbox": [ + 343, + 182, + 407, + 248 + ], + "lines": [ + { + "bbox": [ + 343, + 182, + 407, + 248 + ], + "spans": [ + { + "bbox": [ + 343, + 182, + 407, + 248 + ], + "type": "image", + "image_path": "aaadb08ebd0dd2175e757b7cc0312a03bac03415b9c9354c780f469ea052f716.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 250, + 409, + 259 + ], + "lines": [ + { + "bbox": [ + 343, + 250, + 409, + 259 + ], + "spans": [ + { + "bbox": [ + 343, + 250, + 409, + 259 + ], + "type": "text", + "content": "(d) PoseFormerV2" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 411, + 115, + 477, + 181 + ], + "blocks": [ + { + "bbox": [ + 411, + 115, + 477, + 181 + ], + "lines": [ + { + "bbox": [ + 411, + 115, + 477, + 181 + ], + "spans": [ + { + "bbox": [ + 411, + 115, + 477, + 181 + ], + "type": "image", + "image_path": "0d19798756dd2a26c91ffe0e46824c8a6e4a63fc5bf68828598e21907b0aa15a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 413, + 182, + 477, + 248 + ], + "blocks": [ + { + "bbox": [ + 413, + 182, + 477, + 248 + ], + "lines": [ + { + "bbox": [ + 413, + 182, + 477, + 248 + ], + "spans": [ + { + "bbox": [ + 413, + 182, + 477, + 248 + ], + "type": "image", + "image_path": "5ae4b58a58bddc1fd7fdaa94c17f877c04a37b4d2b781e8b360b40f895eb1a2a.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 250, + 460, + 259 + ], + "lines": [ + { + "bbox": [ + 429, + 250, + 460, + 259 + ], + "spans": [ + { + "bbox": [ + 429, + 250, + 460, + 259 + ], + "type": "text", + "content": "(e) Ours" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 345, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 482, + 429 + ], + "type": "text", + "content": "In our evaluation process, we report two key metrics: MPJPE and Procrustes Aligned-MPJPE (PA-MPJPE), both measured in millimeters. MPJPE assesses the average Euclidean distance between the estimated joint positions and their corresponding ground truth values. In contrast, the PA-MPJPE represents the MPJPE values after performing a rigid alignment procedure involving translation, rotation, and scaling operations to bring the estimated 3D pose as close as possible to the ground truth poses." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 429, + 482, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 482, + 549 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 482, + 549 + ], + "type": "text", + "content": "MPI-INF-3DHP is another large-scale 3D HPE dataset consisting of both constrained indoor and challenging outdoor scenes. It records 8 actors performing 8 activities. It consists of over 1.3 million frames captured from the 14 cameras. To rigorously assess the generalization capabilities of RePOSE, particularly in challenging outdoor environments and scenarios involving occlusions, we conducted a comprehensive evaluation of our model using the MPI-INF-3DHP dataset. Here we utilize 2D pose sequences with a length of 81 frames each as input, as distinct from what it is in Human3.6M. Following previous works [2, 32, 44], we report MPJPE, Percentage of Correct Keypoints (PCK) within the " + }, + { + "bbox": [ + 130, + 429, + 482, + 549 + ], + "type": "inline_equation", + "content": "150\\mathrm{mm}" + }, + { + "bbox": [ + 130, + 429, + 482, + 549 + ], + "type": "text", + "content": " range as well as Area Under the Curve (AUC)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 131, + 563, + 279, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 563, + 279, + 575 + ], + "spans": [ + { + "bbox": [ + 131, + 563, + 279, + 575 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "We have implemented our methodology using PyTorch [24] on a computing platform consisting of two NVIDIA RTX 3090 GPUs. Following previous studies [29, 44, 46], we have configured the video sequence length to 243 frames and leveraged the Stacked Hourglass [23] for the extraction of 2D poses during both training and evaluation phases. Additionally, we have applied horizontal flipping augmentation, consistent with [1, 2, 46]. For the optimization of model parameters, we have employed the AdamW [18] optimizer for 100 training epochs," + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 194, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 194, + 91, + 448, + 102 + ], + "type": "text", + "content": "3D Human Pose Estimation via Depth Relational Consistency" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 211, + 480, + 439 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "text", + "content": "Table 1: Quantitative Comparison on Human3.6M Dataset. The top and bottom show comparison based on MPJPE and Procrustes-Aligned MPJPE in millimeters respectively. In this table, " + }, + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "text", + "content": " represents the number of input frames used in each method. The best-performing methods are highlighted in bold, while the second-best results are underlined for clarity. Different 2D detection methods are indicated as follows: (*) denotes the use of CPN [3], (" + }, + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "text", + "content": ") indicates the use of HRNet [33], and (" + }, + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "text", + "content": ") signifies the use of SH [23] for 2D pose detection." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 211, + 480, + 439 + ], + "lines": [ + { + "bbox": [ + 133, + 211, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 133, + 211, + 480, + 439 + ], + "type": "table", + "html": "
MPJPETDir. Disc. Eat. GreetPhonePhotoPosePurch. Sit SitD.SmokeWaitWalkD.WalkWalkT.
*VideoPose3D (CVPR'19) [26]24345.246.743.345.648.155.144.644.357.365.847.144.049.032.833.9
*PoseFormer (ICCV'21) [45]8141.544.839.842.546.551.642.142.053.360.745.543.346.131.832.2
*MHFormer (CVPR'22) [14]35139.243.140.140.944.951.240.641.353.560.343.741.143.829.830.6
†MixSTE (CVPR'22) [42]24336.739.036.539.440.244.939.836.947.954.839.637.839.329.730.6
*P-STMO (ECCV'22) [27]24338.442.139.840.245.248.940.438.353.857.343.941.642.229.329.3
*PoseFormerV2 (CVPR'23) [44]24341.245.441.444.046.653.742.642.655.164.645.742.845.732.332.8
*STCFformer (CVPR'23) [30]24338.441.236.838.042.750.538.738.252.556.841.838.440.226.227.7
*D3DP (ICCV'23) [28]24337.739.935.738.241.948.839.538.350.553.941.639.439.827.427.5
‡MotionBERT (ICCV'23) [46]24336.137.535.832.140.346.336.135.346.953.939.536.335.825.125.3
‡MoAGFormer (WACV'24) [29]24336.838.535.933.041.148.638.034.849.051.440.337.436.327.227.2
‡RePOSE (Ours)24334.636835.031.338.844.735.335.547.550.938.435.534.923.925.036.5
PA-MPJPETDir. Disc. Eat. GreetPhonePhotoPosePurch. Sit SitD.SmokeWaitWalkD.WalkWalkT.
*VideoPose3D (CVPR'19) [26]24334.136.134.437.236.442.234.433.645.052.537.433.837.825.627.3
*PoseFormer (ICCV'21) [45]8132.534.832.634.635.339.532.132.042.848.534.832.435.324.526.0
*MHFormer (CVPR'22) [14]35131.534.932.833.635.339.632.032.243.548.736.432.634.323.925.1
*MixSTE (CVPR'22) [42]24330.833.130.331.833.139.131.130.542.544.534.030.832.722.122.9
*P-STMO (ECCV'22) [27]24331.335.232.933.935.439.332.531.544.648.236.332.934.423.823.9
*PoseFormerV2 (CVPR'23) [44]24332.335.833.735.835.941.133.132.744.251.937.432.835.525.226.5
*STCFformer (CVPR'23) [30]24329.333.030.730.632.738.229.728.842.245.033.329.431.520.922.3
*D3DP (ICCV'23) [28]24330.632.529.131.031.937.630.329.440.643.633.330.531.421.522.4
‡MotionBERT (ICCV'23) [46]24329.531.332.027.032.837.229.829.641.048.234.330.130.221.221.8
‡MoAGFormer (WACV'24) [29]24331.032.631.027.934.038.731.530.041.445.434.830.831.322.823.2
‡RePOSE (Ours)24329.130.929.426.632.135.828.630.639.944.933.329.029.620.121.2
", + "image_path": "323ae3a81d550612becd749e720e409a35b85352885d75f1386b9afb3389db78.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 464, + 482, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 550 + ], + "type": "text", + "content": "incorporating a weight decay factor of 0.01. The initial learning rate has been set to " + }, + { + "bbox": [ + 130, + 464, + 482, + 550 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 464, + 482, + 550 + ], + "type": "text", + "content": ", accompanied by a learning rate decay strategy with a decay rate of 0.99. Lastly, the hyper-parameters " + }, + { + "bbox": [ + 130, + 464, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\lambda_{S}" + }, + { + "bbox": [ + 130, + 464, + 482, + 550 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 464, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\lambda_{T}" + }, + { + "bbox": [ + 130, + 464, + 482, + 550 + ], + "type": "text", + "content": " are all set to 20. When training our model on Human3.6M, We utilize the stride data sample strategy with a one-third interval of the input length to make sure there are fewer overlapping frames between consecutive sequences. But for the MPI-INF-3DHP dataset, the interval is one-nine of the input length because of fewer samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 570, + 380, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 570, + 380, + 582 + ], + "spans": [ + { + "bbox": [ + 131, + 570, + 380, + 582 + ], + "type": "text", + "content": "4.3 Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Human3.6M dataset. The proposed method has been seriously compared with state-of-the-art (SOTA) approaches, utilizing the Human3.6M dataset as the evaluation benchmark. As illustrated in Tab. 1, our method demonstrates exceptional performance, achieving an MPJPE of 36.5 millimeters and a PA-MPJPE of 30.7 millimeters. These results surpass those achieved by leading SOTA methods, indicating a notable performance improvement." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Z. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 175, + 146, + 436, + 262 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 2: Computational Complexity Comparison. The MPJPEs here are evaluated on Human3.6M using ground-truth 2D joint locations as input." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 175, + 146, + 436, + 262 + ], + "lines": [ + { + "bbox": [ + 175, + 146, + 436, + 262 + ], + "spans": [ + { + "bbox": [ + 175, + 146, + 436, + 262 + ], + "type": "table", + "html": "
MethodParam (M)MACs (G)MPJPE
PoseFormer (ICCV'21) [45]9.52.531.3
MHFormer (CVPR'22) [14]30.97.030.5
MixSTE (CVPR'22) [42]33.6139.025.9
P-STMO (ECCV'22) [27]6.20.729.3
PoseFormerV2 (CVPR'23) [44]14.30.5-
STCFormer (CVPR'23) [30]4.719.621.3
MotionBERT (ICCV'23) [46]42.5174.717.8
MotionAGFormer (WACV'24) [29]19.078.317.3
RePOSE (Ours)16.043.715.7
", + "image_path": "46b7cec7a8be7cfc84e78cca1e8f834f77c7d430cadec3fdcc4ef2619fe6d934.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 179, + 317, + 435, + 436 + ], + "blocks": [ + { + "bbox": [ + 132, + 276, + 480, + 297 + ], + "lines": [ + { + "bbox": [ + 132, + 276, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 132, + 276, + 480, + 297 + ], + "type": "text", + "content": "Table 3: Quantitative comparison on MPI-INF-3DHP dataset. The best and second-best results are bolded and underlined." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 179, + 317, + 435, + 436 + ], + "lines": [ + { + "bbox": [ + 179, + 317, + 435, + 436 + ], + "spans": [ + { + "bbox": [ + 179, + 317, + 435, + 436 + ], + "type": "table", + "html": "
MethodPCK↑AUC↑MPJPE↓
VideoPose3D (CVPR'19) [26]85.551.584.8
PoseFormer (ICCV'21) [45]65.463.257.7
MHFormer (CVPR'22) [14]93.863.358.0
MixSTE (CVPR'22) [42]94.466.554.9
P-STMO (ECCV'22) [27]97.975.832.2
D3DP (ICCV'23) [28]97.777.830.2
PoseFormerV2 (CVPR'23) [44]97.978.827.8
STCFormaler (CVPR'23) [30]98.783.923.1
MotionAGFormer (WACV'24) [29]98.285.316.2
RePOSE (Ours)98.386.715.5
", + "image_path": "be18d433d79bd0a8b944f5a5fc25e49d914f536739bea6164dffd6f63293e16b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 471, + 480, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 480, + 530 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 480, + 530 + ], + "type": "text", + "content": "It is worth noting that even when considering the best results reported by each of the prior studies, our method consistently outperforms them, especially in challenging scenarios involving actions such as \"photo,\" \"sitting,\" and \"sitting down.\" This highlights its potential to effectively address challenges related to occlusion. Qualitative results can be found in Fig. 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 532, + 480, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 532, + 480, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 532, + 480, + 591 + ], + "type": "text", + "content": "We also compare the number of parameters, multiply-accumulate operations (MACs), and MPJPE using the ground-truth 2D pose sequences as input, as shown in Tab. 2. Our model attains the best MPJPE out of other works with limited parameters and algorithm complexity, demonstrating the effectiveness and efficiency of our proposed method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": "MPI-INF-3DHP dataset. Tab. 3 presents a quantitative comparison of our method on the MPI-INF-3DHP dataset. RePOSE achieves impressive results, with an MPJPE of " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "15.5\\mathrm{mm}" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": " and an Area Under Curve (AUC) of " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "86.7\\%" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": ", significantly outperforming other compared methods by margins of " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "1.8\\mathrm{mm}" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": " in MPJPE and " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": " in AUC. However, in terms of the Percentage of Correct Keypoints (PCK), RePOSE reaches " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "98.3\\%" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": ", which is slightly lower by " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": " compared" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 194, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 194, + 91, + 447, + 102 + ], + "type": "text", + "content": "3D Human Pose Estimation via Depth Relational Consistency" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 113, + 203, + 182 + ], + "blocks": [ + { + "bbox": [ + 135, + 113, + 203, + 182 + ], + "lines": [ + { + "bbox": [ + 135, + 113, + 203, + 182 + ], + "spans": [ + { + "bbox": [ + 135, + 113, + 203, + 182 + ], + "type": "image", + "image_path": "5f3f5455109d303d89d43c8c7df680faff1e13a8c0d177ee89254d5af257149f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 136, + 182, + 203, + 248 + ], + "blocks": [ + { + "bbox": [ + 136, + 182, + 203, + 248 + ], + "lines": [ + { + "bbox": [ + 136, + 182, + 203, + 248 + ], + "spans": [ + { + "bbox": [ + 136, + 182, + 203, + 248 + ], + "type": "image", + "image_path": "3a24fc725daba22553b8a94546af8cffd494a94e06b272a28ba4d0f88d1b4a69.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 250, + 186, + 259 + ], + "lines": [ + { + "bbox": [ + 153, + 250, + 186, + 259 + ], + "spans": [ + { + "bbox": [ + 153, + 250, + 186, + 259 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 206, + 116, + 269, + 180 + ], + "blocks": [ + { + "bbox": [ + 206, + 116, + 269, + 180 + ], + "lines": [ + { + "bbox": [ + 206, + 116, + 269, + 180 + ], + "spans": [ + { + "bbox": [ + 206, + 116, + 269, + 180 + ], + "type": "image", + "image_path": "02100a0d79ab8aa956b1a9eab01929261c359d6977c1ef2268f446ca8e3b41e6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 207, + 182, + 269, + 247 + ], + "blocks": [ + { + "bbox": [ + 207, + 182, + 269, + 247 + ], + "lines": [ + { + "bbox": [ + 207, + 182, + 269, + 247 + ], + "spans": [ + { + "bbox": [ + 207, + 182, + 269, + 247 + ], + "type": "image", + "image_path": "39863b86a363cbefee185735abc617c38da40bb3e1932ca6b964aac1db9145b0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 224, + 250, + 252, + 259 + ], + "lines": [ + { + "bbox": [ + 224, + 250, + 252, + 259 + ], + "spans": [ + { + "bbox": [ + 224, + 250, + 252, + 259 + ], + "type": "text", + "content": "(b) GT" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 276, + 116, + 338, + 181 + ], + "blocks": [ + { + "bbox": [ + 276, + 116, + 338, + 181 + ], + "lines": [ + { + "bbox": [ + 276, + 116, + 338, + 181 + ], + "spans": [ + { + "bbox": [ + 276, + 116, + 338, + 181 + ], + "type": "image", + "image_path": "c7970a2ed7a0fcd5e858ff79e1105a51df62fcff18c9efe6c2d58b3703e58f51.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 276, + 182, + 339, + 247 + ], + "blocks": [ + { + "bbox": [ + 276, + 182, + 339, + 247 + ], + "lines": [ + { + "bbox": [ + 276, + 182, + 339, + 247 + ], + "spans": [ + { + "bbox": [ + 276, + 182, + 339, + 247 + ], + "type": "image", + "image_path": "b843a7cab61e261629d0288b08e759075274dc00f6723826f7c7a523ca973d77.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 288, + 250, + 325, + 259 + ], + "lines": [ + { + "bbox": [ + 288, + 250, + 325, + 259 + ], + "spans": [ + { + "bbox": [ + 288, + 250, + 325, + 259 + ], + "type": "text", + "content": "(c) D3DP" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 268, + 480, + 301 + ], + "lines": [ + { + "bbox": [ + 131, + 268, + 480, + 301 + ], + "spans": [ + { + "bbox": [ + 131, + 268, + 480, + 301 + ], + "type": "text", + "content": "Fig. 5: Qualitative comparison on MPI-INF-3DHP. Our method is compared with D3DP [28] and STCFoer [30]. The areas with incorrect results are pointed out with red arrows and circles to exemplify the importance of relational depth consistency." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 345, + 116, + 406, + 181 + ], + "blocks": [ + { + "bbox": [ + 345, + 116, + 406, + 181 + ], + "lines": [ + { + "bbox": [ + 345, + 116, + 406, + 181 + ], + "spans": [ + { + "bbox": [ + 345, + 116, + 406, + 181 + ], + "type": "image", + "image_path": "d5141b30a094ea014d444f13a2f8239c7714756d77904571db0d4c8dc040b4fe.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 345, + 182, + 406, + 247 + ], + "blocks": [ + { + "bbox": [ + 345, + 182, + 406, + 247 + ], + "lines": [ + { + "bbox": [ + 345, + 182, + 406, + 247 + ], + "spans": [ + { + "bbox": [ + 345, + 182, + 406, + 247 + ], + "type": "image", + "image_path": "8912d8606a95cf2b454ce29f8e737086dc0aae73f95dbb8db9f185801f607281.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 250, + 405, + 259 + ], + "lines": [ + { + "bbox": [ + 347, + 250, + 405, + 259 + ], + "spans": [ + { + "bbox": [ + 347, + 250, + 405, + 259 + ], + "type": "text", + "content": "(d) STCFormer" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 413, + 117, + 476, + 181 + ], + "blocks": [ + { + "bbox": [ + 413, + 117, + 476, + 181 + ], + "lines": [ + { + "bbox": [ + 413, + 117, + 476, + 181 + ], + "spans": [ + { + "bbox": [ + 413, + 117, + 476, + 181 + ], + "type": "image", + "image_path": "34d63eba5b5b4b6f98cd66b1215985b913750cb602de1205886d5070af862133.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 413, + 182, + 476, + 247 + ], + "blocks": [ + { + "bbox": [ + 413, + 182, + 476, + 247 + ], + "lines": [ + { + "bbox": [ + 413, + 182, + 476, + 247 + ], + "spans": [ + { + "bbox": [ + 413, + 182, + 476, + 247 + ], + "type": "image", + "image_path": "6b2543511f8dd399d8c942ad6c66f1308fe23b25a082bf1cb32f4bbdeeb98e5f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 250, + 460, + 259 + ], + "lines": [ + { + "bbox": [ + 429, + 250, + 460, + 259 + ], + "spans": [ + { + "bbox": [ + 429, + 250, + 460, + 259 + ], + "type": "text", + "content": "(e) Ours" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 326, + 480, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 480, + 361 + ], + "type": "text", + "content": "to the best performance achieved by existing methods, but we excel in the best balance of all metrics. For a more detailed visual assessment, refer to the qualitative comparisons illustrated in Fig. 5." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 131, + 380, + 238, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 380, + 238, + 393 + ], + "spans": [ + { + "bbox": [ + 131, + 380, + 238, + 393 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 402, + 301, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 301, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 301, + 581 + ], + "type": "text", + "content": "To evaluate the distinct contributions of individual components, we incorporated spatial and temporal relational depth consistency loss functions separately into the baseline network. As shown in Tab. 4, it is clear that without these spatial and temporal depth constraints, the baseline model experiences degradation in both the MPJPE and PA-MPJPE metrics. However, the inclusion of either the spatial or temporal loss functions results in a significant enhancement in the model's performance. This observation underscores the effectiveness of these two" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 582, + 452, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 452, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 452, + 593 + ], + "type": "text", + "content": "loss functions in improving the overall quality of the model's predictions." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": "Fig. 6 provides a comparative visualization of the performance of the RepoSE with and without the incorporation of two relational depth consistency losses. It is evident from the figure that the omission of these losses unfavorably affects the model's capacity to accurately estimate occluded body parts, as exemplified by the red circles. This inadequacy suggests that the model, without these losses, cannot effectively identify the relative depth of body parts during" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 414, + 482, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 482, + 502 + ], + "type": "text", + "content": "Table 4: Ablation Study on Individual Loss Functions. The study reveals that while both spatial and temporal loss functions individually contribute to improved pose estimation, their combined application yields the best results. Performance metrics have been evaluated on the Human3.6M dataset to substantiate this finding." + } + ] + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 308, + 521, + 479, + 570 + ], + "blocks": [ + { + "bbox": [ + 308, + 521, + 479, + 570 + ], + "lines": [ + { + "bbox": [ + 308, + 521, + 479, + 570 + ], + "spans": [ + { + "bbox": [ + 308, + 521, + 479, + 570 + ], + "type": "table", + "html": "
MPJPE Spatial TemporalMPJPE↓PA-MPJPE↓
✓--37.932.5
✓✓-36.931.4
✓-✓37.231.8
✓✓✓36.530.7
", + "image_path": "0f9f08c592643690695593f1d759e80560af84a5a020c45314a2e8115c0cac75.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Z. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 228, + 191 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 228, + 191 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 228, + 191 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 228, + 191 + ], + "type": "image", + "image_path": "0b32b0329018eb45d3a07e32ac68d60a52ef76cc6bcad2f70a6e9eb4d488ef1a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 149, + 192, + 228, + 270 + ], + "blocks": [ + { + "bbox": [ + 149, + 192, + 228, + 270 + ], + "lines": [ + { + "bbox": [ + 149, + 192, + 228, + 270 + ], + "spans": [ + { + "bbox": [ + 149, + 192, + 228, + 270 + ], + "type": "image", + "image_path": "8e1bd81571567d455c0c0b1e3ad64ddd0f3142de0511a4fcc828b8c0c3136744.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 270, + 206, + 280 + ], + "lines": [ + { + "bbox": [ + 171, + 270, + 206, + 280 + ], + "spans": [ + { + "bbox": [ + 171, + 270, + 206, + 280 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 289, + 482, + 334 + ], + "lines": [ + { + "bbox": [ + 131, + 289, + 482, + 334 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 482, + 334 + ], + "type": "text", + "content": "Fig. 6: Visualized Results of Ablation Study. The second column displays poses generated by a model trained without the spatio-temporal relational depth consistency loss, contrasting with the third column, which shows results from the model including this loss. Joints with less accurate estimations are highlighted in red circles for clarity." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 235, + 118, + 306, + 190 + ], + "blocks": [ + { + "bbox": [ + 235, + 118, + 306, + 190 + ], + "lines": [ + { + "bbox": [ + 235, + 118, + 306, + 190 + ], + "spans": [ + { + "bbox": [ + 235, + 118, + 306, + 190 + ], + "type": "image", + "image_path": "6578cb4f6549cacd34706e677dbc302c5522d31096f67e70da93888d6e198cd1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 235, + 194, + 306, + 269 + ], + "blocks": [ + { + "bbox": [ + 235, + 194, + 306, + 269 + ], + "lines": [ + { + "bbox": [ + 235, + 194, + 306, + 269 + ], + "spans": [ + { + "bbox": [ + 235, + 194, + 306, + 269 + ], + "type": "image", + "image_path": "e8fbbfd8791ed3e1bb247458d4f2e5de897c0bc99dc899c2e0a50b4762ff0cea.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 270, + 301, + 280 + ], + "lines": [ + { + "bbox": [ + 235, + 270, + 301, + 280 + ], + "spans": [ + { + "bbox": [ + 235, + 270, + 301, + 280 + ], + "type": "text", + "content": "(b) w/o our losses" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 313, + 118, + 385, + 191 + ], + "blocks": [ + { + "bbox": [ + 313, + 118, + 385, + 191 + ], + "lines": [ + { + "bbox": [ + 313, + 118, + 385, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 118, + 385, + 191 + ], + "type": "image", + "image_path": "2af1bcb8aa2e34e1f78294d691e07d32f2aa12beef27932ed505a995f0d6d30f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 313, + 194, + 384, + 267 + ], + "blocks": [ + { + "bbox": [ + 313, + 194, + 384, + 267 + ], + "lines": [ + { + "bbox": [ + 313, + 194, + 384, + 267 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 384, + 267 + ], + "type": "image", + "image_path": "e3cd59544a8277d065ee6fd435d4ac60cebcfa4354de088be5bba109ba1b80b9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 270, + 370, + 280 + ], + "lines": [ + { + "bbox": [ + 323, + 270, + 370, + 280 + ], + "spans": [ + { + "bbox": [ + 323, + 270, + 370, + 280 + ], + "type": "text", + "content": "(c) RePOSE" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 392, + 118, + 463, + 191 + ], + "blocks": [ + { + "bbox": [ + 392, + 118, + 463, + 191 + ], + "lines": [ + { + "bbox": [ + 392, + 118, + 463, + 191 + ], + "spans": [ + { + "bbox": [ + 392, + 118, + 463, + 191 + ], + "type": "image", + "image_path": "880e335368efc8589ef1c7eeba1669d2dcd5f66f8f706455b8f2eb052bed2a16.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 392, + 194, + 463, + 269 + ], + "blocks": [ + { + "bbox": [ + 392, + 194, + 463, + 269 + ], + "lines": [ + { + "bbox": [ + 392, + 194, + 463, + 269 + ], + "spans": [ + { + "bbox": [ + 392, + 194, + 463, + 269 + ], + "type": "image", + "image_path": "693e73f2ca705f1f3da685dca86cee04166f884acbd0aa3e87f91ee7857eda80.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 393, + 270, + 459, + 280 + ], + "lines": [ + { + "bbox": [ + 393, + 270, + 459, + 280 + ], + "spans": [ + { + "bbox": [ + 393, + 270, + 459, + 280 + ], + "type": "text", + "content": "(d) Ground Truth" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 144, + 399, + 471, + 462 + ], + "blocks": [ + { + "bbox": [ + 131, + 345, + 480, + 379 + ], + "lines": [ + { + "bbox": [ + 131, + 345, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 131, + 345, + 480, + 379 + ], + "type": "text", + "content": "Table 5: Generalization evaluation on Human3.6M with our loss functions. With the integration of our loss functions, all methods exhibit improved performance. " + }, + { + "bbox": [ + 131, + 345, + 480, + 379 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 131, + 345, + 480, + 379 + ], + "type": "text", + "content": " indicates the difference between 'w/ST' and 'w/o'." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 144, + 399, + 471, + 462 + ], + "lines": [ + { + "bbox": [ + 144, + 399, + 471, + 462 + ], + "spans": [ + { + "bbox": [ + 144, + 399, + 471, + 462 + ], + "type": "table", + "html": "
MethodTMPJPE↓PA-MPJPE↓
w/Sw/Tw/STw/oΔw/Sw/Tw/STw/oΔ
P-STMO-S [27]2745.345.544.746.1-1.436.436.635.636.8-1.2
STCFformer [30]2743.843.843.344.1-0.834.134.033.734.8-1.1
MotionBERT [46]24337.037.236.737.5-0.831.431.730.831.8-1.0
", + "image_path": "068457d406f621e08b77dd027f1bc79520c42b10d846c80c557686a81be88f24.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 494, + 482, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 482, + 565 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 482, + 565 + ], + "type": "text", + "content": "the training stage, leading to incorrect results. On the contrary, the integration of our proposed losses into the training phase significantly enhances the model's ability to resolve occlusions, thereby improving its estimation accuracy. This improvement is crucial for applications where precise depth perception is necessary, and it emphasizes the value of our contributions to the field of pose estimation in occlusion situations." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 585, + 292, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 585, + 292, + 596 + ], + "spans": [ + { + "bbox": [ + 132, + 585, + 292, + 596 + ], + "type": "text", + "content": "4.5 Generalization Evaluation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "To validate the versatility and generalization capacity of our proposed loss functions, we extend their application to similar spatio-temporal transformer-based frameworks, including MotionBERT [46], STCFoer [30], and P-STMO [27]. We directly use their official pre-trained models with default setups and fine-tune them with our loss functions." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 194, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 194, + 91, + 447, + 102 + ], + "type": "text", + "content": "3D Human Pose Estimation via Depth Relational Consistency" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "We detail the generalizability of our individual loss functions in Table 5. Overall, integrating both losses ('w/ST') leads to superior performance metrics compared to the original model configurations without our adjustments. Significantly, even when applied separately, each loss function ('w/S' or 'w/T') yields improvements across all evaluated methods, reinforcing their broad applicability. It is important to note that all reported results were achieved without specific tuning of the hyperparameters " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\lambda_{S}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\lambda_{T}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": ". Therefore, there is an opportunity for further enhancements if these parameters were to be systematically optimized." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 236, + 312, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 312, + 249 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 312, + 249 + ], + "type": "text", + "content": "5 Conclusion and Discussion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 268, + 482, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 482, + 435 + ], + "type": "text", + "content": "In this paper, we introduce RePOSE, an approach designed to enhance 3D pose estimation accuracy in video sequences, specifically under conditions where body parts are obscured or occluded. Diverging from conventional methods that predominantly depend on depth information, which can falter when body parts are not visible, RePOSE adopts an innovative strategy centered on the relative positioning of body segments. This methodology not only exhibits superior efficacy in managing occlusions but also boasts a straightforward implementation, requiring minimal code adjustments. Our comprehensive experiments validate that RePOSE outperforms existing state-of-the-art techniques, particularly in scenarios marked by obstructions. These outcomes underscore RePOSE's efficacy and its promising potential in advancing 3D pose estimation, especially in practical settings where occlusions and partial visibilities are common. Additionally, the simplicity of its integration renders RePOSE a viable and effective solution for researchers aiming to refine pose estimation precision." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 437, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 533 + ], + "type": "text", + "content": "Limitations. Despite its significant advantages in occlusion scenarios within 3D human pose estimation, RePOSE encounters limitations, chiefly its dependence on the accuracy of preliminary 2D pose data. Discrepancies in this foundational data can adversely affect subsequent 3D estimations, a challenge inherent in 2D-to-3D conversion techniques. Moreover, although tailored for occlusions, RePOSE's performance may diminish in complex conditions, such as atypical body postures or interactions with objects, and in extremely low-light environments where depth cues are compromised." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "content": "Future Work. The efficacy of our newly proposed loss functions has been established across various spatio-temporal transformer architectures, including MotionBERT [46] and STCFoer [30]. Our next phase of research will explore the extension of these functions to a wider spectrum of models, including those based on GNNs and CNNs. Additionally, we aim to evolve the hyperparameters " + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{S}" + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{T}" + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "content": " from fixed values based on empirical selection to adaptive, learnable parameters optimized through the model's interaction with input data, thereby refining the model's depth dimension understanding in 3D human pose estimation. Future enhancements will focus on augmenting RePOSE's versatility across different input data forms and broadening its efficacy in a more extensive array of challenging conditions." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Z. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 201 + ], + "type": "text", + "content": "This project is supported by the Guangdong Basic and Applied Basic Research Foundation under project No. 2024A1515011995, Guangdong Natural Science Funds for Distinguished Young Scholar under Grant 2023B1515020097, and National Research Foundation Singapore under the AI Singapore Programme under Grant AISG3-GV-2023-011." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 220, + 197, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 220, + 197, + 232 + ], + "spans": [ + { + "bbox": [ + 133, + 220, + 197, + 232 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 245, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 245, + 481, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 481, + 277 + ], + "type": "text", + "content": "1. Cai, Y., Ge, L., Liu, J., Cai, J., Cham, T.J., Yuan, J., Thalmann, N.M.: Exploiting spatial-temporal relationships for 3d pose estimation via graph convolutional networks. In: ICCV. pp. 2272-2281 (2019) 3, 9" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 280, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 280, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 138, + 280, + 481, + 311 + ], + "type": "text", + "content": "2. Chen, T., Fang, C., Shen, X., Zhu, Y., Chen, Z., Luo, J.: Anatomy-aware 3d human pose estimation with bone-based pose decomposition. IEEE TCSVT 32(1), 198-209 (2021) 2, 3, 8, 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 312, + 481, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 312, + 481, + 334 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 481, + 334 + ], + "type": "text", + "content": "3. Chen, Y., Wang, Z., Peng, Y., Zhang, Z., Yu, G., Sun, J.: Cascaded pyramid network for multi-person pose estimation. In: CVPR. pp. 7103-7112 (2018) 3, 10" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 335, + 481, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 335, + 481, + 366 + ], + "spans": [ + { + "bbox": [ + 138, + 335, + 481, + 366 + ], + "type": "text", + "content": "4. Chen, Y., Tu, Z., Ge, L., Zhang, D., Chen, R., Yuan, J.: So-handnet: Self-organizing network for 3d hand pose estimation with semi-supervised learning. In: ICCV. pp. 6961-6970 (2019) 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 368, + 481, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 368, + 481, + 389 + ], + "spans": [ + { + "bbox": [ + 138, + 368, + 481, + 389 + ], + "type": "text", + "content": "5. Ci, H., Wang, C., Ma, X., Wang, Y.: Optimizing network structure for 3d human pose estimation. In: ICCV. pp. 2262-2271 (2019) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 390, + 481, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 390, + 481, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 390, + 481, + 411 + ], + "type": "text", + "content": "6. Ci, H., Wu, M., Zhu, W., Ma, X., Dong, H., Zhong, F., Wang, Y.: Gfpose: Learning 3d human pose prior with gradient fields. In: CVPR. pp. 4800-4810 (2023) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 412, + 481, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 412, + 481, + 433 + ], + "spans": [ + { + "bbox": [ + 138, + 412, + 481, + 433 + ], + "type": "text", + "content": "7. Fang, H.S., Xie, S., Tai, Y.W., Lu, C.: Rumpe: Regional multi-person pose estimation. In: ICCV. pp. 2334-2343 (2017) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 434, + 481, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 481, + 455 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 481, + 455 + ], + "type": "text", + "content": "8. Gong, J., Fan, Z., Ke, Q., Rahmani, H., Liu, J.: Meta agent teaming active learning for pose estimation. In: CVPR. pp. 11079-11089 (2022) 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 456, + 481, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 481, + 477 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 481, + 477 + ], + "type": "text", + "content": "9. Hossain, M.R.I., Little, J.J.: Exploiting temporal information for 3d human pose estimation. In: ECCV. pp. 68-84 (2018) 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 479, + 481, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 481, + 499 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 481, + 499 + ], + "type": "text", + "content": "0. Huang, S., Wang, W., He, S., Lau, R.W.: Egocentric temporal action proposals. IEEE TIP 27(2), 764-777 (2017) 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 501, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 481, + 533 + ], + "type": "text", + "content": "1. Ionescu, C., Papava, D., Olaru, V., Sminchisescu, C.: Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE TPAMI 36(7), 1325-1339 (2013) 3, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 534, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 481, + 555 + ], + "type": "text", + "content": "2. Jiang, Y., Zhou, Y., Liang, Y., Liu, W., Jiao, J., Quan, Y., He, S.: Diffuse3d: Wide-angle 3d photography via bilateral diffusion. In: ICCV. pp. 8998-9008 (2023) 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "type": "text", + "content": "3. Li, W., Liu, H., Ding, R., Liu, M., Wang, P., Yang, W.: Exploiting temporal contexts with strided transformer for 3d human pose estimation. IEEE Transactions on Multimedia 25, 1282-1293 (2022) 2, 3, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 589, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 481, + 620 + ], + "type": "text", + "content": "4. Li, W., Liu, H., Tang, H., Wang, P., Van Gool, L.: Mhformer: Multi-hypothesis transformer for 3d human pose estimation. In: CVPR. pp. 13147-13156 (2022) 2, 3, 5, 10, 11" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 622, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 481, + 643 + ], + "type": "text", + "content": "5. Lin, K., Wang, L., Liu, Z.: End-to-end human pose and mesh reconstruction with transformers. In: CVPR. pp. 1954-1963 (2021) 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "text", + "content": "6. Lin, M., Lin, L., Liang, X., Wang, K., Cheng, H.: Recurrent 3d pose sequence machines. In: CVPR. pp. 810-819 (2017) 3" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 194, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 194, + 91, + 447, + 102 + ], + "type": "text", + "content": "3D Human Pose Estimation via Depth Relational Consistency" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "17. Liu, R., Shen, J., Wang, H., Chen, C., Cheung, S.c., Asari, V.: Attention mechanism exploits temporal contexts: Real-time 3d human pose reconstruction. In: CVPR. pp. 5064-5073 (2020) 2, 3, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 170 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 170 + ], + "type": "text", + "content": "18. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019) 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 480, + 193 + ], + "type": "text", + "content": "19. Martinez, J., Hossain, R., Romero, J., Little, J.J.: A simple yet effective baseline for 3d human pose estimation. In: ICCV. pp. 2640-2649 (2017) 3, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 236 + ], + "type": "text", + "content": "20. Mehta, D., Rhodin, H., Casas, D., Fua, P., Sotnychenko, O., Xu, W., Theobalt, C.: Monocular 3d human pose estimation in the wild using improved cnn supervision. In: 2017 international conference on 3D vision (3DV). pp. 506-516. IEEE (2017) 3, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": "21. Mehta, D., Sridhar, S., Sotnychenko, O., Rhodin, H., Shafiei, M., Seidel, H.P., Xu, W., Casas, D., Theobalt, C.: Vnct: Real-time 3d human pose estimation with a single rgb camera. ACM TOG 36(4), 1-14 (2017) 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "text", + "content": "22. Moon, G., Lee, K.M.: I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In: ECCV. pp. 752-768. Springer (2020) 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "23. Newell, A., Yang, K., Deng, J.: Stacked hourglass networks for human pose estimation. In: ECCV. pp. 483-499. Springer (2016) 9, 10" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "text", + "content": "24. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch. In: NeurIPS (2017) 9" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "text", + "content": "25. Pavlakos, G., Zhou, X., Daniilidis, K.: Ordinal depth supervision for 3d human pose estimation. In: CVPR. pp. 7307-7316 (2018) 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 480, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 412 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 412 + ], + "type": "text", + "content": "26. Pavlo, D., Feichtenhofer, C., Grangier, D., Auli, M.: 3d human pose estimation in video with temporal convolutions and semi-supervised training. In: CVPR. pp. 7753-7762 (2019) 3, 4, 8, 10, 11" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "text", + "content": "27. Shan, W., Liu, Z., Zhang, X., Wang, S., Ma, S., Gao, W.: P-stmo: Pre-trained spatial temporal many-to-one model for 3d human pose estimation. In: ECCV. pp. 461-478. Springer (2022) 2, 3, 5, 6, 10, 11, 13" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "type": "text", + "content": "28. Shan, W., Liu, Z., Zhang, X., Wang, Z., Han, K., Wang, S., Ma, S., Gao, W.: Diffusion-based 3d human pose estimation with multi-hypothesis aggregation. In: ICCV. pp. 14761-14771 (October 2023) 10, 11, 12" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "text", + "content": "29. Soroush Mehraban, Vida Adeli, B.T.: Motionagformer: Enhancing 3d human pose estimation with a transformer-gcnformer network. In: WACV (2024) 4, 9, 10, 11" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "text", + "content": "30. Tang, Z., Qiu, Z., Hao, Y., Hong, R., Yao, T.: 3d human pose estimation with spatio-temporal criss-cross attention. In: CVPR. pp. 4790-4799 (2023) 10, 11, 12, 13, 14" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "text", + "content": "31. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017) 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 556, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 480, + 577 + ], + "type": "text", + "content": "32. Wang, J., Yan, S., Xiong, Y., Lin, D.: Motion guided 3d pose estimation from videos. In: ECCV. pp. 764-780. Springer (2020) 3, 9" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 578, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 609 + ], + "type": "text", + "content": "33. Wang, J., Sun, K., Cheng, T., Jiang, B., Deng, C., Zhao, Y., Liu, D., Mu, Y., Tan, M., Wang, X., et al.: Deep high-resolution representation learning for visual recognition. IEEE TPAMI 43(10), 3349-3364 (2020) 3, 10" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "text", + "content": "34. Xu, Y., Han, C., Qin, J., Xu, X., Han, G., He, S.: Transductive zero-shot action recognition via visually connected graph convolutional networks. IEEE TNNLS 32(8), 3761-3769 (2020) 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "35. Yan, S., Xiong, Y., Lin, D.: Spatial temporal graph convolutional networks for skeleton-based action recognition. In: AAAI. vol. 32 (2018) 1" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Z. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 456 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "36. Yang, X., Xu, K., Chen, S., He, S., Yin, B.Y., Lau, R.: Active matting. NeurIPS 31 (2018) 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "type": "text", + "content": "37. Ye, M., Li, H., Du, B., Shen, J., Shao, L., Hoi, S.C.: Collaborative refining for person re-identification with label noise. IEEE TIP 31, 379-391 (2021) 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 482, + 194 + ], + "type": "text", + "content": "38. Yoon, J.S., Liu, L., Golyanik, V., Sarkar, K., Park, H.S., Theobalt, C.: Pose-guided human animation from a single image in the wild. In: CVPR. pp. 15039-15048 (2021) 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "type": "text", + "content": "39. Yu, B.X., Zhang, Z., Liu, Y., Zhong, S.h., Liu, Y., Chen, C.W.: Gla-gcn: Global-local adaptive graph convolutional network for 3d human pose estimation from monocular video. In: ICCV. pp. 8818-8829 (2023) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 226, + 482, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 258 + ], + "type": "text", + "content": "40. Zhang, C., Yang, T., Weng, J., Cao, M., Wang, J., Zou, Y.: Unsupervised pretraining for temporal action localization tasks. In: CVPR. pp. 14031-14041 (2022) 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "text", + "content": "41. Zhang, J., Ye, G., Tu, Z., Qin, Y., Qin, Q., Zhang, J., Liu, J.: A spatial attentive and temporal dilated (satd) gcn for skeleton-based action recognition. CAAI Transactions on Intelligence Technology 7(1), 46-55 (2022) 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 293, + 482, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 325 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 325 + ], + "type": "text", + "content": "42. Zhang, J., Tu, Z., Yang, J., Chen, Y., Yuan, J.: Mixste: Seq2seq mixed spatiotemporal encoder for 3d human pose estimation in video. In: CVPR. pp. 13232-13242 (2022) 2, 3, 4, 5, 10, 11" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 325, + 482, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 482, + 357 + ], + "type": "text", + "content": "43. Zhao, L., Peng, X., Tian, Y., Kapadia, M., Metaxas, D.N.: Semantic graph convolutional networks for 3d human pose regression. In: CVPR. pp. 3425-3435 (2019) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 357, + 482, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 390 + ], + "type": "text", + "content": "44. Zhao, Q., Zheng, C., Liu, M., Wang, P., Chen, C.: Poseformerv2: Exploring frequency domain for efficient and robust 3d human pose estimation. In: CVPR. pp. 8877-8886 (2023) 2, 3, 5, 9, 10, 11" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 390, + 482, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 423 + ], + "type": "text", + "content": "45. Zheng, C., Zhu, S., Mendieta, M., Yang, T., Chen, C., Ding, Z.: 3d human pose estimation with spatial and temporal transformers. In: ICCV. pp. 11656-11665 (2021) 2, 3, 5, 6, 8, 10, 11" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 423, + 482, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 456 + ], + "type": "text", + "content": "46. Zhu, W., Ma, X., Liu, Z., Liu, L., Wu, W., Wang, Y.: Motionbert: A unified perspective on learning human motion representations. In: ICCV. pp. 15085-15099 (2023) 2, 3, 4, 5, 8, 9, 10, 11, 13, 14" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 194, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 194, + 91, + 447, + 102 + ], + "type": "text", + "content": "3D Human Pose Estimation via Depth Relational Consistency" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/ReSyncer_ Rewiring Style-based Generator for Unified Audio-Visually Synced Facial Performer/6fe8e31e-41da-4533-8af1-187e2be9d843_content_list.json b/2024/ReSyncer_ Rewiring Style-based Generator for Unified Audio-Visually Synced Facial Performer/6fe8e31e-41da-4533-8af1-187e2be9d843_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1ad8159cc0628f88af5d0c2cb19e65904f11f3c2 --- /dev/null +++ b/2024/ReSyncer_ Rewiring Style-based Generator for Unified Audio-Visually Synced Facial Performer/6fe8e31e-41da-4533-8af1-187e2be9d843_content_list.json @@ -0,0 +1,1738 @@ +[ + { + "type": "text", + "text": "ReSyncer: Rewiring Style-based Generator for Unified Audio-Visually Synced Facial Performer", + "text_level": 1, + "bbox": [ + 227, + 141, + 776, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiazhi Guan $^{1,2*}$ , Zhiliang Xu $^{2*}$ , Hang Zhou $^{2\\dagger}$ , Kaisiyuan Wang $^{2}$ , Shengyi He $^{2}$ , Zhanwang Zhang $^{2}$ , Borong Liang $^{2}$ , Haocheng Feng $^{2}$ , Errui Ding $^{2}$ , Jingtuo Liu $^{2}$ , Jingdong Wang $^{2}$ , Youjian Zhao $^{1,3\\dagger}$ , and Ziwei Liu $^{4}$", + "bbox": [ + 215, + 210, + 785, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ BNRist, DCST, Tsinghua University $^{2}$ Baidu Inc. \n $^{3}$ Zhongguancun Laboratory $^{4}$ S-Lab, Nanyang Technological University guanjz20@mails.tsinghua.edu.cn, zhouhang09@baidu.com", + "bbox": [ + 256, + 268, + 743, + 313 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7006e089d5e2a453a0432cdf8dda770d32efa20793899df00e4c54f055d1e327.jpg", + "image_caption": [ + "Fig. 1: Lip-Synced/Speaking-Style-Transferred/Face-Swapping Results by ReSyncer. Our method not only produces high-fidelity lip-synced video according to audio but can further transfer the speaking style and identity of any target person." + ], + "image_footnote": [], + "bbox": [ + 223, + 337, + 777, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Lip-syncing videos with given audio is the foundation for various applications including the creation of virtual presenters or performers. While recent studies explore high-fidelity lip-sync with different techniques, their task-orientated models either require long-term videos for clip-specific training or retain visible artifacts. In this paper, we propose a unified and effective framework ReSyncer, that synchronizes generalized audio-visual facial information. The key design is revisiting and rewiring the Style-based generator to efficiently adopt 3D facial dynamics predicted by a principled style-injected Transformer. By simply re-configuring the information insertion mechanisms within the noise and style space, our framework fuses motion and appearance with unified training. Extensive experiments demonstrate that ReSyncer not", + "bbox": [ + 259, + 635, + 743, + 801 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "\\*Equal Contribution. $\\dagger$ Corresponding Authors.", + "bbox": [ + 238, + 810, + 408, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "only produces high-fidelity lip-synced videos according to audio, but also supports multiple appealing properties that are suitable for creating virtual presenters and performers, including fast personalized fine-tuning, video-driven lip-syncing, the transfer of speaking styles, and even face swapping.", + "bbox": [ + 259, + 147, + 743, + 218 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 238, + 375, + 253 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generating lip-synced facial videos with speech audio is a foundational technique for various applications including the dubbing of movies and the creation of virtual presenters or performers in the form of realistic humans. The needs in virtual performers' creation can hardly be comprehensively met.", + "bbox": [ + 212, + 266, + 785, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While great attention has been paid to creating dynamic lip-synced videos from a few photos [8, 24, 63, 64, 78, 80-82], previous results are still far from realistic. To generate life-like videos, researchers have made explorations [41] with GANs [10,55], NeRF [23,33,47,53], and Diffusion models [48] to edit existing videos. Despite their efforts, most methods rely on a relatively long-term clip of video for training, making them less flexible and efficient. A few methods [10,22, 41,51] enable generalized local editing without training. However, they still tend to create visible flaws on high-quality videos [41,51] Particularly, StyleSync [22] supports not only arbitrary video modeling but can adapt to the speaking style of a target person with short-clip fine-tuning, which balances the generation quality and efficiency. Nevertheless, their results can still be easily affected by mouth movements on their conditional reference frames, leading to unstable results.", + "bbox": [ + 212, + 327, + 785, + 507 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The previous cases reveal the limitations of using low-dimensional audio information to directly modify high-dimensional visual data. On the other hand, intermediate representations like 3D facial models are widely used, but their potential as spatial guidance for efficient, generalized editing remains underexplored. Using only 3DMM [3] coefficients [15,16,19,29,37,55] also suffers from the cross-domain information injection problem. While textured rendering offers a promising alternative [6,55,67], its implementation complexity and susceptibility to errors during 3D modeling or texture fitting can lead to degraded results.", + "bbox": [ + 212, + 508, + 785, + 642 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we revisit and re-wire the powerful Style-based generator [26, 27] to generate faithful lip-synced videos and support virtual performer creation in a concise yet effective framework called ReSyncer. With unified model training, various appealing properties including generalized and personalized lip-sync, speaking style transfer, live streaming, and face swapping can be achieved within one model. Our key is to show how an easy reconfiguration empowers Style-based generator with various powerful audio-visual synchronization abilities by involving 3D facial dynamics. Specifically, we use only roughly fitted 3D facial meshes, instead of coefficients to bridge the audio and image domains. A principled Style-injected Lip-Sync Transformer (Style-SyncFormer) learns stylized point cloud displacements given template meshes and driving audios.", + "bbox": [ + 212, + 643, + 785, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We then perform easy but crucial modifications to Style-based generator with a simple encoder used in previous work [22, 73, 74]. Concretely, while lip-sync", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Guan et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "studies [10,22,41,51,61] mask out the mouth area of the target frame, we overlay reconstructed mesh onto it as strong spatial guidance. Then we compliment the textural and identity information with an attached reference frame, which serves as the source of both the noise and style-related space. With generalized training on massive data, the powerful generator focuses on recovering facial details from the roughly predicted lip-synced white-colored meshes, leading to high-fidelity and stable lip-sync with higher quality.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Additionally, our framework design goes beyond existing solutions by uniformly addressing a broader range of needs in virtual performer creation. These needs include, for instance, preserving personal styles for specific users or changing face identities for privacy protection and diverse character generation. Particularly, our method directly supports few-shot few-step personalized fine-tuning. Moreover, our framework supports *syncing* (swapping) different identities on the same target video with additional training strategies. High-fidelity face-swapping and lip-syncing can hence be simultaneously satisfied.", + "bbox": [ + 212, + 252, + 787, + 373 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows: 1) We propose the ReSyncer framework, which shows the powerfulness of Style-based generator on syncing audio-visual facial information by involving 3D facial meshes with a simple reconfiguration. 2) We propose the Style-SyncFormer which learns stylized 3D facial dynamics with simple Transformer blocks, enabling generalized 3D facial animation. 3) Extensive experiments show that our framework not only enables lip-syncing of higher stability and quality, but also supports various intriguing properties that are essential for creating virtual performers, including fast personalized fine-tuning, video-driven lip-syncing, the transfer of speaking styles, and even face swapping. To the best of our knowledge, we are one of the first methods that support state-of-the-art lip-syncing and face-swapping with one unified model.", + "bbox": [ + 212, + 375, + 795, + 555 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 580, + 388, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Audio-Driven Facial Animation", + "text_level": 1, + "bbox": [ + 214, + 614, + 522, + 630 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Talking Head Generation. A great number of studies focus on predicting the whole human head movements according to audios [7, 8, 23-25, 32, 36, 37, 42, 52, 59, 62, 63, 77, 78, 81]. More recently, one-shot-based methods adopt representations from 3D, videos, and even texts for more controllable results. While other approaches [23, 33, 47, 53, 75, 76] attempt to integrate speech audios into the neural radiance field for personalized audio-to-3D modeling. However, these methods usually fail to achieve stable lip-sync results and show terrible robustness performance when using audio in the wild. Moreover, the form of talking head generation cannot be applied to edit arbitrary video.", + "bbox": [ + 212, + 642, + 787, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lip-Syncing-based Facial Editing. Apart from generating the entire head, certain studies [10, 22, 40, 41, 49, 51, 55, 68] choose to directly edit the mouth shape according to the input audios. Specifically, [41] relies on a well-trained lip-sync discriminator to produce highly synchronized animations. [51] introduces", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "ReSyncer", + "bbox": [ + 665, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "an audio-visual cross-modality framework based on transformer [5]. Particularly, [22] builds a person-agnostic lip-sync framework by StyleGAN2 [27], which also supports personalized optimization. Similarly, [10] also provides identity refinement for the final animation results. Although they have shown accurate lip-sync and comparably high-fidelity results, visible artifacts may still exist due to the latent contradiction between textural and spatial information. Few studies [48, 50] built upon diffusion models [44] are presented recently. Our ReSyncer generates high-fidelity lip-sync videos with high stability and efficiency.", + "bbox": [ + 212, + 146, + 782, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Audio-Driven 3D Facial Animation. The first stage of our work shares a similar setting with methods that generate audio-driven 3D facial animation [13, 18,43,54,69]. Most of them are performed on 3D scan data to predict the point clouds' dynamics. They have explored Transformers [57] and codebooks [56] with different settings on this task. We adopt a highly efficient form to use Transformer in our work.", + "bbox": [ + 212, + 267, + 784, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Face Swapping", + "text_level": 1, + "bbox": [ + 215, + 380, + 385, + 396 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To transfer the identity of one person to the face of another person is a longstanding problem in computer vision. In recent years, GAN [21] has played an important role in this task. RSGAN [38] and FSNet [39] encode the facial region and the non-facial region to different latent codes and perform swapping by replacing these latent representations. [83] proposes a face-swapping method MegaFS built upon the GAN-inversion technique with StyleGAN2, however, the results are not temporal consistent. Similarly, [71, 73] make their attempts to apply the strong prior learned by StyleGAN2 [27] to generate high-resolution swapped faces. Recently, [31] takes inspiration from 3D generative model [4] to capture fine-grained details of face shape and strengthen the robustness under large pose variance by extending the face swapping task into 3D latent space. However, they fail to produce results with accurate poses and expressions when balancing identity similarity and attribution preservation of the target frame.", + "bbox": [ + 212, + 406, + 784, + 601 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Different from these methods, ReSyncer takes advantage of intermediate guidance from 3D mesh to synthesize accurate pose and expression. Moreover, we achieve face-swapping in a compatible with accurate lip-syncing, which allows us to further satisfy more imaginative demands in creating virtual presenters and performers.", + "bbox": [ + 212, + 602, + 784, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 702, + 380, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we introduce the details of our ReSyncer framework as depicted in Fig. 2. It consists of two stages which are all simply designed and easy to implement. In section 3.1 we illustrate our design of the style-injected lip-sync Transformer (Style-SyncFormer) which predicts 3D facial dynamics. Then we show how to reconfigure the Style-based generator to support high-fidelity generalized and personalized lip-sync in section 3.2. Finally, we illustrate our unified design naturally supports face-swapping training in section 3.3.", + "bbox": [ + 212, + 733, + 782, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Guan et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/434071d6863b2949c47f490e6a9ce740f7003eb00ea42586851c7f12e4e6a63b.jpg", + "image_caption": [ + "Fig. 2: The ReSyncer Framework. In the first stage, the Style-SyncFormer takes the style template and audio input to predict 3D facial dynamics. Then the predicted mesh overlays on the target frame to provide strong spatial guidance. The Style-based generator G processes the overlay and reference frames to produce the final result." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 781, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Task Formulation. The task of generalized lip-sync is normally trained on massive in-the-wild videos. Specifically, a training data consists of video clip $\\mathbf{V} = \\{I_1,\\dots ,I_T\\}$ with $T$ frames and the accompanying (processed) audio clip features $\\mathbf{a} = \\{a_{1},\\ldots ,a_{T}\\}$ . The lower half of the target frame $I_{t}$ is masked out by a mask $A$ to form $I_t^A = (\\mathbf{1} - A)*I_t$ as [10,22,41,51]. The normal training objective is to recover the original frame given the audio clip $a_{t}$ and a reference frame $I_{ref}$ . However, the final results might be interfered with by the structure of the reference frame when the estimated animation significantly differs from the reference.", + "bbox": [ + 212, + 380, + 787, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this paper, we involve roughly predicted 3D facial meshes as intermediate representations. For every frame in the training video, we reconstruct the 3D facial meshes $\\mathbf{M} = \\{M_1,\\dots ,M_T\\}$ . Thus our framework can be divided into the stages of predicting 3D facial dynamics from audio and the recovery of frames $I_{t}$ with the assistance of $M_{t}$ .", + "bbox": [ + 212, + 516, + 787, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Style-Injected Lip-Sync Transformer", + "text_level": 1, + "bbox": [ + 214, + 616, + 563, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the first stage, we propose the Style-SyncFormer that takes audio features $\\mathbf{a}_{t_0:t_1} = \\{a_{t_0},\\dots ,a_{t_1}\\}$ extracted from pretrained Wav2Vec2 [1] to predict 3D meshes $\\mathbf{M}_{t_0:t_1} = \\{M_{t_0},\\dots ,M_{t_1}\\}$ . There are $N$ consecutive frames between $t_0$ and $t_1$ . Previous work [18] has explored abilities to learn 3D dynamics but it handles only a fixed number of identities. In contrast, the proposed Style-SyncFormer facilitates zero-shot prediction of 3D dynamics while preserving the original speaking style or transferring it from another person.", + "bbox": [ + 212, + 643, + 787, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data Processing. During training, all pose information is eliminated from the meshes and aligned to the canonical space. Instead of directly learning all vertices' positions, we learn the displacements compared to a template mesh $\\bar{M}$ , the mean of all meshes from this person without expression coefficients. The training target becomes predicting the displacements $\\Delta \\mathbf{M}_{t_0:t_1} = \\{\\Delta M_{t_0},\\dots \\Delta M_{t_1}\\}$ between the template and each reconstructed result.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "ReSyncer", + "bbox": [ + 665, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Network Design. The main network of Style-SyncFormer is built based on only one Transformer [57] encoder block for style information encoding and one decoder block for mesh prediction, where we adopt the periodic positional encoding and biased causal multi-head attention from FaceFormer [18].", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During training, each shifted-right displacement $\\Delta M_{t}$ is mapped to embedding $d_t^m$ and sent into the positional encoding to serve as tokens $f_{t}^{m}$ . Previous studies [18, 43] learn the 3D dynamics of only limited identities with a simple one-hot speaker identity embedding [13, 18]. Differently, we encode the information of a total of $T_{S} = 90$ consecutive $\\Delta M_{ref}$ as speaking style reference. The style information is embedded as a feature $f_{S}^{r}$ from $\\Delta M_{ref}$ using an additional Transformer encoder block. After that, we inject the speaking style $f_{S}^{r}$ by adding it with mesh tokens $f_{t}^{m}$ . Note that when replacing $\\Delta M_{ref}$ with mesh sequences of another identity, speaking style transfer can be achieved.", + "bbox": [ + 212, + 207, + 784, + 342 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the decoder block, the input tokens are sent into multi-head self-attention, cross-attention, and feed-forward layers. The audio features are further encoded as embeddings $f_{t}^{a}$ and leveraged in the cross-attentions. The final output is transformed to $\\Delta \\hat{M}_{t_0:t_1} = \\{\\Delta \\hat{M}_{t_0},\\dots ,\\Delta \\hat{M}_{t_1}\\}$ .", + "bbox": [ + 212, + 343, + 784, + 404 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Loss Function. The training objectives can be written as:", + "bbox": [ + 215, + 404, + 643, + 419 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {3 D} = \\sum_ {t = t _ {0} + 1} ^ {t _ {1}} \\| (\\varDelta \\hat {M} _ {t} - \\varDelta \\hat {M} _ {t - 1}) - (\\varDelta M _ {t} - \\varDelta M _ {t - 1}) \\| _ {2} + \\lambda_ {1} \\sum_ {t = t _ {0}} ^ {t _ {1}} \\| \\varDelta \\hat {M} _ {t} - \\varDelta M _ {t} \\| _ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 431, + 785, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the first term is the temporal consistency loss [55].", + "bbox": [ + 214, + 484, + 627, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The training is performed in a teacher-forcing manner while the network works autoregressively during inference by using previously predicted results. The displacements are finally transferred to 3D meshes $\\hat{\\mathbf{M}}_{t_0:t_1}$ in the canonical space by adding the template mesh $\\bar{M}$ .", + "bbox": [ + 212, + 500, + 784, + 560 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Rewiring Style-based Generator", + "text_level": 1, + "bbox": [ + 215, + 584, + 526, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the second stage, we rewire the widely used Style-based generator [27] to faithfully transfer 3D meshes $\\mathbf{M}$ to high-fidelity facial frames $\\hat{\\mathbf{I}}$ in a simple manner. Previous methods [6,55] resort to accurately textured rendering, which is inefficient from the perspective of implementation and training. Our goal is to find an efficient and effective way. As a result, we leverage the basic white-colored rendering as spatial guidance, which can be easily achieved without textural mapping.", + "bbox": [ + 212, + 611, + 784, + 719 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Network Design. We resort to Style-based generator with noise insertion scheme, which is easy to implement and specializes well in face-related tasks [22, 35,73,74]. Specifically, we project mesh $M_{t}$ on masked target frame $I_{t}^{A}$ 's domain to $M_{t}^{p}$ and overlay to form a 3D-guided image $I_{t}^{AM} = I_{t}^{A} + A * M_{t}^{p}$ . A reference frame $I_{ref}$ is attached for providing textural information in the spatial form. The 6-channel input $[I_{t}^{AM}, I_{ref}]$ is encoded to feature maps $\\mathbf{F}_{f} = \\{f_{f}, F_{1}, \\ldots, F_{L}\\}$ that match the sizes of Style-based generator's each stage and $f_{f}$ being a vector. Then we follow previous studies [22,73,74] and insert $\\mathbf{F}_{f}$ into the generator $G$ .", + "bbox": [ + 212, + 719, + 784, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Guan et al.", + "bbox": [ + 271, + 114, + 364, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As $I_{t}^{AM}$ directly provides the 3D mouth shape information, the reference frame can hardly influence the prediction result as long as it is not the target frame itself. Thus we randomly sample $I_{ref}$ within the nearest 30 frames of $I_{t}$ .", + "bbox": [ + 212, + 145, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A number of previous lip-sync studies [22,32,59,77,81] rely on encoding audio information into the $\\mathcal{W}$ space for animation. However, with explicit guidance, the driving information would play a less important role in the $\\mathcal{W}$ space. Meanwhile, the face-swapping task illustrates that $\\mathcal{W}$ space embedding benefits identity information [35, 73, 83] which we intend to enhance. To avoid involving extra network design and inference procedures, we directly send the encoded $f_{f}$ from $[I_t^{AM}, I_{ref}]$ to the $\\mathcal{W}$ , so that the reference frame provides more facial detail while $I_t^{AM}$ also passes dynamic information.", + "bbox": [ + 212, + 191, + 787, + 311 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Loss Functions. To sum up, in the second stage, the generator takes the output only of a single facial encoder and predicts the target frame itself as $\\hat{I}_t$ . The training objective is much simpler with the 3D guidance. We only adopt the same loss functions as Pix2pix-HD [65], which are the $L_1$ , VGG, and adversarial loss empowered with StyleGAN2's discriminator.", + "bbox": [ + 212, + 311, + 787, + 388 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e c} = \\left\\| \\hat {I} _ {t} - I _ {t} \\right\\| _ {1} + \\lambda_ {2} \\sum_ {m = 1} ^ {N _ {v g g}} \\left\\| \\operatorname {V G G} _ {m} (\\hat {I} _ {t}) - \\operatorname {V G G} _ {m} (I _ {t}) \\right\\| _ {1}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 305, + 398, + 785, + 439 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a d v} = \\underset {\\mathrm {G}} {\\operatorname * {m i n m a x}} _ {\\mathrm {D}} \\left(\\mathbb {E} _ {I _ {t}} \\left[ \\log \\mathrm {D} \\left(I _ {t}\\right) \\right] + \\mathbb {E} _ {\\hat {I} _ {t}} \\left[ \\log \\left(1 - \\mathrm {D} \\left(\\hat {I} _ {t}\\right)\\right) \\right]\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 463, + 784, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As the rendered meshes provide much clearer guidance than audio, the generator can be easily trained without task-oriented losses or specific tuning.", + "bbox": [ + 212, + 493, + 782, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "During inference, we overlay the predicted mesh $\\hat{M}_t^p$ from the first stage to form the input $[\\hat{I}_t^{AM}, I_{ref}]$ . Though the predicted $\\hat{M}_t^p$ cannot perfectly match the reconstructed results, we rely on the powerful generator to compensate for such disturbance.", + "bbox": [ + 212, + 523, + 782, + 582 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Video-Driven Lip-Sync. The generator only depends on mesh information, thus we can transfer the displacements of meshes $\\Delta \\mathbf{M}'$ of another person to the target template $\\bar{M}$ so that video-driven lip-sync can be naturally achieved.", + "bbox": [ + 212, + 584, + 782, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Personalized Fine-tuning. Given thatSYNC information is derived from the 3D facial mesh, the generator in the second stage will not overfit the limited lip patterns present in given clips. This masks our method full-parameter fine-tunable on only seconds of videos.", + "bbox": [ + 212, + 628, + 782, + 689 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Face-Swapping", + "text_level": 1, + "bbox": [ + 214, + 710, + 385, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We now illustrate the power of our comprehensive framework to directly achieve face-swapping and lip-sync with unified training. Our design of leveraging the $\\mathcal{W}$ space for enhancing facial generation considers the similar setting used in previous face swapping studies [35, 72, 73]. Moreover, the identity information of the audio-predicted mesh $\\hat{M}_t$ can be easily transferred by sending different template meshes into the style encoder of the Style-SyncFormer. So that the shape of facial organs can be seamlessly swapped to another person.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "ReSyncer", + "bbox": [ + 665, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5e05de95fed44f81f507b1eeeba50ef9faff99bc4619ac190956d9a3b898e764.jpg", + "image_caption": [ + "Fig. 3: Face-Swapping Pipeline. With input data reconfiguration and additional training losses, we achieve lip-syncing and face-swapping simultaneously." + ], + "image_footnote": [], + "bbox": [ + 302, + 143, + 697, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Face-Swapping Data Stream. The pipeline is depicted in Fig. 3. We first enlarge the masking area $A$ to $A'$ that contains the whole facial part including eyes and noses, which is more suitable for the concerned task. Then we involve a frame $I'$ of another identity and its corresponding mesh $M'$ to provide additional information. We project $M'$ with the pose of $M_t^p$ to $M'^p$ and overlay it on $I_t$ to form $I_t^{A'M'} = (1 - A') * I_t + A' * M'^p$ .", + "bbox": [ + 212, + 373, + 787, + 463 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The face-swapping training consists of two parts, updating only the Style-based generator. The first is the same as Sec 3.2, which reconstructs $I_{t}$ solely from $\\{I_t^{AM}, I_{ref}\\}$ . The second part intends to create an image $\\hat{I}_t'$ , which has $I'$ 's identity on $I_{t}$ 's head with $I_{t}$ 's original tone. This requires the generator G to take $\\mathbf{F}_f' = \\{f_f'', F_1', \\dots, F_L'\\}$ encoded from different sources, including $f_f''$ encoded from $\\{I_t^{A'M'}, I'\\}$ for ID injection in the $\\mathcal{W}$ space and $\\{F_1', \\dots, F_L'\\}$ from $\\{I_t^{A'M'}, I_{ref}\\}$ for expressing facial attributes in the noise space.", + "bbox": [ + 214, + 464, + 787, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training Objectives. Under the cross-identity training setting, the commonly used identity cosine distance loss $\\mathcal{L}_{id} = 1 - \\cos (E_{id}(\\hat{I}_t'),E_{id}(I'))$ where $E_{id}$ is the encoder of ArcFace [14] and feature matching loss with the original $I_{t}$ [9,70,73] in the face-swapping task are leveraged.", + "bbox": [ + 212, + 574, + 787, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Note that after the network is trained in the face-swapping style, which costs more time, high-quality face-swapping and lip-sync abilities can be achieved within one unified model (with the same set of weights). To the best of our knowledge, we are one of the first methods that unifies the two tasks within one model, which is efficient and beneficial to the creation of virtual performers.", + "bbox": [ + 212, + 635, + 787, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 733, + 375, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although lip-sync is our main concern, we thoroughly conduct experiments from the perspectives of both lip-sync and face-swapping to show the effectiveness of our unified framework. For fair comparisons, we present results using only our generalized model, please refer to the supplementary files for personalized experiments.", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Guan et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7bd5cd42c8e6a187f12d217d10b2d9a8684b5f03a0e77624b26071c23d3d21fc.jpg", + "image_caption": [ + "Fig. 4: Qualitative Cross-Sync Results. The top row shows the lip-synced videos of the driving audio. Generation results based on the \"Template\" row should have the same lip shape as the \"Lip-Synced Video\" in the first row." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 522, + 433 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/cc4d1286ad3cbf4138659bd7db49b64ab2b5b588ead0b0ee88d574ae8fe092fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 143, + 785, + 433 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. We train our model with two audio-visual datasets, HDTF [79] and VoxCeleb2 [11]. Both two datasets provide high-quality videos from a variety of identities for lip-syncing and face-swapping training. Notably, we split HDTF into train and test sets without overlap. The test set comprises 100 10-second audio-video pairs. Evaluations on lip-sync are conducted on the official test set of VoxCeleb2 and the introduced test set of HDTF. Face-swapping evaluation is performed on FaceForensics++ (FF++) [45], which contains 1000 face videos. We use the high-quality version and adopt the official video-video pairing. Apart from publicly available datasets, certain visual outcomes in our experiments are produced based on videos sourced from the Internet.", + "bbox": [ + 212, + 527, + 787, + 679 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details. All videos are processed at a frame rate of 25FPS, and facial alignment is conducted based on pre-detected landmarks. Subsequently, all faces are cropped to a size of $256 \\times 256$ . The architecture of G is built upon the success of recent methods [22,28,73,74]. During inference, we split videos into 100-frame consecutive chunks with a 20-frame overlap. Face meshes are predicted in an auto-aggressive manner with the proposed Style-SyncFormer depending on all the previously predicted results within each chunk. For face reconstruction, we train DeepFaceReconstruction [17] on the topology of both Hifi3DFace [2] and BFM. The main results are reported from the Hifi3DFace version but the choice of topology makes little difference. Our models are trained", + "bbox": [ + 212, + 688, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "ReSyncer", + "bbox": [ + 666, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/da8b585601480c42421d681d8834c4a0cb8d017cb45bc7a46b2ad76987b5dedb.jpg", + "table_caption": [ + "Table 1: Quantitative results on HDTF and VoxCeleb2. For LMD and $\\Delta$ Sync the lower the better, and the higher the better for others." + ], + "table_footnote": [], + "table_body": "
MethodHDTFVoxCeleb2
SSIM ↑PSNR ↑LMD ↓ΔSync ↓SSIM ↑PSNR ↑LMD ↓ΔSync ↓
Wav2Lip [41]0.8430.935.411.370.8431.583.442.28
ReTalking [10]0.8131.086.060.740.8331.644.421.50
SadTalker [78]0.7730.866.763.620.7931.365.033.03
StyleSync [22]0.8331.574.970.800.8532.523.281.57
Ours-Sync0.8431.764.340.660.8532.883.190.88
", + "bbox": [ + 220, + 174, + 782, + 296 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "on 4 Tesla A100 GPUs for 2 days with batch size 16. For more implementation details, please refer to the supplementary files.", + "bbox": [ + 212, + 325, + 782, + 356 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Comparison Methods. As for lip-sync evaluation, We compare our method with four different methods including the classic Wav2Lip [41], as well as three recent state-of-the-art (SOTA) methods: ReTalking [10], SadTalker [78], and StyleSync [22]. Note that SadTalker [78] targets talking head generation but is well-known for its good quality. All experiments are conducted on the generalized setting without fine-tuning, and the model without face-swapping training is leveraged. Then we compare our method with four studies on face swapping. The counterparts consist of SimSwap [9], InfoSwap [20], StyleSwap [73] and E4S [34]. Furthermore, we aim to evaluate whether combining two methods from the above two tasks can achieve comparable results with our full model despite the cost. Thus we construct two potential counterparts by combining SOTA methods from both areas, denoted as \"SimSwap + ReTalking\" and \"StyleSwap + StyleSync\".", + "bbox": [ + 212, + 356, + 784, + 537 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Quantitative Comparisons", + "text_level": 1, + "bbox": [ + 215, + 560, + 477, + 574 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Quantitative experiments of lip-sync are conducted following the self-reconstruction setting. We refrain from directly utilizing the ground truth frame as the reference. For face-swapping, we adhere to a protocol as previous studies [30, 73], wherein we uniformly sample 10 frames from each video in $\\mathrm{FF}++$ , resulting in a total of $10\\mathrm{k}$ images for evaluation.", + "bbox": [ + 212, + 585, + 800, + 659 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Metrics. Both lip-sync and face-swapping are broadly studied, we follow previous works [22, 30, 73, 81] to adopt several popular metrics. For lip-sync, we use SSIM [66] and PSNR for measuring visual similarity. Landmark", + "bbox": [ + 212, + 660, + 784, + 707 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/404ec5b450660cc4153a2349f4ffba9df7dcbb68efd79827fe25e89e61c2615f.jpg", + "table_caption": [ + "Table 2: Quantitative results evaluated on Table 3: Quantitative ablations on FF++. VoxCeleb2." + ], + "table_footnote": [], + "table_body": "
MethodID Retrieval↑ ID Sim.↑ PoseError ↓ Exp.Error ↓
SimSwap [9]93.07%0.5781.365.07
InfoSwap [20]95.82%0.6352.546.99
StyleSwap [73]97.05%0.6771.565.28
EAS [34]90.21%0.4952.7811.52
Ours-Swap96.70%0.6651.315.13
", + "bbox": [ + 217, + 762, + 522, + 835 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/dea9476f5558f5703ee76a5db87f740b12a4293ecae199f80ad81559e8ca02b4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSSIM ↑PSNR ↑LMD ↓ΔSync ↓
w/o Mesh-Inject0.8331.573.602.54
w/o Mesh-Style0.8331.803.291.02
TS=500.8532.263.350.95
TS=2000.8532.693.210.88
Ours-Sync0.8532.883.190.88
", + "bbox": [ + 526, + 762, + 777, + 835 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Guan et al.", + "bbox": [ + 271, + 114, + 364, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/ef52110d304742805e54b3a5f5faf20df4ae4e8c02d811e12a564f8b3091bf57.jpg", + "image_caption": [ + "Fig. 5: Qualitative Results of Face-Swap. Identity-swapped results should preserve the expression and lip motion of templates." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 380, + 281 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/5dd25155ad673664f535a3379ff6724071e4a8388727d7a8ed9a73acf877fdb6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 143, + 785, + 281 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "distances around the mouth (LMD), and SyncNet's confidence score [12] are used for lip-sync quality. Particularly, the SyncNet score only indicates how well a video aligns with the learned SyncNet model, thus we use the absolute difference ( $\\Delta$ Sync) between ground truth and prediction to assess whether results are close to the original video. For face-swapping, we leverage a pretrained face-recognition model [60] to measure the ID cosine similarity and ID retrieval scores. In addition, pose error and expression error are measured by two pretrained models [46,58] following previous work [73].", + "bbox": [ + 212, + 340, + 787, + 462 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Lip-Sync. The evaluation results are tabulated in Table 1. Our method outperforms previous works across all metrics, especially concerning lip-sync quality. This is attributed to stylized 3D facial dynamics learned by Style-SyncFormer, which manages to recover the injected personalized speaking style of the target person in the generalized setting. In contrast, previous methods primarily rely on sync-loss [41] and tend to generate a dataset-averaged sync-style.", + "bbox": [ + 212, + 462, + 787, + 551 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Face-Swapping. Results evaluated on $\\mathrm{FF}++$ are listed in Table 2. Our model shows comparable performance on all the metrics and achieves the lowest pose error when compared with SOTA face-swap methods. Particularly, our performances on ID-related metrics are only slightly weaker than StyleSwap.", + "bbox": [ + 212, + 551, + 787, + 613 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Qualitative Comparisons", + "text_level": 1, + "bbox": [ + 215, + 633, + 465, + 648 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Subjective evaluation plays a pivotal role in assessing the capabilities of generative models, especially in the context of videos. We strongly encourage readers to watch our supplementary video for a comprehensive understanding. Our method also enjoys the capability to fast optimize for specific identities, resulting in superior outcomes. Please refer to the supplementary files for the personalized results.", + "bbox": [ + 212, + 657, + 787, + 747 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Lip-Sync. We demonstrate two cross-sync comparisons with SOTA methods in Fig. 4. The lip-synced results are conditioned on audio randomly sampled from the dataset. Wav2Lip has accurate lip shapes, but the mouth part is noticeably blurred. SadTalker achieves better fidelity, but it fails to present natural dynamics. VideoReTalking and StyleSync can generate precise and high-quality results; nevertheless, their results are affected by the opening jaw of the template", + "bbox": [ + 212, + 748, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "ReSyncer", + "bbox": [ + 666, + 114, + 730, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "as shown in the first column of the right figure. Moreover, they exhibit a deficiency in maintaining person-specific attributes around the mouth. Our method directly produces high-fidelity results with accurate lip motion and rich details of the mouth.", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We present cross-sync results, showcasing the transferred speaking style in Fig. 1 and Fig. 9 (in the supplementary files). It can be seen that our method not only accurately generates lip shapes aligned with the driven video but also successfully animates the target person with the speaking style of the source person in the driving video. In addition, our method inherently supports videodriven lip-syncing as introduced in Sec. 3.2. The comparison is demonstrated in Fig. 10 (in the supplementary files), it shows that person-specific attributes can be preserved better with Style-SyncFormer guidance.", + "bbox": [ + 212, + 208, + 787, + 330 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/862bbb804ef064a9f567289f633a614e12bc58e419e920f7ed6d0fb1a0965432.jpg", + "image_caption": [ + "Fig.6: Results of Face-Swapped Lip-Sync. The ID-swapped results are driven by the given audio. We compare with the combinations of two SOTA methods in lip-sync and face-swapping. Results generated by our method preserve details with better fidelity and simultaneously maintain a similar speaking style to the source." + ], + "image_footnote": [], + "bbox": [ + 274, + 358, + 730, + 626 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Face-Swapping. In Fig. 5, we illustrate two face-swapped results for comparison. The generated results in this context should preserve the original expression and lip motion of the template. From the figure, InfoSwap fails to keep a natural brightness across frames. E4S effectively modifies the entire head region but fails to integrate the head part into the background seamlessly. SimSwap and StyleSwap produce good results by preserving precise expression and identity attributes. However, an excessive emphasis on identity feature matching causes them to falter in generating details around certain areas like eyes and mouth.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Guan et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/b6f5a84bdd511830263ccb928c1efc6b46be839ef6082420ee1cea61d4860ec4.jpg", + "table_caption": [ + "Table 4: User study of generalized lip-sync. The scores are ranged from 1 (worst) to 5 (best)." + ], + "table_footnote": [], + "table_body": "
MOS onWav2LipReTalkingSadTalkerStyleSyncOurs-Sync
Lip-Sync Quality3.433.092.983.464.57
Generation Quality1.482.833.133.634.67
Video Realness1.592.412.873.114.54
", + "bbox": [ + 274, + 184, + 722, + 252 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In contrast, our method not only achieves effective identity transfer but also preserves rich details across the entire face. In addition, the videos generated by our method exhibit superior temporal consistency, especially in the mouth region. Please find more results in the supplementary video.", + "bbox": [ + 212, + 282, + 782, + 343 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Furthermore, our method is capable of achieving lip-synced face-swapping conditioned on any audio and identity in a single forward process. We compare our method with two potential baselines \"SimSwap + ReTalking\" and \"StyleSwap + StyleSync\" in Fig. 6. Results generated by our method preserve details with better fidelity and simultaneously maintain a similar speaking style to the source.", + "bbox": [ + 212, + 345, + 784, + 419 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "User Study. We conducted a user study with 30 participants. We adopt the commonly used Mean Opinion Scores (MOS) rating protocol. All users give their ratings, ranging from 1 to 5. On generalized lip-sync, they are asked to evaluate: a) the quality of lip-sync. b) visual quality of the generated video, and c) whether this video looks real. The results averaged from 10 videos are grouped into Table 4. From the table, our method outperforms the four counterparts in all three aspects by clear margins, underscoring the power of our method and the substantial improvement in visual quality and realism over other existing methods. The experiments are similarly conducted with face-swapping and personalized lip-sync, please find the results in the supplementary files.", + "bbox": [ + 212, + 420, + 782, + 571 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 598, + 388, + 614 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Lip-Sync. Here we evaluate four different designs on VoxCeleb2 [11] including: 1) \"w/o Mesh-Inject\": we mask-out the 3D mesh guidance from the encoded feature maps $\\{F_1, \\ldots, F_L\\}$ , thus it cannot provide detailed spatial information during generation; 2) \"w/o Mesh-Style\": 3D mesh content is removed in $I_t^{AM}$ to get a $f_f$ , feature vector that feed into the $\\mathcal{W}$ space contains only the background and reference information; 3) \" $T_S = 50$ \": we reduce the length ( $T_S$ in Sec. 3.1) of the speaking-style reference from 90 to 50; 4) \" $T_S = 200$ \": the length is extended to 200. The results are presented in Table 3. Comparisons between the first two rows and \"Ours-Sync\" reveal the vital role played by mesh guidance in the generation process, where the sync quality is degraded to certain degrees, as indicated by LMD and $\\Delta$ Sync. In addition, the visual quality is also harmed as shown in Fig. 7 (a). In the bottom three rows of Table 3, we compare the performance of different sequence lengths in the speaking-style encoding of Style-SyncFormer. The results indicate that a sequence of 90 frames ( $\\sim$ 4s of", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "ReSyncer", + "bbox": [ + 665, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f0a60a64fef2a460b4b44cdbe1665b37ab6aacd3df6c18b25396c66e43cc9d33.jpg", + "image_caption": [ + "Fig. 7: Ablations. (a) 3D facial mesh provides detailed spatial guidance, empowering superior lip-syncing. (b) The mesh with shapes of facial organs also enhances ID transfer in face-swapping." + ], + "image_footnote": [], + "bbox": [ + 272, + 143, + 728, + 339 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "video) leads to good sync quality, and a further extension of the length does not enhance its performance.", + "bbox": [ + 212, + 415, + 784, + 444 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Face-Swapping. In face-swapping, the predicted 3D facial mesh also provides detailed facial attributes. We provide an intuitive comparison in Fig. 7 (b). When the mesh guidance is removed from the encoded feature map (\"w/o Mesh-Inject\"), we observe a noteworthy degradation in facial dynamics, particularly in lip motion. Please refer to the supplementary video for the comparison. Simultaneously, without the guidance of mesh information, the ID-related metrics of face-swapping are also affected, as indicated at the bottom of Fig. 7 (b) and the same phenomenon applies to \"w/o Mesh-Style\".", + "bbox": [ + 212, + 445, + 787, + 566 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 215, + 589, + 508, + 606 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Ethical Considerations. Our method possesses the capability to create fabricated talks and speeches, even with swapped identities, presenting a potential for malicious use. We will restrict the distribution of our models and the generated results exclusively to research purposes.", + "bbox": [ + 212, + 621, + 784, + 680 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Conclusion. We highlight several important properties of our ReSyncer framework: 1) Our easy reconfiguration further unveils the potential of widely studied structures so that high-quality generalized lip-sync results with mesh representation can be achieved. 2) Our framework is designed to adopt external identity information, thus we achieve face-swapping comparable to the state of the arts while keeping the lip-sync ability within one unified model. 3) Our network supports speaking style transfer, video-driven facial animation, and can be applied to real-time live streamers. These properties complementarily satisfy various needs of virtual performer creation under different circumstances.", + "bbox": [ + 212, + 681, + 787, + 816 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Guan et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments. This work is in part supported by National Natural Science Foundation of China with No. 62394322, Beijing Natural Science Foundation with No. L222024, as well as the Beijing National Research Center for Information Science and Technology (BNRist) key projects. This study is also supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s).", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 291, + 325, + 308 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Baevski, A., Zhou, Y., Mohamed, A., Auli, M.: wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in Neural Information Processing Systems (2020)", + "2. Bao, L., Lin, X., Chen, Y., Zhang, H., Wang, S., Zhe, X., Kang, D., Huang, H., Jiang, X., Wang, J., Yu, D., Zhang, Z.: High-fidelity 3d digital human head creation from rgb-d selfies. ACM Transactions on Graphics (2021)", + "3. Blanz, V., Vetter, T.: A morphable model for the synthesis of 3d faces. In: Proceedings of the 26th annual conference on Computer graphics and interactive techniques. pp. 187-194 (1999)", + "4. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022)", + "5. Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: Maskgit: Masked generative image transformer. arXiv preprint arXiv:2202.04200 (2022)", + "6. Chen, L., Cui, G., Liu, C., Li, Z., Kou, Z., Xu, Y., Xu, C.: Talking-head generation with rhythmic head motion. In: European Conference on Computer Vision. pp. 35-51. Springer (2020)", + "7. Chen, L., Li, Z., Maddox, R.K., Duan, Z., Xu, C.: Lip movements generation at a glance. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 520-535 (2018)", + "8. Chen, L., Maddox, R.K., Duan, Z., Xu, C.: Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7832-7841 (2019)", + "9. Chen, R., Chen, X., Ni, B., Ge, Y.: Simswap: An efficient framework for high fidelity face swapping. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 2003-2011 (2020)", + "0. Cheng, K., Cun, X., Zhang, Y., Xia, M., Yin, F., Zhu, M., Wang, X., Wang, J., Wang, N.: Videoretalking: Audio-based lip synchronization for talking head video editing in the wild (2022)", + "1. Chung, J.S., Nagrani, A., Zisserman, A.: Voxceleb2: Deep speaker recognition. In: INTERSPEECH (2018)", + "12. Chung, J.S., Zisserman, A.: Out of time: automated lip sync in the wild. In: ACCV (2016)", + "13. Cudeiro, D., Bolkart, T., Laidlaw, C., Ranjan, A., Black, M.J.: Capture, learning, and synthesis of 3d speaking styles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10101-10111 (2019)" + ], + "bbox": [ + 218, + 324, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "ReSyncer", + "bbox": [ + 665, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4690-4699 (2019)", + "15. Deng, Y., Yang, J., Chen, D., Wen, F., Tong, X.: Disentangled and controllable face image generation via 3d imitative-contrastive learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)", + "16. Deng, Y., Yang, J., Xiang, J., Tong, X.: Gram: Generative radiance manifolds for 3d-aware image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10673-10683 (June 2022)", + "17. Deng, Y., Yang, J., Xu, S., Chen, D., Jia, Y., Tong, X.: Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019)", + "18. Fan, Y., Lin, Z., Saito, J., Wang, W., Komura, T.: Faceformer: Speech-driven 3d facial animation with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18770-18780 (2022)", + "19. Gafni, G., Thies, J., Zollhöfer, M., Nießner, M.: Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8649-8658 (June 2021)", + "20. Gao, G., Huang, H., Fu, C., Li, Z., He, R.: Information bottleneck disentanglement for identity swapping. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3404-3413 (2021)", + "21. Goodfellow, I.J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. arXiv preprint arXiv:1406.2661 (2014)", + "22. Guan, J., Zhang, Z., Zhou, H., Hu, T., Wang, K., He, D., Feng, H., Liu, J., Ding, E., Liu, Z., Wang, J.: Stylesync: High-fidelity generalized and personalized lip sync in style-based generator. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "23. Guo, Y., Chen, K., Liang, S., Liu, Y., Bao, H., Zhang, J.: Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2021)", + "24. Jamaludin, A., Chung, J.S., Zisserman, A.: You said that?: Synthesising talking faces from audio. International Journal of Computer Vision 127(11), 1767-1779 (2019)", + "25. Ji, X., Zhou, H., Wang, K., Wu, Q., Wu, W., Xu, F., Cao, X.: Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. SIGGRAPH (2022)", + "26. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4401-4410 (2019)", + "27. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of StyleGAN. In: Proc. CVPR (2020)", + "28. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8110-8119 (2020)", + "29. Li, L., Wang, S., Zhang, Z., Ding, Y., Zheng, Y., Yu, X., Fan, C.: Write-a-speaker: Text-based emotional and rhythmic talking-head generation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1911-1920 (2021)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Guan et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "30. Li, L., Bao, J., Yang, H., Chen, D., Wen, F.: Faceshifter: Towards high fidelity and occlusion aware face swapping. arXiv preprint arXiv:1912.13457 (2019)", + "31. Li, Y., Ma, C., Yan, Y., Zhu, W., Yang, X.: 3d-aware face swapping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12705-12714 (2023)", + "32. Liang, B., Pan, Y., Guo, Z., Zhou, H., Hong, Z., Han, X., Han, J., Liu, J., Ding, E., Wang, J.: Expressive talking head generation with granular audio-visual control. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3387-3396 (2022)", + "33. Liu, X., Xu, Y., Wu, Q., Zhou, H., Wu, W., Zhou, B.: Semantic-aware implicit neural audio-driven video portrait generation. ECCV (2022)", + "34. Liu, Z., Li, M., Zhang, Y., Wang, C., Zhang, Q., Wang, J., Nie, Y.: Fine-grained face swapping via regional gan inversion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8578-8587 (2023)", + "35. Luo, Y., Zhu, J., He, K., Chu, W., Tai, Y., Wang, C., Yan, J.: Styleface: Towards identity-disentangled face generation on megapixels. In: European Conference on Computer Vision. pp. 297-312. Springer (2022)", + "36. Ma, Y., Wang, S., Ding, Y., Ma, B., Lv, T., Fan, C., Hu, Z., Deng, Z., Yu, X.: Talkclip: Talking head generation with text-guided expressive speaking styles. arXiv preprint arXiv:2304.00334 (2023)", + "37. Ma, Y., Wang, S., Hu, Z., Fan, C., Lv, T., Ding, Y., Deng, Z., Yu, X.: Styletalk: One-shot talking head generation with controllable speaking styles. AAAI (2023)", + "38. Natsume, R., Yatagawa, T., Morishima, S.: Rsgan: Face swapping and editing using face and hair representation in latent spaces. arxiv 2018. arXiv preprint arXiv:1804.03447 2", + "39. Natsume, R., Yatagawa, T., Morishima, S.: Fsnet: An identity-aware generative model for image-based face swapping. In: Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part VI 14. pp. 117–132. Springer (2019)", + "40. Park, S.J., Kim, M., Hong, J., Choi, J., Ro, Y.M.: Synctalkface: Talking face generation with precise lip-syncing via audio-lip memory. In: AAAI Conference on Artificial Intelligence. Association for the Advancement of Artificial Intelligence (2022)", + "41. Prajwal, K., Mukhopadhyay, R., Namboodiri, V.P., Jawahar, C.: A lip sync expert is all you need for speech to lip generation in the wild. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 484-492 (2020)", + "42. Qiu, H., Chen, Z., Jiang, Y., Zhou, H., Fan, X., Yang, L., Wu, W., Liu, Z.: Relitalk: Relightable talking portrait generation from a single video. International Journal of Computer Vision pp. 1-16 (2024)", + "43. Richard, A., Zollhöfer, M., Wen, Y., de la Torre, F., Sheikh, Y.: Meshtalk: 3d face animation from speech using cross-modality disentanglement. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021)", + "44. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)", + "45. Rössler, A., Cozzolino, D., Verdoliva, L., Riess, C., Thies, J., Nießner, M.: FaceForensics++: Learning to detect manipulated facial images. In: International Conference on Computer Vision (ICCV) (2019)", + "46. Ruiz, N., Chong, E., Rehg, J.M.: Fine-grained head pose estimation without keypoints. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 2074-2083 (2018)" + ], + "bbox": [ + 212, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "ReSyncer", + "bbox": [ + 665, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "47. Shen, S., Li, W., Zhu, Z., Duan, Y., Zhou, J., Lu, J.: Learning dynamic facial radiance fields for few-shot talking head synthesis. In: ECCV (2022)", + "48. Shen, S., Zhao, W., Meng, Z., Li, W., Zhu, Z., Zhou, J., Lu, J.: Difftalk: Crafting diffusion models for generalized audio-driven portraits animation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1982-1991 (2023)", + "49. Song, Y., Zhu, J., Li, D., Wang, X., Qi, H.: Talking face generation by conditional recurrent adversarial network. arXiv preprint arXiv:1804.04786 (2018)", + "50. Stypulkowski, M., Vougioukas, K., He, S., Zieba, M., Petridis, S., Pantic, M.: Diffused heads: Diffusion models beat gans on talking-face generation. arXiv preprint arXiv:2301.03396 (2023)", + "51. Sun, Y., Zhou, H., Wang, K., Wu, Q., Hong, Z., Liu, J., Ding, E., Wang, J., Liu, Z., Hideki, K.: Masked lip-sync prediction by audio-visual contextual exploitation in transformers. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)", + "52. Suwajanakorn, S., Seitz, S.M., Kemelmacher-Shlizerman, I.: Synthesizing obama: learning lip sync from audio. ACM Transactions on Graphics (ToG) 36(4), 1-13 (2017)", + "53. Tang, J., Wang, K., Zhou, H., Chen, X., He, D., Hu, T., Liu, J., Zeng, G., Wang, J.: Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368 (2022)", + "54. Thambiraja, B., Habibie, I., Aliakbarian, S., Cosker, D., Theobalt, C., Thies, J.: Imitator: Personalized speech-driven 3d facial animation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 20621-20631 (2023)", + "55. Thies, J., Elgharib, M., Tewari, A., Theobalt, C., Nießner, M.: Neural voice puppetry: Audio-driven facial reenactment. In: European Conference on Computer Vision. pp. 716-731. Springer (2020)", + "56. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems 30 (2017)", + "57. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)", + "58. Vemulapalli, R., Agarwala, A.: A compact embedding for facial expression similarity. In: proceedings of the IEEE/cvf conference on computer vision and pattern recognition. pp. 5683-5692 (2019)", + "59. Wang, D., Deng, Y., Yin, Z., Shum, H.Y., Wang, B.: Progressive disentangled representation learning for fine-grained controllable talking head synthesis. CVPR (2023)", + "60. Wang, H., Wang, Y., Zhou, Z., Ji, X., Gong, D., Zhou, J., Li, Z., Liu, W.: Cosface: Large margin cosine loss for deep face recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5265-5274 (2018)", + "61. Wang, J., Zhao, K., Ma, Y., Zhang, S., Zhang, Y., Shen, Y., Zhao, D., Zhou, J.: Facecomposer: A unified model for versatile facial content creation. In: Thirty-seventh Conference on Neural Information Processing Systems (2023)", + "62. Wang, K., Wu, Q., Song, L., Yang, Z., Wu, W., Qian, C., He, R., Qiao, Y., Loy, C.C.: Mead: A large-scale audio-visual dataset for emotional talking-face generation. In: ECCV (2020)", + "63. Wang, S., Li, L., Ding, Y., Fan, C., Yu, X.: Audio2head: Audio-driven one-shot talking-head generation with natural head motion. arXiv preprint arXiv:2107.09293 (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "J. Guan et al.", + "bbox": [ + 271, + 114, + 364, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "64. Wang, S., Li, L., Ding, Y., Yu, X.: One-shot talking face generation from single-speaker audio-visual correlation learning. In: Proceedings of the AAAI Conference on Artificial Intelligence (2022)", + "65. Wang, T.C., Liu, M.Y., Zhu, J.Y., Tao, A., Kautz, J., Catanzaro, B.: High-resolution image synthesis and semantic manipulation with conditional gans. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8798-8807 (2018)", + "66. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)", + "67. Wu, H., Jia, J., Wang, H., Dou, Y., Duan, C., Deng, Q.: Imitating arbitrary talking style for realistic audio-driven talking face synthesis. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 1478-1486 (2021)", + "68. Wu, X., Hu, P., Wu, Y., Lyu, X., Cao, Y.P., Shan, Y., Yang, W., Sun, Z., Qi, X.: Speech2lip: High-fidelity speech to lip generation by learning from a short video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22168-22177 (2023)", + "69. Xing, J., Xia, M., Zhang, Y., Cun, X., Wang, J., Wong, T.T.: Codetalker: Speechdriven 3d facial animation with discrete motion prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12780-12790 (June 2023)", + "70. Xu, C., Zhang, J., Hua, M., He, Q., Yi, Z., Liu, Y.: Region-aware face swapping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7632-7641 (2022)", + "71. Xu, Y., Deng, B., Wang, J., Jing, Y., Pan, J., He, S.: High-resolution face swapping via latent semantics disentanglement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7642-7651 (2022)", + "72. Xu, Z., Hong, Z., Ding, C., Zhu, Z., Han, J., Liu, J., Ding, E.: Mobilefaceswap: A lightweight framework for video face swapping. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 2973-2981 (2022)", + "73. Xu, Z., Zhou, H., Hong, Z., Liu, Z., Liu, J., Guo, Z., Han, J., Liu, J., Ding, E., Wang, J.: Styleswap: Style-based generator empowers robust face swapping. In: European Conference on Computer Vision. pp. 661-677. Springer (2022)", + "74. Yang, T., Ren, P., Xie, X., Zhang, L.: Gan prior embedded network for blind face restoration in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 672-681 (2021)", + "75. Yao, S., Zhong, R., Yan, Y., Zhai, G., Yang, X.: Dfa-nerf: Personalized talking head generation via disentangled face attributes neural rendering. arXiv preprint arXiv:2201.00791 (2022)", + "76. Ye, Z., Jiang, Z., Ren, Y., Liu, J., He, J., Zhao, Z.: Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. ICLR (2023)", + "77. Yu, Z., Yin, Z., Zhou, D., Wang, D., Wong, F., Wang, B.: Talking head generation with probabilistic audio-to-visual diffusion priors. ICCV (2023)", + "78. Zhang, W., Cun, X., Wang, X., Zhang, Y., Shen, X., Guo, Y., Shan, Y., Wang, F.: Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8652-8661 (2023)", + "79. Zhang, Z., Li, L., Ding, Y., Fan, C.: Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3661-3670 (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "ReSyncer", + "bbox": [ + 666, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "80. Zhou, H., Liu, Y., Liu, Z., Luo, P., Wang, X.: Talking face generation by adversarially disentangled audio-visual representation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 33, pp. 9299-9306 (2019)", + "81. Zhou, H., Sun, Y., Wu, W., Loy, C.C., Wang, X., Liu, Z.: Pose-controllable talking face generation by implicitly modularized audio-visual representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4176-4186 (2021)", + "82. Zhou, Y., Han, X., Shechtman, E., Echevarria, J., Kalogerakis, E., Li, D.: Makelttalk: speaker-aware talking-head animation. ACM Transactions on Graphics (TOG) 39(6), 1-15 (2020)", + "83. Zhu, Y., Li, Q., Wang, J., Xu, C.Z., Sun, Z.: One shot face swapping on megapixels. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4834-4844 (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 327 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "J. Guan et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/2024/ReSyncer_ Rewiring Style-based Generator for Unified Audio-Visually Synced Facial Performer/6fe8e31e-41da-4533-8af1-187e2be9d843_model.json b/2024/ReSyncer_ Rewiring Style-based Generator for Unified Audio-Visually Synced Facial Performer/6fe8e31e-41da-4533-8af1-187e2be9d843_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e5b9f46add07227760d44ff20e12bb6c5ad8a340 --- /dev/null +++ b/2024/ReSyncer_ Rewiring Style-based Generator for Unified Audio-Visually Synced Facial Performer/6fe8e31e-41da-4533-8af1-187e2be9d843_model.json @@ -0,0 +1,2638 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.228, + 0.142, + 0.777, + 0.187 + ], + "angle": 0, + "content": "ReSyncer: Rewiring Style-based Generator for Unified Audio-Visually Synced Facial Performer" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.212, + 0.787, + 0.26 + ], + "angle": 0, + "content": "Jiazhi Guan\\(^{1,2*}\\), Zhiliang Xu\\(^{2*}\\), Hang Zhou\\(^{2\\dagger}\\), Kaisiyuan Wang\\(^{2}\\), Shengyi He\\(^{2}\\), Zhanwang Zhang\\(^{2}\\), Borong Liang\\(^{2}\\), Haocheng Feng\\(^{2}\\), Errui Ding\\(^{2}\\), Jingtuo Liu\\(^{2}\\), Jingdong Wang\\(^{2}\\), Youjian Zhao\\(^{1,3\\dagger}\\), and Ziwei Liu\\(^{4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.269, + 0.745, + 0.314 + ], + "angle": 0, + "content": "\\(^{1}\\)BNRist, DCST, Tsinghua University \\(^{2}\\)Baidu Inc. \n\\(^{3}\\)Zhongguancun Laboratory \\(^{4}\\)S-Lab, Nanyang Technological University guanjz20@mails.tsinghua.edu.cn, zhouhang09@baidu.com" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.338, + 0.778, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.569, + 0.788, + 0.613 + ], + "angle": 0, + "content": "Fig. 1: Lip-Synced/Speaking-Style-Transferred/Face-Swapping Results by ReSyncer. Our method not only produces high-fidelity lip-synced video according to audio but can further transfer the speaking style and identity of any target person." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.636, + 0.744, + 0.803 + ], + "angle": 0, + "content": "Abstract. Lip-syncing videos with given audio is the foundation for various applications including the creation of virtual presenters or performers. While recent studies explore high-fidelity lip-sync with different techniques, their task-orientated models either require long-term videos for clip-specific training or retain visible artifacts. In this paper, we propose a unified and effective framework ReSyncer, that synchronizes generalized audio-visual facial information. The key design is revisiting and rewiring the Style-based generator to efficiently adopt 3D facial dynamics predicted by a principled style-injected Transformer. By simply re-configuring the information insertion mechanisms within the noise and style space, our framework fuses motion and appearance with unified training. Extensive experiments demonstrate that ReSyncer not" + }, + { + "type": "page_footnote", + "bbox": [ + 0.239, + 0.811, + 0.409, + 0.84 + ], + "angle": 0, + "content": "\\*Equal Contribution. \\(\\dagger\\) Corresponding Authors." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "J. Guan et al." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.148, + 0.744, + 0.219 + ], + "angle": 0, + "content": "only produces high-fidelity lip-synced videos according to audio, but also supports multiple appealing properties that are suitable for creating virtual presenters and performers, including fast personalized fine-tuning, video-driven lip-syncing, the transfer of speaking styles, and even face swapping." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.239, + 0.376, + 0.255 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.267, + 0.787, + 0.327 + ], + "angle": 0, + "content": "Generating lip-synced facial videos with speech audio is a foundational technique for various applications including the dubbing of movies and the creation of virtual presenters or performers in the form of realistic humans. The needs in virtual performers' creation can hardly be comprehensively met." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.328, + 0.787, + 0.508 + ], + "angle": 0, + "content": "While great attention has been paid to creating dynamic lip-synced videos from a few photos [8, 24, 63, 64, 78, 80-82], previous results are still far from realistic. To generate life-like videos, researchers have made explorations [41] with GANs [10,55], NeRF [23,33,47,53], and Diffusion models [48] to edit existing videos. Despite their efforts, most methods rely on a relatively long-term clip of video for training, making them less flexible and efficient. A few methods [10,22, 41,51] enable generalized local editing without training. However, they still tend to create visible flaws on high-quality videos [41,51] Particularly, StyleSync [22] supports not only arbitrary video modeling but can adapt to the speaking style of a target person with short-clip fine-tuning, which balances the generation quality and efficiency. Nevertheless, their results can still be easily affected by mouth movements on their conditional reference frames, leading to unstable results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.787, + 0.643 + ], + "angle": 0, + "content": "The previous cases reveal the limitations of using low-dimensional audio information to directly modify high-dimensional visual data. On the other hand, intermediate representations like 3D facial models are widely used, but their potential as spatial guidance for efficient, generalized editing remains underexplored. Using only 3DMM [3] coefficients [15,16,19,29,37,55] also suffers from the cross-domain information injection problem. While textured rendering offers a promising alternative [6,55,67], its implementation complexity and susceptibility to errors during 3D modeling or texture fitting can lead to degraded results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.787, + 0.81 + ], + "angle": 0, + "content": "In this paper, we revisit and re-wire the powerful Style-based generator [26, 27] to generate faithful lip-synced videos and support virtual performer creation in a concise yet effective framework called ReSyncer. With unified model training, various appealing properties including generalized and personalized lip-sync, speaking style transfer, live streaming, and face swapping can be achieved within one model. Our key is to show how an easy reconfiguration empowers Style-based generator with various powerful audio-visual synchronization abilities by involving 3D facial dynamics. Specifically, we use only roughly fitted 3D facial meshes, instead of coefficients to bridge the audio and image domains. A principled Style-injected Lip-Sync Transformer (Style-SyncFormer) learns stylized point cloud displacements given template meshes and driving audios." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We then perform easy but crucial modifications to Style-based generator with a simple encoder used in previous work [22, 73, 74]. Concretely, while lip-sync" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.666, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ReSyncer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "studies [10,22,41,51,61] mask out the mouth area of the target frame, we overlay reconstructed mesh onto it as strong spatial guidance. Then we compliment the textural and identity information with an attached reference frame, which serves as the source of both the noise and style-related space. With generalized training on massive data, the powerful generator focuses on recovering facial details from the roughly predicted lip-synced white-colored meshes, leading to high-fidelity and stable lip-sync with higher quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.374 + ], + "angle": 0, + "content": "Additionally, our framework design goes beyond existing solutions by uniformly addressing a broader range of needs in virtual performer creation. These needs include, for instance, preserving personal styles for specific users or changing face identities for privacy protection and diverse character generation. Particularly, our method directly supports few-shot few-step personalized fine-tuning. Moreover, our framework supports *syncing* (swapping) different identities on the same target video with additional training strategies. High-fidelity face-swapping and lip-syncing can hence be simultaneously satisfied." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.376, + 0.796, + 0.556 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows: 1) We propose the ReSyncer framework, which shows the powerfulness of Style-based generator on syncing audio-visual facial information by involving 3D facial meshes with a simple reconfiguration. 2) We propose the Style-SyncFormer which learns stylized 3D facial dynamics with simple Transformer blocks, enabling generalized 3D facial animation. 3) Extensive experiments show that our framework not only enables lip-syncing of higher stability and quality, but also supports various intriguing properties that are essential for creating virtual performers, including fast personalized fine-tuning, video-driven lip-syncing, the transfer of speaking styles, and even face swapping. To the best of our knowledge, we are one of the first methods that support state-of-the-art lip-syncing and face-swapping with one unified model." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.581, + 0.39, + 0.598 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.616, + 0.524, + 0.631 + ], + "angle": 0, + "content": "2.1 Audio-Driven Facial Animation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Talking Head Generation. A great number of studies focus on predicting the whole human head movements according to audios [7, 8, 23-25, 32, 36, 37, 42, 52, 59, 62, 63, 77, 78, 81]. More recently, one-shot-based methods adopt representations from 3D, videos, and even texts for more controllable results. While other approaches [23, 33, 47, 53, 75, 76] attempt to integrate speech audios into the neural radiance field for personalized audio-to-3D modeling. However, these methods usually fail to achieve stable lip-sync results and show terrible robustness performance when using audio in the wild. Moreover, the form of talking head generation cannot be applied to edit arbitrary video." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Lip-Syncing-based Facial Editing. Apart from generating the entire head, certain studies [10, 22, 40, 41, 49, 51, 55, 68] choose to directly edit the mouth shape according to the input audios. Specifically, [41] relies on a well-trained lip-sync discriminator to produce highly synchronized animations. [51] introduces" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "J. Guan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.265 + ], + "angle": 0, + "content": "an audio-visual cross-modality framework based on transformer [5]. Particularly, [22] builds a person-agnostic lip-sync framework by StyleGAN2 [27], which also supports personalized optimization. Similarly, [10] also provides identity refinement for the final animation results. Although they have shown accurate lip-sync and comparably high-fidelity results, visible artifacts may still exist due to the latent contradiction between textural and spatial information. Few studies [48, 50] built upon diffusion models [44] are presented recently. Our ReSyncer generates high-fidelity lip-sync videos with high stability and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.785, + 0.356 + ], + "angle": 0, + "content": "Audio-Driven 3D Facial Animation. The first stage of our work shares a similar setting with methods that generate audio-driven 3D facial animation [13, 18,43,54,69]. Most of them are performed on 3D scan data to predict the point clouds' dynamics. They have explored Transformers [57] and codebooks [56] with different settings on this task. We adopt a highly efficient form to use Transformer in our work." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.381, + 0.386, + 0.397 + ], + "angle": 0, + "content": "2.2 Face Swapping" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.407, + 0.785, + 0.602 + ], + "angle": 0, + "content": "To transfer the identity of one person to the face of another person is a longstanding problem in computer vision. In recent years, GAN [21] has played an important role in this task. RSGAN [38] and FSNet [39] encode the facial region and the non-facial region to different latent codes and perform swapping by replacing these latent representations. [83] proposes a face-swapping method MegaFS built upon the GAN-inversion technique with StyleGAN2, however, the results are not temporal consistent. Similarly, [71, 73] make their attempts to apply the strong prior learned by StyleGAN2 [27] to generate high-resolution swapped faces. Recently, [31] takes inspiration from 3D generative model [4] to capture fine-grained details of face shape and strengthen the robustness under large pose variance by extending the face swapping task into 3D latent space. However, they fail to produce results with accurate poses and expressions when balancing identity similarity and attribution preservation of the target frame." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.603, + 0.785, + 0.679 + ], + "angle": 0, + "content": "Different from these methods, ReSyncer takes advantage of intermediate guidance from 3D mesh to synthesize accurate pose and expression. Moreover, we achieve face-swapping in a compatible with accurate lip-syncing, which allows us to further satisfy more imaginative demands in creating virtual presenters and performers." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.703, + 0.381, + 0.72 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.784, + 0.84 + ], + "angle": 0, + "content": "In this section, we introduce the details of our ReSyncer framework as depicted in Fig. 2. It consists of two stages which are all simply designed and easy to implement. In section 3.1 we illustrate our design of the style-injected lip-sync Transformer (Style-SyncFormer) which predicts 3D facial dynamics. Then we show how to reconfigure the Style-based generator to support high-fidelity generalized and personalized lip-sync in section 3.2. Finally, we illustrate our unified design naturally supports face-swapping training in section 3.3." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.666, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ReSyncer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.144, + 0.782, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.295, + 0.788, + 0.351 + ], + "angle": 0, + "content": "Fig. 2: The ReSyncer Framework. In the first stage, the Style-SyncFormer takes the style template and audio input to predict 3D facial dynamics. Then the predicted mesh overlays on the target frame to provide strong spatial guidance. The Style-based generator G processes the overlay and reference frames to produce the final result." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.788, + 0.516 + ], + "angle": 0, + "content": "Task Formulation. The task of generalized lip-sync is normally trained on massive in-the-wild videos. Specifically, a training data consists of video clip \\(\\mathbf{V} = \\{I_1,\\dots ,I_T\\}\\) with \\(T\\) frames and the accompanying (processed) audio clip features \\(\\mathbf{a} = \\{a_{1},\\ldots ,a_{T}\\}\\). The lower half of the target frame \\(I_{t}\\) is masked out by a mask \\(A\\) to form \\(I_t^A = (\\mathbf{1} - A)*I_t\\) as [10,22,41,51]. The normal training objective is to recover the original frame given the audio clip \\(a_{t}\\) and a reference frame \\(I_{ref}\\). However, the final results might be interfered with by the structure of the reference frame when the estimated animation significantly differs from the reference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.517, + 0.788, + 0.593 + ], + "angle": 0, + "content": "In this paper, we involve roughly predicted 3D facial meshes as intermediate representations. For every frame in the training video, we reconstruct the 3D facial meshes \\(\\mathbf{M} = \\{M_1,\\dots ,M_T\\}\\). Thus our framework can be divided into the stages of predicting 3D facial dynamics from audio and the recovery of frames \\(I_{t}\\) with the assistance of \\(M_{t}\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.617, + 0.564, + 0.633 + ], + "angle": 0, + "content": "3.1 Style-Injected Lip-Sync Transformer" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.749 + ], + "angle": 0, + "content": "In the first stage, we propose the Style-SyncFormer that takes audio features \\(\\mathbf{a}_{t_0:t_1} = \\{a_{t_0},\\dots ,a_{t_1}\\}\\) extracted from pretrained Wav2Vec2 [1] to predict 3D meshes \\(\\mathbf{M}_{t_0:t_1} = \\{M_{t_0},\\dots ,M_{t_1}\\}\\). There are \\(N\\) consecutive frames between \\(t_0\\) and \\(t_1\\). Previous work [18] has explored abilities to learn 3D dynamics but it handles only a fixed number of identities. In contrast, the proposed Style-SyncFormer facilitates zero-shot prediction of 3D dynamics while preserving the original speaking style or transferring it from another person." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Data Processing. During training, all pose information is eliminated from the meshes and aligned to the canonical space. Instead of directly learning all vertices' positions, we learn the displacements compared to a template mesh \\(\\bar{M}\\), the mean of all meshes from this person without expression coefficients. The training target becomes predicting the displacements \\(\\Delta \\mathbf{M}_{t_0:t_1} = \\{\\Delta M_{t_0},\\dots \\Delta M_{t_1}\\}\\) between the template and each reconstructed result." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.128 + ], + "angle": 0, + "content": "J. Guan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "Network Design. The main network of Style-SyncFormer is built based on only one Transformer [57] encoder block for style information encoding and one decoder block for mesh prediction, where we adopt the periodic positional encoding and biased causal multi-head attention from FaceFormer [18]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.343 + ], + "angle": 0, + "content": "During training, each shifted-right displacement \\(\\Delta M_{t}\\) is mapped to embedding \\(d_t^m\\) and sent into the positional encoding to serve as tokens \\(f_{t}^{m}\\). Previous studies [18, 43] learn the 3D dynamics of only limited identities with a simple one-hot speaker identity embedding [13, 18]. Differently, we encode the information of a total of \\(T_{S} = 90\\) consecutive \\(\\Delta M_{ref}\\) as speaking style reference. The style information is embedded as a feature \\(f_{S}^{r}\\) from \\(\\Delta M_{ref}\\) using an additional Transformer encoder block. After that, we inject the speaking style \\(f_{S}^{r}\\) by adding it with mesh tokens \\(f_{t}^{m}\\). Note that when replacing \\(\\Delta M_{ref}\\) with mesh sequences of another identity, speaking style transfer can be achieved." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.785, + 0.405 + ], + "angle": 0, + "content": "In the decoder block, the input tokens are sent into multi-head self-attention, cross-attention, and feed-forward layers. The audio features are further encoded as embeddings \\( f_{t}^{a} \\) and leveraged in the cross-attentions. The final output is transformed to \\( \\Delta \\hat{M}_{t_0:t_1} = \\{\\Delta \\hat{M}_{t_0},\\dots ,\\Delta \\hat{M}_{t_1}\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.405, + 0.644, + 0.42 + ], + "angle": 0, + "content": "Loss Function. The training objectives can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.433, + 0.787, + 0.473 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {3 D} = \\sum_ {t = t _ {0} + 1} ^ {t _ {1}} \\| (\\varDelta \\hat {M} _ {t} - \\varDelta \\hat {M} _ {t - 1}) - (\\varDelta M _ {t} - \\varDelta M _ {t - 1}) \\| _ {2} + \\lambda_ {1} \\sum_ {t = t _ {0}} ^ {t _ {1}} \\| \\varDelta \\hat {M} _ {t} - \\varDelta M _ {t} \\| _ {2},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.485, + 0.628, + 0.501 + ], + "angle": 0, + "content": "where the first term is the temporal consistency loss [55]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.501, + 0.785, + 0.561 + ], + "angle": 0, + "content": "The training is performed in a teacher-forcing manner while the network works autoregressively during inference by using previously predicted results. The displacements are finally transferred to 3D meshes \\(\\hat{\\mathbf{M}}_{t_0:t_1}\\) in the canonical space by adding the template mesh \\(\\bar{M}\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.585, + 0.527, + 0.601 + ], + "angle": 0, + "content": "3.2 Rewiring Style-based Generator" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.785, + 0.72 + ], + "angle": 0, + "content": "In the second stage, we rewire the widely used Style-based generator [27] to faithfully transfer 3D meshes \\(\\mathbf{M}\\) to high-fidelity facial frames \\(\\hat{\\mathbf{I}}\\) in a simple manner. Previous methods [6,55] resort to accurately textured rendering, which is inefficient from the perspective of implementation and training. Our goal is to find an efficient and effective way. As a result, we leverage the basic white-colored rendering as spatial guidance, which can be easily achieved without textural mapping." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.785, + 0.841 + ], + "angle": 0, + "content": "Network Design. We resort to Style-based generator with noise insertion scheme, which is easy to implement and specializes well in face-related tasks [22, 35,73,74]. Specifically, we project mesh \\( M_{t} \\) on masked target frame \\( I_{t}^{A} \\)'s domain to \\( M_{t}^{p} \\) and overlay to form a 3D-guided image \\( I_{t}^{AM} = I_{t}^{A} + A * M_{t}^{p} \\). A reference frame \\( I_{ref} \\) is attached for providing textural information in the spatial form. The 6-channel input \\( [I_{t}^{AM}, I_{ref}] \\) is encoded to feature maps \\( \\mathbf{F}_{f} = \\{f_{f}, F_{1}, \\ldots, F_{L}\\} \\) that match the sizes of Style-based generator's each stage and \\( f_{f} \\) being a vector. Then we follow previous studies [22,73,74] and insert \\( \\mathbf{F}_{f} \\) into the generator \\( G \\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.666, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ReSyncer" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.192 + ], + "angle": 0, + "content": "As \\( I_{t}^{AM} \\) directly provides the 3D mouth shape information, the reference frame can hardly influence the prediction result as long as it is not the target frame itself. Thus we randomly sample \\( I_{ref} \\) within the nearest 30 frames of \\( I_{t} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.788, + 0.313 + ], + "angle": 0, + "content": "A number of previous lip-sync studies [22,32,59,77,81] rely on encoding audio information into the \\(\\mathcal{W}\\) space for animation. However, with explicit guidance, the driving information would play a less important role in the \\(\\mathcal{W}\\) space. Meanwhile, the face-swapping task illustrates that \\(\\mathcal{W}\\) space embedding benefits identity information [35, 73, 83] which we intend to enhance. To avoid involving extra network design and inference procedures, we directly send the encoded \\(f_{f}\\) from \\([I_t^{AM}, I_{ref}]\\) to the \\(\\mathcal{W}\\), so that the reference frame provides more facial detail while \\(I_t^{AM}\\) also passes dynamic information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.788, + 0.389 + ], + "angle": 0, + "content": "Loss Functions. To sum up, in the second stage, the generator takes the output only of a single facial encoder and predicts the target frame itself as \\(\\hat{I}_t\\). The training objective is much simpler with the 3D guidance. We only adopt the same loss functions as Pix2pix-HD [65], which are the \\(L_1\\), VGG, and adversarial loss empowered with StyleGAN2's discriminator." + }, + { + "type": "equation", + "bbox": [ + 0.307, + 0.399, + 0.786, + 0.44 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e c} = \\left\\| \\hat {I} _ {t} - I _ {t} \\right\\| _ {1} + \\lambda_ {2} \\sum_ {m = 1} ^ {N _ {v g g}} \\left\\| \\operatorname {V G G} _ {m} (\\hat {I} _ {t}) - \\operatorname {V G G} _ {m} (I _ {t}) \\right\\| _ {1}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.315, + 0.464, + 0.785, + 0.486 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a d v} = \\underset {\\mathrm {G}} {\\operatorname * {m i n m a x}} _ {\\mathrm {D}} \\left(\\mathbb {E} _ {I _ {t}} \\left[ \\log \\mathrm {D} \\left(I _ {t}\\right) \\right] + \\mathbb {E} _ {\\hat {I} _ {t}} \\left[ \\log \\left(1 - \\mathrm {D} \\left(\\hat {I} _ {t}\\right)\\right) \\right]\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.784, + 0.525 + ], + "angle": 0, + "content": "As the rendered meshes provide much clearer guidance than audio, the generator can be easily trained without task-oriented losses or specific tuning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.525, + 0.784, + 0.583 + ], + "angle": 0, + "content": "During inference, we overlay the predicted mesh \\(\\hat{M}_t^p\\) from the first stage to form the input \\([\\hat{I}_t^{AM}, I_{ref}]\\). Though the predicted \\(\\hat{M}_t^p\\) cannot perfectly match the reconstructed results, we rely on the powerful generator to compensate for such disturbance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.784, + 0.629 + ], + "angle": 0, + "content": "Video-Driven Lip-Sync. The generator only depends on mesh information, thus we can transfer the displacements of meshes \\(\\Delta \\mathbf{M}'\\) of another person to the target template \\(\\bar{M}\\) so that video-driven lip-sync can be naturally achieved." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.784, + 0.69 + ], + "angle": 0, + "content": "Personalized Fine-tuning. Given thatSYNC information is derived from the 3D facial mesh, the generator in the second stage will not overfit the limited lip patterns present in given clips. This masks our method full-parameter fine-tunable on only seconds of videos." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.711, + 0.386, + 0.727 + ], + "angle": 0, + "content": "3.3 Face-Swapping" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We now illustrate the power of our comprehensive framework to directly achieve face-swapping and lip-sync with unified training. Our design of leveraging the \\(\\mathcal{W}\\) space for enhancing facial generation considers the similar setting used in previous face swapping studies [35, 72, 73]. Moreover, the identity information of the audio-predicted mesh \\(\\hat{M}_t\\) can be easily transferred by sending different template meshes into the style encoder of the Style-SyncFormer. So that the shape of facial organs can be seamlessly swapped to another person." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "J. Guan et al." + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.144, + 0.699, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.317, + 0.788, + 0.346 + ], + "angle": 0, + "content": "Fig. 3: Face-Swapping Pipeline. With input data reconfiguration and additional training losses, we achieve lip-syncing and face-swapping simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.464 + ], + "angle": 0, + "content": "The Face-Swapping Data Stream. The pipeline is depicted in Fig. 3. We first enlarge the masking area \\( A \\) to \\( A' \\) that contains the whole facial part including eyes and noses, which is more suitable for the concerned task. Then we involve a frame \\( I' \\) of another identity and its corresponding mesh \\( M' \\) to provide additional information. We project \\( M' \\) with the pose of \\( M_t^p \\) to \\( M'^p \\) and overlay it on \\( I_t \\) to form \\( I_t^{A'M'} = (1 - A') * I_t + A' * M'^p \\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.465, + 0.788, + 0.576 + ], + "angle": 0, + "content": "The face-swapping training consists of two parts, updating only the Style-based generator. The first is the same as Sec 3.2, which reconstructs \\( I_{t} \\) solely from \\( \\{I_t^{AM}, I_{ref}\\} \\). The second part intends to create an image \\( \\hat{I}_t' \\), which has \\( I' \\)'s identity on \\( I_{t} \\)'s head with \\( I_{t} \\)'s original tone. This requires the generator G to take \\( \\mathbf{F}_f' = \\{f_f'', F_1', \\dots, F_L'\\} \\) encoded from different sources, including \\( f_f'' \\) encoded from \\( \\{I_t^{A'M'}, I'\\} \\) for ID injection in the \\( \\mathcal{W} \\) space and \\( \\{F_1', \\dots, F_L'\\} \\) from \\( \\{I_t^{A'M'}, I_{ref}\\} \\) for expressing facial attributes in the noise space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.575, + 0.788, + 0.636 + ], + "angle": 0, + "content": "Training Objectives. Under the cross-identity training setting, the commonly used identity cosine distance loss \\(\\mathcal{L}_{id} = 1 - \\cos (E_{id}(\\hat{I}_t'),E_{id}(I'))\\) where \\(E_{id}\\) is the encoder of ArcFace [14] and feature matching loss with the original \\(I_{t}\\) [9,70,73] in the face-swapping task are leveraged." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.636, + 0.788, + 0.711 + ], + "angle": 0, + "content": "Note that after the network is trained in the face-swapping style, which costs more time, high-quality face-swapping and lip-sync abilities can be achieved within one unified model (with the same set of weights). To the best of our knowledge, we are one of the first methods that unifies the two tasks within one model, which is efficient and beneficial to the creation of virtual performers." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.734, + 0.376, + 0.751 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Although lip-sync is our main concern, we thoroughly conduct experiments from the perspectives of both lip-sync and face-swapping to show the effectiveness of our unified framework. For fair comparisons, we present results using only our generalized model, please refer to the supplementary files for personalized experiments." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.667, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ReSyncer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.144, + 0.523, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.144, + 0.787, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.437, + 0.788, + 0.48 + ], + "angle": 0, + "content": "Fig. 4: Qualitative Cross-Sync Results. The top row shows the lip-synced videos of the driving audio. Generation results based on the \"Template\" row should have the same lip shape as the \"Lip-Synced Video\" in the first row." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.528, + 0.788, + 0.68 + ], + "angle": 0, + "content": "Datasets. We train our model with two audio-visual datasets, HDTF [79] and VoxCeleb2 [11]. Both two datasets provide high-quality videos from a variety of identities for lip-syncing and face-swapping training. Notably, we split HDTF into train and test sets without overlap. The test set comprises 100 10-second audio-video pairs. Evaluations on lip-sync are conducted on the official test set of VoxCeleb2 and the introduced test set of HDTF. Face-swapping evaluation is performed on FaceForensics++ (FF++) [45], which contains 1000 face videos. We use the high-quality version and adopt the official video-video pairing. Apart from publicly available datasets, certain visual outcomes in our experiments are produced based on videos sourced from the Internet." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Implementation Details. All videos are processed at a frame rate of 25FPS, and facial alignment is conducted based on pre-detected landmarks. Subsequently, all faces are cropped to a size of \\(256 \\times 256\\). The architecture of G is built upon the success of recent methods [22,28,73,74]. During inference, we split videos into 100-frame consecutive chunks with a 20-frame overlap. Face meshes are predicted in an auto-aggressive manner with the proposed Style-SyncFormer depending on all the previously predicted results within each chunk. For face reconstruction, we train DeepFaceReconstruction [17] on the topology of both Hifi3DFace [2] and BFM. The main results are reported from the Hifi3DFace version but the choice of topology makes little difference. Our models are trained" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.128 + ], + "angle": 0, + "content": "J. Guan et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.787, + 0.173 + ], + "angle": 0, + "content": "Table 1: Quantitative results on HDTF and VoxCeleb2. For LMD and \\(\\Delta\\)Sync the lower the better, and the higher the better for others." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.175, + 0.784, + 0.297 + ], + "angle": 0, + "content": "
MethodHDTFVoxCeleb2
SSIM ↑PSNR ↑LMD ↓ΔSync ↓SSIM ↑PSNR ↑LMD ↓ΔSync ↓
Wav2Lip [41]0.8430.935.411.370.8431.583.442.28
ReTalking [10]0.8131.086.060.740.8331.644.421.50
SadTalker [78]0.7730.866.763.620.7931.365.033.03
StyleSync [22]0.8331.574.970.800.8532.523.281.57
Ours-Sync0.8431.764.340.660.8532.883.190.88
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.327, + 0.784, + 0.357 + ], + "angle": 0, + "content": "on 4 Tesla A100 GPUs for 2 days with batch size 16. For more implementation details, please refer to the supplementary files." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.357, + 0.785, + 0.538 + ], + "angle": 0, + "content": "Comparison Methods. As for lip-sync evaluation, We compare our method with four different methods including the classic Wav2Lip [41], as well as three recent state-of-the-art (SOTA) methods: ReTalking [10], SadTalker [78], and StyleSync [22]. Note that SadTalker [78] targets talking head generation but is well-known for its good quality. All experiments are conducted on the generalized setting without fine-tuning, and the model without face-swapping training is leveraged. Then we compare our method with four studies on face swapping. The counterparts consist of SimSwap [9], InfoSwap [20], StyleSwap [73] and E4S [34]. Furthermore, we aim to evaluate whether combining two methods from the above two tasks can achieve comparable results with our full model despite the cost. Thus we construct two potential counterparts by combining SOTA methods from both areas, denoted as \"SimSwap + ReTalking\" and \"StyleSwap + StyleSync\"." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.561, + 0.478, + 0.575 + ], + "angle": 0, + "content": "4.1 Quantitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.586, + 0.802, + 0.66 + ], + "angle": 0, + "content": "Quantitative experiments of lip-sync are conducted following the self-reconstruction setting. We refrain from directly utilizing the ground truth frame as the reference. For face-swapping, we adhere to a protocol as previous studies [30, 73], wherein we uniformly sample 10 frames from each video in \\(\\mathrm{FF}++\\), resulting in a total of \\(10\\mathrm{k}\\) images for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.661, + 0.785, + 0.708 + ], + "angle": 0, + "content": "Evaluation Metrics. Both lip-sync and face-swapping are broadly studied, we follow previous works [22, 30, 73, 81] to adopt several popular metrics. For lip-sync, we use SSIM [66] and PSNR for measuring visual similarity. Landmark" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.734, + 0.78, + 0.761 + ], + "angle": 0, + "content": "Table 2: Quantitative results evaluated on Table 3: Quantitative ablations on FF++. VoxCeleb2." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.763, + 0.524, + 0.837 + ], + "angle": 0, + "content": "
MethodID Retrieval↑ ID Sim.↑ PoseError ↓ Exp.Error ↓
SimSwap [9]93.07%0.5781.365.07
InfoSwap [20]95.82%0.6352.546.99
StyleSwap [73]97.05%0.6771.565.28
EAS [34]90.21%0.4952.7811.52
Ours-Swap96.70%0.6651.315.13
" + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.763, + 0.778, + 0.837 + ], + "angle": 0, + "content": "
MethodSSIM ↑PSNR ↑LMD ↓ΔSync ↓
w/o Mesh-Inject0.8331.573.602.54
w/o Mesh-Style0.8331.803.291.02
TS=500.8532.263.350.95
TS=2000.8532.693.210.88
Ours-Sync0.8532.883.190.88
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.667, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "ReSyncer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.382, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.144, + 0.787, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.285, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Fig. 5: Qualitative Results of Face-Swap. Identity-swapped results should preserve the expression and lip motion of templates." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.463 + ], + "angle": 0, + "content": "distances around the mouth (LMD), and SyncNet's confidence score [12] are used for lip-sync quality. Particularly, the SyncNet score only indicates how well a video aligns with the learned SyncNet model, thus we use the absolute difference (\\(\\Delta\\)Sync) between ground truth and prediction to assess whether results are close to the original video. For face-swapping, we leverage a pretrained face-recognition model [60] to measure the ID cosine similarity and ID retrieval scores. In addition, pose error and expression error are measured by two pretrained models [46,58] following previous work [73]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.788, + 0.553 + ], + "angle": 0, + "content": "Lip-Sync. The evaluation results are tabulated in Table 1. Our method outperforms previous works across all metrics, especially concerning lip-sync quality. This is attributed to stylized 3D facial dynamics learned by Style-SyncFormer, which manages to recover the injected personalized speaking style of the target person in the generalized setting. In contrast, previous methods primarily rely on sync-loss [41] and tend to generate a dataset-averaged sync-style." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.614 + ], + "angle": 0, + "content": "Face-Swapping. Results evaluated on \\(\\mathrm{FF}++\\) are listed in Table 2. Our model shows comparable performance on all the metrics and achieves the lowest pose error when compared with SOTA face-swap methods. Particularly, our performances on ID-related metrics are only slightly weaker than StyleSwap." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.635, + 0.467, + 0.65 + ], + "angle": 0, + "content": "4.2 Qualitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.748 + ], + "angle": 0, + "content": "Subjective evaluation plays a pivotal role in assessing the capabilities of generative models, especially in the context of videos. We strongly encourage readers to watch our supplementary video for a comprehensive understanding. Our method also enjoys the capability to fast optimize for specific identities, resulting in superior outcomes. Please refer to the supplementary files for the personalized results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Lip-Sync. We demonstrate two cross-sync comparisons with SOTA methods in Fig. 4. The lip-synced results are conditioned on audio randomly sampled from the dataset. Wav2Lip has accurate lip shapes, but the mouth part is noticeably blurred. SadTalker achieves better fidelity, but it fails to present natural dynamics. VideoReTalking and StyleSync can generate precise and high-quality results; nevertheless, their results are affected by the opening jaw of the template" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "J. Guan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "as shown in the first column of the right figure. Moreover, they exhibit a deficiency in maintaining person-specific attributes around the mouth. Our method directly produces high-fidelity results with accurate lip motion and rich details of the mouth." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.209, + 0.789, + 0.331 + ], + "angle": 0, + "content": "We present cross-sync results, showcasing the transferred speaking style in Fig. 1 and Fig. 9 (in the supplementary files). It can be seen that our method not only accurately generates lip shapes aligned with the driven video but also successfully animates the target person with the speaking style of the source person in the driving video. In addition, our method inherently supports videodriven lip-syncing as introduced in Sec. 3.2. The comparison is demonstrated in Fig. 10 (in the supplementary files), it shows that person-specific attributes can be preserved better with Style-SyncFormer guidance." + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.359, + 0.731, + 0.627 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.687 + ], + "angle": 0, + "content": "Fig.6: Results of Face-Swapped Lip-Sync. The ID-swapped results are driven by the given audio. We compare with the combinations of two SOTA methods in lip-sync and face-swapping. Results generated by our method preserve details with better fidelity and simultaneously maintain a similar speaking style to the source." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Face-Swapping. In Fig. 5, we illustrate two face-swapped results for comparison. The generated results in this context should preserve the original expression and lip motion of the template. From the figure, InfoSwap fails to keep a natural brightness across frames. E4S effectively modifies the entire head region but fails to integrate the head part into the background seamlessly. SimSwap and StyleSwap produce good results by preserving precise expression and identity attributes. However, an excessive emphasis on identity feature matching causes them to falter in generating details around certain areas like eyes and mouth." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.666, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ReSyncer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 4: User study of generalized lip-sync. The scores are ranged from 1 (worst) to 5 (best)." + }, + { + "type": "table", + "bbox": [ + 0.276, + 0.185, + 0.723, + 0.253 + ], + "angle": 0, + "content": "
MOS onWav2LipReTalkingSadTalkerStyleSyncOurs-Sync
Lip-Sync Quality3.433.092.983.464.57
Generation Quality1.482.833.133.634.67
Video Realness1.592.412.873.114.54
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.784, + 0.344 + ], + "angle": 0, + "content": "In contrast, our method not only achieves effective identity transfer but also preserves rich details across the entire face. In addition, the videos generated by our method exhibit superior temporal consistency, especially in the mouth region. Please find more results in the supplementary video." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.346, + 0.785, + 0.42 + ], + "angle": 0, + "content": "Furthermore, our method is capable of achieving lip-synced face-swapping conditioned on any audio and identity in a single forward process. We compare our method with two potential baselines \"SimSwap + ReTalking\" and \"StyleSwap + StyleSync\" in Fig. 6. Results generated by our method preserve details with better fidelity and simultaneously maintain a similar speaking style to the source." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.784, + 0.573 + ], + "angle": 0, + "content": "User Study. We conducted a user study with 30 participants. We adopt the commonly used Mean Opinion Scores (MOS) rating protocol. All users give their ratings, ranging from 1 to 5. On generalized lip-sync, they are asked to evaluate: a) the quality of lip-sync. b) visual quality of the generated video, and c) whether this video looks real. The results averaged from 10 videos are grouped into Table 4. From the table, our method outperforms the four counterparts in all three aspects by clear margins, underscoring the power of our method and the substantial improvement in visual quality and realism over other existing methods. The experiments are similarly conducted with face-swapping and personalized lip-sync, please find the results in the supplementary files." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.599, + 0.389, + 0.615 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Lip-Sync. Here we evaluate four different designs on VoxCeleb2 [11] including: 1) \"w/o Mesh-Inject\": we mask-out the 3D mesh guidance from the encoded feature maps \\(\\{F_1, \\ldots, F_L\\}\\), thus it cannot provide detailed spatial information during generation; 2) \"w/o Mesh-Style\": 3D mesh content is removed in \\(I_t^{AM}\\) to get a \\(f_f\\), feature vector that feed into the \\(\\mathcal{W}\\) space contains only the background and reference information; 3) \"\\(T_S = 50\\)\": we reduce the length (\\(T_S\\) in Sec. 3.1) of the speaking-style reference from 90 to 50; 4) \"\\(T_S = 200\\)\": the length is extended to 200. The results are presented in Table 3. Comparisons between the first two rows and \"Ours-Sync\" reveal the vital role played by mesh guidance in the generation process, where the sync quality is degraded to certain degrees, as indicated by LMD and \\(\\Delta\\)Sync. In addition, the visual quality is also harmed as shown in Fig. 7 (a). In the bottom three rows of Table 3, we compare the performance of different sequence lengths in the speaking-style encoding of Style-SyncFormer. The results indicate that a sequence of 90 frames (\\(\\sim\\)4s of" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "J. Guan et al." + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.144, + 0.729, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.343, + 0.788, + 0.388 + ], + "angle": 0, + "content": "Fig. 7: Ablations. (a) 3D facial mesh provides detailed spatial guidance, empowering superior lip-syncing. (b) The mesh with shapes of facial organs also enhances ID transfer in face-swapping." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.785, + 0.445 + ], + "angle": 0, + "content": "video) leads to good sync quality, and a further extension of the length does not enhance its performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.788, + 0.568 + ], + "angle": 0, + "content": "Face-Swapping. In face-swapping, the predicted 3D facial mesh also provides detailed facial attributes. We provide an intuitive comparison in Fig. 7 (b). When the mesh guidance is removed from the encoded feature map (\"w/o Mesh-Inject\"), we observe a noteworthy degradation in facial dynamics, particularly in lip motion. Please refer to the supplementary video for the comparison. Simultaneously, without the guidance of mesh information, the ID-related metrics of face-swapping are also affected, as indicated at the bottom of Fig. 7 (b) and the same phenomenon applies to \"w/o Mesh-Style\"." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.59, + 0.509, + 0.607 + ], + "angle": 0, + "content": "5 Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.622, + 0.785, + 0.681 + ], + "angle": 0, + "content": "Ethical Considerations. Our method possesses the capability to create fabricated talks and speeches, even with swapped identities, presenting a potential for malicious use. We will restrict the distribution of our models and the generated results exclusively to research purposes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.683, + 0.788, + 0.818 + ], + "angle": 0, + "content": "Conclusion. We highlight several important properties of our ReSyncer framework: 1) Our easy reconfiguration further unveils the potential of widely studied structures so that high-quality generalized lip-sync results with mesh representation can be achieved. 2) Our framework is designed to adopt external identity information, thus we achieve face-swapping comparable to the state of the arts while keeping the lip-sync ability within one unified model. 3) Our network supports speaking style transfer, video-driven facial animation, and can be applied to real-time live streamers. These properties complementarily satisfy various needs of virtual performer creation under different circumstances." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.666, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ReSyncer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "Acknowledgments. This work is in part supported by National Natural Science Foundation of China with No. 62394322, Beijing Natural Science Foundation with No. L222024, as well as the Beijing National Research Center for Information Science and Technology (BNRist) key projects. This study is also supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.292, + 0.326, + 0.309 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.325, + 0.785, + 0.367 + ], + "angle": 0, + "content": "1. Baevski, A., Zhou, Y., Mohamed, A., Auli, M.: wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in Neural Information Processing Systems (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.368, + 0.785, + 0.409 + ], + "angle": 0, + "content": "2. Bao, L., Lin, X., Chen, Y., Zhang, H., Wang, S., Zhe, X., Kang, D., Huang, H., Jiang, X., Wang, J., Yu, D., Zhang, Z.: High-fidelity 3d digital human head creation from rgb-d selfies. ACM Transactions on Graphics (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.41, + 0.785, + 0.45 + ], + "angle": 0, + "content": "3. Blanz, V., Vetter, T.: A morphable model for the synthesis of 3d faces. In: Proceedings of the 26th annual conference on Computer graphics and interactive techniques. pp. 187-194 (1999)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.451, + 0.785, + 0.506 + ], + "angle": 0, + "content": "4. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.507, + 0.785, + 0.534 + ], + "angle": 0, + "content": "5. Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: Maskgit: Masked generative image transformer. arXiv preprint arXiv:2202.04200 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.535, + 0.785, + 0.576 + ], + "angle": 0, + "content": "6. Chen, L., Cui, G., Liu, C., Li, Z., Kou, Z., Xu, Y., Xu, C.: Talking-head generation with rhythmic head motion. In: European Conference on Computer Vision. pp. 35-51. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.577, + 0.785, + 0.617 + ], + "angle": 0, + "content": "7. Chen, L., Li, Z., Maddox, R.K., Duan, Z., Xu, C.: Lip movements generation at a glance. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 520-535 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.618, + 0.785, + 0.659 + ], + "angle": 0, + "content": "8. Chen, L., Maddox, R.K., Duan, Z., Xu, C.: Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7832-7841 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.66, + 0.785, + 0.701 + ], + "angle": 0, + "content": "9. Chen, R., Chen, X., Ni, B., Ge, Y.: Simswap: An efficient framework for high fidelity face swapping. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 2003-2011 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.785, + 0.742 + ], + "angle": 0, + "content": "0. Cheng, K., Cun, X., Zhang, Y., Xia, M., Yin, F., Zhu, M., Wang, X., Wang, J., Wang, N.: Videoretalking: Audio-based lip synchronization for talking head video editing in the wild (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.743, + 0.785, + 0.771 + ], + "angle": 0, + "content": "1. Chung, J.S., Nagrani, A., Zisserman, A.: Voxceleb2: Deep speaker recognition. In: INTERSPEECH (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "12. Chung, J.S., Zisserman, A.: Out of time: automated lip sync in the wild. In: ACCV (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "13. Cudeiro, D., Bolkart, T., Laidlaw, C., Ranjan, A., Black, M.J.: Capture, learning, and synthesis of 3d speaking styles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10101-10111 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.325, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "J. Guan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "14. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4690-4699 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.243 + ], + "angle": 0, + "content": "15. Deng, Y., Yang, J., Chen, D., Wen, F., Tong, X.: Disentangled and controllable face image generation via 3d imitative-contrastive learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.243, + 0.785, + 0.285 + ], + "angle": 0, + "content": "16. Deng, Y., Yang, J., Xiang, J., Tong, X.: Gram: Generative radiance manifolds for 3d-aware image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10673-10683 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.785, + 0.338 + ], + "angle": 0, + "content": "17. Deng, Y., Yang, J., Xu, S., Chen, D., Jia, Y., Tong, X.: Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.785, + 0.379 + ], + "angle": 0, + "content": "18. Fan, Y., Lin, Z., Saito, J., Wang, W., Komura, T.: Faceformer: Speech-driven 3d facial animation with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18770-18780 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.434 + ], + "angle": 0, + "content": "19. Gafni, G., Thies, J., Zollhöfer, M., Nießner, M.: Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8649-8658 (June 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.785, + 0.474 + ], + "angle": 0, + "content": "20. Gao, G., Huang, H., Fu, C., Li, Z., He, R.: Information bottleneck disentanglement for identity swapping. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3404-3413 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.785, + 0.514 + ], + "angle": 0, + "content": "21. Goodfellow, I.J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. arXiv preprint arXiv:1406.2661 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.785, + 0.569 + ], + "angle": 0, + "content": "22. Guan, J., Zhang, Z., Zhou, H., Hu, T., Wang, K., He, D., Feng, H., Liu, J., Ding, E., Liu, Z., Wang, J.: Stylesync: High-fidelity generalized and personalized lip sync in style-based generator. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.785, + 0.61 + ], + "angle": 0, + "content": "23. Guo, Y., Chen, K., Liang, S., Liu, Y., Bao, H., Zhang, J.: Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.785, + 0.651 + ], + "angle": 0, + "content": "24. Jamaludin, A., Chung, J.S., Zisserman, A.: You said that?: Synthesising talking faces from audio. International Journal of Computer Vision 127(11), 1767-1779 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.785, + 0.692 + ], + "angle": 0, + "content": "25. Ji, X., Zhou, H., Wang, K., Wu, Q., Wu, W., Xu, F., Cao, X.: Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. SIGGRAPH (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.692, + 0.785, + 0.732 + ], + "angle": 0, + "content": "26. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4401-4410 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.785, + 0.759 + ], + "angle": 0, + "content": "27. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of StyleGAN. In: Proc. CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.8 + ], + "angle": 0, + "content": "28. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8110-8119 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "29. Li, L., Wang, S., Zhang, Z., Ding, Y., Zheng, Y., Yu, X., Fan, C.: Write-a-speaker: Text-based emotional and rhythmic talking-head generation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1911-1920 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.666, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ReSyncer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "30. Li, L., Bao, J., Yang, H., Chen, D., Wen, F.: Faceshifter: Towards high fidelity and occlusion aware face swapping. arXiv preprint arXiv:1912.13457 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.787, + 0.215 + ], + "angle": 0, + "content": "31. Li, Y., Ma, C., Yan, Y., Zhu, W., Yang, X.: 3d-aware face swapping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12705-12714 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.216, + 0.785, + 0.27 + ], + "angle": 0, + "content": "32. Liang, B., Pan, Y., Guo, Z., Zhou, H., Hong, Z., Han, X., Han, J., Liu, J., Ding, E., Wang, J.: Expressive talking head generation with granular audio-visual control. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3387-3396 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.271, + 0.785, + 0.297 + ], + "angle": 0, + "content": "33. Liu, X., Xu, Y., Wu, Q., Zhou, H., Wu, W., Zhou, B.: Semantic-aware implicit neural audio-driven video portrait generation. ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.297, + 0.785, + 0.338 + ], + "angle": 0, + "content": "34. Liu, Z., Li, M., Zhang, Y., Wang, C., Zhang, Q., Wang, J., Nie, Y.: Fine-grained face swapping via regional gan inversion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8578-8587 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.338, + 0.785, + 0.379 + ], + "angle": 0, + "content": "35. Luo, Y., Zhu, J., He, K., Chu, W., Tai, Y., Wang, C., Yan, J.: Styleface: Towards identity-disentangled face generation on megapixels. In: European Conference on Computer Vision. pp. 297-312. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.379, + 0.785, + 0.42 + ], + "angle": 0, + "content": "36. Ma, Y., Wang, S., Ding, Y., Ma, B., Lv, T., Fan, C., Hu, Z., Deng, Z., Yu, X.: Talkclip: Talking head generation with text-guided expressive speaking styles. arXiv preprint arXiv:2304.00334 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.42, + 0.785, + 0.446 + ], + "angle": 0, + "content": "37. Ma, Y., Wang, S., Hu, Z., Fan, C., Lv, T., Ding, Y., Deng, Z., Yu, X.: Styletalk: One-shot talking head generation with controllable speaking styles. AAAI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.446, + 0.785, + 0.486 + ], + "angle": 0, + "content": "38. Natsume, R., Yatagawa, T., Morishima, S.: Rsgan: Face swapping and editing using face and hair representation in latent spaces. arxiv 2018. arXiv preprint arXiv:1804.03447 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.486, + 0.785, + 0.542 + ], + "angle": 0, + "content": "39. Natsume, R., Yatagawa, T., Morishima, S.: Fsnet: An identity-aware generative model for image-based face swapping. In: Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part VI 14. pp. 117–132. Springer (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.542, + 0.785, + 0.595 + ], + "angle": 0, + "content": "40. Park, S.J., Kim, M., Hong, J., Choi, J., Ro, Y.M.: Synctalkface: Talking face generation with precise lip-syncing via audio-lip memory. In: AAAI Conference on Artificial Intelligence. Association for the Advancement of Artificial Intelligence (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.596, + 0.785, + 0.637 + ], + "angle": 0, + "content": "41. Prajwal, K., Mukhopadhyay, R., Namboodiri, V.P., Jawahar, C.: A lip sync expert is all you need for speech to lip generation in the wild. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 484-492 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.637, + 0.785, + 0.678 + ], + "angle": 0, + "content": "42. Qiu, H., Chen, Z., Jiang, Y., Zhou, H., Fan, X., Yang, L., Wu, W., Liu, Z.: Relitalk: Relightable talking portrait generation from a single video. International Journal of Computer Vision pp. 1-16 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.678, + 0.785, + 0.718 + ], + "angle": 0, + "content": "43. Richard, A., Zollhöfer, M., Wen, Y., de la Torre, F., Sheikh, Y.: Meshtalk: 3d face animation from speech using cross-modality disentanglement. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "44. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.759, + 0.785, + 0.8 + ], + "angle": 0, + "content": "45. Rössler, A., Cozzolino, D., Verdoliva, L., Riess, C., Thies, J., Nießner, M.: FaceForensics++: Learning to detect manipulated facial images. In: International Conference on Computer Vision (ICCV) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "46. Ruiz, N., Chong, E., Rehg, J.M.: Fine-grained head pose estimation without keypoints. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 2074-2083 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.128 + ], + "angle": 0, + "content": "J. Guan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "47. Shen, S., Li, W., Zhu, Z., Duan, Y., Zhou, J., Lu, J.: Learning dynamic facial radiance fields for few-shot talking head synthesis. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.787, + 0.232 + ], + "angle": 0, + "content": "48. Shen, S., Zhao, W., Meng, Z., Li, W., Zhu, Z., Zhou, J., Lu, J.: Difftalk: Crafting diffusion models for generalized audio-driven portraits animation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1982-1991 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.785, + 0.26 + ], + "angle": 0, + "content": "49. Song, Y., Zhu, J., Li, D., Wang, X., Qi, H.: Talking face generation by conditional recurrent adversarial network. arXiv preprint arXiv:1804.04786 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.262, + 0.785, + 0.302 + ], + "angle": 0, + "content": "50. Stypulkowski, M., Vougioukas, K., He, S., Zieba, M., Petridis, S., Pantic, M.: Diffused heads: Diffusion models beat gans on talking-face generation. arXiv preprint arXiv:2301.03396 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.785, + 0.346 + ], + "angle": 0, + "content": "51. Sun, Y., Zhou, H., Wang, K., Wu, Q., Hong, Z., Liu, J., Ding, E., Wang, J., Liu, Z., Hideki, K.: Masked lip-sync prediction by audio-visual contextual exploitation in transformers. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.347, + 0.785, + 0.387 + ], + "angle": 0, + "content": "52. Suwajanakorn, S., Seitz, S.M., Kemelmacher-Shlizerman, I.: Synthesizing obama: learning lip sync from audio. ACM Transactions on Graphics (ToG) 36(4), 1-13 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.389, + 0.785, + 0.43 + ], + "angle": 0, + "content": "53. Tang, J., Wang, K., Zhou, H., Chen, X., He, D., Hu, T., Liu, J., Zeng, G., Wang, J.: Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.432, + 0.785, + 0.472 + ], + "angle": 0, + "content": "54. Thambiraja, B., Habibie, I., Aliakbarian, S., Cosker, D., Theobalt, C., Thies, J.: Imitator: Personalized speech-driven 3d facial animation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 20621-20631 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.785, + 0.515 + ], + "angle": 0, + "content": "55. Thies, J., Elgharib, M., Tewari, A., Theobalt, C., Nießner, M.: Neural voice puppetry: Audio-driven facial reenactment. In: European Conference on Computer Vision. pp. 716-731. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.517, + 0.785, + 0.543 + ], + "angle": 0, + "content": "56. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.545, + 0.785, + 0.585 + ], + "angle": 0, + "content": "57. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.785, + 0.628 + ], + "angle": 0, + "content": "58. Vemulapalli, R., Agarwala, A.: A compact embedding for facial expression similarity. In: proceedings of the IEEE/cvf conference on computer vision and pattern recognition. pp. 5683-5692 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.63, + 0.785, + 0.669 + ], + "angle": 0, + "content": "59. Wang, D., Deng, Y., Yin, Z., Shum, H.Y., Wang, B.: Progressive disentangled representation learning for fine-grained controllable talking head synthesis. CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.672, + 0.785, + 0.713 + ], + "angle": 0, + "content": "60. Wang, H., Wang, Y., Zhou, Z., Ji, X., Gong, D., Zhou, J., Li, Z., Liu, W.: Cosface: Large margin cosine loss for deep face recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5265-5274 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.715, + 0.785, + 0.755 + ], + "angle": 0, + "content": "61. Wang, J., Zhao, K., Ma, Y., Zhang, S., Zhang, Y., Shen, Y., Zhao, D., Zhou, J.: Facecomposer: A unified model for versatile facial content creation. In: Thirty-seventh Conference on Neural Information Processing Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.785, + 0.797 + ], + "angle": 0, + "content": "62. Wang, K., Wu, Q., Song, L., Yang, Z., Wu, W., Qian, C., He, R., Qiao, Y., Loy, C.C.: Mead: A large-scale audio-visual dataset for emotional talking-face generation. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "63. Wang, S., Li, L., Ding, Y., Fan, C., Yu, X.: Audio2head: Audio-driven one-shot talking-head generation with natural head motion. arXiv preprint arXiv:2107.09293 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.667, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ReSyncer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "64. Wang, S., Li, L., Ding, Y., Yu, X.: One-shot talking face generation from single-speaker audio-visual correlation learning. In: Proceedings of the AAAI Conference on Artificial Intelligence (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.245 + ], + "angle": 0, + "content": "65. Wang, T.C., Liu, M.Y., Zhu, J.Y., Tao, A., Kautz, J., Catanzaro, B.: High-resolution image synthesis and semantic manipulation with conditional gans. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8798-8807 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.245, + 0.788, + 0.286 + ], + "angle": 0, + "content": "66. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.328 + ], + "angle": 0, + "content": "67. Wu, H., Jia, J., Wang, H., Dou, Y., Duan, C., Deng, Q.: Imitating arbitrary talking style for realistic audio-driven talking face synthesis. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 1478-1486 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.383 + ], + "angle": 0, + "content": "68. Wu, X., Hu, P., Wu, Y., Lyu, X., Cao, Y.P., Shan, Y., Yang, W., Sun, Z., Qi, X.: Speech2lip: High-fidelity speech to lip generation by learning from a short video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22168-22177 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.383, + 0.788, + 0.438 + ], + "angle": 0, + "content": "69. Xing, J., Xia, M., Zhang, Y., Cun, X., Wang, J., Wong, T.T.: Codetalker: Speechdriven 3d facial animation with discrete motion prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12780-12790 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.438, + 0.788, + 0.481 + ], + "angle": 0, + "content": "70. Xu, C., Zhang, J., Hua, M., He, Q., Yi, Z., Liu, Y.: Region-aware face swapping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7632-7641 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.481, + 0.788, + 0.522 + ], + "angle": 0, + "content": "71. Xu, Y., Deng, B., Wang, J., Jing, Y., Pan, J., He, S.: High-resolution face swapping via latent semantics disentanglement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7642-7651 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.563 + ], + "angle": 0, + "content": "72. Xu, Z., Hong, Z., Ding, C., Zhu, Z., Han, J., Liu, J., Ding, E.: Mobilefaceswap: A lightweight framework for video face swapping. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 2973-2981 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.563, + 0.788, + 0.605 + ], + "angle": 0, + "content": "73. Xu, Z., Zhou, H., Hong, Z., Liu, Z., Liu, J., Guo, Z., Han, J., Liu, J., Ding, E., Wang, J.: Styleswap: Style-based generator empowers robust face swapping. In: European Conference on Computer Vision. pp. 661-677. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.605, + 0.788, + 0.647 + ], + "angle": 0, + "content": "74. Yang, T., Ren, P., Xie, X., Zhang, L.: Gan prior embedded network for blind face restoration in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 672-681 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.647, + 0.788, + 0.688 + ], + "angle": 0, + "content": "75. Yao, S., Zhong, R., Yan, Y., Zhai, G., Yang, X.: Dfa-nerf: Personalized talking head generation via disentangled face attributes neural rendering. arXiv preprint arXiv:2201.00791 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.688, + 0.788, + 0.716 + ], + "angle": 0, + "content": "76. Ye, Z., Jiang, Z., Ren, Y., Liu, J., He, J., Zhao, Z.: Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.716, + 0.788, + 0.743 + ], + "angle": 0, + "content": "77. Yu, Z., Yin, Z., Zhou, D., Wang, D., Wong, F., Wang, B.: Talking head generation with probabilistic audio-to-visual diffusion priors. ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.743, + 0.788, + 0.799 + ], + "angle": 0, + "content": "78. Zhang, W., Cun, X., Wang, X., Zhang, Y., Shen, X., Guo, Y., Shan, Y., Wang, F.: Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8652-8661 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.788, + 0.841 + ], + "angle": 0, + "content": "79. Zhang, Z., Li, L., Ding, Y., Fan, C.: Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3661-3670 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "J. Guan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.189 + ], + "angle": 0, + "content": "80. Zhou, H., Liu, Y., Liu, Z., Luo, P., Wang, X.: Talking face generation by adversarially disentangled audio-visual representation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 33, pp. 9299-9306 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "81. Zhou, H., Sun, Y., Wu, W., Loy, C.C., Wang, X., Liu, Z.: Pose-controllable talking face generation by implicitly modularized audio-visual representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4176-4186 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.786, + 0.286 + ], + "angle": 0, + "content": "82. Zhou, Y., Han, X., Shechtman, E., Echevarria, J., Kalogerakis, E., Li, D.: Makelttalk: speaker-aware talking-head animation. ACM Transactions on Graphics (TOG) 39(6), 1-15 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.786, + 0.328 + ], + "angle": 0, + "content": "83. Zhu, Y., Li, Q., Wang, J., Xu, C.Z., Sun, Z.: One shot face swapping on megapixels. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4834-4844 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.328 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/ReSyncer_ Rewiring Style-based Generator for Unified Audio-Visually Synced Facial Performer/6fe8e31e-41da-4533-8af1-187e2be9d843_origin.pdf b/2024/ReSyncer_ Rewiring Style-based Generator for Unified Audio-Visually Synced Facial Performer/6fe8e31e-41da-4533-8af1-187e2be9d843_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bd8ca6d4f8cf9b4fd189a9b5d9a7de5982198a38 --- /dev/null +++ b/2024/ReSyncer_ Rewiring Style-based Generator for Unified Audio-Visually Synced Facial Performer/6fe8e31e-41da-4533-8af1-187e2be9d843_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:656ae8a0bf755ef1d4c35538458d9e6bbae9d440ceb8608085a3439fed424bbc +size 9711129 diff --git a/2024/ReSyncer_ Rewiring Style-based Generator for Unified Audio-Visually Synced Facial Performer/full.md b/2024/ReSyncer_ Rewiring Style-based Generator for Unified Audio-Visually Synced Facial Performer/full.md new file mode 100644 index 0000000000000000000000000000000000000000..805ae38601a98c7038cc899c441b59daecae85dc --- /dev/null +++ b/2024/ReSyncer_ Rewiring Style-based Generator for Unified Audio-Visually Synced Facial Performer/full.md @@ -0,0 +1,304 @@ +# ReSyncer: Rewiring Style-based Generator for Unified Audio-Visually Synced Facial Performer + +Jiazhi Guan $^{1,2*}$ , Zhiliang Xu $^{2*}$ , Hang Zhou $^{2\dagger}$ , Kaisiyuan Wang $^{2}$ , Shengyi He $^{2}$ , Zhanwang Zhang $^{2}$ , Borong Liang $^{2}$ , Haocheng Feng $^{2}$ , Errui Ding $^{2}$ , Jingtuo Liu $^{2}$ , Jingdong Wang $^{2}$ , Youjian Zhao $^{1,3\dagger}$ , and Ziwei Liu $^{4}$ + +$^{1}$ BNRist, DCST, Tsinghua University $^{2}$ Baidu Inc. + $^{3}$ Zhongguancun Laboratory $^{4}$ S-Lab, Nanyang Technological University guanjz20@mails.tsinghua.edu.cn, zhouhang09@baidu.com + +![](images/7006e089d5e2a453a0432cdf8dda770d32efa20793899df00e4c54f055d1e327.jpg) +Fig. 1: Lip-Synced/Speaking-Style-Transferred/Face-Swapping Results by ReSyncer. Our method not only produces high-fidelity lip-synced video according to audio but can further transfer the speaking style and identity of any target person. + +Abstract. Lip-syncing videos with given audio is the foundation for various applications including the creation of virtual presenters or performers. While recent studies explore high-fidelity lip-sync with different techniques, their task-orientated models either require long-term videos for clip-specific training or retain visible artifacts. In this paper, we propose a unified and effective framework ReSyncer, that synchronizes generalized audio-visual facial information. The key design is revisiting and rewiring the Style-based generator to efficiently adopt 3D facial dynamics predicted by a principled style-injected Transformer. By simply re-configuring the information insertion mechanisms within the noise and style space, our framework fuses motion and appearance with unified training. Extensive experiments demonstrate that ReSyncer not + +only produces high-fidelity lip-synced videos according to audio, but also supports multiple appealing properties that are suitable for creating virtual presenters and performers, including fast personalized fine-tuning, video-driven lip-syncing, the transfer of speaking styles, and even face swapping. + +# 1 Introduction + +Generating lip-synced facial videos with speech audio is a foundational technique for various applications including the dubbing of movies and the creation of virtual presenters or performers in the form of realistic humans. The needs in virtual performers' creation can hardly be comprehensively met. + +While great attention has been paid to creating dynamic lip-synced videos from a few photos [8, 24, 63, 64, 78, 80-82], previous results are still far from realistic. To generate life-like videos, researchers have made explorations [41] with GANs [10,55], NeRF [23,33,47,53], and Diffusion models [48] to edit existing videos. Despite their efforts, most methods rely on a relatively long-term clip of video for training, making them less flexible and efficient. A few methods [10,22, 41,51] enable generalized local editing without training. However, they still tend to create visible flaws on high-quality videos [41,51] Particularly, StyleSync [22] supports not only arbitrary video modeling but can adapt to the speaking style of a target person with short-clip fine-tuning, which balances the generation quality and efficiency. Nevertheless, their results can still be easily affected by mouth movements on their conditional reference frames, leading to unstable results. + +The previous cases reveal the limitations of using low-dimensional audio information to directly modify high-dimensional visual data. On the other hand, intermediate representations like 3D facial models are widely used, but their potential as spatial guidance for efficient, generalized editing remains underexplored. Using only 3DMM [3] coefficients [15,16,19,29,37,55] also suffers from the cross-domain information injection problem. While textured rendering offers a promising alternative [6,55,67], its implementation complexity and susceptibility to errors during 3D modeling or texture fitting can lead to degraded results. + +In this paper, we revisit and re-wire the powerful Style-based generator [26, 27] to generate faithful lip-synced videos and support virtual performer creation in a concise yet effective framework called ReSyncer. With unified model training, various appealing properties including generalized and personalized lip-sync, speaking style transfer, live streaming, and face swapping can be achieved within one model. Our key is to show how an easy reconfiguration empowers Style-based generator with various powerful audio-visual synchronization abilities by involving 3D facial dynamics. Specifically, we use only roughly fitted 3D facial meshes, instead of coefficients to bridge the audio and image domains. A principled Style-injected Lip-Sync Transformer (Style-SyncFormer) learns stylized point cloud displacements given template meshes and driving audios. + +We then perform easy but crucial modifications to Style-based generator with a simple encoder used in previous work [22, 73, 74]. Concretely, while lip-sync + +studies [10,22,41,51,61] mask out the mouth area of the target frame, we overlay reconstructed mesh onto it as strong spatial guidance. Then we compliment the textural and identity information with an attached reference frame, which serves as the source of both the noise and style-related space. With generalized training on massive data, the powerful generator focuses on recovering facial details from the roughly predicted lip-synced white-colored meshes, leading to high-fidelity and stable lip-sync with higher quality. + +Additionally, our framework design goes beyond existing solutions by uniformly addressing a broader range of needs in virtual performer creation. These needs include, for instance, preserving personal styles for specific users or changing face identities for privacy protection and diverse character generation. Particularly, our method directly supports few-shot few-step personalized fine-tuning. Moreover, our framework supports *syncing* (swapping) different identities on the same target video with additional training strategies. High-fidelity face-swapping and lip-syncing can hence be simultaneously satisfied. + +Our contributions can be summarized as follows: 1) We propose the ReSyncer framework, which shows the powerfulness of Style-based generator on syncing audio-visual facial information by involving 3D facial meshes with a simple reconfiguration. 2) We propose the Style-SyncFormer which learns stylized 3D facial dynamics with simple Transformer blocks, enabling generalized 3D facial animation. 3) Extensive experiments show that our framework not only enables lip-syncing of higher stability and quality, but also supports various intriguing properties that are essential for creating virtual performers, including fast personalized fine-tuning, video-driven lip-syncing, the transfer of speaking styles, and even face swapping. To the best of our knowledge, we are one of the first methods that support state-of-the-art lip-syncing and face-swapping with one unified model. + +# 2 Related Work + +# 2.1 Audio-Driven Facial Animation + +Talking Head Generation. A great number of studies focus on predicting the whole human head movements according to audios [7, 8, 23-25, 32, 36, 37, 42, 52, 59, 62, 63, 77, 78, 81]. More recently, one-shot-based methods adopt representations from 3D, videos, and even texts for more controllable results. While other approaches [23, 33, 47, 53, 75, 76] attempt to integrate speech audios into the neural radiance field for personalized audio-to-3D modeling. However, these methods usually fail to achieve stable lip-sync results and show terrible robustness performance when using audio in the wild. Moreover, the form of talking head generation cannot be applied to edit arbitrary video. + +Lip-Syncing-based Facial Editing. Apart from generating the entire head, certain studies [10, 22, 40, 41, 49, 51, 55, 68] choose to directly edit the mouth shape according to the input audios. Specifically, [41] relies on a well-trained lip-sync discriminator to produce highly synchronized animations. [51] introduces + +an audio-visual cross-modality framework based on transformer [5]. Particularly, [22] builds a person-agnostic lip-sync framework by StyleGAN2 [27], which also supports personalized optimization. Similarly, [10] also provides identity refinement for the final animation results. Although they have shown accurate lip-sync and comparably high-fidelity results, visible artifacts may still exist due to the latent contradiction between textural and spatial information. Few studies [48, 50] built upon diffusion models [44] are presented recently. Our ReSyncer generates high-fidelity lip-sync videos with high stability and efficiency. + +Audio-Driven 3D Facial Animation. The first stage of our work shares a similar setting with methods that generate audio-driven 3D facial animation [13, 18,43,54,69]. Most of them are performed on 3D scan data to predict the point clouds' dynamics. They have explored Transformers [57] and codebooks [56] with different settings on this task. We adopt a highly efficient form to use Transformer in our work. + +# 2.2 Face Swapping + +To transfer the identity of one person to the face of another person is a longstanding problem in computer vision. In recent years, GAN [21] has played an important role in this task. RSGAN [38] and FSNet [39] encode the facial region and the non-facial region to different latent codes and perform swapping by replacing these latent representations. [83] proposes a face-swapping method MegaFS built upon the GAN-inversion technique with StyleGAN2, however, the results are not temporal consistent. Similarly, [71, 73] make their attempts to apply the strong prior learned by StyleGAN2 [27] to generate high-resolution swapped faces. Recently, [31] takes inspiration from 3D generative model [4] to capture fine-grained details of face shape and strengthen the robustness under large pose variance by extending the face swapping task into 3D latent space. However, they fail to produce results with accurate poses and expressions when balancing identity similarity and attribution preservation of the target frame. + +Different from these methods, ReSyncer takes advantage of intermediate guidance from 3D mesh to synthesize accurate pose and expression. Moreover, we achieve face-swapping in a compatible with accurate lip-syncing, which allows us to further satisfy more imaginative demands in creating virtual presenters and performers. + +# 3 Methodology + +In this section, we introduce the details of our ReSyncer framework as depicted in Fig. 2. It consists of two stages which are all simply designed and easy to implement. In section 3.1 we illustrate our design of the style-injected lip-sync Transformer (Style-SyncFormer) which predicts 3D facial dynamics. Then we show how to reconfigure the Style-based generator to support high-fidelity generalized and personalized lip-sync in section 3.2. Finally, we illustrate our unified design naturally supports face-swapping training in section 3.3. + +![](images/434071d6863b2949c47f490e6a9ce740f7003eb00ea42586851c7f12e4e6a63b.jpg) +Fig. 2: The ReSyncer Framework. In the first stage, the Style-SyncFormer takes the style template and audio input to predict 3D facial dynamics. Then the predicted mesh overlays on the target frame to provide strong spatial guidance. The Style-based generator G processes the overlay and reference frames to produce the final result. + +Task Formulation. The task of generalized lip-sync is normally trained on massive in-the-wild videos. Specifically, a training data consists of video clip $\mathbf{V} = \{I_1,\dots ,I_T\}$ with $T$ frames and the accompanying (processed) audio clip features $\mathbf{a} = \{a_{1},\ldots ,a_{T}\}$ . The lower half of the target frame $I_{t}$ is masked out by a mask $A$ to form $I_t^A = (\mathbf{1} - A)*I_t$ as [10,22,41,51]. The normal training objective is to recover the original frame given the audio clip $a_{t}$ and a reference frame $I_{ref}$ . However, the final results might be interfered with by the structure of the reference frame when the estimated animation significantly differs from the reference. + +In this paper, we involve roughly predicted 3D facial meshes as intermediate representations. For every frame in the training video, we reconstruct the 3D facial meshes $\mathbf{M} = \{M_1,\dots ,M_T\}$ . Thus our framework can be divided into the stages of predicting 3D facial dynamics from audio and the recovery of frames $I_{t}$ with the assistance of $M_{t}$ . + +# 3.1 Style-Injected Lip-Sync Transformer + +In the first stage, we propose the Style-SyncFormer that takes audio features $\mathbf{a}_{t_0:t_1} = \{a_{t_0},\dots ,a_{t_1}\}$ extracted from pretrained Wav2Vec2 [1] to predict 3D meshes $\mathbf{M}_{t_0:t_1} = \{M_{t_0},\dots ,M_{t_1}\}$ . There are $N$ consecutive frames between $t_0$ and $t_1$ . Previous work [18] has explored abilities to learn 3D dynamics but it handles only a fixed number of identities. In contrast, the proposed Style-SyncFormer facilitates zero-shot prediction of 3D dynamics while preserving the original speaking style or transferring it from another person. + +Data Processing. During training, all pose information is eliminated from the meshes and aligned to the canonical space. Instead of directly learning all vertices' positions, we learn the displacements compared to a template mesh $\bar{M}$ , the mean of all meshes from this person without expression coefficients. The training target becomes predicting the displacements $\Delta \mathbf{M}_{t_0:t_1} = \{\Delta M_{t_0},\dots \Delta M_{t_1}\}$ between the template and each reconstructed result. + +Network Design. The main network of Style-SyncFormer is built based on only one Transformer [57] encoder block for style information encoding and one decoder block for mesh prediction, where we adopt the periodic positional encoding and biased causal multi-head attention from FaceFormer [18]. + +During training, each shifted-right displacement $\Delta M_{t}$ is mapped to embedding $d_t^m$ and sent into the positional encoding to serve as tokens $f_{t}^{m}$ . Previous studies [18, 43] learn the 3D dynamics of only limited identities with a simple one-hot speaker identity embedding [13, 18]. Differently, we encode the information of a total of $T_{S} = 90$ consecutive $\Delta M_{ref}$ as speaking style reference. The style information is embedded as a feature $f_{S}^{r}$ from $\Delta M_{ref}$ using an additional Transformer encoder block. After that, we inject the speaking style $f_{S}^{r}$ by adding it with mesh tokens $f_{t}^{m}$ . Note that when replacing $\Delta M_{ref}$ with mesh sequences of another identity, speaking style transfer can be achieved. + +In the decoder block, the input tokens are sent into multi-head self-attention, cross-attention, and feed-forward layers. The audio features are further encoded as embeddings $f_{t}^{a}$ and leveraged in the cross-attentions. The final output is transformed to $\Delta \hat{M}_{t_0:t_1} = \{\Delta \hat{M}_{t_0},\dots ,\Delta \hat{M}_{t_1}\}$ . + +Loss Function. The training objectives can be written as: + +$$ +\mathcal {L} _ {3 D} = \sum_ {t = t _ {0} + 1} ^ {t _ {1}} \| (\varDelta \hat {M} _ {t} - \varDelta \hat {M} _ {t - 1}) - (\varDelta M _ {t} - \varDelta M _ {t - 1}) \| _ {2} + \lambda_ {1} \sum_ {t = t _ {0}} ^ {t _ {1}} \| \varDelta \hat {M} _ {t} - \varDelta M _ {t} \| _ {2}, +$$ + +where the first term is the temporal consistency loss [55]. + +The training is performed in a teacher-forcing manner while the network works autoregressively during inference by using previously predicted results. The displacements are finally transferred to 3D meshes $\hat{\mathbf{M}}_{t_0:t_1}$ in the canonical space by adding the template mesh $\bar{M}$ . + +# 3.2 Rewiring Style-based Generator + +In the second stage, we rewire the widely used Style-based generator [27] to faithfully transfer 3D meshes $\mathbf{M}$ to high-fidelity facial frames $\hat{\mathbf{I}}$ in a simple manner. Previous methods [6,55] resort to accurately textured rendering, which is inefficient from the perspective of implementation and training. Our goal is to find an efficient and effective way. As a result, we leverage the basic white-colored rendering as spatial guidance, which can be easily achieved without textural mapping. + +Network Design. We resort to Style-based generator with noise insertion scheme, which is easy to implement and specializes well in face-related tasks [22, 35,73,74]. Specifically, we project mesh $M_{t}$ on masked target frame $I_{t}^{A}$ 's domain to $M_{t}^{p}$ and overlay to form a 3D-guided image $I_{t}^{AM} = I_{t}^{A} + A * M_{t}^{p}$ . A reference frame $I_{ref}$ is attached for providing textural information in the spatial form. The 6-channel input $[I_{t}^{AM}, I_{ref}]$ is encoded to feature maps $\mathbf{F}_{f} = \{f_{f}, F_{1}, \ldots, F_{L}\}$ that match the sizes of Style-based generator's each stage and $f_{f}$ being a vector. Then we follow previous studies [22,73,74] and insert $\mathbf{F}_{f}$ into the generator $G$ . + +As $I_{t}^{AM}$ directly provides the 3D mouth shape information, the reference frame can hardly influence the prediction result as long as it is not the target frame itself. Thus we randomly sample $I_{ref}$ within the nearest 30 frames of $I_{t}$ . + +A number of previous lip-sync studies [22,32,59,77,81] rely on encoding audio information into the $\mathcal{W}$ space for animation. However, with explicit guidance, the driving information would play a less important role in the $\mathcal{W}$ space. Meanwhile, the face-swapping task illustrates that $\mathcal{W}$ space embedding benefits identity information [35, 73, 83] which we intend to enhance. To avoid involving extra network design and inference procedures, we directly send the encoded $f_{f}$ from $[I_t^{AM}, I_{ref}]$ to the $\mathcal{W}$ , so that the reference frame provides more facial detail while $I_t^{AM}$ also passes dynamic information. + +Loss Functions. To sum up, in the second stage, the generator takes the output only of a single facial encoder and predicts the target frame itself as $\hat{I}_t$ . The training objective is much simpler with the 3D guidance. We only adopt the same loss functions as Pix2pix-HD [65], which are the $L_1$ , VGG, and adversarial loss empowered with StyleGAN2's discriminator. + +$$ +\mathcal {L} _ {r e c} = \left\| \hat {I} _ {t} - I _ {t} \right\| _ {1} + \lambda_ {2} \sum_ {m = 1} ^ {N _ {v g g}} \left\| \operatorname {V G G} _ {m} (\hat {I} _ {t}) - \operatorname {V G G} _ {m} (I _ {t}) \right\| _ {1}, \tag {1} +$$ + +$$ +\mathcal {L} _ {a d v} = \underset {\mathrm {G}} {\operatorname * {m i n m a x}} _ {\mathrm {D}} \left(\mathbb {E} _ {I _ {t}} \left[ \log \mathrm {D} \left(I _ {t}\right) \right] + \mathbb {E} _ {\hat {I} _ {t}} \left[ \log \left(1 - \mathrm {D} \left(\hat {I} _ {t}\right)\right) \right]\right). \tag {2} +$$ + +As the rendered meshes provide much clearer guidance than audio, the generator can be easily trained without task-oriented losses or specific tuning. + +During inference, we overlay the predicted mesh $\hat{M}_t^p$ from the first stage to form the input $[\hat{I}_t^{AM}, I_{ref}]$ . Though the predicted $\hat{M}_t^p$ cannot perfectly match the reconstructed results, we rely on the powerful generator to compensate for such disturbance. + +Video-Driven Lip-Sync. The generator only depends on mesh information, thus we can transfer the displacements of meshes $\Delta \mathbf{M}'$ of another person to the target template $\bar{M}$ so that video-driven lip-sync can be naturally achieved. + +Personalized Fine-tuning. Given thatSYNC information is derived from the 3D facial mesh, the generator in the second stage will not overfit the limited lip patterns present in given clips. This masks our method full-parameter fine-tunable on only seconds of videos. + +# 3.3 Face-Swapping + +We now illustrate the power of our comprehensive framework to directly achieve face-swapping and lip-sync with unified training. Our design of leveraging the $\mathcal{W}$ space for enhancing facial generation considers the similar setting used in previous face swapping studies [35, 72, 73]. Moreover, the identity information of the audio-predicted mesh $\hat{M}_t$ can be easily transferred by sending different template meshes into the style encoder of the Style-SyncFormer. So that the shape of facial organs can be seamlessly swapped to another person. + +![](images/5e05de95fed44f81f507b1eeeba50ef9faff99bc4619ac190956d9a3b898e764.jpg) +Fig. 3: Face-Swapping Pipeline. With input data reconfiguration and additional training losses, we achieve lip-syncing and face-swapping simultaneously. + +The Face-Swapping Data Stream. The pipeline is depicted in Fig. 3. We first enlarge the masking area $A$ to $A'$ that contains the whole facial part including eyes and noses, which is more suitable for the concerned task. Then we involve a frame $I'$ of another identity and its corresponding mesh $M'$ to provide additional information. We project $M'$ with the pose of $M_t^p$ to $M'^p$ and overlay it on $I_t$ to form $I_t^{A'M'} = (1 - A') * I_t + A' * M'^p$ . + +The face-swapping training consists of two parts, updating only the Style-based generator. The first is the same as Sec 3.2, which reconstructs $I_{t}$ solely from $\{I_t^{AM}, I_{ref}\}$ . The second part intends to create an image $\hat{I}_t'$ , which has $I'$ 's identity on $I_{t}$ 's head with $I_{t}$ 's original tone. This requires the generator G to take $\mathbf{F}_f' = \{f_f'', F_1', \dots, F_L'\}$ encoded from different sources, including $f_f''$ encoded from $\{I_t^{A'M'}, I'\}$ for ID injection in the $\mathcal{W}$ space and $\{F_1', \dots, F_L'\}$ from $\{I_t^{A'M'}, I_{ref}\}$ for expressing facial attributes in the noise space. + +Training Objectives. Under the cross-identity training setting, the commonly used identity cosine distance loss $\mathcal{L}_{id} = 1 - \cos (E_{id}(\hat{I}_t'),E_{id}(I'))$ where $E_{id}$ is the encoder of ArcFace [14] and feature matching loss with the original $I_{t}$ [9,70,73] in the face-swapping task are leveraged. + +Note that after the network is trained in the face-swapping style, which costs more time, high-quality face-swapping and lip-sync abilities can be achieved within one unified model (with the same set of weights). To the best of our knowledge, we are one of the first methods that unifies the two tasks within one model, which is efficient and beneficial to the creation of virtual performers. + +# 4 Experiments + +Although lip-sync is our main concern, we thoroughly conduct experiments from the perspectives of both lip-sync and face-swapping to show the effectiveness of our unified framework. For fair comparisons, we present results using only our generalized model, please refer to the supplementary files for personalized experiments. + +![](images/7bd5cd42c8e6a187f12d217d10b2d9a8684b5f03a0e77624b26071c23d3d21fc.jpg) +Fig. 4: Qualitative Cross-Sync Results. The top row shows the lip-synced videos of the driving audio. Generation results based on the "Template" row should have the same lip shape as the "Lip-Synced Video" in the first row. + +![](images/cc4d1286ad3cbf4138659bd7db49b64ab2b5b588ead0b0ee88d574ae8fe092fc.jpg) + +Datasets. We train our model with two audio-visual datasets, HDTF [79] and VoxCeleb2 [11]. Both two datasets provide high-quality videos from a variety of identities for lip-syncing and face-swapping training. Notably, we split HDTF into train and test sets without overlap. The test set comprises 100 10-second audio-video pairs. Evaluations on lip-sync are conducted on the official test set of VoxCeleb2 and the introduced test set of HDTF. Face-swapping evaluation is performed on FaceForensics++ (FF++) [45], which contains 1000 face videos. We use the high-quality version and adopt the official video-video pairing. Apart from publicly available datasets, certain visual outcomes in our experiments are produced based on videos sourced from the Internet. + +Implementation Details. All videos are processed at a frame rate of 25FPS, and facial alignment is conducted based on pre-detected landmarks. Subsequently, all faces are cropped to a size of $256 \times 256$ . The architecture of G is built upon the success of recent methods [22,28,73,74]. During inference, we split videos into 100-frame consecutive chunks with a 20-frame overlap. Face meshes are predicted in an auto-aggressive manner with the proposed Style-SyncFormer depending on all the previously predicted results within each chunk. For face reconstruction, we train DeepFaceReconstruction [17] on the topology of both Hifi3DFace [2] and BFM. The main results are reported from the Hifi3DFace version but the choice of topology makes little difference. Our models are trained + +Table 1: Quantitative results on HDTF and VoxCeleb2. For LMD and $\Delta$ Sync the lower the better, and the higher the better for others. + +
MethodHDTFVoxCeleb2
SSIM ↑PSNR ↑LMD ↓ΔSync ↓SSIM ↑PSNR ↑LMD ↓ΔSync ↓
Wav2Lip [41]0.8430.935.411.370.8431.583.442.28
ReTalking [10]0.8131.086.060.740.8331.644.421.50
SadTalker [78]0.7730.866.763.620.7931.365.033.03
StyleSync [22]0.8331.574.970.800.8532.523.281.57
Ours-Sync0.8431.764.340.660.8532.883.190.88
+ +on 4 Tesla A100 GPUs for 2 days with batch size 16. For more implementation details, please refer to the supplementary files. + +Comparison Methods. As for lip-sync evaluation, We compare our method with four different methods including the classic Wav2Lip [41], as well as three recent state-of-the-art (SOTA) methods: ReTalking [10], SadTalker [78], and StyleSync [22]. Note that SadTalker [78] targets talking head generation but is well-known for its good quality. All experiments are conducted on the generalized setting without fine-tuning, and the model without face-swapping training is leveraged. Then we compare our method with four studies on face swapping. The counterparts consist of SimSwap [9], InfoSwap [20], StyleSwap [73] and E4S [34]. Furthermore, we aim to evaluate whether combining two methods from the above two tasks can achieve comparable results with our full model despite the cost. Thus we construct two potential counterparts by combining SOTA methods from both areas, denoted as "SimSwap + ReTalking" and "StyleSwap + StyleSync". + +# 4.1 Quantitative Comparisons + +Quantitative experiments of lip-sync are conducted following the self-reconstruction setting. We refrain from directly utilizing the ground truth frame as the reference. For face-swapping, we adhere to a protocol as previous studies [30, 73], wherein we uniformly sample 10 frames from each video in $\mathrm{FF}++$ , resulting in a total of $10\mathrm{k}$ images for evaluation. + +Evaluation Metrics. Both lip-sync and face-swapping are broadly studied, we follow previous works [22, 30, 73, 81] to adopt several popular metrics. For lip-sync, we use SSIM [66] and PSNR for measuring visual similarity. Landmark + +Table 2: Quantitative results evaluated on Table 3: Quantitative ablations on FF++. VoxCeleb2. + +
MethodID Retrieval↑ ID Sim.↑ PoseError ↓ Exp.Error ↓
SimSwap [9]93.07%0.5781.365.07
InfoSwap [20]95.82%0.6352.546.99
StyleSwap [73]97.05%0.6771.565.28
EAS [34]90.21%0.4952.7811.52
Ours-Swap96.70%0.6651.315.13
+ +
MethodSSIM ↑PSNR ↑LMD ↓ΔSync ↓
w/o Mesh-Inject0.8331.573.602.54
w/o Mesh-Style0.8331.803.291.02
TS=500.8532.263.350.95
TS=2000.8532.693.210.88
Ours-Sync0.8532.883.190.88
+ +![](images/ef52110d304742805e54b3a5f5faf20df4ae4e8c02d811e12a564f8b3091bf57.jpg) +Fig. 5: Qualitative Results of Face-Swap. Identity-swapped results should preserve the expression and lip motion of templates. + +![](images/5dd25155ad673664f535a3379ff6724071e4a8388727d7a8ed9a73acf877fdb6.jpg) + +distances around the mouth (LMD), and SyncNet's confidence score [12] are used for lip-sync quality. Particularly, the SyncNet score only indicates how well a video aligns with the learned SyncNet model, thus we use the absolute difference ( $\Delta$ Sync) between ground truth and prediction to assess whether results are close to the original video. For face-swapping, we leverage a pretrained face-recognition model [60] to measure the ID cosine similarity and ID retrieval scores. In addition, pose error and expression error are measured by two pretrained models [46,58] following previous work [73]. + +Lip-Sync. The evaluation results are tabulated in Table 1. Our method outperforms previous works across all metrics, especially concerning lip-sync quality. This is attributed to stylized 3D facial dynamics learned by Style-SyncFormer, which manages to recover the injected personalized speaking style of the target person in the generalized setting. In contrast, previous methods primarily rely on sync-loss [41] and tend to generate a dataset-averaged sync-style. + +Face-Swapping. Results evaluated on $\mathrm{FF}++$ are listed in Table 2. Our model shows comparable performance on all the metrics and achieves the lowest pose error when compared with SOTA face-swap methods. Particularly, our performances on ID-related metrics are only slightly weaker than StyleSwap. + +# 4.2 Qualitative Comparisons + +Subjective evaluation plays a pivotal role in assessing the capabilities of generative models, especially in the context of videos. We strongly encourage readers to watch our supplementary video for a comprehensive understanding. Our method also enjoys the capability to fast optimize for specific identities, resulting in superior outcomes. Please refer to the supplementary files for the personalized results. + +Lip-Sync. We demonstrate two cross-sync comparisons with SOTA methods in Fig. 4. The lip-synced results are conditioned on audio randomly sampled from the dataset. Wav2Lip has accurate lip shapes, but the mouth part is noticeably blurred. SadTalker achieves better fidelity, but it fails to present natural dynamics. VideoReTalking and StyleSync can generate precise and high-quality results; nevertheless, their results are affected by the opening jaw of the template + +as shown in the first column of the right figure. Moreover, they exhibit a deficiency in maintaining person-specific attributes around the mouth. Our method directly produces high-fidelity results with accurate lip motion and rich details of the mouth. + +We present cross-sync results, showcasing the transferred speaking style in Fig. 1 and Fig. 9 (in the supplementary files). It can be seen that our method not only accurately generates lip shapes aligned with the driven video but also successfully animates the target person with the speaking style of the source person in the driving video. In addition, our method inherently supports videodriven lip-syncing as introduced in Sec. 3.2. The comparison is demonstrated in Fig. 10 (in the supplementary files), it shows that person-specific attributes can be preserved better with Style-SyncFormer guidance. + +![](images/862bbb804ef064a9f567289f633a614e12bc58e419e920f7ed6d0fb1a0965432.jpg) +Fig.6: Results of Face-Swapped Lip-Sync. The ID-swapped results are driven by the given audio. We compare with the combinations of two SOTA methods in lip-sync and face-swapping. Results generated by our method preserve details with better fidelity and simultaneously maintain a similar speaking style to the source. + +Face-Swapping. In Fig. 5, we illustrate two face-swapped results for comparison. The generated results in this context should preserve the original expression and lip motion of the template. From the figure, InfoSwap fails to keep a natural brightness across frames. E4S effectively modifies the entire head region but fails to integrate the head part into the background seamlessly. SimSwap and StyleSwap produce good results by preserving precise expression and identity attributes. However, an excessive emphasis on identity feature matching causes them to falter in generating details around certain areas like eyes and mouth. + +Table 4: User study of generalized lip-sync. The scores are ranged from 1 (worst) to 5 (best). + +
MOS onWav2LipReTalkingSadTalkerStyleSyncOurs-Sync
Lip-Sync Quality3.433.092.983.464.57
Generation Quality1.482.833.133.634.67
Video Realness1.592.412.873.114.54
+ +In contrast, our method not only achieves effective identity transfer but also preserves rich details across the entire face. In addition, the videos generated by our method exhibit superior temporal consistency, especially in the mouth region. Please find more results in the supplementary video. + +Furthermore, our method is capable of achieving lip-synced face-swapping conditioned on any audio and identity in a single forward process. We compare our method with two potential baselines "SimSwap + ReTalking" and "StyleSwap + StyleSync" in Fig. 6. Results generated by our method preserve details with better fidelity and simultaneously maintain a similar speaking style to the source. + +User Study. We conducted a user study with 30 participants. We adopt the commonly used Mean Opinion Scores (MOS) rating protocol. All users give their ratings, ranging from 1 to 5. On generalized lip-sync, they are asked to evaluate: a) the quality of lip-sync. b) visual quality of the generated video, and c) whether this video looks real. The results averaged from 10 videos are grouped into Table 4. From the table, our method outperforms the four counterparts in all three aspects by clear margins, underscoring the power of our method and the substantial improvement in visual quality and realism over other existing methods. The experiments are similarly conducted with face-swapping and personalized lip-sync, please find the results in the supplementary files. + +# 4.3 Ablation Study + +Lip-Sync. Here we evaluate four different designs on VoxCeleb2 [11] including: 1) "w/o Mesh-Inject": we mask-out the 3D mesh guidance from the encoded feature maps $\{F_1, \ldots, F_L\}$ , thus it cannot provide detailed spatial information during generation; 2) "w/o Mesh-Style": 3D mesh content is removed in $I_t^{AM}$ to get a $f_f$ , feature vector that feed into the $\mathcal{W}$ space contains only the background and reference information; 3) " $T_S = 50$ ": we reduce the length ( $T_S$ in Sec. 3.1) of the speaking-style reference from 90 to 50; 4) " $T_S = 200$ ": the length is extended to 200. The results are presented in Table 3. Comparisons between the first two rows and "Ours-Sync" reveal the vital role played by mesh guidance in the generation process, where the sync quality is degraded to certain degrees, as indicated by LMD and $\Delta$ Sync. In addition, the visual quality is also harmed as shown in Fig. 7 (a). In the bottom three rows of Table 3, we compare the performance of different sequence lengths in the speaking-style encoding of Style-SyncFormer. The results indicate that a sequence of 90 frames ( $\sim$ 4s of + +![](images/f0a60a64fef2a460b4b44cdbe1665b37ab6aacd3df6c18b25396c66e43cc9d33.jpg) +Fig. 7: Ablations. (a) 3D facial mesh provides detailed spatial guidance, empowering superior lip-syncing. (b) The mesh with shapes of facial organs also enhances ID transfer in face-swapping. + +video) leads to good sync quality, and a further extension of the length does not enhance its performance. + +Face-Swapping. In face-swapping, the predicted 3D facial mesh also provides detailed facial attributes. We provide an intuitive comparison in Fig. 7 (b). When the mesh guidance is removed from the encoded feature map ("w/o Mesh-Inject"), we observe a noteworthy degradation in facial dynamics, particularly in lip motion. Please refer to the supplementary video for the comparison. Simultaneously, without the guidance of mesh information, the ID-related metrics of face-swapping are also affected, as indicated at the bottom of Fig. 7 (b) and the same phenomenon applies to "w/o Mesh-Style". + +# 5 Discussion and Conclusion + +Ethical Considerations. Our method possesses the capability to create fabricated talks and speeches, even with swapped identities, presenting a potential for malicious use. We will restrict the distribution of our models and the generated results exclusively to research purposes. + +Conclusion. We highlight several important properties of our ReSyncer framework: 1) Our easy reconfiguration further unveils the potential of widely studied structures so that high-quality generalized lip-sync results with mesh representation can be achieved. 2) Our framework is designed to adopt external identity information, thus we achieve face-swapping comparable to the state of the arts while keeping the lip-sync ability within one unified model. 3) Our network supports speaking style transfer, video-driven facial animation, and can be applied to real-time live streamers. These properties complementarily satisfy various needs of virtual performer creation under different circumstances. + +Acknowledgments. This work is in part supported by National Natural Science Foundation of China with No. 62394322, Beijing Natural Science Foundation with No. L222024, as well as the Beijing National Research Center for Information Science and Technology (BNRist) key projects. This study is also supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). + +# References + +1. Baevski, A., Zhou, Y., Mohamed, A., Auli, M.: wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in Neural Information Processing Systems (2020) +2. Bao, L., Lin, X., Chen, Y., Zhang, H., Wang, S., Zhe, X., Kang, D., Huang, H., Jiang, X., Wang, J., Yu, D., Zhang, Z.: High-fidelity 3d digital human head creation from rgb-d selfies. ACM Transactions on Graphics (2021) +3. Blanz, V., Vetter, T.: A morphable model for the synthesis of 3d faces. In: Proceedings of the 26th annual conference on Computer graphics and interactive techniques. pp. 187-194 (1999) +4. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022) +5. Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: Maskgit: Masked generative image transformer. arXiv preprint arXiv:2202.04200 (2022) +6. Chen, L., Cui, G., Liu, C., Li, Z., Kou, Z., Xu, Y., Xu, C.: Talking-head generation with rhythmic head motion. In: European Conference on Computer Vision. pp. 35-51. Springer (2020) +7. Chen, L., Li, Z., Maddox, R.K., Duan, Z., Xu, C.: Lip movements generation at a glance. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 520-535 (2018) +8. Chen, L., Maddox, R.K., Duan, Z., Xu, C.: Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7832-7841 (2019) +9. Chen, R., Chen, X., Ni, B., Ge, Y.: Simswap: An efficient framework for high fidelity face swapping. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 2003-2011 (2020) +0. Cheng, K., Cun, X., Zhang, Y., Xia, M., Yin, F., Zhu, M., Wang, X., Wang, J., Wang, N.: Videoretalking: Audio-based lip synchronization for talking head video editing in the wild (2022) +1. Chung, J.S., Nagrani, A., Zisserman, A.: Voxceleb2: Deep speaker recognition. In: INTERSPEECH (2018) +12. Chung, J.S., Zisserman, A.: Out of time: automated lip sync in the wild. In: ACCV (2016) +13. Cudeiro, D., Bolkart, T., Laidlaw, C., Ranjan, A., Black, M.J.: Capture, learning, and synthesis of 3d speaking styles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10101-10111 (2019) + +14. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4690-4699 (2019) +15. Deng, Y., Yang, J., Chen, D., Wen, F., Tong, X.: Disentangled and controllable face image generation via 3d imitative-contrastive learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020) +16. Deng, Y., Yang, J., Xiang, J., Tong, X.: Gram: Generative radiance manifolds for 3d-aware image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10673-10683 (June 2022) +17. Deng, Y., Yang, J., Xu, S., Chen, D., Jia, Y., Tong, X.: Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019) +18. Fan, Y., Lin, Z., Saito, J., Wang, W., Komura, T.: Faceformer: Speech-driven 3d facial animation with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18770-18780 (2022) +19. Gafni, G., Thies, J., Zollhöfer, M., Nießner, M.: Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8649-8658 (June 2021) +20. Gao, G., Huang, H., Fu, C., Li, Z., He, R.: Information bottleneck disentanglement for identity swapping. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3404-3413 (2021) +21. Goodfellow, I.J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. arXiv preprint arXiv:1406.2661 (2014) +22. Guan, J., Zhang, Z., Zhou, H., Hu, T., Wang, K., He, D., Feng, H., Liu, J., Ding, E., Liu, Z., Wang, J.: Stylesync: High-fidelity generalized and personalized lip sync in style-based generator. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +23. Guo, Y., Chen, K., Liang, S., Liu, Y., Bao, H., Zhang, J.: Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2021) +24. Jamaludin, A., Chung, J.S., Zisserman, A.: You said that?: Synthesising talking faces from audio. International Journal of Computer Vision 127(11), 1767-1779 (2019) +25. Ji, X., Zhou, H., Wang, K., Wu, Q., Wu, W., Xu, F., Cao, X.: Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. SIGGRAPH (2022) +26. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4401-4410 (2019) +27. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of StyleGAN. In: Proc. CVPR (2020) +28. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8110-8119 (2020) +29. Li, L., Wang, S., Zhang, Z., Ding, Y., Zheng, Y., Yu, X., Fan, C.: Write-a-speaker: Text-based emotional and rhythmic talking-head generation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1911-1920 (2021) + +30. Li, L., Bao, J., Yang, H., Chen, D., Wen, F.: Faceshifter: Towards high fidelity and occlusion aware face swapping. arXiv preprint arXiv:1912.13457 (2019) +31. Li, Y., Ma, C., Yan, Y., Zhu, W., Yang, X.: 3d-aware face swapping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12705-12714 (2023) +32. Liang, B., Pan, Y., Guo, Z., Zhou, H., Hong, Z., Han, X., Han, J., Liu, J., Ding, E., Wang, J.: Expressive talking head generation with granular audio-visual control. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3387-3396 (2022) +33. Liu, X., Xu, Y., Wu, Q., Zhou, H., Wu, W., Zhou, B.: Semantic-aware implicit neural audio-driven video portrait generation. ECCV (2022) +34. Liu, Z., Li, M., Zhang, Y., Wang, C., Zhang, Q., Wang, J., Nie, Y.: Fine-grained face swapping via regional gan inversion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8578-8587 (2023) +35. Luo, Y., Zhu, J., He, K., Chu, W., Tai, Y., Wang, C., Yan, J.: Styleface: Towards identity-disentangled face generation on megapixels. In: European Conference on Computer Vision. pp. 297-312. Springer (2022) +36. Ma, Y., Wang, S., Ding, Y., Ma, B., Lv, T., Fan, C., Hu, Z., Deng, Z., Yu, X.: Talkclip: Talking head generation with text-guided expressive speaking styles. arXiv preprint arXiv:2304.00334 (2023) +37. Ma, Y., Wang, S., Hu, Z., Fan, C., Lv, T., Ding, Y., Deng, Z., Yu, X.: Styletalk: One-shot talking head generation with controllable speaking styles. AAAI (2023) +38. Natsume, R., Yatagawa, T., Morishima, S.: Rsgan: Face swapping and editing using face and hair representation in latent spaces. arxiv 2018. arXiv preprint arXiv:1804.03447 2 +39. Natsume, R., Yatagawa, T., Morishima, S.: Fsnet: An identity-aware generative model for image-based face swapping. In: Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part VI 14. pp. 117–132. Springer (2019) +40. Park, S.J., Kim, M., Hong, J., Choi, J., Ro, Y.M.: Synctalkface: Talking face generation with precise lip-syncing via audio-lip memory. In: AAAI Conference on Artificial Intelligence. Association for the Advancement of Artificial Intelligence (2022) +41. Prajwal, K., Mukhopadhyay, R., Namboodiri, V.P., Jawahar, C.: A lip sync expert is all you need for speech to lip generation in the wild. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 484-492 (2020) +42. Qiu, H., Chen, Z., Jiang, Y., Zhou, H., Fan, X., Yang, L., Wu, W., Liu, Z.: Relitalk: Relightable talking portrait generation from a single video. International Journal of Computer Vision pp. 1-16 (2024) +43. Richard, A., Zollhöfer, M., Wen, Y., de la Torre, F., Sheikh, Y.: Meshtalk: 3d face animation from speech using cross-modality disentanglement. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021) +44. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) +45. Rössler, A., Cozzolino, D., Verdoliva, L., Riess, C., Thies, J., Nießner, M.: FaceForensics++: Learning to detect manipulated facial images. In: International Conference on Computer Vision (ICCV) (2019) +46. Ruiz, N., Chong, E., Rehg, J.M.: Fine-grained head pose estimation without keypoints. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 2074-2083 (2018) + +47. Shen, S., Li, W., Zhu, Z., Duan, Y., Zhou, J., Lu, J.: Learning dynamic facial radiance fields for few-shot talking head synthesis. In: ECCV (2022) +48. Shen, S., Zhao, W., Meng, Z., Li, W., Zhu, Z., Zhou, J., Lu, J.: Difftalk: Crafting diffusion models for generalized audio-driven portraits animation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1982-1991 (2023) +49. Song, Y., Zhu, J., Li, D., Wang, X., Qi, H.: Talking face generation by conditional recurrent adversarial network. arXiv preprint arXiv:1804.04786 (2018) +50. Stypulkowski, M., Vougioukas, K., He, S., Zieba, M., Petridis, S., Pantic, M.: Diffused heads: Diffusion models beat gans on talking-face generation. arXiv preprint arXiv:2301.03396 (2023) +51. Sun, Y., Zhou, H., Wang, K., Wu, Q., Hong, Z., Liu, J., Ding, E., Wang, J., Liu, Z., Hideki, K.: Masked lip-sync prediction by audio-visual contextual exploitation in transformers. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022) +52. Suwajanakorn, S., Seitz, S.M., Kemelmacher-Shlizerman, I.: Synthesizing obama: learning lip sync from audio. ACM Transactions on Graphics (ToG) 36(4), 1-13 (2017) +53. Tang, J., Wang, K., Zhou, H., Chen, X., He, D., Hu, T., Liu, J., Zeng, G., Wang, J.: Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368 (2022) +54. Thambiraja, B., Habibie, I., Aliakbarian, S., Cosker, D., Theobalt, C., Thies, J.: Imitator: Personalized speech-driven 3d facial animation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 20621-20631 (2023) +55. Thies, J., Elgharib, M., Tewari, A., Theobalt, C., Nießner, M.: Neural voice puppetry: Audio-driven facial reenactment. In: European Conference on Computer Vision. pp. 716-731. Springer (2020) +56. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems 30 (2017) +57. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) +58. Vemulapalli, R., Agarwala, A.: A compact embedding for facial expression similarity. In: proceedings of the IEEE/cvf conference on computer vision and pattern recognition. pp. 5683-5692 (2019) +59. Wang, D., Deng, Y., Yin, Z., Shum, H.Y., Wang, B.: Progressive disentangled representation learning for fine-grained controllable talking head synthesis. CVPR (2023) +60. Wang, H., Wang, Y., Zhou, Z., Ji, X., Gong, D., Zhou, J., Li, Z., Liu, W.: Cosface: Large margin cosine loss for deep face recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5265-5274 (2018) +61. Wang, J., Zhao, K., Ma, Y., Zhang, S., Zhang, Y., Shen, Y., Zhao, D., Zhou, J.: Facecomposer: A unified model for versatile facial content creation. In: Thirty-seventh Conference on Neural Information Processing Systems (2023) +62. Wang, K., Wu, Q., Song, L., Yang, Z., Wu, W., Qian, C., He, R., Qiao, Y., Loy, C.C.: Mead: A large-scale audio-visual dataset for emotional talking-face generation. In: ECCV (2020) +63. Wang, S., Li, L., Ding, Y., Fan, C., Yu, X.: Audio2head: Audio-driven one-shot talking-head generation with natural head motion. arXiv preprint arXiv:2107.09293 (2021) + +64. Wang, S., Li, L., Ding, Y., Yu, X.: One-shot talking face generation from single-speaker audio-visual correlation learning. In: Proceedings of the AAAI Conference on Artificial Intelligence (2022) +65. Wang, T.C., Liu, M.Y., Zhu, J.Y., Tao, A., Kautz, J., Catanzaro, B.: High-resolution image synthesis and semantic manipulation with conditional gans. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8798-8807 (2018) +66. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004) +67. Wu, H., Jia, J., Wang, H., Dou, Y., Duan, C., Deng, Q.: Imitating arbitrary talking style for realistic audio-driven talking face synthesis. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 1478-1486 (2021) +68. Wu, X., Hu, P., Wu, Y., Lyu, X., Cao, Y.P., Shan, Y., Yang, W., Sun, Z., Qi, X.: Speech2lip: High-fidelity speech to lip generation by learning from a short video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22168-22177 (2023) +69. Xing, J., Xia, M., Zhang, Y., Cun, X., Wang, J., Wong, T.T.: Codetalker: Speechdriven 3d facial animation with discrete motion prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12780-12790 (June 2023) +70. Xu, C., Zhang, J., Hua, M., He, Q., Yi, Z., Liu, Y.: Region-aware face swapping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7632-7641 (2022) +71. Xu, Y., Deng, B., Wang, J., Jing, Y., Pan, J., He, S.: High-resolution face swapping via latent semantics disentanglement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7642-7651 (2022) +72. Xu, Z., Hong, Z., Ding, C., Zhu, Z., Han, J., Liu, J., Ding, E.: Mobilefaceswap: A lightweight framework for video face swapping. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 2973-2981 (2022) +73. Xu, Z., Zhou, H., Hong, Z., Liu, Z., Liu, J., Guo, Z., Han, J., Liu, J., Ding, E., Wang, J.: Styleswap: Style-based generator empowers robust face swapping. In: European Conference on Computer Vision. pp. 661-677. Springer (2022) +74. Yang, T., Ren, P., Xie, X., Zhang, L.: Gan prior embedded network for blind face restoration in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 672-681 (2021) +75. Yao, S., Zhong, R., Yan, Y., Zhai, G., Yang, X.: Dfa-nerf: Personalized talking head generation via disentangled face attributes neural rendering. arXiv preprint arXiv:2201.00791 (2022) +76. Ye, Z., Jiang, Z., Ren, Y., Liu, J., He, J., Zhao, Z.: Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. ICLR (2023) +77. Yu, Z., Yin, Z., Zhou, D., Wang, D., Wong, F., Wang, B.: Talking head generation with probabilistic audio-to-visual diffusion priors. ICCV (2023) +78. Zhang, W., Cun, X., Wang, X., Zhang, Y., Shen, X., Guo, Y., Shan, Y., Wang, F.: Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8652-8661 (2023) +79. Zhang, Z., Li, L., Ding, Y., Fan, C.: Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3661-3670 (2021) + +80. Zhou, H., Liu, Y., Liu, Z., Luo, P., Wang, X.: Talking face generation by adversarially disentangled audio-visual representation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 33, pp. 9299-9306 (2019) +81. Zhou, H., Sun, Y., Wu, W., Loy, C.C., Wang, X., Liu, Z.: Pose-controllable talking face generation by implicitly modularized audio-visual representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4176-4186 (2021) +82. Zhou, Y., Han, X., Shechtman, E., Echevarria, J., Kalogerakis, E., Li, D.: Makelttalk: speaker-aware talking-head animation. ACM Transactions on Graphics (TOG) 39(6), 1-15 (2020) +83. Zhu, Y., Li, Q., Wang, J., Xu, C.Z., Sun, Z.: One shot face swapping on megapixels. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4834-4844 (2021) \ No newline at end of file diff --git a/2024/ReSyncer_ Rewiring Style-based Generator for Unified Audio-Visually Synced Facial Performer/images.zip b/2024/ReSyncer_ Rewiring Style-based Generator for Unified Audio-Visually Synced Facial Performer/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dee1479f5368cbd6b0ee9582e98c9f11631386aa --- /dev/null +++ b/2024/ReSyncer_ Rewiring Style-based Generator for Unified Audio-Visually Synced Facial Performer/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad31f1f50de90bd14aaa8c684908f37b4f007b3b70469ab8775877ad2448661d +size 652821 diff --git a/2024/ReSyncer_ Rewiring Style-based Generator for Unified Audio-Visually Synced Facial Performer/layout.json b/2024/ReSyncer_ Rewiring Style-based Generator for Unified Audio-Visually Synced Facial Performer/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c655ed8a97a444a8592d7f899a6694a7cf2011c9 --- /dev/null +++ b/2024/ReSyncer_ Rewiring Style-based Generator for Unified Audio-Visually Synced Facial Performer/layout.json @@ -0,0 +1,10443 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 139, + 112, + 475, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 112, + 475, + 148 + ], + "spans": [ + { + "bbox": [ + 139, + 112, + 475, + 148 + ], + "type": "text", + "content": "ReSyncer: Rewiring Style-based Generator for Unified Audio-Visually Synced Facial Performer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "text", + "content": "Jiazhi Guan" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "text", + "content": ", Zhiliang Xu" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "text", + "content": ", Hang Zhou" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "text", + "content": ", Kaisiyuan Wang" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "text", + "content": ", Shengyi He" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "text", + "content": ", Zhanwang Zhang" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "text", + "content": ", Borong Liang" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "text", + "content": ", Haocheng Feng" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "text", + "content": ", Errui Ding" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "text", + "content": ", Jingtuo Liu" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "text", + "content": ", Jingdong Wang" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "text", + "content": ", Youjian Zhao" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{1,3\\dagger}" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "text", + "content": ", and Ziwei Liu" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{4}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 157, + 213, + 455, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 213, + 455, + 248 + ], + "spans": [ + { + "bbox": [ + 157, + 213, + 455, + 248 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 157, + 213, + 455, + 248 + ], + "type": "text", + "content": "BNRist, DCST, Tsinghua University " + }, + { + "bbox": [ + 157, + 213, + 455, + 248 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 157, + 213, + 455, + 248 + ], + "type": "text", + "content": "Baidu Inc. \n" + }, + { + "bbox": [ + 157, + 213, + 455, + 248 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 157, + 213, + 455, + 248 + ], + "type": "text", + "content": "Zhongguancun Laboratory " + }, + { + "bbox": [ + 157, + 213, + 455, + 248 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 157, + 213, + 455, + 248 + ], + "type": "text", + "content": "S-Lab, Nanyang Technological University guanjz20@mails.tsinghua.edu.cn, zhouhang09@baidu.com" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 137, + 267, + 476, + 439 + ], + "blocks": [ + { + "bbox": [ + 137, + 267, + 476, + 439 + ], + "lines": [ + { + "bbox": [ + 137, + 267, + 476, + 439 + ], + "spans": [ + { + "bbox": [ + 137, + 267, + 476, + 439 + ], + "type": "image", + "image_path": "7006e089d5e2a453a0432cdf8dda770d32efa20793899df00e4c54f055d1e327.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 450, + 482, + 485 + ], + "lines": [ + { + "bbox": [ + 132, + 450, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 482, + 485 + ], + "type": "text", + "content": "Fig. 1: Lip-Synced/Speaking-Style-Transferred/Face-Swapping Results by ReSyncer. Our method not only produces high-fidelity lip-synced video according to audio but can further transfer the speaking style and identity of any target person." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 503, + 455, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 503, + 455, + 635 + ], + "spans": [ + { + "bbox": [ + 159, + 503, + 455, + 635 + ], + "type": "text", + "content": "Abstract. Lip-syncing videos with given audio is the foundation for various applications including the creation of virtual presenters or performers. While recent studies explore high-fidelity lip-sync with different techniques, their task-orientated models either require long-term videos for clip-specific training or retain visible artifacts. In this paper, we propose a unified and effective framework ReSyncer, that synchronizes generalized audio-visual facial information. The key design is revisiting and rewiring the Style-based generator to efficiently adopt 3D facial dynamics predicted by a principled style-injected Transformer. By simply re-configuring the information insertion mechanisms within the noise and style space, our framework fuses motion and appearance with unified training. Extensive experiments demonstrate that ReSyncer not" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 642, + 250, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 642, + 250, + 665 + ], + "spans": [ + { + "bbox": [ + 146, + 642, + 250, + 665 + ], + "type": "text", + "content": "\\*Equal Contribution. " + }, + { + "bbox": [ + 146, + 642, + 250, + 665 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 146, + 642, + 250, + 665 + ], + "type": "text", + "content": " Corresponding Authors." + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 159, + 117, + 455, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 117, + 455, + 173 + ], + "spans": [ + { + "bbox": [ + 159, + 117, + 455, + 173 + ], + "type": "text", + "content": "only produces high-fidelity lip-synced videos according to audio, but also supports multiple appealing properties that are suitable for creating virtual presenters and performers, including fast personalized fine-tuning, video-driven lip-syncing, the transfer of speaking styles, and even face swapping." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 189, + 230, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 189, + 230, + 201 + ], + "spans": [ + { + "bbox": [ + 132, + 189, + 230, + 201 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 211, + 481, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 211, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 211, + 481, + 258 + ], + "type": "text", + "content": "Generating lip-synced facial videos with speech audio is a foundational technique for various applications including the dubbing of movies and the creation of virtual presenters or performers in the form of realistic humans. The needs in virtual performers' creation can hardly be comprehensively met." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 259, + 481, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 481, + 402 + ], + "type": "text", + "content": "While great attention has been paid to creating dynamic lip-synced videos from a few photos [8, 24, 63, 64, 78, 80-82], previous results are still far from realistic. To generate life-like videos, researchers have made explorations [41] with GANs [10,55], NeRF [23,33,47,53], and Diffusion models [48] to edit existing videos. Despite their efforts, most methods rely on a relatively long-term clip of video for training, making them less flexible and efficient. A few methods [10,22, 41,51] enable generalized local editing without training. However, they still tend to create visible flaws on high-quality videos [41,51] Particularly, StyleSync [22] supports not only arbitrary video modeling but can adapt to the speaking style of a target person with short-clip fine-tuning, which balances the generation quality and efficiency. Nevertheless, their results can still be easily affected by mouth movements on their conditional reference frames, leading to unstable results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 403, + 481, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 481, + 509 + ], + "type": "text", + "content": "The previous cases reveal the limitations of using low-dimensional audio information to directly modify high-dimensional visual data. On the other hand, intermediate representations like 3D facial models are widely used, but their potential as spatial guidance for efficient, generalized editing remains underexplored. Using only 3DMM [3] coefficients [15,16,19,29,37,55] also suffers from the cross-domain information injection problem. While textured rendering offers a promising alternative [6,55,67], its implementation complexity and susceptibility to errors during 3D modeling or texture fitting can lead to degraded results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 481, + 641 + ], + "type": "text", + "content": "In this paper, we revisit and re-wire the powerful Style-based generator [26, 27] to generate faithful lip-synced videos and support virtual performer creation in a concise yet effective framework called ReSyncer. With unified model training, various appealing properties including generalized and personalized lip-sync, speaking style transfer, live streaming, and face swapping can be achieved within one model. Our key is to show how an easy reconfiguration empowers Style-based generator with various powerful audio-visual synchronization abilities by involving 3D facial dynamics. Specifically, we use only roughly fitted 3D facial meshes, instead of coefficients to bridge the audio and image domains. A principled Style-injected Lip-Sync Transformer (Style-SyncFormer) learns stylized point cloud displacements given template meshes and driving audios." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "We then perform easy but crucial modifications to Style-based generator with a simple encoder used in previous work [22, 73, 74]. Concretely, while lip-sync" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "J. Guan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "studies [10,22,41,51,61] mask out the mouth area of the target frame, we overlay reconstructed mesh onto it as strong spatial guidance. Then we compliment the textural and identity information with an attached reference frame, which serves as the source of both the noise and style-related space. With generalized training on massive data, the powerful generator focuses on recovering facial details from the roughly predicted lip-synced white-colored meshes, leading to high-fidelity and stable lip-sync with higher quality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 482, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 296 + ], + "type": "text", + "content": "Additionally, our framework design goes beyond existing solutions by uniformly addressing a broader range of needs in virtual performer creation. These needs include, for instance, preserving personal styles for specific users or changing face identities for privacy protection and diverse character generation. Particularly, our method directly supports few-shot few-step personalized fine-tuning. Moreover, our framework supports *syncing* (swapping) different identities on the same target video with additional training strategies. High-fidelity face-swapping and lip-syncing can hence be simultaneously satisfied." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 297, + 487, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 487, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 487, + 440 + ], + "type": "text", + "content": "Our contributions can be summarized as follows: 1) We propose the ReSyncer framework, which shows the powerfulness of Style-based generator on syncing audio-visual facial information by involving 3D facial meshes with a simple reconfiguration. 2) We propose the Style-SyncFormer which learns stylized 3D facial dynamics with simple Transformer blocks, enabling generalized 3D facial animation. 3) Extensive experiments show that our framework not only enables lip-syncing of higher stability and quality, but also supports various intriguing properties that are essential for creating virtual performers, including fast personalized fine-tuning, video-driven lip-syncing, the transfer of speaking styles, and even face swapping. To the best of our knowledge, we are one of the first methods that support state-of-the-art lip-syncing and face-swapping with one unified model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 460, + 238, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 460, + 238, + 473 + ], + "spans": [ + { + "bbox": [ + 131, + 460, + 238, + 473 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 487, + 320, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 487, + 320, + 499 + ], + "spans": [ + { + "bbox": [ + 131, + 487, + 320, + 499 + ], + "type": "text", + "content": "2.1 Audio-Driven Facial Animation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 509, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 617 + ], + "type": "text", + "content": "Talking Head Generation. A great number of studies focus on predicting the whole human head movements according to audios [7, 8, 23-25, 32, 36, 37, 42, 52, 59, 62, 63, 77, 78, 81]. More recently, one-shot-based methods adopt representations from 3D, videos, and even texts for more controllable results. While other approaches [23, 33, 47, 53, 75, 76] attempt to integrate speech audios into the neural radiance field for personalized audio-to-3D modeling. However, these methods usually fail to achieve stable lip-sync results and show terrible robustness performance when using audio in the wild. Moreover, the form of talking head generation cannot be applied to edit arbitrary video." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Lip-Syncing-based Facial Editing. Apart from generating the entire head, certain studies [10, 22, 40, 41, 49, 51, 55, 68] choose to directly edit the mouth shape according to the input audios. Specifically, [41] relies on a well-trained lip-sync discriminator to produce highly synchronized animations. [51] introduces" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 407, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 407, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 407, + 91, + 447, + 102 + ], + "type": "text", + "content": "ReSyncer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": "an audio-visual cross-modality framework based on transformer [5]. Particularly, [22] builds a person-agnostic lip-sync framework by StyleGAN2 [27], which also supports personalized optimization. Similarly, [10] also provides identity refinement for the final animation results. Although they have shown accurate lip-sync and comparably high-fidelity results, visible artifacts may still exist due to the latent contradiction between textural and spatial information. Few studies [48, 50] built upon diffusion models [44] are presented recently. Our ReSyncer generates high-fidelity lip-sync videos with high stability and efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 480, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 480, + 281 + ], + "type": "text", + "content": "Audio-Driven 3D Facial Animation. The first stage of our work shares a similar setting with methods that generate audio-driven 3D facial animation [13, 18,43,54,69]. Most of them are performed on 3D scan data to predict the point clouds' dynamics. They have explored Transformers [57] and codebooks [56] with different settings on this task. We adopt a highly efficient form to use Transformer in our work." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 301, + 236, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 301, + 236, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 301, + 236, + 314 + ], + "type": "text", + "content": "2.2 Face Swapping" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 322, + 480, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 480, + 476 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 480, + 476 + ], + "type": "text", + "content": "To transfer the identity of one person to the face of another person is a longstanding problem in computer vision. In recent years, GAN [21] has played an important role in this task. RSGAN [38] and FSNet [39] encode the facial region and the non-facial region to different latent codes and perform swapping by replacing these latent representations. [83] proposes a face-swapping method MegaFS built upon the GAN-inversion technique with StyleGAN2, however, the results are not temporal consistent. Similarly, [71, 73] make their attempts to apply the strong prior learned by StyleGAN2 [27] to generate high-resolution swapped faces. Recently, [31] takes inspiration from 3D generative model [4] to capture fine-grained details of face shape and strengthen the robustness under large pose variance by extending the face swapping task into 3D latent space. However, they fail to produce results with accurate poses and expressions when balancing identity similarity and attribution preservation of the target frame." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 477, + 480, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 477, + 480, + 537 + ], + "spans": [ + { + "bbox": [ + 130, + 477, + 480, + 537 + ], + "type": "text", + "content": "Different from these methods, ReSyncer takes advantage of intermediate guidance from 3D mesh to synthesize accurate pose and expression. Moreover, we achieve face-swapping in a compatible with accurate lip-syncing, which allows us to further satisfy more imaginative demands in creating virtual presenters and performers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 556, + 233, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 233, + 570 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 233, + 570 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "content": "In this section, we introduce the details of our ReSyncer framework as depicted in Fig. 2. It consists of two stages which are all simply designed and easy to implement. In section 3.1 we illustrate our design of the style-injected lip-sync Transformer (Style-SyncFormer) which predicts 3D facial dynamics. Then we show how to reconfigure the Style-based generator to support high-fidelity generalized and personalized lip-sync in section 3.2. Finally, we illustrate our unified design naturally supports face-swapping training in section 3.3." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "J. Guan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 478, + 230 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 478, + 230 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 478, + 230 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 478, + 230 + ], + "type": "image", + "image_path": "434071d6863b2949c47f490e6a9ce740f7003eb00ea42586851c7f12e4e6a63b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 233, + 482, + 277 + ], + "lines": [ + { + "bbox": [ + 130, + 233, + 482, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 482, + 277 + ], + "type": "text", + "content": "Fig. 2: The ReSyncer Framework. In the first stage, the Style-SyncFormer takes the style template and audio input to predict 3D facial dynamics. Then the predicted mesh overlays on the target frame to provide strong spatial guidance. The Style-based generator G processes the overlay and reference frames to produce the final result." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 301, + 482, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 482, + 408 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 482, + 408 + ], + "type": "text", + "content": "Task Formulation. The task of generalized lip-sync is normally trained on massive in-the-wild videos. Specifically, a training data consists of video clip " + }, + { + "bbox": [ + 130, + 301, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = \\{I_1,\\dots ,I_T\\}" + }, + { + "bbox": [ + 130, + 301, + 482, + 408 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 301, + 482, + 408 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 301, + 482, + 408 + ], + "type": "text", + "content": " frames and the accompanying (processed) audio clip features " + }, + { + "bbox": [ + 130, + 301, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\mathbf{a} = \\{a_{1},\\ldots ,a_{T}\\}" + }, + { + "bbox": [ + 130, + 301, + 482, + 408 + ], + "type": "text", + "content": ". The lower half of the target frame " + }, + { + "bbox": [ + 130, + 301, + 482, + 408 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 130, + 301, + 482, + 408 + ], + "type": "text", + "content": " is masked out by a mask " + }, + { + "bbox": [ + 130, + 301, + 482, + 408 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 301, + 482, + 408 + ], + "type": "text", + "content": " to form " + }, + { + "bbox": [ + 130, + 301, + 482, + 408 + ], + "type": "inline_equation", + "content": "I_t^A = (\\mathbf{1} - A)*I_t" + }, + { + "bbox": [ + 130, + 301, + 482, + 408 + ], + "type": "text", + "content": " as [10,22,41,51]. The normal training objective is to recover the original frame given the audio clip " + }, + { + "bbox": [ + 130, + 301, + 482, + 408 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 130, + 301, + 482, + 408 + ], + "type": "text", + "content": " and a reference frame " + }, + { + "bbox": [ + 130, + 301, + 482, + 408 + ], + "type": "inline_equation", + "content": "I_{ref}" + }, + { + "bbox": [ + 130, + 301, + 482, + 408 + ], + "type": "text", + "content": ". However, the final results might be interfered with by the structure of the reference frame when the estimated animation significantly differs from the reference." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 409, + 482, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 409, + 482, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 482, + 469 + ], + "type": "text", + "content": "In this paper, we involve roughly predicted 3D facial meshes as intermediate representations. For every frame in the training video, we reconstruct the 3D facial meshes " + }, + { + "bbox": [ + 130, + 409, + 482, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{M} = \\{M_1,\\dots ,M_T\\}" + }, + { + "bbox": [ + 130, + 409, + 482, + 469 + ], + "type": "text", + "content": ". Thus our framework can be divided into the stages of predicting 3D facial dynamics from audio and the recovery of frames " + }, + { + "bbox": [ + 130, + 409, + 482, + 469 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 130, + 409, + 482, + 469 + ], + "type": "text", + "content": " with the assistance of " + }, + { + "bbox": [ + 130, + 409, + 482, + 469 + ], + "type": "inline_equation", + "content": "M_{t}" + }, + { + "bbox": [ + 130, + 409, + 482, + 469 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 488, + 345, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 488, + 345, + 501 + ], + "spans": [ + { + "bbox": [ + 131, + 488, + 345, + 501 + ], + "type": "text", + "content": "3.1 Style-Injected Lip-Sync Transformer" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "type": "text", + "content": "In the first stage, we propose the Style-SyncFormer that takes audio features " + }, + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{t_0:t_1} = \\{a_{t_0},\\dots ,a_{t_1}\\}" + }, + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "type": "text", + "content": " extracted from pretrained Wav2Vec2 [1] to predict 3D meshes " + }, + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{t_0:t_1} = \\{M_{t_0},\\dots ,M_{t_1}\\}" + }, + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "type": "text", + "content": ". There are " + }, + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "type": "text", + "content": " consecutive frames between " + }, + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "type": "text", + "content": ". Previous work [18] has explored abilities to learn 3D dynamics but it handles only a fixed number of identities. In contrast, the proposed Style-SyncFormer facilitates zero-shot prediction of 3D dynamics while preserving the original speaking style or transferring it from another person." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "Data Processing. During training, all pose information is eliminated from the meshes and aligned to the canonical space. Instead of directly learning all vertices' positions, we learn the displacements compared to a template mesh " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\bar{M}" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": ", the mean of all meshes from this person without expression coefficients. The training target becomes predicting the displacements " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{M}_{t_0:t_1} = \\{\\Delta M_{t_0},\\dots \\Delta M_{t_1}\\}" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " between the template and each reconstructed result." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 407, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 407, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 407, + 91, + 447, + 102 + ], + "type": "text", + "content": "ReSyncer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "Network Design. The main network of Style-SyncFormer is built based on only one Transformer [57] encoder block for style information encoding and one decoder block for mesh prediction, where we adopt the periodic positional encoding and biased causal multi-head attention from FaceFormer [18]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 480, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 271 + ], + "type": "text", + "content": "During training, each shifted-right displacement " + }, + { + "bbox": [ + 130, + 164, + 480, + 271 + ], + "type": "inline_equation", + "content": "\\Delta M_{t}" + }, + { + "bbox": [ + 130, + 164, + 480, + 271 + ], + "type": "text", + "content": " is mapped to embedding " + }, + { + "bbox": [ + 130, + 164, + 480, + 271 + ], + "type": "inline_equation", + "content": "d_t^m" + }, + { + "bbox": [ + 130, + 164, + 480, + 271 + ], + "type": "text", + "content": " and sent into the positional encoding to serve as tokens " + }, + { + "bbox": [ + 130, + 164, + 480, + 271 + ], + "type": "inline_equation", + "content": "f_{t}^{m}" + }, + { + "bbox": [ + 130, + 164, + 480, + 271 + ], + "type": "text", + "content": ". Previous studies [18, 43] learn the 3D dynamics of only limited identities with a simple one-hot speaker identity embedding [13, 18]. Differently, we encode the information of a total of " + }, + { + "bbox": [ + 130, + 164, + 480, + 271 + ], + "type": "inline_equation", + "content": "T_{S} = 90" + }, + { + "bbox": [ + 130, + 164, + 480, + 271 + ], + "type": "text", + "content": " consecutive " + }, + { + "bbox": [ + 130, + 164, + 480, + 271 + ], + "type": "inline_equation", + "content": "\\Delta M_{ref}" + }, + { + "bbox": [ + 130, + 164, + 480, + 271 + ], + "type": "text", + "content": " as speaking style reference. The style information is embedded as a feature " + }, + { + "bbox": [ + 130, + 164, + 480, + 271 + ], + "type": "inline_equation", + "content": "f_{S}^{r}" + }, + { + "bbox": [ + 130, + 164, + 480, + 271 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 164, + 480, + 271 + ], + "type": "inline_equation", + "content": "\\Delta M_{ref}" + }, + { + "bbox": [ + 130, + 164, + 480, + 271 + ], + "type": "text", + "content": " using an additional Transformer encoder block. After that, we inject the speaking style " + }, + { + "bbox": [ + 130, + 164, + 480, + 271 + ], + "type": "inline_equation", + "content": "f_{S}^{r}" + }, + { + "bbox": [ + 130, + 164, + 480, + 271 + ], + "type": "text", + "content": " by adding it with mesh tokens " + }, + { + "bbox": [ + 130, + 164, + 480, + 271 + ], + "type": "inline_equation", + "content": "f_{t}^{m}" + }, + { + "bbox": [ + 130, + 164, + 480, + 271 + ], + "type": "text", + "content": ". Note that when replacing " + }, + { + "bbox": [ + 130, + 164, + 480, + 271 + ], + "type": "inline_equation", + "content": "\\Delta M_{ref}" + }, + { + "bbox": [ + 130, + 164, + 480, + 271 + ], + "type": "text", + "content": " with mesh sequences of another identity, speaking style transfer can be achieved." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 272, + 480, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 480, + 320 + ], + "type": "text", + "content": "In the decoder block, the input tokens are sent into multi-head self-attention, cross-attention, and feed-forward layers. The audio features are further encoded as embeddings " + }, + { + "bbox": [ + 130, + 272, + 480, + 320 + ], + "type": "inline_equation", + "content": "f_{t}^{a}" + }, + { + "bbox": [ + 130, + 272, + 480, + 320 + ], + "type": "text", + "content": " and leveraged in the cross-attentions. The final output is transformed to " + }, + { + "bbox": [ + 130, + 272, + 480, + 320 + ], + "type": "inline_equation", + "content": "\\Delta \\hat{M}_{t_0:t_1} = \\{\\Delta \\hat{M}_{t_0},\\dots ,\\Delta \\hat{M}_{t_1}\\}" + }, + { + "bbox": [ + 130, + 272, + 480, + 320 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 320, + 394, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 320, + 394, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 320, + 394, + 332 + ], + "type": "text", + "content": "Loss Function. The training objectives can be written as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 342, + 481, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 342, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 133, + 342, + 481, + 374 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {3 D} = \\sum_ {t = t _ {0} + 1} ^ {t _ {1}} \\| (\\varDelta \\hat {M} _ {t} - \\varDelta \\hat {M} _ {t - 1}) - (\\varDelta M _ {t} - \\varDelta M _ {t - 1}) \\| _ {2} + \\lambda_ {1} \\sum_ {t = t _ {0}} ^ {t _ {1}} \\| \\varDelta \\hat {M} _ {t} - \\varDelta M _ {t} \\| _ {2},", + "image_path": "21a95bdba6779f5bb57e7322d024f21ce9ded708aa835bc779cd6dd4649592a2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 384, + 384, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 384, + 384, + 396 + ], + "spans": [ + { + "bbox": [ + 131, + 384, + 384, + 396 + ], + "type": "text", + "content": "where the first term is the temporal consistency loss [55]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 396, + 480, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 480, + 444 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 480, + 444 + ], + "type": "text", + "content": "The training is performed in a teacher-forcing manner while the network works autoregressively during inference by using previously predicted results. The displacements are finally transferred to 3D meshes " + }, + { + "bbox": [ + 130, + 396, + 480, + 444 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}_{t_0:t_1}" + }, + { + "bbox": [ + 130, + 396, + 480, + 444 + ], + "type": "text", + "content": " in the canonical space by adding the template mesh " + }, + { + "bbox": [ + 130, + 396, + 480, + 444 + ], + "type": "inline_equation", + "content": "\\bar{M}" + }, + { + "bbox": [ + 130, + 396, + 480, + 444 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 463, + 322, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 463, + 322, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 463, + 322, + 475 + ], + "type": "text", + "content": "3.2 Rewiring Style-based Generator" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 484, + 480, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 480, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 480, + 570 + ], + "type": "text", + "content": "In the second stage, we rewire the widely used Style-based generator [27] to faithfully transfer 3D meshes " + }, + { + "bbox": [ + 130, + 484, + 480, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 484, + 480, + 570 + ], + "type": "text", + "content": " to high-fidelity facial frames " + }, + { + "bbox": [ + 130, + 484, + 480, + 570 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 130, + 484, + 480, + 570 + ], + "type": "text", + "content": " in a simple manner. Previous methods [6,55] resort to accurately textured rendering, which is inefficient from the perspective of implementation and training. Our goal is to find an efficient and effective way. As a result, we leverage the basic white-colored rendering as spatial guidance, which can be easily achieved without textural mapping." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 480, + 666 + ], + "type": "text", + "content": "Network Design. We resort to Style-based generator with noise insertion scheme, which is easy to implement and specializes well in face-related tasks [22, 35,73,74]. Specifically, we project mesh " + }, + { + "bbox": [ + 130, + 570, + 480, + 666 + ], + "type": "inline_equation", + "content": "M_{t}" + }, + { + "bbox": [ + 130, + 570, + 480, + 666 + ], + "type": "text", + "content": " on masked target frame " + }, + { + "bbox": [ + 130, + 570, + 480, + 666 + ], + "type": "inline_equation", + "content": "I_{t}^{A}" + }, + { + "bbox": [ + 130, + 570, + 480, + 666 + ], + "type": "text", + "content": "'s domain to " + }, + { + "bbox": [ + 130, + 570, + 480, + 666 + ], + "type": "inline_equation", + "content": "M_{t}^{p}" + }, + { + "bbox": [ + 130, + 570, + 480, + 666 + ], + "type": "text", + "content": " and overlay to form a 3D-guided image " + }, + { + "bbox": [ + 130, + 570, + 480, + 666 + ], + "type": "inline_equation", + "content": "I_{t}^{AM} = I_{t}^{A} + A * M_{t}^{p}" + }, + { + "bbox": [ + 130, + 570, + 480, + 666 + ], + "type": "text", + "content": ". A reference frame " + }, + { + "bbox": [ + 130, + 570, + 480, + 666 + ], + "type": "inline_equation", + "content": "I_{ref}" + }, + { + "bbox": [ + 130, + 570, + 480, + 666 + ], + "type": "text", + "content": " is attached for providing textural information in the spatial form. The 6-channel input " + }, + { + "bbox": [ + 130, + 570, + 480, + 666 + ], + "type": "inline_equation", + "content": "[I_{t}^{AM}, I_{ref}]" + }, + { + "bbox": [ + 130, + 570, + 480, + 666 + ], + "type": "text", + "content": " is encoded to feature maps " + }, + { + "bbox": [ + 130, + 570, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{f} = \\{f_{f}, F_{1}, \\ldots, F_{L}\\}" + }, + { + "bbox": [ + 130, + 570, + 480, + 666 + ], + "type": "text", + "content": " that match the sizes of Style-based generator's each stage and " + }, + { + "bbox": [ + 130, + 570, + 480, + 666 + ], + "type": "inline_equation", + "content": "f_{f}" + }, + { + "bbox": [ + 130, + 570, + 480, + 666 + ], + "type": "text", + "content": " being a vector. Then we follow previous studies [22,73,74] and insert " + }, + { + "bbox": [ + 130, + 570, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{f}" + }, + { + "bbox": [ + 130, + 570, + 480, + 666 + ], + "type": "text", + "content": " into the generator " + }, + { + "bbox": [ + 130, + 570, + 480, + 666 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 570, + 480, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "text", + "content": "J. Guan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "text", + "content": "As " + }, + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "inline_equation", + "content": "I_{t}^{AM}" + }, + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "text", + "content": " directly provides the 3D mouth shape information, the reference frame can hardly influence the prediction result as long as it is not the target frame itself. Thus we randomly sample " + }, + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "inline_equation", + "content": "I_{ref}" + }, + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "text", + "content": " within the nearest 30 frames of " + }, + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 247 + ], + "type": "text", + "content": "A number of previous lip-sync studies [22,32,59,77,81] rely on encoding audio information into the " + }, + { + "bbox": [ + 130, + 152, + 482, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 130, + 152, + 482, + 247 + ], + "type": "text", + "content": " space for animation. However, with explicit guidance, the driving information would play a less important role in the " + }, + { + "bbox": [ + 130, + 152, + 482, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 130, + 152, + 482, + 247 + ], + "type": "text", + "content": " space. Meanwhile, the face-swapping task illustrates that " + }, + { + "bbox": [ + 130, + 152, + 482, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 130, + 152, + 482, + 247 + ], + "type": "text", + "content": " space embedding benefits identity information [35, 73, 83] which we intend to enhance. To avoid involving extra network design and inference procedures, we directly send the encoded " + }, + { + "bbox": [ + 130, + 152, + 482, + 247 + ], + "type": "inline_equation", + "content": "f_{f}" + }, + { + "bbox": [ + 130, + 152, + 482, + 247 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 152, + 482, + 247 + ], + "type": "inline_equation", + "content": "[I_t^{AM}, I_{ref}]" + }, + { + "bbox": [ + 130, + 152, + 482, + 247 + ], + "type": "text", + "content": " to the " + }, + { + "bbox": [ + 130, + 152, + 482, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 130, + 152, + 482, + 247 + ], + "type": "text", + "content": ", so that the reference frame provides more facial detail while " + }, + { + "bbox": [ + 130, + 152, + 482, + 247 + ], + "type": "inline_equation", + "content": "I_t^{AM}" + }, + { + "bbox": [ + 130, + 152, + 482, + 247 + ], + "type": "text", + "content": " also passes dynamic information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 247, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 308 + ], + "type": "text", + "content": "Loss Functions. To sum up, in the second stage, the generator takes the output only of a single facial encoder and predicts the target frame itself as " + }, + { + "bbox": [ + 130, + 247, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\hat{I}_t" + }, + { + "bbox": [ + 130, + 247, + 482, + 308 + ], + "type": "text", + "content": ". The training objective is much simpler with the 3D guidance. We only adopt the same loss functions as Pix2pix-HD [65], which are the " + }, + { + "bbox": [ + 130, + 247, + 482, + 308 + ], + "type": "inline_equation", + "content": "L_1" + }, + { + "bbox": [ + 130, + 247, + 482, + 308 + ], + "type": "text", + "content": ", VGG, and adversarial loss empowered with StyleGAN2's discriminator." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 187, + 316, + 481, + 348 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 316, + 481, + 348 + ], + "spans": [ + { + "bbox": [ + 187, + 316, + 481, + 348 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e c} = \\left\\| \\hat {I} _ {t} - I _ {t} \\right\\| _ {1} + \\lambda_ {2} \\sum_ {m = 1} ^ {N _ {v g g}} \\left\\| \\operatorname {V G G} _ {m} (\\hat {I} _ {t}) - \\operatorname {V G G} _ {m} (I _ {t}) \\right\\| _ {1}, \\tag {1}", + "image_path": "d6309e84c65406613835af0a65b6165c57a8dc92a37782085e391e216f1657cb.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 192, + 367, + 480, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 367, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 192, + 367, + 480, + 384 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a d v} = \\underset {\\mathrm {G}} {\\operatorname * {m i n m a x}} _ {\\mathrm {D}} \\left(\\mathbb {E} _ {I _ {t}} \\left[ \\log \\mathrm {D} \\left(I _ {t}\\right) \\right] + \\mathbb {E} _ {\\hat {I} _ {t}} \\left[ \\log \\left(1 - \\mathrm {D} \\left(\\hat {I} _ {t}\\right)\\right) \\right]\\right). \\tag {2}", + "image_path": "d97e4948677b513fd5c158c3f64574aa4910176bf1a6d454a6635e32cb780443.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 391, + 479, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 479, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 479, + 415 + ], + "type": "text", + "content": "As the rendered meshes provide much clearer guidance than audio, the generator can be easily trained without task-oriented losses or specific tuning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 415, + 479, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 479, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 479, + 461 + ], + "type": "text", + "content": "During inference, we overlay the predicted mesh " + }, + { + "bbox": [ + 130, + 415, + 479, + 461 + ], + "type": "inline_equation", + "content": "\\hat{M}_t^p" + }, + { + "bbox": [ + 130, + 415, + 479, + 461 + ], + "type": "text", + "content": " from the first stage to form the input " + }, + { + "bbox": [ + 130, + 415, + 479, + 461 + ], + "type": "inline_equation", + "content": "[\\hat{I}_t^{AM}, I_{ref}]" + }, + { + "bbox": [ + 130, + 415, + 479, + 461 + ], + "type": "text", + "content": ". Though the predicted " + }, + { + "bbox": [ + 130, + 415, + 479, + 461 + ], + "type": "inline_equation", + "content": "\\hat{M}_t^p" + }, + { + "bbox": [ + 130, + 415, + 479, + 461 + ], + "type": "text", + "content": " cannot perfectly match the reconstructed results, we rely on the powerful generator to compensate for such disturbance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 463, + 479, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 479, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 479, + 498 + ], + "type": "text", + "content": "Video-Driven Lip-Sync. The generator only depends on mesh information, thus we can transfer the displacements of meshes " + }, + { + "bbox": [ + 130, + 463, + 479, + 498 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{M}'" + }, + { + "bbox": [ + 130, + 463, + 479, + 498 + ], + "type": "text", + "content": " of another person to the target template " + }, + { + "bbox": [ + 130, + 463, + 479, + 498 + ], + "type": "inline_equation", + "content": "\\bar{M}" + }, + { + "bbox": [ + 130, + 463, + 479, + 498 + ], + "type": "text", + "content": " so that video-driven lip-sync can be naturally achieved." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 498, + 479, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 479, + 546 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 479, + 546 + ], + "type": "text", + "content": "Personalized Fine-tuning. Given thatSYNC information is derived from the 3D facial mesh, the generator in the second stage will not overfit the limited lip patterns present in given clips. This masks our method full-parameter fine-tunable on only seconds of videos." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 563, + 236, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 563, + 236, + 575 + ], + "spans": [ + { + "bbox": [ + 131, + 563, + 236, + 575 + ], + "type": "text", + "content": "3.3 Face-Swapping" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "We now illustrate the power of our comprehensive framework to directly achieve face-swapping and lip-sync with unified training. Our design of leveraging the " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " space for enhancing facial generation considers the similar setting used in previous face swapping studies [35, 72, 73]. Moreover, the identity information of the audio-predicted mesh " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{M}_t" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " can be easily transferred by sending different template meshes into the style encoder of the Style-SyncFormer. So that the shape of facial organs can be seamlessly swapped to another person." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 407, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 407, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 407, + 91, + 447, + 102 + ], + "type": "text", + "content": "ReSyncer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 185, + 114, + 427, + 247 + ], + "blocks": [ + { + "bbox": [ + 185, + 114, + 427, + 247 + ], + "lines": [ + { + "bbox": [ + 185, + 114, + 427, + 247 + ], + "spans": [ + { + "bbox": [ + 185, + 114, + 427, + 247 + ], + "type": "image", + "image_path": "5e05de95fed44f81f507b1eeeba50ef9faff99bc4619ac190956d9a3b898e764.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 251, + 482, + 274 + ], + "lines": [ + { + "bbox": [ + 131, + 251, + 482, + 274 + ], + "spans": [ + { + "bbox": [ + 131, + 251, + 482, + 274 + ], + "type": "text", + "content": "Fig. 3: Face-Swapping Pipeline. With input data reconfiguration and additional training losses, we achieve lip-syncing and face-swapping simultaneously." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 296, + 482, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 367 + ], + "type": "text", + "content": "The Face-Swapping Data Stream. The pipeline is depicted in Fig. 3. We first enlarge the masking area " + }, + { + "bbox": [ + 130, + 296, + 482, + 367 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 296, + 482, + 367 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 296, + 482, + 367 + ], + "type": "inline_equation", + "content": "A'" + }, + { + "bbox": [ + 130, + 296, + 482, + 367 + ], + "type": "text", + "content": " that contains the whole facial part including eyes and noses, which is more suitable for the concerned task. Then we involve a frame " + }, + { + "bbox": [ + 130, + 296, + 482, + 367 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 130, + 296, + 482, + 367 + ], + "type": "text", + "content": " of another identity and its corresponding mesh " + }, + { + "bbox": [ + 130, + 296, + 482, + 367 + ], + "type": "inline_equation", + "content": "M'" + }, + { + "bbox": [ + 130, + 296, + 482, + 367 + ], + "type": "text", + "content": " to provide additional information. We project " + }, + { + "bbox": [ + 130, + 296, + 482, + 367 + ], + "type": "inline_equation", + "content": "M'" + }, + { + "bbox": [ + 130, + 296, + 482, + 367 + ], + "type": "text", + "content": " with the pose of " + }, + { + "bbox": [ + 130, + 296, + 482, + 367 + ], + "type": "inline_equation", + "content": "M_t^p" + }, + { + "bbox": [ + 130, + 296, + 482, + 367 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 296, + 482, + 367 + ], + "type": "inline_equation", + "content": "M'^p" + }, + { + "bbox": [ + 130, + 296, + 482, + 367 + ], + "type": "text", + "content": " and overlay it on " + }, + { + "bbox": [ + 130, + 296, + 482, + 367 + ], + "type": "inline_equation", + "content": "I_t" + }, + { + "bbox": [ + 130, + 296, + 482, + 367 + ], + "type": "text", + "content": " to form " + }, + { + "bbox": [ + 130, + 296, + 482, + 367 + ], + "type": "inline_equation", + "content": "I_t^{A'M'} = (1 - A') * I_t + A' * M'^p" + }, + { + "bbox": [ + 130, + 296, + 482, + 367 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 368, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 368, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 131, + 368, + 482, + 456 + ], + "type": "text", + "content": "The face-swapping training consists of two parts, updating only the Style-based generator. The first is the same as Sec 3.2, which reconstructs " + }, + { + "bbox": [ + 131, + 368, + 482, + 456 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 131, + 368, + 482, + 456 + ], + "type": "text", + "content": " solely from " + }, + { + "bbox": [ + 131, + 368, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\{I_t^{AM}, I_{ref}\\}" + }, + { + "bbox": [ + 131, + 368, + 482, + 456 + ], + "type": "text", + "content": ". The second part intends to create an image " + }, + { + "bbox": [ + 131, + 368, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\hat{I}_t'" + }, + { + "bbox": [ + 131, + 368, + 482, + 456 + ], + "type": "text", + "content": ", which has " + }, + { + "bbox": [ + 131, + 368, + 482, + 456 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 131, + 368, + 482, + 456 + ], + "type": "text", + "content": "'s identity on " + }, + { + "bbox": [ + 131, + 368, + 482, + 456 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 131, + 368, + 482, + 456 + ], + "type": "text", + "content": "'s head with " + }, + { + "bbox": [ + 131, + 368, + 482, + 456 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 131, + 368, + 482, + 456 + ], + "type": "text", + "content": "'s original tone. This requires the generator G to take " + }, + { + "bbox": [ + 131, + 368, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_f' = \\{f_f'', F_1', \\dots, F_L'\\}" + }, + { + "bbox": [ + 131, + 368, + 482, + 456 + ], + "type": "text", + "content": " encoded from different sources, including " + }, + { + "bbox": [ + 131, + 368, + 482, + 456 + ], + "type": "inline_equation", + "content": "f_f''" + }, + { + "bbox": [ + 131, + 368, + 482, + 456 + ], + "type": "text", + "content": " encoded from " + }, + { + "bbox": [ + 131, + 368, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\{I_t^{A'M'}, I'\\}" + }, + { + "bbox": [ + 131, + 368, + 482, + 456 + ], + "type": "text", + "content": " for ID injection in the " + }, + { + "bbox": [ + 131, + 368, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 131, + 368, + 482, + 456 + ], + "type": "text", + "content": " space and " + }, + { + "bbox": [ + 131, + 368, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\{F_1', \\dots, F_L'\\}" + }, + { + "bbox": [ + 131, + 368, + 482, + 456 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 131, + 368, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\{I_t^{A'M'}, I_{ref}\\}" + }, + { + "bbox": [ + 131, + 368, + 482, + 456 + ], + "type": "text", + "content": " for expressing facial attributes in the noise space." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 455, + 482, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 455, + 482, + 503 + ], + "spans": [ + { + "bbox": [ + 130, + 455, + 482, + 503 + ], + "type": "text", + "content": "Training Objectives. Under the cross-identity training setting, the commonly used identity cosine distance loss " + }, + { + "bbox": [ + 130, + 455, + 482, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{id} = 1 - \\cos (E_{id}(\\hat{I}_t'),E_{id}(I'))" + }, + { + "bbox": [ + 130, + 455, + 482, + 503 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 455, + 482, + 503 + ], + "type": "inline_equation", + "content": "E_{id}" + }, + { + "bbox": [ + 130, + 455, + 482, + 503 + ], + "type": "text", + "content": " is the encoder of ArcFace [14] and feature matching loss with the original " + }, + { + "bbox": [ + 130, + 455, + 482, + 503 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 130, + 455, + 482, + 503 + ], + "type": "text", + "content": " [9,70,73] in the face-swapping task are leveraged." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 503, + 482, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 482, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 482, + 563 + ], + "type": "text", + "content": "Note that after the network is trained in the face-swapping style, which costs more time, high-quality face-swapping and lip-sync abilities can be achieved within one unified model (with the same set of weights). To the best of our knowledge, we are one of the first methods that unifies the two tasks within one model, which is efficient and beneficial to the creation of virtual performers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 581, + 230, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 581, + 230, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 581, + 230, + 594 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Although lip-sync is our main concern, we thoroughly conduct experiments from the perspectives of both lip-sync and face-swapping to show the effectiveness of our unified framework. For fair comparisons, we present results using only our generalized model, please refer to the supplementary files for personalized experiments." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "J. Guan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 320, + 343 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 320, + 343 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 320, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 320, + 343 + ], + "type": "image", + "image_path": "7bd5cd42c8e6a187f12d217d10b2d9a8684b5f03a0e77624b26071c23d3d21fc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 346, + 482, + 380 + ], + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 380 + ], + "type": "text", + "content": "Fig. 4: Qualitative Cross-Sync Results. The top row shows the lip-synced videos of the driving audio. Generation results based on the \"Template\" row should have the same lip shape as the \"Lip-Synced Video\" in the first row." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 323, + 114, + 481, + 343 + ], + "blocks": [ + { + "bbox": [ + 323, + 114, + 481, + 343 + ], + "lines": [ + { + "bbox": [ + 323, + 114, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 323, + 114, + 481, + 343 + ], + "type": "image", + "image_path": "cc4d1286ad3cbf4138659bd7db49b64ab2b5b588ead0b0ee88d574ae8fe092fc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 418, + 482, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 538 + ], + "type": "text", + "content": "Datasets. We train our model with two audio-visual datasets, HDTF [79] and VoxCeleb2 [11]. Both two datasets provide high-quality videos from a variety of identities for lip-syncing and face-swapping training. Notably, we split HDTF into train and test sets without overlap. The test set comprises 100 10-second audio-video pairs. Evaluations on lip-sync are conducted on the official test set of VoxCeleb2 and the introduced test set of HDTF. Face-swapping evaluation is performed on FaceForensics++ (FF++) [45], which contains 1000 face videos. We use the high-quality version and adopt the official video-video pairing. Apart from publicly available datasets, certain visual outcomes in our experiments are produced based on videos sourced from the Internet." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": "Implementation Details. All videos are processed at a frame rate of 25FPS, and facial alignment is conducted based on pre-detected landmarks. Subsequently, all faces are cropped to a size of " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": ". The architecture of G is built upon the success of recent methods [22,28,73,74]. During inference, we split videos into 100-frame consecutive chunks with a 20-frame overlap. Face meshes are predicted in an auto-aggressive manner with the proposed Style-SyncFormer depending on all the previously predicted results within each chunk. For face reconstruction, we train DeepFaceReconstruction [17] on the topology of both Hifi3DFace [2] and BFM. The main results are reported from the Hifi3DFace version but the choice of topology makes little difference. Our models are trained" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "text", + "content": "ReSyncer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 138, + 479, + 235 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": "Table 1: Quantitative results on HDTF and VoxCeleb2. For LMD and " + }, + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": "Sync the lower the better, and the higher the better for others." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 138, + 479, + 235 + ], + "lines": [ + { + "bbox": [ + 135, + 138, + 479, + 235 + ], + "spans": [ + { + "bbox": [ + 135, + 138, + 479, + 235 + ], + "type": "table", + "html": "
MethodHDTFVoxCeleb2
SSIM ↑PSNR ↑LMD ↓ΔSync ↓SSIM ↑PSNR ↑LMD ↓ΔSync ↓
Wav2Lip [41]0.8430.935.411.370.8431.583.442.28
ReTalking [10]0.8131.086.060.740.8331.644.421.50
SadTalker [78]0.7730.866.763.620.7931.365.033.03
StyleSync [22]0.8331.574.970.800.8532.523.281.57
Ours-Sync0.8431.764.340.660.8532.883.190.88
", + "image_path": "da8b585601480c42421d681d8834c4a0cb8d017cb45bc7a46b2ad76987b5dedb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 258, + 479, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 479, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 479, + 282 + ], + "type": "text", + "content": "on 4 Tesla A100 GPUs for 2 days with batch size 16. For more implementation details, please refer to the supplementary files." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 282, + 480, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 480, + 426 + ], + "type": "text", + "content": "Comparison Methods. As for lip-sync evaluation, We compare our method with four different methods including the classic Wav2Lip [41], as well as three recent state-of-the-art (SOTA) methods: ReTalking [10], SadTalker [78], and StyleSync [22]. Note that SadTalker [78] targets talking head generation but is well-known for its good quality. All experiments are conducted on the generalized setting without fine-tuning, and the model without face-swapping training is leveraged. Then we compare our method with four studies on face swapping. The counterparts consist of SimSwap [9], InfoSwap [20], StyleSwap [73] and E4S [34]. Furthermore, we aim to evaluate whether combining two methods from the above two tasks can achieve comparable results with our full model despite the cost. Thus we construct two potential counterparts by combining SOTA methods from both areas, denoted as \"SimSwap + ReTalking\" and \"StyleSwap + StyleSync\"." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 444, + 292, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 444, + 292, + 455 + ], + "spans": [ + { + "bbox": [ + 132, + 444, + 292, + 455 + ], + "type": "text", + "content": "4.1 Quantitative Comparisons" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 464, + 490, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 490, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 490, + 522 + ], + "type": "text", + "content": "Quantitative experiments of lip-sync are conducted following the self-reconstruction setting. We refrain from directly utilizing the ground truth frame as the reference. For face-swapping, we adhere to a protocol as previous studies [30, 73], wherein we uniformly sample 10 frames from each video in " + }, + { + "bbox": [ + 130, + 464, + 490, + 522 + ], + "type": "inline_equation", + "content": "\\mathrm{FF}++" + }, + { + "bbox": [ + 130, + 464, + 490, + 522 + ], + "type": "text", + "content": ", resulting in a total of " + }, + { + "bbox": [ + 130, + 464, + 490, + 522 + ], + "type": "inline_equation", + "content": "10\\mathrm{k}" + }, + { + "bbox": [ + 130, + 464, + 490, + 522 + ], + "type": "text", + "content": " images for evaluation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 523, + 480, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 523, + 480, + 560 + ], + "spans": [ + { + "bbox": [ + 130, + 523, + 480, + 560 + ], + "type": "text", + "content": "Evaluation Metrics. Both lip-sync and face-swapping are broadly studied, we follow previous works [22, 30, 73, 81] to adopt several popular metrics. For lip-sync, we use SSIM [66] and PSNR for measuring visual similarity. Landmark" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 133, + 604, + 320, + 662 + ], + "blocks": [ + { + "bbox": [ + 132, + 581, + 477, + 602 + ], + "lines": [ + { + "bbox": [ + 132, + 581, + 477, + 602 + ], + "spans": [ + { + "bbox": [ + 132, + 581, + 477, + 602 + ], + "type": "text", + "content": "Table 2: Quantitative results evaluated on Table 3: Quantitative ablations on FF++. VoxCeleb2." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 604, + 320, + 662 + ], + "lines": [ + { + "bbox": [ + 133, + 604, + 320, + 662 + ], + "spans": [ + { + "bbox": [ + 133, + 604, + 320, + 662 + ], + "type": "table", + "html": "
MethodID Retrieval↑ ID Sim.↑ PoseError ↓ Exp.Error ↓
SimSwap [9]93.07%0.5781.365.07
InfoSwap [20]95.82%0.6352.546.99
StyleSwap [73]97.05%0.6771.565.28
EAS [34]90.21%0.4952.7811.52
Ours-Swap96.70%0.6651.315.13
", + "image_path": "404ec5b450660cc4153a2349f4ffba9df7dcbb68efd79827fe25e89e61c2615f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 322, + 604, + 476, + 662 + ], + "blocks": [ + { + "bbox": [ + 322, + 604, + 476, + 662 + ], + "lines": [ + { + "bbox": [ + 322, + 604, + 476, + 662 + ], + "spans": [ + { + "bbox": [ + 322, + 604, + 476, + 662 + ], + "type": "table", + "html": "
MethodSSIM ↑PSNR ↑LMD ↓ΔSync ↓
w/o Mesh-Inject0.8331.573.602.54
w/o Mesh-Style0.8331.803.291.02
TS=500.8532.263.350.95
TS=2000.8532.693.210.88
Ours-Sync0.8532.883.190.88
", + "image_path": "dea9476f5558f5703ee76a5db87f740b12a4293ecae199f80ad81559e8ca02b4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "text", + "content": "J. Guan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 233, + 223 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 233, + 223 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 233, + 223 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 233, + 223 + ], + "type": "image", + "image_path": "ef52110d304742805e54b3a5f5faf20df4ae4e8c02d811e12a564f8b3091bf57.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 225, + 482, + 248 + ], + "lines": [ + { + "bbox": [ + 132, + 225, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 482, + 248 + ], + "type": "text", + "content": "Fig. 5: Qualitative Results of Face-Swap. Identity-swapped results should preserve the expression and lip motion of templates." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 235, + 114, + 481, + 223 + ], + "blocks": [ + { + "bbox": [ + 235, + 114, + 481, + 223 + ], + "lines": [ + { + "bbox": [ + 235, + 114, + 481, + 223 + ], + "spans": [ + { + "bbox": [ + 235, + 114, + 481, + 223 + ], + "type": "image", + "image_path": "5dd25155ad673664f535a3379ff6724071e4a8388727d7a8ed9a73acf877fdb6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 270, + 482, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 366 + ], + "type": "text", + "content": "distances around the mouth (LMD), and SyncNet's confidence score [12] are used for lip-sync quality. Particularly, the SyncNet score only indicates how well a video aligns with the learned SyncNet model, thus we use the absolute difference (" + }, + { + "bbox": [ + 130, + 270, + 482, + 366 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 130, + 270, + 482, + 366 + ], + "type": "text", + "content": "Sync) between ground truth and prediction to assess whether results are close to the original video. For face-swapping, we leverage a pretrained face-recognition model [60] to measure the ID cosine similarity and ID retrieval scores. In addition, pose error and expression error are measured by two pretrained models [46,58] following previous work [73]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 366, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 437 + ], + "type": "text", + "content": "Lip-Sync. The evaluation results are tabulated in Table 1. Our method outperforms previous works across all metrics, especially concerning lip-sync quality. This is attributed to stylized 3D facial dynamics learned by Style-SyncFormer, which manages to recover the injected personalized speaking style of the target person in the generalized setting. In contrast, previous methods primarily rely on sync-loss [41] and tend to generate a dataset-averaged sync-style." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 437, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 486 + ], + "type": "text", + "content": "Face-Swapping. Results evaluated on " + }, + { + "bbox": [ + 130, + 437, + 482, + 486 + ], + "type": "inline_equation", + "content": "\\mathrm{FF}++" + }, + { + "bbox": [ + 130, + 437, + 482, + 486 + ], + "type": "text", + "content": " are listed in Table 2. Our model shows comparable performance on all the metrics and achieves the lowest pose error when compared with SOTA face-swap methods. Particularly, our performances on ID-related metrics are only slightly weaker than StyleSwap." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 502, + 285, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 285, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 285, + 514 + ], + "type": "text", + "content": "4.2 Qualitative Comparisons" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 592 + ], + "type": "text", + "content": "Subjective evaluation plays a pivotal role in assessing the capabilities of generative models, especially in the context of videos. We strongly encourage readers to watch our supplementary video for a comprehensive understanding. Our method also enjoys the capability to fast optimize for specific identities, resulting in superior outcomes. Please refer to the supplementary files for the personalized results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 593, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 665 + ], + "type": "text", + "content": "Lip-Sync. We demonstrate two cross-sync comparisons with SOTA methods in Fig. 4. The lip-synced results are conditioned on audio randomly sampled from the dataset. Wav2Lip has accurate lip shapes, but the mouth part is noticeably blurred. SadTalker achieves better fidelity, but it fails to present natural dynamics. VideoReTalking and StyleSync can generate precise and high-quality results; nevertheless, their results are affected by the opening jaw of the template" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 101 + ], + "type": "text", + "content": "ReSyncer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "as shown in the first column of the right figure. Moreover, they exhibit a deficiency in maintaining person-specific attributes around the mouth. Our method directly produces high-fidelity results with accurate lip motion and rich details of the mouth." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 165, + 482, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 165, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 165, + 482, + 262 + ], + "type": "text", + "content": "We present cross-sync results, showcasing the transferred speaking style in Fig. 1 and Fig. 9 (in the supplementary files). It can be seen that our method not only accurately generates lip shapes aligned with the driven video but also successfully animates the target person with the speaking style of the source person in the driving video. In addition, our method inherently supports videodriven lip-syncing as introduced in Sec. 3.2. The comparison is demonstrated in Fig. 10 (in the supplementary files), it shows that person-specific attributes can be preserved better with Style-SyncFormer guidance." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 168, + 284, + 447, + 496 + ], + "blocks": [ + { + "bbox": [ + 168, + 284, + 447, + 496 + ], + "lines": [ + { + "bbox": [ + 168, + 284, + 447, + 496 + ], + "spans": [ + { + "bbox": [ + 168, + 284, + 447, + 496 + ], + "type": "image", + "image_path": "862bbb804ef064a9f567289f633a614e12bc58e419e920f7ed6d0fb1a0965432.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 498, + 482, + 544 + ], + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 544 + ], + "type": "text", + "content": "Fig.6: Results of Face-Swapped Lip-Sync. The ID-swapped results are driven by the given audio. We compare with the combinations of two SOTA methods in lip-sync and face-swapping. Results generated by our method preserve details with better fidelity and simultaneously maintain a similar speaking style to the source." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Face-Swapping. In Fig. 5, we illustrate two face-swapped results for comparison. The generated results in this context should preserve the original expression and lip motion of the template. From the figure, InfoSwap fails to keep a natural brightness across frames. E4S effectively modifies the entire head region but fails to integrate the head part into the background seamlessly. SimSwap and StyleSwap produce good results by preserving precise expression and identity attributes. However, an excessive emphasis on identity feature matching causes them to falter in generating details around certain areas like eyes and mouth." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "J. Guan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 168, + 146, + 442, + 200 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 4: User study of generalized lip-sync. The scores are ranged from 1 (worst) to 5 (best)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 168, + 146, + 442, + 200 + ], + "lines": [ + { + "bbox": [ + 168, + 146, + 442, + 200 + ], + "spans": [ + { + "bbox": [ + 168, + 146, + 442, + 200 + ], + "type": "table", + "html": "
MOS onWav2LipReTalkingSadTalkerStyleSyncOurs-Sync
Lip-Sync Quality3.433.092.983.464.57
Generation Quality1.482.833.133.634.67
Video Realness1.592.412.873.114.54
", + "image_path": "b6f5a84bdd511830263ccb928c1efc6b46be839ef6082420ee1cea61d4860ec4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 224, + 479, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 479, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 479, + 272 + ], + "type": "text", + "content": "In contrast, our method not only achieves effective identity transfer but also preserves rich details across the entire face. In addition, the videos generated by our method exhibit superior temporal consistency, especially in the mouth region. Please find more results in the supplementary video." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 274, + 480, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 480, + 332 + ], + "type": "text", + "content": "Furthermore, our method is capable of achieving lip-synced face-swapping conditioned on any audio and identity in a single forward process. We compare our method with two potential baselines \"SimSwap + ReTalking\" and \"StyleSwap + StyleSync\" in Fig. 6. Results generated by our method preserve details with better fidelity and simultaneously maintain a similar speaking style to the source." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 333, + 479, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 479, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 479, + 453 + ], + "type": "text", + "content": "User Study. We conducted a user study with 30 participants. We adopt the commonly used Mean Opinion Scores (MOS) rating protocol. All users give their ratings, ranging from 1 to 5. On generalized lip-sync, they are asked to evaluate: a) the quality of lip-sync. b) visual quality of the generated video, and c) whether this video looks real. The results averaged from 10 videos are grouped into Table 4. From the table, our method outperforms the four counterparts in all three aspects by clear margins, underscoring the power of our method and the substantial improvement in visual quality and realism over other existing methods. The experiments are similarly conducted with face-swapping and personalized lip-sync, please find the results in the supplementary files." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 474, + 238, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 474, + 238, + 487 + ], + "spans": [ + { + "bbox": [ + 132, + 474, + 238, + 487 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Lip-Sync. Here we evaluate four different designs on VoxCeleb2 [11] including: 1) \"w/o Mesh-Inject\": we mask-out the 3D mesh guidance from the encoded feature maps " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{F_1, \\ldots, F_L\\}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": ", thus it cannot provide detailed spatial information during generation; 2) \"w/o Mesh-Style\": 3D mesh content is removed in " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "I_t^{AM}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " to get a " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "f_f" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": ", feature vector that feed into the " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " space contains only the background and reference information; 3) \"" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "T_S = 50" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "\": we reduce the length (" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "T_S" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " in Sec. 3.1) of the speaking-style reference from 90 to 50; 4) \"" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "T_S = 200" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "\": the length is extended to 200. The results are presented in Table 3. Comparisons between the first two rows and \"Ours-Sync\" reveal the vital role played by mesh guidance in the generation process, where the sync quality is degraded to certain degrees, as indicated by LMD and " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Sync. In addition, the visual quality is also harmed as shown in Fig. 7 (a). In the bottom three rows of Table 3, we compare the performance of different sequence lengths in the speaking-style encoding of Style-SyncFormer. The results indicate that a sequence of 90 frames (" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "4s of" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 407, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 407, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 407, + 91, + 447, + 102 + ], + "type": "text", + "content": "ReSyncer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 114, + 446, + 269 + ], + "blocks": [ + { + "bbox": [ + 167, + 114, + 446, + 269 + ], + "lines": [ + { + "bbox": [ + 167, + 114, + 446, + 269 + ], + "spans": [ + { + "bbox": [ + 167, + 114, + 446, + 269 + ], + "type": "image", + "image_path": "f0a60a64fef2a460b4b44cdbe1665b37ab6aacd3df6c18b25396c66e43cc9d33.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 271, + 482, + 307 + ], + "lines": [ + { + "bbox": [ + 130, + 271, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 482, + 307 + ], + "type": "text", + "content": "Fig. 7: Ablations. (a) 3D facial mesh provides detailed spatial guidance, empowering superior lip-syncing. (b) The mesh with shapes of facial organs also enhances ID transfer in face-swapping." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 329, + 480, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 480, + 352 + ], + "type": "text", + "content": "video) leads to good sync quality, and a further extension of the length does not enhance its performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 353, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 482, + 449 + ], + "type": "text", + "content": "Face-Swapping. In face-swapping, the predicted 3D facial mesh also provides detailed facial attributes. We provide an intuitive comparison in Fig. 7 (b). When the mesh guidance is removed from the encoded feature map (\"w/o Mesh-Inject\"), we observe a noteworthy degradation in facial dynamics, particularly in lip motion. Please refer to the supplementary video for the comparison. Simultaneously, without the guidance of mesh information, the ID-related metrics of face-swapping are also affected, as indicated at the bottom of Fig. 7 (b) and the same phenomenon applies to \"w/o Mesh-Style\"." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 467, + 311, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 467, + 311, + 480 + ], + "spans": [ + { + "bbox": [ + 132, + 467, + 311, + 480 + ], + "type": "text", + "content": "5 Discussion and Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 492, + 480, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 492, + 480, + 539 + ], + "spans": [ + { + "bbox": [ + 130, + 492, + 480, + 539 + ], + "type": "text", + "content": "Ethical Considerations. Our method possesses the capability to create fabricated talks and speeches, even with swapped identities, presenting a potential for malicious use. We will restrict the distribution of our models and the generated results exclusively to research purposes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 540, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 540, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 540, + 482, + 647 + ], + "type": "text", + "content": "Conclusion. We highlight several important properties of our ReSyncer framework: 1) Our easy reconfiguration further unveils the potential of widely studied structures so that high-quality generalized lip-sync results with mesh representation can be achieved. 2) Our framework is designed to adopt external identity information, thus we achieve face-swapping comparable to the state of the arts while keeping the lip-sync ability within one unified model. 3) Our network supports speaking style transfer, video-driven facial animation, and can be applied to real-time live streamers. These properties complementarily satisfy various needs of virtual performer creation under different circumstances." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "J. Guan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "Acknowledgments. This work is in part supported by National Natural Science Foundation of China with No. 62394322, Beijing Natural Science Foundation with No. L222024, as well as the Beijing National Research Center for Information Science and Technology (BNRist) key projects. This study is also supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 231, + 199, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 231, + 199, + 244 + ], + "spans": [ + { + "bbox": [ + 132, + 231, + 199, + 244 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 257, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 257, + 480, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 257, + 480, + 290 + ], + "spans": [ + { + "bbox": [ + 138, + 257, + 480, + 290 + ], + "type": "text", + "content": "1. Baevski, A., Zhou, Y., Mohamed, A., Auli, M.: wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in Neural Information Processing Systems (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 291, + 480, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 291, + 480, + 323 + ], + "spans": [ + { + "bbox": [ + 138, + 291, + 480, + 323 + ], + "type": "text", + "content": "2. Bao, L., Lin, X., Chen, Y., Zhang, H., Wang, S., Zhe, X., Kang, D., Huang, H., Jiang, X., Wang, J., Yu, D., Zhang, Z.: High-fidelity 3d digital human head creation from rgb-d selfies. ACM Transactions on Graphics (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 324, + 480, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 324, + 480, + 356 + ], + "spans": [ + { + "bbox": [ + 138, + 324, + 480, + 356 + ], + "type": "text", + "content": "3. Blanz, V., Vetter, T.: A morphable model for the synthesis of 3d faces. In: Proceedings of the 26th annual conference on Computer graphics and interactive techniques. pp. 187-194 (1999)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 357, + 480, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 480, + 400 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 480, + 400 + ], + "type": "text", + "content": "4. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 401, + 480, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 401, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 401, + 480, + 422 + ], + "type": "text", + "content": "5. Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: Maskgit: Masked generative image transformer. arXiv preprint arXiv:2202.04200 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 423, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 423, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 423, + 480, + 456 + ], + "type": "text", + "content": "6. Chen, L., Cui, G., Liu, C., Li, Z., Kou, Z., Xu, Y., Xu, C.: Talking-head generation with rhythmic head motion. In: European Conference on Computer Vision. pp. 35-51. Springer (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 456, + 480, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 480, + 488 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 480, + 488 + ], + "type": "text", + "content": "7. Chen, L., Li, Z., Maddox, R.K., Duan, Z., Xu, C.: Lip movements generation at a glance. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 520-535 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 489, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 489, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 489, + 480, + 521 + ], + "type": "text", + "content": "8. Chen, L., Maddox, R.K., Duan, Z., Xu, C.: Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7832-7841 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 522, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 480, + 555 + ], + "type": "text", + "content": "9. Chen, R., Chen, X., Ni, B., Ge, Y.: Simswap: An efficient framework for high fidelity face swapping. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 2003-2011 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 555, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 587 + ], + "type": "text", + "content": "0. Cheng, K., Cun, X., Zhang, Y., Xia, M., Yin, F., Zhu, M., Wang, X., Wang, J., Wang, N.: Videoretalking: Audio-based lip synchronization for talking head video editing in the wild (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 588, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 588, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 588, + 480, + 610 + ], + "type": "text", + "content": "1. Chung, J.S., Nagrani, A., Zisserman, A.: Voxceleb2: Deep speaker recognition. In: INTERSPEECH (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 611, + 480, + 632 + ], + "type": "text", + "content": "12. Chung, J.S., Zisserman, A.: Out of time: automated lip sync in the wild. In: ACCV (2016)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "text", + "content": "13. Cudeiro, D., Bolkart, T., Laidlaw, C., Ranjan, A., Black, M.J.: Capture, learning, and synthesis of 3d speaking styles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10101-10111 (2019)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 407, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 407, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 407, + 91, + 447, + 102 + ], + "type": "text", + "content": "ReSyncer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "14. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4690-4699 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 192 + ], + "type": "text", + "content": "15. Deng, Y., Yang, J., Chen, D., Wen, F., Tong, X.: Disentangled and controllable face image generation via 3d imitative-contrastive learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 192, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 192, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 133, + 192, + 480, + 225 + ], + "type": "text", + "content": "16. Deng, Y., Yang, J., Xiang, J., Tong, X.: Gram: Generative radiance manifolds for 3d-aware image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10673-10683 (June 2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 225, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 480, + 267 + ], + "type": "text", + "content": "17. Deng, Y., Yang, J., Xu, S., Chen, D., Jia, Y., Tong, X.: Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "type": "text", + "content": "18. Fan, Y., Lin, Z., Saito, J., Wang, W., Komura, T.: Faceformer: Speech-driven 3d facial animation with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18770-18780 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 300, + 480, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 343 + ], + "type": "text", + "content": "19. Gafni, G., Thies, J., Zollhöfer, M., Nießner, M.: Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8649-8658 (June 2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 343, + 480, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 480, + 375 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 480, + 375 + ], + "type": "text", + "content": "20. Gao, G., Huang, H., Fu, C., Li, Z., He, R.: Information bottleneck disentanglement for identity swapping. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3404-3413 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "type": "text", + "content": "21. Goodfellow, I.J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. arXiv preprint arXiv:1406.2661 (2014)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 407, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 450 + ], + "type": "text", + "content": "22. Guan, J., Zhang, Z., Zhou, H., Hu, T., Wang, K., He, D., Feng, H., Liu, J., Ding, E., Liu, Z., Wang, J.: Stylesync: High-fidelity generalized and personalized lip sync in style-based generator. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 450, + 480, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 483 + ], + "type": "text", + "content": "23. Guo, Y., Chen, K., Liang, S., Liu, Y., Bao, H., Zhang, J.: Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 483, + 480, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 480, + 515 + ], + "type": "text", + "content": "24. Jamaludin, A., Chung, J.S., Zisserman, A.: You said that?: Synthesising talking faces from audio. International Journal of Computer Vision 127(11), 1767-1779 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 515, + 480, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 480, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 480, + 548 + ], + "type": "text", + "content": "25. Ji, X., Zhou, H., Wang, K., Wu, Q., Wu, W., Xu, F., Cao, X.: Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. SIGGRAPH (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 548, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 480, + 579 + ], + "type": "text", + "content": "26. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4401-4410 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 579, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 480, + 601 + ], + "type": "text", + "content": "27. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of StyleGAN. In: Proc. CVPR (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "text", + "content": "28. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8110-8119 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "29. Li, L., Wang, S., Zhang, Z., Ding, Y., Zheng, Y., Yu, X., Fan, C.: Write-a-speaker: Text-based emotional and rhythmic talking-head generation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1911-1920 (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "J. Guan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "text", + "content": "30. Li, L., Bao, J., Yang, H., Chen, D., Wen, F.: Faceshifter: Towards high fidelity and occlusion aware face swapping. arXiv preprint arXiv:1912.13457 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 481, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 481, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 481, + 170 + ], + "type": "text", + "content": "31. Li, Y., Ma, C., Yan, Y., Zhu, W., Yang, X.: 3d-aware face swapping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12705-12714 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 480, + 213 + ], + "type": "text", + "content": "32. Liang, B., Pan, Y., Guo, Z., Zhou, H., Hong, Z., Han, X., Han, J., Liu, J., Ding, E., Wang, J.: Expressive talking head generation with granular audio-visual control. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3387-3396 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 214, + 480, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 480, + 235 + ], + "type": "text", + "content": "33. Liu, X., Xu, Y., Wu, Q., Zhou, H., Wu, W., Zhou, B.: Semantic-aware implicit neural audio-driven video portrait generation. ECCV (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 235, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 235, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 235, + 480, + 267 + ], + "type": "text", + "content": "34. Liu, Z., Li, M., Zhang, Y., Wang, C., Zhang, Q., Wang, J., Nie, Y.: Fine-grained face swapping via regional gan inversion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8578-8587 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "type": "text", + "content": "35. Luo, Y., Zhu, J., He, K., Chu, W., Tai, Y., Wang, C., Yan, J.: Styleface: Towards identity-disentangled face generation on megapixels. In: European Conference on Computer Vision. pp. 297-312. Springer (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "type": "text", + "content": "36. Ma, Y., Wang, S., Ding, Y., Ma, B., Lv, T., Fan, C., Hu, Z., Deng, Z., Yu, X.: Talkclip: Talking head generation with text-guided expressive speaking styles. arXiv preprint arXiv:2304.00334 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 332, + 480, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 353 + ], + "type": "text", + "content": "37. Ma, Y., Wang, S., Hu, Z., Fan, C., Lv, T., Ding, Y., Deng, Z., Yu, X.: Styletalk: One-shot talking head generation with controllable speaking styles. AAAI (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 353, + 480, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 480, + 384 + ], + "type": "text", + "content": "38. Natsume, R., Yatagawa, T., Morishima, S.: Rsgan: Face swapping and editing using face and hair representation in latent spaces. arxiv 2018. arXiv preprint arXiv:1804.03447 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 384, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 384, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 384, + 480, + 429 + ], + "type": "text", + "content": "39. Natsume, R., Yatagawa, T., Morishima, S.: Fsnet: An identity-aware generative model for image-based face swapping. In: Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part VI 14. pp. 117–132. Springer (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 429, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 480, + 471 + ], + "type": "text", + "content": "40. Park, S.J., Kim, M., Hong, J., Choi, J., Ro, Y.M.: Synctalkface: Talking face generation with precise lip-syncing via audio-lip memory. In: AAAI Conference on Artificial Intelligence. Association for the Advancement of Artificial Intelligence (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "type": "text", + "content": "41. Prajwal, K., Mukhopadhyay, R., Namboodiri, V.P., Jawahar, C.: A lip sync expert is all you need for speech to lip generation in the wild. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 484-492 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "text", + "content": "42. Qiu, H., Chen, Z., Jiang, Y., Zhou, H., Fan, X., Yang, L., Wu, W., Liu, Z.: Relitalk: Relightable talking portrait generation from a single video. International Journal of Computer Vision pp. 1-16 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "text", + "content": "43. Richard, A., Zollhöfer, M., Wen, Y., de la Torre, F., Sheikh, Y.: Meshtalk: 3d face animation from speech using cross-modality disentanglement. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "text", + "content": "44. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "text", + "content": "45. Rössler, A., Cozzolino, D., Verdoliva, L., Riess, C., Thies, J., Nießner, M.: FaceForensics++: Learning to detect manipulated facial images. In: International Conference on Computer Vision (ICCV) (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "46. Ruiz, N., Chong, E., Rehg, J.M.: Fine-grained head pose estimation without keypoints. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 2074-2083 (2018)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 407, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 407, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 407, + 91, + 447, + 102 + ], + "type": "text", + "content": "ReSyncer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "47. Shen, S., Li, W., Zhu, Z., Duan, Y., Zhou, J., Lu, J.: Learning dynamic facial radiance fields for few-shot talking head synthesis. In: ECCV (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 481, + 183 + ], + "type": "text", + "content": "48. Shen, S., Zhao, W., Meng, Z., Li, W., Zhu, Z., Zhou, J., Lu, J.: Difftalk: Crafting diffusion models for generalized audio-driven portraits animation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1982-1991 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 480, + 205 + ], + "type": "text", + "content": "49. Song, Y., Zhu, J., Li, D., Wang, X., Qi, H.: Talking face generation by conditional recurrent adversarial network. arXiv preprint arXiv:1804.04786 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 207, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 480, + 239 + ], + "type": "text", + "content": "50. Stypulkowski, M., Vougioukas, K., He, S., Zieba, M., Petridis, S., Pantic, M.: Diffused heads: Diffusion models beat gans on talking-face generation. arXiv preprint arXiv:2301.03396 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 480, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 480, + 274 + ], + "type": "text", + "content": "51. Sun, Y., Zhou, H., Wang, K., Wu, Q., Hong, Z., Liu, J., Ding, E., Wang, J., Liu, Z., Hideki, K.: Masked lip-sync prediction by audio-visual contextual exploitation in transformers. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 480, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 480, + 306 + ], + "type": "text", + "content": "52. Suwajanakorn, S., Seitz, S.M., Kemelmacher-Shlizerman, I.: Synthesizing obama: learning lip sync from audio. ACM Transactions on Graphics (ToG) 36(4), 1-13 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 308, + 480, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 480, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 480, + 340 + ], + "type": "text", + "content": "53. Tang, J., Wang, K., Zhou, H., Chen, X., He, D., Hu, T., Liu, J., Zeng, G., Wang, J.: Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 342, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 373 + ], + "type": "text", + "content": "54. Thambiraja, B., Habibie, I., Aliakbarian, S., Cosker, D., Theobalt, C., Thies, J.: Imitator: Personalized speech-driven 3d facial animation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 20621-20631 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "type": "text", + "content": "55. Thies, J., Elgharib, M., Tewari, A., Theobalt, C., Nießner, M.: Neural voice puppetry: Audio-driven facial reenactment. In: European Conference on Computer Vision. pp. 716-731. Springer (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 409, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 409, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 409, + 480, + 430 + ], + "type": "text", + "content": "56. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 431, + 480, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 480, + 463 + ], + "type": "text", + "content": "57. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 464, + 480, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 480, + 497 + ], + "type": "text", + "content": "58. Vemulapalli, R., Agarwala, A.: A compact embedding for facial expression similarity. In: proceedings of the IEEE/cvf conference on computer vision and pattern recognition. pp. 5683-5692 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 498, + 480, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 480, + 529 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 480, + 529 + ], + "type": "text", + "content": "59. Wang, D., Deng, Y., Yin, Z., Shum, H.Y., Wang, B.: Progressive disentangled representation learning for fine-grained controllable talking head synthesis. CVPR (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 532, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 480, + 564 + ], + "type": "text", + "content": "60. Wang, H., Wang, Y., Zhou, Z., Ji, X., Gong, D., Zhou, J., Li, Z., Liu, W.: Cosface: Large margin cosine loss for deep face recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5265-5274 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 566, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 566, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 566, + 480, + 597 + ], + "type": "text", + "content": "61. Wang, J., Zhao, K., Ma, Y., Zhang, S., Zhang, Y., Shen, Y., Zhao, D., Zhou, J.: Facecomposer: A unified model for versatile facial content creation. In: Thirty-seventh Conference on Neural Information Processing Systems (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 599, + 480, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 480, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 480, + 631 + ], + "type": "text", + "content": "62. Wang, K., Wu, Q., Song, L., Yang, Z., Wu, W., Qian, C., He, R., Qiao, Y., Loy, C.C.: Mead: A large-scale audio-visual dataset for emotional talking-face generation. In: ECCV (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "63. Wang, S., Li, L., Ding, Y., Fan, C., Yu, X.: Audio2head: Audio-driven one-shot talking-head generation with natural head motion. arXiv preprint arXiv:2107.09293 (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "text", + "content": "J. Guan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "64. Wang, S., Li, L., Ding, Y., Yu, X.: One-shot talking face generation from single-speaker audio-visual correlation learning. In: Proceedings of the AAAI Conference on Artificial Intelligence (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 194 + ], + "type": "text", + "content": "65. Wang, T.C., Liu, M.Y., Zhu, J.Y., Tao, A., Kautz, J., Catanzaro, B.: High-resolution image synthesis and semantic manipulation with conditional gans. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8798-8807 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "type": "text", + "content": "66. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "type": "text", + "content": "67. Wu, H., Jia, J., Wang, H., Dou, Y., Duan, C., Deng, Q.: Imitating arbitrary talking style for realistic audio-driven talking face synthesis. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 1478-1486 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 259, + 482, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 303 + ], + "type": "text", + "content": "68. Wu, X., Hu, P., Wu, Y., Lyu, X., Cao, Y.P., Shan, Y., Yang, W., Sun, Z., Qi, X.: Speech2lip: High-fidelity speech to lip generation by learning from a short video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22168-22177 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 303, + 482, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 346 + ], + "type": "text", + "content": "69. Xing, J., Xia, M., Zhang, Y., Cun, X., Wang, J., Wong, T.T.: Codetalker: Speechdriven 3d facial animation with discrete motion prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12780-12790 (June 2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 346, + 482, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 380 + ], + "type": "text", + "content": "70. Xu, C., Zhang, J., Hua, M., He, Q., Yi, Z., Liu, Y.: Region-aware face swapping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7632-7641 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "type": "text", + "content": "71. Xu, Y., Deng, B., Wang, J., Jing, Y., Pan, J., He, S.: High-resolution face swapping via latent semantics disentanglement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7642-7651 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 413, + 482, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 445 + ], + "type": "text", + "content": "72. Xu, Z., Hong, Z., Ding, C., Zhu, Z., Han, J., Liu, J., Ding, E.: Mobilefaceswap: A lightweight framework for video face swapping. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 2973-2981 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 445, + 482, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 479 + ], + "type": "text", + "content": "73. Xu, Z., Zhou, H., Hong, Z., Liu, Z., Liu, J., Guo, Z., Han, J., Liu, J., Ding, E., Wang, J.: Styleswap: Style-based generator empowers robust face swapping. In: European Conference on Computer Vision. pp. 661-677. Springer (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 479, + 482, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 512 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 512 + ], + "type": "text", + "content": "74. Yang, T., Ren, P., Xie, X., Zhang, L.: Gan prior embedded network for blind face restoration in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 672-681 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 512, + 482, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 482, + 544 + ], + "type": "text", + "content": "75. Yao, S., Zhong, R., Yan, Y., Zhai, G., Yang, X.: Dfa-nerf: Personalized talking head generation via disentangled face attributes neural rendering. arXiv preprint arXiv:2201.00791 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 544, + 482, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 567 + ], + "type": "text", + "content": "76. Ye, Z., Jiang, Z., Ren, Y., Liu, J., He, J., Zhao, Z.: Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. ICLR (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 567, + 482, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 482, + 588 + ], + "type": "text", + "content": "77. Yu, Z., Yin, Z., Zhou, D., Wang, D., Wong, F., Wang, B.: Talking head generation with probabilistic audio-to-visual diffusion priors. ICCV (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 588, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 588, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 588, + 482, + 632 + ], + "type": "text", + "content": "78. Zhang, W., Cun, X., Wang, X., Zhang, Y., Shen, X., Guo, Y., Shan, Y., Wang, F.: Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8652-8661 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "text", + "content": "79. Zhang, Z., Li, L., Ding, Y., Fan, C.: Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3661-3670 (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "text", + "content": "ReSyncer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 259 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "text", + "content": "80. Zhou, H., Liu, Y., Liu, Z., Luo, P., Wang, X.: Talking face generation by adversarially disentangled audio-visual representation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 33, pp. 9299-9306 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 194 + ], + "type": "text", + "content": "81. Zhou, H., Sun, Y., Wu, W., Loy, C.C., Wang, X., Liu, Z.: Pose-controllable talking face generation by implicitly modularized audio-visual representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4176-4186 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "82. Zhou, Y., Han, X., Shechtman, E., Echevarria, J., Kalogerakis, E., Li, D.: Makelttalk: speaker-aware talking-head animation. ACM Transactions on Graphics (TOG) 39(6), 1-15 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "text", + "content": "83. Zhu, Y., Li, Q., Wang, J., Xu, C.Z., Sun, Z.: One shot face swapping on megapixels. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4834-4844 (2021)" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "J. Guan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Real Appearance Modeling for More General Deepfake Detection/2bfe37cc-d80a-473f-a5d8-70fca9af4b74_content_list.json b/2024/Real Appearance Modeling for More General Deepfake Detection/2bfe37cc-d80a-473f-a5d8-70fca9af4b74_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..93939407a03770e78277e92440bb112e4f8d6ba9 --- /dev/null +++ b/2024/Real Appearance Modeling for More General Deepfake Detection/2bfe37cc-d80a-473f-a5d8-70fca9af4b74_content_list.json @@ -0,0 +1,1627 @@ +[ + { + "type": "text", + "text": "Real Appearance Modeling for More General Deepfake Detection", + "text_level": 1, + "bbox": [ + 243, + 140, + 759, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiahe Tian $^{1,2}$ , Cai Yu $^{1,2}$ , Xi Wang $^{3}$ , Peng Chen $^{4}$ , Zihao Xiao $^{4}$ , Jiao Dai $^{2}$ , Jizhong Han $^{2}$ , and Yesheng Chai $^{2}$", + "bbox": [ + 215, + 210, + 785, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Cyber Security, University of Chinese Academy of Sciences, Beijing 100049, China", + "bbox": [ + 236, + 253, + 763, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Institute of Information Engineering, Chinese Academy of Sciences, Beijing 100094, China", + "bbox": [ + 215, + 282, + 782, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Institute of Microelectronics, Chinese Academy of Sciences, Beijing 100094, China \n $^{4}$ RealAI Inc., Beijing 100094, China", + "bbox": [ + 220, + 310, + 781, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Recent studies in deepfake detection have shown promising results when detecting deepfakes of the same type as those present in training. However, their ability to generalize to unseen deepfakes remains limited. This work improves the generalizable deepfake detection from a simple principle: an ideal detector classifies any face that contains anomalies not found in real faces as fake. Namely, detectors should learn consistent real appearances rather than fake patterns in the training set that may not apply to unseen deepfakes. Guided by this principle, we propose a learning task named Real Appearance Modeling (RAM) that guides the model to learn real appearances by recovering original faces from slightly disturbed faces. We further propose Face Disturbance to produce disturbed faces while preserving original information that enables recovery, which aids the model in learning the fine-grained appearance of real faces. Extensive experiments demonstrate the effectiveness of modeling real appearances to spot richer deepfakes. Our method surpasses existing state-of-the-art methods by a large margin on multiple popular deepfake datasets.", + "bbox": [ + 261, + 378, + 740, + 614 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Deepfake Detection $\\cdot$ Face Forgery Detection $\\cdot$ Multimedia Forensic", + "bbox": [ + 259, + 628, + 740, + 656 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 685, + 375, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current deep generative models allow for the generation of realistic fake faces, i.e., deepfakes, making it easy to manipulate face media. Therefore, these techniques can be used for committing fraud, bypassing identity authentication, or spreading false information. The potential misuses raise concerns regarding privacy and reputation, and thus deepfake detectors are needed to counter the risks posed by deepfakes. Moreover, the proliferation of deepfake algorithms presents a challenge to the generalizability of deepfake detection, making it crucial to develop detectors that generalize across various deepfake types.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6ad10ae9a0e37430415dd94175066ffc246a2fe2c628cc6663d6fbfa5b6c7579.jpg", + "image_caption": [ + "(a) Deepfake Detection." + ], + "image_footnote": [], + "bbox": [ + 253, + 146, + 418, + 257 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0878a1942adfae22bff02aa922e0bd116c7c34a1d7aa9e07ebca2190f79d3046.jpg", + "image_caption": [ + "(b) The RAM task.", + "Fig. 1: Illustration of (a) the Deepfake Detection task and (b) the proposed RAM task. Compared to Deepfake Detection, the RAM task guides models to restore disturbed faces back to their original appearance." + ], + "image_footnote": [], + "bbox": [ + 433, + 155, + 754, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous works usually tackled deepfake detection by guiding detectors to recognize specific deepfake patterns, such as the boundaries in swapped faces [1, 2], inconsistency [3, 4], frequency anomalies [5, 6], and movement anomalies [7, 8]. They achieved good performance in detecting deepfakes of the same types as those present in training. However, their performance often significantly deteriorates when detecting deepfakes generated by novel algorithms, which mirrors real-world scenarios where various deepfakes emerge. This is due to different deepfake algorithms producing different fake patterns, and learned patterns are not universally applicable to all deepfakes. Namely, more than knowledge of specific deepfake patterns is required to generalize to unseen types of deepfake.", + "bbox": [ + 212, + 357, + 787, + 507 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Though different deepfakes exhibit different patterns, real faces maintain consistent appearance. Thus, it is better to learn the appearance of real faces to verify their authenticity instead of learning deepfake patterns to spot fakes. Namely, an ideal classifier would classify faces as fake if they exhibit patterns not found in real faces, i.e., anomalies. As the detector learns explicit real face appearance, it should be capable of predicting real faces at pixel-level. This guides us to design a task that densely predicts real faces to learn real appearance. To fully define this task, dense prediction targets and input information that assists the prediction are needed. Specifically, we propose the Real Appearance Modeling (RAM) task that recovers corresponding original faces from disturbed faces, as illustrated in Figure 5, where original faces are prediction targets and the inputted disturbed faces contain information needed for prediction. For the RAM task, Deepfakes are not applicable to serve as training data since their original identities or poses are altered, and the necessary information that enables prediction is lost. We thus propose Face Disturbance to disturb real faces to pseudo-fakes as training data for RAM. In the disturbed faces produced by Face Disturbance, the original information is mostly retained. Then, the detector learns real face appearance by predicting original faces.", + "bbox": [ + 212, + 507, + 787, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To disturb real faces, the proposed Face Disturbance includes two parts, Texture and Structure Disturbance, to introduce anomalies in textures and structures, respectively. This is based on the premise that an image can be regarded as composed of structures and textures [9,10]. Texture Disturbance disturbs partial", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Tian et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "textures to produce anomalies in the form of texture inconsistency. The reason for disturbing partially rather than globally is that distribution of real textures varies, and globally modifying facial textures may fail to produce anomalies. Meanwhile, having inconsistent textures between different facial parts is indeed anomalous. Hence, Texture Disturbance is designed to produce inconsistencies. Structure Disturbance disturbs structure by covering key facial areas, including facial features and boundaries, which are essential components of facial structure. The specific covering method involves copying and tiling the surrounding texture to the disturbed areas to ensure the original texture is intact. Jointly using Texture and Structure Disturbance to produce disturbed faces for RAM helps the detector to learn real structures and textures to spot anomalies in unseen deepfakes.", + "bbox": [ + 212, + 146, + 787, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proposed RAM uses disturbed faces produced by Face Disturbance as inputs and the original faces as prediction targets, making this task unsupervised. Therefore, RAM is adopted as an auxiliary task to enhance deepfake detectors. In training, real and disturbed faces are used as inputs, and the detector is expected to both discriminate disturbed faces and recover their real appearances. Specifically, RAM involves guiding the detector to recover structure and texture anomalies. For structure anomalies, the detector recovers the structure of the disturbed areas with the undisturbed areas as the reference. For texture anomalies, the detector recovers the inner texture based on the texture of the outer face. By leveraging learned real appearances, our model is equipped to detect anomalies within unseen deepfakes. As illustrated in Figure 5, our model diverges from the baseline model by grouping various deepfakes to general anomalies rather than identifying specific deepfake patterns. This indicates RAM narrows the gap across deepfakes and thus improves the generalizability.", + "bbox": [ + 212, + 328, + 787, + 540 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 238, + 541, + 593, + 553 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce Real Appearance Modeling to guide detectors to learn the appearance of real faces by restoring disturbed faces to their original appearances, which helps to improve the generalization of deepfake detectors.", + "- We propose Face Disturbance to disturb real faces in texture and structure views as pseudo-fakes, which aids the detector in learning real texture and structure appearance when predicting original faces.", + "- We conduct extensive experiments using various evaluation protocols to demonstrate the generalizability of our framework, which surpasses previous methods on multiple popular deepfake datasets." + ], + "bbox": [ + 214, + 555, + 784, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 717, + 387, + 732 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Deepfake Detection", + "text_level": 1, + "bbox": [ + 215, + 752, + 423, + 766 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Early works leveraged biological artifacts, including distorted pupils [11], heartbeat frequency [12], and anomalous head pose [13] for deepfake detection. More recent works treated deepfake detection as a binary classification task and focused on detecting specific deepfake patterns. For instance, [14] investigated", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Real Appearance Modeling for More General Deepfake Detection", + "bbox": [ + 297, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mesoscopic features in shallow networks that contain rich features, while [15] employed an attention mechanism to combine RGB and texture features. [6, 16] leveraged high-frequency in the discrete cosine domain as complementary modalities. Another direction involves employing proxy tasks for pretraining to learn high-level representations capable of identifying deepfake videos. Examples of such proxy tasks include lip-reading [8] and audio-visual contrastive learning [17, 18]. Recently, the unsupervised learning object MAE [19] is introduced in deepfake detection [20, 21]. These works usually recover masked facial features through dense prediction to model real faces. As they use masked images as input where information is lost, the set recovery target, i.e., original face, is merely one of the multiple feasible solutions. In contrast, RAM for each disturbed face has a definitive prediction target and thus is not an ill-posed problem with multiple solutions. Compared to MAE, RAM is more fine-grained in supervising detectors to model real face appearance.", + "bbox": [ + 212, + 146, + 787, + 358 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Pseudo-fakes for Deepfake Detection", + "text_level": 1, + "bbox": [ + 215, + 380, + 565, + 395 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instead of training detectors using deepfakes in deepfake datasets, several studies [1,2,4,22-24] suggested constructing pseudo-fakes for training. Since the introduced pattern in pseudo-fakes is controllable, detectors learn prior patterns accurately. Specifically, they introduced texture inconsistencies to simulate swapped faces and guide detectors to learn this prior pattern. For instance, [1] introduced blur to inner faces to replicate clarity inconsistencies. [2,4,22] blended two faces with similar facial landmarks to generate texture inconsistencies. Moreover, [23, 24] utilized augmentations to produce inconsistencies within a single face, achieving good generalization. Although they generalize well to face swap deepfakes, their performance on deepfakes generated by reenactment algorithms [25, 26] or Diffusion model [27] drops where texture inconsistency is not the major artifacts.", + "bbox": [ + 212, + 405, + 787, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 609, + 330, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Different from most deepfake detection methods, which are typically trained to spot specific deepfake patterns, our approach emphasizes learning real face appearances through the RAM task. The Face Disturbance strategy is adopted to produce disturbed faces as training data for RAM, as illustrated in Figure 2. The subsequent subsections delve into the details of the proposed Face Disturbance strategy, the definition of the RAM task, and the design of the proposed Recovery Autoencoder (RAE) model in Figure 3, which simultaneously conducts deepfake detection and RAM in a multi-task learning manner.", + "bbox": [ + 212, + 641, + 787, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Face Disturbance", + "text_level": 1, + "bbox": [ + 215, + 784, + 406, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We aim to disturb real faces to pseudo-fake faces as training data. Previous works devoted to producing texture inconsistency, which is a typical pattern of", + "bbox": [ + 215, + 809, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Tian et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fad3aa5bf056e606ed10f05253b34c2b9abee510e0f9d24b29d676cf3d0f149e.jpg", + "image_caption": [ + "(a) Face Disturbance" + ], + "image_footnote": [], + "bbox": [ + 241, + 148, + 763, + 291 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f6da5cef8653f48697403a959843065768f8a89185217e9c49ecec0db11cde69.jpg", + "image_caption": [ + "(b) Pseudo-fake Examples", + "Fig. 2: (a) Face Disturbance with TD and SD. (b) Pseudo-fake examples, where for each character, from left to right, are real faces, faces with texture anomalies $I_{ta}$ , and faces with structure anomalies $I_{sa}$ . We highlight their disturbed areas." + ], + "image_footnote": [], + "bbox": [ + 246, + 306, + 759, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "face swap deepfakes. Therefore, these detectors exhibited biased generalization to face swap deepfakes and fell short when detecting reenactment deepfakes in [28-30]. Similarly, employing these data for RAM enables the detector to learn real texture appearance only, making the detector struggle to spot reenactment deepfakes with intact textures. Thus, it is necessary to introduce richer anomalies to help detectors handle a wider spectrum of deepfakes. As an image can be regarded as composed of structures and textures [9,10], we categorize anomalies in fake faces into two categories: structure and texture anomalies. Therefore, we implement Texture Disturbance (TD) and Structure Disturbance (SD) in Face Disturbance to produce these anomalies. By adopting this divide-and-conquer strategy, we comprehensively disturb faces and thus make the detector learn richer real appearances from both the texture and structure views.", + "bbox": [ + 212, + 441, + 787, + 623 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Texture Disturbance The aim of TD is to produce texture anomalies. Given the inherent diversity of real facial textures, which vary in color, brightness, and clarity, a reference facial area is needed to discern whether the textures in other facial areas are abnormal. Namely, texture anomalies lie in the texture inconsistency between facial parts. There are two existing approaches to produce such inconsistency: Face Blending [2, 4] using two faces and Partial Augmentation [23, 24] using one face. Face blending involves selecting two faces with similar poses and then merging the inner region of one face into the other face. Partial Augmentation involves disturbing either the inner or outer facial texture through image augmentation. An important advantage of Partial Augmentation is preserving the original identity, which prevents detectors from being overfitted to the training set identities and harming generalization [31]. Thus, the proposed TD adopts a disturbing scheme similar to Partial Augmentation.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Real Appearance Modeling for More General Deepfake Detection", + "bbox": [ + 297, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 2 illustrates the process of TD. The input image $I_{ori}$ is duplicated to $I_{inner}$ and $I_{outer}$ , and one of them is augmented in color, brightness, and clarity. $I_{inner}$ and $I_{outer}$ are then fused using a mask $M$ :", + "bbox": [ + 215, + 146, + 787, + 191 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nI _ {t a} = M \\odot I _ {\\text {i n n e r}} + (1 - M) \\odot I _ {\\text {o u t e r}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 214, + 785, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\odot$ is Hadamard production, $I_{inner}$ is the image that provides the inner face, $I_{outer}$ provides the background, and $I_{ta}$ is the fused face with texture anomaly. $M$ is the mask for the inner face, and its pixel values are between 0 and 1. Specifically, we first apply a landmark detector to the input image to predict its landmarks. Next, we calculate the convex hull of the face area in $I$ to initialize $M$ . To create inconsistent textures between a richer diversity of different areas, we introduce random deformations to $M$ using a 2D elastic transformation. Before using Equation 1 to blend $I_{inner}$ and $I_{outer}$ , we soften $M$ from a binary mask into a soft mask with Gaussian Blur to produce $I_{ta}$ with natural boundaries. This allows the detector to recognize anomalies in long-span texture inconsistencies rather than merely adapt to local boundary artifacts around $M$ .", + "bbox": [ + 215, + 248, + 787, + 415 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Structure Disturbance The aim of SD is to introduce anomalies in the structure view to complement texture anomalies. To disturb facial structures, we disturb the elements that constitute the facial structure. Clearly, these elements are facial boundaries and facial features such as the eyes and mouth. We note that facial landmarks are also positioned at these elements. Therefore, it is intuitive to disturb areas around landmarks to create disturbances in facial structures. Besides, to distinguish with TD, it is also necessary for SD not to disturb the original texture. To disturb facial structures while keeping textures intact, we copy and tile the surrounding textures into areas around a few facial landmarks. The crux lies in the fact that the copy-and-tile operation preserves original textures and thus is the ideal method for disturbing structures. This aids the detector in accurately learning real facial structure in RAM, enabling them to distinguish faces with structural anomalies during detection.", + "bbox": [ + 212, + 458, + 787, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 2 illustrates the process of SD. Given an input face image $I_{ori}$ , we first detect its facial landmarks. We randomly sample several landmark points and draw circles around them as the mask $M$ for disturbed areas. We avoid sampling landmarks of symmetric positions in the face, which enables the detector to rely on the structure information from the symmetric region to perform RAM. We then employ the Fast Marching Method [32], a non-learning image inpainting algorithm, which produces less blur and maintains clearer textures than other textures replicating algorithms, to copy and tile the surrounding textures to the disturbed area. After the surrounding textures are tiled to the disturbed areas, the original structures of the disturbed areas are eliminated and become anomalous. Incorporating SD in Face Disturbance emrichs anomalies in constructed pseudo-fakes, which helps the detector generalize to a wider range of deepfakes.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Tian et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/13953c80bf2e097298f735ce20f5791c3812c1566f29c67d1c6a877705e44e31.jpg", + "image_caption": [ + "Fig. 3: The proposed RAE is optimized using two learning objects, i.e., $\\mathcal{L}_{rec}$ for the RAM and $\\mathcal{L}_{cls}$ for the deepfake detection." + ], + "image_footnote": [], + "bbox": [ + 305, + 142, + 697, + 290 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Real Appearance Modeling", + "text_level": 1, + "bbox": [ + 214, + 357, + 488, + 373 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We believe that guiding models to learn the appearances of real faces is more beneficial for the generalization across deepfake algorithms than learning specific deepfake patterns. Thus, we propose the novel Real Appearance Modeling (RAM) task to guide detectors to learn real face appearances. RAM requires the model to restore anomalous facial areas produced by Face Disturbance to their original appearances. In comparison to existing methods, our model is guided to learn patterns of real faces instead of deepfake patterns in training fake faces. Then, it classifies faces that display anomalies not found in real faces as fake. Therefore, RAM helps to improve generalization across deepfake algorithms.", + "bbox": [ + 212, + 380, + 787, + 517 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Task Definition Given an input face $I_{ori}$ , we obtain pseudo-fake faces $I_{ta}$ with texture anomalies and $I_{sa}$ with structure anomalies after the Face Disturbance strategy. These images are fed into the detector for deepfake detection and RAM, i.e., classification and densely real appearance recovery. In the training procedure, the deepfake detection object and RAM object are optimized in a multi-task learning manner. In deepfake detection, the labels for constructed pseudo-fakes are 1, while for real faces are 0. In RAM, the recovery target $I_{target}$ for $I_{sa}$ is $I$ , while the recovery target $I_{target}$ for $I_{ta}$ is $I_{outer}$ . This is because $I_{ta}$ is derived from $I_{outer}$ and $I_{inner}$ , so there are two potential recovery targets, and we fix the target as $I_{outer}$ . This allows the model to restore the inner texture from $I_{inner}$ using the background texture from $I_{outer}$ as the reference. Thus, the model is guided to align the inner texture with the outer texture. The feasibility of recovering pseudo-fakes to original faces lies in the fact that the pseudo-fakes contains all the information of the corresponding original faces, with only variations between the inner and outer face regions or partially disrupted structure.", + "bbox": [ + 212, + 537, + 787, + 765 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Recovery Autoencoder", + "text_level": 1, + "bbox": [ + 214, + 785, + 450, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In order to perform both deepfake detection and RAM, the detector in our framework requires a classification head and a dense prediction head. Thus, we", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Real Appearance Modeling for More General Deepfake Detection", + "bbox": [ + 297, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "incorporate a dense prediction branch for the RAM object into a classifier. This configuration makes our model resemble an autoencoder, thus named Recovery Autoencoder (RAE). The RAE is illustrated in Figure 3, which is constructed based on a vanilla Vision Transformer Base (ViT Base) as the encoder. We adopt eight transformer layers and a linear projection layer as the decoder, and a Multi-layer Token Fusion (MTF) block to extract features from both the shallow and deep encoder layers. We refer readers to [33] for detailed information of the ViT encoder. Following this, we detail the entire pipeline of RAE.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "After the image $I \\in R^{H \\times W \\times 3}$ is fed to the ViT encoder, it is extracted to a sequence of tokens $x_{l} \\in R^{(N + 1) \\times D}$ in each encoder layer. Here, $l \\in 1, 2, \\dots, L$ is the encoder layer index, $L = 12$ is the depth of the encoder, $N$ is the number of image patches, and $D$ represents the embedding dimension. We note $x_{L}[0]$ is the c1s token [33] for classification. The classification process involves passing the c1s token to a linear layer and a softmax operator to predict the class probabilities $\\hat{y} \\in \\mathbb{R}^2$ , and 2 is the number of classes. Then, the remaining tokens in $x_{l}$ are reshaped to $\\mathbb{R}^{h \\times w \\times D}$ where $h = w = \\sqrt{N}$ , and then inputted to MTF for cross-layer token fusion and the decoder for RAM.", + "bbox": [ + 212, + 267, + 787, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As there are 12 transformer layers in the ViT encoder, inputting all layers in $x_{l}$ to MTF is computationally intensive. Thus, we select $S = 3$ layers at equal intervals and tokens in these layers as the inputs to MTF to provide shallow, mid, and deep features. Formally, the selected tokens of shallow, mid, and deep layers are noted as $x_{s}, x_{m}$ , and $x_{d}$ . In MTF, we first project $x_{s}, x_{m}$ , and $x_{d}$ into the same feature space and concatenate them to $x^{\\prime} \\in R^{S \\times h \\times w \\times D}$ . We employ a linear layer as the projection layer, which is widely used for aligning different spaces. After projection, we further extract and fuse both structural and textural information from different layers. As tokens in different ViT layers have identical dimensions, we adopt a cross-layer attention mechanism for fusion. Specifically, MTF predicts weights $W \\in R^{S \\times h \\times w}$ for each token with an MLP. $W$ is normalized using softmax along the first dimension $S$ , ensuring that the weights for tokens at the same spatial position across $S$ layers sum to 1. $W$ and fused tokens $x_{fus} \\in R^{h \\times w \\times D}$ are derived as:", + "bbox": [ + 212, + 404, + 787, + 616 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {\\prime} = \\operatorname {c o n c a t} \\left(x _ {s}, x _ {m}, x _ {d}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 419, + 628, + 784, + 643 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nW = \\operatorname {s o f t m a x} (\\operatorname {M L P} \\left(x ^ {\\prime}\\right)), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 648, + 784, + 662 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nx _ {f u s} = \\sum_ {i = 0} ^ {S} W [ i ] \\times x ^ {\\prime} [ i ]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 667, + 784, + 693 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "After extracting $x_{fus}$ , we aim to recover the original face with $x_{fus}$ . While it might be intuitive to directly feed $x_{fus}$ into the decoder to recover real faces, this leads to easy mining that outputs the inputted pseudo-fakes directly. This is because the inputted pseudo-fakes and the target real faces are identical in the undisturbed areas $1 - M$ and are similar in disturbed areas $M$ . To avoid such easy mining, we introduce a learnable mask msk token to replace tokens of the disturbed areas. Specifically, given a binary mask $M$ for disturbed areas, it is divided into mask patches $m \\in R^{h \\times w \\times p \\times p}$ , where $p$ is the side length of a patch. If a mask patch is disturbed, the corresponding patch is replaced with msk", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Tian et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "token, and the mask for masked patches is $M_{patch}$ . After masking, the masked tokens $x_{msk} \\in R^{h \\times w \\times D}$ are added with positional encodings and then fed into the decoder to $\\hat{x} \\in R^{N \\times p \\times p \\times 3}$ . Lastly, through a reshape operation, we obtain the recovered face image $\\hat{I} \\in R^{H \\times W \\times 3}$ where $H = W = \\sqrt{N} \\times p$ . We emphasize that the masking is applied after encoding. Thus, the classification head and the predicted $\\hat{y}$ remain unaffected. Moreover, information in masked tokens is fused to unmasked tokens in encoding, thus necessary information that enables the recovery is still conveyed to the decoder.", + "bbox": [ + 212, + 146, + 782, + 263 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In predicting the real appearance of disturbed areas, RAE extracts texture and structure features from undisturbed areas and semantic information from disturbed areas for RAM. Consequently, the decoder can utilize the combination of these features to model the appearance of disturbed areas through pixel-level prediction. This enhances the real appearance learned in the encoder, thereby improving the generalization of the deepfake detection.", + "bbox": [ + 212, + 267, + 784, + 357 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Loss Function", + "text_level": 1, + "bbox": [ + 215, + 378, + 375, + 393 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The proposed RAE is optimized to both discriminate disturbed faces and model real appearance. We use cross-entropy loss $\\mathcal{L}_{cls}$ for classification:", + "bbox": [ + 214, + 404, + 782, + 434 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c l s} = - y \\log \\hat {y} - (1 - y) \\log (1 - \\hat {y}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 446, + 784, + 460 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\hat{y}$ is the predicted class probabilities and $y$ is the class label. Besides, we use the following L1-loss $\\mathcal{L}_{rec}$ for RAM:", + "bbox": [ + 214, + 473, + 782, + 503 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e c} = \\frac {1}{\\left\\| M _ {p a t c h} \\right\\| _ {1}} \\left\\| M _ {p a t c h} \\odot (\\hat {I} - I _ {t a r g e t}) \\right\\| _ {1}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 513, + 782, + 546 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\| \\cdot \\|_1$ is the L1-norm, $M_{patch}$ is the mask for disturbed tokens and thus $\\mathcal{L}_{rec}$ is performed on the disturbed tokens. The total loss $\\mathcal{L}$ for optimizing RAE is formulated as the weighted sum of $\\mathcal{L}_{cls}$ and $\\mathcal{L}_{rec}$ :", + "bbox": [ + 214, + 556, + 782, + 601 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {c l s} + \\alpha \\mathcal {L} _ {r e c}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 437, + 614, + 782, + 628 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We set $\\alpha$ as 0.1 to make $\\mathcal{L}_{cls}$ and $\\mathcal{L}_{rec}$ converge to the same order of magnitude.", + "bbox": [ + 214, + 640, + 782, + 655 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 678, + 375, + 695 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 709, + 455, + 724 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Dataset. We conduct evaluations on six widely used deepfake datasets, i.e., FaceForensics++ (FF++) [34], Celeb-DF v2 (CDF) [35], DFD [34], DFDC [36], DFDC preview (DFDCp) [36], and WildDeepfake (WDF) [37]. FF++ is a deep-fake dataset containing 4,000 forged videos and 1,000 real videos. All videos are provided in three compression levels: raw, high-quality (HQ), and low-quality (LQ). CDF contains 590 real and 5,639 fake videos corresponding to 59 celebrities using an improved face swap algorithm. DFD contains over 3,000 face swap", + "bbox": [ + 214, + 734, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Real Appearance Modeling for More General Deepfake Detection", + "bbox": [ + 297, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "videos. DFDC is a large-scale dataset that contains more than 120,000 deepfake clips generated using eight deepfake techniques. DFDCp is the preview dataset for the Deepfake Detection Challenge. Its fake videos are generated using two face swap algorithms. WDF consists of 10,000 real videos and 10,000 online collected fake videos. For DFDC, we use videos with one identity following [8,17,38] in case of labeling noise. For other datasets, we use the official dataset splits. We use the real videos of $\\mathrm{FF} + + (\\mathrm{HQ})$ to construct pseudo-fakes for training, while other datasets are for testing.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Preprocess. We extract 10 frames at equal temporal intervals from each video in training and 32 frames from each video in testing. Then, the faces are cropped using Retinaface [39], maintaining a $15\\%$ margin around each face. Subsequently, we utilize Dlib [40] to extract 81 facial landmarks for every face and save them to a key-value database to boost the Face Disturbance process.", + "bbox": [ + 212, + 267, + 787, + 343 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Face Disturbance Details. As for the texture augmentations involved in generating $I_{ta}$ , we employ RGBShift, HueSaturationValue, RandomBrightnessContrast, Downscale, and Sharpen to augment textures of either $I_{inner}$ or $I_{outer}$ to produce inconsistencies. As for the disturbed areas in generating $I_{sa}$ , we draw circles around 2 to 6 random landmarks with a radius sampled between 1/14 and 1/16 of the side length. The number and size of the sampled circles for SD ensure have been found empirically to aid in the effectiveness of RAM tasks as shown in supplementary. Then, all images are resized to $224 \\times 224$ as inputs.", + "bbox": [ + 212, + 343, + 787, + 464 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training Details. We use a ViT Base pretrained on ImageNet [41] to initialize our RAE. RAE is trained using SAM [42] optimizer with a learning rate of 8e-4 and a batch size of 192 for 100 epochs, and we use cosine decay as the learning rate scheduler. Each batch consists of 96 real faces, 48 disturbed faces with texture anomalies, and 48 disturbed faces with structure anomalies. In the PyTorch framework, training with automatic mixed precision requires approximately 60GB of GPU memory.", + "bbox": [ + 212, + 464, + 787, + 570 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Cross-Dataset Evaluation", + "text_level": 1, + "bbox": [ + 215, + 590, + 472, + 604 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In cross-dataset evaluation, detectors are trained using faces in $\\mathrm{FF}++(\\mathrm{HQ})$ and evaluated on other datasets to assess their generalization to unknown deepfake techniques. Here, we compare the proposed RAE with recent deepfake detection methods. The results are listed in Table 1, and we annotated the input type and training faces. RAE achieves top generalization performance and outperforms state-of-the-art detectors on CDF, DFDC, and DFDCp by $2.6\\%$ , $3.4\\%$ , and $3.2\\%$ , respectively. Compared to detectors trained using fake faces, such as TALL and IDDDM, the superior generalization of RAE demonstrates the effectiveness of using Face Disturbance to generate pseudo-fakes for training. Compared to detectors trained using pseudo-fakes with texture inconsistency, such as SBI and AltFreezing, RAE exhibits better generalization on face swap datasets such as CDF and DFDCp. As the design of TD is similar to previous works, the effectiveness of RAE against face swap deepfakes indicates RAM helps to spot richer anomalies through learning real appearance. Moreover, previous detectors trained with pseudo-fakes fail to exhibit better performance on", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Tian et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/33551d403954f5fab6745e52b20ce183d6f3c5f68431ed9af11afb692cc37b1c.jpg", + "table_caption": [ + "Table 1: Cross-dataset evaluation of deepfake detection methods on CDF, DFD, DFDC, DFDCP, and WDF. All models are trained using FF++(HQ). All compared results are video-level AUC and cited from the original papers or their subsequences." + ], + "table_footnote": [], + "table_body": "
MethodInput TypeYearTraining SetTest Set AUC (%)
Real FakePseudo-fakeCDFDFDDFDCDFDCpWDF
PCL+I2G [4]Frame2021✓✓90.099.067.574.4-
ENB4+SBI [23]Frame2022✓✓92.998.272.085.5-
Mover [20]Frame2023✓✓87.1--78.982.0
RFFR [21]Frame2023✓✓89.0-67.8--
IDDDM [31]Frame2023✓✓91.2-71.5--
SeeABLE [43]Frame2023✓✓87.3-75.986.3-
FTCN [38]Clip2021✓✓86.994.471.074.0-
RealForensics [17]Clip2022✓✓86.9-75.9--
AltFreezing [24]Clip2023✓✓✓89.598.5---
TALL [44]Clip2023✓✓90.8-76.8--
RAE (Ours)Frame2024✓✓95.599.080.289.588.4
", + "bbox": [ + 220, + 198, + 782, + 368 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/5bde5efbb1cea909bccdf5a05f8466fdfda7bac5822d51f398ac05fd7a32e992.jpg", + "table_caption": [ + "Table 2: In-dataset evaluation on four subsets in FF++(HQ). †indicates reproduced on FF++(HQ)." + ], + "table_footnote": [], + "table_body": "
MethodTest Set AUC(%)
DFF2FFSNTAvg
SLADD [22]----98.4
ENB4+SBI † [23]99.299.199.096.798.5
SeeABLE [43]99.298.899.196.998.5
RAE(Ours)99.699.199.297.698.9
", + "bbox": [ + 223, + 436, + 498, + 516 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/df2047e4ac64919ad9ff2cf51f45fdeaa16fc29881565378c384cecf78e26b18.jpg", + "table_caption": [ + "Table 3: Frame-level FP rate(%) on real face datasets. Lower FP is better. All models are trained on $\\mathrm{FF} + + (\\mathrm{HQ})$" + ], + "table_footnote": [], + "table_body": "
MethodTest Set FP(%)↓
FFHQVoxCeleb
RECCE [45]62.341.7
UIA-ViT [46]40.618.9
ENB4+SBI [23]57.759.2
RAE(ours)36.96.9
", + "bbox": [ + 540, + 436, + 754, + 516 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "DFDC as it contains reenacted faces that are difficult to be imitated by previous pseudo-fakes. Our method demonstrates a major improvement compared to previous methods as SD helps RAE spot structure anomalies in reenacted faces.", + "bbox": [ + 215, + 546, + 782, + 592 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 In-dataset Evaluation", + "text_level": 1, + "bbox": [ + 215, + 616, + 439, + 630 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In addition to cross-dataset evaluation, another frequently used evaluation protocol is in-dataset evaluation. We follow the protocol in [43] to conduct evaluations on the four types of deepfakes in $\\mathrm{FF}++$ , i.e., two types of face swap deepfakes DF and FS, and two types of face reenactment deepfakes F2F and NT, where the training data includes both deepfake and pseudo-fake samples. This evaluation protocol helps to assess the ability of our method in enhancing in-dataset deepfake detection. The results are listed in Table 2. The overall performance of RAE on all four deepfake types surpasses previous works. Specifically, RAE outperforms the compared methods in detecting reenacted faces generated by the two reenactment algorithms, i.e., F2F and NT. RAE achieves a detection AUC of $97.6\\%$ on NT. This indicates that combining SD and the learning target RAM aids in recognizing distorted facial features in NT deepfakes, while previous works trained with texture inconsistency pseudo-fakes fail.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Real Appearance Modeling for More General Deepfake Detection", + "bbox": [ + 297, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Evaluating RAE on Real Face Datasets", + "text_level": 1, + "bbox": [ + 215, + 146, + 584, + 161 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In practical deployment, a low False Positive rate (FP) is important for a deepfake detector, which indicates a low ratio of misclassifying real faces as fakes. For instance, a detector with a low FP rate in an authentication system helps to avoid having users repeatedly submit their portraits, thereby enhancing the user experience. Besides, at the same AUC, a lower FP helps reduce the adjustment of classification thresholds in practical applications. We use the mega-scale face dataset VoxCeleb [47], and the high-quality face dataset FFHQ [48] to test the FP of the proposed RAE on real faces. For comparison, we used three open-source detectors, RECCE [45], SBI [23], and UIA-ViT [46].", + "bbox": [ + 212, + 174, + 785, + 310 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The FP scores are listed in Table 3. Compared to existing arts, RAE achieves the lowest FP on these two datasets. In stark contrast, the average FP scores of the compared RECCE and SBI are larger than $50\\%$ , which indicates they misclassified more than half real faces in these two datasets. This validates that RAM successfully makes the detector learn real face appearances and the trained RAE is less likely to classify a real face as fake.", + "bbox": [ + 212, + 311, + 785, + 401 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Evaluation on Diffusion Deepfakes", + "text_level": 1, + "bbox": [ + 215, + 426, + 542, + 441 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In the cross-dataset evaluation, we validated the generalizability of our framework across five popular deepfake datasets. However, we noticed these datasets primarily consist of face swap and reenactment deepfakes. As a result, the effectiveness of detectors against other types of deepfakes, such as diffusion-generated faces, was not adequately evaluated. Given the fact that a proliferating number of deepfakes are diffusion-generated faces, we conduct evaluations on a diffusion face dataset DeepFakeFace(DFF) [27]. DFF is synthesized by Stable Diffusion Text2Img (T2I), Stable Diffusion Inpaint (Inpaint) [49], and InsightFace (IF) [50]. We conduct evaluation using the three subsets in DFF. For all compared methods, we use their official weights.", + "bbox": [ + 212, + 454, + 785, + 604 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The evaluation results are listed in Table 4. RAE outperforms the compared methods in all three subsets in DFF. More significantly, RAE attains an AUC of $70.8\\%$ on average, which improves existing SOTA methods by $2.2\\%$ . In stark contrast, the AUC scores of RECCE and UIA-ViT on T2I are less than $50\\%$ , which indicates they perform worse than randomly guessing. The results on DFF verify the effectiveness of our proposed method in enhancing the detector's generalization to a wider range of deepfakes.", + "bbox": [ + 212, + 606, + 785, + 710 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.6 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 737, + 387, + 752 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conduct ablations to verify the effectiveness of the RAM task, Face Disturbance strategy, and the MTF block. A ViT Base trained using $I_{ta}$ and classification task is used to serve as the Baseline as it is similar to several previous detectors [1, 23]. We incrementally add the proposed RAM task, SD, and MTF block to the baseline to observe their effectiveness. The results are listed in Table", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Tian et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/08ed0605677945e088fd4d01a3388343a9b0be28f544c5937e5772811a785ee3.jpg", + "table_caption": [ + "Table 4: Frame-level AUC on three subsets of diffusion deepfakes in DFF [27]. All models are trained on $\\mathrm{FF} + + (\\mathrm{HQ})$" + ], + "table_footnote": [], + "table_body": "
MethodTest Set AUC(%)
T2IInpaintIFAvg
RECCE [45]35.151.563.149.9
UIA-ViT [46]45.255.874.058.3
ENB4+SBI [23]62.871.471.568.6
RAE(ours)64.173.075.270.8
", + "bbox": [ + 222, + 198, + 500, + 287 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/14cffd9f061b32f64c874313f1983133ffd073fb2036b398d68728d3affd95b7.jpg", + "table_caption": [ + "Table 5: Ablation studies for the proposed Structure Disturbance, RAM task, and MTF module." + ], + "table_footnote": [], + "table_body": "
MethodTest set AUC(%)
CDFDFDCDFDCp
Baseline92.575.585.9
+RAM94.277.488.6
+Isa94.779.188.4
+MTF95.580.289.5
", + "bbox": [ + 524, + 199, + 771, + 287 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2b2b527d290679c571776faed27b7bb5ee6ebe4fd08c699608eb8c1c0f125442.jpg", + "image_caption": [ + "Fig. 4: Visualizations of RAM. The left are recovery results for $I_{sa}$ while the right are results for $I_{ta}$ ." + ], + "image_footnote": [], + "bbox": [ + 243, + 303, + 759, + 450 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5, which show that each of them contributes to enhancing generalization. Compared to solely using a classification task, adding the RAM task significantly improves generalization by $2.2\\%$ on average. Furthermore, utilizing $I_{sa}$ to model real facial structures mainly improves the generalization on the DFDC dataset, which includes reenactment deepfakes. Besides, the MTF block aids in improving generalization by leveraging multi-layer features for modeling real appearances.", + "bbox": [ + 212, + 523, + 787, + 613 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.7 Visualization", + "text_level": 1, + "bbox": [ + 215, + 642, + 370, + 656 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Face Recovery. We visualize the recovered disturbed faces to verify whether RAE learned real face appearances. The inputted disturbed faces, corresponding original faces, masks for disturbed facial areas, and recovered faces are presented in Figure 4. We observe that our model successfully recovers most of the perturbed areas within the anomalous faces. Though the recovered faces are not explicitly clear, the anomalous patterns are largely eliminated. For faces with structure anomalies, disturbed areas are finely recovered to their original structure appearance. For faces with texture anomalies, the texture pattern of the inner faces is aligned to normal appearance. This demonstrates the feasibility of modeling real appearance by recovering anomalous faces to their original appearance, which provides more fine-grained and interpretable supervision.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Real Appearance Modeling for More General Deepfake Detection", + "bbox": [ + 297, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0ba10af368b17f0890b7db15831c67a5fa13776acccfd4d5e117d27159e683b7.jpg", + "image_caption": [ + "(a) Baseline model." + ], + "image_footnote": [], + "bbox": [ + 276, + 146, + 493, + 315 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/cf9cededb448e7cda57912f606f3dc7e4b2e63d6dea057fe4f332ffe92bf1c69.jpg", + "image_caption": [ + "(b) Our model.", + "Fig. 5: Distributions of learned representations for faces in $\\mathrm{FF} + + (\\mathrm{HQ})$ [34] in the baseline model(a) and our model(b)." + ], + "image_footnote": [], + "bbox": [ + 509, + 146, + 730, + 316 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Feature Distribution. We apply t-SNE [51] to visualize the features in the last layer of the encoder to illustrate the effect of RAM. We use a ViT Base as the baseline detector and train it using real and fake faces in $\\mathrm{FF} + + (\\mathrm{HQ})$ . Then, we use the trained RAE and baseline detector to extract features from the test set of $\\mathrm{FF} + + (\\mathrm{HQ})$ . The results are illustrated in Figure 5. The baseline model learns specific features for each deepfake algorithm, thus separating deepfakes of different types in the feature space. In contrast, the proposed RAE aggregates real faces to a compact cluster in the feature space, which indicates it classifies deepfakes based on the learned general fake patterns. Moreover, RAE mixes different types of deepfakes in the feature space, which confirms that RAM helps to learn general deepfake representations. Compared to the baseline model, the anomalous features learned by RAE are more general across different deepfakes, thus demonstrating better generalization.", + "bbox": [ + 212, + 406, + 787, + 603 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 633, + 359, + 648 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we address the general deepfake detection through guiding the detector to learn real face appearance by recovering original faces from disturbed faces. Anomalies in both the texture and structure views are introduced to faces through the proposed Face Disturbance, and the detector is guided to recover real appearances from these disturbed faces through the proposed RAM task. We then propose the RAE model to conduct both the deepfake detection and the RAM task. Extensive experiments demonstrate the effectiveness of our method against existing methods in general deepfake detection. Future work involves scaling our method with large-scale human face datasets.", + "bbox": [ + 212, + 672, + 787, + 808 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This project is supported by the National Key Research and Development Program of China (No.2022YFB2702500)", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Tian et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Yuezun Li and Siwei Lyu. Exposing deepfake videos by detecting face warping artifacts. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2019, Long Beach, CA, USA, June 16-20, 2019, pages 46-52. Computer Vision Foundation / IEEE, 2019. 2, 4, 12", + "2. Lingzhi Li, Jianmin Bao, Ting Zhang, Hao Yang, Dong Chen, Fang Wen, and Baining Guo. Face x-ray for more general face forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5001-5010, 2020. 2, 4, 5", + "3. Ke Sun, Taiping Yao, Shen Chen, Shouhong Ding, Jilin Li, and Rongrong Ji. Dual contrastive learning for general face forgery detection. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, Thirty-Fourth Conference on Innovative Applications of Artificial Intelligence, IAAI 2022, The Twelveth Symposium on Educational Advances in Artificial Intelligence, EAAI 2022 Virtual Event, February 22 - March 1, 2022, pages 2316-2324. AAAI Press, 2022. 2", + "4. Tianchen Zhao, Xiang Xu, Mingze Xu, Hui Ding, Yuanjun Xiong, and Wei Xia. Learning self-consistency for deepfake detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15023-15033, 2021. 2, 4, 5, 11", + "5. Yuchen Luo, Yong Zhang, Junchi Yan, and Wei Liu. Generalizing face forgery detection with high-frequency features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16317-16326, 2021. 2", + "6. Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In European Conference on Computer Vision, pages 86-103. Springer, 2020. 2, 4", + "7. Yuezun Li, Ming-Ching Chang, and Siwei Lyu. In ictu oculi: Exposing AI created fake videos by detecting eye blinking. In 2018 IEEE International Workshop on Information Forensics and Security, WIFS 2018, Hong Kong, China, December 11-13, 2018, pages 1-7. IEEE, 2018. 2", + "8. Alexandros Haliassos, Konstantinos Vougioukas, Stavros Petridis, and Maja Pan- tic. Lips don't lie: A generalisable and robust approach to face forgery detection. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 5039-5049. Computer Vision Foundation / IEEE, 2021. 2, 4, 10", + "9. Jean-François Aujol, Guy Gilboa, Tony F. Chan, and Stanley J. Osher. Structure-texture image decomposition - modeling, algorithms, and parameter selection. Int. J. Comput. Vis., 67(1):111-136, 2006. 2, 5", + "0. Youngjung Kim, Bumsub Ham, Minh N Do, and Kwanghoon Sohn. Structure-texture image decomposition using deep variational priors. IEEE Transactions on Image Processing, 28(6):2692-2704, 2018. 2, 5", + "1. Yuezun Li, Ming-Ching Chang, and Siwei Lyu. In ictu oculi: Exposing ai created fake videos by detecting eye blinking. In 2018 IEEE International Workshop on Information Forensics and Security (WIFS), pages 1-7. IEEE, 2018. 3", + "2. Umur Aybars Ciftci, Ilke Demir, and Lijun Yin. Fakecatcher: Detection of synthetic portrait videos using biological signals. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 3", + "3. Xin Yang, Yuezun Li, and Siwei Lyu. Exposing deep fakes using inconsistent head poses. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 8261-8265. IEEE, 2019. 3" + ], + "bbox": [ + 225, + 175, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Real Appearance Modeling for More General Deepfake Detection", + "bbox": [ + 297, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Darius Afchar, Vincent Nozick, Junichi Yamagishi, and Isao Echizen. Mesonet: a compact facial video forgery detection network. In 2018 IEEE international workshop on information forensics and security (WIFS), pages 1-7. IEEE, 2018. 3", + "15. Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multi-attentional deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2185–2194, 2021. 4", + "16. Qiqi Gu, Shen Chen, Taiping Yao, Yang Chen, Shouhong Ding, and Ran Yi. Exploiting fine-grained face forgery clues via progressive enhancement learning. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, Thirty-Fourth Conference on Innovative Applications of Artificial Intelligence, IAAI 2022, The Twelveth Symposium on Educational Advances in Artificial Intelligence, EAAI 2022 Virtual Event, February 22 - March 1, 2022, pages 735-743. AAAI Press, 2022. 4", + "17. Alexandros Haliassos, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Leveraging real talking faces via self-supervision for robust forgery detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14950-14962, 2022. 4, 10, 11", + "18. Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Weiming Zhang, and Nenghai Yu. Self-supervised transformer for deepfake detection, 2022. 4", + "19. Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 4", + "20. Juan Hu, Xin Liao, Difei Gao, Satoshi Tsutsui, Qian Wang, Zheng Qin, and Mike Zheng Shou. Mover: Mask and recovery based facial part consistency aware method for deepfake video detection, 2023. 4, 11", + "21. Liang Shi, Jie Zhang, and Shiguang Shan. Real face foundation representation learning for generalized deepfake detection, 2023. 4, 11", + "22. Liang Chen, Yong Zhang, Yibing Song, Lingqiao Liu, and Jue Wang. Self-supervised learning of adversarial examples: Towards good generalizations for deep-fake detections. In CVPR, 2022. 4, 11", + "23. Kaede Shiohara and Toshihiko Yamasaki. Detecting deepfakes with self-blended images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18720-18729, 2022. 4, 5, 11, 12, 13", + "24. Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, and Houqiang Li. Altfreezing for more general video face forgery detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4129-4138, June 2023. 4, 5, 11", + "25. Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. In Hanna M. Wallach, Hugo Larochelle, Alina Beygelzimer, Florence d'Alché-Buc, Emily B. Fox, and Roman Garnett, editors, Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 7135-7145, 2019. 4", + "26. Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Zwei Liu. Pose-controllable talking face generation by implicitly modularized audiovisual representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4176-4186, 2021. 4" + ], + "bbox": [ + 215, + 146, + 787, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Tian et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "27. Haixu Song, Shiyu Huang, Yinpeng Dong, and Wei-Wei Tu. Robustness and generalizability of deepfake detection: A study with diffusion models. arXiv preprint arXiv:2309.02218, 2023. 4, 12, 13", + "28. Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton-Ferrer. The deepfake detection challenge dataset. CoRR, abs/2006.07397, 2020. 5", + "29. Gereon Fox, Wentao Liu, Hyeongwoo Kim, Hans-Peter Seidel, Mohamed Elgharib, and Christian Theobalt. Videoforensicshq: Detecting high-quality manipulated face videos. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 5", + "30. Jiajun Huang, Xueyu Wang, Bo Du, Pei Du, and Chang Xu. Deepfake mnist+: a deepfake facial animation dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1973-1982, 2021. 5", + "31. Baojin Huang, Zhongyuan Wang, Jifan Yang, Jiaxin Ai, Qin Zou, Qian Wang, and Dengpan Ye. Implicit identity driven deepfake face swapping detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4490-4499, 2023. 5, 11", + "32. Alexandru C. Telea. An image inpainting technique based on the fast marching method. J. Graphics, GPU, & Game Tools, 9(1):23-34, 2004. 6", + "33. Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 8", + "34. Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics++: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1-11, 2019. 9, 14", + "35. Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deepfake forensics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3207-3216, 2020. 9", + "36. Brian Dolhansky, Russ Howes, Ben Pflaum, Nicole Baram, and Cristian Canton-Ferrer. The deepfake detection challenge (DFDC) preview dataset. CoRR, abs/1910.08854, 2019. 9", + "37. Bojia Zi, Minghao Chang, Jingjing Chen, Xingjun Ma, and Yu-Gang Jiang. Wilddeepfake: A challenging real-world dataset for deepfake detection. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2382-2390, 2020. 9", + "38. Yinglin Zheng, Jianmin Bao, Dong Chen, Ming Zeng, and Fang Wen. Exploring temporal coherence for more general video face forgery detection. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 15024-15034. IEEE, 2021. 10, 11", + "39. Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-shot multi-level face localisation in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5203-5212, 2020. 10", + "40. Davis E King. Dlib-ml: A machine learning toolkit. The Journal of Machine Learning Research, 10:1755-1758, 2009. 10" + ], + "bbox": [ + 215, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Real Appearance Modeling for More General Deepfake Detection", + "bbox": [ + 297, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "41. Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2009), 20-25 June 2009, Miami, Florida, USA, pages 248-255. IEEE Computer Society, 2009. 10", + "42. Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 10", + "43. Nicolas Larue, Ngoc-Son Vu, Vitomir Struc, Peter Peer, and Vassilis Christophides. Seeable: Soft discrepancies and bounded contrastive learning for exposing deep-fakes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21011-21021, 2023. 11", + "44. Yuting Xu, Jian Liang, Gengyun Jia, Ziming Yang, Yanhao Zhang, and Ran He. Tall: Thumbnail layout for deepfake video detection. arXiv preprint arXiv:2307.07494, 2023. 11", + "45. Junyi Cao, Chao Ma, Taiping Yao, Shen Chen, Shouhong Ding, and Xiaokang Yang. End-to-end reconstruction-classification learning for face forgery detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4113-4122, June 2022. 11, 12, 13", + "46. Wanyi Zhuang, Qi Chu, Zhentao Tan, Qiankun Liu, Haojie Yuan, Changtao Miao, Zixiang Luo, and Nenghai Yu. Uia-vit: Unsupervised inconsistency-aware method based on vision transformer for face forgery detection. In European Conference on Computer Vision (ECCV), 2022. 11, 12, 13", + "47. Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. arXiv preprint arXiv:1806.05622, 2018. 12", + "48. Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. IEEE Trans. Pattern Anal. Mach. Intell., 43(12):4217-4228, 2021. 12", + "49. Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 12", + "50. DeepInsight. insightface. https://github.com/deepinsight/insightface/, 2021. [Online]. 12", + "51. Van Der Maaten Laurens and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9(2605):2579-2605, 2008. 14" + ], + "bbox": [ + 215, + 146, + 785, + 618 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "J. Tian et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Real Appearance Modeling for More General Deepfake Detection/2bfe37cc-d80a-473f-a5d8-70fca9af4b74_model.json b/2024/Real Appearance Modeling for More General Deepfake Detection/2bfe37cc-d80a-473f-a5d8-70fca9af4b74_model.json new file mode 100644 index 0000000000000000000000000000000000000000..29b084e506bf01c9eb04f7dd084b3fd9d2d06b75 --- /dev/null +++ b/2024/Real Appearance Modeling for More General Deepfake Detection/2bfe37cc-d80a-473f-a5d8-70fca9af4b74_model.json @@ -0,0 +1,2282 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.244, + 0.141, + 0.761, + 0.187 + ], + "angle": 0, + "content": "Real Appearance Modeling for More General Deepfake Detection" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.212, + 0.786, + 0.245 + ], + "angle": 0, + "content": "Jiahe Tian\\(^{1,2}\\), Cai Yu\\(^{1,2}\\), Xi Wang\\(^{3}\\), Peng Chen\\(^{4}\\), Zihao Xiao\\(^{4}\\), Jiao Dai\\(^{2}\\), Jizhong Han\\(^{2}\\), and Yesheng Chai\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.255, + 0.764, + 0.284 + ], + "angle": 0, + "content": "\\(^{1}\\) School of Cyber Security, University of Chinese Academy of Sciences, Beijing 100049, China" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.283, + 0.784, + 0.31 + ], + "angle": 0, + "content": "\\(^{2}\\) Institute of Information Engineering, Chinese Academy of Sciences, Beijing 100094, China" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.311, + 0.782, + 0.34 + ], + "angle": 0, + "content": "\\(^{3}\\) Institute of Microelectronics, Chinese Academy of Sciences, Beijing 100094, China \n\\(^{4}\\) RealAI Inc., Beijing 100094, China" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.38, + 0.741, + 0.616 + ], + "angle": 0, + "content": "Abstract. Recent studies in deepfake detection have shown promising results when detecting deepfakes of the same type as those present in training. However, their ability to generalize to unseen deepfakes remains limited. This work improves the generalizable deepfake detection from a simple principle: an ideal detector classifies any face that contains anomalies not found in real faces as fake. Namely, detectors should learn consistent real appearances rather than fake patterns in the training set that may not apply to unseen deepfakes. Guided by this principle, we propose a learning task named Real Appearance Modeling (RAM) that guides the model to learn real appearances by recovering original faces from slightly disturbed faces. We further propose Face Disturbance to produce disturbed faces while preserving original information that enables recovery, which aids the model in learning the fine-grained appearance of real faces. Extensive experiments demonstrate the effectiveness of modeling real appearances to spot richer deepfakes. Our method surpasses existing state-of-the-art methods by a large margin on multiple popular deepfake datasets." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.63, + 0.741, + 0.657 + ], + "angle": 0, + "content": "Keywords: Deepfake Detection \\(\\cdot\\) Face Forgery Detection \\(\\cdot\\) Multimedia Forensic" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.686, + 0.377, + 0.701 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Current deep generative models allow for the generation of realistic fake faces, i.e., deepfakes, making it easy to manipulate face media. Therefore, these techniques can be used for committing fraud, bypassing identity authentication, or spreading false information. The potential misuses raise concerns regarding privacy and reputation, and thus deepfake detectors are needed to counter the risks posed by deepfakes. Moreover, the proliferation of deepfake algorithms presents a challenge to the generalizability of deepfake detection, making it crucial to develop detectors that generalize across various deepfake types." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "J. Tian et al." + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.147, + 0.419, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.268, + 0.261, + 0.408, + 0.274 + ], + "angle": 0, + "content": "(a) Deepfake Detection." + }, + { + "type": "image", + "bbox": [ + 0.434, + 0.156, + 0.756, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.261, + 0.649, + 0.274 + ], + "angle": 0, + "content": "(b) The RAM task." + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.285, + 0.788, + 0.327 + ], + "angle": 0, + "content": "Fig. 1: Illustration of (a) the Deepfake Detection task and (b) the proposed RAM task. Compared to Deepfake Detection, the RAM task guides models to restore disturbed faces back to their original appearance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.788, + 0.508 + ], + "angle": 0, + "content": "Previous works usually tackled deepfake detection by guiding detectors to recognize specific deepfake patterns, such as the boundaries in swapped faces [1, 2], inconsistency [3, 4], frequency anomalies [5, 6], and movement anomalies [7, 8]. They achieved good performance in detecting deepfakes of the same types as those present in training. However, their performance often significantly deteriorates when detecting deepfakes generated by novel algorithms, which mirrors real-world scenarios where various deepfakes emerge. This is due to different deepfake algorithms producing different fake patterns, and learned patterns are not universally applicable to all deepfakes. Namely, more than knowledge of specific deepfake patterns is required to generalize to unseen types of deepfake." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Though different deepfakes exhibit different patterns, real faces maintain consistent appearance. Thus, it is better to learn the appearance of real faces to verify their authenticity instead of learning deepfake patterns to spot fakes. Namely, an ideal classifier would classify faces as fake if they exhibit patterns not found in real faces, i.e., anomalies. As the detector learns explicit real face appearance, it should be capable of predicting real faces at pixel-level. This guides us to design a task that densely predicts real faces to learn real appearance. To fully define this task, dense prediction targets and input information that assists the prediction are needed. Specifically, we propose the Real Appearance Modeling (RAM) task that recovers corresponding original faces from disturbed faces, as illustrated in Figure 5, where original faces are prediction targets and the inputted disturbed faces contain information needed for prediction. For the RAM task, Deepfakes are not applicable to serve as training data since their original identities or poses are altered, and the necessary information that enables prediction is lost. We thus propose Face Disturbance to disturb real faces to pseudo-fakes as training data for RAM. In the disturbed faces produced by Face Disturbance, the original information is mostly retained. Then, the detector learns real face appearance by predicting original faces." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To disturb real faces, the proposed Face Disturbance includes two parts, Texture and Structure Disturbance, to introduce anomalies in textures and structures, respectively. This is based on the premise that an image can be regarded as composed of structures and textures [9,10]. Texture Disturbance disturbs partial" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.298, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Real Appearance Modeling for More General Deepfake Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.327 + ], + "angle": 0, + "content": "textures to produce anomalies in the form of texture inconsistency. The reason for disturbing partially rather than globally is that distribution of real textures varies, and globally modifying facial textures may fail to produce anomalies. Meanwhile, having inconsistent textures between different facial parts is indeed anomalous. Hence, Texture Disturbance is designed to produce inconsistencies. Structure Disturbance disturbs structure by covering key facial areas, including facial features and boundaries, which are essential components of facial structure. The specific covering method involves copying and tiling the surrounding texture to the disturbed areas to ensure the original texture is intact. Jointly using Texture and Structure Disturbance to produce disturbed faces for RAM helps the detector to learn real structures and textures to spot anomalies in unseen deepfakes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.788, + 0.541 + ], + "angle": 0, + "content": "The proposed RAM uses disturbed faces produced by Face Disturbance as inputs and the original faces as prediction targets, making this task unsupervised. Therefore, RAM is adopted as an auxiliary task to enhance deepfake detectors. In training, real and disturbed faces are used as inputs, and the detector is expected to both discriminate disturbed faces and recover their real appearances. Specifically, RAM involves guiding the detector to recover structure and texture anomalies. For structure anomalies, the detector recovers the structure of the disturbed areas with the undisturbed areas as the reference. For texture anomalies, the detector recovers the inner texture based on the texture of the outer face. By leveraging learned real appearances, our model is equipped to detect anomalies within unseen deepfakes. As illustrated in Figure 5, our model diverges from the baseline model by grouping various deepfakes to general anomalies rather than identifying specific deepfake patterns. This indicates RAM narrows the gap across deepfakes and thus improves the generalizability." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.542, + 0.594, + 0.554 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.556, + 0.784, + 0.6 + ], + "angle": 0, + "content": "- We introduce Real Appearance Modeling to guide detectors to learn the appearance of real faces by restoring disturbed faces to their original appearances, which helps to improve the generalization of deepfake detectors." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.602, + 0.785, + 0.647 + ], + "angle": 0, + "content": "- We propose Face Disturbance to disturb real faces in texture and structure views as pseudo-fakes, which aids the detector in learning real texture and structure appearance when predicting original faces." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.648, + 0.785, + 0.692 + ], + "angle": 0, + "content": "- We conduct extensive experiments using various evaluation protocols to demonstrate the generalizability of our framework, which surpasses previous methods on multiple popular deepfake datasets." + }, + { + "type": "list", + "bbox": [ + 0.215, + 0.556, + 0.785, + 0.692 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.718, + 0.388, + 0.733 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.753, + 0.424, + 0.767 + ], + "angle": 0, + "content": "2.1 Deepfake Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Early works leveraged biological artifacts, including distorted pupils [11], heartbeat frequency [12], and anomalous head pose [13] for deepfake detection. More recent works treated deepfake detection as a binary classification task and focused on detecting specific deepfake patterns. For instance, [14] investigated" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "J. Tian et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.359 + ], + "angle": 0, + "content": "mesoscopic features in shallow networks that contain rich features, while [15] employed an attention mechanism to combine RGB and texture features. [6, 16] leveraged high-frequency in the discrete cosine domain as complementary modalities. Another direction involves employing proxy tasks for pretraining to learn high-level representations capable of identifying deepfake videos. Examples of such proxy tasks include lip-reading [8] and audio-visual contrastive learning [17, 18]. Recently, the unsupervised learning object MAE [19] is introduced in deepfake detection [20, 21]. These works usually recover masked facial features through dense prediction to model real faces. As they use masked images as input where information is lost, the set recovery target, i.e., original face, is merely one of the multiple feasible solutions. In contrast, RAM for each disturbed face has a definitive prediction target and thus is not an ill-posed problem with multiple solutions. Compared to MAE, RAM is more fine-grained in supervising detectors to model real face appearance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.381, + 0.566, + 0.396 + ], + "angle": 0, + "content": "2.2 Pseudo-fakes for Deepfake Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.789, + 0.589 + ], + "angle": 0, + "content": "Instead of training detectors using deepfakes in deepfake datasets, several studies [1,2,4,22-24] suggested constructing pseudo-fakes for training. Since the introduced pattern in pseudo-fakes is controllable, detectors learn prior patterns accurately. Specifically, they introduced texture inconsistencies to simulate swapped faces and guide detectors to learn this prior pattern. For instance, [1] introduced blur to inner faces to replicate clarity inconsistencies. [2,4,22] blended two faces with similar facial landmarks to generate texture inconsistencies. Moreover, [23, 24] utilized augmentations to produce inconsistencies within a single face, achieving good generalization. Although they generalize well to face swap deepfakes, their performance on deepfakes generated by reenactment algorithms [25, 26] or Diffusion model [27] drops where texture inconsistency is not the major artifacts." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.61, + 0.331, + 0.626 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.788, + 0.764 + ], + "angle": 0, + "content": "Different from most deepfake detection methods, which are typically trained to spot specific deepfake patterns, our approach emphasizes learning real face appearances through the RAM task. The Face Disturbance strategy is adopted to produce disturbed faces as training data for RAM, as illustrated in Figure 2. The subsequent subsections delve into the details of the proposed Face Disturbance strategy, the definition of the RAM task, and the design of the proposed Recovery Autoencoder (RAE) model in Figure 3, which simultaneously conducts deepfake detection and RAM in a multi-task learning manner." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.785, + 0.407, + 0.799 + ], + "angle": 0, + "content": "3.1 Face Disturbance" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We aim to disturb real faces to pseudo-fake faces as training data. Previous works devoted to producing texture inconsistency, which is a typical pattern of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.298, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Real Appearance Modeling for More General Deepfake Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.149, + 0.764, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.44, + 0.294, + 0.565, + 0.306 + ], + "angle": 0, + "content": "(a) Face Disturbance" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.308, + 0.761, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.348, + 0.58, + 0.361 + ], + "angle": 0, + "content": "(b) Pseudo-fake Examples" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.372, + 0.788, + 0.414 + ], + "angle": 0, + "content": "Fig. 2: (a) Face Disturbance with TD and SD. (b) Pseudo-fake examples, where for each character, from left to right, are real faces, faces with texture anomalies \\( I_{ta} \\), and faces with structure anomalies \\( I_{sa} \\). We highlight their disturbed areas." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.442, + 0.789, + 0.624 + ], + "angle": 0, + "content": "face swap deepfakes. Therefore, these detectors exhibited biased generalization to face swap deepfakes and fell short when detecting reenactment deepfakes in [28-30]. Similarly, employing these data for RAM enables the detector to learn real texture appearance only, making the detector struggle to spot reenactment deepfakes with intact textures. Thus, it is necessary to introduce richer anomalies to help detectors handle a wider spectrum of deepfakes. As an image can be regarded as composed of structures and textures [9,10], we categorize anomalies in fake faces into two categories: structure and texture anomalies. Therefore, we implement Texture Disturbance (TD) and Structure Disturbance (SD) in Face Disturbance to produce these anomalies. By adopting this divide-and-conquer strategy, we comprehensively disturb faces and thus make the detector learn richer real appearances from both the texture and structure views." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Texture Disturbance The aim of TD is to produce texture anomalies. Given the inherent diversity of real facial textures, which vary in color, brightness, and clarity, a reference facial area is needed to discern whether the textures in other facial areas are abnormal. Namely, texture anomalies lie in the texture inconsistency between facial parts. There are two existing approaches to produce such inconsistency: Face Blending [2, 4] using two faces and Partial Augmentation [23, 24] using one face. Face blending involves selecting two faces with similar poses and then merging the inner region of one face into the other face. Partial Augmentation involves disturbing either the inner or outer facial texture through image augmentation. An important advantage of Partial Augmentation is preserving the original identity, which prevents detectors from being overfitted to the training set identities and harming generalization [31]. Thus, the proposed TD adopts a disturbing scheme similar to Partial Augmentation." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "J. Tian et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "Figure 2 illustrates the process of TD. The input image \\( I_{ori} \\) is duplicated to \\( I_{inner} \\) and \\( I_{outer} \\), and one of them is augmented in color, brightness, and clarity. \\( I_{inner} \\) and \\( I_{outer} \\) are then fused using a mask \\( M \\):" + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.215, + 0.786, + 0.23 + ], + "angle": 0, + "content": "\\[\nI _ {t a} = M \\odot I _ {\\text {i n n e r}} + (1 - M) \\odot I _ {\\text {o u t e r}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.249, + 0.789, + 0.416 + ], + "angle": 0, + "content": "where \\(\\odot\\) is Hadamard production, \\(I_{inner}\\) is the image that provides the inner face, \\(I_{outer}\\) provides the background, and \\(I_{ta}\\) is the fused face with texture anomaly. \\(M\\) is the mask for the inner face, and its pixel values are between 0 and 1. Specifically, we first apply a landmark detector to the input image to predict its landmarks. Next, we calculate the convex hull of the face area in \\(I\\) to initialize \\(M\\). To create inconsistent textures between a richer diversity of different areas, we introduce random deformations to \\(M\\) using a 2D elastic transformation. Before using Equation 1 to blend \\(I_{inner}\\) and \\(I_{outer}\\), we soften \\(M\\) from a binary mask into a soft mask with Gaussian Blur to produce \\(I_{ta}\\) with natural boundaries. This allows the detector to recognize anomalies in long-span texture inconsistencies rather than merely adapt to local boundary artifacts around \\(M\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.655 + ], + "angle": 0, + "content": "Structure Disturbance The aim of SD is to introduce anomalies in the structure view to complement texture anomalies. To disturb facial structures, we disturb the elements that constitute the facial structure. Clearly, these elements are facial boundaries and facial features such as the eyes and mouth. We note that facial landmarks are also positioned at these elements. Therefore, it is intuitive to disturb areas around landmarks to create disturbances in facial structures. Besides, to distinguish with TD, it is also necessary for SD not to disturb the original texture. To disturb facial structures while keeping textures intact, we copy and tile the surrounding textures into areas around a few facial landmarks. The crux lies in the fact that the copy-and-tile operation preserves original textures and thus is the ideal method for disturbing structures. This aids the detector in accurately learning real facial structure in RAM, enabling them to distinguish faces with structural anomalies during detection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Figure 2 illustrates the process of SD. Given an input face image \\( I_{ori} \\), we first detect its facial landmarks. We randomly sample several landmark points and draw circles around them as the mask \\( M \\) for disturbed areas. We avoid sampling landmarks of symmetric positions in the face, which enables the detector to rely on the structure information from the symmetric region to perform RAM. We then employ the Fast Marching Method [32], a non-learning image inpainting algorithm, which produces less blur and maintains clearer textures than other textures replicating algorithms, to copy and tile the surrounding textures to the disturbed area. After the surrounding textures are tiled to the disturbed areas, the original structures of the disturbed areas are eliminated and become anomalous. Incorporating SD in Face Disturbance emrichs anomalies in constructed pseudo-fakes, which helps the detector generalize to a wider range of deepfakes." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.298, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Real Appearance Modeling for More General Deepfake Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.143, + 0.699, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.3, + 0.788, + 0.329 + ], + "angle": 0, + "content": "Fig. 3: The proposed RAE is optimized using two learning objects, i.e., \\(\\mathcal{L}_{rec}\\) for the RAM and \\(\\mathcal{L}_{cls}\\) for the deepfake detection." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.358, + 0.489, + 0.374 + ], + "angle": 0, + "content": "3.2 Real Appearance Modeling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.788, + 0.518 + ], + "angle": 0, + "content": "We believe that guiding models to learn the appearances of real faces is more beneficial for the generalization across deepfake algorithms than learning specific deepfake patterns. Thus, we propose the novel Real Appearance Modeling (RAM) task to guide detectors to learn real face appearances. RAM requires the model to restore anomalous facial areas produced by Face Disturbance to their original appearances. In comparison to existing methods, our model is guided to learn patterns of real faces instead of deepfake patterns in training fake faces. Then, it classifies faces that display anomalies not found in real faces as fake. Therefore, RAM helps to improve generalization across deepfake algorithms." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.766 + ], + "angle": 0, + "content": "Task Definition Given an input face \\( I_{ori} \\), we obtain pseudo-fake faces \\( I_{ta} \\) with texture anomalies and \\( I_{sa} \\) with structure anomalies after the Face Disturbance strategy. These images are fed into the detector for deepfake detection and RAM, i.e., classification and densely real appearance recovery. In the training procedure, the deepfake detection object and RAM object are optimized in a multi-task learning manner. In deepfake detection, the labels for constructed pseudo-fakes are 1, while for real faces are 0. In RAM, the recovery target \\( I_{target} \\) for \\( I_{sa} \\) is \\( I \\), while the recovery target \\( I_{target} \\) for \\( I_{ta} \\) is \\( I_{outer} \\). This is because \\( I_{ta} \\) is derived from \\( I_{outer} \\) and \\( I_{inner} \\), so there are two potential recovery targets, and we fix the target as \\( I_{outer} \\). This allows the model to restore the inner texture from \\( I_{inner} \\) using the background texture from \\( I_{outer} \\) as the reference. Thus, the model is guided to align the inner texture with the outer texture. The feasibility of recovering pseudo-fakes to original faces lies in the fact that the pseudo-fakes contains all the information of the corresponding original faces, with only variations between the inner and outer face regions or partially disrupted structure." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.786, + 0.452, + 0.802 + ], + "angle": 0, + "content": "3.3 Recovery Autoencoder" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In order to perform both deepfake detection and RAM, the detector in our framework requires a classification head and a dense prediction head. Thus, we" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "J. Tian et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "incorporate a dense prediction branch for the RAM object into a classifier. This configuration makes our model resemble an autoencoder, thus named Recovery Autoencoder (RAE). The RAE is illustrated in Figure 3, which is constructed based on a vanilla Vision Transformer Base (ViT Base) as the encoder. We adopt eight transformer layers and a linear projection layer as the decoder, and a Multi-layer Token Fusion (MTF) block to extract features from both the shallow and deep encoder layers. We refer readers to [33] for detailed information of the ViT encoder. Following this, we detail the entire pipeline of RAE." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.789, + 0.404 + ], + "angle": 0, + "content": "After the image \\( I \\in R^{H \\times W \\times 3} \\) is fed to the ViT encoder, it is extracted to a sequence of tokens \\( x_{l} \\in R^{(N + 1) \\times D} \\) in each encoder layer. Here, \\( l \\in 1, 2, \\dots, L \\) is the encoder layer index, \\( L = 12 \\) is the depth of the encoder, \\( N \\) is the number of image patches, and \\( D \\) represents the embedding dimension. We note \\( x_{L}[0] \\) is the c1s token [33] for classification. The classification process involves passing the c1s token to a linear layer and a softmax operator to predict the class probabilities \\( \\hat{y} \\in \\mathbb{R}^2 \\), and 2 is the number of classes. Then, the remaining tokens in \\( x_{l} \\) are reshaped to \\( \\mathbb{R}^{h \\times w \\times D} \\) where \\( h = w = \\sqrt{N} \\), and then inputted to MTF for cross-layer token fusion and the decoder for RAM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.789, + 0.617 + ], + "angle": 0, + "content": "As there are 12 transformer layers in the ViT encoder, inputting all layers in \\( x_{l} \\) to MTF is computationally intensive. Thus, we select \\( S = 3 \\) layers at equal intervals and tokens in these layers as the inputs to MTF to provide shallow, mid, and deep features. Formally, the selected tokens of shallow, mid, and deep layers are noted as \\( x_{s}, x_{m} \\), and \\( x_{d} \\). In MTF, we first project \\( x_{s}, x_{m} \\), and \\( x_{d} \\) into the same feature space and concatenate them to \\( x^{\\prime} \\in R^{S \\times h \\times w \\times D} \\). We employ a linear layer as the projection layer, which is widely used for aligning different spaces. After projection, we further extract and fuse both structural and textural information from different layers. As tokens in different ViT layers have identical dimensions, we adopt a cross-layer attention mechanism for fusion. Specifically, MTF predicts weights \\( W \\in R^{S \\times h \\times w} \\) for each token with an MLP. \\( W \\) is normalized using softmax along the first dimension \\( S \\), ensuring that the weights for tokens at the same spatial position across \\( S \\) layers sum to 1. \\( W \\) and fused tokens \\( x_{fus} \\in R^{h \\times w \\times D} \\) are derived as:" + }, + { + "type": "equation", + "bbox": [ + 0.421, + 0.629, + 0.785, + 0.645 + ], + "angle": 0, + "content": "\\[\nx ^ {\\prime} = \\operatorname {c o n c a t} \\left(x _ {s}, x _ {m}, x _ {d}\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.418, + 0.649, + 0.785, + 0.664 + ], + "angle": 0, + "content": "\\[\nW = \\operatorname {s o f t m a x} (\\operatorname {M L P} \\left(x ^ {\\prime}\\right)), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.403, + 0.668, + 0.785, + 0.694 + ], + "angle": 0, + "content": "\\[\nx _ {f u s} = \\sum_ {i = 0} ^ {S} W [ i ] \\times x ^ {\\prime} [ i ]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.842 + ], + "angle": 0, + "content": "After extracting \\( x_{fus} \\), we aim to recover the original face with \\( x_{fus} \\). While it might be intuitive to directly feed \\( x_{fus} \\) into the decoder to recover real faces, this leads to easy mining that outputs the inputted pseudo-fakes directly. This is because the inputted pseudo-fakes and the target real faces are identical in the undisturbed areas \\( 1 - M \\) and are similar in disturbed areas \\( M \\). To avoid such easy mining, we introduce a learnable mask msk token to replace tokens of the disturbed areas. Specifically, given a binary mask \\( M \\) for disturbed areas, it is divided into mask patches \\( m \\in R^{h \\times w \\times p \\times p} \\), where \\( p \\) is the side length of a patch. If a mask patch is disturbed, the corresponding patch is replaced with msk" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.298, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Real Appearance Modeling for More General Deepfake Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.265 + ], + "angle": 0, + "content": "token, and the mask for masked patches is \\( M_{patch} \\). After masking, the masked tokens \\( x_{msk} \\in R^{h \\times w \\times D} \\) are added with positional encodings and then fed into the decoder to \\( \\hat{x} \\in R^{N \\times p \\times p \\times 3} \\). Lastly, through a reshape operation, we obtain the recovered face image \\( \\hat{I} \\in R^{H \\times W \\times 3} \\) where \\( H = W = \\sqrt{N} \\times p \\). We emphasize that the masking is applied after encoding. Thus, the classification head and the predicted \\( \\hat{y} \\) remain unaffected. Moreover, information in masked tokens is fused to unmasked tokens in encoding, thus necessary information that enables the recovery is still conveyed to the decoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.785, + 0.358 + ], + "angle": 0, + "content": "In predicting the real appearance of disturbed areas, RAE extracts texture and structure features from undisturbed areas and semantic information from disturbed areas for RAM. Consequently, the decoder can utilize the combination of these features to model the appearance of disturbed areas through pixel-level prediction. This enhances the real appearance learned in the encoder, thereby improving the generalization of the deepfake detection." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.38, + 0.377, + 0.394 + ], + "angle": 0, + "content": "3.4 Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.405, + 0.784, + 0.435 + ], + "angle": 0, + "content": "The proposed RAE is optimized to both discriminate disturbed faces and model real appearance. We use cross-entropy loss \\(\\mathcal{L}_{cls}\\) for classification:" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.448, + 0.785, + 0.462 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c l s} = - y \\log \\hat {y} - (1 - y) \\log (1 - \\hat {y}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.474, + 0.784, + 0.504 + ], + "angle": 0, + "content": "where \\(\\hat{y}\\) is the predicted class probabilities and \\(y\\) is the class label. Besides, we use the following L1-loss \\(\\mathcal{L}_{rec}\\) for RAM:" + }, + { + "type": "equation", + "bbox": [ + 0.343, + 0.515, + 0.784, + 0.547 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e c} = \\frac {1}{\\left\\| M _ {p a t c h} \\right\\| _ {1}} \\left\\| M _ {p a t c h} \\odot (\\hat {I} - I _ {t a r g e t}) \\right\\| _ {1}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.557, + 0.784, + 0.602 + ], + "angle": 0, + "content": "where \\( \\| \\cdot \\|_1 \\) is the L1-norm, \\( M_{patch} \\) is the mask for disturbed tokens and thus \\( \\mathcal{L}_{rec} \\) is performed on the disturbed tokens. The total loss \\( \\mathcal{L} \\) for optimizing RAE is formulated as the weighted sum of \\( \\mathcal{L}_{cls} \\) and \\( \\mathcal{L}_{rec} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.439, + 0.615, + 0.784, + 0.629 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {c l s} + \\alpha \\mathcal {L} _ {r e c}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.641, + 0.784, + 0.656 + ], + "angle": 0, + "content": "We set \\(\\alpha\\) as 0.1 to make \\(\\mathcal{L}_{cls}\\) and \\(\\mathcal{L}_{rec}\\) converge to the same order of magnitude." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.679, + 0.376, + 0.696 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.71, + 0.457, + 0.725 + ], + "angle": 0, + "content": "4.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.735, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Dataset. We conduct evaluations on six widely used deepfake datasets, i.e., FaceForensics++ (FF++) [34], Celeb-DF v2 (CDF) [35], DFD [34], DFDC [36], DFDC preview (DFDCp) [36], and WildDeepfake (WDF) [37]. FF++ is a deep-fake dataset containing 4,000 forged videos and 1,000 real videos. All videos are provided in three compression levels: raw, high-quality (HQ), and low-quality (LQ). CDF contains 590 real and 5,639 fake videos corresponding to 59 celebrities using an improved face swap algorithm. DFD contains over 3,000 face swap" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "J. Tian et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "videos. DFDC is a large-scale dataset that contains more than 120,000 deepfake clips generated using eight deepfake techniques. DFDCp is the preview dataset for the Deepfake Detection Challenge. Its fake videos are generated using two face swap algorithms. WDF consists of 10,000 real videos and 10,000 online collected fake videos. For DFDC, we use videos with one identity following [8,17,38] in case of labeling noise. For other datasets, we use the official dataset splits. We use the real videos of \\(\\mathrm{FF} + + (\\mathrm{HQ})\\) to construct pseudo-fakes for training, while other datasets are for testing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.788, + 0.344 + ], + "angle": 0, + "content": "Preprocess. We extract 10 frames at equal temporal intervals from each video in training and 32 frames from each video in testing. Then, the faces are cropped using Retinaface [39], maintaining a \\(15\\%\\) margin around each face. Subsequently, we utilize Dlib [40] to extract 81 facial landmarks for every face and save them to a key-value database to boost the Face Disturbance process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.465 + ], + "angle": 0, + "content": "Face Disturbance Details. As for the texture augmentations involved in generating \\( I_{ta} \\), we employ RGBShift, HueSaturationValue, RandomBrightnessContrast, Downscale, and Sharpen to augment textures of either \\( I_{inner} \\) or \\( I_{outer} \\) to produce inconsistencies. As for the disturbed areas in generating \\( I_{sa} \\), we draw circles around 2 to 6 random landmarks with a radius sampled between 1/14 and 1/16 of the side length. The number and size of the sampled circles for SD ensure have been found empirically to aid in the effectiveness of RAM tasks as shown in supplementary. Then, all images are resized to \\( 224 \\times 224 \\) as inputs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.788, + 0.571 + ], + "angle": 0, + "content": "Training Details. We use a ViT Base pretrained on ImageNet [41] to initialize our RAE. RAE is trained using SAM [42] optimizer with a learning rate of 8e-4 and a batch size of 192 for 100 epochs, and we use cosine decay as the learning rate scheduler. Each batch consists of 96 real faces, 48 disturbed faces with texture anomalies, and 48 disturbed faces with structure anomalies. In the PyTorch framework, training with automatic mixed precision requires approximately 60GB of GPU memory." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.591, + 0.473, + 0.605 + ], + "angle": 0, + "content": "4.2 Cross-Dataset Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In cross-dataset evaluation, detectors are trained using faces in \\(\\mathrm{FF}++(\\mathrm{HQ})\\) and evaluated on other datasets to assess their generalization to unknown deepfake techniques. Here, we compare the proposed RAE with recent deepfake detection methods. The results are listed in Table 1, and we annotated the input type and training faces. RAE achieves top generalization performance and outperforms state-of-the-art detectors on CDF, DFDC, and DFDCp by \\(2.6\\%\\), \\(3.4\\%\\), and \\(3.2\\%\\), respectively. Compared to detectors trained using fake faces, such as TALL and IDDDM, the superior generalization of RAE demonstrates the effectiveness of using Face Disturbance to generate pseudo-fakes for training. Compared to detectors trained using pseudo-fakes with texture inconsistency, such as SBI and AltFreezing, RAE exhibits better generalization on face swap datasets such as CDF and DFDCp. As the design of TD is similar to previous works, the effectiveness of RAE against face swap deepfakes indicates RAM helps to spot richer anomalies through learning real appearance. Moreover, previous detectors trained with pseudo-fakes fail to exhibit better performance on" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.298, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Real Appearance Modeling for More General Deepfake Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.187 + ], + "angle": 0, + "content": "Table 1: Cross-dataset evaluation of deepfake detection methods on CDF, DFD, DFDC, DFDCP, and WDF. All models are trained using FF++(HQ). All compared results are video-level AUC and cited from the original papers or their subsequences." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.199, + 0.784, + 0.369 + ], + "angle": 0, + "content": "
MethodInput TypeYearTraining SetTest Set AUC (%)
Real FakePseudo-fakeCDFDFDDFDCDFDCpWDF
PCL+I2G [4]Frame2021✓✓90.099.067.574.4-
ENB4+SBI [23]Frame2022✓✓92.998.272.085.5-
Mover [20]Frame2023✓✓87.1--78.982.0
RFFR [21]Frame2023✓✓89.0-67.8--
IDDDM [31]Frame2023✓✓91.2-71.5--
SeeABLE [43]Frame2023✓✓87.3-75.986.3-
FTCN [38]Clip2021✓✓86.994.471.074.0-
RealForensics [17]Clip2022✓✓86.9-75.9--
AltFreezing [24]Clip2023✓✓✓89.598.5---
TALL [44]Clip2023✓✓90.8-76.8--
RAE (Ours)Frame2024✓✓95.599.080.289.588.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.383, + 0.502, + 0.425 + ], + "angle": 0, + "content": "Table 2: In-dataset evaluation on four subsets in FF++(HQ). †indicates reproduced on FF++(HQ)." + }, + { + "type": "table", + "bbox": [ + 0.224, + 0.437, + 0.499, + 0.517 + ], + "angle": 0, + "content": "
MethodTest Set AUC(%)
DFF2FFSNTAvg
SLADD [22]----98.4
ENB4+SBI † [23]99.299.199.096.798.5
SeeABLE [43]99.298.899.196.998.5
RAE(Ours)99.699.199.297.698.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.383, + 0.785, + 0.425 + ], + "angle": 0, + "content": "Table 3: Frame-level FP rate(%) on real face datasets. Lower FP is better. All models are trained on \\(\\mathrm{FF} + + (\\mathrm{HQ})\\)" + }, + { + "type": "table", + "bbox": [ + 0.542, + 0.437, + 0.756, + 0.517 + ], + "angle": 0, + "content": "
MethodTest Set FP(%)↓
FFHQVoxCeleb
RECCE [45]62.341.7
UIA-ViT [46]40.618.9
ENB4+SBI [23]57.759.2
RAE(ours)36.96.9
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.547, + 0.784, + 0.593 + ], + "angle": 0, + "content": "DFDC as it contains reenacted faces that are difficult to be imitated by previous pseudo-fakes. Our method demonstrates a major improvement compared to previous methods as SD helps RAE spot structure anomalies in reenacted faces." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.617, + 0.44, + 0.631 + ], + "angle": 0, + "content": "4.3 In-dataset Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In addition to cross-dataset evaluation, another frequently used evaluation protocol is in-dataset evaluation. We follow the protocol in [43] to conduct evaluations on the four types of deepfakes in \\(\\mathrm{FF}++\\), i.e., two types of face swap deepfakes DF and FS, and two types of face reenactment deepfakes F2F and NT, where the training data includes both deepfake and pseudo-fake samples. This evaluation protocol helps to assess the ability of our method in enhancing in-dataset deepfake detection. The results are listed in Table 2. The overall performance of RAE on all four deepfake types surpasses previous works. Specifically, RAE outperforms the compared methods in detecting reenacted faces generated by the two reenactment algorithms, i.e., F2F and NT. RAE achieves a detection AUC of \\(97.6\\%\\) on NT. This indicates that combining SD and the learning target RAM aids in recognizing distorted facial features in NT deepfakes, while previous works trained with texture inconsistency pseudo-fakes fail." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "J. Tian et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.586, + 0.162 + ], + "angle": 0, + "content": "4.4 Evaluating RAE on Real Face Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.787, + 0.311 + ], + "angle": 0, + "content": "In practical deployment, a low False Positive rate (FP) is important for a deepfake detector, which indicates a low ratio of misclassifying real faces as fakes. For instance, a detector with a low FP rate in an authentication system helps to avoid having users repeatedly submit their portraits, thereby enhancing the user experience. Besides, at the same AUC, a lower FP helps reduce the adjustment of classification thresholds in practical applications. We use the mega-scale face dataset VoxCeleb [47], and the high-quality face dataset FFHQ [48] to test the FP of the proposed RAE on real faces. For comparison, we used three open-source detectors, RECCE [45], SBI [23], and UIA-ViT [46]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.787, + 0.402 + ], + "angle": 0, + "content": "The FP scores are listed in Table 3. Compared to existing arts, RAE achieves the lowest FP on these two datasets. In stark contrast, the average FP scores of the compared RECCE and SBI are larger than \\(50\\%\\), which indicates they misclassified more than half real faces in these two datasets. This validates that RAM successfully makes the detector learn real face appearances and the trained RAE is less likely to classify a real face as fake." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.427, + 0.544, + 0.442 + ], + "angle": 0, + "content": "4.5 Evaluation on Diffusion Deepfakes" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.787, + 0.606 + ], + "angle": 0, + "content": "In the cross-dataset evaluation, we validated the generalizability of our framework across five popular deepfake datasets. However, we noticed these datasets primarily consist of face swap and reenactment deepfakes. As a result, the effectiveness of detectors against other types of deepfakes, such as diffusion-generated faces, was not adequately evaluated. Given the fact that a proliferating number of deepfakes are diffusion-generated faces, we conduct evaluations on a diffusion face dataset DeepFakeFace(DFF) [27]. DFF is synthesized by Stable Diffusion Text2Img (T2I), Stable Diffusion Inpaint (Inpaint) [49], and InsightFace (IF) [50]. We conduct evaluation using the three subsets in DFF. For all compared methods, we use their official weights." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.607, + 0.787, + 0.712 + ], + "angle": 0, + "content": "The evaluation results are listed in Table 4. RAE outperforms the compared methods in all three subsets in DFF. More significantly, RAE attains an AUC of \\(70.8\\%\\) on average, which improves existing SOTA methods by \\(2.2\\%\\). In stark contrast, the AUC scores of RECCE and UIA-ViT on T2I are less than \\(50\\%\\), which indicates they perform worse than randomly guessing. The results on DFF verify the effectiveness of our proposed method in enhancing the detector's generalization to a wider range of deepfakes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.738, + 0.388, + 0.753 + ], + "angle": 0, + "content": "4.6 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We conduct ablations to verify the effectiveness of the RAM task, Face Disturbance strategy, and the MTF block. A ViT Base trained using \\( I_{ta} \\) and classification task is used to serve as the Baseline as it is similar to several previous detectors [1, 23]. We incrementally add the proposed RAM task, SD, and MTF block to the baseline to observe their effectiveness. The results are listed in Table" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.298, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Real Appearance Modeling for More General Deepfake Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.505, + 0.188 + ], + "angle": 0, + "content": "Table 4: Frame-level AUC on three subsets of diffusion deepfakes in DFF [27]. All models are trained on \\(\\mathrm{FF} + + (\\mathrm{HQ})\\)" + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.199, + 0.501, + 0.289 + ], + "angle": 0, + "content": "
MethodTest Set AUC(%)
T2IInpaintIFAvg
RECCE [45]35.151.563.149.9
UIA-ViT [46]45.255.874.058.3
ENB4+SBI [23]62.871.471.568.6
RAE(ours)64.173.075.270.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 5: Ablation studies for the proposed Structure Disturbance, RAM task, and MTF module." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.2, + 0.772, + 0.289 + ], + "angle": 0, + "content": "
MethodTest set AUC(%)
CDFDFDCDFDCp
Baseline92.575.585.9
+RAM94.277.488.6
+Isa94.779.188.4
+MTF95.580.289.5
" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.304, + 0.761, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.462, + 0.788, + 0.49 + ], + "angle": 0, + "content": "Fig. 4: Visualizations of RAM. The left are recovery results for \\( I_{sa} \\) while the right are results for \\( I_{ta} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.614 + ], + "angle": 0, + "content": "5, which show that each of them contributes to enhancing generalization. Compared to solely using a classification task, adding the RAM task significantly improves generalization by \\(2.2\\%\\) on average. Furthermore, utilizing \\(I_{sa}\\) to model real facial structures mainly improves the generalization on the DFDC dataset, which includes reenactment deepfakes. Besides, the MTF block aids in improving generalization by leveraging multi-layer features for modeling real appearances." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.643, + 0.371, + 0.657 + ], + "angle": 0, + "content": "4.7 Visualization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Face Recovery. We visualize the recovered disturbed faces to verify whether RAE learned real face appearances. The inputted disturbed faces, corresponding original faces, masks for disturbed facial areas, and recovered faces are presented in Figure 4. We observe that our model successfully recovers most of the perturbed areas within the anomalous faces. Though the recovered faces are not explicitly clear, the anomalous patterns are largely eliminated. For faces with structure anomalies, disturbed areas are finely recovered to their original structure appearance. For faces with texture anomalies, the texture pattern of the inner faces is aligned to normal appearance. This demonstrates the feasibility of modeling real appearance by recovering anomalous faces to their original appearance, which provides more fine-grained and interpretable supervision." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "J. Tian et al." + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.147, + 0.494, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.327, + 0.323, + 0.441, + 0.335 + ], + "angle": 0, + "content": "(a) Baseline model." + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.147, + 0.732, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.575, + 0.323, + 0.665, + 0.335 + ], + "angle": 0, + "content": "(b) Our model." + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.346, + 0.788, + 0.375 + ], + "angle": 0, + "content": "Fig. 5: Distributions of learned representations for faces in \\(\\mathrm{FF} + + (\\mathrm{HQ})\\) [34] in the baseline model(a) and our model(b)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.407, + 0.789, + 0.604 + ], + "angle": 0, + "content": "Feature Distribution. We apply t-SNE [51] to visualize the features in the last layer of the encoder to illustrate the effect of RAM. We use a ViT Base as the baseline detector and train it using real and fake faces in \\(\\mathrm{FF} + + (\\mathrm{HQ})\\). Then, we use the trained RAE and baseline detector to extract features from the test set of \\(\\mathrm{FF} + + (\\mathrm{HQ})\\). The results are illustrated in Figure 5. The baseline model learns specific features for each deepfake algorithm, thus separating deepfakes of different types in the feature space. In contrast, the proposed RAE aggregates real faces to a compact cluster in the feature space, which indicates it classifies deepfakes based on the learned general fake patterns. Moreover, RAE mixes different types of deepfakes in the feature space, which confirms that RAM helps to learn general deepfake representations. Compared to the baseline model, the anomalous features learned by RAE are more general across different deepfakes, thus demonstrating better generalization." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.634, + 0.36, + 0.65 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.788, + 0.809 + ], + "angle": 0, + "content": "In this paper, we address the general deepfake detection through guiding the detector to learn real face appearance by recovering original faces from disturbed faces. Anomalies in both the texture and structure views are introduced to faces through the proposed Face Disturbance, and the detector is guided to recover real appearances from these disturbed faces through the proposed RAM task. We then propose the RAE model to conduct both the deepfake detection and the RAM task. Extensive experiments demonstrate the effectiveness of our method against existing methods in general deepfake detection. Future work involves scaling our method with large-scale human face datasets." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "This project is supported by the National Key Research and Development Program of China (No.2022YFB2702500)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.298, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Real Appearance Modeling for More General Deepfake Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.176, + 0.786, + 0.232 + ], + "angle": 0, + "content": "1. Yuezun Li and Siwei Lyu. Exposing deepfake videos by detecting face warping artifacts. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2019, Long Beach, CA, USA, June 16-20, 2019, pages 46-52. Computer Vision Foundation / IEEE, 2019. 2, 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.233, + 0.787, + 0.286 + ], + "angle": 0, + "content": "2. Lingzhi Li, Jianmin Bao, Ting Zhang, Hao Yang, Dong Chen, Fang Wen, and Baining Guo. Face x-ray for more general face forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5001-5010, 2020. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.787, + 0.37 + ], + "angle": 0, + "content": "3. Ke Sun, Taiping Yao, Shen Chen, Shouhong Ding, Jilin Li, and Rongrong Ji. Dual contrastive learning for general face forgery detection. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, Thirty-Fourth Conference on Innovative Applications of Artificial Intelligence, IAAI 2022, The Twelveth Symposium on Educational Advances in Artificial Intelligence, EAAI 2022 Virtual Event, February 22 - March 1, 2022, pages 2316-2324. AAAI Press, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.787, + 0.424 + ], + "angle": 0, + "content": "4. Tianchen Zhao, Xiang Xu, Mingze Xu, Hui Ding, Yuanjun Xiong, and Wei Xia. Learning self-consistency for deepfake detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15023-15033, 2021. 2, 4, 5, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.426, + 0.787, + 0.468 + ], + "angle": 0, + "content": "5. Yuchen Luo, Yong Zhang, Junchi Yan, and Wei Liu. Generalizing face forgery detection with high-frequency features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16317-16326, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.468, + 0.787, + 0.509 + ], + "angle": 0, + "content": "6. Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In European Conference on Computer Vision, pages 86-103. Springer, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.509, + 0.787, + 0.564 + ], + "angle": 0, + "content": "7. Yuezun Li, Ming-Ching Chang, and Siwei Lyu. In ictu oculi: Exposing AI created fake videos by detecting eye blinking. In 2018 IEEE International Workshop on Information Forensics and Security, WIFS 2018, Hong Kong, China, December 11-13, 2018, pages 1-7. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.787, + 0.632 + ], + "angle": 0, + "content": "8. Alexandros Haliassos, Konstantinos Vougioukas, Stavros Petridis, and Maja Pan- tic. Lips don't lie: A generalisable and robust approach to face forgery detection. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 5039-5049. Computer Vision Foundation / IEEE, 2021. 2, 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.787, + 0.675 + ], + "angle": 0, + "content": "9. Jean-François Aujol, Guy Gilboa, Tony F. Chan, and Stanley J. Osher. Structure-texture image decomposition - modeling, algorithms, and parameter selection. Int. J. Comput. Vis., 67(1):111-136, 2006. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.787, + 0.716 + ], + "angle": 0, + "content": "0. Youngjung Kim, Bumsub Ham, Minh N Do, and Kwanghoon Sohn. Structure-texture image decomposition using deep variational priors. IEEE Transactions on Image Processing, 28(6):2692-2704, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.716, + 0.787, + 0.758 + ], + "angle": 0, + "content": "1. Yuezun Li, Ming-Ching Chang, and Siwei Lyu. In ictu oculi: Exposing ai created fake videos by detecting eye blinking. In 2018 IEEE International Workshop on Information Forensics and Security (WIFS), pages 1-7. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "2. Umur Aybars Ciftci, Ilke Demir, and Lijun Yin. Fakecatcher: Detection of synthetic portrait videos using biological signals. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "3. Xin Yang, Yuezun Li, and Siwei Lyu. Exposing deep fakes using inconsistent head poses. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 8261-8265. IEEE, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.176, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "J. Tian et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "14. Darius Afchar, Vincent Nozick, Junichi Yamagishi, and Isao Echizen. Mesonet: a compact facial video forgery detection network. In 2018 IEEE international workshop on information forensics and security (WIFS), pages 1-7. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.788, + 0.245 + ], + "angle": 0, + "content": "15. Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multi-attentional deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2185–2194, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.247, + 0.788, + 0.342 + ], + "angle": 0, + "content": "16. Qiqi Gu, Shen Chen, Taiping Yao, Yang Chen, Shouhong Ding, and Ran Yi. Exploiting fine-grained face forgery clues via progressive enhancement learning. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, Thirty-Fourth Conference on Innovative Applications of Artificial Intelligence, IAAI 2022, The Twelveth Symposium on Educational Advances in Artificial Intelligence, EAAI 2022 Virtual Event, February 22 - March 1, 2022, pages 735-743. AAAI Press, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.345, + 0.788, + 0.4 + ], + "angle": 0, + "content": "17. Alexandros Haliassos, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Leveraging real talking faces via self-supervision for robust forgery detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14950-14962, 2022. 4, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.402, + 0.787, + 0.429 + ], + "angle": 0, + "content": "18. Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Weiming Zhang, and Nenghai Yu. Self-supervised transformer for deepfake detection, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.431, + 0.787, + 0.484 + ], + "angle": 0, + "content": "19. Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.487, + 0.787, + 0.528 + ], + "angle": 0, + "content": "20. Juan Hu, Xin Liao, Difei Gao, Satoshi Tsutsui, Qian Wang, Zheng Qin, and Mike Zheng Shou. Mover: Mask and recovery based facial part consistency aware method for deepfake video detection, 2023. 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.787, + 0.557 + ], + "angle": 0, + "content": "21. Liang Shi, Jie Zhang, and Shiguang Shan. Real face foundation representation learning for generalized deepfake detection, 2023. 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.787, + 0.6 + ], + "angle": 0, + "content": "22. Liang Chen, Yong Zhang, Yibing Song, Lingqiao Liu, and Jue Wang. Self-supervised learning of adversarial examples: Towards good generalizations for deep-fake detections. In CVPR, 2022. 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.602, + 0.787, + 0.643 + ], + "angle": 0, + "content": "23. Kaede Shiohara and Toshihiko Yamasaki. Detecting deepfakes with self-blended images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18720-18729, 2022. 4, 5, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.645, + 0.787, + 0.699 + ], + "angle": 0, + "content": "24. Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, and Houqiang Li. Altfreezing for more general video face forgery detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4129-4138, June 2023. 4, 5, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.701, + 0.787, + 0.783 + ], + "angle": 0, + "content": "25. Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. In Hanna M. Wallach, Hugo Larochelle, Alina Beygelzimer, Florence d'Alché-Buc, Emily B. Fox, and Roman Garnett, editors, Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 7135-7145, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "26. Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Zwei Liu. Pose-controllable talking face generation by implicitly modularized audiovisual representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4176-4186, 2021. 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.298, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Real Appearance Modeling for More General Deepfake Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "27. Haixu Song, Shiyu Huang, Yinpeng Dong, and Wei-Wei Tu. Robustness and generalizability of deepfake detection: A study with diffusion models. arXiv preprint arXiv:2309.02218, 2023. 4, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.231 + ], + "angle": 0, + "content": "28. Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton-Ferrer. The deepfake detection challenge dataset. CoRR, abs/2006.07397, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.787, + 0.286 + ], + "angle": 0, + "content": "29. Gereon Fox, Wentao Liu, Hyeongwoo Kim, Hans-Peter Seidel, Mohamed Elgharib, and Christian Theobalt. Videoforensicshq: Detecting high-quality manipulated face videos. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.328 + ], + "angle": 0, + "content": "30. Jiajun Huang, Xueyu Wang, Bo Du, Pei Du, and Chang Xu. Deepfake mnist+: a deepfake facial animation dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1973-1982, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.383 + ], + "angle": 0, + "content": "31. Baojin Huang, Zhongyuan Wang, Jifan Yang, Jiaxin Ai, Qin Zou, Qian Wang, and Dengpan Ye. Implicit identity driven deepfake face swapping detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4490-4499, 2023. 5, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.411 + ], + "angle": 0, + "content": "32. Alexandru C. Telea. An image inpainting technique based on the fast marching method. J. Graphics, GPU, & Game Tools, 9(1):23-34, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.787, + 0.493 + ], + "angle": 0, + "content": "33. Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.787, + 0.549 + ], + "angle": 0, + "content": "34. Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics++: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1-11, 2019. 9, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.549, + 0.787, + 0.604 + ], + "angle": 0, + "content": "35. Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deepfake forensics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3207-3216, 2020. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.605, + 0.787, + 0.646 + ], + "angle": 0, + "content": "36. Brian Dolhansky, Russ Howes, Ben Pflaum, Nicole Baram, and Cristian Canton-Ferrer. The deepfake detection challenge (DFDC) preview dataset. CoRR, abs/1910.08854, 2019. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.646, + 0.787, + 0.699 + ], + "angle": 0, + "content": "37. Bojia Zi, Minghao Chang, Jingjing Chen, Xingjun Ma, and Yu-Gang Jiang. Wilddeepfake: A challenging real-world dataset for deepfake detection. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2382-2390, 2020. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.701, + 0.787, + 0.757 + ], + "angle": 0, + "content": "38. Yinglin Zheng, Jianmin Bao, Dong Chen, Ming Zeng, and Fang Wen. Exploring temporal coherence for more general video face forgery detection. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 15024-15034. IEEE, 2021. 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.812 + ], + "angle": 0, + "content": "39. Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-shot multi-level face localisation in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5203-5212, 2020. 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.812, + 0.787, + 0.84 + ], + "angle": 0, + "content": "40. Davis E King. Dlib-ml: A machine learning toolkit. The Journal of Machine Learning Research, 10:1755-1758, 2009. 10" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "J. Tian et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "41. Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2009), 20-25 June 2009, Miami, Florida, USA, pages 248-255. IEEE Computer Society, 2009. 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.787, + 0.258 + ], + "angle": 0, + "content": "42. Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.787, + 0.313 + ], + "angle": 0, + "content": "43. Nicolas Larue, Ngoc-Son Vu, Vitomir Struc, Peter Peer, and Vassilis Christophides. Seeable: Soft discrepancies and bounded contrastive learning for exposing deep-fakes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21011-21021, 2023. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "44. Yuting Xu, Jian Liang, Gengyun Jia, Ziming Yang, Yanhao Zhang, and Ran He. Tall: Thumbnail layout for deepfake video detection. arXiv preprint arXiv:2307.07494, 2023. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.787, + 0.411 + ], + "angle": 0, + "content": "45. Junyi Cao, Chao Ma, Taiping Yao, Shen Chen, Shouhong Ding, and Xiaokang Yang. End-to-end reconstruction-classification learning for face forgery detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4113-4122, June 2022. 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.787, + 0.466 + ], + "angle": 0, + "content": "46. Wanyi Zhuang, Qi Chu, Zhentao Tan, Qiankun Liu, Haojie Yuan, Changtao Miao, Zixiang Luo, and Nenghai Yu. Uia-vit: Unsupervised inconsistency-aware method based on vision transformer for face forgery detection. In European Conference on Computer Vision (ECCV), 2022. 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.787, + 0.494 + ], + "angle": 0, + "content": "47. Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. arXiv preprint arXiv:1806.05622, 2018. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.535 + ], + "angle": 0, + "content": "48. Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. IEEE Trans. Pattern Anal. Mach. Intell., 43(12):4217-4228, 2021. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.787, + 0.563 + ], + "angle": 0, + "content": "49. Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.787, + 0.592 + ], + "angle": 0, + "content": "50. DeepInsight. insightface. https://github.com/deepinsight/insightface/, 2021. [Online]. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.787, + 0.619 + ], + "angle": 0, + "content": "51. Van Der Maaten Laurens and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9(2605):2579-2605, 2008. 14" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.619 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Real Appearance Modeling for More General Deepfake Detection/2bfe37cc-d80a-473f-a5d8-70fca9af4b74_origin.pdf b/2024/Real Appearance Modeling for More General Deepfake Detection/2bfe37cc-d80a-473f-a5d8-70fca9af4b74_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..efa76298d664db1840a090c8193852b8684db431 --- /dev/null +++ b/2024/Real Appearance Modeling for More General Deepfake Detection/2bfe37cc-d80a-473f-a5d8-70fca9af4b74_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:74d0c2e7d1777cd2769545e4318ae8d893ace54a957b25027078c94553ea1381 +size 4660297 diff --git a/2024/Real Appearance Modeling for More General Deepfake Detection/full.md b/2024/Real Appearance Modeling for More General Deepfake Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0334a22859322371fc70c1ada40cc76435cb5d74 --- /dev/null +++ b/2024/Real Appearance Modeling for More General Deepfake Detection/full.md @@ -0,0 +1,284 @@ +# Real Appearance Modeling for More General Deepfake Detection + +Jiahe Tian $^{1,2}$ , Cai Yu $^{1,2}$ , Xi Wang $^{3}$ , Peng Chen $^{4}$ , Zihao Xiao $^{4}$ , Jiao Dai $^{2}$ , Jizhong Han $^{2}$ , and Yesheng Chai $^{2}$ + +$^{1}$ School of Cyber Security, University of Chinese Academy of Sciences, Beijing 100049, China + +$^{2}$ Institute of Information Engineering, Chinese Academy of Sciences, Beijing 100094, China + +$^{3}$ Institute of Microelectronics, Chinese Academy of Sciences, Beijing 100094, China + $^{4}$ RealAI Inc., Beijing 100094, China + +Abstract. Recent studies in deepfake detection have shown promising results when detecting deepfakes of the same type as those present in training. However, their ability to generalize to unseen deepfakes remains limited. This work improves the generalizable deepfake detection from a simple principle: an ideal detector classifies any face that contains anomalies not found in real faces as fake. Namely, detectors should learn consistent real appearances rather than fake patterns in the training set that may not apply to unseen deepfakes. Guided by this principle, we propose a learning task named Real Appearance Modeling (RAM) that guides the model to learn real appearances by recovering original faces from slightly disturbed faces. We further propose Face Disturbance to produce disturbed faces while preserving original information that enables recovery, which aids the model in learning the fine-grained appearance of real faces. Extensive experiments demonstrate the effectiveness of modeling real appearances to spot richer deepfakes. Our method surpasses existing state-of-the-art methods by a large margin on multiple popular deepfake datasets. + +Keywords: Deepfake Detection $\cdot$ Face Forgery Detection $\cdot$ Multimedia Forensic + +# 1 Introduction + +Current deep generative models allow for the generation of realistic fake faces, i.e., deepfakes, making it easy to manipulate face media. Therefore, these techniques can be used for committing fraud, bypassing identity authentication, or spreading false information. The potential misuses raise concerns regarding privacy and reputation, and thus deepfake detectors are needed to counter the risks posed by deepfakes. Moreover, the proliferation of deepfake algorithms presents a challenge to the generalizability of deepfake detection, making it crucial to develop detectors that generalize across various deepfake types. + +![](images/6ad10ae9a0e37430415dd94175066ffc246a2fe2c628cc6663d6fbfa5b6c7579.jpg) +(a) Deepfake Detection. + +![](images/0878a1942adfae22bff02aa922e0bd116c7c34a1d7aa9e07ebca2190f79d3046.jpg) +(b) The RAM task. +Fig. 1: Illustration of (a) the Deepfake Detection task and (b) the proposed RAM task. Compared to Deepfake Detection, the RAM task guides models to restore disturbed faces back to their original appearance. + +Previous works usually tackled deepfake detection by guiding detectors to recognize specific deepfake patterns, such as the boundaries in swapped faces [1, 2], inconsistency [3, 4], frequency anomalies [5, 6], and movement anomalies [7, 8]. They achieved good performance in detecting deepfakes of the same types as those present in training. However, their performance often significantly deteriorates when detecting deepfakes generated by novel algorithms, which mirrors real-world scenarios where various deepfakes emerge. This is due to different deepfake algorithms producing different fake patterns, and learned patterns are not universally applicable to all deepfakes. Namely, more than knowledge of specific deepfake patterns is required to generalize to unseen types of deepfake. + +Though different deepfakes exhibit different patterns, real faces maintain consistent appearance. Thus, it is better to learn the appearance of real faces to verify their authenticity instead of learning deepfake patterns to spot fakes. Namely, an ideal classifier would classify faces as fake if they exhibit patterns not found in real faces, i.e., anomalies. As the detector learns explicit real face appearance, it should be capable of predicting real faces at pixel-level. This guides us to design a task that densely predicts real faces to learn real appearance. To fully define this task, dense prediction targets and input information that assists the prediction are needed. Specifically, we propose the Real Appearance Modeling (RAM) task that recovers corresponding original faces from disturbed faces, as illustrated in Figure 5, where original faces are prediction targets and the inputted disturbed faces contain information needed for prediction. For the RAM task, Deepfakes are not applicable to serve as training data since their original identities or poses are altered, and the necessary information that enables prediction is lost. We thus propose Face Disturbance to disturb real faces to pseudo-fakes as training data for RAM. In the disturbed faces produced by Face Disturbance, the original information is mostly retained. Then, the detector learns real face appearance by predicting original faces. + +To disturb real faces, the proposed Face Disturbance includes two parts, Texture and Structure Disturbance, to introduce anomalies in textures and structures, respectively. This is based on the premise that an image can be regarded as composed of structures and textures [9,10]. Texture Disturbance disturbs partial + +textures to produce anomalies in the form of texture inconsistency. The reason for disturbing partially rather than globally is that distribution of real textures varies, and globally modifying facial textures may fail to produce anomalies. Meanwhile, having inconsistent textures between different facial parts is indeed anomalous. Hence, Texture Disturbance is designed to produce inconsistencies. Structure Disturbance disturbs structure by covering key facial areas, including facial features and boundaries, which are essential components of facial structure. The specific covering method involves copying and tiling the surrounding texture to the disturbed areas to ensure the original texture is intact. Jointly using Texture and Structure Disturbance to produce disturbed faces for RAM helps the detector to learn real structures and textures to spot anomalies in unseen deepfakes. + +The proposed RAM uses disturbed faces produced by Face Disturbance as inputs and the original faces as prediction targets, making this task unsupervised. Therefore, RAM is adopted as an auxiliary task to enhance deepfake detectors. In training, real and disturbed faces are used as inputs, and the detector is expected to both discriminate disturbed faces and recover their real appearances. Specifically, RAM involves guiding the detector to recover structure and texture anomalies. For structure anomalies, the detector recovers the structure of the disturbed areas with the undisturbed areas as the reference. For texture anomalies, the detector recovers the inner texture based on the texture of the outer face. By leveraging learned real appearances, our model is equipped to detect anomalies within unseen deepfakes. As illustrated in Figure 5, our model diverges from the baseline model by grouping various deepfakes to general anomalies rather than identifying specific deepfake patterns. This indicates RAM narrows the gap across deepfakes and thus improves the generalizability. + +Our contributions can be summarized as follows: + +- We introduce Real Appearance Modeling to guide detectors to learn the appearance of real faces by restoring disturbed faces to their original appearances, which helps to improve the generalization of deepfake detectors. +- We propose Face Disturbance to disturb real faces in texture and structure views as pseudo-fakes, which aids the detector in learning real texture and structure appearance when predicting original faces. +- We conduct extensive experiments using various evaluation protocols to demonstrate the generalizability of our framework, which surpasses previous methods on multiple popular deepfake datasets. + +# 2 Related Work + +# 2.1 Deepfake Detection + +Early works leveraged biological artifacts, including distorted pupils [11], heartbeat frequency [12], and anomalous head pose [13] for deepfake detection. More recent works treated deepfake detection as a binary classification task and focused on detecting specific deepfake patterns. For instance, [14] investigated + +mesoscopic features in shallow networks that contain rich features, while [15] employed an attention mechanism to combine RGB and texture features. [6, 16] leveraged high-frequency in the discrete cosine domain as complementary modalities. Another direction involves employing proxy tasks for pretraining to learn high-level representations capable of identifying deepfake videos. Examples of such proxy tasks include lip-reading [8] and audio-visual contrastive learning [17, 18]. Recently, the unsupervised learning object MAE [19] is introduced in deepfake detection [20, 21]. These works usually recover masked facial features through dense prediction to model real faces. As they use masked images as input where information is lost, the set recovery target, i.e., original face, is merely one of the multiple feasible solutions. In contrast, RAM for each disturbed face has a definitive prediction target and thus is not an ill-posed problem with multiple solutions. Compared to MAE, RAM is more fine-grained in supervising detectors to model real face appearance. + +# 2.2 Pseudo-fakes for Deepfake Detection + +Instead of training detectors using deepfakes in deepfake datasets, several studies [1,2,4,22-24] suggested constructing pseudo-fakes for training. Since the introduced pattern in pseudo-fakes is controllable, detectors learn prior patterns accurately. Specifically, they introduced texture inconsistencies to simulate swapped faces and guide detectors to learn this prior pattern. For instance, [1] introduced blur to inner faces to replicate clarity inconsistencies. [2,4,22] blended two faces with similar facial landmarks to generate texture inconsistencies. Moreover, [23, 24] utilized augmentations to produce inconsistencies within a single face, achieving good generalization. Although they generalize well to face swap deepfakes, their performance on deepfakes generated by reenactment algorithms [25, 26] or Diffusion model [27] drops where texture inconsistency is not the major artifacts. + +# 3 Method + +Different from most deepfake detection methods, which are typically trained to spot specific deepfake patterns, our approach emphasizes learning real face appearances through the RAM task. The Face Disturbance strategy is adopted to produce disturbed faces as training data for RAM, as illustrated in Figure 2. The subsequent subsections delve into the details of the proposed Face Disturbance strategy, the definition of the RAM task, and the design of the proposed Recovery Autoencoder (RAE) model in Figure 3, which simultaneously conducts deepfake detection and RAM in a multi-task learning manner. + +# 3.1 Face Disturbance + +We aim to disturb real faces to pseudo-fake faces as training data. Previous works devoted to producing texture inconsistency, which is a typical pattern of + +![](images/fad3aa5bf056e606ed10f05253b34c2b9abee510e0f9d24b29d676cf3d0f149e.jpg) +(a) Face Disturbance + +![](images/f6da5cef8653f48697403a959843065768f8a89185217e9c49ecec0db11cde69.jpg) +(b) Pseudo-fake Examples +Fig. 2: (a) Face Disturbance with TD and SD. (b) Pseudo-fake examples, where for each character, from left to right, are real faces, faces with texture anomalies $I_{ta}$ , and faces with structure anomalies $I_{sa}$ . We highlight their disturbed areas. + +face swap deepfakes. Therefore, these detectors exhibited biased generalization to face swap deepfakes and fell short when detecting reenactment deepfakes in [28-30]. Similarly, employing these data for RAM enables the detector to learn real texture appearance only, making the detector struggle to spot reenactment deepfakes with intact textures. Thus, it is necessary to introduce richer anomalies to help detectors handle a wider spectrum of deepfakes. As an image can be regarded as composed of structures and textures [9,10], we categorize anomalies in fake faces into two categories: structure and texture anomalies. Therefore, we implement Texture Disturbance (TD) and Structure Disturbance (SD) in Face Disturbance to produce these anomalies. By adopting this divide-and-conquer strategy, we comprehensively disturb faces and thus make the detector learn richer real appearances from both the texture and structure views. + +Texture Disturbance The aim of TD is to produce texture anomalies. Given the inherent diversity of real facial textures, which vary in color, brightness, and clarity, a reference facial area is needed to discern whether the textures in other facial areas are abnormal. Namely, texture anomalies lie in the texture inconsistency between facial parts. There are two existing approaches to produce such inconsistency: Face Blending [2, 4] using two faces and Partial Augmentation [23, 24] using one face. Face blending involves selecting two faces with similar poses and then merging the inner region of one face into the other face. Partial Augmentation involves disturbing either the inner or outer facial texture through image augmentation. An important advantage of Partial Augmentation is preserving the original identity, which prevents detectors from being overfitted to the training set identities and harming generalization [31]. Thus, the proposed TD adopts a disturbing scheme similar to Partial Augmentation. + +Figure 2 illustrates the process of TD. The input image $I_{ori}$ is duplicated to $I_{inner}$ and $I_{outer}$ , and one of them is augmented in color, brightness, and clarity. $I_{inner}$ and $I_{outer}$ are then fused using a mask $M$ : + +$$ +I _ {t a} = M \odot I _ {\text {i n n e r}} + (1 - M) \odot I _ {\text {o u t e r}}, \tag {1} +$$ + +where $\odot$ is Hadamard production, $I_{inner}$ is the image that provides the inner face, $I_{outer}$ provides the background, and $I_{ta}$ is the fused face with texture anomaly. $M$ is the mask for the inner face, and its pixel values are between 0 and 1. Specifically, we first apply a landmark detector to the input image to predict its landmarks. Next, we calculate the convex hull of the face area in $I$ to initialize $M$ . To create inconsistent textures between a richer diversity of different areas, we introduce random deformations to $M$ using a 2D elastic transformation. Before using Equation 1 to blend $I_{inner}$ and $I_{outer}$ , we soften $M$ from a binary mask into a soft mask with Gaussian Blur to produce $I_{ta}$ with natural boundaries. This allows the detector to recognize anomalies in long-span texture inconsistencies rather than merely adapt to local boundary artifacts around $M$ . + +Structure Disturbance The aim of SD is to introduce anomalies in the structure view to complement texture anomalies. To disturb facial structures, we disturb the elements that constitute the facial structure. Clearly, these elements are facial boundaries and facial features such as the eyes and mouth. We note that facial landmarks are also positioned at these elements. Therefore, it is intuitive to disturb areas around landmarks to create disturbances in facial structures. Besides, to distinguish with TD, it is also necessary for SD not to disturb the original texture. To disturb facial structures while keeping textures intact, we copy and tile the surrounding textures into areas around a few facial landmarks. The crux lies in the fact that the copy-and-tile operation preserves original textures and thus is the ideal method for disturbing structures. This aids the detector in accurately learning real facial structure in RAM, enabling them to distinguish faces with structural anomalies during detection. + +Figure 2 illustrates the process of SD. Given an input face image $I_{ori}$ , we first detect its facial landmarks. We randomly sample several landmark points and draw circles around them as the mask $M$ for disturbed areas. We avoid sampling landmarks of symmetric positions in the face, which enables the detector to rely on the structure information from the symmetric region to perform RAM. We then employ the Fast Marching Method [32], a non-learning image inpainting algorithm, which produces less blur and maintains clearer textures than other textures replicating algorithms, to copy and tile the surrounding textures to the disturbed area. After the surrounding textures are tiled to the disturbed areas, the original structures of the disturbed areas are eliminated and become anomalous. Incorporating SD in Face Disturbance emrichs anomalies in constructed pseudo-fakes, which helps the detector generalize to a wider range of deepfakes. + +![](images/13953c80bf2e097298f735ce20f5791c3812c1566f29c67d1c6a877705e44e31.jpg) +Fig. 3: The proposed RAE is optimized using two learning objects, i.e., $\mathcal{L}_{rec}$ for the RAM and $\mathcal{L}_{cls}$ for the deepfake detection. + +# 3.2 Real Appearance Modeling + +We believe that guiding models to learn the appearances of real faces is more beneficial for the generalization across deepfake algorithms than learning specific deepfake patterns. Thus, we propose the novel Real Appearance Modeling (RAM) task to guide detectors to learn real face appearances. RAM requires the model to restore anomalous facial areas produced by Face Disturbance to their original appearances. In comparison to existing methods, our model is guided to learn patterns of real faces instead of deepfake patterns in training fake faces. Then, it classifies faces that display anomalies not found in real faces as fake. Therefore, RAM helps to improve generalization across deepfake algorithms. + +Task Definition Given an input face $I_{ori}$ , we obtain pseudo-fake faces $I_{ta}$ with texture anomalies and $I_{sa}$ with structure anomalies after the Face Disturbance strategy. These images are fed into the detector for deepfake detection and RAM, i.e., classification and densely real appearance recovery. In the training procedure, the deepfake detection object and RAM object are optimized in a multi-task learning manner. In deepfake detection, the labels for constructed pseudo-fakes are 1, while for real faces are 0. In RAM, the recovery target $I_{target}$ for $I_{sa}$ is $I$ , while the recovery target $I_{target}$ for $I_{ta}$ is $I_{outer}$ . This is because $I_{ta}$ is derived from $I_{outer}$ and $I_{inner}$ , so there are two potential recovery targets, and we fix the target as $I_{outer}$ . This allows the model to restore the inner texture from $I_{inner}$ using the background texture from $I_{outer}$ as the reference. Thus, the model is guided to align the inner texture with the outer texture. The feasibility of recovering pseudo-fakes to original faces lies in the fact that the pseudo-fakes contains all the information of the corresponding original faces, with only variations between the inner and outer face regions or partially disrupted structure. + +# 3.3 Recovery Autoencoder + +In order to perform both deepfake detection and RAM, the detector in our framework requires a classification head and a dense prediction head. Thus, we + +incorporate a dense prediction branch for the RAM object into a classifier. This configuration makes our model resemble an autoencoder, thus named Recovery Autoencoder (RAE). The RAE is illustrated in Figure 3, which is constructed based on a vanilla Vision Transformer Base (ViT Base) as the encoder. We adopt eight transformer layers and a linear projection layer as the decoder, and a Multi-layer Token Fusion (MTF) block to extract features from both the shallow and deep encoder layers. We refer readers to [33] for detailed information of the ViT encoder. Following this, we detail the entire pipeline of RAE. + +After the image $I \in R^{H \times W \times 3}$ is fed to the ViT encoder, it is extracted to a sequence of tokens $x_{l} \in R^{(N + 1) \times D}$ in each encoder layer. Here, $l \in 1, 2, \dots, L$ is the encoder layer index, $L = 12$ is the depth of the encoder, $N$ is the number of image patches, and $D$ represents the embedding dimension. We note $x_{L}[0]$ is the c1s token [33] for classification. The classification process involves passing the c1s token to a linear layer and a softmax operator to predict the class probabilities $\hat{y} \in \mathbb{R}^2$ , and 2 is the number of classes. Then, the remaining tokens in $x_{l}$ are reshaped to $\mathbb{R}^{h \times w \times D}$ where $h = w = \sqrt{N}$ , and then inputted to MTF for cross-layer token fusion and the decoder for RAM. + +As there are 12 transformer layers in the ViT encoder, inputting all layers in $x_{l}$ to MTF is computationally intensive. Thus, we select $S = 3$ layers at equal intervals and tokens in these layers as the inputs to MTF to provide shallow, mid, and deep features. Formally, the selected tokens of shallow, mid, and deep layers are noted as $x_{s}, x_{m}$ , and $x_{d}$ . In MTF, we first project $x_{s}, x_{m}$ , and $x_{d}$ into the same feature space and concatenate them to $x^{\prime} \in R^{S \times h \times w \times D}$ . We employ a linear layer as the projection layer, which is widely used for aligning different spaces. After projection, we further extract and fuse both structural and textural information from different layers. As tokens in different ViT layers have identical dimensions, we adopt a cross-layer attention mechanism for fusion. Specifically, MTF predicts weights $W \in R^{S \times h \times w}$ for each token with an MLP. $W$ is normalized using softmax along the first dimension $S$ , ensuring that the weights for tokens at the same spatial position across $S$ layers sum to 1. $W$ and fused tokens $x_{fus} \in R^{h \times w \times D}$ are derived as: + +$$ +x ^ {\prime} = \operatorname {c o n c a t} \left(x _ {s}, x _ {m}, x _ {d}\right), \tag {2} +$$ + +$$ +W = \operatorname {s o f t m a x} (\operatorname {M L P} \left(x ^ {\prime}\right)), \tag {3} +$$ + +$$ +x _ {f u s} = \sum_ {i = 0} ^ {S} W [ i ] \times x ^ {\prime} [ i ]. \tag {4} +$$ + +After extracting $x_{fus}$ , we aim to recover the original face with $x_{fus}$ . While it might be intuitive to directly feed $x_{fus}$ into the decoder to recover real faces, this leads to easy mining that outputs the inputted pseudo-fakes directly. This is because the inputted pseudo-fakes and the target real faces are identical in the undisturbed areas $1 - M$ and are similar in disturbed areas $M$ . To avoid such easy mining, we introduce a learnable mask msk token to replace tokens of the disturbed areas. Specifically, given a binary mask $M$ for disturbed areas, it is divided into mask patches $m \in R^{h \times w \times p \times p}$ , where $p$ is the side length of a patch. If a mask patch is disturbed, the corresponding patch is replaced with msk + +token, and the mask for masked patches is $M_{patch}$ . After masking, the masked tokens $x_{msk} \in R^{h \times w \times D}$ are added with positional encodings and then fed into the decoder to $\hat{x} \in R^{N \times p \times p \times 3}$ . Lastly, through a reshape operation, we obtain the recovered face image $\hat{I} \in R^{H \times W \times 3}$ where $H = W = \sqrt{N} \times p$ . We emphasize that the masking is applied after encoding. Thus, the classification head and the predicted $\hat{y}$ remain unaffected. Moreover, information in masked tokens is fused to unmasked tokens in encoding, thus necessary information that enables the recovery is still conveyed to the decoder. + +In predicting the real appearance of disturbed areas, RAE extracts texture and structure features from undisturbed areas and semantic information from disturbed areas for RAM. Consequently, the decoder can utilize the combination of these features to model the appearance of disturbed areas through pixel-level prediction. This enhances the real appearance learned in the encoder, thereby improving the generalization of the deepfake detection. + +# 3.4 Loss Function + +The proposed RAE is optimized to both discriminate disturbed faces and model real appearance. We use cross-entropy loss $\mathcal{L}_{cls}$ for classification: + +$$ +\mathcal {L} _ {c l s} = - y \log \hat {y} - (1 - y) \log (1 - \hat {y}), \tag {5} +$$ + +where $\hat{y}$ is the predicted class probabilities and $y$ is the class label. Besides, we use the following L1-loss $\mathcal{L}_{rec}$ for RAM: + +$$ +\mathcal {L} _ {r e c} = \frac {1}{\left\| M _ {p a t c h} \right\| _ {1}} \left\| M _ {p a t c h} \odot (\hat {I} - I _ {t a r g e t}) \right\| _ {1}, \tag {6} +$$ + +where $\| \cdot \|_1$ is the L1-norm, $M_{patch}$ is the mask for disturbed tokens and thus $\mathcal{L}_{rec}$ is performed on the disturbed tokens. The total loss $\mathcal{L}$ for optimizing RAE is formulated as the weighted sum of $\mathcal{L}_{cls}$ and $\mathcal{L}_{rec}$ : + +$$ +\mathcal {L} = \mathcal {L} _ {c l s} + \alpha \mathcal {L} _ {r e c}. \tag {7} +$$ + +We set $\alpha$ as 0.1 to make $\mathcal{L}_{cls}$ and $\mathcal{L}_{rec}$ converge to the same order of magnitude. + +# 4 Experiments + +# 4.1 Implementation Details + +Dataset. We conduct evaluations on six widely used deepfake datasets, i.e., FaceForensics++ (FF++) [34], Celeb-DF v2 (CDF) [35], DFD [34], DFDC [36], DFDC preview (DFDCp) [36], and WildDeepfake (WDF) [37]. FF++ is a deep-fake dataset containing 4,000 forged videos and 1,000 real videos. All videos are provided in three compression levels: raw, high-quality (HQ), and low-quality (LQ). CDF contains 590 real and 5,639 fake videos corresponding to 59 celebrities using an improved face swap algorithm. DFD contains over 3,000 face swap + +videos. DFDC is a large-scale dataset that contains more than 120,000 deepfake clips generated using eight deepfake techniques. DFDCp is the preview dataset for the Deepfake Detection Challenge. Its fake videos are generated using two face swap algorithms. WDF consists of 10,000 real videos and 10,000 online collected fake videos. For DFDC, we use videos with one identity following [8,17,38] in case of labeling noise. For other datasets, we use the official dataset splits. We use the real videos of $\mathrm{FF} + + (\mathrm{HQ})$ to construct pseudo-fakes for training, while other datasets are for testing. + +Preprocess. We extract 10 frames at equal temporal intervals from each video in training and 32 frames from each video in testing. Then, the faces are cropped using Retinaface [39], maintaining a $15\%$ margin around each face. Subsequently, we utilize Dlib [40] to extract 81 facial landmarks for every face and save them to a key-value database to boost the Face Disturbance process. + +Face Disturbance Details. As for the texture augmentations involved in generating $I_{ta}$ , we employ RGBShift, HueSaturationValue, RandomBrightnessContrast, Downscale, and Sharpen to augment textures of either $I_{inner}$ or $I_{outer}$ to produce inconsistencies. As for the disturbed areas in generating $I_{sa}$ , we draw circles around 2 to 6 random landmarks with a radius sampled between 1/14 and 1/16 of the side length. The number and size of the sampled circles for SD ensure have been found empirically to aid in the effectiveness of RAM tasks as shown in supplementary. Then, all images are resized to $224 \times 224$ as inputs. + +Training Details. We use a ViT Base pretrained on ImageNet [41] to initialize our RAE. RAE is trained using SAM [42] optimizer with a learning rate of 8e-4 and a batch size of 192 for 100 epochs, and we use cosine decay as the learning rate scheduler. Each batch consists of 96 real faces, 48 disturbed faces with texture anomalies, and 48 disturbed faces with structure anomalies. In the PyTorch framework, training with automatic mixed precision requires approximately 60GB of GPU memory. + +# 4.2 Cross-Dataset Evaluation + +In cross-dataset evaluation, detectors are trained using faces in $\mathrm{FF}++(\mathrm{HQ})$ and evaluated on other datasets to assess their generalization to unknown deepfake techniques. Here, we compare the proposed RAE with recent deepfake detection methods. The results are listed in Table 1, and we annotated the input type and training faces. RAE achieves top generalization performance and outperforms state-of-the-art detectors on CDF, DFDC, and DFDCp by $2.6\%$ , $3.4\%$ , and $3.2\%$ , respectively. Compared to detectors trained using fake faces, such as TALL and IDDDM, the superior generalization of RAE demonstrates the effectiveness of using Face Disturbance to generate pseudo-fakes for training. Compared to detectors trained using pseudo-fakes with texture inconsistency, such as SBI and AltFreezing, RAE exhibits better generalization on face swap datasets such as CDF and DFDCp. As the design of TD is similar to previous works, the effectiveness of RAE against face swap deepfakes indicates RAM helps to spot richer anomalies through learning real appearance. Moreover, previous detectors trained with pseudo-fakes fail to exhibit better performance on + +Table 1: Cross-dataset evaluation of deepfake detection methods on CDF, DFD, DFDC, DFDCP, and WDF. All models are trained using FF++(HQ). All compared results are video-level AUC and cited from the original papers or their subsequences. + +
MethodInput TypeYearTraining SetTest Set AUC (%)
Real FakePseudo-fakeCDFDFDDFDCDFDCpWDF
PCL+I2G [4]Frame2021✓✓90.099.067.574.4-
ENB4+SBI [23]Frame2022✓✓92.998.272.085.5-
Mover [20]Frame2023✓✓87.1--78.982.0
RFFR [21]Frame2023✓✓89.0-67.8--
IDDDM [31]Frame2023✓✓91.2-71.5--
SeeABLE [43]Frame2023✓✓87.3-75.986.3-
FTCN [38]Clip2021✓✓86.994.471.074.0-
RealForensics [17]Clip2022✓✓86.9-75.9--
AltFreezing [24]Clip2023✓✓✓89.598.5---
TALL [44]Clip2023✓✓90.8-76.8--
RAE (Ours)Frame2024✓✓95.599.080.289.588.4
+ +Table 2: In-dataset evaluation on four subsets in FF++(HQ). †indicates reproduced on FF++(HQ). + +
MethodTest Set AUC(%)
DFF2FFSNTAvg
SLADD [22]----98.4
ENB4+SBI † [23]99.299.199.096.798.5
SeeABLE [43]99.298.899.196.998.5
RAE(Ours)99.699.199.297.698.9
+ +Table 3: Frame-level FP rate(%) on real face datasets. Lower FP is better. All models are trained on $\mathrm{FF} + + (\mathrm{HQ})$ + +
MethodTest Set FP(%)↓
FFHQVoxCeleb
RECCE [45]62.341.7
UIA-ViT [46]40.618.9
ENB4+SBI [23]57.759.2
RAE(ours)36.96.9
+ +DFDC as it contains reenacted faces that are difficult to be imitated by previous pseudo-fakes. Our method demonstrates a major improvement compared to previous methods as SD helps RAE spot structure anomalies in reenacted faces. + +# 4.3 In-dataset Evaluation + +In addition to cross-dataset evaluation, another frequently used evaluation protocol is in-dataset evaluation. We follow the protocol in [43] to conduct evaluations on the four types of deepfakes in $\mathrm{FF}++$ , i.e., two types of face swap deepfakes DF and FS, and two types of face reenactment deepfakes F2F and NT, where the training data includes both deepfake and pseudo-fake samples. This evaluation protocol helps to assess the ability of our method in enhancing in-dataset deepfake detection. The results are listed in Table 2. The overall performance of RAE on all four deepfake types surpasses previous works. Specifically, RAE outperforms the compared methods in detecting reenacted faces generated by the two reenactment algorithms, i.e., F2F and NT. RAE achieves a detection AUC of $97.6\%$ on NT. This indicates that combining SD and the learning target RAM aids in recognizing distorted facial features in NT deepfakes, while previous works trained with texture inconsistency pseudo-fakes fail. + +# 4.4 Evaluating RAE on Real Face Datasets + +In practical deployment, a low False Positive rate (FP) is important for a deepfake detector, which indicates a low ratio of misclassifying real faces as fakes. For instance, a detector with a low FP rate in an authentication system helps to avoid having users repeatedly submit their portraits, thereby enhancing the user experience. Besides, at the same AUC, a lower FP helps reduce the adjustment of classification thresholds in practical applications. We use the mega-scale face dataset VoxCeleb [47], and the high-quality face dataset FFHQ [48] to test the FP of the proposed RAE on real faces. For comparison, we used three open-source detectors, RECCE [45], SBI [23], and UIA-ViT [46]. + +The FP scores are listed in Table 3. Compared to existing arts, RAE achieves the lowest FP on these two datasets. In stark contrast, the average FP scores of the compared RECCE and SBI are larger than $50\%$ , which indicates they misclassified more than half real faces in these two datasets. This validates that RAM successfully makes the detector learn real face appearances and the trained RAE is less likely to classify a real face as fake. + +# 4.5 Evaluation on Diffusion Deepfakes + +In the cross-dataset evaluation, we validated the generalizability of our framework across five popular deepfake datasets. However, we noticed these datasets primarily consist of face swap and reenactment deepfakes. As a result, the effectiveness of detectors against other types of deepfakes, such as diffusion-generated faces, was not adequately evaluated. Given the fact that a proliferating number of deepfakes are diffusion-generated faces, we conduct evaluations on a diffusion face dataset DeepFakeFace(DFF) [27]. DFF is synthesized by Stable Diffusion Text2Img (T2I), Stable Diffusion Inpaint (Inpaint) [49], and InsightFace (IF) [50]. We conduct evaluation using the three subsets in DFF. For all compared methods, we use their official weights. + +The evaluation results are listed in Table 4. RAE outperforms the compared methods in all three subsets in DFF. More significantly, RAE attains an AUC of $70.8\%$ on average, which improves existing SOTA methods by $2.2\%$ . In stark contrast, the AUC scores of RECCE and UIA-ViT on T2I are less than $50\%$ , which indicates they perform worse than randomly guessing. The results on DFF verify the effectiveness of our proposed method in enhancing the detector's generalization to a wider range of deepfakes. + +# 4.6 Ablation Study + +We conduct ablations to verify the effectiveness of the RAM task, Face Disturbance strategy, and the MTF block. A ViT Base trained using $I_{ta}$ and classification task is used to serve as the Baseline as it is similar to several previous detectors [1, 23]. We incrementally add the proposed RAM task, SD, and MTF block to the baseline to observe their effectiveness. The results are listed in Table + +Table 4: Frame-level AUC on three subsets of diffusion deepfakes in DFF [27]. All models are trained on $\mathrm{FF} + + (\mathrm{HQ})$ + +
MethodTest Set AUC(%)
T2IInpaintIFAvg
RECCE [45]35.151.563.149.9
UIA-ViT [46]45.255.874.058.3
ENB4+SBI [23]62.871.471.568.6
RAE(ours)64.173.075.270.8
+ +Table 5: Ablation studies for the proposed Structure Disturbance, RAM task, and MTF module. + +
MethodTest set AUC(%)
CDFDFDCDFDCp
Baseline92.575.585.9
+RAM94.277.488.6
+Isa94.779.188.4
+MTF95.580.289.5
+ +![](images/2b2b527d290679c571776faed27b7bb5ee6ebe4fd08c699608eb8c1c0f125442.jpg) +Fig. 4: Visualizations of RAM. The left are recovery results for $I_{sa}$ while the right are results for $I_{ta}$ . + +5, which show that each of them contributes to enhancing generalization. Compared to solely using a classification task, adding the RAM task significantly improves generalization by $2.2\%$ on average. Furthermore, utilizing $I_{sa}$ to model real facial structures mainly improves the generalization on the DFDC dataset, which includes reenactment deepfakes. Besides, the MTF block aids in improving generalization by leveraging multi-layer features for modeling real appearances. + +# 4.7 Visualization + +Face Recovery. We visualize the recovered disturbed faces to verify whether RAE learned real face appearances. The inputted disturbed faces, corresponding original faces, masks for disturbed facial areas, and recovered faces are presented in Figure 4. We observe that our model successfully recovers most of the perturbed areas within the anomalous faces. Though the recovered faces are not explicitly clear, the anomalous patterns are largely eliminated. For faces with structure anomalies, disturbed areas are finely recovered to their original structure appearance. For faces with texture anomalies, the texture pattern of the inner faces is aligned to normal appearance. This demonstrates the feasibility of modeling real appearance by recovering anomalous faces to their original appearance, which provides more fine-grained and interpretable supervision. + +![](images/0ba10af368b17f0890b7db15831c67a5fa13776acccfd4d5e117d27159e683b7.jpg) +(a) Baseline model. + +![](images/cf9cededb448e7cda57912f606f3dc7e4b2e63d6dea057fe4f332ffe92bf1c69.jpg) +(b) Our model. +Fig. 5: Distributions of learned representations for faces in $\mathrm{FF} + + (\mathrm{HQ})$ [34] in the baseline model(a) and our model(b). + +Feature Distribution. We apply t-SNE [51] to visualize the features in the last layer of the encoder to illustrate the effect of RAM. We use a ViT Base as the baseline detector and train it using real and fake faces in $\mathrm{FF} + + (\mathrm{HQ})$ . Then, we use the trained RAE and baseline detector to extract features from the test set of $\mathrm{FF} + + (\mathrm{HQ})$ . The results are illustrated in Figure 5. The baseline model learns specific features for each deepfake algorithm, thus separating deepfakes of different types in the feature space. In contrast, the proposed RAE aggregates real faces to a compact cluster in the feature space, which indicates it classifies deepfakes based on the learned general fake patterns. Moreover, RAE mixes different types of deepfakes in the feature space, which confirms that RAM helps to learn general deepfake representations. Compared to the baseline model, the anomalous features learned by RAE are more general across different deepfakes, thus demonstrating better generalization. + +# 5 Conclusion + +In this paper, we address the general deepfake detection through guiding the detector to learn real face appearance by recovering original faces from disturbed faces. Anomalies in both the texture and structure views are introduced to faces through the proposed Face Disturbance, and the detector is guided to recover real appearances from these disturbed faces through the proposed RAM task. We then propose the RAE model to conduct both the deepfake detection and the RAM task. Extensive experiments demonstrate the effectiveness of our method against existing methods in general deepfake detection. Future work involves scaling our method with large-scale human face datasets. + +This project is supported by the National Key Research and Development Program of China (No.2022YFB2702500) + +# References + +1. Yuezun Li and Siwei Lyu. Exposing deepfake videos by detecting face warping artifacts. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2019, Long Beach, CA, USA, June 16-20, 2019, pages 46-52. Computer Vision Foundation / IEEE, 2019. 2, 4, 12 +2. Lingzhi Li, Jianmin Bao, Ting Zhang, Hao Yang, Dong Chen, Fang Wen, and Baining Guo. Face x-ray for more general face forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5001-5010, 2020. 2, 4, 5 +3. Ke Sun, Taiping Yao, Shen Chen, Shouhong Ding, Jilin Li, and Rongrong Ji. Dual contrastive learning for general face forgery detection. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, Thirty-Fourth Conference on Innovative Applications of Artificial Intelligence, IAAI 2022, The Twelveth Symposium on Educational Advances in Artificial Intelligence, EAAI 2022 Virtual Event, February 22 - March 1, 2022, pages 2316-2324. AAAI Press, 2022. 2 +4. Tianchen Zhao, Xiang Xu, Mingze Xu, Hui Ding, Yuanjun Xiong, and Wei Xia. Learning self-consistency for deepfake detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15023-15033, 2021. 2, 4, 5, 11 +5. Yuchen Luo, Yong Zhang, Junchi Yan, and Wei Liu. Generalizing face forgery detection with high-frequency features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16317-16326, 2021. 2 +6. Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In European Conference on Computer Vision, pages 86-103. Springer, 2020. 2, 4 +7. Yuezun Li, Ming-Ching Chang, and Siwei Lyu. In ictu oculi: Exposing AI created fake videos by detecting eye blinking. In 2018 IEEE International Workshop on Information Forensics and Security, WIFS 2018, Hong Kong, China, December 11-13, 2018, pages 1-7. IEEE, 2018. 2 +8. Alexandros Haliassos, Konstantinos Vougioukas, Stavros Petridis, and Maja Pan- tic. Lips don't lie: A generalisable and robust approach to face forgery detection. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 5039-5049. Computer Vision Foundation / IEEE, 2021. 2, 4, 10 +9. Jean-François Aujol, Guy Gilboa, Tony F. Chan, and Stanley J. Osher. Structure-texture image decomposition - modeling, algorithms, and parameter selection. Int. J. Comput. Vis., 67(1):111-136, 2006. 2, 5 +0. Youngjung Kim, Bumsub Ham, Minh N Do, and Kwanghoon Sohn. Structure-texture image decomposition using deep variational priors. IEEE Transactions on Image Processing, 28(6):2692-2704, 2018. 2, 5 +1. Yuezun Li, Ming-Ching Chang, and Siwei Lyu. In ictu oculi: Exposing ai created fake videos by detecting eye blinking. In 2018 IEEE International Workshop on Information Forensics and Security (WIFS), pages 1-7. IEEE, 2018. 3 +2. Umur Aybars Ciftci, Ilke Demir, and Lijun Yin. Fakecatcher: Detection of synthetic portrait videos using biological signals. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 3 +3. Xin Yang, Yuezun Li, and Siwei Lyu. Exposing deep fakes using inconsistent head poses. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 8261-8265. IEEE, 2019. 3 + +14. Darius Afchar, Vincent Nozick, Junichi Yamagishi, and Isao Echizen. Mesonet: a compact facial video forgery detection network. In 2018 IEEE international workshop on information forensics and security (WIFS), pages 1-7. IEEE, 2018. 3 +15. Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multi-attentional deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2185–2194, 2021. 4 +16. Qiqi Gu, Shen Chen, Taiping Yao, Yang Chen, Shouhong Ding, and Ran Yi. Exploiting fine-grained face forgery clues via progressive enhancement learning. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, Thirty-Fourth Conference on Innovative Applications of Artificial Intelligence, IAAI 2022, The Twelveth Symposium on Educational Advances in Artificial Intelligence, EAAI 2022 Virtual Event, February 22 - March 1, 2022, pages 735-743. AAAI Press, 2022. 4 +17. Alexandros Haliassos, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Leveraging real talking faces via self-supervision for robust forgery detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14950-14962, 2022. 4, 10, 11 +18. Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Weiming Zhang, and Nenghai Yu. Self-supervised transformer for deepfake detection, 2022. 4 +19. Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 4 +20. Juan Hu, Xin Liao, Difei Gao, Satoshi Tsutsui, Qian Wang, Zheng Qin, and Mike Zheng Shou. Mover: Mask and recovery based facial part consistency aware method for deepfake video detection, 2023. 4, 11 +21. Liang Shi, Jie Zhang, and Shiguang Shan. Real face foundation representation learning for generalized deepfake detection, 2023. 4, 11 +22. Liang Chen, Yong Zhang, Yibing Song, Lingqiao Liu, and Jue Wang. Self-supervised learning of adversarial examples: Towards good generalizations for deep-fake detections. In CVPR, 2022. 4, 11 +23. Kaede Shiohara and Toshihiko Yamasaki. Detecting deepfakes with self-blended images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18720-18729, 2022. 4, 5, 11, 12, 13 +24. Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, and Houqiang Li. Altfreezing for more general video face forgery detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4129-4138, June 2023. 4, 5, 11 +25. Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. In Hanna M. Wallach, Hugo Larochelle, Alina Beygelzimer, Florence d'Alché-Buc, Emily B. Fox, and Roman Garnett, editors, Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 7135-7145, 2019. 4 +26. Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Zwei Liu. Pose-controllable talking face generation by implicitly modularized audiovisual representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4176-4186, 2021. 4 + +27. Haixu Song, Shiyu Huang, Yinpeng Dong, and Wei-Wei Tu. Robustness and generalizability of deepfake detection: A study with diffusion models. arXiv preprint arXiv:2309.02218, 2023. 4, 12, 13 +28. Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton-Ferrer. The deepfake detection challenge dataset. CoRR, abs/2006.07397, 2020. 5 +29. Gereon Fox, Wentao Liu, Hyeongwoo Kim, Hans-Peter Seidel, Mohamed Elgharib, and Christian Theobalt. Videoforensicshq: Detecting high-quality manipulated face videos. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 5 +30. Jiajun Huang, Xueyu Wang, Bo Du, Pei Du, and Chang Xu. Deepfake mnist+: a deepfake facial animation dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1973-1982, 2021. 5 +31. Baojin Huang, Zhongyuan Wang, Jifan Yang, Jiaxin Ai, Qin Zou, Qian Wang, and Dengpan Ye. Implicit identity driven deepfake face swapping detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4490-4499, 2023. 5, 11 +32. Alexandru C. Telea. An image inpainting technique based on the fast marching method. J. Graphics, GPU, & Game Tools, 9(1):23-34, 2004. 6 +33. Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 8 +34. Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics++: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1-11, 2019. 9, 14 +35. Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deepfake forensics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3207-3216, 2020. 9 +36. Brian Dolhansky, Russ Howes, Ben Pflaum, Nicole Baram, and Cristian Canton-Ferrer. The deepfake detection challenge (DFDC) preview dataset. CoRR, abs/1910.08854, 2019. 9 +37. Bojia Zi, Minghao Chang, Jingjing Chen, Xingjun Ma, and Yu-Gang Jiang. Wilddeepfake: A challenging real-world dataset for deepfake detection. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2382-2390, 2020. 9 +38. Yinglin Zheng, Jianmin Bao, Dong Chen, Ming Zeng, and Fang Wen. Exploring temporal coherence for more general video face forgery detection. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 15024-15034. IEEE, 2021. 10, 11 +39. Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-shot multi-level face localisation in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5203-5212, 2020. 10 +40. Davis E King. Dlib-ml: A machine learning toolkit. The Journal of Machine Learning Research, 10:1755-1758, 2009. 10 + +41. Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2009), 20-25 June 2009, Miami, Florida, USA, pages 248-255. IEEE Computer Society, 2009. 10 +42. Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 10 +43. Nicolas Larue, Ngoc-Son Vu, Vitomir Struc, Peter Peer, and Vassilis Christophides. Seeable: Soft discrepancies and bounded contrastive learning for exposing deep-fakes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21011-21021, 2023. 11 +44. Yuting Xu, Jian Liang, Gengyun Jia, Ziming Yang, Yanhao Zhang, and Ran He. Tall: Thumbnail layout for deepfake video detection. arXiv preprint arXiv:2307.07494, 2023. 11 +45. Junyi Cao, Chao Ma, Taiping Yao, Shen Chen, Shouhong Ding, and Xiaokang Yang. End-to-end reconstruction-classification learning for face forgery detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4113-4122, June 2022. 11, 12, 13 +46. Wanyi Zhuang, Qi Chu, Zhentao Tan, Qiankun Liu, Haojie Yuan, Changtao Miao, Zixiang Luo, and Nenghai Yu. Uia-vit: Unsupervised inconsistency-aware method based on vision transformer for face forgery detection. In European Conference on Computer Vision (ECCV), 2022. 11, 12, 13 +47. Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. arXiv preprint arXiv:1806.05622, 2018. 12 +48. Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. IEEE Trans. Pattern Anal. Mach. Intell., 43(12):4217-4228, 2021. 12 +49. Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 12 +50. DeepInsight. insightface. https://github.com/deepinsight/insightface/, 2021. [Online]. 12 +51. Van Der Maaten Laurens and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9(2605):2579-2605, 2008. 14 \ No newline at end of file diff --git a/2024/Real Appearance Modeling for More General Deepfake Detection/images.zip b/2024/Real Appearance Modeling for More General Deepfake Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d5d3558b69cd501893f140ddd2babe20e1b11f5f --- /dev/null +++ b/2024/Real Appearance Modeling for More General Deepfake Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d3c43d6152ecc0114e1d47887f042ee1640f1a24f5123fdd5e2ebbe18e1e01e6 +size 401384 diff --git a/2024/Real Appearance Modeling for More General Deepfake Detection/layout.json b/2024/Real Appearance Modeling for More General Deepfake Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ebe242d0c5daf6dfab98e62d11142dbe71e50544 --- /dev/null +++ b/2024/Real Appearance Modeling for More General Deepfake Detection/layout.json @@ -0,0 +1,9588 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 111, + 465, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 111, + 465, + 148 + ], + "spans": [ + { + "bbox": [ + 149, + 111, + 465, + 148 + ], + "type": "text", + "content": "Real Appearance Modeling for More General Deepfake Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "text", + "content": "Jiahe Tian" + }, + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "text", + "content": ", Cai Yu" + }, + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "text", + "content": ", Xi Wang" + }, + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "text", + "content": ", Peng Chen" + }, + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "text", + "content": ", Zihao Xiao" + }, + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "text", + "content": ", Jiao Dai" + }, + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "text", + "content": ", Jizhong Han" + }, + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "text", + "content": ", and Yesheng Chai" + }, + { + "bbox": [ + 132, + 167, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 145, + 201, + 467, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 201, + 467, + 224 + ], + "spans": [ + { + "bbox": [ + 145, + 201, + 467, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 201, + 467, + 224 + ], + "type": "text", + "content": " School of Cyber Security, University of Chinese Academy of Sciences, Beijing 100049, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 224, + 479, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 479, + 245 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 479, + 245 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 224, + 479, + 245 + ], + "type": "text", + "content": " Institute of Information Engineering, Chinese Academy of Sciences, Beijing 100094, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 246, + 478, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 246, + 478, + 269 + ], + "spans": [ + { + "bbox": [ + 135, + 246, + 478, + 269 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 135, + 246, + 478, + 269 + ], + "type": "text", + "content": " Institute of Microelectronics, Chinese Academy of Sciences, Beijing 100094, China \n" + }, + { + "bbox": [ + 135, + 246, + 478, + 269 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 135, + 246, + 478, + 269 + ], + "type": "text", + "content": " RealAI Inc., Beijing 100094, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 300, + 453, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 300, + 453, + 487 + ], + "spans": [ + { + "bbox": [ + 160, + 300, + 453, + 487 + ], + "type": "text", + "content": "Abstract. Recent studies in deepfake detection have shown promising results when detecting deepfakes of the same type as those present in training. However, their ability to generalize to unseen deepfakes remains limited. This work improves the generalizable deepfake detection from a simple principle: an ideal detector classifies any face that contains anomalies not found in real faces as fake. Namely, detectors should learn consistent real appearances rather than fake patterns in the training set that may not apply to unseen deepfakes. Guided by this principle, we propose a learning task named Real Appearance Modeling (RAM) that guides the model to learn real appearances by recovering original faces from slightly disturbed faces. We further propose Face Disturbance to produce disturbed faces while preserving original information that enables recovery, which aids the model in learning the fine-grained appearance of real faces. Extensive experiments demonstrate the effectiveness of modeling real appearances to spot richer deepfakes. Our method surpasses existing state-of-the-art methods by a large margin on multiple popular deepfake datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 498, + 453, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 498, + 453, + 520 + ], + "spans": [ + { + "bbox": [ + 159, + 498, + 453, + 520 + ], + "type": "text", + "content": "Keywords: Deepfake Detection " + }, + { + "bbox": [ + 159, + 498, + 453, + 520 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 498, + 453, + 520 + ], + "type": "text", + "content": " Face Forgery Detection " + }, + { + "bbox": [ + 159, + 498, + 453, + 520 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 498, + 453, + 520 + ], + "type": "text", + "content": " Multimedia Forensic" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 543, + 230, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 230, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 230, + 555 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Current deep generative models allow for the generation of realistic fake faces, i.e., deepfakes, making it easy to manipulate face media. Therefore, these techniques can be used for committing fraud, bypassing identity authentication, or spreading false information. The potential misuses raise concerns regarding privacy and reputation, and thus deepfake detectors are needed to counter the risks posed by deepfakes. Moreover, the proliferation of deepfake algorithms presents a challenge to the generalizability of deepfake detection, making it crucial to develop detectors that generalize across various deepfake types." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 155, + 116, + 256, + 204 + ], + "blocks": [ + { + "bbox": [ + 155, + 116, + 256, + 204 + ], + "lines": [ + { + "bbox": [ + 155, + 116, + 256, + 204 + ], + "spans": [ + { + "bbox": [ + 155, + 116, + 256, + 204 + ], + "type": "image", + "image_path": "6ad10ae9a0e37430415dd94175066ffc246a2fe2c628cc6663d6fbfa5b6c7579.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 206, + 249, + 217 + ], + "lines": [ + { + "bbox": [ + 164, + 206, + 249, + 217 + ], + "spans": [ + { + "bbox": [ + 164, + 206, + 249, + 217 + ], + "type": "text", + "content": "(a) Deepfake Detection." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 265, + 123, + 462, + 204 + ], + "blocks": [ + { + "bbox": [ + 265, + 123, + 462, + 204 + ], + "lines": [ + { + "bbox": [ + 265, + 123, + 462, + 204 + ], + "spans": [ + { + "bbox": [ + 265, + 123, + 462, + 204 + ], + "type": "image", + "image_path": "0878a1942adfae22bff02aa922e0bd116c7c34a1d7aa9e07ebca2190f79d3046.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 206, + 397, + 217 + ], + "lines": [ + { + "bbox": [ + 326, + 206, + 397, + 217 + ], + "spans": [ + { + "bbox": [ + 326, + 206, + 397, + 217 + ], + "type": "text", + "content": "(b) The RAM task." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 225, + 482, + 258 + ], + "lines": [ + { + "bbox": [ + 131, + 225, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 131, + 225, + 482, + 258 + ], + "type": "text", + "content": "Fig. 1: Illustration of (a) the Deepfake Detection task and (b) the proposed RAM task. Compared to Deepfake Detection, the RAM task guides models to restore disturbed faces back to their original appearance." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 283, + 482, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 402 + ], + "type": "text", + "content": "Previous works usually tackled deepfake detection by guiding detectors to recognize specific deepfake patterns, such as the boundaries in swapped faces [1, 2], inconsistency [3, 4], frequency anomalies [5, 6], and movement anomalies [7, 8]. They achieved good performance in detecting deepfakes of the same types as those present in training. However, their performance often significantly deteriorates when detecting deepfakes generated by novel algorithms, which mirrors real-world scenarios where various deepfakes emerge. This is due to different deepfake algorithms producing different fake patterns, and learned patterns are not universally applicable to all deepfakes. Namely, more than knowledge of specific deepfake patterns is required to generalize to unseen types of deepfake." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 402, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 617 + ], + "type": "text", + "content": "Though different deepfakes exhibit different patterns, real faces maintain consistent appearance. Thus, it is better to learn the appearance of real faces to verify their authenticity instead of learning deepfake patterns to spot fakes. Namely, an ideal classifier would classify faces as fake if they exhibit patterns not found in real faces, i.e., anomalies. As the detector learns explicit real face appearance, it should be capable of predicting real faces at pixel-level. This guides us to design a task that densely predicts real faces to learn real appearance. To fully define this task, dense prediction targets and input information that assists the prediction are needed. Specifically, we propose the Real Appearance Modeling (RAM) task that recovers corresponding original faces from disturbed faces, as illustrated in Figure 5, where original faces are prediction targets and the inputted disturbed faces contain information needed for prediction. For the RAM task, Deepfakes are not applicable to serve as training data since their original identities or poses are altered, and the necessary information that enables prediction is lost. We thus propose Face Disturbance to disturb real faces to pseudo-fakes as training data for RAM. In the disturbed faces produced by Face Disturbance, the original information is mostly retained. Then, the detector learns real face appearance by predicting original faces." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "To disturb real faces, the proposed Face Disturbance includes two parts, Texture and Structure Disturbance, to introduce anomalies in textures and structures, respectively. This is based on the premise that an image can be regarded as composed of structures and textures [9,10]. Texture Disturbance disturbs partial" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "J. Tian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "type": "text", + "content": "textures to produce anomalies in the form of texture inconsistency. The reason for disturbing partially rather than globally is that distribution of real textures varies, and globally modifying facial textures may fail to produce anomalies. Meanwhile, having inconsistent textures between different facial parts is indeed anomalous. Hence, Texture Disturbance is designed to produce inconsistencies. Structure Disturbance disturbs structure by covering key facial areas, including facial features and boundaries, which are essential components of facial structure. The specific covering method involves copying and tiling the surrounding texture to the disturbed areas to ensure the original texture is intact. Jointly using Texture and Structure Disturbance to produce disturbed faces for RAM helps the detector to learn real structures and textures to spot anomalies in unseen deepfakes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 260, + 482, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 482, + 428 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 482, + 428 + ], + "type": "text", + "content": "The proposed RAM uses disturbed faces produced by Face Disturbance as inputs and the original faces as prediction targets, making this task unsupervised. Therefore, RAM is adopted as an auxiliary task to enhance deepfake detectors. In training, real and disturbed faces are used as inputs, and the detector is expected to both discriminate disturbed faces and recover their real appearances. Specifically, RAM involves guiding the detector to recover structure and texture anomalies. For structure anomalies, the detector recovers the structure of the disturbed areas with the undisturbed areas as the reference. For texture anomalies, the detector recovers the inner texture based on the texture of the outer face. By leveraging learned real appearances, our model is equipped to detect anomalies within unseen deepfakes. As illustrated in Figure 5, our model diverges from the baseline model by grouping various deepfakes to general anomalies rather than identifying specific deepfake patterns. This indicates RAM narrows the gap across deepfakes and thus improves the generalizability." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 429, + 363, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 429, + 363, + 438 + ], + "spans": [ + { + "bbox": [ + 146, + 429, + 363, + 438 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 440, + 480, + 548 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 131, + 440, + 479, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 440, + 479, + 475 + ], + "spans": [ + { + "bbox": [ + 131, + 440, + 479, + 475 + ], + "type": "text", + "content": "- We introduce Real Appearance Modeling to guide detectors to learn the appearance of real faces by restoring disturbed faces to their original appearances, which helps to improve the generalization of deepfake detectors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 476, + 480, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 476, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 131, + 476, + 480, + 512 + ], + "type": "text", + "content": "- We propose Face Disturbance to disturb real faces in texture and structure views as pseudo-fakes, which aids the detector in learning real texture and structure appearance when predicting original faces." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 513, + 480, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 513, + 480, + 548 + ], + "spans": [ + { + "bbox": [ + 131, + 513, + 480, + 548 + ], + "type": "text", + "content": "- We conduct extensive experiments using various evaluation protocols to demonstrate the generalizability of our framework, which surpasses previous methods on multiple popular deepfake datasets." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 568, + 237, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 237, + 580 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 237, + 580 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 596, + 259, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 596, + 259, + 607 + ], + "spans": [ + { + "bbox": [ + 132, + 596, + 259, + 607 + ], + "type": "text", + "content": "2.1 Deepfake Detection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "Early works leveraged biological artifacts, including distorted pupils [11], heartbeat frequency [12], and anomalous head pose [13] for deepfake detection. More recent works treated deepfake detection as a binary classification task and focused on detecting specific deepfake patterns. For instance, [14] investigated" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "type": "text", + "content": "Real Appearance Modeling for More General Deepfake Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": "mesoscopic features in shallow networks that contain rich features, while [15] employed an attention mechanism to combine RGB and texture features. [6, 16] leveraged high-frequency in the discrete cosine domain as complementary modalities. Another direction involves employing proxy tasks for pretraining to learn high-level representations capable of identifying deepfake videos. Examples of such proxy tasks include lip-reading [8] and audio-visual contrastive learning [17, 18]. Recently, the unsupervised learning object MAE [19] is introduced in deepfake detection [20, 21]. These works usually recover masked facial features through dense prediction to model real faces. As they use masked images as input where information is lost, the set recovery target, i.e., original face, is merely one of the multiple feasible solutions. In contrast, RAM for each disturbed face has a definitive prediction target and thus is not an ill-posed problem with multiple solutions. Compared to MAE, RAM is more fine-grained in supervising detectors to model real face appearance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 301, + 346, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 301, + 346, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 301, + 346, + 313 + ], + "type": "text", + "content": "2.2 Pseudo-fakes for Deepfake Detection" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 321, + 482, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 482, + 466 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 482, + 466 + ], + "type": "text", + "content": "Instead of training detectors using deepfakes in deepfake datasets, several studies [1,2,4,22-24] suggested constructing pseudo-fakes for training. Since the introduced pattern in pseudo-fakes is controllable, detectors learn prior patterns accurately. Specifically, they introduced texture inconsistencies to simulate swapped faces and guide detectors to learn this prior pattern. For instance, [1] introduced blur to inner faces to replicate clarity inconsistencies. [2,4,22] blended two faces with similar facial landmarks to generate texture inconsistencies. Moreover, [23, 24] utilized augmentations to produce inconsistencies within a single face, achieving good generalization. Although they generalize well to face swap deepfakes, their performance on deepfakes generated by reenactment algorithms [25, 26] or Diffusion model [27] drops where texture inconsistency is not the major artifacts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 483, + 202, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 202, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 202, + 495 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 508, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 482, + 605 + ], + "type": "text", + "content": "Different from most deepfake detection methods, which are typically trained to spot specific deepfake patterns, our approach emphasizes learning real face appearances through the RAM task. The Face Disturbance strategy is adopted to produce disturbed faces as training data for RAM, as illustrated in Figure 2. The subsequent subsections delve into the details of the proposed Face Disturbance strategy, the definition of the RAM task, and the design of the proposed Recovery Autoencoder (RAE) model in Figure 3, which simultaneously conducts deepfake detection and RAM in a multi-task learning manner." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 621, + 249, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 249, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 249, + 632 + ], + "type": "text", + "content": "3.1 Face Disturbance" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 482, + 666 + ], + "type": "text", + "content": "We aim to disturb real faces to pseudo-fake faces as training data. Previous works devoted to producing texture inconsistency, which is a typical pattern of" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "J. Tian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 148, + 118, + 467, + 231 + ], + "blocks": [ + { + "bbox": [ + 148, + 118, + 467, + 231 + ], + "lines": [ + { + "bbox": [ + 148, + 118, + 467, + 231 + ], + "spans": [ + { + "bbox": [ + 148, + 118, + 467, + 231 + ], + "type": "image", + "image_path": "fad3aa5bf056e606ed10f05253b34c2b9abee510e0f9d24b29d676cf3d0f149e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 269, + 232, + 345, + 242 + ], + "lines": [ + { + "bbox": [ + 269, + 232, + 345, + 242 + ], + "spans": [ + { + "bbox": [ + 269, + 232, + 345, + 242 + ], + "type": "text", + "content": "(a) Face Disturbance" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 151, + 243, + 465, + 274 + ], + "blocks": [ + { + "bbox": [ + 151, + 243, + 465, + 274 + ], + "lines": [ + { + "bbox": [ + 151, + 243, + 465, + 274 + ], + "spans": [ + { + "bbox": [ + 151, + 243, + 465, + 274 + ], + "type": "image", + "image_path": "f6da5cef8653f48697403a959843065768f8a89185217e9c49ecec0db11cde69.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 275, + 354, + 285 + ], + "lines": [ + { + "bbox": [ + 260, + 275, + 354, + 285 + ], + "spans": [ + { + "bbox": [ + 260, + 275, + 354, + 285 + ], + "type": "text", + "content": "(b) Pseudo-fake Examples" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 294, + 482, + 327 + ], + "lines": [ + { + "bbox": [ + 131, + 294, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 131, + 294, + 482, + 327 + ], + "type": "text", + "content": "Fig. 2: (a) Face Disturbance with TD and SD. (b) Pseudo-fake examples, where for each character, from left to right, are real faces, faces with texture anomalies " + }, + { + "bbox": [ + 131, + 294, + 482, + 327 + ], + "type": "inline_equation", + "content": "I_{ta}" + }, + { + "bbox": [ + 131, + 294, + 482, + 327 + ], + "type": "text", + "content": ", and faces with structure anomalies " + }, + { + "bbox": [ + 131, + 294, + 482, + 327 + ], + "type": "inline_equation", + "content": "I_{sa}" + }, + { + "bbox": [ + 131, + 294, + 482, + 327 + ], + "type": "text", + "content": ". We highlight their disturbed areas." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 350, + 482, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 494 + ], + "type": "text", + "content": "face swap deepfakes. Therefore, these detectors exhibited biased generalization to face swap deepfakes and fell short when detecting reenactment deepfakes in [28-30]. Similarly, employing these data for RAM enables the detector to learn real texture appearance only, making the detector struggle to spot reenactment deepfakes with intact textures. Thus, it is necessary to introduce richer anomalies to help detectors handle a wider spectrum of deepfakes. As an image can be regarded as composed of structures and textures [9,10], we categorize anomalies in fake faces into two categories: structure and texture anomalies. Therefore, we implement Texture Disturbance (TD) and Structure Disturbance (SD) in Face Disturbance to produce these anomalies. By adopting this divide-and-conquer strategy, we comprehensively disturb faces and thus make the detector learn richer real appearances from both the texture and structure views." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Texture Disturbance The aim of TD is to produce texture anomalies. Given the inherent diversity of real facial textures, which vary in color, brightness, and clarity, a reference facial area is needed to discern whether the textures in other facial areas are abnormal. Namely, texture anomalies lie in the texture inconsistency between facial parts. There are two existing approaches to produce such inconsistency: Face Blending [2, 4] using two faces and Partial Augmentation [23, 24] using one face. Face blending involves selecting two faces with similar poses and then merging the inner region of one face into the other face. Partial Augmentation involves disturbing either the inner or outer facial texture through image augmentation. An important advantage of Partial Augmentation is preserving the original identity, which prevents detectors from being overfitted to the training set identities and harming generalization [31]. Thus, the proposed TD adopts a disturbing scheme similar to Partial Augmentation." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "type": "text", + "content": "Real Appearance Modeling for More General Deepfake Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "content": "Figure 2 illustrates the process of TD. The input image " + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "I_{ori}" + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "content": " is duplicated to " + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "I_{inner}" + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "I_{outer}" + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "content": ", and one of them is augmented in color, brightness, and clarity. " + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "I_{inner}" + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "I_{outer}" + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "content": " are then fused using a mask " + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 227, + 170, + 481, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 170, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 227, + 170, + 481, + 182 + ], + "type": "interline_equation", + "content": "I _ {t a} = M \\odot I _ {\\text {i n n e r}} + (1 - M) \\odot I _ {\\text {o u t e r}}, \\tag {1}", + "image_path": "38a93c3d44bb9996d88cea79de5e2d2541878f43c9f6caaf6c9c1393383124a0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 197, + 482, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 197, + 482, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 197, + 482, + 329 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 197, + 482, + 329 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 132, + 197, + 482, + 329 + ], + "type": "text", + "content": " is Hadamard production, " + }, + { + "bbox": [ + 132, + 197, + 482, + 329 + ], + "type": "inline_equation", + "content": "I_{inner}" + }, + { + "bbox": [ + 132, + 197, + 482, + 329 + ], + "type": "text", + "content": " is the image that provides the inner face, " + }, + { + "bbox": [ + 132, + 197, + 482, + 329 + ], + "type": "inline_equation", + "content": "I_{outer}" + }, + { + "bbox": [ + 132, + 197, + 482, + 329 + ], + "type": "text", + "content": " provides the background, and " + }, + { + "bbox": [ + 132, + 197, + 482, + 329 + ], + "type": "inline_equation", + "content": "I_{ta}" + }, + { + "bbox": [ + 132, + 197, + 482, + 329 + ], + "type": "text", + "content": " is the fused face with texture anomaly. " + }, + { + "bbox": [ + 132, + 197, + 482, + 329 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 132, + 197, + 482, + 329 + ], + "type": "text", + "content": " is the mask for the inner face, and its pixel values are between 0 and 1. Specifically, we first apply a landmark detector to the input image to predict its landmarks. Next, we calculate the convex hull of the face area in " + }, + { + "bbox": [ + 132, + 197, + 482, + 329 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 132, + 197, + 482, + 329 + ], + "type": "text", + "content": " to initialize " + }, + { + "bbox": [ + 132, + 197, + 482, + 329 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 132, + 197, + 482, + 329 + ], + "type": "text", + "content": ". To create inconsistent textures between a richer diversity of different areas, we introduce random deformations to " + }, + { + "bbox": [ + 132, + 197, + 482, + 329 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 132, + 197, + 482, + 329 + ], + "type": "text", + "content": " using a 2D elastic transformation. Before using Equation 1 to blend " + }, + { + "bbox": [ + 132, + 197, + 482, + 329 + ], + "type": "inline_equation", + "content": "I_{inner}" + }, + { + "bbox": [ + 132, + 197, + 482, + 329 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 197, + 482, + 329 + ], + "type": "inline_equation", + "content": "I_{outer}" + }, + { + "bbox": [ + 132, + 197, + 482, + 329 + ], + "type": "text", + "content": ", we soften " + }, + { + "bbox": [ + 132, + 197, + 482, + 329 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 132, + 197, + 482, + 329 + ], + "type": "text", + "content": " from a binary mask into a soft mask with Gaussian Blur to produce " + }, + { + "bbox": [ + 132, + 197, + 482, + 329 + ], + "type": "inline_equation", + "content": "I_{ta}" + }, + { + "bbox": [ + 132, + 197, + 482, + 329 + ], + "type": "text", + "content": " with natural boundaries. This allows the detector to recognize anomalies in long-span texture inconsistencies rather than merely adapt to local boundary artifacts around " + }, + { + "bbox": [ + 132, + 197, + 482, + 329 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 132, + 197, + 482, + 329 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 363, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 518 + ], + "type": "text", + "content": "Structure Disturbance The aim of SD is to introduce anomalies in the structure view to complement texture anomalies. To disturb facial structures, we disturb the elements that constitute the facial structure. Clearly, these elements are facial boundaries and facial features such as the eyes and mouth. We note that facial landmarks are also positioned at these elements. Therefore, it is intuitive to disturb areas around landmarks to create disturbances in facial structures. Besides, to distinguish with TD, it is also necessary for SD not to disturb the original texture. To disturb facial structures while keeping textures intact, we copy and tile the surrounding textures into areas around a few facial landmarks. The crux lies in the fact that the copy-and-tile operation preserves original textures and thus is the ideal method for disturbing structures. This aids the detector in accurately learning real facial structure in RAM, enabling them to distinguish faces with structural anomalies during detection." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Figure 2 illustrates the process of SD. Given an input face image " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "I_{ori}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ", we first detect its facial landmarks. We randomly sample several landmark points and draw circles around them as the mask " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " for disturbed areas. We avoid sampling landmarks of symmetric positions in the face, which enables the detector to rely on the structure information from the symmetric region to perform RAM. We then employ the Fast Marching Method [32], a non-learning image inpainting algorithm, which produces less blur and maintains clearer textures than other textures replicating algorithms, to copy and tile the surrounding textures to the disturbed area. After the surrounding textures are tiled to the disturbed areas, the original structures of the disturbed areas are eliminated and become anomalous. Incorporating SD in Face Disturbance emrichs anomalies in constructed pseudo-fakes, which helps the detector generalize to a wider range of deepfakes." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "J. Tian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 187, + 113, + 427, + 230 + ], + "blocks": [ + { + "bbox": [ + 187, + 113, + 427, + 230 + ], + "lines": [ + { + "bbox": [ + 187, + 113, + 427, + 230 + ], + "spans": [ + { + "bbox": [ + 187, + 113, + 427, + 230 + ], + "type": "image", + "image_path": "13953c80bf2e097298f735ce20f5791c3812c1566f29c67d1c6a877705e44e31.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 237, + 482, + 260 + ], + "lines": [ + { + "bbox": [ + 131, + 237, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 131, + 237, + 482, + 260 + ], + "type": "text", + "content": "Fig. 3: The proposed RAE is optimized using two learning objects, i.e., " + }, + { + "bbox": [ + 131, + 237, + 482, + 260 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 131, + 237, + 482, + 260 + ], + "type": "text", + "content": " for the RAM and " + }, + { + "bbox": [ + 131, + 237, + 482, + 260 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 131, + 237, + 482, + 260 + ], + "type": "text", + "content": " for the deepfake detection." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 283, + 299, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 283, + 299, + 296 + ], + "spans": [ + { + "bbox": [ + 131, + 283, + 299, + 296 + ], + "type": "text", + "content": "3.2 Real Appearance Modeling" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 301, + 482, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 482, + 410 + ], + "type": "text", + "content": "We believe that guiding models to learn the appearances of real faces is more beneficial for the generalization across deepfake algorithms than learning specific deepfake patterns. Thus, we propose the novel Real Appearance Modeling (RAM) task to guide detectors to learn real face appearances. RAM requires the model to restore anomalous facial areas produced by Face Disturbance to their original appearances. In comparison to existing methods, our model is guided to learn patterns of real faces instead of deepfake patterns in training fake faces. Then, it classifies faces that display anomalies not found in real faces as fake. Therefore, RAM helps to improve generalization across deepfake algorithms." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "text", + "content": "Task Definition Given an input face " + }, + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "inline_equation", + "content": "I_{ori}" + }, + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "text", + "content": ", we obtain pseudo-fake faces " + }, + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "inline_equation", + "content": "I_{ta}" + }, + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "text", + "content": " with texture anomalies and " + }, + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "inline_equation", + "content": "I_{sa}" + }, + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "text", + "content": " with structure anomalies after the Face Disturbance strategy. These images are fed into the detector for deepfake detection and RAM, i.e., classification and densely real appearance recovery. In the training procedure, the deepfake detection object and RAM object are optimized in a multi-task learning manner. In deepfake detection, the labels for constructed pseudo-fakes are 1, while for real faces are 0. In RAM, the recovery target " + }, + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "inline_equation", + "content": "I_{target}" + }, + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "inline_equation", + "content": "I_{sa}" + }, + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "text", + "content": ", while the recovery target " + }, + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "inline_equation", + "content": "I_{target}" + }, + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "inline_equation", + "content": "I_{ta}" + }, + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "inline_equation", + "content": "I_{outer}" + }, + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "text", + "content": ". This is because " + }, + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "inline_equation", + "content": "I_{ta}" + }, + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "text", + "content": " is derived from " + }, + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "inline_equation", + "content": "I_{outer}" + }, + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "inline_equation", + "content": "I_{inner}" + }, + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "text", + "content": ", so there are two potential recovery targets, and we fix the target as " + }, + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "inline_equation", + "content": "I_{outer}" + }, + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "text", + "content": ". This allows the model to restore the inner texture from " + }, + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "inline_equation", + "content": "I_{inner}" + }, + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "text", + "content": " using the background texture from " + }, + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "inline_equation", + "content": "I_{outer}" + }, + { + "bbox": [ + 130, + 426, + 482, + 606 + ], + "type": "text", + "content": " as the reference. Thus, the model is guided to align the inner texture with the outer texture. The feasibility of recovering pseudo-fakes to original faces lies in the fact that the pseudo-fakes contains all the information of the corresponding original faces, with only variations between the inner and outer face regions or partially disrupted structure." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 622, + 276, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 622, + 276, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 622, + 276, + 635 + ], + "type": "text", + "content": "3.3 Recovery Autoencoder" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "In order to perform both deepfake detection and RAM, the detector in our framework requires a classification head and a dense prediction head. Thus, we" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "type": "text", + "content": "Real Appearance Modeling for More General Deepfake Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "incorporate a dense prediction branch for the RAM object into a classifier. This configuration makes our model resemble an autoencoder, thus named Recovery Autoencoder (RAE). The RAE is illustrated in Figure 3, which is constructed based on a vanilla Vision Transformer Base (ViT Base) as the encoder. We adopt eight transformer layers and a linear projection layer as the decoder, and a Multi-layer Token Fusion (MTF) block to extract features from both the shallow and deep encoder layers. We refer readers to [33] for detailed information of the ViT encoder. Following this, we detail the entire pipeline of RAE." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 319 + ], + "type": "text", + "content": "After the image " + }, + { + "bbox": [ + 130, + 212, + 482, + 319 + ], + "type": "inline_equation", + "content": "I \\in R^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 212, + 482, + 319 + ], + "type": "text", + "content": " is fed to the ViT encoder, it is extracted to a sequence of tokens " + }, + { + "bbox": [ + 130, + 212, + 482, + 319 + ], + "type": "inline_equation", + "content": "x_{l} \\in R^{(N + 1) \\times D}" + }, + { + "bbox": [ + 130, + 212, + 482, + 319 + ], + "type": "text", + "content": " in each encoder layer. Here, " + }, + { + "bbox": [ + 130, + 212, + 482, + 319 + ], + "type": "inline_equation", + "content": "l \\in 1, 2, \\dots, L" + }, + { + "bbox": [ + 130, + 212, + 482, + 319 + ], + "type": "text", + "content": " is the encoder layer index, " + }, + { + "bbox": [ + 130, + 212, + 482, + 319 + ], + "type": "inline_equation", + "content": "L = 12" + }, + { + "bbox": [ + 130, + 212, + 482, + 319 + ], + "type": "text", + "content": " is the depth of the encoder, " + }, + { + "bbox": [ + 130, + 212, + 482, + 319 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 212, + 482, + 319 + ], + "type": "text", + "content": " is the number of image patches, and " + }, + { + "bbox": [ + 130, + 212, + 482, + 319 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 212, + 482, + 319 + ], + "type": "text", + "content": " represents the embedding dimension. We note " + }, + { + "bbox": [ + 130, + 212, + 482, + 319 + ], + "type": "inline_equation", + "content": "x_{L}[0]" + }, + { + "bbox": [ + 130, + 212, + 482, + 319 + ], + "type": "text", + "content": " is the c1s token [33] for classification. The classification process involves passing the c1s token to a linear layer and a softmax operator to predict the class probabilities " + }, + { + "bbox": [ + 130, + 212, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\hat{y} \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 130, + 212, + 482, + 319 + ], + "type": "text", + "content": ", and 2 is the number of classes. Then, the remaining tokens in " + }, + { + "bbox": [ + 130, + 212, + 482, + 319 + ], + "type": "inline_equation", + "content": "x_{l}" + }, + { + "bbox": [ + 130, + 212, + 482, + 319 + ], + "type": "text", + "content": " are reshaped to " + }, + { + "bbox": [ + 130, + 212, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{h \\times w \\times D}" + }, + { + "bbox": [ + 130, + 212, + 482, + 319 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 212, + 482, + 319 + ], + "type": "inline_equation", + "content": "h = w = \\sqrt{N}" + }, + { + "bbox": [ + 130, + 212, + 482, + 319 + ], + "type": "text", + "content": ", and then inputted to MTF for cross-layer token fusion and the decoder for RAM." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "text", + "content": "As there are 12 transformer layers in the ViT encoder, inputting all layers in " + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "inline_equation", + "content": "x_{l}" + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "text", + "content": " to MTF is computationally intensive. Thus, we select " + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "inline_equation", + "content": "S = 3" + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "text", + "content": " layers at equal intervals and tokens in these layers as the inputs to MTF to provide shallow, mid, and deep features. Formally, the selected tokens of shallow, mid, and deep layers are noted as " + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "inline_equation", + "content": "x_{s}, x_{m}" + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "inline_equation", + "content": "x_{d}" + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "text", + "content": ". In MTF, we first project " + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "inline_equation", + "content": "x_{s}, x_{m}" + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "inline_equation", + "content": "x_{d}" + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "text", + "content": " into the same feature space and concatenate them to " + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "inline_equation", + "content": "x^{\\prime} \\in R^{S \\times h \\times w \\times D}" + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "text", + "content": ". We employ a linear layer as the projection layer, which is widely used for aligning different spaces. After projection, we further extract and fuse both structural and textural information from different layers. As tokens in different ViT layers have identical dimensions, we adopt a cross-layer attention mechanism for fusion. Specifically, MTF predicts weights " + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "inline_equation", + "content": "W \\in R^{S \\times h \\times w}" + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "text", + "content": " for each token with an MLP. " + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "text", + "content": " is normalized using softmax along the first dimension " + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "text", + "content": ", ensuring that the weights for tokens at the same spatial position across " + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "text", + "content": " layers sum to 1. " + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "text", + "content": " and fused tokens " + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "inline_equation", + "content": "x_{fus} \\in R^{h \\times w \\times D}" + }, + { + "bbox": [ + 130, + 320, + 482, + 488 + ], + "type": "text", + "content": " are derived as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 257, + 498, + 480, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 498, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 257, + 498, + 480, + 510 + ], + "type": "interline_equation", + "content": "x ^ {\\prime} = \\operatorname {c o n c a t} \\left(x _ {s}, x _ {m}, x _ {d}\\right), \\tag {2}", + "image_path": "01c5fa1ea778d11a588b2675efd0555daf2bcd92d9c1cf6bcdb5a8066ef227b5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 255, + 514, + 480, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 514, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 255, + 514, + 480, + 525 + ], + "type": "interline_equation", + "content": "W = \\operatorname {s o f t m a x} (\\operatorname {M L P} \\left(x ^ {\\prime}\\right)), \\tag {3}", + "image_path": "4484280d51608898a6000bbccdf5773e5cd64f62b2f5825c83d578ab5b5fef8a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 246, + 529, + 480, + 549 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 529, + 480, + 549 + ], + "spans": [ + { + "bbox": [ + 246, + 529, + 480, + 549 + ], + "type": "interline_equation", + "content": "x _ {f u s} = \\sum_ {i = 0} ^ {S} W [ i ] \\times x ^ {\\prime} [ i ]. \\tag {4}", + "image_path": "88114c3391e7198b864675203fd173f6c9b8b7ebd3d8026c8f4207b7581c8806.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "After extracting " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "x_{fus}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ", we aim to recover the original face with " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "x_{fus}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ". While it might be intuitive to directly feed " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "x_{fus}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " into the decoder to recover real faces, this leads to easy mining that outputs the inputted pseudo-fakes directly. This is because the inputted pseudo-fakes and the target real faces are identical in the undisturbed areas " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "1 - M" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " and are similar in disturbed areas " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ". To avoid such easy mining, we introduce a learnable mask msk token to replace tokens of the disturbed areas. Specifically, given a binary mask " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " for disturbed areas, it is divided into mask patches " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "m \\in R^{h \\times w \\times p \\times p}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " is the side length of a patch. If a mask patch is disturbed, the corresponding patch is replaced with msk" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "J. Tian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": "token, and the mask for masked patches is " + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "inline_equation", + "content": "M_{patch}" + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": ". After masking, the masked tokens " + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "inline_equation", + "content": "x_{msk} \\in R^{h \\times w \\times D}" + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": " are added with positional encodings and then fed into the decoder to " + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "inline_equation", + "content": "\\hat{x} \\in R^{N \\times p \\times p \\times 3}" + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": ". Lastly, through a reshape operation, we obtain the recovered face image " + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "inline_equation", + "content": "\\hat{I} \\in R^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "inline_equation", + "content": "H = W = \\sqrt{N} \\times p" + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": ". We emphasize that the masking is applied after encoding. Thus, the classification head and the predicted " + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": " remain unaffected. Moreover, information in masked tokens is fused to unmasked tokens in encoding, thus necessary information that enables the recovery is still conveyed to the decoder." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 480, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 480, + 283 + ], + "type": "text", + "content": "In predicting the real appearance of disturbed areas, RAE extracts texture and structure features from undisturbed areas and semantic information from disturbed areas for RAM. Consequently, the decoder can utilize the combination of these features to model the appearance of disturbed areas through pixel-level prediction. This enhances the real appearance learned in the encoder, thereby improving the generalization of the deepfake detection." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 300, + 230, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 230, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 230, + 312 + ], + "type": "text", + "content": "3.4 Loss Function" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 320, + 479, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 320, + 479, + 344 + ], + "spans": [ + { + "bbox": [ + 131, + 320, + 479, + 344 + ], + "type": "text", + "content": "The proposed RAE is optimized to both discriminate disturbed faces and model real appearance. We use cross-entropy loss " + }, + { + "bbox": [ + 131, + 320, + 479, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 131, + 320, + 479, + 344 + ], + "type": "text", + "content": " for classification:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 230, + 354, + 480, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 354, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 230, + 354, + 480, + 365 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c l s} = - y \\log \\hat {y} - (1 - y) \\log (1 - \\hat {y}), \\tag {5}", + "image_path": "b5f52bd9d6cb9f71e91c1b195299ac7b5d55153f913bcecbb4039267d8016bd2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 375, + 479, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 375, + 479, + 399 + ], + "spans": [ + { + "bbox": [ + 131, + 375, + 479, + 399 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 375, + 479, + 399 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 131, + 375, + 479, + 399 + ], + "type": "text", + "content": " is the predicted class probabilities and " + }, + { + "bbox": [ + 131, + 375, + 479, + 399 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 131, + 375, + 479, + 399 + ], + "type": "text", + "content": " is the class label. Besides, we use the following L1-loss " + }, + { + "bbox": [ + 131, + 375, + 479, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 131, + 375, + 479, + 399 + ], + "type": "text", + "content": " for RAM:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 209, + 407, + 479, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 407, + 479, + 433 + ], + "spans": [ + { + "bbox": [ + 209, + 407, + 479, + 433 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e c} = \\frac {1}{\\left\\| M _ {p a t c h} \\right\\| _ {1}} \\left\\| M _ {p a t c h} \\odot (\\hat {I} - I _ {t a r g e t}) \\right\\| _ {1}, \\tag {6}", + "image_path": "aedeee5825c446ed2eae9c038dbb443cd519929681f67e868fd8a03f02a5ef9a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 441, + 479, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 441, + 479, + 476 + ], + "spans": [ + { + "bbox": [ + 131, + 441, + 479, + 476 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 441, + 479, + 476 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_1" + }, + { + "bbox": [ + 131, + 441, + 479, + 476 + ], + "type": "text", + "content": " is the L1-norm, " + }, + { + "bbox": [ + 131, + 441, + 479, + 476 + ], + "type": "inline_equation", + "content": "M_{patch}" + }, + { + "bbox": [ + 131, + 441, + 479, + 476 + ], + "type": "text", + "content": " is the mask for disturbed tokens and thus " + }, + { + "bbox": [ + 131, + 441, + 479, + 476 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 131, + 441, + 479, + 476 + ], + "type": "text", + "content": " is performed on the disturbed tokens. The total loss " + }, + { + "bbox": [ + 131, + 441, + 479, + 476 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 131, + 441, + 479, + 476 + ], + "type": "text", + "content": " for optimizing RAE is formulated as the weighted sum of " + }, + { + "bbox": [ + 131, + 441, + 479, + 476 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 131, + 441, + 479, + 476 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 441, + 479, + 476 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 131, + 441, + 479, + 476 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 268, + 487, + 479, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 487, + 479, + 498 + ], + "spans": [ + { + "bbox": [ + 268, + 487, + 479, + 498 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {c l s} + \\alpha \\mathcal {L} _ {r e c}. \\tag {7}", + "image_path": "feb6c2d5096ecb60bb9364f639335e70a12cda74e0e679d7f6c040294d5686d7.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 507, + 479, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 507, + 479, + 519 + ], + "spans": [ + { + "bbox": [ + 131, + 507, + 479, + 519 + ], + "type": "text", + "content": "We set " + }, + { + "bbox": [ + 131, + 507, + 479, + 519 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 507, + 479, + 519 + ], + "type": "text", + "content": " as 0.1 to make " + }, + { + "bbox": [ + 131, + 507, + 479, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 131, + 507, + 479, + 519 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 507, + 479, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 131, + 507, + 479, + 519 + ], + "type": "text", + "content": " converge to the same order of magnitude." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 537, + 230, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 537, + 230, + 551 + ], + "spans": [ + { + "bbox": [ + 132, + 537, + 230, + 551 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 562, + 279, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 562, + 279, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 562, + 279, + 574 + ], + "type": "text", + "content": "4.1 Implementation Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 582, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 582, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 582, + 479, + 665 + ], + "type": "text", + "content": "Dataset. We conduct evaluations on six widely used deepfake datasets, i.e., FaceForensics++ (FF++) [34], Celeb-DF v2 (CDF) [35], DFD [34], DFDC [36], DFDC preview (DFDCp) [36], and WildDeepfake (WDF) [37]. FF++ is a deep-fake dataset containing 4,000 forged videos and 1,000 real videos. All videos are provided in three compression levels: raw, high-quality (HQ), and low-quality (LQ). CDF contains 590 real and 5,639 fake videos corresponding to 59 celebrities using an improved face swap algorithm. DFD contains over 3,000 face swap" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "type": "text", + "content": "Real Appearance Modeling for More General Deepfake Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "videos. DFDC is a large-scale dataset that contains more than 120,000 deepfake clips generated using eight deepfake techniques. DFDCp is the preview dataset for the Deepfake Detection Challenge. Its fake videos are generated using two face swap algorithms. WDF consists of 10,000 real videos and 10,000 online collected fake videos. For DFDC, we use videos with one identity following [8,17,38] in case of labeling noise. For other datasets, we use the official dataset splits. We use the real videos of " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "\\mathrm{FF} + + (\\mathrm{HQ})" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " to construct pseudo-fakes for training, while other datasets are for testing." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 272 + ], + "type": "text", + "content": "Preprocess. We extract 10 frames at equal temporal intervals from each video in training and 32 frames from each video in testing. Then, the faces are cropped using Retinaface [39], maintaining a " + }, + { + "bbox": [ + 130, + 212, + 482, + 272 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 130, + 212, + 482, + 272 + ], + "type": "text", + "content": " margin around each face. Subsequently, we utilize Dlib [40] to extract 81 facial landmarks for every face and save them to a key-value database to boost the Face Disturbance process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 272, + 482, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 368 + ], + "type": "text", + "content": "Face Disturbance Details. As for the texture augmentations involved in generating " + }, + { + "bbox": [ + 130, + 272, + 482, + 368 + ], + "type": "inline_equation", + "content": "I_{ta}" + }, + { + "bbox": [ + 130, + 272, + 482, + 368 + ], + "type": "text", + "content": ", we employ RGBShift, HueSaturationValue, RandomBrightnessContrast, Downscale, and Sharpen to augment textures of either " + }, + { + "bbox": [ + 130, + 272, + 482, + 368 + ], + "type": "inline_equation", + "content": "I_{inner}" + }, + { + "bbox": [ + 130, + 272, + 482, + 368 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 272, + 482, + 368 + ], + "type": "inline_equation", + "content": "I_{outer}" + }, + { + "bbox": [ + 130, + 272, + 482, + 368 + ], + "type": "text", + "content": " to produce inconsistencies. As for the disturbed areas in generating " + }, + { + "bbox": [ + 130, + 272, + 482, + 368 + ], + "type": "inline_equation", + "content": "I_{sa}" + }, + { + "bbox": [ + 130, + 272, + 482, + 368 + ], + "type": "text", + "content": ", we draw circles around 2 to 6 random landmarks with a radius sampled between 1/14 and 1/16 of the side length. The number and size of the sampled circles for SD ensure have been found empirically to aid in the effectiveness of RAM tasks as shown in supplementary. Then, all images are resized to " + }, + { + "bbox": [ + 130, + 272, + 482, + 368 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 272, + 482, + 368 + ], + "type": "text", + "content": " as inputs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 368, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 482, + 452 + ], + "type": "text", + "content": "Training Details. We use a ViT Base pretrained on ImageNet [41] to initialize our RAE. RAE is trained using SAM [42] optimizer with a learning rate of 8e-4 and a batch size of 192 for 100 epochs, and we use cosine decay as the learning rate scheduler. Each batch consists of 96 real faces, 48 disturbed faces with texture anomalies, and 48 disturbed faces with structure anomalies. In the PyTorch framework, training with automatic mixed precision requires approximately 60GB of GPU memory." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 468, + 289, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 289, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 289, + 479 + ], + "type": "text", + "content": "4.2 Cross-Dataset Evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "In cross-dataset evaluation, detectors are trained using faces in " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{FF}++(\\mathrm{HQ})" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " and evaluated on other datasets to assess their generalization to unknown deepfake techniques. Here, we compare the proposed RAE with recent deepfake detection methods. The results are listed in Table 1, and we annotated the input type and training faces. RAE achieves top generalization performance and outperforms state-of-the-art detectors on CDF, DFDC, and DFDCp by " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ", respectively. Compared to detectors trained using fake faces, such as TALL and IDDDM, the superior generalization of RAE demonstrates the effectiveness of using Face Disturbance to generate pseudo-fakes for training. Compared to detectors trained using pseudo-fakes with texture inconsistency, such as SBI and AltFreezing, RAE exhibits better generalization on face swap datasets such as CDF and DFDCp. As the design of TD is similar to previous works, the effectiveness of RAE against face swap deepfakes indicates RAM helps to spot richer anomalies through learning real appearance. Moreover, previous detectors trained with pseudo-fakes fail to exhibit better performance on" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "J. Tian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 157, + 479, + 292 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 148 + ], + "type": "text", + "content": "Table 1: Cross-dataset evaluation of deepfake detection methods on CDF, DFD, DFDC, DFDCP, and WDF. All models are trained using FF++(HQ). All compared results are video-level AUC and cited from the original papers or their subsequences." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 157, + 479, + 292 + ], + "lines": [ + { + "bbox": [ + 135, + 157, + 479, + 292 + ], + "spans": [ + { + "bbox": [ + 135, + 157, + 479, + 292 + ], + "type": "table", + "html": "
MethodInput TypeYearTraining SetTest Set AUC (%)
Real FakePseudo-fakeCDFDFDDFDCDFDCpWDF
PCL+I2G [4]Frame2021✓✓90.099.067.574.4-
ENB4+SBI [23]Frame2022✓✓92.998.272.085.5-
Mover [20]Frame2023✓✓87.1--78.982.0
RFFR [21]Frame2023✓✓89.0-67.8--
IDDDM [31]Frame2023✓✓91.2-71.5--
SeeABLE [43]Frame2023✓✓87.3-75.986.3-
FTCN [38]Clip2021✓✓86.994.471.074.0-
RealForensics [17]Clip2022✓✓86.9-75.9--
AltFreezing [24]Clip2023✓✓✓89.598.5---
TALL [44]Clip2023✓✓90.8-76.8--
RAE (Ours)Frame2024✓✓95.599.080.289.588.4
", + "image_path": "33551d403954f5fab6745e52b20ce183d6f3c5f68431ed9af11afb692cc37b1c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 137, + 346, + 305, + 409 + ], + "blocks": [ + { + "bbox": [ + 132, + 303, + 307, + 336 + ], + "lines": [ + { + "bbox": [ + 132, + 303, + 307, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 307, + 336 + ], + "type": "text", + "content": "Table 2: In-dataset evaluation on four subsets in FF++(HQ). †indicates reproduced on FF++(HQ)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 346, + 305, + 409 + ], + "lines": [ + { + "bbox": [ + 137, + 346, + 305, + 409 + ], + "spans": [ + { + "bbox": [ + 137, + 346, + 305, + 409 + ], + "type": "table", + "html": "
MethodTest Set AUC(%)
DFF2FFSNTAvg
SLADD [22]----98.4
ENB4+SBI † [23]99.299.199.096.798.5
SeeABLE [43]99.298.899.196.998.5
RAE(Ours)99.699.199.297.698.9
", + "image_path": "5bde5efbb1cea909bccdf5a05f8466fdfda7bac5822d51f398ac05fd7a32e992.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 331, + 346, + 462, + 409 + ], + "blocks": [ + { + "bbox": [ + 312, + 303, + 480, + 336 + ], + "lines": [ + { + "bbox": [ + 312, + 303, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 312, + 303, + 480, + 336 + ], + "type": "text", + "content": "Table 3: Frame-level FP rate(%) on real face datasets. Lower FP is better. All models are trained on " + }, + { + "bbox": [ + 312, + 303, + 480, + 336 + ], + "type": "inline_equation", + "content": "\\mathrm{FF} + + (\\mathrm{HQ})" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 346, + 462, + 409 + ], + "lines": [ + { + "bbox": [ + 331, + 346, + 462, + 409 + ], + "spans": [ + { + "bbox": [ + 331, + 346, + 462, + 409 + ], + "type": "table", + "html": "
MethodTest Set FP(%)↓
FFHQVoxCeleb
RECCE [45]62.341.7
UIA-ViT [46]40.618.9
ENB4+SBI [23]57.759.2
RAE(ours)36.96.9
", + "image_path": "df2047e4ac64919ad9ff2cf51f45fdeaa16fc29881565378c384cecf78e26b18.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 433, + 479, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 433, + 479, + 469 + ], + "spans": [ + { + "bbox": [ + 132, + 433, + 479, + 469 + ], + "type": "text", + "content": "DFDC as it contains reenacted faces that are difficult to be imitated by previous pseudo-fakes. Our method demonstrates a major improvement compared to previous methods as SD helps RAE spot structure anomalies in reenacted faces." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 488, + 269, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 488, + 269, + 499 + ], + "spans": [ + { + "bbox": [ + 132, + 488, + 269, + 499 + ], + "type": "text", + "content": "4.3 In-dataset Evaluation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "In addition to cross-dataset evaluation, another frequently used evaluation protocol is in-dataset evaluation. We follow the protocol in [43] to conduct evaluations on the four types of deepfakes in " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{FF}++" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ", i.e., two types of face swap deepfakes DF and FS, and two types of face reenactment deepfakes F2F and NT, where the training data includes both deepfake and pseudo-fake samples. This evaluation protocol helps to assess the ability of our method in enhancing in-dataset deepfake detection. The results are listed in Table 2. The overall performance of RAE on all four deepfake types surpasses previous works. Specifically, RAE outperforms the compared methods in detecting reenacted faces generated by the two reenactment algorithms, i.e., F2F and NT. RAE achieves a detection AUC of " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "97.6\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " on NT. This indicates that combining SD and the learning target RAM aids in recognizing distorted facial features in NT deepfakes, while previous works trained with texture inconsistency pseudo-fakes fail." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "type": "text", + "content": "Real Appearance Modeling for More General Deepfake Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 358, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 358, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 358, + 128 + ], + "type": "text", + "content": "4.4 Evaluating RAE on Real Face Datasets" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 481, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 481, + 246 + ], + "type": "text", + "content": "In practical deployment, a low False Positive rate (FP) is important for a deepfake detector, which indicates a low ratio of misclassifying real faces as fakes. For instance, a detector with a low FP rate in an authentication system helps to avoid having users repeatedly submit their portraits, thereby enhancing the user experience. Besides, at the same AUC, a lower FP helps reduce the adjustment of classification thresholds in practical applications. We use the mega-scale face dataset VoxCeleb [47], and the high-quality face dataset FFHQ [48] to test the FP of the proposed RAE on real faces. For comparison, we used three open-source detectors, RECCE [45], SBI [23], and UIA-ViT [46]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 247, + 481, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 481, + 318 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 481, + 318 + ], + "type": "text", + "content": "The FP scores are listed in Table 3. Compared to existing arts, RAE achieves the lowest FP on these two datasets. In stark contrast, the average FP scores of the compared RECCE and SBI are larger than " + }, + { + "bbox": [ + 130, + 247, + 481, + 318 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 247, + 481, + 318 + ], + "type": "text", + "content": ", which indicates they misclassified more than half real faces in these two datasets. This validates that RAM successfully makes the detector learn real face appearances and the trained RAE is less likely to classify a real face as fake." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 338, + 332, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 338, + 332, + 350 + ], + "spans": [ + { + "bbox": [ + 132, + 338, + 332, + 350 + ], + "type": "text", + "content": "4.5 Evaluation on Diffusion Deepfakes" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 360, + 481, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 481, + 479 + ], + "type": "text", + "content": "In the cross-dataset evaluation, we validated the generalizability of our framework across five popular deepfake datasets. However, we noticed these datasets primarily consist of face swap and reenactment deepfakes. As a result, the effectiveness of detectors against other types of deepfakes, such as diffusion-generated faces, was not adequately evaluated. Given the fact that a proliferating number of deepfakes are diffusion-generated faces, we conduct evaluations on a diffusion face dataset DeepFakeFace(DFF) [27]. DFF is synthesized by Stable Diffusion Text2Img (T2I), Stable Diffusion Inpaint (Inpaint) [49], and InsightFace (IF) [50]. We conduct evaluation using the three subsets in DFF. For all compared methods, we use their official weights." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 480, + 481, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 480, + 481, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 480, + 481, + 563 + ], + "type": "text", + "content": "The evaluation results are listed in Table 4. RAE outperforms the compared methods in all three subsets in DFF. More significantly, RAE attains an AUC of " + }, + { + "bbox": [ + 130, + 480, + 481, + 563 + ], + "type": "inline_equation", + "content": "70.8\\%" + }, + { + "bbox": [ + 130, + 480, + 481, + 563 + ], + "type": "text", + "content": " on average, which improves existing SOTA methods by " + }, + { + "bbox": [ + 130, + 480, + 481, + 563 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 130, + 480, + 481, + 563 + ], + "type": "text", + "content": ". In stark contrast, the AUC scores of RECCE and UIA-ViT on T2I are less than " + }, + { + "bbox": [ + 130, + 480, + 481, + 563 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 480, + 481, + 563 + ], + "type": "text", + "content": ", which indicates they perform worse than randomly guessing. The results on DFF verify the effectiveness of our proposed method in enhancing the detector's generalization to a wider range of deepfakes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 584, + 237, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 584, + 237, + 596 + ], + "spans": [ + { + "bbox": [ + 132, + 584, + 237, + 596 + ], + "type": "text", + "content": "4.6 Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "We conduct ablations to verify the effectiveness of the RAM task, Face Disturbance strategy, and the MTF block. A ViT Base trained using " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "I_{ta}" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": " and classification task is used to serve as the Baseline as it is similar to several previous detectors [1, 23]. We incrementally add the proposed RAM task, SD, and MTF block to the baseline to observe their effectiveness. The results are listed in Table" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "J. Tian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 157, + 306, + 228 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 309, + 148 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 309, + 148 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 309, + 148 + ], + "type": "text", + "content": "Table 4: Frame-level AUC on three subsets of diffusion deepfakes in DFF [27]. All models are trained on " + }, + { + "bbox": [ + 131, + 114, + 309, + 148 + ], + "type": "inline_equation", + "content": "\\mathrm{FF} + + (\\mathrm{HQ})" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 157, + 306, + 228 + ], + "lines": [ + { + "bbox": [ + 136, + 157, + 306, + 228 + ], + "spans": [ + { + "bbox": [ + 136, + 157, + 306, + 228 + ], + "type": "table", + "html": "
MethodTest Set AUC(%)
T2IInpaintIFAvg
RECCE [45]35.151.563.149.9
UIA-ViT [46]45.255.874.058.3
ENB4+SBI [23]62.871.471.568.6
RAE(ours)64.173.075.270.8
", + "image_path": "08ed0605677945e088fd4d01a3388343a9b0be28f544c5937e5772811a785ee3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 321, + 158, + 472, + 228 + ], + "blocks": [ + { + "bbox": [ + 312, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 312, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 312, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 5: Ablation studies for the proposed Structure Disturbance, RAM task, and MTF module." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 158, + 472, + 228 + ], + "lines": [ + { + "bbox": [ + 321, + 158, + 472, + 228 + ], + "spans": [ + { + "bbox": [ + 321, + 158, + 472, + 228 + ], + "type": "table", + "html": "
MethodTest set AUC(%)
CDFDFDCDFDCp
Baseline92.575.585.9
+RAM94.277.488.6
+Isa94.779.188.4
+MTF95.580.289.5
", + "image_path": "14cffd9f061b32f64c874313f1983133ffd073fb2036b398d68728d3affd95b7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 149, + 240, + 465, + 357 + ], + "blocks": [ + { + "bbox": [ + 149, + 240, + 465, + 357 + ], + "lines": [ + { + "bbox": [ + 149, + 240, + 465, + 357 + ], + "spans": [ + { + "bbox": [ + 149, + 240, + 465, + 357 + ], + "type": "image", + "image_path": "2b2b527d290679c571776faed27b7bb5ee6ebe4fd08c699608eb8c1c0f125442.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 365, + 482, + 388 + ], + "lines": [ + { + "bbox": [ + 131, + 365, + 482, + 388 + ], + "spans": [ + { + "bbox": [ + 131, + 365, + 482, + 388 + ], + "type": "text", + "content": "Fig. 4: Visualizations of RAM. The left are recovery results for " + }, + { + "bbox": [ + 131, + 365, + 482, + 388 + ], + "type": "inline_equation", + "content": "I_{sa}" + }, + { + "bbox": [ + 131, + 365, + 482, + 388 + ], + "type": "text", + "content": " while the right are results for " + }, + { + "bbox": [ + 131, + 365, + 482, + 388 + ], + "type": "inline_equation", + "content": "I_{ta}" + }, + { + "bbox": [ + 131, + 365, + 482, + 388 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 415, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 486 + ], + "type": "text", + "content": "5, which show that each of them contributes to enhancing generalization. Compared to solely using a classification task, adding the RAM task significantly improves generalization by " + }, + { + "bbox": [ + 130, + 415, + 482, + 486 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 130, + 415, + 482, + 486 + ], + "type": "text", + "content": " on average. Furthermore, utilizing " + }, + { + "bbox": [ + 130, + 415, + 482, + 486 + ], + "type": "inline_equation", + "content": "I_{sa}" + }, + { + "bbox": [ + 130, + 415, + 482, + 486 + ], + "type": "text", + "content": " to model real facial structures mainly improves the generalization on the DFDC dataset, which includes reenactment deepfakes. Besides, the MTF block aids in improving generalization by leveraging multi-layer features for modeling real appearances." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 509, + 227, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 227, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 227, + 520 + ], + "type": "text", + "content": "4.7 Visualization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Face Recovery. We visualize the recovered disturbed faces to verify whether RAE learned real face appearances. The inputted disturbed faces, corresponding original faces, masks for disturbed facial areas, and recovered faces are presented in Figure 4. We observe that our model successfully recovers most of the perturbed areas within the anomalous faces. Though the recovered faces are not explicitly clear, the anomalous patterns are largely eliminated. For faces with structure anomalies, disturbed areas are finely recovered to their original structure appearance. For faces with texture anomalies, the texture pattern of the inner faces is aligned to normal appearance. This demonstrates the feasibility of modeling real appearance by recovering anomalous faces to their original appearance, which provides more fine-grained and interpretable supervision." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "type": "text", + "content": "Real Appearance Modeling for More General Deepfake Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 169, + 116, + 302, + 250 + ], + "blocks": [ + { + "bbox": [ + 169, + 116, + 302, + 250 + ], + "lines": [ + { + "bbox": [ + 169, + 116, + 302, + 250 + ], + "spans": [ + { + "bbox": [ + 169, + 116, + 302, + 250 + ], + "type": "image", + "image_path": "0ba10af368b17f0890b7db15831c67a5fa13776acccfd4d5e117d27159e683b7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 255, + 269, + 265 + ], + "lines": [ + { + "bbox": [ + 200, + 255, + 269, + 265 + ], + "spans": [ + { + "bbox": [ + 200, + 255, + 269, + 265 + ], + "type": "text", + "content": "(a) Baseline model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 312, + 116, + 447, + 251 + ], + "blocks": [ + { + "bbox": [ + 312, + 116, + 447, + 251 + ], + "lines": [ + { + "bbox": [ + 312, + 116, + 447, + 251 + ], + "spans": [ + { + "bbox": [ + 312, + 116, + 447, + 251 + ], + "type": "image", + "image_path": "cf9cededb448e7cda57912f606f3dc7e4b2e63d6dea057fe4f332ffe92bf1c69.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 255, + 406, + 265 + ], + "lines": [ + { + "bbox": [ + 351, + 255, + 406, + 265 + ], + "spans": [ + { + "bbox": [ + 351, + 255, + 406, + 265 + ], + "type": "text", + "content": "(b) Our model." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 274, + 482, + 297 + ], + "lines": [ + { + "bbox": [ + 131, + 274, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 131, + 274, + 482, + 297 + ], + "type": "text", + "content": "Fig. 5: Distributions of learned representations for faces in " + }, + { + "bbox": [ + 131, + 274, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\mathrm{FF} + + (\\mathrm{HQ})" + }, + { + "bbox": [ + 131, + 274, + 482, + 297 + ], + "type": "text", + "content": " [34] in the baseline model(a) and our model(b)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 322, + 482, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 482, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 482, + 478 + ], + "type": "text", + "content": "Feature Distribution. We apply t-SNE [51] to visualize the features in the last layer of the encoder to illustrate the effect of RAM. We use a ViT Base as the baseline detector and train it using real and fake faces in " + }, + { + "bbox": [ + 130, + 322, + 482, + 478 + ], + "type": "inline_equation", + "content": "\\mathrm{FF} + + (\\mathrm{HQ})" + }, + { + "bbox": [ + 130, + 322, + 482, + 478 + ], + "type": "text", + "content": ". Then, we use the trained RAE and baseline detector to extract features from the test set of " + }, + { + "bbox": [ + 130, + 322, + 482, + 478 + ], + "type": "inline_equation", + "content": "\\mathrm{FF} + + (\\mathrm{HQ})" + }, + { + "bbox": [ + 130, + 322, + 482, + 478 + ], + "type": "text", + "content": ". The results are illustrated in Figure 5. The baseline model learns specific features for each deepfake algorithm, thus separating deepfakes of different types in the feature space. In contrast, the proposed RAE aggregates real faces to a compact cluster in the feature space, which indicates it classifies deepfakes based on the learned general fake patterns. Moreover, RAE mixes different types of deepfakes in the feature space, which confirms that RAM helps to learn general deepfake representations. Compared to the baseline model, the anomalous features learned by RAE are more general across different deepfakes, thus demonstrating better generalization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 502, + 220, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 220, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 220, + 514 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 640 + ], + "type": "text", + "content": "In this paper, we address the general deepfake detection through guiding the detector to learn real face appearance by recovering original faces from disturbed faces. Anomalies in both the texture and structure views are introduced to faces through the proposed Face Disturbance, and the detector is guided to recover real appearances from these disturbed faces through the proposed RAM task. We then propose the RAE model to conduct both the deepfake detection and the RAM task. Extensive experiments demonstrate the effectiveness of our method against existing methods in general deepfake detection. Future work involves scaling our method with large-scale human face datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "This project is supported by the National Key Research and Development Program of China (No.2022YFB2702500)" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "J. Tian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 139, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 139, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 139, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 138, + 139, + 481, + 183 + ], + "type": "text", + "content": "1. Yuezun Li and Siwei Lyu. Exposing deepfake videos by detecting face warping artifacts. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2019, Long Beach, CA, USA, June 16-20, 2019, pages 46-52. Computer Vision Foundation / IEEE, 2019. 2, 4, 12" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 184, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 184, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 184, + 481, + 226 + ], + "type": "text", + "content": "2. Lingzhi Li, Jianmin Bao, Ting Zhang, Hao Yang, Dong Chen, Fang Wen, and Baining Guo. Face x-ray for more general face forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5001-5010, 2020. 2, 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 228, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 481, + 293 + ], + "type": "text", + "content": "3. Ke Sun, Taiping Yao, Shen Chen, Shouhong Ding, Jilin Li, and Rongrong Ji. Dual contrastive learning for general face forgery detection. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, Thirty-Fourth Conference on Innovative Applications of Artificial Intelligence, IAAI 2022, The Twelveth Symposium on Educational Advances in Artificial Intelligence, EAAI 2022 Virtual Event, February 22 - March 1, 2022, pages 2316-2324. AAAI Press, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 293, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 481, + 335 + ], + "type": "text", + "content": "4. Tianchen Zhao, Xiang Xu, Mingze Xu, Hui Ding, Yuanjun Xiong, and Wei Xia. Learning self-consistency for deepfake detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15023-15033, 2021. 2, 4, 5, 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 337, + 481, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 337, + 481, + 370 + ], + "spans": [ + { + "bbox": [ + 138, + 337, + 481, + 370 + ], + "type": "text", + "content": "5. Yuchen Luo, Yong Zhang, Junchi Yan, and Wei Liu. Generalizing face forgery detection with high-frequency features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16317-16326, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 370, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 481, + 403 + ], + "type": "text", + "content": "6. Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In European Conference on Computer Vision, pages 86-103. Springer, 2020. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 403, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 481, + 446 + ], + "type": "text", + "content": "7. Yuezun Li, Ming-Ching Chang, and Siwei Lyu. In ictu oculi: Exposing AI created fake videos by detecting eye blinking. In 2018 IEEE International Workshop on Information Forensics and Security, WIFS 2018, Hong Kong, China, December 11-13, 2018, pages 1-7. IEEE, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 447, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 481, + 500 + ], + "type": "text", + "content": "8. Alexandros Haliassos, Konstantinos Vougioukas, Stavros Petridis, and Maja Pan- tic. Lips don't lie: A generalisable and robust approach to face forgery detection. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 5039-5049. Computer Vision Foundation / IEEE, 2021. 2, 4, 10" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 501, + 481, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 481, + 534 + ], + "type": "text", + "content": "9. Jean-François Aujol, Guy Gilboa, Tony F. Chan, and Stanley J. Osher. Structure-texture image decomposition - modeling, algorithms, and parameter selection. Int. J. Comput. Vis., 67(1):111-136, 2006. 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 534, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 481, + 567 + ], + "type": "text", + "content": "0. Youngjung Kim, Bumsub Ham, Minh N Do, and Kwanghoon Sohn. Structure-texture image decomposition using deep variational priors. IEEE Transactions on Image Processing, 28(6):2692-2704, 2018. 2, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 567, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 481, + 600 + ], + "type": "text", + "content": "1. Yuezun Li, Ming-Ching Chang, and Siwei Lyu. In ictu oculi: Exposing ai created fake videos by detecting eye blinking. In 2018 IEEE International Workshop on Information Forensics and Security (WIFS), pages 1-7. IEEE, 2018. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "type": "text", + "content": "2. Umur Aybars Ciftci, Ilke Demir, and Lijun Yin. Fakecatcher: Detection of synthetic portrait videos using biological signals. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "3. Xin Yang, Yuezun Li, and Siwei Lyu. Exposing deep fakes using inconsistent head poses. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 8261-8265. IEEE, 2019. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "type": "text", + "content": "Real Appearance Modeling for More General Deepfake Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 133, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 482, + 149 + ], + "type": "text", + "content": "14. Darius Afchar, Vincent Nozick, Junichi Yamagishi, and Isao Echizen. Mesonet: a compact facial video forgery detection network. In 2018 IEEE international workshop on information forensics and security (WIFS), pages 1-7. IEEE, 2018. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 482, + 194 + ], + "type": "text", + "content": "15. Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multi-attentional deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2185–2194, 2021. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 195, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 195, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 195, + 482, + 270 + ], + "type": "text", + "content": "16. Qiqi Gu, Shen Chen, Taiping Yao, Yang Chen, Shouhong Ding, and Ran Yi. Exploiting fine-grained face forgery clues via progressive enhancement learning. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, Thirty-Fourth Conference on Innovative Applications of Artificial Intelligence, IAAI 2022, The Twelveth Symposium on Educational Advances in Artificial Intelligence, EAAI 2022 Virtual Event, February 22 - March 1, 2022, pages 735-743. AAAI Press, 2022. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 273, + 482, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 273, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 133, + 273, + 482, + 316 + ], + "type": "text", + "content": "17. Alexandros Haliassos, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Leveraging real talking faces via self-supervision for robust forgery detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14950-14962, 2022. 4, 10, 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 318, + 481, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 318, + 481, + 339 + ], + "spans": [ + { + "bbox": [ + 132, + 318, + 481, + 339 + ], + "type": "text", + "content": "18. Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Weiming Zhang, and Nenghai Yu. Self-supervised transformer for deepfake detection, 2022. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 341, + 481, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 481, + 383 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 481, + 383 + ], + "type": "text", + "content": "19. Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 385, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 481, + 418 + ], + "type": "text", + "content": "20. Juan Hu, Xin Liao, Difei Gao, Satoshi Tsutsui, Qian Wang, Zheng Qin, and Mike Zheng Shou. Mover: Mask and recovery based facial part consistency aware method for deepfake video detection, 2023. 4, 11" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 419, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 481, + 441 + ], + "type": "text", + "content": "21. Liang Shi, Jie Zhang, and Shiguang Shan. Real face foundation representation learning for generalized deepfake detection, 2023. 4, 11" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "type": "text", + "content": "22. Liang Chen, Yong Zhang, Yibing Song, Lingqiao Liu, and Jue Wang. Self-supervised learning of adversarial examples: Towards good generalizations for deep-fake detections. In CVPR, 2022. 4, 11" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 476, + 481, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 476, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 481, + 509 + ], + "type": "text", + "content": "23. Kaede Shiohara and Toshihiko Yamasaki. Detecting deepfakes with self-blended images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18720-18729, 2022. 4, 5, 11, 12, 13" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 510, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 481, + 553 + ], + "type": "text", + "content": "24. Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, and Houqiang Li. Altfreezing for more general video face forgery detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4129-4138, June 2023. 4, 5, 11" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 555, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 620 + ], + "type": "text", + "content": "25. Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. In Hanna M. Wallach, Hugo Larochelle, Alina Beygelzimer, Florence d'Alché-Buc, Emily B. Fox, and Roman Garnett, editors, Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 7135-7145, 2019. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "26. Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Zwei Liu. Pose-controllable talking face generation by implicitly modularized audiovisual representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4176-4186, 2021. 4" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "J. Tian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "27. Haixu Song, Shiyu Huang, Yinpeng Dong, and Wei-Wei Tu. Robustness and generalizability of deepfake detection: A study with diffusion models. arXiv preprint arXiv:2309.02218, 2023. 4, 12, 13" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 182 + ], + "type": "text", + "content": "28. Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton-Ferrer. The deepfake detection challenge dataset. CoRR, abs/2006.07397, 2020. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 226 + ], + "type": "text", + "content": "29. Gereon Fox, Wentao Liu, Hyeongwoo Kim, Hans-Peter Seidel, Mohamed Elgharib, and Christian Theobalt. Videoforensicshq: Detecting high-quality manipulated face videos. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "text", + "content": "30. Jiajun Huang, Xueyu Wang, Bo Du, Pei Du, and Chang Xu. Deepfake mnist+: a deepfake facial animation dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1973-1982, 2021. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 303 + ], + "type": "text", + "content": "31. Baojin Huang, Zhongyuan Wang, Jifan Yang, Jiaxin Ai, Qin Zou, Qian Wang, and Dengpan Ye. Implicit identity driven deepfake face swapping detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4490-4499, 2023. 5, 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "text", + "content": "32. Alexandru C. Telea. An image inpainting technique based on the fast marching method. J. Graphics, GPU, & Game Tools, 9(1):23-34, 2004. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 325, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 390 + ], + "type": "text", + "content": "33. Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 391, + 481, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 481, + 434 + ], + "type": "text", + "content": "34. Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics++: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1-11, 2019. 9, 14" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 434, + 481, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 481, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 481, + 478 + ], + "type": "text", + "content": "35. Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deepfake forensics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3207-3216, 2020. 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "36. Brian Dolhansky, Russ Howes, Ben Pflaum, Nicole Baram, and Cristian Canton-Ferrer. The deepfake detection challenge (DFDC) preview dataset. CoRR, abs/1910.08854, 2019. 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 511, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 511, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 511, + 481, + 553 + ], + "type": "text", + "content": "37. Bojia Zi, Minghao Chang, Jingjing Chen, Xingjun Ma, and Yu-Gang Jiang. Wilddeepfake: A challenging real-world dataset for deepfake detection. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2382-2390, 2020. 9" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 555, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 599 + ], + "type": "text", + "content": "38. Yinglin Zheng, Jianmin Bao, Dong Chen, Ming Zeng, and Fang Wen. Exploring temporal coherence for more general video face forgery detection. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 15024-15034. IEEE, 2021. 10, 11" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "type": "text", + "content": "39. Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-shot multi-level face localisation in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5203-5212, 2020. 10" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "40. Davis E King. Dlib-ml: A machine learning toolkit. The Journal of Machine Learning Research, 10:1755-1758, 2009. 10" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "type": "text", + "content": "Real Appearance Modeling for More General Deepfake Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 490 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "41. Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2009), 20-25 June 2009, Miami, Florida, USA, pages 248-255. IEEE Computer Society, 2009. 10" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 161, + 481, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 481, + 204 + ], + "type": "text", + "content": "42. Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 10" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 247 + ], + "type": "text", + "content": "43. Nicolas Larue, Ngoc-Son Vu, Vitomir Struc, Peter Peer, and Vassilis Christophides. Seeable: Soft discrepancies and bounded contrastive learning for exposing deep-fakes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21011-21021, 2023. 11" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "44. Yuting Xu, Jian Liang, Gengyun Jia, Ziming Yang, Yanhao Zhang, and Ran He. Tall: Thumbnail layout for deepfake video detection. arXiv preprint arXiv:2307.07494, 2023. 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 281, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 325 + ], + "type": "text", + "content": "45. Junyi Cao, Chao Ma, Taiping Yao, Shen Chen, Shouhong Ding, and Xiaokang Yang. End-to-end reconstruction-classification learning for face forgery detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4113-4122, June 2022. 11, 12, 13" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 325, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 369 + ], + "type": "text", + "content": "46. Wanyi Zhuang, Qi Chu, Zhentao Tan, Qiankun Liu, Haojie Yuan, Changtao Miao, Zixiang Luo, and Nenghai Yu. Uia-vit: Unsupervised inconsistency-aware method based on vision transformer for face forgery detection. In European Conference on Computer Vision (ECCV), 2022. 11, 12, 13" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "type": "text", + "content": "47. Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. arXiv preprint arXiv:1806.05622, 2018. 12" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "type": "text", + "content": "48. Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. IEEE Trans. Pattern Anal. Mach. Intell., 43(12):4217-4228, 2021. 12" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 424, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 445 + ], + "type": "text", + "content": "49. Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 12" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 446, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 468 + ], + "type": "text", + "content": "50. DeepInsight. insightface. https://github.com/deepinsight/insightface/, 2021. [Online]. 12" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "text", + "content": "51. Van Der Maaten Laurens and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9(2605):2579-2605, 2008. 14" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "J. Tian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes/6731bc50-147c-4c01-a7c6-498656963ba3_content_list.json b/2024/Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes/6731bc50-147c-4c01-a7c6-498656963ba3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8ade399bfb432b796b32d93e6f03d7c69c5ef307 --- /dev/null +++ b/2024/Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes/6731bc50-147c-4c01-a7c6-498656963ba3_content_list.json @@ -0,0 +1,1633 @@ +[ + { + "type": "text", + "text": "Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes", + "text_level": 1, + "bbox": [ + 274, + 140, + 728, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Siqi Yang $^{1,\\#}$ , Zhaojun Huang $^{2,3,\\#}$ , Yakun Chang $^{5,6}$ , Bin Fan $^{4}$ , Zhaofei Yu $^{1}$ , and Boxin Shi $^{2,3,1,*}$", + "bbox": [ + 274, + 210, + 727, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Institute for Artificial Intelligence, Peking University", + "bbox": [ + 339, + 253, + 661, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 State Key Lab of Multimedia Info. Processing, School of Computer Science, Peking University", + "bbox": [ + 222, + 267, + 777, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Nat'l Eng. Research Ctr. of Visual Technology, School of Computer Science, Peking University", + "bbox": [ + 222, + 277, + 777, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Nat'l Key Lab of General AI, School of Intelligence Science and Technology, Peking University", + "bbox": [ + 222, + 287, + 777, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{5}$ Institute of Information Science, Beijing Jiaotong University", + "bbox": [ + 321, + 297, + 681, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{6}$ Visual Intelligence +X International Cooperation Joint Laboratory of the Ministry of Education {yousiki, huangzhaojun, binfan, yuzf12, shiboxin}@pku.edu.cn, ykchang@bjtu.edu.cn", + "bbox": [ + 217, + 306, + 782, + 328 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The spike camera continuously records scene radiance with high-speed, high dynamic range, and low data redundancy properties, as a promising replacement for frame-based high-speed cameras. Previous methods for reconstructing color videos from monochromatic spikes are constrained in capturing full-temporal color information due to their reliance on compensating colors from low-speed RGB frames. Applying a Bayer-pattern color filter array to the spike sensor yields mosaicked chromatic spikes, which complicates noise distribution in high-speed conditions. By validating that the noise of short-term frames follows a zero-mean distribution, we leverage this hypothesis to develop a self-supervised denoising module trained exclusively on real-world data. Although noise is reduced in short-term frames, the long-term accumulation of incident photons is still necessary to construct HDR frames. Therefore, we introduce a progressive warping module to generate pseudo long-term exposure frames. This approach effectively mitigates motion blur artifacts in high-speed conditions. Integrating these modules forms a real-data-driven reconstruction method for mosaicked chromatic spikes. Extensive experiments conducted on both synthetic and real-world data demonstrate that our approach is effective in reconstructing 2000FPS color HDR videos with significantly reduced noise and motion blur compared to existing methods.", + "bbox": [ + 259, + 364, + 738, + 656 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Spike camera $\\cdot$ High-speed video $\\cdot$ High dynamic range video $\\cdot$ Color demosaicking $\\cdot$ Self-supervised learning", + "bbox": [ + 259, + 670, + 740, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 724, + 375, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As a new class of image sensors inspired by the fovea of retina, the spike camera [3, 14, 46] offers attractive characteristics such as high temporal resolution (20000Hz), high dynamic range (HDR, $>90\\mathrm{dB}$ ), and low data redundancy [14]. In comparison with conventional digital cameras that capture discrete frames,", + "bbox": [ + 212, + 755, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "# Equal contributions. * Corresponding author.", + "bbox": [ + 215, + 824, + 535, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/088473a1bf3a14e9fb809fe8332d9ac56789bb93302929ef88d113bb42985e0f.jpg", + "image_caption": [ + "(b) High-speed color HDR video", + "Fig. 1: High-speed color HDR video reconstruction from mosaicked chromatic spikes. (a) Comparison of different chromatic spike reconstruction (recon.) methodologies, including the adaptation of bio-inspired [45] and supervised learning [39] monochromatic methods to chromatic spikes and our proposed method. (b) A high-quality color HDR video with a frame rate of 2000FPS reconstructed by our method. The corresponding video is available in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 225, + 145, + 782, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "spike cameras continuously accumulate the photons and trigger a single-bit spike of 1 when the detected photons reach a predefined threshold [14]. High-speed readout of spikes provides the spike camera with the potential for photon counting, and the high-dynamic-range ambient light can be estimated through continuous accumulation of spikes in the temporal domain. Therefore, spike cameras exhibit significant advantages in handling low-level vision tasks in high-speed conditions, such as video frame interpolation [33], HDR imaging [11] and high-frame-rate HDR video reconstruction [3].", + "bbox": [ + 212, + 474, + 787, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the higher demand for visual quality in images and videos, there is a strong desire that images or videos reconstructed from spike cameras can reproduce the real world to the greatest extent, with realistic color appearance and extended dynamic range at high frame rate. However, existing approaches [39, 40, 43, 44, 46] mainly focus on image or video reconstruction with monochromatic spikes, and cannot yet recover a high-speed color HDR video. To introduce color, one alternative practice [11] is to build a hybrid spike-RGB camera system, which exploits the colors captured by the conventional RGB camera to colorize the image reconstructed from spikes. Unfortunately, the color information provided by conventional RGB cameras often suffers from underexposure and overexposure, which inevitably lead to color deviations from the real world. Another strategy [3] to alleviate this problem is to replace the conventional RGB camera with an alternate exposure RGB camera to obtain HDR color, thereby enabling high-speed HDR color video reconstruction. However, the above methods have limitations in high-speed scenes, as the RGB camera has a low frame rate (e.g., 60FPS), which results in the loss of colors during the", + "bbox": [ + 212, + 598, + 787, + 843 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "S. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "intervals between frames. In addition, the corresponding hybrid camera system relies on complex temporal synchronization and optical alignment, and the beam splitter with large space footprints would be detrimental to the construction of compact devices.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reconstructing high-speed color HDR videos based only on a single spike camera not only facilitates the acquisition of full-temporal color information but also offers the advantage of lightweight equipment. To this end, a tractable solution is to assemble a Bayer-pattern color filter array (CFA) in the spike camera, thereby producing mosaicked chromatic spikes. Consequently, a pressing challenge that arises is how to reconstruct high-speed color HDR videos from this new data representation. Nevertheless, directly adapting existing monochromatic spike reconstruction methods [39,40,43,44] to mosaicked chromatic spikes suffers from many issues, as shown in Fig. 1 (a). On the one hand, methods based on biology-inspired models [40,43,44] cannot take advantage of deep neural networks and are susceptible to varying degrees of noise contamination. On the other hand, supervised learning methods (e.g., [39]) trained using synthetic datasets have limited generalization ability when transferring to real-world data, because of the imperfection of spike noise modeling and simulation. These issues hinder the development of high-speed color HDR video reconstruction methods based on chromatic spike cameras and greatly limit the effective deployment of existing methods in real scenes.", + "bbox": [ + 212, + 210, + 787, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we propose a real-data-driven mosaicked chromatic spike reconstruction method to recover high-speed color HDR videos, trained solely on readily accessible unlabelled spike streams. The mathematical analysis behind this requires no specific spike noise modeling or calibration, relying only on insights into the relationship between short-term and long-term frames accumulated from chromatic spikes. We design a deep learning framework that allows training directly on real-world data, naturally avoiding the influence of imperfect spike simulation and generalization gap. Specifically, we first propose a self-supervised denoising module to suppress the noise in short-term accumulation of chromatic spikes, which is trained on real-world spike streams without any annotation. Next, considering the limited dynamic range of short-term accumulated frames, we further present a progressive warping module to simulate a pseudolong exposure, which is effective in reconstructing HDR video frames without motion blur. Finally, we integrate these two modules to achieve higher-quality color HDR video reconstruction.", + "bbox": [ + 212, + 473, + 787, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our approach is able to achieve 2000FPS color HDR video reconstruction from mosaicked chromatic spikes with a single device in a real-data-driven manner, as illustrated in Fig. 1 (b), because of the technical contributions from these two complementary aspects:", + "bbox": [ + 212, + 704, + 787, + 765 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We design a self-supervised denoising module to eliminate the noise from short-term accumulation (with high-speed motion cues) without relying on chromatic spike noise modeling and simulation.", + "bbox": [ + 223, + 794, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes", + "bbox": [ + 230, + 113, + 727, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We propose a progressive warping module with pseudo long-term exposure (with expanded dynamic range) to reconstruct HDR frames without introducing motion blur in dynamic scenes.", + "bbox": [ + 223, + 146, + 787, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 215, + 387, + 233 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video reconstruction with monochromatic spikes. To better satisfy human visual perception and facilitate direct processing by machines, reconstructing the corresponding video frame from spikes has been widely studied. Prevalent image reconstruction methods typically employ the temporal statistical properties of spike cameras [43-46]. Zhu et al. [45] explored the spike generation principle and proposed estimating the firing frequency or firing interval of each pixel to reconstruct dynamic scenes, which are known as texture from playback (TFP) and texture from inter-spike-intervals (TFI), respectively. Recently, another line of work has enhanced firing frequency estimates by mimicking human physiological mechanisms, such as retina-like visual imaging [46] and short-term plasticity [43, 44]. Although these methods bring better interpretability, they require careful window selection to balance noise and motion blur, and the reconstruction effect is not yet satisfactory. The powerful representation capabilities of deep learning also inject vitality into spike-based image reconstruction. Spk2ImgNet [39] is the first spike-to-image deep network architecture based on deformable convolution, achieving impressive results. Further, several self-supervised methods [5,6] have also been developed to reduce the dependence on synthetic datasets. However, these methods are tailored for monochromatic spikes and thus cannot be directly applied to mosaicked chromatic spikes.", + "bbox": [ + 212, + 250, + 787, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Demosaicking and denoising with raw images. The purpose of raw image demosaicking is to recover a full-color image from a sub-sampled mosaicked image that contains potential noise. To deal with the ill-posedness of image demosaicking, various image priors have been introduced to regularize the recovery process, such as heuristics [2, 24], sparsity [23, 36], total variation [7, 13], integral gradient [28], self-similarity [22], and residual interpolation [15]. In addition, researchers designed deep neural networks [16, 29, 30] to automatically learn the prior knowledge required for image demosaicking. Moreover, considering the coupling between demosaicking and denoising, some deep learning-based joint demosaicking and denoising network models [9, 18, 34, 38] have also been developed to efficiently remove mosaicks from noisy raw images.", + "bbox": [ + 212, + 537, + 787, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "HDR video with non-conventional sensors. In recent years, a series of nonconventional sensors, such as event cameras [17], spike cameras [14], and quanta image sensors (QIS) [8], have been invented, empowering the ability to capture HDR information in high-speed motion scenes. QIS-based HDR video reconstruction methods [10, 19, 20] were developed to accomplish perceptual HDR illumination. Additionally, the intensity image reconstructed by event cameras [35] or spike cameras [11] can serve as a proxy for enhancing details in low dynamic range (LDR) images. More recently, a hybrid camera system consisting of a spike camera and an alternating-exposure RGB camera is constructed in [3] for $1000 \\times$", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "S. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5645962b9ba6da6cceab1ca074ce59a65902ea8bcdfc7b594e05238c6175edf4.jpg", + "image_caption": [ + "Fig. 2: Chromatic spike accumulation and firing mechanism. We put a Bayer pattern CFA in the spike camera to filter incident light before the photons arrive at the sensor. Each individual pixel corresponds to one of red, green, and blue channels, accumulates the filtered photons, and triggers a spike signal, then resets when it surpasses the threshold $E_{\\mathrm{th}}$ . Reading out the chromatic spikes forms the chromatic spike planes." + ], + "image_footnote": [], + "bbox": [ + 240, + 148, + 776, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "frame rate HDR video reconstruction. Nevertheless, these methods also fall short when dealing with mosaicked chromatic spikes.", + "bbox": [ + 214, + 386, + 782, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 436, + 380, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We initially present a concise review of the operational principles of spike cameras, encompassing both monochromatic spike cameras and those equipped with Color Filter Array (CFA). Subsequently, we discuss the baseline reconstruction techniques for monochromatic spikes.", + "bbox": [ + 214, + 467, + 782, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Monochromatic spikes. The spike camera persistently accumulates electrons produced by incoming photons. When the electrons accumulated at a pixel surpass the predefined threshold $E_{\\mathrm{th}}$ , it triggers a spike signal, immediately resetting the accumulator to initiate a new cycle of accumulation:", + "bbox": [ + 214, + 527, + 784, + 588 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE (\\mathbf {x}, t) = \\int \\alpha L (\\mathbf {x}, t) d t \\mod E _ {\\mathrm {t h}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 598, + 785, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ denotes the conversion ratio and $L(\\mathbf{x},t)$ denotes the incoming light intensity of pixel $\\mathbf{x}$ at time $t$ . Consequently, the spike signal $S(\\mathbf{x},t)$ can be defined as:", + "bbox": [ + 214, + 633, + 782, + 678 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS (\\mathbf {x}, t) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} E (\\mathbf {x}, t) \\geq E _ {\\mathrm {t h}}, \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 689, + 785, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As all the pixels of a spike sensor are arranged in an array, reading out spikes from the pixel array at each time stamp $t$ yields a spike plane of dimensions $h \\times w$ , where $h$ and $w$ are the height and width of the spike data, respectively. Given the readout frequency of 20000Hz, the spike camera [14] allows for capturing spike planes at a frame rate of 20000FPS.", + "bbox": [ + 214, + 734, + 782, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Chromatic spikes. To incorporate color information while capturing HDR scenes with high-speed motion at the same temporal resolution, we can put a", + "bbox": [ + 214, + 809, + 782, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes", + "bbox": [ + 230, + 113, + 727, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CFA in front of the sensor of the spike camera. In a similar way as conventional cameras, the CFA applied to spike sensors presents a Bayer pattern. Each pixel of the spike camera sensor with CFA corresponds to only one channel among red, green, and blue, while the firing mechanism of spikes remains unchanged. As illustrated in Fig. 2, under the influence of the CFA, a spike signal is generated when the cumulative electrons $E_{c}(\\mathbf{x},t)$ produced by the incident photons of the corresponding spectral band meets the threshold $E_{\\mathrm{th}}$ . The chromatic spike signal $S_{c}(\\mathbf{x},t)$ can be recorded as:", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nS _ {c} (\\mathbf {x}, t) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} E _ {c} (\\mathbf {x}, t) \\geq E _ {\\mathrm {t h}}, \\\\ 0, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 279, + 785, + 318 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $c \\in \\{R, G, B\\}$ . Consequently, similar to the firing process of monochromatic spikes, we can obtain a mosaicked chromatic spike plane at each time stamp $t$ .", + "bbox": [ + 212, + 321, + 784, + 366 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline for chromatic frame reconstruction. Preliminary chromatic frame reconstruction can be achieved similarly to [45] by accumulating the spike signals within a defined temporal window:", + "bbox": [ + 212, + 368, + 785, + 412 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nI _ {c} (\\mathbf {x}, t) = \\frac {M _ {c}}{W} \\cdot \\sum_ {\\tau = t - W / 2} ^ {t + W / 2} S _ {c} (\\mathbf {x}, \\tau), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 425, + 785, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $W$ is the size of the accumulation temporal window, $M_{c}$ represents the maximum dynamic range that can be perceived in channel $c$ . Directly applying Eq. (4) to chromatic spikes as a baseline reconstruction algorithm can only obtain mosaicked chromatic frames, which is similar to the RAW image of conventional frame-based sensors. Demosaicking the chromatic frame is challenging since there is a trade-off between noise and superposition of colors. A short-term temporal window $W$ results in severe noise, whereas a long-term $W$ leads to the spatial superposition of varying color information. Furthermore, both noise distribution of chromatic spikes and object motions are more complex, making the direct application of off-the-shelf demosaicking methods insufficient for recovering high-quality color HDR videos.", + "bbox": [ + 212, + 473, + 787, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Real-data-driven chromatic spikes to video", + "text_level": 1, + "bbox": [ + 214, + 659, + 666, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overview. As shown in Fig. 3, we propose a progressive mechanism for high-speed color HDR video reconstruction. The pipeline starts with the accumulation of mosaicked chromatic spikes to reconstruct short-term frames. While these short-term frames are free of motion blur, they suffer from significant noise contamination. Consequently, as shown in Fig. 3 (a), we propose the chromatic spike denoising module (Sec. 4.1) to suppress the noise. The key insight of the denoising module is that a sufficiently long-term frame (e.g., $W = 10000$ , corresponding to 0.5s) of chromatic spikes can produce promising reconstructions in static (or low-speed) scenes, effectively representing the expectations of short-term frames. Given the limited incident photons in short-term frames, which", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "S. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4110ea8fc7ba3608e4170bfb8130d020434bf2380736740d9b794fd46be56169.jpg", + "image_caption": [ + "Fig. 3: Illustration of the overall pipeline. For each sequence of mosaicked chromatic spikes, we split them into four channels according to the RGGB pattern. The four-channel spikes are fed to our progressive warping module, which aligns multiple short-term frames to achieve \"pseudo-long exposure\". (a) Our chromatic spike denoising module utilizes a self-supervised training strategy, which is pretrained on real-world data. (b) Our progressive warping module consists of $L$ steps of frame warping operations, iteratively achieving high-quality HDR frames. At each step of the progressive warping module, we leverage the pretrained chromatic spikes denoising module to facilitate optical flow estimation. $R_{L}$ is the final output color HDR video with 2000FPS." + ], + "image_footnote": [], + "bbox": [ + 218, + 142, + 787, + 425 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "lack the HDR characteristic, extending the accumulation time is a feasible strategy for HDR reconstruction. As shown in Fig. 3 (b), we design the progressive warping module in Sec. 4.2 to establish a pseudo long exposure. At each step of the warping module, chromatic spikes frames are warped according to the estimated optical flow, providing pseudo long-term frames with less noise and higher dynamic range. To leverage a wide range of existing conventional RGB optical flow estimation approaches, we integrate the denoising module into each iteration step of the warping module. After $L$ times of iteration, the progressive warping module produces an HDR color video with a frame rate of 2000FPS.", + "bbox": [ + 212, + 589, + 787, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Denoising of chromatic spikes", + "text_level": 1, + "bbox": [ + 215, + 751, + 506, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Noise distribution of chromatic spikes noise. The noise in monochromatic spikes is primarily categorized into two types: diffuse noise and inherent noise. For the mosaicked chromatic spikes, although the incorporation of a CFA slightly alters the noise distribution for pixels under different filters, pixels corresponding", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes", + "bbox": [ + 230, + 113, + 727, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to the same color exhibit similar noise distribution patterns as monochromatic spikes, except for its expectation and variance. Given the chromatic spikes $S_{c}$ , for each $c \\in \\{R,G,B\\}$ , we use $Q_{d}$ to denote the number of diffuse noise photons arriving at a pixel in a time period of $t$ , and $\\gamma$ to denote the likelihood of $Q_{d}$ diffuse noise photons. The probability of diffuse noise can be modeled as the Poisson distribution [41]:", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} \\left(Q _ {d} | t, \\gamma\\right) = \\frac {\\left(\\gamma t\\right) ^ {Q _ {d}}}{Q _ {d} !} \\exp (- \\gamma W). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 250, + 785, + 284 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hence, the expected number of diffuse noise spikes in temporal window $W$ is", + "bbox": [ + 215, + 289, + 771, + 305 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} \\left[ N _ {d} \\right] = (\\alpha \\gamma W) / E _ {\\mathrm {t h}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 321, + 785, + 339 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For inherent noise, which triggers spikes in the absence of light, we use $T_{i}$ to indicate the interval between the noise spikes. Then, the interval between such noise spikes can be represented by a Gaussian model [41]:", + "bbox": [ + 214, + 345, + 787, + 393 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {N} (T _ {i} | \\mu , \\sigma) = \\frac {1}{\\sigma \\sqrt {2 \\pi}} \\cdot \\exp \\left(- \\frac {1}{2} \\cdot \\left(\\frac {T _ {i} - \\mu}{\\sigma}\\right) ^ {2}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 405, + 785, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We approximate the expectation of the number of inherent noise spikes in the temporal window as", + "bbox": [ + 214, + 441, + 787, + 472 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} [ N _ {i} ] = \\int_ {1} ^ {+ \\infty} \\frac {W}{T _ {i}} \\cdot \\frac {1}{\\sigma \\sqrt {2 \\pi}} \\cdot \\exp (- \\frac {1}{2} \\cdot (\\frac {T _ {i} - \\mu}{\\sigma}) ^ {2}) d T _ {i} = C _ {i} W, (8)\n$$\n", + "text_format": "latex", + "bbox": [ + 292, + 487, + 785, + 520 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $C_i$ denotes the integral after factoring out $W$ from the integral expression. Combining Eq. (6) and Eq. (8), the expectation of the fraction of noise spikes can be formulated as:", + "bbox": [ + 214, + 526, + 787, + 571 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} [ N ] / W = \\left(\\mathbb {E} \\left[ N _ {d} \\right] + \\mathbb {E} \\left[ N _ {i} \\right]\\right) / W = \\alpha \\gamma / E _ {\\mathrm {t h}} + C _ {i}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 333, + 589, + 785, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Instead of the actual incoming light intensity, we estimate the expectation of short-term accumulations $\\mathbb{E}[I_c(\\mathbf{x},t)] = L(\\mathbf{x},t) + \\mathbb{E}[N] / W$ . Thus, the noise between short-term frames and their expectations naturally follows a zero-mean distribution. We demonstrate that the expected fraction of noise spikes is relatively small, especially with sufficient light illumination, and doesn't disturb the visual semantic. As shown in Fig. 4 (a), the noise ratio $(I_{c}(\\mathbf{x},t) / \\mathbb{E}[I_{c}(\\mathbf{x},t)] - 1)$ in the accumulated images $(I_{c}(\\mathbf{x},t))$ gradually decreases as the accumulation window size $W$ increases. Note that the larger window accumulation is equivalent to the average of multiple smaller windows. The convergence demonstrates the existence of expectation and the result faithfully reflects the original light intensity with negligible disturbance. We further validate the zero-mean distribution hypothesis by collecting a series of static scene spike streams with varying light illuminations, and visualize the distribution of noise ratio in Fig. 4 (b). The result shows that the zero-mean assumption holds with sufficient illumination, which supports our self-supervised learning mechanism.", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "S. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/873a2cf93fcb65972ef4faeba1ff17c2e1010b824804e1a25a48be79ee65a821.jpg", + "image_caption": [ + "(a) Temporal accumulation" + ], + "image_footnote": [], + "bbox": [ + 217, + 161, + 576, + 224 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/d30b55fc59d7f7a9e81a99da41cc6da9cc45b46517beb72d70a1bd63f0508782.jpg", + "image_caption": [ + "(b) Noise distribution", + "Fig. 4: Validation of noise on real-world chromatic spikes. As shown in (a), as the accumulation window size $W$ increases, the noise ratio in the accumulated images gradually decreases, making the images more closely resemble the static scene. In (b), we present the statistical noise fraction distribution across three different color channels and different illumination levels." + ], + "image_footnote": [], + "bbox": [ + 578, + 143, + 784, + 228 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Self-supervised chromatic spike denoising module. Taking the inspiration of Batson et al. [1] and Wang et al. [31], which do not require paired noise-contaminated images or clean targets, we design a chromatic spike denoising module trained on only accessible real-world data with self-supervision learning strategy. Specifically, our denoising module operates based on the zero-mean characteristic of short-term frame noise and relies on the blind spot architecture to prevent the network from learning the identity mapping. Hence, the denoising module transforms the noise-contaminated images to their expectations to achieve minimal losses, producing desired clean images. Therefore, as shown in Fig. 3 (a), we begin by extracting the four channels of the chromatic spikes according to the RGGB pattern of the CFA, and the size of each channel is $\\frac{h}{2} \\times \\frac{w}{2}$ . For each channel $c$ , we preliminarily reconstruct a chromatic frame $I_{c}$ by accumulating the chromatic spikes in a temporal window $W$ . Then, $I_{c}$ is further partitioned into $\\frac{h}{2h_s} \\times \\frac{w}{2w_s}$ cells, where $h_s$ and $w_s$ denote height and width of each cell. We sequentially mask one pixel for each cell following the order from top to bottom and left to right. Consequently, we obtain a masked volume with a size of $C \\times (\\frac{h}{2} \\times \\frac{w}{2})$ corresponding to $I_{c}$ , where $C$ is the channel of the masked volume that equals $h_s \\times w_s$ . Each channel of the mask volume acts as the blind spot operation to avoid identity mapping. We then employ a U-Net-inspired fully-convolutional neural network $f_{\\theta}(\\cdot)$ to denoise each masked frame in this volume and merge the denoised frames to the clean image:", + "bbox": [ + 212, + 353, + 787, + 672 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nD \\left(I _ {c}\\right) = \\Phi \\left(f _ {\\theta} \\left(\\Omega_ {c, 0}\\right), f _ {\\theta} \\left(\\Omega_ {c, 1}\\right), \\dots , f _ {\\theta} \\left(\\Omega_ {c, C}\\right)\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 686, + 784, + 704 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\Omega_{c,i}$ denotes the $i$ -th masked frame in volume $\\Omega_c$ , $\\varPhi(\\cdot)$ represents a function for mask-aware averaging:", + "bbox": [ + 212, + 710, + 782, + 742 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi \\left(f _ {\\theta} \\left(\\Omega_ {c, 0}\\right), \\dots\\right) = \\sum_ {i} M _ {i} f _ {\\theta} \\left(\\Omega_ {c, i}\\right) / \\sum_ {i} M _ {i}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 755, + 784, + 785 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "and $M_{i}$ denotes the mask for $\\varOmega_{c,i}(\\varOmega_{c,i}=(1-M_i)I_c)$ . At last, the denoised four-channel images are recombined to form the final mosaicked chromatic frame. It is remarkable that the zero-mean hypothesis holds for arbitrary temporal window", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes", + "bbox": [ + 230, + 113, + 727, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/dd5c16658422c8ea48a85e862587f1e9166987e0def46f1d59f24ebb237a1f7a.jpg", + "image_caption": [ + "Fig. 5: Progressive warping results. We visualize the intermediate frames from our progressive warping module. As the iterative step $l$ increases, the noise in the reconstructed frames is gradually suppressed and the dynamic range is extended." + ], + "image_footnote": [], + "bbox": [ + 218, + 150, + 782, + 220 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "size $W$ , indicating that our neural network can be trained and applied on varying chosen $W$ .", + "bbox": [ + 212, + 306, + 782, + 335 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Loss function. The chromatic spike denoising module is trained using real-world data with realistic noise. Similar to Wang et al. [31], we adopt the regularized re-visible loss:", + "bbox": [ + 212, + 335, + 782, + 380 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r}} = \\mathcal {L} _ {\\mathrm {r e v}} + \\eta \\mathcal {L} _ {\\mathrm {r e g}} = \\| D (I _ {c}) + \\lambda f _ {\\theta} (I _ {c}) - (\\lambda + 1) I _ {c} \\| _ {2} ^ {2} + \\eta \\| D (I _ {c}) - I _ {c} \\| _ {2} ^ {2} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 235, + 406, + 784, + 424 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\eta$ is a constant hyper-parameter, and $\\lambda$ is adjusted in the training progress. We further enhance the denoising module by incorporating an image quality assignment loss, which encourages the predicted score to be close to 1: $\\mathcal{L}_{\\mathrm{i}} = \\| 1 - G_{\\mathrm{iqa}}(D(I_c))\\|_2^2$ , where $G_{\\mathrm{iqa}}$ denotes a pre-trained image quality assignment model [4]. The total training objective combines these losses: $\\mathcal{L} = \\mathcal{L}_{\\mathrm{r}} + \\lambda_{\\mathrm{i}}\\mathcal{L}_{\\mathrm{i}}$ . Note that the training objectives require no clean frames or human annotations, but only readily accessible real-world mosaicked chromatic spike streams.", + "bbox": [ + 212, + 434, + 787, + 541 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Color and HDR via progressive warping", + "text_level": 1, + "bbox": [ + 214, + 560, + 591, + 575 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Although our chromatic spike denoising module can suppress the noise in frames reconstructed from short-term temporal windows, the dynamic range of each short-term frame is still restricted, as the number of photons arriving at the pixels within an extremely short period (e.g., $W = 10$ ) is limited. Moreover, though our denoising module can fit arbitrary temporal window $W$ theoretically, longer-term input frames lead to better output quality. To reconstruct images with less noise and higher dynamic range, it is necessary to select longer windows to preserve sufficient photons. However, directly warping the frames over the longer temporal window (e.g., $W = 100$ ) is challenging since high-speed motions may span a wide range of pixels. Thus, as illustrated in Fig. 3 (b), we develop the progressive warping module, which only warps a small number of $K$ chromatic frames in each step. By progressively applying the warping operation to the output of the previous step, at the $L$ -th step, each output frame is equivalent to the weighted accumulation from $[(K - 1) \\times L + 1] \\times W$ aligned chromatic spike planes. Setting a large value of $L$ for the progressive warping module is equivalent to creating a pseudo-long exposure. The pseudo-long exposure leverages optical flow estimation at each step to obtain HDR frames without motion blur.", + "bbox": [ + 212, + 583, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "S. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As shown in Fig. 3 (b), the progressive warping module involves the application of the pretrained chromatic spike denoising module from Sec. 4.1 and demosaicking for better optical flow estimation. In the initial 0-th step, we split $T$ chromatic spike planes using a short-term temporal window $W$ and accumulate the chromatic spikes in each window to form $\\frac{T}{W}$ mosaicked frames with Eq. (4). We denote these initial mosaicked frames as $I_{c,0}(\\mathbf{x},t)$ , where $c \\in \\{R,G,B\\}$ , which is contaminated by noise given a small temporal window. Afterward, for more accurate optical flow estimation, we feed $I_{c,0}(\\mathbf{x},t)$ to the self-supervised denoising module and the demosaicking module that employs linear interpolation to generate intermediate reconstruction frames:", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\nR _ {0} (t) = \\Psi [ D (I _ {R, 0} (t)), D (I _ {G, 0} (t)), D (I _ {B, 0} (t)) ], \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 321, + 785, + 338 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $\\varPsi$ denotes linear interpolation demosaicking and $\\mathbf{x}$ is omitted. We then estimate the optical flow $F_{0}$ with off-the-shelf methods from $R_0$ [12]. Note that in this initial step, $I_{c,0}(\\mathbf{x},t)$ is not warped and is mainly used to estimate a better initial optical flow $F_{0}$ . Subsequently, $F_{0}$ is leveraged to warp $K\\times W$ spike planes towards the reference frame in the 1-th step, described as:", + "bbox": [ + 214, + 349, + 787, + 426 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\nI _ {c, 1} (\\mathbf {x}, t) = \\frac {M _ {c}}{K W} \\cdot \\sum_ {t ^ {\\prime} = t - K W / 2} ^ {t + K W / 2} S _ {c} (\\mathbf {x} + F _ {0} (\\mathbf {x}, t, t ^ {\\prime}), t ^ {\\prime}). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 445, + 785, + 489 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Note that we utilize chromatic spike planes in the first two steps and mosaicked frames are warped in the later steps. Specifically, for the $l$ -th iteration step, we continuously pass the accumulated frame $I_{c,l-1}$ to the denoising module and the demosaicking module to facilitate the estimation of optical flow $F_{l-1}$ from the previously generated frames $R_{l-1}$ , and warp $I_{c,l-1}$ to the reference time. This operation can further promote noise suppression and dynamic range extension. In particular, we average $K$ accumulated frames $I_{c,l-1}$ obtained in the $(l-1)$ -th step, yielding:", + "bbox": [ + 212, + 503, + 787, + 625 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\nI _ {c, l} (\\mathbf {x}, t) = \\frac {1}{K} \\cdot \\sum_ {i = - K / 2} ^ {K / 2} I _ {c, l - 1} (\\mathbf {x} + F _ {l - 1} (\\mathbf {x}, t, t ^ {\\prime}), t ^ {\\prime}), \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 645, + 785, + 688 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $t' = t + iW$ . The entire procedure is iteratively repeated for $L$ steps. In Fig. 5, we show a set of examples to demonstrate the effectiveness of the progressive warping module. It is evident that as the iteration step $l$ increases, the details in both highlighted dark regions become clearer.", + "bbox": [ + 212, + 700, + 785, + 762 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Finally, we obtain color HDR video frames $R_{L}$ with a frame rate of $\\frac{20000}{W}$ FPS. Note that $W$ is adjustable, which leads to higher frame rates for smaller values and lower frame rates for larger values. In our most experiments, we find that $W = 10$ offers a good balance between temporal resolution and reconstruction quality, which ensures 2000FPS high-speed color HDR video reconstruction.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes", + "bbox": [ + 230, + 113, + 727, + 130 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/cc72206286cdabc1ba3833f93db53c6000e52ff8f7b3cf851429b593a7389210.jpg", + "image_caption": [ + "(a) Synthetic chromatic spike streams and reconstructions" + ], + "image_footnote": [], + "bbox": [ + 228, + 148, + 488, + 202 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3627e88a3ec9de2070838e63d9324d5bf1514617c639f847a1ab1e86bb569626.jpg", + "image_caption": [ + "(b) Real chromatic spike streams and reconstructions", + "Fig. 6: Illustration of our real-world collected chromatic spike streams, synthetic data derived from conventional cameras, and corresponding reconstructions from our proposed method." + ], + "image_footnote": [], + "bbox": [ + 506, + 147, + 771, + 202 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/fde4ca2317f5a8eb2fd541bc18684e12c3e03bbb526393280fab1a6891a1fc14.jpg", + "table_caption": [ + "Table 1: Quantitative evaluation of the image reconstruction quality. The scores are averaged across our synthetic dataset. $\\uparrow$ (↓) means higher (lower) is better." + ], + "table_footnote": [], + "table_body": "
MethodPSNR ↑SSIM ↑LPIPS ↓HDR-VDP3 ↑HDR-VQM ↓
TFI [46]19.2150.2460.7164.6450.7897
TFP(W=10) [46]23.2530.3310.6095.7780.7799
TFP(W=100) [46]28.7910.7200.2787.5450.4994
TFSTP [43]20.9450.2530.6704.5870.7843
MS23 [21]28.4100.6350.3597.1800.5765
Ours30.7800.8830.2087.9910.3023
", + "bbox": [ + 388, + 308, + 782, + 404 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5 Experiment", + "text_level": 1, + "bbox": [ + 215, + 454, + 366, + 470 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.1 Data collection", + "text_level": 1, + "bbox": [ + 215, + 505, + 385, + 518 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Synthetic Data. In line with the methodologies of monochromatic spike simulators (e.g., Zhao et al. [41], SpikeCV [42]), we develop a chromatic spike simulator, following the same Bayer pattern as the real-world chromatic spike camera. We leverage the frames in the GoPro dataset [26] and their corresponding synthetic spike streams to train a supervised learning model and compare it with our self-supervised learning method, which addresses the superiority of our strategy. We interpolate the HDR videos collected by Chang et al. [3] as the evaluation data for quantitative evaluation in our experiments. Some examples of these synthetic scenes are shown in Fig. 6 (a). For the detailed design of our chromatic spike simulator, please refer to our supplementary material.", + "bbox": [ + 212, + 547, + 787, + 700 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Real-world data. To accomplish real-data-driven self-supervised learning and validate the effectiveness of our proposed method, we capture a collection of real-world chromatic spikes using a spike camera equipped with the Bayer pattern. Our dataset comprises over 100 sets of mosaicked chromatic spikes captured from 30 diverse scenes, characterized by varying degrees of high-speed motion and high dynamic range, as depicted in Fig. 6 (b). The real-world data inherently contain the noise characteristics of a real spike camera prototype, posing significant challenges for denoising and demosaicking algorithms. Our real-world dataset is available at https://huggingface.co/datasets/YOUSIKI/chromatic-spikes.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "S. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.2 Quantitative evaluation on synthetic data", + "text_level": 1, + "bbox": [ + 215, + 146, + 606, + 162 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We compare our method with existing spike-based video reconstruction methods, i.e., TFI [46], TFP [46], and TFSTP [43]. We acknowledge that it may not be entirely fair to compare our method with these approaches, as they are designed for monochromatic spikes. However, they can serve as baselines to demonstrate the effectiveness of our denoising and progressive warping modules. Since the data format of the spike camera and the single photon avalanche diodes (SPAD) are both single-bit signals, the color video reconstruction method can be adapted to chromatic spikes, and MS23 [21] is chosen for comparison7. The quantitative evaluation is performed on synthetic data. We select PSNR, SSIM [32], LPIPS [37], HDR-VDP3 [25], and HDR-VQM [27] as the metrics. The quantitative evaluation is shown in Tab. 1. We can see that the proposed method demonstrated superior performance over the competing techniques across all metrics.", + "bbox": [ + 212, + 171, + 787, + 367 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Qualitative evaluation on real data", + "text_level": 1, + "bbox": [ + 215, + 390, + 549, + 404 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To underscore the advantages of our method, we conduct a qualitative evaluation on real and synthetic chromatic spike streams. Fig. 7 presents a quadruple comparison. Our approach not only reconstructs motion blur free frames and recovers color information but also achieves the best balance between noise suppression and detail preservation across diverse scenes. In Fig. 7 (a), we create a colorful and dynamic environment by placing a high-speed electric fan between our chromatic spike camera and a set of toys. The reconstruction from TFP $(W = 100)$ exhibits noticeable motion blur in the blue bounding box. Conversely, TFI, TFP $(W = 10)$ , TFSTP and MS23 [21] suffer from significant noise. On the contrary, our method yields a clean outcome. Fig. 7 (b) showcases flowing water in glass and a rotating toy duck. As indicated by red bounding boxes, TFI, TFP $(W = 10)$ , TFSTP, and MS23 [21] contain different levels of noise. And TFP $(W = 100)$ suffers from heavy motion blur in blue bounding box. Meanwhile, our proposed method produces clean reconstruction without blur. Fig. 7 (c, d) further demonstrates the denoising advantage of our method on synthetic data.", + "bbox": [ + 212, + 415, + 787, + 643 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 665, + 359, + 681 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this paper, we propose a novel real-data-driven approach for high-speed color HDR video reconstruction from mosaicked chromatic spikes. Through the analysis on the noise of chromatic spikes, we find the noise in short-term frames can be regarded as a zero-mean distribution. According to this characteristic, we design a self-supervised denoising module that can be trained on real-world chromatic spikes. We further propose a progressive warping module that enables the reconstruction of HDR frames from a pseudo long exposure. Experiments on", + "bbox": [ + 212, + 696, + 787, + 803 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes", + "bbox": [ + 230, + 113, + 727, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_footnote", + "text": "7 As the source code for MS23 [21] is not publicly accessible, we conduct reimplementation and modification for the evaluation.", + "bbox": [ + 217, + 810, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/bcb7d0f7f9c2c4239cf6e2b6269c113dd2928189bf01bf46dd7c7c8bfb4ebb64.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TFITFP (W = 10)TFP (W = 100)TFSTPMS23Ours
(a)
(b)
(c)
(d)
", + "bbox": [ + 217, + 143, + 785, + 573 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Fig. 7: Visual quality comparison of real (a, b) and synthetic (c, d) data between the proposed method and compared methods. Please zoom-in electronic versions for better details, and watch the videos in the supplementary material.", + "bbox": [ + 215, + 585, + 787, + 628 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "real-world data demonstrate the superiority of our approach for reconstructing 2000FPS color HDR videos.", + "bbox": [ + 212, + 657, + 785, + 686 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations and future work. The Bayer-pattern CFA introduces chromatic information to the spikes but also reduces the intensity of incident light. As a result, compared to monochromatic spikes, the ability to preserve textures in low-light regions is relatively compromised in high-speed conditions. In future research, exploring different types of CFAs to enhance low-light performance would be worth investigating. And the excessive focus on denoising can result in over-smooth downsides, which may lead to the loss of fine details. Future work could involve the development of a more sophisticated denoising module to balance noise suppression and detail preservation.", + "bbox": [ + 212, + 688, + 787, + 825 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "S. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 145, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by National Science and Technology Major Project (Grant No. 2021ZD0109803) and National Natural Science Foundation of China (Grant No. 62088102, 62136001, 62301009). Bin Fan was supported by National Postdoctoral Program for Innovative Talents of China (Grant No.BX20230013).", + "bbox": [ + 215, + 176, + 784, + 236 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 258, + 321, + 273 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Batson, J., Royer, L.: Noise2Self: Blind denoising by self-supervision. In: ICML. pp. 524-533 (2019)", + "2. Buades, A., Coll, B., Morel, J.M., Sbert, C.: Self-similarity driven color demosaicking. IEEE TIP 18(6), 1192-1202 (2009)", + "3. Chang, Y., Zhou, C., Hong, Y., Hu, L., Xu, C., Huang, T., Shi, B.: 1000 FPS HDR video with a spike-RGB hybrid camera. In: CVPR. pp. 22180-22190 (2023)", + "4. Chen, C., Mo, J., Hou, J., Wu, H., Liao, L., Sun, W., Yan, Q., Lin, W.: TOPIQ: A top-down approach from semantics to distortions for image quality assessment. IEEE Transactions on Image Processing (2024)", + "5. Chen, S., Duan, C., Yu, Z., Xiong, R., Huang, T.: Self-supervised mutual learning for dynamic scene reconstruction of spiking camera. In: IJCAI. p. 2859-2866 (2022)", + "6. Chen, S., Yu, Z., Huang, T.: Self-supervised joint dynamic scene reconstruction and optical flow estimation for spiking camera. In: AAAI. vol. 37, pp. 350-358 (2023)", + "7. Condat, L., Mosaddegh, S.: Joint demosaicking and denoising by total variation minimization. In: ICIP. pp. 2781-2784 (2012)", + "8. Fossum, E.R., Ma, J., Masoodian, S., Anzagira, L., Zizza, R.: The quanta image sensor: Every photon counts. Sensors 16(8), 1260 (2016)", + "9. Gharbi, M., Chaurasia, G., Paris, S., Durand, F.: Deep joint demosaicking and denoising. ACM TOG 35(6), 1-12 (2016)", + "0. Gnanasambandam, A., Chan, S.H.: HDR imaging with quanta image sensors: Theoretical limits and optimal reconstruction. IEEE TCI 6, 1571-1585 (2020)", + "11. Han, J., Zhou, C., Duan, P., Tang, Y., Xu, C., Xu, C., Huang, T., Shi, B.: Neuromorphic camera guided high dynamic range imaging. In: CVPR. pp. 1730-1739 (2020)", + "12. Hasinoff, S.W., Sharlet, D., Geiss, R., Adams, A., Barron, J.T., Kainz, F., Chen, J., Levoy, M.: Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM TOG 35(6), 1-12 (2016)", + "13. Heide, F., Rouf, M., Hullin, M.B., Labitzke, B., Heidrich, W., Kolb, A.: High-quality computational imaging through simple lenses. ACM TOG 32(5), 1-14 (2013)", + "14. Huang, T., Zheng, Y., Yu, Z., Chen, R., Li, Y., Xiong, R., Ma, L., Zhao, J., Dong, S., Zhu, L., Li, J., Jia, S., Fu, Y., Shi, B., Wu, S., Tian, Y.: $1000 \\times$ faster camera and machine vision with ordinary devices. Engineering 25, 110-119 (2022)", + "15. Kiku, D., Monno, Y., Tanaka, M., Okutomi, M.: Beyond color difference: Residual interpolation for color image demosaicking. IEEE TIP 25(3), 1288-1300 (2016)", + "16. Kokkinos, F., Lefkimmiatis, S.: Deep image demosaicking using a cascade of convolutional residual denoising networks. In: ECCV. pp. 303-319 (2018)", + "17. Lichtsteiner, P., Posch, C., Delbruck, T.: A $128 \\times 128$ 120 db $15\\mu s$ latency asynchronous temporal contrast vision sensor. IEEE JSSC 43(2), 566-576 (2008)" + ], + "bbox": [ + 218, + 287, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes", + "bbox": [ + 230, + 114, + 727, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Liu, L., Jia, X., Liu, J., Tian, Q.: Joint demosaicing and denoising with self guidance. In: CVPR. pp. 2240-2249 (2020)", + "19. Liu, Y., Gutierrez-Barragan, F., Ingle, A., Gupta, M., Velten, A.: Single-photon camera guided extreme dynamic range imaging. In: WACV. pp. 1575-1585 (2022)", + "20. Ma, S., Gupta, S., Ulku, A.C., Bruschini, C., Charbon, E., Gupta, M.: Quanta burst photography. ACM TOG 39(4), 79-1 (2020)", + "21. Ma, S., Sundar, V., Mos, P., Bruschini, C., Charbon, E., Gupta, M.: Seeing photons in color. ACM TOG 42(4), 1-16 (2023)", + "22. Mairal, J., Bach, F., Ponce, J., Sapiro, G., Zisserman, A.: Non-local sparse models for image restoration. In: ICCV. pp. 2272-2279 (2009)", + "23. Mairal, J., Elad, M., Sapiro, G.: Sparse representation for color image restoration. IEEE TIP 17(1), 53-69 (2007)", + "24. Malvar, H.S., He, L.w., Cutler, R.: High-quality linear interpolation for demosaicing of bayer-patterned color images. In: ICASSP. vol. 3, pp. iii-485 (2004)", + "25. Mantiuk, R.K., Hammou, D., Hanji, P.: HDR-VDP-3: A multi-metric for predicting image differences, quality and contrast distortions in high dynamic range and regular content. arXiv preprint arXiv:2304.13625 (2023)", + "26. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: CVPR. pp. 3883-3891 (2017)", + "27. Narwaria, M., Da Silva, M.P., Le Callet, P.: HDR-VQM: An objective quality measure for high dynamic range video. Signal Processing: Image Communication 35, 46-60 (2015)", + "28. Pekkucuksen, I., Altunbasak, Y.: Gradient based threshold free color filter array interpolation. In: ICIP. pp. 137-140 (2010)", + "29. Tan, D.S., Chen, W.Y., Hua, K.L.: Deepdemosaicking: Adaptive image demosaicking via multiple deep fully convolutional networks. IEEE TIP 27(5), 2408-2419 (2018)", + "30. Tan, R., Zhang, K., Zuo, W., Zhang, L.: Color image demosaicking via deep residual learning. In: ICME. pp. 793-798 (2017)", + "31. Wang, Z., Liu, J., Li, G., Han, H.: Blind2unblind: Self-supervised image denoising with visible blind spots. In: CVPR. pp. 2027-2036 (2022)", + "32. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: From error visibility to structural similarity. IEEE TIP 13(4), 600-612 (2004)", + "33. Xia, L., Zhao, J., Xiong, R., Huang, T.: SVFI: Spiking-based video frame interpolation for high-speed motion. In: AAAI. vol. 37, pp. 2910-2918 (2023)", + "34. Xing, W., Egiazarian, K.: End-to-end learning for joint image demosaicing, denoising and super-resolution. In: CVPR. pp. 3507-3516 (2021)", + "35. Yang, Y., Han, J., Liang, J., Sato, I., Shi, B.: Learning event guided high dynamic range video reconstruction. In: CVPR. pp. 13924-13934 (2023)", + "36. Yu, G., Sapiro, G., Mallat, S.: Solving inverse problems with piecewise linear estimators: From gaussian mixture models to structured sparsity. IEEE TIP 21(5), 2481-2499 (2011)", + "37. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR. pp. 586-595 (2018)", + "38. Zhang, Z., Zheng, H., Hong, R., Xu, M., Yan, S., Wang, M.: Deep color consistent network for low-light image enhancement. In: CVPR. pp. 1899-1908 (2022)", + "39. Zhao, J., Xiong, R., Liu, H., Zhang, J., Huang, T.: Spk2ImgNet: Learning to reconstruct dynamic scene from continuous spike stream. In: CVPR. pp. 11996-12005 (2021)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "S. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "40. Zhao, J., Xiong, R., Xie, J., Shi, B., Yu, Z., Gao, W., Huang, T.: Reconstructing clear image for high-speed motion scene with a retina-inspired spike camera. IEEE TCI 8, 12-27 (2021)", + "41. Zhao, J., Zhang, S., Ma, L., Yu, Z., Huang, T.: SpikingSIM: A bio-inspired spiking simulator. In: IEEE ISCAS. pp. 3003-3007 (2022)", + "42. Zheng, Y., Zhang, J., Zhao, R., Ding, J., Chen, S., Xiong, R., Yu, Z., Huang, T.: SpikeCV: Open a continuous computer vision era. arXiv preprint arXiv:2303.11684 (2023)", + "43. Zheng, Y., Zheng, L., Yu, Z., Huang, T., Wang, S.: Capture the moment: High-speed imaging with spiking cameras through short-term plasticity. IEEE TPAMI 45(7), 8127-8142 (2023)", + "44. Zheng, Y., Zheng, L., Yu, Z., Shi, B., Tian, Y., Huang, T.: High-speed image reconstruction through short-term plasticity for spiking cameras. In: CVPR. pp. 6358-6367 (2021)", + "45. Zhu, L., Dong, S., Huang, T., Tian, Y.: A retina-inspired sampling method for visual texture reconstruction. In: ICME. pp. 1432-1437 (2019)", + "46. Zhu, L., Dong, S., Li, J., Huang, T., Tian, Y.: Retina-like visual image reconstruction via spiking neural model. In: CVPR. pp. 1438-1446 (2020)" + ], + "bbox": [ + 215, + 146, + 787, + 397 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes", + "bbox": [ + 230, + 113, + 727, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes/6731bc50-147c-4c01-a7c6-498656963ba3_model.json b/2024/Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes/6731bc50-147c-4c01-a7c6-498656963ba3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2ddfe99434d2051d59f5e4c5b76395af6a5c0967 --- /dev/null +++ b/2024/Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes/6731bc50-147c-4c01-a7c6-498656963ba3_model.json @@ -0,0 +1,2181 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.275, + 0.141, + 0.73, + 0.187 + ], + "angle": 0, + "content": "Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes" + }, + { + "type": "text", + "bbox": [ + 0.275, + 0.211, + 0.728, + 0.244 + ], + "angle": 0, + "content": "Siqi Yang\\(^{1,\\#}\\), Zhaojun Huang\\(^{2,3,\\#}\\), Yakun Chang\\(^{5,6}\\), Bin Fan\\(^{4}\\), Zhaofei Yu\\(^{1}\\), and Boxin Shi\\(^{2,3,1,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.341, + 0.255, + 0.662, + 0.269 + ], + "angle": 0, + "content": "\\(^{1}\\) Institute for Artificial Intelligence, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.268, + 0.778, + 0.279 + ], + "angle": 0, + "content": "2 State Key Lab of Multimedia Info. Processing, School of Computer Science, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.279, + 0.778, + 0.289 + ], + "angle": 0, + "content": "3 Nat'l Eng. Research Ctr. of Visual Technology, School of Computer Science, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.289, + 0.778, + 0.299 + ], + "angle": 0, + "content": "\\(^{4}\\) Nat'l Key Lab of General AI, School of Intelligence Science and Technology, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.323, + 0.298, + 0.682, + 0.309 + ], + "angle": 0, + "content": "\\(^{5}\\) Institute of Information Science, Beijing Jiaotong University" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.308, + 0.783, + 0.329 + ], + "angle": 0, + "content": "\\(^{6}\\) Visual Intelligence +X International Cooperation Joint Laboratory of the Ministry of Education {yousiki, huangzhaojun, binfan, yuzf12, shiboxin}@pku.edu.cn, ykchang@bjtu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.366, + 0.74, + 0.657 + ], + "angle": 0, + "content": "Abstract. The spike camera continuously records scene radiance with high-speed, high dynamic range, and low data redundancy properties, as a promising replacement for frame-based high-speed cameras. Previous methods for reconstructing color videos from monochromatic spikes are constrained in capturing full-temporal color information due to their reliance on compensating colors from low-speed RGB frames. Applying a Bayer-pattern color filter array to the spike sensor yields mosaicked chromatic spikes, which complicates noise distribution in high-speed conditions. By validating that the noise of short-term frames follows a zero-mean distribution, we leverage this hypothesis to develop a self-supervised denoising module trained exclusively on real-world data. Although noise is reduced in short-term frames, the long-term accumulation of incident photons is still necessary to construct HDR frames. Therefore, we introduce a progressive warping module to generate pseudo long-term exposure frames. This approach effectively mitigates motion blur artifacts in high-speed conditions. Integrating these modules forms a real-data-driven reconstruction method for mosaicked chromatic spikes. Extensive experiments conducted on both synthetic and real-world data demonstrate that our approach is effective in reconstructing 2000FPS color HDR videos with significantly reduced noise and motion blur compared to existing methods." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.671, + 0.741, + 0.699 + ], + "angle": 0, + "content": "Keywords: Spike camera \\(\\cdot\\) High-speed video \\(\\cdot\\) High dynamic range video \\(\\cdot\\) Color demosaicking \\(\\cdot\\) Self-supervised learning" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.725, + 0.376, + 0.74 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.756, + 0.788, + 0.817 + ], + "angle": 0, + "content": "As a new class of image sensors inspired by the fovea of retina, the spike camera [3, 14, 46] offers attractive characteristics such as high temporal resolution (20000Hz), high dynamic range (HDR, \\(>90\\mathrm{dB}\\)), and low data redundancy [14]. In comparison with conventional digital cameras that capture discrete frames," + }, + { + "type": "page_footnote", + "bbox": [ + 0.216, + 0.825, + 0.537, + 0.84 + ], + "angle": 0, + "content": "# Equal contributions. * Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "S. Yang et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.146, + 0.784, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.417, + 0.334, + 0.609, + 0.347 + ], + "angle": 0, + "content": "(b) High-speed color HDR video" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.36, + 0.788, + 0.443 + ], + "angle": 0, + "content": "Fig. 1: High-speed color HDR video reconstruction from mosaicked chromatic spikes. (a) Comparison of different chromatic spike reconstruction (recon.) methodologies, including the adaptation of bio-inspired [45] and supervised learning [39] monochromatic methods to chromatic spikes and our proposed method. (b) A high-quality color HDR video with a frame rate of 2000FPS reconstructed by our method. The corresponding video is available in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.788, + 0.597 + ], + "angle": 0, + "content": "spike cameras continuously accumulate the photons and trigger a single-bit spike of 1 when the detected photons reach a predefined threshold [14]. High-speed readout of spikes provides the spike camera with the potential for photon counting, and the high-dynamic-range ambient light can be estimated through continuous accumulation of spikes in the temporal domain. Therefore, spike cameras exhibit significant advantages in handling low-level vision tasks in high-speed conditions, such as video frame interpolation [33], HDR imaging [11] and high-frame-rate HDR video reconstruction [3]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.789, + 0.844 + ], + "angle": 0, + "content": "With the higher demand for visual quality in images and videos, there is a strong desire that images or videos reconstructed from spike cameras can reproduce the real world to the greatest extent, with realistic color appearance and extended dynamic range at high frame rate. However, existing approaches [39, 40, 43, 44, 46] mainly focus on image or video reconstruction with monochromatic spikes, and cannot yet recover a high-speed color HDR video. To introduce color, one alternative practice [11] is to build a hybrid spike-RGB camera system, which exploits the colors captured by the conventional RGB camera to colorize the image reconstructed from spikes. Unfortunately, the color information provided by conventional RGB cameras often suffers from underexposure and overexposure, which inevitably lead to color deviations from the real world. Another strategy [3] to alleviate this problem is to replace the conventional RGB camera with an alternate exposure RGB camera to obtain HDR color, thereby enabling high-speed HDR color video reconstruction. However, the above methods have limitations in high-speed scenes, as the RGB camera has a low frame rate (e.g., 60FPS), which results in the loss of colors during the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.231, + 0.114, + 0.728, + 0.13 + ], + "angle": 0, + "content": "Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "intervals between frames. In addition, the corresponding hybrid camera system relies on complex temporal synchronization and optical alignment, and the beam splitter with large space footprints would be detrimental to the construction of compact devices." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.212, + 0.789, + 0.468 + ], + "angle": 0, + "content": "Reconstructing high-speed color HDR videos based only on a single spike camera not only facilitates the acquisition of full-temporal color information but also offers the advantage of lightweight equipment. To this end, a tractable solution is to assemble a Bayer-pattern color filter array (CFA) in the spike camera, thereby producing mosaicked chromatic spikes. Consequently, a pressing challenge that arises is how to reconstruct high-speed color HDR videos from this new data representation. Nevertheless, directly adapting existing monochromatic spike reconstruction methods [39,40,43,44] to mosaicked chromatic spikes suffers from many issues, as shown in Fig. 1 (a). On the one hand, methods based on biology-inspired models [40,43,44] cannot take advantage of deep neural networks and are susceptible to varying degrees of noise contamination. On the other hand, supervised learning methods (e.g., [39]) trained using synthetic datasets have limited generalization ability when transferring to real-world data, because of the imperfection of spike noise modeling and simulation. These issues hinder the development of high-speed color HDR video reconstruction methods based on chromatic spike cameras and greatly limit the effective deployment of existing methods in real scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.789, + 0.7 + ], + "angle": 0, + "content": "In this paper, we propose a real-data-driven mosaicked chromatic spike reconstruction method to recover high-speed color HDR videos, trained solely on readily accessible unlabelled spike streams. The mathematical analysis behind this requires no specific spike noise modeling or calibration, relying only on insights into the relationship between short-term and long-term frames accumulated from chromatic spikes. We design a deep learning framework that allows training directly on real-world data, naturally avoiding the influence of imperfect spike simulation and generalization gap. Specifically, we first propose a self-supervised denoising module to suppress the noise in short-term accumulation of chromatic spikes, which is trained on real-world spike streams without any annotation. Next, considering the limited dynamic range of short-term accumulated frames, we further present a progressive warping module to simulate a pseudolong exposure, which is effective in reconstructing HDR video frames without motion blur. Finally, we integrate these two modules to achieve higher-quality color HDR video reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.766 + ], + "angle": 0, + "content": "In summary, our approach is able to achieve 2000FPS color HDR video reconstruction from mosaicked chromatic spikes with a single device in a real-data-driven manner, as illustrated in Fig. 1 (b), because of the technical contributions from these two complementary aspects:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.795, + 0.789, + 0.84 + ], + "angle": 0, + "content": "- We design a self-supervised denoising module to eliminate the noise from short-term accumulation (with high-speed motion cues) without relying on chromatic spike noise modeling and simulation." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "S. Yang et al." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "- We propose a progressive warping module with pseudo long-term exposure (with expanded dynamic range) to reconstruct HDR frames without introducing motion blur in dynamic scenes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.217, + 0.388, + 0.234 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.251, + 0.789, + 0.538 + ], + "angle": 0, + "content": "Video reconstruction with monochromatic spikes. To better satisfy human visual perception and facilitate direct processing by machines, reconstructing the corresponding video frame from spikes has been widely studied. Prevalent image reconstruction methods typically employ the temporal statistical properties of spike cameras [43-46]. Zhu et al. [45] explored the spike generation principle and proposed estimating the firing frequency or firing interval of each pixel to reconstruct dynamic scenes, which are known as texture from playback (TFP) and texture from inter-spike-intervals (TFI), respectively. Recently, another line of work has enhanced firing frequency estimates by mimicking human physiological mechanisms, such as retina-like visual imaging [46] and short-term plasticity [43, 44]. Although these methods bring better interpretability, they require careful window selection to balance noise and motion blur, and the reconstruction effect is not yet satisfactory. The powerful representation capabilities of deep learning also inject vitality into spike-based image reconstruction. Spk2ImgNet [39] is the first spike-to-image deep network architecture based on deformable convolution, achieving impressive results. Further, several self-supervised methods [5,6] have also been developed to reduce the dependence on synthetic datasets. However, these methods are tailored for monochromatic spikes and thus cannot be directly applied to mosaicked chromatic spikes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.789, + 0.704 + ], + "angle": 0, + "content": "Demosaicking and denoising with raw images. The purpose of raw image demosaicking is to recover a full-color image from a sub-sampled mosaicked image that contains potential noise. To deal with the ill-posedness of image demosaicking, various image priors have been introduced to regularize the recovery process, such as heuristics [2, 24], sparsity [23, 36], total variation [7, 13], integral gradient [28], self-similarity [22], and residual interpolation [15]. In addition, researchers designed deep neural networks [16, 29, 30] to automatically learn the prior knowledge required for image demosaicking. Moreover, considering the coupling between demosaicking and denoising, some deep learning-based joint demosaicking and denoising network models [9, 18, 34, 38] have also been developed to efficiently remove mosaicks from noisy raw images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.841 + ], + "angle": 0, + "content": "HDR video with non-conventional sensors. In recent years, a series of nonconventional sensors, such as event cameras [17], spike cameras [14], and quanta image sensors (QIS) [8], have been invented, empowering the ability to capture HDR information in high-speed motion scenes. QIS-based HDR video reconstruction methods [10, 19, 20] were developed to accomplish perceptual HDR illumination. Additionally, the intensity image reconstructed by event cameras [35] or spike cameras [11] can serve as a proxy for enhancing details in low dynamic range (LDR) images. More recently, a hybrid camera system consisting of a spike camera and an alternating-exposure RGB camera is constructed in [3] for \\(1000 \\times\\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.231, + 0.114, + 0.728, + 0.129 + ], + "angle": 0, + "content": "Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.149, + 0.777, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.289, + 0.788, + 0.361 + ], + "angle": 0, + "content": "Fig. 2: Chromatic spike accumulation and firing mechanism. We put a Bayer pattern CFA in the spike camera to filter incident light before the photons arrive at the sensor. Each individual pixel corresponds to one of red, green, and blue channels, accumulates the filtered photons, and triggers a spike signal, then resets when it surpasses the threshold \\( E_{\\mathrm{th}} \\). Reading out the chromatic spikes forms the chromatic spike planes." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.387, + 0.784, + 0.418 + ], + "angle": 0, + "content": "frame rate HDR video reconstruction. Nevertheless, these methods also fall short when dealing with mosaicked chromatic spikes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.438, + 0.382, + 0.455 + ], + "angle": 0, + "content": "3 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.468, + 0.784, + 0.528 + ], + "angle": 0, + "content": "We initially present a concise review of the operational principles of spike cameras, encompassing both monochromatic spike cameras and those equipped with Color Filter Array (CFA). Subsequently, we discuss the baseline reconstruction techniques for monochromatic spikes." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.529, + 0.785, + 0.589 + ], + "angle": 0, + "content": "Monochromatic spikes. The spike camera persistently accumulates electrons produced by incoming photons. When the electrons accumulated at a pixel surpass the predefined threshold \\( E_{\\mathrm{th}} \\), it triggers a spike signal, immediately resetting the accumulator to initiate a new cycle of accumulation:" + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.599, + 0.787, + 0.631 + ], + "angle": 0, + "content": "\\[\nE (\\mathbf {x}, t) = \\int \\alpha L (\\mathbf {x}, t) d t \\mod E _ {\\mathrm {t h}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.635, + 0.784, + 0.679 + ], + "angle": 0, + "content": "where \\(\\alpha\\) denotes the conversion ratio and \\(L(\\mathbf{x},t)\\) denotes the incoming light intensity of pixel \\(\\mathbf{x}\\) at time \\(t\\). Consequently, the spike signal \\(S(\\mathbf{x},t)\\) can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.69, + 0.787, + 0.731 + ], + "angle": 0, + "content": "\\[\nS (\\mathbf {x}, t) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} E (\\mathbf {x}, t) \\geq E _ {\\mathrm {t h}}, \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.735, + 0.784, + 0.809 + ], + "angle": 0, + "content": "As all the pixels of a spike sensor are arranged in an array, reading out spikes from the pixel array at each time stamp \\( t \\) yields a spike plane of dimensions \\( h \\times w \\), where \\( h \\) and \\( w \\) are the height and width of the spike data, respectively. Given the readout frequency of 20000Hz, the spike camera [14] allows for capturing spike planes at a frame rate of 20000FPS." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Chromatic spikes. To incorporate color information while capturing HDR scenes with high-speed motion at the same temporal resolution, we can put a" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "S. Yang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "CFA in front of the sensor of the spike camera. In a similar way as conventional cameras, the CFA applied to spike sensors presents a Bayer pattern. Each pixel of the spike camera sensor with CFA corresponds to only one channel among red, green, and blue, while the firing mechanism of spikes remains unchanged. As illustrated in Fig. 2, under the influence of the CFA, a spike signal is generated when the cumulative electrons \\( E_{c}(\\mathbf{x},t) \\) produced by the incident photons of the corresponding spectral band meets the threshold \\( E_{\\mathrm{th}} \\). The chromatic spike signal \\( S_{c}(\\mathbf{x},t) \\) can be recorded as:" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.28, + 0.786, + 0.319 + ], + "angle": 0, + "content": "\\[\nS _ {c} (\\mathbf {x}, t) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} E _ {c} (\\mathbf {x}, t) \\geq E _ {\\mathrm {t h}}, \\\\ 0, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.323, + 0.785, + 0.367 + ], + "angle": 0, + "content": "where \\( c \\in \\{R, G, B\\} \\). Consequently, similar to the firing process of monochromatic spikes, we can obtain a mosaicked chromatic spike plane at each time stamp \\( t \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.369, + 0.787, + 0.413 + ], + "angle": 0, + "content": "Baseline for chromatic frame reconstruction. Preliminary chromatic frame reconstruction can be achieved similarly to [45] by accumulating the spike signals within a defined temporal window:" + }, + { + "type": "equation", + "bbox": [ + 0.387, + 0.426, + 0.786, + 0.47 + ], + "angle": 0, + "content": "\\[\nI _ {c} (\\mathbf {x}, t) = \\frac {M _ {c}}{W} \\cdot \\sum_ {\\tau = t - W / 2} ^ {t + W / 2} S _ {c} (\\mathbf {x}, \\tau), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.788, + 0.64 + ], + "angle": 0, + "content": "where \\( W \\) is the size of the accumulation temporal window, \\( M_{c} \\) represents the maximum dynamic range that can be perceived in channel \\( c \\). Directly applying Eq. (4) to chromatic spikes as a baseline reconstruction algorithm can only obtain mosaicked chromatic frames, which is similar to the RAW image of conventional frame-based sensors. Demosaicking the chromatic frame is challenging since there is a trade-off between noise and superposition of colors. A short-term temporal window \\( W \\) results in severe noise, whereas a long-term \\( W \\) leads to the spatial superposition of varying color information. Furthermore, both noise distribution of chromatic spikes and object motions are more complex, making the direct application of off-the-shelf demosaicking methods insufficient for recovering high-quality color HDR videos." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.66, + 0.667, + 0.679 + ], + "angle": 0, + "content": "4 Real-data-driven chromatic spikes to video" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Overview. As shown in Fig. 3, we propose a progressive mechanism for high-speed color HDR video reconstruction. The pipeline starts with the accumulation of mosaicked chromatic spikes to reconstruct short-term frames. While these short-term frames are free of motion blur, they suffer from significant noise contamination. Consequently, as shown in Fig. 3 (a), we propose the chromatic spike denoising module (Sec. 4.1) to suppress the noise. The key insight of the denoising module is that a sufficiently long-term frame (e.g., \\( W = 10000 \\), corresponding to 0.5s) of chromatic spikes can produce promising reconstructions in static (or low-speed) scenes, effectively representing the expectations of short-term frames. Given the limited incident photons in short-term frames, which" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.231, + 0.114, + 0.728, + 0.129 + ], + "angle": 0, + "content": "Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.143, + 0.788, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.56 + ], + "angle": 0, + "content": "Fig. 3: Illustration of the overall pipeline. For each sequence of mosaicked chromatic spikes, we split them into four channels according to the RGGB pattern. The four-channel spikes are fed to our progressive warping module, which aligns multiple short-term frames to achieve \"pseudo-long exposure\". (a) Our chromatic spike denoising module utilizes a self-supervised training strategy, which is pretrained on real-world data. (b) Our progressive warping module consists of \\( L \\) steps of frame warping operations, iteratively achieving high-quality HDR frames. At each step of the progressive warping module, we leverage the pretrained chromatic spikes denoising module to facilitate optical flow estimation. \\( R_{L} \\) is the final output color HDR video with 2000FPS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.59, + 0.788, + 0.727 + ], + "angle": 0, + "content": "lack the HDR characteristic, extending the accumulation time is a feasible strategy for HDR reconstruction. As shown in Fig. 3 (b), we design the progressive warping module in Sec. 4.2 to establish a pseudo long exposure. At each step of the warping module, chromatic spikes frames are warped according to the estimated optical flow, providing pseudo long-term frames with less noise and higher dynamic range. To leverage a wide range of existing conventional RGB optical flow estimation approaches, we integrate the denoising module into each iteration step of the warping module. After \\( L \\) times of iteration, the progressive warping module produces an HDR color video with a frame rate of 2000FPS." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.752, + 0.507, + 0.768 + ], + "angle": 0, + "content": "4.1 Denoising of chromatic spikes" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Noise distribution of chromatic spikes noise. The noise in monochromatic spikes is primarily categorized into two types: diffuse noise and inherent noise. For the mosaicked chromatic spikes, although the incorporation of a CFA slightly alters the noise distribution for pixels under different filters, pixels corresponding" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "S. Yang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "to the same color exhibit similar noise distribution patterns as monochromatic spikes, except for its expectation and variance. Given the chromatic spikes \\( S_{c} \\), for each \\( c \\in \\{R,G,B\\} \\), we use \\( Q_{d} \\) to denote the number of diffuse noise photons arriving at a pixel in a time period of \\( t \\), and \\( \\gamma \\) to denote the likelihood of \\( Q_{d} \\) diffuse noise photons. The probability of diffuse noise can be modeled as the Poisson distribution [41]:" + }, + { + "type": "equation", + "bbox": [ + 0.387, + 0.251, + 0.786, + 0.285 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} \\left(Q _ {d} | t, \\gamma\\right) = \\frac {\\left(\\gamma t\\right) ^ {Q _ {d}}}{Q _ {d} !} \\exp (- \\gamma W). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.29, + 0.772, + 0.306 + ], + "angle": 0, + "content": "Hence, the expected number of diffuse noise spikes in temporal window \\( W \\) is" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.323, + 0.786, + 0.34 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} \\left[ N _ {d} \\right] = (\\alpha \\gamma W) / E _ {\\mathrm {t h}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.347, + 0.789, + 0.395 + ], + "angle": 0, + "content": "For inherent noise, which triggers spikes in the absence of light, we use \\( T_{i} \\) to indicate the interval between the noise spikes. Then, the interval between such noise spikes can be represented by a Gaussian model [41]:" + }, + { + "type": "equation", + "bbox": [ + 0.351, + 0.406, + 0.787, + 0.438 + ], + "angle": 0, + "content": "\\[\n\\mathcal {N} (T _ {i} | \\mu , \\sigma) = \\frac {1}{\\sigma \\sqrt {2 \\pi}} \\cdot \\exp \\left(- \\frac {1}{2} \\cdot \\left(\\frac {T _ {i} - \\mu}{\\sigma}\\right) ^ {2}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.443, + 0.788, + 0.473 + ], + "angle": 0, + "content": "We approximate the expectation of the number of inherent noise spikes in the temporal window as" + }, + { + "type": "equation", + "bbox": [ + 0.293, + 0.488, + 0.787, + 0.521 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} [ N _ {i} ] = \\int_ {1} ^ {+ \\infty} \\frac {W}{T _ {i}} \\cdot \\frac {1}{\\sigma \\sqrt {2 \\pi}} \\cdot \\exp (- \\frac {1}{2} \\cdot (\\frac {T _ {i} - \\mu}{\\sigma}) ^ {2}) d T _ {i} = C _ {i} W, (8)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.527, + 0.788, + 0.573 + ], + "angle": 0, + "content": "where \\( C_i \\) denotes the integral after factoring out \\( W \\) from the integral expression. Combining Eq. (6) and Eq. (8), the expectation of the fraction of noise spikes can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.334, + 0.59, + 0.787, + 0.607 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} [ N ] / W = \\left(\\mathbb {E} \\left[ N _ {d} \\right] + \\mathbb {E} \\left[ N _ {i} \\right]\\right) / W = \\alpha \\gamma / E _ {\\mathrm {t h}} + C _ {i}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Instead of the actual incoming light intensity, we estimate the expectation of short-term accumulations \\(\\mathbb{E}[I_c(\\mathbf{x},t)] = L(\\mathbf{x},t) + \\mathbb{E}[N] / W\\). Thus, the noise between short-term frames and their expectations naturally follows a zero-mean distribution. We demonstrate that the expected fraction of noise spikes is relatively small, especially with sufficient light illumination, and doesn't disturb the visual semantic. As shown in Fig. 4 (a), the noise ratio \\((I_{c}(\\mathbf{x},t) / \\mathbb{E}[I_{c}(\\mathbf{x},t)] - 1)\\) in the accumulated images \\((I_{c}(\\mathbf{x},t))\\) gradually decreases as the accumulation window size \\(W\\) increases. Note that the larger window accumulation is equivalent to the average of multiple smaller windows. The convergence demonstrates the existence of expectation and the result faithfully reflects the original light intensity with negligible disturbance. We further validate the zero-mean distribution hypothesis by collecting a series of static scene spike streams with varying light illuminations, and visualize the distribution of noise ratio in Fig. 4 (b). The result shows that the zero-mean assumption holds with sufficient illumination, which supports our self-supervised learning mechanism." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.231, + 0.114, + 0.728, + 0.129 + ], + "angle": 0, + "content": "Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.162, + 0.577, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.33, + 0.23, + 0.474, + 0.242 + ], + "angle": 0, + "content": "(a) Temporal accumulation" + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.145, + 0.785, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.635, + 0.23, + 0.748, + 0.241 + ], + "angle": 0, + "content": "(b) Noise distribution" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.253, + 0.785, + 0.322 + ], + "angle": 0, + "content": "Fig. 4: Validation of noise on real-world chromatic spikes. As shown in (a), as the accumulation window size \\( W \\) increases, the noise ratio in the accumulated images gradually decreases, making the images more closely resemble the static scene. In (b), we present the statistical noise fraction distribution across three different color channels and different illumination levels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.354, + 0.788, + 0.673 + ], + "angle": 0, + "content": "Self-supervised chromatic spike denoising module. Taking the inspiration of Batson et al. [1] and Wang et al. [31], which do not require paired noise-contaminated images or clean targets, we design a chromatic spike denoising module trained on only accessible real-world data with self-supervision learning strategy. Specifically, our denoising module operates based on the zero-mean characteristic of short-term frame noise and relies on the blind spot architecture to prevent the network from learning the identity mapping. Hence, the denoising module transforms the noise-contaminated images to their expectations to achieve minimal losses, producing desired clean images. Therefore, as shown in Fig. 3 (a), we begin by extracting the four channels of the chromatic spikes according to the RGGB pattern of the CFA, and the size of each channel is \\(\\frac{h}{2} \\times \\frac{w}{2}\\). For each channel \\(c\\), we preliminarily reconstruct a chromatic frame \\(I_{c}\\) by accumulating the chromatic spikes in a temporal window \\(W\\). Then, \\(I_{c}\\) is further partitioned into \\(\\frac{h}{2h_s} \\times \\frac{w}{2w_s}\\) cells, where \\(h_s\\) and \\(w_s\\) denote height and width of each cell. We sequentially mask one pixel for each cell following the order from top to bottom and left to right. Consequently, we obtain a masked volume with a size of \\(C \\times (\\frac{h}{2} \\times \\frac{w}{2})\\) corresponding to \\(I_{c}\\), where \\(C\\) is the channel of the masked volume that equals \\(h_s \\times w_s\\). Each channel of the mask volume acts as the blind spot operation to avoid identity mapping. We then employ a U-Net-inspired fully-convolutional neural network \\(f_{\\theta}(\\cdot)\\) to denoise each masked frame in this volume and merge the denoised frames to the clean image:" + }, + { + "type": "equation", + "bbox": [ + 0.343, + 0.687, + 0.785, + 0.705 + ], + "angle": 0, + "content": "\\[\nD \\left(I _ {c}\\right) = \\Phi \\left(f _ {\\theta} \\left(\\Omega_ {c, 0}\\right), f _ {\\theta} \\left(\\Omega_ {c, 1}\\right), \\dots , f _ {\\theta} \\left(\\Omega_ {c, C}\\right)\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.712, + 0.784, + 0.743 + ], + "angle": 0, + "content": "where \\(\\Omega_{c,i}\\) denotes the \\(i\\)-th masked frame in volume \\(\\Omega_c\\), \\(\\varPhi(\\cdot)\\) represents a function for mask-aware averaging:" + }, + { + "type": "equation", + "bbox": [ + 0.35, + 0.756, + 0.785, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\Phi \\left(f _ {\\theta} \\left(\\Omega_ {c, 0}\\right), \\dots\\right) = \\sum_ {i} M _ {i} f _ {\\theta} \\left(\\Omega_ {c, i}\\right) / \\sum_ {i} M _ {i}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "and \\( M_{i} \\) denotes the mask for \\( \\varOmega_{c,i}(\\varOmega_{c,i}=(1-M_i)I_c) \\). At last, the denoised four-channel images are recombined to form the final mosaicked chromatic frame. It is remarkable that the zero-mean hypothesis holds for arbitrary temporal window" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "S. Yang et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.151, + 0.784, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.237, + 0.788, + 0.28 + ], + "angle": 0, + "content": "Fig. 5: Progressive warping results. We visualize the intermediate frames from our progressive warping module. As the iterative step \\( l \\) increases, the noise in the reconstructed frames is gradually suppressed and the dynamic range is extended." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.307, + 0.784, + 0.336 + ], + "angle": 0, + "content": "size \\( W \\), indicating that our neural network can be trained and applied on varying chosen \\( W \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.784, + 0.381 + ], + "angle": 0, + "content": "Loss function. The chromatic spike denoising module is trained using real-world data with realistic noise. Similar to Wang et al. [31], we adopt the regularized re-visible loss:" + }, + { + "type": "equation", + "bbox": [ + 0.236, + 0.407, + 0.785, + 0.425 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {r}} = \\mathcal {L} _ {\\mathrm {r e v}} + \\eta \\mathcal {L} _ {\\mathrm {r e g}} = \\| D (I _ {c}) + \\lambda f _ {\\theta} (I _ {c}) - (\\lambda + 1) I _ {c} \\| _ {2} ^ {2} + \\eta \\| D (I _ {c}) - I _ {c} \\| _ {2} ^ {2} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.788, + 0.542 + ], + "angle": 0, + "content": "where \\(\\eta\\) is a constant hyper-parameter, and \\(\\lambda\\) is adjusted in the training progress. We further enhance the denoising module by incorporating an image quality assignment loss, which encourages the predicted score to be close to 1: \\(\\mathcal{L}_{\\mathrm{i}} = \\| 1 - G_{\\mathrm{iqa}}(D(I_c))\\|_2^2\\), where \\(G_{\\mathrm{iqa}}\\) denotes a pre-trained image quality assignment model [4]. The total training objective combines these losses: \\(\\mathcal{L} = \\mathcal{L}_{\\mathrm{r}} + \\lambda_{\\mathrm{i}}\\mathcal{L}_{\\mathrm{i}}\\). Note that the training objectives require no clean frames or human annotations, but only readily accessible real-world mosaicked chromatic spike streams." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.561, + 0.593, + 0.577 + ], + "angle": 0, + "content": "4.2 Color and HDR via progressive warping" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Although our chromatic spike denoising module can suppress the noise in frames reconstructed from short-term temporal windows, the dynamic range of each short-term frame is still restricted, as the number of photons arriving at the pixels within an extremely short period (e.g., \\( W = 10 \\)) is limited. Moreover, though our denoising module can fit arbitrary temporal window \\( W \\) theoretically, longer-term input frames lead to better output quality. To reconstruct images with less noise and higher dynamic range, it is necessary to select longer windows to preserve sufficient photons. However, directly warping the frames over the longer temporal window (e.g., \\( W = 100 \\)) is challenging since high-speed motions may span a wide range of pixels. Thus, as illustrated in Fig. 3 (b), we develop the progressive warping module, which only warps a small number of \\( K \\) chromatic frames in each step. By progressively applying the warping operation to the output of the previous step, at the \\( L \\)-th step, each output frame is equivalent to the weighted accumulation from \\( [(K - 1) \\times L + 1] \\times W \\) aligned chromatic spike planes. Setting a large value of \\( L \\) for the progressive warping module is equivalent to creating a pseudo-long exposure. The pseudo-long exposure leverages optical flow estimation at each step to obtain HDR frames without motion blur." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.231, + 0.114, + 0.728, + 0.131 + ], + "angle": 0, + "content": "Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "As shown in Fig. 3 (b), the progressive warping module involves the application of the pretrained chromatic spike denoising module from Sec. 4.1 and demosaicking for better optical flow estimation. In the initial 0-th step, we split \\( T \\) chromatic spike planes using a short-term temporal window \\( W \\) and accumulate the chromatic spikes in each window to form \\( \\frac{T}{W} \\) mosaicked frames with Eq. (4). We denote these initial mosaicked frames as \\( I_{c,0}(\\mathbf{x},t) \\), where \\( c \\in \\{R,G,B\\} \\), which is contaminated by noise given a small temporal window. Afterward, for more accurate optical flow estimation, we feed \\( I_{c,0}(\\mathbf{x},t) \\) to the self-supervised denoising module and the demosaicking module that employs linear interpolation to generate intermediate reconstruction frames:" + }, + { + "type": "equation", + "bbox": [ + 0.342, + 0.322, + 0.786, + 0.339 + ], + "angle": 0, + "content": "\\[\nR _ {0} (t) = \\Psi [ D (I _ {R, 0} (t)), D (I _ {G, 0} (t)), D (I _ {B, 0} (t)) ], \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.35, + 0.788, + 0.427 + ], + "angle": 0, + "content": "where \\(\\varPsi\\) denotes linear interpolation demosaicking and \\(\\mathbf{x}\\) is omitted. We then estimate the optical flow \\(F_{0}\\) with off-the-shelf methods from \\(R_0\\) [12]. Note that in this initial step, \\(I_{c,0}(\\mathbf{x},t)\\) is not warped and is mainly used to estimate a better initial optical flow \\(F_{0}\\). Subsequently, \\(F_{0}\\) is leveraged to warp \\(K\\times W\\) spike planes towards the reference frame in the 1-th step, described as:" + }, + { + "type": "equation", + "bbox": [ + 0.324, + 0.446, + 0.786, + 0.49 + ], + "angle": 0, + "content": "\\[\nI _ {c, 1} (\\mathbf {x}, t) = \\frac {M _ {c}}{K W} \\cdot \\sum_ {t ^ {\\prime} = t - K W / 2} ^ {t + K W / 2} S _ {c} (\\mathbf {x} + F _ {0} (\\mathbf {x}, t, t ^ {\\prime}), t ^ {\\prime}). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.788, + 0.626 + ], + "angle": 0, + "content": "Note that we utilize chromatic spike planes in the first two steps and mosaicked frames are warped in the later steps. Specifically, for the \\(l\\)-th iteration step, we continuously pass the accumulated frame \\(I_{c,l-1}\\) to the denoising module and the demosaicking module to facilitate the estimation of optical flow \\(F_{l-1}\\) from the previously generated frames \\(R_{l-1}\\), and warp \\(I_{c,l-1}\\) to the reference time. This operation can further promote noise suppression and dynamic range extension. In particular, we average \\(K\\) accumulated frames \\(I_{c,l-1}\\) obtained in the \\((l-1)\\)-th step, yielding:" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.646, + 0.786, + 0.689 + ], + "angle": 0, + "content": "\\[\nI _ {c, l} (\\mathbf {x}, t) = \\frac {1}{K} \\cdot \\sum_ {i = - K / 2} ^ {K / 2} I _ {c, l - 1} (\\mathbf {x} + F _ {l - 1} (\\mathbf {x}, t, t ^ {\\prime}), t ^ {\\prime}), \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.702, + 0.787, + 0.763 + ], + "angle": 0, + "content": "where \\( t' = t + iW \\). The entire procedure is iteratively repeated for \\( L \\) steps. In Fig. 5, we show a set of examples to demonstrate the effectiveness of the progressive warping module. It is evident that as the iteration step \\( l \\) increases, the details in both highlighted dark regions become clearer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.764, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Finally, we obtain color HDR video frames \\( R_{L} \\) with a frame rate of \\( \\frac{20000}{W} \\) FPS. Note that \\( W \\) is adjustable, which leads to higher frame rates for smaller values and lower frame rates for larger values. In our most experiments, we find that \\( W = 10 \\) offers a good balance between temporal resolution and reconstruction quality, which ensures 2000FPS high-speed color HDR video reconstruction." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "S. Yang et al." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.149, + 0.49, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.205, + 0.505, + 0.218 + ], + "angle": 0, + "content": "(a) Synthetic chromatic spike streams and reconstructions" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.148, + 0.772, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.205, + 0.777, + 0.217 + ], + "angle": 0, + "content": "(b) Real chromatic spike streams and reconstructions" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.23, + 0.785, + 0.272 + ], + "angle": 0, + "content": "Fig. 6: Illustration of our real-world collected chromatic spike streams, synthetic data derived from conventional cameras, and corresponding reconstructions from our proposed method." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.294, + 0.365, + 0.406 + ], + "angle": 0, + "content": "Table 1: Quantitative evaluation of the image reconstruction quality. The scores are averaged across our synthetic dataset. \\(\\uparrow\\) (↓) means higher (lower) is better." + }, + { + "type": "table", + "bbox": [ + 0.389, + 0.309, + 0.784, + 0.405 + ], + "angle": 0, + "content": "
MethodPSNR ↑SSIM ↑LPIPS ↓HDR-VDP3 ↑HDR-VQM ↓
TFI [46]19.2150.2460.7164.6450.7897
TFP(W=10) [46]23.2530.3310.6095.7780.7799
TFP(W=100) [46]28.7910.7200.2787.5450.4994
TFSTP [43]20.9450.2530.6704.5870.7843
MS23 [21]28.4100.6350.3597.1800.5765
Ours30.7800.8830.2087.9910.3023
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.455, + 0.367, + 0.472 + ], + "angle": 0, + "content": "5 Experiment" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.506, + 0.387, + 0.519 + ], + "angle": 0, + "content": "5.1 Data collection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.788, + 0.701 + ], + "angle": 0, + "content": "Synthetic Data. In line with the methodologies of monochromatic spike simulators (e.g., Zhao et al. [41], SpikeCV [42]), we develop a chromatic spike simulator, following the same Bayer pattern as the real-world chromatic spike camera. We leverage the frames in the GoPro dataset [26] and their corresponding synthetic spike streams to train a supervised learning model and compare it with our self-supervised learning method, which addresses the superiority of our strategy. We interpolate the HDR videos collected by Chang et al. [3] as the evaluation data for quantitative evaluation in our experiments. Some examples of these synthetic scenes are shown in Fig. 6 (a). For the detailed design of our chromatic spike simulator, please refer to our supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Real-world data. To accomplish real-data-driven self-supervised learning and validate the effectiveness of our proposed method, we capture a collection of real-world chromatic spikes using a spike camera equipped with the Bayer pattern. Our dataset comprises over 100 sets of mosaicked chromatic spikes captured from 30 diverse scenes, characterized by varying degrees of high-speed motion and high dynamic range, as depicted in Fig. 6 (b). The real-world data inherently contain the noise characteristics of a real spike camera prototype, posing significant challenges for denoising and demosaicking algorithms. Our real-world dataset is available at https://huggingface.co/datasets/YOUSIKI/chromatic-spikes." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.231, + 0.114, + 0.728, + 0.129 + ], + "angle": 0, + "content": "Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.607, + 0.163 + ], + "angle": 0, + "content": "5.2 Quantitative evaluation on synthetic data" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.172, + 0.788, + 0.368 + ], + "angle": 0, + "content": "We compare our method with existing spike-based video reconstruction methods, i.e., TFI [46], TFP [46], and TFSTP [43]. We acknowledge that it may not be entirely fair to compare our method with these approaches, as they are designed for monochromatic spikes. However, they can serve as baselines to demonstrate the effectiveness of our denoising and progressive warping modules. Since the data format of the spike camera and the single photon avalanche diodes (SPAD) are both single-bit signals, the color video reconstruction method can be adapted to chromatic spikes, and MS23 [21] is chosen for comparison7. The quantitative evaluation is performed on synthetic data. We select PSNR, SSIM [32], LPIPS [37], HDR-VDP3 [25], and HDR-VQM [27] as the metrics. The quantitative evaluation is shown in Tab. 1. We can see that the proposed method demonstrated superior performance over the competing techniques across all metrics." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.391, + 0.55, + 0.405 + ], + "angle": 0, + "content": "5.3 Qualitative evaluation on real data" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.789, + 0.644 + ], + "angle": 0, + "content": "To underscore the advantages of our method, we conduct a qualitative evaluation on real and synthetic chromatic spike streams. Fig. 7 presents a quadruple comparison. Our approach not only reconstructs motion blur free frames and recovers color information but also achieves the best balance between noise suppression and detail preservation across diverse scenes. In Fig. 7 (a), we create a colorful and dynamic environment by placing a high-speed electric fan between our chromatic spike camera and a set of toys. The reconstruction from TFP \\((W = 100)\\) exhibits noticeable motion blur in the blue bounding box. Conversely, TFI, TFP \\((W = 10)\\), TFSTP and MS23 [21] suffer from significant noise. On the contrary, our method yields a clean outcome. Fig. 7 (b) showcases flowing water in glass and a rotating toy duck. As indicated by red bounding boxes, TFI, TFP \\((W = 10)\\), TFSTP, and MS23 [21] contain different levels of noise. And TFP \\((W = 100)\\) suffers from heavy motion blur in blue bounding box. Meanwhile, our proposed method produces clean reconstruction without blur. Fig. 7 (c, d) further demonstrates the denoising advantage of our method on synthetic data." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.666, + 0.36, + 0.683 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.697, + 0.788, + 0.804 + ], + "angle": 0, + "content": "In this paper, we propose a novel real-data-driven approach for high-speed color HDR video reconstruction from mosaicked chromatic spikes. Through the analysis on the noise of chromatic spikes, we find the noise in short-term frames can be regarded as a zero-mean distribution. According to this characteristic, we design a self-supervised denoising module that can be trained on real-world chromatic spikes. We further propose a progressive warping module that enables the reconstruction of HDR frames from a pseudo long exposure. Experiments on" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.84 + ], + "angle": 0, + "content": "7 As the source code for MS23 [21] is not publicly accessible, we conduct reimplementation and modification for the evaluation." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "S. Yang et al." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.574 + ], + "angle": 0, + "content": "
TFITFP (W = 10)TFP (W = 100)TFSTPMS23Ours
(a)
(b)
(c)
(d)
" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.586, + 0.788, + 0.63 + ], + "angle": 0, + "content": "Fig. 7: Visual quality comparison of real (a, b) and synthetic (c, d) data between the proposed method and compared methods. Please zoom-in electronic versions for better details, and watch the videos in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.787, + 0.687 + ], + "angle": 0, + "content": "real-world data demonstrate the superiority of our approach for reconstructing 2000FPS color HDR videos." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.826 + ], + "angle": 0, + "content": "Limitations and future work. The Bayer-pattern CFA introduces chromatic information to the spikes but also reduces the intensity of incident light. As a result, compared to monochromatic spikes, the ability to preserve textures in low-light regions is relatively compromised in high-speed conditions. In future research, exploring different types of CFAs to enhance low-light performance would be worth investigating. And the excessive focus on denoising can result in over-smooth downsides, which may lead to the loss of fine details. Future work could involve the development of a more sophisticated denoising module to balance noise suppression and detail preservation." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.231, + 0.115, + 0.728, + 0.129 + ], + "angle": 0, + "content": "Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.146, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.237 + ], + "angle": 0, + "content": "This work was supported by National Science and Technology Major Project (Grant No. 2021ZD0109803) and National Natural Science Foundation of China (Grant No. 62088102, 62136001, 62301009). Bin Fan was supported by National Postdoctoral Program for Innovative Talents of China (Grant No.BX20230013)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.26, + 0.323, + 0.274 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.317 + ], + "angle": 0, + "content": "1. Batson, J., Royer, L.: Noise2Self: Blind denoising by self-supervision. In: ICML. pp. 524-533 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.785, + 0.345 + ], + "angle": 0, + "content": "2. Buades, A., Coll, B., Morel, J.M., Sbert, C.: Self-similarity driven color demosaicking. IEEE TIP 18(6), 1192-1202 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.346, + 0.785, + 0.372 + ], + "angle": 0, + "content": "3. Chang, Y., Zhou, C., Hong, Y., Hu, L., Xu, C., Huang, T., Shi, B.: 1000 FPS HDR video with a spike-RGB hybrid camera. In: CVPR. pp. 22180-22190 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.373, + 0.785, + 0.413 + ], + "angle": 0, + "content": "4. Chen, C., Mo, J., Hou, J., Wu, H., Liao, L., Sun, W., Yan, Q., Lin, W.: TOPIQ: A top-down approach from semantics to distortions for image quality assessment. IEEE Transactions on Image Processing (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.415, + 0.785, + 0.441 + ], + "angle": 0, + "content": "5. Chen, S., Duan, C., Yu, Z., Xiong, R., Huang, T.: Self-supervised mutual learning for dynamic scene reconstruction of spiking camera. In: IJCAI. p. 2859-2866 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.442, + 0.785, + 0.482 + ], + "angle": 0, + "content": "6. Chen, S., Yu, Z., Huang, T.: Self-supervised joint dynamic scene reconstruction and optical flow estimation for spiking camera. In: AAAI. vol. 37, pp. 350-358 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.484, + 0.785, + 0.51 + ], + "angle": 0, + "content": "7. Condat, L., Mosaddegh, S.: Joint demosaicking and denoising by total variation minimization. In: ICIP. pp. 2781-2784 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.511, + 0.785, + 0.537 + ], + "angle": 0, + "content": "8. Fossum, E.R., Ma, J., Masoodian, S., Anzagira, L., Zizza, R.: The quanta image sensor: Every photon counts. Sensors 16(8), 1260 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.538, + 0.785, + 0.565 + ], + "angle": 0, + "content": "9. Gharbi, M., Chaurasia, G., Paris, S., Durand, F.: Deep joint demosaicking and denoising. ACM TOG 35(6), 1-12 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.567, + 0.785, + 0.592 + ], + "angle": 0, + "content": "0. Gnanasambandam, A., Chan, S.H.: HDR imaging with quanta image sensors: Theoretical limits and optimal reconstruction. IEEE TCI 6, 1571-1585 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.594, + 0.785, + 0.633 + ], + "angle": 0, + "content": "11. Han, J., Zhou, C., Duan, P., Tang, Y., Xu, C., Xu, C., Huang, T., Shi, B.: Neuromorphic camera guided high dynamic range imaging. In: CVPR. pp. 1730-1739 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.635, + 0.785, + 0.674 + ], + "angle": 0, + "content": "12. Hasinoff, S.W., Sharlet, D., Geiss, R., Adams, A., Barron, J.T., Kainz, F., Chen, J., Levoy, M.: Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM TOG 35(6), 1-12 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.676, + 0.785, + 0.716 + ], + "angle": 0, + "content": "13. Heide, F., Rouf, M., Hullin, M.B., Labitzke, B., Heidrich, W., Kolb, A.: High-quality computational imaging through simple lenses. ACM TOG 32(5), 1-14 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.717, + 0.785, + 0.757 + ], + "angle": 0, + "content": "14. Huang, T., Zheng, Y., Yu, Z., Chen, R., Li, Y., Xiong, R., Ma, L., Zhao, J., Dong, S., Zhu, L., Li, J., Jia, S., Fu, Y., Shi, B., Wu, S., Tian, Y.: \\(1000 \\times\\) faster camera and machine vision with ordinary devices. Engineering 25, 110-119 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "15. Kiku, D., Monno, Y., Tanaka, M., Okutomi, M.: Beyond color difference: Residual interpolation for color image demosaicking. IEEE TIP 25(3), 1288-1300 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "16. Kokkinos, F., Lefkimmiatis, S.: Deep image demosaicking using a cascade of convolutional residual denoising networks. In: ECCV. pp. 303-319 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "17. Lichtsteiner, P., Posch, C., Delbruck, T.: A \\(128 \\times 128\\) 120 db \\(15\\mu s\\) latency asynchronous temporal contrast vision sensor. IEEE JSSC 43(2), 566-576 (2008)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.289, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "S. Yang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "18. Liu, L., Jia, X., Liu, J., Tian, Q.: Joint demosaicing and denoising with self guidance. In: CVPR. pp. 2240-2249 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "19. Liu, Y., Gutierrez-Barragan, F., Ingle, A., Gupta, M., Velten, A.: Single-photon camera guided extreme dynamic range imaging. In: WACV. pp. 1575-1585 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.785, + 0.233 + ], + "angle": 0, + "content": "20. Ma, S., Gupta, S., Ulku, A.C., Bruschini, C., Charbon, E., Gupta, M.: Quanta burst photography. ACM TOG 39(4), 79-1 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.785, + 0.261 + ], + "angle": 0, + "content": "21. Ma, S., Sundar, V., Mos, P., Bruschini, C., Charbon, E., Gupta, M.: Seeing photons in color. ACM TOG 42(4), 1-16 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.262, + 0.785, + 0.289 + ], + "angle": 0, + "content": "22. Mairal, J., Bach, F., Ponce, J., Sapiro, G., Zisserman, A.: Non-local sparse models for image restoration. In: ICCV. pp. 2272-2279 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.29, + 0.785, + 0.318 + ], + "angle": 0, + "content": "23. Mairal, J., Elad, M., Sapiro, G.: Sparse representation for color image restoration. IEEE TIP 17(1), 53-69 (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.319, + 0.785, + 0.346 + ], + "angle": 0, + "content": "24. Malvar, H.S., He, L.w., Cutler, R.: High-quality linear interpolation for demosaicing of bayer-patterned color images. In: ICASSP. vol. 3, pp. iii-485 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.347, + 0.785, + 0.388 + ], + "angle": 0, + "content": "25. Mantiuk, R.K., Hammou, D., Hanji, P.: HDR-VDP-3: A multi-metric for predicting image differences, quality and contrast distortions in high dynamic range and regular content. arXiv preprint arXiv:2304.13625 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.389, + 0.785, + 0.417 + ], + "angle": 0, + "content": "26. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: CVPR. pp. 3883-3891 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.418, + 0.785, + 0.459 + ], + "angle": 0, + "content": "27. Narwaria, M., Da Silva, M.P., Le Callet, P.: HDR-VQM: An objective quality measure for high dynamic range video. Signal Processing: Image Communication 35, 46-60 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.487 + ], + "angle": 0, + "content": "28. Pekkucuksen, I., Altunbasak, Y.: Gradient based threshold free color filter array interpolation. In: ICIP. pp. 137-140 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.529 + ], + "angle": 0, + "content": "29. Tan, D.S., Chen, W.Y., Hua, K.L.: Deepdemosaicking: Adaptive image demosaicking via multiple deep fully convolutional networks. IEEE TIP 27(5), 2408-2419 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.785, + 0.558 + ], + "angle": 0, + "content": "30. Tan, R., Zhang, K., Zuo, W., Zhang, L.: Color image demosaicking via deep residual learning. In: ICME. pp. 793-798 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.785, + 0.586 + ], + "angle": 0, + "content": "31. Wang, Z., Liu, J., Li, G., Han, H.: Blind2unblind: Self-supervised image denoising with visible blind spots. In: CVPR. pp. 2027-2036 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.785, + 0.614 + ], + "angle": 0, + "content": "32. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: From error visibility to structural similarity. IEEE TIP 13(4), 600-612 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.785, + 0.643 + ], + "angle": 0, + "content": "33. Xia, L., Zhao, J., Xiong, R., Huang, T.: SVFI: Spiking-based video frame interpolation for high-speed motion. In: AAAI. vol. 37, pp. 2910-2918 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.644, + 0.785, + 0.672 + ], + "angle": 0, + "content": "34. Xing, W., Egiazarian, K.: End-to-end learning for joint image demosaicing, denoising and super-resolution. In: CVPR. pp. 3507-3516 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.673, + 0.785, + 0.699 + ], + "angle": 0, + "content": "35. Yang, Y., Han, J., Liang, J., Sato, I., Shi, B.: Learning event guided high dynamic range video reconstruction. In: CVPR. pp. 13924-13934 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.785, + 0.741 + ], + "angle": 0, + "content": "36. Yu, G., Sapiro, G., Mallat, S.: Solving inverse problems with piecewise linear estimators: From gaussian mixture models to structured sparsity. IEEE TIP 21(5), 2481-2499 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.785, + 0.77 + ], + "angle": 0, + "content": "37. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR. pp. 586-595 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.785, + 0.798 + ], + "angle": 0, + "content": "38. Zhang, Z., Zheng, H., Hong, R., Xu, M., Yan, S., Wang, M.: Deep color consistent network for low-light image enhancement. In: CVPR. pp. 1899-1908 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "39. Zhao, J., Xiong, R., Liu, H., Zhang, J., Huang, T.: Spk2ImgNet: Learning to reconstruct dynamic scene from continuous spike stream. In: CVPR. pp. 11996-12005 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.231, + 0.114, + 0.728, + 0.129 + ], + "angle": 0, + "content": "Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "40. Zhao, J., Xiong, R., Xie, J., Shi, B., Yu, Z., Gao, W., Huang, T.: Reconstructing clear image for high-speed motion scene with a retina-inspired spike camera. IEEE TCI 8, 12-27 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "41. Zhao, J., Zhang, S., Ma, L., Yu, Z., Huang, T.: SpikingSIM: A bio-inspired spiking simulator. In: IEEE ISCAS. pp. 3003-3007 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.259 + ], + "angle": 0, + "content": "42. Zheng, Y., Zhang, J., Zhao, R., Ding, J., Chen, S., Xiong, R., Yu, Z., Huang, T.: SpikeCV: Open a continuous computer vision era. arXiv preprint arXiv:2303.11684 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.787, + 0.3 + ], + "angle": 0, + "content": "43. Zheng, Y., Zheng, L., Yu, Z., Huang, T., Wang, S.: Capture the moment: High-speed imaging with spiking cameras through short-term plasticity. IEEE TPAMI 45(7), 8127-8142 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.342 + ], + "angle": 0, + "content": "44. Zheng, Y., Zheng, L., Yu, Z., Shi, B., Tian, Y., Huang, T.: High-speed image reconstruction through short-term plasticity for spiking cameras. In: CVPR. pp. 6358-6367 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.787, + 0.37 + ], + "angle": 0, + "content": "45. Zhu, L., Dong, S., Huang, T., Tian, Y.: A retina-inspired sampling method for visual texture reconstruction. In: ICME. pp. 1432-1437 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.787, + 0.398 + ], + "angle": 0, + "content": "46. Zhu, L., Dong, S., Li, J., Huang, T., Tian, Y.: Retina-like visual image reconstruction via spiking neural model. In: CVPR. pp. 1438-1446 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.398 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes/6731bc50-147c-4c01-a7c6-498656963ba3_origin.pdf b/2024/Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes/6731bc50-147c-4c01-a7c6-498656963ba3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2439f382e06ed02d3c560373a5b89fd04597980d --- /dev/null +++ b/2024/Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes/6731bc50-147c-4c01-a7c6-498656963ba3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:93ef4073d6200c448741feb1a739733b1514d9d1afa29adb1170348e6f76fa75 +size 7596680 diff --git a/2024/Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes/full.md b/2024/Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes/full.md new file mode 100644 index 0000000000000000000000000000000000000000..96ec7f51141cfdcd4ea85f87ffe8434e2ed4d4af --- /dev/null +++ b/2024/Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes/full.md @@ -0,0 +1,291 @@ +# Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes + +Siqi Yang $^{1,\#}$ , Zhaojun Huang $^{2,3,\#}$ , Yakun Chang $^{5,6}$ , Bin Fan $^{4}$ , Zhaofei Yu $^{1}$ , and Boxin Shi $^{2,3,1,*}$ + +$^{1}$ Institute for Artificial Intelligence, Peking University + +2 State Key Lab of Multimedia Info. Processing, School of Computer Science, Peking University + +3 Nat'l Eng. Research Ctr. of Visual Technology, School of Computer Science, Peking University + +$^{4}$ Nat'l Key Lab of General AI, School of Intelligence Science and Technology, Peking University + +$^{5}$ Institute of Information Science, Beijing Jiaotong University + +$^{6}$ Visual Intelligence +X International Cooperation Joint Laboratory of the Ministry of Education {yousiki, huangzhaojun, binfan, yuzf12, shiboxin}@pku.edu.cn, ykchang@bjtu.edu.cn + +Abstract. The spike camera continuously records scene radiance with high-speed, high dynamic range, and low data redundancy properties, as a promising replacement for frame-based high-speed cameras. Previous methods for reconstructing color videos from monochromatic spikes are constrained in capturing full-temporal color information due to their reliance on compensating colors from low-speed RGB frames. Applying a Bayer-pattern color filter array to the spike sensor yields mosaicked chromatic spikes, which complicates noise distribution in high-speed conditions. By validating that the noise of short-term frames follows a zero-mean distribution, we leverage this hypothesis to develop a self-supervised denoising module trained exclusively on real-world data. Although noise is reduced in short-term frames, the long-term accumulation of incident photons is still necessary to construct HDR frames. Therefore, we introduce a progressive warping module to generate pseudo long-term exposure frames. This approach effectively mitigates motion blur artifacts in high-speed conditions. Integrating these modules forms a real-data-driven reconstruction method for mosaicked chromatic spikes. Extensive experiments conducted on both synthetic and real-world data demonstrate that our approach is effective in reconstructing 2000FPS color HDR videos with significantly reduced noise and motion blur compared to existing methods. + +Keywords: Spike camera $\cdot$ High-speed video $\cdot$ High dynamic range video $\cdot$ Color demosaicking $\cdot$ Self-supervised learning + +# 1 Introduction + +As a new class of image sensors inspired by the fovea of retina, the spike camera [3, 14, 46] offers attractive characteristics such as high temporal resolution (20000Hz), high dynamic range (HDR, $>90\mathrm{dB}$ ), and low data redundancy [14]. In comparison with conventional digital cameras that capture discrete frames, + +![](images/088473a1bf3a14e9fb809fe8332d9ac56789bb93302929ef88d113bb42985e0f.jpg) +(b) High-speed color HDR video +Fig. 1: High-speed color HDR video reconstruction from mosaicked chromatic spikes. (a) Comparison of different chromatic spike reconstruction (recon.) methodologies, including the adaptation of bio-inspired [45] and supervised learning [39] monochromatic methods to chromatic spikes and our proposed method. (b) A high-quality color HDR video with a frame rate of 2000FPS reconstructed by our method. The corresponding video is available in the supplementary material. + +spike cameras continuously accumulate the photons and trigger a single-bit spike of 1 when the detected photons reach a predefined threshold [14]. High-speed readout of spikes provides the spike camera with the potential for photon counting, and the high-dynamic-range ambient light can be estimated through continuous accumulation of spikes in the temporal domain. Therefore, spike cameras exhibit significant advantages in handling low-level vision tasks in high-speed conditions, such as video frame interpolation [33], HDR imaging [11] and high-frame-rate HDR video reconstruction [3]. + +With the higher demand for visual quality in images and videos, there is a strong desire that images or videos reconstructed from spike cameras can reproduce the real world to the greatest extent, with realistic color appearance and extended dynamic range at high frame rate. However, existing approaches [39, 40, 43, 44, 46] mainly focus on image or video reconstruction with monochromatic spikes, and cannot yet recover a high-speed color HDR video. To introduce color, one alternative practice [11] is to build a hybrid spike-RGB camera system, which exploits the colors captured by the conventional RGB camera to colorize the image reconstructed from spikes. Unfortunately, the color information provided by conventional RGB cameras often suffers from underexposure and overexposure, which inevitably lead to color deviations from the real world. Another strategy [3] to alleviate this problem is to replace the conventional RGB camera with an alternate exposure RGB camera to obtain HDR color, thereby enabling high-speed HDR color video reconstruction. However, the above methods have limitations in high-speed scenes, as the RGB camera has a low frame rate (e.g., 60FPS), which results in the loss of colors during the + +intervals between frames. In addition, the corresponding hybrid camera system relies on complex temporal synchronization and optical alignment, and the beam splitter with large space footprints would be detrimental to the construction of compact devices. + +Reconstructing high-speed color HDR videos based only on a single spike camera not only facilitates the acquisition of full-temporal color information but also offers the advantage of lightweight equipment. To this end, a tractable solution is to assemble a Bayer-pattern color filter array (CFA) in the spike camera, thereby producing mosaicked chromatic spikes. Consequently, a pressing challenge that arises is how to reconstruct high-speed color HDR videos from this new data representation. Nevertheless, directly adapting existing monochromatic spike reconstruction methods [39,40,43,44] to mosaicked chromatic spikes suffers from many issues, as shown in Fig. 1 (a). On the one hand, methods based on biology-inspired models [40,43,44] cannot take advantage of deep neural networks and are susceptible to varying degrees of noise contamination. On the other hand, supervised learning methods (e.g., [39]) trained using synthetic datasets have limited generalization ability when transferring to real-world data, because of the imperfection of spike noise modeling and simulation. These issues hinder the development of high-speed color HDR video reconstruction methods based on chromatic spike cameras and greatly limit the effective deployment of existing methods in real scenes. + +In this paper, we propose a real-data-driven mosaicked chromatic spike reconstruction method to recover high-speed color HDR videos, trained solely on readily accessible unlabelled spike streams. The mathematical analysis behind this requires no specific spike noise modeling or calibration, relying only on insights into the relationship between short-term and long-term frames accumulated from chromatic spikes. We design a deep learning framework that allows training directly on real-world data, naturally avoiding the influence of imperfect spike simulation and generalization gap. Specifically, we first propose a self-supervised denoising module to suppress the noise in short-term accumulation of chromatic spikes, which is trained on real-world spike streams without any annotation. Next, considering the limited dynamic range of short-term accumulated frames, we further present a progressive warping module to simulate a pseudolong exposure, which is effective in reconstructing HDR video frames without motion blur. Finally, we integrate these two modules to achieve higher-quality color HDR video reconstruction. + +In summary, our approach is able to achieve 2000FPS color HDR video reconstruction from mosaicked chromatic spikes with a single device in a real-data-driven manner, as illustrated in Fig. 1 (b), because of the technical contributions from these two complementary aspects: + +- We design a self-supervised denoising module to eliminate the noise from short-term accumulation (with high-speed motion cues) without relying on chromatic spike noise modeling and simulation. + +- We propose a progressive warping module with pseudo long-term exposure (with expanded dynamic range) to reconstruct HDR frames without introducing motion blur in dynamic scenes. + +# 2 Related Work + +Video reconstruction with monochromatic spikes. To better satisfy human visual perception and facilitate direct processing by machines, reconstructing the corresponding video frame from spikes has been widely studied. Prevalent image reconstruction methods typically employ the temporal statistical properties of spike cameras [43-46]. Zhu et al. [45] explored the spike generation principle and proposed estimating the firing frequency or firing interval of each pixel to reconstruct dynamic scenes, which are known as texture from playback (TFP) and texture from inter-spike-intervals (TFI), respectively. Recently, another line of work has enhanced firing frequency estimates by mimicking human physiological mechanisms, such as retina-like visual imaging [46] and short-term plasticity [43, 44]. Although these methods bring better interpretability, they require careful window selection to balance noise and motion blur, and the reconstruction effect is not yet satisfactory. The powerful representation capabilities of deep learning also inject vitality into spike-based image reconstruction. Spk2ImgNet [39] is the first spike-to-image deep network architecture based on deformable convolution, achieving impressive results. Further, several self-supervised methods [5,6] have also been developed to reduce the dependence on synthetic datasets. However, these methods are tailored for monochromatic spikes and thus cannot be directly applied to mosaicked chromatic spikes. + +Demosaicking and denoising with raw images. The purpose of raw image demosaicking is to recover a full-color image from a sub-sampled mosaicked image that contains potential noise. To deal with the ill-posedness of image demosaicking, various image priors have been introduced to regularize the recovery process, such as heuristics [2, 24], sparsity [23, 36], total variation [7, 13], integral gradient [28], self-similarity [22], and residual interpolation [15]. In addition, researchers designed deep neural networks [16, 29, 30] to automatically learn the prior knowledge required for image demosaicking. Moreover, considering the coupling between demosaicking and denoising, some deep learning-based joint demosaicking and denoising network models [9, 18, 34, 38] have also been developed to efficiently remove mosaicks from noisy raw images. + +HDR video with non-conventional sensors. In recent years, a series of nonconventional sensors, such as event cameras [17], spike cameras [14], and quanta image sensors (QIS) [8], have been invented, empowering the ability to capture HDR information in high-speed motion scenes. QIS-based HDR video reconstruction methods [10, 19, 20] were developed to accomplish perceptual HDR illumination. Additionally, the intensity image reconstructed by event cameras [35] or spike cameras [11] can serve as a proxy for enhancing details in low dynamic range (LDR) images. More recently, a hybrid camera system consisting of a spike camera and an alternating-exposure RGB camera is constructed in [3] for $1000 \times$ + +![](images/5645962b9ba6da6cceab1ca074ce59a65902ea8bcdfc7b594e05238c6175edf4.jpg) +Fig. 2: Chromatic spike accumulation and firing mechanism. We put a Bayer pattern CFA in the spike camera to filter incident light before the photons arrive at the sensor. Each individual pixel corresponds to one of red, green, and blue channels, accumulates the filtered photons, and triggers a spike signal, then resets when it surpasses the threshold $E_{\mathrm{th}}$ . Reading out the chromatic spikes forms the chromatic spike planes. + +frame rate HDR video reconstruction. Nevertheless, these methods also fall short when dealing with mosaicked chromatic spikes. + +# 3 Preliminaries + +We initially present a concise review of the operational principles of spike cameras, encompassing both monochromatic spike cameras and those equipped with Color Filter Array (CFA). Subsequently, we discuss the baseline reconstruction techniques for monochromatic spikes. + +Monochromatic spikes. The spike camera persistently accumulates electrons produced by incoming photons. When the electrons accumulated at a pixel surpass the predefined threshold $E_{\mathrm{th}}$ , it triggers a spike signal, immediately resetting the accumulator to initiate a new cycle of accumulation: + +$$ +E (\mathbf {x}, t) = \int \alpha L (\mathbf {x}, t) d t \mod E _ {\mathrm {t h}}, \tag {1} +$$ + +where $\alpha$ denotes the conversion ratio and $L(\mathbf{x},t)$ denotes the incoming light intensity of pixel $\mathbf{x}$ at time $t$ . Consequently, the spike signal $S(\mathbf{x},t)$ can be defined as: + +$$ +S (\mathbf {x}, t) = \left\{ \begin{array}{l l} 1, & \text {i f} E (\mathbf {x}, t) \geq E _ {\mathrm {t h}}, \\ 0, & \text {o t h e r w i s e .} \end{array} \right. \tag {2} +$$ + +As all the pixels of a spike sensor are arranged in an array, reading out spikes from the pixel array at each time stamp $t$ yields a spike plane of dimensions $h \times w$ , where $h$ and $w$ are the height and width of the spike data, respectively. Given the readout frequency of 20000Hz, the spike camera [14] allows for capturing spike planes at a frame rate of 20000FPS. + +Chromatic spikes. To incorporate color information while capturing HDR scenes with high-speed motion at the same temporal resolution, we can put a + +CFA in front of the sensor of the spike camera. In a similar way as conventional cameras, the CFA applied to spike sensors presents a Bayer pattern. Each pixel of the spike camera sensor with CFA corresponds to only one channel among red, green, and blue, while the firing mechanism of spikes remains unchanged. As illustrated in Fig. 2, under the influence of the CFA, a spike signal is generated when the cumulative electrons $E_{c}(\mathbf{x},t)$ produced by the incident photons of the corresponding spectral band meets the threshold $E_{\mathrm{th}}$ . The chromatic spike signal $S_{c}(\mathbf{x},t)$ can be recorded as: + +$$ +S _ {c} (\mathbf {x}, t) = \left\{ \begin{array}{l l} 1, & \text {i f} E _ {c} (\mathbf {x}, t) \geq E _ {\mathrm {t h}}, \\ 0, & \text {o t h e r w i s e}, \end{array} \right. \tag {3} +$$ + +where $c \in \{R, G, B\}$ . Consequently, similar to the firing process of monochromatic spikes, we can obtain a mosaicked chromatic spike plane at each time stamp $t$ . + +Baseline for chromatic frame reconstruction. Preliminary chromatic frame reconstruction can be achieved similarly to [45] by accumulating the spike signals within a defined temporal window: + +$$ +I _ {c} (\mathbf {x}, t) = \frac {M _ {c}}{W} \cdot \sum_ {\tau = t - W / 2} ^ {t + W / 2} S _ {c} (\mathbf {x}, \tau), \tag {4} +$$ + +where $W$ is the size of the accumulation temporal window, $M_{c}$ represents the maximum dynamic range that can be perceived in channel $c$ . Directly applying Eq. (4) to chromatic spikes as a baseline reconstruction algorithm can only obtain mosaicked chromatic frames, which is similar to the RAW image of conventional frame-based sensors. Demosaicking the chromatic frame is challenging since there is a trade-off between noise and superposition of colors. A short-term temporal window $W$ results in severe noise, whereas a long-term $W$ leads to the spatial superposition of varying color information. Furthermore, both noise distribution of chromatic spikes and object motions are more complex, making the direct application of off-the-shelf demosaicking methods insufficient for recovering high-quality color HDR videos. + +# 4 Real-data-driven chromatic spikes to video + +Overview. As shown in Fig. 3, we propose a progressive mechanism for high-speed color HDR video reconstruction. The pipeline starts with the accumulation of mosaicked chromatic spikes to reconstruct short-term frames. While these short-term frames are free of motion blur, they suffer from significant noise contamination. Consequently, as shown in Fig. 3 (a), we propose the chromatic spike denoising module (Sec. 4.1) to suppress the noise. The key insight of the denoising module is that a sufficiently long-term frame (e.g., $W = 10000$ , corresponding to 0.5s) of chromatic spikes can produce promising reconstructions in static (or low-speed) scenes, effectively representing the expectations of short-term frames. Given the limited incident photons in short-term frames, which + +![](images/4110ea8fc7ba3608e4170bfb8130d020434bf2380736740d9b794fd46be56169.jpg) +Fig. 3: Illustration of the overall pipeline. For each sequence of mosaicked chromatic spikes, we split them into four channels according to the RGGB pattern. The four-channel spikes are fed to our progressive warping module, which aligns multiple short-term frames to achieve "pseudo-long exposure". (a) Our chromatic spike denoising module utilizes a self-supervised training strategy, which is pretrained on real-world data. (b) Our progressive warping module consists of $L$ steps of frame warping operations, iteratively achieving high-quality HDR frames. At each step of the progressive warping module, we leverage the pretrained chromatic spikes denoising module to facilitate optical flow estimation. $R_{L}$ is the final output color HDR video with 2000FPS. + +lack the HDR characteristic, extending the accumulation time is a feasible strategy for HDR reconstruction. As shown in Fig. 3 (b), we design the progressive warping module in Sec. 4.2 to establish a pseudo long exposure. At each step of the warping module, chromatic spikes frames are warped according to the estimated optical flow, providing pseudo long-term frames with less noise and higher dynamic range. To leverage a wide range of existing conventional RGB optical flow estimation approaches, we integrate the denoising module into each iteration step of the warping module. After $L$ times of iteration, the progressive warping module produces an HDR color video with a frame rate of 2000FPS. + +# 4.1 Denoising of chromatic spikes + +Noise distribution of chromatic spikes noise. The noise in monochromatic spikes is primarily categorized into two types: diffuse noise and inherent noise. For the mosaicked chromatic spikes, although the incorporation of a CFA slightly alters the noise distribution for pixels under different filters, pixels corresponding + +to the same color exhibit similar noise distribution patterns as monochromatic spikes, except for its expectation and variance. Given the chromatic spikes $S_{c}$ , for each $c \in \{R,G,B\}$ , we use $Q_{d}$ to denote the number of diffuse noise photons arriving at a pixel in a time period of $t$ , and $\gamma$ to denote the likelihood of $Q_{d}$ diffuse noise photons. The probability of diffuse noise can be modeled as the Poisson distribution [41]: + +$$ +\mathcal {P} \left(Q _ {d} | t, \gamma\right) = \frac {\left(\gamma t\right) ^ {Q _ {d}}}{Q _ {d} !} \exp (- \gamma W). \tag {5} +$$ + +Hence, the expected number of diffuse noise spikes in temporal window $W$ is + +$$ +\mathbb {E} \left[ N _ {d} \right] = (\alpha \gamma W) / E _ {\mathrm {t h}}. \tag {6} +$$ + +For inherent noise, which triggers spikes in the absence of light, we use $T_{i}$ to indicate the interval between the noise spikes. Then, the interval between such noise spikes can be represented by a Gaussian model [41]: + +$$ +\mathcal {N} (T _ {i} | \mu , \sigma) = \frac {1}{\sigma \sqrt {2 \pi}} \cdot \exp \left(- \frac {1}{2} \cdot \left(\frac {T _ {i} - \mu}{\sigma}\right) ^ {2}\right). \tag {7} +$$ + +We approximate the expectation of the number of inherent noise spikes in the temporal window as + +$$ +\mathbb {E} [ N _ {i} ] = \int_ {1} ^ {+ \infty} \frac {W}{T _ {i}} \cdot \frac {1}{\sigma \sqrt {2 \pi}} \cdot \exp (- \frac {1}{2} \cdot (\frac {T _ {i} - \mu}{\sigma}) ^ {2}) d T _ {i} = C _ {i} W, (8) +$$ + +where $C_i$ denotes the integral after factoring out $W$ from the integral expression. Combining Eq. (6) and Eq. (8), the expectation of the fraction of noise spikes can be formulated as: + +$$ +\mathbb {E} [ N ] / W = \left(\mathbb {E} \left[ N _ {d} \right] + \mathbb {E} \left[ N _ {i} \right]\right) / W = \alpha \gamma / E _ {\mathrm {t h}} + C _ {i}. \tag {9} +$$ + +Instead of the actual incoming light intensity, we estimate the expectation of short-term accumulations $\mathbb{E}[I_c(\mathbf{x},t)] = L(\mathbf{x},t) + \mathbb{E}[N] / W$ . Thus, the noise between short-term frames and their expectations naturally follows a zero-mean distribution. We demonstrate that the expected fraction of noise spikes is relatively small, especially with sufficient light illumination, and doesn't disturb the visual semantic. As shown in Fig. 4 (a), the noise ratio $(I_{c}(\mathbf{x},t) / \mathbb{E}[I_{c}(\mathbf{x},t)] - 1)$ in the accumulated images $(I_{c}(\mathbf{x},t))$ gradually decreases as the accumulation window size $W$ increases. Note that the larger window accumulation is equivalent to the average of multiple smaller windows. The convergence demonstrates the existence of expectation and the result faithfully reflects the original light intensity with negligible disturbance. We further validate the zero-mean distribution hypothesis by collecting a series of static scene spike streams with varying light illuminations, and visualize the distribution of noise ratio in Fig. 4 (b). The result shows that the zero-mean assumption holds with sufficient illumination, which supports our self-supervised learning mechanism. + +![](images/873a2cf93fcb65972ef4faeba1ff17c2e1010b824804e1a25a48be79ee65a821.jpg) +(a) Temporal accumulation + +![](images/d30b55fc59d7f7a9e81a99da41cc6da9cc45b46517beb72d70a1bd63f0508782.jpg) +(b) Noise distribution +Fig. 4: Validation of noise on real-world chromatic spikes. As shown in (a), as the accumulation window size $W$ increases, the noise ratio in the accumulated images gradually decreases, making the images more closely resemble the static scene. In (b), we present the statistical noise fraction distribution across three different color channels and different illumination levels. + +Self-supervised chromatic spike denoising module. Taking the inspiration of Batson et al. [1] and Wang et al. [31], which do not require paired noise-contaminated images or clean targets, we design a chromatic spike denoising module trained on only accessible real-world data with self-supervision learning strategy. Specifically, our denoising module operates based on the zero-mean characteristic of short-term frame noise and relies on the blind spot architecture to prevent the network from learning the identity mapping. Hence, the denoising module transforms the noise-contaminated images to their expectations to achieve minimal losses, producing desired clean images. Therefore, as shown in Fig. 3 (a), we begin by extracting the four channels of the chromatic spikes according to the RGGB pattern of the CFA, and the size of each channel is $\frac{h}{2} \times \frac{w}{2}$ . For each channel $c$ , we preliminarily reconstruct a chromatic frame $I_{c}$ by accumulating the chromatic spikes in a temporal window $W$ . Then, $I_{c}$ is further partitioned into $\frac{h}{2h_s} \times \frac{w}{2w_s}$ cells, where $h_s$ and $w_s$ denote height and width of each cell. We sequentially mask one pixel for each cell following the order from top to bottom and left to right. Consequently, we obtain a masked volume with a size of $C \times (\frac{h}{2} \times \frac{w}{2})$ corresponding to $I_{c}$ , where $C$ is the channel of the masked volume that equals $h_s \times w_s$ . Each channel of the mask volume acts as the blind spot operation to avoid identity mapping. We then employ a U-Net-inspired fully-convolutional neural network $f_{\theta}(\cdot)$ to denoise each masked frame in this volume and merge the denoised frames to the clean image: + +$$ +D \left(I _ {c}\right) = \Phi \left(f _ {\theta} \left(\Omega_ {c, 0}\right), f _ {\theta} \left(\Omega_ {c, 1}\right), \dots , f _ {\theta} \left(\Omega_ {c, C}\right)\right), \tag {10} +$$ + +where $\Omega_{c,i}$ denotes the $i$ -th masked frame in volume $\Omega_c$ , $\varPhi(\cdot)$ represents a function for mask-aware averaging: + +$$ +\Phi \left(f _ {\theta} \left(\Omega_ {c, 0}\right), \dots\right) = \sum_ {i} M _ {i} f _ {\theta} \left(\Omega_ {c, i}\right) / \sum_ {i} M _ {i}, \tag {11} +$$ + +and $M_{i}$ denotes the mask for $\varOmega_{c,i}(\varOmega_{c,i}=(1-M_i)I_c)$ . At last, the denoised four-channel images are recombined to form the final mosaicked chromatic frame. It is remarkable that the zero-mean hypothesis holds for arbitrary temporal window + +![](images/dd5c16658422c8ea48a85e862587f1e9166987e0def46f1d59f24ebb237a1f7a.jpg) +Fig. 5: Progressive warping results. We visualize the intermediate frames from our progressive warping module. As the iterative step $l$ increases, the noise in the reconstructed frames is gradually suppressed and the dynamic range is extended. + +size $W$ , indicating that our neural network can be trained and applied on varying chosen $W$ . + +Loss function. The chromatic spike denoising module is trained using real-world data with realistic noise. Similar to Wang et al. [31], we adopt the regularized re-visible loss: + +$$ +\mathcal {L} _ {\mathrm {r}} = \mathcal {L} _ {\mathrm {r e v}} + \eta \mathcal {L} _ {\mathrm {r e g}} = \| D (I _ {c}) + \lambda f _ {\theta} (I _ {c}) - (\lambda + 1) I _ {c} \| _ {2} ^ {2} + \eta \| D (I _ {c}) - I _ {c} \| _ {2} ^ {2} \tag {12} +$$ + +where $\eta$ is a constant hyper-parameter, and $\lambda$ is adjusted in the training progress. We further enhance the denoising module by incorporating an image quality assignment loss, which encourages the predicted score to be close to 1: $\mathcal{L}_{\mathrm{i}} = \| 1 - G_{\mathrm{iqa}}(D(I_c))\|_2^2$ , where $G_{\mathrm{iqa}}$ denotes a pre-trained image quality assignment model [4]. The total training objective combines these losses: $\mathcal{L} = \mathcal{L}_{\mathrm{r}} + \lambda_{\mathrm{i}}\mathcal{L}_{\mathrm{i}}$ . Note that the training objectives require no clean frames or human annotations, but only readily accessible real-world mosaicked chromatic spike streams. + +# 4.2 Color and HDR via progressive warping + +Although our chromatic spike denoising module can suppress the noise in frames reconstructed from short-term temporal windows, the dynamic range of each short-term frame is still restricted, as the number of photons arriving at the pixels within an extremely short period (e.g., $W = 10$ ) is limited. Moreover, though our denoising module can fit arbitrary temporal window $W$ theoretically, longer-term input frames lead to better output quality. To reconstruct images with less noise and higher dynamic range, it is necessary to select longer windows to preserve sufficient photons. However, directly warping the frames over the longer temporal window (e.g., $W = 100$ ) is challenging since high-speed motions may span a wide range of pixels. Thus, as illustrated in Fig. 3 (b), we develop the progressive warping module, which only warps a small number of $K$ chromatic frames in each step. By progressively applying the warping operation to the output of the previous step, at the $L$ -th step, each output frame is equivalent to the weighted accumulation from $[(K - 1) \times L + 1] \times W$ aligned chromatic spike planes. Setting a large value of $L$ for the progressive warping module is equivalent to creating a pseudo-long exposure. The pseudo-long exposure leverages optical flow estimation at each step to obtain HDR frames without motion blur. + +As shown in Fig. 3 (b), the progressive warping module involves the application of the pretrained chromatic spike denoising module from Sec. 4.1 and demosaicking for better optical flow estimation. In the initial 0-th step, we split $T$ chromatic spike planes using a short-term temporal window $W$ and accumulate the chromatic spikes in each window to form $\frac{T}{W}$ mosaicked frames with Eq. (4). We denote these initial mosaicked frames as $I_{c,0}(\mathbf{x},t)$ , where $c \in \{R,G,B\}$ , which is contaminated by noise given a small temporal window. Afterward, for more accurate optical flow estimation, we feed $I_{c,0}(\mathbf{x},t)$ to the self-supervised denoising module and the demosaicking module that employs linear interpolation to generate intermediate reconstruction frames: + +$$ +R _ {0} (t) = \Psi [ D (I _ {R, 0} (t)), D (I _ {G, 0} (t)), D (I _ {B, 0} (t)) ], \tag {13} +$$ + +where $\varPsi$ denotes linear interpolation demosaicking and $\mathbf{x}$ is omitted. We then estimate the optical flow $F_{0}$ with off-the-shelf methods from $R_0$ [12]. Note that in this initial step, $I_{c,0}(\mathbf{x},t)$ is not warped and is mainly used to estimate a better initial optical flow $F_{0}$ . Subsequently, $F_{0}$ is leveraged to warp $K\times W$ spike planes towards the reference frame in the 1-th step, described as: + +$$ +I _ {c, 1} (\mathbf {x}, t) = \frac {M _ {c}}{K W} \cdot \sum_ {t ^ {\prime} = t - K W / 2} ^ {t + K W / 2} S _ {c} (\mathbf {x} + F _ {0} (\mathbf {x}, t, t ^ {\prime}), t ^ {\prime}). \tag {14} +$$ + +Note that we utilize chromatic spike planes in the first two steps and mosaicked frames are warped in the later steps. Specifically, for the $l$ -th iteration step, we continuously pass the accumulated frame $I_{c,l-1}$ to the denoising module and the demosaicking module to facilitate the estimation of optical flow $F_{l-1}$ from the previously generated frames $R_{l-1}$ , and warp $I_{c,l-1}$ to the reference time. This operation can further promote noise suppression and dynamic range extension. In particular, we average $K$ accumulated frames $I_{c,l-1}$ obtained in the $(l-1)$ -th step, yielding: + +$$ +I _ {c, l} (\mathbf {x}, t) = \frac {1}{K} \cdot \sum_ {i = - K / 2} ^ {K / 2} I _ {c, l - 1} (\mathbf {x} + F _ {l - 1} (\mathbf {x}, t, t ^ {\prime}), t ^ {\prime}), \tag {15} +$$ + +where $t' = t + iW$ . The entire procedure is iteratively repeated for $L$ steps. In Fig. 5, we show a set of examples to demonstrate the effectiveness of the progressive warping module. It is evident that as the iteration step $l$ increases, the details in both highlighted dark regions become clearer. + +Finally, we obtain color HDR video frames $R_{L}$ with a frame rate of $\frac{20000}{W}$ FPS. Note that $W$ is adjustable, which leads to higher frame rates for smaller values and lower frame rates for larger values. In our most experiments, we find that $W = 10$ offers a good balance between temporal resolution and reconstruction quality, which ensures 2000FPS high-speed color HDR video reconstruction. + +![](images/cc72206286cdabc1ba3833f93db53c6000e52ff8f7b3cf851429b593a7389210.jpg) +(a) Synthetic chromatic spike streams and reconstructions + +![](images/3627e88a3ec9de2070838e63d9324d5bf1514617c639f847a1ab1e86bb569626.jpg) +(b) Real chromatic spike streams and reconstructions +Fig. 6: Illustration of our real-world collected chromatic spike streams, synthetic data derived from conventional cameras, and corresponding reconstructions from our proposed method. + +Table 1: Quantitative evaluation of the image reconstruction quality. The scores are averaged across our synthetic dataset. $\uparrow$ (↓) means higher (lower) is better. + +
MethodPSNR ↑SSIM ↑LPIPS ↓HDR-VDP3 ↑HDR-VQM ↓
TFI [46]19.2150.2460.7164.6450.7897
TFP(W=10) [46]23.2530.3310.6095.7780.7799
TFP(W=100) [46]28.7910.7200.2787.5450.4994
TFSTP [43]20.9450.2530.6704.5870.7843
MS23 [21]28.4100.6350.3597.1800.5765
Ours30.7800.8830.2087.9910.3023
+ +# 5 Experiment + +# 5.1 Data collection + +Synthetic Data. In line with the methodologies of monochromatic spike simulators (e.g., Zhao et al. [41], SpikeCV [42]), we develop a chromatic spike simulator, following the same Bayer pattern as the real-world chromatic spike camera. We leverage the frames in the GoPro dataset [26] and their corresponding synthetic spike streams to train a supervised learning model and compare it with our self-supervised learning method, which addresses the superiority of our strategy. We interpolate the HDR videos collected by Chang et al. [3] as the evaluation data for quantitative evaluation in our experiments. Some examples of these synthetic scenes are shown in Fig. 6 (a). For the detailed design of our chromatic spike simulator, please refer to our supplementary material. + +Real-world data. To accomplish real-data-driven self-supervised learning and validate the effectiveness of our proposed method, we capture a collection of real-world chromatic spikes using a spike camera equipped with the Bayer pattern. Our dataset comprises over 100 sets of mosaicked chromatic spikes captured from 30 diverse scenes, characterized by varying degrees of high-speed motion and high dynamic range, as depicted in Fig. 6 (b). The real-world data inherently contain the noise characteristics of a real spike camera prototype, posing significant challenges for denoising and demosaicking algorithms. Our real-world dataset is available at https://huggingface.co/datasets/YOUSIKI/chromatic-spikes. + +# 5.2 Quantitative evaluation on synthetic data + +We compare our method with existing spike-based video reconstruction methods, i.e., TFI [46], TFP [46], and TFSTP [43]. We acknowledge that it may not be entirely fair to compare our method with these approaches, as they are designed for monochromatic spikes. However, they can serve as baselines to demonstrate the effectiveness of our denoising and progressive warping modules. Since the data format of the spike camera and the single photon avalanche diodes (SPAD) are both single-bit signals, the color video reconstruction method can be adapted to chromatic spikes, and MS23 [21] is chosen for comparison7. The quantitative evaluation is performed on synthetic data. We select PSNR, SSIM [32], LPIPS [37], HDR-VDP3 [25], and HDR-VQM [27] as the metrics. The quantitative evaluation is shown in Tab. 1. We can see that the proposed method demonstrated superior performance over the competing techniques across all metrics. + +# 5.3 Qualitative evaluation on real data + +To underscore the advantages of our method, we conduct a qualitative evaluation on real and synthetic chromatic spike streams. Fig. 7 presents a quadruple comparison. Our approach not only reconstructs motion blur free frames and recovers color information but also achieves the best balance between noise suppression and detail preservation across diverse scenes. In Fig. 7 (a), we create a colorful and dynamic environment by placing a high-speed electric fan between our chromatic spike camera and a set of toys. The reconstruction from TFP $(W = 100)$ exhibits noticeable motion blur in the blue bounding box. Conversely, TFI, TFP $(W = 10)$ , TFSTP and MS23 [21] suffer from significant noise. On the contrary, our method yields a clean outcome. Fig. 7 (b) showcases flowing water in glass and a rotating toy duck. As indicated by red bounding boxes, TFI, TFP $(W = 10)$ , TFSTP, and MS23 [21] contain different levels of noise. And TFP $(W = 100)$ suffers from heavy motion blur in blue bounding box. Meanwhile, our proposed method produces clean reconstruction without blur. Fig. 7 (c, d) further demonstrates the denoising advantage of our method on synthetic data. + +# 6 Conclusion + +In this paper, we propose a novel real-data-driven approach for high-speed color HDR video reconstruction from mosaicked chromatic spikes. Through the analysis on the noise of chromatic spikes, we find the noise in short-term frames can be regarded as a zero-mean distribution. According to this characteristic, we design a self-supervised denoising module that can be trained on real-world chromatic spikes. We further propose a progressive warping module that enables the reconstruction of HDR frames from a pseudo long exposure. Experiments on + +
TFITFP (W = 10)TFP (W = 100)TFSTPMS23Ours
(a)
(b)
(c)
(d)
+ +Fig. 7: Visual quality comparison of real (a, b) and synthetic (c, d) data between the proposed method and compared methods. Please zoom-in electronic versions for better details, and watch the videos in the supplementary material. + +real-world data demonstrate the superiority of our approach for reconstructing 2000FPS color HDR videos. + +Limitations and future work. The Bayer-pattern CFA introduces chromatic information to the spikes but also reduces the intensity of incident light. As a result, compared to monochromatic spikes, the ability to preserve textures in low-light regions is relatively compromised in high-speed conditions. In future research, exploring different types of CFAs to enhance low-light performance would be worth investigating. And the excessive focus on denoising can result in over-smooth downsides, which may lead to the loss of fine details. Future work could involve the development of a more sophisticated denoising module to balance noise suppression and detail preservation. + +# Acknowledgements + +This work was supported by National Science and Technology Major Project (Grant No. 2021ZD0109803) and National Natural Science Foundation of China (Grant No. 62088102, 62136001, 62301009). Bin Fan was supported by National Postdoctoral Program for Innovative Talents of China (Grant No.BX20230013). + +# References + +1. Batson, J., Royer, L.: Noise2Self: Blind denoising by self-supervision. In: ICML. pp. 524-533 (2019) +2. Buades, A., Coll, B., Morel, J.M., Sbert, C.: Self-similarity driven color demosaicking. IEEE TIP 18(6), 1192-1202 (2009) +3. Chang, Y., Zhou, C., Hong, Y., Hu, L., Xu, C., Huang, T., Shi, B.: 1000 FPS HDR video with a spike-RGB hybrid camera. In: CVPR. pp. 22180-22190 (2023) +4. Chen, C., Mo, J., Hou, J., Wu, H., Liao, L., Sun, W., Yan, Q., Lin, W.: TOPIQ: A top-down approach from semantics to distortions for image quality assessment. IEEE Transactions on Image Processing (2024) +5. Chen, S., Duan, C., Yu, Z., Xiong, R., Huang, T.: Self-supervised mutual learning for dynamic scene reconstruction of spiking camera. In: IJCAI. p. 2859-2866 (2022) +6. Chen, S., Yu, Z., Huang, T.: Self-supervised joint dynamic scene reconstruction and optical flow estimation for spiking camera. In: AAAI. vol. 37, pp. 350-358 (2023) +7. Condat, L., Mosaddegh, S.: Joint demosaicking and denoising by total variation minimization. In: ICIP. pp. 2781-2784 (2012) +8. Fossum, E.R., Ma, J., Masoodian, S., Anzagira, L., Zizza, R.: The quanta image sensor: Every photon counts. Sensors 16(8), 1260 (2016) +9. Gharbi, M., Chaurasia, G., Paris, S., Durand, F.: Deep joint demosaicking and denoising. ACM TOG 35(6), 1-12 (2016) +0. Gnanasambandam, A., Chan, S.H.: HDR imaging with quanta image sensors: Theoretical limits and optimal reconstruction. IEEE TCI 6, 1571-1585 (2020) +11. Han, J., Zhou, C., Duan, P., Tang, Y., Xu, C., Xu, C., Huang, T., Shi, B.: Neuromorphic camera guided high dynamic range imaging. In: CVPR. pp. 1730-1739 (2020) +12. Hasinoff, S.W., Sharlet, D., Geiss, R., Adams, A., Barron, J.T., Kainz, F., Chen, J., Levoy, M.: Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM TOG 35(6), 1-12 (2016) +13. Heide, F., Rouf, M., Hullin, M.B., Labitzke, B., Heidrich, W., Kolb, A.: High-quality computational imaging through simple lenses. ACM TOG 32(5), 1-14 (2013) +14. Huang, T., Zheng, Y., Yu, Z., Chen, R., Li, Y., Xiong, R., Ma, L., Zhao, J., Dong, S., Zhu, L., Li, J., Jia, S., Fu, Y., Shi, B., Wu, S., Tian, Y.: $1000 \times$ faster camera and machine vision with ordinary devices. Engineering 25, 110-119 (2022) +15. Kiku, D., Monno, Y., Tanaka, M., Okutomi, M.: Beyond color difference: Residual interpolation for color image demosaicking. IEEE TIP 25(3), 1288-1300 (2016) +16. Kokkinos, F., Lefkimmiatis, S.: Deep image demosaicking using a cascade of convolutional residual denoising networks. In: ECCV. pp. 303-319 (2018) +17. Lichtsteiner, P., Posch, C., Delbruck, T.: A $128 \times 128$ 120 db $15\mu s$ latency asynchronous temporal contrast vision sensor. IEEE JSSC 43(2), 566-576 (2008) + +18. Liu, L., Jia, X., Liu, J., Tian, Q.: Joint demosaicing and denoising with self guidance. In: CVPR. pp. 2240-2249 (2020) +19. Liu, Y., Gutierrez-Barragan, F., Ingle, A., Gupta, M., Velten, A.: Single-photon camera guided extreme dynamic range imaging. In: WACV. pp. 1575-1585 (2022) +20. Ma, S., Gupta, S., Ulku, A.C., Bruschini, C., Charbon, E., Gupta, M.: Quanta burst photography. ACM TOG 39(4), 79-1 (2020) +21. Ma, S., Sundar, V., Mos, P., Bruschini, C., Charbon, E., Gupta, M.: Seeing photons in color. ACM TOG 42(4), 1-16 (2023) +22. Mairal, J., Bach, F., Ponce, J., Sapiro, G., Zisserman, A.: Non-local sparse models for image restoration. In: ICCV. pp. 2272-2279 (2009) +23. Mairal, J., Elad, M., Sapiro, G.: Sparse representation for color image restoration. IEEE TIP 17(1), 53-69 (2007) +24. Malvar, H.S., He, L.w., Cutler, R.: High-quality linear interpolation for demosaicing of bayer-patterned color images. In: ICASSP. vol. 3, pp. iii-485 (2004) +25. Mantiuk, R.K., Hammou, D., Hanji, P.: HDR-VDP-3: A multi-metric for predicting image differences, quality and contrast distortions in high dynamic range and regular content. arXiv preprint arXiv:2304.13625 (2023) +26. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: CVPR. pp. 3883-3891 (2017) +27. Narwaria, M., Da Silva, M.P., Le Callet, P.: HDR-VQM: An objective quality measure for high dynamic range video. Signal Processing: Image Communication 35, 46-60 (2015) +28. Pekkucuksen, I., Altunbasak, Y.: Gradient based threshold free color filter array interpolation. In: ICIP. pp. 137-140 (2010) +29. Tan, D.S., Chen, W.Y., Hua, K.L.: Deepdemosaicking: Adaptive image demosaicking via multiple deep fully convolutional networks. IEEE TIP 27(5), 2408-2419 (2018) +30. Tan, R., Zhang, K., Zuo, W., Zhang, L.: Color image demosaicking via deep residual learning. In: ICME. pp. 793-798 (2017) +31. Wang, Z., Liu, J., Li, G., Han, H.: Blind2unblind: Self-supervised image denoising with visible blind spots. In: CVPR. pp. 2027-2036 (2022) +32. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: From error visibility to structural similarity. IEEE TIP 13(4), 600-612 (2004) +33. Xia, L., Zhao, J., Xiong, R., Huang, T.: SVFI: Spiking-based video frame interpolation for high-speed motion. In: AAAI. vol. 37, pp. 2910-2918 (2023) +34. Xing, W., Egiazarian, K.: End-to-end learning for joint image demosaicing, denoising and super-resolution. In: CVPR. pp. 3507-3516 (2021) +35. Yang, Y., Han, J., Liang, J., Sato, I., Shi, B.: Learning event guided high dynamic range video reconstruction. In: CVPR. pp. 13924-13934 (2023) +36. Yu, G., Sapiro, G., Mallat, S.: Solving inverse problems with piecewise linear estimators: From gaussian mixture models to structured sparsity. IEEE TIP 21(5), 2481-2499 (2011) +37. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR. pp. 586-595 (2018) +38. Zhang, Z., Zheng, H., Hong, R., Xu, M., Yan, S., Wang, M.: Deep color consistent network for low-light image enhancement. In: CVPR. pp. 1899-1908 (2022) +39. Zhao, J., Xiong, R., Liu, H., Zhang, J., Huang, T.: Spk2ImgNet: Learning to reconstruct dynamic scene from continuous spike stream. In: CVPR. pp. 11996-12005 (2021) + +40. Zhao, J., Xiong, R., Xie, J., Shi, B., Yu, Z., Gao, W., Huang, T.: Reconstructing clear image for high-speed motion scene with a retina-inspired spike camera. IEEE TCI 8, 12-27 (2021) +41. Zhao, J., Zhang, S., Ma, L., Yu, Z., Huang, T.: SpikingSIM: A bio-inspired spiking simulator. In: IEEE ISCAS. pp. 3003-3007 (2022) +42. Zheng, Y., Zhang, J., Zhao, R., Ding, J., Chen, S., Xiong, R., Yu, Z., Huang, T.: SpikeCV: Open a continuous computer vision era. arXiv preprint arXiv:2303.11684 (2023) +43. Zheng, Y., Zheng, L., Yu, Z., Huang, T., Wang, S.: Capture the moment: High-speed imaging with spiking cameras through short-term plasticity. IEEE TPAMI 45(7), 8127-8142 (2023) +44. Zheng, Y., Zheng, L., Yu, Z., Shi, B., Tian, Y., Huang, T.: High-speed image reconstruction through short-term plasticity for spiking cameras. In: CVPR. pp. 6358-6367 (2021) +45. Zhu, L., Dong, S., Huang, T., Tian, Y.: A retina-inspired sampling method for visual texture reconstruction. In: ICME. pp. 1432-1437 (2019) +46. Zhu, L., Dong, S., Li, J., Huang, T., Tian, Y.: Retina-like visual image reconstruction via spiking neural model. In: CVPR. pp. 1438-1446 (2020) \ No newline at end of file diff --git a/2024/Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes/images.zip b/2024/Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..84bce8703a911984b0452da85b6600d1b3c7ea9a --- /dev/null +++ b/2024/Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:697c3e236b8de2145d2aa5bf84dfff0cde4776ce418a5339c99f0c109f86eee3 +size 601468 diff --git a/2024/Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes/layout.json b/2024/Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9a50cc4d824da55d39f64b06fe06a3b679620d20 --- /dev/null +++ b/2024/Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes/layout.json @@ -0,0 +1,9188 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 168, + 111, + 446, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 111, + 446, + 148 + ], + "spans": [ + { + "bbox": [ + 168, + 111, + 446, + 148 + ], + "type": "text", + "content": "Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 168, + 167, + 445, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 167, + 445, + 193 + ], + "spans": [ + { + "bbox": [ + 168, + 167, + 445, + 193 + ], + "type": "text", + "content": "Siqi Yang" + }, + { + "bbox": [ + 168, + 167, + 445, + 193 + ], + "type": "inline_equation", + "content": "^{1,\\#}" + }, + { + "bbox": [ + 168, + 167, + 445, + 193 + ], + "type": "text", + "content": ", Zhaojun Huang" + }, + { + "bbox": [ + 168, + 167, + 445, + 193 + ], + "type": "inline_equation", + "content": "^{2,3,\\#}" + }, + { + "bbox": [ + 168, + 167, + 445, + 193 + ], + "type": "text", + "content": ", Yakun Chang" + }, + { + "bbox": [ + 168, + 167, + 445, + 193 + ], + "type": "inline_equation", + "content": "^{5,6}" + }, + { + "bbox": [ + 168, + 167, + 445, + 193 + ], + "type": "text", + "content": ", Bin Fan" + }, + { + "bbox": [ + 168, + 167, + 445, + 193 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 168, + 167, + 445, + 193 + ], + "type": "text", + "content": ", Zhaofei Yu" + }, + { + "bbox": [ + 168, + 167, + 445, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 168, + 167, + 445, + 193 + ], + "type": "text", + "content": ", and Boxin Shi" + }, + { + "bbox": [ + 168, + 167, + 445, + 193 + ], + "type": "inline_equation", + "content": "^{2,3,1,*}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 208, + 201, + 405, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 201, + 405, + 213 + ], + "spans": [ + { + "bbox": [ + 208, + 201, + 405, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 208, + 201, + 405, + 213 + ], + "type": "text", + "content": " Institute for Artificial Intelligence, Peking University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 212, + 476, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 212, + 476, + 220 + ], + "spans": [ + { + "bbox": [ + 136, + 212, + 476, + 220 + ], + "type": "text", + "content": "2 State Key Lab of Multimedia Info. Processing, School of Computer Science, Peking University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 220, + 476, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 220, + 476, + 228 + ], + "spans": [ + { + "bbox": [ + 136, + 220, + 476, + 228 + ], + "type": "text", + "content": "3 Nat'l Eng. Research Ctr. of Visual Technology, School of Computer Science, Peking University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 228, + 476, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 228, + 476, + 236 + ], + "spans": [ + { + "bbox": [ + 136, + 228, + 476, + 236 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 136, + 228, + 476, + 236 + ], + "type": "text", + "content": " Nat'l Key Lab of General AI, School of Intelligence Science and Technology, Peking University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 197, + 236, + 417, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 236, + 417, + 244 + ], + "spans": [ + { + "bbox": [ + 197, + 236, + 417, + 244 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 197, + 236, + 417, + 244 + ], + "type": "text", + "content": " Institute of Information Science, Beijing Jiaotong University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 243, + 479, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 243, + 479, + 260 + ], + "spans": [ + { + "bbox": [ + 133, + 243, + 479, + 260 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 133, + 243, + 479, + 260 + ], + "type": "text", + "content": " Visual Intelligence +X International Cooperation Joint Laboratory of the Ministry of Education {yousiki, huangzhaojun, binfan, yuzf12, shiboxin}@pku.edu.cn, ykchang@bjtu.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 289, + 452, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 289, + 452, + 520 + ], + "spans": [ + { + "bbox": [ + 159, + 289, + 452, + 520 + ], + "type": "text", + "content": "Abstract. The spike camera continuously records scene radiance with high-speed, high dynamic range, and low data redundancy properties, as a promising replacement for frame-based high-speed cameras. Previous methods for reconstructing color videos from monochromatic spikes are constrained in capturing full-temporal color information due to their reliance on compensating colors from low-speed RGB frames. Applying a Bayer-pattern color filter array to the spike sensor yields mosaicked chromatic spikes, which complicates noise distribution in high-speed conditions. By validating that the noise of short-term frames follows a zero-mean distribution, we leverage this hypothesis to develop a self-supervised denoising module trained exclusively on real-world data. Although noise is reduced in short-term frames, the long-term accumulation of incident photons is still necessary to construct HDR frames. Therefore, we introduce a progressive warping module to generate pseudo long-term exposure frames. This approach effectively mitigates motion blur artifacts in high-speed conditions. Integrating these modules forms a real-data-driven reconstruction method for mosaicked chromatic spikes. Extensive experiments conducted on both synthetic and real-world data demonstrate that our approach is effective in reconstructing 2000FPS color HDR videos with significantly reduced noise and motion blur compared to existing methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 159, + 531, + 453, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 531, + 453, + 553 + ], + "spans": [ + { + "bbox": [ + 159, + 531, + 453, + 553 + ], + "type": "text", + "content": "Keywords: Spike camera " + }, + { + "bbox": [ + 159, + 531, + 453, + 553 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 531, + 453, + 553 + ], + "type": "text", + "content": " High-speed video " + }, + { + "bbox": [ + 159, + 531, + 453, + 553 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 531, + 453, + 553 + ], + "type": "text", + "content": " High dynamic range video " + }, + { + "bbox": [ + 159, + 531, + 453, + 553 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 531, + 453, + 553 + ], + "type": "text", + "content": " Color demosaicking " + }, + { + "bbox": [ + 159, + 531, + 453, + 553 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 531, + 453, + 553 + ], + "type": "text", + "content": " Self-supervised learning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 574, + 230, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 230, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 230, + 586 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": "As a new class of image sensors inspired by the fovea of retina, the spike camera [3, 14, 46] offers attractive characteristics such as high temporal resolution (20000Hz), high dynamic range (HDR, " + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "inline_equation", + "content": ">90\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": "), and low data redundancy [14]. In comparison with conventional digital cameras that capture discrete frames," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 132, + 653, + 328, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 653, + 328, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 653, + 328, + 665 + ], + "type": "text", + "content": "# Equal contributions. * Corresponding author." + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 115, + 479, + 264 + ], + "blocks": [ + { + "bbox": [ + 138, + 115, + 479, + 264 + ], + "lines": [ + { + "bbox": [ + 138, + 115, + 479, + 264 + ], + "spans": [ + { + "bbox": [ + 138, + 115, + 479, + 264 + ], + "type": "image", + "image_path": "088473a1bf3a14e9fb809fe8332d9ac56789bb93302929ef88d113bb42985e0f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 264, + 372, + 274 + ], + "lines": [ + { + "bbox": [ + 255, + 264, + 372, + 274 + ], + "spans": [ + { + "bbox": [ + 255, + 264, + 372, + 274 + ], + "type": "text", + "content": "(b) High-speed color HDR video" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 285, + 482, + 350 + ], + "lines": [ + { + "bbox": [ + 131, + 285, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 131, + 285, + 482, + 350 + ], + "type": "text", + "content": "Fig. 1: High-speed color HDR video reconstruction from mosaicked chromatic spikes. (a) Comparison of different chromatic spike reconstruction (recon.) methodologies, including the adaptation of bio-inspired [45] and supervised learning [39] monochromatic methods to chromatic spikes and our proposed method. (b) A high-quality color HDR video with a frame rate of 2000FPS reconstructed by our method. The corresponding video is available in the supplementary material." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "text", + "content": "spike cameras continuously accumulate the photons and trigger a single-bit spike of 1 when the detected photons reach a predefined threshold [14]. High-speed readout of spikes provides the spike camera with the potential for photon counting, and the high-dynamic-range ambient light can be estimated through continuous accumulation of spikes in the temporal domain. Therefore, spike cameras exhibit significant advantages in handling low-level vision tasks in high-speed conditions, such as video frame interpolation [33], HDR imaging [11] and high-frame-rate HDR video reconstruction [3]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "content": "With the higher demand for visual quality in images and videos, there is a strong desire that images or videos reconstructed from spike cameras can reproduce the real world to the greatest extent, with realistic color appearance and extended dynamic range at high frame rate. However, existing approaches [39, 40, 43, 44, 46] mainly focus on image or video reconstruction with monochromatic spikes, and cannot yet recover a high-speed color HDR video. To introduce color, one alternative practice [11] is to build a hybrid spike-RGB camera system, which exploits the colors captured by the conventional RGB camera to colorize the image reconstructed from spikes. Unfortunately, the color information provided by conventional RGB cameras often suffers from underexposure and overexposure, which inevitably lead to color deviations from the real world. Another strategy [3] to alleviate this problem is to replace the conventional RGB camera with an alternate exposure RGB camera to obtain HDR color, thereby enabling high-speed HDR color video reconstruction. However, the above methods have limitations in high-speed scenes, as the RGB camera has a low frame rate (e.g., 60FPS), which results in the loss of colors during the" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "S. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "intervals between frames. In addition, the corresponding hybrid camera system relies on complex temporal synchronization and optical alignment, and the beam splitter with large space footprints would be detrimental to the construction of compact devices." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 167, + 482, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 167, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 167, + 482, + 370 + ], + "type": "text", + "content": "Reconstructing high-speed color HDR videos based only on a single spike camera not only facilitates the acquisition of full-temporal color information but also offers the advantage of lightweight equipment. To this end, a tractable solution is to assemble a Bayer-pattern color filter array (CFA) in the spike camera, thereby producing mosaicked chromatic spikes. Consequently, a pressing challenge that arises is how to reconstruct high-speed color HDR videos from this new data representation. Nevertheless, directly adapting existing monochromatic spike reconstruction methods [39,40,43,44] to mosaicked chromatic spikes suffers from many issues, as shown in Fig. 1 (a). On the one hand, methods based on biology-inspired models [40,43,44] cannot take advantage of deep neural networks and are susceptible to varying degrees of noise contamination. On the other hand, supervised learning methods (e.g., [39]) trained using synthetic datasets have limited generalization ability when transferring to real-world data, because of the imperfection of spike noise modeling and simulation. These issues hinder the development of high-speed color HDR video reconstruction methods based on chromatic spike cameras and greatly limit the effective deployment of existing methods in real scenes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 375, + 482, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 482, + 554 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 482, + 554 + ], + "type": "text", + "content": "In this paper, we propose a real-data-driven mosaicked chromatic spike reconstruction method to recover high-speed color HDR videos, trained solely on readily accessible unlabelled spike streams. The mathematical analysis behind this requires no specific spike noise modeling or calibration, relying only on insights into the relationship between short-term and long-term frames accumulated from chromatic spikes. We design a deep learning framework that allows training directly on real-world data, naturally avoiding the influence of imperfect spike simulation and generalization gap. Specifically, we first propose a self-supervised denoising module to suppress the noise in short-term accumulation of chromatic spikes, which is trained on real-world spike streams without any annotation. Next, considering the limited dynamic range of short-term accumulated frames, we further present a progressive warping module to simulate a pseudolong exposure, which is effective in reconstructing HDR video frames without motion blur. Finally, we integrate these two modules to achieve higher-quality color HDR video reconstruction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 558, + 482, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 606 + ], + "type": "text", + "content": "In summary, our approach is able to achieve 2000FPS color HDR video reconstruction from mosaicked chromatic spikes with a single device in a real-data-driven manner, as illustrated in Fig. 1 (b), because of the technical contributions from these two complementary aspects:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 137, + 629, + 482, + 665 + ], + "type": "text", + "content": "- We design a self-supervised denoising module to eliminate the noise from short-term accumulation (with high-speed motion cues) without relying on chromatic spike noise modeling and simulation." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 90, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 90, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 141, + 90, + 445, + 102 + ], + "type": "text", + "content": "Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 482, + 152 + ], + "type": "text", + "content": "- We propose a progressive warping module with pseudo long-term exposure (with expanded dynamic range) to reconstruct HDR frames without introducing motion blur in dynamic scenes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 171, + 237, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 237, + 185 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 237, + 185 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 198, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 198, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 482, + 426 + ], + "type": "text", + "content": "Video reconstruction with monochromatic spikes. To better satisfy human visual perception and facilitate direct processing by machines, reconstructing the corresponding video frame from spikes has been widely studied. Prevalent image reconstruction methods typically employ the temporal statistical properties of spike cameras [43-46]. Zhu et al. [45] explored the spike generation principle and proposed estimating the firing frequency or firing interval of each pixel to reconstruct dynamic scenes, which are known as texture from playback (TFP) and texture from inter-spike-intervals (TFI), respectively. Recently, another line of work has enhanced firing frequency estimates by mimicking human physiological mechanisms, such as retina-like visual imaging [46] and short-term plasticity [43, 44]. Although these methods bring better interpretability, they require careful window selection to balance noise and motion blur, and the reconstruction effect is not yet satisfactory. The powerful representation capabilities of deep learning also inject vitality into spike-based image reconstruction. Spk2ImgNet [39] is the first spike-to-image deep network architecture based on deformable convolution, achieving impressive results. Further, several self-supervised methods [5,6] have also been developed to reduce the dependence on synthetic datasets. However, these methods are tailored for monochromatic spikes and thus cannot be directly applied to mosaicked chromatic spikes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 426, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 557 + ], + "type": "text", + "content": "Demosaicking and denoising with raw images. The purpose of raw image demosaicking is to recover a full-color image from a sub-sampled mosaicked image that contains potential noise. To deal with the ill-posedness of image demosaicking, various image priors have been introduced to regularize the recovery process, such as heuristics [2, 24], sparsity [23, 36], total variation [7, 13], integral gradient [28], self-similarity [22], and residual interpolation [15]. In addition, researchers designed deep neural networks [16, 29, 30] to automatically learn the prior knowledge required for image demosaicking. Moreover, considering the coupling between demosaicking and denoising, some deep learning-based joint demosaicking and denoising network models [9, 18, 34, 38] have also been developed to efficiently remove mosaicks from noisy raw images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "HDR video with non-conventional sensors. In recent years, a series of nonconventional sensors, such as event cameras [17], spike cameras [14], and quanta image sensors (QIS) [8], have been invented, empowering the ability to capture HDR information in high-speed motion scenes. QIS-based HDR video reconstruction methods [10, 19, 20] were developed to accomplish perceptual HDR illumination. Additionally, the intensity image reconstructed by event cameras [35] or spike cameras [11] can serve as a proxy for enhancing details in low dynamic range (LDR) images. More recently, a hybrid camera system consisting of a spike camera and an alternating-exposure RGB camera is constructed in [3] for " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "1000 \\times" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "S. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 147, + 118, + 475, + 219 + ], + "blocks": [ + { + "bbox": [ + 147, + 118, + 475, + 219 + ], + "lines": [ + { + "bbox": [ + 147, + 118, + 475, + 219 + ], + "spans": [ + { + "bbox": [ + 147, + 118, + 475, + 219 + ], + "type": "image", + "image_path": "5645962b9ba6da6cceab1ca074ce59a65902ea8bcdfc7b594e05238c6175edf4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 228, + 482, + 285 + ], + "lines": [ + { + "bbox": [ + 131, + 228, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 131, + 228, + 482, + 285 + ], + "type": "text", + "content": "Fig. 2: Chromatic spike accumulation and firing mechanism. We put a Bayer pattern CFA in the spike camera to filter incident light before the photons arrive at the sensor. Each individual pixel corresponds to one of red, green, and blue channels, accumulates the filtered photons, and triggers a spike signal, then resets when it surpasses the threshold " + }, + { + "bbox": [ + 131, + 228, + 482, + 285 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{th}}" + }, + { + "bbox": [ + 131, + 228, + 482, + 285 + ], + "type": "text", + "content": ". Reading out the chromatic spikes forms the chromatic spike planes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 306, + 479, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 306, + 479, + 331 + ], + "spans": [ + { + "bbox": [ + 131, + 306, + 479, + 331 + ], + "type": "text", + "content": "frame rate HDR video reconstruction. Nevertheless, these methods also fall short when dealing with mosaicked chromatic spikes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 346, + 233, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 346, + 233, + 360 + ], + "spans": [ + { + "bbox": [ + 132, + 346, + 233, + 360 + ], + "type": "text", + "content": "3 Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 370, + 479, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 370, + 479, + 418 + ], + "spans": [ + { + "bbox": [ + 131, + 370, + 479, + 418 + ], + "type": "text", + "content": "We initially present a concise review of the operational principles of spike cameras, encompassing both monochromatic spike cameras and those equipped with Color Filter Array (CFA). Subsequently, we discuss the baseline reconstruction techniques for monochromatic spikes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 418, + 480, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 418, + 480, + 466 + ], + "spans": [ + { + "bbox": [ + 131, + 418, + 480, + 466 + ], + "type": "text", + "content": "Monochromatic spikes. The spike camera persistently accumulates electrons produced by incoming photons. When the electrons accumulated at a pixel surpass the predefined threshold " + }, + { + "bbox": [ + 131, + 418, + 480, + 466 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{th}}" + }, + { + "bbox": [ + 131, + 418, + 480, + 466 + ], + "type": "text", + "content": ", it triggers a spike signal, immediately resetting the accumulator to initiate a new cycle of accumulation:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 227, + 474, + 481, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 474, + 481, + 499 + ], + "spans": [ + { + "bbox": [ + 227, + 474, + 481, + 499 + ], + "type": "interline_equation", + "content": "E (\\mathbf {x}, t) = \\int \\alpha L (\\mathbf {x}, t) d t \\mod E _ {\\mathrm {t h}}, \\tag {1}", + "image_path": "d325a9474300add6d9def3c1f3be187a7683e06b425866af57d6539760a68a78.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 502, + 479, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 502, + 479, + 537 + ], + "spans": [ + { + "bbox": [ + 131, + 502, + 479, + 537 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 502, + 479, + 537 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 502, + 479, + 537 + ], + "type": "text", + "content": " denotes the conversion ratio and " + }, + { + "bbox": [ + 131, + 502, + 479, + 537 + ], + "type": "inline_equation", + "content": "L(\\mathbf{x},t)" + }, + { + "bbox": [ + 131, + 502, + 479, + 537 + ], + "type": "text", + "content": " denotes the incoming light intensity of pixel " + }, + { + "bbox": [ + 131, + 502, + 479, + 537 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 131, + 502, + 479, + 537 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 131, + 502, + 479, + 537 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 502, + 479, + 537 + ], + "type": "text", + "content": ". Consequently, the spike signal " + }, + { + "bbox": [ + 131, + 502, + 479, + 537 + ], + "type": "inline_equation", + "content": "S(\\mathbf{x},t)" + }, + { + "bbox": [ + 131, + 502, + 479, + 537 + ], + "type": "text", + "content": " can be defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 235, + 546, + 481, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 546, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 235, + 546, + 481, + 578 + ], + "type": "interline_equation", + "content": "S (\\mathbf {x}, t) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} E (\\mathbf {x}, t) \\geq E _ {\\mathrm {t h}}, \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {2}", + "image_path": "dfb7f6de87b0588397aee1a5c1bd6bc40f76ad5d6e6048520500a09a96f294ff.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 582, + 479, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 582, + 479, + 640 + ], + "spans": [ + { + "bbox": [ + 131, + 582, + 479, + 640 + ], + "type": "text", + "content": "As all the pixels of a spike sensor are arranged in an array, reading out spikes from the pixel array at each time stamp " + }, + { + "bbox": [ + 131, + 582, + 479, + 640 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 582, + 479, + 640 + ], + "type": "text", + "content": " yields a spike plane of dimensions " + }, + { + "bbox": [ + 131, + 582, + 479, + 640 + ], + "type": "inline_equation", + "content": "h \\times w" + }, + { + "bbox": [ + 131, + 582, + 479, + 640 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 582, + 479, + 640 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 131, + 582, + 479, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 582, + 479, + 640 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 131, + 582, + 479, + 640 + ], + "type": "text", + "content": " are the height and width of the spike data, respectively. Given the readout frequency of 20000Hz, the spike camera [14] allows for capturing spike planes at a frame rate of 20000FPS." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": "Chromatic spikes. To incorporate color information while capturing HDR scenes with high-speed motion at the same temporal resolution, we can put a" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 90, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 90, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 141, + 90, + 445, + 102 + ], + "type": "text", + "content": "Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "CFA in front of the sensor of the spike camera. In a similar way as conventional cameras, the CFA applied to spike sensors presents a Bayer pattern. Each pixel of the spike camera sensor with CFA corresponds to only one channel among red, green, and blue, while the firing mechanism of spikes remains unchanged. As illustrated in Fig. 2, under the influence of the CFA, a spike signal is generated when the cumulative electrons " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "E_{c}(\\mathbf{x},t)" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " produced by the incident photons of the corresponding spectral band meets the threshold " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{th}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": ". The chromatic spike signal " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "S_{c}(\\mathbf{x},t)" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " can be recorded as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 233, + 221, + 481, + 252 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 221, + 481, + 252 + ], + "spans": [ + { + "bbox": [ + 233, + 221, + 481, + 252 + ], + "type": "interline_equation", + "content": "S _ {c} (\\mathbf {x}, t) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} E _ {c} (\\mathbf {x}, t) \\geq E _ {\\mathrm {t h}}, \\\\ 0, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {3}", + "image_path": "142215513703ed62ecfe473ef98e0c3b6518e7dc61fc1ec89e3e298fb7a3cded.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 255, + 480, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 480, + 290 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 480, + 290 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 255, + 480, + 290 + ], + "type": "inline_equation", + "content": "c \\in \\{R, G, B\\}" + }, + { + "bbox": [ + 130, + 255, + 480, + 290 + ], + "type": "text", + "content": ". Consequently, similar to the firing process of monochromatic spikes, we can obtain a mosaicked chromatic spike plane at each time stamp " + }, + { + "bbox": [ + 130, + 255, + 480, + 290 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 255, + 480, + 290 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 292, + 481, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 292, + 481, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 481, + 327 + ], + "type": "text", + "content": "Baseline for chromatic frame reconstruction. Preliminary chromatic frame reconstruction can be achieved similarly to [45] by accumulating the spike signals within a defined temporal window:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 236, + 337, + 481, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 337, + 481, + 372 + ], + "spans": [ + { + "bbox": [ + 236, + 337, + 481, + 372 + ], + "type": "interline_equation", + "content": "I _ {c} (\\mathbf {x}, t) = \\frac {M _ {c}}{W} \\cdot \\sum_ {\\tau = t - W / 2} ^ {t + W / 2} S _ {c} (\\mathbf {x}, \\tau), \\tag {4}", + "image_path": "59926aad24e622fa844b8697ecd4a6472eea308fbf593bb7a58b0d5b69c7e0b0.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "text", + "content": " is the size of the accumulation temporal window, " + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "inline_equation", + "content": "M_{c}" + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "text", + "content": " represents the maximum dynamic range that can be perceived in channel " + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "text", + "content": ". Directly applying Eq. (4) to chromatic spikes as a baseline reconstruction algorithm can only obtain mosaicked chromatic frames, which is similar to the RAW image of conventional frame-based sensors. Demosaicking the chromatic frame is challenging since there is a trade-off between noise and superposition of colors. A short-term temporal window " + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "text", + "content": " results in severe noise, whereas a long-term " + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "text", + "content": " leads to the spatial superposition of varying color information. Furthermore, both noise distribution of chromatic spikes and object motions are more complex, making the direct application of off-the-shelf demosaicking methods insufficient for recovering high-quality color HDR videos." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 522, + 408, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 522, + 408, + 537 + ], + "spans": [ + { + "bbox": [ + 131, + 522, + 408, + 537 + ], + "type": "text", + "content": "4 Real-data-driven chromatic spikes to video" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Overview. As shown in Fig. 3, we propose a progressive mechanism for high-speed color HDR video reconstruction. The pipeline starts with the accumulation of mosaicked chromatic spikes to reconstruct short-term frames. While these short-term frames are free of motion blur, they suffer from significant noise contamination. Consequently, as shown in Fig. 3 (a), we propose the chromatic spike denoising module (Sec. 4.1) to suppress the noise. The key insight of the denoising module is that a sufficiently long-term frame (e.g., " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "W = 10000" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ", corresponding to 0.5s) of chromatic spikes can produce promising reconstructions in static (or low-speed) scenes, effectively representing the expectations of short-term frames. Given the limited incident photons in short-term frames, which" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "S. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 113, + 482, + 337 + ], + "blocks": [ + { + "bbox": [ + 134, + 113, + 482, + 337 + ], + "lines": [ + { + "bbox": [ + 134, + 113, + 482, + 337 + ], + "spans": [ + { + "bbox": [ + 134, + 113, + 482, + 337 + ], + "type": "image", + "image_path": "4110ea8fc7ba3608e4170bfb8130d020434bf2380736740d9b794fd46be56169.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 343, + 482, + 443 + ], + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 443 + ], + "type": "text", + "content": "Fig. 3: Illustration of the overall pipeline. For each sequence of mosaicked chromatic spikes, we split them into four channels according to the RGGB pattern. The four-channel spikes are fed to our progressive warping module, which aligns multiple short-term frames to achieve \"pseudo-long exposure\". (a) Our chromatic spike denoising module utilizes a self-supervised training strategy, which is pretrained on real-world data. (b) Our progressive warping module consists of " + }, + { + "bbox": [ + 130, + 343, + 482, + 443 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 343, + 482, + 443 + ], + "type": "text", + "content": " steps of frame warping operations, iteratively achieving high-quality HDR frames. At each step of the progressive warping module, we leverage the pretrained chromatic spikes denoising module to facilitate optical flow estimation. " + }, + { + "bbox": [ + 130, + 343, + 482, + 443 + ], + "type": "inline_equation", + "content": "R_{L}" + }, + { + "bbox": [ + 130, + 343, + 482, + 443 + ], + "type": "text", + "content": " is the final output color HDR video with 2000FPS." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 467, + 482, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 467, + 482, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 467, + 482, + 575 + ], + "type": "text", + "content": "lack the HDR characteristic, extending the accumulation time is a feasible strategy for HDR reconstruction. As shown in Fig. 3 (b), we design the progressive warping module in Sec. 4.2 to establish a pseudo long exposure. At each step of the warping module, chromatic spikes frames are warped according to the estimated optical flow, providing pseudo long-term frames with less noise and higher dynamic range. To leverage a wide range of existing conventional RGB optical flow estimation approaches, we integrate the denoising module into each iteration step of the warping module. After " + }, + { + "bbox": [ + 130, + 467, + 482, + 575 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 467, + 482, + 575 + ], + "type": "text", + "content": " times of iteration, the progressive warping module produces an HDR color video with a frame rate of 2000FPS." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 595, + 310, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 595, + 310, + 608 + ], + "spans": [ + { + "bbox": [ + 132, + 595, + 310, + 608 + ], + "type": "text", + "content": "4.1 Denoising of chromatic spikes" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Noise distribution of chromatic spikes noise. The noise in monochromatic spikes is primarily categorized into two types: diffuse noise and inherent noise. For the mosaicked chromatic spikes, although the incorporation of a CFA slightly alters the noise distribution for pixels under different filters, pixels corresponding" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 90, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 90, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 141, + 90, + 445, + 102 + ], + "type": "text", + "content": "Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "to the same color exhibit similar noise distribution patterns as monochromatic spikes, except for its expectation and variance. Given the chromatic spikes " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "S_{c}" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": ", for each " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "c \\in \\{R,G,B\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": ", we use " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "Q_{d}" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " to denote the number of diffuse noise photons arriving at a pixel in a time period of " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " to denote the likelihood of " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "Q_{d}" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " diffuse noise photons. The probability of diffuse noise can be modeled as the Poisson distribution [41]:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 236, + 198, + 481, + 225 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 198, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 236, + 198, + 481, + 225 + ], + "type": "interline_equation", + "content": "\\mathcal {P} \\left(Q _ {d} | t, \\gamma\\right) = \\frac {\\left(\\gamma t\\right) ^ {Q _ {d}}}{Q _ {d} !} \\exp (- \\gamma W). \\tag {5}", + "image_path": "f6b365cbbfba79586e828b2d9a30b5538dd355927eaa1d47bc39b96b581bb568.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 229, + 472, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 472, + 242 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 472, + 242 + ], + "type": "text", + "content": "Hence, the expected number of diffuse noise spikes in temporal window " + }, + { + "bbox": [ + 132, + 229, + 472, + 242 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 132, + 229, + 472, + 242 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 261, + 255, + 481, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 255, + 481, + 269 + ], + "spans": [ + { + "bbox": [ + 261, + 255, + 481, + 269 + ], + "type": "interline_equation", + "content": "\\mathbb {E} \\left[ N _ {d} \\right] = (\\alpha \\gamma W) / E _ {\\mathrm {t h}}. \\tag {6}", + "image_path": "b99ba8a94ca231b2bee867607f2159a6b529a6fb39514dfbabc73197925bd0ca.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 274, + 482, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 274, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 131, + 274, + 482, + 312 + ], + "type": "text", + "content": "For inherent noise, which triggers spikes in the absence of light, we use " + }, + { + "bbox": [ + 131, + 274, + 482, + 312 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 131, + 274, + 482, + 312 + ], + "type": "text", + "content": " to indicate the interval between the noise spikes. Then, the interval between such noise spikes can be represented by a Gaussian model [41]:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 214, + 321, + 481, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 321, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 214, + 321, + 481, + 346 + ], + "type": "interline_equation", + "content": "\\mathcal {N} (T _ {i} | \\mu , \\sigma) = \\frac {1}{\\sigma \\sqrt {2 \\pi}} \\cdot \\exp \\left(- \\frac {1}{2} \\cdot \\left(\\frac {T _ {i} - \\mu}{\\sigma}\\right) ^ {2}\\right). \\tag {7}", + "image_path": "3a03c04f574a8cbcd776def7276a7e86b3fa567a10b8eb18274c13b44cbd3f8d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 350, + 482, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 350, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 131, + 350, + 482, + 374 + ], + "type": "text", + "content": "We approximate the expectation of the number of inherent noise spikes in the temporal window as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 179, + 386, + 481, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 386, + 481, + 412 + ], + "spans": [ + { + "bbox": [ + 179, + 386, + 481, + 412 + ], + "type": "interline_equation", + "content": "\\mathbb {E} [ N _ {i} ] = \\int_ {1} ^ {+ \\infty} \\frac {W}{T _ {i}} \\cdot \\frac {1}{\\sigma \\sqrt {2 \\pi}} \\cdot \\exp (- \\frac {1}{2} \\cdot (\\frac {T _ {i} - \\mu}{\\sigma}) ^ {2}) d T _ {i} = C _ {i} W, (8)", + "image_path": "05b7603515a3fea25d9a51a2fc98a9c2e74f354c316cedf9c388634b68d3eb5e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 417, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 417, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 131, + 417, + 482, + 453 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 417, + 482, + 453 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 131, + 417, + 482, + 453 + ], + "type": "text", + "content": " denotes the integral after factoring out " + }, + { + "bbox": [ + 131, + 417, + 482, + 453 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 131, + 417, + 482, + 453 + ], + "type": "text", + "content": " from the integral expression. Combining Eq. (6) and Eq. (8), the expectation of the fraction of noise spikes can be formulated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 204, + 467, + 481, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 467, + 481, + 480 + ], + "spans": [ + { + "bbox": [ + 204, + 467, + 481, + 480 + ], + "type": "interline_equation", + "content": "\\mathbb {E} [ N ] / W = \\left(\\mathbb {E} \\left[ N _ {d} \\right] + \\mathbb {E} \\left[ N _ {i} \\right]\\right) / W = \\alpha \\gamma / E _ {\\mathrm {t h}} + C _ {i}. \\tag {9}", + "image_path": "d15eb67b5479888176d76ec869e9feb4bf162e78f3580707f50b89f582ed9d35.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "Instead of the actual incoming light intensity, we estimate the expectation of short-term accumulations " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[I_c(\\mathbf{x},t)] = L(\\mathbf{x},t) + \\mathbb{E}[N] / W" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ". Thus, the noise between short-term frames and their expectations naturally follows a zero-mean distribution. We demonstrate that the expected fraction of noise spikes is relatively small, especially with sufficient light illumination, and doesn't disturb the visual semantic. As shown in Fig. 4 (a), the noise ratio " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "(I_{c}(\\mathbf{x},t) / \\mathbb{E}[I_{c}(\\mathbf{x},t)] - 1)" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " in the accumulated images " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "(I_{c}(\\mathbf{x},t))" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " gradually decreases as the accumulation window size " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " increases. Note that the larger window accumulation is equivalent to the average of multiple smaller windows. The convergence demonstrates the existence of expectation and the result faithfully reflects the original light intensity with negligible disturbance. We further validate the zero-mean distribution hypothesis by collecting a series of static scene spike streams with varying light illuminations, and visualize the distribution of noise ratio in Fig. 4 (b). The result shows that the zero-mean assumption holds with sufficient illumination, which supports our self-supervised learning mechanism." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "S. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 128, + 353, + 178 + ], + "blocks": [ + { + "bbox": [ + 133, + 128, + 353, + 178 + ], + "lines": [ + { + "bbox": [ + 133, + 128, + 353, + 178 + ], + "spans": [ + { + "bbox": [ + 133, + 128, + 353, + 178 + ], + "type": "image", + "image_path": "873a2cf93fcb65972ef4faeba1ff17c2e1010b824804e1a25a48be79ee65a821.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 201, + 182, + 290, + 191 + ], + "lines": [ + { + "bbox": [ + 201, + 182, + 290, + 191 + ], + "spans": [ + { + "bbox": [ + 201, + 182, + 290, + 191 + ], + "type": "text", + "content": "(a) Temporal accumulation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 354, + 114, + 480, + 181 + ], + "blocks": [ + { + "bbox": [ + 354, + 114, + 480, + 181 + ], + "lines": [ + { + "bbox": [ + 354, + 114, + 480, + 181 + ], + "spans": [ + { + "bbox": [ + 354, + 114, + 480, + 181 + ], + "type": "image", + "image_path": "d30b55fc59d7f7a9e81a99da41cc6da9cc45b46517beb72d70a1bd63f0508782.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 182, + 457, + 190 + ], + "lines": [ + { + "bbox": [ + 388, + 182, + 457, + 190 + ], + "spans": [ + { + "bbox": [ + 388, + 182, + 457, + 190 + ], + "type": "text", + "content": "(b) Noise distribution" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 200, + 480, + 255 + ], + "lines": [ + { + "bbox": [ + 130, + 200, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 480, + 255 + ], + "type": "text", + "content": "Fig. 4: Validation of noise on real-world chromatic spikes. As shown in (a), as the accumulation window size " + }, + { + "bbox": [ + 130, + 200, + 480, + 255 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 200, + 480, + 255 + ], + "type": "text", + "content": " increases, the noise ratio in the accumulated images gradually decreases, making the images more closely resemble the static scene. In (b), we present the statistical noise fraction distribution across three different color channels and different illumination levels." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 280, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 280, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 482, + 533 + ], + "type": "text", + "content": "Self-supervised chromatic spike denoising module. Taking the inspiration of Batson et al. [1] and Wang et al. [31], which do not require paired noise-contaminated images or clean targets, we design a chromatic spike denoising module trained on only accessible real-world data with self-supervision learning strategy. Specifically, our denoising module operates based on the zero-mean characteristic of short-term frame noise and relies on the blind spot architecture to prevent the network from learning the identity mapping. Hence, the denoising module transforms the noise-contaminated images to their expectations to achieve minimal losses, producing desired clean images. Therefore, as shown in Fig. 3 (a), we begin by extracting the four channels of the chromatic spikes according to the RGGB pattern of the CFA, and the size of each channel is " + }, + { + "bbox": [ + 130, + 280, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\frac{h}{2} \\times \\frac{w}{2}" + }, + { + "bbox": [ + 130, + 280, + 482, + 533 + ], + "type": "text", + "content": ". For each channel " + }, + { + "bbox": [ + 130, + 280, + 482, + 533 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 280, + 482, + 533 + ], + "type": "text", + "content": ", we preliminarily reconstruct a chromatic frame " + }, + { + "bbox": [ + 130, + 280, + 482, + 533 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 130, + 280, + 482, + 533 + ], + "type": "text", + "content": " by accumulating the chromatic spikes in a temporal window " + }, + { + "bbox": [ + 130, + 280, + 482, + 533 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 280, + 482, + 533 + ], + "type": "text", + "content": ". Then, " + }, + { + "bbox": [ + 130, + 280, + 482, + 533 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 130, + 280, + 482, + 533 + ], + "type": "text", + "content": " is further partitioned into " + }, + { + "bbox": [ + 130, + 280, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\frac{h}{2h_s} \\times \\frac{w}{2w_s}" + }, + { + "bbox": [ + 130, + 280, + 482, + 533 + ], + "type": "text", + "content": " cells, where " + }, + { + "bbox": [ + 130, + 280, + 482, + 533 + ], + "type": "inline_equation", + "content": "h_s" + }, + { + "bbox": [ + 130, + 280, + 482, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 280, + 482, + 533 + ], + "type": "inline_equation", + "content": "w_s" + }, + { + "bbox": [ + 130, + 280, + 482, + 533 + ], + "type": "text", + "content": " denote height and width of each cell. We sequentially mask one pixel for each cell following the order from top to bottom and left to right. Consequently, we obtain a masked volume with a size of " + }, + { + "bbox": [ + 130, + 280, + 482, + 533 + ], + "type": "inline_equation", + "content": "C \\times (\\frac{h}{2} \\times \\frac{w}{2})" + }, + { + "bbox": [ + 130, + 280, + 482, + 533 + ], + "type": "text", + "content": " corresponding to " + }, + { + "bbox": [ + 130, + 280, + 482, + 533 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 130, + 280, + 482, + 533 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 280, + 482, + 533 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 280, + 482, + 533 + ], + "type": "text", + "content": " is the channel of the masked volume that equals " + }, + { + "bbox": [ + 130, + 280, + 482, + 533 + ], + "type": "inline_equation", + "content": "h_s \\times w_s" + }, + { + "bbox": [ + 130, + 280, + 482, + 533 + ], + "type": "text", + "content": ". Each channel of the mask volume acts as the blind spot operation to avoid identity mapping. We then employ a U-Net-inspired fully-convolutional neural network " + }, + { + "bbox": [ + 130, + 280, + 482, + 533 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 130, + 280, + 482, + 533 + ], + "type": "text", + "content": " to denoise each masked frame in this volume and merge the denoised frames to the clean image:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 209, + 544, + 480, + 558 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 544, + 480, + 558 + ], + "spans": [ + { + "bbox": [ + 209, + 544, + 480, + 558 + ], + "type": "interline_equation", + "content": "D \\left(I _ {c}\\right) = \\Phi \\left(f _ {\\theta} \\left(\\Omega_ {c, 0}\\right), f _ {\\theta} \\left(\\Omega_ {c, 1}\\right), \\dots , f _ {\\theta} \\left(\\Omega_ {c, C}\\right)\\right), \\tag {10}", + "image_path": "3db83a8ef89b8792562609812d1420bc1abfe9c40f32de43b1b55180d98e5be9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 563, + 479, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 479, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 479, + 588 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 563, + 479, + 588 + ], + "type": "inline_equation", + "content": "\\Omega_{c,i}" + }, + { + "bbox": [ + 130, + 563, + 479, + 588 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 130, + 563, + 479, + 588 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 563, + 479, + 588 + ], + "type": "text", + "content": "-th masked frame in volume " + }, + { + "bbox": [ + 130, + 563, + 479, + 588 + ], + "type": "inline_equation", + "content": "\\Omega_c" + }, + { + "bbox": [ + 130, + 563, + 479, + 588 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 563, + 479, + 588 + ], + "type": "inline_equation", + "content": "\\varPhi(\\cdot)" + }, + { + "bbox": [ + 130, + 563, + 479, + 588 + ], + "type": "text", + "content": " represents a function for mask-aware averaging:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 214, + 598, + 480, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 598, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 214, + 598, + 480, + 622 + ], + "type": "interline_equation", + "content": "\\Phi \\left(f _ {\\theta} \\left(\\Omega_ {c, 0}\\right), \\dots\\right) = \\sum_ {i} M _ {i} f _ {\\theta} \\left(\\Omega_ {c, i}\\right) / \\sum_ {i} M _ {i}, \\tag {11}", + "image_path": "c0667f00a7a746b493c976147f8a63469cdcdd8151e9ae7317a85ffabbb74958.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": " denotes the mask for " + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\varOmega_{c,i}(\\varOmega_{c,i}=(1-M_i)I_c)" + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": ". At last, the denoised four-channel images are recombined to form the final mosaicked chromatic frame. It is remarkable that the zero-mean hypothesis holds for arbitrary temporal window" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 90, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 90, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 141, + 90, + 445, + 102 + ], + "type": "text", + "content": "Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 119, + 479, + 175 + ], + "blocks": [ + { + "bbox": [ + 134, + 119, + 479, + 175 + ], + "lines": [ + { + "bbox": [ + 134, + 119, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 134, + 119, + 479, + 175 + ], + "type": "image", + "image_path": "dd5c16658422c8ea48a85e862587f1e9166987e0def46f1d59f24ebb237a1f7a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 187, + 482, + 221 + ], + "lines": [ + { + "bbox": [ + 130, + 187, + 482, + 221 + ], + "spans": [ + { + "bbox": [ + 130, + 187, + 482, + 221 + ], + "type": "text", + "content": "Fig. 5: Progressive warping results. We visualize the intermediate frames from our progressive warping module. As the iterative step " + }, + { + "bbox": [ + 130, + 187, + 482, + 221 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 187, + 482, + 221 + ], + "type": "text", + "content": " increases, the noise in the reconstructed frames is gradually suppressed and the dynamic range is extended." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 243, + 479, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 479, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 479, + 266 + ], + "type": "text", + "content": "size " + }, + { + "bbox": [ + 130, + 243, + 479, + 266 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 243, + 479, + 266 + ], + "type": "text", + "content": ", indicating that our neural network can be trained and applied on varying chosen " + }, + { + "bbox": [ + 130, + 243, + 479, + 266 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 243, + 479, + 266 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 266, + 479, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 479, + 301 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 479, + 301 + ], + "type": "text", + "content": "Loss function. The chromatic spike denoising module is trained using real-world data with realistic noise. Similar to Wang et al. [31], we adopt the regularized re-visible loss:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 144, + 322, + 480, + 336 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 322, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 144, + 322, + 480, + 336 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {r}} = \\mathcal {L} _ {\\mathrm {r e v}} + \\eta \\mathcal {L} _ {\\mathrm {r e g}} = \\| D (I _ {c}) + \\lambda f _ {\\theta} (I _ {c}) - (\\lambda + 1) I _ {c} \\| _ {2} ^ {2} + \\eta \\| D (I _ {c}) - I _ {c} \\| _ {2} ^ {2} \\tag {12}", + "image_path": "11385bbc02f044ff149e9a732251b8ca128b37e977e880f0444ec0d213da1903.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "text", + "content": " is a constant hyper-parameter, and " + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "text", + "content": " is adjusted in the training progress. We further enhance the denoising module by incorporating an image quality assignment loss, which encourages the predicted score to be close to 1: " + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{i}} = \\| 1 - G_{\\mathrm{iqa}}(D(I_c))\\|_2^2" + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{iqa}}" + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "text", + "content": " denotes a pre-trained image quality assignment model [4]. The total training objective combines these losses: " + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{\\mathrm{r}} + \\lambda_{\\mathrm{i}}\\mathcal{L}_{\\mathrm{i}}" + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "text", + "content": ". Note that the training objectives require no clean frames or human annotations, but only readily accessible real-world mosaicked chromatic spike streams." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 444, + 362, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 444, + 362, + 456 + ], + "spans": [ + { + "bbox": [ + 131, + 444, + 362, + 456 + ], + "type": "text", + "content": "4.2 Color and HDR via progressive warping" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": "Although our chromatic spike denoising module can suppress the noise in frames reconstructed from short-term temporal windows, the dynamic range of each short-term frame is still restricted, as the number of photons arriving at the pixels within an extremely short period (e.g., " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "W = 10" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": ") is limited. Moreover, though our denoising module can fit arbitrary temporal window " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": " theoretically, longer-term input frames lead to better output quality. To reconstruct images with less noise and higher dynamic range, it is necessary to select longer windows to preserve sufficient photons. However, directly warping the frames over the longer temporal window (e.g., " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "W = 100" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": ") is challenging since high-speed motions may span a wide range of pixels. Thus, as illustrated in Fig. 3 (b), we develop the progressive warping module, which only warps a small number of " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": " chromatic frames in each step. By progressively applying the warping operation to the output of the previous step, at the " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": "-th step, each output frame is equivalent to the weighted accumulation from " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "[(K - 1) \\times L + 1] \\times W" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": " aligned chromatic spike planes. Setting a large value of " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": " for the progressive warping module is equivalent to creating a pseudo-long exposure. The pseudo-long exposure leverages optical flow estimation at each step to obtain HDR frames without motion blur." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "S. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "As shown in Fig. 3 (b), the progressive warping module involves the application of the pretrained chromatic spike denoising module from Sec. 4.1 and demosaicking for better optical flow estimation. In the initial 0-th step, we split " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " chromatic spike planes using a short-term temporal window " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " and accumulate the chromatic spikes in each window to form " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\frac{T}{W}" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " mosaicked frames with Eq. (4). We denote these initial mosaicked frames as " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "I_{c,0}(\\mathbf{x},t)" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "c \\in \\{R,G,B\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": ", which is contaminated by noise given a small temporal window. Afterward, for more accurate optical flow estimation, we feed " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "I_{c,0}(\\mathbf{x},t)" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " to the self-supervised denoising module and the demosaicking module that employs linear interpolation to generate intermediate reconstruction frames:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 209, + 255, + 481, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 255, + 481, + 268 + ], + "spans": [ + { + "bbox": [ + 209, + 255, + 481, + 268 + ], + "type": "interline_equation", + "content": "R _ {0} (t) = \\Psi [ D (I _ {R, 0} (t)), D (I _ {G, 0} (t)), D (I _ {B, 0} (t)) ], \\tag {13}", + "image_path": "a40856821591b3d4a23e4e168b9ea16309c933700d0a1ce54c774611c6c41572.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 277, + 482, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 277, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 131, + 277, + 482, + 338 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 277, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\varPsi" + }, + { + "bbox": [ + 131, + 277, + 482, + 338 + ], + "type": "text", + "content": " denotes linear interpolation demosaicking and " + }, + { + "bbox": [ + 131, + 277, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 131, + 277, + 482, + 338 + ], + "type": "text", + "content": " is omitted. We then estimate the optical flow " + }, + { + "bbox": [ + 131, + 277, + 482, + 338 + ], + "type": "inline_equation", + "content": "F_{0}" + }, + { + "bbox": [ + 131, + 277, + 482, + 338 + ], + "type": "text", + "content": " with off-the-shelf methods from " + }, + { + "bbox": [ + 131, + 277, + 482, + 338 + ], + "type": "inline_equation", + "content": "R_0" + }, + { + "bbox": [ + 131, + 277, + 482, + 338 + ], + "type": "text", + "content": " [12]. Note that in this initial step, " + }, + { + "bbox": [ + 131, + 277, + 482, + 338 + ], + "type": "inline_equation", + "content": "I_{c,0}(\\mathbf{x},t)" + }, + { + "bbox": [ + 131, + 277, + 482, + 338 + ], + "type": "text", + "content": " is not warped and is mainly used to estimate a better initial optical flow " + }, + { + "bbox": [ + 131, + 277, + 482, + 338 + ], + "type": "inline_equation", + "content": "F_{0}" + }, + { + "bbox": [ + 131, + 277, + 482, + 338 + ], + "type": "text", + "content": ". Subsequently, " + }, + { + "bbox": [ + 131, + 277, + 482, + 338 + ], + "type": "inline_equation", + "content": "F_{0}" + }, + { + "bbox": [ + 131, + 277, + 482, + 338 + ], + "type": "text", + "content": " is leveraged to warp " + }, + { + "bbox": [ + 131, + 277, + 482, + 338 + ], + "type": "inline_equation", + "content": "K\\times W" + }, + { + "bbox": [ + 131, + 277, + 482, + 338 + ], + "type": "text", + "content": " spike planes towards the reference frame in the 1-th step, described as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 198, + 353, + 481, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 353, + 481, + 388 + ], + "spans": [ + { + "bbox": [ + 198, + 353, + 481, + 388 + ], + "type": "interline_equation", + "content": "I _ {c, 1} (\\mathbf {x}, t) = \\frac {M _ {c}}{K W} \\cdot \\sum_ {t ^ {\\prime} = t - K W / 2} ^ {t + K W / 2} S _ {c} (\\mathbf {x} + F _ {0} (\\mathbf {x}, t, t ^ {\\prime}), t ^ {\\prime}). \\tag {14}", + "image_path": "3d56383ceb29013965c597f7b6ab67d58a9c6ab049d37374a8724bf0dab6647b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "text", + "content": "Note that we utilize chromatic spike planes in the first two steps and mosaicked frames are warped in the later steps. Specifically, for the " + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "text", + "content": "-th iteration step, we continuously pass the accumulated frame " + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "inline_equation", + "content": "I_{c,l-1}" + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "text", + "content": " to the denoising module and the demosaicking module to facilitate the estimation of optical flow " + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "inline_equation", + "content": "F_{l-1}" + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "text", + "content": " from the previously generated frames " + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "inline_equation", + "content": "R_{l-1}" + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "text", + "content": ", and warp " + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "inline_equation", + "content": "I_{c,l-1}" + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "text", + "content": " to the reference time. This operation can further promote noise suppression and dynamic range extension. In particular, we average " + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "text", + "content": " accumulated frames " + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "inline_equation", + "content": "I_{c,l-1}" + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "text", + "content": " obtained in the " + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "inline_equation", + "content": "(l-1)" + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "text", + "content": "-th step, yielding:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 200, + 511, + 481, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 511, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 200, + 511, + 481, + 545 + ], + "type": "interline_equation", + "content": "I _ {c, l} (\\mathbf {x}, t) = \\frac {1}{K} \\cdot \\sum_ {i = - K / 2} ^ {K / 2} I _ {c, l - 1} (\\mathbf {x} + F _ {l - 1} (\\mathbf {x}, t, t ^ {\\prime}), t ^ {\\prime}), \\tag {15}", + "image_path": "fa2bde078b668ad3dcc16301812fd9f64d5ac22988acaff0f16763539b7dece3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 555, + 481, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 481, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 481, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 555, + 481, + 604 + ], + "type": "inline_equation", + "content": "t' = t + iW" + }, + { + "bbox": [ + 130, + 555, + 481, + 604 + ], + "type": "text", + "content": ". The entire procedure is iteratively repeated for " + }, + { + "bbox": [ + 130, + 555, + 481, + 604 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 555, + 481, + 604 + ], + "type": "text", + "content": " steps. In Fig. 5, we show a set of examples to demonstrate the effectiveness of the progressive warping module. It is evident that as the iteration step " + }, + { + "bbox": [ + 130, + 555, + 481, + 604 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 555, + 481, + 604 + ], + "type": "text", + "content": " increases, the details in both highlighted dark regions become clearer." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Finally, we obtain color HDR video frames " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "R_{L}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " with a frame rate of " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\frac{20000}{W}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " FPS. Note that " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " is adjustable, which leads to higher frame rates for smaller values and lower frame rates for larger values. In our most experiments, we find that " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "W = 10" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " offers a good balance between temporal resolution and reconstruction quality, which ensures 2000FPS high-speed color HDR video reconstruction." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 90, + 445, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 90, + 445, + 103 + ], + "spans": [ + { + "bbox": [ + 141, + 90, + 445, + 103 + ], + "type": "text", + "content": "Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 118, + 299, + 160 + ], + "blocks": [ + { + "bbox": [ + 140, + 118, + 299, + 160 + ], + "lines": [ + { + "bbox": [ + 140, + 118, + 299, + 160 + ], + "spans": [ + { + "bbox": [ + 140, + 118, + 299, + 160 + ], + "type": "image", + "image_path": "cc72206286cdabc1ba3833f93db53c6000e52ff8f7b3cf851429b593a7389210.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 162, + 309, + 172 + ], + "lines": [ + { + "bbox": [ + 133, + 162, + 309, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 309, + 172 + ], + "type": "text", + "content": "(a) Synthetic chromatic spike streams and reconstructions" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 310, + 117, + 472, + 160 + ], + "blocks": [ + { + "bbox": [ + 310, + 117, + 472, + 160 + ], + "lines": [ + { + "bbox": [ + 310, + 117, + 472, + 160 + ], + "spans": [ + { + "bbox": [ + 310, + 117, + 472, + 160 + ], + "type": "image", + "image_path": "3627e88a3ec9de2070838e63d9324d5bf1514617c639f847a1ab1e86bb569626.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 162, + 475, + 171 + ], + "lines": [ + { + "bbox": [ + 315, + 162, + 475, + 171 + ], + "spans": [ + { + "bbox": [ + 315, + 162, + 475, + 171 + ], + "type": "text", + "content": "(b) Real chromatic spike streams and reconstructions" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 182, + 480, + 215 + ], + "lines": [ + { + "bbox": [ + 130, + 182, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 480, + 215 + ], + "type": "text", + "content": "Fig. 6: Illustration of our real-world collected chromatic spike streams, synthetic data derived from conventional cameras, and corresponding reconstructions from our proposed method." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 238, + 244, + 479, + 320 + ], + "blocks": [ + { + "bbox": [ + 132, + 232, + 223, + 321 + ], + "lines": [ + { + "bbox": [ + 132, + 232, + 223, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 232, + 223, + 321 + ], + "type": "text", + "content": "Table 1: Quantitative evaluation of the image reconstruction quality. The scores are averaged across our synthetic dataset. " + }, + { + "bbox": [ + 132, + 232, + 223, + 321 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 132, + 232, + 223, + 321 + ], + "type": "text", + "content": " (↓) means higher (lower) is better." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 238, + 244, + 479, + 320 + ], + "lines": [ + { + "bbox": [ + 238, + 244, + 479, + 320 + ], + "spans": [ + { + "bbox": [ + 238, + 244, + 479, + 320 + ], + "type": "table", + "html": "
MethodPSNR ↑SSIM ↑LPIPS ↓HDR-VDP3 ↑HDR-VQM ↓
TFI [46]19.2150.2460.7164.6450.7897
TFP(W=10) [46]23.2530.3310.6095.7780.7799
TFP(W=100) [46]28.7910.7200.2787.5450.4994
TFSTP [43]20.9450.2530.6704.5870.7843
MS23 [21]28.4100.6350.3597.1800.5765
Ours30.7800.8830.2087.9910.3023
", + "image_path": "fde4ca2317f5a8eb2fd541bc18684e12c3e03bbb526393280fab1a6891a1fc14.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 360, + 224, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 360, + 224, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 360, + 224, + 373 + ], + "type": "text", + "content": "5 Experiment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 400, + 236, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 400, + 236, + 411 + ], + "spans": [ + { + "bbox": [ + 132, + 400, + 236, + 411 + ], + "type": "text", + "content": "5.1 Data collection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 434, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 555 + ], + "type": "text", + "content": "Synthetic Data. In line with the methodologies of monochromatic spike simulators (e.g., Zhao et al. [41], SpikeCV [42]), we develop a chromatic spike simulator, following the same Bayer pattern as the real-world chromatic spike camera. We leverage the frames in the GoPro dataset [26] and their corresponding synthetic spike streams to train a supervised learning model and compare it with our self-supervised learning method, which addresses the superiority of our strategy. We interpolate the HDR videos collected by Chang et al. [3] as the evaluation data for quantitative evaluation in our experiments. Some examples of these synthetic scenes are shown in Fig. 6 (a). For the detailed design of our chromatic spike simulator, please refer to our supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Real-world data. To accomplish real-data-driven self-supervised learning and validate the effectiveness of our proposed method, we capture a collection of real-world chromatic spikes using a spike camera equipped with the Bayer pattern. Our dataset comprises over 100 sets of mosaicked chromatic spikes captured from 30 diverse scenes, characterized by varying degrees of high-speed motion and high dynamic range, as depicted in Fig. 6 (b). The real-world data inherently contain the noise characteristics of a real spike camera prototype, posing significant challenges for denoising and demosaicking algorithms. Our real-world dataset is available at https://huggingface.co/datasets/YOUSIKI/chromatic-spikes." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "S. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 371, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 371, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 371, + 129 + ], + "type": "text", + "content": "5.2 Quantitative evaluation on synthetic data" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 136, + 482, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 136, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 482, + 291 + ], + "type": "text", + "content": "We compare our method with existing spike-based video reconstruction methods, i.e., TFI [46], TFP [46], and TFSTP [43]. We acknowledge that it may not be entirely fair to compare our method with these approaches, as they are designed for monochromatic spikes. However, they can serve as baselines to demonstrate the effectiveness of our denoising and progressive warping modules. Since the data format of the spike camera and the single photon avalanche diodes (SPAD) are both single-bit signals, the color video reconstruction method can be adapted to chromatic spikes, and MS23 [21] is chosen for comparison7. The quantitative evaluation is performed on synthetic data. We select PSNR, SSIM [32], LPIPS [37], HDR-VDP3 [25], and HDR-VQM [27] as the metrics. The quantitative evaluation is shown in Tab. 1. We can see that the proposed method demonstrated superior performance over the competing techniques across all metrics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 309, + 336, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 309, + 336, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 309, + 336, + 320 + ], + "type": "text", + "content": "5.3 Qualitative evaluation on real data" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 329, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 482, + 510 + ], + "type": "text", + "content": "To underscore the advantages of our method, we conduct a qualitative evaluation on real and synthetic chromatic spike streams. Fig. 7 presents a quadruple comparison. Our approach not only reconstructs motion blur free frames and recovers color information but also achieves the best balance between noise suppression and detail preservation across diverse scenes. In Fig. 7 (a), we create a colorful and dynamic environment by placing a high-speed electric fan between our chromatic spike camera and a set of toys. The reconstruction from TFP " + }, + { + "bbox": [ + 130, + 329, + 482, + 510 + ], + "type": "inline_equation", + "content": "(W = 100)" + }, + { + "bbox": [ + 130, + 329, + 482, + 510 + ], + "type": "text", + "content": " exhibits noticeable motion blur in the blue bounding box. Conversely, TFI, TFP " + }, + { + "bbox": [ + 130, + 329, + 482, + 510 + ], + "type": "inline_equation", + "content": "(W = 10)" + }, + { + "bbox": [ + 130, + 329, + 482, + 510 + ], + "type": "text", + "content": ", TFSTP and MS23 [21] suffer from significant noise. On the contrary, our method yields a clean outcome. Fig. 7 (b) showcases flowing water in glass and a rotating toy duck. As indicated by red bounding boxes, TFI, TFP " + }, + { + "bbox": [ + 130, + 329, + 482, + 510 + ], + "type": "inline_equation", + "content": "(W = 10)" + }, + { + "bbox": [ + 130, + 329, + 482, + 510 + ], + "type": "text", + "content": ", TFSTP, and MS23 [21] contain different levels of noise. And TFP " + }, + { + "bbox": [ + 130, + 329, + 482, + 510 + ], + "type": "inline_equation", + "content": "(W = 100)" + }, + { + "bbox": [ + 130, + 329, + 482, + 510 + ], + "type": "text", + "content": " suffers from heavy motion blur in blue bounding box. Meanwhile, our proposed method produces clean reconstruction without blur. Fig. 7 (c, d) further demonstrates the denoising advantage of our method on synthetic data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 527, + 220, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 527, + 220, + 540 + ], + "spans": [ + { + "bbox": [ + 132, + 527, + 220, + 540 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 552, + 482, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 482, + 636 + ], + "type": "text", + "content": "In this paper, we propose a novel real-data-driven approach for high-speed color HDR video reconstruction from mosaicked chromatic spikes. Through the analysis on the noise of chromatic spikes, we find the noise in short-term frames can be regarded as a zero-mean distribution. According to this characteristic, we design a self-supervised denoising module that can be trained on real-world chromatic spikes. We further propose a progressive warping module that enables the reconstruction of HDR frames from a pseudo long exposure. Experiments on" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 90, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 90, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 141, + 90, + 445, + 102 + ], + "type": "text", + "content": "Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "text", + "content": "7 As the source code for MS23 [21] is not publicly accessible, we conduct reimplementation and modification for the evaluation." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 114, + 481, + 454 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 454 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 454 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 454 + ], + "type": "table", + "html": "
TFITFP (W = 10)TFP (W = 100)TFSTPMS23Ours
(a)
(b)
(c)
(d)
", + "image_path": "bcb7d0f7f9c2c4239cf6e2b6269c113dd2928189bf01bf46dd7c7c8bfb4ebb64.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 464, + 482, + 498 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 482, + 498 + ], + "type": "text", + "content": "Fig. 7: Visual quality comparison of real (a, b) and synthetic (c, d) data between the proposed method and compared methods. Please zoom-in electronic versions for better details, and watch the videos in the supplementary material." + } + ] + } + ], + "index": 3, + "type": "text" + }, + { + "bbox": [ + 130, + 521, + 481, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 544 + ], + "type": "text", + "content": "real-world data demonstrate the superiority of our approach for reconstructing 2000FPS color HDR videos." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 545, + 482, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 654 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 654 + ], + "type": "text", + "content": "Limitations and future work. The Bayer-pattern CFA introduces chromatic information to the spikes but also reduces the intensity of incident light. As a result, compared to monochromatic spikes, the ability to preserve textures in low-light regions is relatively compromised in high-speed conditions. In future research, exploring different types of CFAs to enhance low-light performance would be worth investigating. And the excessive focus on denoising can result in over-smooth downsides, which may lead to the loss of fine details. Future work could involve the development of a more sophisticated denoising module to balance noise suppression and detail preservation." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "S. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 115, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 115, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 187 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 187 + ], + "type": "text", + "content": "This work was supported by National Science and Technology Major Project (Grant No. 2021ZD0109803) and National Natural Science Foundation of China (Grant No. 62088102, 62136001, 62301009). Bin Fan was supported by National Postdoctoral Program for Innovative Talents of China (Grant No.BX20230013)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 228, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 228, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 251 + ], + "type": "text", + "content": "1. Batson, J., Royer, L.: Noise2Self: Blind denoising by self-supervision. In: ICML. pp. 524-533 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 251, + 480, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 480, + 273 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 480, + 273 + ], + "type": "text", + "content": "2. Buades, A., Coll, B., Morel, J.M., Sbert, C.: Self-similarity driven color demosaicking. IEEE TIP 18(6), 1192-1202 (2009)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 274, + 480, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 274, + 480, + 294 + ], + "spans": [ + { + "bbox": [ + 138, + 274, + 480, + 294 + ], + "type": "text", + "content": "3. Chang, Y., Zhou, C., Hong, Y., Hu, L., Xu, C., Huang, T., Shi, B.: 1000 FPS HDR video with a spike-RGB hybrid camera. In: CVPR. pp. 22180-22190 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 295, + 480, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 295, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 138, + 295, + 480, + 327 + ], + "type": "text", + "content": "4. Chen, C., Mo, J., Hou, J., Wu, H., Liao, L., Sun, W., Yan, Q., Lin, W.: TOPIQ: A top-down approach from semantics to distortions for image quality assessment. IEEE Transactions on Image Processing (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 328, + 480, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 328, + 480, + 349 + ], + "spans": [ + { + "bbox": [ + 138, + 328, + 480, + 349 + ], + "type": "text", + "content": "5. Chen, S., Duan, C., Yu, Z., Xiong, R., Huang, T.: Self-supervised mutual learning for dynamic scene reconstruction of spiking camera. In: IJCAI. p. 2859-2866 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 350, + 480, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 350, + 480, + 381 + ], + "spans": [ + { + "bbox": [ + 138, + 350, + 480, + 381 + ], + "type": "text", + "content": "6. Chen, S., Yu, Z., Huang, T.: Self-supervised joint dynamic scene reconstruction and optical flow estimation for spiking camera. In: AAAI. vol. 37, pp. 350-358 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 383, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 383, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 383, + 480, + 403 + ], + "type": "text", + "content": "7. Condat, L., Mosaddegh, S.: Joint demosaicking and denoising by total variation minimization. In: ICIP. pp. 2781-2784 (2012)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 404, + 480, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 404, + 480, + 425 + ], + "spans": [ + { + "bbox": [ + 138, + 404, + 480, + 425 + ], + "type": "text", + "content": "8. Fossum, E.R., Ma, J., Masoodian, S., Anzagira, L., Zizza, R.: The quanta image sensor: Every photon counts. Sensors 16(8), 1260 (2016)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 426, + 480, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 426, + 480, + 447 + ], + "spans": [ + { + "bbox": [ + 138, + 426, + 480, + 447 + ], + "type": "text", + "content": "9. Gharbi, M., Chaurasia, G., Paris, S., Durand, F.: Deep joint demosaicking and denoising. ACM TOG 35(6), 1-12 (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 449, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 480, + 468 + ], + "type": "text", + "content": "0. Gnanasambandam, A., Chan, S.H.: HDR imaging with quanta image sensors: Theoretical limits and optimal reconstruction. IEEE TCI 6, 1571-1585 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 470, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 470, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 134, + 470, + 480, + 501 + ], + "type": "text", + "content": "11. Han, J., Zhou, C., Duan, P., Tang, Y., Xu, C., Xu, C., Huang, T., Shi, B.: Neuromorphic camera guided high dynamic range imaging. In: CVPR. pp. 1730-1739 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 502, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 502, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 134, + 502, + 480, + 533 + ], + "type": "text", + "content": "12. Hasinoff, S.W., Sharlet, D., Geiss, R., Adams, A., Barron, J.T., Kainz, F., Chen, J., Levoy, M.: Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM TOG 35(6), 1-12 (2016)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 535, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 535, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 134, + 535, + 480, + 567 + ], + "type": "text", + "content": "13. Heide, F., Rouf, M., Hullin, M.B., Labitzke, B., Heidrich, W., Kolb, A.: High-quality computational imaging through simple lenses. ACM TOG 32(5), 1-14 (2013)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 134, + 567, + 480, + 599 + ], + "type": "text", + "content": "14. Huang, T., Zheng, Y., Yu, Z., Chen, R., Li, Y., Xiong, R., Ma, L., Zhao, J., Dong, S., Zhu, L., Li, J., Jia, S., Fu, Y., Shi, B., Wu, S., Tian, Y.: " + }, + { + "bbox": [ + 134, + 567, + 480, + 599 + ], + "type": "inline_equation", + "content": "1000 \\times" + }, + { + "bbox": [ + 134, + 567, + 480, + 599 + ], + "type": "text", + "content": " faster camera and machine vision with ordinary devices. Engineering 25, 110-119 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 480, + 621 + ], + "type": "text", + "content": "15. Kiku, D., Monno, Y., Tanaka, M., Okutomi, M.: Beyond color difference: Residual interpolation for color image demosaicking. IEEE TIP 25(3), 1288-1300 (2016)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 134, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 134, + 622, + 480, + 643 + ], + "type": "text", + "content": "16. Kokkinos, F., Lefkimmiatis, S.: Deep image demosaicking using a cascade of convolutional residual denoising networks. In: ECCV. pp. 303-319 (2018)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "type": "text", + "content": "17. Lichtsteiner, P., Posch, C., Delbruck, T.: A " + }, + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "type": "text", + "content": " 120 db " + }, + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "type": "inline_equation", + "content": "15\\mu s" + }, + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "type": "text", + "content": " latency asynchronous temporal contrast vision sensor. IEEE JSSC 43(2), 566-576 (2008)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 91, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 91, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 141, + 91, + 445, + 102 + ], + "type": "text", + "content": "Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "18. Liu, L., Jia, X., Liu, J., Tian, Q.: Joint demosaicing and denoising with self guidance. In: CVPR. pp. 2240-2249 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 161 + ], + "type": "text", + "content": "19. Liu, Y., Gutierrez-Barragan, F., Ingle, A., Gupta, M., Velten, A.: Single-photon camera guided extreme dynamic range imaging. In: WACV. pp. 1575-1585 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 162, + 480, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 480, + 184 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 480, + 184 + ], + "type": "text", + "content": "20. Ma, S., Gupta, S., Ulku, A.C., Bruschini, C., Charbon, E., Gupta, M.: Quanta burst photography. ACM TOG 39(4), 79-1 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 185, + 480, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 480, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 480, + 206 + ], + "type": "text", + "content": "21. Ma, S., Sundar, V., Mos, P., Bruschini, C., Charbon, E., Gupta, M.: Seeing photons in color. ACM TOG 42(4), 1-16 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 207, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 480, + 228 + ], + "type": "text", + "content": "22. Mairal, J., Bach, F., Ponce, J., Sapiro, G., Zisserman, A.: Non-local sparse models for image restoration. In: ICCV. pp. 2272-2279 (2009)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 229, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 480, + 251 + ], + "type": "text", + "content": "23. Mairal, J., Elad, M., Sapiro, G.: Sparse representation for color image restoration. IEEE TIP 17(1), 53-69 (2007)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 252, + 480, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 252, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 252, + 480, + 274 + ], + "type": "text", + "content": "24. Malvar, H.S., He, L.w., Cutler, R.: High-quality linear interpolation for demosaicing of bayer-patterned color images. In: ICASSP. vol. 3, pp. iii-485 (2004)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 274, + 480, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 480, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 480, + 307 + ], + "type": "text", + "content": "25. Mantiuk, R.K., Hammou, D., Hanji, P.: HDR-VDP-3: A multi-metric for predicting image differences, quality and contrast distortions in high dynamic range and regular content. arXiv preprint arXiv:2304.13625 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 308, + 480, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 480, + 330 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 480, + 330 + ], + "type": "text", + "content": "26. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: CVPR. pp. 3883-3891 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 331, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 480, + 363 + ], + "type": "text", + "content": "27. Narwaria, M., Da Silva, M.P., Le Callet, P.: HDR-VQM: An objective quality measure for high dynamic range video. Signal Processing: Image Communication 35, 46-60 (2015)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "type": "text", + "content": "28. Pekkucuksen, I., Altunbasak, Y.: Gradient based threshold free color filter array interpolation. In: ICIP. pp. 137-140 (2010)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "text", + "content": "29. Tan, D.S., Chen, W.Y., Hua, K.L.: Deepdemosaicking: Adaptive image demosaicking via multiple deep fully convolutional networks. IEEE TIP 27(5), 2408-2419 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 419, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 480, + 441 + ], + "type": "text", + "content": "30. Tan, R., Zhang, K., Zuo, W., Zhang, L.: Color image demosaicking via deep residual learning. In: ICME. pp. 793-798 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 442, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 480, + 464 + ], + "type": "text", + "content": "31. Wang, Z., Liu, J., Li, G., Han, H.: Blind2unblind: Self-supervised image denoising with visible blind spots. In: CVPR. pp. 2027-2036 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 464, + 480, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 480, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 480, + 486 + ], + "type": "text", + "content": "32. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: From error visibility to structural similarity. IEEE TIP 13(4), 600-612 (2004)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 487, + 480, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 480, + 509 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 480, + 509 + ], + "type": "text", + "content": "33. Xia, L., Zhao, J., Xiong, R., Huang, T.: SVFI: Spiking-based video frame interpolation for high-speed motion. In: AAAI. vol. 37, pp. 2910-2918 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 510, + 480, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 480, + 532 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 480, + 532 + ], + "type": "text", + "content": "34. Xing, W., Egiazarian, K.: End-to-end learning for joint image demosaicing, denoising and super-resolution. In: CVPR. pp. 3507-3516 (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 533, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 480, + 553 + ], + "type": "text", + "content": "35. Yang, Y., Han, J., Liang, J., Sato, I., Shi, B.: Learning event guided high dynamic range video reconstruction. In: CVPR. pp. 13924-13934 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 554, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 480, + 586 + ], + "type": "text", + "content": "36. Yu, G., Sapiro, G., Mallat, S.: Solving inverse problems with piecewise linear estimators: From gaussian mixture models to structured sparsity. IEEE TIP 21(5), 2481-2499 (2011)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 587, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 480, + 609 + ], + "type": "text", + "content": "37. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR. pp. 586-595 (2018)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "type": "text", + "content": "38. Zhang, Z., Zheng, H., Hong, R., Xu, M., Yan, S., Wang, M.: Deep color consistent network for low-light image enhancement. In: CVPR. pp. 1899-1908 (2022)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "39. Zhao, J., Xiong, R., Liu, H., Zhang, J., Huang, T.: Spk2ImgNet: Learning to reconstruct dynamic scene from continuous spike stream. In: CVPR. pp. 11996-12005 (2021)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "S. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 315 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "40. Zhao, J., Xiong, R., Xie, J., Shi, B., Yu, Z., Gao, W., Huang, T.: Reconstructing clear image for high-speed motion scene with a retina-inspired spike camera. IEEE TCI 8, 12-27 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "text", + "content": "41. Zhao, J., Zhang, S., Ma, L., Yu, Z., Huang, T.: SpikingSIM: A bio-inspired spiking simulator. In: IEEE ISCAS. pp. 3003-3007 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "text", + "content": "42. Zheng, Y., Zhang, J., Zhao, R., Ding, J., Chen, S., Xiong, R., Yu, Z., Huang, T.: SpikeCV: Open a continuous computer vision era. arXiv preprint arXiv:2303.11684 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "43. Zheng, Y., Zheng, L., Yu, Z., Huang, T., Wang, S.: Capture the moment: High-speed imaging with spiking cameras through short-term plasticity. IEEE TPAMI 45(7), 8127-8142 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "text", + "content": "44. Zheng, Y., Zheng, L., Yu, Z., Shi, B., Tian, Y., Huang, T.: High-speed image reconstruction through short-term plasticity for spiking cameras. In: CVPR. pp. 6358-6367 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "type": "text", + "content": "45. Zhu, L., Dong, S., Huang, T., Tian, Y.: A retina-inspired sampling method for visual texture reconstruction. In: ICME. pp. 1432-1437 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 315 + ], + "type": "text", + "content": "46. Zhu, L., Dong, S., Li, J., Huang, T., Tian, Y.: Retina-like visual image reconstruction via spiking neural model. In: CVPR. pp. 1438-1446 (2020)" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 90, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 90, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 141, + 90, + 445, + 102 + ], + "type": "text", + "content": "Real-data-driven 2000 FPS Color Video from Mosaicked Chromatic Spikes" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Real-time 3D-aware Portrait Editing from a Single Image/c7f6568e-666e-490d-b059-9d93ade8c6f3_content_list.json b/2024/Real-time 3D-aware Portrait Editing from a Single Image/c7f6568e-666e-490d-b059-9d93ade8c6f3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..eec5aef3f77ef7f485db2a8ddc759d79fd91b9ff --- /dev/null +++ b/2024/Real-time 3D-aware Portrait Editing from a Single Image/c7f6568e-666e-490d-b059-9d93ade8c6f3_content_list.json @@ -0,0 +1,1822 @@ +[ + { + "type": "text", + "text": "Real-time 3D-aware Portrait Editing from a Single Image", + "text_level": 1, + "bbox": [ + 290, + 140, + 710, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qingyan Bai $^{1,2}$ , Zifan Shi $^{1}$ , Yinghao Xu $^{3}$ , Hao Ouyang $^{1,2}$", + "bbox": [ + 290, + 210, + 707, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qiuyu Wang $^{2}$ , Ceyuan Yang $^{4}$ , Xuan Wang $^{2}$", + "bbox": [ + 341, + 228, + 658, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gordon Wetzstein3, Yujun Shen2, and Qifeng Chen1", + "bbox": [ + 305, + 243, + 694, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 The Hong Kong University of Science and Technology", + "bbox": [ + 313, + 268, + 687, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Ant Group", + "bbox": [ + 455, + 282, + 545, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Stanford University", + "bbox": [ + 426, + 297, + 573, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Shanghai AI Laboratory", + "bbox": [ + 411, + 311, + 589, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. This work presents 3DPE, a practical method that can efficiently edit a face image following given prompts, like reference images or text descriptions, in a 3D-aware manner. To this end, a lightweight module is distilled from a 3D portrait generator and a text-to-image model, which provide prior knowledge of face geometry and superior editing capability, respectively. Such a design brings two compelling advantages over existing approaches. First, our method achieves real-time editing with a feedforward network (i.e., $\\sim 0.04$ s per image), over $100\\times$ faster than the second competitor. Second, thanks to the powerful priors, our module could focus on the learning of editing-related variations, such that it manages to handle various types of editing simultaneously in the training phase and further supports fast adaptation to user-specified customized types of editing during inference (e.g., with $\\sim 5$ min fine-tuning per style). Project page can be found here.", + "bbox": [ + 259, + 359, + 738, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D-aware portrait $\\cdot$ Efficient editing", + "bbox": [ + 261, + 566, + 589, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 604, + 374, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inferring the geometry and appearance from a single-view portrait image has become mature and practical [22, 37, 42, 66, 70, 74, 76], largely attributed to the utilization of priors in various 2D/3D generative models. However, only performing geometry reconstruction is insufficient. The significance of 3D portrait editing, driven by user intentions, and the need for streamlined efficiency in the editing process has been steadily increasing. This is particularly crucial in real-world applications such as AR/VR, 3D telepresence, and video conferencing, where real-time editing is often essential. Consequently, a key question arises: How can we effectively address the challenge of attaining high-fidelity portrait editing while ensuring real-time efficiency?", + "bbox": [ + 212, + 635, + 785, + 785 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditional methods for 3D portrait editing [17] typically rely on template facial models [6,51], which has limitations in handling the substantial geometry", + "bbox": [ + 212, + 786, + 785, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding authors.", + "bbox": [ + 217, + 824, + 390, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ae23707245db9249b6c133432e8b7729bf5940f9afcb2fa0ff5ab65b2a3394c8.jpg", + "image_caption": [ + "Fig. 1: Photorealistic editing results produced by our proposed 3DPE, which allows users to perform 3D-aware portrait editing using image or text prompts. In comparison with baseline methods, such as InstructPix2Pix (IP2P) [7] + Live3D [70] and PTI [55] + StyleCLIP [50] (details are illustrated in Sec. 4), our approach accurately follows the guidance from reference prompts and maintains sufficiently better efficiency." + ], + "image_footnote": [], + "bbox": [ + 222, + 145, + 504, + 320 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/215ff6e676dd1831c92569bc9b0864817bfff9e03a3fe21b5f1dd19979eb5290.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 513, + 146, + 787, + 320 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "changes because it often overlooks precise features like hair and beard. Recent 3D GANs [13, 14, 58, 62, 75] show remarkable capabilities in generating high-fidelity 3D portraits. They can serve as powerful generative priors for 3D portrait editing when coupled with GAN inversion techniques [2, 3, 61, 86, 87]. However, these methods may encounter issues related to geometry distortion or exhibit slow speed, and the editing is constrained by the limited latent attributes. Besides, 2D diffusion models can offer a strong prior with Score Distillation Sampling (SDS) loss [52] for editing purposes [30]. Nonetheless, they often require step-by-step optimization and become a major bottleneck for real-time applications.", + "bbox": [ + 212, + 439, + 787, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we present 3DPE, a real-time 3D-aware Portrait Editing method driven by user-defined prompts. As shown in Fig. 1, when provided with a single-view portrait image, our approach empowers a versatile range of editing styles with one model through flexible instructions, including images and texts. We leverage the powerful 3D prior from a 3D-aware face generator and achieve a high-fidelity 3D reconstruction of the portrait image. Subsequently, we distill editing knowledge from a text-guided image editing model into a lightweight module integrated with the 3D-aware generator. This module, characterized by its minimal computational cost, allows our method to maintain real-time inference and excel in handling various types of editing, while ensuring good 3D consistency. An additional advantage is that our model supports customization through user-specified prompts with fast adaptation speed. This empowers users to build their own editing model at a minimal cost, enabling our method to cater to a broader audience.", + "bbox": [ + 212, + 580, + 787, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method achieves real-time 3D-aware portrait editing through the utilization of a feedforward network, with a processing time of $40\\mathrm{ms}$ on a standard consumer GPU. Additionally, we present a comprehensive evaluation of", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Q. Bai et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "our method using various prompts both quantitatively and qualitatively. Our design choices are also validated through comparisons with ablated variants of our method. Compared with baseline methods, our approach demonstrates superior 3D consistency, precise texture alignment, and a substantial improvement in inference time, as reflected by the evaluation metrics. We demonstrate the versatility of our method by showcasing its capacity to perform a wide variety of edits, including text and image prompts, on portrait images.", + "bbox": [ + 212, + 146, + 782, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, the contributions of our work include:", + "bbox": [ + 238, + 253, + 607, + 266 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a lightweight module to distill knowledge from 3D GANs and diffusion models for 3D-aware editing from a single image. Due to the minimal cost of the new module, our model maintains real-time performance.", + "- Our model supports fast adaptation to user-specified editing, requiring only 10 image pairs and 5 minutes for the adaptation.", + "- Our framework can accommodate various control signals, including text and image prompts." + ], + "bbox": [ + 225, + 277, + 782, + 382 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 406, + 385, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Generative Face Priors", + "text_level": 1, + "bbox": [ + 215, + 436, + 450, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generative models aim at modeling the underlying distribution of the data, containing a wealth of prior knowledge. Recently, 3D GANs [13, 14, 25, 47, 49, 58, 59, 63-65, 69, 71, 77] are mostly adopted to learn 3D faces from single-view image dataset. The interior rich domain-specific geometry priors enable various face-related applications such as image editing [9, 66, 67, 70, 79, 85] and domain adaptation [1, 34, 36, 46, 82]. Large-scale diffusion models [56], in contrast, encode knowledge of huge datasets and thus can provide general prior information. Such priors are broadly leveraged for tasks such as image editing [7, 10, 24, 45, 83], customization [31, 38, 44, 57], and video editing [11, 12, 23, 43, 48, 53, 73, 80]. In our approach, we capitalize on the advantages of both the geometry prior derived from 3D GANs and the broader editing prior offered by large-scale text-to-image diffusion models, instead of relying solely on a single type of prior.", + "bbox": [ + 212, + 460, + 784, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Portrait Editing from a Single Image", + "text_level": 1, + "bbox": [ + 215, + 664, + 566, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although many methods work well for reconstructing [22, 26] or generating [13] faces, editing has become a necessary interface to connect these methods with real-world applications. Previously, portrait editing given a single-view image was mostly completed in 2D image space facilitated with GAN inversion [2, 3, 5, 28, 55, 60, 61, 78, 86], which enables fast editing by exploring the trajectories in the GAN's latent space. Recent diffusion-based portrait editing [7, 83] can support various editing types with texts as guidance. However, most of them are done in the 2D space, and thus there is no guarantee for the underlying 3D consistency. Therefore, 3D-aware portrait editing is crucial to achieve the goal. Some methods [15, 33, 39, 42, 66, 67, 74, 76, 81] rely on the latent space", + "bbox": [ + 212, + 688, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "3DPE", + "bbox": [ + 687, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of 3D GANs to perform editing through walking in the latent space, but are limited by the number of editable latent attributes. A few methods, such as Instruct-NeRF2NeRF [27], ClipFace [4], and LENeRF [32], attempt to leverage the large-scale models [54, 56] to edit on 3D representations with texts as guidance but require heavy iterative refinement of the edited results. Similarly, InstructPix2NeRF [40] leverages text-to-image models and text prompts to edit within the latent space of the GAN-based 3D generator. However, it is held back by the multi-step diffusion inference and the heavy structure of a GAN-based generator. Ours, in contrast, benefits from the lightweight design of the module and achieves real-time editing with ease. With such a design, our model can also adapt fast to the user-defined editing prompts. Moreover, texts cannot always illustrate the desired effects precisely, and sometimes an image prompt serves as better guidance for image editing [20, 57]. Our method can support not only text prompts but also image prompts for editing, providing a more flexible and user-friendly interface.", + "bbox": [ + 212, + 146, + 787, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 396, + 331, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our approach takes a portrait image $\\mathbf{I}$ and its camera pose $\\mathbf{c}$ , obtained via the face pose estimator [18]. Additionally, the referenced prompts $\\mathbf{P}$ (images or texts) serve as the editing instructions. The outcome of our model is an edited version of the 3D portraits characterized by NeRF in accordance with the prompts. The edited image is denoted as $\\mathbf{I}_p$ . Within our framework, we achieve this objective by leveraging the knowledge of 3D GANs and a text-guided image editing model, which is distilled into a lightweight module. Following distillation, our method enables real-time 3D portrait editing and efficient adaptation to user-specified prompts. In Sec. 3.1, we offer an illustration of 3D GANs and diffusion models. The process of distilling these priors into a lightweight module is outlined in Sec. 3.2. Finally, the details for model training and inference, along with the fast adaptation for customized prompts, are presented in Sec. 3.3.", + "bbox": [ + 212, + 429, + 789, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminary", + "text_level": 1, + "bbox": [ + 215, + 632, + 362, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D GAN Prior for Portrait Reconstruction. The 3D-aware GANs showcase the ability to synthesize photorealistic 3D images using a collection of single-view images. Notably, EG3D [13] introduces an efficient triplane 3D representation, demonstrating high-quality 3D-aware image rendering. Once trained, the generator of EG3D can be applied for single-image 3D reconstruction via GAN inversion. However, existing 3D GAN inversion methods often encounter geometry distortion or exhibit slow inference speed. To tackle these challenges, we utilize Live3D [70], a state-of-the-art single-image portrait 3D reconstruction model built upon EG3D, preserving geometry quality while ensuring real-time performance. Live3D employs a two-branch encoder, $\\mathrm{E}_{high}(\\cdot)$ and $\\mathrm{E}_{low}(\\cdot)$ , to extract different resolution features from the input. It then utilizes a ViT-based", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 127 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Q. Bai et al.", + "bbox": [ + 271, + 114, + 357, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9074058ef0a16f926b765e7d2abeb7ca60bafbd451e32105f57698762302864c.jpg", + "image_caption": [ + "Fig. 2: Overview of our method. We distill the prior in the diffusion model and 3D GAN for real-time 3D-aware editing. Our approach is fine-tuned from Live3D [70], where we extract features from the input portrait I using $\\mathrm{E}_{high}(\\cdot)$ and $\\mathrm{E}_{low}(\\cdot)$ . The prompt embedding is generated with $\\mathrm{E}_p(\\cdot)$ and injected with the input features from $\\mathrm{E}_{high}(\\cdot)$ through a cross-attention mechanism. Our model is trained to mimic the output from the diffusion prior to acquire editing knowledge and enforce geometry constraints through triplane, multi-view images, and depth supervision from the 3D prior. In this context, InstructPix2Pix [7] and Live3D serve as the diffusion and 3D prior, respectively. It's noteworthy that only $\\mathrm{E}_p(\\cdot)$ and $\\mathrm{E}_t(\\cdot)$ are learnable during training, while all other parameters remain frozen." + ], + "image_footnote": [], + "bbox": [ + 236, + 146, + 772, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathrm{E}_t(\\cdot)$ decoder [19] to transform the fused encoder features into the triplane $\\mathbf{T}$ :", + "bbox": [ + 215, + 547, + 784, + 564 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} = \\mathrm {E} _ {t} \\left(\\mathrm {E} _ {\\text {h i g h}} (\\mathbf {I}), \\mathrm {E} _ {\\text {l o w}} (\\mathbf {I})\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 594, + 785, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This triplane $\\mathbf{T}$ is subsequently used in conjunction with the volume rendering module and upsampler of EG3D to generate photorealistic view synthesis given the camera pose c. For clarity, we use $\\mathrm{R}(\\cdot)$ to denote this process.", + "bbox": [ + 212, + 626, + 784, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Diffusion Prior for Portrait Editing. Only performing 3D reconstruction is insufficient as our goal is to edit portrait images. Large-scale diffusion models [56], trained on vast text-image pairs, can synthesize realistic photos with text input, offering powerful editing prior. Despite their effectiveness for face images, the editing process is often slow due to step-by-step optimization or multiple inferences in diffusion models. The scarcity of high-quality and multi-view consistent paired data significantly hinders the practical application of 3D-aware editing from a single image, which is hard to acquire with diffusion models. Therefore, our goal is to distill editing knowledge from the diffusion model, integrate it with the 3D prior in Live3D into a lightweight module, and employ it for real-time portrait editing.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "3DPE", + "bbox": [ + 687, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f81ee83c8db8f852b73649c43459d539b6ad26c055042c6a1c778eb05c796b5f.jpg", + "image_caption": [ + "Fig. 3: Disentanglement in Live3D features. We separately disable the features from $\\mathrm{E}_{high}(\\cdot)$ and $\\mathrm{E}_{low}(\\cdot)$ to infer the reconstructed image. Without $\\mathrm{E}_{high}(\\cdot)$ , the output retains the coarse structure but loses its appearance. Conversely, when $\\mathrm{E}_{low}(\\cdot)$ is deactivated, the reconstructed portraits preserve the texture (such as the blue and purple reflection on the glasses) but fail to capture the geometry." + ], + "image_footnote": [], + "bbox": [ + 250, + 148, + 754, + 392 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Distilling Priors into a Lightweight Module", + "text_level": 1, + "bbox": [ + 214, + 506, + 617, + 522 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method aims to perform real-time 3D-aware editing for various prompts for the single-view portrait. Thus, it needs powerful 3D knowledge for geometry reconstruction and editing prior to handling various control signals. We leverage the strengths of both 3D GANs and diffusion models. In the following, we provide detailed presentations on how the knowledge of these two types of models is distilled into a lightweight module.", + "bbox": [ + 212, + 551, + 787, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Feature Representation in Live3D. We carefully study the Live3D model and discover that, as a two-branch triplane-based 3D reconstruction model, the low-resolution and high-resolution features from the Live3D encoder $\\mathrm{E}_{low}(\\cdot)$ and $\\mathrm{E}_{high}(\\cdot)$ tend to learn various levels of information without explicit guidance. We conduct a study by disabling one of the branches and inferring the reconstructed images. As illustrated in Fig. 3, when the high-resolution encoder $\\mathrm{E}_{high}(\\cdot)$ is disabled with random weights, the inference image retains a similar structure but loses its detailed appearance. Conversely, when the low-resolution encoder $\\mathrm{E}_{low}(\\cdot)$ is disabled, the reconstructed portraits preserve some of the texture from the input but struggle to capture the geometry. Based on this analysis, the features of the two branches separately model low-frequency and high-frequency information. This insight motivates our model design in terms of how to distill knowledge from the diffusion model and 3D GANs.", + "bbox": [ + 212, + 643, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Q. Bai et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Geometry Prediction for Inputs I. In our framework, we use the prompts $\\mathbf{P}$ to refine the input portrait $\\mathbf{I}$ into the edited image $\\mathbf{I}_p$ . The $\\mathbf{I}_p$ inherits a similar structure to the input $\\mathbf{I}$ , providing coarse geometry or structural guidance in the editing process. As discussed in the above section, the encoder $\\mathrm{E}_{low}(\\cdot)$ consistently generates low-frequency features to represent geometry cues, making it well-suited for preserving structured information in the input portrait. As a result, we freeze $\\mathrm{E}_{low}(\\cdot)$ and leverage it to produce structure features $\\mathbf{F}_g = \\mathrm{E}_{low}(\\mathbf{I})$ for coarse geometry prediction.", + "bbox": [ + 212, + 146, + 782, + 263 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Injecting Prompts P as Condition. In contrast to input images, prompts typically offer more high-frequency and texture information to guide and control the editing process. Accordingly, it is appropriate to incorporate the prompts into the high-level branch $\\mathrm{E}_{high}(\\cdot)$ in Live3D to extract the high-frequency features. To incorporate prompts, we utilize the prompt encoder $\\mathrm{E}_p(\\cdot)$ to create prompt embeddings that are fused into our model through cross attention, which is similar to the strategy in Stable Diffusion [56]. Specifically, we add a cross-attention layer after $\\mathrm{E}_{high}(\\cdot)$ to obtain the feature updated with the prompt embeddings:", + "bbox": [ + 212, + 267, + 784, + 401 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {a} = \\operatorname {C r o s s A t t e n t i o n} \\left(\\operatorname {E} _ {\\text {h i g h}} (\\mathbf {I}), \\operatorname {E} _ {p} (\\mathbf {P})\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 412, + 784, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where the encoder $\\mathbf{E}_p(\\cdot)$ is a transformer (MAE [29] for image and CLIP [54] for text). The $\\mathrm{E}_{high}(\\mathbf{I})$ serves as the query and $\\mathrm{E}_p(\\mathbf{P})$ is employed as key and value. The updated feature $\\mathbf{F}_a$ and the geometry feature $\\mathbf{F}_g$ are then feed into the decoder $\\mathrm{E}_t(\\cdot)$ to infer the triplane $\\mathbf{T_p}$ :", + "bbox": [ + 212, + 433, + 782, + 494 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} _ {\\mathbf {p}} = \\operatorname {E} _ {t} \\left(\\mathbf {F} _ {a}, \\mathbf {F} _ {g}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 439, + 503, + 784, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Distilling Diffusion and 3D GAN Prior. Unlike previous methods that require large amounts of 3D data, our model operates with only 2D paired data generated by the 2D editing model [7]. For both visual and textual prompts, we generate edited images $\\mathbf{I}_{gt}$ with the diffusion-based image editing model [7] as the pseudo labels. The details of data generation can be found in Sec. 4.1.", + "bbox": [ + 212, + 523, + 782, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With the predicted triplane $\\mathbf{T}_{\\mathbf{p}}$ , we can render the image using the pretrained EG3D model and then calculate the reconstruction loss as follows:", + "bbox": [ + 212, + 599, + 782, + 630 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {2 d} = \\ell_ {\\mathbf {I}} \\left(\\mathbf {I} _ {g t}, \\mathrm {R} \\left(\\mathbf {T} _ {\\mathbf {p}}, \\mathbf {c}\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 418, + 638, + 784, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathrm{R}(\\cdot)$ is the rendering module of EG3D, $\\mathbf{c}$ is the camera pose of input portrait, and $\\ell_1(\\cdot)$ is an image reconstruction loss penalizing the difference between the ground truth $\\mathbf{I}_{gt}$ and the rendering $\\mathbf{I}_p = \\mathrm{R}(\\mathbf{T_p},\\mathbf{c})$ . It's important to note that $\\mathcal{L}_{2d}$ , used to reconstruct the $\\mathbf{I}_{gt}$ , essentially distills the knowledge of editing from the diffusion model.", + "bbox": [ + 212, + 661, + 782, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We observe that $\\mathcal{L}_{2d}$ helps the model to reconstruct well on the input camera view but suffers from geometry distortion during novel-view synthesis. To fully exploit the 3D properties of Live3D, we propose to distill the 3D knowledge of Live3D. Specifically, we leverage the pretrained Live3D to infer the triplane $\\mathbf{T}_{gt}$ , and multi-view depths $\\mathcal{D}_{gt}$ and images $\\mathcal{T}_{gt}$ of the pseudo-label image $\\mathbf{I}_{gt}$ :", + "bbox": [ + 212, + 736, + 782, + 811 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} _ {g t}, \\mathcal {D} _ {g t} = \\mathrm {G} (\\mathbf {I} _ {g t}, \\mathcal {C}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 428, + 821, + 784, + 835 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "3DPE", + "bbox": [ + 687, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathrm{G}(\\cdot)$ represents the inference process of Live3D, where $\\mathcal{C} = \\{\\mathbf{c}_1,\\dots,\\mathbf{c}_n\\}$ is the camera set, and $n$ denotes the number of cameras. We also render the multi-view images $\\mathcal{I}_p$ and depths $\\mathcal{D}_p$ of the edited image $\\mathrm{I}_p$ from the triplane $\\mathbf{T}_p$ and then define the objective:", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {3 d} = \\ell_ {\\mathcal {I}} \\left(\\mathcal {I} _ {p}, \\mathcal {I} _ {g t}\\right) + \\ell_ {\\mathbf {T}} \\left(\\mathbf {T} _ {p}, \\mathbf {T} _ {g t}\\right) + \\ell_ {\\mathcal {D}} \\left(\\mathcal {D} _ {p}, \\mathcal {D} _ {g t}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 219, + 785, + 237 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\ell_{\\mathcal{I}}(\\cdot), \\ell_{\\mathbf{T}}(\\cdot)$ and $\\ell_{\\mathcal{D}}(\\cdot)$ is the reconstruction loss penalizing the difference in image, triplane, and depth between $\\mathrm{I}_p$ and $\\mathrm{I}_{gt}$ .", + "bbox": [ + 212, + 246, + 785, + 277 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Training and Inference", + "text_level": 1, + "bbox": [ + 215, + 300, + 450, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training. During the training phase, we sample a triplet consisting of the input portrait $\\mathbf{I}$ along with its camera pose $\\mathbf{c}$ , prompts $\\mathbf{P}$ , and the pseudo-label image $\\mathbf{I}_{gt}$ generated by the 2D diffusion model. We leverage the Live3D model as a pretrained model and add an additional prompts encoder $\\mathrm{E}_p$ to extract prompt embeddings. The overall learning objective can be described as follows:", + "bbox": [ + 212, + 340, + 787, + 416 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL = \\lambda_ {1} L _ {2 D} + \\lambda_ {2} L _ {3 D}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 430, + 785, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ and $\\lambda_{2}$ are loss weights. In our setting, $\\lambda_{1}$ and $\\lambda_{2}$ are both set to 1.0. For the reconstruction loss, $\\ell_{\\mathcal{I}}(\\cdot)$ , $\\ell_{\\mathrm{I}}(\\cdot)$ are combinations of $\\mathrm{L}_2$ loss and LPIPS loss [84], with loss weights being 1 and 2, respectively. $\\ell_{\\mathcal{D}}(\\cdot)$ and $\\ell_{\\mathrm{T}}(\\cdot)$ are $\\mathrm{L}_1$ loss. Notably, based on the study on Live3D feature representation, during training only $\\mathrm{E}_p(\\cdot)$ , $\\mathrm{E}_t(\\cdot)$ are set to be learnable, while other modules are frozen. It allows our model to leverage Live3D knowledge as much as possible and converge at a very fast speed.", + "bbox": [ + 212, + 455, + 787, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inference. For inference, users can provide a single 2D portrait image and choose an image prompt or text prompt. Our model is able to generate the edited 3D NeRF along with photorealistic view synthesis.", + "bbox": [ + 212, + 561, + 787, + 607 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Customized Prompts Adaptation. To accommodate customized styles provided by users, we propose a method to adapt our pretrained encoder to user-defined styles. We increase the tuning efficiency by optimizing only $\\mathrm{E}_p(\\cdot)$ and the normalization layers in $\\mathrm{E}_t(\\cdot)$ with the same learning objective Eq. 7. This method allows us to limit the training data to only 10 image pairs and the learning time to 5 minutes on a single GPU.", + "bbox": [ + 212, + 607, + 787, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 722, + 375, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 753, + 428, + 770 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training Settings. For real faces, we adopt the FFHQ dataset [35] at $512 \\times 512$ resolution with camera parameters aligned by EG3D [13]. In order to obtain the stylized images as pseudo labels, and the corresponding textual and visual prompts, we leverage the text-guided editing model InstructPix2Pix (IP2P) [7] to", + "bbox": [ + 212, + 779, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Q. Bai et al.", + "bbox": [ + 271, + 114, + 356, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fac216f03f4f959144cd3e01c13277758aa34d7a06948f1663c2948f73e71b77.jpg", + "image_caption": [ + "mnd" + ], + "image_footnote": [], + "bbox": [ + 254, + 145, + 356, + 199 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b43743ade296fac25742324ff4fa9a9df10d49b1af67b1428a50f37c9aee2567.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 392, + 145, + 491, + 199 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/dd1544dd3fe5b6fe651dd1d5be29ecea4c6ee7378a529a456927ab81272ee084.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 145, + 629, + 198 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/061d13bb5dff2ada5f8b148d1dc34ad11fc1e99106b14272343871933008d364.jpg", + "image_caption": [ + "Recolor it to a Gothic color scheme." + ], + "image_footnote": [], + "bbox": [ + 663, + 145, + 732, + 199 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/90d5bc38436b57d70008b667788d2b87ebaab64700b2695fa04eafcb62090631.jpg", + "image_caption": [ + "P 12345678901001100110011001100110011001100110011001100110011001100110011001100110011001100110011001100110011001100" + ], + "image_footnote": [], + "bbox": [ + 241, + 200, + 370, + 412 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/30e99cad967c55e84925a0c9e13c7325270c8f7988317ca53a515906bbe71ed2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 200, + 506, + 412 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/705204cfb497a2a8f845cc8ab25e4e1106f2351b510e8fa0362e906930690810.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 513, + 200, + 645, + 412 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/9768e369ab754eb13086513b97f1af71fd66067a6fb161bf59c1dc6581884197.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 650, + 200, + 782, + 412 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/26263c704c8728002033135f47f0b7e7721fa49d6d2cca01d12fd9c457879a32.jpg", + "image_caption": [ + "0000000000000000000000000000000000000" + ], + "image_footnote": [], + "bbox": [ + 240, + 414, + 370, + 465 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/fc936265282935b0a001880613bddc357adb22e9187d45e8eb7a1b616e360b41.jpg", + "image_caption": [ + "Fig. 4: Qualitative comparisons. We compare the results of several baselines with image prompts and text prompts. In each case, we include the edited portraits as well as their novel view renderings. Our method generates high-quality edited portraits with better 3D quality and alignment with the referenced prompts." + ], + "image_footnote": [], + "bbox": [ + 374, + 414, + 506, + 465 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/4e1bcb7134c60dbc4d990b8785227a03ac8cc1fbed1a333ea034658db9587a0e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 513, + 414, + 645, + 465 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/cf3df868e514f235686b90c0d843bcec5a1b6f8ac7da4e878adc7f336dd682ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 650, + 414, + 782, + 465 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "edit the real faces. Specifically, we manually generate and select textual prompts, and send them to IP2P together with the real images to obtain the edited 2D images. The text prompts of our training and testing set are obtained by augmenting the aforementioned textual prompts with the large language model GPT [8] following IP2P [7]. While the image prompts of our training and testing set are picked from the edited 2D images. We conduct experiments on 20 styles, and for each style, we synthesize 1000 images, resulting in 20,000 image pairs with text prompts for model training. For each style, we use 8 textual prompts for training and 5 textual prompts for testing. We adopt a learning rate of 5e-5 and optimize the model for 60k iterations with a batch size of 32. The entire training procedure is completed over a period of 40 hours using 8 NVIDIA A100 GPUs. Baselines. We conduct comparisons against two categories of methods that achieve analogous outcomes: 1) 3D reconstruction coupled with 3D editing; 2) 2D editing followed by 3D reconstruction. For the first category, we implement three baseline approaches, wherein we firstly leverage PTI [55] or Live3D [70] for reconstruction and edit both of them following StyleCLIP [50]. Specifically, for experiments in PTI + StyleCLIP, we first conduct pivotal tuning [55] based on the predicted latent codes predicted by e4e [68] to reconstruct the given image in", + "bbox": [ + 212, + 566, + 787, + 843 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "3DPE", + "bbox": [ + 687, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/3964f252442b6d3fee030b588de29b0765bf4ed32401a3d84728302861a136ee.jpg", + "table_caption": [ + "Table 1: Quantitative comparisons. We compare several baselines on the 100 images of FFHQ dataset. It's important to note that we exclude $CLIP_r$ for PTI+StyleCLIP and Live3D+StyleCLIP since these models utilize CLIP for optimization. Our model excels in 3D quality and achieves a remarkable 250x improvement (compared with IP2P+Live3D) in inference speed, achieving real-time performance." + ], + "table_footnote": [], + "table_body": "
MethodIDt↑CLIPr↑3D↑Time↓
PTI+StyleCLIP0.11-0.73360s
Live3D+StyleCLIP0.63-0.7230s
IP2P+Live3D0.440.620.7510s
Ours0.470.730.760.04s
", + "bbox": [ + 272, + 226, + 728, + 306 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "a 3D-aware manner. Then we perform editing on the latent codes following [50], where CLIP loss [54] is utilized between the model output and the prompt (no matter whether it is an image or text). The setting of Live3D + StyleCLIP is similar to PTI + StyleCLIP except that we obtain the triplane by Live3D and perform editing on this 3D representation. SPI [81] + StyleGAN-NADA [21] (StyleNADA for the abbreviation in Fig. 4) serves as the third baseline in the first category, where SPI is used for 3D reconstruction and StyleGAN-NADA is utilized to optimize the 3D generator for editing. In the second category, we employ a diffusion-based text-guided 2D image editing method IP2P [7] prior to executing 3D reconstruction via Live3D. IP2P being a text-instructed editing model, we need to first extract text prompts from the reference image with the image caption algorithm BLIP [41] when image prompts are input. The 2D edited results could then be obtained by sending the input image and the text prompt to IP2P. The final 3D results are acquired by inverting the 2D editing results with Live3D.", + "bbox": [ + 215, + 378, + 785, + 604 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Criteria. We evaluate all metrics on 100 pairs of images processed from FFHQ. We conduct a comprehensive evaluation of the editing performance in the following four aspects: identity preservation, reference alignment, 3D consistency, and inference speed. 1) Identity preservation $(\\mathbf{ID}_{\\mathbf{t}})$ aims to measure the preservation of the original identity by calculating the cosine similarity between the identity feature of the input image $\\mathbf{I}$ and that of the edited image $\\mathbf{I}_p$ . We use the ArcFace model [16] to extract identity features. 2) Reference alignment $(\\mathrm{CLIP_r})$ targets at assessing the alignment of the output editing styles to the desired input prompt by computing the cosine similarity in the CLIP [50] feature space. 3) 3D consistency (3D) on the edited outputs is measured following the evaluation protocols established by EG3D [13]. This involves calculating the identity similarity across multiple views. 4) Inference speed (Time) is measured on a single NVIDIA A6000 GPU with an average of 100 samples.", + "bbox": [ + 215, + 628, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Q. Bai et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/f6c3f00a93f75379fee3f937e97790f28c47b3c39e0f7d18dc7e9c660cfd69bb.jpg", + "image_caption": [ + "Fig. 5: (a) Generated results of customized prompt adaptation and (b) its learning process. We show the intermediate testing results at 10s, 1min, 2min and 5min during adaptation for the novel style golden statue." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 782, + 268 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/840bba9faad9d542b7aabe52638d1a08a584d4deda24a2b2360447a1dbdcf07e.jpg", + "image_caption": [ + "Fig.6: Qualitative comparison for ablations on (a) the distillation loss of 3D GAN $(\\mathcal{L}_{3D})$ and (b) diffusion models $(\\mathcal{L}_{2D})$ ." + ], + "image_footnote": [], + "bbox": [ + 217, + 339, + 782, + 426 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Efficient 3D-aware Portrait Editing", + "text_level": 1, + "bbox": [ + 215, + 501, + 552, + 517 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We make an in-depth analysis of our efficient portrait editing method both quantitatively and qualitatively. The results are included in Tab. 1 and Fig. 4. For reference alignment, we do not report $CLIP_{r}$ for methods that leverage CLIP for optimization since it is evaluated with CLIP as well. The standout feature of our method is its efficiency. Our method achieves an inference speed of merely 40ms, which improves over 100 times compared to the fastest existing baselines, which require around 10 seconds. Because of the efficient knowledge distillation, our approach is also good in preserving the identity, adhering more closely to the reference prompts and achieving the best 3D consistency.", + "bbox": [ + 212, + 536, + 787, + 671 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As in Fig. 4, compared with PTI + CLIP methods, our model enables precise alignment with inputs and prompts, while the baseline always generates low-fidelity textures, and the resulting geometry has many artifacts. Although Live3D + CLIP achieves the best $ID_{t}$ score, the edited results are basically unchanged compared to the input. We suspect that Live3D does not possess a latent space and loses the editing priors, making it challenging for the results of CLIP optimization to align well with prompts. SPI + StyleGAN-NADA performs relatively slowly (18 minutes per instance) because both the 3D reconstruction and the editing are achieved by optimizing the generator. Although it can produce realistic textures, it fails to reconstruct fine details and to edit following the prompts. In contrast to the 2D editing and subsequent 3D reconstruction", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "3DPE", + "bbox": [ + 687, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/2b428d3392c8dc930ecfe504fc5d33621eb1194707c783952335a33ef29c0332.jpg", + "table_caption": [ + "Table 2: Ablation study on the number of styles used for pretraining. We report LPIPS score after 0.1min, 1min, 2min, and 5min fine-tuning for evaluation." + ], + "table_footnote": [], + "table_body": "
#Style/Time0.1min1min2min5min
30.77580.63700.53120.5220
100.59320.57340.51060.5012
200.57030.52900.50220.4849
", + "bbox": [ + 290, + 184, + 710, + 241 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/95e797f49ac6c1caea0803ad325b1a3e98abe0155c5168fc3137fa0ae42c6b82.jpg", + "table_caption": [ + "Table 3: Ablation study on the number of data pairs used for novel prompt adaptation. We report LPIPS score after 0.1min, 1min, 2min, and 5min fine-tuning for evaluation." + ], + "table_footnote": [], + "table_body": "
#Pair/Time0.1min1min2min5min
20.61910.54310.52130.5212
50.61480.53240.50640.4971
100.60760.52420.49830.4935
200.59810.55250.49540.4895
500.58910.57180.49220.4797
", + "bbox": [ + 292, + 309, + 709, + 387 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "pipeline (IP2P + Live3D), our method produces textures more consistent with the prompt and aligns better with the input structure.", + "bbox": [ + 215, + 414, + 785, + 445 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Adaptation for Customized Editing", + "text_level": 1, + "bbox": [ + 215, + 465, + 553, + 482 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "With the trained editing network, our method already supports a variety of 3D-aware editing styles. Following DreamBooth [57] and to expand the selection and better conform to users' preferences, we offer an efficient method for fast adaptation to the user-defined customized prompts. Users can personalize the editing network by providing a modest set of 10 reference editing pairs. These pairs can be either handpicked from artist-created examples or produced using text-guided image editing models. As described in Sec. 3.3, only the partial of the encoder would be optimized. The adaptation process itself is remarkably fast, requiring only about 5 minutes to accomplish the learning of the customized prompts. Please also note that, once the user-defined style has been trained, all the incoming testing results can be obtained in a real-time manner. In Fig. 5, we present the testing results and learning process of the adaptation. Our model can quickly master the customized knowledge in about 2 minutes. With further training (e.g., 5 minutes), the edited results become more stylized, demonstrating a trade-off between authenticity and stylization. Upon completion of the adaptation, the method allows users to edit testing inputs in these newly learned styles with a minimal inference time of 0.04s. This rapid performance indicates that the method is well-optimized for real-time applications, providing a seamless and efficient user experience.", + "bbox": [ + 215, + 489, + 785, + 777 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 800, + 401, + 814 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We analyze our model and validate our design choice by ablating the components.", + "bbox": [ + 215, + 824, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Q. Bai et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/23913448181367f3e27cd2f751b0607ddf42b7e0911b6618b2faea15af9fe67c.jpg", + "image_caption": [ + "Fig. 7: Video editing results with 3DPE. We use text prompts \"Make him a bronze statue\" to edit the input video. Our method can accurately reconstruct and preserve challenging facial expressions and achieve high-quality novel view renderings." + ], + "image_footnote": [], + "bbox": [ + 230, + 146, + 767, + 385 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{3D}$ for Distillation of Live3D. We compare our model trained with and without $\\mathcal{L}_{3D}$ . As shown in Fig. 6, we observe that the model without $\\mathcal{L}_{3D}$ results in very flat geometry, and the synthesis of the side view loses the normal structure with many artifacts. This highlights the importance of the 3D knowledge from Live3D in training our model with 2D paired data, enabling the inference of reasonable face geometry without relying on any 3D data sources.", + "bbox": [ + 212, + 482, + 787, + 573 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{2D}$ for Distillation of Diffusion Models. We also study the effect of $\\mathcal{L}_{2D}$ , which represents using 2D editing priors from the diffusion model [7] to optimize the encoder. From Fig. 6, we observe that $\\mathcal{L}_{2D}$ is critical for our model, especially in preserving detailed texture. Without it, the edited image tends to lose the appearance information from prompts and exhibits structural artifacts around the nose.", + "bbox": [ + 212, + 577, + 787, + 667 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation on Pretraining Data Scale for Customized Prompt Adaptations. In this part, we conduct ablation studies on the pretraining scale and investigate its influence on customized propmts. Specifically, we choose various 3DPE models pretrained with different style amounts (3, 10, and 20 styles respectively) and perform novel prompt adaptation on a single A6000 GPU. We choose the golden statue as the customized style and report LPIPS score between the predicted stylized images and the pseudo stylized labels produced by IP2P [7]. Experimental results in Tab. 2 reveal that the proposed method could benefit from the improving pretraining data scale and adapt faster and eventually perform better with more data, which suggests the potential of the model for future scaling up.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "3DPE", + "bbox": [ + 687, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation on Tuning Data Scale for Customized Prompt Adaptations.", + "text_level": 1, + "bbox": [ + 215, + 146, + 785, + 161 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "To investigate the ability of our method to adapt to user-defined novel prompts, we conduct an ablation study on the number of novel-prompt data pairs required for training. Similar to the prior experiments on the pretraining scale, we perform adaptation with 2, 5, 10, 20, or 50 pairs of novel-prompt data on a single A6000 GPU. We experimentally find that our model can converge in about 5 minutes, and therefore, we report the perceptual distance between the model output and the provided ground truth at $0.1\\mathrm{min}$ , $1\\mathrm{min}$ , $2\\mathrm{min}$ , and $5\\mathrm{min}$ for comparison. As shown in Tab. 3, our model can adapt to the new style in around 2 minutes. As the number of provided data pairs increases, the model can adapt faster at the beginning.", + "bbox": [ + 215, + 162, + 785, + 311 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.5 Applications", + "text_level": 1, + "bbox": [ + 215, + 335, + 364, + 349 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "One application of our pipeline is to edit videos, requiring accurate reconstruction and efficient editing simultaneously. We demonstrate the effectiveness of our pipeline on talking face videos, as illustrated in Fig. 7. In this experiment, we perform sequence editing in a frame-by-frame manner by sending an image sequence from TalkingHead-1KH [72] dataset together with the text prompt \"Make him a bronze statue\" to 3DPE. The three rows in Fig. 7 respectively illustrate the input frames, the edited frames, and the novel views. Our method accurately reconstructs facial expressions, while achieving high-quality editing results. Since our method is 3D-aware, we can naturally view videos from novel perspectives. To better experience our method, we design an interactive system that allows users for interactive editing. Users are required to provide the image or text prompt to indicate the desired style and input an input image that is to be edited, and the system will apply our method on the given image and prompt to output the edited image as well as novel-view edited results.", + "bbox": [ + 215, + 361, + 785, + 571 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Discussion", + "text_level": 1, + "bbox": [ + 215, + 595, + 352, + 611 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. Despite achieving state-of-the-art performance in quality and efficiency, our method exhibits inconsistencies in details for novel-view rendering because the EG3D framework relies on a super-resolution module. Additionally, when our method is applied to video editing, it presents flickering artifacts since our model is originally designed for image-based editing.", + "bbox": [ + 215, + 628, + 785, + 703 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Conclusion. In this work, we introduce 3DPE for real-time 3D-aware portrait editing. By distilling the powerful knowledge of diffusion models and 3D GANs into a lightweight module, our model significantly reduces editing time while ensuring quality. Benefiting from the distillation process, 3DPE could handle multiple styles in a single model and perform both image-guided and text-guided editing. The advantages of our method empower us to generate photorealistic 3D portraits, a capability crucial for the visual effects industry, AR/VR systems, and teleconferencing, among other applications, while we firmly oppose any misuse that could compromise privacy and security.", + "bbox": [ + 215, + 704, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Q. Bai et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This project was supported by the National Key R&D Program of China under grant number 2022ZD0161501.", + "bbox": [ + 212, + 178, + 785, + 209 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 233, + 323, + 250 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abdal, R., Lee, H.Y., Zhu, P., Chai, M., Siarohin, A., Wonka, P., Tulyakov, S.: 3DAvatarGAN: Bridging domains for personalized editable avatars. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)", + "2. Abdal, R., Qin, Y., Wonka, P.: Image2StyleGAN: How to embed images into the StyleGAN latent space? In: Int. Conf. Comput. Vis. (2019)", + "3. Abdal, R., Qin, Y., Wonka, P.: Image2StyleGAN++: How to edit the embedded images? In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)", + "4. Aneja, S., Thies, J., Dai, A., Nießner, M.: ClipFace: Text-guided editing of textured 3d morphable models. In: ACM SIGGRAPH Conf. Proc. (2023)", + "5. Bai, Q., Xu, Y., Zhu, J., Xia, W., Yang, Y., Shen, Y.: High-fidelity GAN inversion with padding space. In: Eur. Conf. Comput. Vis. (2022)", + "6. Blanz, V., Vetter, T.: A morphable model for the synthesis of 3d faces. In: Seminal Graphics Papers: Pushing the Boundaries (2023)", + "7. Brooks, T., Holynski, A., Efros, A.A.: InstructPix2Pix: Learning to follow image editing instructions. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)", + "8. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: Adv. Neural Inform. Process. Syst. (2020)", + "9. Cai, S., Obukhov, A., Dai, D., Van Gool, L.: Pix2NeRF: Unsupervised conditional p-gan for single image to neural radiance fields translation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)", + "0. Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., Zheng, Y.: MasaCtrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In: Int. Conf. Comput. Vis. (2023)", + "1. Ceylan, D., Huang, C.H.P., Mitra, N.J.: Pix2Video: Video editing using image diffusion. In: Int. Conf. Comput. Vis. (2023)", + "2. Chai, W., Guo, X., Wang, G., Lu, Y.: StableVideo: Text-driven consistency-aware diffusion video editing. In: Int. Conf. Comput. Vis. (2023)", + "3. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)", + "4. Chan, E.R., Monteiro, M., Kellnhofer, P., Wu, J., Wetzstein, G.: pi-GAN: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)", + "5. Chang, S., Kim, G., Kim, H.: HairNeRF: Geometry-aware image synthesis for hairstyle transfer. In: Int. Conf. Comput. Vis. (2023)", + "6. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: ArcFace: Additive angular margin loss for deep face recognition. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019)", + "7. Deng, Y., Yang, J., Chen, D., Wen, F., Tong, X.: Disentangled and controllable face image generation via 3d imitative-contrastive learning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)" + ], + "bbox": [ + 225, + 266, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "3DPE", + "bbox": [ + 687, + 114, + 732, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Deng, Y., Yang, J., Xu, S., Chen, D., Jia, Y., Tong, X.: Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019)", + "19. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Int. Conf. Learn. Represent. (2021)", + "20. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. In: Int. Conf. Learn. Represent. (2023)", + "21. Gal, R., Patashnik, O., Maron, H., Bermano, A.H., Chechik, G., Cohen-Or, D.: StyleGAN-NADA: Clip-guided domain adaptation of image generators. ACM Trans. Graph. (2022)", + "22. Gao, C., Shih, Y., Lai, W.S., Liang, C.K., Huang, J.B.: Portrait neural radiance fields from a single image. arXiv preprint arXiv:2012.05903 (2020)", + "23. Geyer, M., Bar-Tal, O., Bagon, S., Dekel, T.: TokenFlow: Consistent diffusion features for consistent video editing. In: Int. Conf. Learn. Represent. (2024)", + "24. Graikos, A., Malkin, N., Jojic, N., Samaras, D.: Diffusion models as plug-and-play priors. In: Adv. Neural Inform. Process. Syst. (2022)", + "25. Gu, J., Liu, L., Wang, P., Theobalt, C.: StyleNeRF: A style-based 3d-aware generator for high-resolution image synthesis. In: Int. Conf. Learn. Represent. (2022)", + "26. Guo, J., Yu, J., Lattas, A., Deng, J.: Perspective reconstruction of human faces by joint mesh and landmark regression. In: Eur. Conf. Comput. Vis. (2022)", + "27. Haque, A., Tancik, M., Efros, A.A., Holynski, A., Kanazawa, A.: Instruct-NeRF2NeRF: Editing 3d scenes with instructions. In: Int. Conf. Comput. Vis. (2023)", + "28. Harkonen, E., Hertzmann, A., Lehtinen, J., Paris, S.: GANSpace: Discovering interpretable GAN controls. Adv. Neural Inform. Process. Syst. (2020)", + "29. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)", + "30. Hertz, A., Aberman, K., Cohen-Or, D.: Delta denoising score. In: Int. Conf. Comput. Vis. (2023)", + "31. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: Int. Conf. Learn. Represent. (2022)", + "32. Hyung, J., Hwang, S., Kim, D., Lee, H., Choo, J.: Local 3d editing via 3d distillation of clip knowledge. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)", + "33. Jiang, K., Chen, S.Y., Fu, H., Gao, L.: NeRFFaceLighting: Implicit and disentangled face lighting representation leveraging generative prior in neural radiance fields. ACM Trans. Graph. (2023)", + "34. Jin, W., Ryu, N., Kim, G., Baek, S.H., Cho, S.: Dr. 3D: Adapting 3d GANs to artistic drawings. In: ACM SIGGRAPH Asia Conf. Proc. (2022)", + "35. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019)", + "36. Kim, G., Chun, S.Y.: DATID-3D: Diversity-preserved domain adaptation using text-to-image diffusion for 3d generative model. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)", + "37. Ko, J., Cho, K., Choi, D., Ryoo, K., Kim, S.: 3D GAN inversion with pose optimization. In: IEEE Winter Conf. Appl. Comput. Vis. (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Q. Bai et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "38. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)", + "39. Li, J., Li, J., Zhang, H., Liu, S., Wang, Z., Xiao, Z., Zheng, K., Zhu, J.: PREIM3D: 3d consistent precise image attribute editing from a single image. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)", + "40. Li, J., Liu, S., Liu, Z., Wang, Y., Zheng, K., Xu, J., Li, J., Zhu, J.: Instruct-Pix2NeRF: Instructed 3d portrait editing from a single image. arXiv preprint arXiv:2311.02826 (2023)", + "41. Li, J., Li, D., Xiong, C., Hoi, S.: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: Int. Conf. Mach. Learn. (2022)", + "42. Lin, C., Lindell, D., Chan, E., Wetzstein, G.: 3D GAN inversion for controllable portrait image animation. In: Eur. Conf. Comput. Vis. Workshop (2022)", + "43. Liu, S., Zhang, Y., Li, W., Lin, Z., Jia, J.: Video-P2P: Video editing with cross-attention control. arXiv preprint arXiv:2303.04761 (2023)", + "44. Liu, Z., Feng, R., Zhu, K., Zhang, Y., Zheng, K., Liu, Y., Zhao, D., Zhou, J., Cao, Y.: Cones: Concept neurons in diffusion models for customized generation. In: Int. Conf. Mach. Learn. (2023)", + "45. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: Int. Conf. Learn. Represent. (2022)", + "46. Nitzan, Y., Gharbi, M., Zhang, R., Park, T., Zhu, J.Y., Cohen-Or, D., Shechtman, E.: Domain expansion of image generators. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)", + "47. Or-El, R., Luo, X., Shan, M., Shechtman, E., Park, J.J., Kemelmacher-Shlizerman, I.: StyleSDF: High-resolution 3d-consistent image and geometry generation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)", + "48. Ouyang, H., Wang, Q., Xiao, Y., Bai, Q., Zhang, J., Zheng, K., Zhou, X., Chen, Q., Shen, Y.: CoDeF: Content deformation fields for temporally consistent video processing. arXiv preprint arXiv:2308.07926 (2023)", + "49. Pan, X., Xu, X., Loy, C.C., Theobalt, C., Dai, B.: A shading-guided generative implicit model for shape-accurate 3d-aware image synthesis. Adv. Neural Inform. Process. Syst. (2021)", + "50. Patashnik, O., Wu, Z., Shechtman, E., Cohen-Or, D., Lischinski, D.: StyleCLIP: Text-driven manipulation of StyleGAN imagery. In: Int. Conf. Comput. Vis. (2021)", + "51. Paysan, P., Knothe, R., Amberg, B., Romdhani, S., Vetter, T.: A 3d face model for pose and illumination invariant face recognition. In: IEEE international conference on advanced video and signal based surveillance (2009)", + "52. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: DreamFusion: Text-to-3d using 2d diffusion. In: Int. Conf. Learn. Represent. (2023)", + "53. Qi, C., Cun, X., Zhang, Y., Lei, C., Wang, X., Shan, Y., Chen, Q.: FateZero: Fusing attentions for zero-shot text-based video editing. In: Int. Conf. Comput. Vis. (2023)", + "54. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Int. Conf. Mach. Learn. (2021)", + "55. Roich, D., Mokady, R., Bermano, A.H., Cohen-Or, D.: Pivotal tuning for latent-based editing of real images. ACM Trans. Graph. (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "3DPE", + "bbox": [ + 687, + 114, + 732, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "56. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)", + "57. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dream-Booth: Fine tuning text-to-image diffusion models for subject-driven generation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)", + "58. Schwarz, K., Liao, Y., Niemeyer, M., Geiger, A.: GRAF: Generative radiance fields for 3d-aware image synthesis. In: Adv. Neural Inform. Process. Syst. (2020)", + "59. Schwarz, K., Sauer, A., Niemeyer, M., Liao, Y., Geiger, A.: VoxGRAF: Fast 3d-aware image synthesis with sparse voxel grids. In: Adv. Neural Inform. Process. Syst. (2022)", + "60. Shen, Y., Gu, J., Tang, X., Zhou, B.: Interpreting the latent space of GANs for semantic face editing. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)", + "61. Shen, Y., Yang, C., Tang, X., Zhou, B.: InterFaceGAN: Interpreting the disentangled face representation learned by GANs. IEEE Trans. Pattern Anal. Mach. Intell. (2020)", + "62. Shi, Z., Peng, S., Xu, Y., Geiger, A., Liao, Y., Shen, Y.: Deep generative models on 3d representations: A survey. arXiv preprint arXiv:2210.15663 (2022)", + "63. Shi, Z., Shen, Y., Xu, Y., Peng, S., Liao, Y., Guo, S., Chen, Q., Yeung, D.Y.: Learning 3d-aware image synthesis with unknown pose distribution. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)", + "64. Shi, Z., Xu, Y., Shen, Y., Zhao, D., Chen, Q., Yeung, D.Y.: Improving 3d-aware image synthesis with a geometry-aware discriminator. In: Adv. Neural Inform. Process. Syst. (2022)", + "65. Skorokhodov, I., Tulyakov, S., Wang, Y., Wonka, P.: EpiGRAF: Rethinking training of 3d gans. In: Adv. Neural Inform. Process. Syst. (2022)", + "66. Sun, J., Wang, X., Shi, Y., Wang, L., Wang, J., Liu, Y.: IDE-3D: Interactive disentangled editing for high-resolution 3d-aware portrait synthesis. ACM Trans. Graph. (2022)", + "67. Sun, J., Wang, X., Zhang, Y., Li, X., Zhang, Q., Liu, Y., Wang, J.: FENeRF: Face editing in neural radiance fields. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)", + "68. Tov, O., Alaluf, Y., Nitzan, Y., Patashnik, O., Cohen-Or, D.: Designing an encoder for StyleGAN image manipulation. ACM Trans. Graph. (2021)", + "69. Tran, P., Zakharov, E., Ho, L.N., Tran, A.T., Hu, L., Li, H.: VOODOO 3D: Volumetric portrait disentanglement for one-shot 3d head reenactment. In: IEEE Conf. Comput. Vis. Pattern Recog. (2024)", + "70. Trevithick, A., Chan, M., Stengel, M., Chan, E., Liu, C., Yu, Z., Khamis, S., Chandraker, M., Ramamoorthi, R., Nagano, K.: Real-time radiance fields for single-image portrait view synthesis. ACM Trans. Graph. (2023)", + "71. Wang, Q., Shi, Z., Zheng, K., Xu, Y., Peng, S., Shen, Y.: Benchmarking and analyzing 3d-aware image synthesis with a modularized codebase. In: Adv. Neural Inform. Process. Syst. (2023)", + "72. Wang, T.C., Mallya, A., Liu, M.Y.: One-shot free-view neural talking-head synthesis for video conferencing. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)", + "73. Wang, W., Jiang, Y., Xie, K., Liu, Z., Chen, H., Cao, Y., Wang, X., Shen, C.: Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599 (2023)", + "74. Wang, Y., Xu, T., Wu, Y., Li, M., Chen, W., Xu, L., Yu, J.: NARRATE: A normal assisted free-view portrait stylizer. arXiv preprint arXiv:2207.00974 (2022)" + ], + "bbox": [ + 215, + 147, + 784, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Q. Bai et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "75. Xia, W., Xue, J.H.: A survey on deep generative 3d-aware image synthesis. ACM Computing Surveys (2023)", + "76. Xie, J., Ouyang, H., Piao, J., Lei, C., Chen, Q.: High-fidelity 3d GAN inversion by pseudo-multi-view optimization. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)", + "77. Xu, Y., Peng, S., Yang, C., Shen, Y., Zhou, B.: 3d-aware image synthesis via learning structural and textural representations. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)", + "78. Xu, Y., Shen, Y., Zhu, J., Yang, C., Zhou, B.: Generative hierarchical features from synthesizing images. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)", + "79. Yang, S., Jiang, L., Liu, Z., Loy, C.C.: Pastiche master: Exemplar-based high-resolution portrait style transfer. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)", + "80. Yang, S., Zhou, Y., Liu, Z.,, Loy, C.C.: Rerender a video: Zero-shot text-guided video-to-video translation. In: ACM SIGGRAPH Asia Conf. Proc. (2023)", + "81. Yin, F., Zhang, Y., Wang, X., Wang, T., Li, X., Gong, Y., Fan, Y., Cun, X., Shan, Y., Oztireli, C., et al.: 3D GAN inversion with facial symmetry prior. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)", + "82. Zhang, J., Lan, Y., Yang, S., Hong, F., Wang, Q., Yeo, C.K., Liu, Z., Loy, C.C.: DeformToon3D: Deformable neural radiance fields for 3d toonification. In: Int. Conf. Comput. Vis. (2023)", + "83. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Int. Conf. Comput. Vis. (2023)", + "84. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018)", + "85. Zhou, P., Xie, L., Ni, B., Tian, Q.: CIPS-3D: A 3d-aware generator of gans based on conditionally-independent pixel synthesis. arXiv preprint arXiv:2110.09788 (2021)", + "86. Zhu, J., Shen, Y., Zhao, D., Zhou, B.: In-domain GAN inversion for real image editing. In: Eur. Conf. Comput. Vis. (2020)", + "87. Zhu, J.Y., Krahenbuhl, P., Shechtman, E., Efros, A.A.: Generative visual manipulation on the natural image manifold. In: Eur. Conf. Comput. Vis. (2016)" + ], + "bbox": [ + 212, + 146, + 784, + 590 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "3DPE", + "bbox": [ + 687, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Real-time 3D-aware Portrait Editing from a Single Image/c7f6568e-666e-490d-b059-9d93ade8c6f3_model.json b/2024/Real-time 3D-aware Portrait Editing from a Single Image/c7f6568e-666e-490d-b059-9d93ade8c6f3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fa5836acfbe655cc732e3f4659cbc50cda2e5bf7 --- /dev/null +++ b/2024/Real-time 3D-aware Portrait Editing from a Single Image/c7f6568e-666e-490d-b059-9d93ade8c6f3_model.json @@ -0,0 +1,2801 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.292, + 0.141, + 0.711, + 0.187 + ], + "angle": 0, + "content": "Real-time 3D-aware Portrait Editing from a Single Image" + }, + { + "type": "text", + "bbox": [ + 0.292, + 0.211, + 0.709, + 0.23 + ], + "angle": 0, + "content": "Qingyan Bai\\(^{1,2}\\), Zifan Shi\\(^{1}\\), Yinghao Xu\\(^{3}\\), Hao Ouyang\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.229, + 0.66, + 0.244 + ], + "angle": 0, + "content": "Qiuyu Wang\\(^{2}\\), Ceyuan Yang\\(^{4}\\), Xuan Wang\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.306, + 0.244, + 0.696, + 0.259 + ], + "angle": 0, + "content": "Gordon Wetzstein3, Yujun Shen2, and Qifeng Chen1" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.27, + 0.688, + 0.285 + ], + "angle": 0, + "content": "1 The Hong Kong University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.457, + 0.284, + 0.546, + 0.298 + ], + "angle": 0, + "content": "2 Ant Group" + }, + { + "type": "text", + "bbox": [ + 0.428, + 0.298, + 0.575, + 0.312 + ], + "angle": 0, + "content": "3 Stanford University" + }, + { + "type": "text", + "bbox": [ + 0.413, + 0.312, + 0.59, + 0.326 + ], + "angle": 0, + "content": "4 Shanghai AI Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.36, + 0.74, + 0.553 + ], + "angle": 0, + "content": "Abstract. This work presents 3DPE, a practical method that can efficiently edit a face image following given prompts, like reference images or text descriptions, in a 3D-aware manner. To this end, a lightweight module is distilled from a 3D portrait generator and a text-to-image model, which provide prior knowledge of face geometry and superior editing capability, respectively. Such a design brings two compelling advantages over existing approaches. First, our method achieves real-time editing with a feedforward network (i.e., \\(\\sim 0.04\\)s per image), over \\(100\\times\\) faster than the second competitor. Second, thanks to the powerful priors, our module could focus on the learning of editing-related variations, such that it manages to handle various types of editing simultaneously in the training phase and further supports fast adaptation to user-specified customized types of editing during inference (e.g., with \\(\\sim 5\\) min fine-tuning per style). Project page can be found here." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.567, + 0.591, + 0.58 + ], + "angle": 0, + "content": "Keywords: 3D-aware portrait \\(\\cdot\\) Efficient editing" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.605, + 0.375, + 0.621 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.636, + 0.787, + 0.786 + ], + "angle": 0, + "content": "Inferring the geometry and appearance from a single-view portrait image has become mature and practical [22, 37, 42, 66, 70, 74, 76], largely attributed to the utilization of priors in various 2D/3D generative models. However, only performing geometry reconstruction is insufficient. The significance of 3D portrait editing, driven by user intentions, and the need for streamlined efficiency in the editing process has been steadily increasing. This is particularly crucial in real-world applications such as AR/VR, 3D telepresence, and video conferencing, where real-time editing is often essential. Consequently, a key question arises: How can we effectively address the challenge of attaining high-fidelity portrait editing while ensuring real-time efficiency?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.787, + 0.787, + 0.817 + ], + "angle": 0, + "content": "Traditional methods for 3D portrait editing [17] typically rely on template facial models [6,51], which has limitations in handling the substantial geometry" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.392, + 0.84 + ], + "angle": 0, + "content": "* Corresponding authors." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Q. Bai et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.146, + 0.506, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.147, + 0.788, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.334, + 0.788, + 0.406 + ], + "angle": 0, + "content": "Fig. 1: Photorealistic editing results produced by our proposed 3DPE, which allows users to perform 3D-aware portrait editing using image or text prompts. In comparison with baseline methods, such as InstructPix2Pix (IP2P) [7] + Live3D [70] and PTI [55] + StyleCLIP [50] (details are illustrated in Sec. 4), our approach accurately follows the guidance from reference prompts and maintains sufficiently better efficiency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.44, + 0.788, + 0.577 + ], + "angle": 0, + "content": "changes because it often overlooks precise features like hair and beard. Recent 3D GANs [13, 14, 58, 62, 75] show remarkable capabilities in generating high-fidelity 3D portraits. They can serve as powerful generative priors for 3D portrait editing when coupled with GAN inversion techniques [2, 3, 61, 86, 87]. However, these methods may encounter issues related to geometry distortion or exhibit slow speed, and the editing is constrained by the limited latent attributes. Besides, 2D diffusion models can offer a strong prior with Score Distillation Sampling (SDS) loss [52] for editing purposes [30]. Nonetheless, they often require step-by-step optimization and become a major bottleneck for real-time applications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.581, + 0.789, + 0.791 + ], + "angle": 0, + "content": "To this end, we present 3DPE, a real-time 3D-aware Portrait Editing method driven by user-defined prompts. As shown in Fig. 1, when provided with a single-view portrait image, our approach empowers a versatile range of editing styles with one model through flexible instructions, including images and texts. We leverage the powerful 3D prior from a 3D-aware face generator and achieve a high-fidelity 3D reconstruction of the portrait image. Subsequently, we distill editing knowledge from a text-guided image editing model into a lightweight module integrated with the 3D-aware generator. This module, characterized by its minimal computational cost, allows our method to maintain real-time inference and excel in handling various types of editing, while ensuring good 3D consistency. An additional advantage is that our model supports customization through user-specified prompts with fast adaptation speed. This empowers users to build their own editing model at a minimal cost, enabling our method to cater to a broader audience." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Our method achieves real-time 3D-aware portrait editing through the utilization of a feedforward network, with a processing time of \\(40\\mathrm{ms}\\) on a standard consumer GPU. Additionally, we present a comprehensive evaluation of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.689, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "3DPE" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.253 + ], + "angle": 0, + "content": "our method using various prompts both quantitatively and qualitatively. Our design choices are also validated through comparisons with ablated variants of our method. Compared with baseline methods, our approach demonstrates superior 3D consistency, precise texture alignment, and a substantial improvement in inference time, as reflected by the evaluation metrics. We demonstrate the versatility of our method by showcasing its capacity to perform a wide variety of edits, including text and image prompts, on portrait images." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.254, + 0.609, + 0.267 + ], + "angle": 0, + "content": "In summary, the contributions of our work include:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.279, + 0.784, + 0.323 + ], + "angle": 0, + "content": "- We propose a lightweight module to distill knowledge from 3D GANs and diffusion models for 3D-aware editing from a single image. Due to the minimal cost of the new module, our model maintains real-time performance." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.324, + 0.784, + 0.353 + ], + "angle": 0, + "content": "- Our model supports fast adaptation to user-specified editing, requiring only 10 image pairs and 5 minutes for the adaptation." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.354, + 0.784, + 0.383 + ], + "angle": 0, + "content": "- Our framework can accommodate various control signals, including text and image prompts." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.279, + 0.784, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.407, + 0.387, + 0.423 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.438, + 0.452, + 0.451 + ], + "angle": 0, + "content": "2.1 Generative Face Priors" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.785, + 0.644 + ], + "angle": 0, + "content": "Generative models aim at modeling the underlying distribution of the data, containing a wealth of prior knowledge. Recently, 3D GANs [13, 14, 25, 47, 49, 58, 59, 63-65, 69, 71, 77] are mostly adopted to learn 3D faces from single-view image dataset. The interior rich domain-specific geometry priors enable various face-related applications such as image editing [9, 66, 67, 70, 79, 85] and domain adaptation [1, 34, 36, 46, 82]. Large-scale diffusion models [56], in contrast, encode knowledge of huge datasets and thus can provide general prior information. Such priors are broadly leveraged for tasks such as image editing [7, 10, 24, 45, 83], customization [31, 38, 44, 57], and video editing [11, 12, 23, 43, 48, 53, 73, 80]. In our approach, we capitalize on the advantages of both the geometry prior derived from 3D GANs and the broader editing prior offered by large-scale text-to-image diffusion models, instead of relying solely on a single type of prior." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.665, + 0.567, + 0.68 + ], + "angle": 0, + "content": "2.2 Portrait Editing from a Single Image" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Although many methods work well for reconstructing [22, 26] or generating [13] faces, editing has become a necessary interface to connect these methods with real-world applications. Previously, portrait editing given a single-view image was mostly completed in 2D image space facilitated with GAN inversion [2, 3, 5, 28, 55, 60, 61, 78, 86], which enables fast editing by exploring the trajectories in the GAN's latent space. Recent diffusion-based portrait editing [7, 83] can support various editing types with texts as guidance. However, most of them are done in the 2D space, and thus there is no guarantee for the underlying 3D consistency. Therefore, 3D-aware portrait editing is crucial to achieve the goal. Some methods [15, 33, 39, 42, 66, 67, 74, 76, 81] rely on the latent space" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.128 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.129 + ], + "angle": 0, + "content": "Q. Bai et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.375 + ], + "angle": 0, + "content": "of 3D GANs to perform editing through walking in the latent space, but are limited by the number of editable latent attributes. A few methods, such as Instruct-NeRF2NeRF [27], ClipFace [4], and LENeRF [32], attempt to leverage the large-scale models [54, 56] to edit on 3D representations with texts as guidance but require heavy iterative refinement of the edited results. Similarly, InstructPix2NeRF [40] leverages text-to-image models and text prompts to edit within the latent space of the GAN-based 3D generator. However, it is held back by the multi-step diffusion inference and the heavy structure of a GAN-based generator. Ours, in contrast, benefits from the lightweight design of the module and achieves real-time editing with ease. With such a design, our model can also adapt fast to the user-defined editing prompts. Moreover, texts cannot always illustrate the desired effects precisely, and sometimes an image prompt serves as better guidance for image editing [20, 57]. Our method can support not only text prompts but also image prompts for editing, providing a more flexible and user-friendly interface." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.397, + 0.333, + 0.414 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.79, + 0.612 + ], + "angle": 0, + "content": "Our approach takes a portrait image \\(\\mathbf{I}\\) and its camera pose \\(\\mathbf{c}\\), obtained via the face pose estimator [18]. Additionally, the referenced prompts \\(\\mathbf{P}\\) (images or texts) serve as the editing instructions. The outcome of our model is an edited version of the 3D portraits characterized by NeRF in accordance with the prompts. The edited image is denoted as \\(\\mathbf{I}_p\\). Within our framework, we achieve this objective by leveraging the knowledge of 3D GANs and a text-guided image editing model, which is distilled into a lightweight module. Following distillation, our method enables real-time 3D portrait editing and efficient adaptation to user-specified prompts. In Sec. 3.1, we offer an illustration of 3D GANs and diffusion models. The process of distilling these priors into a lightweight module is outlined in Sec. 3.2. Finally, the details for model training and inference, along with the fast adaptation for customized prompts, are presented in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.633, + 0.363, + 0.65 + ], + "angle": 0, + "content": "3.1 Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.842 + ], + "angle": 0, + "content": "3D GAN Prior for Portrait Reconstruction. The 3D-aware GANs showcase the ability to synthesize photorealistic 3D images using a collection of single-view images. Notably, EG3D [13] introduces an efficient triplane 3D representation, demonstrating high-quality 3D-aware image rendering. Once trained, the generator of EG3D can be applied for single-image 3D reconstruction via GAN inversion. However, existing 3D GAN inversion methods often encounter geometry distortion or exhibit slow inference speed. To tackle these challenges, we utilize Live3D [70], a state-of-the-art single-image portrait 3D reconstruction model built upon EG3D, preserving geometry quality while ensuring real-time performance. Live3D employs a two-branch encoder, \\(\\mathrm{E}_{high}(\\cdot)\\) and \\(\\mathrm{E}_{low}(\\cdot)\\), to extract different resolution features from the input. It then utilizes a ViT-based" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.689, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "3DPE" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.147, + 0.773, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.375, + 0.788, + 0.513 + ], + "angle": 0, + "content": "Fig. 2: Overview of our method. We distill the prior in the diffusion model and 3D GAN for real-time 3D-aware editing. Our approach is fine-tuned from Live3D [70], where we extract features from the input portrait I using \\(\\mathrm{E}_{high}(\\cdot)\\) and \\(\\mathrm{E}_{low}(\\cdot)\\). The prompt embedding is generated with \\(\\mathrm{E}_p(\\cdot)\\) and injected with the input features from \\(\\mathrm{E}_{high}(\\cdot)\\) through a cross-attention mechanism. Our model is trained to mimic the output from the diffusion prior to acquire editing knowledge and enforce geometry constraints through triplane, multi-view images, and depth supervision from the 3D prior. In this context, InstructPix2Pix [7] and Live3D serve as the diffusion and 3D prior, respectively. It's noteworthy that only \\(\\mathrm{E}_p(\\cdot)\\) and \\(\\mathrm{E}_t(\\cdot)\\) are learnable during training, while all other parameters remain frozen." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.548, + 0.785, + 0.565 + ], + "angle": 0, + "content": "\\(\\mathrm{E}_t(\\cdot)\\) decoder [19] to transform the fused encoder features into the triplane \\(\\mathbf{T}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.595, + 0.786, + 0.612 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} = \\mathrm {E} _ {t} \\left(\\mathrm {E} _ {\\text {h i g h}} (\\mathbf {I}), \\mathrm {E} _ {\\text {l o w}} (\\mathbf {I})\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.785, + 0.672 + ], + "angle": 0, + "content": "This triplane \\(\\mathbf{T}\\) is subsequently used in conjunction with the volume rendering module and upsampler of EG3D to generate photorealistic view synthesis given the camera pose c. For clarity, we use \\(\\mathrm{R}(\\cdot)\\) to denote this process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Diffusion Prior for Portrait Editing. Only performing 3D reconstruction is insufficient as our goal is to edit portrait images. Large-scale diffusion models [56], trained on vast text-image pairs, can synthesize realistic photos with text input, offering powerful editing prior. Despite their effectiveness for face images, the editing process is often slow due to step-by-step optimization or multiple inferences in diffusion models. The scarcity of high-quality and multi-view consistent paired data significantly hinders the practical application of 3D-aware editing from a single image, which is hard to acquire with diffusion models. Therefore, our goal is to distill editing knowledge from the diffusion model, integrate it with the 3D prior in Live3D into a lightweight module, and employ it for real-time portrait editing." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "Q. Bai et al." + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.149, + 0.756, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.405, + 0.788, + 0.475 + ], + "angle": 0, + "content": "Fig. 3: Disentanglement in Live3D features. We separately disable the features from \\(\\mathrm{E}_{high}(\\cdot)\\) and \\(\\mathrm{E}_{low}(\\cdot)\\) to infer the reconstructed image. Without \\(\\mathrm{E}_{high}(\\cdot)\\), the output retains the coarse structure but loses its appearance. Conversely, when \\(\\mathrm{E}_{low}(\\cdot)\\) is deactivated, the reconstructed portraits preserve the texture (such as the blue and purple reflection on the glasses) but fail to capture the geometry." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.507, + 0.618, + 0.523 + ], + "angle": 0, + "content": "3.2 Distilling Priors into a Lightweight Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.643 + ], + "angle": 0, + "content": "Our method aims to perform real-time 3D-aware editing for various prompts for the single-view portrait. Thus, it needs powerful 3D knowledge for geometry reconstruction and editing prior to handling various control signals. We leverage the strengths of both 3D GANs and diffusion models. In the following, we provide detailed presentations on how the knowledge of these two types of models is distilled into a lightweight module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Feature Representation in Live3D. We carefully study the Live3D model and discover that, as a two-branch triplane-based 3D reconstruction model, the low-resolution and high-resolution features from the Live3D encoder \\(\\mathrm{E}_{low}(\\cdot)\\) and \\(\\mathrm{E}_{high}(\\cdot)\\) tend to learn various levels of information without explicit guidance. We conduct a study by disabling one of the branches and inferring the reconstructed images. As illustrated in Fig. 3, when the high-resolution encoder \\(\\mathrm{E}_{high}(\\cdot)\\) is disabled with random weights, the inference image retains a similar structure but loses its detailed appearance. Conversely, when the low-resolution encoder \\(\\mathrm{E}_{low}(\\cdot)\\) is disabled, the reconstructed portraits preserve some of the texture from the input but struggle to capture the geometry. Based on this analysis, the features of the two branches separately model low-frequency and high-frequency information. This insight motivates our model design in terms of how to distill knowledge from the diffusion model and 3D GANs." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.688, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "3DPE" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.265 + ], + "angle": 0, + "content": "Geometry Prediction for Inputs I. In our framework, we use the prompts \\(\\mathbf{P}\\) to refine the input portrait \\(\\mathbf{I}\\) into the edited image \\(\\mathbf{I}_p\\). The \\(\\mathbf{I}_p\\) inherits a similar structure to the input \\(\\mathbf{I}\\), providing coarse geometry or structural guidance in the editing process. As discussed in the above section, the encoder \\(\\mathrm{E}_{low}(\\cdot)\\) consistently generates low-frequency features to represent geometry cues, making it well-suited for preserving structured information in the input portrait. As a result, we freeze \\(\\mathrm{E}_{low}(\\cdot)\\) and leverage it to produce structure features \\(\\mathbf{F}_g = \\mathrm{E}_{low}(\\mathbf{I})\\) for coarse geometry prediction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.785, + 0.402 + ], + "angle": 0, + "content": "Injecting Prompts P as Condition. In contrast to input images, prompts typically offer more high-frequency and texture information to guide and control the editing process. Accordingly, it is appropriate to incorporate the prompts into the high-level branch \\(\\mathrm{E}_{high}(\\cdot)\\) in Live3D to extract the high-frequency features. To incorporate prompts, we utilize the prompt encoder \\(\\mathrm{E}_p(\\cdot)\\) to create prompt embeddings that are fused into our model through cross attention, which is similar to the strategy in Stable Diffusion [56]. Specifically, we add a cross-attention layer after \\(\\mathrm{E}_{high}(\\cdot)\\) to obtain the feature updated with the prompt embeddings:" + }, + { + "type": "equation", + "bbox": [ + 0.365, + 0.413, + 0.785, + 0.428 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {a} = \\operatorname {C r o s s A t t e n t i o n} \\left(\\operatorname {E} _ {\\text {h i g h}} (\\mathbf {I}), \\operatorname {E} _ {p} (\\mathbf {P})\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.784, + 0.496 + ], + "angle": 0, + "content": "where the encoder \\(\\mathbf{E}_p(\\cdot)\\) is a transformer (MAE [29] for image and CLIP [54] for text). The \\(\\mathrm{E}_{high}(\\mathbf{I})\\) serves as the query and \\(\\mathrm{E}_p(\\mathbf{P})\\) is employed as key and value. The updated feature \\(\\mathbf{F}_a\\) and the geometry feature \\(\\mathbf{F}_g\\) are then feed into the decoder \\(\\mathrm{E}_t(\\cdot)\\) to infer the triplane \\(\\mathbf{T_p}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.441, + 0.504, + 0.785, + 0.519 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} _ {\\mathbf {p}} = \\operatorname {E} _ {t} \\left(\\mathbf {F} _ {a}, \\mathbf {F} _ {g}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.525, + 0.784, + 0.6 + ], + "angle": 0, + "content": "Distilling Diffusion and 3D GAN Prior. Unlike previous methods that require large amounts of 3D data, our model operates with only 2D paired data generated by the 2D editing model [7]. For both visual and textual prompts, we generate edited images \\(\\mathbf{I}_{gt}\\) with the diffusion-based image editing model [7] as the pseudo labels. The details of data generation can be found in Sec. 4.1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.601, + 0.784, + 0.631 + ], + "angle": 0, + "content": "With the predicted triplane \\(\\mathbf{T}_{\\mathbf{p}}\\), we can render the image using the pretrained EG3D model and then calculate the reconstruction loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.419, + 0.64, + 0.785, + 0.655 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {2 d} = \\ell_ {\\mathbf {I}} \\left(\\mathbf {I} _ {g t}, \\mathrm {R} \\left(\\mathbf {T} _ {\\mathbf {p}}, \\mathbf {c}\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.662, + 0.784, + 0.734 + ], + "angle": 0, + "content": "where \\( \\mathrm{R}(\\cdot) \\) is the rendering module of EG3D, \\( \\mathbf{c} \\) is the camera pose of input portrait, and \\( \\ell_1(\\cdot) \\) is an image reconstruction loss penalizing the difference between the ground truth \\( \\mathbf{I}_{gt} \\) and the rendering \\( \\mathbf{I}_p = \\mathrm{R}(\\mathbf{T_p},\\mathbf{c}) \\). It's important to note that \\( \\mathcal{L}_{2d} \\), used to reconstruct the \\( \\mathbf{I}_{gt} \\), essentially distills the knowledge of editing from the diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.737, + 0.784, + 0.813 + ], + "angle": 0, + "content": "We observe that \\(\\mathcal{L}_{2d}\\) helps the model to reconstruct well on the input camera view but suffers from geometry distortion during novel-view synthesis. To fully exploit the 3D properties of Live3D, we propose to distill the 3D knowledge of Live3D. Specifically, we leverage the pretrained Live3D to infer the triplane \\(\\mathbf{T}_{gt}\\), and multi-view depths \\(\\mathcal{D}_{gt}\\) and images \\(\\mathcal{T}_{gt}\\) of the pseudo-label image \\(\\mathbf{I}_{gt}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.429, + 0.822, + 0.785, + 0.837 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} _ {g t}, \\mathcal {D} _ {g t} = \\mathrm {G} (\\mathbf {I} _ {g t}, \\mathcal {C}), \\tag {5}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.357, + 0.129 + ], + "angle": 0, + "content": "Q. Bai et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "where \\(\\mathrm{G}(\\cdot)\\) represents the inference process of Live3D, where \\(\\mathcal{C} = \\{\\mathbf{c}_1,\\dots,\\mathbf{c}_n\\}\\) is the camera set, and \\(n\\) denotes the number of cameras. We also render the multi-view images \\(\\mathcal{I}_p\\) and depths \\(\\mathcal{D}_p\\) of the edited image \\(\\mathrm{I}_p\\) from the triplane \\(\\mathbf{T}_p\\) and then define the objective:" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.22, + 0.786, + 0.238 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {3 d} = \\ell_ {\\mathcal {I}} \\left(\\mathcal {I} _ {p}, \\mathcal {I} _ {g t}\\right) + \\ell_ {\\mathbf {T}} \\left(\\mathbf {T} _ {p}, \\mathbf {T} _ {g t}\\right) + \\ell_ {\\mathcal {D}} \\left(\\mathcal {D} _ {p}, \\mathcal {D} _ {g t}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.247, + 0.786, + 0.279 + ], + "angle": 0, + "content": "where \\(\\ell_{\\mathcal{I}}(\\cdot), \\ell_{\\mathbf{T}}(\\cdot)\\) and \\(\\ell_{\\mathcal{D}}(\\cdot)\\) is the reconstruction loss penalizing the difference in image, triplane, and depth between \\(\\mathrm{I}_p\\) and \\(\\mathrm{I}_{gt}\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.301, + 0.452, + 0.317 + ], + "angle": 0, + "content": "3.3 Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.341, + 0.788, + 0.417 + ], + "angle": 0, + "content": "Training. During the training phase, we sample a triplet consisting of the input portrait \\(\\mathbf{I}\\) along with its camera pose \\(\\mathbf{c}\\), prompts \\(\\mathbf{P}\\), and the pseudo-label image \\(\\mathbf{I}_{gt}\\) generated by the 2D diffusion model. We leverage the Live3D model as a pretrained model and add an additional prompts encoder \\(\\mathrm{E}_p\\) to extract prompt embeddings. The overall learning objective can be described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.431, + 0.786, + 0.448 + ], + "angle": 0, + "content": "\\[\nL = \\lambda_ {1} L _ {2 D} + \\lambda_ {2} L _ {3 D}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.562 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are loss weights. In our setting, \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are both set to 1.0. For the reconstruction loss, \\(\\ell_{\\mathcal{I}}(\\cdot)\\), \\(\\ell_{\\mathrm{I}}(\\cdot)\\) are combinations of \\(\\mathrm{L}_2\\) loss and LPIPS loss [84], with loss weights being 1 and 2, respectively. \\(\\ell_{\\mathcal{D}}(\\cdot)\\) and \\(\\ell_{\\mathrm{T}}(\\cdot)\\) are \\(\\mathrm{L}_1\\) loss. Notably, based on the study on Live3D feature representation, during training only \\(\\mathrm{E}_p(\\cdot)\\), \\(\\mathrm{E}_t(\\cdot)\\) are set to be learnable, while other modules are frozen. It allows our model to leverage Live3D knowledge as much as possible and converge at a very fast speed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.563, + 0.788, + 0.608 + ], + "angle": 0, + "content": "Inference. For inference, users can provide a single 2D portrait image and choose an image prompt or text prompt. Our model is able to generate the edited 3D NeRF along with photorealistic view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.788, + 0.699 + ], + "angle": 0, + "content": "Customized Prompts Adaptation. To accommodate customized styles provided by users, we propose a method to adapt our pretrained encoder to user-defined styles. We increase the tuning efficiency by optimizing only \\(\\mathrm{E}_p(\\cdot)\\) and the normalization layers in \\(\\mathrm{E}_t(\\cdot)\\) with the same learning objective Eq. 7. This method allows us to limit the training data to only 10 image pairs and the learning time to 5 minutes on a single GPU." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.723, + 0.377, + 0.74 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.755, + 0.429, + 0.771 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Training Settings. For real faces, we adopt the FFHQ dataset [35] at \\(512 \\times 512\\) resolution with camera parameters aligned by EG3D [13]. In order to obtain the stylized images as pseudo labels, and the corresponding textual and visual prompts, we leverage the text-guided editing model InstructPix2Pix (IP2P) [7] to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.689, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "3DPE" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.161, + 0.231, + 0.179 + ], + "angle": 0, + "content": "mnd" + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.146, + 0.357, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.393, + 0.146, + 0.493, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.146, + 0.63, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.146, + 0.733, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.733, + 0.177, + 0.776, + 0.198 + ], + "angle": 0, + "content": "Recolor it to a Gothic color scheme." + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.203, + 0.238, + 0.41 + ], + "angle": 0, + "content": "P 12345678901001100110011001100110011001100110011001100110011001100110011001100110011001100110011001100110011001100" + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.202, + 0.372, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.202, + 0.508, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.202, + 0.646, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.202, + 0.783, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.219, + 0.433, + 0.239, + 0.451 + ], + "angle": 0, + "content": "0000000000000000000000000000000000000" + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.415, + 0.372, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.415, + 0.508, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.415, + 0.646, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.415, + 0.783, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.48, + 0.788, + 0.538 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparisons. We compare the results of several baselines with image prompts and text prompts. In each case, we include the edited portraits as well as their novel view renderings. Our method generates high-quality edited portraits with better 3D quality and alignment with the referenced prompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.844 + ], + "angle": 0, + "content": "edit the real faces. Specifically, we manually generate and select textual prompts, and send them to IP2P together with the real images to obtain the edited 2D images. The text prompts of our training and testing set are obtained by augmenting the aforementioned textual prompts with the large language model GPT [8] following IP2P [7]. While the image prompts of our training and testing set are picked from the edited 2D images. We conduct experiments on 20 styles, and for each style, we synthesize 1000 images, resulting in 20,000 image pairs with text prompts for model training. For each style, we use 8 textual prompts for training and 5 textual prompts for testing. We adopt a learning rate of 5e-5 and optimize the model for 60k iterations with a batch size of 32. The entire training procedure is completed over a period of 40 hours using 8 NVIDIA A100 GPUs. Baselines. We conduct comparisons against two categories of methods that achieve analogous outcomes: 1) 3D reconstruction coupled with 3D editing; 2) 2D editing followed by 3D reconstruction. For the first category, we implement three baseline approaches, wherein we firstly leverage PTI [55] or Live3D [70] for reconstruction and edit both of them following StyleCLIP [50]. Specifically, for experiments in PTI + StyleCLIP, we first conduct pivotal tuning [55] based on the predicted latent codes predicted by e4e [68] to reconstruct the given image in" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Q. Bai et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.787, + 0.215 + ], + "angle": 0, + "content": "Table 1: Quantitative comparisons. We compare several baselines on the 100 images of FFHQ dataset. It's important to note that we exclude \\(CLIP_r\\) for PTI+StyleCLIP and Live3D+StyleCLIP since these models utilize CLIP for optimization. Our model excels in 3D quality and achieves a remarkable 250x improvement (compared with IP2P+Live3D) in inference speed, achieving real-time performance." + }, + { + "type": "table", + "bbox": [ + 0.274, + 0.227, + 0.73, + 0.307 + ], + "angle": 0, + "content": "
MethodIDt↑CLIPr↑3D↑Time↓
PTI+StyleCLIP0.11-0.73360s
Live3D+StyleCLIP0.63-0.7230s
IP2P+Live3D0.440.620.7510s
Ours0.470.730.760.04s
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.38, + 0.787, + 0.606 + ], + "angle": 0, + "content": "a 3D-aware manner. Then we perform editing on the latent codes following [50], where CLIP loss [54] is utilized between the model output and the prompt (no matter whether it is an image or text). The setting of Live3D + StyleCLIP is similar to PTI + StyleCLIP except that we obtain the triplane by Live3D and perform editing on this 3D representation. SPI [81] + StyleGAN-NADA [21] (StyleNADA for the abbreviation in Fig. 4) serves as the third baseline in the first category, where SPI is used for 3D reconstruction and StyleGAN-NADA is utilized to optimize the 3D generator for editing. In the second category, we employ a diffusion-based text-guided 2D image editing method IP2P [7] prior to executing 3D reconstruction via Live3D. IP2P being a text-instructed editing model, we need to first extract text prompts from the reference image with the image caption algorithm BLIP [41] when image prompts are input. The 2D edited results could then be obtained by sending the input image and the text prompt to IP2P. The final 3D results are acquired by inverting the 2D editing results with Live3D." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.629, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Evaluation Criteria. We evaluate all metrics on 100 pairs of images processed from FFHQ. We conduct a comprehensive evaluation of the editing performance in the following four aspects: identity preservation, reference alignment, 3D consistency, and inference speed. 1) Identity preservation \\((\\mathbf{ID}_{\\mathbf{t}})\\) aims to measure the preservation of the original identity by calculating the cosine similarity between the identity feature of the input image \\(\\mathbf{I}\\) and that of the edited image \\(\\mathbf{I}_p\\). We use the ArcFace model [16] to extract identity features. 2) Reference alignment \\((\\mathrm{CLIP_r})\\) targets at assessing the alignment of the output editing styles to the desired input prompt by computing the cosine similarity in the CLIP [50] feature space. 3) 3D consistency (3D) on the edited outputs is measured following the evaluation protocols established by EG3D [13]. This involves calculating the identity similarity across multiple views. 4) Inference speed (Time) is measured on a single NVIDIA A6000 GPU with an average of 100 samples." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.689, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "3DPE" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.784, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.281, + 0.785, + 0.325 + ], + "angle": 0, + "content": "Fig. 5: (a) Generated results of customized prompt adaptation and (b) its learning process. We show the intermediate testing results at 10s, 1min, 2min and 5min during adaptation for the novel style golden statue." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.34, + 0.784, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.44, + 0.785, + 0.469 + ], + "angle": 0, + "content": "Fig.6: Qualitative comparison for ablations on (a) the distillation loss of 3D GAN \\((\\mathcal{L}_{3D})\\) and (b) diffusion models \\((\\mathcal{L}_{2D})\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.502, + 0.553, + 0.518 + ], + "angle": 0, + "content": "4.2 Efficient 3D-aware Portrait Editing" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.788, + 0.672 + ], + "angle": 0, + "content": "We make an in-depth analysis of our efficient portrait editing method both quantitatively and qualitatively. The results are included in Tab. 1 and Fig. 4. For reference alignment, we do not report \\(CLIP_{r}\\) for methods that leverage CLIP for optimization since it is evaluated with CLIP as well. The standout feature of our method is its efficiency. Our method achieves an inference speed of merely 40ms, which improves over 100 times compared to the fastest existing baselines, which require around 10 seconds. Because of the efficient knowledge distillation, our approach is also good in preserving the identity, adhering more closely to the reference prompts and achieving the best 3D consistency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "As in Fig. 4, compared with PTI + CLIP methods, our model enables precise alignment with inputs and prompts, while the baseline always generates low-fidelity textures, and the resulting geometry has many artifacts. Although Live3D + CLIP achieves the best \\( ID_{t} \\) score, the edited results are basically unchanged compared to the input. We suspect that Live3D does not possess a latent space and loses the editing priors, making it challenging for the results of CLIP optimization to align well with prompts. SPI + StyleGAN-NADA performs relatively slowly (18 minutes per instance) because both the 3D reconstruction and the editing are achieved by optimizing the generator. Although it can produce realistic textures, it fails to reconstruct fine details and to edit following the prompts. In contrast to the 2D editing and subsequent 3D reconstruction" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Q. Bai et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 2: Ablation study on the number of styles used for pretraining. We report LPIPS score after 0.1min, 1min, 2min, and 5min fine-tuning for evaluation." + }, + { + "type": "table", + "bbox": [ + 0.292, + 0.185, + 0.711, + 0.242 + ], + "angle": 0, + "content": "
#Style/Time0.1min1min2min5min
30.77580.63700.53120.5220
100.59320.57340.51060.5012
200.57030.52900.50220.4849
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.255, + 0.787, + 0.297 + ], + "angle": 0, + "content": "Table 3: Ablation study on the number of data pairs used for novel prompt adaptation. We report LPIPS score after 0.1min, 1min, 2min, and 5min fine-tuning for evaluation." + }, + { + "type": "table", + "bbox": [ + 0.294, + 0.31, + 0.71, + 0.388 + ], + "angle": 0, + "content": "
#Pair/Time0.1min1min2min5min
20.61910.54310.52130.5212
50.61480.53240.50640.4971
100.60760.52420.49830.4935
200.59810.55250.49540.4895
500.58910.57180.49220.4797
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.415, + 0.787, + 0.446 + ], + "angle": 0, + "content": "pipeline (IP2P + Live3D), our method produces textures more consistent with the prompt and aligns better with the input structure." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.467, + 0.555, + 0.483 + ], + "angle": 0, + "content": "4.3 Adaptation for Customized Editing" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.491, + 0.787, + 0.779 + ], + "angle": 0, + "content": "With the trained editing network, our method already supports a variety of 3D-aware editing styles. Following DreamBooth [57] and to expand the selection and better conform to users' preferences, we offer an efficient method for fast adaptation to the user-defined customized prompts. Users can personalize the editing network by providing a modest set of 10 reference editing pairs. These pairs can be either handpicked from artist-created examples or produced using text-guided image editing models. As described in Sec. 3.3, only the partial of the encoder would be optimized. The adaptation process itself is remarkably fast, requiring only about 5 minutes to accomplish the learning of the customized prompts. Please also note that, once the user-defined style has been trained, all the incoming testing results can be obtained in a real-time manner. In Fig. 5, we present the testing results and learning process of the adaptation. Our model can quickly master the customized knowledge in about 2 minutes. With further training (e.g., 5 minutes), the edited results become more stylized, demonstrating a trade-off between authenticity and stylization. Upon completion of the adaptation, the method allows users to edit testing inputs in these newly learned styles with a minimal inference time of 0.04s. This rapid performance indicates that the method is well-optimized for real-time applications, providing a seamless and efficient user experience." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.801, + 0.403, + 0.815 + ], + "angle": 0, + "content": "4.4 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.825, + 0.787, + 0.841 + ], + "angle": 0, + "content": "We analyze our model and validate our design choice by ablating the components." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.689, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "3DPE" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.147, + 0.769, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.4, + 0.788, + 0.444 + ], + "angle": 0, + "content": "Fig. 7: Video editing results with 3DPE. We use text prompts \"Make him a bronze statue\" to edit the input video. Our method can accurately reconstruct and preserve challenging facial expressions and achieve high-quality novel view renderings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.483, + 0.788, + 0.574 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{3D}\\) for Distillation of Live3D. We compare our model trained with and without \\(\\mathcal{L}_{3D}\\). As shown in Fig. 6, we observe that the model without \\(\\mathcal{L}_{3D}\\) results in very flat geometry, and the synthesis of the side view loses the normal structure with many artifacts. This highlights the importance of the 3D knowledge from Live3D in training our model with 2D paired data, enabling the inference of reasonable face geometry without relying on any 3D data sources." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.578, + 0.788, + 0.669 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{2D}\\) for Distillation of Diffusion Models. We also study the effect of \\(\\mathcal{L}_{2D}\\), which represents using 2D editing priors from the diffusion model [7] to optimize the encoder. From Fig. 6, we observe that \\(\\mathcal{L}_{2D}\\) is critical for our model, especially in preserving detailed texture. Without it, the edited image tends to lose the appearance information from prompts and exhibits structural artifacts around the nose." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Ablation on Pretraining Data Scale for Customized Prompt Adaptations. In this part, we conduct ablation studies on the pretraining scale and investigate its influence on customized propmts. Specifically, we choose various 3DPE models pretrained with different style amounts (3, 10, and 20 styles respectively) and perform novel prompt adaptation on a single A6000 GPU. We choose the golden statue as the customized style and report LPIPS score between the predicted stylized images and the pseudo stylized labels produced by IP2P [7]. Experimental results in Tab. 2 reveal that the proposed method could benefit from the improving pretraining data scale and adapt faster and eventually perform better with more data, which suggests the potential of the model for future scaling up." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Q. Bai et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.162 + ], + "angle": 0, + "content": "Ablation on Tuning Data Scale for Customized Prompt Adaptations." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.163, + 0.787, + 0.313 + ], + "angle": 0, + "content": "To investigate the ability of our method to adapt to user-defined novel prompts, we conduct an ablation study on the number of novel-prompt data pairs required for training. Similar to the prior experiments on the pretraining scale, we perform adaptation with 2, 5, 10, 20, or 50 pairs of novel-prompt data on a single A6000 GPU. We experimentally find that our model can converge in about 5 minutes, and therefore, we report the perceptual distance between the model output and the provided ground truth at \\(0.1\\mathrm{min}\\), \\(1\\mathrm{min}\\), \\(2\\mathrm{min}\\), and \\(5\\mathrm{min}\\) for comparison. As shown in Tab. 3, our model can adapt to the new style in around 2 minutes. As the number of provided data pairs increases, the model can adapt faster at the beginning." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.336, + 0.365, + 0.351 + ], + "angle": 0, + "content": "4.5 Applications" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.362, + 0.787, + 0.573 + ], + "angle": 0, + "content": "One application of our pipeline is to edit videos, requiring accurate reconstruction and efficient editing simultaneously. We demonstrate the effectiveness of our pipeline on talking face videos, as illustrated in Fig. 7. In this experiment, we perform sequence editing in a frame-by-frame manner by sending an image sequence from TalkingHead-1KH [72] dataset together with the text prompt \"Make him a bronze statue\" to 3DPE. The three rows in Fig. 7 respectively illustrate the input frames, the edited frames, and the novel views. Our method accurately reconstructs facial expressions, while achieving high-quality editing results. Since our method is 3D-aware, we can naturally view videos from novel perspectives. To better experience our method, we design an interactive system that allows users for interactive editing. Users are required to provide the image or text prompt to indicate the desired style and input an input image that is to be edited, and the system will apply our method on the given image and prompt to output the edited image as well as novel-view edited results." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.597, + 0.353, + 0.612 + ], + "angle": 0, + "content": "5 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.629, + 0.786, + 0.704 + ], + "angle": 0, + "content": "Limitations. Despite achieving state-of-the-art performance in quality and efficiency, our method exhibits inconsistencies in details for novel-view rendering because the EG3D framework relies on a super-resolution module. Additionally, when our method is applied to video editing, it presents flickering artifacts since our model is originally designed for image-based editing." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.705, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Conclusion. In this work, we introduce 3DPE for real-time 3D-aware portrait editing. By distilling the powerful knowledge of diffusion models and 3D GANs into a lightweight module, our model significantly reduces editing time while ensuring quality. Benefiting from the distillation process, 3DPE could handle multiple styles in a single model and perform both image-guided and text-guided editing. The advantages of our method empower us to generate photorealistic 3D portraits, a capability crucial for the visual effects industry, AR/VR systems, and teleconferencing, among other applications, while we firmly oppose any misuse that could compromise privacy and security." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.689, + 0.115, + 0.733, + 0.127 + ], + "angle": 0, + "content": "3DPE" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.786, + 0.21 + ], + "angle": 0, + "content": "This project was supported by the National Key R&D Program of China under grant number 2022ZD0161501." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.234, + 0.324, + 0.25 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.267, + 0.785, + 0.308 + ], + "angle": 0, + "content": "1. Abdal, R., Lee, H.Y., Zhu, P., Chai, M., Siarohin, A., Wonka, P., Tulyakov, S.: 3DAvatarGAN: Bridging domains for personalized editable avatars. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.785, + 0.337 + ], + "angle": 0, + "content": "2. Abdal, R., Qin, Y., Wonka, P.: Image2StyleGAN: How to embed images into the StyleGAN latent space? In: Int. Conf. Comput. Vis. (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.338, + 0.785, + 0.365 + ], + "angle": 0, + "content": "3. Abdal, R., Qin, Y., Wonka, P.: Image2StyleGAN++: How to edit the embedded images? In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.367, + 0.785, + 0.393 + ], + "angle": 0, + "content": "4. Aneja, S., Thies, J., Dai, A., Nießner, M.: ClipFace: Text-guided editing of textured 3d morphable models. In: ACM SIGGRAPH Conf. Proc. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.395, + 0.785, + 0.421 + ], + "angle": 0, + "content": "5. Bai, Q., Xu, Y., Zhu, J., Xia, W., Yang, Y., Shen, Y.: High-fidelity GAN inversion with padding space. In: Eur. Conf. Comput. Vis. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.423, + 0.785, + 0.449 + ], + "angle": 0, + "content": "6. Blanz, V., Vetter, T.: A morphable model for the synthesis of 3d faces. In: Seminal Graphics Papers: Pushing the Boundaries (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.451, + 0.785, + 0.477 + ], + "angle": 0, + "content": "7. Brooks, T., Holynski, A., Efros, A.A.: InstructPix2Pix: Learning to follow image editing instructions. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.479, + 0.785, + 0.518 + ], + "angle": 0, + "content": "8. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: Adv. Neural Inform. Process. Syst. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.52, + 0.785, + 0.56 + ], + "angle": 0, + "content": "9. Cai, S., Obukhov, A., Dai, D., Van Gool, L.: Pix2NeRF: Unsupervised conditional p-gan for single image to neural radiance fields translation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.563, + 0.785, + 0.602 + ], + "angle": 0, + "content": "0. Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., Zheng, Y.: MasaCtrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In: Int. Conf. Comput. Vis. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.604, + 0.785, + 0.631 + ], + "angle": 0, + "content": "1. Ceylan, D., Huang, C.H.P., Mitra, N.J.: Pix2Video: Video editing using image diffusion. In: Int. Conf. Comput. Vis. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.785, + 0.658 + ], + "angle": 0, + "content": "2. Chai, W., Guo, X., Wang, G., Lu, Y.: StableVideo: Text-driven consistency-aware diffusion video editing. In: Int. Conf. Comput. Vis. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.66, + 0.785, + 0.701 + ], + "angle": 0, + "content": "3. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.785, + 0.742 + ], + "angle": 0, + "content": "4. Chan, E.R., Monteiro, M., Kellnhofer, P., Wu, J., Wetzstein, G.: pi-GAN: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.785, + 0.77 + ], + "angle": 0, + "content": "5. Chang, S., Kim, G., Kim, H.: HairNeRF: Geometry-aware image synthesis for hairstyle transfer. In: Int. Conf. Comput. Vis. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.798 + ], + "angle": 0, + "content": "6. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: ArcFace: Additive angular margin loss for deep face recognition. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Deng, Y., Yang, J., Chen, D., Wen, F., Tong, X.: Disentangled and controllable face image generation via 3d imitative-contrastive learning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.267, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Q. Bai et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "18. Deng, Y., Yang, J., Xu, S., Chen, D., Jia, Y., Tong, X.: Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "19. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Int. Conf. Learn. Represent. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.247, + 0.786, + 0.288 + ], + "angle": 0, + "content": "20. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. In: Int. Conf. Learn. Represent. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.289, + 0.787, + 0.33 + ], + "angle": 0, + "content": "21. Gal, R., Patashnik, O., Maron, H., Bermano, A.H., Chechik, G., Cohen-Or, D.: StyleGAN-NADA: Clip-guided domain adaptation of image generators. ACM Trans. Graph. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.331, + 0.787, + 0.359 + ], + "angle": 0, + "content": "22. Gao, C., Shih, Y., Lai, W.S., Liang, C.K., Huang, J.B.: Portrait neural radiance fields from a single image. arXiv preprint arXiv:2012.05903 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.36, + 0.787, + 0.387 + ], + "angle": 0, + "content": "23. Geyer, M., Bar-Tal, O., Bagon, S., Dekel, T.: TokenFlow: Consistent diffusion features for consistent video editing. In: Int. Conf. Learn. Represent. (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.388, + 0.787, + 0.416 + ], + "angle": 0, + "content": "24. Graikos, A., Malkin, N., Jojic, N., Samaras, D.: Diffusion models as plug-and-play priors. In: Adv. Neural Inform. Process. Syst. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.417, + 0.787, + 0.444 + ], + "angle": 0, + "content": "25. Gu, J., Liu, L., Wang, P., Theobalt, C.: StyleNeRF: A style-based 3d-aware generator for high-resolution image synthesis. In: Int. Conf. Learn. Represent. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.445, + 0.787, + 0.472 + ], + "angle": 0, + "content": "26. Guo, J., Yu, J., Lattas, A., Deng, J.: Perspective reconstruction of human faces by joint mesh and landmark regression. In: Eur. Conf. Comput. Vis. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.514 + ], + "angle": 0, + "content": "27. Haque, A., Tancik, M., Efros, A.A., Holynski, A., Kanazawa, A.: Instruct-NeRF2NeRF: Editing 3d scenes with instructions. In: Int. Conf. Comput. Vis. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.787, + 0.543 + ], + "angle": 0, + "content": "28. Harkonen, E., Hertzmann, A., Lehtinen, J., Paris, S.: GANSpace: Discovering interpretable GAN controls. Adv. Neural Inform. Process. Syst. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.544, + 0.787, + 0.572 + ], + "angle": 0, + "content": "29. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.787, + 0.6 + ], + "angle": 0, + "content": "30. Hertz, A., Aberman, K., Cohen-Or, D.: Delta denoising score. In: Int. Conf. Comput. Vis. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.787, + 0.642 + ], + "angle": 0, + "content": "31. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: Int. Conf. Learn. Represent. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.643, + 0.787, + 0.671 + ], + "angle": 0, + "content": "32. Hyung, J., Hwang, S., Kim, D., Lee, H., Choo, J.: Local 3d editing via 3d distillation of clip knowledge. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.672, + 0.787, + 0.712 + ], + "angle": 0, + "content": "33. Jiang, K., Chen, S.Y., Fu, H., Gao, L.: NeRFFaceLighting: Implicit and disentangled face lighting representation leveraging generative prior in neural radiance fields. ACM Trans. Graph. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.713, + 0.787, + 0.741 + ], + "angle": 0, + "content": "34. Jin, W., Ryu, N., Kim, G., Baek, S.H., Cho, S.: Dr. 3D: Adapting 3d GANs to artistic drawings. In: ACM SIGGRAPH Asia Conf. Proc. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.787, + 0.77 + ], + "angle": 0, + "content": "35. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.787, + 0.812 + ], + "angle": 0, + "content": "36. Kim, G., Chun, S.Y.: DATID-3D: Diversity-preserved domain adaptation using text-to-image diffusion for 3d generative model. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "37. Ko, J., Cho, K., Choi, D., Ryoo, K., Kim, S.: 3D GAN inversion with pose optimization. In: IEEE Winter Conf. Appl. Comput. Vis. (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.689, + 0.115, + 0.733, + 0.127 + ], + "angle": 0, + "content": "3DPE" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "38. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "39. Li, J., Li, J., Zhang, H., Liu, S., Wang, Z., Xiao, Z., Zheng, K., Zhu, J.: PREIM3D: 3d consistent precise image attribute editing from a single image. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.233, + 0.788, + 0.274 + ], + "angle": 0, + "content": "40. Li, J., Liu, S., Liu, Z., Wang, Y., Zheng, K., Xu, J., Li, J., Zhu, J.: Instruct-Pix2NeRF: Instructed 3d portrait editing from a single image. arXiv preprint arXiv:2311.02826 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.276, + 0.788, + 0.317 + ], + "angle": 0, + "content": "41. Li, J., Li, D., Xiong, C., Hoi, S.: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: Int. Conf. Mach. Learn. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.318, + 0.788, + 0.346 + ], + "angle": 0, + "content": "42. Lin, C., Lindell, D., Chan, E., Wetzstein, G.: 3D GAN inversion for controllable portrait image animation. In: Eur. Conf. Comput. Vis. Workshop (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.347, + 0.788, + 0.373 + ], + "angle": 0, + "content": "43. Liu, S., Zhang, Y., Li, W., Lin, Z., Jia, J.: Video-P2P: Video editing with cross-attention control. arXiv preprint arXiv:2303.04761 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.375, + 0.788, + 0.416 + ], + "angle": 0, + "content": "44. Liu, Z., Feng, R., Zhu, K., Zhang, Y., Zheng, K., Liu, Y., Zhao, D., Zhou, J., Cao, Y.: Cones: Concept neurons in diffusion models for customized generation. In: Int. Conf. Mach. Learn. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.418, + 0.788, + 0.459 + ], + "angle": 0, + "content": "45. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: Int. Conf. Learn. Represent. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.46, + 0.788, + 0.5 + ], + "angle": 0, + "content": "46. Nitzan, Y., Gharbi, M., Zhang, R., Park, T., Zhu, J.Y., Cohen-Or, D., Shechtman, E.: Domain expansion of image generators. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.502, + 0.788, + 0.543 + ], + "angle": 0, + "content": "47. Or-El, R., Luo, X., Shan, M., Shechtman, E., Park, J.J., Kemelmacher-Shlizerman, I.: StyleSDF: High-resolution 3d-consistent image and geometry generation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.545, + 0.788, + 0.585 + ], + "angle": 0, + "content": "48. Ouyang, H., Wang, Q., Xiao, Y., Bai, Q., Zhang, J., Zheng, K., Zhou, X., Chen, Q., Shen, Y.: CoDeF: Content deformation fields for temporally consistent video processing. arXiv preprint arXiv:2308.07926 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.587, + 0.788, + 0.628 + ], + "angle": 0, + "content": "49. Pan, X., Xu, X., Loy, C.C., Theobalt, C., Dai, B.: A shading-guided generative implicit model for shape-accurate 3d-aware image synthesis. Adv. Neural Inform. Process. Syst. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.657 + ], + "angle": 0, + "content": "50. Patashnik, O., Wu, Z., Shechtman, E., Cohen-Or, D., Lischinski, D.: StyleCLIP: Text-driven manipulation of StyleGAN imagery. In: Int. Conf. Comput. Vis. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.658, + 0.788, + 0.699 + ], + "angle": 0, + "content": "51. Paysan, P., Knothe, R., Amberg, B., Romdhani, S., Vetter, T.: A 3d face model for pose and illumination invariant face recognition. In: IEEE international conference on advanced video and signal based surveillance (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.7, + 0.788, + 0.727 + ], + "angle": 0, + "content": "52. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: DreamFusion: Text-to-3d using 2d diffusion. In: Int. Conf. Learn. Represent. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.728, + 0.788, + 0.77 + ], + "angle": 0, + "content": "53. Qi, C., Cun, X., Zhang, Y., Lei, C., Wang, X., Shan, Y., Chen, Q.: FateZero: Fusing attentions for zero-shot text-based video editing. In: Int. Conf. Comput. Vis. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.771, + 0.788, + 0.812 + ], + "angle": 0, + "content": "54. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Int. Conf. Mach. Learn. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.813, + 0.788, + 0.84 + ], + "angle": 0, + "content": "55. Roich, D., Mokady, R., Bermano, A.H., Cohen-Or, D.: Pivotal tuning for latent-based editing of real images. ACM Trans. Graph. (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Q. Bai et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "56. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "57. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dream-Booth: Fine tuning text-to-image diffusion models for subject-driven generation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "58. Schwarz, K., Liao, Y., Niemeyer, M., Geiger, A.: GRAF: Generative radiance fields for 3d-aware image synthesis. In: Adv. Neural Inform. Process. Syst. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.3 + ], + "angle": 0, + "content": "59. Schwarz, K., Sauer, A., Niemeyer, M., Liao, Y., Geiger, A.: VoxGRAF: Fast 3d-aware image synthesis with sparse voxel grids. In: Adv. Neural Inform. Process. Syst. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.785, + 0.328 + ], + "angle": 0, + "content": "60. Shen, Y., Gu, J., Tang, X., Zhou, B.: Interpreting the latent space of GANs for semantic face editing. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.329, + 0.785, + 0.37 + ], + "angle": 0, + "content": "61. Shen, Y., Yang, C., Tang, X., Zhou, B.: InterFaceGAN: Interpreting the disentangled face representation learned by GANs. IEEE Trans. Pattern Anal. Mach. Intell. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.398 + ], + "angle": 0, + "content": "62. Shi, Z., Peng, S., Xu, Y., Geiger, A., Liao, Y., Shen, Y.: Deep generative models on 3d representations: A survey. arXiv preprint arXiv:2210.15663 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.399, + 0.785, + 0.438 + ], + "angle": 0, + "content": "63. Shi, Z., Shen, Y., Xu, Y., Peng, S., Liao, Y., Guo, S., Chen, Q., Yeung, D.Y.: Learning 3d-aware image synthesis with unknown pose distribution. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "64. Shi, Z., Xu, Y., Shen, Y., Zhao, D., Chen, Q., Yeung, D.Y.: Improving 3d-aware image synthesis with a geometry-aware discriminator. In: Adv. Neural Inform. Process. Syst. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.508 + ], + "angle": 0, + "content": "65. Skorokhodov, I., Tulyakov, S., Wang, Y., Wonka, P.: EpiGRAF: Rethinking training of 3d gans. In: Adv. Neural Inform. Process. Syst. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "66. Sun, J., Wang, X., Shi, Y., Wang, L., Wang, J., Liu, Y.: IDE-3D: Interactive disentangled editing for high-resolution 3d-aware portrait synthesis. ACM Trans. Graph. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "67. Sun, J., Wang, X., Zhang, Y., Li, X., Zhang, Q., Liu, Y., Wang, J.: FENeRF: Face editing in neural radiance fields. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.619 + ], + "angle": 0, + "content": "68. Tov, O., Alaluf, Y., Nitzan, Y., Patashnik, O., Cohen-Or, D.: Designing an encoder for StyleGAN image manipulation. ACM Trans. Graph. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.785, + 0.66 + ], + "angle": 0, + "content": "69. Tran, P., Zakharov, E., Ho, L.N., Tran, A.T., Hu, L., Li, H.: VOODOO 3D: Volumetric portrait disentanglement for one-shot 3d head reenactment. In: IEEE Conf. Comput. Vis. Pattern Recog. (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "70. Trevithick, A., Chan, M., Stengel, M., Chan, E., Liu, C., Yu, Z., Khamis, S., Chandraker, M., Ramamoorthi, R., Nagano, K.: Real-time radiance fields for single-image portrait view synthesis. ACM Trans. Graph. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.743 + ], + "angle": 0, + "content": "71. Wang, Q., Shi, Z., Zheng, K., Xu, Y., Peng, S., Shen, Y.: Benchmarking and analyzing 3d-aware image synthesis with a modularized codebase. In: Adv. Neural Inform. Process. Syst. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.771 + ], + "angle": 0, + "content": "72. Wang, T.C., Mallya, A., Liu, M.Y.: One-shot free-view neural talking-head synthesis for video conferencing. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "73. Wang, W., Jiang, Y., Xie, K., Liu, Z., Chen, H., Cao, Y., Wang, X., Shen, C.: Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.841 + ], + "angle": 0, + "content": "74. Wang, Y., Xu, T., Wu, Y., Li, M., Chen, W., Xu, L., Yu, J.: NARRATE: A normal assisted free-view portrait stylizer. arXiv preprint arXiv:2207.00974 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.689, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "3DPE" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "75. Xia, W., Xue, J.H.: A survey on deep generative 3d-aware image synthesis. ACM Computing Surveys (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.217 + ], + "angle": 0, + "content": "76. Xie, J., Ouyang, H., Piao, J., Lei, C., Chen, Q.: High-fidelity 3d GAN inversion by pseudo-multi-view optimization. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "77. Xu, Y., Peng, S., Yang, C., Shen, Y., Zhou, B.: 3d-aware image synthesis via learning structural and textural representations. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.26, + 0.785, + 0.286 + ], + "angle": 0, + "content": "78. Xu, Y., Shen, Y., Zhu, J., Yang, C., Zhou, B.: Generative hierarchical features from synthesizing images. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "79. Yang, S., Jiang, L., Liu, Z., Loy, C.C.: Pastiche master: Exemplar-based high-resolution portrait style transfer. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.328, + 0.785, + 0.355 + ], + "angle": 0, + "content": "80. Yang, S., Zhou, Y., Liu, Z.,, Loy, C.C.: Rerender a video: Zero-shot text-guided video-to-video translation. In: ACM SIGGRAPH Asia Conf. Proc. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.356, + 0.785, + 0.397 + ], + "angle": 0, + "content": "81. Yin, F., Zhang, Y., Wang, X., Wang, T., Li, X., Gong, Y., Fan, Y., Cun, X., Shan, Y., Oztireli, C., et al.: 3D GAN inversion with facial symmetry prior. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "82. Zhang, J., Lan, Y., Yang, S., Hong, F., Wang, Q., Yeo, C.K., Liu, Z., Loy, C.C.: DeformToon3D: Deformable neural radiance fields for 3d toonification. In: Int. Conf. Comput. Vis. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.439, + 0.785, + 0.466 + ], + "angle": 0, + "content": "83. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Int. Conf. Comput. Vis. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "84. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "85. Zhou, P., Xie, L., Ni, B., Tian, Q.: CIPS-3D: A 3d-aware generator of gans based on conditionally-independent pixel synthesis. arXiv preprint arXiv:2110.09788 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.537, + 0.785, + 0.563 + ], + "angle": 0, + "content": "86. Zhu, J., Shen, Y., Zhao, D., Zhou, B.: In-domain GAN inversion for real image editing. In: Eur. Conf. Comput. Vis. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.591 + ], + "angle": 0, + "content": "87. Zhu, J.Y., Krahenbuhl, P., Shechtman, E., Efros, A.A.: Generative visual manipulation on the natural image manifold. In: Eur. Conf. Comput. Vis. (2016)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.591 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Real-time 3D-aware Portrait Editing from a Single Image/c7f6568e-666e-490d-b059-9d93ade8c6f3_origin.pdf b/2024/Real-time 3D-aware Portrait Editing from a Single Image/c7f6568e-666e-490d-b059-9d93ade8c6f3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cd6c618b5f6868381090e0fc3f87addb465adbbe --- /dev/null +++ b/2024/Real-time 3D-aware Portrait Editing from a Single Image/c7f6568e-666e-490d-b059-9d93ade8c6f3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a319349920e55da5adb251c54c03e9f8633064cd401ed7182db89a709e5851f +size 7748584 diff --git a/2024/Real-time 3D-aware Portrait Editing from a Single Image/full.md b/2024/Real-time 3D-aware Portrait Editing from a Single Image/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0edf997258c9db84eaeeb1bbd1e9c74cdfc63607 --- /dev/null +++ b/2024/Real-time 3D-aware Portrait Editing from a Single Image/full.md @@ -0,0 +1,334 @@ +# Real-time 3D-aware Portrait Editing from a Single Image + +Qingyan Bai $^{1,2}$ , Zifan Shi $^{1}$ , Yinghao Xu $^{3}$ , Hao Ouyang $^{1,2}$ + +Qiuyu Wang $^{2}$ , Ceyuan Yang $^{4}$ , Xuan Wang $^{2}$ + +Gordon Wetzstein3, Yujun Shen2, and Qifeng Chen1 + +1 The Hong Kong University of Science and Technology + +2 Ant Group + +3 Stanford University + +4 Shanghai AI Laboratory + +Abstract. This work presents 3DPE, a practical method that can efficiently edit a face image following given prompts, like reference images or text descriptions, in a 3D-aware manner. To this end, a lightweight module is distilled from a 3D portrait generator and a text-to-image model, which provide prior knowledge of face geometry and superior editing capability, respectively. Such a design brings two compelling advantages over existing approaches. First, our method achieves real-time editing with a feedforward network (i.e., $\sim 0.04$ s per image), over $100\times$ faster than the second competitor. Second, thanks to the powerful priors, our module could focus on the learning of editing-related variations, such that it manages to handle various types of editing simultaneously in the training phase and further supports fast adaptation to user-specified customized types of editing during inference (e.g., with $\sim 5$ min fine-tuning per style). Project page can be found here. + +Keywords: 3D-aware portrait $\cdot$ Efficient editing + +# 1 Introduction + +Inferring the geometry and appearance from a single-view portrait image has become mature and practical [22, 37, 42, 66, 70, 74, 76], largely attributed to the utilization of priors in various 2D/3D generative models. However, only performing geometry reconstruction is insufficient. The significance of 3D portrait editing, driven by user intentions, and the need for streamlined efficiency in the editing process has been steadily increasing. This is particularly crucial in real-world applications such as AR/VR, 3D telepresence, and video conferencing, where real-time editing is often essential. Consequently, a key question arises: How can we effectively address the challenge of attaining high-fidelity portrait editing while ensuring real-time efficiency? + +Traditional methods for 3D portrait editing [17] typically rely on template facial models [6,51], which has limitations in handling the substantial geometry + +![](images/ae23707245db9249b6c133432e8b7729bf5940f9afcb2fa0ff5ab65b2a3394c8.jpg) +Fig. 1: Photorealistic editing results produced by our proposed 3DPE, which allows users to perform 3D-aware portrait editing using image or text prompts. In comparison with baseline methods, such as InstructPix2Pix (IP2P) [7] + Live3D [70] and PTI [55] + StyleCLIP [50] (details are illustrated in Sec. 4), our approach accurately follows the guidance from reference prompts and maintains sufficiently better efficiency. + +![](images/215ff6e676dd1831c92569bc9b0864817bfff9e03a3fe21b5f1dd19979eb5290.jpg) + +changes because it often overlooks precise features like hair and beard. Recent 3D GANs [13, 14, 58, 62, 75] show remarkable capabilities in generating high-fidelity 3D portraits. They can serve as powerful generative priors for 3D portrait editing when coupled with GAN inversion techniques [2, 3, 61, 86, 87]. However, these methods may encounter issues related to geometry distortion or exhibit slow speed, and the editing is constrained by the limited latent attributes. Besides, 2D diffusion models can offer a strong prior with Score Distillation Sampling (SDS) loss [52] for editing purposes [30]. Nonetheless, they often require step-by-step optimization and become a major bottleneck for real-time applications. + +To this end, we present 3DPE, a real-time 3D-aware Portrait Editing method driven by user-defined prompts. As shown in Fig. 1, when provided with a single-view portrait image, our approach empowers a versatile range of editing styles with one model through flexible instructions, including images and texts. We leverage the powerful 3D prior from a 3D-aware face generator and achieve a high-fidelity 3D reconstruction of the portrait image. Subsequently, we distill editing knowledge from a text-guided image editing model into a lightweight module integrated with the 3D-aware generator. This module, characterized by its minimal computational cost, allows our method to maintain real-time inference and excel in handling various types of editing, while ensuring good 3D consistency. An additional advantage is that our model supports customization through user-specified prompts with fast adaptation speed. This empowers users to build their own editing model at a minimal cost, enabling our method to cater to a broader audience. + +Our method achieves real-time 3D-aware portrait editing through the utilization of a feedforward network, with a processing time of $40\mathrm{ms}$ on a standard consumer GPU. Additionally, we present a comprehensive evaluation of + +our method using various prompts both quantitatively and qualitatively. Our design choices are also validated through comparisons with ablated variants of our method. Compared with baseline methods, our approach demonstrates superior 3D consistency, precise texture alignment, and a substantial improvement in inference time, as reflected by the evaluation metrics. We demonstrate the versatility of our method by showcasing its capacity to perform a wide variety of edits, including text and image prompts, on portrait images. + +In summary, the contributions of our work include: + +- We propose a lightweight module to distill knowledge from 3D GANs and diffusion models for 3D-aware editing from a single image. Due to the minimal cost of the new module, our model maintains real-time performance. +- Our model supports fast adaptation to user-specified editing, requiring only 10 image pairs and 5 minutes for the adaptation. +- Our framework can accommodate various control signals, including text and image prompts. + +# 2 Related Work + +# 2.1 Generative Face Priors + +Generative models aim at modeling the underlying distribution of the data, containing a wealth of prior knowledge. Recently, 3D GANs [13, 14, 25, 47, 49, 58, 59, 63-65, 69, 71, 77] are mostly adopted to learn 3D faces from single-view image dataset. The interior rich domain-specific geometry priors enable various face-related applications such as image editing [9, 66, 67, 70, 79, 85] and domain adaptation [1, 34, 36, 46, 82]. Large-scale diffusion models [56], in contrast, encode knowledge of huge datasets and thus can provide general prior information. Such priors are broadly leveraged for tasks such as image editing [7, 10, 24, 45, 83], customization [31, 38, 44, 57], and video editing [11, 12, 23, 43, 48, 53, 73, 80]. In our approach, we capitalize on the advantages of both the geometry prior derived from 3D GANs and the broader editing prior offered by large-scale text-to-image diffusion models, instead of relying solely on a single type of prior. + +# 2.2 Portrait Editing from a Single Image + +Although many methods work well for reconstructing [22, 26] or generating [13] faces, editing has become a necessary interface to connect these methods with real-world applications. Previously, portrait editing given a single-view image was mostly completed in 2D image space facilitated with GAN inversion [2, 3, 5, 28, 55, 60, 61, 78, 86], which enables fast editing by exploring the trajectories in the GAN's latent space. Recent diffusion-based portrait editing [7, 83] can support various editing types with texts as guidance. However, most of them are done in the 2D space, and thus there is no guarantee for the underlying 3D consistency. Therefore, 3D-aware portrait editing is crucial to achieve the goal. Some methods [15, 33, 39, 42, 66, 67, 74, 76, 81] rely on the latent space + +of 3D GANs to perform editing through walking in the latent space, but are limited by the number of editable latent attributes. A few methods, such as Instruct-NeRF2NeRF [27], ClipFace [4], and LENeRF [32], attempt to leverage the large-scale models [54, 56] to edit on 3D representations with texts as guidance but require heavy iterative refinement of the edited results. Similarly, InstructPix2NeRF [40] leverages text-to-image models and text prompts to edit within the latent space of the GAN-based 3D generator. However, it is held back by the multi-step diffusion inference and the heavy structure of a GAN-based generator. Ours, in contrast, benefits from the lightweight design of the module and achieves real-time editing with ease. With such a design, our model can also adapt fast to the user-defined editing prompts. Moreover, texts cannot always illustrate the desired effects precisely, and sometimes an image prompt serves as better guidance for image editing [20, 57]. Our method can support not only text prompts but also image prompts for editing, providing a more flexible and user-friendly interface. + +# 3 Method + +Our approach takes a portrait image $\mathbf{I}$ and its camera pose $\mathbf{c}$ , obtained via the face pose estimator [18]. Additionally, the referenced prompts $\mathbf{P}$ (images or texts) serve as the editing instructions. The outcome of our model is an edited version of the 3D portraits characterized by NeRF in accordance with the prompts. The edited image is denoted as $\mathbf{I}_p$ . Within our framework, we achieve this objective by leveraging the knowledge of 3D GANs and a text-guided image editing model, which is distilled into a lightweight module. Following distillation, our method enables real-time 3D portrait editing and efficient adaptation to user-specified prompts. In Sec. 3.1, we offer an illustration of 3D GANs and diffusion models. The process of distilling these priors into a lightweight module is outlined in Sec. 3.2. Finally, the details for model training and inference, along with the fast adaptation for customized prompts, are presented in Sec. 3.3. + +# 3.1 Preliminary + +3D GAN Prior for Portrait Reconstruction. The 3D-aware GANs showcase the ability to synthesize photorealistic 3D images using a collection of single-view images. Notably, EG3D [13] introduces an efficient triplane 3D representation, demonstrating high-quality 3D-aware image rendering. Once trained, the generator of EG3D can be applied for single-image 3D reconstruction via GAN inversion. However, existing 3D GAN inversion methods often encounter geometry distortion or exhibit slow inference speed. To tackle these challenges, we utilize Live3D [70], a state-of-the-art single-image portrait 3D reconstruction model built upon EG3D, preserving geometry quality while ensuring real-time performance. Live3D employs a two-branch encoder, $\mathrm{E}_{high}(\cdot)$ and $\mathrm{E}_{low}(\cdot)$ , to extract different resolution features from the input. It then utilizes a ViT-based + +![](images/9074058ef0a16f926b765e7d2abeb7ca60bafbd451e32105f57698762302864c.jpg) +Fig. 2: Overview of our method. We distill the prior in the diffusion model and 3D GAN for real-time 3D-aware editing. Our approach is fine-tuned from Live3D [70], where we extract features from the input portrait I using $\mathrm{E}_{high}(\cdot)$ and $\mathrm{E}_{low}(\cdot)$ . The prompt embedding is generated with $\mathrm{E}_p(\cdot)$ and injected with the input features from $\mathrm{E}_{high}(\cdot)$ through a cross-attention mechanism. Our model is trained to mimic the output from the diffusion prior to acquire editing knowledge and enforce geometry constraints through triplane, multi-view images, and depth supervision from the 3D prior. In this context, InstructPix2Pix [7] and Live3D serve as the diffusion and 3D prior, respectively. It's noteworthy that only $\mathrm{E}_p(\cdot)$ and $\mathrm{E}_t(\cdot)$ are learnable during training, while all other parameters remain frozen. + +$\mathrm{E}_t(\cdot)$ decoder [19] to transform the fused encoder features into the triplane $\mathbf{T}$ : + +$$ +\mathbf {T} = \mathrm {E} _ {t} \left(\mathrm {E} _ {\text {h i g h}} (\mathbf {I}), \mathrm {E} _ {\text {l o w}} (\mathbf {I})\right). \tag {1} +$$ + +This triplane $\mathbf{T}$ is subsequently used in conjunction with the volume rendering module and upsampler of EG3D to generate photorealistic view synthesis given the camera pose c. For clarity, we use $\mathrm{R}(\cdot)$ to denote this process. + +Diffusion Prior for Portrait Editing. Only performing 3D reconstruction is insufficient as our goal is to edit portrait images. Large-scale diffusion models [56], trained on vast text-image pairs, can synthesize realistic photos with text input, offering powerful editing prior. Despite their effectiveness for face images, the editing process is often slow due to step-by-step optimization or multiple inferences in diffusion models. The scarcity of high-quality and multi-view consistent paired data significantly hinders the practical application of 3D-aware editing from a single image, which is hard to acquire with diffusion models. Therefore, our goal is to distill editing knowledge from the diffusion model, integrate it with the 3D prior in Live3D into a lightweight module, and employ it for real-time portrait editing. + +![](images/f81ee83c8db8f852b73649c43459d539b6ad26c055042c6a1c778eb05c796b5f.jpg) +Fig. 3: Disentanglement in Live3D features. We separately disable the features from $\mathrm{E}_{high}(\cdot)$ and $\mathrm{E}_{low}(\cdot)$ to infer the reconstructed image. Without $\mathrm{E}_{high}(\cdot)$ , the output retains the coarse structure but loses its appearance. Conversely, when $\mathrm{E}_{low}(\cdot)$ is deactivated, the reconstructed portraits preserve the texture (such as the blue and purple reflection on the glasses) but fail to capture the geometry. + +# 3.2 Distilling Priors into a Lightweight Module + +Our method aims to perform real-time 3D-aware editing for various prompts for the single-view portrait. Thus, it needs powerful 3D knowledge for geometry reconstruction and editing prior to handling various control signals. We leverage the strengths of both 3D GANs and diffusion models. In the following, we provide detailed presentations on how the knowledge of these two types of models is distilled into a lightweight module. + +Feature Representation in Live3D. We carefully study the Live3D model and discover that, as a two-branch triplane-based 3D reconstruction model, the low-resolution and high-resolution features from the Live3D encoder $\mathrm{E}_{low}(\cdot)$ and $\mathrm{E}_{high}(\cdot)$ tend to learn various levels of information without explicit guidance. We conduct a study by disabling one of the branches and inferring the reconstructed images. As illustrated in Fig. 3, when the high-resolution encoder $\mathrm{E}_{high}(\cdot)$ is disabled with random weights, the inference image retains a similar structure but loses its detailed appearance. Conversely, when the low-resolution encoder $\mathrm{E}_{low}(\cdot)$ is disabled, the reconstructed portraits preserve some of the texture from the input but struggle to capture the geometry. Based on this analysis, the features of the two branches separately model low-frequency and high-frequency information. This insight motivates our model design in terms of how to distill knowledge from the diffusion model and 3D GANs. + +Geometry Prediction for Inputs I. In our framework, we use the prompts $\mathbf{P}$ to refine the input portrait $\mathbf{I}$ into the edited image $\mathbf{I}_p$ . The $\mathbf{I}_p$ inherits a similar structure to the input $\mathbf{I}$ , providing coarse geometry or structural guidance in the editing process. As discussed in the above section, the encoder $\mathrm{E}_{low}(\cdot)$ consistently generates low-frequency features to represent geometry cues, making it well-suited for preserving structured information in the input portrait. As a result, we freeze $\mathrm{E}_{low}(\cdot)$ and leverage it to produce structure features $\mathbf{F}_g = \mathrm{E}_{low}(\mathbf{I})$ for coarse geometry prediction. + +Injecting Prompts P as Condition. In contrast to input images, prompts typically offer more high-frequency and texture information to guide and control the editing process. Accordingly, it is appropriate to incorporate the prompts into the high-level branch $\mathrm{E}_{high}(\cdot)$ in Live3D to extract the high-frequency features. To incorporate prompts, we utilize the prompt encoder $\mathrm{E}_p(\cdot)$ to create prompt embeddings that are fused into our model through cross attention, which is similar to the strategy in Stable Diffusion [56]. Specifically, we add a cross-attention layer after $\mathrm{E}_{high}(\cdot)$ to obtain the feature updated with the prompt embeddings: + +$$ +\mathbf {F} _ {a} = \operatorname {C r o s s A t t e n t i o n} \left(\operatorname {E} _ {\text {h i g h}} (\mathbf {I}), \operatorname {E} _ {p} (\mathbf {P})\right), \tag {2} +$$ + +where the encoder $\mathbf{E}_p(\cdot)$ is a transformer (MAE [29] for image and CLIP [54] for text). The $\mathrm{E}_{high}(\mathbf{I})$ serves as the query and $\mathrm{E}_p(\mathbf{P})$ is employed as key and value. The updated feature $\mathbf{F}_a$ and the geometry feature $\mathbf{F}_g$ are then feed into the decoder $\mathrm{E}_t(\cdot)$ to infer the triplane $\mathbf{T_p}$ : + +$$ +\mathbf {T} _ {\mathbf {p}} = \operatorname {E} _ {t} \left(\mathbf {F} _ {a}, \mathbf {F} _ {g}\right). \tag {3} +$$ + +Distilling Diffusion and 3D GAN Prior. Unlike previous methods that require large amounts of 3D data, our model operates with only 2D paired data generated by the 2D editing model [7]. For both visual and textual prompts, we generate edited images $\mathbf{I}_{gt}$ with the diffusion-based image editing model [7] as the pseudo labels. The details of data generation can be found in Sec. 4.1. + +With the predicted triplane $\mathbf{T}_{\mathbf{p}}$ , we can render the image using the pretrained EG3D model and then calculate the reconstruction loss as follows: + +$$ +\mathcal {L} _ {2 d} = \ell_ {\mathbf {I}} \left(\mathbf {I} _ {g t}, \mathrm {R} \left(\mathbf {T} _ {\mathbf {p}}, \mathbf {c}\right)\right), \tag {4} +$$ + +where $\mathrm{R}(\cdot)$ is the rendering module of EG3D, $\mathbf{c}$ is the camera pose of input portrait, and $\ell_1(\cdot)$ is an image reconstruction loss penalizing the difference between the ground truth $\mathbf{I}_{gt}$ and the rendering $\mathbf{I}_p = \mathrm{R}(\mathbf{T_p},\mathbf{c})$ . It's important to note that $\mathcal{L}_{2d}$ , used to reconstruct the $\mathbf{I}_{gt}$ , essentially distills the knowledge of editing from the diffusion model. + +We observe that $\mathcal{L}_{2d}$ helps the model to reconstruct well on the input camera view but suffers from geometry distortion during novel-view synthesis. To fully exploit the 3D properties of Live3D, we propose to distill the 3D knowledge of Live3D. Specifically, we leverage the pretrained Live3D to infer the triplane $\mathbf{T}_{gt}$ , and multi-view depths $\mathcal{D}_{gt}$ and images $\mathcal{T}_{gt}$ of the pseudo-label image $\mathbf{I}_{gt}$ : + +$$ +\mathbf {T} _ {g t}, \mathcal {D} _ {g t} = \mathrm {G} (\mathbf {I} _ {g t}, \mathcal {C}), \tag {5} +$$ + +where $\mathrm{G}(\cdot)$ represents the inference process of Live3D, where $\mathcal{C} = \{\mathbf{c}_1,\dots,\mathbf{c}_n\}$ is the camera set, and $n$ denotes the number of cameras. We also render the multi-view images $\mathcal{I}_p$ and depths $\mathcal{D}_p$ of the edited image $\mathrm{I}_p$ from the triplane $\mathbf{T}_p$ and then define the objective: + +$$ +\mathcal {L} _ {3 d} = \ell_ {\mathcal {I}} \left(\mathcal {I} _ {p}, \mathcal {I} _ {g t}\right) + \ell_ {\mathbf {T}} \left(\mathbf {T} _ {p}, \mathbf {T} _ {g t}\right) + \ell_ {\mathcal {D}} \left(\mathcal {D} _ {p}, \mathcal {D} _ {g t}\right), \tag {6} +$$ + +where $\ell_{\mathcal{I}}(\cdot), \ell_{\mathbf{T}}(\cdot)$ and $\ell_{\mathcal{D}}(\cdot)$ is the reconstruction loss penalizing the difference in image, triplane, and depth between $\mathrm{I}_p$ and $\mathrm{I}_{gt}$ . + +# 3.3 Training and Inference + +Training. During the training phase, we sample a triplet consisting of the input portrait $\mathbf{I}$ along with its camera pose $\mathbf{c}$ , prompts $\mathbf{P}$ , and the pseudo-label image $\mathbf{I}_{gt}$ generated by the 2D diffusion model. We leverage the Live3D model as a pretrained model and add an additional prompts encoder $\mathrm{E}_p$ to extract prompt embeddings. The overall learning objective can be described as follows: + +$$ +L = \lambda_ {1} L _ {2 D} + \lambda_ {2} L _ {3 D}, \tag {7} +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ are loss weights. In our setting, $\lambda_{1}$ and $\lambda_{2}$ are both set to 1.0. For the reconstruction loss, $\ell_{\mathcal{I}}(\cdot)$ , $\ell_{\mathrm{I}}(\cdot)$ are combinations of $\mathrm{L}_2$ loss and LPIPS loss [84], with loss weights being 1 and 2, respectively. $\ell_{\mathcal{D}}(\cdot)$ and $\ell_{\mathrm{T}}(\cdot)$ are $\mathrm{L}_1$ loss. Notably, based on the study on Live3D feature representation, during training only $\mathrm{E}_p(\cdot)$ , $\mathrm{E}_t(\cdot)$ are set to be learnable, while other modules are frozen. It allows our model to leverage Live3D knowledge as much as possible and converge at a very fast speed. + +Inference. For inference, users can provide a single 2D portrait image and choose an image prompt or text prompt. Our model is able to generate the edited 3D NeRF along with photorealistic view synthesis. + +Customized Prompts Adaptation. To accommodate customized styles provided by users, we propose a method to adapt our pretrained encoder to user-defined styles. We increase the tuning efficiency by optimizing only $\mathrm{E}_p(\cdot)$ and the normalization layers in $\mathrm{E}_t(\cdot)$ with the same learning objective Eq. 7. This method allows us to limit the training data to only 10 image pairs and the learning time to 5 minutes on a single GPU. + +# 4 Experiments + +# 4.1 Experimental Setup + +Training Settings. For real faces, we adopt the FFHQ dataset [35] at $512 \times 512$ resolution with camera parameters aligned by EG3D [13]. In order to obtain the stylized images as pseudo labels, and the corresponding textual and visual prompts, we leverage the text-guided editing model InstructPix2Pix (IP2P) [7] to + +![](images/fac216f03f4f959144cd3e01c13277758aa34d7a06948f1663c2948f73e71b77.jpg) +mnd + +![](images/b43743ade296fac25742324ff4fa9a9df10d49b1af67b1428a50f37c9aee2567.jpg) + +![](images/dd1544dd3fe5b6fe651dd1d5be29ecea4c6ee7378a529a456927ab81272ee084.jpg) + +![](images/061d13bb5dff2ada5f8b148d1dc34ad11fc1e99106b14272343871933008d364.jpg) +Recolor it to a Gothic color scheme. + +![](images/90d5bc38436b57d70008b667788d2b87ebaab64700b2695fa04eafcb62090631.jpg) +P 12345678901001100110011001100110011001100110011001100110011001100110011001100110011001100110011001100110011001100 + +![](images/30e99cad967c55e84925a0c9e13c7325270c8f7988317ca53a515906bbe71ed2.jpg) + +![](images/705204cfb497a2a8f845cc8ab25e4e1106f2351b510e8fa0362e906930690810.jpg) + +![](images/9768e369ab754eb13086513b97f1af71fd66067a6fb161bf59c1dc6581884197.jpg) + +![](images/26263c704c8728002033135f47f0b7e7721fa49d6d2cca01d12fd9c457879a32.jpg) +0000000000000000000000000000000000000 + +![](images/fc936265282935b0a001880613bddc357adb22e9187d45e8eb7a1b616e360b41.jpg) +Fig. 4: Qualitative comparisons. We compare the results of several baselines with image prompts and text prompts. In each case, we include the edited portraits as well as their novel view renderings. Our method generates high-quality edited portraits with better 3D quality and alignment with the referenced prompts. + +![](images/4e1bcb7134c60dbc4d990b8785227a03ac8cc1fbed1a333ea034658db9587a0e.jpg) + +![](images/cf3df868e514f235686b90c0d843bcec5a1b6f8ac7da4e878adc7f336dd682ba.jpg) + +edit the real faces. Specifically, we manually generate and select textual prompts, and send them to IP2P together with the real images to obtain the edited 2D images. The text prompts of our training and testing set are obtained by augmenting the aforementioned textual prompts with the large language model GPT [8] following IP2P [7]. While the image prompts of our training and testing set are picked from the edited 2D images. We conduct experiments on 20 styles, and for each style, we synthesize 1000 images, resulting in 20,000 image pairs with text prompts for model training. For each style, we use 8 textual prompts for training and 5 textual prompts for testing. We adopt a learning rate of 5e-5 and optimize the model for 60k iterations with a batch size of 32. The entire training procedure is completed over a period of 40 hours using 8 NVIDIA A100 GPUs. Baselines. We conduct comparisons against two categories of methods that achieve analogous outcomes: 1) 3D reconstruction coupled with 3D editing; 2) 2D editing followed by 3D reconstruction. For the first category, we implement three baseline approaches, wherein we firstly leverage PTI [55] or Live3D [70] for reconstruction and edit both of them following StyleCLIP [50]. Specifically, for experiments in PTI + StyleCLIP, we first conduct pivotal tuning [55] based on the predicted latent codes predicted by e4e [68] to reconstruct the given image in + +Table 1: Quantitative comparisons. We compare several baselines on the 100 images of FFHQ dataset. It's important to note that we exclude $CLIP_r$ for PTI+StyleCLIP and Live3D+StyleCLIP since these models utilize CLIP for optimization. Our model excels in 3D quality and achieves a remarkable 250x improvement (compared with IP2P+Live3D) in inference speed, achieving real-time performance. + +
MethodIDt↑CLIPr↑3D↑Time↓
PTI+StyleCLIP0.11-0.73360s
Live3D+StyleCLIP0.63-0.7230s
IP2P+Live3D0.440.620.7510s
Ours0.470.730.760.04s
+ +a 3D-aware manner. Then we perform editing on the latent codes following [50], where CLIP loss [54] is utilized between the model output and the prompt (no matter whether it is an image or text). The setting of Live3D + StyleCLIP is similar to PTI + StyleCLIP except that we obtain the triplane by Live3D and perform editing on this 3D representation. SPI [81] + StyleGAN-NADA [21] (StyleNADA for the abbreviation in Fig. 4) serves as the third baseline in the first category, where SPI is used for 3D reconstruction and StyleGAN-NADA is utilized to optimize the 3D generator for editing. In the second category, we employ a diffusion-based text-guided 2D image editing method IP2P [7] prior to executing 3D reconstruction via Live3D. IP2P being a text-instructed editing model, we need to first extract text prompts from the reference image with the image caption algorithm BLIP [41] when image prompts are input. The 2D edited results could then be obtained by sending the input image and the text prompt to IP2P. The final 3D results are acquired by inverting the 2D editing results with Live3D. + +Evaluation Criteria. We evaluate all metrics on 100 pairs of images processed from FFHQ. We conduct a comprehensive evaluation of the editing performance in the following four aspects: identity preservation, reference alignment, 3D consistency, and inference speed. 1) Identity preservation $(\mathbf{ID}_{\mathbf{t}})$ aims to measure the preservation of the original identity by calculating the cosine similarity between the identity feature of the input image $\mathbf{I}$ and that of the edited image $\mathbf{I}_p$ . We use the ArcFace model [16] to extract identity features. 2) Reference alignment $(\mathrm{CLIP_r})$ targets at assessing the alignment of the output editing styles to the desired input prompt by computing the cosine similarity in the CLIP [50] feature space. 3) 3D consistency (3D) on the edited outputs is measured following the evaluation protocols established by EG3D [13]. This involves calculating the identity similarity across multiple views. 4) Inference speed (Time) is measured on a single NVIDIA A6000 GPU with an average of 100 samples. + +![](images/f6c3f00a93f75379fee3f937e97790f28c47b3c39e0f7d18dc7e9c660cfd69bb.jpg) +Fig. 5: (a) Generated results of customized prompt adaptation and (b) its learning process. We show the intermediate testing results at 10s, 1min, 2min and 5min during adaptation for the novel style golden statue. + +![](images/840bba9faad9d542b7aabe52638d1a08a584d4deda24a2b2360447a1dbdcf07e.jpg) +Fig.6: Qualitative comparison for ablations on (a) the distillation loss of 3D GAN $(\mathcal{L}_{3D})$ and (b) diffusion models $(\mathcal{L}_{2D})$ . + +# 4.2 Efficient 3D-aware Portrait Editing + +We make an in-depth analysis of our efficient portrait editing method both quantitatively and qualitatively. The results are included in Tab. 1 and Fig. 4. For reference alignment, we do not report $CLIP_{r}$ for methods that leverage CLIP for optimization since it is evaluated with CLIP as well. The standout feature of our method is its efficiency. Our method achieves an inference speed of merely 40ms, which improves over 100 times compared to the fastest existing baselines, which require around 10 seconds. Because of the efficient knowledge distillation, our approach is also good in preserving the identity, adhering more closely to the reference prompts and achieving the best 3D consistency. + +As in Fig. 4, compared with PTI + CLIP methods, our model enables precise alignment with inputs and prompts, while the baseline always generates low-fidelity textures, and the resulting geometry has many artifacts. Although Live3D + CLIP achieves the best $ID_{t}$ score, the edited results are basically unchanged compared to the input. We suspect that Live3D does not possess a latent space and loses the editing priors, making it challenging for the results of CLIP optimization to align well with prompts. SPI + StyleGAN-NADA performs relatively slowly (18 minutes per instance) because both the 3D reconstruction and the editing are achieved by optimizing the generator. Although it can produce realistic textures, it fails to reconstruct fine details and to edit following the prompts. In contrast to the 2D editing and subsequent 3D reconstruction + +Table 2: Ablation study on the number of styles used for pretraining. We report LPIPS score after 0.1min, 1min, 2min, and 5min fine-tuning for evaluation. + +
#Style/Time0.1min1min2min5min
30.77580.63700.53120.5220
100.59320.57340.51060.5012
200.57030.52900.50220.4849
+ +Table 3: Ablation study on the number of data pairs used for novel prompt adaptation. We report LPIPS score after 0.1min, 1min, 2min, and 5min fine-tuning for evaluation. + +
#Pair/Time0.1min1min2min5min
20.61910.54310.52130.5212
50.61480.53240.50640.4971
100.60760.52420.49830.4935
200.59810.55250.49540.4895
500.58910.57180.49220.4797
+ +pipeline (IP2P + Live3D), our method produces textures more consistent with the prompt and aligns better with the input structure. + +# 4.3 Adaptation for Customized Editing + +With the trained editing network, our method already supports a variety of 3D-aware editing styles. Following DreamBooth [57] and to expand the selection and better conform to users' preferences, we offer an efficient method for fast adaptation to the user-defined customized prompts. Users can personalize the editing network by providing a modest set of 10 reference editing pairs. These pairs can be either handpicked from artist-created examples or produced using text-guided image editing models. As described in Sec. 3.3, only the partial of the encoder would be optimized. The adaptation process itself is remarkably fast, requiring only about 5 minutes to accomplish the learning of the customized prompts. Please also note that, once the user-defined style has been trained, all the incoming testing results can be obtained in a real-time manner. In Fig. 5, we present the testing results and learning process of the adaptation. Our model can quickly master the customized knowledge in about 2 minutes. With further training (e.g., 5 minutes), the edited results become more stylized, demonstrating a trade-off between authenticity and stylization. Upon completion of the adaptation, the method allows users to edit testing inputs in these newly learned styles with a minimal inference time of 0.04s. This rapid performance indicates that the method is well-optimized for real-time applications, providing a seamless and efficient user experience. + +# 4.4 Ablation Studies + +We analyze our model and validate our design choice by ablating the components. + +![](images/23913448181367f3e27cd2f751b0607ddf42b7e0911b6618b2faea15af9fe67c.jpg) +Fig. 7: Video editing results with 3DPE. We use text prompts "Make him a bronze statue" to edit the input video. Our method can accurately reconstruct and preserve challenging facial expressions and achieve high-quality novel view renderings. + +$\mathcal{L}_{3D}$ for Distillation of Live3D. We compare our model trained with and without $\mathcal{L}_{3D}$ . As shown in Fig. 6, we observe that the model without $\mathcal{L}_{3D}$ results in very flat geometry, and the synthesis of the side view loses the normal structure with many artifacts. This highlights the importance of the 3D knowledge from Live3D in training our model with 2D paired data, enabling the inference of reasonable face geometry without relying on any 3D data sources. + +$\mathcal{L}_{2D}$ for Distillation of Diffusion Models. We also study the effect of $\mathcal{L}_{2D}$ , which represents using 2D editing priors from the diffusion model [7] to optimize the encoder. From Fig. 6, we observe that $\mathcal{L}_{2D}$ is critical for our model, especially in preserving detailed texture. Without it, the edited image tends to lose the appearance information from prompts and exhibits structural artifacts around the nose. + +Ablation on Pretraining Data Scale for Customized Prompt Adaptations. In this part, we conduct ablation studies on the pretraining scale and investigate its influence on customized propmts. Specifically, we choose various 3DPE models pretrained with different style amounts (3, 10, and 20 styles respectively) and perform novel prompt adaptation on a single A6000 GPU. We choose the golden statue as the customized style and report LPIPS score between the predicted stylized images and the pseudo stylized labels produced by IP2P [7]. Experimental results in Tab. 2 reveal that the proposed method could benefit from the improving pretraining data scale and adapt faster and eventually perform better with more data, which suggests the potential of the model for future scaling up. + +# Ablation on Tuning Data Scale for Customized Prompt Adaptations. + +To investigate the ability of our method to adapt to user-defined novel prompts, we conduct an ablation study on the number of novel-prompt data pairs required for training. Similar to the prior experiments on the pretraining scale, we perform adaptation with 2, 5, 10, 20, or 50 pairs of novel-prompt data on a single A6000 GPU. We experimentally find that our model can converge in about 5 minutes, and therefore, we report the perceptual distance between the model output and the provided ground truth at $0.1\mathrm{min}$ , $1\mathrm{min}$ , $2\mathrm{min}$ , and $5\mathrm{min}$ for comparison. As shown in Tab. 3, our model can adapt to the new style in around 2 minutes. As the number of provided data pairs increases, the model can adapt faster at the beginning. + +# 4.5 Applications + +One application of our pipeline is to edit videos, requiring accurate reconstruction and efficient editing simultaneously. We demonstrate the effectiveness of our pipeline on talking face videos, as illustrated in Fig. 7. In this experiment, we perform sequence editing in a frame-by-frame manner by sending an image sequence from TalkingHead-1KH [72] dataset together with the text prompt "Make him a bronze statue" to 3DPE. The three rows in Fig. 7 respectively illustrate the input frames, the edited frames, and the novel views. Our method accurately reconstructs facial expressions, while achieving high-quality editing results. Since our method is 3D-aware, we can naturally view videos from novel perspectives. To better experience our method, we design an interactive system that allows users for interactive editing. Users are required to provide the image or text prompt to indicate the desired style and input an input image that is to be edited, and the system will apply our method on the given image and prompt to output the edited image as well as novel-view edited results. + +# 5 Discussion + +Limitations. Despite achieving state-of-the-art performance in quality and efficiency, our method exhibits inconsistencies in details for novel-view rendering because the EG3D framework relies on a super-resolution module. Additionally, when our method is applied to video editing, it presents flickering artifacts since our model is originally designed for image-based editing. + +Conclusion. In this work, we introduce 3DPE for real-time 3D-aware portrait editing. By distilling the powerful knowledge of diffusion models and 3D GANs into a lightweight module, our model significantly reduces editing time while ensuring quality. Benefiting from the distillation process, 3DPE could handle multiple styles in a single model and perform both image-guided and text-guided editing. The advantages of our method empower us to generate photorealistic 3D portraits, a capability crucial for the visual effects industry, AR/VR systems, and teleconferencing, among other applications, while we firmly oppose any misuse that could compromise privacy and security. + +# Acknowledgements + +This project was supported by the National Key R&D Program of China under grant number 2022ZD0161501. + +# References + +1. Abdal, R., Lee, H.Y., Zhu, P., Chai, M., Siarohin, A., Wonka, P., Tulyakov, S.: 3DAvatarGAN: Bridging domains for personalized editable avatars. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) +2. Abdal, R., Qin, Y., Wonka, P.: Image2StyleGAN: How to embed images into the StyleGAN latent space? In: Int. Conf. Comput. Vis. (2019) +3. Abdal, R., Qin, Y., Wonka, P.: Image2StyleGAN++: How to edit the embedded images? In: IEEE Conf. Comput. Vis. Pattern Recog. (2020) +4. Aneja, S., Thies, J., Dai, A., Nießner, M.: ClipFace: Text-guided editing of textured 3d morphable models. In: ACM SIGGRAPH Conf. Proc. (2023) +5. Bai, Q., Xu, Y., Zhu, J., Xia, W., Yang, Y., Shen, Y.: High-fidelity GAN inversion with padding space. In: Eur. Conf. Comput. Vis. (2022) +6. Blanz, V., Vetter, T.: A morphable model for the synthesis of 3d faces. In: Seminal Graphics Papers: Pushing the Boundaries (2023) +7. Brooks, T., Holynski, A., Efros, A.A.: InstructPix2Pix: Learning to follow image editing instructions. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) +8. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: Adv. Neural Inform. Process. Syst. (2020) +9. Cai, S., Obukhov, A., Dai, D., Van Gool, L.: Pix2NeRF: Unsupervised conditional p-gan for single image to neural radiance fields translation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) +0. Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., Zheng, Y.: MasaCtrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In: Int. Conf. Comput. Vis. (2023) +1. Ceylan, D., Huang, C.H.P., Mitra, N.J.: Pix2Video: Video editing using image diffusion. In: Int. Conf. Comput. Vis. (2023) +2. Chai, W., Guo, X., Wang, G., Lu, Y.: StableVideo: Text-driven consistency-aware diffusion video editing. In: Int. Conf. Comput. Vis. (2023) +3. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) +4. Chan, E.R., Monteiro, M., Kellnhofer, P., Wu, J., Wetzstein, G.: pi-GAN: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021) +5. Chang, S., Kim, G., Kim, H.: HairNeRF: Geometry-aware image synthesis for hairstyle transfer. In: Int. Conf. Comput. Vis. (2023) +6. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: ArcFace: Additive angular margin loss for deep face recognition. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019) +7. Deng, Y., Yang, J., Chen, D., Wen, F., Tong, X.: Disentangled and controllable face image generation via 3d imitative-contrastive learning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020) + +18. Deng, Y., Yang, J., Xu, S., Chen, D., Jia, Y., Tong, X.: Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019) +19. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Int. Conf. Learn. Represent. (2021) +20. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. In: Int. Conf. Learn. Represent. (2023) +21. Gal, R., Patashnik, O., Maron, H., Bermano, A.H., Chechik, G., Cohen-Or, D.: StyleGAN-NADA: Clip-guided domain adaptation of image generators. ACM Trans. Graph. (2022) +22. Gao, C., Shih, Y., Lai, W.S., Liang, C.K., Huang, J.B.: Portrait neural radiance fields from a single image. arXiv preprint arXiv:2012.05903 (2020) +23. Geyer, M., Bar-Tal, O., Bagon, S., Dekel, T.: TokenFlow: Consistent diffusion features for consistent video editing. In: Int. Conf. Learn. Represent. (2024) +24. Graikos, A., Malkin, N., Jojic, N., Samaras, D.: Diffusion models as plug-and-play priors. In: Adv. Neural Inform. Process. Syst. (2022) +25. Gu, J., Liu, L., Wang, P., Theobalt, C.: StyleNeRF: A style-based 3d-aware generator for high-resolution image synthesis. In: Int. Conf. Learn. Represent. (2022) +26. Guo, J., Yu, J., Lattas, A., Deng, J.: Perspective reconstruction of human faces by joint mesh and landmark regression. In: Eur. Conf. Comput. Vis. (2022) +27. Haque, A., Tancik, M., Efros, A.A., Holynski, A., Kanazawa, A.: Instruct-NeRF2NeRF: Editing 3d scenes with instructions. In: Int. Conf. Comput. Vis. (2023) +28. Harkonen, E., Hertzmann, A., Lehtinen, J., Paris, S.: GANSpace: Discovering interpretable GAN controls. Adv. Neural Inform. Process. Syst. (2020) +29. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) +30. Hertz, A., Aberman, K., Cohen-Or, D.: Delta denoising score. In: Int. Conf. Comput. Vis. (2023) +31. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: Int. Conf. Learn. Represent. (2022) +32. Hyung, J., Hwang, S., Kim, D., Lee, H., Choo, J.: Local 3d editing via 3d distillation of clip knowledge. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) +33. Jiang, K., Chen, S.Y., Fu, H., Gao, L.: NeRFFaceLighting: Implicit and disentangled face lighting representation leveraging generative prior in neural radiance fields. ACM Trans. Graph. (2023) +34. Jin, W., Ryu, N., Kim, G., Baek, S.H., Cho, S.: Dr. 3D: Adapting 3d GANs to artistic drawings. In: ACM SIGGRAPH Asia Conf. Proc. (2022) +35. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019) +36. Kim, G., Chun, S.Y.: DATID-3D: Diversity-preserved domain adaptation using text-to-image diffusion for 3d generative model. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) +37. Ko, J., Cho, K., Choi, D., Ryoo, K., Kim, S.: 3D GAN inversion with pose optimization. In: IEEE Winter Conf. Appl. Comput. Vis. (2023) + +38. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) +39. Li, J., Li, J., Zhang, H., Liu, S., Wang, Z., Xiao, Z., Zheng, K., Zhu, J.: PREIM3D: 3d consistent precise image attribute editing from a single image. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) +40. Li, J., Liu, S., Liu, Z., Wang, Y., Zheng, K., Xu, J., Li, J., Zhu, J.: Instruct-Pix2NeRF: Instructed 3d portrait editing from a single image. arXiv preprint arXiv:2311.02826 (2023) +41. Li, J., Li, D., Xiong, C., Hoi, S.: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: Int. Conf. Mach. Learn. (2022) +42. Lin, C., Lindell, D., Chan, E., Wetzstein, G.: 3D GAN inversion for controllable portrait image animation. In: Eur. Conf. Comput. Vis. Workshop (2022) +43. Liu, S., Zhang, Y., Li, W., Lin, Z., Jia, J.: Video-P2P: Video editing with cross-attention control. arXiv preprint arXiv:2303.04761 (2023) +44. Liu, Z., Feng, R., Zhu, K., Zhang, Y., Zheng, K., Liu, Y., Zhao, D., Zhou, J., Cao, Y.: Cones: Concept neurons in diffusion models for customized generation. In: Int. Conf. Mach. Learn. (2023) +45. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: Int. Conf. Learn. Represent. (2022) +46. Nitzan, Y., Gharbi, M., Zhang, R., Park, T., Zhu, J.Y., Cohen-Or, D., Shechtman, E.: Domain expansion of image generators. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) +47. Or-El, R., Luo, X., Shan, M., Shechtman, E., Park, J.J., Kemelmacher-Shlizerman, I.: StyleSDF: High-resolution 3d-consistent image and geometry generation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) +48. Ouyang, H., Wang, Q., Xiao, Y., Bai, Q., Zhang, J., Zheng, K., Zhou, X., Chen, Q., Shen, Y.: CoDeF: Content deformation fields for temporally consistent video processing. arXiv preprint arXiv:2308.07926 (2023) +49. Pan, X., Xu, X., Loy, C.C., Theobalt, C., Dai, B.: A shading-guided generative implicit model for shape-accurate 3d-aware image synthesis. Adv. Neural Inform. Process. Syst. (2021) +50. Patashnik, O., Wu, Z., Shechtman, E., Cohen-Or, D., Lischinski, D.: StyleCLIP: Text-driven manipulation of StyleGAN imagery. In: Int. Conf. Comput. Vis. (2021) +51. Paysan, P., Knothe, R., Amberg, B., Romdhani, S., Vetter, T.: A 3d face model for pose and illumination invariant face recognition. In: IEEE international conference on advanced video and signal based surveillance (2009) +52. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: DreamFusion: Text-to-3d using 2d diffusion. In: Int. Conf. Learn. Represent. (2023) +53. Qi, C., Cun, X., Zhang, Y., Lei, C., Wang, X., Shan, Y., Chen, Q.: FateZero: Fusing attentions for zero-shot text-based video editing. In: Int. Conf. Comput. Vis. (2023) +54. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Int. Conf. Mach. Learn. (2021) +55. Roich, D., Mokady, R., Bermano, A.H., Cohen-Or, D.: Pivotal tuning for latent-based editing of real images. ACM Trans. Graph. (2022) + +56. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) +57. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dream-Booth: Fine tuning text-to-image diffusion models for subject-driven generation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) +58. Schwarz, K., Liao, Y., Niemeyer, M., Geiger, A.: GRAF: Generative radiance fields for 3d-aware image synthesis. In: Adv. Neural Inform. Process. Syst. (2020) +59. Schwarz, K., Sauer, A., Niemeyer, M., Liao, Y., Geiger, A.: VoxGRAF: Fast 3d-aware image synthesis with sparse voxel grids. In: Adv. Neural Inform. Process. Syst. (2022) +60. Shen, Y., Gu, J., Tang, X., Zhou, B.: Interpreting the latent space of GANs for semantic face editing. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020) +61. Shen, Y., Yang, C., Tang, X., Zhou, B.: InterFaceGAN: Interpreting the disentangled face representation learned by GANs. IEEE Trans. Pattern Anal. Mach. Intell. (2020) +62. Shi, Z., Peng, S., Xu, Y., Geiger, A., Liao, Y., Shen, Y.: Deep generative models on 3d representations: A survey. arXiv preprint arXiv:2210.15663 (2022) +63. Shi, Z., Shen, Y., Xu, Y., Peng, S., Liao, Y., Guo, S., Chen, Q., Yeung, D.Y.: Learning 3d-aware image synthesis with unknown pose distribution. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) +64. Shi, Z., Xu, Y., Shen, Y., Zhao, D., Chen, Q., Yeung, D.Y.: Improving 3d-aware image synthesis with a geometry-aware discriminator. In: Adv. Neural Inform. Process. Syst. (2022) +65. Skorokhodov, I., Tulyakov, S., Wang, Y., Wonka, P.: EpiGRAF: Rethinking training of 3d gans. In: Adv. Neural Inform. Process. Syst. (2022) +66. Sun, J., Wang, X., Shi, Y., Wang, L., Wang, J., Liu, Y.: IDE-3D: Interactive disentangled editing for high-resolution 3d-aware portrait synthesis. ACM Trans. Graph. (2022) +67. Sun, J., Wang, X., Zhang, Y., Li, X., Zhang, Q., Liu, Y., Wang, J.: FENeRF: Face editing in neural radiance fields. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) +68. Tov, O., Alaluf, Y., Nitzan, Y., Patashnik, O., Cohen-Or, D.: Designing an encoder for StyleGAN image manipulation. ACM Trans. Graph. (2021) +69. Tran, P., Zakharov, E., Ho, L.N., Tran, A.T., Hu, L., Li, H.: VOODOO 3D: Volumetric portrait disentanglement for one-shot 3d head reenactment. In: IEEE Conf. Comput. Vis. Pattern Recog. (2024) +70. Trevithick, A., Chan, M., Stengel, M., Chan, E., Liu, C., Yu, Z., Khamis, S., Chandraker, M., Ramamoorthi, R., Nagano, K.: Real-time radiance fields for single-image portrait view synthesis. ACM Trans. Graph. (2023) +71. Wang, Q., Shi, Z., Zheng, K., Xu, Y., Peng, S., Shen, Y.: Benchmarking and analyzing 3d-aware image synthesis with a modularized codebase. In: Adv. Neural Inform. Process. Syst. (2023) +72. Wang, T.C., Mallya, A., Liu, M.Y.: One-shot free-view neural talking-head synthesis for video conferencing. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021) +73. Wang, W., Jiang, Y., Xie, K., Liu, Z., Chen, H., Cao, Y., Wang, X., Shen, C.: Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599 (2023) +74. Wang, Y., Xu, T., Wu, Y., Li, M., Chen, W., Xu, L., Yu, J.: NARRATE: A normal assisted free-view portrait stylizer. arXiv preprint arXiv:2207.00974 (2022) + +75. Xia, W., Xue, J.H.: A survey on deep generative 3d-aware image synthesis. ACM Computing Surveys (2023) +76. Xie, J., Ouyang, H., Piao, J., Lei, C., Chen, Q.: High-fidelity 3d GAN inversion by pseudo-multi-view optimization. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) +77. Xu, Y., Peng, S., Yang, C., Shen, Y., Zhou, B.: 3d-aware image synthesis via learning structural and textural representations. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) +78. Xu, Y., Shen, Y., Zhu, J., Yang, C., Zhou, B.: Generative hierarchical features from synthesizing images. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021) +79. Yang, S., Jiang, L., Liu, Z., Loy, C.C.: Pastiche master: Exemplar-based high-resolution portrait style transfer. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) +80. Yang, S., Zhou, Y., Liu, Z.,, Loy, C.C.: Rerender a video: Zero-shot text-guided video-to-video translation. In: ACM SIGGRAPH Asia Conf. Proc. (2023) +81. Yin, F., Zhang, Y., Wang, X., Wang, T., Li, X., Gong, Y., Fan, Y., Cun, X., Shan, Y., Oztireli, C., et al.: 3D GAN inversion with facial symmetry prior. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) +82. Zhang, J., Lan, Y., Yang, S., Hong, F., Wang, Q., Yeo, C.K., Liu, Z., Loy, C.C.: DeformToon3D: Deformable neural radiance fields for 3d toonification. In: Int. Conf. Comput. Vis. (2023) +83. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Int. Conf. Comput. Vis. (2023) +84. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018) +85. Zhou, P., Xie, L., Ni, B., Tian, Q.: CIPS-3D: A 3d-aware generator of gans based on conditionally-independent pixel synthesis. arXiv preprint arXiv:2110.09788 (2021) +86. Zhu, J., Shen, Y., Zhao, D., Zhou, B.: In-domain GAN inversion for real image editing. In: Eur. Conf. Comput. Vis. (2020) +87. Zhu, J.Y., Krahenbuhl, P., Shechtman, E., Efros, A.A.: Generative visual manipulation on the natural image manifold. In: Eur. Conf. Comput. Vis. (2016) \ No newline at end of file diff --git a/2024/Real-time 3D-aware Portrait Editing from a Single Image/images.zip b/2024/Real-time 3D-aware Portrait Editing from a Single Image/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..08841883d1996c60c12089cd3d13f76138479aee --- /dev/null +++ b/2024/Real-time 3D-aware Portrait Editing from a Single Image/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a2e815562f93877ca7769231789375db9174b625be696f189af3842523917ef +size 523168 diff --git a/2024/Real-time 3D-aware Portrait Editing from a Single Image/layout.json b/2024/Real-time 3D-aware Portrait Editing from a Single Image/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..45747790106f220133e2f76cefc9beb920087307 --- /dev/null +++ b/2024/Real-time 3D-aware Portrait Editing from a Single Image/layout.json @@ -0,0 +1,10968 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 178, + 111, + 435, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 111, + 435, + 148 + ], + "spans": [ + { + "bbox": [ + 178, + 111, + 435, + 148 + ], + "type": "text", + "content": "Real-time 3D-aware Portrait Editing from a Single Image" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 178, + 167, + 433, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 167, + 433, + 182 + ], + "spans": [ + { + "bbox": [ + 178, + 167, + 433, + 182 + ], + "type": "text", + "content": "Qingyan Bai" + }, + { + "bbox": [ + 178, + 167, + 433, + 182 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 178, + 167, + 433, + 182 + ], + "type": "text", + "content": ", Zifan Shi" + }, + { + "bbox": [ + 178, + 167, + 433, + 182 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 178, + 167, + 433, + 182 + ], + "type": "text", + "content": ", Yinghao Xu" + }, + { + "bbox": [ + 178, + 167, + 433, + 182 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 178, + 167, + 433, + 182 + ], + "type": "text", + "content": ", Hao Ouyang" + }, + { + "bbox": [ + 178, + 167, + 433, + 182 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 209, + 181, + 403, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 181, + 403, + 193 + ], + "spans": [ + { + "bbox": [ + 209, + 181, + 403, + 193 + ], + "type": "text", + "content": "Qiuyu Wang" + }, + { + "bbox": [ + 209, + 181, + 403, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 209, + 181, + 403, + 193 + ], + "type": "text", + "content": ", Ceyuan Yang" + }, + { + "bbox": [ + 209, + 181, + 403, + 193 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 209, + 181, + 403, + 193 + ], + "type": "text", + "content": ", Xuan Wang" + }, + { + "bbox": [ + 209, + 181, + 403, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 187, + 193, + 425, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 193, + 425, + 205 + ], + "spans": [ + { + "bbox": [ + 187, + 193, + 425, + 205 + ], + "type": "text", + "content": "Gordon Wetzstein3, Yujun Shen2, and Qifeng Chen1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 192, + 213, + 421, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 213, + 421, + 225 + ], + "spans": [ + { + "bbox": [ + 192, + 213, + 421, + 225 + ], + "type": "text", + "content": "1 The Hong Kong University of Science and Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 279, + 224, + 334, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 224, + 334, + 236 + ], + "spans": [ + { + "bbox": [ + 279, + 224, + 334, + 236 + ], + "type": "text", + "content": "2 Ant Group" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 261, + 236, + 351, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 236, + 351, + 247 + ], + "spans": [ + { + "bbox": [ + 261, + 236, + 351, + 247 + ], + "type": "text", + "content": "3 Stanford University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 252, + 247, + 361, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 247, + 361, + 258 + ], + "spans": [ + { + "bbox": [ + 252, + 247, + 361, + 258 + ], + "type": "text", + "content": "4 Shanghai AI Laboratory" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 285, + 452, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 285, + 452, + 437 + ], + "spans": [ + { + "bbox": [ + 159, + 285, + 452, + 437 + ], + "type": "text", + "content": "Abstract. This work presents 3DPE, a practical method that can efficiently edit a face image following given prompts, like reference images or text descriptions, in a 3D-aware manner. To this end, a lightweight module is distilled from a 3D portrait generator and a text-to-image model, which provide prior knowledge of face geometry and superior editing capability, respectively. Such a design brings two compelling advantages over existing approaches. First, our method achieves real-time editing with a feedforward network (i.e., " + }, + { + "bbox": [ + 159, + 285, + 452, + 437 + ], + "type": "inline_equation", + "content": "\\sim 0.04" + }, + { + "bbox": [ + 159, + 285, + 452, + 437 + ], + "type": "text", + "content": "s per image), over " + }, + { + "bbox": [ + 159, + 285, + 452, + 437 + ], + "type": "inline_equation", + "content": "100\\times" + }, + { + "bbox": [ + 159, + 285, + 452, + 437 + ], + "type": "text", + "content": " faster than the second competitor. Second, thanks to the powerful priors, our module could focus on the learning of editing-related variations, such that it manages to handle various types of editing simultaneously in the training phase and further supports fast adaptation to user-specified customized types of editing during inference (e.g., with " + }, + { + "bbox": [ + 159, + 285, + 452, + 437 + ], + "type": "inline_equation", + "content": "\\sim 5" + }, + { + "bbox": [ + 159, + 285, + 452, + 437 + ], + "type": "text", + "content": " min fine-tuning per style). Project page can be found here." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 160, + 449, + 361, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 449, + 361, + 459 + ], + "spans": [ + { + "bbox": [ + 160, + 449, + 361, + 459 + ], + "type": "text", + "content": "Keywords: 3D-aware portrait " + }, + { + "bbox": [ + 160, + 449, + 361, + 459 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 449, + 361, + 459 + ], + "type": "text", + "content": " Efficient editing" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 479, + 229, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 229, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 229, + 491 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 503, + 481, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 481, + 622 + ], + "type": "text", + "content": "Inferring the geometry and appearance from a single-view portrait image has become mature and practical [22, 37, 42, 66, 70, 74, 76], largely attributed to the utilization of priors in various 2D/3D generative models. However, only performing geometry reconstruction is insufficient. The significance of 3D portrait editing, driven by user intentions, and the need for streamlined efficiency in the editing process has been steadily increasing. This is particularly crucial in real-world applications such as AR/VR, 3D telepresence, and video conferencing, where real-time editing is often essential. Consequently, a key question arises: How can we effectively address the challenge of attaining high-fidelity portrait editing while ensuring real-time efficiency?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 623, + 481, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 623, + 481, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 623, + 481, + 647 + ], + "type": "text", + "content": "Traditional methods for 3D portrait editing [17] typically rely on template facial models [6,51], which has limitations in handling the substantial geometry" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 239, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 239, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 239, + 665 + ], + "type": "text", + "content": "* Corresponding authors." + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 115, + 309, + 254 + ], + "blocks": [ + { + "bbox": [ + 136, + 115, + 309, + 254 + ], + "lines": [ + { + "bbox": [ + 136, + 115, + 309, + 254 + ], + "spans": [ + { + "bbox": [ + 136, + 115, + 309, + 254 + ], + "type": "image", + "image_path": "ae23707245db9249b6c133432e8b7729bf5940f9afcb2fa0ff5ab65b2a3394c8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 264, + 482, + 321 + ], + "lines": [ + { + "bbox": [ + 130, + 264, + 482, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 482, + 321 + ], + "type": "text", + "content": "Fig. 1: Photorealistic editing results produced by our proposed 3DPE, which allows users to perform 3D-aware portrait editing using image or text prompts. In comparison with baseline methods, such as InstructPix2Pix (IP2P) [7] + Live3D [70] and PTI [55] + StyleCLIP [50] (details are illustrated in Sec. 4), our approach accurately follows the guidance from reference prompts and maintains sufficiently better efficiency." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 314, + 116, + 482, + 254 + ], + "blocks": [ + { + "bbox": [ + 314, + 116, + 482, + 254 + ], + "lines": [ + { + "bbox": [ + 314, + 116, + 482, + 254 + ], + "spans": [ + { + "bbox": [ + 314, + 116, + 482, + 254 + ], + "type": "image", + "image_path": "215ff6e676dd1831c92569bc9b0864817bfff9e03a3fe21b5f1dd19979eb5290.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 348, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 482, + 456 + ], + "type": "text", + "content": "changes because it often overlooks precise features like hair and beard. Recent 3D GANs [13, 14, 58, 62, 75] show remarkable capabilities in generating high-fidelity 3D portraits. They can serve as powerful generative priors for 3D portrait editing when coupled with GAN inversion techniques [2, 3, 61, 86, 87]. However, these methods may encounter issues related to geometry distortion or exhibit slow speed, and the editing is constrained by the limited latent attributes. Besides, 2D diffusion models can offer a strong prior with Score Distillation Sampling (SDS) loss [52] for editing purposes [30]. Nonetheless, they often require step-by-step optimization and become a major bottleneck for real-time applications." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 460, + 482, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 626 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 626 + ], + "type": "text", + "content": "To this end, we present 3DPE, a real-time 3D-aware Portrait Editing method driven by user-defined prompts. As shown in Fig. 1, when provided with a single-view portrait image, our approach empowers a versatile range of editing styles with one model through flexible instructions, including images and texts. We leverage the powerful 3D prior from a 3D-aware face generator and achieve a high-fidelity 3D reconstruction of the portrait image. Subsequently, we distill editing knowledge from a text-guided image editing model into a lightweight module integrated with the 3D-aware generator. This module, characterized by its minimal computational cost, allows our method to maintain real-time inference and excel in handling various types of editing, while ensuring good 3D consistency. An additional advantage is that our model supports customization through user-specified prompts with fast adaptation speed. This empowers users to build their own editing model at a minimal cost, enabling our method to cater to a broader audience." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Our method achieves real-time 3D-aware portrait editing through the utilization of a feedforward network, with a processing time of " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "40\\mathrm{ms}" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " on a standard consumer GPU. Additionally, we present a comprehensive evaluation of" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Bai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": "our method using various prompts both quantitatively and qualitatively. Our design choices are also validated through comparisons with ablated variants of our method. Compared with baseline methods, our approach demonstrates superior 3D consistency, precise texture alignment, and a substantial improvement in inference time, as reflected by the evaluation metrics. We demonstrate the versatility of our method by showcasing its capacity to perform a wide variety of edits, including text and image prompts, on portrait images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 201, + 372, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 201, + 372, + 211 + ], + "spans": [ + { + "bbox": [ + 146, + 201, + 372, + 211 + ], + "type": "text", + "content": "In summary, the contributions of our work include:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 220, + 479, + 303 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 220, + 479, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 220, + 479, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 220, + 479, + 255 + ], + "type": "text", + "content": "- We propose a lightweight module to distill knowledge from 3D GANs and diffusion models for 3D-aware editing from a single image. Due to the minimal cost of the new module, our model maintains real-time performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 256, + 479, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 256, + 479, + 279 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 479, + 279 + ], + "type": "text", + "content": "- Our model supports fast adaptation to user-specified editing, requiring only 10 image pairs and 5 minutes for the adaptation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 280, + 479, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 280, + 479, + 303 + ], + "spans": [ + { + "bbox": [ + 138, + 280, + 479, + 303 + ], + "type": "text", + "content": "- Our framework can accommodate various control signals, including text and image prompts." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 322, + 236, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 236, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 236, + 335 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 346, + 276, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 346, + 276, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 346, + 276, + 357 + ], + "type": "text", + "content": "2.1 Generative Face Priors" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 365, + 480, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 480, + 510 + ], + "type": "text", + "content": "Generative models aim at modeling the underlying distribution of the data, containing a wealth of prior knowledge. Recently, 3D GANs [13, 14, 25, 47, 49, 58, 59, 63-65, 69, 71, 77] are mostly adopted to learn 3D faces from single-view image dataset. The interior rich domain-specific geometry priors enable various face-related applications such as image editing [9, 66, 67, 70, 79, 85] and domain adaptation [1, 34, 36, 46, 82]. Large-scale diffusion models [56], in contrast, encode knowledge of huge datasets and thus can provide general prior information. Such priors are broadly leveraged for tasks such as image editing [7, 10, 24, 45, 83], customization [31, 38, 44, 57], and video editing [11, 12, 23, 43, 48, 53, 73, 80]. In our approach, we capitalize on the advantages of both the geometry prior derived from 3D GANs and the broader editing prior offered by large-scale text-to-image diffusion models, instead of relying solely on a single type of prior." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 526, + 347, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 526, + 347, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 526, + 347, + 538 + ], + "type": "text", + "content": "2.2 Portrait Editing from a Single Image" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 545, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 479, + 665 + ], + "type": "text", + "content": "Although many methods work well for reconstructing [22, 26] or generating [13] faces, editing has become a necessary interface to connect these methods with real-world applications. Previously, portrait editing given a single-view image was mostly completed in 2D image space facilitated with GAN inversion [2, 3, 5, 28, 55, 60, 61, 78, 86], which enables fast editing by exploring the trajectories in the GAN's latent space. Recent diffusion-based portrait editing [7, 83] can support various editing types with texts as guidance. However, most of them are done in the 2D space, and thus there is no guarantee for the underlying 3D consistency. Therefore, 3D-aware portrait editing is crucial to achieve the goal. Some methods [15, 33, 39, 42, 66, 67, 74, 76, 81] rely on the latent space" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "text", + "content": "3DPE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": "of 3D GANs to perform editing through walking in the latent space, but are limited by the number of editable latent attributes. A few methods, such as Instruct-NeRF2NeRF [27], ClipFace [4], and LENeRF [32], attempt to leverage the large-scale models [54, 56] to edit on 3D representations with texts as guidance but require heavy iterative refinement of the edited results. Similarly, InstructPix2NeRF [40] leverages text-to-image models and text prompts to edit within the latent space of the GAN-based 3D generator. However, it is held back by the multi-step diffusion inference and the heavy structure of a GAN-based generator. Ours, in contrast, benefits from the lightweight design of the module and achieves real-time editing with ease. With such a design, our model can also adapt fast to the user-defined editing prompts. Moreover, texts cannot always illustrate the desired effects precisely, and sometimes an image prompt serves as better guidance for image editing [20, 57]. Our method can support not only text prompts but also image prompts for editing, providing a more flexible and user-friendly interface." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 314, + 203, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 314, + 203, + 327 + ], + "spans": [ + { + "bbox": [ + 132, + 314, + 203, + 327 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 340, + 483, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 483, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 483, + 484 + ], + "type": "text", + "content": "Our approach takes a portrait image " + }, + { + "bbox": [ + 130, + 340, + 483, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 130, + 340, + 483, + 484 + ], + "type": "text", + "content": " and its camera pose " + }, + { + "bbox": [ + 130, + 340, + 483, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 340, + 483, + 484 + ], + "type": "text", + "content": ", obtained via the face pose estimator [18]. Additionally, the referenced prompts " + }, + { + "bbox": [ + 130, + 340, + 483, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 130, + 340, + 483, + 484 + ], + "type": "text", + "content": " (images or texts) serve as the editing instructions. The outcome of our model is an edited version of the 3D portraits characterized by NeRF in accordance with the prompts. The edited image is denoted as " + }, + { + "bbox": [ + 130, + 340, + 483, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_p" + }, + { + "bbox": [ + 130, + 340, + 483, + 484 + ], + "type": "text", + "content": ". Within our framework, we achieve this objective by leveraging the knowledge of 3D GANs and a text-guided image editing model, which is distilled into a lightweight module. Following distillation, our method enables real-time 3D portrait editing and efficient adaptation to user-specified prompts. In Sec. 3.1, we offer an illustration of 3D GANs and diffusion models. The process of distilling these priors into a lightweight module is outlined in Sec. 3.2. Finally, the details for model training and inference, along with the fast adaptation for customized prompts, are presented in Sec. 3.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 501, + 222, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 222, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 222, + 514 + ], + "type": "text", + "content": "3.1 Preliminary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "3D GAN Prior for Portrait Reconstruction. The 3D-aware GANs showcase the ability to synthesize photorealistic 3D images using a collection of single-view images. Notably, EG3D [13] introduces an efficient triplane 3D representation, demonstrating high-quality 3D-aware image rendering. Once trained, the generator of EG3D can be applied for single-image 3D reconstruction via GAN inversion. However, existing 3D GAN inversion methods often encounter geometry distortion or exhibit slow inference speed. To tackle these challenges, we utilize Live3D [70], a state-of-the-art single-image portrait 3D reconstruction model built upon EG3D, preserving geometry quality while ensuring real-time performance. Live3D employs a two-branch encoder, " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_{high}(\\cdot)" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_{low}(\\cdot)" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", to extract different resolution features from the input. It then utilizes a ViT-based" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 101 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 101 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "text", + "content": "Q. Bai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 116, + 473, + 288 + ], + "blocks": [ + { + "bbox": [ + 145, + 116, + 473, + 288 + ], + "lines": [ + { + "bbox": [ + 145, + 116, + 473, + 288 + ], + "spans": [ + { + "bbox": [ + 145, + 116, + 473, + 288 + ], + "type": "image", + "image_path": "9074058ef0a16f926b765e7d2abeb7ca60bafbd451e32105f57698762302864c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 297, + 482, + 406 + ], + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 406 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 406 + ], + "type": "text", + "content": "Fig. 2: Overview of our method. We distill the prior in the diffusion model and 3D GAN for real-time 3D-aware editing. Our approach is fine-tuned from Live3D [70], where we extract features from the input portrait I using " + }, + { + "bbox": [ + 130, + 297, + 482, + 406 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_{high}(\\cdot)" + }, + { + "bbox": [ + 130, + 297, + 482, + 406 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 297, + 482, + 406 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_{low}(\\cdot)" + }, + { + "bbox": [ + 130, + 297, + 482, + 406 + ], + "type": "text", + "content": ". The prompt embedding is generated with " + }, + { + "bbox": [ + 130, + 297, + 482, + 406 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_p(\\cdot)" + }, + { + "bbox": [ + 130, + 297, + 482, + 406 + ], + "type": "text", + "content": " and injected with the input features from " + }, + { + "bbox": [ + 130, + 297, + 482, + 406 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_{high}(\\cdot)" + }, + { + "bbox": [ + 130, + 297, + 482, + 406 + ], + "type": "text", + "content": " through a cross-attention mechanism. Our model is trained to mimic the output from the diffusion prior to acquire editing knowledge and enforce geometry constraints through triplane, multi-view images, and depth supervision from the 3D prior. In this context, InstructPix2Pix [7] and Live3D serve as the diffusion and 3D prior, respectively. It's noteworthy that only " + }, + { + "bbox": [ + 130, + 297, + 482, + 406 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_p(\\cdot)" + }, + { + "bbox": [ + 130, + 297, + 482, + 406 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 297, + 482, + 406 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_t(\\cdot)" + }, + { + "bbox": [ + 130, + 297, + 482, + 406 + ], + "type": "text", + "content": " are learnable during training, while all other parameters remain frozen." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 434, + 480, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 480, + 447 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 480, + 447 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_t(\\cdot)" + }, + { + "bbox": [ + 132, + 434, + 480, + 447 + ], + "type": "text", + "content": " decoder [19] to transform the fused encoder features into the triplane " + }, + { + "bbox": [ + 132, + 434, + 480, + 447 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 132, + 434, + 480, + 447 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 247, + 471, + 481, + 484 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 471, + 481, + 484 + ], + "spans": [ + { + "bbox": [ + 247, + 471, + 481, + 484 + ], + "type": "interline_equation", + "content": "\\mathbf {T} = \\mathrm {E} _ {t} \\left(\\mathrm {E} _ {\\text {h i g h}} (\\mathbf {I}), \\mathrm {E} _ {\\text {l o w}} (\\mathbf {I})\\right). \\tag {1}", + "image_path": "c3f98b4fa01256ede8e2638f3a1ac318126bc097b2127b2309baf4da93ebd471.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 496, + 480, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 480, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 480, + 532 + ], + "type": "text", + "content": "This triplane " + }, + { + "bbox": [ + 130, + 496, + 480, + 532 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 130, + 496, + 480, + 532 + ], + "type": "text", + "content": " is subsequently used in conjunction with the volume rendering module and upsampler of EG3D to generate photorealistic view synthesis given the camera pose c. For clarity, we use " + }, + { + "bbox": [ + 130, + 496, + 480, + 532 + ], + "type": "inline_equation", + "content": "\\mathrm{R}(\\cdot)" + }, + { + "bbox": [ + 130, + 496, + 480, + 532 + ], + "type": "text", + "content": " to denote this process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Diffusion Prior for Portrait Editing. Only performing 3D reconstruction is insufficient as our goal is to edit portrait images. Large-scale diffusion models [56], trained on vast text-image pairs, can synthesize realistic photos with text input, offering powerful editing prior. Despite their effectiveness for face images, the editing process is often slow due to step-by-step optimization or multiple inferences in diffusion models. The scarcity of high-quality and multi-view consistent paired data significantly hinders the practical application of 3D-aware editing from a single image, which is hard to acquire with diffusion models. Therefore, our goal is to distill editing knowledge from the diffusion model, integrate it with the 3D prior in Live3D into a lightweight module, and employ it for real-time portrait editing." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "text", + "content": "3DPE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 118, + 462, + 311 + ], + "blocks": [ + { + "bbox": [ + 153, + 118, + 462, + 311 + ], + "lines": [ + { + "bbox": [ + 153, + 118, + 462, + 311 + ], + "spans": [ + { + "bbox": [ + 153, + 118, + 462, + 311 + ], + "type": "image", + "image_path": "f81ee83c8db8f852b73649c43459d539b6ad26c055042c6a1c778eb05c796b5f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 320, + 482, + 376 + ], + "lines": [ + { + "bbox": [ + 130, + 320, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 482, + 376 + ], + "type": "text", + "content": "Fig. 3: Disentanglement in Live3D features. We separately disable the features from " + }, + { + "bbox": [ + 130, + 320, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_{high}(\\cdot)" + }, + { + "bbox": [ + 130, + 320, + 482, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 320, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_{low}(\\cdot)" + }, + { + "bbox": [ + 130, + 320, + 482, + 376 + ], + "type": "text", + "content": " to infer the reconstructed image. Without " + }, + { + "bbox": [ + 130, + 320, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_{high}(\\cdot)" + }, + { + "bbox": [ + 130, + 320, + 482, + 376 + ], + "type": "text", + "content": ", the output retains the coarse structure but loses its appearance. Conversely, when " + }, + { + "bbox": [ + 130, + 320, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_{low}(\\cdot)" + }, + { + "bbox": [ + 130, + 320, + 482, + 376 + ], + "type": "text", + "content": " is deactivated, the reconstructed portraits preserve the texture (such as the blue and purple reflection on the glasses) but fail to capture the geometry." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 401, + 378, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 401, + 378, + 414 + ], + "spans": [ + { + "bbox": [ + 131, + 401, + 378, + 414 + ], + "type": "text", + "content": "3.2 Distilling Priors into a Lightweight Module" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 437, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 509 + ], + "type": "text", + "content": "Our method aims to perform real-time 3D-aware editing for various prompts for the single-view portrait. Thus, it needs powerful 3D knowledge for geometry reconstruction and editing prior to handling various control signals. We leverage the strengths of both 3D GANs and diffusion models. In the following, we provide detailed presentations on how the knowledge of these two types of models is distilled into a lightweight module." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": "Feature Representation in Live3D. We carefully study the Live3D model and discover that, as a two-branch triplane-based 3D reconstruction model, the low-resolution and high-resolution features from the Live3D encoder " + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_{low}(\\cdot)" + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_{high}(\\cdot)" + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": " tend to learn various levels of information without explicit guidance. We conduct a study by disabling one of the branches and inferring the reconstructed images. As illustrated in Fig. 3, when the high-resolution encoder " + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_{high}(\\cdot)" + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": " is disabled with random weights, the inference image retains a similar structure but loses its detailed appearance. Conversely, when the low-resolution encoder " + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_{low}(\\cdot)" + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": " is disabled, the reconstructed portraits preserve some of the texture from the input but struggle to capture the geometry. Based on this analysis, the features of the two branches separately model low-frequency and high-frequency information. This insight motivates our model design in terms of how to distill knowledge from the diffusion model and 3D GANs." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Q. Bai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": "Geometry Prediction for Inputs I. In our framework, we use the prompts " + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": " to refine the input portrait " + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": " into the edited image " + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_p" + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_p" + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": " inherits a similar structure to the input " + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": ", providing coarse geometry or structural guidance in the editing process. As discussed in the above section, the encoder " + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_{low}(\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": " consistently generates low-frequency features to represent geometry cues, making it well-suited for preserving structured information in the input portrait. As a result, we freeze " + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_{low}(\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": " and leverage it to produce structure features " + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_g = \\mathrm{E}_{low}(\\mathbf{I})" + }, + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": " for coarse geometry prediction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 480, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 480, + 318 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 480, + 318 + ], + "type": "text", + "content": "Injecting Prompts P as Condition. In contrast to input images, prompts typically offer more high-frequency and texture information to guide and control the editing process. Accordingly, it is appropriate to incorporate the prompts into the high-level branch " + }, + { + "bbox": [ + 130, + 212, + 480, + 318 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_{high}(\\cdot)" + }, + { + "bbox": [ + 130, + 212, + 480, + 318 + ], + "type": "text", + "content": " in Live3D to extract the high-frequency features. To incorporate prompts, we utilize the prompt encoder " + }, + { + "bbox": [ + 130, + 212, + 480, + 318 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_p(\\cdot)" + }, + { + "bbox": [ + 130, + 212, + 480, + 318 + ], + "type": "text", + "content": " to create prompt embeddings that are fused into our model through cross attention, which is similar to the strategy in Stable Diffusion [56]. Specifically, we add a cross-attention layer after " + }, + { + "bbox": [ + 130, + 212, + 480, + 318 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_{high}(\\cdot)" + }, + { + "bbox": [ + 130, + 212, + 480, + 318 + ], + "type": "text", + "content": " to obtain the feature updated with the prompt embeddings:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 223, + 327, + 480, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 327, + 480, + 338 + ], + "spans": [ + { + "bbox": [ + 223, + 327, + 480, + 338 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {a} = \\operatorname {C r o s s A t t e n t i o n} \\left(\\operatorname {E} _ {\\text {h i g h}} (\\mathbf {I}), \\operatorname {E} _ {p} (\\mathbf {P})\\right), \\tag {2}", + "image_path": "14f6eba4f01e72b81a953158589d712f2c9c2ceb2ec6b48d50747265f0c9536b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 343, + 479, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 479, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 479, + 392 + ], + "type": "text", + "content": "where the encoder " + }, + { + "bbox": [ + 130, + 343, + 479, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_p(\\cdot)" + }, + { + "bbox": [ + 130, + 343, + 479, + 392 + ], + "type": "text", + "content": " is a transformer (MAE [29] for image and CLIP [54] for text). The " + }, + { + "bbox": [ + 130, + 343, + 479, + 392 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_{high}(\\mathbf{I})" + }, + { + "bbox": [ + 130, + 343, + 479, + 392 + ], + "type": "text", + "content": " serves as the query and " + }, + { + "bbox": [ + 130, + 343, + 479, + 392 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_p(\\mathbf{P})" + }, + { + "bbox": [ + 130, + 343, + 479, + 392 + ], + "type": "text", + "content": " is employed as key and value. The updated feature " + }, + { + "bbox": [ + 130, + 343, + 479, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_a" + }, + { + "bbox": [ + 130, + 343, + 479, + 392 + ], + "type": "text", + "content": " and the geometry feature " + }, + { + "bbox": [ + 130, + 343, + 479, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_g" + }, + { + "bbox": [ + 130, + 343, + 479, + 392 + ], + "type": "text", + "content": " are then feed into the decoder " + }, + { + "bbox": [ + 130, + 343, + 479, + 392 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_t(\\cdot)" + }, + { + "bbox": [ + 130, + 343, + 479, + 392 + ], + "type": "text", + "content": " to infer the triplane " + }, + { + "bbox": [ + 130, + 343, + 479, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{T_p}" + }, + { + "bbox": [ + 130, + 343, + 479, + 392 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 269, + 399, + 480, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 399, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 269, + 399, + 480, + 411 + ], + "type": "interline_equation", + "content": "\\mathbf {T} _ {\\mathbf {p}} = \\operatorname {E} _ {t} \\left(\\mathbf {F} _ {a}, \\mathbf {F} _ {g}\\right). \\tag {3}", + "image_path": "c8dd569b9de1575ed4cc2c983d401da2134e9a66c210003b2d1b53c32bfa6eb2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 415, + 479, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 479, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 479, + 475 + ], + "type": "text", + "content": "Distilling Diffusion and 3D GAN Prior. Unlike previous methods that require large amounts of 3D data, our model operates with only 2D paired data generated by the 2D editing model [7]. For both visual and textual prompts, we generate edited images " + }, + { + "bbox": [ + 130, + 415, + 479, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{gt}" + }, + { + "bbox": [ + 130, + 415, + 479, + 475 + ], + "type": "text", + "content": " with the diffusion-based image editing model [7] as the pseudo labels. The details of data generation can be found in Sec. 4.1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 475, + 479, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 479, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 479, + 499 + ], + "type": "text", + "content": "With the predicted triplane " + }, + { + "bbox": [ + 130, + 475, + 479, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathbf{p}}" + }, + { + "bbox": [ + 130, + 475, + 479, + 499 + ], + "type": "text", + "content": ", we can render the image using the pretrained EG3D model and then calculate the reconstruction loss as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 256, + 506, + 480, + 518 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 506, + 480, + 518 + ], + "spans": [ + { + "bbox": [ + 256, + 506, + 480, + 518 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {2 d} = \\ell_ {\\mathbf {I}} \\left(\\mathbf {I} _ {g t}, \\mathrm {R} \\left(\\mathbf {T} _ {\\mathbf {p}}, \\mathbf {c}\\right)\\right), \\tag {4}", + "image_path": "3dc4799cce0ea45588ce76e484084e8455ac19a42ead97d305aeb60b57ee50ff.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 524, + 479, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 524, + 479, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 524, + 479, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 524, + 479, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{R}(\\cdot)" + }, + { + "bbox": [ + 130, + 524, + 479, + 581 + ], + "type": "text", + "content": " is the rendering module of EG3D, " + }, + { + "bbox": [ + 130, + 524, + 479, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 524, + 479, + 581 + ], + "type": "text", + "content": " is the camera pose of input portrait, and " + }, + { + "bbox": [ + 130, + 524, + 479, + 581 + ], + "type": "inline_equation", + "content": "\\ell_1(\\cdot)" + }, + { + "bbox": [ + 130, + 524, + 479, + 581 + ], + "type": "text", + "content": " is an image reconstruction loss penalizing the difference between the ground truth " + }, + { + "bbox": [ + 130, + 524, + 479, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{gt}" + }, + { + "bbox": [ + 130, + 524, + 479, + 581 + ], + "type": "text", + "content": " and the rendering " + }, + { + "bbox": [ + 130, + 524, + 479, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_p = \\mathrm{R}(\\mathbf{T_p},\\mathbf{c})" + }, + { + "bbox": [ + 130, + 524, + 479, + 581 + ], + "type": "text", + "content": ". It's important to note that " + }, + { + "bbox": [ + 130, + 524, + 479, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{2d}" + }, + { + "bbox": [ + 130, + 524, + 479, + 581 + ], + "type": "text", + "content": ", used to reconstruct the " + }, + { + "bbox": [ + 130, + 524, + 479, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{gt}" + }, + { + "bbox": [ + 130, + 524, + 479, + 581 + ], + "type": "text", + "content": ", essentially distills the knowledge of editing from the diffusion model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 583, + 479, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 583, + 479, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 583, + 479, + 643 + ], + "type": "text", + "content": "We observe that " + }, + { + "bbox": [ + 130, + 583, + 479, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{2d}" + }, + { + "bbox": [ + 130, + 583, + 479, + 643 + ], + "type": "text", + "content": " helps the model to reconstruct well on the input camera view but suffers from geometry distortion during novel-view synthesis. To fully exploit the 3D properties of Live3D, we propose to distill the 3D knowledge of Live3D. Specifically, we leverage the pretrained Live3D to infer the triplane " + }, + { + "bbox": [ + 130, + 583, + 479, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{gt}" + }, + { + "bbox": [ + 130, + 583, + 479, + 643 + ], + "type": "text", + "content": ", and multi-view depths " + }, + { + "bbox": [ + 130, + 583, + 479, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{gt}" + }, + { + "bbox": [ + 130, + 583, + 479, + 643 + ], + "type": "text", + "content": " and images " + }, + { + "bbox": [ + 130, + 583, + 479, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{gt}" + }, + { + "bbox": [ + 130, + 583, + 479, + 643 + ], + "type": "text", + "content": " of the pseudo-label image " + }, + { + "bbox": [ + 130, + 583, + 479, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{gt}" + }, + { + "bbox": [ + 130, + 583, + 479, + 643 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 262, + 651, + 480, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 651, + 480, + 662 + ], + "spans": [ + { + "bbox": [ + 262, + 651, + 480, + 662 + ], + "type": "interline_equation", + "content": "\\mathbf {T} _ {g t}, \\mathcal {D} _ {g t} = \\mathrm {G} (\\mathbf {I} _ {g t}, \\mathcal {C}), \\tag {5}", + "image_path": "1e69ca00692d86d6e61e342c65c336800c584ae9b75d96ab9b47e42986e42fd9.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "text", + "content": "3DPE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\mathrm{G}(\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " represents the inference process of Live3D, where " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{\\mathbf{c}_1,\\dots,\\mathbf{c}_n\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " is the camera set, and " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " denotes the number of cameras. We also render the multi-view images " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_p" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " and depths " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_p" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " of the edited image " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_p" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " from the triplane " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_p" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " and then define the objective:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 205, + 174, + 481, + 188 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 174, + 481, + 188 + ], + "spans": [ + { + "bbox": [ + 205, + 174, + 481, + 188 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {3 d} = \\ell_ {\\mathcal {I}} \\left(\\mathcal {I} _ {p}, \\mathcal {I} _ {g t}\\right) + \\ell_ {\\mathbf {T}} \\left(\\mathbf {T} _ {p}, \\mathbf {T} _ {g t}\\right) + \\ell_ {\\mathcal {D}} \\left(\\mathcal {D} _ {p}, \\mathcal {D} _ {g t}\\right), \\tag {6}", + "image_path": "f7cdea59c1d15bc6a321f85d26aac4a032aa443518e39e29961ce93a9b6b5366.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 195, + 481, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 195, + 481, + 220 + ], + "spans": [ + { + "bbox": [ + 130, + 195, + 481, + 220 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 195, + 481, + 220 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathcal{I}}(\\cdot), \\ell_{\\mathbf{T}}(\\cdot)" + }, + { + "bbox": [ + 130, + 195, + 481, + 220 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 195, + 481, + 220 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathcal{D}}(\\cdot)" + }, + { + "bbox": [ + 130, + 195, + 481, + 220 + ], + "type": "text", + "content": " is the reconstruction loss penalizing the difference in image, triplane, and depth between " + }, + { + "bbox": [ + 130, + 195, + 481, + 220 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_p" + }, + { + "bbox": [ + 130, + 195, + 481, + 220 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 195, + 481, + 220 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_{gt}" + }, + { + "bbox": [ + 130, + 195, + 481, + 220 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 238, + 276, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 276, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 276, + 251 + ], + "type": "text", + "content": "3.3 Training and Inference" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 270, + 482, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 330 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 330 + ], + "type": "text", + "content": "Training. During the training phase, we sample a triplet consisting of the input portrait " + }, + { + "bbox": [ + 130, + 270, + 482, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 130, + 270, + 482, + 330 + ], + "type": "text", + "content": " along with its camera pose " + }, + { + "bbox": [ + 130, + 270, + 482, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 270, + 482, + 330 + ], + "type": "text", + "content": ", prompts " + }, + { + "bbox": [ + 130, + 270, + 482, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 130, + 270, + 482, + 330 + ], + "type": "text", + "content": ", and the pseudo-label image " + }, + { + "bbox": [ + 130, + 270, + 482, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{gt}" + }, + { + "bbox": [ + 130, + 270, + 482, + 330 + ], + "type": "text", + "content": " generated by the 2D diffusion model. We leverage the Live3D model as a pretrained model and add an additional prompts encoder " + }, + { + "bbox": [ + 130, + 270, + 482, + 330 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_p" + }, + { + "bbox": [ + 130, + 270, + 482, + 330 + ], + "type": "text", + "content": " to extract prompt embeddings. The overall learning objective can be described as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 261, + 341, + 481, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 341, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 261, + 341, + 481, + 354 + ], + "type": "interline_equation", + "content": "L = \\lambda_ {1} L _ {2 D} + \\lambda_ {2} L _ {3 D}, \\tag {7}", + "image_path": "d8a19c99862ff62ce79e2d13c143bf3adbd0434d6b3b570f34db280fbca41076.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "text", + "content": " are loss weights. In our setting, " + }, + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "text", + "content": " are both set to 1.0. For the reconstruction loss, " + }, + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathcal{I}}(\\cdot)" + }, + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{I}}(\\cdot)" + }, + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "text", + "content": " are combinations of " + }, + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_2" + }, + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "text", + "content": " loss and LPIPS loss [84], with loss weights being 1 and 2, respectively. " + }, + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathcal{D}}(\\cdot)" + }, + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{T}}(\\cdot)" + }, + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_1" + }, + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "text", + "content": " loss. Notably, based on the study on Live3D feature representation, during training only " + }, + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_p(\\cdot)" + }, + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_t(\\cdot)" + }, + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "text", + "content": " are set to be learnable, while other modules are frozen. It allows our model to leverage Live3D knowledge as much as possible and converge at a very fast speed." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 445, + 482, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 481 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 481 + ], + "type": "text", + "content": "Inference. For inference, users can provide a single 2D portrait image and choose an image prompt or text prompt. Our model is able to generate the edited 3D NeRF along with photorealistic view synthesis." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 481, + 482, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 482, + 553 + ], + "type": "text", + "content": "Customized Prompts Adaptation. To accommodate customized styles provided by users, we propose a method to adapt our pretrained encoder to user-defined styles. We increase the tuning efficiency by optimizing only " + }, + { + "bbox": [ + 130, + 481, + 482, + 553 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_p(\\cdot)" + }, + { + "bbox": [ + 130, + 481, + 482, + 553 + ], + "type": "text", + "content": " and the normalization layers in " + }, + { + "bbox": [ + 130, + 481, + 482, + 553 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_t(\\cdot)" + }, + { + "bbox": [ + 130, + 481, + 482, + 553 + ], + "type": "text", + "content": " with the same learning objective Eq. 7. This method allows us to limit the training data to only 10 image pairs and the learning time to 5 minutes on a single GPU." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 572, + 230, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 572, + 230, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 572, + 230, + 586 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 597, + 262, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 262, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 262, + 610 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": "Training Settings. For real faces, we adopt the FFHQ dataset [35] at " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": " resolution with camera parameters aligned by EG3D [13]. In order to obtain the stylized images as pseudo labels, and the corresponding textual and visual prompts, we leverage the text-guided editing model InstructPix2Pix (IP2P) [7] to" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 218, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 218, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 218, + 102 + ], + "type": "text", + "content": "Q. Bai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 156, + 115, + 218, + 158 + ], + "blocks": [ + { + "bbox": [ + 133, + 127, + 141, + 141 + ], + "lines": [ + { + "bbox": [ + 133, + 127, + 141, + 141 + ], + "spans": [ + { + "bbox": [ + 133, + 127, + 141, + 141 + ], + "type": "text", + "content": "mnd" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 156, + 115, + 218, + 158 + ], + "lines": [ + { + "bbox": [ + 156, + 115, + 218, + 158 + ], + "spans": [ + { + "bbox": [ + 156, + 115, + 218, + 158 + ], + "type": "image", + "image_path": "fac216f03f4f959144cd3e01c13277758aa34d7a06948f1663c2948f73e71b77.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 240, + 115, + 301, + 158 + ], + "blocks": [ + { + "bbox": [ + 240, + 115, + 301, + 158 + ], + "lines": [ + { + "bbox": [ + 240, + 115, + 301, + 158 + ], + "spans": [ + { + "bbox": [ + 240, + 115, + 301, + 158 + ], + "type": "image", + "image_path": "b43743ade296fac25742324ff4fa9a9df10d49b1af67b1428a50f37c9aee2567.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 324, + 115, + 385, + 157 + ], + "blocks": [ + { + "bbox": [ + 324, + 115, + 385, + 157 + ], + "lines": [ + { + "bbox": [ + 324, + 115, + 385, + 157 + ], + "spans": [ + { + "bbox": [ + 324, + 115, + 385, + 157 + ], + "type": "image", + "image_path": "dd1544dd3fe5b6fe651dd1d5be29ecea4c6ee7378a529a456927ab81272ee084.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 406, + 115, + 448, + 158 + ], + "blocks": [ + { + "bbox": [ + 406, + 115, + 448, + 158 + ], + "lines": [ + { + "bbox": [ + 406, + 115, + 448, + 158 + ], + "spans": [ + { + "bbox": [ + 406, + 115, + 448, + 158 + ], + "type": "image", + "image_path": "061d13bb5dff2ada5f8b148d1dc34ad11fc1e99106b14272343871933008d364.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 448, + 140, + 474, + 156 + ], + "lines": [ + { + "bbox": [ + 448, + 140, + 474, + 156 + ], + "spans": [ + { + "bbox": [ + 448, + 140, + 474, + 156 + ], + "type": "text", + "content": "Recolor it to a Gothic color scheme." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 148, + 159, + 227, + 327 + ], + "blocks": [ + { + "bbox": [ + 132, + 160, + 145, + 324 + ], + "lines": [ + { + "bbox": [ + 132, + 160, + 145, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 145, + 324 + ], + "type": "text", + "content": "P 12345678901001100110011001100110011001100110011001100110011001100110011001100110011001100110011001100110011001100" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 148, + 159, + 227, + 327 + ], + "lines": [ + { + "bbox": [ + 148, + 159, + 227, + 327 + ], + "spans": [ + { + "bbox": [ + 148, + 159, + 227, + 327 + ], + "type": "image", + "image_path": "90d5bc38436b57d70008b667788d2b87ebaab64700b2695fa04eafcb62090631.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 229, + 159, + 310, + 327 + ], + "blocks": [ + { + "bbox": [ + 229, + 159, + 310, + 327 + ], + "lines": [ + { + "bbox": [ + 229, + 159, + 310, + 327 + ], + "spans": [ + { + "bbox": [ + 229, + 159, + 310, + 327 + ], + "type": "image", + "image_path": "30e99cad967c55e84925a0c9e13c7325270c8f7988317ca53a515906bbe71ed2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 314, + 159, + 395, + 327 + ], + "blocks": [ + { + "bbox": [ + 314, + 159, + 395, + 327 + ], + "lines": [ + { + "bbox": [ + 314, + 159, + 395, + 327 + ], + "spans": [ + { + "bbox": [ + 314, + 159, + 395, + 327 + ], + "type": "image", + "image_path": "705204cfb497a2a8f845cc8ab25e4e1106f2351b510e8fa0362e906930690810.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 398, + 159, + 479, + 327 + ], + "blocks": [ + { + "bbox": [ + 398, + 159, + 479, + 327 + ], + "lines": [ + { + "bbox": [ + 398, + 159, + 479, + 327 + ], + "spans": [ + { + "bbox": [ + 398, + 159, + 479, + 327 + ], + "type": "image", + "image_path": "9768e369ab754eb13086513b97f1af71fd66067a6fb161bf59c1dc6581884197.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 147, + 328, + 227, + 369 + ], + "blocks": [ + { + "bbox": [ + 134, + 342, + 146, + 357 + ], + "lines": [ + { + "bbox": [ + 134, + 342, + 146, + 357 + ], + "spans": [ + { + "bbox": [ + 134, + 342, + 146, + 357 + ], + "type": "text", + "content": "0000000000000000000000000000000000000" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 147, + 328, + 227, + 369 + ], + "lines": [ + { + "bbox": [ + 147, + 328, + 227, + 369 + ], + "spans": [ + { + "bbox": [ + 147, + 328, + 227, + 369 + ], + "type": "image", + "image_path": "26263c704c8728002033135f47f0b7e7721fa49d6d2cca01d12fd9c457879a32.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 229, + 328, + 310, + 369 + ], + "blocks": [ + { + "bbox": [ + 229, + 328, + 310, + 369 + ], + "lines": [ + { + "bbox": [ + 229, + 328, + 310, + 369 + ], + "spans": [ + { + "bbox": [ + 229, + 328, + 310, + 369 + ], + "type": "image", + "image_path": "fc936265282935b0a001880613bddc357adb22e9187d45e8eb7a1b616e360b41.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 380, + 482, + 426 + ], + "lines": [ + { + "bbox": [ + 131, + 380, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 131, + 380, + 482, + 426 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparisons. We compare the results of several baselines with image prompts and text prompts. In each case, we include the edited portraits as well as their novel view renderings. Our method generates high-quality edited portraits with better 3D quality and alignment with the referenced prompts." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 314, + 328, + 395, + 369 + ], + "blocks": [ + { + "bbox": [ + 314, + 328, + 395, + 369 + ], + "lines": [ + { + "bbox": [ + 314, + 328, + 395, + 369 + ], + "spans": [ + { + "bbox": [ + 314, + 328, + 395, + 369 + ], + "type": "image", + "image_path": "4e1bcb7134c60dbc4d990b8785227a03ac8cc1fbed1a333ea034658db9587a0e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 398, + 328, + 479, + 369 + ], + "blocks": [ + { + "bbox": [ + 398, + 328, + 479, + 369 + ], + "lines": [ + { + "bbox": [ + 398, + 328, + 479, + 369 + ], + "spans": [ + { + "bbox": [ + 398, + 328, + 479, + 369 + ], + "type": "image", + "image_path": "cf3df868e514f235686b90c0d843bcec5a1b6f8ac7da4e878adc7f336dd682ba.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "text", + "content": "edit the real faces. Specifically, we manually generate and select textual prompts, and send them to IP2P together with the real images to obtain the edited 2D images. The text prompts of our training and testing set are obtained by augmenting the aforementioned textual prompts with the large language model GPT [8] following IP2P [7]. While the image prompts of our training and testing set are picked from the edited 2D images. We conduct experiments on 20 styles, and for each style, we synthesize 1000 images, resulting in 20,000 image pairs with text prompts for model training. For each style, we use 8 textual prompts for training and 5 textual prompts for testing. We adopt a learning rate of 5e-5 and optimize the model for 60k iterations with a batch size of 32. The entire training procedure is completed over a period of 40 hours using 8 NVIDIA A100 GPUs. Baselines. We conduct comparisons against two categories of methods that achieve analogous outcomes: 1) 3D reconstruction coupled with 3D editing; 2) 2D editing followed by 3D reconstruction. For the first category, we implement three baseline approaches, wherein we firstly leverage PTI [55] or Live3D [70] for reconstruction and edit both of them following StyleCLIP [50]. Specifically, for experiments in PTI + StyleCLIP, we first conduct pivotal tuning [55] based on the predicted latent codes predicted by e4e [68] to reconstruct the given image in" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "text", + "content": "3DPE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 167, + 179, + 446, + 243 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 170 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 170 + ], + "type": "text", + "content": "Table 1: Quantitative comparisons. We compare several baselines on the 100 images of FFHQ dataset. It's important to note that we exclude " + }, + { + "bbox": [ + 132, + 114, + 481, + 170 + ], + "type": "inline_equation", + "content": "CLIP_r" + }, + { + "bbox": [ + 132, + 114, + 481, + 170 + ], + "type": "text", + "content": " for PTI+StyleCLIP and Live3D+StyleCLIP since these models utilize CLIP for optimization. Our model excels in 3D quality and achieves a remarkable 250x improvement (compared with IP2P+Live3D) in inference speed, achieving real-time performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 179, + 446, + 243 + ], + "lines": [ + { + "bbox": [ + 167, + 179, + 446, + 243 + ], + "spans": [ + { + "bbox": [ + 167, + 179, + 446, + 243 + ], + "type": "table", + "html": "
MethodIDt↑CLIPr↑3D↑Time↓
PTI+StyleCLIP0.11-0.73360s
Live3D+StyleCLIP0.63-0.7230s
IP2P+Live3D0.440.620.7510s
Ours0.470.730.760.04s
", + "image_path": "3964f252442b6d3fee030b588de29b0765bf4ed32401a3d84728302861a136ee.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 300, + 481, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 479 + ], + "type": "text", + "content": "a 3D-aware manner. Then we perform editing on the latent codes following [50], where CLIP loss [54] is utilized between the model output and the prompt (no matter whether it is an image or text). The setting of Live3D + StyleCLIP is similar to PTI + StyleCLIP except that we obtain the triplane by Live3D and perform editing on this 3D representation. SPI [81] + StyleGAN-NADA [21] (StyleNADA for the abbreviation in Fig. 4) serves as the third baseline in the first category, where SPI is used for 3D reconstruction and StyleGAN-NADA is utilized to optimize the 3D generator for editing. In the second category, we employ a diffusion-based text-guided 2D image editing method IP2P [7] prior to executing 3D reconstruction via Live3D. IP2P being a text-instructed editing model, we need to first extract text prompts from the reference image with the image caption algorithm BLIP [41] when image prompts are input. The 2D edited results could then be obtained by sending the input image and the text prompt to IP2P. The final 3D results are acquired by inverting the 2D editing results with Live3D." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 498, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 665 + ], + "type": "text", + "content": "Evaluation Criteria. We evaluate all metrics on 100 pairs of images processed from FFHQ. We conduct a comprehensive evaluation of the editing performance in the following four aspects: identity preservation, reference alignment, 3D consistency, and inference speed. 1) Identity preservation " + }, + { + "bbox": [ + 132, + 498, + 481, + 665 + ], + "type": "inline_equation", + "content": "(\\mathbf{ID}_{\\mathbf{t}})" + }, + { + "bbox": [ + 132, + 498, + 481, + 665 + ], + "type": "text", + "content": " aims to measure the preservation of the original identity by calculating the cosine similarity between the identity feature of the input image " + }, + { + "bbox": [ + 132, + 498, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 132, + 498, + 481, + 665 + ], + "type": "text", + "content": " and that of the edited image " + }, + { + "bbox": [ + 132, + 498, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_p" + }, + { + "bbox": [ + 132, + 498, + 481, + 665 + ], + "type": "text", + "content": ". We use the ArcFace model [16] to extract identity features. 2) Reference alignment " + }, + { + "bbox": [ + 132, + 498, + 481, + 665 + ], + "type": "inline_equation", + "content": "(\\mathrm{CLIP_r})" + }, + { + "bbox": [ + 132, + 498, + 481, + 665 + ], + "type": "text", + "content": " targets at assessing the alignment of the output editing styles to the desired input prompt by computing the cosine similarity in the CLIP [50] feature space. 3) 3D consistency (3D) on the edited outputs is measured following the evaluation protocols established by EG3D [13]. This involves calculating the identity similarity across multiple views. 4) Inference speed (Time) is measured on a single NVIDIA A6000 GPU with an average of 100 samples." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Bai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 479, + 213 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 479, + 213 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 479, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 479, + 213 + ], + "type": "image", + "image_path": "f6c3f00a93f75379fee3f937e97790f28c47b3c39e0f7d18dc7e9c660cfd69bb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 222, + 480, + 257 + ], + "lines": [ + { + "bbox": [ + 130, + 222, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 222, + 480, + 257 + ], + "type": "text", + "content": "Fig. 5: (a) Generated results of customized prompt adaptation and (b) its learning process. We show the intermediate testing results at 10s, 1min, 2min and 5min during adaptation for the novel style golden statue." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 269, + 479, + 338 + ], + "blocks": [ + { + "bbox": [ + 133, + 269, + 479, + 338 + ], + "lines": [ + { + "bbox": [ + 133, + 269, + 479, + 338 + ], + "spans": [ + { + "bbox": [ + 133, + 269, + 479, + 338 + ], + "type": "image", + "image_path": "840bba9faad9d542b7aabe52638d1a08a584d4deda24a2b2360447a1dbdcf07e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 348, + 480, + 371 + ], + "lines": [ + { + "bbox": [ + 130, + 348, + 480, + 371 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 480, + 371 + ], + "type": "text", + "content": "Fig.6: Qualitative comparison for ablations on (a) the distillation loss of 3D GAN " + }, + { + "bbox": [ + 130, + 348, + 480, + 371 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{3D})" + }, + { + "bbox": [ + 130, + 348, + 480, + 371 + ], + "type": "text", + "content": " and (b) diffusion models " + }, + { + "bbox": [ + 130, + 348, + 480, + 371 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{2D})" + }, + { + "bbox": [ + 130, + 348, + 480, + 371 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 397, + 338, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 338, + 410 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 338, + 410 + ], + "type": "text", + "content": "4.2 Efficient 3D-aware Portrait Editing" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 425, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 532 + ], + "type": "text", + "content": "We make an in-depth analysis of our efficient portrait editing method both quantitatively and qualitatively. The results are included in Tab. 1 and Fig. 4. For reference alignment, we do not report " + }, + { + "bbox": [ + 130, + 425, + 482, + 532 + ], + "type": "inline_equation", + "content": "CLIP_{r}" + }, + { + "bbox": [ + 130, + 425, + 482, + 532 + ], + "type": "text", + "content": " for methods that leverage CLIP for optimization since it is evaluated with CLIP as well. The standout feature of our method is its efficiency. Our method achieves an inference speed of merely 40ms, which improves over 100 times compared to the fastest existing baselines, which require around 10 seconds. Because of the efficient knowledge distillation, our approach is also good in preserving the identity, adhering more closely to the reference prompts and achieving the best 3D consistency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "As in Fig. 4, compared with PTI + CLIP methods, our model enables precise alignment with inputs and prompts, while the baseline always generates low-fidelity textures, and the resulting geometry has many artifacts. Although Live3D + CLIP achieves the best " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "ID_{t}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " score, the edited results are basically unchanged compared to the input. We suspect that Live3D does not possess a latent space and loses the editing priors, making it challenging for the results of CLIP optimization to align well with prompts. SPI + StyleGAN-NADA performs relatively slowly (18 minutes per instance) because both the 3D reconstruction and the editing are achieved by optimizing the generator. Although it can produce realistic textures, it fails to reconstruct fine details and to edit following the prompts. In contrast to the 2D editing and subsequent 3D reconstruction" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "text", + "content": "3DPE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 178, + 146, + 435, + 191 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 2: Ablation study on the number of styles used for pretraining. We report LPIPS score after 0.1min, 1min, 2min, and 5min fine-tuning for evaluation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 178, + 146, + 435, + 191 + ], + "lines": [ + { + "bbox": [ + 178, + 146, + 435, + 191 + ], + "spans": [ + { + "bbox": [ + 178, + 146, + 435, + 191 + ], + "type": "table", + "html": "
#Style/Time0.1min1min2min5min
30.77580.63700.53120.5220
100.59320.57340.51060.5012
200.57030.52900.50220.4849
", + "image_path": "2b428d3392c8dc930ecfe504fc5d33621eb1194707c783952335a33ef29c0332.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 179, + 245, + 434, + 307 + ], + "blocks": [ + { + "bbox": [ + 132, + 201, + 481, + 235 + ], + "lines": [ + { + "bbox": [ + 132, + 201, + 481, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 201, + 481, + 235 + ], + "type": "text", + "content": "Table 3: Ablation study on the number of data pairs used for novel prompt adaptation. We report LPIPS score after 0.1min, 1min, 2min, and 5min fine-tuning for evaluation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 179, + 245, + 434, + 307 + ], + "lines": [ + { + "bbox": [ + 179, + 245, + 434, + 307 + ], + "spans": [ + { + "bbox": [ + 179, + 245, + 434, + 307 + ], + "type": "table", + "html": "
#Pair/Time0.1min1min2min5min
20.61910.54310.52130.5212
50.61480.53240.50640.4971
100.60760.52420.49830.4935
200.59810.55250.49540.4895
500.58910.57180.49220.4797
", + "image_path": "95e797f49ac6c1caea0803ad325b1a3e98abe0155c5168fc3137fa0ae42c6b82.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 328, + 481, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 328, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 328, + 481, + 353 + ], + "type": "text", + "content": "pipeline (IP2P + Live3D), our method produces textures more consistent with the prompt and aligns better with the input structure." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 369, + 339, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 339, + 382 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 339, + 382 + ], + "type": "text", + "content": "4.3 Adaptation for Customized Editing" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 388, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 388, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 132, + 388, + 481, + 616 + ], + "type": "text", + "content": "With the trained editing network, our method already supports a variety of 3D-aware editing styles. Following DreamBooth [57] and to expand the selection and better conform to users' preferences, we offer an efficient method for fast adaptation to the user-defined customized prompts. Users can personalize the editing network by providing a modest set of 10 reference editing pairs. These pairs can be either handpicked from artist-created examples or produced using text-guided image editing models. As described in Sec. 3.3, only the partial of the encoder would be optimized. The adaptation process itself is remarkably fast, requiring only about 5 minutes to accomplish the learning of the customized prompts. Please also note that, once the user-defined style has been trained, all the incoming testing results can be obtained in a real-time manner. In Fig. 5, we present the testing results and learning process of the adaptation. Our model can quickly master the customized knowledge in about 2 minutes. With further training (e.g., 5 minutes), the edited results become more stylized, demonstrating a trade-off between authenticity and stylization. Upon completion of the adaptation, the method allows users to edit testing inputs in these newly learned styles with a minimal inference time of 0.04s. This rapid performance indicates that the method is well-optimized for real-time applications, providing a seamless and efficient user experience." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 634, + 246, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 634, + 246, + 645 + ], + "spans": [ + { + "bbox": [ + 132, + 634, + 246, + 645 + ], + "type": "text", + "content": "4.4 Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 653, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 653, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 653, + 481, + 666 + ], + "type": "text", + "content": "We analyze our model and validate our design choice by ablating the components." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Bai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 116, + 470, + 305 + ], + "blocks": [ + { + "bbox": [ + 141, + 116, + 470, + 305 + ], + "lines": [ + { + "bbox": [ + 141, + 116, + 470, + 305 + ], + "spans": [ + { + "bbox": [ + 141, + 116, + 470, + 305 + ], + "type": "image", + "image_path": "23913448181367f3e27cd2f751b0607ddf42b7e0911b6618b2faea15af9fe67c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 316, + 482, + 351 + ], + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 351 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 351 + ], + "type": "text", + "content": "Fig. 7: Video editing results with 3DPE. We use text prompts \"Make him a bronze statue\" to edit the input video. Our method can accurately reconstruct and preserve challenging facial expressions and achieve high-quality novel view renderings." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 382, + 482, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 382, + 482, + 454 + ], + "spans": [ + { + "bbox": [ + 130, + 382, + 482, + 454 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{3D}" + }, + { + "bbox": [ + 130, + 382, + 482, + 454 + ], + "type": "text", + "content": " for Distillation of Live3D. We compare our model trained with and without " + }, + { + "bbox": [ + 130, + 382, + 482, + 454 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{3D}" + }, + { + "bbox": [ + 130, + 382, + 482, + 454 + ], + "type": "text", + "content": ". As shown in Fig. 6, we observe that the model without " + }, + { + "bbox": [ + 130, + 382, + 482, + 454 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{3D}" + }, + { + "bbox": [ + 130, + 382, + 482, + 454 + ], + "type": "text", + "content": " results in very flat geometry, and the synthesis of the side view loses the normal structure with many artifacts. This highlights the importance of the 3D knowledge from Live3D in training our model with 2D paired data, enabling the inference of reasonable face geometry without relying on any 3D data sources." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 457, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{2D}" + }, + { + "bbox": [ + 130, + 457, + 482, + 529 + ], + "type": "text", + "content": " for Distillation of Diffusion Models. We also study the effect of " + }, + { + "bbox": [ + 130, + 457, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{2D}" + }, + { + "bbox": [ + 130, + 457, + 482, + 529 + ], + "type": "text", + "content": ", which represents using 2D editing priors from the diffusion model [7] to optimize the encoder. From Fig. 6, we observe that " + }, + { + "bbox": [ + 130, + 457, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{2D}" + }, + { + "bbox": [ + 130, + 457, + 482, + 529 + ], + "type": "text", + "content": " is critical for our model, especially in preserving detailed texture. Without it, the edited image tends to lose the appearance information from prompts and exhibits structural artifacts around the nose." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Ablation on Pretraining Data Scale for Customized Prompt Adaptations. In this part, we conduct ablation studies on the pretraining scale and investigate its influence on customized propmts. Specifically, we choose various 3DPE models pretrained with different style amounts (3, 10, and 20 styles respectively) and perform novel prompt adaptation on a single A6000 GPU. We choose the golden statue as the customized style and report LPIPS score between the predicted stylized images and the pseudo stylized labels produced by IP2P [7]. Experimental results in Tab. 2 reveal that the proposed method could benefit from the improving pretraining data scale and adapt faster and eventually perform better with more data, which suggests the potential of the model for future scaling up." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "text", + "content": "3DPE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 128 + ], + "type": "text", + "content": "Ablation on Tuning Data Scale for Customized Prompt Adaptations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 129, + 481, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 129, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 129, + 481, + 247 + ], + "type": "text", + "content": "To investigate the ability of our method to adapt to user-defined novel prompts, we conduct an ablation study on the number of novel-prompt data pairs required for training. Similar to the prior experiments on the pretraining scale, we perform adaptation with 2, 5, 10, 20, or 50 pairs of novel-prompt data on a single A6000 GPU. We experimentally find that our model can converge in about 5 minutes, and therefore, we report the perceptual distance between the model output and the provided ground truth at " + }, + { + "bbox": [ + 132, + 129, + 481, + 247 + ], + "type": "inline_equation", + "content": "0.1\\mathrm{min}" + }, + { + "bbox": [ + 132, + 129, + 481, + 247 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 129, + 481, + 247 + ], + "type": "inline_equation", + "content": "1\\mathrm{min}" + }, + { + "bbox": [ + 132, + 129, + 481, + 247 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 129, + 481, + 247 + ], + "type": "inline_equation", + "content": "2\\mathrm{min}" + }, + { + "bbox": [ + 132, + 129, + 481, + 247 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 132, + 129, + 481, + 247 + ], + "type": "inline_equation", + "content": "5\\mathrm{min}" + }, + { + "bbox": [ + 132, + 129, + 481, + 247 + ], + "type": "text", + "content": " for comparison. As shown in Tab. 3, our model can adapt to the new style in around 2 minutes. As the number of provided data pairs increases, the model can adapt faster at the beginning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 266, + 223, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 266, + 223, + 277 + ], + "spans": [ + { + "bbox": [ + 132, + 266, + 223, + 277 + ], + "type": "text", + "content": "4.5 Applications" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 286, + 481, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 286, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 286, + 481, + 453 + ], + "type": "text", + "content": "One application of our pipeline is to edit videos, requiring accurate reconstruction and efficient editing simultaneously. We demonstrate the effectiveness of our pipeline on talking face videos, as illustrated in Fig. 7. In this experiment, we perform sequence editing in a frame-by-frame manner by sending an image sequence from TalkingHead-1KH [72] dataset together with the text prompt \"Make him a bronze statue\" to 3DPE. The three rows in Fig. 7 respectively illustrate the input frames, the edited frames, and the novel views. Our method accurately reconstructs facial expressions, while achieving high-quality editing results. Since our method is 3D-aware, we can naturally view videos from novel perspectives. To better experience our method, we design an interactive system that allows users for interactive editing. Users are required to provide the image or text prompt to indicate the desired style and input an input image that is to be edited, and the system will apply our method on the given image and prompt to output the edited image as well as novel-view edited results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 472, + 216, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 216, + 484 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 216, + 484 + ], + "type": "text", + "content": "5 Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 498, + 481, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 557 + ], + "type": "text", + "content": "Limitations. Despite achieving state-of-the-art performance in quality and efficiency, our method exhibits inconsistencies in details for novel-view rendering because the EG3D framework relies on a super-resolution module. Additionally, when our method is applied to video editing, it presents flickering artifacts since our model is originally designed for image-based editing." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 558, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 481, + 665 + ], + "type": "text", + "content": "Conclusion. In this work, we introduce 3DPE for real-time 3D-aware portrait editing. By distilling the powerful knowledge of diffusion models and 3D GANs into a lightweight module, our model significantly reduces editing time while ensuring quality. Benefiting from the distillation process, 3DPE could handle multiple styles in a single model and perform both image-guided and text-guided editing. The advantages of our method empower us to generate photorealistic 3D portraits, a capability crucial for the visual effects industry, AR/VR systems, and teleconferencing, among other applications, while we firmly oppose any misuse that could compromise privacy and security." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Bai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 481, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 481, + 166 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 481, + 166 + ], + "type": "text", + "content": "This project was supported by the National Key R&D Program of China under grant number 2022ZD0161501." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 198, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 198, + 198 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 198, + 198 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 211, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 211, + 480, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 211, + 480, + 243 + ], + "spans": [ + { + "bbox": [ + 138, + 211, + 480, + 243 + ], + "type": "text", + "content": "1. Abdal, R., Lee, H.Y., Zhu, P., Chai, M., Siarohin, A., Wonka, P., Tulyakov, S.: 3DAvatarGAN: Bridging domains for personalized editable avatars. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 245, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 480, + 266 + ], + "type": "text", + "content": "2. Abdal, R., Qin, Y., Wonka, P.: Image2StyleGAN: How to embed images into the StyleGAN latent space? In: Int. Conf. Comput. Vis. (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 267, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 480, + 289 + ], + "type": "text", + "content": "3. Abdal, R., Qin, Y., Wonka, P.: Image2StyleGAN++: How to edit the embedded images? In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 290, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 290, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 138, + 290, + 480, + 311 + ], + "type": "text", + "content": "4. Aneja, S., Thies, J., Dai, A., Nießner, M.: ClipFace: Text-guided editing of textured 3d morphable models. In: ACM SIGGRAPH Conf. Proc. (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 312, + 480, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 312, + 480, + 333 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 480, + 333 + ], + "type": "text", + "content": "5. Bai, Q., Xu, Y., Zhu, J., Xia, W., Yang, Y., Shen, Y.: High-fidelity GAN inversion with padding space. In: Eur. Conf. Comput. Vis. (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 335, + 480, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 335, + 480, + 355 + ], + "spans": [ + { + "bbox": [ + 138, + 335, + 480, + 355 + ], + "type": "text", + "content": "6. Blanz, V., Vetter, T.: A morphable model for the synthesis of 3d faces. In: Seminal Graphics Papers: Pushing the Boundaries (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 357, + 480, + 377 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 480, + 377 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 480, + 377 + ], + "type": "text", + "content": "7. Brooks, T., Holynski, A., Efros, A.A.: InstructPix2Pix: Learning to follow image editing instructions. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 379, + 480, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 379, + 480, + 410 + ], + "spans": [ + { + "bbox": [ + 138, + 379, + 480, + 410 + ], + "type": "text", + "content": "8. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: Adv. Neural Inform. Process. Syst. (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 411, + 480, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 480, + 443 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 480, + 443 + ], + "type": "text", + "content": "9. Cai, S., Obukhov, A., Dai, D., Van Gool, L.: Pix2NeRF: Unsupervised conditional p-gan for single image to neural radiance fields translation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 445, + 480, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 445, + 480, + 476 + ], + "spans": [ + { + "bbox": [ + 138, + 445, + 480, + 476 + ], + "type": "text", + "content": "0. Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., Zheng, Y.: MasaCtrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In: Int. Conf. Comput. Vis. (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 478, + 480, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 478, + 480, + 499 + ], + "spans": [ + { + "bbox": [ + 138, + 478, + 480, + 499 + ], + "type": "text", + "content": "1. Ceylan, D., Huang, C.H.P., Mitra, N.J.: Pix2Video: Video editing using image diffusion. In: Int. Conf. Comput. Vis. (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 501, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 480, + 521 + ], + "type": "text", + "content": "2. Chai, W., Guo, X., Wang, G., Lu, Y.: StableVideo: Text-driven consistency-aware diffusion video editing. In: Int. Conf. Comput. Vis. (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 522, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 480, + 555 + ], + "type": "text", + "content": "3. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 555, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 587 + ], + "type": "text", + "content": "4. Chan, E.R., Monteiro, M., Kellnhofer, P., Wu, J., Wetzstein, G.: pi-GAN: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 589, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 480, + 609 + ], + "type": "text", + "content": "5. Chang, S., Kim, G., Kim, H.: HairNeRF: Geometry-aware image synthesis for hairstyle transfer. In: Int. Conf. Comput. Vis. (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "text", + "content": "6. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: ArcFace: Additive angular margin loss for deep face recognition. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "7. Deng, Y., Yang, J., Chen, D., Wen, F., Tong, X.: Disentangled and controllable face image generation via 3d imitative-contrastive learning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 421, + 91, + 448, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 91, + 448, + 100 + ], + "spans": [ + { + "bbox": [ + 421, + 91, + 448, + 100 + ], + "type": "text", + "content": "3DPE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "18. Deng, Y., Yang, J., Xu, S., Chen, D., Jia, Y., Tong, X.: Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 194 + ], + "type": "text", + "content": "19. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Int. Conf. Learn. Represent. (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 195, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 195, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 481, + 228 + ], + "type": "text", + "content": "20. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. In: Int. Conf. Learn. Represent. (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 228, + 481, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 481, + 261 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 481, + 261 + ], + "type": "text", + "content": "21. Gal, R., Patashnik, O., Maron, H., Bermano, A.H., Chechik, G., Cohen-Or, D.: StyleGAN-NADA: Clip-guided domain adaptation of image generators. ACM Trans. Graph. (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 262, + 481, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 481, + 284 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 481, + 284 + ], + "type": "text", + "content": "22. Gao, C., Shih, Y., Lai, W.S., Liang, C.K., Huang, J.B.: Portrait neural radiance fields from a single image. arXiv preprint arXiv:2012.05903 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 285, + 481, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 481, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 481, + 306 + ], + "type": "text", + "content": "23. Geyer, M., Bar-Tal, O., Bagon, S., Dekel, T.: TokenFlow: Consistent diffusion features for consistent video editing. In: Int. Conf. Learn. Represent. (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 307, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 307, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 307, + 481, + 329 + ], + "type": "text", + "content": "24. Graikos, A., Malkin, N., Jojic, N., Samaras, D.: Diffusion models as plug-and-play priors. In: Adv. Neural Inform. Process. Syst. (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 330, + 481, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 481, + 351 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 481, + 351 + ], + "type": "text", + "content": "25. Gu, J., Liu, L., Wang, P., Theobalt, C.: StyleNeRF: A style-based 3d-aware generator for high-resolution image synthesis. In: Int. Conf. Learn. Represent. (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 352, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 481, + 373 + ], + "type": "text", + "content": "26. Guo, J., Yu, J., Lattas, A., Deng, J.: Perspective reconstruction of human faces by joint mesh and landmark regression. In: Eur. Conf. Comput. Vis. (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "text", + "content": "27. Haque, A., Tancik, M., Efros, A.A., Holynski, A., Kanazawa, A.: Instruct-NeRF2NeRF: Editing 3d scenes with instructions. In: Int. Conf. Comput. Vis. (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 407, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 430 + ], + "type": "text", + "content": "28. Harkonen, E., Hertzmann, A., Lehtinen, J., Paris, S.: GANSpace: Discovering interpretable GAN controls. Adv. Neural Inform. Process. Syst. (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 430, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 481, + 453 + ], + "type": "text", + "content": "29. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 453, + 481, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 481, + 475 + ], + "type": "text", + "content": "30. Hertz, A., Aberman, K., Cohen-Or, D.: Delta denoising score. In: Int. Conf. Comput. Vis. (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 475, + 481, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 481, + 508 + ], + "type": "text", + "content": "31. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: Int. Conf. Learn. Represent. (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 509, + 481, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 481, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 481, + 531 + ], + "type": "text", + "content": "32. Hyung, J., Hwang, S., Kim, D., Lee, H., Choo, J.: Local 3d editing via 3d distillation of clip knowledge. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 532, + 481, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 481, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 481, + 563 + ], + "type": "text", + "content": "33. Jiang, K., Chen, S.Y., Fu, H., Gao, L.: NeRFFaceLighting: Implicit and disentangled face lighting representation leveraging generative prior in neural radiance fields. ACM Trans. Graph. (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 564, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 564, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 564, + 481, + 586 + ], + "type": "text", + "content": "34. Jin, W., Ryu, N., Kim, G., Baek, S.H., Cho, S.: Dr. 3D: Adapting 3d GANs to artistic drawings. In: ACM SIGGRAPH Asia Conf. Proc. (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 587, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 481, + 609 + ], + "type": "text", + "content": "35. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "text", + "content": "36. Kim, G., Chun, S.Y.: DATID-3D: Diversity-preserved domain adaptation using text-to-image diffusion for 3d generative model. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "37. Ko, J., Cho, K., Choi, D., Ryoo, K., Kim, S.: 3D GAN inversion with pose optimization. In: IEEE Winter Conf. Appl. Comput. Vis. (2023)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Bai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "38. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "39. Li, J., Li, J., Zhang, H., Liu, S., Wang, Z., Xiao, Z., Zheng, K., Zhu, J.: PREIM3D: 3d consistent precise image attribute editing from a single image. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 184, + 482, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 184, + 482, + 217 + ], + "spans": [ + { + "bbox": [ + 130, + 184, + 482, + 217 + ], + "type": "text", + "content": "40. Li, J., Liu, S., Liu, Z., Wang, Y., Zheng, K., Xu, J., Li, J., Zhu, J.: Instruct-Pix2NeRF: Instructed 3d portrait editing from a single image. arXiv preprint arXiv:2311.02826 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 218, + 482, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 218, + 482, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 218, + 482, + 251 + ], + "type": "text", + "content": "41. Li, J., Li, D., Xiong, C., Hoi, S.: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: Int. Conf. Mach. Learn. (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 251, + 482, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 274 + ], + "type": "text", + "content": "42. Lin, C., Lindell, D., Chan, E., Wetzstein, G.: 3D GAN inversion for controllable portrait image animation. In: Eur. Conf. Comput. Vis. Workshop (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 274, + 482, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 295 + ], + "type": "text", + "content": "43. Liu, S., Zhang, Y., Li, W., Lin, Z., Jia, J.: Video-P2P: Video editing with cross-attention control. arXiv preprint arXiv:2303.04761 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 297, + 482, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 329 + ], + "type": "text", + "content": "44. Liu, Z., Feng, R., Zhu, K., Zhang, Y., Zheng, K., Liu, Y., Zhao, D., Zhou, J., Cao, Y.: Cones: Concept neurons in diffusion models for customized generation. In: Int. Conf. Mach. Learn. (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 331, + 482, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 363 + ], + "type": "text", + "content": "45. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: Int. Conf. Learn. Represent. (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 364, + 482, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 482, + 396 + ], + "type": "text", + "content": "46. Nitzan, Y., Gharbi, M., Zhang, R., Park, T., Zhu, J.Y., Cohen-Or, D., Shechtman, E.: Domain expansion of image generators. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 397, + 482, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 482, + 430 + ], + "type": "text", + "content": "47. Or-El, R., Luo, X., Shan, M., Shechtman, E., Park, J.J., Kemelmacher-Shlizerman, I.: StyleSDF: High-resolution 3d-consistent image and geometry generation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 431, + 482, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 431, + 482, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 431, + 482, + 463 + ], + "type": "text", + "content": "48. Ouyang, H., Wang, Q., Xiao, Y., Bai, Q., Zhang, J., Zheng, K., Zhou, X., Chen, Q., Shen, Y.: CoDeF: Content deformation fields for temporally consistent video processing. arXiv preprint arXiv:2308.07926 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 464, + 482, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 497 + ], + "type": "text", + "content": "49. Pan, X., Xu, X., Loy, C.C., Theobalt, C., Dai, B.: A shading-guided generative implicit model for shape-accurate 3d-aware image synthesis. Adv. Neural Inform. Process. Syst. (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 498, + 482, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 520 + ], + "type": "text", + "content": "50. Patashnik, O., Wu, Z., Shechtman, E., Cohen-Or, D., Lischinski, D.: StyleCLIP: Text-driven manipulation of StyleGAN imagery. In: Int. Conf. Comput. Vis. (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 521, + 482, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 553 + ], + "type": "text", + "content": "51. Paysan, P., Knothe, R., Amberg, B., Romdhani, S., Vetter, T.: A 3d face model for pose and illumination invariant face recognition. In: IEEE international conference on advanced video and signal based surveillance (2009)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 554, + 482, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 554, + 482, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 554, + 482, + 575 + ], + "type": "text", + "content": "52. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: DreamFusion: Text-to-3d using 2d diffusion. In: Int. Conf. Learn. Represent. (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 576, + 482, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 576, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 576, + 482, + 609 + ], + "type": "text", + "content": "53. Qi, C., Cun, X., Zhang, Y., Lei, C., Wang, X., Shan, Y., Chen, Q.: FateZero: Fusing attentions for zero-shot text-based video editing. In: Int. Conf. Comput. Vis. (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 610, + 482, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 610, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 610, + 482, + 643 + ], + "type": "text", + "content": "54. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Int. Conf. Mach. Learn. (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 643, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 643, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 643, + 482, + 665 + ], + "type": "text", + "content": "55. Roich, D., Mokady, R., Bermano, A.H., Cohen-Or, D.: Pivotal tuning for latent-based editing of real images. ACM Trans. Graph. (2022)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 421, + 91, + 448, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 91, + 448, + 100 + ], + "spans": [ + { + "bbox": [ + 421, + 91, + 448, + 100 + ], + "type": "text", + "content": "3DPE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "56. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "text", + "content": "57. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dream-Booth: Fine tuning text-to-image diffusion models for subject-driven generation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "text", + "content": "58. Schwarz, K., Liao, Y., Niemeyer, M., Geiger, A.: GRAF: Generative radiance fields for 3d-aware image synthesis. In: Adv. Neural Inform. Process. Syst. (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "59. Schwarz, K., Sauer, A., Niemeyer, M., Liao, Y., Geiger, A.: VoxGRAF: Fast 3d-aware image synthesis with sparse voxel grids. In: Adv. Neural Inform. Process. Syst. (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 480, + 259 + ], + "type": "text", + "content": "60. Shen, Y., Gu, J., Tang, X., Zhou, B.: Interpreting the latent space of GANs for semantic face editing. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 260, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 480, + 293 + ], + "type": "text", + "content": "61. Shen, Y., Yang, C., Tang, X., Zhou, B.: InterFaceGAN: Interpreting the disentangled face representation learned by GANs. IEEE Trans. Pattern Anal. Mach. Intell. (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "type": "text", + "content": "62. Shi, Z., Peng, S., Xu, Y., Geiger, A., Liao, Y., Shen, Y.: Deep generative models on 3d representations: A survey. arXiv preprint arXiv:2210.15663 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 316, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 316, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 316, + 480, + 346 + ], + "type": "text", + "content": "63. Shi, Z., Shen, Y., Xu, Y., Peng, S., Liao, Y., Guo, S., Chen, Q., Yeung, D.Y.: Learning 3d-aware image synthesis with unknown pose distribution. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "64. Shi, Z., Xu, Y., Shen, Y., Zhao, D., Chen, Q., Yeung, D.Y.: Improving 3d-aware image synthesis with a geometry-aware discriminator. In: Adv. Neural Inform. Process. Syst. (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "text", + "content": "65. Skorokhodov, I., Tulyakov, S., Wang, Y., Wonka, P.: EpiGRAF: Rethinking training of 3d gans. In: Adv. Neural Inform. Process. Syst. (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "text", + "content": "66. Sun, J., Wang, X., Shi, Y., Wang, L., Wang, J., Liu, Y.: IDE-3D: Interactive disentangled editing for high-resolution 3d-aware portrait synthesis. ACM Trans. Graph. (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "text", + "content": "67. Sun, J., Wang, X., Zhang, Y., Li, X., Zhang, Q., Liu, Y., Wang, J.: FENeRF: Face editing in neural radiance fields. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "text", + "content": "68. Tov, O., Alaluf, Y., Nitzan, Y., Patashnik, O., Cohen-Or, D.: Designing an encoder for StyleGAN image manipulation. ACM Trans. Graph. (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "type": "text", + "content": "69. Tran, P., Zakharov, E., Ho, L.N., Tran, A.T., Hu, L., Li, H.: VOODOO 3D: Volumetric portrait disentanglement for one-shot 3d head reenactment. In: IEEE Conf. Comput. Vis. Pattern Recog. (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "text", + "content": "70. Trevithick, A., Chan, M., Stengel, M., Chan, E., Liu, C., Yu, Z., Khamis, S., Chandraker, M., Ramamoorthi, R., Nagano, K.: Real-time radiance fields for single-image portrait view synthesis. ACM Trans. Graph. (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "text", + "content": "71. Wang, Q., Shi, Z., Zheng, K., Xu, Y., Peng, S., Shen, Y.: Benchmarking and analyzing 3d-aware image synthesis with a modularized codebase. In: Adv. Neural Inform. Process. Syst. (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "type": "text", + "content": "72. Wang, T.C., Mallya, A., Liu, M.Y.: One-shot free-view neural talking-head synthesis for video conferencing. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "text", + "content": "73. Wang, W., Jiang, Y., Xie, K., Liu, Z., Chen, H., Cao, Y., Wang, X., Shen, C.: Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "text", + "content": "74. Wang, Y., Xu, T., Wu, Y., Li, M., Chen, W., Xu, L., Yu, J.: NARRATE: A normal assisted free-view portrait stylizer. arXiv preprint arXiv:2207.00974 (2022)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Bai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 468 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "text", + "content": "75. Xia, W., Xue, J.H.: A survey on deep generative 3d-aware image synthesis. ACM Computing Surveys (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "type": "text", + "content": "76. Xie, J., Ouyang, H., Piao, J., Lei, C., Chen, Q.: High-fidelity 3d GAN inversion by pseudo-multi-view optimization. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "text", + "content": "77. Xu, Y., Peng, S., Yang, C., Shen, Y., Zhou, B.: 3d-aware image synthesis via learning structural and textural representations. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 480, + 226 + ], + "type": "text", + "content": "78. Xu, Y., Shen, Y., Zhu, J., Yang, C., Zhou, B.: Generative hierarchical features from synthesizing images. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 133, + 227, + 480, + 258 + ], + "type": "text", + "content": "79. Yang, S., Jiang, L., Liu, Z., Loy, C.C.: Pastiche master: Exemplar-based high-resolution portrait style transfer. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 133, + 259, + 480, + 281 + ], + "type": "text", + "content": "80. Yang, S., Zhou, Y., Liu, Z.,, Loy, C.C.: Rerender a video: Zero-shot text-guided video-to-video translation. In: ACM SIGGRAPH Asia Conf. Proc. (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 281, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 281, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 133, + 281, + 480, + 314 + ], + "type": "text", + "content": "81. Yin, F., Zhang, Y., Wang, X., Wang, T., Li, X., Gong, Y., Fan, Y., Cun, X., Shan, Y., Oztireli, C., et al.: 3D GAN inversion with facial symmetry prior. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 133, + 315, + 480, + 346 + ], + "type": "text", + "content": "82. Zhang, J., Lan, Y., Yang, S., Hong, F., Wang, Q., Yeo, C.K., Liu, Z., Loy, C.C.: DeformToon3D: Deformable neural radiance fields for 3d toonification. In: Int. Conf. Comput. Vis. (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 347, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 347, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 133, + 347, + 480, + 369 + ], + "type": "text", + "content": "83. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Int. Conf. Comput. Vis. (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 133, + 369, + 480, + 402 + ], + "type": "text", + "content": "84. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 133, + 403, + 480, + 424 + ], + "type": "text", + "content": "85. Zhou, P., Xie, L., Ni, B., Tian, Q.: CIPS-3D: A 3d-aware generator of gans based on conditionally-independent pixel synthesis. arXiv preprint arXiv:2110.09788 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 425, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 425, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 133, + 425, + 480, + 445 + ], + "type": "text", + "content": "86. Zhu, J., Shen, Y., Zhao, D., Zhou, B.: In-domain GAN inversion for real image editing. In: Eur. Conf. Comput. Vis. (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "type": "text", + "content": "87. Zhu, J.Y., Krahenbuhl, P., Shechtman, E., Efros, A.A.: Generative visual manipulation on the natural image manifold. In: Eur. Conf. Comput. Vis. (2016)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "text", + "content": "3DPE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Real-time Holistic Robot Pose Estimation with Unknown States/0b349d89-54a2-48da-a5a9-0c888f46a6ff_content_list.json b/2024/Real-time Holistic Robot Pose Estimation with Unknown States/0b349d89-54a2-48da-a5a9-0c888f46a6ff_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..272cb2c81800fad70a0621097c52ee165a65ddfc --- /dev/null +++ b/2024/Real-time Holistic Robot Pose Estimation with Unknown States/0b349d89-54a2-48da-a5a9-0c888f46a6ff_content_list.json @@ -0,0 +1,1701 @@ +[ + { + "type": "text", + "text": "Real-time Holistic Robot Pose Estimation with Unknown States", + "text_level": 1, + "bbox": [ + 232, + 140, + 772, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shikun Ban1, Juling Fan1, Xiaoxuan Ma1, Wentao Zhu1,2, Yu Qiao5, and Yizhou Wang1,2,3,4", + "bbox": [ + 292, + 210, + 710, + 244 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1 Center on Frontiers of Computing Studies, School of Computer Science, Peking University", + "2 Inst. for Artificial Intelligence, Peking University", + "3 Nat'l Eng. Research Center of Visual Technology", + "4 Nat'l Key Lab of General Artificial Intelligence" + ], + "bbox": [ + 254, + 253, + 746, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{5}$ Department of Automation, School of Electronics, Information and Electrical Engineering, Shanghai Jiao Tong University", + "bbox": [ + 236, + 324, + 764, + 352 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Estimating robot pose from RGB images is a crucial problem in computer vision and robotics. While previous methods have achieved promising performance, most of them presume full knowledge of robot internal states, e.g. ground-truth robot joint angles. However, this assumption is not always valid in practical situations. In real-world applications such as multi-robot collaboration or human-robot interaction, the robot joint states might not be shared or could be unreliable. On the other hand, existing approaches that estimate robot pose without joint state priors suffer from heavy computation burdens and thus cannot support real-time applications. This work introduces an efficient framework for real-time robot pose estimation from RGB images without requiring known robot states. Our method estimates camera-to-robot rotation, robot state parameters, keypoint locations, and root depth, employing a neural network module for each task to facilitate learning and sim-to-real transfer. Notably, it achieves inference in a single feed-forward pass without iterative optimization. Our approach offers a $12 \\times$ speed increase with state-of-the-art accuracy, enabling real-time holistic robot pose estimation for the first time. Code and models are available at https://github.com/Oliverbansk/Holistic-Robot-Pose-Estimation", + "bbox": [ + 259, + 386, + 740, + 651 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 696, + 375, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Robot pose estimation is a fundamental task in computer vision and robotics as it not only lays the groundwork for multiple downstream tasks, e.g. visually-guided object grasping and manipulation [21,40,49], but also paves the way for a variety of applications, such as autonomous navigation [1,13,33,34], human-robot interaction [3,39,43,44,46,48] and multi-robot collaboration [20,29,35,45].", + "bbox": [ + 212, + 726, + 787, + 805 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Correspondence to: Wentao Zhu and Yu Qiao .", + "bbox": [ + 228, + 810, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/400b567273e0a3757470601efe968e5aad3887625d0ab8d4e346edfc1927a0d6.jpg", + "image_caption": [ + "Fig. 1: The majority of previous robot pose estimation methods assume known robot joint states and focus solely on estimating the camera-to-robot pose, i.e. camera-to-robot rotation and translation. In contrast, the holistic robot pose estimation problem requires estimating both joint states and the camera-to-robot pose, given only an RGB image without known joint states. For holistic robot pose estimation, RoboPose [15] uses costly test-time optimization (Render-and-Compare) to iteratively refine the predictions. In contrast, our feed-forward method achieves state-of-the-art accuracy with a $12 \\times$ speed boost." + ], + "image_footnote": [], + "bbox": [ + 218, + 147, + 781, + 281 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nonetheless, most previous works for monocular robot pose estimation [17,18,22,40] rely on a known set of robot joint angles as an additional input to alleviate the difficulty of the estimation process. Yet, this prerequisite information of the robot state is not always available in practical scenarios:", + "bbox": [ + 212, + 428, + 787, + 491 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- For instance, low-cost robotic arms without sophisticated sensors or internal monitors can not provide the visual calibration process with readily accessible robot state measurements [4,10,49].", + "- In the context of multi-robot collaboration, the real-time synchronization of joint state data may be hindered by environmental limitations [24], necessitating pose estimation without robot joint state information, and additional monitoring of the joint states.", + "- In domains such as human-robot interaction [36,41] and competitive robot gaming [32,42], the reliability of the state information from counterpart robots is not always guaranteed, as trust issues may arise in robotic systems [11,12,14]. In this sense, failing to obtain independent estimation of the counterpart robot joint state may lead to potential safety issues." + ], + "bbox": [ + 223, + 498, + 787, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These scenarios further emphasize the urgent need for monocular robot pose estimation with unknown robot joint state. In this work, we address the challenge of Holistic Robot Pose Estimation with unknown internal states for articulated robots. Specifically, given a monocular RGB image, we aim to simultaneously solve for 6D robot pose (3D rotation and translation relative to the camera) and fine-grained robot joint state parameters (including joint angles for revolute joints and displacement for prismatic joints).", + "bbox": [ + 212, + 689, + 787, + 795 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nonetheless, learning holistic robot pose estimation without joint state information presents significant challenges. The primary challenge is the dual ambiguity of unknown joint states and 6D pose. Specifically, estimating the 6D pose", + "bbox": [ + 212, + 795, + 787, + 842 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "S. Ban et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of an articulated robot is more complex than for a rigid object due to shape and scale variability from unknown joint states. Conversely, estimating joint states is complicated by the need for 2D-to-3D correspondences and the unknown 6D pose. This intertwined issue requires novel solutions to decouple joint states from 6D pose estimation. Furthermore, the sophisticated robot morphologies vary from model to model, requiring non-trivial model design. In addition, The robot parts are frequently invisible in the monocular image as a result of self-occlusion and truncation, posing more challenges for robust estimation. Finally, the domain gap between synthetic training data and real-world images introduces additional challenges for learning-based approaches.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It is worth noting that some previous works [15,49] have also addressed holistic estimation with unknown states, tackling some challenges. However, these methods all necessitate iterative optimization at inference, which is computationally intensive and unsuitable for real-time applications. Motivated by the need for efficient holistic robot pose estimation without knowledge of the robot state, we propose an end-to-end framework that directly estimates camera-to-robot pose and robot joint state parameters, requiring only a single feed-forward pass during inference.", + "bbox": [ + 212, + 300, + 787, + 421 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In pursuit of this goal, we factorize the holistic robot pose estimation task into several sub-tasks, including the estimation of camera-to-robot rotation, robot joint states, root depth, and root-relative keypoint locations. We then design corresponding neural network modules for each sub-task. Specifically, the standalone root depth estimator disentangles the estimation of scale-variant and scale-invariant variables. The keypoint estimator bridges the gap between the pixel space and the robotic parameter space. Additionally, we utilize self-supervised consistency regularizations which could be performed on the real images without ground truth (GT).", + "bbox": [ + 212, + 425, + 787, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We demonstrate that our designs of modular feed-forward networks improve model generalization while maintaining high computation efficiency. Consequently, our approach achieves state-of-the-art estimation accuracy with unknown robot states and a $12 \\times$ speedup compared to the iterative Render-and-Compare (RnC) [15] approach, as depicted in Fig. 1. The key insight behind our approach is that we could get rid of the cumbersome RnC [15] and error-prone Perspective-n-Points (PnP) [18,22,40] schemes to establish 2D-3D correspondences. Instead, we demonstrate that feed-forward networks could well handle the problem with proper problem decomposition and meticulous design.", + "bbox": [ + 212, + 564, + 787, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our contributions are three-fold:", + "bbox": [ + 238, + 704, + 566, + 718 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We present an efficient end-to-end learning framework for the holistic robot pose estimation problem with unknown robot states.", + "2. We propose to factorize and decompose the holistic robot pose estimation problem into several sub-tasks and tackle them with respective modules.", + "3. We achieve state-of-the-art performance on various robot models while delivering a significant $12 \\times$ speedup, paving the way for real-time applications." + ], + "bbox": [ + 222, + 742, + 782, + 837 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Holistic Robot Pose", + "bbox": [ + 596, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 388, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Hand-eye calibration", + "text_level": 1, + "bbox": [ + 215, + 171, + 434, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The goal of Hand-Eye Calibration (HEC) is to determine the transformation between the position of a robot's end-effector (hand) and its vision system (eye). This task can be categorized into two types: eye-in-hand where the camera is fixed at the robot end-effector [27], and eye-to-hand which refers to the configurations where the camera is placed at a distance from the robot [5,23,30]. This work focuses on a similar setting with eye-to-hand. Traditional HEC methods involve attaching a fiducial marker to the end effector [6,7,28], placing it in several known positions, detecting the marker positions in images, and then solving for the camera-to-robot transformation. Despite its broad applicability, the classical HEC pipeline necessitates a complete offline calibration process every time the camera-robot transformation changes, which can be tedious in real-world applications. Consequently, recent works, such as DREAM [18], propose an online, markerless calibration problem, as discussed below.", + "bbox": [ + 212, + 191, + 787, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Image-based robot pose estimation", + "text_level": 1, + "bbox": [ + 215, + 406, + 552, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous works can be divided into two categories based on whether the robot states are known or not. Most previous works estimate the 6D camera-to-robot pose with known robot states. For example, DREAM [18] first trains a 2D keypoint estimator and then applies Perspective- $n$ -Point (PnP) [19] to obtain camera-to-robot pose, showing comparable results with the classical methods. Some work further utilizes segmentation masks and edges to aid pose estimation [8,16]. Chen et al. [2] apply rendering-based camera pose optimization and consistency-based joint space exploration. SGTAPose [40] leverages robot structure prior and temporal information. Lu et al. [22] propose to train on real-world data using foreground segmentation and differentiable rendering in a self-supervised manner.", + "bbox": [ + 212, + 426, + 787, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recently, several works have shifted to estimating robot pose with unknown joint states. Zuo et al. [49] estimate the robot pose and joint angles of a simple 4-DoF robotic arm from a single RGB image by solving a nonlinear nonconvex optimization problem. However, this method is not well-suited for robot arms with higher complexity. Labbé et al. [15] propose a render-and-compare approach to iteratively refine the estimation for both known and unknown states, and apply the method to different robot models up to 15-DoF. However, the iterative optimization process limits its applications in real-time scenarios. In contrast, our approach can be deployed with a single feed-forward at inference and achieves real-time efficiency.", + "bbox": [ + 212, + 593, + 789, + 743 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 762, + 330, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Overview", + "text_level": 1, + "bbox": [ + 215, + 789, + 341, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an RGB image $\\mathbf{I}$ , we employ a suite of neural network modules, including DepthNet, JointNet, RotationNet, and KeypointNet. These modules are used", + "bbox": [ + 215, + 809, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "S. Ban et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/30c7f4508dca82009277bbbef491c2aa93c01e4d95b128047e0361da616dfe0a.jpg", + "image_caption": [ + "Fig. 2: Framework overview. The JointNet and the RotationNet regress joint state parameters $\\mathbf{q}$ and camera-to-robot rotation $\\mathbf{R}$ , respectively. The KeypointNet estimates root-relative 3D keypoint locations $\\mathbf{P}^r$ . The DepthNet's estimation of root depth $d$ is combined with $\\mathbf{P}^r$ to acquire absolute 3D keypoint locations $\\mathbf{P}'$ and camera-to-robot translation $\\mathbf{t}$ . Joint state parameters $\\mathbf{q}$ , rotation $\\mathbf{R}$ and translation $\\mathbf{t}$ are used to compute 3D keypoint locations $\\mathbf{P}$ via forward-kinematics." + ], + "image_footnote": [], + "bbox": [ + 217, + 145, + 785, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to respectively estimate the absolute root depth $d$ , joint state parameters $\\mathbf{q}$ , camera-to-robot rotation $\\mathbf{R} \\in SO(3)$ , and root-relative 3D keypoint locations $\\mathbf{P}^r \\in \\mathbb{R}^{N \\times 3}$ , where $N$ represents the total number of keypoints. The keypoints are defined at the joint articulations, in line with previous works [15,18,22], and one keypoint is designated as the robot root. The root joint is critical for computing the camera-to-robot rotation $\\mathbf{R}$ and translation $\\mathbf{t}$ , as well as serving as the origin for the forward kinematics (FK) process.", + "bbox": [ + 212, + 496, + 787, + 603 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our modularized design offers several key insights. Firstly, DepthNet is tailored to learn the scale-variant camera-robot spatial relationship, while the other modules concentrate on the scale-invariant robot features. Secondly, while the joint state parameters $\\mathbf{q}$ provide a compact representation of robot morphologies subject to robotic structure constraints, they are highly abstract and thus challenging to directly regress from image features. To address this, we utilize the 3D keypoint locations as an intermediary. On the one hand, they can be effectively estimated with pixel-aligned image features; on the other hand, they are inherently linked to $\\mathbf{q}$ through an FK process.", + "bbox": [ + 212, + 608, + 787, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "An overview of the proposed framework is illustrated in Fig. 2. Specifically, DepthNet estimates the absolute root depth $d$ , which, combined with KeypointNet's output $\\mathbf{P}^r \\in \\mathbb{R}^{N \\times 3}$ , provides the absolute 3D keypoint locations $\\mathbf{P}' \\in \\mathbb{R}^{N \\times 3}$ . The root keypoint location in $\\mathbf{P}'$ is the camera-to-robot translation $\\mathbf{t}$ . The estimated joint state parameters $\\mathbf{q}$ , along with the camera-to-robot rotation $\\mathbf{R}$ and translation $\\mathbf{t}$ , are then used to perform FK, yielding the final", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Holistic Robot Pose", + "bbox": [ + 596, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D keypoint estimations $\\mathbf{P} \\in \\mathbb{R}^{N \\times 3}$ . Please refer to the appendix for a detailed illustration of the relations of these variables.", + "bbox": [ + 212, + 145, + 782, + 175 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the following, we will first introduce the architecture design of the network modules in Sec. 3.2, and then explain the training loss design consisting of ground-truth supervision and self-supervision in Sec. 3.3.", + "bbox": [ + 212, + 176, + 787, + 224 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Architecture", + "text_level": 1, + "bbox": [ + 215, + 248, + 367, + 262 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DepthNet Monocular object depth estimation suffers from significant ambiguities, as the absolute depth value and camera intrinsics are deeply intertwined [25]. For instance, a robot in a fixed pose relative to the camera may appear larger in photos taken with cameras having larger focal lengths. Scaling both the absolute depth and focal lengths equally results in identical object size and appearance in different images. To address this challenge, we implement DepthNet following the approach in [26] to disentangle depth from camera intrinsics. Initially, we use camera intrinsics to calculate a normalized, coarse depth value $d_{c}$ :", + "bbox": [ + 212, + 276, + 787, + 398 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nd _ {c} = \\sqrt {\\frac {f _ {x} \\cdot f _ {y} \\cdot A _ {r e a l}}{A _ {b b o x}}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 411, + 785, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $f_{x}$ and $f_{y}$ represent the focal lengths along the x- and y-axes, $A_{real}$ is the area of the robot in the x-y plane of the camera frame (in $mm^2$ ), and $A_{bbox}$ is the area of the robot's bounding box (in $pixel^2$ ). This method of calculating $d_{c}$ normalizes it in 3D space and provides an approximate depth estimation for subsequent refinement. The primary function of DepthNet is to predict a correction factor $\\lambda$ based on the input image. We multiply the estimated $\\lambda$ by $d_{c}$ to obtain the final absolute root depth $d$ :", + "bbox": [ + 212, + 454, + 787, + 561 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nd = \\lambda \\cdot d _ {c}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 465, + 575, + 784, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With a fixed camera, the relationship between the normalized coarse depth $d_{c}$ and the robot's size in the image is linear. However, robots can assume poses with the same depth but different sizes in the image, making $d_{c}$ inadequate for accurate root depth representation. The correction factor $\\lambda$ introduces non-linear flexibility to the depth estimation, addressing this issue by adjusting $d_{c}$ based on image features.", + "bbox": [ + 212, + 602, + 787, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "JointNet The goal of JointNet is to estimate joint state parameters $\\mathbf{q} = (q_{1}, q_{2}, \\dots, q_{J})$ , where $J$ denotes the number of robot joints. As this work deals with 1-DoF joints (revolute or prismatic), $J$ equals the robot's DoF. JointNet shares the same feature extractor with RotationNet and KeypointNet. We extract feature map $\\mathbf{F} \\in \\mathbb{R}^{C \\times H \\times W}$ (where $C$ denotes the number of channels, and $H \\times W$ denotes the spatial resolution) and use global average pooling to obtain $\\mathbf{f} \\in \\mathbb{R}^{C}$ . Finally, we employ several MLP blocks to regress the joint state parameters $\\mathbf{q}$ from $\\mathbf{f}$ .", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "S. Ban et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "RotationNet The goal of RotationNet is to estimate the camera-to-robot rotation $\\mathbf{R} \\in SO(3)$ . We employ an MLP to estimate $\\mathbf{R}$ from the image feature $\\mathbf{f}$ . We utilize the 6D continuous rotation representation [47] to parameterize $\\mathbf{R}$ .", + "bbox": [ + 214, + 146, + 787, + 195 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "KeypointNet The goal of KeypointNet is to estimate root-relative 3D keypoint locations $\\mathbf{P}^r\\in \\mathbb{R}^{N\\times 3}$ for $N$ robot keypoints. We transform the feature map $\\mathbf{F}$ to a 3D heatmap $\\mathbf{H}\\in \\mathbb{R}^{N\\times D\\times H'\\times W'}$ through convolutional layers, where $D\\times H'\\times W'$ is the spatial resolution discretizing the 3D space. The 3D heatmap encodes per-voxel likelihood of containing each robot keypoint. We derive the root-relative 3D positions of each keypoint $\\mathbf{P}_n^r$ by performing spatial integral techniques [38] over the corresponding heatmap $\\mathbf{H}_n$ in a differentiable manner:", + "bbox": [ + 214, + 210, + 787, + 316 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {n} ^ {r} = \\sum_ {k = 1} ^ {D} \\sum_ {i = 1} ^ {H ^ {\\prime}} \\sum_ {j = 1} ^ {W ^ {\\prime}} (k, i, j) \\cdot \\mathbf {H} _ {n} (k, i, j). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 324, + 785, + 367 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "By combining $\\mathbf{P}^r$ with the estimated root depth $d$ , we obtain the regressed 3D keypoint locations $\\mathbf{P}'$ and root translation $\\mathbf{t}$ . We further calculate the FK-based 3D keypoint locations $\\mathbf{P}$ with differentiable FK parameterized by $\\mathbf{q}$ , $\\mathbf{R}$ , and $\\mathbf{t}$ .", + "bbox": [ + 214, + 373, + 787, + 420 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Training loss", + "text_level": 1, + "bbox": [ + 214, + 436, + 370, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ground-truth supervision For training on synthetic datasets where we have access to all the ground-truth information, we apply separate supervision for all the sub-tasks to train the corresponding network modules:", + "bbox": [ + 214, + 459, + 787, + 506 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d e p t h}} = \\| d - \\hat {d} \\| _ {1}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 436, + 518, + 785, + 536 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {j o i n t}} = \\| \\mathbf {q} - \\hat {\\mathbf {q}} \\| _ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 437, + 555, + 784, + 571 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r o t}} = \\left\\| \\mathbf {R} - \\hat {\\mathbf {R}} \\right\\| _ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 437, + 589, + 784, + 606 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t r a n s}} = \\| \\mathbf {t} - \\hat {\\mathbf {t}} \\| _ {2}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 437, + 625, + 784, + 641 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {k p t s}} = \\| \\mathbf {P} - \\hat {\\mathbf {P}} \\| _ {2} + \\| \\mathbf {p} - \\hat {\\mathbf {p}} \\| _ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 659, + 784, + 676 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {k p t s}} ^ {\\prime} = \\left\\| \\mathbf {P} ^ {\\prime} - \\hat {\\mathbf {P}} ^ {\\prime} \\right\\| _ {2} + \\left\\| \\mathbf {p} ^ {\\prime} - \\hat {\\mathbf {p}} ^ {\\prime} \\right\\| _ {2}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 691, + 784, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $d$ , $\\mathbf{q}$ , $\\mathbf{R}$ , $\\mathbf{t}$ , $\\mathbf{P}$ , and $\\mathbf{P}'$ are GT root depth, joint states, camera-to-robot rotation and translation, FK-based keypoint locations and regressed keypoint locations, respectively; $\\mathbf{p}$ and $\\mathbf{p}'$ are 2D projections of $\\mathbf{P}$ and $\\mathbf{P}'$ using the camera intrinsics; $\\hat{}$ denotes model estimations.", + "bbox": [ + 214, + 717, + 787, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We then train all the network modules end-to-end using", + "bbox": [ + 238, + 777, + 648, + 794 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {G T}} = \\mathcal {L} _ {\\text {j o i n t}} + \\mathcal {L} _ {\\text {r o t}} + \\mathcal {L} _ {\\text {t r a n s}} + \\lambda \\left(\\mathcal {L} _ {\\text {k p t s}} + \\mathcal {L} _ {\\text {k p t s}} ^ {\\prime}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 800, + 784, + 820 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where we set $\\lambda = 10.0$ to balance the loss terms.", + "bbox": [ + 214, + 824, + 566, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Holistic Robot Pose", + "bbox": [ + 596, + 114, + 732, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Self-supervision Apart from GT supervision on synthetic data, our framework also enables scaling the training pipeline to unlabeled real-world data in a self-supervised manner [22]. We employ two consistency-based training regularizations to improve model generalization.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Despite the absence of GT joint states and keypoint locations, we regularize the consistency of the FK-based keypoints $\\hat{\\mathbf{P}}$ and regressed keypoints $\\hat{\\mathbf{P}}'$ by:", + "bbox": [ + 214, + 207, + 784, + 239 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {k c}} = \\left\\| \\hat {\\mathbf {P}} - \\hat {\\mathbf {P}} ^ {\\prime} \\right\\| _ {2}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 439, + 252, + 784, + 270 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce this keypoint consistency regularization to effectively reconcile the predictions subject to robot structure constraints $\\hat{\\mathbf{P}}$ and predictions based on the pixel-aligned feature $\\hat{\\mathbf{P}}'$ . In addition, similar to [22], we also regularize the consistency between robot rendering mask $\\mathbf{M}_{\\mathrm{render}}$ and foreground segmentation mask $\\mathbf{M}_{\\mathrm{seg}}$ using their Intersection-over-Union (IoU):", + "bbox": [ + 214, + 282, + 784, + 359 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {m c}} = 1 - \\frac {\\mathcal {S} _ {\\text {i n t e r s e c t i o n}}}{\\mathcal {S} _ {\\text {u n i o n}}}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 419, + 372, + 784, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $S_{\\text{intersection}}$ and $S_{\\text{union}}$ is the area of the intersection and the union of $\\mathbf{M}_{\\text{render}}$ and $\\mathbf{M}_{\\text{seg}}$ , respectively. $\\mathbf{M}_{\\text{render}}$ is obtained with differentiable rendering using $\\mathbf{q}$ , $\\mathbf{R}$ , and $\\mathbf{t}$ [15]; $\\mathbf{M}_{\\text{seg}}$ is obtained with an image segmentation model [22].", + "bbox": [ + 214, + 409, + 785, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The overall self-supervision training objective is as follows:", + "bbox": [ + 238, + 455, + 663, + 469 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s e l f}} = \\mathcal {L} _ {\\mathrm {k c}} + \\lambda_ {\\mathrm {m c}} \\mathcal {L} _ {\\mathrm {m c}} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 484, + 784, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where we set $\\lambda_{\\mathrm{mc}} = 1.0$ . We apply end-to-end self-supervised training for the sim-to-real domain adaptation stage.", + "bbox": [ + 214, + 511, + 782, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 566, + 375, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Dataset and metrics", + "text_level": 1, + "bbox": [ + 215, + 601, + 429, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We utilize the benchmark dataset DREAM [18], for both training and testing in our study. This dataset comprises images from three widely used robotic arms, namely, Franka Emika Panda (Panda), Kuka iiwa7 (Kuka), and Rethink Robotics Baxter (Baxter). The training dataset for each robot consists of approximately 100k synthetic images generated using the domain randomization (DR) technique. These images encompass scenarios with varying camera-to-robot poses and joint states. The testing datasets include both DR-generate synthetic images for each robot and photorealistic synthetic images (denoted as \"Photo\") for Panda and Kuka. Additionally, real-world image datasets of Panda are provided. Panda-3Cam comprises a total of 17k image frames captured by three different cameras, namely Azure Kinect (AK), XBOX 360 Kinect (XK), and Realsense (RS). This dataset is presented as three sub-datasets, each corresponding to one camera. Another real-world dataset, Panda-ORB, contains 32,315 images captured from 27 viewpoints using a Realsense camera.", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "S. Ban et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate the accuracy of our estimations using the Average Distance (ADD) metric, which represents the average Euclidean distances (in millimeters) between predicted 3D keypoint locations and their corresponding ground truth locations. Our reported metrics include both the Area Under the Curve (AUC) value and the mean value of ADD. The AUC integrates ADD over different thresholds (with the maximum threshold set at $100\\mathrm{mm}$ ), while the mean value effectively represents the general prediction accuracy.", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Implementation details", + "text_level": 1, + "bbox": [ + 215, + 279, + 454, + 295 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We first pre-train DepthNet on the synthetic training dataset for 100 epochs, employing a learning rate of $1e - 4$ and relying solely on ground truth supervision for depth. Subsequently, we conduct training for our entire model based on the pretrained DepthNet for an additional 100 epochs with a learning rate of $1e - 4$ and a decay rate of 0.95. The Adam optimizer is employed in each stage to optimize the network parameters, with the momentum set to 0.9. As depth influences the global offset of all keypoints, a pre-trained DepthNet serves as a reasonable starting point for training other networks. On the Panda real datasets, we further apply self-supervised training on real-world images to overcome the sim-to-real domain gap with a learning rate of $1e - 6$ . This fine-tuning process uses only image data and no ground-truth labels.", + "bbox": [ + 212, + 309, + 789, + 474 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Evaluation and Comparison", + "text_level": 1, + "bbox": [ + 215, + 502, + 493, + 518 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/ea71af5dc6561dc3230bb46287923c5aa5352473126706983888768e556c09a2.jpg", + "table_caption": [ + "Table 1: Comparison of AUC $\\uparrow$ of the ADD distribution curve and mean $\\downarrow$ of the ADD, on each dataset in the DREAM datasets. Panda 3CAM datasets and Panda ORB are real-world datasets and the rest are synthetic datasets. † denotes using ground-truth joint state parameters." + ], + "table_footnote": [], + "table_body": "
MethodMetricBaxter DRKuka DRKuka PhotoPanda DRPanda PhotoPanda 3CAM-AKPanda 3CAM-XKPanda 3CAM-RSPanda ORB
† DREAM-F 18AUC ↑---81.379.568.924.476.161.9
† DREAM-Q 18AUC ↑75.5--77.874.352.437.578.057.1
† DREAM-H 18AUC ↑-73.372.182.981.160.564.078.869.1
RoboPose 15AUC ↑32.780.273.282.979.770.477.674.370.4
OursAUC ↑58.875.173.982.782.082.276.075.275.2
† DREAM-F 18Mean ↓---552.0785.811413.1491911.42077.495319.1
† DREAM-Q 18Mean ↓59.0--359.0569.078089.354178.227.264247.6
† DREAM-H 18Mean ↓-5613.211099.4325.6518.856.57381.623.625685.3
RoboPose 15Mean ↓87.934.289.021.432.234.322.326.030.1
OursMean ↓45.028.029.018.018.418.924.024.824.8
", + "bbox": [ + 218, + 640, + 787, + 809 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Holistic Robot Pose", + "bbox": [ + 596, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Accuracy We first quantitatively evaluate the estimation accuracy of our method and compare it with the state-of-the-art (SOTA) approaches. To the best of our knowledge, among prior works, only RoboPose [15] addresses a scenario similar to ours, where both the pose and joint state parameters for an industry-level robot are estimated from a single RGB image. We also present the results of DREAM [18], a representative feed-forward method using ground-truth joint state parameters, as a reference. Table [1] presents the comparisons of AUC and the mean value of ADD. Our method outperforms SOTA on most datasets, significantly reducing joint localization errors. Specifically, DREAM [18] detects the 2D keypoint locations first, then employs a PnP process to solve the 2D-to-3D transformation. However, once the 2D keypoint detector fails due to truncations and occlusions, the PnP process produces unreliable numerical solutions. RoboPose [15] utilizes iterative render-and-compare, which is prone to falling into local optima on complex robot models with symmetric structures (e.g., Baxter). In contrast, our approach directly regresses both the robot pose and joint state parameters, enabling end-to-end optimization from multi-facet supervision signals. We also visualize the complete ADD distribution. As shown in Fig. [3], our approach achieves higher accuracy across most ADD thresholds. Notably, our approach attains near $100\\%$ accuracy with a $60\\mathrm{mm}$ threshold, indicating more reliable estimation results on challenging cases with fewer outliers.", + "bbox": [ + 212, + 146, + 792, + 450 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/e842e3e6f8250fee4f6c64e0d874ee0a105c168976b83f1753ecbd0058f850ce.jpg", + "image_caption": [ + "Fig. 3: Comparison of ADD distributions on the real-world datasets between our approach and RoboPose [15]. The y-axis represents the accuracy of our estimation at different ADD thresholds." + ], + "image_footnote": [], + "bbox": [ + 217, + 470, + 357, + 573 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/5070e1baded4945afba15df5224f21065d183de7284bc65f1eade51e3f032cb9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 470, + 501, + 571 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/ef0cffca512ef254468f10e0b926f4ecb66ac706a3ab91f6a249aa9fddcdc01c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 470, + 642, + 571 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/8181d5b4fbddedf7342b7d41ae4fdf8af9e8c28d4e4e2f7ba68cc2105f8bce84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 470, + 784, + 571 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Finally, we visualize the predictions of our approach and compare them with SOTA to better understand the performance improvement. Figure 4 illustrates our approach's ability to produce high-quality estimates across different robot morphologies. Particularly, our approach demonstrates superior performance in highly challenging cases, including self-occlusions, truncations, and extreme lighting conditions. Additionally, we compare the model performance qualitatively on in-the-wild lab images and quantitatively under various levels of truncation severity, illustrating the superior generalization ability and robustness of our approach. Please refer to the appendix for more detailed results.", + "bbox": [ + 212, + 652, + 789, + 791 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Efficiency Furthermore, we benchmark the computation efficiency of our method and compare it with previous works. We measure the average inference time,", + "bbox": [ + 212, + 809, + 792, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "S. Ban et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/8637dd38ae82d1780bbd07474339ee3f3025b679262675d32c873284ab750df1.jpg", + "table_caption": [ + "Table 2: Comparison of computation efficiency on the Panda 3CAM-AK dataset. † denotes using ground-truth joint state parameters." + ], + "table_footnote": [], + "table_body": "
MethodAUC ↑Time (ms) ↓FPS ↑FLOPS (G) ↓
† DREAM-F 1868.966.015.190.5
RoboPose 1570.4570.81.857.3
Ours82.244.322.625.2
", + "bbox": [ + 333, + 196, + 674, + 260 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "frame per second (FPS), and Floating Point Operations (FLOPs). For a fair comparison, we benchmark all methods on a Linux machine with CPU Intel(R) Xeon(R) Gold 6240 CPU @ 2.60GHz and GPU NVIDIA Tesla V100S. Batch size is set to be 1 for all methods. Table 2 shows that our method exhibits a remarkable advantage in terms of inference speed and computation loads, enabling real-time inference while achieving state-of-the-art accuracy. This advantage stems from the fact that our method involves only a single feed-forward pass without costly iterative optimization during inference. We demonstrate that our approach serves as an efficient and effective solution for real-time holistic robot pose estimation, which holds great potential in real-world applications.", + "bbox": [ + 212, + 287, + 787, + 441 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Comparison under known joints states", + "text_level": 1, + "bbox": [ + 215, + 463, + 578, + 479 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Although the primary focus of this work is holistic robot pose estimation with unknown states, our approach can also be easily adapted to the known joint states setting. We conduct additional experiments on four real datasets for robot pose estimation with ground-truth joint states available. We simply replace the predicted joint state with the ground-truth joint state during the self-supervised training process. As shown in Tab. 3 our approach also achieves comparable performance with previous SOTA under this well-explored setting.", + "bbox": [ + 212, + 488, + 787, + 595 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.5 Ablation study", + "text_level": 1, + "bbox": [ + 215, + 617, + 385, + 633 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ablation on network modules We first conduct an ablation study on the proposed network modules to investigate their effectiveness. Since JointNet and RotationNet are the major regression targets of our task, we focus on the ablation of two assistive network modules, DepthNet and KeypointNet. We conduct the experiments on the synthetic datasets, utilizing only ground-truth supervision to simplify the process.", + "bbox": [ + 212, + 643, + 787, + 734 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Firstly, we delve into the implementation of DepthNet. In Tab. 4, we evaluate the design choice of whether to share the backbone of DepthNet with other modules or not. The results indicate that separating the DepthNet backbone consistently improves performance. This improvement could be explained by the fact that DepthNet involves learning scale-variant features, while the other network modules all focus on learning scale-invariant features. We also investigate whether to pre-train the DepthNet or jointly train it with other modules", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Holistic Robot Pose", + "bbox": [ + 596, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/9f874de395080ec4b75c1d1634dec1741bb915b2f3e63142c721a5f1b3bb5382.jpg", + "table_caption": [ + "Table 3: Comparison of AUC $\\uparrow$ of the ADD distribution curve and Mean $\\downarrow$ of the ADD, on the real-world datasets of DREAM, with all methods using known ground-truth joint state parameters." + ], + "table_footnote": [], + "table_body": "
MethodMetricPanda 3CAM-AKPanda 3CAM-XKPanda 3CAM-RSPanda ORB
DREAM-F18AUC↑68.924.476.161.9
DREAM-Q18AUC↑52.437.578.057.1
DREAM-H18AUC↑60.564.078.869.1
RoboPose15AUC↑76.586.076.980.5
CtRNet22AUC↑89.979.590.885.3
SGTAPose40AUC↑67.82.187.672.3
OursAUC↑90.281.291.987.6
DREAM-F18Mean↓11413.1491911.42077.495319.1
DREAM-Q18Mean↓78089.354178.227.264247.6
DREAM-H18Mean↓56.57381.623.625685.3
RoboPose15Mean↓24.214.023.119.4
CtRNet22Mean↓13.032.010.021.0
SGTAPose40Mean↓35.7157.712.835.0
OursMean↓12.919.911.015.9
", + "bbox": [ + 233, + 210, + 772, + 412 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/760208ca9cfdffeb13012641e11666047dcfcbc0770942ae5531ae93cbc201e9.jpg", + "table_caption": [ + "Table 4: Ablation studies on the effect of separating DepthNet image feature extractor." + ], + "table_footnote": [], + "table_body": "
DatasetSeparateAUC ↑Mean (mm)↓
Panda DR (a)-79.9821.9
Panda DR (b)+82.6818.0
Panda Photo (a)-79.6622.0
Panda Photo (b)+81.9818.4
", + "bbox": [ + 220, + 496, + 486, + 577 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/494b3ed769b728038574cbafce56a5b0abbd174e18703b52d91708584373d038.jpg", + "table_caption": [ + "Table 5: Ablation studies on the effect of pre-training DepthNet in advance of jointly training." + ], + "table_footnote": [], + "table_body": "
DatasetPretrainAUC ↑Mean (mm)↓
Panda DR (a)-74.2926.8
Panda DR (b)+82.6818.0
Panda Photo (a)-73.3427.9
Panda Photo (b)+81.9818.4
", + "bbox": [ + 519, + 496, + 785, + 577 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "from scratch. Table 5 suggests that a pre-training stage leads to considerable performance improvement. This improvement likely occurs because a well-trained DepthNet reliably aids other modules in better training, as it determines the global translation of all joints. Additionally, we select the keypoint closest to the robot center as the root keypoint requiring depth prediction by DepthNet. Please refer to the S2.3 section of the appendix for rationale behind this selection.", + "bbox": [ + 212, + 609, + 787, + 700 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Moreover, we conduct ablation experiments to investigate the impact of KeypointNet on other network modules, namely JointNet and RotationNet. We utilize Mean Joint Error, which represents the mean error of the joint state parameter predictions, and the Mean Rotation Angle Error [15], which represents the mean error of the Euler angle predictions $(\\theta_{x}^{pred}, \\theta_{y}^{pred}, \\theta_{z}^{pred})$ for the camera-to-robot rotation, to quantify the accuracy of the JointNet and RotationNet predictions, respectively. It is important to note that a Panda robot consists of seven revolute joints with angle parameters (representing one-DoF rotations) and one prismatic joint with a length parameter (representing one-DoF transla", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "S. Ban et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/19d02b552fdc9ecf9b92395bac0cc91b86a44add76ceb99ba1f0b9e4674624f6.jpg", + "image_caption": [ + "Fig. 4: Qualitative comparison between our method and RoboPose [15] on both real and synthetic datasets." + ], + "image_footnote": [], + "bbox": [ + 246, + 146, + 748, + 441 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "tion). Therefore, we calculate and present the Mean Joint Error separately for the revolute and prismatic joints of Panda. We present the detailed ablation results in the S2.1 section of the appendix, in which Tables 6 and 7 show that the presence of KeypointNet significantly improves the prediction accuracy of JointNet and RotationNet. By sharing a feature extractor with JointNet and RotationNet, KeypointNet guides the network towards learning pixel-aligned robot visual features, which further assists in estimating robot pose and joint states.", + "bbox": [ + 212, + 506, + 787, + 611 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Finally, we examine the overall effect of DepthNet and KeypointNet. To replace DepthNet, we implement a straightforward alternative, directly regressing the camera-to-robot translation $\\mathbf{t}$ from image features $\\mathbf{f}$ using an MLP, similar to RotationNet. Figure 5a illustrates that both modules play an important role in prediction accuracy. The DepthNet notably enhances accuracy by improving root depth estimation, a global offset that affects the precision of all joints. Meanwhile, KeypointNet further benefits the estimation results.", + "bbox": [ + 212, + 612, + 787, + 717 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The contribution of these modules to the overall performance validates our design. DepthNet disentangles depth from camera intrinsic, facilitating a learning-based approach to depth estimation. This method exhibits greater expressive capability compared to the analytical, non-parameterized PnP methods employed in previous work [18,22,40]. On the other hand, KeypointNet enhances both JointNet and RotationNet by steering the shared backbone towards learning more pixel-aligned features, thus complementing the learning objectives defined in the robot kinematic space.", + "bbox": [ + 212, + 719, + 795, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Holistic Robot Pose", + "bbox": [ + 596, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation on self-supervision We also perform an ablation study of our self-supervised learning objectives. We conduct the experiments on the Panda real datasets. Figure 5b demonstrates that both the keypoint consistency regularization and the mask consistency regularization contribute to the overall performance. Specifically, the mask consistency regularization provides pixel-wise supervision signal with the help of differentiable rendering, while the keypoint consistency regularization unifies the predictions in the pixel space and that in the robot kinematic space. We suppose that pixel-aligned 2D tasks, e.g. keypoint detection and mask segmentation, generally generalize better, and could therefore improve the 3D estimation quality through consistency regularization, further improving the overall performance. As a result, our multi-stage training scheme yields favorable convergence ability. Please refer to the S4 section of the appendix for the error analysis of each module across each training stage.", + "bbox": [ + 212, + 146, + 787, + 344 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/719af2c204af9ac0f711ce0a150a758f556cbdee2843edde3af6a75bcc74a3c9.jpg", + "image_caption": [ + "(a) Ablation study on network modules. 'Baseline' represents direct regression, 'DN' and 'KN' refer to DepthNet and KeypointNet, respectively.", + "Fig. 5: Ablation studies on network modules and self-supervision." + ], + "image_footnote": [], + "bbox": [ + 222, + 376, + 480, + 506 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/e9262e5a7a8d0f0543a2c87f1cd39756565c3c08b6a00917d201889743b3b883.jpg", + "image_caption": [ + "(b) Ablation study on self-supervision. 'Baseline' represents no self-supervision, 'MC' and 'KC' refer to mask and keypoint consistency regularization, respectively." + ], + "image_footnote": [], + "bbox": [ + 521, + 376, + 779, + 506 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 646, + 359, + 662 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work presents a novel end-to-end framework for real-time holistic robot pose estimation without prior knowledge of internal robot states, providing substantial assistance in real-world scenarios where the joint state of the estimation target is unavailable or unreliable. By decomposing this complex task into specific sub-tasks and tailoring neural network modules for each, the method effectively addresses the challenges while retaining high computation efficiency. Our approach, which circumvents iterative optimization, significantly accelerates the estimation process and achieves state-of-the-art accuracy. This advancement holds promise for enhancing real-world robotics applications, offering a practical solution for real-time robot pose estimation and joint state monitoring.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "S. Ban et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is supported by National Science and Technology Major Project (2022ZD0114904). We thank Hai Ci for the valuable discussions during the early stages of this work. We thank Haoran Lu and Ruihai Wu for their assistance with real-world image captures.", + "bbox": [ + 215, + 172, + 785, + 233 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 253, + 321, + 270 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bultmann, S., Memmesheimer, R., Behnke, S.: External camera-based mobile robot pose estimation for collaborative perception with smart edge sensors. In: ICRA (2023)", + "2. Chen, L., Qin, Y., Zhou, X., Su, H.: EasyHeC: Accurate and Automatic Hand-eye Calibration via Differentiable Rendering and Space Exploration. In: IEEE Robotics and Automation Letters (2023)", + "3. Christen, S., Yang, W., Pérez-D'Arpino, C., Hilliges, O., Fox, D., Chao, Y.W.: Learning human-to-robot handovers from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9654–9664 (2023)", + "4. Deisenroth, M.P., Rasmussen, C.E., Fox, D.: Learning to control a low-cost manipulator using data-efficient reinforcement learning. Robotics: Science and Systems VII 7, 57-64 (2011)", + "5. Feniello, A., Dang, H.N., Birchfield, S.: Program synthesis by examples for object repositioning tasks. 2014 IEEE/RSJ International Conference on Intelligent Robots and Systems pp. 4428-4435 (2014)", + "6. Fiala, M.: Artag, a fiducial marker system using digital techniques. In: CVPR (2005)", + "7. Garrido-Jurado, S., Muñoz-Salinas, R., Madrid-Cuevas, F.J., Marín-Jiménez, M.J.: Automatic generation and detection of highly reliable fiducial markers under occlusion. Pattern Recognition (2014)", + "8. Hao, R., Özgün, O., Çavusoglu, M.C.: Vision-based surgical tool pose estimation for the da Vinci® robotic surgical system. In: 2018 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 1298-1305. IEEE (2018)", + "9. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)", + "0. Icli, C., Stepanenko, O., Bonev, I.: New method and portable measurement device for the calibration of industrial robots. Sensors 20(20) (2020)", + "1. Kim, W., Kim, N., Lyons, J.B., Nam, C.S.: Factors affecting trust in high-vulnerability human-robot interaction contexts: A structural equation modelling approach. Applied Ergonomics 85, 103056 (2020)", + "2. Kok, B.C., Soh, H.: Trust in robots: Challenges and opportunities. Current Robotics Reports 1, 297-309 (2020)", + "3. Kothari, N., Gupta, M., Vachhani, L., Arya, H.: Pose estimation for an autonomous vehicle using monocular vision. In: 2017 Indian Control Conference (ICC). pp. 424-431 (2017)", + "4. Kuipers, B.: How can we trust a robot? Communications of the ACM 61(3), 86-95 (2018)", + "5. Labbé, Y., Carpentier, J., Aubry, M., Sivic, J.: Single-view robot pose and joint angle estimation via render & compare. In: Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + ], + "bbox": [ + 225, + 282, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Holistic Robot Pose", + "bbox": [ + 596, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Lambrecht, J., Grosenick, P., Meusel, M.: Optimizing keypoint-based single-shot camera-to-robot pose estimation through shape segmentation. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 13843-13849. IEEE (2021)", + "17. Lambrecht, J., Kästner, L.: Towards the usage of synthetic data for marker-less pose estimation of articulated robots in rgb images. In: ICAR (2019)", + "18. Lee, T.E., Tremblay, J., To, T., Cheng, J., Mosier, T., Kroemer, O., Fox, D., Birchfield, S.: Camera-to-robot pose estimation from a single image. In: International Conference on Robotics and Automation (ICRA) (2020)", + "19. Lepetit, V., Moreno-Noguer, F., Fua, P.: Epnp: An accurate o (n) solution to the pnp problem. IJCV (2009)", + "20. Li, S., De Wagter, C., De Croon, G.C.: Self-supervised monocular multi-robot relative localization with efficient deep neural networks. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 9689-9695. IEEE (2022)", + "21. Li, X., Zhang, X., Zhou, X., Chen, I.M.: Upg: 3d vision-based prediction framework for robotic grasping in multi-object scenes. Knowledge-Based Systems 270, 110491 (2023)", + "22. Lu, J., Richter, F., Yip, M.C.: Markerless camera-to-robot pose estimation via self-supervised sim-to-real transfer. In: Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "23. Mahler, J., Matl, M., Satish, V., Danielczuk, M., DeRose, B., McKinley, S., Goldberg, K.: Learning ambidextrous robot grasping policies. Science Robotics 4(26), eaau4984 (2019)", + "24. Masehian, E., Jannati, M., Hekmatfar, T.: Cooperative mapping of unknown environments by multiple heterogeneous mobile robots with limited sensing. Robotics and Autonomous Systems 87, 188-218 (2017)", + "25. Ming, Y., Meng, X., Fan, C., Yu, H.: Deep learning for monocular depth estimation: A review. Neurocomputing 438, 14-33 (2021)", + "26. Moon, G., Chang, J., Lee, K.M.: Camera distance-aware top-down approach for 3d multi-person pose estimation from a single rgb image. In: The IEEE Conference on International Conference on Computer Vision (ICCV) (2019)", + "27. Morrison, D., Corke, P., Leitner, J.: Closing the Loop for Robotic Grasping: A Real-time, Generative Grasp Synthesis Approach. In: Proc. of Robotics: Science and Systems (RSS) (2018)", + "28. Olson, E.: Apriltag: A robust and flexible visual fiducial system. In: ICRA (2011)", + "29. Papadimitriou, A., Mansouri, S.S., Nikolakopoulos, G.: Range-aided ego-centric collaborative pose estimation for multiple robots. Expert Systems with Applications 202, 117052 (2022)", + "30. Park, D., Seo, Y., Chun, S.Y.: Real-time, highly accurate robotic grasp detection using fully convolutional neural networks with high-resolution images (2019)", + "31. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch (2017)", + "32. Pinto, L., Davidson, J., Gupta, A.: Supervision via competition: Robot adversaries for learning tasks. In: 2017 IEEE International Conference on Robotics and Automation (ICRA). pp. 1601-1608 (2017)", + "33. Prusak, A., Melnychuk, O., Roth, H., Schiller, I., Koch, R.: Pose estimation and map building with a time-of-flight-camera for robot navigation. International Journal of Intelligent Systems Technologies and Applications 5(3-4), 355-364 (2008)", + "34. Rana, A., Vulpi, F., Galati, R., Milella, A., Petitti, A.: A pose estimation algorithm for agricultural mobile robots using an rgb-d camera. In: 2022 International Con-" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "S. Ban et al.", + "bbox": [ + 271, + 114, + 356, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ference on Electrical, Computer, Communications and Mechatronics Engineering (ICECCME). pp. 1-5 (2022)", + "35. Rizk, Y., Awad, M., Tunstel, E.W.: Cooperative heterogeneous multi-robot systems. ACM Computing Surveys (CSUR) 52, 1 - 31 (2019)", + "36. Salvini, P., Paez-Granados, D.F., Billard, A.: Safety concerns emerging from robots navigating in crowded pedestrian areas. International Journal of Social Robotics 14, 441 - 462 (2021)", + "37. Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learning for human pose estimation. In: CVPR (2019)", + "38. Sun, X., Xiao, B., Wei, F., Liang, S., Wei, Y.: Integral human pose regression. In: Proceedings of the European Conference on Computer Vision (ECCV) (2018)", + "39. Svenstrup, M., Tranberg, S., Andersen, H.J., Bak, T.: Pose estimation and adaptive robot behaviour for human-robot interaction. In: 2009 IEEE International Conference on Robotics and Automation. pp. 3571-3576 (2009)", + "40. Tian, Y., Zhang, J., Yin, Z., Dong, H.: Robot structure prior guided temporal attention for camera-to-robot pose estimation from image sequence. In: Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "41. Tzafestas, S.G.: 15 - mobile robots at work. In: Tzafestas, S.G. (ed.) Introduction to Mobile Robot Control, pp. 635-663. Elsevier, Oxford (2014)", + "42. Uchibe, E., Asada, M., Noda, S., Takahashi, Y., Hosoda, K.: Vision-based reinforcement learning for robocup: Towards real robot competition. In: Proc. of IROS. vol. 96 (1996)", + "43. Xia, Y., Tang, Y., Hu, Y., Hoffman, G.: Privacy-preserving pose estimation for human-robot interaction. CoRR abs/2011.07387 (2020)", + "44. Xu, C., Yu, X., Wang, Z., Ou, L.: Multi-view human pose estimation in humanrobot interaction. In: IECON 2020 The 46th Annual Conference of the IEEE Industrial Electronics Society. pp. 4769-4775 (2020)", + "45. Xun, Z., Huang, J., Li, Z., Xu, C., Gao, F., Cao, Y.: Crepes: Cooperative relative pose estimation towards real-world multi-robot systems. arXiv preprint arXiv:2302.01036 (2023)", + "46. Yang, W., Paxton, C., Mousavian, A., Chao, Y.W., Cakmak, M., Fox, D.: Reactive human-to-robot handovers of arbitrary objects. In: IEEE International Conference on Robotics and Automation (ICRA). IEEE (2021)", + "47. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5745-5753 (2019)", + "48. Zhu, W., Ma, X., Ro, D., Ci, H., Zhang, J., Shi, J., Gao, F., Tian, Q., Wang, Y.: Human motion generation: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)", + "49. Zuo, Y., Qiu, W., Xie, L., Zhong, F., Wang, Y., Yuille, A.L.: Craves: Controlling robotic arm with a vision-based economic system. In: CVPR (2019)" + ], + "bbox": [ + 215, + 147, + 785, + 715 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Holistic Robot Pose", + "bbox": [ + 596, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Real-time Holistic Robot Pose Estimation with Unknown States/0b349d89-54a2-48da-a5a9-0c888f46a6ff_model.json b/2024/Real-time Holistic Robot Pose Estimation with Unknown States/0b349d89-54a2-48da-a5a9-0c888f46a6ff_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0d827fc08cb28ad65bd879c4520a97c3404c3fe6 --- /dev/null +++ b/2024/Real-time Holistic Robot Pose Estimation with Unknown States/0b349d89-54a2-48da-a5a9-0c888f46a6ff_model.json @@ -0,0 +1,2368 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.233, + 0.141, + 0.773, + 0.185 + ], + "angle": 0, + "content": "Real-time Holistic Robot Pose Estimation with Unknown States" + }, + { + "type": "text", + "bbox": [ + 0.294, + 0.211, + 0.711, + 0.245 + ], + "angle": 0, + "content": "Shikun Ban1, Juling Fan1, Xiaoxuan Ma1, Wentao Zhu1,2, Yu Qiao5, and Yizhou Wang1,2,3,4" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.254, + 0.747, + 0.283 + ], + "angle": 0, + "content": "1 Center on Frontiers of Computing Studies, School of Computer Science, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.332, + 0.283, + 0.672, + 0.298 + ], + "angle": 0, + "content": "2 Inst. for Artificial Intelligence, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.332, + 0.297, + 0.672, + 0.311 + ], + "angle": 0, + "content": "3 Nat'l Eng. Research Center of Visual Technology" + }, + { + "type": "text", + "bbox": [ + 0.332, + 0.311, + 0.666, + 0.325 + ], + "angle": 0, + "content": "4 Nat'l Key Lab of General Artificial Intelligence" + }, + { + "type": "list", + "bbox": [ + 0.255, + 0.254, + 0.747, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.237, + 0.325, + 0.766, + 0.353 + ], + "angle": 0, + "content": "\\(^{5}\\) Department of Automation, School of Electronics, Information and Electrical Engineering, Shanghai Jiao Tong University" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.387, + 0.741, + 0.652 + ], + "angle": 0, + "content": "Abstract. Estimating robot pose from RGB images is a crucial problem in computer vision and robotics. While previous methods have achieved promising performance, most of them presume full knowledge of robot internal states, e.g. ground-truth robot joint angles. However, this assumption is not always valid in practical situations. In real-world applications such as multi-robot collaboration or human-robot interaction, the robot joint states might not be shared or could be unreliable. On the other hand, existing approaches that estimate robot pose without joint state priors suffer from heavy computation burdens and thus cannot support real-time applications. This work introduces an efficient framework for real-time robot pose estimation from RGB images without requiring known robot states. Our method estimates camera-to-robot rotation, robot state parameters, keypoint locations, and root depth, employing a neural network module for each task to facilitate learning and sim-to-real transfer. Notably, it achieves inference in a single feed-forward pass without iterative optimization. Our approach offers a \\(12 \\times\\) speed increase with state-of-the-art accuracy, enabling real-time holistic robot pose estimation for the first time. Code and models are available at https://github.com/Oliverbansk/Holistic-Robot-Pose-Estimation" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.697, + 0.376, + 0.712 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.727, + 0.788, + 0.806 + ], + "angle": 0, + "content": "Robot pose estimation is a fundamental task in computer vision and robotics as it not only lays the groundwork for multiple downstream tasks, e.g. visually-guided object grasping and manipulation [21,40,49], but also paves the way for a variety of applications, such as autonomous navigation [1,13,33,34], human-robot interaction [3,39,43,44,46,48] and multi-robot collaboration [20,29,35,45]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.23, + 0.811, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Correspondence to: Wentao Zhu and Yu Qiao ." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "S. Ban et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.148, + 0.782, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.295, + 0.789, + 0.406 + ], + "angle": 0, + "content": "Fig. 1: The majority of previous robot pose estimation methods assume known robot joint states and focus solely on estimating the camera-to-robot pose, i.e. camera-to-robot rotation and translation. In contrast, the holistic robot pose estimation problem requires estimating both joint states and the camera-to-robot pose, given only an RGB image without known joint states. For holistic robot pose estimation, RoboPose [15] uses costly test-time optimization (Render-and-Compare) to iteratively refine the predictions. In contrast, our feed-forward method achieves state-of-the-art accuracy with a \\(12 \\times\\) speed boost." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.789, + 0.492 + ], + "angle": 0, + "content": "Nonetheless, most previous works for monocular robot pose estimation [17,18,22,40] rely on a known set of robot joint angles as an additional input to alleviate the difficulty of the estimation process. Yet, this prerequisite information of the robot state is not always available in practical scenarios:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.499, + 0.787, + 0.546 + ], + "angle": 0, + "content": "- For instance, low-cost robotic arms without sophisticated sensors or internal monitors can not provide the visual calibration process with readily accessible robot state measurements [4,10,49]." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.546, + 0.788, + 0.605 + ], + "angle": 0, + "content": "- In the context of multi-robot collaboration, the real-time synchronization of joint state data may be hindered by environmental limitations [24], necessitating pose estimation without robot joint state information, and additional monitoring of the joint states." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.605, + 0.788, + 0.682 + ], + "angle": 0, + "content": "- In domains such as human-robot interaction [36,41] and competitive robot gaming [32,42], the reliability of the state information from counterpart robots is not always guaranteed, as trust issues may arise in robotic systems [11,12,14]. In this sense, failing to obtain independent estimation of the counterpart robot joint state may lead to potential safety issues." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.499, + 0.788, + 0.682 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.789, + 0.796 + ], + "angle": 0, + "content": "These scenarios further emphasize the urgent need for monocular robot pose estimation with unknown robot joint state. In this work, we address the challenge of Holistic Robot Pose Estimation with unknown internal states for articulated robots. Specifically, given a monocular RGB image, we aim to simultaneously solve for 6D robot pose (3D rotation and translation relative to the camera) and fine-grained robot joint state parameters (including joint angles for revolute joints and displacement for prismatic joints)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Nonetheless, learning holistic robot pose estimation without joint state information presents significant challenges. The primary challenge is the dual ambiguity of unknown joint states and 6D pose. Specifically, estimating the 6D pose" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.597, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Holistic Robot Pose" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.299 + ], + "angle": 0, + "content": "of an articulated robot is more complex than for a rigid object due to shape and scale variability from unknown joint states. Conversely, estimating joint states is complicated by the need for 2D-to-3D correspondences and the unknown 6D pose. This intertwined issue requires novel solutions to decouple joint states from 6D pose estimation. Furthermore, the sophisticated robot morphologies vary from model to model, requiring non-trivial model design. In addition, The robot parts are frequently invisible in the monocular image as a result of self-occlusion and truncation, posing more challenges for robust estimation. Finally, the domain gap between synthetic training data and real-world images introduces additional challenges for learning-based approaches." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.301, + 0.788, + 0.422 + ], + "angle": 0, + "content": "It is worth noting that some previous works [15,49] have also addressed holistic estimation with unknown states, tackling some challenges. However, these methods all necessitate iterative optimization at inference, which is computationally intensive and unsuitable for real-time applications. Motivated by the need for efficient holistic robot pose estimation without knowledge of the robot state, we propose an end-to-end framework that directly estimates camera-to-robot pose and robot joint state parameters, requiring only a single feed-forward pass during inference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.426, + 0.788, + 0.562 + ], + "angle": 0, + "content": "In pursuit of this goal, we factorize the holistic robot pose estimation task into several sub-tasks, including the estimation of camera-to-robot rotation, robot joint states, root depth, and root-relative keypoint locations. We then design corresponding neural network modules for each sub-task. Specifically, the standalone root depth estimator disentangles the estimation of scale-variant and scale-invariant variables. The keypoint estimator bridges the gap between the pixel space and the robotic parameter space. Additionally, we utilize self-supervised consistency regularizations which could be performed on the real images without ground truth (GT)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.788, + 0.701 + ], + "angle": 0, + "content": "We demonstrate that our designs of modular feed-forward networks improve model generalization while maintaining high computation efficiency. Consequently, our approach achieves state-of-the-art estimation accuracy with unknown robot states and a \\(12 \\times\\) speedup compared to the iterative Render-and-Compare (RnC) [15] approach, as depicted in Fig. 1. The key insight behind our approach is that we could get rid of the cumbersome RnC [15] and error-prone Perspective-n-Points (PnP) [18,22,40] schemes to establish 2D-3D correspondences. Instead, we demonstrate that feed-forward networks could well handle the problem with proper problem decomposition and meticulous design." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.705, + 0.568, + 0.719 + ], + "angle": 0, + "content": "In summary, our contributions are three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.743, + 0.784, + 0.771 + ], + "angle": 0, + "content": "1. We present an efficient end-to-end learning framework for the holistic robot pose estimation problem with unknown robot states." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.777, + 0.784, + 0.804 + ], + "angle": 0, + "content": "2. We propose to factorize and decompose the holistic robot pose estimation problem into several sub-tasks and tackle them with respective modules." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.81, + 0.784, + 0.838 + ], + "angle": 0, + "content": "3. We achieve state-of-the-art performance on various robot models while delivering a significant \\(12 \\times\\) speedup, paving the way for real-time applications." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.743, + 0.784, + 0.838 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "S. Ban et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.389, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.172, + 0.436, + 0.187 + ], + "angle": 0, + "content": "2.1 Hand-eye calibration" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.789, + 0.39 + ], + "angle": 0, + "content": "The goal of Hand-Eye Calibration (HEC) is to determine the transformation between the position of a robot's end-effector (hand) and its vision system (eye). This task can be categorized into two types: eye-in-hand where the camera is fixed at the robot end-effector [27], and eye-to-hand which refers to the configurations where the camera is placed at a distance from the robot [5,23,30]. This work focuses on a similar setting with eye-to-hand. Traditional HEC methods involve attaching a fiducial marker to the end effector [6,7,28], placing it in several known positions, detecting the marker positions in images, and then solving for the camera-to-robot transformation. Despite its broad applicability, the classical HEC pipeline necessitates a complete offline calibration process every time the camera-robot transformation changes, which can be tedious in real-world applications. Consequently, recent works, such as DREAM [18], propose an online, markerless calibration problem, as discussed below." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.407, + 0.553, + 0.423 + ], + "angle": 0, + "content": "2.2 Image-based robot pose estimation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.427, + 0.789, + 0.592 + ], + "angle": 0, + "content": "Previous works can be divided into two categories based on whether the robot states are known or not. Most previous works estimate the 6D camera-to-robot pose with known robot states. For example, DREAM [18] first trains a 2D keypoint estimator and then applies Perspective-\\(n\\)-Point (PnP) [19] to obtain camera-to-robot pose, showing comparable results with the classical methods. Some work further utilizes segmentation masks and edges to aid pose estimation [8,16]. Chen et al. [2] apply rendering-based camera pose optimization and consistency-based joint space exploration. SGTAPose [40] leverages robot structure prior and temporal information. Lu et al. [22] propose to train on real-world data using foreground segmentation and differentiable rendering in a self-supervised manner." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.79, + 0.744 + ], + "angle": 0, + "content": "Recently, several works have shifted to estimating robot pose with unknown joint states. Zuo et al. [49] estimate the robot pose and joint angles of a simple 4-DoF robotic arm from a single RGB image by solving a nonlinear nonconvex optimization problem. However, this method is not well-suited for robot arms with higher complexity. Labbé et al. [15] propose a render-and-compare approach to iteratively refine the estimation for both known and unknown states, and apply the method to different robot models up to 15-DoF. However, the iterative optimization process limits its applications in real-time scenarios. In contrast, our approach can be deployed with a single feed-forward at inference and achieves real-time efficiency." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.763, + 0.331, + 0.779 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.79, + 0.342, + 0.804 + ], + "angle": 0, + "content": "3.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Given an RGB image \\(\\mathbf{I}\\), we employ a suite of neural network modules, including DepthNet, JointNet, RotationNet, and KeypointNet. These modules are used" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.598, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Holistic Robot Pose" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.146, + 0.787, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.373, + 0.788, + 0.456 + ], + "angle": 0, + "content": "Fig. 2: Framework overview. The JointNet and the RotationNet regress joint state parameters \\(\\mathbf{q}\\) and camera-to-robot rotation \\(\\mathbf{R}\\), respectively. The KeypointNet estimates root-relative 3D keypoint locations \\(\\mathbf{P}^r\\). The DepthNet's estimation of root depth \\(d\\) is combined with \\(\\mathbf{P}^r\\) to acquire absolute 3D keypoint locations \\(\\mathbf{P}'\\) and camera-to-robot translation \\(\\mathbf{t}\\). Joint state parameters \\(\\mathbf{q}\\), rotation \\(\\mathbf{R}\\) and translation \\(\\mathbf{t}\\) are used to compute 3D keypoint locations \\(\\mathbf{P}\\) via forward-kinematics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.497, + 0.788, + 0.604 + ], + "angle": 0, + "content": "to respectively estimate the absolute root depth \\(d\\), joint state parameters \\(\\mathbf{q}\\), camera-to-robot rotation \\(\\mathbf{R} \\in SO(3)\\), and root-relative 3D keypoint locations \\(\\mathbf{P}^r \\in \\mathbb{R}^{N \\times 3}\\), where \\(N\\) represents the total number of keypoints. The keypoints are defined at the joint articulations, in line with previous works [15,18,22], and one keypoint is designated as the robot root. The root joint is critical for computing the camera-to-robot rotation \\(\\mathbf{R}\\) and translation \\(\\mathbf{t}\\), as well as serving as the origin for the forward kinematics (FK) process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.609, + 0.788, + 0.746 + ], + "angle": 0, + "content": "Our modularized design offers several key insights. Firstly, DepthNet is tailored to learn the scale-variant camera-robot spatial relationship, while the other modules concentrate on the scale-invariant robot features. Secondly, while the joint state parameters \\(\\mathbf{q}\\) provide a compact representation of robot morphologies subject to robotic structure constraints, they are highly abstract and thus challenging to directly regress from image features. To address this, we utilize the 3D keypoint locations as an intermediary. On the one hand, they can be effectively estimated with pixel-aligned image features; on the other hand, they are inherently linked to \\(\\mathbf{q}\\) through an FK process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "An overview of the proposed framework is illustrated in Fig. 2. Specifically, DepthNet estimates the absolute root depth \\( d \\), which, combined with KeypointNet's output \\( \\mathbf{P}^r \\in \\mathbb{R}^{N \\times 3} \\), provides the absolute 3D keypoint locations \\( \\mathbf{P}' \\in \\mathbb{R}^{N \\times 3} \\). The root keypoint location in \\( \\mathbf{P}' \\) is the camera-to-robot translation \\( \\mathbf{t} \\). The estimated joint state parameters \\( \\mathbf{q} \\), along with the camera-to-robot rotation \\( \\mathbf{R} \\) and translation \\( \\mathbf{t} \\), are then used to perform FK, yielding the final" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "S. Ban et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.176 + ], + "angle": 0, + "content": "3D keypoint estimations \\(\\mathbf{P} \\in \\mathbb{R}^{N \\times 3}\\). Please refer to the appendix for a detailed illustration of the relations of these variables." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.226 + ], + "angle": 0, + "content": "In the following, we will first introduce the architecture design of the network modules in Sec. 3.2, and then explain the training loss design consisting of ground-truth supervision and self-supervision in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.249, + 0.368, + 0.263 + ], + "angle": 0, + "content": "3.2 Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.277, + 0.789, + 0.399 + ], + "angle": 0, + "content": "DepthNet Monocular object depth estimation suffers from significant ambiguities, as the absolute depth value and camera intrinsics are deeply intertwined [25]. For instance, a robot in a fixed pose relative to the camera may appear larger in photos taken with cameras having larger focal lengths. Scaling both the absolute depth and focal lengths equally results in identical object size and appearance in different images. To address this challenge, we implement DepthNet following the approach in [26] to disentangle depth from camera intrinsics. Initially, we use camera intrinsics to calculate a normalized, coarse depth value \\( d_{c} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.425, + 0.412, + 0.786, + 0.445 + ], + "angle": 0, + "content": "\\[\nd _ {c} = \\sqrt {\\frac {f _ {x} \\cdot f _ {y} \\cdot A _ {r e a l}}{A _ {b b o x}}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.788, + 0.562 + ], + "angle": 0, + "content": "where \\( f_{x} \\) and \\( f_{y} \\) represent the focal lengths along the x- and y-axes, \\( A_{real} \\) is the area of the robot in the x-y plane of the camera frame (in \\( mm^2 \\)), and \\( A_{bbox} \\) is the area of the robot's bounding box (in \\( pixel^2 \\)). This method of calculating \\( d_{c} \\) normalizes it in 3D space and provides an approximate depth estimation for subsequent refinement. The primary function of DepthNet is to predict a correction factor \\( \\lambda \\) based on the input image. We multiply the estimated \\( \\lambda \\) by \\( d_{c} \\) to obtain the final absolute root depth \\( d \\):" + }, + { + "type": "equation", + "bbox": [ + 0.467, + 0.576, + 0.785, + 0.59 + ], + "angle": 0, + "content": "\\[\nd = \\lambda \\cdot d _ {c}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.603, + 0.788, + 0.695 + ], + "angle": 0, + "content": "With a fixed camera, the relationship between the normalized coarse depth \\( d_{c} \\) and the robot's size in the image is linear. However, robots can assume poses with the same depth but different sizes in the image, making \\( d_{c} \\) inadequate for accurate root depth representation. The correction factor \\( \\lambda \\) introduces non-linear flexibility to the depth estimation, addressing this issue by adjusting \\( d_{c} \\) based on image features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.84 + ], + "angle": 0, + "content": "JointNet The goal of JointNet is to estimate joint state parameters \\(\\mathbf{q} = (q_{1}, q_{2}, \\dots, q_{J})\\), where \\(J\\) denotes the number of robot joints. As this work deals with 1-DoF joints (revolute or prismatic), \\(J\\) equals the robot's DoF. JointNet shares the same feature extractor with RotationNet and KeypointNet. We extract feature map \\(\\mathbf{F} \\in \\mathbb{R}^{C \\times H \\times W}\\) (where \\(C\\) denotes the number of channels, and \\(H \\times W\\) denotes the spatial resolution) and use global average pooling to obtain \\(\\mathbf{f} \\in \\mathbb{R}^{C}\\). Finally, we employ several MLP blocks to regress the joint state parameters \\(\\mathbf{q}\\) from \\(\\mathbf{f}\\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.597, + 0.115, + 0.733, + 0.127 + ], + "angle": 0, + "content": "Holistic Robot Pose" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.196 + ], + "angle": 0, + "content": "RotationNet The goal of RotationNet is to estimate the camera-to-robot rotation \\(\\mathbf{R} \\in SO(3)\\). We employ an MLP to estimate \\(\\mathbf{R}\\) from the image feature \\(\\mathbf{f}\\). We utilize the 6D continuous rotation representation [47] to parameterize \\(\\mathbf{R}\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.211, + 0.789, + 0.318 + ], + "angle": 0, + "content": "KeypointNet The goal of KeypointNet is to estimate root-relative 3D keypoint locations \\(\\mathbf{P}^r\\in \\mathbb{R}^{N\\times 3}\\) for \\(N\\) robot keypoints. We transform the feature map \\(\\mathbf{F}\\) to a 3D heatmap \\(\\mathbf{H}\\in \\mathbb{R}^{N\\times D\\times H'\\times W'}\\) through convolutional layers, where \\(D\\times H'\\times W'\\) is the spatial resolution discretizing the 3D space. The 3D heatmap encodes per-voxel likelihood of containing each robot keypoint. We derive the root-relative 3D positions of each keypoint \\(\\mathbf{P}_n^r\\) by performing spatial integral techniques [38] over the corresponding heatmap \\(\\mathbf{H}_n\\) in a differentiable manner:" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.325, + 0.786, + 0.368 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {n} ^ {r} = \\sum_ {k = 1} ^ {D} \\sum_ {i = 1} ^ {H ^ {\\prime}} \\sum_ {j = 1} ^ {W ^ {\\prime}} (k, i, j) \\cdot \\mathbf {H} _ {n} (k, i, j). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.374, + 0.789, + 0.421 + ], + "angle": 0, + "content": "By combining \\(\\mathbf{P}^r\\) with the estimated root depth \\(d\\), we obtain the regressed 3D keypoint locations \\(\\mathbf{P}'\\) and root translation \\(\\mathbf{t}\\). We further calculate the FK-based 3D keypoint locations \\(\\mathbf{P}\\) with differentiable FK parameterized by \\(\\mathbf{q}\\), \\(\\mathbf{R}\\), and \\(\\mathbf{t}\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.438, + 0.371, + 0.455 + ], + "angle": 0, + "content": "3.3 Training loss" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.46, + 0.789, + 0.507 + ], + "angle": 0, + "content": "Ground-truth supervision For training on synthetic datasets where we have access to all the ground-truth information, we apply separate supervision for all the sub-tasks to train the corresponding network modules:" + }, + { + "type": "equation", + "bbox": [ + 0.437, + 0.519, + 0.786, + 0.537 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d e p t h}} = \\| d - \\hat {d} \\| _ {1}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.439, + 0.556, + 0.785, + 0.572 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {j o i n t}} = \\| \\mathbf {q} - \\hat {\\mathbf {q}} \\| _ {2}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.439, + 0.59, + 0.785, + 0.607 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r o t}} = \\left\\| \\mathbf {R} - \\hat {\\mathbf {R}} \\right\\| _ {2}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.439, + 0.625, + 0.785, + 0.642 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t r a n s}} = \\| \\mathbf {t} - \\hat {\\mathbf {t}} \\| _ {2}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.394, + 0.66, + 0.785, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {k p t s}} = \\| \\mathbf {P} - \\hat {\\mathbf {P}} \\| _ {2} + \\| \\mathbf {p} - \\hat {\\mathbf {p}} \\| _ {2}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.693, + 0.785, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {k p t s}} ^ {\\prime} = \\left\\| \\mathbf {P} ^ {\\prime} - \\hat {\\mathbf {P}} ^ {\\prime} \\right\\| _ {2} + \\left\\| \\mathbf {p} ^ {\\prime} - \\hat {\\mathbf {p}} ^ {\\prime} \\right\\| _ {2}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.718, + 0.788, + 0.778 + ], + "angle": 0, + "content": "where \\(d\\), \\(\\mathbf{q}\\), \\(\\mathbf{R}\\), \\(\\mathbf{t}\\), \\(\\mathbf{P}\\), and \\(\\mathbf{P}'\\) are GT root depth, joint states, camera-to-robot rotation and translation, FK-based keypoint locations and regressed keypoint locations, respectively; \\(\\mathbf{p}\\) and \\(\\mathbf{p}'\\) are 2D projections of \\(\\mathbf{P}\\) and \\(\\mathbf{P}'\\) using the camera intrinsics; \\(\\hat{}\\) denotes model estimations." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.779, + 0.649, + 0.795 + ], + "angle": 0, + "content": "We then train all the network modules end-to-end using" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.801, + 0.785, + 0.821 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {G T}} = \\mathcal {L} _ {\\text {j o i n t}} + \\mathcal {L} _ {\\text {r o t}} + \\mathcal {L} _ {\\text {t r a n s}} + \\lambda \\left(\\mathcal {L} _ {\\text {k p t s}} + \\mathcal {L} _ {\\text {k p t s}} ^ {\\prime}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.825, + 0.568, + 0.84 + ], + "angle": 0, + "content": "where we set \\(\\lambda = 10.0\\) to balance the loss terms." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "S. Ban et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "Self-supervision Apart from GT supervision on synthetic data, our framework also enables scaling the training pipeline to unlabeled real-world data in a self-supervised manner [22]. We employ two consistency-based training regularizations to improve model generalization." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.208, + 0.785, + 0.24 + ], + "angle": 0, + "content": "Despite the absence of GT joint states and keypoint locations, we regularize the consistency of the FK-based keypoints \\(\\hat{\\mathbf{P}}\\) and regressed keypoints \\(\\hat{\\mathbf{P}}'\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.44, + 0.253, + 0.785, + 0.271 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {k c}} = \\left\\| \\hat {\\mathbf {P}} - \\hat {\\mathbf {P}} ^ {\\prime} \\right\\| _ {2}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.283, + 0.785, + 0.361 + ], + "angle": 0, + "content": "We introduce this keypoint consistency regularization to effectively reconcile the predictions subject to robot structure constraints \\(\\hat{\\mathbf{P}}\\) and predictions based on the pixel-aligned feature \\(\\hat{\\mathbf{P}}'\\). In addition, similar to [22], we also regularize the consistency between robot rendering mask \\(\\mathbf{M}_{\\mathrm{render}}\\) and foreground segmentation mask \\(\\mathbf{M}_{\\mathrm{seg}}\\) using their Intersection-over-Union (IoU):" + }, + { + "type": "equation", + "bbox": [ + 0.421, + 0.373, + 0.785, + 0.404 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {m c}} = 1 - \\frac {\\mathcal {S} _ {\\text {i n t e r s e c t i o n}}}{\\mathcal {S} _ {\\text {u n i o n}}}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.41, + 0.787, + 0.456 + ], + "angle": 0, + "content": "where \\( S_{\\text{intersection}} \\) and \\( S_{\\text{union}} \\) is the area of the intersection and the union of \\( \\mathbf{M}_{\\text{render}} \\) and \\( \\mathbf{M}_{\\text{seg}} \\), respectively. \\( \\mathbf{M}_{\\text{render}} \\) is obtained with differentiable rendering using \\( \\mathbf{q} \\), \\( \\mathbf{R} \\), and \\( \\mathbf{t} \\) [15]; \\( \\mathbf{M}_{\\text{seg}} \\) is obtained with an image segmentation model [22]." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.456, + 0.665, + 0.47 + ], + "angle": 0, + "content": "The overall self-supervision training objective is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.427, + 0.485, + 0.785, + 0.5 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s e l f}} = \\mathcal {L} _ {\\mathrm {k c}} + \\lambda_ {\\mathrm {m c}} \\mathcal {L} _ {\\mathrm {m c}} \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.512, + 0.784, + 0.543 + ], + "angle": 0, + "content": "where we set \\(\\lambda_{\\mathrm{mc}} = 1.0\\). We apply end-to-end self-supervised training for the sim-to-real domain adaptation stage." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.568, + 0.376, + 0.586 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.602, + 0.431, + 0.616 + ], + "angle": 0, + "content": "4.1 Dataset and metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We utilize the benchmark dataset DREAM [18], for both training and testing in our study. This dataset comprises images from three widely used robotic arms, namely, Franka Emika Panda (Panda), Kuka iiwa7 (Kuka), and Rethink Robotics Baxter (Baxter). The training dataset for each robot consists of approximately 100k synthetic images generated using the domain randomization (DR) technique. These images encompass scenarios with varying camera-to-robot poses and joint states. The testing datasets include both DR-generate synthetic images for each robot and photorealistic synthetic images (denoted as \"Photo\") for Panda and Kuka. Additionally, real-world image datasets of Panda are provided. Panda-3Cam comprises a total of 17k image frames captured by three different cameras, namely Azure Kinect (AK), XBOX 360 Kinect (XK), and Realsense (RS). This dataset is presented as three sub-datasets, each corresponding to one camera. Another real-world dataset, Panda-ORB, contains 32,315 images captured from 27 viewpoints using a Realsense camera." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.598, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Holistic Robot Pose" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "We evaluate the accuracy of our estimations using the Average Distance (ADD) metric, which represents the average Euclidean distances (in millimeters) between predicted 3D keypoint locations and their corresponding ground truth locations. Our reported metrics include both the Area Under the Curve (AUC) value and the mean value of ADD. The AUC integrates ADD over different thresholds (with the maximum threshold set at \\(100\\mathrm{mm}\\)), while the mean value effectively represents the general prediction accuracy." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.28, + 0.455, + 0.296 + ], + "angle": 0, + "content": "4.2 Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.31, + 0.79, + 0.476 + ], + "angle": 0, + "content": "We first pre-train DepthNet on the synthetic training dataset for 100 epochs, employing a learning rate of \\( 1e - 4 \\) and relying solely on ground truth supervision for depth. Subsequently, we conduct training for our entire model based on the pretrained DepthNet for an additional 100 epochs with a learning rate of \\( 1e - 4 \\) and a decay rate of 0.95. The Adam optimizer is employed in each stage to optimize the network parameters, with the momentum set to 0.9. As depth influences the global offset of all keypoints, a pre-trained DepthNet serves as a reasonable starting point for training other networks. On the Panda real datasets, we further apply self-supervised training on real-world images to overcome the sim-to-real domain gap with a learning rate of \\( 1e - 6 \\). This fine-tuning process uses only image data and no ground-truth labels." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.503, + 0.494, + 0.519 + ], + "angle": 0, + "content": "4.3 Evaluation and Comparison" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.573, + 0.789, + 0.63 + ], + "angle": 0, + "content": "Table 1: Comparison of AUC \\( \\uparrow \\) of the ADD distribution curve and mean \\( \\downarrow \\) of the ADD, on each dataset in the DREAM datasets. Panda 3CAM datasets and Panda ORB are real-world datasets and the rest are synthetic datasets. † denotes using ground-truth joint state parameters." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.641, + 0.789, + 0.81 + ], + "angle": 0, + "content": "
MethodMetricBaxter DRKuka DRKuka PhotoPanda DRPanda PhotoPanda 3CAM-AKPanda 3CAM-XKPanda 3CAM-RSPanda ORB
† DREAM-F 18AUC ↑---81.379.568.924.476.161.9
† DREAM-Q 18AUC ↑75.5--77.874.352.437.578.057.1
† DREAM-H 18AUC ↑-73.372.182.981.160.564.078.869.1
RoboPose 15AUC ↑32.780.273.282.979.770.477.674.370.4
OursAUC ↑58.875.173.982.782.082.276.075.275.2
† DREAM-F 18Mean ↓---552.0785.811413.1491911.42077.495319.1
† DREAM-Q 18Mean ↓59.0--359.0569.078089.354178.227.264247.6
† DREAM-H 18Mean ↓-5613.211099.4325.6518.856.57381.623.625685.3
RoboPose 15Mean ↓87.934.289.021.432.234.322.326.030.1
OursMean ↓45.028.029.018.018.418.924.024.824.8
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "S. Ban et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.793, + 0.451 + ], + "angle": 0, + "content": "Accuracy We first quantitatively evaluate the estimation accuracy of our method and compare it with the state-of-the-art (SOTA) approaches. To the best of our knowledge, among prior works, only RoboPose [15] addresses a scenario similar to ours, where both the pose and joint state parameters for an industry-level robot are estimated from a single RGB image. We also present the results of DREAM [18], a representative feed-forward method using ground-truth joint state parameters, as a reference. Table [1] presents the comparisons of AUC and the mean value of ADD. Our method outperforms SOTA on most datasets, significantly reducing joint localization errors. Specifically, DREAM [18] detects the 2D keypoint locations first, then employs a PnP process to solve the 2D-to-3D transformation. However, once the 2D keypoint detector fails due to truncations and occlusions, the PnP process produces unreliable numerical solutions. RoboPose [15] utilizes iterative render-and-compare, which is prone to falling into local optima on complex robot models with symmetric structures (e.g., Baxter). In contrast, our approach directly regresses both the robot pose and joint state parameters, enabling end-to-end optimization from multi-facet supervision signals. We also visualize the complete ADD distribution. As shown in Fig. [3], our approach achieves higher accuracy across most ADD thresholds. Notably, our approach attains near \\(100\\%\\) accuracy with a \\(60\\mathrm{mm}\\) threshold, indicating more reliable estimation results on challenging cases with fewer outliers." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.471, + 0.359, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.472, + 0.502, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.472, + 0.643, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.472, + 0.785, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.626 + ], + "angle": 0, + "content": "Fig. 3: Comparison of ADD distributions on the real-world datasets between our approach and RoboPose [15]. The y-axis represents the accuracy of our estimation at different ADD thresholds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.654, + 0.79, + 0.792 + ], + "angle": 0, + "content": "Finally, we visualize the predictions of our approach and compare them with SOTA to better understand the performance improvement. Figure 4 illustrates our approach's ability to produce high-quality estimates across different robot morphologies. Particularly, our approach demonstrates superior performance in highly challenging cases, including self-occlusions, truncations, and extreme lighting conditions. Additionally, we compare the model performance qualitatively on in-the-wild lab images and quantitatively under various levels of truncation severity, illustrating the superior generalization ability and robustness of our approach. Please refer to the appendix for more detailed results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.794, + 0.842 + ], + "angle": 0, + "content": "Efficiency Furthermore, we benchmark the computation efficiency of our method and compare it with previous works. We measure the average inference time," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.597, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Holistic Robot Pose" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.158, + 0.785, + 0.186 + ], + "angle": 0, + "content": "Table 2: Comparison of computation efficiency on the Panda 3CAM-AK dataset. † denotes using ground-truth joint state parameters." + }, + { + "type": "table", + "bbox": [ + 0.334, + 0.198, + 0.675, + 0.261 + ], + "angle": 0, + "content": "
MethodAUC ↑Time (ms) ↓FPS ↑FLOPS (G) ↓
† DREAM-F 1868.966.015.190.5
RoboPose 1570.4570.81.857.3
Ours82.244.322.625.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.289, + 0.788, + 0.442 + ], + "angle": 0, + "content": "frame per second (FPS), and Floating Point Operations (FLOPs). For a fair comparison, we benchmark all methods on a Linux machine with CPU Intel(R) Xeon(R) Gold 6240 CPU @ 2.60GHz and GPU NVIDIA Tesla V100S. Batch size is set to be 1 for all methods. Table 2 shows that our method exhibits a remarkable advantage in terms of inference speed and computation loads, enabling real-time inference while achieving state-of-the-art accuracy. This advantage stems from the fact that our method involves only a single feed-forward pass without costly iterative optimization during inference. We demonstrate that our approach serves as an efficient and effective solution for real-time holistic robot pose estimation, which holds great potential in real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.464, + 0.58, + 0.48 + ], + "angle": 0, + "content": "4.4 Comparison under known joints states" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.788, + 0.596 + ], + "angle": 0, + "content": "Although the primary focus of this work is holistic robot pose estimation with unknown states, our approach can also be easily adapted to the known joint states setting. We conduct additional experiments on four real datasets for robot pose estimation with ground-truth joint states available. We simply replace the predicted joint state with the ground-truth joint state during the self-supervised training process. As shown in Tab. 3 our approach also achieves comparable performance with previous SOTA under this well-explored setting." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.618, + 0.387, + 0.634 + ], + "angle": 0, + "content": "4.5 Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.735 + ], + "angle": 0, + "content": "Ablation on network modules We first conduct an ablation study on the proposed network modules to investigate their effectiveness. Since JointNet and RotationNet are the major regression targets of our task, we focus on the ablation of two assistive network modules, DepthNet and KeypointNet. We conduct the experiments on the synthetic datasets, utilizing only ground-truth supervision to simplify the process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Firstly, we delve into the implementation of DepthNet. In Tab. 4, we evaluate the design choice of whether to share the backbone of DepthNet with other modules or not. The results indicate that separating the DepthNet backbone consistently improves performance. This improvement could be explained by the fact that DepthNet involves learning scale-variant features, while the other network modules all focus on learning scale-invariant features. We also investigate whether to pre-train the DepthNet or jointly train it with other modules" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "S. Ban et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.157, + 0.788, + 0.2 + ], + "angle": 0, + "content": "Table 3: Comparison of AUC \\(\\uparrow\\) of the ADD distribution curve and Mean \\(\\downarrow\\) of the ADD, on the real-world datasets of DREAM, with all methods using known ground-truth joint state parameters." + }, + { + "type": "table", + "bbox": [ + 0.235, + 0.212, + 0.773, + 0.413 + ], + "angle": 0, + "content": "
MethodMetricPanda 3CAM-AKPanda 3CAM-XKPanda 3CAM-RSPanda ORB
DREAM-F18AUC↑68.924.476.161.9
DREAM-Q18AUC↑52.437.578.057.1
DREAM-H18AUC↑60.564.078.869.1
RoboPose15AUC↑76.586.076.980.5
CtRNet22AUC↑89.979.590.885.3
SGTAPose40AUC↑67.82.187.672.3
OursAUC↑90.281.291.987.6
DREAM-F18Mean↓11413.1491911.42077.495319.1
DREAM-Q18Mean↓78089.354178.227.264247.6
DREAM-H18Mean↓56.57381.623.625685.3
RoboPose15Mean↓24.214.023.119.4
CtRNet22Mean↓13.032.010.021.0
SGTAPose40Mean↓35.7157.712.835.0
OursMean↓12.919.911.015.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.442, + 0.489, + 0.484 + ], + "angle": 0, + "content": "Table 4: Ablation studies on the effect of separating DepthNet image feature extractor." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.497, + 0.487, + 0.578 + ], + "angle": 0, + "content": "
DatasetSeparateAUC ↑Mean (mm)↓
Panda DR (a)-79.9821.9
Panda DR (b)+82.6818.0
Panda Photo (a)-79.6622.0
Panda Photo (b)+81.9818.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.442, + 0.789, + 0.485 + ], + "angle": 0, + "content": "Table 5: Ablation studies on the effect of pre-training DepthNet in advance of jointly training." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.497, + 0.787, + 0.578 + ], + "angle": 0, + "content": "
DatasetPretrainAUC ↑Mean (mm)↓
Panda DR (a)-74.2926.8
Panda DR (b)+82.6818.0
Panda Photo (a)-73.3427.9
Panda Photo (b)+81.9818.4
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.611, + 0.788, + 0.702 + ], + "angle": 0, + "content": "from scratch. Table 5 suggests that a pre-training stage leads to considerable performance improvement. This improvement likely occurs because a well-trained DepthNet reliably aids other modules in better training, as it determines the global translation of all joints. Additionally, we select the keypoint closest to the robot center as the root keypoint requiring depth prediction by DepthNet. Please refer to the S2.3 section of the appendix for rationale behind this selection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Moreover, we conduct ablation experiments to investigate the impact of KeypointNet on other network modules, namely JointNet and RotationNet. We utilize Mean Joint Error, which represents the mean error of the joint state parameter predictions, and the Mean Rotation Angle Error [15], which represents the mean error of the Euler angle predictions \\((\\theta_{x}^{pred}, \\theta_{y}^{pred}, \\theta_{z}^{pred})\\) for the camera-to-robot rotation, to quantify the accuracy of the JointNet and RotationNet predictions, respectively. It is important to note that a Panda robot consists of seven revolute joints with angle parameters (representing one-DoF rotations) and one prismatic joint with a length parameter (representing one-DoF transla" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.598, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Holistic Robot Pose" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.147, + 0.749, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.452, + 0.788, + 0.481 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparison between our method and RoboPose [15] on both real and synthetic datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.789, + 0.612 + ], + "angle": 0, + "content": "tion). Therefore, we calculate and present the Mean Joint Error separately for the revolute and prismatic joints of Panda. We present the detailed ablation results in the S2.1 section of the appendix, in which Tables 6 and 7 show that the presence of KeypointNet significantly improves the prediction accuracy of JointNet and RotationNet. By sharing a feature extractor with JointNet and RotationNet, KeypointNet guides the network towards learning pixel-aligned robot visual features, which further assists in estimating robot pose and joint states." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.789, + 0.718 + ], + "angle": 0, + "content": "Finally, we examine the overall effect of DepthNet and KeypointNet. To replace DepthNet, we implement a straightforward alternative, directly regressing the camera-to-robot translation \\(\\mathbf{t}\\) from image features \\(\\mathbf{f}\\) using an MLP, similar to RotationNet. Figure 5a illustrates that both modules play an important role in prediction accuracy. The DepthNet notably enhances accuracy by improving root depth estimation, a global offset that affects the precision of all joints. Meanwhile, KeypointNet further benefits the estimation results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.796, + 0.841 + ], + "angle": 0, + "content": "The contribution of these modules to the overall performance validates our design. DepthNet disentangles depth from camera intrinsic, facilitating a learning-based approach to depth estimation. This method exhibits greater expressive capability compared to the analytical, non-parameterized PnP methods employed in previous work [18,22,40]. On the other hand, KeypointNet enhances both JointNet and RotationNet by steering the shared backbone towards learning more pixel-aligned features, thus complementing the learning objectives defined in the robot kinematic space." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "S. Ban et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.345 + ], + "angle": 0, + "content": "Ablation on self-supervision We also perform an ablation study of our self-supervised learning objectives. We conduct the experiments on the Panda real datasets. Figure 5b demonstrates that both the keypoint consistency regularization and the mask consistency regularization contribute to the overall performance. Specifically, the mask consistency regularization provides pixel-wise supervision signal with the help of differentiable rendering, while the keypoint consistency regularization unifies the predictions in the pixel space and that in the robot kinematic space. We suppose that pixel-aligned 2D tasks, e.g. keypoint detection and mask segmentation, generally generalize better, and could therefore improve the 3D estimation quality through consistency regularization, further improving the overall performance. As a result, our multi-stage training scheme yields favorable convergence ability. Please refer to the S4 section of the appendix for the error analysis of each module across each training stage." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.377, + 0.482, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.513, + 0.488, + 0.556 + ], + "angle": 0, + "content": "(a) Ablation study on network modules. 'Baseline' represents direct regression, 'DN' and 'KN' refer to DepthNet and KeypointNet, respectively." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.377, + 0.78, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.513, + 0.788, + 0.556 + ], + "angle": 0, + "content": "(b) Ablation study on self-supervision. 'Baseline' represents no self-supervision, 'MC' and 'KC' refer to mask and keypoint consistency regularization, respectively." + }, + { + "type": "image_caption", + "bbox": [ + 0.281, + 0.567, + 0.722, + 0.582 + ], + "angle": 0, + "content": "Fig. 5: Ablation studies on network modules and self-supervision." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.647, + 0.36, + 0.663 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "This work presents a novel end-to-end framework for real-time holistic robot pose estimation without prior knowledge of internal robot states, providing substantial assistance in real-world scenarios where the joint state of the estimation target is unavailable or unreliable. By decomposing this complex task into specific sub-tasks and tailoring neural network modules for each, the method effectively addresses the challenges while retaining high computation efficiency. Our approach, which circumvents iterative optimization, significantly accelerates the estimation process and achieves state-of-the-art accuracy. This advancement holds promise for enhancing real-world robotics applications, offering a practical solution for real-time robot pose estimation and joint state monitoring." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.598, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Holistic Robot Pose" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.174, + 0.787, + 0.234 + ], + "angle": 0, + "content": "This work is supported by National Science and Technology Major Project (2022ZD0114904). We thank Hai Ci for the valuable discussions during the early stages of this work. We thank Haoran Lu and Ruihai Wu for their assistance with real-world image captures." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.255, + 0.323, + 0.271 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.283, + 0.786, + 0.324 + ], + "angle": 0, + "content": "1. Bultmann, S., Memmesheimer, R., Behnke, S.: External camera-based mobile robot pose estimation for collaborative perception with smart edge sensors. In: ICRA (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.325, + 0.786, + 0.365 + ], + "angle": 0, + "content": "2. Chen, L., Qin, Y., Zhou, X., Su, H.: EasyHeC: Accurate and Automatic Hand-eye Calibration via Differentiable Rendering and Space Exploration. In: IEEE Robotics and Automation Letters (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.365, + 0.786, + 0.419 + ], + "angle": 0, + "content": "3. Christen, S., Yang, W., Pérez-D'Arpino, C., Hilliges, O., Fox, D., Chao, Y.W.: Learning human-to-robot handovers from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9654–9664 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.42, + 0.786, + 0.46 + ], + "angle": 0, + "content": "4. Deisenroth, M.P., Rasmussen, C.E., Fox, D.: Learning to control a low-cost manipulator using data-efficient reinforcement learning. Robotics: Science and Systems VII 7, 57-64 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.461, + 0.786, + 0.502 + ], + "angle": 0, + "content": "5. Feniello, A., Dang, H.N., Birchfield, S.: Program synthesis by examples for object repositioning tasks. 2014 IEEE/RSJ International Conference on Intelligent Robots and Systems pp. 4428-4435 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.502, + 0.786, + 0.529 + ], + "angle": 0, + "content": "6. Fiala, M.: Artag, a fiducial marker system using digital techniques. In: CVPR (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.529, + 0.786, + 0.569 + ], + "angle": 0, + "content": "7. Garrido-Jurado, S., Muñoz-Salinas, R., Madrid-Cuevas, F.J., Marín-Jiménez, M.J.: Automatic generation and detection of highly reliable fiducial markers under occlusion. Pattern Recognition (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.569, + 0.786, + 0.61 + ], + "angle": 0, + "content": "8. Hao, R., Özgün, O., Çavusoglu, M.C.: Vision-based surgical tool pose estimation for the da Vinci® robotic surgical system. In: 2018 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 1298-1305. IEEE (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.611, + 0.786, + 0.637 + ], + "angle": 0, + "content": "9. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.637, + 0.786, + 0.664 + ], + "angle": 0, + "content": "0. Icli, C., Stepanenko, O., Bonev, I.: New method and portable measurement device for the calibration of industrial robots. Sensors 20(20) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.664, + 0.786, + 0.705 + ], + "angle": 0, + "content": "1. Kim, W., Kim, N., Lyons, J.B., Nam, C.S.: Factors affecting trust in high-vulnerability human-robot interaction contexts: A structural equation modelling approach. Applied Ergonomics 85, 103056 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.705, + 0.786, + 0.732 + ], + "angle": 0, + "content": "2. Kok, B.C., Soh, H.: Trust in robots: Challenges and opportunities. Current Robotics Reports 1, 297-309 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.733, + 0.786, + 0.772 + ], + "angle": 0, + "content": "3. Kothari, N., Gupta, M., Vachhani, L., Arya, H.: Pose estimation for an autonomous vehicle using monocular vision. In: 2017 Indian Control Conference (ICC). pp. 424-431 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.786, + 0.8 + ], + "angle": 0, + "content": "4. Kuipers, B.: How can we trust a robot? Communications of the ACM 61(3), 86-95 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.786, + 0.84 + ], + "angle": 0, + "content": "5. Labbé, Y., Carpentier, J., Aubry, M., Sivic, J.: Single-view robot pose and joint angle estimation via render & compare. In: Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.283, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.357, + 0.127 + ], + "angle": 0, + "content": "S. Ban et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.204 + ], + "angle": 0, + "content": "16. Lambrecht, J., Grosenick, P., Meusel, M.: Optimizing keypoint-based single-shot camera-to-robot pose estimation through shape segmentation. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 13843-13849. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.205, + 0.787, + 0.231 + ], + "angle": 0, + "content": "17. Lambrecht, J., Kästner, L.: Towards the usage of synthetic data for marker-less pose estimation of articulated robots in rgb images. In: ICAR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.787, + 0.272 + ], + "angle": 0, + "content": "18. Lee, T.E., Tremblay, J., To, T., Cheng, J., Mosier, T., Kroemer, O., Fox, D., Birchfield, S.: Camera-to-robot pose estimation from a single image. In: International Conference on Robotics and Automation (ICRA) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.273, + 0.787, + 0.3 + ], + "angle": 0, + "content": "19. Lepetit, V., Moreno-Noguer, F., Fua, P.: Epnp: An accurate o (n) solution to the pnp problem. IJCV (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.301, + 0.787, + 0.342 + ], + "angle": 0, + "content": "20. Li, S., De Wagter, C., De Croon, G.C.: Self-supervised monocular multi-robot relative localization with efficient deep neural networks. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 9689-9695. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.787, + 0.383 + ], + "angle": 0, + "content": "21. Li, X., Zhang, X., Zhou, X., Chen, I.M.: Upg: 3d vision-based prediction framework for robotic grasping in multi-object scenes. Knowledge-Based Systems 270, 110491 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.425 + ], + "angle": 0, + "content": "22. Lu, J., Richter, F., Yip, M.C.: Markerless camera-to-robot pose estimation via self-supervised sim-to-real transfer. In: Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.426, + 0.787, + 0.466 + ], + "angle": 0, + "content": "23. Mahler, J., Matl, M., Satish, V., Danielczuk, M., DeRose, B., McKinley, S., Goldberg, K.: Learning ambidextrous robot grasping policies. Science Robotics 4(26), eaau4984 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.787, + 0.508 + ], + "angle": 0, + "content": "24. Masehian, E., Jannati, M., Hekmatfar, T.: Cooperative mapping of unknown environments by multiple heterogeneous mobile robots with limited sensing. Robotics and Autonomous Systems 87, 188-218 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.787, + 0.536 + ], + "angle": 0, + "content": "25. Ming, Y., Meng, X., Fan, C., Yu, H.: Deep learning for monocular depth estimation: A review. Neurocomputing 438, 14-33 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.787, + 0.577 + ], + "angle": 0, + "content": "26. Moon, G., Chang, J., Lee, K.M.: Camera distance-aware top-down approach for 3d multi-person pose estimation from a single rgb image. In: The IEEE Conference on International Conference on Computer Vision (ICCV) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.619 + ], + "angle": 0, + "content": "27. Morrison, D., Corke, P., Leitner, J.: Closing the Loop for Robotic Grasping: A Real-time, Generative Grasp Synthesis Approach. In: Proc. of Robotics: Science and Systems (RSS) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.787, + 0.633 + ], + "angle": 0, + "content": "28. Olson, E.: Apriltag: A robust and flexible visual fiducial system. In: ICRA (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.634, + 0.787, + 0.673 + ], + "angle": 0, + "content": "29. Papadimitriou, A., Mansouri, S.S., Nikolakopoulos, G.: Range-aided ego-centric collaborative pose estimation for multiple robots. Expert Systems with Applications 202, 117052 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.702 + ], + "angle": 0, + "content": "30. Park, D., Seo, Y., Chun, S.Y.: Real-time, highly accurate robotic grasp detection using fully convolutional neural networks with high-resolution images (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.703, + 0.787, + 0.73 + ], + "angle": 0, + "content": "31. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.731, + 0.787, + 0.77 + ], + "angle": 0, + "content": "32. Pinto, L., Davidson, J., Gupta, A.: Supervision via competition: Robot adversaries for learning tasks. In: 2017 IEEE International Conference on Robotics and Automation (ICRA). pp. 1601-1608 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.812 + ], + "angle": 0, + "content": "33. Prusak, A., Melnychuk, O., Roth, H., Schiller, I., Koch, R.: Pose estimation and map building with a time-of-flight-camera for robot navigation. International Journal of Intelligent Systems Technologies and Applications 5(3-4), 355-364 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "34. Rana, A., Vulpi, F., Galati, R., Milella, A., Petitti, A.: A pose estimation algorithm for agricultural mobile robots using an rgb-d camera. In: 2022 International Con-" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.598, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Holistic Robot Pose" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.148, + 0.786, + 0.175 + ], + "angle": 0, + "content": "ference on Electrical, Computer, Communications and Mechatronics Engineering (ICECCME). pp. 1-5 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.786, + 0.203 + ], + "angle": 0, + "content": "35. Rizk, Y., Awad, M., Tunstel, E.W.: Cooperative heterogeneous multi-robot systems. ACM Computing Surveys (CSUR) 52, 1 - 31 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "36. Salvini, P., Paez-Granados, D.F., Billard, A.: Safety concerns emerging from robots navigating in crowded pedestrian areas. International Journal of Social Robotics 14, 441 - 462 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.786, + 0.273 + ], + "angle": 0, + "content": "37. Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learning for human pose estimation. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.786, + 0.3 + ], + "angle": 0, + "content": "38. Sun, X., Xiao, B., Wei, F., Liang, S., Wei, Y.: Integral human pose regression. In: Proceedings of the European Conference on Computer Vision (ECCV) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.786, + 0.342 + ], + "angle": 0, + "content": "39. Svenstrup, M., Tranberg, S., Andersen, H.J., Bak, T.: Pose estimation and adaptive robot behaviour for human-robot interaction. In: 2009 IEEE International Conference on Robotics and Automation. pp. 3571-3576 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.786, + 0.384 + ], + "angle": 0, + "content": "40. Tian, Y., Zhang, J., Yin, Z., Dong, H.: Robot structure prior guided temporal attention for camera-to-robot pose estimation from image sequence. In: Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.786, + 0.411 + ], + "angle": 0, + "content": "41. Tzafestas, S.G.: 15 - mobile robots at work. In: Tzafestas, S.G. (ed.) Introduction to Mobile Robot Control, pp. 635-663. Elsevier, Oxford (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.786, + 0.452 + ], + "angle": 0, + "content": "42. Uchibe, E., Asada, M., Noda, S., Takahashi, Y., Hosoda, K.: Vision-based reinforcement learning for robocup: Towards real robot competition. In: Proc. of IROS. vol. 96 (1996)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.786, + 0.479 + ], + "angle": 0, + "content": "43. Xia, Y., Tang, Y., Hu, Y., Hoffman, G.: Privacy-preserving pose estimation for human-robot interaction. CoRR abs/2011.07387 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.48, + 0.786, + 0.522 + ], + "angle": 0, + "content": "44. Xu, C., Yu, X., Wang, Z., Ou, L.: Multi-view human pose estimation in humanrobot interaction. In: IECON 2020 The 46th Annual Conference of the IEEE Industrial Electronics Society. pp. 4769-4775 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.786, + 0.563 + ], + "angle": 0, + "content": "45. Xun, Z., Huang, J., Li, Z., Xu, C., Gao, F., Cao, Y.: Crepes: Cooperative relative pose estimation towards real-world multi-robot systems. arXiv preprint arXiv:2302.01036 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.786, + 0.605 + ], + "angle": 0, + "content": "46. Yang, W., Paxton, C., Mousavian, A., Chao, Y.W., Cakmak, M., Fox, D.: Reactive human-to-robot handovers of arbitrary objects. In: IEEE International Conference on Robotics and Automation (ICRA). IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.786, + 0.646 + ], + "angle": 0, + "content": "47. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5745-5753 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.646, + 0.786, + 0.688 + ], + "angle": 0, + "content": "48. Zhu, W., Ma, X., Ro, D., Ci, H., Zhang, J., Shi, J., Gao, F., Tian, Q., Wang, Y.: Human motion generation: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.688, + 0.786, + 0.716 + ], + "angle": 0, + "content": "49. Zuo, Y., Qiu, W., Xie, L., Zhong, F., Wang, Y., Yuille, A.L.: Craves: Controlling robotic arm with a vision-based economic system. In: CVPR (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.716 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Real-time Holistic Robot Pose Estimation with Unknown States/0b349d89-54a2-48da-a5a9-0c888f46a6ff_origin.pdf b/2024/Real-time Holistic Robot Pose Estimation with Unknown States/0b349d89-54a2-48da-a5a9-0c888f46a6ff_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4024a1859b3fa66d79edc26c3b3adb3c249eed26 --- /dev/null +++ b/2024/Real-time Holistic Robot Pose Estimation with Unknown States/0b349d89-54a2-48da-a5a9-0c888f46a6ff_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6315c671a57b2187a778c529f79fcaf62e3f0c96c87232b894930cd105b01f3e +size 4440627 diff --git a/2024/Real-time Holistic Robot Pose Estimation with Unknown States/full.md b/2024/Real-time Holistic Robot Pose Estimation with Unknown States/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7877dd279bb49d78063dc34f697e064f43b334f0 --- /dev/null +++ b/2024/Real-time Holistic Robot Pose Estimation with Unknown States/full.md @@ -0,0 +1,310 @@ +# Real-time Holistic Robot Pose Estimation with Unknown States + +Shikun Ban1, Juling Fan1, Xiaoxuan Ma1, Wentao Zhu1,2, Yu Qiao5, and Yizhou Wang1,2,3,4 + +1 Center on Frontiers of Computing Studies, School of Computer Science, Peking University +2 Inst. for Artificial Intelligence, Peking University +3 Nat'l Eng. Research Center of Visual Technology +4 Nat'l Key Lab of General Artificial Intelligence + +$^{5}$ Department of Automation, School of Electronics, Information and Electrical Engineering, Shanghai Jiao Tong University + +Abstract. Estimating robot pose from RGB images is a crucial problem in computer vision and robotics. While previous methods have achieved promising performance, most of them presume full knowledge of robot internal states, e.g. ground-truth robot joint angles. However, this assumption is not always valid in practical situations. In real-world applications such as multi-robot collaboration or human-robot interaction, the robot joint states might not be shared or could be unreliable. On the other hand, existing approaches that estimate robot pose without joint state priors suffer from heavy computation burdens and thus cannot support real-time applications. This work introduces an efficient framework for real-time robot pose estimation from RGB images without requiring known robot states. Our method estimates camera-to-robot rotation, robot state parameters, keypoint locations, and root depth, employing a neural network module for each task to facilitate learning and sim-to-real transfer. Notably, it achieves inference in a single feed-forward pass without iterative optimization. Our approach offers a $12 \times$ speed increase with state-of-the-art accuracy, enabling real-time holistic robot pose estimation for the first time. Code and models are available at https://github.com/Oliverbansk/Holistic-Robot-Pose-Estimation + +# 1 Introduction + +Robot pose estimation is a fundamental task in computer vision and robotics as it not only lays the groundwork for multiple downstream tasks, e.g. visually-guided object grasping and manipulation [21,40,49], but also paves the way for a variety of applications, such as autonomous navigation [1,13,33,34], human-robot interaction [3,39,43,44,46,48] and multi-robot collaboration [20,29,35,45]. + +![](images/400b567273e0a3757470601efe968e5aad3887625d0ab8d4e346edfc1927a0d6.jpg) +Fig. 1: The majority of previous robot pose estimation methods assume known robot joint states and focus solely on estimating the camera-to-robot pose, i.e. camera-to-robot rotation and translation. In contrast, the holistic robot pose estimation problem requires estimating both joint states and the camera-to-robot pose, given only an RGB image without known joint states. For holistic robot pose estimation, RoboPose [15] uses costly test-time optimization (Render-and-Compare) to iteratively refine the predictions. In contrast, our feed-forward method achieves state-of-the-art accuracy with a $12 \times$ speed boost. + +Nonetheless, most previous works for monocular robot pose estimation [17,18,22,40] rely on a known set of robot joint angles as an additional input to alleviate the difficulty of the estimation process. Yet, this prerequisite information of the robot state is not always available in practical scenarios: + +- For instance, low-cost robotic arms without sophisticated sensors or internal monitors can not provide the visual calibration process with readily accessible robot state measurements [4,10,49]. +- In the context of multi-robot collaboration, the real-time synchronization of joint state data may be hindered by environmental limitations [24], necessitating pose estimation without robot joint state information, and additional monitoring of the joint states. +- In domains such as human-robot interaction [36,41] and competitive robot gaming [32,42], the reliability of the state information from counterpart robots is not always guaranteed, as trust issues may arise in robotic systems [11,12,14]. In this sense, failing to obtain independent estimation of the counterpart robot joint state may lead to potential safety issues. + +These scenarios further emphasize the urgent need for monocular robot pose estimation with unknown robot joint state. In this work, we address the challenge of Holistic Robot Pose Estimation with unknown internal states for articulated robots. Specifically, given a monocular RGB image, we aim to simultaneously solve for 6D robot pose (3D rotation and translation relative to the camera) and fine-grained robot joint state parameters (including joint angles for revolute joints and displacement for prismatic joints). + +Nonetheless, learning holistic robot pose estimation without joint state information presents significant challenges. The primary challenge is the dual ambiguity of unknown joint states and 6D pose. Specifically, estimating the 6D pose + +of an articulated robot is more complex than for a rigid object due to shape and scale variability from unknown joint states. Conversely, estimating joint states is complicated by the need for 2D-to-3D correspondences and the unknown 6D pose. This intertwined issue requires novel solutions to decouple joint states from 6D pose estimation. Furthermore, the sophisticated robot morphologies vary from model to model, requiring non-trivial model design. In addition, The robot parts are frequently invisible in the monocular image as a result of self-occlusion and truncation, posing more challenges for robust estimation. Finally, the domain gap between synthetic training data and real-world images introduces additional challenges for learning-based approaches. + +It is worth noting that some previous works [15,49] have also addressed holistic estimation with unknown states, tackling some challenges. However, these methods all necessitate iterative optimization at inference, which is computationally intensive and unsuitable for real-time applications. Motivated by the need for efficient holistic robot pose estimation without knowledge of the robot state, we propose an end-to-end framework that directly estimates camera-to-robot pose and robot joint state parameters, requiring only a single feed-forward pass during inference. + +In pursuit of this goal, we factorize the holistic robot pose estimation task into several sub-tasks, including the estimation of camera-to-robot rotation, robot joint states, root depth, and root-relative keypoint locations. We then design corresponding neural network modules for each sub-task. Specifically, the standalone root depth estimator disentangles the estimation of scale-variant and scale-invariant variables. The keypoint estimator bridges the gap between the pixel space and the robotic parameter space. Additionally, we utilize self-supervised consistency regularizations which could be performed on the real images without ground truth (GT). + +We demonstrate that our designs of modular feed-forward networks improve model generalization while maintaining high computation efficiency. Consequently, our approach achieves state-of-the-art estimation accuracy with unknown robot states and a $12 \times$ speedup compared to the iterative Render-and-Compare (RnC) [15] approach, as depicted in Fig. 1. The key insight behind our approach is that we could get rid of the cumbersome RnC [15] and error-prone Perspective-n-Points (PnP) [18,22,40] schemes to establish 2D-3D correspondences. Instead, we demonstrate that feed-forward networks could well handle the problem with proper problem decomposition and meticulous design. + +In summary, our contributions are three-fold: + +1. We present an efficient end-to-end learning framework for the holistic robot pose estimation problem with unknown robot states. +2. We propose to factorize and decompose the holistic robot pose estimation problem into several sub-tasks and tackle them with respective modules. +3. We achieve state-of-the-art performance on various robot models while delivering a significant $12 \times$ speedup, paving the way for real-time applications. + +# 2 Related Work + +# 2.1 Hand-eye calibration + +The goal of Hand-Eye Calibration (HEC) is to determine the transformation between the position of a robot's end-effector (hand) and its vision system (eye). This task can be categorized into two types: eye-in-hand where the camera is fixed at the robot end-effector [27], and eye-to-hand which refers to the configurations where the camera is placed at a distance from the robot [5,23,30]. This work focuses on a similar setting with eye-to-hand. Traditional HEC methods involve attaching a fiducial marker to the end effector [6,7,28], placing it in several known positions, detecting the marker positions in images, and then solving for the camera-to-robot transformation. Despite its broad applicability, the classical HEC pipeline necessitates a complete offline calibration process every time the camera-robot transformation changes, which can be tedious in real-world applications. Consequently, recent works, such as DREAM [18], propose an online, markerless calibration problem, as discussed below. + +# 2.2 Image-based robot pose estimation + +Previous works can be divided into two categories based on whether the robot states are known or not. Most previous works estimate the 6D camera-to-robot pose with known robot states. For example, DREAM [18] first trains a 2D keypoint estimator and then applies Perspective- $n$ -Point (PnP) [19] to obtain camera-to-robot pose, showing comparable results with the classical methods. Some work further utilizes segmentation masks and edges to aid pose estimation [8,16]. Chen et al. [2] apply rendering-based camera pose optimization and consistency-based joint space exploration. SGTAPose [40] leverages robot structure prior and temporal information. Lu et al. [22] propose to train on real-world data using foreground segmentation and differentiable rendering in a self-supervised manner. + +Recently, several works have shifted to estimating robot pose with unknown joint states. Zuo et al. [49] estimate the robot pose and joint angles of a simple 4-DoF robotic arm from a single RGB image by solving a nonlinear nonconvex optimization problem. However, this method is not well-suited for robot arms with higher complexity. Labbé et al. [15] propose a render-and-compare approach to iteratively refine the estimation for both known and unknown states, and apply the method to different robot models up to 15-DoF. However, the iterative optimization process limits its applications in real-time scenarios. In contrast, our approach can be deployed with a single feed-forward at inference and achieves real-time efficiency. + +# 3 Method + +# 3.1 Overview + +Given an RGB image $\mathbf{I}$ , we employ a suite of neural network modules, including DepthNet, JointNet, RotationNet, and KeypointNet. These modules are used + +![](images/30c7f4508dca82009277bbbef491c2aa93c01e4d95b128047e0361da616dfe0a.jpg) +Fig. 2: Framework overview. The JointNet and the RotationNet regress joint state parameters $\mathbf{q}$ and camera-to-robot rotation $\mathbf{R}$ , respectively. The KeypointNet estimates root-relative 3D keypoint locations $\mathbf{P}^r$ . The DepthNet's estimation of root depth $d$ is combined with $\mathbf{P}^r$ to acquire absolute 3D keypoint locations $\mathbf{P}'$ and camera-to-robot translation $\mathbf{t}$ . Joint state parameters $\mathbf{q}$ , rotation $\mathbf{R}$ and translation $\mathbf{t}$ are used to compute 3D keypoint locations $\mathbf{P}$ via forward-kinematics. + +to respectively estimate the absolute root depth $d$ , joint state parameters $\mathbf{q}$ , camera-to-robot rotation $\mathbf{R} \in SO(3)$ , and root-relative 3D keypoint locations $\mathbf{P}^r \in \mathbb{R}^{N \times 3}$ , where $N$ represents the total number of keypoints. The keypoints are defined at the joint articulations, in line with previous works [15,18,22], and one keypoint is designated as the robot root. The root joint is critical for computing the camera-to-robot rotation $\mathbf{R}$ and translation $\mathbf{t}$ , as well as serving as the origin for the forward kinematics (FK) process. + +Our modularized design offers several key insights. Firstly, DepthNet is tailored to learn the scale-variant camera-robot spatial relationship, while the other modules concentrate on the scale-invariant robot features. Secondly, while the joint state parameters $\mathbf{q}$ provide a compact representation of robot morphologies subject to robotic structure constraints, they are highly abstract and thus challenging to directly regress from image features. To address this, we utilize the 3D keypoint locations as an intermediary. On the one hand, they can be effectively estimated with pixel-aligned image features; on the other hand, they are inherently linked to $\mathbf{q}$ through an FK process. + +An overview of the proposed framework is illustrated in Fig. 2. Specifically, DepthNet estimates the absolute root depth $d$ , which, combined with KeypointNet's output $\mathbf{P}^r \in \mathbb{R}^{N \times 3}$ , provides the absolute 3D keypoint locations $\mathbf{P}' \in \mathbb{R}^{N \times 3}$ . The root keypoint location in $\mathbf{P}'$ is the camera-to-robot translation $\mathbf{t}$ . The estimated joint state parameters $\mathbf{q}$ , along with the camera-to-robot rotation $\mathbf{R}$ and translation $\mathbf{t}$ , are then used to perform FK, yielding the final + +3D keypoint estimations $\mathbf{P} \in \mathbb{R}^{N \times 3}$ . Please refer to the appendix for a detailed illustration of the relations of these variables. + +In the following, we will first introduce the architecture design of the network modules in Sec. 3.2, and then explain the training loss design consisting of ground-truth supervision and self-supervision in Sec. 3.3. + +# 3.2 Architecture + +DepthNet Monocular object depth estimation suffers from significant ambiguities, as the absolute depth value and camera intrinsics are deeply intertwined [25]. For instance, a robot in a fixed pose relative to the camera may appear larger in photos taken with cameras having larger focal lengths. Scaling both the absolute depth and focal lengths equally results in identical object size and appearance in different images. To address this challenge, we implement DepthNet following the approach in [26] to disentangle depth from camera intrinsics. Initially, we use camera intrinsics to calculate a normalized, coarse depth value $d_{c}$ : + +$$ +d _ {c} = \sqrt {\frac {f _ {x} \cdot f _ {y} \cdot A _ {r e a l}}{A _ {b b o x}}}, \tag {1} +$$ + +where $f_{x}$ and $f_{y}$ represent the focal lengths along the x- and y-axes, $A_{real}$ is the area of the robot in the x-y plane of the camera frame (in $mm^2$ ), and $A_{bbox}$ is the area of the robot's bounding box (in $pixel^2$ ). This method of calculating $d_{c}$ normalizes it in 3D space and provides an approximate depth estimation for subsequent refinement. The primary function of DepthNet is to predict a correction factor $\lambda$ based on the input image. We multiply the estimated $\lambda$ by $d_{c}$ to obtain the final absolute root depth $d$ : + +$$ +d = \lambda \cdot d _ {c}. \tag {2} +$$ + +With a fixed camera, the relationship between the normalized coarse depth $d_{c}$ and the robot's size in the image is linear. However, robots can assume poses with the same depth but different sizes in the image, making $d_{c}$ inadequate for accurate root depth representation. The correction factor $\lambda$ introduces non-linear flexibility to the depth estimation, addressing this issue by adjusting $d_{c}$ based on image features. + +JointNet The goal of JointNet is to estimate joint state parameters $\mathbf{q} = (q_{1}, q_{2}, \dots, q_{J})$ , where $J$ denotes the number of robot joints. As this work deals with 1-DoF joints (revolute or prismatic), $J$ equals the robot's DoF. JointNet shares the same feature extractor with RotationNet and KeypointNet. We extract feature map $\mathbf{F} \in \mathbb{R}^{C \times H \times W}$ (where $C$ denotes the number of channels, and $H \times W$ denotes the spatial resolution) and use global average pooling to obtain $\mathbf{f} \in \mathbb{R}^{C}$ . Finally, we employ several MLP blocks to regress the joint state parameters $\mathbf{q}$ from $\mathbf{f}$ . + +RotationNet The goal of RotationNet is to estimate the camera-to-robot rotation $\mathbf{R} \in SO(3)$ . We employ an MLP to estimate $\mathbf{R}$ from the image feature $\mathbf{f}$ . We utilize the 6D continuous rotation representation [47] to parameterize $\mathbf{R}$ . + +KeypointNet The goal of KeypointNet is to estimate root-relative 3D keypoint locations $\mathbf{P}^r\in \mathbb{R}^{N\times 3}$ for $N$ robot keypoints. We transform the feature map $\mathbf{F}$ to a 3D heatmap $\mathbf{H}\in \mathbb{R}^{N\times D\times H'\times W'}$ through convolutional layers, where $D\times H'\times W'$ is the spatial resolution discretizing the 3D space. The 3D heatmap encodes per-voxel likelihood of containing each robot keypoint. We derive the root-relative 3D positions of each keypoint $\mathbf{P}_n^r$ by performing spatial integral techniques [38] over the corresponding heatmap $\mathbf{H}_n$ in a differentiable manner: + +$$ +\mathbf {P} _ {n} ^ {r} = \sum_ {k = 1} ^ {D} \sum_ {i = 1} ^ {H ^ {\prime}} \sum_ {j = 1} ^ {W ^ {\prime}} (k, i, j) \cdot \mathbf {H} _ {n} (k, i, j). \tag {3} +$$ + +By combining $\mathbf{P}^r$ with the estimated root depth $d$ , we obtain the regressed 3D keypoint locations $\mathbf{P}'$ and root translation $\mathbf{t}$ . We further calculate the FK-based 3D keypoint locations $\mathbf{P}$ with differentiable FK parameterized by $\mathbf{q}$ , $\mathbf{R}$ , and $\mathbf{t}$ . + +# 3.3 Training loss + +Ground-truth supervision For training on synthetic datasets where we have access to all the ground-truth information, we apply separate supervision for all the sub-tasks to train the corresponding network modules: + +$$ +\mathcal {L} _ {\text {d e p t h}} = \| d - \hat {d} \| _ {1}, \tag {4} +$$ + +$$ +\mathcal {L} _ {\text {j o i n t}} = \| \mathbf {q} - \hat {\mathbf {q}} \| _ {2}, \tag {5} +$$ + +$$ +\mathcal {L} _ {\text {r o t}} = \left\| \mathbf {R} - \hat {\mathbf {R}} \right\| _ {2}, \tag {6} +$$ + +$$ +\mathcal {L} _ {\text {t r a n s}} = \| \mathbf {t} - \hat {\mathbf {t}} \| _ {2}, \tag {7} +$$ + +$$ +\mathcal {L} _ {\mathrm {k p t s}} = \| \mathbf {P} - \hat {\mathbf {P}} \| _ {2} + \| \mathbf {p} - \hat {\mathbf {p}} \| _ {2}, \tag {8} +$$ + +$$ +\mathcal {L} _ {\mathrm {k p t s}} ^ {\prime} = \left\| \mathbf {P} ^ {\prime} - \hat {\mathbf {P}} ^ {\prime} \right\| _ {2} + \left\| \mathbf {p} ^ {\prime} - \hat {\mathbf {p}} ^ {\prime} \right\| _ {2}, \tag {9} +$$ + +where $d$ , $\mathbf{q}$ , $\mathbf{R}$ , $\mathbf{t}$ , $\mathbf{P}$ , and $\mathbf{P}'$ are GT root depth, joint states, camera-to-robot rotation and translation, FK-based keypoint locations and regressed keypoint locations, respectively; $\mathbf{p}$ and $\mathbf{p}'$ are 2D projections of $\mathbf{P}$ and $\mathbf{P}'$ using the camera intrinsics; $\hat{}$ denotes model estimations. + +We then train all the network modules end-to-end using + +$$ +\mathcal {L} _ {\mathrm {G T}} = \mathcal {L} _ {\text {j o i n t}} + \mathcal {L} _ {\text {r o t}} + \mathcal {L} _ {\text {t r a n s}} + \lambda \left(\mathcal {L} _ {\text {k p t s}} + \mathcal {L} _ {\text {k p t s}} ^ {\prime}\right), \tag {10} +$$ + +where we set $\lambda = 10.0$ to balance the loss terms. + +Self-supervision Apart from GT supervision on synthetic data, our framework also enables scaling the training pipeline to unlabeled real-world data in a self-supervised manner [22]. We employ two consistency-based training regularizations to improve model generalization. + +Despite the absence of GT joint states and keypoint locations, we regularize the consistency of the FK-based keypoints $\hat{\mathbf{P}}$ and regressed keypoints $\hat{\mathbf{P}}'$ by: + +$$ +\mathcal {L} _ {\mathrm {k c}} = \left\| \hat {\mathbf {P}} - \hat {\mathbf {P}} ^ {\prime} \right\| _ {2}. \tag {11} +$$ + +We introduce this keypoint consistency regularization to effectively reconcile the predictions subject to robot structure constraints $\hat{\mathbf{P}}$ and predictions based on the pixel-aligned feature $\hat{\mathbf{P}}'$ . In addition, similar to [22], we also regularize the consistency between robot rendering mask $\mathbf{M}_{\mathrm{render}}$ and foreground segmentation mask $\mathbf{M}_{\mathrm{seg}}$ using their Intersection-over-Union (IoU): + +$$ +\mathcal {L} _ {\mathrm {m c}} = 1 - \frac {\mathcal {S} _ {\text {i n t e r s e c t i o n}}}{\mathcal {S} _ {\text {u n i o n}}}, \tag {12} +$$ + +where $S_{\text{intersection}}$ and $S_{\text{union}}$ is the area of the intersection and the union of $\mathbf{M}_{\text{render}}$ and $\mathbf{M}_{\text{seg}}$ , respectively. $\mathbf{M}_{\text{render}}$ is obtained with differentiable rendering using $\mathbf{q}$ , $\mathbf{R}$ , and $\mathbf{t}$ [15]; $\mathbf{M}_{\text{seg}}$ is obtained with an image segmentation model [22]. + +The overall self-supervision training objective is as follows: + +$$ +\mathcal {L} _ {\text {s e l f}} = \mathcal {L} _ {\mathrm {k c}} + \lambda_ {\mathrm {m c}} \mathcal {L} _ {\mathrm {m c}} \tag {13} +$$ + +where we set $\lambda_{\mathrm{mc}} = 1.0$ . We apply end-to-end self-supervised training for the sim-to-real domain adaptation stage. + +# 4 Experiments + +# 4.1 Dataset and metrics + +We utilize the benchmark dataset DREAM [18], for both training and testing in our study. This dataset comprises images from three widely used robotic arms, namely, Franka Emika Panda (Panda), Kuka iiwa7 (Kuka), and Rethink Robotics Baxter (Baxter). The training dataset for each robot consists of approximately 100k synthetic images generated using the domain randomization (DR) technique. These images encompass scenarios with varying camera-to-robot poses and joint states. The testing datasets include both DR-generate synthetic images for each robot and photorealistic synthetic images (denoted as "Photo") for Panda and Kuka. Additionally, real-world image datasets of Panda are provided. Panda-3Cam comprises a total of 17k image frames captured by three different cameras, namely Azure Kinect (AK), XBOX 360 Kinect (XK), and Realsense (RS). This dataset is presented as three sub-datasets, each corresponding to one camera. Another real-world dataset, Panda-ORB, contains 32,315 images captured from 27 viewpoints using a Realsense camera. + +We evaluate the accuracy of our estimations using the Average Distance (ADD) metric, which represents the average Euclidean distances (in millimeters) between predicted 3D keypoint locations and their corresponding ground truth locations. Our reported metrics include both the Area Under the Curve (AUC) value and the mean value of ADD. The AUC integrates ADD over different thresholds (with the maximum threshold set at $100\mathrm{mm}$ ), while the mean value effectively represents the general prediction accuracy. + +# 4.2 Implementation details + +We first pre-train DepthNet on the synthetic training dataset for 100 epochs, employing a learning rate of $1e - 4$ and relying solely on ground truth supervision for depth. Subsequently, we conduct training for our entire model based on the pretrained DepthNet for an additional 100 epochs with a learning rate of $1e - 4$ and a decay rate of 0.95. The Adam optimizer is employed in each stage to optimize the network parameters, with the momentum set to 0.9. As depth influences the global offset of all keypoints, a pre-trained DepthNet serves as a reasonable starting point for training other networks. On the Panda real datasets, we further apply self-supervised training on real-world images to overcome the sim-to-real domain gap with a learning rate of $1e - 6$ . This fine-tuning process uses only image data and no ground-truth labels. + +# 4.3 Evaluation and Comparison + +Table 1: Comparison of AUC $\uparrow$ of the ADD distribution curve and mean $\downarrow$ of the ADD, on each dataset in the DREAM datasets. Panda 3CAM datasets and Panda ORB are real-world datasets and the rest are synthetic datasets. † denotes using ground-truth joint state parameters. + +
MethodMetricBaxter DRKuka DRKuka PhotoPanda DRPanda PhotoPanda 3CAM-AKPanda 3CAM-XKPanda 3CAM-RSPanda ORB
† DREAM-F 18AUC ↑---81.379.568.924.476.161.9
† DREAM-Q 18AUC ↑75.5--77.874.352.437.578.057.1
† DREAM-H 18AUC ↑-73.372.182.981.160.564.078.869.1
RoboPose 15AUC ↑32.780.273.282.979.770.477.674.370.4
OursAUC ↑58.875.173.982.782.082.276.075.275.2
† DREAM-F 18Mean ↓---552.0785.811413.1491911.42077.495319.1
† DREAM-Q 18Mean ↓59.0--359.0569.078089.354178.227.264247.6
† DREAM-H 18Mean ↓-5613.211099.4325.6518.856.57381.623.625685.3
RoboPose 15Mean ↓87.934.289.021.432.234.322.326.030.1
OursMean ↓45.028.029.018.018.418.924.024.824.8
+ +Accuracy We first quantitatively evaluate the estimation accuracy of our method and compare it with the state-of-the-art (SOTA) approaches. To the best of our knowledge, among prior works, only RoboPose [15] addresses a scenario similar to ours, where both the pose and joint state parameters for an industry-level robot are estimated from a single RGB image. We also present the results of DREAM [18], a representative feed-forward method using ground-truth joint state parameters, as a reference. Table [1] presents the comparisons of AUC and the mean value of ADD. Our method outperforms SOTA on most datasets, significantly reducing joint localization errors. Specifically, DREAM [18] detects the 2D keypoint locations first, then employs a PnP process to solve the 2D-to-3D transformation. However, once the 2D keypoint detector fails due to truncations and occlusions, the PnP process produces unreliable numerical solutions. RoboPose [15] utilizes iterative render-and-compare, which is prone to falling into local optima on complex robot models with symmetric structures (e.g., Baxter). In contrast, our approach directly regresses both the robot pose and joint state parameters, enabling end-to-end optimization from multi-facet supervision signals. We also visualize the complete ADD distribution. As shown in Fig. [3], our approach achieves higher accuracy across most ADD thresholds. Notably, our approach attains near $100\%$ accuracy with a $60\mathrm{mm}$ threshold, indicating more reliable estimation results on challenging cases with fewer outliers. + +![](images/e842e3e6f8250fee4f6c64e0d874ee0a105c168976b83f1753ecbd0058f850ce.jpg) +Fig. 3: Comparison of ADD distributions on the real-world datasets between our approach and RoboPose [15]. The y-axis represents the accuracy of our estimation at different ADD thresholds. + +![](images/5070e1baded4945afba15df5224f21065d183de7284bc65f1eade51e3f032cb9.jpg) + +![](images/ef0cffca512ef254468f10e0b926f4ecb66ac706a3ab91f6a249aa9fddcdc01c.jpg) + +![](images/8181d5b4fbddedf7342b7d41ae4fdf8af9e8c28d4e4e2f7ba68cc2105f8bce84.jpg) + +Finally, we visualize the predictions of our approach and compare them with SOTA to better understand the performance improvement. Figure 4 illustrates our approach's ability to produce high-quality estimates across different robot morphologies. Particularly, our approach demonstrates superior performance in highly challenging cases, including self-occlusions, truncations, and extreme lighting conditions. Additionally, we compare the model performance qualitatively on in-the-wild lab images and quantitatively under various levels of truncation severity, illustrating the superior generalization ability and robustness of our approach. Please refer to the appendix for more detailed results. + +Efficiency Furthermore, we benchmark the computation efficiency of our method and compare it with previous works. We measure the average inference time, + +Table 2: Comparison of computation efficiency on the Panda 3CAM-AK dataset. † denotes using ground-truth joint state parameters. + +
MethodAUC ↑Time (ms) ↓FPS ↑FLOPS (G) ↓
† DREAM-F 1868.966.015.190.5
RoboPose 1570.4570.81.857.3
Ours82.244.322.625.2
+ +frame per second (FPS), and Floating Point Operations (FLOPs). For a fair comparison, we benchmark all methods on a Linux machine with CPU Intel(R) Xeon(R) Gold 6240 CPU @ 2.60GHz and GPU NVIDIA Tesla V100S. Batch size is set to be 1 for all methods. Table 2 shows that our method exhibits a remarkable advantage in terms of inference speed and computation loads, enabling real-time inference while achieving state-of-the-art accuracy. This advantage stems from the fact that our method involves only a single feed-forward pass without costly iterative optimization during inference. We demonstrate that our approach serves as an efficient and effective solution for real-time holistic robot pose estimation, which holds great potential in real-world applications. + +# 4.4 Comparison under known joints states + +Although the primary focus of this work is holistic robot pose estimation with unknown states, our approach can also be easily adapted to the known joint states setting. We conduct additional experiments on four real datasets for robot pose estimation with ground-truth joint states available. We simply replace the predicted joint state with the ground-truth joint state during the self-supervised training process. As shown in Tab. 3 our approach also achieves comparable performance with previous SOTA under this well-explored setting. + +# 4.5 Ablation study + +Ablation on network modules We first conduct an ablation study on the proposed network modules to investigate their effectiveness. Since JointNet and RotationNet are the major regression targets of our task, we focus on the ablation of two assistive network modules, DepthNet and KeypointNet. We conduct the experiments on the synthetic datasets, utilizing only ground-truth supervision to simplify the process. + +Firstly, we delve into the implementation of DepthNet. In Tab. 4, we evaluate the design choice of whether to share the backbone of DepthNet with other modules or not. The results indicate that separating the DepthNet backbone consistently improves performance. This improvement could be explained by the fact that DepthNet involves learning scale-variant features, while the other network modules all focus on learning scale-invariant features. We also investigate whether to pre-train the DepthNet or jointly train it with other modules + +Table 3: Comparison of AUC $\uparrow$ of the ADD distribution curve and Mean $\downarrow$ of the ADD, on the real-world datasets of DREAM, with all methods using known ground-truth joint state parameters. + +
MethodMetricPanda 3CAM-AKPanda 3CAM-XKPanda 3CAM-RSPanda ORB
DREAM-F18AUC↑68.924.476.161.9
DREAM-Q18AUC↑52.437.578.057.1
DREAM-H18AUC↑60.564.078.869.1
RoboPose15AUC↑76.586.076.980.5
CtRNet22AUC↑89.979.590.885.3
SGTAPose40AUC↑67.82.187.672.3
OursAUC↑90.281.291.987.6
DREAM-F18Mean↓11413.1491911.42077.495319.1
DREAM-Q18Mean↓78089.354178.227.264247.6
DREAM-H18Mean↓56.57381.623.625685.3
RoboPose15Mean↓24.214.023.119.4
CtRNet22Mean↓13.032.010.021.0
SGTAPose40Mean↓35.7157.712.835.0
OursMean↓12.919.911.015.9
+ +Table 4: Ablation studies on the effect of separating DepthNet image feature extractor. + +
DatasetSeparateAUC ↑Mean (mm)↓
Panda DR (a)-79.9821.9
Panda DR (b)+82.6818.0
Panda Photo (a)-79.6622.0
Panda Photo (b)+81.9818.4
+ +Table 5: Ablation studies on the effect of pre-training DepthNet in advance of jointly training. + +
DatasetPretrainAUC ↑Mean (mm)↓
Panda DR (a)-74.2926.8
Panda DR (b)+82.6818.0
Panda Photo (a)-73.3427.9
Panda Photo (b)+81.9818.4
+ +from scratch. Table 5 suggests that a pre-training stage leads to considerable performance improvement. This improvement likely occurs because a well-trained DepthNet reliably aids other modules in better training, as it determines the global translation of all joints. Additionally, we select the keypoint closest to the robot center as the root keypoint requiring depth prediction by DepthNet. Please refer to the S2.3 section of the appendix for rationale behind this selection. + +Moreover, we conduct ablation experiments to investigate the impact of KeypointNet on other network modules, namely JointNet and RotationNet. We utilize Mean Joint Error, which represents the mean error of the joint state parameter predictions, and the Mean Rotation Angle Error [15], which represents the mean error of the Euler angle predictions $(\theta_{x}^{pred}, \theta_{y}^{pred}, \theta_{z}^{pred})$ for the camera-to-robot rotation, to quantify the accuracy of the JointNet and RotationNet predictions, respectively. It is important to note that a Panda robot consists of seven revolute joints with angle parameters (representing one-DoF rotations) and one prismatic joint with a length parameter (representing one-DoF transla + +![](images/19d02b552fdc9ecf9b92395bac0cc91b86a44add76ceb99ba1f0b9e4674624f6.jpg) +Fig. 4: Qualitative comparison between our method and RoboPose [15] on both real and synthetic datasets. + +tion). Therefore, we calculate and present the Mean Joint Error separately for the revolute and prismatic joints of Panda. We present the detailed ablation results in the S2.1 section of the appendix, in which Tables 6 and 7 show that the presence of KeypointNet significantly improves the prediction accuracy of JointNet and RotationNet. By sharing a feature extractor with JointNet and RotationNet, KeypointNet guides the network towards learning pixel-aligned robot visual features, which further assists in estimating robot pose and joint states. + +Finally, we examine the overall effect of DepthNet and KeypointNet. To replace DepthNet, we implement a straightforward alternative, directly regressing the camera-to-robot translation $\mathbf{t}$ from image features $\mathbf{f}$ using an MLP, similar to RotationNet. Figure 5a illustrates that both modules play an important role in prediction accuracy. The DepthNet notably enhances accuracy by improving root depth estimation, a global offset that affects the precision of all joints. Meanwhile, KeypointNet further benefits the estimation results. + +The contribution of these modules to the overall performance validates our design. DepthNet disentangles depth from camera intrinsic, facilitating a learning-based approach to depth estimation. This method exhibits greater expressive capability compared to the analytical, non-parameterized PnP methods employed in previous work [18,22,40]. On the other hand, KeypointNet enhances both JointNet and RotationNet by steering the shared backbone towards learning more pixel-aligned features, thus complementing the learning objectives defined in the robot kinematic space. + +Ablation on self-supervision We also perform an ablation study of our self-supervised learning objectives. We conduct the experiments on the Panda real datasets. Figure 5b demonstrates that both the keypoint consistency regularization and the mask consistency regularization contribute to the overall performance. Specifically, the mask consistency regularization provides pixel-wise supervision signal with the help of differentiable rendering, while the keypoint consistency regularization unifies the predictions in the pixel space and that in the robot kinematic space. We suppose that pixel-aligned 2D tasks, e.g. keypoint detection and mask segmentation, generally generalize better, and could therefore improve the 3D estimation quality through consistency regularization, further improving the overall performance. As a result, our multi-stage training scheme yields favorable convergence ability. Please refer to the S4 section of the appendix for the error analysis of each module across each training stage. + +![](images/719af2c204af9ac0f711ce0a150a758f556cbdee2843edde3af6a75bcc74a3c9.jpg) +(a) Ablation study on network modules. 'Baseline' represents direct regression, 'DN' and 'KN' refer to DepthNet and KeypointNet, respectively. +Fig. 5: Ablation studies on network modules and self-supervision. + +![](images/e9262e5a7a8d0f0543a2c87f1cd39756565c3c08b6a00917d201889743b3b883.jpg) +(b) Ablation study on self-supervision. 'Baseline' represents no self-supervision, 'MC' and 'KC' refer to mask and keypoint consistency regularization, respectively. + +# 5 Conclusion + +This work presents a novel end-to-end framework for real-time holistic robot pose estimation without prior knowledge of internal robot states, providing substantial assistance in real-world scenarios where the joint state of the estimation target is unavailable or unreliable. By decomposing this complex task into specific sub-tasks and tailoring neural network modules for each, the method effectively addresses the challenges while retaining high computation efficiency. Our approach, which circumvents iterative optimization, significantly accelerates the estimation process and achieves state-of-the-art accuracy. This advancement holds promise for enhancing real-world robotics applications, offering a practical solution for real-time robot pose estimation and joint state monitoring. + +# Acknowledgments + +This work is supported by National Science and Technology Major Project (2022ZD0114904). We thank Hai Ci for the valuable discussions during the early stages of this work. We thank Haoran Lu and Ruihai Wu for their assistance with real-world image captures. + +# References + +1. Bultmann, S., Memmesheimer, R., Behnke, S.: External camera-based mobile robot pose estimation for collaborative perception with smart edge sensors. In: ICRA (2023) +2. Chen, L., Qin, Y., Zhou, X., Su, H.: EasyHeC: Accurate and Automatic Hand-eye Calibration via Differentiable Rendering and Space Exploration. In: IEEE Robotics and Automation Letters (2023) +3. Christen, S., Yang, W., Pérez-D'Arpino, C., Hilliges, O., Fox, D., Chao, Y.W.: Learning human-to-robot handovers from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9654–9664 (2023) +4. Deisenroth, M.P., Rasmussen, C.E., Fox, D.: Learning to control a low-cost manipulator using data-efficient reinforcement learning. Robotics: Science and Systems VII 7, 57-64 (2011) +5. Feniello, A., Dang, H.N., Birchfield, S.: Program synthesis by examples for object repositioning tasks. 2014 IEEE/RSJ International Conference on Intelligent Robots and Systems pp. 4428-4435 (2014) +6. Fiala, M.: Artag, a fiducial marker system using digital techniques. In: CVPR (2005) +7. Garrido-Jurado, S., Muñoz-Salinas, R., Madrid-Cuevas, F.J., Marín-Jiménez, M.J.: Automatic generation and detection of highly reliable fiducial markers under occlusion. Pattern Recognition (2014) +8. Hao, R., Özgün, O., Çavusoglu, M.C.: Vision-based surgical tool pose estimation for the da Vinci® robotic surgical system. In: 2018 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 1298-1305. IEEE (2018) +9. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016) +0. Icli, C., Stepanenko, O., Bonev, I.: New method and portable measurement device for the calibration of industrial robots. Sensors 20(20) (2020) +1. Kim, W., Kim, N., Lyons, J.B., Nam, C.S.: Factors affecting trust in high-vulnerability human-robot interaction contexts: A structural equation modelling approach. Applied Ergonomics 85, 103056 (2020) +2. Kok, B.C., Soh, H.: Trust in robots: Challenges and opportunities. Current Robotics Reports 1, 297-309 (2020) +3. Kothari, N., Gupta, M., Vachhani, L., Arya, H.: Pose estimation for an autonomous vehicle using monocular vision. In: 2017 Indian Control Conference (ICC). pp. 424-431 (2017) +4. Kuipers, B.: How can we trust a robot? Communications of the ACM 61(3), 86-95 (2018) +5. Labbé, Y., Carpentier, J., Aubry, M., Sivic, J.: Single-view robot pose and joint angle estimation via render & compare. In: Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR) (2021) + +16. Lambrecht, J., Grosenick, P., Meusel, M.: Optimizing keypoint-based single-shot camera-to-robot pose estimation through shape segmentation. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 13843-13849. IEEE (2021) +17. Lambrecht, J., Kästner, L.: Towards the usage of synthetic data for marker-less pose estimation of articulated robots in rgb images. In: ICAR (2019) +18. Lee, T.E., Tremblay, J., To, T., Cheng, J., Mosier, T., Kroemer, O., Fox, D., Birchfield, S.: Camera-to-robot pose estimation from a single image. In: International Conference on Robotics and Automation (ICRA) (2020) +19. Lepetit, V., Moreno-Noguer, F., Fua, P.: Epnp: An accurate o (n) solution to the pnp problem. IJCV (2009) +20. Li, S., De Wagter, C., De Croon, G.C.: Self-supervised monocular multi-robot relative localization with efficient deep neural networks. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 9689-9695. IEEE (2022) +21. Li, X., Zhang, X., Zhou, X., Chen, I.M.: Upg: 3d vision-based prediction framework for robotic grasping in multi-object scenes. Knowledge-Based Systems 270, 110491 (2023) +22. Lu, J., Richter, F., Yip, M.C.: Markerless camera-to-robot pose estimation via self-supervised sim-to-real transfer. In: Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +23. Mahler, J., Matl, M., Satish, V., Danielczuk, M., DeRose, B., McKinley, S., Goldberg, K.: Learning ambidextrous robot grasping policies. Science Robotics 4(26), eaau4984 (2019) +24. Masehian, E., Jannati, M., Hekmatfar, T.: Cooperative mapping of unknown environments by multiple heterogeneous mobile robots with limited sensing. Robotics and Autonomous Systems 87, 188-218 (2017) +25. Ming, Y., Meng, X., Fan, C., Yu, H.: Deep learning for monocular depth estimation: A review. Neurocomputing 438, 14-33 (2021) +26. Moon, G., Chang, J., Lee, K.M.: Camera distance-aware top-down approach for 3d multi-person pose estimation from a single rgb image. In: The IEEE Conference on International Conference on Computer Vision (ICCV) (2019) +27. Morrison, D., Corke, P., Leitner, J.: Closing the Loop for Robotic Grasping: A Real-time, Generative Grasp Synthesis Approach. In: Proc. of Robotics: Science and Systems (RSS) (2018) +28. Olson, E.: Apriltag: A robust and flexible visual fiducial system. In: ICRA (2011) +29. Papadimitriou, A., Mansouri, S.S., Nikolakopoulos, G.: Range-aided ego-centric collaborative pose estimation for multiple robots. Expert Systems with Applications 202, 117052 (2022) +30. Park, D., Seo, Y., Chun, S.Y.: Real-time, highly accurate robotic grasp detection using fully convolutional neural networks with high-resolution images (2019) +31. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch (2017) +32. Pinto, L., Davidson, J., Gupta, A.: Supervision via competition: Robot adversaries for learning tasks. In: 2017 IEEE International Conference on Robotics and Automation (ICRA). pp. 1601-1608 (2017) +33. Prusak, A., Melnychuk, O., Roth, H., Schiller, I., Koch, R.: Pose estimation and map building with a time-of-flight-camera for robot navigation. International Journal of Intelligent Systems Technologies and Applications 5(3-4), 355-364 (2008) +34. Rana, A., Vulpi, F., Galati, R., Milella, A., Petitti, A.: A pose estimation algorithm for agricultural mobile robots using an rgb-d camera. In: 2022 International Con- + +ference on Electrical, Computer, Communications and Mechatronics Engineering (ICECCME). pp. 1-5 (2022) +35. Rizk, Y., Awad, M., Tunstel, E.W.: Cooperative heterogeneous multi-robot systems. ACM Computing Surveys (CSUR) 52, 1 - 31 (2019) +36. Salvini, P., Paez-Granados, D.F., Billard, A.: Safety concerns emerging from robots navigating in crowded pedestrian areas. International Journal of Social Robotics 14, 441 - 462 (2021) +37. Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learning for human pose estimation. In: CVPR (2019) +38. Sun, X., Xiao, B., Wei, F., Liang, S., Wei, Y.: Integral human pose regression. In: Proceedings of the European Conference on Computer Vision (ECCV) (2018) +39. Svenstrup, M., Tranberg, S., Andersen, H.J., Bak, T.: Pose estimation and adaptive robot behaviour for human-robot interaction. In: 2009 IEEE International Conference on Robotics and Automation. pp. 3571-3576 (2009) +40. Tian, Y., Zhang, J., Yin, Z., Dong, H.: Robot structure prior guided temporal attention for camera-to-robot pose estimation from image sequence. In: Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +41. Tzafestas, S.G.: 15 - mobile robots at work. In: Tzafestas, S.G. (ed.) Introduction to Mobile Robot Control, pp. 635-663. Elsevier, Oxford (2014) +42. Uchibe, E., Asada, M., Noda, S., Takahashi, Y., Hosoda, K.: Vision-based reinforcement learning for robocup: Towards real robot competition. In: Proc. of IROS. vol. 96 (1996) +43. Xia, Y., Tang, Y., Hu, Y., Hoffman, G.: Privacy-preserving pose estimation for human-robot interaction. CoRR abs/2011.07387 (2020) +44. Xu, C., Yu, X., Wang, Z., Ou, L.: Multi-view human pose estimation in humanrobot interaction. In: IECON 2020 The 46th Annual Conference of the IEEE Industrial Electronics Society. pp. 4769-4775 (2020) +45. Xun, Z., Huang, J., Li, Z., Xu, C., Gao, F., Cao, Y.: Crepes: Cooperative relative pose estimation towards real-world multi-robot systems. arXiv preprint arXiv:2302.01036 (2023) +46. Yang, W., Paxton, C., Mousavian, A., Chao, Y.W., Cakmak, M., Fox, D.: Reactive human-to-robot handovers of arbitrary objects. In: IEEE International Conference on Robotics and Automation (ICRA). IEEE (2021) +47. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5745-5753 (2019) +48. Zhu, W., Ma, X., Ro, D., Ci, H., Zhang, J., Shi, J., Gao, F., Tian, Q., Wang, Y.: Human motion generation: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) +49. Zuo, Y., Qiu, W., Xie, L., Zhong, F., Wang, Y., Yuille, A.L.: Craves: Controlling robotic arm with a vision-based economic system. In: CVPR (2019) \ No newline at end of file diff --git a/2024/Real-time Holistic Robot Pose Estimation with Unknown States/images.zip b/2024/Real-time Holistic Robot Pose Estimation with Unknown States/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a7a733246c7f20d3adb51ee362fad51d0c9d7bd6 --- /dev/null +++ b/2024/Real-time Holistic Robot Pose Estimation with Unknown States/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d43ac30764811c25acc5193c7769363ae0df146f7a33ebf5ad0a7acc2778f4dc +size 427892 diff --git a/2024/Real-time Holistic Robot Pose Estimation with Unknown States/layout.json b/2024/Real-time Holistic Robot Pose Estimation with Unknown States/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..98b035a2a7c24277db141c92ed9cbd9208b45add --- /dev/null +++ b/2024/Real-time Holistic Robot Pose Estimation with Unknown States/layout.json @@ -0,0 +1,9671 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 142, + 111, + 473, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 111, + 473, + 146 + ], + "spans": [ + { + "bbox": [ + 142, + 111, + 473, + 146 + ], + "type": "text", + "content": "Real-time Holistic Robot Pose Estimation with Unknown States" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 179, + 167, + 435, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 167, + 435, + 194 + ], + "spans": [ + { + "bbox": [ + 179, + 167, + 435, + 194 + ], + "type": "text", + "content": "Shikun Ban1, Juling Fan1, Xiaoxuan Ma1, Wentao Zhu1,2, Yu Qiao5, and Yizhou Wang1,2,3,4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 156, + 201, + 457, + 257 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 156, + 201, + 457, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 201, + 457, + 224 + ], + "spans": [ + { + "bbox": [ + 156, + 201, + 457, + 224 + ], + "type": "text", + "content": "1 Center on Frontiers of Computing Studies, School of Computer Science, Peking University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 203, + 224, + 411, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 224, + 411, + 236 + ], + "spans": [ + { + "bbox": [ + 203, + 224, + 411, + 236 + ], + "type": "text", + "content": "2 Inst. for Artificial Intelligence, Peking University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 203, + 235, + 411, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 235, + 411, + 246 + ], + "spans": [ + { + "bbox": [ + 203, + 235, + 411, + 246 + ], + "type": "text", + "content": "3 Nat'l Eng. Research Center of Visual Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 203, + 246, + 407, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 246, + 407, + 257 + ], + "spans": [ + { + "bbox": [ + 203, + 246, + 407, + 257 + ], + "type": "text", + "content": "4 Nat'l Key Lab of General Artificial Intelligence" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 145, + 257, + 468, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 257, + 468, + 279 + ], + "spans": [ + { + "bbox": [ + 145, + 257, + 468, + 279 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 145, + 257, + 468, + 279 + ], + "type": "text", + "content": " Department of Automation, School of Electronics, Information and Electrical Engineering, Shanghai Jiao Tong University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 306, + 453, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 306, + 453, + 516 + ], + "spans": [ + { + "bbox": [ + 159, + 306, + 453, + 516 + ], + "type": "text", + "content": "Abstract. Estimating robot pose from RGB images is a crucial problem in computer vision and robotics. While previous methods have achieved promising performance, most of them presume full knowledge of robot internal states, e.g. ground-truth robot joint angles. However, this assumption is not always valid in practical situations. In real-world applications such as multi-robot collaboration or human-robot interaction, the robot joint states might not be shared or could be unreliable. On the other hand, existing approaches that estimate robot pose without joint state priors suffer from heavy computation burdens and thus cannot support real-time applications. This work introduces an efficient framework for real-time robot pose estimation from RGB images without requiring known robot states. Our method estimates camera-to-robot rotation, robot state parameters, keypoint locations, and root depth, employing a neural network module for each task to facilitate learning and sim-to-real transfer. Notably, it achieves inference in a single feed-forward pass without iterative optimization. Our approach offers a " + }, + { + "bbox": [ + 159, + 306, + 453, + 516 + ], + "type": "inline_equation", + "content": "12 \\times" + }, + { + "bbox": [ + 159, + 306, + 453, + 516 + ], + "type": "text", + "content": " speed increase with state-of-the-art accuracy, enabling real-time holistic robot pose estimation for the first time. Code and models are available at https://github.com/Oliverbansk/Holistic-Robot-Pose-Estimation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 552, + 230, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 552, + 230, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 552, + 230, + 563 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 575, + 482, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 575, + 482, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 575, + 482, + 638 + ], + "type": "text", + "content": "Robot pose estimation is a fundamental task in computer vision and robotics as it not only lays the groundwork for multiple downstream tasks, e.g. visually-guided object grasping and manipulation [21,40,49], but also paves the way for a variety of applications, such as autonomous navigation [1,13,33,34], human-robot interaction [3,39,43,44,46,48] and multi-robot collaboration [20,29,35,45]." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 140, + 642, + 482, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 642, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 140, + 642, + 482, + 666 + ], + "type": "text", + "content": "Correspondence to: Wentao Zhu and Yu Qiao ." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 117, + 478, + 223 + ], + "blocks": [ + { + "bbox": [ + 134, + 117, + 478, + 223 + ], + "lines": [ + { + "bbox": [ + 134, + 117, + 478, + 223 + ], + "spans": [ + { + "bbox": [ + 134, + 117, + 478, + 223 + ], + "type": "image", + "image_path": "400b567273e0a3757470601efe968e5aad3887625d0ab8d4e346edfc1927a0d6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 233, + 482, + 321 + ], + "lines": [ + { + "bbox": [ + 130, + 233, + 482, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 482, + 321 + ], + "type": "text", + "content": "Fig. 1: The majority of previous robot pose estimation methods assume known robot joint states and focus solely on estimating the camera-to-robot pose, i.e. camera-to-robot rotation and translation. In contrast, the holistic robot pose estimation problem requires estimating both joint states and the camera-to-robot pose, given only an RGB image without known joint states. For holistic robot pose estimation, RoboPose [15] uses costly test-time optimization (Render-and-Compare) to iteratively refine the predictions. In contrast, our feed-forward method achieves state-of-the-art accuracy with a " + }, + { + "bbox": [ + 130, + 233, + 482, + 321 + ], + "type": "inline_equation", + "content": "12 \\times" + }, + { + "bbox": [ + 130, + 233, + 482, + 321 + ], + "type": "text", + "content": " speed boost." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 339, + 482, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 482, + 389 + ], + "type": "text", + "content": "Nonetheless, most previous works for monocular robot pose estimation [17,18,22,40] rely on a known set of robot joint angles as an additional input to alleviate the difficulty of the estimation process. Yet, this prerequisite information of the robot state is not always available in practical scenarios:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 395, + 482, + 540 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 137, + 395, + 481, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 395, + 481, + 432 + ], + "spans": [ + { + "bbox": [ + 137, + 395, + 481, + 432 + ], + "type": "text", + "content": "- For instance, low-cost robotic arms without sophisticated sensors or internal monitors can not provide the visual calibration process with readily accessible robot state measurements [4,10,49]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 432, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 432, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 432, + 482, + 479 + ], + "type": "text", + "content": "- In the context of multi-robot collaboration, the real-time synchronization of joint state data may be hindered by environmental limitations [24], necessitating pose estimation without robot joint state information, and additional monitoring of the joint states." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 479, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 479, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 137, + 479, + 482, + 540 + ], + "type": "text", + "content": "- In domains such as human-robot interaction [36,41] and competitive robot gaming [32,42], the reliability of the state information from counterpart robots is not always guaranteed, as trust issues may arise in robotic systems [11,12,14]. In this sense, failing to obtain independent estimation of the counterpart robot joint state may lead to potential safety issues." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 546, + 482, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 630 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 630 + ], + "type": "text", + "content": "These scenarios further emphasize the urgent need for monocular robot pose estimation with unknown robot joint state. In this work, we address the challenge of Holistic Robot Pose Estimation with unknown internal states for articulated robots. Specifically, given a monocular RGB image, we aim to simultaneously solve for 6D robot pose (3D rotation and translation relative to the camera) and fine-grained robot joint state parameters (including joint angles for revolute joints and displacement for prismatic joints)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 630, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 667 + ], + "type": "text", + "content": "Nonetheless, learning holistic robot pose estimation without joint state information presents significant challenges. The primary challenge is the dual ambiguity of unknown joint states and 6D pose. Specifically, estimating the 6D pose" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "S. Ban et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "of an articulated robot is more complex than for a rigid object due to shape and scale variability from unknown joint states. Conversely, estimating joint states is complicated by the need for 2D-to-3D correspondences and the unknown 6D pose. This intertwined issue requires novel solutions to decouple joint states from 6D pose estimation. Furthermore, the sophisticated robot morphologies vary from model to model, requiring non-trivial model design. In addition, The robot parts are frequently invisible in the monocular image as a result of self-occlusion and truncation, posing more challenges for robust estimation. Finally, the domain gap between synthetic training data and real-world images introduces additional challenges for learning-based approaches." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 238, + 482, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 238, + 482, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 238, + 482, + 334 + ], + "type": "text", + "content": "It is worth noting that some previous works [15,49] have also addressed holistic estimation with unknown states, tackling some challenges. However, these methods all necessitate iterative optimization at inference, which is computationally intensive and unsuitable for real-time applications. Motivated by the need for efficient holistic robot pose estimation without knowledge of the robot state, we propose an end-to-end framework that directly estimates camera-to-robot pose and robot joint state parameters, requiring only a single feed-forward pass during inference." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 337, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 337, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 337, + 482, + 445 + ], + "type": "text", + "content": "In pursuit of this goal, we factorize the holistic robot pose estimation task into several sub-tasks, including the estimation of camera-to-robot rotation, robot joint states, root depth, and root-relative keypoint locations. We then design corresponding neural network modules for each sub-task. Specifically, the standalone root depth estimator disentangles the estimation of scale-variant and scale-invariant variables. The keypoint estimator bridges the gap between the pixel space and the robotic parameter space. Additionally, we utilize self-supervised consistency regularizations which could be performed on the real images without ground truth (GT)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 447, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 482, + 555 + ], + "type": "text", + "content": "We demonstrate that our designs of modular feed-forward networks improve model generalization while maintaining high computation efficiency. Consequently, our approach achieves state-of-the-art estimation accuracy with unknown robot states and a " + }, + { + "bbox": [ + 130, + 447, + 482, + 555 + ], + "type": "inline_equation", + "content": "12 \\times" + }, + { + "bbox": [ + 130, + 447, + 482, + 555 + ], + "type": "text", + "content": " speedup compared to the iterative Render-and-Compare (RnC) [15] approach, as depicted in Fig. 1. The key insight behind our approach is that we could get rid of the cumbersome RnC [15] and error-prone Perspective-n-Points (PnP) [18,22,40] schemes to establish 2D-3D correspondences. Instead, we demonstrate that feed-forward networks could well handle the problem with proper problem decomposition and meticulous design." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 146, + 558, + 347, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 558, + 347, + 569 + ], + "spans": [ + { + "bbox": [ + 146, + 558, + 347, + 569 + ], + "type": "text", + "content": "In summary, our contributions are three-fold:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 588, + 479, + 663 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 136, + 588, + 479, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 588, + 479, + 610 + ], + "spans": [ + { + "bbox": [ + 136, + 588, + 479, + 610 + ], + "type": "text", + "content": "1. We present an efficient end-to-end learning framework for the holistic robot pose estimation problem with unknown robot states." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 615, + 479, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 615, + 479, + 636 + ], + "spans": [ + { + "bbox": [ + 136, + 615, + 479, + 636 + ], + "type": "text", + "content": "2. We propose to factorize and decompose the holistic robot pose estimation problem into several sub-tasks and tackle them with respective modules." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 136, + 641, + 479, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 641, + 479, + 663 + ], + "spans": [ + { + "bbox": [ + 136, + 641, + 479, + 663 + ], + "type": "text", + "content": "3. We achieve state-of-the-art performance on various robot models while delivering a significant " + }, + { + "bbox": [ + 136, + 641, + 479, + 663 + ], + "type": "inline_equation", + "content": "12 \\times" + }, + { + "bbox": [ + 136, + 641, + 479, + 663 + ], + "type": "text", + "content": " speedup, paving the way for real-time applications." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 365, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 365, + 91, + 447, + 100 + ], + "type": "text", + "content": "Holistic Robot Pose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 136, + 266, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 136, + 266, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 136, + 266, + 148 + ], + "type": "text", + "content": "2.1 Hand-eye calibration" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 152, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 308 + ], + "type": "text", + "content": "The goal of Hand-Eye Calibration (HEC) is to determine the transformation between the position of a robot's end-effector (hand) and its vision system (eye). This task can be categorized into two types: eye-in-hand where the camera is fixed at the robot end-effector [27], and eye-to-hand which refers to the configurations where the camera is placed at a distance from the robot [5,23,30]. This work focuses on a similar setting with eye-to-hand. Traditional HEC methods involve attaching a fiducial marker to the end effector [6,7,28], placing it in several known positions, detecting the marker positions in images, and then solving for the camera-to-robot transformation. Despite its broad applicability, the classical HEC pipeline necessitates a complete offline calibration process every time the camera-robot transformation changes, which can be tedious in real-world applications. Consequently, recent works, such as DREAM [18], propose an online, markerless calibration problem, as discussed below." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 322, + 338, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 338, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 338, + 335 + ], + "type": "text", + "content": "2.2 Image-based robot pose estimation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 338, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 468 + ], + "type": "text", + "content": "Previous works can be divided into two categories based on whether the robot states are known or not. Most previous works estimate the 6D camera-to-robot pose with known robot states. For example, DREAM [18] first trains a 2D keypoint estimator and then applies Perspective-" + }, + { + "bbox": [ + 130, + 338, + 482, + 468 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 338, + 482, + 468 + ], + "type": "text", + "content": "-Point (PnP) [19] to obtain camera-to-robot pose, showing comparable results with the classical methods. Some work further utilizes segmentation masks and edges to aid pose estimation [8,16]. Chen et al. [2] apply rendering-based camera pose optimization and consistency-based joint space exploration. SGTAPose [40] leverages robot structure prior and temporal information. Lu et al. [22] propose to train on real-world data using foreground segmentation and differentiable rendering in a self-supervised manner." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 470, + 483, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 483, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 483, + 589 + ], + "type": "text", + "content": "Recently, several works have shifted to estimating robot pose with unknown joint states. Zuo et al. [49] estimate the robot pose and joint angles of a simple 4-DoF robotic arm from a single RGB image by solving a nonlinear nonconvex optimization problem. However, this method is not well-suited for robot arms with higher complexity. Labbé et al. [15] propose a render-and-compare approach to iteratively refine the estimation for both known and unknown states, and apply the method to different robot models up to 15-DoF. However, the iterative optimization process limits its applications in real-time scenarios. In contrast, our approach can be deployed with a single feed-forward at inference and achieves real-time efficiency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 604, + 202, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 604, + 202, + 616 + ], + "spans": [ + { + "bbox": [ + 132, + 604, + 202, + 616 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 625, + 209, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 625, + 209, + 636 + ], + "spans": [ + { + "bbox": [ + 132, + 625, + 209, + 636 + ], + "type": "text", + "content": "3.1 Overview" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 482, + 665 + ], + "type": "text", + "content": "Given an RGB image " + }, + { + "bbox": [ + 132, + 641, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 132, + 641, + 482, + 665 + ], + "type": "text", + "content": ", we employ a suite of neural network modules, including DepthNet, JointNet, RotationNet, and KeypointNet. These modules are used" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "S. Ban et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 115, + 481, + 285 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 481, + 285 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 481, + 285 + ], + "type": "image", + "image_path": "30c7f4508dca82009277bbbef491c2aa93c01e4d95b128047e0361da616dfe0a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 295, + 482, + 361 + ], + "lines": [ + { + "bbox": [ + 130, + 295, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 482, + 361 + ], + "type": "text", + "content": "Fig. 2: Framework overview. The JointNet and the RotationNet regress joint state parameters " + }, + { + "bbox": [ + 130, + 295, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 130, + 295, + 482, + 361 + ], + "type": "text", + "content": " and camera-to-robot rotation " + }, + { + "bbox": [ + 130, + 295, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 130, + 295, + 482, + 361 + ], + "type": "text", + "content": ", respectively. The KeypointNet estimates root-relative 3D keypoint locations " + }, + { + "bbox": [ + 130, + 295, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^r" + }, + { + "bbox": [ + 130, + 295, + 482, + 361 + ], + "type": "text", + "content": ". The DepthNet's estimation of root depth " + }, + { + "bbox": [ + 130, + 295, + 482, + 361 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 295, + 482, + 361 + ], + "type": "text", + "content": " is combined with " + }, + { + "bbox": [ + 130, + 295, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^r" + }, + { + "bbox": [ + 130, + 295, + 482, + 361 + ], + "type": "text", + "content": " to acquire absolute 3D keypoint locations " + }, + { + "bbox": [ + 130, + 295, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{P}'" + }, + { + "bbox": [ + 130, + 295, + 482, + 361 + ], + "type": "text", + "content": " and camera-to-robot translation " + }, + { + "bbox": [ + 130, + 295, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 130, + 295, + 482, + 361 + ], + "type": "text", + "content": ". Joint state parameters " + }, + { + "bbox": [ + 130, + 295, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 130, + 295, + 482, + 361 + ], + "type": "text", + "content": ", rotation " + }, + { + "bbox": [ + 130, + 295, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 130, + 295, + 482, + 361 + ], + "type": "text", + "content": " and translation " + }, + { + "bbox": [ + 130, + 295, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 130, + 295, + 482, + 361 + ], + "type": "text", + "content": " are used to compute 3D keypoint locations " + }, + { + "bbox": [ + 130, + 295, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 130, + 295, + 482, + 361 + ], + "type": "text", + "content": " via forward-kinematics." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 393, + 482, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 393, + 482, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 393, + 482, + 478 + ], + "type": "text", + "content": "to respectively estimate the absolute root depth " + }, + { + "bbox": [ + 130, + 393, + 482, + 478 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 393, + 482, + 478 + ], + "type": "text", + "content": ", joint state parameters " + }, + { + "bbox": [ + 130, + 393, + 482, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 130, + 393, + 482, + 478 + ], + "type": "text", + "content": ", camera-to-robot rotation " + }, + { + "bbox": [ + 130, + 393, + 482, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in SO(3)" + }, + { + "bbox": [ + 130, + 393, + 482, + 478 + ], + "type": "text", + "content": ", and root-relative 3D keypoint locations " + }, + { + "bbox": [ + 130, + 393, + 482, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^r \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 130, + 393, + 482, + 478 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 393, + 482, + 478 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 393, + 482, + 478 + ], + "type": "text", + "content": " represents the total number of keypoints. The keypoints are defined at the joint articulations, in line with previous works [15,18,22], and one keypoint is designated as the robot root. The root joint is critical for computing the camera-to-robot rotation " + }, + { + "bbox": [ + 130, + 393, + 482, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 130, + 393, + 482, + 478 + ], + "type": "text", + "content": " and translation " + }, + { + "bbox": [ + 130, + 393, + 482, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 130, + 393, + 482, + 478 + ], + "type": "text", + "content": ", as well as serving as the origin for the forward kinematics (FK) process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 482, + 482, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 482, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 482, + 482, + 590 + ], + "type": "text", + "content": "Our modularized design offers several key insights. Firstly, DepthNet is tailored to learn the scale-variant camera-robot spatial relationship, while the other modules concentrate on the scale-invariant robot features. Secondly, while the joint state parameters " + }, + { + "bbox": [ + 130, + 482, + 482, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 130, + 482, + 482, + 590 + ], + "type": "text", + "content": " provide a compact representation of robot morphologies subject to robotic structure constraints, they are highly abstract and thus challenging to directly regress from image features. To address this, we utilize the 3D keypoint locations as an intermediary. On the one hand, they can be effectively estimated with pixel-aligned image features; on the other hand, they are inherently linked to " + }, + { + "bbox": [ + 130, + 482, + 482, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 130, + 482, + 482, + 590 + ], + "type": "text", + "content": " through an FK process." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "An overview of the proposed framework is illustrated in Fig. 2. Specifically, DepthNet estimates the absolute root depth " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", which, combined with KeypointNet's output " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^r \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", provides the absolute 3D keypoint locations " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{P}' \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ". The root keypoint location in " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{P}'" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " is the camera-to-robot translation " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ". The estimated joint state parameters " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", along with the camera-to-robot rotation " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " and translation " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", are then used to perform FK, yielding the final" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 365, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 365, + 91, + 447, + 100 + ], + "type": "text", + "content": "Holistic Robot Pose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 139 + ], + "type": "text", + "content": "3D keypoint estimations " + }, + { + "bbox": [ + 130, + 115, + 479, + 139 + ], + "type": "inline_equation", + "content": "\\mathbf{P} \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 130, + 115, + 479, + 139 + ], + "type": "text", + "content": ". Please refer to the appendix for a detailed illustration of the relations of these variables." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 178 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 178 + ], + "type": "text", + "content": "In the following, we will first introduce the architecture design of the network modules in Sec. 3.2, and then explain the training loss design consisting of ground-truth supervision and self-supervision in Sec. 3.3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 197, + 225, + 208 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 197, + 225, + 208 + ], + "spans": [ + { + "bbox": [ + 132, + 197, + 225, + 208 + ], + "type": "text", + "content": "3.2 Architecture" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 219, + 482, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 219, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 219, + 482, + 316 + ], + "type": "text", + "content": "DepthNet Monocular object depth estimation suffers from significant ambiguities, as the absolute depth value and camera intrinsics are deeply intertwined [25]. For instance, a robot in a fixed pose relative to the camera may appear larger in photos taken with cameras having larger focal lengths. Scaling both the absolute depth and focal lengths equally results in identical object size and appearance in different images. To address this challenge, we implement DepthNet following the approach in [26] to disentangle depth from camera intrinsics. Initially, we use camera intrinsics to calculate a normalized, coarse depth value " + }, + { + "bbox": [ + 130, + 219, + 482, + 316 + ], + "type": "inline_equation", + "content": "d_{c}" + }, + { + "bbox": [ + 130, + 219, + 482, + 316 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 260, + 326, + 481, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 326, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 260, + 326, + 481, + 352 + ], + "type": "interline_equation", + "content": "d _ {c} = \\sqrt {\\frac {f _ {x} \\cdot f _ {y} \\cdot A _ {r e a l}}{A _ {b b o x}}}, \\tag {1}", + "image_path": "1906cf0c19eccbefe57e2ec8a824bc65a21aa0d01c60183b2cde402fc87b368b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "inline_equation", + "content": "f_{x}" + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "inline_equation", + "content": "f_{y}" + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "text", + "content": " represent the focal lengths along the x- and y-axes, " + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "inline_equation", + "content": "A_{real}" + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "text", + "content": " is the area of the robot in the x-y plane of the camera frame (in " + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "inline_equation", + "content": "mm^2" + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "text", + "content": "), and " + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "inline_equation", + "content": "A_{bbox}" + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "text", + "content": " is the area of the robot's bounding box (in " + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "inline_equation", + "content": "pixel^2" + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "text", + "content": "). This method of calculating " + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "inline_equation", + "content": "d_{c}" + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "text", + "content": " normalizes it in 3D space and provides an approximate depth estimation for subsequent refinement. The primary function of DepthNet is to predict a correction factor " + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "text", + "content": " based on the input image. We multiply the estimated " + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "inline_equation", + "content": "d_{c}" + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "text", + "content": " to obtain the final absolute root depth " + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 285, + 456, + 480, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 285, + 456, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 285, + 456, + 480, + 467 + ], + "type": "interline_equation", + "content": "d = \\lambda \\cdot d _ {c}. \\tag {2}", + "image_path": "e9df6cad89f1e4d5d776082ee19a03d778ceacf2d7722267616fcc5f93c2bab9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 477, + 482, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 477, + 482, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 477, + 482, + 550 + ], + "type": "text", + "content": "With a fixed camera, the relationship between the normalized coarse depth " + }, + { + "bbox": [ + 130, + 477, + 482, + 550 + ], + "type": "inline_equation", + "content": "d_{c}" + }, + { + "bbox": [ + 130, + 477, + 482, + 550 + ], + "type": "text", + "content": " and the robot's size in the image is linear. However, robots can assume poses with the same depth but different sizes in the image, making " + }, + { + "bbox": [ + 130, + 477, + 482, + 550 + ], + "type": "inline_equation", + "content": "d_{c}" + }, + { + "bbox": [ + 130, + 477, + 482, + 550 + ], + "type": "text", + "content": " inadequate for accurate root depth representation. The correction factor " + }, + { + "bbox": [ + 130, + 477, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 477, + 482, + 550 + ], + "type": "text", + "content": " introduces non-linear flexibility to the depth estimation, addressing this issue by adjusting " + }, + { + "bbox": [ + 130, + 477, + 482, + 550 + ], + "type": "inline_equation", + "content": "d_{c}" + }, + { + "bbox": [ + 130, + 477, + 482, + 550 + ], + "type": "text", + "content": " based on image features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "JointNet The goal of JointNet is to estimate joint state parameters " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{q} = (q_{1}, q_{2}, \\dots, q_{J})" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " denotes the number of robot joints. As this work deals with 1-DoF joints (revolute or prismatic), " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " equals the robot's DoF. JointNet shares the same feature extractor with RotationNet and KeypointNet. We extract feature map " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{F} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " denotes the number of channels, and " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " denotes the spatial resolution) and use global average pooling to obtain " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{f} \\in \\mathbb{R}^{C}" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": ". Finally, we employ several MLP blocks to regress the joint state parameters " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "S. Ban et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 155 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 155 + ], + "type": "text", + "content": "RotationNet The goal of RotationNet is to estimate the camera-to-robot rotation " + }, + { + "bbox": [ + 131, + 116, + 482, + 155 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in SO(3)" + }, + { + "bbox": [ + 131, + 116, + 482, + 155 + ], + "type": "text", + "content": ". We employ an MLP to estimate " + }, + { + "bbox": [ + 131, + 116, + 482, + 155 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 131, + 116, + 482, + 155 + ], + "type": "text", + "content": " from the image feature " + }, + { + "bbox": [ + 131, + 116, + 482, + 155 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 131, + 116, + 482, + 155 + ], + "type": "text", + "content": ". We utilize the 6D continuous rotation representation [47] to parameterize " + }, + { + "bbox": [ + 131, + 116, + 482, + 155 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 131, + 116, + 482, + 155 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 167, + 482, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 167, + 482, + 251 + ], + "spans": [ + { + "bbox": [ + 131, + 167, + 482, + 251 + ], + "type": "text", + "content": "KeypointNet The goal of KeypointNet is to estimate root-relative 3D keypoint locations " + }, + { + "bbox": [ + 131, + 167, + 482, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^r\\in \\mathbb{R}^{N\\times 3}" + }, + { + "bbox": [ + 131, + 167, + 482, + 251 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 131, + 167, + 482, + 251 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 167, + 482, + 251 + ], + "type": "text", + "content": " robot keypoints. We transform the feature map " + }, + { + "bbox": [ + 131, + 167, + 482, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 131, + 167, + 482, + 251 + ], + "type": "text", + "content": " to a 3D heatmap " + }, + { + "bbox": [ + 131, + 167, + 482, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{H}\\in \\mathbb{R}^{N\\times D\\times H'\\times W'}" + }, + { + "bbox": [ + 131, + 167, + 482, + 251 + ], + "type": "text", + "content": " through convolutional layers, where " + }, + { + "bbox": [ + 131, + 167, + 482, + 251 + ], + "type": "inline_equation", + "content": "D\\times H'\\times W'" + }, + { + "bbox": [ + 131, + 167, + 482, + 251 + ], + "type": "text", + "content": " is the spatial resolution discretizing the 3D space. The 3D heatmap encodes per-voxel likelihood of containing each robot keypoint. We derive the root-relative 3D positions of each keypoint " + }, + { + "bbox": [ + 131, + 167, + 482, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_n^r" + }, + { + "bbox": [ + 131, + 167, + 482, + 251 + ], + "type": "text", + "content": " by performing spatial integral techniques [38] over the corresponding heatmap " + }, + { + "bbox": [ + 131, + 167, + 482, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_n" + }, + { + "bbox": [ + 131, + 167, + 482, + 251 + ], + "type": "text", + "content": " in a differentiable manner:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 228, + 257, + 481, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 257, + 481, + 291 + ], + "spans": [ + { + "bbox": [ + 228, + 257, + 481, + 291 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {n} ^ {r} = \\sum_ {k = 1} ^ {D} \\sum_ {i = 1} ^ {H ^ {\\prime}} \\sum_ {j = 1} ^ {W ^ {\\prime}} (k, i, j) \\cdot \\mathbf {H} _ {n} (k, i, j). \\tag {3}", + "image_path": "ec2c2a6ed3c99fcab825ae6658c6705de5c3b48d24a742169f1401865022178f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 296, + 482, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 296, + 482, + 333 + ], + "spans": [ + { + "bbox": [ + 131, + 296, + 482, + 333 + ], + "type": "text", + "content": "By combining " + }, + { + "bbox": [ + 131, + 296, + 482, + 333 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^r" + }, + { + "bbox": [ + 131, + 296, + 482, + 333 + ], + "type": "text", + "content": " with the estimated root depth " + }, + { + "bbox": [ + 131, + 296, + 482, + 333 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 131, + 296, + 482, + 333 + ], + "type": "text", + "content": ", we obtain the regressed 3D keypoint locations " + }, + { + "bbox": [ + 131, + 296, + 482, + 333 + ], + "type": "inline_equation", + "content": "\\mathbf{P}'" + }, + { + "bbox": [ + 131, + 296, + 482, + 333 + ], + "type": "text", + "content": " and root translation " + }, + { + "bbox": [ + 131, + 296, + 482, + 333 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 131, + 296, + 482, + 333 + ], + "type": "text", + "content": ". We further calculate the FK-based 3D keypoint locations " + }, + { + "bbox": [ + 131, + 296, + 482, + 333 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 131, + 296, + 482, + 333 + ], + "type": "text", + "content": " with differentiable FK parameterized by " + }, + { + "bbox": [ + 131, + 296, + 482, + 333 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 131, + 296, + 482, + 333 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 296, + 482, + 333 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 131, + 296, + 482, + 333 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 296, + 482, + 333 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 131, + 296, + 482, + 333 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 346, + 227, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 346, + 227, + 360 + ], + "spans": [ + { + "bbox": [ + 131, + 346, + 227, + 360 + ], + "type": "text", + "content": "3.3 Training loss" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 364, + 482, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 364, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 131, + 364, + 482, + 401 + ], + "type": "text", + "content": "Ground-truth supervision For training on synthetic datasets where we have access to all the ground-truth information, we apply separate supervision for all the sub-tasks to train the corresponding network modules:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 267, + 411, + 481, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 411, + 481, + 425 + ], + "spans": [ + { + "bbox": [ + 267, + 411, + 481, + 425 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d e p t h}} = \\| d - \\hat {d} \\| _ {1}, \\tag {4}", + "image_path": "6832166968634bd54e1ef3d41a8ae9037652baa8cc33f1313c6c7d78278199c9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 268, + 440, + 480, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 440, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 268, + 440, + 480, + 453 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {j o i n t}} = \\| \\mathbf {q} - \\hat {\\mathbf {q}} \\| _ {2}, \\tag {5}", + "image_path": "5affdea7b3ac2063412c18202545ef626af2c73441b2abbc27d05496c1a29659.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 268, + 467, + 480, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 467, + 480, + 480 + ], + "spans": [ + { + "bbox": [ + 268, + 467, + 480, + 480 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r o t}} = \\left\\| \\mathbf {R} - \\hat {\\mathbf {R}} \\right\\| _ {2}, \\tag {6}", + "image_path": "179d967ec5027b14cab8ac5b126bc951e98c8383ba886305a94c7b7a23dbc241.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 268, + 495, + 480, + 508 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 495, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 268, + 495, + 480, + 508 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t r a n s}} = \\| \\mathbf {t} - \\hat {\\mathbf {t}} \\| _ {2}, \\tag {7}", + "image_path": "66afcadc04db3daaaf8d9d50a5a3cf7462053b48ddd81aca8503a974bbf45c88.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 241, + 522, + 480, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 522, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 241, + 522, + 480, + 536 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {k p t s}} = \\| \\mathbf {P} - \\hat {\\mathbf {P}} \\| _ {2} + \\| \\mathbf {p} - \\hat {\\mathbf {p}} \\| _ {2}, \\tag {8}", + "image_path": "28cf50b31d2c851124c8b4cae12985b85c7b51980b83583ecda53678bfdf8e22.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 235, + 548, + 480, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 548, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 235, + 548, + 480, + 564 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {k p t s}} ^ {\\prime} = \\left\\| \\mathbf {P} ^ {\\prime} - \\hat {\\mathbf {P}} ^ {\\prime} \\right\\| _ {2} + \\left\\| \\mathbf {p} ^ {\\prime} - \\hat {\\mathbf {p}} ^ {\\prime} \\right\\| _ {2}, \\tag {9}", + "image_path": "ad791de4c44626c0e451cba8a6053c81743809a2221d3464efcb22e1434f7278.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 568, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 568, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 131, + 568, + 482, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 568, + 482, + 616 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 131, + 568, + 482, + 616 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 568, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 131, + 568, + 482, + 616 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 568, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 131, + 568, + 482, + 616 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 568, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 131, + 568, + 482, + 616 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 568, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 131, + 568, + 482, + 616 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 568, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{P}'" + }, + { + "bbox": [ + 131, + 568, + 482, + 616 + ], + "type": "text", + "content": " are GT root depth, joint states, camera-to-robot rotation and translation, FK-based keypoint locations and regressed keypoint locations, respectively; " + }, + { + "bbox": [ + 131, + 568, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 131, + 568, + 482, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 568, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{p}'" + }, + { + "bbox": [ + 131, + 568, + 482, + 616 + ], + "type": "text", + "content": " are 2D projections of " + }, + { + "bbox": [ + 131, + 568, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 131, + 568, + 482, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 568, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{P}'" + }, + { + "bbox": [ + 131, + 568, + 482, + 616 + ], + "type": "text", + "content": " using the camera intrinsics; " + }, + { + "bbox": [ + 131, + 568, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\hat{}" + }, + { + "bbox": [ + 131, + 568, + 482, + 616 + ], + "type": "text", + "content": " denotes model estimations." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 146, + 616, + 397, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 616, + 397, + 629 + ], + "spans": [ + { + "bbox": [ + 146, + 616, + 397, + 629 + ], + "type": "text", + "content": "We then train all the network modules end-to-end using" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 203, + 634, + 480, + 650 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 634, + 480, + 650 + ], + "spans": [ + { + "bbox": [ + 203, + 634, + 480, + 650 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {G T}} = \\mathcal {L} _ {\\text {j o i n t}} + \\mathcal {L} _ {\\text {r o t}} + \\mathcal {L} _ {\\text {t r a n s}} + \\lambda \\left(\\mathcal {L} _ {\\text {k p t s}} + \\mathcal {L} _ {\\text {k p t s}} ^ {\\prime}\\right), \\tag {10}", + "image_path": "a7fde8d75934873c75a8530a31aea96636d1461a443fd7e859d3b31022915a05.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 653, + 347, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 653, + 347, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 653, + 347, + 665 + ], + "type": "text", + "content": "where we set " + }, + { + "bbox": [ + 131, + 653, + 347, + 665 + ], + "type": "inline_equation", + "content": "\\lambda = 10.0" + }, + { + "bbox": [ + 131, + 653, + 347, + 665 + ], + "type": "text", + "content": " to balance the loss terms." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 365, + 91, + 448, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 91, + 448, + 100 + ], + "spans": [ + { + "bbox": [ + 365, + 91, + 448, + 100 + ], + "type": "text", + "content": "Holistic Robot Pose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "Self-supervision Apart from GT supervision on synthetic data, our framework also enables scaling the training pipeline to unlabeled real-world data in a self-supervised manner [22]. We employ two consistency-based training regularizations to improve model generalization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 164, + 480, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 164, + 480, + 190 + ], + "spans": [ + { + "bbox": [ + 131, + 164, + 480, + 190 + ], + "type": "text", + "content": "Despite the absence of GT joint states and keypoint locations, we regularize the consistency of the FK-based keypoints " + }, + { + "bbox": [ + 131, + 164, + 480, + 190 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}" + }, + { + "bbox": [ + 131, + 164, + 480, + 190 + ], + "type": "text", + "content": " and regressed keypoints " + }, + { + "bbox": [ + 131, + 164, + 480, + 190 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}'" + }, + { + "bbox": [ + 131, + 164, + 480, + 190 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 269, + 200, + 480, + 214 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 200, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 269, + 200, + 480, + 214 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {k c}} = \\left\\| \\hat {\\mathbf {P}} - \\hat {\\mathbf {P}} ^ {\\prime} \\right\\| _ {2}. \\tag {11}", + "image_path": "2214dc84e6c5b0d698a01b336e969ec2d846198c5d969ae7f76701bef345b58f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 224, + 480, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 224, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 131, + 224, + 480, + 285 + ], + "type": "text", + "content": "We introduce this keypoint consistency regularization to effectively reconcile the predictions subject to robot structure constraints " + }, + { + "bbox": [ + 131, + 224, + 480, + 285 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}" + }, + { + "bbox": [ + 131, + 224, + 480, + 285 + ], + "type": "text", + "content": " and predictions based on the pixel-aligned feature " + }, + { + "bbox": [ + 131, + 224, + 480, + 285 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}'" + }, + { + "bbox": [ + 131, + 224, + 480, + 285 + ], + "type": "text", + "content": ". In addition, similar to [22], we also regularize the consistency between robot rendering mask " + }, + { + "bbox": [ + 131, + 224, + 480, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{render}}" + }, + { + "bbox": [ + 131, + 224, + 480, + 285 + ], + "type": "text", + "content": " and foreground segmentation mask " + }, + { + "bbox": [ + 131, + 224, + 480, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{seg}}" + }, + { + "bbox": [ + 131, + 224, + 480, + 285 + ], + "type": "text", + "content": " using their Intersection-over-Union (IoU):" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 257, + 295, + 480, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 295, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 257, + 295, + 480, + 319 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {m c}} = 1 - \\frac {\\mathcal {S} _ {\\text {i n t e r s e c t i o n}}}{\\mathcal {S} _ {\\text {u n i o n}}}, \\tag {12}", + "image_path": "3d0cf85b0a0a45d15016e0830f7410b40c02353089dcd37e26f5f0f03085b0d4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 324, + 481, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 324, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 131, + 324, + 481, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 324, + 481, + 361 + ], + "type": "inline_equation", + "content": "S_{\\text{intersection}}" + }, + { + "bbox": [ + 131, + 324, + 481, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 324, + 481, + 361 + ], + "type": "inline_equation", + "content": "S_{\\text{union}}" + }, + { + "bbox": [ + 131, + 324, + 481, + 361 + ], + "type": "text", + "content": " is the area of the intersection and the union of " + }, + { + "bbox": [ + 131, + 324, + 481, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\text{render}}" + }, + { + "bbox": [ + 131, + 324, + 481, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 324, + 481, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\text{seg}}" + }, + { + "bbox": [ + 131, + 324, + 481, + 361 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 131, + 324, + 481, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\text{render}}" + }, + { + "bbox": [ + 131, + 324, + 481, + 361 + ], + "type": "text", + "content": " is obtained with differentiable rendering using " + }, + { + "bbox": [ + 131, + 324, + 481, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 131, + 324, + 481, + 361 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 324, + 481, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 131, + 324, + 481, + 361 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 324, + 481, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 131, + 324, + 481, + 361 + ], + "type": "text", + "content": " [15]; " + }, + { + "bbox": [ + 131, + 324, + 481, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\text{seg}}" + }, + { + "bbox": [ + 131, + 324, + 481, + 361 + ], + "type": "text", + "content": " is obtained with an image segmentation model [22]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 146, + 361, + 406, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 361, + 406, + 372 + ], + "spans": [ + { + "bbox": [ + 146, + 361, + 406, + 372 + ], + "type": "text", + "content": "The overall self-supervision training objective is as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 261, + 384, + 480, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 384, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 261, + 384, + 480, + 396 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s e l f}} = \\mathcal {L} _ {\\mathrm {k c}} + \\lambda_ {\\mathrm {m c}} \\mathcal {L} _ {\\mathrm {m c}} \\tag {13}", + "image_path": "ce092f075627a42dd6fff5857630b30c04a07a6b33d51a54f3b74d431957f050.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 405, + 479, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 405, + 479, + 430 + ], + "spans": [ + { + "bbox": [ + 131, + 405, + 479, + 430 + ], + "type": "text", + "content": "where we set " + }, + { + "bbox": [ + 131, + 405, + 479, + 430 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{mc}} = 1.0" + }, + { + "bbox": [ + 131, + 405, + 479, + 430 + ], + "type": "text", + "content": ". We apply end-to-end self-supervised training for the sim-to-real domain adaptation stage." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 449, + 230, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 230, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 230, + 464 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 476, + 263, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 476, + 263, + 487 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 263, + 487 + ], + "type": "text", + "content": "4.1 Dataset and metrics" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "We utilize the benchmark dataset DREAM [18], for both training and testing in our study. This dataset comprises images from three widely used robotic arms, namely, Franka Emika Panda (Panda), Kuka iiwa7 (Kuka), and Rethink Robotics Baxter (Baxter). The training dataset for each robot consists of approximately 100k synthetic images generated using the domain randomization (DR) technique. These images encompass scenarios with varying camera-to-robot poses and joint states. The testing datasets include both DR-generate synthetic images for each robot and photorealistic synthetic images (denoted as \"Photo\") for Panda and Kuka. Additionally, real-world image datasets of Panda are provided. Panda-3Cam comprises a total of 17k image frames captured by three different cameras, namely Azure Kinect (AK), XBOX 360 Kinect (XK), and Realsense (RS). This dataset is presented as three sub-datasets, each corresponding to one camera. Another real-world dataset, Panda-ORB, contains 32,315 images captured from 27 viewpoints using a Realsense camera." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "S. Ban et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "We evaluate the accuracy of our estimations using the Average Distance (ADD) metric, which represents the average Euclidean distances (in millimeters) between predicted 3D keypoint locations and their corresponding ground truth locations. Our reported metrics include both the Area Under the Curve (AUC) value and the mean value of ADD. The AUC integrates ADD over different thresholds (with the maximum threshold set at " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "100\\mathrm{mm}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "), while the mean value effectively represents the general prediction accuracy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 221, + 278, + 234 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 221, + 278, + 234 + ], + "spans": [ + { + "bbox": [ + 132, + 221, + 278, + 234 + ], + "type": "text", + "content": "4.2 Implementation details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 245, + 483, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 245, + 483, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 483, + 376 + ], + "type": "text", + "content": "We first pre-train DepthNet on the synthetic training dataset for 100 epochs, employing a learning rate of " + }, + { + "bbox": [ + 130, + 245, + 483, + 376 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 130, + 245, + 483, + 376 + ], + "type": "text", + "content": " and relying solely on ground truth supervision for depth. Subsequently, we conduct training for our entire model based on the pretrained DepthNet for an additional 100 epochs with a learning rate of " + }, + { + "bbox": [ + 130, + 245, + 483, + 376 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 130, + 245, + 483, + 376 + ], + "type": "text", + "content": " and a decay rate of 0.95. The Adam optimizer is employed in each stage to optimize the network parameters, with the momentum set to 0.9. As depth influences the global offset of all keypoints, a pre-trained DepthNet serves as a reasonable starting point for training other networks. On the Panda real datasets, we further apply self-supervised training on real-world images to overcome the sim-to-real domain gap with a learning rate of " + }, + { + "bbox": [ + 130, + 245, + 483, + 376 + ], + "type": "inline_equation", + "content": "1e - 6" + }, + { + "bbox": [ + 130, + 245, + 483, + 376 + ], + "type": "text", + "content": ". This fine-tuning process uses only image data and no ground-truth labels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 398, + 302, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 398, + 302, + 411 + ], + "spans": [ + { + "bbox": [ + 132, + 398, + 302, + 411 + ], + "type": "text", + "content": "4.3 Evaluation and Comparison" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 134, + 507, + 482, + 641 + ], + "blocks": [ + { + "bbox": [ + 130, + 453, + 482, + 498 + ], + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 498 + ], + "type": "text", + "content": "Table 1: Comparison of AUC " + }, + { + "bbox": [ + 130, + 453, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 130, + 453, + 482, + 498 + ], + "type": "text", + "content": " of the ADD distribution curve and mean " + }, + { + "bbox": [ + 130, + 453, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 130, + 453, + 482, + 498 + ], + "type": "text", + "content": " of the ADD, on each dataset in the DREAM datasets. Panda 3CAM datasets and Panda ORB are real-world datasets and the rest are synthetic datasets. † denotes using ground-truth joint state parameters." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 507, + 482, + 641 + ], + "lines": [ + { + "bbox": [ + 134, + 507, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 134, + 507, + 482, + 641 + ], + "type": "table", + "html": "
MethodMetricBaxter DRKuka DRKuka PhotoPanda DRPanda PhotoPanda 3CAM-AKPanda 3CAM-XKPanda 3CAM-RSPanda ORB
† DREAM-F 18AUC ↑---81.379.568.924.476.161.9
† DREAM-Q 18AUC ↑75.5--77.874.352.437.578.057.1
† DREAM-H 18AUC ↑-73.372.182.981.160.564.078.869.1
RoboPose 15AUC ↑32.780.273.282.979.770.477.674.370.4
OursAUC ↑58.875.173.982.782.082.276.075.275.2
† DREAM-F 18Mean ↓---552.0785.811413.1491911.42077.495319.1
† DREAM-Q 18Mean ↓59.0--359.0569.078089.354178.227.264247.6
† DREAM-H 18Mean ↓-5613.211099.4325.6518.856.57381.623.625685.3
RoboPose 15Mean ↓87.934.289.021.432.234.322.326.030.1
OursMean ↓45.028.029.018.018.418.924.024.824.8
", + "image_path": "ea71af5dc6561dc3230bb46287923c5aa5352473126706983888768e556c09a2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 365, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 365, + 91, + 447, + 100 + ], + "type": "text", + "content": "Holistic Robot Pose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 485, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 485, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 485, + 357 + ], + "type": "text", + "content": "Accuracy We first quantitatively evaluate the estimation accuracy of our method and compare it with the state-of-the-art (SOTA) approaches. To the best of our knowledge, among prior works, only RoboPose [15] addresses a scenario similar to ours, where both the pose and joint state parameters for an industry-level robot are estimated from a single RGB image. We also present the results of DREAM [18], a representative feed-forward method using ground-truth joint state parameters, as a reference. Table [1] presents the comparisons of AUC and the mean value of ADD. Our method outperforms SOTA on most datasets, significantly reducing joint localization errors. Specifically, DREAM [18] detects the 2D keypoint locations first, then employs a PnP process to solve the 2D-to-3D transformation. However, once the 2D keypoint detector fails due to truncations and occlusions, the PnP process produces unreliable numerical solutions. RoboPose [15] utilizes iterative render-and-compare, which is prone to falling into local optima on complex robot models with symmetric structures (e.g., Baxter). In contrast, our approach directly regresses both the robot pose and joint state parameters, enabling end-to-end optimization from multi-facet supervision signals. We also visualize the complete ADD distribution. As shown in Fig. [3], our approach achieves higher accuracy across most ADD thresholds. Notably, our approach attains near " + }, + { + "bbox": [ + 130, + 116, + 485, + 357 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 116, + 485, + 357 + ], + "type": "text", + "content": " accuracy with a " + }, + { + "bbox": [ + 130, + 116, + 485, + 357 + ], + "type": "inline_equation", + "content": "60\\mathrm{mm}" + }, + { + "bbox": [ + 130, + 116, + 485, + 357 + ], + "type": "text", + "content": " threshold, indicating more reliable estimation results on challenging cases with fewer outliers." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 373, + 219, + 454 + ], + "blocks": [ + { + "bbox": [ + 133, + 373, + 219, + 454 + ], + "lines": [ + { + "bbox": [ + 133, + 373, + 219, + 454 + ], + "spans": [ + { + "bbox": [ + 133, + 373, + 219, + 454 + ], + "type": "image", + "image_path": "e842e3e6f8250fee4f6c64e0d874ee0a105c168976b83f1753ecbd0058f850ce.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 462, + 482, + 495 + ], + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 495 + ], + "type": "text", + "content": "Fig. 3: Comparison of ADD distributions on the real-world datasets between our approach and RoboPose [15]. The y-axis represents the accuracy of our estimation at different ADD thresholds." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 220, + 373, + 307, + 453 + ], + "blocks": [ + { + "bbox": [ + 220, + 373, + 307, + 453 + ], + "lines": [ + { + "bbox": [ + 220, + 373, + 307, + 453 + ], + "spans": [ + { + "bbox": [ + 220, + 373, + 307, + 453 + ], + "type": "image", + "image_path": "5070e1baded4945afba15df5224f21065d183de7284bc65f1eade51e3f032cb9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 373, + 393, + 453 + ], + "blocks": [ + { + "bbox": [ + 307, + 373, + 393, + 453 + ], + "lines": [ + { + "bbox": [ + 307, + 373, + 393, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 373, + 393, + 453 + ], + "type": "image", + "image_path": "ef0cffca512ef254468f10e0b926f4ecb66ac706a3ab91f6a249aa9fddcdc01c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 394, + 373, + 480, + 453 + ], + "blocks": [ + { + "bbox": [ + 394, + 373, + 480, + 453 + ], + "lines": [ + { + "bbox": [ + 394, + 373, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 394, + 373, + 480, + 453 + ], + "type": "image", + "image_path": "8181d5b4fbddedf7342b7d41ae4fdf8af9e8c28d4e4e2f7ba68cc2105f8bce84.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 517, + 483, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 483, + 627 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 483, + 627 + ], + "type": "text", + "content": "Finally, we visualize the predictions of our approach and compare them with SOTA to better understand the performance improvement. Figure 4 illustrates our approach's ability to produce high-quality estimates across different robot morphologies. Particularly, our approach demonstrates superior performance in highly challenging cases, including self-occlusions, truncations, and extreme lighting conditions. Additionally, we compare the model performance qualitatively on in-the-wild lab images and quantitatively under various levels of truncation severity, illustrating the superior generalization ability and robustness of our approach. Please refer to the appendix for more detailed results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 485, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 485, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 485, + 666 + ], + "type": "text", + "content": "Efficiency Furthermore, we benchmark the computation efficiency of our method and compare it with previous works. We measure the average inference time," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "S. Ban et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 204, + 156, + 413, + 206 + ], + "blocks": [ + { + "bbox": [ + 132, + 125, + 480, + 147 + ], + "lines": [ + { + "bbox": [ + 132, + 125, + 480, + 147 + ], + "spans": [ + { + "bbox": [ + 132, + 125, + 480, + 147 + ], + "type": "text", + "content": "Table 2: Comparison of computation efficiency on the Panda 3CAM-AK dataset. † denotes using ground-truth joint state parameters." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 204, + 156, + 413, + 206 + ], + "lines": [ + { + "bbox": [ + 204, + 156, + 413, + 206 + ], + "spans": [ + { + "bbox": [ + 204, + 156, + 413, + 206 + ], + "type": "table", + "html": "
MethodAUC ↑Time (ms) ↓FPS ↑FLOPS (G) ↓
† DREAM-F 1868.966.015.190.5
RoboPose 1570.4570.81.857.3
Ours82.244.322.625.2
", + "image_path": "8637dd38ae82d1780bbd07474339ee3f3025b679262675d32c873284ab750df1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 228, + 482, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 482, + 350 + ], + "type": "text", + "content": "frame per second (FPS), and Floating Point Operations (FLOPs). For a fair comparison, we benchmark all methods on a Linux machine with CPU Intel(R) Xeon(R) Gold 6240 CPU @ 2.60GHz and GPU NVIDIA Tesla V100S. Batch size is set to be 1 for all methods. Table 2 shows that our method exhibits a remarkable advantage in terms of inference speed and computation loads, enabling real-time inference while achieving state-of-the-art accuracy. This advantage stems from the fact that our method involves only a single feed-forward pass without costly iterative optimization during inference. We demonstrate that our approach serves as an efficient and effective solution for real-time holistic robot pose estimation, which holds great potential in real-world applications." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 367, + 354, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 367, + 354, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 367, + 354, + 380 + ], + "type": "text", + "content": "4.4 Comparison under known joints states" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 387, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 472 + ], + "type": "text", + "content": "Although the primary focus of this work is holistic robot pose estimation with unknown states, our approach can also be easily adapted to the known joint states setting. We conduct additional experiments on four real datasets for robot pose estimation with ground-truth joint states available. We simply replace the predicted joint state with the ground-truth joint state during the self-supervised training process. As shown in Tab. 3 our approach also achieves comparable performance with previous SOTA under this well-explored setting." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 489, + 236, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 489, + 236, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 489, + 236, + 502 + ], + "type": "text", + "content": "4.5 Ablation study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 582 + ], + "type": "text", + "content": "Ablation on network modules We first conduct an ablation study on the proposed network modules to investigate their effectiveness. Since JointNet and RotationNet are the major regression targets of our task, we focus on the ablation of two assistive network modules, DepthNet and KeypointNet. We conduct the experiments on the synthetic datasets, utilizing only ground-truth supervision to simplify the process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Firstly, we delve into the implementation of DepthNet. In Tab. 4, we evaluate the design choice of whether to share the backbone of DepthNet with other modules or not. The results indicate that separating the DepthNet backbone consistently improves performance. This improvement could be explained by the fact that DepthNet involves learning scale-variant features, while the other network modules all focus on learning scale-invariant features. We also investigate whether to pre-train the DepthNet or jointly train it with other modules" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 365, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 365, + 91, + 447, + 100 + ], + "type": "text", + "content": "Holistic Robot Pose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 167, + 473, + 327 + ], + "blocks": [ + { + "bbox": [ + 132, + 124, + 482, + 158 + ], + "lines": [ + { + "bbox": [ + 132, + 124, + 482, + 158 + ], + "spans": [ + { + "bbox": [ + 132, + 124, + 482, + 158 + ], + "type": "text", + "content": "Table 3: Comparison of AUC " + }, + { + "bbox": [ + 132, + 124, + 482, + 158 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 132, + 124, + 482, + 158 + ], + "type": "text", + "content": " of the ADD distribution curve and Mean " + }, + { + "bbox": [ + 132, + 124, + 482, + 158 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 132, + 124, + 482, + 158 + ], + "type": "text", + "content": " of the ADD, on the real-world datasets of DREAM, with all methods using known ground-truth joint state parameters." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 167, + 473, + 327 + ], + "lines": [ + { + "bbox": [ + 143, + 167, + 473, + 327 + ], + "spans": [ + { + "bbox": [ + 143, + 167, + 473, + 327 + ], + "type": "table", + "html": "
MethodMetricPanda 3CAM-AKPanda 3CAM-XKPanda 3CAM-RSPanda ORB
DREAM-F18AUC↑68.924.476.161.9
DREAM-Q18AUC↑52.437.578.057.1
DREAM-H18AUC↑60.564.078.869.1
RoboPose15AUC↑76.586.076.980.5
CtRNet22AUC↑89.979.590.885.3
SGTAPose40AUC↑67.82.187.672.3
OursAUC↑90.281.291.987.6
DREAM-F18Mean↓11413.1491911.42077.495319.1
DREAM-Q18Mean↓78089.354178.227.264247.6
DREAM-H18Mean↓56.57381.623.625685.3
RoboPose15Mean↓24.214.023.119.4
CtRNet22Mean↓13.032.010.021.0
SGTAPose40Mean↓35.7157.712.835.0
OursMean↓12.919.911.015.9
", + "image_path": "9f874de395080ec4b75c1d1634dec1741bb915b2f3e63142c721a5f1b3bb5382.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 393, + 298, + 457 + ], + "blocks": [ + { + "bbox": [ + 132, + 350, + 299, + 383 + ], + "lines": [ + { + "bbox": [ + 132, + 350, + 299, + 383 + ], + "spans": [ + { + "bbox": [ + 132, + 350, + 299, + 383 + ], + "type": "text", + "content": "Table 4: Ablation studies on the effect of separating DepthNet image feature extractor." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 393, + 298, + 457 + ], + "lines": [ + { + "bbox": [ + 135, + 393, + 298, + 457 + ], + "spans": [ + { + "bbox": [ + 135, + 393, + 298, + 457 + ], + "type": "table", + "html": "
DatasetSeparateAUC ↑Mean (mm)↓
Panda DR (a)-79.9821.9
Panda DR (b)+82.6818.0
Panda Photo (a)-79.6622.0
Panda Photo (b)+81.9818.4
", + "image_path": "760208ca9cfdffeb13012641e11666047dcfcbc0770942ae5531ae93cbc201e9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 318, + 393, + 481, + 457 + ], + "blocks": [ + { + "bbox": [ + 314, + 350, + 482, + 384 + ], + "lines": [ + { + "bbox": [ + 314, + 350, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 314, + 350, + 482, + 384 + ], + "type": "text", + "content": "Table 5: Ablation studies on the effect of pre-training DepthNet in advance of jointly training." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 393, + 481, + 457 + ], + "lines": [ + { + "bbox": [ + 318, + 393, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 318, + 393, + 481, + 457 + ], + "type": "table", + "html": "
DatasetPretrainAUC ↑Mean (mm)↓
Panda DR (a)-74.2926.8
Panda DR (b)+82.6818.0
Panda Photo (a)-73.3427.9
Panda Photo (b)+81.9818.4
", + "image_path": "494b3ed769b728038574cbafce56a5b0abbd174e18703b52d91708584373d038.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 483, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 482, + 555 + ], + "type": "text", + "content": "from scratch. Table 5 suggests that a pre-training stage leads to considerable performance improvement. This improvement likely occurs because a well-trained DepthNet reliably aids other modules in better training, as it determines the global translation of all joints. Additionally, we select the keypoint closest to the robot center as the root keypoint requiring depth prediction by DepthNet. Please refer to the S2.3 section of the appendix for rationale behind this selection." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Moreover, we conduct ablation experiments to investigate the impact of KeypointNet on other network modules, namely JointNet and RotationNet. We utilize Mean Joint Error, which represents the mean error of the joint state parameter predictions, and the Mean Rotation Angle Error [15], which represents the mean error of the Euler angle predictions " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\theta_{x}^{pred}, \\theta_{y}^{pred}, \\theta_{z}^{pred})" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " for the camera-to-robot rotation, to quantify the accuracy of the JointNet and RotationNet predictions, respectively. It is important to note that a Panda robot consists of seven revolute joints with angle parameters (representing one-DoF rotations) and one prismatic joint with a length parameter (representing one-DoF transla" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "S. Ban et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 116, + 458, + 350 + ], + "blocks": [ + { + "bbox": [ + 151, + 116, + 458, + 350 + ], + "lines": [ + { + "bbox": [ + 151, + 116, + 458, + 350 + ], + "spans": [ + { + "bbox": [ + 151, + 116, + 458, + 350 + ], + "type": "image", + "image_path": "19d02b552fdc9ecf9b92395bac0cc91b86a44add76ceb99ba1f0b9e4674624f6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 357, + 482, + 380 + ], + "lines": [ + { + "bbox": [ + 131, + 357, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 131, + 357, + 482, + 380 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparison between our method and RoboPose [15] on both real and synthetic datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 401, + 482, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 484 + ], + "type": "text", + "content": "tion). Therefore, we calculate and present the Mean Joint Error separately for the revolute and prismatic joints of Panda. We present the detailed ablation results in the S2.1 section of the appendix, in which Tables 6 and 7 show that the presence of KeypointNet significantly improves the prediction accuracy of JointNet and RotationNet. By sharing a feature extractor with JointNet and RotationNet, KeypointNet guides the network towards learning pixel-aligned robot visual features, which further assists in estimating robot pose and joint states." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "type": "text", + "content": "Finally, we examine the overall effect of DepthNet and KeypointNet. To replace DepthNet, we implement a straightforward alternative, directly regressing the camera-to-robot translation " + }, + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "type": "text", + "content": " from image features " + }, + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "type": "text", + "content": " using an MLP, similar to RotationNet. Figure 5a illustrates that both modules play an important role in prediction accuracy. The DepthNet notably enhances accuracy by improving root depth estimation, a global offset that affects the precision of all joints. Meanwhile, KeypointNet further benefits the estimation results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 487, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 487, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 487, + 666 + ], + "type": "text", + "content": "The contribution of these modules to the overall performance validates our design. DepthNet disentangles depth from camera intrinsic, facilitating a learning-based approach to depth estimation. This method exhibits greater expressive capability compared to the analytical, non-parameterized PnP methods employed in previous work [18,22,40]. On the other hand, KeypointNet enhances both JointNet and RotationNet by steering the shared backbone towards learning more pixel-aligned features, thus complementing the learning objectives defined in the robot kinematic space." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 365, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 365, + 91, + 447, + 100 + ], + "type": "text", + "content": "Holistic Robot Pose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": "Ablation on self-supervision We also perform an ablation study of our self-supervised learning objectives. We conduct the experiments on the Panda real datasets. Figure 5b demonstrates that both the keypoint consistency regularization and the mask consistency regularization contribute to the overall performance. Specifically, the mask consistency regularization provides pixel-wise supervision signal with the help of differentiable rendering, while the keypoint consistency regularization unifies the predictions in the pixel space and that in the robot kinematic space. We suppose that pixel-aligned 2D tasks, e.g. keypoint detection and mask segmentation, generally generalize better, and could therefore improve the 3D estimation quality through consistency regularization, further improving the overall performance. As a result, our multi-stage training scheme yields favorable convergence ability. Please refer to the S4 section of the appendix for the error analysis of each module across each training stage." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 136, + 298, + 294, + 401 + ], + "blocks": [ + { + "bbox": [ + 136, + 298, + 294, + 401 + ], + "lines": [ + { + "bbox": [ + 136, + 298, + 294, + 401 + ], + "spans": [ + { + "bbox": [ + 136, + 298, + 294, + 401 + ], + "type": "image", + "image_path": "719af2c204af9ac0f711ce0a150a758f556cbdee2843edde3af6a75bcc74a3c9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 406, + 298, + 440 + ], + "lines": [ + { + "bbox": [ + 132, + 406, + 298, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 406, + 298, + 440 + ], + "type": "text", + "content": "(a) Ablation study on network modules. 'Baseline' represents direct regression, 'DN' and 'KN' refer to DepthNet and KeypointNet, respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 171, + 449, + 441, + 460 + ], + "lines": [ + { + "bbox": [ + 171, + 449, + 441, + 460 + ], + "spans": [ + { + "bbox": [ + 171, + 449, + 441, + 460 + ], + "type": "text", + "content": "Fig. 5: Ablation studies on network modules and self-supervision." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 319, + 298, + 477, + 401 + ], + "blocks": [ + { + "bbox": [ + 319, + 298, + 477, + 401 + ], + "lines": [ + { + "bbox": [ + 319, + 298, + 477, + 401 + ], + "spans": [ + { + "bbox": [ + 319, + 298, + 477, + 401 + ], + "type": "image", + "image_path": "e9262e5a7a8d0f0543a2c87f1cd39756565c3c08b6a00917d201889743b3b883.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 406, + 482, + 440 + ], + "lines": [ + { + "bbox": [ + 315, + 406, + 482, + 440 + ], + "spans": [ + { + "bbox": [ + 315, + 406, + 482, + 440 + ], + "type": "text", + "content": "(b) Ablation study on self-supervision. 'Baseline' represents no self-supervision, 'MC' and 'KC' refer to mask and keypoint consistency regularization, respectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 512, + 220, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 220, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 220, + 525 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "This work presents a novel end-to-end framework for real-time holistic robot pose estimation without prior knowledge of internal robot states, providing substantial assistance in real-world scenarios where the joint state of the estimation target is unavailable or unreliable. By decomposing this complex task into specific sub-tasks and tailoring neural network modules for each, the method effectively addresses the challenges while retaining high computation efficiency. Our approach, which circumvents iterative optimization, significantly accelerates the estimation process and achieves state-of-the-art accuracy. This advancement holds promise for enhancing real-world robotics applications, offering a practical solution for real-time robot pose estimation and joint state monitoring." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "S. Ban et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 137, + 481, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 137, + 481, + 185 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 481, + 185 + ], + "type": "text", + "content": "This work is supported by National Science and Technology Major Project (2022ZD0114904). We thank Hai Ci for the valuable discussions during the early stages of this work. We thank Haoran Lu and Ruihai Wu for their assistance with real-world image captures." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 201, + 197, + 214 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 201, + 197, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 201, + 197, + 214 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 224, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 224, + 481, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 481, + 256 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 481, + 256 + ], + "type": "text", + "content": "1. Bultmann, S., Memmesheimer, R., Behnke, S.: External camera-based mobile robot pose estimation for collaborative perception with smart edge sensors. In: ICRA (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 257, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 257, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 257, + 481, + 289 + ], + "type": "text", + "content": "2. Chen, L., Qin, Y., Zhou, X., Su, H.: EasyHeC: Accurate and Automatic Hand-eye Calibration via Differentiable Rendering and Space Exploration. In: IEEE Robotics and Automation Letters (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 289, + 481, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 481, + 331 + ], + "type": "text", + "content": "3. Christen, S., Yang, W., Pérez-D'Arpino, C., Hilliges, O., Fox, D., Chao, Y.W.: Learning human-to-robot handovers from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9654–9664 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 332, + 481, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 481, + 364 + ], + "type": "text", + "content": "4. Deisenroth, M.P., Rasmussen, C.E., Fox, D.: Learning to control a low-cost manipulator using data-efficient reinforcement learning. Robotics: Science and Systems VII 7, 57-64 (2011)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 365, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 481, + 397 + ], + "type": "text", + "content": "5. Feniello, A., Dang, H.N., Birchfield, S.: Program synthesis by examples for object repositioning tasks. 2014 IEEE/RSJ International Conference on Intelligent Robots and Systems pp. 4428-4435 (2014)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 397, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 397, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 397, + 481, + 418 + ], + "type": "text", + "content": "6. Fiala, M.: Artag, a fiducial marker system using digital techniques. In: CVPR (2005)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 418, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 481, + 450 + ], + "type": "text", + "content": "7. Garrido-Jurado, S., Muñoz-Salinas, R., Madrid-Cuevas, F.J., Marín-Jiménez, M.J.: Automatic generation and detection of highly reliable fiducial markers under occlusion. Pattern Recognition (2014)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 450, + 481, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 450, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 138, + 450, + 481, + 483 + ], + "type": "text", + "content": "8. Hao, R., Özgün, O., Çavusoglu, M.C.: Vision-based surgical tool pose estimation for the da Vinci® robotic surgical system. In: 2018 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 1298-1305. IEEE (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 483, + 481, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 483, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 138, + 483, + 481, + 504 + ], + "type": "text", + "content": "9. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 504, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 504, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 504, + 481, + 525 + ], + "type": "text", + "content": "0. Icli, C., Stepanenko, O., Bonev, I.: New method and portable measurement device for the calibration of industrial robots. Sensors 20(20) (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 525, + 481, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 481, + 558 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 481, + 558 + ], + "type": "text", + "content": "1. Kim, W., Kim, N., Lyons, J.B., Nam, C.S.: Factors affecting trust in high-vulnerability human-robot interaction contexts: A structural equation modelling approach. Applied Ergonomics 85, 103056 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 558, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 558, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 138, + 558, + 481, + 579 + ], + "type": "text", + "content": "2. Kok, B.C., Soh, H.: Trust in robots: Challenges and opportunities. Current Robotics Reports 1, 297-309 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 580, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 580, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 580, + 481, + 611 + ], + "type": "text", + "content": "3. Kothari, N., Gupta, M., Vachhani, L., Arya, H.: Pose estimation for an autonomous vehicle using monocular vision. In: 2017 Indian Control Conference (ICC). pp. 424-431 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 611, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 481, + 633 + ], + "type": "text", + "content": "4. Kuipers, B.: How can we trust a robot? Communications of the ACM 61(3), 86-95 (2018)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "text", + "content": "5. Labbé, Y., Carpentier, J., Aubry, M., Sivic, J.: Single-view robot pose and joint angle estimation via render & compare. In: Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 365, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 365, + 91, + 447, + 100 + ], + "type": "text", + "content": "Holistic Robot Pose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 481, + 161 + ], + "type": "text", + "content": "16. Lambrecht, J., Grosenick, P., Meusel, M.: Optimizing keypoint-based single-shot camera-to-robot pose estimation through shape segmentation. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 13843-13849. IEEE (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 162, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 162, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 481, + 182 + ], + "type": "text", + "content": "17. Lambrecht, J., Kästner, L.: Towards the usage of synthetic data for marker-less pose estimation of articulated robots in rgb images. In: ICAR (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 183, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 481, + 215 + ], + "type": "text", + "content": "18. Lee, T.E., Tremblay, J., To, T., Cheng, J., Mosier, T., Kroemer, O., Fox, D., Birchfield, S.: Camera-to-robot pose estimation from a single image. In: International Conference on Robotics and Automation (ICRA) (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 481, + 237 + ], + "type": "text", + "content": "19. Lepetit, V., Moreno-Noguer, F., Fua, P.: Epnp: An accurate o (n) solution to the pnp problem. IJCV (2009)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 238, + 481, + 270 + ], + "type": "text", + "content": "20. Li, S., De Wagter, C., De Croon, G.C.: Self-supervised monocular multi-robot relative localization with efficient deep neural networks. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 9689-9695. IEEE (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "text", + "content": "21. Li, X., Zhang, X., Zhou, X., Chen, I.M.: Upg: 3d vision-based prediction framework for robotic grasping in multi-object scenes. Knowledge-Based Systems 270, 110491 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "text", + "content": "22. Lu, J., Richter, F., Yip, M.C.: Markerless camera-to-robot pose estimation via self-supervised sim-to-real transfer. In: Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 337, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 337, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 337, + 481, + 369 + ], + "type": "text", + "content": "23. Mahler, J., Matl, M., Satish, V., Danielczuk, M., DeRose, B., McKinley, S., Goldberg, K.: Learning ambidextrous robot grasping policies. Science Robotics 4(26), eaau4984 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "text", + "content": "24. Masehian, E., Jannati, M., Hekmatfar, T.: Cooperative mapping of unknown environments by multiple heterogeneous mobile robots with limited sensing. Robotics and Autonomous Systems 87, 188-218 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "type": "text", + "content": "25. Ming, Y., Meng, X., Fan, C., Yu, H.: Deep learning for monocular depth estimation: A review. Neurocomputing 438, 14-33 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 425, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 481, + 456 + ], + "type": "text", + "content": "26. Moon, G., Chang, J., Lee, K.M.: Camera distance-aware top-down approach for 3d multi-person pose estimation from a single rgb image. In: The IEEE Conference on International Conference on Computer Vision (ICCV) (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "text", + "content": "27. Morrison, D., Corke, P., Leitner, J.: Closing the Loop for Robotic Grasping: A Real-time, Generative Grasp Synthesis Approach. In: Proc. of Robotics: Science and Systems (RSS) (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 491, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 501 + ], + "type": "text", + "content": "28. Olson, E.: Apriltag: A robust and flexible visual fiducial system. In: ICRA (2011)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 502, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 481, + 533 + ], + "type": "text", + "content": "29. Papadimitriou, A., Mansouri, S.S., Nikolakopoulos, G.: Range-aided ego-centric collaborative pose estimation for multiple robots. Expert Systems with Applications 202, 117052 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "type": "text", + "content": "30. Park, D., Seo, Y., Chun, S.Y.: Real-time, highly accurate robotic grasp detection using fully convolutional neural networks with high-resolution images (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 556, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 481, + 578 + ], + "type": "text", + "content": "31. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 578, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 609 + ], + "type": "text", + "content": "32. Pinto, L., Davidson, J., Gupta, A.: Supervision via competition: Robot adversaries for learning tasks. In: 2017 IEEE International Conference on Robotics and Automation (ICRA). pp. 1601-1608 (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "text", + "content": "33. Prusak, A., Melnychuk, O., Roth, H., Schiller, I., Koch, R.: Pose estimation and map building with a time-of-flight-camera for robot navigation. International Journal of Intelligent Systems Technologies and Applications 5(3-4), 355-364 (2008)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "34. Rana, A., Vulpi, F., Galati, R., Milella, A., Petitti, A.: A pose estimation algorithm for agricultural mobile robots using an rgb-d camera. In: 2022 International Con-" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 218, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 218, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 218, + 100 + ], + "type": "text", + "content": "S. Ban et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 567 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 149, + 117, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 117, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 149, + 117, + 481, + 138 + ], + "type": "text", + "content": "ference on Electrical, Computer, Communications and Mechatronics Engineering (ICECCME). pp. 1-5 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 481, + 160 + ], + "type": "text", + "content": "35. Rizk, Y., Awad, M., Tunstel, E.W.: Cooperative heterogeneous multi-robot systems. ACM Computing Surveys (CSUR) 52, 1 - 31 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "type": "text", + "content": "36. Salvini, P., Paez-Granados, D.F., Billard, A.: Safety concerns emerging from robots navigating in crowded pedestrian areas. International Journal of Social Robotics 14, 441 - 462 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "text", + "content": "37. Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learning for human pose estimation. In: CVPR (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "text", + "content": "38. Sun, X., Xiao, B., Wei, F., Liang, S., Wei, Y.: Integral human pose regression. In: Proceedings of the European Conference on Computer Vision (ECCV) (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "text", + "content": "39. Svenstrup, M., Tranberg, S., Andersen, H.J., Bak, T.: Pose estimation and adaptive robot behaviour for human-robot interaction. In: 2009 IEEE International Conference on Robotics and Automation. pp. 3571-3576 (2009)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 270, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 304 + ], + "type": "text", + "content": "40. Tian, Y., Zhang, J., Yin, Z., Dong, H.: Robot structure prior guided temporal attention for camera-to-robot pose estimation from image sequence. In: Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "text", + "content": "41. Tzafestas, S.G.: 15 - mobile robots at work. In: Tzafestas, S.G. (ed.) Introduction to Mobile Robot Control, pp. 635-663. Elsevier, Oxford (2014)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "type": "text", + "content": "42. Uchibe, E., Asada, M., Noda, S., Takahashi, Y., Hosoda, K.: Vision-based reinforcement learning for robocup: Towards real robot competition. In: Proc. of IROS. vol. 96 (1996)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "type": "text", + "content": "43. Xia, Y., Tang, Y., Hu, Y., Hoffman, G.: Privacy-preserving pose estimation for human-robot interaction. CoRR abs/2011.07387 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "44. Xu, C., Yu, X., Wang, Z., Ou, L.: Multi-view human pose estimation in humanrobot interaction. In: IECON 2020 The 46th Annual Conference of the IEEE Industrial Electronics Society. pp. 4769-4775 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "type": "text", + "content": "45. Xun, Z., Huang, J., Li, Z., Xu, C., Gao, F., Cao, Y.: Crepes: Cooperative relative pose estimation towards real-world multi-robot systems. arXiv preprint arXiv:2302.01036 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "text", + "content": "46. Yang, W., Paxton, C., Mousavian, A., Chao, Y.W., Cakmak, M., Fox, D.: Reactive human-to-robot handovers of arbitrary objects. In: IEEE International Conference on Robotics and Automation (ICRA). IEEE (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "47. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5745-5753 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 511, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 511, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 511, + 481, + 544 + ], + "type": "text", + "content": "48. Zhu, W., Ma, X., Ro, D., Ci, H., Zhang, J., Shi, J., Gao, F., Tian, Q., Wang, Y.: Human motion generation: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 544, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 481, + 567 + ], + "type": "text", + "content": "49. Zuo, Y., Qiu, W., Xie, L., Zhong, F., Wang, Y., Yuille, A.L.: Craves: Controlling robotic arm with a vision-based economic system. In: CVPR (2019)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 365, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 365, + 91, + 447, + 100 + ], + "type": "text", + "content": "Holistic Robot Pose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/RealGen_ Retrieval Augmented Generation for Controllable Traffic Scenarios/4b2a9c03-0a68-49b3-9375-225cd793a8b3_content_list.json b/2024/RealGen_ Retrieval Augmented Generation for Controllable Traffic Scenarios/4b2a9c03-0a68-49b3-9375-225cd793a8b3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..aaa1429eca95408a5a066f91ba94a373e6cb344f --- /dev/null +++ b/2024/RealGen_ Retrieval Augmented Generation for Controllable Traffic Scenarios/4b2a9c03-0a68-49b3-9375-225cd793a8b3_content_list.json @@ -0,0 +1,1830 @@ +[ + { + "type": "text", + "text": "RealGen: Retrieval Augmented Generation for Controllable Traffic Scenarios", + "text_level": 1, + "bbox": [ + 235, + 140, + 767, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenhao Ding\\*1,2, Yulong Cao\\*2, Ding Zhao\\*1, Chaowei Xiao\\*2,3, and Marco Pavone\\*2,4", + "bbox": [ + 230, + 210, + 771, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Carnegie Mellon University", + "bbox": [ + 398, + 253, + 602, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 NVIDIA Research", + "bbox": [ + 431, + 268, + 570, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 UW-Madison", + "bbox": [ + 449, + 282, + 553, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Stanford University", + "bbox": [ + 428, + 296, + 575, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wenhaod, yulongc}@nvidia.com", + "bbox": [ + 388, + 311, + 614, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Simulation plays a crucial role in the development of autonomous vehicles (AVs) due to the potential risks associated with real-world testing. Although significant progress has been made in the visual aspects of simulators, generating complex behavior among agents remains a formidable challenge. It is not only imperative to ensure realism in the scenarios generated but also essential to incorporate preferences and conditions to facilitate controllable generation for AV training and evaluation. Traditional methods, which rely mainly on memorizing the distribution of training datasets, often fail to generate unseen scenarios. Inspired by the success of retrieval augmented generation in large language models, we present RealGen, a novel retrieval-based in-context learning framework for traffic scenario generation. RealGen synthesizes new scenarios by combining behaviors from multiple retrieved examples in a gradient-free way, which may originate from templates or tagged scenarios. This in-context learning framework endows versatile generative capabilities, including the ability to edit scenarios, compose various behaviors, and produce critical scenarios. Evaluations show that RealGen offers considerable flexibility and controllability, marking a new direction in the field of controllable traffic scenario generation. Check our project website for more information: https://realgen.github.io.", + "bbox": [ + 259, + 349, + 743, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 646, + 375, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Simulation is indispensable in the development of autonomous vehicles (AVs), primarily due to the considerable risks associated with training and evaluating these systems in real-world conditions. The biggest challenge in simulations lies in achieving realistic driving scenarios, as this realism influences the discrepancy between AV performance in simulated and actual environments. Although advancements in high-quality graphical engines have significantly enhanced the perception quality of simulators, the realism of agent behavior remains constrained because of the complicated interactions among naturalistic agents. To counteract this issue, data-driven simulation has emerged as a promising approach in the realm of autonomous driving, which leverages real-world scenario datasets to accurately generate the behaviors of agents.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rapid achievement of deep generative models [29] and imitation learning algorithms [32], current data-driven simulations [23, 38] can generate scenarios that closely mimic human driver behavior. However, the effectiveness of these simulations in accelerating the development of AV is limited. This limitation stems from the need for scenarios that meet specific conditions tailored for targeted training and evaluation. Achieving such controllability in simulations is challenging due to the complex nature of driving scenarios, which involve intricate interactions, diverse road layouts, and varying traffic regulations.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In pursuit of controllability, existing work applies additional guidance, typically in the form of constraint functions [16, 64] or languages [49, 63], to pre-trained scenario generative models. Regularization of the generation process through these tools is straightforward and effective, yet it encounters two principal challenges. First, the training of scenario generative models typically utilizes naturalistic datasets, which might not encompass the specific scenarios desired as per the control signals. Even if such", + "bbox": [ + 212, + 281, + 431, + 553 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c2692d7e4ef746f27023fb6009f1edbf303e3631ea23102ce352a5bae027f12e.jpg", + "image_caption": [ + "(a) Conventional method" + ], + "image_footnote": [], + "bbox": [ + 442, + 284, + 560, + 430 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/17d61e5d3fabb3e7ce926a63f363956d5811a6f3db4fdf611e42045720b3588f.jpg", + "image_caption": [ + "(b) Our method", + "Fig. 1: (a) Conventional methods make the model memorize the data distribution for generating. (b) In contrast, our method employs a retriever to query datasets (including external data obtained after training) and uses a generative model to generate scenarios by integrating the information from the retrieved scenarios." + ], + "image_footnote": [], + "bbox": [ + 571, + 285, + 784, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "scenarios exist within the dataset, they are often omitted because of the rarity of long-tail data. The second challenge is that the representation of the guidance to the generative model may not be sufficiently expressive to accurately depict complex scenarios, such as specifying intricate interactions among multiple vehicles, using language. These limitations underscore the need for a more sophisticated and nuanced framework for controllable scenario generation.", + "bbox": [ + 212, + 553, + 787, + 645 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Retrieval Augmented Generation (RAG) [8, 25], which enhances the generative process by querying related information from external databases, represents great potential in the domain of large language models [37]. In contrast to conventional models that memorize all knowledge within their parameters, as shown in Figure 1(a), RAG models, shown in Figure 1(b), learn to generate comprehensive outputs by retrieving pertinent knowledge from a database, based on the input provided. A notable aspect of RAG is the ability of the database to undergo updates even after the model has been trained, allowing continuous improvement and adaptation. This flexible framework offers possibilities for controllable scenario generation by using appropriate template scenarios as input and facilitating the generation that is not only realistic but also aligned with specific training and evaluation requirements.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "W. Ding et al.", + "bbox": [ + 271, + 114, + 369, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we present RealGen, a retrieval augmented generation framework for generating traffic scenarios. This framework, as shown in Figure 2, begins with the training of an encoder through contrastive self-supervised learning [42] to allow the retrieval process to query similar scenarios in a latent embedding space. Leveraging this latent representation, we subsequently train a generative model that combines retrieved scenarios to create novel scenarios. The key contributions of this paper are summarized below.", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We develop a novel contrastive autoencoder model to extract scenario embeddings as latent representations, which can be used for a wide range of downstream tasks.", + "- We propose the first retrieval augmented generation framework using the latent representation tailored for controllable driving scenario generation.", + "- We validate our framework through qualitative and quantitative metrics, demonstrating strong flexibility and controllability of generated scenarios." + ], + "bbox": [ + 225, + 261, + 785, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 390, + 387, + 406 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Traffic Modeling and Scenario Generation", + "text_level": 1, + "bbox": [ + 215, + 422, + 606, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Research on traffic modeling using generative models has received considerable attention in recent works. Notably, ScenarioNet [38] and Waymax [23] employ large-scale data to train imitation learning models, facilitating the generation of realistic multi-agent scenarios in simulations. SceneGen [50] utilizes the Long Short Term Memory [30] module to autoregressively generate the trajectories for vehicles and pedestrians based on provided maps. Additionally, TrafficSim [47] learns the multi-agent behaviors from real-world data, and TrafficGen [18] proposes a transformer-based autoencoder architecture to accurately model complex interactions involving multiple agents. MixSim [48] builds a reactive digital twin and finds safety-critical scenarios with black-box optimization. RTR [61] models reactive behaviors of vehicles using the combination of reinforcement learning and imitation learning.", + "bbox": [ + 212, + 446, + 787, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Numerous prior studies have employed adversarial generation to synthesize rare yet critical scenarios. L2C [17] utilizes a reinforcement learning framework, where the reward for scenario generation is based on the collision rate. To enhance the realism in purely adversarial generation methods, MMG [15] incorporates the data distribution as regularization. Further developments such as AdvSim [54], AdvDO [7], and KING [26] integrate vehicle dynamics to directly optimize the trajectory to find critical scenarios.", + "bbox": [ + 212, + 628, + 787, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proliferation of large language models (LLMs) has led to recent approaches in generating traffic scenarios with language as conditions to follow instructions from humans. CTG++ [63] replaces the gradient guidance process of CTG [64] with cost functions generated by an LLM. Leveraging the power of GPT-4 [59], this framework can generate diverse motion data with language conditions. Additionally, LCTGen [49] also harnesses the strengths of GPT-4 to generate a heuristic intermediate representation of scenarios from language", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "RealGen: Retrieval Augmented Generation for Controllable Traffic Scenarios", + "bbox": [ + 220, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "inputs. Subsequently, a generative model pre-trained on open-source datasets generates various scenarios with this representation as inputs.", + "bbox": [ + 212, + 146, + 784, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Retrieval-augmented Generation", + "text_level": 1, + "bbox": [ + 214, + 196, + 532, + 213 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Information retrieval (IR) [33] is the procedure of representing and searching a collection of data with the goal of extracting knowledge to satisfy the query of users. Predominantly utilized in search engines and digital libraries [43], IR primarily deals with the information stored in textual formats. Recently, IR has extended across diverse sectors, notably enhancing the quality of outputs in areas such as language modeling [3, 39], question answering [25, 62], image creation [2, 11], and the generation of molecular [55].", + "bbox": [ + 212, + 220, + 787, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One intuitive usage of the retrieved information is to enhance input data through several methods, such as merging the original data and retrieved data [37], employing attention mechanism [3], or extracting skeleton [6]. The rationale behind retrieval systems stems from the impracticality of encoding all knowledge within model parameters, especially considering the dynamic nature of knowledge that evolves with human activities. Consequently, the capability to access external knowledge databases can significantly improve the precision and quality of generated responses, as evidenced in applications involving LLMs [3, 37].", + "bbox": [ + 212, + 327, + 789, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Another usage of retrieved data, particularly pertinent to this study, involves the controllable generation by integrating desired features retrieved from the dataset. In [35], the authors generate product reviews with controllable information about the user, product, and rating. Similarly, the process of story creation can be viewed as a blend of external story databases with selected text fragments [57] Furthermore, [55] explores the controllable generation of molecules to meet various constraints, utilizing a database comprising simple molecules that typically meet only one of these constraints.", + "bbox": [ + 212, + 448, + 787, + 569 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Self-supervised Feature Learning", + "text_level": 1, + "bbox": [ + 214, + 590, + 532, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The necessity for manual labeling in supervised learning often results in human biases, extraneous noises, and labor-intensive efforts. Therefore, the paradigm of self-supervised learning (SSL) [46] is garnering heightened interest, particularly for its promising applications in language modeling and image interpretation. SSL algorithms usually learn implicit representations from extensive pools of unlabeled data without relying on human annotations. Generally, this line of research can be bifurcated into two categories [60]: generative SSL and discriminative SSL.", + "bbox": [ + 212, + 613, + 787, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the domain of generative SSL, models employ an autoencoder to convert input data into a latent representation, followed by a reconstruction process. As an illustrative example, Denoising Autoencoder [53], an early example of this approach, reconstructs images from their noisy version to extract representative features. Another technique in this field is masked modeling, extensively utilized in language models such as GPT-3 [4] and BERT [14], which predicts the masked token to gain a semantic understanding of the text. More recently, the", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "W. Ding et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Masked Autoencoder [27] has emerged as a powerful framework, demonstrating its efficacy across a variety of downstream tasks.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discriminative SSL typically optimizes a discriminative loss to learn representations from sets of anchor, positive, and negative samples. Without ground-truth labels, these pairs are often constructed through solving jigsaw puzzles [41] or making geometry-based prediction [20]. A prominent example of discriminative SSL is contrastive learning, which brings samples from the same class closer while distancing those from different classes. Representative works such as MoCo [28] and SimCLR [10] learn embeddings by capturing invariant features between original data and its augmented variants. Additionally, InfoNCE [42], a method grounded in noise contrastive estimation [24], is widely used due to its effectiveness.", + "bbox": [ + 212, + 176, + 787, + 328 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 A Latent Representation of Scenario", + "text_level": 1, + "bbox": [ + 215, + 354, + 606, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A crucial element within the retrieval system is the selection metric for data retrieval, which is typically implemented through a distance function between the query sample and the candidate samples in the database. Unlike text, which can be converted into word embeddings, traffic scenarios encompass sequential behaviors and intricate interactions among entities, complicating the establishment of a similarity metric for these scenarios. Consequently, in this section, we introduce a scenario autoencoder to extract latent representations that facilitate the assessment of similarity between various traffic scenarios, a vital component in our RAG framework.", + "bbox": [ + 212, + 388, + 787, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Scenario Definition", + "text_level": 1, + "bbox": [ + 215, + 547, + 419, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Each scenario is characterized by the trajectories $\\tau \\in \\mathbb{R}^{M\\times T\\times 5}$ , encompassing $M$ agents over a maximum of $T$ time steps. The trajectory of each agent is composed of parameters $[x,y,v,c,s]$ , signifying position $x$ , position $y$ , velocity $v$ , cosine of the heading $c$ , and sine of the heading $s$ . The initial state of these entities is denoted as $\\tau_0\\in \\mathbb{R}^{M\\times 5}$ . Furthermore, the map is encapsulated by $m\\in \\mathbb{R}^{S\\times 4}$ , composed of $S$ lane segments. The attributes of these points $[x_s,y_s,x_e,y_e]$ correspond to the starting and ending positions of each segment.", + "bbox": [ + 212, + 575, + 787, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Scenario Autoencoder", + "text_level": 1, + "bbox": [ + 215, + 707, + 444, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Autoencoders [44] are used to learn compressed latent representations of high-dimensional data, where an encoder projects the data into latent code and a decoder reconstructs the code in the data space. Given that traffic scenarios encompass spatial and temporal dynamics from multiple agents, we design a hierarchical encoder structure alternating between spatial and temporal layers based on transformer architecture [51]. As depicted on the left of Figure 2, our design incorporates three encoders for behavior, map, and initial position.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "RealGen: Retrieval Augmented Generation for Controllable Traffic Scenarios", + "bbox": [ + 220, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f11185e87a54575c4943521e59e96c8a8d4a3217b9d0d8eedc46c35bc908207b.jpg", + "image_caption": [ + "Fig. 2: Left: the training pipeline for the encoders and decoder aimed at learning latent embeddings of scenarios. A contrastive loss is applied to behavior embeddings to ensure invariance to absolute positions. Middle: the training pipeline for the combiner with frozen encoder and decoder parameters. We use K-Nearest Neighbors (KNN) to retrieve scenarios similar to a template scenario in the dataset and use the retrieved behaviors to reconstruct the template scenario. Right: the generation pipeline with a retriever and a generator." + ], + "image_footnote": [], + "bbox": [ + 217, + 142, + 787, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "First, we design a behavior encoder $E_{b}$ with a spatial-temporal transformer structure that comprises $L_{e}$ temporal transformer encoders $(\\mathrm{Encoder}_{t}^{i})$ and $L_{e}$ spatial transformer encoders $(\\mathrm{Encoder}_{s}^{i})$ , where $i \\in \\{1, \\dots, L_{e}\\}$ . The trajectory $\\tau$ is first transformed into a latent embedding $z_{b}$ , where $H$ denotes the hidden dimension, through a Multiple Linear Perception (MLP) module. Subsequently, an alternating encoding procedure extracts spatial and temporal features from $z_{b}$ . To preserve temporal information, we add a sinusoidal positional embedding (PE) to the embedding before employing the temporal transformer. To retain distinct agent information and further compress the behavior feature, we calculate the mean of the latent embedding across the temporal dimension, resulting in the final behavior embedding $z_{b} \\in \\mathbb{R}^{M \\times H}$ . Besides the behavior encoder, we also process the map information represented by lane vectors with the attention mechanism [51]. Similarly to the first step in $E_{b}$ , we project $m$ to the latent space with an MLP module. Then we apply a multi-head attention module (MHA) with layer normalization (LN) [1] to acquire the map embedding $z_{m} \\in \\mathbb{R}^{S \\times H}$ with a learnable query embedding $q_{m}$ .", + "bbox": [ + 212, + 439, + 787, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the decoder, a spatial-temporal transformer architecture, similar to the behavior encoder, is designed to obtain the behavior embedding $z_{b}$ through $L_{d}$ times encoding. Given that $z_{b}$ lacks the temporal dimension, we first replicate it $T$ times and then add a PE before entering it into the temporal transformer encoder. Throughout the decoding process, the map embedding $z_{m}$ is injected by a cross-attention mechanism: $z_{b} \\gets z_{b} + \\mathrm{MHA}(z_{b}, z_{m}, z_{m})$ . MHA( $Q, K, V$ ) is the multi-head attention with $Q, K, V$ representing query, key, and value:", + "bbox": [ + 212, + 683, + 789, + 789 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {M H A} (Q, K, V) = \\text {C o n c a t e n a t e} \\left(h _ {1}, \\dots , h _ {i}\\right) W ^ {O},\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 803, + 785, + 830 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nh _ {i} = \\mathrm {A t t e n t i o n} (Q W _ {i} ^ {Q}, K W _ {i} ^ {K}, V W _ {i} ^ {V}),\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 825, + 607, + 843 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "W. Ding et al.", + "bbox": [ + 271, + 114, + 369, + 128 + ], + "page_idx": 5 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Details of Encoder and Decoder" + ], + "code_body": "1 Behavior Encoder $E_{b}$ $(\\tau)$ : 14 Initial Pose Encoder $E_{i}$ $(\\tau_{0})$ : \n2 $z_{b} \\gets \\mathrm{MLP}(\\tau) // \\text{projection}$ 15 $z_{i} \\gets \\mathrm{MLP}(\\tau_{0}) // \\text{projection}$ \n3 for $i$ in $[1, \\dots, L_{e}]$ do 16 return initial pose embedding $z_{i}$ \n4 $z_{b} \\gets \\mathrm{Encoder}_{t}^{i}(\\mathrm{PE} + z_{b})$ 17 Decoder $D(z_{i}, z_{b}, z_{m})$ : \n5 $z_{b} \\gets \\mathrm{Encoder}_{s}^{i}(z_{b})$ 18 $z_{r} \\gets z_{b} + \\mathrm{MHA}(z_{b}, z_{i}, z_{i})$ \n6 $z_{b} \\gets \\text{mean}(z_{b})$ 19 for $i$ in $[1, \\dots, L_{d}]$ do \n7 return behavior embedding $z_{b}$ 20 $z_{r} \\gets \\mathrm{Encoder}_{t}^{i}(\\mathrm{PE} + z_{r})$ \n8 Map Encoder $E_{m}$ : 21 \n9 Initialize a learnable query $q_{m}$ 22 \n10 $z_{m} \\gets \\mathrm{MLP}(m) // \\text{projection}$ 23 $\\hat{\\tau} \\gets \\mathrm{MLP}(z_{r}) // \\text{projection}$ \n11 $z_{m} \\gets \\mathrm{LN}(\\mathrm{MHA}(q_{m}, z_{m}, z_{m}))$ 24 return reconstructed trajectory $\\hat{\\tau}$ \n12 $z_{m} \\gets \\mathrm{LN}(z_{m} + \\mathrm{MLP}(z_{m}))$ \n13 return map embedding $z_{m}$", + "bbox": [ + 218, + 166, + 787, + 361 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $W^{O}$ , $W_{i}^{Q}$ , $W_{i}^{K}$ , and $W_{i}^{V}$ are learnable parameters. The last component of the decoder is an MLP that projects the hidden embedding back into the data domain, resulting in the reconstructed trajectory $\\hat{\\tau}$ . To train the encoders and decoder, the mean square error is used as the reconstruction loss, expressed as $\\mathcal{L}_r = |\\hat{\\tau} -\\tau |_2$ . Comprehensive descriptions of the encoding and decoding processes are provided in Algorithm 1.", + "bbox": [ + 212, + 393, + 787, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It is important to note that, while the current design of the autoencoder is sufficient for acquiring compressed representations of scenarios, these representations are not invariant to the absolute coordinates and the order of the agents. This limitation can lead to substantial embedding distances between similar scenarios. To address these issues, two enhancements are proposed and detailed in the following section.", + "bbox": [ + 212, + 486, + 787, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Invariant feature with contrastive loss", + "text_level": 1, + "bbox": [ + 215, + 601, + 576, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To enhance the representation of scenario similarity in the behavior embedding $z_{b}$ , we employ contrastive learning to acquire invariant features. Specifically, we integrate InfoNCE [42] as an additional loss function $\\mathcal{L}_c$ , which optimizes categorical cross-entropy to distinguish a positive sample from a batch of negative samples. In practice, for a query embedding $z_{b}$ , a positive sample $z_{b}^{+}$ is generated by applying random rotation and translation to the original scenario $\\tau, m$ . Meanwhile, negative samples $Z_{b}^{-}$ are selected from the remaining samples in the same batch. In conventional InfoNCE, to bring $z_{b}$ and $z_{b}^{+}$ closer together, the inner product is calculated between the embedding $z_{b}$ and the set $z_{b}^{+}, Z_{b}^{-}$ , employing cross-entropy loss to identify the index of the positive sample.", + "bbox": [ + 212, + 627, + 787, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To ensure that the ordering of the agents does not affect the outcomes, $z_{b} \\in \\mathbb{R}^{M \\times H}$ should exhibit permutation invariance across the dimension $M$ . Otherwise, simply stacking them into a one-dimensional vector could erroneously represent distinct scenarios as significantly different. To establish permutation", + "bbox": [ + 212, + 780, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "RealGen: Retrieval Augmented Generation for Controllable Traffic Scenarios", + "bbox": [ + 220, + 113, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "invariance, we adopt the Wasserstein distance $W_{2}$ [52] as a more suitable metric than cosine distance for assessing the similarity. Taking $z_{b}$ as a distribution representing $M$ individual behaviors, the intuition behind this choice is that the Wasserstein distance identifies the minimal adjustments necessary for behaviors in one scenario to reflect those in another. With such a powerful tool, we have the following contrastive loss:", + "bbox": [ + 212, + 146, + 787, + 237 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} = - \\sum_ {z _ {b}} \\log \\frac {\\exp \\left[ - W _ {2} \\left(z _ {b} , z _ {b} ^ {+}\\right) \\right]}{\\sum_ {z ^ {\\prime} \\in \\left\\{z _ {b} ^ {+} , Z _ {b} ^ {-} \\right\\}} \\exp \\left[ - W _ {2} \\left(z _ {b} , z ^ {\\prime}\\right) \\right]}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 243, + 785, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The implementation of this loss results in the behavior embedding $z_{b}$ lacking absolute coordinate data, making the decoder incapable of accurately reconstructing the precise trajectory. To handle this problem, we add the initial poses $\\tau_0$ of all agents as a supplementary input for the decoder. We encode these poses into $z_{i}$ with an MLP encoder $\\mathrm{Enc}_i$ , and then integrate it into the decoder with an MHA module $z_{r} \\gets z_{b} + \\mathrm{MHA}(z_{b},z_{i},z_{i})$ . Additional specifics regarding the initial pose encoder are detailed in Algorithm 1. In the training stage, the objective is to minimize a combined loss function $\\mathcal{L} = \\mathcal{L}_r + \\lambda \\mathcal{L}_c$ , where $\\lambda$ serves as a weighting factor for contrastive loss and is uniformly set at 0.1 for all experiments.", + "bbox": [ + 214, + 289, + 789, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Retrieval Augmented Scenario Generation", + "text_level": 1, + "bbox": [ + 214, + 446, + 663, + 464 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The autoencoder structure introduced in Section 3 builds a one-to-one generation framework, which still needs additional modules to support the retrieval augmented generation, which is a many-to-one framework. In this section, we introduce a module called Combiner, which takes input behavior embeddings from multiple scenarios and outputs the combined embedding. To train this module, we designed a KNN-based training pipeline that forces the model to learn to combine and edit existing scenarios.", + "bbox": [ + 212, + 474, + 787, + 582 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 The Training of Combiner", + "text_level": 1, + "bbox": [ + 214, + 599, + 478, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unlike the trajectory prediction task [45] which has ground truth as the optimization target, the objective in training the combiner is to learn the alignment of behaviors in the retrieved scenarios with the initial pose and the specified map. A well-trained combiner should be able to compose behaviors from these retrieved scenarios, thereby generating new scenarios that have a resemblance to all the retrieved scenarios. This type of objective aligns with the concept of meta-learning, or 'learning to learn', as described in [31].", + "bbox": [ + 212, + 621, + 787, + 727 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Within our training framework, for a given query scenario $\\tau, m$ in the dataset, we initially utilize the behavior encoder to obtain $z_{b}$ and then use KNN to identify $K$ similar behavior embeddings from the database, denoted $z_{ret} = [z_{b,1}, \\dots, z_{b,K}]$ . Building on this, we propose a model comprising two MHA modules as the combiner:", + "bbox": [ + 214, + 727, + 787, + 801 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nz _ {r a g} \\leftarrow z _ {i} + \\mathrm {M H A} \\left(z _ {i}, z _ {r e t}, z _ {r e t}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 808, + 785, + 833 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nz _ {r a g} \\leftarrow z _ {r a g} + \\mathrm {M H A} (z _ {r a g}, z _ {m}, z _ {m}),\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 828, + 624, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "W. Ding et al.", + "bbox": [ + 271, + 114, + 369, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $z_{i} \\gets E_{i}(x_{0})$ and $z_{m} \\gets E_{m}(m)$ are the initial pose embedding and the map embedding of the query scenario. Gradients are stopped for $z_{i}$ and $z_{m}$ . Assuming that the $K$ nearest scenarios adequately represent the query scenario, it should be feasible to reconstruct the behavior of the query scenario, $z_{b}$ , using the retrieved scenario, $z_{rag}$ . Consequently, we employ the following loss function as our training goal:", + "bbox": [ + 212, + 146, + 782, + 238 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r a g} = \\left\\| D \\left(z _ {i}, z _ {r a g}, z _ {m}\\right) - \\tau \\right\\| _ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 246, + 785, + 263 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "wherein the parameters of the decoder $D\\phi$ remain fixed during training. Essentially, this training method is designed to learn an \"inverse\" operation of the KNN, aiming to reconstruct a query scenario that closely resembles all $K$ retrieved scenarios in the database.", + "bbox": [ + 212, + 268, + 787, + 330 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 The generation pipeline", + "text_level": 1, + "bbox": [ + 215, + 349, + 459, + 366 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The pipeline for scenario generation, as delineated in Figure 2, is categorized into retriever and generator components. The generation process is expedited by pre-processing all scenarios in the database using a behavior encoder, which yields behavior embeddings that facilitate efficient similarity computation.", + "bbox": [ + 212, + 371, + 782, + 431 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The retriever enhances the versatility of generation by dividing the process into two stages. In the initial stage, users can choose a set of template scenarios that depict specific conditions. These may include manually annotated scenarios with tags denoting actions like left or right turns, enabling the generation of additional scenarios under similar tags or even a combination thereof. In addition, templates can include critical and interesting scenarios collected from real-world data. Solely relying on these templates for generation could be limited, which is mitigated by incorporating a secondary phase, which employs a KNN approach to fetch high-quality scenarios from a vast and unlabeled database to augment adaptability.", + "bbox": [ + 212, + 431, + 784, + 582 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Subsequently, the generator, comprising a combiner and a decoder, follows the same inference as combiner training with the initial pose and lane map specified by the user. We obtain the RAG embedding $z_{rag}$ with Eq. (3), and then infer the generated scenario through $\\tau_{rag} \\gets D(z_i, z_{rag}, z_m)$ .", + "bbox": [ + 212, + 583, + 785, + 645 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiment", + "text_level": 1, + "bbox": [ + 215, + 662, + 366, + 681 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we start with an overview of the experimental setup and details regarding the implementation of the RealGEN framework. Then we discuss the baselines, including different types of autoencoders and variants of RealGen, used for comparison. In the discussion of the findings, we respectively evaluate the quality of scenario embedding and the generation capability of RAG.", + "bbox": [ + 212, + 691, + 782, + 768 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Settings and Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 787, + 566, + 803 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We conducted all training and evaluation with the nuScenes [5] dataset using the trajdata [34] package for data loading and processing. Each scenario spans", + "bbox": [ + 212, + 809, + 784, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "RealGen: Retrieval Augmented Generation for Controllable Traffic Scenarios", + "bbox": [ + 220, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "a duration of 8 seconds, operating at a frequency of $2\\mathrm{Hz}$ , and encompasses a maximum of 11 agents. We filter out agents that travel less than 3 meters in 8 seconds and select 11 agents closest to the ego vehicle. The map contains 100 lanes (each lane has 20 points) that are ordered according to the distance between the center of the lane and the center of the ego vehicle.", + "bbox": [ + 212, + 146, + 787, + 222 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To train RealGen, we use Adam [36] as the optimizer to update the parameters of the autoencoder and the combiner. To make the calculation of the Wasserstein distance efficient, we use the Sinkhorn distance [13], an entropy regularized approximation of the Wasserstein distance [52], with implementation in the GeomLoss [19] package.", + "bbox": [ + 212, + 222, + 787, + 297 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2Baselines", + "text_level": 1, + "bbox": [ + 214, + 318, + 339, + 332 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In the experimental section, we evaluate the following reconstruction-based generative models as baselines for comparison. Autoencoder (AE) shares the same behavior encoder, map encoder, and decoder structures as RealGen, serving as the most straightforward baseline for scenario reconstruction. Contrastive AE mirrors the structure of the RealGen autoencoder but omits the initial pose as absolute information. Masked AE is a self-supervised learning baseline, which has been investigated for trajectory data as described in [9, 12, 56, 58]. To evaluate the controllable generation capability, we select a state-of-the-art modelLCTGen [49] as the baseline, which takes a high-level agent representation $z$ and a vector map $m$ as input. We also consider LCTGen w/o $z$ , a variant model proposed in the original paper to show the reconstruction results without using the agent information. Given that our method has two phases, we refer to the autoencoder component specifically as RealGen-AE and the complete model as RealGen.", + "bbox": [ + 212, + 342, + 789, + 554 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.3 Evaluation of Behavior Embedding", + "text_level": 1, + "bbox": [ + 214, + 574, + 552, + 589 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We first evaluate the quality of the learned representation of behavior, which is critical for the following retrieval and generation processes.", + "bbox": [ + 212, + 598, + 785, + 628 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Visualizing similar and dissimilar scenarios. After training the autoencoder with contrastive loss, the distance between behavior embeddings can be used as an indicator of the similarity of the two scenarios. To validate this statement, we visualize qualitative examples of using a query to find the most similar (minimal $W_{2}$ distance) and dissimilar (maximal $W_{2}$ distance) scenarios in Figure 3. We observe that the most similar scenario contains the same behavior and number of vehicles as in the query scenario.", + "bbox": [ + 212, + 628, + 787, + 734 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Classifying Scene ID with behavior embedding. To further provide quantitative results on how well the behavior embedding encodes behavior information, we leverage the Scene ID of the nuScenes dataset. Each scene typically lasts 20 seconds and our scenario segments last 8 seconds, so we get multiple segments belonging to the same Scene ID. We assume that the segments having the same Scene ID have similar behaviors so that we can calculate the accuracy by using the distance to find the closest segment. In Figure 4, we summarize", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "W. Ding et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/8d94447a233949b044fad76d42fd238d99fd05b3f83b1513c868e46241024162.jpg", + "image_caption": [ + "Fig.3: Qualitative evaluation of similar and dissimilar scenarios calculated by our scenario embedding. Rectangles represent the initial poses of vehicles and the lines represent the future trajectories." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 787, + 373 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "the results in the left part, where top- $k$ means we calculate the accuracy for the closest $k$ segments. To validate the effectiveness of $W_{2}$ distance, we also permute the order of agents in the behavior embedding, which is denoted as cosine-permuted and $W_{2}$ -permuted. We find that the cosine distance cannot deal with the permuted setting, as the order is important when stacking the embedding of agents. Our method performs well in top-1 and top-5 settings but badly in others, which could be explained by the fact that segments in one scene have very different behaviors. To validate this, we plot the distance matrix between segments for 11 scenes in the right part of Figure 4. We find sub-blocks in the diagonal blocks, indicating that segments are similar in a small time interval but could be different when the time interval is large.", + "bbox": [ + 212, + 440, + 787, + 606 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Linear probing of behavior embedding. Linear probing, a commonly employed method to assess representations in self-supervised learn", + "bbox": [ + 212, + 607, + 486, + 666 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/94d128ad46f2eeaeed89c0ab1c83cb054705f13facbc77bc567ed9f002bf504a.jpg", + "table_caption": [ + "Table 1: Accuracy of linear probing" + ], + "table_footnote": [], + "table_body": "
AEContrastive AEMasked AERealGen-AE
82.5%67.1%86.2%87.8%
", + "bbox": [ + 501, + 623, + 784, + 660 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "ing (SSL), involves training a linear classifier using the derived embeddings [27]. To train this classifier, we implemented heuristic rules to assign basic behavioral labels (acceleration, deceleration, stopping, keeping speed, left/right turn) to each agent's embedding. The outcomes, along with comparisons with baseline models, are presented in Table 1. These findings demonstrate that the embeddings generated by RealGen surpass all baseline models in terms of accuracy.", + "bbox": [ + 212, + 667, + 787, + 758 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.4 Evaluation of Retrieval Augmented Generation", + "text_level": 1, + "bbox": [ + 212, + 782, + 651, + 797 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluating scene-level controllability and quality of scenario generation is an open problem due to the lack of quantitative metrics. Following most previous", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "RealGen: Retrieval Augmented Generation for Controllable Traffic Scenarios", + "bbox": [ + 220, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0ba044be2984e857b489034882666ad9d2bee681a44e58d26f7fdabdbd50ed4c.jpg", + "image_caption": [ + "Fig. 4: (a) Scene ID accuracy using the behavior embedding with difference distance metrics. (b) A matrix shows the Wasserstein distance between scenario segments, where each block contains the segments that belong to the same Scene ID." + ], + "image_footnote": [], + "bbox": [ + 287, + 143, + 709, + 349 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/94dd70d1ecc510fdb390c5a64f22d1ae2628492b1557dd526c1b9523d2d1a0df.jpg", + "table_caption": [ + "Table 2: Results of realism metrics. Recon.-based and Retrieval-based means using the target scenario and retrieved scenarios as input for generation, respectively." + ], + "table_footnote": [], + "table_body": "
CategoryMethodmADEmFDESpeedHeadingSCRORR
Recon.-basedAE0.18±0.030.41±0.060.04±0.010.10±0.010.02±0.000.02±0.00
Masked AE0.16±0.010.39±0.010.04±0.010.09±0.010.03±0.000.02±0.00
Contrastive AE0.92±0.021.47±0.040.12±0.000.36±0.020.04±0.000.04±0.00
RealGen-AE0.31±0.010.53±0.010.08±0.000.15±0.010.03±0.000.02±0.00
Retrieval-basedAE-KNN14.3±0.0316.4±0.050.57±0.010.59±0.020.15±0.010.15±0.01
LCTGen4.76±0.096.24±0.080.52±0.060.57±0.030.07±0.010.07±0.01
LCTGen w/o z14.2±0.0716.7±0.092.04±0.041.42±0.000.16±0.020.13±0.04
RealGen-AE-KNN13.1±0.0614.1±0.030.46±0.010.44±0.000.12±0.010.11±0.00
RealGen1.54±0.041.21±0.030.21±0.030.21±0.010.05±0.000.04±0.00
", + "bbox": [ + 225, + 433, + 767, + 556 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "work [18, 49, 64], we provide the results of the realism metrics and show the qualitative results of using RealGen for various downstream tasks. More results can be found in the appendix.", + "bbox": [ + 212, + 566, + 785, + 612 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Realism of generated scenario. To evaluate the realism of generated scenarios, we consider the following metrics and show the results in Table 2. We use the maximum mean discrepancy (MMD) [22] to measure the similarity in velocity and direction between the original scenarios and the generated scenarios. We also compare the mean average displacement error (mADE) and the mean final displacement error (mFDE) for the average reconstruction performance. To evaluate scene-level realism, we calculate the scene collision rate (SCR) and the off-road rate (ORR) following the metrics defined in [49]. The recon-based generation methods in Table 2 use the behavior of the target scenario as input. For this category, we compare RealGen-AE, only using the encoder and decoder modules, with three baseline methods. Due to the additional contrastive term, RealGen-AE achieves slightly worse performance than AE and Masked AE. However, RealGen is designed for retrieval-based generation, which uses retrieved scenarios rather than the target scenario as input. To fairly compare the generation performance, we design two baselines named AE-KNN and RealGen-AE-KNN,", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "W. Ding et al.", + "bbox": [ + 271, + 114, + 369, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f7e4b1d8ea5b2ab2e2a60bf5374f0576dc882d1b1ccc08b123363cef7d066d9b.jpg", + "image_caption": [ + "Fig. 5: Examples of tag-retrieved scenarios generated by RealGen for six different tags." + ], + "image_footnote": [], + "bbox": [ + 217, + 142, + 408, + 292 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/4d6dfbb28922c1ab5aae663a722917cc60ab7f09df7e9b4cee7401a5baa3e1f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 142, + 598, + 292 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8082e18975e3ca40ce7807079e934502c0f2402a76e01254a32374b487371eb9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 142, + 785, + 292 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "which use KNN to find the most similar behavior embedding to the target scenario and use this embedding as input to the decoder for generation. According to the results, we find that RealGen achieves comparable performance as recombined generation and outperforms baselines, which indicates the important role of the combiner in fusing the information of the retrieved scenarios.", + "bbox": [ + 212, + 338, + 784, + 412 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Generating tag-retrieved scenarios. Now we explore the qualitative performance of using RealGen for tag-retrieved generation. Given a target behavior tag, we first obtain several template scenarios from a small dataset and use them to retrieve more scenarios from the training database, which will be used for generation in the combiner. Since there is no existing dataset with tags, we manually labeled six tags - U-Turn, Overtaking, Left Lane Change, Right Lane Change, Left Turn, Right Turn - for", + "bbox": [ + 212, + 416, + 486, + 626 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/fad1fdbdbf1c8bd46fc0d4bf839287bca04e3454b860efd0c427fb504ffb442b.jpg", + "image_caption": [ + "Fig. 6: Examples of generating crash scenarios from RealGen. The shadow rectangles represent the initial positions of agents." + ], + "image_footnote": [], + "bbox": [ + 500, + 421, + 785, + 569 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "the nuScenes dataset to get template 1349 scenarios. We plot generated scenarios for each tag in Figure 5, where the left part of each example shows the given initial pose and map and the right part of each example shows the generated scenario from RealGen.", + "bbox": [ + 212, + 627, + 782, + 686 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Safety-critical scenario gen", + "text_level": 1, + "bbox": [ + 240, + 689, + 465, + 704 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "eration. Beyond the tag mentioned above, RealGen demonstrates its capacity for in-context", + "bbox": [ + 212, + 704, + 464, + 748 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "learning by generating critical and unseen crash scenarios, divergent from those of the training datasets. This is initiated by manually creating several crash scenario templates, guided by the scenarios recorded in the NHTSA Crash Report [40]. Subsequently, we generate crash scenarios using existing initial poses and maps of the dataset. Figure 6 illustrates six instances in which the shadowed rectangles denote the initial positions of agents and the red box highlights", + "bbox": [ + 212, + 750, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/56c74e7baa16b454f2f07d4fae748ebd85421f9ae3f3c9a826b30e838346e732.jpg", + "table_caption": [ + "Table 3: Safety-critical scenario generation." + ], + "table_footnote": [], + "table_body": "
MethodRealgen-AE-RRealGen-RRealGen
Collision Rate (↑)0.920.830.59
", + "bbox": [ + 470, + 708, + 787, + 744 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "RealGen: Retrieval Augmented Generation for Controllable Traffic Scenarios", + "bbox": [ + 220, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/ba309a4592db3904aa2298c1cfefddd52ff94edfda753a95e5220d70202aee9b.jpg", + "table_caption": [ + "Table 4: Results of human evaluation of controllability. (Details in Appendix ??)" + ], + "table_footnote": [], + "table_body": "
CategoryLeft-TurnRight-TurnLeft-Lane-ChangeRight-Lane-ChangeStraight
RealGen Preferred (%)81.891.797.893.3100.0
RealGen Score (0-5)4.27±1.054.27±0.693.96±1.944.17±1.933.94±2.27
LCTGen Score (0-5)2.15±1.442.08±1.192.42±1.952.0±1.962.14±2.31
", + "bbox": [ + 217, + 164, + 785, + 222 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "the point of collision. To quantitatively assess the performance of safety-critical scenarios, we compare RealGen with two baselines. RealGen-AE-R random samples behavior embeddings in the Realgen-AE model, and RealGen-R random retrieves behavior embeddings. According to the results in Table 3, we find that the scenarios generated by RealGen using crash scenarios as templates achieve the highest collision rate, which means that RealGen has more efficiency.", + "bbox": [ + 212, + 231, + 784, + 321 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Human evaluation of controllability. As there is no automatic way to evaluate controllability, we follow the protocol in [49] to perform A/B testing using human evaluation. We report the ratio of our method preferred as well as the absolute score (0-5) of both our method and the baseline. In Table 4, we find that scenarios generated by RealGen are highly preferred in most categories. The absolute score of RealGen is also much higher than that of LCTGen.", + "bbox": [ + 212, + 321, + 785, + 412 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Downstream task evaluation. A direct downstream task of our method is to use the generated data to augment the training dataset of trajectory", + "bbox": [ + 212, + 414, + 447, + 488 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/64e9fd20cd57f4e88c96916a0fd95c7279797f0194186514f8f2e7fd77cde6f7.jpg", + "table_caption": [ + "Table 5: Downstream task evaluation." + ], + "table_footnote": [], + "table_body": "
MethodOriginalRandom Aug.RealGen Aug.
mADE (↓)3.5442.9202.309
collision rate (↓)0.0490.0370.018
", + "bbox": [ + 460, + 431, + 784, + 479 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "prediction models. We use Autobots [21] as a predictor and report the results trained on different datasets in Table 5. Original means using the original data in nuScenes [5], Random Aug. means augmenting the original dataset with Gaussian noise, and RealGen Aug. means augmenting the original dataset with scenarios from RealGen. We observe that the model trained with the RealGen dataset achieves the lowest mADE and collision rate.", + "bbox": [ + 212, + 488, + 785, + 579 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion and Limitation", + "text_level": 1, + "bbox": [ + 214, + 590, + 509, + 606 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper proposes RealGen, a novel framework for traffic scenario generation that utilizes retrieval-augmented generation. Unlike previous approaches, which primarily rely on models replicating training distributions, RealGen demonstrates in-context learning abilities that synthesize scenarios by combining and modifying provided examples, enabling controlled generation. These scenarios can be automatically obtained from a retrieval system, which only requires the users to provide a few template scenarios as examples. We comprehensively evaluated the similarity autoencoder model for retrieval and the combiner model for generation. The findings indicate that RealGen achieves low reconstruction error and high generation quality.", + "bbox": [ + 212, + 613, + 785, + 763 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The primary limitation of our current method is that the behavior encoder focuses solely on agent trajectories, neglecting the intricate interactions between agents and lane maps in complex behaviors. Advancing the feature representation within the behavior encoder could significantly broaden RealGen's capacity to generate diverse and controllable traffic scenarios.", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "W. Ding et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Wenhao Ding contributed to this paper while being an intern at NVIDIA Research. Ding Zhao was partially supported by the National Science Foundation under grant CNS-2047454.", + "bbox": [ + 212, + 172, + 785, + 219 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 239, + 321, + 255 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016)", + "2. Blattmann, A., Rombach, R., Oktay, K., Müller, J., Ommer, B.: Retrievalaugmented diffusion models. Advances in Neural Information Processing Systems 35, 15309-15324 (2022)", + "3. Borgeaud, S., Mensch, A., Hoffmann, J., Cai, T., Rutherford, E., Millican, K., Van Den Driessche, G.B., Lespiau, J.B., Damoc, B., Clark, A., et al.: Improving language models by retrieving from trillions of tokens. In: International conference on machine learning. pp. 2206-2240. PMLR (2022)", + "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)", + "5. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)", + "6. Cai, D., Wang, Y., Bi, W., Tu, Z., Liu, X., Shi, S.: Retrieval-guided dialogue response generation via a matching-to-generation framework. In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). pp. 1866-1875 (2019)", + "7. Cao, Y., Xiao, C., Anandkumar, A., Xu, D., Pavone, M.: Advdo: Realistic adversarial attacks for trajectory prediction. In: European Conference on Computer Vision. pp. 36-52. Springer (2022)", + "8. Chen, D., Fisch, A., Weston, J., Bordes, A.: Reading wikipedia to answer open-domain questions. arXiv preprint arXiv:1704.00051 (2017)", + "9. Chen, H., Wang, J., Shao, K., Liu, F., Hao, J., Guan, C., Chen, G., Heng, P.A.: Traj-mae: Masked autoencoders for trajectory prediction. arXiv preprint arXiv:2303.06697 (2023)", + "10. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020)", + "11. Chen, W., Hu, H., Sahara, C., Cohen, W.W.: Re-imagen: Retrieval-augmented text-to-image generator. arXiv preprint arXiv:2209.14491 (2022)", + "12. Cheng, J., Mei, X., Liu, M.: Forecast-mae: Self-supervised pre-training for motion forecasting with masked autoencoders. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8679-8689 (2023)", + "13. Cuturi, M.: Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems 26 (2013)", + "14. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + ], + "bbox": [ + 218, + 266, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "RealGen: Retrieval Augmented Generation for Controllable Traffic Scenarios", + "bbox": [ + 220, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Ding, W., Chen, B., Li, B., Eun, K.J., Zhao, D.: Multimodal safety-critical scenarios generation for decision-making algorithms evaluation. IEEE Robotics and Automation Letters 6(2), 1551-1558 (2021)", + "16. Ding, W., Lin, H., Li, B., Eun, K.J., Zhao, D.: Semantically adversarial driving scenario generation with explicit knowledge integration. arXiv preprint arXiv:2106.04066 (2021)", + "17. Ding, W., Xu, M., Zhao, D.: Learning to collide: An adaptive safety-critical scenarios generating method. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE (2020)", + "18. Feng, L., Li, Q., Peng, Z., Tan, S., Zhou, B.: Trafficgen: Learning to generate diverse and realistic traffic scenarios. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 3567-3575. IEEE (2023)", + "19. Feydy, J., Séjourne, T., Vialard, F.X., Amari, S.i., Trouve, A., Peyré, G.: Interpolating between optimal transport and mmd using sinkhorn divergences. In: The 22nd International Conference on Artificial Intelligence and Statistics. pp. 2681-2690 (2019)", + "20. Gidaris, S., Singh, P., Komodakis, N.: Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728 (2018)", + "21. Girgis, R., Golemo, F., Codevilla, F., Weiss, M., D'Souza, J.A., Kahou, S.E., Heide, F., Pal, C.: Latent variable sequential set transformers for joint multi-agent motion prediction. arXiv preprint arXiv:2104.00563 (2021)", + "22. Gretton, A., Borgwardt, K.M., Rasch, M.J., Scholkopf, B., Smola, A.: A kernel two-sample test. The Journal of Machine Learning Research 13(1), 723-773 (2012)", + "23. Gulino, C., Fu, J., Luo, W., Tucker, G., Bronstein, E., Lu, Y., Harb, J., Pan, X., Wang, Y., Chen, X., et al.: Waymax: An accelerated, data-driven simulator for large-scale autonomous driving research. arXiv preprint arXiv:2310.08710 (2023)", + "24. Gutmann, M., Hyvarinen, A.: Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In: Proceedings of the thirteenth international conference on artificial intelligence and statistics. pp. 297-304. JMLR Workshop and Conference Proceedings (2010)", + "25. Guu, K., Lee, K., Tung, Z., Pasupat, P., Chang, M.: Retrieval augmented language model pre-training. In: International conference on machine learning. pp. 3929-3938. PMLR (2020)", + "26. Hanselmann, N., Renz, K., Chitta, K., Bhattacharyya, A., Geiger, A.: King: Generating safety-critical driving scenarios for robust imitation via kinematics gradients. In: European Conference on Computer Vision. pp. 335-352. Springer (2022)", + "27. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)", + "28. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020)", + "29. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)", + "30. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural computation 9(8), 1735-1780 (1997)", + "31. Hospedales, T., Antoniou, A., Micaelli, P., Storkey, A.: Meta-learning in neural networks: A survey. IEEE transactions on pattern analysis and machine intelligence 44(9), 5149-5169 (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "W. Ding et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Hu, A., Corrado, G., Griffiths, N., Murez, Z., Gurau, C., Yeo, H., Kendall, A., Cipolla, R., Shotton, J.: Model-based imitation learning for urban driving. Advances in Neural Information Processing Systems 35, 20703-20716 (2022)", + "33. Ibrihich, S., Oussous, A., Ibrihich, O., Esghir, M.: A review on recent research in information retrieval. Procedia Computer Science 201, 777-782 (2022)", + "34. Ivanovic, B., Song, G., Gilitschenski, I., Pavone, M.: trajdata: A unified interface to multiple human trajectory datasets. arXiv preprint arXiv:2307.13924 (2023)", + "35. Kim, J., Choi, S., Amplayo, R.K., Hwang, S.w.: Retrieval-augmented controllable review generation. In: Proceedings of the 28th International Conference on Computational Linguistics. pp. 2284-2295 (2020)", + "36. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)", + "37. Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Kuttler, H., Lewis, M., Yih, W.t., Rocktäschel, T., et al.: Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems 33, 9459-9474 (2020)", + "38. Li, Q., Peng, Z., Feng, L., Duan, C., Mo, W., Zhou, B., et al.: Scenarionet: Open-source platform for large-scale traffic scenario simulation and modeling. arXiv preprint arXiv:2306.12241 (2023)", + "39. Liu, Q., Yogatama, D., Blunsom, P.: Relational memory-augmented language models. Transactions of the Association for Computational Linguistics 10, 555-572 (2022)", + "40. NHTSA: Nhtsa crash viewer (2023), https://crashviewer.nhtsa.dot.gov/", + "41. Noroozi, M., Favaro, P.: Unsupervised learning of visual representations by solving jigsaw puzzles. In: European conference on computer vision. pp. 69-84. Springer (2016)", + "42. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)", + "43. Roshdi, A., Roohparvar, A.: Information retrieval techniques and applications. International Journal of Computer Networks and Communications Security 3(9), 373-377 (2015)", + "44. Schmidhuber, J.: Deep learning in neural networks: An overview. Neural networks 61, 85-117 (2015)", + "45. Shi, S., Jiang, L., Dai, D., Schiele, B.: Motion transformer with global intention localization and local movement refinement. Advances in Neural Information Processing Systems 35, 6531-6543 (2022)", + "46. Shurrab, S., Duwairi, R.: Self-supervised learning methods and applications in medical imaging analysis: A survey. PeerJ Computer Science 8, e1045 (2022)", + "47. Suo, S., Regalado, S., Casas, S., Urtasun, R.: Trafficsim: Learning to simulate realistic multi-agent behaviors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10400-10409 (2021)", + "48. Suo, S., Wong, K., Xu, J., Tu, J., Cui, A., Casas, S., Urtasun, R.: Mixsim: A hierarchical framework for mixed reality traffic simulation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9622-9631 (2023)", + "49. Tan, S., Ivanovic, B., Weng, X., Pavone, M., Kraehenbuehl, P.: Language conditioned traffic generation. In: Conference on Robot Learning. pp. 2714-2752. PMLR (2023)", + "50. Tan, S., Wong, K., Wang, S., Manivasagam, S., Ren, M., Urtasun, R.: Scenegen: Learning to generate realistic traffic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 892-901 (2021)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "RealGen: Retrieval Augmented Generation for Controllable Traffic Scenarios", + "bbox": [ + 222, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "51. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)", + "52. Villani, C., et al.: Optimal transport: old and new, vol. 338. Springer (2009)", + "53. Vincent, P., Larochelle, H., Bengio, Y., Manzagol, P.A.: Extracting and composing robust features with denoising autoencoders. In: Proceedings of the 25th international conference on Machine learning. pp. 1096-1103 (2008)", + "54. Wang, J., Pun, A., Tu, J., Manivasagam, S., Sadat, A., Casas, S., Ren, M., Urtasun, R.: Advsim: Generating safety-critical scenarios for self-driving vehicles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9909-9918 (2021)", + "55. Wang, Z., Nie, W., Qiao, Z., Xiao, C., Baraniuk, R., Anandkumar, A.: Retrieval-based controllable molecule generation. arXiv preprint arXiv:2208.11126 (2022)", + "56. Wu, P., Majumdar, A., Stone, K., Lin, Y., Mordatch, I., Abbeel, P., Rajeswaran, A.: Masked trajectory models for prediction, representation, and control. arXiv preprint arXiv:2305.02968 (2023)", + "57. Xu, P., Patwary, M., Shoeybi, M., Puri, R., Fung, P., Anandkumar, A., Catanzaro, B.: Megatron-cntrl: Controllable story generation with external knowledge using large-scale language models. arXiv preprint arXiv:2010.00840 (2020)", + "58. Yang, Y., Zhang, Q., Gilles, T., Batool, N., Folkesson, J.: Rmp: A random mask pretrain framework for motion prediction. arXiv preprint arXiv:2309.08989 (2023)", + "59. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of lmms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421 (2023)", + "60. Zhang, C., Zhang, C., Song, J., Yi, J.S.K., Zhang, K., Kweon, I.S.: A survey on masked autoencoder for self-supervised learning in vision and beyond. arXiv preprint arXiv:2208.00173 (2022)", + "61. Zhang, C., Tu, J., Zhang, L., Wong, K., Suo, S., Urtasun, R.: Learning realistic traffic agents in closed-loop. In: 7th Annual Conference on Robot Learning (2023)", + "62. Zhang, X., Bosselut, A., Yasunaga, M., Ren, H., Liang, P., Manning, C.D., Leskovec, J.: Greaselm: Graph reasoning enhanced language models. In: International conference on learning representations (2021)", + "63. Zhong, Z., Rempe, D., Chen, Y., Ivanovic, B., Cao, Y., Xu, D., Pavone, M., Ray, B.: Language-guided traffic simulation via scene-level diffusion. arXiv preprint arXiv:2306.06344 (2023)", + "64. Zhong, Z., Rempe, D., Xu, D., Chen, Y., Veer, S., Che, T., Ray, B., Pavone, M.: Guided conditional diffusion for controllable traffic simulation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 3560-3566. IEEE (2023)" + ], + "bbox": [ + 215, + 147, + 785, + 686 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "W. Ding et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/RealGen_ Retrieval Augmented Generation for Controllable Traffic Scenarios/4b2a9c03-0a68-49b3-9375-225cd793a8b3_model.json b/2024/RealGen_ Retrieval Augmented Generation for Controllable Traffic Scenarios/4b2a9c03-0a68-49b3-9375-225cd793a8b3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..119a60e6b06f8a0d97a0970426ba2a478cb30c75 --- /dev/null +++ b/2024/RealGen_ Retrieval Augmented Generation for Controllable Traffic Scenarios/4b2a9c03-0a68-49b3-9375-225cd793a8b3_model.json @@ -0,0 +1,2590 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.236, + 0.141, + 0.768, + 0.185 + ], + "angle": 0, + "content": "RealGen: Retrieval Augmented Generation for Controllable Traffic Scenarios" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.212, + 0.772, + 0.243 + ], + "angle": 0, + "content": "Wenhao Ding\\*1,2, Yulong Cao\\*2, Ding Zhao\\*1, Chaowei Xiao\\*2,3, and Marco Pavone\\*2,4" + }, + { + "type": "text", + "bbox": [ + 0.4, + 0.255, + 0.603, + 0.269 + ], + "angle": 0, + "content": "1 Carnegie Mellon University" + }, + { + "type": "text", + "bbox": [ + 0.433, + 0.269, + 0.571, + 0.282 + ], + "angle": 0, + "content": "2 NVIDIA Research" + }, + { + "type": "text", + "bbox": [ + 0.45, + 0.283, + 0.555, + 0.296 + ], + "angle": 0, + "content": "3 UW-Madison" + }, + { + "type": "text", + "bbox": [ + 0.429, + 0.297, + 0.576, + 0.311 + ], + "angle": 0, + "content": "4 Stanford University" + }, + { + "type": "text", + "bbox": [ + 0.39, + 0.312, + 0.615, + 0.325 + ], + "angle": 0, + "content": "{wenhaod, yulongc}@nvidia.com" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.35, + 0.744, + 0.628 + ], + "angle": 0, + "content": "Abstract. Simulation plays a crucial role in the development of autonomous vehicles (AVs) due to the potential risks associated with real-world testing. Although significant progress has been made in the visual aspects of simulators, generating complex behavior among agents remains a formidable challenge. It is not only imperative to ensure realism in the scenarios generated but also essential to incorporate preferences and conditions to facilitate controllable generation for AV training and evaluation. Traditional methods, which rely mainly on memorizing the distribution of training datasets, often fail to generate unseen scenarios. Inspired by the success of retrieval augmented generation in large language models, we present RealGen, a novel retrieval-based in-context learning framework for traffic scenario generation. RealGen synthesizes new scenarios by combining behaviors from multiple retrieved examples in a gradient-free way, which may originate from templates or tagged scenarios. This in-context learning framework endows versatile generative capabilities, including the ability to edit scenarios, compose various behaviors, and produce critical scenarios. Evaluations show that RealGen offers considerable flexibility and controllability, marking a new direction in the field of controllable traffic scenario generation. Check our project website for more information: https://realgen.github.io." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.647, + 0.377, + 0.663 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Simulation is indispensable in the development of autonomous vehicles (AVs), primarily due to the considerable risks associated with training and evaluating these systems in real-world conditions. The biggest challenge in simulations lies in achieving realistic driving scenarios, as this realism influences the discrepancy between AV performance in simulated and actual environments. Although advancements in high-quality graphical engines have significantly enhanced the perception quality of simulators, the realism of agent behavior remains constrained because of the complicated interactions among naturalistic agents. To counteract this issue, data-driven simulation has emerged as a promising approach in the realm of autonomous driving, which leverages real-world scenario datasets to accurately generate the behaviors of agents." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.129 + ], + "angle": 0, + "content": "W. Ding et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "With the rapid achievement of deep generative models [29] and imitation learning algorithms [32], current data-driven simulations [23, 38] can generate scenarios that closely mimic human driver behavior. However, the effectiveness of these simulations in accelerating the development of AV is limited. This limitation stems from the need for scenarios that meet specific conditions tailored for targeted training and evaluation. Achieving such controllability in simulations is challenging due to the complex nature of driving scenarios, which involve intricate interactions, diverse road layouts, and varying traffic regulations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.282, + 0.432, + 0.554 + ], + "angle": 0, + "content": "In pursuit of controllability, existing work applies additional guidance, typically in the form of constraint functions [16, 64] or languages [49, 63], to pre-trained scenario generative models. Regularization of the generation process through these tools is straightforward and effective, yet it encounters two principal challenges. First, the training of scenario generative models typically utilizes naturalistic datasets, which might not encompass the specific scenarios desired as per the control signals. Even if such" + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.285, + 0.561, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.445, + 0.432, + 0.56, + 0.442 + ], + "angle": 0, + "content": "(a) Conventional method" + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.286, + 0.785, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.433, + 0.71, + 0.442 + ], + "angle": 0, + "content": "(b) Our method" + }, + { + "type": "image_caption", + "bbox": [ + 0.44, + 0.45, + 0.788, + 0.547 + ], + "angle": 0, + "content": "Fig. 1: (a) Conventional methods make the model memorize the data distribution for generating. (b) In contrast, our method employs a retriever to query datasets (including external data obtained after training) and uses a generative model to generate scenarios by integrating the information from the retrieved scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.646 + ], + "angle": 0, + "content": "scenarios exist within the dataset, they are often omitted because of the rarity of long-tail data. The second challenge is that the representation of the guidance to the generative model may not be sufficiently expressive to accurately depict complex scenarios, such as specifying intricate interactions among multiple vehicles, using language. These limitations underscore the need for a more sophisticated and nuanced framework for controllable scenario generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Retrieval Augmented Generation (RAG) [8, 25], which enhances the generative process by querying related information from external databases, represents great potential in the domain of large language models [37]. In contrast to conventional models that memorize all knowledge within their parameters, as shown in Figure 1(a), RAG models, shown in Figure 1(b), learn to generate comprehensive outputs by retrieving pertinent knowledge from a database, based on the input provided. A notable aspect of RAG is the ability of the database to undergo updates even after the model has been trained, allowing continuous improvement and adaptation. This flexible framework offers possibilities for controllable scenario generation by using appropriate template scenarios as input and facilitating the generation that is not only realistic but also aligned with specific training and evaluation requirements." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.222, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RealGen: Retrieval Augmented Generation for Controllable Traffic Scenarios" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "In this work, we present RealGen, a retrieval augmented generation framework for generating traffic scenarios. This framework, as shown in Figure 2, begins with the training of an encoder through contrastive self-supervised learning [42] to allow the retrieval process to query similar scenarios in a latent embedding space. Leveraging this latent representation, we subsequently train a generative model that combines retrieved scenarios to create novel scenarios. The key contributions of this paper are summarized below." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.262, + 0.787, + 0.306 + ], + "angle": 0, + "content": "- We develop a novel contrastive autoencoder model to extract scenario embeddings as latent representations, which can be used for a wide range of downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.308, + 0.785, + 0.337 + ], + "angle": 0, + "content": "- We propose the first retrieval augmented generation framework using the latent representation tailored for controllable driving scenario generation." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.338, + 0.785, + 0.368 + ], + "angle": 0, + "content": "- We validate our framework through qualitative and quantitative metrics, demonstrating strong flexibility and controllability of generated scenarios." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.262, + 0.787, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.391, + 0.388, + 0.407 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.423, + 0.607, + 0.438 + ], + "angle": 0, + "content": "2.1 Traffic Modeling and Scenario Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.629 + ], + "angle": 0, + "content": "Research on traffic modeling using generative models has received considerable attention in recent works. Notably, ScenarioNet [38] and Waymax [23] employ large-scale data to train imitation learning models, facilitating the generation of realistic multi-agent scenarios in simulations. SceneGen [50] utilizes the Long Short Term Memory [30] module to autoregressively generate the trajectories for vehicles and pedestrians based on provided maps. Additionally, TrafficSim [47] learns the multi-agent behaviors from real-world data, and TrafficGen [18] proposes a transformer-based autoencoder architecture to accurately model complex interactions involving multiple agents. MixSim [48] builds a reactive digital twin and finds safety-critical scenarios with black-box optimization. RTR [61] models reactive behaviors of vehicles using the combination of reinforcement learning and imitation learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.735 + ], + "angle": 0, + "content": "Numerous prior studies have employed adversarial generation to synthesize rare yet critical scenarios. L2C [17] utilizes a reinforcement learning framework, where the reward for scenario generation is based on the collision rate. To enhance the realism in purely adversarial generation methods, MMG [15] incorporates the data distribution as regularization. Further developments such as AdvSim [54], AdvDO [7], and KING [26] integrate vehicle dynamics to directly optimize the trajectory to find critical scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "The proliferation of large language models (LLMs) has led to recent approaches in generating traffic scenarios with language as conditions to follow instructions from humans. CTG++ [63] replaces the gradient guidance process of CTG [64] with cost functions generated by an LLM. Leveraging the power of GPT-4 [59], this framework can generate diverse motion data with language conditions. Additionally, LCTGen [49] also harnesses the strengths of GPT-4 to generate a heuristic intermediate representation of scenarios from language" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "W. Ding et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.178 + ], + "angle": 0, + "content": "inputs. Subsequently, a generative model pre-trained on open-source datasets generates various scenarios with this representation as inputs." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.198, + 0.533, + 0.214 + ], + "angle": 0, + "content": "2.2 Retrieval-augmented Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.222, + 0.788, + 0.328 + ], + "angle": 0, + "content": "Information retrieval (IR) [33] is the procedure of representing and searching a collection of data with the goal of extracting knowledge to satisfy the query of users. Predominantly utilized in search engines and digital libraries [43], IR primarily deals with the information stored in textual formats. Recently, IR has extended across diverse sectors, notably enhancing the quality of outputs in areas such as language modeling [3, 39], question answering [25, 62], image creation [2, 11], and the generation of molecular [55]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.328, + 0.79, + 0.448 + ], + "angle": 0, + "content": "One intuitive usage of the retrieved information is to enhance input data through several methods, such as merging the original data and retrieved data [37], employing attention mechanism [3], or extracting skeleton [6]. The rationale behind retrieval systems stems from the impracticality of encoding all knowledge within model parameters, especially considering the dynamic nature of knowledge that evolves with human activities. Consequently, the capability to access external knowledge databases can significantly improve the precision and quality of generated responses, as evidenced in applications involving LLMs [3, 37]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.449, + 0.788, + 0.57 + ], + "angle": 0, + "content": "Another usage of retrieved data, particularly pertinent to this study, involves the controllable generation by integrating desired features retrieved from the dataset. In [35], the authors generate product reviews with controllable information about the user, product, and rating. Similarly, the process of story creation can be viewed as a blend of external story databases with selected text fragments [57] Furthermore, [55] explores the controllable generation of molecules to meet various constraints, utilizing a database comprising simple molecules that typically meet only one of these constraints." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.591, + 0.533, + 0.607 + ], + "angle": 0, + "content": "2.3 Self-supervised Feature Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.734 + ], + "angle": 0, + "content": "The necessity for manual labeling in supervised learning often results in human biases, extraneous noises, and labor-intensive efforts. Therefore, the paradigm of self-supervised learning (SSL) [46] is garnering heightened interest, particularly for its promising applications in language modeling and image interpretation. SSL algorithms usually learn implicit representations from extensive pools of unlabeled data without relying on human annotations. Generally, this line of research can be bifurcated into two categories [60]: generative SSL and discriminative SSL." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In the domain of generative SSL, models employ an autoencoder to convert input data into a latent representation, followed by a reconstruction process. As an illustrative example, Denoising Autoencoder [53], an early example of this approach, reconstructs images from their noisy version to extract representative features. Another technique in this field is masked modeling, extensively utilized in language models such as GPT-3 [4] and BERT [14], which predicts the masked token to gain a semantic understanding of the text. More recently, the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.222, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "RealGen: Retrieval Augmented Generation for Controllable Traffic Scenarios" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "Masked Autoencoder [27] has emerged as a powerful framework, demonstrating its efficacy across a variety of downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.329 + ], + "angle": 0, + "content": "Discriminative SSL typically optimizes a discriminative loss to learn representations from sets of anchor, positive, and negative samples. Without ground-truth labels, these pairs are often constructed through solving jigsaw puzzles [41] or making geometry-based prediction [20]. A prominent example of discriminative SSL is contrastive learning, which brings samples from the same class closer while distancing those from different classes. Representative works such as MoCo [28] and SimCLR [10] learn embeddings by capturing invariant features between original data and its augmented variants. Additionally, InfoNCE [42], a method grounded in noise contrastive estimation [24], is widely used due to its effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.355, + 0.607, + 0.373 + ], + "angle": 0, + "content": "3 A Latent Representation of Scenario" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.788, + 0.525 + ], + "angle": 0, + "content": "A crucial element within the retrieval system is the selection metric for data retrieval, which is typically implemented through a distance function between the query sample and the candidate samples in the database. Unlike text, which can be converted into word embeddings, traffic scenarios encompass sequential behaviors and intricate interactions among entities, complicating the establishment of a similarity metric for these scenarios. Consequently, in this section, we introduce a scenario autoencoder to extract latent representations that facilitate the assessment of similarity between various traffic scenarios, a vital component in our RAG framework." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.549, + 0.421, + 0.563 + ], + "angle": 0, + "content": "3.1 Scenario Definition" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.576, + 0.788, + 0.684 + ], + "angle": 0, + "content": "Each scenario is characterized by the trajectories \\(\\tau \\in \\mathbb{R}^{M\\times T\\times 5}\\), encompassing \\(M\\) agents over a maximum of \\(T\\) time steps. The trajectory of each agent is composed of parameters \\([x,y,v,c,s]\\), signifying position \\(x\\), position \\(y\\), velocity \\(v\\), cosine of the heading \\(c\\), and sine of the heading \\(s\\). The initial state of these entities is denoted as \\(\\tau_0\\in \\mathbb{R}^{M\\times 5}\\). Furthermore, the map is encapsulated by \\(m\\in \\mathbb{R}^{S\\times 4}\\), composed of \\(S\\) lane segments. The attributes of these points \\([x_s,y_s,x_e,y_e]\\) correspond to the starting and ending positions of each segment." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.708, + 0.445, + 0.722 + ], + "angle": 0, + "content": "3.2 Scenario Autoencoder" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Autoencoders [44] are used to learn compressed latent representations of high-dimensional data, where an encoder projects the data into latent code and a decoder reconstructs the code in the data space. Given that traffic scenarios encompass spatial and temporal dynamics from multiple agents, we design a hierarchical encoder structure alternating between spatial and temporal layers based on transformer architecture [51]. As depicted on the left of Figure 2, our design incorporates three encoders for behavior, map, and initial position." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.129 + ], + "angle": 0, + "content": "W. Ding et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.143, + 0.789, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.327, + 0.789, + 0.427 + ], + "angle": 0, + "content": "Fig. 2: Left: the training pipeline for the encoders and decoder aimed at learning latent embeddings of scenarios. A contrastive loss is applied to behavior embeddings to ensure invariance to absolute positions. Middle: the training pipeline for the combiner with frozen encoder and decoder parameters. We use K-Nearest Neighbors (KNN) to retrieve scenarios similar to a template scenario in the dataset and use the retrieved behaviors to reconstruct the template scenario. Right: the generation pipeline with a retriever and a generator." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.44, + 0.789, + 0.683 + ], + "angle": 0, + "content": "First, we design a behavior encoder \\( E_{b} \\) with a spatial-temporal transformer structure that comprises \\( L_{e} \\) temporal transformer encoders \\( (\\mathrm{Encoder}_{t}^{i}) \\) and \\( L_{e} \\) spatial transformer encoders \\( (\\mathrm{Encoder}_{s}^{i}) \\), where \\( i \\in \\{1, \\dots, L_{e}\\} \\). The trajectory \\( \\tau \\) is first transformed into a latent embedding \\( z_{b} \\), where \\( H \\) denotes the hidden dimension, through a Multiple Linear Perception (MLP) module. Subsequently, an alternating encoding procedure extracts spatial and temporal features from \\( z_{b} \\). To preserve temporal information, we add a sinusoidal positional embedding (PE) to the embedding before employing the temporal transformer. To retain distinct agent information and further compress the behavior feature, we calculate the mean of the latent embedding across the temporal dimension, resulting in the final behavior embedding \\( z_{b} \\in \\mathbb{R}^{M \\times H} \\). Besides the behavior encoder, we also process the map information represented by lane vectors with the attention mechanism [51]. Similarly to the first step in \\( E_{b} \\), we project \\( m \\) to the latent space with an MLP module. Then we apply a multi-head attention module (MHA) with layer normalization (LN) [1] to acquire the map embedding \\( z_{m} \\in \\mathbb{R}^{S \\times H} \\) with a learnable query embedding \\( q_{m} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.684, + 0.79, + 0.79 + ], + "angle": 0, + "content": "In the decoder, a spatial-temporal transformer architecture, similar to the behavior encoder, is designed to obtain the behavior embedding \\( z_{b} \\) through \\( L_{d} \\) times encoding. Given that \\( z_{b} \\) lacks the temporal dimension, we first replicate it \\( T \\) times and then add a PE before entering it into the temporal transformer encoder. Throughout the decoding process, the map embedding \\( z_{m} \\) is injected by a cross-attention mechanism: \\( z_{b} \\gets z_{b} + \\mathrm{MHA}(z_{b}, z_{m}, z_{m}) \\). MHA(\\( Q, K, V \\)) is the multi-head attention with \\( Q, K, V \\) representing query, key, and value:" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.804, + 0.786, + 0.831 + ], + "angle": 0, + "content": "\\[\n\\mathrm {M H A} (Q, K, V) = \\text {C o n c a t e n a t e} \\left(h _ {1}, \\dots , h _ {i}\\right) W ^ {O},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.826, + 0.609, + 0.844 + ], + "angle": 0, + "content": "\\[\nh _ {i} = \\mathrm {A t t e n t i o n} (Q W _ {i} ^ {Q}, K W _ {i} ^ {K}, V W _ {i} ^ {V}),\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.222, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "RealGen: Retrieval Augmented Generation for Controllable Traffic Scenarios" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "code_caption", + "bbox": [ + 0.227, + 0.15, + 0.572, + 0.165 + ], + "angle": 0, + "content": "Algorithm 1: Details of Encoder and Decoder" + }, + { + "type": "algorithm", + "bbox": [ + 0.22, + 0.167, + 0.789, + 0.362 + ], + "angle": 0, + "content": "1 Behavior Encoder \\(E_{b}\\) \\((\\tau)\\): 14 Initial Pose Encoder \\(E_{i}\\) \\((\\tau_{0})\\): \n2 \\(z_{b} \\gets \\mathrm{MLP}(\\tau) // \\text{projection}\\) 15 \\(z_{i} \\gets \\mathrm{MLP}(\\tau_{0}) // \\text{projection}\\) \n3 for \\(i\\) in \\([1, \\dots, L_{e}]\\) do 16 return initial pose embedding \\(z_{i}\\) \n4 \\(z_{b} \\gets \\mathrm{Encoder}_{t}^{i}(\\mathrm{PE} + z_{b})\\) 17 Decoder \\(D(z_{i}, z_{b}, z_{m})\\): \n5 \\(z_{b} \\gets \\mathrm{Encoder}_{s}^{i}(z_{b})\\) 18 \\(z_{r} \\gets z_{b} + \\mathrm{MHA}(z_{b}, z_{i}, z_{i})\\) \n6 \\(z_{b} \\gets \\text{mean}(z_{b})\\) 19 for \\(i\\) in \\([1, \\dots, L_{d}]\\) do \n7 return behavior embedding \\(z_{b}\\) 20 \\(z_{r} \\gets \\mathrm{Encoder}_{t}^{i}(\\mathrm{PE} + z_{r})\\) \n8 Map Encoder \\(E_{m}\\): 21 \n9 Initialize a learnable query \\(q_{m}\\) 22 \n10 \\(z_{m} \\gets \\mathrm{MLP}(m) // \\text{projection}\\) 23 \\(\\hat{\\tau} \\gets \\mathrm{MLP}(z_{r}) // \\text{projection}\\) \n11 \\(z_{m} \\gets \\mathrm{LN}(\\mathrm{MHA}(q_{m}, z_{m}, z_{m}))\\) 24 return reconstructed trajectory \\(\\hat{\\tau}\\) \n12 \\(z_{m} \\gets \\mathrm{LN}(z_{m} + \\mathrm{MLP}(z_{m}))\\) \n13 return map embedding \\(z_{m}\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.789, + 0.486 + ], + "angle": 0, + "content": "where \\( W^{O} \\), \\( W_{i}^{Q} \\), \\( W_{i}^{K} \\), and \\( W_{i}^{V} \\) are learnable parameters. The last component of the decoder is an MLP that projects the hidden embedding back into the data domain, resulting in the reconstructed trajectory \\( \\hat{\\tau} \\). To train the encoders and decoder, the mean square error is used as the reconstruction loss, expressed as \\( \\mathcal{L}_r = |\\hat{\\tau} -\\tau |_2 \\). Comprehensive descriptions of the encoding and decoding processes are provided in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.789, + 0.577 + ], + "angle": 0, + "content": "It is important to note that, while the current design of the autoencoder is sufficient for acquiring compressed representations of scenarios, these representations are not invariant to the absolute coordinates and the order of the agents. This limitation can lead to substantial embedding distances between similar scenarios. To address these issues, two enhancements are proposed and detailed in the following section." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.602, + 0.577, + 0.616 + ], + "angle": 0, + "content": "3.3 Invariant feature with contrastive loss" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.789, + 0.78 + ], + "angle": 0, + "content": "To enhance the representation of scenario similarity in the behavior embedding \\( z_{b} \\), we employ contrastive learning to acquire invariant features. Specifically, we integrate InfoNCE [42] as an additional loss function \\( \\mathcal{L}_c \\), which optimizes categorical cross-entropy to distinguish a positive sample from a batch of negative samples. In practice, for a query embedding \\( z_{b} \\), a positive sample \\( z_{b}^{+} \\) is generated by applying random rotation and translation to the original scenario \\( \\tau, m \\). Meanwhile, negative samples \\( Z_{b}^{-} \\) are selected from the remaining samples in the same batch. In conventional InfoNCE, to bring \\( z_{b} \\) and \\( z_{b}^{+} \\) closer together, the inner product is calculated between the embedding \\( z_{b} \\) and the set \\( z_{b}^{+}, Z_{b}^{-} \\), employing cross-entropy loss to identify the index of the positive sample." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.781, + 0.789, + 0.841 + ], + "angle": 0, + "content": "To ensure that the ordering of the agents does not affect the outcomes, \\( z_{b} \\in \\mathbb{R}^{M \\times H} \\) should exhibit permutation invariance across the dimension \\( M \\). Otherwise, simply stacking them into a one-dimensional vector could erroneously represent distinct scenarios as significantly different. To establish permutation" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.129 + ], + "angle": 0, + "content": "W. Ding et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.238 + ], + "angle": 0, + "content": "invariance, we adopt the Wasserstein distance \\( W_{2} \\) [52] as a more suitable metric than cosine distance for assessing the similarity. Taking \\( z_{b} \\) as a distribution representing \\( M \\) individual behaviors, the intuition behind this choice is that the Wasserstein distance identifies the minimal adjustments necessary for behaviors in one scenario to reflect those in another. With such a powerful tool, we have the following contrastive loss:" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.244, + 0.786, + 0.284 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} = - \\sum_ {z _ {b}} \\log \\frac {\\exp \\left[ - W _ {2} \\left(z _ {b} , z _ {b} ^ {+}\\right) \\right]}{\\sum_ {z ^ {\\prime} \\in \\left\\{z _ {b} ^ {+} , Z _ {b} ^ {-} \\right\\}} \\exp \\left[ - W _ {2} \\left(z _ {b} , z ^ {\\prime}\\right) \\right]}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.29, + 0.79, + 0.427 + ], + "angle": 0, + "content": "The implementation of this loss results in the behavior embedding \\( z_{b} \\) lacking absolute coordinate data, making the decoder incapable of accurately reconstructing the precise trajectory. To handle this problem, we add the initial poses \\( \\tau_0 \\) of all agents as a supplementary input for the decoder. We encode these poses into \\( z_{i} \\) with an MLP encoder \\( \\mathrm{Enc}_i \\), and then integrate it into the decoder with an MHA module \\( z_{r} \\gets z_{b} + \\mathrm{MHA}(z_{b},z_{i},z_{i}) \\). Additional specifics regarding the initial pose encoder are detailed in Algorithm 1. In the training stage, the objective is to minimize a combined loss function \\( \\mathcal{L} = \\mathcal{L}_r + \\lambda \\mathcal{L}_c \\), where \\( \\lambda \\) serves as a weighting factor for contrastive loss and is uniformly set at 0.1 for all experiments." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.447, + 0.664, + 0.465 + ], + "angle": 0, + "content": "4 Retrieval Augmented Scenario Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.788, + 0.583 + ], + "angle": 0, + "content": "The autoencoder structure introduced in Section 3 builds a one-to-one generation framework, which still needs additional modules to support the retrieval augmented generation, which is a many-to-one framework. In this section, we introduce a module called Combiner, which takes input behavior embeddings from multiple scenarios and outputs the combined embedding. To train this module, we designed a KNN-based training pipeline that forces the model to learn to combine and edit existing scenarios." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.6, + 0.48, + 0.616 + ], + "angle": 0, + "content": "4.1 The Training of Combiner" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.622, + 0.788, + 0.728 + ], + "angle": 0, + "content": "Unlike the trajectory prediction task [45] which has ground truth as the optimization target, the objective in training the combiner is to learn the alignment of behaviors in the retrieved scenarios with the initial pose and the specified map. A well-trained combiner should be able to compose behaviors from these retrieved scenarios, thereby generating new scenarios that have a resemblance to all the retrieved scenarios. This type of objective aligns with the concept of meta-learning, or 'learning to learn', as described in [31]." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.728, + 0.788, + 0.803 + ], + "angle": 0, + "content": "Within our training framework, for a given query scenario \\(\\tau, m\\) in the dataset, we initially utilize the behavior encoder to obtain \\(z_{b}\\) and then use KNN to identify \\(K\\) similar behavior embeddings from the database, denoted \\(z_{ret} = [z_{b,1}, \\dots, z_{b,K}]\\). Building on this, we propose a model comprising two MHA modules as the combiner:" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.809, + 0.786, + 0.834 + ], + "angle": 0, + "content": "\\[\nz _ {r a g} \\leftarrow z _ {i} + \\mathrm {M H A} \\left(z _ {i}, z _ {r e t}, z _ {r e t}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.829, + 0.625, + 0.843 + ], + "angle": 0, + "content": "\\[\nz _ {r a g} \\leftarrow z _ {r a g} + \\mathrm {M H A} (z _ {r a g}, z _ {m}, z _ {m}),\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.222, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "RealGen: Retrieval Augmented Generation for Controllable Traffic Scenarios" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.239 + ], + "angle": 0, + "content": "where \\( z_{i} \\gets E_{i}(x_{0}) \\) and \\( z_{m} \\gets E_{m}(m) \\) are the initial pose embedding and the map embedding of the query scenario. Gradients are stopped for \\( z_{i} \\) and \\( z_{m} \\). Assuming that the \\( K \\) nearest scenarios adequately represent the query scenario, it should be feasible to reconstruct the behavior of the query scenario, \\( z_{b} \\), using the retrieved scenario, \\( z_{rag} \\). Consequently, we employ the following loss function as our training goal:" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.247, + 0.786, + 0.264 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r a g} = \\left\\| D \\left(z _ {i}, z _ {r a g}, z _ {m}\\right) - \\tau \\right\\| _ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.27, + 0.788, + 0.331 + ], + "angle": 0, + "content": "wherein the parameters of the decoder \\( D\\phi \\) remain fixed during training. Essentially, this training method is designed to learn an \"inverse\" operation of the KNN, aiming to reconstruct a query scenario that closely resembles all \\( K \\) retrieved scenarios in the database." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.35, + 0.46, + 0.367 + ], + "angle": 0, + "content": "4.2 The generation pipeline" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.784, + 0.432 + ], + "angle": 0, + "content": "The pipeline for scenario generation, as delineated in Figure 2, is categorized into retriever and generator components. The generation process is expedited by pre-processing all scenarios in the database using a behavior encoder, which yields behavior embeddings that facilitate efficient similarity computation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.785, + 0.583 + ], + "angle": 0, + "content": "The retriever enhances the versatility of generation by dividing the process into two stages. In the initial stage, users can choose a set of template scenarios that depict specific conditions. These may include manually annotated scenarios with tags denoting actions like left or right turns, enabling the generation of additional scenarios under similar tags or even a combination thereof. In addition, templates can include critical and interesting scenarios collected from real-world data. Solely relying on these templates for generation could be limited, which is mitigated by incorporating a secondary phase, which employs a KNN approach to fetch high-quality scenarios from a vast and unlabeled database to augment adaptability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.786, + 0.646 + ], + "angle": 0, + "content": "Subsequently, the generator, comprising a combiner and a decoder, follows the same inference as combiner training with the initial pose and lane map specified by the user. We obtain the RAG embedding \\( z_{rag} \\) with Eq. (3), and then infer the generated scenario through \\( \\tau_{rag} \\gets D(z_i, z_{rag}, z_m) \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.664, + 0.367, + 0.683 + ], + "angle": 0, + "content": "5 Experiment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.693, + 0.784, + 0.77 + ], + "angle": 0, + "content": "In this section, we start with an overview of the experimental setup and details regarding the implementation of the RealGEN framework. Then we discuss the baselines, including different types of autoencoders and variants of RealGen, used for comparison. In the discussion of the findings, we respectively evaluate the quality of scenario embedding and the generation capability of RAG." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.567, + 0.804 + ], + "angle": 0, + "content": "5.1 Settings and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.842 + ], + "angle": 0, + "content": "We conducted all training and evaluation with the nuScenes [5] dataset using the trajdata [34] package for data loading and processing. Each scenario spans" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "W. Ding et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.223 + ], + "angle": 0, + "content": "a duration of 8 seconds, operating at a frequency of \\(2\\mathrm{Hz}\\), and encompasses a maximum of 11 agents. We filter out agents that travel less than 3 meters in 8 seconds and select 11 agents closest to the ego vehicle. The map contains 100 lanes (each lane has 20 points) that are ordered according to the distance between the center of the lane and the center of the ego vehicle." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.788, + 0.299 + ], + "angle": 0, + "content": "To train RealGen, we use Adam [36] as the optimizer to update the parameters of the autoencoder and the combiner. To make the calculation of the Wasserstein distance efficient, we use the Sinkhorn distance [13], an entropy regularized approximation of the Wasserstein distance [52], with implementation in the GeomLoss [19] package." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.319, + 0.341, + 0.333 + ], + "angle": 0, + "content": "5.2Baselines" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.79, + 0.555 + ], + "angle": 0, + "content": "In the experimental section, we evaluate the following reconstruction-based generative models as baselines for comparison. Autoencoder (AE) shares the same behavior encoder, map encoder, and decoder structures as RealGen, serving as the most straightforward baseline for scenario reconstruction. Contrastive AE mirrors the structure of the RealGen autoencoder but omits the initial pose as absolute information. Masked AE is a self-supervised learning baseline, which has been investigated for trajectory data as described in [9, 12, 56, 58]. To evaluate the controllable generation capability, we select a state-of-the-art modelLCTGen [49] as the baseline, which takes a high-level agent representation \\( z \\) and a vector map \\( m \\) as input. We also consider LCTGen w/o \\( z \\), a variant model proposed in the original paper to show the reconstruction results without using the agent information. Given that our method has two phases, we refer to the autoencoder component specifically as RealGen-AE and the complete model as RealGen." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.575, + 0.553, + 0.59 + ], + "angle": 0, + "content": "5.3 Evaluation of Behavior Embedding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.629 + ], + "angle": 0, + "content": "We first evaluate the quality of the learned representation of behavior, which is critical for the following retrieval and generation processes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.735 + ], + "angle": 0, + "content": "Visualizing similar and dissimilar scenarios. After training the autoencoder with contrastive loss, the distance between behavior embeddings can be used as an indicator of the similarity of the two scenarios. To validate this statement, we visualize qualitative examples of using a query to find the most similar (minimal \\( W_{2} \\) distance) and dissimilar (maximal \\( W_{2} \\) distance) scenarios in Figure 3. We observe that the most similar scenario contains the same behavior and number of vehicles as in the query scenario." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Classifying Scene ID with behavior embedding. To further provide quantitative results on how well the behavior embedding encodes behavior information, we leverage the Scene ID of the nuScenes dataset. Each scene typically lasts 20 seconds and our scenario segments last 8 seconds, so we get multiple segments belonging to the same Scene ID. We assume that the segments having the same Scene ID have similar behaviors so that we can calculate the accuracy by using the distance to find the closest segment. In Figure 4, we summarize" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.222, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "RealGen: Retrieval Augmented Generation for Controllable Traffic Scenarios" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.788, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.379, + 0.788, + 0.421 + ], + "angle": 0, + "content": "Fig.3: Qualitative evaluation of similar and dissimilar scenarios calculated by our scenario embedding. Rectangles represent the initial poses of vehicles and the lines represent the future trajectories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.441, + 0.788, + 0.607 + ], + "angle": 0, + "content": "the results in the left part, where top-\\(k\\) means we calculate the accuracy for the closest \\(k\\) segments. To validate the effectiveness of \\(W_{2}\\) distance, we also permute the order of agents in the behavior embedding, which is denoted as cosine-permuted and \\(W_{2}\\)-permuted. We find that the cosine distance cannot deal with the permuted setting, as the order is important when stacking the embedding of agents. Our method performs well in top-1 and top-5 settings but badly in others, which could be explained by the fact that segments in one scene have very different behaviors. To validate this, we plot the distance matrix between segments for 11 scenes in the right part of Figure 4. We find sub-blocks in the diagonal blocks, indicating that segments are similar in a small time interval but could be different when the time interval is large." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.487, + 0.667 + ], + "angle": 0, + "content": "Linear probing of behavior embedding. Linear probing, a commonly employed method to assess representations in self-supervised learn" + }, + { + "type": "table_caption", + "bbox": [ + 0.519, + 0.609, + 0.766, + 0.623 + ], + "angle": 0, + "content": "Table 1: Accuracy of linear probing" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.624, + 0.785, + 0.661 + ], + "angle": 0, + "content": "
AEContrastive AEMasked AERealGen-AE
82.5%67.1%86.2%87.8%
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.668, + 0.788, + 0.76 + ], + "angle": 0, + "content": "ing (SSL), involves training a linear classifier using the derived embeddings [27]. To train this classifier, we implemented heuristic rules to assign basic behavioral labels (acceleration, deceleration, stopping, keeping speed, left/right turn) to each agent's embedding. The outcomes, along with comparisons with baseline models, are presented in Table 1. These findings demonstrate that the embeddings generated by RealGen surpass all baseline models in terms of accuracy." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.783, + 0.652, + 0.799 + ], + "angle": 0, + "content": "5.4 Evaluation of Retrieval Augmented Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Evaluating scene-level controllability and quality of scenario generation is an open problem due to the lack of quantitative metrics. Following most previous" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.129 + ], + "angle": 0, + "content": "W. Ding et al." + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.145, + 0.71, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.354, + 0.788, + 0.397 + ], + "angle": 0, + "content": "Fig. 4: (a) Scene ID accuracy using the behavior embedding with difference distance metrics. (b) A matrix shows the Wasserstein distance between scenario segments, where each block contains the segments that belong to the same Scene ID." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.4, + 0.789, + 0.429 + ], + "angle": 0, + "content": "Table 2: Results of realism metrics. Recon.-based and Retrieval-based means using the target scenario and retrieved scenarios as input for generation, respectively." + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.434, + 0.769, + 0.557 + ], + "angle": 0, + "content": "
CategoryMethodmADEmFDESpeedHeadingSCRORR
Recon.-basedAE0.18±0.030.41±0.060.04±0.010.10±0.010.02±0.000.02±0.00
Masked AE0.16±0.010.39±0.010.04±0.010.09±0.010.03±0.000.02±0.00
Contrastive AE0.92±0.021.47±0.040.12±0.000.36±0.020.04±0.000.04±0.00
RealGen-AE0.31±0.010.53±0.010.08±0.000.15±0.010.03±0.000.02±0.00
Retrieval-basedAE-KNN14.3±0.0316.4±0.050.57±0.010.59±0.020.15±0.010.15±0.01
LCTGen4.76±0.096.24±0.080.52±0.060.57±0.030.07±0.010.07±0.01
LCTGen w/o z14.2±0.0716.7±0.092.04±0.041.42±0.000.16±0.020.13±0.04
RealGen-AE-KNN13.1±0.0614.1±0.030.46±0.010.44±0.000.12±0.010.11±0.00
RealGen1.54±0.041.21±0.030.21±0.030.21±0.010.05±0.000.04±0.00
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.787, + 0.613 + ], + "angle": 0, + "content": "work [18, 49, 64], we provide the results of the realism metrics and show the qualitative results of using RealGen for various downstream tasks. More results can be found in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Realism of generated scenario. To evaluate the realism of generated scenarios, we consider the following metrics and show the results in Table 2. We use the maximum mean discrepancy (MMD) [22] to measure the similarity in velocity and direction between the original scenarios and the generated scenarios. We also compare the mean average displacement error (mADE) and the mean final displacement error (mFDE) for the average reconstruction performance. To evaluate scene-level realism, we calculate the scene collision rate (SCR) and the off-road rate (ORR) following the metrics defined in [49]. The recon-based generation methods in Table 2 use the behavior of the target scenario as input. For this category, we compare RealGen-AE, only using the encoder and decoder modules, with three baseline methods. Due to the additional contrastive term, RealGen-AE achieves slightly worse performance than AE and Masked AE. However, RealGen is designed for retrieval-based generation, which uses retrieved scenarios rather than the target scenario as input. To fairly compare the generation performance, we design two baselines named AE-KNN and RealGen-AE-KNN," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.222, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "RealGen: Retrieval Augmented Generation for Controllable Traffic Scenarios" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.143, + 0.409, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.143, + 0.599, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.143, + 0.787, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.298, + 0.785, + 0.313 + ], + "angle": 0, + "content": "Fig. 5: Examples of tag-retrieved scenarios generated by RealGen for six different tags." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.785, + 0.414 + ], + "angle": 0, + "content": "which use KNN to find the most similar behavior embedding to the target scenario and use this embedding as input to the decoder for generation. According to the results, we find that RealGen achieves comparable performance as recombined generation and outperforms baselines, which indicates the important role of the combiner in fusing the information of the retrieved scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.487, + 0.627 + ], + "angle": 0, + "content": "Generating tag-retrieved scenarios. Now we explore the qualitative performance of using RealGen for tag-retrieved generation. Given a target behavior tag, we first obtain several template scenarios from a small dataset and use them to retrieve more scenarios from the training database, which will be used for generation in the combiner. Since there is no existing dataset with tags, we manually labeled six tags - U-Turn, Overtaking, Left Lane Change, Right Lane Change, Left Turn, Right Turn - for" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.422, + 0.787, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.577, + 0.785, + 0.619 + ], + "angle": 0, + "content": "Fig. 6: Examples of generating crash scenarios from RealGen. The shadow rectangles represent the initial positions of agents." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.784, + 0.687 + ], + "angle": 0, + "content": "the nuScenes dataset to get template 1349 scenarios. We plot generated scenarios for each tag in Figure 5, where the left part of each example shows the given initial pose and map and the right part of each example shows the generated scenario from RealGen." + }, + { + "type": "title", + "bbox": [ + 0.241, + 0.69, + 0.466, + 0.705 + ], + "angle": 0, + "content": "Safety-critical scenario gen" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.465, + 0.749 + ], + "angle": 0, + "content": "eration. Beyond the tag mentioned above, RealGen demonstrates its capacity for in-context" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "learning by generating critical and unseen crash scenarios, divergent from those of the training datasets. This is initiated by manually creating several crash scenario templates, guided by the scenarios recorded in the NHTSA Crash Report [40]. Subsequently, we generate crash scenarios using existing initial poses and maps of the dataset. Figure 6 illustrates six instances in which the shadowed rectangles denote the initial positions of agents and the red box highlights" + }, + { + "type": "table_caption", + "bbox": [ + 0.48, + 0.695, + 0.776, + 0.708 + ], + "angle": 0, + "content": "Table 3: Safety-critical scenario generation." + }, + { + "type": "table", + "bbox": [ + 0.472, + 0.709, + 0.789, + 0.746 + ], + "angle": 0, + "content": "
MethodRealgen-AE-RRealGen-RRealGen
Collision Rate (↑)0.920.830.59
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "W. Ding et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.228, + 0.145, + 0.774, + 0.16 + ], + "angle": 0, + "content": "Table 4: Results of human evaluation of controllability. (Details in Appendix ??)" + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.165, + 0.787, + 0.223 + ], + "angle": 0, + "content": "
CategoryLeft-TurnRight-TurnLeft-Lane-ChangeRight-Lane-ChangeStraight
RealGen Preferred (%)81.891.797.893.3100.0
RealGen Score (0-5)4.27±1.054.27±0.693.96±1.944.17±1.933.94±2.27
LCTGen Score (0-5)2.15±1.442.08±1.192.42±1.952.0±1.962.14±2.31
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.232, + 0.785, + 0.323 + ], + "angle": 0, + "content": "the point of collision. To quantitatively assess the performance of safety-critical scenarios, we compare RealGen with two baselines. RealGen-AE-R random samples behavior embeddings in the Realgen-AE model, and RealGen-R random retrieves behavior embeddings. According to the results in Table 3, we find that the scenarios generated by RealGen using crash scenarios as templates achieve the highest collision rate, which means that RealGen has more efficiency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.323, + 0.787, + 0.413 + ], + "angle": 0, + "content": "Human evaluation of controllability. As there is no automatic way to evaluate controllability, we follow the protocol in [49] to perform A/B testing using human evaluation. We report the ratio of our method preferred as well as the absolute score (0-5) of both our method and the baseline. In Table 4, we find that scenarios generated by RealGen are highly preferred in most categories. The absolute score of RealGen is also much higher than that of LCTGen." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.449, + 0.489 + ], + "angle": 0, + "content": "Downstream task evaluation. A direct downstream task of our method is to use the generated data to augment the training dataset of trajectory" + }, + { + "type": "table_caption", + "bbox": [ + 0.492, + 0.419, + 0.753, + 0.431 + ], + "angle": 0, + "content": "Table 5: Downstream task evaluation." + }, + { + "type": "table", + "bbox": [ + 0.461, + 0.432, + 0.785, + 0.481 + ], + "angle": 0, + "content": "
MethodOriginalRandom Aug.RealGen Aug.
mADE (↓)3.5442.9202.309
collision rate (↓)0.0490.0370.018
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.787, + 0.58 + ], + "angle": 0, + "content": "prediction models. We use Autobots [21] as a predictor and report the results trained on different datasets in Table 5. Original means using the original data in nuScenes [5], Random Aug. means augmenting the original dataset with Gaussian noise, and RealGen Aug. means augmenting the original dataset with scenarios from RealGen. We observe that the model trained with the RealGen dataset achieves the lowest mADE and collision rate." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.591, + 0.51, + 0.607 + ], + "angle": 0, + "content": "6 Conclusion and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.787, + 0.765 + ], + "angle": 0, + "content": "This paper proposes RealGen, a novel framework for traffic scenario generation that utilizes retrieval-augmented generation. Unlike previous approaches, which primarily rely on models replicating training distributions, RealGen demonstrates in-context learning abilities that synthesize scenarios by combining and modifying provided examples, enabling controlled generation. These scenarios can be automatically obtained from a retrieval system, which only requires the users to provide a few template scenarios as examples. We comprehensively evaluated the similarity autoencoder model for retrieval and the combiner model for generation. The findings indicate that RealGen achieves low reconstruction error and high generation quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "The primary limitation of our current method is that the behavior encoder focuses solely on agent trajectories, neglecting the intricate interactions between agents and lane maps in complex behaviors. Advancing the feature representation within the behavior encoder could significantly broaden RealGen's capacity to generate diverse and controllable traffic scenarios." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.222, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RealGen: Retrieval Augmented Generation for Controllable Traffic Scenarios" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.786, + 0.22 + ], + "angle": 0, + "content": "Wenhao Ding contributed to this paper while being an intern at NVIDIA Research. Ding Zhao was partially supported by the National Science Foundation under grant CNS-2047454." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.24, + 0.323, + 0.256 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.267, + 0.785, + 0.295 + ], + "angle": 0, + "content": "1. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.296, + 0.785, + 0.336 + ], + "angle": 0, + "content": "2. Blattmann, A., Rombach, R., Oktay, K., Müller, J., Ommer, B.: Retrievalaugmented diffusion models. Advances in Neural Information Processing Systems 35, 15309-15324 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.337, + 0.785, + 0.391 + ], + "angle": 0, + "content": "3. Borgeaud, S., Mensch, A., Hoffmann, J., Cai, T., Rutherford, E., Millican, K., Van Den Driessche, G.B., Lespiau, J.B., Damoc, B., Clark, A., et al.: Improving language models by retrieving from trillions of tokens. In: International conference on machine learning. pp. 2206-2240. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.392, + 0.785, + 0.433 + ], + "angle": 0, + "content": "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.785, + 0.487 + ], + "angle": 0, + "content": "5. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.488, + 0.785, + 0.555 + ], + "angle": 0, + "content": "6. Cai, D., Wang, Y., Bi, W., Tu, Z., Liu, X., Shi, S.: Retrieval-guided dialogue response generation via a matching-to-generation framework. In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). pp. 1866-1875 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.556, + 0.785, + 0.597 + ], + "angle": 0, + "content": "7. Cao, Y., Xiao, C., Anandkumar, A., Xu, D., Pavone, M.: Advdo: Realistic adversarial attacks for trajectory prediction. In: European Conference on Computer Vision. pp. 36-52. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.597, + 0.785, + 0.623 + ], + "angle": 0, + "content": "8. Chen, D., Fisch, A., Weston, J., Bordes, A.: Reading wikipedia to answer open-domain questions. arXiv preprint arXiv:1704.00051 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.624, + 0.785, + 0.663 + ], + "angle": 0, + "content": "9. Chen, H., Wang, J., Shao, K., Liu, F., Hao, J., Guan, C., Chen, G., Heng, P.A.: Traj-mae: Masked autoencoders for trajectory prediction. arXiv preprint arXiv:2303.06697 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.664, + 0.785, + 0.704 + ], + "angle": 0, + "content": "10. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.705, + 0.785, + 0.732 + ], + "angle": 0, + "content": "11. Chen, W., Hu, H., Sahara, C., Cohen, W.W.: Re-imagen: Retrieval-augmented text-to-image generator. arXiv preprint arXiv:2209.14491 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.733, + 0.785, + 0.772 + ], + "angle": 0, + "content": "12. Cheng, J., Mei, X., Liu, M.: Forecast-mae: Self-supervised pre-training for motion forecasting with masked autoencoders. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8679-8689 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.773, + 0.785, + 0.799 + ], + "angle": 0, + "content": "13. Cuturi, M.: Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems 26 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "14. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.267, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "W. Ding et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "15. Ding, W., Chen, B., Li, B., Eun, K.J., Zhao, D.: Multimodal safety-critical scenarios generation for decision-making algorithms evaluation. IEEE Robotics and Automation Letters 6(2), 1551-1558 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "16. Ding, W., Lin, H., Li, B., Eun, K.J., Zhao, D.: Semantically adversarial driving scenario generation with explicit knowledge integration. arXiv preprint arXiv:2106.04066 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.787, + 0.274 + ], + "angle": 0, + "content": "17. Ding, W., Xu, M., Zhao, D.: Learning to collide: An adaptive safety-critical scenarios generating method. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.787, + 0.317 + ], + "angle": 0, + "content": "18. Feng, L., Li, Q., Peng, Z., Tan, S., Zhou, B.: Trafficgen: Learning to generate diverse and realistic traffic scenarios. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 3567-3575. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.787, + 0.373 + ], + "angle": 0, + "content": "19. Feydy, J., Séjourne, T., Vialard, F.X., Amari, S.i., Trouve, A., Peyré, G.: Interpolating between optimal transport and mmd using sinkhorn divergences. In: The 22nd International Conference on Artificial Intelligence and Statistics. pp. 2681-2690 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.374, + 0.787, + 0.402 + ], + "angle": 0, + "content": "20. Gidaris, S., Singh, P., Komodakis, N.: Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.403, + 0.787, + 0.444 + ], + "angle": 0, + "content": "21. Girgis, R., Golemo, F., Codevilla, F., Weiss, M., D'Souza, J.A., Kahou, S.E., Heide, F., Pal, C.: Latent variable sequential set transformers for joint multi-agent motion prediction. arXiv preprint arXiv:2104.00563 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.445, + 0.787, + 0.472 + ], + "angle": 0, + "content": "22. Gretton, A., Borgwardt, K.M., Rasch, M.J., Scholkopf, B., Smola, A.: A kernel two-sample test. The Journal of Machine Learning Research 13(1), 723-773 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.787, + 0.515 + ], + "angle": 0, + "content": "23. Gulino, C., Fu, J., Luo, W., Tucker, G., Bronstein, E., Lu, Y., Harb, J., Pan, X., Wang, Y., Chen, X., et al.: Waymax: An accelerated, data-driven simulator for large-scale autonomous driving research. arXiv preprint arXiv:2310.08710 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.787, + 0.572 + ], + "angle": 0, + "content": "24. Gutmann, M., Hyvarinen, A.: Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In: Proceedings of the thirteenth international conference on artificial intelligence and statistics. pp. 297-304. JMLR Workshop and Conference Proceedings (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.787, + 0.613 + ], + "angle": 0, + "content": "25. Guu, K., Lee, K., Tung, Z., Pasupat, P., Chang, M.: Retrieval augmented language model pre-training. In: International conference on machine learning. pp. 3929-3938. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.614, + 0.787, + 0.655 + ], + "angle": 0, + "content": "26. Hanselmann, N., Renz, K., Chitta, K., Bhattacharyya, A., Geiger, A.: King: Generating safety-critical driving scenarios for robust imitation via kinematics gradients. In: European Conference on Computer Vision. pp. 335-352. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.787, + 0.698 + ], + "angle": 0, + "content": "27. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.787, + 0.741 + ], + "angle": 0, + "content": "28. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.787, + 0.769 + ], + "angle": 0, + "content": "29. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.77, + 0.787, + 0.798 + ], + "angle": 0, + "content": "30. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural computation 9(8), 1735-1780 (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "31. Hospedales, T., Antoniou, A., Micaelli, P., Storkey, A.: Meta-learning in neural networks: A survey. IEEE transactions on pattern analysis and machine intelligence 44(9), 5149-5169 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.223, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RealGen: Retrieval Augmented Generation for Controllable Traffic Scenarios" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "32. Hu, A., Corrado, G., Griffiths, N., Murez, Z., Gurau, C., Yeo, H., Kendall, A., Cipolla, R., Shotton, J.: Model-based imitation learning for urban driving. Advances in Neural Information Processing Systems 35, 20703-20716 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.217 + ], + "angle": 0, + "content": "33. Ibrihich, S., Oussous, A., Ibrihich, O., Esghir, M.: A review on recent research in information retrieval. Procedia Computer Science 201, 777-782 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.217, + 0.785, + 0.243 + ], + "angle": 0, + "content": "34. Ivanovic, B., Song, G., Gilitschenski, I., Pavone, M.: trajdata: A unified interface to multiple human trajectory datasets. arXiv preprint arXiv:2307.13924 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.785, + 0.284 + ], + "angle": 0, + "content": "35. Kim, J., Choi, S., Amplayo, R.K., Hwang, S.w.: Retrieval-augmented controllable review generation. In: Proceedings of the 28th International Conference on Computational Linguistics. pp. 2284-2295 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.785, + 0.311 + ], + "angle": 0, + "content": "36. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.785, + 0.365 + ], + "angle": 0, + "content": "37. Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Kuttler, H., Lewis, M., Yih, W.t., Rocktäschel, T., et al.: Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems 33, 9459-9474 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.365, + 0.785, + 0.407 + ], + "angle": 0, + "content": "38. Li, Q., Peng, Z., Feng, L., Duan, C., Mo, W., Zhou, B., et al.: Scenarionet: Open-source platform for large-scale traffic scenario simulation and modeling. arXiv preprint arXiv:2306.12241 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.407, + 0.785, + 0.447 + ], + "angle": 0, + "content": "39. Liu, Q., Yogatama, D., Blunsom, P.: Relational memory-augmented language models. Transactions of the Association for Computational Linguistics 10, 555-572 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.753, + 0.46 + ], + "angle": 0, + "content": "40. NHTSA: Nhtsa crash viewer (2023), https://crashviewer.nhtsa.dot.gov/" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.785, + 0.5 + ], + "angle": 0, + "content": "41. Noroozi, M., Favaro, P.: Unsupervised learning of visual representations by solving jigsaw puzzles. In: European conference on computer vision. pp. 69-84. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.785, + 0.528 + ], + "angle": 0, + "content": "42. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.785, + 0.568 + ], + "angle": 0, + "content": "43. Roshdi, A., Roohparvar, A.: Information retrieval techniques and applications. International Journal of Computer Networks and Communications Security 3(9), 373-377 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.568, + 0.785, + 0.595 + ], + "angle": 0, + "content": "44. Schmidhuber, J.: Deep learning in neural networks: An overview. Neural networks 61, 85-117 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.595, + 0.785, + 0.636 + ], + "angle": 0, + "content": "45. Shi, S., Jiang, L., Dai, D., Schiele, B.: Motion transformer with global intention localization and local movement refinement. Advances in Neural Information Processing Systems 35, 6531-6543 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.636, + 0.785, + 0.663 + ], + "angle": 0, + "content": "46. Shurrab, S., Duwairi, R.: Self-supervised learning methods and applications in medical imaging analysis: A survey. PeerJ Computer Science 8, e1045 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.663, + 0.785, + 0.703 + ], + "angle": 0, + "content": "47. Suo, S., Regalado, S., Casas, S., Urtasun, R.: Trafficsim: Learning to simulate realistic multi-agent behaviors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10400-10409 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.703, + 0.785, + 0.758 + ], + "angle": 0, + "content": "48. Suo, S., Wong, K., Xu, J., Tu, J., Cui, A., Casas, S., Urtasun, R.: Mixsim: A hierarchical framework for mixed reality traffic simulation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9622-9631 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "49. Tan, S., Ivanovic, B., Weng, X., Pavone, M., Kraehenbuehl, P.: Language conditioned traffic generation. In: Conference on Robot Learning. pp. 2714-2752. PMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "50. Tan, S., Wong, K., Wang, S., Manivasagam, S., Ren, M., Urtasun, R.: Scenegen: Learning to generate realistic traffic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 892-901 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "W. Ding et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "51. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.751, + 0.203 + ], + "angle": 0, + "content": "52. Villani, C., et al.: Optimal transport: old and new, vol. 338. Springer (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.786, + 0.245 + ], + "angle": 0, + "content": "53. Vincent, P., Larochelle, H., Bengio, Y., Manzagol, P.A.: Extracting and composing robust features with denoising autoencoders. In: Proceedings of the 25th international conference on Machine learning. pp. 1096-1103 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.786, + 0.3 + ], + "angle": 0, + "content": "54. Wang, J., Pun, A., Tu, J., Manivasagam, S., Sadat, A., Casas, S., Ren, M., Urtasun, R.: Advsim: Generating safety-critical scenarios for self-driving vehicles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9909-9918 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.786, + 0.328 + ], + "angle": 0, + "content": "55. Wang, Z., Nie, W., Qiao, Z., Xiao, C., Baraniuk, R., Anandkumar, A.: Retrieval-based controllable molecule generation. arXiv preprint arXiv:2208.11126 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.329, + 0.786, + 0.37 + ], + "angle": 0, + "content": "56. Wu, P., Majumdar, A., Stone, K., Lin, Y., Mordatch, I., Abbeel, P., Rajeswaran, A.: Masked trajectory models for prediction, representation, and control. arXiv preprint arXiv:2305.02968 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.786, + 0.411 + ], + "angle": 0, + "content": "57. Xu, P., Patwary, M., Shoeybi, M., Puri, R., Fung, P., Anandkumar, A., Catanzaro, B.: Megatron-cntrl: Controllable story generation with external knowledge using large-scale language models. arXiv preprint arXiv:2010.00840 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.786, + 0.439 + ], + "angle": 0, + "content": "58. Yang, Y., Zhang, Q., Gilles, T., Batool, N., Folkesson, J.: Rmp: A random mask pretrain framework for motion prediction. arXiv preprint arXiv:2309.08989 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.44, + 0.786, + 0.481 + ], + "angle": 0, + "content": "59. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of lmms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.786, + 0.522 + ], + "angle": 0, + "content": "60. Zhang, C., Zhang, C., Song, J., Yi, J.S.K., Zhang, K., Kweon, I.S.: A survey on masked autoencoder for self-supervised learning in vision and beyond. arXiv preprint arXiv:2208.00173 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.786, + 0.55 + ], + "angle": 0, + "content": "61. Zhang, C., Tu, J., Zhang, L., Wong, K., Suo, S., Urtasun, R.: Learning realistic traffic agents in closed-loop. In: 7th Annual Conference on Robot Learning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.551, + 0.786, + 0.591 + ], + "angle": 0, + "content": "62. Zhang, X., Bosselut, A., Yasunaga, M., Ren, H., Liang, P., Manning, C.D., Leskovec, J.: Greaselm: Graph reasoning enhanced language models. In: International conference on learning representations (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.786, + 0.632 + ], + "angle": 0, + "content": "63. Zhong, Z., Rempe, D., Chen, Y., Ivanovic, B., Cao, Y., Xu, D., Pavone, M., Ray, B.: Language-guided traffic simulation via scene-level diffusion. arXiv preprint arXiv:2306.06344 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.786, + 0.688 + ], + "angle": 0, + "content": "64. Zhong, Z., Rempe, D., Xu, D., Chen, Y., Veer, S., Che, T., Ray, B., Pavone, M.: Guided conditional diffusion for controllable traffic simulation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 3560-3566. IEEE (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.786, + 0.688 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/RealGen_ Retrieval Augmented Generation for Controllable Traffic Scenarios/4b2a9c03-0a68-49b3-9375-225cd793a8b3_origin.pdf b/2024/RealGen_ Retrieval Augmented Generation for Controllable Traffic Scenarios/4b2a9c03-0a68-49b3-9375-225cd793a8b3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5f7616ad4744a9b3170be1c04ecbefb0efe4b664 --- /dev/null +++ b/2024/RealGen_ Retrieval Augmented Generation for Controllable Traffic Scenarios/4b2a9c03-0a68-49b3-9375-225cd793a8b3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0567c38f71f37eeadc6294d830926f979489508471b3e64142ede11a58f1eaec +size 4849639 diff --git a/2024/RealGen_ Retrieval Augmented Generation for Controllable Traffic Scenarios/full.md b/2024/RealGen_ Retrieval Augmented Generation for Controllable Traffic Scenarios/full.md new file mode 100644 index 0000000000000000000000000000000000000000..58e2f94ec7ef115e42ebeebf7e0ab1a8b9747949 --- /dev/null +++ b/2024/RealGen_ Retrieval Augmented Generation for Controllable Traffic Scenarios/full.md @@ -0,0 +1,338 @@ +# RealGen: Retrieval Augmented Generation for Controllable Traffic Scenarios + +Wenhao Ding\*1,2, Yulong Cao\*2, Ding Zhao\*1, Chaowei Xiao\*2,3, and Marco Pavone\*2,4 + +1 Carnegie Mellon University + +2 NVIDIA Research + +3 UW-Madison + +4 Stanford University + +{wenhaod, yulongc}@nvidia.com + +Abstract. Simulation plays a crucial role in the development of autonomous vehicles (AVs) due to the potential risks associated with real-world testing. Although significant progress has been made in the visual aspects of simulators, generating complex behavior among agents remains a formidable challenge. It is not only imperative to ensure realism in the scenarios generated but also essential to incorporate preferences and conditions to facilitate controllable generation for AV training and evaluation. Traditional methods, which rely mainly on memorizing the distribution of training datasets, often fail to generate unseen scenarios. Inspired by the success of retrieval augmented generation in large language models, we present RealGen, a novel retrieval-based in-context learning framework for traffic scenario generation. RealGen synthesizes new scenarios by combining behaviors from multiple retrieved examples in a gradient-free way, which may originate from templates or tagged scenarios. This in-context learning framework endows versatile generative capabilities, including the ability to edit scenarios, compose various behaviors, and produce critical scenarios. Evaluations show that RealGen offers considerable flexibility and controllability, marking a new direction in the field of controllable traffic scenario generation. Check our project website for more information: https://realgen.github.io. + +# 1 Introduction + +Simulation is indispensable in the development of autonomous vehicles (AVs), primarily due to the considerable risks associated with training and evaluating these systems in real-world conditions. The biggest challenge in simulations lies in achieving realistic driving scenarios, as this realism influences the discrepancy between AV performance in simulated and actual environments. Although advancements in high-quality graphical engines have significantly enhanced the perception quality of simulators, the realism of agent behavior remains constrained because of the complicated interactions among naturalistic agents. To counteract this issue, data-driven simulation has emerged as a promising approach in the realm of autonomous driving, which leverages real-world scenario datasets to accurately generate the behaviors of agents. + +With the rapid achievement of deep generative models [29] and imitation learning algorithms [32], current data-driven simulations [23, 38] can generate scenarios that closely mimic human driver behavior. However, the effectiveness of these simulations in accelerating the development of AV is limited. This limitation stems from the need for scenarios that meet specific conditions tailored for targeted training and evaluation. Achieving such controllability in simulations is challenging due to the complex nature of driving scenarios, which involve intricate interactions, diverse road layouts, and varying traffic regulations. + +In pursuit of controllability, existing work applies additional guidance, typically in the form of constraint functions [16, 64] or languages [49, 63], to pre-trained scenario generative models. Regularization of the generation process through these tools is straightforward and effective, yet it encounters two principal challenges. First, the training of scenario generative models typically utilizes naturalistic datasets, which might not encompass the specific scenarios desired as per the control signals. Even if such + +![](images/c2692d7e4ef746f27023fb6009f1edbf303e3631ea23102ce352a5bae027f12e.jpg) +(a) Conventional method + +![](images/17d61e5d3fabb3e7ce926a63f363956d5811a6f3db4fdf611e42045720b3588f.jpg) +(b) Our method +Fig. 1: (a) Conventional methods make the model memorize the data distribution for generating. (b) In contrast, our method employs a retriever to query datasets (including external data obtained after training) and uses a generative model to generate scenarios by integrating the information from the retrieved scenarios. + +scenarios exist within the dataset, they are often omitted because of the rarity of long-tail data. The second challenge is that the representation of the guidance to the generative model may not be sufficiently expressive to accurately depict complex scenarios, such as specifying intricate interactions among multiple vehicles, using language. These limitations underscore the need for a more sophisticated and nuanced framework for controllable scenario generation. + +Retrieval Augmented Generation (RAG) [8, 25], which enhances the generative process by querying related information from external databases, represents great potential in the domain of large language models [37]. In contrast to conventional models that memorize all knowledge within their parameters, as shown in Figure 1(a), RAG models, shown in Figure 1(b), learn to generate comprehensive outputs by retrieving pertinent knowledge from a database, based on the input provided. A notable aspect of RAG is the ability of the database to undergo updates even after the model has been trained, allowing continuous improvement and adaptation. This flexible framework offers possibilities for controllable scenario generation by using appropriate template scenarios as input and facilitating the generation that is not only realistic but also aligned with specific training and evaluation requirements. + +In this work, we present RealGen, a retrieval augmented generation framework for generating traffic scenarios. This framework, as shown in Figure 2, begins with the training of an encoder through contrastive self-supervised learning [42] to allow the retrieval process to query similar scenarios in a latent embedding space. Leveraging this latent representation, we subsequently train a generative model that combines retrieved scenarios to create novel scenarios. The key contributions of this paper are summarized below. + +- We develop a novel contrastive autoencoder model to extract scenario embeddings as latent representations, which can be used for a wide range of downstream tasks. +- We propose the first retrieval augmented generation framework using the latent representation tailored for controllable driving scenario generation. +- We validate our framework through qualitative and quantitative metrics, demonstrating strong flexibility and controllability of generated scenarios. + +# 2 Related Work + +# 2.1 Traffic Modeling and Scenario Generation + +Research on traffic modeling using generative models has received considerable attention in recent works. Notably, ScenarioNet [38] and Waymax [23] employ large-scale data to train imitation learning models, facilitating the generation of realistic multi-agent scenarios in simulations. SceneGen [50] utilizes the Long Short Term Memory [30] module to autoregressively generate the trajectories for vehicles and pedestrians based on provided maps. Additionally, TrafficSim [47] learns the multi-agent behaviors from real-world data, and TrafficGen [18] proposes a transformer-based autoencoder architecture to accurately model complex interactions involving multiple agents. MixSim [48] builds a reactive digital twin and finds safety-critical scenarios with black-box optimization. RTR [61] models reactive behaviors of vehicles using the combination of reinforcement learning and imitation learning. + +Numerous prior studies have employed adversarial generation to synthesize rare yet critical scenarios. L2C [17] utilizes a reinforcement learning framework, where the reward for scenario generation is based on the collision rate. To enhance the realism in purely adversarial generation methods, MMG [15] incorporates the data distribution as regularization. Further developments such as AdvSim [54], AdvDO [7], and KING [26] integrate vehicle dynamics to directly optimize the trajectory to find critical scenarios. + +The proliferation of large language models (LLMs) has led to recent approaches in generating traffic scenarios with language as conditions to follow instructions from humans. CTG++ [63] replaces the gradient guidance process of CTG [64] with cost functions generated by an LLM. Leveraging the power of GPT-4 [59], this framework can generate diverse motion data with language conditions. Additionally, LCTGen [49] also harnesses the strengths of GPT-4 to generate a heuristic intermediate representation of scenarios from language + +inputs. Subsequently, a generative model pre-trained on open-source datasets generates various scenarios with this representation as inputs. + +# 2.2 Retrieval-augmented Generation + +Information retrieval (IR) [33] is the procedure of representing and searching a collection of data with the goal of extracting knowledge to satisfy the query of users. Predominantly utilized in search engines and digital libraries [43], IR primarily deals with the information stored in textual formats. Recently, IR has extended across diverse sectors, notably enhancing the quality of outputs in areas such as language modeling [3, 39], question answering [25, 62], image creation [2, 11], and the generation of molecular [55]. + +One intuitive usage of the retrieved information is to enhance input data through several methods, such as merging the original data and retrieved data [37], employing attention mechanism [3], or extracting skeleton [6]. The rationale behind retrieval systems stems from the impracticality of encoding all knowledge within model parameters, especially considering the dynamic nature of knowledge that evolves with human activities. Consequently, the capability to access external knowledge databases can significantly improve the precision and quality of generated responses, as evidenced in applications involving LLMs [3, 37]. + +Another usage of retrieved data, particularly pertinent to this study, involves the controllable generation by integrating desired features retrieved from the dataset. In [35], the authors generate product reviews with controllable information about the user, product, and rating. Similarly, the process of story creation can be viewed as a blend of external story databases with selected text fragments [57] Furthermore, [55] explores the controllable generation of molecules to meet various constraints, utilizing a database comprising simple molecules that typically meet only one of these constraints. + +# 2.3 Self-supervised Feature Learning + +The necessity for manual labeling in supervised learning often results in human biases, extraneous noises, and labor-intensive efforts. Therefore, the paradigm of self-supervised learning (SSL) [46] is garnering heightened interest, particularly for its promising applications in language modeling and image interpretation. SSL algorithms usually learn implicit representations from extensive pools of unlabeled data without relying on human annotations. Generally, this line of research can be bifurcated into two categories [60]: generative SSL and discriminative SSL. + +In the domain of generative SSL, models employ an autoencoder to convert input data into a latent representation, followed by a reconstruction process. As an illustrative example, Denoising Autoencoder [53], an early example of this approach, reconstructs images from their noisy version to extract representative features. Another technique in this field is masked modeling, extensively utilized in language models such as GPT-3 [4] and BERT [14], which predicts the masked token to gain a semantic understanding of the text. More recently, the + +Masked Autoencoder [27] has emerged as a powerful framework, demonstrating its efficacy across a variety of downstream tasks. + +Discriminative SSL typically optimizes a discriminative loss to learn representations from sets of anchor, positive, and negative samples. Without ground-truth labels, these pairs are often constructed through solving jigsaw puzzles [41] or making geometry-based prediction [20]. A prominent example of discriminative SSL is contrastive learning, which brings samples from the same class closer while distancing those from different classes. Representative works such as MoCo [28] and SimCLR [10] learn embeddings by capturing invariant features between original data and its augmented variants. Additionally, InfoNCE [42], a method grounded in noise contrastive estimation [24], is widely used due to its effectiveness. + +# 3 A Latent Representation of Scenario + +A crucial element within the retrieval system is the selection metric for data retrieval, which is typically implemented through a distance function between the query sample and the candidate samples in the database. Unlike text, which can be converted into word embeddings, traffic scenarios encompass sequential behaviors and intricate interactions among entities, complicating the establishment of a similarity metric for these scenarios. Consequently, in this section, we introduce a scenario autoencoder to extract latent representations that facilitate the assessment of similarity between various traffic scenarios, a vital component in our RAG framework. + +# 3.1 Scenario Definition + +Each scenario is characterized by the trajectories $\tau \in \mathbb{R}^{M\times T\times 5}$ , encompassing $M$ agents over a maximum of $T$ time steps. The trajectory of each agent is composed of parameters $[x,y,v,c,s]$ , signifying position $x$ , position $y$ , velocity $v$ , cosine of the heading $c$ , and sine of the heading $s$ . The initial state of these entities is denoted as $\tau_0\in \mathbb{R}^{M\times 5}$ . Furthermore, the map is encapsulated by $m\in \mathbb{R}^{S\times 4}$ , composed of $S$ lane segments. The attributes of these points $[x_s,y_s,x_e,y_e]$ correspond to the starting and ending positions of each segment. + +# 3.2 Scenario Autoencoder + +Autoencoders [44] are used to learn compressed latent representations of high-dimensional data, where an encoder projects the data into latent code and a decoder reconstructs the code in the data space. Given that traffic scenarios encompass spatial and temporal dynamics from multiple agents, we design a hierarchical encoder structure alternating between spatial and temporal layers based on transformer architecture [51]. As depicted on the left of Figure 2, our design incorporates three encoders for behavior, map, and initial position. + +![](images/f11185e87a54575c4943521e59e96c8a8d4a3217b9d0d8eedc46c35bc908207b.jpg) +Fig. 2: Left: the training pipeline for the encoders and decoder aimed at learning latent embeddings of scenarios. A contrastive loss is applied to behavior embeddings to ensure invariance to absolute positions. Middle: the training pipeline for the combiner with frozen encoder and decoder parameters. We use K-Nearest Neighbors (KNN) to retrieve scenarios similar to a template scenario in the dataset and use the retrieved behaviors to reconstruct the template scenario. Right: the generation pipeline with a retriever and a generator. + +First, we design a behavior encoder $E_{b}$ with a spatial-temporal transformer structure that comprises $L_{e}$ temporal transformer encoders $(\mathrm{Encoder}_{t}^{i})$ and $L_{e}$ spatial transformer encoders $(\mathrm{Encoder}_{s}^{i})$ , where $i \in \{1, \dots, L_{e}\}$ . The trajectory $\tau$ is first transformed into a latent embedding $z_{b}$ , where $H$ denotes the hidden dimension, through a Multiple Linear Perception (MLP) module. Subsequently, an alternating encoding procedure extracts spatial and temporal features from $z_{b}$ . To preserve temporal information, we add a sinusoidal positional embedding (PE) to the embedding before employing the temporal transformer. To retain distinct agent information and further compress the behavior feature, we calculate the mean of the latent embedding across the temporal dimension, resulting in the final behavior embedding $z_{b} \in \mathbb{R}^{M \times H}$ . Besides the behavior encoder, we also process the map information represented by lane vectors with the attention mechanism [51]. Similarly to the first step in $E_{b}$ , we project $m$ to the latent space with an MLP module. Then we apply a multi-head attention module (MHA) with layer normalization (LN) [1] to acquire the map embedding $z_{m} \in \mathbb{R}^{S \times H}$ with a learnable query embedding $q_{m}$ . + +In the decoder, a spatial-temporal transformer architecture, similar to the behavior encoder, is designed to obtain the behavior embedding $z_{b}$ through $L_{d}$ times encoding. Given that $z_{b}$ lacks the temporal dimension, we first replicate it $T$ times and then add a PE before entering it into the temporal transformer encoder. Throughout the decoding process, the map embedding $z_{m}$ is injected by a cross-attention mechanism: $z_{b} \gets z_{b} + \mathrm{MHA}(z_{b}, z_{m}, z_{m})$ . MHA( $Q, K, V$ ) is the multi-head attention with $Q, K, V$ representing query, key, and value: + +$$ +\mathrm {M H A} (Q, K, V) = \text {C o n c a t e n a t e} \left(h _ {1}, \dots , h _ {i}\right) W ^ {O}, +$$ + +$$ +h _ {i} = \mathrm {A t t e n t i o n} (Q W _ {i} ^ {Q}, K W _ {i} ^ {K}, V W _ {i} ^ {V}), +$$ + +Algorithm 1: Details of Encoder and Decoder +1 Behavior Encoder $E_{b}$ $(\tau)$ : 14 Initial Pose Encoder $E_{i}$ $(\tau_{0})$ : +2 $z_{b} \gets \mathrm{MLP}(\tau) // \text{projection}$ 15 $z_{i} \gets \mathrm{MLP}(\tau_{0}) // \text{projection}$ +3 for $i$ in $[1, \dots, L_{e}]$ do 16 return initial pose embedding $z_{i}$ +4 $z_{b} \gets \mathrm{Encoder}_{t}^{i}(\mathrm{PE} + z_{b})$ 17 Decoder $D(z_{i}, z_{b}, z_{m})$ : +5 $z_{b} \gets \mathrm{Encoder}_{s}^{i}(z_{b})$ 18 $z_{r} \gets z_{b} + \mathrm{MHA}(z_{b}, z_{i}, z_{i})$ +6 $z_{b} \gets \text{mean}(z_{b})$ 19 for $i$ in $[1, \dots, L_{d}]$ do +7 return behavior embedding $z_{b}$ 20 $z_{r} \gets \mathrm{Encoder}_{t}^{i}(\mathrm{PE} + z_{r})$ +8 Map Encoder $E_{m}$ : 21 +9 Initialize a learnable query $q_{m}$ 22 +10 $z_{m} \gets \mathrm{MLP}(m) // \text{projection}$ 23 $\hat{\tau} \gets \mathrm{MLP}(z_{r}) // \text{projection}$ +11 $z_{m} \gets \mathrm{LN}(\mathrm{MHA}(q_{m}, z_{m}, z_{m}))$ 24 return reconstructed trajectory $\hat{\tau}$ +12 $z_{m} \gets \mathrm{LN}(z_{m} + \mathrm{MLP}(z_{m}))$ +13 return map embedding $z_{m}$ + +where $W^{O}$ , $W_{i}^{Q}$ , $W_{i}^{K}$ , and $W_{i}^{V}$ are learnable parameters. The last component of the decoder is an MLP that projects the hidden embedding back into the data domain, resulting in the reconstructed trajectory $\hat{\tau}$ . To train the encoders and decoder, the mean square error is used as the reconstruction loss, expressed as $\mathcal{L}_r = |\hat{\tau} -\tau |_2$ . Comprehensive descriptions of the encoding and decoding processes are provided in Algorithm 1. + +It is important to note that, while the current design of the autoencoder is sufficient for acquiring compressed representations of scenarios, these representations are not invariant to the absolute coordinates and the order of the agents. This limitation can lead to substantial embedding distances between similar scenarios. To address these issues, two enhancements are proposed and detailed in the following section. + +# 3.3 Invariant feature with contrastive loss + +To enhance the representation of scenario similarity in the behavior embedding $z_{b}$ , we employ contrastive learning to acquire invariant features. Specifically, we integrate InfoNCE [42] as an additional loss function $\mathcal{L}_c$ , which optimizes categorical cross-entropy to distinguish a positive sample from a batch of negative samples. In practice, for a query embedding $z_{b}$ , a positive sample $z_{b}^{+}$ is generated by applying random rotation and translation to the original scenario $\tau, m$ . Meanwhile, negative samples $Z_{b}^{-}$ are selected from the remaining samples in the same batch. In conventional InfoNCE, to bring $z_{b}$ and $z_{b}^{+}$ closer together, the inner product is calculated between the embedding $z_{b}$ and the set $z_{b}^{+}, Z_{b}^{-}$ , employing cross-entropy loss to identify the index of the positive sample. + +To ensure that the ordering of the agents does not affect the outcomes, $z_{b} \in \mathbb{R}^{M \times H}$ should exhibit permutation invariance across the dimension $M$ . Otherwise, simply stacking them into a one-dimensional vector could erroneously represent distinct scenarios as significantly different. To establish permutation + +invariance, we adopt the Wasserstein distance $W_{2}$ [52] as a more suitable metric than cosine distance for assessing the similarity. Taking $z_{b}$ as a distribution representing $M$ individual behaviors, the intuition behind this choice is that the Wasserstein distance identifies the minimal adjustments necessary for behaviors in one scenario to reflect those in another. With such a powerful tool, we have the following contrastive loss: + +$$ +\mathcal {L} _ {c} = - \sum_ {z _ {b}} \log \frac {\exp \left[ - W _ {2} \left(z _ {b} , z _ {b} ^ {+}\right) \right]}{\sum_ {z ^ {\prime} \in \left\{z _ {b} ^ {+} , Z _ {b} ^ {-} \right\}} \exp \left[ - W _ {2} \left(z _ {b} , z ^ {\prime}\right) \right]}. \tag {2} +$$ + +The implementation of this loss results in the behavior embedding $z_{b}$ lacking absolute coordinate data, making the decoder incapable of accurately reconstructing the precise trajectory. To handle this problem, we add the initial poses $\tau_0$ of all agents as a supplementary input for the decoder. We encode these poses into $z_{i}$ with an MLP encoder $\mathrm{Enc}_i$ , and then integrate it into the decoder with an MHA module $z_{r} \gets z_{b} + \mathrm{MHA}(z_{b},z_{i},z_{i})$ . Additional specifics regarding the initial pose encoder are detailed in Algorithm 1. In the training stage, the objective is to minimize a combined loss function $\mathcal{L} = \mathcal{L}_r + \lambda \mathcal{L}_c$ , where $\lambda$ serves as a weighting factor for contrastive loss and is uniformly set at 0.1 for all experiments. + +# 4 Retrieval Augmented Scenario Generation + +The autoencoder structure introduced in Section 3 builds a one-to-one generation framework, which still needs additional modules to support the retrieval augmented generation, which is a many-to-one framework. In this section, we introduce a module called Combiner, which takes input behavior embeddings from multiple scenarios and outputs the combined embedding. To train this module, we designed a KNN-based training pipeline that forces the model to learn to combine and edit existing scenarios. + +# 4.1 The Training of Combiner + +Unlike the trajectory prediction task [45] which has ground truth as the optimization target, the objective in training the combiner is to learn the alignment of behaviors in the retrieved scenarios with the initial pose and the specified map. A well-trained combiner should be able to compose behaviors from these retrieved scenarios, thereby generating new scenarios that have a resemblance to all the retrieved scenarios. This type of objective aligns with the concept of meta-learning, or 'learning to learn', as described in [31]. + +Within our training framework, for a given query scenario $\tau, m$ in the dataset, we initially utilize the behavior encoder to obtain $z_{b}$ and then use KNN to identify $K$ similar behavior embeddings from the database, denoted $z_{ret} = [z_{b,1}, \dots, z_{b,K}]$ . Building on this, we propose a model comprising two MHA modules as the combiner: + +$$ +z _ {r a g} \leftarrow z _ {i} + \mathrm {M H A} \left(z _ {i}, z _ {r e t}, z _ {r e t}\right), \tag {3} +$$ + +$$ +z _ {r a g} \leftarrow z _ {r a g} + \mathrm {M H A} (z _ {r a g}, z _ {m}, z _ {m}), +$$ + +where $z_{i} \gets E_{i}(x_{0})$ and $z_{m} \gets E_{m}(m)$ are the initial pose embedding and the map embedding of the query scenario. Gradients are stopped for $z_{i}$ and $z_{m}$ . Assuming that the $K$ nearest scenarios adequately represent the query scenario, it should be feasible to reconstruct the behavior of the query scenario, $z_{b}$ , using the retrieved scenario, $z_{rag}$ . Consequently, we employ the following loss function as our training goal: + +$$ +\mathcal {L} _ {r a g} = \left\| D \left(z _ {i}, z _ {r a g}, z _ {m}\right) - \tau \right\| _ {2}, \tag {4} +$$ + +wherein the parameters of the decoder $D\phi$ remain fixed during training. Essentially, this training method is designed to learn an "inverse" operation of the KNN, aiming to reconstruct a query scenario that closely resembles all $K$ retrieved scenarios in the database. + +# 4.2 The generation pipeline + +The pipeline for scenario generation, as delineated in Figure 2, is categorized into retriever and generator components. The generation process is expedited by pre-processing all scenarios in the database using a behavior encoder, which yields behavior embeddings that facilitate efficient similarity computation. + +The retriever enhances the versatility of generation by dividing the process into two stages. In the initial stage, users can choose a set of template scenarios that depict specific conditions. These may include manually annotated scenarios with tags denoting actions like left or right turns, enabling the generation of additional scenarios under similar tags or even a combination thereof. In addition, templates can include critical and interesting scenarios collected from real-world data. Solely relying on these templates for generation could be limited, which is mitigated by incorporating a secondary phase, which employs a KNN approach to fetch high-quality scenarios from a vast and unlabeled database to augment adaptability. + +Subsequently, the generator, comprising a combiner and a decoder, follows the same inference as combiner training with the initial pose and lane map specified by the user. We obtain the RAG embedding $z_{rag}$ with Eq. (3), and then infer the generated scenario through $\tau_{rag} \gets D(z_i, z_{rag}, z_m)$ . + +# 5 Experiment + +In this section, we start with an overview of the experimental setup and details regarding the implementation of the RealGEN framework. Then we discuss the baselines, including different types of autoencoders and variants of RealGen, used for comparison. In the discussion of the findings, we respectively evaluate the quality of scenario embedding and the generation capability of RAG. + +# 5.1 Settings and Implementation Details + +We conducted all training and evaluation with the nuScenes [5] dataset using the trajdata [34] package for data loading and processing. Each scenario spans + +a duration of 8 seconds, operating at a frequency of $2\mathrm{Hz}$ , and encompasses a maximum of 11 agents. We filter out agents that travel less than 3 meters in 8 seconds and select 11 agents closest to the ego vehicle. The map contains 100 lanes (each lane has 20 points) that are ordered according to the distance between the center of the lane and the center of the ego vehicle. + +To train RealGen, we use Adam [36] as the optimizer to update the parameters of the autoencoder and the combiner. To make the calculation of the Wasserstein distance efficient, we use the Sinkhorn distance [13], an entropy regularized approximation of the Wasserstein distance [52], with implementation in the GeomLoss [19] package. + +# 5.2Baselines + +In the experimental section, we evaluate the following reconstruction-based generative models as baselines for comparison. Autoencoder (AE) shares the same behavior encoder, map encoder, and decoder structures as RealGen, serving as the most straightforward baseline for scenario reconstruction. Contrastive AE mirrors the structure of the RealGen autoencoder but omits the initial pose as absolute information. Masked AE is a self-supervised learning baseline, which has been investigated for trajectory data as described in [9, 12, 56, 58]. To evaluate the controllable generation capability, we select a state-of-the-art modelLCTGen [49] as the baseline, which takes a high-level agent representation $z$ and a vector map $m$ as input. We also consider LCTGen w/o $z$ , a variant model proposed in the original paper to show the reconstruction results without using the agent information. Given that our method has two phases, we refer to the autoencoder component specifically as RealGen-AE and the complete model as RealGen. + +# 5.3 Evaluation of Behavior Embedding + +We first evaluate the quality of the learned representation of behavior, which is critical for the following retrieval and generation processes. + +Visualizing similar and dissimilar scenarios. After training the autoencoder with contrastive loss, the distance between behavior embeddings can be used as an indicator of the similarity of the two scenarios. To validate this statement, we visualize qualitative examples of using a query to find the most similar (minimal $W_{2}$ distance) and dissimilar (maximal $W_{2}$ distance) scenarios in Figure 3. We observe that the most similar scenario contains the same behavior and number of vehicles as in the query scenario. + +Classifying Scene ID with behavior embedding. To further provide quantitative results on how well the behavior embedding encodes behavior information, we leverage the Scene ID of the nuScenes dataset. Each scene typically lasts 20 seconds and our scenario segments last 8 seconds, so we get multiple segments belonging to the same Scene ID. We assume that the segments having the same Scene ID have similar behaviors so that we can calculate the accuracy by using the distance to find the closest segment. In Figure 4, we summarize + +![](images/8d94447a233949b044fad76d42fd238d99fd05b3f83b1513c868e46241024162.jpg) +Fig.3: Qualitative evaluation of similar and dissimilar scenarios calculated by our scenario embedding. Rectangles represent the initial poses of vehicles and the lines represent the future trajectories. + +the results in the left part, where top- $k$ means we calculate the accuracy for the closest $k$ segments. To validate the effectiveness of $W_{2}$ distance, we also permute the order of agents in the behavior embedding, which is denoted as cosine-permuted and $W_{2}$ -permuted. We find that the cosine distance cannot deal with the permuted setting, as the order is important when stacking the embedding of agents. Our method performs well in top-1 and top-5 settings but badly in others, which could be explained by the fact that segments in one scene have very different behaviors. To validate this, we plot the distance matrix between segments for 11 scenes in the right part of Figure 4. We find sub-blocks in the diagonal blocks, indicating that segments are similar in a small time interval but could be different when the time interval is large. + +Linear probing of behavior embedding. Linear probing, a commonly employed method to assess representations in self-supervised learn + +Table 1: Accuracy of linear probing + +
AEContrastive AEMasked AERealGen-AE
82.5%67.1%86.2%87.8%
+ +ing (SSL), involves training a linear classifier using the derived embeddings [27]. To train this classifier, we implemented heuristic rules to assign basic behavioral labels (acceleration, deceleration, stopping, keeping speed, left/right turn) to each agent's embedding. The outcomes, along with comparisons with baseline models, are presented in Table 1. These findings demonstrate that the embeddings generated by RealGen surpass all baseline models in terms of accuracy. + +# 5.4 Evaluation of Retrieval Augmented Generation + +Evaluating scene-level controllability and quality of scenario generation is an open problem due to the lack of quantitative metrics. Following most previous + +![](images/0ba044be2984e857b489034882666ad9d2bee681a44e58d26f7fdabdbd50ed4c.jpg) +Fig. 4: (a) Scene ID accuracy using the behavior embedding with difference distance metrics. (b) A matrix shows the Wasserstein distance between scenario segments, where each block contains the segments that belong to the same Scene ID. + +Table 2: Results of realism metrics. Recon.-based and Retrieval-based means using the target scenario and retrieved scenarios as input for generation, respectively. + +
CategoryMethodmADEmFDESpeedHeadingSCRORR
Recon.-basedAE0.18±0.030.41±0.060.04±0.010.10±0.010.02±0.000.02±0.00
Masked AE0.16±0.010.39±0.010.04±0.010.09±0.010.03±0.000.02±0.00
Contrastive AE0.92±0.021.47±0.040.12±0.000.36±0.020.04±0.000.04±0.00
RealGen-AE0.31±0.010.53±0.010.08±0.000.15±0.010.03±0.000.02±0.00
Retrieval-basedAE-KNN14.3±0.0316.4±0.050.57±0.010.59±0.020.15±0.010.15±0.01
LCTGen4.76±0.096.24±0.080.52±0.060.57±0.030.07±0.010.07±0.01
LCTGen w/o z14.2±0.0716.7±0.092.04±0.041.42±0.000.16±0.020.13±0.04
RealGen-AE-KNN13.1±0.0614.1±0.030.46±0.010.44±0.000.12±0.010.11±0.00
RealGen1.54±0.041.21±0.030.21±0.030.21±0.010.05±0.000.04±0.00
+ +work [18, 49, 64], we provide the results of the realism metrics and show the qualitative results of using RealGen for various downstream tasks. More results can be found in the appendix. + +Realism of generated scenario. To evaluate the realism of generated scenarios, we consider the following metrics and show the results in Table 2. We use the maximum mean discrepancy (MMD) [22] to measure the similarity in velocity and direction between the original scenarios and the generated scenarios. We also compare the mean average displacement error (mADE) and the mean final displacement error (mFDE) for the average reconstruction performance. To evaluate scene-level realism, we calculate the scene collision rate (SCR) and the off-road rate (ORR) following the metrics defined in [49]. The recon-based generation methods in Table 2 use the behavior of the target scenario as input. For this category, we compare RealGen-AE, only using the encoder and decoder modules, with three baseline methods. Due to the additional contrastive term, RealGen-AE achieves slightly worse performance than AE and Masked AE. However, RealGen is designed for retrieval-based generation, which uses retrieved scenarios rather than the target scenario as input. To fairly compare the generation performance, we design two baselines named AE-KNN and RealGen-AE-KNN, + +![](images/f7e4b1d8ea5b2ab2e2a60bf5374f0576dc882d1b1ccc08b123363cef7d066d9b.jpg) +Fig. 5: Examples of tag-retrieved scenarios generated by RealGen for six different tags. + +![](images/4d6dfbb28922c1ab5aae663a722917cc60ab7f09df7e9b4cee7401a5baa3e1f2.jpg) + +![](images/8082e18975e3ca40ce7807079e934502c0f2402a76e01254a32374b487371eb9.jpg) + +which use KNN to find the most similar behavior embedding to the target scenario and use this embedding as input to the decoder for generation. According to the results, we find that RealGen achieves comparable performance as recombined generation and outperforms baselines, which indicates the important role of the combiner in fusing the information of the retrieved scenarios. + +Generating tag-retrieved scenarios. Now we explore the qualitative performance of using RealGen for tag-retrieved generation. Given a target behavior tag, we first obtain several template scenarios from a small dataset and use them to retrieve more scenarios from the training database, which will be used for generation in the combiner. Since there is no existing dataset with tags, we manually labeled six tags - U-Turn, Overtaking, Left Lane Change, Right Lane Change, Left Turn, Right Turn - for + +![](images/fad1fdbdbf1c8bd46fc0d4bf839287bca04e3454b860efd0c427fb504ffb442b.jpg) +Fig. 6: Examples of generating crash scenarios from RealGen. The shadow rectangles represent the initial positions of agents. + +the nuScenes dataset to get template 1349 scenarios. We plot generated scenarios for each tag in Figure 5, where the left part of each example shows the given initial pose and map and the right part of each example shows the generated scenario from RealGen. + +# Safety-critical scenario gen + +eration. Beyond the tag mentioned above, RealGen demonstrates its capacity for in-context + +learning by generating critical and unseen crash scenarios, divergent from those of the training datasets. This is initiated by manually creating several crash scenario templates, guided by the scenarios recorded in the NHTSA Crash Report [40]. Subsequently, we generate crash scenarios using existing initial poses and maps of the dataset. Figure 6 illustrates six instances in which the shadowed rectangles denote the initial positions of agents and the red box highlights + +Table 3: Safety-critical scenario generation. + +
MethodRealgen-AE-RRealGen-RRealGen
Collision Rate (↑)0.920.830.59
+ +Table 4: Results of human evaluation of controllability. (Details in Appendix ??) + +
CategoryLeft-TurnRight-TurnLeft-Lane-ChangeRight-Lane-ChangeStraight
RealGen Preferred (%)81.891.797.893.3100.0
RealGen Score (0-5)4.27±1.054.27±0.693.96±1.944.17±1.933.94±2.27
LCTGen Score (0-5)2.15±1.442.08±1.192.42±1.952.0±1.962.14±2.31
+ +the point of collision. To quantitatively assess the performance of safety-critical scenarios, we compare RealGen with two baselines. RealGen-AE-R random samples behavior embeddings in the Realgen-AE model, and RealGen-R random retrieves behavior embeddings. According to the results in Table 3, we find that the scenarios generated by RealGen using crash scenarios as templates achieve the highest collision rate, which means that RealGen has more efficiency. + +Human evaluation of controllability. As there is no automatic way to evaluate controllability, we follow the protocol in [49] to perform A/B testing using human evaluation. We report the ratio of our method preferred as well as the absolute score (0-5) of both our method and the baseline. In Table 4, we find that scenarios generated by RealGen are highly preferred in most categories. The absolute score of RealGen is also much higher than that of LCTGen. + +Downstream task evaluation. A direct downstream task of our method is to use the generated data to augment the training dataset of trajectory + +Table 5: Downstream task evaluation. + +
MethodOriginalRandom Aug.RealGen Aug.
mADE (↓)3.5442.9202.309
collision rate (↓)0.0490.0370.018
+ +prediction models. We use Autobots [21] as a predictor and report the results trained on different datasets in Table 5. Original means using the original data in nuScenes [5], Random Aug. means augmenting the original dataset with Gaussian noise, and RealGen Aug. means augmenting the original dataset with scenarios from RealGen. We observe that the model trained with the RealGen dataset achieves the lowest mADE and collision rate. + +# 6 Conclusion and Limitation + +This paper proposes RealGen, a novel framework for traffic scenario generation that utilizes retrieval-augmented generation. Unlike previous approaches, which primarily rely on models replicating training distributions, RealGen demonstrates in-context learning abilities that synthesize scenarios by combining and modifying provided examples, enabling controlled generation. These scenarios can be automatically obtained from a retrieval system, which only requires the users to provide a few template scenarios as examples. We comprehensively evaluated the similarity autoencoder model for retrieval and the combiner model for generation. The findings indicate that RealGen achieves low reconstruction error and high generation quality. + +The primary limitation of our current method is that the behavior encoder focuses solely on agent trajectories, neglecting the intricate interactions between agents and lane maps in complex behaviors. Advancing the feature representation within the behavior encoder could significantly broaden RealGen's capacity to generate diverse and controllable traffic scenarios. + +# Acknowledgments + +Wenhao Ding contributed to this paper while being an intern at NVIDIA Research. Ding Zhao was partially supported by the National Science Foundation under grant CNS-2047454. + +# References + +1. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016) +2. Blattmann, A., Rombach, R., Oktay, K., Müller, J., Ommer, B.: Retrievalaugmented diffusion models. Advances in Neural Information Processing Systems 35, 15309-15324 (2022) +3. Borgeaud, S., Mensch, A., Hoffmann, J., Cai, T., Rutherford, E., Millican, K., Van Den Driessche, G.B., Lespiau, J.B., Damoc, B., Clark, A., et al.: Improving language models by retrieving from trillions of tokens. In: International conference on machine learning. pp. 2206-2240. PMLR (2022) +4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) +5. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020) +6. Cai, D., Wang, Y., Bi, W., Tu, Z., Liu, X., Shi, S.: Retrieval-guided dialogue response generation via a matching-to-generation framework. In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). pp. 1866-1875 (2019) +7. Cao, Y., Xiao, C., Anandkumar, A., Xu, D., Pavone, M.: Advdo: Realistic adversarial attacks for trajectory prediction. In: European Conference on Computer Vision. pp. 36-52. Springer (2022) +8. Chen, D., Fisch, A., Weston, J., Bordes, A.: Reading wikipedia to answer open-domain questions. arXiv preprint arXiv:1704.00051 (2017) +9. Chen, H., Wang, J., Shao, K., Liu, F., Hao, J., Guan, C., Chen, G., Heng, P.A.: Traj-mae: Masked autoencoders for trajectory prediction. arXiv preprint arXiv:2303.06697 (2023) +10. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020) +11. Chen, W., Hu, H., Sahara, C., Cohen, W.W.: Re-imagen: Retrieval-augmented text-to-image generator. arXiv preprint arXiv:2209.14491 (2022) +12. Cheng, J., Mei, X., Liu, M.: Forecast-mae: Self-supervised pre-training for motion forecasting with masked autoencoders. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8679-8689 (2023) +13. Cuturi, M.: Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems 26 (2013) +14. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) + +15. Ding, W., Chen, B., Li, B., Eun, K.J., Zhao, D.: Multimodal safety-critical scenarios generation for decision-making algorithms evaluation. IEEE Robotics and Automation Letters 6(2), 1551-1558 (2021) +16. Ding, W., Lin, H., Li, B., Eun, K.J., Zhao, D.: Semantically adversarial driving scenario generation with explicit knowledge integration. arXiv preprint arXiv:2106.04066 (2021) +17. Ding, W., Xu, M., Zhao, D.: Learning to collide: An adaptive safety-critical scenarios generating method. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE (2020) +18. Feng, L., Li, Q., Peng, Z., Tan, S., Zhou, B.: Trafficgen: Learning to generate diverse and realistic traffic scenarios. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 3567-3575. IEEE (2023) +19. Feydy, J., Séjourne, T., Vialard, F.X., Amari, S.i., Trouve, A., Peyré, G.: Interpolating between optimal transport and mmd using sinkhorn divergences. In: The 22nd International Conference on Artificial Intelligence and Statistics. pp. 2681-2690 (2019) +20. Gidaris, S., Singh, P., Komodakis, N.: Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728 (2018) +21. Girgis, R., Golemo, F., Codevilla, F., Weiss, M., D'Souza, J.A., Kahou, S.E., Heide, F., Pal, C.: Latent variable sequential set transformers for joint multi-agent motion prediction. arXiv preprint arXiv:2104.00563 (2021) +22. Gretton, A., Borgwardt, K.M., Rasch, M.J., Scholkopf, B., Smola, A.: A kernel two-sample test. The Journal of Machine Learning Research 13(1), 723-773 (2012) +23. Gulino, C., Fu, J., Luo, W., Tucker, G., Bronstein, E., Lu, Y., Harb, J., Pan, X., Wang, Y., Chen, X., et al.: Waymax: An accelerated, data-driven simulator for large-scale autonomous driving research. arXiv preprint arXiv:2310.08710 (2023) +24. Gutmann, M., Hyvarinen, A.: Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In: Proceedings of the thirteenth international conference on artificial intelligence and statistics. pp. 297-304. JMLR Workshop and Conference Proceedings (2010) +25. Guu, K., Lee, K., Tung, Z., Pasupat, P., Chang, M.: Retrieval augmented language model pre-training. In: International conference on machine learning. pp. 3929-3938. PMLR (2020) +26. Hanselmann, N., Renz, K., Chitta, K., Bhattacharyya, A., Geiger, A.: King: Generating safety-critical driving scenarios for robust imitation via kinematics gradients. In: European Conference on Computer Vision. pp. 335-352. Springer (2022) +27. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022) +28. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020) +29. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020) +30. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural computation 9(8), 1735-1780 (1997) +31. Hospedales, T., Antoniou, A., Micaelli, P., Storkey, A.: Meta-learning in neural networks: A survey. IEEE transactions on pattern analysis and machine intelligence 44(9), 5149-5169 (2021) + +32. Hu, A., Corrado, G., Griffiths, N., Murez, Z., Gurau, C., Yeo, H., Kendall, A., Cipolla, R., Shotton, J.: Model-based imitation learning for urban driving. Advances in Neural Information Processing Systems 35, 20703-20716 (2022) +33. Ibrihich, S., Oussous, A., Ibrihich, O., Esghir, M.: A review on recent research in information retrieval. Procedia Computer Science 201, 777-782 (2022) +34. Ivanovic, B., Song, G., Gilitschenski, I., Pavone, M.: trajdata: A unified interface to multiple human trajectory datasets. arXiv preprint arXiv:2307.13924 (2023) +35. Kim, J., Choi, S., Amplayo, R.K., Hwang, S.w.: Retrieval-augmented controllable review generation. In: Proceedings of the 28th International Conference on Computational Linguistics. pp. 2284-2295 (2020) +36. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) +37. Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Kuttler, H., Lewis, M., Yih, W.t., Rocktäschel, T., et al.: Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems 33, 9459-9474 (2020) +38. Li, Q., Peng, Z., Feng, L., Duan, C., Mo, W., Zhou, B., et al.: Scenarionet: Open-source platform for large-scale traffic scenario simulation and modeling. arXiv preprint arXiv:2306.12241 (2023) +39. Liu, Q., Yogatama, D., Blunsom, P.: Relational memory-augmented language models. Transactions of the Association for Computational Linguistics 10, 555-572 (2022) +40. NHTSA: Nhtsa crash viewer (2023), https://crashviewer.nhtsa.dot.gov/ +41. Noroozi, M., Favaro, P.: Unsupervised learning of visual representations by solving jigsaw puzzles. In: European conference on computer vision. pp. 69-84. Springer (2016) +42. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018) +43. Roshdi, A., Roohparvar, A.: Information retrieval techniques and applications. International Journal of Computer Networks and Communications Security 3(9), 373-377 (2015) +44. Schmidhuber, J.: Deep learning in neural networks: An overview. Neural networks 61, 85-117 (2015) +45. Shi, S., Jiang, L., Dai, D., Schiele, B.: Motion transformer with global intention localization and local movement refinement. Advances in Neural Information Processing Systems 35, 6531-6543 (2022) +46. Shurrab, S., Duwairi, R.: Self-supervised learning methods and applications in medical imaging analysis: A survey. PeerJ Computer Science 8, e1045 (2022) +47. Suo, S., Regalado, S., Casas, S., Urtasun, R.: Trafficsim: Learning to simulate realistic multi-agent behaviors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10400-10409 (2021) +48. Suo, S., Wong, K., Xu, J., Tu, J., Cui, A., Casas, S., Urtasun, R.: Mixsim: A hierarchical framework for mixed reality traffic simulation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9622-9631 (2023) +49. Tan, S., Ivanovic, B., Weng, X., Pavone, M., Kraehenbuehl, P.: Language conditioned traffic generation. In: Conference on Robot Learning. pp. 2714-2752. PMLR (2023) +50. Tan, S., Wong, K., Wang, S., Manivasagam, S., Ren, M., Urtasun, R.: Scenegen: Learning to generate realistic traffic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 892-901 (2021) + +51. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) +52. Villani, C., et al.: Optimal transport: old and new, vol. 338. Springer (2009) +53. Vincent, P., Larochelle, H., Bengio, Y., Manzagol, P.A.: Extracting and composing robust features with denoising autoencoders. In: Proceedings of the 25th international conference on Machine learning. pp. 1096-1103 (2008) +54. Wang, J., Pun, A., Tu, J., Manivasagam, S., Sadat, A., Casas, S., Ren, M., Urtasun, R.: Advsim: Generating safety-critical scenarios for self-driving vehicles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9909-9918 (2021) +55. Wang, Z., Nie, W., Qiao, Z., Xiao, C., Baraniuk, R., Anandkumar, A.: Retrieval-based controllable molecule generation. arXiv preprint arXiv:2208.11126 (2022) +56. Wu, P., Majumdar, A., Stone, K., Lin, Y., Mordatch, I., Abbeel, P., Rajeswaran, A.: Masked trajectory models for prediction, representation, and control. arXiv preprint arXiv:2305.02968 (2023) +57. Xu, P., Patwary, M., Shoeybi, M., Puri, R., Fung, P., Anandkumar, A., Catanzaro, B.: Megatron-cntrl: Controllable story generation with external knowledge using large-scale language models. arXiv preprint arXiv:2010.00840 (2020) +58. Yang, Y., Zhang, Q., Gilles, T., Batool, N., Folkesson, J.: Rmp: A random mask pretrain framework for motion prediction. arXiv preprint arXiv:2309.08989 (2023) +59. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of lmms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421 (2023) +60. Zhang, C., Zhang, C., Song, J., Yi, J.S.K., Zhang, K., Kweon, I.S.: A survey on masked autoencoder for self-supervised learning in vision and beyond. arXiv preprint arXiv:2208.00173 (2022) +61. Zhang, C., Tu, J., Zhang, L., Wong, K., Suo, S., Urtasun, R.: Learning realistic traffic agents in closed-loop. In: 7th Annual Conference on Robot Learning (2023) +62. Zhang, X., Bosselut, A., Yasunaga, M., Ren, H., Liang, P., Manning, C.D., Leskovec, J.: Greaselm: Graph reasoning enhanced language models. In: International conference on learning representations (2021) +63. Zhong, Z., Rempe, D., Chen, Y., Ivanovic, B., Cao, Y., Xu, D., Pavone, M., Ray, B.: Language-guided traffic simulation via scene-level diffusion. arXiv preprint arXiv:2306.06344 (2023) +64. Zhong, Z., Rempe, D., Xu, D., Chen, Y., Veer, S., Che, T., Ray, B., Pavone, M.: Guided conditional diffusion for controllable traffic simulation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 3560-3566. IEEE (2023) \ No newline at end of file diff --git a/2024/RealGen_ Retrieval Augmented Generation for Controllable Traffic Scenarios/images.zip b/2024/RealGen_ Retrieval Augmented Generation for Controllable Traffic Scenarios/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c71ebd73b0d6ca9d7945650ba75805a59934ac1e --- /dev/null +++ b/2024/RealGen_ Retrieval Augmented Generation for Controllable Traffic Scenarios/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be91b70c39d6ee4ef81b7ab2ef5dceccfe4a6b314cc919d550e3978a392b70df +size 451203 diff --git a/2024/RealGen_ Retrieval Augmented Generation for Controllable Traffic Scenarios/layout.json b/2024/RealGen_ Retrieval Augmented Generation for Controllable Traffic Scenarios/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..72ae00367baa256e4625f8946bd9c10434ae170c --- /dev/null +++ b/2024/RealGen_ Retrieval Augmented Generation for Controllable Traffic Scenarios/layout.json @@ -0,0 +1,10226 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 144, + 111, + 470, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 111, + 470, + 146 + ], + "spans": [ + { + "bbox": [ + 144, + 111, + 470, + 146 + ], + "type": "text", + "content": "RealGen: Retrieval Augmented Generation for Controllable Traffic Scenarios" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 141, + 167, + 472, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 167, + 472, + 192 + ], + "spans": [ + { + "bbox": [ + 141, + 167, + 472, + 192 + ], + "type": "text", + "content": "Wenhao Ding\\*1,2, Yulong Cao\\*2, Ding Zhao\\*1, Chaowei Xiao\\*2,3, and Marco Pavone\\*2,4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 244, + 201, + 369, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 201, + 369, + 213 + ], + "spans": [ + { + "bbox": [ + 244, + 201, + 369, + 213 + ], + "type": "text", + "content": "1 Carnegie Mellon University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 264, + 213, + 349, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 213, + 349, + 223 + ], + "spans": [ + { + "bbox": [ + 264, + 213, + 349, + 223 + ], + "type": "text", + "content": "2 NVIDIA Research" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 275, + 224, + 339, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 224, + 339, + 234 + ], + "spans": [ + { + "bbox": [ + 275, + 224, + 339, + 234 + ], + "type": "text", + "content": "3 UW-Madison" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 262, + 235, + 352, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 235, + 352, + 246 + ], + "spans": [ + { + "bbox": [ + 262, + 235, + 352, + 246 + ], + "type": "text", + "content": "4 Stanford University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 238, + 247, + 376, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 247, + 376, + 257 + ], + "spans": [ + { + "bbox": [ + 238, + 247, + 376, + 257 + ], + "type": "text", + "content": "{wenhaod, yulongc}@nvidia.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 277, + 455, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 277, + 455, + 497 + ], + "spans": [ + { + "bbox": [ + 159, + 277, + 455, + 497 + ], + "type": "text", + "content": "Abstract. Simulation plays a crucial role in the development of autonomous vehicles (AVs) due to the potential risks associated with real-world testing. Although significant progress has been made in the visual aspects of simulators, generating complex behavior among agents remains a formidable challenge. It is not only imperative to ensure realism in the scenarios generated but also essential to incorporate preferences and conditions to facilitate controllable generation for AV training and evaluation. Traditional methods, which rely mainly on memorizing the distribution of training datasets, often fail to generate unseen scenarios. Inspired by the success of retrieval augmented generation in large language models, we present RealGen, a novel retrieval-based in-context learning framework for traffic scenario generation. RealGen synthesizes new scenarios by combining behaviors from multiple retrieved examples in a gradient-free way, which may originate from templates or tagged scenarios. This in-context learning framework endows versatile generative capabilities, including the ability to edit scenarios, compose various behaviors, and produce critical scenarios. Evaluations show that RealGen offers considerable flexibility and controllability, marking a new direction in the field of controllable traffic scenario generation. Check our project website for more information: https://realgen.github.io." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 512, + 230, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 512, + 230, + 525 + ], + "spans": [ + { + "bbox": [ + 133, + 512, + 230, + 525 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Simulation is indispensable in the development of autonomous vehicles (AVs), primarily due to the considerable risks associated with training and evaluating these systems in real-world conditions. The biggest challenge in simulations lies in achieving realistic driving scenarios, as this realism influences the discrepancy between AV performance in simulated and actual environments. Although advancements in high-quality graphical engines have significantly enhanced the perception quality of simulators, the realism of agent behavior remains constrained because of the complicated interactions among naturalistic agents. To counteract this issue, data-driven simulation has emerged as a promising approach in the realm of autonomous driving, which leverages real-world scenario datasets to accurately generate the behaviors of agents." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "With the rapid achievement of deep generative models [29] and imitation learning algorithms [32], current data-driven simulations [23, 38] can generate scenarios that closely mimic human driver behavior. However, the effectiveness of these simulations in accelerating the development of AV is limited. This limitation stems from the need for scenarios that meet specific conditions tailored for targeted training and evaluation. Achieving such controllability in simulations is challenging due to the complex nature of driving scenarios, which involve intricate interactions, diverse road layouts, and varying traffic regulations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 223, + 264, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 223, + 264, + 438 + ], + "spans": [ + { + "bbox": [ + 130, + 223, + 264, + 438 + ], + "type": "text", + "content": "In pursuit of controllability, existing work applies additional guidance, typically in the form of constraint functions [16, 64] or languages [49, 63], to pre-trained scenario generative models. Regularization of the generation process through these tools is straightforward and effective, yet it encounters two principal challenges. First, the training of scenario generative models typically utilizes naturalistic datasets, which might not encompass the specific scenarios desired as per the control signals. Even if such" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 271, + 225, + 343, + 341 + ], + "blocks": [ + { + "bbox": [ + 271, + 225, + 343, + 341 + ], + "lines": [ + { + "bbox": [ + 271, + 225, + 343, + 341 + ], + "spans": [ + { + "bbox": [ + 271, + 225, + 343, + 341 + ], + "type": "image", + "image_path": "c2692d7e4ef746f27023fb6009f1edbf303e3631ea23102ce352a5bae027f12e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 272, + 342, + 342, + 350 + ], + "lines": [ + { + "bbox": [ + 272, + 342, + 342, + 350 + ], + "spans": [ + { + "bbox": [ + 272, + 342, + 342, + 350 + ], + "type": "text", + "content": "(a) Conventional method" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 350, + 226, + 480, + 342 + ], + "blocks": [ + { + "bbox": [ + 350, + 226, + 480, + 342 + ], + "lines": [ + { + "bbox": [ + 350, + 226, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 350, + 226, + 480, + 342 + ], + "type": "image", + "image_path": "17d61e5d3fabb3e7ce926a63f363956d5811a6f3db4fdf611e42045720b3588f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 390, + 342, + 434, + 350 + ], + "lines": [ + { + "bbox": [ + 390, + 342, + 434, + 350 + ], + "spans": [ + { + "bbox": [ + 390, + 342, + 434, + 350 + ], + "type": "text", + "content": "(b) Our method" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 269, + 356, + 482, + 433 + ], + "lines": [ + { + "bbox": [ + 269, + 356, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 269, + 356, + 482, + 433 + ], + "type": "text", + "content": "Fig. 1: (a) Conventional methods make the model memorize the data distribution for generating. (b) In contrast, our method employs a retriever to query datasets (including external data obtained after training) and uses a generative model to generate scenarios by integrating the information from the retrieved scenarios." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 438, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 511 + ], + "type": "text", + "content": "scenarios exist within the dataset, they are often omitted because of the rarity of long-tail data. The second challenge is that the representation of the guidance to the generative model may not be sufficiently expressive to accurately depict complex scenarios, such as specifying intricate interactions among multiple vehicles, using language. These limitations underscore the need for a more sophisticated and nuanced framework for controllable scenario generation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Retrieval Augmented Generation (RAG) [8, 25], which enhances the generative process by querying related information from external databases, represents great potential in the domain of large language models [37]. In contrast to conventional models that memorize all knowledge within their parameters, as shown in Figure 1(a), RAG models, shown in Figure 1(b), learn to generate comprehensive outputs by retrieving pertinent knowledge from a database, based on the input provided. A notable aspect of RAG is the ability of the database to undergo updates even after the model has been trained, allowing continuous improvement and adaptation. This flexible framework offers possibilities for controllable scenario generation by using appropriate template scenarios as input and facilitating the generation that is not only realistic but also aligned with specific training and evaluation requirements." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "text", + "content": "W. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "In this work, we present RealGen, a retrieval augmented generation framework for generating traffic scenarios. This framework, as shown in Figure 2, begins with the training of an encoder through contrastive self-supervised learning [42] to allow the retrieval process to query similar scenarios in a latent embedding space. Leveraging this latent representation, we subsequently train a generative model that combines retrieved scenarios to create novel scenarios. The key contributions of this paper are summarized below." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 207, + 481, + 291 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 207, + 481, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 207, + 481, + 242 + ], + "spans": [ + { + "bbox": [ + 138, + 207, + 481, + 242 + ], + "type": "text", + "content": "- We develop a novel contrastive autoencoder model to extract scenario embeddings as latent representations, which can be used for a wide range of downstream tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 243, + 480, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 243, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 243, + 480, + 266 + ], + "type": "text", + "content": "- We propose the first retrieval augmented generation framework using the latent representation tailored for controllable driving scenario generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 267, + 480, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 480, + 291 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 480, + 291 + ], + "type": "text", + "content": "- We validate our framework through qualitative and quantitative metrics, demonstrating strong flexibility and controllability of generated scenarios." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 309, + 237, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 309, + 237, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 309, + 237, + 322 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 335, + 371, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 335, + 371, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 335, + 371, + 346 + ], + "type": "text", + "content": "2.1 Traffic Modeling and Scenario Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 354, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 498 + ], + "type": "text", + "content": "Research on traffic modeling using generative models has received considerable attention in recent works. Notably, ScenarioNet [38] and Waymax [23] employ large-scale data to train imitation learning models, facilitating the generation of realistic multi-agent scenarios in simulations. SceneGen [50] utilizes the Long Short Term Memory [30] module to autoregressively generate the trajectories for vehicles and pedestrians based on provided maps. Additionally, TrafficSim [47] learns the multi-agent behaviors from real-world data, and TrafficGen [18] proposes a transformer-based autoencoder architecture to accurately model complex interactions involving multiple agents. MixSim [48] builds a reactive digital twin and finds safety-critical scenarios with black-box optimization. RTR [61] models reactive behaviors of vehicles using the combination of reinforcement learning and imitation learning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 498, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 582 + ], + "type": "text", + "content": "Numerous prior studies have employed adversarial generation to synthesize rare yet critical scenarios. L2C [17] utilizes a reinforcement learning framework, where the reward for scenario generation is based on the collision rate. To enhance the realism in purely adversarial generation methods, MMG [15] incorporates the data distribution as regularization. Further developments such as AdvSim [54], AdvDO [7], and KING [26] integrate vehicle dynamics to directly optimize the trajectory to find critical scenarios." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "The proliferation of large language models (LLMs) has led to recent approaches in generating traffic scenarios with language as conditions to follow instructions from humans. CTG++ [63] replaces the gradient guidance process of CTG [64] with cost functions generated by an LLM. Leveraging the power of GPT-4 [59], this framework can generate diverse motion data with language conditions. Additionally, LCTGen [49] also harnesses the strengths of GPT-4 to generate a heuristic intermediate representation of scenarios from language" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "type": "text", + "content": "RealGen: Retrieval Augmented Generation for Controllable Traffic Scenarios" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 140 + ], + "type": "text", + "content": "inputs. Subsequently, a generative model pre-trained on open-source datasets generates various scenarios with this representation as inputs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 156, + 326, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 156, + 326, + 169 + ], + "spans": [ + { + "bbox": [ + 131, + 156, + 326, + 169 + ], + "type": "text", + "content": "2.2 Retrieval-augmented Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 175, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 175, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 482, + 259 + ], + "type": "text", + "content": "Information retrieval (IR) [33] is the procedure of representing and searching a collection of data with the goal of extracting knowledge to satisfy the query of users. Predominantly utilized in search engines and digital libraries [43], IR primarily deals with the information stored in textual formats. Recently, IR has extended across diverse sectors, notably enhancing the quality of outputs in areas such as language modeling [3, 39], question answering [25, 62], image creation [2, 11], and the generation of molecular [55]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 259, + 483, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 483, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 483, + 354 + ], + "type": "text", + "content": "One intuitive usage of the retrieved information is to enhance input data through several methods, such as merging the original data and retrieved data [37], employing attention mechanism [3], or extracting skeleton [6]. The rationale behind retrieval systems stems from the impracticality of encoding all knowledge within model parameters, especially considering the dynamic nature of knowledge that evolves with human activities. Consequently, the capability to access external knowledge databases can significantly improve the precision and quality of generated responses, as evidenced in applications involving LLMs [3, 37]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 355, + 482, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 355, + 482, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 355, + 482, + 451 + ], + "type": "text", + "content": "Another usage of retrieved data, particularly pertinent to this study, involves the controllable generation by integrating desired features retrieved from the dataset. In [35], the authors generate product reviews with controllable information about the user, product, and rating. Similarly, the process of story creation can be viewed as a blend of external story databases with selected text fragments [57] Furthermore, [55] explores the controllable generation of molecules to meet various constraints, utilizing a database comprising simple molecules that typically meet only one of these constraints." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 468, + 326, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 468, + 326, + 480 + ], + "spans": [ + { + "bbox": [ + 131, + 468, + 326, + 480 + ], + "type": "text", + "content": "2.3 Self-supervised Feature Learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 486, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 581 + ], + "type": "text", + "content": "The necessity for manual labeling in supervised learning often results in human biases, extraneous noises, and labor-intensive efforts. Therefore, the paradigm of self-supervised learning (SSL) [46] is garnering heightened interest, particularly for its promising applications in language modeling and image interpretation. SSL algorithms usually learn implicit representations from extensive pools of unlabeled data without relying on human annotations. Generally, this line of research can be bifurcated into two categories [60]: generative SSL and discriminative SSL." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "In the domain of generative SSL, models employ an autoencoder to convert input data into a latent representation, followed by a reconstruction process. As an illustrative example, Denoising Autoencoder [53], an early example of this approach, reconstructs images from their noisy version to extract representative features. Another technique in this field is masked modeling, extensively utilized in language models such as GPT-3 [4] and BERT [14], which predicts the masked token to gain a semantic understanding of the text. More recently, the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "W. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "Masked Autoencoder [27] has emerged as a powerful framework, demonstrating its efficacy across a variety of downstream tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 260 + ], + "type": "text", + "content": "Discriminative SSL typically optimizes a discriminative loss to learn representations from sets of anchor, positive, and negative samples. Without ground-truth labels, these pairs are often constructed through solving jigsaw puzzles [41] or making geometry-based prediction [20]. A prominent example of discriminative SSL is contrastive learning, which brings samples from the same class closer while distancing those from different classes. Representative works such as MoCo [28] and SimCLR [10] learn embeddings by capturing invariant features between original data and its augmented variants. Additionally, InfoNCE [42], a method grounded in noise contrastive estimation [24], is widely used due to its effectiveness." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 281, + 371, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 371, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 371, + 295 + ], + "type": "text", + "content": "3 A Latent Representation of Scenario" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 308, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 415 + ], + "type": "text", + "content": "A crucial element within the retrieval system is the selection metric for data retrieval, which is typically implemented through a distance function between the query sample and the candidate samples in the database. Unlike text, which can be converted into word embeddings, traffic scenarios encompass sequential behaviors and intricate interactions among entities, complicating the establishment of a similarity metric for these scenarios. Consequently, in this section, we introduce a scenario autoencoder to extract latent representations that facilitate the assessment of similarity between various traffic scenarios, a vital component in our RAG framework." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 434, + 257, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 257, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 257, + 445 + ], + "type": "text", + "content": "3.1 Scenario Definition" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "text", + "content": "Each scenario is characterized by the trajectories " + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "inline_equation", + "content": "\\tau \\in \\mathbb{R}^{M\\times T\\times 5}" + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "text", + "content": ", encompassing " + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "text", + "content": " agents over a maximum of " + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "text", + "content": " time steps. The trajectory of each agent is composed of parameters " + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "inline_equation", + "content": "[x,y,v,c,s]" + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "text", + "content": ", signifying position " + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "text", + "content": ", position " + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "text", + "content": ", velocity " + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "text", + "content": ", cosine of the heading " + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "text", + "content": ", and sine of the heading " + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "text", + "content": ". The initial state of these entities is denoted as " + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "inline_equation", + "content": "\\tau_0\\in \\mathbb{R}^{M\\times 5}" + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "text", + "content": ". Furthermore, the map is encapsulated by " + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "inline_equation", + "content": "m\\in \\mathbb{R}^{S\\times 4}" + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "text", + "content": ", composed of " + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "text", + "content": " lane segments. The attributes of these points " + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "inline_equation", + "content": "[x_s,y_s,x_e,y_e]" + }, + { + "bbox": [ + 130, + 456, + 482, + 541 + ], + "type": "text", + "content": " correspond to the starting and ending positions of each segment." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 560, + 272, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 560, + 272, + 571 + ], + "spans": [ + { + "bbox": [ + 132, + 560, + 272, + 571 + ], + "type": "text", + "content": "3.2 Scenario Autoencoder" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Autoencoders [44] are used to learn compressed latent representations of high-dimensional data, where an encoder projects the data into latent code and a decoder reconstructs the code in the data space. Given that traffic scenarios encompass spatial and temporal dynamics from multiple agents, we design a hierarchical encoder structure alternating between spatial and temporal layers based on transformer architecture [51]. As depicted on the left of Figure 2, our design incorporates three encoders for behavior, map, and initial position." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 135, + 91, + 448, + 102 + ], + "type": "text", + "content": "RealGen: Retrieval Augmented Generation for Controllable Traffic Scenarios" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 113, + 482, + 255 + ], + "blocks": [ + { + "bbox": [ + 133, + 113, + 482, + 255 + ], + "lines": [ + { + "bbox": [ + 133, + 113, + 482, + 255 + ], + "spans": [ + { + "bbox": [ + 133, + 113, + 482, + 255 + ], + "type": "image", + "image_path": "f11185e87a54575c4943521e59e96c8a8d4a3217b9d0d8eedc46c35bc908207b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 258, + 482, + 338 + ], + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 338 + ], + "type": "text", + "content": "Fig. 2: Left: the training pipeline for the encoders and decoder aimed at learning latent embeddings of scenarios. A contrastive loss is applied to behavior embeddings to ensure invariance to absolute positions. Middle: the training pipeline for the combiner with frozen encoder and decoder parameters. We use K-Nearest Neighbors (KNN) to retrieve scenarios similar to a template scenario in the dataset and use the retrieved behaviors to reconstruct the template scenario. Right: the generation pipeline with a retriever and a generator." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "text", + "content": "First, we design a behavior encoder " + }, + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "inline_equation", + "content": "E_{b}" + }, + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "text", + "content": " with a spatial-temporal transformer structure that comprises " + }, + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "inline_equation", + "content": "L_{e}" + }, + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "text", + "content": " temporal transformer encoders " + }, + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "inline_equation", + "content": "(\\mathrm{Encoder}_{t}^{i})" + }, + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "inline_equation", + "content": "L_{e}" + }, + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "text", + "content": " spatial transformer encoders " + }, + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "inline_equation", + "content": "(\\mathrm{Encoder}_{s}^{i})" + }, + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "inline_equation", + "content": "i \\in \\{1, \\dots, L_{e}\\}" + }, + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "text", + "content": ". The trajectory " + }, + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "text", + "content": " is first transformed into a latent embedding " + }, + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "inline_equation", + "content": "z_{b}" + }, + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "text", + "content": " denotes the hidden dimension, through a Multiple Linear Perception (MLP) module. Subsequently, an alternating encoding procedure extracts spatial and temporal features from " + }, + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "inline_equation", + "content": "z_{b}" + }, + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "text", + "content": ". To preserve temporal information, we add a sinusoidal positional embedding (PE) to the embedding before employing the temporal transformer. To retain distinct agent information and further compress the behavior feature, we calculate the mean of the latent embedding across the temporal dimension, resulting in the final behavior embedding " + }, + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "inline_equation", + "content": "z_{b} \\in \\mathbb{R}^{M \\times H}" + }, + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "text", + "content": ". Besides the behavior encoder, we also process the map information represented by lane vectors with the attention mechanism [51]. Similarly to the first step in " + }, + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "inline_equation", + "content": "E_{b}" + }, + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "text", + "content": ", we project " + }, + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "text", + "content": " to the latent space with an MLP module. Then we apply a multi-head attention module (MHA) with layer normalization (LN) [1] to acquire the map embedding " + }, + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "inline_equation", + "content": "z_{m} \\in \\mathbb{R}^{S \\times H}" + }, + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "text", + "content": " with a learnable query embedding " + }, + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "inline_equation", + "content": "q_{m}" + }, + { + "bbox": [ + 130, + 348, + 482, + 540 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 541, + 483, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 541, + 483, + 625 + ], + "spans": [ + { + "bbox": [ + 130, + 541, + 483, + 625 + ], + "type": "text", + "content": "In the decoder, a spatial-temporal transformer architecture, similar to the behavior encoder, is designed to obtain the behavior embedding " + }, + { + "bbox": [ + 130, + 541, + 483, + 625 + ], + "type": "inline_equation", + "content": "z_{b}" + }, + { + "bbox": [ + 130, + 541, + 483, + 625 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 130, + 541, + 483, + 625 + ], + "type": "inline_equation", + "content": "L_{d}" + }, + { + "bbox": [ + 130, + 541, + 483, + 625 + ], + "type": "text", + "content": " times encoding. Given that " + }, + { + "bbox": [ + 130, + 541, + 483, + 625 + ], + "type": "inline_equation", + "content": "z_{b}" + }, + { + "bbox": [ + 130, + 541, + 483, + 625 + ], + "type": "text", + "content": " lacks the temporal dimension, we first replicate it " + }, + { + "bbox": [ + 130, + 541, + 483, + 625 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 541, + 483, + 625 + ], + "type": "text", + "content": " times and then add a PE before entering it into the temporal transformer encoder. Throughout the decoding process, the map embedding " + }, + { + "bbox": [ + 130, + 541, + 483, + 625 + ], + "type": "inline_equation", + "content": "z_{m}" + }, + { + "bbox": [ + 130, + 541, + 483, + 625 + ], + "type": "text", + "content": " is injected by a cross-attention mechanism: " + }, + { + "bbox": [ + 130, + 541, + 483, + 625 + ], + "type": "inline_equation", + "content": "z_{b} \\gets z_{b} + \\mathrm{MHA}(z_{b}, z_{m}, z_{m})" + }, + { + "bbox": [ + 130, + 541, + 483, + 625 + ], + "type": "text", + "content": ". MHA(" + }, + { + "bbox": [ + 130, + 541, + 483, + 625 + ], + "type": "inline_equation", + "content": "Q, K, V" + }, + { + "bbox": [ + 130, + 541, + 483, + 625 + ], + "type": "text", + "content": ") is the multi-head attention with " + }, + { + "bbox": [ + 130, + 541, + 483, + 625 + ], + "type": "inline_equation", + "content": "Q, K, V" + }, + { + "bbox": [ + 130, + 541, + 483, + 625 + ], + "type": "text", + "content": " representing query, key, and value:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 205, + 636, + 481, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 636, + 481, + 658 + ], + "spans": [ + { + "bbox": [ + 205, + 636, + 481, + 658 + ], + "type": "interline_equation", + "content": "\\mathrm {M H A} (Q, K, V) = \\text {C o n c a t e n a t e} \\left(h _ {1}, \\dots , h _ {i}\\right) W ^ {O},", + "image_path": "15ac241f7c97efb0cf2c7da573ac646aa9836dab14bbc962264b3e160f472de1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 207, + 654, + 372, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 654, + 372, + 668 + ], + "spans": [ + { + "bbox": [ + 207, + 654, + 372, + 668 + ], + "type": "interline_equation", + "content": "h _ {i} = \\mathrm {A t t e n t i o n} (Q W _ {i} ^ {Q}, K W _ {i} ^ {K}, V W _ {i} ^ {V}),", + "image_path": "08c90101421f7f9e1f434480aaf0eec6814cf8653c70e63b02379e7da9948443.jpg" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "text", + "content": "W. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 134, + 132, + 482, + 286 + ], + "blocks": [ + { + "bbox": [ + 138, + 118, + 350, + 130 + ], + "lines": [ + { + "bbox": [ + 138, + 118, + 350, + 130 + ], + "spans": [ + { + "bbox": [ + 138, + 118, + 350, + 130 + ], + "type": "text", + "content": "Algorithm 1: Details of Encoder and Decoder" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "lines": [ + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "spans": [ + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "text", + "content": "1 Behavior Encoder " + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "inline_equation", + "content": "E_{b}" + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "inline_equation", + "content": "(\\tau)" + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "text", + "content": ": 14 Initial Pose Encoder " + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "inline_equation", + "content": "(\\tau_{0})" + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "text", + "content": ": \n2 " + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "inline_equation", + "content": "z_{b} \\gets \\mathrm{MLP}(\\tau) // \\text{projection}" + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "text", + "content": " 15 " + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "inline_equation", + "content": "z_{i} \\gets \\mathrm{MLP}(\\tau_{0}) // \\text{projection}" + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "text", + "content": " \n3 for " + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "inline_equation", + "content": "[1, \\dots, L_{e}]" + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "text", + "content": " do 16 return initial pose embedding " + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "text", + "content": " \n4 " + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "inline_equation", + "content": "z_{b} \\gets \\mathrm{Encoder}_{t}^{i}(\\mathrm{PE} + z_{b})" + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "text", + "content": " 17 Decoder " + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "inline_equation", + "content": "D(z_{i}, z_{b}, z_{m})" + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "text", + "content": ": \n5 " + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "inline_equation", + "content": "z_{b} \\gets \\mathrm{Encoder}_{s}^{i}(z_{b})" + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "text", + "content": " 18 " + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "inline_equation", + "content": "z_{r} \\gets z_{b} + \\mathrm{MHA}(z_{b}, z_{i}, z_{i})" + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "text", + "content": " \n6 " + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "inline_equation", + "content": "z_{b} \\gets \\text{mean}(z_{b})" + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "text", + "content": " 19 for " + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "inline_equation", + "content": "[1, \\dots, L_{d}]" + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "text", + "content": " do \n7 return behavior embedding " + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "inline_equation", + "content": "z_{b}" + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "text", + "content": " 20 " + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "inline_equation", + "content": "z_{r} \\gets \\mathrm{Encoder}_{t}^{i}(\\mathrm{PE} + z_{r})" + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "text", + "content": " \n8 Map Encoder " + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "inline_equation", + "content": "E_{m}" + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "text", + "content": ": 21 \n9 Initialize a learnable query " + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "inline_equation", + "content": "q_{m}" + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "text", + "content": " 22 \n10 " + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "inline_equation", + "content": "z_{m} \\gets \\mathrm{MLP}(m) // \\text{projection}" + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "text", + "content": " 23 " + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "inline_equation", + "content": "\\hat{\\tau} \\gets \\mathrm{MLP}(z_{r}) // \\text{projection}" + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "text", + "content": " \n11 " + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "inline_equation", + "content": "z_{m} \\gets \\mathrm{LN}(\\mathrm{MHA}(q_{m}, z_{m}, z_{m}))" + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "text", + "content": " 24 return reconstructed trajectory " + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "inline_equation", + "content": "\\hat{\\tau}" + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "text", + "content": " \n12 " + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "inline_equation", + "content": "z_{m} \\gets \\mathrm{LN}(z_{m} + \\mathrm{MLP}(z_{m}))" + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "text", + "content": " \n13 return map embedding " + }, + { + "bbox": [ + 134, + 132, + 482, + 286 + ], + "type": "inline_equation", + "content": "z_{m}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 312, + 482, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 384 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 312, + 482, + 384 + ], + "type": "inline_equation", + "content": "W^{O}" + }, + { + "bbox": [ + 130, + 312, + 482, + 384 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 312, + 482, + 384 + ], + "type": "inline_equation", + "content": "W_{i}^{Q}" + }, + { + "bbox": [ + 130, + 312, + 482, + 384 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 312, + 482, + 384 + ], + "type": "inline_equation", + "content": "W_{i}^{K}" + }, + { + "bbox": [ + 130, + 312, + 482, + 384 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 312, + 482, + 384 + ], + "type": "inline_equation", + "content": "W_{i}^{V}" + }, + { + "bbox": [ + 130, + 312, + 482, + 384 + ], + "type": "text", + "content": " are learnable parameters. The last component of the decoder is an MLP that projects the hidden embedding back into the data domain, resulting in the reconstructed trajectory " + }, + { + "bbox": [ + 130, + 312, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\hat{\\tau}" + }, + { + "bbox": [ + 130, + 312, + 482, + 384 + ], + "type": "text", + "content": ". To train the encoders and decoder, the mean square error is used as the reconstruction loss, expressed as " + }, + { + "bbox": [ + 130, + 312, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r = |\\hat{\\tau} -\\tau |_2" + }, + { + "bbox": [ + 130, + 312, + 482, + 384 + ], + "type": "text", + "content": ". Comprehensive descriptions of the encoding and decoding processes are provided in Algorithm 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 385, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 482, + 456 + ], + "type": "text", + "content": "It is important to note that, while the current design of the autoencoder is sufficient for acquiring compressed representations of scenarios, these representations are not invariant to the absolute coordinates and the order of the agents. This limitation can lead to substantial embedding distances between similar scenarios. To address these issues, two enhancements are proposed and detailed in the following section." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 476, + 353, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 476, + 353, + 487 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 353, + 487 + ], + "type": "text", + "content": "3.3 Invariant feature with contrastive loss" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "content": "To enhance the representation of scenario similarity in the behavior embedding " + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "inline_equation", + "content": "z_{b}" + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "content": ", we employ contrastive learning to acquire invariant features. Specifically, we integrate InfoNCE [42] as an additional loss function " + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "content": ", which optimizes categorical cross-entropy to distinguish a positive sample from a batch of negative samples. In practice, for a query embedding " + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "inline_equation", + "content": "z_{b}" + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "content": ", a positive sample " + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "inline_equation", + "content": "z_{b}^{+}" + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "content": " is generated by applying random rotation and translation to the original scenario " + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\tau, m" + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "content": ". Meanwhile, negative samples " + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "inline_equation", + "content": "Z_{b}^{-}" + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "content": " are selected from the remaining samples in the same batch. In conventional InfoNCE, to bring " + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "inline_equation", + "content": "z_{b}" + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "inline_equation", + "content": "z_{b}^{+}" + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "content": " closer together, the inner product is calculated between the embedding " + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "inline_equation", + "content": "z_{b}" + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "content": " and the set " + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "inline_equation", + "content": "z_{b}^{+}, Z_{b}^{-}" + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "content": ", employing cross-entropy loss to identify the index of the positive sample." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "type": "text", + "content": "To ensure that the ordering of the agents does not affect the outcomes, " + }, + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "type": "inline_equation", + "content": "z_{b} \\in \\mathbb{R}^{M \\times H}" + }, + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "type": "text", + "content": " should exhibit permutation invariance across the dimension " + }, + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "type": "text", + "content": ". Otherwise, simply stacking them into a one-dimensional vector could erroneously represent distinct scenarios as significantly different. To establish permutation" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 135, + 90, + 448, + 102 + ], + "type": "text", + "content": "RealGen: Retrieval Augmented Generation for Controllable Traffic Scenarios" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "content": "invariance, we adopt the Wasserstein distance " + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "inline_equation", + "content": "W_{2}" + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "content": " [52] as a more suitable metric than cosine distance for assessing the similarity. Taking " + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "inline_equation", + "content": "z_{b}" + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "content": " as a distribution representing " + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "content": " individual behaviors, the intuition behind this choice is that the Wasserstein distance identifies the minimal adjustments necessary for behaviors in one scenario to reflect those in another. With such a powerful tool, we have the following contrastive loss:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 207, + 193, + 481, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 193, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 207, + 193, + 481, + 224 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} = - \\sum_ {z _ {b}} \\log \\frac {\\exp \\left[ - W _ {2} \\left(z _ {b} , z _ {b} ^ {+}\\right) \\right]}{\\sum_ {z ^ {\\prime} \\in \\left\\{z _ {b} ^ {+} , Z _ {b} ^ {-} \\right\\}} \\exp \\left[ - W _ {2} \\left(z _ {b} , z ^ {\\prime}\\right) \\right]}. \\tag {2}", + "image_path": "7e4633ec71b0aedf0167a2ea952692c406bcf02df41bdb8368f6c62b39731007.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 229, + 483, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 229, + 483, + 338 + ], + "spans": [ + { + "bbox": [ + 131, + 229, + 483, + 338 + ], + "type": "text", + "content": "The implementation of this loss results in the behavior embedding " + }, + { + "bbox": [ + 131, + 229, + 483, + 338 + ], + "type": "inline_equation", + "content": "z_{b}" + }, + { + "bbox": [ + 131, + 229, + 483, + 338 + ], + "type": "text", + "content": " lacking absolute coordinate data, making the decoder incapable of accurately reconstructing the precise trajectory. To handle this problem, we add the initial poses " + }, + { + "bbox": [ + 131, + 229, + 483, + 338 + ], + "type": "inline_equation", + "content": "\\tau_0" + }, + { + "bbox": [ + 131, + 229, + 483, + 338 + ], + "type": "text", + "content": " of all agents as a supplementary input for the decoder. We encode these poses into " + }, + { + "bbox": [ + 131, + 229, + 483, + 338 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 131, + 229, + 483, + 338 + ], + "type": "text", + "content": " with an MLP encoder " + }, + { + "bbox": [ + 131, + 229, + 483, + 338 + ], + "type": "inline_equation", + "content": "\\mathrm{Enc}_i" + }, + { + "bbox": [ + 131, + 229, + 483, + 338 + ], + "type": "text", + "content": ", and then integrate it into the decoder with an MHA module " + }, + { + "bbox": [ + 131, + 229, + 483, + 338 + ], + "type": "inline_equation", + "content": "z_{r} \\gets z_{b} + \\mathrm{MHA}(z_{b},z_{i},z_{i})" + }, + { + "bbox": [ + 131, + 229, + 483, + 338 + ], + "type": "text", + "content": ". Additional specifics regarding the initial pose encoder are detailed in Algorithm 1. In the training stage, the objective is to minimize a combined loss function " + }, + { + "bbox": [ + 131, + 229, + 483, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_r + \\lambda \\mathcal{L}_c" + }, + { + "bbox": [ + 131, + 229, + 483, + 338 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 229, + 483, + 338 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 131, + 229, + 483, + 338 + ], + "type": "text", + "content": " serves as a weighting factor for contrastive loss and is uniformly set at 0.1 for all experiments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 354, + 406, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 354, + 406, + 368 + ], + "spans": [ + { + "bbox": [ + 131, + 354, + 406, + 368 + ], + "type": "text", + "content": "4 Retrieval Augmented Scenario Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 376, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 461 + ], + "type": "text", + "content": "The autoencoder structure introduced in Section 3 builds a one-to-one generation framework, which still needs additional modules to support the retrieval augmented generation, which is a many-to-one framework. In this section, we introduce a module called Combiner, which takes input behavior embeddings from multiple scenarios and outputs the combined embedding. To train this module, we designed a KNN-based training pipeline that forces the model to learn to combine and edit existing scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 475, + 293, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 475, + 293, + 487 + ], + "spans": [ + { + "bbox": [ + 131, + 475, + 293, + 487 + ], + "type": "text", + "content": "4.1 The Training of Combiner" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 492, + 482, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 492, + 482, + 576 + ], + "spans": [ + { + "bbox": [ + 130, + 492, + 482, + 576 + ], + "type": "text", + "content": "Unlike the trajectory prediction task [45] which has ground truth as the optimization target, the objective in training the combiner is to learn the alignment of behaviors in the retrieved scenarios with the initial pose and the specified map. A well-trained combiner should be able to compose behaviors from these retrieved scenarios, thereby generating new scenarios that have a resemblance to all the retrieved scenarios. This type of objective aligns with the concept of meta-learning, or 'learning to learn', as described in [31]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 576, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 576, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 576, + 482, + 635 + ], + "type": "text", + "content": "Within our training framework, for a given query scenario " + }, + { + "bbox": [ + 131, + 576, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\tau, m" + }, + { + "bbox": [ + 131, + 576, + 482, + 635 + ], + "type": "text", + "content": " in the dataset, we initially utilize the behavior encoder to obtain " + }, + { + "bbox": [ + 131, + 576, + 482, + 635 + ], + "type": "inline_equation", + "content": "z_{b}" + }, + { + "bbox": [ + 131, + 576, + 482, + 635 + ], + "type": "text", + "content": " and then use KNN to identify " + }, + { + "bbox": [ + 131, + 576, + 482, + 635 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 576, + 482, + 635 + ], + "type": "text", + "content": " similar behavior embeddings from the database, denoted " + }, + { + "bbox": [ + 131, + 576, + 482, + 635 + ], + "type": "inline_equation", + "content": "z_{ret} = [z_{b,1}, \\dots, z_{b,K}]" + }, + { + "bbox": [ + 131, + 576, + 482, + 635 + ], + "type": "text", + "content": ". Building on this, we propose a model comprising two MHA modules as the combiner:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 233, + 640, + 481, + 660 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 640, + 481, + 660 + ], + "spans": [ + { + "bbox": [ + 233, + 640, + 481, + 660 + ], + "type": "interline_equation", + "content": "z _ {r a g} \\leftarrow z _ {i} + \\mathrm {M H A} \\left(z _ {i}, z _ {r e t}, z _ {r e t}\\right), \\tag {3}", + "image_path": "2829f7986911e0b0e4631af5effa5c4f6b855e439223d05f8227266b448ba682.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 233, + 656, + 382, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 656, + 382, + 667 + ], + "spans": [ + { + "bbox": [ + 233, + 656, + 382, + 667 + ], + "type": "interline_equation", + "content": "z _ {r a g} \\leftarrow z _ {r a g} + \\mathrm {M H A} (z _ {r a g}, z _ {m}, z _ {m}),", + "image_path": "13ff0fe5b7dcbb970a96538affcf73944b6051adf45e48ed942f4bb724177969.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "text", + "content": "W. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 189 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 189 + ], + "type": "inline_equation", + "content": "z_{i} \\gets E_{i}(x_{0})" + }, + { + "bbox": [ + 130, + 116, + 479, + 189 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 189 + ], + "type": "inline_equation", + "content": "z_{m} \\gets E_{m}(m)" + }, + { + "bbox": [ + 130, + 116, + 479, + 189 + ], + "type": "text", + "content": " are the initial pose embedding and the map embedding of the query scenario. Gradients are stopped for " + }, + { + "bbox": [ + 130, + 116, + 479, + 189 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 130, + 116, + 479, + 189 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 189 + ], + "type": "inline_equation", + "content": "z_{m}" + }, + { + "bbox": [ + 130, + 116, + 479, + 189 + ], + "type": "text", + "content": ". Assuming that the " + }, + { + "bbox": [ + 130, + 116, + 479, + 189 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 116, + 479, + 189 + ], + "type": "text", + "content": " nearest scenarios adequately represent the query scenario, it should be feasible to reconstruct the behavior of the query scenario, " + }, + { + "bbox": [ + 130, + 116, + 479, + 189 + ], + "type": "inline_equation", + "content": "z_{b}" + }, + { + "bbox": [ + 130, + 116, + 479, + 189 + ], + "type": "text", + "content": ", using the retrieved scenario, " + }, + { + "bbox": [ + 130, + 116, + 479, + 189 + ], + "type": "inline_equation", + "content": "z_{rag}" + }, + { + "bbox": [ + 130, + 116, + 479, + 189 + ], + "type": "text", + "content": ". Consequently, we employ the following loss function as our training goal:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 241, + 195, + 481, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 195, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 241, + 195, + 481, + 209 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r a g} = \\left\\| D \\left(z _ {i}, z _ {r a g}, z _ {m}\\right) - \\tau \\right\\| _ {2}, \\tag {4}", + "image_path": "71afd1bec5d4ebb53722ed45136482404632dfea5b6eba2be419dff2622e63f0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 213, + 482, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 262 + ], + "type": "text", + "content": "wherein the parameters of the decoder " + }, + { + "bbox": [ + 130, + 213, + 482, + 262 + ], + "type": "inline_equation", + "content": "D\\phi" + }, + { + "bbox": [ + 130, + 213, + 482, + 262 + ], + "type": "text", + "content": " remain fixed during training. Essentially, this training method is designed to learn an \"inverse\" operation of the KNN, aiming to reconstruct a query scenario that closely resembles all " + }, + { + "bbox": [ + 130, + 213, + 482, + 262 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 213, + 482, + 262 + ], + "type": "text", + "content": " retrieved scenarios in the database." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 277, + 281, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 277, + 281, + 290 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 281, + 290 + ], + "type": "text", + "content": "4.2 The generation pipeline" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 294, + 479, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 479, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 479, + 342 + ], + "type": "text", + "content": "The pipeline for scenario generation, as delineated in Figure 2, is categorized into retriever and generator components. The generation process is expedited by pre-processing all scenarios in the database using a behavior encoder, which yields behavior embeddings that facilitate efficient similarity computation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 342, + 480, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 480, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 480, + 461 + ], + "type": "text", + "content": "The retriever enhances the versatility of generation by dividing the process into two stages. In the initial stage, users can choose a set of template scenarios that depict specific conditions. These may include manually annotated scenarios with tags denoting actions like left or right turns, enabling the generation of additional scenarios under similar tags or even a combination thereof. In addition, templates can include critical and interesting scenarios collected from real-world data. Solely relying on these templates for generation could be limited, which is mitigated by incorporating a secondary phase, which employs a KNN approach to fetch high-quality scenarios from a vast and unlabeled database to augment adaptability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 462, + 481, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 481, + 511 + ], + "type": "text", + "content": "Subsequently, the generator, comprising a combiner and a decoder, follows the same inference as combiner training with the initial pose and lane map specified by the user. We obtain the RAG embedding " + }, + { + "bbox": [ + 130, + 462, + 481, + 511 + ], + "type": "inline_equation", + "content": "z_{rag}" + }, + { + "bbox": [ + 130, + 462, + 481, + 511 + ], + "type": "text", + "content": " with Eq. (3), and then infer the generated scenario through " + }, + { + "bbox": [ + 130, + 462, + 481, + 511 + ], + "type": "inline_equation", + "content": "\\tau_{rag} \\gets D(z_i, z_{rag}, z_m)" + }, + { + "bbox": [ + 130, + 462, + 481, + 511 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 525, + 224, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 224, + 540 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 224, + 540 + ], + "type": "text", + "content": "5 Experiment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 548, + 479, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 479, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 479, + 609 + ], + "type": "text", + "content": "In this section, we start with an overview of the experimental setup and details regarding the implementation of the RealGEN framework. Then we discuss the baselines, including different types of autoencoders and variants of RealGen, used for comparison. In the discussion of the findings, we respectively evaluate the quality of scenario embedding and the generation capability of RAG." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 624, + 347, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 347, + 636 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 347, + 636 + ], + "type": "text", + "content": "5.1 Settings and Implementation Details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "text", + "content": "We conducted all training and evaluation with the nuScenes [5] dataset using the trajdata [34] package for data loading and processing. Each scenario spans" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 135, + 91, + 448, + 102 + ], + "type": "text", + "content": "RealGen: Retrieval Augmented Generation for Controllable Traffic Scenarios" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": "a duration of 8 seconds, operating at a frequency of " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "2\\mathrm{Hz}" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": ", and encompasses a maximum of 11 agents. We filter out agents that travel less than 3 meters in 8 seconds and select 11 agents closest to the ego vehicle. The map contains 100 lanes (each lane has 20 points) that are ordered according to the distance between the center of the lane and the center of the ego vehicle." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 482, + 236 + ], + "type": "text", + "content": "To train RealGen, we use Adam [36] as the optimizer to update the parameters of the autoencoder and the combiner. To make the calculation of the Wasserstein distance efficient, we use the Sinkhorn distance [13], an entropy regularized approximation of the Wasserstein distance [52], with implementation in the GeomLoss [19] package." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 252, + 208, + 263 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 252, + 208, + 263 + ], + "spans": [ + { + "bbox": [ + 131, + 252, + 208, + 263 + ], + "type": "text", + "content": "5.2Baselines" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 271, + 483, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 483, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 483, + 439 + ], + "type": "text", + "content": "In the experimental section, we evaluate the following reconstruction-based generative models as baselines for comparison. Autoencoder (AE) shares the same behavior encoder, map encoder, and decoder structures as RealGen, serving as the most straightforward baseline for scenario reconstruction. Contrastive AE mirrors the structure of the RealGen autoencoder but omits the initial pose as absolute information. Masked AE is a self-supervised learning baseline, which has been investigated for trajectory data as described in [9, 12, 56, 58]. To evaluate the controllable generation capability, we select a state-of-the-art modelLCTGen [49] as the baseline, which takes a high-level agent representation " + }, + { + "bbox": [ + 130, + 271, + 483, + 439 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 271, + 483, + 439 + ], + "type": "text", + "content": " and a vector map " + }, + { + "bbox": [ + 130, + 271, + 483, + 439 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 271, + 483, + 439 + ], + "type": "text", + "content": " as input. We also consider LCTGen w/o " + }, + { + "bbox": [ + 130, + 271, + 483, + 439 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 271, + 483, + 439 + ], + "type": "text", + "content": ", a variant model proposed in the original paper to show the reconstruction results without using the agent information. Given that our method has two phases, we refer to the autoencoder component specifically as RealGen-AE and the complete model as RealGen." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 455, + 338, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 455, + 338, + 467 + ], + "spans": [ + { + "bbox": [ + 131, + 455, + 338, + 467 + ], + "type": "text", + "content": "5.3 Evaluation of Behavior Embedding" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 481, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 498 + ], + "type": "text", + "content": "We first evaluate the quality of the learned representation of behavior, which is critical for the following retrieval and generation processes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 582 + ], + "type": "text", + "content": "Visualizing similar and dissimilar scenarios. After training the autoencoder with contrastive loss, the distance between behavior embeddings can be used as an indicator of the similarity of the two scenarios. To validate this statement, we visualize qualitative examples of using a query to find the most similar (minimal " + }, + { + "bbox": [ + 130, + 498, + 482, + 582 + ], + "type": "inline_equation", + "content": "W_{2}" + }, + { + "bbox": [ + 130, + 498, + 482, + 582 + ], + "type": "text", + "content": " distance) and dissimilar (maximal " + }, + { + "bbox": [ + 130, + 498, + 482, + 582 + ], + "type": "inline_equation", + "content": "W_{2}" + }, + { + "bbox": [ + 130, + 498, + 482, + 582 + ], + "type": "text", + "content": " distance) scenarios in Figure 3. We observe that the most similar scenario contains the same behavior and number of vehicles as in the query scenario." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Classifying Scene ID with behavior embedding. To further provide quantitative results on how well the behavior embedding encodes behavior information, we leverage the Scene ID of the nuScenes dataset. Each scene typically lasts 20 seconds and our scenario segments last 8 seconds, so we get multiple segments belonging to the same Scene ID. We assume that the segments having the same Scene ID have similar behaviors so that we can calculate the accuracy by using the distance to find the closest segment. In Figure 4, we summarize" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "W. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 482, + 296 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 482, + 296 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 482, + 296 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 482, + 296 + ], + "type": "image", + "image_path": "8d94447a233949b044fad76d42fd238d99fd05b3f83b1513c868e46241024162.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 300, + 482, + 333 + ], + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 333 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 333 + ], + "type": "text", + "content": "Fig.3: Qualitative evaluation of similar and dissimilar scenarios calculated by our scenario embedding. Rectangles represent the initial poses of vehicles and the lines represent the future trajectories." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 349, + 482, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 349, + 482, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 349, + 482, + 480 + ], + "type": "text", + "content": "the results in the left part, where top-" + }, + { + "bbox": [ + 130, + 349, + 482, + 480 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 349, + 482, + 480 + ], + "type": "text", + "content": " means we calculate the accuracy for the closest " + }, + { + "bbox": [ + 130, + 349, + 482, + 480 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 349, + 482, + 480 + ], + "type": "text", + "content": " segments. To validate the effectiveness of " + }, + { + "bbox": [ + 130, + 349, + 482, + 480 + ], + "type": "inline_equation", + "content": "W_{2}" + }, + { + "bbox": [ + 130, + 349, + 482, + 480 + ], + "type": "text", + "content": " distance, we also permute the order of agents in the behavior embedding, which is denoted as cosine-permuted and " + }, + { + "bbox": [ + 130, + 349, + 482, + 480 + ], + "type": "inline_equation", + "content": "W_{2}" + }, + { + "bbox": [ + 130, + 349, + 482, + 480 + ], + "type": "text", + "content": "-permuted. We find that the cosine distance cannot deal with the permuted setting, as the order is important when stacking the embedding of agents. Our method performs well in top-1 and top-5 settings but badly in others, which could be explained by the fact that segments in one scene have very different behaviors. To validate this, we plot the distance matrix between segments for 11 scenes in the right part of Figure 4. We find sub-blocks in the diagonal blocks, indicating that segments are similar in a small time interval but could be different when the time interval is large." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 481, + 298, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 298, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 298, + 528 + ], + "type": "text", + "content": "Linear probing of behavior embedding. Linear probing, a commonly employed method to assess representations in self-supervised learn" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 494, + 480, + 523 + ], + "blocks": [ + { + "bbox": [ + 317, + 482, + 468, + 493 + ], + "lines": [ + { + "bbox": [ + 317, + 482, + 468, + 493 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 468, + 493 + ], + "type": "text", + "content": "Table 1: Accuracy of linear probing" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 494, + 480, + 523 + ], + "lines": [ + { + "bbox": [ + 307, + 494, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 307, + 494, + 480, + 523 + ], + "type": "table", + "html": "
AEContrastive AEMasked AERealGen-AE
82.5%67.1%86.2%87.8%
", + "image_path": "94d128ad46f2eeaeed89c0ab1c83cb054705f13facbc77bc567ed9f002bf504a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 529, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 482, + 601 + ], + "type": "text", + "content": "ing (SSL), involves training a linear classifier using the derived embeddings [27]. To train this classifier, we implemented heuristic rules to assign basic behavioral labels (acceleration, deceleration, stopping, keeping speed, left/right turn) to each agent's embedding. The outcomes, along with comparisons with baseline models, are presented in Table 1. These findings demonstrate that the embeddings generated by RealGen surpass all baseline models in terms of accuracy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 620, + 399, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 620, + 399, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 620, + 399, + 632 + ], + "type": "text", + "content": "5.4 Evaluation of Retrieval Augmented Generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Evaluating scene-level controllability and quality of scenario generation is an open problem due to the lack of quantitative metrics. Following most previous" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 135, + 91, + 448, + 102 + ], + "type": "text", + "content": "RealGen: Retrieval Augmented Generation for Controllable Traffic Scenarios" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 176, + 114, + 434, + 277 + ], + "blocks": [ + { + "bbox": [ + 176, + 114, + 434, + 277 + ], + "lines": [ + { + "bbox": [ + 176, + 114, + 434, + 277 + ], + "spans": [ + { + "bbox": [ + 176, + 114, + 434, + 277 + ], + "type": "image", + "image_path": "0ba044be2984e857b489034882666ad9d2bee681a44e58d26f7fdabdbd50ed4c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 280, + 482, + 314 + ], + "lines": [ + { + "bbox": [ + 130, + 280, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 482, + 314 + ], + "type": "text", + "content": "Fig. 4: (a) Scene ID accuracy using the behavior embedding with difference distance metrics. (b) A matrix shows the Wasserstein distance between scenario segments, where each block contains the segments that belong to the same Scene ID." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 138, + 343, + 470, + 441 + ], + "blocks": [ + { + "bbox": [ + 132, + 316, + 482, + 339 + ], + "lines": [ + { + "bbox": [ + 132, + 316, + 482, + 339 + ], + "spans": [ + { + "bbox": [ + 132, + 316, + 482, + 339 + ], + "type": "text", + "content": "Table 2: Results of realism metrics. Recon.-based and Retrieval-based means using the target scenario and retrieved scenarios as input for generation, respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 343, + 470, + 441 + ], + "lines": [ + { + "bbox": [ + 138, + 343, + 470, + 441 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 470, + 441 + ], + "type": "table", + "html": "
CategoryMethodmADEmFDESpeedHeadingSCRORR
Recon.-basedAE0.18±0.030.41±0.060.04±0.010.10±0.010.02±0.000.02±0.00
Masked AE0.16±0.010.39±0.010.04±0.010.09±0.010.03±0.000.02±0.00
Contrastive AE0.92±0.021.47±0.040.12±0.000.36±0.020.04±0.000.04±0.00
RealGen-AE0.31±0.010.53±0.010.08±0.000.15±0.010.03±0.000.02±0.00
Retrieval-basedAE-KNN14.3±0.0316.4±0.050.57±0.010.59±0.020.15±0.010.15±0.01
LCTGen4.76±0.096.24±0.080.52±0.060.57±0.030.07±0.010.07±0.01
LCTGen w/o z14.2±0.0716.7±0.092.04±0.041.42±0.000.16±0.020.13±0.04
RealGen-AE-KNN13.1±0.0614.1±0.030.46±0.010.44±0.000.12±0.010.11±0.00
RealGen1.54±0.041.21±0.030.21±0.030.21±0.010.05±0.000.04±0.00
", + "image_path": "94dd70d1ecc510fdb390c5a64f22d1ae2628492b1557dd526c1b9523d2d1a0df.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 449, + 481, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 481, + 485 + ], + "type": "text", + "content": "work [18, 49, 64], we provide the results of the realism metrics and show the qualitative results of using RealGen for various downstream tasks. More results can be found in the appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "Realism of generated scenario. To evaluate the realism of generated scenarios, we consider the following metrics and show the results in Table 2. We use the maximum mean discrepancy (MMD) [22] to measure the similarity in velocity and direction between the original scenarios and the generated scenarios. We also compare the mean average displacement error (mADE) and the mean final displacement error (mFDE) for the average reconstruction performance. To evaluate scene-level realism, we calculate the scene collision rate (SCR) and the off-road rate (ORR) following the metrics defined in [49]. The recon-based generation methods in Table 2 use the behavior of the target scenario as input. For this category, we compare RealGen-AE, only using the encoder and decoder modules, with three baseline methods. Due to the additional contrastive term, RealGen-AE achieves slightly worse performance than AE and Masked AE. However, RealGen is designed for retrieval-based generation, which uses retrieved scenarios rather than the target scenario as input. To fairly compare the generation performance, we design two baselines named AE-KNN and RealGen-AE-KNN," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "text", + "content": "W. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 113, + 250, + 232 + ], + "blocks": [ + { + "bbox": [ + 133, + 113, + 250, + 232 + ], + "lines": [ + { + "bbox": [ + 133, + 113, + 250, + 232 + ], + "spans": [ + { + "bbox": [ + 133, + 113, + 250, + 232 + ], + "type": "image", + "image_path": "f7e4b1d8ea5b2ab2e2a60bf5374f0576dc882d1b1ccc08b123363cef7d066d9b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 236, + 480, + 247 + ], + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 247 + ], + "type": "text", + "content": "Fig. 5: Examples of tag-retrieved scenarios generated by RealGen for six different tags." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 250, + 113, + 366, + 232 + ], + "blocks": [ + { + "bbox": [ + 250, + 113, + 366, + 232 + ], + "lines": [ + { + "bbox": [ + 250, + 113, + 366, + 232 + ], + "spans": [ + { + "bbox": [ + 250, + 113, + 366, + 232 + ], + "type": "image", + "image_path": "4d6dfbb28922c1ab5aae663a722917cc60ab7f09df7e9b4cee7401a5baa3e1f2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 366, + 113, + 481, + 232 + ], + "blocks": [ + { + "bbox": [ + 366, + 113, + 481, + 232 + ], + "lines": [ + { + "bbox": [ + 366, + 113, + 481, + 232 + ], + "spans": [ + { + "bbox": [ + 366, + 113, + 481, + 232 + ], + "type": "image", + "image_path": "8082e18975e3ca40ce7807079e934502c0f2402a76e01254a32374b487371eb9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 268, + 480, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 480, + 327 + ], + "type": "text", + "content": "which use KNN to find the most similar behavior embedding to the target scenario and use this embedding as input to the decoder for generation. According to the results, we find that RealGen achieves comparable performance as recombined generation and outperforms baselines, which indicates the important role of the combiner in fusing the information of the retrieved scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 330, + 298, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 298, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 298, + 496 + ], + "type": "text", + "content": "Generating tag-retrieved scenarios. Now we explore the qualitative performance of using RealGen for tag-retrieved generation. Given a target behavior tag, we first obtain several template scenarios from a small dataset and use them to retrieve more scenarios from the training database, which will be used for generation in the combiner. Since there is no existing dataset with tags, we manually labeled six tags - U-Turn, Overtaking, Left Lane Change, Right Lane Change, Left Turn, Right Turn - for" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 334, + 481, + 451 + ], + "blocks": [ + { + "bbox": [ + 306, + 334, + 481, + 451 + ], + "lines": [ + { + "bbox": [ + 306, + 334, + 481, + 451 + ], + "spans": [ + { + "bbox": [ + 306, + 334, + 481, + 451 + ], + "type": "image", + "image_path": "fad1fdbdbf1c8bd46fc0d4bf839287bca04e3454b860efd0c427fb504ffb442b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 456, + 480, + 490 + ], + "lines": [ + { + "bbox": [ + 304, + 456, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 480, + 490 + ], + "type": "text", + "content": "Fig. 6: Examples of generating crash scenarios from RealGen. The shadow rectangles represent the initial positions of agents." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 497, + 479, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 479, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 479, + 544 + ], + "type": "text", + "content": "the nuScenes dataset to get template 1349 scenarios. We plot generated scenarios for each tag in Figure 5, where the left part of each example shows the given initial pose and map and the right part of each example shows the generated scenario from RealGen." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 147, + 546, + 285, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 546, + 285, + 558 + ], + "spans": [ + { + "bbox": [ + 147, + 546, + 285, + 558 + ], + "type": "text", + "content": "Safety-critical scenario gen" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 558, + 284, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 284, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 284, + 593 + ], + "type": "text", + "content": "eration. Beyond the tag mentioned above, RealGen demonstrates its capacity for in-context" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": "learning by generating critical and unseen crash scenarios, divergent from those of the training datasets. This is initiated by manually creating several crash scenario templates, guided by the scenarios recorded in the NHTSA Crash Report [40]. Subsequently, we generate crash scenarios using existing initial poses and maps of the dataset. Figure 6 illustrates six instances in which the shadowed rectangles denote the initial positions of agents and the red box highlights" + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 288, + 561, + 482, + 590 + ], + "blocks": [ + { + "bbox": [ + 293, + 550, + 474, + 560 + ], + "lines": [ + { + "bbox": [ + 293, + 550, + 474, + 560 + ], + "spans": [ + { + "bbox": [ + 293, + 550, + 474, + 560 + ], + "type": "text", + "content": "Table 3: Safety-critical scenario generation." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 288, + 561, + 482, + 590 + ], + "lines": [ + { + "bbox": [ + 288, + 561, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 288, + 561, + 482, + 590 + ], + "type": "table", + "html": "
MethodRealgen-AE-RRealGen-RRealGen
Collision Rate (↑)0.920.830.59
", + "image_path": "56c74e7baa16b454f2f07d4fae748ebd85421f9ae3f3c9a826b30e838346e732.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 135, + 91, + 448, + 102 + ], + "type": "text", + "content": "RealGen: Retrieval Augmented Generation for Controllable Traffic Scenarios" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 130, + 481, + 176 + ], + "blocks": [ + { + "bbox": [ + 139, + 114, + 473, + 126 + ], + "lines": [ + { + "bbox": [ + 139, + 114, + 473, + 126 + ], + "spans": [ + { + "bbox": [ + 139, + 114, + 473, + 126 + ], + "type": "text", + "content": "Table 4: Results of human evaluation of controllability. (Details in Appendix ??)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 130, + 481, + 176 + ], + "lines": [ + { + "bbox": [ + 133, + 130, + 481, + 176 + ], + "spans": [ + { + "bbox": [ + 133, + 130, + 481, + 176 + ], + "type": "table", + "html": "
CategoryLeft-TurnRight-TurnLeft-Lane-ChangeRight-Lane-ChangeStraight
RealGen Preferred (%)81.891.797.893.3100.0
RealGen Score (0-5)4.27±1.054.27±0.693.96±1.944.17±1.933.94±2.27
LCTGen Score (0-5)2.15±1.442.08±1.192.42±1.952.0±1.962.14±2.31
", + "image_path": "ba309a4592db3904aa2298c1cfefddd52ff94edfda753a95e5220d70202aee9b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 480, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 480, + 255 + ], + "type": "text", + "content": "the point of collision. To quantitatively assess the performance of safety-critical scenarios, we compare RealGen with two baselines. RealGen-AE-R random samples behavior embeddings in the Realgen-AE model, and RealGen-R random retrieves behavior embeddings. According to the results in Table 3, we find that the scenarios generated by RealGen using crash scenarios as templates achieve the highest collision rate, which means that RealGen has more efficiency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 255, + 481, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 481, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 481, + 327 + ], + "type": "text", + "content": "Human evaluation of controllability. As there is no automatic way to evaluate controllability, we follow the protocol in [49] to perform A/B testing using human evaluation. We report the ratio of our method preferred as well as the absolute score (0-5) of both our method and the baseline. In Table 4, we find that scenarios generated by RealGen are highly preferred in most categories. The absolute score of RealGen is also much higher than that of LCTGen." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 328, + 274, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 274, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 274, + 387 + ], + "type": "text", + "content": "Downstream task evaluation. A direct downstream task of our method is to use the generated data to augment the training dataset of trajectory" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 282, + 342, + 480, + 380 + ], + "blocks": [ + { + "bbox": [ + 301, + 331, + 460, + 341 + ], + "lines": [ + { + "bbox": [ + 301, + 331, + 460, + 341 + ], + "spans": [ + { + "bbox": [ + 301, + 331, + 460, + 341 + ], + "type": "text", + "content": "Table 5: Downstream task evaluation." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 282, + 342, + 480, + 380 + ], + "lines": [ + { + "bbox": [ + 282, + 342, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 282, + 342, + 480, + 380 + ], + "type": "table", + "html": "
MethodOriginalRandom Aug.RealGen Aug.
mADE (↓)3.5442.9202.309
collision rate (↓)0.0490.0370.018
", + "image_path": "64e9fd20cd57f4e88c96916a0fd95c7279797f0194186514f8f2e7fd77cde6f7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 387, + 481, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 481, + 459 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 481, + 459 + ], + "type": "text", + "content": "prediction models. We use Autobots [21] as a predictor and report the results trained on different datasets in Table 5. Original means using the original data in nuScenes [5], Random Aug. means augmenting the original dataset with Gaussian noise, and RealGen Aug. means augmenting the original dataset with scenarios from RealGen. We observe that the model trained with the RealGen dataset achieves the lowest mADE and collision rate." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 468, + 312, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 468, + 312, + 480 + ], + "spans": [ + { + "bbox": [ + 131, + 468, + 312, + 480 + ], + "type": "text", + "content": "6 Conclusion and Limitation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 486, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 481, + 605 + ], + "type": "text", + "content": "This paper proposes RealGen, a novel framework for traffic scenario generation that utilizes retrieval-augmented generation. Unlike previous approaches, which primarily rely on models replicating training distributions, RealGen demonstrates in-context learning abilities that synthesize scenarios by combining and modifying provided examples, enabling controlled generation. These scenarios can be automatically obtained from a retrieval system, which only requires the users to provide a few template scenarios as examples. We comprehensively evaluated the similarity autoencoder model for retrieval and the combiner model for generation. The findings indicate that RealGen achieves low reconstruction error and high generation quality." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "The primary limitation of our current method is that the behavior encoder focuses solely on agent trajectories, neglecting the intricate interactions between agents and lane maps in complex behaviors. Advancing the feature representation within the behavior encoder could significantly broaden RealGen's capacity to generate diverse and controllable traffic scenarios." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "W. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 481, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 481, + 174 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 481, + 174 + ], + "type": "text", + "content": "Wenhao Ding contributed to this paper while being an intern at NVIDIA Research. Ding Zhao was partially supported by the National Science Foundation under grant CNS-2047454." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 190, + 197, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 190, + 197, + 202 + ], + "spans": [ + { + "bbox": [ + 133, + 190, + 197, + 202 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 211, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 211, + 480, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 211, + 480, + 233 + ], + "spans": [ + { + "bbox": [ + 138, + 211, + 480, + 233 + ], + "type": "text", + "content": "1. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 234, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 234, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 234, + 480, + 266 + ], + "type": "text", + "content": "2. Blattmann, A., Rombach, R., Oktay, K., Müller, J., Ommer, B.: Retrievalaugmented diffusion models. Advances in Neural Information Processing Systems 35, 15309-15324 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 266, + 480, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 480, + 309 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 480, + 309 + ], + "type": "text", + "content": "3. Borgeaud, S., Mensch, A., Hoffmann, J., Cai, T., Rutherford, E., Millican, K., Van Den Driessche, G.B., Lespiau, J.B., Damoc, B., Clark, A., et al.: Improving language models by retrieving from trillions of tokens. In: International conference on machine learning. pp. 2206-2240. PMLR (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 310, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 310, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 310, + 480, + 342 + ], + "type": "text", + "content": "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 343, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 480, + 385 + ], + "type": "text", + "content": "5. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 386, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 386, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 138, + 386, + 480, + 439 + ], + "type": "text", + "content": "6. Cai, D., Wang, Y., Bi, W., Tu, Z., Liu, X., Shi, S.: Retrieval-guided dialogue response generation via a matching-to-generation framework. In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). pp. 1866-1875 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 440, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 440, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 138, + 440, + 480, + 472 + ], + "type": "text", + "content": "7. Cao, Y., Xiao, C., Anandkumar, A., Xu, D., Pavone, M.: Advdo: Realistic adversarial attacks for trajectory prediction. In: European Conference on Computer Vision. pp. 36-52. Springer (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 472, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 472, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 480, + 493 + ], + "type": "text", + "content": "8. Chen, D., Fisch, A., Weston, J., Bordes, A.: Reading wikipedia to answer open-domain questions. arXiv preprint arXiv:1704.00051 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 494, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 494, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 494, + 480, + 525 + ], + "type": "text", + "content": "9. Chen, H., Wang, J., Shao, K., Liu, F., Hao, J., Guan, C., Chen, G., Heng, P.A.: Traj-mae: Masked autoencoders for trajectory prediction. arXiv preprint arXiv:2303.06697 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 525, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 525, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 134, + 525, + 480, + 557 + ], + "type": "text", + "content": "10. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 558, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 558, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 134, + 558, + 480, + 579 + ], + "type": "text", + "content": "11. Chen, W., Hu, H., Sahara, C., Cohen, W.W.: Re-imagen: Retrieval-augmented text-to-image generator. arXiv preprint arXiv:2209.14491 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 580, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 580, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 134, + 580, + 480, + 611 + ], + "type": "text", + "content": "12. Cheng, J., Mei, X., Liu, M.: Forecast-mae: Self-supervised pre-training for motion forecasting with masked autoencoders. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8679-8689 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 612, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 612, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 612, + 480, + 632 + ], + "type": "text", + "content": "13. Cuturi, M.: Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems 26 (2013)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "text", + "content": "14. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "type": "text", + "content": "RealGen: Retrieval Augmented Generation for Controllable Traffic Scenarios" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "15. Ding, W., Chen, B., Li, B., Eun, K.J., Zhao, D.: Multimodal safety-critical scenarios generation for decision-making algorithms evaluation. IEEE Robotics and Automation Letters 6(2), 1551-1558 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "text", + "content": "16. Ding, W., Lin, H., Li, B., Eun, K.J., Zhao, D.: Semantically adversarial driving scenario generation with explicit knowledge integration. arXiv preprint arXiv:2106.04066 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "type": "text", + "content": "17. Ding, W., Xu, M., Zhao, D.: Learning to collide: An adaptive safety-critical scenarios generating method. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 251 + ], + "type": "text", + "content": "18. Feng, L., Li, Q., Peng, Z., Tan, S., Zhou, B.: Trafficgen: Learning to generate diverse and realistic traffic scenarios. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 3567-3575. IEEE (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 251, + 481, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 481, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 481, + 295 + ], + "type": "text", + "content": "19. Feydy, J., Séjourne, T., Vialard, F.X., Amari, S.i., Trouve, A., Peyré, G.: Interpolating between optimal transport and mmd using sinkhorn divergences. In: The 22nd International Conference on Artificial Intelligence and Statistics. pp. 2681-2690 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 296, + 481, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 296, + 481, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 296, + 481, + 318 + ], + "type": "text", + "content": "20. Gidaris, S., Singh, P., Komodakis, N.: Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 319, + 481, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 481, + 351 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 481, + 351 + ], + "type": "text", + "content": "21. Girgis, R., Golemo, F., Codevilla, F., Weiss, M., D'Souza, J.A., Kahou, S.E., Heide, F., Pal, C.: Latent variable sequential set transformers for joint multi-agent motion prediction. arXiv preprint arXiv:2104.00563 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 352, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 481, + 373 + ], + "type": "text", + "content": "22. Gretton, A., Borgwardt, K.M., Rasch, M.J., Scholkopf, B., Smola, A.: A kernel two-sample test. The Journal of Machine Learning Research 13(1), 723-773 (2012)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "type": "text", + "content": "23. Gulino, C., Fu, J., Luo, W., Tucker, G., Bronstein, E., Lu, Y., Harb, J., Pan, X., Wang, Y., Chen, X., et al.: Waymax: An accelerated, data-driven simulator for large-scale autonomous driving research. arXiv preprint arXiv:2310.08710 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 408, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 481, + 453 + ], + "type": "text", + "content": "24. Gutmann, M., Hyvarinen, A.: Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In: Proceedings of the thirteenth international conference on artificial intelligence and statistics. pp. 297-304. JMLR Workshop and Conference Proceedings (2010)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 453, + 481, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 481, + 485 + ], + "type": "text", + "content": "25. Guu, K., Lee, K., Tung, Z., Pasupat, P., Chang, M.: Retrieval augmented language model pre-training. In: International conference on machine learning. pp. 3929-3938. PMLR (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 486, + 481, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 481, + 518 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 481, + 518 + ], + "type": "text", + "content": "26. Hanselmann, N., Renz, K., Chitta, K., Bhattacharyya, A., Geiger, A.: King: Generating safety-critical driving scenarios for robust imitation via kinematics gradients. In: European Conference on Computer Vision. pp. 335-352. Springer (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 520, + 481, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 481, + 552 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 481, + 552 + ], + "type": "text", + "content": "27. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 554, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 481, + 586 + ], + "type": "text", + "content": "28. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 587, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 481, + 609 + ], + "type": "text", + "content": "29. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 609, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 481, + 632 + ], + "type": "text", + "content": "30. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural computation 9(8), 1735-1780 (1997)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "31. Hospedales, T., Antoniou, A., Micaelli, P., Storkey, A.: Meta-learning in neural networks: A survey. IEEE transactions on pattern analysis and machine intelligence 44(9), 5149-5169 (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "W. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "32. Hu, A., Corrado, G., Griffiths, N., Murez, Z., Gurau, C., Yeo, H., Kendall, A., Cipolla, R., Shotton, J.: Model-based imitation learning for urban driving. Advances in Neural Information Processing Systems 35, 20703-20716 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "text", + "content": "33. Ibrihich, S., Oussous, A., Ibrihich, O., Esghir, M.: A review on recent research in information retrieval. Procedia Computer Science 201, 777-782 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 480, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 480, + 192 + ], + "type": "text", + "content": "34. Ivanovic, B., Song, G., Gilitschenski, I., Pavone, M.: trajdata: A unified interface to multiple human trajectory datasets. arXiv preprint arXiv:2307.13924 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "type": "text", + "content": "35. Kim, J., Choi, S., Amplayo, R.K., Hwang, S.w.: Retrieval-augmented controllable review generation. In: Proceedings of the 28th International Conference on Computational Linguistics. pp. 2284-2295 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "type": "text", + "content": "36. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 246, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 480, + 289 + ], + "type": "text", + "content": "37. Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Kuttler, H., Lewis, M., Yih, W.t., Rocktäschel, T., et al.: Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems 33, 9459-9474 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 289, + 480, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 322 + ], + "type": "text", + "content": "38. Li, Q., Peng, Z., Feng, L., Duan, C., Mo, W., Zhou, B., et al.: Scenarionet: Open-source platform for large-scale traffic scenario simulation and modeling. arXiv preprint arXiv:2306.12241 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 322, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 480, + 354 + ], + "type": "text", + "content": "39. Liu, Q., Yogatama, D., Blunsom, P.: Relational memory-augmented language models. Transactions of the Association for Computational Linguistics 10, 555-572 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 354, + 460, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 460, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 460, + 364 + ], + "type": "text", + "content": "40. NHTSA: Nhtsa crash viewer (2023), https://crashviewer.nhtsa.dot.gov/" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 365, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 396 + ], + "type": "text", + "content": "41. Noroozi, M., Favaro, P.: Unsupervised learning of visual representations by solving jigsaw puzzles. In: European conference on computer vision. pp. 69-84. Springer (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "type": "text", + "content": "42. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 418, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 449 + ], + "type": "text", + "content": "43. Roshdi, A., Roohparvar, A.: Information retrieval techniques and applications. International Journal of Computer Networks and Communications Security 3(9), 373-377 (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 449, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 480, + 471 + ], + "type": "text", + "content": "44. Schmidhuber, J.: Deep learning in neural networks: An overview. Neural networks 61, 85-117 (2015)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 471, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 480, + 503 + ], + "type": "text", + "content": "45. Shi, S., Jiang, L., Dai, D., Schiele, B.: Motion transformer with global intention localization and local movement refinement. Advances in Neural Information Processing Systems 35, 6531-6543 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 503, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 503, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 503, + 480, + 525 + ], + "type": "text", + "content": "46. Shurrab, S., Duwairi, R.: Self-supervised learning methods and applications in medical imaging analysis: A survey. PeerJ Computer Science 8, e1045 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 525, + 480, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 556 + ], + "type": "text", + "content": "47. Suo, S., Regalado, S., Casas, S., Urtasun, R.: Trafficsim: Learning to simulate realistic multi-agent behaviors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10400-10409 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 556, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 480, + 600 + ], + "type": "text", + "content": "48. Suo, S., Wong, K., Xu, J., Tu, J., Cui, A., Casas, S., Urtasun, R.: Mixsim: A hierarchical framework for mixed reality traffic simulation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9622-9631 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "text", + "content": "49. Tan, S., Ivanovic, B., Weng, X., Pavone, M., Kraehenbuehl, P.: Language conditioned traffic generation. In: Conference on Robot Learning. pp. 2714-2752. PMLR (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "50. Tan, S., Wong, K., Wang, S., Manivasagam, S., Ren, M., Urtasun, R.: Scenegen: Learning to generate realistic traffic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 892-901 (2021)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 136, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 136, + 91, + 447, + 102 + ], + "type": "text", + "content": "RealGen: Retrieval Augmented Generation for Controllable Traffic Scenarios" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 544 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 149 + ], + "type": "text", + "content": "51. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 459, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 459, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 459, + 160 + ], + "type": "text", + "content": "52. Villani, C., et al.: Optimal transport: old and new, vol. 338. Springer (2009)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "text", + "content": "53. Vincent, P., Larochelle, H., Bengio, Y., Manzagol, P.A.: Extracting and composing robust features with denoising autoencoders. In: Proceedings of the 25th international conference on Machine learning. pp. 1096-1103 (2008)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 237 + ], + "type": "text", + "content": "54. Wang, J., Pun, A., Tu, J., Manivasagam, S., Sadat, A., Casas, S., Ren, M., Urtasun, R.: Advsim: Generating safety-critical scenarios for self-driving vehicles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9909-9918 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 259 + ], + "type": "text", + "content": "55. Wang, Z., Nie, W., Qiao, Z., Xiao, C., Baraniuk, R., Anandkumar, A.: Retrieval-based controllable molecule generation. arXiv preprint arXiv:2208.11126 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 260, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 481, + 293 + ], + "type": "text", + "content": "56. Wu, P., Majumdar, A., Stone, K., Lin, Y., Mordatch, I., Abbeel, P., Rajeswaran, A.: Masked trajectory models for prediction, representation, and control. arXiv preprint arXiv:2305.02968 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "text", + "content": "57. Xu, P., Patwary, M., Shoeybi, M., Puri, R., Fung, P., Anandkumar, A., Catanzaro, B.: Megatron-cntrl: Controllable story generation with external knowledge using large-scale language models. arXiv preprint arXiv:2010.00840 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "type": "text", + "content": "58. Yang, Y., Zhang, Q., Gilles, T., Batool, N., Folkesson, J.: Rmp: A random mask pretrain framework for motion prediction. arXiv preprint arXiv:2309.08989 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "type": "text", + "content": "59. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of lmms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "60. Zhang, C., Zhang, C., Song, J., Yi, J.S.K., Zhang, K., Kweon, I.S.: A survey on masked autoencoder for self-supervised learning in vision and beyond. arXiv preprint arXiv:2208.00173 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 414, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 435 + ], + "type": "text", + "content": "61. Zhang, C., Tu, J., Zhang, L., Wong, K., Suo, S., Urtasun, R.: Learning realistic traffic agents in closed-loop. In: 7th Annual Conference on Robot Learning (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 436, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 436, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 481, + 468 + ], + "type": "text", + "content": "62. Zhang, X., Bosselut, A., Yasunaga, M., Ren, H., Liang, P., Manning, C.D., Leskovec, J.: Greaselm: Graph reasoning enhanced language models. In: International conference on learning representations (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "type": "text", + "content": "63. Zhong, Z., Rempe, D., Chen, Y., Ivanovic, B., Cao, Y., Xu, D., Pavone, M., Ray, B.: Language-guided traffic simulation via scene-level diffusion. arXiv preprint arXiv:2306.06344 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 501, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 544 + ], + "type": "text", + "content": "64. Zhong, Z., Rempe, D., Xu, D., Chen, Y., Veer, S., Che, T., Ray, B., Pavone, M.: Guided conditional diffusion for controllable traffic simulation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 3560-3566. IEEE (2023)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "W. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/RealViformer_ Investigating Attention for Real-World Video Super-Resolution/39444a96-c968-4d03-9b6c-b416cbbaab18_content_list.json b/2024/RealViformer_ Investigating Attention for Real-World Video Super-Resolution/39444a96-c968-4d03-9b6c-b416cbbaab18_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5fa1f5e5a0838e2c2a81c8bd6c206d4c621cee --- /dev/null +++ b/2024/RealViformer_ Investigating Attention for Real-World Video Super-Resolution/39444a96-c968-4d03-9b6c-b416cbbaab18_content_list.json @@ -0,0 +1,1720 @@ +[ + { + "type": "text", + "text": "RealVformer: Investigating Attention for Real-World Video Super-Resolution", + "text_level": 1, + "bbox": [ + 264, + 140, + 743, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuehan Zhang and Angela Yao", + "bbox": [ + 375, + 212, + 625, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "National University of Singapore {zyuehan,ayao}@comp.nus.edu.sg", + "bbox": [ + 383, + 239, + 620, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In real-world video super-resolution (VSR), videos suffer from in-the-wild degradations and artifacts. VSR methods, especially recurrent ones, tend to propagate artifacts over time in the real-world setting and are more vulnerable than image super-resolution. This paper investigates the influence of artifacts on commonly used covariance-based attention mechanisms in VSR. Comparing the widely-used spatial attention, which computes covariance over space, versus the channel attention, we observe that the latter is less sensitive to artifacts. However, channel attention leads to feature redundancy, as evidenced by the higher covariance among output channels. As such, we explore simple techniques such as the squeeze-excite mechanism and covariance-based rescaling to counter the effects of high channel covariance. Based on our findings, we propose RealVformer. This channel-attention-based real-world VSR framework surpasses state-of-the-art on two real-world VSR datasets with fewer parameters and faster runtimes. The source code is available at https://github.com/Yuehan717/RealVformer.", + "bbox": [ + 261, + 304, + 743, + 526 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 551, + 375, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video super-resolution (VSR) recovers a high-resolution (HR) sequence of frames from its low-resolution (LR) counterpart. Recurrent convolutional approaches are commonly used in VSR with standard settings, assuming the LR frames are downsampled from HR frames with known kernels. [3, 10, 15]. However, in real-world VSR (RWVSR), the low-resolution videos are not simply downsampled versions of their high-resolution counterparts. Instead, they feature complex degradations that arise from the camera imaging system, compression, internet transmission, and other factors. These degradations make architecture design for RWVSR challenging, as artifacts and degradations tend to propagate and get exaggerated over the recurrent connection [5, 36].", + "bbox": [ + 212, + 582, + 787, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, transformer architectures have replaced convolutional architectures as state-of-the-art for standard VSR. While attention mechanisms have replaced convolution operations, most methods retain the recurrent connection to aggregate the information over time [22, 30]. Yet, such architectures do not always perform well on RWVSR. For instance, a Swin-based model [21] designed for standard VSR, when applied to a real-world input frame, as shown in Fig. 1a, generates more artificial lines than the convolutional model RealBasicVSR [5].", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/110ffaa2f809c8fbaaf8966a162e998bca4cf29b8f6bfdc7e6bbbdf0a92444b6.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 272, + 142, + 364, + 195 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/456b3246abbab733088a604dbc47e275c35a887c21e6f92da2873f6033cbd83a.jpg", + "image_caption": [ + "RealBasicVSR" + ], + "image_footnote": [], + "bbox": [ + 367, + 143, + 460, + 195 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/84f5cc884e9e4d8e0b056fc7d3d08be8a7ec48437a2e83bf577871e3316a5ea4.jpg", + "image_caption": [ + "Swin-based VSR" + ], + "image_footnote": [], + "bbox": [ + 272, + 212, + 364, + 265 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bf3b02d4f19ffbccf955eee78c521b1bbb179ef40d04958aeeb88a13715baad0.jpg", + "image_caption": [ + "Ours: RealVformer" + ], + "image_footnote": [], + "bbox": [ + 367, + 212, + 460, + 265 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/63c21eeb4eedbfec929404f0e2ce863925a932efc469e60750bb571e63ddffbb.jpg", + "image_caption": [ + "(a) Visual comparisons.", + "W", + "Spatial Attention" + ], + "image_footnote": [], + "bbox": [ + 491, + 155, + 614, + 243 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/13508f59141d52d449e745094a9abb701d287e42aa305cbada4bb50442108113.jpg", + "image_caption": [ + "H", + "Channel Attention", + "(b) Schematics of two attentions.", + "Fig. 1: (a) Designing a RWVSR transformer is not trivial. A Swin-based transformer suited for standard VSR hallucinates more lines than a RealBasicVSR, a convolutional state-of-the-art. We propose RealVformer based on our investigation of attention under the RWVSR setting. RealVformer generates details with fewer artifacts than RealBasicVSR [3] and the Swin-based VSR model. (b) Schematic for spatial and channel attention. Spatial attention aggregates features based on pixel representations. Channel attention takes $H \\times W$ feature map for matching across channels." + ], + "image_footnote": [], + "bbox": [ + 617, + 155, + 743, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Why should transformers perform well on (synthetic) standard VSR but poorly in real-world cases? We speculate that standard VSR transformers benefit from the similarity-based matching of the attention mechanism, which accurately aggregates information spatially and temporally [21,22,30]. However, when input degradation exists, the aggregated information becomes less reliable because the attention queries may be derived from both true source video and artifacts.", + "bbox": [ + 212, + 449, + 784, + 541 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This work investigates and sheds light on the sensitivity of attention in real-world settings. We compare two covariance-based attention mechanisms used in low-level Transformers: spatial attention [21,30] and channel attention [39]. Spatial attention takes pixel-wise features as keys and queries and estimates their covariances across spatial positions. The most popular form is window-based attention [6]; The shift-window scheme from Swin Transformers [23] enables the model to access distant spatial ranges [2,21,30] without computational blow-up. Spatial attention is widely used for video and image super-resolution, albeit in the standard setting. Channel attention [39] estimates covariances across channels (see Fig. 1b) and collapses the spatial extent of a feature map. It defines the number of queries and keys by channel numbers rather than spatial resolution. Although established for deblurring or denoising [39], channel attention's efficacy in super-resolution remains to be determined.", + "bbox": [ + 212, + 546, + 784, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our experiments show that channel attention is less sensitive to artifacts than the spatial counterparts, resulting in higher performance gains in RWVSR. However, it is also revealed that channel attention leads to feature channels with higher covariance. From a learning perspective, a high covariance is undesirable because it is a strong indicator of feature redundancy [1,8,13]. Therefore, using channel attention naively will have limited improvements over existing RWVSR", + "bbox": [ + 212, + 750, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Zhang and A. Yao", + "bbox": [ + 271, + 114, + 416, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "state-of-the-art methods. Such a finding has wide-reaching impacts as channel attention is used increasingly for low-level vision.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To verify our findings, we explore established mechanisms to counter the effects of feature redundancy - simple techniques such as squeeze-excite and covariance-based rescaling improve the vanilla channel attention design. From these outcomes, we propose RealVformer, a new transformer real-world VSR model. RealVformer performs channel attention between the current frame feature and the propagated hidden state to limit model-produced artifacts. The model then reconstructs features through improved channel attention modules featuring squeeze-and-excite and covariance-based channel rescaling mechanisms. With our effective designs, RealVformer achieves state-of-the-art performance with fewer parameters on challenging synthetic video datasets and two real-world video datasets collected from different scenes.", + "bbox": [ + 212, + 176, + 787, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Summarizing our contributions in order of importance, our paper", + "bbox": [ + 238, + 344, + 712, + 358 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- investigates the differences between spatial and channel attention for RWVSR. Spatial attention, although widely used, is revealed to be highly sensitive to the noise and degradations common in RWVSR sequences, while channel attention is more robust.", + "- reveals that naively applying channel attention increases channel covariance, which is problematic for learning; this overlooked fact has a wide-reaching impact as channel attention becomes more used in low-level vision.", + "- empirically verifies the negative effect of high channel covariance by countering it with established techniques, based on which we develop the RealVformer for RWVSR. Our simple modification surpasses state-of-the-art despite using less compute." + ], + "bbox": [ + 225, + 369, + 790, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 563, + 387, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Standard Video Super-Resolution models focus on architecture design to use temporal information better [2-4, 15]. Previous research starts from sliding-window-based [15,33] to recurrent-based frameworks [3,4] for using distant-frame information. Recent works introduce Transformer blocks into existing recurrent frameworks to overcome the locality limitation of convolution and accurately match abundant information for feature reconstruction [20,22,30].", + "bbox": [ + 212, + 597, + 787, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Real-world video super-resolution focuses on modeling, removing, and limiting the impact of real-world degradations. Existing works are convolutional models and focus on designing losses or modules for degradation processing. DBVSR [28] explicitly estimates the degradation kernel through a sub-network. RealBasicVSR [5] tries to 'clean' artifacts through a processing module for BasicVSR [3]. In a similar approach, FastRealVSR [36] borrows an external pool of blur and sharpening filters to 'clean' the hidden states. Other recent works [16,31] advance the synthesis method for paired training data. Instead, we focus on investigating the function of attention in RWVSR rather than architecture design or data synthesis.", + "bbox": [ + 212, + 688, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "RealViformer", + "bbox": [ + 638, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Attention mechanisms have been widely applied for low-level vision tasks [9, 25, 43, 43]. Transformers with covariance-based attention are the most prevalent [21, 30, 39] for standard VSR. Most existing methods adopt shift-window-based spatial attention [23] to aggregate information from other positions within or across frames. In contrast, Restormer [39] computes the covariance among channels and shows its effectiveness for multiple image restoration tasks. More recent works stitch spatial and channel attention together to enlarge the receptive field [7, 32] for standard image super-resolution. Instead, we investigate attention mechanisms in terms of their sensitivity to real-world degradation for the first time and develop an effective real-world VSR Transformer.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Explorations on Attention", + "text_level": 1, + "bbox": [ + 215, + 319, + 504, + 337 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sec. 3.1 defines the VSR task and the two attention mechanisms. Sec. 3.2 compares the channel and spatial attentions' sensitivity to query artifacts and effects on real-world VSR performance. Sec. 3.3 reveals that channel attention leads to higher covariance among channels and explores mitigating options.", + "bbox": [ + 212, + 349, + 787, + 411 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 430, + 372, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Standard vs. Real-World VSR. Given a low-resolution (LR) video sequence with $T$ frames $I^L \\in \\mathcal{R}^{T \\times H \\times W \\times K}$ , VSR models reconstruct a high-resolution sequence $I^H \\in \\mathcal{R}^{T \\times sH \\times sW \\times K}$ , where $H \\times W$ is the input spatial resolution, $K$ is the number of input channels and $s$ is the scaling factor. In the standard setting of VSR, $I_t^L$ , where $t \\in \\{0, \\dots, T\\}$ , is assumed as the downsampled version of $I_t^H$ , defined as: $I_t^L = (I_t^H) \\downarrow_{\\frac{1}{s}}$ . Both training and testing datasets follow this formulation to generate LR-HR pairs given $I^H$ . In a real-world setting, there is no closed formulation for the relationship between $I_t^L$ and $I_t^H$ due to the unknown distribution of real-world degradations. Real-ESRGAN [34] proposed a widely-used training setting that randomly applies synthesized blur, noise, compression, and resizing to the HR frames to generate paired LR frames with complex degradations. The testing datasets are either synthesized by the same pipeline in training or collected from diverse real-world sources [5, 37]. The synthesized testing sets have paired ground-truth sequences and are evaluated by full-reference metrics, e.g. PSNR and LPIPS [42]; real-world datasets are always without ground truth and require no-reference metrics, such as NRQM [24].", + "bbox": [ + 212, + 453, + 787, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attention Definitions. In Transformers, the attention modules project layer-normalized tensor $X \\in \\mathbb{R}^{C \\times H \\times W}$ to query $Q$ , and tensor $Y \\in \\mathbb{R}^{\\hat{C} \\times \\hat{H} \\times \\hat{W}}$ to key $K$ and value $V$ , where $H \\times W$ and $\\hat{H} \\times \\hat{W}$ are the spatial resolution of the normalized tensors. The attention map $\\mathcal{A}$ is generated by calculating the covariance between $Q$ and $K$ , followed by a softmax function, before being applied to the value $V$ to produce output $O$ . Spatial and channel attention differ in tensor dimension taken for the covariance calculation.", + "bbox": [ + 212, + 699, + 787, + 806 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Zhang and A. Yao", + "bbox": [ + 271, + 114, + 415, + 127 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1 Note we define mutual attention as the default; self-attention is a special form of mutual attention where $Y = X$ .", + "bbox": [ + 217, + 810, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Spatial attention generates an $\\mathbb{R}^{HW\\times \\hat{H}\\hat{W}}$ attention map, $\\mathcal{A}_s$ , by computing the covariance between features in the query $(Q_{s})$ and key $(K_{s})$ at each spatial position. The query, key and values are computed as $Q_{s} = W_{s}^{Q}X$ , $K_{s} = W_{s}^{K}Y$ , $V_{s} = W_{s}^{V}Y$ , where $W_{s}^{Q}\\in \\mathbb{R}^{D_{s}\\times C}$ , $\\{W_s^K,W_s^V\\} \\in \\mathbb{R}^{D_s\\times \\hat{C}}$ , and $D_{s}$ is the dimension of projections. The attention map $\\mathcal{A}_s$ and output attention features $O_{s}$ are:", + "bbox": [ + 214, + 145, + 787, + 224 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} _ {s} = \\operatorname {s o f t m a x} \\left(Q _ {s} ^ {T} K _ {s} / \\sqrt {D _ {s}}\\right), O _ {s} = \\mathcal {A} _ {s} V _ {s} ^ {T}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 351, + 232, + 785, + 251 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Q_{s}(K_{s})$ is reshaped to matrix of $\\mathbb{R}^{D_s\\times HW}$ $(\\mathbb{R}^{D_s\\times \\hat{H}\\hat{W}})$ . We omit the later reshaping operation for brevity. $\\mathcal{A}_s$ exhaustively relates every spatial location to every other location in space. A window-based implementation [21] limits the correlations to windows of $\\omega \\times \\omega$ , reducing the map size to $\\mathbb{R}^{\\omega^2\\times \\omega^2}$ .", + "bbox": [ + 212, + 258, + 782, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Channel attention applies convolutions on $X$ to get query $(Q_{c})$ and $Y$ for key $(K_{c})$ and value $(V_{c})$ [39], after which features are the same in spatial resolution (assumed to be $H \\times W$ in the following context). It estimates covariance across channels to yield a $C \\times \\hat{C}$ sized map $\\mathcal{A}_{c}$ and output features $O_{c}$ :", + "bbox": [ + 214, + 321, + 782, + 383 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} _ {c} = \\operatorname {s o f t m a x} \\left(Q _ {c} K _ {c} ^ {T} / \\alpha\\right), O _ {c} = \\mathcal {A} _ {c} V _ {c}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 392, + 785, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Q_{c}$ is reshaped to $\\mathbb{R}^{C\\times HW}$ , $K_{c}$ and $V_{c}$ are reshaped to $\\mathbb{R}^{\\hat{C}\\times HW}$ and $\\alpha$ is learnable scaling parameter. Because channel attention computes correlations between features of size $\\mathbb{R}^{1\\times HW}$ , it has a larger spatial context than the $\\mathbb{R}^{D_s\\times 1}$ -sized features in spatial attention (see Fig. 1b).", + "bbox": [ + 212, + 417, + 784, + 481 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Attention in Real-World VSR", + "text_level": 1, + "bbox": [ + 214, + 500, + 508, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sensitivity to Query Artifacts. In standard VSR, attention is used to aggregate information temporally and spatially [22, 30]. The assumption is that queries can match beneficial cues for super-resolution. But what if the queries themselves are unreliable? We speculate this to be true in real-world VSR, where inputs have artifacts and degradations.", + "bbox": [ + 212, + 522, + 782, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "How affected are spatial and channel attention mechanisms by query artifacts? Fig. 2 shows how we compare, using cosine similarity, the attention outputs based on queries from the same frame with and without certain degradations. Using encoding layers from a standard convolutional VSR model, we perform the attention operations defined in Sec. 3.1 by taking embeddings of frame $I_{t}$ as query and $I_{t-1}$ as key and value. $I_{t}$ and $I_{t-1}$ are clean without degradation other than downsampling, and we represent the output feature of the attention module as $O$ . When additional degradation $D_{i}$ is applied to $I_{t}$ , we represent the corresponding output as $O_{D_{i}}$ . A smaller deviation of $O_{D_{i}}$ from $O$ indicates lower sensitivity to artifacts in the query.", + "bbox": [ + 212, + 598, + 784, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tab. 1 shows the cosine similarity between $O$ and $O_{D_i}$ for spatial and channel attention modules. Curiously, $O_{D_i}$ based on the channel attention module is more similar to the $O$ matched by the degradation-free query, indicating that channel attention is less sensitive to query artifacts. Intuitively, the lower sensitivity of channel attention is related to the larger spatial context used for feature matching. Given a deep feature of size $\\mathbb{R}^{C\\times H\\times W}$ , channel attention uses feature sized", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "RealViformer", + "bbox": [ + 638, + 114, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/304df22f113c50b298215525302e7655239221a7f78137f645f39c0b2edcf9c3.jpg", + "image_caption": [ + "Fig.2: Schematic for sensitivity comparison. $I_{t-1}, I_t$ are downsampled but clean frames at times $t$ and $t-1$ . $D_i(.)$ apply degradations to $I_t$ , where $D_i \\in \\{\\text{blur, noise, compression}\\}$ . $O$ and $O_{D_i}$ are output features of the attention module. Queries are from the embedding at time $t$ , and keys and values are from time $t-1$ . Higher cosine similarities $S$ between attention output features $O$ and $O_{D_i}$ reflect less sensitivity to artifacts in queries." + ], + "image_footnote": [], + "bbox": [ + 362, + 142, + 638, + 247 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a72e11bdc6f0f7b6efd9979c0ebe5205839e5adcb40ca158be758d001306ac2a.jpg", + "table_caption": [ + "Table 1: Cosine similarity between $O$ and ${O}_{{D}_{i}}$ ,attention outputs without and with query degradation. Outputs of channel attention change less under query degradation." + ], + "table_footnote": [], + "table_body": "
ModuleDi
blurnoisecompression
Spatial attention0.750.920.84
Channel attention0.980.990.99
", + "bbox": [ + 357, + 398, + 640, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\mathbb{R}^{1 \\times HW}$ to calculate the covariance across channels. As such, feature aggregation is based on global information observed in a large normalized spatial context. Instead, the covariance of spatial attention is for features at each location sized $\\mathbb{R}^{C \\times 1}$ , so it is likely more sensitive to local value changes from artifacts.", + "bbox": [ + 212, + 487, + 785, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Impact on VSR Models. While channel attention output features are less sensitive to query artifacts than spatial attention, how well do they fare for real-world VSR? We experiment by incorporating these attention variants into VSR models. We focus specifically on recurrent pipelines due to their popularity in standard and real-world VSR [5,22,36]. In recurrent pipelines, artifacts in hidden states may propagate over time, get exaggerated, and negatively influence overall performance [5, 36]. Therefore, we explicitly add the attention module between features of the current frame and the propagated hidden state. Ideally, the added attention will select relative information from hidden states for the current frame and reduce the propagation of model-produced artifacts.", + "bbox": [ + 212, + 550, + 787, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 3a shows the baseline model which concatenates $f_{t}$ , the shallow feature of at time $t$ , and $\\hat{h}_{t-1}$ , the spatially aligned hidden state at time $t-1$ . We experiment with two variants by using a channel-attention module $G_{c}$ or a spatial-attention module $G_{s}$ to replace the concatenation in the baseline. As shown in Fig. 3b, the query is generated by $f_{t}$ , and the key and value are predicted from $\\hat{h}_{t-1}$ . The attention output $O_{t}^{A}$ is concatenated with $f_{t}$ as the input for $\\mathcal{R}$ .", + "bbox": [ + 212, + 702, + 787, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All models are trained on the REDS dataset [27] with the random degradation pipeline from Real-ESRGAN [34]. We test on REDS4 [27] with different types and extents of degradations, including Gaussian blur, Gaussian noise, and JPEG", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Zhang and A. Yao", + "bbox": [ + 271, + 114, + 416, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/35e515ed0362ecb528cc6ce0b23a682ab670ee9ec95f37a8a860fa37e88d83df.jpg", + "image_caption": [ + "(a) Recurrent Baseline" + ], + "image_footnote": [], + "bbox": [ + 292, + 143, + 506, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/62244e9bb5b9c9de03ad155c4fe629c5dc01b194dcc6355fc3b5318a6e7cb16c.jpg", + "image_caption": [ + "(b) Attention" + ], + "image_footnote": [], + "bbox": [ + 568, + 151, + 767, + 239 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bed2f5adfd4a7b2449c7e08a60b2a56d3bcb866ffe0709cdd0166b87ecb4fb98.jpg", + "image_caption": [ + "Fig. 3: (a) The recurrent baseline in Sec. 3.2 has a shallow mapping module $\\mathcal{F}$ , reconstruction module $\\mathcal{R}$ , upsampling module $\\mathcal{U}$ and warping function $W$ . $W$ aligns the hidden state $h_{t-1}$ to feature at $t$ based on optical flow $s_{(t-1)\\rightarrow t}^{f}$ . All residual blocks are convolutional. The concatenation between $f_{t}$ and $\\hat{h}_{t-1}$ are replaced with the spatial or channel attention modules in (b) to compare the effect of attention. (b) The attention module first applies layer normalization to $f_{t}$ and $\\hat{h}_{t-1}$ and then performs channel or spatial attention according to Sec. 3.1. The output feature $O_{t}^{A}$ concatenated with $f_{t}$ is processed by the module $\\mathcal{R}$ in (a)." + ], + "image_footnote": [], + "bbox": [ + 218, + 402, + 401, + 521 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/897c696d2d587de5dcf9c83c5f2b51e8c89185ba5156b647fa03e5f671ae4d3f.jpg", + "image_caption": [ + "Fig. 4: Comparison of spatial and channel attention through impact on the performance of real-world VSR model. The Y-axis shows improvements compared to the convolutional baseline. A lower LPIPS score is better. The channel attention module is the best except for the PSNR score of highly blurred inputs." + ], + "image_footnote": [], + "bbox": [ + 403, + 402, + 581, + 521 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ca67627848bee6f9524c8382dc94841d5c12b544f459775e56c1e468838bbfac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 583, + 402, + 772, + 520 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "compression. Fig. 4 displays changes in PSNR and LPIPS compared with the baseline model. Trained on the same dataset and degradation setting, channel attention between temporal information yields better objective and perceptual reconstruction quality than spatial attention and baseline for noise and JPEG compression inputs. Channel attention still achieves better scores for blurred inputs, except for the PSNR performance of severely blurred inputs.", + "bbox": [ + 212, + 619, + 787, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Limitations of Channel Attention", + "text_level": 1, + "bbox": [ + 214, + 736, + 537, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results in Sec. 3.2 indicate that channel attention is better suited for propagating information over time. However, it also has an inherent flaw - the correlation among output channels will increase since each channel in the attention output is a weighted summation over channels of the value $V_{c}$ . We calculate the covariance following VICReg [1] for a quantitative measurement. Given deep", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "RealViformer", + "bbox": [ + 638, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "feature $z^n \\in \\mathbb{R}^{C \\times H \\times W}$ , where $n \\in \\{1, N\\}$ is the index of a sample, we reshape $z^n$ to $\\mathbb{R}^{C \\times HW}$ and define the covariance matrix over $Z = \\{z^1, \\dots, z^N\\}$ as:", + "bbox": [ + 215, + 145, + 784, + 176 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nC o v (Z) = \\frac {1}{N - 1} \\sum_ {n = 1} ^ {N} \\left(z _ {n} - \\bar {z}\\right) \\left(z _ {n} - \\bar {z}\\right) ^ {T}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 188, + 785, + 231 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\bar{z} = \\frac{1}{N}\\sum_{n=1}^{N}z_{n}$ . We then define the indicator for the covariance matrix as $ac(Z) = \\frac{1}{d}\\sum_{i\\neq j}|Cov(Z)|_{i,j}$ , i.e. the average of absolute off-diagonal coefficients of $Cov(Z)$ , where $d$ is the number of off-diagonal coefficients. Function $ac(Z)$ encodes the covariance among feature channels. Taking $O$ in Fig. 2 for comparison, $ac(O)$ with $N = 400$ based on channel attention is 0.87 and significantly higher than the input features ( $\\approx 0.15$ ). Instead, $ac(O)$ of spatial attention remains similar to the input features. Previous works on representation learning [1] and overfitting [8] propose that the high covariance of feature channels indicates redundancy and tends to hinder informative predictions.", + "bbox": [ + 212, + 242, + 782, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Similarly, we speculate the redundancy effects will hinder the prediction of HR outputs when adopting channel attention in building blocks. For a closer look, we investigate the standard VSR task, which is artifact-free. Specifically, we build $M_{c}$ and $M_{s}$ by replacing the Residual Block in Fig. 3a with channel attention blocks [39] and spatial attention blocks [21]. Models are trained on the REDS dataset without extra degradation. For evaluation, we choose SSIM [35], which focuses on structural information. The SSIM score of $M_{c}$ (0.8338) is lower than $M_{s}$ (0.8432); $ac(\\cdot)$ of the last features before the upsampling module is higher for channel attention than for spatial attention, i.e. 0.199 vs. 0.147.", + "bbox": [ + 212, + 380, + 784, + 516 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We adopt two simple modifications to channel attention and boost informative features to verify the redundancy effect empirically. Fig. 5 shows the Improved Channel Attention (ICA) module. Our approach features two key steps to enhance the quality. First, we use the squeeze-and-excite mechanism to predict new information. The features are squeezed channel-wise to extract meaningful", + "bbox": [ + 212, + 517, + 785, + 592 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1749f415e521704d711882a6df1461ef863dc7988a82718fb792f2f63de5a078.jpg", + "image_caption": [ + "Fig. 5: Improved Channel Attention Module (ICA), showing self-attention for simplicity. The 'squeeze' convolution compresses the number of input feature channels $X \\in \\mathbb{R}^{C \\times H \\times W}$ by ratio $r$ . The features are then rescaled by weights predicted from the $\\frac{C}{r} \\times \\frac{C}{r}$ attention map before being expanded by the 'excite' convolution back to the original number of input channels." + ], + "image_footnote": [], + "bbox": [ + 334, + 621, + 669, + 753 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Zhang and A. Yao", + "bbox": [ + 271, + 114, + 416, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e4aba276499b6b85f5db0256db27dfe9fbb641783a645a51bccee794f785f330.jpg", + "image_caption": [ + "(c) Reconstruction Module", + "Fig. 6: The framework of RealVformer. (a) Overview of RealVformer, following a unidirectional recurrent framework. The outputs of the Forward module are propagated to the next time step and upsampled by module $\\mathcal{U}$ to get HR frames. (b) Explanation of the Forward module in (a), where $W$ denotes the warping function. The reconstruction module $\\mathcal{R}$ takes current frame $I_t^L$ and warped hidden state $\\hat{h}_{t-1}$ as inputs. (c) Reconstruction module $\\mathcal{R}$ . The shallow feature of $I_t^L$ and $\\hat{h}_{t-1}$ are fused by CAF and then forwarded to Transformer blocks with U-shape connection [39]. Module GDFN follows Restormer [39]. Details of CAF and ICA modules are stated in Fig. 7 and Fig. 5." + ], + "image_footnote": [], + "bbox": [ + 328, + 140, + 669, + 314 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "information and then expanded into new channels based on the attention outputs. Secondly, we rescale channels in attention outputs by scalar weights predicted from the attention map. The attention map measures relationships across channels and encourages the associated convolutions in the 'excite' operation to yield more precise and discriminating features. Our designs are inspired by the SE Network [12] but with two key distinctions. First, we apply the squeeze-and-excite mechanism to generate new information, while SE Network uses it to help with scalar weight prediction. Second, the rescaling weights prediction takes the attention map as a cue rather than the naive pooling results of channels. In Sec. 4, we show the efficacy of new designs on the real-world VSR task and compare channel correlation with and without new designs in ablations.", + "bbox": [ + 212, + 484, + 787, + 652 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 RealVifromer", + "text_level": 1, + "bbox": [ + 215, + 675, + 383, + 690 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Overview", + "text_level": 1, + "bbox": [ + 215, + 708, + 339, + 722 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We apply our findings in Sec. 3.2 to the real-world VSR task as further support. To that end, we design RealVformer to incorporate channel attention as the basic processing module and the modification to boost informative features. Our emphasis here is not novel architecture design but a showcase of effectiveness brought by applying our findings in Sec. 3.2.", + "bbox": [ + 212, + 733, + 785, + 808 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "RealVformer is a recurrent Transformer network with channel attention modules. Fig. 6 shows the model architecture. To reduce the computational", + "bbox": [ + 214, + 810, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "RealViformer", + "bbox": [ + 638, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/71297bf3b076a0ea814a16d09bf0c2e24b20f50ea114312c4bb2e3709fbd802c.jpg", + "image_caption": [ + "Channel Attention Fusion module (CAF)", + "Fig. 7: Details of Channel Attention Fusion (CAF) module. CAF gets the query from current frame feature $f_{t}$ and {key, value} from hidden state $\\hat{h}_{t-1}$ . The attention output is concatenated with $f_{t}$ to process for module output $O_{t}$ ." + ], + "image_footnote": [ + "1x1 Norm 3x3 Depth-wise $\\mathbb{C}$ Concatenation Reshape $\\bigotimes$ Matrix multiplication Conv" + ], + "bbox": [ + 313, + 160, + 679, + 258 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "cost, we follow a widely used recurrent architecture [3] in a unidirectional setting. The model first estimates optical flow $s_{(t - 1)\\rightarrow t}^{f}$ from $I_{t - 1}^{L}$ to $I_{t}^{L}$ through Spynet [29] and warps previous hidden state $h_{t - 1}$ to current time step based on the flow. Frame $I_t^L$ and spatially aligned hidden state $\\hat{h}_{t - 1}$ are processed in the reconstruction module $\\mathcal{R}$ . The hidden state is updated with outputs of $\\mathcal{R}$ and further processed by the upsampling module $\\mathcal{U}$ to output high-resolution frames.", + "bbox": [ + 212, + 382, + 787, + 478 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The reconstruction module $\\mathcal{R}$ uses channel attention in two ways. First, the Channel Attention Fusion (CAF) module fuses the temporal information to limit the produced artifacts in the hidden state. CAF queries the aligned hidden state $\\hat{h}_{t-1}$ by the shallow feature $f_t$ . Secondly, we take the Improved Channel Attention Module (ICA) in Fig. 5 to build the Transformer blocks for better HR reconstructions.", + "bbox": [ + 212, + 481, + 789, + 571 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Channel Attention Fusion (CAF) Module", + "text_level": 1, + "bbox": [ + 214, + 611, + 606, + 628 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As explored in Sec. 3.2, adding channel attention promotes the real-world VSR performance compared to spatial attention and the simple concatenation baseline. Thus, we keep this design in our model and put it as the Channel Attention Fusion (CAF) module for temporal aggregation. Fig. 7 show details of how CAF perform channel attention between $f_{t}$ , the feature of the current frame, and $\\hat{h}_{t-1}$ , the aligned propagated hidden state. The query $(Q_{t})$ is generated as $Q_{t} = K_{3\\times 3} * \\text{LayerNorm}(f_{t})$ , where $K_{3\\times 3} *$ refers to $3 \\times 3$ convolution operation. Similarly, we process hidden state $\\hat{h}_{t-1}$ by $\\tilde{h}_{t-1} = K_{3\\times 3}^{d} * K_{1\\times 1} * \\text{LayerNorm}(\\hat{h}_{t-1})$ , where $K_{3\\times 3}^{d} *$ is the $3 \\times 3$ depth-wise convolution which double the channel number. Chunking $\\tilde{h}_{t-1}$ gives key $(K_{t})$ and value $(V_{t})$ . The calculation of attention map $A_{t} \\in \\mathbb{R}^{C \\times C}$ follows Eq. (2). The final output $O_{t}$ is computed as $O_{t} = K_{1\\times 1} * K_{3\\times 3}^{d} * K_{1\\times 1} * \\mathbf{C}[A_{t}V_{t}; f_{t}]$ , where $\\mathbf{C}[\\cdot; \\cdot]$ denotes concatenation.", + "bbox": [ + 212, + 655, + 787, + 843 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Zhang and A. Yao", + "bbox": [ + 271, + 114, + 416, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Improved Channel Attention (ICA) Module", + "text_level": 1, + "bbox": [ + 215, + 146, + 624, + 162 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The design of ICA follows Fig. 5 for empirically verifying findings in Sec. 3.3. Squeeze and Excite follows the squeeze-and-excite mechanism in Fig. 5 and helps to predict channels with new information. It squeezes the channels of input feature $X \\in \\mathbb{R}^{C \\times H \\times W}$ by factor $r$ . The attention map refers to self-attention on $X$ and of size $\\mathbb{R}^{\\frac{C}{r} \\times \\frac{C}{r}}$ . Excitation layers expand outputs back to size $\\mathbb{R}^{C \\times H \\times W}$ . Correlation-based channel weighting weights the output channel based on scalars predicted from the attention map. This design emphasizes channels for predicting more discriminative features in 'excite' operation. The attention map $A_r \\in \\mathbb{R}^{\\frac{C}{r} \\times \\frac{C}{r}}$ is taken for calculating the average and max values along the rows. The average and max values are combined and mapped to weights of size $\\mathbb{R}^{\\frac{C}{r} \\times 1}$ through linear layers and a sigmoid function.", + "bbox": [ + 212, + 171, + 787, + 339 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 359, + 455, + 375 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Model details. RealVformer uses SPyNet [29] for flow estimation. After the CAF module, the reconstruction module $\\mathcal{R}$ has a three-level encoder-decoder architecture. From level 1 to level 3, there are [2,3,4] transformer blocks with [48,96,192] channels. There are [1,2,4] attention heads in the ICA, all with a squeeze factor of 4. Supplementary B.1 gives the detailed architecture of $\\mathcal{R}$ .", + "bbox": [ + 212, + 383, + 785, + 460 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Training details. We train using the REDS dataset [27] and follow RealBasicVSR [5] in applying random combinations of blur, noise, JPEG compression, and video compression for synthesizing input degradations. We load 15 frames as an input sequence. The spatial size of inputs is cropped to $64 \\times 64$ , and the batch size is 16. We use a pre-trained flow estimation model SPyNet, the parameter of which is fixed for the first 5K iterations and tuned with other modules later.", + "bbox": [ + 212, + 460, + 785, + 549 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Following RealBasicVSR [5], we perform two-stage training. The first stage trains the model with a Charbonnier loss [19] and SSIM [35] loss for 300K iterations. In the second stage, the model is trained for another 130K iterations with the Charbonnier loss, SSIM loss, perceptual loss [18] and GAN loss [11] together, weighted by 1, 0.001, 1, and 0.005, respectively. The implementations of perceptual loss, GAN loss, and discriminator follow RealBasicVSR [5]. We implement all experiments on 4 Quadro RTX 8000 GPUs with PyTorch. Other details of the training settings are in Supplementary B.2.", + "bbox": [ + 212, + 550, + 785, + 674 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.5 Experimental Results", + "text_level": 1, + "bbox": [ + 215, + 694, + 441, + 709 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We compare our model on four datasets, VideoLQ [5], RealVSR [37], REDS4 [27] and UDM10 [38], with RealBasicVSR [5], Real-ESRGAN [34], BSRGAN [40], DBVSR [28], RealVSR [37], DAN [14], and RealSR [17]. VideoLQ and RealVSR are collected from real-world scenarios. VideoLQ is an unpaired dataset. RealVSR has same-size paired low-quality (LQ) and high-quality (HQ) frames. Our model super-resolves the input frames spatially. Although downsampling LQ enables paired data, it alters the original degradation. Thus, we test on the original LQ and do not use the HQ for evaluation. REDS4 and UDM10 have", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "RealViformer", + "bbox": [ + 638, + 114, + 730, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/bbec08fb949c97929cafd7cf6c1c6da610c575bf3509ad7dcb91c858a390957b.jpg", + "table_caption": [ + "Table 2: Quantitative comparisons with existing methods with best and second-best results. Our method achieves the best ILNIQE and NRQM scores over VideoLQ and RealVSR datasets and the best PSNR, SSIM, and LPIPS over synthetic datasets REDS4 and UDM10 with relatively few parameters and the lowest run-time." + ], + "table_footnote": [], + "table_body": "
RealSRDANRealVSRDBVSRBSRGANReal-ESRGANRealBasicVSROurs
Params (M)16.74.32.725.516.716.76.35.3
Runtime (ms)180250772-1801967349
VideoLQILNIQE↓26.6328.2831.9427.8527.4927.9725.9825.94
NRQM↑6.0543.7423.4603.8516.1566.0576.3066.338
RealVSRILNIQE↓32.8132.2934.39-32.6531.9330.3728.61
NRQM↑5.6103.5233.795-6.1526.2456.5826.588
REDS4PSNR↑22.0222.6718.3022.3522.9421.5623.0923.34
SSIM↑0.50970.55710.49000.55300.57500.55560.60760.6079
LPIPS↓0.59910.63150.72400.62110.37660.35330.29910.2877
UDM10PSNR↑25.3725.9023.3525.0825.9724.9625.9626.42
SSIM↑0.66580.72290.71150.71120.75680.74320.74910.7609
LPIPS↓0.48110.47810.47610.47560.33880.33950.32090.3063
", + "bbox": [ + 220, + 212, + 787, + 366 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "ground-truth images, and we synthesize low-quality inputs through the same pipeline during training. All tested recurrent-based VSR models load the half sequence for videos in RealVSR and the whole sequence for others each time. The quantitative and qualitative results are discussed below.", + "bbox": [ + 212, + 402, + 787, + 464 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Quantitative Results. For evaluation without reference, we apply the ILNIQE [41], the improved version of NIQE [26], and NRQM [24] metrics based on each output sequence's first, middle, and last frames in RGB format [5]. These metrics appear less biased towards oversharpened features; details are given in Supplementary B.3. For evaluation on REDS4 and UDM10, we report more reliable metrics, PSNR, SSIM, and LPIPS [42]. We collect released models of all compared methods and generate sequences. As shown in Tab. 2, RealVformer performs better than other methods with smaller parameters than the most competitive RealBasicVSR [5] and the shortest runtime.", + "bbox": [ + 212, + 468, + 787, + 604 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7154e1d49620a554ebc3214d43090340255fe01d6f563fb14ab126472cc6e278.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 220, + 638, + 330, + 782 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/79300de9e973112bf493dfba011b44c494ff1c3d93bc9e56c00fb17c687553f2.jpg", + "image_caption": [ + "(b) RE" + ], + "image_footnote": [], + "bbox": [ + 333, + 638, + 442, + 782 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/754704fc04ad4b48fa413195c389fc83f074ec91c3db1c74d0d58cc64e20e0a1.jpg", + "image_caption": [ + "(c) RB", + "Fig. 8: Qualitative comparisons on synthetic datasets. Our method produces clearer than RealESRGAN (RE) [34] and RealBasicVSR (RB) [5] for very hard inputs." + ], + "image_footnote": [], + "bbox": [ + 446, + 638, + 555, + 782 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b4b7fcdd3a127285dfa26501028d6ba6ed6f799dd9054c8c5a593e99e6bfd4e7.jpg", + "image_caption": [ + "(d) Ours" + ], + "image_footnote": [], + "bbox": [ + 558, + 638, + 666, + 782 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1cf4db5470437749b3c26ad8ad3585bc5933cd921158cd7c4bffff6fec5300b9.jpg", + "image_caption": [ + "(e) GT" + ], + "image_footnote": [], + "bbox": [ + 669, + 638, + 779, + 782 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Zhang and A. Yao", + "bbox": [ + 271, + 114, + 416, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b60ce4e5e99cf6e1031caf5b5ce7b5e252fbb35e8b12b05b2fe1eefd517135e6.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 359, + 284 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/5bddf52abdf27bd053de4eb6034dfb839fa72ecf7a7bb857dec2a97e6c1232d3.jpg", + "image_caption": [ + "(b) RE" + ], + "image_footnote": [], + "bbox": [ + 361, + 143, + 500, + 284 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/87acee6190ccde9bce59c79d4b1e0f2792378f80e62b05e06062a55971effd69.jpg", + "image_caption": [ + "(c) RB", + "Fig. 9: Qualitative comparisons on real-world datasets. Our method produces less high-frequency artifacts and overshoot effects than RealESRGAN [34] and RealBasicVSR [5]." + ], + "image_footnote": [], + "bbox": [ + 501, + 143, + 640, + 284 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/47abac7bb4302e1cd80e6bbd5fb217d77e3703c2037e2053d5baeb1af4918a53.jpg", + "image_caption": [ + "(d) Ours" + ], + "image_footnote": [], + "bbox": [ + 640, + 143, + 782, + 284 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Qualitative Results. We show qualitative comparisons on synthetic (see Fig. 8) and real-world (see Fig. 9) datasets. Compared to the listed methods, RealViformer generates clear structures with much fewer high-frequency artifacts. More visual comparisons are in Supplementary B.5.", + "bbox": [ + 212, + 368, + 784, + 428 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "User Study. We also performed a user study and asked 30 evaluators on Amazon MTurk to score reconstructions for 85 frames sampled from datasets VideoLQ [5], RealVSR [37], REDS4 [27] and UDM10 [38]. Each worker saw five HR results of the same frame and rated them based on the visual quality, from 1 (the worst) to 5 (the best); as shown in Fig. 10, our method surpasses BSRGAN [40], Real-ESRGAN [34], RealSR [17], and RealBasicVSR [5].", + "bbox": [ + 212, + 429, + 784, + 521 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/69fc77c9b7463039c19a06fe44f3f9f4c460eb54e7081c79daeca8b43c650419.jpg", + "image_caption": [ + "Fig.10: User study results from 30 evaluators on 85 frames. Our method achieves the best among all five methods regrading mean option scores (MOS)." + ], + "image_footnote": [], + "bbox": [ + 230, + 547, + 431, + 664 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/64b854cd570e9665dc5fb6907d03a1982abd2918a8565450bb47afbbfe497d41.jpg", + "table_caption": [ + "Table 3: Ablations of CAF and ICA modules. Channel-attention baseline (ch-baseline) performs better than the spatial-attention baseline (sp-baseline). The CAF module improves the performances for both datasets, and ICA further improves the performances to the state-of-the-art." + ], + "table_footnote": [], + "table_body": "
MethodCAFICAVideoLQ NRQM↑UDM10 LPIPS↓
Sp-baseline--6.0610.3482
Ch-baselineXX6.1810.3085
RealVformer-✓X6.1960.2933
RealVformer✓✓6.3380.2877
", + "bbox": [ + 468, + 652, + 766, + 742 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation Studies. We conduct ablations to validate the advantage of channel attention and the efficacy of the CAF and ICA. We build a spatial-attention baseline (sp-baseline) by replacing the reconstruction module of BasicVSR [3] with SwinIR [21]. Channel-attention baseline (ch-baseline) is built with the same", + "bbox": [ + 212, + 779, + 784, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "RealViformer", + "bbox": [ + 638, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9d56bebcd12991c96ab8b84c91e0d6d0131003d9a103dfffc935862074792869.jpg", + "image_caption": [ + "Fig. 11: (a) Visual comparison between RealVformer and its ablations. Red circles highlight the improved details. (b) Radial Power Spectrum (RPS) of model predictions. Using ICA improves the power of high-frequency components (blue region)." + ], + "image_footnote": [], + "bbox": [ + 228, + 148, + 764, + 256 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "overall architecture as RealVformer but replaces CAF with simple concatenation and substitutes ICA with the channel attention block in Restormer [39]. RealVformer $^{-}$ applies CAF on ch-baseline. All models are trained with the same settings in Sec. 4.4. We report NRQM scores for the VideoLQ [5] dataset and LPIPS scores for the UDM10 [38] da set. As shown in Tab. 3, using the original channel attention module, ch-baseline already yields better performances than sp-baseline. The CAF and ICA modules further improve RealVformer to state-of-the-art performance with channel correlation of propagated information decreasing from 0.436 to 0.422. Fig. 11a visually compares RealVformer with its ablations. Applying CAF reduces artifacts, while ICA provides further improvements. Fig. 11b supplements the Radial Power Spectrum of model predictions. The blur region shows ICA increases the power in the high-frequency region.", + "bbox": [ + 212, + 340, + 787, + 523 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 544, + 359, + 559 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper proposes a real-world VSR model, RealVformer, based on findings from investigating channel and spatial attention in a real-world setting. Explorations reveal that channel attention is less sensitive to the artifacts in query and better serves as a temporal aggregation module to limit model-produced artifacts in hidden states. Additionally, we observe the higher covariance of channel attention outputs and propose the Improved Channel Attention (ICA) Module with a squeeze-and-excite and a covariance-based rescaling mechanism. Based on our findings, we build RealVformer, a channel-attention-based recurrent model for real-world VSR. We propose the CAF module to limit artifact propagation and use the ICA module to achieve better reconstructions. RealVformer performs state-of-the-art on two real-world video datasets with fewer parameters and shorter runtime. On the other hand, we value our findings w.r.t. comparisons between channel and spatial attention and exploration of covariance in channel attention as inspiration for further real-world VSR research.", + "bbox": [ + 212, + 575, + 787, + 787 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement This research is supported by the National Research Foundation, Singapore under its NRF Fellowship for AI (NRF-NRFFAI1-2019-0001).", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Zhang and A. Yao", + "bbox": [ + 271, + 114, + 416, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of National Research Foundation, Singapore.", + "bbox": [ + 215, + 146, + 787, + 193 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 220, + 325, + 237 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bardes, A., Ponce, J., LeCun, Y.: Vicreg: Variance-invariance-covariance regularization for self-supervised learning. arXiv preprint arXiv:2105.04906 (2021)", + "2. Cao, J., Li, Y., Zhang, K., Van Gool, L.: Video super-resolution transformer. arXiv preprint arXiv:2106.06847 (2021)", + "3. Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: Basicvsr: The search for essential components in video super-resolution and beyond. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4947-4956 (2021)", + "4. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5972-5981 (2022)", + "5. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Investigating tradeoffs in real-world video super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5962-5971 (2022)", + "6. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12299-12310 (2021)", + "7. Chen, X., Wang, X., Zhou, J., Qiao, Y., Dong, C.: Activating more pixels in image super-resolution transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22367-22377 (2023)", + "8. Cogswell, M., Ahmed, F., Girshick, R., Zitnick, L., Batra, D.: Reducing overfitting in deep networks by decorrelating representations. arXiv preprint arXiv:1511.06068 (2015)", + "9. Dai, T., Cai, J., Zhang, Y., Xia, S.T., Zhang, L.: Second-order attention network for single image super-resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11065-11074 (2019)", + "0. Fuoli, D., Gu, S., Timofte, R.: Efficient video super-resolution through recurrent latent space propagation. In: 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW). pp. 3476-3485. IEEE (2019)", + "1. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020)", + "2. Hu, J., Shen, L., Sun, G.: Squeeze-and-excitation networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7132-7141 (2018)", + "3. Hua, T., Wang, W., Xue, Z., Ren, S., Wang, Y., Zhao, H.: On feature decorrelation in self-supervised learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9598-9608 (2021)", + "4. Huang, Y., Li, S., Wang, L., Tan, T., et al.: Unfolding the alternating optimization for blind super resolution. Advances in Neural Information Processing Systems 33, 5632-5643 (2020)" + ], + "bbox": [ + 225, + 257, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "RealViformer", + "bbox": [ + 638, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Isobe, T., Jia, X., Gu, S., Li, S., Wang, S., Tian, Q.: Video super-resolution with recurrent structure-detail network. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16. pp. 645-660. Springer (2020)", + "16. Jeelani, M., Cheema, N., Illgner-Fehns, K., Slusallek, P., Jaiswal, S., et al.: Expanding synthetic real-world degradations for blind video super resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1199-1208 (2023)", + "17. Ji, X., Cao, Y., Tai, Y., Wang, C., Li, J., Huang, F.: Real-world super-resolution via kernel estimation and noise injection. In: proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 466-467 (2020)", + "18. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 694-711. Springer (2016)", + "19. Lai, W.S., Huang, J.B., Ahuja, N., Yang, M.H.: Fast and accurate image superresolution with deep laplacian pyramid networks. IEEE transactions on pattern analysis and machine intelligence 41(11), 2599-2613 (2018)", + "20. Liang, J., Cao, J., Fan, Y., Zhang, K., Ranjan, R., Li, Y., Timofte, R., Van Gool, L.: Vrt: A video restoration transformer. arXiv preprint arXiv:2201.12288 (2022)", + "21. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1833-1844 (2021)", + "22. Liang, J., Fan, Y., Xiang, X., Ranjan, R., Ilg, E., Green, S., Cao, J., Zhang, K., Timofte, R., Gool, L.V.: Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems 35, 378-393 (2022)", + "23. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)", + "24. Ma, C., Yang, C.Y., Yang, X., Yang, M.H.: Learning a no-reference quality metric for single-image super-resolution. Computer Vision and Image Understanding 158, 1-16 (2017)", + "25. Mei, Y., Fan, Y., Zhou, Y.: Image super-resolution with non-local sparse attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3517-3526 (2021)", + "26. Mittal, A., Soundararajan, R., Bovik, A.C.: Making a \"completely blind\" image quality analyzer. IEEE Signal processing letters 20(3), 209-212 (2012)", + "27. Nah, S., Baik, S., Hong, S., Moon, G., Son, S., Timofte, R., Mu Lee, K.: Ntire 2019 challenge on video deblurring and super-resolution: Dataset and study. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019)", + "28. Pan, J., Bai, H., Dong, J., Zhang, J., Tang, J.: Deep blind video super-resolution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4811-4820 (2021)", + "29. Ranjan, A., Black, M.J.: Optical flow estimation using a spatial pyramid network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4161-4170 (2017)", + "30. Shi, S., Gu, J., Xie, L., Wang, X., Yang, Y., Dong, C.: Rethinking alignment in video super-resolution transformers. arXiv preprint arXiv:2207.08494 (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Zhang and A. Yao", + "bbox": [ + 271, + 114, + 415, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Song, Y., Wang, M., Yang, Z., Xian, X., Shi, Y.: Negvsr: Augmenting negatives for generalized noise modeling in real-world video super-resolution. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 10705-10713 (2024)", + "32. Wang, H., Chen, X., Ni, B., Liu, Y., Liu, J.: Omni aggregation networks for lightweight image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22378-22387 (2023)", + "33. Wang, X., Chan, K.C., Yu, K., Dong, C., Change Loy, C.: Edvr: Video restoration with enhanced deformable convolutional networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019)", + "34. Wang, X., Xie, L., Dong, C., Shan, Y.: Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1905-1914 (2021)", + "35. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)", + "36. Xie, L., Wang, X., Shi, S., Gu, J., Dong, C., Shan, Y.: Mitigating artifacts in real-world video super-resolution models. arXiv preprint arXiv:2212.07339 (2022)", + "37. Yang, X., Xiang, W., Zeng, H., Zhang, L.: Real-world video super-resolution: A benchmark dataset and a decomposition based learning scheme. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4781-4790 (2021)", + "38. Yi, P., Wang, Z., Jiang, K., Jiang, J., Ma, J.: Progressive fusion video superresolution network via exploiting non-local spatio-temporal correlations. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3106-3115 (2019)", + "39. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5728-5739 (2022)", + "40. Zhang, K., Liang, J., Van Gool, L., Timofte, R.: Designing a practical degradation model for deep blind image super-resolution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4791-4800 (2021)", + "41. Zhang, L., Zhang, L., Bovik, A.C.: A feature-enriched completely blind image quality evaluator. IEEE Transactions on Image Processing 24(8), 2579-2591 (2015)", + "42. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)", + "43. Zhang, Y., Li, K., Li, K., Wang, L., Zhong, B., Fu, Y.: Image super-resolution using very deep residual channel attention networks. In: Proceedings of the European conference on computer vision (ECCV). pp. 286-301 (2018)" + ], + "bbox": [ + 212, + 146, + 787, + 715 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "RealViformer", + "bbox": [ + 638, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/RealViformer_ Investigating Attention for Real-World Video Super-Resolution/39444a96-c968-4d03-9b6c-b416cbbaab18_model.json b/2024/RealViformer_ Investigating Attention for Real-World Video Super-Resolution/39444a96-c968-4d03-9b6c-b416cbbaab18_model.json new file mode 100644 index 0000000000000000000000000000000000000000..18dea784e24d768c89b31bba8c451a48014b11a4 --- /dev/null +++ b/2024/RealViformer_ Investigating Attention for Real-World Video Super-Resolution/39444a96-c968-4d03-9b6c-b416cbbaab18_model.json @@ -0,0 +1,2478 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.266, + 0.141, + 0.744, + 0.187 + ], + "angle": 0, + "content": "RealVformer: Investigating Attention for Real-World Video Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.377, + 0.213, + 0.626, + 0.228 + ], + "angle": 0, + "content": "Yuehan Zhang and Angela Yao" + }, + { + "type": "text", + "bbox": [ + 0.384, + 0.24, + 0.621, + 0.269 + ], + "angle": 0, + "content": "National University of Singapore {zyuehan,ayao}@comp.nus.edu.sg" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.305, + 0.744, + 0.527 + ], + "angle": 0, + "content": "Abstract. In real-world video super-resolution (VSR), videos suffer from in-the-wild degradations and artifacts. VSR methods, especially recurrent ones, tend to propagate artifacts over time in the real-world setting and are more vulnerable than image super-resolution. This paper investigates the influence of artifacts on commonly used covariance-based attention mechanisms in VSR. Comparing the widely-used spatial attention, which computes covariance over space, versus the channel attention, we observe that the latter is less sensitive to artifacts. However, channel attention leads to feature redundancy, as evidenced by the higher covariance among output channels. As such, we explore simple techniques such as the squeeze-excite mechanism and covariance-based rescaling to counter the effects of high channel covariance. Based on our findings, we propose RealVformer. This channel-attention-based real-world VSR framework surpasses state-of-the-art on two real-world VSR datasets with fewer parameters and faster runtimes. The source code is available at https://github.com/Yuehan717/RealVformer." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.552, + 0.377, + 0.568 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.735 + ], + "angle": 0, + "content": "Video super-resolution (VSR) recovers a high-resolution (HR) sequence of frames from its low-resolution (LR) counterpart. Recurrent convolutional approaches are commonly used in VSR with standard settings, assuming the LR frames are downsampled from HR frames with known kernels. [3, 10, 15]. However, in real-world VSR (RWVSR), the low-resolution videos are not simply downsampled versions of their high-resolution counterparts. Instead, they feature complex degradations that arise from the camera imaging system, compression, internet transmission, and other factors. These degradations make architecture design for RWVSR challenging, as artifacts and degradations tend to propagate and get exaggerated over the recurrent connection [5, 36]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Recently, transformer architectures have replaced convolutional architectures as state-of-the-art for standard VSR. While attention mechanisms have replaced convolution operations, most methods retain the recurrent connection to aggregate the information over time [22, 30]. Yet, such architectures do not always perform well on RWVSR. For instance, a Swin-based model [21] designed for standard VSR, when applied to a real-world input frame, as shown in Fig. 1a, generates more artificial lines than the convolutional model RealBasicVSR [5]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.417, + 0.128 + ], + "angle": 0, + "content": "Y. Zhang and A. Yao" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.143, + 0.366, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.306, + 0.197, + 0.333, + 0.207 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.144, + 0.462, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.38, + 0.197, + 0.449, + 0.207 + ], + "angle": 0, + "content": "RealBasicVSR" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.213, + 0.366, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.283, + 0.269, + 0.359, + 0.28 + ], + "angle": 0, + "content": "Swin-based VSR" + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.213, + 0.462, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.269, + 0.46, + 0.28 + ], + "angle": 0, + "content": "Ours: RealVformer" + }, + { + "type": "image_caption", + "bbox": [ + 0.302, + 0.287, + 0.441, + 0.299 + ], + "angle": 0, + "content": "(a) Visual comparisons." + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.156, + 0.616, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.248, + 0.546, + 0.256 + ], + "angle": 0, + "content": "W" + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.256, + 0.583, + 0.265 + ], + "angle": 0, + "content": "Spatial Attention" + }, + { + "type": "image_caption", + "bbox": [ + 0.624, + 0.208, + 0.634, + 0.22 + ], + "angle": 0, + "content": "H" + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.156, + 0.745, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.255, + 0.718, + 0.265 + ], + "angle": 0, + "content": "Channel Attention" + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.287, + 0.714, + 0.299 + ], + "angle": 0, + "content": "(b) Schematics of two attentions." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.31, + 0.785, + 0.407 + ], + "angle": 0, + "content": "Fig. 1: (a) Designing a RWVSR transformer is not trivial. A Swin-based transformer suited for standard VSR hallucinates more lines than a RealBasicVSR, a convolutional state-of-the-art. We propose RealVformer based on our investigation of attention under the RWVSR setting. RealVformer generates details with fewer artifacts than RealBasicVSR [3] and the Swin-based VSR model. (b) Schematic for spatial and channel attention. Spatial attention aggregates features based on pixel representations. Channel attention takes \\( H \\times W \\) feature map for matching across channels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.785, + 0.542 + ], + "angle": 0, + "content": "Why should transformers perform well on (synthetic) standard VSR but poorly in real-world cases? We speculate that standard VSR transformers benefit from the similarity-based matching of the attention mechanism, which accurately aggregates information spatially and temporally [21,22,30]. However, when input degradation exists, the aggregated information becomes less reliable because the attention queries may be derived from both true source video and artifacts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.547, + 0.785, + 0.744 + ], + "angle": 0, + "content": "This work investigates and sheds light on the sensitivity of attention in real-world settings. We compare two covariance-based attention mechanisms used in low-level Transformers: spatial attention [21,30] and channel attention [39]. Spatial attention takes pixel-wise features as keys and queries and estimates their covariances across spatial positions. The most popular form is window-based attention [6]; The shift-window scheme from Swin Transformers [23] enables the model to access distant spatial ranges [2,21,30] without computational blow-up. Spatial attention is widely used for video and image super-resolution, albeit in the standard setting. Channel attention [39] estimates covariances across channels (see Fig. 1b) and collapses the spatial extent of a feature map. It defines the number of queries and keys by channel numbers rather than spatial resolution. Although established for deblurring or denoising [39], channel attention's efficacy in super-resolution remains to be determined." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Our experiments show that channel attention is less sensitive to artifacts than the spatial counterparts, resulting in higher performance gains in RWVSR. However, it is also revealed that channel attention leads to feature channels with higher covariance. From a learning perspective, a high covariance is undesirable because it is a strong indicator of feature redundancy [1,8,13]. Therefore, using channel attention naively will have limited improvements over existing RWVSR" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.64, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "RealViformer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "state-of-the-art methods. Such a finding has wide-reaching impacts as channel attention is used increasingly for low-level vision." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.342 + ], + "angle": 0, + "content": "To verify our findings, we explore established mechanisms to counter the effects of feature redundancy - simple techniques such as squeeze-excite and covariance-based rescaling improve the vanilla channel attention design. From these outcomes, we propose RealVformer, a new transformer real-world VSR model. RealVformer performs channel attention between the current frame feature and the propagated hidden state to limit model-produced artifacts. The model then reconstructs features through improved channel attention modules featuring squeeze-and-excite and covariance-based channel rescaling mechanisms. With our effective designs, RealVformer achieves state-of-the-art performance with fewer parameters on challenging synthetic video datasets and two real-world video datasets collected from different scenes." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.345, + 0.714, + 0.359 + ], + "angle": 0, + "content": "Summarizing our contributions in order of importance, our paper" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.371, + 0.791, + 0.431 + ], + "angle": 0, + "content": "- investigates the differences between spatial and channel attention for RWVSR. Spatial attention, although widely used, is revealed to be highly sensitive to the noise and degradations common in RWVSR sequences, while channel attention is more robust." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.433, + 0.787, + 0.477 + ], + "angle": 0, + "content": "- reveals that naively applying channel attention increases channel covariance, which is problematic for learning; this overlooked fact has a wide-reaching impact as channel attention becomes more used in low-level vision." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.478, + 0.787, + 0.539 + ], + "angle": 0, + "content": "- empirically verifies the negative effect of high channel covariance by countering it with established techniques, based on which we develop the RealVformer for RWVSR. Our simple modification surpasses state-of-the-art despite using less compute." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.371, + 0.791, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.564, + 0.388, + 0.58 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.788, + 0.688 + ], + "angle": 0, + "content": "Standard Video Super-Resolution models focus on architecture design to use temporal information better [2-4, 15]. Previous research starts from sliding-window-based [15,33] to recurrent-based frameworks [3,4] for using distant-frame information. Recent works introduce Transformer blocks into existing recurrent frameworks to overcome the locality limitation of convolution and accurately match abundant information for feature reconstruction [20,22,30]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Real-world video super-resolution focuses on modeling, removing, and limiting the impact of real-world degradations. Existing works are convolutional models and focus on designing losses or modules for degradation processing. DBVSR [28] explicitly estimates the degradation kernel through a sub-network. RealBasicVSR [5] tries to 'clean' artifacts through a processing module for BasicVSR [3]. In a similar approach, FastRealVSR [36] borrows an external pool of blur and sharpening filters to 'clean' the hidden states. Other recent works [16,31] advance the synthesis method for paired training data. Instead, we focus on investigating the function of attention in RWVSR rather than architecture design or data synthesis." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.416, + 0.128 + ], + "angle": 0, + "content": "Y. Zhang and A. Yao" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "Attention mechanisms have been widely applied for low-level vision tasks [9, 25, 43, 43]. Transformers with covariance-based attention are the most prevalent [21, 30, 39] for standard VSR. Most existing methods adopt shift-window-based spatial attention [23] to aggregate information from other positions within or across frames. In contrast, Restormer [39] computes the covariance among channels and shows its effectiveness for multiple image restoration tasks. More recent works stitch spatial and channel attention together to enlarge the receptive field [7, 32] for standard image super-resolution. Instead, we investigate attention mechanisms in terms of their sensitivity to real-world degradation for the first time and develop an effective real-world VSR Transformer." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.32, + 0.506, + 0.338 + ], + "angle": 0, + "content": "3 Explorations on Attention" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.35, + 0.788, + 0.412 + ], + "angle": 0, + "content": "Sec. 3.1 defines the VSR task and the two attention mechanisms. Sec. 3.2 compares the channel and spatial attentions' sensitivity to query artifacts and effects on real-world VSR performance. Sec. 3.3 reveals that channel attention leads to higher covariance among channels and explores mitigating options." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.431, + 0.373, + 0.446 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.454, + 0.788, + 0.699 + ], + "angle": 0, + "content": "Standard vs. Real-World VSR. Given a low-resolution (LR) video sequence with \\(T\\) frames \\(I^L \\in \\mathcal{R}^{T \\times H \\times W \\times K}\\), VSR models reconstruct a high-resolution sequence \\(I^H \\in \\mathcal{R}^{T \\times sH \\times sW \\times K}\\), where \\(H \\times W\\) is the input spatial resolution, \\(K\\) is the number of input channels and \\(s\\) is the scaling factor. In the standard setting of VSR, \\(I_t^L\\), where \\(t \\in \\{0, \\dots, T\\}\\), is assumed as the downsampled version of \\(I_t^H\\), defined as: \\(I_t^L = (I_t^H) \\downarrow_{\\frac{1}{s}}\\). Both training and testing datasets follow this formulation to generate LR-HR pairs given \\(I^H\\). In a real-world setting, there is no closed formulation for the relationship between \\(I_t^L\\) and \\(I_t^H\\) due to the unknown distribution of real-world degradations. Real-ESRGAN [34] proposed a widely-used training setting that randomly applies synthesized blur, noise, compression, and resizing to the HR frames to generate paired LR frames with complex degradations. The testing datasets are either synthesized by the same pipeline in training or collected from diverse real-world sources [5, 37]. The synthesized testing sets have paired ground-truth sequences and are evaluated by full-reference metrics, e.g. PSNR and LPIPS [42]; real-world datasets are always without ground truth and require no-reference metrics, such as NRQM [24]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.7, + 0.788, + 0.807 + ], + "angle": 0, + "content": "Attention Definitions. In Transformers, the attention modules project layer-normalized tensor \\( X \\in \\mathbb{R}^{C \\times H \\times W} \\) to query \\( Q \\), and tensor \\( Y \\in \\mathbb{R}^{\\hat{C} \\times \\hat{H} \\times \\hat{W}} \\) to key \\( K \\) and value \\( V \\), where \\( H \\times W \\) and \\( \\hat{H} \\times \\hat{W} \\) are the spatial resolution of the normalized tensors. The attention map \\( \\mathcal{A} \\) is generated by calculating the covariance between \\( Q \\) and \\( K \\), followed by a softmax function, before being applied to the value \\( V \\) to produce output \\( O \\). Spatial and channel attention differ in tensor dimension taken for the covariance calculation." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.788, + 0.84 + ], + "angle": 0, + "content": "1 Note we define mutual attention as the default; self-attention is a special form of mutual attention where \\( Y = X \\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.64, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "RealViformer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.146, + 0.788, + 0.226 + ], + "angle": 0, + "content": "Spatial attention generates an \\(\\mathbb{R}^{HW\\times \\hat{H}\\hat{W}}\\) attention map, \\(\\mathcal{A}_s\\), by computing the covariance between features in the query \\((Q_{s})\\) and key \\((K_{s})\\) at each spatial position. The query, key and values are computed as \\(Q_{s} = W_{s}^{Q}X\\), \\(K_{s} = W_{s}^{K}Y\\), \\(V_{s} = W_{s}^{V}Y\\), where \\(W_{s}^{Q}\\in \\mathbb{R}^{D_{s}\\times C}\\), \\(\\{W_s^K,W_s^V\\} \\in \\mathbb{R}^{D_s\\times \\hat{C}}\\), and \\(D_{s}\\) is the dimension of projections. The attention map \\(\\mathcal{A}_s\\) and output attention features \\(O_{s}\\) are:" + }, + { + "type": "equation", + "bbox": [ + 0.352, + 0.233, + 0.786, + 0.252 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} _ {s} = \\operatorname {s o f t m a x} \\left(Q _ {s} ^ {T} K _ {s} / \\sqrt {D _ {s}}\\right), O _ {s} = \\mathcal {A} _ {s} V _ {s} ^ {T}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.259, + 0.784, + 0.323 + ], + "angle": 0, + "content": "where \\( Q_{s}(K_{s}) \\) is reshaped to matrix of \\( \\mathbb{R}^{D_s\\times HW} \\) \\( (\\mathbb{R}^{D_s\\times \\hat{H}\\hat{W}}) \\). We omit the later reshaping operation for brevity. \\( \\mathcal{A}_s \\) exhaustively relates every spatial location to every other location in space. A window-based implementation [21] limits the correlations to windows of \\( \\omega \\times \\omega \\), reducing the map size to \\( \\mathbb{R}^{\\omega^2\\times \\omega^2} \\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.323, + 0.784, + 0.385 + ], + "angle": 0, + "content": "Channel attention applies convolutions on \\( X \\) to get query \\( (Q_{c}) \\) and \\( Y \\) for key \\( (K_{c}) \\) and value \\( (V_{c}) \\) [39], after which features are the same in spatial resolution (assumed to be \\( H \\times W \\) in the following context). It estimates covariance across channels to yield a \\( C \\times \\hat{C} \\) sized map \\( \\mathcal{A}_{c} \\) and output features \\( O_{c} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.369, + 0.393, + 0.786, + 0.41 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} _ {c} = \\operatorname {s o f t m a x} \\left(Q _ {c} K _ {c} ^ {T} / \\alpha\\right), O _ {c} = \\mathcal {A} _ {c} V _ {c}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.785, + 0.482 + ], + "angle": 0, + "content": "where \\(Q_{c}\\) is reshaped to \\(\\mathbb{R}^{C\\times HW}\\), \\(K_{c}\\) and \\(V_{c}\\) are reshaped to \\(\\mathbb{R}^{\\hat{C}\\times HW}\\) and \\(\\alpha\\) is learnable scaling parameter. Because channel attention computes correlations between features of size \\(\\mathbb{R}^{1\\times HW}\\), it has a larger spatial context than the \\(\\mathbb{R}^{D_s\\times 1}\\)-sized features in spatial attention (see Fig. 1b)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.5, + 0.509, + 0.515 + ], + "angle": 0, + "content": "3.2 Attention in Real-World VSR" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.784, + 0.598 + ], + "angle": 0, + "content": "Sensitivity to Query Artifacts. In standard VSR, attention is used to aggregate information temporally and spatially [22, 30]. The assumption is that queries can match beneficial cues for super-resolution. But what if the queries themselves are unreliable? We speculate this to be true in real-world VSR, where inputs have artifacts and degradations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.785, + 0.749 + ], + "angle": 0, + "content": "How affected are spatial and channel attention mechanisms by query artifacts? Fig. 2 shows how we compare, using cosine similarity, the attention outputs based on queries from the same frame with and without certain degradations. Using encoding layers from a standard convolutional VSR model, we perform the attention operations defined in Sec. 3.1 by taking embeddings of frame \\( I_{t} \\) as query and \\( I_{t-1} \\) as key and value. \\( I_{t} \\) and \\( I_{t-1} \\) are clean without degradation other than downsampling, and we represent the output feature of the attention module as \\( O \\). When additional degradation \\( D_{i} \\) is applied to \\( I_{t} \\), we represent the corresponding output as \\( O_{D_{i}} \\). A smaller deviation of \\( O_{D_{i}} \\) from \\( O \\) indicates lower sensitivity to artifacts in the query." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Tab. 1 shows the cosine similarity between \\( O \\) and \\( O_{D_i} \\) for spatial and channel attention modules. Curiously, \\( O_{D_i} \\) based on the channel attention module is more similar to the \\( O \\) matched by the degradation-free query, indicating that channel attention is less sensitive to query artifacts. Intuitively, the lower sensitivity of channel attention is related to the larger spatial context used for feature matching. Given a deep feature of size \\( \\mathbb{R}^{C\\times H\\times W} \\), channel attention uses feature sized" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.418, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang and A. Yao" + }, + { + "type": "image", + "bbox": [ + 0.363, + 0.143, + 0.64, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.258, + 0.788, + 0.342 + ], + "angle": 0, + "content": "Fig.2: Schematic for sensitivity comparison. \\( I_{t-1}, I_t \\) are downsampled but clean frames at times \\( t \\) and \\( t-1 \\). \\( D_i(.) \\) apply degradations to \\( I_t \\), where \\( D_i \\in \\{\\text{blur, noise, compression}\\} \\). \\( O \\) and \\( O_{D_i} \\) are output features of the attention module. Queries are from the embedding at time \\( t \\), and keys and values are from time \\( t-1 \\). Higher cosine similarities \\( S \\) between attention output features \\( O \\) and \\( O_{D_i} \\) reflect less sensitivity to artifacts in queries." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.389 + ], + "angle": 0, + "content": "Table 1: Cosine similarity between \\( O \\) and \\( {O}_{{D}_{i}} \\) ,attention outputs without and with query degradation. Outputs of channel attention change less under query degradation." + }, + { + "type": "table", + "bbox": [ + 0.359, + 0.4, + 0.642, + 0.459 + ], + "angle": 0, + "content": "
ModuleDi
blurnoisecompression
Spatial attention0.750.920.84
Channel attention0.980.990.99
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.787, + 0.549 + ], + "angle": 0, + "content": "\\(\\mathbb{R}^{1 \\times HW}\\) to calculate the covariance across channels. As such, feature aggregation is based on global information observed in a large normalized spatial context. Instead, the covariance of spatial attention is for features at each location sized \\(\\mathbb{R}^{C \\times 1}\\), so it is likely more sensitive to local value changes from artifacts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.788, + 0.702 + ], + "angle": 0, + "content": "Impact on VSR Models. While channel attention output features are less sensitive to query artifacts than spatial attention, how well do they fare for real-world VSR? We experiment by incorporating these attention variants into VSR models. We focus specifically on recurrent pipelines due to their popularity in standard and real-world VSR [5,22,36]. In recurrent pipelines, artifacts in hidden states may propagate over time, get exaggerated, and negatively influence overall performance [5, 36]. Therefore, we explicitly add the attention module between features of the current frame and the propagated hidden state. Ideally, the added attention will select relative information from hidden states for the current frame and reduce the propagation of model-produced artifacts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.703, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Fig. 3a shows the baseline model which concatenates \\( f_{t} \\), the shallow feature of at time \\( t \\), and \\( \\hat{h}_{t-1} \\), the spatially aligned hidden state at time \\( t-1 \\). We experiment with two variants by using a channel-attention module \\( G_{c} \\) or a spatial-attention module \\( G_{s} \\) to replace the concatenation in the baseline. As shown in Fig. 3b, the query is generated by \\( f_{t} \\), and the key and value are predicted from \\( \\hat{h}_{t-1} \\). The attention output \\( O_{t}^{A} \\) is concatenated with \\( f_{t} \\) as the input for \\( \\mathcal{R} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "All models are trained on the REDS dataset [27] with the random degradation pipeline from Real-ESRGAN [34]. We test on REDS4 [27] with different types and extents of degradations, including Gaussian blur, Gaussian noise, and JPEG" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.64, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RealViformer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.294, + 0.145, + 0.507, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.249, + 0.468, + 0.261 + ], + "angle": 0, + "content": "(a) Recurrent Baseline" + }, + { + "type": "image", + "bbox": [ + 0.57, + 0.152, + 0.768, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.627, + 0.242, + 0.708, + 0.253 + ], + "angle": 0, + "content": "(b) Attention" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.389 + ], + "angle": 0, + "content": "Fig. 3: (a) The recurrent baseline in Sec. 3.2 has a shallow mapping module \\(\\mathcal{F}\\), reconstruction module \\(\\mathcal{R}\\), upsampling module \\(\\mathcal{U}\\) and warping function \\(W\\). \\(W\\) aligns the hidden state \\(h_{t-1}\\) to feature at \\(t\\) based on optical flow \\(s_{(t-1)\\rightarrow t}^{f}\\). All residual blocks are convolutional. The concatenation between \\(f_{t}\\) and \\(\\hat{h}_{t-1}\\) are replaced with the spatial or channel attention modules in (b) to compare the effect of attention. (b) The attention module first applies layer normalization to \\(f_{t}\\) and \\(\\hat{h}_{t-1}\\) and then performs channel or spatial attention according to Sec. 3.1. The output feature \\(O_{t}^{A}\\) concatenated with \\(f_{t}\\) is processed by the module \\(\\mathcal{R}\\) in (a)." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.404, + 0.402, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.405, + 0.404, + 0.583, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.404, + 0.774, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.533, + 0.788, + 0.591 + ], + "angle": 0, + "content": "Fig. 4: Comparison of spatial and channel attention through impact on the performance of real-world VSR model. The Y-axis shows improvements compared to the convolutional baseline. A lower LPIPS score is better. The channel attention module is the best except for the PSNR score of highly blurred inputs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.62, + 0.788, + 0.713 + ], + "angle": 0, + "content": "compression. Fig. 4 displays changes in PSNR and LPIPS compared with the baseline model. Trained on the same dataset and degradation setting, channel attention between temporal information yields better objective and perceptual reconstruction quality than spatial attention and baseline for noise and JPEG compression inputs. Channel attention still achieves better scores for blurred inputs, except for the PSNR performance of severely blurred inputs." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.737, + 0.538, + 0.751 + ], + "angle": 0, + "content": "3.3 Limitations of Channel Attention" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "The results in Sec. 3.2 indicate that channel attention is better suited for propagating information over time. However, it also has an inherent flaw - the correlation among output channels will increase since each channel in the attention output is a weighted summation over channels of the value \\( V_{c} \\). We calculate the covariance following VICReg [1] for a quantitative measurement. Given deep" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.418, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang and A. Yao" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.146, + 0.785, + 0.178 + ], + "angle": 0, + "content": "feature \\(z^n \\in \\mathbb{R}^{C \\times H \\times W}\\), where \\(n \\in \\{1, N\\}\\) is the index of a sample, we reshape \\(z^n\\) to \\(\\mathbb{R}^{C \\times HW}\\) and define the covariance matrix over \\(Z = \\{z^1, \\dots, z^N\\}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.36, + 0.189, + 0.786, + 0.232 + ], + "angle": 0, + "content": "\\[\nC o v (Z) = \\frac {1}{N - 1} \\sum_ {n = 1} ^ {N} \\left(z _ {n} - \\bar {z}\\right) \\left(z _ {n} - \\bar {z}\\right) ^ {T}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.243, + 0.784, + 0.381 + ], + "angle": 0, + "content": "where \\(\\bar{z} = \\frac{1}{N}\\sum_{n=1}^{N}z_{n}\\). We then define the indicator for the covariance matrix as \\(ac(Z) = \\frac{1}{d}\\sum_{i\\neq j}|Cov(Z)|_{i,j}\\), i.e. the average of absolute off-diagonal coefficients of \\(Cov(Z)\\), where \\(d\\) is the number of off-diagonal coefficients. Function \\(ac(Z)\\) encodes the covariance among feature channels. Taking \\(O\\) in Fig. 2 for comparison, \\(ac(O)\\) with \\(N = 400\\) based on channel attention is 0.87 and significantly higher than the input features (\\(\\approx 0.15\\)). Instead, \\(ac(O)\\) of spatial attention remains similar to the input features. Previous works on representation learning [1] and overfitting [8] propose that the high covariance of feature channels indicates redundancy and tends to hinder informative predictions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.785, + 0.517 + ], + "angle": 0, + "content": "Similarly, we speculate the redundancy effects will hinder the prediction of HR outputs when adopting channel attention in building blocks. For a closer look, we investigate the standard VSR task, which is artifact-free. Specifically, we build \\( M_{c} \\) and \\( M_{s} \\) by replacing the Residual Block in Fig. 3a with channel attention blocks [39] and spatial attention blocks [21]. Models are trained on the REDS dataset without extra degradation. For evaluation, we choose SSIM [35], which focuses on structural information. The SSIM score of \\( M_{c} \\) (0.8338) is lower than \\( M_{s} \\) (0.8432); \\( ac(\\cdot) \\) of the last features before the upsampling module is higher for channel attention than for spatial attention, i.e. 0.199 vs. 0.147." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.787, + 0.593 + ], + "angle": 0, + "content": "We adopt two simple modifications to channel attention and boost informative features to verify the redundancy effect empirically. Fig. 5 shows the Improved Channel Attention (ICA) module. Our approach features two key steps to enhance the quality. First, we use the squeeze-and-excite mechanism to predict new information. The features are squeezed channel-wise to extract meaningful" + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.622, + 0.67, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.767, + 0.788, + 0.837 + ], + "angle": 0, + "content": "Fig. 5: Improved Channel Attention Module (ICA), showing self-attention for simplicity. The 'squeeze' convolution compresses the number of input feature channels \\( X \\in \\mathbb{R}^{C \\times H \\times W} \\) by ratio \\( r \\). The features are then rescaled by weights predicted from the \\( \\frac{C}{r} \\times \\frac{C}{r} \\) attention map before being expanded by the 'excite' convolution back to the original number of input channels." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.64, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RealViformer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.33, + 0.141, + 0.67, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.321, + 0.558, + 0.331 + ], + "angle": 0, + "content": "(c) Reconstruction Module" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.456 + ], + "angle": 0, + "content": "Fig. 6: The framework of RealVformer. (a) Overview of RealVformer, following a unidirectional recurrent framework. The outputs of the Forward module are propagated to the next time step and upsampled by module \\(\\mathcal{U}\\) to get HR frames. (b) Explanation of the Forward module in (a), where \\(W\\) denotes the warping function. The reconstruction module \\(\\mathcal{R}\\) takes current frame \\(I_t^L\\) and warped hidden state \\(\\hat{h}_{t-1}\\) as inputs. (c) Reconstruction module \\(\\mathcal{R}\\). The shallow feature of \\(I_t^L\\) and \\(\\hat{h}_{t-1}\\) are fused by CAF and then forwarded to Transformer blocks with U-shape connection [39]. Module GDFN follows Restormer [39]. Details of CAF and ICA modules are stated in Fig. 7 and Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.788, + 0.653 + ], + "angle": 0, + "content": "information and then expanded into new channels based on the attention outputs. Secondly, we rescale channels in attention outputs by scalar weights predicted from the attention map. The attention map measures relationships across channels and encourages the associated convolutions in the 'excite' operation to yield more precise and discriminating features. Our designs are inspired by the SE Network [12] but with two key distinctions. First, we apply the squeeze-and-excite mechanism to generate new information, while SE Network uses it to help with scalar weight prediction. Second, the rescaling weights prediction takes the attention map as a cue rather than the naive pooling results of channels. In Sec. 4, we show the efficacy of new designs on the real-world VSR task and compare channel correlation with and without new designs in ablations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.676, + 0.384, + 0.691 + ], + "angle": 0, + "content": "4 RealVifromer" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.709, + 0.341, + 0.723 + ], + "angle": 0, + "content": "4.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.786, + 0.809 + ], + "angle": 0, + "content": "We apply our findings in Sec. 3.2 to the real-world VSR task as further support. To that end, we design RealVformer to incorporate channel attention as the basic processing module and the modification to boost informative features. Our emphasis here is not novel architecture design but a showcase of effectiveness brought by applying our findings in Sec. 3.2." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.811, + 0.787, + 0.84 + ], + "angle": 0, + "content": "RealVformer is a recurrent Transformer network with channel attention modules. Fig. 6 shows the model architecture. To reduce the computational" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.418, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang and A. Yao" + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.148, + 0.484, + 0.159 + ], + "angle": 0, + "content": "Channel Attention Fusion module (CAF)" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.161, + 0.681, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.318, + 0.263, + 0.695, + 0.287 + ], + "angle": 0, + "content": "1x1 Norm 3x3 Depth-wise \\(\\mathbb{C}\\) Concatenation Reshape \\(\\bigotimes\\) Matrix multiplication Conv" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.304, + 0.789, + 0.348 + ], + "angle": 0, + "content": "Fig. 7: Details of Channel Attention Fusion (CAF) module. CAF gets the query from current frame feature \\( f_{t} \\) and {key, value} from hidden state \\( \\hat{h}_{t-1} \\). The attention output is concatenated with \\( f_{t} \\) to process for module output \\( O_{t} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.383, + 0.789, + 0.479 + ], + "angle": 0, + "content": "cost, we follow a widely used recurrent architecture [3] in a unidirectional setting. The model first estimates optical flow \\( s_{(t - 1)\\rightarrow t}^{f} \\) from \\( I_{t - 1}^{L} \\) to \\( I_{t}^{L} \\) through Spynet [29] and warps previous hidden state \\( h_{t - 1} \\) to current time step based on the flow. Frame \\( I_t^L \\) and spatially aligned hidden state \\( \\hat{h}_{t - 1} \\) are processed in the reconstruction module \\( \\mathcal{R} \\). The hidden state is updated with outputs of \\( \\mathcal{R} \\) and further processed by the upsampling module \\( \\mathcal{U} \\) to output high-resolution frames." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.482, + 0.79, + 0.572 + ], + "angle": 0, + "content": "The reconstruction module \\(\\mathcal{R}\\) uses channel attention in two ways. First, the Channel Attention Fusion (CAF) module fuses the temporal information to limit the produced artifacts in the hidden state. CAF queries the aligned hidden state \\(\\hat{h}_{t-1}\\) by the shallow feature \\(f_t\\). Secondly, we take the Improved Channel Attention Module (ICA) in Fig. 5 to build the Transformer blocks for better HR reconstructions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.612, + 0.607, + 0.629 + ], + "angle": 0, + "content": "4.2 Channel Attention Fusion (CAF) Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.656, + 0.789, + 0.844 + ], + "angle": 0, + "content": "As explored in Sec. 3.2, adding channel attention promotes the real-world VSR performance compared to spatial attention and the simple concatenation baseline. Thus, we keep this design in our model and put it as the Channel Attention Fusion (CAF) module for temporal aggregation. Fig. 7 show details of how CAF perform channel attention between \\( f_{t} \\), the feature of the current frame, and \\( \\hat{h}_{t-1} \\), the aligned propagated hidden state. The query \\( (Q_{t}) \\) is generated as \\( Q_{t} = K_{3\\times 3} * \\text{LayerNorm}(f_{t}) \\), where \\( K_{3\\times 3} * \\) refers to \\( 3 \\times 3 \\) convolution operation. Similarly, we process hidden state \\( \\hat{h}_{t-1} \\) by \\( \\tilde{h}_{t-1} = K_{3\\times 3}^{d} * K_{1\\times 1} * \\text{LayerNorm}(\\hat{h}_{t-1}) \\), where \\( K_{3\\times 3}^{d} * \\) is the \\( 3 \\times 3 \\) depth-wise convolution which double the channel number. Chunking \\( \\tilde{h}_{t-1} \\) gives key \\( (K_{t}) \\) and value \\( (V_{t}) \\). The calculation of attention map \\( A_{t} \\in \\mathbb{R}^{C \\times C} \\) follows Eq. (2). The final output \\( O_{t} \\) is computed as \\( O_{t} = K_{1\\times 1} * K_{3\\times 3}^{d} * K_{1\\times 1} * \\mathbf{C}[A_{t}V_{t}; f_{t}] \\), where \\( \\mathbf{C}[\\cdot; \\cdot] \\) denotes concatenation." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.64, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "RealViformer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.625, + 0.163 + ], + "angle": 0, + "content": "4.3 Improved Channel Attention (ICA) Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.172, + 0.788, + 0.34 + ], + "angle": 0, + "content": "The design of ICA follows Fig. 5 for empirically verifying findings in Sec. 3.3. Squeeze and Excite follows the squeeze-and-excite mechanism in Fig. 5 and helps to predict channels with new information. It squeezes the channels of input feature \\( X \\in \\mathbb{R}^{C \\times H \\times W} \\) by factor \\( r \\). The attention map refers to self-attention on \\( X \\) and of size \\( \\mathbb{R}^{\\frac{C}{r} \\times \\frac{C}{r}} \\). Excitation layers expand outputs back to size \\( \\mathbb{R}^{C \\times H \\times W} \\). Correlation-based channel weighting weights the output channel based on scalars predicted from the attention map. This design emphasizes channels for predicting more discriminative features in 'excite' operation. The attention map \\( A_r \\in \\mathbb{R}^{\\frac{C}{r} \\times \\frac{C}{r}} \\) is taken for calculating the average and max values along the rows. The average and max values are combined and mapped to weights of size \\( \\mathbb{R}^{\\frac{C}{r} \\times 1} \\) through linear layers and a sigmoid function." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.361, + 0.457, + 0.375 + ], + "angle": 0, + "content": "4.4 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.787, + 0.461 + ], + "angle": 0, + "content": "Model details. RealVformer uses SPyNet [29] for flow estimation. After the CAF module, the reconstruction module \\(\\mathcal{R}\\) has a three-level encoder-decoder architecture. From level 1 to level 3, there are [2,3,4] transformer blocks with [48,96,192] channels. There are [1,2,4] attention heads in the ICA, all with a squeeze factor of 4. Supplementary B.1 gives the detailed architecture of \\(\\mathcal{R}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.787, + 0.55 + ], + "angle": 0, + "content": "Training details. We train using the REDS dataset [27] and follow RealBasicVSR [5] in applying random combinations of blur, noise, JPEG compression, and video compression for synthesizing input degradations. We load 15 frames as an input sequence. The spatial size of inputs is cropped to \\(64 \\times 64\\), and the batch size is 16. We use a pre-trained flow estimation model SPyNet, the parameter of which is fixed for the first 5K iterations and tuned with other modules later." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.787, + 0.675 + ], + "angle": 0, + "content": "Following RealBasicVSR [5], we perform two-stage training. The first stage trains the model with a Charbonnier loss [19] and SSIM [35] loss for 300K iterations. In the second stage, the model is trained for another 130K iterations with the Charbonnier loss, SSIM loss, perceptual loss [18] and GAN loss [11] together, weighted by 1, 0.001, 1, and 0.005, respectively. The implementations of perceptual loss, GAN loss, and discriminator follow RealBasicVSR [5]. We implement all experiments on 4 Quadro RTX 8000 GPUs with PyTorch. Other details of the training settings are in Supplementary B.2." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.695, + 0.442, + 0.71 + ], + "angle": 0, + "content": "4.5 Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We compare our model on four datasets, VideoLQ [5], RealVSR [37], REDS4 [27] and UDM10 [38], with RealBasicVSR [5], Real-ESRGAN [34], BSRGAN [40], DBVSR [28], RealVSR [37], DAN [14], and RealSR [17]. VideoLQ and RealVSR are collected from real-world scenarios. VideoLQ is an unpaired dataset. RealVSR has same-size paired low-quality (LQ) and high-quality (HQ) frames. Our model super-resolves the input frames spatially. Although downsampling LQ enables paired data, it alters the original degradation. Thus, we test on the original LQ and do not use the HQ for evaluation. REDS4 and UDM10 have" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.418, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang and A. Yao" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 2: Quantitative comparisons with existing methods with best and second-best results. Our method achieves the best ILNIQE and NRQM scores over VideoLQ and RealVSR datasets and the best PSNR, SSIM, and LPIPS over synthetic datasets REDS4 and UDM10 with relatively few parameters and the lowest run-time." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.213, + 0.788, + 0.367 + ], + "angle": 0, + "content": "
RealSRDANRealVSRDBVSRBSRGANReal-ESRGANRealBasicVSROurs
Params (M)16.74.32.725.516.716.76.35.3
Runtime (ms)180250772-1801967349
VideoLQILNIQE↓26.6328.2831.9427.8527.4927.9725.9825.94
NRQM↑6.0543.7423.4603.8516.1566.0576.3066.338
RealVSRILNIQE↓32.8132.2934.39-32.6531.9330.3728.61
NRQM↑5.6103.5233.795-6.1526.2456.5826.588
REDS4PSNR↑22.0222.6718.3022.3522.9421.5623.0923.34
SSIM↑0.50970.55710.49000.55300.57500.55560.60760.6079
LPIPS↓0.59910.63150.72400.62110.37660.35330.29910.2877
UDM10PSNR↑25.3725.9023.3525.0825.9724.9625.9626.42
SSIM↑0.66580.72290.71150.71120.75680.74320.74910.7609
LPIPS↓0.48110.47810.47610.47560.33880.33950.32090.3063
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.789, + 0.465 + ], + "angle": 0, + "content": "ground-truth images, and we synthesize low-quality inputs through the same pipeline during training. All tested recurrent-based VSR models load the half sequence for videos in RealVSR and the whole sequence for others each time. The quantitative and qualitative results are discussed below." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.469, + 0.789, + 0.606 + ], + "angle": 0, + "content": "Quantitative Results. For evaluation without reference, we apply the ILNIQE [41], the improved version of NIQE [26], and NRQM [24] metrics based on each output sequence's first, middle, and last frames in RGB format [5]. These metrics appear less biased towards oversharpened features; details are given in Supplementary B.3. For evaluation on REDS4 and UDM10, we report more reliable metrics, PSNR, SSIM, and LPIPS [42]. We collect released models of all compared methods and generate sequences. As shown in Tab. 2, RealVformer performs better than other methods with smaller parameters than the most competitive RealBasicVSR [5] and the shortest runtime." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.639, + 0.331, + 0.783 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.249, + 0.785, + 0.306, + 0.796 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.639, + 0.444, + 0.783 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.785, + 0.412, + 0.796 + ], + "angle": 0, + "content": "(b) RE" + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.639, + 0.556, + 0.783 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.481, + 0.785, + 0.523, + 0.796 + ], + "angle": 0, + "content": "(c) RB" + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.639, + 0.668, + 0.783 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.588, + 0.785, + 0.642, + 0.796 + ], + "angle": 0, + "content": "(d) Ours" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.639, + 0.781, + 0.783 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.706, + 0.785, + 0.749, + 0.796 + ], + "angle": 0, + "content": "(e) GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.808, + 0.787, + 0.838 + ], + "angle": 0, + "content": "Fig. 8: Qualitative comparisons on synthetic datasets. Our method produces clearer than RealESRGAN (RE) [34] and RealBasicVSR (RB) [5] for very hard inputs." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.64, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RealViformer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.144, + 0.36, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.287, + 0.32, + 0.299 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.144, + 0.501, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.41, + 0.287, + 0.454, + 0.299 + ], + "angle": 0, + "content": "(b) RE" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.144, + 0.641, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.287, + 0.593, + 0.299 + ], + "angle": 0, + "content": "(c) RB" + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.144, + 0.783, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.687, + 0.287, + 0.74, + 0.299 + ], + "angle": 0, + "content": "(d) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.31, + 0.784, + 0.34 + ], + "angle": 0, + "content": "Fig. 9: Qualitative comparisons on real-world datasets. Our method produces less high-frequency artifacts and overshoot effects than RealESRGAN [34] and RealBasicVSR [5]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.369, + 0.785, + 0.429 + ], + "angle": 0, + "content": "Qualitative Results. We show qualitative comparisons on synthetic (see Fig. 8) and real-world (see Fig. 9) datasets. Compared to the listed methods, RealViformer generates clear structures with much fewer high-frequency artifacts. More visual comparisons are in Supplementary B.5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.785, + 0.522 + ], + "angle": 0, + "content": "User Study. We also performed a user study and asked 30 evaluators on Amazon MTurk to score reconstructions for 85 frames sampled from datasets VideoLQ [5], RealVSR [37], REDS4 [27] and UDM10 [38]. Each worker saw five HR results of the same frame and rated them based on the visual quality, from 1 (the worst) to 5 (the best); as shown in Fig. 10, our method surpasses BSRGAN [40], Real-ESRGAN [34], RealSR [17], and RealBasicVSR [5]." + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.549, + 0.432, + 0.665 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.679, + 0.447, + 0.749 + ], + "angle": 0, + "content": "Fig.10: User study results from 30 evaluators on 85 frames. Our method achieves the best among all five methods regrading mean option scores (MOS)." + }, + { + "type": "table_caption", + "bbox": [ + 0.452, + 0.568, + 0.785, + 0.652 + ], + "angle": 0, + "content": "Table 3: Ablations of CAF and ICA modules. Channel-attention baseline (ch-baseline) performs better than the spatial-attention baseline (sp-baseline). The CAF module improves the performances for both datasets, and ICA further improves the performances to the state-of-the-art." + }, + { + "type": "table", + "bbox": [ + 0.47, + 0.653, + 0.767, + 0.743 + ], + "angle": 0, + "content": "
MethodCAFICAVideoLQ NRQM↑UDM10 LPIPS↓
Sp-baseline--6.0610.3482
Ch-baselineXX6.1810.3085
RealVformer-✓X6.1960.2933
RealVformer✓✓6.3380.2877
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.841 + ], + "angle": 0, + "content": "Ablation Studies. We conduct ablations to validate the advantage of channel attention and the efficacy of the CAF and ICA. We build a spatial-attention baseline (sp-baseline) by replacing the reconstruction module of BasicVSR [3] with SwinIR [21]. Channel-attention baseline (ch-baseline) is built with the same" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.418, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang and A. Yao" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.149, + 0.765, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.269, + 0.788, + 0.313 + ], + "angle": 0, + "content": "Fig. 11: (a) Visual comparison between RealVformer and its ablations. Red circles highlight the improved details. (b) Radial Power Spectrum (RPS) of model predictions. Using ICA improves the power of high-frequency components (blue region)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.341, + 0.789, + 0.525 + ], + "angle": 0, + "content": "overall architecture as RealVformer but replaces CAF with simple concatenation and substitutes ICA with the channel attention block in Restormer [39]. RealVformer\\(^{-}\\) applies CAF on ch-baseline. All models are trained with the same settings in Sec. 4.4. We report NRQM scores for the VideoLQ [5] dataset and LPIPS scores for the UDM10 [38] da set. As shown in Tab. 3, using the original channel attention module, ch-baseline already yields better performances than sp-baseline. The CAF and ICA modules further improve RealVformer to state-of-the-art performance with channel correlation of propagated information decreasing from 0.436 to 0.422. Fig. 11a visually compares RealVformer with its ablations. Applying CAF reduces artifacts, while ICA provides further improvements. Fig. 11b supplements the Radial Power Spectrum of model predictions. The blur region shows ICA increases the power in the high-frequency region." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.545, + 0.36, + 0.56 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.577, + 0.789, + 0.789 + ], + "angle": 0, + "content": "This paper proposes a real-world VSR model, RealVformer, based on findings from investigating channel and spatial attention in a real-world setting. Explorations reveal that channel attention is less sensitive to the artifacts in query and better serves as a temporal aggregation module to limit model-produced artifacts in hidden states. Additionally, we observe the higher covariance of channel attention outputs and propose the Improved Channel Attention (ICA) Module with a squeeze-and-excite and a covariance-based rescaling mechanism. Based on our findings, we build RealVformer, a channel-attention-based recurrent model for real-world VSR. We propose the CAF module to limit artifact propagation and use the ICA module to achieve better reconstructions. RealVformer performs state-of-the-art on two real-world video datasets with fewer parameters and shorter runtime. On the other hand, we value our findings w.r.t. comparisons between channel and spatial attention and exploration of covariance in channel attention as inspiration for further real-world VSR research." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Acknowledgement This research is supported by the National Research Foundation, Singapore under its NRF Fellowship for AI (NRF-NRFFAI1-2019-0001)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.64, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "RealViformer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.194 + ], + "angle": 0, + "content": "Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of National Research Foundation, Singapore." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.221, + 0.326, + 0.238 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.258, + 0.785, + 0.286 + ], + "angle": 0, + "content": "1. Bardes, A., Ponce, J., LeCun, Y.: Vicreg: Variance-invariance-covariance regularization for self-supervised learning. arXiv preprint arXiv:2105.04906 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.785, + 0.315 + ], + "angle": 0, + "content": "2. Cao, J., Li, Y., Zhang, K., Van Gool, L.: Video super-resolution transformer. arXiv preprint arXiv:2106.06847 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.317, + 0.785, + 0.372 + ], + "angle": 0, + "content": "3. Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: Basicvsr: The search for essential components in video super-resolution and beyond. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4947-4956 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.373, + 0.785, + 0.428 + ], + "angle": 0, + "content": "4. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5972-5981 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.43, + 0.785, + 0.47 + ], + "angle": 0, + "content": "5. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Investigating tradeoffs in real-world video super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5962-5971 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.472, + 0.785, + 0.514 + ], + "angle": 0, + "content": "6. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12299-12310 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.515, + 0.785, + 0.556 + ], + "angle": 0, + "content": "7. Chen, X., Wang, X., Zhou, J., Qiao, Y., Dong, C.: Activating more pixels in image super-resolution transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22367-22377 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.558, + 0.785, + 0.598 + ], + "angle": 0, + "content": "8. Cogswell, M., Ahmed, F., Girshick, R., Zitnick, L., Batra, D.: Reducing overfitting in deep networks by decorrelating representations. arXiv preprint arXiv:1511.06068 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.6, + 0.785, + 0.641 + ], + "angle": 0, + "content": "9. Dai, T., Cai, J., Zhang, Y., Xia, S.T., Zhang, L.: Second-order attention network for single image super-resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11065-11074 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.643, + 0.785, + 0.684 + ], + "angle": 0, + "content": "0. Fuoli, D., Gu, S., Timofte, R.: Efficient video super-resolution through recurrent latent space propagation. In: 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW). pp. 3476-3485. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.686, + 0.785, + 0.727 + ], + "angle": 0, + "content": "1. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.728, + 0.785, + 0.755 + ], + "angle": 0, + "content": "2. Hu, J., Shen, L., Sun, G.: Squeeze-and-excitation networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7132-7141 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "3. Hua, T., Wang, W., Xue, Z., Ren, S., Wang, Y., Zhao, H.: On feature decorrelation in self-supervised learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9598-9608 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "4. Huang, Y., Li, S., Wang, L., Tan, T., et al.: Unfolding the alternating optimization for blind super resolution. Advances in Neural Information Processing Systems 33, 5632-5643 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.258, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.416, + 0.128 + ], + "angle": 0, + "content": "Y. Zhang and A. Yao" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "15. Isobe, T., Jia, X., Gu, S., Li, S., Wang, S., Tian, Q.: Video super-resolution with recurrent structure-detail network. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16. pp. 645-660. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.787, + 0.257 + ], + "angle": 0, + "content": "16. Jeelani, M., Cheema, N., Illgner-Fehns, K., Slusallek, P., Jaiswal, S., et al.: Expanding synthetic real-world degradations for blind video super resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1199-1208 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.257, + 0.787, + 0.299 + ], + "angle": 0, + "content": "17. Ji, X., Cao, Y., Tai, Y., Wang, C., Li, J., Huang, F.: Real-world super-resolution via kernel estimation and noise injection. In: proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 466-467 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.298, + 0.787, + 0.354 + ], + "angle": 0, + "content": "18. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 694-711. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.354, + 0.787, + 0.393 + ], + "angle": 0, + "content": "19. Lai, W.S., Huang, J.B., Ahuja, N., Yang, M.H.: Fast and accurate image superresolution with deep laplacian pyramid networks. IEEE transactions on pattern analysis and machine intelligence 41(11), 2599-2613 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.787, + 0.42 + ], + "angle": 0, + "content": "20. Liang, J., Cao, J., Fan, Y., Zhang, K., Ranjan, R., Li, Y., Timofte, R., Van Gool, L.: Vrt: A video restoration transformer. arXiv preprint arXiv:2201.12288 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.787, + 0.461 + ], + "angle": 0, + "content": "21. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1833-1844 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.787, + 0.514 + ], + "angle": 0, + "content": "22. Liang, J., Fan, Y., Xiang, X., Ranjan, R., Ilg, E., Green, S., Cao, J., Zhang, K., Timofte, R., Gool, L.V.: Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems 35, 378-393 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.787, + 0.57 + ], + "angle": 0, + "content": "23. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.787, + 0.61 + ], + "angle": 0, + "content": "24. Ma, C., Yang, C.Y., Yang, X., Yang, M.H.: Learning a no-reference quality metric for single-image super-resolution. Computer Vision and Image Understanding 158, 1-16 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.787, + 0.651 + ], + "angle": 0, + "content": "25. Mei, Y., Fan, Y., Zhou, Y.: Image super-resolution with non-local sparse attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3517-3526 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.787, + 0.678 + ], + "angle": 0, + "content": "26. Mittal, A., Soundararajan, R., Bovik, A.C.: Making a \"completely blind\" image quality analyzer. IEEE Signal processing letters 20(3), 209-212 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.787, + 0.733 + ], + "angle": 0, + "content": "27. Nah, S., Baik, S., Hong, S., Moon, G., Son, S., Timofte, R., Mu Lee, K.: Ntire 2019 challenge on video deblurring and super-resolution: Dataset and study. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.733, + 0.787, + 0.772 + ], + "angle": 0, + "content": "28. Pan, J., Bai, H., Dong, J., Zhang, J., Tang, J.: Deep blind video super-resolution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4811-4820 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.813 + ], + "angle": 0, + "content": "29. Ranjan, A., Black, M.J.: Optical flow estimation using a spatial pyramid network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4161-4170 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.841 + ], + "angle": 0, + "content": "30. Shi, S., Gu, J., Xie, L., Wang, X., Yang, Y., Dong, C.: Rethinking alignment in video super-resolution transformers. arXiv preprint arXiv:2207.08494 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.64, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RealViformer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "31. Song, Y., Wang, M., Yang, Z., Xian, X., Shi, Y.: Negvsr: Augmenting negatives for generalized noise modeling in real-world video super-resolution. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 10705-10713 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "32. Wang, H., Chen, X., Ni, B., Liu, Y., Liu, J.: Omni aggregation networks for lightweight image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22378-22387 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.787, + 0.286 + ], + "angle": 0, + "content": "33. Wang, X., Chan, K.C., Yu, K., Dong, C., Change Loy, C.: Edvr: Video restoration with enhanced deformable convolutional networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.787, + 0.328 + ], + "angle": 0, + "content": "34. Wang, X., Xie, L., Dong, C., Shan, Y.: Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1905-1914 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "35. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.37, + 0.787, + 0.397 + ], + "angle": 0, + "content": "36. Xie, L., Wang, X., Shi, S., Gu, J., Dong, C., Shan, Y.: Mitigating artifacts in real-world video super-resolution models. arXiv preprint arXiv:2212.07339 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.397, + 0.787, + 0.453 + ], + "angle": 0, + "content": "37. Yang, X., Xiang, W., Zeng, H., Zhang, L.: Real-world video super-resolution: A benchmark dataset and a decomposition based learning scheme. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4781-4790 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.453, + 0.787, + 0.508 + ], + "angle": 0, + "content": "38. Yi, P., Wang, Z., Jiang, K., Jiang, J., Ma, J.: Progressive fusion video superresolution network via exploiting non-local spatio-temporal correlations. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3106-3115 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.508, + 0.787, + 0.563 + ], + "angle": 0, + "content": "39. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5728-5739 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.564, + 0.787, + 0.605 + ], + "angle": 0, + "content": "40. Zhang, K., Liang, J., Van Gool, L., Timofte, R.: Designing a practical degradation model for deep blind image super-resolution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4791-4800 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.605, + 0.787, + 0.633 + ], + "angle": 0, + "content": "41. Zhang, L., Zhang, L., Bovik, A.C.: A feature-enriched completely blind image quality evaluator. IEEE Transactions on Image Processing 24(8), 2579-2591 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.633, + 0.787, + 0.674 + ], + "angle": 0, + "content": "42. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.716 + ], + "angle": 0, + "content": "43. Zhang, Y., Li, K., Li, K., Wang, L., Zhong, B., Fu, Y.: Image super-resolution using very deep residual channel attention networks. In: Proceedings of the European conference on computer vision (ECCV). pp. 286-301 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.716 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/RealViformer_ Investigating Attention for Real-World Video Super-Resolution/39444a96-c968-4d03-9b6c-b416cbbaab18_origin.pdf b/2024/RealViformer_ Investigating Attention for Real-World Video Super-Resolution/39444a96-c968-4d03-9b6c-b416cbbaab18_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2a492ba4b9920e07f45bc9546a61b5d536774646 --- /dev/null +++ b/2024/RealViformer_ Investigating Attention for Real-World Video Super-Resolution/39444a96-c968-4d03-9b6c-b416cbbaab18_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:04d69e9a4457b4c9a74f702cc33a9055888e81a39cc3c85f14506a857c176e93 +size 7996537 diff --git a/2024/RealViformer_ Investigating Attention for Real-World Video Super-Resolution/full.md b/2024/RealViformer_ Investigating Attention for Real-World Video Super-Resolution/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f1d3abbb9ca036b9250fa504b926135000f1d3c6 --- /dev/null +++ b/2024/RealViformer_ Investigating Attention for Real-World Video Super-Resolution/full.md @@ -0,0 +1,298 @@ +# RealVformer: Investigating Attention for Real-World Video Super-Resolution + +Yuehan Zhang and Angela Yao + +National University of Singapore {zyuehan,ayao}@comp.nus.edu.sg + +Abstract. In real-world video super-resolution (VSR), videos suffer from in-the-wild degradations and artifacts. VSR methods, especially recurrent ones, tend to propagate artifacts over time in the real-world setting and are more vulnerable than image super-resolution. This paper investigates the influence of artifacts on commonly used covariance-based attention mechanisms in VSR. Comparing the widely-used spatial attention, which computes covariance over space, versus the channel attention, we observe that the latter is less sensitive to artifacts. However, channel attention leads to feature redundancy, as evidenced by the higher covariance among output channels. As such, we explore simple techniques such as the squeeze-excite mechanism and covariance-based rescaling to counter the effects of high channel covariance. Based on our findings, we propose RealVformer. This channel-attention-based real-world VSR framework surpasses state-of-the-art on two real-world VSR datasets with fewer parameters and faster runtimes. The source code is available at https://github.com/Yuehan717/RealVformer. + +# 1 Introduction + +Video super-resolution (VSR) recovers a high-resolution (HR) sequence of frames from its low-resolution (LR) counterpart. Recurrent convolutional approaches are commonly used in VSR with standard settings, assuming the LR frames are downsampled from HR frames with known kernels. [3, 10, 15]. However, in real-world VSR (RWVSR), the low-resolution videos are not simply downsampled versions of their high-resolution counterparts. Instead, they feature complex degradations that arise from the camera imaging system, compression, internet transmission, and other factors. These degradations make architecture design for RWVSR challenging, as artifacts and degradations tend to propagate and get exaggerated over the recurrent connection [5, 36]. + +Recently, transformer architectures have replaced convolutional architectures as state-of-the-art for standard VSR. While attention mechanisms have replaced convolution operations, most methods retain the recurrent connection to aggregate the information over time [22, 30]. Yet, such architectures do not always perform well on RWVSR. For instance, a Swin-based model [21] designed for standard VSR, when applied to a real-world input frame, as shown in Fig. 1a, generates more artificial lines than the convolutional model RealBasicVSR [5]. + +![](images/110ffaa2f809c8fbaaf8966a162e998bca4cf29b8f6bfdc7e6bbbdf0a92444b6.jpg) +Input + +![](images/456b3246abbab733088a604dbc47e275c35a887c21e6f92da2873f6033cbd83a.jpg) +RealBasicVSR + +![](images/84f5cc884e9e4d8e0b056fc7d3d08be8a7ec48437a2e83bf577871e3316a5ea4.jpg) +Swin-based VSR + +![](images/bf3b02d4f19ffbccf955eee78c521b1bbb179ef40d04958aeeb88a13715baad0.jpg) +Ours: RealVformer + +![](images/63c21eeb4eedbfec929404f0e2ce863925a932efc469e60750bb571e63ddffbb.jpg) +(a) Visual comparisons. +W +Spatial Attention + +![](images/13508f59141d52d449e745094a9abb701d287e42aa305cbada4bb50442108113.jpg) +H +Channel Attention +(b) Schematics of two attentions. +Fig. 1: (a) Designing a RWVSR transformer is not trivial. A Swin-based transformer suited for standard VSR hallucinates more lines than a RealBasicVSR, a convolutional state-of-the-art. We propose RealVformer based on our investigation of attention under the RWVSR setting. RealVformer generates details with fewer artifacts than RealBasicVSR [3] and the Swin-based VSR model. (b) Schematic for spatial and channel attention. Spatial attention aggregates features based on pixel representations. Channel attention takes $H \times W$ feature map for matching across channels. + +Why should transformers perform well on (synthetic) standard VSR but poorly in real-world cases? We speculate that standard VSR transformers benefit from the similarity-based matching of the attention mechanism, which accurately aggregates information spatially and temporally [21,22,30]. However, when input degradation exists, the aggregated information becomes less reliable because the attention queries may be derived from both true source video and artifacts. + +This work investigates and sheds light on the sensitivity of attention in real-world settings. We compare two covariance-based attention mechanisms used in low-level Transformers: spatial attention [21,30] and channel attention [39]. Spatial attention takes pixel-wise features as keys and queries and estimates their covariances across spatial positions. The most popular form is window-based attention [6]; The shift-window scheme from Swin Transformers [23] enables the model to access distant spatial ranges [2,21,30] without computational blow-up. Spatial attention is widely used for video and image super-resolution, albeit in the standard setting. Channel attention [39] estimates covariances across channels (see Fig. 1b) and collapses the spatial extent of a feature map. It defines the number of queries and keys by channel numbers rather than spatial resolution. Although established for deblurring or denoising [39], channel attention's efficacy in super-resolution remains to be determined. + +Our experiments show that channel attention is less sensitive to artifacts than the spatial counterparts, resulting in higher performance gains in RWVSR. However, it is also revealed that channel attention leads to feature channels with higher covariance. From a learning perspective, a high covariance is undesirable because it is a strong indicator of feature redundancy [1,8,13]. Therefore, using channel attention naively will have limited improvements over existing RWVSR + +state-of-the-art methods. Such a finding has wide-reaching impacts as channel attention is used increasingly for low-level vision. + +To verify our findings, we explore established mechanisms to counter the effects of feature redundancy - simple techniques such as squeeze-excite and covariance-based rescaling improve the vanilla channel attention design. From these outcomes, we propose RealVformer, a new transformer real-world VSR model. RealVformer performs channel attention between the current frame feature and the propagated hidden state to limit model-produced artifacts. The model then reconstructs features through improved channel attention modules featuring squeeze-and-excite and covariance-based channel rescaling mechanisms. With our effective designs, RealVformer achieves state-of-the-art performance with fewer parameters on challenging synthetic video datasets and two real-world video datasets collected from different scenes. + +Summarizing our contributions in order of importance, our paper + +- investigates the differences between spatial and channel attention for RWVSR. Spatial attention, although widely used, is revealed to be highly sensitive to the noise and degradations common in RWVSR sequences, while channel attention is more robust. +- reveals that naively applying channel attention increases channel covariance, which is problematic for learning; this overlooked fact has a wide-reaching impact as channel attention becomes more used in low-level vision. +- empirically verifies the negative effect of high channel covariance by countering it with established techniques, based on which we develop the RealVformer for RWVSR. Our simple modification surpasses state-of-the-art despite using less compute. + +# 2 Related Work + +Standard Video Super-Resolution models focus on architecture design to use temporal information better [2-4, 15]. Previous research starts from sliding-window-based [15,33] to recurrent-based frameworks [3,4] for using distant-frame information. Recent works introduce Transformer blocks into existing recurrent frameworks to overcome the locality limitation of convolution and accurately match abundant information for feature reconstruction [20,22,30]. + +Real-world video super-resolution focuses on modeling, removing, and limiting the impact of real-world degradations. Existing works are convolutional models and focus on designing losses or modules for degradation processing. DBVSR [28] explicitly estimates the degradation kernel through a sub-network. RealBasicVSR [5] tries to 'clean' artifacts through a processing module for BasicVSR [3]. In a similar approach, FastRealVSR [36] borrows an external pool of blur and sharpening filters to 'clean' the hidden states. Other recent works [16,31] advance the synthesis method for paired training data. Instead, we focus on investigating the function of attention in RWVSR rather than architecture design or data synthesis. + +Attention mechanisms have been widely applied for low-level vision tasks [9, 25, 43, 43]. Transformers with covariance-based attention are the most prevalent [21, 30, 39] for standard VSR. Most existing methods adopt shift-window-based spatial attention [23] to aggregate information from other positions within or across frames. In contrast, Restormer [39] computes the covariance among channels and shows its effectiveness for multiple image restoration tasks. More recent works stitch spatial and channel attention together to enlarge the receptive field [7, 32] for standard image super-resolution. Instead, we investigate attention mechanisms in terms of their sensitivity to real-world degradation for the first time and develop an effective real-world VSR Transformer. + +# 3 Explorations on Attention + +Sec. 3.1 defines the VSR task and the two attention mechanisms. Sec. 3.2 compares the channel and spatial attentions' sensitivity to query artifacts and effects on real-world VSR performance. Sec. 3.3 reveals that channel attention leads to higher covariance among channels and explores mitigating options. + +# 3.1 Preliminaries + +Standard vs. Real-World VSR. Given a low-resolution (LR) video sequence with $T$ frames $I^L \in \mathcal{R}^{T \times H \times W \times K}$ , VSR models reconstruct a high-resolution sequence $I^H \in \mathcal{R}^{T \times sH \times sW \times K}$ , where $H \times W$ is the input spatial resolution, $K$ is the number of input channels and $s$ is the scaling factor. In the standard setting of VSR, $I_t^L$ , where $t \in \{0, \dots, T\}$ , is assumed as the downsampled version of $I_t^H$ , defined as: $I_t^L = (I_t^H) \downarrow_{\frac{1}{s}}$ . Both training and testing datasets follow this formulation to generate LR-HR pairs given $I^H$ . In a real-world setting, there is no closed formulation for the relationship between $I_t^L$ and $I_t^H$ due to the unknown distribution of real-world degradations. Real-ESRGAN [34] proposed a widely-used training setting that randomly applies synthesized blur, noise, compression, and resizing to the HR frames to generate paired LR frames with complex degradations. The testing datasets are either synthesized by the same pipeline in training or collected from diverse real-world sources [5, 37]. The synthesized testing sets have paired ground-truth sequences and are evaluated by full-reference metrics, e.g. PSNR and LPIPS [42]; real-world datasets are always without ground truth and require no-reference metrics, such as NRQM [24]. + +Attention Definitions. In Transformers, the attention modules project layer-normalized tensor $X \in \mathbb{R}^{C \times H \times W}$ to query $Q$ , and tensor $Y \in \mathbb{R}^{\hat{C} \times \hat{H} \times \hat{W}}$ to key $K$ and value $V$ , where $H \times W$ and $\hat{H} \times \hat{W}$ are the spatial resolution of the normalized tensors. The attention map $\mathcal{A}$ is generated by calculating the covariance between $Q$ and $K$ , followed by a softmax function, before being applied to the value $V$ to produce output $O$ . Spatial and channel attention differ in tensor dimension taken for the covariance calculation. + +Spatial attention generates an $\mathbb{R}^{HW\times \hat{H}\hat{W}}$ attention map, $\mathcal{A}_s$ , by computing the covariance between features in the query $(Q_{s})$ and key $(K_{s})$ at each spatial position. The query, key and values are computed as $Q_{s} = W_{s}^{Q}X$ , $K_{s} = W_{s}^{K}Y$ , $V_{s} = W_{s}^{V}Y$ , where $W_{s}^{Q}\in \mathbb{R}^{D_{s}\times C}$ , $\{W_s^K,W_s^V\} \in \mathbb{R}^{D_s\times \hat{C}}$ , and $D_{s}$ is the dimension of projections. The attention map $\mathcal{A}_s$ and output attention features $O_{s}$ are: + +$$ +\mathcal {A} _ {s} = \operatorname {s o f t m a x} \left(Q _ {s} ^ {T} K _ {s} / \sqrt {D _ {s}}\right), O _ {s} = \mathcal {A} _ {s} V _ {s} ^ {T}, \tag {1} +$$ + +where $Q_{s}(K_{s})$ is reshaped to matrix of $\mathbb{R}^{D_s\times HW}$ $(\mathbb{R}^{D_s\times \hat{H}\hat{W}})$ . We omit the later reshaping operation for brevity. $\mathcal{A}_s$ exhaustively relates every spatial location to every other location in space. A window-based implementation [21] limits the correlations to windows of $\omega \times \omega$ , reducing the map size to $\mathbb{R}^{\omega^2\times \omega^2}$ . + +Channel attention applies convolutions on $X$ to get query $(Q_{c})$ and $Y$ for key $(K_{c})$ and value $(V_{c})$ [39], after which features are the same in spatial resolution (assumed to be $H \times W$ in the following context). It estimates covariance across channels to yield a $C \times \hat{C}$ sized map $\mathcal{A}_{c}$ and output features $O_{c}$ : + +$$ +\mathcal {A} _ {c} = \operatorname {s o f t m a x} \left(Q _ {c} K _ {c} ^ {T} / \alpha\right), O _ {c} = \mathcal {A} _ {c} V _ {c}, \tag {2} +$$ + +where $Q_{c}$ is reshaped to $\mathbb{R}^{C\times HW}$ , $K_{c}$ and $V_{c}$ are reshaped to $\mathbb{R}^{\hat{C}\times HW}$ and $\alpha$ is learnable scaling parameter. Because channel attention computes correlations between features of size $\mathbb{R}^{1\times HW}$ , it has a larger spatial context than the $\mathbb{R}^{D_s\times 1}$ -sized features in spatial attention (see Fig. 1b). + +# 3.2 Attention in Real-World VSR + +Sensitivity to Query Artifacts. In standard VSR, attention is used to aggregate information temporally and spatially [22, 30]. The assumption is that queries can match beneficial cues for super-resolution. But what if the queries themselves are unreliable? We speculate this to be true in real-world VSR, where inputs have artifacts and degradations. + +How affected are spatial and channel attention mechanisms by query artifacts? Fig. 2 shows how we compare, using cosine similarity, the attention outputs based on queries from the same frame with and without certain degradations. Using encoding layers from a standard convolutional VSR model, we perform the attention operations defined in Sec. 3.1 by taking embeddings of frame $I_{t}$ as query and $I_{t-1}$ as key and value. $I_{t}$ and $I_{t-1}$ are clean without degradation other than downsampling, and we represent the output feature of the attention module as $O$ . When additional degradation $D_{i}$ is applied to $I_{t}$ , we represent the corresponding output as $O_{D_{i}}$ . A smaller deviation of $O_{D_{i}}$ from $O$ indicates lower sensitivity to artifacts in the query. + +Tab. 1 shows the cosine similarity between $O$ and $O_{D_i}$ for spatial and channel attention modules. Curiously, $O_{D_i}$ based on the channel attention module is more similar to the $O$ matched by the degradation-free query, indicating that channel attention is less sensitive to query artifacts. Intuitively, the lower sensitivity of channel attention is related to the larger spatial context used for feature matching. Given a deep feature of size $\mathbb{R}^{C\times H\times W}$ , channel attention uses feature sized + +![](images/304df22f113c50b298215525302e7655239221a7f78137f645f39c0b2edcf9c3.jpg) +Fig.2: Schematic for sensitivity comparison. $I_{t-1}, I_t$ are downsampled but clean frames at times $t$ and $t-1$ . $D_i(.)$ apply degradations to $I_t$ , where $D_i \in \{\text{blur, noise, compression}\}$ . $O$ and $O_{D_i}$ are output features of the attention module. Queries are from the embedding at time $t$ , and keys and values are from time $t-1$ . Higher cosine similarities $S$ between attention output features $O$ and $O_{D_i}$ reflect less sensitivity to artifacts in queries. + +Table 1: Cosine similarity between $O$ and ${O}_{{D}_{i}}$ ,attention outputs without and with query degradation. Outputs of channel attention change less under query degradation. + +
ModuleDi
blurnoisecompression
Spatial attention0.750.920.84
Channel attention0.980.990.99
+ +$\mathbb{R}^{1 \times HW}$ to calculate the covariance across channels. As such, feature aggregation is based on global information observed in a large normalized spatial context. Instead, the covariance of spatial attention is for features at each location sized $\mathbb{R}^{C \times 1}$ , so it is likely more sensitive to local value changes from artifacts. + +Impact on VSR Models. While channel attention output features are less sensitive to query artifacts than spatial attention, how well do they fare for real-world VSR? We experiment by incorporating these attention variants into VSR models. We focus specifically on recurrent pipelines due to their popularity in standard and real-world VSR [5,22,36]. In recurrent pipelines, artifacts in hidden states may propagate over time, get exaggerated, and negatively influence overall performance [5, 36]. Therefore, we explicitly add the attention module between features of the current frame and the propagated hidden state. Ideally, the added attention will select relative information from hidden states for the current frame and reduce the propagation of model-produced artifacts. + +Fig. 3a shows the baseline model which concatenates $f_{t}$ , the shallow feature of at time $t$ , and $\hat{h}_{t-1}$ , the spatially aligned hidden state at time $t-1$ . We experiment with two variants by using a channel-attention module $G_{c}$ or a spatial-attention module $G_{s}$ to replace the concatenation in the baseline. As shown in Fig. 3b, the query is generated by $f_{t}$ , and the key and value are predicted from $\hat{h}_{t-1}$ . The attention output $O_{t}^{A}$ is concatenated with $f_{t}$ as the input for $\mathcal{R}$ . + +All models are trained on the REDS dataset [27] with the random degradation pipeline from Real-ESRGAN [34]. We test on REDS4 [27] with different types and extents of degradations, including Gaussian blur, Gaussian noise, and JPEG + +![](images/35e515ed0362ecb528cc6ce0b23a682ab670ee9ec95f37a8a860fa37e88d83df.jpg) +(a) Recurrent Baseline + +![](images/62244e9bb5b9c9de03ad155c4fe629c5dc01b194dcc6355fc3b5318a6e7cb16c.jpg) +(b) Attention + +![](images/bed2f5adfd4a7b2449c7e08a60b2a56d3bcb866ffe0709cdd0166b87ecb4fb98.jpg) +Fig. 3: (a) The recurrent baseline in Sec. 3.2 has a shallow mapping module $\mathcal{F}$ , reconstruction module $\mathcal{R}$ , upsampling module $\mathcal{U}$ and warping function $W$ . $W$ aligns the hidden state $h_{t-1}$ to feature at $t$ based on optical flow $s_{(t-1)\rightarrow t}^{f}$ . All residual blocks are convolutional. The concatenation between $f_{t}$ and $\hat{h}_{t-1}$ are replaced with the spatial or channel attention modules in (b) to compare the effect of attention. (b) The attention module first applies layer normalization to $f_{t}$ and $\hat{h}_{t-1}$ and then performs channel or spatial attention according to Sec. 3.1. The output feature $O_{t}^{A}$ concatenated with $f_{t}$ is processed by the module $\mathcal{R}$ in (a). + +![](images/897c696d2d587de5dcf9c83c5f2b51e8c89185ba5156b647fa03e5f671ae4d3f.jpg) +Fig. 4: Comparison of spatial and channel attention through impact on the performance of real-world VSR model. The Y-axis shows improvements compared to the convolutional baseline. A lower LPIPS score is better. The channel attention module is the best except for the PSNR score of highly blurred inputs. + +![](images/ca67627848bee6f9524c8382dc94841d5c12b544f459775e56c1e468838bbfac.jpg) + +compression. Fig. 4 displays changes in PSNR and LPIPS compared with the baseline model. Trained on the same dataset and degradation setting, channel attention between temporal information yields better objective and perceptual reconstruction quality than spatial attention and baseline for noise and JPEG compression inputs. Channel attention still achieves better scores for blurred inputs, except for the PSNR performance of severely blurred inputs. + +# 3.3 Limitations of Channel Attention + +The results in Sec. 3.2 indicate that channel attention is better suited for propagating information over time. However, it also has an inherent flaw - the correlation among output channels will increase since each channel in the attention output is a weighted summation over channels of the value $V_{c}$ . We calculate the covariance following VICReg [1] for a quantitative measurement. Given deep + +feature $z^n \in \mathbb{R}^{C \times H \times W}$ , where $n \in \{1, N\}$ is the index of a sample, we reshape $z^n$ to $\mathbb{R}^{C \times HW}$ and define the covariance matrix over $Z = \{z^1, \dots, z^N\}$ as: + +$$ +C o v (Z) = \frac {1}{N - 1} \sum_ {n = 1} ^ {N} \left(z _ {n} - \bar {z}\right) \left(z _ {n} - \bar {z}\right) ^ {T}, \tag {3} +$$ + +where $\bar{z} = \frac{1}{N}\sum_{n=1}^{N}z_{n}$ . We then define the indicator for the covariance matrix as $ac(Z) = \frac{1}{d}\sum_{i\neq j}|Cov(Z)|_{i,j}$ , i.e. the average of absolute off-diagonal coefficients of $Cov(Z)$ , where $d$ is the number of off-diagonal coefficients. Function $ac(Z)$ encodes the covariance among feature channels. Taking $O$ in Fig. 2 for comparison, $ac(O)$ with $N = 400$ based on channel attention is 0.87 and significantly higher than the input features ( $\approx 0.15$ ). Instead, $ac(O)$ of spatial attention remains similar to the input features. Previous works on representation learning [1] and overfitting [8] propose that the high covariance of feature channels indicates redundancy and tends to hinder informative predictions. + +Similarly, we speculate the redundancy effects will hinder the prediction of HR outputs when adopting channel attention in building blocks. For a closer look, we investigate the standard VSR task, which is artifact-free. Specifically, we build $M_{c}$ and $M_{s}$ by replacing the Residual Block in Fig. 3a with channel attention blocks [39] and spatial attention blocks [21]. Models are trained on the REDS dataset without extra degradation. For evaluation, we choose SSIM [35], which focuses on structural information. The SSIM score of $M_{c}$ (0.8338) is lower than $M_{s}$ (0.8432); $ac(\cdot)$ of the last features before the upsampling module is higher for channel attention than for spatial attention, i.e. 0.199 vs. 0.147. + +We adopt two simple modifications to channel attention and boost informative features to verify the redundancy effect empirically. Fig. 5 shows the Improved Channel Attention (ICA) module. Our approach features two key steps to enhance the quality. First, we use the squeeze-and-excite mechanism to predict new information. The features are squeezed channel-wise to extract meaningful + +![](images/1749f415e521704d711882a6df1461ef863dc7988a82718fb792f2f63de5a078.jpg) +Fig. 5: Improved Channel Attention Module (ICA), showing self-attention for simplicity. The 'squeeze' convolution compresses the number of input feature channels $X \in \mathbb{R}^{C \times H \times W}$ by ratio $r$ . The features are then rescaled by weights predicted from the $\frac{C}{r} \times \frac{C}{r}$ attention map before being expanded by the 'excite' convolution back to the original number of input channels. + +![](images/e4aba276499b6b85f5db0256db27dfe9fbb641783a645a51bccee794f785f330.jpg) +(c) Reconstruction Module +Fig. 6: The framework of RealVformer. (a) Overview of RealVformer, following a unidirectional recurrent framework. The outputs of the Forward module are propagated to the next time step and upsampled by module $\mathcal{U}$ to get HR frames. (b) Explanation of the Forward module in (a), where $W$ denotes the warping function. The reconstruction module $\mathcal{R}$ takes current frame $I_t^L$ and warped hidden state $\hat{h}_{t-1}$ as inputs. (c) Reconstruction module $\mathcal{R}$ . The shallow feature of $I_t^L$ and $\hat{h}_{t-1}$ are fused by CAF and then forwarded to Transformer blocks with U-shape connection [39]. Module GDFN follows Restormer [39]. Details of CAF and ICA modules are stated in Fig. 7 and Fig. 5. + +information and then expanded into new channels based on the attention outputs. Secondly, we rescale channels in attention outputs by scalar weights predicted from the attention map. The attention map measures relationships across channels and encourages the associated convolutions in the 'excite' operation to yield more precise and discriminating features. Our designs are inspired by the SE Network [12] but with two key distinctions. First, we apply the squeeze-and-excite mechanism to generate new information, while SE Network uses it to help with scalar weight prediction. Second, the rescaling weights prediction takes the attention map as a cue rather than the naive pooling results of channels. In Sec. 4, we show the efficacy of new designs on the real-world VSR task and compare channel correlation with and without new designs in ablations. + +# 4 RealVifromer + +# 4.1 Overview + +We apply our findings in Sec. 3.2 to the real-world VSR task as further support. To that end, we design RealVformer to incorporate channel attention as the basic processing module and the modification to boost informative features. Our emphasis here is not novel architecture design but a showcase of effectiveness brought by applying our findings in Sec. 3.2. + +RealVformer is a recurrent Transformer network with channel attention modules. Fig. 6 shows the model architecture. To reduce the computational + +Channel Attention Fusion module (CAF) +Fig. 7: Details of Channel Attention Fusion (CAF) module. CAF gets the query from current frame feature $f_{t}$ and {key, value} from hidden state $\hat{h}_{t-1}$ . The attention output is concatenated with $f_{t}$ to process for module output $O_{t}$ . +![](images/71297bf3b076a0ea814a16d09bf0c2e24b20f50ea114312c4bb2e3709fbd802c.jpg) +1x1 Norm 3x3 Depth-wise $\mathbb{C}$ Concatenation Reshape $\bigotimes$ Matrix multiplication Conv + +cost, we follow a widely used recurrent architecture [3] in a unidirectional setting. The model first estimates optical flow $s_{(t - 1)\rightarrow t}^{f}$ from $I_{t - 1}^{L}$ to $I_{t}^{L}$ through Spynet [29] and warps previous hidden state $h_{t - 1}$ to current time step based on the flow. Frame $I_t^L$ and spatially aligned hidden state $\hat{h}_{t - 1}$ are processed in the reconstruction module $\mathcal{R}$ . The hidden state is updated with outputs of $\mathcal{R}$ and further processed by the upsampling module $\mathcal{U}$ to output high-resolution frames. + +The reconstruction module $\mathcal{R}$ uses channel attention in two ways. First, the Channel Attention Fusion (CAF) module fuses the temporal information to limit the produced artifacts in the hidden state. CAF queries the aligned hidden state $\hat{h}_{t-1}$ by the shallow feature $f_t$ . Secondly, we take the Improved Channel Attention Module (ICA) in Fig. 5 to build the Transformer blocks for better HR reconstructions. + +# 4.2 Channel Attention Fusion (CAF) Module + +As explored in Sec. 3.2, adding channel attention promotes the real-world VSR performance compared to spatial attention and the simple concatenation baseline. Thus, we keep this design in our model and put it as the Channel Attention Fusion (CAF) module for temporal aggregation. Fig. 7 show details of how CAF perform channel attention between $f_{t}$ , the feature of the current frame, and $\hat{h}_{t-1}$ , the aligned propagated hidden state. The query $(Q_{t})$ is generated as $Q_{t} = K_{3\times 3} * \text{LayerNorm}(f_{t})$ , where $K_{3\times 3} *$ refers to $3 \times 3$ convolution operation. Similarly, we process hidden state $\hat{h}_{t-1}$ by $\tilde{h}_{t-1} = K_{3\times 3}^{d} * K_{1\times 1} * \text{LayerNorm}(\hat{h}_{t-1})$ , where $K_{3\times 3}^{d} *$ is the $3 \times 3$ depth-wise convolution which double the channel number. Chunking $\tilde{h}_{t-1}$ gives key $(K_{t})$ and value $(V_{t})$ . The calculation of attention map $A_{t} \in \mathbb{R}^{C \times C}$ follows Eq. (2). The final output $O_{t}$ is computed as $O_{t} = K_{1\times 1} * K_{3\times 3}^{d} * K_{1\times 1} * \mathbf{C}[A_{t}V_{t}; f_{t}]$ , where $\mathbf{C}[\cdot; \cdot]$ denotes concatenation. + +# 4.3 Improved Channel Attention (ICA) Module + +The design of ICA follows Fig. 5 for empirically verifying findings in Sec. 3.3. Squeeze and Excite follows the squeeze-and-excite mechanism in Fig. 5 and helps to predict channels with new information. It squeezes the channels of input feature $X \in \mathbb{R}^{C \times H \times W}$ by factor $r$ . The attention map refers to self-attention on $X$ and of size $\mathbb{R}^{\frac{C}{r} \times \frac{C}{r}}$ . Excitation layers expand outputs back to size $\mathbb{R}^{C \times H \times W}$ . Correlation-based channel weighting weights the output channel based on scalars predicted from the attention map. This design emphasizes channels for predicting more discriminative features in 'excite' operation. The attention map $A_r \in \mathbb{R}^{\frac{C}{r} \times \frac{C}{r}}$ is taken for calculating the average and max values along the rows. The average and max values are combined and mapped to weights of size $\mathbb{R}^{\frac{C}{r} \times 1}$ through linear layers and a sigmoid function. + +# 4.4 Implementation Details + +Model details. RealVformer uses SPyNet [29] for flow estimation. After the CAF module, the reconstruction module $\mathcal{R}$ has a three-level encoder-decoder architecture. From level 1 to level 3, there are [2,3,4] transformer blocks with [48,96,192] channels. There are [1,2,4] attention heads in the ICA, all with a squeeze factor of 4. Supplementary B.1 gives the detailed architecture of $\mathcal{R}$ . + +Training details. We train using the REDS dataset [27] and follow RealBasicVSR [5] in applying random combinations of blur, noise, JPEG compression, and video compression for synthesizing input degradations. We load 15 frames as an input sequence. The spatial size of inputs is cropped to $64 \times 64$ , and the batch size is 16. We use a pre-trained flow estimation model SPyNet, the parameter of which is fixed for the first 5K iterations and tuned with other modules later. + +Following RealBasicVSR [5], we perform two-stage training. The first stage trains the model with a Charbonnier loss [19] and SSIM [35] loss for 300K iterations. In the second stage, the model is trained for another 130K iterations with the Charbonnier loss, SSIM loss, perceptual loss [18] and GAN loss [11] together, weighted by 1, 0.001, 1, and 0.005, respectively. The implementations of perceptual loss, GAN loss, and discriminator follow RealBasicVSR [5]. We implement all experiments on 4 Quadro RTX 8000 GPUs with PyTorch. Other details of the training settings are in Supplementary B.2. + +# 4.5 Experimental Results + +We compare our model on four datasets, VideoLQ [5], RealVSR [37], REDS4 [27] and UDM10 [38], with RealBasicVSR [5], Real-ESRGAN [34], BSRGAN [40], DBVSR [28], RealVSR [37], DAN [14], and RealSR [17]. VideoLQ and RealVSR are collected from real-world scenarios. VideoLQ is an unpaired dataset. RealVSR has same-size paired low-quality (LQ) and high-quality (HQ) frames. Our model super-resolves the input frames spatially. Although downsampling LQ enables paired data, it alters the original degradation. Thus, we test on the original LQ and do not use the HQ for evaluation. REDS4 and UDM10 have + +Table 2: Quantitative comparisons with existing methods with best and second-best results. Our method achieves the best ILNIQE and NRQM scores over VideoLQ and RealVSR datasets and the best PSNR, SSIM, and LPIPS over synthetic datasets REDS4 and UDM10 with relatively few parameters and the lowest run-time. + +
RealSRDANRealVSRDBVSRBSRGANReal-ESRGANRealBasicVSROurs
Params (M)16.74.32.725.516.716.76.35.3
Runtime (ms)180250772-1801967349
VideoLQILNIQE↓26.6328.2831.9427.8527.4927.9725.9825.94
NRQM↑6.0543.7423.4603.8516.1566.0576.3066.338
RealVSRILNIQE↓32.8132.2934.39-32.6531.9330.3728.61
NRQM↑5.6103.5233.795-6.1526.2456.5826.588
REDS4PSNR↑22.0222.6718.3022.3522.9421.5623.0923.34
SSIM↑0.50970.55710.49000.55300.57500.55560.60760.6079
LPIPS↓0.59910.63150.72400.62110.37660.35330.29910.2877
UDM10PSNR↑25.3725.9023.3525.0825.9724.9625.9626.42
SSIM↑0.66580.72290.71150.71120.75680.74320.74910.7609
LPIPS↓0.48110.47810.47610.47560.33880.33950.32090.3063
+ +ground-truth images, and we synthesize low-quality inputs through the same pipeline during training. All tested recurrent-based VSR models load the half sequence for videos in RealVSR and the whole sequence for others each time. The quantitative and qualitative results are discussed below. + +Quantitative Results. For evaluation without reference, we apply the ILNIQE [41], the improved version of NIQE [26], and NRQM [24] metrics based on each output sequence's first, middle, and last frames in RGB format [5]. These metrics appear less biased towards oversharpened features; details are given in Supplementary B.3. For evaluation on REDS4 and UDM10, we report more reliable metrics, PSNR, SSIM, and LPIPS [42]. We collect released models of all compared methods and generate sequences. As shown in Tab. 2, RealVformer performs better than other methods with smaller parameters than the most competitive RealBasicVSR [5] and the shortest runtime. + +![](images/7154e1d49620a554ebc3214d43090340255fe01d6f563fb14ab126472cc6e278.jpg) +(a) Input + +![](images/79300de9e973112bf493dfba011b44c494ff1c3d93bc9e56c00fb17c687553f2.jpg) +(b) RE + +![](images/754704fc04ad4b48fa413195c389fc83f074ec91c3db1c74d0d58cc64e20e0a1.jpg) +(c) RB +Fig. 8: Qualitative comparisons on synthetic datasets. Our method produces clearer than RealESRGAN (RE) [34] and RealBasicVSR (RB) [5] for very hard inputs. + +![](images/b4b7fcdd3a127285dfa26501028d6ba6ed6f799dd9054c8c5a593e99e6bfd4e7.jpg) +(d) Ours + +![](images/1cf4db5470437749b3c26ad8ad3585bc5933cd921158cd7c4bffff6fec5300b9.jpg) +(e) GT + +![](images/b60ce4e5e99cf6e1031caf5b5ce7b5e252fbb35e8b12b05b2fe1eefd517135e6.jpg) +(a) Input + +![](images/5bddf52abdf27bd053de4eb6034dfb839fa72ecf7a7bb857dec2a97e6c1232d3.jpg) +(b) RE + +![](images/87acee6190ccde9bce59c79d4b1e0f2792378f80e62b05e06062a55971effd69.jpg) +(c) RB +Fig. 9: Qualitative comparisons on real-world datasets. Our method produces less high-frequency artifacts and overshoot effects than RealESRGAN [34] and RealBasicVSR [5]. + +![](images/47abac7bb4302e1cd80e6bbd5fb217d77e3703c2037e2053d5baeb1af4918a53.jpg) +(d) Ours + +Qualitative Results. We show qualitative comparisons on synthetic (see Fig. 8) and real-world (see Fig. 9) datasets. Compared to the listed methods, RealViformer generates clear structures with much fewer high-frequency artifacts. More visual comparisons are in Supplementary B.5. + +User Study. We also performed a user study and asked 30 evaluators on Amazon MTurk to score reconstructions for 85 frames sampled from datasets VideoLQ [5], RealVSR [37], REDS4 [27] and UDM10 [38]. Each worker saw five HR results of the same frame and rated them based on the visual quality, from 1 (the worst) to 5 (the best); as shown in Fig. 10, our method surpasses BSRGAN [40], Real-ESRGAN [34], RealSR [17], and RealBasicVSR [5]. + +![](images/69fc77c9b7463039c19a06fe44f3f9f4c460eb54e7081c79daeca8b43c650419.jpg) +Fig.10: User study results from 30 evaluators on 85 frames. Our method achieves the best among all five methods regrading mean option scores (MOS). + +Table 3: Ablations of CAF and ICA modules. Channel-attention baseline (ch-baseline) performs better than the spatial-attention baseline (sp-baseline). The CAF module improves the performances for both datasets, and ICA further improves the performances to the state-of-the-art. + +
MethodCAFICAVideoLQ NRQM↑UDM10 LPIPS↓
Sp-baseline--6.0610.3482
Ch-baselineXX6.1810.3085
RealVformer-✓X6.1960.2933
RealVformer✓✓6.3380.2877
+ +Ablation Studies. We conduct ablations to validate the advantage of channel attention and the efficacy of the CAF and ICA. We build a spatial-attention baseline (sp-baseline) by replacing the reconstruction module of BasicVSR [3] with SwinIR [21]. Channel-attention baseline (ch-baseline) is built with the same + +![](images/9d56bebcd12991c96ab8b84c91e0d6d0131003d9a103dfffc935862074792869.jpg) +Fig. 11: (a) Visual comparison between RealVformer and its ablations. Red circles highlight the improved details. (b) Radial Power Spectrum (RPS) of model predictions. Using ICA improves the power of high-frequency components (blue region). + +overall architecture as RealVformer but replaces CAF with simple concatenation and substitutes ICA with the channel attention block in Restormer [39]. RealVformer $^{-}$ applies CAF on ch-baseline. All models are trained with the same settings in Sec. 4.4. We report NRQM scores for the VideoLQ [5] dataset and LPIPS scores for the UDM10 [38] da set. As shown in Tab. 3, using the original channel attention module, ch-baseline already yields better performances than sp-baseline. The CAF and ICA modules further improve RealVformer to state-of-the-art performance with channel correlation of propagated information decreasing from 0.436 to 0.422. Fig. 11a visually compares RealVformer with its ablations. Applying CAF reduces artifacts, while ICA provides further improvements. Fig. 11b supplements the Radial Power Spectrum of model predictions. The blur region shows ICA increases the power in the high-frequency region. + +# 5 Conclusion + +This paper proposes a real-world VSR model, RealVformer, based on findings from investigating channel and spatial attention in a real-world setting. Explorations reveal that channel attention is less sensitive to the artifacts in query and better serves as a temporal aggregation module to limit model-produced artifacts in hidden states. Additionally, we observe the higher covariance of channel attention outputs and propose the Improved Channel Attention (ICA) Module with a squeeze-and-excite and a covariance-based rescaling mechanism. Based on our findings, we build RealVformer, a channel-attention-based recurrent model for real-world VSR. We propose the CAF module to limit artifact propagation and use the ICA module to achieve better reconstructions. RealVformer performs state-of-the-art on two real-world video datasets with fewer parameters and shorter runtime. On the other hand, we value our findings w.r.t. comparisons between channel and spatial attention and exploration of covariance in channel attention as inspiration for further real-world VSR research. + +Acknowledgement This research is supported by the National Research Foundation, Singapore under its NRF Fellowship for AI (NRF-NRFFAI1-2019-0001). + +Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of National Research Foundation, Singapore. + +# References + +1. Bardes, A., Ponce, J., LeCun, Y.: Vicreg: Variance-invariance-covariance regularization for self-supervised learning. arXiv preprint arXiv:2105.04906 (2021) +2. Cao, J., Li, Y., Zhang, K., Van Gool, L.: Video super-resolution transformer. arXiv preprint arXiv:2106.06847 (2021) +3. Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: Basicvsr: The search for essential components in video super-resolution and beyond. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4947-4956 (2021) +4. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5972-5981 (2022) +5. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Investigating tradeoffs in real-world video super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5962-5971 (2022) +6. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12299-12310 (2021) +7. Chen, X., Wang, X., Zhou, J., Qiao, Y., Dong, C.: Activating more pixels in image super-resolution transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22367-22377 (2023) +8. Cogswell, M., Ahmed, F., Girshick, R., Zitnick, L., Batra, D.: Reducing overfitting in deep networks by decorrelating representations. arXiv preprint arXiv:1511.06068 (2015) +9. Dai, T., Cai, J., Zhang, Y., Xia, S.T., Zhang, L.: Second-order attention network for single image super-resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11065-11074 (2019) +0. Fuoli, D., Gu, S., Timofte, R.: Efficient video super-resolution through recurrent latent space propagation. In: 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW). pp. 3476-3485. IEEE (2019) +1. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020) +2. Hu, J., Shen, L., Sun, G.: Squeeze-and-excitation networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7132-7141 (2018) +3. Hua, T., Wang, W., Xue, Z., Ren, S., Wang, Y., Zhao, H.: On feature decorrelation in self-supervised learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9598-9608 (2021) +4. Huang, Y., Li, S., Wang, L., Tan, T., et al.: Unfolding the alternating optimization for blind super resolution. Advances in Neural Information Processing Systems 33, 5632-5643 (2020) + +15. Isobe, T., Jia, X., Gu, S., Li, S., Wang, S., Tian, Q.: Video super-resolution with recurrent structure-detail network. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16. pp. 645-660. Springer (2020) +16. Jeelani, M., Cheema, N., Illgner-Fehns, K., Slusallek, P., Jaiswal, S., et al.: Expanding synthetic real-world degradations for blind video super resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1199-1208 (2023) +17. Ji, X., Cao, Y., Tai, Y., Wang, C., Li, J., Huang, F.: Real-world super-resolution via kernel estimation and noise injection. In: proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 466-467 (2020) +18. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 694-711. Springer (2016) +19. Lai, W.S., Huang, J.B., Ahuja, N., Yang, M.H.: Fast and accurate image superresolution with deep laplacian pyramid networks. IEEE transactions on pattern analysis and machine intelligence 41(11), 2599-2613 (2018) +20. Liang, J., Cao, J., Fan, Y., Zhang, K., Ranjan, R., Li, Y., Timofte, R., Van Gool, L.: Vrt: A video restoration transformer. arXiv preprint arXiv:2201.12288 (2022) +21. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1833-1844 (2021) +22. Liang, J., Fan, Y., Xiang, X., Ranjan, R., Ilg, E., Green, S., Cao, J., Zhang, K., Timofte, R., Gool, L.V.: Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems 35, 378-393 (2022) +23. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021) +24. Ma, C., Yang, C.Y., Yang, X., Yang, M.H.: Learning a no-reference quality metric for single-image super-resolution. Computer Vision and Image Understanding 158, 1-16 (2017) +25. Mei, Y., Fan, Y., Zhou, Y.: Image super-resolution with non-local sparse attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3517-3526 (2021) +26. Mittal, A., Soundararajan, R., Bovik, A.C.: Making a "completely blind" image quality analyzer. IEEE Signal processing letters 20(3), 209-212 (2012) +27. Nah, S., Baik, S., Hong, S., Moon, G., Son, S., Timofte, R., Mu Lee, K.: Ntire 2019 challenge on video deblurring and super-resolution: Dataset and study. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019) +28. Pan, J., Bai, H., Dong, J., Zhang, J., Tang, J.: Deep blind video super-resolution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4811-4820 (2021) +29. Ranjan, A., Black, M.J.: Optical flow estimation using a spatial pyramid network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4161-4170 (2017) +30. Shi, S., Gu, J., Xie, L., Wang, X., Yang, Y., Dong, C.: Rethinking alignment in video super-resolution transformers. arXiv preprint arXiv:2207.08494 (2022) + +31. Song, Y., Wang, M., Yang, Z., Xian, X., Shi, Y.: Negvsr: Augmenting negatives for generalized noise modeling in real-world video super-resolution. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 10705-10713 (2024) +32. Wang, H., Chen, X., Ni, B., Liu, Y., Liu, J.: Omni aggregation networks for lightweight image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22378-22387 (2023) +33. Wang, X., Chan, K.C., Yu, K., Dong, C., Change Loy, C.: Edvr: Video restoration with enhanced deformable convolutional networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019) +34. Wang, X., Xie, L., Dong, C., Shan, Y.: Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1905-1914 (2021) +35. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004) +36. Xie, L., Wang, X., Shi, S., Gu, J., Dong, C., Shan, Y.: Mitigating artifacts in real-world video super-resolution models. arXiv preprint arXiv:2212.07339 (2022) +37. Yang, X., Xiang, W., Zeng, H., Zhang, L.: Real-world video super-resolution: A benchmark dataset and a decomposition based learning scheme. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4781-4790 (2021) +38. Yi, P., Wang, Z., Jiang, K., Jiang, J., Ma, J.: Progressive fusion video superresolution network via exploiting non-local spatio-temporal correlations. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3106-3115 (2019) +39. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5728-5739 (2022) +40. Zhang, K., Liang, J., Van Gool, L., Timofte, R.: Designing a practical degradation model for deep blind image super-resolution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4791-4800 (2021) +41. Zhang, L., Zhang, L., Bovik, A.C.: A feature-enriched completely blind image quality evaluator. IEEE Transactions on Image Processing 24(8), 2579-2591 (2015) +42. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018) +43. Zhang, Y., Li, K., Li, K., Wang, L., Zhong, B., Fu, Y.: Image super-resolution using very deep residual channel attention networks. In: Proceedings of the European conference on computer vision (ECCV). pp. 286-301 (2018) \ No newline at end of file diff --git a/2024/RealViformer_ Investigating Attention for Real-World Video Super-Resolution/images.zip b/2024/RealViformer_ Investigating Attention for Real-World Video Super-Resolution/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dd6e2d9b28b093e6969a8974d4724359d8d18485 --- /dev/null +++ b/2024/RealViformer_ Investigating Attention for Real-World Video Super-Resolution/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac6edd08010d7701c3529e64f6730421d8099e193ee3fe5b0bf0678f3d5c168d +size 440584 diff --git a/2024/RealViformer_ Investigating Attention for Real-World Video Super-Resolution/layout.json b/2024/RealViformer_ Investigating Attention for Real-World Video Super-Resolution/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9cbbcd3d08d85d4a62b3ee3e26a2dba616d9c95d --- /dev/null +++ b/2024/RealViformer_ Investigating Attention for Real-World Video Super-Resolution/layout.json @@ -0,0 +1,11736 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 162, + 111, + 455, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 111, + 455, + 148 + ], + "spans": [ + { + "bbox": [ + 162, + 111, + 455, + 148 + ], + "type": "text", + "content": "RealVformer: Investigating Attention for Real-World Video Super-Resolution" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 230, + 168, + 383, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 168, + 383, + 180 + ], + "spans": [ + { + "bbox": [ + 230, + 168, + 383, + 180 + ], + "type": "text", + "content": "Yuehan Zhang and Angela Yao" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 235, + 190, + 380, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 190, + 380, + 213 + ], + "spans": [ + { + "bbox": [ + 235, + 190, + 380, + 213 + ], + "type": "text", + "content": "National University of Singapore {zyuehan,ayao}@comp.nus.edu.sg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 241, + 455, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 241, + 455, + 417 + ], + "spans": [ + { + "bbox": [ + 160, + 241, + 455, + 417 + ], + "type": "text", + "content": "Abstract. In real-world video super-resolution (VSR), videos suffer from in-the-wild degradations and artifacts. VSR methods, especially recurrent ones, tend to propagate artifacts over time in the real-world setting and are more vulnerable than image super-resolution. This paper investigates the influence of artifacts on commonly used covariance-based attention mechanisms in VSR. Comparing the widely-used spatial attention, which computes covariance over space, versus the channel attention, we observe that the latter is less sensitive to artifacts. However, channel attention leads to feature redundancy, as evidenced by the higher covariance among output channels. As such, we explore simple techniques such as the squeeze-excite mechanism and covariance-based rescaling to counter the effects of high channel covariance. Based on our findings, we propose RealVformer. This channel-attention-based real-world VSR framework surpasses state-of-the-art on two real-world VSR datasets with fewer parameters and faster runtimes. The source code is available at https://github.com/Yuehan717/RealVformer." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 437, + 230, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 437, + 230, + 449 + ], + "spans": [ + { + "bbox": [ + 133, + 437, + 230, + 449 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 461, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 582 + ], + "type": "text", + "content": "Video super-resolution (VSR) recovers a high-resolution (HR) sequence of frames from its low-resolution (LR) counterpart. Recurrent convolutional approaches are commonly used in VSR with standard settings, assuming the LR frames are downsampled from HR frames with known kernels. [3, 10, 15]. However, in real-world VSR (RWVSR), the low-resolution videos are not simply downsampled versions of their high-resolution counterparts. Instead, they feature complex degradations that arise from the camera imaging system, compression, internet transmission, and other factors. These degradations make architecture design for RWVSR challenging, as artifacts and degradations tend to propagate and get exaggerated over the recurrent connection [5, 36]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Recently, transformer architectures have replaced convolutional architectures as state-of-the-art for standard VSR. While attention mechanisms have replaced convolution operations, most methods retain the recurrent connection to aggregate the information over time [22, 30]. Yet, such architectures do not always perform well on RWVSR. For instance, a Swin-based model [21] designed for standard VSR, when applied to a real-world input frame, as shown in Fig. 1a, generates more artificial lines than the convolutional model RealBasicVSR [5]." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 113, + 223, + 155 + ], + "blocks": [ + { + "bbox": [ + 167, + 113, + 223, + 155 + ], + "lines": [ + { + "bbox": [ + 167, + 113, + 223, + 155 + ], + "spans": [ + { + "bbox": [ + 167, + 113, + 223, + 155 + ], + "type": "image", + "image_path": "110ffaa2f809c8fbaaf8966a162e998bca4cf29b8f6bfdc7e6bbbdf0a92444b6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 156, + 203, + 163 + ], + "lines": [ + { + "bbox": [ + 187, + 156, + 203, + 163 + ], + "spans": [ + { + "bbox": [ + 187, + 156, + 203, + 163 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 225, + 114, + 282, + 155 + ], + "blocks": [ + { + "bbox": [ + 225, + 114, + 282, + 155 + ], + "lines": [ + { + "bbox": [ + 225, + 114, + 282, + 155 + ], + "spans": [ + { + "bbox": [ + 225, + 114, + 282, + 155 + ], + "type": "image", + "image_path": "456b3246abbab733088a604dbc47e275c35a887c21e6f92da2873f6033cbd83a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 232, + 156, + 274, + 163 + ], + "lines": [ + { + "bbox": [ + 232, + 156, + 274, + 163 + ], + "spans": [ + { + "bbox": [ + 232, + 156, + 274, + 163 + ], + "type": "text", + "content": "RealBasicVSR" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 167, + 168, + 223, + 210 + ], + "blocks": [ + { + "bbox": [ + 167, + 168, + 223, + 210 + ], + "lines": [ + { + "bbox": [ + 167, + 168, + 223, + 210 + ], + "spans": [ + { + "bbox": [ + 167, + 168, + 223, + 210 + ], + "type": "image", + "image_path": "84f5cc884e9e4d8e0b056fc7d3d08be8a7ec48437a2e83bf577871e3316a5ea4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 213, + 219, + 221 + ], + "lines": [ + { + "bbox": [ + 173, + 213, + 219, + 221 + ], + "spans": [ + { + "bbox": [ + 173, + 213, + 219, + 221 + ], + "type": "text", + "content": "Swin-based VSR" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 225, + 168, + 282, + 210 + ], + "blocks": [ + { + "bbox": [ + 225, + 168, + 282, + 210 + ], + "lines": [ + { + "bbox": [ + 225, + 168, + 282, + 210 + ], + "spans": [ + { + "bbox": [ + 225, + 168, + 282, + 210 + ], + "type": "image", + "image_path": "bf3b02d4f19ffbccf955eee78c521b1bbb179ef40d04958aeeb88a13715baad0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 213, + 281, + 221 + ], + "lines": [ + { + "bbox": [ + 227, + 213, + 281, + 221 + ], + "spans": [ + { + "bbox": [ + 227, + 213, + 281, + 221 + ], + "type": "text", + "content": "Ours: RealVformer" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 301, + 123, + 376, + 193 + ], + "blocks": [ + { + "bbox": [ + 184, + 227, + 269, + 236 + ], + "lines": [ + { + "bbox": [ + 184, + 227, + 269, + 236 + ], + "spans": [ + { + "bbox": [ + 184, + 227, + 269, + 236 + ], + "type": "text", + "content": "(a) Visual comparisons." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 301, + 123, + 376, + 193 + ], + "lines": [ + { + "bbox": [ + 301, + 123, + 376, + 193 + ], + "spans": [ + { + "bbox": [ + 301, + 123, + 376, + 193 + ], + "type": "image", + "image_path": "63c21eeb4eedbfec929404f0e2ce863925a932efc469e60750bb571e63ddffbb.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 196, + 334, + 202 + ], + "lines": [ + { + "bbox": [ + 326, + 196, + 334, + 202 + ], + "spans": [ + { + "bbox": [ + 326, + 196, + 334, + 202 + ], + "type": "text", + "content": "W" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 202, + 356, + 209 + ], + "lines": [ + { + "bbox": [ + 310, + 202, + 356, + 209 + ], + "spans": [ + { + "bbox": [ + 310, + 202, + 356, + 209 + ], + "type": "text", + "content": "Spatial Attention" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 378, + 123, + 455, + 194 + ], + "blocks": [ + { + "bbox": [ + 381, + 164, + 388, + 174 + ], + "lines": [ + { + "bbox": [ + 381, + 164, + 388, + 174 + ], + "spans": [ + { + "bbox": [ + 381, + 164, + 388, + 174 + ], + "type": "text", + "content": "H" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 378, + 123, + 455, + 194 + ], + "lines": [ + { + "bbox": [ + 378, + 123, + 455, + 194 + ], + "spans": [ + { + "bbox": [ + 378, + 123, + 455, + 194 + ], + "type": "image", + "image_path": "13508f59141d52d449e745094a9abb701d287e42aa305cbada4bb50442108113.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 201, + 439, + 209 + ], + "lines": [ + { + "bbox": [ + 389, + 201, + 439, + 209 + ], + "spans": [ + { + "bbox": [ + 389, + 201, + 439, + 209 + ], + "type": "text", + "content": "Channel Attention" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 318, + 227, + 436, + 236 + ], + "lines": [ + { + "bbox": [ + 318, + 227, + 436, + 236 + ], + "spans": [ + { + "bbox": [ + 318, + 227, + 436, + 236 + ], + "type": "text", + "content": "(b) Schematics of two attentions." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 245, + 480, + 322 + ], + "lines": [ + { + "bbox": [ + 130, + 245, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 480, + 322 + ], + "type": "text", + "content": "Fig. 1: (a) Designing a RWVSR transformer is not trivial. A Swin-based transformer suited for standard VSR hallucinates more lines than a RealBasicVSR, a convolutional state-of-the-art. We propose RealVformer based on our investigation of attention under the RWVSR setting. RealVformer generates details with fewer artifacts than RealBasicVSR [3] and the Swin-based VSR model. (b) Schematic for spatial and channel attention. Spatial attention aggregates features based on pixel representations. Channel attention takes " + }, + { + "bbox": [ + 130, + 245, + 480, + 322 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 130, + 245, + 480, + 322 + ], + "type": "text", + "content": " feature map for matching across channels." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 356, + 480, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 480, + 429 + ], + "type": "text", + "content": "Why should transformers perform well on (synthetic) standard VSR but poorly in real-world cases? We speculate that standard VSR transformers benefit from the similarity-based matching of the attention mechanism, which accurately aggregates information spatially and temporally [21,22,30]. However, when input degradation exists, the aggregated information becomes less reliable because the attention queries may be derived from both true source video and artifacts." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 433, + 480, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 480, + 589 + ], + "type": "text", + "content": "This work investigates and sheds light on the sensitivity of attention in real-world settings. We compare two covariance-based attention mechanisms used in low-level Transformers: spatial attention [21,30] and channel attention [39]. Spatial attention takes pixel-wise features as keys and queries and estimates their covariances across spatial positions. The most popular form is window-based attention [6]; The shift-window scheme from Swin Transformers [23] enables the model to access distant spatial ranges [2,21,30] without computational blow-up. Spatial attention is widely used for video and image super-resolution, albeit in the standard setting. Channel attention [39] estimates covariances across channels (see Fig. 1b) and collapses the spatial extent of a feature map. It defines the number of queries and keys by channel numbers rather than spatial resolution. Although established for deblurring or denoising [39], channel attention's efficacy in super-resolution remains to be determined." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": "Our experiments show that channel attention is less sensitive to artifacts than the spatial counterparts, resulting in higher performance gains in RWVSR. However, it is also revealed that channel attention leads to feature channels with higher covariance. From a learning perspective, a high covariance is undesirable because it is a strong indicator of feature redundancy [1,8,13]. Therefore, using channel attention naively will have limited improvements over existing RWVSR" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 255, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 255, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 255, + 101 + ], + "type": "text", + "content": "Y. Zhang and A. Yao" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "state-of-the-art methods. Such a finding has wide-reaching impacts as channel attention is used increasingly for low-level vision." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 270 + ], + "type": "text", + "content": "To verify our findings, we explore established mechanisms to counter the effects of feature redundancy - simple techniques such as squeeze-excite and covariance-based rescaling improve the vanilla channel attention design. From these outcomes, we propose RealVformer, a new transformer real-world VSR model. RealVformer performs channel attention between the current frame feature and the propagated hidden state to limit model-produced artifacts. The model then reconstructs features through improved channel attention modules featuring squeeze-and-excite and covariance-based channel rescaling mechanisms. With our effective designs, RealVformer achieves state-of-the-art performance with fewer parameters on challenging synthetic video datasets and two real-world video datasets collected from different scenes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 273, + 436, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 273, + 436, + 284 + ], + "spans": [ + { + "bbox": [ + 146, + 273, + 436, + 284 + ], + "type": "text", + "content": "Summarizing our contributions in order of importance, our paper" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 293, + 484, + 426 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 293, + 484, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 484, + 341 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 484, + 341 + ], + "type": "text", + "content": "- investigates the differences between spatial and channel attention for RWVSR. Spatial attention, although widely used, is revealed to be highly sensitive to the noise and degradations common in RWVSR sequences, while channel attention is more robust." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 342, + 481, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 481, + 377 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 481, + 377 + ], + "type": "text", + "content": "- reveals that naively applying channel attention increases channel covariance, which is problematic for learning; this overlooked fact has a wide-reaching impact as channel attention becomes more used in low-level vision." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 378, + 481, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 378, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 138, + 378, + 481, + 426 + ], + "type": "text", + "content": "- empirically verifies the negative effect of high channel covariance by countering it with established techniques, based on which we develop the RealVformer for RWVSR. Our simple modification surpasses state-of-the-art despite using less compute." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 446, + 237, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 237, + 459 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 237, + 459 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 473, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 544 + ], + "type": "text", + "content": "Standard Video Super-Resolution models focus on architecture design to use temporal information better [2-4, 15]. Previous research starts from sliding-window-based [15,33] to recurrent-based frameworks [3,4] for using distant-frame information. Recent works introduce Transformer blocks into existing recurrent frameworks to overcome the locality limitation of convolution and accurately match abundant information for feature reconstruction [20,22,30]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": "Real-world video super-resolution focuses on modeling, removing, and limiting the impact of real-world degradations. Existing works are convolutional models and focus on designing losses or modules for degradation processing. DBVSR [28] explicitly estimates the degradation kernel through a sub-network. RealBasicVSR [5] tries to 'clean' artifacts through a processing module for BasicVSR [3]. In a similar approach, FastRealVSR [36] borrows an external pool of blur and sharpening filters to 'clean' the hidden states. Other recent works [16,31] advance the synthesis method for paired training data. Instead, we focus on investigating the function of attention in RWVSR rather than architecture design or data synthesis." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "type": "text", + "content": "RealViformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "Attention mechanisms have been widely applied for low-level vision tasks [9, 25, 43, 43]. Transformers with covariance-based attention are the most prevalent [21, 30, 39] for standard VSR. Most existing methods adopt shift-window-based spatial attention [23] to aggregate information from other positions within or across frames. In contrast, Restormer [39] computes the covariance among channels and shows its effectiveness for multiple image restoration tasks. More recent works stitch spatial and channel attention together to enlarge the receptive field [7, 32] for standard image super-resolution. Instead, we investigate attention mechanisms in terms of their sensitivity to real-world degradation for the first time and develop an effective real-world VSR Transformer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 253, + 309, + 267 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 253, + 309, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 253, + 309, + 267 + ], + "type": "text", + "content": "3 Explorations on Attention" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 277, + 482, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 326 + ], + "type": "text", + "content": "Sec. 3.1 defines the VSR task and the two attention mechanisms. Sec. 3.2 compares the channel and spatial attentions' sensitivity to query artifacts and effects on real-world VSR performance. Sec. 3.3 reveals that channel attention leads to higher covariance among channels and explores mitigating options." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 341, + 228, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 228, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 228, + 353 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 359, + 482, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 359, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 359, + 482, + 553 + ], + "type": "text", + "content": "Standard vs. Real-World VSR. Given a low-resolution (LR) video sequence with " + }, + { + "bbox": [ + 130, + 359, + 482, + 553 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 359, + 482, + 553 + ], + "type": "text", + "content": " frames " + }, + { + "bbox": [ + 130, + 359, + 482, + 553 + ], + "type": "inline_equation", + "content": "I^L \\in \\mathcal{R}^{T \\times H \\times W \\times K}" + }, + { + "bbox": [ + 130, + 359, + 482, + 553 + ], + "type": "text", + "content": ", VSR models reconstruct a high-resolution sequence " + }, + { + "bbox": [ + 130, + 359, + 482, + 553 + ], + "type": "inline_equation", + "content": "I^H \\in \\mathcal{R}^{T \\times sH \\times sW \\times K}" + }, + { + "bbox": [ + 130, + 359, + 482, + 553 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 359, + 482, + 553 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 130, + 359, + 482, + 553 + ], + "type": "text", + "content": " is the input spatial resolution, " + }, + { + "bbox": [ + 130, + 359, + 482, + 553 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 359, + 482, + 553 + ], + "type": "text", + "content": " is the number of input channels and " + }, + { + "bbox": [ + 130, + 359, + 482, + 553 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 359, + 482, + 553 + ], + "type": "text", + "content": " is the scaling factor. In the standard setting of VSR, " + }, + { + "bbox": [ + 130, + 359, + 482, + 553 + ], + "type": "inline_equation", + "content": "I_t^L" + }, + { + "bbox": [ + 130, + 359, + 482, + 553 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 359, + 482, + 553 + ], + "type": "inline_equation", + "content": "t \\in \\{0, \\dots, T\\}" + }, + { + "bbox": [ + 130, + 359, + 482, + 553 + ], + "type": "text", + "content": ", is assumed as the downsampled version of " + }, + { + "bbox": [ + 130, + 359, + 482, + 553 + ], + "type": "inline_equation", + "content": "I_t^H" + }, + { + "bbox": [ + 130, + 359, + 482, + 553 + ], + "type": "text", + "content": ", defined as: " + }, + { + "bbox": [ + 130, + 359, + 482, + 553 + ], + "type": "inline_equation", + "content": "I_t^L = (I_t^H) \\downarrow_{\\frac{1}{s}}" + }, + { + "bbox": [ + 130, + 359, + 482, + 553 + ], + "type": "text", + "content": ". Both training and testing datasets follow this formulation to generate LR-HR pairs given " + }, + { + "bbox": [ + 130, + 359, + 482, + 553 + ], + "type": "inline_equation", + "content": "I^H" + }, + { + "bbox": [ + 130, + 359, + 482, + 553 + ], + "type": "text", + "content": ". In a real-world setting, there is no closed formulation for the relationship between " + }, + { + "bbox": [ + 130, + 359, + 482, + 553 + ], + "type": "inline_equation", + "content": "I_t^L" + }, + { + "bbox": [ + 130, + 359, + 482, + 553 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 359, + 482, + 553 + ], + "type": "inline_equation", + "content": "I_t^H" + }, + { + "bbox": [ + 130, + 359, + 482, + 553 + ], + "type": "text", + "content": " due to the unknown distribution of real-world degradations. Real-ESRGAN [34] proposed a widely-used training setting that randomly applies synthesized blur, noise, compression, and resizing to the HR frames to generate paired LR frames with complex degradations. The testing datasets are either synthesized by the same pipeline in training or collected from diverse real-world sources [5, 37]. The synthesized testing sets have paired ground-truth sequences and are evaluated by full-reference metrics, e.g. PSNR and LPIPS [42]; real-world datasets are always without ground truth and require no-reference metrics, such as NRQM [24]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "type": "text", + "content": "Attention Definitions. In Transformers, the attention modules project layer-normalized tensor " + }, + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "type": "text", + "content": " to query " + }, + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "type": "text", + "content": ", and tensor " + }, + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "type": "inline_equation", + "content": "Y \\in \\mathbb{R}^{\\hat{C} \\times \\hat{H} \\times \\hat{W}}" + }, + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "type": "text", + "content": " to key " + }, + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\hat{H} \\times \\hat{W}" + }, + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "type": "text", + "content": " are the spatial resolution of the normalized tensors. The attention map " + }, + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "type": "text", + "content": " is generated by calculating the covariance between " + }, + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "type": "text", + "content": ", followed by a softmax function, before being applied to the value " + }, + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "type": "text", + "content": " to produce output " + }, + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "type": "text", + "content": ". Spatial and channel attention differ in tensor dimension taken for the covariance calculation." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 254, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 254, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 254, + 101 + ], + "type": "text", + "content": "Y. Zhang and A. Yao" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 482, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 482, + 665 + ], + "type": "text", + "content": "1 Note we define mutual attention as the default; self-attention is a special form of mutual attention where " + }, + { + "bbox": [ + 133, + 642, + 482, + 665 + ], + "type": "inline_equation", + "content": "Y = X" + }, + { + "bbox": [ + 133, + 642, + 482, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 115, + 482, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 115, + 482, + 178 + ], + "spans": [ + { + "bbox": [ + 131, + 115, + 482, + 178 + ], + "type": "text", + "content": "Spatial attention generates an " + }, + { + "bbox": [ + 131, + 115, + 482, + 178 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{HW\\times \\hat{H}\\hat{W}}" + }, + { + "bbox": [ + 131, + 115, + 482, + 178 + ], + "type": "text", + "content": " attention map, " + }, + { + "bbox": [ + 131, + 115, + 482, + 178 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_s" + }, + { + "bbox": [ + 131, + 115, + 482, + 178 + ], + "type": "text", + "content": ", by computing the covariance between features in the query " + }, + { + "bbox": [ + 131, + 115, + 482, + 178 + ], + "type": "inline_equation", + "content": "(Q_{s})" + }, + { + "bbox": [ + 131, + 115, + 482, + 178 + ], + "type": "text", + "content": " and key " + }, + { + "bbox": [ + 131, + 115, + 482, + 178 + ], + "type": "inline_equation", + "content": "(K_{s})" + }, + { + "bbox": [ + 131, + 115, + 482, + 178 + ], + "type": "text", + "content": " at each spatial position. The query, key and values are computed as " + }, + { + "bbox": [ + 131, + 115, + 482, + 178 + ], + "type": "inline_equation", + "content": "Q_{s} = W_{s}^{Q}X" + }, + { + "bbox": [ + 131, + 115, + 482, + 178 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 115, + 482, + 178 + ], + "type": "inline_equation", + "content": "K_{s} = W_{s}^{K}Y" + }, + { + "bbox": [ + 131, + 115, + 482, + 178 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 115, + 482, + 178 + ], + "type": "inline_equation", + "content": "V_{s} = W_{s}^{V}Y" + }, + { + "bbox": [ + 131, + 115, + 482, + 178 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 115, + 482, + 178 + ], + "type": "inline_equation", + "content": "W_{s}^{Q}\\in \\mathbb{R}^{D_{s}\\times C}" + }, + { + "bbox": [ + 131, + 115, + 482, + 178 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 115, + 482, + 178 + ], + "type": "inline_equation", + "content": "\\{W_s^K,W_s^V\\} \\in \\mathbb{R}^{D_s\\times \\hat{C}}" + }, + { + "bbox": [ + 131, + 115, + 482, + 178 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 115, + 482, + 178 + ], + "type": "inline_equation", + "content": "D_{s}" + }, + { + "bbox": [ + 131, + 115, + 482, + 178 + ], + "type": "text", + "content": " is the dimension of projections. The attention map " + }, + { + "bbox": [ + 131, + 115, + 482, + 178 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_s" + }, + { + "bbox": [ + 131, + 115, + 482, + 178 + ], + "type": "text", + "content": " and output attention features " + }, + { + "bbox": [ + 131, + 115, + 482, + 178 + ], + "type": "inline_equation", + "content": "O_{s}" + }, + { + "bbox": [ + 131, + 115, + 482, + 178 + ], + "type": "text", + "content": " are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 215, + 184, + 481, + 199 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 184, + 481, + 199 + ], + "spans": [ + { + "bbox": [ + 215, + 184, + 481, + 199 + ], + "type": "interline_equation", + "content": "\\mathcal {A} _ {s} = \\operatorname {s o f t m a x} \\left(Q _ {s} ^ {T} K _ {s} / \\sqrt {D _ {s}}\\right), O _ {s} = \\mathcal {A} _ {s} V _ {s} ^ {T}, \\tag {1}", + "image_path": "3da8484142e90214a5d9ca4c1d67754a3b32a783f5ef093ca725843e9fb7af0f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 205, + 479, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 479, + 255 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 479, + 255 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 205, + 479, + 255 + ], + "type": "inline_equation", + "content": "Q_{s}(K_{s})" + }, + { + "bbox": [ + 130, + 205, + 479, + 255 + ], + "type": "text", + "content": " is reshaped to matrix of " + }, + { + "bbox": [ + 130, + 205, + 479, + 255 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{D_s\\times HW}" + }, + { + "bbox": [ + 130, + 205, + 479, + 255 + ], + "type": "inline_equation", + "content": "(\\mathbb{R}^{D_s\\times \\hat{H}\\hat{W}})" + }, + { + "bbox": [ + 130, + 205, + 479, + 255 + ], + "type": "text", + "content": ". We omit the later reshaping operation for brevity. " + }, + { + "bbox": [ + 130, + 205, + 479, + 255 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_s" + }, + { + "bbox": [ + 130, + 205, + 479, + 255 + ], + "type": "text", + "content": " exhaustively relates every spatial location to every other location in space. A window-based implementation [21] limits the correlations to windows of " + }, + { + "bbox": [ + 130, + 205, + 479, + 255 + ], + "type": "inline_equation", + "content": "\\omega \\times \\omega" + }, + { + "bbox": [ + 130, + 205, + 479, + 255 + ], + "type": "text", + "content": ", reducing the map size to " + }, + { + "bbox": [ + 130, + 205, + 479, + 255 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{\\omega^2\\times \\omega^2}" + }, + { + "bbox": [ + 130, + 205, + 479, + 255 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 255, + 479, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 255, + 479, + 304 + ], + "spans": [ + { + "bbox": [ + 131, + 255, + 479, + 304 + ], + "type": "text", + "content": "Channel attention applies convolutions on " + }, + { + "bbox": [ + 131, + 255, + 479, + 304 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 131, + 255, + 479, + 304 + ], + "type": "text", + "content": " to get query " + }, + { + "bbox": [ + 131, + 255, + 479, + 304 + ], + "type": "inline_equation", + "content": "(Q_{c})" + }, + { + "bbox": [ + 131, + 255, + 479, + 304 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 255, + 479, + 304 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 131, + 255, + 479, + 304 + ], + "type": "text", + "content": " for key " + }, + { + "bbox": [ + 131, + 255, + 479, + 304 + ], + "type": "inline_equation", + "content": "(K_{c})" + }, + { + "bbox": [ + 131, + 255, + 479, + 304 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 131, + 255, + 479, + 304 + ], + "type": "inline_equation", + "content": "(V_{c})" + }, + { + "bbox": [ + 131, + 255, + 479, + 304 + ], + "type": "text", + "content": " [39], after which features are the same in spatial resolution (assumed to be " + }, + { + "bbox": [ + 131, + 255, + 479, + 304 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 131, + 255, + 479, + 304 + ], + "type": "text", + "content": " in the following context). It estimates covariance across channels to yield a " + }, + { + "bbox": [ + 131, + 255, + 479, + 304 + ], + "type": "inline_equation", + "content": "C \\times \\hat{C}" + }, + { + "bbox": [ + 131, + 255, + 479, + 304 + ], + "type": "text", + "content": " sized map " + }, + { + "bbox": [ + 131, + 255, + 479, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{c}" + }, + { + "bbox": [ + 131, + 255, + 479, + 304 + ], + "type": "text", + "content": " and output features " + }, + { + "bbox": [ + 131, + 255, + 479, + 304 + ], + "type": "inline_equation", + "content": "O_{c}" + }, + { + "bbox": [ + 131, + 255, + 479, + 304 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 225, + 311, + 481, + 324 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 311, + 481, + 324 + ], + "spans": [ + { + "bbox": [ + 225, + 311, + 481, + 324 + ], + "type": "interline_equation", + "content": "\\mathcal {A} _ {c} = \\operatorname {s o f t m a x} \\left(Q _ {c} K _ {c} ^ {T} / \\alpha\\right), O _ {c} = \\mathcal {A} _ {c} V _ {c}, \\tag {2}", + "image_path": "79463cf581119cae8c4b832cc4d0e4a2475891ba1c86f5e347b9fcdad48571f5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 331, + 480, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 480, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 480, + 381 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 331, + 480, + 381 + ], + "type": "inline_equation", + "content": "Q_{c}" + }, + { + "bbox": [ + 130, + 331, + 480, + 381 + ], + "type": "text", + "content": " is reshaped to " + }, + { + "bbox": [ + 130, + 331, + 480, + 381 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C\\times HW}" + }, + { + "bbox": [ + 130, + 331, + 480, + 381 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 331, + 480, + 381 + ], + "type": "inline_equation", + "content": "K_{c}" + }, + { + "bbox": [ + 130, + 331, + 480, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 331, + 480, + 381 + ], + "type": "inline_equation", + "content": "V_{c}" + }, + { + "bbox": [ + 130, + 331, + 480, + 381 + ], + "type": "text", + "content": " are reshaped to " + }, + { + "bbox": [ + 130, + 331, + 480, + 381 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{\\hat{C}\\times HW}" + }, + { + "bbox": [ + 130, + 331, + 480, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 331, + 480, + 381 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 331, + 480, + 381 + ], + "type": "text", + "content": " is learnable scaling parameter. Because channel attention computes correlations between features of size " + }, + { + "bbox": [ + 130, + 331, + 480, + 381 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{1\\times HW}" + }, + { + "bbox": [ + 130, + 331, + 480, + 381 + ], + "type": "text", + "content": ", it has a larger spatial context than the " + }, + { + "bbox": [ + 130, + 331, + 480, + 381 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{D_s\\times 1}" + }, + { + "bbox": [ + 130, + 331, + 480, + 381 + ], + "type": "text", + "content": "-sized features in spatial attention (see Fig. 1b)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 396, + 311, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 396, + 311, + 407 + ], + "spans": [ + { + "bbox": [ + 131, + 396, + 311, + 407 + ], + "type": "text", + "content": "3.2 Attention in Real-World VSR" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 414, + 479, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 479, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 479, + 473 + ], + "type": "text", + "content": "Sensitivity to Query Artifacts. In standard VSR, attention is used to aggregate information temporally and spatially [22, 30]. The assumption is that queries can match beneficial cues for super-resolution. But what if the queries themselves are unreliable? We speculate this to be true in real-world VSR, where inputs have artifacts and degradations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 474, + 480, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 480, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 480, + 593 + ], + "type": "text", + "content": "How affected are spatial and channel attention mechanisms by query artifacts? Fig. 2 shows how we compare, using cosine similarity, the attention outputs based on queries from the same frame with and without certain degradations. Using encoding layers from a standard convolutional VSR model, we perform the attention operations defined in Sec. 3.1 by taking embeddings of frame " + }, + { + "bbox": [ + 130, + 474, + 480, + 593 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 130, + 474, + 480, + 593 + ], + "type": "text", + "content": " as query and " + }, + { + "bbox": [ + 130, + 474, + 480, + 593 + ], + "type": "inline_equation", + "content": "I_{t-1}" + }, + { + "bbox": [ + 130, + 474, + 480, + 593 + ], + "type": "text", + "content": " as key and value. " + }, + { + "bbox": [ + 130, + 474, + 480, + 593 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 130, + 474, + 480, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 474, + 480, + 593 + ], + "type": "inline_equation", + "content": "I_{t-1}" + }, + { + "bbox": [ + 130, + 474, + 480, + 593 + ], + "type": "text", + "content": " are clean without degradation other than downsampling, and we represent the output feature of the attention module as " + }, + { + "bbox": [ + 130, + 474, + 480, + 593 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 130, + 474, + 480, + 593 + ], + "type": "text", + "content": ". When additional degradation " + }, + { + "bbox": [ + 130, + 474, + 480, + 593 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 130, + 474, + 480, + 593 + ], + "type": "text", + "content": " is applied to " + }, + { + "bbox": [ + 130, + 474, + 480, + 593 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 130, + 474, + 480, + 593 + ], + "type": "text", + "content": ", we represent the corresponding output as " + }, + { + "bbox": [ + 130, + 474, + 480, + 593 + ], + "type": "inline_equation", + "content": "O_{D_{i}}" + }, + { + "bbox": [ + 130, + 474, + 480, + 593 + ], + "type": "text", + "content": ". A smaller deviation of " + }, + { + "bbox": [ + 130, + 474, + 480, + 593 + ], + "type": "inline_equation", + "content": "O_{D_{i}}" + }, + { + "bbox": [ + 130, + 474, + 480, + 593 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 474, + 480, + 593 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 130, + 474, + 480, + 593 + ], + "type": "text", + "content": " indicates lower sensitivity to artifacts in the query." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Tab. 1 shows the cosine similarity between " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "O_{D_i}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " for spatial and channel attention modules. Curiously, " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "O_{D_i}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " based on the channel attention module is more similar to the " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " matched by the degradation-free query, indicating that channel attention is less sensitive to query artifacts. Intuitively, the lower sensitivity of channel attention is related to the larger spatial context used for feature matching. Given a deep feature of size " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C\\times H\\times W}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", channel attention uses feature sized" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "type": "text", + "content": "RealViformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 222, + 113, + 391, + 196 + ], + "blocks": [ + { + "bbox": [ + 222, + 113, + 391, + 196 + ], + "lines": [ + { + "bbox": [ + 222, + 113, + 391, + 196 + ], + "spans": [ + { + "bbox": [ + 222, + 113, + 391, + 196 + ], + "type": "image", + "image_path": "304df22f113c50b298215525302e7655239221a7f78137f645f39c0b2edcf9c3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "lines": [ + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "text", + "content": "Fig.2: Schematic for sensitivity comparison. " + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "inline_equation", + "content": "I_{t-1}, I_t" + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "text", + "content": " are downsampled but clean frames at times " + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "inline_equation", + "content": "t-1" + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "inline_equation", + "content": "D_i(.)" + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "text", + "content": " apply degradations to " + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "inline_equation", + "content": "I_t" + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "inline_equation", + "content": "D_i \\in \\{\\text{blur, noise, compression}\\}" + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "inline_equation", + "content": "O_{D_i}" + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "text", + "content": " are output features of the attention module. Queries are from the embedding at time " + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "text", + "content": ", and keys and values are from time " + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "inline_equation", + "content": "t-1" + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "text", + "content": ". Higher cosine similarities " + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "text", + "content": " between attention output features " + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "inline_equation", + "content": "O_{D_i}" + }, + { + "bbox": [ + 130, + 204, + 482, + 270 + ], + "type": "text", + "content": " reflect less sensitivity to artifacts in queries." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 219, + 316, + 392, + 363 + ], + "blocks": [ + { + "bbox": [ + 130, + 284, + 482, + 308 + ], + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 308 + ], + "type": "text", + "content": "Table 1: Cosine similarity between " + }, + { + "bbox": [ + 130, + 284, + 482, + 308 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 130, + 284, + 482, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 284, + 482, + 308 + ], + "type": "inline_equation", + "content": "{O}_{{D}_{i}}" + }, + { + "bbox": [ + 130, + 284, + 482, + 308 + ], + "type": "text", + "content": " ,attention outputs without and with query degradation. Outputs of channel attention change less under query degradation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 219, + 316, + 392, + 363 + ], + "lines": [ + { + "bbox": [ + 219, + 316, + 392, + 363 + ], + "spans": [ + { + "bbox": [ + 219, + 316, + 392, + 363 + ], + "type": "table", + "html": "
ModuleDi
blurnoisecompression
Spatial attention0.750.920.84
Channel attention0.980.990.99
", + "image_path": "a72e11bdc6f0f7b6efd9979c0ebe5205839e5adcb40ca158be758d001306ac2a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 386, + 481, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 481, + 434 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{1 \\times HW}" + }, + { + "bbox": [ + 130, + 386, + 481, + 434 + ], + "type": "text", + "content": " to calculate the covariance across channels. As such, feature aggregation is based on global information observed in a large normalized spatial context. Instead, the covariance of spatial attention is for features at each location sized " + }, + { + "bbox": [ + 130, + 386, + 481, + 434 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C \\times 1}" + }, + { + "bbox": [ + 130, + 386, + 481, + 434 + ], + "type": "text", + "content": ", so it is likely more sensitive to local value changes from artifacts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 436, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 555 + ], + "type": "text", + "content": "Impact on VSR Models. While channel attention output features are less sensitive to query artifacts than spatial attention, how well do they fare for real-world VSR? We experiment by incorporating these attention variants into VSR models. We focus specifically on recurrent pipelines due to their popularity in standard and real-world VSR [5,22,36]. In recurrent pipelines, artifacts in hidden states may propagate over time, get exaggerated, and negatively influence overall performance [5, 36]. Therefore, we explicitly add the attention module between features of the current frame and the propagated hidden state. Ideally, the added attention will select relative information from hidden states for the current frame and reduce the propagation of model-produced artifacts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "text", + "content": "Fig. 3a shows the baseline model which concatenates " + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "text", + "content": ", the shallow feature of at time " + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\hat{h}_{t-1}" + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "text", + "content": ", the spatially aligned hidden state at time " + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "inline_equation", + "content": "t-1" + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "text", + "content": ". We experiment with two variants by using a channel-attention module " + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "inline_equation", + "content": "G_{c}" + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "text", + "content": " or a spatial-attention module " + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "inline_equation", + "content": "G_{s}" + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "text", + "content": " to replace the concatenation in the baseline. As shown in Fig. 3b, the query is generated by " + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "text", + "content": ", and the key and value are predicted from " + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\hat{h}_{t-1}" + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "text", + "content": ". The attention output " + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "inline_equation", + "content": "O_{t}^{A}" + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "text", + "content": " is concatenated with " + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "text", + "content": " as the input for " + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "All models are trained on the REDS dataset [27] with the random degradation pipeline from Real-ESRGAN [34]. We test on REDS4 [27] with different types and extents of degradations, including Gaussian blur, Gaussian noise, and JPEG" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "type": "text", + "content": "Y. Zhang and A. Yao" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 179, + 114, + 310, + 192 + ], + "blocks": [ + { + "bbox": [ + 179, + 114, + 310, + 192 + ], + "lines": [ + { + "bbox": [ + 179, + 114, + 310, + 192 + ], + "spans": [ + { + "bbox": [ + 179, + 114, + 310, + 192 + ], + "type": "image", + "image_path": "35e515ed0362ecb528cc6ce0b23a682ab670ee9ec95f37a8a860fa37e88d83df.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 204, + 197, + 286, + 206 + ], + "lines": [ + { + "bbox": [ + 204, + 197, + 286, + 206 + ], + "spans": [ + { + "bbox": [ + 204, + 197, + 286, + 206 + ], + "type": "text", + "content": "(a) Recurrent Baseline" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 348, + 120, + 470, + 190 + ], + "blocks": [ + { + "bbox": [ + 348, + 120, + 470, + 190 + ], + "lines": [ + { + "bbox": [ + 348, + 120, + 470, + 190 + ], + "spans": [ + { + "bbox": [ + 348, + 120, + 470, + 190 + ], + "type": "image", + "image_path": "62244e9bb5b9c9de03ad155c4fe629c5dc01b194dcc6355fc3b5318a6e7cb16c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 191, + 433, + 200 + ], + "lines": [ + { + "bbox": [ + 383, + 191, + 433, + 200 + ], + "spans": [ + { + "bbox": [ + 383, + 191, + 433, + 200 + ], + "type": "text", + "content": "(b) Attention" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 134, + 319, + 246, + 413 + ], + "blocks": [ + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "type": "text", + "content": "Fig. 3: (a) The recurrent baseline in Sec. 3.2 has a shallow mapping module " + }, + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "type": "text", + "content": ", reconstruction module " + }, + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "type": "text", + "content": ", upsampling module " + }, + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "type": "text", + "content": " and warping function " + }, + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "type": "text", + "content": " aligns the hidden state " + }, + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "type": "inline_equation", + "content": "h_{t-1}" + }, + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "type": "text", + "content": " to feature at " + }, + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "type": "text", + "content": " based on optical flow " + }, + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "type": "inline_equation", + "content": "s_{(t-1)\\rightarrow t}^{f}" + }, + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "type": "text", + "content": ". All residual blocks are convolutional. The concatenation between " + }, + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\hat{h}_{t-1}" + }, + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "type": "text", + "content": " are replaced with the spatial or channel attention modules in (b) to compare the effect of attention. (b) The attention module first applies layer normalization to " + }, + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\hat{h}_{t-1}" + }, + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "type": "text", + "content": " and then performs channel or spatial attention according to Sec. 3.1. The output feature " + }, + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "type": "inline_equation", + "content": "O_{t}^{A}" + }, + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "type": "text", + "content": " concatenated with " + }, + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "type": "text", + "content": " is processed by the module " + }, + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 130, + 215, + 482, + 308 + ], + "type": "text", + "content": " in (a)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 134, + 319, + 246, + 413 + ], + "lines": [ + { + "bbox": [ + 134, + 319, + 246, + 413 + ], + "spans": [ + { + "bbox": [ + 134, + 319, + 246, + 413 + ], + "type": "image", + "image_path": "bed2f5adfd4a7b2449c7e08a60b2a56d3bcb866ffe0709cdd0166b87ecb4fb98.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 247, + 319, + 356, + 413 + ], + "blocks": [ + { + "bbox": [ + 247, + 319, + 356, + 413 + ], + "lines": [ + { + "bbox": [ + 247, + 319, + 356, + 413 + ], + "spans": [ + { + "bbox": [ + 247, + 319, + 356, + 413 + ], + "type": "image", + "image_path": "897c696d2d587de5dcf9c83c5f2b51e8c89185ba5156b647fa03e5f671ae4d3f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 422, + 482, + 468 + ], + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 468 + ], + "type": "text", + "content": "Fig. 4: Comparison of spatial and channel attention through impact on the performance of real-world VSR model. The Y-axis shows improvements compared to the convolutional baseline. A lower LPIPS score is better. The channel attention module is the best except for the PSNR score of highly blurred inputs." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 357, + 319, + 473, + 412 + ], + "blocks": [ + { + "bbox": [ + 357, + 319, + 473, + 412 + ], + "lines": [ + { + "bbox": [ + 357, + 319, + 473, + 412 + ], + "spans": [ + { + "bbox": [ + 357, + 319, + 473, + 412 + ], + "type": "image", + "image_path": "ca67627848bee6f9524c8382dc94841d5c12b544f459775e56c1e468838bbfac.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 491, + 482, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 564 + ], + "type": "text", + "content": "compression. Fig. 4 displays changes in PSNR and LPIPS compared with the baseline model. Trained on the same dataset and degradation setting, channel attention between temporal information yields better objective and perceptual reconstruction quality than spatial attention and baseline for noise and JPEG compression inputs. Channel attention still achieves better scores for blurred inputs, except for the PSNR performance of severely blurred inputs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 583, + 329, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 583, + 329, + 594 + ], + "spans": [ + { + "bbox": [ + 131, + 583, + 329, + 594 + ], + "type": "text", + "content": "3.3 Limitations of Channel Attention" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "The results in Sec. 3.2 indicate that channel attention is better suited for propagating information over time. However, it also has an inherent flaw - the correlation among output channels will increase since each channel in the attention output is a weighted summation over channels of the value " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "V_{c}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ". We calculate the covariance following VICReg [1] for a quantitative measurement. Given deep" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "text", + "content": "RealViformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 480, + 140 + ], + "type": "text", + "content": "feature " + }, + { + "bbox": [ + 132, + 115, + 480, + 140 + ], + "type": "inline_equation", + "content": "z^n \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 132, + 115, + 480, + 140 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 132, + 115, + 480, + 140 + ], + "type": "inline_equation", + "content": "n \\in \\{1, N\\}" + }, + { + "bbox": [ + 132, + 115, + 480, + 140 + ], + "type": "text", + "content": " is the index of a sample, we reshape " + }, + { + "bbox": [ + 132, + 115, + 480, + 140 + ], + "type": "inline_equation", + "content": "z^n" + }, + { + "bbox": [ + 132, + 115, + 480, + 140 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 115, + 480, + 140 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C \\times HW}" + }, + { + "bbox": [ + 132, + 115, + 480, + 140 + ], + "type": "text", + "content": " and define the covariance matrix over " + }, + { + "bbox": [ + 132, + 115, + 480, + 140 + ], + "type": "inline_equation", + "content": "Z = \\{z^1, \\dots, z^N\\}" + }, + { + "bbox": [ + 132, + 115, + 480, + 140 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 220, + 149, + 481, + 183 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 149, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 220, + 149, + 481, + 183 + ], + "type": "interline_equation", + "content": "C o v (Z) = \\frac {1}{N - 1} \\sum_ {n = 1} ^ {N} \\left(z _ {n} - \\bar {z}\\right) \\left(z _ {n} - \\bar {z}\\right) ^ {T}, \\tag {3}", + "image_path": "52ab80e3c40ec6c38b269496bd43a399f78e2785af135c116e74f7ffee7c56e3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 192, + 479, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 192, + 479, + 301 + ], + "spans": [ + { + "bbox": [ + 130, + 192, + 479, + 301 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 192, + 479, + 301 + ], + "type": "inline_equation", + "content": "\\bar{z} = \\frac{1}{N}\\sum_{n=1}^{N}z_{n}" + }, + { + "bbox": [ + 130, + 192, + 479, + 301 + ], + "type": "text", + "content": ". We then define the indicator for the covariance matrix as " + }, + { + "bbox": [ + 130, + 192, + 479, + 301 + ], + "type": "inline_equation", + "content": "ac(Z) = \\frac{1}{d}\\sum_{i\\neq j}|Cov(Z)|_{i,j}" + }, + { + "bbox": [ + 130, + 192, + 479, + 301 + ], + "type": "text", + "content": ", i.e. the average of absolute off-diagonal coefficients of " + }, + { + "bbox": [ + 130, + 192, + 479, + 301 + ], + "type": "inline_equation", + "content": "Cov(Z)" + }, + { + "bbox": [ + 130, + 192, + 479, + 301 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 192, + 479, + 301 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 192, + 479, + 301 + ], + "type": "text", + "content": " is the number of off-diagonal coefficients. Function " + }, + { + "bbox": [ + 130, + 192, + 479, + 301 + ], + "type": "inline_equation", + "content": "ac(Z)" + }, + { + "bbox": [ + 130, + 192, + 479, + 301 + ], + "type": "text", + "content": " encodes the covariance among feature channels. Taking " + }, + { + "bbox": [ + 130, + 192, + 479, + 301 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 130, + 192, + 479, + 301 + ], + "type": "text", + "content": " in Fig. 2 for comparison, " + }, + { + "bbox": [ + 130, + 192, + 479, + 301 + ], + "type": "inline_equation", + "content": "ac(O)" + }, + { + "bbox": [ + 130, + 192, + 479, + 301 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 192, + 479, + 301 + ], + "type": "inline_equation", + "content": "N = 400" + }, + { + "bbox": [ + 130, + 192, + 479, + 301 + ], + "type": "text", + "content": " based on channel attention is 0.87 and significantly higher than the input features (" + }, + { + "bbox": [ + 130, + 192, + 479, + 301 + ], + "type": "inline_equation", + "content": "\\approx 0.15" + }, + { + "bbox": [ + 130, + 192, + 479, + 301 + ], + "type": "text", + "content": "). Instead, " + }, + { + "bbox": [ + 130, + 192, + 479, + 301 + ], + "type": "inline_equation", + "content": "ac(O)" + }, + { + "bbox": [ + 130, + 192, + 479, + 301 + ], + "type": "text", + "content": " of spatial attention remains similar to the input features. Previous works on representation learning [1] and overfitting [8] propose that the high covariance of feature channels indicates redundancy and tends to hinder informative predictions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 301, + 480, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 480, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 480, + 409 + ], + "type": "text", + "content": "Similarly, we speculate the redundancy effects will hinder the prediction of HR outputs when adopting channel attention in building blocks. For a closer look, we investigate the standard VSR task, which is artifact-free. Specifically, we build " + }, + { + "bbox": [ + 130, + 301, + 480, + 409 + ], + "type": "inline_equation", + "content": "M_{c}" + }, + { + "bbox": [ + 130, + 301, + 480, + 409 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 301, + 480, + 409 + ], + "type": "inline_equation", + "content": "M_{s}" + }, + { + "bbox": [ + 130, + 301, + 480, + 409 + ], + "type": "text", + "content": " by replacing the Residual Block in Fig. 3a with channel attention blocks [39] and spatial attention blocks [21]. Models are trained on the REDS dataset without extra degradation. For evaluation, we choose SSIM [35], which focuses on structural information. The SSIM score of " + }, + { + "bbox": [ + 130, + 301, + 480, + 409 + ], + "type": "inline_equation", + "content": "M_{c}" + }, + { + "bbox": [ + 130, + 301, + 480, + 409 + ], + "type": "text", + "content": " (0.8338) is lower than " + }, + { + "bbox": [ + 130, + 301, + 480, + 409 + ], + "type": "inline_equation", + "content": "M_{s}" + }, + { + "bbox": [ + 130, + 301, + 480, + 409 + ], + "type": "text", + "content": " (0.8432); " + }, + { + "bbox": [ + 130, + 301, + 480, + 409 + ], + "type": "inline_equation", + "content": "ac(\\cdot)" + }, + { + "bbox": [ + 130, + 301, + 480, + 409 + ], + "type": "text", + "content": " of the last features before the upsampling module is higher for channel attention than for spatial attention, i.e. 0.199 vs. 0.147." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 410, + 481, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 481, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 481, + 469 + ], + "type": "text", + "content": "We adopt two simple modifications to channel attention and boost informative features to verify the redundancy effect empirically. Fig. 5 shows the Improved Channel Attention (ICA) module. Our approach features two key steps to enhance the quality. First, we use the squeeze-and-excite mechanism to predict new information. The features are squeezed channel-wise to extract meaningful" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 205, + 492, + 410, + 597 + ], + "blocks": [ + { + "bbox": [ + 205, + 492, + 410, + 597 + ], + "lines": [ + { + "bbox": [ + 205, + 492, + 410, + 597 + ], + "spans": [ + { + "bbox": [ + 205, + 492, + 410, + 597 + ], + "type": "image", + "image_path": "1749f415e521704d711882a6df1461ef863dc7988a82718fb792f2f63de5a078.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 607, + 482, + 662 + ], + "lines": [ + { + "bbox": [ + 132, + 607, + 482, + 662 + ], + "spans": [ + { + "bbox": [ + 132, + 607, + 482, + 662 + ], + "type": "text", + "content": "Fig. 5: Improved Channel Attention Module (ICA), showing self-attention for simplicity. The 'squeeze' convolution compresses the number of input feature channels " + }, + { + "bbox": [ + 132, + 607, + 482, + 662 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 132, + 607, + 482, + 662 + ], + "type": "text", + "content": " by ratio " + }, + { + "bbox": [ + 132, + 607, + 482, + 662 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 132, + 607, + 482, + 662 + ], + "type": "text", + "content": ". The features are then rescaled by weights predicted from the " + }, + { + "bbox": [ + 132, + 607, + 482, + 662 + ], + "type": "inline_equation", + "content": "\\frac{C}{r} \\times \\frac{C}{r}" + }, + { + "bbox": [ + 132, + 607, + 482, + 662 + ], + "type": "text", + "content": " attention map before being expanded by the 'excite' convolution back to the original number of input channels." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "type": "text", + "content": "Y. Zhang and A. Yao" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 201, + 111, + 410, + 249 + ], + "blocks": [ + { + "bbox": [ + 201, + 111, + 410, + 249 + ], + "lines": [ + { + "bbox": [ + 201, + 111, + 410, + 249 + ], + "spans": [ + { + "bbox": [ + 201, + 111, + 410, + 249 + ], + "type": "image", + "image_path": "e4aba276499b6b85f5db0256db27dfe9fbb641783a645a51bccee794f785f330.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 270, + 254, + 341, + 262 + ], + "lines": [ + { + "bbox": [ + 270, + 254, + 341, + 262 + ], + "spans": [ + { + "bbox": [ + 270, + 254, + 341, + 262 + ], + "type": "text", + "content": "(c) Reconstruction Module" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 272, + 482, + 361 + ], + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 361 + ], + "type": "text", + "content": "Fig. 6: The framework of RealVformer. (a) Overview of RealVformer, following a unidirectional recurrent framework. The outputs of the Forward module are propagated to the next time step and upsampled by module " + }, + { + "bbox": [ + 130, + 272, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 130, + 272, + 482, + 361 + ], + "type": "text", + "content": " to get HR frames. (b) Explanation of the Forward module in (a), where " + }, + { + "bbox": [ + 130, + 272, + 482, + 361 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 272, + 482, + 361 + ], + "type": "text", + "content": " denotes the warping function. The reconstruction module " + }, + { + "bbox": [ + 130, + 272, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 130, + 272, + 482, + 361 + ], + "type": "text", + "content": " takes current frame " + }, + { + "bbox": [ + 130, + 272, + 482, + 361 + ], + "type": "inline_equation", + "content": "I_t^L" + }, + { + "bbox": [ + 130, + 272, + 482, + 361 + ], + "type": "text", + "content": " and warped hidden state " + }, + { + "bbox": [ + 130, + 272, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\hat{h}_{t-1}" + }, + { + "bbox": [ + 130, + 272, + 482, + 361 + ], + "type": "text", + "content": " as inputs. (c) Reconstruction module " + }, + { + "bbox": [ + 130, + 272, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 130, + 272, + 482, + 361 + ], + "type": "text", + "content": ". The shallow feature of " + }, + { + "bbox": [ + 130, + 272, + 482, + 361 + ], + "type": "inline_equation", + "content": "I_t^L" + }, + { + "bbox": [ + 130, + 272, + 482, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 272, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\hat{h}_{t-1}" + }, + { + "bbox": [ + 130, + 272, + 482, + 361 + ], + "type": "text", + "content": " are fused by CAF and then forwarded to Transformer blocks with U-shape connection [39]. Module GDFN follows Restormer [39]. Details of CAF and ICA modules are stated in Fig. 7 and Fig. 5." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 384, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 517 + ], + "type": "text", + "content": "information and then expanded into new channels based on the attention outputs. Secondly, we rescale channels in attention outputs by scalar weights predicted from the attention map. The attention map measures relationships across channels and encourages the associated convolutions in the 'excite' operation to yield more precise and discriminating features. Our designs are inspired by the SE Network [12] but with two key distinctions. First, we apply the squeeze-and-excite mechanism to generate new information, while SE Network uses it to help with scalar weight prediction. Second, the rescaling weights prediction takes the attention map as a cue rather than the naive pooling results of channels. In Sec. 4, we show the efficacy of new designs on the real-world VSR task and compare channel correlation with and without new designs in ablations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 535, + 235, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 535, + 235, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 535, + 235, + 547 + ], + "type": "text", + "content": "4 RealVifromer" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 561, + 208, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 561, + 208, + 572 + ], + "spans": [ + { + "bbox": [ + 132, + 561, + 208, + 572 + ], + "type": "text", + "content": "4.1 Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 481, + 640 + ], + "type": "text", + "content": "We apply our findings in Sec. 3.2 to the real-world VSR task as further support. To that end, we design RealVformer to incorporate channel attention as the basic processing module and the modification to boost informative features. Our emphasis here is not novel architecture design but a showcase of effectiveness brought by applying our findings in Sec. 3.2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 642, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 642, + 481, + 665 + ], + "type": "text", + "content": "RealVformer is a recurrent Transformer network with channel attention modules. Fig. 6 shows the model architecture. To reduce the computational" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "text", + "content": "RealViformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 192, + 127, + 416, + 205 + ], + "blocks": [ + { + "bbox": [ + 192, + 117, + 296, + 125 + ], + "lines": [ + { + "bbox": [ + 192, + 117, + 296, + 125 + ], + "spans": [ + { + "bbox": [ + 192, + 117, + 296, + 125 + ], + "type": "text", + "content": "Channel Attention Fusion module (CAF)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 192, + 127, + 416, + 205 + ], + "lines": [ + { + "bbox": [ + 192, + 127, + 416, + 205 + ], + "spans": [ + { + "bbox": [ + 192, + 127, + 416, + 205 + ], + "type": "image", + "image_path": "71297bf3b076a0ea814a16d09bf0c2e24b20f50ea114312c4bb2e3709fbd802c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 208, + 425, + 227 + ], + "lines": [ + { + "bbox": [ + 194, + 208, + 425, + 227 + ], + "spans": [ + { + "bbox": [ + 194, + 208, + 425, + 227 + ], + "type": "text", + "content": "1x1 Norm 3x3 Depth-wise " + }, + { + "bbox": [ + 194, + 208, + 425, + 227 + ], + "type": "inline_equation", + "content": "\\mathbb{C}" + }, + { + "bbox": [ + 194, + 208, + 425, + 227 + ], + "type": "text", + "content": " Concatenation Reshape " + }, + { + "bbox": [ + 194, + 208, + 425, + 227 + ], + "type": "inline_equation", + "content": "\\bigotimes" + }, + { + "bbox": [ + 194, + 208, + 425, + 227 + ], + "type": "text", + "content": " Matrix multiplication Conv" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 131, + 240, + 482, + 275 + ], + "lines": [ + { + "bbox": [ + 131, + 240, + 482, + 275 + ], + "spans": [ + { + "bbox": [ + 131, + 240, + 482, + 275 + ], + "type": "text", + "content": "Fig. 7: Details of Channel Attention Fusion (CAF) module. CAF gets the query from current frame feature " + }, + { + "bbox": [ + 131, + 240, + 482, + 275 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 131, + 240, + 482, + 275 + ], + "type": "text", + "content": " and {key, value} from hidden state " + }, + { + "bbox": [ + 131, + 240, + 482, + 275 + ], + "type": "inline_equation", + "content": "\\hat{h}_{t-1}" + }, + { + "bbox": [ + 131, + 240, + 482, + 275 + ], + "type": "text", + "content": ". The attention output is concatenated with " + }, + { + "bbox": [ + 131, + 240, + 482, + 275 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 131, + 240, + 482, + 275 + ], + "type": "text", + "content": " to process for module output " + }, + { + "bbox": [ + 131, + 240, + 482, + 275 + ], + "type": "inline_equation", + "content": "O_{t}" + }, + { + "bbox": [ + 131, + 240, + 482, + 275 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 303, + 482, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 379 + ], + "type": "text", + "content": "cost, we follow a widely used recurrent architecture [3] in a unidirectional setting. The model first estimates optical flow " + }, + { + "bbox": [ + 130, + 303, + 482, + 379 + ], + "type": "inline_equation", + "content": "s_{(t - 1)\\rightarrow t}^{f}" + }, + { + "bbox": [ + 130, + 303, + 482, + 379 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 303, + 482, + 379 + ], + "type": "inline_equation", + "content": "I_{t - 1}^{L}" + }, + { + "bbox": [ + 130, + 303, + 482, + 379 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 303, + 482, + 379 + ], + "type": "inline_equation", + "content": "I_{t}^{L}" + }, + { + "bbox": [ + 130, + 303, + 482, + 379 + ], + "type": "text", + "content": " through Spynet [29] and warps previous hidden state " + }, + { + "bbox": [ + 130, + 303, + 482, + 379 + ], + "type": "inline_equation", + "content": "h_{t - 1}" + }, + { + "bbox": [ + 130, + 303, + 482, + 379 + ], + "type": "text", + "content": " to current time step based on the flow. Frame " + }, + { + "bbox": [ + 130, + 303, + 482, + 379 + ], + "type": "inline_equation", + "content": "I_t^L" + }, + { + "bbox": [ + 130, + 303, + 482, + 379 + ], + "type": "text", + "content": " and spatially aligned hidden state " + }, + { + "bbox": [ + 130, + 303, + 482, + 379 + ], + "type": "inline_equation", + "content": "\\hat{h}_{t - 1}" + }, + { + "bbox": [ + 130, + 303, + 482, + 379 + ], + "type": "text", + "content": " are processed in the reconstruction module " + }, + { + "bbox": [ + 130, + 303, + 482, + 379 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 130, + 303, + 482, + 379 + ], + "type": "text", + "content": ". The hidden state is updated with outputs of " + }, + { + "bbox": [ + 130, + 303, + 482, + 379 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 130, + 303, + 482, + 379 + ], + "type": "text", + "content": " and further processed by the upsampling module " + }, + { + "bbox": [ + 130, + 303, + 482, + 379 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 130, + 303, + 482, + 379 + ], + "type": "text", + "content": " to output high-resolution frames." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 381, + 483, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 381, + 483, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 483, + 453 + ], + "type": "text", + "content": "The reconstruction module " + }, + { + "bbox": [ + 130, + 381, + 483, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 130, + 381, + 483, + 453 + ], + "type": "text", + "content": " uses channel attention in two ways. First, the Channel Attention Fusion (CAF) module fuses the temporal information to limit the produced artifacts in the hidden state. CAF queries the aligned hidden state " + }, + { + "bbox": [ + 130, + 381, + 483, + 453 + ], + "type": "inline_equation", + "content": "\\hat{h}_{t-1}" + }, + { + "bbox": [ + 130, + 381, + 483, + 453 + ], + "type": "text", + "content": " by the shallow feature " + }, + { + "bbox": [ + 130, + 381, + 483, + 453 + ], + "type": "inline_equation", + "content": "f_t" + }, + { + "bbox": [ + 130, + 381, + 483, + 453 + ], + "type": "text", + "content": ". Secondly, we take the Improved Channel Attention Module (ICA) in Fig. 5 to build the Transformer blocks for better HR reconstructions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 484, + 371, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 484, + 371, + 498 + ], + "spans": [ + { + "bbox": [ + 131, + 484, + 371, + 498 + ], + "type": "text", + "content": "4.2 Channel Attention Fusion (CAF) Module" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "text", + "content": "As explored in Sec. 3.2, adding channel attention promotes the real-world VSR performance compared to spatial attention and the simple concatenation baseline. Thus, we keep this design in our model and put it as the Channel Attention Fusion (CAF) module for temporal aggregation. Fig. 7 show details of how CAF perform channel attention between " + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "text", + "content": ", the feature of the current frame, and " + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\hat{h}_{t-1}" + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "text", + "content": ", the aligned propagated hidden state. The query " + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "inline_equation", + "content": "(Q_{t})" + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "text", + "content": " is generated as " + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "inline_equation", + "content": "Q_{t} = K_{3\\times 3} * \\text{LayerNorm}(f_{t})" + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "inline_equation", + "content": "K_{3\\times 3} *" + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "text", + "content": " refers to " + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "text", + "content": " convolution operation. Similarly, we process hidden state " + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\hat{h}_{t-1}" + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\tilde{h}_{t-1} = K_{3\\times 3}^{d} * K_{1\\times 1} * \\text{LayerNorm}(\\hat{h}_{t-1})" + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "inline_equation", + "content": "K_{3\\times 3}^{d} *" + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "text", + "content": " depth-wise convolution which double the channel number. Chunking " + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\tilde{h}_{t-1}" + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "text", + "content": " gives key " + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "inline_equation", + "content": "(K_{t})" + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "inline_equation", + "content": "(V_{t})" + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "text", + "content": ". The calculation of attention map " + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "inline_equation", + "content": "A_{t} \\in \\mathbb{R}^{C \\times C}" + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "text", + "content": " follows Eq. (2). The final output " + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "inline_equation", + "content": "O_{t}" + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "text", + "content": " is computed as " + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "inline_equation", + "content": "O_{t} = K_{1\\times 1} * K_{3\\times 3}^{d} * K_{1\\times 1} * \\mathbf{C}[A_{t}V_{t}; f_{t}]" + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{C}[\\cdot; \\cdot]" + }, + { + "bbox": [ + 130, + 519, + 482, + 668 + ], + "type": "text", + "content": " denotes concatenation." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "type": "text", + "content": "Y. Zhang and A. Yao" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 382, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 382, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 382, + 129 + ], + "type": "text", + "content": "4.3 Improved Channel Attention (ICA) Module" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 136, + 482, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 136, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 482, + 269 + ], + "type": "text", + "content": "The design of ICA follows Fig. 5 for empirically verifying findings in Sec. 3.3. Squeeze and Excite follows the squeeze-and-excite mechanism in Fig. 5 and helps to predict channels with new information. It squeezes the channels of input feature " + }, + { + "bbox": [ + 130, + 136, + 482, + 269 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 130, + 136, + 482, + 269 + ], + "type": "text", + "content": " by factor " + }, + { + "bbox": [ + 130, + 136, + 482, + 269 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 136, + 482, + 269 + ], + "type": "text", + "content": ". The attention map refers to self-attention on " + }, + { + "bbox": [ + 130, + 136, + 482, + 269 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 136, + 482, + 269 + ], + "type": "text", + "content": " and of size " + }, + { + "bbox": [ + 130, + 136, + 482, + 269 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{\\frac{C}{r} \\times \\frac{C}{r}}" + }, + { + "bbox": [ + 130, + 136, + 482, + 269 + ], + "type": "text", + "content": ". Excitation layers expand outputs back to size " + }, + { + "bbox": [ + 130, + 136, + 482, + 269 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 130, + 136, + 482, + 269 + ], + "type": "text", + "content": ". Correlation-based channel weighting weights the output channel based on scalars predicted from the attention map. This design emphasizes channels for predicting more discriminative features in 'excite' operation. The attention map " + }, + { + "bbox": [ + 130, + 136, + 482, + 269 + ], + "type": "inline_equation", + "content": "A_r \\in \\mathbb{R}^{\\frac{C}{r} \\times \\frac{C}{r}}" + }, + { + "bbox": [ + 130, + 136, + 482, + 269 + ], + "type": "text", + "content": " is taken for calculating the average and max values along the rows. The average and max values are combined and mapped to weights of size " + }, + { + "bbox": [ + 130, + 136, + 482, + 269 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{\\frac{C}{r} \\times 1}" + }, + { + "bbox": [ + 130, + 136, + 482, + 269 + ], + "type": "text", + "content": " through linear layers and a sigmoid function." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 285, + 279, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 279, + 297 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 279, + 297 + ], + "type": "text", + "content": "4.4 Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 304, + 481, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 481, + 365 + ], + "type": "text", + "content": "Model details. RealVformer uses SPyNet [29] for flow estimation. After the CAF module, the reconstruction module " + }, + { + "bbox": [ + 130, + 304, + 481, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 130, + 304, + 481, + 365 + ], + "type": "text", + "content": " has a three-level encoder-decoder architecture. From level 1 to level 3, there are [2,3,4] transformer blocks with [48,96,192] channels. There are [1,2,4] attention heads in the ICA, all with a squeeze factor of 4. Supplementary B.1 gives the detailed architecture of " + }, + { + "bbox": [ + 130, + 304, + 481, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 130, + 304, + 481, + 365 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 365, + 481, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 481, + 435 + ], + "type": "text", + "content": "Training details. We train using the REDS dataset [27] and follow RealBasicVSR [5] in applying random combinations of blur, noise, JPEG compression, and video compression for synthesizing input degradations. We load 15 frames as an input sequence. The spatial size of inputs is cropped to " + }, + { + "bbox": [ + 130, + 365, + 481, + 435 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 130, + 365, + 481, + 435 + ], + "type": "text", + "content": ", and the batch size is 16. We use a pre-trained flow estimation model SPyNet, the parameter of which is fixed for the first 5K iterations and tuned with other modules later." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 436, + 481, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 481, + 534 + ], + "type": "text", + "content": "Following RealBasicVSR [5], we perform two-stage training. The first stage trains the model with a Charbonnier loss [19] and SSIM [35] loss for 300K iterations. In the second stage, the model is trained for another 130K iterations with the Charbonnier loss, SSIM loss, perceptual loss [18] and GAN loss [11] together, weighted by 1, 0.001, 1, and 0.005, respectively. The implementations of perceptual loss, GAN loss, and discriminator follow RealBasicVSR [5]. We implement all experiments on 4 Quadro RTX 8000 GPUs with PyTorch. Other details of the training settings are in Supplementary B.2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 550, + 270, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 550, + 270, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 550, + 270, + 562 + ], + "type": "text", + "content": "4.5 Experimental Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "We compare our model on four datasets, VideoLQ [5], RealVSR [37], REDS4 [27] and UDM10 [38], with RealBasicVSR [5], Real-ESRGAN [34], BSRGAN [40], DBVSR [28], RealVSR [37], DAN [14], and RealSR [17]. VideoLQ and RealVSR are collected from real-world scenarios. VideoLQ is an unpaired dataset. RealVSR has same-size paired low-quality (LQ) and high-quality (HQ) frames. Our model super-resolves the input frames spatially. Although downsampling LQ enables paired data, it alters the original degradation. Thus, we test on the original LQ and do not use the HQ for evaluation. REDS4 and UDM10 have" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "type": "text", + "content": "RealViformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 168, + 482, + 290 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 2: Quantitative comparisons with existing methods with best and second-best results. Our method achieves the best ILNIQE and NRQM scores over VideoLQ and RealVSR datasets and the best PSNR, SSIM, and LPIPS over synthetic datasets REDS4 and UDM10 with relatively few parameters and the lowest run-time." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 168, + 482, + 290 + ], + "lines": [ + { + "bbox": [ + 135, + 168, + 482, + 290 + ], + "spans": [ + { + "bbox": [ + 135, + 168, + 482, + 290 + ], + "type": "table", + "html": "
RealSRDANRealVSRDBVSRBSRGANReal-ESRGANRealBasicVSROurs
Params (M)16.74.32.725.516.716.76.35.3
Runtime (ms)180250772-1801967349
VideoLQILNIQE↓26.6328.2831.9427.8527.4927.9725.9825.94
NRQM↑6.0543.7423.4603.8516.1566.0576.3066.338
RealVSRILNIQE↓32.8132.2934.39-32.6531.9330.3728.61
NRQM↑5.6103.5233.795-6.1526.2456.5826.588
REDS4PSNR↑22.0222.6718.3022.3522.9421.5623.0923.34
SSIM↑0.50970.55710.49000.55300.57500.55560.60760.6079
LPIPS↓0.59910.63150.72400.62110.37660.35330.29910.2877
UDM10PSNR↑25.3725.9023.3525.0825.9724.9625.9626.42
SSIM↑0.66580.72290.71150.71120.75680.74320.74910.7609
LPIPS↓0.48110.47810.47610.47560.33880.33950.32090.3063
", + "image_path": "bbec08fb949c97929cafd7cf6c1c6da610c575bf3509ad7dcb91c858a390957b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 319, + 482, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 368 + ], + "type": "text", + "content": "ground-truth images, and we synthesize low-quality inputs through the same pipeline during training. All tested recurrent-based VSR models load the half sequence for videos in RealVSR and the whole sequence for others each time. The quantitative and qualitative results are discussed below." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 371, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 371, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 482, + 479 + ], + "type": "text", + "content": "Quantitative Results. For evaluation without reference, we apply the ILNIQE [41], the improved version of NIQE [26], and NRQM [24] metrics based on each output sequence's first, middle, and last frames in RGB format [5]. These metrics appear less biased towards oversharpened features; details are given in Supplementary B.3. For evaluation on REDS4 and UDM10, we report more reliable metrics, PSNR, SSIM, and LPIPS [42]. We collect released models of all compared methods and generate sequences. As shown in Tab. 2, RealVformer performs better than other methods with smaller parameters than the most competitive RealBasicVSR [5] and the shortest runtime." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 135, + 506, + 202, + 620 + ], + "blocks": [ + { + "bbox": [ + 135, + 506, + 202, + 620 + ], + "lines": [ + { + "bbox": [ + 135, + 506, + 202, + 620 + ], + "spans": [ + { + "bbox": [ + 135, + 506, + 202, + 620 + ], + "type": "image", + "image_path": "7154e1d49620a554ebc3214d43090340255fe01d6f563fb14ab126472cc6e278.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 152, + 621, + 187, + 630 + ], + "lines": [ + { + "bbox": [ + 152, + 621, + 187, + 630 + ], + "spans": [ + { + "bbox": [ + 152, + 621, + 187, + 630 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 204, + 506, + 271, + 620 + ], + "blocks": [ + { + "bbox": [ + 204, + 506, + 271, + 620 + ], + "lines": [ + { + "bbox": [ + 204, + 506, + 271, + 620 + ], + "spans": [ + { + "bbox": [ + 204, + 506, + 271, + 620 + ], + "type": "image", + "image_path": "79300de9e973112bf493dfba011b44c494ff1c3d93bc9e56c00fb17c687553f2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 621, + 252, + 630 + ], + "lines": [ + { + "bbox": [ + 225, + 621, + 252, + 630 + ], + "spans": [ + { + "bbox": [ + 225, + 621, + 252, + 630 + ], + "type": "text", + "content": "(b) RE" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 273, + 506, + 340, + 620 + ], + "blocks": [ + { + "bbox": [ + 273, + 506, + 340, + 620 + ], + "lines": [ + { + "bbox": [ + 273, + 506, + 340, + 620 + ], + "spans": [ + { + "bbox": [ + 273, + 506, + 340, + 620 + ], + "type": "image", + "image_path": "754704fc04ad4b48fa413195c389fc83f074ec91c3db1c74d0d58cc64e20e0a1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 294, + 621, + 320, + 630 + ], + "lines": [ + { + "bbox": [ + 294, + 621, + 320, + 630 + ], + "spans": [ + { + "bbox": [ + 294, + 621, + 320, + 630 + ], + "type": "text", + "content": "(c) RB" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 639, + 481, + 663 + ], + "lines": [ + { + "bbox": [ + 131, + 639, + 481, + 663 + ], + "spans": [ + { + "bbox": [ + 131, + 639, + 481, + 663 + ], + "type": "text", + "content": "Fig. 8: Qualitative comparisons on synthetic datasets. Our method produces clearer than RealESRGAN (RE) [34] and RealBasicVSR (RB) [5] for very hard inputs." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 342, + 506, + 408, + 620 + ], + "blocks": [ + { + "bbox": [ + 342, + 506, + 408, + 620 + ], + "lines": [ + { + "bbox": [ + 342, + 506, + 408, + 620 + ], + "spans": [ + { + "bbox": [ + 342, + 506, + 408, + 620 + ], + "type": "image", + "image_path": "b4b7fcdd3a127285dfa26501028d6ba6ed6f799dd9054c8c5a593e99e6bfd4e7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 621, + 392, + 630 + ], + "lines": [ + { + "bbox": [ + 359, + 621, + 392, + 630 + ], + "spans": [ + { + "bbox": [ + 359, + 621, + 392, + 630 + ], + "type": "text", + "content": "(d) Ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 410, + 506, + 477, + 620 + ], + "blocks": [ + { + "bbox": [ + 410, + 506, + 477, + 620 + ], + "lines": [ + { + "bbox": [ + 410, + 506, + 477, + 620 + ], + "spans": [ + { + "bbox": [ + 410, + 506, + 477, + 620 + ], + "type": "image", + "image_path": "1cf4db5470437749b3c26ad8ad3585bc5933cd921158cd7c4bffff6fec5300b9.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 432, + 621, + 458, + 630 + ], + "lines": [ + { + "bbox": [ + 432, + 621, + 458, + 630 + ], + "spans": [ + { + "bbox": [ + 432, + 621, + 458, + 630 + ], + "type": "text", + "content": "(e) GT" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "type": "text", + "content": "Y. Zhang and A. Yao" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 220, + 225 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 220, + 225 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 220, + 225 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 220, + 225 + ], + "type": "image", + "image_path": "b60ce4e5e99cf6e1031caf5b5ce7b5e252fbb35e8b12b05b2fe1eefd517135e6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 161, + 227, + 195, + 236 + ], + "lines": [ + { + "bbox": [ + 161, + 227, + 195, + 236 + ], + "spans": [ + { + "bbox": [ + 161, + 227, + 195, + 236 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 221, + 114, + 306, + 225 + ], + "blocks": [ + { + "bbox": [ + 221, + 114, + 306, + 225 + ], + "lines": [ + { + "bbox": [ + 221, + 114, + 306, + 225 + ], + "spans": [ + { + "bbox": [ + 221, + 114, + 306, + 225 + ], + "type": "image", + "image_path": "5bddf52abdf27bd053de4eb6034dfb839fa72ecf7a7bb857dec2a97e6c1232d3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 250, + 227, + 277, + 236 + ], + "lines": [ + { + "bbox": [ + 250, + 227, + 277, + 236 + ], + "spans": [ + { + "bbox": [ + 250, + 227, + 277, + 236 + ], + "type": "text", + "content": "(b) RE" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 392, + 225 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 392, + 225 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 392, + 225 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 392, + 225 + ], + "type": "image", + "image_path": "87acee6190ccde9bce59c79d4b1e0f2792378f80e62b05e06062a55971effd69.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 227, + 362, + 236 + ], + "lines": [ + { + "bbox": [ + 337, + 227, + 362, + 236 + ], + "spans": [ + { + "bbox": [ + 337, + 227, + 362, + 236 + ], + "type": "text", + "content": "(c) RB" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 245, + 479, + 269 + ], + "lines": [ + { + "bbox": [ + 132, + 245, + 479, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 245, + 479, + 269 + ], + "type": "text", + "content": "Fig. 9: Qualitative comparisons on real-world datasets. Our method produces less high-frequency artifacts and overshoot effects than RealESRGAN [34] and RealBasicVSR [5]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 392, + 114, + 479, + 225 + ], + "blocks": [ + { + "bbox": [ + 392, + 114, + 479, + 225 + ], + "lines": [ + { + "bbox": [ + 392, + 114, + 479, + 225 + ], + "spans": [ + { + "bbox": [ + 392, + 114, + 479, + 225 + ], + "type": "image", + "image_path": "47abac7bb4302e1cd80e6bbd5fb217d77e3703c2037e2053d5baeb1af4918a53.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 420, + 227, + 452, + 236 + ], + "lines": [ + { + "bbox": [ + 420, + 227, + 452, + 236 + ], + "spans": [ + { + "bbox": [ + 420, + 227, + 452, + 236 + ], + "type": "text", + "content": "(d) Ours" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 292, + 480, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 292, + 480, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 480, + 339 + ], + "type": "text", + "content": "Qualitative Results. We show qualitative comparisons on synthetic (see Fig. 8) and real-world (see Fig. 9) datasets. Compared to the listed methods, RealViformer generates clear structures with much fewer high-frequency artifacts. More visual comparisons are in Supplementary B.5." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 340, + 480, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 480, + 413 + ], + "type": "text", + "content": "User Study. We also performed a user study and asked 30 evaluators on Amazon MTurk to score reconstructions for 85 frames sampled from datasets VideoLQ [5], RealVSR [37], REDS4 [27] and UDM10 [38]. Each worker saw five HR results of the same frame and rated them based on the visual quality, from 1 (the worst) to 5 (the best); as shown in Fig. 10, our method surpasses BSRGAN [40], Real-ESRGAN [34], RealSR [17], and RealBasicVSR [5]." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 141, + 434, + 264, + 526 + ], + "blocks": [ + { + "bbox": [ + 141, + 434, + 264, + 526 + ], + "lines": [ + { + "bbox": [ + 141, + 434, + 264, + 526 + ], + "spans": [ + { + "bbox": [ + 141, + 434, + 264, + 526 + ], + "type": "image", + "image_path": "69fc77c9b7463039c19a06fe44f3f9f4c460eb54e7081c79daeca8b43c650419.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 537, + 273, + 593 + ], + "lines": [ + { + "bbox": [ + 131, + 537, + 273, + 593 + ], + "spans": [ + { + "bbox": [ + 131, + 537, + 273, + 593 + ], + "type": "text", + "content": "Fig.10: User study results from 30 evaluators on 85 frames. Our method achieves the best among all five methods regrading mean option scores (MOS)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 287, + 517, + 469, + 588 + ], + "blocks": [ + { + "bbox": [ + 276, + 449, + 480, + 516 + ], + "lines": [ + { + "bbox": [ + 276, + 449, + 480, + 516 + ], + "spans": [ + { + "bbox": [ + 276, + 449, + 480, + 516 + ], + "type": "text", + "content": "Table 3: Ablations of CAF and ICA modules. Channel-attention baseline (ch-baseline) performs better than the spatial-attention baseline (sp-baseline). The CAF module improves the performances for both datasets, and ICA further improves the performances to the state-of-the-art." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 287, + 517, + 469, + 588 + ], + "lines": [ + { + "bbox": [ + 287, + 517, + 469, + 588 + ], + "spans": [ + { + "bbox": [ + 287, + 517, + 469, + 588 + ], + "type": "table", + "html": "
MethodCAFICAVideoLQ NRQM↑UDM10 LPIPS↓
Sp-baseline--6.0610.3482
Ch-baselineXX6.1810.3085
RealVformer-✓X6.1960.2933
RealVformer✓✓6.3380.2877
", + "image_path": "64b854cd570e9665dc5fb6907d03a1982abd2918a8565450bb47afbbfe497d41.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "type": "text", + "content": "Ablation Studies. We conduct ablations to validate the advantage of channel attention and the efficacy of the CAF and ICA. We build a spatial-attention baseline (sp-baseline) by replacing the reconstruction module of BasicVSR [3] with SwinIR [21]. Channel-attention baseline (ch-baseline) is built with the same" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "text", + "content": "RealViformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 118, + 468, + 203 + ], + "blocks": [ + { + "bbox": [ + 140, + 118, + 468, + 203 + ], + "lines": [ + { + "bbox": [ + 140, + 118, + 468, + 203 + ], + "spans": [ + { + "bbox": [ + 140, + 118, + 468, + 203 + ], + "type": "image", + "image_path": "9d56bebcd12991c96ab8b84c91e0d6d0131003d9a103dfffc935862074792869.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 213, + 482, + 247 + ], + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 247 + ], + "type": "text", + "content": "Fig. 11: (a) Visual comparison between RealVformer and its ablations. Red circles highlight the improved details. (b) Radial Power Spectrum (RPS) of model predictions. Using ICA improves the power of high-frequency components (blue region)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 270, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 415 + ], + "type": "text", + "content": "overall architecture as RealVformer but replaces CAF with simple concatenation and substitutes ICA with the channel attention block in Restormer [39]. RealVformer" + }, + { + "bbox": [ + 130, + 270, + 482, + 415 + ], + "type": "inline_equation", + "content": "^{-}" + }, + { + "bbox": [ + 130, + 270, + 482, + 415 + ], + "type": "text", + "content": " applies CAF on ch-baseline. All models are trained with the same settings in Sec. 4.4. We report NRQM scores for the VideoLQ [5] dataset and LPIPS scores for the UDM10 [38] da set. As shown in Tab. 3, using the original channel attention module, ch-baseline already yields better performances than sp-baseline. The CAF and ICA modules further improve RealVformer to state-of-the-art performance with channel correlation of propagated information decreasing from 0.436 to 0.422. Fig. 11a visually compares RealVformer with its ablations. Applying CAF reduces artifacts, while ICA provides further improvements. Fig. 11b supplements the Radial Power Spectrum of model predictions. The blur region shows ICA increases the power in the high-frequency region." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 431, + 220, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 220, + 443 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 220, + 443 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 456, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 624 + ], + "type": "text", + "content": "This paper proposes a real-world VSR model, RealVformer, based on findings from investigating channel and spatial attention in a real-world setting. Explorations reveal that channel attention is less sensitive to the artifacts in query and better serves as a temporal aggregation module to limit model-produced artifacts in hidden states. Additionally, we observe the higher covariance of channel attention outputs and propose the Improved Channel Attention (ICA) Module with a squeeze-and-excite and a covariance-based rescaling mechanism. Based on our findings, we build RealVformer, a channel-attention-based recurrent model for real-world VSR. We propose the CAF module to limit artifact propagation and use the ICA module to achieve better reconstructions. RealVformer performs state-of-the-art on two real-world video datasets with fewer parameters and shorter runtime. On the other hand, we value our findings w.r.t. comparisons between channel and spatial attention and exploration of covariance in channel attention as inspiration for further real-world VSR research." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Acknowledgement This research is supported by the National Research Foundation, Singapore under its NRF Fellowship for AI (NRF-NRFFAI1-2019-0001)." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "type": "text", + "content": "Y. Zhang and A. Yao" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 153 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 153 + ], + "type": "text", + "content": "Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of National Research Foundation, Singapore." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 175, + 199, + 188 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 175, + 199, + 188 + ], + "spans": [ + { + "bbox": [ + 132, + 175, + 199, + 188 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 204, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 204, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 204, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 204, + 480, + 226 + ], + "type": "text", + "content": "1. Bardes, A., Ponce, J., LeCun, Y.: Vicreg: Variance-invariance-covariance regularization for self-supervised learning. arXiv preprint arXiv:2105.04906 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 228, + 480, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 249 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 249 + ], + "type": "text", + "content": "2. Cao, J., Li, Y., Zhang, K., Van Gool, L.: Video super-resolution transformer. arXiv preprint arXiv:2106.06847 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 251, + 480, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 480, + 294 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 480, + 294 + ], + "type": "text", + "content": "3. Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: Basicvsr: The search for essential components in video super-resolution and beyond. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4947-4956 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 295, + 480, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 295, + 480, + 338 + ], + "spans": [ + { + "bbox": [ + 138, + 295, + 480, + 338 + ], + "type": "text", + "content": "4. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5972-5981 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 340, + 480, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 340, + 480, + 372 + ], + "spans": [ + { + "bbox": [ + 138, + 340, + 480, + 372 + ], + "type": "text", + "content": "5. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Investigating tradeoffs in real-world video super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5962-5971 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 373, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 373, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 373, + 480, + 407 + ], + "type": "text", + "content": "6. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12299-12310 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 407, + 480, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 480, + 440 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 480, + 440 + ], + "type": "text", + "content": "7. Chen, X., Wang, X., Zhou, J., Qiao, Y., Dong, C.: Activating more pixels in image super-resolution transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22367-22377 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 441, + 480, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 441, + 480, + 473 + ], + "spans": [ + { + "bbox": [ + 138, + 441, + 480, + 473 + ], + "type": "text", + "content": "8. Cogswell, M., Ahmed, F., Girshick, R., Zitnick, L., Batra, D.: Reducing overfitting in deep networks by decorrelating representations. arXiv preprint arXiv:1511.06068 (2015)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 475, + 480, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 475, + 480, + 507 + ], + "spans": [ + { + "bbox": [ + 138, + 475, + 480, + 507 + ], + "type": "text", + "content": "9. Dai, T., Cai, J., Zhang, Y., Xia, S.T., Zhang, L.: Second-order attention network for single image super-resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11065-11074 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 509, + 480, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 509, + 480, + 541 + ], + "spans": [ + { + "bbox": [ + 138, + 509, + 480, + 541 + ], + "type": "text", + "content": "0. Fuoli, D., Gu, S., Timofte, R.: Efficient video super-resolution through recurrent latent space propagation. In: 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW). pp. 3476-3485. IEEE (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 543, + 480, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 543, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 138, + 543, + 480, + 575 + ], + "type": "text", + "content": "1. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 576, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 576, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 576, + 480, + 597 + ], + "type": "text", + "content": "2. Hu, J., Shen, L., Sun, G.: Squeeze-and-excitation networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7132-7141 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 480, + 632 + ], + "type": "text", + "content": "3. Hua, T., Wang, W., Xue, Z., Ren, S., Wang, Y., Zhao, H.: On feature decorrelation in self-supervised learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9598-9608 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "4. Huang, Y., Li, S., Wang, L., Tan, T., et al.: Unfolding the alternating optimization for blind super resolution. Advances in Neural Information Processing Systems 33, 5632-5643 (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "type": "text", + "content": "RealViformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "15. Isobe, T., Jia, X., Gu, S., Li, S., Wang, S., Tian, Q.: Video super-resolution with recurrent structure-detail network. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16. pp. 645-660. Springer (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 481, + 203 + ], + "type": "text", + "content": "16. Jeelani, M., Cheema, N., Illgner-Fehns, K., Slusallek, P., Jaiswal, S., et al.: Expanding synthetic real-world degradations for blind video super resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1199-1208 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 203, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 203, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 203, + 481, + 236 + ], + "type": "text", + "content": "17. Ji, X., Cao, Y., Tai, Y., Wang, C., Li, J., Huang, F.: Real-world super-resolution via kernel estimation and noise injection. In: proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 466-467 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 236, + 481, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 481, + 280 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 481, + 280 + ], + "type": "text", + "content": "18. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 694-711. Springer (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 280, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 280, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 133, + 280, + 481, + 311 + ], + "type": "text", + "content": "19. Lai, W.S., Huang, J.B., Ahuja, N., Yang, M.H.: Fast and accurate image superresolution with deep laplacian pyramid networks. IEEE transactions on pattern analysis and machine intelligence 41(11), 2599-2613 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 311, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 481, + 332 + ], + "type": "text", + "content": "20. Liang, J., Cao, J., Fan, Y., Zhang, K., Ranjan, R., Li, Y., Timofte, R., Van Gool, L.: Vrt: A video restoration transformer. arXiv preprint arXiv:2201.12288 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "type": "text", + "content": "21. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1833-1844 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 365, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 481, + 407 + ], + "type": "text", + "content": "22. Liang, J., Fan, Y., Xiang, X., Ranjan, R., Ilg, E., Green, S., Cao, J., Zhang, K., Timofte, R., Gool, L.V.: Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems 35, 378-393 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 407, + 481, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 451 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 451 + ], + "type": "text", + "content": "23. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 450, + 481, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 481, + 483 + ], + "type": "text", + "content": "24. Ma, C., Yang, C.Y., Yang, X., Yang, M.H.: Learning a no-reference quality metric for single-image super-resolution. Computer Vision and Image Understanding 158, 1-16 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 483, + 481, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 481, + 515 + ], + "type": "text", + "content": "25. Mei, Y., Fan, Y., Zhou, Y.: Image super-resolution with non-local sparse attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3517-3526 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 515, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 481, + 536 + ], + "type": "text", + "content": "26. Mittal, A., Soundararajan, R., Bovik, A.C.: Making a \"completely blind\" image quality analyzer. IEEE Signal processing letters 20(3), 209-212 (2012)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 536, + 481, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 580 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 580 + ], + "type": "text", + "content": "27. Nah, S., Baik, S., Hong, S., Moon, G., Son, S., Timofte, R., Mu Lee, K.: Ntire 2019 challenge on video deblurring and super-resolution: Dataset and study. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 580, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 580, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 580, + 481, + 611 + ], + "type": "text", + "content": "28. Pan, J., Bai, H., Dong, J., Zhang, J., Tang, J.: Deep blind video super-resolution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4811-4820 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "text", + "content": "29. Ranjan, A., Black, M.J.: Optical flow estimation using a spatial pyramid network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4161-4170 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "30. Shi, S., Gu, J., Xie, L., Wang, X., Yang, Y., Dong, C.: Rethinking alignment in video super-resolution transformers. arXiv preprint arXiv:2207.08494 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 254, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 254, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 254, + 101 + ], + "type": "text", + "content": "Y. Zhang and A. Yao" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 567 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "31. Song, Y., Wang, M., Yang, Z., Xian, X., Shi, Y.: Negvsr: Augmenting negatives for generalized noise modeling in real-world video super-resolution. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 10705-10713 (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "text", + "content": "32. Wang, H., Chen, X., Ni, B., Liu, Y., Liu, J.: Omni aggregation networks for lightweight image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22378-22387 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 481, + 226 + ], + "type": "text", + "content": "33. Wang, X., Chan, K.C., Yu, K., Dong, C., Change Loy, C.: Edvr: Video restoration with enhanced deformable convolutional networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 226, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 481, + 259 + ], + "type": "text", + "content": "34. Wang, X., Xie, L., Dong, C., Shan, Y.: Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1905-1914 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 481, + 293 + ], + "type": "text", + "content": "35. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 293, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 481, + 314 + ], + "type": "text", + "content": "36. Xie, L., Wang, X., Shi, S., Gu, J., Dong, C., Shan, Y.: Mitigating artifacts in real-world video super-resolution models. arXiv preprint arXiv:2212.07339 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 314, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 481, + 358 + ], + "type": "text", + "content": "37. Yang, X., Xiang, W., Zeng, H., Zhang, L.: Real-world video super-resolution: A benchmark dataset and a decomposition based learning scheme. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4781-4790 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 358, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 481, + 402 + ], + "type": "text", + "content": "38. Yi, P., Wang, Z., Jiang, K., Jiang, J., Ma, J.: Progressive fusion video superresolution network via exploiting non-local spatio-temporal correlations. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3106-3115 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 402, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 481, + 445 + ], + "type": "text", + "content": "39. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5728-5739 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 446, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 481, + 479 + ], + "type": "text", + "content": "40. Zhang, K., Liang, J., Van Gool, L., Timofte, R.: Designing a practical degradation model for deep blind image super-resolution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4791-4800 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 479, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 481, + 501 + ], + "type": "text", + "content": "41. Zhang, L., Zhang, L., Bovik, A.C.: A feature-enriched completely blind image quality evaluator. IEEE Transactions on Image Processing 24(8), 2579-2591 (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 501, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 481, + 533 + ], + "type": "text", + "content": "42. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 533, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 567 + ], + "type": "text", + "content": "43. Zhang, Y., Li, K., Li, K., Wang, L., Zhong, B., Fu, Y.: Image super-resolution using very deep residual channel attention networks. In: Proceedings of the European conference on computer vision (ECCV). pp. 286-301 (2018)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "text", + "content": "RealViformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Realistic Human Motion Generation with Cross-Diffusion Models/5b1805dd-89b3-4df9-9e0a-1ed375e3f44c_content_list.json b/2024/Realistic Human Motion Generation with Cross-Diffusion Models/5b1805dd-89b3-4df9-9e0a-1ed375e3f44c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ee1daaa1c870d7e311bbbf25ddcefb331c6f9ab9 --- /dev/null +++ b/2024/Realistic Human Motion Generation with Cross-Diffusion Models/5b1805dd-89b3-4df9-9e0a-1ed375e3f44c_content_list.json @@ -0,0 +1,1712 @@ +[ + { + "type": "text", + "text": "Realistic Human Motion Generation with Cross-Diffusion Models", + "text_level": 1, + "bbox": [ + 263, + 140, + 743, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zeping Ren $^{1*}$ , Shaoli Huang $^{2\\dagger}$ , Xiu Li $^{1\\dagger}$", + "bbox": [ + 339, + 210, + 663, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tsinghua Shenzhen International Graduate School \n $^{2}$ Tencent AI Lab", + "bbox": [ + 326, + 238, + 676, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In this work, we introduce the Cross Human Motion Diffusion Model (CrossDiff $^3$ ), a novel approach for generating high-quality human motion based on textual descriptions. Our method integrates 3D and 2D information using a shared transformer network within the training of the diffusion model, unifying motion noise into a single feature space. This enables cross-decoding of features into both 3D and 2D motion representations, regardless of their original dimension. The primary advantage of CrossDiff is its cross-diffusion mechanism, which allows the model to reverse either 2D or 3D noise into clean motion during training. This capability leverages the complementary information in both motion representations, capturing intricate human movement details often missed by models relying solely on 3D information. Consequently, CrossDiff effectively combines the strengths of both representations to generate more realistic motion sequences. In our experiments, our model demonstrates competitive state-of-the-art performance on text-to-motion benchmarks. Moreover, our method consistently provides enhanced motion generation quality, capturing complex full-body movement intricacies. Additionally, with a pre-trained model, our approach accommodates using in-the-wild 2D motion data without 3D motion ground truth during training to generate 3D motion, highlighting its potential for broader applications and efficient use of available data resources.", + "bbox": [ + 261, + 303, + 743, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 619, + 375, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, the field of human motion synthesis [7,9,10,14,15,18,41,50] has witnessed significant advancements, primarily driven by the growing demand for high-quality, realistic motion generation in applications such as gaming, virtual reality, and robotics.", + "bbox": [ + 212, + 651, + 787, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A crucial aspect in this research area is generating human motion based on textual descriptions, enabling contextually accurate and natural movements [40]. However, current methods [4,8,26,40] predominantly rely on 3D motion information during training, leading to an inability to capture the full spectrum of intricacies associated with human motion. When using only 3D representation,", + "bbox": [ + 212, + 712, + 787, + 787 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Work done during an internship at Tencent AI Lab.", + "bbox": [ + 202, + 796, + 563, + 811 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 204, + 811, + 369, + 825 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "3 https://wonderno.github.io/CrossDiff-webpage/", + "bbox": [ + 204, + 825, + 535, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aa7e76b8920840cc33a9d3d660034946fef1934e781406e8d1c390bdcddfccfa.jpg", + "image_caption": [ + "Fig. 1: Our method utilizing the cross-diffusion mechanism (Left) exhibits more full-body details compared to existing methods (Right)." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 787, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the generation model may struggle to relate text semantics to some body part movements with very small movement variations compared to others, which can lead to overlooking important motion details. This is because the model might focus on more dominant or larger movements within the 3D space, leaving subtle nuances underrepresented. For example, when given a prompt such as \"a person is dancing eloquently,\" as illustrated in Figure 1, the generated motion might lack vitality, display a limited range of movements, and contain minimal local motion details.", + "bbox": [ + 212, + 339, + 787, + 459 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To effectively address the limitations and accurately capture the nuances of full-body movement, we introduce the Cross Human Motion Diffusion Model (CrossDiff). This innovative approach seamlessly integrates and leverages both 3D and 2D motion information to generate high-quality human motion sequences. The 2D data representation effectively illustrates the intricacies of human body movements from various viewing angle projections. Due to different view projections in 2D data, small body part movements can be magnified in certain projections, making them more noticeable and easier to capture. This helps the text-to-motion generation models to better associate text descriptions with a wider range of human body motion details, including subtle movements that might have been overlooked when relying solely on 3D representation.", + "bbox": [ + 212, + 460, + 787, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As a result, incorporating 2D information with 3D enables the diffusion model to establish more connections between motion and text prompts, ultimately enhancing the motion synthesis process. The CrossDiff learning framework consists of two main components: unified encoding and cross-decoding. These components work together to achieve more precise and realistic motion synthesis. Furthermore, it is essential to transfer the knowledge acquired in the 2D domain to 3D motion, which leads to an overall improvement in the model's performance.", + "bbox": [ + 212, + 627, + 787, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unified encoding fuses motion noise from both 3D and 2D sources into a single feature space, facilitating cross-decoding of features into either 3D or 2D motion representations, regardless of their original dimension. The distinctive innovation of our approach stems from the cross-diffusion mechanism, which enables the model to transform 2D or 3D noise into clean motion during the training process. This capability allows the model to harness the complementary information present in both motion representations, effectively capturing", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z. Ren et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "intricate details of human movement that are often missed by models relying exclusively on 3D data.", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In experiments, we demonstrate our model achieves competitive state-of-the-art performance on several text-to-motion benchmarks, outperforming existing diffusion-based approaches that rely solely on 3D motion information during training. Furthermore, our method consistently delivers enhanced motion generation quality, capturing complex full-body movement intricacies essential for realistic motion synthesis. A notable advantage of our approach is its ability to utilize 2D motion data without necessitating 3D motion ground truth during training, enabling the generation of 3D motion. This feature underscores the potential of the CrossDiff model for a wide range of applications and efficient use of available data resources.", + "bbox": [ + 212, + 176, + 787, + 328 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 354, + 387, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Human Motion Generation", + "text_level": 1, + "bbox": [ + 214, + 388, + 486, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Human motion generation is the process of synthesizing human motion either unconditionally or conditioned by signals such as text, audio, or action labels. Early works [7, 16, 18, 24] treated this as a deterministic mapping problem, generating a single motion from a specific signal using neural networks. However, human motion is inherently stochastic, even under certain conditions, leading to the adoption of deep generative models in more recent research.", + "bbox": [ + 212, + 415, + 782, + 505 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For instance, Dancing2music [14] employed GANs to generate motion under corresponding conditions. ACTOR [25] introduced a framework based on transformers [44] and VAEs, which, although designed for action-to-motion tasks, can be easily adapted for text-to-motion tasks as demonstrated in TEMOS [26]. Since text and audio are time-series data, natural language processing approaches are commonly used. Works by [6], [1], and [8] utilized GRU-based language models to process motion data along the time axis.", + "bbox": [ + 212, + 506, + 784, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MotionCLIP [39] uses the shared text-image space learned by CLIP [28] to align the feature space of human motion with that of CLIP. MotionGPT [35,36] directly treats motion as language and addresses the motion generation task as a translation problem. However, conditions like language and human motion differ significantly in terms of distribution and expression, making accurate alignment challenging.", + "bbox": [ + 212, + 612, + 784, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To overcome this issue, T2M-GPT [52] and TM2T [9] encode motion using VQ-VAE [43] and generate motion embeddings with generative pretrained transformers. MotionDiffuse [53] is the first application of diffusion models in text-to-motion tasks. MDM [40] employs a simple diffusion framework to diffuse raw motion data, while MLD [4] encodes motion using a VAE model and diffuses it in the latent space. ReMoDiffuse [54] retrieves the motion related to the text to assist in motion generation. Meanwhile, Fg-T2M [45] utilizes a fine-grained method to extract neighborhood and overall semantic linguistic features. Although these methods attain success, they depend exclusively on 3D", + "bbox": [ + 212, + 703, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Realistic Human Motion Generation with Cross-Diffusion Models", + "bbox": [ + 295, + 114, + 732, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/86acb77782775995898ad7825a170fffd7001afd59889499e67f8f4d9acc0857.jpg", + "image_caption": [ + "Fig. 2: Overview of our CrossDiff framework for generating human motion from textual descriptions. The framework incorporates both 3D and 2D motion data, using unified encoding and cross-decoding components to process mixed representations obtained from random projection." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 782, + 296 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "motion data during training, which results in a failure to capture sufficient complexities associated with human motion. In contrast, our approach utilizes a cross-diffusion mechanism to leverage the complementary information found in both 2D and 3D motion representations.", + "bbox": [ + 212, + 406, + 787, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Diffusion Models", + "text_level": 1, + "bbox": [ + 215, + 489, + 401, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion generative models [11, 35, 36], based on stochastic diffusion processes in Thermodynamics, involve a forward process where samples from the data distribution are progressively noised towards a Gaussian distribution and a reverse process where the model learns to denoise Gaussian samples. These models have achieved success in various domains, including image synthesis [29, 31, 32, 34, 42], video generation [11, 21, 49], adversarial attacks [23, 55], motion prediction [3, 46], music-to-dance synthesis [17, 41], and text-to-motion generation [4, 30, 40, 51, 53].", + "bbox": [ + 212, + 515, + 787, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Classifier-Free Guidance [12] enables conditioned generation without additional model training, balancing fidelity and diversity. In the image domain, inpainting methods [5,20,37] iteratively incorporate known information into the diffusion process, maintaining constant image parts while learning to inpaint others. Similar approaches have been applied to motion editing [13,33,40]. Differing from these works, our work focuses on leveraging multiple data representations to enhance diffusion model learning.", + "bbox": [ + 212, + 621, + 787, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 750, + 330, + 765 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Overview", + "text_level": 1, + "bbox": [ + 215, + 782, + 339, + 796 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a textual description $c$ , our objective is to generate multiple human motion sequences $x^{1:N} = \\{x^i\\}_{i=1}^N$ , each with a length of $N$ . As illustrated in Figure 2, our", + "bbox": [ + 212, + 806, + 785, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z. Ren et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "method is carefully designed to efficiently incorporate both 3D and 2D motion data within the learning process of the diffusion model.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the training phase, we first obtain mixed representations of the data from the provided 3D input using a random projection technique. Afterward, the 2D and 3D data representations are independently diffused and processed through our learning framework, CrossDiff, which primarily consists of unified encoding and cross-decoding components.", + "bbox": [ + 212, + 176, + 784, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The unified encoding module maps both the 2D and 3D data into a shared feature space. These features are then passed through the cross-decoding component, resulting in the generation of two motion representations. These representations are subsequently employed for loss calculation and model learning. In the inference phase, our approach supports not only standard sampling but also mixture sampling.", + "bbox": [ + 212, + 252, + 784, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Preliminary. Denoising diffusion probabilistic models (DDPM) [11] can iteratively eliminate noise from a gaussian distribution to approximate a true data distribution. This technique has had a significant impact on the field of generative research, including text-to-motion applications. In this study, we have adapted DDPM and trained a transformer-based model to gradually reduce noise and generate motion sequences.", + "bbox": [ + 212, + 343, + 784, + 433 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Diffusion is modeled as a Markov noising process $\\{x_{t}^{1:N}\\}_{t=0}^{T}$ of $T$ steps. For simplicity, we use $x_{t}$ to denote $x_{t}^{1:N}$ in the following discussion. Starting with a motion sequence $x_{0}$ in original data distribution, the noising process can be described as", + "bbox": [ + 212, + 433, + 784, + 492 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(x _ {t} \\mid x _ {t - 1}\\right) = \\mathcal {N} \\left(x _ {t}; \\sqrt {\\alpha_ {t}} x _ {t - 1}, (1 - \\alpha_ {t}) \\mathbf {I}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 494, + 784, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha_{t} \\in (0,1)$ is constant hyper-parameters. When $\\alpha_{T}$ is small enough, we can approximate $x_{T} \\in \\mathcal{N}(0,\\mathbf{I})$ . The reverse process is to progressively denoise $x_{T}$ from a gaussian distribution to obtain the clean motion $x_0$ . Following [29,40], we predict the clean motion $x_0$ itself on textual condition $c$ as $\\hat{x}_0 = G(x_t,t,c)$ . We apply the simple objective [11]", + "bbox": [ + 212, + 517, + 784, + 594 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s i m p l e}} = \\mathbb {E} _ {t \\sim [ 1, T ]} \\| x _ {0} - G (x _ {t}, t, c) \\| _ {2} ^ {2}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 607, + 784, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Mixed Representations", + "text_level": 1, + "bbox": [ + 214, + 646, + 452, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As the naive diffusion model is trained only on one data distribution (3D poses), we have trained it on a mixture representation of 3D and 2D poses. To obtain 2D data that is closer to the real distribution, we randomly projected the 3D poses into 2D planes in four directions (front, left, right, and back). The 3D poses $x_{3D}^{i} \\in \\mathbb{R}^{d_{3D}}$ and 2D poses $x_{2D}^{i} \\in \\mathbb{R}^{d_{2D}}$ are represented respectively by $d_{3D}$ -dimensional and $d_{2D}$ -dimensional redundant features, respectively, as suggested by [8]. The pose $x^{i}$ is defined by a tuple of $(r,j^{p},j^{v},j^{r},c^{f})$ , where $(r_{3D},j_{3D}^{p},j_{3D}^{v},j_{3D}^{r},c_{3D}^{f})$ is identical to [8]. In addition, $r_{2D} \\in \\mathbb{R}^2$ represents 2D root velocity. $j_{2D}^{p} \\in \\mathbb{R}^{2j},j_{2D}^{v} \\in \\mathbb{R}^{2j}$ and $j_{2D}^{r} \\in \\mathbb{R}^{2j}$ represent the local joints positions, velocities and rotations, respectively, with $j$ denoting the number of joints besides the root. $c_{2D}^{f} \\in \\mathbb{R}^{4}$ is a set of binary features obtained by thresholding the heel and toe", + "bbox": [ + 212, + 672, + 785, + 843 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Realistic Human Motion Generation with Cross-Diffusion Models", + "bbox": [ + 295, + 114, + 732, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "joint velocities. Notably, the rotation representation is made up of the sine and cosine values of the angle.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Cross Motion Diffusion Model", + "text_level": 1, + "bbox": [ + 214, + 196, + 511, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "**Framework.** Our pipeline is illustrated in Figure 2. CLIP [28] is a widely recognized text encoder, and we use it to encode the text prompt $c$ . The encoded text feature and time-step $t$ are projected into transformer dimension and summed together as the condition token $z_{tk}$ . The 2D and 3D motion sequences are projected into the same dimension, concatenated with condition token $z_{tk}$ in time axis and summed with a standard positional embedding. We aim to unify the two domain features in one space but it is too difficult for one linear layer. A straightforward idea is to encode via two separate encoders:", + "bbox": [ + 212, + 220, + 784, + 340 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nz _ {3 D} ^ {0} = \\mathcal {E} _ {3 D} \\left(x _ {3 D, t}, t, c\\right), z _ {2 D} ^ {0} = \\mathcal {E} _ {2 D} \\left(x _ {2 D, t}, t, c\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 349, + 785, + 368 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{E}_{3D / 2D}(\\cdot)$ are 3D/2D $L_{1}$ -layer transformer encoders [44]. However, we find it more efficient to add another shared-weight encoder to extract shared feature:", + "bbox": [ + 212, + 376, + 782, + 419 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{z _ {3 D / 2 D} ^ {i} \\right\\} _ {i = 1} ^ {L _ {2}} = \\mathcal {E} _ {\\text {s h a r e}} \\left(z _ {3 D / 2 D} ^ {0}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 419, + 784, + 440 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{E}_{\\text{share}}(\\cdot)$ is a shared-weight $L_{2}$ -layer transformer encoder, and $\\{z_{3D/2D}^{i}\\}_{i=1}^{L_{2}}$ are the outputs of each shared-weight layer. The whole process is defined as unified encoding.", + "bbox": [ + 212, + 446, + 782, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To output motion in two modality, we use independent $L_{2}$ -layer transformer decoders [44] for 2D and 3D data. Starting from 2D/3D learnable token embeddings $Tok_{2D/3D}$ , each decoder layer takes the output of the previous layer as queries and the output of same-level shared layer $z^{i}$ as keys and values instead of the last layer. The starting point is to make decoder layers follow the extracting patterns of shared-weight layers rather than gaining deeper embeddings. Finally, a linear layer is added to map the features to the dimensions of the motions. This cross-decoding can be integrated as:", + "bbox": [ + 212, + 496, + 784, + 617 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} _ {3 D, 0} = \\mathcal {D} _ {3 D} \\left(\\left\\{z _ {3 D / 2 D} ^ {i} \\right\\} _ {i = 1} ^ {L _ {2}}\\right), \\hat {x} _ {2 D, 0} = \\mathcal {D} _ {2 D} \\left(\\left\\{z _ {3 D / 2 D} ^ {i} \\right\\} _ {i = 1} ^ {L _ {2}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 303, + 626, + 784, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{D}_{3D / 2D}$ are 3D/2D decoders including learnable tokens; and $\\hat{x}_{3D / 2D,0}$ are predicted clean 3D/2D motion sequences. In summary, with a pair of 3D motion and 2D projected motion, CrossDiff outputs four results via", + "bbox": [ + 212, + 652, + 782, + 696 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} _ {i D \\rightarrow j D, 0} = G _ {i D \\rightarrow j D} \\left(x _ {i D, t}, t, c\\right) = \\mathcal {D} _ {j D} \\left(\\mathcal {E} _ {\\text {s h a r e}} \\left(\\mathcal {E} _ {i D} \\left(x _ {i D, t}, t, c\\right)\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 277, + 708, + 784, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{x}_{iD\\to jD,0}$ are predicted $j$ -dimension clean motion $\\hat{x}_{jD,0}$ from $i$ -dimension motion noise $x_{iD,t}$ with $i,j\\in \\{2,3\\}$ .", + "bbox": [ + 212, + 733, + 782, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training. As mentioned in Section 3.1, we apply a simple objective (Eq. 2) for all outputs:", + "bbox": [ + 212, + 784, + 782, + 815 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i D \\rightarrow j D} = \\mathbb {E} _ {t \\sim [ 1, T ]} \\| x _ {j D, 0} - G _ {i D \\rightarrow j D} (x _ {i D, t}, t, c) \\| _ {2} ^ {2}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 823, + 784, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z. Ren et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train our model in two stage. In stage I, CrossDiff is forced to learn the reverse process, motion characteristic of texts in both domains and the motion connection of two distribution via the loss", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s t a g e I}} = \\mathcal {L} _ {3 D \\rightarrow 3 D} + w _ {2 3} \\mathcal {L} _ {2 D \\rightarrow 3 D} + w _ {3 2} \\mathcal {L} _ {3 D \\rightarrow 2 D} + w _ {2 2} \\mathcal {L} _ {2 D \\rightarrow 2 D}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 269, + 204, + 785, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $w_{23}, w_{32}, w_{22}$ are relative weights. In stage II, there is only a 3D generation loss:", + "bbox": [ + 212, + 232, + 782, + 258 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s t a g e I I}} = \\mathcal {L} _ {3 D \\rightarrow 3 D}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 262, + 785, + 279 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This helps the model focus on the 3D denoising process and eliminate the uncertainty of the 2D and 3D mapping relationship while retaining the knowledge of diverse motion features.", + "bbox": [ + 212, + 285, + 782, + 329 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Mixture Sampling", + "text_level": 1, + "bbox": [ + 215, + 352, + 413, + 368 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c881a86b105e53ff8bfc2fd03eaa5271b7b38f105f1c4c97cd08023400335077.jpg", + "image_caption": [ + "Fig. 3: Overview of Mixture Sampling. The original noise is sampled from a 2D gaussian distribution. From time-step $T$ to $\\alpha$ , CrossDiff predicts the clean 2D motion $\\hat{x}_{2D,0}$ and diffuses it back to $x_{2D,t-1}$ . In the remaining $\\alpha$ steps, CrossDiff denoises in the 3D domain and finally obtains the clean 3D motion." + ], + "image_footnote": [], + "bbox": [ + 218, + 404, + 779, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After training, one can sample a motion sequence conditioned on a text prompt in an iterative manner. The standard method [40] gradually anneals the 3D noise from a gaussian distribution, which we still use. We predict the clean sample $\\hat{x}_{3D,0}$ and noise it back to $x_{3D,t-1}$ for $T$ steps until $x_{3D,0}$ .", + "bbox": [ + 212, + 580, + 782, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, utilizing the CrossDiff architecture, we propose a novel two-domain sampling approach. As shown in Figure 3, We first sample 2D gaussian noise which is then denoised with $G_{2D\\to 2D}(x_{2D,t},t,c)$ until time-step $\\alpha$ . Next, we project the denoised 2D noise onto the 3D domain using $G_{2D\\to 3D}(x_{2D,t},t,c)$ and continue the denoising process with $G_{3D\\to 3D}(x_{3D,t},t,c)$ for the remaining $\\alpha$ steps. Our experiments in supplementary demonstrate the difference between mixture sampling and the vanilla method.", + "bbox": [ + 212, + 642, + 784, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.5 Learning 3D Motion Generation from 2D Data", + "text_level": 1, + "bbox": [ + 212, + 768, + 648, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given the complexity and cost associated with collecting high-quality 3D motion data, generating 3D motion from 2D motion data is an attractive alternative. Moreover, generating 3D motion from textual descriptions in an out-of-domain", + "bbox": [ + 212, + 794, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Realistic Human Motion Generation with Cross-Diffusion Models", + "bbox": [ + 295, + 114, + 732, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "scenario is approximately a zero-shot task. To achieve this, we first estimated 2D motion from text-related videos using an off-the-shelf model. We then utilized the pretrained model in stage I $G_{2D \\to 3D}(x_{2D,t}, t, c)$ to generate corresponding 3D clean motion, with $t$ set to 0 and $c$ set to null condition $\\emptyset$ . A motion filter is applied to smooth the generated motion. We assume that 2D pose estimation is relatively precise, allowing the processed 2D/3D motion to serve as pseudolabels for training. The model is fine-tuned with the same objective as stage I, but with different weight hyper-parameters. After training, our model can generate diverse motion according to out-of-domain textual descriptions using mixture sampling (Sec. 3.4).", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During training with 2D motion estimated from videos, we encounter significant errors in root estimation due to the uncertainty of camera movement. To address this issue, we decouple the root information $r_{3D/2D}$ and generate it based on other pose features. Please refer to supplementary for more details.", + "bbox": [ + 212, + 297, + 787, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 386, + 375, + 404 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our focus is on the text-to-motion generation task, and we conduct experiments on two standard datasets: HumanML3D [8] and KIT Motion-Language (KIT-ML) [27]. In Section 4.1, we introduce these standard datasets and the evaluation metrics used in our experiments. In Section 4.2, we present comparable quantitative results with state-of-the-art methods and outperform diffused methods on the HumanML3D dataset. We also use upper and lower body indices, combined with visualization effects, to illustrate our superior performance in capturing whole-body details. Moreover, we demonstrate that CrossDiff supports training with additional 2D motion sequences in Section 4.3. Finally, we discuss ablation studies in Section 4.4 to analyze the contributions of each component of our proposed method.", + "bbox": [ + 212, + 421, + 787, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 215, + 614, + 532, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets. The HumanML3D [8] and KIT-ML datasets [27] are widely used in research. KIT-ML provides 6,353 textual descriptions for 3,911 motion sequences, which are all down-sampled to 12.5 FPS. HumanML3D is currently the largest 3D human motion dataset with textual descriptions. The motions are originally from two motion capture datasets, AMASS [22] and HumanAct12 [10], and are rescaled to 20 frames per second. It contains 14,616 motions annotated with 44,970 sequence-level textual descriptions. To enable fair comparisons with previous works [4,8,40], we use the redundant motion representation proposed by [8]. This representation involves re-targeting the joint positions to a default human skeleton template, setting the initial pose at the same position $(\\mathrm{X} = 0,\\mathrm{Z} = 0)$ facing the $\\mathbf{Z}+$ direction, and including root global positions, local positions, joint rotations, joint velocities, and foot contact labels. We use the same motion representation for KIT-ML.", + "bbox": [ + 212, + 642, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z. Ren et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To obtain corresponding 2D motion, we utilize orthogonal projection to project 3D motion into 2D planes for four directions (front, left, right, and back). Then, we process the redundant 2D motion representation as described in Section 3.2.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "UFC101 [38] is an influential action recognition dataset that contains paired videos and action labels. We obtain 2D joint positions using the ViTPose [48] model and process them into redundant 2D motion representation. We filter out fuzzy data and ultimately select 24 action labels, with each label related to 10-50 motion sequences. We then annotate around five textual descriptions for each label.", + "bbox": [ + 212, + 205, + 784, + 295 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation Metrics. We compare our results with previous works using the same metrics as [8, 40]. These metrics involve evaluating quality with Frechet Inception Distance (FID), precision with R-Precision and Multi-modal Distance (MM Dist), and diversity with Diversity (DIV) and Multimodality (MModality). These measurements assess generated motion distribution, retrieval ranking accuracy, Euclidean distances between text and motion features, and variance across features and within a single text.", + "bbox": [ + 212, + 318, + 784, + 424 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Besides using the evaluation model from [8], we introduce a new metric measuring the FID (Fréchet Inception Distance) of upper and lower body movements, denoted as FID-U and FID-L. This enables a fine-grained analysis of human motion and better comprehension of upper and lower body dynamics. We split joints into two groups using the root joint as a boundary and train separate evaluators, following a similar approach to [8]. This effectively evaluates generated motion quality for both body segments, offering a deeper understanding of human motion complexities and advancing research on new motion generation models.", + "bbox": [ + 212, + 425, + 785, + 559 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details. We use the AdamW [19] optimizer, setting the learning rate to 1e-4 and 1e-5 in stages I and II, respectively. Our model is trained for 4k epochs in stage I and 1k epochs in stage II, using 8 Tesla V100 GPUs with a mini-batch size of 32. The number of diffusion steps is set to 1K. The loss weights $(w_{23}, w_{32}, w_{33})$ is set to $(1, 1, 1)$ in stage I and $(0.1, 0.1, 1)$ in training on UFC101 datasets.", + "bbox": [ + 212, + 582, + 784, + 672 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Comparisons on Text-to-motion", + "text_level": 1, + "bbox": [ + 214, + 695, + 522, + 709 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Comparative Analysis of Standard Metrics. In our evaluation, we test our models 20 times and compare their performance with existing state-of-the-art methods. These methods include Language2Pose [1], T2G [2], Hier [6], T2M [8], MotionDiffuse [53], Fg-T2M [45], MDM [40], T2M-GPT [52], MLD [4] and ReMoDiffuse [54]. As illustrated in Table 1, our model exhibits competitive performance when compared to these leading methods. However, it is important to note that the KIT-ML dataset primarily consists of \"walk\" movements and lacks intricate details. Consequently, this dataset does not present the same challenges", + "bbox": [ + 212, + 719, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Realistic Human Motion Generation with Cross-Diffusion Models", + "bbox": [ + 295, + 114, + 732, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/6864ab268ad4ca97c37472495468440de21e4c3191692a607a12e773c7c9db9d.jpg", + "table_caption": [ + "Table 1: Quantitative results on the HumanML3D and KIT-ML test set. The overall results on KIT-ML are shown on the right, while the results of both widely-used and newly-proposed metrics on HumanML3D are shown on the left. The red and blue colors indicate the best and second-best results, respectively." + ], + "table_footnote": [], + "table_body": "
MethodsHumanML3DKIT-ML
R Precision (top 3)↑FID↓MM Dist↓DIV→ MModality↑FID-U↓FID-L↓R Precision (top 3)↑FID↓MM Dist↓DIV→ MModality↑FID-U↓FID-L↓
Real0.7970.0022.9749.503--0.7790.0312.78811.08--
Language2Pose0.48611.025.2967.676--0.4836.5455.1479.073--
T2G0.3457.6646.0306.409--0.33812.126.9649.334--
Hier0.5526.5325.0128.332--0.5315.2034.9869.563--
T2M0.7401.0673.3409.1882.090-0.6932.7703.40110.911.482-
MotionDiffuse0.7820.6303.1139.4101.553-0.7391.9542.95811.100.730-
Fg-T2M0.7830.2433.1099.2781.614-0.7450.5713.11410.931.019-
MDM0.6110.5445.5669.5592.7990.8250.4840.4979.19110.8471.9070.925
T2M-GPT0.7750.1413.1219.7221.8310.1450.6070.7450.5143.0071.5700.602
MLD0.7720.4733.1969.7242.4130.5410.5530.7340.4043.2042.1920.563
ReMoDiffuse0.7950.1032.9749.0181.7950.1250.5650.7650.1552.8141.8080.205
Ours0.7300.1623.3589.5772.6200.1180.2810.7040.4743.30810.770.434
", + "bbox": [ + 217, + 224, + 787, + 348 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "that our method is specifically designed to address. It indicates we can steadily generate high-quality and precise motion while pay attention to rich diversity. In other words, our generated results are not only consistent with the textual description, but also more expressive.", + "bbox": [ + 212, + 385, + 787, + 446 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Comparative Analysis of Fine-grained Metrics. We compare the fine-grained metrics for our upper and lower body with those from four recent studies [4, 40, 52, 54]. As demonstrated in Table 1, our generated motion is more robust and detailed. Our low FID scores for both the upper and lower body indicate that our synthesized motion effectively captures full-body movement rather than focusing solely on specific semantic parts. In contrast, ReMoDiffuse and T2M-GPT achieves a low FID score for the upper body but a high score for the lower body. This suggests that their generation process exhibits unbalanced attention towards different body parts, primarily translating textual descriptions into upper body characteristics rather than capturing the entire body's motion.", + "bbox": [ + 212, + 465, + 787, + 616 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Figure 4 displays qualitative comparisons with existing methods. Our method can \"march\" with arm swings, \"wobble\" using hands for balancing and alternate between defense and attack in a \"fight\". MDM exhibits a certain rigidity, while T2M-GPT and ReMoDiffuse appear to lack dynamism. For direct comparisons, please refer to the supplementary videos provided. We conducted a user study on motion performance, in which participants were asked two questions to assess the vitality and diversity of the motions. The results, presented in Figure 5(a), confirm our analysis. In summary, our method demonstrates a superior ability to interpret semantic information and generate more accurate and expressive motions.", + "bbox": [ + 212, + 617, + 787, + 767 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Learning from 2D Data", + "text_level": 1, + "bbox": [ + 214, + 787, + 457, + 801 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "After being trained on a 3D dataset, our model can learn 3D motion generation from 2D data. By fine-tuning the model with the UCF101 dataset [38], we ef", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z. Ren et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/a8f5cabae1da8913966082f8e909fe530cc00beab367e9af967ff3a758e0a8fb.jpg", + "image_caption": [ + "Fig. 4: Qualitative results on HumanML3D dataset. We compare our method with MDM [40], T2M-GPT [52] and MLD [4]. We find that our generated actions better convey the intended semantics." + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 782, + 444 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "fectively address a zero-shot problem arising from the absence of ground-truth 3D motion. Our sampling strategy reaches optimal performance when $\\alpha = 1$ . As depicted in Figure 6, the generated motions for various activities, such as pulling up, biking, table tennis, and baseball, are showcased alongside their textual prompts. Notably, pulling up and biking are entirely outside the HumanML3D motion domain. Although playing table tennis and baseball share similarities with actions like slapping or throwing in HumanML3D, the original model is unable to synthesize motion under those specific descriptions. However, after fine-tuning, our generated motion more closely resembles the given sports rather than merely mimicking the actions. Despite some activities being beyond the scope of the HumanML3D domain, our fine-tuned model successfully synthesizes specific motions by leveraging the weak 2D data. This demonstrates its remarkable adaptability and potential for efficient use of available motion data since 2D motion and related textual descriptions are easier to obtain than 3D motion.", + "bbox": [ + 212, + 539, + 787, + 767 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 786, + 401, + 800 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Our model features separate pipelines for 2D and 3D inputs, allowing us to train solely on 3D motion sequences, which is an improvement over MDM [40].", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Realistic Human Motion Generation with Cross-Diffusion Models", + "bbox": [ + 295, + 114, + 732, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/917d6d57989f0277642179d6280abda59d4250989c1a571a9700021312078382.jpg", + "image_caption": [ + "Fig. 5: (a) The result of the user study. (b) Difference between 3D and 2D motion data distribution. The time axis is represented on the x-axis, while the normalized joint velocity is represented on the y-axis. The 3D motion is represented by a blue full line, while the 2D motion is represented by red and green dashed lines, indicating the front and left view, respectively." + ], + "image_footnote": [], + "bbox": [ + 217, + 148, + 406, + 262 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/65ff17a8c208517b3e98ced8ff77c8ea391c8c655c106f9d9891d633d4b7ae30.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 447, + 142, + 787, + 263 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b7b6c6c3cba1bacdeb542dfbcdb1adfceabd06e88f8f1dbd76282b3a486a8ef6.jpg", + "image_caption": [ + "Fig. 6: Generating 3D movements without training on paired 3D motion and textual descriptions." + ], + "image_footnote": [], + "bbox": [ + 217, + 366, + 787, + 526 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We investigate the impact of introducing 2D data on the performance of 3D generation and demonstrate the effectiveness of using a shared-weights encoder.", + "bbox": [ + 212, + 607, + 787, + 640 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Why 2D motion help? To explain the benefits of 2D motions, we compared the distribution differences between 3D and 2D motion data. Hand and feet movements, which are primary indicators of motion, were visualized in both 3D and 2D levels, and their velocities were normalized along the joints dimension. In Figure 5(b), we can clearly see that around the 20th frame, the 2D velocity of the left hand in the front view reaches a higher value while the 3D velocity is quite low, indicating that hand movements in 2D are more prominent than in 3D. The results show that 2D motion captures different details from 3D motion, suggesting that the CrossDiff model can lead 3D motion to learn from the knowledge that 2D motion acquired from text prompts. Specifically, for the given sample, 2D motion might learn \"animated hand motions\" while 3D motion focuses only on \"walking\". 2D motion is an explicit feature that we artificially extract to", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z. Ren et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/4faea49aa7cb6dddfce3225448f9e48c3a39d5292cde7752bacb0c88305d7180.jpg", + "table_caption": [ + "Table 2: Evaluation of our models with different settings on the HumanML3D dataset. Bold indicates best result. The symbol % indicates the percentage of data being used. From top to bottom, we present MDM as baselines, the impact of training with 2D representations, with(w/) or without(w/o) shared-weights encoder." + ], + "table_footnote": [], + "table_body": "
MethodsR Precision (top 3)↑FID↓MM Dist↓DIV→
MDM0.6110.5445.5669.559
Ours0.7300.1623.3589.577
50% 3D0.6660.5863.8949.513
100% 3D0.6850.2243.6909.445
50% 3D + 100% 2D0.6720.4223.7089.345
100% 3D + 100% 2D0.7300.1623.3589.577
w/o shared-weights encoder0.7140.1873.4969.488
w/ shared-weights encoder0.7300.1623.3589.577
1 view(front)0.7220.1863.4679.798
1 view(left)0.7150.1813.4129.834
4 views0.7300.1623.3589.577
5 views0.6950.2023.6139.502
", + "bbox": [ + 282, + 224, + 720, + 429 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "aid the model's learning, and this approach can help improve performance when dealing with arbitrary data.", + "bbox": [ + 212, + 470, + 784, + 502 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Influence of 2D Representation Table 2 presents the results from four different experiment settings. The control groups of \"100% 3D\" and \"100% 3D + 100% 2D\" demonstrate that when training with paired 3D motion and text, projecting the 3D motion to 2D and building a connection between the 2D motion and text can help boost performance. The visualizations in Figure 1 further highlight the enhanced quality of our generated outputs. The control groups of \"50% 3D\" and \"50% 3D + 100% 2D\" prove that additional 2D data can also help improve performance. The additional 2D data indicates other 2D motion without ground truth 3D motion. The experiment in Section 4.3 shows learning 2D motion in the wild can also help with out-of-domain 3D motion learning. As we can see, the combined learning of 2D motion has great potential.", + "bbox": [ + 212, + 527, + 787, + 694 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Shared-weights Encoder Without the shared-weights encoder, the model is a simple encoder-decoder framework with two modalities. However, we believe that this is not sufficient to fully fuse the 3D and 2D motion features. Inspired by [47], we found that when learning with data from two modalities, extracting separate and fused feature layer-by-layer is more efficient. The shared-weights encoder serves as a fusing module, while the 3D/2D motion decoder acts as a separate decoder module. The goal is to ensure that the decoder layers follow the same extraction patterns as the shared-weight layers, rather than simply gaining", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Realistic Human Motion Generation with Cross-Diffusion Models", + "bbox": [ + 295, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "deeper embeddings. The results presented in Table 2 demonstrate the efficiency of using a shared-weights encoder.", + "bbox": [ + 212, + 146, + 784, + 176 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "2D Representation from different views To evaluate the impact of different views, we conducted tests on four settings: a) only the front view, b) only the left view, c) four views including front, left, back, and right, and d) five views with an additional top view. When multiple views were used, the 3D motion was paired with a random view projection to formulate the loss. As shown in Table 2, it is reasonable to assume that four views contain more 2D information and outperform one view. Furthermore, the front view and left view do not have much distinction in terms of performance. However, the performance actually decreased with the additional top view. This could be due to the fact that the 2D information becomes more difficult to classify without the condition of the camera view. We did not pass along the camera information because it is difficult to estimate the camera view of in-the-wild videos, and injecting camera information could disrupt the structure of the text-to-2D model, making it difficult for the text-to-3D model to follow. In conclusion, the number of views serves as a practical hyperparameter that can be adjusted through enumeration experiments.", + "bbox": [ + 212, + 196, + 787, + 439 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 458, + 359, + 474 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In conclusion, the Cross Human Motion Diffusion Model (CrossDiff) presents a promising advancement in the field of human motion synthesis by effectively integrating and leveraging both 3D and 2D motion information for high-quality motion generation. The unified encoding and cross-decoding components of the CrossDiff learning framework enable the capture of intricate details of human movement that are often overlooked by models relying solely on 3D data. Our experiments validate the superior performance of the proposed model on various text-to-motion benchmarks.", + "bbox": [ + 212, + 487, + 787, + 607 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Despite its promising results, the CrossDiff model is not without potential weaknesses. One limitation may arise from the model's ability to generalize to unseen or rare motion patterns, especially when trained only on 2D motion data. Additionally, the computational complexity of the model might hinder its real-time application in certain scenarios, such as interactive gaming and virtual reality environments.", + "bbox": [ + 212, + 608, + 787, + 698 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Future work could explore methods to enhance the model's generalization capabilities, such as incorporating unsupervised or semi-supervised learning techniques. To further advance our understanding, we propose the accumulation of a sizable 2D motion dataset coupled with relevant textual prompts, enabling the training of a unified motion generation model.", + "bbox": [ + 212, + 699, + 787, + 776 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments. This research was partly supported by Shenzhen Key Laboratory of next generation interactive media innovative technology (Grant No: ZDSYS20210623092001004).", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z. Ren et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ahuja, C., Morency, L.P.: Language2pose: Natural language grounded pose forecasting. In: 2019 International Conference on 3D Vision (3DV). pp. 719-728. IEEE (2019)", + "2. Bhattacharya, U., Rewkowski, N., Banerjee, A., Guhan, P., Bera, A., Manocha, D.: Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In: 2021 IEEE virtual reality and 3D user interfaces (VR). pp. 1-10. IEEE (2021)", + "3. Chen, L.H., Zhang, J., Li, Y., Pang, Y., Xia, X., Liu, T.: Humanmac: Masked motion completion for human motion prediction. arXiv preprint arXiv:2302.03665 (2023)", + "4. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18000-18010 (2023)", + "5. Chung, H., Sim, B., Ye, J.C.: Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12413-12422 (2022)", + "6. Ghosh, A., Cheema, N., Oguz, C., Theobalt, C., Slusallek, P.: Synthesis of compositional animations from textual descriptions. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1396-1406 (2021)", + "7. Ghosh, P., Song, J., Aksan, E., Hilliges, O.: Learning human motion models for long-term predictions. In: 2017 International Conference on 3D Vision (3DV). pp. 458-466. IEEE (2017)", + "8. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5152-5161 (2022)", + "9. Guo, C., Zuo, X., Wang, S., Cheng, L.: Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In: European Conference on Computer Vision. pp. 580-597. Springer (2022)", + "10. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 2021-2029 (2020)", + "11. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)", + "12. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)", + "13. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Gmd: Controllable human motion synthesis via guided diffusion models. arXiv preprint arXiv:2305.12577 (2023)", + "14. Lee, H.Y., Yang, X., Liu, M.Y., Wang, T.C., Lu, Y.D., Yang, M.H., Kautz, J.: Dancing to music. Advances in neural information processing systems 32 (2019)", + "15. Li, J., Kang, D., Pei, W., Zhe, X., Zhang, Y., He, Z., Bao, L.: Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11293-11302 (2021)", + "16. Li, M., Chen, S., Zhao, Y., Zhang, Y., Wang, Y., Tian, Q.: Dynamic multiscale graph neural networks for 3d skeleton based human motion prediction. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 214-223 (2020)" + ], + "bbox": [ + 218, + 172, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Realistic Human Motion Generation with Cross-Diffusion Models", + "bbox": [ + 295, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Li, R., Zhao, J., Zhang, Y., Su, M., Ren, Z., Zhang, H., Tang, Y., Li, X.: Finedance: A fine-grained choreography dataset for 3d full body dance generation. arXiv preprint arXiv:2212.03741 (2023)", + "18. Li, Z., Zhou, Y., Xiao, S., He, C., Huang, Z., Li, H.: Auto-conditioned recurrent networks for extended complex human motion synthesis. arXiv preprint arXiv:1707.05363 (2017)", + "19. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)", + "20. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11461-11471 (2022)", + "21. Luo, Z., Chen, D., Zhang, Y., Huang, Y., Wang, L., Shen, Y., Zhao, D., Zhou, J., Tan, T.: Videofusion: Decomposed diffusion models for high-quality video generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10209-10218 (2023)", + "22. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5442-5451 (2019)", + "23. Nie, W., Guo, B., Huang, Y., Xiao, C., Vahdat, A., Anandkumar, A.: Diffusion models for adversarial purification. arXiv preprint arXiv:2205.07460 (2022)", + "24. Pavllo, D., Grangier, D., Auli, M.: Quaternet: A quaternion-based recurrent model for human motion. arXiv preprint arXiv:1805.06485 (2018)", + "25. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3d human motion synthesis with transformer vae. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10985-10995 (2021)", + "26. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: European Conference on Computer Vision. pp. 480-497. Springer (2022)", + "27. Plappert, M., Mandery, C., Asfour, T.: The kit motion-language dataset. Big data 4(4), 236-252 (2016)", + "28. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "29. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022)", + "30. Ren, Z., Pan, Z., Zhou, X., Kang, L.: Diffusion motion: Generate text-guided 3d human motion by diffusion model. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023)", + "31. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)", + "32. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)", + "33. Shafir, Y., Tevet, G., Kapon, R., Bermano, A.H.: Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418 (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z. Ren et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Sinha, A., Song, J., Meng, C., Ermon, S.: D2c: Diffusion-decoding models for few-shot conditional generation. Advances in Neural Information Processing Systems 34, 12533-12548 (2021)", + "35. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256-2265. PMLR (2015)", + "36. Song, Y., Ermon, S.: Improved techniques for training score-based generative models. Advances in neural information processing systems 33, 12438-12448 (2020)", + "37. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)", + "38. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012)", + "39. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: European Conference on Computer Vision. pp. 358-374. Springer (2022)", + "40. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. arXiv preprint arXiv:2209.14916 (2022)", + "41. Tseng, J., Castellon, R., Liu, K.: Edge: Editable dance generation from music. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 448-458 (2023)", + "42. Vahdat, A., Kreis, K., Kautz, J.: Score-based generative modeling in latent space. Advances in Neural Information Processing Systems 34, 11287-11302 (2021)", + "43. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems 30 (2017)", + "44. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)", + "45. Wang, Y., Leng, Z., Li, F.W., Wu, S.C., Liang, X.: Fg-t2m: Fine-grained text-driven human motion generation via diffusion model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22035-22044 (2023)", + "46. Wei, D., Sun, H., Li, B., Lu, J., Li, W., Sun, X., Hu, S.: Human joint kinematics diffusion-refinement for stochastic motion prediction. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 6110-6118 (2023)", + "47. Xu, X., Wu, C., Rosenman, S., Lal, V., Che, W., Duan, N.: Bridgetower: Building bridges between encoders in vision-language representation learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 10637-10647 (2023)", + "48. Xu, Y., Zhang, J., Zhang, Q., Tao, D.: Vitpose: Simple vision transformer baselines for human pose estimation. Advances in Neural Information Processing Systems 35, 38571-38584 (2022)", + "49. Yang, R., Srivastava, P., Mandt, S.: Diffusion probabilistic modeling for video generation. arXiv preprint arXiv:2203.09481 (2022)", + "50. Yoon, Y., Cha, B., Lee, J.H., Jang, M., Lee, J., Kim, J., Lee, G.: Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics (TOG) 39(6), 1-16 (2020)", + "51. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: Physdiff: Physics-guided human motion diffusion model. arXiv preprint arXiv:2212.02500 (2022)", + "52. Zhang, J., Zhang, Y., Cun, X., Huang, S., Zhang, Y., Zhao, H., Lu, H., Shen, X.: T2m-gpt: Generating human motion from textual descriptions with discrete representations. arXiv preprint arXiv:2301.06052 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Realistic Human Motion Generation with Cross-Diffusion Models", + "bbox": [ + 295, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "53. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022)", + "54. Zhang, M., Guo, X., Pan, L., Cai, Z., Hong, F., Li, H., Yang, L., Liu, Z.: Remodif- \nfuse: Retrieval-augmented motion diffusion model. arXiv preprint arXiv:2304.01116 (2023)", + "55. Zhuang, H., Zhang, Y., Liu, S.: A pilot study of query-free adversarial attack against stable diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2384-2391 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 272 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Z. Ren et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Realistic Human Motion Generation with Cross-Diffusion Models/5b1805dd-89b3-4df9-9e0a-1ed375e3f44c_model.json b/2024/Realistic Human Motion Generation with Cross-Diffusion Models/5b1805dd-89b3-4df9-9e0a-1ed375e3f44c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..17bca7d2475b85d1b7ef5ad2ab6c3c9b9b7c7141 --- /dev/null +++ b/2024/Realistic Human Motion Generation with Cross-Diffusion Models/5b1805dd-89b3-4df9-9e0a-1ed375e3f44c_model.json @@ -0,0 +1,2326 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.264, + 0.141, + 0.744, + 0.187 + ], + "angle": 0, + "content": "Realistic Human Motion Generation with Cross-Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.34, + 0.212, + 0.664, + 0.229 + ], + "angle": 0, + "content": "Zeping Ren\\(^{1*}\\), Shaoli Huang\\(^{2\\dagger}\\), Xiu Li\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.327, + 0.239, + 0.677, + 0.268 + ], + "angle": 0, + "content": "\\(^{1}\\) Tsinghua Shenzhen International Graduate School \n\\(^{2}\\) Tencent AI Lab" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.304, + 0.744, + 0.596 + ], + "angle": 0, + "content": "Abstract. In this work, we introduce the Cross Human Motion Diffusion Model (CrossDiff\\(^3\\)), a novel approach for generating high-quality human motion based on textual descriptions. Our method integrates 3D and 2D information using a shared transformer network within the training of the diffusion model, unifying motion noise into a single feature space. This enables cross-decoding of features into both 3D and 2D motion representations, regardless of their original dimension. The primary advantage of CrossDiff is its cross-diffusion mechanism, which allows the model to reverse either 2D or 3D noise into clean motion during training. This capability leverages the complementary information in both motion representations, capturing intricate human movement details often missed by models relying solely on 3D information. Consequently, CrossDiff effectively combines the strengths of both representations to generate more realistic motion sequences. In our experiments, our model demonstrates competitive state-of-the-art performance on text-to-motion benchmarks. Moreover, our method consistently provides enhanced motion generation quality, capturing complex full-body movement intricacies. Additionally, with a pre-trained model, our approach accommodates using in-the-wild 2D motion data without 3D motion ground truth during training to generate 3D motion, highlighting its potential for broader applications and efficient use of available data resources." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.621, + 0.377, + 0.637 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.652, + 0.788, + 0.711 + ], + "angle": 0, + "content": "In recent years, the field of human motion synthesis [7,9,10,14,15,18,41,50] has witnessed significant advancements, primarily driven by the growing demand for high-quality, realistic motion generation in applications such as gaming, virtual reality, and robotics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.713, + 0.788, + 0.789 + ], + "angle": 0, + "content": "A crucial aspect in this research area is generating human motion based on textual descriptions, enabling contextually accurate and natural movements [40]. However, current methods [4,8,26,40] predominantly rely on 3D motion information during training, leading to an inability to capture the full spectrum of intricacies associated with human motion. When using only 3D representation," + }, + { + "type": "page_footnote", + "bbox": [ + 0.203, + 0.797, + 0.564, + 0.812 + ], + "angle": 0, + "content": "* Work done during an internship at Tencent AI Lab." + }, + { + "type": "page_footnote", + "bbox": [ + 0.205, + 0.812, + 0.37, + 0.826 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.205, + 0.826, + 0.536, + 0.84 + ], + "angle": 0, + "content": "3 https://wonderno.github.io/CrossDiff-webpage/" + }, + { + "type": "list", + "bbox": [ + 0.203, + 0.797, + 0.564, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "Z. Ren et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.788, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.281, + 0.785, + 0.31 + ], + "angle": 0, + "content": "Fig. 1: Our method utilizing the cross-diffusion mechanism (Left) exhibits more full-body details compared to existing methods (Right)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.788, + 0.46 + ], + "angle": 0, + "content": "the generation model may struggle to relate text semantics to some body part movements with very small movement variations compared to others, which can lead to overlooking important motion details. This is because the model might focus on more dominant or larger movements within the 3D space, leaving subtle nuances underrepresented. For example, when given a prompt such as \"a person is dancing eloquently,\" as illustrated in Figure 1, the generated motion might lack vitality, display a limited range of movements, and contain minimal local motion details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.627 + ], + "angle": 0, + "content": "To effectively address the limitations and accurately capture the nuances of full-body movement, we introduce the Cross Human Motion Diffusion Model (CrossDiff). This innovative approach seamlessly integrates and leverages both 3D and 2D motion information to generate high-quality human motion sequences. The 2D data representation effectively illustrates the intricacies of human body movements from various viewing angle projections. Due to different view projections in 2D data, small body part movements can be magnified in certain projections, making them more noticeable and easier to capture. This helps the text-to-motion generation models to better associate text descriptions with a wider range of human body motion details, including subtle movements that might have been overlooked when relying solely on 3D representation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.734 + ], + "angle": 0, + "content": "As a result, incorporating 2D information with 3D enables the diffusion model to establish more connections between motion and text prompts, ultimately enhancing the motion synthesis process. The CrossDiff learning framework consists of two main components: unified encoding and cross-decoding. These components work together to achieve more precise and realistic motion synthesis. Furthermore, it is essential to transfer the knowledge acquired in the 2D domain to 3D motion, which leads to an overall improvement in the model's performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Unified encoding fuses motion noise from both 3D and 2D sources into a single feature space, facilitating cross-decoding of features into either 3D or 2D motion representations, regardless of their original dimension. The distinctive innovation of our approach stems from the cross-diffusion mechanism, which enables the model to transform 2D or 3D noise into clean motion during the training process. This capability allows the model to harness the complementary information present in both motion representations, effectively capturing" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.297, + 0.115, + 0.733, + 0.128 + ], + "angle": 0, + "content": "Realistic Human Motion Generation with Cross-Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "intricate details of human movement that are often missed by models relying exclusively on 3D data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.329 + ], + "angle": 0, + "content": "In experiments, we demonstrate our model achieves competitive state-of-the-art performance on several text-to-motion benchmarks, outperforming existing diffusion-based approaches that rely solely on 3D motion information during training. Furthermore, our method consistently delivers enhanced motion generation quality, capturing complex full-body movement intricacies essential for realistic motion synthesis. A notable advantage of our approach is its ability to utilize 2D motion data without necessitating 3D motion ground truth during training, enabling the generation of 3D motion. This feature underscores the potential of the CrossDiff model for a wide range of applications and efficient use of available data resources." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.355, + 0.388, + 0.37 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.389, + 0.487, + 0.403 + ], + "angle": 0, + "content": "2.1 Human Motion Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.784, + 0.506 + ], + "angle": 0, + "content": "Human motion generation is the process of synthesizing human motion either unconditionally or conditioned by signals such as text, audio, or action labels. Early works [7, 16, 18, 24] treated this as a deterministic mapping problem, generating a single motion from a specific signal using neural networks. However, human motion is inherently stochastic, even under certain conditions, leading to the adoption of deep generative models in more recent research." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.785, + 0.612 + ], + "angle": 0, + "content": "For instance, Dancing2music [14] employed GANs to generate motion under corresponding conditions. ACTOR [25] introduced a framework based on transformers [44] and VAEs, which, although designed for action-to-motion tasks, can be easily adapted for text-to-motion tasks as demonstrated in TEMOS [26]. Since text and audio are time-series data, natural language processing approaches are commonly used. Works by [6], [1], and [8] utilized GRU-based language models to process motion data along the time axis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.785, + 0.703 + ], + "angle": 0, + "content": "MotionCLIP [39] uses the shared text-image space learned by CLIP [28] to align the feature space of human motion with that of CLIP. MotionGPT [35,36] directly treats motion as language and addresses the motion generation task as a translation problem. However, conditions like language and human motion differ significantly in terms of distribution and expression, making accurate alignment challenging." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To overcome this issue, T2M-GPT [52] and TM2T [9] encode motion using VQ-VAE [43] and generate motion embeddings with generative pretrained transformers. MotionDiffuse [53] is the first application of diffusion models in text-to-motion tasks. MDM [40] employs a simple diffusion framework to diffuse raw motion data, while MLD [4] encodes motion using a VAE model and diffuses it in the latent space. ReMoDiffuse [54] retrieves the motion related to the text to assist in motion generation. Meanwhile, Fg-T2M [45] utilizes a fine-grained method to extract neighborhood and overall semantic linguistic features. Although these methods attain success, they depend exclusively on 3D" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "Z. Ren et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.784, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.321, + 0.788, + 0.378 + ], + "angle": 0, + "content": "Fig. 2: Overview of our CrossDiff framework for generating human motion from textual descriptions. The framework incorporates both 3D and 2D motion data, using unified encoding and cross-decoding components to process mixed representations obtained from random projection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.407, + 0.789, + 0.468 + ], + "angle": 0, + "content": "motion data during training, which results in a failure to capture sufficient complexities associated with human motion. In contrast, our approach utilizes a cross-diffusion mechanism to leverage the complementary information found in both 2D and 3D motion representations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.49, + 0.403, + 0.504 + ], + "angle": 0, + "content": "2.2 Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.516, + 0.788, + 0.621 + ], + "angle": 0, + "content": "Diffusion generative models [11, 35, 36], based on stochastic diffusion processes in Thermodynamics, involve a forward process where samples from the data distribution are progressively noised towards a Gaussian distribution and a reverse process where the model learns to denoise Gaussian samples. These models have achieved success in various domains, including image synthesis [29, 31, 32, 34, 42], video generation [11, 21, 49], adversarial attacks [23, 55], motion prediction [3, 46], music-to-dance synthesis [17, 41], and text-to-motion generation [4, 30, 40, 51, 53]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.622, + 0.788, + 0.727 + ], + "angle": 0, + "content": "Classifier-Free Guidance [12] enables conditioned generation without additional model training, balancing fidelity and diversity. In the image domain, inpainting methods [5,20,37] iteratively incorporate known information into the diffusion process, maintaining constant image parts while learning to inpaint others. Similar approaches have been applied to motion editing [13,33,40]. Differing from these works, our work focuses on leveraging multiple data representations to enhance diffusion model learning." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.751, + 0.331, + 0.766 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.783, + 0.341, + 0.797 + ], + "angle": 0, + "content": "3.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.808, + 0.787, + 0.843 + ], + "angle": 0, + "content": "Given a textual description \\( c \\), our objective is to generate multiple human motion sequences \\( x^{1:N} = \\{x^i\\}_{i=1}^N \\), each with a length of \\( N \\). As illustrated in Figure 2, our" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.297, + 0.115, + 0.733, + 0.128 + ], + "angle": 0, + "content": "Realistic Human Motion Generation with Cross-Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "method is carefully designed to efficiently incorporate both 3D and 2D motion data within the learning process of the diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.253 + ], + "angle": 0, + "content": "During the training phase, we first obtain mixed representations of the data from the provided 3D input using a random projection technique. Afterward, the 2D and 3D data representations are independently diffused and processed through our learning framework, CrossDiff, which primarily consists of unified encoding and cross-decoding components." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.785, + 0.344 + ], + "angle": 0, + "content": "The unified encoding module maps both the 2D and 3D data into a shared feature space. These features are then passed through the cross-decoding component, resulting in the generation of two motion representations. These representations are subsequently employed for loss calculation and model learning. In the inference phase, our approach supports not only standard sampling but also mixture sampling." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.785, + 0.434 + ], + "angle": 0, + "content": "Preliminary. Denoising diffusion probabilistic models (DDPM) [11] can iteratively eliminate noise from a gaussian distribution to approximate a true data distribution. This technique has had a significant impact on the field of generative research, including text-to-motion applications. In this study, we have adapted DDPM and trained a transformer-based model to gradually reduce noise and generate motion sequences." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.785, + 0.493 + ], + "angle": 0, + "content": "Diffusion is modeled as a Markov noising process \\(\\{x_{t}^{1:N}\\}_{t=0}^{T}\\) of \\(T\\) steps. For simplicity, we use \\(x_{t}\\) to denote \\(x_{t}^{1:N}\\) in the following discussion. Starting with a motion sequence \\(x_{0}\\) in original data distribution, the noising process can be described as" + }, + { + "type": "equation", + "bbox": [ + 0.363, + 0.496, + 0.785, + 0.512 + ], + "angle": 0, + "content": "\\[\nq \\left(x _ {t} \\mid x _ {t - 1}\\right) = \\mathcal {N} \\left(x _ {t}; \\sqrt {\\alpha_ {t}} x _ {t - 1}, (1 - \\alpha_ {t}) \\mathbf {I}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.785, + 0.595 + ], + "angle": 0, + "content": "where \\(\\alpha_{t} \\in (0,1)\\) is constant hyper-parameters. When \\(\\alpha_{T}\\) is small enough, we can approximate \\(x_{T} \\in \\mathcal{N}(0,\\mathbf{I})\\). The reverse process is to progressively denoise \\(x_{T}\\) from a gaussian distribution to obtain the clean motion \\(x_0\\). Following [29,40], we predict the clean motion \\(x_0\\) itself on textual condition \\(c\\) as \\(\\hat{x}_0 = G(x_t,t,c)\\). We apply the simple objective [11]" + }, + { + "type": "equation", + "bbox": [ + 0.372, + 0.608, + 0.785, + 0.626 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s i m p l e}} = \\mathbb {E} _ {t \\sim [ 1, T ]} \\| x _ {0} - G (x _ {t}, t, c) \\| _ {2} ^ {2}. \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.647, + 0.453, + 0.663 + ], + "angle": 0, + "content": "3.2 Mixed Representations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.787, + 0.844 + ], + "angle": 0, + "content": "As the naive diffusion model is trained only on one data distribution (3D poses), we have trained it on a mixture representation of 3D and 2D poses. To obtain 2D data that is closer to the real distribution, we randomly projected the 3D poses into 2D planes in four directions (front, left, right, and back). The 3D poses \\( x_{3D}^{i} \\in \\mathbb{R}^{d_{3D}} \\) and 2D poses \\( x_{2D}^{i} \\in \\mathbb{R}^{d_{2D}} \\) are represented respectively by \\( d_{3D} \\)-dimensional and \\( d_{2D} \\)-dimensional redundant features, respectively, as suggested by [8]. The pose \\( x^{i} \\) is defined by a tuple of \\( (r,j^{p},j^{v},j^{r},c^{f}) \\), where \\( (r_{3D},j_{3D}^{p},j_{3D}^{v},j_{3D}^{r},c_{3D}^{f}) \\) is identical to [8]. In addition, \\( r_{2D} \\in \\mathbb{R}^2 \\) represents 2D root velocity. \\( j_{2D}^{p} \\in \\mathbb{R}^{2j},j_{2D}^{v} \\in \\mathbb{R}^{2j} \\) and \\( j_{2D}^{r} \\in \\mathbb{R}^{2j} \\) represent the local joints positions, velocities and rotations, respectively, with \\( j \\) denoting the number of joints besides the root. \\( c_{2D}^{f} \\in \\mathbb{R}^{4} \\) is a set of binary features obtained by thresholding the heel and toe" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "Z. Ren et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "joint velocities. Notably, the rotation representation is made up of the sine and cosine values of the angle." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.198, + 0.513, + 0.213 + ], + "angle": 0, + "content": "3.3 Cross Motion Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.221, + 0.785, + 0.342 + ], + "angle": 0, + "content": "**Framework.** Our pipeline is illustrated in Figure 2. CLIP [28] is a widely recognized text encoder, and we use it to encode the text prompt \\( c \\). The encoded text feature and time-step \\( t \\) are projected into transformer dimension and summed together as the condition token \\( z_{tk} \\). The 2D and 3D motion sequences are projected into the same dimension, concatenated with condition token \\( z_{tk} \\) in time axis and summed with a standard positional embedding. We aim to unify the two domain features in one space but it is too difficult for one linear layer. A straightforward idea is to encode via two separate encoders:" + }, + { + "type": "equation", + "bbox": [ + 0.346, + 0.351, + 0.786, + 0.369 + ], + "angle": 0, + "content": "\\[\nz _ {3 D} ^ {0} = \\mathcal {E} _ {3 D} \\left(x _ {3 D, t}, t, c\\right), z _ {2 D} ^ {0} = \\mathcal {E} _ {2 D} \\left(x _ {2 D, t}, t, c\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.784, + 0.42 + ], + "angle": 0, + "content": "where \\(\\mathcal{E}_{3D / 2D}(\\cdot)\\) are 3D/2D \\(L_{1}\\)-layer transformer encoders [44]. However, we find it more efficient to add another shared-weight encoder to extract shared feature:" + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.42, + 0.785, + 0.441 + ], + "angle": 0, + "content": "\\[\n\\left\\{z _ {3 D / 2 D} ^ {i} \\right\\} _ {i = 1} ^ {L _ {2}} = \\mathcal {E} _ {\\text {s h a r e}} \\left(z _ {3 D / 2 D} ^ {0}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.784, + 0.496 + ], + "angle": 0, + "content": "where \\(\\mathcal{E}_{\\text{share}}(\\cdot)\\) is a shared-weight \\(L_{2}\\)-layer transformer encoder, and \\(\\{z_{3D/2D}^{i}\\}_{i=1}^{L_{2}}\\) are the outputs of each shared-weight layer. The whole process is defined as unified encoding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.497, + 0.785, + 0.618 + ], + "angle": 0, + "content": "To output motion in two modality, we use independent \\( L_{2} \\)-layer transformer decoders [44] for 2D and 3D data. Starting from 2D/3D learnable token embeddings \\( Tok_{2D/3D} \\), each decoder layer takes the output of the previous layer as queries and the output of same-level shared layer \\( z^{i} \\) as keys and values instead of the last layer. The starting point is to make decoder layers follow the extracting patterns of shared-weight layers rather than gaining deeper embeddings. Finally, a linear layer is added to map the features to the dimensions of the motions. This cross-decoding can be integrated as:" + }, + { + "type": "equation", + "bbox": [ + 0.305, + 0.627, + 0.785, + 0.646 + ], + "angle": 0, + "content": "\\[\n\\hat {x} _ {3 D, 0} = \\mathcal {D} _ {3 D} \\left(\\left\\{z _ {3 D / 2 D} ^ {i} \\right\\} _ {i = 1} ^ {L _ {2}}\\right), \\hat {x} _ {2 D, 0} = \\mathcal {D} _ {2 D} \\left(\\left\\{z _ {3 D / 2 D} ^ {i} \\right\\} _ {i = 1} ^ {L _ {2}}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.653, + 0.784, + 0.698 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}_{3D / 2D}\\) are 3D/2D decoders including learnable tokens; and \\(\\hat{x}_{3D / 2D,0}\\) are predicted clean 3D/2D motion sequences. In summary, with a pair of 3D motion and 2D projected motion, CrossDiff outputs four results via" + }, + { + "type": "equation", + "bbox": [ + 0.279, + 0.709, + 0.785, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\hat {x} _ {i D \\rightarrow j D, 0} = G _ {i D \\rightarrow j D} \\left(x _ {i D, t}, t, c\\right) = \\mathcal {D} _ {j D} \\left(\\mathcal {E} _ {\\text {s h a r e}} \\left(\\mathcal {E} _ {i D} \\left(x _ {i D, t}, t, c\\right)\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.784, + 0.765 + ], + "angle": 0, + "content": "where \\(\\hat{x}_{iD\\to jD,0}\\) are predicted \\(j\\) -dimension clean motion \\(\\hat{x}_{jD,0}\\) from \\(i\\) -dimension motion noise \\(x_{iD,t}\\) with \\(i,j\\in \\{2,3\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.785, + 0.784, + 0.816 + ], + "angle": 0, + "content": "Training. As mentioned in Section 3.1, we apply a simple objective (Eq. 2) for all outputs:" + }, + { + "type": "equation", + "bbox": [ + 0.325, + 0.824, + 0.785, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i D \\rightarrow j D} = \\mathbb {E} _ {t \\sim [ 1, T ]} \\| x _ {j D, 0} - G _ {i D \\rightarrow j D} (x _ {i D, t}, t, c) \\| _ {2} ^ {2}. \\tag {7}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.297, + 0.115, + 0.733, + 0.128 + ], + "angle": 0, + "content": "Realistic Human Motion Generation with Cross-Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "We train our model in two stage. In stage I, CrossDiff is forced to learn the reverse process, motion characteristic of texts in both domains and the motion connection of two distribution via the loss" + }, + { + "type": "equation", + "bbox": [ + 0.271, + 0.205, + 0.786, + 0.222 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s t a g e I}} = \\mathcal {L} _ {3 D \\rightarrow 3 D} + w _ {2 3} \\mathcal {L} _ {2 D \\rightarrow 3 D} + w _ {3 2} \\mathcal {L} _ {3 D \\rightarrow 2 D} + w _ {2 2} \\mathcal {L} _ {2 D \\rightarrow 2 D}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.233, + 0.784, + 0.26 + ], + "angle": 0, + "content": "where \\( w_{23}, w_{32}, w_{22} \\) are relative weights. In stage II, there is only a 3D generation loss:" + }, + { + "type": "equation", + "bbox": [ + 0.427, + 0.263, + 0.786, + 0.28 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s t a g e I I}} = \\mathcal {L} _ {3 D \\rightarrow 3 D}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.286, + 0.784, + 0.33 + ], + "angle": 0, + "content": "This helps the model focus on the 3D denoising process and eliminate the uncertainty of the 2D and 3D mapping relationship while retaining the knowledge of diverse motion features." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.353, + 0.414, + 0.369 + ], + "angle": 0, + "content": "3.4 Mixture Sampling" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.405, + 0.78, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.495, + 0.785, + 0.552 + ], + "angle": 0, + "content": "Fig. 3: Overview of Mixture Sampling. The original noise is sampled from a 2D gaussian distribution. From time-step \\( T \\) to \\( \\alpha \\), CrossDiff predicts the clean 2D motion \\( \\hat{x}_{2D,0} \\) and diffuses it back to \\( x_{2D,t-1} \\). In the remaining \\( \\alpha \\) steps, CrossDiff denoises in the 3D domain and finally obtains the clean 3D motion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.784, + 0.642 + ], + "angle": 0, + "content": "After training, one can sample a motion sequence conditioned on a text prompt in an iterative manner. The standard method [40] gradually anneals the 3D noise from a gaussian distribution, which we still use. We predict the clean sample \\(\\hat{x}_{3D,0}\\) and noise it back to \\(x_{3D,t-1}\\) for \\(T\\) steps until \\(x_{3D,0}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.785, + 0.747 + ], + "angle": 0, + "content": "Furthermore, utilizing the CrossDiff architecture, we propose a novel two-domain sampling approach. As shown in Figure 3, We first sample 2D gaussian noise which is then denoised with \\( G_{2D\\to 2D}(x_{2D,t},t,c) \\) until time-step \\( \\alpha \\). Next, we project the denoised 2D noise onto the 3D domain using \\( G_{2D\\to 3D}(x_{2D,t},t,c) \\) and continue the denoising process with \\( G_{3D\\to 3D}(x_{3D,t},t,c) \\) for the remaining \\( \\alpha \\) steps. Our experiments in supplementary demonstrate the difference between mixture sampling and the vanilla method." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.77, + 0.65, + 0.785 + ], + "angle": 0, + "content": "3.5 Learning 3D Motion Generation from 2D Data" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Given the complexity and cost associated with collecting high-quality 3D motion data, generating 3D motion from 2D motion data is an attractive alternative. Moreover, generating 3D motion from textual descriptions in an out-of-domain" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "Z. Ren et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "scenario is approximately a zero-shot task. To achieve this, we first estimated 2D motion from text-related videos using an off-the-shelf model. We then utilized the pretrained model in stage I \\( G_{2D \\to 3D}(x_{2D,t}, t, c) \\) to generate corresponding 3D clean motion, with \\( t \\) set to 0 and \\( c \\) set to null condition \\( \\emptyset \\). A motion filter is applied to smooth the generated motion. We assume that 2D pose estimation is relatively precise, allowing the processed 2D/3D motion to serve as pseudolabels for training. The model is fine-tuned with the same objective as stage I, but with different weight hyper-parameters. After training, our model can generate diverse motion according to out-of-domain textual descriptions using mixture sampling (Sec. 3.4)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.788, + 0.361 + ], + "angle": 0, + "content": "During training with 2D motion estimated from videos, we encounter significant errors in root estimation due to the uncertainty of camera movement. To address this issue, we decouple the root information \\( r_{3D/2D} \\) and generate it based on other pose features. Please refer to supplementary for more details." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.387, + 0.377, + 0.405 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.788, + 0.589 + ], + "angle": 0, + "content": "Our focus is on the text-to-motion generation task, and we conduct experiments on two standard datasets: HumanML3D [8] and KIT Motion-Language (KIT-ML) [27]. In Section 4.1, we introduce these standard datasets and the evaluation metrics used in our experiments. In Section 4.2, we present comparable quantitative results with state-of-the-art methods and outperform diffused methods on the HumanML3D dataset. We also use upper and lower body indices, combined with visualization effects, to illustrate our superior performance in capturing whole-body details. Moreover, we demonstrate that CrossDiff supports training with additional 2D motion sequences in Section 4.3. Finally, we discuss ablation studies in Section 4.4 to analyze the contributions of each component of our proposed method." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.615, + 0.533, + 0.63 + ], + "angle": 0, + "content": "4.1 Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Datasets. The HumanML3D [8] and KIT-ML datasets [27] are widely used in research. KIT-ML provides 6,353 textual descriptions for 3,911 motion sequences, which are all down-sampled to 12.5 FPS. HumanML3D is currently the largest 3D human motion dataset with textual descriptions. The motions are originally from two motion capture datasets, AMASS [22] and HumanAct12 [10], and are rescaled to 20 frames per second. It contains 14,616 motions annotated with 44,970 sequence-level textual descriptions. To enable fair comparisons with previous works [4,8,40], we use the redundant motion representation proposed by [8]. This representation involves re-targeting the joint positions to a default human skeleton template, setting the initial pose at the same position \\((\\mathrm{X} = 0,\\mathrm{Z} = 0)\\) facing the \\(\\mathbf{Z}+\\) direction, and including root global positions, local positions, joint rotations, joint velocities, and foot contact labels. We use the same motion representation for KIT-ML." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.297, + 0.115, + 0.733, + 0.128 + ], + "angle": 0, + "content": "Realistic Human Motion Generation with Cross-Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.206 + ], + "angle": 0, + "content": "To obtain corresponding 2D motion, we utilize orthogonal projection to project 3D motion into 2D planes for four directions (front, left, right, and back). Then, we process the redundant 2D motion representation as described in Section 3.2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.207, + 0.785, + 0.296 + ], + "angle": 0, + "content": "UFC101 [38] is an influential action recognition dataset that contains paired videos and action labels. We obtain 2D joint positions using the ViTPose [48] model and process them into redundant 2D motion representation. We filter out fuzzy data and ultimately select 24 action labels, with each label related to 10-50 motion sequences. We then annotate around five textual descriptions for each label." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.319, + 0.785, + 0.425 + ], + "angle": 0, + "content": "Evaluation Metrics. We compare our results with previous works using the same metrics as [8, 40]. These metrics involve evaluating quality with Frechet Inception Distance (FID), precision with R-Precision and Multi-modal Distance (MM Dist), and diversity with Diversity (DIV) and Multimodality (MModality). These measurements assess generated motion distribution, retrieval ranking accuracy, Euclidean distances between text and motion features, and variance across features and within a single text." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.426, + 0.787, + 0.56 + ], + "angle": 0, + "content": "Besides using the evaluation model from [8], we introduce a new metric measuring the FID (Fréchet Inception Distance) of upper and lower body movements, denoted as FID-U and FID-L. This enables a fine-grained analysis of human motion and better comprehension of upper and lower body dynamics. We split joints into two groups using the root joint as a boundary and train separate evaluators, following a similar approach to [8]. This effectively evaluates generated motion quality for both body segments, offering a deeper understanding of human motion complexities and advancing research on new motion generation models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.785, + 0.674 + ], + "angle": 0, + "content": "Implementation Details. We use the AdamW [19] optimizer, setting the learning rate to 1e-4 and 1e-5 in stages I and II, respectively. Our model is trained for 4k epochs in stage I and 1k epochs in stage II, using 8 Tesla V100 GPUs with a mini-batch size of 32. The number of diffusion steps is set to 1K. The loss weights \\((w_{23}, w_{32}, w_{33})\\) is set to \\((1, 1, 1)\\) in stage I and \\((0.1, 0.1, 1)\\) in training on UFC101 datasets." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.696, + 0.524, + 0.71 + ], + "angle": 0, + "content": "4.2 Comparisons on Text-to-motion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Comparative Analysis of Standard Metrics. In our evaluation, we test our models 20 times and compare their performance with existing state-of-the-art methods. These methods include Language2Pose [1], T2G [2], Hier [6], T2M [8], MotionDiffuse [53], Fg-T2M [45], MDM [40], T2M-GPT [52], MLD [4] and ReMoDiffuse [54]. As illustrated in Table 1, our model exhibits competitive performance when compared to these leading methods. However, it is important to note that the KIT-ML dataset primarily consists of \"walk\" movements and lacks intricate details. Consequently, this dataset does not present the same challenges" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "Z. Ren et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 1: Quantitative results on the HumanML3D and KIT-ML test set. The overall results on KIT-ML are shown on the right, while the results of both widely-used and newly-proposed metrics on HumanML3D are shown on the left. The red and blue colors indicate the best and second-best results, respectively." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.226, + 0.789, + 0.349 + ], + "angle": 0, + "content": "
MethodsHumanML3DKIT-ML
R Precision (top 3)↑FID↓MM Dist↓DIV→ MModality↑FID-U↓FID-L↓R Precision (top 3)↑FID↓MM Dist↓DIV→ MModality↑FID-U↓FID-L↓
Real0.7970.0022.9749.503--0.7790.0312.78811.08--
Language2Pose0.48611.025.2967.676--0.4836.5455.1479.073--
T2G0.3457.6646.0306.409--0.33812.126.9649.334--
Hier0.5526.5325.0128.332--0.5315.2034.9869.563--
T2M0.7401.0673.3409.1882.090-0.6932.7703.40110.911.482-
MotionDiffuse0.7820.6303.1139.4101.553-0.7391.9542.95811.100.730-
Fg-T2M0.7830.2433.1099.2781.614-0.7450.5713.11410.931.019-
MDM0.6110.5445.5669.5592.7990.8250.4840.4979.19110.8471.9070.925
T2M-GPT0.7750.1413.1219.7221.8310.1450.6070.7450.5143.0071.5700.602
MLD0.7720.4733.1969.7242.4130.5410.5530.7340.4043.2042.1920.563
ReMoDiffuse0.7950.1032.9749.0181.7950.1250.5650.7650.1552.8141.8080.205
Ours0.7300.1623.3589.5772.6200.1180.2810.7040.4743.30810.770.434
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.386, + 0.788, + 0.448 + ], + "angle": 0, + "content": "that our method is specifically designed to address. It indicates we can steadily generate high-quality and precise motion while pay attention to rich diversity. In other words, our generated results are not only consistent with the textual description, but also more expressive." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.466, + 0.788, + 0.617 + ], + "angle": 0, + "content": "Comparative Analysis of Fine-grained Metrics. We compare the fine-grained metrics for our upper and lower body with those from four recent studies [4, 40, 52, 54]. As demonstrated in Table 1, our generated motion is more robust and detailed. Our low FID scores for both the upper and lower body indicate that our synthesized motion effectively captures full-body movement rather than focusing solely on specific semantic parts. In contrast, ReMoDiffuse and T2M-GPT achieves a low FID score for the upper body but a high score for the lower body. This suggests that their generation process exhibits unbalanced attention towards different body parts, primarily translating textual descriptions into upper body characteristics rather than capturing the entire body's motion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.618, + 0.788, + 0.768 + ], + "angle": 0, + "content": "Figure 4 displays qualitative comparisons with existing methods. Our method can \"march\" with arm swings, \"wobble\" using hands for balancing and alternate between defense and attack in a \"fight\". MDM exhibits a certain rigidity, while T2M-GPT and ReMoDiffuse appear to lack dynamism. For direct comparisons, please refer to the supplementary videos provided. We conducted a user study on motion performance, in which participants were asked two questions to assess the vitality and diversity of the motions. The results, presented in Figure 5(a), confirm our analysis. In summary, our method demonstrates a superior ability to interpret semantic information and generate more accurate and expressive motions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.788, + 0.458, + 0.803 + ], + "angle": 0, + "content": "4.3 Learning from 2D Data" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "After being trained on a 3D dataset, our model can learn 3D motion generation from 2D data. By fine-tuning the model with the UCF101 dataset [38], we ef" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.297, + 0.115, + 0.733, + 0.128 + ], + "angle": 0, + "content": "Realistic Human Motion Generation with Cross-Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.784, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.47, + 0.788, + 0.514 + ], + "angle": 0, + "content": "Fig. 4: Qualitative results on HumanML3D dataset. We compare our method with MDM [40], T2M-GPT [52] and MLD [4]. We find that our generated actions better convey the intended semantics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.54, + 0.788, + 0.768 + ], + "angle": 0, + "content": "fectively address a zero-shot problem arising from the absence of ground-truth 3D motion. Our sampling strategy reaches optimal performance when \\(\\alpha = 1\\). As depicted in Figure 6, the generated motions for various activities, such as pulling up, biking, table tennis, and baseball, are showcased alongside their textual prompts. Notably, pulling up and biking are entirely outside the HumanML3D motion domain. Although playing table tennis and baseball share similarities with actions like slapping or throwing in HumanML3D, the original model is unable to synthesize motion under those specific descriptions. However, after fine-tuning, our generated motion more closely resembles the given sports rather than merely mimicking the actions. Despite some activities being beyond the scope of the HumanML3D domain, our fine-tuned model successfully synthesizes specific motions by leveraging the weak 2D data. This demonstrates its remarkable adaptability and potential for efficient use of available motion data since 2D motion and related textual descriptions are easier to obtain than 3D motion." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.787, + 0.402, + 0.801 + ], + "angle": 0, + "content": "4.4 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Our model features separate pipelines for 2D and 3D inputs, allowing us to train solely on 3D motion sequences, which is an improvement over MDM [40]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "Z. Ren et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.149, + 0.408, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.448, + 0.143, + 0.788, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.285, + 0.789, + 0.357 + ], + "angle": 0, + "content": "Fig. 5: (a) The result of the user study. (b) Difference between 3D and 2D motion data distribution. The time axis is represented on the x-axis, while the normalized joint velocity is represented on the y-axis. The 3D motion is represented by a blue full line, while the 2D motion is represented by red and green dashed lines, indicating the front and left view, respectively." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.367, + 0.789, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.55, + 0.788, + 0.581 + ], + "angle": 0, + "content": "Fig. 6: Generating 3D movements without training on paired 3D motion and textual descriptions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.788, + 0.641 + ], + "angle": 0, + "content": "We investigate the impact of introducing 2D data on the performance of 3D generation and demonstrate the effectiveness of using a shared-weights encoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Why 2D motion help? To explain the benefits of 2D motions, we compared the distribution differences between 3D and 2D motion data. Hand and feet movements, which are primary indicators of motion, were visualized in both 3D and 2D levels, and their velocities were normalized along the joints dimension. In Figure 5(b), we can clearly see that around the 20th frame, the 2D velocity of the left hand in the front view reaches a higher value while the 3D velocity is quite low, indicating that hand movements in 2D are more prominent than in 3D. The results show that 2D motion captures different details from 3D motion, suggesting that the CrossDiff model can lead 3D motion to learn from the knowledge that 2D motion acquired from text prompts. Specifically, for the given sample, 2D motion might learn \"animated hand motions\" while 3D motion focuses only on \"walking\". 2D motion is an explicit feature that we artificially extract to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.297, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Realistic Human Motion Generation with Cross-Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 2: Evaluation of our models with different settings on the HumanML3D dataset. Bold indicates best result. The symbol % indicates the percentage of data being used. From top to bottom, we present MDM as baselines, the impact of training with 2D representations, with(w/) or without(w/o) shared-weights encoder." + }, + { + "type": "table", + "bbox": [ + 0.284, + 0.226, + 0.721, + 0.43 + ], + "angle": 0, + "content": "
MethodsR Precision (top 3)↑FID↓MM Dist↓DIV→
MDM0.6110.5445.5669.559
Ours0.7300.1623.3589.577
50% 3D0.6660.5863.8949.513
100% 3D0.6850.2243.6909.445
50% 3D + 100% 2D0.6720.4223.7089.345
100% 3D + 100% 2D0.7300.1623.3589.577
w/o shared-weights encoder0.7140.1873.4969.488
w/ shared-weights encoder0.7300.1623.3589.577
1 view(front)0.7220.1863.4679.798
1 view(left)0.7150.1813.4129.834
4 views0.7300.1623.3589.577
5 views0.6950.2023.6139.502
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.471, + 0.785, + 0.503 + ], + "angle": 0, + "content": "aid the model's learning, and this approach can help improve performance when dealing with arbitrary data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.528, + 0.788, + 0.695 + ], + "angle": 0, + "content": "Influence of 2D Representation Table 2 presents the results from four different experiment settings. The control groups of \"100% 3D\" and \"100% 3D + 100% 2D\" demonstrate that when training with paired 3D motion and text, projecting the 3D motion to 2D and building a connection between the 2D motion and text can help boost performance. The visualizations in Figure 1 further highlight the enhanced quality of our generated outputs. The control groups of \"50% 3D\" and \"50% 3D + 100% 2D\" prove that additional 2D data can also help improve performance. The additional 2D data indicates other 2D motion without ground truth 3D motion. The experiment in Section 4.3 shows learning 2D motion in the wild can also help with out-of-domain 3D motion learning. As we can see, the combined learning of 2D motion has great potential." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Shared-weights Encoder Without the shared-weights encoder, the model is a simple encoder-decoder framework with two modalities. However, we believe that this is not sufficient to fully fuse the 3D and 2D motion features. Inspired by [47], we found that when learning with data from two modalities, extracting separate and fused feature layer-by-layer is more efficient. The shared-weights encoder serves as a fusing module, while the 3D/2D motion decoder acts as a separate decoder module. The goal is to ensure that the decoder layers follow the same extraction patterns as the shared-weight layers, rather than simply gaining" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "Z. Ren et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.178 + ], + "angle": 0, + "content": "deeper embeddings. The results presented in Table 2 demonstrate the efficiency of using a shared-weights encoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.197, + 0.788, + 0.44 + ], + "angle": 0, + "content": "2D Representation from different views To evaluate the impact of different views, we conducted tests on four settings: a) only the front view, b) only the left view, c) four views including front, left, back, and right, and d) five views with an additional top view. When multiple views were used, the 3D motion was paired with a random view projection to formulate the loss. As shown in Table 2, it is reasonable to assume that four views contain more 2D information and outperform one view. Furthermore, the front view and left view do not have much distinction in terms of performance. However, the performance actually decreased with the additional top view. This could be due to the fact that the 2D information becomes more difficult to classify without the condition of the camera view. We did not pass along the camera information because it is difficult to estimate the camera view of in-the-wild videos, and injecting camera information could disrupt the structure of the text-to-2D model, making it difficult for the text-to-3D model to follow. In conclusion, the number of views serves as a practical hyperparameter that can be adjusted through enumeration experiments." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.459, + 0.36, + 0.475 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.788, + 0.608 + ], + "angle": 0, + "content": "In conclusion, the Cross Human Motion Diffusion Model (CrossDiff) presents a promising advancement in the field of human motion synthesis by effectively integrating and leveraging both 3D and 2D motion information for high-quality motion generation. The unified encoding and cross-decoding components of the CrossDiff learning framework enable the capture of intricate details of human movement that are often overlooked by models relying solely on 3D data. Our experiments validate the superior performance of the proposed model on various text-to-motion benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.609, + 0.788, + 0.699 + ], + "angle": 0, + "content": "Despite its promising results, the CrossDiff model is not without potential weaknesses. One limitation may arise from the model's ability to generalize to unseen or rare motion patterns, especially when trained only on 2D motion data. Additionally, the computational complexity of the model might hinder its real-time application in certain scenarios, such as interactive gaming and virtual reality environments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.7, + 0.788, + 0.777 + ], + "angle": 0, + "content": "Future work could explore methods to enhance the model's generalization capabilities, such as incorporating unsupervised or semi-supervised learning techniques. To further advance our understanding, we propose the accumulation of a sizable 2D motion dataset coupled with relevant textual prompts, enabling the training of a unified motion generation model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Acknowledgments. This research was partly supported by Shenzhen Key Laboratory of next generation interactive media innovative technology (Grant No: ZDSYS20210623092001004)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.297, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Realistic Human Motion Generation with Cross-Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.173, + 0.787, + 0.214 + ], + "angle": 0, + "content": "1. Ahuja, C., Morency, L.P.: Language2pose: Natural language grounded pose forecasting. In: 2019 International Conference on 3D Vision (3DV). pp. 719-728. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.215, + 0.787, + 0.269 + ], + "angle": 0, + "content": "2. Bhattacharya, U., Rewkowski, N., Banerjee, A., Guhan, P., Bera, A., Manocha, D.: Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In: 2021 IEEE virtual reality and 3D user interfaces (VR). pp. 1-10. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.27, + 0.787, + 0.31 + ], + "angle": 0, + "content": "3. Chen, L.H., Zhang, J., Li, Y., Pang, Y., Xia, X., Liu, T.: Humanmac: Masked motion completion for human motion prediction. arXiv preprint arXiv:2302.03665 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.786, + 0.351 + ], + "angle": 0, + "content": "4. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18000-18010 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.352, + 0.787, + 0.405 + ], + "angle": 0, + "content": "5. Chung, H., Sim, B., Ye, J.C.: Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12413-12422 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.787, + 0.446 + ], + "angle": 0, + "content": "6. Ghosh, A., Cheema, N., Oguz, C., Theobalt, C., Slusallek, P.: Synthesis of compositional animations from textual descriptions. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1396-1406 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.447, + 0.787, + 0.487 + ], + "angle": 0, + "content": "7. Ghosh, P., Song, J., Aksan, E., Hilliges, O.: Learning human motion models for long-term predictions. In: 2017 International Conference on 3D Vision (3DV). pp. 458-466. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.488, + 0.787, + 0.528 + ], + "angle": 0, + "content": "8. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5152-5161 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.529, + 0.787, + 0.569 + ], + "angle": 0, + "content": "9. Guo, C., Zuo, X., Wang, S., Cheng, L.: Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In: European Conference on Computer Vision. pp. 580-597. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.569, + 0.787, + 0.61 + ], + "angle": 0, + "content": "10. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 2021-2029 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.611, + 0.786, + 0.636 + ], + "angle": 0, + "content": "11. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.637, + 0.786, + 0.663 + ], + "angle": 0, + "content": "12. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.664, + 0.786, + 0.704 + ], + "angle": 0, + "content": "13. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Gmd: Controllable human motion synthesis via guided diffusion models. arXiv preprint arXiv:2305.12577 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.705, + 0.786, + 0.731 + ], + "angle": 0, + "content": "14. Lee, H.Y., Yang, X., Liu, M.Y., Wang, T.C., Lu, Y.D., Yang, M.H., Kautz, J.: Dancing to music. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.732, + 0.786, + 0.785 + ], + "angle": 0, + "content": "15. Li, J., Kang, D., Pei, W., Zhe, X., Zhang, Y., He, Z., Bao, L.: Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11293-11302 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.786, + 0.786, + 0.84 + ], + "angle": 0, + "content": "16. Li, M., Chen, S., Zhao, Y., Zhang, Y., Wang, Y., Tian, Q.: Dynamic multiscale graph neural networks for 3d skeleton based human motion prediction. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 214-223 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.219, + 0.173, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "Z. Ren et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "17. Li, R., Zhao, J., Zhang, Y., Su, M., Ren, Z., Zhang, H., Tang, Y., Li, X.: Finedance: A fine-grained choreography dataset for 3d full body dance generation. arXiv preprint arXiv:2212.03741 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "18. Li, Z., Zhou, Y., Xiao, S., He, C., Huang, Z., Li, H.: Auto-conditioned recurrent networks for extended complex human motion synthesis. arXiv preprint arXiv:1707.05363 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "19. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.313 + ], + "angle": 0, + "content": "20. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11461-11471 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.37 + ], + "angle": 0, + "content": "21. Luo, Z., Chen, D., Zhang, Y., Huang, Y., Wang, L., Shen, Y., Zhao, D., Zhou, J., Tan, T.: Videofusion: Decomposed diffusion models for high-quality video generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10209-10218 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.412 + ], + "angle": 0, + "content": "22. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5442-5451 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.439 + ], + "angle": 0, + "content": "23. Nie, W., Guo, B., Huang, Y., Xiao, C., Vahdat, A., Anandkumar, A.: Diffusion models for adversarial purification. arXiv preprint arXiv:2205.07460 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.44, + 0.785, + 0.466 + ], + "angle": 0, + "content": "24. Pavllo, D., Grangier, D., Auli, M.: Quaternet: A quaternion-based recurrent model for human motion. arXiv preprint arXiv:1805.06485 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.785, + 0.509 + ], + "angle": 0, + "content": "25. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3d human motion synthesis with transformer vae. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10985-10995 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "26. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: European Conference on Computer Vision. pp. 480-497. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.577 + ], + "angle": 0, + "content": "27. Plappert, M., Mandery, C., Asfour, T.: The kit motion-language dataset. Big data 4(4), 236-252 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.632 + ], + "angle": 0, + "content": "28. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.675 + ], + "angle": 0, + "content": "29. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.716 + ], + "angle": 0, + "content": "30. Ren, Z., Pan, Z., Zhou, X., Kang, L.: Diffusion motion: Generate text-guided 3d human motion by diffusion model. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.785, + 0.758 + ], + "angle": 0, + "content": "31. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.813 + ], + "angle": 0, + "content": "32. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "33. Shafir, Y., Tevet, G., Kapon, R., Bermano, A.H.: Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.297, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Realistic Human Motion Generation with Cross-Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "34. Sinha, A., Song, J., Meng, C., Ermon, S.: D2c: Diffusion-decoding models for few-shot conditional generation. Advances in Neural Information Processing Systems 34, 12533-12548 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.23 + ], + "angle": 0, + "content": "35. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256-2265. PMLR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.23, + 0.788, + 0.257 + ], + "angle": 0, + "content": "36. Song, Y., Ermon, S.: Improved techniques for training score-based generative models. Advances in neural information processing systems 33, 12438-12448 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.257, + 0.788, + 0.298 + ], + "angle": 0, + "content": "37. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.298, + 0.788, + 0.325 + ], + "angle": 0, + "content": "38. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.325, + 0.788, + 0.366 + ], + "angle": 0, + "content": "39. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: European Conference on Computer Vision. pp. 358-374. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.366, + 0.788, + 0.393 + ], + "angle": 0, + "content": "40. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. arXiv preprint arXiv:2209.14916 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.393, + 0.788, + 0.434 + ], + "angle": 0, + "content": "41. Tseng, J., Castellon, R., Liu, K.: Edge: Editable dance generation from music. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 448-458 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.461 + ], + "angle": 0, + "content": "42. Vahdat, A., Kreis, K., Kautz, J.: Score-based generative modeling in latent space. Advances in Neural Information Processing Systems 34, 11287-11302 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.461, + 0.788, + 0.487 + ], + "angle": 0, + "content": "43. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.487, + 0.788, + 0.528 + ], + "angle": 0, + "content": "44. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.528, + 0.788, + 0.569 + ], + "angle": 0, + "content": "45. Wang, Y., Leng, Z., Li, F.W., Wu, S.C., Liang, X.: Fg-t2m: Fine-grained text-driven human motion generation via diffusion model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22035-22044 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.61 + ], + "angle": 0, + "content": "46. Wei, D., Sun, H., Li, B., Lu, J., Li, W., Sun, X., Hu, S.: Human joint kinematics diffusion-refinement for stochastic motion prediction. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 6110-6118 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.61, + 0.788, + 0.664 + ], + "angle": 0, + "content": "47. Xu, X., Wu, C., Rosenman, S., Lal, V., Che, W., Duan, N.: Bridgetower: Building bridges between encoders in vision-language representation learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 10637-10647 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.664, + 0.788, + 0.704 + ], + "angle": 0, + "content": "48. Xu, Y., Zhang, J., Zhang, Q., Tao, D.: Vitpose: Simple vision transformer baselines for human pose estimation. Advances in Neural Information Processing Systems 35, 38571-38584 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.732 + ], + "angle": 0, + "content": "49. Yang, R., Srivastava, P., Mandt, S.: Diffusion probabilistic modeling for video generation. arXiv preprint arXiv:2203.09481 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.732, + 0.788, + 0.772 + ], + "angle": 0, + "content": "50. Yoon, Y., Cha, B., Lee, J.H., Jang, M., Lee, J., Kim, J., Lee, G.: Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics (TOG) 39(6), 1-16 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.772, + 0.788, + 0.8 + ], + "angle": 0, + "content": "51. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: Physdiff: Physics-guided human motion diffusion model. arXiv preprint arXiv:2212.02500 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.8, + 0.788, + 0.841 + ], + "angle": 0, + "content": "52. Zhang, J., Zhang, Y., Cun, X., Huang, S., Zhang, Y., Zhao, H., Lu, H., Shen, X.: T2m-gpt: Generating human motion from textual descriptions with discrete representations. arXiv preprint arXiv:2301.06052 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "Z. Ren et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.189 + ], + "angle": 0, + "content": "53. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.231 + ], + "angle": 0, + "content": "54. Zhang, M., Guo, X., Pan, L., Cai, Z., Hong, F., Li, H., Yang, L., Liu, Z.: Remodif- \nfuse: Retrieval-augmented motion diffusion model. arXiv preprint arXiv:2304.01116 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.786, + 0.273 + ], + "angle": 0, + "content": "55. Zhuang, H., Zhang, Y., Liu, S.: A pilot study of query-free adversarial attack against stable diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2384-2391 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.273 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Realistic Human Motion Generation with Cross-Diffusion Models/5b1805dd-89b3-4df9-9e0a-1ed375e3f44c_origin.pdf b/2024/Realistic Human Motion Generation with Cross-Diffusion Models/5b1805dd-89b3-4df9-9e0a-1ed375e3f44c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1c7ba5c91c51f09a3a3e5b4811134e38726b1a56 --- /dev/null +++ b/2024/Realistic Human Motion Generation with Cross-Diffusion Models/5b1805dd-89b3-4df9-9e0a-1ed375e3f44c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a71663978f9e2a85b5481476fc6c8767a6239a73e16a64a9698a2768bbf0581f +size 2194543 diff --git a/2024/Realistic Human Motion Generation with Cross-Diffusion Models/full.md b/2024/Realistic Human Motion Generation with Cross-Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1eddc07b57eb17088ad44b1310eac14094a6a18b --- /dev/null +++ b/2024/Realistic Human Motion Generation with Cross-Diffusion Models/full.md @@ -0,0 +1,294 @@ +# Realistic Human Motion Generation with Cross-Diffusion Models + +Zeping Ren $^{1*}$ , Shaoli Huang $^{2\dagger}$ , Xiu Li $^{1\dagger}$ + +$^{1}$ Tsinghua Shenzhen International Graduate School + $^{2}$ Tencent AI Lab + +Abstract. In this work, we introduce the Cross Human Motion Diffusion Model (CrossDiff $^3$ ), a novel approach for generating high-quality human motion based on textual descriptions. Our method integrates 3D and 2D information using a shared transformer network within the training of the diffusion model, unifying motion noise into a single feature space. This enables cross-decoding of features into both 3D and 2D motion representations, regardless of their original dimension. The primary advantage of CrossDiff is its cross-diffusion mechanism, which allows the model to reverse either 2D or 3D noise into clean motion during training. This capability leverages the complementary information in both motion representations, capturing intricate human movement details often missed by models relying solely on 3D information. Consequently, CrossDiff effectively combines the strengths of both representations to generate more realistic motion sequences. In our experiments, our model demonstrates competitive state-of-the-art performance on text-to-motion benchmarks. Moreover, our method consistently provides enhanced motion generation quality, capturing complex full-body movement intricacies. Additionally, with a pre-trained model, our approach accommodates using in-the-wild 2D motion data without 3D motion ground truth during training to generate 3D motion, highlighting its potential for broader applications and efficient use of available data resources. + +# 1 Introduction + +In recent years, the field of human motion synthesis [7,9,10,14,15,18,41,50] has witnessed significant advancements, primarily driven by the growing demand for high-quality, realistic motion generation in applications such as gaming, virtual reality, and robotics. + +A crucial aspect in this research area is generating human motion based on textual descriptions, enabling contextually accurate and natural movements [40]. However, current methods [4,8,26,40] predominantly rely on 3D motion information during training, leading to an inability to capture the full spectrum of intricacies associated with human motion. When using only 3D representation, + +![](images/aa7e76b8920840cc33a9d3d660034946fef1934e781406e8d1c390bdcddfccfa.jpg) +Fig. 1: Our method utilizing the cross-diffusion mechanism (Left) exhibits more full-body details compared to existing methods (Right). + +the generation model may struggle to relate text semantics to some body part movements with very small movement variations compared to others, which can lead to overlooking important motion details. This is because the model might focus on more dominant or larger movements within the 3D space, leaving subtle nuances underrepresented. For example, when given a prompt such as "a person is dancing eloquently," as illustrated in Figure 1, the generated motion might lack vitality, display a limited range of movements, and contain minimal local motion details. + +To effectively address the limitations and accurately capture the nuances of full-body movement, we introduce the Cross Human Motion Diffusion Model (CrossDiff). This innovative approach seamlessly integrates and leverages both 3D and 2D motion information to generate high-quality human motion sequences. The 2D data representation effectively illustrates the intricacies of human body movements from various viewing angle projections. Due to different view projections in 2D data, small body part movements can be magnified in certain projections, making them more noticeable and easier to capture. This helps the text-to-motion generation models to better associate text descriptions with a wider range of human body motion details, including subtle movements that might have been overlooked when relying solely on 3D representation. + +As a result, incorporating 2D information with 3D enables the diffusion model to establish more connections between motion and text prompts, ultimately enhancing the motion synthesis process. The CrossDiff learning framework consists of two main components: unified encoding and cross-decoding. These components work together to achieve more precise and realistic motion synthesis. Furthermore, it is essential to transfer the knowledge acquired in the 2D domain to 3D motion, which leads to an overall improvement in the model's performance. + +Unified encoding fuses motion noise from both 3D and 2D sources into a single feature space, facilitating cross-decoding of features into either 3D or 2D motion representations, regardless of their original dimension. The distinctive innovation of our approach stems from the cross-diffusion mechanism, which enables the model to transform 2D or 3D noise into clean motion during the training process. This capability allows the model to harness the complementary information present in both motion representations, effectively capturing + +intricate details of human movement that are often missed by models relying exclusively on 3D data. + +In experiments, we demonstrate our model achieves competitive state-of-the-art performance on several text-to-motion benchmarks, outperforming existing diffusion-based approaches that rely solely on 3D motion information during training. Furthermore, our method consistently delivers enhanced motion generation quality, capturing complex full-body movement intricacies essential for realistic motion synthesis. A notable advantage of our approach is its ability to utilize 2D motion data without necessitating 3D motion ground truth during training, enabling the generation of 3D motion. This feature underscores the potential of the CrossDiff model for a wide range of applications and efficient use of available data resources. + +# 2 Related Work + +# 2.1 Human Motion Generation + +Human motion generation is the process of synthesizing human motion either unconditionally or conditioned by signals such as text, audio, or action labels. Early works [7, 16, 18, 24] treated this as a deterministic mapping problem, generating a single motion from a specific signal using neural networks. However, human motion is inherently stochastic, even under certain conditions, leading to the adoption of deep generative models in more recent research. + +For instance, Dancing2music [14] employed GANs to generate motion under corresponding conditions. ACTOR [25] introduced a framework based on transformers [44] and VAEs, which, although designed for action-to-motion tasks, can be easily adapted for text-to-motion tasks as demonstrated in TEMOS [26]. Since text and audio are time-series data, natural language processing approaches are commonly used. Works by [6], [1], and [8] utilized GRU-based language models to process motion data along the time axis. + +MotionCLIP [39] uses the shared text-image space learned by CLIP [28] to align the feature space of human motion with that of CLIP. MotionGPT [35,36] directly treats motion as language and addresses the motion generation task as a translation problem. However, conditions like language and human motion differ significantly in terms of distribution and expression, making accurate alignment challenging. + +To overcome this issue, T2M-GPT [52] and TM2T [9] encode motion using VQ-VAE [43] and generate motion embeddings with generative pretrained transformers. MotionDiffuse [53] is the first application of diffusion models in text-to-motion tasks. MDM [40] employs a simple diffusion framework to diffuse raw motion data, while MLD [4] encodes motion using a VAE model and diffuses it in the latent space. ReMoDiffuse [54] retrieves the motion related to the text to assist in motion generation. Meanwhile, Fg-T2M [45] utilizes a fine-grained method to extract neighborhood and overall semantic linguistic features. Although these methods attain success, they depend exclusively on 3D + +![](images/86acb77782775995898ad7825a170fffd7001afd59889499e67f8f4d9acc0857.jpg) +Fig. 2: Overview of our CrossDiff framework for generating human motion from textual descriptions. The framework incorporates both 3D and 2D motion data, using unified encoding and cross-decoding components to process mixed representations obtained from random projection. + +motion data during training, which results in a failure to capture sufficient complexities associated with human motion. In contrast, our approach utilizes a cross-diffusion mechanism to leverage the complementary information found in both 2D and 3D motion representations. + +# 2.2 Diffusion Models + +Diffusion generative models [11, 35, 36], based on stochastic diffusion processes in Thermodynamics, involve a forward process where samples from the data distribution are progressively noised towards a Gaussian distribution and a reverse process where the model learns to denoise Gaussian samples. These models have achieved success in various domains, including image synthesis [29, 31, 32, 34, 42], video generation [11, 21, 49], adversarial attacks [23, 55], motion prediction [3, 46], music-to-dance synthesis [17, 41], and text-to-motion generation [4, 30, 40, 51, 53]. + +Classifier-Free Guidance [12] enables conditioned generation without additional model training, balancing fidelity and diversity. In the image domain, inpainting methods [5,20,37] iteratively incorporate known information into the diffusion process, maintaining constant image parts while learning to inpaint others. Similar approaches have been applied to motion editing [13,33,40]. Differing from these works, our work focuses on leveraging multiple data representations to enhance diffusion model learning. + +# 3 Method + +# 3.1 Overview + +Given a textual description $c$ , our objective is to generate multiple human motion sequences $x^{1:N} = \{x^i\}_{i=1}^N$ , each with a length of $N$ . As illustrated in Figure 2, our + +method is carefully designed to efficiently incorporate both 3D and 2D motion data within the learning process of the diffusion model. + +During the training phase, we first obtain mixed representations of the data from the provided 3D input using a random projection technique. Afterward, the 2D and 3D data representations are independently diffused and processed through our learning framework, CrossDiff, which primarily consists of unified encoding and cross-decoding components. + +The unified encoding module maps both the 2D and 3D data into a shared feature space. These features are then passed through the cross-decoding component, resulting in the generation of two motion representations. These representations are subsequently employed for loss calculation and model learning. In the inference phase, our approach supports not only standard sampling but also mixture sampling. + +Preliminary. Denoising diffusion probabilistic models (DDPM) [11] can iteratively eliminate noise from a gaussian distribution to approximate a true data distribution. This technique has had a significant impact on the field of generative research, including text-to-motion applications. In this study, we have adapted DDPM and trained a transformer-based model to gradually reduce noise and generate motion sequences. + +Diffusion is modeled as a Markov noising process $\{x_{t}^{1:N}\}_{t=0}^{T}$ of $T$ steps. For simplicity, we use $x_{t}$ to denote $x_{t}^{1:N}$ in the following discussion. Starting with a motion sequence $x_{0}$ in original data distribution, the noising process can be described as + +$$ +q \left(x _ {t} \mid x _ {t - 1}\right) = \mathcal {N} \left(x _ {t}; \sqrt {\alpha_ {t}} x _ {t - 1}, (1 - \alpha_ {t}) \mathbf {I}\right) \tag {1} +$$ + +where $\alpha_{t} \in (0,1)$ is constant hyper-parameters. When $\alpha_{T}$ is small enough, we can approximate $x_{T} \in \mathcal{N}(0,\mathbf{I})$ . The reverse process is to progressively denoise $x_{T}$ from a gaussian distribution to obtain the clean motion $x_0$ . Following [29,40], we predict the clean motion $x_0$ itself on textual condition $c$ as $\hat{x}_0 = G(x_t,t,c)$ . We apply the simple objective [11] + +$$ +\mathcal {L} _ {\text {s i m p l e}} = \mathbb {E} _ {t \sim [ 1, T ]} \| x _ {0} - G (x _ {t}, t, c) \| _ {2} ^ {2}. \tag {2} +$$ + +# 3.2 Mixed Representations + +As the naive diffusion model is trained only on one data distribution (3D poses), we have trained it on a mixture representation of 3D and 2D poses. To obtain 2D data that is closer to the real distribution, we randomly projected the 3D poses into 2D planes in four directions (front, left, right, and back). The 3D poses $x_{3D}^{i} \in \mathbb{R}^{d_{3D}}$ and 2D poses $x_{2D}^{i} \in \mathbb{R}^{d_{2D}}$ are represented respectively by $d_{3D}$ -dimensional and $d_{2D}$ -dimensional redundant features, respectively, as suggested by [8]. The pose $x^{i}$ is defined by a tuple of $(r,j^{p},j^{v},j^{r},c^{f})$ , where $(r_{3D},j_{3D}^{p},j_{3D}^{v},j_{3D}^{r},c_{3D}^{f})$ is identical to [8]. In addition, $r_{2D} \in \mathbb{R}^2$ represents 2D root velocity. $j_{2D}^{p} \in \mathbb{R}^{2j},j_{2D}^{v} \in \mathbb{R}^{2j}$ and $j_{2D}^{r} \in \mathbb{R}^{2j}$ represent the local joints positions, velocities and rotations, respectively, with $j$ denoting the number of joints besides the root. $c_{2D}^{f} \in \mathbb{R}^{4}$ is a set of binary features obtained by thresholding the heel and toe + +joint velocities. Notably, the rotation representation is made up of the sine and cosine values of the angle. + +# 3.3 Cross Motion Diffusion Model + +**Framework.** Our pipeline is illustrated in Figure 2. CLIP [28] is a widely recognized text encoder, and we use it to encode the text prompt $c$ . The encoded text feature and time-step $t$ are projected into transformer dimension and summed together as the condition token $z_{tk}$ . The 2D and 3D motion sequences are projected into the same dimension, concatenated with condition token $z_{tk}$ in time axis and summed with a standard positional embedding. We aim to unify the two domain features in one space but it is too difficult for one linear layer. A straightforward idea is to encode via two separate encoders: + +$$ +z _ {3 D} ^ {0} = \mathcal {E} _ {3 D} \left(x _ {3 D, t}, t, c\right), z _ {2 D} ^ {0} = \mathcal {E} _ {2 D} \left(x _ {2 D, t}, t, c\right), \tag {3} +$$ + +where $\mathcal{E}_{3D / 2D}(\cdot)$ are 3D/2D $L_{1}$ -layer transformer encoders [44]. However, we find it more efficient to add another shared-weight encoder to extract shared feature: + +$$ +\left\{z _ {3 D / 2 D} ^ {i} \right\} _ {i = 1} ^ {L _ {2}} = \mathcal {E} _ {\text {s h a r e}} \left(z _ {3 D / 2 D} ^ {0}\right), \tag {4} +$$ + +where $\mathcal{E}_{\text{share}}(\cdot)$ is a shared-weight $L_{2}$ -layer transformer encoder, and $\{z_{3D/2D}^{i}\}_{i=1}^{L_{2}}$ are the outputs of each shared-weight layer. The whole process is defined as unified encoding. + +To output motion in two modality, we use independent $L_{2}$ -layer transformer decoders [44] for 2D and 3D data. Starting from 2D/3D learnable token embeddings $Tok_{2D/3D}$ , each decoder layer takes the output of the previous layer as queries and the output of same-level shared layer $z^{i}$ as keys and values instead of the last layer. The starting point is to make decoder layers follow the extracting patterns of shared-weight layers rather than gaining deeper embeddings. Finally, a linear layer is added to map the features to the dimensions of the motions. This cross-decoding can be integrated as: + +$$ +\hat {x} _ {3 D, 0} = \mathcal {D} _ {3 D} \left(\left\{z _ {3 D / 2 D} ^ {i} \right\} _ {i = 1} ^ {L _ {2}}\right), \hat {x} _ {2 D, 0} = \mathcal {D} _ {2 D} \left(\left\{z _ {3 D / 2 D} ^ {i} \right\} _ {i = 1} ^ {L _ {2}}\right), \tag {5} +$$ + +where $\mathcal{D}_{3D / 2D}$ are 3D/2D decoders including learnable tokens; and $\hat{x}_{3D / 2D,0}$ are predicted clean 3D/2D motion sequences. In summary, with a pair of 3D motion and 2D projected motion, CrossDiff outputs four results via + +$$ +\hat {x} _ {i D \rightarrow j D, 0} = G _ {i D \rightarrow j D} \left(x _ {i D, t}, t, c\right) = \mathcal {D} _ {j D} \left(\mathcal {E} _ {\text {s h a r e}} \left(\mathcal {E} _ {i D} \left(x _ {i D, t}, t, c\right)\right)\right), \tag {6} +$$ + +where $\hat{x}_{iD\to jD,0}$ are predicted $j$ -dimension clean motion $\hat{x}_{jD,0}$ from $i$ -dimension motion noise $x_{iD,t}$ with $i,j\in \{2,3\}$ . + +Training. As mentioned in Section 3.1, we apply a simple objective (Eq. 2) for all outputs: + +$$ +\mathcal {L} _ {i D \rightarrow j D} = \mathbb {E} _ {t \sim [ 1, T ]} \| x _ {j D, 0} - G _ {i D \rightarrow j D} (x _ {i D, t}, t, c) \| _ {2} ^ {2}. \tag {7} +$$ + +We train our model in two stage. In stage I, CrossDiff is forced to learn the reverse process, motion characteristic of texts in both domains and the motion connection of two distribution via the loss + +$$ +\mathcal {L} _ {\text {s t a g e I}} = \mathcal {L} _ {3 D \rightarrow 3 D} + w _ {2 3} \mathcal {L} _ {2 D \rightarrow 3 D} + w _ {3 2} \mathcal {L} _ {3 D \rightarrow 2 D} + w _ {2 2} \mathcal {L} _ {2 D \rightarrow 2 D}, \tag {8} +$$ + +where $w_{23}, w_{32}, w_{22}$ are relative weights. In stage II, there is only a 3D generation loss: + +$$ +\mathcal {L} _ {\text {s t a g e I I}} = \mathcal {L} _ {3 D \rightarrow 3 D}. \tag {9} +$$ + +This helps the model focus on the 3D denoising process and eliminate the uncertainty of the 2D and 3D mapping relationship while retaining the knowledge of diverse motion features. + +# 3.4 Mixture Sampling + +![](images/c881a86b105e53ff8bfc2fd03eaa5271b7b38f105f1c4c97cd08023400335077.jpg) +Fig. 3: Overview of Mixture Sampling. The original noise is sampled from a 2D gaussian distribution. From time-step $T$ to $\alpha$ , CrossDiff predicts the clean 2D motion $\hat{x}_{2D,0}$ and diffuses it back to $x_{2D,t-1}$ . In the remaining $\alpha$ steps, CrossDiff denoises in the 3D domain and finally obtains the clean 3D motion. + +After training, one can sample a motion sequence conditioned on a text prompt in an iterative manner. The standard method [40] gradually anneals the 3D noise from a gaussian distribution, which we still use. We predict the clean sample $\hat{x}_{3D,0}$ and noise it back to $x_{3D,t-1}$ for $T$ steps until $x_{3D,0}$ . + +Furthermore, utilizing the CrossDiff architecture, we propose a novel two-domain sampling approach. As shown in Figure 3, We first sample 2D gaussian noise which is then denoised with $G_{2D\to 2D}(x_{2D,t},t,c)$ until time-step $\alpha$ . Next, we project the denoised 2D noise onto the 3D domain using $G_{2D\to 3D}(x_{2D,t},t,c)$ and continue the denoising process with $G_{3D\to 3D}(x_{3D,t},t,c)$ for the remaining $\alpha$ steps. Our experiments in supplementary demonstrate the difference between mixture sampling and the vanilla method. + +# 3.5 Learning 3D Motion Generation from 2D Data + +Given the complexity and cost associated with collecting high-quality 3D motion data, generating 3D motion from 2D motion data is an attractive alternative. Moreover, generating 3D motion from textual descriptions in an out-of-domain + +scenario is approximately a zero-shot task. To achieve this, we first estimated 2D motion from text-related videos using an off-the-shelf model. We then utilized the pretrained model in stage I $G_{2D \to 3D}(x_{2D,t}, t, c)$ to generate corresponding 3D clean motion, with $t$ set to 0 and $c$ set to null condition $\emptyset$ . A motion filter is applied to smooth the generated motion. We assume that 2D pose estimation is relatively precise, allowing the processed 2D/3D motion to serve as pseudolabels for training. The model is fine-tuned with the same objective as stage I, but with different weight hyper-parameters. After training, our model can generate diverse motion according to out-of-domain textual descriptions using mixture sampling (Sec. 3.4). + +During training with 2D motion estimated from videos, we encounter significant errors in root estimation due to the uncertainty of camera movement. To address this issue, we decouple the root information $r_{3D/2D}$ and generate it based on other pose features. Please refer to supplementary for more details. + +# 4 Experiments + +Our focus is on the text-to-motion generation task, and we conduct experiments on two standard datasets: HumanML3D [8] and KIT Motion-Language (KIT-ML) [27]. In Section 4.1, we introduce these standard datasets and the evaluation metrics used in our experiments. In Section 4.2, we present comparable quantitative results with state-of-the-art methods and outperform diffused methods on the HumanML3D dataset. We also use upper and lower body indices, combined with visualization effects, to illustrate our superior performance in capturing whole-body details. Moreover, we demonstrate that CrossDiff supports training with additional 2D motion sequences in Section 4.3. Finally, we discuss ablation studies in Section 4.4 to analyze the contributions of each component of our proposed method. + +# 4.1 Datasets and Evaluation Metrics + +Datasets. The HumanML3D [8] and KIT-ML datasets [27] are widely used in research. KIT-ML provides 6,353 textual descriptions for 3,911 motion sequences, which are all down-sampled to 12.5 FPS. HumanML3D is currently the largest 3D human motion dataset with textual descriptions. The motions are originally from two motion capture datasets, AMASS [22] and HumanAct12 [10], and are rescaled to 20 frames per second. It contains 14,616 motions annotated with 44,970 sequence-level textual descriptions. To enable fair comparisons with previous works [4,8,40], we use the redundant motion representation proposed by [8]. This representation involves re-targeting the joint positions to a default human skeleton template, setting the initial pose at the same position $(\mathrm{X} = 0,\mathrm{Z} = 0)$ facing the $\mathbf{Z}+$ direction, and including root global positions, local positions, joint rotations, joint velocities, and foot contact labels. We use the same motion representation for KIT-ML. + +To obtain corresponding 2D motion, we utilize orthogonal projection to project 3D motion into 2D planes for four directions (front, left, right, and back). Then, we process the redundant 2D motion representation as described in Section 3.2. + +UFC101 [38] is an influential action recognition dataset that contains paired videos and action labels. We obtain 2D joint positions using the ViTPose [48] model and process them into redundant 2D motion representation. We filter out fuzzy data and ultimately select 24 action labels, with each label related to 10-50 motion sequences. We then annotate around five textual descriptions for each label. + +Evaluation Metrics. We compare our results with previous works using the same metrics as [8, 40]. These metrics involve evaluating quality with Frechet Inception Distance (FID), precision with R-Precision and Multi-modal Distance (MM Dist), and diversity with Diversity (DIV) and Multimodality (MModality). These measurements assess generated motion distribution, retrieval ranking accuracy, Euclidean distances between text and motion features, and variance across features and within a single text. + +Besides using the evaluation model from [8], we introduce a new metric measuring the FID (Fréchet Inception Distance) of upper and lower body movements, denoted as FID-U and FID-L. This enables a fine-grained analysis of human motion and better comprehension of upper and lower body dynamics. We split joints into two groups using the root joint as a boundary and train separate evaluators, following a similar approach to [8]. This effectively evaluates generated motion quality for both body segments, offering a deeper understanding of human motion complexities and advancing research on new motion generation models. + +Implementation Details. We use the AdamW [19] optimizer, setting the learning rate to 1e-4 and 1e-5 in stages I and II, respectively. Our model is trained for 4k epochs in stage I and 1k epochs in stage II, using 8 Tesla V100 GPUs with a mini-batch size of 32. The number of diffusion steps is set to 1K. The loss weights $(w_{23}, w_{32}, w_{33})$ is set to $(1, 1, 1)$ in stage I and $(0.1, 0.1, 1)$ in training on UFC101 datasets. + +# 4.2 Comparisons on Text-to-motion + +Comparative Analysis of Standard Metrics. In our evaluation, we test our models 20 times and compare their performance with existing state-of-the-art methods. These methods include Language2Pose [1], T2G [2], Hier [6], T2M [8], MotionDiffuse [53], Fg-T2M [45], MDM [40], T2M-GPT [52], MLD [4] and ReMoDiffuse [54]. As illustrated in Table 1, our model exhibits competitive performance when compared to these leading methods. However, it is important to note that the KIT-ML dataset primarily consists of "walk" movements and lacks intricate details. Consequently, this dataset does not present the same challenges + +Table 1: Quantitative results on the HumanML3D and KIT-ML test set. The overall results on KIT-ML are shown on the right, while the results of both widely-used and newly-proposed metrics on HumanML3D are shown on the left. The red and blue colors indicate the best and second-best results, respectively. + +
MethodsHumanML3DKIT-ML
R Precision (top 3)↑FID↓MM Dist↓DIV→ MModality↑FID-U↓FID-L↓R Precision (top 3)↑FID↓MM Dist↓DIV→ MModality↑FID-U↓FID-L↓
Real0.7970.0022.9749.503--0.7790.0312.78811.08--
Language2Pose0.48611.025.2967.676--0.4836.5455.1479.073--
T2G0.3457.6646.0306.409--0.33812.126.9649.334--
Hier0.5526.5325.0128.332--0.5315.2034.9869.563--
T2M0.7401.0673.3409.1882.090-0.6932.7703.40110.911.482-
MotionDiffuse0.7820.6303.1139.4101.553-0.7391.9542.95811.100.730-
Fg-T2M0.7830.2433.1099.2781.614-0.7450.5713.11410.931.019-
MDM0.6110.5445.5669.5592.7990.8250.4840.4979.19110.8471.9070.925
T2M-GPT0.7750.1413.1219.7221.8310.1450.6070.7450.5143.0071.5700.602
MLD0.7720.4733.1969.7242.4130.5410.5530.7340.4043.2042.1920.563
ReMoDiffuse0.7950.1032.9749.0181.7950.1250.5650.7650.1552.8141.8080.205
Ours0.7300.1623.3589.5772.6200.1180.2810.7040.4743.30810.770.434
+ +that our method is specifically designed to address. It indicates we can steadily generate high-quality and precise motion while pay attention to rich diversity. In other words, our generated results are not only consistent with the textual description, but also more expressive. + +Comparative Analysis of Fine-grained Metrics. We compare the fine-grained metrics for our upper and lower body with those from four recent studies [4, 40, 52, 54]. As demonstrated in Table 1, our generated motion is more robust and detailed. Our low FID scores for both the upper and lower body indicate that our synthesized motion effectively captures full-body movement rather than focusing solely on specific semantic parts. In contrast, ReMoDiffuse and T2M-GPT achieves a low FID score for the upper body but a high score for the lower body. This suggests that their generation process exhibits unbalanced attention towards different body parts, primarily translating textual descriptions into upper body characteristics rather than capturing the entire body's motion. + +Figure 4 displays qualitative comparisons with existing methods. Our method can "march" with arm swings, "wobble" using hands for balancing and alternate between defense and attack in a "fight". MDM exhibits a certain rigidity, while T2M-GPT and ReMoDiffuse appear to lack dynamism. For direct comparisons, please refer to the supplementary videos provided. We conducted a user study on motion performance, in which participants were asked two questions to assess the vitality and diversity of the motions. The results, presented in Figure 5(a), confirm our analysis. In summary, our method demonstrates a superior ability to interpret semantic information and generate more accurate and expressive motions. + +# 4.3 Learning from 2D Data + +After being trained on a 3D dataset, our model can learn 3D motion generation from 2D data. By fine-tuning the model with the UCF101 dataset [38], we ef + +![](images/a8f5cabae1da8913966082f8e909fe530cc00beab367e9af967ff3a758e0a8fb.jpg) +Fig. 4: Qualitative results on HumanML3D dataset. We compare our method with MDM [40], T2M-GPT [52] and MLD [4]. We find that our generated actions better convey the intended semantics. + +fectively address a zero-shot problem arising from the absence of ground-truth 3D motion. Our sampling strategy reaches optimal performance when $\alpha = 1$ . As depicted in Figure 6, the generated motions for various activities, such as pulling up, biking, table tennis, and baseball, are showcased alongside their textual prompts. Notably, pulling up and biking are entirely outside the HumanML3D motion domain. Although playing table tennis and baseball share similarities with actions like slapping or throwing in HumanML3D, the original model is unable to synthesize motion under those specific descriptions. However, after fine-tuning, our generated motion more closely resembles the given sports rather than merely mimicking the actions. Despite some activities being beyond the scope of the HumanML3D domain, our fine-tuned model successfully synthesizes specific motions by leveraging the weak 2D data. This demonstrates its remarkable adaptability and potential for efficient use of available motion data since 2D motion and related textual descriptions are easier to obtain than 3D motion. + +# 4.4 Ablation Studies + +Our model features separate pipelines for 2D and 3D inputs, allowing us to train solely on 3D motion sequences, which is an improvement over MDM [40]. + +![](images/917d6d57989f0277642179d6280abda59d4250989c1a571a9700021312078382.jpg) +Fig. 5: (a) The result of the user study. (b) Difference between 3D and 2D motion data distribution. The time axis is represented on the x-axis, while the normalized joint velocity is represented on the y-axis. The 3D motion is represented by a blue full line, while the 2D motion is represented by red and green dashed lines, indicating the front and left view, respectively. + +![](images/65ff17a8c208517b3e98ced8ff77c8ea391c8c655c106f9d9891d633d4b7ae30.jpg) + +![](images/b7b6c6c3cba1bacdeb542dfbcdb1adfceabd06e88f8f1dbd76282b3a486a8ef6.jpg) +Fig. 6: Generating 3D movements without training on paired 3D motion and textual descriptions. + +We investigate the impact of introducing 2D data on the performance of 3D generation and demonstrate the effectiveness of using a shared-weights encoder. + +Why 2D motion help? To explain the benefits of 2D motions, we compared the distribution differences between 3D and 2D motion data. Hand and feet movements, which are primary indicators of motion, were visualized in both 3D and 2D levels, and their velocities were normalized along the joints dimension. In Figure 5(b), we can clearly see that around the 20th frame, the 2D velocity of the left hand in the front view reaches a higher value while the 3D velocity is quite low, indicating that hand movements in 2D are more prominent than in 3D. The results show that 2D motion captures different details from 3D motion, suggesting that the CrossDiff model can lead 3D motion to learn from the knowledge that 2D motion acquired from text prompts. Specifically, for the given sample, 2D motion might learn "animated hand motions" while 3D motion focuses only on "walking". 2D motion is an explicit feature that we artificially extract to + +Table 2: Evaluation of our models with different settings on the HumanML3D dataset. Bold indicates best result. The symbol % indicates the percentage of data being used. From top to bottom, we present MDM as baselines, the impact of training with 2D representations, with(w/) or without(w/o) shared-weights encoder. + +
MethodsR Precision (top 3)↑FID↓MM Dist↓DIV→
MDM0.6110.5445.5669.559
Ours0.7300.1623.3589.577
50% 3D0.6660.5863.8949.513
100% 3D0.6850.2243.6909.445
50% 3D + 100% 2D0.6720.4223.7089.345
100% 3D + 100% 2D0.7300.1623.3589.577
w/o shared-weights encoder0.7140.1873.4969.488
w/ shared-weights encoder0.7300.1623.3589.577
1 view(front)0.7220.1863.4679.798
1 view(left)0.7150.1813.4129.834
4 views0.7300.1623.3589.577
5 views0.6950.2023.6139.502
+ +aid the model's learning, and this approach can help improve performance when dealing with arbitrary data. + +Influence of 2D Representation Table 2 presents the results from four different experiment settings. The control groups of "100% 3D" and "100% 3D + 100% 2D" demonstrate that when training with paired 3D motion and text, projecting the 3D motion to 2D and building a connection between the 2D motion and text can help boost performance. The visualizations in Figure 1 further highlight the enhanced quality of our generated outputs. The control groups of "50% 3D" and "50% 3D + 100% 2D" prove that additional 2D data can also help improve performance. The additional 2D data indicates other 2D motion without ground truth 3D motion. The experiment in Section 4.3 shows learning 2D motion in the wild can also help with out-of-domain 3D motion learning. As we can see, the combined learning of 2D motion has great potential. + +Shared-weights Encoder Without the shared-weights encoder, the model is a simple encoder-decoder framework with two modalities. However, we believe that this is not sufficient to fully fuse the 3D and 2D motion features. Inspired by [47], we found that when learning with data from two modalities, extracting separate and fused feature layer-by-layer is more efficient. The shared-weights encoder serves as a fusing module, while the 3D/2D motion decoder acts as a separate decoder module. The goal is to ensure that the decoder layers follow the same extraction patterns as the shared-weight layers, rather than simply gaining + +deeper embeddings. The results presented in Table 2 demonstrate the efficiency of using a shared-weights encoder. + +2D Representation from different views To evaluate the impact of different views, we conducted tests on four settings: a) only the front view, b) only the left view, c) four views including front, left, back, and right, and d) five views with an additional top view. When multiple views were used, the 3D motion was paired with a random view projection to formulate the loss. As shown in Table 2, it is reasonable to assume that four views contain more 2D information and outperform one view. Furthermore, the front view and left view do not have much distinction in terms of performance. However, the performance actually decreased with the additional top view. This could be due to the fact that the 2D information becomes more difficult to classify without the condition of the camera view. We did not pass along the camera information because it is difficult to estimate the camera view of in-the-wild videos, and injecting camera information could disrupt the structure of the text-to-2D model, making it difficult for the text-to-3D model to follow. In conclusion, the number of views serves as a practical hyperparameter that can be adjusted through enumeration experiments. + +# 5 Conclusion + +In conclusion, the Cross Human Motion Diffusion Model (CrossDiff) presents a promising advancement in the field of human motion synthesis by effectively integrating and leveraging both 3D and 2D motion information for high-quality motion generation. The unified encoding and cross-decoding components of the CrossDiff learning framework enable the capture of intricate details of human movement that are often overlooked by models relying solely on 3D data. Our experiments validate the superior performance of the proposed model on various text-to-motion benchmarks. + +Despite its promising results, the CrossDiff model is not without potential weaknesses. One limitation may arise from the model's ability to generalize to unseen or rare motion patterns, especially when trained only on 2D motion data. Additionally, the computational complexity of the model might hinder its real-time application in certain scenarios, such as interactive gaming and virtual reality environments. + +Future work could explore methods to enhance the model's generalization capabilities, such as incorporating unsupervised or semi-supervised learning techniques. To further advance our understanding, we propose the accumulation of a sizable 2D motion dataset coupled with relevant textual prompts, enabling the training of a unified motion generation model. + +Acknowledgments. This research was partly supported by Shenzhen Key Laboratory of next generation interactive media innovative technology (Grant No: ZDSYS20210623092001004). + +# References + +1. Ahuja, C., Morency, L.P.: Language2pose: Natural language grounded pose forecasting. In: 2019 International Conference on 3D Vision (3DV). pp. 719-728. IEEE (2019) +2. Bhattacharya, U., Rewkowski, N., Banerjee, A., Guhan, P., Bera, A., Manocha, D.: Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In: 2021 IEEE virtual reality and 3D user interfaces (VR). pp. 1-10. IEEE (2021) +3. Chen, L.H., Zhang, J., Li, Y., Pang, Y., Xia, X., Liu, T.: Humanmac: Masked motion completion for human motion prediction. arXiv preprint arXiv:2302.03665 (2023) +4. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18000-18010 (2023) +5. Chung, H., Sim, B., Ye, J.C.: Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12413-12422 (2022) +6. Ghosh, A., Cheema, N., Oguz, C., Theobalt, C., Slusallek, P.: Synthesis of compositional animations from textual descriptions. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1396-1406 (2021) +7. Ghosh, P., Song, J., Aksan, E., Hilliges, O.: Learning human motion models for long-term predictions. In: 2017 International Conference on 3D Vision (3DV). pp. 458-466. IEEE (2017) +8. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5152-5161 (2022) +9. Guo, C., Zuo, X., Wang, S., Cheng, L.: Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In: European Conference on Computer Vision. pp. 580-597. Springer (2022) +10. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 2021-2029 (2020) +11. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020) +12. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022) +13. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Gmd: Controllable human motion synthesis via guided diffusion models. arXiv preprint arXiv:2305.12577 (2023) +14. Lee, H.Y., Yang, X., Liu, M.Y., Wang, T.C., Lu, Y.D., Yang, M.H., Kautz, J.: Dancing to music. Advances in neural information processing systems 32 (2019) +15. Li, J., Kang, D., Pei, W., Zhe, X., Zhang, Y., He, Z., Bao, L.: Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11293-11302 (2021) +16. Li, M., Chen, S., Zhao, Y., Zhang, Y., Wang, Y., Tian, Q.: Dynamic multiscale graph neural networks for 3d skeleton based human motion prediction. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 214-223 (2020) + +17. Li, R., Zhao, J., Zhang, Y., Su, M., Ren, Z., Zhang, H., Tang, Y., Li, X.: Finedance: A fine-grained choreography dataset for 3d full body dance generation. arXiv preprint arXiv:2212.03741 (2023) +18. Li, Z., Zhou, Y., Xiao, S., He, C., Huang, Z., Li, H.: Auto-conditioned recurrent networks for extended complex human motion synthesis. arXiv preprint arXiv:1707.05363 (2017) +19. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017) +20. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11461-11471 (2022) +21. Luo, Z., Chen, D., Zhang, Y., Huang, Y., Wang, L., Shen, Y., Zhao, D., Zhou, J., Tan, T.: Videofusion: Decomposed diffusion models for high-quality video generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10209-10218 (2023) +22. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5442-5451 (2019) +23. Nie, W., Guo, B., Huang, Y., Xiao, C., Vahdat, A., Anandkumar, A.: Diffusion models for adversarial purification. arXiv preprint arXiv:2205.07460 (2022) +24. Pavllo, D., Grangier, D., Auli, M.: Quaternet: A quaternion-based recurrent model for human motion. arXiv preprint arXiv:1805.06485 (2018) +25. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3d human motion synthesis with transformer vae. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10985-10995 (2021) +26. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: European Conference on Computer Vision. pp. 480-497. Springer (2022) +27. Plappert, M., Mandery, C., Asfour, T.: The kit motion-language dataset. Big data 4(4), 236-252 (2016) +28. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +29. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022) +30. Ren, Z., Pan, Z., Zhou, X., Kang, L.: Diffusion motion: Generate text-guided 3d human motion by diffusion model. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023) +31. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) +32. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022) +33. Shafir, Y., Tevet, G., Kapon, R., Bermano, A.H.: Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418 (2023) + +34. Sinha, A., Song, J., Meng, C., Ermon, S.: D2c: Diffusion-decoding models for few-shot conditional generation. Advances in Neural Information Processing Systems 34, 12533-12548 (2021) +35. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256-2265. PMLR (2015) +36. Song, Y., Ermon, S.: Improved techniques for training score-based generative models. Advances in neural information processing systems 33, 12438-12448 (2020) +37. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020) +38. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012) +39. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: European Conference on Computer Vision. pp. 358-374. Springer (2022) +40. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. arXiv preprint arXiv:2209.14916 (2022) +41. Tseng, J., Castellon, R., Liu, K.: Edge: Editable dance generation from music. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 448-458 (2023) +42. Vahdat, A., Kreis, K., Kautz, J.: Score-based generative modeling in latent space. Advances in Neural Information Processing Systems 34, 11287-11302 (2021) +43. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems 30 (2017) +44. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) +45. Wang, Y., Leng, Z., Li, F.W., Wu, S.C., Liang, X.: Fg-t2m: Fine-grained text-driven human motion generation via diffusion model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22035-22044 (2023) +46. Wei, D., Sun, H., Li, B., Lu, J., Li, W., Sun, X., Hu, S.: Human joint kinematics diffusion-refinement for stochastic motion prediction. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 6110-6118 (2023) +47. Xu, X., Wu, C., Rosenman, S., Lal, V., Che, W., Duan, N.: Bridgetower: Building bridges between encoders in vision-language representation learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 10637-10647 (2023) +48. Xu, Y., Zhang, J., Zhang, Q., Tao, D.: Vitpose: Simple vision transformer baselines for human pose estimation. Advances in Neural Information Processing Systems 35, 38571-38584 (2022) +49. Yang, R., Srivastava, P., Mandt, S.: Diffusion probabilistic modeling for video generation. arXiv preprint arXiv:2203.09481 (2022) +50. Yoon, Y., Cha, B., Lee, J.H., Jang, M., Lee, J., Kim, J., Lee, G.: Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics (TOG) 39(6), 1-16 (2020) +51. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: Physdiff: Physics-guided human motion diffusion model. arXiv preprint arXiv:2212.02500 (2022) +52. Zhang, J., Zhang, Y., Cun, X., Huang, S., Zhang, Y., Zhao, H., Lu, H., Shen, X.: T2m-gpt: Generating human motion from textual descriptions with discrete representations. arXiv preprint arXiv:2301.06052 (2023) + +53. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022) +54. Zhang, M., Guo, X., Pan, L., Cai, Z., Hong, F., Li, H., Yang, L., Liu, Z.: Remodif- +fuse: Retrieval-augmented motion diffusion model. arXiv preprint arXiv:2304.01116 (2023) +55. Zhuang, H., Zhang, Y., Liu, S.: A pilot study of query-free adversarial attack against stable diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2384-2391 (2023) \ No newline at end of file diff --git a/2024/Realistic Human Motion Generation with Cross-Diffusion Models/images.zip b/2024/Realistic Human Motion Generation with Cross-Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..49acc984ffb6764993b2f64783a110b8defbdc99 --- /dev/null +++ b/2024/Realistic Human Motion Generation with Cross-Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b905a1a8b4ffb648a992d74712d4f9dfd23664d2bf0ce54af04d41a13d5257ef +size 402739 diff --git a/2024/Realistic Human Motion Generation with Cross-Diffusion Models/layout.json b/2024/Realistic Human Motion Generation with Cross-Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..465ce5b3c2f0a7b13f1e8fe18326f04344b507eb --- /dev/null +++ b/2024/Realistic Human Motion Generation with Cross-Diffusion Models/layout.json @@ -0,0 +1,8658 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 161, + 111, + 455, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 111, + 455, + 148 + ], + "spans": [ + { + "bbox": [ + 161, + 111, + 455, + 148 + ], + "type": "text", + "content": "Realistic Human Motion Generation with Cross-Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 208, + 167, + 406, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 167, + 406, + 181 + ], + "spans": [ + { + "bbox": [ + 208, + 167, + 406, + 181 + ], + "type": "text", + "content": "Zeping Ren" + }, + { + "bbox": [ + 208, + 167, + 406, + 181 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 208, + 167, + 406, + 181 + ], + "type": "text", + "content": ", Shaoli Huang" + }, + { + "bbox": [ + 208, + 167, + 406, + 181 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 208, + 167, + 406, + 181 + ], + "type": "text", + "content": ", Xiu Li" + }, + { + "bbox": [ + 208, + 167, + 406, + 181 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 200, + 189, + 414, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 189, + 414, + 212 + ], + "spans": [ + { + "bbox": [ + 200, + 189, + 414, + 212 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 200, + 189, + 414, + 212 + ], + "type": "text", + "content": " Tsinghua Shenzhen International Graduate School \n" + }, + { + "bbox": [ + 200, + 189, + 414, + 212 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 200, + 189, + 414, + 212 + ], + "type": "text", + "content": " Tencent AI Lab" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 240, + 455, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 240, + 455, + 472 + ], + "spans": [ + { + "bbox": [ + 160, + 240, + 455, + 472 + ], + "type": "text", + "content": "Abstract. In this work, we introduce the Cross Human Motion Diffusion Model (CrossDiff" + }, + { + "bbox": [ + 160, + 240, + 455, + 472 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 160, + 240, + 455, + 472 + ], + "type": "text", + "content": "), a novel approach for generating high-quality human motion based on textual descriptions. Our method integrates 3D and 2D information using a shared transformer network within the training of the diffusion model, unifying motion noise into a single feature space. This enables cross-decoding of features into both 3D and 2D motion representations, regardless of their original dimension. The primary advantage of CrossDiff is its cross-diffusion mechanism, which allows the model to reverse either 2D or 3D noise into clean motion during training. This capability leverages the complementary information in both motion representations, capturing intricate human movement details often missed by models relying solely on 3D information. Consequently, CrossDiff effectively combines the strengths of both representations to generate more realistic motion sequences. In our experiments, our model demonstrates competitive state-of-the-art performance on text-to-motion benchmarks. Moreover, our method consistently provides enhanced motion generation quality, capturing complex full-body movement intricacies. Additionally, with a pre-trained model, our approach accommodates using in-the-wild 2D motion data without 3D motion ground truth during training to generate 3D motion, highlighting its potential for broader applications and efficient use of available data resources." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 491, + 230, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 230, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 230, + 504 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 516, + 482, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 516, + 482, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 516, + 482, + 563 + ], + "type": "text", + "content": "In recent years, the field of human motion synthesis [7,9,10,14,15,18,41,50] has witnessed significant advancements, primarily driven by the growing demand for high-quality, realistic motion generation in applications such as gaming, virtual reality, and robotics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 564, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 564, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 564, + 482, + 624 + ], + "type": "text", + "content": "A crucial aspect in this research area is generating human motion based on textual descriptions, enabling contextually accurate and natural movements [40]. However, current methods [4,8,26,40] predominantly rely on 3D motion information during training, leading to an inability to capture the full spectrum of intricacies associated with human motion. When using only 3D representation," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 124, + 631, + 345, + 643 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 631, + 345, + 643 + ], + "spans": [ + { + "bbox": [ + 124, + 631, + 345, + 643 + ], + "type": "text", + "content": "* Work done during an internship at Tencent AI Lab." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 125, + 643, + 226, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 643, + 226, + 654 + ], + "spans": [ + { + "bbox": [ + 125, + 643, + 226, + 654 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 125, + 654, + 328, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 654, + 328, + 665 + ], + "spans": [ + { + "bbox": [ + 125, + 654, + 328, + 665 + ], + "type": "text", + "content": "3 https://wonderno.github.io/CrossDiff-webpage/" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 482, + 203 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 482, + 203 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 482, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 482, + 203 + ], + "type": "image", + "image_path": "aa7e76b8920840cc33a9d3d660034946fef1934e781406e8d1c390bdcddfccfa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 222, + 480, + 245 + ], + "lines": [ + { + "bbox": [ + 131, + 222, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 131, + 222, + 480, + 245 + ], + "type": "text", + "content": "Fig. 1: Our method utilizing the cross-diffusion mechanism (Left) exhibits more full-body details compared to existing methods (Right)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 269, + 482, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 482, + 364 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 482, + 364 + ], + "type": "text", + "content": "the generation model may struggle to relate text semantics to some body part movements with very small movement variations compared to others, which can lead to overlooking important motion details. This is because the model might focus on more dominant or larger movements within the 3D space, leaving subtle nuances underrepresented. For example, when given a prompt such as \"a person is dancing eloquently,\" as illustrated in Figure 1, the generated motion might lack vitality, display a limited range of movements, and contain minimal local motion details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 365, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 496 + ], + "type": "text", + "content": "To effectively address the limitations and accurately capture the nuances of full-body movement, we introduce the Cross Human Motion Diffusion Model (CrossDiff). This innovative approach seamlessly integrates and leverages both 3D and 2D motion information to generate high-quality human motion sequences. The 2D data representation effectively illustrates the intricacies of human body movements from various viewing angle projections. Due to different view projections in 2D data, small body part movements can be magnified in certain projections, making them more noticeable and easier to capture. This helps the text-to-motion generation models to better associate text descriptions with a wider range of human body motion details, including subtle movements that might have been overlooked when relying solely on 3D representation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 497, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 581 + ], + "type": "text", + "content": "As a result, incorporating 2D information with 3D enables the diffusion model to establish more connections between motion and text prompts, ultimately enhancing the motion synthesis process. The CrossDiff learning framework consists of two main components: unified encoding and cross-decoding. These components work together to achieve more precise and realistic motion synthesis. Furthermore, it is essential to transfer the knowledge acquired in the 2D domain to 3D motion, which leads to an overall improvement in the model's performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Unified encoding fuses motion noise from both 3D and 2D sources into a single feature space, facilitating cross-decoding of features into either 3D or 2D motion representations, regardless of their original dimension. The distinctive innovation of our approach stems from the cross-diffusion mechanism, which enables the model to transform 2D or 3D noise into clean motion during the training process. This capability allows the model to harness the complementary information present in both motion representations, effectively capturing" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Z. Ren et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "intricate details of human movement that are often missed by models relying exclusively on 3D data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 260 + ], + "type": "text", + "content": "In experiments, we demonstrate our model achieves competitive state-of-the-art performance on several text-to-motion benchmarks, outperforming existing diffusion-based approaches that rely solely on 3D motion information during training. Furthermore, our method consistently delivers enhanced motion generation quality, capturing complex full-body movement intricacies essential for realistic motion synthesis. A notable advantage of our approach is its ability to utilize 2D motion data without necessitating 3D motion ground truth during training, enabling the generation of 3D motion. This feature underscores the potential of the CrossDiff model for a wide range of applications and efficient use of available data resources." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 281, + 237, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 237, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 237, + 293 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 308, + 298, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 308, + 298, + 319 + ], + "spans": [ + { + "bbox": [ + 131, + 308, + 298, + 319 + ], + "type": "text", + "content": "2.1 Human Motion Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 329, + 479, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 479, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 479, + 400 + ], + "type": "text", + "content": "Human motion generation is the process of synthesizing human motion either unconditionally or conditioned by signals such as text, audio, or action labels. Early works [7, 16, 18, 24] treated this as a deterministic mapping problem, generating a single motion from a specific signal using neural networks. However, human motion is inherently stochastic, even under certain conditions, leading to the adoption of deep generative models in more recent research." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 401, + 480, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 480, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 480, + 484 + ], + "type": "text", + "content": "For instance, Dancing2music [14] employed GANs to generate motion under corresponding conditions. ACTOR [25] introduced a framework based on transformers [44] and VAEs, which, although designed for action-to-motion tasks, can be easily adapted for text-to-motion tasks as demonstrated in TEMOS [26]. Since text and audio are time-series data, natural language processing approaches are commonly used. Works by [6], [1], and [8] utilized GRU-based language models to process motion data along the time axis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 485, + 480, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 480, + 556 + ], + "type": "text", + "content": "MotionCLIP [39] uses the shared text-image space learned by CLIP [28] to align the feature space of human motion with that of CLIP. MotionGPT [35,36] directly treats motion as language and addresses the motion generation task as a translation problem. However, conditions like language and human motion differ significantly in terms of distribution and expression, making accurate alignment challenging." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "text", + "content": "To overcome this issue, T2M-GPT [52] and TM2T [9] encode motion using VQ-VAE [43] and generate motion embeddings with generative pretrained transformers. MotionDiffuse [53] is the first application of diffusion models in text-to-motion tasks. MDM [40] employs a simple diffusion framework to diffuse raw motion data, while MLD [4] encodes motion using a VAE model and diffuses it in the latent space. ReMoDiffuse [54] retrieves the motion related to the text to assist in motion generation. Meanwhile, Fg-T2M [45] utilizes a fine-grained method to extract neighborhood and overall semantic linguistic features. Although these methods attain success, they depend exclusively on 3D" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 181, + 91, + 448, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 91, + 448, + 101 + ], + "spans": [ + { + "bbox": [ + 181, + 91, + 448, + 101 + ], + "type": "text", + "content": "Realistic Human Motion Generation with Cross-Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 479, + 235 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 479, + 235 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 479, + 235 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 479, + 235 + ], + "type": "image", + "image_path": "86acb77782775995898ad7825a170fffd7001afd59889499e67f8f4d9acc0857.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 254, + 482, + 299 + ], + "lines": [ + { + "bbox": [ + 130, + 254, + 482, + 299 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 482, + 299 + ], + "type": "text", + "content": "Fig. 2: Overview of our CrossDiff framework for generating human motion from textual descriptions. The framework incorporates both 3D and 2D motion data, using unified encoding and cross-decoding components to process mixed representations obtained from random projection." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 322, + 482, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 482, + 370 + ], + "type": "text", + "content": "motion data during training, which results in a failure to capture sufficient complexities associated with human motion. In contrast, our approach utilizes a cross-diffusion mechanism to leverage the complementary information found in both 2D and 3D motion representations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 388, + 246, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 388, + 246, + 399 + ], + "spans": [ + { + "bbox": [ + 132, + 388, + 246, + 399 + ], + "type": "text", + "content": "2.2 Diffusion Models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 408, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 482, + 491 + ], + "type": "text", + "content": "Diffusion generative models [11, 35, 36], based on stochastic diffusion processes in Thermodynamics, involve a forward process where samples from the data distribution are progressively noised towards a Gaussian distribution and a reverse process where the model learns to denoise Gaussian samples. These models have achieved success in various domains, including image synthesis [29, 31, 32, 34, 42], video generation [11, 21, 49], adversarial attacks [23, 55], motion prediction [3, 46], music-to-dance synthesis [17, 41], and text-to-motion generation [4, 30, 40, 51, 53]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 492, + 482, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 492, + 482, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 492, + 482, + 575 + ], + "type": "text", + "content": "Classifier-Free Guidance [12] enables conditioned generation without additional model training, balancing fidelity and diversity. In the image domain, inpainting methods [5,20,37] iteratively incorporate known information into the diffusion process, maintaining constant image parts while learning to inpaint others. Similar approaches have been applied to motion editing [13,33,40]. Differing from these works, our work focuses on leveraging multiple data representations to enhance diffusion model learning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 594, + 202, + 606 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 202, + 606 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 202, + 606 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 620, + 208, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 208, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 208, + 631 + ], + "type": "text", + "content": "3.1 Overview" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 639, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 639, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 639, + 481, + 667 + ], + "type": "text", + "content": "Given a textual description " + }, + { + "bbox": [ + 130, + 639, + 481, + 667 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 639, + 481, + 667 + ], + "type": "text", + "content": ", our objective is to generate multiple human motion sequences " + }, + { + "bbox": [ + 130, + 639, + 481, + 667 + ], + "type": "inline_equation", + "content": "x^{1:N} = \\{x^i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 639, + 481, + 667 + ], + "type": "text", + "content": ", each with a length of " + }, + { + "bbox": [ + 130, + 639, + 481, + 667 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 639, + 481, + 667 + ], + "type": "text", + "content": ". As illustrated in Figure 2, our" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Z. Ren et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "method is carefully designed to efficiently incorporate both 3D and 2D motion data within the learning process of the diffusion model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": "During the training phase, we first obtain mixed representations of the data from the provided 3D input using a random projection technique. Afterward, the 2D and 3D data representations are independently diffused and processed through our learning framework, CrossDiff, which primarily consists of unified encoding and cross-decoding components." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 200, + 480, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 480, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 480, + 272 + ], + "type": "text", + "content": "The unified encoding module maps both the 2D and 3D data into a shared feature space. These features are then passed through the cross-decoding component, resulting in the generation of two motion representations. These representations are subsequently employed for loss calculation and model learning. In the inference phase, our approach supports not only standard sampling but also mixture sampling." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 272, + 480, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 480, + 343 + ], + "type": "text", + "content": "Preliminary. Denoising diffusion probabilistic models (DDPM) [11] can iteratively eliminate noise from a gaussian distribution to approximate a true data distribution. This technique has had a significant impact on the field of generative research, including text-to-motion applications. In this study, we have adapted DDPM and trained a transformer-based model to gradually reduce noise and generate motion sequences." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 343, + 480, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 480, + 390 + ], + "type": "text", + "content": "Diffusion is modeled as a Markov noising process " + }, + { + "bbox": [ + 130, + 343, + 480, + 390 + ], + "type": "inline_equation", + "content": "\\{x_{t}^{1:N}\\}_{t=0}^{T}" + }, + { + "bbox": [ + 130, + 343, + 480, + 390 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 343, + 480, + 390 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 343, + 480, + 390 + ], + "type": "text", + "content": " steps. For simplicity, we use " + }, + { + "bbox": [ + 130, + 343, + 480, + 390 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 130, + 343, + 480, + 390 + ], + "type": "text", + "content": " to denote " + }, + { + "bbox": [ + 130, + 343, + 480, + 390 + ], + "type": "inline_equation", + "content": "x_{t}^{1:N}" + }, + { + "bbox": [ + 130, + 343, + 480, + 390 + ], + "type": "text", + "content": " in the following discussion. Starting with a motion sequence " + }, + { + "bbox": [ + 130, + 343, + 480, + 390 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 130, + 343, + 480, + 390 + ], + "type": "text", + "content": " in original data distribution, the noising process can be described as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 222, + 392, + 480, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 392, + 480, + 405 + ], + "spans": [ + { + "bbox": [ + 222, + 392, + 480, + 405 + ], + "type": "interline_equation", + "content": "q \\left(x _ {t} \\mid x _ {t - 1}\\right) = \\mathcal {N} \\left(x _ {t}; \\sqrt {\\alpha_ {t}} x _ {t - 1}, (1 - \\alpha_ {t}) \\mathbf {I}\\right) \\tag {1}", + "image_path": "ad3754b9b171a0ec4dbdf8040a4a5857f8003a81b9454fa8d111f2c91af45aae.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 410, + 480, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 480, + 471 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 410, + 480, + 471 + ], + "type": "inline_equation", + "content": "\\alpha_{t} \\in (0,1)" + }, + { + "bbox": [ + 130, + 410, + 480, + 471 + ], + "type": "text", + "content": " is constant hyper-parameters. When " + }, + { + "bbox": [ + 130, + 410, + 480, + 471 + ], + "type": "inline_equation", + "content": "\\alpha_{T}" + }, + { + "bbox": [ + 130, + 410, + 480, + 471 + ], + "type": "text", + "content": " is small enough, we can approximate " + }, + { + "bbox": [ + 130, + 410, + 480, + 471 + ], + "type": "inline_equation", + "content": "x_{T} \\in \\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 130, + 410, + 480, + 471 + ], + "type": "text", + "content": ". The reverse process is to progressively denoise " + }, + { + "bbox": [ + 130, + 410, + 480, + 471 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 130, + 410, + 480, + 471 + ], + "type": "text", + "content": " from a gaussian distribution to obtain the clean motion " + }, + { + "bbox": [ + 130, + 410, + 480, + 471 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 130, + 410, + 480, + 471 + ], + "type": "text", + "content": ". Following [29,40], we predict the clean motion " + }, + { + "bbox": [ + 130, + 410, + 480, + 471 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 130, + 410, + 480, + 471 + ], + "type": "text", + "content": " itself on textual condition " + }, + { + "bbox": [ + 130, + 410, + 480, + 471 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 410, + 480, + 471 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 410, + 480, + 471 + ], + "type": "inline_equation", + "content": "\\hat{x}_0 = G(x_t,t,c)" + }, + { + "bbox": [ + 130, + 410, + 480, + 471 + ], + "type": "text", + "content": ". We apply the simple objective [11]" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 227, + 481, + 480, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 481, + 480, + 495 + ], + "spans": [ + { + "bbox": [ + 227, + 481, + 480, + 495 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s i m p l e}} = \\mathbb {E} _ {t \\sim [ 1, T ]} \\| x _ {0} - G (x _ {t}, t, c) \\| _ {2} ^ {2}. \\tag {2}", + "image_path": "1eb520be3ea3707b9ffa30d72dcfdb2213f375392341b0e8f94f6f5c3d5a21cf.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 512, + 277, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 512, + 277, + 525 + ], + "spans": [ + { + "bbox": [ + 131, + 512, + 277, + 525 + ], + "type": "text", + "content": "3.2 Mixed Representations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 533, + 481, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 668 + ], + "type": "text", + "content": "As the naive diffusion model is trained only on one data distribution (3D poses), we have trained it on a mixture representation of 3D and 2D poses. To obtain 2D data that is closer to the real distribution, we randomly projected the 3D poses into 2D planes in four directions (front, left, right, and back). The 3D poses " + }, + { + "bbox": [ + 130, + 533, + 481, + 668 + ], + "type": "inline_equation", + "content": "x_{3D}^{i} \\in \\mathbb{R}^{d_{3D}}" + }, + { + "bbox": [ + 130, + 533, + 481, + 668 + ], + "type": "text", + "content": " and 2D poses " + }, + { + "bbox": [ + 130, + 533, + 481, + 668 + ], + "type": "inline_equation", + "content": "x_{2D}^{i} \\in \\mathbb{R}^{d_{2D}}" + }, + { + "bbox": [ + 130, + 533, + 481, + 668 + ], + "type": "text", + "content": " are represented respectively by " + }, + { + "bbox": [ + 130, + 533, + 481, + 668 + ], + "type": "inline_equation", + "content": "d_{3D}" + }, + { + "bbox": [ + 130, + 533, + 481, + 668 + ], + "type": "text", + "content": "-dimensional and " + }, + { + "bbox": [ + 130, + 533, + 481, + 668 + ], + "type": "inline_equation", + "content": "d_{2D}" + }, + { + "bbox": [ + 130, + 533, + 481, + 668 + ], + "type": "text", + "content": "-dimensional redundant features, respectively, as suggested by [8]. The pose " + }, + { + "bbox": [ + 130, + 533, + 481, + 668 + ], + "type": "inline_equation", + "content": "x^{i}" + }, + { + "bbox": [ + 130, + 533, + 481, + 668 + ], + "type": "text", + "content": " is defined by a tuple of " + }, + { + "bbox": [ + 130, + 533, + 481, + 668 + ], + "type": "inline_equation", + "content": "(r,j^{p},j^{v},j^{r},c^{f})" + }, + { + "bbox": [ + 130, + 533, + 481, + 668 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 533, + 481, + 668 + ], + "type": "inline_equation", + "content": "(r_{3D},j_{3D}^{p},j_{3D}^{v},j_{3D}^{r},c_{3D}^{f})" + }, + { + "bbox": [ + 130, + 533, + 481, + 668 + ], + "type": "text", + "content": " is identical to [8]. In addition, " + }, + { + "bbox": [ + 130, + 533, + 481, + 668 + ], + "type": "inline_equation", + "content": "r_{2D} \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 130, + 533, + 481, + 668 + ], + "type": "text", + "content": " represents 2D root velocity. " + }, + { + "bbox": [ + 130, + 533, + 481, + 668 + ], + "type": "inline_equation", + "content": "j_{2D}^{p} \\in \\mathbb{R}^{2j},j_{2D}^{v} \\in \\mathbb{R}^{2j}" + }, + { + "bbox": [ + 130, + 533, + 481, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 481, + 668 + ], + "type": "inline_equation", + "content": "j_{2D}^{r} \\in \\mathbb{R}^{2j}" + }, + { + "bbox": [ + 130, + 533, + 481, + 668 + ], + "type": "text", + "content": " represent the local joints positions, velocities and rotations, respectively, with " + }, + { + "bbox": [ + 130, + 533, + 481, + 668 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 533, + 481, + 668 + ], + "type": "text", + "content": " denoting the number of joints besides the root. " + }, + { + "bbox": [ + 130, + 533, + 481, + 668 + ], + "type": "inline_equation", + "content": "c_{2D}^{f} \\in \\mathbb{R}^{4}" + }, + { + "bbox": [ + 130, + 533, + 481, + 668 + ], + "type": "text", + "content": " is a set of binary features obtained by thresholding the heel and toe" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 181, + 91, + 448, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 91, + 448, + 101 + ], + "spans": [ + { + "bbox": [ + 181, + 91, + 448, + 101 + ], + "type": "text", + "content": "Realistic Human Motion Generation with Cross-Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "joint velocities. Notably, the rotation representation is made up of the sine and cosine values of the angle." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 156, + 313, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 156, + 313, + 168 + ], + "spans": [ + { + "bbox": [ + 131, + 156, + 313, + 168 + ], + "type": "text", + "content": "3.3 Cross Motion Diffusion Model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 175, + 480, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 175, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 480, + 270 + ], + "type": "text", + "content": "**Framework.** Our pipeline is illustrated in Figure 2. CLIP [28] is a widely recognized text encoder, and we use it to encode the text prompt " + }, + { + "bbox": [ + 130, + 175, + 480, + 270 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 175, + 480, + 270 + ], + "type": "text", + "content": ". The encoded text feature and time-step " + }, + { + "bbox": [ + 130, + 175, + 480, + 270 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 175, + 480, + 270 + ], + "type": "text", + "content": " are projected into transformer dimension and summed together as the condition token " + }, + { + "bbox": [ + 130, + 175, + 480, + 270 + ], + "type": "inline_equation", + "content": "z_{tk}" + }, + { + "bbox": [ + 130, + 175, + 480, + 270 + ], + "type": "text", + "content": ". The 2D and 3D motion sequences are projected into the same dimension, concatenated with condition token " + }, + { + "bbox": [ + 130, + 175, + 480, + 270 + ], + "type": "inline_equation", + "content": "z_{tk}" + }, + { + "bbox": [ + 130, + 175, + 480, + 270 + ], + "type": "text", + "content": " in time axis and summed with a standard positional embedding. We aim to unify the two domain features in one space but it is too difficult for one linear layer. A straightforward idea is to encode via two separate encoders:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 211, + 277, + 481, + 292 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 277, + 481, + 292 + ], + "spans": [ + { + "bbox": [ + 211, + 277, + 481, + 292 + ], + "type": "interline_equation", + "content": "z _ {3 D} ^ {0} = \\mathcal {E} _ {3 D} \\left(x _ {3 D, t}, t, c\\right), z _ {2 D} ^ {0} = \\mathcal {E} _ {2 D} \\left(x _ {2 D, t}, t, c\\right), \\tag {3}", + "image_path": "b67e21dc8a8e5df6e6c95bc2a609d9294c0cb1ea07dc35e8168e88b330bafbf7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 298, + 479, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 479, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 479, + 332 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 298, + 479, + 332 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{3D / 2D}(\\cdot)" + }, + { + "bbox": [ + 130, + 298, + 479, + 332 + ], + "type": "text", + "content": " are 3D/2D " + }, + { + "bbox": [ + 130, + 298, + 479, + 332 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 130, + 298, + 479, + 332 + ], + "type": "text", + "content": "-layer transformer encoders [44]. However, we find it more efficient to add another shared-weight encoder to extract shared feature:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 240, + 332, + 480, + 349 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 332, + 480, + 349 + ], + "spans": [ + { + "bbox": [ + 240, + 332, + 480, + 349 + ], + "type": "interline_equation", + "content": "\\left\\{z _ {3 D / 2 D} ^ {i} \\right\\} _ {i = 1} ^ {L _ {2}} = \\mathcal {E} _ {\\text {s h a r e}} \\left(z _ {3 D / 2 D} ^ {0}\\right), \\tag {4}", + "image_path": "b879b25f5f329fff63c8dc496c243a483f1d5306548c490b31c017504f494941.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 354, + 479, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 479, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 479, + 392 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 354, + 479, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\text{share}}(\\cdot)" + }, + { + "bbox": [ + 130, + 354, + 479, + 392 + ], + "type": "text", + "content": " is a shared-weight " + }, + { + "bbox": [ + 130, + 354, + 479, + 392 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 130, + 354, + 479, + 392 + ], + "type": "text", + "content": "-layer transformer encoder, and " + }, + { + "bbox": [ + 130, + 354, + 479, + 392 + ], + "type": "inline_equation", + "content": "\\{z_{3D/2D}^{i}\\}_{i=1}^{L_{2}}" + }, + { + "bbox": [ + 130, + 354, + 479, + 392 + ], + "type": "text", + "content": " are the outputs of each shared-weight layer. The whole process is defined as unified encoding." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 393, + 480, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 393, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 393, + 480, + 489 + ], + "type": "text", + "content": "To output motion in two modality, we use independent " + }, + { + "bbox": [ + 130, + 393, + 480, + 489 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 130, + 393, + 480, + 489 + ], + "type": "text", + "content": "-layer transformer decoders [44] for 2D and 3D data. Starting from 2D/3D learnable token embeddings " + }, + { + "bbox": [ + 130, + 393, + 480, + 489 + ], + "type": "inline_equation", + "content": "Tok_{2D/3D}" + }, + { + "bbox": [ + 130, + 393, + 480, + 489 + ], + "type": "text", + "content": ", each decoder layer takes the output of the previous layer as queries and the output of same-level shared layer " + }, + { + "bbox": [ + 130, + 393, + 480, + 489 + ], + "type": "inline_equation", + "content": "z^{i}" + }, + { + "bbox": [ + 130, + 393, + 480, + 489 + ], + "type": "text", + "content": " as keys and values instead of the last layer. The starting point is to make decoder layers follow the extracting patterns of shared-weight layers rather than gaining deeper embeddings. Finally, a linear layer is added to map the features to the dimensions of the motions. This cross-decoding can be integrated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 186, + 496, + 480, + 511 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 496, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 186, + 496, + 480, + 511 + ], + "type": "interline_equation", + "content": "\\hat {x} _ {3 D, 0} = \\mathcal {D} _ {3 D} \\left(\\left\\{z _ {3 D / 2 D} ^ {i} \\right\\} _ {i = 1} ^ {L _ {2}}\\right), \\hat {x} _ {2 D, 0} = \\mathcal {D} _ {2 D} \\left(\\left\\{z _ {3 D / 2 D} ^ {i} \\right\\} _ {i = 1} ^ {L _ {2}}\\right), \\tag {5}", + "image_path": "5db1c2e26d59024b10cd9004aefc9f70d774925e15decb29c35dc5bdadb03d05.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 517, + 479, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 479, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 479, + 552 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 517, + 479, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{3D / 2D}" + }, + { + "bbox": [ + 130, + 517, + 479, + 552 + ], + "type": "text", + "content": " are 3D/2D decoders including learnable tokens; and " + }, + { + "bbox": [ + 130, + 517, + 479, + 552 + ], + "type": "inline_equation", + "content": "\\hat{x}_{3D / 2D,0}" + }, + { + "bbox": [ + 130, + 517, + 479, + 552 + ], + "type": "text", + "content": " are predicted clean 3D/2D motion sequences. In summary, with a pair of 3D motion and 2D projected motion, CrossDiff outputs four results via" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 170, + 561, + 480, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 561, + 480, + 574 + ], + "spans": [ + { + "bbox": [ + 170, + 561, + 480, + 574 + ], + "type": "interline_equation", + "content": "\\hat {x} _ {i D \\rightarrow j D, 0} = G _ {i D \\rightarrow j D} \\left(x _ {i D, t}, t, c\\right) = \\mathcal {D} _ {j D} \\left(\\mathcal {E} _ {\\text {s h a r e}} \\left(\\mathcal {E} _ {i D} \\left(x _ {i D, t}, t, c\\right)\\right)\\right), \\tag {6}", + "image_path": "c07e702bb558311906afd7336566a779e919908264b0e105832ed692ca6320c2.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 581, + 479, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 479, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 479, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 581, + 479, + 605 + ], + "type": "inline_equation", + "content": "\\hat{x}_{iD\\to jD,0}" + }, + { + "bbox": [ + 130, + 581, + 479, + 605 + ], + "type": "text", + "content": " are predicted " + }, + { + "bbox": [ + 130, + 581, + 479, + 605 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 581, + 479, + 605 + ], + "type": "text", + "content": " -dimension clean motion " + }, + { + "bbox": [ + 130, + 581, + 479, + 605 + ], + "type": "inline_equation", + "content": "\\hat{x}_{jD,0}" + }, + { + "bbox": [ + 130, + 581, + 479, + 605 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 581, + 479, + 605 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 581, + 479, + 605 + ], + "type": "text", + "content": " -dimension motion noise " + }, + { + "bbox": [ + 130, + 581, + 479, + 605 + ], + "type": "inline_equation", + "content": "x_{iD,t}" + }, + { + "bbox": [ + 130, + 581, + 479, + 605 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 581, + 479, + 605 + ], + "type": "inline_equation", + "content": "i,j\\in \\{2,3\\}" + }, + { + "bbox": [ + 130, + 581, + 479, + 605 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 621, + 479, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 621, + 479, + 646 + ], + "spans": [ + { + "bbox": [ + 130, + 621, + 479, + 646 + ], + "type": "text", + "content": "Training. As mentioned in Section 3.1, we apply a simple objective (Eq. 2) for all outputs:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 198, + 652, + 480, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 652, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 198, + 652, + 480, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i D \\rightarrow j D} = \\mathbb {E} _ {t \\sim [ 1, T ]} \\| x _ {j D, 0} - G _ {i D \\rightarrow j D} (x _ {i D, t}, t, c) \\| _ {2} ^ {2}. \\tag {7}", + "image_path": "29888cf1dcb4bde10102855a0a63243c20bfec3d4aa3de3df08d9ec140378d43.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "Z. Ren et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "We train our model in two stage. In stage I, CrossDiff is forced to learn the reverse process, motion characteristic of texts in both domains and the motion connection of two distribution via the loss" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 165, + 162, + 481, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 162, + 481, + 175 + ], + "spans": [ + { + "bbox": [ + 165, + 162, + 481, + 175 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s t a g e I}} = \\mathcal {L} _ {3 D \\rightarrow 3 D} + w _ {2 3} \\mathcal {L} _ {2 D \\rightarrow 3 D} + w _ {3 2} \\mathcal {L} _ {3 D \\rightarrow 2 D} + w _ {2 2} \\mathcal {L} _ {2 D \\rightarrow 2 D}, \\tag {8}", + "image_path": "15768ea5b9b5d9142ee45dacc716cf597bd463f56129197648d507dd4ddad121.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 184, + 479, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 184, + 479, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 184, + 479, + 205 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 184, + 479, + 205 + ], + "type": "inline_equation", + "content": "w_{23}, w_{32}, w_{22}" + }, + { + "bbox": [ + 130, + 184, + 479, + 205 + ], + "type": "text", + "content": " are relative weights. In stage II, there is only a 3D generation loss:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 261, + 208, + 481, + 221 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 208, + 481, + 221 + ], + "spans": [ + { + "bbox": [ + 261, + 208, + 481, + 221 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s t a g e I I}} = \\mathcal {L} _ {3 D \\rightarrow 3 D}. \\tag {9}", + "image_path": "11dd19826070468ad7ffcf33f09225cacc5d3538e80864462774d2aca1e7b6d0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 226, + 479, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 479, + 261 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 479, + 261 + ], + "type": "text", + "content": "This helps the model focus on the 3D denoising process and eliminate the uncertainty of the 2D and 3D mapping relationship while retaining the knowledge of diverse motion features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 279, + 253, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 279, + 253, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 279, + 253, + 292 + ], + "type": "text", + "content": "3.4 Mixture Sampling" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 134, + 320, + 477, + 373 + ], + "blocks": [ + { + "bbox": [ + 134, + 320, + 477, + 373 + ], + "lines": [ + { + "bbox": [ + 134, + 320, + 477, + 373 + ], + "spans": [ + { + "bbox": [ + 134, + 320, + 477, + 373 + ], + "type": "image", + "image_path": "c881a86b105e53ff8bfc2fd03eaa5271b7b38f105f1c4c97cd08023400335077.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 392, + 480, + 437 + ], + "lines": [ + { + "bbox": [ + 130, + 392, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 480, + 437 + ], + "type": "text", + "content": "Fig. 3: Overview of Mixture Sampling. The original noise is sampled from a 2D gaussian distribution. From time-step " + }, + { + "bbox": [ + 130, + 392, + 480, + 437 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 392, + 480, + 437 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 392, + 480, + 437 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 392, + 480, + 437 + ], + "type": "text", + "content": ", CrossDiff predicts the clean 2D motion " + }, + { + "bbox": [ + 130, + 392, + 480, + 437 + ], + "type": "inline_equation", + "content": "\\hat{x}_{2D,0}" + }, + { + "bbox": [ + 130, + 392, + 480, + 437 + ], + "type": "text", + "content": " and diffuses it back to " + }, + { + "bbox": [ + 130, + 392, + 480, + 437 + ], + "type": "inline_equation", + "content": "x_{2D,t-1}" + }, + { + "bbox": [ + 130, + 392, + 480, + 437 + ], + "type": "text", + "content": ". In the remaining " + }, + { + "bbox": [ + 130, + 392, + 480, + 437 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 392, + 480, + 437 + ], + "type": "text", + "content": " steps, CrossDiff denoises in the 3D domain and finally obtains the clean 3D motion." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 460, + 479, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 479, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 479, + 508 + ], + "type": "text", + "content": "After training, one can sample a motion sequence conditioned on a text prompt in an iterative manner. The standard method [40] gradually anneals the 3D noise from a gaussian distribution, which we still use. We predict the clean sample " + }, + { + "bbox": [ + 130, + 460, + 479, + 508 + ], + "type": "inline_equation", + "content": "\\hat{x}_{3D,0}" + }, + { + "bbox": [ + 130, + 460, + 479, + 508 + ], + "type": "text", + "content": " and noise it back to " + }, + { + "bbox": [ + 130, + 460, + 479, + 508 + ], + "type": "inline_equation", + "content": "x_{3D,t-1}" + }, + { + "bbox": [ + 130, + 460, + 479, + 508 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 460, + 479, + 508 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 460, + 479, + 508 + ], + "type": "text", + "content": " steps until " + }, + { + "bbox": [ + 130, + 460, + 479, + 508 + ], + "type": "inline_equation", + "content": "x_{3D,0}" + }, + { + "bbox": [ + 130, + 460, + 479, + 508 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 509, + 480, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 480, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 480, + 591 + ], + "type": "text", + "content": "Furthermore, utilizing the CrossDiff architecture, we propose a novel two-domain sampling approach. As shown in Figure 3, We first sample 2D gaussian noise which is then denoised with " + }, + { + "bbox": [ + 130, + 509, + 480, + 591 + ], + "type": "inline_equation", + "content": "G_{2D\\to 2D}(x_{2D,t},t,c)" + }, + { + "bbox": [ + 130, + 509, + 480, + 591 + ], + "type": "text", + "content": " until time-step " + }, + { + "bbox": [ + 130, + 509, + 480, + 591 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 509, + 480, + 591 + ], + "type": "text", + "content": ". Next, we project the denoised 2D noise onto the 3D domain using " + }, + { + "bbox": [ + 130, + 509, + 480, + 591 + ], + "type": "inline_equation", + "content": "G_{2D\\to 3D}(x_{2D,t},t,c)" + }, + { + "bbox": [ + 130, + 509, + 480, + 591 + ], + "type": "text", + "content": " and continue the denoising process with " + }, + { + "bbox": [ + 130, + 509, + 480, + 591 + ], + "type": "inline_equation", + "content": "G_{3D\\to 3D}(x_{3D,t},t,c)" + }, + { + "bbox": [ + 130, + 509, + 480, + 591 + ], + "type": "text", + "content": " for the remaining " + }, + { + "bbox": [ + 130, + 509, + 480, + 591 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 509, + 480, + 591 + ], + "type": "text", + "content": " steps. Our experiments in supplementary demonstrate the difference between mixture sampling and the vanilla method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 609, + 397, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 609, + 397, + 621 + ], + "spans": [ + { + "bbox": [ + 130, + 609, + 397, + 621 + ], + "type": "text", + "content": "3.5 Learning 3D Motion Generation from 2D Data" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": "Given the complexity and cost associated with collecting high-quality 3D motion data, generating 3D motion from 2D motion data is an attractive alternative. Moreover, generating 3D motion from textual descriptions in an out-of-domain" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 181, + 91, + 448, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 91, + 448, + 101 + ], + "spans": [ + { + "bbox": [ + 181, + 91, + 448, + 101 + ], + "type": "text", + "content": "Realistic Human Motion Generation with Cross-Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "scenario is approximately a zero-shot task. To achieve this, we first estimated 2D motion from text-related videos using an off-the-shelf model. We then utilized the pretrained model in stage I " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "G_{2D \\to 3D}(x_{2D,t}, t, c)" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " to generate corresponding 3D clean motion, with " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " set to 0 and " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " set to null condition " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\emptyset" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": ". A motion filter is applied to smooth the generated motion. We assume that 2D pose estimation is relatively precise, allowing the processed 2D/3D motion to serve as pseudolabels for training. The model is fine-tuned with the same objective as stage I, but with different weight hyper-parameters. After training, our model can generate diverse motion according to out-of-domain textual descriptions using mixture sampling (Sec. 3.4)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 482, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 285 + ], + "type": "text", + "content": "During training with 2D motion estimated from videos, we encounter significant errors in root estimation due to the uncertainty of camera movement. To address this issue, we decouple the root information " + }, + { + "bbox": [ + 130, + 236, + 482, + 285 + ], + "type": "inline_equation", + "content": "r_{3D/2D}" + }, + { + "bbox": [ + 130, + 236, + 482, + 285 + ], + "type": "text", + "content": " and generate it based on other pose features. Please refer to supplementary for more details." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 306, + 230, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 306, + 230, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 306, + 230, + 320 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 334, + 482, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 482, + 466 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 482, + 466 + ], + "type": "text", + "content": "Our focus is on the text-to-motion generation task, and we conduct experiments on two standard datasets: HumanML3D [8] and KIT Motion-Language (KIT-ML) [27]. In Section 4.1, we introduce these standard datasets and the evaluation metrics used in our experiments. In Section 4.2, we present comparable quantitative results with state-of-the-art methods and outperform diffused methods on the HumanML3D dataset. We also use upper and lower body indices, combined with visualization effects, to illustrate our superior performance in capturing whole-body details. Moreover, we demonstrate that CrossDiff supports training with additional 2D motion sequences in Section 4.3. Finally, we discuss ablation studies in Section 4.4 to analyze the contributions of each component of our proposed method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 487, + 326, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 326, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 326, + 498 + ], + "type": "text", + "content": "4.1 Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": "Datasets. The HumanML3D [8] and KIT-ML datasets [27] are widely used in research. KIT-ML provides 6,353 textual descriptions for 3,911 motion sequences, which are all down-sampled to 12.5 FPS. HumanML3D is currently the largest 3D human motion dataset with textual descriptions. The motions are originally from two motion capture datasets, AMASS [22] and HumanAct12 [10], and are rescaled to 20 frames per second. It contains 14,616 motions annotated with 44,970 sequence-level textual descriptions. To enable fair comparisons with previous works [4,8,40], we use the redundant motion representation proposed by [8]. This representation involves re-targeting the joint positions to a default human skeleton template, setting the initial pose at the same position " + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathrm{X} = 0,\\mathrm{Z} = 0)" + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": " facing the " + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}+" + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": " direction, and including root global positions, local positions, joint rotations, joint velocities, and foot contact labels. We use the same motion representation for KIT-ML." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Z. Ren et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "To obtain corresponding 2D motion, we utilize orthogonal projection to project 3D motion into 2D planes for four directions (front, left, right, and back). Then, we process the redundant 2D motion representation as described in Section 3.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 163, + 480, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 163, + 480, + 234 + ], + "spans": [ + { + "bbox": [ + 130, + 163, + 480, + 234 + ], + "type": "text", + "content": "UFC101 [38] is an influential action recognition dataset that contains paired videos and action labels. We obtain 2D joint positions using the ViTPose [48] model and process them into redundant 2D motion representation. We filter out fuzzy data and ultimately select 24 action labels, with each label related to 10-50 motion sequences. We then annotate around five textual descriptions for each label." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 252, + 480, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 252, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 130, + 252, + 480, + 336 + ], + "type": "text", + "content": "Evaluation Metrics. We compare our results with previous works using the same metrics as [8, 40]. These metrics involve evaluating quality with Frechet Inception Distance (FID), precision with R-Precision and Multi-modal Distance (MM Dist), and diversity with Diversity (DIV) and Multimodality (MModality). These measurements assess generated motion distribution, retrieval ranking accuracy, Euclidean distances between text and motion features, and variance across features and within a single text." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 337, + 481, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 337, + 481, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 337, + 481, + 443 + ], + "type": "text", + "content": "Besides using the evaluation model from [8], we introduce a new metric measuring the FID (Fréchet Inception Distance) of upper and lower body movements, denoted as FID-U and FID-L. This enables a fine-grained analysis of human motion and better comprehension of upper and lower body dynamics. We split joints into two groups using the root joint as a boundary and train separate evaluators, following a similar approach to [8]. This effectively evaluates generated motion quality for both body segments, offering a deeper understanding of human motion complexities and advancing research on new motion generation models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 461, + 480, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 480, + 533 + ], + "type": "text", + "content": "Implementation Details. We use the AdamW [19] optimizer, setting the learning rate to 1e-4 and 1e-5 in stages I and II, respectively. Our model is trained for 4k epochs in stage I and 1k epochs in stage II, using 8 Tesla V100 GPUs with a mini-batch size of 32. The number of diffusion steps is set to 1K. The loss weights " + }, + { + "bbox": [ + 130, + 461, + 480, + 533 + ], + "type": "inline_equation", + "content": "(w_{23}, w_{32}, w_{33})" + }, + { + "bbox": [ + 130, + 461, + 480, + 533 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 130, + 461, + 480, + 533 + ], + "type": "inline_equation", + "content": "(1, 1, 1)" + }, + { + "bbox": [ + 130, + 461, + 480, + 533 + ], + "type": "text", + "content": " in stage I and " + }, + { + "bbox": [ + 130, + 461, + 480, + 533 + ], + "type": "inline_equation", + "content": "(0.1, 0.1, 1)" + }, + { + "bbox": [ + 130, + 461, + 480, + 533 + ], + "type": "text", + "content": " in training on UFC101 datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 551, + 320, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 551, + 320, + 562 + ], + "spans": [ + { + "bbox": [ + 131, + 551, + 320, + 562 + ], + "type": "text", + "content": "4.2 Comparisons on Text-to-motion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": "Comparative Analysis of Standard Metrics. In our evaluation, we test our models 20 times and compare their performance with existing state-of-the-art methods. These methods include Language2Pose [1], T2G [2], Hier [6], T2M [8], MotionDiffuse [53], Fg-T2M [45], MDM [40], T2M-GPT [52], MLD [4] and ReMoDiffuse [54]. As illustrated in Table 1, our model exhibits competitive performance when compared to these leading methods. However, it is important to note that the KIT-ML dataset primarily consists of \"walk\" movements and lacks intricate details. Consequently, this dataset does not present the same challenges" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 181, + 91, + 448, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 91, + 448, + 101 + ], + "spans": [ + { + "bbox": [ + 181, + 91, + 448, + 101 + ], + "type": "text", + "content": "Realistic Human Motion Generation with Cross-Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 178, + 482, + 276 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 1: Quantitative results on the HumanML3D and KIT-ML test set. The overall results on KIT-ML are shown on the right, while the results of both widely-used and newly-proposed metrics on HumanML3D are shown on the left. The red and blue colors indicate the best and second-best results, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 178, + 482, + 276 + ], + "lines": [ + { + "bbox": [ + 133, + 178, + 482, + 276 + ], + "spans": [ + { + "bbox": [ + 133, + 178, + 482, + 276 + ], + "type": "table", + "html": "
MethodsHumanML3DKIT-ML
R Precision (top 3)↑FID↓MM Dist↓DIV→ MModality↑FID-U↓FID-L↓R Precision (top 3)↑FID↓MM Dist↓DIV→ MModality↑FID-U↓FID-L↓
Real0.7970.0022.9749.503--0.7790.0312.78811.08--
Language2Pose0.48611.025.2967.676--0.4836.5455.1479.073--
T2G0.3457.6646.0306.409--0.33812.126.9649.334--
Hier0.5526.5325.0128.332--0.5315.2034.9869.563--
T2M0.7401.0673.3409.1882.090-0.6932.7703.40110.911.482-
MotionDiffuse0.7820.6303.1139.4101.553-0.7391.9542.95811.100.730-
Fg-T2M0.7830.2433.1099.2781.614-0.7450.5713.11410.931.019-
MDM0.6110.5445.5669.5592.7990.8250.4840.4979.19110.8471.9070.925
T2M-GPT0.7750.1413.1219.7221.8310.1450.6070.7450.5143.0071.5700.602
MLD0.7720.4733.1969.7242.4130.5410.5530.7340.4043.2042.1920.563
ReMoDiffuse0.7950.1032.9749.0181.7950.1250.5650.7650.1552.8141.8080.205
Ours0.7300.1623.3589.5772.6200.1180.2810.7040.4743.30810.770.434
", + "image_path": "6864ab268ad4ca97c37472495468440de21e4c3191692a607a12e773c7c9db9d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 305, + 482, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 305, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 305, + 482, + 354 + ], + "type": "text", + "content": "that our method is specifically designed to address. It indicates we can steadily generate high-quality and precise motion while pay attention to rich diversity. In other words, our generated results are not only consistent with the textual description, but also more expressive." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 369, + 482, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 488 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 488 + ], + "type": "text", + "content": "Comparative Analysis of Fine-grained Metrics. We compare the fine-grained metrics for our upper and lower body with those from four recent studies [4, 40, 52, 54]. As demonstrated in Table 1, our generated motion is more robust and detailed. Our low FID scores for both the upper and lower body indicate that our synthesized motion effectively captures full-body movement rather than focusing solely on specific semantic parts. In contrast, ReMoDiffuse and T2M-GPT achieves a low FID score for the upper body but a high score for the lower body. This suggests that their generation process exhibits unbalanced attention towards different body parts, primarily translating textual descriptions into upper body characteristics rather than capturing the entire body's motion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 489, + 482, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 489, + 482, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 489, + 482, + 608 + ], + "type": "text", + "content": "Figure 4 displays qualitative comparisons with existing methods. Our method can \"march\" with arm swings, \"wobble\" using hands for balancing and alternate between defense and attack in a \"fight\". MDM exhibits a certain rigidity, while T2M-GPT and ReMoDiffuse appear to lack dynamism. For direct comparisons, please refer to the supplementary videos provided. We conducted a user study on motion performance, in which participants were asked two questions to assess the vitality and diversity of the motions. The results, presented in Figure 5(a), confirm our analysis. In summary, our method demonstrates a superior ability to interpret semantic information and generate more accurate and expressive motions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 624, + 280, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 280, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 280, + 635 + ], + "type": "text", + "content": "4.3 Learning from 2D Data" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "After being trained on a 3D dataset, our model can learn 3D motion generation from 2D data. By fine-tuning the model with the UCF101 dataset [38], we ef" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Z. Ren et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 479, + 352 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 479, + 352 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 479, + 352 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 479, + 352 + ], + "type": "image", + "image_path": "a8f5cabae1da8913966082f8e909fe530cc00beab367e9af967ff3a758e0a8fb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 372, + 482, + 407 + ], + "lines": [ + { + "bbox": [ + 130, + 372, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 372, + 482, + 407 + ], + "type": "text", + "content": "Fig. 4: Qualitative results on HumanML3D dataset. We compare our method with MDM [40], T2M-GPT [52] and MLD [4]. We find that our generated actions better convey the intended semantics." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 427, + 482, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 427, + 482, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 427, + 482, + 608 + ], + "type": "text", + "content": "fectively address a zero-shot problem arising from the absence of ground-truth 3D motion. Our sampling strategy reaches optimal performance when " + }, + { + "bbox": [ + 130, + 427, + 482, + 608 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 130, + 427, + 482, + 608 + ], + "type": "text", + "content": ". As depicted in Figure 6, the generated motions for various activities, such as pulling up, biking, table tennis, and baseball, are showcased alongside their textual prompts. Notably, pulling up and biking are entirely outside the HumanML3D motion domain. Although playing table tennis and baseball share similarities with actions like slapping or throwing in HumanML3D, the original model is unable to synthesize motion under those specific descriptions. However, after fine-tuning, our generated motion more closely resembles the given sports rather than merely mimicking the actions. Despite some activities being beyond the scope of the HumanML3D domain, our fine-tuned model successfully synthesizes specific motions by leveraging the weak 2D data. This demonstrates its remarkable adaptability and potential for efficient use of available motion data since 2D motion and related textual descriptions are easier to obtain than 3D motion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 623, + 246, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 623, + 246, + 634 + ], + "spans": [ + { + "bbox": [ + 132, + 623, + 246, + 634 + ], + "type": "text", + "content": "4.4 Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Our model features separate pipelines for 2D and 3D inputs, allowing us to train solely on 3D motion sequences, which is an improvement over MDM [40]." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 181, + 91, + 448, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 91, + 448, + 101 + ], + "spans": [ + { + "bbox": [ + 181, + 91, + 448, + 101 + ], + "type": "text", + "content": "Realistic Human Motion Generation with Cross-Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 118, + 249, + 208 + ], + "blocks": [ + { + "bbox": [ + 133, + 118, + 249, + 208 + ], + "lines": [ + { + "bbox": [ + 133, + 118, + 249, + 208 + ], + "spans": [ + { + "bbox": [ + 133, + 118, + 249, + 208 + ], + "type": "image", + "image_path": "917d6d57989f0277642179d6280abda59d4250989c1a571a9700021312078382.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 225, + 482, + 282 + ], + "lines": [ + { + "bbox": [ + 130, + 225, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 482, + 282 + ], + "type": "text", + "content": "Fig. 5: (a) The result of the user study. (b) Difference between 3D and 2D motion data distribution. The time axis is represented on the x-axis, while the normalized joint velocity is represented on the y-axis. The 3D motion is represented by a blue full line, while the 2D motion is represented by red and green dashed lines, indicating the front and left view, respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 274, + 113, + 482, + 209 + ], + "blocks": [ + { + "bbox": [ + 274, + 113, + 482, + 209 + ], + "lines": [ + { + "bbox": [ + 274, + 113, + 482, + 209 + ], + "spans": [ + { + "bbox": [ + 274, + 113, + 482, + 209 + ], + "type": "image", + "image_path": "65ff17a8c208517b3e98ced8ff77c8ea391c8c655c106f9d9891d633d4b7ae30.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 290, + 482, + 417 + ], + "blocks": [ + { + "bbox": [ + 133, + 290, + 482, + 417 + ], + "lines": [ + { + "bbox": [ + 133, + 290, + 482, + 417 + ], + "spans": [ + { + "bbox": [ + 133, + 290, + 482, + 417 + ], + "type": "image", + "image_path": "b7b6c6c3cba1bacdeb542dfbcdb1adfceabd06e88f8f1dbd76282b3a486a8ef6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 435, + 482, + 460 + ], + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 460 + ], + "type": "text", + "content": "Fig. 6: Generating 3D movements without training on paired 3D motion and textual descriptions." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 481, + 482, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 482, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 482, + 507 + ], + "type": "text", + "content": "We investigate the impact of introducing 2D data on the performance of 3D generation and demonstrate the effectiveness of using a shared-weights encoder." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Why 2D motion help? To explain the benefits of 2D motions, we compared the distribution differences between 3D and 2D motion data. Hand and feet movements, which are primary indicators of motion, were visualized in both 3D and 2D levels, and their velocities were normalized along the joints dimension. In Figure 5(b), we can clearly see that around the 20th frame, the 2D velocity of the left hand in the front view reaches a higher value while the 3D velocity is quite low, indicating that hand movements in 2D are more prominent than in 3D. The results show that 2D motion captures different details from 3D motion, suggesting that the CrossDiff model can lead 3D motion to learn from the knowledge that 2D motion acquired from text prompts. Specifically, for the given sample, 2D motion might learn \"animated hand motions\" while 3D motion focuses only on \"walking\". 2D motion is an explicit feature that we artificially extract to" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Z. Ren et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 173, + 178, + 441, + 340 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": "Table 2: Evaluation of our models with different settings on the HumanML3D dataset. Bold indicates best result. The symbol % indicates the percentage of data being used. From top to bottom, we present MDM as baselines, the impact of training with 2D representations, with(w/) or without(w/o) shared-weights encoder." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 173, + 178, + 441, + 340 + ], + "lines": [ + { + "bbox": [ + 173, + 178, + 441, + 340 + ], + "spans": [ + { + "bbox": [ + 173, + 178, + 441, + 340 + ], + "type": "table", + "html": "
MethodsR Precision (top 3)↑FID↓MM Dist↓DIV→
MDM0.6110.5445.5669.559
Ours0.7300.1623.3589.577
50% 3D0.6660.5863.8949.513
100% 3D0.6850.2243.6909.445
50% 3D + 100% 2D0.6720.4223.7089.345
100% 3D + 100% 2D0.7300.1623.3589.577
w/o shared-weights encoder0.7140.1873.4969.488
w/ shared-weights encoder0.7300.1623.3589.577
1 view(front)0.7220.1863.4679.798
1 view(left)0.7150.1813.4129.834
4 views0.7300.1623.3589.577
5 views0.6950.2023.6139.502
", + "image_path": "4faea49aa7cb6dddfce3225448f9e48c3a39d5292cde7752bacb0c88305d7180.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 373, + 480, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 480, + 398 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 480, + 398 + ], + "type": "text", + "content": "aid the model's learning, and this approach can help improve performance when dealing with arbitrary data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 418, + 482, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 550 + ], + "type": "text", + "content": "Influence of 2D Representation Table 2 presents the results from four different experiment settings. The control groups of \"100% 3D\" and \"100% 3D + 100% 2D\" demonstrate that when training with paired 3D motion and text, projecting the 3D motion to 2D and building a connection between the 2D motion and text can help boost performance. The visualizations in Figure 1 further highlight the enhanced quality of our generated outputs. The control groups of \"50% 3D\" and \"50% 3D + 100% 2D\" prove that additional 2D data can also help improve performance. The additional 2D data indicates other 2D motion without ground truth 3D motion. The experiment in Section 4.3 shows learning 2D motion in the wild can also help with out-of-domain 3D motion learning. As we can see, the combined learning of 2D motion has great potential." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Shared-weights Encoder Without the shared-weights encoder, the model is a simple encoder-decoder framework with two modalities. However, we believe that this is not sufficient to fully fuse the 3D and 2D motion features. Inspired by [47], we found that when learning with data from two modalities, extracting separate and fused feature layer-by-layer is more efficient. The shared-weights encoder serves as a fusing module, while the 3D/2D motion decoder acts as a separate decoder module. The goal is to ensure that the decoder layers follow the same extraction patterns as the shared-weight layers, rather than simply gaining" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 181, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 181, + 91, + 447, + 101 + ], + "type": "text", + "content": "Realistic Human Motion Generation with Cross-Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 140 + ], + "type": "text", + "content": "deeper embeddings. The results presented in Table 2 demonstrate the efficiency of using a shared-weights encoder." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 156, + 482, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 156, + 482, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 156, + 482, + 348 + ], + "type": "text", + "content": "2D Representation from different views To evaluate the impact of different views, we conducted tests on four settings: a) only the front view, b) only the left view, c) four views including front, left, back, and right, and d) five views with an additional top view. When multiple views were used, the 3D motion was paired with a random view projection to formulate the loss. As shown in Table 2, it is reasonable to assume that four views contain more 2D information and outperform one view. Furthermore, the front view and left view do not have much distinction in terms of performance. However, the performance actually decreased with the additional top view. This could be due to the fact that the 2D information becomes more difficult to classify without the condition of the camera view. We did not pass along the camera information because it is difficult to estimate the camera view of in-the-wild videos, and injecting camera information could disrupt the structure of the text-to-2D model, making it difficult for the text-to-3D model to follow. In conclusion, the number of views serves as a practical hyperparameter that can be adjusted through enumeration experiments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 363, + 220, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 220, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 220, + 376 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 386, + 482, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 481 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 481 + ], + "type": "text", + "content": "In conclusion, the Cross Human Motion Diffusion Model (CrossDiff) presents a promising advancement in the field of human motion synthesis by effectively integrating and leveraging both 3D and 2D motion information for high-quality motion generation. The unified encoding and cross-decoding components of the CrossDiff learning framework enable the capture of intricate details of human movement that are often overlooked by models relying solely on 3D data. Our experiments validate the superior performance of the proposed model on various text-to-motion benchmarks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 482, + 482, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 482, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 482, + 482, + 553 + ], + "type": "text", + "content": "Despite its promising results, the CrossDiff model is not without potential weaknesses. One limitation may arise from the model's ability to generalize to unseen or rare motion patterns, especially when trained only on 2D motion data. Additionally, the computational complexity of the model might hinder its real-time application in certain scenarios, such as interactive gaming and virtual reality environments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 554, + 482, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 554, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 554, + 482, + 615 + ], + "type": "text", + "content": "Future work could explore methods to enhance the model's generalization capabilities, such as incorporating unsupervised or semi-supervised learning techniques. To further advance our understanding, we propose the accumulation of a sizable 2D motion dataset coupled with relevant textual prompts, enabling the training of a unified motion generation model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Acknowledgments. This research was partly supported by Shenzhen Key Laboratory of next generation interactive media innovative technology (Grant No: ZDSYS20210623092001004)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Z. Ren et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 137, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 136, + 137, + 481, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 137, + 481, + 169 + ], + "spans": [ + { + "bbox": [ + 136, + 137, + 481, + 169 + ], + "type": "text", + "content": "1. Ahuja, C., Morency, L.P.: Language2pose: Natural language grounded pose forecasting. In: 2019 International Conference on 3D Vision (3DV). pp. 719-728. IEEE (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 170, + 481, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 170, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 136, + 170, + 481, + 213 + ], + "type": "text", + "content": "2. Bhattacharya, U., Rewkowski, N., Banerjee, A., Guhan, P., Bera, A., Manocha, D.: Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In: 2021 IEEE virtual reality and 3D user interfaces (VR). pp. 1-10. IEEE (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 213, + 481, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 213, + 481, + 245 + ], + "spans": [ + { + "bbox": [ + 137, + 213, + 481, + 245 + ], + "type": "text", + "content": "3. Chen, L.H., Zhang, J., Li, Y., Pang, Y., Xia, X., Liu, T.: Humanmac: Masked motion completion for human motion prediction. arXiv preprint arXiv:2302.03665 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 246, + 481, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 481, + 277 + ], + "type": "text", + "content": "4. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18000-18010 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 278, + 481, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 278, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 278, + 481, + 320 + ], + "type": "text", + "content": "5. Chung, H., Sim, B., Ye, J.C.: Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12413-12422 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 321, + 481, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 481, + 353 + ], + "type": "text", + "content": "6. Ghosh, A., Cheema, N., Oguz, C., Theobalt, C., Slusallek, P.: Synthesis of compositional animations from textual descriptions. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1396-1406 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 354, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 481, + 385 + ], + "type": "text", + "content": "7. Ghosh, P., Song, J., Aksan, E., Hilliges, O.: Learning human motion models for long-term predictions. In: 2017 International Conference on 3D Vision (3DV). pp. 458-466. IEEE (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 386, + 481, + 418 + ], + "type": "text", + "content": "8. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5152-5161 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 418, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 481, + 450 + ], + "type": "text", + "content": "9. Guo, C., Zuo, X., Wang, S., Cheng, L.: Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In: European Conference on Computer Vision. pp. 580-597. Springer (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 450, + 481, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 450, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 134, + 450, + 481, + 483 + ], + "type": "text", + "content": "10. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 2021-2029 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 483, + 481, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 483, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 134, + 483, + 481, + 503 + ], + "type": "text", + "content": "11. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 504, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 504, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 134, + 504, + 481, + 525 + ], + "type": "text", + "content": "12. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 525, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 525, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 134, + 525, + 481, + 557 + ], + "type": "text", + "content": "13. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Gmd: Controllable human motion synthesis via guided diffusion models. arXiv preprint arXiv:2305.12577 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 558, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 558, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 134, + 558, + 481, + 578 + ], + "type": "text", + "content": "14. Lee, H.Y., Yang, X., Liu, M.Y., Wang, T.C., Lu, Y.D., Yang, M.H., Kautz, J.: Dancing to music. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 579, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 579, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 134, + 579, + 481, + 621 + ], + "type": "text", + "content": "15. Li, J., Kang, D., Pei, W., Zhe, X., Zhang, Y., He, Z., Bao, L.: Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11293-11302 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 622, + 481, + 665 + ], + "type": "text", + "content": "16. Li, M., Chen, S., Zhao, Y., Zhang, Y., Wang, Y., Tian, Q.: Dynamic multiscale graph neural networks for 3d skeleton based human motion prediction. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 214-223 (2020)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 181, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 181, + 91, + 447, + 101 + ], + "type": "text", + "content": "Realistic Human Motion Generation with Cross-Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "17. Li, R., Zhao, J., Zhang, Y., Su, M., Ren, Z., Zhang, H., Tang, Y., Li, X.: Finedance: A fine-grained choreography dataset for 3d full body dance generation. arXiv preprint arXiv:2212.03741 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "text", + "content": "18. Li, Z., Zhou, Y., Xiao, S., He, C., Huang, Z., Li, H.: Auto-conditioned recurrent networks for extended complex human motion synthesis. arXiv preprint arXiv:1707.05363 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "text", + "content": "19. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 247 + ], + "type": "text", + "content": "20. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11461-11471 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 293 + ], + "type": "text", + "content": "21. Luo, Z., Chen, D., Zhang, Y., Huang, Y., Wang, L., Shen, Y., Zhao, D., Zhou, J., Tan, T.: Videofusion: Decomposed diffusion models for high-quality video generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10209-10218 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 326 + ], + "type": "text", + "content": "22. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5442-5451 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 347 + ], + "type": "text", + "content": "23. Nie, W., Guo, B., Huang, Y., Xiao, C., Vahdat, A., Anandkumar, A.: Diffusion models for adversarial purification. arXiv preprint arXiv:2205.07460 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 348, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 480, + 369 + ], + "type": "text", + "content": "24. Pavllo, D., Grangier, D., Auli, M.: Quaternet: A quaternion-based recurrent model for human motion. arXiv preprint arXiv:1805.06485 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 370, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 480, + 403 + ], + "type": "text", + "content": "25. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3d human motion synthesis with transformer vae. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10985-10995 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "text", + "content": "26. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: European Conference on Computer Vision. pp. 480-497. Springer (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "text", + "content": "27. Plappert, M., Mandery, C., Asfour, T.: The kit motion-language dataset. Big data 4(4), 236-252 (2016)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 500 + ], + "type": "text", + "content": "28. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 501, + 480, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 534 + ], + "type": "text", + "content": "29. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 567 + ], + "type": "text", + "content": "30. Ren, Z., Pan, Z., Zhou, X., Kang, L.: Diffusion motion: Generate text-guided 3d human motion by diffusion model. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "text", + "content": "31. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 601, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 643 + ], + "type": "text", + "content": "32. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "text", + "content": "33. Shafir, Y., Tevet, G., Kapon, R., Bermano, A.H.: Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Z. Ren et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "34. Sinha, A., Song, J., Meng, C., Ermon, S.: D2c: Diffusion-decoding models for few-shot conditional generation. Advances in Neural Information Processing Systems 34, 12533-12548 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "text", + "content": "35. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256-2265. PMLR (2015)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 203 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 203 + ], + "type": "text", + "content": "36. Song, Y., Ermon, S.: Improved techniques for training score-based generative models. Advances in neural information processing systems 33, 12438-12448 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 203, + 482, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 203, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 203, + 482, + 236 + ], + "type": "text", + "content": "37. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 236, + 482, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 257 + ], + "type": "text", + "content": "38. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 257, + 482, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 482, + 289 + ], + "type": "text", + "content": "39. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: European Conference on Computer Vision. pp. 358-374. Springer (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 289, + 482, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 311 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 311 + ], + "type": "text", + "content": "40. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. arXiv preprint arXiv:2209.14916 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 311, + 482, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 482, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 482, + 343 + ], + "type": "text", + "content": "41. Tseng, J., Castellon, R., Liu, K.: Edge: Editable dance generation from music. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 448-458 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 343, + 482, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 365 + ], + "type": "text", + "content": "42. Vahdat, A., Kreis, K., Kautz, J.: Score-based generative modeling in latent space. Advances in Neural Information Processing Systems 34, 11287-11302 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 365, + 482, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 385 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 385 + ], + "type": "text", + "content": "43. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 385, + 482, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 482, + 418 + ], + "type": "text", + "content": "44. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 418, + 482, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 450 + ], + "type": "text", + "content": "45. Wang, Y., Leng, Z., Li, F.W., Wu, S.C., Liang, X.: Fg-t2m: Fine-grained text-driven human motion generation via diffusion model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22035-22044 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 450, + 482, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 483 + ], + "type": "text", + "content": "46. Wei, D., Sun, H., Li, B., Lu, J., Li, W., Sun, X., Hu, S.: Human joint kinematics diffusion-refinement for stochastic motion prediction. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 6110-6118 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 483, + 482, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 482, + 525 + ], + "type": "text", + "content": "47. Xu, X., Wu, C., Rosenman, S., Lal, V., Che, W., Duan, N.: Bridgetower: Building bridges between encoders in vision-language representation learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 10637-10647 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 525, + 482, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 482, + 557 + ], + "type": "text", + "content": "48. Xu, Y., Zhang, J., Zhang, Q., Tao, D.: Vitpose: Simple vision transformer baselines for human pose estimation. Advances in Neural Information Processing Systems 35, 38571-38584 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 558, + 482, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 579 + ], + "type": "text", + "content": "49. Yang, R., Srivastava, P., Mandt, S.: Diffusion probabilistic modeling for video generation. arXiv preprint arXiv:2203.09481 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 579, + 482, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 579, + 482, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 579, + 482, + 611 + ], + "type": "text", + "content": "50. Yoon, Y., Cha, B., Lee, J.H., Jang, M., Lee, J., Kim, J., Lee, G.: Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics (TOG) 39(6), 1-16 (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 611, + 482, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 482, + 633 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 482, + 633 + ], + "type": "text", + "content": "51. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: Physdiff: Physics-guided human motion diffusion model. arXiv preprint arXiv:2212.02500 (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 633, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 633, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 633, + 482, + 666 + ], + "type": "text", + "content": "52. Zhang, J., Zhang, Y., Cun, X., Huang, S., Zhang, Y., Zhao, H., Lu, H., Shen, X.: T2m-gpt: Generating human motion from textual descriptions with discrete representations. arXiv preprint arXiv:2301.06052 (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 181, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 181, + 91, + 447, + 101 + ], + "type": "text", + "content": "Realistic Human Motion Generation with Cross-Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 216 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "text", + "content": "53. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 182 + ], + "type": "text", + "content": "54. Zhang, M., Guo, X., Pan, L., Cai, Z., Hong, F., Li, H., Yang, L., Liu, Z.: Remodif- \nfuse: Retrieval-augmented motion diffusion model. arXiv preprint arXiv:2304.01116 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "text", + "content": "55. Zhuang, H., Zhang, Y., Liu, S.: A pilot study of query-free adversarial attack against stable diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2384-2391 (2023)" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Z. Ren et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Reason2Drive_ Towards Interpretable and Chain-based Reasoning for Autonomous Driving/66f968e7-ebed-40e1-a3ff-b107a9813c0e_content_list.json b/2024/Reason2Drive_ Towards Interpretable and Chain-based Reasoning for Autonomous Driving/66f968e7-ebed-40e1-a3ff-b107a9813c0e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c98b63730307295fd90ddf0e496d182b9dd0ab38 --- /dev/null +++ b/2024/Reason2Drive_ Towards Interpretable and Chain-based Reasoning for Autonomous Driving/66f968e7-ebed-40e1-a3ff-b107a9813c0e_content_list.json @@ -0,0 +1,1819 @@ +[ + { + "type": "text", + "text": "Reason2Drive: Towards Interpretable and Chain-based Reasoning for Autonomous Driving", + "text_level": 1, + "bbox": [ + 225, + 140, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ming Nie $^{1}$ , Renyuan Peng $^{1}$ , Chunwei Wang $^{2}$ , Xinyue Cai $^{2}$ , Jianhua Han $^{2}$ , Hang Xu $^{2\\star}$ , and Li Zhang $^{1\\star}$", + "bbox": [ + 233, + 210, + 769, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Data Science, Fudan University $^{2}$ Huawei Noah's Ark Lab", + "bbox": [ + 352, + 253, + 647, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/fudan-zvg/reason2drive", + "bbox": [ + 354, + 297, + 647, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Large vision-language models (VLMs) have garnered increasing interest in autonomous driving areas, due to their advanced capabilities in complex reasoning tasks essential for highly autonomous vehicle behavior. Despite their potential, research in autonomous systems is hindered by the lack of datasets with annotated reasoning chains that explain the decision-making processes in driving. To bridge this gap, we present Reason2Drive, a benchmark dataset with over 600K video-text pairs, aimed at facilitating the study of interpretable reasoning in complex driving environments. We distinctly characterize the autonomous driving process as a sequential combination of perception, prediction, and reasoning steps, and the question-answer pairs are automatically collected from a diverse range of open-source outdoor driving datasets, including nuScenes, Waymo and ONCE. Moreover, we introduce a novel aggregated evaluation metric to assess chain-based reasoning performance in autonomous systems, addressing the reasoning ambiguities of existing metrics such as BLEU and CIDEr. Based on the proposed benchmark, we conduct experiments to assess various existing VLMs, revealing insights into their reasoning capabilities. Additionally, we develop an efficient approach to empower VLMs to leverage object-level perceptual elements in both feature extraction and prediction, further enhancing their reasoning accuracy. Extendable experiments demonstrate the supportive effect of Reason2Drive towards visual reasoning and downstream planning tasks.", + "bbox": [ + 261, + 348, + 743, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 695, + 375, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Modern autonomous driving systems face challenges related to generalization issues across diverse scenarios, which is often attributed to the reliance on empirical and intricate rules involved in decision-making. To reduce dependence on such rules, recent end-to-end approaches [5, 20] have been developed to derive control signals directly from sensor inputs, treating the system as a black", + "bbox": [ + 212, + 727, + 787, + 803 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Li Zhang (lizhangfd@fudan.edu.cn) and Hang Xu (chromexbjxh@gmail.com) are the corresponding authors.", + "bbox": [ + 217, + 810, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "box that requires extensive data for training. However, this approach tends to obscure the underlying logic of decisions, complicating failure diagnosis in real-world applications. In contrast, Large Vision-Language Models (VLMs) offer a promising alternative, potentially enhancing interpretability and generalization for these systems. With their broad world knowledge and advanced reasoning abilities, as illustrated in Fig. 1(a), VLMs have the potential to provide a more thorough understanding and explicit explanation for reliable decision-making.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nonetheless, existing works [35, 40] primarily focused on the straightforward adaptation of question-answering tasks to the autonomous driving; how to exploit VLMs to facilitate the reasoning abilities of autonomous systems is still under exploration.", + "bbox": [ + 212, + 252, + 491, + 357 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One reason that hinders the research in this field lies in the scarcity of datasets, especially those chained-based reasoning labels that elucidate the decision-making process. Most existing datasets [11, 35, 41] often oversimplify the complex processes of driving into straightforward question-answering tasks with only a few specific tasks covered. As depicted in Fig. 1(b), they typically provide closed-form annotations constrained", + "bbox": [ + 212, + 364, + 491, + 545 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to boolean (i.e., yes or no) answers or limited multiple-choice responses (e.g., stopped, parked, and moving). However, autonomous driving transcends a simplistic QA process. It encompasses a multi-step approach involving perception, prediction, and reasoning, each of which plays an indispensable role in the decision-making. Therefore, it is crucial to introduce a novel benchmark annotated with detailed decision-making reasoning for assessing the reasoning abilities of current VLMs.", + "bbox": [ + 212, + 546, + 787, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we introduce Reason2Drive, a new benchmark comprising over 600K video-text pairs, characterized by intricate driving instructions and a series of perception, prediction and reasoning steps. Our benchmark builds upon widely-used open-source driving datasets including nuScenes [2], Waymo [37], and ONCE [28], utilizing an extensible annotation schema. Specifically, we structure the comprehensive annotations into object-centric database and integrate it into manually crafted templates to create paired data for VLMs at both object and scenario levels. To enhance diversity, GPT-4 and manual instructions are employed for verification and enrichment purposes. Notably, Reason2Drive is the most extensive dataset available to date, outperforming existing datasets in scale and the complexity of reasoning chains included, which is a distinctive attribute not present in other datasets.", + "bbox": [ + 212, + 657, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/36d359993802073393053a61f1d11b3beadec8fc472636ff12379307d698ff20.jpg", + "image_caption": [ + "Fig.1: (a) Different decision-making processes in autonomous driving. (b) Language-based dataset comparison." + ], + "image_footnote": [], + "bbox": [ + 504, + 282, + 789, + 473 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Ming Nie, et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, we observe a fundamental flaw in the current evaluation of VLMs on autonomous driving tasks, due to the inherent reasoning ambiguities of traditional caption-based metrics like BLEU [31] and CIDEr [39]. These metrics mainly measure text generation from a holistic perspective, without considering the causal relationship between the reasoning steps and the final conclusion. For example, when the VLM suggests ego vehicle turning left, we cannot ascertain from these metrics whether its reasoning steps effectively support the final decision. To address this issue, we propose a new aggregated evaluation metric, ADRScore, specifically designed to measure chain-based reasoning performance in autonomous systems, which aims to resolve the reasoning ambiguities associated with current metrics.", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Utilizing the proposed benchmark, we undertake experiments to assess various existing VLMs, thereby unveiling valuable insights into their reasoning capabilities. We find that most methods struggle to effectively leverage perceptual priors, resulting in subpar reasoning performance. Additionally, constrained by the language model functioning solely as a decoder, these methods often fail to deliver accurate perceptual results, which is a crucial component for verifying a model's spatial reasoning capability. To alleviate this dilemma, we present a simple yet efficient framework, augmenting existing VLMs with two new components: a prior tokenizer and an instructed vision decoder, which aim to bolster the models' visual localization capabilities within the encoder and decoder, respectively. Extendable experiments demonstrate the supportive effect of Reason2Drive towards visual reasoning and downstream planning tasks.", + "bbox": [ + 212, + 311, + 787, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The contributions of this paper are summarized as follows: (i) We publish a novel visual instruction tuning dataset aimed at facilitating interpretable and chain-based reasoning autonomous systems. (ii) We introduce a novel evaluation metric, ADRScore, to assess chain-based reasoning performance in autonomous driving, effectively addressing the reasoning ambiguities present in existing metrics. (iii) We conduct experiments to assess a range of existing VLMs, revealing valuable insights into their reasoning capabilities. (iv) To address the challenges posed by inefficient prior feature extraction and inaccurate perceptual predictions, we introduce an efficient approach to integrate these elements into VLMs. This results in a substantial improvement in reasoning accuracy and provides remarkable support for downstream planning tasks. Our method surpasses all baselines, notably achieving impressive generalization in unseen scenarios.", + "bbox": [ + 212, + 494, + 787, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 700, + 387, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multimodal large language model. The current state of large language models provides remarkable abilities in natural language understanding and generation ([6, 7, 30, 38]). Inspired by the potential of large language models, a multitude of multimodal models has emerged, intended to enhance these models' capabilities in achieving multi-modal comprehension. Blip-2 [24] aligns visual and language features by utilizing a learnable Q-former. LLaVA [25] and MiniGPT-4 [46] initially align image-text features and then proceed with in", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Reason2Drive", + "bbox": [ + 637, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/48d2d27052970c57e8dd3a8226dfb8e7af4194e5f0c9748d58ab05e2626edee6.jpg", + "image_caption": [ + "Fig. 2: Schema of Our Reason2Drive Dataset. The upper part illustrates the pipeline for the automated construction of datasets. The lower part shows detailed instances of perception, prediction, and reasoning, accompanied by outcomes after applying GPT-4 for data augmentation. The special tokens hold distinct definitions: represents a specified instance, signifies a forecasted sequence of trajectory coordinates, and denotes positional coordinates. The colors associated with these tokens correspond to the highlighted objects in the upper-left image's boxes." + ], + "image_footnote": [], + "bbox": [ + 220, + 148, + 785, + 401 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "struction tuning. Additionally, Video-LLaMA [44] and ImageBind-LLM [18] integrate multiple modalities into the input, aligning features from various sources like images, videos, audio, and point clouds, consolidating them into the space of language features. Kosmos-2 [33] and Shikra [4] perform object detection based on instructions and also accomplish grounded visual question answering. DetGPT [34] connects a fixed multi-modal LLM with a customizable detector based on user instructions. LISA [23] efficiently embeds segmentation abilities into multi-modal LLMs, showcasing self-reasoning for current perception systems. The previous works have demonstrated that current large-scale multimodal models can achieve cross-modal alignment, enabling comprehension and inference towards images and more. These models can not only perform perceptual tasks like detection but also accomplish preliminary reasoning tasks.", + "bbox": [ + 212, + 518, + 787, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vision language tasks in autonomous driving. Currently, VLMs have demonstrated robust capabilities in scene perception and understanding. Extensive efforts have been dedicated to the realm of autonomous driving, leveraging VLM to achieve comprehensive scene understanding and perform diverse tasks [13, 16, 29, 42]. Simultaneously, substantial works are in progress to create datasets and models tailored to various tasks. Talk2Car [11] proposes the first object referral dataset for grounding commands for self-driving cars in free natural language into the visual context. But it exclusively contains information about visible objects. While DRAMA [27] outlines the overall scene risk, it lacks pre", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Ming Nie, et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "cise perception annotation. NuPrompt [41] and Refer-KITTI [40] offer language prompt sets for driving scenes but primarily concentrate on multi-object tracking tasks. NuScenesQA [35] and DriveLM [9] build visual question-answering (VQA) datasets for scenario understanding. However, their primary emphasis is on the perceptual information in the scene, lacking annotations for the analysis and complex reasoning of the entire scenario. To address the limitations of existing works, we construct a thorough dataset covering perception, prediction, and complex reasoning, additionally with an improved vision-language model for better analyzing autonomous driving scenarios.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Reason2Drive Dataset", + "text_level": 1, + "bbox": [ + 215, + 306, + 467, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We introduce Reason2Drive, a dataset that comprises comprehensive driving instructions and a chain-based reasoning framework for decision-making. Our dataset is characterized by the following key aspects:", + "bbox": [ + 212, + 339, + 785, + 385 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Quantity: It stands out as the largest language-based driving dataset available, collated from prominent publicly accessible datasets worldwide.", + "- **Quality:** Reason2Drive offers a more precise representation of driving activities, including perception, prediction and reasoning, with a reliable auto-annotation schema for data collection.", + "- Diversity: (i) The dataset exhibits a broader range of scenes, encompassing both object-level and scenario-level data. This diversity includes object types, visual and motion attributes, object locations, and relationships relative to the ego-vehicle. (ii) It includes more intricate question-answer pairs, enhanced by GPT-4, along with longer text passages featuring step-by-step reasoning.", + "- Protocols: A novel evaluation metric is introduced to assess the reasoning capabilities of VLMs. Different from those widely used in the NLP community, it takes into account not only perception results but also reasoning ambiguities, providing a more holistic evaluation of the VLM's reasoning capacity for autonomous driving scenarios." + ], + "bbox": [ + 225, + 402, + 782, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Further details regarding the data collection process, statistical data analysis, and benchmark protocols are provided in the subsequent section.", + "bbox": [ + 214, + 669, + 785, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Dataset collection", + "text_level": 1, + "bbox": [ + 215, + 723, + 410, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, we employ an extensible annotation schema, constructing data in the forms of question-answer pairs. Specifically, we first leverage a diverse array of publicly available datasets collected in different regions worldwide, including nuScenes, Waymo, and ONCE, and then parse their comprehensive annotations into object-centric database. The database is organized in key frames, and each frame stores object entries containing various details pertaining to its", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Reason2Drive", + "bbox": [ + 637, + 114, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a09e21784b2f65f73f0283aea84067e30d0ab5a21e46bc8915374d1d9cd66015.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetDescriptionSource Datasets
Talk2Car [11]Object referralnuScenes
CityFlow-NL [15]Tracking & retrievalCityFlow
CARLA-NAV [21]Segmentation & predictionCARLA Simulator
NuPrompt [41]Multi-object trackingnuScenes
NuScenes-QA [35]PerceptionnuScenes
Refer-KITTI [40]Multi-object trackingKITTI
Talk2BEV [12]Visual understandingnuScenes
DRAMA [27]Risk localizationself-collected
Rank2Tell [36]Risk localization & rankingself-collected
Reason2DrivePerceptionnuScenes
PredictionWaymo
ReasoningONCE
", + "bbox": [ + 233, + 143, + 517, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1: The comparison between our Reason2Drive dataset and other prompt-based datasets. ■ means dataset not published.", + "bbox": [ + 215, + 311, + 531, + 353 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4834d66a5938ded50e5fe869078d9bbf875b2d58a647a8b912d184c30b13e0aa.jpg", + "image_caption": [ + "Fig. 3: Data quality comparison. Reason2Drive is larger in scale, richer in data content, and more diverse in scenarios." + ], + "image_footnote": [], + "bbox": [ + 571, + 156, + 759, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "driving actions, including location, category, attributes and more. Afterwards, the structured database is required to answer the manually crafted question templates, which are divided into different tasks (i.e., perception, prediction and reasoning) at both object-level and scenario-level. Subsequently, GPT-4 is involved for verification and enrichment purposes. The example of GPT augmented data and manual instructions are provided in the appendix.", + "bbox": [ + 212, + 368, + 784, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Due to the complexity of autonomous driving activities, we categorize the tasks into three distinct groups to acquire diversified data: perception, prediction and reasoning. The specifics and distinctions of these three types of tasks are elaborated as follows:", + "bbox": [ + 212, + 458, + 785, + 518 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Perception task is designed to identify objects within the driving scenario, assessing the fundamental perceptual capabilities of VLMs in outdoor environments.", + "- Prediction task entails the prediction of future states of key objects within the perceptual range, challenging VLMs to infer the intentions of objects with video input.", + "- Reasoning task prompts the analysis of the current perceptual and predicted states step by step, requiring the deduction of reasoned inferences and decisions through a chain of thoughts (COT) approach." + ], + "bbox": [ + 225, + 530, + 784, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For each task, we further categorize the data into object-level and scenario-level. In more detail,", + "bbox": [ + 212, + 676, + 784, + 708 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Object-level data is formatted to benchmark the foundational capabilities of specific objects. As for perception, we address the location and attributes of objects such as moving status and distance to ego, while for prediction, future motion and merging-in/out status are considered.", + "- Scenario-level data is organized from a global perspective towards driving environment and ego-driving instructions. It focuses on whether there is an object worth noting currently (perception), whether there is an object worth noting in the future (prediction) and why (reasoning). For example," + ], + "bbox": [ + 223, + 719, + 784, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Ming Nie, et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d62edde5d1b5ac8f3529cf28c302e8770bd37e8424f0fca8c9f487722bc10952.jpg", + "image_caption": [ + "Fig. 4: Statistical distribution of different tasks in Reason2Drive." + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 421, + 262 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cc437172ff7a9cd5da62bda56cb2cbd12a9894df1c1848394755a2c23d7e4920.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Task TargetPerception (PE) Prediction (PR) Reasoning (RE)Total
Ego vehicle22629171731886858670
Single object718826166772006205555
Multi objects1020129450252175248689
Scenario495894279527657120041
Total246112216137170706632955
", + "bbox": [ + 473, + 165, + 785, + 263 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2: The statistics of different tasks in Reason2Drive dataset.", + "bbox": [ + 468, + 276, + 784, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "as illustrated in Fig. 2, models are asked to identify distances, merging states and other risks from the whole scene. It verifies the agent's ability to perceive the entire driving scene rather than specifying objects, thus more challenging.", + "bbox": [ + 241, + 330, + 782, + 376 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Dataset analysis", + "text_level": 1, + "bbox": [ + 215, + 414, + 395, + 429 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 1 and Fig. 3 demonstrate the comparison between our Reason2Drive dataset and existing benchmarks. It is noteworthy that our benchmark stands as the largest dataset to date, surpassing others in terms of both dataset size and the inclusion of extensive long-text chain-based reasoning references.", + "bbox": [ + 212, + 441, + 782, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further investigate the property of Reason2Drive dataset, we count the distribution and sample numbers of our dataset in Fig. 4 and Tab. 2. We split the dataset according to the task and target. The benchmark exhibits a balanced distribution, with multi-object tasks constituting the majority. Single-object and scenario-level questions are of similar quantities. The fewest questions are related to the ego-vehicle. Additionally, perception, prediction and reasoning questions are distributed as $39\\%$ , $34\\%$ , and $27\\%$ , respectively. More dataset details are provided in the appendix.", + "bbox": [ + 212, + 502, + 784, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Benchmark protocol", + "text_level": 1, + "bbox": [ + 215, + 646, + 429, + 661 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It is worth noting that previous works [12, 27, 35] simply utilize metric scores widely used in the NLP community, including BLEU [31], CIDEr [39] and ME-TEOR [1]. However, these metrics mainly measure text generation from a holistic perspective, without considering the causal relationship between the reasoning steps and the final conclusion. While they perform well in translation and captioning, their efficacy is limited when it comes to reasoning. Moreover, the existing evaluation system is confined to assessing the semantic quality of the results. It fails to effectively evaluate the quality of the perceived results, a crucial aspect supporting automatic driving decision-making. To address these dilemmas, inspired by [43] and [17], we develop a novel evaluation protocol, ADRScore, to measure the correctness of the reasoning chains towards autonomous driving.", + "bbox": [ + 212, + 672, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Reason2Drive", + "bbox": [ + 637, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b1d1d82a41fd4b26479b0c0c4d605067948d0c83ad641f894b2da666ff1fbac3.jpg", + "image_caption": [ + "Fig. 5: Illustration of ADRScore and its visual adaptation. When substituting semantic similarities with actual geometric similarities, the score decreases." + ], + "image_footnote": [], + "bbox": [ + 241, + 145, + 787, + 368 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Preliminary. To begin with, we denote the generated reasoning steps as hypothesis $\\pmb{h} = \\{h_1, \\dots, h_N\\}$ , and the gold annotation as reference $\\pmb{r} = \\{r_1, \\dots, r_K\\}$ .", + "bbox": [ + 214, + 411, + 782, + 443 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "At the core of reasoning metrics is the reasoning alignment vector from the $N$ -step hypothesis $h$ to the $K$ -step reference:", + "bbox": [ + 215, + 443, + 782, + 472 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {a l i g n} (\\boldsymbol {h} \\rightarrow \\boldsymbol {r}) = \\left\\{\\alpha_ {1}, \\dots , \\alpha_ {N} \\right\\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 489, + 784, + 507 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where alignment value $\\alpha_{i}$ represents the semantic similarity between the corresponding hypothesis step and the most similar reference step:", + "bbox": [ + 214, + 513, + 782, + 545 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i} = \\max _ {j = 1} ^ {K} s _ {i, j},\n$$\n", + "text_format": "latex", + "bbox": [ + 436, + 556, + 562, + 575 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i, j} = \\cos \\left(h _ {i}, r _ {j}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 439, + 569, + 784, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\alpha_{i} \\in [0,1]$ explicitly measures the grounding of the step-wise reasoning with respect to the reference, and $\\cos(\\cdot)$ denotes the cosine similarity between the corresponding sentence embeddings, which are extracted a pre-trained bert model. Based on the above reasoning alignment vector, we propose the following metrics to thoroughly measure the quality of reasoning steps.", + "bbox": [ + 214, + 607, + 782, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Reasoning alignment. The most straightforward way to evaluate the correctness of the hypothesis reasoning chain is to compare the degree of overlap between the hypothesis and the reference. One way of doing that is to measure the reasoning alignment between them:", + "bbox": [ + 214, + 684, + 782, + 744 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nR A = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\operatorname {a l i g n} \\left(h _ {i} \\rightarrow r\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 756, + 784, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Redundancy. To find chains that contain information that is not required to solve the problem (i.e., redundant steps), we identify those hypothesis steps that", + "bbox": [ + 214, + 809, + 782, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Ming Nie, et al.", + "bbox": [ + 271, + 114, + 379, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "are least aligned with the reference steps. This metric punishes the chain with steps that are not required for the correct solution.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nR D = \\min _ {i = 1} ^ {N} \\operatorname {a l i g n} \\left(h _ {i} \\rightarrow \\boldsymbol {r}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 186, + 785, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Missing step. To identify steps that are missing from the hypothesis but could be required to solve the problem, we look at the alignment between reference and the hypothesis, similar to Redundancy. However, here we go through each step in the reference, and check if there is a similar step in the hypothesis:", + "bbox": [ + 214, + 215, + 782, + 276 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nM S = \\min _ {i = 1} ^ {K} \\operatorname {a l i g n} \\left(r _ {i} \\rightarrow \\boldsymbol {h}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 287, + 785, + 305 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Finally, the aggregated metric score is the average of the above performance, which is:", + "bbox": [ + 214, + 315, + 782, + 340 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nA D R S c o r e = \\frac {1}{3} (R A + R D + M S). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 340, + 785, + 369 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Adapted to visual elements. To further adapt to the realistic driving process, we promote the above metric to the situation with visual elements, named ADRScore-S. Specifically, as illustrated in Fig. 5, when the hypothesis step $h_i$ and reference step $r_j$ contains visual elements, i.e., the locations and motions predicted for further reasoning, the similarity score becomes:", + "bbox": [ + 214, + 373, + 784, + 449 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i, j} = \\frac {\\tau - M \\left(h _ {i} , r _ {j}\\right)}{\\beta}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 460, + 785, + 491 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $M(\\cdot)$ measures the mean square error between two perceptual elements. And $\\tau$ and $\\beta$ are used to normalize it to [0,1] to match the distribution of semantic-level similarity. ADRScore-S more harshly measures the performance on spatial reasoning as the metric calculates the error of visual elements in spatial instead of text semantic. The latter is too lenient for visual predictions in language.", + "bbox": [ + 214, + 497, + 785, + 587 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Methodology", + "text_level": 1, + "bbox": [ + 214, + 609, + 380, + 628 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section we introduce our framework in Sec. 4.1, followed by the training details provided in Sec. 4.2.", + "bbox": [ + 212, + 642, + 782, + 672 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Model architecture", + "text_level": 1, + "bbox": [ + 214, + 694, + 419, + 709 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We observe that most VLMs struggle to effectively handle object-level perceptual information, including the input of visual priors and predictions of object locations, which are indispensable in autonomous driving scenarios. The limitation is primarily due to (i) the lack of a targeted tokenizer and (ii) decoder solely composed of a language model, resulting in subpar reasoning performance.", + "bbox": [ + 212, + 719, + 782, + 794 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To address this challenge, as illustrated in Fig. 6, we introduce a straightforward yet effective framework that enhances existing VLMs with two new components: a prior tokenizer and an instructed vision decoder. Notably, the design", + "bbox": [ + 212, + 795, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Reason2Drive", + "bbox": [ + 637, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/2636b6b6bb2a3c6b070ed52f9a633f9e27a2f5b52f1d54dceedd0ae7b7023d21.jpg", + "image_caption": [ + "Fig. 6: The pipeline of our framework. The input video and perceptual priors are tokenized using the vision encoder and prior tokenizer. A Q-former then aligns them to the text's feature space. The LLM and instructed vision decoder predict answers with precise perception results for user queries. The highlighted yellow box and red curve in the perception results respectively represent the visualization of $\\langle \\mathsf{LOC} \\rangle$ and $\\langle \\mathsf{MOT} \\rangle$ ." + ], + "image_footnote": [], + "bbox": [ + 215, + 142, + 787, + 296 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "of the components is not for detection but to aid interpretable visual reasoning based on available perception inputs. These modules aim to strengthen the capabilities of the model to utilize object-level perceptual elements in both extracting visual priors and generating perceptual predictions for visual reasoning.", + "bbox": [ + 212, + 386, + 787, + 446 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Vision encoder. Our model accepts both video frames and text inputs, along with perceptual priors, and tokenizes them into embeddings. For a sequence of video frames $(V_{1}, V_{2}, \\dots, V_{N})$ , features are extracted using a pretrained Blip-2 visual encoder [24] $F_{v}$ and aggregated through concatenation:", + "bbox": [ + 212, + 449, + 789, + 510 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nf _ {v} = F _ {v} \\left(V _ {1}\\right) \\oplus F _ {v} \\left(V _ {2}\\right) \\oplus \\dots \\oplus F _ {v} \\left(V _ {N}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 531, + 785, + 549 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Prior tokenizer. We propose a novel tokenization strategy tailored to taking advantage of visual cues. The motivation is grounded in the acknowledgement that extracting and aligning visual features is considerably simpler and more suitable compared to compelling the LLM to comprehend ambiguous positional descriptions. Direct textual input to the LLM may result in challenges such as information loss, as textual representation may not fully capture image details and context, especially in complex scenarios with dynamic object positions and velocities. To tackle this issue, we design a novel tokenizer $F_{p}$ , implemented as a two-layer MLP, to independently extract local image features and positional embeddings from visual priors:", + "bbox": [ + 212, + 561, + 787, + 715 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nf _ {p} = F _ {p} (f _ {r} + E (P)), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 734, + 785, + 753 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $f_{r}$ represents the region-level features extracted from the image-level features $f_{v}$ according to the precise locations of perception priors $P$ . These features are aligned to $7 \\times 7$ size using the RoIAlign [14] operation and fused into a single embedding $f_{r}$ . And $E(\\cdot)$ is a positional encoding function mapping the geometry locations and motions into the same dimension of $f_{r}$ .", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Ming Nie, et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "LLM. After we tokenize the video and perception priors into embedding $f_{v}$ and $f_{p}$ , a projector $Q$ (Q-former [24] in this work) is adopted to align the non-text features into textual domain:", + "bbox": [ + 214, + 146, + 782, + 191 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\nf _ {q} = Q \\left(f _ {v}, f _ {p}\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 449, + 208, + 784, + 222 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Then, to generate the final text output, we utilize the LLM for further language processing with the extracted text embedding $f_{t}$ :", + "bbox": [ + 214, + 228, + 782, + 260 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} _ {t} = F \\left(f _ {t}, f _ {q}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 450, + 276, + 784, + 290 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Instructed vision decoder. Current works [11, 16] treat the LLM as a versatile tool to generate answers and inferences without intermediate reasoning steps, let alone considering the perceptions of the agent toward driving scenes. However, the perception ability of the agent towards driving scenarios is an indispensable part of a reliable driving procedure. Moreover, recent works [23] have demonstrated that, rather than training with textualized perceptual sequences, incorporating the perception abilities into the multi-modal LLM brings a significant improvement. To this end, inspired by [23], we integrate new perception capabilities into the multi-modal LLM. Specifically, we expand the original LLM vocabulary by introducing new tokens as placeholders, denoted as $\\langle \\mathsf{LOC} \\rangle$ and $\\langle \\mathsf{MOT} \\rangle$ , to signify the request for the perception output. When the LLM aims to generate a specific perception, the output $\\hat{y}_t$ should include a designed token. We then extract the last-layer textual features corresponding to the specific token and apply an MLP projection layer to obtain the hidden embedding $f_h$ . Finally, the textual embedding and visual features are fed into the instructed vision decoder to decode the predictions:", + "bbox": [ + 214, + 295, + 784, + 537 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {P} = D \\left(f _ {v}, f _ {h}\\right). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 450, + 553, + 782, + 568 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "This module is comprised of a transformer decoder for features alignment [3] and task-specific heads designed to generate object locations and motions.", + "bbox": [ + 214, + 574, + 782, + 606 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Training details", + "text_level": 1, + "bbox": [ + 214, + 627, + 390, + 642 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Training objectives. The model is trained end-to-end using the text generation loss $\\mathcal{L}_{txt}$ and the perception output loss $\\mathcal{L}_{per}$ :", + "bbox": [ + 214, + 651, + 782, + 683 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {t x t} + \\lambda_ {p e r} \\mathcal {L} _ {p e r}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 428, + 698, + 782, + 712 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $\\lambda_{per}$ is the balancing term. Specifically, $\\mathcal{L}_{txt}$ is the auto-regressive cross-entropy loss for text generation, and $\\mathcal{L}_{per}$ encourages the instructed vision decoder to generate accurate locations and motions, which is similar to traditional detection loss and is employed with the combination of binary cross-entropy loss and MSE loss. More details are included in the appendix.", + "bbox": [ + 212, + 719, + 782, + 794 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Tuning strategy. Our tuning strategy consists of two stages: the pre-training stage and the fine-tuning stage. In the pre-training stage, we initialize the weights from instructBLIP [10], including the pre-trained vision encoder, Q-former and", + "bbox": [ + 214, + 795, + 782, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Reason2Drive", + "bbox": [ + 637, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/760d851588df512f61deb8076f2e54d4ee8141c870d2ef2e2172636d77243387.jpg", + "table_caption": [ + "Table 3: Results of different models on the Reason2Drive validation set. We evaluate with ADRScore as well as captioning-based metrics." + ], + "table_footnote": [], + "table_body": "
MethodsLLMReasoning metricCaptioning metric
ADRScoreADRScore-SB@4METEORROUGECIDEr
Blip-2 [24]OPT-2.7B [45]0.2960.1620.3610.2490.4430.174
FlanT5-XL [8]0.3100.1710.3680.2560.4510.187
InstructBLIP [10]FlanT5-XL0.3290.1870.3760.2690.4620.196
Vicuna-7B [32]0.3510.2140.4080.2940.4840.211
MiniGPT-4 [46]Vicuna-7B0.3380.2030.3960.2860.4750.219
OursFlanT5-XL0.4570.4200.4510.3490.5200.292
Vicuna-7B0.4630.4320.4570.3560.5290.298
", + "bbox": [ + 277, + 176, + 730, + 339 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/2fb5b36e415bc116d6e9fbcfc52cebc4bca9d2b0bc2d683c6b5a5baafb7dfd0c.jpg", + "table_caption": [ + "Table 4: Ablations on different combinations of training tasks." + ], + "table_footnote": [], + "table_body": "
TasksReasoning metricCaptioning metric
PerceptionPredictionReasoningADRScoreADRScore-SB@4METEORROUGECIDEr
✓0.2820.2530.4220.3070.4790.226
✓✓0.2970.2640.4190.3100.4790.228
✓0.3510.3230.4300.3250.4950.263
✓✓0.4070.3640.4350.3370.5010.274
✓✓✓0.4630.4320.4570.3560.5290.298
", + "bbox": [ + 277, + 361, + 730, + 483 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "LLM, and freeze the parameters of LLM and vision tokenizer $F_{v}$ . We train the prior tokenizer $F_{p}$ and Q-former $Q$ to align visual priors with text, along with the instructed vision decoder $D$ to enhance visual localization capabilities. The fine-tuning phase equips the LLM with reasoning abilities in autonomous driving using the instructed vision decoder. To retain pre-trained LLM generalization, we employ efficient fine-tuning with LoRA [19]. The vision encoder and prior tokenizer $F_{p}$ remain fixed, while the instructed vision decoder $D$ is fully fine-tuned. Word embeddings of the LLM and Q-former are also trainable.", + "bbox": [ + 212, + 497, + 787, + 619 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 642, + 375, + 659 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We benchmark various baseline models and present our method on Reason2Drive dataset. Sec. 5.1 covers implementation details. We assess reasoning performance using our proposed metric in Sec. 5.2, perform ablation studies in Sec. 5.3 and provide more ablations and qualitative results in the appendix.", + "bbox": [ + 212, + 672, + 787, + 734 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.1 Experimental setting", + "text_level": 1, + "bbox": [ + 215, + 753, + 434, + 771 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Our main experiments are carried out on the complete Reason2Drive benchmark. The dataset is collected from three different source datasets: nuScenes [2], Waymo [37], and ONCE [28]. It is divided into training and validation sets based on segments, with $70\\%$ allocated to the training set and $30\\%$ to the validation", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Ming Nie, et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/2a5ef19aec2a016a4b402cbdcd6bd69f0c2bc7ba9f37364c0594af5ff0e9689b.jpg", + "table_caption": [ + "Table 5: The quality of predicted visual elements." + ], + "table_footnote": [], + "table_body": "
PredictionsMetricMiniGPT-4Kosmos-2 [33]Ours
Bounding boxAccuracy ↑0.7230.7450.806
TrajectoryADE ↓2.3342.5631.875
", + "bbox": [ + 277, + 143, + 730, + 196 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "set, ensuring no overlap in scenes between them. The validation set also contains all the declared tasks and has been augmented. The input consists of 5 frames of cropped images with a size of $224 \\times 224$ pixels. During training, we leverage the AdamW [26] optimizer with a weight decay of 0.01. We adopt a cosine learning rate decay scheduler with a max value of 3e-4 and a linear warm-up for the first 1000 iterations. The weight of perception loss $\\lambda_{p}$ is set to 1.0. The normalization parameters $\\tau$ and $\\beta$ are selected to be 15 and 10 after empirical practice. Our models are trained with a batch size of 8 on 8 V100 GPUs. The baseline models are fine-tuned following the official tuning strategy, with the inputs consistent with our approach. Specifically, the perception priors are also provided as textual inputs to baseline models for fair comparison.", + "bbox": [ + 212, + 203, + 787, + 368 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/d1a03333d33809ad2f947317d1f1cd368e47c3b13117ce629ae56cba1da9ae79.jpg", + "table_caption": [ + "Table 6: Ablations on visual input and Table 7: Ablations on different settings of instructed vision decoder." + ], + "table_footnote": [], + "table_body": "
Visual features\nimage-level video-levelPerceptual priors\nregion-level positionalADRScoreADRScore-S
✓0.4140.379
0.4310.394
✓0.4470.418
✓0.4630.432
", + "bbox": [ + 243, + 419, + 500, + 489 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/94a828ffb28d311907c9d015a9086f5db39bef0dd0abca03c3c8edd7679204f4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Pre-train text embedding MLPADRScoreADRScore-S
0.3870.361
✓0.4210.396
✓✓0.4550.425
✓✓✓0.4630.432
", + "bbox": [ + 509, + 419, + 753, + 491 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.2 Reasoning performance evaluation", + "text_level": 1, + "bbox": [ + 215, + 500, + 545, + 515 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "As demonstrated in Tab. 3, we evaluate both ADRScore and traditional caption-based performance of different models on our benchmark. It is worth noting that our method outperforms others comprehensively in all metrics. We also observe that, despite there is a correlation between ADRScore and traditional metrics, while on the other hand, the performance gap is more pronounced in our metrics, specially revealed by ADRScore-S. The results further substantiate reasoning ambiguities in traditional metrics, constraining the differentiation in benchmarking model performance. Specifically, models with varying reasoning capabilities exhibit minimal disparities when evaluated using traditional metrics.", + "bbox": [ + 212, + 523, + 787, + 659 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Ablation study", + "text_level": 1, + "bbox": [ + 215, + 680, + 385, + 695 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Task contributions. To investigate the synergies between different tasks, we separate and evaluate various types of tasks independently. As shown in Tab. 4, we train our model on different combinations of tasks. Most notably, training on reasoning tasks plays the most important role, indicating the necessity of relevant reasoning data for instructional tuning. Based on reasoning tasks, perception and prediction tasks additionally enhance the models for visual reasoning, showing specific improvements of $4.1\\%$ and $6.8\\%$ , respectively.", + "bbox": [ + 212, + 704, + 787, + 809 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Quality of predicted visual elements. To validate the quality of predicted perceptual elements, we also conduct experiments to evaluate the bounding", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Reason2Drive", + "bbox": [ + 637, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/ef3a495f797ddbb68ae5c5ea799968516979631752781678ac07b1444a3b8277.jpg", + "table_caption": [ + "Table 8: Evaluation of control signals. B: B@4. M: METHOR." + ], + "table_footnote": [], + "table_body": "
MethodLLMB↑M↑RMSE
Speed↓Steer↓
Directly fine-tuned with control signals:
InstructBLIPVicuna-7B0.1660.2013.7435.926
Additionally pre-trained on Reason2Drive:
InstructBLIPVicuna-7B0.1920.2373.0865.151
OursVicuna-7B0.2130.2692.8424.866
", + "bbox": [ + 276, + 159, + 730, + 291 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "boxes and trajectories separately, as shown in Tab. 5. To ensure a fair comparison, we implement Kosmos-2 [33] due to its advantageous grounding capabilities. The model is fine-tuned on our dataset following the official strategy. Experimental results confirm the high quality of our predicted visual elements, with particular emphasis on the accuracy of the trajectories.", + "bbox": [ + 212, + 292, + 782, + 367 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The effects of tokenizers. To verify the effectiveness of the tokenizers, we conduct ablation studies to pinpoint where the improvements come from in Tab. 6. Visual features from single frame to multi-frame bring $1.5\\%$ improvement in ADRScore-S. Perceptual priors, i.e., region-level features and positional embeddings bring $2.4\\%$ and $1.4\\%$ development.", + "bbox": [ + 212, + 368, + 784, + 441 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The effects of instructed vision decoder. To verify the efficiency of our instructed vision decoder, we conduct an ablation study to compare it with other methods. As demonstrated in Tab. 7, pre-training and textual embedding bring the major contribution (3.5% and 2.9% in ADRScore-S).", + "bbox": [ + 212, + 443, + 784, + 503 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Downstream tasks. We are also interested in understanding how our benchmark will contribute to downstream tasks, such as predicting control signals. Following the approach in ADriver-I [22], we generate control signals on nuScenes and fine-tune InstructBLIP [10] directly for planning signal prediction. To ablate the influence of our dataset, we also pre-train models on Reason2Drive before fine-tuning with control signals. The results presented in Tab. 8 demonstrate the supportive effect of Reason2Drive on downstream planning tasks.", + "bbox": [ + 212, + 503, + 784, + 609 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 630, + 359, + 645 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In summary, Large Vision-Language Models (VLMs) have sparked interest in autonomous driving for their advanced reasoning capabilities. However, the absence of datasets explaining decision-making processes hinders progress. To tackle this, we introduce Reason2Drive benchmark, comprising $600\\mathrm{K}+$ video-text pairs for interpretable reasoning in complex driving scenarios. It outperforms existing datasets in scale, sources and task diversities. We also propose a novel evaluation protocol for chain-based reasoning, addressing existing semantic ambiguities. To uncover insights into their reasoning abilities, our work evaluates various VLMs and proposes an efficient method to boost the ability of models to utilize object-level perceptual elements in both the encoder and decoder. We expect our work could propel further advancements in interpretable reasoning for autonomous systems.", + "bbox": [ + 212, + 657, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Ming Nie, et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported in part by National Natural Science Foundation of China (Grant No. 62106050 and 62376060), Natural Science Foundation of Shanghai (Grant No. 22ZR1407500), USyd-Fudan BISA Flagship Research Program and Lingang Laboratory (Grant No. LG-QS-202202-07).", + "bbox": [ + 212, + 176, + 787, + 239 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 263, + 323, + 277 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Banerjee, S., Lavie, A.: Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In: ACL workshop (2005)", + "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: CVPR (2020)", + "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV (2020)", + "4. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal lvm's referential dialogue magic. arXiv preprint (2023)", + "5. Chen, L., Wu, P., Chitta, K., Jaeger, B., Geiger, A., Li, H.: End-to-end autonomous driving: Challenges and frontiers. arXiv preprint (2023)", + "6. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023)", + "7. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. arXiv preprint (2022)", + "8. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint (2022)", + "9. Contributors, D.: Drivelm: Drive on language. https://github.com/OpenDriveLab/DriveLM (2023)", + "0. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B.A., Fung, P., Hoi, S.C.H.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint (2023)", + "1. Deruyttere, T., Grujicic, D., Blaschko, M.B., Moens, M.F.: Talk2car: Predicting physical trajectories for natural language commands. IEEE Access (2022)", + "2. Dewangan, V., Choudhary, T., Chandhok, S., Priyadarshan, S., Jain, A., Singh, A.K., Srivastava, S., Jatavallabhula, K.M., Krishna, K.M.: Talk2bev: Language-enhanced bird's-eye view maps for autonomous driving. arXiv preprint (2023)", + "3. Ding, X., Han, J., Xu, H., Zhang, W., Li, X.: Hilm-d: Towards high-resolution understanding in multimodal large language models for autonomous driving. arXiv preprint (2023)", + "4. Dollar, K., Girshick, R.: Mask r-cnn. In: ICCV (2017)", + "5. Feng, Q., Ablavsky, V., Sclaroff, S.: Cityflow-nl: Tracking and retrieval of vehicles at city scale by natural language descriptions. arXiv preprint (2021)", + "6. Fu, D., Li, X., Wen, L., Dou, M., Cai, P., Shi, B., Qiao, Y.: Drive like a human: Rethinking autonomous driving with large language models. arXiv preprint (2023)" + ], + "bbox": [ + 225, + 295, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Reason2Drive", + "bbox": [ + 637, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Golovneva, O., Chen, M., Poff, S., Corredor, M., Zettlemoyer, L., Fazel-Zarandi, M., Celikyilmaz, A.: Roscoe: A suite of metrics for scoring step-by-step reasoning. arXiv preprint (2022)", + "18. Han, J., Zhang, R., Shao, W., Gao, P., Xu, P., Xiao, H., Zhang, K., Liu, C., Wen, S., Guo, Z., et al.: Imagebind-llm: Multi-modality instruction tuning. arXiv preprint (2023)", + "19. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint (2021)", + "20. Hu, Y., Yang, J., Chen, L., Li, K., Sima, C., Zhu, X., Chai, S., Du, S., Lin, T., Wang, W., et al.: Planning-oriented autonomous driving. In: CVPR (2023)", + "21. Jain, K., Chhangani, V., Tiwari, A., Krishna, K.M., Gandhi, V.: Ground then navigate: Language-guided navigation in dynamic scenes. In: ICRA (2023)", + "22. Jia, F., Mao, W., Liu, Y., Zhao, Y., Wen, Y., Zhang, C., Zhang, X., Wang, T.: Adriver-i: A general world model for autonomous driving. arXiv preprint (2023)", + "23. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: Lisa: Reasoning segmentation via large language model. arXiv preprint (2023)", + "24. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint (2023)", + "25. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NIPS (2023)", + "26. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2017)", + "27. Malla, S., Choi, C., Dwivedi, I., Choi, J.H., Li, J.: Drama: Joint risk localization and captioning in driving. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (2023)", + "28. Mao, J., Niu, M., Jiang, C., Liang, H., Chen, J., Liang, X., Li, Y., Ye, C., Zhang, W., Li, Z., et al.: One million scenes for autonomous driving: Once dataset. arXiv preprint (2021)", + "29. Mao, J., Qian, Y., Zhao, H., Wang, Y.: Gpt-driver: Learning to drive with gpt. arXiv preprint (2023)", + "30. OpenAI: Gpt-4: A large-scale transformer-based language model (2023), https://www.openai.com/research/gpt-4, https://www.openai.com/research/gpt-4", + "31. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a method for automatic evaluation of machine translation. In: ACL (2002)", + "32. Peng, B., Li, C., He, P., Galley, M., Gao, J.: Instruction tuning with gpt-4. arXiv preprint (2023)", + "33. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint (2023)", + "34. Pi, R., Gao, J., Diao, S., Pan, R., Dong, H., Zhang, J., Yao, L., Han, J., Xu, H., Zhang, L.K.T.: Detgpt: Detect what you need via reasoning. arXiv preprint (2023)", + "35. Qian, T., Chen, J., Zhuo, L., Jiao, Y., Jiang, Y.G.: Nuscenes-qa: A multi-modal visual question answering benchmark for autonomous driving scenario. arXiv preprint (2023)", + "36. Sachdeva, E., Agarwal, N., Chundi, S., Roelofs, S., Li, J., Dariush, B., Choi, C., Kochenderfer, M.: Rank2tell: A multimodal driving dataset for joint importance ranking and reasoning. arXiv preprint (2023)", + "37. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: CVPR (2020)", + "38. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Ming Nie, et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "39. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: CVPR (2015)", + "40. Wu, D., Han, W., Wang, T., Dong, X., Zhang, X., Shen, J.: Referring multi-object tracking. In: CVPR (2023)", + "41. Wu, D., Han, W., Wang, T., Liu, Y., Zhang, X., Shen, J.: Language prompt for autonomous driving. arXiv preprint (2023)", + "42. Xu, Z., Zhang, Y., Xie, E., Zhao, Z., Guo, Y., Wong, K.K., Li, Z., Zhao, H.: Drivegpt4: Interpretable end-to-end autonomous driving via large language model. arXiv preprint (2023)", + "43. Yu, P., Wang, T., Golovneva, O., Alkhamissy, B., Ghosh, G., Diab, M., Celikyilmaz, A.: Alert: Adapting language models to reasoning tasks. arXiv preprint (2022)", + "44. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint (2023)", + "45. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint (2022)", + "46. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 410 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Reason2Drive", + "bbox": [ + 637, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Reason2Drive_ Towards Interpretable and Chain-based Reasoning for Autonomous Driving/66f968e7-ebed-40e1-a3ff-b107a9813c0e_model.json b/2024/Reason2Drive_ Towards Interpretable and Chain-based Reasoning for Autonomous Driving/66f968e7-ebed-40e1-a3ff-b107a9813c0e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..10ac1d011838ca23168889d3e93f1b151d69013e --- /dev/null +++ b/2024/Reason2Drive_ Towards Interpretable and Chain-based Reasoning for Autonomous Driving/66f968e7-ebed-40e1-a3ff-b107a9813c0e_model.json @@ -0,0 +1,2423 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.227, + 0.141, + 0.78, + 0.187 + ], + "angle": 0, + "content": "Reason2Drive: Towards Interpretable and Chain-based Reasoning for Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.212, + 0.77, + 0.245 + ], + "angle": 0, + "content": "Ming Nie\\(^{1}\\), Renyuan Peng\\(^{1}\\), Chunwei Wang\\(^{2}\\), Xinyue Cai\\(^{2}\\), Jianhua Han\\(^{2}\\), Hang Xu\\(^{2\\star}\\), and Li Zhang\\(^{1\\star}\\)" + }, + { + "type": "text", + "bbox": [ + 0.354, + 0.255, + 0.648, + 0.283 + ], + "angle": 0, + "content": "\\(^{1}\\) School of Data Science, Fudan University \\(^{2}\\) Huawei Noah's Ark Lab" + }, + { + "type": "text", + "bbox": [ + 0.355, + 0.298, + 0.648, + 0.312 + ], + "angle": 0, + "content": "https://github.com/fudan-zvg/reason2drive" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.349, + 0.744, + 0.655 + ], + "angle": 0, + "content": "Abstract. Large vision-language models (VLMs) have garnered increasing interest in autonomous driving areas, due to their advanced capabilities in complex reasoning tasks essential for highly autonomous vehicle behavior. Despite their potential, research in autonomous systems is hindered by the lack of datasets with annotated reasoning chains that explain the decision-making processes in driving. To bridge this gap, we present Reason2Drive, a benchmark dataset with over 600K video-text pairs, aimed at facilitating the study of interpretable reasoning in complex driving environments. We distinctly characterize the autonomous driving process as a sequential combination of perception, prediction, and reasoning steps, and the question-answer pairs are automatically collected from a diverse range of open-source outdoor driving datasets, including nuScenes, Waymo and ONCE. Moreover, we introduce a novel aggregated evaluation metric to assess chain-based reasoning performance in autonomous systems, addressing the reasoning ambiguities of existing metrics such as BLEU and CIDEr. Based on the proposed benchmark, we conduct experiments to assess various existing VLMs, revealing insights into their reasoning capabilities. Additionally, we develop an efficient approach to empower VLMs to leverage object-level perceptual elements in both feature extraction and prediction, further enhancing their reasoning accuracy. Extendable experiments demonstrate the supportive effect of Reason2Drive towards visual reasoning and downstream planning tasks." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.696, + 0.377, + 0.712 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.728, + 0.788, + 0.804 + ], + "angle": 0, + "content": "Modern autonomous driving systems face challenges related to generalization issues across diverse scenarios, which is often attributed to the reliance on empirical and intricate rules involved in decision-making. To reduce dependence on such rules, recent end-to-end approaches [5, 20] have been developed to derive control signals directly from sensor inputs, treating the system as a black" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.788, + 0.841 + ], + "angle": 0, + "content": "* Li Zhang (lizhangfd@fudan.edu.cn) and Hang Xu (chromexbjxh@gmail.com) are the corresponding authors." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "Ming Nie, et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "box that requires extensive data for training. However, this approach tends to obscure the underlying logic of decisions, complicating failure diagnosis in real-world applications. In contrast, Large Vision-Language Models (VLMs) offer a promising alternative, potentially enhancing interpretability and generalization for these systems. With their broad world knowledge and advanced reasoning abilities, as illustrated in Fig. 1(a), VLMs have the potential to provide a more thorough understanding and explicit explanation for reliable decision-making." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.492, + 0.358 + ], + "angle": 0, + "content": "Nonetheless, existing works [35, 40] primarily focused on the straightforward adaptation of question-answering tasks to the autonomous driving; how to exploit VLMs to facilitate the reasoning abilities of autonomous systems is still under exploration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.366, + 0.493, + 0.546 + ], + "angle": 0, + "content": "One reason that hinders the research in this field lies in the scarcity of datasets, especially those chained-based reasoning labels that elucidate the decision-making process. Most existing datasets [11, 35, 41] often oversimplify the complex processes of driving into straightforward question-answering tasks with only a few specific tasks covered. As depicted in Fig. 1(b), they typically provide closed-form annotations constrained" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.547, + 0.788, + 0.651 + ], + "angle": 0, + "content": "to boolean (i.e., yes or no) answers or limited multiple-choice responses (e.g., stopped, parked, and moving). However, autonomous driving transcends a simplistic QA process. It encompasses a multi-step approach involving perception, prediction, and reasoning, each of which plays an indispensable role in the decision-making. Therefore, it is crucial to introduce a novel benchmark annotated with detailed decision-making reasoning for assessing the reasoning abilities of current VLMs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To this end, we introduce Reason2Drive, a new benchmark comprising over 600K video-text pairs, characterized by intricate driving instructions and a series of perception, prediction and reasoning steps. Our benchmark builds upon widely-used open-source driving datasets including nuScenes [2], Waymo [37], and ONCE [28], utilizing an extensible annotation schema. Specifically, we structure the comprehensive annotations into object-centric database and integrate it into manually crafted templates to create paired data for VLMs at both object and scenario levels. To enhance diversity, GPT-4 and manual instructions are employed for verification and enrichment purposes. Notably, Reason2Drive is the most extensive dataset available to date, outperforming existing datasets in scale and the complexity of reasoning chains included, which is a distinctive attribute not present in other datasets." + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.284, + 0.79, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.486, + 0.788, + 0.528 + ], + "angle": 0, + "content": "Fig.1: (a) Different decision-making processes in autonomous driving. (b) Language-based dataset comparison." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.638, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Reason2Drive" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.312 + ], + "angle": 0, + "content": "Furthermore, we observe a fundamental flaw in the current evaluation of VLMs on autonomous driving tasks, due to the inherent reasoning ambiguities of traditional caption-based metrics like BLEU [31] and CIDEr [39]. These metrics mainly measure text generation from a holistic perspective, without considering the causal relationship between the reasoning steps and the final conclusion. For example, when the VLM suggests ego vehicle turning left, we cannot ascertain from these metrics whether its reasoning steps effectively support the final decision. To address this issue, we propose a new aggregated evaluation metric, ADRScore, specifically designed to measure chain-based reasoning performance in autonomous systems, which aims to resolve the reasoning ambiguities associated with current metrics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.788, + 0.495 + ], + "angle": 0, + "content": "Utilizing the proposed benchmark, we undertake experiments to assess various existing VLMs, thereby unveiling valuable insights into their reasoning capabilities. We find that most methods struggle to effectively leverage perceptual priors, resulting in subpar reasoning performance. Additionally, constrained by the language model functioning solely as a decoder, these methods often fail to deliver accurate perceptual results, which is a crucial component for verifying a model's spatial reasoning capability. To alleviate this dilemma, we present a simple yet efficient framework, augmenting existing VLMs with two new components: a prior tokenizer and an instructed vision decoder, which aim to bolster the models' visual localization capabilities within the encoder and decoder, respectively. Extendable experiments demonstrate the supportive effect of Reason2Drive towards visual reasoning and downstream planning tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.496, + 0.788, + 0.678 + ], + "angle": 0, + "content": "The contributions of this paper are summarized as follows: (i) We publish a novel visual instruction tuning dataset aimed at facilitating interpretable and chain-based reasoning autonomous systems. (ii) We introduce a novel evaluation metric, ADRScore, to assess chain-based reasoning performance in autonomous driving, effectively addressing the reasoning ambiguities present in existing metrics. (iii) We conduct experiments to assess a range of existing VLMs, revealing valuable insights into their reasoning capabilities. (iv) To address the challenges posed by inefficient prior feature extraction and inaccurate perceptual predictions, we introduce an efficient approach to integrate these elements into VLMs. This results in a substantial improvement in reasoning accuracy and provides remarkable support for downstream planning tasks. Our method surpasses all baselines, notably achieving impressive generalization in unseen scenarios." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.701, + 0.388, + 0.716 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Multimodal large language model. The current state of large language models provides remarkable abilities in natural language understanding and generation ([6, 7, 30, 38]). Inspired by the potential of large language models, a multitude of multimodal models has emerged, intended to enhance these models' capabilities in achieving multi-modal comprehension. Blip-2 [24] aligns visual and language features by utilizing a learnable Q-former. LLaVA [25] and MiniGPT-4 [46] initially align image-text features and then proceed with in" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "Ming Nie, et al." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.15, + 0.787, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.411, + 0.788, + 0.509 + ], + "angle": 0, + "content": "Fig. 2: Schema of Our Reason2Drive Dataset. The upper part illustrates the pipeline for the automated construction of datasets. The lower part shows detailed instances of perception, prediction, and reasoning, accompanied by outcomes after applying GPT-4 for data augmentation. The special tokens hold distinct definitions: represents a specified instance, signifies a forecasted sequence of trajectory coordinates, and denotes positional coordinates. The colors associated with these tokens correspond to the highlighted objects in the upper-left image's boxes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.788, + 0.702 + ], + "angle": 0, + "content": "struction tuning. Additionally, Video-LLaMA [44] and ImageBind-LLM [18] integrate multiple modalities into the input, aligning features from various sources like images, videos, audio, and point clouds, consolidating them into the space of language features. Kosmos-2 [33] and Shikra [4] perform object detection based on instructions and also accomplish grounded visual question answering. DetGPT [34] connects a fixed multi-modal LLM with a customizable detector based on user instructions. LISA [23] efficiently embeds segmentation abilities into multi-modal LLMs, showcasing self-reasoning for current perception systems. The previous works have demonstrated that current large-scale multimodal models can achieve cross-modal alignment, enabling comprehension and inference towards images and more. These models can not only perform perceptual tasks like detection but also accomplish preliminary reasoning tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Vision language tasks in autonomous driving. Currently, VLMs have demonstrated robust capabilities in scene perception and understanding. Extensive efforts have been dedicated to the realm of autonomous driving, leveraging VLM to achieve comprehensive scene understanding and perform diverse tasks [13, 16, 29, 42]. Simultaneously, substantial works are in progress to create datasets and models tailored to various tasks. Talk2Car [11] proposes the first object referral dataset for grounding commands for self-driving cars in free natural language into the visual context. But it exclusively contains information about visible objects. While DRAMA [27] outlines the overall scene risk, it lacks pre" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.638, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Reason2Drive" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "cise perception annotation. NuPrompt [41] and Refer-KITTI [40] offer language prompt sets for driving scenes but primarily concentrate on multi-object tracking tasks. NuScenesQA [35] and DriveLM [9] build visual question-answering (VQA) datasets for scenario understanding. However, their primary emphasis is on the perceptual information in the scene, lacking annotations for the analysis and complex reasoning of the entire scenario. To address the limitations of existing works, we construct a thorough dataset covering perception, prediction, and complex reasoning, additionally with an improved vision-language model for better analyzing autonomous driving scenarios." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.308, + 0.468, + 0.323 + ], + "angle": 0, + "content": "3 Reason2Drive Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.787, + 0.386 + ], + "angle": 0, + "content": "We introduce Reason2Drive, a dataset that comprises comprehensive driving instructions and a chain-based reasoning framework for decision-making. Our dataset is characterized by the following key aspects:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.403, + 0.784, + 0.431 + ], + "angle": 0, + "content": "- Quantity: It stands out as the largest language-based driving dataset available, collated from prominent publicly accessible datasets worldwide." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.434, + 0.784, + 0.476 + ], + "angle": 0, + "content": "- **Quality:** Reason2Drive offers a more precise representation of driving activities, including perception, prediction and reasoning, with a reliable auto-annotation schema for data collection." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.485, + 0.784, + 0.573 + ], + "angle": 0, + "content": "- Diversity: (i) The dataset exhibits a broader range of scenes, encompassing both object-level and scenario-level data. This diversity includes object types, visual and motion attributes, object locations, and relationships relative to the ego-vehicle. (ii) It includes more intricate question-answer pairs, enhanced by GPT-4, along with longer text passages featuring step-by-step reasoning." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.584, + 0.784, + 0.657 + ], + "angle": 0, + "content": "- Protocols: A novel evaluation metric is introduced to assess the reasoning capabilities of VLMs. Different from those widely used in the NLP community, it takes into account not only perception results but also reasoning ambiguities, providing a more holistic evaluation of the VLM's reasoning capacity for autonomous driving scenarios." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.403, + 0.784, + 0.657 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.67, + 0.786, + 0.701 + ], + "angle": 0, + "content": "Further details regarding the data collection process, statistical data analysis, and benchmark protocols are provided in the subsequent section." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.724, + 0.411, + 0.737 + ], + "angle": 0, + "content": "3.1 Dataset collection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, we employ an extensible annotation schema, constructing data in the forms of question-answer pairs. Specifically, we first leverage a diverse array of publicly available datasets collected in different regions worldwide, including nuScenes, Waymo, and ONCE, and then parse their comprehensive annotations into object-centric database. The database is organized in key frames, and each frame stores object entries containing various details pertaining to its" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "Ming Nie, et al." + }, + { + "type": "table", + "bbox": [ + 0.235, + 0.145, + 0.519, + 0.302 + ], + "angle": 0, + "content": "
DatasetDescriptionSource Datasets
Talk2Car [11]Object referralnuScenes
CityFlow-NL [15]Tracking & retrievalCityFlow
CARLA-NAV [21]Segmentation & predictionCARLA Simulator
NuPrompt [41]Multi-object trackingnuScenes
NuScenes-QA [35]PerceptionnuScenes
Refer-KITTI [40]Multi-object trackingKITTI
Talk2BEV [12]Visual understandingnuScenes
DRAMA [27]Risk localizationself-collected
Rank2Tell [36]Risk localization & rankingself-collected
Reason2DrivePerceptionnuScenes
PredictionWaymo
ReasoningONCE
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.312, + 0.532, + 0.354 + ], + "angle": 0, + "content": "Table 1: The comparison between our Reason2Drive dataset and other prompt-based datasets. ■ means dataset not published." + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.157, + 0.761, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.301, + 0.781, + 0.356 + ], + "angle": 0, + "content": "Fig. 3: Data quality comparison. Reason2Drive is larger in scale, richer in data content, and more diverse in scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.369, + 0.785, + 0.459 + ], + "angle": 0, + "content": "driving actions, including location, category, attributes and more. Afterwards, the structured database is required to answer the manually crafted question templates, which are divided into different tasks (i.e., perception, prediction and reasoning) at both object-level and scenario-level. Subsequently, GPT-4 is involved for verification and enrichment purposes. The example of GPT augmented data and manual instructions are provided in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.787, + 0.519 + ], + "angle": 0, + "content": "Due to the complexity of autonomous driving activities, we categorize the tasks into three distinct groups to acquire diversified data: perception, prediction and reasoning. The specifics and distinctions of these three types of tasks are elaborated as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.531, + 0.784, + 0.574 + ], + "angle": 0, + "content": "- Perception task is designed to identify objects within the driving scenario, assessing the fundamental perceptual capabilities of VLMs in outdoor environments." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.577, + 0.785, + 0.62 + ], + "angle": 0, + "content": "- Prediction task entails the prediction of future states of key objects within the perceptual range, challenging VLMs to infer the intentions of objects with video input." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.622, + 0.784, + 0.667 + ], + "angle": 0, + "content": "- Reasoning task prompts the analysis of the current perceptual and predicted states step by step, requiring the deduction of reasoned inferences and decisions through a chain of thoughts (COT) approach." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.531, + 0.785, + 0.667 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.678, + 0.785, + 0.709 + ], + "angle": 0, + "content": "For each task, we further categorize the data into object-level and scenario-level. In more detail," + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.72, + 0.784, + 0.779 + ], + "angle": 0, + "content": "- Object-level data is formatted to benchmark the foundational capabilities of specific objects. As for perception, we address the location and attributes of objects such as moving status and distance to ego, while for prediction, future motion and merging-in/out status are considered." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.78, + 0.785, + 0.84 + ], + "angle": 0, + "content": "- Scenario-level data is organized from a global perspective towards driving environment and ego-driving instructions. It focuses on whether there is an object worth noting currently (perception), whether there is an object worth noting in the future (prediction) and why (reasoning). For example," + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.72, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.638, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Reason2Drive" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.144, + 0.422, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.277, + 0.448, + 0.303 + ], + "angle": 0, + "content": "Fig. 4: Statistical distribution of different tasks in Reason2Drive." + }, + { + "type": "table", + "bbox": [ + 0.474, + 0.166, + 0.787, + 0.264 + ], + "angle": 0, + "content": "
Task TargetPerception (PE) Prediction (PR) Reasoning (RE)Total
Ego vehicle22629171731886858670
Single object718826166772006205555
Multi objects1020129450252175248689
Scenario495894279527657120041
Total246112216137170706632955
" + }, + { + "type": "table_caption", + "bbox": [ + 0.47, + 0.277, + 0.785, + 0.303 + ], + "angle": 0, + "content": "Table 2: The statistics of different tasks in Reason2Drive dataset." + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.332, + 0.784, + 0.377 + ], + "angle": 0, + "content": "as illustrated in Fig. 2, models are asked to identify distances, merging states and other risks from the whole scene. It verifies the agent's ability to perceive the entire driving scene rather than specifying objects, thus more challenging." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.415, + 0.396, + 0.43 + ], + "angle": 0, + "content": "3.2 Dataset analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.442, + 0.784, + 0.501 + ], + "angle": 0, + "content": "Tab. 1 and Fig. 3 demonstrate the comparison between our Reason2Drive dataset and existing benchmarks. It is noteworthy that our benchmark stands as the largest dataset to date, surpassing others in terms of both dataset size and the inclusion of extensive long-text chain-based reasoning references." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.785, + 0.623 + ], + "angle": 0, + "content": "To further investigate the property of Reason2Drive dataset, we count the distribution and sample numbers of our dataset in Fig. 4 and Tab. 2. We split the dataset according to the task and target. The benchmark exhibits a balanced distribution, with multi-object tasks constituting the majority. Single-object and scenario-level questions are of similar quantities. The fewest questions are related to the ego-vehicle. Additionally, perception, prediction and reasoning questions are distributed as \\(39\\%\\), \\(34\\%\\), and \\(27\\%\\), respectively. More dataset details are provided in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.647, + 0.43, + 0.662 + ], + "angle": 0, + "content": "3.3 Benchmark protocol" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.84 + ], + "angle": 0, + "content": "It is worth noting that previous works [12, 27, 35] simply utilize metric scores widely used in the NLP community, including BLEU [31], CIDEr [39] and ME-TEOR [1]. However, these metrics mainly measure text generation from a holistic perspective, without considering the causal relationship between the reasoning steps and the final conclusion. While they perform well in translation and captioning, their efficacy is limited when it comes to reasoning. Moreover, the existing evaluation system is confined to assessing the semantic quality of the results. It fails to effectively evaluate the quality of the perceived results, a crucial aspect supporting automatic driving decision-making. To address these dilemmas, inspired by [43] and [17], we develop a novel evaluation protocol, ADRScore, to measure the correctness of the reasoning chains towards autonomous driving." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.38, + 0.129 + ], + "angle": 0, + "content": "Ming Nie, et al." + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.146, + 0.788, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.378, + 0.789, + 0.406 + ], + "angle": 0, + "content": "Fig. 5: Illustration of ADRScore and its visual adaptation. When substituting semantic similarities with actual geometric similarities, the score decreases." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.412, + 0.784, + 0.444 + ], + "angle": 0, + "content": "Preliminary. To begin with, we denote the generated reasoning steps as hypothesis \\( \\pmb{h} = \\{h_1, \\dots, h_N\\} \\), and the gold annotation as reference \\( \\pmb{r} = \\{r_1, \\dots, r_K\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.444, + 0.784, + 0.473 + ], + "angle": 0, + "content": "At the core of reasoning metrics is the reasoning alignment vector from the \\(N\\)-step hypothesis \\(h\\) to the \\(K\\)-step reference:" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.491, + 0.785, + 0.508 + ], + "angle": 0, + "content": "\\[\n\\operatorname {a l i g n} (\\boldsymbol {h} \\rightarrow \\boldsymbol {r}) = \\left\\{\\alpha_ {1}, \\dots , \\alpha_ {N} \\right\\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.515, + 0.784, + 0.546 + ], + "angle": 0, + "content": "where alignment value \\(\\alpha_{i}\\) represents the semantic similarity between the corresponding hypothesis step and the most similar reference step:" + }, + { + "type": "equation", + "bbox": [ + 0.437, + 0.557, + 0.563, + 0.577 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i} = \\max _ {j = 1} ^ {K} s _ {i, j},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.44, + 0.57, + 0.785, + 0.596 + ], + "angle": 0, + "content": "\\[\ns _ {i, j} = \\cos \\left(h _ {i}, r _ {j}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.608, + 0.784, + 0.683 + ], + "angle": 0, + "content": "\\(\\alpha_{i} \\in [0,1]\\) explicitly measures the grounding of the step-wise reasoning with respect to the reference, and \\(\\cos(\\cdot)\\) denotes the cosine similarity between the corresponding sentence embeddings, which are extracted a pre-trained bert model. Based on the above reasoning alignment vector, we propose the following metrics to thoroughly measure the quality of reasoning steps." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.685, + 0.784, + 0.745 + ], + "angle": 0, + "content": "Reasoning alignment. The most straightforward way to evaluate the correctness of the hypothesis reasoning chain is to compare the degree of overlap between the hypothesis and the reference. One way of doing that is to measure the reasoning alignment between them:" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "\\[\nR A = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\operatorname {a l i g n} \\left(h _ {i} \\rightarrow r\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Redundancy. To find chains that contain information that is not required to solve the problem (i.e., redundant steps), we identify those hypothesis steps that" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.638, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Reason2Drive" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "are least aligned with the reference steps. This metric punishes the chain with steps that are not required for the correct solution." + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.188, + 0.786, + 0.207 + ], + "angle": 0, + "content": "\\[\nR D = \\min _ {i = 1} ^ {N} \\operatorname {a l i g n} \\left(h _ {i} \\rightarrow \\boldsymbol {r}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.216, + 0.784, + 0.277 + ], + "angle": 0, + "content": "Missing step. To identify steps that are missing from the hypothesis but could be required to solve the problem, we look at the alignment between reference and the hypothesis, similar to Redundancy. However, here we go through each step in the reference, and check if there is a similar step in the hypothesis:" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.288, + 0.786, + 0.306 + ], + "angle": 0, + "content": "\\[\nM S = \\min _ {i = 1} ^ {K} \\operatorname {a l i g n} \\left(r _ {i} \\rightarrow \\boldsymbol {h}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.316, + 0.784, + 0.342 + ], + "angle": 0, + "content": "Finally, the aggregated metric score is the average of the above performance, which is:" + }, + { + "type": "equation", + "bbox": [ + 0.376, + 0.342, + 0.786, + 0.371 + ], + "angle": 0, + "content": "\\[\nA D R S c o r e = \\frac {1}{3} (R A + R D + M S). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.374, + 0.785, + 0.45 + ], + "angle": 0, + "content": "Adapted to visual elements. To further adapt to the realistic driving process, we promote the above metric to the situation with visual elements, named ADRScore-S. Specifically, as illustrated in Fig. 5, when the hypothesis step \\( h_i \\) and reference step \\( r_j \\) contains visual elements, i.e., the locations and motions predicted for further reasoning, the similarity score becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.427, + 0.462, + 0.786, + 0.492 + ], + "angle": 0, + "content": "\\[\ns _ {i, j} = \\frac {\\tau - M \\left(h _ {i} , r _ {j}\\right)}{\\beta}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.498, + 0.786, + 0.588 + ], + "angle": 0, + "content": "where \\( M(\\cdot) \\) measures the mean square error between two perceptual elements. And \\( \\tau \\) and \\( \\beta \\) are used to normalize it to [0,1] to match the distribution of semantic-level similarity. ADRScore-S more harshly measures the performance on spatial reasoning as the metric calculates the error of visual elements in spatial instead of text semantic. The latter is too lenient for visual predictions in language." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.611, + 0.381, + 0.629 + ], + "angle": 0, + "content": "4 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.784, + 0.673 + ], + "angle": 0, + "content": "In this section we introduce our framework in Sec. 4.1, followed by the training details provided in Sec. 4.2." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.695, + 0.421, + 0.71 + ], + "angle": 0, + "content": "4.1 Model architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.784, + 0.795 + ], + "angle": 0, + "content": "We observe that most VLMs struggle to effectively handle object-level perceptual information, including the input of visual priors and predictions of object locations, which are indispensable in autonomous driving scenarios. The limitation is primarily due to (i) the lack of a targeted tokenizer and (ii) decoder solely composed of a language model, resulting in subpar reasoning performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.786, + 0.841 + ], + "angle": 0, + "content": "To address this challenge, as illustrated in Fig. 6, we introduce a straightforward yet effective framework that enhances existing VLMs with two new components: a prior tokenizer and an instructed vision decoder. Notably, the design" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "Ming Nie, et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.143, + 0.788, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.305, + 0.789, + 0.377 + ], + "angle": 0, + "content": "Fig. 6: The pipeline of our framework. The input video and perceptual priors are tokenized using the vision encoder and prior tokenizer. A Q-former then aligns them to the text's feature space. The LLM and instructed vision decoder predict answers with precise perception results for user queries. The highlighted yellow box and red curve in the perception results respectively represent the visualization of \\( \\langle \\mathsf{LOC} \\rangle \\) and \\( \\langle \\mathsf{MOT} \\rangle \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.789, + 0.448 + ], + "angle": 0, + "content": "of the components is not for detection but to aid interpretable visual reasoning based on available perception inputs. These modules aim to strengthen the capabilities of the model to utilize object-level perceptual elements in both extracting visual priors and generating perceptual predictions for visual reasoning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.79, + 0.511 + ], + "angle": 0, + "content": "Vision encoder. Our model accepts both video frames and text inputs, along with perceptual priors, and tokenizes them into embeddings. For a sequence of video frames \\((V_{1}, V_{2}, \\dots, V_{N})\\), features are extracted using a pretrained Blip-2 visual encoder [24] \\(F_{v}\\) and aggregated through concatenation:" + }, + { + "type": "equation", + "bbox": [ + 0.366, + 0.532, + 0.787, + 0.55 + ], + "angle": 0, + "content": "\\[\nf _ {v} = F _ {v} \\left(V _ {1}\\right) \\oplus F _ {v} \\left(V _ {2}\\right) \\oplus \\dots \\oplus F _ {v} \\left(V _ {N}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.563, + 0.789, + 0.716 + ], + "angle": 0, + "content": "Prior tokenizer. We propose a novel tokenization strategy tailored to taking advantage of visual cues. The motivation is grounded in the acknowledgement that extracting and aligning visual features is considerably simpler and more suitable compared to compelling the LLM to comprehend ambiguous positional descriptions. Direct textual input to the LLM may result in challenges such as information loss, as textual representation may not fully capture image details and context, especially in complex scenarios with dynamic object positions and velocities. To tackle this issue, we design a novel tokenizer \\( F_{p} \\), implemented as a two-layer MLP, to independently extract local image features and positional embeddings from visual priors:" + }, + { + "type": "equation", + "bbox": [ + 0.425, + 0.736, + 0.787, + 0.755 + ], + "angle": 0, + "content": "\\[\nf _ {p} = F _ {p} (f _ {r} + E (P)), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.843 + ], + "angle": 0, + "content": "where \\( f_{r} \\) represents the region-level features extracted from the image-level features \\( f_{v} \\) according to the precise locations of perception priors \\( P \\). These features are aligned to \\( 7 \\times 7 \\) size using the RoIAlign [14] operation and fused into a single embedding \\( f_{r} \\). And \\( E(\\cdot) \\) is a positional encoding function mapping the geometry locations and motions into the same dimension of \\( f_{r} \\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.638, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Reason2Drive" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "LLM. After we tokenize the video and perception priors into embedding \\( f_{v} \\) and \\( f_{p} \\), a projector \\( Q \\) (Q-former [24] in this work) is adopted to align the non-text features into textual domain:" + }, + { + "type": "equation", + "bbox": [ + 0.45, + 0.209, + 0.785, + 0.223 + ], + "angle": 0, + "content": "\\[\nf _ {q} = Q \\left(f _ {v}, f _ {p}\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.229, + 0.784, + 0.261 + ], + "angle": 0, + "content": "Then, to generate the final text output, we utilize the LLM for further language processing with the extracted text embedding \\( f_{t} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.452, + 0.277, + 0.785, + 0.291 + ], + "angle": 0, + "content": "\\[\n\\hat {y} _ {t} = F \\left(f _ {t}, f _ {q}\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.296, + 0.785, + 0.538 + ], + "angle": 0, + "content": "Instructed vision decoder. Current works [11, 16] treat the LLM as a versatile tool to generate answers and inferences without intermediate reasoning steps, let alone considering the perceptions of the agent toward driving scenes. However, the perception ability of the agent towards driving scenarios is an indispensable part of a reliable driving procedure. Moreover, recent works [23] have demonstrated that, rather than training with textualized perceptual sequences, incorporating the perception abilities into the multi-modal LLM brings a significant improvement. To this end, inspired by [23], we integrate new perception capabilities into the multi-modal LLM. Specifically, we expand the original LLM vocabulary by introducing new tokens as placeholders, denoted as \\( \\langle \\mathsf{LOC} \\rangle \\) and \\( \\langle \\mathsf{MOT} \\rangle \\), to signify the request for the perception output. When the LLM aims to generate a specific perception, the output \\( \\hat{y}_t \\) should include a designed token. We then extract the last-layer textual features corresponding to the specific token and apply an MLP projection layer to obtain the hidden embedding \\( f_h \\). Finally, the textual embedding and visual features are fed into the instructed vision decoder to decode the predictions:" + }, + { + "type": "equation", + "bbox": [ + 0.451, + 0.554, + 0.784, + 0.569 + ], + "angle": 0, + "content": "\\[\n\\tilde {P} = D \\left(f _ {v}, f _ {h}\\right). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.575, + 0.784, + 0.607 + ], + "angle": 0, + "content": "This module is comprised of a transformer decoder for features alignment [3] and task-specific heads designed to generate object locations and motions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.628, + 0.391, + 0.643 + ], + "angle": 0, + "content": "4.2 Training details" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.652, + 0.784, + 0.684 + ], + "angle": 0, + "content": "Training objectives. The model is trained end-to-end using the text generation loss \\(\\mathcal{L}_{txt}\\) and the perception output loss \\(\\mathcal{L}_{per}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.429, + 0.699, + 0.784, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {t x t} + \\lambda_ {p e r} \\mathcal {L} _ {p e r}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.784, + 0.795 + ], + "angle": 0, + "content": "where \\(\\lambda_{per}\\) is the balancing term. Specifically, \\(\\mathcal{L}_{txt}\\) is the auto-regressive cross-entropy loss for text generation, and \\(\\mathcal{L}_{per}\\) encourages the instructed vision decoder to generate accurate locations and motions, which is similar to traditional detection loss and is employed with the combination of binary cross-entropy loss and MSE loss. More details are included in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.796, + 0.784, + 0.841 + ], + "angle": 0, + "content": "Tuning strategy. Our tuning strategy consists of two stages: the pre-training stage and the fine-tuning stage. In the pre-training stage, we initialize the weights from instructBLIP [10], including the pre-trained vision encoder, Q-former and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "Ming Nie, et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 3: Results of different models on the Reason2Drive validation set. We evaluate with ADRScore as well as captioning-based metrics." + }, + { + "type": "table", + "bbox": [ + 0.278, + 0.178, + 0.731, + 0.34 + ], + "angle": 0, + "content": "
MethodsLLMReasoning metricCaptioning metric
ADRScoreADRScore-SB@4METEORROUGECIDEr
Blip-2 [24]OPT-2.7B [45]0.2960.1620.3610.2490.4430.174
FlanT5-XL [8]0.3100.1710.3680.2560.4510.187
InstructBLIP [10]FlanT5-XL0.3290.1870.3760.2690.4620.196
Vicuna-7B [32]0.3510.2140.4080.2940.4840.211
MiniGPT-4 [46]Vicuna-7B0.3380.2030.3960.2860.4750.219
OursFlanT5-XL0.4570.4200.4510.3490.5200.292
Vicuna-7B0.4630.4320.4570.3560.5290.298
" + }, + { + "type": "table_caption", + "bbox": [ + 0.29, + 0.344, + 0.711, + 0.358 + ], + "angle": 0, + "content": "Table 4: Ablations on different combinations of training tasks." + }, + { + "type": "table", + "bbox": [ + 0.278, + 0.362, + 0.731, + 0.484 + ], + "angle": 0, + "content": "
TasksReasoning metricCaptioning metric
PerceptionPredictionReasoningADRScoreADRScore-SB@4METEORROUGECIDEr
✓0.2820.2530.4220.3070.4790.226
✓✓0.2970.2640.4190.3100.4790.228
✓0.3510.3230.4300.3250.4950.263
✓✓0.4070.3640.4350.3370.5010.274
✓✓✓0.4630.4320.4570.3560.5290.298
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.498, + 0.788, + 0.621 + ], + "angle": 0, + "content": "LLM, and freeze the parameters of LLM and vision tokenizer \\( F_{v} \\). We train the prior tokenizer \\( F_{p} \\) and Q-former \\( Q \\) to align visual priors with text, along with the instructed vision decoder \\( D \\) to enhance visual localization capabilities. The fine-tuning phase equips the LLM with reasoning abilities in autonomous driving using the instructed vision decoder. To retain pre-trained LLM generalization, we employ efficient fine-tuning with LoRA [19]. The vision encoder and prior tokenizer \\( F_{p} \\) remain fixed, while the instructed vision decoder \\( D \\) is fully fine-tuned. Word embeddings of the LLM and Q-former are also trainable." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.643, + 0.377, + 0.66 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.788, + 0.735 + ], + "angle": 0, + "content": "We benchmark various baseline models and present our method on Reason2Drive dataset. Sec. 5.1 covers implementation details. We assess reasoning performance using our proposed metric in Sec. 5.2, perform ablation studies in Sec. 5.3 and provide more ablations and qualitative results in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.755, + 0.436, + 0.772 + ], + "angle": 0, + "content": "5.1 Experimental setting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Our main experiments are carried out on the complete Reason2Drive benchmark. The dataset is collected from three different source datasets: nuScenes [2], Waymo [37], and ONCE [28]. It is divided into training and validation sets based on segments, with \\(70\\%\\) allocated to the training set and \\(30\\%\\) to the validation" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.638, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Reason2Drive" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.332, + 0.128, + 0.673, + 0.141 + ], + "angle": 0, + "content": "Table 5: The quality of predicted visual elements." + }, + { + "type": "table", + "bbox": [ + 0.278, + 0.145, + 0.731, + 0.198 + ], + "angle": 0, + "content": "
PredictionsMetricMiniGPT-4Kosmos-2 [33]Ours
Bounding boxAccuracy ↑0.7230.7450.806
TrajectoryADE ↓2.3342.5631.875
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.204, + 0.788, + 0.369 + ], + "angle": 0, + "content": "set, ensuring no overlap in scenes between them. The validation set also contains all the declared tasks and has been augmented. The input consists of 5 frames of cropped images with a size of \\(224 \\times 224\\) pixels. During training, we leverage the AdamW [26] optimizer with a weight decay of 0.01. We adopt a cosine learning rate decay scheduler with a max value of 3e-4 and a linear warm-up for the first 1000 iterations. The weight of perception loss \\(\\lambda_{p}\\) is set to 1.0. The normalization parameters \\(\\tau\\) and \\(\\beta\\) are selected to be 15 and 10 after empirical practice. Our models are trained with a batch size of 8 on 8 V100 GPUs. The baseline models are fine-tuned following the official tuning strategy, with the inputs consistent with our approach. Specifically, the perception priors are also provided as textual inputs to baseline models for fair comparison." + }, + { + "type": "table_caption", + "bbox": [ + 0.24, + 0.38, + 0.761, + 0.407 + ], + "angle": 0, + "content": "Table 6: Ablations on visual input and Table 7: Ablations on different settings of instructed vision decoder." + }, + { + "type": "table", + "bbox": [ + 0.245, + 0.42, + 0.501, + 0.491 + ], + "angle": 0, + "content": "
Visual features\nimage-level video-levelPerceptual priors\nregion-level positionalADRScoreADRScore-S
✓0.4140.379
0.4310.394
✓0.4470.418
✓0.4630.432
" + }, + { + "type": "table", + "bbox": [ + 0.51, + 0.42, + 0.754, + 0.492 + ], + "angle": 0, + "content": "
Pre-train text embedding MLPADRScoreADRScore-S
0.3870.361
✓0.4210.396
✓✓0.4550.425
✓✓✓0.4630.432
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.501, + 0.546, + 0.516 + ], + "angle": 0, + "content": "5.2 Reasoning performance evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.66 + ], + "angle": 0, + "content": "As demonstrated in Tab. 3, we evaluate both ADRScore and traditional caption-based performance of different models on our benchmark. It is worth noting that our method outperforms others comprehensively in all metrics. We also observe that, despite there is a correlation between ADRScore and traditional metrics, while on the other hand, the performance gap is more pronounced in our metrics, specially revealed by ADRScore-S. The results further substantiate reasoning ambiguities in traditional metrics, constraining the differentiation in benchmarking model performance. Specifically, models with varying reasoning capabilities exhibit minimal disparities when evaluated using traditional metrics." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.681, + 0.387, + 0.696 + ], + "angle": 0, + "content": "5.3 Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Task contributions. To investigate the synergies between different tasks, we separate and evaluate various types of tasks independently. As shown in Tab. 4, we train our model on different combinations of tasks. Most notably, training on reasoning tasks plays the most important role, indicating the necessity of relevant reasoning data for instructional tuning. Based on reasoning tasks, perception and prediction tasks additionally enhance the models for visual reasoning, showing specific improvements of \\(4.1\\%\\) and \\(6.8\\%\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Quality of predicted visual elements. To validate the quality of predicted perceptual elements, we also conduct experiments to evaluate the bounding" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "Ming Nie, et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.289, + 0.145, + 0.711, + 0.159 + ], + "angle": 0, + "content": "Table 8: Evaluation of control signals. B: B@4. M: METHOR." + }, + { + "type": "table", + "bbox": [ + 0.277, + 0.16, + 0.731, + 0.292 + ], + "angle": 0, + "content": "
MethodLLMB↑M↑RMSE
Speed↓Steer↓
Directly fine-tuned with control signals:
InstructBLIPVicuna-7B0.1660.2013.7435.926
Additionally pre-trained on Reason2Drive:
InstructBLIPVicuna-7B0.1920.2373.0865.151
OursVicuna-7B0.2130.2692.8424.866
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.294, + 0.784, + 0.368 + ], + "angle": 0, + "content": "boxes and trajectories separately, as shown in Tab. 5. To ensure a fair comparison, we implement Kosmos-2 [33] due to its advantageous grounding capabilities. The model is fine-tuned on our dataset following the official strategy. Experimental results confirm the high quality of our predicted visual elements, with particular emphasis on the accuracy of the trajectories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.369, + 0.785, + 0.443 + ], + "angle": 0, + "content": "The effects of tokenizers. To verify the effectiveness of the tokenizers, we conduct ablation studies to pinpoint where the improvements come from in Tab. 6. Visual features from single frame to multi-frame bring \\(1.5\\%\\) improvement in ADRScore-S. Perceptual priors, i.e., region-level features and positional embeddings bring \\(2.4\\%\\) and \\(1.4\\%\\) development." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.785, + 0.504 + ], + "angle": 0, + "content": "The effects of instructed vision decoder. To verify the efficiency of our instructed vision decoder, we conduct an ablation study to compare it with other methods. As demonstrated in Tab. 7, pre-training and textual embedding bring the major contribution (3.5% and 2.9% in ADRScore-S)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.785, + 0.61 + ], + "angle": 0, + "content": "Downstream tasks. We are also interested in understanding how our benchmark will contribute to downstream tasks, such as predicting control signals. Following the approach in ADriver-I [22], we generate control signals on nuScenes and fine-tune InstructBLIP [10] directly for planning signal prediction. To ablate the influence of our dataset, we also pre-train models on Reason2Drive before fine-tuning with control signals. The results presented in Tab. 8 demonstrate the supportive effect of Reason2Drive on downstream planning tasks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.631, + 0.36, + 0.646 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In summary, Large Vision-Language Models (VLMs) have sparked interest in autonomous driving for their advanced reasoning capabilities. However, the absence of datasets explaining decision-making processes hinders progress. To tackle this, we introduce Reason2Drive benchmark, comprising \\(600\\mathrm{K}+\\) video-text pairs for interpretable reasoning in complex driving scenarios. It outperforms existing datasets in scale, sources and task diversities. We also propose a novel evaluation protocol for chain-based reasoning, addressing existing semantic ambiguities. To uncover insights into their reasoning abilities, our work evaluates various VLMs and proposes an efficient method to boost the ability of models to utilize object-level perceptual elements in both the encoder and decoder. We expect our work could propel further advancements in interpretable reasoning for autonomous systems." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.638, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Reason2Drive" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.24 + ], + "angle": 0, + "content": "This work was supported in part by National Natural Science Foundation of China (Grant No. 62106050 and 62376060), Natural Science Foundation of Shanghai (Grant No. 22ZR1407500), USyd-Fudan BISA Flagship Research Program and Lingang Laboratory (Grant No. LG-QS-202202-07)." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.264, + 0.324, + 0.279 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.296, + 0.785, + 0.325 + ], + "angle": 0, + "content": "1. Banerjee, S., Lavie, A.: Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In: ACL workshop (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.326, + 0.785, + 0.366 + ], + "angle": 0, + "content": "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.367, + 0.785, + 0.394 + ], + "angle": 0, + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.395, + 0.785, + 0.422 + ], + "angle": 0, + "content": "4. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal lvm's referential dialogue magic. arXiv preprint (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.423, + 0.785, + 0.45 + ], + "angle": 0, + "content": "5. Chen, L., Wu, P., Chitta, K., Jaeger, B., Geiger, A., Li, H.: End-to-end autonomous driving: Challenges and frontiers. arXiv preprint (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.451, + 0.785, + 0.505 + ], + "angle": 0, + "content": "6. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.506, + 0.785, + 0.547 + ], + "angle": 0, + "content": "7. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. arXiv preprint (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.548, + 0.785, + 0.589 + ], + "angle": 0, + "content": "8. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.59, + 0.785, + 0.617 + ], + "angle": 0, + "content": "9. Contributors, D.: Drivelm: Drive on language. https://github.com/OpenDriveLab/DriveLM (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.618, + 0.785, + 0.659 + ], + "angle": 0, + "content": "0. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B.A., Fung, P., Hoi, S.C.H.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.66, + 0.785, + 0.687 + ], + "angle": 0, + "content": "1. Deruyttere, T., Grujicic, D., Blaschko, M.B., Moens, M.F.: Talk2car: Predicting physical trajectories for natural language commands. IEEE Access (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.688, + 0.785, + 0.729 + ], + "angle": 0, + "content": "2. Dewangan, V., Choudhary, T., Chandhok, S., Priyadarshan, S., Jain, A., Singh, A.K., Srivastava, S., Jatavallabhula, K.M., Krishna, K.M.: Talk2bev: Language-enhanced bird's-eye view maps for autonomous driving. arXiv preprint (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.785, + 0.771 + ], + "angle": 0, + "content": "3. Ding, X., Han, J., Xu, H., Zhang, W., Li, X.: Hilm-d: Towards high-resolution understanding in multimodal large language models for autonomous driving. arXiv preprint (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.602, + 0.785 + ], + "angle": 0, + "content": "4. Dollar, K., Girshick, R.: Mask r-cnn. In: ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "5. Feng, Q., Ablavsky, V., Sclaroff, S.: Cityflow-nl: Tracking and retrieval of vehicles at city scale by natural language descriptions. arXiv preprint (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "6. Fu, D., Li, X., Wen, L., Dou, M., Cai, P., Shi, B., Qiao, Y.: Drive like a human: Rethinking autonomous driving with large language models. arXiv preprint (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.296, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "Ming Nie, et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "17. Golovneva, O., Chen, M., Poff, S., Corredor, M., Zettlemoyer, L., Fazel-Zarandi, M., Celikyilmaz, A.: Roscoe: A suite of metrics for scoring step-by-step reasoning. arXiv preprint (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.23 + ], + "angle": 0, + "content": "18. Han, J., Zhang, R., Shao, W., Gao, P., Xu, P., Xiao, H., Zhang, K., Liu, C., Wen, S., Guo, Z., et al.: Imagebind-llm: Multi-modality instruction tuning. arXiv preprint (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.258 + ], + "angle": 0, + "content": "19. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.258, + 0.785, + 0.284 + ], + "angle": 0, + "content": "20. Hu, Y., Yang, J., Chen, L., Li, K., Sima, C., Zhu, X., Chai, S., Du, S., Lin, T., Wang, W., et al.: Planning-oriented autonomous driving. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.785, + 0.311 + ], + "angle": 0, + "content": "21. Jain, K., Chhangani, V., Tiwari, A., Krishna, K.M., Gandhi, V.: Ground then navigate: Language-guided navigation in dynamic scenes. In: ICRA (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.312, + 0.785, + 0.339 + ], + "angle": 0, + "content": "22. Jia, F., Mao, W., Liu, Y., Zhao, Y., Wen, Y., Zhang, C., Zhang, X., Wang, T.: Adriver-i: A general world model for autonomous driving. arXiv preprint (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.339, + 0.785, + 0.366 + ], + "angle": 0, + "content": "23. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: Lisa: Reasoning segmentation via large language model. arXiv preprint (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.366, + 0.785, + 0.406 + ], + "angle": 0, + "content": "24. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.406, + 0.785, + 0.42 + ], + "angle": 0, + "content": "25. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.433 + ], + "angle": 0, + "content": "26. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.785, + 0.473 + ], + "angle": 0, + "content": "27. Malla, S., Choi, C., Dwivedi, I., Choi, J.H., Li, J.: Drama: Joint risk localization and captioning in driving. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.514 + ], + "angle": 0, + "content": "28. Mao, J., Niu, M., Jiang, C., Liang, H., Chen, J., Liang, X., Li, Y., Ye, C., Zhang, W., Li, Z., et al.: One million scenes for autonomous driving: Once dataset. arXiv preprint (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.785, + 0.542 + ], + "angle": 0, + "content": "29. Mao, J., Qian, Y., Zhao, H., Wang, Y.: Gpt-driver: Learning to drive with gpt. arXiv preprint (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.785, + 0.568 + ], + "angle": 0, + "content": "30. OpenAI: Gpt-4: A large-scale transformer-based language model (2023), https://www.openai.com/research/gpt-4, https://www.openai.com/research/gpt-4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.568, + 0.785, + 0.595 + ], + "angle": 0, + "content": "31. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a method for automatic evaluation of machine translation. In: ACL (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.595, + 0.785, + 0.622 + ], + "angle": 0, + "content": "32. Peng, B., Li, C., He, P., Galley, M., Gao, J.: Instruction tuning with gpt-4. arXiv preprint (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.622, + 0.785, + 0.65 + ], + "angle": 0, + "content": "33. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.785, + 0.676 + ], + "angle": 0, + "content": "34. Pi, R., Gao, J., Diao, S., Pan, R., Dong, H., Zhang, J., Yao, L., Han, J., Xu, H., Zhang, L.K.T.: Detgpt: Detect what you need via reasoning. arXiv preprint (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.677, + 0.785, + 0.717 + ], + "angle": 0, + "content": "35. Qian, T., Chen, J., Zhuo, L., Jiao, Y., Jiang, Y.G.: Nuscenes-qa: A multi-modal visual question answering benchmark for autonomous driving scenario. arXiv preprint (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "36. Sachdeva, E., Agarwal, N., Chundi, S., Roelofs, S., Li, J., Dariush, B., Choi, C., Kochenderfer, M.: Rank2tell: A multimodal driving dataset for joint importance ranking and reasoning. arXiv preprint (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.8 + ], + "angle": 0, + "content": "37. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "38. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.638, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Reason2Drive" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "39. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: CVPR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.203 + ], + "angle": 0, + "content": "40. Wu, D., Han, W., Wang, T., Dong, X., Zhang, X., Shen, J.: Referring multi-object tracking. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.788, + 0.231 + ], + "angle": 0, + "content": "41. Wu, D., Han, W., Wang, T., Liu, Y., Zhang, X., Shen, J.: Language prompt for autonomous driving. arXiv preprint (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.272 + ], + "angle": 0, + "content": "42. Xu, Z., Zhang, Y., Xie, E., Zhao, Z., Guo, Y., Wong, K.K., Li, Z., Zhao, H.: Drivegpt4: Interpretable end-to-end autonomous driving via large language model. arXiv preprint (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.3 + ], + "angle": 0, + "content": "43. Yu, P., Wang, T., Golovneva, O., Alkhamissy, B., Ghosh, G., Diab, M., Celikyilmaz, A.: Alert: Adapting language models to reasoning tasks. arXiv preprint (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.328 + ], + "angle": 0, + "content": "44. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.369 + ], + "angle": 0, + "content": "45. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.411 + ], + "angle": 0, + "content": "46. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.411 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Reason2Drive_ Towards Interpretable and Chain-based Reasoning for Autonomous Driving/66f968e7-ebed-40e1-a3ff-b107a9813c0e_origin.pdf b/2024/Reason2Drive_ Towards Interpretable and Chain-based Reasoning for Autonomous Driving/66f968e7-ebed-40e1-a3ff-b107a9813c0e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..261843c08fd987579eae6a425542cc6e6565c4d4 --- /dev/null +++ b/2024/Reason2Drive_ Towards Interpretable and Chain-based Reasoning for Autonomous Driving/66f968e7-ebed-40e1-a3ff-b107a9813c0e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f1d88426c6671c873767feeb6f906908323f46c00e3ab2d74f8a64e37068eda8 +size 726938 diff --git a/2024/Reason2Drive_ Towards Interpretable and Chain-based Reasoning for Autonomous Driving/full.md b/2024/Reason2Drive_ Towards Interpretable and Chain-based Reasoning for Autonomous Driving/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b7d7384dd859948fc1975b573b551f9c7e2c386a --- /dev/null +++ b/2024/Reason2Drive_ Towards Interpretable and Chain-based Reasoning for Autonomous Driving/full.md @@ -0,0 +1,330 @@ +# Reason2Drive: Towards Interpretable and Chain-based Reasoning for Autonomous Driving + +Ming Nie $^{1}$ , Renyuan Peng $^{1}$ , Chunwei Wang $^{2}$ , Xinyue Cai $^{2}$ , Jianhua Han $^{2}$ , Hang Xu $^{2\star}$ , and Li Zhang $^{1\star}$ + +$^{1}$ School of Data Science, Fudan University $^{2}$ Huawei Noah's Ark Lab + +https://github.com/fudan-zvg/reason2drive + +Abstract. Large vision-language models (VLMs) have garnered increasing interest in autonomous driving areas, due to their advanced capabilities in complex reasoning tasks essential for highly autonomous vehicle behavior. Despite their potential, research in autonomous systems is hindered by the lack of datasets with annotated reasoning chains that explain the decision-making processes in driving. To bridge this gap, we present Reason2Drive, a benchmark dataset with over 600K video-text pairs, aimed at facilitating the study of interpretable reasoning in complex driving environments. We distinctly characterize the autonomous driving process as a sequential combination of perception, prediction, and reasoning steps, and the question-answer pairs are automatically collected from a diverse range of open-source outdoor driving datasets, including nuScenes, Waymo and ONCE. Moreover, we introduce a novel aggregated evaluation metric to assess chain-based reasoning performance in autonomous systems, addressing the reasoning ambiguities of existing metrics such as BLEU and CIDEr. Based on the proposed benchmark, we conduct experiments to assess various existing VLMs, revealing insights into their reasoning capabilities. Additionally, we develop an efficient approach to empower VLMs to leverage object-level perceptual elements in both feature extraction and prediction, further enhancing their reasoning accuracy. Extendable experiments demonstrate the supportive effect of Reason2Drive towards visual reasoning and downstream planning tasks. + +# 1 Introduction + +Modern autonomous driving systems face challenges related to generalization issues across diverse scenarios, which is often attributed to the reliance on empirical and intricate rules involved in decision-making. To reduce dependence on such rules, recent end-to-end approaches [5, 20] have been developed to derive control signals directly from sensor inputs, treating the system as a black + +box that requires extensive data for training. However, this approach tends to obscure the underlying logic of decisions, complicating failure diagnosis in real-world applications. In contrast, Large Vision-Language Models (VLMs) offer a promising alternative, potentially enhancing interpretability and generalization for these systems. With their broad world knowledge and advanced reasoning abilities, as illustrated in Fig. 1(a), VLMs have the potential to provide a more thorough understanding and explicit explanation for reliable decision-making. + +Nonetheless, existing works [35, 40] primarily focused on the straightforward adaptation of question-answering tasks to the autonomous driving; how to exploit VLMs to facilitate the reasoning abilities of autonomous systems is still under exploration. + +One reason that hinders the research in this field lies in the scarcity of datasets, especially those chained-based reasoning labels that elucidate the decision-making process. Most existing datasets [11, 35, 41] often oversimplify the complex processes of driving into straightforward question-answering tasks with only a few specific tasks covered. As depicted in Fig. 1(b), they typically provide closed-form annotations constrained + +to boolean (i.e., yes or no) answers or limited multiple-choice responses (e.g., stopped, parked, and moving). However, autonomous driving transcends a simplistic QA process. It encompasses a multi-step approach involving perception, prediction, and reasoning, each of which plays an indispensable role in the decision-making. Therefore, it is crucial to introduce a novel benchmark annotated with detailed decision-making reasoning for assessing the reasoning abilities of current VLMs. + +To this end, we introduce Reason2Drive, a new benchmark comprising over 600K video-text pairs, characterized by intricate driving instructions and a series of perception, prediction and reasoning steps. Our benchmark builds upon widely-used open-source driving datasets including nuScenes [2], Waymo [37], and ONCE [28], utilizing an extensible annotation schema. Specifically, we structure the comprehensive annotations into object-centric database and integrate it into manually crafted templates to create paired data for VLMs at both object and scenario levels. To enhance diversity, GPT-4 and manual instructions are employed for verification and enrichment purposes. Notably, Reason2Drive is the most extensive dataset available to date, outperforming existing datasets in scale and the complexity of reasoning chains included, which is a distinctive attribute not present in other datasets. + +![](images/36d359993802073393053a61f1d11b3beadec8fc472636ff12379307d698ff20.jpg) +Fig.1: (a) Different decision-making processes in autonomous driving. (b) Language-based dataset comparison. + +Furthermore, we observe a fundamental flaw in the current evaluation of VLMs on autonomous driving tasks, due to the inherent reasoning ambiguities of traditional caption-based metrics like BLEU [31] and CIDEr [39]. These metrics mainly measure text generation from a holistic perspective, without considering the causal relationship between the reasoning steps and the final conclusion. For example, when the VLM suggests ego vehicle turning left, we cannot ascertain from these metrics whether its reasoning steps effectively support the final decision. To address this issue, we propose a new aggregated evaluation metric, ADRScore, specifically designed to measure chain-based reasoning performance in autonomous systems, which aims to resolve the reasoning ambiguities associated with current metrics. + +Utilizing the proposed benchmark, we undertake experiments to assess various existing VLMs, thereby unveiling valuable insights into their reasoning capabilities. We find that most methods struggle to effectively leverage perceptual priors, resulting in subpar reasoning performance. Additionally, constrained by the language model functioning solely as a decoder, these methods often fail to deliver accurate perceptual results, which is a crucial component for verifying a model's spatial reasoning capability. To alleviate this dilemma, we present a simple yet efficient framework, augmenting existing VLMs with two new components: a prior tokenizer and an instructed vision decoder, which aim to bolster the models' visual localization capabilities within the encoder and decoder, respectively. Extendable experiments demonstrate the supportive effect of Reason2Drive towards visual reasoning and downstream planning tasks. + +The contributions of this paper are summarized as follows: (i) We publish a novel visual instruction tuning dataset aimed at facilitating interpretable and chain-based reasoning autonomous systems. (ii) We introduce a novel evaluation metric, ADRScore, to assess chain-based reasoning performance in autonomous driving, effectively addressing the reasoning ambiguities present in existing metrics. (iii) We conduct experiments to assess a range of existing VLMs, revealing valuable insights into their reasoning capabilities. (iv) To address the challenges posed by inefficient prior feature extraction and inaccurate perceptual predictions, we introduce an efficient approach to integrate these elements into VLMs. This results in a substantial improvement in reasoning accuracy and provides remarkable support for downstream planning tasks. Our method surpasses all baselines, notably achieving impressive generalization in unseen scenarios. + +# 2 Related Work + +Multimodal large language model. The current state of large language models provides remarkable abilities in natural language understanding and generation ([6, 7, 30, 38]). Inspired by the potential of large language models, a multitude of multimodal models has emerged, intended to enhance these models' capabilities in achieving multi-modal comprehension. Blip-2 [24] aligns visual and language features by utilizing a learnable Q-former. LLaVA [25] and MiniGPT-4 [46] initially align image-text features and then proceed with in + +![](images/48d2d27052970c57e8dd3a8226dfb8e7af4194e5f0c9748d58ab05e2626edee6.jpg) +Fig. 2: Schema of Our Reason2Drive Dataset. The upper part illustrates the pipeline for the automated construction of datasets. The lower part shows detailed instances of perception, prediction, and reasoning, accompanied by outcomes after applying GPT-4 for data augmentation. The special tokens hold distinct definitions: represents a specified instance, signifies a forecasted sequence of trajectory coordinates, and denotes positional coordinates. The colors associated with these tokens correspond to the highlighted objects in the upper-left image's boxes. + +struction tuning. Additionally, Video-LLaMA [44] and ImageBind-LLM [18] integrate multiple modalities into the input, aligning features from various sources like images, videos, audio, and point clouds, consolidating them into the space of language features. Kosmos-2 [33] and Shikra [4] perform object detection based on instructions and also accomplish grounded visual question answering. DetGPT [34] connects a fixed multi-modal LLM with a customizable detector based on user instructions. LISA [23] efficiently embeds segmentation abilities into multi-modal LLMs, showcasing self-reasoning for current perception systems. The previous works have demonstrated that current large-scale multimodal models can achieve cross-modal alignment, enabling comprehension and inference towards images and more. These models can not only perform perceptual tasks like detection but also accomplish preliminary reasoning tasks. + +Vision language tasks in autonomous driving. Currently, VLMs have demonstrated robust capabilities in scene perception and understanding. Extensive efforts have been dedicated to the realm of autonomous driving, leveraging VLM to achieve comprehensive scene understanding and perform diverse tasks [13, 16, 29, 42]. Simultaneously, substantial works are in progress to create datasets and models tailored to various tasks. Talk2Car [11] proposes the first object referral dataset for grounding commands for self-driving cars in free natural language into the visual context. But it exclusively contains information about visible objects. While DRAMA [27] outlines the overall scene risk, it lacks pre + +cise perception annotation. NuPrompt [41] and Refer-KITTI [40] offer language prompt sets for driving scenes but primarily concentrate on multi-object tracking tasks. NuScenesQA [35] and DriveLM [9] build visual question-answering (VQA) datasets for scenario understanding. However, their primary emphasis is on the perceptual information in the scene, lacking annotations for the analysis and complex reasoning of the entire scenario. To address the limitations of existing works, we construct a thorough dataset covering perception, prediction, and complex reasoning, additionally with an improved vision-language model for better analyzing autonomous driving scenarios. + +# 3 Reason2Drive Dataset + +We introduce Reason2Drive, a dataset that comprises comprehensive driving instructions and a chain-based reasoning framework for decision-making. Our dataset is characterized by the following key aspects: + +- Quantity: It stands out as the largest language-based driving dataset available, collated from prominent publicly accessible datasets worldwide. +- **Quality:** Reason2Drive offers a more precise representation of driving activities, including perception, prediction and reasoning, with a reliable auto-annotation schema for data collection. +- Diversity: (i) The dataset exhibits a broader range of scenes, encompassing both object-level and scenario-level data. This diversity includes object types, visual and motion attributes, object locations, and relationships relative to the ego-vehicle. (ii) It includes more intricate question-answer pairs, enhanced by GPT-4, along with longer text passages featuring step-by-step reasoning. +- Protocols: A novel evaluation metric is introduced to assess the reasoning capabilities of VLMs. Different from those widely used in the NLP community, it takes into account not only perception results but also reasoning ambiguities, providing a more holistic evaluation of the VLM's reasoning capacity for autonomous driving scenarios. + +Further details regarding the data collection process, statistical data analysis, and benchmark protocols are provided in the subsequent section. + +# 3.1 Dataset collection + +As illustrated in Fig. 2, we employ an extensible annotation schema, constructing data in the forms of question-answer pairs. Specifically, we first leverage a diverse array of publicly available datasets collected in different regions worldwide, including nuScenes, Waymo, and ONCE, and then parse their comprehensive annotations into object-centric database. The database is organized in key frames, and each frame stores object entries containing various details pertaining to its + +
DatasetDescriptionSource Datasets
Talk2Car [11]Object referralnuScenes
CityFlow-NL [15]Tracking & retrievalCityFlow
CARLA-NAV [21]Segmentation & predictionCARLA Simulator
NuPrompt [41]Multi-object trackingnuScenes
NuScenes-QA [35]PerceptionnuScenes
Refer-KITTI [40]Multi-object trackingKITTI
Talk2BEV [12]Visual understandingnuScenes
DRAMA [27]Risk localizationself-collected
Rank2Tell [36]Risk localization & rankingself-collected
Reason2DrivePerceptionnuScenes
PredictionWaymo
ReasoningONCE
+ +Table 1: The comparison between our Reason2Drive dataset and other prompt-based datasets. ■ means dataset not published. + +![](images/4834d66a5938ded50e5fe869078d9bbf875b2d58a647a8b912d184c30b13e0aa.jpg) +Fig. 3: Data quality comparison. Reason2Drive is larger in scale, richer in data content, and more diverse in scenarios. + +driving actions, including location, category, attributes and more. Afterwards, the structured database is required to answer the manually crafted question templates, which are divided into different tasks (i.e., perception, prediction and reasoning) at both object-level and scenario-level. Subsequently, GPT-4 is involved for verification and enrichment purposes. The example of GPT augmented data and manual instructions are provided in the appendix. + +Due to the complexity of autonomous driving activities, we categorize the tasks into three distinct groups to acquire diversified data: perception, prediction and reasoning. The specifics and distinctions of these three types of tasks are elaborated as follows: + +- Perception task is designed to identify objects within the driving scenario, assessing the fundamental perceptual capabilities of VLMs in outdoor environments. +- Prediction task entails the prediction of future states of key objects within the perceptual range, challenging VLMs to infer the intentions of objects with video input. +- Reasoning task prompts the analysis of the current perceptual and predicted states step by step, requiring the deduction of reasoned inferences and decisions through a chain of thoughts (COT) approach. + +For each task, we further categorize the data into object-level and scenario-level. In more detail, + +- Object-level data is formatted to benchmark the foundational capabilities of specific objects. As for perception, we address the location and attributes of objects such as moving status and distance to ego, while for prediction, future motion and merging-in/out status are considered. +- Scenario-level data is organized from a global perspective towards driving environment and ego-driving instructions. It focuses on whether there is an object worth noting currently (perception), whether there is an object worth noting in the future (prediction) and why (reasoning). For example, + +![](images/d62edde5d1b5ac8f3529cf28c302e8770bd37e8424f0fca8c9f487722bc10952.jpg) +Fig. 4: Statistical distribution of different tasks in Reason2Drive. + +
Task TargetPerception (PE) Prediction (PR) Reasoning (RE)Total
Ego vehicle22629171731886858670
Single object718826166772006205555
Multi objects1020129450252175248689
Scenario495894279527657120041
Total246112216137170706632955
+ +Table 2: The statistics of different tasks in Reason2Drive dataset. + +as illustrated in Fig. 2, models are asked to identify distances, merging states and other risks from the whole scene. It verifies the agent's ability to perceive the entire driving scene rather than specifying objects, thus more challenging. + +# 3.2 Dataset analysis + +Tab. 1 and Fig. 3 demonstrate the comparison between our Reason2Drive dataset and existing benchmarks. It is noteworthy that our benchmark stands as the largest dataset to date, surpassing others in terms of both dataset size and the inclusion of extensive long-text chain-based reasoning references. + +To further investigate the property of Reason2Drive dataset, we count the distribution and sample numbers of our dataset in Fig. 4 and Tab. 2. We split the dataset according to the task and target. The benchmark exhibits a balanced distribution, with multi-object tasks constituting the majority. Single-object and scenario-level questions are of similar quantities. The fewest questions are related to the ego-vehicle. Additionally, perception, prediction and reasoning questions are distributed as $39\%$ , $34\%$ , and $27\%$ , respectively. More dataset details are provided in the appendix. + +# 3.3 Benchmark protocol + +It is worth noting that previous works [12, 27, 35] simply utilize metric scores widely used in the NLP community, including BLEU [31], CIDEr [39] and ME-TEOR [1]. However, these metrics mainly measure text generation from a holistic perspective, without considering the causal relationship between the reasoning steps and the final conclusion. While they perform well in translation and captioning, their efficacy is limited when it comes to reasoning. Moreover, the existing evaluation system is confined to assessing the semantic quality of the results. It fails to effectively evaluate the quality of the perceived results, a crucial aspect supporting automatic driving decision-making. To address these dilemmas, inspired by [43] and [17], we develop a novel evaluation protocol, ADRScore, to measure the correctness of the reasoning chains towards autonomous driving. + +![](images/b1d1d82a41fd4b26479b0c0c4d605067948d0c83ad641f894b2da666ff1fbac3.jpg) +Fig. 5: Illustration of ADRScore and its visual adaptation. When substituting semantic similarities with actual geometric similarities, the score decreases. + +Preliminary. To begin with, we denote the generated reasoning steps as hypothesis $\pmb{h} = \{h_1, \dots, h_N\}$ , and the gold annotation as reference $\pmb{r} = \{r_1, \dots, r_K\}$ . + +At the core of reasoning metrics is the reasoning alignment vector from the $N$ -step hypothesis $h$ to the $K$ -step reference: + +$$ +\operatorname {a l i g n} (\boldsymbol {h} \rightarrow \boldsymbol {r}) = \left\{\alpha_ {1}, \dots , \alpha_ {N} \right\}, \tag {1} +$$ + +where alignment value $\alpha_{i}$ represents the semantic similarity between the corresponding hypothesis step and the most similar reference step: + +$$ +\alpha_ {i} = \max _ {j = 1} ^ {K} s _ {i, j}, +$$ + +$$ +s _ {i, j} = \cos \left(h _ {i}, r _ {j}\right). \tag {2} +$$ + +$\alpha_{i} \in [0,1]$ explicitly measures the grounding of the step-wise reasoning with respect to the reference, and $\cos(\cdot)$ denotes the cosine similarity between the corresponding sentence embeddings, which are extracted a pre-trained bert model. Based on the above reasoning alignment vector, we propose the following metrics to thoroughly measure the quality of reasoning steps. + +Reasoning alignment. The most straightforward way to evaluate the correctness of the hypothesis reasoning chain is to compare the degree of overlap between the hypothesis and the reference. One way of doing that is to measure the reasoning alignment between them: + +$$ +R A = \frac {1}{N} \sum_ {i = 1} ^ {N} \operatorname {a l i g n} \left(h _ {i} \rightarrow r\right). \tag {3} +$$ + +Redundancy. To find chains that contain information that is not required to solve the problem (i.e., redundant steps), we identify those hypothesis steps that + +are least aligned with the reference steps. This metric punishes the chain with steps that are not required for the correct solution. + +$$ +R D = \min _ {i = 1} ^ {N} \operatorname {a l i g n} \left(h _ {i} \rightarrow \boldsymbol {r}\right). \tag {4} +$$ + +Missing step. To identify steps that are missing from the hypothesis but could be required to solve the problem, we look at the alignment between reference and the hypothesis, similar to Redundancy. However, here we go through each step in the reference, and check if there is a similar step in the hypothesis: + +$$ +M S = \min _ {i = 1} ^ {K} \operatorname {a l i g n} \left(r _ {i} \rightarrow \boldsymbol {h}\right). \tag {5} +$$ + +Finally, the aggregated metric score is the average of the above performance, which is: + +$$ +A D R S c o r e = \frac {1}{3} (R A + R D + M S). \tag {6} +$$ + +Adapted to visual elements. To further adapt to the realistic driving process, we promote the above metric to the situation with visual elements, named ADRScore-S. Specifically, as illustrated in Fig. 5, when the hypothesis step $h_i$ and reference step $r_j$ contains visual elements, i.e., the locations and motions predicted for further reasoning, the similarity score becomes: + +$$ +s _ {i, j} = \frac {\tau - M \left(h _ {i} , r _ {j}\right)}{\beta}, \tag {7} +$$ + +where $M(\cdot)$ measures the mean square error between two perceptual elements. And $\tau$ and $\beta$ are used to normalize it to [0,1] to match the distribution of semantic-level similarity. ADRScore-S more harshly measures the performance on spatial reasoning as the metric calculates the error of visual elements in spatial instead of text semantic. The latter is too lenient for visual predictions in language. + +# 4 Methodology + +In this section we introduce our framework in Sec. 4.1, followed by the training details provided in Sec. 4.2. + +# 4.1 Model architecture + +We observe that most VLMs struggle to effectively handle object-level perceptual information, including the input of visual priors and predictions of object locations, which are indispensable in autonomous driving scenarios. The limitation is primarily due to (i) the lack of a targeted tokenizer and (ii) decoder solely composed of a language model, resulting in subpar reasoning performance. + +To address this challenge, as illustrated in Fig. 6, we introduce a straightforward yet effective framework that enhances existing VLMs with two new components: a prior tokenizer and an instructed vision decoder. Notably, the design + +![](images/2636b6b6bb2a3c6b070ed52f9a633f9e27a2f5b52f1d54dceedd0ae7b7023d21.jpg) +Fig. 6: The pipeline of our framework. The input video and perceptual priors are tokenized using the vision encoder and prior tokenizer. A Q-former then aligns them to the text's feature space. The LLM and instructed vision decoder predict answers with precise perception results for user queries. The highlighted yellow box and red curve in the perception results respectively represent the visualization of $\langle \mathsf{LOC} \rangle$ and $\langle \mathsf{MOT} \rangle$ . + +of the components is not for detection but to aid interpretable visual reasoning based on available perception inputs. These modules aim to strengthen the capabilities of the model to utilize object-level perceptual elements in both extracting visual priors and generating perceptual predictions for visual reasoning. + +Vision encoder. Our model accepts both video frames and text inputs, along with perceptual priors, and tokenizes them into embeddings. For a sequence of video frames $(V_{1}, V_{2}, \dots, V_{N})$ , features are extracted using a pretrained Blip-2 visual encoder [24] $F_{v}$ and aggregated through concatenation: + +$$ +f _ {v} = F _ {v} \left(V _ {1}\right) \oplus F _ {v} \left(V _ {2}\right) \oplus \dots \oplus F _ {v} \left(V _ {N}\right). \tag {8} +$$ + +Prior tokenizer. We propose a novel tokenization strategy tailored to taking advantage of visual cues. The motivation is grounded in the acknowledgement that extracting and aligning visual features is considerably simpler and more suitable compared to compelling the LLM to comprehend ambiguous positional descriptions. Direct textual input to the LLM may result in challenges such as information loss, as textual representation may not fully capture image details and context, especially in complex scenarios with dynamic object positions and velocities. To tackle this issue, we design a novel tokenizer $F_{p}$ , implemented as a two-layer MLP, to independently extract local image features and positional embeddings from visual priors: + +$$ +f _ {p} = F _ {p} (f _ {r} + E (P)), \tag {9} +$$ + +where $f_{r}$ represents the region-level features extracted from the image-level features $f_{v}$ according to the precise locations of perception priors $P$ . These features are aligned to $7 \times 7$ size using the RoIAlign [14] operation and fused into a single embedding $f_{r}$ . And $E(\cdot)$ is a positional encoding function mapping the geometry locations and motions into the same dimension of $f_{r}$ . + +LLM. After we tokenize the video and perception priors into embedding $f_{v}$ and $f_{p}$ , a projector $Q$ (Q-former [24] in this work) is adopted to align the non-text features into textual domain: + +$$ +f _ {q} = Q \left(f _ {v}, f _ {p}\right). \tag {10} +$$ + +Then, to generate the final text output, we utilize the LLM for further language processing with the extracted text embedding $f_{t}$ : + +$$ +\hat {y} _ {t} = F \left(f _ {t}, f _ {q}\right). \tag {11} +$$ + +Instructed vision decoder. Current works [11, 16] treat the LLM as a versatile tool to generate answers and inferences without intermediate reasoning steps, let alone considering the perceptions of the agent toward driving scenes. However, the perception ability of the agent towards driving scenarios is an indispensable part of a reliable driving procedure. Moreover, recent works [23] have demonstrated that, rather than training with textualized perceptual sequences, incorporating the perception abilities into the multi-modal LLM brings a significant improvement. To this end, inspired by [23], we integrate new perception capabilities into the multi-modal LLM. Specifically, we expand the original LLM vocabulary by introducing new tokens as placeholders, denoted as $\langle \mathsf{LOC} \rangle$ and $\langle \mathsf{MOT} \rangle$ , to signify the request for the perception output. When the LLM aims to generate a specific perception, the output $\hat{y}_t$ should include a designed token. We then extract the last-layer textual features corresponding to the specific token and apply an MLP projection layer to obtain the hidden embedding $f_h$ . Finally, the textual embedding and visual features are fed into the instructed vision decoder to decode the predictions: + +$$ +\tilde {P} = D \left(f _ {v}, f _ {h}\right). \tag {12} +$$ + +This module is comprised of a transformer decoder for features alignment [3] and task-specific heads designed to generate object locations and motions. + +# 4.2 Training details + +Training objectives. The model is trained end-to-end using the text generation loss $\mathcal{L}_{txt}$ and the perception output loss $\mathcal{L}_{per}$ : + +$$ +\mathcal {L} = \mathcal {L} _ {t x t} + \lambda_ {p e r} \mathcal {L} _ {p e r}, \tag {13} +$$ + +where $\lambda_{per}$ is the balancing term. Specifically, $\mathcal{L}_{txt}$ is the auto-regressive cross-entropy loss for text generation, and $\mathcal{L}_{per}$ encourages the instructed vision decoder to generate accurate locations and motions, which is similar to traditional detection loss and is employed with the combination of binary cross-entropy loss and MSE loss. More details are included in the appendix. + +Tuning strategy. Our tuning strategy consists of two stages: the pre-training stage and the fine-tuning stage. In the pre-training stage, we initialize the weights from instructBLIP [10], including the pre-trained vision encoder, Q-former and + +Table 3: Results of different models on the Reason2Drive validation set. We evaluate with ADRScore as well as captioning-based metrics. + +
MethodsLLMReasoning metricCaptioning metric
ADRScoreADRScore-SB@4METEORROUGECIDEr
Blip-2 [24]OPT-2.7B [45]0.2960.1620.3610.2490.4430.174
FlanT5-XL [8]0.3100.1710.3680.2560.4510.187
InstructBLIP [10]FlanT5-XL0.3290.1870.3760.2690.4620.196
Vicuna-7B [32]0.3510.2140.4080.2940.4840.211
MiniGPT-4 [46]Vicuna-7B0.3380.2030.3960.2860.4750.219
OursFlanT5-XL0.4570.4200.4510.3490.5200.292
Vicuna-7B0.4630.4320.4570.3560.5290.298
+ +Table 4: Ablations on different combinations of training tasks. + +
TasksReasoning metricCaptioning metric
PerceptionPredictionReasoningADRScoreADRScore-SB@4METEORROUGECIDEr
✓0.2820.2530.4220.3070.4790.226
✓✓0.2970.2640.4190.3100.4790.228
✓0.3510.3230.4300.3250.4950.263
✓✓0.4070.3640.4350.3370.5010.274
✓✓✓0.4630.4320.4570.3560.5290.298
+ +LLM, and freeze the parameters of LLM and vision tokenizer $F_{v}$ . We train the prior tokenizer $F_{p}$ and Q-former $Q$ to align visual priors with text, along with the instructed vision decoder $D$ to enhance visual localization capabilities. The fine-tuning phase equips the LLM with reasoning abilities in autonomous driving using the instructed vision decoder. To retain pre-trained LLM generalization, we employ efficient fine-tuning with LoRA [19]. The vision encoder and prior tokenizer $F_{p}$ remain fixed, while the instructed vision decoder $D$ is fully fine-tuned. Word embeddings of the LLM and Q-former are also trainable. + +# 5 Experiments + +We benchmark various baseline models and present our method on Reason2Drive dataset. Sec. 5.1 covers implementation details. We assess reasoning performance using our proposed metric in Sec. 5.2, perform ablation studies in Sec. 5.3 and provide more ablations and qualitative results in the appendix. + +# 5.1 Experimental setting + +Our main experiments are carried out on the complete Reason2Drive benchmark. The dataset is collected from three different source datasets: nuScenes [2], Waymo [37], and ONCE [28]. It is divided into training and validation sets based on segments, with $70\%$ allocated to the training set and $30\%$ to the validation + +Table 5: The quality of predicted visual elements. + +
PredictionsMetricMiniGPT-4Kosmos-2 [33]Ours
Bounding boxAccuracy ↑0.7230.7450.806
TrajectoryADE ↓2.3342.5631.875
+ +set, ensuring no overlap in scenes between them. The validation set also contains all the declared tasks and has been augmented. The input consists of 5 frames of cropped images with a size of $224 \times 224$ pixels. During training, we leverage the AdamW [26] optimizer with a weight decay of 0.01. We adopt a cosine learning rate decay scheduler with a max value of 3e-4 and a linear warm-up for the first 1000 iterations. The weight of perception loss $\lambda_{p}$ is set to 1.0. The normalization parameters $\tau$ and $\beta$ are selected to be 15 and 10 after empirical practice. Our models are trained with a batch size of 8 on 8 V100 GPUs. The baseline models are fine-tuned following the official tuning strategy, with the inputs consistent with our approach. Specifically, the perception priors are also provided as textual inputs to baseline models for fair comparison. + +Table 6: Ablations on visual input and Table 7: Ablations on different settings of instructed vision decoder. + +
Visual features +image-level video-levelPerceptual priors +region-level positionalADRScoreADRScore-S
✓0.4140.379
0.4310.394
✓0.4470.418
✓0.4630.432
+ +
Pre-train text embedding MLPADRScoreADRScore-S
0.3870.361
✓0.4210.396
✓✓0.4550.425
✓✓✓0.4630.432
+ +# 5.2 Reasoning performance evaluation + +As demonstrated in Tab. 3, we evaluate both ADRScore and traditional caption-based performance of different models on our benchmark. It is worth noting that our method outperforms others comprehensively in all metrics. We also observe that, despite there is a correlation between ADRScore and traditional metrics, while on the other hand, the performance gap is more pronounced in our metrics, specially revealed by ADRScore-S. The results further substantiate reasoning ambiguities in traditional metrics, constraining the differentiation in benchmarking model performance. Specifically, models with varying reasoning capabilities exhibit minimal disparities when evaluated using traditional metrics. + +# 5.3 Ablation study + +Task contributions. To investigate the synergies between different tasks, we separate and evaluate various types of tasks independently. As shown in Tab. 4, we train our model on different combinations of tasks. Most notably, training on reasoning tasks plays the most important role, indicating the necessity of relevant reasoning data for instructional tuning. Based on reasoning tasks, perception and prediction tasks additionally enhance the models for visual reasoning, showing specific improvements of $4.1\%$ and $6.8\%$ , respectively. + +Quality of predicted visual elements. To validate the quality of predicted perceptual elements, we also conduct experiments to evaluate the bounding + +Table 8: Evaluation of control signals. B: B@4. M: METHOR. + +
MethodLLMB↑M↑RMSE
Speed↓Steer↓
Directly fine-tuned with control signals:
InstructBLIPVicuna-7B0.1660.2013.7435.926
Additionally pre-trained on Reason2Drive:
InstructBLIPVicuna-7B0.1920.2373.0865.151
OursVicuna-7B0.2130.2692.8424.866
+ +boxes and trajectories separately, as shown in Tab. 5. To ensure a fair comparison, we implement Kosmos-2 [33] due to its advantageous grounding capabilities. The model is fine-tuned on our dataset following the official strategy. Experimental results confirm the high quality of our predicted visual elements, with particular emphasis on the accuracy of the trajectories. + +The effects of tokenizers. To verify the effectiveness of the tokenizers, we conduct ablation studies to pinpoint where the improvements come from in Tab. 6. Visual features from single frame to multi-frame bring $1.5\%$ improvement in ADRScore-S. Perceptual priors, i.e., region-level features and positional embeddings bring $2.4\%$ and $1.4\%$ development. + +The effects of instructed vision decoder. To verify the efficiency of our instructed vision decoder, we conduct an ablation study to compare it with other methods. As demonstrated in Tab. 7, pre-training and textual embedding bring the major contribution (3.5% and 2.9% in ADRScore-S). + +Downstream tasks. We are also interested in understanding how our benchmark will contribute to downstream tasks, such as predicting control signals. Following the approach in ADriver-I [22], we generate control signals on nuScenes and fine-tune InstructBLIP [10] directly for planning signal prediction. To ablate the influence of our dataset, we also pre-train models on Reason2Drive before fine-tuning with control signals. The results presented in Tab. 8 demonstrate the supportive effect of Reason2Drive on downstream planning tasks. + +# 6 Conclusion + +In summary, Large Vision-Language Models (VLMs) have sparked interest in autonomous driving for their advanced reasoning capabilities. However, the absence of datasets explaining decision-making processes hinders progress. To tackle this, we introduce Reason2Drive benchmark, comprising $600\mathrm{K}+$ video-text pairs for interpretable reasoning in complex driving scenarios. It outperforms existing datasets in scale, sources and task diversities. We also propose a novel evaluation protocol for chain-based reasoning, addressing existing semantic ambiguities. To uncover insights into their reasoning abilities, our work evaluates various VLMs and proposes an efficient method to boost the ability of models to utilize object-level perceptual elements in both the encoder and decoder. We expect our work could propel further advancements in interpretable reasoning for autonomous systems. + +# Acknowledgments + +This work was supported in part by National Natural Science Foundation of China (Grant No. 62106050 and 62376060), Natural Science Foundation of Shanghai (Grant No. 22ZR1407500), USyd-Fudan BISA Flagship Research Program and Lingang Laboratory (Grant No. LG-QS-202202-07). + +# References + +1. Banerjee, S., Lavie, A.: Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In: ACL workshop (2005) +2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: CVPR (2020) +3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV (2020) +4. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal lvm's referential dialogue magic. arXiv preprint (2023) +5. Chen, L., Wu, P., Chitta, K., Jaeger, B., Geiger, A., Li, H.: End-to-end autonomous driving: Challenges and frontiers. arXiv preprint (2023) +6. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023) +7. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. arXiv preprint (2022) +8. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint (2022) +9. Contributors, D.: Drivelm: Drive on language. https://github.com/OpenDriveLab/DriveLM (2023) +0. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B.A., Fung, P., Hoi, S.C.H.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint (2023) +1. Deruyttere, T., Grujicic, D., Blaschko, M.B., Moens, M.F.: Talk2car: Predicting physical trajectories for natural language commands. IEEE Access (2022) +2. Dewangan, V., Choudhary, T., Chandhok, S., Priyadarshan, S., Jain, A., Singh, A.K., Srivastava, S., Jatavallabhula, K.M., Krishna, K.M.: Talk2bev: Language-enhanced bird's-eye view maps for autonomous driving. arXiv preprint (2023) +3. Ding, X., Han, J., Xu, H., Zhang, W., Li, X.: Hilm-d: Towards high-resolution understanding in multimodal large language models for autonomous driving. arXiv preprint (2023) +4. Dollar, K., Girshick, R.: Mask r-cnn. In: ICCV (2017) +5. Feng, Q., Ablavsky, V., Sclaroff, S.: Cityflow-nl: Tracking and retrieval of vehicles at city scale by natural language descriptions. arXiv preprint (2021) +6. Fu, D., Li, X., Wen, L., Dou, M., Cai, P., Shi, B., Qiao, Y.: Drive like a human: Rethinking autonomous driving with large language models. arXiv preprint (2023) + +17. Golovneva, O., Chen, M., Poff, S., Corredor, M., Zettlemoyer, L., Fazel-Zarandi, M., Celikyilmaz, A.: Roscoe: A suite of metrics for scoring step-by-step reasoning. arXiv preprint (2022) +18. Han, J., Zhang, R., Shao, W., Gao, P., Xu, P., Xiao, H., Zhang, K., Liu, C., Wen, S., Guo, Z., et al.: Imagebind-llm: Multi-modality instruction tuning. arXiv preprint (2023) +19. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint (2021) +20. Hu, Y., Yang, J., Chen, L., Li, K., Sima, C., Zhu, X., Chai, S., Du, S., Lin, T., Wang, W., et al.: Planning-oriented autonomous driving. In: CVPR (2023) +21. Jain, K., Chhangani, V., Tiwari, A., Krishna, K.M., Gandhi, V.: Ground then navigate: Language-guided navigation in dynamic scenes. In: ICRA (2023) +22. Jia, F., Mao, W., Liu, Y., Zhao, Y., Wen, Y., Zhang, C., Zhang, X., Wang, T.: Adriver-i: A general world model for autonomous driving. arXiv preprint (2023) +23. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: Lisa: Reasoning segmentation via large language model. arXiv preprint (2023) +24. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint (2023) +25. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NIPS (2023) +26. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2017) +27. Malla, S., Choi, C., Dwivedi, I., Choi, J.H., Li, J.: Drama: Joint risk localization and captioning in driving. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (2023) +28. Mao, J., Niu, M., Jiang, C., Liang, H., Chen, J., Liang, X., Li, Y., Ye, C., Zhang, W., Li, Z., et al.: One million scenes for autonomous driving: Once dataset. arXiv preprint (2021) +29. Mao, J., Qian, Y., Zhao, H., Wang, Y.: Gpt-driver: Learning to drive with gpt. arXiv preprint (2023) +30. OpenAI: Gpt-4: A large-scale transformer-based language model (2023), https://www.openai.com/research/gpt-4, https://www.openai.com/research/gpt-4 +31. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a method for automatic evaluation of machine translation. In: ACL (2002) +32. Peng, B., Li, C., He, P., Galley, M., Gao, J.: Instruction tuning with gpt-4. arXiv preprint (2023) +33. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint (2023) +34. Pi, R., Gao, J., Diao, S., Pan, R., Dong, H., Zhang, J., Yao, L., Han, J., Xu, H., Zhang, L.K.T.: Detgpt: Detect what you need via reasoning. arXiv preprint (2023) +35. Qian, T., Chen, J., Zhuo, L., Jiao, Y., Jiang, Y.G.: Nuscenes-qa: A multi-modal visual question answering benchmark for autonomous driving scenario. arXiv preprint (2023) +36. Sachdeva, E., Agarwal, N., Chundi, S., Roelofs, S., Li, J., Dariush, B., Choi, C., Kochenderfer, M.: Rank2tell: A multimodal driving dataset for joint importance ranking and reasoning. arXiv preprint (2023) +37. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: CVPR (2020) +38. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint (2023) + +39. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: CVPR (2015) +40. Wu, D., Han, W., Wang, T., Dong, X., Zhang, X., Shen, J.: Referring multi-object tracking. In: CVPR (2023) +41. Wu, D., Han, W., Wang, T., Liu, Y., Zhang, X., Shen, J.: Language prompt for autonomous driving. arXiv preprint (2023) +42. Xu, Z., Zhang, Y., Xie, E., Zhao, Z., Guo, Y., Wong, K.K., Li, Z., Zhao, H.: Drivegpt4: Interpretable end-to-end autonomous driving via large language model. arXiv preprint (2023) +43. Yu, P., Wang, T., Golovneva, O., Alkhamissy, B., Ghosh, G., Diab, M., Celikyilmaz, A.: Alert: Adapting language models to reasoning tasks. arXiv preprint (2022) +44. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint (2023) +45. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint (2022) +46. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint (2023) \ No newline at end of file diff --git a/2024/Reason2Drive_ Towards Interpretable and Chain-based Reasoning for Autonomous Driving/images.zip b/2024/Reason2Drive_ Towards Interpretable and Chain-based Reasoning for Autonomous Driving/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a3bbe344d08dd6fbe360a55fd02db95b21e5bb24 --- /dev/null +++ b/2024/Reason2Drive_ Towards Interpretable and Chain-based Reasoning for Autonomous Driving/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:360e97451c5d1300953e2c441f24b0ff3788255f5dc36370e2ecee3c9f0c5893 +size 515626 diff --git a/2024/Reason2Drive_ Towards Interpretable and Chain-based Reasoning for Autonomous Driving/layout.json b/2024/Reason2Drive_ Towards Interpretable and Chain-based Reasoning for Autonomous Driving/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8d5ccccde48f5da4417a21240563b0047c23746d --- /dev/null +++ b/2024/Reason2Drive_ Towards Interpretable and Chain-based Reasoning for Autonomous Driving/layout.json @@ -0,0 +1,8780 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "type": "text", + "content": "Reason2Drive: Towards Interpretable and Chain-based Reasoning for Autonomous Driving" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 143, + 167, + 471, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 167, + 471, + 194 + ], + "spans": [ + { + "bbox": [ + 143, + 167, + 471, + 194 + ], + "type": "text", + "content": "Ming Nie" + }, + { + "bbox": [ + 143, + 167, + 471, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 143, + 167, + 471, + 194 + ], + "type": "text", + "content": ", Renyuan Peng" + }, + { + "bbox": [ + 143, + 167, + 471, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 143, + 167, + 471, + 194 + ], + "type": "text", + "content": ", Chunwei Wang" + }, + { + "bbox": [ + 143, + 167, + 471, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 143, + 167, + 471, + 194 + ], + "type": "text", + "content": ", Xinyue Cai" + }, + { + "bbox": [ + 143, + 167, + 471, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 143, + 167, + 471, + 194 + ], + "type": "text", + "content": ", Jianhua Han" + }, + { + "bbox": [ + 143, + 167, + 471, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 143, + 167, + 471, + 194 + ], + "type": "text", + "content": ", Hang Xu" + }, + { + "bbox": [ + 143, + 167, + 471, + 194 + ], + "type": "inline_equation", + "content": "^{2\\star}" + }, + { + "bbox": [ + 143, + 167, + 471, + 194 + ], + "type": "text", + "content": ", and Li Zhang" + }, + { + "bbox": [ + 143, + 167, + 471, + 194 + ], + "type": "inline_equation", + "content": "^{1\\star}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 216, + 201, + 396, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 201, + 396, + 224 + ], + "spans": [ + { + "bbox": [ + 216, + 201, + 396, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 216, + 201, + 396, + 224 + ], + "type": "text", + "content": " School of Data Science, Fudan University " + }, + { + "bbox": [ + 216, + 201, + 396, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 216, + 201, + 396, + 224 + ], + "type": "text", + "content": " Huawei Noah's Ark Lab" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 217, + 236, + 396, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 236, + 396, + 247 + ], + "spans": [ + { + "bbox": [ + 217, + 236, + 396, + 247 + ], + "type": "text", + "content": "https://github.com/fudan-zvg/reason2drive" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 276, + 455, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 276, + 455, + 518 + ], + "spans": [ + { + "bbox": [ + 160, + 276, + 455, + 518 + ], + "type": "text", + "content": "Abstract. Large vision-language models (VLMs) have garnered increasing interest in autonomous driving areas, due to their advanced capabilities in complex reasoning tasks essential for highly autonomous vehicle behavior. Despite their potential, research in autonomous systems is hindered by the lack of datasets with annotated reasoning chains that explain the decision-making processes in driving. To bridge this gap, we present Reason2Drive, a benchmark dataset with over 600K video-text pairs, aimed at facilitating the study of interpretable reasoning in complex driving environments. We distinctly characterize the autonomous driving process as a sequential combination of perception, prediction, and reasoning steps, and the question-answer pairs are automatically collected from a diverse range of open-source outdoor driving datasets, including nuScenes, Waymo and ONCE. Moreover, we introduce a novel aggregated evaluation metric to assess chain-based reasoning performance in autonomous systems, addressing the reasoning ambiguities of existing metrics such as BLEU and CIDEr. Based on the proposed benchmark, we conduct experiments to assess various existing VLMs, revealing insights into their reasoning capabilities. Additionally, we develop an efficient approach to empower VLMs to leverage object-level perceptual elements in both feature extraction and prediction, further enhancing their reasoning accuracy. Extendable experiments demonstrate the supportive effect of Reason2Drive towards visual reasoning and downstream planning tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 551, + 230, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 551, + 230, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 551, + 230, + 563 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 576, + 482, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 576, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 576, + 482, + 636 + ], + "type": "text", + "content": "Modern autonomous driving systems face challenges related to generalization issues across diverse scenarios, which is often attributed to the reliance on empirical and intricate rules involved in decision-making. To reduce dependence on such rules, recent end-to-end approaches [5, 20] have been developed to derive control signals directly from sensor inputs, treating the system as a black" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 642, + 482, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 482, + 666 + ], + "type": "text", + "content": "* Li Zhang (lizhangfd@fudan.edu.cn) and Hang Xu (chromexbjxh@gmail.com) are the corresponding authors." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "box that requires extensive data for training. However, this approach tends to obscure the underlying logic of decisions, complicating failure diagnosis in real-world applications. In contrast, Large Vision-Language Models (VLMs) offer a promising alternative, potentially enhancing interpretability and generalization for these systems. With their broad world knowledge and advanced reasoning abilities, as illustrated in Fig. 1(a), VLMs have the potential to provide a more thorough understanding and explicit explanation for reliable decision-making." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 301, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 301, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 301, + 283 + ], + "type": "text", + "content": "Nonetheless, existing works [35, 40] primarily focused on the straightforward adaptation of question-answering tasks to the autonomous driving; how to exploit VLMs to facilitate the reasoning abilities of autonomous systems is still under exploration." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 289, + 301, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 301, + 432 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 301, + 432 + ], + "type": "text", + "content": "One reason that hinders the research in this field lies in the scarcity of datasets, especially those chained-based reasoning labels that elucidate the decision-making process. Most existing datasets [11, 35, 41] often oversimplify the complex processes of driving into straightforward question-answering tasks with only a few specific tasks covered. As depicted in Fig. 1(b), they typically provide closed-form annotations constrained" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 433, + 482, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 482, + 515 + ], + "type": "text", + "content": "to boolean (i.e., yes or no) answers or limited multiple-choice responses (e.g., stopped, parked, and moving). However, autonomous driving transcends a simplistic QA process. It encompasses a multi-step approach involving perception, prediction, and reasoning, each of which plays an indispensable role in the decision-making. Therefore, it is crucial to introduce a novel benchmark annotated with detailed decision-making reasoning for assessing the reasoning abilities of current VLMs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "content": "To this end, we introduce Reason2Drive, a new benchmark comprising over 600K video-text pairs, characterized by intricate driving instructions and a series of perception, prediction and reasoning steps. Our benchmark builds upon widely-used open-source driving datasets including nuScenes [2], Waymo [37], and ONCE [28], utilizing an extensible annotation schema. Specifically, we structure the comprehensive annotations into object-centric database and integrate it into manually crafted templates to create paired data for VLMs at both object and scenario levels. To enhance diversity, GPT-4 and manual instructions are employed for verification and enrichment purposes. Notably, Reason2Drive is the most extensive dataset available to date, outperforming existing datasets in scale and the complexity of reasoning chains included, which is a distinctive attribute not present in other datasets." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 224, + 483, + 375 + ], + "blocks": [ + { + "bbox": [ + 309, + 224, + 483, + 375 + ], + "lines": [ + { + "bbox": [ + 309, + 224, + 483, + 375 + ], + "spans": [ + { + "bbox": [ + 309, + 224, + 483, + 375 + ], + "type": "image", + "image_path": "36d359993802073393053a61f1d11b3beadec8fc472636ff12379307d698ff20.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 384, + 482, + 418 + ], + "lines": [ + { + "bbox": [ + 307, + 384, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 307, + 384, + 482, + 418 + ], + "type": "text", + "content": "Fig.1: (a) Different decision-making processes in autonomous driving. (b) Language-based dataset comparison." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "Ming Nie, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "Furthermore, we observe a fundamental flaw in the current evaluation of VLMs on autonomous driving tasks, due to the inherent reasoning ambiguities of traditional caption-based metrics like BLEU [31] and CIDEr [39]. These metrics mainly measure text generation from a holistic perspective, without considering the causal relationship between the reasoning steps and the final conclusion. For example, when the VLM suggests ego vehicle turning left, we cannot ascertain from these metrics whether its reasoning steps effectively support the final decision. To address this issue, we propose a new aggregated evaluation metric, ADRScore, specifically designed to measure chain-based reasoning performance in autonomous systems, which aims to resolve the reasoning ambiguities associated with current metrics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 247, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 392 + ], + "type": "text", + "content": "Utilizing the proposed benchmark, we undertake experiments to assess various existing VLMs, thereby unveiling valuable insights into their reasoning capabilities. We find that most methods struggle to effectively leverage perceptual priors, resulting in subpar reasoning performance. Additionally, constrained by the language model functioning solely as a decoder, these methods often fail to deliver accurate perceptual results, which is a crucial component for verifying a model's spatial reasoning capability. To alleviate this dilemma, we present a simple yet efficient framework, augmenting existing VLMs with two new components: a prior tokenizer and an instructed vision decoder, which aim to bolster the models' visual localization capabilities within the encoder and decoder, respectively. Extendable experiments demonstrate the supportive effect of Reason2Drive towards visual reasoning and downstream planning tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 392, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 536 + ], + "type": "text", + "content": "The contributions of this paper are summarized as follows: (i) We publish a novel visual instruction tuning dataset aimed at facilitating interpretable and chain-based reasoning autonomous systems. (ii) We introduce a novel evaluation metric, ADRScore, to assess chain-based reasoning performance in autonomous driving, effectively addressing the reasoning ambiguities present in existing metrics. (iii) We conduct experiments to assess a range of existing VLMs, revealing valuable insights into their reasoning capabilities. (iv) To address the challenges posed by inefficient prior feature extraction and inaccurate perceptual predictions, we introduce an efficient approach to integrate these elements into VLMs. This results in a substantial improvement in reasoning accuracy and provides remarkable support for downstream planning tasks. Our method surpasses all baselines, notably achieving impressive generalization in unseen scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 555, + 237, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 237, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 237, + 567 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Multimodal large language model. The current state of large language models provides remarkable abilities in natural language understanding and generation ([6, 7, 30, 38]). Inspired by the potential of large language models, a multitude of multimodal models has emerged, intended to enhance these models' capabilities in achieving multi-modal comprehension. Blip-2 [24] aligns visual and language features by utilizing a learnable Q-former. LLaVA [25] and MiniGPT-4 [46] initially align image-text features and then proceed with in" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "text", + "content": "Reason2Drive" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 118, + 481, + 318 + ], + "blocks": [ + { + "bbox": [ + 135, + 118, + 481, + 318 + ], + "lines": [ + { + "bbox": [ + 135, + 118, + 481, + 318 + ], + "spans": [ + { + "bbox": [ + 135, + 118, + 481, + 318 + ], + "type": "image", + "image_path": "48d2d27052970c57e8dd3a8226dfb8e7af4194e5f0c9748d58ab05e2626edee6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 325, + 482, + 403 + ], + "lines": [ + { + "bbox": [ + 130, + 325, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 482, + 403 + ], + "type": "text", + "content": "Fig. 2: Schema of Our Reason2Drive Dataset. The upper part illustrates the pipeline for the automated construction of datasets. The lower part shows detailed instances of perception, prediction, and reasoning, accompanied by outcomes after applying GPT-4 for data augmentation. The special tokens hold distinct definitions: represents a specified instance, signifies a forecasted sequence of trajectory coordinates, and denotes positional coordinates. The colors associated with these tokens correspond to the highlighted objects in the upper-left image's boxes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 411, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 555 + ], + "type": "text", + "content": "struction tuning. Additionally, Video-LLaMA [44] and ImageBind-LLM [18] integrate multiple modalities into the input, aligning features from various sources like images, videos, audio, and point clouds, consolidating them into the space of language features. Kosmos-2 [33] and Shikra [4] perform object detection based on instructions and also accomplish grounded visual question answering. DetGPT [34] connects a fixed multi-modal LLM with a customizable detector based on user instructions. LISA [23] efficiently embeds segmentation abilities into multi-modal LLMs, showcasing self-reasoning for current perception systems. The previous works have demonstrated that current large-scale multimodal models can achieve cross-modal alignment, enabling comprehension and inference towards images and more. These models can not only perform perceptual tasks like detection but also accomplish preliminary reasoning tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Vision language tasks in autonomous driving. Currently, VLMs have demonstrated robust capabilities in scene perception and understanding. Extensive efforts have been dedicated to the realm of autonomous driving, leveraging VLM to achieve comprehensive scene understanding and perform diverse tasks [13, 16, 29, 42]. Simultaneously, substantial works are in progress to create datasets and models tailored to various tasks. Talk2Car [11] proposes the first object referral dataset for grounding commands for self-driving cars in free natural language into the visual context. But it exclusively contains information about visible objects. While DRAMA [27] outlines the overall scene risk, it lacks pre" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "Ming Nie, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "cise perception annotation. NuPrompt [41] and Refer-KITTI [40] offer language prompt sets for driving scenes but primarily concentrate on multi-object tracking tasks. NuScenesQA [35] and DriveLM [9] build visual question-answering (VQA) datasets for scenario understanding. However, their primary emphasis is on the perceptual information in the scene, lacking annotations for the analysis and complex reasoning of the entire scenario. To address the limitations of existing works, we construct a thorough dataset covering perception, prediction, and complex reasoning, additionally with an improved vision-language model for better analyzing autonomous driving scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 243, + 286, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 243, + 286, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 243, + 286, + 255 + ], + "type": "text", + "content": "3 Reason2Drive Dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 269, + 481, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 481, + 305 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 481, + 305 + ], + "type": "text", + "content": "We introduce Reason2Drive, a dataset that comprises comprehensive driving instructions and a chain-based reasoning framework for decision-making. Our dataset is characterized by the following key aspects:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 319, + 479, + 520 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 319, + 479, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 319, + 479, + 341 + ], + "spans": [ + { + "bbox": [ + 138, + 319, + 479, + 341 + ], + "type": "text", + "content": "- Quantity: It stands out as the largest language-based driving dataset available, collated from prominent publicly accessible datasets worldwide." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 343, + 479, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 479, + 376 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 479, + 376 + ], + "type": "text", + "content": "- **Quality:** Reason2Drive offers a more precise representation of driving activities, including perception, prediction and reasoning, with a reliable auto-annotation schema for data collection." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 384, + 479, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 384, + 479, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 384, + 479, + 453 + ], + "type": "text", + "content": "- Diversity: (i) The dataset exhibits a broader range of scenes, encompassing both object-level and scenario-level data. This diversity includes object types, visual and motion attributes, object locations, and relationships relative to the ego-vehicle. (ii) It includes more intricate question-answer pairs, enhanced by GPT-4, along with longer text passages featuring step-by-step reasoning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 462, + 479, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 462, + 479, + 520 + ], + "spans": [ + { + "bbox": [ + 138, + 462, + 479, + 520 + ], + "type": "text", + "content": "- Protocols: A novel evaluation metric is introduced to assess the reasoning capabilities of VLMs. Different from those widely used in the NLP community, it takes into account not only perception results but also reasoning ambiguities, providing a more holistic evaluation of the VLM's reasoning capacity for autonomous driving scenarios." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 530, + 481, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 530, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 131, + 530, + 481, + 555 + ], + "type": "text", + "content": "Further details regarding the data collection process, statistical data analysis, and benchmark protocols are provided in the subsequent section." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 573, + 251, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 573, + 251, + 583 + ], + "spans": [ + { + "bbox": [ + 132, + 573, + 251, + 583 + ], + "type": "text", + "content": "3.1 Dataset collection" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "As illustrated in Fig. 2, we employ an extensible annotation schema, constructing data in the forms of question-answer pairs. Specifically, we first leverage a diverse array of publicly available datasets collected in different regions worldwide, including nuScenes, Waymo, and ONCE, and then parse their comprehensive annotations into object-centric database. The database is organized in key frames, and each frame stores object entries containing various details pertaining to its" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 390, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 390, + 91, + 447, + 101 + ], + "type": "text", + "content": "Reason2Drive" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 114, + 317, + 239 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 317, + 239 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 317, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 317, + 239 + ], + "type": "table", + "html": "
DatasetDescriptionSource Datasets
Talk2Car [11]Object referralnuScenes
CityFlow-NL [15]Tracking & retrievalCityFlow
CARLA-NAV [21]Segmentation & predictionCARLA Simulator
NuPrompt [41]Multi-object trackingnuScenes
NuScenes-QA [35]PerceptionnuScenes
Refer-KITTI [40]Multi-object trackingKITTI
Talk2BEV [12]Visual understandingnuScenes
DRAMA [27]Risk localizationself-collected
Rank2Tell [36]Risk localization & rankingself-collected
Reason2DrivePerceptionnuScenes
PredictionWaymo
ReasoningONCE
", + "image_path": "a09e21784b2f65f73f0283aea84067e30d0ab5a21e46bc8915374d1d9cd66015.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 247, + 325, + 280 + ], + "lines": [ + { + "bbox": [ + 132, + 247, + 325, + 280 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 325, + 280 + ], + "type": "text", + "content": "Table 1: The comparison between our Reason2Drive dataset and other prompt-based datasets. ■ means dataset not published." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 350, + 124, + 465, + 232 + ], + "blocks": [ + { + "bbox": [ + 350, + 124, + 465, + 232 + ], + "lines": [ + { + "bbox": [ + 350, + 124, + 465, + 232 + ], + "spans": [ + { + "bbox": [ + 350, + 124, + 465, + 232 + ], + "type": "image", + "image_path": "4834d66a5938ded50e5fe869078d9bbf875b2d58a647a8b912d184c30b13e0aa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 238, + 477, + 281 + ], + "lines": [ + { + "bbox": [ + 335, + 238, + 477, + 281 + ], + "spans": [ + { + "bbox": [ + 335, + 238, + 477, + 281 + ], + "type": "text", + "content": "Fig. 3: Data quality comparison. Reason2Drive is larger in scale, richer in data content, and more diverse in scenarios." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 292, + 480, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 292, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 480, + 363 + ], + "type": "text", + "content": "driving actions, including location, category, attributes and more. Afterwards, the structured database is required to answer the manually crafted question templates, which are divided into different tasks (i.e., perception, prediction and reasoning) at both object-level and scenario-level. Subsequently, GPT-4 is involved for verification and enrichment purposes. The example of GPT augmented data and manual instructions are provided in the appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 363, + 481, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 481, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 481, + 411 + ], + "type": "text", + "content": "Due to the complexity of autonomous driving activities, we categorize the tasks into three distinct groups to acquire diversified data: perception, prediction and reasoning. The specifics and distinctions of these three types of tasks are elaborated as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 420, + 480, + 528 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 138, + 420, + 479, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 420, + 479, + 454 + ], + "spans": [ + { + "bbox": [ + 138, + 420, + 479, + 454 + ], + "type": "text", + "content": "- Perception task is designed to identify objects within the driving scenario, assessing the fundamental perceptual capabilities of VLMs in outdoor environments." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 456, + 480, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 480, + 491 + ], + "type": "text", + "content": "- Prediction task entails the prediction of future states of key objects within the perceptual range, challenging VLMs to infer the intentions of objects with video input." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 492, + 479, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 492, + 479, + 528 + ], + "spans": [ + { + "bbox": [ + 138, + 492, + 479, + 528 + ], + "type": "text", + "content": "- Reasoning task prompts the analysis of the current perceptual and predicted states step by step, requiring the deduction of reasoned inferences and decisions through a chain of thoughts (COT) approach." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 536, + 480, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 480, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 480, + 561 + ], + "type": "text", + "content": "For each task, we further categorize the data into object-level and scenario-level. In more detail," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 137, + 570, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 137, + 570, + 479, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 570, + 479, + 616 + ], + "spans": [ + { + "bbox": [ + 137, + 570, + 479, + 616 + ], + "type": "text", + "content": "- Object-level data is formatted to benchmark the foundational capabilities of specific objects. As for perception, we address the location and attributes of objects such as moving status and distance to ego, while for prediction, future motion and merging-in/out status are considered." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 137, + 617, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 617, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 137, + 617, + 480, + 665 + ], + "type": "text", + "content": "- Scenario-level data is organized from a global perspective towards driving environment and ego-driving instructions. It focuses on whether there is an object worth noting currently (perception), whether there is an object worth noting in the future (prediction) and why (reasoning). For example," + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "Ming Nie, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 258, + 208 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 258, + 208 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 258, + 208 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 258, + 208 + ], + "type": "image", + "image_path": "d62edde5d1b5ac8f3529cf28c302e8770bd37e8424f0fca8c9f487722bc10952.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 219, + 274, + 239 + ], + "lines": [ + { + "bbox": [ + 132, + 219, + 274, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 219, + 274, + 239 + ], + "type": "text", + "content": "Fig. 4: Statistical distribution of different tasks in Reason2Drive." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 290, + 131, + 481, + 209 + ], + "blocks": [ + { + "bbox": [ + 290, + 131, + 481, + 209 + ], + "lines": [ + { + "bbox": [ + 290, + 131, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 290, + 131, + 481, + 209 + ], + "type": "table", + "html": "
Task TargetPerception (PE) Prediction (PR) Reasoning (RE)Total
Ego vehicle22629171731886858670
Single object718826166772006205555
Multi objects1020129450252175248689
Scenario495894279527657120041
Total246112216137170706632955
", + "image_path": "cc437172ff7a9cd5da62bda56cb2cbd12a9894df1c1848394755a2c23d7e4920.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 287, + 219, + 480, + 239 + ], + "lines": [ + { + "bbox": [ + 287, + 219, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 287, + 219, + 480, + 239 + ], + "type": "text", + "content": "Table 2: The statistics of different tasks in Reason2Drive dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 148, + 262, + 479, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 262, + 479, + 298 + ], + "spans": [ + { + "bbox": [ + 148, + 262, + 479, + 298 + ], + "type": "text", + "content": "as illustrated in Fig. 2, models are asked to identify distances, merging states and other risks from the whole scene. It verifies the agent's ability to perceive the entire driving scene rather than specifying objects, thus more challenging." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 328, + 242, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 328, + 242, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 328, + 242, + 340 + ], + "type": "text", + "content": "3.2 Dataset analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 350, + 479, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 479, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 479, + 396 + ], + "type": "text", + "content": "Tab. 1 and Fig. 3 demonstrate the comparison between our Reason2Drive dataset and existing benchmarks. It is noteworthy that our benchmark stands as the largest dataset to date, surpassing others in terms of both dataset size and the inclusion of extensive long-text chain-based reasoning references." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 398, + 480, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 480, + 493 + ], + "type": "text", + "content": "To further investigate the property of Reason2Drive dataset, we count the distribution and sample numbers of our dataset in Fig. 4 and Tab. 2. We split the dataset according to the task and target. The benchmark exhibits a balanced distribution, with multi-object tasks constituting the majority. Single-object and scenario-level questions are of similar quantities. The fewest questions are related to the ego-vehicle. Additionally, perception, prediction and reasoning questions are distributed as " + }, + { + "bbox": [ + 130, + 398, + 480, + 493 + ], + "type": "inline_equation", + "content": "39\\%" + }, + { + "bbox": [ + 130, + 398, + 480, + 493 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 398, + 480, + 493 + ], + "type": "inline_equation", + "content": "34\\%" + }, + { + "bbox": [ + 130, + 398, + 480, + 493 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 398, + 480, + 493 + ], + "type": "inline_equation", + "content": "27\\%" + }, + { + "bbox": [ + 130, + 398, + 480, + 493 + ], + "type": "text", + "content": ", respectively. More dataset details are provided in the appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 512, + 263, + 524 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 263, + 524 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 263, + 524 + ], + "type": "text", + "content": "3.3 Benchmark protocol" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 533, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 665 + ], + "type": "text", + "content": "It is worth noting that previous works [12, 27, 35] simply utilize metric scores widely used in the NLP community, including BLEU [31], CIDEr [39] and ME-TEOR [1]. However, these metrics mainly measure text generation from a holistic perspective, without considering the causal relationship between the reasoning steps and the final conclusion. While they perform well in translation and captioning, their efficacy is limited when it comes to reasoning. Moreover, the existing evaluation system is confined to assessing the semantic quality of the results. It fails to effectively evaluate the quality of the perceived results, a crucial aspect supporting automatic driving decision-making. To address these dilemmas, inspired by [43] and [17], we develop a novel evaluation protocol, ADRScore, to measure the correctness of the reasoning chains towards autonomous driving." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "text", + "content": "Reason2Drive" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 148, + 115, + 482, + 292 + ], + "blocks": [ + { + "bbox": [ + 148, + 115, + 482, + 292 + ], + "lines": [ + { + "bbox": [ + 148, + 115, + 482, + 292 + ], + "spans": [ + { + "bbox": [ + 148, + 115, + 482, + 292 + ], + "type": "image", + "image_path": "b1d1d82a41fd4b26479b0c0c4d605067948d0c83ad641f894b2da666ff1fbac3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 299, + 482, + 321 + ], + "lines": [ + { + "bbox": [ + 131, + 299, + 482, + 321 + ], + "spans": [ + { + "bbox": [ + 131, + 299, + 482, + 321 + ], + "type": "text", + "content": "Fig. 5: Illustration of ADRScore and its visual adaptation. When substituting semantic similarities with actual geometric similarities, the score decreases." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 326, + 479, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 326, + 479, + 351 + ], + "spans": [ + { + "bbox": [ + 131, + 326, + 479, + 351 + ], + "type": "text", + "content": "Preliminary. To begin with, we denote the generated reasoning steps as hypothesis " + }, + { + "bbox": [ + 131, + 326, + 479, + 351 + ], + "type": "inline_equation", + "content": "\\pmb{h} = \\{h_1, \\dots, h_N\\}" + }, + { + "bbox": [ + 131, + 326, + 479, + 351 + ], + "type": "text", + "content": ", and the gold annotation as reference " + }, + { + "bbox": [ + 131, + 326, + 479, + 351 + ], + "type": "inline_equation", + "content": "\\pmb{r} = \\{r_1, \\dots, r_K\\}" + }, + { + "bbox": [ + 131, + 326, + 479, + 351 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 351, + 479, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 351, + 479, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 351, + 479, + 374 + ], + "type": "text", + "content": "At the core of reasoning metrics is the reasoning alignment vector from the " + }, + { + "bbox": [ + 132, + 351, + 479, + 374 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 132, + 351, + 479, + 374 + ], + "type": "text", + "content": "-step hypothesis " + }, + { + "bbox": [ + 132, + 351, + 479, + 374 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 132, + 351, + 479, + 374 + ], + "type": "text", + "content": " to the " + }, + { + "bbox": [ + 132, + 351, + 479, + 374 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 132, + 351, + 479, + 374 + ], + "type": "text", + "content": "-step reference:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 241, + 388, + 480, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 388, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 241, + 388, + 480, + 402 + ], + "type": "interline_equation", + "content": "\\operatorname {a l i g n} (\\boldsymbol {h} \\rightarrow \\boldsymbol {r}) = \\left\\{\\alpha_ {1}, \\dots , \\alpha_ {N} \\right\\}, \\tag {1}", + "image_path": "394aa6db237d67b1561bf04b4abd3b7a57e0b2224acbe8a53b4aaff855803487.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 407, + 479, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 407, + 479, + 432 + ], + "spans": [ + { + "bbox": [ + 131, + 407, + 479, + 432 + ], + "type": "text", + "content": "where alignment value " + }, + { + "bbox": [ + 131, + 407, + 479, + 432 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 131, + 407, + 479, + 432 + ], + "type": "text", + "content": " represents the semantic similarity between the corresponding hypothesis step and the most similar reference step:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 267, + 441, + 344, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 441, + 344, + 456 + ], + "spans": [ + { + "bbox": [ + 267, + 441, + 344, + 456 + ], + "type": "interline_equation", + "content": "\\alpha_ {i} = \\max _ {j = 1} ^ {K} s _ {i, j},", + "image_path": "50dae903322d878a74afc5936c0fd2b3f2d16fb7b678c1ffd6e98e42ac2dc5fe.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 269, + 451, + 480, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 451, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 269, + 451, + 480, + 472 + ], + "type": "interline_equation", + "content": "s _ {i, j} = \\cos \\left(h _ {i}, r _ {j}\\right). \\tag {2}", + "image_path": "6b3662fd60c733bc27b9e2fc095e7c35361e133d7f555eddf46fb6edcf63014c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 481, + 479, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 481, + 479, + 540 + ], + "spans": [ + { + "bbox": [ + 131, + 481, + 479, + 540 + ], + "type": "inline_equation", + "content": "\\alpha_{i} \\in [0,1]" + }, + { + "bbox": [ + 131, + 481, + 479, + 540 + ], + "type": "text", + "content": " explicitly measures the grounding of the step-wise reasoning with respect to the reference, and " + }, + { + "bbox": [ + 131, + 481, + 479, + 540 + ], + "type": "inline_equation", + "content": "\\cos(\\cdot)" + }, + { + "bbox": [ + 131, + 481, + 479, + 540 + ], + "type": "text", + "content": " denotes the cosine similarity between the corresponding sentence embeddings, which are extracted a pre-trained bert model. Based on the above reasoning alignment vector, we propose the following metrics to thoroughly measure the quality of reasoning steps." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 542, + 479, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 542, + 479, + 590 + ], + "spans": [ + { + "bbox": [ + 131, + 542, + 479, + 590 + ], + "type": "text", + "content": "Reasoning alignment. The most straightforward way to evaluate the correctness of the hypothesis reasoning chain is to compare the degree of overlap between the hypothesis and the reference. One way of doing that is to measure the reasoning alignment between them:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 244, + 599, + 480, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 244, + 599, + 480, + 632 + ], + "type": "interline_equation", + "content": "R A = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\operatorname {a l i g n} \\left(h _ {i} \\rightarrow r\\right). \\tag {3}", + "image_path": "aa40063c2692d9ecfca97111a1c30120bc9f77ea29f4ad79a2048aafc1ff0937.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": "Redundancy. To find chains that contain information that is not required to solve the problem (i.e., redundant steps), we identify those hypothesis steps that" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "text", + "content": "Ming Nie, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "are least aligned with the reference steps. This metric punishes the chain with steps that are not required for the correct solution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 244, + 148, + 481, + 163 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 148, + 481, + 163 + ], + "spans": [ + { + "bbox": [ + 244, + 148, + 481, + 163 + ], + "type": "interline_equation", + "content": "R D = \\min _ {i = 1} ^ {N} \\operatorname {a l i g n} \\left(h _ {i} \\rightarrow \\boldsymbol {r}\\right). \\tag {4}", + "image_path": "64303802e8e47138609b6c698620d9c434a7776b669ef716be9f610910dcc0a9.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 171, + 479, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 171, + 479, + 219 + ], + "spans": [ + { + "bbox": [ + 131, + 171, + 479, + 219 + ], + "type": "text", + "content": "Missing step. To identify steps that are missing from the hypothesis but could be required to solve the problem, we look at the alignment between reference and the hypothesis, similar to Redundancy. However, here we go through each step in the reference, and check if there is a similar step in the hypothesis:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 244, + 228, + 481, + 242 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 228, + 481, + 242 + ], + "spans": [ + { + "bbox": [ + 244, + 228, + 481, + 242 + ], + "type": "interline_equation", + "content": "M S = \\min _ {i = 1} ^ {K} \\operatorname {a l i g n} \\left(r _ {i} \\rightarrow \\boldsymbol {h}\\right). \\tag {5}", + "image_path": "5db592e01f39a4a0d04247df41d471e05cf1ab8c118ed2de05f78c8e710ddf20.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 250, + 479, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 250, + 479, + 270 + ], + "spans": [ + { + "bbox": [ + 131, + 250, + 479, + 270 + ], + "type": "text", + "content": "Finally, the aggregated metric score is the average of the above performance, which is:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 230, + 270, + 481, + 293 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 270, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 230, + 270, + 481, + 293 + ], + "type": "interline_equation", + "content": "A D R S c o r e = \\frac {1}{3} (R A + R D + M S). \\tag {6}", + "image_path": "5baa930f74e484487e03b57197147de91e3f919497945b69964dcd8161feed9d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 296, + 480, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 296, + 480, + 356 + ], + "spans": [ + { + "bbox": [ + 131, + 296, + 480, + 356 + ], + "type": "text", + "content": "Adapted to visual elements. To further adapt to the realistic driving process, we promote the above metric to the situation with visual elements, named ADRScore-S. Specifically, as illustrated in Fig. 5, when the hypothesis step " + }, + { + "bbox": [ + 131, + 296, + 480, + 356 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 131, + 296, + 480, + 356 + ], + "type": "text", + "content": " and reference step " + }, + { + "bbox": [ + 131, + 296, + 480, + 356 + ], + "type": "inline_equation", + "content": "r_j" + }, + { + "bbox": [ + 131, + 296, + 480, + 356 + ], + "type": "text", + "content": " contains visual elements, i.e., the locations and motions predicted for further reasoning, the similarity score becomes:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 261, + 365, + 481, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 365, + 481, + 389 + ], + "spans": [ + { + "bbox": [ + 261, + 365, + 481, + 389 + ], + "type": "interline_equation", + "content": "s _ {i, j} = \\frac {\\tau - M \\left(h _ {i} , r _ {j}\\right)}{\\beta}, \\tag {7}", + "image_path": "d277d79a97dd2a5296c37303797a4be098db8ea5eaf5df7e85acaebac73d15c1.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 394, + 481, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 394, + 481, + 465 + ], + "spans": [ + { + "bbox": [ + 131, + 394, + 481, + 465 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 394, + 481, + 465 + ], + "type": "inline_equation", + "content": "M(\\cdot)" + }, + { + "bbox": [ + 131, + 394, + 481, + 465 + ], + "type": "text", + "content": " measures the mean square error between two perceptual elements. And " + }, + { + "bbox": [ + 131, + 394, + 481, + 465 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 131, + 394, + 481, + 465 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 394, + 481, + 465 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 131, + 394, + 481, + 465 + ], + "type": "text", + "content": " are used to normalize it to [0,1] to match the distribution of semantic-level similarity. ADRScore-S more harshly measures the performance on spatial reasoning as the metric calculates the error of visual elements in spatial instead of text semantic. The latter is too lenient for visual predictions in language." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 483, + 233, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 483, + 233, + 498 + ], + "spans": [ + { + "bbox": [ + 131, + 483, + 233, + 498 + ], + "type": "text", + "content": "4 Methodology" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 509, + 479, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 479, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 479, + 533 + ], + "type": "text", + "content": "In this section we introduce our framework in Sec. 4.1, followed by the training details provided in Sec. 4.2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 550, + 257, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 550, + 257, + 562 + ], + "spans": [ + { + "bbox": [ + 131, + 550, + 257, + 562 + ], + "type": "text", + "content": "4.1 Model architecture" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 570, + 479, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 479, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 479, + 629 + ], + "type": "text", + "content": "We observe that most VLMs struggle to effectively handle object-level perceptual information, including the input of visual priors and predictions of object locations, which are indispensable in autonomous driving scenarios. The limitation is primarily due to (i) the lack of a targeted tokenizer and (ii) decoder solely composed of a language model, resulting in subpar reasoning performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "content": "To address this challenge, as illustrated in Fig. 6, we introduce a straightforward yet effective framework that enhances existing VLMs with two new components: a prior tokenizer and an instructed vision decoder. Notably, the design" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 390, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 390, + 91, + 447, + 101 + ], + "type": "text", + "content": "Reason2Drive" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 113, + 482, + 235 + ], + "blocks": [ + { + "bbox": [ + 132, + 113, + 482, + 235 + ], + "lines": [ + { + "bbox": [ + 132, + 113, + 482, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 113, + 482, + 235 + ], + "type": "image", + "image_path": "2636b6b6bb2a3c6b070ed52f9a633f9e27a2f5b52f1d54dceedd0ae7b7023d21.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 241, + 482, + 298 + ], + "lines": [ + { + "bbox": [ + 130, + 241, + 482, + 298 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 482, + 298 + ], + "type": "text", + "content": "Fig. 6: The pipeline of our framework. The input video and perceptual priors are tokenized using the vision encoder and prior tokenizer. A Q-former then aligns them to the text's feature space. The LLM and instructed vision decoder predict answers with precise perception results for user queries. The highlighted yellow box and red curve in the perception results respectively represent the visualization of " + }, + { + "bbox": [ + 130, + 241, + 482, + 298 + ], + "type": "inline_equation", + "content": "\\langle \\mathsf{LOC} \\rangle" + }, + { + "bbox": [ + 130, + 241, + 482, + 298 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 241, + 482, + 298 + ], + "type": "inline_equation", + "content": "\\langle \\mathsf{MOT} \\rangle" + }, + { + "bbox": [ + 130, + 241, + 482, + 298 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 306, + 482, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 482, + 354 + ], + "type": "text", + "content": "of the components is not for detection but to aid interpretable visual reasoning based on available perception inputs. These modules aim to strengthen the capabilities of the model to utilize object-level perceptual elements in both extracting visual priors and generating perceptual predictions for visual reasoning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 356, + 483, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 483, + 404 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 483, + 404 + ], + "type": "text", + "content": "Vision encoder. Our model accepts both video frames and text inputs, along with perceptual priors, and tokenizes them into embeddings. For a sequence of video frames " + }, + { + "bbox": [ + 130, + 356, + 483, + 404 + ], + "type": "inline_equation", + "content": "(V_{1}, V_{2}, \\dots, V_{N})" + }, + { + "bbox": [ + 130, + 356, + 483, + 404 + ], + "type": "text", + "content": ", features are extracted using a pretrained Blip-2 visual encoder [24] " + }, + { + "bbox": [ + 130, + 356, + 483, + 404 + ], + "type": "inline_equation", + "content": "F_{v}" + }, + { + "bbox": [ + 130, + 356, + 483, + 404 + ], + "type": "text", + "content": " and aggregated through concatenation:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 223, + 421, + 481, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 421, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 223, + 421, + 481, + 435 + ], + "type": "interline_equation", + "content": "f _ {v} = F _ {v} \\left(V _ {1}\\right) \\oplus F _ {v} \\left(V _ {2}\\right) \\oplus \\dots \\oplus F _ {v} \\left(V _ {N}\\right). \\tag {8}", + "image_path": "cdee82d02af77da7f9e963d7318707e51ac992e874bc81ea2be44f8699180221.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 445, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 567 + ], + "type": "text", + "content": "Prior tokenizer. We propose a novel tokenization strategy tailored to taking advantage of visual cues. The motivation is grounded in the acknowledgement that extracting and aligning visual features is considerably simpler and more suitable compared to compelling the LLM to comprehend ambiguous positional descriptions. Direct textual input to the LLM may result in challenges such as information loss, as textual representation may not fully capture image details and context, especially in complex scenarios with dynamic object positions and velocities. To tackle this issue, we design a novel tokenizer " + }, + { + "bbox": [ + 130, + 445, + 482, + 567 + ], + "type": "inline_equation", + "content": "F_{p}" + }, + { + "bbox": [ + 130, + 445, + 482, + 567 + ], + "type": "text", + "content": ", implemented as a two-layer MLP, to independently extract local image features and positional embeddings from visual priors:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 260, + 582, + 481, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 582, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 260, + 582, + 481, + 597 + ], + "type": "interline_equation", + "content": "f _ {p} = F _ {p} (f _ {r} + E (P)), \\tag {9}", + "image_path": "976719265ce5ceaf2b71dad8079579a1d79f1e429c71cc40101689ad71d4a868.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "f_{r}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " represents the region-level features extracted from the image-level features " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " according to the precise locations of perception priors " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ". These features are aligned to " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " size using the RoIAlign [14] operation and fused into a single embedding " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "f_{r}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ". And " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "E(\\cdot)" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " is a positional encoding function mapping the geometry locations and motions into the same dimension of " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "f_{r}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "Ming Nie, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 152 + ], + "type": "text", + "content": "LLM. After we tokenize the video and perception priors into embedding " + }, + { + "bbox": [ + 131, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 131, + 116, + 479, + 152 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "f_{p}" + }, + { + "bbox": [ + 131, + 116, + 479, + 152 + ], + "type": "text", + "content": ", a projector " + }, + { + "bbox": [ + 131, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 131, + 116, + 479, + 152 + ], + "type": "text", + "content": " (Q-former [24] in this work) is adopted to align the non-text features into textual domain:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 275, + 165, + 480, + 176 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 165, + 480, + 176 + ], + "spans": [ + { + "bbox": [ + 275, + 165, + 480, + 176 + ], + "type": "interline_equation", + "content": "f _ {q} = Q \\left(f _ {v}, f _ {p}\\right). \\tag {10}", + "image_path": "e5e66bd3d660dc42cac1972ea8b559954e813c281c4dc7ad99002cfd84c8fc55.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 181, + 479, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 181, + 479, + 206 + ], + "spans": [ + { + "bbox": [ + 131, + 181, + 479, + 206 + ], + "type": "text", + "content": "Then, to generate the final text output, we utilize the LLM for further language processing with the extracted text embedding " + }, + { + "bbox": [ + 131, + 181, + 479, + 206 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 131, + 181, + 479, + 206 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 276, + 219, + 480, + 230 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 219, + 480, + 230 + ], + "spans": [ + { + "bbox": [ + 276, + 219, + 480, + 230 + ], + "type": "interline_equation", + "content": "\\hat {y} _ {t} = F \\left(f _ {t}, f _ {q}\\right). \\tag {11}", + "image_path": "257761e84e7d43a90f75d76056364c50550a7e6b9b7b1f4a876e76e8b03b086d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 234, + 480, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 234, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 131, + 234, + 480, + 426 + ], + "type": "text", + "content": "Instructed vision decoder. Current works [11, 16] treat the LLM as a versatile tool to generate answers and inferences without intermediate reasoning steps, let alone considering the perceptions of the agent toward driving scenes. However, the perception ability of the agent towards driving scenarios is an indispensable part of a reliable driving procedure. Moreover, recent works [23] have demonstrated that, rather than training with textualized perceptual sequences, incorporating the perception abilities into the multi-modal LLM brings a significant improvement. To this end, inspired by [23], we integrate new perception capabilities into the multi-modal LLM. Specifically, we expand the original LLM vocabulary by introducing new tokens as placeholders, denoted as " + }, + { + "bbox": [ + 131, + 234, + 480, + 426 + ], + "type": "inline_equation", + "content": "\\langle \\mathsf{LOC} \\rangle" + }, + { + "bbox": [ + 131, + 234, + 480, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 234, + 480, + 426 + ], + "type": "inline_equation", + "content": "\\langle \\mathsf{MOT} \\rangle" + }, + { + "bbox": [ + 131, + 234, + 480, + 426 + ], + "type": "text", + "content": ", to signify the request for the perception output. When the LLM aims to generate a specific perception, the output " + }, + { + "bbox": [ + 131, + 234, + 480, + 426 + ], + "type": "inline_equation", + "content": "\\hat{y}_t" + }, + { + "bbox": [ + 131, + 234, + 480, + 426 + ], + "type": "text", + "content": " should include a designed token. We then extract the last-layer textual features corresponding to the specific token and apply an MLP projection layer to obtain the hidden embedding " + }, + { + "bbox": [ + 131, + 234, + 480, + 426 + ], + "type": "inline_equation", + "content": "f_h" + }, + { + "bbox": [ + 131, + 234, + 480, + 426 + ], + "type": "text", + "content": ". Finally, the textual embedding and visual features are fed into the instructed vision decoder to decode the predictions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 276, + 438, + 479, + 450 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 438, + 479, + 450 + ], + "spans": [ + { + "bbox": [ + 276, + 438, + 479, + 450 + ], + "type": "interline_equation", + "content": "\\tilde {P} = D \\left(f _ {v}, f _ {h}\\right). \\tag {12}", + "image_path": "4c12fd1601a6b7a43d12bec0d42f4cdf7b8ae70a374e2d6901d5e99e139d0a7c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 455, + 479, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 455, + 479, + 480 + ], + "spans": [ + { + "bbox": [ + 131, + 455, + 479, + 480 + ], + "type": "text", + "content": "This module is comprised of a transformer decoder for features alignment [3] and task-specific heads designed to generate object locations and motions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 497, + 239, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 497, + 239, + 509 + ], + "spans": [ + { + "bbox": [ + 131, + 497, + 239, + 509 + ], + "type": "text", + "content": "4.2 Training details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 516, + 479, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 516, + 479, + 541 + ], + "spans": [ + { + "bbox": [ + 131, + 516, + 479, + 541 + ], + "type": "text", + "content": "Training objectives. The model is trained end-to-end using the text generation loss " + }, + { + "bbox": [ + 131, + 516, + 479, + 541 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{txt}" + }, + { + "bbox": [ + 131, + 516, + 479, + 541 + ], + "type": "text", + "content": " and the perception output loss " + }, + { + "bbox": [ + 131, + 516, + 479, + 541 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{per}" + }, + { + "bbox": [ + 131, + 516, + 479, + 541 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 262, + 553, + 479, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 553, + 479, + 564 + ], + "spans": [ + { + "bbox": [ + 262, + 553, + 479, + 564 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {t x t} + \\lambda_ {p e r} \\mathcal {L} _ {p e r}, \\tag {13}", + "image_path": "96689dc3730a7111054b39bfb1258f9c8f6ff7fb7f447fc3fd3c9f6eef4b4bb9.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 570, + 479, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 479, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 479, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 570, + 479, + 629 + ], + "type": "inline_equation", + "content": "\\lambda_{per}" + }, + { + "bbox": [ + 130, + 570, + 479, + 629 + ], + "type": "text", + "content": " is the balancing term. Specifically, " + }, + { + "bbox": [ + 130, + 570, + 479, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{txt}" + }, + { + "bbox": [ + 130, + 570, + 479, + 629 + ], + "type": "text", + "content": " is the auto-regressive cross-entropy loss for text generation, and " + }, + { + "bbox": [ + 130, + 570, + 479, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{per}" + }, + { + "bbox": [ + 130, + 570, + 479, + 629 + ], + "type": "text", + "content": " encourages the instructed vision decoder to generate accurate locations and motions, which is similar to traditional detection loss and is employed with the combination of binary cross-entropy loss and MSE loss. More details are included in the appendix." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 630, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 630, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 630, + 479, + 666 + ], + "type": "text", + "content": "Tuning strategy. Our tuning strategy consists of two stages: the pre-training stage and the fine-tuning stage. In the pre-training stage, we initialize the weights from instructBLIP [10], including the pre-trained vision encoder, Q-former and" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "text", + "content": "Reason2Drive" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 170, + 140, + 447, + 269 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 3: Results of different models on the Reason2Drive validation set. We evaluate with ADRScore as well as captioning-based metrics." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 170, + 140, + 447, + 269 + ], + "lines": [ + { + "bbox": [ + 170, + 140, + 447, + 269 + ], + "spans": [ + { + "bbox": [ + 170, + 140, + 447, + 269 + ], + "type": "table", + "html": "
MethodsLLMReasoning metricCaptioning metric
ADRScoreADRScore-SB@4METEORROUGECIDEr
Blip-2 [24]OPT-2.7B [45]0.2960.1620.3610.2490.4430.174
FlanT5-XL [8]0.3100.1710.3680.2560.4510.187
InstructBLIP [10]FlanT5-XL0.3290.1870.3760.2690.4620.196
Vicuna-7B [32]0.3510.2140.4080.2940.4840.211
MiniGPT-4 [46]Vicuna-7B0.3380.2030.3960.2860.4750.219
OursFlanT5-XL0.4570.4200.4510.3490.5200.292
Vicuna-7B0.4630.4320.4570.3560.5290.298
", + "image_path": "760d851588df512f61deb8076f2e54d4ee8141c870d2ef2e2172636d77243387.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 170, + 286, + 447, + 383 + ], + "blocks": [ + { + "bbox": [ + 177, + 272, + 435, + 283 + ], + "lines": [ + { + "bbox": [ + 177, + 272, + 435, + 283 + ], + "spans": [ + { + "bbox": [ + 177, + 272, + 435, + 283 + ], + "type": "text", + "content": "Table 4: Ablations on different combinations of training tasks." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 170, + 286, + 447, + 383 + ], + "lines": [ + { + "bbox": [ + 170, + 286, + 447, + 383 + ], + "spans": [ + { + "bbox": [ + 170, + 286, + 447, + 383 + ], + "type": "table", + "html": "
TasksReasoning metricCaptioning metric
PerceptionPredictionReasoningADRScoreADRScore-SB@4METEORROUGECIDEr
✓0.2820.2530.4220.3070.4790.226
✓✓0.2970.2640.4190.3100.4790.228
✓0.3510.3230.4300.3250.4950.263
✓✓0.4070.3640.4350.3370.5010.274
✓✓✓0.4630.4320.4570.3560.5290.298
", + "image_path": "2fb5b36e415bc116d6e9fbcfc52cebc4bca9d2b0bc2d683c6b5a5baafb7dfd0c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 394, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 482, + 491 + ], + "type": "text", + "content": "LLM, and freeze the parameters of LLM and vision tokenizer " + }, + { + "bbox": [ + 130, + 394, + 482, + 491 + ], + "type": "inline_equation", + "content": "F_{v}" + }, + { + "bbox": [ + 130, + 394, + 482, + 491 + ], + "type": "text", + "content": ". We train the prior tokenizer " + }, + { + "bbox": [ + 130, + 394, + 482, + 491 + ], + "type": "inline_equation", + "content": "F_{p}" + }, + { + "bbox": [ + 130, + 394, + 482, + 491 + ], + "type": "text", + "content": " and Q-former " + }, + { + "bbox": [ + 130, + 394, + 482, + 491 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 394, + 482, + 491 + ], + "type": "text", + "content": " to align visual priors with text, along with the instructed vision decoder " + }, + { + "bbox": [ + 130, + 394, + 482, + 491 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 394, + 482, + 491 + ], + "type": "text", + "content": " to enhance visual localization capabilities. The fine-tuning phase equips the LLM with reasoning abilities in autonomous driving using the instructed vision decoder. To retain pre-trained LLM generalization, we employ efficient fine-tuning with LoRA [19]. The vision encoder and prior tokenizer " + }, + { + "bbox": [ + 130, + 394, + 482, + 491 + ], + "type": "inline_equation", + "content": "F_{p}" + }, + { + "bbox": [ + 130, + 394, + 482, + 491 + ], + "type": "text", + "content": " remain fixed, while the instructed vision decoder " + }, + { + "bbox": [ + 130, + 394, + 482, + 491 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 394, + 482, + 491 + ], + "type": "text", + "content": " is fully fine-tuned. Word embeddings of the LLM and Q-former are also trainable." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 509, + 230, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 230, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 230, + 522 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 582 + ], + "type": "text", + "content": "We benchmark various baseline models and present our method on Reason2Drive dataset. Sec. 5.1 covers implementation details. We assess reasoning performance using our proposed metric in Sec. 5.2, perform ablation studies in Sec. 5.3 and provide more ablations and qualitative results in the appendix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 597, + 266, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 266, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 266, + 611 + ], + "type": "text", + "content": "5.1 Experimental setting" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Our main experiments are carried out on the complete Reason2Drive benchmark. The dataset is collected from three different source datasets: nuScenes [2], Waymo [37], and ONCE [28]. It is divided into training and validation sets based on segments, with " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": " allocated to the training set and " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": " to the validation" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "Ming Nie, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 170, + 114, + 447, + 156 + ], + "blocks": [ + { + "bbox": [ + 203, + 101, + 411, + 111 + ], + "lines": [ + { + "bbox": [ + 203, + 101, + 411, + 111 + ], + "spans": [ + { + "bbox": [ + 203, + 101, + 411, + 111 + ], + "type": "text", + "content": "Table 5: The quality of predicted visual elements." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 170, + 114, + 447, + 156 + ], + "lines": [ + { + "bbox": [ + 170, + 114, + 447, + 156 + ], + "spans": [ + { + "bbox": [ + 170, + 114, + 447, + 156 + ], + "type": "table", + "html": "
PredictionsMetricMiniGPT-4Kosmos-2 [33]Ours
Bounding boxAccuracy ↑0.7230.7450.806
TrajectoryADE ↓2.3342.5631.875
", + "image_path": "2a5ef19aec2a016a4b402cbdcd6bd69f0c2bc7ba9f37364c0594af5ff0e9689b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 161, + 482, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 482, + 292 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 482, + 292 + ], + "type": "text", + "content": "set, ensuring no overlap in scenes between them. The validation set also contains all the declared tasks and has been augmented. The input consists of 5 frames of cropped images with a size of " + }, + { + "bbox": [ + 130, + 161, + 482, + 292 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 161, + 482, + 292 + ], + "type": "text", + "content": " pixels. During training, we leverage the AdamW [26] optimizer with a weight decay of 0.01. We adopt a cosine learning rate decay scheduler with a max value of 3e-4 and a linear warm-up for the first 1000 iterations. The weight of perception loss " + }, + { + "bbox": [ + 130, + 161, + 482, + 292 + ], + "type": "inline_equation", + "content": "\\lambda_{p}" + }, + { + "bbox": [ + 130, + 161, + 482, + 292 + ], + "type": "text", + "content": " is set to 1.0. The normalization parameters " + }, + { + "bbox": [ + 130, + 161, + 482, + 292 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 161, + 482, + 292 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 161, + 482, + 292 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 161, + 482, + 292 + ], + "type": "text", + "content": " are selected to be 15 and 10 after empirical practice. Our models are trained with a batch size of 8 on 8 V100 GPUs. The baseline models are fine-tuned following the official tuning strategy, with the inputs consistent with our approach. Specifically, the perception priors are also provided as textual inputs to baseline models for fair comparison." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 149, + 332, + 306, + 388 + ], + "blocks": [ + { + "bbox": [ + 146, + 300, + 465, + 322 + ], + "lines": [ + { + "bbox": [ + 146, + 300, + 465, + 322 + ], + "spans": [ + { + "bbox": [ + 146, + 300, + 465, + 322 + ], + "type": "text", + "content": "Table 6: Ablations on visual input and Table 7: Ablations on different settings of instructed vision decoder." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 332, + 306, + 388 + ], + "lines": [ + { + "bbox": [ + 149, + 332, + 306, + 388 + ], + "spans": [ + { + "bbox": [ + 149, + 332, + 306, + 388 + ], + "type": "table", + "html": "
Visual features\nimage-level video-levelPerceptual priors\nregion-level positionalADRScoreADRScore-S
✓0.4140.379
0.4310.394
✓0.4470.418
✓0.4630.432
", + "image_path": "d1a03333d33809ad2f947317d1f1cd368e47c3b13117ce629ae56cba1da9ae79.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 312, + 332, + 461, + 389 + ], + "blocks": [ + { + "bbox": [ + 312, + 332, + 461, + 389 + ], + "lines": [ + { + "bbox": [ + 312, + 332, + 461, + 389 + ], + "spans": [ + { + "bbox": [ + 312, + 332, + 461, + 389 + ], + "type": "table", + "html": "
Pre-train text embedding MLPADRScoreADRScore-S
0.3870.361
✓0.4210.396
✓✓0.4550.425
✓✓✓0.4630.432
", + "image_path": "94a828ffb28d311907c9d015a9086f5db39bef0dd0abca03c3c8edd7679204f4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 396, + 334, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 334, + 408 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 334, + 408 + ], + "type": "text", + "content": "5.2 Reasoning performance evaluation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 415, + 482, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 522 + ], + "type": "text", + "content": "As demonstrated in Tab. 3, we evaluate both ADRScore and traditional caption-based performance of different models on our benchmark. It is worth noting that our method outperforms others comprehensively in all metrics. We also observe that, despite there is a correlation between ADRScore and traditional metrics, while on the other hand, the performance gap is more pronounced in our metrics, specially revealed by ADRScore-S. The results further substantiate reasoning ambiguities in traditional metrics, constraining the differentiation in benchmarking model performance. Specifically, models with varying reasoning capabilities exhibit minimal disparities when evaluated using traditional metrics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 539, + 236, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 539, + 236, + 551 + ], + "spans": [ + { + "bbox": [ + 132, + 539, + 236, + 551 + ], + "type": "text", + "content": "5.3 Ablation study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 558, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 641 + ], + "type": "text", + "content": "Task contributions. To investigate the synergies between different tasks, we separate and evaluate various types of tasks independently. As shown in Tab. 4, we train our model on different combinations of tasks. Most notably, training on reasoning tasks plays the most important role, indicating the necessity of relevant reasoning data for instructional tuning. Based on reasoning tasks, perception and prediction tasks additionally enhance the models for visual reasoning, showing specific improvements of " + }, + { + "bbox": [ + 130, + 558, + 482, + 641 + ], + "type": "inline_equation", + "content": "4.1\\%" + }, + { + "bbox": [ + 130, + 558, + 482, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 558, + 482, + 641 + ], + "type": "inline_equation", + "content": "6.8\\%" + }, + { + "bbox": [ + 130, + 558, + 482, + 641 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "Quality of predicted visual elements. To validate the quality of predicted perceptual elements, we also conduct experiments to evaluate the bounding" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "text", + "content": "Reason2Drive" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 169, + 126, + 447, + 231 + ], + "blocks": [ + { + "bbox": [ + 176, + 114, + 435, + 125 + ], + "lines": [ + { + "bbox": [ + 176, + 114, + 435, + 125 + ], + "spans": [ + { + "bbox": [ + 176, + 114, + 435, + 125 + ], + "type": "text", + "content": "Table 8: Evaluation of control signals. B: B@4. M: METHOR." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 169, + 126, + 447, + 231 + ], + "lines": [ + { + "bbox": [ + 169, + 126, + 447, + 231 + ], + "spans": [ + { + "bbox": [ + 169, + 126, + 447, + 231 + ], + "type": "table", + "html": "
MethodLLMB↑M↑RMSE
Speed↓Steer↓
Directly fine-tuned with control signals:
InstructBLIPVicuna-7B0.1660.2013.7435.926
Additionally pre-trained on Reason2Drive:
InstructBLIPVicuna-7B0.1920.2373.0865.151
OursVicuna-7B0.2130.2692.8424.866
", + "image_path": "ef3a495f797ddbb68ae5c5ea799968516979631752781678ac07b1444a3b8277.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 232, + 479, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 232, + 479, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 479, + 291 + ], + "type": "text", + "content": "boxes and trajectories separately, as shown in Tab. 5. To ensure a fair comparison, we implement Kosmos-2 [33] due to its advantageous grounding capabilities. The model is fine-tuned on our dataset following the official strategy. Experimental results confirm the high quality of our predicted visual elements, with particular emphasis on the accuracy of the trajectories." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 292, + 480, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 292, + 480, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 480, + 350 + ], + "type": "text", + "content": "The effects of tokenizers. To verify the effectiveness of the tokenizers, we conduct ablation studies to pinpoint where the improvements come from in Tab. 6. Visual features from single frame to multi-frame bring " + }, + { + "bbox": [ + 130, + 292, + 480, + 350 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 130, + 292, + 480, + 350 + ], + "type": "text", + "content": " improvement in ADRScore-S. Perceptual priors, i.e., region-level features and positional embeddings bring " + }, + { + "bbox": [ + 130, + 292, + 480, + 350 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 130, + 292, + 480, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 292, + 480, + 350 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 130, + 292, + 480, + 350 + ], + "type": "text", + "content": " development." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 351, + 480, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 480, + 399 + ], + "type": "text", + "content": "The effects of instructed vision decoder. To verify the efficiency of our instructed vision decoder, we conduct an ablation study to compare it with other methods. As demonstrated in Tab. 7, pre-training and textual embedding bring the major contribution (3.5% and 2.9% in ADRScore-S)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 399, + 480, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 480, + 483 + ], + "type": "text", + "content": "Downstream tasks. We are also interested in understanding how our benchmark will contribute to downstream tasks, such as predicting control signals. Following the approach in ADriver-I [22], we generate control signals on nuScenes and fine-tune InstructBLIP [10] directly for planning signal prediction. To ablate the influence of our dataset, we also pre-train models on Reason2Drive before fine-tuning with control signals. The results presented in Tab. 8 demonstrate the supportive effect of Reason2Drive on downstream planning tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 499, + 220, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 499, + 220, + 511 + ], + "spans": [ + { + "bbox": [ + 131, + 499, + 220, + 511 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "content": "In summary, Large Vision-Language Models (VLMs) have sparked interest in autonomous driving for their advanced reasoning capabilities. However, the absence of datasets explaining decision-making processes hinders progress. To tackle this, we introduce Reason2Drive benchmark, comprising " + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "inline_equation", + "content": "600\\mathrm{K}+" + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "content": " video-text pairs for interpretable reasoning in complex driving scenarios. It outperforms existing datasets in scale, sources and task diversities. We also propose a novel evaluation protocol for chain-based reasoning, addressing existing semantic ambiguities. To uncover insights into their reasoning abilities, our work evaluates various VLMs and proposes an efficient method to boost the ability of models to utilize object-level perceptual elements in both the encoder and decoder. We expect our work could propel further advancements in interpretable reasoning for autonomous systems." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "Ming Nie, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 190 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 190 + ], + "type": "text", + "content": "This work was supported in part by National Natural Science Foundation of China (Grant No. 62106050 and 62376060), Natural Science Foundation of Shanghai (Grant No. 22ZR1407500), USyd-Fudan BISA Flagship Research Program and Lingang Laboratory (Grant No. LG-QS-202202-07)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 209, + 198, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 209, + 198, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 209, + 198, + 220 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 234, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 234, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 234, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 138, + 234, + 480, + 257 + ], + "type": "text", + "content": "1. Banerjee, S., Lavie, A.: Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In: ACL workshop (2005)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 258, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 258, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 258, + 480, + 289 + ], + "type": "text", + "content": "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: CVPR (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 290, + 480, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 290, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 138, + 290, + 480, + 312 + ], + "type": "text", + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 312, + 480, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 312, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 480, + 334 + ], + "type": "text", + "content": "4. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal lvm's referential dialogue magic. arXiv preprint (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 335, + 480, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 335, + 480, + 356 + ], + "spans": [ + { + "bbox": [ + 138, + 335, + 480, + 356 + ], + "type": "text", + "content": "5. Chen, L., Wu, P., Chitta, K., Jaeger, B., Geiger, A., Li, H.: End-to-end autonomous driving: Challenges and frontiers. arXiv preprint (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 357, + 480, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 480, + 399 + ], + "type": "text", + "content": "6. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 138, + 357, + 480, + 399 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 138, + 357, + 480, + 399 + ], + "type": "text", + "content": " chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 400, + 480, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 400, + 480, + 433 + ], + "spans": [ + { + "bbox": [ + 138, + 400, + 480, + 433 + ], + "type": "text", + "content": "7. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. arXiv preprint (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 434, + 480, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 480, + 466 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 480, + 466 + ], + "type": "text", + "content": "8. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 467, + 480, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 467, + 480, + 488 + ], + "spans": [ + { + "bbox": [ + 138, + 467, + 480, + 488 + ], + "type": "text", + "content": "9. Contributors, D.: Drivelm: Drive on language. https://github.com/OpenDriveLab/DriveLM (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 489, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 489, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 489, + 480, + 521 + ], + "type": "text", + "content": "0. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B.A., Fung, P., Hoi, S.C.H.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 522, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 480, + 544 + ], + "type": "text", + "content": "1. Deruyttere, T., Grujicic, D., Blaschko, M.B., Moens, M.F.: Talk2car: Predicting physical trajectories for natural language commands. IEEE Access (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 544, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 480, + 577 + ], + "type": "text", + "content": "2. Dewangan, V., Choudhary, T., Chandhok, S., Priyadarshan, S., Jain, A., Singh, A.K., Srivastava, S., Jatavallabhula, K.M., Krishna, K.M.: Talk2bev: Language-enhanced bird's-eye view maps for autonomous driving. arXiv preprint (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 577, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 480, + 610 + ], + "type": "text", + "content": "3. Ding, X., Han, J., Xu, H., Zhang, W., Li, X.: Hilm-d: Towards high-resolution understanding in multimodal large language models for autonomous driving. arXiv preprint (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 611, + 368, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 368, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 368, + 621 + ], + "type": "text", + "content": "4. Dollar, K., Girshick, R.: Mask r-cnn. In: ICCV (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "text", + "content": "5. Feng, Q., Ablavsky, V., Sclaroff, S.: Cityflow-nl: Tracking and retrieval of vehicles at city scale by natural language descriptions. arXiv preprint (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "text", + "content": "6. Fu, D., Li, X., Wen, L., Dou, M., Cai, P., Shi, B., Qiao, Y.: Drive like a human: Rethinking autonomous driving with large language models. arXiv preprint (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "text", + "content": "Reason2Drive" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "17. Golovneva, O., Chen, M., Poff, S., Corredor, M., Zettlemoyer, L., Fazel-Zarandi, M., Celikyilmaz, A.: Roscoe: A suite of metrics for scoring step-by-step reasoning. arXiv preprint (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "text", + "content": "18. Han, J., Zhang, R., Shao, W., Gao, P., Xu, P., Xiao, H., Zhang, K., Liu, C., Wen, S., Guo, Z., et al.: Imagebind-llm: Multi-modality instruction tuning. arXiv preprint (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 204 + ], + "type": "text", + "content": "19. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 204, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 204, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 204, + 480, + 224 + ], + "type": "text", + "content": "20. Hu, Y., Yang, J., Chen, L., Li, K., Sima, C., Zhu, X., Chai, S., Du, S., Lin, T., Wang, W., et al.: Planning-oriented autonomous driving. In: CVPR (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 225, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 480, + 246 + ], + "type": "text", + "content": "21. Jain, K., Chhangani, V., Tiwari, A., Krishna, K.M., Gandhi, V.: Ground then navigate: Language-guided navigation in dynamic scenes. In: ICRA (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 247, + 480, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 480, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 480, + 268 + ], + "type": "text", + "content": "22. Jia, F., Mao, W., Liu, Y., Zhao, Y., Wen, Y., Zhang, C., Zhang, X., Wang, T.: Adriver-i: A general world model for autonomous driving. arXiv preprint (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "type": "text", + "content": "23. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: Lisa: Reasoning segmentation via large language model. arXiv preprint (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 289, + 480, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 321 + ], + "type": "text", + "content": "24. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 321, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 480, + 332 + ], + "type": "text", + "content": "25. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NIPS (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 332, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 342 + ], + "type": "text", + "content": "26. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "type": "text", + "content": "27. Malla, S., Choi, C., Dwivedi, I., Choi, J.H., Li, J.: Drama: Joint risk localization and captioning in driving. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "text", + "content": "28. Mao, J., Niu, M., Jiang, C., Liang, H., Chen, J., Liang, X., Li, Y., Ye, C., Zhang, W., Li, Z., et al.: One million scenes for autonomous driving: Once dataset. arXiv preprint (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 407, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 429 + ], + "type": "text", + "content": "29. Mao, J., Qian, Y., Zhao, H., Wang, Y.: Gpt-driver: Learning to drive with gpt. arXiv preprint (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 429, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 480, + 449 + ], + "type": "text", + "content": "30. OpenAI: Gpt-4: A large-scale transformer-based language model (2023), https://www.openai.com/research/gpt-4, https://www.openai.com/research/gpt-4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 449, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 480, + 471 + ], + "type": "text", + "content": "31. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a method for automatic evaluation of machine translation. In: ACL (2002)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 471, + 480, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 480, + 492 + ], + "type": "text", + "content": "32. Peng, B., Li, C., He, P., Galley, M., Gao, J.: Instruction tuning with gpt-4. arXiv preprint (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 492, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 492, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 492, + 480, + 514 + ], + "type": "text", + "content": "33. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 514, + 480, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 480, + 535 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 480, + 535 + ], + "type": "text", + "content": "34. Pi, R., Gao, J., Diao, S., Pan, R., Dong, H., Zhang, J., Yao, L., Han, J., Xu, H., Zhang, L.K.T.: Detgpt: Detect what you need via reasoning. arXiv preprint (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 536, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 567 + ], + "type": "text", + "content": "35. Qian, T., Chen, J., Zhuo, L., Jiao, Y., Jiang, Y.G.: Nuscenes-qa: A multi-modal visual question answering benchmark for autonomous driving scenario. arXiv preprint (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "text", + "content": "36. Sachdeva, E., Agarwal, N., Chundi, S., Roelofs, S., Li, J., Dariush, B., Choi, C., Kochenderfer, M.: Rank2tell: A multimodal driving dataset for joint importance ranking and reasoning. arXiv preprint (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "text", + "content": "37. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: CVPR (2020)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "38. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint (2023)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "Ming Nie, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 325 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "39. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: CVPR (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "type": "text", + "content": "40. Wu, D., Han, W., Wang, T., Dong, X., Zhang, X., Shen, J.: Referring multi-object tracking. In: CVPR (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 482, + 182 + ], + "type": "text", + "content": "41. Wu, D., Han, W., Wang, T., Liu, Y., Zhang, X., Shen, J.: Language prompt for autonomous driving. arXiv preprint (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "text", + "content": "42. Xu, Z., Zhang, Y., Xie, E., Zhao, Z., Guo, Y., Wong, K.K., Li, Z., Zhao, H.: Drivegpt4: Interpretable end-to-end autonomous driving via large language model. arXiv preprint (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 215, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 237 + ], + "type": "text", + "content": "43. Yu, P., Wang, T., Golovneva, O., Alkhamissy, B., Ghosh, G., Diab, M., Celikyilmaz, A.: Alert: Adapting language models to reasoning tasks. arXiv preprint (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "type": "text", + "content": "44. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 259, + 482, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 292 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 292 + ], + "type": "text", + "content": "45. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 293, + 482, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 325 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 325 + ], + "type": "text", + "content": "46. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint (2023)" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "text", + "content": "Reason2Drive" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Rebalancing Using Estimated Class Distribution for Imbalanced Semi-Supervised Learning under Class Distribution Mismatch/c1f56d15-2542-45a0-80c7-7ebab6520f43_content_list.json b/2024/Rebalancing Using Estimated Class Distribution for Imbalanced Semi-Supervised Learning under Class Distribution Mismatch/c1f56d15-2542-45a0-80c7-7ebab6520f43_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3fc97b92de32bb78ba44a0d54c989c1b18724096 --- /dev/null +++ b/2024/Rebalancing Using Estimated Class Distribution for Imbalanced Semi-Supervised Learning under Class Distribution Mismatch/c1f56d15-2542-45a0-80c7-7ebab6520f43_content_list.json @@ -0,0 +1,1901 @@ +[ + { + "type": "text", + "text": "Rebalancing Using Estimated Class Distribution for Imbalanced Semi-Supervised Learning under Class Distribution Mismatch", + "text_level": 1, + "bbox": [ + 225, + 140, + 779, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Taemin Park\\**, Hyuck Lee\\**, and Heeyoung Kim", + "bbox": [ + 307, + 234, + 692, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Department of Industrial and Systems Engineering, KAIST", + "bbox": [ + 302, + 262, + 699, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Daejeon 34141, Republic of Korea", + "bbox": [ + 387, + 277, + 614, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{ptm1001,dlgur0921,heeyoungkim}@kaist.ac.kr", + "bbox": [ + 334, + 291, + 666, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Despite significant advancements in class-imbalanced semi-supervised learning (CISSL), many existing algorithms explicitly or implicitly assume that the class distribution of unlabeled data matches that of labeled data. However, when this assumption fails in practice, the classification performance of such algorithms may degrade due to incorrectly assigned weight to each class during training. We propose a novel CISSL algorithm called Rebalancing Using Estimated Class Distribution (RECD). RECD estimates the unknown class distribution of unlabeled data through Monte Carlo approximation, leveraging predicted class probabilities for unlabeled samples, and subsequently rebalances the classifier based on the estimated class distribution. Additionally, we propose an extension of feature clusters compression in the context of CISSL to mitigate feature map imbalance by densifying minority class clusters. Experimental results on four benchmark datasets demonstrate that RECD achieves state-of-the-art classification performance in CISSL.", + "bbox": [ + 259, + 358, + 743, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Class-imbalanced semi-supervised learning $\\cdot$ Long-tailed learning $\\cdot$ Auxiliary Balanced Classifier", + "bbox": [ + 259, + 579, + 756, + 608 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 633, + 375, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In real-world datasets, the number of samples often varies across different classes. Deep neural network algorithms trained on such datasets tend to exhibit bias towards majority classes, which have a larger number of samples in the dataset [15, 31, 41]. To mitigate the bias, various class-imbalanced learning (CIL) techniques have been proposed, including re-sampling [4, 10, 30, 42], re-weighting [8, 29, 36], decoupled learning [19, 35, 53], and multi-expert learning [43, 48, 52].", + "bbox": [ + 212, + 666, + 787, + 757 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although the CIL techniques significantly mitigate the bias, they typically assume that the entire training set is labeled. However, in reality, labeling samples requires substantial resources, resulting in many real-world datasets containing", + "bbox": [ + 212, + 757, + 787, + 803 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution", + "bbox": [ + 230, + 811, + 374, + 825 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 230, + 825, + 392, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a significant proportion of unlabeled data. Because the CIL techniques assign importance to each data point based on its corresponding label, they cannot be directly applied when the training set includes unlabeled samples.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, several class-imbalanced semi-supervised learning (CISSL) algorithms, including ABC [27], UDAL [24], CReST [46], DARP [20], DASO [34], CoSSL [11], and SAW [23], have emerged to address the bias towards majority classes in scenarios where the training set consists of both labeled and unlabeled data. However, ABC, UDAL, and CReST explicitly assume that the class distribution of the unlabeled data matches that of the labeled data. Similarly, DARP, DASO, and SAW can be seen as implicitly assuming the class distribution match between the labeled and unlabeled data, as they incorporate CIL techniques such as LA [32] and cRT [19], which rely on the same class distribution between the labeled and unlabeled data. Additionally, CoSSL solely considers the class distribution of labeled data in its main component, TFE.", + "bbox": [ + 212, + 191, + 784, + 356 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, the class distribution of the unlabeled set is often unknown and may differ from that of the labeled set [34, 40, 47]. In this case, the classification performance of the aforementioned algorithms may degrade because the unknown class distribution of the unlabeled set is not considered in the rebalancing process. Fig. 1b illustrates the class predictions on the test set of CIFAR-10 [21] using ReMixMatch+ABC, an existing CISSL algorithm, trained on the long-tailed version of the CIFAR-10 [8] training set where the class distributions of labeled and unlabeled sets mismatch, as presented in Fig. 1a. From Fig. 1b, we can observe that many test samples are mispredicted as Class 10 (the most minor class). This may be because ABC rebalances the classifier excessively, considering only the class distribution of the labeled set.", + "bbox": [ + 212, + 357, + 785, + 523 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e1202bd77e9c19ecee6374f042f9c2d15ec144c619b9484c1d49e46de9ddfeb2.jpg", + "image_caption": [ + "(a) Class distributions" + ], + "image_footnote": [], + "bbox": [ + 217, + 546, + 390, + 638 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c84920ac757b6feb2e5d7b92219ee731492bdb788be55a9fb45bd01ae22453f0.jpg", + "image_caption": [ + "(b) ReMixMatch+ABC", + "Fig. 1: Comparison of the predictions of ReMixMatch+ABC and the proposed algorithm (ReMixMatch+RECD) trained on an imbalanced version of CIFAR-10 where the class distributions of labeled and unlabeled sets mismatch, as illustrated in Fig. 1a." + ], + "image_footnote": [], + "bbox": [ + 415, + 546, + 586, + 638 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8c66110307eb62a2654e3e21c59902955c2a9699d7b114a1ce8d60cf5a1659c9.jpg", + "image_caption": [ + "(c) Proposed algorithm" + ], + "image_footnote": [], + "bbox": [ + 612, + 547, + 784, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To rebalance the classifier appropriately when the class distribution of the unlabeled set mismatches that of the labeled set, we propose a novel CISSL algorithm called Rebalancing Using Estimated Class Distribution (RECD), which estimates the unknown class distribution using Monte Carlo (MC) approximation and rebalances the training loss based on the estimated class distribution. Specifically, RECD is a modular algorithm that operates in conjunction with previous CISSL algorithms. While RECD can generally be integrated with most", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Park et al.", + "bbox": [ + 271, + 114, + 343, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "CISSL algorithms by providing additional information about the class distribution of the unlabeled set, in this paper, we specifically combine RECD with ABC due to its simple and effective framework compared to other CISSL algorithms.", + "bbox": [ + 212, + 146, + 784, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To estimate the class distribution of unlabeled data, RECD first computes the average class predictions of ABC for the entire unlabeled set. Because generating class predictions for the entire unlabeled set can be computationally intensive, RECD generates class predictions on an unlabeled minibatch for each iteration and updates the exponential moving average (EMA) of the class predictions on the unlabeled set. Utilizing MC approximation, this average class prediction serves as the classifier's prior, representing the estimated unknown class distribution of the unlabeled set. Then, RECD rebalances the training loss on unlabeled samples for ABC based on the estimated class distribution of the unlabeled set, and trains ABC using this rebalanced unlabeled loss in addition to the training loss on labeled samples for ABC. As depicted in Fig. 1c, ReMix-Match+RECD generates nearly balanced class predictions on the test set.", + "bbox": [ + 212, + 191, + 785, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition to mitigating imbalance in the classifier, RECD also mitigates imbalance in the feature map. While existing CISSL algorithms primarily focus on mitigating classifier bias, there also exists imbalance in the feature map [5,18]. Specifically, feature clusters of minority classes tend to be sparser compared to the dense clusters of majority classes. These sparse features often overlap with clusters of other classes in the feature map, posing challenges in classifying minority class features [28]. To address this issue by densifying the sparse feature clusters of minority classes, we extend feature clusters compression (FCC) [28], a previous CIL technique, into the context of CISSL by leveraging the unknown class distribution of the unlabeled data estimated by RECD. We refer to this extended version of FCC as the adaptively adjusted feature multiplier (AAFM).", + "bbox": [ + 212, + 373, + 785, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our experimental results on CIFAR-10-LT [8], CIFAR-100-LT [8], STL-10-LT [20], and Small-ImageNet-127 [11] demonstrate the superior performance of the proposed algorithm compared to the baseline algorithms, regardless of whether the class distributions of labeled and unlabeled sets match or mismatch. Through detailed analyses, we verify that RECD effectively estimates the true class distribution of the unlabeled set and uses it to mitigate class imbalance. Additionally, we validate that AAFM not only reduces the test loss for minority classes but also stabilizes the training process. Finally, by conducting experiments that combine RECD with DARP [20], we validate the effectiveness of RECD beyond its application solely with ABC. The source codes for RECD are available at https://github.com/taemin-park/RECD.", + "bbox": [ + 212, + 539, + 785, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 215, + 729, + 382, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Many CISSL algorithms have been proposed to address class imbalance in scenarios where the training set consists of both labeled and unlabeled data. For example, ABC [27] trains an auxiliary balanced classifier with a training loss that is rebalanced based on the class distribution of the labeled data. CReST [46] systematically enlarges the labeled dataset by assigning the labels to unlabeled", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "RECD", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "samples predicted as minority classes more frequently than those predicted as majority classes. UDAL [24] combines progressive distribution alignment and logit-adjustment [32] to adjust the training loss. ABC, CReST, and UDAL have effectively mitigated bias towards majority classes. However, they assumed that the class distribution of the unlabeled data is the same as that of the labeled data, which may pose practical limitation as described in Sec. 1.", + "bbox": [ + 212, + 146, + 782, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "DARP [20] solves a convex optimization problem to refine biased pseudo labels. DASO [34] blends class predictions of a similarity-based classifier and a linear classifier to obtain unbiased pseudo-labels. SAW [23] estimates the learning difficulty of each class and smoothly reweights the training loss based on the estimated learning difficulty. However, DARP, DASO, and SAW implicitly assume the same class distributions of the unlabeled and labeled datasets, as they incorporate CIL techniques such as LA [32] and cRT [19], which rebalance the classifier based on the class distribution of the labeled data.", + "bbox": [ + 212, + 237, + 784, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CoSSL [11] integrates Tail-class Feature Enhancement (TFE) and class-balanced sampling to rebalance the classifier. InPL [50] introduces energy-based pseudolabeling to effectively discern correctly predicted unlabeled samples of minority classes. DePL [44] mitigates pseudo-label bias using counterfactual reasoning and adaptive marginal loss. L2AC [40] mitigates training bias by using a bias adaptive classifier, which consists of a bias attractor and a linear classifier. Adsh [14] sets an adaptive confidence threshold for each class to utilize unlabeled samples predicted as minority classes more frequently than those predicted as majority classes. ACR [47] estimates the class distribution of the unlabeled set by comparing class predictions with class distribution prototypes, and then adaptively adjusts the training loss based on the estimated class distribution. CDMAD [26] refines the predictions for unlabeled and test samples based on the logits for solid color image. Although these algorithms have significantly improved CISSL, this area remains relatively less explored compared to SSL and CIL.", + "bbox": [ + 212, + 357, + 803, + 569 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 590, + 382, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Problem settings", + "text_level": 1, + "bbox": [ + 215, + 619, + 401, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We have a training dataset that consists of the labeled set $\\mathcal{X} = \\{(x_n, y_n) : n \\in (1, \\ldots, N)\\}$ and unlabeled set $\\mathcal{U} = \\{(u_m) : m \\in (1, \\ldots, M)\\}$ , where $x_n \\in R^d$ and $y_n \\in R^K$ are the $n$ th labeled sample and its corresponding label, respectively, and $u_m \\in R^d$ is the $m$ th unlabeled sample. For each class $k$ , $N_k$ and $M_k$ denote the numbers of the labeled and unlabeled samples, respectively, i.e., $\\sum N_k = N$ and $\\sum M_k = M$ . In CISSL, it is typically assumed that $N_1 \\geq N_2 \\geq \\ldots \\geq N_K$ and $M_1 \\geq M_2 \\geq \\ldots \\geq M_K$ . We denote the class imbalance ratios of the labeled and unlabeled sets as $\\gamma_l = \\frac{N_1}{N_K}$ and $\\gamma_u = \\frac{M_1}{M_K}$ , respectively. When $M_k$ is unknown, $\\gamma_u$ would also be unknown, and it may differ from $\\gamma_l$ . For each training iteration, we generate minibatches $\\mathcal{MB}_{\\mathcal{X}} = \\{(x_b^m, y_b^m) : b \\in (1, \\ldots, B_l)\\} \\subset \\mathcal{X}$ and $\\mathcal{MB}_{\\mathcal{U}} = \\{(u_b^m) : b \\in (1, \\ldots, B_u)\\} \\subset \\mathcal{U}$ from $\\mathcal{X}$ and $\\mathcal{U}$ , respectively, where $B_l$ and $B_u$ are the sizes of the labeled and unlabeled minibatches, respectively. Using $\\mathcal{MB}_{\\mathcal{X}}$ and $\\mathcal{MB}_{\\mathcal{U}}$ , we aim to train an algorithm $f_\\theta: R^d \\to R^K$ that accurately predicts the", + "bbox": [ + 215, + 643, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Park et al.", + "bbox": [ + 271, + 114, + 343, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "class probabilities for unseen samples, where $\\theta$ denotes the network parameters. We denote the extracted features before the classification layer as $\\xi (\\cdot)$", + "bbox": [ + 214, + 146, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Backbone SSL algorithm", + "text_level": 1, + "bbox": [ + 214, + 199, + 465, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following previous CISSL algorithms [11, 20, 23, 27, 34, 50], either FixMatch [38] or ReMixMatch [1] is used as a backbone SSL algorithm for RECD. To utilize unlabeled samples, FixMatch and ReMixMatch conduct consistency regularization using weak data augmentation techniques $\\alpha(\\cdot)$ (flipping and cropping) and strong data augmentation techniques $\\mathcal{A}(\\cdot)$ (Cutout [9] and RandomAugment [6]).", + "bbox": [ + 212, + 223, + 784, + 299 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, FixMatch first predicts the class probabilities for $\\alpha(u_b^m)$ as $f(\\alpha(u_b^m))$ and then generates the pseudo-label as $q_b^m = \\arg \\max_k (f(\\alpha(u_b^m))_k)$ when the confidence exceeds the confidence threshold $\\epsilon$ , i.e., $\\max_k (f(\\alpha(u_b^m))_k) \\geq \\epsilon$ , where $(\\cdot)_k$ denotes the $k$ th element of $(\\cdot)$ . Then, FixMatch predicts the class probabilities for $\\mathcal{A}(u_b^m)$ as $f(\\mathcal{A}(u_b^m))$ and ensures $f(\\mathcal{A}(u_b^m))$ to be close to the pseudo-label $q_b^m$ by minimizing a loss that penalizes the distance between $f(\\mathcal{A}(u_b^m))$ and $q_b^m$ . In addition, FixMatch minimizes the prediction loss calculated between weakly augmented labeled sample $\\alpha(x_b^m)$ and $p_b^m$ , which is the one-hot version of $y_b^m$ .", + "bbox": [ + 212, + 299, + 803, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarly, ReMixMatch first predicts the class probabilities of $\\alpha(u_b^m)$ as $q_b^m = f(\\alpha(u_b^m))$ . Then ReMixMatch applies a distribution alignment technique as $\\tilde{q}_b^m = \\mathrm{Normalize}(q_b^m \\times q(y) / \\tilde{q}(y))$ , where $\\mathrm{Normalize}(x)_i = x_i / \\sum_j x_j$ , $\\tilde{q}(y)$ is the moving average of class predictions on unlabeled samples, and $q(y)$ indicates the class distribution of the training set. Finally, ReMixMatch sharpens $\\tilde{q}_b^m$ as $\\bar{q}_b^m = \\mathrm{Normalize}((\\tilde{q}_b^m)^{1/T})$ , where $T$ is the temperature for sharpening. Using $\\bar{q}_b^m$ as pseudo-labels, ReMixMatch conducts consistency regularization [33,37], entropy minimization [13,25] and mixup regularization [2,39,51]. ReMixMatch also minimizes the rotation loss [12] for self-supervised learning. The training loss of the backbone SSL algorithm, $L_{back}$ , is detailed in Appendix B.", + "bbox": [ + 212, + 420, + 785, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 Auxiliary balanced classifier (ABC)", + "text_level": 1, + "bbox": [ + 214, + 593, + 553, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To use a balanced classifier trained on a class-balanced subset of the training set, while benefiting from the high-quality representations learned from the entire training set, Lee and Kim [27] introduced an auxiliary balanced classifier (ABC) attached to the representation layer of a backbone SSL algorithm. To ensure that ABC is trained on a class-balanced subset of the training set, ABC generates a $0/1$ mask $M(\\cdot)$ for each training sample, where the $0/1$ mask determines whether each sample contributes to the training loss. Specifically, ABC samples a $0/1$ mask from a Bernoulli distribution for each labeled and unlabeled sample as", + "bbox": [ + 212, + 619, + 782, + 739 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM (x _ {b} ^ {m}) = \\mathcal {B} (\\frac {N _ {K}}{N _ {\\hat {y} _ {b} ^ {m}}}) \\quad \\mathrm {a n d} \\quad M (u _ {b} ^ {m}) = \\mathcal {B} (\\frac {N _ {K}}{N _ {\\hat {q} _ {b} ^ {m}}}). \\qquad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 750, + 784, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathcal{B}(\\cdot)$ denotes a Bernoulli distribution and $\\hat{q}_b^m = \\arg \\max (q_b^m)$ , where $q_{b}^{m}$ is the class prediction of the ABC for an unlabeled sample $u_{b}^{m}$ . Note that in Eq. (1), the probability of sampling 1 is adjusted to be inversely proportional to the number", + "bbox": [ + 214, + 794, + 784, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "RECD", + "bbox": [ + 683, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of training samples belonging to each class. This sampling strategy ensures that ABC is trained with an equal frequency from each class's data.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With the $0/1$ mask, ABC calculates the classification loss, $L_{cls}$ , between weakly augmented labeled sample $\\alpha(x_b^m)$ and its corresponding one-hot version label $p_b^m$ . In addition, ABC also calculates a consistency regularization loss, $L_{con}$ , between weakly augmented unlabeled sample $\\alpha(u_b^m)$ and strongly augmented unlabeled sample $\\mathcal{A}(u_b^m)$ , which is minimized to ensure that decision boundaries lie in low-density regions. The training losses $L_{cls}$ and $L_{con}$ are calculated as:", + "bbox": [ + 212, + 176, + 784, + 267 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c l s} = \\frac {1}{B _ {l}} \\sum_ {b = 1} ^ {B _ {l}} M \\left(x _ {b} ^ {m}\\right) \\mathcal {H} \\left(p _ {b} ^ {m}, f _ {a b c} (\\xi \\left(\\alpha \\left(x _ {b} ^ {m}\\right)\\right))\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 280, + 785, + 319 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c o n} = \\frac {1}{B _ {u}} \\sum_ {b = 1} ^ {B _ {u}} \\sum_ {i = 1} ^ {2} M \\left(u _ {b} ^ {m}\\right) \\mathbb {1} \\left(\\max \\left(q _ {b} ^ {m}\\right) \\geq \\epsilon\\right) \\mathcal {H} \\left(q _ {b} ^ {m}, f _ {a b c} \\left(\\xi \\left(\\mathcal {A} _ {i} \\left(u _ {b} ^ {m}\\right)\\right)\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 264, + 334, + 784, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{H}$ is the cross-entropy loss, $q_{b}^{m} = f_{abc}(\\xi (\\alpha (u_{b}^{m})))$ , $f_{abc}$ denotes the auxiliary classifier attached to the representation layer, and $\\mathbb{1}(\\cdot)$ denotes an indicator function. With the $L_{cls}$ , $L_{con}$ and the training loss of backbone SSL algorithm $L_{back}$ , total training loss of ABC is calculated as:", + "bbox": [ + 214, + 380, + 782, + 441 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {A B C} = L _ {c l s} + L _ {c o n} + L _ {b a c k}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 454, + 784, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ABC [27] effectively mitigated class imbalance when the class distributions of the labeled and unlabeled sets match. However, as we can observe in the second equation of the Eq. (1), because the $0/1$ masks for unlabeled samples are generated based on the class distribution of the labeled set without considering that of the unlabeled set, $f_{abc}$ may not be properly rebalanced when the unknown class distribution of the unlabeled set significantly mismatches that of the labeled set.", + "bbox": [ + 212, + 479, + 787, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4 Feature clusters compression (FCC) [28]", + "text_level": 1, + "bbox": [ + 214, + 593, + 593, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To separate features extracted from samples of minority classes from those of majority classes by densifying the feature clusters, Li et al. [28] introduced FCC. For a training sample $x$ of class $k$ , FCC multiplies the extracted feature $\\xi(x)$ by a scaling factor $\\tau_k$ , and uses $\\tau_k \\times \\xi(x)$ as input for the classification layer $\\psi$ , as", + "bbox": [ + 212, + 619, + 782, + 679 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\theta} (x) = \\psi \\left(\\tau_ {k} \\times \\xi (x)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 693, + 784, + 708 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\tau_{k}$ decreases from majority to minority classes in a naive manner as follows:", + "bbox": [ + 214, + 720, + 782, + 734 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tau_ {k} = 1 + \\eta \\times \\left(1 - \\frac {k}{K}\\right), k = 1, \\dots , K, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 746, + 784, + 775 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\eta$ is a scale hyperparameter. Since the features of majority classes are multiplied by relatively large scaling factors compared to those of minority classes, their feature clusters occupy larger regions on the feature map during training. In this context, to prevent minority class features from crossing the decision", + "bbox": [ + 212, + 779, + 782, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Park et al.", + "bbox": [ + 271, + 114, + 343, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "boundary, the feature extractor $\\xi(\\cdot)$ would be trained to map minority class features into denser regions. Note that during the test phase, the features extracted from test samples are not multiplied by the scaling factor, ensuring that the classifier predicts classes for test samples with densified features. According to Li et al. [28], FCC effectively enhances the separability of feature clusters of minority classes in CIL. However, in CISSL, it may fail to appropriately densify minority class features because the class distribution of the unlabeled set cannot be considered in the scaling factor $\\tau_{k}$ . In this paper, we refer to the scaling factor in Eq. (6) as the naive feature multiplier.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Methodology", + "text_level": 1, + "bbox": [ + 215, + 305, + 380, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Rebalancing using estimated class distribution (RECD)", + "text_level": 1, + "bbox": [ + 214, + 337, + 718, + 353 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As described in Sec. 1, RECD incorporates the unknown class distribution of the unlabeled data by rebalancing the classifier based on the estimated class distribution of the unlabeled data. RECD is integrated with ABC [27], which rebalances the training loss using the $0/1$ mask, as described in Eq. (1). Whereas ABC generates $0/1$ masks for unlabeled samples based on the class distribution of the labeled set, RECD replaces the parameter of $0/1$ masks for unlabeled samples using the estimated class distribution of the unlabeled set.", + "bbox": [ + 212, + 362, + 787, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, to estimate the unknown class distribution of the unlabeled set, $p(y_{u})$ , RECD first computes the average class predictions of ABC on the entire unlabeled set as $\\frac{1}{M}\\sum_{m = 1}^{M}p_{\\theta}(y_u|u_m)$ , where $p_{\\theta}(y_u|u_m) = f_{abc}(\\xi (\\alpha (u_m)))$ . Then, using MC approximation, $\\frac{1}{M}\\sum_{m = 1}^{M}p_{\\theta}(y_u|u_m)$ becomes $p_{\\theta}(y_u)$ as follows:", + "bbox": [ + 212, + 468, + 787, + 535 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {1}{M} \\sum p _ {\\theta} \\left(y _ {u} \\mid u _ {m}\\right) \\approx \\int p _ {\\theta} \\left(y _ {u} \\mid u\\right) p (u) d u = p _ {\\theta} \\left(y _ {u}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 546, + 784, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Eq. (7), $p_{\\theta}(y_u)$ can be regarded as the classifier's prior for the unlabeled samples, and we use this classifier's prior as the estimation of the class distribution of the unlabeled set $p(y_u)$ based on that the classifier's prior is highly correlated with the class distribution, as analyzed in Fig. 3.", + "bbox": [ + 212, + 585, + 787, + 647 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Using the estimated class distribution $p_{\\theta}(y_u)$ , RECD modifies the sampling distribution of the $0/1$ mask for unlabeled samples in Eq. (1) as follows:", + "bbox": [ + 212, + 647, + 787, + 678 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nM \\left(u _ {b} ^ {m}\\right) = \\mathcal {B} \\left(\\frac {\\operatorname* {m i n} \\left(p _ {\\theta} \\left(y _ {u}\\right)\\right)}{p _ {\\theta} \\left(y _ {u}\\right) \\hat {q} _ {b} ^ {m}}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 690, + 784, + 722 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "By setting the parameter of the $0/1$ mask based on $p_{\\theta}(y_u)$ , RECD rebalances the training loss of the unlabeled samples to an appropriate extent even when the class distributions of the labeled and unlabeled sets severely mismatch. The total training loss of RECD is described in detail in Sec. 4.3.", + "bbox": [ + 212, + 733, + 787, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "RECD may increase computational complexity compared to the original ABC because, as shown in Eq. (7), computing class predictions of ABC for the entire unlabeled set for each iteration or epoch would require a substantial amount of", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "RECD", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "computation, especially for large-scale datasets such as LSUN [49] or iNaturallist [7]. To address this issue, RECD replaces the naive average $\\frac{1}{M}\\sum p_{\\theta}(y_u|u_m)$ with an EMA of the class predictions on the unlabeled samples, which is updated for each iteration of the training as follows:", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(y _ {u}\\right) _ {n e w} = \\rho \\times p _ {\\theta} \\left(y _ {u}\\right) _ {o l d} + (1 - \\rho) \\times p _ {\\theta} \\left(y _ {u}\\right) _ {b a t c h}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 220, + 785, + 238 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\rho$ is a hyperparameter for the EMA update, $p_{\\theta}(y_u)_{old}$ is $p_{\\theta}(y_u)$ before the update, and $p_{\\theta}(y_u)_{batch}$ is the average class prediction on the weakly augmented unlabeled minibatch, $\\frac{1}{B_u}\\sum_{b = 1}^{B_u}q_b^m = \\frac{1}{B_u}\\sum_{b = 1}^{B_u}f_{abc}(\\xi (\\alpha (u_b^m)))$ . Updating the EMA of the class predictions on unlabeled samples adds negligible computational cost compared to the original ABC because RECD reuses $q_b^m$ that is calculated to conduct consistency regularization for ABC, as in Eq. (3).", + "bbox": [ + 212, + 250, + 787, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Adaptively adjusted feature multiplier (AAFM)", + "text_level": 1, + "bbox": [ + 214, + 364, + 656, + 382 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although the naive feature multiplier effectively densifies feature clusters of the minority classes in CIL, its efficacy may diminish in CISSL due to its lack of consideration for the class distribution of the unlabeled set. In addition, the naive feature multiplier linearly reduces from majority to minority classes as shown in Eq. (6), without considering the degree of class imbalance in the training set. These limitations may lead to inappropriate scaling factors for the extracted features, resulting in insufficient densification or unstable training, as discussed in Sec. 5.3. To address this issue by incorporating both the class imbalance ratio of the labeled set and the unknown class distribution of the unlabeled set, we introduce an advanced version of the naive feature multiplier, AAFM. For the labeled samples, AAFM for the $k$ th class, $\\tau_k^l$ , is calculated as follows:", + "bbox": [ + 212, + 393, + 787, + 561 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\tau_ {k} ^ {l} = 1 + \\eta \\times \\left(\\frac {N _ {k}}{N _ {1}} \\times C _ {l k} ^ {t}\\right), k = 1, \\dots , K, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 571, + 785, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\eta$ is a scale hyperparameter similar to $\\eta$ in Eq. (6), and $t$ is a hyperparameter for $C_{lk}$ , which represents the EMA of confidence for labeled samples of the $k$ th class, which is updated for each iteration of the training process as follows:", + "bbox": [ + 212, + 614, + 787, + 659 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nC _ {l k} = \\omega \\times C _ {l k} ^ {\\text {o l d}} + (1 - \\omega) \\times C _ {l k} ^ {\\text {b a t c h}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 671, + 785, + 690 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\omega$ is a hyperparameter for the EMA update, $C_{lk}^{old}$ is $C_{lk}$ before the update, and $C_{lk}^{batch}$ is the average confidence for class $k$ on a minibatch, which is calculated as $\\frac{1}{n_k}\\sum_{b = 1}^{B_l}f_{abc}(\\xi (\\alpha (x_b^m)))_k\\times \\mathbb{1}(y_b^m = k)$ . Here, $n_k$ denotes the number of samples for the $k$ th class on a minibatch. Note that $C_{lk}$ is updated only when $n_k\\geq 1$ . By multiplying $C_{lk}^{t}$ to the second term on the right-hand side of Eq. (10), we stabilize the training process, assigning smaller AAFMs to classes whose feature clusters are located near the decision boundary.", + "bbox": [ + 212, + 700, + 787, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also multiply AAFM $\\tau_{k}^{u}$ with the features extracted from strongly augmented unlabeled samples, where $\\tau_{k}^{u}$ incorporates the unknown class distribution", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Park et al.", + "bbox": [ + 271, + 114, + 343, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of the unlabeled set by reusing the estimated class distribution $p_{\\theta}(y_u)$ as follows:", + "bbox": [ + 214, + 146, + 782, + 161 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\tau_ {k} ^ {u} = 1 + \\eta \\times \\left(\\frac {p _ {\\theta} \\left(y _ {u}\\right) _ {k}}{\\operatorname* {m a x} \\left(p _ {\\theta} \\left(y _ {u}\\right)\\right)} \\times C _ {u k} ^ {t}\\right), k = 1, \\dots , K. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 178, + 784, + 209 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In Eq. (12), $C_{uk}$ is an EMA of confidence for unlabeled samples predicted as the $k$ th class, which is updated for each iteration of the training process as follows:", + "bbox": [ + 214, + 218, + 782, + 250 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nC _ {u k} = \\omega \\times C _ {u k} ^ {\\text {o l d}} + (1 - \\omega) \\times C _ {u k} ^ {\\text {b a t c h}}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 262, + 784, + 280 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $C_{uk}^{batch} = \\frac{1}{m_k}\\sum_{b = 1}^{B_u}f_{abc}(\\xi (\\alpha (u_b^m)))_k\\times \\mathbb{1}(\\hat{q}_b^m = k)$ , with $m_{k} = \\sum_{b = 1}^{B_{u}}\\mathbb{1}(\\hat{q}_{b}^{m} = k)$ and $\\hat{q}_b^m = \\arg \\max_k(q_b^m)_k$ . $C_{uk}$ is updated only when $m_{k}\\geq 1$ . Since $\\tau_k^u$ incorporates the unknown class distribution of the unlabeled set, AAFM can densify feature clusters of minority classes to an appropriate extent even when the class distributions of the labeled and unlabeled sets significantly mismatch.", + "bbox": [ + 214, + 292, + 784, + 372 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 End-to-end training of RECD", + "text_level": 1, + "bbox": [ + 215, + 402, + 506, + 417 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Fig. 2 shows the end-to-end training process of the proposed algorithm for labeled and unlabeled data, respectively. The total training loss of RECD, denoted as $L_{RECD}$ , closely resembles that of ABC in Eq. (4), and is formulated as follows:", + "bbox": [ + 214, + 434, + 782, + 479 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {R E C D} = L _ {c l s} + L _ {c o n} + L _ {b a c k}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 494, + 784, + 510 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $L_{cls}$ and $L_{con}$ are improved by applying RECD and AAFM as follows:", + "bbox": [ + 214, + 523, + 769, + 540 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c l s} = \\frac {1}{B _ {l}} \\sum_ {b = 1} ^ {B _ {l}} M \\left(x _ {b} ^ {m}\\right) \\mathcal {H} \\left(p _ {b} ^ {m}, f _ {a b c} \\left(\\xi \\left(\\alpha \\left(x _ {b} ^ {m}\\right)\\right) \\times \\tau_ {y _ {b} ^ {m}} ^ {l}\\right)\\right), \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 320, + 555, + 784, + 594 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c o n} = \\frac {1}{B _ {u}} \\sum_ {b = 1} ^ {B _ {u}} \\sum_ {i = 1} ^ {2} M \\left(u _ {b} ^ {m}\\right) \\mathcal {H} \\left(q _ {b} ^ {m}, f _ {a b c} \\left(\\xi \\left(\\mathcal {A} _ {i} \\left(u _ {b} ^ {m}\\right)\\right) \\times \\tau_ {\\hat {q} _ {b} ^ {m}} ^ {u}\\right)\\right), \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 300, + 616, + 784, + 656 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nM \\left(x _ {b} ^ {m}\\right) = \\mathcal {B} \\left(\\frac {N _ {K}}{N _ {y _ {b} ^ {m}}}\\right) \\quad \\text {a n d} \\quad M \\left(u _ {b} ^ {m}\\right) = \\mathcal {B} \\left(\\frac {\\min \\left(p _ {\\theta} \\left(y _ {u}\\right)\\right)}{p _ {\\theta} \\left(y _ {u}\\right) \\hat {q} _ {b} ^ {m}}\\right), \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 307, + 672, + 784, + 707 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $p_b^m$ is the one-hot version of $y_b^m$ , $q_b^m = f_{abc}(\\xi(\\alpha(u_b^m)))$ , and $\\tau_{y_b^m}^l$ and $\\tau_{\\hat{q}_b^m}^u$ are AAFMs corresponding to $y_b^m$ and $\\hat{q}_b^m$ , respectively. To improve the quality of features learned by the backbone SSL algorithm, $L_{back}$ is also adjusted with two modifications. First, AAFM is applied to further densify feature clusters of minority classes during the learning of the backbone SSL algorithm. Second, to train the backbone algorithm with unbiased pseudo-labels, pseudo-labels of the backbone algorithm are generated by $f_{abc}$ instead of the classifier of the backbone algorithm. The pseudo-code of RECD is presented in Appendix D.", + "bbox": [ + 214, + 717, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "RECD", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/1fc43e714bf55d86b81286c90986c5acdd3f53a84fe6f24b363135a6d6a80d50.jpg", + "image_caption": [ + "(a) Balanced training with labeled data" + ], + "image_footnote": [], + "bbox": [ + 272, + 142, + 519, + 257 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/fa73259374f230d43d91cee017ef1b4125b3aeaac7bf67b8dda584b3d0191815.jpg", + "image_caption": [ + "(b) Balanced training with unlabeled data", + "Fig. 2: End-to-end training of the proposed algorithm" + ], + "image_footnote": [], + "bbox": [ + 526, + 142, + 807, + 258 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 321, + 374, + 339 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Experimental setup", + "text_level": 1, + "bbox": [ + 215, + 354, + 423, + 371 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Following previous CISSL studies [11, 20, 23, 34, 40], we conducted experiments on CIFAR-10-LT [8], CIFAR-100-LT [8], STL10-LT [20], and Small-ImageNet-127 [11]. The detailed descriptions for the datasets are provided in Appendix E.1. We evaluated the classification performance using balanced accuracy (bACC) (referred to as the averaged class recall in [11]) [15] and geometric mean (GM) [22]. The performance of the proposed algorithm was compared with the following baseline algorithms: vanilla algorithm, CIL algorithms [3, 16, 19], SSL algorithms [1, 38], and CISSL algorithms [11, 14, 20, 23, 24, 27, 40, 44, 46]. Algorithms such as DASO [34] and ACR [47], whose classification performances were measured under different experimental settings from ours, were compared with RECD in Appendix G. The detailed descriptions of the baseline algorithms and training setups are provided in Appendix E.2.", + "bbox": [ + 212, + 381, + 787, + 564 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Experimental results", + "text_level": 1, + "bbox": [ + 215, + 585, + 433, + 602 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Tab. 1 summarizes the classification performances of the baseline algorithms and RECD on CIFAR-10-LT under $\\gamma_l = \\gamma_u$ . First, the CISSL algorithms achieved higher performances than the vanilla algorithm, CIL algorithms, and SSL algorithms. These results verify the importance of mitigating class imbalance and utilizing unlabeled samples. Furthermore, RECD achieved the highest performance among the CISSL algorithms. Specifically, RECD showed a tendency to outperform other CISSL algorithms by a large margin as the class imbalance ratio increased, indicating its effectiveness in mitigating class imbalance. These results may be attributed to the unique ability of the proposed algorithm to mitigate class imbalance in the feature map, as we analyze in Sec. 5.3.", + "bbox": [ + 212, + 612, + 787, + 763 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Tab. 2 summarizes the performances of the baseline CISSL algorithms and RECD on CIFAR-100-LT. RECD outperformed the baseline algorithms. Given that CIFAR-100-LT has 100 classes, these results suggest that RECD may also be suitable for datasets with a large number of classes. Considering that classes with the fewest labeled data points have only three labeled samples under $\\gamma = 50$ ,", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Park et al.", + "bbox": [ + 271, + 114, + 344, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/733078891b924ee6580e601c2cba91670d16f9284a32e273df4c6fba53625abe.jpg", + "table_caption": [ + "Table 1: Comparison of bACC/GM on CIFAR-10-LT under $\\gamma = {\\gamma }_{l} = {\\gamma }_{u}$" + ], + "table_footnote": [], + "table_body": "
CIFAR-10-LT (γ = γt = γu)
Algorithmγ = 50γ = 100γ = 150
Vanilla (Cross-Entropy Loss)65.2±0.05/61.1±0.0958.8±0.13/58.2±0.1155.6±0.43/
Re-sampling [17]64.3±0.48/60.6±0.6755.8±0.47/45.1±0.3052.2±0.05/
LDAM-DRV [3]68.9±0.07/67.0±0.0862.8±0.17/58.9±0.6057.9±0.20/
cRT [19]67.8±0.13/66.3±0.1563.2±0.45/59.9±0.4059.3±0.10/
FixMatch [38]79.2±0.33/77.8±0.3671.5±0.72/66.8±1.5168.4±0.15/
FixMatch+DARP+cRT [19, 20]85.8±0.43/85.6±0.5682.4±0.26/81.8±0.1779.6±0.42/
FixMatch+CRST+LA [32, 46]85.6±0.36/81.9±0.4581.2±0.70/74.5±0.9971.9±2.24/
FixMatch+ABC [27]85.6±0.26/85.2±0.2981.1±1.14/80.3±1.2977.3±1.25/
FixMatch+CoSSL [11]86.8±0.30/86.6±0.2583.2±0.49/82.7±0.6080.3±0.55/
FixMatch+SAW+LA [23, 32]86.2±0.15/83.9±0.3580.7±0.15/77.5±0.2173.7±0.06/
FixMatch+Adsh [14]83.4±0.06/-76.5±0.35/-71.5±0.30/
FixMatch+DebiasPL [44]- / --80.6±0.50/-- / -
FixMatch+UDAL [24]86.5±0.29/-81.4±0.39/-77.9±0.33/
FixMatch+L2AC [40]- / --82.1±0.57/81.5±0.6477.6±0.53/
FixMatch+RECD87.3±0.18/87.2±0.1984.0±0.13/83.6±0.1680.6±0.53/
ReMixMatch [1]81.5±0.26/80.2±0.3273.8±0.38/69.5±0.8469.9±0.47/
ReMixMatch+DARP+cRT [19, 20]87.3±0.61/87.0±0.1183.5±0.07/83.1±0.0979.7±0.54/
ReMixMatch+CRST+LA [32, 46]84.2±0.11/-81.3±0.34/-79.2±0.31/
ReMixMatch+ABC [27]87.9±0.47/87.6±0.5184.5±0.32/84.1±0.3680.5±1.18/
ReMixMatch+CoSSL [11]87.7±0.21/87.6±0.2584.1±0.56/83.7±0.6681.3±0.83/
ReMixMatch+SAW+cRT [19, 23]87.6±0.21/87.4±0.2685.4±0.32/83.9±0.2179.9±0.15/
ReMixMatch+RECD88.1±0.13/87.9±0.1285.4±0.40/85.2±0.4182.5±0.33/
", + "bbox": [ + 277, + 171, + 728, + 397 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "these results also demonstrate the possibility that RECD can outperform the baseline algorithms even when labeled samples are extremely scarce.", + "bbox": [ + 212, + 429, + 784, + 460 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Tab. 3 summarizes the performances of the baseline CISSL algorithms and RECD on Small-ImageNet-127. For both $32 \\times 32$ and $64 \\times 64$ version, RECD significantly outperformed the baseline algorithms. Given that the class distribution of the test set of Small-ImageNet-127 is also imbalanced like the training set, these results show that RECD may also be suitable for datasets with imbalanced test sets. Moreover, given that Small-ImageNet-127 is an large-scale dataset, these results verify that RECD can also be applied to large-scale datasets.", + "bbox": [ + 212, + 460, + 784, + 566 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Tab. 4 summarizes the performances of the baseline algorithms and RECD evaluated on CIFAR-10-LT under $\\gamma_l \\neq \\gamma_u$ and STL-10-LT under unknown", + "bbox": [ + 212, + 568, + 785, + 599 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/f97159c492961423c0379a77cc655d04a396e837eedb34cd8b97a4fd14fcafbc.jpg", + "table_caption": [ + "Table 2: Comparison of bACC on CIFAR-100-LT under $\\gamma = {\\gamma }_{l} = {\\gamma }_{u}$" + ], + "table_footnote": [], + "table_body": "
CIFAR-100-LT (γ = γt = γu)
Algorithmγ = 20γ = 50
FixMatch [38]49.6±0.7842.1±0.33
FixMatch+DARP [20]50.8±0.7743.1±0.54
FixMatch+DARP+cRT [19, 20]51.4±0.6844.9±0.54
FixMatch+CReST [46]51.8±0.1244.9±0.50
FixMatch+CReST+LA [32, 46]52.9±0.0747.3±0.17
FixMatch+ABC [27]53.3±0.7946.7±0.26
FixMatch+CoSSL [11]53.9±0.7847.6±0.57
FixMatch+RECD54.6±0.3647.8±0.17
ReMixMatch [1]51.6±0.4344.2±0.59
ReMixMatch+DARP [20]51.9±0.3544.7±0.66
ReMixMatch+DARP+cRT [19, 20]54.5±0.4248.5±0.91
ReMixMatch+CReST [46]51.3±0.3445.5±0.76
ReMixMatch+CReST+LA [32, 46]51.9±0.6046.6±1.14
ReMixMatch+ABC [27]55.6±0.3547.9±0.10
ReMixMatch+CoSSL [11]55.8±0.6248.9±0.61
ReMixMatch+RECD55.9±0.3649.5±0.21
", + "bbox": [ + 364, + 655, + 633, + 837 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "RECD", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/50e9c35407365d3b9af57d5b7c614d858a50b4d852c8c67d1a8d187a7307573a.jpg", + "table_caption": [ + "Table 3: Comparison of bACC on Small-ImageNet-127" + ], + "table_footnote": [], + "table_body": "
Small-ImageNet-127 (γ = γl = γu)
Algorithm32 × 3264 × 64
FixMatch [38]29.742.3
FixMatch+DARP [20]30.542.5
FixMatch+DARP +cRT [19,20]39.751.0
FixMatch+CReST+ [46]32.544.7
FixMatch+CReST++LA [32,46]40.955.9
FixMatch+CoSSL [11]43.753.8
FixMatch+RECD47.359.5
", + "bbox": [ + 377, + 170, + 619, + 263 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "$\\gamma_{u}$ . RECD outperformed the baseline algorithms by large margins. These results demonstrate that RECD effectively mitigates class imbalance even when $\\gamma_{l} \\neq \\gamma_{u}$ , by considering the unknown class distribution of the unlabeled set. Notably, ReMixMatch+RECD outperformed the baseline algorithms that use ReMixMatch* [20] (ReMixMatch that uses the estimated class distribution of the unlabeled set) as the backbone SSL algorithm. These results implicitly demonstrate that RECD estimates class distribution better than the previous estimation algorithm. Additionally, under severe class distribution mismatch between the labeled and unlabeled sets, LA degraded the performance of ReMixMatch*+DARP and ReMixMatch*+SAW. This may be because the unknown class distribution of the unlabeled set cannot be incorporated into LA, emphasizing the importance of considering the class distribution of the unlabeled set.", + "bbox": [ + 212, + 303, + 787, + 484 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/4a64f8df6452e05e3a7e16af98c202f8e88e4fb1ab057904fa1b73cde81daca2.jpg", + "table_caption": [ + "Table 4: Comparison of bACC/GM on CIFAR-10-LT and STL-10-LT under $\\gamma_l \\neq \\gamma_u$ ." + ], + "table_footnote": [], + "table_body": "
CIFAR-10-LT (γl=100)STL-10-LT (γu=Unknown)
Algorithmγu=1γu=50γu=150γl=10γl=20
FixMatch [38]68.9±1.95/42.8±8.1173.9±0.25/70.5±0.5269.6±0.60/62.6±1.1172.9±0.09/69.6±0.0163.4±0.21/52.6±0.09
+DARP [20]85.4±0.5585.0±0.6577.3±0.17/75.5±0.2172.9±0.24/69.5±0.1877.8±0.33/76.5±0.4069.9±1.77/65.4±0.97
+DARP+LA [20,32]86.6±1.11/86.2±1.1582.3±0.32/81.5±0.2978.9±0.23/77.7±0.0678.6±0.30/77.4±0.4071.9±0.49/68.7±0.51
+DARP+cRT [19,20]87.0±0.70/86.8±0.6782.7±0.21/82.3±0.2580.7±0.44/80.2±0.6179.3±0.23/78.7±0.2174.1±0.61/73.1±1.21
+ABC [27]82.7±0.49/81.9±0.6882.7±0.64/82.0±0.7678.4±0.87/77.2±1.0779.1±0.46/78.1±0.5773.8±0.15/72.1±0.15
+SAW [23]81.2±0.68/80.2±0.9179.8±0.25/79.1±0.3274.5±0.97/72.5±1.3778.3±0.25/77.0±0.1971.9±0.81/69.0±0.81
+SAW+LA [23,32]84.5±0.68/84.1±0.7882.9±0.38/82.6±0.3879.1±0.81/78.6±0.91-/--/-
+SAW+cRT [19,23]84.6±0.23/84.4±0.2681.6±0.38/81.3±0.3277.6±0.40/77.1±0.41-/--/-
+RECD90.2±0.57/90.0±0.6485.6±0.15/85.3±0.1582.3±0.39/81.8±0.4581.4±0.28/80.6±0.3879.0±0.48/78.1±0.54
ReMixMatch [1]48.3±0.14/19.5±0.8575.1±0.43/71.9±0.7772.5±0.10/68.2±0.3267.8±0.45/61.1±0.9260.1±1.18/44.9±1.52
+ABC [27]76.4±5.34/74.8±6.0585.2±0.20/84.7±0.2580.4±0.40/80.0±0.4476.8±0.52/74.8±0.6471.2±1.37/67.4±1.89
ReMixMatch* [1]85.0±1.35/84.3±1.5577.0±0.12/74.7±0.0472.8±0.10/68.8±0.2176.7±0.15/73.9±0.3267.7±0.46/60.3±0.76
+DARP [20]86.9±0.10/86.4±0.1577.4±0.22/75.0±0.2573.2±0.11/69.2±0.3179.4±0.07/78.2±0.1070.9±0.44/67.0±1.62
+DARP+LA [20,32]81.8±0.45/80.9±0.4083.9±0.42/83.4±0.4581.1±0.20/80.3±0.2680.6±0.45/79.6±0.5576.8±0.60/74.8±0.68
+DARP+cRT [19,20]88.7±0.25/88.5±0.2583.5±0.53/83.1±0.5180.9±0.25/80.3±0.3180.9±0.53/80.0±0.4676.7±0.50/74.9±0.70
+SAW [23]87.0±0.75/86.4±0.8580.6±1.57/79.2±1.1977.6±0.76/76.0±0.9382.0±0.55/81.0±0.6479.2±0.44/77.9±0.52
+SAW+LA [23,32]74.2±1.49/65.1±2.3684.8±1.07/82.4±2.3281.3±2.42/80.9±2.47-/--/-
+SAW+cRT [19,23]88.8±0.79/88.6±0.8384.5±0.78/83.6±1.2782.4±0.10/82.0±0.10-/--/-
ReMixMatch+RECD90.3±0.40/90.2±0.4186.8±0.17/86.6±0.1883.9±0.11/83.7±0.1284.9±0.41/84.4±0.4782.5±0.27/81.7±0.31
", + "bbox": [ + 218, + 550, + 785, + 739 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In addition to the above results, we demonstrate the effectiveness of RECD with fewer labeled samples in Appendix J. Furthermore, we verify that RECD can be effectively combined with recent SSL algorithms by conducting experiments using FreeMatch [45] as the backbone SSL algorithm in Appendix I.", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Park et al.", + "bbox": [ + 271, + 114, + 343, + 126 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Detailed analyses on the proposed algorithm", + "text_level": 1, + "bbox": [ + 215, + 146, + 630, + 162 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We argued that RECD effectively estimates the unknown class distribution of the unlabeled set and mitigates class imbalance using this estimation. To verify this argument, we plotted the true class distribution of the unlabeled set, the class distribution of the unlabeled set estimated by RECD, and the class distribution of the labeled set in Fig. 3. We can observe that RECD accurately estimated the unknown class distribution of the unlabeled set, even when it significantly differed from that of the labeled set. This accurate class distribution estimation enabled the proposed algorithm to be trained with a properly rebalanced training loss and accurately calculated AAFM, mitigating class imbalance as in Fig. 4.", + "bbox": [ + 212, + 172, + 787, + 310 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a07408fa87cca9fa5fefce2e1db239452512b8ad72d069f97c1dfe26d9e9df3f.jpg", + "image_caption": [ + "(a) FixMatch+RECD" + ], + "image_footnote": [], + "bbox": [ + 282, + 335, + 488, + 438 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/72704aba64fa1b059fe50bbd0905d4c3b7df3f12579516ed56d8f36c223244ab.jpg", + "image_caption": [ + "(b) RemixMatch+RECD", + "Fig. 3: True class distributions of the labeled and unlabeled sets, along with the class distribution of the unlabeled set estimated by RECD. FixMatch+RECD and Remix-Match+RECD were trained on CIFAR-10-LT with $\\gamma_{l} = 100$ and $\\gamma_{u} = 1$ ." + ], + "image_footnote": [], + "bbox": [ + 511, + 335, + 723, + 438 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Fig. 4 presents the confusion matrices of class predictions on the test set of CIFAR-10 generated by each algorithm trained on the training set of CIFAR-10-LT under $\\gamma_l = 100$ and $\\gamma_u = 1$ . The $(i,j)$ position of each confusion matrix represents the proportion of the $i$ th class samples classified as the $j$ th class. In the figures, the left three algorithms often misclassified minority classes as majority classes. On the other hand, with both RECD and AAFM, the algorithm classified most samples of minority classes accurately. These results demonstrate that both RECD and AAFM effectively mitigate class imbalance when the class distribution of the labeled and unlabeled sets significantly mismatch.", + "bbox": [ + 212, + 537, + 787, + 672 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Additionally, we argued that AAFM not only reduces the test losses for minority classes but also stabilizes the training process. To verify this, we measured the overall test loss and the test losses of minority classes for FixMatch+RECD without AAFM and FixMatch+RECD with AAFM trained on CIFAR-10-LT. From Fig. 5, we can observe that the test losses for minority classes tended to increase as the training proceeds when AAFM was not used. On the other hand, the overall test loss and the test losses for the minority classes consistently decreased when AAFM was used. We can also observe that whereas the test losses for minority classes fluctuated when AAFM was not used, they stably decreased when AAFM was used. In addition, we compared the performance of the naive feature multiplier and AAFM with FixMatch+RECD in Appendix K. Finally,", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "RECD", + "bbox": [ + 683, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f7455dbc685d9de10c884e4c212d2a1fd40ccaa9fb8896a8d37d959170fd0d6c.jpg", + "image_caption": [ + "(a) FixMatch" + ], + "image_footnote": [], + "bbox": [ + 240, + 143, + 370, + 244 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/0bb3d89f794a8e74e448bab48f42a7a0336aa390bf8b71a5a020a3af1ef82366.jpg", + "image_caption": [ + "(b) Without RECD" + ], + "image_footnote": [], + "bbox": [ + 372, + 143, + 501, + 242 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/f0fee412f9bae0708ed2d5bfd441210ae21eb645a618d3edc9f116ffc1742a29.jpg", + "image_caption": [ + "(c) Without AAFM", + "Fig. 4: Confusion matrices of class predictions on the test set of CIFAR-10 generated by each algorithm (FixMatch+RECD with the described condition) trained on the training set of CIFAR-10-LT under $\\gamma_l = 100$ and $\\gamma_u = 1$ ." + ], + "image_footnote": [], + "bbox": [ + 504, + 143, + 633, + 242 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/a61fb43238383e5559756ccb71379879588fbe06a6b2f25a11640c80cf14345e.jpg", + "image_caption": [ + "(d) $\\mathrm{RECD + AAFM}$" + ], + "image_footnote": [], + "bbox": [ + 638, + 143, + 767, + 242 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "we argue that RECD also can be combined with various CISSL algorithms. To verify this, we conducted additional experiments using DARP, and demonstrated RECD-based DARP outperformed the original DARP in Appendix F.", + "bbox": [ + 212, + 338, + 784, + 383 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/f3c70fd891ef136643eb01f0e2478815210e2f0d919dce16685c4262a38ba34e.jpg", + "image_caption": [ + "(a) WithoutAAFM" + ], + "image_footnote": [], + "bbox": [ + 281, + 407, + 488, + 510 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/9a34d14dfc37409986f4662696f612674b20f6cb24f466f44396d2a0f19842f8.jpg", + "image_caption": [ + "(b) WithAAFM", + "Fig.5: Overall test loss and test losses for minority classes (Classes 8 and 9). The algorithms were trained on CIFAR-10-LT under $\\gamma_{l} = 100$ and $\\gamma_{u} = 1$ ." + ], + "image_footnote": [], + "bbox": [ + 511, + 407, + 722, + 511 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 613, + 359, + 630 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We proposed a novel CISSL algorithm, RECD, which addresses the class distribution mismatch between labeled and unlabeled sets. RECD estimates the class distribution of the unlabeled set through MC approximation and uses it for rebalancing the classifier. Additionally, RECD reuses the estimated class distribution for AAFM to mitigate class imbalance in the feature map by densifying feature clusters of minority classes. Experimental results demonstrate that RECD outperforms the baseline algorithms under both scenarios where the class distributions of the labeled and unlabeled sets match and mismatch. Detailed analyses on RECD verify that the effectiveness of each component of RECD for mitigating class imbalance. However, the estimated class distribution of unlabeled set through MC approximation might be inaccurate when unlabeled training samples are scarce. In future work, we aim to improve the robustness of class distribution estimation, even with a limited amount of unlabeled samples.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Park et al.", + "bbox": [ + 271, + 114, + 344, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This research was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (2023R1A2C2005453, RS-2023-00218913).", + "bbox": [ + 215, + 176, + 784, + 222 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 247, + 321, + 263 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Berthelot, D., Carlini, N., an Alex Kurakin, E.D.C., Sohn, K., Zhang, H., Raffel, C.: Remixmatch: Semi-supervised learning with distribution matching and augmentation anchoring. In: ICLR (2020)", + "2. Berthelot, D., Carlini, N., Goodfellow, I., Papernot, N., Oliver, A., Raffel, C.: Mixmatch: A holistic approach to semi-supervised learning. In: NeurIPS (2019)", + "3. Cao, K., Wei, C., Gaidon, A., Arechiga, N., Ma, T.: Learning imbalanced datasets with label-distribution-aware margin loss. In: Wallach, H., Larochelle, H., Beygelzimer, A., d'Alché-Buc, F., Fox, E., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 32. Curran Associates, Inc. (2019)", + "4. Chawla, N.V., Bowyer, K.W., Hall, L.O., Kegelmeyer, W.P.: Smote: Synthetic minority over-sampling technique. Journal of Artificial Intelligence Research pp. 321-357 (2002)", + "5. Chu, P., Bian, X., Liu, S., Ling, H.: Feature space augmentation for long-tailed data. In: ECCV. vol. 12374, pp. 694-710 (2020)", + "6. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.: Randaugment: Practical automated data augmentation with a reduced search space. In: NeurIPS (2020)", + "7. Cui, Y., Jia, M., Lin, T.Y., Song, Y., Belongie, S.: Large scale fine-grained categorization and domain-specific transfer learning. In: CVPR. pp. 4109-4118 (2018)", + "8. Cui, Y., Jia, M., Lin, T.Y., Song, Y., Belongie, S.: Class-balanced loss based on effective number of samples. In: CVPR. pp. 9268-9277 (2019)", + "9. DeVries, T., Taylor, G.W.: Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552 (2017)", + "10. Estabrooks, A., Jo, T., Japkowicz, N.: A multiple resampling method for learning from imbalanced data sets. Computational Intelligence 20(1), 18-36 (2004)", + "1. Fan, Y., Dai, D., Kukleva, A., Schiele, B.: Cossl: Co-learning of representation and classifier for imbalanced semi-supervised learning. In: CVPR. pp. 14574-14584 (2022)", + "2. Gidaris, S., Singh, P., Komodakis, N.: Unsupervised representation learning by predicting image rotations. In: ICLR (2018)", + "3. Grandvalet, Y., Bengio, Y.: Semi-supervised learning by entropy minimization. In: NeurIPS (2004)", + "4. Guo, L.Z., Li, Y.F.: Class-imbalanced semi-supervised learning with adaptive thresholding. In: ICML. vol. 162, pp. 8082-8094 (2022)", + "5. Huang, C., Li, Y., Loy, C.C., Tang, X.: Learning deep representation for imbalanced classification. In: CVPR. pp. 5375-5384 (2016)", + "6. JAPKOWICZ, N.: The class imbalance problem: Significance and strategies. In: Proc. 2000 International Conference on Artificial Intelligence. vol. 1, pp. 111-117 (2000)", + "7. Japkowicz, N.: The class imbalance problem: Significance and strategies. In: In Proceedings of the International Conference on Artificial Intelligence (2000)" + ], + "bbox": [ + 225, + 280, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "RECD", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Kang, B., Li, Y., Xie, S., Yuan, Z., Feng, J.: Exploring balanced feature spaces for representation learning. In: ICLR (2021)", + "19. Kang, B., Xie, S., Rohrbach, M., Yan, Z., Gordo, A., Feng, J., Kalantidis, Y.: Decoupling representation and classifier for long-tailed recognition. In: ICLR (2020)", + "20. Kim, J., Hur, Y., Park, S., Yang, E., Hwang, S., Shin, J.: Distribution aligning refinery of pseudo-label for imbalanced semi-supervised learning. In: NeurIPS (2020)", + "21. Krizhevsky, A.: Learning multiple layers of features from tiny images. Technical report, Department of Computer Science, University of Toronto (2009)", + "22. Kubat, M., Matwin, S., et al.: Addressing the curse of imbalanced training sets: one-sided selection. In: Icml. vol. 97, p. 179. Citeseer (1997)", + "23. Lai, Z., Wang, C., Gunawan, H., Cheung, S.C., Chuah, C.N.: Smoothed adaptive weighting for imbalanced semi-supervised learning: Improve reliability against unknown distribution data. In: ICML. vol. 162, pp. 11828-11843 (2022)", + "24. Lazarow, J., Sohn, K., Lee, C.Y., Li, C.L., Zhang, Z., Pfister, T.: Unifying distribution alignment as a loss for imbalanced semi-supervised learning. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) pp. 5644-5653 (2023)", + "25. Lee, D.H.: Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In: In Workshop on challenges in representation learning (ICML) (2013)", + "26. Lee, H., Kim, H.: Cdmad: Class-distribution-mismatch-aware debiasing for class-imbalanced semi-supervised learning. In: CVPR. pp. 23891-23900 (2024)", + "27. Lee, H., Shin, S., Kim, H.: Abc: Auxiliary balanced classifier for class-imbalanced semi-supervised learning. In: NeurIPS (2021)", + "28. Li, J., Meng, Z., Shi, D., Song, R., Diao, X., Wang, J., Xu, H.: Fcc: Feature clusters compression for long-tailed visual recognition. In: CVPR. pp. 24080-24089 (2023)", + "29. Li, K., Kong, X., Lu, Z., Wenyin, L., Yin, J.: Boosting weighted elm for imbalanced learning. Neurocomputing 128, 15-21 (2014)", + "30. Liu, X.Y., Wu, J., Zhou, Z.H.: Exploratory undersampling for class-imbalance learning. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics) 39(2), 539-550 (2008)", + "31. Liu, Z., Miao, Z., Zhan, X., Wang, J., Gong, B., Yu, S.X.: Large-scale long-tailed recognition in an open world. In: CVPR. pp. 2537-2546 (2019)", + "32. Menon, A.K., Jayasumana, S., Rawat, A.S., Jain, H., Veit, A., Kumar, S.: Long-tail learning via logit adjustment. In: ICLR (2021)", + "33. Miyato, T., ichi Maeda, S., Koyama, M., Ishii, S.: Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE Transactions on Pattern Analysis and Machine Intelligence 41(8), 1979-1993 (2018)", + "34. Oh, Y., Kim, D.J., Kweon, I.S.: Daso: Distribution-aware semantics-oriented pseudo-label for imbalanced semi-supervised learning. In: CVPR. pp. 9786-9796 (2022)", + "35. Ren, J., Yu, C., Sheng, S., Ma, X., Zhao, H., Yi, S., Li, H.: Balanced meta-softmax for long-tailed visual recognition. In: NeurIPS (2020)", + "36. Ren, M., Zeng, W., Yang, B., Urtasun, R.: Learning to reweight examples for robust deep learning. In: ICML. vol. 80, pp. 4334-4343 (2018)", + "37. Sajjadi, M., Javanmardi, M., Tasdizen, T.: Regularization with stochastic transformations and perturbations for deep semi-supervised learning. In: NeurIPS (2016)", + "38. Sohn, K., Berthelot, D., Li, C.L., Zhang, Z., Carlini, N., Cubuk, E.D., Kurakin, A., Zhang, H., Raffel, C.: Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In: NeurIPS (2020)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Park et al.", + "bbox": [ + 271, + 114, + 343, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "39. Verma, V., Kawaguchi, K., Lamb, A., Kannala, J., Solin, A., Bengio, Y., Lopez-Paz, D.: Interpolation consistency training for semi-supervised learning. Neural Networks 145, 90-106 (2022)", + "40. Wang, R., Jia, X., Wang, Q., Wu, Y., Meng, D.: Imbalanced semi-supervised learning with bias adaptive classifier. In: ICLR (2023)", + "41. Wang, S., Minku, L.L., Yao, X.: A learning framework for online class imbalance learning. IEEE Symposium on Computational Intelligence and Ensemble Learning (CIEL) pp. 36-45 (2013)", + "42. Wang, S., Minku, L.L., Yao, X.: Resampling-based ensemble methods for online class imbalance learning. IEEE Transactions on Knowledge and Data Engineering 27(5), 1356-1368 (2015)", + "43. Wang, X., Lian, L., Miao, Z., Liu, Z., Yu, S.X.: Long-tailed recognition by routing diverse distribution-aware experts. In: ICLR (2021)", + "44. Wang, X., Wu, Z., Lian, L., Yu, S.X.: Debiased learning from naturally imbalanced pseudo-labels. In: CVPR. pp. 14647-14657 (2022)", + "45. Wang, Y., Chen, H., Heng, Q., Hou, W., Fan, Y., Wu, Z., Wang, J., Savvides, M., Shinozaki, T., Raj, B., Schiele, B., Xie, X.: Freematch: Self-adaptive thresholding for semi-supervised learning. In: ICLR (2023)", + "46. Wei, C., Sohn, K., Mellina, C., Yuille, A., Yang, F.: Crest: A class-rebalancing self-training framework for imbalanced semi-supervised learning. In: CVPR. pp. 10857-10866 (2021)", + "47. Wei, T., Gan, K.: Towards realistic long-tailed semi-supervised learning: Consistency is all you need. In: CVPR. pp. 3469-3478 (2023)", + "48. Xiang, L., Ding, G., Han, J.: Learning from multiple experts: Self-paced knowledge distillation for long-tailed classification. In: ECCV. pp. 247-263 (2020)", + "49. Yu, F., Seff, A., Zhang, Y., Song, S., Funkhouser, T., Xiao, J.: Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365 (2015)", + "50. Yu, Z., Li, Y., Lee, Y.J.: Inpl: Pseudo-labeling the inliers first for imbalanced semi-supervised learning. In: ICLR (2023)", + "51. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. In: ICLR (2018)", + "52. Zhang, Y., Hooi, B., Hong, L., Feng, J.: Self-supervised aggregation of diverse experts for test-agnostic long-tailed recognition. In: NeurIPS (2022)", + "53. Zhong, Z., Cui, J., Liu, S., Jia, J.: Improving calibration for long-tailed recognition. In: CVPR. pp. 16489-16498 (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 646 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "RECD", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Rebalancing Using Estimated Class Distribution for Imbalanced Semi-Supervised Learning under Class Distribution Mismatch/c1f56d15-2542-45a0-80c7-7ebab6520f43_model.json b/2024/Rebalancing Using Estimated Class Distribution for Imbalanced Semi-Supervised Learning under Class Distribution Mismatch/c1f56d15-2542-45a0-80c7-7ebab6520f43_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7f3b0d80622b86b9a44ac11567698d985cb60bc3 --- /dev/null +++ b/2024/Rebalancing Using Estimated Class Distribution for Imbalanced Semi-Supervised Learning under Class Distribution Mismatch/c1f56d15-2542-45a0-80c7-7ebab6520f43_model.json @@ -0,0 +1,2588 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.226, + 0.141, + 0.78, + 0.207 + ], + "angle": 0, + "content": "Rebalancing Using Estimated Class Distribution for Imbalanced Semi-Supervised Learning under Class Distribution Mismatch" + }, + { + "type": "text", + "bbox": [ + 0.308, + 0.235, + 0.694, + 0.252 + ], + "angle": 0, + "content": "Taemin Park\\**, Hyuck Lee\\**, and Heeyoung Kim" + }, + { + "type": "text", + "bbox": [ + 0.303, + 0.263, + 0.7, + 0.278 + ], + "angle": 0, + "content": "Department of Industrial and Systems Engineering, KAIST" + }, + { + "type": "text", + "bbox": [ + 0.388, + 0.278, + 0.616, + 0.291 + ], + "angle": 0, + "content": "Daejeon 34141, Republic of Korea" + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.292, + 0.668, + 0.305 + ], + "angle": 0, + "content": "{ptm1001,dlgur0921,heeyoungkim}@kaist.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.359, + 0.744, + 0.568 + ], + "angle": 0, + "content": "Abstract. Despite significant advancements in class-imbalanced semi-supervised learning (CISSL), many existing algorithms explicitly or implicitly assume that the class distribution of unlabeled data matches that of labeled data. However, when this assumption fails in practice, the classification performance of such algorithms may degrade due to incorrectly assigned weight to each class during training. We propose a novel CISSL algorithm called Rebalancing Using Estimated Class Distribution (RECD). RECD estimates the unknown class distribution of unlabeled data through Monte Carlo approximation, leveraging predicted class probabilities for unlabeled samples, and subsequently rebalances the classifier based on the estimated class distribution. Additionally, we propose an extension of feature clusters compression in the context of CISSL to mitigate feature map imbalance by densifying minority class clusters. Experimental results on four benchmark datasets demonstrate that RECD achieves state-of-the-art classification performance in CISSL." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.58, + 0.758, + 0.609 + ], + "angle": 0, + "content": "Keywords: Class-imbalanced semi-supervised learning \\(\\cdot\\) Long-tailed learning \\(\\cdot\\) Auxiliary Balanced Classifier" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.635, + 0.377, + 0.65 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.667, + 0.788, + 0.758 + ], + "angle": 0, + "content": "In real-world datasets, the number of samples often varies across different classes. Deep neural network algorithms trained on such datasets tend to exhibit bias towards majority classes, which have a larger number of samples in the dataset [15, 31, 41]. To mitigate the bias, various class-imbalanced learning (CIL) techniques have been proposed, including re-sampling [4, 10, 30, 42], re-weighting [8, 29, 36], decoupled learning [19, 35, 53], and multi-expert learning [43, 48, 52]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.758, + 0.788, + 0.804 + ], + "angle": 0, + "content": "Although the CIL techniques significantly mitigate the bias, they typically assume that the entire training set is labeled. However, in reality, labeling samples requires substantial resources, resulting in many real-world datasets containing" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.812, + 0.375, + 0.826 + ], + "angle": 0, + "content": "* Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.826, + 0.393, + 0.84 + ], + "angle": 0, + "content": "† Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.344, + 0.127 + ], + "angle": 0, + "content": "Park et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "a significant proportion of unlabeled data. Because the CIL techniques assign importance to each data point based on its corresponding label, they cannot be directly applied when the training set includes unlabeled samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.785, + 0.357 + ], + "angle": 0, + "content": "Recently, several class-imbalanced semi-supervised learning (CISSL) algorithms, including ABC [27], UDAL [24], CReST [46], DARP [20], DASO [34], CoSSL [11], and SAW [23], have emerged to address the bias towards majority classes in scenarios where the training set consists of both labeled and unlabeled data. However, ABC, UDAL, and CReST explicitly assume that the class distribution of the unlabeled data matches that of the labeled data. Similarly, DARP, DASO, and SAW can be seen as implicitly assuming the class distribution match between the labeled and unlabeled data, as they incorporate CIL techniques such as LA [32] and cRT [19], which rely on the same class distribution between the labeled and unlabeled data. Additionally, CoSSL solely considers the class distribution of labeled data in its main component, TFE." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.787, + 0.525 + ], + "angle": 0, + "content": "However, the class distribution of the unlabeled set is often unknown and may differ from that of the labeled set [34, 40, 47]. In this case, the classification performance of the aforementioned algorithms may degrade because the unknown class distribution of the unlabeled set is not considered in the rebalancing process. Fig. 1b illustrates the class predictions on the test set of CIFAR-10 [21] using ReMixMatch+ABC, an existing CISSL algorithm, trained on the long-tailed version of the CIFAR-10 [8] training set where the class distributions of labeled and unlabeled sets mismatch, as presented in Fig. 1a. From Fig. 1b, we can observe that many test samples are mispredicted as Class 10 (the most minor class). This may be because ABC rebalances the classifier excessively, considering only the class distribution of the labeled set." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.547, + 0.391, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.238, + 0.641, + 0.371, + 0.653 + ], + "angle": 0, + "content": "(a) Class distributions" + }, + { + "type": "image", + "bbox": [ + 0.416, + 0.547, + 0.588, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.434, + 0.641, + 0.571, + 0.653 + ], + "angle": 0, + "content": "(b) ReMixMatch+ABC" + }, + { + "type": "image", + "bbox": [ + 0.614, + 0.548, + 0.785, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.641, + 0.768, + 0.654 + ], + "angle": 0, + "content": "(c) Proposed algorithm" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.664, + 0.788, + 0.708 + ], + "angle": 0, + "content": "Fig. 1: Comparison of the predictions of ReMixMatch+ABC and the proposed algorithm (ReMixMatch+RECD) trained on an imbalanced version of CIFAR-10 where the class distributions of labeled and unlabeled sets mismatch, as illustrated in Fig. 1a." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To rebalance the classifier appropriately when the class distribution of the unlabeled set mismatches that of the labeled set, we propose a novel CISSL algorithm called Rebalancing Using Estimated Class Distribution (RECD), which estimates the unknown class distribution using Monte Carlo (MC) approximation and rebalances the training loss based on the estimated class distribution. Specifically, RECD is a modular algorithm that operates in conjunction with previous CISSL algorithms. While RECD can generally be integrated with most" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.685, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RECD" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.192 + ], + "angle": 0, + "content": "CISSL algorithms by providing additional information about the class distribution of the unlabeled set, in this paper, we specifically combine RECD with ABC due to its simple and effective framework compared to other CISSL algorithms." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.787, + 0.373 + ], + "angle": 0, + "content": "To estimate the class distribution of unlabeled data, RECD first computes the average class predictions of ABC for the entire unlabeled set. Because generating class predictions for the entire unlabeled set can be computationally intensive, RECD generates class predictions on an unlabeled minibatch for each iteration and updates the exponential moving average (EMA) of the class predictions on the unlabeled set. Utilizing MC approximation, this average class prediction serves as the classifier's prior, representing the estimated unknown class distribution of the unlabeled set. Then, RECD rebalances the training loss on unlabeled samples for ABC based on the estimated class distribution of the unlabeled set, and trains ABC using this rebalanced unlabeled loss in addition to the training loss on labeled samples for ABC. As depicted in Fig. 1c, ReMix-Match+RECD generates nearly balanced class predictions on the test set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.787, + 0.54 + ], + "angle": 0, + "content": "In addition to mitigating imbalance in the classifier, RECD also mitigates imbalance in the feature map. While existing CISSL algorithms primarily focus on mitigating classifier bias, there also exists imbalance in the feature map [5,18]. Specifically, feature clusters of minority classes tend to be sparser compared to the dense clusters of majority classes. These sparse features often overlap with clusters of other classes in the feature map, posing challenges in classifying minority class features [28]. To address this issue by densifying the sparse feature clusters of minority classes, we extend feature clusters compression (FCC) [28], a previous CIL technique, into the context of CISSL by leveraging the unknown class distribution of the unlabeled data estimated by RECD. We refer to this extended version of FCC as the adaptively adjusted feature multiplier (AAFM)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.54, + 0.787, + 0.707 + ], + "angle": 0, + "content": "Our experimental results on CIFAR-10-LT [8], CIFAR-100-LT [8], STL-10-LT [20], and Small-ImageNet-127 [11] demonstrate the superior performance of the proposed algorithm compared to the baseline algorithms, regardless of whether the class distributions of labeled and unlabeled sets match or mismatch. Through detailed analyses, we verify that RECD effectively estimates the true class distribution of the unlabeled set and uses it to mitigate class imbalance. Additionally, we validate that AAFM not only reduces the test loss for minority classes but also stabilizes the training process. Finally, by conducting experiments that combine RECD with DARP [20], we validate the effectiveness of RECD beyond its application solely with ABC. The source codes for RECD are available at https://github.com/taemin-park/RECD." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.731, + 0.383, + 0.748 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Many CISSL algorithms have been proposed to address class imbalance in scenarios where the training set consists of both labeled and unlabeled data. For example, ABC [27] trains an auxiliary balanced classifier with a training loss that is rebalanced based on the class distribution of the labeled data. CReST [46] systematically enlarges the labeled dataset by assigning the labels to unlabeled" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.344, + 0.127 + ], + "angle": 0, + "content": "Park et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.236 + ], + "angle": 0, + "content": "samples predicted as minority classes more frequently than those predicted as majority classes. UDAL [24] combines progressive distribution alignment and logit-adjustment [32] to adjust the training loss. ABC, CReST, and UDAL have effectively mitigated bias towards majority classes. However, they assumed that the class distribution of the unlabeled data is the same as that of the labeled data, which may pose practical limitation as described in Sec. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.785, + 0.356 + ], + "angle": 0, + "content": "DARP [20] solves a convex optimization problem to refine biased pseudo labels. DASO [34] blends class predictions of a similarity-based classifier and a linear classifier to obtain unbiased pseudo-labels. SAW [23] estimates the learning difficulty of each class and smoothly reweights the training loss based on the estimated learning difficulty. However, DARP, DASO, and SAW implicitly assume the same class distributions of the unlabeled and labeled datasets, as they incorporate CIL techniques such as LA [32] and cRT [19], which rebalance the classifier based on the class distribution of the labeled data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.804, + 0.57 + ], + "angle": 0, + "content": "CoSSL [11] integrates Tail-class Feature Enhancement (TFE) and class-balanced sampling to rebalance the classifier. InPL [50] introduces energy-based pseudolabeling to effectively discern correctly predicted unlabeled samples of minority classes. DePL [44] mitigates pseudo-label bias using counterfactual reasoning and adaptive marginal loss. L2AC [40] mitigates training bias by using a bias adaptive classifier, which consists of a bias attractor and a linear classifier. Adsh [14] sets an adaptive confidence threshold for each class to utilize unlabeled samples predicted as minority classes more frequently than those predicted as majority classes. ACR [47] estimates the class distribution of the unlabeled set by comparing class predictions with class distribution prototypes, and then adaptively adjusts the training loss based on the estimated class distribution. CDMAD [26] refines the predictions for unlabeled and test samples based on the logits for solid color image. Although these algorithms have significantly improved CISSL, this area remains relatively less explored compared to SSL and CIL." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.591, + 0.383, + 0.607 + ], + "angle": 0, + "content": "3 Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.621, + 0.403, + 0.636 + ], + "angle": 0, + "content": "3.1 Problem settings" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We have a training dataset that consists of the labeled set \\(\\mathcal{X} = \\{(x_n, y_n) : n \\in (1, \\ldots, N)\\}\\) and unlabeled set \\(\\mathcal{U} = \\{(u_m) : m \\in (1, \\ldots, M)\\}\\), where \\(x_n \\in R^d\\) and \\(y_n \\in R^K\\) are the \\(n\\)th labeled sample and its corresponding label, respectively, and \\(u_m \\in R^d\\) is the \\(m\\)th unlabeled sample. For each class \\(k\\), \\(N_k\\) and \\(M_k\\) denote the numbers of the labeled and unlabeled samples, respectively, i.e., \\(\\sum N_k = N\\) and \\(\\sum M_k = M\\). In CISSL, it is typically assumed that \\(N_1 \\geq N_2 \\geq \\ldots \\geq N_K\\) and \\(M_1 \\geq M_2 \\geq \\ldots \\geq M_K\\). We denote the class imbalance ratios of the labeled and unlabeled sets as \\(\\gamma_l = \\frac{N_1}{N_K}\\) and \\(\\gamma_u = \\frac{M_1}{M_K}\\), respectively. When \\(M_k\\) is unknown, \\(\\gamma_u\\) would also be unknown, and it may differ from \\(\\gamma_l\\). For each training iteration, we generate minibatches \\(\\mathcal{MB}_{\\mathcal{X}} = \\{(x_b^m, y_b^m) : b \\in (1, \\ldots, B_l)\\} \\subset \\mathcal{X}\\) and \\(\\mathcal{MB}_{\\mathcal{U}} = \\{(u_b^m) : b \\in (1, \\ldots, B_u)\\} \\subset \\mathcal{U}\\) from \\(\\mathcal{X}\\) and \\(\\mathcal{U}\\), respectively, where \\(B_l\\) and \\(B_u\\) are the sizes of the labeled and unlabeled minibatches, respectively. Using \\(\\mathcal{MB}_{\\mathcal{X}}\\) and \\(\\mathcal{MB}_{\\mathcal{U}}\\), we aim to train an algorithm \\(f_\\theta: R^d \\to R^K\\) that accurately predicts the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.684, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RECD" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "class probabilities for unseen samples, where \\(\\theta\\) denotes the network parameters. We denote the extracted features before the classification layer as \\(\\xi (\\cdot)\\)" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.2, + 0.467, + 0.215 + ], + "angle": 0, + "content": "3.2 Backbone SSL algorithm" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.224, + 0.785, + 0.3 + ], + "angle": 0, + "content": "Following previous CISSL algorithms [11, 20, 23, 27, 34, 50], either FixMatch [38] or ReMixMatch [1] is used as a backbone SSL algorithm for RECD. To utilize unlabeled samples, FixMatch and ReMixMatch conduct consistency regularization using weak data augmentation techniques \\(\\alpha(\\cdot)\\) (flipping and cropping) and strong data augmentation techniques \\(\\mathcal{A}(\\cdot)\\) (Cutout [9] and RandomAugment [6])." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.3, + 0.804, + 0.421 + ], + "angle": 0, + "content": "Specifically, FixMatch first predicts the class probabilities for \\(\\alpha(u_b^m)\\) as \\(f(\\alpha(u_b^m))\\) and then generates the pseudo-label as \\(q_b^m = \\arg \\max_k (f(\\alpha(u_b^m))_k)\\) when the confidence exceeds the confidence threshold \\(\\epsilon\\), i.e., \\(\\max_k (f(\\alpha(u_b^m))_k) \\geq \\epsilon\\), where \\((\\cdot)_k\\) denotes the \\(k\\)th element of \\((\\cdot)\\). Then, FixMatch predicts the class probabilities for \\(\\mathcal{A}(u_b^m)\\) as \\(f(\\mathcal{A}(u_b^m))\\) and ensures \\(f(\\mathcal{A}(u_b^m))\\) to be close to the pseudo-label \\(q_b^m\\) by minimizing a loss that penalizes the distance between \\(f(\\mathcal{A}(u_b^m))\\) and \\(q_b^m\\). In addition, FixMatch minimizes the prediction loss calculated between weakly augmented labeled sample \\(\\alpha(x_b^m)\\) and \\(p_b^m\\), which is the one-hot version of \\(y_b^m\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.787, + 0.572 + ], + "angle": 0, + "content": "Similarly, ReMixMatch first predicts the class probabilities of \\(\\alpha(u_b^m)\\) as \\(q_b^m = f(\\alpha(u_b^m))\\). Then ReMixMatch applies a distribution alignment technique as \\(\\tilde{q}_b^m = \\mathrm{Normalize}(q_b^m \\times q(y) / \\tilde{q}(y))\\), where \\(\\mathrm{Normalize}(x)_i = x_i / \\sum_j x_j\\), \\(\\tilde{q}(y)\\) is the moving average of class predictions on unlabeled samples, and \\(q(y)\\) indicates the class distribution of the training set. Finally, ReMixMatch sharpens \\(\\tilde{q}_b^m\\) as \\(\\bar{q}_b^m = \\mathrm{Normalize}((\\tilde{q}_b^m)^{1/T})\\), where \\(T\\) is the temperature for sharpening. Using \\(\\bar{q}_b^m\\) as pseudo-labels, ReMixMatch conducts consistency regularization [33,37], entropy minimization [13,25] and mixup regularization [2,39,51]. ReMixMatch also minimizes the rotation loss [12] for self-supervised learning. The training loss of the backbone SSL algorithm, \\(L_{back}\\), is detailed in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.594, + 0.555, + 0.61 + ], + "angle": 0, + "content": "3.3 Auxiliary balanced classifier (ABC)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.62, + 0.784, + 0.74 + ], + "angle": 0, + "content": "To use a balanced classifier trained on a class-balanced subset of the training set, while benefiting from the high-quality representations learned from the entire training set, Lee and Kim [27] introduced an auxiliary balanced classifier (ABC) attached to the representation layer of a backbone SSL algorithm. To ensure that ABC is trained on a class-balanced subset of the training set, ABC generates a \\(0/1\\) mask \\(M(\\cdot)\\) for each training sample, where the \\(0/1\\) mask determines whether each sample contributes to the training loss. Specifically, ABC samples a \\(0/1\\) mask from a Bernoulli distribution for each labeled and unlabeled sample as" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.751, + 0.785, + 0.786 + ], + "angle": 0, + "content": "\\[\nM (x _ {b} ^ {m}) = \\mathcal {B} (\\frac {N _ {K}}{N _ {\\hat {y} _ {b} ^ {m}}}) \\quad \\mathrm {a n d} \\quad M (u _ {b} ^ {m}) = \\mathcal {B} (\\frac {N _ {K}}{N _ {\\hat {q} _ {b} ^ {m}}}). \\qquad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.785, + 0.841 + ], + "angle": 0, + "content": "\\(\\mathcal{B}(\\cdot)\\) denotes a Bernoulli distribution and \\(\\hat{q}_b^m = \\arg \\max (q_b^m)\\), where \\(q_{b}^{m}\\) is the class prediction of the ABC for an unlabeled sample \\(u_{b}^{m}\\). Note that in Eq. (1), the probability of sampling 1 is adjusted to be inversely proportional to the number" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.344, + 0.128 + ], + "angle": 0, + "content": "Park et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "of training samples belonging to each class. This sampling strategy ensures that ABC is trained with an equal frequency from each class's data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.268 + ], + "angle": 0, + "content": "With the \\(0/1\\) mask, ABC calculates the classification loss, \\(L_{cls}\\), between weakly augmented labeled sample \\(\\alpha(x_b^m)\\) and its corresponding one-hot version label \\(p_b^m\\). In addition, ABC also calculates a consistency regularization loss, \\(L_{con}\\), between weakly augmented unlabeled sample \\(\\alpha(u_b^m)\\) and strongly augmented unlabeled sample \\(\\mathcal{A}(u_b^m)\\), which is minimized to ensure that decision boundaries lie in low-density regions. The training losses \\(L_{cls}\\) and \\(L_{con}\\) are calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.343, + 0.281, + 0.786, + 0.32 + ], + "angle": 0, + "content": "\\[\nL _ {c l s} = \\frac {1}{B _ {l}} \\sum_ {b = 1} ^ {B _ {l}} M \\left(x _ {b} ^ {m}\\right) \\mathcal {H} \\left(p _ {b} ^ {m}, f _ {a b c} (\\xi \\left(\\alpha \\left(x _ {b} ^ {m}\\right)\\right))\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.265, + 0.335, + 0.785, + 0.375 + ], + "angle": 0, + "content": "\\[\nL _ {c o n} = \\frac {1}{B _ {u}} \\sum_ {b = 1} ^ {B _ {u}} \\sum_ {i = 1} ^ {2} M \\left(u _ {b} ^ {m}\\right) \\mathbb {1} \\left(\\max \\left(q _ {b} ^ {m}\\right) \\geq \\epsilon\\right) \\mathcal {H} \\left(q _ {b} ^ {m}, f _ {a b c} \\left(\\xi \\left(\\mathcal {A} _ {i} \\left(u _ {b} ^ {m}\\right)\\right)\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.381, + 0.784, + 0.442 + ], + "angle": 0, + "content": "where \\(\\mathcal{H}\\) is the cross-entropy loss, \\(q_{b}^{m} = f_{abc}(\\xi (\\alpha (u_{b}^{m})))\\), \\(f_{abc}\\) denotes the auxiliary classifier attached to the representation layer, and \\(\\mathbb{1}(\\cdot)\\) denotes an indicator function. With the \\(L_{cls}\\), \\(L_{con}\\) and the training loss of backbone SSL algorithm \\(L_{back}\\), total training loss of ABC is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.4, + 0.455, + 0.785, + 0.47 + ], + "angle": 0, + "content": "\\[\nL _ {A B C} = L _ {c l s} + L _ {c o n} + L _ {b a c k}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.481, + 0.788, + 0.572 + ], + "angle": 0, + "content": "ABC [27] effectively mitigated class imbalance when the class distributions of the labeled and unlabeled sets match. However, as we can observe in the second equation of the Eq. (1), because the \\(0/1\\) masks for unlabeled samples are generated based on the class distribution of the labeled set without considering that of the unlabeled set, \\(f_{abc}\\) may not be properly rebalanced when the unknown class distribution of the unlabeled set significantly mismatches that of the labeled set." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.594, + 0.594, + 0.611 + ], + "angle": 0, + "content": "3.4 Feature clusters compression (FCC) [28]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.62, + 0.784, + 0.68 + ], + "angle": 0, + "content": "To separate features extracted from samples of minority classes from those of majority classes by densifying the feature clusters, Li et al. [28] introduced FCC. For a training sample \\( x \\) of class \\( k \\), FCC multiplies the extracted feature \\( \\xi(x) \\) by a scaling factor \\( \\tau_k \\), and uses \\( \\tau_k \\times \\xi(x) \\) as input for the classification layer \\( \\psi \\), as" + }, + { + "type": "equation", + "bbox": [ + 0.427, + 0.694, + 0.785, + 0.709 + ], + "angle": 0, + "content": "\\[\nf _ {\\theta} (x) = \\psi \\left(\\tau_ {k} \\times \\xi (x)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.721, + 0.784, + 0.736 + ], + "angle": 0, + "content": "where \\(\\tau_{k}\\) decreases from majority to minority classes in a naive manner as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.747, + 0.785, + 0.776 + ], + "angle": 0, + "content": "\\[\n\\tau_ {k} = 1 + \\eta \\times \\left(1 - \\frac {k}{K}\\right), k = 1, \\dots , K, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.784, + 0.84 + ], + "angle": 0, + "content": "where \\(\\eta\\) is a scale hyperparameter. Since the features of majority classes are multiplied by relatively large scaling factors compared to those of minority classes, their feature clusters occupy larger regions on the feature map during training. In this context, to prevent minority class features from crossing the decision" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.685, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RECD" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "boundary, the feature extractor \\(\\xi(\\cdot)\\) would be trained to map minority class features into denser regions. Note that during the test phase, the features extracted from test samples are not multiplied by the scaling factor, ensuring that the classifier predicts classes for test samples with densified features. According to Li et al. [28], FCC effectively enhances the separability of feature clusters of minority classes in CIL. However, in CISSL, it may fail to appropriately densify minority class features because the class distribution of the unlabeled set cannot be considered in the scaling factor \\(\\tau_{k}\\). In this paper, we refer to the scaling factor in Eq. (6) as the naive feature multiplier." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.306, + 0.381, + 0.323 + ], + "angle": 0, + "content": "4 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.338, + 0.719, + 0.354 + ], + "angle": 0, + "content": "4.1 Rebalancing using estimated class distribution (RECD)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.363, + 0.788, + 0.468 + ], + "angle": 0, + "content": "As described in Sec. 1, RECD incorporates the unknown class distribution of the unlabeled data by rebalancing the classifier based on the estimated class distribution of the unlabeled data. RECD is integrated with ABC [27], which rebalances the training loss using the \\(0/1\\) mask, as described in Eq. (1). Whereas ABC generates \\(0/1\\) masks for unlabeled samples based on the class distribution of the labeled set, RECD replaces the parameter of \\(0/1\\) masks for unlabeled samples using the estimated class distribution of the unlabeled set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.469, + 0.788, + 0.536 + ], + "angle": 0, + "content": "Specifically, to estimate the unknown class distribution of the unlabeled set, \\( p(y_{u}) \\), RECD first computes the average class predictions of ABC on the entire unlabeled set as \\( \\frac{1}{M}\\sum_{m = 1}^{M}p_{\\theta}(y_u|u_m) \\), where \\( p_{\\theta}(y_u|u_m) = f_{abc}(\\xi (\\alpha (u_m))) \\). Then, using MC approximation, \\( \\frac{1}{M}\\sum_{m = 1}^{M}p_{\\theta}(y_u|u_m) \\) becomes \\( p_{\\theta}(y_u) \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.547, + 0.785, + 0.577 + ], + "angle": 0, + "content": "\\[\n\\frac {1}{M} \\sum p _ {\\theta} \\left(y _ {u} \\mid u _ {m}\\right) \\approx \\int p _ {\\theta} \\left(y _ {u} \\mid u\\right) p (u) d u = p _ {\\theta} \\left(y _ {u}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.788, + 0.648 + ], + "angle": 0, + "content": "In Eq. (7), \\( p_{\\theta}(y_u) \\) can be regarded as the classifier's prior for the unlabeled samples, and we use this classifier's prior as the estimation of the class distribution of the unlabeled set \\( p(y_u) \\) based on that the classifier's prior is highly correlated with the class distribution, as analyzed in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.648, + 0.788, + 0.679 + ], + "angle": 0, + "content": "Using the estimated class distribution \\( p_{\\theta}(y_u) \\), RECD modifies the sampling distribution of the \\( 0/1 \\) mask for unlabeled samples in Eq. (1) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.691, + 0.785, + 0.723 + ], + "angle": 0, + "content": "\\[\nM \\left(u _ {b} ^ {m}\\right) = \\mathcal {B} \\left(\\frac {\\operatorname* {m i n} \\left(p _ {\\theta} \\left(y _ {u}\\right)\\right)}{p _ {\\theta} \\left(y _ {u}\\right) \\hat {q} _ {b} ^ {m}}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.794 + ], + "angle": 0, + "content": "By setting the parameter of the \\(0/1\\) mask based on \\(p_{\\theta}(y_u)\\), RECD rebalances the training loss of the unlabeled samples to an appropriate extent even when the class distributions of the labeled and unlabeled sets severely mismatch. The total training loss of RECD is described in detail in Sec. 4.3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "RECD may increase computational complexity compared to the original ABC because, as shown in Eq. (7), computing class predictions of ABC for the entire unlabeled set for each iteration or epoch would require a substantial amount of" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.344, + 0.128 + ], + "angle": 0, + "content": "Park et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "computation, especially for large-scale datasets such as LSUN [49] or iNaturallist [7]. To address this issue, RECD replaces the naive average \\(\\frac{1}{M}\\sum p_{\\theta}(y_u|u_m)\\) with an EMA of the class predictions on the unlabeled samples, which is updated for each iteration of the training as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.325, + 0.222, + 0.786, + 0.239 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(y _ {u}\\right) _ {n e w} = \\rho \\times p _ {\\theta} \\left(y _ {u}\\right) _ {o l d} + (1 - \\rho) \\times p _ {\\theta} \\left(y _ {u}\\right) _ {b a t c h}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.25, + 0.788, + 0.342 + ], + "angle": 0, + "content": "where \\(\\rho\\) is a hyperparameter for the EMA update, \\(p_{\\theta}(y_u)_{old}\\) is \\(p_{\\theta}(y_u)\\) before the update, and \\(p_{\\theta}(y_u)_{batch}\\) is the average class prediction on the weakly augmented unlabeled minibatch, \\(\\frac{1}{B_u}\\sum_{b = 1}^{B_u}q_b^m = \\frac{1}{B_u}\\sum_{b = 1}^{B_u}f_{abc}(\\xi (\\alpha (u_b^m)))\\). Updating the EMA of the class predictions on unlabeled samples adds negligible computational cost compared to the original ABC because RECD reuses \\(q_b^m\\) that is calculated to conduct consistency regularization for ABC, as in Eq. (3)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.366, + 0.657, + 0.383 + ], + "angle": 0, + "content": "4.2 Adaptively adjusted feature multiplier (AAFM)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.789, + 0.562 + ], + "angle": 0, + "content": "Although the naive feature multiplier effectively densifies feature clusters of the minority classes in CIL, its efficacy may diminish in CISSL due to its lack of consideration for the class distribution of the unlabeled set. In addition, the naive feature multiplier linearly reduces from majority to minority classes as shown in Eq. (6), without considering the degree of class imbalance in the training set. These limitations may lead to inappropriate scaling factors for the extracted features, resulting in insufficient densification or unstable training, as discussed in Sec. 5.3. To address this issue by incorporating both the class imbalance ratio of the labeled set and the unknown class distribution of the unlabeled set, we introduce an advanced version of the naive feature multiplier, AAFM. For the labeled samples, AAFM for the \\(k\\)th class, \\(\\tau_k^l\\), is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.36, + 0.573, + 0.787, + 0.604 + ], + "angle": 0, + "content": "\\[\n\\tau_ {k} ^ {l} = 1 + \\eta \\times \\left(\\frac {N _ {k}}{N _ {1}} \\times C _ {l k} ^ {t}\\right), k = 1, \\dots , K, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.788, + 0.66 + ], + "angle": 0, + "content": "where \\(\\eta\\) is a scale hyperparameter similar to \\(\\eta\\) in Eq. (6), and \\(t\\) is a hyperparameter for \\(C_{lk}\\), which represents the EMA of confidence for labeled samples of the \\(k\\)th class, which is updated for each iteration of the training process as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.378, + 0.672, + 0.787, + 0.691 + ], + "angle": 0, + "content": "\\[\nC _ {l k} = \\omega \\times C _ {l k} ^ {\\text {o l d}} + (1 - \\omega) \\times C _ {l k} ^ {\\text {b a t c h}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.702, + 0.788, + 0.809 + ], + "angle": 0, + "content": "where \\(\\omega\\) is a hyperparameter for the EMA update, \\(C_{lk}^{old}\\) is \\(C_{lk}\\) before the update, and \\(C_{lk}^{batch}\\) is the average confidence for class \\(k\\) on a minibatch, which is calculated as \\(\\frac{1}{n_k}\\sum_{b = 1}^{B_l}f_{abc}(\\xi (\\alpha (x_b^m)))_k\\times \\mathbb{1}(y_b^m = k)\\). Here, \\(n_k\\) denotes the number of samples for the \\(k\\)th class on a minibatch. Note that \\(C_{lk}\\) is updated only when \\(n_k\\geq 1\\). By multiplying \\(C_{lk}^{t}\\) to the second term on the right-hand side of Eq. (10), we stabilize the training process, assigning smaller AAFMs to classes whose feature clusters are located near the decision boundary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "We also multiply AAFM \\(\\tau_{k}^{u}\\) with the features extracted from strongly augmented unlabeled samples, where \\(\\tau_{k}^{u}\\) incorporates the unknown class distribution" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.685, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RECD" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.162 + ], + "angle": 0, + "content": "of the unlabeled set by reusing the estimated class distribution \\( p_{\\theta}(y_u) \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.323, + 0.179, + 0.785, + 0.21 + ], + "angle": 0, + "content": "\\[\n\\tau_ {k} ^ {u} = 1 + \\eta \\times \\left(\\frac {p _ {\\theta} \\left(y _ {u}\\right) _ {k}}{\\operatorname* {m a x} \\left(p _ {\\theta} \\left(y _ {u}\\right)\\right)} \\times C _ {u k} ^ {t}\\right), k = 1, \\dots , K. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.219, + 0.784, + 0.25 + ], + "angle": 0, + "content": "In Eq. (12), \\( C_{uk} \\) is an EMA of confidence for unlabeled samples predicted as the \\( k \\)th class, which is updated for each iteration of the training process as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.376, + 0.263, + 0.785, + 0.281 + ], + "angle": 0, + "content": "\\[\nC _ {u k} = \\omega \\times C _ {u k} ^ {\\text {o l d}} + (1 - \\omega) \\times C _ {u k} ^ {\\text {b a t c h}}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.294, + 0.785, + 0.373 + ], + "angle": 0, + "content": "where \\( C_{uk}^{batch} = \\frac{1}{m_k}\\sum_{b = 1}^{B_u}f_{abc}(\\xi (\\alpha (u_b^m)))_k\\times \\mathbb{1}(\\hat{q}_b^m = k) \\), with \\( m_{k} = \\sum_{b = 1}^{B_{u}}\\mathbb{1}(\\hat{q}_{b}^{m} = k) \\) and \\( \\hat{q}_b^m = \\arg \\max_k(q_b^m)_k \\). \\( C_{uk} \\) is updated only when \\( m_{k}\\geq 1 \\). Since \\( \\tau_k^u \\) incorporates the unknown class distribution of the unlabeled set, AAFM can densify feature clusters of minority classes to an appropriate extent even when the class distributions of the labeled and unlabeled sets significantly mismatch." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.403, + 0.507, + 0.418 + ], + "angle": 0, + "content": "4.3 End-to-end training of RECD" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.435, + 0.784, + 0.48 + ], + "angle": 0, + "content": "Fig. 2 shows the end-to-end training process of the proposed algorithm for labeled and unlabeled data, respectively. The total training loss of RECD, denoted as \\( L_{RECD} \\), closely resembles that of ABC in Eq. (4), and is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.496, + 0.785, + 0.511 + ], + "angle": 0, + "content": "\\[\nL _ {R E C D} = L _ {c l s} + L _ {c o n} + L _ {b a c k}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.525, + 0.77, + 0.541 + ], + "angle": 0, + "content": "where \\(L_{cls}\\) and \\(L_{con}\\) are improved by applying RECD and AAFM as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.321, + 0.556, + 0.785, + 0.595 + ], + "angle": 0, + "content": "\\[\nL _ {c l s} = \\frac {1}{B _ {l}} \\sum_ {b = 1} ^ {B _ {l}} M \\left(x _ {b} ^ {m}\\right) \\mathcal {H} \\left(p _ {b} ^ {m}, f _ {a b c} \\left(\\xi \\left(\\alpha \\left(x _ {b} ^ {m}\\right)\\right) \\times \\tau_ {y _ {b} ^ {m}} ^ {l}\\right)\\right), \\tag {15}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.301, + 0.617, + 0.785, + 0.657 + ], + "angle": 0, + "content": "\\[\nL _ {c o n} = \\frac {1}{B _ {u}} \\sum_ {b = 1} ^ {B _ {u}} \\sum_ {i = 1} ^ {2} M \\left(u _ {b} ^ {m}\\right) \\mathcal {H} \\left(q _ {b} ^ {m}, f _ {a b c} \\left(\\xi \\left(\\mathcal {A} _ {i} \\left(u _ {b} ^ {m}\\right)\\right) \\times \\tau_ {\\hat {q} _ {b} ^ {m}} ^ {u}\\right)\\right), \\tag {16}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.308, + 0.674, + 0.785, + 0.708 + ], + "angle": 0, + "content": "\\[\nM \\left(x _ {b} ^ {m}\\right) = \\mathcal {B} \\left(\\frac {N _ {K}}{N _ {y _ {b} ^ {m}}}\\right) \\quad \\text {a n d} \\quad M \\left(u _ {b} ^ {m}\\right) = \\mathcal {B} \\left(\\frac {\\min \\left(p _ {\\theta} \\left(y _ {u}\\right)\\right)}{p _ {\\theta} \\left(y _ {u}\\right) \\hat {q} _ {b} ^ {m}}\\right), \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.718, + 0.787, + 0.84 + ], + "angle": 0, + "content": "where \\( p_b^m \\) is the one-hot version of \\( y_b^m \\), \\( q_b^m = f_{abc}(\\xi(\\alpha(u_b^m))) \\), and \\( \\tau_{y_b^m}^l \\) and \\( \\tau_{\\hat{q}_b^m}^u \\) are AAFMs corresponding to \\( y_b^m \\) and \\( \\hat{q}_b^m \\), respectively. To improve the quality of features learned by the backbone SSL algorithm, \\( L_{back} \\) is also adjusted with two modifications. First, AAFM is applied to further densify feature clusters of minority classes during the learning of the backbone SSL algorithm. Second, to train the backbone algorithm with unbiased pseudo-labels, pseudo-labels of the backbone algorithm are generated by \\( f_{abc} \\) instead of the classifier of the backbone algorithm. The pseudo-code of RECD is presented in Appendix D." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.345, + 0.127 + ], + "angle": 0, + "content": "Park et al." + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.143, + 0.52, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.303, + 0.259, + 0.493, + 0.269 + ], + "angle": 0, + "content": "(a) Balanced training with labeled data" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.143, + 0.808, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.259, + 0.735, + 0.269 + ], + "angle": 0, + "content": "(b) Balanced training with unlabeled data" + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.28, + 0.682, + 0.294 + ], + "angle": 0, + "content": "Fig. 2: End-to-end training of the proposed algorithm" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.323, + 0.375, + 0.34 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.356, + 0.424, + 0.372 + ], + "angle": 0, + "content": "5.1 Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.789, + 0.565 + ], + "angle": 0, + "content": "Following previous CISSL studies [11, 20, 23, 34, 40], we conducted experiments on CIFAR-10-LT [8], CIFAR-100-LT [8], STL10-LT [20], and Small-ImageNet-127 [11]. The detailed descriptions for the datasets are provided in Appendix E.1. We evaluated the classification performance using balanced accuracy (bACC) (referred to as the averaged class recall in [11]) [15] and geometric mean (GM) [22]. The performance of the proposed algorithm was compared with the following baseline algorithms: vanilla algorithm, CIL algorithms [3, 16, 19], SSL algorithms [1, 38], and CISSL algorithms [11, 14, 20, 23, 24, 27, 40, 44, 46]. Algorithms such as DASO [34] and ACR [47], whose classification performances were measured under different experimental settings from ours, were compared with RECD in Appendix G. The detailed descriptions of the baseline algorithms and training setups are provided in Appendix E.2." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.587, + 0.434, + 0.603 + ], + "angle": 0, + "content": "5.2 Experimental results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.764 + ], + "angle": 0, + "content": "Tab. 1 summarizes the classification performances of the baseline algorithms and RECD on CIFAR-10-LT under \\(\\gamma_l = \\gamma_u\\). First, the CISSL algorithms achieved higher performances than the vanilla algorithm, CIL algorithms, and SSL algorithms. These results verify the importance of mitigating class imbalance and utilizing unlabeled samples. Furthermore, RECD achieved the highest performance among the CISSL algorithms. Specifically, RECD showed a tendency to outperform other CISSL algorithms by a large margin as the class imbalance ratio increased, indicating its effectiveness in mitigating class imbalance. These results may be attributed to the unique ability of the proposed algorithm to mitigate class imbalance in the feature map, as we analyze in Sec. 5.3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Tab. 2 summarizes the performances of the baseline CISSL algorithms and RECD on CIFAR-100-LT. RECD outperformed the baseline algorithms. Given that CIFAR-100-LT has 100 classes, these results suggest that RECD may also be suitable for datasets with a large number of classes. Considering that classes with the fewest labeled data points have only three labeled samples under \\(\\gamma = 50\\)," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.685, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RECD" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.258, + 0.145, + 0.746, + 0.16 + ], + "angle": 0, + "content": "Table 1: Comparison of bACC/GM on CIFAR-10-LT under \\( \\gamma = {\\gamma }_{l} = {\\gamma }_{u} \\)" + }, + { + "type": "table", + "bbox": [ + 0.279, + 0.172, + 0.73, + 0.398 + ], + "angle": 0, + "content": "
CIFAR-10-LT (γ = γt = γu)
Algorithmγ = 50γ = 100γ = 150
Vanilla (Cross-Entropy Loss)65.2±0.05/61.1±0.0958.8±0.13/58.2±0.1155.6±0.43/
Re-sampling [17]64.3±0.48/60.6±0.6755.8±0.47/45.1±0.3052.2±0.05/
LDAM-DRV [3]68.9±0.07/67.0±0.0862.8±0.17/58.9±0.6057.9±0.20/
cRT [19]67.8±0.13/66.3±0.1563.2±0.45/59.9±0.4059.3±0.10/
FixMatch [38]79.2±0.33/77.8±0.3671.5±0.72/66.8±1.5168.4±0.15/
FixMatch+DARP+cRT [19, 20]85.8±0.43/85.6±0.5682.4±0.26/81.8±0.1779.6±0.42/
FixMatch+CRST+LA [32, 46]85.6±0.36/81.9±0.4581.2±0.70/74.5±0.9971.9±2.24/
FixMatch+ABC [27]85.6±0.26/85.2±0.2981.1±1.14/80.3±1.2977.3±1.25/
FixMatch+CoSSL [11]86.8±0.30/86.6±0.2583.2±0.49/82.7±0.6080.3±0.55/
FixMatch+SAW+LA [23, 32]86.2±0.15/83.9±0.3580.7±0.15/77.5±0.2173.7±0.06/
FixMatch+Adsh [14]83.4±0.06/-76.5±0.35/-71.5±0.30/
FixMatch+DebiasPL [44]- / --80.6±0.50/-- / -
FixMatch+UDAL [24]86.5±0.29/-81.4±0.39/-77.9±0.33/
FixMatch+L2AC [40]- / --82.1±0.57/81.5±0.6477.6±0.53/
FixMatch+RECD87.3±0.18/87.2±0.1984.0±0.13/83.6±0.1680.6±0.53/
ReMixMatch [1]81.5±0.26/80.2±0.3273.8±0.38/69.5±0.8469.9±0.47/
ReMixMatch+DARP+cRT [19, 20]87.3±0.61/87.0±0.1183.5±0.07/83.1±0.0979.7±0.54/
ReMixMatch+CRST+LA [32, 46]84.2±0.11/-81.3±0.34/-79.2±0.31/
ReMixMatch+ABC [27]87.9±0.47/87.6±0.5184.5±0.32/84.1±0.3680.5±1.18/
ReMixMatch+CoSSL [11]87.7±0.21/87.6±0.2584.1±0.56/83.7±0.6681.3±0.83/
ReMixMatch+SAW+cRT [19, 23]87.6±0.21/87.4±0.2685.4±0.32/83.9±0.2179.9±0.15/
ReMixMatch+RECD88.1±0.13/87.9±0.1285.4±0.40/85.2±0.4182.5±0.33/
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.785, + 0.461 + ], + "angle": 0, + "content": "these results also demonstrate the possibility that RECD can outperform the baseline algorithms even when labeled samples are extremely scarce." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.785, + 0.568 + ], + "angle": 0, + "content": "Tab. 3 summarizes the performances of the baseline CISSL algorithms and RECD on Small-ImageNet-127. For both \\(32 \\times 32\\) and \\(64 \\times 64\\) version, RECD significantly outperformed the baseline algorithms. Given that the class distribution of the test set of Small-ImageNet-127 is also imbalanced like the training set, these results show that RECD may also be suitable for datasets with imbalanced test sets. Moreover, given that Small-ImageNet-127 is an large-scale dataset, these results verify that RECD can also be applied to large-scale datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.787, + 0.601 + ], + "angle": 0, + "content": "Tab. 4 summarizes the performances of the baseline algorithms and RECD evaluated on CIFAR-10-LT under \\(\\gamma_l \\neq \\gamma_u\\) and STL-10-LT under unknown" + }, + { + "type": "table_caption", + "bbox": [ + 0.27, + 0.629, + 0.733, + 0.644 + ], + "angle": 0, + "content": "Table 2: Comparison of bACC on CIFAR-100-LT under \\( \\gamma = {\\gamma }_{l} = {\\gamma }_{u} \\)" + }, + { + "type": "table", + "bbox": [ + 0.365, + 0.656, + 0.635, + 0.838 + ], + "angle": 0, + "content": "
CIFAR-100-LT (γ = γt = γu)
Algorithmγ = 20γ = 50
FixMatch [38]49.6±0.7842.1±0.33
FixMatch+DARP [20]50.8±0.7743.1±0.54
FixMatch+DARP+cRT [19, 20]51.4±0.6844.9±0.54
FixMatch+CReST [46]51.8±0.1244.9±0.50
FixMatch+CReST+LA [32, 46]52.9±0.0747.3±0.17
FixMatch+ABC [27]53.3±0.7946.7±0.26
FixMatch+CoSSL [11]53.9±0.7847.6±0.57
FixMatch+RECD54.6±0.3647.8±0.17
ReMixMatch [1]51.6±0.4344.2±0.59
ReMixMatch+DARP [20]51.9±0.3544.7±0.66
ReMixMatch+DARP+cRT [19, 20]54.5±0.4248.5±0.91
ReMixMatch+CReST [46]51.3±0.3445.5±0.76
ReMixMatch+CReST+LA [32, 46]51.9±0.6046.6±1.14
ReMixMatch+ABC [27]55.6±0.3547.9±0.10
ReMixMatch+CoSSL [11]55.8±0.6248.9±0.61
ReMixMatch+RECD55.9±0.3649.5±0.21
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.344, + 0.127 + ], + "angle": 0, + "content": "Park et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.315, + 0.145, + 0.688, + 0.16 + ], + "angle": 0, + "content": "Table 3: Comparison of bACC on Small-ImageNet-127" + }, + { + "type": "table", + "bbox": [ + 0.379, + 0.171, + 0.62, + 0.264 + ], + "angle": 0, + "content": "
Small-ImageNet-127 (γ = γl = γu)
Algorithm32 × 3264 × 64
FixMatch [38]29.742.3
FixMatch+DARP [20]30.542.5
FixMatch+DARP +cRT [19,20]39.751.0
FixMatch+CReST+ [46]32.544.7
FixMatch+CReST++LA [32,46]40.955.9
FixMatch+CoSSL [11]43.753.8
FixMatch+RECD47.359.5
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.304, + 0.788, + 0.486 + ], + "angle": 0, + "content": "\\(\\gamma_{u}\\). RECD outperformed the baseline algorithms by large margins. These results demonstrate that RECD effectively mitigates class imbalance even when \\(\\gamma_{l} \\neq \\gamma_{u}\\), by considering the unknown class distribution of the unlabeled set. Notably, ReMixMatch+RECD outperformed the baseline algorithms that use ReMixMatch* [20] (ReMixMatch that uses the estimated class distribution of the unlabeled set) as the backbone SSL algorithm. These results implicitly demonstrate that RECD estimates class distribution better than the previous estimation algorithm. Additionally, under severe class distribution mismatch between the labeled and unlabeled sets, LA degraded the performance of ReMixMatch*+DARP and ReMixMatch*+SAW. This may be because the unknown class distribution of the unlabeled set cannot be incorporated into LA, emphasizing the importance of considering the class distribution of the unlabeled set." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.524, + 0.786, + 0.54 + ], + "angle": 0, + "content": "Table 4: Comparison of bACC/GM on CIFAR-10-LT and STL-10-LT under \\( \\gamma_l \\neq \\gamma_u \\)." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.551, + 0.787, + 0.74 + ], + "angle": 0, + "content": "
CIFAR-10-LT (γl=100)STL-10-LT (γu=Unknown)
Algorithmγu=1γu=50γu=150γl=10γl=20
FixMatch [38]68.9±1.95/42.8±8.1173.9±0.25/70.5±0.5269.6±0.60/62.6±1.1172.9±0.09/69.6±0.0163.4±0.21/52.6±0.09
+DARP [20]85.4±0.5585.0±0.6577.3±0.17/75.5±0.2172.9±0.24/69.5±0.1877.8±0.33/76.5±0.4069.9±1.77/65.4±0.97
+DARP+LA [20,32]86.6±1.11/86.2±1.1582.3±0.32/81.5±0.2978.9±0.23/77.7±0.0678.6±0.30/77.4±0.4071.9±0.49/68.7±0.51
+DARP+cRT [19,20]87.0±0.70/86.8±0.6782.7±0.21/82.3±0.2580.7±0.44/80.2±0.6179.3±0.23/78.7±0.2174.1±0.61/73.1±1.21
+ABC [27]82.7±0.49/81.9±0.6882.7±0.64/82.0±0.7678.4±0.87/77.2±1.0779.1±0.46/78.1±0.5773.8±0.15/72.1±0.15
+SAW [23]81.2±0.68/80.2±0.9179.8±0.25/79.1±0.3274.5±0.97/72.5±1.3778.3±0.25/77.0±0.1971.9±0.81/69.0±0.81
+SAW+LA [23,32]84.5±0.68/84.1±0.7882.9±0.38/82.6±0.3879.1±0.81/78.6±0.91-/--/-
+SAW+cRT [19,23]84.6±0.23/84.4±0.2681.6±0.38/81.3±0.3277.6±0.40/77.1±0.41-/--/-
+RECD90.2±0.57/90.0±0.6485.6±0.15/85.3±0.1582.3±0.39/81.8±0.4581.4±0.28/80.6±0.3879.0±0.48/78.1±0.54
ReMixMatch [1]48.3±0.14/19.5±0.8575.1±0.43/71.9±0.7772.5±0.10/68.2±0.3267.8±0.45/61.1±0.9260.1±1.18/44.9±1.52
+ABC [27]76.4±5.34/74.8±6.0585.2±0.20/84.7±0.2580.4±0.40/80.0±0.4476.8±0.52/74.8±0.6471.2±1.37/67.4±1.89
ReMixMatch* [1]85.0±1.35/84.3±1.5577.0±0.12/74.7±0.0472.8±0.10/68.8±0.2176.7±0.15/73.9±0.3267.7±0.46/60.3±0.76
+DARP [20]86.9±0.10/86.4±0.1577.4±0.22/75.0±0.2573.2±0.11/69.2±0.3179.4±0.07/78.2±0.1070.9±0.44/67.0±1.62
+DARP+LA [20,32]81.8±0.45/80.9±0.4083.9±0.42/83.4±0.4581.1±0.20/80.3±0.2680.6±0.45/79.6±0.5576.8±0.60/74.8±0.68
+DARP+cRT [19,20]88.7±0.25/88.5±0.2583.5±0.53/83.1±0.5180.9±0.25/80.3±0.3180.9±0.53/80.0±0.4676.7±0.50/74.9±0.70
+SAW [23]87.0±0.75/86.4±0.8580.6±1.57/79.2±1.1977.6±0.76/76.0±0.9382.0±0.55/81.0±0.6479.2±0.44/77.9±0.52
+SAW+LA [23,32]74.2±1.49/65.1±2.3684.8±1.07/82.4±2.3281.3±2.42/80.9±2.47-/--/-
+SAW+cRT [19,23]88.8±0.79/88.6±0.8384.5±0.78/83.6±1.2782.4±0.10/82.0±0.10-/--/-
ReMixMatch+RECD90.3±0.40/90.2±0.4186.8±0.17/86.6±0.1883.9±0.11/83.7±0.1284.9±0.41/84.4±0.4782.5±0.27/81.7±0.31
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.842 + ], + "angle": 0, + "content": "In addition to the above results, we demonstrate the effectiveness of RECD with fewer labeled samples in Appendix J. Furthermore, we verify that RECD can be effectively combined with recent SSL algorithms by conducting experiments using FreeMatch [45] as the backbone SSL algorithm in Appendix I." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.684, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RECD" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.631, + 0.163 + ], + "angle": 0, + "content": "5.3 Detailed analyses on the proposed algorithm" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.788, + 0.311 + ], + "angle": 0, + "content": "We argued that RECD effectively estimates the unknown class distribution of the unlabeled set and mitigates class imbalance using this estimation. To verify this argument, we plotted the true class distribution of the unlabeled set, the class distribution of the unlabeled set estimated by RECD, and the class distribution of the labeled set in Fig. 3. We can observe that RECD accurately estimated the unknown class distribution of the unlabeled set, even when it significantly differed from that of the labeled set. This accurate class distribution estimation enabled the proposed algorithm to be trained with a properly rebalanced training loss and accurately calculated AAFM, mitigating class imbalance as in Fig. 4." + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.337, + 0.49, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.323, + 0.441, + 0.451, + 0.453 + ], + "angle": 0, + "content": "(a) FixMatch+RECD" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.337, + 0.724, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.441, + 0.692, + 0.453 + ], + "angle": 0, + "content": "(b) RemixMatch+RECD" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.465, + 0.788, + 0.507 + ], + "angle": 0, + "content": "Fig. 3: True class distributions of the labeled and unlabeled sets, along with the class distribution of the unlabeled set estimated by RECD. FixMatch+RECD and Remix-Match+RECD were trained on CIFAR-10-LT with \\(\\gamma_{l} = 100\\) and \\(\\gamma_{u} = 1\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Fig. 4 presents the confusion matrices of class predictions on the test set of CIFAR-10 generated by each algorithm trained on the training set of CIFAR-10-LT under \\(\\gamma_l = 100\\) and \\(\\gamma_u = 1\\). The \\((i,j)\\) position of each confusion matrix represents the proportion of the \\(i\\)th class samples classified as the \\(j\\)th class. In the figures, the left three algorithms often misclassified minority classes as majority classes. On the other hand, with both RECD and AAFM, the algorithm classified most samples of minority classes accurately. These results demonstrate that both RECD and AAFM effectively mitigate class imbalance when the class distribution of the labeled and unlabeled sets significantly mismatch." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Additionally, we argued that AAFM not only reduces the test losses for minority classes but also stabilizes the training process. To verify this, we measured the overall test loss and the test losses of minority classes for FixMatch+RECD without AAFM and FixMatch+RECD with AAFM trained on CIFAR-10-LT. From Fig. 5, we can observe that the test losses for minority classes tended to increase as the training proceeds when AAFM was not used. On the other hand, the overall test loss and the test losses for the minority classes consistently decreased when AAFM was used. We can also observe that whereas the test losses for minority classes fluctuated when AAFM was not used, they stably decreased when AAFM was used. In addition, we compared the performance of the naive feature multiplier and AAFM with FixMatch+RECD in Appendix K. Finally," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.345, + 0.127 + ], + "angle": 0, + "content": "Park et al." + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.145, + 0.371, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.247, + 0.346, + 0.258 + ], + "angle": 0, + "content": "(a) FixMatch" + }, + { + "type": "image", + "bbox": [ + 0.374, + 0.144, + 0.503, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.382, + 0.246, + 0.495, + 0.258 + ], + "angle": 0, + "content": "(b) Without RECD" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.144, + 0.635, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.246, + 0.628, + 0.258 + ], + "angle": 0, + "content": "(c) Without AAFM" + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.144, + 0.768, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.246, + 0.759, + 0.258 + ], + "angle": 0, + "content": "(d) \\(\\mathrm{RECD + AAFM}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.269, + 0.787, + 0.312 + ], + "angle": 0, + "content": "Fig. 4: Confusion matrices of class predictions on the test set of CIFAR-10 generated by each algorithm (FixMatch+RECD with the described condition) trained on the training set of CIFAR-10-LT under \\(\\gamma_l = 100\\) and \\(\\gamma_u = 1\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.785, + 0.385 + ], + "angle": 0, + "content": "we argue that RECD also can be combined with various CISSL algorithms. To verify this, we conducted additional experiments using DARP, and demonstrated RECD-based DARP outperformed the original DARP in Appendix F." + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.409, + 0.49, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.513, + 0.445, + 0.525 + ], + "angle": 0, + "content": "(a) WithoutAAFM" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.409, + 0.723, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.513, + 0.668, + 0.526 + ], + "angle": 0, + "content": "(b) WithAAFM" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.537, + 0.785, + 0.566 + ], + "angle": 0, + "content": "Fig.5: Overall test loss and test losses for minority classes (Classes 8 and 9). The algorithms were trained on CIFAR-10-LT under \\(\\gamma_{l} = 100\\) and \\(\\gamma_{u} = 1\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.614, + 0.36, + 0.631 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We proposed a novel CISSL algorithm, RECD, which addresses the class distribution mismatch between labeled and unlabeled sets. RECD estimates the class distribution of the unlabeled set through MC approximation and uses it for rebalancing the classifier. Additionally, RECD reuses the estimated class distribution for AAFM to mitigate class imbalance in the feature map by densifying feature clusters of minority classes. Experimental results demonstrate that RECD outperforms the baseline algorithms under both scenarios where the class distributions of the labeled and unlabeled sets match and mismatch. Detailed analyses on RECD verify that the effectiveness of each component of RECD for mitigating class imbalance. However, the estimated class distribution of unlabeled set through MC approximation might be inaccurate when unlabeled training samples are scarce. In future work, we aim to improve the robustness of class distribution estimation, even with a limited amount of unlabeled samples." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.685, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RECD" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.178, + 0.785, + 0.223 + ], + "angle": 0, + "content": "This research was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (2023R1A2C2005453, RS-2023-00218913)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.248, + 0.323, + 0.264 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.281, + 0.785, + 0.323 + ], + "angle": 0, + "content": "1. Berthelot, D., Carlini, N., an Alex Kurakin, E.D.C., Sohn, K., Zhang, H., Raffel, C.: Remixmatch: Semi-supervised learning with distribution matching and augmentation anchoring. In: ICLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.324, + 0.785, + 0.351 + ], + "angle": 0, + "content": "2. Berthelot, D., Carlini, N., Goodfellow, I., Papernot, N., Oliver, A., Raffel, C.: Mixmatch: A holistic approach to semi-supervised learning. In: NeurIPS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.352, + 0.785, + 0.407 + ], + "angle": 0, + "content": "3. Cao, K., Wei, C., Gaidon, A., Arechiga, N., Ma, T.: Learning imbalanced datasets with label-distribution-aware margin loss. In: Wallach, H., Larochelle, H., Beygelzimer, A., d'Alché-Buc, F., Fox, E., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 32. Curran Associates, Inc. (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.408, + 0.785, + 0.449 + ], + "angle": 0, + "content": "4. Chawla, N.V., Bowyer, K.W., Hall, L.O., Kegelmeyer, W.P.: Smote: Synthetic minority over-sampling technique. Journal of Artificial Intelligence Research pp. 321-357 (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.449, + 0.785, + 0.477 + ], + "angle": 0, + "content": "5. Chu, P., Bian, X., Liu, S., Ling, H.: Feature space augmentation for long-tailed data. In: ECCV. vol. 12374, pp. 694-710 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.478, + 0.785, + 0.505 + ], + "angle": 0, + "content": "6. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.: Randaugment: Practical automated data augmentation with a reduced search space. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.505, + 0.785, + 0.533 + ], + "angle": 0, + "content": "7. Cui, Y., Jia, M., Lin, T.Y., Song, Y., Belongie, S.: Large scale fine-grained categorization and domain-specific transfer learning. In: CVPR. pp. 4109-4118 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.533, + 0.785, + 0.561 + ], + "angle": 0, + "content": "8. Cui, Y., Jia, M., Lin, T.Y., Song, Y., Belongie, S.: Class-balanced loss based on effective number of samples. In: CVPR. pp. 9268-9277 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.561, + 0.785, + 0.589 + ], + "angle": 0, + "content": "9. DeVries, T., Taylor, G.W.: Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.589, + 0.785, + 0.617 + ], + "angle": 0, + "content": "10. Estabrooks, A., Jo, T., Japkowicz, N.: A multiple resampling method for learning from imbalanced data sets. Computational Intelligence 20(1), 18-36 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.617, + 0.785, + 0.658 + ], + "angle": 0, + "content": "1. Fan, Y., Dai, D., Kukleva, A., Schiele, B.: Cossl: Co-learning of representation and classifier for imbalanced semi-supervised learning. In: CVPR. pp. 14574-14584 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.785, + 0.687 + ], + "angle": 0, + "content": "2. Gidaris, S., Singh, P., Komodakis, N.: Unsupervised representation learning by predicting image rotations. In: ICLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.687, + 0.785, + 0.714 + ], + "angle": 0, + "content": "3. Grandvalet, Y., Bengio, Y.: Semi-supervised learning by entropy minimization. In: NeurIPS (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.714, + 0.785, + 0.743 + ], + "angle": 0, + "content": "4. Guo, L.Z., Li, Y.F.: Class-imbalanced semi-supervised learning with adaptive thresholding. In: ICML. vol. 162, pp. 8082-8094 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.743, + 0.785, + 0.771 + ], + "angle": 0, + "content": "5. Huang, C., Li, Y., Loy, C.C., Tang, X.: Learning deep representation for imbalanced classification. In: CVPR. pp. 5375-5384 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.785, + 0.812 + ], + "angle": 0, + "content": "6. JAPKOWICZ, N.: The class imbalance problem: Significance and strategies. In: Proc. 2000 International Conference on Artificial Intelligence. vol. 1, pp. 111-117 (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.812, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Japkowicz, N.: The class imbalance problem: Significance and strategies. In: In Proceedings of the International Conference on Artificial Intelligence (2000)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.281, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.344, + 0.127 + ], + "angle": 0, + "content": "Park et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "18. Kang, B., Li, Y., Xie, S., Yuan, Z., Feng, J.: Exploring balanced feature spaces for representation learning. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "19. Kang, B., Xie, S., Rohrbach, M., Yan, Z., Gordo, A., Feng, J., Kalantidis, Y.: Decoupling representation and classifier for long-tailed recognition. In: ICLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.205, + 0.785, + 0.231 + ], + "angle": 0, + "content": "20. Kim, J., Hur, Y., Park, S., Yang, E., Hwang, S., Shin, J.: Distribution aligning refinery of pseudo-label for imbalanced semi-supervised learning. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.785, + 0.259 + ], + "angle": 0, + "content": "21. Krizhevsky, A.: Learning multiple layers of features from tiny images. Technical report, Department of Computer Science, University of Toronto (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.261, + 0.785, + 0.286 + ], + "angle": 0, + "content": "22. Kubat, M., Matwin, S., et al.: Addressing the curse of imbalanced training sets: one-sided selection. In: Icml. vol. 97, p. 179. Citeseer (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.288, + 0.785, + 0.327 + ], + "angle": 0, + "content": "23. Lai, Z., Wang, C., Gunawan, H., Cheung, S.C., Chuah, C.N.: Smoothed adaptive weighting for imbalanced semi-supervised learning: Improve reliability against unknown distribution data. In: ICML. vol. 162, pp. 11828-11843 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.329, + 0.785, + 0.383 + ], + "angle": 0, + "content": "24. Lazarow, J., Sohn, K., Lee, C.Y., Li, C.L., Zhang, Z., Pfister, T.: Unifying distribution alignment as a loss for imbalanced semi-supervised learning. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) pp. 5644-5653 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.385, + 0.785, + 0.424 + ], + "angle": 0, + "content": "25. Lee, D.H.: Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In: In Workshop on challenges in representation learning (ICML) (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.426, + 0.785, + 0.452 + ], + "angle": 0, + "content": "26. Lee, H., Kim, H.: Cdmad: Class-distribution-mismatch-aware debiasing for class-imbalanced semi-supervised learning. In: CVPR. pp. 23891-23900 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.454, + 0.785, + 0.479 + ], + "angle": 0, + "content": "27. Lee, H., Shin, S., Kim, H.: Abc: Auxiliary balanced classifier for class-imbalanced semi-supervised learning. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.508 + ], + "angle": 0, + "content": "28. Li, J., Meng, Z., Shi, D., Song, R., Diao, X., Wang, J., Xu, H.: Fcc: Feature clusters compression for long-tailed visual recognition. In: CVPR. pp. 24080-24089 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "29. Li, K., Kong, X., Lu, Z., Wenyin, L., Yin, J.: Boosting weighted elm for imbalanced learning. Neurocomputing 128, 15-21 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.785, + 0.577 + ], + "angle": 0, + "content": "30. Liu, X.Y., Wu, J., Zhou, Z.H.: Exploratory undersampling for class-imbalance learning. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics) 39(2), 539-550 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.579, + 0.785, + 0.604 + ], + "angle": 0, + "content": "31. Liu, Z., Miao, Z., Zhan, X., Wang, J., Gong, B., Yu, S.X.: Large-scale long-tailed recognition in an open world. In: CVPR. pp. 2537-2546 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.632 + ], + "angle": 0, + "content": "32. Menon, A.K., Jayasumana, S., Rawat, A.S., Jain, H., Veit, A., Kumar, S.: Long-tail learning via logit adjustment. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.634, + 0.785, + 0.673 + ], + "angle": 0, + "content": "33. Miyato, T., ichi Maeda, S., Koyama, M., Ishii, S.: Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE Transactions on Pattern Analysis and Machine Intelligence 41(8), 1979-1993 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.676, + 0.785, + 0.715 + ], + "angle": 0, + "content": "34. Oh, Y., Kim, D.J., Kweon, I.S.: Daso: Distribution-aware semantics-oriented pseudo-label for imbalanced semi-supervised learning. In: CVPR. pp. 9786-9796 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.785, + 0.743 + ], + "angle": 0, + "content": "35. Ren, J., Yu, C., Sheng, S., Ma, X., Zhao, H., Yi, S., Li, H.: Balanced meta-softmax for long-tailed visual recognition. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.785, + 0.771 + ], + "angle": 0, + "content": "36. Ren, M., Zeng, W., Yang, B., Urtasun, R.: Learning to reweight examples for robust deep learning. In: ICML. vol. 80, pp. 4334-4343 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.773, + 0.785, + 0.798 + ], + "angle": 0, + "content": "37. Sajjadi, M., Javanmardi, M., Tasdizen, T.: Regularization with stochastic transformations and perturbations for deep semi-supervised learning. In: NeurIPS (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "38. Sohn, K., Berthelot, D., Li, C.L., Zhang, Z., Carlini, N., Cubuk, E.D., Kurakin, A., Zhang, H., Raffel, C.: Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In: NeurIPS (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.685, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RECD" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "39. Verma, V., Kawaguchi, K., Lamb, A., Kannala, J., Solin, A., Bengio, Y., Lopez-Paz, D.: Interpolation consistency training for semi-supervised learning. Neural Networks 145, 90-106 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.217 + ], + "angle": 0, + "content": "40. Wang, R., Jia, X., Wang, Q., Wu, Y., Meng, D.: Imbalanced semi-supervised learning with bias adaptive classifier. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.787, + 0.259 + ], + "angle": 0, + "content": "41. Wang, S., Minku, L.L., Yao, X.: A learning framework for online class imbalance learning. IEEE Symposium on Computational Intelligence and Ensemble Learning (CIEL) pp. 36-45 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.787, + 0.3 + ], + "angle": 0, + "content": "42. Wang, S., Minku, L.L., Yao, X.: Resampling-based ensemble methods for online class imbalance learning. IEEE Transactions on Knowledge and Data Engineering 27(5), 1356-1368 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.787, + 0.328 + ], + "angle": 0, + "content": "43. Wang, X., Lian, L., Miao, Z., Liu, Z., Yu, S.X.: Long-tailed recognition by routing diverse distribution-aware experts. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.329, + 0.787, + 0.356 + ], + "angle": 0, + "content": "44. Wang, X., Wu, Z., Lian, L., Yu, S.X.: Debiased learning from naturally imbalanced pseudo-labels. In: CVPR. pp. 14647-14657 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.357, + 0.787, + 0.397 + ], + "angle": 0, + "content": "45. Wang, Y., Chen, H., Heng, Q., Hou, W., Fan, Y., Wu, Z., Wang, J., Savvides, M., Shinozaki, T., Raj, B., Schiele, B., Xie, X.: Freematch: Self-adaptive thresholding for semi-supervised learning. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.398, + 0.787, + 0.438 + ], + "angle": 0, + "content": "46. Wei, C., Sohn, K., Mellina, C., Yuille, A., Yang, F.: Crest: A class-rebalancing self-training framework for imbalanced semi-supervised learning. In: CVPR. pp. 10857-10866 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.466 + ], + "angle": 0, + "content": "47. Wei, T., Gan, K.: Towards realistic long-tailed semi-supervised learning: Consistency is all you need. In: CVPR. pp. 3469-3478 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.787, + 0.494 + ], + "angle": 0, + "content": "48. Xiang, L., Ding, G., Han, J.: Learning from multiple experts: Self-paced knowledge distillation for long-tailed classification. In: ECCV. pp. 247-263 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.495, + 0.787, + 0.536 + ], + "angle": 0, + "content": "49. Yu, F., Seff, A., Zhang, Y., Song, S., Funkhouser, T., Xiao, J.: Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.537, + 0.787, + 0.563 + ], + "angle": 0, + "content": "50. Yu, Z., Li, Y., Lee, Y.J.: Inpl: Pseudo-labeling the inliers first for imbalanced semi-supervised learning. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.787, + 0.591 + ], + "angle": 0, + "content": "51. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. In: ICLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.787, + 0.619 + ], + "angle": 0, + "content": "52. Zhang, Y., Hooi, B., Hong, L., Feng, J.: Self-supervised aggregation of diverse experts for test-agnostic long-tailed recognition. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.787, + 0.647 + ], + "angle": 0, + "content": "53. Zhong, Z., Cui, J., Liu, S., Jia, J.: Improving calibration for long-tailed recognition. In: CVPR. pp. 16489-16498 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.647 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Rebalancing Using Estimated Class Distribution for Imbalanced Semi-Supervised Learning under Class Distribution Mismatch/c1f56d15-2542-45a0-80c7-7ebab6520f43_origin.pdf b/2024/Rebalancing Using Estimated Class Distribution for Imbalanced Semi-Supervised Learning under Class Distribution Mismatch/c1f56d15-2542-45a0-80c7-7ebab6520f43_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..24c15929aea2a9b411f3437258d147aa6f5c5281 --- /dev/null +++ b/2024/Rebalancing Using Estimated Class Distribution for Imbalanced Semi-Supervised Learning under Class Distribution Mismatch/c1f56d15-2542-45a0-80c7-7ebab6520f43_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2aa4ad586d2917764c1c578115c2add416c6e71fd4e5f43a5be681feb58e6458 +size 2404266 diff --git a/2024/Rebalancing Using Estimated Class Distribution for Imbalanced Semi-Supervised Learning under Class Distribution Mismatch/full.md b/2024/Rebalancing Using Estimated Class Distribution for Imbalanced Semi-Supervised Learning under Class Distribution Mismatch/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ac9b7df7f36e6710e9718b5e280d4a80ec717dce --- /dev/null +++ b/2024/Rebalancing Using Estimated Class Distribution for Imbalanced Semi-Supervised Learning under Class Distribution Mismatch/full.md @@ -0,0 +1,355 @@ +# Rebalancing Using Estimated Class Distribution for Imbalanced Semi-Supervised Learning under Class Distribution Mismatch + +Taemin Park\**, Hyuck Lee\**, and Heeyoung Kim + +Department of Industrial and Systems Engineering, KAIST + +Daejeon 34141, Republic of Korea + +{ptm1001,dlgur0921,heeyoungkim}@kaist.ac.kr + +Abstract. Despite significant advancements in class-imbalanced semi-supervised learning (CISSL), many existing algorithms explicitly or implicitly assume that the class distribution of unlabeled data matches that of labeled data. However, when this assumption fails in practice, the classification performance of such algorithms may degrade due to incorrectly assigned weight to each class during training. We propose a novel CISSL algorithm called Rebalancing Using Estimated Class Distribution (RECD). RECD estimates the unknown class distribution of unlabeled data through Monte Carlo approximation, leveraging predicted class probabilities for unlabeled samples, and subsequently rebalances the classifier based on the estimated class distribution. Additionally, we propose an extension of feature clusters compression in the context of CISSL to mitigate feature map imbalance by densifying minority class clusters. Experimental results on four benchmark datasets demonstrate that RECD achieves state-of-the-art classification performance in CISSL. + +Keywords: Class-imbalanced semi-supervised learning $\cdot$ Long-tailed learning $\cdot$ Auxiliary Balanced Classifier + +# 1 Introduction + +In real-world datasets, the number of samples often varies across different classes. Deep neural network algorithms trained on such datasets tend to exhibit bias towards majority classes, which have a larger number of samples in the dataset [15, 31, 41]. To mitigate the bias, various class-imbalanced learning (CIL) techniques have been proposed, including re-sampling [4, 10, 30, 42], re-weighting [8, 29, 36], decoupled learning [19, 35, 53], and multi-expert learning [43, 48, 52]. + +Although the CIL techniques significantly mitigate the bias, they typically assume that the entire training set is labeled. However, in reality, labeling samples requires substantial resources, resulting in many real-world datasets containing + +a significant proportion of unlabeled data. Because the CIL techniques assign importance to each data point based on its corresponding label, they cannot be directly applied when the training set includes unlabeled samples. + +Recently, several class-imbalanced semi-supervised learning (CISSL) algorithms, including ABC [27], UDAL [24], CReST [46], DARP [20], DASO [34], CoSSL [11], and SAW [23], have emerged to address the bias towards majority classes in scenarios where the training set consists of both labeled and unlabeled data. However, ABC, UDAL, and CReST explicitly assume that the class distribution of the unlabeled data matches that of the labeled data. Similarly, DARP, DASO, and SAW can be seen as implicitly assuming the class distribution match between the labeled and unlabeled data, as they incorporate CIL techniques such as LA [32] and cRT [19], which rely on the same class distribution between the labeled and unlabeled data. Additionally, CoSSL solely considers the class distribution of labeled data in its main component, TFE. + +However, the class distribution of the unlabeled set is often unknown and may differ from that of the labeled set [34, 40, 47]. In this case, the classification performance of the aforementioned algorithms may degrade because the unknown class distribution of the unlabeled set is not considered in the rebalancing process. Fig. 1b illustrates the class predictions on the test set of CIFAR-10 [21] using ReMixMatch+ABC, an existing CISSL algorithm, trained on the long-tailed version of the CIFAR-10 [8] training set where the class distributions of labeled and unlabeled sets mismatch, as presented in Fig. 1a. From Fig. 1b, we can observe that many test samples are mispredicted as Class 10 (the most minor class). This may be because ABC rebalances the classifier excessively, considering only the class distribution of the labeled set. + +![](images/e1202bd77e9c19ecee6374f042f9c2d15ec144c619b9484c1d49e46de9ddfeb2.jpg) +(a) Class distributions + +![](images/c84920ac757b6feb2e5d7b92219ee731492bdb788be55a9fb45bd01ae22453f0.jpg) +(b) ReMixMatch+ABC +Fig. 1: Comparison of the predictions of ReMixMatch+ABC and the proposed algorithm (ReMixMatch+RECD) trained on an imbalanced version of CIFAR-10 where the class distributions of labeled and unlabeled sets mismatch, as illustrated in Fig. 1a. + +![](images/8c66110307eb62a2654e3e21c59902955c2a9699d7b114a1ce8d60cf5a1659c9.jpg) +(c) Proposed algorithm + +To rebalance the classifier appropriately when the class distribution of the unlabeled set mismatches that of the labeled set, we propose a novel CISSL algorithm called Rebalancing Using Estimated Class Distribution (RECD), which estimates the unknown class distribution using Monte Carlo (MC) approximation and rebalances the training loss based on the estimated class distribution. Specifically, RECD is a modular algorithm that operates in conjunction with previous CISSL algorithms. While RECD can generally be integrated with most + +CISSL algorithms by providing additional information about the class distribution of the unlabeled set, in this paper, we specifically combine RECD with ABC due to its simple and effective framework compared to other CISSL algorithms. + +To estimate the class distribution of unlabeled data, RECD first computes the average class predictions of ABC for the entire unlabeled set. Because generating class predictions for the entire unlabeled set can be computationally intensive, RECD generates class predictions on an unlabeled minibatch for each iteration and updates the exponential moving average (EMA) of the class predictions on the unlabeled set. Utilizing MC approximation, this average class prediction serves as the classifier's prior, representing the estimated unknown class distribution of the unlabeled set. Then, RECD rebalances the training loss on unlabeled samples for ABC based on the estimated class distribution of the unlabeled set, and trains ABC using this rebalanced unlabeled loss in addition to the training loss on labeled samples for ABC. As depicted in Fig. 1c, ReMix-Match+RECD generates nearly balanced class predictions on the test set. + +In addition to mitigating imbalance in the classifier, RECD also mitigates imbalance in the feature map. While existing CISSL algorithms primarily focus on mitigating classifier bias, there also exists imbalance in the feature map [5,18]. Specifically, feature clusters of minority classes tend to be sparser compared to the dense clusters of majority classes. These sparse features often overlap with clusters of other classes in the feature map, posing challenges in classifying minority class features [28]. To address this issue by densifying the sparse feature clusters of minority classes, we extend feature clusters compression (FCC) [28], a previous CIL technique, into the context of CISSL by leveraging the unknown class distribution of the unlabeled data estimated by RECD. We refer to this extended version of FCC as the adaptively adjusted feature multiplier (AAFM). + +Our experimental results on CIFAR-10-LT [8], CIFAR-100-LT [8], STL-10-LT [20], and Small-ImageNet-127 [11] demonstrate the superior performance of the proposed algorithm compared to the baseline algorithms, regardless of whether the class distributions of labeled and unlabeled sets match or mismatch. Through detailed analyses, we verify that RECD effectively estimates the true class distribution of the unlabeled set and uses it to mitigate class imbalance. Additionally, we validate that AAFM not only reduces the test loss for minority classes but also stabilizes the training process. Finally, by conducting experiments that combine RECD with DARP [20], we validate the effectiveness of RECD beyond its application solely with ABC. The source codes for RECD are available at https://github.com/taemin-park/RECD. + +# 2 Related work + +Many CISSL algorithms have been proposed to address class imbalance in scenarios where the training set consists of both labeled and unlabeled data. For example, ABC [27] trains an auxiliary balanced classifier with a training loss that is rebalanced based on the class distribution of the labeled data. CReST [46] systematically enlarges the labeled dataset by assigning the labels to unlabeled + +samples predicted as minority classes more frequently than those predicted as majority classes. UDAL [24] combines progressive distribution alignment and logit-adjustment [32] to adjust the training loss. ABC, CReST, and UDAL have effectively mitigated bias towards majority classes. However, they assumed that the class distribution of the unlabeled data is the same as that of the labeled data, which may pose practical limitation as described in Sec. 1. + +DARP [20] solves a convex optimization problem to refine biased pseudo labels. DASO [34] blends class predictions of a similarity-based classifier and a linear classifier to obtain unbiased pseudo-labels. SAW [23] estimates the learning difficulty of each class and smoothly reweights the training loss based on the estimated learning difficulty. However, DARP, DASO, and SAW implicitly assume the same class distributions of the unlabeled and labeled datasets, as they incorporate CIL techniques such as LA [32] and cRT [19], which rebalance the classifier based on the class distribution of the labeled data. + +CoSSL [11] integrates Tail-class Feature Enhancement (TFE) and class-balanced sampling to rebalance the classifier. InPL [50] introduces energy-based pseudolabeling to effectively discern correctly predicted unlabeled samples of minority classes. DePL [44] mitigates pseudo-label bias using counterfactual reasoning and adaptive marginal loss. L2AC [40] mitigates training bias by using a bias adaptive classifier, which consists of a bias attractor and a linear classifier. Adsh [14] sets an adaptive confidence threshold for each class to utilize unlabeled samples predicted as minority classes more frequently than those predicted as majority classes. ACR [47] estimates the class distribution of the unlabeled set by comparing class predictions with class distribution prototypes, and then adaptively adjusts the training loss based on the estimated class distribution. CDMAD [26] refines the predictions for unlabeled and test samples based on the logits for solid color image. Although these algorithms have significantly improved CISSL, this area remains relatively less explored compared to SSL and CIL. + +# 3 Preliminaries + +# 3.1 Problem settings + +We have a training dataset that consists of the labeled set $\mathcal{X} = \{(x_n, y_n) : n \in (1, \ldots, N)\}$ and unlabeled set $\mathcal{U} = \{(u_m) : m \in (1, \ldots, M)\}$ , where $x_n \in R^d$ and $y_n \in R^K$ are the $n$ th labeled sample and its corresponding label, respectively, and $u_m \in R^d$ is the $m$ th unlabeled sample. For each class $k$ , $N_k$ and $M_k$ denote the numbers of the labeled and unlabeled samples, respectively, i.e., $\sum N_k = N$ and $\sum M_k = M$ . In CISSL, it is typically assumed that $N_1 \geq N_2 \geq \ldots \geq N_K$ and $M_1 \geq M_2 \geq \ldots \geq M_K$ . We denote the class imbalance ratios of the labeled and unlabeled sets as $\gamma_l = \frac{N_1}{N_K}$ and $\gamma_u = \frac{M_1}{M_K}$ , respectively. When $M_k$ is unknown, $\gamma_u$ would also be unknown, and it may differ from $\gamma_l$ . For each training iteration, we generate minibatches $\mathcal{MB}_{\mathcal{X}} = \{(x_b^m, y_b^m) : b \in (1, \ldots, B_l)\} \subset \mathcal{X}$ and $\mathcal{MB}_{\mathcal{U}} = \{(u_b^m) : b \in (1, \ldots, B_u)\} \subset \mathcal{U}$ from $\mathcal{X}$ and $\mathcal{U}$ , respectively, where $B_l$ and $B_u$ are the sizes of the labeled and unlabeled minibatches, respectively. Using $\mathcal{MB}_{\mathcal{X}}$ and $\mathcal{MB}_{\mathcal{U}}$ , we aim to train an algorithm $f_\theta: R^d \to R^K$ that accurately predicts the + +class probabilities for unseen samples, where $\theta$ denotes the network parameters. We denote the extracted features before the classification layer as $\xi (\cdot)$ + +# 3.2 Backbone SSL algorithm + +Following previous CISSL algorithms [11, 20, 23, 27, 34, 50], either FixMatch [38] or ReMixMatch [1] is used as a backbone SSL algorithm for RECD. To utilize unlabeled samples, FixMatch and ReMixMatch conduct consistency regularization using weak data augmentation techniques $\alpha(\cdot)$ (flipping and cropping) and strong data augmentation techniques $\mathcal{A}(\cdot)$ (Cutout [9] and RandomAugment [6]). + +Specifically, FixMatch first predicts the class probabilities for $\alpha(u_b^m)$ as $f(\alpha(u_b^m))$ and then generates the pseudo-label as $q_b^m = \arg \max_k (f(\alpha(u_b^m))_k)$ when the confidence exceeds the confidence threshold $\epsilon$ , i.e., $\max_k (f(\alpha(u_b^m))_k) \geq \epsilon$ , where $(\cdot)_k$ denotes the $k$ th element of $(\cdot)$ . Then, FixMatch predicts the class probabilities for $\mathcal{A}(u_b^m)$ as $f(\mathcal{A}(u_b^m))$ and ensures $f(\mathcal{A}(u_b^m))$ to be close to the pseudo-label $q_b^m$ by minimizing a loss that penalizes the distance between $f(\mathcal{A}(u_b^m))$ and $q_b^m$ . In addition, FixMatch minimizes the prediction loss calculated between weakly augmented labeled sample $\alpha(x_b^m)$ and $p_b^m$ , which is the one-hot version of $y_b^m$ . + +Similarly, ReMixMatch first predicts the class probabilities of $\alpha(u_b^m)$ as $q_b^m = f(\alpha(u_b^m))$ . Then ReMixMatch applies a distribution alignment technique as $\tilde{q}_b^m = \mathrm{Normalize}(q_b^m \times q(y) / \tilde{q}(y))$ , where $\mathrm{Normalize}(x)_i = x_i / \sum_j x_j$ , $\tilde{q}(y)$ is the moving average of class predictions on unlabeled samples, and $q(y)$ indicates the class distribution of the training set. Finally, ReMixMatch sharpens $\tilde{q}_b^m$ as $\bar{q}_b^m = \mathrm{Normalize}((\tilde{q}_b^m)^{1/T})$ , where $T$ is the temperature for sharpening. Using $\bar{q}_b^m$ as pseudo-labels, ReMixMatch conducts consistency regularization [33,37], entropy minimization [13,25] and mixup regularization [2,39,51]. ReMixMatch also minimizes the rotation loss [12] for self-supervised learning. The training loss of the backbone SSL algorithm, $L_{back}$ , is detailed in Appendix B. + +# 3.3 Auxiliary balanced classifier (ABC) + +To use a balanced classifier trained on a class-balanced subset of the training set, while benefiting from the high-quality representations learned from the entire training set, Lee and Kim [27] introduced an auxiliary balanced classifier (ABC) attached to the representation layer of a backbone SSL algorithm. To ensure that ABC is trained on a class-balanced subset of the training set, ABC generates a $0/1$ mask $M(\cdot)$ for each training sample, where the $0/1$ mask determines whether each sample contributes to the training loss. Specifically, ABC samples a $0/1$ mask from a Bernoulli distribution for each labeled and unlabeled sample as + +$$ +M (x _ {b} ^ {m}) = \mathcal {B} (\frac {N _ {K}}{N _ {\hat {y} _ {b} ^ {m}}}) \quad \mathrm {a n d} \quad M (u _ {b} ^ {m}) = \mathcal {B} (\frac {N _ {K}}{N _ {\hat {q} _ {b} ^ {m}}}). \qquad (1) +$$ + +$\mathcal{B}(\cdot)$ denotes a Bernoulli distribution and $\hat{q}_b^m = \arg \max (q_b^m)$ , where $q_{b}^{m}$ is the class prediction of the ABC for an unlabeled sample $u_{b}^{m}$ . Note that in Eq. (1), the probability of sampling 1 is adjusted to be inversely proportional to the number + +of training samples belonging to each class. This sampling strategy ensures that ABC is trained with an equal frequency from each class's data. + +With the $0/1$ mask, ABC calculates the classification loss, $L_{cls}$ , between weakly augmented labeled sample $\alpha(x_b^m)$ and its corresponding one-hot version label $p_b^m$ . In addition, ABC also calculates a consistency regularization loss, $L_{con}$ , between weakly augmented unlabeled sample $\alpha(u_b^m)$ and strongly augmented unlabeled sample $\mathcal{A}(u_b^m)$ , which is minimized to ensure that decision boundaries lie in low-density regions. The training losses $L_{cls}$ and $L_{con}$ are calculated as: + +$$ +L _ {c l s} = \frac {1}{B _ {l}} \sum_ {b = 1} ^ {B _ {l}} M \left(x _ {b} ^ {m}\right) \mathcal {H} \left(p _ {b} ^ {m}, f _ {a b c} (\xi \left(\alpha \left(x _ {b} ^ {m}\right)\right))\right), \tag {2} +$$ + +$$ +L _ {c o n} = \frac {1}{B _ {u}} \sum_ {b = 1} ^ {B _ {u}} \sum_ {i = 1} ^ {2} M \left(u _ {b} ^ {m}\right) \mathbb {1} \left(\max \left(q _ {b} ^ {m}\right) \geq \epsilon\right) \mathcal {H} \left(q _ {b} ^ {m}, f _ {a b c} \left(\xi \left(\mathcal {A} _ {i} \left(u _ {b} ^ {m}\right)\right)\right)\right), \tag {3} +$$ + +where $\mathcal{H}$ is the cross-entropy loss, $q_{b}^{m} = f_{abc}(\xi (\alpha (u_{b}^{m})))$ , $f_{abc}$ denotes the auxiliary classifier attached to the representation layer, and $\mathbb{1}(\cdot)$ denotes an indicator function. With the $L_{cls}$ , $L_{con}$ and the training loss of backbone SSL algorithm $L_{back}$ , total training loss of ABC is calculated as: + +$$ +L _ {A B C} = L _ {c l s} + L _ {c o n} + L _ {b a c k}. \tag {4} +$$ + +ABC [27] effectively mitigated class imbalance when the class distributions of the labeled and unlabeled sets match. However, as we can observe in the second equation of the Eq. (1), because the $0/1$ masks for unlabeled samples are generated based on the class distribution of the labeled set without considering that of the unlabeled set, $f_{abc}$ may not be properly rebalanced when the unknown class distribution of the unlabeled set significantly mismatches that of the labeled set. + +# 3.4 Feature clusters compression (FCC) [28] + +To separate features extracted from samples of minority classes from those of majority classes by densifying the feature clusters, Li et al. [28] introduced FCC. For a training sample $x$ of class $k$ , FCC multiplies the extracted feature $\xi(x)$ by a scaling factor $\tau_k$ , and uses $\tau_k \times \xi(x)$ as input for the classification layer $\psi$ , as + +$$ +f _ {\theta} (x) = \psi \left(\tau_ {k} \times \xi (x)\right), \tag {5} +$$ + +where $\tau_{k}$ decreases from majority to minority classes in a naive manner as follows: + +$$ +\tau_ {k} = 1 + \eta \times \left(1 - \frac {k}{K}\right), k = 1, \dots , K, \tag {6} +$$ + +where $\eta$ is a scale hyperparameter. Since the features of majority classes are multiplied by relatively large scaling factors compared to those of minority classes, their feature clusters occupy larger regions on the feature map during training. In this context, to prevent minority class features from crossing the decision + +boundary, the feature extractor $\xi(\cdot)$ would be trained to map minority class features into denser regions. Note that during the test phase, the features extracted from test samples are not multiplied by the scaling factor, ensuring that the classifier predicts classes for test samples with densified features. According to Li et al. [28], FCC effectively enhances the separability of feature clusters of minority classes in CIL. However, in CISSL, it may fail to appropriately densify minority class features because the class distribution of the unlabeled set cannot be considered in the scaling factor $\tau_{k}$ . In this paper, we refer to the scaling factor in Eq. (6) as the naive feature multiplier. + +# 4 Methodology + +# 4.1 Rebalancing using estimated class distribution (RECD) + +As described in Sec. 1, RECD incorporates the unknown class distribution of the unlabeled data by rebalancing the classifier based on the estimated class distribution of the unlabeled data. RECD is integrated with ABC [27], which rebalances the training loss using the $0/1$ mask, as described in Eq. (1). Whereas ABC generates $0/1$ masks for unlabeled samples based on the class distribution of the labeled set, RECD replaces the parameter of $0/1$ masks for unlabeled samples using the estimated class distribution of the unlabeled set. + +Specifically, to estimate the unknown class distribution of the unlabeled set, $p(y_{u})$ , RECD first computes the average class predictions of ABC on the entire unlabeled set as $\frac{1}{M}\sum_{m = 1}^{M}p_{\theta}(y_u|u_m)$ , where $p_{\theta}(y_u|u_m) = f_{abc}(\xi (\alpha (u_m)))$ . Then, using MC approximation, $\frac{1}{M}\sum_{m = 1}^{M}p_{\theta}(y_u|u_m)$ becomes $p_{\theta}(y_u)$ as follows: + +$$ +\frac {1}{M} \sum p _ {\theta} \left(y _ {u} \mid u _ {m}\right) \approx \int p _ {\theta} \left(y _ {u} \mid u\right) p (u) d u = p _ {\theta} \left(y _ {u}\right). \tag {7} +$$ + +In Eq. (7), $p_{\theta}(y_u)$ can be regarded as the classifier's prior for the unlabeled samples, and we use this classifier's prior as the estimation of the class distribution of the unlabeled set $p(y_u)$ based on that the classifier's prior is highly correlated with the class distribution, as analyzed in Fig. 3. + +Using the estimated class distribution $p_{\theta}(y_u)$ , RECD modifies the sampling distribution of the $0/1$ mask for unlabeled samples in Eq. (1) as follows: + +$$ +M \left(u _ {b} ^ {m}\right) = \mathcal {B} \left(\frac {\operatorname* {m i n} \left(p _ {\theta} \left(y _ {u}\right)\right)}{p _ {\theta} \left(y _ {u}\right) \hat {q} _ {b} ^ {m}}\right). \tag {8} +$$ + +By setting the parameter of the $0/1$ mask based on $p_{\theta}(y_u)$ , RECD rebalances the training loss of the unlabeled samples to an appropriate extent even when the class distributions of the labeled and unlabeled sets severely mismatch. The total training loss of RECD is described in detail in Sec. 4.3. + +RECD may increase computational complexity compared to the original ABC because, as shown in Eq. (7), computing class predictions of ABC for the entire unlabeled set for each iteration or epoch would require a substantial amount of + +computation, especially for large-scale datasets such as LSUN [49] or iNaturallist [7]. To address this issue, RECD replaces the naive average $\frac{1}{M}\sum p_{\theta}(y_u|u_m)$ with an EMA of the class predictions on the unlabeled samples, which is updated for each iteration of the training as follows: + +$$ +p _ {\theta} \left(y _ {u}\right) _ {n e w} = \rho \times p _ {\theta} \left(y _ {u}\right) _ {o l d} + (1 - \rho) \times p _ {\theta} \left(y _ {u}\right) _ {b a t c h}, \tag {9} +$$ + +where $\rho$ is a hyperparameter for the EMA update, $p_{\theta}(y_u)_{old}$ is $p_{\theta}(y_u)$ before the update, and $p_{\theta}(y_u)_{batch}$ is the average class prediction on the weakly augmented unlabeled minibatch, $\frac{1}{B_u}\sum_{b = 1}^{B_u}q_b^m = \frac{1}{B_u}\sum_{b = 1}^{B_u}f_{abc}(\xi (\alpha (u_b^m)))$ . Updating the EMA of the class predictions on unlabeled samples adds negligible computational cost compared to the original ABC because RECD reuses $q_b^m$ that is calculated to conduct consistency regularization for ABC, as in Eq. (3). + +# 4.2 Adaptively adjusted feature multiplier (AAFM) + +Although the naive feature multiplier effectively densifies feature clusters of the minority classes in CIL, its efficacy may diminish in CISSL due to its lack of consideration for the class distribution of the unlabeled set. In addition, the naive feature multiplier linearly reduces from majority to minority classes as shown in Eq. (6), without considering the degree of class imbalance in the training set. These limitations may lead to inappropriate scaling factors for the extracted features, resulting in insufficient densification or unstable training, as discussed in Sec. 5.3. To address this issue by incorporating both the class imbalance ratio of the labeled set and the unknown class distribution of the unlabeled set, we introduce an advanced version of the naive feature multiplier, AAFM. For the labeled samples, AAFM for the $k$ th class, $\tau_k^l$ , is calculated as follows: + +$$ +\tau_ {k} ^ {l} = 1 + \eta \times \left(\frac {N _ {k}}{N _ {1}} \times C _ {l k} ^ {t}\right), k = 1, \dots , K, \tag {10} +$$ + +where $\eta$ is a scale hyperparameter similar to $\eta$ in Eq. (6), and $t$ is a hyperparameter for $C_{lk}$ , which represents the EMA of confidence for labeled samples of the $k$ th class, which is updated for each iteration of the training process as follows: + +$$ +C _ {l k} = \omega \times C _ {l k} ^ {\text {o l d}} + (1 - \omega) \times C _ {l k} ^ {\text {b a t c h}}, \tag {11} +$$ + +where $\omega$ is a hyperparameter for the EMA update, $C_{lk}^{old}$ is $C_{lk}$ before the update, and $C_{lk}^{batch}$ is the average confidence for class $k$ on a minibatch, which is calculated as $\frac{1}{n_k}\sum_{b = 1}^{B_l}f_{abc}(\xi (\alpha (x_b^m)))_k\times \mathbb{1}(y_b^m = k)$ . Here, $n_k$ denotes the number of samples for the $k$ th class on a minibatch. Note that $C_{lk}$ is updated only when $n_k\geq 1$ . By multiplying $C_{lk}^{t}$ to the second term on the right-hand side of Eq. (10), we stabilize the training process, assigning smaller AAFMs to classes whose feature clusters are located near the decision boundary. + +We also multiply AAFM $\tau_{k}^{u}$ with the features extracted from strongly augmented unlabeled samples, where $\tau_{k}^{u}$ incorporates the unknown class distribution + +of the unlabeled set by reusing the estimated class distribution $p_{\theta}(y_u)$ as follows: + +$$ +\tau_ {k} ^ {u} = 1 + \eta \times \left(\frac {p _ {\theta} \left(y _ {u}\right) _ {k}}{\operatorname* {m a x} \left(p _ {\theta} \left(y _ {u}\right)\right)} \times C _ {u k} ^ {t}\right), k = 1, \dots , K. \tag {12} +$$ + +In Eq. (12), $C_{uk}$ is an EMA of confidence for unlabeled samples predicted as the $k$ th class, which is updated for each iteration of the training process as follows: + +$$ +C _ {u k} = \omega \times C _ {u k} ^ {\text {o l d}} + (1 - \omega) \times C _ {u k} ^ {\text {b a t c h}}, \tag {13} +$$ + +where $C_{uk}^{batch} = \frac{1}{m_k}\sum_{b = 1}^{B_u}f_{abc}(\xi (\alpha (u_b^m)))_k\times \mathbb{1}(\hat{q}_b^m = k)$ , with $m_{k} = \sum_{b = 1}^{B_{u}}\mathbb{1}(\hat{q}_{b}^{m} = k)$ and $\hat{q}_b^m = \arg \max_k(q_b^m)_k$ . $C_{uk}$ is updated only when $m_{k}\geq 1$ . Since $\tau_k^u$ incorporates the unknown class distribution of the unlabeled set, AAFM can densify feature clusters of minority classes to an appropriate extent even when the class distributions of the labeled and unlabeled sets significantly mismatch. + +# 4.3 End-to-end training of RECD + +Fig. 2 shows the end-to-end training process of the proposed algorithm for labeled and unlabeled data, respectively. The total training loss of RECD, denoted as $L_{RECD}$ , closely resembles that of ABC in Eq. (4), and is formulated as follows: + +$$ +L _ {R E C D} = L _ {c l s} + L _ {c o n} + L _ {b a c k}, \tag {14} +$$ + +where $L_{cls}$ and $L_{con}$ are improved by applying RECD and AAFM as follows: + +$$ +L _ {c l s} = \frac {1}{B _ {l}} \sum_ {b = 1} ^ {B _ {l}} M \left(x _ {b} ^ {m}\right) \mathcal {H} \left(p _ {b} ^ {m}, f _ {a b c} \left(\xi \left(\alpha \left(x _ {b} ^ {m}\right)\right) \times \tau_ {y _ {b} ^ {m}} ^ {l}\right)\right), \tag {15} +$$ + +$$ +L _ {c o n} = \frac {1}{B _ {u}} \sum_ {b = 1} ^ {B _ {u}} \sum_ {i = 1} ^ {2} M \left(u _ {b} ^ {m}\right) \mathcal {H} \left(q _ {b} ^ {m}, f _ {a b c} \left(\xi \left(\mathcal {A} _ {i} \left(u _ {b} ^ {m}\right)\right) \times \tau_ {\hat {q} _ {b} ^ {m}} ^ {u}\right)\right), \tag {16} +$$ + +$$ +M \left(x _ {b} ^ {m}\right) = \mathcal {B} \left(\frac {N _ {K}}{N _ {y _ {b} ^ {m}}}\right) \quad \text {a n d} \quad M \left(u _ {b} ^ {m}\right) = \mathcal {B} \left(\frac {\min \left(p _ {\theta} \left(y _ {u}\right)\right)}{p _ {\theta} \left(y _ {u}\right) \hat {q} _ {b} ^ {m}}\right), \tag {17} +$$ + +where $p_b^m$ is the one-hot version of $y_b^m$ , $q_b^m = f_{abc}(\xi(\alpha(u_b^m)))$ , and $\tau_{y_b^m}^l$ and $\tau_{\hat{q}_b^m}^u$ are AAFMs corresponding to $y_b^m$ and $\hat{q}_b^m$ , respectively. To improve the quality of features learned by the backbone SSL algorithm, $L_{back}$ is also adjusted with two modifications. First, AAFM is applied to further densify feature clusters of minority classes during the learning of the backbone SSL algorithm. Second, to train the backbone algorithm with unbiased pseudo-labels, pseudo-labels of the backbone algorithm are generated by $f_{abc}$ instead of the classifier of the backbone algorithm. The pseudo-code of RECD is presented in Appendix D. + +![](images/1fc43e714bf55d86b81286c90986c5acdd3f53a84fe6f24b363135a6d6a80d50.jpg) +(a) Balanced training with labeled data + +![](images/fa73259374f230d43d91cee017ef1b4125b3aeaac7bf67b8dda584b3d0191815.jpg) +(b) Balanced training with unlabeled data +Fig. 2: End-to-end training of the proposed algorithm + +# 5 Experiments + +# 5.1 Experimental setup + +Following previous CISSL studies [11, 20, 23, 34, 40], we conducted experiments on CIFAR-10-LT [8], CIFAR-100-LT [8], STL10-LT [20], and Small-ImageNet-127 [11]. The detailed descriptions for the datasets are provided in Appendix E.1. We evaluated the classification performance using balanced accuracy (bACC) (referred to as the averaged class recall in [11]) [15] and geometric mean (GM) [22]. The performance of the proposed algorithm was compared with the following baseline algorithms: vanilla algorithm, CIL algorithms [3, 16, 19], SSL algorithms [1, 38], and CISSL algorithms [11, 14, 20, 23, 24, 27, 40, 44, 46]. Algorithms such as DASO [34] and ACR [47], whose classification performances were measured under different experimental settings from ours, were compared with RECD in Appendix G. The detailed descriptions of the baseline algorithms and training setups are provided in Appendix E.2. + +# 5.2 Experimental results + +Tab. 1 summarizes the classification performances of the baseline algorithms and RECD on CIFAR-10-LT under $\gamma_l = \gamma_u$ . First, the CISSL algorithms achieved higher performances than the vanilla algorithm, CIL algorithms, and SSL algorithms. These results verify the importance of mitigating class imbalance and utilizing unlabeled samples. Furthermore, RECD achieved the highest performance among the CISSL algorithms. Specifically, RECD showed a tendency to outperform other CISSL algorithms by a large margin as the class imbalance ratio increased, indicating its effectiveness in mitigating class imbalance. These results may be attributed to the unique ability of the proposed algorithm to mitigate class imbalance in the feature map, as we analyze in Sec. 5.3. + +Tab. 2 summarizes the performances of the baseline CISSL algorithms and RECD on CIFAR-100-LT. RECD outperformed the baseline algorithms. Given that CIFAR-100-LT has 100 classes, these results suggest that RECD may also be suitable for datasets with a large number of classes. Considering that classes with the fewest labeled data points have only three labeled samples under $\gamma = 50$ , + +Table 1: Comparison of bACC/GM on CIFAR-10-LT under $\gamma = {\gamma }_{l} = {\gamma }_{u}$ + +
CIFAR-10-LT (γ = γt = γu)
Algorithmγ = 50γ = 100γ = 150
Vanilla (Cross-Entropy Loss)65.2±0.05/61.1±0.0958.8±0.13/58.2±0.1155.6±0.43/
Re-sampling [17]64.3±0.48/60.6±0.6755.8±0.47/45.1±0.3052.2±0.05/
LDAM-DRV [3]68.9±0.07/67.0±0.0862.8±0.17/58.9±0.6057.9±0.20/
cRT [19]67.8±0.13/66.3±0.1563.2±0.45/59.9±0.4059.3±0.10/
FixMatch [38]79.2±0.33/77.8±0.3671.5±0.72/66.8±1.5168.4±0.15/
FixMatch+DARP+cRT [19, 20]85.8±0.43/85.6±0.5682.4±0.26/81.8±0.1779.6±0.42/
FixMatch+CRST+LA [32, 46]85.6±0.36/81.9±0.4581.2±0.70/74.5±0.9971.9±2.24/
FixMatch+ABC [27]85.6±0.26/85.2±0.2981.1±1.14/80.3±1.2977.3±1.25/
FixMatch+CoSSL [11]86.8±0.30/86.6±0.2583.2±0.49/82.7±0.6080.3±0.55/
FixMatch+SAW+LA [23, 32]86.2±0.15/83.9±0.3580.7±0.15/77.5±0.2173.7±0.06/
FixMatch+Adsh [14]83.4±0.06/-76.5±0.35/-71.5±0.30/
FixMatch+DebiasPL [44]- / --80.6±0.50/-- / -
FixMatch+UDAL [24]86.5±0.29/-81.4±0.39/-77.9±0.33/
FixMatch+L2AC [40]- / --82.1±0.57/81.5±0.6477.6±0.53/
FixMatch+RECD87.3±0.18/87.2±0.1984.0±0.13/83.6±0.1680.6±0.53/
ReMixMatch [1]81.5±0.26/80.2±0.3273.8±0.38/69.5±0.8469.9±0.47/
ReMixMatch+DARP+cRT [19, 20]87.3±0.61/87.0±0.1183.5±0.07/83.1±0.0979.7±0.54/
ReMixMatch+CRST+LA [32, 46]84.2±0.11/-81.3±0.34/-79.2±0.31/
ReMixMatch+ABC [27]87.9±0.47/87.6±0.5184.5±0.32/84.1±0.3680.5±1.18/
ReMixMatch+CoSSL [11]87.7±0.21/87.6±0.2584.1±0.56/83.7±0.6681.3±0.83/
ReMixMatch+SAW+cRT [19, 23]87.6±0.21/87.4±0.2685.4±0.32/83.9±0.2179.9±0.15/
ReMixMatch+RECD88.1±0.13/87.9±0.1285.4±0.40/85.2±0.4182.5±0.33/
+ +these results also demonstrate the possibility that RECD can outperform the baseline algorithms even when labeled samples are extremely scarce. + +Tab. 3 summarizes the performances of the baseline CISSL algorithms and RECD on Small-ImageNet-127. For both $32 \times 32$ and $64 \times 64$ version, RECD significantly outperformed the baseline algorithms. Given that the class distribution of the test set of Small-ImageNet-127 is also imbalanced like the training set, these results show that RECD may also be suitable for datasets with imbalanced test sets. Moreover, given that Small-ImageNet-127 is an large-scale dataset, these results verify that RECD can also be applied to large-scale datasets. + +Tab. 4 summarizes the performances of the baseline algorithms and RECD evaluated on CIFAR-10-LT under $\gamma_l \neq \gamma_u$ and STL-10-LT under unknown + +Table 2: Comparison of bACC on CIFAR-100-LT under $\gamma = {\gamma }_{l} = {\gamma }_{u}$ + +
CIFAR-100-LT (γ = γt = γu)
Algorithmγ = 20γ = 50
FixMatch [38]49.6±0.7842.1±0.33
FixMatch+DARP [20]50.8±0.7743.1±0.54
FixMatch+DARP+cRT [19, 20]51.4±0.6844.9±0.54
FixMatch+CReST [46]51.8±0.1244.9±0.50
FixMatch+CReST+LA [32, 46]52.9±0.0747.3±0.17
FixMatch+ABC [27]53.3±0.7946.7±0.26
FixMatch+CoSSL [11]53.9±0.7847.6±0.57
FixMatch+RECD54.6±0.3647.8±0.17
ReMixMatch [1]51.6±0.4344.2±0.59
ReMixMatch+DARP [20]51.9±0.3544.7±0.66
ReMixMatch+DARP+cRT [19, 20]54.5±0.4248.5±0.91
ReMixMatch+CReST [46]51.3±0.3445.5±0.76
ReMixMatch+CReST+LA [32, 46]51.9±0.6046.6±1.14
ReMixMatch+ABC [27]55.6±0.3547.9±0.10
ReMixMatch+CoSSL [11]55.8±0.6248.9±0.61
ReMixMatch+RECD55.9±0.3649.5±0.21
+ +Table 3: Comparison of bACC on Small-ImageNet-127 + +
Small-ImageNet-127 (γ = γl = γu)
Algorithm32 × 3264 × 64
FixMatch [38]29.742.3
FixMatch+DARP [20]30.542.5
FixMatch+DARP +cRT [19,20]39.751.0
FixMatch+CReST+ [46]32.544.7
FixMatch+CReST++LA [32,46]40.955.9
FixMatch+CoSSL [11]43.753.8
FixMatch+RECD47.359.5
+ +$\gamma_{u}$ . RECD outperformed the baseline algorithms by large margins. These results demonstrate that RECD effectively mitigates class imbalance even when $\gamma_{l} \neq \gamma_{u}$ , by considering the unknown class distribution of the unlabeled set. Notably, ReMixMatch+RECD outperformed the baseline algorithms that use ReMixMatch* [20] (ReMixMatch that uses the estimated class distribution of the unlabeled set) as the backbone SSL algorithm. These results implicitly demonstrate that RECD estimates class distribution better than the previous estimation algorithm. Additionally, under severe class distribution mismatch between the labeled and unlabeled sets, LA degraded the performance of ReMixMatch*+DARP and ReMixMatch*+SAW. This may be because the unknown class distribution of the unlabeled set cannot be incorporated into LA, emphasizing the importance of considering the class distribution of the unlabeled set. + +Table 4: Comparison of bACC/GM on CIFAR-10-LT and STL-10-LT under $\gamma_l \neq \gamma_u$ . + +
CIFAR-10-LT (γl=100)STL-10-LT (γu=Unknown)
Algorithmγu=1γu=50γu=150γl=10γl=20
FixMatch [38]68.9±1.95/42.8±8.1173.9±0.25/70.5±0.5269.6±0.60/62.6±1.1172.9±0.09/69.6±0.0163.4±0.21/52.6±0.09
+DARP [20]85.4±0.5585.0±0.6577.3±0.17/75.5±0.2172.9±0.24/69.5±0.1877.8±0.33/76.5±0.4069.9±1.77/65.4±0.97
+DARP+LA [20,32]86.6±1.11/86.2±1.1582.3±0.32/81.5±0.2978.9±0.23/77.7±0.0678.6±0.30/77.4±0.4071.9±0.49/68.7±0.51
+DARP+cRT [19,20]87.0±0.70/86.8±0.6782.7±0.21/82.3±0.2580.7±0.44/80.2±0.6179.3±0.23/78.7±0.2174.1±0.61/73.1±1.21
+ABC [27]82.7±0.49/81.9±0.6882.7±0.64/82.0±0.7678.4±0.87/77.2±1.0779.1±0.46/78.1±0.5773.8±0.15/72.1±0.15
+SAW [23]81.2±0.68/80.2±0.9179.8±0.25/79.1±0.3274.5±0.97/72.5±1.3778.3±0.25/77.0±0.1971.9±0.81/69.0±0.81
+SAW+LA [23,32]84.5±0.68/84.1±0.7882.9±0.38/82.6±0.3879.1±0.81/78.6±0.91-/--/-
+SAW+cRT [19,23]84.6±0.23/84.4±0.2681.6±0.38/81.3±0.3277.6±0.40/77.1±0.41-/--/-
+RECD90.2±0.57/90.0±0.6485.6±0.15/85.3±0.1582.3±0.39/81.8±0.4581.4±0.28/80.6±0.3879.0±0.48/78.1±0.54
ReMixMatch [1]48.3±0.14/19.5±0.8575.1±0.43/71.9±0.7772.5±0.10/68.2±0.3267.8±0.45/61.1±0.9260.1±1.18/44.9±1.52
+ABC [27]76.4±5.34/74.8±6.0585.2±0.20/84.7±0.2580.4±0.40/80.0±0.4476.8±0.52/74.8±0.6471.2±1.37/67.4±1.89
ReMixMatch* [1]85.0±1.35/84.3±1.5577.0±0.12/74.7±0.0472.8±0.10/68.8±0.2176.7±0.15/73.9±0.3267.7±0.46/60.3±0.76
+DARP [20]86.9±0.10/86.4±0.1577.4±0.22/75.0±0.2573.2±0.11/69.2±0.3179.4±0.07/78.2±0.1070.9±0.44/67.0±1.62
+DARP+LA [20,32]81.8±0.45/80.9±0.4083.9±0.42/83.4±0.4581.1±0.20/80.3±0.2680.6±0.45/79.6±0.5576.8±0.60/74.8±0.68
+DARP+cRT [19,20]88.7±0.25/88.5±0.2583.5±0.53/83.1±0.5180.9±0.25/80.3±0.3180.9±0.53/80.0±0.4676.7±0.50/74.9±0.70
+SAW [23]87.0±0.75/86.4±0.8580.6±1.57/79.2±1.1977.6±0.76/76.0±0.9382.0±0.55/81.0±0.6479.2±0.44/77.9±0.52
+SAW+LA [23,32]74.2±1.49/65.1±2.3684.8±1.07/82.4±2.3281.3±2.42/80.9±2.47-/--/-
+SAW+cRT [19,23]88.8±0.79/88.6±0.8384.5±0.78/83.6±1.2782.4±0.10/82.0±0.10-/--/-
ReMixMatch+RECD90.3±0.40/90.2±0.4186.8±0.17/86.6±0.1883.9±0.11/83.7±0.1284.9±0.41/84.4±0.4782.5±0.27/81.7±0.31
+ +In addition to the above results, we demonstrate the effectiveness of RECD with fewer labeled samples in Appendix J. Furthermore, we verify that RECD can be effectively combined with recent SSL algorithms by conducting experiments using FreeMatch [45] as the backbone SSL algorithm in Appendix I. + +# 5.3 Detailed analyses on the proposed algorithm + +We argued that RECD effectively estimates the unknown class distribution of the unlabeled set and mitigates class imbalance using this estimation. To verify this argument, we plotted the true class distribution of the unlabeled set, the class distribution of the unlabeled set estimated by RECD, and the class distribution of the labeled set in Fig. 3. We can observe that RECD accurately estimated the unknown class distribution of the unlabeled set, even when it significantly differed from that of the labeled set. This accurate class distribution estimation enabled the proposed algorithm to be trained with a properly rebalanced training loss and accurately calculated AAFM, mitigating class imbalance as in Fig. 4. + +![](images/a07408fa87cca9fa5fefce2e1db239452512b8ad72d069f97c1dfe26d9e9df3f.jpg) +(a) FixMatch+RECD + +![](images/72704aba64fa1b059fe50bbd0905d4c3b7df3f12579516ed56d8f36c223244ab.jpg) +(b) RemixMatch+RECD +Fig. 3: True class distributions of the labeled and unlabeled sets, along with the class distribution of the unlabeled set estimated by RECD. FixMatch+RECD and Remix-Match+RECD were trained on CIFAR-10-LT with $\gamma_{l} = 100$ and $\gamma_{u} = 1$ . + +Fig. 4 presents the confusion matrices of class predictions on the test set of CIFAR-10 generated by each algorithm trained on the training set of CIFAR-10-LT under $\gamma_l = 100$ and $\gamma_u = 1$ . The $(i,j)$ position of each confusion matrix represents the proportion of the $i$ th class samples classified as the $j$ th class. In the figures, the left three algorithms often misclassified minority classes as majority classes. On the other hand, with both RECD and AAFM, the algorithm classified most samples of minority classes accurately. These results demonstrate that both RECD and AAFM effectively mitigate class imbalance when the class distribution of the labeled and unlabeled sets significantly mismatch. + +Additionally, we argued that AAFM not only reduces the test losses for minority classes but also stabilizes the training process. To verify this, we measured the overall test loss and the test losses of minority classes for FixMatch+RECD without AAFM and FixMatch+RECD with AAFM trained on CIFAR-10-LT. From Fig. 5, we can observe that the test losses for minority classes tended to increase as the training proceeds when AAFM was not used. On the other hand, the overall test loss and the test losses for the minority classes consistently decreased when AAFM was used. We can also observe that whereas the test losses for minority classes fluctuated when AAFM was not used, they stably decreased when AAFM was used. In addition, we compared the performance of the naive feature multiplier and AAFM with FixMatch+RECD in Appendix K. Finally, + +![](images/f7455dbc685d9de10c884e4c212d2a1fd40ccaa9fb8896a8d37d959170fd0d6c.jpg) +(a) FixMatch + +![](images/0bb3d89f794a8e74e448bab48f42a7a0336aa390bf8b71a5a020a3af1ef82366.jpg) +(b) Without RECD + +![](images/f0fee412f9bae0708ed2d5bfd441210ae21eb645a618d3edc9f116ffc1742a29.jpg) +(c) Without AAFM +Fig. 4: Confusion matrices of class predictions on the test set of CIFAR-10 generated by each algorithm (FixMatch+RECD with the described condition) trained on the training set of CIFAR-10-LT under $\gamma_l = 100$ and $\gamma_u = 1$ . + +![](images/a61fb43238383e5559756ccb71379879588fbe06a6b2f25a11640c80cf14345e.jpg) +(d) $\mathrm{RECD + AAFM}$ + +we argue that RECD also can be combined with various CISSL algorithms. To verify this, we conducted additional experiments using DARP, and demonstrated RECD-based DARP outperformed the original DARP in Appendix F. + +![](images/f3c70fd891ef136643eb01f0e2478815210e2f0d919dce16685c4262a38ba34e.jpg) +(a) WithoutAAFM + +![](images/9a34d14dfc37409986f4662696f612674b20f6cb24f466f44396d2a0f19842f8.jpg) +(b) WithAAFM +Fig.5: Overall test loss and test losses for minority classes (Classes 8 and 9). The algorithms were trained on CIFAR-10-LT under $\gamma_{l} = 100$ and $\gamma_{u} = 1$ . + +# 6 Conclusion + +We proposed a novel CISSL algorithm, RECD, which addresses the class distribution mismatch between labeled and unlabeled sets. RECD estimates the class distribution of the unlabeled set through MC approximation and uses it for rebalancing the classifier. Additionally, RECD reuses the estimated class distribution for AAFM to mitigate class imbalance in the feature map by densifying feature clusters of minority classes. Experimental results demonstrate that RECD outperforms the baseline algorithms under both scenarios where the class distributions of the labeled and unlabeled sets match and mismatch. Detailed analyses on RECD verify that the effectiveness of each component of RECD for mitigating class imbalance. However, the estimated class distribution of unlabeled set through MC approximation might be inaccurate when unlabeled training samples are scarce. In future work, we aim to improve the robustness of class distribution estimation, even with a limited amount of unlabeled samples. + +# Acknowledgements + +This research was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (2023R1A2C2005453, RS-2023-00218913). + +# References + +1. Berthelot, D., Carlini, N., an Alex Kurakin, E.D.C., Sohn, K., Zhang, H., Raffel, C.: Remixmatch: Semi-supervised learning with distribution matching and augmentation anchoring. In: ICLR (2020) +2. Berthelot, D., Carlini, N., Goodfellow, I., Papernot, N., Oliver, A., Raffel, C.: Mixmatch: A holistic approach to semi-supervised learning. In: NeurIPS (2019) +3. Cao, K., Wei, C., Gaidon, A., Arechiga, N., Ma, T.: Learning imbalanced datasets with label-distribution-aware margin loss. In: Wallach, H., Larochelle, H., Beygelzimer, A., d'Alché-Buc, F., Fox, E., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 32. Curran Associates, Inc. (2019) +4. Chawla, N.V., Bowyer, K.W., Hall, L.O., Kegelmeyer, W.P.: Smote: Synthetic minority over-sampling technique. Journal of Artificial Intelligence Research pp. 321-357 (2002) +5. Chu, P., Bian, X., Liu, S., Ling, H.: Feature space augmentation for long-tailed data. In: ECCV. vol. 12374, pp. 694-710 (2020) +6. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.: Randaugment: Practical automated data augmentation with a reduced search space. In: NeurIPS (2020) +7. Cui, Y., Jia, M., Lin, T.Y., Song, Y., Belongie, S.: Large scale fine-grained categorization and domain-specific transfer learning. In: CVPR. pp. 4109-4118 (2018) +8. Cui, Y., Jia, M., Lin, T.Y., Song, Y., Belongie, S.: Class-balanced loss based on effective number of samples. In: CVPR. pp. 9268-9277 (2019) +9. DeVries, T., Taylor, G.W.: Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552 (2017) +10. Estabrooks, A., Jo, T., Japkowicz, N.: A multiple resampling method for learning from imbalanced data sets. Computational Intelligence 20(1), 18-36 (2004) +1. Fan, Y., Dai, D., Kukleva, A., Schiele, B.: Cossl: Co-learning of representation and classifier for imbalanced semi-supervised learning. In: CVPR. pp. 14574-14584 (2022) +2. Gidaris, S., Singh, P., Komodakis, N.: Unsupervised representation learning by predicting image rotations. In: ICLR (2018) +3. Grandvalet, Y., Bengio, Y.: Semi-supervised learning by entropy minimization. In: NeurIPS (2004) +4. Guo, L.Z., Li, Y.F.: Class-imbalanced semi-supervised learning with adaptive thresholding. In: ICML. vol. 162, pp. 8082-8094 (2022) +5. Huang, C., Li, Y., Loy, C.C., Tang, X.: Learning deep representation for imbalanced classification. In: CVPR. pp. 5375-5384 (2016) +6. JAPKOWICZ, N.: The class imbalance problem: Significance and strategies. In: Proc. 2000 International Conference on Artificial Intelligence. vol. 1, pp. 111-117 (2000) +7. Japkowicz, N.: The class imbalance problem: Significance and strategies. In: In Proceedings of the International Conference on Artificial Intelligence (2000) + +18. Kang, B., Li, Y., Xie, S., Yuan, Z., Feng, J.: Exploring balanced feature spaces for representation learning. In: ICLR (2021) +19. Kang, B., Xie, S., Rohrbach, M., Yan, Z., Gordo, A., Feng, J., Kalantidis, Y.: Decoupling representation and classifier for long-tailed recognition. In: ICLR (2020) +20. Kim, J., Hur, Y., Park, S., Yang, E., Hwang, S., Shin, J.: Distribution aligning refinery of pseudo-label for imbalanced semi-supervised learning. In: NeurIPS (2020) +21. Krizhevsky, A.: Learning multiple layers of features from tiny images. Technical report, Department of Computer Science, University of Toronto (2009) +22. Kubat, M., Matwin, S., et al.: Addressing the curse of imbalanced training sets: one-sided selection. In: Icml. vol. 97, p. 179. Citeseer (1997) +23. Lai, Z., Wang, C., Gunawan, H., Cheung, S.C., Chuah, C.N.: Smoothed adaptive weighting for imbalanced semi-supervised learning: Improve reliability against unknown distribution data. In: ICML. vol. 162, pp. 11828-11843 (2022) +24. Lazarow, J., Sohn, K., Lee, C.Y., Li, C.L., Zhang, Z., Pfister, T.: Unifying distribution alignment as a loss for imbalanced semi-supervised learning. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) pp. 5644-5653 (2023) +25. Lee, D.H.: Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In: In Workshop on challenges in representation learning (ICML) (2013) +26. Lee, H., Kim, H.: Cdmad: Class-distribution-mismatch-aware debiasing for class-imbalanced semi-supervised learning. In: CVPR. pp. 23891-23900 (2024) +27. Lee, H., Shin, S., Kim, H.: Abc: Auxiliary balanced classifier for class-imbalanced semi-supervised learning. In: NeurIPS (2021) +28. Li, J., Meng, Z., Shi, D., Song, R., Diao, X., Wang, J., Xu, H.: Fcc: Feature clusters compression for long-tailed visual recognition. In: CVPR. pp. 24080-24089 (2023) +29. Li, K., Kong, X., Lu, Z., Wenyin, L., Yin, J.: Boosting weighted elm for imbalanced learning. Neurocomputing 128, 15-21 (2014) +30. Liu, X.Y., Wu, J., Zhou, Z.H.: Exploratory undersampling for class-imbalance learning. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics) 39(2), 539-550 (2008) +31. Liu, Z., Miao, Z., Zhan, X., Wang, J., Gong, B., Yu, S.X.: Large-scale long-tailed recognition in an open world. In: CVPR. pp. 2537-2546 (2019) +32. Menon, A.K., Jayasumana, S., Rawat, A.S., Jain, H., Veit, A., Kumar, S.: Long-tail learning via logit adjustment. In: ICLR (2021) +33. Miyato, T., ichi Maeda, S., Koyama, M., Ishii, S.: Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE Transactions on Pattern Analysis and Machine Intelligence 41(8), 1979-1993 (2018) +34. Oh, Y., Kim, D.J., Kweon, I.S.: Daso: Distribution-aware semantics-oriented pseudo-label for imbalanced semi-supervised learning. In: CVPR. pp. 9786-9796 (2022) +35. Ren, J., Yu, C., Sheng, S., Ma, X., Zhao, H., Yi, S., Li, H.: Balanced meta-softmax for long-tailed visual recognition. In: NeurIPS (2020) +36. Ren, M., Zeng, W., Yang, B., Urtasun, R.: Learning to reweight examples for robust deep learning. In: ICML. vol. 80, pp. 4334-4343 (2018) +37. Sajjadi, M., Javanmardi, M., Tasdizen, T.: Regularization with stochastic transformations and perturbations for deep semi-supervised learning. In: NeurIPS (2016) +38. Sohn, K., Berthelot, D., Li, C.L., Zhang, Z., Carlini, N., Cubuk, E.D., Kurakin, A., Zhang, H., Raffel, C.: Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In: NeurIPS (2020) + +39. Verma, V., Kawaguchi, K., Lamb, A., Kannala, J., Solin, A., Bengio, Y., Lopez-Paz, D.: Interpolation consistency training for semi-supervised learning. Neural Networks 145, 90-106 (2022) +40. Wang, R., Jia, X., Wang, Q., Wu, Y., Meng, D.: Imbalanced semi-supervised learning with bias adaptive classifier. In: ICLR (2023) +41. Wang, S., Minku, L.L., Yao, X.: A learning framework for online class imbalance learning. IEEE Symposium on Computational Intelligence and Ensemble Learning (CIEL) pp. 36-45 (2013) +42. Wang, S., Minku, L.L., Yao, X.: Resampling-based ensemble methods for online class imbalance learning. IEEE Transactions on Knowledge and Data Engineering 27(5), 1356-1368 (2015) +43. Wang, X., Lian, L., Miao, Z., Liu, Z., Yu, S.X.: Long-tailed recognition by routing diverse distribution-aware experts. In: ICLR (2021) +44. Wang, X., Wu, Z., Lian, L., Yu, S.X.: Debiased learning from naturally imbalanced pseudo-labels. In: CVPR. pp. 14647-14657 (2022) +45. Wang, Y., Chen, H., Heng, Q., Hou, W., Fan, Y., Wu, Z., Wang, J., Savvides, M., Shinozaki, T., Raj, B., Schiele, B., Xie, X.: Freematch: Self-adaptive thresholding for semi-supervised learning. In: ICLR (2023) +46. Wei, C., Sohn, K., Mellina, C., Yuille, A., Yang, F.: Crest: A class-rebalancing self-training framework for imbalanced semi-supervised learning. In: CVPR. pp. 10857-10866 (2021) +47. Wei, T., Gan, K.: Towards realistic long-tailed semi-supervised learning: Consistency is all you need. In: CVPR. pp. 3469-3478 (2023) +48. Xiang, L., Ding, G., Han, J.: Learning from multiple experts: Self-paced knowledge distillation for long-tailed classification. In: ECCV. pp. 247-263 (2020) +49. Yu, F., Seff, A., Zhang, Y., Song, S., Funkhouser, T., Xiao, J.: Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365 (2015) +50. Yu, Z., Li, Y., Lee, Y.J.: Inpl: Pseudo-labeling the inliers first for imbalanced semi-supervised learning. In: ICLR (2023) +51. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. In: ICLR (2018) +52. Zhang, Y., Hooi, B., Hong, L., Feng, J.: Self-supervised aggregation of diverse experts for test-agnostic long-tailed recognition. In: NeurIPS (2022) +53. Zhong, Z., Cui, J., Liu, S., Jia, J.: Improving calibration for long-tailed recognition. In: CVPR. pp. 16489-16498 (2021) \ No newline at end of file diff --git a/2024/Rebalancing Using Estimated Class Distribution for Imbalanced Semi-Supervised Learning under Class Distribution Mismatch/images.zip b/2024/Rebalancing Using Estimated Class Distribution for Imbalanced Semi-Supervised Learning under Class Distribution Mismatch/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0a1055ebbf53f880f0f7b6ddbe28d6686fe895b1 --- /dev/null +++ b/2024/Rebalancing Using Estimated Class Distribution for Imbalanced Semi-Supervised Learning under Class Distribution Mismatch/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1b7760e558ea749de11395931ee65bc5f4ef8714fc6d80bbb80f1724292e81d5 +size 524559 diff --git a/2024/Rebalancing Using Estimated Class Distribution for Imbalanced Semi-Supervised Learning under Class Distribution Mismatch/layout.json b/2024/Rebalancing Using Estimated Class Distribution for Imbalanced Semi-Supervised Learning under Class Distribution Mismatch/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..da6741521fedafb4651d70978af71712a1290a2f --- /dev/null +++ b/2024/Rebalancing Using Estimated Class Distribution for Imbalanced Semi-Supervised Learning under Class Distribution Mismatch/layout.json @@ -0,0 +1,11597 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 138, + 111, + 477, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 111, + 477, + 163 + ], + "spans": [ + { + "bbox": [ + 138, + 111, + 477, + 163 + ], + "type": "text", + "content": "Rebalancing Using Estimated Class Distribution for Imbalanced Semi-Supervised Learning under Class Distribution Mismatch" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 188, + 186, + 424, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 186, + 424, + 199 + ], + "spans": [ + { + "bbox": [ + 188, + 186, + 424, + 199 + ], + "type": "text", + "content": "Taemin Park\\**, Hyuck Lee\\**, and Heeyoung Kim" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 185, + 208, + 428, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 208, + 428, + 220 + ], + "spans": [ + { + "bbox": [ + 185, + 208, + 428, + 220 + ], + "type": "text", + "content": "Department of Industrial and Systems Engineering, KAIST" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 237, + 220, + 376, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 220, + 376, + 230 + ], + "spans": [ + { + "bbox": [ + 237, + 220, + 376, + 230 + ], + "type": "text", + "content": "Daejeon 34141, Republic of Korea" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 205, + 231, + 408, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 231, + 408, + 241 + ], + "spans": [ + { + "bbox": [ + 205, + 231, + 408, + 241 + ], + "type": "text", + "content": "{ptm1001,dlgur0921,heeyoungkim}@kaist.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 284, + 455, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 284, + 455, + 449 + ], + "spans": [ + { + "bbox": [ + 159, + 284, + 455, + 449 + ], + "type": "text", + "content": "Abstract. Despite significant advancements in class-imbalanced semi-supervised learning (CISSL), many existing algorithms explicitly or implicitly assume that the class distribution of unlabeled data matches that of labeled data. However, when this assumption fails in practice, the classification performance of such algorithms may degrade due to incorrectly assigned weight to each class during training. We propose a novel CISSL algorithm called Rebalancing Using Estimated Class Distribution (RECD). RECD estimates the unknown class distribution of unlabeled data through Monte Carlo approximation, leveraging predicted class probabilities for unlabeled samples, and subsequently rebalances the classifier based on the estimated class distribution. Additionally, we propose an extension of feature clusters compression in the context of CISSL to mitigate feature map imbalance by densifying minority class clusters. Experimental results on four benchmark datasets demonstrate that RECD achieves state-of-the-art classification performance in CISSL." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 459, + 463, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 459, + 463, + 482 + ], + "spans": [ + { + "bbox": [ + 159, + 459, + 463, + 482 + ], + "type": "text", + "content": "Keywords: Class-imbalanced semi-supervised learning " + }, + { + "bbox": [ + 159, + 459, + 463, + 482 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 459, + 463, + 482 + ], + "type": "text", + "content": " Long-tailed learning " + }, + { + "bbox": [ + 159, + 459, + 463, + 482 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 459, + 463, + 482 + ], + "type": "text", + "content": " Auxiliary Balanced Classifier" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 502, + 230, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 230, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 230, + 514 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 528, + 482, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 482, + 600 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 482, + 600 + ], + "type": "text", + "content": "In real-world datasets, the number of samples often varies across different classes. Deep neural network algorithms trained on such datasets tend to exhibit bias towards majority classes, which have a larger number of samples in the dataset [15, 31, 41]. To mitigate the bias, various class-imbalanced learning (CIL) techniques have been proposed, including re-sampling [4, 10, 30, 42], re-weighting [8, 29, 36], decoupled learning [19, 35, 53], and multi-expert learning [43, 48, 52]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 600, + 482, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 600, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 600, + 482, + 636 + ], + "type": "text", + "content": "Although the CIL techniques significantly mitigate the bias, they typically assume that the entire training set is labeled. However, in reality, labeling samples requires substantial resources, resulting in many real-world datasets containing" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 643, + 229, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 643, + 229, + 654 + ], + "spans": [ + { + "bbox": [ + 141, + 643, + 229, + 654 + ], + "type": "text", + "content": "* Equal contribution" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 654, + 240, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 240, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 240, + 665 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "a significant proportion of unlabeled data. Because the CIL techniques assign importance to each data point based on its corresponding label, they cannot be directly applied when the training set includes unlabeled samples." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 282 + ], + "type": "text", + "content": "Recently, several class-imbalanced semi-supervised learning (CISSL) algorithms, including ABC [27], UDAL [24], CReST [46], DARP [20], DASO [34], CoSSL [11], and SAW [23], have emerged to address the bias towards majority classes in scenarios where the training set consists of both labeled and unlabeled data. However, ABC, UDAL, and CReST explicitly assume that the class distribution of the unlabeled data matches that of the labeled data. Similarly, DARP, DASO, and SAW can be seen as implicitly assuming the class distribution match between the labeled and unlabeled data, as they incorporate CIL techniques such as LA [32] and cRT [19], which rely on the same class distribution between the labeled and unlabeled data. Additionally, CoSSL solely considers the class distribution of labeled data in its main component, TFE." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 283, + 481, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 481, + 415 + ], + "type": "text", + "content": "However, the class distribution of the unlabeled set is often unknown and may differ from that of the labeled set [34, 40, 47]. In this case, the classification performance of the aforementioned algorithms may degrade because the unknown class distribution of the unlabeled set is not considered in the rebalancing process. Fig. 1b illustrates the class predictions on the test set of CIFAR-10 [21] using ReMixMatch+ABC, an existing CISSL algorithm, trained on the long-tailed version of the CIFAR-10 [8] training set where the class distributions of labeled and unlabeled sets mismatch, as presented in Fig. 1a. From Fig. 1b, we can observe that many test samples are mispredicted as Class 10 (the most minor class). This may be because ABC rebalances the classifier excessively, considering only the class distribution of the labeled set." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 133, + 433, + 239, + 506 + ], + "blocks": [ + { + "bbox": [ + 133, + 433, + 239, + 506 + ], + "lines": [ + { + "bbox": [ + 133, + 433, + 239, + 506 + ], + "spans": [ + { + "bbox": [ + 133, + 433, + 239, + 506 + ], + "type": "image", + "image_path": "e1202bd77e9c19ecee6374f042f9c2d15ec144c619b9484c1d49e46de9ddfeb2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 507, + 227, + 517 + ], + "lines": [ + { + "bbox": [ + 145, + 507, + 227, + 517 + ], + "spans": [ + { + "bbox": [ + 145, + 507, + 227, + 517 + ], + "type": "text", + "content": "(a) Class distributions" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 254, + 433, + 359, + 506 + ], + "blocks": [ + { + "bbox": [ + 254, + 433, + 359, + 506 + ], + "lines": [ + { + "bbox": [ + 254, + 433, + 359, + 506 + ], + "spans": [ + { + "bbox": [ + 254, + 433, + 359, + 506 + ], + "type": "image", + "image_path": "c84920ac757b6feb2e5d7b92219ee731492bdb788be55a9fb45bd01ae22453f0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 265, + 507, + 349, + 517 + ], + "lines": [ + { + "bbox": [ + 265, + 507, + 349, + 517 + ], + "spans": [ + { + "bbox": [ + 265, + 507, + 349, + 517 + ], + "type": "text", + "content": "(b) ReMixMatch+ABC" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 525, + 482, + 560 + ], + "lines": [ + { + "bbox": [ + 130, + 525, + 482, + 560 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 482, + 560 + ], + "type": "text", + "content": "Fig. 1: Comparison of the predictions of ReMixMatch+ABC and the proposed algorithm (ReMixMatch+RECD) trained on an imbalanced version of CIFAR-10 where the class distributions of labeled and unlabeled sets mismatch, as illustrated in Fig. 1a." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 375, + 434, + 480, + 506 + ], + "blocks": [ + { + "bbox": [ + 375, + 434, + 480, + 506 + ], + "lines": [ + { + "bbox": [ + 375, + 434, + 480, + 506 + ], + "spans": [ + { + "bbox": [ + 375, + 434, + 480, + 506 + ], + "type": "image", + "image_path": "8c66110307eb62a2654e3e21c59902955c2a9699d7b114a1ce8d60cf5a1659c9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 507, + 470, + 517 + ], + "lines": [ + { + "bbox": [ + 386, + 507, + 470, + 517 + ], + "spans": [ + { + "bbox": [ + 386, + 507, + 470, + 517 + ], + "type": "text", + "content": "(c) Proposed algorithm" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "To rebalance the classifier appropriately when the class distribution of the unlabeled set mismatches that of the labeled set, we propose a novel CISSL algorithm called Rebalancing Using Estimated Class Distribution (RECD), which estimates the unknown class distribution using Monte Carlo (MC) approximation and rebalances the training loss based on the estimated class distribution. Specifically, RECD is a modular algorithm that operates in conjunction with previous CISSL algorithms. While RECD can generally be integrated with most" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "type": "text", + "content": "Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "content": "CISSL algorithms by providing additional information about the class distribution of the unlabeled set, in this paper, we specifically combine RECD with ABC due to its simple and effective framework compared to other CISSL algorithms." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 481, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 481, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 481, + 295 + ], + "type": "text", + "content": "To estimate the class distribution of unlabeled data, RECD first computes the average class predictions of ABC for the entire unlabeled set. Because generating class predictions for the entire unlabeled set can be computationally intensive, RECD generates class predictions on an unlabeled minibatch for each iteration and updates the exponential moving average (EMA) of the class predictions on the unlabeled set. Utilizing MC approximation, this average class prediction serves as the classifier's prior, representing the estimated unknown class distribution of the unlabeled set. Then, RECD rebalances the training loss on unlabeled samples for ABC based on the estimated class distribution of the unlabeled set, and trains ABC using this rebalanced unlabeled loss in addition to the training loss on labeled samples for ABC. As depicted in Fig. 1c, ReMix-Match+RECD generates nearly balanced class predictions on the test set." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 296, + 481, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 481, + 427 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 481, + 427 + ], + "type": "text", + "content": "In addition to mitigating imbalance in the classifier, RECD also mitigates imbalance in the feature map. While existing CISSL algorithms primarily focus on mitigating classifier bias, there also exists imbalance in the feature map [5,18]. Specifically, feature clusters of minority classes tend to be sparser compared to the dense clusters of majority classes. These sparse features often overlap with clusters of other classes in the feature map, posing challenges in classifying minority class features [28]. To address this issue by densifying the sparse feature clusters of minority classes, we extend feature clusters compression (FCC) [28], a previous CIL technique, into the context of CISSL by leveraging the unknown class distribution of the unlabeled data estimated by RECD. We refer to this extended version of FCC as the adaptively adjusted feature multiplier (AAFM)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 427, + 481, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 427, + 481, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 427, + 481, + 559 + ], + "type": "text", + "content": "Our experimental results on CIFAR-10-LT [8], CIFAR-100-LT [8], STL-10-LT [20], and Small-ImageNet-127 [11] demonstrate the superior performance of the proposed algorithm compared to the baseline algorithms, regardless of whether the class distributions of labeled and unlabeled sets match or mismatch. Through detailed analyses, we verify that RECD effectively estimates the true class distribution of the unlabeled set and uses it to mitigate class imbalance. Additionally, we validate that AAFM not only reduces the test loss for minority classes but also stabilizes the training process. Finally, by conducting experiments that combine RECD with DARP [20], we validate the effectiveness of RECD beyond its application solely with ABC. The source codes for RECD are available at https://github.com/taemin-park/RECD." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 578, + 234, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 234, + 592 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 234, + 592 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "Many CISSL algorithms have been proposed to address class imbalance in scenarios where the training set consists of both labeled and unlabeled data. For example, ABC [27] trains an auxiliary balanced classifier with a training loss that is rebalanced based on the class distribution of the labeled data. CReST [46] systematically enlarges the labeled dataset by assigning the labels to unlabeled" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "RECD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": "samples predicted as minority classes more frequently than those predicted as majority classes. UDAL [24] combines progressive distribution alignment and logit-adjustment [32] to adjust the training loss. ABC, CReST, and UDAL have effectively mitigated bias towards majority classes. However, they assumed that the class distribution of the unlabeled data is the same as that of the labeled data, which may pose practical limitation as described in Sec. 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 480, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 480, + 281 + ], + "type": "text", + "content": "DARP [20] solves a convex optimization problem to refine biased pseudo labels. DASO [34] blends class predictions of a similarity-based classifier and a linear classifier to obtain unbiased pseudo-labels. SAW [23] estimates the learning difficulty of each class and smoothly reweights the training loss based on the estimated learning difficulty. However, DARP, DASO, and SAW implicitly assume the same class distributions of the unlabeled and labeled datasets, as they incorporate CIL techniques such as LA [32] and cRT [19], which rebalance the classifier based on the class distribution of the labeled data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 283, + 492, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 492, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 492, + 451 + ], + "type": "text", + "content": "CoSSL [11] integrates Tail-class Feature Enhancement (TFE) and class-balanced sampling to rebalance the classifier. InPL [50] introduces energy-based pseudolabeling to effectively discern correctly predicted unlabeled samples of minority classes. DePL [44] mitigates pseudo-label bias using counterfactual reasoning and adaptive marginal loss. L2AC [40] mitigates training bias by using a bias adaptive classifier, which consists of a bias attractor and a linear classifier. Adsh [14] sets an adaptive confidence threshold for each class to utilize unlabeled samples predicted as minority classes more frequently than those predicted as majority classes. ACR [47] estimates the class distribution of the unlabeled set by comparing class predictions with class distribution prototypes, and then adaptively adjusts the training loss based on the estimated class distribution. CDMAD [26] refines the predictions for unlabeled and test samples based on the logits for solid color image. Although these algorithms have significantly improved CISSL, this area remains relatively less explored compared to SSL and CIL." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 468, + 234, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 234, + 480 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 234, + 480 + ], + "type": "text", + "content": "3 Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 491, + 246, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 246, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 246, + 503 + ], + "type": "text", + "content": "3.1 Problem settings" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "content": "We have a training dataset that consists of the labeled set " + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\{(x_n, y_n) : n \\in (1, \\ldots, N)\\}" + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "content": " and unlabeled set " + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{U} = \\{(u_m) : m \\in (1, \\ldots, M)\\}" + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "x_n \\in R^d" + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "y_n \\in R^K" + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "content": " are the " + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "content": "th labeled sample and its corresponding label, respectively, and " + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "u_m \\in R^d" + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "content": "th unlabeled sample. For each class " + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "N_k" + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "M_k" + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "content": " denote the numbers of the labeled and unlabeled samples, respectively, i.e., " + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sum N_k = N" + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sum M_k = M" + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "content": ". In CISSL, it is typically assumed that " + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "N_1 \\geq N_2 \\geq \\ldots \\geq N_K" + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "M_1 \\geq M_2 \\geq \\ldots \\geq M_K" + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "content": ". We denote the class imbalance ratios of the labeled and unlabeled sets as " + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\gamma_l = \\frac{N_1}{N_K}" + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\gamma_u = \\frac{M_1}{M_K}" + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "content": ", respectively. When " + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "M_k" + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "content": " is unknown, " + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\gamma_u" + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "content": " would also be unknown, and it may differ from " + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\gamma_l" + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "content": ". For each training iteration, we generate minibatches " + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{MB}_{\\mathcal{X}} = \\{(x_b^m, y_b^m) : b \\in (1, \\ldots, B_l)\\} \\subset \\mathcal{X}" + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{MB}_{\\mathcal{U}} = \\{(u_b^m) : b \\in (1, \\ldots, B_u)\\} \\subset \\mathcal{U}" + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "content": ", respectively, where " + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "B_l" + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "B_u" + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "content": " are the sizes of the labeled and unlabeled minibatches, respectively. Using " + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{MB}_{\\mathcal{X}}" + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{MB}_{\\mathcal{U}}" + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "content": ", we aim to train an algorithm " + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "f_\\theta: R^d \\to R^K" + }, + { + "bbox": [ + 132, + 510, + 482, + 666 + ], + "type": "text", + "content": " that accurately predicts the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "type": "text", + "content": "Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": "class probabilities for unseen samples, where " + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": " denotes the network parameters. We denote the extracted features before the classification layer as " + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\xi (\\cdot)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 158, + 285, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 158, + 285, + 170 + ], + "spans": [ + { + "bbox": [ + 131, + 158, + 285, + 170 + ], + "type": "text", + "content": "3.2 Backbone SSL algorithm" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 177, + 480, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 177, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 177, + 480, + 237 + ], + "type": "text", + "content": "Following previous CISSL algorithms [11, 20, 23, 27, 34, 50], either FixMatch [38] or ReMixMatch [1] is used as a backbone SSL algorithm for RECD. To utilize unlabeled samples, FixMatch and ReMixMatch conduct consistency regularization using weak data augmentation techniques " + }, + { + "bbox": [ + 130, + 177, + 480, + 237 + ], + "type": "inline_equation", + "content": "\\alpha(\\cdot)" + }, + { + "bbox": [ + 130, + 177, + 480, + 237 + ], + "type": "text", + "content": " (flipping and cropping) and strong data augmentation techniques " + }, + { + "bbox": [ + 130, + 177, + 480, + 237 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(\\cdot)" + }, + { + "bbox": [ + 130, + 177, + 480, + 237 + ], + "type": "text", + "content": " (Cutout [9] and RandomAugment [6])." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "text", + "content": "Specifically, FixMatch first predicts the class probabilities for " + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "inline_equation", + "content": "\\alpha(u_b^m)" + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "inline_equation", + "content": "f(\\alpha(u_b^m))" + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "text", + "content": " and then generates the pseudo-label as " + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "inline_equation", + "content": "q_b^m = \\arg \\max_k (f(\\alpha(u_b^m))_k)" + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "text", + "content": " when the confidence exceeds the confidence threshold " + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "inline_equation", + "content": "\\max_k (f(\\alpha(u_b^m))_k) \\geq \\epsilon" + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "inline_equation", + "content": "(\\cdot)_k" + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "text", + "content": "th element of " + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "text", + "content": ". Then, FixMatch predicts the class probabilities for " + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(u_b^m)" + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "inline_equation", + "content": "f(\\mathcal{A}(u_b^m))" + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "text", + "content": " and ensures " + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "inline_equation", + "content": "f(\\mathcal{A}(u_b^m))" + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "text", + "content": " to be close to the pseudo-label " + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "inline_equation", + "content": "q_b^m" + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "text", + "content": " by minimizing a loss that penalizes the distance between " + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "inline_equation", + "content": "f(\\mathcal{A}(u_b^m))" + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "inline_equation", + "content": "q_b^m" + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "text", + "content": ". In addition, FixMatch minimizes the prediction loss calculated between weakly augmented labeled sample " + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "inline_equation", + "content": "\\alpha(x_b^m)" + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "inline_equation", + "content": "p_b^m" + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "text", + "content": ", which is the one-hot version of " + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "inline_equation", + "content": "y_b^m" + }, + { + "bbox": [ + 130, + 237, + 492, + 333 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 333, + 481, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 481, + 453 + ], + "type": "text", + "content": "Similarly, ReMixMatch first predicts the class probabilities of " + }, + { + "bbox": [ + 130, + 333, + 481, + 453 + ], + "type": "inline_equation", + "content": "\\alpha(u_b^m)" + }, + { + "bbox": [ + 130, + 333, + 481, + 453 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 333, + 481, + 453 + ], + "type": "inline_equation", + "content": "q_b^m = f(\\alpha(u_b^m))" + }, + { + "bbox": [ + 130, + 333, + 481, + 453 + ], + "type": "text", + "content": ". Then ReMixMatch applies a distribution alignment technique as " + }, + { + "bbox": [ + 130, + 333, + 481, + 453 + ], + "type": "inline_equation", + "content": "\\tilde{q}_b^m = \\mathrm{Normalize}(q_b^m \\times q(y) / \\tilde{q}(y))" + }, + { + "bbox": [ + 130, + 333, + 481, + 453 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 333, + 481, + 453 + ], + "type": "inline_equation", + "content": "\\mathrm{Normalize}(x)_i = x_i / \\sum_j x_j" + }, + { + "bbox": [ + 130, + 333, + 481, + 453 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 333, + 481, + 453 + ], + "type": "inline_equation", + "content": "\\tilde{q}(y)" + }, + { + "bbox": [ + 130, + 333, + 481, + 453 + ], + "type": "text", + "content": " is the moving average of class predictions on unlabeled samples, and " + }, + { + "bbox": [ + 130, + 333, + 481, + 453 + ], + "type": "inline_equation", + "content": "q(y)" + }, + { + "bbox": [ + 130, + 333, + 481, + 453 + ], + "type": "text", + "content": " indicates the class distribution of the training set. Finally, ReMixMatch sharpens " + }, + { + "bbox": [ + 130, + 333, + 481, + 453 + ], + "type": "inline_equation", + "content": "\\tilde{q}_b^m" + }, + { + "bbox": [ + 130, + 333, + 481, + 453 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 333, + 481, + 453 + ], + "type": "inline_equation", + "content": "\\bar{q}_b^m = \\mathrm{Normalize}((\\tilde{q}_b^m)^{1/T})" + }, + { + "bbox": [ + 130, + 333, + 481, + 453 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 333, + 481, + 453 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 333, + 481, + 453 + ], + "type": "text", + "content": " is the temperature for sharpening. Using " + }, + { + "bbox": [ + 130, + 333, + 481, + 453 + ], + "type": "inline_equation", + "content": "\\bar{q}_b^m" + }, + { + "bbox": [ + 130, + 333, + 481, + 453 + ], + "type": "text", + "content": " as pseudo-labels, ReMixMatch conducts consistency regularization [33,37], entropy minimization [13,25] and mixup regularization [2,39,51]. ReMixMatch also minimizes the rotation loss [12] for self-supervised learning. The training loss of the backbone SSL algorithm, " + }, + { + "bbox": [ + 130, + 333, + 481, + 453 + ], + "type": "inline_equation", + "content": "L_{back}" + }, + { + "bbox": [ + 130, + 333, + 481, + 453 + ], + "type": "text", + "content": ", is detailed in Appendix B." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 470, + 339, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 470, + 339, + 483 + ], + "spans": [ + { + "bbox": [ + 131, + 470, + 339, + 483 + ], + "type": "text", + "content": "3.3 Auxiliary balanced classifier (ABC)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 491, + 479, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 479, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 479, + 586 + ], + "type": "text", + "content": "To use a balanced classifier trained on a class-balanced subset of the training set, while benefiting from the high-quality representations learned from the entire training set, Lee and Kim [27] introduced an auxiliary balanced classifier (ABC) attached to the representation layer of a backbone SSL algorithm. To ensure that ABC is trained on a class-balanced subset of the training set, ABC generates a " + }, + { + "bbox": [ + 130, + 491, + 479, + 586 + ], + "type": "inline_equation", + "content": "0/1" + }, + { + "bbox": [ + 130, + 491, + 479, + 586 + ], + "type": "text", + "content": " mask " + }, + { + "bbox": [ + 130, + 491, + 479, + 586 + ], + "type": "inline_equation", + "content": "M(\\cdot)" + }, + { + "bbox": [ + 130, + 491, + 479, + 586 + ], + "type": "text", + "content": " for each training sample, where the " + }, + { + "bbox": [ + 130, + 491, + 479, + 586 + ], + "type": "inline_equation", + "content": "0/1" + }, + { + "bbox": [ + 130, + 491, + 479, + 586 + ], + "type": "text", + "content": " mask determines whether each sample contributes to the training loss. Specifically, ABC samples a " + }, + { + "bbox": [ + 130, + 491, + 479, + 586 + ], + "type": "inline_equation", + "content": "0/1" + }, + { + "bbox": [ + 130, + 491, + 479, + 586 + ], + "type": "text", + "content": " mask from a Bernoulli distribution for each labeled and unlabeled sample as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 205, + 594, + 480, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 594, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 205, + 594, + 480, + 622 + ], + "type": "interline_equation", + "content": "M (x _ {b} ^ {m}) = \\mathcal {B} (\\frac {N _ {K}}{N _ {\\hat {y} _ {b} ^ {m}}}) \\quad \\mathrm {a n d} \\quad M (u _ {b} ^ {m}) = \\mathcal {B} (\\frac {N _ {K}}{N _ {\\hat {q} _ {b} ^ {m}}}). \\qquad (1)", + "image_path": "5b5f3fc74060223f6b2f49a8a2efd1145361f92cacdc587f50e018f5fac9e90f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 629, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{B}(\\cdot)" + }, + { + "bbox": [ + 131, + 629, + 480, + 666 + ], + "type": "text", + "content": " denotes a Bernoulli distribution and " + }, + { + "bbox": [ + 131, + 629, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\hat{q}_b^m = \\arg \\max (q_b^m)" + }, + { + "bbox": [ + 131, + 629, + 480, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 629, + 480, + 666 + ], + "type": "inline_equation", + "content": "q_{b}^{m}" + }, + { + "bbox": [ + 131, + 629, + 480, + 666 + ], + "type": "text", + "content": " is the class prediction of the ABC for an unlabeled sample " + }, + { + "bbox": [ + 131, + 629, + 480, + 666 + ], + "type": "inline_equation", + "content": "u_{b}^{m}" + }, + { + "bbox": [ + 131, + 629, + 480, + 666 + ], + "type": "text", + "content": ". Note that in Eq. (1), the probability of sampling 1 is adjusted to be inversely proportional to the number" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "text", + "content": "RECD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "of training samples belonging to each class. This sampling strategy ensures that ABC is trained with an equal frequency from each class's data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": "With the " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "0/1" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": " mask, ABC calculates the classification loss, " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "L_{cls}" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": ", between weakly augmented labeled sample " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "\\alpha(x_b^m)" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": " and its corresponding one-hot version label " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "p_b^m" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": ". In addition, ABC also calculates a consistency regularization loss, " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "L_{con}" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": ", between weakly augmented unlabeled sample " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "\\alpha(u_b^m)" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": " and strongly augmented unlabeled sample " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(u_b^m)" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": ", which is minimized to ensure that decision boundaries lie in low-density regions. The training losses " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "L_{cls}" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "L_{con}" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": " are calculated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 209, + 222, + 481, + 253 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 222, + 481, + 253 + ], + "spans": [ + { + "bbox": [ + 209, + 222, + 481, + 253 + ], + "type": "interline_equation", + "content": "L _ {c l s} = \\frac {1}{B _ {l}} \\sum_ {b = 1} ^ {B _ {l}} M \\left(x _ {b} ^ {m}\\right) \\mathcal {H} \\left(p _ {b} ^ {m}, f _ {a b c} (\\xi \\left(\\alpha \\left(x _ {b} ^ {m}\\right)\\right))\\right), \\tag {2}", + "image_path": "2561e46bb49d473abe165037be02b976c38a42440d2325dea1a6826b7f4d0f38.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 162, + 265, + 480, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 265, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 162, + 265, + 480, + 297 + ], + "type": "interline_equation", + "content": "L _ {c o n} = \\frac {1}{B _ {u}} \\sum_ {b = 1} ^ {B _ {u}} \\sum_ {i = 1} ^ {2} M \\left(u _ {b} ^ {m}\\right) \\mathbb {1} \\left(\\max \\left(q _ {b} ^ {m}\\right) \\geq \\epsilon\\right) \\mathcal {H} \\left(q _ {b} ^ {m}, f _ {a b c} \\left(\\xi \\left(\\mathcal {A} _ {i} \\left(u _ {b} ^ {m}\\right)\\right)\\right)\\right), \\tag {3}", + "image_path": "c5f844140627b3ae68cbb96a2fde5ad89f809d8f5f4add11b8651ac182356217.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 301, + 479, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 301, + 479, + 350 + ], + "spans": [ + { + "bbox": [ + 131, + 301, + 479, + 350 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 301, + 479, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 131, + 301, + 479, + 350 + ], + "type": "text", + "content": " is the cross-entropy loss, " + }, + { + "bbox": [ + 131, + 301, + 479, + 350 + ], + "type": "inline_equation", + "content": "q_{b}^{m} = f_{abc}(\\xi (\\alpha (u_{b}^{m})))" + }, + { + "bbox": [ + 131, + 301, + 479, + 350 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 301, + 479, + 350 + ], + "type": "inline_equation", + "content": "f_{abc}" + }, + { + "bbox": [ + 131, + 301, + 479, + 350 + ], + "type": "text", + "content": " denotes the auxiliary classifier attached to the representation layer, and " + }, + { + "bbox": [ + 131, + 301, + 479, + 350 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(\\cdot)" + }, + { + "bbox": [ + 131, + 301, + 479, + 350 + ], + "type": "text", + "content": " denotes an indicator function. With the " + }, + { + "bbox": [ + 131, + 301, + 479, + 350 + ], + "type": "inline_equation", + "content": "L_{cls}" + }, + { + "bbox": [ + 131, + 301, + 479, + 350 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 301, + 479, + 350 + ], + "type": "inline_equation", + "content": "L_{con}" + }, + { + "bbox": [ + 131, + 301, + 479, + 350 + ], + "type": "text", + "content": " and the training loss of backbone SSL algorithm " + }, + { + "bbox": [ + 131, + 301, + 479, + 350 + ], + "type": "inline_equation", + "content": "L_{back}" + }, + { + "bbox": [ + 131, + 301, + 479, + 350 + ], + "type": "text", + "content": ", total training loss of ABC is calculated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 244, + 360, + 480, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 360, + 480, + 372 + ], + "spans": [ + { + "bbox": [ + 244, + 360, + 480, + 372 + ], + "type": "interline_equation", + "content": "L _ {A B C} = L _ {c l s} + L _ {c o n} + L _ {b a c k}. \\tag {4}", + "image_path": "5885e53a5c03c61b49e4be9af4bd70cee31e7362b2c540f9428de624bb8db818.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 380, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 453 + ], + "type": "text", + "content": "ABC [27] effectively mitigated class imbalance when the class distributions of the labeled and unlabeled sets match. However, as we can observe in the second equation of the Eq. (1), because the " + }, + { + "bbox": [ + 130, + 380, + 482, + 453 + ], + "type": "inline_equation", + "content": "0/1" + }, + { + "bbox": [ + 130, + 380, + 482, + 453 + ], + "type": "text", + "content": " masks for unlabeled samples are generated based on the class distribution of the labeled set without considering that of the unlabeled set, " + }, + { + "bbox": [ + 130, + 380, + 482, + 453 + ], + "type": "inline_equation", + "content": "f_{abc}" + }, + { + "bbox": [ + 130, + 380, + 482, + 453 + ], + "type": "text", + "content": " may not be properly rebalanced when the unknown class distribution of the unlabeled set significantly mismatches that of the labeled set." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 470, + 363, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 470, + 363, + 483 + ], + "spans": [ + { + "bbox": [ + 131, + 470, + 363, + 483 + ], + "type": "text", + "content": "3.4 Feature clusters compression (FCC) [28]" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 491, + 479, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 479, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 479, + 538 + ], + "type": "text", + "content": "To separate features extracted from samples of minority classes from those of majority classes by densifying the feature clusters, Li et al. [28] introduced FCC. For a training sample " + }, + { + "bbox": [ + 130, + 491, + 479, + 538 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 491, + 479, + 538 + ], + "type": "text", + "content": " of class " + }, + { + "bbox": [ + 130, + 491, + 479, + 538 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 491, + 479, + 538 + ], + "type": "text", + "content": ", FCC multiplies the extracted feature " + }, + { + "bbox": [ + 130, + 491, + 479, + 538 + ], + "type": "inline_equation", + "content": "\\xi(x)" + }, + { + "bbox": [ + 130, + 491, + 479, + 538 + ], + "type": "text", + "content": " by a scaling factor " + }, + { + "bbox": [ + 130, + 491, + 479, + 538 + ], + "type": "inline_equation", + "content": "\\tau_k" + }, + { + "bbox": [ + 130, + 491, + 479, + 538 + ], + "type": "text", + "content": ", and uses " + }, + { + "bbox": [ + 130, + 491, + 479, + 538 + ], + "type": "inline_equation", + "content": "\\tau_k \\times \\xi(x)" + }, + { + "bbox": [ + 130, + 491, + 479, + 538 + ], + "type": "text", + "content": " as input for the classification layer " + }, + { + "bbox": [ + 130, + 491, + 479, + 538 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 130, + 491, + 479, + 538 + ], + "type": "text", + "content": ", as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 261, + 549, + 480, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 549, + 480, + 561 + ], + "spans": [ + { + "bbox": [ + 261, + 549, + 480, + 561 + ], + "type": "interline_equation", + "content": "f _ {\\theta} (x) = \\psi \\left(\\tau_ {k} \\times \\xi (x)\\right), \\tag {5}", + "image_path": "465c8a2ac47c2421e1b571f4e61af0d9691e3b7fc6e8738711533e80c68345b8.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 571, + 479, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 571, + 479, + 582 + ], + "spans": [ + { + "bbox": [ + 131, + 571, + 479, + 582 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 571, + 479, + 582 + ], + "type": "inline_equation", + "content": "\\tau_{k}" + }, + { + "bbox": [ + 131, + 571, + 479, + 582 + ], + "type": "text", + "content": " decreases from majority to minority classes in a naive manner as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 227, + 591, + 480, + 614 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 591, + 480, + 614 + ], + "spans": [ + { + "bbox": [ + 227, + 591, + 480, + 614 + ], + "type": "interline_equation", + "content": "\\tau_ {k} = 1 + \\eta \\times \\left(1 - \\frac {k}{K}\\right), k = 1, \\dots , K, \\tag {6}", + "image_path": "32738c7959002f00a4bec8e8a671fe910061ae0ef7ba3d3736b7a1b35048c7fa.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "content": " is a scale hyperparameter. Since the features of majority classes are multiplied by relatively large scaling factors compared to those of minority classes, their feature clusters occupy larger regions on the feature map during training. In this context, to prevent minority class features from crossing the decision" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 210, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 210, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 210, + 101 + ], + "type": "text", + "content": "Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "boundary, the feature extractor " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\xi(\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " would be trained to map minority class features into denser regions. Note that during the test phase, the features extracted from test samples are not multiplied by the scaling factor, ensuring that the classifier predicts classes for test samples with densified features. According to Li et al. [28], FCC effectively enhances the separability of feature clusters of minority classes in CIL. However, in CISSL, it may fail to appropriately densify minority class features because the class distribution of the unlabeled set cannot be considered in the scaling factor " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\tau_{k}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ". In this paper, we refer to the scaling factor in Eq. (6) as the naive feature multiplier." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 242, + 233, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 242, + 233, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 242, + 233, + 255 + ], + "type": "text", + "content": "4 Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 267, + 440, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 267, + 440, + 280 + ], + "spans": [ + { + "bbox": [ + 131, + 267, + 440, + 280 + ], + "type": "text", + "content": "4.1 Rebalancing using estimated class distribution (RECD)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 287, + 482, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 287, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 287, + 482, + 370 + ], + "type": "text", + "content": "As described in Sec. 1, RECD incorporates the unknown class distribution of the unlabeled data by rebalancing the classifier based on the estimated class distribution of the unlabeled data. RECD is integrated with ABC [27], which rebalances the training loss using the " + }, + { + "bbox": [ + 130, + 287, + 482, + 370 + ], + "type": "inline_equation", + "content": "0/1" + }, + { + "bbox": [ + 130, + 287, + 482, + 370 + ], + "type": "text", + "content": " mask, as described in Eq. (1). Whereas ABC generates " + }, + { + "bbox": [ + 130, + 287, + 482, + 370 + ], + "type": "inline_equation", + "content": "0/1" + }, + { + "bbox": [ + 130, + 287, + 482, + 370 + ], + "type": "text", + "content": " masks for unlabeled samples based on the class distribution of the labeled set, RECD replaces the parameter of " + }, + { + "bbox": [ + 130, + 287, + 482, + 370 + ], + "type": "inline_equation", + "content": "0/1" + }, + { + "bbox": [ + 130, + 287, + 482, + 370 + ], + "type": "text", + "content": " masks for unlabeled samples using the estimated class distribution of the unlabeled set." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 371, + 482, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 371, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 482, + 424 + ], + "type": "text", + "content": "Specifically, to estimate the unknown class distribution of the unlabeled set, " + }, + { + "bbox": [ + 130, + 371, + 482, + 424 + ], + "type": "inline_equation", + "content": "p(y_{u})" + }, + { + "bbox": [ + 130, + 371, + 482, + 424 + ], + "type": "text", + "content": ", RECD first computes the average class predictions of ABC on the entire unlabeled set as " + }, + { + "bbox": [ + 130, + 371, + 482, + 424 + ], + "type": "inline_equation", + "content": "\\frac{1}{M}\\sum_{m = 1}^{M}p_{\\theta}(y_u|u_m)" + }, + { + "bbox": [ + 130, + 371, + 482, + 424 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 371, + 482, + 424 + ], + "type": "inline_equation", + "content": "p_{\\theta}(y_u|u_m) = f_{abc}(\\xi (\\alpha (u_m)))" + }, + { + "bbox": [ + 130, + 371, + 482, + 424 + ], + "type": "text", + "content": ". Then, using MC approximation, " + }, + { + "bbox": [ + 130, + 371, + 482, + 424 + ], + "type": "inline_equation", + "content": "\\frac{1}{M}\\sum_{m = 1}^{M}p_{\\theta}(y_u|u_m)" + }, + { + "bbox": [ + 130, + 371, + 482, + 424 + ], + "type": "text", + "content": " becomes " + }, + { + "bbox": [ + 130, + 371, + 482, + 424 + ], + "type": "inline_equation", + "content": "p_{\\theta}(y_u)" + }, + { + "bbox": [ + 130, + 371, + 482, + 424 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 203, + 433, + 480, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 433, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 203, + 433, + 480, + 456 + ], + "type": "interline_equation", + "content": "\\frac {1}{M} \\sum p _ {\\theta} \\left(y _ {u} \\mid u _ {m}\\right) \\approx \\int p _ {\\theta} \\left(y _ {u} \\mid u\\right) p (u) d u = p _ {\\theta} \\left(y _ {u}\\right). \\tag {7}", + "image_path": "ebdd3ae167997413afc0de156e3c3abdf85d6676966b248bb68af4b40b82e16f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 464, + 482, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 513 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 513 + ], + "type": "text", + "content": "In Eq. (7), " + }, + { + "bbox": [ + 130, + 464, + 482, + 513 + ], + "type": "inline_equation", + "content": "p_{\\theta}(y_u)" + }, + { + "bbox": [ + 130, + 464, + 482, + 513 + ], + "type": "text", + "content": " can be regarded as the classifier's prior for the unlabeled samples, and we use this classifier's prior as the estimation of the class distribution of the unlabeled set " + }, + { + "bbox": [ + 130, + 464, + 482, + 513 + ], + "type": "inline_equation", + "content": "p(y_u)" + }, + { + "bbox": [ + 130, + 464, + 482, + 513 + ], + "type": "text", + "content": " based on that the classifier's prior is highly correlated with the class distribution, as analyzed in Fig. 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 513, + 482, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 513, + 482, + 537 + ], + "spans": [ + { + "bbox": [ + 130, + 513, + 482, + 537 + ], + "type": "text", + "content": "Using the estimated class distribution " + }, + { + "bbox": [ + 130, + 513, + 482, + 537 + ], + "type": "inline_equation", + "content": "p_{\\theta}(y_u)" + }, + { + "bbox": [ + 130, + 513, + 482, + 537 + ], + "type": "text", + "content": ", RECD modifies the sampling distribution of the " + }, + { + "bbox": [ + 130, + 513, + 482, + 537 + ], + "type": "inline_equation", + "content": "0/1" + }, + { + "bbox": [ + 130, + 513, + 482, + 537 + ], + "type": "text", + "content": " mask for unlabeled samples in Eq. (1) as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 244, + 547, + 480, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 547, + 480, + 572 + ], + "spans": [ + { + "bbox": [ + 244, + 547, + 480, + 572 + ], + "type": "interline_equation", + "content": "M \\left(u _ {b} ^ {m}\\right) = \\mathcal {B} \\left(\\frac {\\operatorname* {m i n} \\left(p _ {\\theta} \\left(y _ {u}\\right)\\right)}{p _ {\\theta} \\left(y _ {u}\\right) \\hat {q} _ {b} ^ {m}}\\right). \\tag {8}", + "image_path": "4fcc45fbe0962e07a9d5e697c1f3a7431d6157fefe545f4418d3d0a0a4ffbcce.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 581, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 628 + ], + "type": "text", + "content": "By setting the parameter of the " + }, + { + "bbox": [ + 130, + 581, + 482, + 628 + ], + "type": "inline_equation", + "content": "0/1" + }, + { + "bbox": [ + 130, + 581, + 482, + 628 + ], + "type": "text", + "content": " mask based on " + }, + { + "bbox": [ + 130, + 581, + 482, + 628 + ], + "type": "inline_equation", + "content": "p_{\\theta}(y_u)" + }, + { + "bbox": [ + 130, + 581, + 482, + 628 + ], + "type": "text", + "content": ", RECD rebalances the training loss of the unlabeled samples to an appropriate extent even when the class distributions of the labeled and unlabeled sets severely mismatch. The total training loss of RECD is described in detail in Sec. 4.3." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "RECD may increase computational complexity compared to the original ABC because, as shown in Eq. (7), computing class predictions of ABC for the entire unlabeled set for each iteration or epoch would require a substantial amount of" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "RECD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "computation, especially for large-scale datasets such as LSUN [49] or iNaturallist [7]. To address this issue, RECD replaces the naive average " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\frac{1}{M}\\sum p_{\\theta}(y_u|u_m)" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " with an EMA of the class predictions on the unlabeled samples, which is updated for each iteration of the training as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 198, + 175, + 481, + 189 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 175, + 481, + 189 + ], + "spans": [ + { + "bbox": [ + 198, + 175, + 481, + 189 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(y _ {u}\\right) _ {n e w} = \\rho \\times p _ {\\theta} \\left(y _ {u}\\right) _ {o l d} + (1 - \\rho) \\times p _ {\\theta} \\left(y _ {u}\\right) _ {b a t c h}, \\tag {9}", + "image_path": "96ce3ef13dae63e1de136a3a2b8f8264a7dcf0360467e1628ae63d9218f7dd4a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 198, + 482, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 198, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 482, + 270 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 198, + 482, + 270 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 130, + 198, + 482, + 270 + ], + "type": "text", + "content": " is a hyperparameter for the EMA update, " + }, + { + "bbox": [ + 130, + 198, + 482, + 270 + ], + "type": "inline_equation", + "content": "p_{\\theta}(y_u)_{old}" + }, + { + "bbox": [ + 130, + 198, + 482, + 270 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 130, + 198, + 482, + 270 + ], + "type": "inline_equation", + "content": "p_{\\theta}(y_u)" + }, + { + "bbox": [ + 130, + 198, + 482, + 270 + ], + "type": "text", + "content": " before the update, and " + }, + { + "bbox": [ + 130, + 198, + 482, + 270 + ], + "type": "inline_equation", + "content": "p_{\\theta}(y_u)_{batch}" + }, + { + "bbox": [ + 130, + 198, + 482, + 270 + ], + "type": "text", + "content": " is the average class prediction on the weakly augmented unlabeled minibatch, " + }, + { + "bbox": [ + 130, + 198, + 482, + 270 + ], + "type": "inline_equation", + "content": "\\frac{1}{B_u}\\sum_{b = 1}^{B_u}q_b^m = \\frac{1}{B_u}\\sum_{b = 1}^{B_u}f_{abc}(\\xi (\\alpha (u_b^m)))" + }, + { + "bbox": [ + 130, + 198, + 482, + 270 + ], + "type": "text", + "content": ". Updating the EMA of the class predictions on unlabeled samples adds negligible computational cost compared to the original ABC because RECD reuses " + }, + { + "bbox": [ + 130, + 198, + 482, + 270 + ], + "type": "inline_equation", + "content": "q_b^m" + }, + { + "bbox": [ + 130, + 198, + 482, + 270 + ], + "type": "text", + "content": " that is calculated to conduct consistency regularization for ABC, as in Eq. (3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 289, + 402, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 289, + 402, + 303 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 402, + 303 + ], + "type": "text", + "content": "4.2 Adaptively adjusted feature multiplier (AAFM)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 312, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 445 + ], + "type": "text", + "content": "Although the naive feature multiplier effectively densifies feature clusters of the minority classes in CIL, its efficacy may diminish in CISSL due to its lack of consideration for the class distribution of the unlabeled set. In addition, the naive feature multiplier linearly reduces from majority to minority classes as shown in Eq. (6), without considering the degree of class imbalance in the training set. These limitations may lead to inappropriate scaling factors for the extracted features, resulting in insufficient densification or unstable training, as discussed in Sec. 5.3. To address this issue by incorporating both the class imbalance ratio of the labeled set and the unknown class distribution of the unlabeled set, we introduce an advanced version of the naive feature multiplier, AAFM. For the labeled samples, AAFM for the " + }, + { + "bbox": [ + 130, + 312, + 482, + 445 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 312, + 482, + 445 + ], + "type": "text", + "content": "th class, " + }, + { + "bbox": [ + 130, + 312, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\tau_k^l" + }, + { + "bbox": [ + 130, + 312, + 482, + 445 + ], + "type": "text", + "content": ", is calculated as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 220, + 453, + 481, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 453, + 481, + 478 + ], + "spans": [ + { + "bbox": [ + 220, + 453, + 481, + 478 + ], + "type": "interline_equation", + "content": "\\tau_ {k} ^ {l} = 1 + \\eta \\times \\left(\\frac {N _ {k}}{N _ {1}} \\times C _ {l k} ^ {t}\\right), k = 1, \\dots , K, \\tag {10}", + "image_path": "8335bc5d6b24110c053d0a18843cef37425bab134b0dc877ca191e8cac579187.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 487, + 482, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 522 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 487, + 482, + 522 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 487, + 482, + 522 + ], + "type": "text", + "content": " is a scale hyperparameter similar to " + }, + { + "bbox": [ + 130, + 487, + 482, + 522 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 487, + 482, + 522 + ], + "type": "text", + "content": " in Eq. (6), and " + }, + { + "bbox": [ + 130, + 487, + 482, + 522 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 487, + 482, + 522 + ], + "type": "text", + "content": " is a hyperparameter for " + }, + { + "bbox": [ + 130, + 487, + 482, + 522 + ], + "type": "inline_equation", + "content": "C_{lk}" + }, + { + "bbox": [ + 130, + 487, + 482, + 522 + ], + "type": "text", + "content": ", which represents the EMA of confidence for labeled samples of the " + }, + { + "bbox": [ + 130, + 487, + 482, + 522 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 487, + 482, + 522 + ], + "type": "text", + "content": "th class, which is updated for each iteration of the training process as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 231, + 532, + 481, + 547 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 532, + 481, + 547 + ], + "spans": [ + { + "bbox": [ + 231, + 532, + 481, + 547 + ], + "type": "interline_equation", + "content": "C _ {l k} = \\omega \\times C _ {l k} ^ {\\text {o l d}} + (1 - \\omega) \\times C _ {l k} ^ {\\text {b a t c h}}, \\tag {11}", + "image_path": "3297f2ac5e902115cd376f2b34bfaa888b6209f15d663191117718e7b49baad7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 555, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 482, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 555, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 555, + 482, + 640 + ], + "type": "text", + "content": " is a hyperparameter for the EMA update, " + }, + { + "bbox": [ + 130, + 555, + 482, + 640 + ], + "type": "inline_equation", + "content": "C_{lk}^{old}" + }, + { + "bbox": [ + 130, + 555, + 482, + 640 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 130, + 555, + 482, + 640 + ], + "type": "inline_equation", + "content": "C_{lk}" + }, + { + "bbox": [ + 130, + 555, + 482, + 640 + ], + "type": "text", + "content": " before the update, and " + }, + { + "bbox": [ + 130, + 555, + 482, + 640 + ], + "type": "inline_equation", + "content": "C_{lk}^{batch}" + }, + { + "bbox": [ + 130, + 555, + 482, + 640 + ], + "type": "text", + "content": " is the average confidence for class " + }, + { + "bbox": [ + 130, + 555, + 482, + 640 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 555, + 482, + 640 + ], + "type": "text", + "content": " on a minibatch, which is calculated as " + }, + { + "bbox": [ + 130, + 555, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\frac{1}{n_k}\\sum_{b = 1}^{B_l}f_{abc}(\\xi (\\alpha (x_b^m)))_k\\times \\mathbb{1}(y_b^m = k)" + }, + { + "bbox": [ + 130, + 555, + 482, + 640 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 555, + 482, + 640 + ], + "type": "inline_equation", + "content": "n_k" + }, + { + "bbox": [ + 130, + 555, + 482, + 640 + ], + "type": "text", + "content": " denotes the number of samples for the " + }, + { + "bbox": [ + 130, + 555, + 482, + 640 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 555, + 482, + 640 + ], + "type": "text", + "content": "th class on a minibatch. Note that " + }, + { + "bbox": [ + 130, + 555, + 482, + 640 + ], + "type": "inline_equation", + "content": "C_{lk}" + }, + { + "bbox": [ + 130, + 555, + 482, + 640 + ], + "type": "text", + "content": " is updated only when " + }, + { + "bbox": [ + 130, + 555, + 482, + 640 + ], + "type": "inline_equation", + "content": "n_k\\geq 1" + }, + { + "bbox": [ + 130, + 555, + 482, + 640 + ], + "type": "text", + "content": ". By multiplying " + }, + { + "bbox": [ + 130, + 555, + 482, + 640 + ], + "type": "inline_equation", + "content": "C_{lk}^{t}" + }, + { + "bbox": [ + 130, + 555, + 482, + 640 + ], + "type": "text", + "content": " to the second term on the right-hand side of Eq. (10), we stabilize the training process, assigning smaller AAFMs to classes whose feature clusters are located near the decision boundary." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "We also multiply AAFM " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tau_{k}^{u}" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " with the features extracted from strongly augmented unlabeled samples, where " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tau_{k}^{u}" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " incorporates the unknown class distribution" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 210, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 210, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 210, + 101 + ], + "type": "text", + "content": "Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 128 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 128 + ], + "type": "text", + "content": "of the unlabeled set by reusing the estimated class distribution " + }, + { + "bbox": [ + 131, + 116, + 479, + 128 + ], + "type": "inline_equation", + "content": "p_{\\theta}(y_u)" + }, + { + "bbox": [ + 131, + 116, + 479, + 128 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 197, + 141, + 480, + 166 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 141, + 480, + 166 + ], + "spans": [ + { + "bbox": [ + 197, + 141, + 480, + 166 + ], + "type": "interline_equation", + "content": "\\tau_ {k} ^ {u} = 1 + \\eta \\times \\left(\\frac {p _ {\\theta} \\left(y _ {u}\\right) _ {k}}{\\operatorname* {m a x} \\left(p _ {\\theta} \\left(y _ {u}\\right)\\right)} \\times C _ {u k} ^ {t}\\right), k = 1, \\dots , K. \\tag {12}", + "image_path": "cd7da58b405472835d5b86871871eaa1201993c22eb1551bef5800f1f59477da.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 173, + 479, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 173, + 479, + 198 + ], + "spans": [ + { + "bbox": [ + 131, + 173, + 479, + 198 + ], + "type": "text", + "content": "In Eq. (12), " + }, + { + "bbox": [ + 131, + 173, + 479, + 198 + ], + "type": "inline_equation", + "content": "C_{uk}" + }, + { + "bbox": [ + 131, + 173, + 479, + 198 + ], + "type": "text", + "content": " is an EMA of confidence for unlabeled samples predicted as the " + }, + { + "bbox": [ + 131, + 173, + 479, + 198 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 173, + 479, + 198 + ], + "type": "text", + "content": "th class, which is updated for each iteration of the training process as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 230, + 208, + 480, + 222 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 208, + 480, + 222 + ], + "spans": [ + { + "bbox": [ + 230, + 208, + 480, + 222 + ], + "type": "interline_equation", + "content": "C _ {u k} = \\omega \\times C _ {u k} ^ {\\text {o l d}} + (1 - \\omega) \\times C _ {u k} ^ {\\text {b a t c h}}, \\tag {13}", + "image_path": "46d3d12373b5ea570ab391543fd858f99174d4fa68e3b64b47cdb9f7df15927c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 232, + 480, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 232, + 480, + 295 + ], + "spans": [ + { + "bbox": [ + 131, + 232, + 480, + 295 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 232, + 480, + 295 + ], + "type": "inline_equation", + "content": "C_{uk}^{batch} = \\frac{1}{m_k}\\sum_{b = 1}^{B_u}f_{abc}(\\xi (\\alpha (u_b^m)))_k\\times \\mathbb{1}(\\hat{q}_b^m = k)" + }, + { + "bbox": [ + 131, + 232, + 480, + 295 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 131, + 232, + 480, + 295 + ], + "type": "inline_equation", + "content": "m_{k} = \\sum_{b = 1}^{B_{u}}\\mathbb{1}(\\hat{q}_{b}^{m} = k)" + }, + { + "bbox": [ + 131, + 232, + 480, + 295 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 232, + 480, + 295 + ], + "type": "inline_equation", + "content": "\\hat{q}_b^m = \\arg \\max_k(q_b^m)_k" + }, + { + "bbox": [ + 131, + 232, + 480, + 295 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 232, + 480, + 295 + ], + "type": "inline_equation", + "content": "C_{uk}" + }, + { + "bbox": [ + 131, + 232, + 480, + 295 + ], + "type": "text", + "content": " is updated only when " + }, + { + "bbox": [ + 131, + 232, + 480, + 295 + ], + "type": "inline_equation", + "content": "m_{k}\\geq 1" + }, + { + "bbox": [ + 131, + 232, + 480, + 295 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 131, + 232, + 480, + 295 + ], + "type": "inline_equation", + "content": "\\tau_k^u" + }, + { + "bbox": [ + 131, + 232, + 480, + 295 + ], + "type": "text", + "content": " incorporates the unknown class distribution of the unlabeled set, AAFM can densify feature clusters of minority classes to an appropriate extent even when the class distributions of the labeled and unlabeled sets significantly mismatch." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 319, + 310, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 310, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 310, + 331 + ], + "type": "text", + "content": "4.3 End-to-end training of RECD" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 344, + 479, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 344, + 479, + 380 + ], + "spans": [ + { + "bbox": [ + 131, + 344, + 479, + 380 + ], + "type": "text", + "content": "Fig. 2 shows the end-to-end training process of the proposed algorithm for labeled and unlabeled data, respectively. The total training loss of RECD, denoted as " + }, + { + "bbox": [ + 131, + 344, + 479, + 380 + ], + "type": "inline_equation", + "content": "L_{RECD}" + }, + { + "bbox": [ + 131, + 344, + 479, + 380 + ], + "type": "text", + "content": ", closely resembles that of ABC in Eq. (4), and is formulated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 240, + 392, + 480, + 404 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 392, + 480, + 404 + ], + "spans": [ + { + "bbox": [ + 240, + 392, + 480, + 404 + ], + "type": "interline_equation", + "content": "L _ {R E C D} = L _ {c l s} + L _ {c o n} + L _ {b a c k}, \\tag {14}", + "image_path": "da07071ff4006909680b8ab0989ec604f9b7518393a1070f8228456681bf482a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 415, + 471, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 415, + 471, + 428 + ], + "spans": [ + { + "bbox": [ + 131, + 415, + 471, + 428 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 415, + 471, + 428 + ], + "type": "inline_equation", + "content": "L_{cls}" + }, + { + "bbox": [ + 131, + 415, + 471, + 428 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 415, + 471, + 428 + ], + "type": "inline_equation", + "content": "L_{con}" + }, + { + "bbox": [ + 131, + 415, + 471, + 428 + ], + "type": "text", + "content": " are improved by applying RECD and AAFM as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 196, + 440, + 480, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 440, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 196, + 440, + 480, + 471 + ], + "type": "interline_equation", + "content": "L _ {c l s} = \\frac {1}{B _ {l}} \\sum_ {b = 1} ^ {B _ {l}} M \\left(x _ {b} ^ {m}\\right) \\mathcal {H} \\left(p _ {b} ^ {m}, f _ {a b c} \\left(\\xi \\left(\\alpha \\left(x _ {b} ^ {m}\\right)\\right) \\times \\tau_ {y _ {b} ^ {m}} ^ {l}\\right)\\right), \\tag {15}", + "image_path": "57992a72d9889de233368469e6a48fe69f15e110353b12e103c53f091354611a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 184, + 488, + 480, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 488, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 184, + 488, + 480, + 520 + ], + "type": "interline_equation", + "content": "L _ {c o n} = \\frac {1}{B _ {u}} \\sum_ {b = 1} ^ {B _ {u}} \\sum_ {i = 1} ^ {2} M \\left(u _ {b} ^ {m}\\right) \\mathcal {H} \\left(q _ {b} ^ {m}, f _ {a b c} \\left(\\xi \\left(\\mathcal {A} _ {i} \\left(u _ {b} ^ {m}\\right)\\right) \\times \\tau_ {\\hat {q} _ {b} ^ {m}} ^ {u}\\right)\\right), \\tag {16}", + "image_path": "884719891b5397c649859bfd0e55b170d986ba35b089a6fa6be3186a2097ba4e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 188, + 533, + 480, + 560 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 533, + 480, + 560 + ], + "spans": [ + { + "bbox": [ + 188, + 533, + 480, + 560 + ], + "type": "interline_equation", + "content": "M \\left(x _ {b} ^ {m}\\right) = \\mathcal {B} \\left(\\frac {N _ {K}}{N _ {y _ {b} ^ {m}}}\\right) \\quad \\text {a n d} \\quad M \\left(u _ {b} ^ {m}\\right) = \\mathcal {B} \\left(\\frac {\\min \\left(p _ {\\theta} \\left(y _ {u}\\right)\\right)}{p _ {\\theta} \\left(y _ {u}\\right) \\hat {q} _ {b} ^ {m}}\\right), \\tag {17}", + "image_path": "630d2a01e1665090337eb10cfa3da89abceb5fdab9bf1a0881d0fe1e8ec21651.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 568, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 568, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 568, + 481, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 568, + 481, + 665 + ], + "type": "inline_equation", + "content": "p_b^m" + }, + { + "bbox": [ + 131, + 568, + 481, + 665 + ], + "type": "text", + "content": " is the one-hot version of " + }, + { + "bbox": [ + 131, + 568, + 481, + 665 + ], + "type": "inline_equation", + "content": "y_b^m" + }, + { + "bbox": [ + 131, + 568, + 481, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 568, + 481, + 665 + ], + "type": "inline_equation", + "content": "q_b^m = f_{abc}(\\xi(\\alpha(u_b^m)))" + }, + { + "bbox": [ + 131, + 568, + 481, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 568, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\tau_{y_b^m}^l" + }, + { + "bbox": [ + 131, + 568, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 568, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\tau_{\\hat{q}_b^m}^u" + }, + { + "bbox": [ + 131, + 568, + 481, + 665 + ], + "type": "text", + "content": " are AAFMs corresponding to " + }, + { + "bbox": [ + 131, + 568, + 481, + 665 + ], + "type": "inline_equation", + "content": "y_b^m" + }, + { + "bbox": [ + 131, + 568, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 568, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\hat{q}_b^m" + }, + { + "bbox": [ + 131, + 568, + 481, + 665 + ], + "type": "text", + "content": ", respectively. To improve the quality of features learned by the backbone SSL algorithm, " + }, + { + "bbox": [ + 131, + 568, + 481, + 665 + ], + "type": "inline_equation", + "content": "L_{back}" + }, + { + "bbox": [ + 131, + 568, + 481, + 665 + ], + "type": "text", + "content": " is also adjusted with two modifications. First, AAFM is applied to further densify feature clusters of minority classes during the learning of the backbone SSL algorithm. Second, to train the backbone algorithm with unbiased pseudo-labels, pseudo-labels of the backbone algorithm are generated by " + }, + { + "bbox": [ + 131, + 568, + 481, + 665 + ], + "type": "inline_equation", + "content": "f_{abc}" + }, + { + "bbox": [ + 131, + 568, + 481, + 665 + ], + "type": "text", + "content": " instead of the classifier of the backbone algorithm. The pseudo-code of RECD is presented in Appendix D." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "RECD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 113, + 318, + 204 + ], + "blocks": [ + { + "bbox": [ + 167, + 113, + 318, + 204 + ], + "lines": [ + { + "bbox": [ + 167, + 113, + 318, + 204 + ], + "spans": [ + { + "bbox": [ + 167, + 113, + 318, + 204 + ], + "type": "image", + "image_path": "1fc43e714bf55d86b81286c90986c5acdd3f53a84fe6f24b363135a6d6a80d50.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 185, + 205, + 301, + 213 + ], + "lines": [ + { + "bbox": [ + 185, + 205, + 301, + 213 + ], + "spans": [ + { + "bbox": [ + 185, + 205, + 301, + 213 + ], + "type": "text", + "content": "(a) Balanced training with labeled data" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 322, + 113, + 494, + 205 + ], + "blocks": [ + { + "bbox": [ + 322, + 113, + 494, + 205 + ], + "lines": [ + { + "bbox": [ + 322, + 113, + 494, + 205 + ], + "spans": [ + { + "bbox": [ + 322, + 113, + 494, + 205 + ], + "type": "image", + "image_path": "fa73259374f230d43d91cee017ef1b4125b3aeaac7bf67b8dda584b3d0191815.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 205, + 449, + 213 + ], + "lines": [ + { + "bbox": [ + 326, + 205, + 449, + 213 + ], + "spans": [ + { + "bbox": [ + 326, + 205, + 449, + 213 + ], + "type": "text", + "content": "(b) Balanced training with unlabeled data" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 195, + 221, + 417, + 232 + ], + "lines": [ + { + "bbox": [ + 195, + 221, + 417, + 232 + ], + "spans": [ + { + "bbox": [ + 195, + 221, + 417, + 232 + ], + "type": "text", + "content": "Fig. 2: End-to-end training of the proposed algorithm" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 255, + 229, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 255, + 229, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 229, + 269 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 259, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 259, + 294 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 259, + 294 + ], + "type": "text", + "content": "5.1 Experimental setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 302, + 482, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 482, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 482, + 447 + ], + "type": "text", + "content": "Following previous CISSL studies [11, 20, 23, 34, 40], we conducted experiments on CIFAR-10-LT [8], CIFAR-100-LT [8], STL10-LT [20], and Small-ImageNet-127 [11]. The detailed descriptions for the datasets are provided in Appendix E.1. We evaluated the classification performance using balanced accuracy (bACC) (referred to as the averaged class recall in [11]) [15] and geometric mean (GM) [22]. The performance of the proposed algorithm was compared with the following baseline algorithms: vanilla algorithm, CIL algorithms [3, 16, 19], SSL algorithms [1, 38], and CISSL algorithms [11, 14, 20, 23, 24, 27, 40, 44, 46]. Algorithms such as DASO [34] and ACR [47], whose classification performances were measured under different experimental settings from ours, were compared with RECD in Appendix G. The detailed descriptions of the baseline algorithms and training setups are provided in Appendix E.2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 464, + 265, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 265, + 477 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 265, + 477 + ], + "type": "text", + "content": "5.2 Experimental results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 485, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 605 + ], + "type": "text", + "content": "Tab. 1 summarizes the classification performances of the baseline algorithms and RECD on CIFAR-10-LT under " + }, + { + "bbox": [ + 130, + 485, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\gamma_l = \\gamma_u" + }, + { + "bbox": [ + 130, + 485, + 482, + 605 + ], + "type": "text", + "content": ". First, the CISSL algorithms achieved higher performances than the vanilla algorithm, CIL algorithms, and SSL algorithms. These results verify the importance of mitigating class imbalance and utilizing unlabeled samples. Furthermore, RECD achieved the highest performance among the CISSL algorithms. Specifically, RECD showed a tendency to outperform other CISSL algorithms by a large margin as the class imbalance ratio increased, indicating its effectiveness in mitigating class imbalance. These results may be attributed to the unique ability of the proposed algorithm to mitigate class imbalance in the feature map, as we analyze in Sec. 5.3." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Tab. 2 summarizes the performances of the baseline CISSL algorithms and RECD on CIFAR-100-LT. RECD outperformed the baseline algorithms. Given that CIFAR-100-LT has 100 classes, these results suggest that RECD may also be suitable for datasets with a large number of classes. Considering that classes with the fewest labeled data points have only three labeled samples under " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\gamma = 50" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "text", + "content": "Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 170, + 136, + 446, + 315 + ], + "blocks": [ + { + "bbox": [ + 157, + 114, + 456, + 126 + ], + "lines": [ + { + "bbox": [ + 157, + 114, + 456, + 126 + ], + "spans": [ + { + "bbox": [ + 157, + 114, + 456, + 126 + ], + "type": "text", + "content": "Table 1: Comparison of bACC/GM on CIFAR-10-LT under " + }, + { + "bbox": [ + 157, + 114, + 456, + 126 + ], + "type": "inline_equation", + "content": "\\gamma = {\\gamma }_{l} = {\\gamma }_{u}" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 170, + 136, + 446, + 315 + ], + "lines": [ + { + "bbox": [ + 170, + 136, + 446, + 315 + ], + "spans": [ + { + "bbox": [ + 170, + 136, + 446, + 315 + ], + "type": "table", + "html": "
CIFAR-10-LT (γ = γt = γu)
Algorithmγ = 50γ = 100γ = 150
Vanilla (Cross-Entropy Loss)65.2±0.05/61.1±0.0958.8±0.13/58.2±0.1155.6±0.43/
Re-sampling [17]64.3±0.48/60.6±0.6755.8±0.47/45.1±0.3052.2±0.05/
LDAM-DRV [3]68.9±0.07/67.0±0.0862.8±0.17/58.9±0.6057.9±0.20/
cRT [19]67.8±0.13/66.3±0.1563.2±0.45/59.9±0.4059.3±0.10/
FixMatch [38]79.2±0.33/77.8±0.3671.5±0.72/66.8±1.5168.4±0.15/
FixMatch+DARP+cRT [19, 20]85.8±0.43/85.6±0.5682.4±0.26/81.8±0.1779.6±0.42/
FixMatch+CRST+LA [32, 46]85.6±0.36/81.9±0.4581.2±0.70/74.5±0.9971.9±2.24/
FixMatch+ABC [27]85.6±0.26/85.2±0.2981.1±1.14/80.3±1.2977.3±1.25/
FixMatch+CoSSL [11]86.8±0.30/86.6±0.2583.2±0.49/82.7±0.6080.3±0.55/
FixMatch+SAW+LA [23, 32]86.2±0.15/83.9±0.3580.7±0.15/77.5±0.2173.7±0.06/
FixMatch+Adsh [14]83.4±0.06/-76.5±0.35/-71.5±0.30/
FixMatch+DebiasPL [44]- / --80.6±0.50/-- / -
FixMatch+UDAL [24]86.5±0.29/-81.4±0.39/-77.9±0.33/
FixMatch+L2AC [40]- / --82.1±0.57/81.5±0.6477.6±0.53/
FixMatch+RECD87.3±0.18/87.2±0.1984.0±0.13/83.6±0.1680.6±0.53/
ReMixMatch [1]81.5±0.26/80.2±0.3273.8±0.38/69.5±0.8469.9±0.47/
ReMixMatch+DARP+cRT [19, 20]87.3±0.61/87.0±0.1183.5±0.07/83.1±0.0979.7±0.54/
ReMixMatch+CRST+LA [32, 46]84.2±0.11/-81.3±0.34/-79.2±0.31/
ReMixMatch+ABC [27]87.9±0.47/87.6±0.5184.5±0.32/84.1±0.3680.5±1.18/
ReMixMatch+CoSSL [11]87.7±0.21/87.6±0.2584.1±0.56/83.7±0.6681.3±0.83/
ReMixMatch+SAW+cRT [19, 23]87.6±0.21/87.4±0.2685.4±0.32/83.9±0.2179.9±0.15/
ReMixMatch+RECD88.1±0.13/87.9±0.1285.4±0.40/85.2±0.4182.5±0.33/
", + "image_path": "733078891b924ee6580e601c2cba91670d16f9284a32e273df4c6fba53625abe.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 340, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 480, + 365 + ], + "type": "text", + "content": "these results also demonstrate the possibility that RECD can outperform the baseline algorithms even when labeled samples are extremely scarce." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 365, + 480, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 480, + 449 + ], + "type": "text", + "content": "Tab. 3 summarizes the performances of the baseline CISSL algorithms and RECD on Small-ImageNet-127. For both " + }, + { + "bbox": [ + 130, + 365, + 480, + 449 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 130, + 365, + 480, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 365, + 480, + 449 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 130, + 365, + 480, + 449 + ], + "type": "text", + "content": " version, RECD significantly outperformed the baseline algorithms. Given that the class distribution of the test set of Small-ImageNet-127 is also imbalanced like the training set, these results show that RECD may also be suitable for datasets with imbalanced test sets. Moreover, given that Small-ImageNet-127 is an large-scale dataset, these results verify that RECD can also be applied to large-scale datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 450, + 481, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 481, + 475 + ], + "type": "text", + "content": "Tab. 4 summarizes the performances of the baseline algorithms and RECD evaluated on CIFAR-10-LT under " + }, + { + "bbox": [ + 130, + 450, + 481, + 475 + ], + "type": "inline_equation", + "content": "\\gamma_l \\neq \\gamma_u" + }, + { + "bbox": [ + 130, + 450, + 481, + 475 + ], + "type": "text", + "content": " and STL-10-LT under unknown" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 223, + 519, + 388, + 663 + ], + "blocks": [ + { + "bbox": [ + 165, + 498, + 448, + 510 + ], + "lines": [ + { + "bbox": [ + 165, + 498, + 448, + 510 + ], + "spans": [ + { + "bbox": [ + 165, + 498, + 448, + 510 + ], + "type": "text", + "content": "Table 2: Comparison of bACC on CIFAR-100-LT under " + }, + { + "bbox": [ + 165, + 498, + 448, + 510 + ], + "type": "inline_equation", + "content": "\\gamma = {\\gamma }_{l} = {\\gamma }_{u}" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 223, + 519, + 388, + 663 + ], + "lines": [ + { + "bbox": [ + 223, + 519, + 388, + 663 + ], + "spans": [ + { + "bbox": [ + 223, + 519, + 388, + 663 + ], + "type": "table", + "html": "
CIFAR-100-LT (γ = γt = γu)
Algorithmγ = 20γ = 50
FixMatch [38]49.6±0.7842.1±0.33
FixMatch+DARP [20]50.8±0.7743.1±0.54
FixMatch+DARP+cRT [19, 20]51.4±0.6844.9±0.54
FixMatch+CReST [46]51.8±0.1244.9±0.50
FixMatch+CReST+LA [32, 46]52.9±0.0747.3±0.17
FixMatch+ABC [27]53.3±0.7946.7±0.26
FixMatch+CoSSL [11]53.9±0.7847.6±0.57
FixMatch+RECD54.6±0.3647.8±0.17
ReMixMatch [1]51.6±0.4344.2±0.59
ReMixMatch+DARP [20]51.9±0.3544.7±0.66
ReMixMatch+DARP+cRT [19, 20]54.5±0.4248.5±0.91
ReMixMatch+CReST [46]51.3±0.3445.5±0.76
ReMixMatch+CReST+LA [32, 46]51.9±0.6046.6±1.14
ReMixMatch+ABC [27]55.6±0.3547.9±0.10
ReMixMatch+CoSSL [11]55.8±0.6248.9±0.61
ReMixMatch+RECD55.9±0.3649.5±0.21
", + "image_path": "f97159c492961423c0379a77cc655d04a396e837eedb34cd8b97a4fd14fcafbc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "RECD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 231, + 135, + 379, + 209 + ], + "blocks": [ + { + "bbox": [ + 192, + 114, + 421, + 126 + ], + "lines": [ + { + "bbox": [ + 192, + 114, + 421, + 126 + ], + "spans": [ + { + "bbox": [ + 192, + 114, + 421, + 126 + ], + "type": "text", + "content": "Table 3: Comparison of bACC on Small-ImageNet-127" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 231, + 135, + 379, + 209 + ], + "lines": [ + { + "bbox": [ + 231, + 135, + 379, + 209 + ], + "spans": [ + { + "bbox": [ + 231, + 135, + 379, + 209 + ], + "type": "table", + "html": "
Small-ImageNet-127 (γ = γl = γu)
Algorithm32 × 3264 × 64
FixMatch [38]29.742.3
FixMatch+DARP [20]30.542.5
FixMatch+DARP +cRT [19,20]39.751.0
FixMatch+CReST+ [46]32.544.7
FixMatch+CReST++LA [32,46]40.955.9
FixMatch+CoSSL [11]43.753.8
FixMatch+RECD47.359.5
", + "image_path": "50e9c35407365d3b9af57d5b7c614d858a50b4d852c8c67d1a8d187a7307573a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 240, + 482, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 240, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\gamma_{u}" + }, + { + "bbox": [ + 130, + 240, + 482, + 384 + ], + "type": "text", + "content": ". RECD outperformed the baseline algorithms by large margins. These results demonstrate that RECD effectively mitigates class imbalance even when " + }, + { + "bbox": [ + 130, + 240, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\gamma_{l} \\neq \\gamma_{u}" + }, + { + "bbox": [ + 130, + 240, + 482, + 384 + ], + "type": "text", + "content": ", by considering the unknown class distribution of the unlabeled set. Notably, ReMixMatch+RECD outperformed the baseline algorithms that use ReMixMatch* [20] (ReMixMatch that uses the estimated class distribution of the unlabeled set) as the backbone SSL algorithm. These results implicitly demonstrate that RECD estimates class distribution better than the previous estimation algorithm. Additionally, under severe class distribution mismatch between the labeled and unlabeled sets, LA degraded the performance of ReMixMatch*+DARP and ReMixMatch*+SAW. This may be because the unknown class distribution of the unlabeled set cannot be incorporated into LA, emphasizing the importance of considering the class distribution of the unlabeled set." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 134, + 436, + 481, + 586 + ], + "blocks": [ + { + "bbox": [ + 132, + 415, + 481, + 427 + ], + "lines": [ + { + "bbox": [ + 132, + 415, + 481, + 427 + ], + "spans": [ + { + "bbox": [ + 132, + 415, + 481, + 427 + ], + "type": "text", + "content": "Table 4: Comparison of bACC/GM on CIFAR-10-LT and STL-10-LT under " + }, + { + "bbox": [ + 132, + 415, + 481, + 427 + ], + "type": "inline_equation", + "content": "\\gamma_l \\neq \\gamma_u" + }, + { + "bbox": [ + 132, + 415, + 481, + 427 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 436, + 481, + 586 + ], + "lines": [ + { + "bbox": [ + 134, + 436, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 134, + 436, + 481, + 586 + ], + "type": "table", + "html": "
CIFAR-10-LT (γl=100)STL-10-LT (γu=Unknown)
Algorithmγu=1γu=50γu=150γl=10γl=20
FixMatch [38]68.9±1.95/42.8±8.1173.9±0.25/70.5±0.5269.6±0.60/62.6±1.1172.9±0.09/69.6±0.0163.4±0.21/52.6±0.09
+DARP [20]85.4±0.5585.0±0.6577.3±0.17/75.5±0.2172.9±0.24/69.5±0.1877.8±0.33/76.5±0.4069.9±1.77/65.4±0.97
+DARP+LA [20,32]86.6±1.11/86.2±1.1582.3±0.32/81.5±0.2978.9±0.23/77.7±0.0678.6±0.30/77.4±0.4071.9±0.49/68.7±0.51
+DARP+cRT [19,20]87.0±0.70/86.8±0.6782.7±0.21/82.3±0.2580.7±0.44/80.2±0.6179.3±0.23/78.7±0.2174.1±0.61/73.1±1.21
+ABC [27]82.7±0.49/81.9±0.6882.7±0.64/82.0±0.7678.4±0.87/77.2±1.0779.1±0.46/78.1±0.5773.8±0.15/72.1±0.15
+SAW [23]81.2±0.68/80.2±0.9179.8±0.25/79.1±0.3274.5±0.97/72.5±1.3778.3±0.25/77.0±0.1971.9±0.81/69.0±0.81
+SAW+LA [23,32]84.5±0.68/84.1±0.7882.9±0.38/82.6±0.3879.1±0.81/78.6±0.91-/--/-
+SAW+cRT [19,23]84.6±0.23/84.4±0.2681.6±0.38/81.3±0.3277.6±0.40/77.1±0.41-/--/-
+RECD90.2±0.57/90.0±0.6485.6±0.15/85.3±0.1582.3±0.39/81.8±0.4581.4±0.28/80.6±0.3879.0±0.48/78.1±0.54
ReMixMatch [1]48.3±0.14/19.5±0.8575.1±0.43/71.9±0.7772.5±0.10/68.2±0.3267.8±0.45/61.1±0.9260.1±1.18/44.9±1.52
+ABC [27]76.4±5.34/74.8±6.0585.2±0.20/84.7±0.2580.4±0.40/80.0±0.4476.8±0.52/74.8±0.6471.2±1.37/67.4±1.89
ReMixMatch* [1]85.0±1.35/84.3±1.5577.0±0.12/74.7±0.0472.8±0.10/68.8±0.2176.7±0.15/73.9±0.3267.7±0.46/60.3±0.76
+DARP [20]86.9±0.10/86.4±0.1577.4±0.22/75.0±0.2573.2±0.11/69.2±0.3179.4±0.07/78.2±0.1070.9±0.44/67.0±1.62
+DARP+LA [20,32]81.8±0.45/80.9±0.4083.9±0.42/83.4±0.4581.1±0.20/80.3±0.2680.6±0.45/79.6±0.5576.8±0.60/74.8±0.68
+DARP+cRT [19,20]88.7±0.25/88.5±0.2583.5±0.53/83.1±0.5180.9±0.25/80.3±0.3180.9±0.53/80.0±0.4676.7±0.50/74.9±0.70
+SAW [23]87.0±0.75/86.4±0.8580.6±1.57/79.2±1.1977.6±0.76/76.0±0.9382.0±0.55/81.0±0.6479.2±0.44/77.9±0.52
+SAW+LA [23,32]74.2±1.49/65.1±2.3684.8±1.07/82.4±2.3281.3±2.42/80.9±2.47-/--/-
+SAW+cRT [19,23]88.8±0.79/88.6±0.8384.5±0.78/83.6±1.2782.4±0.10/82.0±0.10-/--/-
ReMixMatch+RECD90.3±0.40/90.2±0.4186.8±0.17/86.6±0.1883.9±0.11/83.7±0.1284.9±0.41/84.4±0.4782.5±0.27/81.7±0.31
", + "image_path": "4a64f8df6452e05e3a7e16af98c202f8e88e4fb1ab057904fa1b73cde81daca2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "In addition to the above results, we demonstrate the effectiveness of RECD with fewer labeled samples in Appendix J. Furthermore, we verify that RECD can be effectively combined with recent SSL algorithms by conducting experiments using FreeMatch [45] as the backbone SSL algorithm in Appendix I." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "type": "text", + "content": "Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 386, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 386, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 386, + 129 + ], + "type": "text", + "content": "5.3 Detailed analyses on the proposed algorithm" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 246 + ], + "type": "text", + "content": "We argued that RECD effectively estimates the unknown class distribution of the unlabeled set and mitigates class imbalance using this estimation. To verify this argument, we plotted the true class distribution of the unlabeled set, the class distribution of the unlabeled set estimated by RECD, and the class distribution of the labeled set in Fig. 3. We can observe that RECD accurately estimated the unknown class distribution of the unlabeled set, even when it significantly differed from that of the labeled set. This accurate class distribution estimation enabled the proposed algorithm to be trained with a properly rebalanced training loss and accurately calculated AAFM, mitigating class imbalance as in Fig. 4." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 173, + 266, + 299, + 347 + ], + "blocks": [ + { + "bbox": [ + 173, + 266, + 299, + 347 + ], + "lines": [ + { + "bbox": [ + 173, + 266, + 299, + 347 + ], + "spans": [ + { + "bbox": [ + 173, + 266, + 299, + 347 + ], + "type": "image", + "image_path": "a07408fa87cca9fa5fefce2e1db239452512b8ad72d069f97c1dfe26d9e9df3f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 197, + 349, + 276, + 358 + ], + "lines": [ + { + "bbox": [ + 197, + 349, + 276, + 358 + ], + "spans": [ + { + "bbox": [ + 197, + 349, + 276, + 358 + ], + "type": "text", + "content": "(a) FixMatch+RECD" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 313, + 266, + 443, + 347 + ], + "blocks": [ + { + "bbox": [ + 313, + 266, + 443, + 347 + ], + "lines": [ + { + "bbox": [ + 313, + 266, + 443, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 443, + 347 + ], + "type": "image", + "image_path": "72704aba64fa1b059fe50bbd0905d4c3b7df3f12579516ed56d8f36c223244ab.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 349, + 423, + 358 + ], + "lines": [ + { + "bbox": [ + 333, + 349, + 423, + 358 + ], + "spans": [ + { + "bbox": [ + 333, + 349, + 423, + 358 + ], + "type": "text", + "content": "(b) RemixMatch+RECD" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 368, + 482, + 401 + ], + "lines": [ + { + "bbox": [ + 130, + 368, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 482, + 401 + ], + "type": "text", + "content": "Fig. 3: True class distributions of the labeled and unlabeled sets, along with the class distribution of the unlabeled set estimated by RECD. FixMatch+RECD and Remix-Match+RECD were trained on CIFAR-10-LT with " + }, + { + "bbox": [ + 130, + 368, + 482, + 401 + ], + "type": "inline_equation", + "content": "\\gamma_{l} = 100" + }, + { + "bbox": [ + 130, + 368, + 482, + 401 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 368, + 482, + 401 + ], + "type": "inline_equation", + "content": "\\gamma_{u} = 1" + }, + { + "bbox": [ + 130, + 368, + 482, + 401 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 426, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 533 + ], + "type": "text", + "content": "Fig. 4 presents the confusion matrices of class predictions on the test set of CIFAR-10 generated by each algorithm trained on the training set of CIFAR-10-LT under " + }, + { + "bbox": [ + 130, + 426, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\gamma_l = 100" + }, + { + "bbox": [ + 130, + 426, + 482, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 426, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\gamma_u = 1" + }, + { + "bbox": [ + 130, + 426, + 482, + 533 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 130, + 426, + 482, + 533 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 130, + 426, + 482, + 533 + ], + "type": "text", + "content": " position of each confusion matrix represents the proportion of the " + }, + { + "bbox": [ + 130, + 426, + 482, + 533 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 426, + 482, + 533 + ], + "type": "text", + "content": "th class samples classified as the " + }, + { + "bbox": [ + 130, + 426, + 482, + 533 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 426, + 482, + 533 + ], + "type": "text", + "content": "th class. In the figures, the left three algorithms often misclassified minority classes as majority classes. On the other hand, with both RECD and AAFM, the algorithm classified most samples of minority classes accurately. These results demonstrate that both RECD and AAFM effectively mitigate class imbalance when the class distribution of the labeled and unlabeled sets significantly mismatch." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "Additionally, we argued that AAFM not only reduces the test losses for minority classes but also stabilizes the training process. To verify this, we measured the overall test loss and the test losses of minority classes for FixMatch+RECD without AAFM and FixMatch+RECD with AAFM trained on CIFAR-10-LT. From Fig. 5, we can observe that the test losses for minority classes tended to increase as the training proceeds when AAFM was not used. On the other hand, the overall test loss and the test losses for the minority classes consistently decreased when AAFM was used. We can also observe that whereas the test losses for minority classes fluctuated when AAFM was not used, they stably decreased when AAFM was used. In addition, we compared the performance of the naive feature multiplier and AAFM with FixMatch+RECD in Appendix K. Finally," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "text", + "content": "RECD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 147, + 114, + 227, + 194 + ], + "blocks": [ + { + "bbox": [ + 147, + 114, + 227, + 194 + ], + "lines": [ + { + "bbox": [ + 147, + 114, + 227, + 194 + ], + "spans": [ + { + "bbox": [ + 147, + 114, + 227, + 194 + ], + "type": "image", + "image_path": "f7455dbc685d9de10c884e4c212d2a1fd40ccaa9fb8896a8d37d959170fd0d6c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 163, + 195, + 211, + 204 + ], + "lines": [ + { + "bbox": [ + 163, + 195, + 211, + 204 + ], + "spans": [ + { + "bbox": [ + 163, + 195, + 211, + 204 + ], + "type": "text", + "content": "(a) FixMatch" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 228, + 114, + 307, + 192 + ], + "blocks": [ + { + "bbox": [ + 228, + 114, + 307, + 192 + ], + "lines": [ + { + "bbox": [ + 228, + 114, + 307, + 192 + ], + "spans": [ + { + "bbox": [ + 228, + 114, + 307, + 192 + ], + "type": "image", + "image_path": "0bb3d89f794a8e74e448bab48f42a7a0336aa390bf8b71a5a020a3af1ef82366.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 194, + 302, + 204 + ], + "lines": [ + { + "bbox": [ + 233, + 194, + 302, + 204 + ], + "spans": [ + { + "bbox": [ + 233, + 194, + 302, + 204 + ], + "type": "text", + "content": "(b) Without RECD" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 388, + 192 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 388, + 192 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 388, + 192 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 388, + 192 + ], + "type": "image", + "image_path": "f0fee412f9bae0708ed2d5bfd441210ae21eb645a618d3edc9f116ffc1742a29.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 194, + 384, + 204 + ], + "lines": [ + { + "bbox": [ + 315, + 194, + 384, + 204 + ], + "spans": [ + { + "bbox": [ + 315, + 194, + 384, + 204 + ], + "type": "text", + "content": "(c) Without AAFM" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 213, + 481, + 247 + ], + "lines": [ + { + "bbox": [ + 130, + 213, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 481, + 247 + ], + "type": "text", + "content": "Fig. 4: Confusion matrices of class predictions on the test set of CIFAR-10 generated by each algorithm (FixMatch+RECD with the described condition) trained on the training set of CIFAR-10-LT under " + }, + { + "bbox": [ + 130, + 213, + 481, + 247 + ], + "type": "inline_equation", + "content": "\\gamma_l = 100" + }, + { + "bbox": [ + 130, + 213, + 481, + 247 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 213, + 481, + 247 + ], + "type": "inline_equation", + "content": "\\gamma_u = 1" + }, + { + "bbox": [ + 130, + 213, + 481, + 247 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 391, + 114, + 470, + 192 + ], + "blocks": [ + { + "bbox": [ + 391, + 114, + 470, + 192 + ], + "lines": [ + { + "bbox": [ + 391, + 114, + 470, + 192 + ], + "spans": [ + { + "bbox": [ + 391, + 114, + 470, + 192 + ], + "type": "image", + "image_path": "a61fb43238383e5559756ccb71379879588fbe06a6b2f25a11640c80cf14345e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 396, + 194, + 464, + 204 + ], + "lines": [ + { + "bbox": [ + 396, + 194, + 464, + 204 + ], + "spans": [ + { + "bbox": [ + 396, + 194, + 464, + 204 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 396, + 194, + 464, + 204 + ], + "type": "inline_equation", + "content": "\\mathrm{RECD + AAFM}" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 268, + 480, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 480, + 304 + ], + "type": "text", + "content": "we argue that RECD also can be combined with various CISSL algorithms. To verify this, we conducted additional experiments using DARP, and demonstrated RECD-based DARP outperformed the original DARP in Appendix F." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 172, + 323, + 299, + 404 + ], + "blocks": [ + { + "bbox": [ + 172, + 323, + 299, + 404 + ], + "lines": [ + { + "bbox": [ + 172, + 323, + 299, + 404 + ], + "spans": [ + { + "bbox": [ + 172, + 323, + 299, + 404 + ], + "type": "image", + "image_path": "f3c70fd891ef136643eb01f0e2478815210e2f0d919dce16685c4262a38ba34e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 406, + 272, + 415 + ], + "lines": [ + { + "bbox": [ + 200, + 406, + 272, + 415 + ], + "spans": [ + { + "bbox": [ + 200, + 406, + 272, + 415 + ], + "type": "text", + "content": "(a) WithoutAAFM" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 313, + 323, + 442, + 405 + ], + "blocks": [ + { + "bbox": [ + 313, + 323, + 442, + 405 + ], + "lines": [ + { + "bbox": [ + 313, + 323, + 442, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 442, + 405 + ], + "type": "image", + "image_path": "9a34d14dfc37409986f4662696f612674b20f6cb24f466f44396d2a0f19842f8.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 406, + 408, + 416 + ], + "lines": [ + { + "bbox": [ + 347, + 406, + 408, + 416 + ], + "spans": [ + { + "bbox": [ + 347, + 406, + 408, + 416 + ], + "type": "text", + "content": "(b) WithAAFM" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 425, + 480, + 448 + ], + "lines": [ + { + "bbox": [ + 130, + 425, + 480, + 448 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 480, + 448 + ], + "type": "text", + "content": "Fig.5: Overall test loss and test losses for minority classes (Classes 8 and 9). The algorithms were trained on CIFAR-10-LT under " + }, + { + "bbox": [ + 130, + 425, + 480, + 448 + ], + "type": "inline_equation", + "content": "\\gamma_{l} = 100" + }, + { + "bbox": [ + 130, + 425, + 480, + 448 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 425, + 480, + 448 + ], + "type": "inline_equation", + "content": "\\gamma_{u} = 1" + }, + { + "bbox": [ + 130, + 425, + 480, + 448 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 486, + 220, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 486, + 220, + 499 + ], + "spans": [ + { + "bbox": [ + 131, + 486, + 220, + 499 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "We proposed a novel CISSL algorithm, RECD, which addresses the class distribution mismatch between labeled and unlabeled sets. RECD estimates the class distribution of the unlabeled set through MC approximation and uses it for rebalancing the classifier. Additionally, RECD reuses the estimated class distribution for AAFM to mitigate class imbalance in the feature map by densifying feature clusters of minority classes. Experimental results demonstrate that RECD outperforms the baseline algorithms under both scenarios where the class distributions of the labeled and unlabeled sets match and mismatch. Detailed analyses on RECD verify that the effectiveness of each component of RECD for mitigating class imbalance. However, the estimated class distribution of unlabeled set through MC approximation might be inaccurate when unlabeled training samples are scarce. In future work, we aim to improve the robustness of class distribution estimation, even with a limited amount of unlabeled samples." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "text", + "content": "Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 176 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 176 + ], + "type": "text", + "content": "This research was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (2023R1A2C2005453, RS-2023-00218913)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 196, + 197, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 196, + 197, + 209 + ], + "spans": [ + { + "bbox": [ + 133, + 196, + 197, + 209 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 222, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 222, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 222, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 222, + 480, + 255 + ], + "type": "text", + "content": "1. Berthelot, D., Carlini, N., an Alex Kurakin, E.D.C., Sohn, K., Zhang, H., Raffel, C.: Remixmatch: Semi-supervised learning with distribution matching and augmentation anchoring. In: ICLR (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 256, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 256, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 480, + 277 + ], + "type": "text", + "content": "2. Berthelot, D., Carlini, N., Goodfellow, I., Papernot, N., Oliver, A., Raffel, C.: Mixmatch: A holistic approach to semi-supervised learning. In: NeurIPS (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 278, + 480, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 278, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 138, + 278, + 480, + 322 + ], + "type": "text", + "content": "3. Cao, K., Wei, C., Gaidon, A., Arechiga, N., Ma, T.: Learning imbalanced datasets with label-distribution-aware margin loss. In: Wallach, H., Larochelle, H., Beygelzimer, A., d'Alché-Buc, F., Fox, E., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 32. Curran Associates, Inc. (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 323, + 480, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 323, + 480, + 355 + ], + "spans": [ + { + "bbox": [ + 138, + 323, + 480, + 355 + ], + "type": "text", + "content": "4. Chawla, N.V., Bowyer, K.W., Hall, L.O., Kegelmeyer, W.P.: Smote: Synthetic minority over-sampling technique. Journal of Artificial Intelligence Research pp. 321-357 (2002)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 355, + 480, + 377 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 355, + 480, + 377 + ], + "spans": [ + { + "bbox": [ + 138, + 355, + 480, + 377 + ], + "type": "text", + "content": "5. Chu, P., Bian, X., Liu, S., Ling, H.: Feature space augmentation for long-tailed data. In: ECCV. vol. 12374, pp. 694-710 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 378, + 480, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 378, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 138, + 378, + 480, + 399 + ], + "type": "text", + "content": "6. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.: Randaugment: Practical automated data augmentation with a reduced search space. In: NeurIPS (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 399, + 480, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 480, + 422 + ], + "type": "text", + "content": "7. Cui, Y., Jia, M., Lin, T.Y., Song, Y., Belongie, S.: Large scale fine-grained categorization and domain-specific transfer learning. In: CVPR. pp. 4109-4118 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 422, + 480, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 422, + 480, + 444 + ], + "spans": [ + { + "bbox": [ + 138, + 422, + 480, + 444 + ], + "type": "text", + "content": "8. Cui, Y., Jia, M., Lin, T.Y., Song, Y., Belongie, S.: Class-balanced loss based on effective number of samples. In: CVPR. pp. 9268-9277 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 444, + 480, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 444, + 480, + 466 + ], + "spans": [ + { + "bbox": [ + 138, + 444, + 480, + 466 + ], + "type": "text", + "content": "9. DeVries, T., Taylor, G.W.: Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 466, + 480, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 466, + 480, + 488 + ], + "spans": [ + { + "bbox": [ + 138, + 466, + 480, + 488 + ], + "type": "text", + "content": "10. Estabrooks, A., Jo, T., Japkowicz, N.: A multiple resampling method for learning from imbalanced data sets. Computational Intelligence 20(1), 18-36 (2004)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 488, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 488, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 488, + 480, + 521 + ], + "type": "text", + "content": "1. Fan, Y., Dai, D., Kukleva, A., Schiele, B.: Cossl: Co-learning of representation and classifier for imbalanced semi-supervised learning. In: CVPR. pp. 14574-14584 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 521, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 480, + 544 + ], + "type": "text", + "content": "2. Gidaris, S., Singh, P., Komodakis, N.: Unsupervised representation learning by predicting image rotations. In: ICLR (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 544, + 480, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 480, + 565 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 480, + 565 + ], + "type": "text", + "content": "3. Grandvalet, Y., Bengio, Y.: Semi-supervised learning by entropy minimization. In: NeurIPS (2004)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 565, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 565, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 565, + 480, + 588 + ], + "type": "text", + "content": "4. Guo, L.Z., Li, Y.F.: Class-imbalanced semi-supervised learning with adaptive thresholding. In: ICML. vol. 162, pp. 8082-8094 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 588, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 588, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 588, + 480, + 610 + ], + "type": "text", + "content": "5. Huang, C., Li, Y., Loy, C.C., Tang, X.: Learning deep representation for imbalanced classification. In: CVPR. pp. 5375-5384 (2016)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "text", + "content": "6. JAPKOWICZ, N.: The class imbalance problem: Significance and strategies. In: Proc. 2000 International Conference on Artificial Intelligence. vol. 1, pp. 111-117 (2000)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "7. Japkowicz, N.: The class imbalance problem: Significance and strategies. In: In Proceedings of the International Conference on Artificial Intelligence (2000)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "RECD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "18. Kang, B., Li, Y., Xie, S., Yuan, Z., Feng, J.: Exploring balanced feature spaces for representation learning. In: ICLR (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "text", + "content": "19. Kang, B., Xie, S., Rohrbach, M., Yan, Z., Gordo, A., Feng, J., Kalantidis, Y.: Decoupling representation and classifier for long-tailed recognition. In: ICLR (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 162, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 162, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 480, + 182 + ], + "type": "text", + "content": "20. Kim, J., Hur, Y., Park, S., Yang, E., Hwang, S., Shin, J.: Distribution aligning refinery of pseudo-label for imbalanced semi-supervised learning. In: NeurIPS (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 184, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 480, + 205 + ], + "type": "text", + "content": "21. Krizhevsky, A.: Learning multiple layers of features from tiny images. Technical report, Department of Computer Science, University of Toronto (2009)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 206, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 480, + 226 + ], + "type": "text", + "content": "22. Kubat, M., Matwin, S., et al.: Addressing the curse of imbalanced training sets: one-sided selection. In: Icml. vol. 97, p. 179. Citeseer (1997)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 228, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 480, + 258 + ], + "type": "text", + "content": "23. Lai, Z., Wang, C., Gunawan, H., Cheung, S.C., Chuah, C.N.: Smoothed adaptive weighting for imbalanced semi-supervised learning: Improve reliability against unknown distribution data. In: ICML. vol. 162, pp. 11828-11843 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 260, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 480, + 303 + ], + "type": "text", + "content": "24. Lazarow, J., Sohn, K., Lee, C.Y., Li, C.L., Zhang, Z., Pfister, T.: Unifying distribution alignment as a loss for imbalanced semi-supervised learning. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) pp. 5644-5653 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "text", + "content": "25. Lee, D.H.: Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In: In Workshop on challenges in representation learning (ICML) (2013)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 337, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 337, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 337, + 480, + 357 + ], + "type": "text", + "content": "26. Lee, H., Kim, H.: Cdmad: Class-distribution-mismatch-aware debiasing for class-imbalanced semi-supervised learning. In: CVPR. pp. 23891-23900 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 359, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 359, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 359, + 480, + 379 + ], + "type": "text", + "content": "27. Lee, H., Shin, S., Kim, H.: Abc: Auxiliary balanced classifier for class-imbalanced semi-supervised learning. In: NeurIPS (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "text", + "content": "28. Li, J., Meng, Z., Shi, D., Song, R., Diao, X., Wang, J., Xu, H.: Fcc: Feature clusters compression for long-tailed visual recognition. In: CVPR. pp. 24080-24089 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "text", + "content": "29. Li, K., Kong, X., Lu, Z., Wenyin, L., Yin, J.: Boosting weighted elm for imbalanced learning. Neurocomputing 128, 15-21 (2014)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "type": "text", + "content": "30. Liu, X.Y., Wu, J., Zhou, Z.H.: Exploratory undersampling for class-imbalance learning. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics) 39(2), 539-550 (2008)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 458, + 480, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 458, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 458, + 480, + 478 + ], + "type": "text", + "content": "31. Liu, Z., Miao, Z., Zhan, X., Wang, J., Gong, B., Yu, S.X.: Large-scale long-tailed recognition in an open world. In: CVPR. pp. 2537-2546 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "type": "text", + "content": "32. Menon, A.K., Jayasumana, S., Rawat, A.S., Jain, H., Veit, A., Kumar, S.: Long-tail learning via logit adjustment. In: ICLR (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "text", + "content": "33. Miyato, T., ichi Maeda, S., Koyama, M., Ishii, S.: Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE Transactions on Pattern Analysis and Machine Intelligence 41(8), 1979-1993 (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 535, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 535, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 535, + 480, + 566 + ], + "type": "text", + "content": "34. Oh, Y., Kim, D.J., Kweon, I.S.: Daso: Distribution-aware semantics-oriented pseudo-label for imbalanced semi-supervised learning. In: CVPR. pp. 9786-9796 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "text", + "content": "35. Ren, J., Yu, C., Sheng, S., Ma, X., Zhao, H., Yi, S., Li, H.: Balanced meta-softmax for long-tailed visual recognition. In: NeurIPS (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 590, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 610 + ], + "type": "text", + "content": "36. Ren, M., Zeng, W., Yang, B., Urtasun, R.: Learning to reweight examples for robust deep learning. In: ICML. vol. 80, pp. 4334-4343 (2018)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 612, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 612, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 612, + 480, + 632 + ], + "type": "text", + "content": "37. Sajjadi, M., Javanmardi, M., Tasdizen, T.: Regularization with stochastic transformations and perturbations for deep semi-supervised learning. In: NeurIPS (2016)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "38. Sohn, K., Berthelot, D., Li, C.L., Zhang, Z., Carlini, N., Cubuk, E.D., Kurakin, A., Zhang, H., Raffel, C.: Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In: NeurIPS (2020)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 210, + 100 + ], + "type": "text", + "content": "Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 512 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "39. Verma, V., Kawaguchi, K., Lamb, A., Kannala, J., Solin, A., Bengio, Y., Lopez-Paz, D.: Interpolation consistency training for semi-supervised learning. Neural Networks 145, 90-106 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "type": "text", + "content": "40. Wang, R., Jia, X., Wang, Q., Wu, Y., Meng, D.: Imbalanced semi-supervised learning with bias adaptive classifier. In: ICLR (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "text", + "content": "41. Wang, S., Minku, L.L., Yao, X.: A learning framework for online class imbalance learning. IEEE Symposium on Computational Intelligence and Ensemble Learning (CIEL) pp. 36-45 (2013)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "42. Wang, S., Minku, L.L., Yao, X.: Resampling-based ensemble methods for online class imbalance learning. IEEE Transactions on Knowledge and Data Engineering 27(5), 1356-1368 (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 259 + ], + "type": "text", + "content": "43. Wang, X., Lian, L., Miao, Z., Liu, Z., Yu, S.X.: Long-tailed recognition by routing diverse distribution-aware experts. In: ICLR (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 260, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 481, + 281 + ], + "type": "text", + "content": "44. Wang, X., Wu, Z., Lian, L., Yu, S.X.: Debiased learning from naturally imbalanced pseudo-labels. In: CVPR. pp. 14647-14657 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 282, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 282, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 133, + 282, + 481, + 314 + ], + "type": "text", + "content": "45. Wang, Y., Chen, H., Heng, Q., Hou, W., Fan, Y., Wu, Z., Wang, J., Savvides, M., Shinozaki, T., Raj, B., Schiele, B., Xie, X.: Freematch: Self-adaptive thresholding for semi-supervised learning. In: ICLR (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 315, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 315, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 133, + 315, + 481, + 346 + ], + "type": "text", + "content": "46. Wei, C., Sohn, K., Mellina, C., Yuille, A., Yang, F.: Crest: A class-rebalancing self-training framework for imbalanced semi-supervised learning. In: CVPR. pp. 10857-10866 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "text", + "content": "47. Wei, T., Gan, K.: Towards realistic long-tailed semi-supervised learning: Consistency is all you need. In: CVPR. pp. 3469-3478 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "type": "text", + "content": "48. Xiang, L., Ding, G., Han, J.: Learning from multiple experts: Self-paced knowledge distillation for long-tailed classification. In: ECCV. pp. 247-263 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 392, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 392, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 133, + 392, + 481, + 424 + ], + "type": "text", + "content": "49. Yu, F., Seff, A., Zhang, Y., Song, S., Funkhouser, T., Xiao, J.: Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365 (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 425, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 425, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 133, + 425, + 481, + 445 + ], + "type": "text", + "content": "50. Yu, Z., Li, Y., Lee, Y.J.: Inpl: Pseudo-labeling the inliers first for imbalanced semi-supervised learning. In: ICLR (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "type": "text", + "content": "51. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. In: ICLR (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "text", + "content": "52. Zhang, Y., Hooi, B., Hong, L., Feng, J.: Self-supervised aggregation of diverse experts for test-agnostic long-tailed recognition. In: NeurIPS (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 491, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 512 + ], + "type": "text", + "content": "53. Zhong, Z., Cui, J., Liu, S., Jia, J.: Improving calibration for long-tailed recognition. In: CVPR. pp. 16489-16498 (2021)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "RECD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file